diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_0/index_tensor([6954, 6954])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_0/index_tensor([6954, 6954])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..34ef0b6b8d6eccdad98eed1c51a6aaddfd77d298 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_0/index_tensor([6954, 6954])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This ultra-high-speed medium shot, captured from a static low-angle perspective, gazes upon the intricate, spiny underside of a massive Victoria Amazon water lily, capturing the scene in extreme slow motion. The entire frame is submerged in murky green-brown water, while filtered sunlight penetrates from above, creating an ethereal crepuscular glow that illuminates floating silt particles. The lily's surface is an intricate network of thick, reddish-purple vascular ribs capped with sharp spines, creating dramatic patterns of light and deep shadow. As a delicate stream of shimmering oxygen bubbles begins to condense from the plant's surface, a magnificent process of gas exchange unfolds. Each pearlescent bubble slowly forms, clinging precariously to the spiny texture in a tense moment, its perfectly spherical shape slightly distorted by surface tension before bursting. Then, one by one, they detach and begin a tranquil, silent ascent, rising through the hazy water like a string of tiny, glowing gems. As they float steadily and deliberately upward and out of the frame, they catch the overhead light, creating an atmosphere of profound, hidden wonder and nature's quiet grandeur.", + "seed": 18434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_0/index_6954_seed_18434_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040415395074523984, + "vlm_logprob": -0.040415395074523984, + "vlm_yes_no_margin": 4.1749999761581424, + "vlm_token_credit": -0.040415395074523984, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054057758301496506 + ], + "opposite_logprob_tokens": [ + -3.1790578365325928 + ], + "logprob_reward": -0.054057758301496506, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.054057758301496506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18824811279773712, + 0.05003443360328674, + 0.18814049661159515, + 0.08620981872081757, + 0.06346452981233597, + 0.13089078664779663, + 0.15820513665676117, + 0.06414086371660233, + 0.03639927878975868, + 0.034266531467437744 + ], + "credit_map": [ + 0.015271011739969254, + 0.023141682147979736, + 0.013169319368898869, + 0.031335413455963135, + 0.021738674491643906, + 0.03607115522027016, + 0.018494190648198128, + 0.029026668518781662, + 0.007020561955869198, + 0.006855063606053591, + 0.006516135763376951, + 0.006947794463485479, + 0.004154327791184187, + 0.005028022453188896, + 0.005928706843405962, + 0.007583819329738617, + 0.018298635259270668, + 0.02351568639278412, + 0.01991458050906658, + 0.011605419218540192, + 0.038748279213905334, + 0.02794366516172886, + 0.015762630850076675, + 0.032351598143577576, + 0.009164930321276188, + 0.01258180569857359, + 0.008121668361127377, + 0.007644866593182087, + 0.02013704925775528, + 0.01138144452124834, + 0.007246783934533596, + 0.009931262582540512, + 0.007469634059816599, + 0.009490418247878551, + 0.004296161234378815, + 0.006071855314075947, + 0.014954556711018085, + 0.01025741919875145, + 0.005283046513795853, + 0.005641444120556116, + 0.016558608040213585, + 0.029124999418854713, + 0.009727215394377708, + 0.013894678093492985, + 0.027548013255000114, + 0.02006266824901104, + 0.007851574569940567, + 0.0061230165883898735, + 0.022678274661302567, + 0.02504728175699711, + 0.00995508674532175, + 0.0074731153436005116, + 0.03977102413773537, + 0.026250101625919342, + 0.009636396542191505, + 0.017393846064805984, + 0.00959381926804781, + 0.006586809642612934, + 0.005655036773532629, + 0.0038081961683928967, + 0.01616247184574604, + 0.009853350929915905, + 0.006259473040699959, + 0.006221706513315439, + 0.004749880637973547, + 0.004930391907691956, + 0.0033241647761315107, + 0.0032392290886491537, + 0.004471521358937025, + 0.00597444036975503, + 0.00341204134747386, + 0.006297606509178877, + 0.002789426827803254, + 0.004148213658481836, + 0.0030696268659085035, + 0.0030955688562244177, + 0.008679159916937351, + 0.005221827421337366, + 0.003555367235094309, + 0.0037073399871587753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.054057758301496506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702758610248566 + ], + "opposite_logprob_tokens": [ + -2.1270275115966797 + ], + "logprob_reward": -0.12702758610248566, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12702758610248566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10597197711467743, + 0.06443285942077637, + 0.0684572383761406, + 0.2786697745323181, + 0.10612324625253677, + 0.0870281308889389, + 0.15683455765247345, + 0.09741412103176117, + 0.01679830625653267, + 0.018269803375005722 + ], + "credit_map": [ + 0.008446978405117989, + 0.015462804585695267, + 0.007685764227062464, + 0.02651802822947502, + 0.009932850487530231, + 0.01784537173807621, + 0.010105207562446594, + 0.009974981658160686, + 0.007565110921859741, + 0.011819321662187576, + 0.005694916937500238, + 0.004904504399746656, + 0.009416385553777218, + 0.009391024708747864, + 0.005957815796136856, + 0.009683789685368538, + 0.006103056017309427, + 0.013483262620866299, + 0.006164555437862873, + 0.006520762573927641, + 0.014481556601822376, + 0.010000187903642654, + 0.006586626637727022, + 0.0051172287203371525, + 0.026323478668928146, + 0.05621550977230072, + 0.013937784358859062, + 0.017113545909523964, + 0.07580193877220154, + 0.05063151568174362, + 0.017912473529577255, + 0.020733542740345, + 0.01512320525944233, + 0.01669844426214695, + 0.008581908419728279, + 0.00733767356723547, + 0.02605310268700123, + 0.01695280522108078, + 0.007402032148092985, + 0.007974074222147465, + 0.011145081371068954, + 0.017283206805586815, + 0.004565148614346981, + 0.010313997976481915, + 0.019033892080187798, + 0.01212681457400322, + 0.00735348928719759, + 0.0052064997144043446, + 0.023493997752666473, + 0.025234881788492203, + 0.009564600884914398, + 0.009963450022041798, + 0.037899140268564224, + 0.024549026042222977, + 0.008680603466928005, + 0.017448870465159416, + 0.014609077014029026, + 0.011897709220647812, + 0.007423032075166702, + 0.00454632518813014, + 0.025911767035722733, + 0.01573421061038971, + 0.008771792985498905, + 0.008520212024450302, + 0.0021039461717009544, + 0.0020142430439591408, + 0.0022034153807908297, + 0.0014336323365569115, + 0.001759490231052041, + 0.002142468001693487, + 0.0026392308063805103, + 0.002501879818737507, + 0.001824580947868526, + 0.003005894599482417, + 0.0010236086091026664, + 0.00156520982272923, + 0.005124593153595924, + 0.0027364895213395357, + 0.0012155214790254831, + 0.0017739039612933993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.12702758610248566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539356015622616 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539356015622616, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539356015622616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5080356597900391, + 0.06505903601646423, + 0.05398988351225853, + 0.08312855660915375, + 0.02156255580484867, + 0.13952742516994476, + 0.03636418655514717, + 0.02527010068297386, + 0.030119605362415314, + 0.03694300353527069 + ], + "credit_map": [ + 0.04212186485528946, + 0.07591941207647324, + 0.03853656351566315, + 0.09959272295236588, + 0.05738024413585663, + 0.0758952647447586, + 0.045687444508075714, + 0.07290210574865341, + 0.007589215412735939, + 0.011374062858521938, + 0.006851299200206995, + 0.0060617257840931416, + 0.008864854462444782, + 0.010033185593783855, + 0.005864681210368872, + 0.0084200045093894, + 0.0055732689797878265, + 0.006100834347307682, + 0.0058293212205171585, + 0.003656029934063554, + 0.010852414183318615, + 0.007526547182351351, + 0.004145332612097263, + 0.010306129232048988, + 0.009184914641082287, + 0.014294066466391087, + 0.005745590664446354, + 0.007173785474151373, + 0.019934291020035744, + 0.012427115812897682, + 0.005886940285563469, + 0.008481841534376144, + 0.002521843183785677, + 0.003457212820649147, + 0.0013608674053102732, + 0.0014091032790020108, + 0.006402713246643543, + 0.00339095969684422, + 0.0016152388416230679, + 0.001404616516083479, + 0.01760009117424488, + 0.021857311949133873, + 0.00969846360385418, + 0.014032580889761448, + 0.034255608916282654, + 0.021408233791589737, + 0.012281760573387146, + 0.008393359370529652, + 0.0058199153281748295, + 0.005330667831003666, + 0.0022207784932106733, + 0.0029942463152110577, + 0.008481043390929699, + 0.005642186384648085, + 0.0025642632972449064, + 0.003311083884909749, + 0.0038733945693820715, + 0.002928137080743909, + 0.0024096767883747816, + 0.0014877936337143183, + 0.004758200608193874, + 0.0042279865592718124, + 0.002105234656482935, + 0.003479676553979516, + 0.0031857213471084833, + 0.0026366012170910835, + 0.004764878191053867, + 0.0024236426688730717, + 0.0037006756756454706, + 0.004128524102270603, + 0.004856019280850887, + 0.004423539619892836, + 0.0032761560287326574, + 0.00582816731184721, + 0.0024237586185336113, + 0.002757979091256857, + 0.010740636847913265, + 0.0052963318303227425, + 0.0030841685365885496, + 0.003535802476108074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.012539356015622616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017144280718639493 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.0017144280718639493, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017144280718639493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1252489537000656, + 0.2424982488155365, + 0.08725523948669434, + 0.07610287517309189, + 0.053421128541231155, + 0.1740441471338272, + 0.07860009372234344, + 0.07180256396532059, + 0.02540825679898262, + 0.06561850756406784 + ], + "credit_map": [ + 0.011268327943980694, + 0.01616561785340309, + 0.007166666444391012, + 0.019295621663331985, + 0.016698086634278297, + 0.02809515967965126, + 0.013785168528556824, + 0.012774295173585415, + 0.030782435089349747, + 0.03938719630241394, + 0.02172037586569786, + 0.018285701051354408, + 0.033528026193380356, + 0.029889091849327087, + 0.024255456402897835, + 0.044649966061115265, + 0.008100446313619614, + 0.018072353675961494, + 0.006951631046831608, + 0.006474263966083527, + 0.019393160939216614, + 0.013856751844286919, + 0.006803346332162619, + 0.007603284902870655, + 0.006637320388108492, + 0.012923676520586014, + 0.005578514654189348, + 0.006597379222512245, + 0.018067805096507072, + 0.010682354681193829, + 0.0062525211833417416, + 0.009363302029669285, + 0.005872043780982494, + 0.008961998857557774, + 0.0031328690238296986, + 0.0032310131937265396, + 0.016345249488949776, + 0.008574782870709896, + 0.0034152136649936438, + 0.0038879557978361845, + 0.019622066989541054, + 0.030503109097480774, + 0.013954131864011288, + 0.0212089940905571, + 0.03344828262925148, + 0.023066524416208267, + 0.019818900153040886, + 0.01242213137447834, + 0.01066182553768158, + 0.013270030729472637, + 0.004094305913895369, + 0.004391896538436413, + 0.020563902333378792, + 0.013396804220974445, + 0.004682037979364395, + 0.007539288140833378, + 0.014095143415033817, + 0.008237701840698719, + 0.007270446512848139, + 0.0035496053751558065, + 0.016349617391824722, + 0.01087093260139227, + 0.007235181983560324, + 0.004193935543298721, + 0.002493047621101141, + 0.002768187317997217, + 0.0038176625967025757, + 0.0021251814905554056, + 0.0033967799972742796, + 0.003319407347589731, + 0.0037489149253815413, + 0.0037390750367194414, + 0.005718170665204525, + 0.009045804850757122, + 0.004178968258202076, + 0.004371084272861481, + 0.020668892189860344, + 0.010948811657726765, + 0.005306691862642765, + 0.005380083341151476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.0017144280718639493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737846881151199 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737846881151199, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737846881151199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.266306072473526, + 0.020195096731185913, + 0.09229832887649536, + 0.2802281677722931, + 0.12796135246753693, + 0.03120625577867031, + 0.05894960090517998, + 0.08700120449066162, + 0.017471851781010628, + 0.018382083624601364 + ], + "credit_map": [ + 0.022755803540349007, + 0.043638136237859726, + 0.020754285156726837, + 0.0522317960858345, + 0.03008337691426277, + 0.031500618904829025, + 0.02436242066323757, + 0.04097965359687805, + 0.0023575611412525177, + 0.0041055213660001755, + 0.0018957515712827444, + 0.0017639946890994906, + 0.0030025390442460775, + 0.0023913902696222067, + 0.002302871784195304, + 0.002375468146055937, + 0.008950147777795792, + 0.011446508578956127, + 0.009130164980888367, + 0.00670208316296339, + 0.01776244305074215, + 0.013332059606909752, + 0.010461078956723213, + 0.014513837173581123, + 0.030755063518881798, + 0.05536176264286041, + 0.014733570627868176, + 0.01805208995938301, + 0.07386396825313568, + 0.05107099935412407, + 0.016773942857980728, + 0.019616741687059402, + 0.01825929619371891, + 0.021587194874882698, + 0.00851889606565237, + 0.007019053678959608, + 0.035306841135025024, + 0.021412819623947144, + 0.0068788244388997555, + 0.00897842925041914, + 0.004663397092372179, + 0.0044196611270308495, + 0.0027993004769086838, + 0.0023919963277876377, + 0.008243328891694546, + 0.004863972309976816, + 0.002116102259606123, + 0.0017084968276321888, + 0.008780306205153465, + 0.012117685750126839, + 0.0026502099353820086, + 0.0028936099261045456, + 0.015735123306512833, + 0.010350016877055168, + 0.002916777040809393, + 0.0035058725625276566, + 0.015970684587955475, + 0.010141788050532341, + 0.008868901990354061, + 0.004096423275768757, + 0.020370883867144585, + 0.013524604961276054, + 0.008660421706736088, + 0.005367495119571686, + 0.0016576568596065044, + 0.0019227593438699841, + 0.0027068385388702154, + 0.0012501192977651954, + 0.0022234877105802298, + 0.0023060468956828117, + 0.002496548928320408, + 0.002908393507823348, + 0.0019931704737246037, + 0.0031810167711228132, + 0.00131609954405576, + 0.0015534976264461875, + 0.0042530507780611515, + 0.002681600395590067, + 0.0011573905358090997, + 0.002246258081868291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.006737846881151199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This ultra-high-speed medium shot, captured from a static low-angle perspective, gazes upon the intricate, spiny underside of a massive Victoria Amazon water lily, capturing the scene in extreme slow motion. The entire frame is submerged in murky green-brown water, while filtered sunlight penetrates from above, creating an ethereal crepuscular glow that illuminates floating silt particles. The lily's surface is an intricate network of thick, reddish-purple vascular ribs capped with sharp spines, creating dramatic patterns of light and deep shadow. As a delicate stream of shimmering oxygen bubbles begins to condense from the plant's surface, a magnificent process of gas exchange unfolds. Each pearlescent bubble slowly forms, clinging precariously to the spiny texture in a tense moment, its perfectly spherical shape slightly distorted by surface tension before bursting. Then, one by one, they detach and begin a tranquil, silent ascent, rising through the hazy water like a string of tiny, glowing gems. As they float steadily and deliberately upward and out of the frame, they catch the overhead light, creating an atmosphere of profound, hidden wonder and nature's quiet grandeur.", + "seed": 18435, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_0/index_6954_seed_18435_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14224722364451736, + "vlm_logprob": -0.14224722364451736, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.14224722364451736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03023352660238743 + ], + "opposite_logprob_tokens": [ + -4.155233383178711 + ], + "logprob_reward": -0.03023352660238743, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03023352660238743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20265047252178192, + 0.13701458275318146, + 0.09531022608280182, + 0.07222924381494522, + 0.14070704579353333, + 0.11507944017648697, + 0.0466509684920311, + 0.028232978656888008, + 0.10177754610776901, + 0.06034752354025841 + ], + "credit_map": [ + 0.03531736508011818, + 0.025825176388025284, + 0.024777818471193314, + 0.01978129707276821, + 0.031780682504177094, + 0.030817639082670212, + 0.016635354608297348, + 0.017715130001306534, + 0.02498326078057289, + 0.016345255076885223, + 0.014322794042527676, + 0.016439419239759445, + 0.016141235828399658, + 0.022589925676584244, + 0.01121070608496666, + 0.014981983229517937, + 0.012447190470993519, + 0.01071664597839117, + 0.007977868430316448, + 0.011920087970793247, + 0.012231470085680485, + 0.021185262128710747, + 0.010270923376083374, + 0.008560781367123127, + 0.011229991912841797, + 0.008436114527285099, + 0.007257731631398201, + 0.006079006474465132, + 0.011129855178296566, + 0.010086895897984505, + 0.009162640199065208, + 0.008847011253237724, + 0.02617424540221691, + 0.04226960614323616, + 0.00929533876478672, + 0.011042222380638123, + 0.016181249171495438, + 0.015001054853200912, + 0.010513323359191418, + 0.010230018757283688, + 0.02534644678235054, + 0.023344745859503746, + 0.010338443331420422, + 0.012484929524362087, + 0.011648260988295078, + 0.011580610647797585, + 0.008866486139595509, + 0.011469517834484577, + 0.008699797093868256, + 0.009405924938619137, + 0.0039232345297932625, + 0.0061498358845710754, + 0.005689890589565039, + 0.004492837004363537, + 0.0037847936619073153, + 0.004504654556512833, + 0.006390260998159647, + 0.004408351145684719, + 0.002282927045598626, + 0.0025970349088311195, + 0.00484686903655529, + 0.0031690436881035566, + 0.002339959144592285, + 0.0021985319908708334, + 0.020276060327887535, + 0.0334538072347641, + 0.009961295872926712, + 0.007804150693118572, + 0.00954196322709322, + 0.006055093370378017, + 0.0070586116053164005, + 0.007626563310623169, + 0.008687281049787998, + 0.014686480164527893, + 0.00541511457413435, + 0.004653165582567453, + 0.009848770685493946, + 0.006729099899530411, + 0.005246441345661879, + 0.005081167444586754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.03023352660238743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632794201374054 + ], + "opposite_logprob_tokens": [ + -0.757794201374054 + ], + "logprob_reward": -0.632794201374054, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632794201374054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2692941725254059, + 0.1648070216178894, + 0.17251300811767578, + 0.034981854259967804, + 0.10727723687887192, + 0.04201371967792511, + 0.09980132430791855, + 0.01976495049893856, + 0.06414351612329483, + 0.025403207167983055 + ], + "credit_map": [ + 0.0517311617732048, + 0.052440300583839417, + 0.021494701504707336, + 0.025841787457466125, + 0.0474492684006691, + 0.02423872984945774, + 0.02132134884595871, + 0.0247768796980381, + 0.029577497392892838, + 0.022219659760594368, + 0.015819313004612923, + 0.01657644845545292, + 0.01899358071386814, + 0.024254005402326584, + 0.01854885369539261, + 0.018817655742168427, + 0.025994354858994484, + 0.021855246275663376, + 0.014501058496534824, + 0.014752024784684181, + 0.03693615645170212, + 0.020700259134173393, + 0.021657951176166534, + 0.016115950420498848, + 0.006253738421946764, + 0.004293710924685001, + 0.0031198945362120867, + 0.0035465219989418983, + 0.00504610501229763, + 0.006108307279646397, + 0.003315781941637397, + 0.003297796007245779, + 0.023824818432331085, + 0.026640141382813454, + 0.006739587057381868, + 0.011554574593901634, + 0.010929933749139309, + 0.0076476698741316795, + 0.010200155898928642, + 0.009740359149873257, + 0.010017262771725655, + 0.005387341603636742, + 0.003465053392574191, + 0.005562400445342064, + 0.005685858894139528, + 0.004793842788785696, + 0.0030987593345344067, + 0.004003198817372322, + 0.021599551662802696, + 0.023203445598483086, + 0.007749850861728191, + 0.011322909966111183, + 0.011902401223778725, + 0.0090655367821455, + 0.007279321551322937, + 0.00767831364646554, + 0.002984205959364772, + 0.0031649756710976362, + 0.0018841734854504466, + 0.0021656611934304237, + 0.003812729613855481, + 0.002241109497845173, + 0.001860130694694817, + 0.0016519655473530293, + 0.012664848007261753, + 0.021168405190110207, + 0.0055984496138989925, + 0.00484372116625309, + 0.00584475276991725, + 0.005013773683458567, + 0.004146886989474297, + 0.004862678237259388, + 0.003934063017368317, + 0.005968797951936722, + 0.002535184845328331, + 0.0021046274341642857, + 0.0035523339174687862, + 0.002920178696513176, + 0.0023756909649819136, + 0.002012330573052168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.632794201374054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410179328173399 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.00410179328173399, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410179328173399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20409564673900604, + 0.1832011342048645, + 0.14053098857402802, + 0.06955870240926743, + 0.07642575353384018, + 0.05298842489719391, + 0.07062675803899765, + 0.06292557716369629, + 0.07984302937984467, + 0.059803999960422516 + ], + "credit_map": [ + 0.030257364735007286, + 0.04440825432538986, + 0.02165995165705681, + 0.027371002361178398, + 0.024394871667027473, + 0.02019573748111725, + 0.014735918492078781, + 0.021072519943118095, + 0.027810387313365936, + 0.02410075068473816, + 0.021338418126106262, + 0.02065163664519787, + 0.01919274404644966, + 0.028370434418320656, + 0.02160119079053402, + 0.020135559141635895, + 0.01758386567234993, + 0.014523432590067387, + 0.011431715451180935, + 0.016447244212031364, + 0.02120816335082054, + 0.02850707806646824, + 0.01711387187242508, + 0.013715610839426517, + 0.013113894499838352, + 0.009222671389579773, + 0.005639293231070042, + 0.004341913387179375, + 0.0110769709572196, + 0.009995970875024796, + 0.009748740121722221, + 0.006419248413294554, + 0.011058893986046314, + 0.012812556698918343, + 0.007346449885517359, + 0.011225823312997818, + 0.006885452661663294, + 0.010027306154370308, + 0.010028762742877007, + 0.007040506228804588, + 0.010192492976784706, + 0.010249907150864601, + 0.004028664901852608, + 0.004550690297037363, + 0.006497715134173632, + 0.006901001092046499, + 0.004921393934637308, + 0.0056465547531843185, + 0.015491747297346592, + 0.019009528681635857, + 0.005468085873872042, + 0.006601947825402021, + 0.007419591303914785, + 0.0061431946232914925, + 0.004719763994216919, + 0.005772891454398632, + 0.012415825389325619, + 0.0143090495839715, + 0.0039909896440804005, + 0.006408486049622297, + 0.009949172846972942, + 0.005279875360429287, + 0.004887778777629137, + 0.005684399511665106, + 0.015090708620846272, + 0.019206272438168526, + 0.008308290503919125, + 0.006923663895577192, + 0.00948286522179842, + 0.006851636804640293, + 0.006035732105374336, + 0.007943852804601192, + 0.007551061920821667, + 0.009244224056601524, + 0.005868225824087858, + 0.006268300116062164, + 0.013860129751265049, + 0.0060751475393772125, + 0.004648059606552124, + 0.006288846954703331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00410179328173399 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010446811793372035 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010446811793372035, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010446811793372035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17819595336914062, + 0.07730013132095337, + 0.1763070970773697, + 0.08449143171310425, + 0.16539129614830017, + 0.04676782712340355, + 0.047200653702020645, + 0.044905148446559906, + 0.0433456227183342, + 0.1360948383808136 + ], + "credit_map": [ + 0.023991020396351814, + 0.02997271716594696, + 0.02054499089717865, + 0.0186978317797184, + 0.034119173884391785, + 0.01969943940639496, + 0.014900111593306065, + 0.016270672902464867, + 0.013830122537910938, + 0.00862610712647438, + 0.008972348645329475, + 0.007875395938754082, + 0.010179156437516212, + 0.011205393821001053, + 0.008744096383452415, + 0.007867509499192238, + 0.018256427720189095, + 0.01911158673465252, + 0.014052868820726871, + 0.021895239129662514, + 0.02651459537446499, + 0.04190654680132866, + 0.018316298723220825, + 0.016253521665930748, + 0.01383424922823906, + 0.011397548951208591, + 0.007820391096174717, + 0.008340763859450817, + 0.011449085548520088, + 0.014508654363453388, + 0.009743094444274902, + 0.0073976400308310986, + 0.023970672860741615, + 0.025781996548175812, + 0.014287256635725498, + 0.02566780336201191, + 0.019915005192160606, + 0.020211845636367798, + 0.015320082195103168, + 0.020236631855368614, + 0.009917412884533405, + 0.006802947726100683, + 0.004822868853807449, + 0.0036377040669322014, + 0.005931445397436619, + 0.006994062103331089, + 0.00421149842441082, + 0.004449887201189995, + 0.008125277236104012, + 0.00767040578648448, + 0.003517944598570466, + 0.006010602694004774, + 0.007487313821911812, + 0.00500142015516758, + 0.00440175412222743, + 0.004985938314348459, + 0.009801341220736504, + 0.012464156374335289, + 0.0025885235518217087, + 0.0035225776955485344, + 0.006229540333151817, + 0.005634659435600042, + 0.002090430585667491, + 0.0025739192496985197, + 0.00779141578823328, + 0.007436547894030809, + 0.004253160674124956, + 0.004127031192183495, + 0.005953251384198666, + 0.004238239489495754, + 0.005619082599878311, + 0.0039268918335437775, + 0.0248632300645113, + 0.019624857231974602, + 0.010435501113533974, + 0.009361417964100838, + 0.036660004407167435, + 0.014927194453775883, + 0.011010570451617241, + 0.009212055243551731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0010446811793372035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043061915785074234 + ], + "opposite_logprob_tokens": [ + -3.1680619716644287 + ], + "logprob_reward": -0.043061915785074234, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043061915785074234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26493486762046814, + 0.05052667483687401, + 0.08338180929422379, + 0.06201476976275444, + 0.1917012482881546, + 0.05905124545097351, + 0.051015034317970276, + 0.10606694966554642, + 0.07072948664426804, + 0.06057796627283096 + ], + "credit_map": [ + 0.06799294054508209, + 0.04004460573196411, + 0.020484112203121185, + 0.019502250477671623, + 0.05037911981344223, + 0.024734877049922943, + 0.021019132807850838, + 0.020777808502316475, + 0.008044218644499779, + 0.006389164365828037, + 0.005399447865784168, + 0.0047255391255021095, + 0.007444009650498629, + 0.006602637004107237, + 0.0057686553336679935, + 0.006153002846986055, + 0.010581844486296177, + 0.00876445323228836, + 0.007298278156667948, + 0.012272306717932224, + 0.0126983392983675, + 0.014147753827273846, + 0.008193508721888065, + 0.009425323456525803, + 0.011218445375561714, + 0.007538336329162121, + 0.005291705019772053, + 0.005761957727372646, + 0.008840113878250122, + 0.009729084558784962, + 0.008254953660070896, + 0.005380168091505766, + 0.027678320184350014, + 0.04153059422969818, + 0.015348168089985847, + 0.032760970294475555, + 0.018595173954963684, + 0.019805407151579857, + 0.014114164747297764, + 0.021868430078029633, + 0.011258242651820183, + 0.008522072806954384, + 0.004316824022680521, + 0.006703919265419245, + 0.007916064001619816, + 0.007824016734957695, + 0.00653027044609189, + 0.005979830399155617, + 0.009703353978693485, + 0.009610049426555634, + 0.004388617817312479, + 0.0049642762169241905, + 0.008656909689307213, + 0.00498045003041625, + 0.0040842327289283276, + 0.004627141635864973, + 0.02257814258337021, + 0.021804384887218475, + 0.007533700671046972, + 0.011870509013533592, + 0.017919017001986504, + 0.010204252786934376, + 0.007110613863915205, + 0.007046322338283062, + 0.013986313715577126, + 0.024586772546172142, + 0.006498836446553469, + 0.005321022123098373, + 0.007572268135845661, + 0.004776400048285723, + 0.0030205745715647936, + 0.004967294633388519, + 0.009955760091543198, + 0.008139369077980518, + 0.005193115212023258, + 0.006279728841036558, + 0.014655321836471558, + 0.006624265108257532, + 0.005115950480103493, + 0.004614448640495539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.043061915785074234 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This ultra-high-speed medium shot, captured from a static low-angle perspective, gazes upon the intricate, spiny underside of a massive Victoria Amazon water lily, capturing the scene in extreme slow motion. The entire frame is submerged in murky green-brown water, while filtered sunlight penetrates from above, creating an ethereal crepuscular glow that illuminates floating silt particles. The lily's surface is an intricate network of thick, reddish-purple vascular ribs capped with sharp spines, creating dramatic patterns of light and deep shadow. As a delicate stream of shimmering oxygen bubbles begins to condense from the plant's surface, a magnificent process of gas exchange unfolds. Each pearlescent bubble slowly forms, clinging precariously to the spiny texture in a tense moment, its perfectly spherical shape slightly distorted by surface tension before bursting. Then, one by one, they detach and begin a tranquil, silent ascent, rising through the hazy water like a string of tiny, glowing gems. As they float steadily and deliberately upward and out of the frame, they catch the overhead light, creating an atmosphere of profound, hidden wonder and nature's quiet grandeur.", + "seed": 18436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_0/index_6954_seed_18436_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07837150087580085, + "vlm_logprob": -0.07837150087580085, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.07837150087580085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0925007238984108 + ], + "opposite_logprob_tokens": [ + -2.717500686645508 + ], + "logprob_reward": -0.0925007238984108, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0925007238984108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2024717777967453, + 0.19683340191841125, + 0.15847453474998474, + 0.041256777942180634, + 0.07277759909629822, + 0.061930570751428604, + 0.07111921906471252, + 0.1234714537858963, + 0.04155155271291733, + 0.030113134533166885 + ], + "credit_map": [ + 0.024344850331544876, + 0.025195853784680367, + 0.015462123788893223, + 0.016801461577415466, + 0.04288670793175697, + 0.03641239181160927, + 0.014515073038637638, + 0.026853319257497787, + 0.013277814723551273, + 0.017499186098575592, + 0.003767133690416813, + 0.004746125545352697, + 0.0769944116473198, + 0.06810055673122406, + 0.005817726254463196, + 0.006630434654653072, + 0.018150977790355682, + 0.021142669022083282, + 0.008555956184864044, + 0.015584463253617287, + 0.03467839956283569, + 0.03588924556970596, + 0.011208603158593178, + 0.013264219276607037, + 0.004729032516479492, + 0.006307699717581272, + 0.0030764134135097265, + 0.005206073634326458, + 0.00722733186557889, + 0.004838736727833748, + 0.005491807125508785, + 0.004379684571176767, + 0.008654537610709667, + 0.009533952921628952, + 0.006037695799022913, + 0.009159574285149574, + 0.01176619902253151, + 0.00882006622850895, + 0.00803744699805975, + 0.01076812669634819, + 0.009525591507554054, + 0.008392905816435814, + 0.006582908798009157, + 0.00601681973785162, + 0.008220267482101917, + 0.010808201506733894, + 0.004392223432660103, + 0.007991652004420757, + 0.011377173475921154, + 0.009824290871620178, + 0.004599456675350666, + 0.007493687327951193, + 0.010500175878405571, + 0.012851904146373272, + 0.006371836178004742, + 0.008100698702037334, + 0.013959183357656002, + 0.020334536209702492, + 0.008831319399178028, + 0.008647345937788486, + 0.030756719410419464, + 0.014681881293654442, + 0.010406897403299809, + 0.015853574499487877, + 0.006735044531524181, + 0.0058466121554374695, + 0.0033547987695783377, + 0.004755490459501743, + 0.007715738378465176, + 0.004958630073815584, + 0.003195799421519041, + 0.004989436827600002, + 0.0033921287395060062, + 0.004772949498146772, + 0.00289939041249454, + 0.002952782204374671, + 0.0038196018431335688, + 0.00621928833425045, + 0.0019570759031921625, + 0.0040999166667461395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0925007238984108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281290203332901 + ], + "opposite_logprob_tokens": [ + -1.4062901735305786 + ], + "logprob_reward": -0.281290203332901, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281290203332901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5932063460350037, + 0.048623837530612946, + 0.06464319676160812, + 0.045850563794374466, + 0.023655569180846214, + 0.027410423383116722, + 0.03765200451016426, + 0.02341300994157791, + 0.11426600813865662, + 0.02127906121313572 + ], + "credit_map": [ + 0.03780084848403931, + 0.08033490926027298, + 0.012253668159246445, + 0.018416473641991615, + 0.2141188383102417, + 0.19276531040668488, + 0.01680983603000641, + 0.02070646546781063, + 0.004085328429937363, + 0.004514914005994797, + 0.0017493482446298003, + 0.001846783678047359, + 0.01631929725408554, + 0.015187466517090797, + 0.0023224197793751955, + 0.0025982798542827368, + 0.00810321606695652, + 0.010021342895925045, + 0.003601780626922846, + 0.007136224303394556, + 0.011780335567891598, + 0.012848042882978916, + 0.005166703835129738, + 0.005985545460134745, + 0.00624775979667902, + 0.006839493755251169, + 0.004138259682804346, + 0.004819875117391348, + 0.007912291213870049, + 0.005617454182356596, + 0.004378648940473795, + 0.005896780174225569, + 0.003649506950750947, + 0.0029900174122303724, + 0.0019483461510390043, + 0.002512999577447772, + 0.0035915288608521223, + 0.00255623459815979, + 0.0021523367613554, + 0.004254596773535013, + 0.004489773418754339, + 0.0033588316291570663, + 0.0033745516557246447, + 0.0025292474310845137, + 0.0038325886707752943, + 0.004304736386984587, + 0.0022683795541524887, + 0.0032523125410079956, + 0.0045349751599133015, + 0.00529450410977006, + 0.002929629758000374, + 0.002642502076923847, + 0.0045520951971411705, + 0.007501649204641581, + 0.0031933626160025597, + 0.007003281731158495, + 0.003291037864983082, + 0.003767963033169508, + 0.0013905797386541963, + 0.0021711939480155706, + 0.005411084741353989, + 0.002746036509051919, + 0.001966900657862425, + 0.0026682098396122456, + 0.012774649076163769, + 0.017798857763409615, + 0.008190032094717026, + 0.010948658920824528, + 0.024478277191519737, + 0.012630779296159744, + 0.010157388634979725, + 0.0172873642295599, + 0.002506258664652705, + 0.002898368751630187, + 0.0018534379778429866, + 0.0021621359046548605, + 0.003464748617261648, + 0.0035763687919825315, + 0.0018231357680633664, + 0.0029946044087409973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.281290203332901 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951540544629097 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951540544629097, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951540544629097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13321039080619812, + 0.37265366315841675, + 0.04771772772073746, + 0.08042462170124054, + 0.1041099801659584, + 0.03984576836228371, + 0.06278429180383682, + 0.03183092549443245, + 0.0993761271238327, + 0.028046591207385063 + ], + "credit_map": [ + 0.012975314632058144, + 0.01860218308866024, + 0.01083432324230671, + 0.00989671889692545, + 0.03501531854271889, + 0.024389738216996193, + 0.009240055456757545, + 0.01225671824067831, + 0.025839192792773247, + 0.03307401016354561, + 0.004697074182331562, + 0.005794096738100052, + 0.15183579921722412, + 0.13180187344551086, + 0.00824216939508915, + 0.011369415558874607, + 0.005654879845678806, + 0.007922629825770855, + 0.0030474578961730003, + 0.005137386731803417, + 0.01081147976219654, + 0.006672751624137163, + 0.004493024665862322, + 0.003978113178163767, + 0.009375754743814468, + 0.009286205284297466, + 0.007122111041098833, + 0.01346708182245493, + 0.011263523250818253, + 0.008343449793756008, + 0.01330464705824852, + 0.00826184544712305, + 0.01385272666811943, + 0.018716150894761086, + 0.008729745633900166, + 0.010480169206857681, + 0.020643679425120354, + 0.013651756569743156, + 0.00783331785351038, + 0.010202421806752682, + 0.007010455708950758, + 0.004804688505828381, + 0.0044551449827849865, + 0.003661384806036949, + 0.005443301983177662, + 0.00676687341183424, + 0.003044964512810111, + 0.004658951889723539, + 0.008352696895599365, + 0.008604966104030609, + 0.004743044730275869, + 0.0053376974537968636, + 0.008497876115143299, + 0.008254307322204113, + 0.007778501603752375, + 0.011215196922421455, + 0.0036291906144469976, + 0.004771660082042217, + 0.002582667162641883, + 0.002514622174203396, + 0.006454731337726116, + 0.004715785849839449, + 0.0031735613010823727, + 0.003988704644143581, + 0.01187269575893879, + 0.015615230426192284, + 0.006699824705719948, + 0.009270125068724155, + 0.02182649075984955, + 0.011022006161510944, + 0.008649410679936409, + 0.014420337043702602, + 0.002970051020383835, + 0.006513563450425863, + 0.0019160601077601314, + 0.0026546986773610115, + 0.002697192830964923, + 0.007290366105735302, + 0.0017809567507356405, + 0.002223699353635311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.005951540544629097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010422994382679462 + ], + "opposite_logprob_tokens": [ + -6.876042366027832 + ], + "logprob_reward": -0.0010422994382679462, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010422994382679462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6505916118621826, + 0.08130531758069992, + 0.018064597621560097, + 0.024640483781695366, + 0.09210082143545151, + 0.031581200659275055, + 0.020002899691462517, + 0.018574248999357224, + 0.037254698574543, + 0.0258841123431921 + ], + "credit_map": [ + 0.04139844328165054, + 0.08604778349399567, + 0.016730252653360367, + 0.023886334151029587, + 0.2281946837902069, + 0.20575980842113495, + 0.01950770802795887, + 0.029066666960716248, + 0.005448608193546534, + 0.007282548118382692, + 0.0018903753953054547, + 0.002802118891850114, + 0.029760193079710007, + 0.026291081681847572, + 0.004649859853088856, + 0.0031805294565856457, + 0.001582668861374259, + 0.0028101769275963306, + 0.0010372297838330269, + 0.002161686308681965, + 0.004443102516233921, + 0.0027403521817177534, + 0.0020210519433021545, + 0.0012683282839134336, + 0.0031083819922059774, + 0.003055393695831299, + 0.0023296047002077103, + 0.004017426632344723, + 0.002636047312989831, + 0.002598891034722328, + 0.003942420240491629, + 0.0029523184057325125, + 0.0127104502171278, + 0.016899442300200462, + 0.009690587408840656, + 0.008052460849285126, + 0.014189502224326134, + 0.010138102807104588, + 0.007950295694172382, + 0.012469982728362083, + 0.0052797067910432816, + 0.0056227524764835835, + 0.001901583862490952, + 0.002808743854984641, + 0.004187525250017643, + 0.006764998659491539, + 0.0019444696372374892, + 0.0030714161694049835, + 0.0030036396346986294, + 0.0022231987677514553, + 0.0017481350805610418, + 0.00211959145963192, + 0.0024029258638620377, + 0.0024560019373893738, + 0.002675611525774002, + 0.0033737949561327696, + 0.0024774009361863136, + 0.0030268938280642033, + 0.001097199390642345, + 0.0015831220662221313, + 0.00446360046043992, + 0.0023496500216424465, + 0.0016185928834602237, + 0.0019577904604375362, + 0.003907542675733566, + 0.0057870177552104, + 0.002548959106206894, + 0.003696952946484089, + 0.007678938563913107, + 0.004575937986373901, + 0.0034668887965381145, + 0.005592464469373226, + 0.003278901567682624, + 0.0037942680064588785, + 0.0017796194879338145, + 0.002582285087555647, + 0.004535276908427477, + 0.0042894198559224606, + 0.002194356871768832, + 0.003429984673857689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0010422994382679462 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072737164795399 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072737164795399, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072737164795399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33160385489463806, + 0.22843702137470245, + 0.042246777564287186, + 0.04434998333454132, + 0.0370134674012661, + 0.06111839786171913, + 0.05648481100797653, + 0.0930396169424057, + 0.07030639797449112, + 0.03539961576461792 + ], + "credit_map": [ + 0.0335993655025959, + 0.048405781388282776, + 0.021927818655967712, + 0.024151131510734558, + 0.08329992741346359, + 0.05948704853653908, + 0.02290484495460987, + 0.03782794252038002, + 0.01632538251578808, + 0.020684629678726196, + 0.003785419277846813, + 0.0040572090074419975, + 0.09107346087694168, + 0.07886632531881332, + 0.006158196832984686, + 0.0074864043854177, + 0.004584657959640026, + 0.006510626059025526, + 0.003544370410963893, + 0.0039195469580590725, + 0.00966443121433258, + 0.005856537725776434, + 0.004002195782959461, + 0.004164410289376974, + 0.004546147771179676, + 0.007257434539496899, + 0.0032145488075912, + 0.007181308697909117, + 0.006367615424096584, + 0.004261249676346779, + 0.005490520503371954, + 0.0060311583802104, + 0.005418301559984684, + 0.00688838679343462, + 0.0029562669806182384, + 0.004230109974741936, + 0.006129192188382149, + 0.004541821777820587, + 0.0030206693336367607, + 0.003828719025477767, + 0.007500272244215012, + 0.013760476373136044, + 0.003924776799976826, + 0.005079311318695545, + 0.00842871330678463, + 0.012727580033242702, + 0.004095686133950949, + 0.00560158072039485, + 0.007142507936805487, + 0.007774345111101866, + 0.004096785560250282, + 0.004484750796109438, + 0.009617255069315434, + 0.011911955662071705, + 0.004609944298863411, + 0.006847265642136335, + 0.006661990191787481, + 0.018550949171185493, + 0.004414353519678116, + 0.003895349567756057, + 0.036965690553188324, + 0.012058679014444351, + 0.004446757957339287, + 0.006045849062502384, + 0.008678106591105461, + 0.010856992565095425, + 0.0049486844800412655, + 0.0064203473739326, + 0.014258576557040215, + 0.00784032791852951, + 0.006360874976962805, + 0.010942489840090275, + 0.004946464207023382, + 0.004707067273557186, + 0.0024749301373958588, + 0.0035997626837342978, + 0.006856505759060383, + 0.0056465258821845055, + 0.0035242962185293436, + 0.0036440619733184576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011072737164795399 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This ultra-high-speed medium shot, captured from a static low-angle perspective, gazes upon the intricate, spiny underside of a massive Victoria Amazon water lily, capturing the scene in extreme slow motion. The entire frame is submerged in murky green-brown water, while filtered sunlight penetrates from above, creating an ethereal crepuscular glow that illuminates floating silt particles. The lily's surface is an intricate network of thick, reddish-purple vascular ribs capped with sharp spines, creating dramatic patterns of light and deep shadow. As a delicate stream of shimmering oxygen bubbles begins to condense from the plant's surface, a magnificent process of gas exchange unfolds. Each pearlescent bubble slowly forms, clinging precariously to the spiny texture in a tense moment, its perfectly spherical shape slightly distorted by surface tension before bursting. Then, one by one, they detach and begin a tranquil, silent ascent, rising through the hazy water like a string of tiny, glowing gems. As they float steadily and deliberately upward and out of the frame, they catch the overhead light, creating an atmosphere of profound, hidden wonder and nature's quiet grandeur.", + "seed": 18437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_0/index_6954_seed_18437_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.053410820476710794, + "vlm_logprob": -0.053410820476710794, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.053410820476710794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046134814620018005 + ], + "opposite_logprob_tokens": [ + -3.4211347103118896 + ], + "logprob_reward": -0.046134814620018005, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.046134814620018005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09905602037906647, + 0.16177965700626373, + 0.138625830411911, + 0.08657783269882202, + 0.10358516871929169, + 0.09907054156064987, + 0.07840397208929062, + 0.07691037654876709, + 0.10192310065031052, + 0.05406753718852997 + ], + "credit_map": [ + 0.030356863513588905, + 0.0115975895896554, + 0.019858283922076225, + 0.00743088498711586, + 0.006857478525489569, + 0.010338867083191872, + 0.006382242310792208, + 0.006233805324882269, + 0.011039255186915398, + 0.06401420384645462, + 0.006287282332777977, + 0.006966033484786749, + 0.024724463000893593, + 0.03290349617600441, + 0.007324986159801483, + 0.008519947528839111, + 0.010498082265257835, + 0.021366238594055176, + 0.009665993973612785, + 0.013763057999312878, + 0.03539085388183594, + 0.019810611382126808, + 0.014590797945857048, + 0.013540190644562244, + 0.007913338951766491, + 0.011033626273274422, + 0.007421627640724182, + 0.012355729937553406, + 0.010983959771692753, + 0.017802299931645393, + 0.009867014363408089, + 0.009200229309499264, + 0.010503497906029224, + 0.008849043399095535, + 0.015101239085197449, + 0.01444519218057394, + 0.012524470686912537, + 0.012118343263864517, + 0.01705850660800934, + 0.012984862551093102, + 0.009172770194709301, + 0.014633029699325562, + 0.007893214002251625, + 0.012392085045576096, + 0.020913291722536087, + 0.012357242405414581, + 0.013707771897315979, + 0.008001135662198067, + 0.013800788670778275, + 0.012079230509698391, + 0.004267853684723377, + 0.00521173095330596, + 0.016337944194674492, + 0.016010256484150887, + 0.005357899237424135, + 0.005338267423212528, + 0.010989004746079445, + 0.01106877252459526, + 0.004973164293915033, + 0.00678729172796011, + 0.017378875985741615, + 0.011240542866289616, + 0.007057500537484884, + 0.007415231317281723, + 0.010684038512408733, + 0.01400547381490469, + 0.00501349987462163, + 0.011408471502363682, + 0.025731505826115608, + 0.016530541703104973, + 0.012269864790141582, + 0.006279702298343182, + 0.0053251455537974834, + 0.00681212218478322, + 0.003738906467333436, + 0.006721746642142534, + 0.01205977238714695, + 0.008492317982017994, + 0.005821987986564636, + 0.005095536820590496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.046134814620018005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153886079788208 + ], + "opposite_logprob_tokens": [ + -1.7015388011932373 + ], + "logprob_reward": -0.20153886079788208, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153886079788208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1463657170534134, + 0.23583269119262695, + 0.1203782707452774, + 0.07337524741888046, + 0.11487413197755814, + 0.05820073187351227, + 0.10520975291728973, + 0.031034542247653008, + 0.05290509760379791, + 0.061823803931474686 + ], + "credit_map": [ + 0.02958417870104313, + 0.014860554598271847, + 0.03240674361586571, + 0.012181795202195644, + 0.009574266150593758, + 0.022366471588611603, + 0.013901726342737675, + 0.011489976197481155, + 0.01606718637049198, + 0.09540127962827682, + 0.008236815221607685, + 0.007955418899655342, + 0.03495177999138832, + 0.05086146295070648, + 0.011241807602345943, + 0.01111692376434803, + 0.009700321592390537, + 0.028142286464571953, + 0.006683436222374439, + 0.007494618650525808, + 0.036036327481269836, + 0.017349278554320335, + 0.008190159685909748, + 0.0067818365059792995, + 0.0071752299554646015, + 0.010024284943938255, + 0.0039520105347037315, + 0.007564599625766277, + 0.01552044227719307, + 0.017593294382095337, + 0.005763014778494835, + 0.005782371386885643, + 0.016992853954434395, + 0.012633273378014565, + 0.00998782180249691, + 0.014205843210220337, + 0.019487187266349792, + 0.01702186092734337, + 0.015127064660191536, + 0.009418225847184658, + 0.006450806278735399, + 0.007786815986037254, + 0.0038338322192430496, + 0.008219851180911064, + 0.013052720576524734, + 0.007894097827374935, + 0.005856608506292105, + 0.005106000229716301, + 0.025167809799313545, + 0.013306447304785252, + 0.005422133486717939, + 0.006152020301669836, + 0.02390393428504467, + 0.01802627183496952, + 0.005882397294044495, + 0.007348738145083189, + 0.0034730664920061827, + 0.0038396616000682116, + 0.0021625999361276627, + 0.004356179386377335, + 0.004295353777706623, + 0.003917166963219643, + 0.003621704876422882, + 0.005368808284401894, + 0.0065117222256958485, + 0.006449966691434383, + 0.0032827381510287523, + 0.005376701708883047, + 0.013242102228105068, + 0.007954824715852737, + 0.006724565755575895, + 0.003362473100423813, + 0.006575264967978001, + 0.008138683624565601, + 0.0037428049836307764, + 0.0067941355518996716, + 0.01698019728064537, + 0.009417071007192135, + 0.005195725709199905, + 0.004979920107871294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.20153886079788208 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640880040824413 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640880040824413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640880040824413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09608852118253708, + 0.22940433025360107, + 0.09180687367916107, + 0.3173101842403412, + 0.07642946392297745, + 0.049647748470306396, + 0.014950519427657127, + 0.027526021003723145, + 0.026945311576128006, + 0.06989095360040665 + ], + "credit_map": [ + 0.01586976647377014, + 0.014905406162142754, + 0.012702587060630322, + 0.010303057730197906, + 0.008942871354520321, + 0.01597859151661396, + 0.007944459095597267, + 0.009441791102290154, + 0.015993814915418625, + 0.08787254244089127, + 0.012495367787778378, + 0.009371945634484291, + 0.035021983087062836, + 0.041462041437625885, + 0.01496982667595148, + 0.012216824106872082, + 0.00791135709732771, + 0.015037636272609234, + 0.006960838567465544, + 0.010249152779579163, + 0.02453482151031494, + 0.010641984641551971, + 0.009777677245438099, + 0.00669341254979372, + 0.02150074951350689, + 0.04162575677037239, + 0.02135077863931656, + 0.07162696868181229, + 0.04573458433151245, + 0.04318744316697121, + 0.05691620334982872, + 0.015367715619504452, + 0.0068288990296423435, + 0.010663378983736038, + 0.007025643717497587, + 0.009608456864953041, + 0.014081929810345173, + 0.008805083110928535, + 0.011947493068873882, + 0.0074685863219201565, + 0.0047494228929281235, + 0.005894786212593317, + 0.004243320785462856, + 0.0065051172859966755, + 0.009791525080800056, + 0.005773433018475771, + 0.006016716826707125, + 0.006673433352261782, + 0.0016491453861817718, + 0.0018105097115039825, + 0.0014678360894322395, + 0.0014828736893832684, + 0.0030777049250900745, + 0.001921153045259416, + 0.0018443618901073933, + 0.0016969365533441305, + 0.003655991517007351, + 0.004288524389266968, + 0.0019287460017949343, + 0.0030080550350248814, + 0.005361124407500029, + 0.0037113584112375975, + 0.0034287634771317244, + 0.0021434612572193146, + 0.0027602016925811768, + 0.0034331001807004213, + 0.002728541847318411, + 0.003522943239659071, + 0.003216359531506896, + 0.003651498816907406, + 0.0044434210285544395, + 0.00318924686871469, + 0.006899996194988489, + 0.008521286770701408, + 0.00756773492321372, + 0.00947930384427309, + 0.01148545928299427, + 0.009424581192433834, + 0.007940909825265408, + 0.008571689948439598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.004640880040824413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198542468249798 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198542468249798, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198542468249798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08893755078315735, + 0.22397476434707642, + 0.13293901085853577, + 0.24122120440006256, + 0.11947315186262131, + 0.04203171655535698, + 0.0543147511780262, + 0.043972261250019073, + 0.01920822076499462, + 0.0339273065328598 + ], + "credit_map": [ + 0.029088173061609268, + 0.011092675849795341, + 0.01333533227443695, + 0.006925567984580994, + 0.006755037698894739, + 0.009966088458895683, + 0.006443863734602928, + 0.005330825224518776, + 0.013780886307358742, + 0.09050996601581573, + 0.00813464354723692, + 0.007634143810719252, + 0.03181916102766991, + 0.05126272886991501, + 0.010694997385144234, + 0.010138256475329399, + 0.009026700630784035, + 0.022009801119565964, + 0.01358127687126398, + 0.013881866820156574, + 0.03266752511262894, + 0.01871773973107338, + 0.0120746660977602, + 0.010979448445141315, + 0.015592886134982109, + 0.03278360515832901, + 0.017825989052653313, + 0.060440659523010254, + 0.02994714304804802, + 0.025045858696103096, + 0.047355614602565765, + 0.012229477055370808, + 0.013238910585641861, + 0.015262724831700325, + 0.01058154460042715, + 0.013870871625840664, + 0.02314177341759205, + 0.01850295625627041, + 0.015487433411180973, + 0.009386943653225899, + 0.006688020657747984, + 0.005233654286712408, + 0.0029430140275508165, + 0.004483005963265896, + 0.008239503018558025, + 0.005547972861677408, + 0.004431603942066431, + 0.004464944824576378, + 0.012050308287143707, + 0.007106343749910593, + 0.003691112156957388, + 0.0033041401766240597, + 0.011204930953681469, + 0.009812472388148308, + 0.003402421250939369, + 0.0037430215161293745, + 0.0055021122097969055, + 0.006321310997009277, + 0.002829871140420437, + 0.0036249307449907064, + 0.011258628219366074, + 0.007351486478000879, + 0.004067855421453714, + 0.003016073489561677, + 0.002935956232249737, + 0.0022439854219555855, + 0.0017308766255155206, + 0.001922283903695643, + 0.0030537054408341646, + 0.0026407449040561914, + 0.00230035581625998, + 0.002380314050242305, + 0.0032827951945364475, + 0.0037355758249759674, + 0.003051429521292448, + 0.004104462917894125, + 0.006933846976608038, + 0.005454209633171558, + 0.0030198534950613976, + 0.004345138557255268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.002198542468249798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject the spiny underside of a massive Victoria Amazon water lily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lily's ribs reddish-purple and capped with sharp spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do oxygen bubbles form on the lily, detach, and rise upward out of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541004456579685 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012541004456579685, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541004456579685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058887071907520294, + 0.0547039657831192, + 0.3185408115386963, + 0.38717716932296753, + 0.03750164434313774, + 0.03807021677494049, + 0.02278229035437107, + 0.03502868860960007, + 0.020622704178094864, + 0.026685472577810287 + ], + "credit_map": [ + 0.014698267914354801, + 0.008095408789813519, + 0.01386245060712099, + 0.004537208005785942, + 0.002688275184482336, + 0.00806095078587532, + 0.003613673150539398, + 0.003330835374072194, + 0.004714015871286392, + 0.012452542781829834, + 0.004709210712462664, + 0.004740344360470772, + 0.0053465901874005795, + 0.008102008141577244, + 0.0077377245761454105, + 0.006901526357978582, + 0.02931646816432476, + 0.060598138719797134, + 0.025431953370571136, + 0.037419527769088745, + 0.03437658026814461, + 0.04682356119155884, + 0.03732905164361, + 0.04724550619721413, + 0.02674158848822117, + 0.05021774023771286, + 0.029497694224119186, + 0.10318297892808914, + 0.04233914613723755, + 0.033013489097356796, + 0.08153359591960907, + 0.020650943741202354, + 0.0039453813806176186, + 0.004578074906021357, + 0.003929072059690952, + 0.004987328313291073, + 0.0066559999249875546, + 0.004142926540225744, + 0.00537724606692791, + 0.00388561119325459, + 0.005322811659425497, + 0.006295337807387114, + 0.00248083658516407, + 0.003123221453279257, + 0.009782686829566956, + 0.005892706569284201, + 0.00242712558247149, + 0.0027454865630716085, + 0.004873627796769142, + 0.0026349315885454416, + 0.001312128733843565, + 0.0013936478644609451, + 0.005216551478952169, + 0.004104852210730314, + 0.001731325057335198, + 0.001515223877504468, + 0.003605979261919856, + 0.005556332413107157, + 0.001359609654173255, + 0.0021161020267754793, + 0.010357867926359177, + 0.007239094935357571, + 0.002812470542266965, + 0.0019812327809631824, + 0.0025069043040275574, + 0.0026831193827092648, + 0.0017108976608142257, + 0.002113606547936797, + 0.003844979451969266, + 0.002803963841870427, + 0.0030444844160228968, + 0.00191474927123636, + 0.0029519018717110157, + 0.003446738002821803, + 0.002036161720752716, + 0.0027263457886874676, + 0.005988230463117361, + 0.003929352853447199, + 0.0022461179178208113, + 0.0033606216311454773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.012541004456579685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_tensor([7238, 7238])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_tensor([7238, 7238])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..46857cd86137fbba113c0b1d4c269c1d0ec5c53b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_tensor([7238, 7238])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Static, long takes employ a surreal macrophotographic aesthetic to create a scene of intense, naturalistic terror. The camera, positioned at a distance and employing a powerful telephoto macro lens, captures the giant Galapagos centipede with unsettling intimacy. The creature is defensively coiled in a tight, menacing spiral on the jagged, porous surface of black volcanic rock. Every part of its powerful body is rendered with crystalline, formidable clarity; the overlapping reddish-brown chitin plates of its exoskeleton take on a subtle, oily sheen in the harsh, direct sunlight. Its countless, vibrant yellow legs are held rigidly, erected like a spiny defense mechanism. At the center of the coil, its head is tucked in, yet its long, delicate tentacles twitch almost imperceptibly, sensing the air with palpable alertness. The shallow depth of field casts the expansive background in a complete, painterly blur of soft ochres and desolate greens, effectively isolating the centipede and emphasizing its alien presence. Harsh lighting carves deep, sharp shadows across its body, highlighting the terrifying textures and architecture of the predator frozen in pure defensive threat. The overall atmosphere is one of profound isolation and silent, intimidating power.", + "seed": 9507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_7238_seed_9507_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08510330012068153, + "vlm_logprob": -0.08510330012068153, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.08510330012068153, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08548081666231155 + ], + "opposite_logprob_tokens": [ + -2.5854809284210205 + ], + "logprob_reward": -0.08548081666231155, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08548081666231155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0809471607208252, + 0.229249969124794, + 0.04066718369722366, + 0.025602778419852257, + 0.041124604642391205, + 0.013033004477620125, + 0.08194712549448013, + 0.21024949848651886, + 0.25453653931617737, + 0.02264217473566532 + ], + "credit_map": [ + 0.009232673794031143, + 0.008237685076892376, + 0.009056000038981438, + 0.0071343109011650085, + 0.014298542402684689, + 0.01340428926050663, + 0.009056470356881618, + 0.010527176782488823, + 0.01384531706571579, + 0.016386955976486206, + 0.011978225782513618, + 0.00943685881793499, + 0.07766595482826233, + 0.0648096352815628, + 0.02279309369623661, + 0.012333898805081844, + 0.003787276102229953, + 0.003087699180468917, + 0.002000091364607215, + 0.0017913891933858395, + 0.013528862036764622, + 0.011511045508086681, + 0.003048867918550968, + 0.0019119484350085258, + 0.00216458598151803, + 0.0018087126081809402, + 0.0012613595463335514, + 0.0015179404290392995, + 0.008065367117524147, + 0.006463600788265467, + 0.002307430375367403, + 0.0020137785468250513, + 0.0034139740746468306, + 0.003583459882065654, + 0.002439730567857623, + 0.0031737470999360085, + 0.011662578210234642, + 0.009113692678511143, + 0.003923123236745596, + 0.00381429772824049, + 0.0021897407714277506, + 0.001411833567544818, + 0.000881942396517843, + 0.0010036975145339966, + 0.0023733589332550764, + 0.002143161604180932, + 0.0016031782142817974, + 0.0014260906027629972, + 0.0052943723276257515, + 0.004139089025557041, + 0.0017594696255400777, + 0.0019363543251529336, + 0.03289385512471199, + 0.03180193528532982, + 0.002653268165886402, + 0.00146877474617213, + 0.01192659791558981, + 0.010192545130848885, + 0.006679472979158163, + 0.004386332351714373, + 0.08357544243335724, + 0.07870115339756012, + 0.007718948181718588, + 0.007068990264087915, + 0.017222875729203224, + 0.011273588985204697, + 0.005674658343195915, + 0.006509274244308472, + 0.10071240365505219, + 0.09725983440876007, + 0.009299702011048794, + 0.00658419169485569, + 0.002763621974736452, + 0.001419209293089807, + 0.002075125463306904, + 0.0018695852486416698, + 0.004790091887116432, + 0.00520961033180356, + 0.0031146188266575336, + 0.0014003103133291006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.08548081666231155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330298162996769 + ], + "opposite_logprob_tokens": [ + -3.7733030319213867 + ], + "logprob_reward": -0.02330298162996769, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330298162996769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04456586390733719, + 0.09687570482492447, + 0.038649674504995346, + 0.02420686185359955, + 0.021140096709132195, + 0.01789204031229019, + 0.36577096581459045, + 0.027158455923199654, + 0.08673729002475739, + 0.27700304985046387 + ], + "credit_map": [ + 0.0056328908540308475, + 0.005254886578768492, + 0.003624217351898551, + 0.003850779961794615, + 0.008609666489064693, + 0.0093429209664464, + 0.0037407532799988985, + 0.00450974702835083, + 0.006546402350068092, + 0.009405543096363544, + 0.002527041593566537, + 0.003362080780789256, + 0.03593461960554123, + 0.026941703632473946, + 0.00838752742856741, + 0.0037707716692239046, + 0.004584923852235079, + 0.0035320103634148836, + 0.0031603407114744186, + 0.0024727885611355305, + 0.009691323153674603, + 0.009430934675037861, + 0.0032918527722358704, + 0.0024854964576661587, + 0.0019670906476676464, + 0.0020634334068745375, + 0.0014711208641529083, + 0.0019891555421054363, + 0.0071701910346746445, + 0.005940270144492388, + 0.0023776497691869736, + 0.0012279473012313247, + 0.0026798935141414404, + 0.002643158193677664, + 0.001776242977939546, + 0.002228460041806102, + 0.0021946311462670565, + 0.0037908463273197412, + 0.002920679748058319, + 0.0029061820823699236, + 0.00134277215693146, + 0.0011331404093652964, + 0.0008992517250590026, + 0.0007720111752860248, + 0.005985700525343418, + 0.005406545475125313, + 0.0014059430686756968, + 0.0009466734481975436, + 0.022304702550172806, + 0.01638784259557724, + 0.005831086076796055, + 0.004664469510316849, + 0.15699075162410736, + 0.14900852739810944, + 0.0062984926626086235, + 0.004285066854208708, + 0.003296398324891925, + 0.0031240745447576046, + 0.003226682310923934, + 0.002755293855443597, + 0.004026822280138731, + 0.0053589604794979095, + 0.002843437483534217, + 0.0025267829187214375, + 0.006685735657811165, + 0.003472707699984312, + 0.0015503410249948502, + 0.0019470903789624572, + 0.03471966087818146, + 0.03381044790148735, + 0.0024999654851853848, + 0.0020513245835900307, + 0.016805367544293404, + 0.014297115616500378, + 0.004308512434363365, + 0.0029120505787432194, + 0.11766749620437622, + 0.11252889782190323, + 0.005571465939283371, + 0.002912136260420084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.02330298162996769 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159585773944855 + ], + "opposite_logprob_tokens": [ + -1.7015959024429321 + ], + "logprob_reward": -0.20159585773944855, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159585773944855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01710166595876217, + 0.04018261283636093, + 0.02502085454761982, + 0.02693803980946541, + 0.013516063801944256, + 0.005231163930147886, + 0.022708067670464516, + 0.026021072641015053, + 0.7789449691772461, + 0.04433555528521538 + ], + "credit_map": [ + 0.0026640493888407946, + 0.001977117732167244, + 0.0017617861740291119, + 0.0012447740882635117, + 0.002264605835080147, + 0.002696253592148423, + 0.0020546226296573877, + 0.002438455121591687, + 0.0025343962479382753, + 0.0024983608163893223, + 0.0018632024293765426, + 0.0015500823501497507, + 0.013892357237637043, + 0.011647514067590237, + 0.004238672088831663, + 0.0019580270163714886, + 0.0025054789148271084, + 0.0023044031113386154, + 0.0022533980663865805, + 0.001704612048342824, + 0.0062802014872431755, + 0.004943774547427893, + 0.0026304475031793118, + 0.0023985346779227257, + 0.0016718173865228891, + 0.0012388996547088027, + 0.0006573644932359457, + 0.0008615501574240625, + 0.010277429595589638, + 0.009860401041805744, + 0.001676502637565136, + 0.0006940704770386219, + 0.0016110800206661224, + 0.001747327740304172, + 0.0014343574875965714, + 0.0013983267126604915, + 0.0018443955341354012, + 0.0014335504965856671, + 0.002281952416524291, + 0.0017650715308263898, + 0.0005739172920584679, + 0.0006520649767480791, + 0.0005122395232319832, + 0.0007135446649044752, + 0.000795105064753443, + 0.0007739995489828289, + 0.0005633163964375854, + 0.0006469759391620755, + 0.001419559819623828, + 0.0010931749129667878, + 0.00039816444041207433, + 0.0004281510482542217, + 0.009425570257008076, + 0.009033259004354477, + 0.0005349272978492081, + 0.00037525914376601577, + 0.0016706091118976474, + 0.0017526509473100305, + 0.000833175319712609, + 0.0010396428406238556, + 0.009708018973469734, + 0.00929246749728918, + 0.000985151855275035, + 0.0007393527776002884, + 0.04680725932121277, + 0.028312740847468376, + 0.009534509852528572, + 0.01334453746676445, + 0.3335973024368286, + 0.3239738941192627, + 0.01580066792666912, + 0.007573913317173719, + 0.002806340577080846, + 0.002339352620765567, + 0.0010715139796957374, + 0.0018380003748461604, + 0.01699812524020672, + 0.01607322320342064, + 0.0018964115297421813, + 0.0013125849654898047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.20159585773944855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293937265872955 + ], + "opposite_logprob_tokens": [ + -2.2379393577575684 + ], + "logprob_reward": -0.11293937265872955, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293937265872955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025820821523666382, + 0.02591465413570404, + 0.02662709169089794, + 0.02586141973733902, + 0.019468849524855614, + 0.003348250174894929, + 0.10163471102714539, + 0.015827683731913567, + 0.6663029789924622, + 0.08919361233711243 + ], + "credit_map": [ + 0.002373481635004282, + 0.0023582426365464926, + 0.002524566138163209, + 0.0018277793424203992, + 0.005420234054327011, + 0.00606575096026063, + 0.002193732652813196, + 0.0030570337548851967, + 0.001652460196055472, + 0.00213017244823277, + 0.0008892891928553581, + 0.000860919535625726, + 0.009453952312469482, + 0.007636093068867922, + 0.002105731051415205, + 0.0011860349914059043, + 0.0017523267306387424, + 0.0014753063442185521, + 0.0011406178819015622, + 0.0010344355832785368, + 0.009811420924961567, + 0.00874212384223938, + 0.0013305718312039971, + 0.0013402870390564203, + 0.001806774758733809, + 0.0015241638757288456, + 0.0013982075033709407, + 0.0013343656901270151, + 0.007516131270676851, + 0.007244373671710491, + 0.0031957626342773438, + 0.0018416395178064704, + 0.002099759876728058, + 0.0025498871691524982, + 0.0016895928420126438, + 0.0024234491866081953, + 0.002575167454779148, + 0.002384625840932131, + 0.0029189479537308216, + 0.0028274187352508307, + 0.00043653754983097315, + 0.00040376544347964227, + 0.00017942898557521403, + 0.00021786056458950043, + 0.0007829234236851335, + 0.0006905023474246264, + 0.0003134889411740005, + 0.00032374271540902555, + 0.006300223525613546, + 0.004752091597765684, + 0.001757709076628089, + 0.001627570018172264, + 0.04276973009109497, + 0.04086711257696152, + 0.002040034392848611, + 0.0015202437061816454, + 0.0012167986715212464, + 0.0010542416712269187, + 0.0004311819502618164, + 0.0005060471594333649, + 0.005936178378760815, + 0.00561662670224905, + 0.0005558709963224828, + 0.0005107392207719386, + 0.040000032633543015, + 0.024102430790662766, + 0.008132023736834526, + 0.01137488428503275, + 0.2855050265789032, + 0.27722254395484924, + 0.013501176610589027, + 0.006464870180934668, + 0.005185319576412439, + 0.004654350690543652, + 0.0014540855772793293, + 0.0009841097053140402, + 0.03774101287126541, + 0.03634453937411308, + 0.0019097777549177408, + 0.00092042435426265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.11293937265872955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021974719129502773 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021974719129502773, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021974719129502773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07671765983104706, + 0.09241267293691635, + 0.030812358483672142, + 0.038468122482299805, + 0.07058265060186386, + 0.01000475138425827, + 0.1171947494149208, + 0.043316155672073364, + 0.29157954454421997, + 0.228911355137825 + ], + "credit_map": [ + 0.009285856038331985, + 0.00697738490998745, + 0.006597972009330988, + 0.00719669321551919, + 0.020824791863560677, + 0.01146078109741211, + 0.0050698923878371716, + 0.009304288774728775, + 0.0068085333332419395, + 0.006607117597013712, + 0.002909976290538907, + 0.0038085817359387875, + 0.03137456253170967, + 0.02745911292731762, + 0.008810190483927727, + 0.004634594079107046, + 0.003750798525288701, + 0.002753442618995905, + 0.0022241855040192604, + 0.002260088687762618, + 0.008401859551668167, + 0.006578459870070219, + 0.0027806980069726706, + 0.0020628266502171755, + 0.004998937249183655, + 0.003340289229527116, + 0.0021866450551897287, + 0.0023775282315909863, + 0.007633879315108061, + 0.009570249356329441, + 0.00460205739364028, + 0.0037585378158837557, + 0.006799268536269665, + 0.005329736974090338, + 0.0036672332789748907, + 0.004484063480049372, + 0.01994963362812996, + 0.01655588485300541, + 0.006844391580671072, + 0.006952434778213501, + 0.0011203264584764838, + 0.0013699372066184878, + 0.0007544350228272378, + 0.0008262274204753339, + 0.0016677030362188816, + 0.002290453063324094, + 0.0009431032231077552, + 0.001032565487548709, + 0.007094436325132847, + 0.005828221328556538, + 0.0026145714800804853, + 0.0024139098823070526, + 0.0477270632982254, + 0.04530875012278557, + 0.004044536035507917, + 0.0021632674615830183, + 0.004164000973105431, + 0.004466161131858826, + 0.0031745247542858124, + 0.002452246146276593, + 0.010944994166493416, + 0.010983174666762352, + 0.0042161704041063786, + 0.0029148831963539124, + 0.017937347292900085, + 0.01082546915858984, + 0.003941074479371309, + 0.00531281903386116, + 0.12358083575963974, + 0.11960682272911072, + 0.006215559784322977, + 0.004159609321504831, + 0.013246141374111176, + 0.011627263389527798, + 0.003677073400467634, + 0.0026169840712100267, + 0.09733749181032181, + 0.09317722916603088, + 0.004884739406406879, + 0.0023444255348294973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.0021974719129502773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, long takes employ a surreal macrophotographic aesthetic to create a scene of intense, naturalistic terror. The camera, positioned at a distance and employing a powerful telephoto macro lens, captures the giant Galapagos centipede with unsettling intimacy. The creature is defensively coiled in a tight, menacing spiral on the jagged, porous surface of black volcanic rock. Every part of its powerful body is rendered with crystalline, formidable clarity; the overlapping reddish-brown chitin plates of its exoskeleton take on a subtle, oily sheen in the harsh, direct sunlight. Its countless, vibrant yellow legs are held rigidly, erected like a spiny defense mechanism. At the center of the coil, its head is tucked in, yet its long, delicate tentacles twitch almost imperceptibly, sensing the air with palpable alertness. The shallow depth of field casts the expansive background in a complete, painterly blur of soft ochres and desolate greens, effectively isolating the centipede and emphasizing its alien presence. Harsh lighting carves deep, sharp shadows across its body, highlighting the terrifying textures and architecture of the predator frozen in pure defensive threat. The overall atmosphere is one of profound isolation and silent, intimidating power.", + "seed": 9508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_7238_seed_9508_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17828455781564118, + "vlm_logprob": -0.17828455781564118, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.17828455781564118, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6339542269706726 + ], + "opposite_logprob_tokens": [ + -0.7589542269706726 + ], + "logprob_reward": -0.6339542269706726, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6339542269706726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19467830657958984, + 0.30408963561058044, + 0.031782735139131546, + 0.03126358613371849, + 0.05572851747274399, + 0.0327332429587841, + 0.035363126546144485, + 0.011849850416183472, + 0.2503221929073334, + 0.052188772708177567 + ], + "credit_map": [ + 0.0199162345379591, + 0.03099009580910206, + 0.017849985510110855, + 0.020879186689853668, + 0.03573663905262947, + 0.022447476163506508, + 0.02633834071457386, + 0.020520346239209175, + 0.03480260446667671, + 0.04765433445572853, + 0.020898375660181046, + 0.050092995166778564, + 0.049020737409591675, + 0.03429768979549408, + 0.038912635296583176, + 0.028410250321030617, + 0.004128520376980305, + 0.007767679635435343, + 0.001975756837055087, + 0.0038706022314727306, + 0.0040496764704585075, + 0.004219795577228069, + 0.0028799823485314846, + 0.002890720497816801, + 0.002910773502662778, + 0.003959393594413996, + 0.004150877241045237, + 0.004238096531480551, + 0.0036580839660018682, + 0.003543480532243848, + 0.004788105376064777, + 0.004014774691313505, + 0.0061980923637747765, + 0.00967771839350462, + 0.005863209255039692, + 0.007504734676331282, + 0.008084364235401154, + 0.006240554619580507, + 0.007085583172738552, + 0.00507426168769598, + 0.0027470721397548914, + 0.005913871806114912, + 0.0035978297237306833, + 0.004158595111221075, + 0.004674127791076899, + 0.0038719933945685625, + 0.003481210907921195, + 0.004288541153073311, + 0.0034465373028069735, + 0.0037440943997353315, + 0.0028733049985021353, + 0.00487907649949193, + 0.0043250080198049545, + 0.00460090721026063, + 0.004284332972019911, + 0.007209863048046827, + 0.0012792079942300916, + 0.0014587410259991884, + 0.0016243371646851301, + 0.0020950587932020426, + 0.00129714235663414, + 0.001352903083898127, + 0.0015423194272443652, + 0.0012001401046290994, + 0.027783352881669998, + 0.03749120607972145, + 0.02402089163661003, + 0.028467055410146713, + 0.02614637278020382, + 0.029408082365989685, + 0.03410591930150986, + 0.04289931058883667, + 0.005718780215829611, + 0.010380961000919342, + 0.0038917004130780697, + 0.005455414764583111, + 0.00836688932031393, + 0.00622938759624958, + 0.006460502278059721, + 0.005685137584805489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.6339542269706726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026367327198386192 + ], + "opposite_logprob_tokens": [ + -3.651367425918579 + ], + "logprob_reward": -0.026367327198386192, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026367327198386192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12119785696268082, + 0.20505185425281525, + 0.07749713957309723, + 0.07183105498552322, + 0.05936666950583458, + 0.06647426635026932, + 0.03090200014412403, + 0.0289822556078434, + 0.2780188024044037, + 0.06067811697721481 + ], + "credit_map": [ + 0.012472153641283512, + 0.018167242407798767, + 0.018219217658042908, + 0.010947722010314465, + 0.018354877829551697, + 0.01422202680259943, + 0.012075523845851421, + 0.016739092767238617, + 0.02172015979886055, + 0.0272444486618042, + 0.014734014868736267, + 0.03883896395564079, + 0.02454327419400215, + 0.0297397430986166, + 0.030729634687304497, + 0.017501618713140488, + 0.009229703806340694, + 0.016684435307979584, + 0.00980245042592287, + 0.009663810953497887, + 0.008790063671767712, + 0.009211147204041481, + 0.007062448654323816, + 0.007053079549223185, + 0.006137373391538858, + 0.008439532481133938, + 0.011332346126437187, + 0.010481179691851139, + 0.008232537657022476, + 0.008538271300494671, + 0.009987497702240944, + 0.008682317100465298, + 0.00577967707067728, + 0.01058167964220047, + 0.00671409722417593, + 0.00841256882995367, + 0.00845547765493393, + 0.007222218904644251, + 0.0071939630433917046, + 0.005006986670196056, + 0.0065453858114778996, + 0.011155984364449978, + 0.0071510737761855125, + 0.008325597271323204, + 0.009479092434048653, + 0.007586301304399967, + 0.0089433453977108, + 0.007287485525012016, + 0.004085956607013941, + 0.004008576273918152, + 0.002867390401661396, + 0.0034270314499735832, + 0.004256300162523985, + 0.004328187555074692, + 0.003274344140663743, + 0.00465421425178647, + 0.0030354615300893784, + 0.0034014820121228695, + 0.003717333311215043, + 0.004937179386615753, + 0.00388238113373518, + 0.003596930531784892, + 0.0033903636503964663, + 0.0030211254488676786, + 0.03111746348440647, + 0.036851055920124054, + 0.031458038836717606, + 0.03330458700656891, + 0.026410402730107307, + 0.031731098890304565, + 0.03836815431714058, + 0.04877801984548569, + 0.006533773150295019, + 0.013114218600094318, + 0.004419268574565649, + 0.005969437770545483, + 0.010734830982983112, + 0.007539670914411545, + 0.005661186762154102, + 0.00670572929084301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.026367327198386192 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992809414863586 + ], + "opposite_logprob_tokens": [ + -1.8049280643463135 + ], + "logprob_reward": -0.17992809414863586, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992809414863586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11739982664585114, + 0.2946932017803192, + 0.064759761095047, + 0.08605489879846573, + 0.08053901791572571, + 0.08647900819778442, + 0.03229382261633873, + 0.023374225944280624, + 0.13689970970153809, + 0.07750649005174637 + ], + "credit_map": [ + 0.015607396140694618, + 0.017101990059018135, + 0.010261722840368748, + 0.012254673056304455, + 0.020721720531582832, + 0.013726544566452503, + 0.015407683327794075, + 0.012318105436861515, + 0.028020814061164856, + 0.06153705716133118, + 0.023122048005461693, + 0.04619475081562996, + 0.04835150018334389, + 0.032332099974155426, + 0.03176918625831604, + 0.023365749046206474, + 0.005891342181712389, + 0.008418391458690166, + 0.004383789375424385, + 0.01234576664865017, + 0.008996129035949707, + 0.008077984675765038, + 0.011257506906986237, + 0.00538885872811079, + 0.00800247211009264, + 0.01065504364669323, + 0.013815728016197681, + 0.010755935683846474, + 0.011383693665266037, + 0.009847428649663925, + 0.010351781733334064, + 0.0112428218126297, + 0.00923746358603239, + 0.01237990241497755, + 0.006244986783713102, + 0.00878293439745903, + 0.011510984972119331, + 0.009710161946713924, + 0.012460939586162567, + 0.010211651213467121, + 0.009944325312972069, + 0.02343130111694336, + 0.004089007619768381, + 0.011567410081624985, + 0.012990027666091919, + 0.011197230778634548, + 0.007528276182711124, + 0.005731438752263784, + 0.005344249773770571, + 0.004770857281982899, + 0.002667575143277645, + 0.002656172029674053, + 0.004032819997519255, + 0.004095070529729128, + 0.0028155429754406214, + 0.005911539308726788, + 0.002320079365745187, + 0.00282902829349041, + 0.0022373292595148087, + 0.003673048922792077, + 0.003251422429457307, + 0.003253264119848609, + 0.003065928118303418, + 0.002744128694757819, + 0.017858998849987984, + 0.017181580886244774, + 0.007672533858567476, + 0.02465534582734108, + 0.014612656086683273, + 0.018148699775338173, + 0.021162396296858788, + 0.015607497654855251, + 0.0083828866481781, + 0.007997564971446991, + 0.015285824425518513, + 0.007386700250208378, + 0.006119273137301207, + 0.00895073264837265, + 0.006817455869168043, + 0.016566066071391106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.17992809414863586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868203029036522 + ], + "opposite_logprob_tokens": [ + -3.0486819744110107 + ], + "logprob_reward": -0.04868203029036522, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868203029036522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16172875463962555, + 0.10710234940052032, + 0.07759333401918411, + 0.04204653203487396, + 0.1071246862411499, + 0.1249835342168808, + 0.03941413760185242, + 0.0853118896484375, + 0.1756005883216858, + 0.07909408211708069 + ], + "credit_map": [ + 0.016134118661284447, + 0.02447439916431904, + 0.018950678408145905, + 0.013596932403743267, + 0.028990309685468674, + 0.01970669999718666, + 0.018229199573397636, + 0.021646425127983093, + 0.011236926540732384, + 0.011754948645830154, + 0.01236130390316248, + 0.018167125061154366, + 0.013820127584040165, + 0.01192336156964302, + 0.01699194125831127, + 0.010846615768969059, + 0.008425937034189701, + 0.016022685915231705, + 0.00766994571313262, + 0.011669312603771687, + 0.009330240078270435, + 0.009361145086586475, + 0.007027957122772932, + 0.008086113259196281, + 0.004492809530347586, + 0.00978743378072977, + 0.0040668039582669735, + 0.0034095924347639084, + 0.00494740903377533, + 0.005261068232357502, + 0.003917715046554804, + 0.006163700949400663, + 0.00942198559641838, + 0.023722708225250244, + 0.0075376261956989765, + 0.014793816953897476, + 0.016289735212922096, + 0.013207368552684784, + 0.012014912441372871, + 0.010136536322534084, + 0.013909989036619663, + 0.021379504352808, + 0.014258340001106262, + 0.015352939255535603, + 0.01638738438487053, + 0.014664758928120136, + 0.014429103583097458, + 0.014601512812077999, + 0.004607310518622398, + 0.005417616572231054, + 0.004754174035042524, + 0.004960314836353064, + 0.004617622587829828, + 0.004959449637681246, + 0.005218416918069124, + 0.004879231099039316, + 0.009623861871659756, + 0.012859943322837353, + 0.010417753830552101, + 0.009677436202764511, + 0.009884843602776527, + 0.011435147374868393, + 0.01171070896089077, + 0.009702196344733238, + 0.022261139005422592, + 0.02558842860162258, + 0.01932840421795845, + 0.014726592227816582, + 0.01746446080505848, + 0.01990465819835663, + 0.021499117836356163, + 0.03482779115438461, + 0.010369865223765373, + 0.018322836607694626, + 0.004160815384238958, + 0.007676843553781509, + 0.0126084815710783, + 0.009457673877477646, + 0.008204030804336071, + 0.008293533697724342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.04868203029036522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024911104701459408 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.0024911104701459408, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024911104701459408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16102416813373566, + 0.17554740607738495, + 0.060322560369968414, + 0.06578017771244049, + 0.05260661244392395, + 0.0361025370657444, + 0.03603503108024597, + 0.04739347845315933, + 0.32116249203681946, + 0.044025443494319916 + ], + "credit_map": [ + 0.015210958197712898, + 0.02488996833562851, + 0.016407374292612076, + 0.016308248043060303, + 0.027852946892380714, + 0.018480226397514343, + 0.022592004388570786, + 0.01928246021270752, + 0.0231963898986578, + 0.017569690942764282, + 0.012578082270920277, + 0.031547512859106064, + 0.024233968928456306, + 0.02272002026438713, + 0.026487769559025764, + 0.01721399463713169, + 0.00786120817065239, + 0.012940926477313042, + 0.005019051022827625, + 0.007326419930905104, + 0.007897158153355122, + 0.007824832573533058, + 0.006015582475811243, + 0.005437388084828854, + 0.005652254447340965, + 0.008925181813538074, + 0.007293867878615856, + 0.009695128537714481, + 0.006977566983550787, + 0.007090813014656305, + 0.008158382028341293, + 0.0119869913905859, + 0.00515442481264472, + 0.007367659825831652, + 0.003795818192884326, + 0.008910962380468845, + 0.008400991559028625, + 0.006197068840265274, + 0.007341018877923489, + 0.005438675172626972, + 0.003130705561488867, + 0.006720090284943581, + 0.005732021294534206, + 0.003691570833325386, + 0.004974078852683306, + 0.00405533192679286, + 0.0034239524975419044, + 0.0043747881427407265, + 0.0034772618673741817, + 0.0037491738330572844, + 0.0026760592591017485, + 0.004988917615264654, + 0.004035121761262417, + 0.004043335560709238, + 0.004331206902861595, + 0.008733958937227726, + 0.005070963408797979, + 0.006377130746841431, + 0.007499647792428732, + 0.00563594326376915, + 0.0044707125052809715, + 0.0053587257862091064, + 0.005696749780327082, + 0.007283610291779041, + 0.03351506218314171, + 0.04224659875035286, + 0.02165113389492035, + 0.053992416709661484, + 0.0335763543844223, + 0.04091142117977142, + 0.05161530151963234, + 0.04365423321723938, + 0.004550686571747065, + 0.005447568371891975, + 0.006463471334427595, + 0.004359912127256393, + 0.0056890565901994705, + 0.0055568101815879345, + 0.005251715425401926, + 0.006706223823130131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0024911104701459408 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, long takes employ a surreal macrophotographic aesthetic to create a scene of intense, naturalistic terror. The camera, positioned at a distance and employing a powerful telephoto macro lens, captures the giant Galapagos centipede with unsettling intimacy. The creature is defensively coiled in a tight, menacing spiral on the jagged, porous surface of black volcanic rock. Every part of its powerful body is rendered with crystalline, formidable clarity; the overlapping reddish-brown chitin plates of its exoskeleton take on a subtle, oily sheen in the harsh, direct sunlight. Its countless, vibrant yellow legs are held rigidly, erected like a spiny defense mechanism. At the center of the coil, its head is tucked in, yet its long, delicate tentacles twitch almost imperceptibly, sensing the air with palpable alertness. The shallow depth of field casts the expansive background in a complete, painterly blur of soft ochres and desolate greens, effectively isolating the centipede and emphasizing its alien presence. Harsh lighting carves deep, sharp shadows across its body, highlighting the terrifying textures and architecture of the predator frozen in pure defensive threat. The overall atmosphere is one of profound isolation and silent, intimidating power.", + "seed": 9509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_7238_seed_9509_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09841111358255147, + "vlm_logprob": -0.09841111358255147, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.09841111358255147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12023626267910004 + ], + "opposite_logprob_tokens": [ + -2.2452361583709717 + ], + "logprob_reward": -0.12023626267910004, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12023626267910004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09168580174446106, + 0.17987556755542755, + 0.10012493282556534, + 0.23369838297367096, + 0.07225515693426132, + 0.07604575902223587, + 0.11074690520763397, + 0.07864521443843842, + 0.024689948186278343, + 0.032232288271188736 + ], + "credit_map": [ + 0.01720470003783703, + 0.014185754582285881, + 0.005994878243654966, + 0.010920541360974312, + 0.007217471953481436, + 0.013568623922765255, + 0.010658558458089828, + 0.01193529088050127, + 0.02484380640089512, + 0.020250149071216583, + 0.01228791568428278, + 0.018542397767305374, + 0.015406765975058079, + 0.01724478229880333, + 0.04969603568315506, + 0.021603722125291824, + 0.016202954575419426, + 0.017753884196281433, + 0.011982626281678677, + 0.012094898149371147, + 0.009587086737155914, + 0.010145879350602627, + 0.013595388270914555, + 0.008762223646044731, + 0.02919968217611313, + 0.02245347574353218, + 0.02311197854578495, + 0.03433586284518242, + 0.023277290165424347, + 0.029476668685674667, + 0.04090718924999237, + 0.03093625232577324, + 0.012050164863467216, + 0.012497122399508953, + 0.005561555735766888, + 0.012924096547067165, + 0.005453689955174923, + 0.008434418588876724, + 0.008188612759113312, + 0.007145510520786047, + 0.01876700110733509, + 0.019874248653650284, + 0.0048088193871080875, + 0.00781470350921154, + 0.005505071021616459, + 0.0062850723043084145, + 0.007240073289722204, + 0.005750773474574089, + 0.04068365693092346, + 0.03332360461354256, + 0.0036722763907164335, + 0.004959783051162958, + 0.006083732005208731, + 0.00970597192645073, + 0.0056749386712908745, + 0.006642953027039766, + 0.03088194690644741, + 0.010127679444849491, + 0.0041139861568808556, + 0.004853877238929272, + 0.005097773391753435, + 0.006472082808613777, + 0.007151658181101084, + 0.00994622241705656, + 0.005050531588494778, + 0.0027669784612953663, + 0.0016536260955035686, + 0.0031806642655283213, + 0.0017990138148888946, + 0.0031080732587724924, + 0.0038232251536101103, + 0.0033078386913985014, + 0.012353520840406418, + 0.0036310891155153513, + 0.0012921197339892387, + 0.0027110388036817312, + 0.0022487526293843985, + 0.0029622006695717573, + 0.0022375781554728746, + 0.004795993212610483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.12023626267910004 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018208960071206093 + ], + "opposite_logprob_tokens": [ + -4.018208980560303 + ], + "logprob_reward": -0.018208960071206093, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018208960071206093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11949852108955383, + 0.04054700955748558, + 0.057468850165605545, + 0.23705007135868073, + 0.27809685468673706, + 0.053053855895996094, + 0.06491290777921677, + 0.0949615016579628, + 0.0250140018761158, + 0.02939639240503311 + ], + "credit_map": [ + 0.0154653862118721, + 0.021151384338736534, + 0.011225983500480652, + 0.019930994138121605, + 0.010550174862146378, + 0.013571234419941902, + 0.013961729593575, + 0.013641627505421638, + 0.005369715392589569, + 0.005211383569985628, + 0.004062848165631294, + 0.00480335159227252, + 0.004143702331930399, + 0.005893462337553501, + 0.006309994496405125, + 0.004752551671117544, + 0.00669722119346261, + 0.008118825033307076, + 0.005111789330840111, + 0.00997898355126381, + 0.0039229486137628555, + 0.006406456232070923, + 0.0100579634308815, + 0.00717466464266181, + 0.029886608943343163, + 0.02407398261129856, + 0.02585606276988983, + 0.03308875113725662, + 0.023025527596473694, + 0.03179173544049263, + 0.0397627130150795, + 0.02956469915807247, + 0.031183287501335144, + 0.0339282788336277, + 0.021400779485702515, + 0.03844011574983597, + 0.01699882559478283, + 0.032085493206977844, + 0.047872524708509445, + 0.05618754029273987, + 0.008017003536224365, + 0.007409147918224335, + 0.0047276439145207405, + 0.008812084794044495, + 0.004949739668518305, + 0.006405975669622421, + 0.006897346582263708, + 0.0058349124155938625, + 0.00973003264516592, + 0.0117599256336689, + 0.005474796984344721, + 0.0050089070573449135, + 0.005139050539582968, + 0.010069725103676319, + 0.006724806502461433, + 0.01100566703826189, + 0.042547788470983505, + 0.012717320583760738, + 0.003167050890624523, + 0.0031772516667842865, + 0.006140190176665783, + 0.008341962471604347, + 0.005621661432087421, + 0.013248275965452194, + 0.007037212140858173, + 0.002575820544734597, + 0.001255546580068767, + 0.0026565827429294586, + 0.001989945536479354, + 0.004866694565862417, + 0.0022630984894931316, + 0.002369101159274578, + 0.006644189357757568, + 0.0030107584316283464, + 0.0015292076859623194, + 0.002614052966237068, + 0.0020531504414975643, + 0.006861071567982435, + 0.0030298379715532064, + 0.0036541265435516834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.018208960071206093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813723683357239 + ], + "opposite_logprob_tokens": [ + -1.406372308731079 + ], + "logprob_reward": -0.2813723683357239, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813723683357239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16582812368869781, + 0.08800631761550903, + 0.26889780163764954, + 0.07307770848274231, + 0.09856107085943222, + 0.03500450402498245, + 0.11692295223474503, + 0.0616605319082737, + 0.022056791931390762, + 0.06998414546251297 + ], + "credit_map": [ + 0.02032284438610077, + 0.028861934319138527, + 0.013324016705155373, + 0.02603241056203842, + 0.014472694136202335, + 0.018119899556040764, + 0.025177108123898506, + 0.019517211243510246, + 0.011427994817495346, + 0.009664274752140045, + 0.006445454899221659, + 0.010560286231338978, + 0.008022885769605637, + 0.01038671936839819, + 0.02126162126660347, + 0.01023708377033472, + 0.03865283727645874, + 0.0459468849003315, + 0.030229663476347923, + 0.03260577842593193, + 0.024828262627124786, + 0.03040277026593685, + 0.040390584617853165, + 0.025841012597084045, + 0.009784378111362457, + 0.005779668688774109, + 0.004670607857406139, + 0.00801588874310255, + 0.005580469965934753, + 0.011669625528156757, + 0.012187467887997627, + 0.015389608219265938, + 0.015078206546604633, + 0.009165468625724316, + 0.007290998473763466, + 0.012474577873945236, + 0.007013712543994188, + 0.014407153241336346, + 0.012004035525023937, + 0.021126924082636833, + 0.007109563332051039, + 0.004947146400809288, + 0.003046009922400117, + 0.0036632732953876257, + 0.002975562121719122, + 0.005508235655725002, + 0.0034026920329779387, + 0.004352021496742964, + 0.03908294066786766, + 0.03470657020807266, + 0.004770924337208271, + 0.00576179102063179, + 0.007492981385439634, + 0.011116371490061283, + 0.006322228349745274, + 0.007669145241379738, + 0.025618871673941612, + 0.006510896608233452, + 0.0028186647687107325, + 0.003667144337669015, + 0.004032369703054428, + 0.0056020417250692844, + 0.004577355924993753, + 0.008833183906972408, + 0.003492983290925622, + 0.002179451286792755, + 0.0013740936992689967, + 0.001964493189007044, + 0.0019407345680519938, + 0.005370609927922487, + 0.0035270066000521183, + 0.0022074184380471706, + 0.02840477228164673, + 0.0076536862179636955, + 0.0022643418051302433, + 0.0035501469392329454, + 0.004617974627763033, + 0.007008482702076435, + 0.0047370921820402145, + 0.011747647076845169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.2813723683357239 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003860175609589 + ], + "opposite_logprob_tokens": [ + -2.6950385570526123 + ], + "logprob_reward": -0.07003860175609589, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003860175609589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11500082910060883, + 0.08974481374025345, + 0.08986254036426544, + 0.12562134861946106, + 0.2582733929157257, + 0.07269351929426193, + 0.0359342098236084, + 0.12939491868019104, + 0.051016513258218765, + 0.0324578620493412 + ], + "credit_map": [ + 0.02054508775472641, + 0.014534075744450092, + 0.008556640706956387, + 0.014036656357347965, + 0.010020670481026173, + 0.018176548182964325, + 0.014050262048840523, + 0.015080885961651802, + 0.012146792374551296, + 0.011423927731812, + 0.00806827936321497, + 0.01192433200776577, + 0.0075461165979504585, + 0.012233948335051537, + 0.015223992988467216, + 0.011177422478795052, + 0.013770109042525291, + 0.012299034744501114, + 0.01159586850553751, + 0.01432203408330679, + 0.00613253191113472, + 0.007825510576367378, + 0.014120891690254211, + 0.00979656632989645, + 0.017017584294080734, + 0.016950977966189384, + 0.015678593888878822, + 0.023875238373875618, + 0.00914541445672512, + 0.010562639683485031, + 0.015717193484306335, + 0.016673704609274864, + 0.038813941180706024, + 0.04254372790455818, + 0.020833183079957962, + 0.05987080931663513, + 0.01786154881119728, + 0.019500162452459335, + 0.034078869968652725, + 0.024771159514784813, + 0.009794033132493496, + 0.007168926298618317, + 0.006385888904333115, + 0.01127712707966566, + 0.0062146312557160854, + 0.011043696664273739, + 0.010665297508239746, + 0.010143916122615337, + 0.004862707573920488, + 0.005028205458074808, + 0.003398920875042677, + 0.0037223624531179667, + 0.00309646874666214, + 0.006665044464170933, + 0.0041581252589821815, + 0.005002375692129135, + 0.056272998452186584, + 0.016941189765930176, + 0.00544012850150466, + 0.007593653164803982, + 0.00862867757678032, + 0.009641597978770733, + 0.007990623824298382, + 0.016886049881577492, + 0.008201733231544495, + 0.004584330599755049, + 0.003577578580006957, + 0.006705062929540873, + 0.004195150453597307, + 0.00873561855405569, + 0.009386819787323475, + 0.00563021982088685, + 0.008025565184652805, + 0.002546470146626234, + 0.0023565776646137238, + 0.004789721220731735, + 0.0033460797276347876, + 0.0058926064521074295, + 0.0024025992024689913, + 0.003098242450505495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.07003860175609589 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199375070631504 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.002199375070631504, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199375070631504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15017050504684448, + 0.03493940085172653, + 0.03307552635669708, + 0.12102674692869186, + 0.13120773434638977, + 0.12471747398376465, + 0.24049262702465057, + 0.06224217638373375, + 0.05410941690206528, + 0.0480184406042099 + ], + "credit_map": [ + 0.031071791425347328, + 0.022004680708050728, + 0.011989612132310867, + 0.02331213280558586, + 0.01098267175257206, + 0.018025366589426994, + 0.016101831570267677, + 0.016682414337992668, + 0.0048543233424425125, + 0.003936926368623972, + 0.001930219936184585, + 0.0038250174839049578, + 0.0031089310068637133, + 0.00371548137627542, + 0.009432408027350903, + 0.004136092495173216, + 0.0049638524651527405, + 0.0036219204775989056, + 0.0026385567616671324, + 0.005725128576159477, + 0.0027940161526203156, + 0.004251529462635517, + 0.00504637137055397, + 0.004034151788800955, + 0.01472373865544796, + 0.014554771594703197, + 0.012449808418750763, + 0.01715470664203167, + 0.011228926479816437, + 0.017254125326871872, + 0.01864871382713318, + 0.015011953189969063, + 0.0153536107391119, + 0.017606470733880997, + 0.01039922796189785, + 0.01583855040371418, + 0.007599564269185066, + 0.018230265006422997, + 0.017454205080866814, + 0.028725838288664818, + 0.01967727765440941, + 0.015425223857164383, + 0.01101165171712637, + 0.01788465864956379, + 0.011609126813709736, + 0.0172925665974617, + 0.016117844730615616, + 0.015699120238423347, + 0.05371468514204025, + 0.05545654892921448, + 0.014332523569464684, + 0.014239056967198849, + 0.022193579003214836, + 0.03493031486868858, + 0.01855425164103508, + 0.027071673423051834, + 0.0233994759619236, + 0.007435832172632217, + 0.0028724134899675846, + 0.0030497959814965725, + 0.004110999405384064, + 0.009581910446286201, + 0.004086791537702084, + 0.007704955525696278, + 0.007630030158907175, + 0.004335084930062294, + 0.004564904607832432, + 0.008151788264513016, + 0.004331429488956928, + 0.0063893795013427734, + 0.011563763953745365, + 0.007143034134060144, + 0.018841858953237534, + 0.005506985355168581, + 0.0021019845735281706, + 0.001986111979931593, + 0.003659923793748021, + 0.006685788277536631, + 0.003057009307667613, + 0.006178777664899826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.002199375070631504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, long takes employ a surreal macrophotographic aesthetic to create a scene of intense, naturalistic terror. The camera, positioned at a distance and employing a powerful telephoto macro lens, captures the giant Galapagos centipede with unsettling intimacy. The creature is defensively coiled in a tight, menacing spiral on the jagged, porous surface of black volcanic rock. Every part of its powerful body is rendered with crystalline, formidable clarity; the overlapping reddish-brown chitin plates of its exoskeleton take on a subtle, oily sheen in the harsh, direct sunlight. Its countless, vibrant yellow legs are held rigidly, erected like a spiny defense mechanism. At the center of the coil, its head is tucked in, yet its long, delicate tentacles twitch almost imperceptibly, sensing the air with palpable alertness. The shallow depth of field casts the expansive background in a complete, painterly blur of soft ochres and desolate greens, effectively isolating the centipede and emphasizing its alien presence. Harsh lighting carves deep, sharp shadows across its body, highlighting the terrifying textures and architecture of the predator frozen in pure defensive threat. The overall atmosphere is one of profound isolation and silent, intimidating power.", + "seed": 9510, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_1/index_7238_seed_9510_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1663128669373691, + "vlm_logprob": -0.1663128669373691, + "vlm_yes_no_margin": 2.575000023841858, + "vlm_token_credit": -0.1663128669373691, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3524189293384552 + ], + "opposite_logprob_tokens": [ + -1.2274188995361328 + ], + "logprob_reward": -0.3524189293384552, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3524189293384552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17214439809322357, + 0.04706612974405289, + 0.0681881308555603, + 0.026549631729722023, + 0.2625349760055542, + 0.0650501698255539, + 0.07655216008424759, + 0.24084646999835968, + 0.024198513478040695, + 0.016869382932782173 + ], + "credit_map": [ + 0.01853896863758564, + 0.013222095556557178, + 0.00977321807295084, + 0.010817133821547031, + 0.024729589000344276, + 0.03958577290177345, + 0.018544834107160568, + 0.0369328074157238, + 0.0060244896449148655, + 0.007960707880556583, + 0.003887294325977564, + 0.00480547733604908, + 0.007971139624714851, + 0.006867916788905859, + 0.004748559556901455, + 0.004800545517355204, + 0.007716750260442495, + 0.006498917005956173, + 0.004669304471462965, + 0.007264130748808384, + 0.007367533165961504, + 0.0145516786724329, + 0.010468450374901295, + 0.009651369415223598, + 0.0037237887736409903, + 0.002548313234001398, + 0.0032662420999258757, + 0.003192467149347067, + 0.0031817969866096973, + 0.003770771436393261, + 0.0033365951385349035, + 0.0035296559799462557, + 0.02889864519238472, + 0.014871558174490929, + 0.005801717285066843, + 0.010776598937809467, + 0.0270000621676445, + 0.09847772121429443, + 0.05662410333752632, + 0.020084574818611145, + 0.0063800206407904625, + 0.007062101736664772, + 0.00536081800237298, + 0.007747388444840908, + 0.014903099276125431, + 0.0086759552359581, + 0.00835292786359787, + 0.006567863747477531, + 0.007047448772937059, + 0.006035058293491602, + 0.0032518524676561356, + 0.006629748269915581, + 0.01676511950790882, + 0.01909262128174305, + 0.01285083219408989, + 0.004879482556134462, + 0.038607511669397354, + 0.0187112744897604, + 0.012604559771716595, + 0.01616930030286312, + 0.021107610315084457, + 0.07708024978637695, + 0.035986077040433884, + 0.020579898729920387, + 0.003565615275874734, + 0.0030497193802148104, + 0.002285187831148505, + 0.0022841589525341988, + 0.0028119334019720554, + 0.004657443147152662, + 0.0025321899447590113, + 0.003012267407029867, + 0.0021515493281185627, + 0.0015343055129051208, + 0.0019150173757225275, + 0.0018549467204138637, + 0.0016784267500042915, + 0.004056681878864765, + 0.0024569614324718714, + 0.0012214957969263196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.3524189293384552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016088193282485008 + ], + "opposite_logprob_tokens": [ + -4.141088008880615 + ], + "logprob_reward": -0.016088193282485008, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016088193282485008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2970799207687378, + 0.11014246940612793, + 0.15244421362876892, + 0.051935795694589615, + 0.08335521072149277, + 0.045588329434394836, + 0.017759395763278008, + 0.05520819127559662, + 0.11445354670286179, + 0.07203284651041031 + ], + "credit_map": [ + 0.03001975268125534, + 0.02291441336274147, + 0.020521752536296844, + 0.02525351569056511, + 0.030817752704024315, + 0.06971127539873123, + 0.032252538949251175, + 0.06558892130851746, + 0.020506523549556732, + 0.015460767783224583, + 0.009812557138502598, + 0.0091866971924901, + 0.016157077625393867, + 0.0172483641654253, + 0.00968091655522585, + 0.01208957377821207, + 0.018065517768263817, + 0.01409584004431963, + 0.013147043064236641, + 0.014388412237167358, + 0.017609501257538795, + 0.03859197720885277, + 0.016796518117189407, + 0.019749419763684273, + 0.005470640026032925, + 0.008371704258024693, + 0.0036689939443022013, + 0.006080079358071089, + 0.007675277069211006, + 0.005691224709153175, + 0.005720457527786493, + 0.0092574218288064, + 0.010049893520772457, + 0.004401068203151226, + 0.0022008977830410004, + 0.004335338715463877, + 0.008822796866297722, + 0.030492868274450302, + 0.016477739438414574, + 0.006574616301804781, + 0.007271582260727882, + 0.004383310675621033, + 0.0054030222818255424, + 0.005628861021250486, + 0.00529944384470582, + 0.00572594627737999, + 0.00607821810990572, + 0.005797951947897673, + 0.0026151498313993216, + 0.002523591974750161, + 0.001616411260329187, + 0.0017949908506125212, + 0.0019718282856047153, + 0.0026255252305418253, + 0.0021536746062338352, + 0.0024582247715443373, + 0.00782318226993084, + 0.007525932043790817, + 0.003768092254176736, + 0.006824981886893511, + 0.005290526896715164, + 0.008817429654300213, + 0.008960578590631485, + 0.0061974734999239445, + 0.01363587286323309, + 0.00839176494628191, + 0.013597358949482441, + 0.01162254624068737, + 0.01926247775554657, + 0.017775045707821846, + 0.01846393756568432, + 0.011704562231898308, + 0.012651574797928333, + 0.006064278073608875, + 0.008018972352147102, + 0.006264712195843458, + 0.006498062051832676, + 0.016695350408554077, + 0.007887396961450577, + 0.007952507585287094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.016088193282485008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31349003314971924 + ], + "opposite_logprob_tokens": [ + -1.3134900331497192 + ], + "logprob_reward": -0.31349003314971924, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31349003314971924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23839697241783142, + 0.07260990887880325, + 0.05757136642932892, + 0.026859357953071594, + 0.08583694696426392, + 0.09814102947711945, + 0.05330004170536995, + 0.22071526944637299, + 0.07293854653835297, + 0.07363050431013107 + ], + "credit_map": [ + 0.02709585800766945, + 0.02234932780265808, + 0.01977161131799221, + 0.021461132913827896, + 0.04202849790453911, + 0.03753408044576645, + 0.027396632358431816, + 0.0407598502933979, + 0.010526743717491627, + 0.012756098993122578, + 0.007088224869221449, + 0.0062402402982115746, + 0.00734745804220438, + 0.014664847403764725, + 0.006036971230059862, + 0.007949327118694782, + 0.007871059700846672, + 0.00886708963662386, + 0.005609688349068165, + 0.00708761066198349, + 0.008052386343479156, + 0.007128812372684479, + 0.005693580489605665, + 0.0072611430659890175, + 0.0030455237720161676, + 0.0024253404699265957, + 0.002479453105479479, + 0.005118491593748331, + 0.0025994712486863136, + 0.003819982521235943, + 0.003934728447347879, + 0.0034363684244453907, + 0.009465344250202179, + 0.006513637024909258, + 0.004111096728593111, + 0.00654583191499114, + 0.008809424005448818, + 0.026791075244545937, + 0.01650911383330822, + 0.007091430481523275, + 0.013559145852923393, + 0.0107308030128479, + 0.013574615120887756, + 0.010548988357186317, + 0.0197439007461071, + 0.011014081537723541, + 0.011510946787893772, + 0.007458553183823824, + 0.006473167333751917, + 0.004753928165882826, + 0.003244269173592329, + 0.0046777380630373955, + 0.012352545745670795, + 0.012209959328174591, + 0.006043794099241495, + 0.0035446439869701862, + 0.024133436381816864, + 0.014949433505535126, + 0.014952500350773335, + 0.0243983194231987, + 0.017484497278928757, + 0.056808799505233765, + 0.04490543156862259, + 0.023082870990037918, + 0.011447811499238014, + 0.008487626910209656, + 0.00715352687984705, + 0.004965729545801878, + 0.007953541353344917, + 0.017517928034067154, + 0.00642947293817997, + 0.008982907980680466, + 0.013391761109232903, + 0.008806280791759491, + 0.007408882491290569, + 0.0069784075021743774, + 0.008395197801291943, + 0.013161322101950645, + 0.008856536820530891, + 0.006632114760577679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.31349003314971924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428298056125641 + ], + "opposite_logprob_tokens": [ + -2.0178298950195312 + ], + "logprob_reward": -0.1428298056125641, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428298056125641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20133467018604279, + 0.0638689249753952, + 0.06484200060367584, + 0.04137643799185753, + 0.3006468415260315, + 0.08837052434682846, + 0.03571433201432228, + 0.06771009415388107, + 0.07630287110805511, + 0.059833332896232605 + ], + "credit_map": [ + 0.029596587643027306, + 0.03736624866724014, + 0.0185791477560997, + 0.015465257689356804, + 0.03226393088698387, + 0.02843288704752922, + 0.021250726655125618, + 0.018379874527454376, + 0.013735979795455933, + 0.008390643633902073, + 0.006624420639127493, + 0.006105654872953892, + 0.008668340742588043, + 0.007030990440398455, + 0.00614324864000082, + 0.007169641554355621, + 0.0076207360252738, + 0.007069042883813381, + 0.005245584528893232, + 0.0074333306401968, + 0.007008896674960852, + 0.01293871272355318, + 0.00920243188738823, + 0.008323260582983494, + 0.004189515486359596, + 0.004908775445073843, + 0.0031424418557435274, + 0.006375396158546209, + 0.0053137014620006084, + 0.005064586643129587, + 0.004895707592368126, + 0.007486309390515089, + 0.030359631404280663, + 0.0149433184415102, + 0.007622705306857824, + 0.01656298339366913, + 0.03152840957045555, + 0.10989297926425934, + 0.06578688323497772, + 0.02394992485642433, + 0.010794216766953468, + 0.008181953802704811, + 0.01198570616543293, + 0.011507982388138771, + 0.014460353180766106, + 0.009712422266602516, + 0.012027389369904995, + 0.00970049574971199, + 0.004830435384064913, + 0.00411960668861866, + 0.004780994262546301, + 0.002873270772397518, + 0.006132751703262329, + 0.0059838551096618176, + 0.003516832133755088, + 0.003476582933217287, + 0.008633028715848923, + 0.007484630215913057, + 0.006446790415793657, + 0.006903005763888359, + 0.006641087122261524, + 0.012503629550337791, + 0.010701175779104233, + 0.00839674100279808, + 0.011673484928905964, + 0.011516797356307507, + 0.007227208930999041, + 0.0075046103447675705, + 0.011366271413862705, + 0.012439862824976444, + 0.0070336065255105495, + 0.0075410292483866215, + 0.011029353365302086, + 0.006025443784892559, + 0.005697022192180157, + 0.005883967503905296, + 0.006202637683600187, + 0.00817026849836111, + 0.010390874929726124, + 0.006433764006942511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.1428298056125641 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Galapagos centipede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centipede coiled in a tight spiral on a jagged, porous black volcanic rock surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the centipede's legs vibrant yellow and held rigidly erected?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede's head tucked in the center of the coil with long, delicate tentacles twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a painterly blur of soft ochres and desolate greens with shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737373303622007 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737373303622007, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737373303622007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12832994759082794, + 0.12571296095848083, + 0.06669337302446365, + 0.025348493829369545, + 0.18506808578968048, + 0.03362039476633072, + 0.030413327738642693, + 0.07819220423698425, + 0.12368930131196976, + 0.20293200016021729 + ], + "credit_map": [ + 0.015067986212670803, + 0.01311031449586153, + 0.014414328150451183, + 0.012680219486355782, + 0.013230467215180397, + 0.02521863766014576, + 0.013507945463061333, + 0.021100033074617386, + 0.02327018603682518, + 0.022993717342615128, + 0.008002839982509613, + 0.009392306208610535, + 0.015670280903577805, + 0.022175494581460953, + 0.010737147182226181, + 0.013470973819494247, + 0.007327903062105179, + 0.005277973599731922, + 0.0043658362701535225, + 0.005234478507190943, + 0.008210614323616028, + 0.018333368003368378, + 0.008977226912975311, + 0.008965962566435337, + 0.003726365277543664, + 0.002808736404404044, + 0.0023573555517941713, + 0.002730004722252488, + 0.0051410384476184845, + 0.002831477904692292, + 0.0024964313488453627, + 0.0032570799812674522, + 0.019408036023378372, + 0.008030521683394909, + 0.003302821423858404, + 0.006655509117990732, + 0.019077951088547707, + 0.06941939890384674, + 0.04369663819670677, + 0.01547718420624733, + 0.004872613586485386, + 0.0027899695560336113, + 0.0026532644405961037, + 0.004166814498603344, + 0.006631691474467516, + 0.00533425435423851, + 0.0038704967591911554, + 0.0033012889325618744, + 0.00407032435759902, + 0.003609426086768508, + 0.0035471871960908175, + 0.003355218330398202, + 0.004156508482992649, + 0.005092914681881666, + 0.0036356037016958, + 0.0029461397789418697, + 0.009387166239321232, + 0.006795112509280443, + 0.006127949338406324, + 0.008129140362143517, + 0.0070677269250154495, + 0.019836202263832092, + 0.01308029517531395, + 0.007768593728542328, + 0.017713505774736404, + 0.020008038729429245, + 0.013117351569235325, + 0.01221479196101427, + 0.020984549075365067, + 0.016168372705578804, + 0.013989799655973911, + 0.009492874145507812, + 0.02859817072749138, + 0.030997512862086296, + 0.01311526820063591, + 0.014509334228932858, + 0.02075461857020855, + 0.04800499975681305, + 0.021366456523537636, + 0.025585614144802094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.006737373303622007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_tensor([5521, 5521])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_tensor([5521, 5521])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..29ce0bedb643e06a0ffc5d11bb0fd9c165fbc0f5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_tensor([5521, 5521])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a slightly upward angle and handheld, creates a constant, subtle tremor as the environment oscillates, creating a powerful sense of presence. Under incredibly realistic path-traced lighting, a resolute old man stands tall and proud in the center of a dense, impenetrable bamboo forest. His graying hair and beard flutter in the wind, his face deeply lined with wrinkles as if carved by a knife. His deep eyes gaze piercingly ahead, revealing an unyielding resolve. He wears a dark gray, rough linen robe, its sleeves rustling in the wind, yet his feet remain firmly planted on the moist earth, carpeted with withered, yellowing leaves. Around him, countless massive, sturdy, emerald green bamboos bend, sway, and collide violently like the surging waves of the sea, their leaves emitting a sharp whistling sound. Light struggles to penetrate through the dense bamboo foliage, forming countless sharp, dynamically shifting beams, casting a variegated, ever-changing light and shadow on the elderly man and the swaying ground. The scene, dominated by deep greens and grays, features intense contrasts of light and dark, and the constant tremors of the painting, creating a tense atmosphere of confrontation between man and nature, imbued with a sense of primal power.", + "seed": 27591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_5521_seed_27591_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08910999745130539, + "vlm_logprob": -0.08910999745130539, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.08910999745130539, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03475872799754143 + ], + "opposite_logprob_tokens": [ + -3.4097588062286377 + ], + "logprob_reward": -0.03475872799754143, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03475872799754143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14223399758338928, + 0.1352851688861847, + 0.046249356120824814, + 0.03415665403008461, + 0.13339726626873016, + 0.05422889068722725, + 0.3888607323169708, + 0.020217392593622208, + 0.016358422115445137, + 0.029012158513069153 + ], + "credit_map": [ + 0.013180015608668327, + 0.009495998732745647, + 0.01869898848235607, + 0.020343579351902008, + 0.014846921898424625, + 0.02003861963748932, + 0.025966765359044075, + 0.01966310665011406, + 0.010133231990039349, + 0.015161566436290741, + 0.010460267774760723, + 0.017931008711457253, + 0.022004272788763046, + 0.013731361366808414, + 0.026702582836151123, + 0.019160885363817215, + 0.004010758362710476, + 0.008822930976748466, + 0.004093200899660587, + 0.00582448486238718, + 0.006590839941054583, + 0.003718082793056965, + 0.005870840046554804, + 0.007318219635635614, + 0.0021041017025709152, + 0.0043544708751142025, + 0.004639750346541405, + 0.006565207149833441, + 0.002463726093992591, + 0.002885077614337206, + 0.0058306739665567875, + 0.0053136455826461315, + 0.010030421428382397, + 0.0255482979118824, + 0.008477768860757351, + 0.020327376201748848, + 0.012425298802554607, + 0.014172622002661228, + 0.019991617649793625, + 0.022423865273594856, + 0.002514729741960764, + 0.008485876023769379, + 0.005503891967236996, + 0.009434063918888569, + 0.00526826735585928, + 0.006421410944312811, + 0.006197074893862009, + 0.010403572581708431, + 0.01702999509871006, + 0.012271632440388203, + 0.007595837116241455, + 0.1516772359609604, + 0.01750027947127819, + 0.00552948284894228, + 0.15147507190704346, + 0.025781208649277687, + 0.0012855530949309468, + 0.0032627235632389784, + 0.0012940916931256652, + 0.0034081265330314636, + 0.003695884719491005, + 0.0016505110543221235, + 0.0031524342484772205, + 0.0024680690839886665, + 0.0010629635071381927, + 0.0030957928393036127, + 0.0011868044966831803, + 0.0028917379677295685, + 0.003208481939509511, + 0.0013843461638316512, + 0.0018743323162198067, + 0.001653962885029614, + 0.002034546574577689, + 0.003434186801314354, + 0.002063509076833725, + 0.004750615451484919, + 0.005269977729767561, + 0.003179998369887471, + 0.0053358497098088264, + 0.002943475032225251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.03475872799754143 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033716656267642975 + ], + "opposite_logprob_tokens": [ + -3.4087166786193848 + ], + "logprob_reward": -0.033716656267642975, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033716656267642975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09056847542524338, + 0.057695887982845306, + 0.015206404030323029, + 0.01834246702492237, + 0.015915030613541603, + 0.026180902495980263, + 0.7388679385185242, + 0.012051217257976532, + 0.012878728099167347, + 0.012292885221540928 + ], + "credit_map": [ + 0.008785348385572433, + 0.006277418229728937, + 0.013827748596668243, + 0.013567156158387661, + 0.006885874085128307, + 0.0122871994972229, + 0.016223588958382607, + 0.012714153155684471, + 0.00342037552036345, + 0.007322617340832949, + 0.004374490585178137, + 0.007518671918660402, + 0.009714466519653797, + 0.006941911298781633, + 0.010003641247749329, + 0.008399720303714275, + 0.001256718416698277, + 0.0022563948296010494, + 0.0012306385906413198, + 0.0025207246653735638, + 0.0018280621152371168, + 0.001475221011787653, + 0.002183993812650442, + 0.002454651752486825, + 0.0012573887361213565, + 0.0023928454611450434, + 0.0029380435589700937, + 0.0028525979723781347, + 0.0017984113655984402, + 0.0025634521152824163, + 0.0019490027334541082, + 0.0025907251983880997, + 0.0011900695972144604, + 0.002829477656632662, + 0.0010152902686968446, + 0.002341420156881213, + 0.001349155092611909, + 0.0019412657711654902, + 0.0029848450794816017, + 0.002263508504256606, + 0.0014798840275034308, + 0.002293205354362726, + 0.003528109984472394, + 0.0043712155893445015, + 0.0023256940767169, + 0.003850096371024847, + 0.0038654268719255924, + 0.004467272665351629, + 0.03210360184311867, + 0.0232721958309412, + 0.014339040964841843, + 0.2877453863620758, + 0.033174946904182434, + 0.010295260697603226, + 0.2889642119407654, + 0.04897336661815643, + 0.00066784379305318, + 0.0017118144314736128, + 0.0006993357092142105, + 0.0028955868910998106, + 0.0012936940183863044, + 0.0012585505610331893, + 0.0018989453092217445, + 0.0016254482325166464, + 0.0007244774606078863, + 0.0019563857931643724, + 0.0009127648081630468, + 0.00305401929654181, + 0.0020097303204238415, + 0.00095453328685835, + 0.0018102002795785666, + 0.0014566178433597088, + 0.0010619295062497258, + 0.0014373785816133022, + 0.001166415517218411, + 0.0018767311703413725, + 0.0017605824396014214, + 0.0018240876961499453, + 0.0018476180266588926, + 0.0013181432150304317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.033716656267642975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000403106212616 + ], + "opposite_logprob_tokens": [ + -2.6950039863586426 + ], + "logprob_reward": -0.07000403106212616, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000403106212616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040691107511520386, + 0.03215949609875679, + 0.01667196862399578, + 0.011121312156319618, + 0.01468505710363388, + 0.03697757422924042, + 0.8011212944984436, + 0.023309586569666862, + 0.012294957414269447, + 0.010967681184411049 + ], + "credit_map": [ + 0.004360157530754805, + 0.0035619905684143305, + 0.005084292497485876, + 0.004996455740183592, + 0.003492161398753524, + 0.006272953003644943, + 0.006831408943980932, + 0.006091690622270107, + 0.0019938477780669928, + 0.0036940798163414, + 0.0026566041633486748, + 0.0040403129532933235, + 0.0051989564672112465, + 0.0033376901410520077, + 0.00669793551787734, + 0.004540068097412586, + 0.0016473432769998908, + 0.0020686464849859476, + 0.0016127877170220017, + 0.0020640019793063402, + 0.0012195034651085734, + 0.002340290229767561, + 0.002870238618925214, + 0.0028491581324487925, + 0.0006026297924108803, + 0.00168404052965343, + 0.0007932194275781512, + 0.0018161882180720568, + 0.0011122421128675342, + 0.0014996855752542615, + 0.001782862120307982, + 0.0018304458353668451, + 0.0010118887294083834, + 0.003081294009461999, + 0.0015511162346228957, + 0.0019459383329376578, + 0.00141387979965657, + 0.0018587651429697871, + 0.0015514475526288152, + 0.002270729513838887, + 0.002150312066078186, + 0.002980007790029049, + 0.005005340091884136, + 0.006525724194943905, + 0.003090432146564126, + 0.005314242094755173, + 0.00561529491096735, + 0.006296223495155573, + 0.03564835339784622, + 0.025531914085149765, + 0.01568157970905304, + 0.3139026165008545, + 0.03586193919181824, + 0.011519384570419788, + 0.30922719836235046, + 0.05374833196401596, + 0.0011309151304885745, + 0.0026261918246746063, + 0.0012198531767353415, + 0.007263936568051577, + 0.0027810002211481333, + 0.0018604840151965618, + 0.003514078678563237, + 0.0029131299816071987, + 0.0006574240978807211, + 0.0017778079491108656, + 0.0008075484656728804, + 0.0030894973315298557, + 0.0020365433301776648, + 0.0009078193688765168, + 0.0016528264386579394, + 0.0013654923532158136, + 0.0010947557166218758, + 0.0014078106032684445, + 0.0010521693620830774, + 0.0014101981651037931, + 0.0013820391613990068, + 0.0018681039800867438, + 0.00137545564211905, + 0.00137714936863631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.07000403106212616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965127259492874 + ], + "opposite_logprob_tokens": [ + -2.929965019226074 + ], + "logprob_reward": -0.054965127259492874, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965127259492874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18182390928268433, + 0.07484295964241028, + 0.04473719373345375, + 0.0330192893743515, + 0.050947874784469604, + 0.04324983060359955, + 0.49297672510147095, + 0.0327836275100708, + 0.014976833015680313, + 0.030641810968518257 + ], + "credit_map": [ + 0.035222865641117096, + 0.01238185539841652, + 0.024511409923434258, + 0.021680021658539772, + 0.009631970897316933, + 0.02934134379029274, + 0.026276471093297005, + 0.022777961567044258, + 0.003992516081780195, + 0.01068189088255167, + 0.007549677975475788, + 0.009830256924033165, + 0.013789110817015171, + 0.008619657717645168, + 0.010158475488424301, + 0.010221370495855808, + 0.0044434261508286, + 0.006562444847077131, + 0.0048577021807432175, + 0.005124653689563274, + 0.004698087926954031, + 0.005029779858887196, + 0.006825522985309362, + 0.007195573300123215, + 0.003285733750090003, + 0.003967935685068369, + 0.003303894540295005, + 0.005119402892887592, + 0.002609718358144164, + 0.005083177704364061, + 0.004666728898882866, + 0.004982695449143648, + 0.0034040017053484917, + 0.008612719364464283, + 0.00505378982052207, + 0.0074684289284050465, + 0.0048551857471466064, + 0.007521577645093203, + 0.007075154688209295, + 0.006957019213587046, + 0.0022919594775885344, + 0.004637519363313913, + 0.004567165859043598, + 0.007661431096494198, + 0.0036908865440636873, + 0.004773405380547047, + 0.00644620880484581, + 0.009181254543364048, + 0.021507348865270615, + 0.01609569601714611, + 0.010128829628229141, + 0.1911906599998474, + 0.02229337953031063, + 0.007989509962499142, + 0.19102135300636292, + 0.03274991735816002, + 0.002034352393820882, + 0.006469773128628731, + 0.0022667970042675734, + 0.004767230711877346, + 0.006289822049438953, + 0.003112750593572855, + 0.004315173719078302, + 0.0035277276765555143, + 0.0009410015773028135, + 0.00210764491930604, + 0.0014235085109248757, + 0.002849829848855734, + 0.002533579245209694, + 0.0015124274650588632, + 0.002094422234222293, + 0.0015144201461225748, + 0.0021770396269857883, + 0.003958643414080143, + 0.0024466076865792274, + 0.006676563993096352, + 0.004029655829071999, + 0.0033489635679870844, + 0.005446579772979021, + 0.0025577566120773554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.054965127259492874 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521054446697235 + ], + "opposite_logprob_tokens": [ + -1.502105474472046 + ], + "logprob_reward": -0.2521054446697235, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521054446697235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02268916182219982, + 0.011126554571092129, + 0.012686525471508503, + 0.010049284435808659, + 0.010629219934344292, + 0.03081609681248665, + 0.8604283332824707, + 0.02752780169248581, + 0.0099659888073802, + 0.0040811155922710896 + ], + "credit_map": [ + 0.002905074506998062, + 0.002116321586072445, + 0.002222904236987233, + 0.0029687893111258745, + 0.0021021671127527952, + 0.0036256101448088884, + 0.00333055155351758, + 0.003417740808799863, + 0.0008609858341515064, + 0.0011095247464254498, + 0.0010482523357495666, + 0.001477995072491467, + 0.001474378746934235, + 0.0010362514294683933, + 0.0024523816537111998, + 0.0016667828895151615, + 0.0010878430912271142, + 0.001818108488805592, + 0.001137906569056213, + 0.0016889761900529265, + 0.0013967712875455618, + 0.0014779696939513087, + 0.0022457384038716555, + 0.0018332108156755567, + 0.0006045992486178875, + 0.0010782142635434866, + 0.0007106591947376728, + 0.0023494933266192675, + 0.0005496739177033305, + 0.0006663950043730438, + 0.001655028434470296, + 0.0024352199397981167, + 0.0009324945276603103, + 0.0016077449545264244, + 0.0009860459249466658, + 0.0014340095221996307, + 0.000984958023764193, + 0.0010257454123347998, + 0.0020650303922593594, + 0.00159319001249969, + 0.0018374515930190682, + 0.0027143056504428387, + 0.004146881867200136, + 0.005150449927896261, + 0.002817451721057296, + 0.0043670134618878365, + 0.005053869914263487, + 0.004728671163320541, + 0.037834566086530685, + 0.02718643844127655, + 0.016795499250292778, + 0.3357155919075012, + 0.03881337493658066, + 0.012129279784858227, + 0.33486536145210266, + 0.05708814412355423, + 0.0012276527704671025, + 0.0031161014921963215, + 0.0012996767181903124, + 0.008384053595364094, + 0.003927603363990784, + 0.0020146381575614214, + 0.004177473485469818, + 0.0033805989660322666, + 0.0005229654489085078, + 0.0014640740118920803, + 0.0005883790436200798, + 0.0026729765813797712, + 0.0015380778349936008, + 0.0006588213727809489, + 0.0013551884330809116, + 0.0011655051494017243, + 0.00028210878372192383, + 0.0005618734285235405, + 0.00032477249624207616, + 0.0008608130738139153, + 0.0006930716917850077, + 0.00040429120417684317, + 0.00045423934352584183, + 0.0004999447846785188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.2521054446697235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle and handheld, creates a constant, subtle tremor as the environment oscillates, creating a powerful sense of presence. Under incredibly realistic path-traced lighting, a resolute old man stands tall and proud in the center of a dense, impenetrable bamboo forest. His graying hair and beard flutter in the wind, his face deeply lined with wrinkles as if carved by a knife. His deep eyes gaze piercingly ahead, revealing an unyielding resolve. He wears a dark gray, rough linen robe, its sleeves rustling in the wind, yet his feet remain firmly planted on the moist earth, carpeted with withered, yellowing leaves. Around him, countless massive, sturdy, emerald green bamboos bend, sway, and collide violently like the surging waves of the sea, their leaves emitting a sharp whistling sound. Light struggles to penetrate through the dense bamboo foliage, forming countless sharp, dynamically shifting beams, casting a variegated, ever-changing light and shadow on the elderly man and the swaying ground. The scene, dominated by deep greens and grays, features intense contrasts of light and dark, and the constant tremors of the painting, creating a tense atmosphere of confrontation between man and nature, imbued with a sense of primal power.", + "seed": 27592, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_5521_seed_27592_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18872929960489274, + "vlm_logprob": -0.18872929960489274, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.18872929960489274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05600063130259514 + ], + "opposite_logprob_tokens": [ + -2.9310007095336914 + ], + "logprob_reward": -0.05600063130259514, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05600063130259514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1883765012025833, + 0.04805345833301544, + 0.0616915225982666, + 0.07386800646781921, + 0.28797265887260437, + 0.03799824416637421, + 0.09132944047451019, + 0.06764337420463562, + 0.03942170739173889, + 0.10364507883787155 + ], + "credit_map": [ + 0.027157142758369446, + 0.015328316017985344, + 0.02278916724026203, + 0.020089898258447647, + 0.017281997948884964, + 0.025342348963022232, + 0.03009309619665146, + 0.03029453381896019, + 0.003894423134624958, + 0.0031315558589994907, + 0.0031476847361773252, + 0.0033638072200119495, + 0.0048615713603794575, + 0.0033687306568026543, + 0.012736030854284763, + 0.013549650087952614, + 0.00825726892799139, + 0.0059337676502764225, + 0.00567613635212183, + 0.007651790976524353, + 0.007899959571659565, + 0.007500495761632919, + 0.008677860721945763, + 0.010094238445162773, + 0.009132222272455692, + 0.008709018118679523, + 0.006257389206439257, + 0.008398851379752159, + 0.012189784087240696, + 0.00996096059679985, + 0.010090052150189877, + 0.009129727259278297, + 0.03118722513318062, + 0.014450021088123322, + 0.015610016882419586, + 0.03226858004927635, + 0.01739807240664959, + 0.029607342556118965, + 0.06961402297019958, + 0.07783737033605576, + 0.0040321359410882, + 0.0032534643542021513, + 0.003391705919057131, + 0.0050359126180410385, + 0.0036991522647440434, + 0.005714917089790106, + 0.006052963901311159, + 0.006817990448325872, + 0.00932718813419342, + 0.014084530994296074, + 0.010903585702180862, + 0.01201891154050827, + 0.010863229632377625, + 0.008805668912827969, + 0.013191591016948223, + 0.012134729884564877, + 0.005695462692528963, + 0.010513034649193287, + 0.006022930145263672, + 0.009649674408137798, + 0.005231236107647419, + 0.00565717788413167, + 0.011108345352113247, + 0.013765504583716393, + 0.004884510766714811, + 0.006166305858641863, + 0.0035500461235642433, + 0.00534390052780509, + 0.0040469602681696415, + 0.0038786036893725395, + 0.005846399813890457, + 0.005704971961677074, + 0.011849885806441307, + 0.01794266328215599, + 0.006411145441234112, + 0.015100948512554169, + 0.011839700862765312, + 0.01089651882648468, + 0.015279828570783138, + 0.01432438101619482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.05600063130259514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496704578399658 + ], + "opposite_logprob_tokens": [ + -2.929966926574707 + ], + "logprob_reward": -0.05496704578399658, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496704578399658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.230293408036232, + 0.06132743880152702, + 0.08013477176427841, + 0.1409004032611847, + 0.08279062062501907, + 0.08862201869487762, + 0.07204151898622513, + 0.09321220219135284, + 0.09257546067237854, + 0.05810218304395676 + ], + "credit_map": [ + 0.02321028523147106, + 0.027806540951132774, + 0.02054757811129093, + 0.030104661360383034, + 0.03686908632516861, + 0.022507034242153168, + 0.03785504028201103, + 0.03139318898320198, + 0.005637376103550196, + 0.004419432021677494, + 0.004599791020154953, + 0.006527849473059177, + 0.004730932880192995, + 0.006117898505181074, + 0.014717510901391506, + 0.014576644636690617, + 0.008714701980352402, + 0.006365783978253603, + 0.006265208590775728, + 0.014012331143021584, + 0.007594941183924675, + 0.008046860806643963, + 0.015255914069712162, + 0.013879029080271721, + 0.017083410173654556, + 0.010001321323215961, + 0.008869034238159657, + 0.010704870335757732, + 0.013474239967763424, + 0.017990030348300934, + 0.03380277752876282, + 0.028974713757634163, + 0.005797744728624821, + 0.0069624800235033035, + 0.007741699460893869, + 0.01120135746896267, + 0.011131953448057175, + 0.007459076587110758, + 0.015525776892900467, + 0.016970541328191757, + 0.008613958954811096, + 0.0067871385253965855, + 0.007958848029375076, + 0.009451184421777725, + 0.007366917096078396, + 0.016048338264226913, + 0.013540089130401611, + 0.018855547532439232, + 0.008861935697495937, + 0.009719948284327984, + 0.009590987116098404, + 0.008559854701161385, + 0.008323059417307377, + 0.010734841227531433, + 0.00799315981566906, + 0.008257738314568996, + 0.010121201165020466, + 0.01255075540393591, + 0.007656053174287081, + 0.015261414460837841, + 0.010965385474264622, + 0.008843853138387203, + 0.015658292919397354, + 0.012155245058238506, + 0.01210328284651041, + 0.011098151095211506, + 0.006580393295735121, + 0.00852627120912075, + 0.0065905265510082245, + 0.00963932927697897, + 0.015645500272512436, + 0.022391999140381813, + 0.004954679869115353, + 0.004871482029557228, + 0.0032155730295926332, + 0.005832243710756302, + 0.004239760804921389, + 0.006794323679059744, + 0.011133641935884953, + 0.017060477286577225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.05496704578399658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153652131557465 + ], + "opposite_logprob_tokens": [ + -1.7015365362167358 + ], + "logprob_reward": -0.20153652131557465, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153652131557465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16250677406787872, + 0.05909183248877525, + 0.07087651640176773, + 0.08850771933794022, + 0.12182432413101196, + 0.13721826672554016, + 0.12112388014793396, + 0.09097447246313095, + 0.05643090233206749, + 0.09144534170627594 + ], + "credit_map": [ + 0.020690040662884712, + 0.020888544619083405, + 0.020466094836592674, + 0.018322380259633064, + 0.017344849184155464, + 0.019082220271229744, + 0.020353516563773155, + 0.02535911835730076, + 0.007574849762022495, + 0.005356039851903915, + 0.005544286221265793, + 0.007504444103688002, + 0.005702866706997156, + 0.008590040728449821, + 0.009747280739247799, + 0.009072023443877697, + 0.009233001619577408, + 0.007725242525339127, + 0.00725915003567934, + 0.01012760866433382, + 0.007615251466631889, + 0.007842088118195534, + 0.009793750941753387, + 0.011280418373644352, + 0.012390096671879292, + 0.008480072021484375, + 0.0077543193474411964, + 0.006765140686184168, + 0.006764947436749935, + 0.013209427706897259, + 0.018163656815886497, + 0.014980065636336803, + 0.013756437227129936, + 0.007395163644105196, + 0.008548763580620289, + 0.01463143341243267, + 0.009850422851741314, + 0.012243077158927917, + 0.026221374049782753, + 0.029177658259868622, + 0.012464507482945919, + 0.01123859453946352, + 0.011511399410665035, + 0.014803783036768436, + 0.011735767126083374, + 0.02352031320333481, + 0.021859604865312576, + 0.03008430078625679, + 0.0170399509370327, + 0.012993385083973408, + 0.01566239446401596, + 0.012749656103551388, + 0.013658138923346996, + 0.02178214304149151, + 0.012489888817071915, + 0.014748325571417809, + 0.01450190506875515, + 0.006823529023677111, + 0.015482313930988312, + 0.009925251826643944, + 0.004558946937322617, + 0.016042249277234077, + 0.010724480263888836, + 0.012915793806314468, + 0.008397573605179787, + 0.008081629872322083, + 0.005379781126976013, + 0.006659433711320162, + 0.005980067420750856, + 0.007931515574455261, + 0.0061930580995976925, + 0.007807846646755934, + 0.010936630889773369, + 0.01253155805170536, + 0.006489653140306473, + 0.011043787933886051, + 0.007452387362718582, + 0.014105130918323994, + 0.012635167688131332, + 0.016251027584075928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.20153652131557465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495576187968254 + ], + "opposite_logprob_tokens": [ + -2.9299557209014893 + ], + "logprob_reward": -0.05495576187968254, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495576187968254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26021552085876465, + 0.1818968802690506, + 0.0537094920873642, + 0.0846644714474678, + 0.06850439310073853, + 0.057839393615722656, + 0.07784755527973175, + 0.08315775543451309, + 0.0795293003320694, + 0.05263524129986763 + ], + "credit_map": [ + 0.03808756545186043, + 0.02556893788278103, + 0.0331495963037014, + 0.029556043446063995, + 0.027014175429940224, + 0.03613267093896866, + 0.03794687241315842, + 0.03275967016816139, + 0.014832543209195137, + 0.009833068586885929, + 0.012015331536531448, + 0.013054734095931053, + 0.0067986599169671535, + 0.013191737234592438, + 0.05366814136505127, + 0.05850265920162201, + 0.00681546563282609, + 0.0075499736703932285, + 0.005770104005932808, + 0.007887322455644608, + 0.0038689475040882826, + 0.0059670149348676205, + 0.008640183135867119, + 0.007210482377558947, + 0.010908251628279686, + 0.008661344647407532, + 0.005843952763825655, + 0.007599066477268934, + 0.010968095622956753, + 0.010209859348833561, + 0.018147466704249382, + 0.012326432392001152, + 0.007381732575595379, + 0.007536794524639845, + 0.008943086490035057, + 0.010725262574851513, + 0.008074990473687649, + 0.008046237751841545, + 0.009840898215770721, + 0.007955389097332954, + 0.006581360008567572, + 0.006596661638468504, + 0.0062610432505607605, + 0.005293103400617838, + 0.006361109670251608, + 0.010778747498989105, + 0.007555077318102121, + 0.008412289433181286, + 0.01245530042797327, + 0.009750244207680225, + 0.007766357157379389, + 0.008461636491119862, + 0.008322170004248619, + 0.015196247026324272, + 0.008123829029500484, + 0.007771771401166916, + 0.012754996307194233, + 0.006463803816586733, + 0.007786306086927652, + 0.008095075376331806, + 0.005199308041483164, + 0.014604579657316208, + 0.011947317980229855, + 0.01630636490881443, + 0.009522702544927597, + 0.007463315036147833, + 0.0054901777766644955, + 0.0048055280931293964, + 0.00540595268830657, + 0.010082877241075039, + 0.014184622094035149, + 0.02257412299513817, + 0.004786944016814232, + 0.003820509882643819, + 0.0030064049642533064, + 0.006729962769895792, + 0.0042184689082205296, + 0.008134248666465282, + 0.009661623276770115, + 0.012277079746127129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.05495576187968254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761865377426147 + ], + "opposite_logprob_tokens": [ + -0.8261865377426147 + ], + "logprob_reward": -0.5761865377426147, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761865377426147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22633172571659088, + 0.07147763669490814, + 0.04831195995211601, + 0.14478501677513123, + 0.2003883719444275, + 0.059580858796834946, + 0.06759575009346008, + 0.043748755007982254, + 0.08311966061592102, + 0.05466023087501526 + ], + "credit_map": [ + 0.02023395523428917, + 0.024129310622811317, + 0.026782531291246414, + 0.02645057998597622, + 0.023268161341547966, + 0.02185644581913948, + 0.042229052633047104, + 0.04138168320059776, + 0.00938537809997797, + 0.010793093591928482, + 0.007598768919706345, + 0.009041126817464828, + 0.006838630884885788, + 0.007100867107510567, + 0.01080522034317255, + 0.009914549067616463, + 0.005515031050890684, + 0.004463969264179468, + 0.003856983734294772, + 0.009080188348889351, + 0.004751795902848244, + 0.00552018778398633, + 0.00765032134950161, + 0.007473484147340059, + 0.01862478069961071, + 0.013376760296523571, + 0.010223720222711563, + 0.01386566087603569, + 0.01830815151333809, + 0.022262830287218094, + 0.024632003158330917, + 0.023491110652685165, + 0.022708643227815628, + 0.009588619694113731, + 0.011258416809141636, + 0.02140510454773903, + 0.01163044199347496, + 0.019440755248069763, + 0.04961515590548515, + 0.05474124476313591, + 0.007855067029595375, + 0.005520204082131386, + 0.005508802365511656, + 0.006264161318540573, + 0.006401504389941692, + 0.00951247476041317, + 0.007753564976155758, + 0.010765081271529198, + 0.010684631764888763, + 0.007038517855107784, + 0.008201293647289276, + 0.008819812908768654, + 0.0063368999399244785, + 0.011994182132184505, + 0.006562328431755304, + 0.00795808993279934, + 0.004168514162302017, + 0.005816083401441574, + 0.0031172398012131453, + 0.009509582072496414, + 0.004062512889504433, + 0.0048186094500124454, + 0.006075873505324125, + 0.006180338095873594, + 0.010102215223014355, + 0.008027651347219944, + 0.005850169342011213, + 0.006977819837629795, + 0.004504552111029625, + 0.009587856009602547, + 0.015582378022372723, + 0.02248702198266983, + 0.00608227401971817, + 0.009764301590621471, + 0.0033408273011446, + 0.0078493757173419, + 0.005311279091984034, + 0.006675702054053545, + 0.008320583030581474, + 0.007315885275602341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.5761865377426147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle and handheld, creates a constant, subtle tremor as the environment oscillates, creating a powerful sense of presence. Under incredibly realistic path-traced lighting, a resolute old man stands tall and proud in the center of a dense, impenetrable bamboo forest. His graying hair and beard flutter in the wind, his face deeply lined with wrinkles as if carved by a knife. His deep eyes gaze piercingly ahead, revealing an unyielding resolve. He wears a dark gray, rough linen robe, its sleeves rustling in the wind, yet his feet remain firmly planted on the moist earth, carpeted with withered, yellowing leaves. Around him, countless massive, sturdy, emerald green bamboos bend, sway, and collide violently like the surging waves of the sea, their leaves emitting a sharp whistling sound. Light struggles to penetrate through the dense bamboo foliage, forming countless sharp, dynamically shifting beams, casting a variegated, ever-changing light and shadow on the elderly man and the swaying ground. The scene, dominated by deep greens and grays, features intense contrasts of light and dark, and the constant tremors of the painting, creating a tense atmosphere of confrontation between man and nature, imbued with a sense of primal power.", + "seed": 27593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_5521_seed_27593_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10233375914394856, + "vlm_logprob": -0.10233375914394856, + "vlm_yes_no_margin": 2.974999988079071, + "vlm_token_credit": -0.10233375914394856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021643631160259247 + ], + "opposite_logprob_tokens": [ + -3.89664363861084 + ], + "logprob_reward": -0.021643631160259247, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021643631160259247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27737393975257874, + 0.3081291913986206, + 0.10562767088413239, + 0.04213123768568039, + 0.06142110377550125, + 0.039533741772174835, + 0.02562495693564415, + 0.04301925003528595, + 0.04207457974553108, + 0.0550643615424633 + ], + "credit_map": [ + 0.015221977606415749, + 0.02037365734577179, + 0.035252735018730164, + 0.03828595578670502, + 0.05846790596842766, + 0.030734358355402946, + 0.042758774012327194, + 0.03627857193350792, + 0.04357394203543663, + 0.013810972683131695, + 0.03555557504296303, + 0.03446486219763756, + 0.019923685118556023, + 0.06932877749204636, + 0.05217054486274719, + 0.0393008217215538, + 0.009238328784704208, + 0.008545458316802979, + 0.0069237276911735535, + 0.016790442168712616, + 0.024893490597605705, + 0.014491683803498745, + 0.01322057656943798, + 0.011523962020874023, + 0.004882401321083307, + 0.004377422854304314, + 0.007755822502076626, + 0.006397646386176348, + 0.004900569561868906, + 0.004015420563519001, + 0.004235564265400171, + 0.005566386505961418, + 0.004511130973696709, + 0.0049352627247571945, + 0.012165981344878674, + 0.006075734738260508, + 0.007304273545742035, + 0.009767532348632812, + 0.00936624314635992, + 0.007294944953173399, + 0.0040764291770756245, + 0.0037042407784610987, + 0.006373114883899689, + 0.005179659463465214, + 0.0033254516310989857, + 0.008650433272123337, + 0.004009951837360859, + 0.004214458633214235, + 0.003098585642874241, + 0.002811928978189826, + 0.003072221763432026, + 0.0041947937570512295, + 0.003108005505055189, + 0.003209629561752081, + 0.003929798025637865, + 0.0021999927703291178, + 0.009990484453737736, + 0.004239856731146574, + 0.0036941785365343094, + 0.003989205230027437, + 0.004632230848073959, + 0.006050129886716604, + 0.007395547349005938, + 0.003027614438906312, + 0.005029512569308281, + 0.005518433637917042, + 0.005986828822642565, + 0.006769445259124041, + 0.0032420288771390915, + 0.007614754140377045, + 0.004507684614509344, + 0.0034058885648846626, + 0.007970254868268967, + 0.005730267148464918, + 0.007209285162389278, + 0.00716880289837718, + 0.004139990545809269, + 0.007881736382842064, + 0.007298785261809826, + 0.007665238343179226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.021643631160259247 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059783972799778 + ], + "opposite_logprob_tokens": [ + -3.8955979347229004 + ], + "logprob_reward": -0.02059783972799778, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059783972799778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2274436205625534, + 0.16027456521987915, + 0.09976880252361298, + 0.09218709915876389, + 0.08404289186000824, + 0.06451600044965744, + 0.021478433161973953, + 0.08170300722122192, + 0.09085969626903534, + 0.0777258574962616 + ], + "credit_map": [ + 0.017641417682170868, + 0.021729040890932083, + 0.03210998326539993, + 0.034075718373060226, + 0.028949232771992683, + 0.02716691419482231, + 0.032736144959926605, + 0.033035166561603546, + 0.013412214815616608, + 0.01424518134444952, + 0.014263810589909554, + 0.013369070366024971, + 0.009744318202137947, + 0.05366813763976097, + 0.01774255558848381, + 0.023829272016882896, + 0.012565979734063148, + 0.012004655785858631, + 0.008087283931672573, + 0.012145832180976868, + 0.015743885189294815, + 0.013954613357782364, + 0.013059630058705807, + 0.012206923216581345, + 0.011851047165691853, + 0.011678609065711498, + 0.01618099771440029, + 0.015040404163300991, + 0.006727789528667927, + 0.010711563751101494, + 0.01068542618304491, + 0.009311260655522346, + 0.006077265366911888, + 0.007769403047859669, + 0.020229307934641838, + 0.006532348692417145, + 0.006130623631179333, + 0.022118166089057922, + 0.009972011670470238, + 0.005213761702179909, + 0.0072468058206140995, + 0.005698165390640497, + 0.012845653109252453, + 0.008368182927370071, + 0.004019130486994982, + 0.012194073759019375, + 0.006745210848748684, + 0.00739878136664629, + 0.003454481950029731, + 0.0018140091560781002, + 0.0023724394850432873, + 0.0025971655268222094, + 0.002943319734185934, + 0.003164933994412422, + 0.002856707898899913, + 0.002275376580655575, + 0.014753700233995914, + 0.011121593415737152, + 0.009678645059466362, + 0.009010226465761662, + 0.006646404508501291, + 0.014659624546766281, + 0.009922320954501629, + 0.005910490173846483, + 0.01103003229945898, + 0.014978596940636635, + 0.010474313981831074, + 0.012158554047346115, + 0.0064473338425159454, + 0.009655117988586426, + 0.01397650595754385, + 0.012139239348471165, + 0.011765524744987488, + 0.008047946728765965, + 0.009711891412734985, + 0.012740031816065311, + 0.005789260845631361, + 0.011227371171116829, + 0.008712472394108772, + 0.009731363505125046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02059783972799778 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369014710187912 + ], + "opposite_logprob_tokens": [ + -3.4086902141571045 + ], + "logprob_reward": -0.03369014710187912, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369014710187912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12707093358039856, + 0.2695259749889374, + 0.07375144213438034, + 0.058064054697752, + 0.14660222828388214, + 0.07982753217220306, + 0.04433578625321388, + 0.07996592670679092, + 0.05244503542780876, + 0.06841105967760086 + ], + "credit_map": [ + 0.015173126943409443, + 0.01539800874888897, + 0.017792722210288048, + 0.014571445994079113, + 0.014644457027316093, + 0.018453646451234818, + 0.015946179628372192, + 0.015091353096067905, + 0.027151135727763176, + 0.01618952304124832, + 0.02029697597026825, + 0.028274638578295708, + 0.014458153396844864, + 0.0934649258852005, + 0.04373705014586449, + 0.02595359832048416, + 0.012776768766343594, + 0.010756967589259148, + 0.008572020567953587, + 0.00891822762787342, + 0.009709644131362438, + 0.009484706446528435, + 0.008257610723376274, + 0.005275497678667307, + 0.007778718136250973, + 0.007905567064881325, + 0.010538836941123009, + 0.008280789479613304, + 0.004138991702347994, + 0.005622950382530689, + 0.005946903955191374, + 0.00785130076110363, + 0.010114140808582306, + 0.009010215289890766, + 0.05410383269190788, + 0.008764104917645454, + 0.004518477711826563, + 0.03895988315343857, + 0.013718191534280777, + 0.007413397077471018, + 0.00821737851947546, + 0.005128615535795689, + 0.017586518079042435, + 0.006918573752045631, + 0.003025323385372758, + 0.02778233028948307, + 0.006558965891599655, + 0.004609833005815744, + 0.006524926982820034, + 0.002561718225479126, + 0.004195101093500853, + 0.004277391824871302, + 0.0031330292113125324, + 0.013450722210109234, + 0.005314825102686882, + 0.004878074862062931, + 0.011899005621671677, + 0.013121712021529675, + 0.01176441926509142, + 0.008593404665589333, + 0.006897042039781809, + 0.01123126596212387, + 0.01087560411542654, + 0.005583486519753933, + 0.006563814356923103, + 0.009068814106285572, + 0.006658237427473068, + 0.006942916661500931, + 0.003305630525574088, + 0.0056959958747029305, + 0.00789575558155775, + 0.006313877645879984, + 0.008548439480364323, + 0.006899673026055098, + 0.009900067001581192, + 0.01024689432233572, + 0.005650738719850779, + 0.01054881326854229, + 0.007466827053576708, + 0.009149612858891487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.03369014710187912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865477979183197 + ], + "opposite_logprob_tokens": [ + -3.048654794692993 + ], + "logprob_reward": -0.04865477979183197, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865477979183197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15040861070156097, + 0.1251491755247116, + 0.050385963171720505, + 0.044505443423986435, + 0.08516856282949448, + 0.061975229531526566, + 0.08373938500881195, + 0.21907223761081696, + 0.10856648534536362, + 0.07102896273136139 + ], + "credit_map": [ + 0.011344951577484608, + 0.014463658444583416, + 0.021243490278720856, + 0.02073133923113346, + 0.019711632281541824, + 0.019081834703683853, + 0.019490018486976624, + 0.024341672658920288, + 0.012399278581142426, + 0.009092532098293304, + 0.00911648478358984, + 0.012269998900592327, + 0.006834528874605894, + 0.04440780729055405, + 0.018409211188554764, + 0.012619316577911377, + 0.005208985880017281, + 0.004599007312208414, + 0.00449082488194108, + 0.009289532899856567, + 0.008382647298276424, + 0.006039996165782213, + 0.006122342776507139, + 0.006252622231841087, + 0.005780741106718779, + 0.006031101103872061, + 0.007058493793010712, + 0.005680037662386894, + 0.004540731199085712, + 0.004355469718575478, + 0.005423309747129679, + 0.005635558627545834, + 0.007099274080246687, + 0.01642342284321785, + 0.0137630021199584, + 0.009601474739611149, + 0.006245969794690609, + 0.011893564835190773, + 0.01128745824098587, + 0.008854388259351254, + 0.006322952453047037, + 0.00354921817779541, + 0.017480557784438133, + 0.0046648853458464146, + 0.003046222496777773, + 0.01781706139445305, + 0.004951858893036842, + 0.004142465069890022, + 0.010716094635426998, + 0.01072611566632986, + 0.007409502752125263, + 0.008943289518356323, + 0.004994464572519064, + 0.02289463020861149, + 0.009043330326676369, + 0.009011954069137573, + 0.07436036318540573, + 0.023552928119897842, + 0.021019507199525833, + 0.014996934682130814, + 0.012837509624660015, + 0.02951001562178135, + 0.022097963839769363, + 0.02069699577987194, + 0.014429071918129921, + 0.016031352803111076, + 0.014848916791379452, + 0.017153700813651085, + 0.006170238368213177, + 0.01824367046356201, + 0.012174021452665329, + 0.009515502490103245, + 0.009111461229622364, + 0.007284135092049837, + 0.011013149283826351, + 0.009060622192919254, + 0.004523064009845257, + 0.014349900186061859, + 0.007283635903149843, + 0.00840299017727375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.04865477979183197 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38708239793777466 + ], + "opposite_logprob_tokens": [ + -1.1370823383331299 + ], + "logprob_reward": -0.38708239793777466, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38708239793777466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14867860078811646, + 0.1151685044169426, + 0.05845791846513748, + 0.06544986367225647, + 0.1944003701210022, + 0.14479270577430725, + 0.06920472532510757, + 0.04973214492201805, + 0.07170629501342773, + 0.08240889012813568 + ], + "credit_map": [ + 0.009421505965292454, + 0.018317775800824165, + 0.02008691616356373, + 0.023275408893823624, + 0.01817142404615879, + 0.01791735738515854, + 0.021101079881191254, + 0.020387131720781326, + 0.019521377980709076, + 0.01280162576586008, + 0.021142099052667618, + 0.00841799657791853, + 0.008414240553975105, + 0.01666717790067196, + 0.01034941803663969, + 0.01785457134246826, + 0.00941530242562294, + 0.006166468374431133, + 0.005284165497869253, + 0.011420872062444687, + 0.008755206130445004, + 0.005992667283862829, + 0.006961887236684561, + 0.004461353644728661, + 0.008312143385410309, + 0.009904013015329838, + 0.009470286779105663, + 0.008981798775494099, + 0.007266572676599026, + 0.006477014627307653, + 0.0064908238127827644, + 0.008547203615307808, + 0.012945841997861862, + 0.00975601002573967, + 0.07012961059808731, + 0.012897119857370853, + 0.007916714064776897, + 0.05438803508877754, + 0.017935315147042274, + 0.008431728929281235, + 0.01420070044696331, + 0.013271217234432697, + 0.018822038546204567, + 0.020583532750606537, + 0.008161698468029499, + 0.03660948574542999, + 0.013870345428586006, + 0.01927369460463524, + 0.010441084392368793, + 0.0037502292543649673, + 0.004657842218875885, + 0.006013777572661638, + 0.004475467372685671, + 0.024876471608877182, + 0.007210932672023773, + 0.007778919767588377, + 0.01498645730316639, + 0.004408689681440592, + 0.005234007257968187, + 0.005682103335857391, + 0.0034301024861633778, + 0.005804280750453472, + 0.005704618990421295, + 0.004481884650886059, + 0.0068735117092728615, + 0.010573575273156166, + 0.008489986881613731, + 0.01583016850054264, + 0.00420710863545537, + 0.007762590888887644, + 0.011177750304341316, + 0.006791603751480579, + 0.01113293506205082, + 0.007095590699464083, + 0.010965054854750633, + 0.01680593006312847, + 0.006224189419299364, + 0.010179456323385239, + 0.011958597227931023, + 0.008047135546803474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.38708239793777466 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle and handheld, creates a constant, subtle tremor as the environment oscillates, creating a powerful sense of presence. Under incredibly realistic path-traced lighting, a resolute old man stands tall and proud in the center of a dense, impenetrable bamboo forest. His graying hair and beard flutter in the wind, his face deeply lined with wrinkles as if carved by a knife. His deep eyes gaze piercingly ahead, revealing an unyielding resolve. He wears a dark gray, rough linen robe, its sleeves rustling in the wind, yet his feet remain firmly planted on the moist earth, carpeted with withered, yellowing leaves. Around him, countless massive, sturdy, emerald green bamboos bend, sway, and collide violently like the surging waves of the sea, their leaves emitting a sharp whistling sound. Light struggles to penetrate through the dense bamboo foliage, forming countless sharp, dynamically shifting beams, casting a variegated, ever-changing light and shadow on the elderly man and the swaying ground. The scene, dominated by deep greens and grays, features intense contrasts of light and dark, and the constant tremors of the painting, creating a tense atmosphere of confrontation between man and nature, imbued with a sense of primal power.", + "seed": 27594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_10/index_5521_seed_27594_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1883312650024891, + "vlm_logprob": -0.1883312650024891, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.1883312650024891, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2533252239227295 + ], + "opposite_logprob_tokens": [ + -1.5033252239227295 + ], + "logprob_reward": -0.2533252239227295, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2533252239227295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19620592892169952, + 0.034466277807950974, + 0.05458461120724678, + 0.07821892201900482, + 0.2009020447731018, + 0.07672816514968872, + 0.03907820209860802, + 0.12016432732343674, + 0.10656198114156723, + 0.09308964014053345 + ], + "credit_map": [ + 0.007915973663330078, + 0.014587945304811, + 0.013895238749682903, + 0.014478062279522419, + 0.03204204887151718, + 0.048558566719293594, + 0.022445186972618103, + 0.04228288680315018, + 0.00363958440721035, + 0.0028678267262876034, + 0.0032378362957388163, + 0.0052368370816111565, + 0.003779510036110878, + 0.005648261867463589, + 0.005433748476207256, + 0.004622672218829393, + 0.002531129866838455, + 0.0073909456841647625, + 0.004747972823679447, + 0.008104030042886734, + 0.008273354731500149, + 0.006971075665205717, + 0.009426161646842957, + 0.007139931432902813, + 0.005033875349909067, + 0.008737952448427677, + 0.007433900143951178, + 0.011435314081609249, + 0.011496984399855137, + 0.016432246193289757, + 0.007138540968298912, + 0.010510101914405823, + 0.007565010339021683, + 0.040373481810092926, + 0.010908311232924461, + 0.040101319551467896, + 0.027410011738538742, + 0.02646510675549507, + 0.024183711037039757, + 0.023895077407360077, + 0.012412413023412228, + 0.005890730302780867, + 0.011790141463279724, + 0.011998310685157776, + 0.008694438263773918, + 0.010150554589927197, + 0.008074906654655933, + 0.007716666907072067, + 0.002320597879588604, + 0.004322509281337261, + 0.002530420199036598, + 0.004996397532522678, + 0.004933428484946489, + 0.00929997768253088, + 0.004549051634967327, + 0.006125813815742731, + 0.006808335427194834, + 0.010070455260574818, + 0.00882020965218544, + 0.01462581567466259, + 0.012330724857747555, + 0.03484952822327614, + 0.011190880089998245, + 0.02146836370229721, + 0.006475144997239113, + 0.014468830078840256, + 0.0104958675801754, + 0.013165101408958435, + 0.014109469018876553, + 0.015465826727449894, + 0.016546348109841347, + 0.015835391357541084, + 0.0054925596341490746, + 0.012960009276866913, + 0.004834845662117004, + 0.01609158329665661, + 0.010163074359297752, + 0.01211876142770052, + 0.016318146139383316, + 0.015110651962459087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.2533252239227295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054975058883428574 + ], + "opposite_logprob_tokens": [ + -2.9299750328063965 + ], + "logprob_reward": -0.054975058883428574, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054975058883428574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29329657554626465, + 0.08656686544418335, + 0.07573701441287994, + 0.059251029044389725, + 0.12495633214712143, + 0.05556212738156319, + 0.07085336744785309, + 0.09711425751447678, + 0.1031385064125061, + 0.03352394327521324 + ], + "credit_map": [ + 0.014162122271955013, + 0.023989757522940636, + 0.026387222111225128, + 0.029584957286715508, + 0.05436961352825165, + 0.05317210033535957, + 0.036881279200315475, + 0.05474948510527611, + 0.013777006417512894, + 0.007057508919388056, + 0.008502058684825897, + 0.00881963036954403, + 0.009496173821389675, + 0.014333347789943218, + 0.009998525492846966, + 0.014582611620426178, + 0.0058650067076087, + 0.007314811460673809, + 0.006764912512153387, + 0.012909844517707825, + 0.008928217925131321, + 0.011313216760754585, + 0.01319308765232563, + 0.009447906166315079, + 0.003373067593201995, + 0.0047956230118870735, + 0.005610832944512367, + 0.008406342007219791, + 0.008406609296798706, + 0.0130841713398695, + 0.0069542499259114265, + 0.008620129898190498, + 0.004394324496388435, + 0.022701002657413483, + 0.006916014011949301, + 0.026340464130043983, + 0.016102982684969902, + 0.017843997105956078, + 0.014947759918868542, + 0.0157097727060318, + 0.004881690721958876, + 0.005311344750225544, + 0.0065673235803842545, + 0.012458885088562965, + 0.006763782352209091, + 0.007538301404565573, + 0.00570006063207984, + 0.00634072907269001, + 0.003317601280286908, + 0.006494047585874796, + 0.006003641057759523, + 0.01291132066398859, + 0.007883152924478054, + 0.014644687995314598, + 0.008944215252995491, + 0.010654693469405174, + 0.006334994453936815, + 0.00861157476902008, + 0.007552189286798239, + 0.013081161305308342, + 0.010712282732129097, + 0.02450481802225113, + 0.009861909784376621, + 0.01645531691610813, + 0.006042531691491604, + 0.013348078355193138, + 0.010392187163233757, + 0.015511225908994675, + 0.011586316861212254, + 0.014255693182349205, + 0.017812442034482956, + 0.014190012589097023, + 0.002911157673224807, + 0.004204804077744484, + 0.0026598921976983547, + 0.004393148235976696, + 0.004772383254021406, + 0.005750851705670357, + 0.004008645191788673, + 0.004823057912290096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.054975058883428574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902136236429214 + ], + "opposite_logprob_tokens": [ + -2.4640214443206787 + ], + "logprob_reward": -0.08902136236429214, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902136236429214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15439555048942566, + 0.15420624613761902, + 0.12065508961677551, + 0.06738913804292679, + 0.13159900903701782, + 0.10672949254512787, + 0.040947988629341125, + 0.10547647625207901, + 0.08686351031064987, + 0.0317375548183918 + ], + "credit_map": [ + 0.011528792791068554, + 0.01668590120971203, + 0.019705191254615784, + 0.016351819038391113, + 0.02361721731722355, + 0.02344582788646221, + 0.020125150680541992, + 0.022935643792152405, + 0.019236601889133453, + 0.014198370277881622, + 0.01912165805697441, + 0.016849590465426445, + 0.014878230169415474, + 0.027031054720282555, + 0.018857210874557495, + 0.024033525958657265, + 0.011619632132351398, + 0.011699610389769077, + 0.012203049845993519, + 0.020234545692801476, + 0.012285142205655575, + 0.020377129316329956, + 0.016537752002477646, + 0.01569822058081627, + 0.004572744481265545, + 0.004786578472703695, + 0.007492241449654102, + 0.008719644509255886, + 0.009170902892947197, + 0.016050327569246292, + 0.007253891788423061, + 0.009342804551124573, + 0.00697789853438735, + 0.01174833532422781, + 0.013445081189274788, + 0.016435913741588593, + 0.022892575711011887, + 0.02905823104083538, + 0.01345216017216444, + 0.017588820308446884, + 0.007039955817162991, + 0.011833480559289455, + 0.011450331658124924, + 0.01577397622168064, + 0.017629120498895645, + 0.014504860155284405, + 0.015830347314476967, + 0.012667423114180565, + 0.002431169617921114, + 0.004309412091970444, + 0.003559112548828125, + 0.006283140275627375, + 0.004223954398185015, + 0.010048873722553253, + 0.004203834570944309, + 0.005888494197279215, + 0.006415217183530331, + 0.011922289617359638, + 0.008517301641404629, + 0.015406664460897446, + 0.010557510890066624, + 0.026418117806315422, + 0.010544905439019203, + 0.01569446735084057, + 0.005344908218830824, + 0.011926405131816864, + 0.011727648787200451, + 0.011484670452773571, + 0.008851551450788975, + 0.012630125507712364, + 0.012813758105039597, + 0.012084435671567917, + 0.0021969748195260763, + 0.0029696417041122913, + 0.0026200092397630215, + 0.004967902321368456, + 0.003889144631102681, + 0.006258578039705753, + 0.003960126079618931, + 0.004875176586210728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.08902136236429214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002648711204529 + ], + "opposite_logprob_tokens": [ + -2.695026397705078 + ], + "logprob_reward": -0.07002648711204529, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002648711204529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23841547966003418, + 0.09422600269317627, + 0.05794805660843849, + 0.09022065997123718, + 0.07482076436281204, + 0.06796814501285553, + 0.05913941189646721, + 0.1528819501399994, + 0.1439572274684906, + 0.0204223170876503 + ], + "credit_map": [ + 0.00900193676352501, + 0.014196346513926983, + 0.014505036175251007, + 0.014524001628160477, + 0.024462783709168434, + 0.0750190019607544, + 0.01703965663909912, + 0.06966667622327805, + 0.01347745768725872, + 0.008987491019070148, + 0.011075214482843876, + 0.011752001009881496, + 0.00941037479788065, + 0.014467978850007057, + 0.010113096795976162, + 0.014942381531000137, + 0.004394451156258583, + 0.006920403800904751, + 0.005035149399191141, + 0.006396133918315172, + 0.0068250009790062904, + 0.012015745043754578, + 0.0069177886471152306, + 0.009443377144634724, + 0.0053093344904482365, + 0.006443793885409832, + 0.008227569051086903, + 0.010267386212944984, + 0.012954779900610447, + 0.023260708898305893, + 0.00899603869765997, + 0.014761037193238735, + 0.0037740920670330524, + 0.010270601138472557, + 0.006681274622678757, + 0.01559348776936531, + 0.007542791776359081, + 0.011784135363996029, + 0.009644072502851486, + 0.009530307725071907, + 0.004934868775308132, + 0.006325825583189726, + 0.008295152336359024, + 0.010954759083688259, + 0.00904902908951044, + 0.010561969131231308, + 0.009592912159860134, + 0.008253623731434345, + 0.0028558396734297276, + 0.007119840011000633, + 0.00385768199339509, + 0.009738589636981487, + 0.006040457636117935, + 0.01348444726318121, + 0.0062665073201060295, + 0.00977603904902935, + 0.010400664992630482, + 0.017014941200613976, + 0.012695936486124992, + 0.020280228927731514, + 0.01502601895481348, + 0.03994959965348244, + 0.015048296190798283, + 0.022466253489255905, + 0.007757149171084166, + 0.02451830543577671, + 0.00944917555898428, + 0.01828671619296074, + 0.02311640977859497, + 0.019838476553559303, + 0.021443188190460205, + 0.019547807052731514, + 0.0020858626812696457, + 0.0020489103626459837, + 0.0021184959914535284, + 0.0026569983456283808, + 0.0025510063860565424, + 0.003342660376802087, + 0.0023428366985172033, + 0.003275544149801135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.07002648711204529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a resolute old man with graying hair, a beard, and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, rough linen robe with sleeves rustling in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a dense bamboo forest with massive, emerald green bamboos swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with moist earth and withered, yellowing leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a constant, subtle tremor causing the environment to oscillate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47430819272994995 + ], + "opposite_logprob_tokens": [ + -0.97430819272995 + ], + "logprob_reward": -0.47430819272994995, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47430819272994995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28070399165153503, + 0.08875172585248947, + 0.11643989384174347, + 0.06106347590684891, + 0.05720875784754753, + 0.0652928501367569, + 0.06031830981373787, + 0.07275707274675369, + 0.14127936959266663, + 0.05618452653288841 + ], + "credit_map": [ + 0.010600611567497253, + 0.009449368342757225, + 0.0162640530616045, + 0.015448424965143204, + 0.021479211747646332, + 0.09643074870109558, + 0.016416415572166443, + 0.09461517632007599, + 0.007535175420343876, + 0.008103150874376297, + 0.00786746758967638, + 0.011107898317277431, + 0.01075019035488367, + 0.016985919326543808, + 0.012202287092804909, + 0.014199634082615376, + 0.0066104005090892315, + 0.011855952441692352, + 0.007055864203721285, + 0.020522188395261765, + 0.014759854413568974, + 0.01568378321826458, + 0.02315075509250164, + 0.016801100224256516, + 0.0038792560808360577, + 0.007160336710512638, + 0.0052795796655118465, + 0.00897261407226324, + 0.009262790903449059, + 0.009904172271490097, + 0.008220652118325233, + 0.008384074084460735, + 0.0023441403172910213, + 0.005640896037220955, + 0.0031670802272856236, + 0.01307558361440897, + 0.008169126696884632, + 0.010029864497482777, + 0.00729134539142251, + 0.007490721996873617, + 0.00537912966683507, + 0.00796921644359827, + 0.007535446900874376, + 0.010189227759838104, + 0.009783613495528698, + 0.007950455881655216, + 0.009092206135392189, + 0.007393557578325272, + 0.003333719912916422, + 0.005659470800310373, + 0.003691899124532938, + 0.011103135533630848, + 0.00790451094508171, + 0.01258394680917263, + 0.008027163334190845, + 0.008014465682208538, + 0.004092930816113949, + 0.005772589705884457, + 0.0065493457950651646, + 0.012133397161960602, + 0.00753442058339715, + 0.018108809366822243, + 0.0072608329355716705, + 0.01130474079400301, + 0.006380237638950348, + 0.015063690952956676, + 0.014081367291510105, + 0.028124947100877762, + 0.019796881824731827, + 0.018094180151820183, + 0.020774397999048233, + 0.018963666632771492, + 0.006480618845671415, + 0.0059954761527478695, + 0.005957827437669039, + 0.00789687130600214, + 0.006563282106071711, + 0.008947566151618958, + 0.006371165160089731, + 0.007971718907356262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.47430819272994995 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_11/index_tensor([2179, 2179])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_11/index_tensor([2179, 2179])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..469f026b385c554ac7239350a1d467abf83951ea --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_11/index_tensor([2179, 2179])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a slow-motion shot from a horizontal angle, the camera pans with remarkable fluidity from left to right, capturing a quiet, modern, open-plan office. A beam of warm afternoon sunlight streams through the vast floor-to-ceiling windows, creating a distinct Tyndall effect in the air. Countless tiny dust particles dance slowly within the beam, creating the delicate atmosphere of light and shadow that is Makoto Shinkai's hallmark. As the camera focuses on a workstation screen made of translucent frosted glass, set in a thin silver metal frame, it suddenly, without warning, erupts in a series of lightning-like cracks, starting at its center. These cracks spread outward at a visible speed, and then the entire screen undergoes a violent, brittle collapse, sending thousands of fragments of varying sizes gracefully shattering through the air like enchanted crystal. In extreme slow motion, each fragment captures and refracts the afternoon sunlight, flashing with rainbow-like halos and dazzling starlight. They slowly drift and tumble in the air, intertwined with the dust in the light column, forming a silent and gorgeous visual ballet. The entire dramatic collapse process is fully recorded by this smooth single shot, and the picture is full of broken beauty and a quiet and melancholy atmosphere.", + "seed": 9284, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_11/index_2179_seed_9284_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10016323525924235, + "vlm_logprob": -0.10016323525924235, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.10016323525924235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09000924229621887 + ], + "opposite_logprob_tokens": [ + -2.7150092124938965 + ], + "logprob_reward": -0.09000924229621887, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09000924229621887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15396882593631744, + 0.02020317316055298, + 0.09923627227544785, + 0.06907317787408829, + 0.08599641919136047, + 0.04267401248216629, + 0.04184931889176369, + 0.42726585268974304, + 0.030330179259181023, + 0.02940274029970169 + ], + "credit_map": [ + 0.016137050464749336, + 0.023608161136507988, + 0.027462154626846313, + 0.03586455062031746, + 0.011544739827513695, + 0.011410362087190151, + 0.01347985677421093, + 0.014461944811046124, + 0.0029607105534523726, + 0.001736641046591103, + 0.004120389465242624, + 0.003185031935572624, + 0.001377186505123973, + 0.0020241462625563145, + 0.0027617814484983683, + 0.002037285827100277, + 0.012334932573139668, + 0.03330310061573982, + 0.009981587529182434, + 0.008071986958384514, + 0.011600468307733536, + 0.010307570919394493, + 0.007394781336188316, + 0.006241855677217245, + 0.00407559797167778, + 0.01754145137965679, + 0.007348755374550819, + 0.005831015761941671, + 0.01723363623023033, + 0.00642307847738266, + 0.006388898938894272, + 0.004230746533721685, + 0.009978032670915127, + 0.013115945272147655, + 0.015378138981759548, + 0.008200914599001408, + 0.010468607768416405, + 0.013231591321527958, + 0.008995261043310165, + 0.006627921015024185, + 0.005584515631198883, + 0.005777114070951939, + 0.004603487905114889, + 0.004090055823326111, + 0.005237299017608166, + 0.0061431583017110825, + 0.004647472407668829, + 0.0065909093245863914, + 0.00471634091809392, + 0.006502646021544933, + 0.003011920489370823, + 0.0032663203310221434, + 0.006404584739357233, + 0.008749129250645638, + 0.004671504255384207, + 0.0045268721878528595, + 0.01228492520749569, + 0.1720714569091797, + 0.01241378765553236, + 0.009387901052832603, + 0.18051613867282867, + 0.02003411389887333, + 0.009841536171734333, + 0.010716015473008156, + 0.0029129409231245518, + 0.0035412332508713007, + 0.003994302358478308, + 0.004688948858529329, + 0.0035408069379627705, + 0.003966720774769783, + 0.0036125099286437035, + 0.004072715528309345, + 0.0035459899809211493, + 0.0027199094183743, + 0.0040106684900820255, + 0.0029687033966183662, + 0.002954361028969288, + 0.004606273025274277, + 0.0043813190422952175, + 0.004215516149997711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.09000924229621887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017514378996565938 + ], + "opposite_logprob_tokens": [ + -6.37675142288208 + ], + "logprob_reward": -0.0017514378996565938, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017514378996565938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05918395519256592, + 0.02228625677525997, + 0.03472475707530975, + 0.07683193683624268, + 0.023684874176979065, + 0.025259656831622124, + 0.015215867199003696, + 0.70158851146698, + 0.022381586953997612, + 0.018842661753296852 + ], + "credit_map": [ + 0.005284827668219805, + 0.009573588147759438, + 0.007177914958447218, + 0.012843962758779526, + 0.005764390807598829, + 0.0063293492421507835, + 0.005771401803940535, + 0.0064385212026536465, + 0.003380276495590806, + 0.002048170892521739, + 0.004486515652388334, + 0.0035264769103378057, + 0.0015702306991443038, + 0.0021434978116303682, + 0.0029508869629353285, + 0.0021802012342959642, + 0.004668171983212233, + 0.005191754084080458, + 0.006605119444429874, + 0.0052515724673867226, + 0.0023210346698760986, + 0.0031908571254462004, + 0.0032146181911230087, + 0.004281631670892239, + 0.005565506871789694, + 0.015000829473137856, + 0.008369394578039646, + 0.007898754440248013, + 0.014698479324579239, + 0.0067725456319749355, + 0.013520405627787113, + 0.005006014369428158, + 0.0019589553121477365, + 0.0032583593856543303, + 0.004057126119732857, + 0.0024469508789479733, + 0.002713277004659176, + 0.0032022749073803425, + 0.003840208752080798, + 0.0022077211178839207, + 0.0025189206935465336, + 0.0022725537419319153, + 0.003414214588701725, + 0.0037179209757596254, + 0.0022892928682267666, + 0.0031440905295312405, + 0.0028653384651988745, + 0.005037324968725443, + 0.0018108247313648462, + 0.0025395783595740795, + 0.0016796671552583575, + 0.0015734527260065079, + 0.00161787960678339, + 0.002529117977246642, + 0.001507437671534717, + 0.0019579092040657997, + 0.019950568675994873, + 0.2839052677154541, + 0.0193612240254879, + 0.016259504482150078, + 0.2972697615623474, + 0.03289642557501793, + 0.014763379469513893, + 0.017182378098368645, + 0.0022190632298588753, + 0.0028997219633311033, + 0.002941942075267434, + 0.003204320091754198, + 0.0025539114139974117, + 0.0025375240948051214, + 0.002950520720332861, + 0.0030745824333280325, + 0.0022371888626366854, + 0.0021319882944226265, + 0.0021156449802219868, + 0.0021060877479612827, + 0.002310017589479685, + 0.002877400489524007, + 0.002689352026209235, + 0.0023749808315187693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.0017514378996565938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962323863059282 + ], + "opposite_logprob_tokens": [ + -5.130962371826172 + ], + "logprob_reward": -0.005962323863059282, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962323863059282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03379599377512932, + 0.025604957714676857, + 0.0802234336733818, + 0.1370561718940735, + 0.03320196643471718, + 0.03887305408716202, + 0.055741820484399796, + 0.5419089794158936, + 0.035341404378414154, + 0.018252182751893997 + ], + "credit_map": [ + 0.003107645083218813, + 0.00420044781640172, + 0.005283363629132509, + 0.006365714129060507, + 0.0025045371148735285, + 0.0030690545681864023, + 0.003355663502588868, + 0.0059095690958201885, + 0.0037723619025200605, + 0.002634977223351598, + 0.004816622007638216, + 0.005690252408385277, + 0.001499430276453495, + 0.002131112851202488, + 0.002503895666450262, + 0.0025563037488609552, + 0.007053894456475973, + 0.019839119166135788, + 0.01930040307343006, + 0.009140548296272755, + 0.007350398693233728, + 0.005948581732809544, + 0.005082476418465376, + 0.006508017890155315, + 0.009007567539811134, + 0.03169802948832512, + 0.01349507737904787, + 0.013434011489152908, + 0.031169453635811806, + 0.013164041563868523, + 0.017819520086050034, + 0.0072684758342802525, + 0.0029851198196411133, + 0.0041710683144629, + 0.00763695128262043, + 0.003950315061956644, + 0.004097078926861286, + 0.003454167628660798, + 0.0035639596171677113, + 0.003343306016176939, + 0.004413140006363392, + 0.00372018082998693, + 0.005867288913577795, + 0.005426185671240091, + 0.0035797685850411654, + 0.006988557055592537, + 0.003928841557353735, + 0.004949093330651522, + 0.00393786933273077, + 0.018630104139447212, + 0.002885678317397833, + 0.0027633982244879007, + 0.017107488587498665, + 0.0037358065601438284, + 0.003600314725190401, + 0.003081158036366105, + 0.015862075611948967, + 0.21934406459331512, + 0.015215880237519741, + 0.012661375105381012, + 0.22900722920894623, + 0.024994349107146263, + 0.011374750174582005, + 0.013449241407215595, + 0.0028387235943228006, + 0.0038263010792434216, + 0.00354580907151103, + 0.009770648553967476, + 0.0035169224720448256, + 0.0036973420064896345, + 0.004453417845070362, + 0.0036922376602888107, + 0.0020276547875255346, + 0.0022765472531318665, + 0.0023691316600888968, + 0.0019507340621203184, + 0.0019003800116479397, + 0.002835845109075308, + 0.002576900413259864, + 0.002314990386366844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.005962323863059282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065670177340508 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065670177340508, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065670177340508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03239426761865616, + 0.019271787256002426, + 0.03443123400211334, + 0.15338321030139923, + 0.0454239584505558, + 0.0404692143201828, + 0.024602487683296204, + 0.6150725483894348, + 0.018196510151028633, + 0.01675482466816902 + ], + "credit_map": [ + 0.003045374294742942, + 0.005217297933995724, + 0.004422582220286131, + 0.004843541420996189, + 0.002547385636717081, + 0.004162146244198084, + 0.0034494560677558184, + 0.004706478212028742, + 0.003401469439268112, + 0.002009618328884244, + 0.0035765497013926506, + 0.0038619618862867355, + 0.0008828401332721114, + 0.001343857147730887, + 0.0017184442840516567, + 0.0024770444724708796, + 0.004856558050960302, + 0.003393835388123989, + 0.007475972641259432, + 0.006353984586894512, + 0.0020176218822598457, + 0.0031278689857572317, + 0.003250590991228819, + 0.00395479891449213, + 0.007704692892730236, + 0.05101245641708374, + 0.006136470474302769, + 0.008649863302707672, + 0.05111541599035263, + 0.008284886367619038, + 0.01517249085009098, + 0.005306916311383247, + 0.0035396264865994453, + 0.0039036606904119253, + 0.008758360520005226, + 0.006015482824295759, + 0.005318505689501762, + 0.004696589894592762, + 0.00835692510008812, + 0.004834805149585009, + 0.003552244044840336, + 0.003969141282141209, + 0.005020536482334137, + 0.004539025016129017, + 0.00479365698993206, + 0.00437538418918848, + 0.0038735452108085155, + 0.010345677845180035, + 0.0018316941568627954, + 0.007732299156486988, + 0.0015605963999405503, + 0.0014836477348580956, + 0.006482691969722509, + 0.002274254569783807, + 0.0016776489792391658, + 0.0015596522716805339, + 0.017709437757730484, + 0.24861374497413635, + 0.017373107373714447, + 0.01378068421036005, + 0.2602371275424957, + 0.028859121724963188, + 0.013261108659207821, + 0.015238151885569096, + 0.0014651374658569694, + 0.0020125822629779577, + 0.0025276318192481995, + 0.004166679456830025, + 0.0014735847944393754, + 0.0018257051706314087, + 0.0019869671668857336, + 0.0027382203843444586, + 0.001704114256426692, + 0.0017793793231248856, + 0.0023316005244851112, + 0.0020978825632482767, + 0.0017369348788633943, + 0.0024829634930938482, + 0.0025953566655516624, + 0.0020265921484678984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.016065670177340508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870275020599365 + ], + "opposite_logprob_tokens": [ + -1.1370275020599365 + ], + "logprob_reward": -0.3870275020599365, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870275020599365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03292149305343628, + 0.017251893877983093, + 0.040159448981285095, + 0.10123606771230698, + 0.03103060834109783, + 0.02182689681649208, + 0.06884009391069412, + 0.6666790843009949, + 0.008347835391759872, + 0.01170665118843317 + ], + "credit_map": [ + 0.0030887937173247337, + 0.00640524085611105, + 0.003937542904168367, + 0.004422164522111416, + 0.001692265155725181, + 0.004620571620762348, + 0.0033798643853515387, + 0.005375043023377657, + 0.0022498369216918945, + 0.0030923846643418074, + 0.002377548022195697, + 0.0025898320600390434, + 0.0013032519491389394, + 0.002063547959551215, + 0.0018941001035273075, + 0.0016813885886222124, + 0.003521159989759326, + 0.009829005226492882, + 0.011867248453199863, + 0.004230977967381477, + 0.004205825738608837, + 0.0021582741755992174, + 0.0015402712160721421, + 0.0028066837694495916, + 0.005491519812494516, + 0.032785814255476, + 0.0050032539293169975, + 0.006409360561519861, + 0.03212079405784607, + 0.006092413794249296, + 0.00931879784911871, + 0.004014102276414633, + 0.003010793821886182, + 0.003152082907035947, + 0.005150044336915016, + 0.003566908882930875, + 0.003577788593247533, + 0.004641506355255842, + 0.003810553578659892, + 0.004120926838368177, + 0.002492124680429697, + 0.0029485037084668875, + 0.002426914405077696, + 0.0020838724449276924, + 0.002879881067201495, + 0.003620419418439269, + 0.0021864764858037233, + 0.0031887032091617584, + 0.00468846783041954, + 0.025215622037649155, + 0.0030887406319379807, + 0.002849029144272208, + 0.021002937108278275, + 0.005314678885042667, + 0.003309224033728242, + 0.003371388418599963, + 0.01926017738878727, + 0.26856347918510437, + 0.0185666773468256, + 0.014910900965332985, + 0.2822118401527405, + 0.03292107209563255, + 0.01447059866040945, + 0.01577422209084034, + 0.0007282203296199441, + 0.0008546614553779364, + 0.001107370131649077, + 0.001480201375670731, + 0.0009134140564128757, + 0.0010251301573589444, + 0.0010623978450894356, + 0.00117643887642771, + 0.0013500521890819073, + 0.0015267551643773913, + 0.0011159817222505808, + 0.001163019216619432, + 0.0012729198206216097, + 0.0018136288272216916, + 0.002219538437202573, + 0.0012447539484128356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.3870275020599365 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow-motion shot from a horizontal angle, the camera pans with remarkable fluidity from left to right, capturing a quiet, modern, open-plan office. A beam of warm afternoon sunlight streams through the vast floor-to-ceiling windows, creating a distinct Tyndall effect in the air. Countless tiny dust particles dance slowly within the beam, creating the delicate atmosphere of light and shadow that is Makoto Shinkai's hallmark. As the camera focuses on a workstation screen made of translucent frosted glass, set in a thin silver metal frame, it suddenly, without warning, erupts in a series of lightning-like cracks, starting at its center. These cracks spread outward at a visible speed, and then the entire screen undergoes a violent, brittle collapse, sending thousands of fragments of varying sizes gracefully shattering through the air like enchanted crystal. In extreme slow motion, each fragment captures and refracts the afternoon sunlight, flashing with rainbow-like halos and dazzling starlight. They slowly drift and tumble in the air, intertwined with the dust in the light column, forming a silent and gorgeous visual ballet. The entire dramatic collapse process is fully recorded by this smooth single shot, and the picture is full of broken beauty and a quiet and melancholy atmosphere.", + "seed": 9285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_11/index_2179_seed_9285_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08547953062225133, + "vlm_logprob": -0.08547953062225133, + "vlm_yes_no_margin": 4.550000011920929, + "vlm_token_credit": -0.08547953062225133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016381768509745598 + ], + "opposite_logprob_tokens": [ + -6.141381740570068 + ], + "logprob_reward": -0.016381768509745598, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.016381768509745598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19302991032600403, + 0.052185624837875366, + 0.02954607456922531, + 0.031038539484143257, + 0.23966439068317413, + 0.1296289712190628, + 0.09034024924039841, + 0.08295387029647827, + 0.08624965697526932, + 0.06536272913217545 + ], + "credit_map": [ + 0.023539144545793533, + 0.02307216078042984, + 0.03081386163830757, + 0.019153498113155365, + 0.02009589597582817, + 0.024783924221992493, + 0.02418362908065319, + 0.027387799695134163, + 0.006701358128339052, + 0.008630547672510147, + 0.007072379346936941, + 0.004682793747633696, + 0.006278908811509609, + 0.006070585921406746, + 0.005385003052651882, + 0.007364046759903431, + 0.003480195300653577, + 0.0042678858153522015, + 0.003743348876014352, + 0.00414626207202673, + 0.0036697443574666977, + 0.003054231870919466, + 0.003968308679759502, + 0.003216101322323084, + 0.0039747473783791065, + 0.003788310568779707, + 0.0029820373747497797, + 0.00414697453379631, + 0.003038299037143588, + 0.0032674523536115885, + 0.006537512876093388, + 0.003303204895928502, + 0.023836154490709305, + 0.0178374033421278, + 0.0203230120241642, + 0.051780328154563904, + 0.01792687363922596, + 0.022256962954998016, + 0.057471394538879395, + 0.02823225036263466, + 0.014071407727897167, + 0.013031693175435066, + 0.015259260311722755, + 0.014208856038749218, + 0.016120729967951775, + 0.014053383842110634, + 0.026371298357844353, + 0.016512349247932434, + 0.007299435790628195, + 0.012628376483917236, + 0.009208337403833866, + 0.014841114170849323, + 0.00856354646384716, + 0.010371232405304909, + 0.014462695457041264, + 0.012965508736670017, + 0.007555787451565266, + 0.008511594496667385, + 0.007800590712577105, + 0.018054449930787086, + 0.008710791356861591, + 0.009205532260239124, + 0.014495735056698322, + 0.008619391359388828, + 0.011711429804563522, + 0.017820969223976135, + 0.010230939835309982, + 0.009158454835414886, + 0.010425198823213577, + 0.011098768562078476, + 0.0104984724894166, + 0.005305426195263863, + 0.0053090439178049564, + 0.00595544558018446, + 0.007623686455190182, + 0.01321781799197197, + 0.0064622764475643635, + 0.006700900848954916, + 0.011790855787694454, + 0.008302704431116581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.016381768509745598 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017316836165264249 + ], + "opposite_logprob_tokens": [ + -6.376731872558594 + ], + "logprob_reward": -0.0017316836165264249, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017316836165264249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15320932865142822, + 0.07951560616493225, + 0.07452677190303802, + 0.04595731198787689, + 0.14856885373592377, + 0.08572451025247574, + 0.13294756412506104, + 0.06509820371866226, + 0.04246469587087631, + 0.1719871163368225 + ], + "credit_map": [ + 0.017448194324970245, + 0.02345193177461624, + 0.0305838193744421, + 0.015611960552632809, + 0.018032975494861603, + 0.019497571513056755, + 0.011838162317872047, + 0.01674470491707325, + 0.010580244474112988, + 0.011131283827126026, + 0.012546321377158165, + 0.008626294322311878, + 0.009770932607352734, + 0.0097225122153759, + 0.007272294722497463, + 0.00986572541296482, + 0.009580224752426147, + 0.008866210468113422, + 0.01255873218178749, + 0.009375965222716331, + 0.008020910434424877, + 0.00958055816590786, + 0.007456724997609854, + 0.00908744614571333, + 0.005592228379100561, + 0.005719473119825125, + 0.006051997188478708, + 0.005244110710918903, + 0.005394659470766783, + 0.005198793485760689, + 0.0065157110802829266, + 0.006240339484065771, + 0.01281785499304533, + 0.011319873854517937, + 0.012036768719553947, + 0.035715386271476746, + 0.012994218617677689, + 0.013974281959235668, + 0.033369049429893494, + 0.016341427341103554, + 0.012032737955451012, + 0.01029309630393982, + 0.009093621745705605, + 0.011425173841416836, + 0.00770322373136878, + 0.008800914511084557, + 0.014943025074899197, + 0.011432713828980923, + 0.014209719374775887, + 0.016034062951803207, + 0.01806149072945118, + 0.01924922503530979, + 0.013857726007699966, + 0.015530243515968323, + 0.019296322017908096, + 0.01670878380537033, + 0.006101807113736868, + 0.009296710602939129, + 0.006173843517899513, + 0.012851875275373459, + 0.008576503954827785, + 0.00609741872176528, + 0.009973010048270226, + 0.006027032155543566, + 0.003949441947042942, + 0.00446121022105217, + 0.00713015254586935, + 0.007047233637422323, + 0.0038471922744065523, + 0.004963444545865059, + 0.006601112894713879, + 0.004464906174689531, + 0.018642472103238106, + 0.031245434656739235, + 0.016722045838832855, + 0.021146271377801895, + 0.02534293197095394, + 0.017953351140022278, + 0.023804835975170135, + 0.017129763960838318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0017316836165264249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101912025362253 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101912025362253, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101912025362253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3287857174873352, + 0.10767612606287003, + 0.031224828213453293, + 0.03537081554532051, + 0.12119728326797485, + 0.08685024827718735, + 0.07045070827007294, + 0.0676078125834465, + 0.047119446098804474, + 0.10371699184179306 + ], + "credit_map": [ + 0.05731618404388428, + 0.04060560464859009, + 0.04127012565732002, + 0.02751472406089306, + 0.03011378087103367, + 0.04614696279168129, + 0.030956411734223366, + 0.05486193299293518, + 0.017477624118328094, + 0.01285691000521183, + 0.012696223333477974, + 0.00792255625128746, + 0.009744836948812008, + 0.01632278598845005, + 0.008868350647389889, + 0.02178684063255787, + 0.0036301533691585064, + 0.004269811324775219, + 0.003839511889964342, + 0.003058342495933175, + 0.004030319396406412, + 0.004327056929469109, + 0.0030610926914960146, + 0.005008538719266653, + 0.003329220227897167, + 0.0029972929041832685, + 0.0036697830073535442, + 0.0044423663057386875, + 0.0027215289883315563, + 0.0040541235357522964, + 0.009688745252788067, + 0.004467756021767855, + 0.010238680988550186, + 0.008059170097112656, + 0.010421836748719215, + 0.03151292726397514, + 0.009998922236263752, + 0.010345171205699444, + 0.027654508128762245, + 0.012966062873601913, + 0.010018441826105118, + 0.009481018409132957, + 0.01185513287782669, + 0.00895464513450861, + 0.009960388764739037, + 0.010140128433704376, + 0.015354890376329422, + 0.011085603386163712, + 0.006757104303687811, + 0.00942474789917469, + 0.006430159788578749, + 0.010455220006406307, + 0.01193628553301096, + 0.007792920805513859, + 0.010607958771288395, + 0.007046309765428305, + 0.0072619556449353695, + 0.007474946323782206, + 0.006838370580226183, + 0.015691768378019333, + 0.005968081299215555, + 0.007522795349359512, + 0.010589633136987686, + 0.006260263733565807, + 0.004084230400621891, + 0.004624706692993641, + 0.005932897329330444, + 0.007136291358619928, + 0.005274005234241486, + 0.004627871327102184, + 0.01036380510777235, + 0.0050756377167999744, + 0.01235435251146555, + 0.014686616137623787, + 0.01209472306072712, + 0.016509678214788437, + 0.012039005756378174, + 0.011711815372109413, + 0.013255289755761623, + 0.011065508238971233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.004101912025362253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176065757870674 + ], + "opposite_logprob_tokens": [ + -4.018176078796387 + ], + "logprob_reward": -0.018176065757870674, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176065757870674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13677304983139038, + 0.03921346366405487, + 0.033672768622636795, + 0.05144774168729782, + 0.2759344279766083, + 0.14637994766235352, + 0.0859275832772255, + 0.11109760403633118, + 0.058637574315071106, + 0.06091582402586937 + ], + "credit_map": [ + 0.01499374583363533, + 0.014223870821297169, + 0.018448922783136368, + 0.011521901935338974, + 0.015664467588067055, + 0.019130557775497437, + 0.01681741699576378, + 0.025972163304686546, + 0.0037714748177677393, + 0.004125303588807583, + 0.005448986776173115, + 0.0034933825954794884, + 0.004753272980451584, + 0.0057066320441663265, + 0.004170277621597052, + 0.007744126487523317, + 0.00423094816505909, + 0.004697317723184824, + 0.003945792093873024, + 0.003051899606361985, + 0.004162253346294165, + 0.005235800985246897, + 0.004482567775994539, + 0.0038661856669932604, + 0.006714144255965948, + 0.007452256046235561, + 0.004973500967025757, + 0.005236213095486164, + 0.005684495437890291, + 0.006125284358859062, + 0.008803963661193848, + 0.006457881070673466, + 0.02700716070830822, + 0.020769713446497917, + 0.024655554443597794, + 0.06732123345136642, + 0.024542739614844322, + 0.026179686188697815, + 0.05142950266599655, + 0.03402883559465408, + 0.013677243143320084, + 0.014795307070016861, + 0.015499298460781574, + 0.021041344851255417, + 0.015740536153316498, + 0.015976356342434883, + 0.02940352074801922, + 0.020246325060725212, + 0.008891785517334938, + 0.010392495431005955, + 0.008572524413466454, + 0.011538859456777573, + 0.0134833212941885, + 0.00899504218250513, + 0.013785514049232006, + 0.01026803720742464, + 0.012518447823822498, + 0.01715477928519249, + 0.00958989281207323, + 0.018242526799440384, + 0.01667305827140808, + 0.010378503240644932, + 0.01601230539381504, + 0.010528085753321648, + 0.0064544761553406715, + 0.00604649493470788, + 0.008783247321844101, + 0.008089853450655937, + 0.00661311624571681, + 0.006093068514019251, + 0.010139930061995983, + 0.006417383905500174, + 0.00529865687713027, + 0.00773794436827302, + 0.008139069192111492, + 0.008420329540967941, + 0.006665593944489956, + 0.005877731833606958, + 0.01288911234587431, + 0.005887384060770273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.018176065757870674 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870062232017517 + ], + "opposite_logprob_tokens": [ + -1.1370062828063965 + ], + "logprob_reward": -0.3870062232017517, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870062232017517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1273767501115799, + 0.04199231415987015, + 0.04203645884990692, + 0.05425213649868965, + 0.11948516964912415, + 0.14288869500160217, + 0.07218076288700104, + 0.10775377601385117, + 0.09414362907409668, + 0.19789034128189087 + ], + "credit_map": [ + 0.01347599271684885, + 0.01668531820178032, + 0.015978854149580002, + 0.011541567742824554, + 0.013678868301212788, + 0.015392527915537357, + 0.017761286348104477, + 0.022862328216433525, + 0.004561454523354769, + 0.0038721899036318064, + 0.0059403046034276485, + 0.00551491929218173, + 0.00418963935226202, + 0.00631319172680378, + 0.003887246595695615, + 0.007713364902883768, + 0.004479023162275553, + 0.004701226484030485, + 0.004489742685109377, + 0.006026189308613539, + 0.004301753360778093, + 0.005871673580259085, + 0.005407451186329126, + 0.006759398616850376, + 0.008591154590249062, + 0.007105865981429815, + 0.007030507083982229, + 0.0053582326509058475, + 0.0057412320747971535, + 0.005816989578306675, + 0.008384810760617256, + 0.006223342847079039, + 0.011887168511748314, + 0.007979734800755978, + 0.009907673113048077, + 0.02729972079396248, + 0.00803807470947504, + 0.012074245139956474, + 0.028041694313287735, + 0.014256858266890049, + 0.0184018611907959, + 0.015493171289563179, + 0.019229469820857048, + 0.01814642734825611, + 0.014750013127923012, + 0.016153933480381966, + 0.02310195192694664, + 0.01761186309158802, + 0.0059382435865700245, + 0.009148919023573399, + 0.008213549852371216, + 0.013051800429821014, + 0.007064143195748329, + 0.0077145700342953205, + 0.013408863916993141, + 0.007640674244612455, + 0.008584079332649708, + 0.01191688422113657, + 0.013417809270322323, + 0.01943417266011238, + 0.010963642969727516, + 0.012026659213006496, + 0.0160079225897789, + 0.015402608551084995, + 0.008570680394768715, + 0.006249359808862209, + 0.012826251797378063, + 0.015304280444979668, + 0.009150815196335316, + 0.010699199512600899, + 0.016093660145998, + 0.015249384567141533, + 0.021390696987509727, + 0.03579767793416977, + 0.0182438213378191, + 0.023425176739692688, + 0.024623097851872444, + 0.019726889207959175, + 0.031917620450258255, + 0.02276536263525486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.3870062232017517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow-motion shot from a horizontal angle, the camera pans with remarkable fluidity from left to right, capturing a quiet, modern, open-plan office. A beam of warm afternoon sunlight streams through the vast floor-to-ceiling windows, creating a distinct Tyndall effect in the air. Countless tiny dust particles dance slowly within the beam, creating the delicate atmosphere of light and shadow that is Makoto Shinkai's hallmark. As the camera focuses on a workstation screen made of translucent frosted glass, set in a thin silver metal frame, it suddenly, without warning, erupts in a series of lightning-like cracks, starting at its center. These cracks spread outward at a visible speed, and then the entire screen undergoes a violent, brittle collapse, sending thousands of fragments of varying sizes gracefully shattering through the air like enchanted crystal. In extreme slow motion, each fragment captures and refracts the afternoon sunlight, flashing with rainbow-like halos and dazzling starlight. They slowly drift and tumble in the air, intertwined with the dust in the light column, forming a silent and gorgeous visual ballet. The entire dramatic collapse process is fully recorded by this smooth single shot, and the picture is full of broken beauty and a quiet and melancholy atmosphere.", + "seed": 9286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_11/index_2179_seed_9286_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07731569726020097, + "vlm_logprob": -0.07731569726020097, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07731569726020097, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043748728930950165 + ], + "opposite_logprob_tokens": [ + -3.793748617172241 + ], + "logprob_reward": -0.043748728930950165, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.043748728930950165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39779841899871826, + 0.07225635647773743, + 0.03920481353998184, + 0.055264610797166824, + 0.0457555428147316, + 0.05908118188381195, + 0.08597775548696518, + 0.13610240817070007, + 0.05718934163451195, + 0.051369499415159225 + ], + "credit_map": [ + 0.04064900800585747, + 0.05247850716114044, + 0.02977207861840725, + 0.032386414706707, + 0.09347550570964813, + 0.07672750949859619, + 0.03759791702032089, + 0.034711532294750214, + 0.009340010583400726, + 0.008667968213558197, + 0.005264452658593655, + 0.01279052346944809, + 0.010744070634245872, + 0.01034251507371664, + 0.01044494565576315, + 0.004661878105252981, + 0.005047007463872433, + 0.0062774005346000195, + 0.003223465057089925, + 0.004559673368930817, + 0.005115419160574675, + 0.006253354717046022, + 0.004502946976572275, + 0.004225549753755331, + 0.006759692449122667, + 0.009081889875233173, + 0.0031947260722517967, + 0.004639329854398966, + 0.007137614767998457, + 0.0111928042024374, + 0.006998406257480383, + 0.006260154768824577, + 0.005156662315130234, + 0.007008103653788567, + 0.003590910229831934, + 0.0066230869852006435, + 0.005855214782059193, + 0.007357023656368256, + 0.005951246712356806, + 0.004213300067931414, + 0.007401263806968927, + 0.01151927001774311, + 0.004809897858649492, + 0.004399589728564024, + 0.007624803576618433, + 0.010998143814504147, + 0.007139953784644604, + 0.005188265815377235, + 0.009869820438325405, + 0.01872890256345272, + 0.005866777151823044, + 0.00855977926403284, + 0.009316554293036461, + 0.017285138368606567, + 0.010766328312456608, + 0.005584468133747578, + 0.014794889837503433, + 0.02446100115776062, + 0.011306338012218475, + 0.014166335575282574, + 0.01579701527953148, + 0.029720351099967957, + 0.01767801120877266, + 0.00817848276346922, + 0.005496329162269831, + 0.00933795515447855, + 0.003984193783253431, + 0.004146820865571499, + 0.0062581743113696575, + 0.011752809397876263, + 0.010623317211866379, + 0.00558974826708436, + 0.005713385064154863, + 0.00905510876327753, + 0.004714220762252808, + 0.005176587495952845, + 0.004755900241434574, + 0.008178886026144028, + 0.00867894385010004, + 0.0050964755937457085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.043748728930950165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008684013038873672 + ], + "opposite_logprob_tokens": [ + -4.758684158325195 + ], + "logprob_reward": -0.008684013038873672, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008684013038873672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5751892924308777, + 0.08629823476076126, + 0.0251613799482584, + 0.022077905014157295, + 0.11470769345760345, + 0.03676517307758331, + 0.03929377719759941, + 0.037487514317035675, + 0.04055939242243767, + 0.02245967648923397 + ], + "credit_map": [ + 0.07524695992469788, + 0.08097419887781143, + 0.03294123709201813, + 0.048593536019325256, + 0.07531309127807617, + 0.15426942706108093, + 0.04275213181972504, + 0.06509870290756226, + 0.007438489235937595, + 0.01059780828654766, + 0.006746858824044466, + 0.01590699702501297, + 0.012629673816263676, + 0.013407542370259762, + 0.01219155639410019, + 0.007379310671240091, + 0.003242560662329197, + 0.004465488251298666, + 0.002151251770555973, + 0.003633270738646388, + 0.003331981599330902, + 0.004323006607592106, + 0.0021910900250077248, + 0.0018227295950055122, + 0.0028402346652001143, + 0.003997489344328642, + 0.0013530058786273003, + 0.002152802888303995, + 0.002466481411829591, + 0.004070129711180925, + 0.003640606766566634, + 0.0015571546973660588, + 0.006219964474439621, + 0.010350143536925316, + 0.004936343990266323, + 0.009202035143971443, + 0.03756086155772209, + 0.007710463833063841, + 0.03227754682302475, + 0.006450336426496506, + 0.0045901695266366005, + 0.007359047885984182, + 0.0020760120823979378, + 0.004367042798548937, + 0.004873861558735371, + 0.007273009978234768, + 0.003631278406828642, + 0.0025947506073862314, + 0.0044029466807842255, + 0.007849892601370811, + 0.003329056315124035, + 0.003775379154831171, + 0.003730782074853778, + 0.008279351517558098, + 0.00485308188945055, + 0.0030732867307960987, + 0.003366201650351286, + 0.00608142651617527, + 0.0019775398541241884, + 0.0035612157080322504, + 0.004965210799127817, + 0.008289060555398464, + 0.00604340061545372, + 0.003203461179509759, + 0.003601911710575223, + 0.007874701172113419, + 0.0025033357087522745, + 0.003642901312559843, + 0.004346106667071581, + 0.009338891133666039, + 0.00532690342515707, + 0.0039246403612196445, + 0.0030814234632998705, + 0.004983965307474136, + 0.0017132738139480352, + 0.0024694791063666344, + 0.002181967720389366, + 0.003643944626674056, + 0.002691259142011404, + 0.0016943649388849735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.008684013038873672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795927442610264 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.009795927442610264, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795927442610264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16087056696414948, + 0.1958029419183731, + 0.03935135528445244, + 0.05738931521773338, + 0.14012017846107483, + 0.1023322269320488, + 0.0800209492444992, + 0.12303287535905838, + 0.029551759362220764, + 0.07152785360813141 + ], + "credit_map": [ + 0.022478988394141197, + 0.02526962012052536, + 0.009008650667965412, + 0.01313546858727932, + 0.02248810976743698, + 0.03458714857697487, + 0.014860616065561771, + 0.01904195547103882, + 0.014923694543540478, + 0.023978227749466896, + 0.01546918973326683, + 0.037569303065538406, + 0.024373602122068405, + 0.023244095966219902, + 0.03889627754688263, + 0.017348552122712135, + 0.0047586411237716675, + 0.007332193199545145, + 0.0030538474675267935, + 0.006589418742805719, + 0.004127189051359892, + 0.005510623101145029, + 0.0044780694879591465, + 0.0035013710148632526, + 0.005311358254402876, + 0.007731385063380003, + 0.004008250776678324, + 0.01083639357239008, + 0.007120957598090172, + 0.007512642536312342, + 0.00960170105099678, + 0.005266626365482807, + 0.011906657367944717, + 0.00919873546808958, + 0.007253377232700586, + 0.02475591003894806, + 0.03152116760611534, + 0.011265059001743793, + 0.03497501090168953, + 0.009244243614375591, + 0.010846462100744247, + 0.020539049059152603, + 0.006340736988931894, + 0.010927300900220871, + 0.013521269895136356, + 0.02066168561577797, + 0.012300614267587662, + 0.007195106241852045, + 0.0066987089812755585, + 0.009235515259206295, + 0.009169019758701324, + 0.011942947283387184, + 0.008460691198706627, + 0.009287849068641663, + 0.01755147986114025, + 0.007674728520214558, + 0.010854504071176052, + 0.016464557498693466, + 0.00753382034599781, + 0.011273367330431938, + 0.01729787327349186, + 0.017314091324806213, + 0.0313381589949131, + 0.01095649041235447, + 0.0038019681815057993, + 0.005313186440616846, + 0.002257400657981634, + 0.0028986125253140926, + 0.003438354469835758, + 0.005253835115581751, + 0.0025332565419375896, + 0.0040551433339715, + 0.008670249953866005, + 0.016012350097298622, + 0.007042701356112957, + 0.007717328146100044, + 0.006264751777052879, + 0.011895366944372654, + 0.007804821245372295, + 0.00612027570605278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.009795927442610264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305529221892357 + ], + "opposite_logprob_tokens": [ + -3.168055295944214 + ], + "logprob_reward": -0.04305529221892357, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305529221892357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17347075045108795, + 0.20020022988319397, + 0.08593542873859406, + 0.0527474544942379, + 0.13988269865512848, + 0.09960650652647018, + 0.056305501610040665, + 0.09058467298746109, + 0.05124013498425484, + 0.050026606768369675 + ], + "credit_map": [ + 0.024564646184444427, + 0.024890147149562836, + 0.014034144580364227, + 0.017574137076735497, + 0.022344041615724564, + 0.029342424124479294, + 0.018975764513015747, + 0.02174544706940651, + 0.018150784075260162, + 0.02547594904899597, + 0.017161333933472633, + 0.038207054138183594, + 0.02801108919084072, + 0.028589610010385513, + 0.028257766738533974, + 0.016346648335456848, + 0.008345826528966427, + 0.012104666791856289, + 0.007735534571111202, + 0.01737232878804207, + 0.011897271499037743, + 0.011227956041693687, + 0.01040471438318491, + 0.0068471385166049, + 0.0072433906607329845, + 0.007665577344596386, + 0.0034611925948411226, + 0.00600424874573946, + 0.00582886952906847, + 0.007121668662875891, + 0.00638689985498786, + 0.009035610593855381, + 0.015032720752060413, + 0.018023090437054634, + 0.008525627665221691, + 0.028437772765755653, + 0.015896916389465332, + 0.01887110248208046, + 0.023532871156930923, + 0.011562598869204521, + 0.00973526667803526, + 0.021256135776638985, + 0.0050947112031280994, + 0.010156789794564247, + 0.013899898156523705, + 0.02270699478685856, + 0.010876549407839775, + 0.005880165845155716, + 0.004931356757879257, + 0.006122824270278215, + 0.006665588356554508, + 0.007523911539465189, + 0.005676389671862125, + 0.0076876129023730755, + 0.011630772612988949, + 0.0060670506209135056, + 0.007422599475830793, + 0.011966890655457973, + 0.007089375983923674, + 0.010170956142246723, + 0.013045509345829487, + 0.011040866374969482, + 0.022666485980153084, + 0.00718199135735631, + 0.0054742516949772835, + 0.010763532482087612, + 0.003613291075453162, + 0.004125650506466627, + 0.006267985794693232, + 0.011311697773635387, + 0.006109955720603466, + 0.0035737729631364346, + 0.00561936479061842, + 0.011042903177440166, + 0.002998253796249628, + 0.007979234680533409, + 0.004993868991732597, + 0.007410969119518995, + 0.006062756292521954, + 0.0039192624390125275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.04305529221892357 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812945246696472 + ], + "opposite_logprob_tokens": [ + -1.406294584274292 + ], + "logprob_reward": -0.2812945246696472, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812945246696472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10383037477731705, + 0.23512840270996094, + 0.07836966216564178, + 0.07466616481542587, + 0.12726236879825592, + 0.04911134019494057, + 0.05557975172996521, + 0.05057797580957413, + 0.1532454490661621, + 0.07222843170166016 + ], + "credit_map": [ + 0.011457212269306183, + 0.015161571092903614, + 0.009023251011967659, + 0.008657203987240791, + 0.01786191202700138, + 0.020558377727866173, + 0.010379644110798836, + 0.01073120441287756, + 0.016771281138062477, + 0.02181243896484375, + 0.019325532019138336, + 0.052470799535512924, + 0.029638094827532768, + 0.023900823667645454, + 0.04861775040626526, + 0.022591684013605118, + 0.007333744317293167, + 0.009839723818004131, + 0.006503962446004152, + 0.020140940323472023, + 0.010352308861911297, + 0.008260882459580898, + 0.010835195891559124, + 0.005102905910462141, + 0.006727247033268213, + 0.010174177587032318, + 0.005251851864159107, + 0.015246579423546791, + 0.009883170947432518, + 0.011501166969537735, + 0.010787858627736568, + 0.0050941091030836105, + 0.011596638709306717, + 0.015318596735596657, + 0.00905144214630127, + 0.028196657076478004, + 0.017474183812737465, + 0.009362994693219662, + 0.027457011863589287, + 0.008804851211607456, + 0.0036202999763190746, + 0.007615491282194853, + 0.003595703514292836, + 0.011780069209635258, + 0.004979318007826805, + 0.006921959109604359, + 0.006572125945240259, + 0.004026374313980341, + 0.005082517396658659, + 0.00826380867511034, + 0.007492065895348787, + 0.0065490445122122765, + 0.0062173642218112946, + 0.008345155976712704, + 0.007018874399363995, + 0.006610918790102005, + 0.0057054986245930195, + 0.009398565627634525, + 0.0032141562551259995, + 0.004199264105409384, + 0.006975051015615463, + 0.008209318853914738, + 0.0070237331092357635, + 0.005852388683706522, + 0.014770470559597015, + 0.03095763921737671, + 0.00963485799729824, + 0.013514063321053982, + 0.017930161207914352, + 0.03416895866394043, + 0.02322060987353325, + 0.009048686362802982, + 0.010612483136355877, + 0.016146475449204445, + 0.006981921847909689, + 0.00638238713145256, + 0.007280400954186916, + 0.008463381789624691, + 0.009024360217154026, + 0.007337021641433239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.2812945246696472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow-motion shot from a horizontal angle, the camera pans with remarkable fluidity from left to right, capturing a quiet, modern, open-plan office. A beam of warm afternoon sunlight streams through the vast floor-to-ceiling windows, creating a distinct Tyndall effect in the air. Countless tiny dust particles dance slowly within the beam, creating the delicate atmosphere of light and shadow that is Makoto Shinkai's hallmark. As the camera focuses on a workstation screen made of translucent frosted glass, set in a thin silver metal frame, it suddenly, without warning, erupts in a series of lightning-like cracks, starting at its center. These cracks spread outward at a visible speed, and then the entire screen undergoes a violent, brittle collapse, sending thousands of fragments of varying sizes gracefully shattering through the air like enchanted crystal. In extreme slow motion, each fragment captures and refracts the afternoon sunlight, flashing with rainbow-like halos and dazzling starlight. They slowly drift and tumble in the air, intertwined with the dust in the light column, forming a silent and gorgeous visual ballet. The entire dramatic collapse process is fully recorded by this smooth single shot, and the picture is full of broken beauty and a quiet and melancholy atmosphere.", + "seed": 9287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_11/index_2179_seed_9287_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01242841719649732, + "vlm_logprob": -0.01242841719649732, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.01242841719649732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01223876141011715 + ], + "opposite_logprob_tokens": [ + -5.6372389793396 + ], + "logprob_reward": -0.01223876141011715, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.01223876141011715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07124468684196472, + 0.04963868483901024, + 0.04736320301890373, + 0.04402327537536621, + 0.06155328452587128, + 0.07766666263341904, + 0.3513629138469696, + 0.1614290326833725, + 0.07595865428447723, + 0.059759609401226044 + ], + "credit_map": [ + 0.0065283505246043205, + 0.008860649541020393, + 0.011216863058507442, + 0.011005324311554432, + 0.007161432411521673, + 0.006773593835532665, + 0.011994889937341213, + 0.007703574374318123, + 0.00449571805074811, + 0.007324988953769207, + 0.006305142771452665, + 0.009851795621216297, + 0.006355748511850834, + 0.0037537598982453346, + 0.005714507307857275, + 0.005837019998580217, + 0.00655945623293519, + 0.008639070205390453, + 0.003452985780313611, + 0.0056020114570856094, + 0.00942658819258213, + 0.004734206013381481, + 0.004706578329205513, + 0.004242302849888802, + 0.003547225845977664, + 0.005989882629364729, + 0.0034211978781968355, + 0.005489947274327278, + 0.00650846678763628, + 0.003453571582213044, + 0.008264966309070587, + 0.0073480126447975636, + 0.009280913509428501, + 0.009577232412993908, + 0.00590968644246459, + 0.007291918620467186, + 0.007347587961703539, + 0.005959826055914164, + 0.00791223905980587, + 0.008273876272141933, + 0.012401090003550053, + 0.021145857870578766, + 0.0062891049310564995, + 0.009809376671910286, + 0.009697148576378822, + 0.0059466236270964146, + 0.006458393298089504, + 0.005919068120419979, + 0.03069050796329975, + 0.10035857558250427, + 0.016988538205623627, + 0.12663233280181885, + 0.013474746607244015, + 0.0061822631396353245, + 0.031855400651693344, + 0.025180527940392494, + 0.024962319061160088, + 0.02613847889006138, + 0.019485188648104668, + 0.017130089923739433, + 0.0148472860455513, + 0.014092481695115566, + 0.01632431522011757, + 0.028448861092329025, + 0.010292572900652885, + 0.008175790309906006, + 0.009946522302925587, + 0.012755848467350006, + 0.009679478593170643, + 0.007605132181197405, + 0.009879221208393574, + 0.0076240855269134045, + 0.010534757748246193, + 0.009392818436026573, + 0.005618975032120943, + 0.006328895688056946, + 0.0038741088937968016, + 0.005118852946907282, + 0.006958737038075924, + 0.01193245779722929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.01223876141011715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025050232652574778 + ], + "opposite_logprob_tokens": [ + -6.002504825592041 + ], + "logprob_reward": -0.0025050232652574778, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025050232652574778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08967545628547668, + 0.050003781914711, + 0.08361940830945969, + 0.08041638880968094, + 0.06907542794942856, + 0.05577293410897255, + 0.1495577096939087, + 0.039526164531707764, + 0.1891801357269287, + 0.1931726187467575 + ], + "credit_map": [ + 0.007554108742624521, + 0.010848226957023144, + 0.012859698385000229, + 0.013400817289948463, + 0.01001562550663948, + 0.009876878000795841, + 0.01260307990014553, + 0.012517017312347889, + 0.004189764149487019, + 0.008028128184378147, + 0.0056161219254136086, + 0.00708366883918643, + 0.007487549912184477, + 0.0036838981322944164, + 0.007943484000861645, + 0.005971164908260107, + 0.007068475242704153, + 0.010749656707048416, + 0.005026216618716717, + 0.017318420112133026, + 0.01711040921509266, + 0.006506516598165035, + 0.010238218121230602, + 0.00960149522870779, + 0.010815063491463661, + 0.011499391868710518, + 0.006843581795692444, + 0.010514596477150917, + 0.009833564050495625, + 0.006399073638021946, + 0.014666293747723103, + 0.009844813495874405, + 0.007527452893555164, + 0.007673012558370829, + 0.006516534369438887, + 0.009104794822633266, + 0.00877055898308754, + 0.006655259057879448, + 0.010460908524692059, + 0.012366910465061665, + 0.006635453086346388, + 0.010377757251262665, + 0.004250286612659693, + 0.004941553343087435, + 0.007570791058242321, + 0.004654460586607456, + 0.010839083231985569, + 0.006503549870103598, + 0.01343019399791956, + 0.02482462301850319, + 0.009968933649361134, + 0.05646586790680885, + 0.010645966976881027, + 0.006014376413077116, + 0.01754491776227951, + 0.010662822984158993, + 0.007160336710512638, + 0.004367574583739042, + 0.004584068898111582, + 0.0037207521963864565, + 0.002868448384106159, + 0.002611047588288784, + 0.00737812789157033, + 0.00683581130579114, + 0.04039822891354561, + 0.026596350595355034, + 0.027379930019378662, + 0.019978731870651245, + 0.017152290791273117, + 0.014763719402253628, + 0.023915860801935196, + 0.01899503357708454, + 0.031893301755189896, + 0.03801030293107033, + 0.01622920297086239, + 0.015424525365233421, + 0.012602276168763638, + 0.014042947441339493, + 0.020267760381102562, + 0.044702302664518356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.0025050232652574778 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950473714619875 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950473714619875, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950473714619875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09519835561513901, + 0.04493438079953194, + 0.09681041538715363, + 0.045925263315439224, + 0.10108949989080429, + 0.06576107442378998, + 0.3913118839263916, + 0.05904329568147659, + 0.028896726667881012, + 0.07102908939123154 + ], + "credit_map": [ + 0.009161479771137238, + 0.012904899194836617, + 0.01277034543454647, + 0.015302736312150955, + 0.012546539306640625, + 0.009503704495728016, + 0.013659307733178139, + 0.009349342435598373, + 0.004860203247517347, + 0.006658459082245827, + 0.004569277632981539, + 0.005656526889652014, + 0.007963947020471096, + 0.0035638054832816124, + 0.006012927275151014, + 0.005649232771247625, + 0.008193223737180233, + 0.018205352127552032, + 0.007429091725498438, + 0.021182291209697723, + 0.01157770212739706, + 0.006272655911743641, + 0.012346629984676838, + 0.011603467166423798, + 0.005020791199058294, + 0.009373657405376434, + 0.004011937882751226, + 0.004684744402766228, + 0.006650018040090799, + 0.004075254313647747, + 0.005724577698856592, + 0.0063842786476016045, + 0.01989421620965004, + 0.022040659561753273, + 0.014748279005289078, + 0.0088631771504879, + 0.011221003718674183, + 0.007831505499780178, + 0.009965009056031704, + 0.006525649223476648, + 0.007583224214613438, + 0.007865392602980137, + 0.005570967681705952, + 0.008274617604911327, + 0.006817254703491926, + 0.006056440062820911, + 0.015787523239850998, + 0.007805649656802416, + 0.02560141310095787, + 0.09267126768827438, + 0.02212737873196602, + 0.15023942291736603, + 0.02254057489335537, + 0.011525839567184448, + 0.04145798832178116, + 0.025148015469312668, + 0.008088493719696999, + 0.009748993441462517, + 0.006923943758010864, + 0.006712041329592466, + 0.005267584230750799, + 0.005395019426941872, + 0.006194380111992359, + 0.010712837800383568, + 0.004971289541572332, + 0.005465389229357243, + 0.002716024173423648, + 0.004168297164142132, + 0.0029371886048465967, + 0.0017836791230365634, + 0.003587916260585189, + 0.003266941988840699, + 0.010889135301113129, + 0.015964357182383537, + 0.006999405566602945, + 0.00726731400936842, + 0.005408009514212608, + 0.004189271479845047, + 0.010209945030510426, + 0.01010164525359869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.005950473714619875 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018174543976783752 + ], + "opposite_logprob_tokens": [ + -4.018174648284912 + ], + "logprob_reward": -0.018174543976783752, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018174543976783752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09314610064029694, + 0.058449212461709976, + 0.10677655041217804, + 0.11839209496974945, + 0.04070424661040306, + 0.0755184069275856, + 0.31324443221092224, + 0.11604949086904526, + 0.03807934746146202, + 0.039640072733163834 + ], + "credit_map": [ + 0.008626931346952915, + 0.011065704748034477, + 0.01346791721880436, + 0.014775141142308712, + 0.009035799652338028, + 0.007760311011224985, + 0.014346947893500328, + 0.014067348092794418, + 0.005199988838285208, + 0.006603793706744909, + 0.006324068643152714, + 0.00953458622097969, + 0.007533749099820852, + 0.003758192528039217, + 0.009644263423979282, + 0.00985056720674038, + 0.01022245828062296, + 0.018452901393175125, + 0.00812577549368143, + 0.02175411954522133, + 0.013915497809648514, + 0.007834224961698055, + 0.01282672118395567, + 0.013644857332110405, + 0.00949398335069418, + 0.015942314639687538, + 0.013077713549137115, + 0.01990286074578762, + 0.012026950716972351, + 0.009730436839163303, + 0.018190499395132065, + 0.02002733014523983, + 0.005542890168726444, + 0.00583566864952445, + 0.004143692087382078, + 0.004454647656530142, + 0.0035191981587558985, + 0.00369557598605752, + 0.005932260304689407, + 0.007580310571938753, + 0.013177666813135147, + 0.019050566479563713, + 0.004949669353663921, + 0.0071993861347436905, + 0.009579265490174294, + 0.00788239948451519, + 0.007635957095772028, + 0.006043495610356331, + 0.030256018042564392, + 0.09610746800899506, + 0.014333670027554035, + 0.11078867316246033, + 0.01182274054735899, + 0.004781100433319807, + 0.025185830891132355, + 0.019968952983617783, + 0.02449929341673851, + 0.014899887144565582, + 0.01938799023628235, + 0.01104737725108862, + 0.008569947443902493, + 0.006852444261312485, + 0.016735030338168144, + 0.01405752170830965, + 0.007038452662527561, + 0.005955043714493513, + 0.005148699972778559, + 0.004310733638703823, + 0.0035870831925421953, + 0.00354974577203393, + 0.005031683016568422, + 0.00345790502615273, + 0.006745263002812862, + 0.008754595182836056, + 0.0031433820258826017, + 0.003830553265288472, + 0.005094602238386869, + 0.0030010920017957687, + 0.005139396525919437, + 0.003931188955903053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.018174543976783752 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a quiet, modern, open-plan office with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm afternoon sunlight visible streaming through the windows with a distinct Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a workstation screen made of translucent frosted glass in a thin silver metal frame appear and then erupt in lightning-like cracks starting from the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen collapse violently into thousands of fragments that shatter through the air in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot horizontal and does it pan smoothly from left to right throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327328361570835 + ], + "opposite_logprob_tokens": [ + -3.773273229598999 + ], + "logprob_reward": -0.02327328361570835, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327328361570835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12038711458444595, + 0.0438535213470459, + 0.07181043922901154, + 0.05955996364355087, + 0.08554109185934067, + 0.08740565180778503, + 0.3840236961841583, + 0.08943299204111099, + 0.044554706662893295, + 0.013430869206786156 + ], + "credit_map": [ + 0.014037766493856907, + 0.01740553416311741, + 0.01629369892179966, + 0.017083793878555298, + 0.016780458390712738, + 0.011364332400262356, + 0.013621130958199501, + 0.013800393790006638, + 0.0036490545608103275, + 0.0054425522685050964, + 0.005635335110127926, + 0.006382203195244074, + 0.003944643773138523, + 0.0034654990304261446, + 0.006599822081625462, + 0.008734408766031265, + 0.006240111775696278, + 0.008152355439960957, + 0.006216814275830984, + 0.01927877590060234, + 0.008195972070097923, + 0.0044782268814742565, + 0.010386006906628609, + 0.00886217225342989, + 0.007526971399784088, + 0.011334479786455631, + 0.003810245543718338, + 0.008446202613413334, + 0.0073408265598118305, + 0.005442756228148937, + 0.007027077954262495, + 0.008631397038698196, + 0.016851043328642845, + 0.015187840908765793, + 0.00834565982222557, + 0.008969612419605255, + 0.009336053393781185, + 0.007511272560805082, + 0.010638110339641571, + 0.008701493963599205, + 0.015445572324097157, + 0.018532732501626015, + 0.008068634197115898, + 0.011171134188771248, + 0.01089190412312746, + 0.008231550455093384, + 0.00800385233014822, + 0.00706026516854763, + 0.027797149494290352, + 0.10392795503139496, + 0.01978650875389576, + 0.14671550691127777, + 0.014764226973056793, + 0.006596319377422333, + 0.038242436945438385, + 0.02619357965886593, + 0.015977561473846436, + 0.015117185190320015, + 0.013186180032789707, + 0.00964963436126709, + 0.009093602187931538, + 0.007995322346687317, + 0.009505482390522957, + 0.008908017538487911, + 0.005242977757006884, + 0.011095352470874786, + 0.004603737965226173, + 0.006101096514612436, + 0.0031774977687746286, + 0.0022301310673356056, + 0.004513539839535952, + 0.007590367924422026, + 0.0022973224986344576, + 0.0033967546187341213, + 0.0008627544157207012, + 0.000919144309591502, + 0.0013254594523459673, + 0.0008610871736891568, + 0.0016765138134360313, + 0.002091833157464862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.02327328361570835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_12/index_tensor([3770, 3770])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_12/index_tensor([3770, 3770])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a4d51b18ddb4942b30708ee44051ea70f1a3d4aa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_12/index_tensor([3770, 3770])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level wide-angle shot captures the scene with the candid, unposed quality of an informal snapshot, its fixed perspective emphasizing the wild, acrobatic attack unfolding within the frame. A primitive, feline hunter, a powerfully built, humanoid figure with lean, roped-in muscles, tumbles and swirls through the dense jungle canopy. Wild, unkempt black hair flies as they move, adorned with weathered bone trophies: a necklace of tiny, yellowed skulls clinks across their chest, a shield fashioned from the jaws of a massive beast strapped to their shoulders. In a ferocious blur of movement, the hunters launch explosively from a dense, mossy branch, their bodies swirling and splaying wildly through the air, their expressions the mask of predation. They fluidly grasp dangling vines for a spin, continuing their unruly trajectory through a chaotic grid of gnarled branches and enormous, waxy leaves. Dappled sunlight filters through the leaves, creating a dynamic play of bright hotspots and deep, shifting shadows that streak and distort with the hunters’ rapid movements, creating a tense and visceral atmosphere of raw, untamed energy.", + "seed": 8789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_12/index_3770_seed_8789_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07548876712098718, + "vlm_logprob": -0.07548876712098718, + "vlm_yes_no_margin": 3.874999988079071, + "vlm_token_credit": -0.07548876712098718, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029191358014941216 + ], + "opposite_logprob_tokens": [ + -3.654191255569458 + ], + "logprob_reward": -0.029191358014941216, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029191358014941216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06936301290988922, + 0.13656927645206451, + 0.03275257349014282, + 0.12235073000192642, + 0.08421686291694641, + 0.1265600621700287, + 0.10894493013620377, + 0.08548980951309204, + 0.18130317330360413, + 0.05244956538081169 + ], + "credit_map": [ + 0.0041715847328305244, + 0.0059544178657233715, + 0.013891369104385376, + 0.007038023788481951, + 0.007184683345258236, + 0.004914066754281521, + 0.0062883286736905575, + 0.01992054097354412, + 0.008839109912514687, + 0.012708796188235283, + 0.013762149028480053, + 0.027027778327465057, + 0.01095258817076683, + 0.014285653829574585, + 0.02543335221707821, + 0.023559844121336937, + 0.003169713541865349, + 0.004827600903809071, + 0.004983818624168634, + 0.003965973854064941, + 0.005807843990623951, + 0.0030026135500520468, + 0.0036354800686240196, + 0.003359532682225108, + 0.00877348706126213, + 0.024113669991493225, + 0.014117354527115822, + 0.02514510042965412, + 0.009810053743422031, + 0.012284915894269943, + 0.01816195622086525, + 0.009944187477231026, + 0.00900076050311327, + 0.007427689619362354, + 0.007594681810587645, + 0.020627642050385475, + 0.009575476869940758, + 0.00798061117529869, + 0.012251432985067368, + 0.00975856278091669, + 0.012331049889326096, + 0.009419620037078857, + 0.008966667577624321, + 0.018773499876260757, + 0.02963646687567234, + 0.009248105809092522, + 0.026963530108332634, + 0.011221121996641159, + 0.008392869494855404, + 0.01032145693898201, + 0.022208506241440773, + 0.01640438102185726, + 0.009754934348165989, + 0.008861803449690342, + 0.01628073863685131, + 0.016720248386263847, + 0.008078471757471561, + 0.007127040531486273, + 0.010714552365243435, + 0.010319176129996777, + 0.01741642691195011, + 0.008689312264323235, + 0.011724412441253662, + 0.01142041478306055, + 0.010730866342782974, + 0.025622013956308365, + 0.02089080587029457, + 0.0258832685649395, + 0.02171211503446102, + 0.030418841168284416, + 0.03602510318160057, + 0.01002015545964241, + 0.004530294332653284, + 0.00669321371242404, + 0.006958758924156427, + 0.006212120410054922, + 0.005126116331666708, + 0.010850477032363415, + 0.004714000504463911, + 0.007364582270383835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.029191358014941216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314859718084335 + ], + "opposite_logprob_tokens": [ + -3.773314952850342 + ], + "logprob_reward": -0.023314859718084335, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314859718084335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10977586358785629, + 0.08058223873376846, + 0.05616973340511322, + 0.1123151034116745, + 0.1376030594110489, + 0.06861820071935654, + 0.1839461624622345, + 0.06582976877689362, + 0.07215768843889236, + 0.1130022257566452 + ], + "credit_map": [ + 0.007369110360741615, + 0.024683687835931778, + 0.015776071697473526, + 0.013403540477156639, + 0.010212930850684643, + 0.00862293504178524, + 0.0092699583619833, + 0.02043762430548668, + 0.006615994032472372, + 0.012589572928845882, + 0.008252439089119434, + 0.012228716164827347, + 0.007988722063601017, + 0.011401425115764141, + 0.011536401696503162, + 0.009968966245651245, + 0.0044517251662909985, + 0.009263256564736366, + 0.006822885014116764, + 0.00833488442003727, + 0.007610979489982128, + 0.0076147462241351604, + 0.005457309074699879, + 0.006613945588469505, + 0.009807449765503407, + 0.02037588506937027, + 0.013252422213554382, + 0.019185613840818405, + 0.01413414254784584, + 0.013387830927968025, + 0.013803289271891117, + 0.008368473500013351, + 0.01509061362594366, + 0.011364254169166088, + 0.010021170601248741, + 0.03811407834291458, + 0.014199822209775448, + 0.011606762185692787, + 0.018844567239284515, + 0.018361780792474747, + 0.006810768041759729, + 0.0076333098113536835, + 0.005909297149628401, + 0.011842187494039536, + 0.009338336065411568, + 0.007232438772916794, + 0.013869741931557655, + 0.0059821209870278835, + 0.015194492414593697, + 0.0066288430243730545, + 0.05640439689159393, + 0.011190660297870636, + 0.013280995190143585, + 0.01433798111975193, + 0.011918261647224426, + 0.054990530014038086, + 0.005254270508885384, + 0.005531289614737034, + 0.006956526078283787, + 0.009675320237874985, + 0.011088093742728233, + 0.0077231028117239475, + 0.009970695711672306, + 0.0096304751932621, + 0.005799947772175074, + 0.0092287203297019, + 0.010799107141792774, + 0.009627429768443108, + 0.010335328988730907, + 0.009531131945550442, + 0.011162440292537212, + 0.005673584528267384, + 0.009946241974830627, + 0.014700484462082386, + 0.014272450469434261, + 0.013097216375172138, + 0.011415977030992508, + 0.024682313203811646, + 0.0100795803591609, + 0.014807961881160736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.023314859718084335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784949012100697 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009784949012100697, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784949012100697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08752132207155228, + 0.10246530920267105, + 0.06993817538022995, + 0.10546590387821198, + 0.11784247308969498, + 0.15177051723003387, + 0.0690295398235321, + 0.14353184401988983, + 0.07354318350553513, + 0.07889169454574585 + ], + "credit_map": [ + 0.008088952861726284, + 0.009430814534425735, + 0.011986693367362022, + 0.011889697052538395, + 0.010241430252790451, + 0.01084459200501442, + 0.011517725884914398, + 0.013521412387490273, + 0.00723959319293499, + 0.013580996543169022, + 0.02396722324192524, + 0.011938505806028843, + 0.009470400400459766, + 0.011195026338100433, + 0.016415957361459732, + 0.008657604455947876, + 0.00731651159003377, + 0.009640004485845566, + 0.01228663232177496, + 0.009840057231485844, + 0.007520387414842844, + 0.007434995844960213, + 0.00782904401421547, + 0.008070546202361584, + 0.009320421144366264, + 0.021766221150755882, + 0.01235493365675211, + 0.019133681431412697, + 0.009164836257696152, + 0.012411970645189285, + 0.012000344693660736, + 0.009313488379120827, + 0.010523511096835136, + 0.0096435546875, + 0.010523350909352303, + 0.026592226698994637, + 0.01828586310148239, + 0.01034194603562355, + 0.020378226414322853, + 0.011553794145584106, + 0.015204020775854588, + 0.011281415820121765, + 0.010034500621259212, + 0.01784120500087738, + 0.031884197145700455, + 0.011052927002310753, + 0.04200257733464241, + 0.012469670735299587, + 0.006620543077588081, + 0.003704322502017021, + 0.016812216490507126, + 0.007646401412785053, + 0.007540395017713308, + 0.005303355399519205, + 0.008272165432572365, + 0.013130136765539646, + 0.013545745983719826, + 0.01870918460190296, + 0.014933214522898197, + 0.027354463934898376, + 0.01898580975830555, + 0.01746312342584133, + 0.018679533153772354, + 0.013860776089131832, + 0.0034748229663819075, + 0.008379519917070866, + 0.01102710235863924, + 0.012845426797866821, + 0.008780080825090408, + 0.011289665475487709, + 0.012599240057170391, + 0.005147326271981001, + 0.006263184826821089, + 0.011630061082541943, + 0.009342135861515999, + 0.01116977445781231, + 0.007711621001362801, + 0.009835850447416306, + 0.009181314148008823, + 0.01375774759799242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.009784949012100697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017120479606091976 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017120479606091976, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017120479606091976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13839086890220642, + 0.11020959913730621, + 0.09371983259916306, + 0.056694403290748596, + 0.06264650076627731, + 0.06009788066148758, + 0.1829446256160736, + 0.07018574327230453, + 0.10808034241199493, + 0.11703019589185715 + ], + "credit_map": [ + 0.013812342658638954, + 0.017296617850661278, + 0.016879957169294357, + 0.019613614305853844, + 0.020364712923765182, + 0.015769602730870247, + 0.016601132228970528, + 0.01805288903415203, + 0.006344716064631939, + 0.00923149473965168, + 0.03736472874879837, + 0.01562959887087345, + 0.006030041724443436, + 0.006520652677863836, + 0.018326828256249428, + 0.010761539451777935, + 0.009418784640729427, + 0.0184891689568758, + 0.013866414315998554, + 0.011959454976022243, + 0.01218689139932394, + 0.007618370000272989, + 0.010484558530151844, + 0.009696182794868946, + 0.004076783545315266, + 0.009630839340388775, + 0.006460166536271572, + 0.011154122650623322, + 0.008355852216482162, + 0.006427912972867489, + 0.0064033446833491325, + 0.004185383673757315, + 0.007123623974621296, + 0.00552158011123538, + 0.006303489673882723, + 0.011285125277936459, + 0.009121827781200409, + 0.005674674641340971, + 0.008089052513241768, + 0.009527127258479595, + 0.006735112052410841, + 0.006022993940860033, + 0.004258276894688606, + 0.01104812417179346, + 0.013946463353931904, + 0.004897636827081442, + 0.008473927155137062, + 0.004715349525213242, + 0.016811465844511986, + 0.009428875520825386, + 0.05527699738740921, + 0.011705287732183933, + 0.030257824808359146, + 0.014446881599724293, + 0.012711567804217339, + 0.03230573236942291, + 0.008106144145131111, + 0.007391786202788353, + 0.005038877949118614, + 0.014310996048152447, + 0.0095716817304492, + 0.007013486698269844, + 0.010866583324968815, + 0.007886188104748726, + 0.006011524703353643, + 0.01579710654914379, + 0.010317019186913967, + 0.01843106560409069, + 0.012689848430454731, + 0.015673790127038956, + 0.02052973210811615, + 0.008630253374576569, + 0.008748122490942478, + 0.01516417320817709, + 0.014543484896421432, + 0.018334440886974335, + 0.011952459812164307, + 0.016309702768921852, + 0.013261590152978897, + 0.018716219812631607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.0017120479606091976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344062089920044 + ], + "opposite_logprob_tokens": [ + -1.3134405612945557 + ], + "logprob_reward": -0.31344062089920044, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31344062089920044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08710210025310516, + 0.14621040225028992, + 0.06994390487670898, + 0.13122393190860748, + 0.11826971918344498, + 0.074603371322155, + 0.1450563669204712, + 0.07953611761331558, + 0.08962375670671463, + 0.05843038484454155 + ], + "credit_map": [ + 0.0048105246387422085, + 0.012809819541871548, + 0.018465619534254074, + 0.007153064012527466, + 0.01291636936366558, + 0.008576348423957825, + 0.00861090887337923, + 0.013759446330368519, + 0.008481458760797977, + 0.0167219340801239, + 0.032216690480709076, + 0.030821871012449265, + 0.009600970894098282, + 0.010539633221924305, + 0.023761171847581863, + 0.014066675677895546, + 0.0070760236121714115, + 0.013560297898948193, + 0.008569279685616493, + 0.009500513784587383, + 0.0075909411534667015, + 0.006713837385177612, + 0.008648424409329891, + 0.008284586481750011, + 0.01109496783465147, + 0.024198489263653755, + 0.01725432090461254, + 0.027229411527514458, + 0.009640815667808056, + 0.013551861047744751, + 0.01687382534146309, + 0.011380236595869064, + 0.01148905698210001, + 0.01012807060033083, + 0.009568987414240837, + 0.02497507631778717, + 0.021132666617631912, + 0.009725519455969334, + 0.015590349212288857, + 0.015659993514418602, + 0.007117485627532005, + 0.006745682097971439, + 0.005365265533328056, + 0.010458954609930515, + 0.017705535516142845, + 0.006607995368540287, + 0.014439104124903679, + 0.006163352634757757, + 0.011778941377997398, + 0.011675652116537094, + 0.03630400076508522, + 0.014130810275673866, + 0.012974818237125874, + 0.01087411679327488, + 0.011333051137626171, + 0.03598497062921524, + 0.006786026526242495, + 0.007041347678750753, + 0.00764301884919405, + 0.014247247949242592, + 0.010430066846311092, + 0.007079548668116331, + 0.010606760159134865, + 0.015702098608016968, + 0.005544112995266914, + 0.010784124955534935, + 0.012225267477333546, + 0.014518269337713718, + 0.012262368574738503, + 0.01330814603716135, + 0.013719788752496243, + 0.00726167531684041, + 0.005635854788124561, + 0.007937698625028133, + 0.007830263115465641, + 0.008313793689012527, + 0.005132886115461588, + 0.01017860509455204, + 0.0064527918584644794, + 0.006948490627110004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.31344062089920044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures the scene with the candid, unposed quality of an informal snapshot, its fixed perspective emphasizing the wild, acrobatic attack unfolding within the frame. A primitive, feline hunter, a powerfully built, humanoid figure with lean, roped-in muscles, tumbles and swirls through the dense jungle canopy. Wild, unkempt black hair flies as they move, adorned with weathered bone trophies: a necklace of tiny, yellowed skulls clinks across their chest, a shield fashioned from the jaws of a massive beast strapped to their shoulders. In a ferocious blur of movement, the hunters launch explosively from a dense, mossy branch, their bodies swirling and splaying wildly through the air, their expressions the mask of predation. They fluidly grasp dangling vines for a spin, continuing their unruly trajectory through a chaotic grid of gnarled branches and enormous, waxy leaves. Dappled sunlight filters through the leaves, creating a dynamic play of bright hotspots and deep, shifting shadows that streak and distort with the hunters’ rapid movements, creating a tense and visceral atmosphere of raw, untamed energy.", + "seed": 8790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_12/index_3770_seed_8790_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05910266770515591, + "vlm_logprob": -0.05910266770515591, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.05910266770515591, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0313204787671566 + ], + "opposite_logprob_tokens": [ + -3.531320571899414 + ], + "logprob_reward": -0.0313204787671566, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0313204787671566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07273265719413757, + 0.042791254818439484, + 0.046921756118535995, + 0.11415320634841919, + 0.13820385932922363, + 0.024258216843008995, + 0.09041111916303635, + 0.3091444671154022, + 0.0690072625875473, + 0.0923762172460556 + ], + "credit_map": [ + 0.005620268173515797, + 0.0059790234081447124, + 0.011317568831145763, + 0.016169417649507523, + 0.00785486213862896, + 0.004339598119258881, + 0.009436503984034061, + 0.01201541256159544, + 0.005192004144191742, + 0.004780180752277374, + 0.008302506990730762, + 0.006527376361191273, + 0.0025930169504135847, + 0.0045597143471241, + 0.0031422809697687626, + 0.007694173138588667, + 0.005705667659640312, + 0.008670232258737087, + 0.005206944886595011, + 0.007221240550279617, + 0.0033111642114818096, + 0.005225684028118849, + 0.007235828321427107, + 0.004344991408288479, + 0.00896750669926405, + 0.030661942437291145, + 0.025232477113604546, + 0.02063325233757496, + 0.006760904565453529, + 0.006103094667196274, + 0.008226790465414524, + 0.007567235734313726, + 0.0130686704069376, + 0.013179224915802479, + 0.018765386193990707, + 0.030519919469952583, + 0.02001185528934002, + 0.015110062435269356, + 0.014707941561937332, + 0.01284078136086464, + 0.0028446500655263662, + 0.0015597044257447124, + 0.003114471212029457, + 0.0032805074006319046, + 0.003654069732874632, + 0.0022675390355288982, + 0.0046119168400764465, + 0.0029253563843667507, + 0.008472989313304424, + 0.006800912786275148, + 0.012618349865078926, + 0.018852420151233673, + 0.009369328618049622, + 0.010738908313214779, + 0.012883731164038181, + 0.010674484074115753, + 0.03485589846968651, + 0.04419296607375145, + 0.040654994547367096, + 0.072859987616539, + 0.02946031093597412, + 0.025047991424798965, + 0.027365190908312798, + 0.03470709174871445, + 0.013650867156684399, + 0.013775778003036976, + 0.006035268772393465, + 0.015574319288134575, + 0.0033837161026895046, + 0.005256053060293198, + 0.00430601742118597, + 0.007025238126516342, + 0.015381796285510063, + 0.011053462512791157, + 0.009964749217033386, + 0.009099394083023071, + 0.00949760153889656, + 0.011219191364943981, + 0.012747358530759811, + 0.013412658125162125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.0313204787671566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370346546173096 + ], + "opposite_logprob_tokens": [ + -3.6513702869415283 + ], + "logprob_reward": -0.026370346546173096, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370346546173096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10435548424720764, + 0.08149110525846481, + 0.07639734447002411, + 0.13675178587436676, + 0.0867784172296524, + 0.07331304252147675, + 0.08095171302556992, + 0.11338265240192413, + 0.07101456820964813, + 0.17556391656398773 + ], + "credit_map": [ + 0.012737158685922623, + 0.0071412972174584866, + 0.022973326966166496, + 0.013364872895181179, + 0.01146471407264471, + 0.006640935782343149, + 0.012549054808914661, + 0.01748412474989891, + 0.0073150573298335075, + 0.012648822739720345, + 0.01461803913116455, + 0.01469231303781271, + 0.006135801784694195, + 0.006972128059715033, + 0.007838894613087177, + 0.011270047165453434, + 0.010338916443288326, + 0.017626214772462845, + 0.008101026527583599, + 0.009560528211295605, + 0.006027702707797289, + 0.008877652697265148, + 0.009063937701284885, + 0.006801365874707699, + 0.008976129814982414, + 0.04164062440395355, + 0.028815610334277153, + 0.02159998007118702, + 0.009337141178548336, + 0.005063233897089958, + 0.011945908889174461, + 0.00937314797192812, + 0.00830148532986641, + 0.016355987638235092, + 0.010825994424521923, + 0.012510483153164387, + 0.011180492118000984, + 0.0083164619281888, + 0.009887905791401863, + 0.009399603120982647, + 0.009098878130316734, + 0.005418580491095781, + 0.011182313784956932, + 0.009454041719436646, + 0.010088655166327953, + 0.005556062329560518, + 0.013097771443426609, + 0.009416740387678146, + 0.007228533737361431, + 0.006121320184320211, + 0.0088060162961483, + 0.01880486309528351, + 0.007675113622099161, + 0.008091443218290806, + 0.016143329441547394, + 0.008081093430519104, + 0.014292528852820396, + 0.016565395519137383, + 0.014432232826948166, + 0.018452759832143784, + 0.010380524210631847, + 0.012439589947462082, + 0.014717791229486465, + 0.012101829051971436, + 0.01524066086858511, + 0.012866147793829441, + 0.007526993285864592, + 0.01240425556898117, + 0.004637622740119696, + 0.006750980857759714, + 0.005614652298390865, + 0.005973250605165958, + 0.027534665539860725, + 0.022540058940649033, + 0.013264470733702183, + 0.018555674701929092, + 0.026606284081935883, + 0.01668650098145008, + 0.023133302107453346, + 0.02724294923245907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.026370346546173096 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076156049966812 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011076156049966812, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076156049966812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13566851615905762, + 0.13109026849269867, + 0.09872714430093765, + 0.10043451189994812, + 0.07806719094514847, + 0.038457661867141724, + 0.17168660461902618, + 0.08415930718183517, + 0.0809658020734787, + 0.08074301481246948 + ], + "credit_map": [ + 0.019837290048599243, + 0.009089571423828602, + 0.03536447137594223, + 0.019248362630605698, + 0.007948501966893673, + 0.0062606059946119785, + 0.01359013095498085, + 0.02432958036661148, + 0.009352872148156166, + 0.013881462626159191, + 0.03752661496400833, + 0.012497158721089363, + 0.01010701060295105, + 0.00870844628661871, + 0.006990713533014059, + 0.032025981694459915, + 0.012000245973467827, + 0.021895848214626312, + 0.010009230114519596, + 0.013039655983448029, + 0.0055679576471447945, + 0.008157294243574142, + 0.009117750450968742, + 0.018939169123768806, + 0.007504624780267477, + 0.023047499358654022, + 0.013079185970127583, + 0.021940650418400764, + 0.007883097045123577, + 0.007191738113760948, + 0.008688314817845821, + 0.011099402792751789, + 0.00860658660531044, + 0.010730726644396782, + 0.012042917311191559, + 0.015107895247638226, + 0.00835323240607977, + 0.007710454054176807, + 0.008111441507935524, + 0.007403937168419361, + 0.0033589135855436325, + 0.0029382803477346897, + 0.00635643070563674, + 0.007238212041556835, + 0.004381509032100439, + 0.003107355674728751, + 0.005965117830783129, + 0.00511183962225914, + 0.01856510527431965, + 0.013435001485049725, + 0.02309080772101879, + 0.03464609012007713, + 0.018734468147158623, + 0.014573519118130207, + 0.026929475367069244, + 0.02171213924884796, + 0.008394486270844936, + 0.00806492194533348, + 0.010857721790671349, + 0.016263611614704132, + 0.010419112630188465, + 0.01181544829159975, + 0.008630067110061646, + 0.009713941253721714, + 0.014992870390415192, + 0.016062632203102112, + 0.006458563264459372, + 0.018191948533058167, + 0.004329105373471975, + 0.006937094964087009, + 0.0057760379277169704, + 0.008217556402087212, + 0.010586254298686981, + 0.013682885095477104, + 0.010339326225221157, + 0.00876772403717041, + 0.007155939936637878, + 0.00868492666631937, + 0.011218277737498283, + 0.010307678021490574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.011076156049966812 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177690108306706 + ], + "opposite_logprob_tokens": [ + -6.751177787780762 + ], + "logprob_reward": -0.001177690108306706, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177690108306706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16629089415073395, + 0.043064482510089874, + 0.0640576183795929, + 0.06177607551217079, + 0.07791236788034439, + 0.0767841786146164, + 0.2327597588300705, + 0.10592260956764221, + 0.06513837724924088, + 0.10629360377788544 + ], + "credit_map": [ + 0.017468631267547607, + 0.009174993261694908, + 0.034654151648283005, + 0.017315801233053207, + 0.01497067604213953, + 0.011122629977762699, + 0.03216691315174103, + 0.029417116194963455, + 0.004648801404982805, + 0.004792288411408663, + 0.008390862494707108, + 0.006622734945267439, + 0.0034996678587049246, + 0.005448104348033667, + 0.0033746520057320595, + 0.006287374999374151, + 0.010498929768800735, + 0.00866176187992096, + 0.006850583478808403, + 0.008214359171688557, + 0.0037930402904748917, + 0.005249686073511839, + 0.006588340736925602, + 0.014200919307768345, + 0.004866533447057009, + 0.016626453027129173, + 0.007879415526986122, + 0.010812501423060894, + 0.0054030632600188255, + 0.005120395217090845, + 0.0056085423566401005, + 0.005459177307784557, + 0.007897217757999897, + 0.00765575934201479, + 0.011696359142661095, + 0.016739586368203163, + 0.009130698628723621, + 0.007938212715089321, + 0.008186831139028072, + 0.008667712099850178, + 0.007788268383592367, + 0.0052251266315579414, + 0.010701363906264305, + 0.010365774855017662, + 0.01271833200007677, + 0.00871254038065672, + 0.011211210861802101, + 0.01006157137453556, + 0.02204813063144684, + 0.01684686541557312, + 0.026631992310285568, + 0.051537711173295975, + 0.02637508511543274, + 0.02881462126970291, + 0.03656204789876938, + 0.02394331991672516, + 0.008944902569055557, + 0.012638295069336891, + 0.011121190153062344, + 0.020560309290885925, + 0.02048085816204548, + 0.008507685735821724, + 0.012765749357640743, + 0.010903628543019295, + 0.013850775547325611, + 0.006137016229331493, + 0.0057803490199148655, + 0.014023760333657265, + 0.004569727927446365, + 0.006185164675116539, + 0.006075579207390547, + 0.008516008965671062, + 0.015363116748631, + 0.021200483664870262, + 0.011595821008086205, + 0.012236984446644783, + 0.008927959017455578, + 0.008260845206677914, + 0.012014500796794891, + 0.01669391058385372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.001177690108306706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556866705417633 + ], + "opposite_logprob_tokens": [ + -1.6005686521530151 + ], + "logprob_reward": -0.22556866705417633, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556866705417633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07966354489326477, + 0.11967098712921143, + 0.06723224371671677, + 0.09443492442369461, + 0.1066511869430542, + 0.10965539515018463, + 0.1633671373128891, + 0.07152358442544937, + 0.0491831973195076, + 0.13861778378486633 + ], + "credit_map": [ + 0.007673497777432203, + 0.005828354973345995, + 0.011209413409233093, + 0.008472035638988018, + 0.009925005957484245, + 0.006210309453308582, + 0.01665615662932396, + 0.01368876826018095, + 0.011194786056876183, + 0.013315717689692974, + 0.028818191960453987, + 0.01629646122455597, + 0.006856278982013464, + 0.01165177021175623, + 0.0067066471092402935, + 0.024831144139170647, + 0.009662212803959846, + 0.012879230082035065, + 0.00847750622779131, + 0.011839215643703938, + 0.0036976896226406097, + 0.005070266779512167, + 0.006691249553114176, + 0.008914872072637081, + 0.00610013585537672, + 0.02589336223900318, + 0.027419231832027435, + 0.01196693442761898, + 0.005153401754796505, + 0.0038287045899778605, + 0.0074140639044344425, + 0.006659093778580427, + 0.009740781970322132, + 0.01377852726727724, + 0.012632818892598152, + 0.01661924459040165, + 0.010263876989483833, + 0.01772185042500496, + 0.011658529751002789, + 0.01423555426299572, + 0.011470085009932518, + 0.006634320132434368, + 0.015702126547694206, + 0.015241875313222408, + 0.019078794866800308, + 0.010161920450627804, + 0.017744462937116623, + 0.013621807098388672, + 0.017102262005209923, + 0.011938512325286865, + 0.015342136844992638, + 0.043795015662908554, + 0.013423364609479904, + 0.01592732034623623, + 0.02410103753209114, + 0.02173748053610325, + 0.00866386853158474, + 0.0074881697073578835, + 0.007719927933067083, + 0.01424303837120533, + 0.012358875945210457, + 0.005804102402180433, + 0.007275339215993881, + 0.007970261387526989, + 0.008582043461501598, + 0.007162205409258604, + 0.004298053681850433, + 0.007417643442749977, + 0.0036425874568521976, + 0.004987954627722502, + 0.0040186564438045025, + 0.009074053727090359, + 0.01718433015048504, + 0.026357097551226616, + 0.020437220111489296, + 0.023598872125148773, + 0.011931500397622585, + 0.011999262496829033, + 0.012792114168405533, + 0.014317388646304607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.22556866705417633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures the scene with the candid, unposed quality of an informal snapshot, its fixed perspective emphasizing the wild, acrobatic attack unfolding within the frame. A primitive, feline hunter, a powerfully built, humanoid figure with lean, roped-in muscles, tumbles and swirls through the dense jungle canopy. Wild, unkempt black hair flies as they move, adorned with weathered bone trophies: a necklace of tiny, yellowed skulls clinks across their chest, a shield fashioned from the jaws of a massive beast strapped to their shoulders. In a ferocious blur of movement, the hunters launch explosively from a dense, mossy branch, their bodies swirling and splaying wildly through the air, their expressions the mask of predation. They fluidly grasp dangling vines for a spin, continuing their unruly trajectory through a chaotic grid of gnarled branches and enormous, waxy leaves. Dappled sunlight filters through the leaves, creating a dynamic play of bright hotspots and deep, shifting shadows that streak and distort with the hunters’ rapid movements, creating a tense and visceral atmosphere of raw, untamed energy.", + "seed": 8791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_12/index_3770_seed_8791_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12203897666186095, + "vlm_logprob": -0.12203897666186095, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.12203897666186095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05114620923995972 + ], + "opposite_logprob_tokens": [ + -3.0511462688446045 + ], + "logprob_reward": -0.05114620923995972, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05114620923995972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1411856710910797, + 0.08000101149082184, + 0.052405569702386856, + 0.15984070301055908, + 0.12109638750553131, + 0.0473153181374073, + 0.266787588596344, + 0.037725210189819336, + 0.057404592633247375, + 0.03623794764280319 + ], + "credit_map": [ + 0.01262888964265585, + 0.014580823481082916, + 0.012818822637200356, + 0.022868040949106216, + 0.024022722616791725, + 0.01868652179837227, + 0.01774255931377411, + 0.017837291583418846, + 0.006756181828677654, + 0.007942129857838154, + 0.009287475608289242, + 0.009672896936535835, + 0.012463390827178955, + 0.012307053431868553, + 0.011705752462148666, + 0.009866122156381607, + 0.00352274882607162, + 0.005522471386939287, + 0.00958013441413641, + 0.01036616787314415, + 0.005061066243797541, + 0.005527146626263857, + 0.006532169878482819, + 0.006293665152043104, + 0.012346888892352581, + 0.017151212319731712, + 0.019198931753635406, + 0.01553307380527258, + 0.0293169803917408, + 0.017784753814339638, + 0.0319063737988472, + 0.016602477058768272, + 0.009157449938356876, + 0.009952540509402752, + 0.0324239656329155, + 0.012848182581365108, + 0.0068596480414271355, + 0.028372209519147873, + 0.009594383649528027, + 0.011887995526194572, + 0.004271949175745249, + 0.006568798329681158, + 0.005474220495671034, + 0.007070990279316902, + 0.007867819629609585, + 0.006202156189829111, + 0.005398072302341461, + 0.004461315460503101, + 0.018162798136472702, + 0.01464812457561493, + 0.060613278299570084, + 0.01947702094912529, + 0.023774614557623863, + 0.037238504737615585, + 0.04378638416528702, + 0.04908687248826027, + 0.0031785995233803988, + 0.005005871411412954, + 0.004382092040032148, + 0.008284038864076138, + 0.0039175949059426785, + 0.006201022770255804, + 0.004202006384730339, + 0.0025539810303598642, + 0.006041870918124914, + 0.006121054291725159, + 0.008886034600436687, + 0.009677151218056679, + 0.004296823404729366, + 0.009333010762929916, + 0.0068109165877103806, + 0.006237733643501997, + 0.003650803118944168, + 0.0036622467450797558, + 0.0037066403310745955, + 0.0056396108120679855, + 0.0033822250552475452, + 0.004943441599607468, + 0.005375949200242758, + 0.0058770314790308475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.05114620923995972 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026375338435173035 + ], + "opposite_logprob_tokens": [ + -3.6513752937316895 + ], + "logprob_reward": -0.026375338435173035, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026375338435173035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07528722286224365, + 0.07790803909301758, + 0.09088458865880966, + 0.0799805223941803, + 0.1005038246512413, + 0.058796659111976624, + 0.21908733248710632, + 0.0534219928085804, + 0.2089497447013855, + 0.035180069506168365 + ], + "credit_map": [ + 0.008489279076457024, + 0.008516943082213402, + 0.009250270202755928, + 0.012984356842935085, + 0.010094253346323967, + 0.00931607186794281, + 0.00808162335306406, + 0.008554423227906227, + 0.006576693616807461, + 0.009117786772549152, + 0.010135907679796219, + 0.009853975847363472, + 0.012973221018910408, + 0.010743044316768646, + 0.009028881788253784, + 0.009478527121245861, + 0.007197828497737646, + 0.009558405727148056, + 0.015741176903247833, + 0.016880840063095093, + 0.00912833958864212, + 0.008848459459841251, + 0.014414017088711262, + 0.009115519933402538, + 0.01044897548854351, + 0.007477855775505304, + 0.01129197422415018, + 0.010428536683321, + 0.005544155836105347, + 0.008425622247159481, + 0.009614573791623116, + 0.016748832538723946, + 0.007543929386883974, + 0.00945806410163641, + 0.013951178640127182, + 0.009244571439921856, + 0.0074250418692827225, + 0.025790566578507423, + 0.008011369965970516, + 0.01907910592854023, + 0.005696083884686232, + 0.007807957008481026, + 0.005976004060357809, + 0.011324958875775337, + 0.009356549009680748, + 0.006794939283281565, + 0.006628510542213917, + 0.005211659241467714, + 0.018342595547437668, + 0.01666221208870411, + 0.06669365614652634, + 0.015893734991550446, + 0.014083878137171268, + 0.02088521048426628, + 0.01943274214863777, + 0.047093309462070465, + 0.0038309323135763407, + 0.0071077970787882805, + 0.00716523127630353, + 0.00964803621172905, + 0.004790030419826508, + 0.008851726539433002, + 0.007061990909278393, + 0.004966247361153364, + 0.01905936747789383, + 0.019377050921320915, + 0.030199507251381874, + 0.04751922935247421, + 0.015531633980572224, + 0.03222129866480827, + 0.024514101445674896, + 0.02052754908800125, + 0.00311490916647017, + 0.0035447287373244762, + 0.004256613552570343, + 0.004955539014190435, + 0.0028184622060507536, + 0.004220023285597563, + 0.00510143768042326, + 0.007168359588831663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.026375338435173035 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11899454891681671, + 0.10706180334091187, + 0.11366719007492065, + 0.1085236668586731, + 0.10306806117296219, + 0.05979544296860695, + 0.21670524775981903, + 0.06197301670908928, + 0.07406378537416458, + 0.03614727035164833 + ], + "credit_map": [ + 0.011639051139354706, + 0.013904638588428497, + 0.012525792233645916, + 0.019733263179659843, + 0.019461272284388542, + 0.014768479391932487, + 0.01437039952725172, + 0.012591639533638954, + 0.00699181342497468, + 0.01369202695786953, + 0.014519348740577698, + 0.014859724789857864, + 0.013742725364863873, + 0.017940659075975418, + 0.012455616146326065, + 0.012859886512160301, + 0.007869777269661427, + 0.007568515837192535, + 0.027098698541522026, + 0.02622676081955433, + 0.01121173333376646, + 0.009895572438836098, + 0.013772249221801758, + 0.010023877024650574, + 0.009824853390455246, + 0.008046996779739857, + 0.012495853938162327, + 0.01378918532282114, + 0.01404385082423687, + 0.011011666618287563, + 0.01580836810171604, + 0.023502890020608902, + 0.008924207650125027, + 0.0088772252202034, + 0.008800627663731575, + 0.01944628171622753, + 0.008715387433767319, + 0.014333381317555904, + 0.008618154563009739, + 0.02535279467701912, + 0.00482231704518199, + 0.00847261119633913, + 0.007194644771516323, + 0.00837075337767601, + 0.010051791556179523, + 0.00721316784620285, + 0.008577966131269932, + 0.005092189181596041, + 0.01309921219944954, + 0.01744561456143856, + 0.05866234749555588, + 0.011166323907673359, + 0.028238888829946518, + 0.01935519278049469, + 0.025190383195877075, + 0.04354725778102875, + 0.00496088108047843, + 0.008618200197815895, + 0.007908857427537441, + 0.008773400448262691, + 0.00631861062720418, + 0.010493128560483456, + 0.00865283515304327, + 0.0062470934353768826, + 0.00730079784989357, + 0.009324515238404274, + 0.012424447573721409, + 0.011658388189971447, + 0.006769934203475714, + 0.010723995976150036, + 0.00934164971113205, + 0.006520054303109646, + 0.002954046241939068, + 0.003604981815442443, + 0.00362136191688478, + 0.003669226774945855, + 0.0030252388678491116, + 0.008474192582070827, + 0.0029683622997254133, + 0.007829858921468258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.007631433662027121 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017120479606091976 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017120479606091976, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017120479606091976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13636690378189087, + 0.08362074941396713, + 0.06270672380924225, + 0.07382647693157196, + 0.09010209888219833, + 0.04057056084275246, + 0.16728532314300537, + 0.06886329501867294, + 0.20925846695899963, + 0.06739941239356995 + ], + "credit_map": [ + 0.012987752445042133, + 0.014973120763897896, + 0.01288739126175642, + 0.020238297060132027, + 0.02361534908413887, + 0.016977254301309586, + 0.01635337993502617, + 0.01833435334265232, + 0.007621895056217909, + 0.0114369448274374, + 0.011040321551263332, + 0.01382838562130928, + 0.010269620455801487, + 0.011514862067997456, + 0.008426676504313946, + 0.009482042863965034, + 0.005007956176996231, + 0.008655814453959465, + 0.00980610866099596, + 0.010108108632266521, + 0.006909891031682491, + 0.009057891555130482, + 0.008642982691526413, + 0.0045179747976362705, + 0.008031298406422138, + 0.00788066629320383, + 0.007958739064633846, + 0.010462344624102116, + 0.006315745413303375, + 0.008141119964420795, + 0.009985508397221565, + 0.01505105197429657, + 0.014859722927212715, + 0.010565923526883125, + 0.024509916082024574, + 0.00891731958836317, + 0.00488476874306798, + 0.012054148130118847, + 0.00633565429598093, + 0.007974646985530853, + 0.0032787483651190996, + 0.006580401211977005, + 0.004619219806045294, + 0.005967557895928621, + 0.005741121247410774, + 0.00694976095110178, + 0.003980448003858328, + 0.003453303361311555, + 0.01220965851098299, + 0.01711183972656727, + 0.040975421667099, + 0.014592024497687817, + 0.01241418905556202, + 0.01696786656975746, + 0.02391398698091507, + 0.02910032868385315, + 0.006804357748478651, + 0.01265790406614542, + 0.005934067070484161, + 0.008155443705618382, + 0.01186232827603817, + 0.008189652115106583, + 0.007646107114851475, + 0.007613431196659803, + 0.02195602096617222, + 0.025390125811100006, + 0.037722960114479065, + 0.02237984910607338, + 0.015585314482450485, + 0.03545564040541649, + 0.024208545684814453, + 0.026560010388493538, + 0.006513427942991257, + 0.008378262631595135, + 0.006786596029996872, + 0.007325996644794941, + 0.005544138606637716, + 0.013470654375851154, + 0.007812134455889463, + 0.011568204499781132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.0017120479606091976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233298540115356 + ], + "opposite_logprob_tokens": [ + -0.8983298540115356 + ], + "logprob_reward": -0.5233298540115356, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233298540115356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07043049484491348, + 0.05410405620932579, + 0.09414605796337128, + 0.04982009530067444, + 0.11957155168056488, + 0.07609192281961441, + 0.13576772809028625, + 0.21251267194747925, + 0.13130615651607513, + 0.056249331682920456 + ], + "credit_map": [ + 0.006732565816491842, + 0.007529000286012888, + 0.007426859810948372, + 0.01161898486316204, + 0.00767767708748579, + 0.014106344431638718, + 0.0067855301313102245, + 0.008553517051041126, + 0.004985290113836527, + 0.006790248677134514, + 0.007325654849410057, + 0.008019999600946903, + 0.005317550618201494, + 0.008817224763333797, + 0.00551763828843832, + 0.0073304446414113045, + 0.005991135723888874, + 0.007787984795868397, + 0.022713178768754005, + 0.021614985540509224, + 0.008371607400476933, + 0.009138358756899834, + 0.011099564842879772, + 0.007429235614836216, + 0.0037218283396214247, + 0.00504818931221962, + 0.006808558013290167, + 0.005662490148097277, + 0.007285987492650747, + 0.004641824401915073, + 0.008997201919555664, + 0.007654006127268076, + 0.008259308524429798, + 0.0061563048511743546, + 0.020706191658973694, + 0.01567378081381321, + 0.005810390692204237, + 0.024528682231903076, + 0.008736854419112206, + 0.02970002219080925, + 0.005824805703014135, + 0.008277621120214462, + 0.007816661149263382, + 0.017238358035683632, + 0.011152759194374084, + 0.010750900954008102, + 0.007988229393959045, + 0.007042573764920235, + 0.008815251290798187, + 0.012672396376729012, + 0.03356855735182762, + 0.01230792235583067, + 0.010520530864596367, + 0.020508328452706337, + 0.01317238062620163, + 0.02420235052704811, + 0.017934169620275497, + 0.020214049145579338, + 0.024560052901506424, + 0.029936594888567924, + 0.026685630902647972, + 0.039381302893161774, + 0.028201503679156303, + 0.025599336251616478, + 0.011175289750099182, + 0.013445843942463398, + 0.016755007207393646, + 0.031704582273960114, + 0.008946560323238373, + 0.020258966833353043, + 0.014440164901316166, + 0.014579707756638527, + 0.005716565530747175, + 0.006617871578782797, + 0.005418816581368446, + 0.006851489655673504, + 0.006803950294852257, + 0.010049430653452873, + 0.007653435692191124, + 0.0071377623826265335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.5233298540115356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures the scene with the candid, unposed quality of an informal snapshot, its fixed perspective emphasizing the wild, acrobatic attack unfolding within the frame. A primitive, feline hunter, a powerfully built, humanoid figure with lean, roped-in muscles, tumbles and swirls through the dense jungle canopy. Wild, unkempt black hair flies as they move, adorned with weathered bone trophies: a necklace of tiny, yellowed skulls clinks across their chest, a shield fashioned from the jaws of a massive beast strapped to their shoulders. In a ferocious blur of movement, the hunters launch explosively from a dense, mossy branch, their bodies swirling and splaying wildly through the air, their expressions the mask of predation. They fluidly grasp dangling vines for a spin, continuing their unruly trajectory through a chaotic grid of gnarled branches and enormous, waxy leaves. Dappled sunlight filters through the leaves, creating a dynamic play of bright hotspots and deep, shifting shadows that streak and distort with the hunters’ rapid movements, creating a tense and visceral atmosphere of raw, untamed energy.", + "seed": 8792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_12/index_3770_seed_8792_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06406353230122477, + "vlm_logprob": -0.06406353230122477, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.06406353230122477, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040050897747278214 + ], + "opposite_logprob_tokens": [ + -3.290050983428955 + ], + "logprob_reward": -0.040050897747278214, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.040050897747278214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12110798805952072, + 0.09126201272010803, + 0.0642552524805069, + 0.126065194606781, + 0.16733597218990326, + 0.15244117379188538, + 0.07331206649541855, + 0.10211101919412613, + 0.030168695375323296, + 0.07194069027900696 + ], + "credit_map": [ + 0.00810195691883564, + 0.016519421711564064, + 0.012895160354673862, + 0.006769078783690929, + 0.007666419260203838, + 0.009464642964303493, + 0.019868966192007065, + 0.03982232138514519, + 0.007698774337768555, + 0.010253937914967537, + 0.018264714628458023, + 0.012089736759662628, + 0.009153047576546669, + 0.011135412380099297, + 0.008897853083908558, + 0.013768523000180721, + 0.007349540013819933, + 0.006081493105739355, + 0.008168675005435944, + 0.007545267231762409, + 0.0052010491490364075, + 0.014925000257790089, + 0.008018712513148785, + 0.006965510547161102, + 0.019462520256638527, + 0.016897542402148247, + 0.023121634498238564, + 0.010952343232929707, + 0.013321558944880962, + 0.01063684280961752, + 0.017985550686717033, + 0.013687202706933022, + 0.02128535695374012, + 0.02086593210697174, + 0.017312806099653244, + 0.02317957766354084, + 0.026932958513498306, + 0.02331169880926609, + 0.017603911459445953, + 0.01684371940791607, + 0.011681102216243744, + 0.032790154218673706, + 0.01810475066304207, + 0.02961195819079876, + 0.013669874519109726, + 0.013442113995552063, + 0.01686445251107216, + 0.016276752576231956, + 0.006459006108343601, + 0.005443614441901445, + 0.018580274656414986, + 0.006079638842493296, + 0.004006196279078722, + 0.004734039306640625, + 0.006325946189463139, + 0.021683339029550552, + 0.01359913032501936, + 0.034198611974716187, + 0.00829198770225048, + 0.01909761317074299, + 0.006401309743523598, + 0.005914703477174044, + 0.007266643922775984, + 0.007341016549617052, + 0.0030307830311357975, + 0.004845479968935251, + 0.0033046749886125326, + 0.0039031021296977997, + 0.0031138123013079166, + 0.004599018022418022, + 0.002560689812526107, + 0.0048111299984157085, + 0.008376912213861942, + 0.007240639068186283, + 0.007780900225043297, + 0.007636398542672396, + 0.007717603351920843, + 0.009336739778518677, + 0.007986856624484062, + 0.015864631161093712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.040050897747278214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820732094347477 + ], + "opposite_logprob_tokens": [ + -4.018207550048828 + ], + "logprob_reward": -0.01820732094347477, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820732094347477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08755861222743988, + 0.0983445942401886, + 0.09784268587827682, + 0.25980088114738464, + 0.13145308196544647, + 0.07864668220281601, + 0.1079777404665947, + 0.06368126720190048, + 0.03627125173807144, + 0.03842323645949364 + ], + "credit_map": [ + 0.005915598012506962, + 0.01378058921545744, + 0.007225001696497202, + 0.005384872201830149, + 0.007339872885495424, + 0.007321882527321577, + 0.0131246168166399, + 0.02746618166565895, + 0.00904424861073494, + 0.009655412286520004, + 0.017335934564471245, + 0.017102135345339775, + 0.010273242369294167, + 0.009972394444048405, + 0.012378356419503689, + 0.01258287113159895, + 0.010462460108101368, + 0.008653218857944012, + 0.011021850630640984, + 0.012228835374116898, + 0.007780073210597038, + 0.024887870997190475, + 0.011886450462043285, + 0.010921924374997616, + 0.07135136425495148, + 0.026655547320842743, + 0.0753554031252861, + 0.02038000151515007, + 0.016183363273739815, + 0.01890897937119007, + 0.019005171954631805, + 0.01196107268333435, + 0.01288985088467598, + 0.016564568504691124, + 0.011593901552259922, + 0.02357614040374756, + 0.01906842552125454, + 0.018944263458251953, + 0.016510363668203354, + 0.012305558659136295, + 0.007363980170339346, + 0.014819075353443623, + 0.010426375083625317, + 0.013575074262917042, + 0.007777459919452667, + 0.007975080981850624, + 0.008995525538921356, + 0.0077141099609434605, + 0.009415884502232075, + 0.009651951491832733, + 0.027766382321715355, + 0.009506741538643837, + 0.005614123307168484, + 0.006722948048263788, + 0.009134429506957531, + 0.030165281146764755, + 0.01028375793248415, + 0.01833432726562023, + 0.0064815036021173, + 0.008713428862392902, + 0.004323406610637903, + 0.004045890644192696, + 0.005496022757142782, + 0.00600292906165123, + 0.003933032043278217, + 0.004924483131617308, + 0.00446027796715498, + 0.003899306757375598, + 0.0042980932630598545, + 0.005170844960957766, + 0.0035052511375397444, + 0.0060799615457654, + 0.002907970454543829, + 0.004130163695663214, + 0.0035777282901108265, + 0.004048808012157679, + 0.008671246469020844, + 0.004752088338136673, + 0.004513325169682503, + 0.005821902304887772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.01820732094347477 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864111352711916 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.00864111352711916, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864111352711916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059188053011894226, + 0.07613909989595413, + 0.0694814920425415, + 0.3918992578983307, + 0.07697582244873047, + 0.08749201893806458, + 0.0492590069770813, + 0.05840146914124489, + 0.03986983746290207, + 0.09129398316144943 + ], + "credit_map": [ + 0.005519572179764509, + 0.006121355574578047, + 0.006347126327455044, + 0.004657937213778496, + 0.006758936680853367, + 0.009843353182077408, + 0.010419025085866451, + 0.00952074583619833, + 0.005726655479520559, + 0.005597658455371857, + 0.008831021375954151, + 0.009649422019720078, + 0.006699967663735151, + 0.007449025753885508, + 0.011678350158035755, + 0.020507000386714935, + 0.007229022681713104, + 0.008577825501561165, + 0.00836391281336546, + 0.008346798829734325, + 0.00484732910990715, + 0.016266295686364174, + 0.008531204424798489, + 0.007319103926420212, + 0.11817613244056702, + 0.027620472013950348, + 0.132803276181221, + 0.02035394124686718, + 0.01678261160850525, + 0.02163826860487461, + 0.0328749418258667, + 0.021649610251188278, + 0.008601123467087746, + 0.008561462163925171, + 0.0063826898112893105, + 0.010949851013720036, + 0.009905348531901836, + 0.011459868401288986, + 0.01213541254401207, + 0.00898006372153759, + 0.008556583896279335, + 0.015190299600362778, + 0.013403553515672684, + 0.015426247380673885, + 0.008507288061082363, + 0.009107169695198536, + 0.008282100781798363, + 0.009018772281706333, + 0.004217113833874464, + 0.005053589586168528, + 0.01104674767702818, + 0.005260745529085398, + 0.002761000534519553, + 0.0029989827889949083, + 0.004431780893355608, + 0.013489044271409512, + 0.008034811355173588, + 0.018210845068097115, + 0.004556856118142605, + 0.008631776086986065, + 0.004699323792010546, + 0.003941708244383335, + 0.005388251040130854, + 0.004937893245369196, + 0.004480371251702309, + 0.005458729341626167, + 0.004929247312247753, + 0.004229838494211435, + 0.004382542334496975, + 0.005847549065947533, + 0.004492986481636763, + 0.006048570387065411, + 0.008524073287844658, + 0.00875239260494709, + 0.0093481270596385, + 0.009595461189746857, + 0.011278298683464527, + 0.012377097271382809, + 0.011192790232598782, + 0.02022574655711651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.00864111352711916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001334729720838368 + ], + "opposite_logprob_tokens": [ + -6.6263346672058105 + ], + "logprob_reward": -0.001334729720838368, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001334729720838368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12906533479690552, + 0.11879754066467285, + 0.06027532368898392, + 0.177000492811203, + 0.10909732431173325, + 0.10390176624059677, + 0.07984820008277893, + 0.05679861828684807, + 0.0674392506480217, + 0.09777615964412689 + ], + "credit_map": [ + 0.017666978761553764, + 0.013139968737959862, + 0.01698451116681099, + 0.014020659029483795, + 0.016110654920339584, + 0.016763785853981972, + 0.017572056502103806, + 0.016806717962026596, + 0.010983839631080627, + 0.009877808392047882, + 0.012802116572856903, + 0.01486419141292572, + 0.011619480326771736, + 0.018439382314682007, + 0.01507695484906435, + 0.025133773684501648, + 0.0069320425391197205, + 0.004143924918025732, + 0.008443400263786316, + 0.004150847904384136, + 0.004893879406154156, + 0.017682692036032677, + 0.0052913338877260685, + 0.008737199939787388, + 0.04988998919725418, + 0.011843821965157986, + 0.05137431249022484, + 0.018194003030657768, + 0.007982420735061169, + 0.013278264552354813, + 0.013443263247609138, + 0.010994421318173409, + 0.01063638087362051, + 0.013034048490226269, + 0.00966385193169117, + 0.0199351254850626, + 0.012285872362554073, + 0.013665322214365005, + 0.012726125307381153, + 0.017150599509477615, + 0.011251982301473618, + 0.01865004561841488, + 0.015722140669822693, + 0.019097264856100082, + 0.009652586653828621, + 0.009298943914473057, + 0.009794232435524464, + 0.010434569790959358, + 0.009217659011483192, + 0.011262794025242329, + 0.015810756012797356, + 0.007760763168334961, + 0.006140540819615126, + 0.005092816427350044, + 0.006735449656844139, + 0.017827415838837624, + 0.008325627073645592, + 0.00786648876965046, + 0.005530221853405237, + 0.010183091275393963, + 0.005103304982185364, + 0.007038655690848827, + 0.006003780290484428, + 0.006747448351234198, + 0.005790384020656347, + 0.010061106644570827, + 0.008602508343756199, + 0.007258845493197441, + 0.008672692812979221, + 0.008898762986063957, + 0.0074257440865039825, + 0.010729197412729263, + 0.012802141718566418, + 0.01133286114782095, + 0.01036159135401249, + 0.009369110688567162, + 0.015002784319221973, + 0.011148033663630486, + 0.013868344947695732, + 0.013891291804611683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.001334729720838368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid figure with lean, roped-in muscles and wild, unkempt black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure adorned with a necklace of tiny, yellowed skulls and a shield made from beast jaws on their shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure launch explosively from a dense, mossy branch and tumble/swirl through the air grasping vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense jungle canopy with gnarled branches and enormous, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level wide-angle perspective with dappled sunlight creating bright hotspots and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208359956741333 + ], + "opposite_logprob_tokens": [ + -1.5020835399627686 + ], + "logprob_reward": -0.25208359956741333, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208359956741333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17782482504844666, + 0.16954007744789124, + 0.03791314363479614, + 0.17207391560077667, + 0.04674657806754112, + 0.04592900723218918, + 0.09772740304470062, + 0.10212374478578568, + 0.050109781324863434, + 0.10001152753829956 + ], + "credit_map": [ + 0.028187833726406097, + 0.01804579235613346, + 0.016008907929062843, + 0.00998115073889494, + 0.014791040681302547, + 0.017376096919178963, + 0.04070887714624405, + 0.03272511810064316, + 0.015467644669115543, + 0.01722666248679161, + 0.0255922582000494, + 0.03081616573035717, + 0.016033899039030075, + 0.019466865807771683, + 0.01678406074643135, + 0.028152504935860634, + 0.004189756233245134, + 0.0035401363857090473, + 0.005053797736763954, + 0.0034209969453513622, + 0.00381821789778769, + 0.00978555902838707, + 0.003949146252125502, + 0.004155533853918314, + 0.052836161106824875, + 0.011408105492591858, + 0.05725933238863945, + 0.011532205156981945, + 0.008371993899345398, + 0.0119505375623703, + 0.011085567995905876, + 0.007630001753568649, + 0.0062789395451545715, + 0.004888310097157955, + 0.004786440636962652, + 0.005215544253587723, + 0.006088295020163059, + 0.0067001949064433575, + 0.005449914839118719, + 0.0073389350436627865, + 0.004375775344669819, + 0.006116658914834261, + 0.007625562138855457, + 0.007290741428732872, + 0.004926189314574003, + 0.00424196245148778, + 0.006906024180352688, + 0.004446091596037149, + 0.007979546673595905, + 0.008358214981853962, + 0.02402222901582718, + 0.007875598035752773, + 0.006207106169313192, + 0.006396587938070297, + 0.007193727884441614, + 0.029694393277168274, + 0.018246691673994064, + 0.012914035469293594, + 0.0141949113458395, + 0.01651713252067566, + 0.009402753785252571, + 0.009627130813896656, + 0.009928063489496708, + 0.011293020099401474, + 0.009637463837862015, + 0.006189136765897274, + 0.005381355993449688, + 0.008896108716726303, + 0.0031710194889456034, + 0.0066574751399457455, + 0.005217582918703556, + 0.0049596428871154785, + 0.014317527413368225, + 0.013461864553391933, + 0.009189378470182419, + 0.014252357184886932, + 0.01218988187611103, + 0.011525268666446209, + 0.011612452566623688, + 0.013462796807289124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.25208359956741333 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_tensor([1969, 1969])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_tensor([1969, 1969])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4aa96679e9fa82fa047614dae37117068ff39a80 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_tensor([1969, 1969])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angled, full-body shot captures the scene in the style of an imperial portrait, focusing on a divine warrior standing steadfast and calm. His flowing white hair flutters gently, as if in a breeze, and his chiseled, timeless face wears a calm, contemplative expression. He clads in a voluminous, multi-layered silver robe woven from a fabric that shimmers like liquid moonlight and embroidered with subtle, luminous celestial patterns. Against a backdrop of a dark, stormy sky filled with swirling, roiling clouds of deep indigo and charcoal gray, a relentless stream of shadowy projectiles, like jagged shards of pure darkness, trailing ethereal black smoke, rapidly hurtles toward him. As they approach his body, an invisible energy field repels each one; the air around the warrior subtly ripples and twists, the projectiles being smoothly and deliberately deflected, their aggressive trajectories curving into graceful arcs that curve into the darkness. The entire action unfolds without the warrior moving a muscle, his calm posture and the dramatic, calm lighting from his own glowing robes creating a powerful, epic, unshakable atmosphere of divine power amidst the chaos.", + "seed": 26015, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_1969_seed_26015_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011438482906669378, + "vlm_logprob": -0.011438482906669378, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.011438482906669378, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0079382648691535 + ], + "opposite_logprob_tokens": [ + -5.007938385009766 + ], + "logprob_reward": -0.0079382648691535, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0079382648691535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10642091929912567, + 0.4295806586742401, + 0.0691533014178276, + 0.06734471023082733, + 0.03442101553082466, + 0.028050273656845093, + 0.08861447125673294, + 0.06802088767290115, + 0.057627756148576736, + 0.05076604709029198 + ], + "credit_map": [ + 0.013617691583931446, + 0.015574472025036812, + 0.018210439011454582, + 0.009637703187763691, + 0.014565547928214073, + 0.012927076779305935, + 0.011091107502579689, + 0.010796880349516869, + 0.08085738867521286, + 0.02465382218360901, + 0.04844895750284195, + 0.015137982554733753, + 0.01820596680045128, + 0.1641550213098526, + 0.052217986434698105, + 0.025903526693582535, + 0.007996446453034878, + 0.007924537174403667, + 0.013418870978057384, + 0.005914856214076281, + 0.005705970339477062, + 0.011907657608389854, + 0.008209080435335636, + 0.008075880818068981, + 0.009838031604886055, + 0.007266232278198004, + 0.014008399099111557, + 0.004607230890542269, + 0.007287475746124983, + 0.009419728070497513, + 0.0075129675678908825, + 0.007404636591672897, + 0.004826968535780907, + 0.0040699392557144165, + 0.007338980678468943, + 0.0027718953788280487, + 0.003560266923159361, + 0.0035086406860500574, + 0.003767638234421611, + 0.004576683975756168, + 0.003373703919351101, + 0.002866863738745451, + 0.004619203973561525, + 0.0034220258239656687, + 0.0034607520792633295, + 0.003302321769297123, + 0.0032247244380414486, + 0.0037806741893291473, + 0.011348438449203968, + 0.01212633028626442, + 0.015958420932292938, + 0.01184376422315836, + 0.011582576669752598, + 0.008268052712082863, + 0.008304256945848465, + 0.00918262917548418, + 0.007438310422003269, + 0.009247186593711376, + 0.011883283033967018, + 0.006409960333257914, + 0.007640816271305084, + 0.010448763146996498, + 0.0064855050295591354, + 0.008467059582471848, + 0.006555790547281504, + 0.0047906250692903996, + 0.010116035118699074, + 0.006157034542411566, + 0.0044733816757798195, + 0.012346449308097363, + 0.006385906599462032, + 0.006802530027925968, + 0.0037966829258948565, + 0.006413809023797512, + 0.006761649623513222, + 0.008036276325583458, + 0.004944588523358107, + 0.007720646448433399, + 0.008498328737914562, + 0.004594061989337206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0079382648691535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003205758286640048 + ], + "opposite_logprob_tokens": [ + -5.7532057762146 + ], + "logprob_reward": -0.003205758286640048, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003205758286640048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0687786266207695, + 0.6039689779281616, + 0.09688311070203781, + 0.06629196554422379, + 0.030286122113466263, + 0.013040159828960896, + 0.020745599642395973, + 0.04875948652625084, + 0.029317915439605713, + 0.021928057074546814 + ], + "credit_map": [ + 0.01019559521228075, + 0.008629248477518559, + 0.009975751861929893, + 0.00722514558583498, + 0.007586001884192228, + 0.006900655571371317, + 0.011289493180811405, + 0.00697673112154007, + 0.11498213559389114, + 0.03409557044506073, + 0.06785807758569717, + 0.015463119372725487, + 0.02177800051867962, + 0.23477989435195923, + 0.07842454314231873, + 0.036587588489055634, + 0.011279704980552197, + 0.011466321535408497, + 0.0220455601811409, + 0.006691556889563799, + 0.007267896085977554, + 0.020885301753878593, + 0.007937469519674778, + 0.009309290908277035, + 0.008143752813339233, + 0.007618434261530638, + 0.02055889368057251, + 0.004342535510659218, + 0.004579300060868263, + 0.00967372301965952, + 0.004710845183581114, + 0.00666447589173913, + 0.0036107490304857492, + 0.004963413812220097, + 0.0062382821924984455, + 0.002589586190879345, + 0.0029901363886892796, + 0.003659613663330674, + 0.0029366028029471636, + 0.0032977371010929346, + 0.0016552337910979986, + 0.001532547059468925, + 0.002387396525591612, + 0.0013287148904055357, + 0.001479830709286034, + 0.0013547667767852545, + 0.0017209835350513458, + 0.0015806866576895118, + 0.002876691985875368, + 0.002534259809181094, + 0.004442659672349691, + 0.002492987783625722, + 0.0019680620171129704, + 0.0025488720275461674, + 0.0016921067144721746, + 0.002189959865063429, + 0.005761878099292517, + 0.005978207103908062, + 0.007862264290452003, + 0.005203316919505596, + 0.004897912964224815, + 0.006824319716542959, + 0.006340065039694309, + 0.005891520529985428, + 0.0031894503626972437, + 0.002973230555653572, + 0.0048185111954808235, + 0.0038363607600331306, + 0.0026944573037326336, + 0.0049567753449082375, + 0.0038216607645154, + 0.0030274682212620974, + 0.002019008854404092, + 0.002916883211582899, + 0.003967962693423033, + 0.001981825567781925, + 0.0019153032917529345, + 0.0041300831362605095, + 0.0028805937618017197, + 0.002116394927725196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.003205758286640048 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361985132098198 + ], + "opposite_logprob_tokens": [ + -3.6513619422912598 + ], + "logprob_reward": -0.026361985132098198, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361985132098198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08817856013774872, + 0.5290661454200745, + 0.10701531171798706, + 0.10327920317649841, + 0.02794606238603592, + 0.009110300801694393, + 0.029196765273809433, + 0.02146238274872303, + 0.027013424783945084, + 0.05773177370429039 + ], + "credit_map": [ + 0.015100521966814995, + 0.008989853784441948, + 0.014876098372042179, + 0.008483516983687878, + 0.007882151752710342, + 0.014519214630126953, + 0.011384343728423119, + 0.006942866835743189, + 0.09983473271131516, + 0.030171332880854607, + 0.05821222811937332, + 0.01718769036233425, + 0.019690698012709618, + 0.20339134335517883, + 0.0678020715713501, + 0.032776083797216415, + 0.012688767164945602, + 0.012538565322756767, + 0.023577464744448662, + 0.006579115055501461, + 0.009188457392156124, + 0.023627718910574913, + 0.00787669699639082, + 0.010938537307083607, + 0.01674266718327999, + 0.007933003827929497, + 0.03755655139684677, + 0.004497603978961706, + 0.006556074600666761, + 0.015034042298793793, + 0.006796612869948149, + 0.0081626633182168, + 0.0042330240830779076, + 0.004277394153177738, + 0.00570636847987771, + 0.0019451744155958295, + 0.0022697895765304565, + 0.004220666363835335, + 0.002339709084481001, + 0.002953939139842987, + 0.0014101805863901973, + 0.0012503298930823803, + 0.002069406444206834, + 0.0006973595009185374, + 0.0008791330619715154, + 0.0008999360725283623, + 0.0007388066733255982, + 0.0011651493841782212, + 0.004816248081624508, + 0.0035334357526153326, + 0.0049382662400603294, + 0.0023613078519701958, + 0.002877582563087344, + 0.0036650998517870903, + 0.0025383250322192907, + 0.004466503392904997, + 0.0026941162068396807, + 0.002433492336422205, + 0.0029113744385540485, + 0.002068293048068881, + 0.001577345421537757, + 0.0042521036230027676, + 0.0019622989930212498, + 0.003563361242413521, + 0.0030367623548954725, + 0.003316114190965891, + 0.005160636268556118, + 0.002991100074723363, + 0.002801429945975542, + 0.0031734195072203875, + 0.0031201716046780348, + 0.003413793630897999, + 0.004748312756419182, + 0.009252767078578472, + 0.006251113954931498, + 0.008163153193891048, + 0.006539182737469673, + 0.008484276942908764, + 0.009226840920746326, + 0.005066129378974438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.026361985132098198 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036206187214702368 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036206187214702368, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036206187214702368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03199644014239311, + 0.8616753220558167, + 0.03498290479183197, + 0.02228621020913124, + 0.010199104435741901, + 0.005245064850896597, + 0.008380756713449955, + 0.006434719078242779, + 0.008327270857989788, + 0.010472210124135017 + ], + "credit_map": [ + 0.00453032786026597, + 0.003113877261057496, + 0.005344450008124113, + 0.003300210228189826, + 0.0029816078022122383, + 0.005726140458136797, + 0.004104699473828077, + 0.0028951282147318125, + 0.16463692486286163, + 0.04886825010180473, + 0.09691905975341797, + 0.023325549438595772, + 0.03011680580675602, + 0.3352898359298706, + 0.11010617017745972, + 0.0524127371609211, + 0.0038380713667720556, + 0.004129320383071899, + 0.006721903569996357, + 0.002909607719630003, + 0.0034464977215975523, + 0.0073976279236376286, + 0.0037537184543907642, + 0.0027861574199050665, + 0.00305515737272799, + 0.0019799303263425827, + 0.0075360010378062725, + 0.0013012923300266266, + 0.0017272306140512228, + 0.0033546402119100094, + 0.001676261774264276, + 0.0016556972404941916, + 0.0011881249956786633, + 0.0016414052806794643, + 0.0023532146587967873, + 0.0009115854627452791, + 0.0008085466106422246, + 0.0014310512924566865, + 0.0008562419097870588, + 0.0010089336428791285, + 0.0007966553675942123, + 0.0004953997558914125, + 0.001097386353649199, + 0.0005378688219934702, + 0.0005768632399849594, + 0.0006744374986737967, + 0.0004718516720458865, + 0.000594602432101965, + 0.001570983207784593, + 0.0008227460202760994, + 0.0019611495081335306, + 0.0005029201274737716, + 0.0007669891347177327, + 0.0010736682452261448, + 0.0005490189651027322, + 0.001133281853981316, + 0.0008619540021754801, + 0.0005249006790108979, + 0.0011117197573184967, + 0.00048560480354353786, + 0.0005976665415801108, + 0.0012603278737515211, + 0.0005150812212377787, + 0.001077463966794312, + 0.0007807755609974265, + 0.0011316400486975908, + 0.0012625906383618712, + 0.0009284959523938596, + 0.0009400617564097047, + 0.0013525881804525852, + 0.00074424734339118, + 0.0011868707370012999, + 0.0008568182238377631, + 0.0014406448462978005, + 0.0014702193439006805, + 0.0013230626937001944, + 0.0012564169010147452, + 0.0016120844520628452, + 0.0014094602083787322, + 0.0011035037459805608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0036206187214702368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606578752398491 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.01606578752398491, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606578752398491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01189436949789524, + 0.8887488842010498, + 0.07417993992567062, + 0.014566071331501007, + 0.0023787051904946566, + 0.002318435348570347, + 0.001241096993908286, + 0.0025111958384513855, + 0.0014295767759904265, + 0.0007316346163861454 + ], + "credit_map": [ + 0.001794461510144174, + 0.0012995885917916894, + 0.002115031471475959, + 0.0011091515189036727, + 0.0010794587433338165, + 0.0018713781610131264, + 0.001453247619792819, + 0.0011720528127625585, + 0.17050297558307648, + 0.050415121018886566, + 0.09992963820695877, + 0.022897686809301376, + 0.031219301745295525, + 0.345935195684433, + 0.11398667097091675, + 0.053862396627664566, + 0.009182065725326538, + 0.006830194499343634, + 0.01976477913558483, + 0.005001841578632593, + 0.005349725019186735, + 0.016399133950471878, + 0.005374573636800051, + 0.006277634296566248, + 0.0020942832343280315, + 0.0014953449135646224, + 0.004673463758081198, + 0.0006382326246239245, + 0.0010923530207946897, + 0.0021189332474023104, + 0.000880780047737062, + 0.0015726815909147263, + 0.0003756022488232702, + 0.0003805227461270988, + 0.0006027783965691924, + 0.00017232839309144765, + 0.000208489247597754, + 0.00022633367916569114, + 0.0001983418333111331, + 0.00021430864580906928, + 0.0003332757332827896, + 0.00039206555811688304, + 0.00052473513642326, + 0.0001766721106832847, + 0.00019286054885014892, + 0.0002587689959909767, + 0.0001518366188975051, + 0.0002882207918446511, + 0.00015510039520449936, + 0.00012967904331162572, + 0.0002515492669772357, + 0.00012256672198418528, + 0.00014918751548975706, + 0.00016134361794684082, + 0.00012141190381953493, + 0.00015025849279481918, + 0.0003076240827795118, + 0.00031344222952611744, + 0.000645069289021194, + 0.00019991224689874798, + 0.0001919494679896161, + 0.00041835970478132367, + 0.00017866712005343288, + 0.00025617191568017006, + 0.0001693590311333537, + 0.00011559634003788233, + 0.0003345343575347215, + 0.00015618355246260762, + 0.0001231925270985812, + 0.00022674628417007625, + 0.0001347262441413477, + 0.00016923855582717806, + 8.485303987981752e-05, + 6.876199768157676e-05, + 0.00018468998314347118, + 6.55797339277342e-05, + 5.785613757325336e-05, + 9.515813871985301e-05, + 7.257386459968984e-05, + 0.00010216177179245278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.01606578752398491 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, full-body shot captures the scene in the style of an imperial portrait, focusing on a divine warrior standing steadfast and calm. His flowing white hair flutters gently, as if in a breeze, and his chiseled, timeless face wears a calm, contemplative expression. He clads in a voluminous, multi-layered silver robe woven from a fabric that shimmers like liquid moonlight and embroidered with subtle, luminous celestial patterns. Against a backdrop of a dark, stormy sky filled with swirling, roiling clouds of deep indigo and charcoal gray, a relentless stream of shadowy projectiles, like jagged shards of pure darkness, trailing ethereal black smoke, rapidly hurtles toward him. As they approach his body, an invisible energy field repels each one; the air around the warrior subtly ripples and twists, the projectiles being smoothly and deliberately deflected, their aggressive trajectories curving into graceful arcs that curve into the darkness. The entire action unfolds without the warrior moving a muscle, his calm posture and the dramatic, calm lighting from his own glowing robes creating a powerful, epic, unshakable atmosphere of divine power amidst the chaos.", + "seed": 26016, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_1969_seed_26016_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.005846363166347146, + "vlm_logprob": -0.005846363166347146, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.005846363166347146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004544881638139486 + ], + "opposite_logprob_tokens": [ + -5.754544734954834 + ], + "logprob_reward": -0.004544881638139486, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004544881638139486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42656171321868896, + 0.08665593713521957, + 0.07620685547590256, + 0.10562925040721893, + 0.07121703773736954, + 0.05596129968762398, + 0.03764132410287857, + 0.049591902643442154, + 0.047386664897203445, + 0.043148063123226166 + ], + "credit_map": [ + 0.0339852049946785, + 0.02707144431769848, + 0.03146735578775406, + 0.08599837869405746, + 0.016837935894727707, + 0.0625065267086029, + 0.04185599088668823, + 0.12683884799480438, + 0.011901350691914558, + 0.01031150296330452, + 0.012246353551745415, + 0.008863496594130993, + 0.007438193075358868, + 0.011791559867560863, + 0.011163156479597092, + 0.012940313667058945, + 0.005755293183028698, + 0.010722244158387184, + 0.007783138658851385, + 0.009512316435575485, + 0.006506071425974369, + 0.00816758256405592, + 0.01494942232966423, + 0.012810789048671722, + 0.010484482161700726, + 0.01366608776152134, + 0.014796596020460129, + 0.016140880063176155, + 0.007949131540954113, + 0.006986001506447792, + 0.026421504095196724, + 0.00918456632643938, + 0.009155099280178547, + 0.006644972134381533, + 0.00914706103503704, + 0.008788682520389557, + 0.006141170859336853, + 0.00817255862057209, + 0.015136493369936943, + 0.008031001314520836, + 0.008389043621718884, + 0.008639355190098286, + 0.007441997062414885, + 0.006256510969251394, + 0.005197532009333372, + 0.005230872426182032, + 0.009183168411254883, + 0.00562282232567668, + 0.005780351813882589, + 0.004332711920142174, + 0.004907389637082815, + 0.0037044393830001354, + 0.0032596851233392954, + 0.006024658679962158, + 0.005599985830485821, + 0.0040321010164916515, + 0.004964435938745737, + 0.007042385637760162, + 0.0077338265255093575, + 0.006975830532610416, + 0.00543083343654871, + 0.0035110479220747948, + 0.008993212133646011, + 0.004940330516546965, + 0.005059198010712862, + 0.00667999405413866, + 0.007896509021520615, + 0.005472821183502674, + 0.005534936673939228, + 0.0039312285371124744, + 0.008931242860853672, + 0.003880735021084547, + 0.005507056601345539, + 0.005042813252657652, + 0.006857110187411308, + 0.004721856210380793, + 0.00586359016597271, + 0.005669568665325642, + 0.00601178128272295, + 0.0034742853604257107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.004544881638139486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025027638766914606 + ], + "opposite_logprob_tokens": [ + -6.002502918243408 + ], + "logprob_reward": -0.0025027638766914606, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025027638766914606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.57027667760849, + 0.05605066195130348, + 0.06709349900484085, + 0.04733986034989357, + 0.04300748556852341, + 0.03362884372472763, + 0.06297223269939423, + 0.04299939423799515, + 0.041764453053474426, + 0.03486694023013115 + ], + "credit_map": [ + 0.04527759179472923, + 0.028871292248368263, + 0.03567526489496231, + 0.1381702423095703, + 0.018392914906144142, + 0.07784351706504822, + 0.05196903645992279, + 0.17407678067684174, + 0.006829730700701475, + 0.008377834223210812, + 0.008015763945877552, + 0.0052518476732075214, + 0.004251609090715647, + 0.007882404141128063, + 0.006716384552419186, + 0.008725090883672237, + 0.006191140506416559, + 0.008530215360224247, + 0.007589379325509071, + 0.008582338690757751, + 0.00600710092112422, + 0.008984940126538277, + 0.013860856182873249, + 0.007347525097429752, + 0.004990240093320608, + 0.007436501327902079, + 0.007706211414188147, + 0.00469278683885932, + 0.004803231451660395, + 0.0038374990690499544, + 0.007091956213116646, + 0.006781433708965778, + 0.005340819247066975, + 0.004004816059023142, + 0.005982715170830488, + 0.004711554851382971, + 0.0049974191933870316, + 0.004928800277411938, + 0.008674872107803822, + 0.004366491921246052, + 0.0029168243054300547, + 0.0059712957590818405, + 0.0040018148720264435, + 0.003984709270298481, + 0.0036486799363046885, + 0.0029874895699322224, + 0.006762308068573475, + 0.0033557203132659197, + 0.007138910237699747, + 0.006217495072633028, + 0.010064782574772835, + 0.007837191224098206, + 0.005769927520304918, + 0.006904597859829664, + 0.013595351949334145, + 0.005443976726382971, + 0.005090038757771254, + 0.004335100296884775, + 0.004778020549565554, + 0.005537578836083412, + 0.004771634936332703, + 0.005626928061246872, + 0.008008218370378017, + 0.004851874429732561, + 0.0049914405681192875, + 0.004171598702669144, + 0.00695962505415082, + 0.005400510970503092, + 0.004361375235021114, + 0.0038482556119561195, + 0.0069041782990098, + 0.005127464886754751, + 0.003424671245738864, + 0.0050322310999035835, + 0.004249945282936096, + 0.003862655023112893, + 0.004056369885802269, + 0.004879795480519533, + 0.0065460242331027985, + 0.0028152484446763992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0025027638766914606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075171530246735 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075171530246735, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075171530246735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6165953278541565, + 0.08732166886329651, + 0.03166232630610466, + 0.04454873129725456, + 0.055637918412685394, + 0.03179656341671944, + 0.0507059283554554, + 0.019980531185865402, + 0.048049069941043854, + 0.01370189618319273 + ], + "credit_map": [ + 0.04865181818604469, + 0.031110359355807304, + 0.0393797792494297, + 0.15119005739688873, + 0.019014647230505943, + 0.08163966983556747, + 0.05696649104356766, + 0.18864251673221588, + 0.012604491785168648, + 0.010610947385430336, + 0.011781065724790096, + 0.007986840792000294, + 0.007953205145895481, + 0.012283495627343655, + 0.008580976165831089, + 0.015520649962127209, + 0.0036447567399591208, + 0.003942169714719057, + 0.0034582274965941906, + 0.003298982745036483, + 0.002747737802565098, + 0.005549595225602388, + 0.004603327717632055, + 0.0044175321236252785, + 0.005090674385428429, + 0.007296346593648195, + 0.003916903864592314, + 0.00475251954048872, + 0.003398055909201503, + 0.004965639673173428, + 0.007677736692130566, + 0.00745085533708334, + 0.007013820111751556, + 0.0046973018907010555, + 0.007497016806155443, + 0.006933612748980522, + 0.0059671709313988686, + 0.0055908639915287495, + 0.011761367321014404, + 0.0061767688021063805, + 0.0034405856858938932, + 0.00544552830979228, + 0.004356728866696358, + 0.003577049355953932, + 0.0034644207917153835, + 0.0035445785615593195, + 0.005204395391047001, + 0.002763276221230626, + 0.0071664187125861645, + 0.0049169003032147884, + 0.008276693522930145, + 0.005956994369626045, + 0.00462826993316412, + 0.005062106531113386, + 0.009475433267652988, + 0.005223116371780634, + 0.002258632332086563, + 0.002616289770230651, + 0.0017839678330346942, + 0.002345031127333641, + 0.0027955833356827497, + 0.0023376471363008022, + 0.003868039697408676, + 0.001975342398509383, + 0.005557587835937738, + 0.005410904996097088, + 0.007467072457075119, + 0.005887956358492374, + 0.0053595565259456635, + 0.00441804388538003, + 0.009394600987434387, + 0.004553352016955614, + 0.0016358980210497975, + 0.0018301468808203936, + 0.0020061142276972532, + 0.0014588444028049707, + 0.001680489513091743, + 0.0017710513202473521, + 0.0021257149055600166, + 0.0011936372611671686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.016075171530246735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024924185127019882 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024924185127019882, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024924185127019882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4433368146419525, + 0.10837986320257187, + 0.06294704228639603, + 0.05306890234351158, + 0.06977126002311707, + 0.07169134169816971, + 0.04099716991186142, + 0.04539182409644127, + 0.07914016395807266, + 0.025275561958551407 + ], + "credit_map": [ + 0.03387070819735527, + 0.025157354772090912, + 0.03026961162686348, + 0.10819021612405777, + 0.0182072464376688, + 0.05925726890563965, + 0.04386823624372482, + 0.12451621890068054, + 0.012799986638128757, + 0.008314928039908409, + 0.012507391162216663, + 0.012239610776305199, + 0.009563141502439976, + 0.013743196614086628, + 0.018509596586227417, + 0.020702015608549118, + 0.005249431356787682, + 0.007762682158499956, + 0.007170677185058594, + 0.009453239850699902, + 0.005896385293453932, + 0.006034843157976866, + 0.015217700973153114, + 0.006162086967378855, + 0.0058680386282503605, + 0.006337198428809643, + 0.006205109879374504, + 0.006480011623352766, + 0.004432161338627338, + 0.005445193964987993, + 0.010612173937261105, + 0.007689016405493021, + 0.008973288349807262, + 0.005689841695129871, + 0.009535057470202446, + 0.008848875761032104, + 0.007538144011050463, + 0.0066860695369541645, + 0.014847595244646072, + 0.0076524014584720135, + 0.012325888499617577, + 0.007863172329962254, + 0.007725230883806944, + 0.006814529653638601, + 0.007281388156116009, + 0.011828914284706116, + 0.0077515835873782635, + 0.010100641287863255, + 0.00556653318926692, + 0.005447059869766235, + 0.00455976789817214, + 0.004126981366425753, + 0.003797302721068263, + 0.0069864485412836075, + 0.006575143896043301, + 0.003937937784940004, + 0.005014512687921524, + 0.005028177984058857, + 0.004915193654596806, + 0.0058016348630189896, + 0.0053132460452616215, + 0.00600480055436492, + 0.007817387580871582, + 0.005496873985975981, + 0.009867113083600998, + 0.009752748534083366, + 0.010104664601385593, + 0.012055044993758202, + 0.007861291989684105, + 0.007355790119618177, + 0.01348956860601902, + 0.008653948083519936, + 0.0023328608367592096, + 0.003455036086961627, + 0.0024732821621000767, + 0.0034997519105672836, + 0.0029065103735774755, + 0.002968463348224759, + 0.0056067961268126965, + 0.002032862976193428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0024924185127019882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036165802739560604 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036165802739560604, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036165802739560604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5805767178535461, + 0.03927995637059212, + 0.05831748992204666, + 0.03307854011654854, + 0.039912983775138855, + 0.05652286112308502, + 0.011485596187412739, + 0.11329033225774765, + 0.03382726386189461, + 0.03370831534266472 + ], + "credit_map": [ + 0.05206802487373352, + 0.03058508038520813, + 0.03851267322897911, + 0.13262632489204407, + 0.020308110862970352, + 0.08261574059724808, + 0.05169909447431564, + 0.17216160893440247, + 0.004660866688936949, + 0.0044988724403083324, + 0.003976649139076471, + 0.0042819855734705925, + 0.003411015495657921, + 0.005692834034562111, + 0.007018666248768568, + 0.005739064887166023, + 0.006347925867885351, + 0.005829242058098316, + 0.006540164817124605, + 0.006687362678349018, + 0.005615689791738987, + 0.008694512769579887, + 0.0104189682751894, + 0.008183615282177925, + 0.003747624810785055, + 0.004845666233450174, + 0.003375991480425, + 0.003113384824246168, + 0.003046276280656457, + 0.0036598832812160254, + 0.00529876071959734, + 0.005990950390696526, + 0.0048501091077923775, + 0.0042225392535328865, + 0.004174139350652695, + 0.0038962592370808125, + 0.004145151469856501, + 0.007166646886616945, + 0.0049478644505143166, + 0.006510270293802023, + 0.01103335153311491, + 0.0047663976438343525, + 0.005971672013401985, + 0.005004263482987881, + 0.005163656547665596, + 0.009088397957384586, + 0.006437897682189941, + 0.009057221934199333, + 0.0011161919683218002, + 0.0017746081575751305, + 0.0015539873857051134, + 0.0012245866237208247, + 0.0013669540639966726, + 0.0015218264888972044, + 0.0019425794016569853, + 0.0009848602348938584, + 0.012921039946377277, + 0.014115288853645325, + 0.014480500482022762, + 0.012290828861296177, + 0.014928151853382587, + 0.016741806641221046, + 0.01629728265106678, + 0.011515420861542225, + 0.004301948007196188, + 0.0030530965887010098, + 0.0037805140018463135, + 0.004673033952713013, + 0.0027449706103652716, + 0.00519830035045743, + 0.004740896634757519, + 0.005334500223398209, + 0.0040735239163041115, + 0.003781314706429839, + 0.004396654665470123, + 0.003430904820561409, + 0.0047462135553359985, + 0.005411169026046991, + 0.004381115082651377, + 0.0034874174743890762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0036165802739560604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, full-body shot captures the scene in the style of an imperial portrait, focusing on a divine warrior standing steadfast and calm. His flowing white hair flutters gently, as if in a breeze, and his chiseled, timeless face wears a calm, contemplative expression. He clads in a voluminous, multi-layered silver robe woven from a fabric that shimmers like liquid moonlight and embroidered with subtle, luminous celestial patterns. Against a backdrop of a dark, stormy sky filled with swirling, roiling clouds of deep indigo and charcoal gray, a relentless stream of shadowy projectiles, like jagged shards of pure darkness, trailing ethereal black smoke, rapidly hurtles toward him. As they approach his body, an invisible energy field repels each one; the air around the warrior subtly ripples and twists, the projectiles being smoothly and deliberately deflected, their aggressive trajectories curving into graceful arcs that curve into the darkness. The entire action unfolds without the warrior moving a muscle, his calm posture and the dramatic, calm lighting from his own glowing robes creating a powerful, epic, unshakable atmosphere of divine power amidst the chaos.", + "seed": 26017, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_1969_seed_26017_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.006675838399678469, + "vlm_logprob": -0.006675838399678469, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.006675838399678469, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0023664822801947594 + ], + "opposite_logprob_tokens": [ + -6.752366542816162 + ], + "logprob_reward": -0.0023664822801947594, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0023664822801947594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10493379831314087, + 0.09193368256092072, + 0.03337664157152176, + 0.04780590161681175, + 0.06630006432533264, + 0.03904971852898598, + 0.03439301624894142, + 0.31551793217658997, + 0.21986347436904907, + 0.04682575538754463 + ], + "credit_map": [ + 0.010730043984949589, + 0.016889341175556183, + 0.01570940390229225, + 0.01265658624470234, + 0.017825458198785782, + 0.012425117194652557, + 0.0071905627846717834, + 0.011507278308272362, + 0.009315776638686657, + 0.015490356832742691, + 0.014010792598128319, + 0.009072701446712017, + 0.016493147239089012, + 0.00989074818789959, + 0.008115233853459358, + 0.009544925764203072, + 0.0034060580655932426, + 0.0045490870252251625, + 0.004685650579631329, + 0.003049952909350395, + 0.005595772061496973, + 0.004107935354113579, + 0.0030078308191150427, + 0.0049743507988750935, + 0.004792907740920782, + 0.006313341204077005, + 0.00595234427601099, + 0.005041527561843395, + 0.006800529081374407, + 0.006016048137098551, + 0.004875747486948967, + 0.008013454265892506, + 0.008611451834440231, + 0.009490950964391232, + 0.010252675972878933, + 0.00526099419221282, + 0.007743430323898792, + 0.007203138433396816, + 0.007948013953864574, + 0.00978941097855568, + 0.004802921321243048, + 0.00466633215546608, + 0.00560497771948576, + 0.00343134393915534, + 0.00532269524410367, + 0.004735077731311321, + 0.003745882073417306, + 0.006740488111972809, + 0.004110075067728758, + 0.0034550612326711416, + 0.0057058678939938545, + 0.0038636724930256605, + 0.0035584873985499144, + 0.004090632311999798, + 0.005041154101490974, + 0.004568064119666815, + 0.023827608674764633, + 0.02292337268590927, + 0.03838355466723442, + 0.010991767980158329, + 0.01562776416540146, + 0.10718953609466553, + 0.014858383685350418, + 0.08171594887971878, + 0.084583580493927, + 0.008576203137636185, + 0.028079206123948097, + 0.006914900150150061, + 0.008434521034359932, + 0.06210334599018097, + 0.009988460689783096, + 0.011183253489434719, + 0.009334764443337917, + 0.005042144097387791, + 0.005391471553593874, + 0.0035375040024518967, + 0.005971295293420553, + 0.008901773951947689, + 0.00376197905279696, + 0.004884825088083744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0023664822801947594 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104761406779289 + ], + "opposite_logprob_tokens": [ + -5.5041046142578125 + ], + "logprob_reward": -0.004104761406779289, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104761406779289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10610423237085342, + 0.08603204786777496, + 0.08068511635065079, + 0.0453515350818634, + 0.0822264552116394, + 0.05458831787109375, + 0.026998324319720268, + 0.262370765209198, + 0.16182494163513184, + 0.09381823241710663 + ], + "credit_map": [ + 0.009492394514381886, + 0.012950834818184376, + 0.017200147733092308, + 0.012519045732915401, + 0.01723404973745346, + 0.011156348511576653, + 0.011743659153580666, + 0.013807743787765503, + 0.009450054727494717, + 0.012119505554437637, + 0.016496244817972183, + 0.0071646119467914104, + 0.010847092606127262, + 0.009901386685669422, + 0.008230151608586311, + 0.011822997592389584, + 0.00853428803384304, + 0.011436131782829762, + 0.010895715095102787, + 0.010940901935100555, + 0.01083243265748024, + 0.009365778416395187, + 0.009596058167517185, + 0.009083811193704605, + 0.00542551651597023, + 0.007528022862970829, + 0.005669345613569021, + 0.004733023699373007, + 0.006597541272640228, + 0.0050337002612650394, + 0.004351433366537094, + 0.006012951023876667, + 0.015265047550201416, + 0.00853709690272808, + 0.013879591599106789, + 0.005871891975402832, + 0.0063479081727564335, + 0.010030565783381462, + 0.006887174677103758, + 0.01540718600153923, + 0.006659395527094603, + 0.005252630449831486, + 0.012764721177518368, + 0.004425174091011286, + 0.003941535484045744, + 0.007085525896400213, + 0.005181631073355675, + 0.009277703240513802, + 0.003235753858461976, + 0.0026683148462325335, + 0.004714303649961948, + 0.0028154635801911354, + 0.0033989124931395054, + 0.003478509606793523, + 0.004141214769333601, + 0.0025458517484366894, + 0.02073649689555168, + 0.023549124598503113, + 0.03321804851293564, + 0.01247114222496748, + 0.014593780972063541, + 0.08211918920278549, + 0.015557508915662766, + 0.06012549251317978, + 0.04240687936544418, + 0.011424073949456215, + 0.035477619618177414, + 0.013287634588778019, + 0.010668978095054626, + 0.016396038234233856, + 0.011590059846639633, + 0.020573673769831657, + 0.022464409470558167, + 0.007450094446539879, + 0.01230460312217474, + 0.005900693591684103, + 0.00820624828338623, + 0.019451355561614037, + 0.00829299259930849, + 0.009747840464115143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.004104761406779289 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607399806380272 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.01607399806380272, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607399806380272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061277758330106735, + 0.04924563691020012, + 0.03207501396536827, + 0.04337901622056961, + 0.24649691581726074, + 0.0935245007276535, + 0.029378492385149002, + 0.20061823725700378, + 0.21116375923156738, + 0.03284056857228279 + ], + "credit_map": [ + 0.011290107853710651, + 0.007971507497131824, + 0.007231875788420439, + 0.005998146254569292, + 0.00868014246225357, + 0.007303295657038689, + 0.00495872413739562, + 0.007843966595828533, + 0.00657934183254838, + 0.00513082928955555, + 0.008884903974831104, + 0.004396846983581781, + 0.0061540380120277405, + 0.005314181558787823, + 0.0040873149409890175, + 0.008698183111846447, + 0.005161843728274107, + 0.0035945724230259657, + 0.004921731539070606, + 0.0036974663380533457, + 0.0037758992984890938, + 0.0040389178320765495, + 0.0031466365326195955, + 0.003737953258678317, + 0.0038654992822557688, + 0.006982759106904268, + 0.005236064549535513, + 0.004811204504221678, + 0.005604086443781853, + 0.004662263672798872, + 0.00395794166252017, + 0.00825920607894659, + 0.04148641228675842, + 0.01964660733938217, + 0.05072209984064102, + 0.01600160263478756, + 0.021162696182727814, + 0.03345631808042526, + 0.01605650968849659, + 0.047964707016944885, + 0.011269859969615936, + 0.010942380875349045, + 0.017774008214473724, + 0.00847549457103014, + 0.00926145538687706, + 0.01292182132601738, + 0.009039437398314476, + 0.013840055093169212, + 0.003302768338471651, + 0.004045357462018728, + 0.006094747222959995, + 0.0024314364418387413, + 0.0025978833436965942, + 0.003330764127895236, + 0.002305530710145831, + 0.005270008463412523, + 0.017059853300452232, + 0.007195035926997662, + 0.025009606033563614, + 0.0058000339195132256, + 0.007032888010144234, + 0.07623350620269775, + 0.00882829912006855, + 0.0534590445458889, + 0.07244449853897095, + 0.011372403241693974, + 0.03686083108186722, + 0.010816330090165138, + 0.00831160694360733, + 0.0424216166138649, + 0.010363838635385036, + 0.01857266202569008, + 0.005470354110002518, + 0.0029213433153927326, + 0.005274411290884018, + 0.0020618310663849115, + 0.0040383352898061275, + 0.005194544326514006, + 0.002702117431908846, + 0.005177633836865425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.01607399806380272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040994188748300076 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.0040994188748300076, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040994188748300076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.093055859208107, + 0.08271817862987518, + 0.03269714489579201, + 0.03685883805155754, + 0.12444741278886795, + 0.08687309920787811, + 0.047714538872241974, + 0.17452776432037354, + 0.29846134781837463, + 0.022645847871899605 + ], + "credit_map": [ + 0.01998314820230007, + 0.009243017062544823, + 0.011316953226923943, + 0.009964548982679844, + 0.013580841943621635, + 0.011622955091297626, + 0.0075686718337237835, + 0.00977571401745081, + 0.011569674126803875, + 0.009901260025799274, + 0.014006504788994789, + 0.007198619190603495, + 0.010519501753151417, + 0.009441993199288845, + 0.006157957017421722, + 0.01392266433686018, + 0.004525065887719393, + 0.00370520306751132, + 0.0050957039929926395, + 0.003871156135573983, + 0.00506976805627346, + 0.003784280503168702, + 0.0033080221619457006, + 0.0033379467204213142, + 0.00439621414989233, + 0.006103243678808212, + 0.0048258742317557335, + 0.0044091371819376945, + 0.005357676185667515, + 0.004162601660937071, + 0.003301482181996107, + 0.004302607383579016, + 0.023780345916748047, + 0.011880612000823021, + 0.021323557943105698, + 0.008228454738855362, + 0.010924826376140118, + 0.01610919088125229, + 0.009701059199869633, + 0.02249935455620289, + 0.009318082593381405, + 0.01100409310311079, + 0.014768377877771854, + 0.007981713861227036, + 0.00882179755717516, + 0.01000477746129036, + 0.009620552882552147, + 0.015353702940046787, + 0.006905004847794771, + 0.004590989090502262, + 0.009079376235604286, + 0.0035982972476631403, + 0.0043829637579619884, + 0.00747337331995368, + 0.0034089379478245974, + 0.008275595493614674, + 0.01624455489218235, + 0.008804913610219955, + 0.02198740839958191, + 0.006194166373461485, + 0.007940164767205715, + 0.0648762658238411, + 0.009077941067516804, + 0.039402350783348083, + 0.09518203884363174, + 0.016367094591259956, + 0.055168211460113525, + 0.0172471534460783, + 0.013439948670566082, + 0.0514151006937027, + 0.015898223966360092, + 0.03374355658888817, + 0.002993660746142268, + 0.0029411225114017725, + 0.0036169374361634254, + 0.002126396168023348, + 0.003458134364336729, + 0.0022804427426308393, + 0.002260497771203518, + 0.0029686561319977045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.0040994188748300076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734531372785568 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734531372785568, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734531372785568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17545326054096222, + 0.03774948790669441, + 0.13507084548473358, + 0.14863379299640656, + 0.05457252264022827, + 0.08408572524785995, + 0.0810006782412529, + 0.031237492337822914, + 0.12359676510095596, + 0.12859949469566345 + ], + "credit_map": [ + 0.04662428796291351, + 0.020105326548218727, + 0.022326216101646423, + 0.01566506363451481, + 0.017088349908590317, + 0.024943996220827103, + 0.011284292675554752, + 0.017415711656212807, + 0.005670497659593821, + 0.004004419315606356, + 0.00546561274677515, + 0.0035847139079123735, + 0.005174280609935522, + 0.004396962933242321, + 0.004339052829891443, + 0.0051139467395842075, + 0.02156646177172661, + 0.015692338347434998, + 0.01843240298330784, + 0.012986847199499607, + 0.01389575470238924, + 0.015607225708663464, + 0.013524141162633896, + 0.023365655913949013, + 0.019771482795476913, + 0.011540994048118591, + 0.026589632034301758, + 0.010228138417005539, + 0.008711094036698341, + 0.020258594304323196, + 0.010244163684546947, + 0.041289668530225754, + 0.011114878579974174, + 0.004871728830039501, + 0.009898023679852486, + 0.003444206900894642, + 0.003518833080306649, + 0.008065931499004364, + 0.003620619187131524, + 0.01003829762339592, + 0.017432307824492455, + 0.007338578812777996, + 0.012933914549648762, + 0.006276682019233704, + 0.009915024973452091, + 0.00904111284762621, + 0.008017884567379951, + 0.013130213133990765, + 0.015448318794369698, + 0.006630406249314547, + 0.014621352776885033, + 0.00595758156850934, + 0.004647360183298588, + 0.013220294378697872, + 0.007104905322194099, + 0.013370453380048275, + 0.00299322372302413, + 0.0014876804780215025, + 0.0033052244689315557, + 0.0013706928584724665, + 0.0016813133843243122, + 0.007634546607732773, + 0.0020643474999815226, + 0.010700461454689503, + 0.041354287415742874, + 0.007014526519924402, + 0.020443331450223923, + 0.006192860193550587, + 0.006832975894212723, + 0.02373683452606201, + 0.0069495500065386295, + 0.01107238419353962, + 0.026526248082518578, + 0.0076393019407987595, + 0.018860578536987305, + 0.0071473149582743645, + 0.011614374816417694, + 0.025931790471076965, + 0.007368216756731272, + 0.02351164072751999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.006734531372785568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, full-body shot captures the scene in the style of an imperial portrait, focusing on a divine warrior standing steadfast and calm. His flowing white hair flutters gently, as if in a breeze, and his chiseled, timeless face wears a calm, contemplative expression. He clads in a voluminous, multi-layered silver robe woven from a fabric that shimmers like liquid moonlight and embroidered with subtle, luminous celestial patterns. Against a backdrop of a dark, stormy sky filled with swirling, roiling clouds of deep indigo and charcoal gray, a relentless stream of shadowy projectiles, like jagged shards of pure darkness, trailing ethereal black smoke, rapidly hurtles toward him. As they approach his body, an invisible energy field repels each one; the air around the warrior subtly ripples and twists, the projectiles being smoothly and deliberately deflected, their aggressive trajectories curving into graceful arcs that curve into the darkness. The entire action unfolds without the warrior moving a muscle, his calm posture and the dramatic, calm lighting from his own glowing robes creating a powerful, epic, unshakable atmosphere of divine power amidst the chaos.", + "seed": 26018, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_13/index_1969_seed_26018_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014114431664347648, + "vlm_logprob": -0.014114431664347648, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.014114431664347648, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0027659989427775145 + ], + "opposite_logprob_tokens": [ + -6.377766132354736 + ], + "logprob_reward": -0.0027659989427775145, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0027659989427775145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19681866466999054, + 0.10232974588871002, + 0.056110650300979614, + 0.07114365696907043, + 0.05574651435017586, + 0.04403082653880119, + 0.056622158735990524, + 0.03934074938297272, + 0.29687514901161194, + 0.08098187297582626 + ], + "credit_map": [ + 0.01731973886489868, + 0.02738398313522339, + 0.016455192118883133, + 0.023959500715136528, + 0.04347950220108032, + 0.01570863276720047, + 0.030298976227641106, + 0.022213155403733253, + 0.00853682216256857, + 0.015341833233833313, + 0.008701980113983154, + 0.011225225403904915, + 0.022841980680823326, + 0.008222749456763268, + 0.016203705221414566, + 0.011255457065999508, + 0.004233676940202713, + 0.005657913628965616, + 0.00378919905051589, + 0.005154665559530258, + 0.01333689596503973, + 0.0074710845947265625, + 0.006201930344104767, + 0.010265285149216652, + 0.005106262397021055, + 0.007565192878246307, + 0.006799211259931326, + 0.0045969667844474316, + 0.01920059509575367, + 0.009198613464832306, + 0.006695886608213186, + 0.011980927549302578, + 0.00449594808742404, + 0.00455006817355752, + 0.008588721044361591, + 0.005272816400974989, + 0.009851349517703056, + 0.007181569933891296, + 0.0071161356754601, + 0.008689912036061287, + 0.0032273735851049423, + 0.00996236503124237, + 0.003022190183401108, + 0.00472868699580431, + 0.0096751619130373, + 0.0045297592878341675, + 0.004264726769179106, + 0.004620563704520464, + 0.005397465080022812, + 0.007527771405875683, + 0.005488513503223658, + 0.004607419017702341, + 0.00899431947618723, + 0.00658006826415658, + 0.005793170537799597, + 0.012233435176312923, + 0.002809712430462241, + 0.005830470472574234, + 0.0035056746564805508, + 0.004561010282486677, + 0.008956466801464558, + 0.0036514061503112316, + 0.006411756854504347, + 0.003614256391301751, + 0.03707953914999962, + 0.013442164286971092, + 0.010742186568677425, + 0.06151103973388672, + 0.016033316031098366, + 0.09064516425132751, + 0.03390612080693245, + 0.03351564332842827, + 0.006791636813431978, + 0.010149700567126274, + 0.006775890477001667, + 0.009536925703287125, + 0.0183404628187418, + 0.008838544599711895, + 0.0120966462418437, + 0.00845207367092371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0027659989427775145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002501455834135413 + ], + "opposite_logprob_tokens": [ + -6.002501487731934 + ], + "logprob_reward": -0.002501455834135413, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002501455834135413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18085917830467224, + 0.06643905490636826, + 0.12870419025421143, + 0.0676693245768547, + 0.04834016412496567, + 0.04826534911990166, + 0.045105624943971634, + 0.03247729688882828, + 0.34098541736602783, + 0.041154392063617706 + ], + "credit_map": [ + 0.016414165496826172, + 0.02112109400331974, + 0.018045809119939804, + 0.01639675348997116, + 0.0260541420429945, + 0.024199634790420532, + 0.024019954726099968, + 0.034607626497745514, + 0.006568127777427435, + 0.006936526391655207, + 0.005749742966145277, + 0.0070964801125228405, + 0.010116773657500744, + 0.008448155596852303, + 0.010103173553943634, + 0.01142007578164339, + 0.008519412018358707, + 0.00941387191414833, + 0.008454514667391777, + 0.011349250562489033, + 0.020637786015868187, + 0.021068084985017776, + 0.0184628888964653, + 0.03079839050769806, + 0.005005669314414263, + 0.005662090610712767, + 0.007251932751387358, + 0.004478374496102333, + 0.01625835709273815, + 0.00983375869691372, + 0.006566446740180254, + 0.012612695805728436, + 0.0035094302147626877, + 0.0046871318481862545, + 0.007112693041563034, + 0.00533578684553504, + 0.008013143204152584, + 0.006925428286194801, + 0.006334488280117512, + 0.006422062404453754, + 0.0038293637335300446, + 0.00680778780952096, + 0.002943955594673753, + 0.005564241204410791, + 0.011256534606218338, + 0.005531522445380688, + 0.004870956297963858, + 0.007460988592356443, + 0.005013709422200918, + 0.005985709372907877, + 0.004063208121806383, + 0.0035029796417802572, + 0.0069589694030582905, + 0.005604684818536043, + 0.004117715172469616, + 0.009858646430075169, + 0.002620352664962411, + 0.004361175000667572, + 0.003160873893648386, + 0.00308528495952487, + 0.004958296660333872, + 0.003628801554441452, + 0.0044065252877771854, + 0.006255988962948322, + 0.039956528693437576, + 0.021320663392543793, + 0.01285810861736536, + 0.06628557294607162, + 0.020208243280649185, + 0.09939379245042801, + 0.04150756075978279, + 0.039454951882362366, + 0.005016440991312265, + 0.00291991513222456, + 0.002864283975213766, + 0.003942564129829407, + 0.004235369153320789, + 0.009186379611492157, + 0.005133762490004301, + 0.007855677977204323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.002501455834135413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357341557741165 + ], + "opposite_logprob_tokens": [ + -3.651357412338257 + ], + "logprob_reward": -0.026357341557741165, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357341557741165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10557787120342255, + 0.09684279561042786, + 0.08560574054718018, + 0.08218294382095337, + 0.06487302482128143, + 0.20432786643505096, + 0.11168559640645981, + 0.07423682510852814, + 0.12275279313325882, + 0.05191456899046898 + ], + "credit_map": [ + 0.012753847055137157, + 0.013174722902476788, + 0.012666554190218449, + 0.009889231994748116, + 0.020933838561177254, + 0.009611059911549091, + 0.015578626655042171, + 0.010969987139105797, + 0.008182638324797153, + 0.010909034870564938, + 0.008055956102907658, + 0.009440070949494839, + 0.01762279123067856, + 0.011640550568699837, + 0.014826797880232334, + 0.016164954751729965, + 0.004804094787687063, + 0.006093249190598726, + 0.006128828041255474, + 0.006941635627299547, + 0.009144722484052181, + 0.01701567880809307, + 0.010844521224498749, + 0.0246330127120018, + 0.0065472847782075405, + 0.007287641055881977, + 0.009732327423989773, + 0.00666346400976181, + 0.018446560949087143, + 0.011661137454211712, + 0.009319574572145939, + 0.012524956837296486, + 0.004364104941487312, + 0.0067308335565030575, + 0.007710603531450033, + 0.008680981583893299, + 0.009920235723257065, + 0.007740359287708998, + 0.010614750906825066, + 0.009111152030527592, + 0.018331874161958694, + 0.03174610808491707, + 0.01766897737979889, + 0.021456176415085793, + 0.029771452769637108, + 0.029920611530542374, + 0.019347170367836952, + 0.036085497587919235, + 0.0160059854388237, + 0.011150573380291462, + 0.010684769600629807, + 0.009085016325116158, + 0.01255052536725998, + 0.013905980624258518, + 0.012635564431548119, + 0.025667181238532066, + 0.005407807417213917, + 0.00971835944801569, + 0.006216845940798521, + 0.008193900808691978, + 0.010618606582283974, + 0.007387843448668718, + 0.01253217551857233, + 0.014161285944283009, + 0.01242771465331316, + 0.010754229500889778, + 0.009026668034493923, + 0.025834687054157257, + 0.012818660587072372, + 0.024315472692251205, + 0.01779738999903202, + 0.00977796595543623, + 0.006109042093157768, + 0.004257287830114365, + 0.0036046088207513094, + 0.005552226677536964, + 0.0069918702356517315, + 0.010806884616613388, + 0.006802729796618223, + 0.007789920084178448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.026357341557741165 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052565219812095165 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052565219812095165, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052565219812095165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21052871644496918, + 0.06216546520590782, + 0.084776371717453, + 0.07186547666788101, + 0.04034389555454254, + 0.14503373205661774, + 0.05958490073680878, + 0.025584882125258446, + 0.2536397874355316, + 0.046476706862449646 + ], + "credit_map": [ + 0.02242383547127247, + 0.027237316593527794, + 0.024707812815904617, + 0.018612321466207504, + 0.03630329295992851, + 0.022512998431921005, + 0.028483007103204727, + 0.0302481297403574, + 0.0052598160691559315, + 0.0071096159517765045, + 0.004929340910166502, + 0.006337382365018129, + 0.012494922615587711, + 0.006393572315573692, + 0.010221153497695923, + 0.009419665671885014, + 0.005112715531140566, + 0.007814113050699234, + 0.006054417230188847, + 0.009235701523721218, + 0.01046570111066103, + 0.015112070366740227, + 0.011082185432314873, + 0.01989947259426117, + 0.005957117769867182, + 0.005920592229813337, + 0.0076085468754172325, + 0.005499834194779396, + 0.016036495566368103, + 0.01074923574924469, + 0.0072999014519155025, + 0.012793751433491707, + 0.0024680844508111477, + 0.005207815207540989, + 0.005188029259443283, + 0.004965954925864935, + 0.005538605619221926, + 0.0042153531685471535, + 0.0059949299320578575, + 0.00676512299105525, + 0.0116051584482193, + 0.024898309260606766, + 0.01397731713950634, + 0.017160527408123016, + 0.02311856299638748, + 0.017653604969382286, + 0.016684729605913162, + 0.019935522228479385, + 0.0061111124232411385, + 0.008235304616391659, + 0.006006957031786442, + 0.004621205385774374, + 0.010268301703035831, + 0.006946256849914789, + 0.004933499731123447, + 0.01246226392686367, + 0.0021905023604631424, + 0.002387635177001357, + 0.0019212636398151517, + 0.0027317337226122618, + 0.0028703093994408846, + 0.0037917618174105883, + 0.0037441772874444723, + 0.005947499070316553, + 0.030673662200570107, + 0.011977867223322392, + 0.010641110129654408, + 0.05063917115330696, + 0.014023685827851295, + 0.07609768956899643, + 0.029590921476483345, + 0.029995689168572426, + 0.004900630097836256, + 0.005686859600245953, + 0.0033966898918151855, + 0.005243135616183281, + 0.007569541223347187, + 0.0059338822029531, + 0.006783238146454096, + 0.006962731014937162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.0052565219812095165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a divine warrior with flowing white hair and a calm, contemplative expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing a voluminous, multi-layered silver robe with luminous celestial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are shadowy projectiles resembling jagged shards of darkness trailing black smoke rapidly approaching the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an invisible energy field repel the projectiles, causing them to curve into graceful arcs without the warrior moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, stormy sky filled with swirling clouds of deep indigo and charcoal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033690840005874634 + ], + "opposite_logprob_tokens": [ + -3.408690929412842 + ], + "logprob_reward": -0.033690840005874634, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033690840005874634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07561656832695007, + 0.1307045966386795, + 0.10698602348566055, + 0.26553982496261597, + 0.23038242757320404, + 0.02774617075920105, + 0.03333581984043121, + 0.02842607907950878, + 0.04084564000368118, + 0.060416869819164276 + ], + "credit_map": [ + 0.008989227004349232, + 0.008634490892291069, + 0.006671676877886057, + 0.008652355521917343, + 0.014821184799075127, + 0.00988870207220316, + 0.01058348547667265, + 0.007375448476523161, + 0.017450151965022087, + 0.01320841908454895, + 0.012261259369552135, + 0.011113733984529972, + 0.020925596356391907, + 0.022572796791791916, + 0.015400469303131104, + 0.017772171646356583, + 0.007677631918340921, + 0.005900765769183636, + 0.006494380533695221, + 0.00903923436999321, + 0.013314967975020409, + 0.023475313559174538, + 0.011898668482899666, + 0.029185064136981964, + 0.02429950423538685, + 0.020270874723792076, + 0.03302634507417679, + 0.017303666099905968, + 0.055142346769571304, + 0.04541650786995888, + 0.024638913571834564, + 0.045441657304763794, + 0.020108738914132118, + 0.017941709607839584, + 0.04284191504120827, + 0.01819070428609848, + 0.03563432767987251, + 0.04243141785264015, + 0.02573215216398239, + 0.027501458302140236, + 0.0024356436915695667, + 0.004232930485159159, + 0.002107800915837288, + 0.0025032588746398687, + 0.006000336725264788, + 0.0034833254758268595, + 0.0023057262878865004, + 0.004677148535847664, + 0.00414968142285943, + 0.0030742003582417965, + 0.004703050944954157, + 0.0023948941379785538, + 0.0034219124354422092, + 0.004143678583204746, + 0.003394874045625329, + 0.008053526282310486, + 0.0027492514345794916, + 0.0030544400215148926, + 0.002817460335791111, + 0.0024662313517183065, + 0.003887179074808955, + 0.0036652779672294855, + 0.003837920492514968, + 0.005948318168520927, + 0.004233677405864, + 0.003911127801984549, + 0.002267028670758009, + 0.008893448859453201, + 0.004746207036077976, + 0.006874347571283579, + 0.006919731851667166, + 0.0030000710394233465, + 0.0069792140275239944, + 0.006437988951802254, + 0.00408134562894702, + 0.0062347231432795525, + 0.00980408862233162, + 0.012728205882012844, + 0.007338954601436853, + 0.006812350358814001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.033690840005874634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_14/index_tensor([2422, 2422])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_14/index_tensor([2422, 2422])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0d16980d7fc446c1af085503407bc375b6f9a90d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_14/index_tensor([2422, 2422])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Captured in stunning, high-frame detail, this static, slightly low-angle, medium-close-up image focuses on an old-fashioned oscillating desk fan on a worn wooden workbench. The fan, with its off-white, slightly yellowed plastic body and chrome safety cage, rotates steadily from side to side in the dimly lit room. As it vibrates, its once-hard plastic blades (originally pure white) begin to tremble imperceptibly under the warm, focused light source just outside the frame. The blades' sharp edges soften and begin to gleam, slowly and disconcertingly drooping downward as they rotate, bending under their own weight like withered petals. The plastic transforms into a thick, translucent, honey-colored syrup. With each rotation, viscous droplets deliberately form at the drooping tips, stretching in mesmerizing macroscopic detail before detaching and falling in surreal slow motion, creating tiny, silent splashes on the fan's base and the dusty surface of the bench below, creating a strange and hypnotic atmosphere of physical impossibility.", + "seed": 31410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_14/index_2422_seed_31410_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.008894092310220003, + "vlm_logprob": -0.008894092310220003, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.008894092310220003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010606229305267334 + ], + "opposite_logprob_tokens": [ + -5.135606288909912 + ], + "logprob_reward": -0.010606229305267334, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010606229305267334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19495566189289093, + 0.15169955790042877, + 0.0682741329073906, + 0.14054174721240997, + 0.07306528091430664, + 0.1670123189687729, + 0.032777100801467896, + 0.13104817271232605, + 0.020219027996063232, + 0.020406942814588547 + ], + "credit_map": [ + 0.019597772508859634, + 0.017314914613962173, + 0.03973915055394173, + 0.014731851406395435, + 0.024318227544426918, + 0.019171828404068947, + 0.0390067994594574, + 0.021075110882520676, + 0.01412466261535883, + 0.05240093171596527, + 0.04487204924225807, + 0.009776677936315536, + 0.010710992850363255, + 0.006161308381706476, + 0.006586567033082247, + 0.007066376972943544, + 0.007342168129980564, + 0.020160343497991562, + 0.006938342936336994, + 0.00569454301148653, + 0.010569385252892971, + 0.005945548880845308, + 0.006875346880406141, + 0.004748457111418247, + 0.014397704973816872, + 0.03932733088731766, + 0.020592793822288513, + 0.012933097779750824, + 0.016095038503408432, + 0.011115233413875103, + 0.01373467780649662, + 0.012345869094133377, + 0.007184789516031742, + 0.012922707945108414, + 0.006846562027931213, + 0.005712361540645361, + 0.01760479435324669, + 0.007894548587501049, + 0.0092347776517272, + 0.0056647444143891335, + 0.013181659393012524, + 0.06802905350923538, + 0.032817136496305466, + 0.007772418204694986, + 0.027337435632944107, + 0.006375068332999945, + 0.005171784665435553, + 0.0063277725130319595, + 0.003963124006986618, + 0.004365694243460894, + 0.005881540477275848, + 0.0034783678129315376, + 0.003982329275459051, + 0.0019040537299588323, + 0.003534561488777399, + 0.0056674303486943245, + 0.010428749024868011, + 0.048940859735012054, + 0.0338982455432415, + 0.006720483768731356, + 0.01237321738153696, + 0.0036183211486786604, + 0.010070552118122578, + 0.004997743759304285, + 0.0018557258881628513, + 0.004010876640677452, + 0.0032605486921966076, + 0.0021699038334190845, + 0.0022196690551936626, + 0.0014662789180874825, + 0.002972071524709463, + 0.002263953909277916, + 0.0020224363543093204, + 0.0026351776905357838, + 0.0033603650517761707, + 0.0025928346440196037, + 0.003174013691022992, + 0.0015122272307053208, + 0.0024653824511915445, + 0.002644504653289914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.010606229305267334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008672785945236683 + ], + "opposite_logprob_tokens": [ + -4.758672714233398 + ], + "logprob_reward": -0.008672785945236683, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008672785945236683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03423784300684929, + 0.443779855966568, + 0.06310831010341644, + 0.01649746298789978, + 0.062106505036354065, + 0.06943223625421524, + 0.008783599361777306, + 0.26656052470207214, + 0.022637678310275078, + 0.012855959124863148 + ], + "credit_map": [ + 0.003908846992999315, + 0.0041413744911551476, + 0.004632391035556793, + 0.0035255125258117914, + 0.003793930634856224, + 0.0028539695776998997, + 0.006846671458333731, + 0.004535146523267031, + 0.04080896079540253, + 0.15505293011665344, + 0.1311725527048111, + 0.028286941349506378, + 0.030903570353984833, + 0.017720673233270645, + 0.01843980699777603, + 0.021394463256001472, + 0.006462108809500933, + 0.024843433871865273, + 0.014231405220925808, + 0.0031845434568822384, + 0.0062040057964622974, + 0.0028812906239181757, + 0.0029691571835428476, + 0.0023323632776737213, + 0.0017902279505506158, + 0.0028677976224571466, + 0.001504771877080202, + 0.001742887543514371, + 0.0027135512791574, + 0.002117612399160862, + 0.002123023383319378, + 0.0016375919803977013, + 0.006307190749794245, + 0.011993332765996456, + 0.006873768288642168, + 0.005303339101374149, + 0.012156033888459206, + 0.006664430722594261, + 0.008415285497903824, + 0.004393128678202629, + 0.007867474108934402, + 0.014294472523033619, + 0.011183024384081364, + 0.005497443489730358, + 0.013069487176835537, + 0.004362389445304871, + 0.007071197032928467, + 0.006086746696382761, + 0.0008448341395705938, + 0.0010867983801290393, + 0.0015265019610524178, + 0.0008189513464458287, + 0.0012998883612453938, + 0.0007237729732878506, + 0.0010442741913720965, + 0.0014385782415047288, + 0.02143033966422081, + 0.09953217953443527, + 0.06924574822187424, + 0.013225673697888851, + 0.02638554759323597, + 0.007518779952079058, + 0.019484950229525566, + 0.00973727460950613, + 0.0018659929046407342, + 0.00391230545938015, + 0.0029457849450409412, + 0.0030897518154233694, + 0.0028546880930662155, + 0.0020279178861528635, + 0.003705274313688278, + 0.002235964871942997, + 0.0010284052696079016, + 0.0018047633348032832, + 0.0014291114639490843, + 0.0010792555985972285, + 0.003694012062624097, + 0.0010553490137681365, + 0.0016158547950908542, + 0.0011492081684991717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.008672785945236683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866604968905449 + ], + "opposite_logprob_tokens": [ + -4.758666038513184 + ], + "logprob_reward": -0.00866604968905449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866604968905449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16529469192028046, + 0.115022212266922, + 0.06604062020778656, + 0.03518414869904518, + 0.12265162169933319, + 0.18815387785434723, + 0.023741185665130615, + 0.2249857783317566, + 0.041591089218854904, + 0.017334748059511185 + ], + "credit_map": [ + 0.015771200880408287, + 0.014558223076164722, + 0.036709826439619064, + 0.012096798047423363, + 0.02052016742527485, + 0.01599336974322796, + 0.03359126299619675, + 0.01605386845767498, + 0.010723593644797802, + 0.036716729402542114, + 0.030125882476568222, + 0.008091380819678307, + 0.008103428408503532, + 0.005493215750902891, + 0.007555452641099691, + 0.008212541230022907, + 0.006312170997262001, + 0.023788616061210632, + 0.018150733783841133, + 0.003648140700533986, + 0.005241112317889929, + 0.0029496564529836178, + 0.0031521462369710207, + 0.002798045752570033, + 0.003458264749497175, + 0.00896480306982994, + 0.004510522820055485, + 0.0031107200775295496, + 0.0048389434814453125, + 0.003172909142449498, + 0.003610247280448675, + 0.0035177404060959816, + 0.011659996584057808, + 0.021255159750580788, + 0.013315742835402489, + 0.009518592618405819, + 0.02660193480551243, + 0.01280742697417736, + 0.01740890182554722, + 0.01008386630564928, + 0.01600896567106247, + 0.0730435699224472, + 0.037285853177309036, + 0.009873093105852604, + 0.028045158833265305, + 0.007741166744381189, + 0.00784433912485838, + 0.00831176433712244, + 0.0036422053817659616, + 0.0030863769352436066, + 0.002765461103990674, + 0.0024331470485776663, + 0.00472588324919343, + 0.002023387933149934, + 0.0026442145463079214, + 0.002420511096715927, + 0.019826993346214294, + 0.08076164871454239, + 0.05481303483247757, + 0.011492742225527763, + 0.023347793146967888, + 0.007145575247704983, + 0.017898518592119217, + 0.009699482470750809, + 0.003377345623448491, + 0.0074387285858392715, + 0.0050231232307851315, + 0.004707899875938892, + 0.007651252206414938, + 0.003276808885857463, + 0.005196563899517059, + 0.004919370170682669, + 0.0013388274237513542, + 0.0025831342209130526, + 0.0017072211485356092, + 0.0015618379693478346, + 0.005133879370987415, + 0.001457695965655148, + 0.0020016066264361143, + 0.0015505474293604493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.00866604968905449 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786483831703663 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786483831703663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786483831703663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1473841667175293, + 0.37514135241508484, + 0.11805157363414764, + 0.02503237873315811, + 0.10988248884677887, + 0.046794719994068146, + 0.026741210371255875, + 0.07123422622680664, + 0.06931658834218979, + 0.010421225801110268 + ], + "credit_map": [ + 0.014111633412539959, + 0.025610974058508873, + 0.020809531211853027, + 0.013493710197508335, + 0.018408693373203278, + 0.01063654012978077, + 0.024924039840698242, + 0.019389064982533455, + 0.03399151563644409, + 0.1279875636100769, + 0.10755696892738342, + 0.02365850657224655, + 0.025693092495203018, + 0.015887834131717682, + 0.020235031843185425, + 0.020130889490246773, + 0.0116517823189497, + 0.04555073380470276, + 0.030433351173996925, + 0.006086799781769514, + 0.010386672802269459, + 0.005820135585963726, + 0.004241642076522112, + 0.003880460746586323, + 0.003678609151393175, + 0.005072002299129963, + 0.0025920444168150425, + 0.0021841772831976414, + 0.003923194482922554, + 0.0022125705145299435, + 0.00304439221508801, + 0.0023253897670656443, + 0.011343380436301231, + 0.024324093014001846, + 0.010985215194523335, + 0.007866510190069675, + 0.023328540846705437, + 0.010680175386369228, + 0.013303670100867748, + 0.008050910197198391, + 0.00481821782886982, + 0.013518117368221283, + 0.005355515982955694, + 0.003296721028164029, + 0.008521710522472858, + 0.0030224614311009645, + 0.004271700046956539, + 0.003990276250988245, + 0.0039026373997330666, + 0.0035534899216145277, + 0.0036091383080929518, + 0.002585778944194317, + 0.004122026264667511, + 0.002280836459249258, + 0.003316999413073063, + 0.003370307618752122, + 0.007021027151495218, + 0.022269150242209435, + 0.015127689577639103, + 0.0039885747246444225, + 0.007107634097337723, + 0.003387410892173648, + 0.008083086460828781, + 0.004249660298228264, + 0.0057313572615385056, + 0.010505642741918564, + 0.007358143571764231, + 0.008083533495664597, + 0.009387253783643246, + 0.005188465118408203, + 0.012636172585189342, + 0.01042602863162756, + 0.001118914340622723, + 0.0013817993458360434, + 0.001361767528578639, + 0.0011078509269282222, + 0.001216139062307775, + 0.0007227043970488012, + 0.0016865638317540288, + 0.0018254879396408796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.009786483831703663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738912779837847 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006738912779837847, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738912779837847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08400978893041611, + 0.48479124903678894, + 0.04822145029902458, + 0.05518258735537529, + 0.07269929349422455, + 0.08808942884206772, + 0.00970054604113102, + 0.12383314222097397, + 0.016167471185326576, + 0.017304986715316772 + ], + "credit_map": [ + 0.009513950906693935, + 0.01002376340329647, + 0.012852386571466923, + 0.006085507106035948, + 0.012117685750126839, + 0.010389503091573715, + 0.015174231491982937, + 0.007852762006223202, + 0.04403887316584587, + 0.1685342937707901, + 0.14178615808486938, + 0.03014584258198738, + 0.03361496329307556, + 0.018648799508810043, + 0.023110132664442062, + 0.024912208318710327, + 0.004855158738791943, + 0.016586702316999435, + 0.010853097774088383, + 0.0026374743320047855, + 0.004324344918131828, + 0.0024520312435925007, + 0.003260683501139283, + 0.0032519602682441473, + 0.005081967916339636, + 0.01368566695600748, + 0.009204293601214886, + 0.004278416745364666, + 0.00826861709356308, + 0.005532410927116871, + 0.005183476023375988, + 0.003947740886360407, + 0.007010309956967831, + 0.014011764898896217, + 0.007358693517744541, + 0.005313617177307606, + 0.016240933910012245, + 0.00748631963506341, + 0.009866591542959213, + 0.005411071702837944, + 0.007119910791516304, + 0.03113856166601181, + 0.022048011422157288, + 0.004645312670618296, + 0.01048625074326992, + 0.004747811704874039, + 0.0041115847416222095, + 0.003791992785409093, + 0.0010687786852940917, + 0.0013675502268597484, + 0.0014367157127708197, + 0.0010411737021058798, + 0.0013717284891754389, + 0.00066842429805547, + 0.001145529909990728, + 0.0016006457153707743, + 0.010539287701249123, + 0.04436693713068962, + 0.029549095779657364, + 0.007320523262023926, + 0.011936142109334469, + 0.003564483718946576, + 0.010492060333490372, + 0.006064625922590494, + 0.0012395285302773118, + 0.0033878120593726635, + 0.0020420742221176624, + 0.0017618890851736069, + 0.0023515485227108, + 0.0009587692329660058, + 0.0020654357504099607, + 0.0023604154121130705, + 0.0019430796382948756, + 0.0021353571210056543, + 0.0035169324837625027, + 0.0018853545188903809, + 0.0023242090828716755, + 0.0009965457720682025, + 0.0018740290543064475, + 0.0026294789277017117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.006738912779837847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in stunning, high-frame detail, this static, slightly low-angle, medium-close-up image focuses on an old-fashioned oscillating desk fan on a worn wooden workbench. The fan, with its off-white, slightly yellowed plastic body and chrome safety cage, rotates steadily from side to side in the dimly lit room. As it vibrates, its once-hard plastic blades (originally pure white) begin to tremble imperceptibly under the warm, focused light source just outside the frame. The blades' sharp edges soften and begin to gleam, slowly and disconcertingly drooping downward as they rotate, bending under their own weight like withered petals. The plastic transforms into a thick, translucent, honey-colored syrup. With each rotation, viscous droplets deliberately form at the drooping tips, stretching in mesmerizing macroscopic detail before detaching and falling in surreal slow motion, creating tiny, silent splashes on the fan's base and the dusty surface of the bench below, creating a strange and hypnotic atmosphere of physical impossibility.", + "seed": 31411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_14/index_2422_seed_31411_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.003609253070317209, + "vlm_logprob": -0.003609253070317209, + "vlm_yes_no_margin": 6.0, + "vlm_token_credit": -0.003609253070317209, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005613990128040314 + ], + "opposite_logprob_tokens": [ + -6.505613803863525 + ], + "logprob_reward": -0.005613990128040314, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.005613990128040314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13823986053466797, + 0.07913841307163239, + 0.05471331626176834, + 0.046942442655563354, + 0.18037018179893494, + 0.046347323805093765, + 0.08477050811052322, + 0.06610962003469467, + 0.19528451561927795, + 0.1080838292837143 + ], + "credit_map": [ + 0.019006645306944847, + 0.020195115357637405, + 0.01612282358109951, + 0.01368727721273899, + 0.01893724501132965, + 0.016683274880051613, + 0.01722371019423008, + 0.016383778303861618, + 0.010642753913998604, + 0.012316391803324223, + 0.006611671298742294, + 0.00739889545366168, + 0.01440359465777874, + 0.01653190888464451, + 0.006105702370405197, + 0.005127499811351299, + 0.0036288718692958355, + 0.009180155582726002, + 0.003222427796572447, + 0.003567209467291832, + 0.016608579084277153, + 0.008014907129108906, + 0.005356498993933201, + 0.005134668666869402, + 0.004592486657202244, + 0.008929980918765068, + 0.003510566893965006, + 0.004473408218473196, + 0.011551589705049992, + 0.005780912935733795, + 0.004010600037872791, + 0.00409289775416255, + 0.017963971942663193, + 0.02202676609158516, + 0.010799168609082699, + 0.015640590339899063, + 0.055799469351768494, + 0.020682010799646378, + 0.015221324749290943, + 0.02223687618970871, + 0.006090824492275715, + 0.0045745959505438805, + 0.003246191656216979, + 0.003080046037212014, + 0.013976720161736012, + 0.005949887912720442, + 0.003565158462151885, + 0.005863898899406195, + 0.010538952425122261, + 0.0078587606549263, + 0.00628351466730237, + 0.009296699427068233, + 0.022890692576766014, + 0.011864524334669113, + 0.007343208882957697, + 0.008694162592291832, + 0.007137060631066561, + 0.009147142060101032, + 0.0044016106985509396, + 0.006673098541796207, + 0.01816016435623169, + 0.009441185742616653, + 0.005878498312085867, + 0.005270865745842457, + 0.014818485826253891, + 0.05760682374238968, + 0.010711472481489182, + 0.011315112002193928, + 0.0402696430683136, + 0.036785900592803955, + 0.01322904322296381, + 0.010548027232289314, + 0.01269408781081438, + 0.013095978647470474, + 0.013255355879664421, + 0.009397254325449467, + 0.02420933172106743, + 0.0189400352537632, + 0.00799690093845129, + 0.008494877256453037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.005613990128040314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002831618534401059 + ], + "opposite_logprob_tokens": [ + -5.87783145904541 + ], + "logprob_reward": -0.002831618534401059, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002831618534401059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14018815755844116, + 0.1644776612520218, + 0.043158385902643204, + 0.06184469908475876, + 0.18565580248832703, + 0.025994136929512024, + 0.0931953564286232, + 0.027160566300153732, + 0.17830683290958405, + 0.08001840114593506 + ], + "credit_map": [ + 0.019033314660191536, + 0.018330084159970284, + 0.01857350766658783, + 0.011696569621562958, + 0.022524403408169746, + 0.015740443021059036, + 0.0167932640761137, + 0.017496570944786072, + 0.017149748280644417, + 0.03057532012462616, + 0.014676038175821304, + 0.01718159019947052, + 0.026752600446343422, + 0.03027152456343174, + 0.01325178612023592, + 0.014619049616158009, + 0.003360533621162176, + 0.007011239882558584, + 0.0037348538171499968, + 0.0033900721464306116, + 0.011702341958880424, + 0.006378860678523779, + 0.004176478832960129, + 0.0034040019381791353, + 0.0068662031553685665, + 0.007958891801536083, + 0.004126954823732376, + 0.004948899615556002, + 0.016459347680211067, + 0.007673707325011492, + 0.008432462811470032, + 0.005378232337534428, + 0.018305763602256775, + 0.025118043646216393, + 0.009194977581501007, + 0.016201963648200035, + 0.06101094186306, + 0.02202644571661949, + 0.015639103949069977, + 0.0181585643440485, + 0.002980234334245324, + 0.0030483626760542393, + 0.002473170170560479, + 0.0030128727667033672, + 0.005230541341006756, + 0.0033967127092182636, + 0.0026839084457606077, + 0.003168335882946849, + 0.011460723355412483, + 0.011202245019376278, + 0.005873457062989473, + 0.007533789146691561, + 0.028477739542722702, + 0.011298279277980328, + 0.008717800490558147, + 0.008631317876279354, + 0.002756548812612891, + 0.003931696992367506, + 0.002007101895287633, + 0.0025593063328415155, + 0.005604841746389866, + 0.004019073210656643, + 0.0025501055642962456, + 0.0037318901158869267, + 0.01608297787606716, + 0.044995587319135666, + 0.009085315279662609, + 0.008265837095677853, + 0.03982432559132576, + 0.031564924865961075, + 0.0128590427339077, + 0.01562880165874958, + 0.01570221036672592, + 0.009784446097910404, + 0.0053590224124491215, + 0.003930429928004742, + 0.016217701137065887, + 0.017453890293836594, + 0.005056113004684448, + 0.006514583248645067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.002831618534401059 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001348063233308494 + ], + "opposite_logprob_tokens": [ + -6.62634801864624 + ], + "logprob_reward": -0.001348063233308494, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001348063233308494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07372906059026718, + 0.04719119146466255, + 0.06389638036489487, + 0.03312604874372482, + 0.31158533692359924, + 0.020224055275321007, + 0.12009090930223465, + 0.020617207512259483, + 0.1970086246728897, + 0.1125311627984047 + ], + "credit_map": [ + 0.013006417080760002, + 0.005878028925508261, + 0.012307988479733467, + 0.007198586594313383, + 0.009035457856953144, + 0.008777341805398464, + 0.00838834885507822, + 0.009136888198554516, + 0.0050238692201673985, + 0.006947972811758518, + 0.0051483879797160625, + 0.004968453664332628, + 0.008757879957556725, + 0.006862972863018513, + 0.004652176983654499, + 0.004829481244087219, + 0.004431259818375111, + 0.011564655229449272, + 0.005387882236391306, + 0.006472647190093994, + 0.014145969413220882, + 0.008398503996431828, + 0.00868428684771061, + 0.0048111784271895885, + 0.0034734876826405525, + 0.004777493886649609, + 0.0019796055275946856, + 0.002351987175643444, + 0.007743225432932377, + 0.003971085883677006, + 0.0033835540525615215, + 0.005445612128823996, + 0.0301884226500988, + 0.04302186891436577, + 0.015709305182099342, + 0.026897475123405457, + 0.10102225095033646, + 0.03683633729815483, + 0.02617061324417591, + 0.03173907846212387, + 0.0026143433060497046, + 0.0022828602232038975, + 0.0017718569142743945, + 0.002359851961955428, + 0.004289946518838406, + 0.0028107238467782736, + 0.0016541482182219625, + 0.0024403254501521587, + 0.014891298487782478, + 0.012755274772644043, + 0.0077773649245500565, + 0.011583326384425163, + 0.03600533679127693, + 0.015200408175587654, + 0.011286403052508831, + 0.010591495782136917, + 0.0021511386148631573, + 0.002404439728707075, + 0.0016590050654485822, + 0.002018189523369074, + 0.004985394887626171, + 0.003097972832620144, + 0.0020263127516955137, + 0.0022747530601918697, + 0.017810383811593056, + 0.04838544502854347, + 0.010683842934668064, + 0.008356579579412937, + 0.044915180653333664, + 0.034470025449991226, + 0.015466992743313313, + 0.01692017912864685, + 0.019615691155195236, + 0.014903015457093716, + 0.0057786256074905396, + 0.0057197920978069305, + 0.027178196236491203, + 0.025167658925056458, + 0.008513111621141434, + 0.005655067507177591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.001348063233308494 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10559006035327911, + 0.09896428883075714, + 0.08629520982503891, + 0.028957920148968697, + 0.16283254325389862, + 0.12648840248584747, + 0.0597962848842144, + 0.02911609597504139, + 0.17302915453910828, + 0.12893004715442657 + ], + "credit_map": [ + 0.01727340929210186, + 0.007914029993116856, + 0.01749759539961815, + 0.012136850506067276, + 0.012598233297467232, + 0.01247151754796505, + 0.014152954332530499, + 0.01154547743499279, + 0.008254142478108406, + 0.017233604565262794, + 0.006384414620697498, + 0.01094527542591095, + 0.02300388738512993, + 0.017511898651719093, + 0.0097091905772686, + 0.005921876989305019, + 0.0057836128398776054, + 0.01651998609304428, + 0.006786361802369356, + 0.010057132691144943, + 0.015810325741767883, + 0.011913754977285862, + 0.01261142548173666, + 0.0068126157857477665, + 0.0028068888932466507, + 0.005941973999142647, + 0.0014171975199133158, + 0.003131269244477153, + 0.005789679940789938, + 0.004128155764192343, + 0.003371583530679345, + 0.002371174516156316, + 0.01530237216502428, + 0.023688463494181633, + 0.009037203155457973, + 0.014379609376192093, + 0.052351221442222595, + 0.019086742773652077, + 0.01440527755767107, + 0.01458165142685175, + 0.01805974915623665, + 0.01958918385207653, + 0.008893362246453762, + 0.011612989939749241, + 0.03268808871507645, + 0.01566815748810768, + 0.009247515350580215, + 0.010729363188147545, + 0.005041450262069702, + 0.00916159339249134, + 0.0036815109197050333, + 0.006090578623116016, + 0.014852216467261314, + 0.0075339642353355885, + 0.0081078065559268, + 0.005327161401510239, + 0.0034722061827778816, + 0.0030389358289539814, + 0.0023344748187810183, + 0.0024049857165664434, + 0.007306314073503017, + 0.003870051819831133, + 0.002661596518009901, + 0.004027531016618013, + 0.016852760687470436, + 0.035632967948913574, + 0.010107542388141155, + 0.009586760774254799, + 0.036256078630685806, + 0.027671724557876587, + 0.014549016021192074, + 0.022372305393218994, + 0.01683824695646763, + 0.01703602448105812, + 0.009145280346274376, + 0.0070354831404984, + 0.03395749256014824, + 0.02719627134501934, + 0.010621009394526482, + 0.007100245915353298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.0036145609337836504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046380325220525265 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.0046380325220525265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046380325220525265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322081834077835, + 0.0907917246222496, + 0.06343693286180496, + 0.03869752213358879, + 0.3904062807559967, + 0.04905130714178085, + 0.017743712291121483, + 0.02422424778342247, + 0.11955832690000534, + 0.07388179004192352 + ], + "credit_map": [ + 0.019845938310027122, + 0.020593132823705673, + 0.014283870346844196, + 0.011197912506759167, + 0.023575032129883766, + 0.016274575144052505, + 0.013686521910130978, + 0.012751206755638123, + 0.009495463222265244, + 0.02058175392448902, + 0.008266700431704521, + 0.006829905789345503, + 0.013478281907737255, + 0.017114650458097458, + 0.0057630352675914764, + 0.009261936880648136, + 0.00432029040530324, + 0.01188688725233078, + 0.004284766968339682, + 0.004725987557321787, + 0.018166912719607353, + 0.01008113194257021, + 0.006477950140833855, + 0.0034930079709738493, + 0.003750996431335807, + 0.00825819093734026, + 0.002680729376152158, + 0.003423080313950777, + 0.006789016537368298, + 0.005868328735232353, + 0.003854044945910573, + 0.00407313322648406, + 0.03989369422197342, + 0.04930839687585831, + 0.02180471643805504, + 0.034386634826660156, + 0.12346500903367996, + 0.04774333909153938, + 0.029739810153841972, + 0.04406467452645302, + 0.0065217409282922745, + 0.007420953828841448, + 0.004231767728924751, + 0.004545012954622507, + 0.010588277131319046, + 0.0058220140635967255, + 0.0037269527092576027, + 0.006194588262587786, + 0.001601261319592595, + 0.002913559088483453, + 0.0015273633180186152, + 0.0028166614938527346, + 0.003012242494150996, + 0.002143033780157566, + 0.002145718317478895, + 0.0015838724793866277, + 0.002329368144273758, + 0.003488508053123951, + 0.001783801824785769, + 0.002641396364197135, + 0.005103573203086853, + 0.003596455790102482, + 0.0021833598148077726, + 0.0030977828428149223, + 0.01119877491146326, + 0.028497794643044472, + 0.005715429317206144, + 0.005229415837675333, + 0.027489542961120605, + 0.021563246846199036, + 0.008077112026512623, + 0.011787008494138718, + 0.014847875572741032, + 0.007790224626660347, + 0.005196470767259598, + 0.0037266130093485117, + 0.014299937523901463, + 0.01669790968298912, + 0.004322395194321871, + 0.007000366225838661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0046380325220525265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in stunning, high-frame detail, this static, slightly low-angle, medium-close-up image focuses on an old-fashioned oscillating desk fan on a worn wooden workbench. The fan, with its off-white, slightly yellowed plastic body and chrome safety cage, rotates steadily from side to side in the dimly lit room. As it vibrates, its once-hard plastic blades (originally pure white) begin to tremble imperceptibly under the warm, focused light source just outside the frame. The blades' sharp edges soften and begin to gleam, slowly and disconcertingly drooping downward as they rotate, bending under their own weight like withered petals. The plastic transforms into a thick, translucent, honey-colored syrup. With each rotation, viscous droplets deliberately form at the drooping tips, stretching in mesmerizing macroscopic detail before detaching and falling in surreal slow motion, creating tiny, silent splashes on the fan's base and the dusty surface of the bench below, creating a strange and hypnotic atmosphere of physical impossibility.", + "seed": 31412, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_14/index_2422_seed_31412_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.005159719381481409, + "vlm_logprob": -0.005159719381481409, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.005159719381481409, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00889045000076294 + ], + "opposite_logprob_tokens": [ + -5.633890628814697 + ], + "logprob_reward": -0.00889045000076294, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00889045000076294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17750543355941772, + 0.07760588079690933, + 0.02872600220143795, + 0.19951948523521423, + 0.05962233617901802, + 0.021295834332704544, + 0.25792765617370605, + 0.10378266870975494, + 0.0301040168851614, + 0.043910760432481766 + ], + "credit_map": [ + 0.01137769129127264, + 0.027081158012151718, + 0.02294798754155636, + 0.024663735181093216, + 0.03783004730939865, + 0.01625547558069229, + 0.017879126593470573, + 0.01947021670639515, + 0.005376835353672504, + 0.016893014311790466, + 0.011486384086310863, + 0.013583280146121979, + 0.010313003323972225, + 0.005651412531733513, + 0.005747889168560505, + 0.008554064668715, + 0.0014220622833818197, + 0.005821194965392351, + 0.0024328248109668493, + 0.003168466268107295, + 0.006208898965269327, + 0.002212817082181573, + 0.002432670211419463, + 0.005027069244533777, + 0.006957391742616892, + 0.01808254048228264, + 0.011512182652950287, + 0.014887375757098198, + 0.0689944252371788, + 0.009553715586662292, + 0.01016221009194851, + 0.05936964601278305, + 0.005620259791612625, + 0.008604222908616066, + 0.010233404114842415, + 0.008857516571879387, + 0.0073659769259393215, + 0.005639175418764353, + 0.0071264454163610935, + 0.006175338290631771, + 0.001387562952004373, + 0.0033552218228578568, + 0.0028622900135815144, + 0.002781463088467717, + 0.003854179522022605, + 0.0018092014361172915, + 0.002655493561178446, + 0.002590422285720706, + 0.024692196398973465, + 0.017841584980487823, + 0.02703351341187954, + 0.022395147010684013, + 0.06335795670747757, + 0.025036200881004333, + 0.029640773311257362, + 0.047930262982845306, + 0.015337108634412289, + 0.014056351967155933, + 0.012718726880848408, + 0.009836554527282715, + 0.0171393770724535, + 0.009855888783931732, + 0.01350858062505722, + 0.011330078355967999, + 0.002186513040214777, + 0.007955743931233883, + 0.003295833943411708, + 0.003999700769782066, + 0.003967615310102701, + 0.0028224841225892305, + 0.0029169933404773474, + 0.0029591333586722612, + 0.002309957053512335, + 0.008218067698180676, + 0.004235280677676201, + 0.006783855147659779, + 0.011756625957787037, + 0.0029178098775446415, + 0.0033627215307205915, + 0.004326443187892437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.00889045000076294 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00676318584010005 + ], + "opposite_logprob_tokens": [ + -5.006762981414795 + ], + "logprob_reward": -0.00676318584010005, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00676318584010005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1962515115737915, + 0.11322597414255142, + 0.10333254188299179, + 0.11450128257274628, + 0.09041199833154678, + 0.05052595213055611, + 0.1676788181066513, + 0.07434037327766418, + 0.05136749520897865, + 0.03836408630013466 + ], + "credit_map": [ + 0.012010677717626095, + 0.030295168980956078, + 0.028358660638332367, + 0.021248342469334602, + 0.044911667704582214, + 0.016577810049057007, + 0.020065709948539734, + 0.022783469408750534, + 0.006571175996214151, + 0.022200940176844597, + 0.00934579223394394, + 0.023827403783798218, + 0.01403362862765789, + 0.012737187556922436, + 0.01350080594420433, + 0.011009033769369125, + 0.004709762055426836, + 0.01386389508843422, + 0.007825949229300022, + 0.007652721367776394, + 0.02905884012579918, + 0.0074662696570158005, + 0.00673579890280962, + 0.02601928822696209, + 0.008840715512633324, + 0.03327469155192375, + 0.013315428979694843, + 0.013459913432598114, + 0.01577400229871273, + 0.010381873697042465, + 0.010965344496071339, + 0.00848930049687624, + 0.00865154154598713, + 0.022430192679166794, + 0.013581951148808002, + 0.013143307529389858, + 0.009153076447546482, + 0.007331580854952335, + 0.007517999969422817, + 0.008602341637015343, + 0.0044205524027347565, + 0.005101115442812443, + 0.006910017225891352, + 0.0076835365034639835, + 0.010311083868145943, + 0.0039252289570868015, + 0.003854867070913315, + 0.008319548331201077, + 0.01682209223508835, + 0.012273094616830349, + 0.017363369464874268, + 0.014434438198804855, + 0.041882243007421494, + 0.016012433916330338, + 0.017663517966866493, + 0.03122762404382229, + 0.008650503121316433, + 0.010620657354593277, + 0.00773767102509737, + 0.007341114804148674, + 0.012493529357016087, + 0.006977396085858345, + 0.010729864239692688, + 0.009789632633328438, + 0.0034377756528556347, + 0.013783697970211506, + 0.006090978160500526, + 0.007477860897779465, + 0.0072648427449166775, + 0.003745031775906682, + 0.003718286519870162, + 0.005849020555615425, + 0.002243140246719122, + 0.006777877453714609, + 0.004044091794639826, + 0.0063231284730136395, + 0.010385546833276749, + 0.002702099271118641, + 0.0028062080964446068, + 0.0030819943640381098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.00676318584010005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001955146435648203 + ], + "opposite_logprob_tokens": [ + -6.251955032348633 + ], + "logprob_reward": -0.001955146435648203, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001955146435648203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10909131169319153, + 0.1446595937013626, + 0.09897914528846741, + 0.2213865965604782, + 0.041194748133420944, + 0.06234727054834366, + 0.10637587308883667, + 0.028948258608579636, + 0.1318213790655136, + 0.05519586801528931 + ], + "credit_map": [ + 0.007921393029391766, + 0.023288680240511894, + 0.012331326492130756, + 0.013579841703176498, + 0.024907654151320457, + 0.007770891301333904, + 0.007073358166962862, + 0.012218164280056953, + 0.00692661851644516, + 0.03145090118050575, + 0.013374017551541328, + 0.04049476236104965, + 0.01960468478500843, + 0.009224473498761654, + 0.013952941633760929, + 0.009631188586354256, + 0.005343429744243622, + 0.020617123693227768, + 0.014460686594247818, + 0.010599842295050621, + 0.01665237918496132, + 0.011156558059155941, + 0.010744122788310051, + 0.009405000135302544, + 0.008000891655683517, + 0.024907300248742104, + 0.013133110478520393, + 0.01890961080789566, + 0.07179843634366989, + 0.011042305268347263, + 0.012780147604644299, + 0.06081479787826538, + 0.0029986239969730377, + 0.0065377335995435715, + 0.0072284117341041565, + 0.006136440671980381, + 0.006354542914777994, + 0.0034541303757578135, + 0.004446741659194231, + 0.004038127604871988, + 0.0032818103209137917, + 0.004970775917172432, + 0.006089011672884226, + 0.008499951101839542, + 0.015849996358156204, + 0.0038294722326099873, + 0.0050293453969061375, + 0.014796910807490349, + 0.009606615640223026, + 0.008105364628136158, + 0.010967637412250042, + 0.011937994509935379, + 0.024674121290445328, + 0.009809582494199276, + 0.013526293449103832, + 0.017748264595866203, + 0.003376912558451295, + 0.005833125207573175, + 0.003180047031491995, + 0.003318813629448414, + 0.0037283501587808132, + 0.0027560319285839796, + 0.0026793302968144417, + 0.0040756468661129475, + 0.007993482053279877, + 0.03670046105980873, + 0.016224505379796028, + 0.020604515448212624, + 0.018739869818091393, + 0.010191424749791622, + 0.01071250718086958, + 0.010654611513018608, + 0.003209848888218403, + 0.011029117740690708, + 0.004645229317247868, + 0.008764668367803097, + 0.014524522237479687, + 0.003963768482208252, + 0.003889797255396843, + 0.0051689138635993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.001955146435648203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094432573765516 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094432573765516, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094432573765516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15063759684562683, + 0.06761427223682404, + 0.13755232095718384, + 0.11073974519968033, + 0.12445487827062607, + 0.05775810033082962, + 0.20772594213485718, + 0.03959706798195839, + 0.03676006942987442, + 0.0671599954366684 + ], + "credit_map": [ + 0.011336036026477814, + 0.023513268679380417, + 0.02015812136232853, + 0.0279143787920475, + 0.021819427609443665, + 0.012073005549609661, + 0.017004216089844704, + 0.016819152981042862, + 0.004872375633567572, + 0.017605600878596306, + 0.0059085567481815815, + 0.009344321675598621, + 0.007957274094223976, + 0.008296597748994827, + 0.007020547986030579, + 0.006608998402953148, + 0.007568173576146364, + 0.020775914192199707, + 0.011867394670844078, + 0.013517441228032112, + 0.034298744052648544, + 0.010285765863955021, + 0.01313108205795288, + 0.026107804849743843, + 0.006150467786937952, + 0.023813463747501373, + 0.010928604751825333, + 0.009783510118722916, + 0.027322936803102493, + 0.007018385920673609, + 0.005055098328739405, + 0.02066728286445141, + 0.011898921802639961, + 0.01646393910050392, + 0.02012033388018608, + 0.01807558164000511, + 0.014208664186298847, + 0.012975133024156094, + 0.017944522202014923, + 0.012767776846885681, + 0.006094502285122871, + 0.007146446034312248, + 0.012322002090513706, + 0.008550064638257027, + 0.00995759665966034, + 0.0036176752764731646, + 0.004697970114648342, + 0.005371843930333853, + 0.017152197659015656, + 0.01416490226984024, + 0.020570727065205574, + 0.01627369411289692, + 0.05394783616065979, + 0.018245676532387733, + 0.025253133848309517, + 0.04211777448654175, + 0.005406047683209181, + 0.004024284891784191, + 0.005128258373588324, + 0.004705092404037714, + 0.0061742523685097694, + 0.003590347245335579, + 0.005838671699166298, + 0.0047301133163273335, + 0.0028163010720163584, + 0.01120488066226244, + 0.004548698663711548, + 0.004064721520990133, + 0.005185296293348074, + 0.002551222452893853, + 0.0025599573273211718, + 0.003828989341855049, + 0.0038323896005749702, + 0.014288348145782948, + 0.006269386969506741, + 0.012045162729918957, + 0.015357116237282753, + 0.0044173551723361015, + 0.004881297238171101, + 0.00606894725933671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.004094432573765516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095382057130337 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095382057130337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095382057130337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07242284715175629, + 0.06808406114578247, + 0.024367639794945717, + 0.1420188993215561, + 0.044941991567611694, + 0.09392470121383667, + 0.45137888193130493, + 0.041679199784994125, + 0.04983959719538689, + 0.011342188343405724 + ], + "credit_map": [ + 0.004217241425067186, + 0.01595279388129711, + 0.008163620717823505, + 0.014490899629890919, + 0.01168903335928917, + 0.0048673017881810665, + 0.005326281301677227, + 0.007715675514191389, + 0.004641957115381956, + 0.017525652423501015, + 0.007891254499554634, + 0.01070460956543684, + 0.008803087286651134, + 0.006850752048194408, + 0.0047104801051318645, + 0.006956270895898342, + 0.001682196045294404, + 0.004538135603070259, + 0.0030666051898151636, + 0.003466167487204075, + 0.004108179826289415, + 0.002557984320446849, + 0.001898542046546936, + 0.0030498295091092587, + 0.006769127678126097, + 0.026649540290236473, + 0.00946084875613451, + 0.010614345781505108, + 0.0388125404715538, + 0.007990261539816856, + 0.009223723784089088, + 0.03249850124120712, + 0.0034075933508574963, + 0.00921662524342537, + 0.006653401535004377, + 0.007352405693382025, + 0.006762216798961163, + 0.0036803779657930136, + 0.004177785478532314, + 0.0036915852688252926, + 0.0049384525045752525, + 0.010082046501338482, + 0.010602423921227455, + 0.009020784869790077, + 0.023781636729836464, + 0.006378582678735256, + 0.00819542445242405, + 0.020925357937812805, + 0.04089796170592308, + 0.03233965486288071, + 0.0452260822057724, + 0.04018688574433327, + 0.10821397602558136, + 0.04581340774893761, + 0.056919023394584656, + 0.08178187161684036, + 0.007195322308689356, + 0.0049903616309165955, + 0.006087210960686207, + 0.004498637281358242, + 0.007113483268767595, + 0.003779605496674776, + 0.0031754306983202696, + 0.004839148838073015, + 0.002662368817254901, + 0.011952371336519718, + 0.005878667812794447, + 0.008266155607998371, + 0.006905789487063885, + 0.0032504266127943993, + 0.0042047989554703236, + 0.006719019263982773, + 0.0009350551990792155, + 0.0017284192144870758, + 0.0015446365578100085, + 0.0016537740593776107, + 0.0018605772638693452, + 0.001072809798642993, + 0.0013783618342131376, + 0.0011685551144182682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.004095382057130337 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in stunning, high-frame detail, this static, slightly low-angle, medium-close-up image focuses on an old-fashioned oscillating desk fan on a worn wooden workbench. The fan, with its off-white, slightly yellowed plastic body and chrome safety cage, rotates steadily from side to side in the dimly lit room. As it vibrates, its once-hard plastic blades (originally pure white) begin to tremble imperceptibly under the warm, focused light source just outside the frame. The blades' sharp edges soften and begin to gleam, slowly and disconcertingly drooping downward as they rotate, bending under their own weight like withered petals. The plastic transforms into a thick, translucent, honey-colored syrup. With each rotation, viscous droplets deliberately form at the drooping tips, stretching in mesmerizing macroscopic detail before detaching and falling in surreal slow motion, creating tiny, silent splashes on the fan's base and the dusty surface of the bench below, creating a strange and hypnotic atmosphere of physical impossibility.", + "seed": 31413, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_14/index_2422_seed_31413_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.004905362147837877, + "vlm_logprob": -0.004905362147837877, + "vlm_yes_no_margin": 5.6, + "vlm_token_credit": -0.004905362147837877, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00845152884721756 + ], + "opposite_logprob_tokens": [ + -5.758451461791992 + ], + "logprob_reward": -0.00845152884721756, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00845152884721756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08855491876602173, + 0.08881993591785431, + 0.033594515174627304, + 0.07590056955814362, + 0.1257704645395279, + 0.044475480914115906, + 0.06373819708824158, + 0.29856836795806885, + 0.039256494492292404, + 0.1413210928440094 + ], + "credit_map": [ + 0.00661756144836545, + 0.008145221509039402, + 0.01353073213249445, + 0.004935780074447393, + 0.011150214821100235, + 0.013221900910139084, + 0.005804288201034069, + 0.02514921873807907, + 0.006541952956467867, + 0.00680004246532917, + 0.005452618934214115, + 0.00495415972545743, + 0.03218282386660576, + 0.010408101603388786, + 0.004736275412142277, + 0.01774395816028118, + 0.003573667025193572, + 0.0037057832814753056, + 0.0037982615176588297, + 0.0021547337528318167, + 0.004856630694121122, + 0.00544206565245986, + 0.00236971746198833, + 0.007693654391914606, + 0.010933833196759224, + 0.012057633139193058, + 0.007854800671339035, + 0.005968982819467783, + 0.011826610192656517, + 0.0067648813128471375, + 0.005371124017983675, + 0.015122715383768082, + 0.01357175037264824, + 0.01805247738957405, + 0.01718536578118801, + 0.008029303513467312, + 0.018080491572618484, + 0.015620261430740356, + 0.013272787444293499, + 0.021958013996481895, + 0.00396219315007329, + 0.005866647232323885, + 0.004197864327579737, + 0.003774243174120784, + 0.0084711704403162, + 0.006222105585038662, + 0.0058324034325778484, + 0.00614885613322258, + 0.006018562242388725, + 0.0077603962272405624, + 0.007569809909909964, + 0.0043833376839756966, + 0.012885980308055878, + 0.007646289188414812, + 0.004821492359042168, + 0.012652333825826645, + 0.01219404861330986, + 0.018837690353393555, + 0.012848987244069576, + 0.012224266305565834, + 0.11623769998550415, + 0.027296895161271095, + 0.014181547798216343, + 0.08474723994731903, + 0.0034320191480219364, + 0.003391921753063798, + 0.00535799702629447, + 0.003041256917640567, + 0.007130458950996399, + 0.007101747207343578, + 0.0036697082687169313, + 0.006131387781351805, + 0.008401368744671345, + 0.017295043915510178, + 0.014240561984479427, + 0.01855679601430893, + 0.03099457360804081, + 0.019248969852924347, + 0.01628943346440792, + 0.016294356435537338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.00845152884721756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032062334939837456 + ], + "opposite_logprob_tokens": [ + -5.753206253051758 + ], + "logprob_reward": -0.0032062334939837456, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032062334939837456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15955308079719543, + 0.06302794069051743, + 0.08794644474983215, + 0.10986866801977158, + 0.18487882614135742, + 0.0235879085958004, + 0.054597433656454086, + 0.20108123123645782, + 0.04041365161538124, + 0.07504481077194214 + ], + "credit_map": [ + 0.009888719767332077, + 0.010536367073655128, + 0.019493723288178444, + 0.007057409267872572, + 0.019902221858501434, + 0.03968410938978195, + 0.011345051229000092, + 0.041645485907793045, + 0.005789397284388542, + 0.0060847001150250435, + 0.006003523711115122, + 0.005306574981659651, + 0.016531407833099365, + 0.007040770258754492, + 0.0066713388077914715, + 0.009600235149264336, + 0.006123576778918505, + 0.009771117009222507, + 0.016033941879868507, + 0.00792459212243557, + 0.009387659840285778, + 0.007700982969254255, + 0.006071144249290228, + 0.02493344061076641, + 0.016120631247758865, + 0.016837842762470245, + 0.0111469067633152, + 0.008875684812664986, + 0.017553644254803658, + 0.010939344763755798, + 0.008104871958494186, + 0.020289747044444084, + 0.02033686824142933, + 0.02709132246673107, + 0.025174103677272797, + 0.011350363492965698, + 0.027162697166204453, + 0.022495603188872337, + 0.017731111496686935, + 0.03353675454854965, + 0.0022696489468216896, + 0.0027853979263454676, + 0.0026584495790302753, + 0.0020152085926383734, + 0.00445034634321928, + 0.0027192747220396996, + 0.003893575631082058, + 0.0027960084844380617, + 0.005019506439566612, + 0.0080728055909276, + 0.0071706511080265045, + 0.005973908584564924, + 0.008967315778136253, + 0.004771070089191198, + 0.005055755842477083, + 0.009566421620547771, + 0.009102695621550083, + 0.012772935442626476, + 0.009579367004334927, + 0.00850349385291338, + 0.07698594033718109, + 0.01899692602455616, + 0.010263348929584026, + 0.05487654358148575, + 0.0033413192722946405, + 0.0037944577634334564, + 0.003798497375100851, + 0.0035895942710340023, + 0.008166761137545109, + 0.007370934356004, + 0.00356923695653677, + 0.006782849784940481, + 0.006332424469292164, + 0.00750910397619009, + 0.006169687956571579, + 0.016485029831528664, + 0.01111441757529974, + 0.00830028485506773, + 0.009416685439646244, + 0.009717182256281376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0032062334939837456 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028292411006987095 + ], + "opposite_logprob_tokens": [ + -5.877829074859619 + ], + "logprob_reward": -0.0028292411006987095, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028292411006987095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17672769725322723, + 0.07990080118179321, + 0.06837601214647293, + 0.06997209787368774, + 0.08257713913917542, + 0.03251199051737785, + 0.0308462455868721, + 0.24465207755565643, + 0.014392760582268238, + 0.20004324615001678 + ], + "credit_map": [ + 0.012728255242109299, + 0.015697430819272995, + 0.02130923978984356, + 0.01045655645430088, + 0.04589264839887619, + 0.029920537024736404, + 0.010778164491057396, + 0.029944850131869316, + 0.007028589025139809, + 0.0066397348418831825, + 0.005489904899150133, + 0.004716489464044571, + 0.028051462024450302, + 0.008966458030045033, + 0.0048880986869335175, + 0.014120063744485378, + 0.005335537251085043, + 0.00592644140124321, + 0.01225497480481863, + 0.0033375306520611048, + 0.011727755889296532, + 0.004549519624561071, + 0.0030155868735164404, + 0.022228660061955452, + 0.009786617942154408, + 0.009630756452679634, + 0.0077172331511974335, + 0.008575246669352055, + 0.009882570244371891, + 0.006114891730248928, + 0.004866485018283129, + 0.013398290611803532, + 0.009215201251208782, + 0.01226913183927536, + 0.011233470402657986, + 0.005254318937659264, + 0.011942991986870766, + 0.010034698061645031, + 0.007905418053269386, + 0.014721906743943691, + 0.0030092899687588215, + 0.004456889349967241, + 0.0030172737315297127, + 0.0029266246128827333, + 0.007197065744549036, + 0.003463056404143572, + 0.005118494853377342, + 0.0033232939895242453, + 0.0028461653273552656, + 0.00544997863471508, + 0.003813433228060603, + 0.003121789311990142, + 0.0055327811278402805, + 0.0028865330386906862, + 0.002738681621849537, + 0.004456879571080208, + 0.010322236455976963, + 0.015455755405128002, + 0.00983278825879097, + 0.00961758941411972, + 0.0953894555568695, + 0.021847639232873917, + 0.011432315222918987, + 0.0707542896270752, + 0.0013771457597613335, + 0.0014884364791214466, + 0.0015764366835355759, + 0.0012175784213468432, + 0.0024314008187502623, + 0.00284545193426311, + 0.0014370718272402883, + 0.0020192372612655163, + 0.014321569353342056, + 0.024764880537986755, + 0.01674051024019718, + 0.028022268787026405, + 0.043585408478975296, + 0.023984035477042198, + 0.02659936435520649, + 0.022025195881724358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.0028292411006987095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092770162969828 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092770162969828, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092770162969828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14108572900295258, + 0.10178142786026001, + 0.06099981442093849, + 0.0723327100276947, + 0.10113558173179626, + 0.06624787300825119, + 0.04192783311009407, + 0.1089189201593399, + 0.022607849910855293, + 0.2829623222351074 + ], + "credit_map": [ + 0.011405025608837605, + 0.00850581843405962, + 0.013277143239974976, + 0.006885072216391563, + 0.024596987292170525, + 0.03561221808195114, + 0.007608969695866108, + 0.03319448605179787, + 0.01086145918816328, + 0.009358295239508152, + 0.01040177047252655, + 0.008535592816770077, + 0.027226680889725685, + 0.011701849289238453, + 0.010465405881404877, + 0.013230359181761742, + 0.0050582620315253735, + 0.007180844899266958, + 0.007525580003857613, + 0.004819195717573166, + 0.01155057456344366, + 0.005766765214502811, + 0.004408990498632193, + 0.014689592644572258, + 0.009085147641599178, + 0.011182457208633423, + 0.006588537711650133, + 0.007717346306890249, + 0.011389154009521008, + 0.007455133367329836, + 0.005549952853471041, + 0.013364965096116066, + 0.010498552583158016, + 0.013183887116611004, + 0.012636516243219376, + 0.006993626710027456, + 0.01431167870759964, + 0.01350174006074667, + 0.013307875022292137, + 0.01670169271528721, + 0.006382760126143694, + 0.0069937086664140224, + 0.006783244665712118, + 0.00557943107560277, + 0.013577839359641075, + 0.010006665252149105, + 0.008943615481257439, + 0.00798059906810522, + 0.004099109210073948, + 0.007524269167333841, + 0.005124656017869711, + 0.004206203855574131, + 0.0066365632228553295, + 0.004585674498230219, + 0.0055237882770597935, + 0.00422756327316165, + 0.005555034149438143, + 0.006641420070081949, + 0.006069710478186607, + 0.004689125809818506, + 0.04083521291613579, + 0.009315651841461658, + 0.0051518711261451244, + 0.030660880729556084, + 0.0024388672318309546, + 0.002130580134689808, + 0.0020560109987854958, + 0.0019993989262729883, + 0.005097344983369112, + 0.00400159414857626, + 0.0023512805346399546, + 0.0025327696930617094, + 0.01889454759657383, + 0.03347751870751381, + 0.02621631510555744, + 0.03795121982693672, + 0.06121718883514404, + 0.03511577099561691, + 0.03844248503446579, + 0.031647276133298874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.004092770162969828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an old-fashioned oscillating desk fan on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the fan have an off-white, slightly yellowed plastic body with a chrome safety cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fan blades transform from hard plastic into a thick, translucent, honey-colored syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do viscous droplets form at the drooping tips of the syrup blades and fall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly low-angle, medium-close-up view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947037134319544 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947037134319544, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947037134319544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17397171258926392, + 0.1224614605307579, + 0.07885067909955978, + 0.0626184493303299, + 0.03697335347533226, + 0.08194395154714584, + 0.07133428007364273, + 0.15082380175590515, + 0.041563089936971664, + 0.17945913970470428 + ], + "credit_map": [ + 0.01155775599181652, + 0.0105575667694211, + 0.02101396583020687, + 0.006788739003241062, + 0.014398557133972645, + 0.042767029255628586, + 0.0081993592903018, + 0.05868874490261078, + 0.010614623315632343, + 0.011358389630913734, + 0.008227409794926643, + 0.006679989863187075, + 0.04012583941221237, + 0.017981508746743202, + 0.008015941828489304, + 0.01945776492357254, + 0.006169953383505344, + 0.007865623570978642, + 0.008213138207793236, + 0.0038796078879386187, + 0.018663324415683746, + 0.008128195069730282, + 0.005618411116302013, + 0.020312435925006866, + 0.006106792949140072, + 0.006613109726458788, + 0.007702788338065147, + 0.011106082238256931, + 0.00786276813596487, + 0.00720494519919157, + 0.006773738656193018, + 0.009248223155736923, + 0.004881417844444513, + 0.006284560076892376, + 0.003318053437396884, + 0.002721854019910097, + 0.004783961921930313, + 0.0066720894537866116, + 0.003894893219694495, + 0.004416524898260832, + 0.008355067111551762, + 0.006660923361778259, + 0.009378517977893353, + 0.00930030271410942, + 0.010144312866032124, + 0.013836724683642387, + 0.009797688573598862, + 0.014470416121184826, + 0.007647041697055101, + 0.008612360805273056, + 0.00852159783244133, + 0.005009019747376442, + 0.013085531070828438, + 0.007574934512376785, + 0.007981415838003159, + 0.012902378104627132, + 0.0064427414909005165, + 0.009770920500159264, + 0.0071221329271793365, + 0.0070485700853168964, + 0.05617232993245125, + 0.016321290284395218, + 0.0078514339402318, + 0.040094394236803055, + 0.004751300439238548, + 0.003975654486566782, + 0.0033708682749420404, + 0.0034025562927126884, + 0.010152787901461124, + 0.006979894358664751, + 0.003978715743869543, + 0.004951317794620991, + 0.015941835939884186, + 0.01884806528687477, + 0.020839734002947807, + 0.03325524926185608, + 0.03163350746035576, + 0.01942087896168232, + 0.018653379753232002, + 0.020866505801677704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.005947037134319544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_15/index_tensor([2609, 2609])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_15/index_tensor([2609, 2609])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a820597eba76425376a95b562c54e43143ce98ee --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_15/index_tensor([2609, 2609])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful, dynamic swooping shot, the camera rapidly descends from a high vantage point and zooms forward, closely following the protagonist, Hodaka. Under a damp, neon-dappled city overpass at night, Hodaka, his short, neatly cut black hair ruffled by the wind, is at the controls of a sleek, heavy-duty black motorcycle. Clad in a fitted black leather motorcycle jacket, his features chiseled, his brow furrowed, his gaze fixed piercingly ahead. Suddenly, in a desperate attempt to control his loss of control, he leans back and slams down the brakes with all his might, his feet practically touching the ground to maintain his balance. This decisive, violent movement causes the motorcycle's rear wheel to lock, sending the vehicle sliding sideways at a precarious angle across the rough asphalt. The intense friction between the tires and the road creates a long, dazzling streak of orange-red sparks, instantly illuminating his focused, tense profile and the motorcycle's cold metal components. The entire picture is filled with high-contrast light and shadow, with the interweaving of dim street lights and dazzling car lights, rendering the sharp dynamics and critical tension of modern martial arts style to the extreme.", + "seed": 17041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_15/index_2609_seed_17041_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009786979318596423, + "vlm_logprob": -0.009786979318596423, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.009786979318596423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026922861114144325 + ], + "opposite_logprob_tokens": [ + -3.7769229412078857 + ], + "logprob_reward": -0.026922861114144325, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026922861114144325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07064533978700638, + 0.047136563807725906, + 0.06963831931352615, + 0.41506850719451904, + 0.0765625610947609, + 0.04509711265563965, + 0.030854618176817894, + 0.054525405168533325, + 0.0904621109366417, + 0.100009486079216 + ], + "credit_map": [ + 0.009116172790527344, + 0.008666690438985825, + 0.011388902552425861, + 0.009535172022879124, + 0.007035210728645325, + 0.00840881559997797, + 0.007559375371783972, + 0.008935000747442245, + 0.003799854312092066, + 0.006729160901159048, + 0.00665672030299902, + 0.006177498959004879, + 0.01019754633307457, + 0.0044803195632994175, + 0.004572154488414526, + 0.004523306619375944, + 0.006140061188489199, + 0.011483229696750641, + 0.009769649244844913, + 0.008975925855338573, + 0.011799287050962448, + 0.010289305821061134, + 0.006164728198200464, + 0.005016134586185217, + 0.03943300619721413, + 0.045550622045993805, + 0.032772041857242584, + 0.047088731080293655, + 0.08223070949316025, + 0.06546109914779663, + 0.05856122821569443, + 0.043971046805381775, + 0.006993715651333332, + 0.006013830192387104, + 0.007623546756803989, + 0.005251673050224781, + 0.007805937901139259, + 0.020946547389030457, + 0.01659102737903595, + 0.005336281843483448, + 0.007209832780063152, + 0.007815718650817871, + 0.00417902460321784, + 0.008400782011449337, + 0.003504369407892227, + 0.004841725341975689, + 0.00499369390308857, + 0.0041519636288285255, + 0.0032977613154798746, + 0.004122427199035883, + 0.0033104978501796722, + 0.0032739960588514805, + 0.0031756984535604715, + 0.0075330669060349464, + 0.002781903836876154, + 0.003359268419444561, + 0.006579816807061434, + 0.009342717938125134, + 0.00700236763805151, + 0.006109747104346752, + 0.010773754678666592, + 0.005519743077456951, + 0.004728490021079779, + 0.0044687665067613125, + 0.009658395312726498, + 0.012905706651508808, + 0.021104464307427406, + 0.012447051703929901, + 0.00966164842247963, + 0.01466822437942028, + 0.005867775995284319, + 0.004148843232542276, + 0.01243378035724163, + 0.022761400789022446, + 0.015899136662483215, + 0.01038806326687336, + 0.010096314363181591, + 0.007998770102858543, + 0.014447525143623352, + 0.005984501913189888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.026922861114144325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024924185127019882 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024924185127019882, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024924185127019882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12509280443191528, + 0.09536650776863098, + 0.06620357930660248, + 0.06349686533212662, + 0.10748326033353806, + 0.2120487242937088, + 0.028951924294233322, + 0.03723210096359253, + 0.2055884152650833, + 0.05853581055998802 + ], + "credit_map": [ + 0.01640833541750908, + 0.014200814068317413, + 0.022018922492861748, + 0.021416975185275078, + 0.009722208604216576, + 0.013880983926355839, + 0.012739594094455242, + 0.014704967848956585, + 0.007669356185942888, + 0.013135664165019989, + 0.01363705936819315, + 0.012118708342313766, + 0.021957257762551308, + 0.009060406126081944, + 0.008515323512256145, + 0.009272734634578228, + 0.009806761518120766, + 0.008448954671621323, + 0.008040571585297585, + 0.0063813310116529465, + 0.008643805049359798, + 0.009599054232239723, + 0.005257081240415573, + 0.010026024654507637, + 0.00835334975272417, + 0.007288351655006409, + 0.006087728776037693, + 0.007609773892909288, + 0.0071461619809269905, + 0.009300049394369125, + 0.0065073287114501, + 0.011204121634364128, + 0.010824357159435749, + 0.0056816269643604755, + 0.009847580455243587, + 0.005551440641283989, + 0.010758679360151291, + 0.03716716542840004, + 0.021135136485099792, + 0.006517274305224419, + 0.027271732687950134, + 0.038040876388549805, + 0.019015299156308174, + 0.028722187504172325, + 0.026419684290885925, + 0.023839181289076805, + 0.0323120541870594, + 0.016427719965577126, + 0.004823153372853994, + 0.0034673470072448254, + 0.0032990602776408195, + 0.0029493533074855804, + 0.0025249074678868055, + 0.006748134270310402, + 0.002636950695887208, + 0.002503017894923687, + 0.006582396104931831, + 0.0058860876597464085, + 0.005755485966801643, + 0.0041440557688474655, + 0.00399903254583478, + 0.0036377557553350925, + 0.0031206493731588125, + 0.004106641747057438, + 0.026153409853577614, + 0.011229402385652065, + 0.0554138645529747, + 0.006356445606797934, + 0.013899377547204494, + 0.08295607566833496, + 0.00449324119836092, + 0.005086610559374094, + 0.010013123974204063, + 0.0073767476715147495, + 0.01225859671831131, + 0.004814230836927891, + 0.006751568056643009, + 0.005665041506290436, + 0.006666258908808231, + 0.004990246146917343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0024924185127019882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947037134319544 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947037134319544, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947037134319544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1042027696967125, + 0.18635182082653046, + 0.08734153211116791, + 0.1374061107635498, + 0.09599317610263824, + 0.07942721247673035, + 0.05157626420259476, + 0.1627236157655716, + 0.042377207428216934, + 0.05260029062628746 + ], + "credit_map": [ + 0.019108042120933533, + 0.01655890978872776, + 0.019924510270357132, + 0.013977156952023506, + 0.006056451238691807, + 0.009574370458722115, + 0.00737758306786418, + 0.011625758372247219, + 0.012716366909444332, + 0.02402724325656891, + 0.02709961123764515, + 0.025456519797444344, + 0.034934479743242264, + 0.019718004390597343, + 0.019098902121186256, + 0.02330067753791809, + 0.009711951948702335, + 0.00956081971526146, + 0.020256727933883667, + 0.01137414202094078, + 0.009318656288087368, + 0.012637875974178314, + 0.006449411157518625, + 0.008031955920159817, + 0.01156666874885559, + 0.018446872010827065, + 0.008285539224743843, + 0.012785620987415314, + 0.031947121024131775, + 0.02331164851784706, + 0.02351558580994606, + 0.007547055371105671, + 0.011569495312869549, + 0.008437597192823887, + 0.018363742157816887, + 0.009442641399800777, + 0.009012134745717049, + 0.01393856480717659, + 0.015695884823799133, + 0.009533115662634373, + 0.007502414286136627, + 0.007182820234447718, + 0.00899695884436369, + 0.02306421846151352, + 0.0057224007323384285, + 0.006268542725592852, + 0.011233620345592499, + 0.009456234984099865, + 0.006082497071474791, + 0.012159688398241997, + 0.007002987898886204, + 0.006201534997671843, + 0.004280895460397005, + 0.005060430150479078, + 0.0039951433427631855, + 0.006793087348341942, + 0.02197563461959362, + 0.03429686278104782, + 0.0192364864051342, + 0.0168838519603014, + 0.020709143951535225, + 0.01996598206460476, + 0.014865030534565449, + 0.014790624380111694, + 0.009462921880185604, + 0.005632766056805849, + 0.010611345991492271, + 0.003205652115866542, + 0.005185642279684544, + 0.0031686550937592983, + 0.002797296503558755, + 0.002312927506864071, + 0.006168486550450325, + 0.007106020115315914, + 0.007530274800956249, + 0.004346787929534912, + 0.008939376100897789, + 0.005001259967684746, + 0.00890156626701355, + 0.004606513772159815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.005947037134319544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253123302012682 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.01253123302012682, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253123302012682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0429815798997879, + 0.11414770781993866, + 0.01763973757624626, + 0.23205077648162842, + 0.1054980456829071, + 0.05132686346769333, + 0.042611975222826004, + 0.057802919298410416, + 0.15452708303928375, + 0.18141329288482666 + ], + "credit_map": [ + 0.006542132701724768, + 0.004683858714997768, + 0.00823174137622118, + 0.006114509887993336, + 0.003603300778195262, + 0.00451422855257988, + 0.004179357085376978, + 0.005112451501190662, + 0.009014028124511242, + 0.01593157649040222, + 0.01693582907319069, + 0.015760865062475204, + 0.023410100489854813, + 0.010769149288535118, + 0.01007333304733038, + 0.01225281972438097, + 0.0020126011222600937, + 0.0020185839384794235, + 0.0016588420839980245, + 0.0031375198159366846, + 0.0023945029824972153, + 0.00235213921405375, + 0.0021013275254517794, + 0.0019642221741378307, + 0.021463187411427498, + 0.02515193447470665, + 0.016725022345781326, + 0.02636628784239292, + 0.047586590051651, + 0.035737406462430954, + 0.033494748175144196, + 0.025525592267513275, + 0.010235118679702282, + 0.006702123209834099, + 0.010077483020722866, + 0.005307219922542572, + 0.010383233428001404, + 0.034168701618909836, + 0.023115698248147964, + 0.005508468486368656, + 0.005159290973097086, + 0.007160107605159283, + 0.006684361957013607, + 0.006508675403892994, + 0.006941787898540497, + 0.007636712864041328, + 0.008056330494582653, + 0.003179599065333605, + 0.006675658281892538, + 0.005105836316943169, + 0.004194199573248625, + 0.006604332942515612, + 0.005302738398313522, + 0.006535772234201431, + 0.0045897020027041435, + 0.0036037378013134003, + 0.007709803991019726, + 0.0096722561866045, + 0.007321591954678297, + 0.006215821020305157, + 0.008432208560407162, + 0.008440743200480938, + 0.004572933539748192, + 0.005437562707811594, + 0.020131338387727737, + 0.008976863697171211, + 0.03929625451564789, + 0.004502080846577883, + 0.011559971608221531, + 0.06121229752898216, + 0.004011661279946566, + 0.004836625419557095, + 0.026091650128364563, + 0.03089141845703125, + 0.028011372312903404, + 0.01440410315990448, + 0.022332075983285904, + 0.016126420348882675, + 0.02385290525853634, + 0.01970333606004715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.01253123302012682 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010413468116894364 + ], + "opposite_logprob_tokens": [ + -6.876041412353516 + ], + "logprob_reward": -0.0010413468116894364, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010413468116894364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049798861145973206, + 0.10726746916770935, + 0.04743344709277153, + 0.06030808389186859, + 0.15440614521503448, + 0.23964346945285797, + 0.03361206501722336, + 0.046011049300432205, + 0.16068178415298462, + 0.1008375734090805 + ], + "credit_map": [ + 0.006632273551076651, + 0.005069531500339508, + 0.008657272905111313, + 0.006524203810840845, + 0.00447001401335001, + 0.005207037087529898, + 0.006622808985412121, + 0.006615720223635435, + 0.009117101319134235, + 0.014941736124455929, + 0.01591983623802662, + 0.012316850014030933, + 0.022934002801775932, + 0.010690136812627316, + 0.010006969794631004, + 0.011340835131704807, + 0.006217875052243471, + 0.005098654422909021, + 0.007835489697754383, + 0.0076110148802399635, + 0.00619784789159894, + 0.006075683981180191, + 0.004583383444696665, + 0.0038134988863021135, + 0.005374955013394356, + 0.007395185995846987, + 0.005132666323333979, + 0.006272571161389351, + 0.012758882716298103, + 0.010034749284386635, + 0.008238287642598152, + 0.005100787151604891, + 0.015241305343806744, + 0.008884871378540993, + 0.013886099681258202, + 0.008133289404213428, + 0.014886570163071156, + 0.05347888916730881, + 0.031606629490852356, + 0.008288494311273098, + 0.03313738480210304, + 0.04465796798467636, + 0.021115845069289207, + 0.03854178264737129, + 0.02422465942800045, + 0.025778256356716156, + 0.03252396732568741, + 0.019663605839014053, + 0.0061609637923538685, + 0.0052142031490802765, + 0.0042596235871315, + 0.004634330980479717, + 0.0034450406674295664, + 0.004068068228662014, + 0.0033582376781851053, + 0.002471594838425517, + 0.00754969846457243, + 0.006474228110164404, + 0.006945926696062088, + 0.00524402828887105, + 0.006555067840963602, + 0.004455483518540859, + 0.0039484756998717785, + 0.004838138818740845, + 0.022181373089551926, + 0.013647589832544327, + 0.03805410489439964, + 0.006080919411033392, + 0.011103599332273006, + 0.06035695597529411, + 0.004394076764583588, + 0.004863174632191658, + 0.017642086371779442, + 0.01461632177233696, + 0.01454070396721363, + 0.009737304411828518, + 0.010779210366308689, + 0.007790802977979183, + 0.012768113054335117, + 0.012963025830686092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.0010413468116894364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, dynamic swooping shot, the camera rapidly descends from a high vantage point and zooms forward, closely following the protagonist, Hodaka. Under a damp, neon-dappled city overpass at night, Hodaka, his short, neatly cut black hair ruffled by the wind, is at the controls of a sleek, heavy-duty black motorcycle. Clad in a fitted black leather motorcycle jacket, his features chiseled, his brow furrowed, his gaze fixed piercingly ahead. Suddenly, in a desperate attempt to control his loss of control, he leans back and slams down the brakes with all his might, his feet practically touching the ground to maintain his balance. This decisive, violent movement causes the motorcycle's rear wheel to lock, sending the vehicle sliding sideways at a precarious angle across the rough asphalt. The intense friction between the tires and the road creates a long, dazzling streak of orange-red sparks, instantly illuminating his focused, tense profile and the motorcycle's cold metal components. The entire picture is filled with high-contrast light and shadow, with the interweaving of dim street lights and dazzling car lights, rendering the sharp dynamics and critical tension of modern martial arts style to the extreme.", + "seed": 17042, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_15/index_2609_seed_17042_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010512557439506054, + "vlm_logprob": -0.010512557439506054, + "vlm_yes_no_margin": 5.325, + "vlm_token_credit": -0.010512557439506054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029985174536705017 + ], + "opposite_logprob_tokens": [ + -3.654985189437866 + ], + "logprob_reward": -0.029985174536705017, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029985174536705017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05953437462449074, + 0.029683949425816536, + 0.06280961632728577, + 0.1656816154718399, + 0.11464295536279678, + 0.0488935261964798, + 0.12021011114120483, + 0.2099161446094513, + 0.02829740010201931, + 0.16033026576042175 + ], + "credit_map": [ + 0.008720689453184605, + 0.009107454679906368, + 0.005813471972942352, + 0.009820947423577309, + 0.004759155213832855, + 0.007145035546272993, + 0.008229144848883152, + 0.005938480142503977, + 0.003354542888700962, + 0.002626103349030018, + 0.0037849093787372112, + 0.004324599634855986, + 0.003849601838737726, + 0.0025573689490556717, + 0.006632757373154163, + 0.0025540697388350964, + 0.007239357568323612, + 0.006214698310941458, + 0.011139506474137306, + 0.00806894525885582, + 0.0076505038887262344, + 0.006460860371589661, + 0.007608701009303331, + 0.008427048102021217, + 0.009664071723818779, + 0.008213276974856853, + 0.034101564437150955, + 0.05294474959373474, + 0.008034220896661282, + 0.00859605148434639, + 0.02957170084118843, + 0.014555993489921093, + 0.013062337413430214, + 0.012341994792222977, + 0.01482225488871336, + 0.01382485032081604, + 0.01978597603738308, + 0.008696794509887695, + 0.018954813480377197, + 0.01315394602715969, + 0.005574223585426807, + 0.0066224257461726665, + 0.004986042622476816, + 0.004833625629544258, + 0.010409437119960785, + 0.005367714911699295, + 0.005003639496862888, + 0.006096421275287867, + 0.012942735105752945, + 0.010106140747666359, + 0.017950210720300674, + 0.016993816941976547, + 0.011788972653448582, + 0.01709205098450184, + 0.017117023468017578, + 0.01621917635202408, + 0.014888939447700977, + 0.016504762694239616, + 0.01762971468269825, + 0.06176937371492386, + 0.00793418288230896, + 0.01147434115409851, + 0.020040081813931465, + 0.05967475846409798, + 0.002768703503534198, + 0.0022095567546784878, + 0.003665877040475607, + 0.004601381719112396, + 0.002834330778568983, + 0.004317352548241615, + 0.004094795789569616, + 0.0038054047618061304, + 0.012927303090691566, + 0.011330166831612587, + 0.014325890690088272, + 0.05567610263824463, + 0.03003724105656147, + 0.009231393225491047, + 0.011585719883441925, + 0.015216456726193428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 9 + ], + "selected_reward": -0.029985174536705017 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017189502250403166 + ], + "opposite_logprob_tokens": [ + -6.376718997955322 + ], + "logprob_reward": -0.0017189502250403166, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017189502250403166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1267557293176651, + 0.03862810879945755, + 0.10745035111904144, + 0.09260739386081696, + 0.0640525370836258, + 0.03023720160126686, + 0.06616590172052383, + 0.08983388543128967, + 0.03675662353634834, + 0.34751227498054504 + ], + "credit_map": [ + 0.023374337702989578, + 0.011419137939810753, + 0.014922799542546272, + 0.0156580600887537, + 0.011868045665323734, + 0.020517298951745033, + 0.011304721236228943, + 0.01769133470952511, + 0.004256139043718576, + 0.003181462176144123, + 0.007056558970361948, + 0.004396050237119198, + 0.004093355964869261, + 0.0048290095292031765, + 0.006137653719633818, + 0.004677880089730024, + 0.012191513553261757, + 0.009295035153627396, + 0.018534071743488312, + 0.022778594866394997, + 0.011380667798221111, + 0.01045737974345684, + 0.011245290748775005, + 0.011567790992558002, + 0.011843058280646801, + 0.00672308960929513, + 0.013188142329454422, + 0.020096344873309135, + 0.006585748866200447, + 0.00735875777900219, + 0.011189451441168785, + 0.015622798353433609, + 0.007886186242103577, + 0.006708980072289705, + 0.009191221557557583, + 0.010097474791109562, + 0.009032659232616425, + 0.005403060466051102, + 0.007437798660248518, + 0.00829515140503645, + 0.003773726290091872, + 0.004332840908318758, + 0.00310921692289412, + 0.0030221424531191587, + 0.004852410405874252, + 0.0036219165194779634, + 0.003275870578363538, + 0.0042490786872804165, + 0.0113596823066473, + 0.006443108897656202, + 0.008937391452491283, + 0.007102959789335728, + 0.005286837927997112, + 0.009406579658389091, + 0.008052434772253036, + 0.00957689993083477, + 0.006751845590770245, + 0.007203145418316126, + 0.00808908324688673, + 0.025887873023748398, + 0.0035593025386333466, + 0.005815151147544384, + 0.008113647811114788, + 0.024413837119936943, + 0.005955968052148819, + 0.0034250791650265455, + 0.006180427502840757, + 0.004787737503647804, + 0.004818528890609741, + 0.0040618726052343845, + 0.004037503153085709, + 0.003489505033940077, + 0.02550000511109829, + 0.025079522281885147, + 0.026006892323493958, + 0.11341788619756699, + 0.08718349039554596, + 0.0220124963670969, + 0.022208455950021744, + 0.026103515177965164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0017189502250403166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195299707353115 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195299707353115, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195299707353115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13044562935829163, + 0.054631877690553665, + 0.1386161595582962, + 0.19045588374137878, + 0.11770254373550415, + 0.06600791215896606, + 0.10226114839315414, + 0.06924861669540405, + 0.05728280916810036, + 0.07334747165441513 + ], + "credit_map": [ + 0.030918480828404427, + 0.014624546281993389, + 0.009377170354127884, + 0.01656262017786503, + 0.012625000439584255, + 0.013767932541668415, + 0.012139326892793179, + 0.020430538803339005, + 0.007776210084557533, + 0.006083531770855188, + 0.005468838382512331, + 0.007232829928398132, + 0.00648000231012702, + 0.004679048899561167, + 0.012741023674607277, + 0.00417038518935442, + 0.007131366990506649, + 0.006767077371478081, + 0.04361400753259659, + 0.043258991092443466, + 0.009647134691476822, + 0.009317057207226753, + 0.00990291777998209, + 0.00897759199142456, + 0.012795106507837772, + 0.012394490651786327, + 0.032654792070388794, + 0.058898553252220154, + 0.010096006095409393, + 0.010794628411531448, + 0.03287102282047272, + 0.019951267167925835, + 0.01141201052814722, + 0.015703750774264336, + 0.011032103560864925, + 0.014838457107543945, + 0.020804567262530327, + 0.007047225255519152, + 0.01721630059182644, + 0.019648125395178795, + 0.006485862657427788, + 0.007536169607192278, + 0.006754213944077492, + 0.006546613294631243, + 0.013715872541069984, + 0.008201300166547298, + 0.008070859126746655, + 0.008697019889950752, + 0.010609835386276245, + 0.008242572657763958, + 0.015160035341978073, + 0.009732091799378395, + 0.007329075131565332, + 0.01824551448225975, + 0.017493629828095436, + 0.015448380261659622, + 0.006101374514400959, + 0.009132720530033112, + 0.0076367720030248165, + 0.017118221148848534, + 0.004093274008482695, + 0.00435362895950675, + 0.004448787774890661, + 0.016363834962248802, + 0.0074449325911700726, + 0.005465509369969368, + 0.009094414301216602, + 0.008138651959598064, + 0.008300301618874073, + 0.007826896384358406, + 0.005986002739518881, + 0.005026095546782017, + 0.007542431354522705, + 0.00586905749514699, + 0.0058782026171684265, + 0.01740434020757675, + 0.01748495362699032, + 0.0073681543581187725, + 0.005239731632173061, + 0.006560590583831072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.014195299707353115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059441933408379555 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.0059441933408379555, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059441933408379555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07596267014741898, + 0.025230690836906433, + 0.15394774079322815, + 0.14191220700740814, + 0.05481366813182831, + 0.029689887538552284, + 0.09165642410516739, + 0.16665686666965485, + 0.07001883536577225, + 0.19011099636554718 + ], + "credit_map": [ + 0.013146583922207355, + 0.011024895124137402, + 0.0078118713572621346, + 0.013379804790019989, + 0.006733932066708803, + 0.010250712744891644, + 0.008104157634079456, + 0.005510708782821894, + 0.002885874593630433, + 0.002565708477050066, + 0.003247979562729597, + 0.00348273990675807, + 0.0028648825827986, + 0.001997855491936207, + 0.005408772733062506, + 0.002776877023279667, + 0.006254571955651045, + 0.005140830762684345, + 0.06076228618621826, + 0.05773312970995903, + 0.006996202748268843, + 0.005031708162277937, + 0.007226820103824139, + 0.004802193026989698, + 0.0206027552485466, + 0.013347218744456768, + 0.018933555111289024, + 0.029328199103474617, + 0.008173591457307339, + 0.013574154116213322, + 0.013861091807484627, + 0.024091646075248718, + 0.006191394757479429, + 0.005419919267296791, + 0.0069556222297251225, + 0.007333520799875259, + 0.010494094341993332, + 0.00523961428552866, + 0.007085724733769894, + 0.006093778647482395, + 0.0027018655091524124, + 0.005827595014125109, + 0.0027189371176064014, + 0.0028705967124551535, + 0.005118832923471928, + 0.004216263070702553, + 0.002620195969939232, + 0.0036156002897769213, + 0.010509636253118515, + 0.006504957098513842, + 0.012325339950621128, + 0.009355872869491577, + 0.009715838357806206, + 0.014274181798100471, + 0.015634994953870773, + 0.013335605151951313, + 0.020703591406345367, + 0.016144324094057083, + 0.02975170500576496, + 0.02707546018064022, + 0.012391510419547558, + 0.020400114357471466, + 0.019757429137825966, + 0.020432719960808754, + 0.006354140117764473, + 0.0063837142661213875, + 0.007724575698375702, + 0.007537690922617912, + 0.00955201592296362, + 0.008833764120936394, + 0.01201875600963831, + 0.01161417830735445, + 0.015228448435664177, + 0.014750520698726177, + 0.013664470054209232, + 0.062447331845760345, + 0.04615907371044159, + 0.011205476708710194, + 0.01183269638568163, + 0.014822982251644135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.0059441933408379555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007191693875938654 + ], + "opposite_logprob_tokens": [ + -7.25071907043457 + ], + "logprob_reward": -0.0007191693875938654, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007191693875938654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11689015477895737, + 0.024467650800943375, + 0.22458191215991974, + 0.09787674993276596, + 0.07733958214521408, + 0.03700009360909462, + 0.06060051918029785, + 0.11801522225141525, + 0.03269138187170029, + 0.21053673326969147 + ], + "credit_map": [ + 0.029877029359340668, + 0.00992651842534542, + 0.009240152314305305, + 0.009601294994354248, + 0.008232816122472286, + 0.016329485923051834, + 0.010607070289552212, + 0.023075785487890244, + 0.0036819458473473787, + 0.0018060751026496291, + 0.003702992806211114, + 0.0039818002842366695, + 0.0030177209991961718, + 0.0034183086827397346, + 0.0033480122219771147, + 0.0015107952058315277, + 0.010702184401452541, + 0.008137345314025879, + 0.0805998146533966, + 0.08342397212982178, + 0.008466202765703201, + 0.010867229662835598, + 0.01154395006597042, + 0.010841215029358864, + 0.0074874842539429665, + 0.006968164350837469, + 0.0187103021889925, + 0.024994751438498497, + 0.00484774773940444, + 0.008842911571264267, + 0.016510577872395515, + 0.00951481144875288, + 0.009453856386244297, + 0.007596542127430439, + 0.010194807313382626, + 0.008575000800192356, + 0.009951004758477211, + 0.006212921347469091, + 0.010941010899841785, + 0.01441444456577301, + 0.0029806809034198523, + 0.003712836652994156, + 0.0032058488577604294, + 0.0043454039841890335, + 0.00809370819479227, + 0.00459953211247921, + 0.005320473574101925, + 0.004741608165204525, + 0.009542636573314667, + 0.006048389244824648, + 0.007563407067209482, + 0.0063164979219436646, + 0.004368518479168415, + 0.008148560300469398, + 0.009001116268336773, + 0.009611395187675953, + 0.007344959303736687, + 0.009377443231642246, + 0.009244188666343689, + 0.037512924522161484, + 0.00415142672136426, + 0.004687710199505091, + 0.009977812878787518, + 0.035718757659196854, + 0.004197895061224699, + 0.0026600651908665895, + 0.004703431390225887, + 0.004415267612785101, + 0.003780702129006386, + 0.004528749734163284, + 0.00503672705963254, + 0.0033685408998280764, + 0.01639201119542122, + 0.015573730692267418, + 0.017525097355246544, + 0.07177581638097763, + 0.04534447565674782, + 0.012055965140461922, + 0.013895959593355656, + 0.017973683774471283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.0007191693875938654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, dynamic swooping shot, the camera rapidly descends from a high vantage point and zooms forward, closely following the protagonist, Hodaka. Under a damp, neon-dappled city overpass at night, Hodaka, his short, neatly cut black hair ruffled by the wind, is at the controls of a sleek, heavy-duty black motorcycle. Clad in a fitted black leather motorcycle jacket, his features chiseled, his brow furrowed, his gaze fixed piercingly ahead. Suddenly, in a desperate attempt to control his loss of control, he leans back and slams down the brakes with all his might, his feet practically touching the ground to maintain his balance. This decisive, violent movement causes the motorcycle's rear wheel to lock, sending the vehicle sliding sideways at a precarious angle across the rough asphalt. The intense friction between the tires and the road creates a long, dazzling streak of orange-red sparks, instantly illuminating his focused, tense profile and the motorcycle's cold metal components. The entire picture is filled with high-contrast light and shadow, with the interweaving of dim street lights and dazzling car lights, rendering the sharp dynamics and critical tension of modern martial arts style to the extreme.", + "seed": 17043, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_15/index_2609_seed_17043_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018132525496184826, + "vlm_logprob": -0.018132525496184826, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.018132525496184826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06871341913938522 + ], + "opposite_logprob_tokens": [ + -2.818713426589966 + ], + "logprob_reward": -0.06871341913938522, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06871341913938522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09543182700872421, + 0.19995419681072235, + 0.06463102996349335, + 0.1274873912334442, + 0.05749134346842766, + 0.11540301144123077, + 0.04129164665937424, + 0.10697243362665176, + 0.16139814257621765, + 0.029939036816358566 + ], + "credit_map": [ + 0.010469355620443821, + 0.011583366431295872, + 0.012982997111976147, + 0.015856236219406128, + 0.012499314732849598, + 0.008948584087193012, + 0.013218839652836323, + 0.009873121976852417, + 0.035879988223314285, + 0.02671131119132042, + 0.01310561504215002, + 0.02792949415743351, + 0.021583108231425285, + 0.02397160232067108, + 0.03008306212723255, + 0.020690010860562325, + 0.006031816825270653, + 0.00547056645154953, + 0.007877003401517868, + 0.010037454776465893, + 0.013619940727949142, + 0.007009346038103104, + 0.008549441583454609, + 0.00603545643389225, + 0.016678232699632645, + 0.018925150856375694, + 0.010631326586008072, + 0.014431535266339779, + 0.010374362580478191, + 0.008076648227870464, + 0.029018361121416092, + 0.019351759925484657, + 0.010039753280580044, + 0.009921376593410969, + 0.004566750023514032, + 0.007681280840188265, + 0.004696074407547712, + 0.00677099172025919, + 0.007932248525321484, + 0.005882862955331802, + 0.015823988243937492, + 0.016884153708815575, + 0.010011130943894386, + 0.010895878076553345, + 0.021919619292020798, + 0.01409163512289524, + 0.01570799946784973, + 0.010068601928651333, + 0.005060017108917236, + 0.005463636014610529, + 0.005164031870663166, + 0.0040590474382042885, + 0.006138361990451813, + 0.005030873231589794, + 0.00631096912547946, + 0.004064710345119238, + 0.011790352873504162, + 0.018310656771063805, + 0.01623690500855446, + 0.009273100644350052, + 0.014121207408607006, + 0.012917444109916687, + 0.011374784633517265, + 0.012947973795235157, + 0.020897341892123222, + 0.024030566215515137, + 0.014681452885270119, + 0.0343523733317852, + 0.019804438576102257, + 0.019041314721107483, + 0.013308107852935791, + 0.015282539650797844, + 0.0033045285381376743, + 0.0020111515186727047, + 0.0041476343758404255, + 0.003615083871409297, + 0.004029454197734594, + 0.005733903031796217, + 0.0037795163225382566, + 0.0033177630975842476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.06871341913938522 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632380351424217 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632380351424217, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632380351424217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13817714154720306, + 0.20623424649238586, + 0.09757910668849945, + 0.14238478243350983, + 0.11113544553518295, + 0.07222029566764832, + 0.031940165907144547, + 0.055750057101249695, + 0.0922139436006546, + 0.05236484110355377 + ], + "credit_map": [ + 0.02055465616285801, + 0.013437440618872643, + 0.02100639045238495, + 0.015741625800728798, + 0.01608719862997532, + 0.02763350121676922, + 0.008619753643870354, + 0.015096567571163177, + 0.03279266133904457, + 0.02866940014064312, + 0.012537899427115917, + 0.028769204393029213, + 0.0181259848177433, + 0.02935449406504631, + 0.03561580553650856, + 0.02036878652870655, + 0.010026815347373486, + 0.009207475930452347, + 0.016419554129242897, + 0.013190694153308868, + 0.012912108562886715, + 0.01060909777879715, + 0.014265472069382668, + 0.010947889648377895, + 0.02348521165549755, + 0.02595081366598606, + 0.012889103963971138, + 0.013233490288257599, + 0.009428475983440876, + 0.009676906280219555, + 0.02768051251769066, + 0.02004026621580124, + 0.012356551364064217, + 0.015192368067800999, + 0.011374286375939846, + 0.013057556003332138, + 0.01164131797850132, + 0.01825666055083275, + 0.013993950560688972, + 0.015262750908732414, + 0.015757432207465172, + 0.012342238798737526, + 0.005969955585896969, + 0.00945065077394247, + 0.006967776454985142, + 0.008192501030862331, + 0.008401529863476753, + 0.005138209089636803, + 0.0036935515236109495, + 0.0027760930825024843, + 0.004310183692723513, + 0.0027611355762928724, + 0.003671871032565832, + 0.0037223459221422672, + 0.005217999219894409, + 0.005786987487226725, + 0.0059048887342214584, + 0.0069441720843315125, + 0.00958127062767744, + 0.005218537524342537, + 0.006591996643692255, + 0.006963689811527729, + 0.007981423288583755, + 0.0065640779212117195, + 0.010137567296624184, + 0.011333681643009186, + 0.007883886806666851, + 0.020989755168557167, + 0.012972572818398476, + 0.011468912474811077, + 0.007475692313164473, + 0.009951867163181305, + 0.0056867473758757114, + 0.006040363106876612, + 0.005279242061078548, + 0.006203314755111933, + 0.009169754572212696, + 0.006356849800795317, + 0.007401931565254927, + 0.006226636003702879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.007632380351424217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638513274490833 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638513274490833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638513274490833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09566807746887207, + 0.26826032996177673, + 0.03945222496986389, + 0.062385592609643936, + 0.09517383575439453, + 0.072472482919693, + 0.13210076093673706, + 0.08461268991231918, + 0.10556478053331375, + 0.04430922120809555 + ], + "credit_map": [ + 0.009206161834299564, + 0.013042319566011429, + 0.013843510299921036, + 0.01363330613821745, + 0.01092508528381586, + 0.013265362940728664, + 0.011128379963338375, + 0.010623952373862267, + 0.050285786390304565, + 0.02741028368473053, + 0.019644329324364662, + 0.04147041589021683, + 0.022825686261057854, + 0.03236984461545944, + 0.04617368057370186, + 0.028080282732844353, + 0.006580732762813568, + 0.004562674555927515, + 0.004321458749473095, + 0.00542279751971364, + 0.004935396369546652, + 0.003486187430098653, + 0.005817200988531113, + 0.004325775429606438, + 0.007084963843226433, + 0.008804530836641788, + 0.007287539541721344, + 0.008160521276295185, + 0.006471080239862204, + 0.004958954639732838, + 0.010178717784583569, + 0.00943928211927414, + 0.014036813750863075, + 0.01626068726181984, + 0.007907452993094921, + 0.016406742855906487, + 0.009185480885207653, + 0.010200785472989082, + 0.012934336438775063, + 0.008241537027060986, + 0.011838019825518131, + 0.011195148341357708, + 0.006457341369241476, + 0.0063819726929068565, + 0.008641289547085762, + 0.011300165206193924, + 0.010755257681012154, + 0.005903289653360844, + 0.01713014580309391, + 0.015057225711643696, + 0.018309572711586952, + 0.011475143022835255, + 0.011187799274921417, + 0.01646759919822216, + 0.016107827425003052, + 0.026365455240011215, + 0.01220727525651455, + 0.012492888607084751, + 0.01564212515950203, + 0.006415525451302528, + 0.008812112733721733, + 0.010275056585669518, + 0.009000514633953571, + 0.00976718869060278, + 0.012627849355340004, + 0.01429698709398508, + 0.009133766405284405, + 0.02523978427052498, + 0.01141861267387867, + 0.011792607605457306, + 0.009435503743588924, + 0.01161966659128666, + 0.005454496946185827, + 0.00474190479144454, + 0.004172417800873518, + 0.004954481963068247, + 0.0068658143281936646, + 0.006743865553289652, + 0.006276159547269344, + 0.005100079346448183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.008638513274490833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636015277355909 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004636015277355909, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636015277355909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049427349120378494, + 0.22436779737472534, + 0.018477890640497208, + 0.06438256055116653, + 0.04776297137141228, + 0.10667763650417328, + 0.0472666434943676, + 0.0670776292681694, + 0.24741224944591522, + 0.12714719772338867 + ], + "credit_map": [ + 0.005792492069303989, + 0.006530180107802153, + 0.006415367592126131, + 0.007056946866214275, + 0.007345911581069231, + 0.00675647146999836, + 0.004192873369902372, + 0.005337111186236143, + 0.04389796406030655, + 0.021234961226582527, + 0.01655585691332817, + 0.028536861762404442, + 0.024957984685897827, + 0.029908016324043274, + 0.03345595300197601, + 0.025820212438702583, + 0.0022924458608031273, + 0.0021174519788473845, + 0.0022471649572253227, + 0.002300581894814968, + 0.003863817546516657, + 0.0015127742663025856, + 0.0026683572214096785, + 0.0014752978459000587, + 0.008210449479520321, + 0.013141336850821972, + 0.007144052069634199, + 0.005666166078299284, + 0.0058159055188298225, + 0.0039491187781095505, + 0.01218861248344183, + 0.008266923949122429, + 0.005019681993871927, + 0.0048394715413451195, + 0.007784122135490179, + 0.005543156526982784, + 0.00519241439178586, + 0.005734168458729982, + 0.007148417644202709, + 0.006501543335616589, + 0.017815005034208298, + 0.01379443984478712, + 0.009069977328181267, + 0.01493341475725174, + 0.014070279896259308, + 0.015911268070340157, + 0.013399042189121246, + 0.007684213574975729, + 0.0064552174881100655, + 0.004537611734122038, + 0.006476827431470156, + 0.004612546414136887, + 0.004669182002544403, + 0.005560850724577904, + 0.006502348463982344, + 0.00845206156373024, + 0.008033489808440208, + 0.007180883549153805, + 0.009202905930578709, + 0.006538442801684141, + 0.007870325818657875, + 0.008551695384085178, + 0.01061440072953701, + 0.009085492230951786, + 0.02766207419335842, + 0.03324698656797409, + 0.02232225239276886, + 0.050001360476017, + 0.0365922786295414, + 0.02881225198507309, + 0.022392546758055687, + 0.026382528245449066, + 0.013907236978411674, + 0.014761107973754406, + 0.0134734520688653, + 0.01670912466943264, + 0.014988509006798267, + 0.018404047936201096, + 0.020744632929563522, + 0.014159104786813259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.004636015277355909 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010422994382679462 + ], + "opposite_logprob_tokens": [ + -6.876042366027832 + ], + "logprob_reward": -0.0010422994382679462, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010422994382679462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11526339501142502, + 0.21371816098690033, + 0.06793815642595291, + 0.10789830982685089, + 0.09535172581672668, + 0.08369235694408417, + 0.07564686238765717, + 0.06740299612283707, + 0.12318882346153259, + 0.049899160861968994 + ], + "credit_map": [ + 0.014535621739923954, + 0.010879572480916977, + 0.018590401858091354, + 0.015219966880977154, + 0.012017135508358479, + 0.021581901237368584, + 0.00926254317164421, + 0.013176258653402328, + 0.03469334915280342, + 0.031725287437438965, + 0.012966091744601727, + 0.03145061805844307, + 0.021543744951486588, + 0.025707663968205452, + 0.035363804548978806, + 0.020267607644200325, + 0.00950413104146719, + 0.00694111455231905, + 0.0077547235414385796, + 0.008361040614545345, + 0.011342158541083336, + 0.007481927517801523, + 0.009119024500250816, + 0.007434044498950243, + 0.014850575476884842, + 0.02337043732404709, + 0.011990402825176716, + 0.009357634000480175, + 0.00839154701679945, + 0.008049845695495605, + 0.018124539405107498, + 0.013763329945504665, + 0.012112879194319248, + 0.016179444268345833, + 0.010121267288923264, + 0.016115330159664154, + 0.00798158347606659, + 0.011650252155959606, + 0.011493056081235409, + 0.00969791878014803, + 0.020018834620714188, + 0.012578423134982586, + 0.0063843270763754845, + 0.008561519905924797, + 0.007018519099801779, + 0.0132669722661376, + 0.010459342040121555, + 0.0054044234566390514, + 0.015838561579585075, + 0.009939231909811497, + 0.007013975642621517, + 0.006084580440074205, + 0.00788623746484518, + 0.008560863323509693, + 0.010188549757003784, + 0.010134861804544926, + 0.007822372950613499, + 0.00764632411301136, + 0.009611242450773716, + 0.0059082601219415665, + 0.007652970962226391, + 0.00954862218350172, + 0.011459030210971832, + 0.007754174061119556, + 0.014814990572631359, + 0.01481909304857254, + 0.011438178829848766, + 0.02363332360982895, + 0.01837141253054142, + 0.016299862414598465, + 0.010682235471904278, + 0.013129734434187412, + 0.006178588606417179, + 0.00438636913895607, + 0.006120544392615557, + 0.007550135720521212, + 0.005898453760892153, + 0.006973046809434891, + 0.006649905815720558, + 0.006142116617411375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0010422994382679462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, dynamic swooping shot, the camera rapidly descends from a high vantage point and zooms forward, closely following the protagonist, Hodaka. Under a damp, neon-dappled city overpass at night, Hodaka, his short, neatly cut black hair ruffled by the wind, is at the controls of a sleek, heavy-duty black motorcycle. Clad in a fitted black leather motorcycle jacket, his features chiseled, his brow furrowed, his gaze fixed piercingly ahead. Suddenly, in a desperate attempt to control his loss of control, he leans back and slams down the brakes with all his might, his feet practically touching the ground to maintain his balance. This decisive, violent movement causes the motorcycle's rear wheel to lock, sending the vehicle sliding sideways at a precarious angle across the rough asphalt. The intense friction between the tires and the road creates a long, dazzling streak of orange-red sparks, instantly illuminating his focused, tense profile and the motorcycle's cold metal components. The entire picture is filled with high-contrast light and shadow, with the interweaving of dim street lights and dazzling car lights, rendering the sharp dynamics and critical tension of modern martial arts style to the extreme.", + "seed": 17044, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_15/index_2609_seed_17044_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016057304467540234, + "vlm_logprob": -0.016057304467540234, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.016057304467540234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05958480015397072 + ], + "opposite_logprob_tokens": [ + -2.934584856033325 + ], + "logprob_reward": -0.05958480015397072, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05958480015397072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11675619333982468, + 0.05206120014190674, + 0.1898675560951233, + 0.1547447144985199, + 0.23126764595508575, + 0.034984882920980453, + 0.04738942161202431, + 0.0644170418381691, + 0.07947392761707306, + 0.029037389904260635 + ], + "credit_map": [ + 0.016405876725912094, + 0.013742643408477306, + 0.017321916297078133, + 0.012917432934045792, + 0.01126687228679657, + 0.017127282917499542, + 0.014307258650660515, + 0.013666915707290173, + 0.008776826784014702, + 0.005168385803699493, + 0.005886666011065245, + 0.005865579005330801, + 0.005146197509020567, + 0.007248330395668745, + 0.007784769870340824, + 0.00618444848805666, + 0.019936857745051384, + 0.02611745521426201, + 0.03299421817064285, + 0.01359023991972208, + 0.026171689853072166, + 0.024060338735580444, + 0.013182347640395164, + 0.03381442651152611, + 0.009995677508413792, + 0.04894081503152847, + 0.010959233157336712, + 0.020746858790516853, + 0.015857545658946037, + 0.016924375668168068, + 0.013645747676491737, + 0.01767447404563427, + 0.026170959696173668, + 0.03669052943587303, + 0.02742585353553295, + 0.028209365904331207, + 0.03756609931588173, + 0.024050764739513397, + 0.025333458557724953, + 0.025820618495345116, + 0.002888713264837861, + 0.004167549312114716, + 0.005190819501876831, + 0.004552262835204601, + 0.005072346422821283, + 0.002721763215959072, + 0.005349438171833754, + 0.005041992757469416, + 0.005592711735516787, + 0.004161202814429998, + 0.010362711735069752, + 0.0031959819607436657, + 0.0042685819789767265, + 0.006893925368785858, + 0.004638738930225372, + 0.008275570347905159, + 0.00714331679046154, + 0.007519349921494722, + 0.00697854720056057, + 0.008974759839475155, + 0.0110083669424057, + 0.0062012989073991776, + 0.011642752215266228, + 0.004948653746396303, + 0.0065665096044540405, + 0.00796655286103487, + 0.0260508731007576, + 0.014351804740726948, + 0.004840220790356398, + 0.011482834815979004, + 0.003876107046380639, + 0.004339027218520641, + 0.002810643520206213, + 0.006065662484616041, + 0.0036508808843791485, + 0.0038491208106279373, + 0.0032572122290730476, + 0.003455480793491006, + 0.004059791099280119, + 0.0018885976169258356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.05958480015397072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036216876469552517 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036216876469552517, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036216876469552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13673771917819977, + 0.09307311475276947, + 0.08112838119268417, + 0.24667099118232727, + 0.1463063508272171, + 0.044361140578985214, + 0.1119958832859993, + 0.031025955453515053, + 0.06249489262700081, + 0.0462055541574955 + ], + "credit_map": [ + 0.020913975313305855, + 0.017057353630661964, + 0.019290149211883545, + 0.01831451803445816, + 0.01472381316125393, + 0.014288832433521748, + 0.01882016845047474, + 0.013328916393220425, + 0.01207036804407835, + 0.011251542717218399, + 0.010637111030519009, + 0.00956119317561388, + 0.012265794910490513, + 0.019211068749427795, + 0.008404628373682499, + 0.009671404026448727, + 0.010846269316971302, + 0.013497351668775082, + 0.010102012194693089, + 0.0075737666338682175, + 0.011195111088454723, + 0.011074747890233994, + 0.007415443193167448, + 0.00942367222160101, + 0.02497001551091671, + 0.04045205935835838, + 0.018215486779808998, + 0.027143636718392372, + 0.03690961003303528, + 0.039596542716026306, + 0.03476913273334503, + 0.02461451292037964, + 0.01986539363861084, + 0.019951552152633667, + 0.019065164029598236, + 0.01918039657175541, + 0.021195460110902786, + 0.01566406525671482, + 0.017283832654356956, + 0.014100500382483006, + 0.004629035014659166, + 0.004345148801803589, + 0.008375213481485844, + 0.004761307965964079, + 0.004969953093677759, + 0.003844421822577715, + 0.0060681989416480064, + 0.0073678577318787575, + 0.013152418658137321, + 0.015456628985702991, + 0.01641712710261345, + 0.017804818227887154, + 0.015424701385200024, + 0.013277608901262283, + 0.010917861945927143, + 0.00954472366720438, + 0.0033580849412828684, + 0.003010161453858018, + 0.002898965962231159, + 0.003824791172519326, + 0.005508236587047577, + 0.0037564176600426435, + 0.004500010050833225, + 0.004169285763055086, + 0.005329141393303871, + 0.01327485404908657, + 0.01647607982158661, + 0.00915976520627737, + 0.005066583398729563, + 0.005933215841650963, + 0.0034790164791047573, + 0.0037762396968901157, + 0.0047965580597519875, + 0.005037541501224041, + 0.00458849873393774, + 0.005925754085183144, + 0.0052254414185881615, + 0.010219528339803219, + 0.0063019911758601665, + 0.0041102394461631775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.0036216876469552517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639813400804996 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639813400804996, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639813400804996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11096305400133133, + 0.03782346844673157, + 0.09470352530479431, + 0.048261649906635284, + 0.43099159002304077, + 0.09413933008909225, + 0.05412618815898895, + 0.015663886442780495, + 0.0867784395813942, + 0.02654888667166233 + ], + "credit_map": [ + 0.015813063830137253, + 0.01616324670612812, + 0.017199592664837837, + 0.014960278756916523, + 0.011276889592409134, + 0.010859503410756588, + 0.01280196662992239, + 0.011888501234352589, + 0.005651264451444149, + 0.004022552631795406, + 0.005931055173277855, + 0.003437966341152787, + 0.003339777933433652, + 0.004478215705603361, + 0.004072040319442749, + 0.006890594027936459, + 0.010354826226830482, + 0.013790282420814037, + 0.01461495365947485, + 0.007607676088809967, + 0.017167942598462105, + 0.008061077445745468, + 0.012591533362865448, + 0.010515226051211357, + 0.003320768941193819, + 0.009113688953220844, + 0.0030062871519476175, + 0.0068116867914795876, + 0.007630982436239719, + 0.007172813639044762, + 0.005850162822753191, + 0.005355257540941238, + 0.0519896000623703, + 0.070261150598526, + 0.05057113617658615, + 0.0542602576315403, + 0.06555545330047607, + 0.04576842114329338, + 0.04750480502843857, + 0.045080725103616714, + 0.009849813766777515, + 0.011246571317315102, + 0.012049288488924503, + 0.014110853895545006, + 0.011874713934957981, + 0.01069719810038805, + 0.010813643224537373, + 0.013497233390808105, + 0.005938698071986437, + 0.006793180480599403, + 0.008161798119544983, + 0.00620687659829855, + 0.005720330402255058, + 0.0061213611625134945, + 0.009512526914477348, + 0.005671414081007242, + 0.001476766774430871, + 0.0017598969861865044, + 0.0013768295757472515, + 0.002091183327138424, + 0.0027548596262931824, + 0.002129909349605441, + 0.0022468669340014458, + 0.0018275730544701219, + 0.006034331861883402, + 0.012659831903874874, + 0.028755610808730125, + 0.014413349330425262, + 0.004278465639799833, + 0.013232789002358913, + 0.0028410900849848986, + 0.004562963265925646, + 0.0023067165166139603, + 0.005319776013493538, + 0.0024586692452430725, + 0.005107899196445942, + 0.002274363534525037, + 0.003483329201117158, + 0.004104192368686199, + 0.0014939375687390566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.008639813400804996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625400088727474 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625400088727474, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625400088727474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10839546471834183, + 0.10886380821466446, + 0.08829715102910995, + 0.13604402542114258, + 0.19549718499183655, + 0.05297242850065231, + 0.0752999484539032, + 0.02841498889029026, + 0.1768130362033844, + 0.029402023181319237 + ], + "credit_map": [ + 0.013713539578020573, + 0.009421908296644688, + 0.013665108010172844, + 0.010908885858952999, + 0.01712045632302761, + 0.020491698756814003, + 0.014415335841476917, + 0.008658532053232193, + 0.02404564619064331, + 0.008429705165326595, + 0.01121348887681961, + 0.011849991977214813, + 0.010167166590690613, + 0.015396892093122005, + 0.015538861975073814, + 0.012222051620483398, + 0.011477237567305565, + 0.014898772351443768, + 0.00834549218416214, + 0.007778619881719351, + 0.012546238489449024, + 0.01212329138070345, + 0.010570110753178596, + 0.010557388886809349, + 0.010890042409300804, + 0.030584001913666725, + 0.010980764403939247, + 0.01704445853829384, + 0.01591036282479763, + 0.0176894161850214, + 0.015167138539254665, + 0.017777839675545692, + 0.021045738831162453, + 0.025355560705065727, + 0.0239564199000597, + 0.023797787725925446, + 0.03499915450811386, + 0.02041797526180744, + 0.023108195513486862, + 0.02281634695827961, + 0.004091707523912191, + 0.004574829712510109, + 0.005272038280963898, + 0.008347483351826668, + 0.007031163666397333, + 0.004780990071594715, + 0.009896038100123405, + 0.008978176862001419, + 0.008359256200492382, + 0.005448098294436932, + 0.015972957015037537, + 0.011072301305830479, + 0.008762196637690067, + 0.009096779860556126, + 0.0076514119282364845, + 0.008936943486332893, + 0.0036894201766699553, + 0.0033994417171925306, + 0.002995549701154232, + 0.0034734068904072046, + 0.004837788641452789, + 0.0033053376246243715, + 0.004070242866873741, + 0.0026438029017299414, + 0.014314684085547924, + 0.022972259670495987, + 0.038744792342185974, + 0.02558104135096073, + 0.023106876760721207, + 0.020329749211668968, + 0.014706059359014034, + 0.01705758459866047, + 0.0032286534551531076, + 0.005762044340372086, + 0.0035091545432806015, + 0.004320531617850065, + 0.0030836889054626226, + 0.0041227382607758045, + 0.003148293122649193, + 0.002226919401437044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.007625400088727474 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist, Hodaka, visible riding a sleek, heavy-duty black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, neon-dappled city overpass at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly descend from a high vantage point and zoom forward while following the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle slide sideways on rough asphalt while leaving a long streak of orange-red sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast light and shadow with interweaving dim street lights and dazzling car lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008148210472427309 + ], + "opposite_logprob_tokens": [ + -7.125814914703369 + ], + "logprob_reward": -0.0008148210472427309, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008148210472427309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.131049245595932, + 0.07022452354431152, + 0.15572649240493774, + 0.09652520716190338, + 0.061642903834581375, + 0.12087948620319366, + 0.08823686838150024, + 0.04686475172638893, + 0.16698826849460602, + 0.06186224892735481 + ], + "credit_map": [ + 0.019282463937997818, + 0.017100416123867035, + 0.020685339346528053, + 0.016234319657087326, + 0.012099282816052437, + 0.014414102770388126, + 0.016731059178709984, + 0.014502253383398056, + 0.009951314888894558, + 0.009127178229391575, + 0.007296718657016754, + 0.010433322750031948, + 0.0067745959386229515, + 0.009587100706994534, + 0.009328450076282024, + 0.007725846022367477, + 0.01556265540421009, + 0.022696897387504578, + 0.016191797330975533, + 0.011801548302173615, + 0.02784992754459381, + 0.02555415779352188, + 0.011626831255853176, + 0.024442682042717934, + 0.012460164725780487, + 0.013440834358334541, + 0.007393003907054663, + 0.010538801550865173, + 0.011181365698575974, + 0.02152634784579277, + 0.00862162746489048, + 0.011363065801560879, + 0.004597538150846958, + 0.007487810682505369, + 0.010554029606282711, + 0.009034750051796436, + 0.00814337283372879, + 0.007213280536234379, + 0.008390181697905064, + 0.006221942137926817, + 0.011849112808704376, + 0.019976148381829262, + 0.01926378719508648, + 0.014560291543602943, + 0.014913209713995457, + 0.012749920599162579, + 0.012219235301017761, + 0.015347780659794807, + 0.009769932366907597, + 0.010050653479993343, + 0.010687438771128654, + 0.014308149926364422, + 0.012983501888811588, + 0.01048718485981226, + 0.010591142810881138, + 0.009358861483633518, + 0.004944647196680307, + 0.005013166461139917, + 0.0037940095644444227, + 0.005438601598143578, + 0.008271532133221626, + 0.006163168698549271, + 0.007707655429840088, + 0.005531969480216503, + 0.012887297198176384, + 0.019922200590372086, + 0.051443926990032196, + 0.02535785175859928, + 0.009881566278636456, + 0.031397365033626556, + 0.006494589615613222, + 0.009603467769920826, + 0.006260250695049763, + 0.009544385597109795, + 0.00849571917206049, + 0.007294720970094204, + 0.007635089568793774, + 0.010085778310894966, + 0.005530688911676407, + 0.007015617098659277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.0008148210472427309 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_tensor([3046, 3046])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_tensor([3046, 3046])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..911261de21c65caaa8c85cc7a9d57288adc85dfc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_tensor([3046, 3046])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, continuous, circling tracking shot captures a medium shot at eye level, steadily rotating around a man engaged in an intense balance exercise. He is in an abstract world composed entirely of clean, flat polygons, atop a patchwork of vast, translucent geometric shapes suspended in an endless, dark void. The man, with an athletic build, wears a sleek, form-fitting, emerald-green high-tech bodysuit that perfectly outlines every inch of his taut muscles. In the center of his chest, a geometric, abstract symbol emits a soft, rhythmic cyan light, casting a cool halo around his focused jaw. He carefully stands on one leg on a narrow, hexagonal platform with a faintly luminescent edge, his other leg extended behind him for balance, his arms raised, his fingers making minute, precise adjustments to his body's subtle movements. Beneath his sweat-stained short hair, his eyes gazed unwaveringly at a single point in the void ahead. His tightly pursed lips and taut jawline betrayed his extraordinary concentration and willpower. The scene was dominated by a stark palette of black, green, and cyan. The cool white light emanating from the polygonal edges and the cyan glow of the emblem on his chest provided the sole illumination, creating a tranquil atmosphere that felt futuristic, intensely focused, and filled with inner tension.", + "seed": 24442, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_3046_seed_24442_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02972012381069362, + "vlm_logprob": -0.02972012381069362, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02972012381069362, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08798239380121231 + ], + "opposite_logprob_tokens": [ + -2.712982416152954 + ], + "logprob_reward": -0.08798239380121231, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08798239380121231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26147156953811646, + 0.0509292371571064, + 0.08571173995733261, + 0.08719935268163681, + 0.05346830189228058, + 0.06634746491909027, + 0.050656333565711975, + 0.16744954884052277, + 0.06995531916618347, + 0.10681122541427612 + ], + "credit_map": [ + 0.02368989773094654, + 0.023431144654750824, + 0.018687494099140167, + 0.03518281877040863, + 0.032216206192970276, + 0.0364648774266243, + 0.04238050803542137, + 0.04941857233643532, + 0.005670784041285515, + 0.008260724134743214, + 0.00683389138430357, + 0.004556531552225351, + 0.007753463927656412, + 0.004686340689659119, + 0.006801785435527563, + 0.006365709938108921, + 0.013033307157456875, + 0.018781889230012894, + 0.012966734357178211, + 0.012964318506419659, + 0.009125977754592896, + 0.003969541750848293, + 0.00834355317056179, + 0.0065264091826975346, + 0.013059958815574646, + 0.01323716901242733, + 0.012952892109751701, + 0.011864722706377506, + 0.0077611906453967094, + 0.009482856839895248, + 0.007890162989497185, + 0.010950391180813313, + 0.005973224528133869, + 0.010026896372437477, + 0.0077462405897676945, + 0.007905873470008373, + 0.004830266814678907, + 0.005952079780399799, + 0.006016873754560947, + 0.005016836803406477, + 0.0069938646629452705, + 0.008138557896018028, + 0.010924667119979858, + 0.011385413818061352, + 0.0066135115921497345, + 0.007399012800306082, + 0.006927149835973978, + 0.00796528346836567, + 0.007726103998720646, + 0.005536020267754793, + 0.0075597092509269714, + 0.00444093719124794, + 0.005384367890655994, + 0.005928804632276297, + 0.0080899428576231, + 0.005990442354232073, + 0.01975281909108162, + 0.01808181032538414, + 0.030021702870726585, + 0.028502361848950386, + 0.020986201241612434, + 0.01379663310945034, + 0.017440371215343475, + 0.018867630511522293, + 0.006995059084147215, + 0.007761192508041859, + 0.010640738531947136, + 0.012685294263064861, + 0.009201787412166595, + 0.006678445264697075, + 0.009799106977880001, + 0.006193685811012983, + 0.006954212207347155, + 0.009357658214867115, + 0.015621992759406567, + 0.009893745183944702, + 0.021292826160788536, + 0.011709067970514297, + 0.02047272026538849, + 0.011508985422551632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.08798239380121231 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006773842498660088 + ], + "opposite_logprob_tokens": [ + -5.006773948669434 + ], + "logprob_reward": -0.006773842498660088, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006773842498660088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20528946816921234, + 0.07940000295639038, + 0.11996447294950485, + 0.09287840127944946, + 0.048378732055425644, + 0.046237435191869736, + 0.1224086657166481, + 0.14286968111991882, + 0.07037036120891571, + 0.07220283150672913 + ], + "credit_map": [ + 0.027359988540410995, + 0.015114777721464634, + 0.018003657460212708, + 0.03420316055417061, + 0.025012165307998657, + 0.0286756232380867, + 0.032581791281700134, + 0.024338295683264732, + 0.012749351561069489, + 0.01257762499153614, + 0.008812734857201576, + 0.007731129880994558, + 0.009654447436332703, + 0.006410762667655945, + 0.012117777019739151, + 0.009346169419586658, + 0.016992727294564247, + 0.032137930393218994, + 0.010201769880950451, + 0.01997740939259529, + 0.011374070309102535, + 0.00721985287964344, + 0.010821596719324589, + 0.01123911701142788, + 0.012096774764358997, + 0.012389588169753551, + 0.01515668909996748, + 0.01287092361599207, + 0.007785001304000616, + 0.010565835051238537, + 0.009377541020512581, + 0.012636051513254642, + 0.005059558432549238, + 0.008578912355005741, + 0.007420200854539871, + 0.0075980303809046745, + 0.006181218661367893, + 0.004025336354970932, + 0.00662409421056509, + 0.0028913815040141344, + 0.007605529390275478, + 0.007258888799697161, + 0.008077568374574184, + 0.005941222421824932, + 0.006404415238648653, + 0.0028932220302522182, + 0.005259292665868998, + 0.002797299763187766, + 0.009388020262122154, + 0.009382679127156734, + 0.020159544423222542, + 0.011217964813113213, + 0.013209831900894642, + 0.016309797763824463, + 0.02020801603794098, + 0.022532818838953972, + 0.02321796864271164, + 0.01765935868024826, + 0.024988794699311256, + 0.023898739367723465, + 0.016166482120752335, + 0.010081824846565723, + 0.014416655525565147, + 0.012439863756299019, + 0.005667957477271557, + 0.00814882107079029, + 0.006392511073499918, + 0.010357718914747238, + 0.009589267894625664, + 0.009765027090907097, + 0.014537646435201168, + 0.005911409389227629, + 0.008386786095798016, + 0.009741106070578098, + 0.012219009920954704, + 0.006079473532736301, + 0.011153412982821465, + 0.008212101645767689, + 0.010461020283401012, + 0.005949912592768669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.006773842498660088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547949329018593 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547949329018593, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547949329018593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1488407999277115, + 0.060242779552936554, + 0.25904950499534607, + 0.07002425193786621, + 0.044109031558036804, + 0.051461361348629, + 0.051451992243528366, + 0.2047552764415741, + 0.031681548804044724, + 0.07838346064090729 + ], + "credit_map": [ + 0.012347547337412834, + 0.011715464293956757, + 0.008903060108423233, + 0.018668703734874725, + 0.026108531281352043, + 0.014109502546489239, + 0.034834496676921844, + 0.022153504192829132, + 0.008008192293345928, + 0.007525538094341755, + 0.005692868493497372, + 0.005171541124582291, + 0.008789380080997944, + 0.0056686317548155785, + 0.011935637332499027, + 0.0074509866535663605, + 0.03742842748761177, + 0.059880536049604416, + 0.02888449653983116, + 0.04496581479907036, + 0.02701069787144661, + 0.013432065024971962, + 0.027706408873200417, + 0.019741078838706017, + 0.008825226686894894, + 0.010317560285329819, + 0.009877152740955353, + 0.008876271545886993, + 0.00841725617647171, + 0.005127186421304941, + 0.011624936014413834, + 0.0069586606696248055, + 0.003796694800257683, + 0.006133885122835636, + 0.007274611387401819, + 0.011468644253909588, + 0.00391008285805583, + 0.002946852007880807, + 0.0053255693055689335, + 0.003252688329666853, + 0.006367019843310118, + 0.005686277057975531, + 0.009318152442574501, + 0.007210979238152504, + 0.006875219754874706, + 0.0037887885700911283, + 0.0084813442081213, + 0.0037335797678679228, + 0.0052207233384251595, + 0.0044142152182757854, + 0.00809026975184679, + 0.0068329996429383755, + 0.005463607609272003, + 0.006637653335928917, + 0.008214612491428852, + 0.0065779127180576324, + 0.027376722544431686, + 0.034963492304086685, + 0.028385141864418983, + 0.02901138737797737, + 0.02018221840262413, + 0.025848388671875, + 0.02409539930522442, + 0.014892533421516418, + 0.002522979862987995, + 0.004576405510306358, + 0.0031644373666495085, + 0.004392171278595924, + 0.0051789660938084126, + 0.0038437077309936285, + 0.005174513440579176, + 0.002828366355970502, + 0.005882955621927977, + 0.009568602778017521, + 0.016803503036499023, + 0.00859820656478405, + 0.007695819716900587, + 0.007879180833697319, + 0.01361856795847416, + 0.008336629718542099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.012547949329018593 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810315206646919 + ], + "opposite_logprob_tokens": [ + -3.2881031036376953 + ], + "logprob_reward": -0.03810315206646919, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810315206646919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07678595185279846, + 0.03958479315042496, + 0.06119190901517868, + 0.2076994776725769, + 0.07878011465072632, + 0.0835447683930397, + 0.09014081954956055, + 0.27455025911331177, + 0.045369137078523636, + 0.04235276207327843 + ], + "credit_map": [ + 0.00810972973704338, + 0.0084104398265481, + 0.007092406041920185, + 0.010532064363360405, + 0.008332216180860996, + 0.009948335587978363, + 0.01248534582555294, + 0.011875423602759838, + 0.0070181419141590595, + 0.006040758918970823, + 0.005044123623520136, + 0.004424385726451874, + 0.0030323632527142763, + 0.003282022662460804, + 0.005239507183432579, + 0.005503492895513773, + 0.011261843144893646, + 0.007683372125029564, + 0.012038514018058777, + 0.006850008852779865, + 0.007222842425107956, + 0.0030350976157933474, + 0.006327375769615173, + 0.006772855296730995, + 0.02539687789976597, + 0.023085901513695717, + 0.027766644954681396, + 0.03306473419070244, + 0.025608360767364502, + 0.024057762697339058, + 0.025164633989334106, + 0.023554563522338867, + 0.006706566084176302, + 0.011449412442743778, + 0.013903998769819736, + 0.01984436996281147, + 0.005645338911563158, + 0.005435931961983442, + 0.00925479270517826, + 0.006539711728692055, + 0.01665230095386505, + 0.012753380462527275, + 0.015818845480680466, + 0.014066076837480068, + 0.004962621722370386, + 0.005454965867102146, + 0.007053298410028219, + 0.006783280521631241, + 0.016546588391065598, + 0.010724913328886032, + 0.013982050120830536, + 0.010238878428936005, + 0.008465620689094067, + 0.008369009010493755, + 0.009586620144546032, + 0.012227139435708523, + 0.02974710427224636, + 0.051980674266815186, + 0.04082544520497322, + 0.04973158612847328, + 0.03272994980216026, + 0.027893098071217537, + 0.020101219415664673, + 0.021541200578212738, + 0.0044607906602323055, + 0.0048929122276604176, + 0.007203483954071999, + 0.0107728261500597, + 0.0040133302100002766, + 0.003969014156609774, + 0.00529093760997057, + 0.004765849094837904, + 0.0037805454339832067, + 0.006919047329574823, + 0.008843251504004002, + 0.004655234981328249, + 0.0038449091371148825, + 0.004575476981699467, + 0.00575271574780345, + 0.00398158747702837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.03810315206646919 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031932813581079245 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031932813581079245, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031932813581079245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15162305533885956, + 0.043515682220458984, + 0.10606518387794495, + 0.08949452638626099, + 0.0477483794093132, + 0.17451947927474976, + 0.034221649169921875, + 0.10606341063976288, + 0.05327490717172623, + 0.1934737116098404 + ], + "credit_map": [ + 0.020218972116708755, + 0.015974586829543114, + 0.014815113507211208, + 0.016357311978936195, + 0.01964757591485977, + 0.022815754637122154, + 0.01877146027982235, + 0.023022286593914032, + 0.007663513533771038, + 0.006693757139146328, + 0.0056567867286503315, + 0.00428002281114459, + 0.004917019046843052, + 0.003848816966637969, + 0.004723154939711094, + 0.005732612684369087, + 0.021480554714798927, + 0.020665109157562256, + 0.01942630670964718, + 0.014522523619234562, + 0.00816341396421194, + 0.005278442986309528, + 0.008450067602097988, + 0.008078771643340588, + 0.014792739413678646, + 0.01147592905908823, + 0.0124352490529418, + 0.013642785139381886, + 0.007863929495215416, + 0.011125982739031315, + 0.008493381552398205, + 0.009664532728493214, + 0.0046915533021092415, + 0.009195744059979916, + 0.007473133970052004, + 0.009613723494112492, + 0.005101855378597975, + 0.003956938162446022, + 0.004592821467667818, + 0.0031226102728396654, + 0.03407147526741028, + 0.03253354877233505, + 0.03279139846563339, + 0.024277864024043083, + 0.011795107275247574, + 0.01353747770190239, + 0.011687936261296272, + 0.01382468268275261, + 0.006694023031741381, + 0.0036389692686498165, + 0.005703760776668787, + 0.002866995520889759, + 0.003105695592239499, + 0.0052477773278951645, + 0.003431734163314104, + 0.003532692324370146, + 0.015584465116262436, + 0.00874850433319807, + 0.018596380949020386, + 0.023611560463905334, + 0.010773153975605965, + 0.009628313593566418, + 0.012105477973818779, + 0.007015552371740341, + 0.004078748635947704, + 0.006815867964178324, + 0.007310561370104551, + 0.0075742146000266075, + 0.0077069224789738655, + 0.005916869733482599, + 0.008761836215853691, + 0.005109887104481459, + 0.023511676117777824, + 0.02205665409564972, + 0.03218884766101837, + 0.018449150025844574, + 0.027241287752985954, + 0.022953452542424202, + 0.026715492829680443, + 0.020357150584459305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0031932813581079245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, continuous, circling tracking shot captures a medium shot at eye level, steadily rotating around a man engaged in an intense balance exercise. He is in an abstract world composed entirely of clean, flat polygons, atop a patchwork of vast, translucent geometric shapes suspended in an endless, dark void. The man, with an athletic build, wears a sleek, form-fitting, emerald-green high-tech bodysuit that perfectly outlines every inch of his taut muscles. In the center of his chest, a geometric, abstract symbol emits a soft, rhythmic cyan light, casting a cool halo around his focused jaw. He carefully stands on one leg on a narrow, hexagonal platform with a faintly luminescent edge, his other leg extended behind him for balance, his arms raised, his fingers making minute, precise adjustments to his body's subtle movements. Beneath his sweat-stained short hair, his eyes gazed unwaveringly at a single point in the void ahead. His tightly pursed lips and taut jawline betrayed his extraordinary concentration and willpower. The scene was dominated by a stark palette of black, green, and cyan. The cool white light emanating from the polygonal edges and the cyan glow of the emblem on his chest provided the sole illumination, creating a tranquil atmosphere that felt futuristic, intensely focused, and filled with inner tension.", + "seed": 24443, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_3046_seed_24443_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029841855308040975, + "vlm_logprob": -0.029841855308040975, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.029841855308040975, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07296279072761536 + ], + "opposite_logprob_tokens": [ + -2.947962760925293 + ], + "logprob_reward": -0.07296279072761536, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07296279072761536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14640071988105774, + 0.05979228392243385, + 0.25390511751174927, + 0.06624515354633331, + 0.0935339480638504, + 0.05524323880672455, + 0.06872785091400146, + 0.07429835945367813, + 0.12286007404327393, + 0.05899316444993019 + ], + "credit_map": [ + 0.020021634176373482, + 0.01603454351425171, + 0.01760128140449524, + 0.012774090282619, + 0.026950569823384285, + 0.020700568333268166, + 0.014278934337198734, + 0.01803911291062832, + 0.010892054997384548, + 0.006975007709115744, + 0.010573373176157475, + 0.00963064655661583, + 0.005423435475677252, + 0.004644982982426882, + 0.004919665865600109, + 0.0067331185564398766, + 0.01372618693858385, + 0.009259367361664772, + 0.08945497125387192, + 0.020951738581061363, + 0.011625725775957108, + 0.06525735557079315, + 0.03274574130773544, + 0.010884063318371773, + 0.006612060125917196, + 0.005798473954200745, + 0.00798938237130642, + 0.011361945420503616, + 0.007127068471163511, + 0.0060102203860878944, + 0.01135787833482027, + 0.00998813845217228, + 0.008401826024055481, + 0.010721812024712563, + 0.012951220385730267, + 0.0154481902718544, + 0.01258448138833046, + 0.006063899490982294, + 0.016078030690550804, + 0.011284500360488892, + 0.006664956919848919, + 0.00497313030064106, + 0.009690477512776852, + 0.004456235561519861, + 0.005923583172261715, + 0.007022480946034193, + 0.004914219956845045, + 0.011598157696425915, + 0.006679181940853596, + 0.004887199494987726, + 0.019013304263353348, + 0.004310351796448231, + 0.005359121598303318, + 0.01820910908281803, + 0.006210930645465851, + 0.004058656748384237, + 0.006450970657169819, + 0.007739000953733921, + 0.023395011201500893, + 0.008536946959793568, + 0.0060716127045452595, + 0.009393631480634212, + 0.006235545966774225, + 0.00647564884275198, + 0.010296144522726536, + 0.015608823858201504, + 0.014907622709870338, + 0.009673284366726875, + 0.02294336073100567, + 0.017085853964090347, + 0.019931627437472343, + 0.012413366697728634, + 0.006081043742597103, + 0.008186932653188705, + 0.0060836635529994965, + 0.0056959656067192554, + 0.009795823134481907, + 0.008066145703196526, + 0.007260625716298819, + 0.007822967134416103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.07296279072761536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008675740100443363 + ], + "opposite_logprob_tokens": [ + -4.758675575256348 + ], + "logprob_reward": -0.008675740100443363, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008675740100443363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24295255541801453, + 0.09463566541671753, + 0.1643708348274231, + 0.05214085802435875, + 0.10362639278173447, + 0.04828912019729614, + 0.06009411811828613, + 0.07054407149553299, + 0.10065785050392151, + 0.06268849968910217 + ], + "credit_map": [ + 0.04579447954893112, + 0.01892097480595112, + 0.0267118439078331, + 0.02526959218084812, + 0.03137602657079697, + 0.035175763070583344, + 0.0249320138245821, + 0.03477185219526291, + 0.009496178478002548, + 0.008555258624255657, + 0.012652966193854809, + 0.016193734481930733, + 0.012987000867724419, + 0.007362755481153727, + 0.01160361897200346, + 0.015784161165356636, + 0.009971564635634422, + 0.01006191223859787, + 0.048060860484838486, + 0.014604405499994755, + 0.015230256132781506, + 0.03481766954064369, + 0.022836579009890556, + 0.00878758355975151, + 0.005271874368190765, + 0.0063541908748447895, + 0.004793970379978418, + 0.00684224022552371, + 0.006469703279435635, + 0.011297298595309258, + 0.006576259154826403, + 0.004535321611911058, + 0.01462150365114212, + 0.015271446667611599, + 0.012811101973056793, + 0.014161545783281326, + 0.009747061878442764, + 0.007040877360850573, + 0.015069853514432907, + 0.014903008937835693, + 0.004071165341883898, + 0.007079631090164185, + 0.005320975091308355, + 0.007471112068742514, + 0.00664269644767046, + 0.0054583437740802765, + 0.00617635203525424, + 0.006068845745176077, + 0.004826923832297325, + 0.007136629428714514, + 0.012895307503640652, + 0.006771740037947893, + 0.007067847531288862, + 0.007437300402671099, + 0.009453237056732178, + 0.004505135118961334, + 0.005892885383218527, + 0.008092928677797318, + 0.008866493590176105, + 0.014273558743298054, + 0.00814205501228571, + 0.006054084748029709, + 0.01026071049273014, + 0.008961357176303864, + 0.0096508814021945, + 0.011275135912001133, + 0.012637641280889511, + 0.010583935305476189, + 0.01322084665298462, + 0.01375878881663084, + 0.01771569438278675, + 0.011814916506409645, + 0.0054327757097780704, + 0.00570361502468586, + 0.004445651546120644, + 0.005620598793029785, + 0.011337519623339176, + 0.009661175310611725, + 0.009132915176451206, + 0.01135424617677927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.008675740100443363 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792386554181576 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792386554181576, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792386554181576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13458503782749176, + 0.040123965591192245, + 0.14458459615707397, + 0.3078206777572632, + 0.08067793399095535, + 0.033646535128355026, + 0.06637071073055267, + 0.07384087145328522, + 0.07797996699810028, + 0.04036970064043999 + ], + "credit_map": [ + 0.020579868927598, + 0.019165273755788803, + 0.013086038641631603, + 0.011753862723708153, + 0.018907418474555016, + 0.015435165725648403, + 0.01472171489149332, + 0.020935697481036186, + 0.006075767334550619, + 0.006689154077321291, + 0.00550453457981348, + 0.005284471902996302, + 0.006141336169093847, + 0.0027804591227322817, + 0.0038173221983015537, + 0.0038309211377054453, + 0.011066291481256485, + 0.010868729092180729, + 0.041367195546627045, + 0.01463430654257536, + 0.009875383228063583, + 0.03028082475066185, + 0.0165658351033926, + 0.009926027618348598, + 0.02269153669476509, + 0.03619691729545593, + 0.03463711589574814, + 0.040468115359544754, + 0.05192101374268532, + 0.03035384975373745, + 0.044343188405036926, + 0.04720892757177353, + 0.007868231274187565, + 0.00780410785228014, + 0.014722773805260658, + 0.01222644466906786, + 0.01247312780469656, + 0.005588768515735865, + 0.010206373408436775, + 0.009788105264306068, + 0.004163176286965609, + 0.004330975469201803, + 0.0069685219787061214, + 0.004093404393643141, + 0.0027874803636223078, + 0.003745108610019088, + 0.003036735812202096, + 0.004521132912486792, + 0.006285650189965963, + 0.004808488767594099, + 0.01715443842113018, + 0.003765176283195615, + 0.005508927628397942, + 0.01944664865732193, + 0.005370463244616985, + 0.004030921496450901, + 0.006285015493631363, + 0.009634831920266151, + 0.01875275932252407, + 0.011753753758966923, + 0.00779018085449934, + 0.007304975297302008, + 0.007274741772562265, + 0.005044613964855671, + 0.007389554288238287, + 0.01164202205836773, + 0.00878793653100729, + 0.006373397074639797, + 0.01129878219217062, + 0.012118014506995678, + 0.011601211503148079, + 0.008769048377871513, + 0.0034683842677623034, + 0.00526476139202714, + 0.0028345633763819933, + 0.0027783673722296953, + 0.007114874664694071, + 0.006270131096243858, + 0.0062906998209655285, + 0.00634791562333703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.009792386554181576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549577958881855 + ], + "opposite_logprob_tokens": [ + -2.929957866668701 + ], + "logprob_reward": -0.0549577958881855, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549577958881855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14026308059692383, + 0.045944079756736755, + 0.16000248491764069, + 0.0867476612329483, + 0.05112301558256149, + 0.04651721566915512, + 0.2685382664203644, + 0.07479637861251831, + 0.07811255753040314, + 0.04795529320836067 + ], + "credit_map": [ + 0.024062925949692726, + 0.01597333699464798, + 0.01529745664447546, + 0.01804889179766178, + 0.019061116501688957, + 0.017199749127030373, + 0.016627879813313484, + 0.013991734012961388, + 0.004851112142205238, + 0.005050416570156813, + 0.005653925705701113, + 0.007154511287808418, + 0.008040010929107666, + 0.0035394062288105488, + 0.00615679332986474, + 0.005497904494404793, + 0.008033432997763157, + 0.005618975963443518, + 0.057107001543045044, + 0.01340993121266365, + 0.00781206926330924, + 0.04065818339586258, + 0.020639361813664436, + 0.0067235324531793594, + 0.007100283168256283, + 0.008114110678434372, + 0.009601759724318981, + 0.012591609731316566, + 0.008680799975991249, + 0.012330523692071438, + 0.01803174614906311, + 0.01029682345688343, + 0.00619784090667963, + 0.0063227759674191475, + 0.00984741561114788, + 0.007090046536177397, + 0.00434590270742774, + 0.0030181866604834795, + 0.005765623413026333, + 0.008535224944353104, + 0.00433920556679368, + 0.005773729644715786, + 0.005524787120521069, + 0.00690569169819355, + 0.00652327761054039, + 0.0055219815112650394, + 0.006304692476987839, + 0.005623849108815193, + 0.01927156187593937, + 0.012190110050141811, + 0.08902621269226074, + 0.013753021135926247, + 0.012694108299911022, + 0.08859354257583618, + 0.022273065522313118, + 0.010736634023487568, + 0.0076911416836082935, + 0.007378840819001198, + 0.022045986726880074, + 0.009471101686358452, + 0.005951488856226206, + 0.008579058572649956, + 0.007074081338942051, + 0.006604672875255346, + 0.007325455546379089, + 0.008754651993513107, + 0.012867559678852558, + 0.006809891667217016, + 0.011537225916981697, + 0.010763433761894703, + 0.011694971472024918, + 0.008359366096556187, + 0.005453931167721748, + 0.004678704310208559, + 0.003742990782484412, + 0.00401452137157321, + 0.006456305272877216, + 0.007560955826193094, + 0.006107383407652378, + 0.009940503165125847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.0549577958881855 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820563269779086 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.002820563269779086, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820563269779086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13243715465068817, + 0.06498809158802032, + 0.31959590315818787, + 0.04680575802922249, + 0.056721609085798264, + 0.04291052743792534, + 0.11935794353485107, + 0.05215724930167198, + 0.085502028465271, + 0.0795237123966217 + ], + "credit_map": [ + 0.02014530822634697, + 0.015643853694200516, + 0.014352946542203426, + 0.018405362963676453, + 0.016089338809251785, + 0.01807917095720768, + 0.01613760180771351, + 0.013583571650087833, + 0.011815477162599564, + 0.010447399690747261, + 0.010352526791393757, + 0.011083768680691719, + 0.005624833982437849, + 0.004513975698500872, + 0.005259550642222166, + 0.00589055847376585, + 0.018529225140810013, + 0.014001908712089062, + 0.10937655717134476, + 0.02753792330622673, + 0.016363350674510002, + 0.07880276441574097, + 0.040881119668483734, + 0.014103071764111519, + 0.005337756592780352, + 0.004619714803993702, + 0.006052284501492977, + 0.006726015359163284, + 0.005966550204902887, + 0.0037303846329450607, + 0.007810791488736868, + 0.00656225997954607, + 0.005581514909863472, + 0.008458076976239681, + 0.007399316411465406, + 0.008238033391535282, + 0.005979400128126144, + 0.003142351983115077, + 0.008146329782903194, + 0.0097765838727355, + 0.007064686622470617, + 0.006074856035411358, + 0.00625207182019949, + 0.0040192510932683945, + 0.005994030274450779, + 0.004136944189667702, + 0.0036352300085127354, + 0.00573345459997654, + 0.013077789917588234, + 0.010385156609117985, + 0.02955533191561699, + 0.01035917829722166, + 0.008479882031679153, + 0.022342637181282043, + 0.010255677625536919, + 0.01490228995680809, + 0.005356972571462393, + 0.007080945186316967, + 0.007427194155752659, + 0.010694739408791065, + 0.006140479352325201, + 0.004107422661036253, + 0.00629190681502223, + 0.005057591944932938, + 0.006713234353810549, + 0.009690036997199059, + 0.010275201871991158, + 0.007507792208343744, + 0.01038945559412241, + 0.013255558907985687, + 0.009792404249310493, + 0.017878340557217598, + 0.009237231686711311, + 0.00999678485095501, + 0.00757917296141386, + 0.0076814028434455395, + 0.01074374932795763, + 0.012202238664031029, + 0.009931374341249466, + 0.012151760049164295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.002820563269779086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, continuous, circling tracking shot captures a medium shot at eye level, steadily rotating around a man engaged in an intense balance exercise. He is in an abstract world composed entirely of clean, flat polygons, atop a patchwork of vast, translucent geometric shapes suspended in an endless, dark void. The man, with an athletic build, wears a sleek, form-fitting, emerald-green high-tech bodysuit that perfectly outlines every inch of his taut muscles. In the center of his chest, a geometric, abstract symbol emits a soft, rhythmic cyan light, casting a cool halo around his focused jaw. He carefully stands on one leg on a narrow, hexagonal platform with a faintly luminescent edge, his other leg extended behind him for balance, his arms raised, his fingers making minute, precise adjustments to his body's subtle movements. Beneath his sweat-stained short hair, his eyes gazed unwaveringly at a single point in the void ahead. His tightly pursed lips and taut jawline betrayed his extraordinary concentration and willpower. The scene was dominated by a stark palette of black, green, and cyan. The cool white light emanating from the polygonal edges and the cyan glow of the emblem on his chest provided the sole illumination, creating a tranquil atmosphere that felt futuristic, intensely focused, and filled with inner tension.", + "seed": 24444, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_3046_seed_24444_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02576679168269038, + "vlm_logprob": -0.02576679168269038, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02576679168269038, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06121606007218361 + ], + "opposite_logprob_tokens": [ + -3.311216115951538 + ], + "logprob_reward": -0.06121606007218361, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06121606007218361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14824867248535156, + 0.07448213547468185, + 0.24550576508045197, + 0.07443956285715103, + 0.08566548675298691, + 0.047248564660549164, + 0.04671947658061981, + 0.0634213238954544, + 0.14089582860469818, + 0.07337315380573273 + ], + "credit_map": [ + 0.019633622840046883, + 0.01567702740430832, + 0.026780307292938232, + 0.025579486042261124, + 0.013316942378878593, + 0.011105292476713657, + 0.014957493171095848, + 0.02119849994778633, + 0.010524679906666279, + 0.00797999370843172, + 0.010718228295445442, + 0.009542702697217464, + 0.008204278536140919, + 0.014547211118042469, + 0.0065576680935919285, + 0.006407379172742367, + 0.02163299173116684, + 0.03249719738960266, + 0.023082295432686806, + 0.0253074299544096, + 0.0462469607591629, + 0.03876841813325882, + 0.0317160040140152, + 0.02625446580350399, + 0.010809505358338356, + 0.010922761633992195, + 0.013670198619365692, + 0.008094885386526585, + 0.007362755481153727, + 0.00658104894682765, + 0.008893479593098164, + 0.008104929700493813, + 0.013202635571360588, + 0.008997215889394283, + 0.012208058498799801, + 0.010985258035361767, + 0.008353975601494312, + 0.01240175124257803, + 0.0087866997346282, + 0.010729897767305374, + 0.009138929657638073, + 0.006539736874401569, + 0.005550206173211336, + 0.0044958326034247875, + 0.0033628924284130335, + 0.009661080315709114, + 0.003412750782445073, + 0.0050871348939836025, + 0.006003289949148893, + 0.006309068761765957, + 0.007151724770665169, + 0.005641827825456858, + 0.007332980167120695, + 0.006169445812702179, + 0.003872535889968276, + 0.004238598980009556, + 0.009064985439181328, + 0.006497221067547798, + 0.005487813148647547, + 0.005585600156337023, + 0.011661275289952755, + 0.009831798262894154, + 0.00854004081338644, + 0.006752589717507362, + 0.013496555387973785, + 0.02289443090558052, + 0.014380275271832943, + 0.027493024244904518, + 0.015293659642338753, + 0.016240153461694717, + 0.019864734262228012, + 0.011232986114919186, + 0.00525027746334672, + 0.009143760427832603, + 0.018572760745882988, + 0.009387630969285965, + 0.012209758162498474, + 0.006676012650132179, + 0.0070874616503715515, + 0.005045488942414522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.06121606007218361 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016096873208880424 + ], + "opposite_logprob_tokens": [ + -4.141097068786621 + ], + "logprob_reward": -0.016096873208880424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096873208880424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11602336168289185, + 0.025270162150263786, + 0.049719199538230896, + 0.12180005013942719, + 0.024925442412495613, + 0.03613940626382828, + 0.04089818149805069, + 0.22876538336277008, + 0.2192370444536209, + 0.1372217833995819 + ], + "credit_map": [ + 0.01666967011988163, + 0.009614783339202404, + 0.016571534797549248, + 0.01517800334841013, + 0.007722507230937481, + 0.023501558229327202, + 0.009325560182332993, + 0.017439749091863632, + 0.0026895420160144567, + 0.005357895977795124, + 0.0022597205825150013, + 0.0030555317644029856, + 0.003332919906824827, + 0.0028819444123655558, + 0.003144162241369486, + 0.002548446413129568, + 0.0063506136648356915, + 0.005630459636449814, + 0.005583839025348425, + 0.0052899862639606, + 0.007290968671441078, + 0.007612098474055529, + 0.005809728521853685, + 0.006151502951979637, + 0.012683550827205181, + 0.00970305036753416, + 0.025415407493710518, + 0.011376572772860527, + 0.00651790713891387, + 0.008164316415786743, + 0.033456411212682724, + 0.014482831582427025, + 0.0035234668757766485, + 0.0032119217794388533, + 0.0026589282788336277, + 0.003074155654758215, + 0.0033291832078248262, + 0.0038030301220715046, + 0.0025184755213558674, + 0.002806280739605427, + 0.004082137253135443, + 0.00696608005091548, + 0.005007889587432146, + 0.0041158851236104965, + 0.0038766036741435528, + 0.004977047443389893, + 0.003047477686777711, + 0.004066283814609051, + 0.004647052846848965, + 0.00911908969283104, + 0.0033441116102039814, + 0.003982922527939081, + 0.006043522153049707, + 0.004128857981413603, + 0.003838371019810438, + 0.005794253665953875, + 0.0286166500300169, + 0.02637222222983837, + 0.025675591081380844, + 0.033562008291482925, + 0.031634967774152756, + 0.026972221210598946, + 0.029127001762390137, + 0.026804717257618904, + 0.01934979483485222, + 0.03416653349995613, + 0.02211330458521843, + 0.05255161598324776, + 0.02137313224375248, + 0.022615354508161545, + 0.032094039022922516, + 0.014973266050219536, + 0.008877858519554138, + 0.019128834828734398, + 0.025814900174736977, + 0.017889399081468582, + 0.02408618852496147, + 0.012956456281244755, + 0.016976727172732353, + 0.011491428129374981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.016096873208880424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204349368810654 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204349368810654, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204349368810654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07300622761249542, + 0.031565554440021515, + 0.04788936302065849, + 0.4718506634235382, + 0.06647581607103348, + 0.03641147166490555, + 0.03483140468597412, + 0.03995051607489586, + 0.14655140042304993, + 0.05146759748458862 + ], + "credit_map": [ + 0.006434181239455938, + 0.007697906345129013, + 0.008595560677349567, + 0.009167992509901524, + 0.00679077347740531, + 0.010890696197748184, + 0.007652483880519867, + 0.015776634216308594, + 0.0031053104903548956, + 0.005473276600241661, + 0.0030613804701715708, + 0.004789017606526613, + 0.004327347036451101, + 0.0035288846120238304, + 0.004660929553210735, + 0.002619406208395958, + 0.00482779648154974, + 0.005830229725688696, + 0.005390769802033901, + 0.00451484601944685, + 0.009107672609388828, + 0.006964306812733412, + 0.004595268052071333, + 0.006658474914729595, + 0.04888739436864853, + 0.028603939339518547, + 0.11805795878171921, + 0.03474944829940796, + 0.021671874448657036, + 0.022279080003499985, + 0.13966041803359985, + 0.05794055759906769, + 0.010104609653353691, + 0.008640429936349392, + 0.007551838643848896, + 0.008134808391332626, + 0.008487340062856674, + 0.008750899694859982, + 0.006665725260972977, + 0.00814016256481409, + 0.004903546068817377, + 0.004957664757966995, + 0.005109529010951519, + 0.004124004393815994, + 0.005475878715515137, + 0.003669044701382518, + 0.004248300101608038, + 0.003923502750694752, + 0.0037652812898159027, + 0.00648800702765584, + 0.003577792551368475, + 0.0032964556012302637, + 0.0060866945423185825, + 0.00374341057613492, + 0.003005723934620619, + 0.0048680379986763, + 0.004892112221568823, + 0.004984082654118538, + 0.003558507189154625, + 0.006671017501503229, + 0.005009765736758709, + 0.005492046009749174, + 0.005437520332634449, + 0.003905463730916381, + 0.011298309080302715, + 0.018517281860113144, + 0.014161486178636551, + 0.03597616031765938, + 0.014180628582835197, + 0.016225513070821762, + 0.023342423141002655, + 0.012849598191678524, + 0.003551850561052561, + 0.00695373397320509, + 0.01135823130607605, + 0.007687569595873356, + 0.007903068326413631, + 0.005524878390133381, + 0.004680189769715071, + 0.0038080739323049784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.014204349368810654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370270878076553 + ], + "opposite_logprob_tokens": [ + -3.4087026119232178 + ], + "logprob_reward": -0.03370270878076553, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370270878076553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05170045793056488, + 0.039911869913339615, + 0.22320473194122314, + 0.3808561861515045, + 0.02405438758432865, + 0.049190640449523926, + 0.03034435771405697, + 0.07399395108222961, + 0.10553295910358429, + 0.021210428327322006 + ], + "credit_map": [ + 0.00821646861732006, + 0.006528747733682394, + 0.006840339861810207, + 0.007828684523701668, + 0.0033667385578155518, + 0.00837714783847332, + 0.0043877423740923405, + 0.006154598668217659, + 0.004325286019593477, + 0.0067454115487635136, + 0.005733998958021402, + 0.005393695086240768, + 0.004032233264297247, + 0.005261440295726061, + 0.00402494752779603, + 0.004394857678562403, + 0.03089410811662674, + 0.021870197728276253, + 0.03330788388848305, + 0.02291952259838581, + 0.028753632679581642, + 0.03905686363577843, + 0.02464151941239834, + 0.021761013194918633, + 0.03735256567597389, + 0.03217941150069237, + 0.0925392210483551, + 0.030623754486441612, + 0.022145044058561325, + 0.022780295461416245, + 0.09823025017976761, + 0.04500564932823181, + 0.0026051083113998175, + 0.0026548190508037806, + 0.0028777108527719975, + 0.0025892299599945545, + 0.003443680237978697, + 0.0035572098568081856, + 0.0032687149941921234, + 0.003057915950194001, + 0.005066459998488426, + 0.006788875907659531, + 0.006671255454421043, + 0.00597158819437027, + 0.0060424781404435635, + 0.0069095962680876255, + 0.005700594279915094, + 0.006039793603122234, + 0.0034939562901854515, + 0.004848902579396963, + 0.0035820321645587683, + 0.003019706578925252, + 0.003928217571228743, + 0.004849092569202185, + 0.003363825613632798, + 0.0032586255110800266, + 0.008628590032458305, + 0.009312042966485023, + 0.008375662378966808, + 0.00931557361036539, + 0.010838940739631653, + 0.00936610996723175, + 0.009146200492978096, + 0.009010830894112587, + 0.00935451127588749, + 0.015829794108867645, + 0.01109881792217493, + 0.023256685584783554, + 0.009809024631977081, + 0.012068222276866436, + 0.015225931070744991, + 0.008889979682862759, + 0.001648669014684856, + 0.002427071798592806, + 0.0035673596430569887, + 0.0027485284954309464, + 0.0044149188324809074, + 0.0028119469061493874, + 0.001897970912978053, + 0.0016939652850851417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.03370270878076553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036139669828116894 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036139669828116894, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036139669828116894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12972743809223175, + 0.04614640027284622, + 0.20819403231143951, + 0.2763062119483948, + 0.05633058026432991, + 0.10575974732637405, + 0.02859930880367756, + 0.03889848664402962, + 0.04265544191002846, + 0.0673823356628418 + ], + "credit_map": [ + 0.017261920496821404, + 0.009895037859678268, + 0.008604953996837139, + 0.012335256673395634, + 0.008406195789575577, + 0.03634147718548775, + 0.009421606548130512, + 0.027461007237434387, + 0.005284520331770182, + 0.00575591903179884, + 0.005470026284456253, + 0.006619506981223822, + 0.004877247381955385, + 0.006215016357600689, + 0.006154757458716631, + 0.00576940830796957, + 0.039301589131355286, + 0.01913483254611492, + 0.03553728014230728, + 0.02144620195031166, + 0.023899367079138756, + 0.029996691271662712, + 0.02030479907989502, + 0.018573308363556862, + 0.02860451489686966, + 0.02146608941257, + 0.05788808688521385, + 0.02996869757771492, + 0.02090548351407051, + 0.02736041508615017, + 0.05192209407687187, + 0.03819083422422409, + 0.00907118245959282, + 0.006107652094215155, + 0.00716582965105772, + 0.005802494939416647, + 0.00659091817215085, + 0.008175499737262726, + 0.005449766293168068, + 0.007967241108417511, + 0.010532306507229805, + 0.013401266187429428, + 0.012815212830901146, + 0.013686981052160263, + 0.01573904976248741, + 0.01318465918302536, + 0.013271079398691654, + 0.013129203580319881, + 0.0028776740655303, + 0.005982575938105583, + 0.003302955301478505, + 0.0024653165601193905, + 0.0051016598008573055, + 0.0030030757188796997, + 0.002821259433403611, + 0.0030447938479483128, + 0.004857333842664957, + 0.0038844309747219086, + 0.0037289189640432596, + 0.004008392803370953, + 0.005358600057661533, + 0.006345807109028101, + 0.0058289626613259315, + 0.004886047914624214, + 0.0045261322520673275, + 0.005070900544524193, + 0.00507457135245204, + 0.008489261381328106, + 0.004182689357548952, + 0.004566548392176628, + 0.005851222202181816, + 0.004894122947007418, + 0.005459578707814217, + 0.007763952948153019, + 0.013800977729260921, + 0.007167227566242218, + 0.010114035569131374, + 0.008518432267010212, + 0.007347106002271175, + 0.007211024407297373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0036139669828116894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, continuous, circling tracking shot captures a medium shot at eye level, steadily rotating around a man engaged in an intense balance exercise. He is in an abstract world composed entirely of clean, flat polygons, atop a patchwork of vast, translucent geometric shapes suspended in an endless, dark void. The man, with an athletic build, wears a sleek, form-fitting, emerald-green high-tech bodysuit that perfectly outlines every inch of his taut muscles. In the center of his chest, a geometric, abstract symbol emits a soft, rhythmic cyan light, casting a cool halo around his focused jaw. He carefully stands on one leg on a narrow, hexagonal platform with a faintly luminescent edge, his other leg extended behind him for balance, his arms raised, his fingers making minute, precise adjustments to his body's subtle movements. Beneath his sweat-stained short hair, his eyes gazed unwaveringly at a single point in the void ahead. His tightly pursed lips and taut jawline betrayed his extraordinary concentration and willpower. The scene was dominated by a stark palette of black, green, and cyan. The cool white light emanating from the polygonal edges and the cyan glow of the emblem on his chest provided the sole illumination, creating a tranquil atmosphere that felt futuristic, intensely focused, and filled with inner tension.", + "seed": 24445, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_2/index_3046_seed_24445_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0316363365156576, + "vlm_logprob": -0.0316363365156576, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.0316363365156576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07627185434103012 + ], + "opposite_logprob_tokens": [ + -2.8262717723846436 + ], + "logprob_reward": -0.07627185434103012, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07627185434103012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2971819341182709, + 0.064351886510849, + 0.08020897209644318, + 0.07783134281635284, + 0.07589355856180191, + 0.11101263016462326, + 0.0822288766503334, + 0.0510856993496418, + 0.08203001320362091, + 0.07817511260509491 + ], + "credit_map": [ + 0.020434938371181488, + 0.06429615616798401, + 0.021195251494646072, + 0.040499527007341385, + 0.060218002647161484, + 0.04472936689853668, + 0.023804308846592903, + 0.022004371508955956, + 0.006091836374253035, + 0.007960110902786255, + 0.008353030309081078, + 0.010952826589345932, + 0.008638085797429085, + 0.008019092492759228, + 0.007689591962844133, + 0.006647318601608276, + 0.008536696434020996, + 0.007764511276036501, + 0.008906872011721134, + 0.013426958583295345, + 0.016372112557291985, + 0.011564347892999649, + 0.00696842186152935, + 0.006669049151241779, + 0.009804590605199337, + 0.007146991789340973, + 0.010404675267636776, + 0.01567324809730053, + 0.011503355577588081, + 0.009687834419310093, + 0.006455597002059221, + 0.007155050989240408, + 0.009878016076982021, + 0.01178136095404625, + 0.004250787198543549, + 0.0121479956433177, + 0.006654334720224142, + 0.004581954795867205, + 0.012586617842316628, + 0.014012490399181843, + 0.011575253680348396, + 0.011553131975233555, + 0.013463259674608707, + 0.011544744484126568, + 0.020683245733380318, + 0.01420858409255743, + 0.012561420910060406, + 0.015422982163727283, + 0.008539645932614803, + 0.008747334592044353, + 0.010276741348206997, + 0.008360383100807667, + 0.01620914600789547, + 0.012227804400026798, + 0.0076461536809802055, + 0.010221668519079685, + 0.005995563697069883, + 0.004618891514837742, + 0.005667233839631081, + 0.008882109075784683, + 0.005289154127240181, + 0.004840669687837362, + 0.012872898951172829, + 0.0029191793873906136, + 0.011923552490770817, + 0.013657732866704464, + 0.010009347461163998, + 0.00813754741102457, + 0.008861551992595196, + 0.012011576443910599, + 0.0076034823432564735, + 0.009825224988162518, + 0.013976577669382095, + 0.009230163879692554, + 0.009321660734713078, + 0.010775957256555557, + 0.009859863668680191, + 0.0072326366789639, + 0.009661931544542313, + 0.00811631791293621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.07627185434103012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004677663091570139 + ], + "opposite_logprob_tokens": [ + -5.379677772521973 + ], + "logprob_reward": -0.004677663091570139, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004677663091570139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09013117104768753, + 0.0770374983549118, + 0.10015693306922913, + 0.03729555383324623, + 0.047761496156454086, + 0.10975348949432373, + 0.06505310535430908, + 0.04998897388577461, + 0.39467713236808777, + 0.02814461663365364 + ], + "credit_map": [ + 0.006126932334154844, + 0.01873028464615345, + 0.006208440288901329, + 0.014973757788538933, + 0.01690180040895939, + 0.015421871095895767, + 0.005408371798694134, + 0.0063597108237445354, + 0.008249489590525627, + 0.009787526912987232, + 0.009849409572780132, + 0.010314824990928173, + 0.011986270546913147, + 0.01197098009288311, + 0.008519253693521023, + 0.006359744351357222, + 0.01356928888708353, + 0.010284771211445332, + 0.009393698535859585, + 0.015137859620153904, + 0.017562799155712128, + 0.01023914385586977, + 0.01584472693502903, + 0.008124643936753273, + 0.004463603720068932, + 0.004048009868711233, + 0.004599323961883783, + 0.004933719057589769, + 0.005965657997876406, + 0.0052368370816111565, + 0.004129130858927965, + 0.0039192731492221355, + 0.005879237316548824, + 0.006805124692618847, + 0.0019445712678134441, + 0.007656336762011051, + 0.005336939822882414, + 0.003167699556797743, + 0.00839268323034048, + 0.008578901179134846, + 0.010229035280644894, + 0.013574551790952682, + 0.00887175090610981, + 0.01031777448952198, + 0.02112348936498165, + 0.01769210211932659, + 0.014395647682249546, + 0.013549136929214, + 0.006687275134027004, + 0.007514666765928268, + 0.00630075391381979, + 0.0060256896540522575, + 0.009752641431987286, + 0.01004512794315815, + 0.00780178839340806, + 0.01092515978962183, + 0.005690376274287701, + 0.004045224282890558, + 0.005572895519435406, + 0.0084930919110775, + 0.00433052983134985, + 0.004854314494878054, + 0.013057035394012928, + 0.003945501986891031, + 0.028307365253567696, + 0.030053244903683662, + 0.04852943494915962, + 0.06444251537322998, + 0.05392666906118393, + 0.04720049723982811, + 0.06234067678451538, + 0.05987674742937088, + 0.004636181518435478, + 0.004134202841669321, + 0.0035226864274591208, + 0.0032421122305095196, + 0.003862787503749132, + 0.0021987026557326317, + 0.004918274469673634, + 0.0016296692192554474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.004677663091570139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005265890154987574 + ], + "opposite_logprob_tokens": [ + -5.255265712738037 + ], + "logprob_reward": -0.005265890154987574, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265890154987574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08526281267404556, + 0.08058236539363861, + 0.05463024228811264, + 0.04016624391078949, + 0.04347163066267967, + 0.07299590110778809, + 0.1147107481956482, + 0.05686962977051735, + 0.4185161590576172, + 0.032794300466775894 + ], + "credit_map": [ + 0.008010783232748508, + 0.012151035480201244, + 0.007886342704296112, + 0.010045602917671204, + 0.015965797007083893, + 0.018142718821763992, + 0.006917716935276985, + 0.0061428118497133255, + 0.008376603946089745, + 0.00954191293567419, + 0.011773843318223953, + 0.009975842200219631, + 0.012232639826834202, + 0.012586384080350399, + 0.007257532794028521, + 0.008837606757879257, + 0.007674405816942453, + 0.00498268473893404, + 0.005661932285875082, + 0.007443625014275312, + 0.010820885188877583, + 0.006435133516788483, + 0.006122733000665903, + 0.005488836672157049, + 0.00402827886864543, + 0.003849655855447054, + 0.005726813804358244, + 0.004529581870883703, + 0.006934020202606916, + 0.005294907372444868, + 0.004741271026432514, + 0.0050617121160030365, + 0.00574827054515481, + 0.005928784143179655, + 0.002518748166039586, + 0.004415867384523153, + 0.004727659281343222, + 0.0036529116332530975, + 0.0062537966296076775, + 0.010225587524473667, + 0.007455772254616022, + 0.008805576711893082, + 0.006280974950641394, + 0.006963549647480249, + 0.014206387102603912, + 0.01206026691943407, + 0.007954414933919907, + 0.009268955327570438, + 0.012081429362297058, + 0.013454499654471874, + 0.011049924418330193, + 0.009548474103212357, + 0.01767602190375328, + 0.018839705735445023, + 0.009972247295081615, + 0.022088443860411644, + 0.006790886167436838, + 0.005627182312309742, + 0.0066809263080358505, + 0.009228461422026157, + 0.0062216902151703835, + 0.004531370475888252, + 0.013468565419316292, + 0.004320542793720961, + 0.02432587370276451, + 0.0392499640583992, + 0.04157973825931549, + 0.049642328172922134, + 0.07091538608074188, + 0.05440029129385948, + 0.0897899940609932, + 0.0486125573515892, + 0.006520989816635847, + 0.005719072185456753, + 0.004114218056201935, + 0.0038125254213809967, + 0.004518199246376753, + 0.0029864441603422165, + 0.0032480673398822546, + 0.001874779467470944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.005265890154987574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002492994070053 + ], + "opposite_logprob_tokens": [ + -2.6950249671936035 + ], + "logprob_reward": -0.07002492994070053, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002492994070053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11488266289234161, + 0.05663924664258957, + 0.10130904614925385, + 0.07649564743041992, + 0.03647364303469658, + 0.10929835587739944, + 0.05429283529520035, + 0.07192046195268631, + 0.07267577946186066, + 0.3060123324394226 + ], + "credit_map": [ + 0.011211760342121124, + 0.016895348206162453, + 0.009311511181294918, + 0.014604375697672367, + 0.021188315004110336, + 0.022281479090452194, + 0.008111421950161457, + 0.01127846073359251, + 0.00519247492775321, + 0.0063600726425647736, + 0.005651081446558237, + 0.005196274723857641, + 0.012148335576057434, + 0.009305177256464958, + 0.005822341423481703, + 0.006963489577174187, + 0.014192436821758747, + 0.011752950958907604, + 0.014366243034601212, + 0.015760304406285286, + 0.009625720791518688, + 0.01576562225818634, + 0.011554272845387459, + 0.008291497826576233, + 0.008391798473894596, + 0.007603608071804047, + 0.007618618197739124, + 0.010041574016213417, + 0.011267120949923992, + 0.011015546508133411, + 0.00831646379083395, + 0.012240921147167683, + 0.004114832263439894, + 0.0052103581838309765, + 0.0026007788255810738, + 0.006604344584047794, + 0.0048251342959702015, + 0.0022585263941437006, + 0.006219167727977037, + 0.00464050006121397, + 0.007378765847533941, + 0.017250925302505493, + 0.009289335459470749, + 0.012173924595117569, + 0.01734687201678753, + 0.015100978314876556, + 0.015518655069172382, + 0.015238896943628788, + 0.005770056042820215, + 0.005027368199080229, + 0.005746135953813791, + 0.004210062325000763, + 0.008357684127986431, + 0.00957998726516962, + 0.006452525965869427, + 0.009149016812443733, + 0.00823430810123682, + 0.007302809040993452, + 0.009172504767775536, + 0.008584742434322834, + 0.008704510517418385, + 0.006670461967587471, + 0.017054155468940735, + 0.006196967326104641, + 0.0059642549604177475, + 0.0071860007010400295, + 0.008558299392461777, + 0.010618000291287899, + 0.006501613184809685, + 0.0071713803336024284, + 0.015702469274401665, + 0.01097375713288784, + 0.05644218996167183, + 0.045054469257593155, + 0.02867000177502632, + 0.04825469106435776, + 0.03239099308848381, + 0.025586245581507683, + 0.039695367217063904, + 0.029918378219008446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.07002492994070053 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, continuous, circling tracking shot around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a sleek, form-fitting, emerald-green high-tech bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a geometric, abstract symbol on the man's chest emitting a soft, rhythmic cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man standing on one leg on a narrow, hexagonal platform with a faintly luminescent edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment composed of clean, flat polygons and translucent geometric shapes suspended in a dark void?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019413450499996543 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.0019413450499996543, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019413450499996543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15216664969921112, + 0.11584216356277466, + 0.08956074714660645, + 0.05053792893886566, + 0.05107525736093521, + 0.14586490392684937, + 0.04444211721420288, + 0.12765055894851685, + 0.12058291584253311, + 0.10227672010660172 + ], + "credit_map": [ + 0.01655687391757965, + 0.016742592677474022, + 0.012567608617246151, + 0.017926808446645737, + 0.03439819812774658, + 0.0325230173766613, + 0.009178077802062035, + 0.012273475527763367, + 0.012918654829263687, + 0.015603603795170784, + 0.011815067380666733, + 0.009217333048582077, + 0.021980412304401398, + 0.02056231163442135, + 0.012550406157970428, + 0.011194374412298203, + 0.00972494762390852, + 0.008363724686205387, + 0.008827866055071354, + 0.011100664734840393, + 0.013395841233432293, + 0.015176056884229183, + 0.011692102067172527, + 0.01127955038100481, + 0.0056768180802464485, + 0.005283001344650984, + 0.0062525151297450066, + 0.0077633922919631, + 0.0068342844024300575, + 0.0067242905497550964, + 0.00510036526247859, + 0.006903265602886677, + 0.007393878884613514, + 0.00699648167937994, + 0.003966483753174543, + 0.008359416387975216, + 0.00618578027933836, + 0.0034331532660871744, + 0.006920429412275553, + 0.007819636724889278, + 0.011984596028923988, + 0.016662025824189186, + 0.010650347918272018, + 0.015266046859323978, + 0.017658010125160217, + 0.03243494778871536, + 0.01728331856429577, + 0.023925626650452614, + 0.004657475743442774, + 0.004539075307548046, + 0.00572620565071702, + 0.005433851853013039, + 0.008369001559913158, + 0.007016786374151707, + 0.004025828093290329, + 0.004673892166465521, + 0.01631820760667324, + 0.010288908146321774, + 0.027979612350463867, + 0.013835642486810684, + 0.011576937511563301, + 0.015363287180662155, + 0.012693645432591438, + 0.019594326615333557, + 0.01465430948883295, + 0.023165125399827957, + 0.018482547253370285, + 0.010180733166635036, + 0.011817734688520432, + 0.017266377806663513, + 0.011341078206896782, + 0.01367500051856041, + 0.022937843576073647, + 0.01824130490422249, + 0.009563334286212921, + 0.014380356296896935, + 0.011120197363197803, + 0.008553523570299149, + 0.010557197965681553, + 0.0069229560904204845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0019413450499996543 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_tensor([4458, 4458])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_tensor([4458, 4458])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..950acde0c73d991ceff636b96d4fd77ce69c5093 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_tensor([4458, 4458])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up shot, captured with a macro lens and a cross-sectional perspective reminiscent of a popular science program, focuses on the lock of an antique camphorwood chest. The chest, a deep reddish-brown, its wood grain mellow and lustered by age, is inlaid with a meticulously crafted brass lock. Its surface, intricately carved with a curly grass pattern, has a warm matte finish, enhanced by age, with only a soft metallic sheen reflected at the edges. Illuminated by bright, focused overhead lighting, the entire image is vibrantly detailed, creating an atmosphere of focused concentration and precision. As the lock cylinder is slowly unscrewed by an invisible force, the camera simultaneously reveals its incredible inner workings: a series of tiny brass gears activate, meshing silently and smoothly, each tooth gleaming with the precision of craftsmanship. These gears, through a connecting rod, steadily and effortlessly drive a wooden tenon concealed within the chest, allowing it to slide smoothly and precisely out of its perfectly fitted mortise. The entire process is clearly visible in ultra-high-definition image quality. The fiber of the wood contrasts sharply with the cold texture of the metal, perfectly demonstrating the ingenious combination of ancient mortise and tenon craftsmanship and precision machinery, and is full of awe and admiration for ancient wisdom.", + "seed": 1264, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_4458_seed_1264_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08991079265251756, + "vlm_logprob": -0.08991079265251756, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.08991079265251756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02014533430337906 + ], + "opposite_logprob_tokens": [ + -4.270145416259766 + ], + "logprob_reward": -0.02014533430337906, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02014533430337906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1504540741443634, + 0.161848783493042, + 0.17498958110809326, + 0.11216942965984344, + 0.0476861298084259, + 0.049736108630895615, + 0.11166355013847351, + 0.06570255756378174, + 0.0914703980088234, + 0.034279435873031616 + ], + "credit_map": [ + 0.01133678201586008, + 0.02407091297209263, + 0.02421940676867962, + 0.019851213321089745, + 0.010375208221375942, + 0.011589779518544674, + 0.024140499532222748, + 0.0248702522367239, + 0.019456077367067337, + 0.017260322347283363, + 0.03758694976568222, + 0.02658069133758545, + 0.0070932707749307156, + 0.009148657321929932, + 0.024046652019023895, + 0.020676154643297195, + 0.021805880591273308, + 0.017335552722215652, + 0.03070247359573841, + 0.026941003277897835, + 0.011144638992846012, + 0.010788552463054657, + 0.022772744297981262, + 0.03349873796105385, + 0.011264431290328503, + 0.015035866759717464, + 0.012913912534713745, + 0.01243474893271923, + 0.008887500502169132, + 0.008994088508188725, + 0.02414887212216854, + 0.018489999696612358, + 0.005623185075819492, + 0.004017861559987068, + 0.006997647695243359, + 0.004960335325449705, + 0.002849995857104659, + 0.003509335219860077, + 0.013080519624054432, + 0.006647242698818445, + 0.00645265681669116, + 0.005374170374125242, + 0.0065653580240905285, + 0.0073186540976166725, + 0.002873179968446493, + 0.003264972008764744, + 0.00829419493675232, + 0.009592918679118156, + 0.016043871641159058, + 0.0122501440346241, + 0.012675262056291103, + 0.019775390625, + 0.009871910326182842, + 0.008102593012154102, + 0.015099376440048218, + 0.017844997346401215, + 0.008286227472126484, + 0.009534556418657303, + 0.010393008589744568, + 0.01430059690028429, + 0.0038628692273050547, + 0.0032000772189348936, + 0.007559335790574551, + 0.008565881289541721, + 0.009873293340206146, + 0.008044728077948093, + 0.01028499100357294, + 0.01372629776597023, + 0.007370397914201021, + 0.007054856978356838, + 0.017768284305930138, + 0.017347538843750954, + 0.0037665937561541796, + 0.0022390715312212706, + 0.004527752753347158, + 0.005307134240865707, + 0.002365534659475088, + 0.0027317204512655735, + 0.0053943912498652935, + 0.007947235368192196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.02014533430337906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113825023174286 + ], + "opposite_logprob_tokens": [ + -3.288113832473755 + ], + "logprob_reward": -0.038113825023174286, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113825023174286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11817006766796112, + 0.05871664360165596, + 0.08310519903898239, + 0.1533665508031845, + 0.08842187374830246, + 0.05112416669726372, + 0.06931959837675095, + 0.1245075985789299, + 0.10960347950458527, + 0.1436648666858673 + ], + "credit_map": [ + 0.01381062064319849, + 0.014594298787415028, + 0.01291019469499588, + 0.016779564321041107, + 0.009148528799414635, + 0.014550519175827503, + 0.015347152948379517, + 0.021029183641076088, + 0.0072067552246153355, + 0.0061228894628584385, + 0.00964012835174799, + 0.008597943000495434, + 0.0034324589651077986, + 0.004899987950921059, + 0.008102073334157467, + 0.010714400559663773, + 0.008844711817800999, + 0.007512880023568869, + 0.012354149483144283, + 0.013647460378706455, + 0.004778513219207525, + 0.005274997092783451, + 0.012626423500478268, + 0.018066061660647392, + 0.018357152119278908, + 0.026485761627554893, + 0.025034533813595772, + 0.026551926508545876, + 0.008701061829924583, + 0.008906844072043896, + 0.01892082765698433, + 0.02040843293070793, + 0.013763715513050556, + 0.008327484130859375, + 0.011726259253919125, + 0.011027463711798191, + 0.006800299044698477, + 0.006488136947154999, + 0.015527973882853985, + 0.014760538004338741, + 0.008727844804525375, + 0.005083850584924221, + 0.005789550952613354, + 0.006722469814121723, + 0.002675434807315469, + 0.0038858135230839252, + 0.008368038572371006, + 0.009871160611510277, + 0.010811555199325085, + 0.006155016832053661, + 0.011871382594108582, + 0.01174174528568983, + 0.005326248239725828, + 0.0043481881730258465, + 0.009478403255343437, + 0.009587052278220654, + 0.014182349666953087, + 0.012901733629405499, + 0.010083966888487339, + 0.018971601501107216, + 0.009039221331477165, + 0.011022667400538921, + 0.02653544768691063, + 0.021770603954792023, + 0.01152410265058279, + 0.012380523607134819, + 0.01224211510270834, + 0.019899364560842514, + 0.007663859985768795, + 0.008017896674573421, + 0.016656456515192986, + 0.021219152957201004, + 0.015232197009027004, + 0.010118132457137108, + 0.014636492356657982, + 0.017288632690906525, + 0.00879680272191763, + 0.010231210850179195, + 0.02940746583044529, + 0.037953924387693405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.038113825023174286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263399798423052 + ], + "opposite_logprob_tokens": [ + -5.255263328552246 + ], + "logprob_reward": -0.005263399798423052, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263399798423052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1531371921300888, + 0.23860624432563782, + 0.19333864748477936, + 0.08312433212995529, + 0.08081775158643723, + 0.07669385522603989, + 0.0292695090174675, + 0.041946493089199066, + 0.03860623762011528, + 0.06445970386266708 + ], + "credit_map": [ + 0.014975341968238354, + 0.019052773714065552, + 0.02024046517908573, + 0.02040286175906658, + 0.00925557129085064, + 0.011601255275309086, + 0.0293719582259655, + 0.028236960992217064, + 0.026549063622951508, + 0.022777380421757698, + 0.058495718985795975, + 0.03580760583281517, + 0.010461789555847645, + 0.012344720773398876, + 0.039242684841156006, + 0.032927267253398895, + 0.022399961948394775, + 0.026646101847290993, + 0.03617455065250397, + 0.033206213265657425, + 0.01028252113610506, + 0.01006321795284748, + 0.02459721267223358, + 0.029968855902552605, + 0.009561248123645782, + 0.010259163565933704, + 0.011400244198739529, + 0.00861956737935543, + 0.004497138783335686, + 0.004121759440749884, + 0.01956237107515335, + 0.015102828852832317, + 0.006837046705186367, + 0.007500419858843088, + 0.011280410923063755, + 0.009351965971291065, + 0.0036578106228262186, + 0.005318705458194017, + 0.021490223705768585, + 0.015381162986159325, + 0.009726030752062798, + 0.00716323172673583, + 0.009099096059799194, + 0.009244340471923351, + 0.004290090873837471, + 0.003979936707764864, + 0.01594816893339157, + 0.017242955043911934, + 0.004033692646771669, + 0.0035309193190187216, + 0.002990719163790345, + 0.004743071738630533, + 0.002490338869392872, + 0.0025565451942384243, + 0.004726616200059652, + 0.004197604022920132, + 0.00444423221051693, + 0.0028731937054544687, + 0.006000288296490908, + 0.0058353315107524395, + 0.003112359205260873, + 0.0031231294851750135, + 0.008358550257980824, + 0.008199404925107956, + 0.0047573428601026535, + 0.002931836061179638, + 0.00806901790201664, + 0.005006114020943642, + 0.002351603237912059, + 0.0026385330129414797, + 0.006498947273939848, + 0.006352838594466448, + 0.007151521276682615, + 0.005628215614706278, + 0.01034470833837986, + 0.011235899291932583, + 0.004260494839400053, + 0.005251728929579258, + 0.009424224495887756, + 0.011162908747792244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.005263399798423052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160390242934227 + ], + "opposite_logprob_tokens": [ + -1.9103902578353882 + ], + "logprob_reward": -0.160390242934227, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160390242934227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15711309015750885, + 0.09530405700206757, + 0.2713816165924072, + 0.10276644676923752, + 0.06105055287480354, + 0.04300248995423317, + 0.060710128396749496, + 0.07941990345716476, + 0.08957050740718842, + 0.03968127444386482 + ], + "credit_map": [ + 0.011942470446228981, + 0.02088918350636959, + 0.019122490659356117, + 0.02252880670130253, + 0.008809447288513184, + 0.010501370765268803, + 0.031768739223480225, + 0.0315505750477314, + 0.010677520185709, + 0.013804774731397629, + 0.01619846001267433, + 0.01706458069384098, + 0.004426546860486269, + 0.006017057690769434, + 0.013376046903431416, + 0.013739067129790783, + 0.03383108973503113, + 0.02715533785521984, + 0.04894821345806122, + 0.04790282994508743, + 0.01668717712163925, + 0.016785012558102608, + 0.034479208290576935, + 0.045592762529850006, + 0.013944859616458416, + 0.010768108069896698, + 0.01810271665453911, + 0.011272204108536243, + 0.0067861126735806465, + 0.007410891819745302, + 0.020538242533802986, + 0.013943313620984554, + 0.00731716537848115, + 0.005592076573520899, + 0.009044178761541843, + 0.005799267906695604, + 0.003635444212704897, + 0.004610569681972265, + 0.015114720910787582, + 0.00993712805211544, + 0.006238370668143034, + 0.00405863719061017, + 0.00575645174831152, + 0.006344381719827652, + 0.00255500222556293, + 0.004082439001649618, + 0.006812200881540775, + 0.0071550048887729645, + 0.008919124491512775, + 0.005983744747936726, + 0.007560086902230978, + 0.011221089400351048, + 0.003961451817303896, + 0.004337044432759285, + 0.00919617060571909, + 0.009531415067613125, + 0.009554986841976643, + 0.007305547595024109, + 0.009867505170404911, + 0.013291149400174618, + 0.006350877229124308, + 0.006483863107860088, + 0.010692661628127098, + 0.015873311087489128, + 0.008946835994720459, + 0.006751478649675846, + 0.015605529770255089, + 0.01605965383350849, + 0.006361972540616989, + 0.0076667251996695995, + 0.011814632453024387, + 0.01636367477476597, + 0.00354287913069129, + 0.003898059017956257, + 0.006570467259734869, + 0.008832021616399288, + 0.0020489958114922047, + 0.0020622529555112123, + 0.00652271555736661, + 0.006203881464898586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.160390242934227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256411612033844 + ], + "opposite_logprob_tokens": [ + -1.600641131401062 + ], + "logprob_reward": -0.2256411612033844, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256411612033844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1464793086051941, + 0.06014419347047806, + 0.11857513338327408, + 0.06551402807235718, + 0.07776812463998795, + 0.10092713683843613, + 0.04970413073897362, + 0.18493519723415375, + 0.11520406603813171, + 0.08074871450662613 + ], + "credit_map": [ + 0.013267740607261658, + 0.017560908570885658, + 0.02383730560541153, + 0.01851075142621994, + 0.008611042983829975, + 0.009140348061919212, + 0.028092194348573685, + 0.027459019795060158, + 0.005869970191270113, + 0.007000333163887262, + 0.013861353509128094, + 0.007911103777587414, + 0.004078212659806013, + 0.004379577003419399, + 0.009831305593252182, + 0.0072123389691114426, + 0.016682090237736702, + 0.010211360640823841, + 0.016038386151194572, + 0.015009443275630474, + 0.009319759905338287, + 0.00931762158870697, + 0.02289484813809395, + 0.019101623445749283, + 0.008134052157402039, + 0.005405495874583721, + 0.01520108338445425, + 0.009410670958459377, + 0.003309779567644, + 0.00496685178950429, + 0.008049193769693375, + 0.011036898009479046, + 0.008713608607649803, + 0.00681063299998641, + 0.013308079913258553, + 0.008914431557059288, + 0.004614060278981924, + 0.005418488755822182, + 0.016062315553426743, + 0.013926508836448193, + 0.01322182733565569, + 0.007474479265511036, + 0.016207154840230942, + 0.014926033094525337, + 0.00453207828104496, + 0.007152556907385588, + 0.01970626786351204, + 0.01770673505961895, + 0.006785763427615166, + 0.0033703686203807592, + 0.007565928157418966, + 0.006314527243375778, + 0.002899641403928399, + 0.004140892531722784, + 0.008559013716876507, + 0.010067994706332684, + 0.022498777136206627, + 0.015657098963856697, + 0.022705262526869774, + 0.032252490520477295, + 0.015411083586513996, + 0.015750639140605927, + 0.0250414926558733, + 0.03561835736036301, + 0.011819092556834221, + 0.009252688847482204, + 0.018445879220962524, + 0.021543558686971664, + 0.00878180842846632, + 0.011489491909742355, + 0.014091327786445618, + 0.019780224189162254, + 0.008001267910003662, + 0.005310415290296078, + 0.009781007654964924, + 0.007770042400807142, + 0.0055123851634562016, + 0.006380747072398663, + 0.018209848552942276, + 0.019783003255724907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.2256411612033844 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot, captured with a macro lens and a cross-sectional perspective reminiscent of a popular science program, focuses on the lock of an antique camphorwood chest. The chest, a deep reddish-brown, its wood grain mellow and lustered by age, is inlaid with a meticulously crafted brass lock. Its surface, intricately carved with a curly grass pattern, has a warm matte finish, enhanced by age, with only a soft metallic sheen reflected at the edges. Illuminated by bright, focused overhead lighting, the entire image is vibrantly detailed, creating an atmosphere of focused concentration and precision. As the lock cylinder is slowly unscrewed by an invisible force, the camera simultaneously reveals its incredible inner workings: a series of tiny brass gears activate, meshing silently and smoothly, each tooth gleaming with the precision of craftsmanship. These gears, through a connecting rod, steadily and effortlessly drive a wooden tenon concealed within the chest, allowing it to slide smoothly and precisely out of its perfectly fitted mortise. The entire process is clearly visible in ultra-high-definition image quality. The fiber of the wood contrasts sharply with the cold texture of the metal, perfectly demonstrating the ingenious combination of ancient mortise and tenon craftsmanship and precision machinery, and is full of awe and admiration for ancient wisdom.", + "seed": 1265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_4458_seed_1265_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.034045522147789595, + "vlm_logprob": -0.034045522147789595, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.034045522147789595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012255012057721615 + ], + "opposite_logprob_tokens": [ + -4.7622551918029785 + ], + "logprob_reward": -0.012255012057721615, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012255012057721615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22630950808525085, + 0.12277451157569885, + 0.0316166914999485, + 0.048754166811704636, + 0.0962657555937767, + 0.08774306625127792, + 0.06832117587327957, + 0.07512255012989044, + 0.09910309314727783, + 0.1439894437789917 + ], + "credit_map": [ + 0.047807756811380386, + 0.021075721830129623, + 0.03122233971953392, + 0.01978454925119877, + 0.01954433135688305, + 0.017011279240250587, + 0.029271218925714493, + 0.040592338889837265, + 0.01988915354013443, + 0.016830967739224434, + 0.017088141292333603, + 0.010279475711286068, + 0.012949892319738865, + 0.009604930877685547, + 0.013341479934751987, + 0.02279048040509224, + 0.004151250701397657, + 0.004430034663528204, + 0.003390695434063673, + 0.003087088465690613, + 0.003989724908024073, + 0.002817863831296563, + 0.004266566131263971, + 0.005483470857143402, + 0.005904851481318474, + 0.007240056060254574, + 0.00430081645026803, + 0.005347609054297209, + 0.008412226103246212, + 0.005584414582699537, + 0.005553392227739096, + 0.006410802714526653, + 0.021818015724420547, + 0.010778311640024185, + 0.01429258193820715, + 0.005380811635404825, + 0.007853941060602665, + 0.007378467824310064, + 0.008714570663869381, + 0.02004905417561531, + 0.015410060994327068, + 0.007613517809659243, + 0.008397393859922886, + 0.007003736682236195, + 0.010228201746940613, + 0.006670928560197353, + 0.013292760588228703, + 0.019126474857330322, + 0.0128127196803689, + 0.005890136118978262, + 0.0051557756960392, + 0.005841877777129412, + 0.008097616955637932, + 0.008822493255138397, + 0.010527685284614563, + 0.011172876693308353, + 0.009799712337553501, + 0.006478676572442055, + 0.005902974866330624, + 0.003647205885499716, + 0.013022392056882381, + 0.007980093359947205, + 0.010624310001730919, + 0.017667189240455627, + 0.017026256769895554, + 0.012963817454874516, + 0.01078647468239069, + 0.014169854111969471, + 0.005759727209806442, + 0.007702744100242853, + 0.016991419717669487, + 0.013702803291380405, + 0.01275708433240652, + 0.011643105186522007, + 0.015269480645656586, + 0.020715514197945595, + 0.018512245267629623, + 0.028226133435964584, + 0.009966565296053886, + 0.02689933404326439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.012255012057721615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308326914906502 + ], + "opposite_logprob_tokens": [ + -3.1680831909179688 + ], + "logprob_reward": -0.04308326914906502, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308326914906502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16182225942611694, + 0.13525360822677612, + 0.07404515147209167, + 0.07357479631900787, + 0.05810699611902237, + 0.09973213821649551, + 0.05294546112418175, + 0.05562352389097214, + 0.07379214465618134, + 0.2151038944721222 + ], + "credit_map": [ + 0.0228128544986248, + 0.022227613255381584, + 0.01958266645669937, + 0.01600019820034504, + 0.017611250281333923, + 0.013385936617851257, + 0.017948970198631287, + 0.03225276246666908, + 0.017574619501829147, + 0.020207617431879044, + 0.014663067646324635, + 0.014661884866654873, + 0.014414021745324135, + 0.010481900535523891, + 0.018681015819311142, + 0.02456948347389698, + 0.017776355147361755, + 0.0068997652269899845, + 0.00999447051435709, + 0.00465778075158596, + 0.004925626330077648, + 0.004758340772241354, + 0.007915961556136608, + 0.017116853967308998, + 0.01777311973273754, + 0.007369650527834892, + 0.007678392343223095, + 0.004922804888337851, + 0.00743499118834734, + 0.005571477580815554, + 0.009484394453465939, + 0.013339963741600513, + 0.011823171749711037, + 0.00732669560238719, + 0.007732282858341932, + 0.0038795473519712687, + 0.006605812348425388, + 0.0049835373647511005, + 0.00514986552298069, + 0.010606082156300545, + 0.015518534928560257, + 0.011710017919540405, + 0.007894095033407211, + 0.007070601452142, + 0.020023709163069725, + 0.01195490825921297, + 0.008953520096838474, + 0.01660674810409546, + 0.008539428934454918, + 0.00708438316360116, + 0.005268187262117863, + 0.006199272349476814, + 0.005712909623980522, + 0.004665574990212917, + 0.009606343694031239, + 0.005869362968951464, + 0.006502858828753233, + 0.004957018420100212, + 0.004816188011318445, + 0.003645916236564517, + 0.009385479614138603, + 0.004892296623438597, + 0.006498766131699085, + 0.01492499839514494, + 0.011172741651535034, + 0.007901035249233246, + 0.008774640038609505, + 0.010965215042233467, + 0.004049480427056551, + 0.007010419387370348, + 0.012027347460389137, + 0.01189125794917345, + 0.0347483791410923, + 0.01661834493279457, + 0.025414858013391495, + 0.0284101665019989, + 0.031833089888095856, + 0.03350158408284187, + 0.01720644161105156, + 0.027371034026145935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.04308326914906502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002202229807153344 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.002202229807153344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002202229807153344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39320674538612366, + 0.08816279470920563, + 0.06279683858156204, + 0.06621363759040833, + 0.051348403096199036, + 0.11829382181167603, + 0.040893543511629105, + 0.04703221842646599, + 0.027248533442616463, + 0.10480339825153351 + ], + "credit_map": [ + 0.06405967473983765, + 0.054956451058387756, + 0.05063880607485771, + 0.03558172285556793, + 0.04199308529496193, + 0.029542498290538788, + 0.044957175850868225, + 0.07147735357284546, + 0.014030493795871735, + 0.01178718265146017, + 0.01206129789352417, + 0.006956093478947878, + 0.010260035283863544, + 0.007272628601640463, + 0.009670213796198368, + 0.01612485572695732, + 0.01578158140182495, + 0.006133960094302893, + 0.008014323189854622, + 0.004414914175868034, + 0.0038262330926954746, + 0.004509889520704746, + 0.005568370223045349, + 0.014547573402523994, + 0.018236668780446053, + 0.005153432488441467, + 0.008392472751438618, + 0.0037440636660903692, + 0.0042524454183876514, + 0.004581035580486059, + 0.006886601448059082, + 0.014966921880841255, + 0.01164156012237072, + 0.004751854110509157, + 0.007638238370418549, + 0.003275945084169507, + 0.00385249312967062, + 0.0036004239227622747, + 0.005624822340905666, + 0.010963072068989277, + 0.016571717336773872, + 0.013487102463841438, + 0.009724589064717293, + 0.010004709474742413, + 0.02573120780289173, + 0.014172452501952648, + 0.013162507675588131, + 0.015439546667039394, + 0.0072756242007017136, + 0.0030520097352564335, + 0.003401894820854068, + 0.004285912495106459, + 0.0037344687152653933, + 0.005020268261432648, + 0.007205020636320114, + 0.006918347906321287, + 0.00711255194619298, + 0.004552014637738466, + 0.004249301739037037, + 0.0030225699301809072, + 0.006674484815448523, + 0.00468188151717186, + 0.006215298548340797, + 0.01052411925047636, + 0.0038973467890173197, + 0.0027698816265910864, + 0.0032075115013867617, + 0.005027217324823141, + 0.0014835374895483255, + 0.0028095939196646214, + 0.004083063453435898, + 0.003970382269471884, + 0.016624944284558296, + 0.00688542565330863, + 0.014340673573315144, + 0.014997590333223343, + 0.013366260565817356, + 0.017905455082654953, + 0.007842397317290306, + 0.012840666808187962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.002202229807153344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370305523276329 + ], + "opposite_logprob_tokens": [ + -3.408703088760376 + ], + "logprob_reward": -0.03370305523276329, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370305523276329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10203299671411514, + 0.09660882502794266, + 0.04736388847231865, + 0.06614389270544052, + 0.02291754074394703, + 0.18585650622844696, + 0.05326613783836365, + 0.11119728535413742, + 0.05868164449930191, + 0.2559312582015991 + ], + "credit_map": [ + 0.01812869682908058, + 0.012217308394610882, + 0.013979918323457241, + 0.009686090052127838, + 0.012327553704380989, + 0.008077948354184628, + 0.0125321876257658, + 0.015083294361829758, + 0.020751958712935448, + 0.010849927552044392, + 0.012095862999558449, + 0.0080695366486907, + 0.007711774669587612, + 0.006984807550907135, + 0.010681853629648685, + 0.019463103264570236, + 0.01038265135139227, + 0.0051523917354643345, + 0.005914664827287197, + 0.004556300584226847, + 0.003386119846254587, + 0.0034133736044168472, + 0.004789115395396948, + 0.009769272990524769, + 0.012847092933952808, + 0.008378707803785801, + 0.006776121910661459, + 0.006005189847201109, + 0.007739158812910318, + 0.005786546505987644, + 0.007267190143465996, + 0.0113438805565238, + 0.0031451964750885963, + 0.003400459885597229, + 0.001989427488297224, + 0.0024503094609826803, + 0.003738541156053543, + 0.0024227285757660866, + 0.0027813243214040995, + 0.0029895526822656393, + 0.02640039660036564, + 0.02229197695851326, + 0.01469352561980486, + 0.0123008843511343, + 0.03029780462384224, + 0.018077434971928596, + 0.02160533517599106, + 0.040189146995544434, + 0.00885117519646883, + 0.00447035999968648, + 0.005194543395191431, + 0.0056806630454957485, + 0.005663530435413122, + 0.005114696454256773, + 0.009368490427732468, + 0.008922676555812359, + 0.017139511182904243, + 0.009974547661840916, + 0.009824808686971664, + 0.00707788672298193, + 0.019539086148142815, + 0.009845257736742496, + 0.013593855313956738, + 0.02420232445001602, + 0.01241455227136612, + 0.006101751700043678, + 0.006742578465491533, + 0.007636006455868483, + 0.003248555352911353, + 0.00582903204485774, + 0.008309590630233288, + 0.008399576880037785, + 0.04186515137553215, + 0.01720430701971054, + 0.034833624958992004, + 0.04055412486195564, + 0.034095343202352524, + 0.03694668784737587, + 0.019626997411251068, + 0.03080500289797783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.03370305523276329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898404449224472 + ], + "opposite_logprob_tokens": [ + -2.578984022140503 + ], + "logprob_reward": -0.07898404449224472, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898404449224472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24351926147937775, + 0.05771540850400925, + 0.027785293757915497, + 0.024593358859419823, + 0.03519347682595253, + 0.1259389966726303, + 0.05778311565518379, + 0.09578229486942291, + 0.0872143805027008, + 0.2444744110107422 + ], + "credit_map": [ + 0.04917239770293236, + 0.030944686383008957, + 0.03259424492716789, + 0.017241036519408226, + 0.02642466127872467, + 0.01892620511353016, + 0.02807372435927391, + 0.040142305195331573, + 0.011394624598324299, + 0.00650556106120348, + 0.007023866754025221, + 0.004967804532498121, + 0.005092103499919176, + 0.004310920834541321, + 0.007307668216526508, + 0.011112859472632408, + 0.0032935854978859425, + 0.003785684471949935, + 0.0025480338372290134, + 0.0038743638433516026, + 0.0033502005971968174, + 0.0024695247411727905, + 0.004356319550424814, + 0.004107579588890076, + 0.004476281348615885, + 0.002876865677535534, + 0.0022499398328363895, + 0.0028300618287175894, + 0.0026992515195161104, + 0.001985198585316539, + 0.0034663863480091095, + 0.004009373486042023, + 0.007617845665663481, + 0.00446586636826396, + 0.0048894379287958145, + 0.0025543365627527237, + 0.00316150882281363, + 0.002640310674905777, + 0.003652367275208235, + 0.006211802829056978, + 0.015139518305659294, + 0.018487771973013878, + 0.008997105062007904, + 0.00955408439040184, + 0.02859494648873806, + 0.01562831923365593, + 0.008868635632097721, + 0.02066861093044281, + 0.009695622138679028, + 0.0040692053735256195, + 0.004408721812069416, + 0.004327402450144291, + 0.006024365313351154, + 0.00770079018548131, + 0.011746504344046116, + 0.009810504503548145, + 0.0154028395190835, + 0.008118600584566593, + 0.007088307291269302, + 0.006359538994729519, + 0.013522233814001083, + 0.008585680276155472, + 0.011870334856212139, + 0.024834763258695602, + 0.015611507929861546, + 0.008418465964496136, + 0.010158857330679893, + 0.011823827400803566, + 0.004510965198278427, + 0.008217698894441128, + 0.015075535513460636, + 0.013397520408034325, + 0.034727729856967926, + 0.018928781151771545, + 0.026645055040717125, + 0.026371197775006294, + 0.04041273146867752, + 0.0461370013654232, + 0.016475671902298927, + 0.034776221960783005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.07898404449224472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot, captured with a macro lens and a cross-sectional perspective reminiscent of a popular science program, focuses on the lock of an antique camphorwood chest. The chest, a deep reddish-brown, its wood grain mellow and lustered by age, is inlaid with a meticulously crafted brass lock. Its surface, intricately carved with a curly grass pattern, has a warm matte finish, enhanced by age, with only a soft metallic sheen reflected at the edges. Illuminated by bright, focused overhead lighting, the entire image is vibrantly detailed, creating an atmosphere of focused concentration and precision. As the lock cylinder is slowly unscrewed by an invisible force, the camera simultaneously reveals its incredible inner workings: a series of tiny brass gears activate, meshing silently and smoothly, each tooth gleaming with the precision of craftsmanship. These gears, through a connecting rod, steadily and effortlessly drive a wooden tenon concealed within the chest, allowing it to slide smoothly and precisely out of its perfectly fitted mortise. The entire process is clearly visible in ultra-high-definition image quality. The fiber of the wood contrasts sharply with the cold texture of the metal, perfectly demonstrating the ingenious combination of ancient mortise and tenon craftsmanship and precision machinery, and is full of awe and admiration for ancient wisdom.", + "seed": 1266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_4458_seed_1266_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04078228604048491, + "vlm_logprob": -0.04078228604048491, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.04078228604048491, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011717895045876503 + ], + "opposite_logprob_tokens": [ + -4.886717796325684 + ], + "logprob_reward": -0.011717895045876503, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011717895045876503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26305070519447327, + 0.09986428916454315, + 0.058401934802532196, + 0.027530912309885025, + 0.03313127160072327, + 0.0383598692715168, + 0.025280505418777466, + 0.052092358469963074, + 0.17145805060863495, + 0.2308301478624344 + ], + "credit_map": [ + 0.02645198069512844, + 0.03129721060395241, + 0.024959148839116096, + 0.03567499294877052, + 0.05067151039838791, + 0.03673310577869415, + 0.029864346608519554, + 0.027398405596613884, + 0.007194115314632654, + 0.013222139328718185, + 0.010398275218904018, + 0.017160559073090553, + 0.016605379059910774, + 0.012747296132147312, + 0.011454383842647076, + 0.01108213048428297, + 0.004528400953859091, + 0.007803893182426691, + 0.005985345225781202, + 0.008547917939722538, + 0.010347520001232624, + 0.008841237984597683, + 0.00646853307262063, + 0.005879080388695002, + 0.0026691826060414314, + 0.0026810530107468367, + 0.003367034485563636, + 0.0037912363186478615, + 0.003054360393434763, + 0.0035600464325398207, + 0.003998563624918461, + 0.004409432411193848, + 0.003583513433113694, + 0.0033230690751224756, + 0.0037473367992788553, + 0.00345996324904263, + 0.00571134127676487, + 0.005131499841809273, + 0.0042189969681203365, + 0.003955547232180834, + 0.004380532540380955, + 0.004598214756697416, + 0.0038693146780133247, + 0.004139396827667952, + 0.007865045219659805, + 0.006473802961409092, + 0.004221323411911726, + 0.002812235848978162, + 0.0024369671009480953, + 0.002831933554261923, + 0.0035126067232340574, + 0.0030105337500572205, + 0.002678454387933016, + 0.002934330375865102, + 0.004319621250033379, + 0.0035560557153075933, + 0.003116822335869074, + 0.003996707033365965, + 0.003975431434810162, + 0.003730101278051734, + 0.0150748360902071, + 0.016750557348132133, + 0.0028810801450163126, + 0.002566821873188019, + 0.018396412953734398, + 0.017827853560447693, + 0.019896313548088074, + 0.014092208817601204, + 0.026664238423109055, + 0.01939845085144043, + 0.025241410359740257, + 0.029941154643893242, + 0.024731041863560677, + 0.0404309518635273, + 0.021309679374098778, + 0.03557709977030754, + 0.04043395817279816, + 0.0307515449821949, + 0.020412854850292206, + 0.017182987183332443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.011717895045876503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370939567685127 + ], + "opposite_logprob_tokens": [ + -3.4087092876434326 + ], + "logprob_reward": -0.03370939567685127, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370939567685127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.169008269906044, + 0.09959618747234344, + 0.05502897873520851, + 0.05338333919644356, + 0.02168978750705719, + 0.026925934478640556, + 0.04318288341164589, + 0.04341137409210205, + 0.25054609775543213, + 0.23722711205482483 + ], + "credit_map": [ + 0.01854332536458969, + 0.020488908514380455, + 0.02298274077475071, + 0.030312057584524155, + 0.02413131296634674, + 0.020322410389780998, + 0.016459379345178604, + 0.015768133103847504, + 0.01019227597862482, + 0.011229236610233784, + 0.008357401005923748, + 0.0072198286652565, + 0.02339025028049946, + 0.016430934891104698, + 0.011508064344525337, + 0.011268203146755695, + 0.006252479273825884, + 0.004660732578486204, + 0.0072876461781561375, + 0.005406816489994526, + 0.009084722958505154, + 0.00713254464790225, + 0.006664397194981575, + 0.008539642207324505, + 0.005154954735189676, + 0.0068182568065822124, + 0.0050928727723658085, + 0.008419736288487911, + 0.00822347030043602, + 0.006129245739430189, + 0.00530975079163909, + 0.008235050365328789, + 0.0022496599704027176, + 0.0016717554535716772, + 0.0028409198857843876, + 0.0019575224723666906, + 0.002750996733084321, + 0.0027187259402126074, + 0.0030716690234839916, + 0.0044285389594733715, + 0.0025282339192926884, + 0.002780930371955037, + 0.0033265852835029364, + 0.0028997729532420635, + 0.003396042389795184, + 0.004692777059972286, + 0.0035203611478209496, + 0.0037812304217368364, + 0.004163563717156649, + 0.005487381014972925, + 0.004724869504570961, + 0.003972472157329321, + 0.008182334713637829, + 0.005880780518054962, + 0.005751156248152256, + 0.0050203269347548485, + 0.004118379205465317, + 0.005727238021790981, + 0.0050707305781543255, + 0.005619035568088293, + 0.006490823347121477, + 0.005532907322049141, + 0.004389951936900616, + 0.0064623067155480385, + 0.023745112121105194, + 0.032371122390031815, + 0.0337250716984272, + 0.030753469094634056, + 0.034822046756744385, + 0.031789977103471756, + 0.03380540385842323, + 0.02953389659523964, + 0.021073540672659874, + 0.0467848926782608, + 0.036170605570077896, + 0.02719162032008171, + 0.03582756221294403, + 0.03468405455350876, + 0.018897036090493202, + 0.016597796231508255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.03370939567685127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031963707879185677 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.0031963707879185677, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031963707879185677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26071062684059143, + 0.17765437066555023, + 0.0304384995251894, + 0.03648635745048523, + 0.03638157621026039, + 0.014518197625875473, + 0.11104642599821091, + 0.024401845410466194, + 0.08397181332111359, + 0.22439026832580566 + ], + "credit_map": [ + 0.021447818726301193, + 0.03662322089076042, + 0.026436656713485718, + 0.04770026355981827, + 0.03740815073251724, + 0.03051292710006237, + 0.030582139268517494, + 0.029999449849128723, + 0.012415697798132896, + 0.024818869307637215, + 0.01462771836668253, + 0.03182174265384674, + 0.029432659968733788, + 0.028159959241747856, + 0.018891628831624985, + 0.017486117780208588, + 0.0023877613712102175, + 0.0036167774815112352, + 0.002999331336468458, + 0.0027440295089036226, + 0.00728420028463006, + 0.0059696976095438, + 0.003132767276838422, + 0.002303937217220664, + 0.00276850420050323, + 0.005361500196158886, + 0.003171250456944108, + 0.005090299062430859, + 0.008469048887491226, + 0.006314706988632679, + 0.0032774419523775578, + 0.002033611061051488, + 0.002387545071542263, + 0.004519262816756964, + 0.0032822638750076294, + 0.0060018738731741905, + 0.006147051230072975, + 0.00600259006023407, + 0.003798218909651041, + 0.004242774564772844, + 0.0015639076009392738, + 0.0021635706070810556, + 0.0019692699424922466, + 0.0021894173696637154, + 0.0015553185949102044, + 0.0016450285911560059, + 0.0016768573550507426, + 0.0017548281466588378, + 0.00658538518473506, + 0.012700770981609821, + 0.00928573403507471, + 0.014932246878743172, + 0.021701443940401077, + 0.02153203636407852, + 0.012713862583041191, + 0.011594963259994984, + 0.0017699776217341423, + 0.0022883256897330284, + 0.00194455252494663, + 0.0025580439250916243, + 0.006156759336590767, + 0.006496245972812176, + 0.0014869875740259886, + 0.0017009546281769872, + 0.009616675786674023, + 0.00993187353014946, + 0.01064762007445097, + 0.0123393964022398, + 0.014292479492723942, + 0.011246000416576862, + 0.008595186285674572, + 0.007302592508494854, + 0.02404315024614334, + 0.05098370462656021, + 0.020454518496990204, + 0.03104814887046814, + 0.03547581657767296, + 0.033211760222911835, + 0.01715182512998581, + 0.012021351605653763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0031963707879185677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496332421898842 + ], + "opposite_logprob_tokens": [ + -2.9299633502960205 + ], + "logprob_reward": -0.05496332421898842, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496332421898842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24788683652877808, + 0.10676375031471252, + 0.04770462587475777, + 0.042467743158340454, + 0.025993529707193375, + 0.020465508103370667, + 0.049301840364933014, + 0.05916375666856766, + 0.09646964073181152, + 0.30378279089927673 + ], + "credit_map": [ + 0.021861832588911057, + 0.02271580696105957, + 0.039789795875549316, + 0.017297958955168724, + 0.04762620851397514, + 0.04386045038700104, + 0.02532871626317501, + 0.029406074434518814, + 0.007451181765645742, + 0.013189765624701977, + 0.011100409552454948, + 0.015114416368305683, + 0.019586816430091858, + 0.01786733977496624, + 0.010116140358150005, + 0.012337686493992805, + 0.0037046854849904776, + 0.007036041468381882, + 0.0048904502764344215, + 0.008382752537727356, + 0.0069622560404241085, + 0.006472927983850241, + 0.004799807444214821, + 0.005455705802887678, + 0.0032400486525148153, + 0.00585029274225235, + 0.0049627916887402534, + 0.006534019485116005, + 0.0055242558009922504, + 0.005286772735416889, + 0.004713109694421291, + 0.006356457248330116, + 0.001856961753219366, + 0.0026490171439945698, + 0.0029329368844628334, + 0.0027231210842728615, + 0.004285014234483242, + 0.004230261780321598, + 0.003145852591842413, + 0.0041703651659190655, + 0.0017640898004174232, + 0.0018330171005800366, + 0.002476408611983061, + 0.0019019762985408306, + 0.00344947655685246, + 0.003556195180863142, + 0.0025675338692963123, + 0.0029168114997446537, + 0.003137768479064107, + 0.006105071399360895, + 0.0044425479136407375, + 0.00710675586014986, + 0.008433863520622253, + 0.00912270788103342, + 0.005267094820737839, + 0.005686034914106131, + 0.0033006975427269936, + 0.0058060442097485065, + 0.005321293603628874, + 0.0065874806605279446, + 0.013021855615079403, + 0.013800375163555145, + 0.004815276246517897, + 0.006510732229799032, + 0.008520756848156452, + 0.012594999745488167, + 0.010851808823645115, + 0.01191890612244606, + 0.020764607936143875, + 0.01218077726662159, + 0.010434162802994251, + 0.009203612804412842, + 0.028758645057678223, + 0.06805247068405151, + 0.02817619778215885, + 0.04039823263883591, + 0.05508837103843689, + 0.040373269468545914, + 0.024726087227463722, + 0.018209507688879967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.05496332421898842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032444447278976 + ], + "opposite_logprob_tokens": [ + -2.3503243923187256 + ], + "logprob_reward": -0.10032444447278976, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032444447278976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27560630440711975, + 0.14323045313358307, + 0.02800232172012329, + 0.03891637921333313, + 0.04351762682199478, + 0.027967464178800583, + 0.04206583648920059, + 0.07466154545545578, + 0.11148206144571304, + 0.2145499438047409 + ], + "credit_map": [ + 0.020171336829662323, + 0.030216021463274956, + 0.030643075704574585, + 0.02855491451919079, + 0.0553412102162838, + 0.05018426477909088, + 0.029712826013565063, + 0.030782688409090042, + 0.010129514150321484, + 0.018072159960865974, + 0.01316728163510561, + 0.01987786404788494, + 0.027932293713092804, + 0.02541252411901951, + 0.013262614607810974, + 0.015376219525933266, + 0.0019322887528687716, + 0.003833831287920475, + 0.003019458381459117, + 0.004187798593193293, + 0.0041822283528745174, + 0.0034213850740343332, + 0.002949725603684783, + 0.004475609865039587, + 0.0029565722215920687, + 0.003732955316081643, + 0.005132032558321953, + 0.003498774254694581, + 0.0058807735331356525, + 0.004836149513721466, + 0.005487105343490839, + 0.007392019499093294, + 0.0030392054468393326, + 0.0042512984946370125, + 0.0050526089034974575, + 0.004537766799330711, + 0.00757383881136775, + 0.006934921257197857, + 0.004935200791805983, + 0.007192789576947689, + 0.002442810218781233, + 0.002834667917340994, + 0.0035449799615889788, + 0.003171595511958003, + 0.00440467381849885, + 0.004290975630283356, + 0.003245086409151554, + 0.004032676573842764, + 0.003680681809782982, + 0.004048809874802828, + 0.004482530523091555, + 0.004604271613061428, + 0.00973413698375225, + 0.008105897344648838, + 0.0035575940273702145, + 0.003851920133456588, + 0.00521314051002264, + 0.00842256098985672, + 0.007830717600882053, + 0.009379524737596512, + 0.013718152418732643, + 0.01365689281374216, + 0.0068594589829444885, + 0.009581107646226883, + 0.011396492831408978, + 0.012147371657192707, + 0.014775151386857033, + 0.017274118959903717, + 0.012987111695110798, + 0.012194824405014515, + 0.012157794088125229, + 0.018549205735325813, + 0.02404598332941532, + 0.043707940727472305, + 0.021826231852173805, + 0.030866311863064766, + 0.03741668909788132, + 0.026553558185696602, + 0.01577610895037651, + 0.01435713842511177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.10032444447278976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot, captured with a macro lens and a cross-sectional perspective reminiscent of a popular science program, focuses on the lock of an antique camphorwood chest. The chest, a deep reddish-brown, its wood grain mellow and lustered by age, is inlaid with a meticulously crafted brass lock. Its surface, intricately carved with a curly grass pattern, has a warm matte finish, enhanced by age, with only a soft metallic sheen reflected at the edges. Illuminated by bright, focused overhead lighting, the entire image is vibrantly detailed, creating an atmosphere of focused concentration and precision. As the lock cylinder is slowly unscrewed by an invisible force, the camera simultaneously reveals its incredible inner workings: a series of tiny brass gears activate, meshing silently and smoothly, each tooth gleaming with the precision of craftsmanship. These gears, through a connecting rod, steadily and effortlessly drive a wooden tenon concealed within the chest, allowing it to slide smoothly and precisely out of its perfectly fitted mortise. The entire process is clearly visible in ultra-high-definition image quality. The fiber of the wood contrasts sharply with the cold texture of the metal, perfectly demonstrating the ingenious combination of ancient mortise and tenon craftsmanship and precision machinery, and is full of awe and admiration for ancient wisdom.", + "seed": 1267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_3/index_4458_seed_1267_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08681891732849181, + "vlm_logprob": -0.08681891732849181, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.08681891732849181, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027946414425969124 + ], + "opposite_logprob_tokens": [ + -4.027946472167969 + ], + "logprob_reward": -0.027946414425969124, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.027946414425969124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17466458678245544, + 0.4027565121650696, + 0.06822666525840759, + 0.027710719034075737, + 0.05573977902531624, + 0.02598102204501629, + 0.01850035786628723, + 0.06103683263063431, + 0.0675690695643425, + 0.09781452268362045 + ], + "credit_map": [ + 0.016340214759111404, + 0.010418941266834736, + 0.015509866178035736, + 0.014891617000102997, + 0.011389769613742828, + 0.03979775682091713, + 0.016786132007837296, + 0.04953029006719589, + 0.10101967304944992, + 0.02102741412818432, + 0.06024162098765373, + 0.028348375111818314, + 0.032685067504644394, + 0.07273712009191513, + 0.03432079404592514, + 0.052376434206962585, + 0.011816763319075108, + 0.0047303675673902035, + 0.0076834941282868385, + 0.006674149073660374, + 0.005735104437917471, + 0.012637052685022354, + 0.006350382696837187, + 0.01259935088455677, + 0.0030161424074321985, + 0.001966508338227868, + 0.002787984674796462, + 0.003302342724055052, + 0.002374229719862342, + 0.005613740999251604, + 0.002778618596494198, + 0.005871148779988289, + 0.007721626199781895, + 0.003951461054384708, + 0.005207170732319355, + 0.004476235713809729, + 0.0037740974221378565, + 0.007908523082733154, + 0.006683745887130499, + 0.016016919165849686, + 0.004896859638392925, + 0.0025898092426359653, + 0.0026613252703100443, + 0.003315100446343422, + 0.0017292103730142117, + 0.004479147959500551, + 0.003051494713872671, + 0.0032580734696239233, + 0.0024986285716295242, + 0.00282592442817986, + 0.001911115599796176, + 0.003347822232171893, + 0.0016283374279737473, + 0.0019151329761371017, + 0.0025822508614510298, + 0.0017911471659317613, + 0.006717650685459375, + 0.008813825435936451, + 0.008310078643262386, + 0.01485539972782135, + 0.004900140687823296, + 0.006569668184965849, + 0.005769174080342054, + 0.005100896582007408, + 0.006830692756921053, + 0.006100812926888466, + 0.011511269025504589, + 0.0118555948138237, + 0.007313765585422516, + 0.009214566089212894, + 0.006726873107254505, + 0.008015497587621212, + 0.007938362658023834, + 0.014490929432213306, + 0.01324321236461401, + 0.01718033291399479, + 0.012264514341950417, + 0.010160659439861774, + 0.012576408684253693, + 0.009960097260773182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.027946414425969124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043092403560876846 + ], + "opposite_logprob_tokens": [ + -3.1680924892425537 + ], + "logprob_reward": -0.043092403560876846, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043092403560876846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10203739255666733, + 0.6906814575195312, + 0.029948409646749496, + 0.03163597360253334, + 0.03786446526646614, + 0.010858908295631409, + 0.01740724965929985, + 0.020357714965939522, + 0.03324293717741966, + 0.025965504348278046 + ], + "credit_map": [ + 0.024322399869561195, + 0.008609365671873093, + 0.014563239179551601, + 0.012457954697310925, + 0.006314814556390047, + 0.014607343822717667, + 0.010070154443383217, + 0.011092105880379677, + 0.16078424453735352, + 0.03651749715209007, + 0.09773281961679459, + 0.05389661714434624, + 0.056967951357364655, + 0.12825456261634827, + 0.0591169074177742, + 0.09741081297397614, + 0.008461047895252705, + 0.0022776180412620306, + 0.0050262222066521645, + 0.002249845303595066, + 0.0021260343492031097, + 0.0030988019425421953, + 0.0029005776159465313, + 0.003808261128142476, + 0.007418741937726736, + 0.0020447729621082544, + 0.004367001820355654, + 0.0026101975236088037, + 0.0023751426488161087, + 0.00573152769356966, + 0.003111782018095255, + 0.00397680327296257, + 0.004660656210035086, + 0.0017709876410663128, + 0.003542570397257805, + 0.0016505164094269276, + 0.0028520715422928333, + 0.009131315164268017, + 0.003456908743828535, + 0.01079943310469389, + 0.0016933252336457372, + 0.0014566171448677778, + 0.0011057008523494005, + 0.0012549642706289887, + 0.0009534243145026267, + 0.0015398546820506454, + 0.0014908509328961372, + 0.001364170340821147, + 0.0018781258258968592, + 0.0021201337222009897, + 0.001721467007882893, + 0.004133968148380518, + 0.0022059727925807238, + 0.0019120124634355307, + 0.0015542223118245602, + 0.0018813458736985922, + 0.002468873979523778, + 0.0023822332732379436, + 0.002749526174739003, + 0.004171077627688646, + 0.0023374895099550486, + 0.0022773337550461292, + 0.001910924562253058, + 0.0020602543372660875, + 0.003515716642141342, + 0.0034939036704599857, + 0.00570959597826004, + 0.0049720690585672855, + 0.003989091608673334, + 0.002905614674091339, + 0.004532843828201294, + 0.004124096594750881, + 0.003411826677620411, + 0.00335460202768445, + 0.003110801102593541, + 0.004907676484435797, + 0.0019345046021044254, + 0.0028711373452097178, + 0.002565149450674653, + 0.003809806192293763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.043092403560876846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032038569916039705 + ], + "opposite_logprob_tokens": [ + -5.753203868865967 + ], + "logprob_reward": -0.0032038569916039705, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032038569916039705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19739170372486115, + 0.5688618421554565, + 0.029342997819185257, + 0.029002148658037186, + 0.023749202489852905, + 0.014669031836092472, + 0.02708076313138008, + 0.00879562459886074, + 0.0483575239777565, + 0.05274910479784012 + ], + "credit_map": [ + 0.025746630504727364, + 0.009984969161450863, + 0.019879961386322975, + 0.015512332320213318, + 0.012922489084303379, + 0.047066472470760345, + 0.017685765400528908, + 0.04859308525919914, + 0.1337270885705948, + 0.029465047642588615, + 0.08063727617263794, + 0.04486343264579773, + 0.04645204171538353, + 0.10549169778823853, + 0.04881102964282036, + 0.07941427081823349, + 0.006413506809622049, + 0.0022500816266983747, + 0.0037880069576203823, + 0.002885506022721529, + 0.002261826302856207, + 0.004676597658544779, + 0.0026096522342413664, + 0.004457817878574133, + 0.004312862176448107, + 0.0018921938026323915, + 0.003306290367618203, + 0.003122753696516156, + 0.0025354470126330853, + 0.005764026660472155, + 0.0027748604770749807, + 0.005293715745210648, + 0.0028154992032796144, + 0.0028360597789287567, + 0.002672938397154212, + 0.002586331916972995, + 0.0022635450586676598, + 0.003721108427271247, + 0.003373250365257263, + 0.003480470273643732, + 0.0016839908203110099, + 0.0017485457938164473, + 0.0013546154368668795, + 0.0028744309674948454, + 0.0017671328969299793, + 0.0016761538572609425, + 0.001724445610307157, + 0.0018397157546132803, + 0.0036116233095526695, + 0.0045729223638772964, + 0.0034905648790299892, + 0.005153511185199022, + 0.002279408508911729, + 0.003401856403797865, + 0.0023881900124251842, + 0.002182686235755682, + 0.0012726340210065246, + 0.0008869192097336054, + 0.0011482423869892955, + 0.0019955302122980356, + 0.0007544921827502549, + 0.000764431431889534, + 0.000809514313004911, + 0.0011638612486422062, + 0.00506097124889493, + 0.005258976016193628, + 0.01124501507729292, + 0.006974949035793543, + 0.004470851272344589, + 0.004232087172567844, + 0.005205150227993727, + 0.005909522995352745, + 0.006620754487812519, + 0.005688686389476061, + 0.007750533055514097, + 0.009016668424010277, + 0.006022158078849316, + 0.006505787838250399, + 0.005436766427010298, + 0.00570775056257844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0032038569916039705 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991943657398224 + ], + "opposite_logprob_tokens": [ + -1.8049194812774658 + ], + "logprob_reward": -0.17991943657398224, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991943657398224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18576616048812866, + 0.31546926498413086, + 0.062306132167577744, + 0.02053903415799141, + 0.05217936262488365, + 0.016244055703282356, + 0.05826882645487785, + 0.055874697864055634, + 0.11388281732797623, + 0.11946961283683777 + ], + "credit_map": [ + 0.018393274396657944, + 0.009930064901709557, + 0.018199695274233818, + 0.014202991500496864, + 0.011736320331692696, + 0.040187012404203415, + 0.018198678269982338, + 0.05491812154650688, + 0.07261161506175995, + 0.017671767622232437, + 0.04422605782747269, + 0.026783747598528862, + 0.02570558339357376, + 0.058155469596385956, + 0.026679357513785362, + 0.04363567754626274, + 0.013226534239947796, + 0.005048870109021664, + 0.008728442713618279, + 0.005267622880637646, + 0.005227981135249138, + 0.01071744691580534, + 0.006634813733398914, + 0.007454420439898968, + 0.0021663298830389977, + 0.0023067581932991743, + 0.002200316870585084, + 0.0030145079363137484, + 0.0016106718685477972, + 0.002493213629350066, + 0.003665492869913578, + 0.003081743838265538, + 0.005418260116130114, + 0.003248963737860322, + 0.004463698249310255, + 0.0037957539316266775, + 0.003783596446737647, + 0.013492995873093605, + 0.004849896300584078, + 0.013126200996339321, + 0.0022702282294631004, + 0.0018748281290754676, + 0.0017491671023890376, + 0.0032153145875781775, + 0.0015387864550575614, + 0.00226508523337543, + 0.0014432993484660983, + 0.0018873477820307016, + 0.0091292979195714, + 0.008937547914683819, + 0.008642292581498623, + 0.00819146167486906, + 0.005117391236126423, + 0.006803898140788078, + 0.005989226512610912, + 0.005457710940390825, + 0.006444168742746115, + 0.006314761471003294, + 0.008145150728523731, + 0.013246634043753147, + 0.0055041443556547165, + 0.005730040371417999, + 0.005455402657389641, + 0.0050343964248895645, + 0.012610426172614098, + 0.011289315298199654, + 0.030928101390600204, + 0.020494570955634117, + 0.008323993533849716, + 0.009150683879852295, + 0.01078631728887558, + 0.01029941625893116, + 0.014829087071120739, + 0.020003506913781166, + 0.014042053371667862, + 0.018884586170315742, + 0.01010443177074194, + 0.01220304612070322, + 0.013366298750042915, + 0.016036605462431908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.17991943657398224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an antique camphorwood chest lock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chest wood deep reddish-brown with a curly grass carving pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lock made of brass and inlaid into the chest surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock cylinder unscrew to reveal tiny brass gears activating and meshing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the gears drive a wooden tenon to slide out of its mortise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993247509002686 + ], + "opposite_logprob_tokens": [ + -1.8049324750900269 + ], + "logprob_reward": -0.17993247509002686, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993247509002686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21537964046001434, + 0.22441436350345612, + 0.07544834166765213, + 0.0684792771935463, + 0.10892293602228165, + 0.036438748240470886, + 0.03915118798613548, + 0.09777872264385223, + 0.0561375692486763, + 0.07784918695688248 + ], + "credit_map": [ + 0.023742886260151863, + 0.010760175995528698, + 0.021523043513298035, + 0.01818646304309368, + 0.014138022437691689, + 0.04649212583899498, + 0.022653374820947647, + 0.05788355693221092, + 0.05633404850959778, + 0.01327577792108059, + 0.03398537263274193, + 0.015355014242231846, + 0.018845288082957268, + 0.04077373072504997, + 0.01839420758187771, + 0.027450915426015854, + 0.019868135452270508, + 0.004821542650461197, + 0.010794434696435928, + 0.0071724895387887955, + 0.0050850664265453815, + 0.010372268036007881, + 0.007620702963322401, + 0.009713712148368359, + 0.012344742193818092, + 0.006602640729397535, + 0.0064367689192295074, + 0.008896920830011368, + 0.0047227381728589535, + 0.01110558770596981, + 0.008324435912072659, + 0.010045450180768967, + 0.01322228368371725, + 0.0069440919905900955, + 0.009187338873744011, + 0.00708039989694953, + 0.00785593781620264, + 0.0247928649187088, + 0.010283335112035275, + 0.029556695371866226, + 0.004189017694443464, + 0.0056785354390740395, + 0.0036851102486252785, + 0.006359859369695187, + 0.002805985976010561, + 0.004482191987335682, + 0.005047825165092945, + 0.004190223757177591, + 0.005864225793629885, + 0.004686939064413309, + 0.005745626520365477, + 0.006435325834900141, + 0.004414086230099201, + 0.004372507333755493, + 0.0034561266656965017, + 0.004176354967057705, + 0.010281266644597054, + 0.013988628052175045, + 0.01296169962733984, + 0.024193057790398598, + 0.00819957535713911, + 0.009979642927646637, + 0.009918838739395142, + 0.008256015367805958, + 0.0069364202208817005, + 0.004405391868203878, + 0.014207565225660801, + 0.00880014430731535, + 0.005632967222481966, + 0.007109330967068672, + 0.004567502066493034, + 0.00447825388982892, + 0.006252521649003029, + 0.013474265113472939, + 0.00962233915925026, + 0.012461645528674126, + 0.007979659363627434, + 0.00834609940648079, + 0.010246933437883854, + 0.009465728886425495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.17993247509002686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_4/index_tensor([8608, 8608])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_4/index_tensor([8608, 8608])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f09d861995718d28e11c452b138d50b53e227f37 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_4/index_tensor([8608, 8608])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, visceral macro shot captures a moment of intense, natural drama, focusing on a tiny tree frog, its skin a vibrant, almost neon lime green, shimmering with light. The frog desperately clings to the edge of the velvety, ruffled petals of a deep purple iris, its body splayed out in a gesture of pure tension. A single, harsh light source, sharply positioned from the side, pierces the darkness, creating a dramatic black-and-white contrast. This intense light casts deep, impenetrable shadows that obscure much of the flower, yet vividly illuminates the frog's form, capturing bright highlights on its damp skin and sculpting the contours of every muscle in its leg. The camera remains steady, allowing the viewer to see the translucent, perfectly round suction pads on its toes pressing firmly against the soft, royal purple surface of the petals, revealing the immense effort exerted by its precarious grip and creating a powerful, tense atmosphere of utter focus on this primal moment of existence.", + "seed": 8385, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_4/index_8608_seed_8385_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17519344394095243, + "vlm_logprob": -0.17519344394095243, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.17519344394095243, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021913539618253708 + ], + "opposite_logprob_tokens": [ + -3.896913528442383 + ], + "logprob_reward": -0.021913539618253708, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021913539618253708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14511913061141968, + 0.13463878631591797, + 0.06720700114965439, + 0.1535973697900772, + 0.3298792541027069, + 0.0653734803199768, + 0.01706273853778839, + 0.02941974438726902, + 0.034373603761196136, + 0.02332892268896103 + ], + "credit_map": [ + 0.024318097159266472, + 0.010640540160238743, + 0.015412766486406326, + 0.015808308497071266, + 0.01708044484257698, + 0.023069603368639946, + 0.022822503000497818, + 0.015966853126883507, + 0.015267938375473022, + 0.027527667582035065, + 0.008115353062748909, + 0.017702998593449593, + 0.012992985546588898, + 0.020417844876646996, + 0.01286463811993599, + 0.01974935084581375, + 0.00871779303997755, + 0.008134458214044571, + 0.005742328241467476, + 0.006811392959207296, + 0.011620596051216125, + 0.009555085562169552, + 0.006678899750113487, + 0.009946438483893871, + 0.020809920504689217, + 0.015922391787171364, + 0.012546460144221783, + 0.015467386692762375, + 0.015041973441839218, + 0.027269642800092697, + 0.016505252569913864, + 0.030034329742193222, + 0.03846365958452225, + 0.029414553195238113, + 0.027449555695056915, + 0.049865346401929855, + 0.030395731329917908, + 0.04655173048377037, + 0.04699178412556648, + 0.06074684485793114, + 0.008845383301377296, + 0.010749293491244316, + 0.005620100535452366, + 0.007521027699112892, + 0.00579906115308404, + 0.009810256771743298, + 0.007307584397494793, + 0.009720769710838795, + 0.0022135525941848755, + 0.0029493754263967276, + 0.000933858216740191, + 0.00245360704138875, + 0.0014347892720252275, + 0.0024914955720305443, + 0.0022913969587534666, + 0.0022946626413613558, + 0.0034267441369593143, + 0.004294959362596273, + 0.001662701484747231, + 0.003277731128036976, + 0.0030824702698737383, + 0.0044115460477769375, + 0.004331229254603386, + 0.00493236118927598, + 0.0038019169587641954, + 0.004541275091469288, + 0.0017690311651676893, + 0.0049335043877363205, + 0.0036291908472776413, + 0.0050654602237045765, + 0.006048273295164108, + 0.00458495132625103, + 0.002345987595617771, + 0.003444420639425516, + 0.001434107543900609, + 0.004339377861469984, + 0.0018316636560484767, + 0.0032941712997853756, + 0.004249589052051306, + 0.0023896056227385998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.021913539618253708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031977968756109476 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.0031977968756109476, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031977968756109476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17794325947761536, + 0.1693863719701767, + 0.11016461253166199, + 0.2427143156528473, + 0.036861877888441086, + 0.06456729024648666, + 0.07571770995855331, + 0.03835233673453331, + 0.05655018985271454, + 0.027742110192775726 + ], + "credit_map": [ + 0.0302326288074255, + 0.01966436393558979, + 0.018336670473217964, + 0.020510602742433548, + 0.022595932707190514, + 0.031575001776218414, + 0.01797136850655079, + 0.017056673765182495, + 0.02272515371441841, + 0.03434513136744499, + 0.010045904666185379, + 0.019007762894034386, + 0.02014337107539177, + 0.031097155064344406, + 0.015513305552303791, + 0.01650858297944069, + 0.013726496137678623, + 0.01749185100197792, + 0.01063430868089199, + 0.01296208705753088, + 0.015579602681100368, + 0.015665428712964058, + 0.011526991613209248, + 0.012577841058373451, + 0.04080889746546745, + 0.034052636474370956, + 0.015791501849889755, + 0.029959507286548615, + 0.021479608491063118, + 0.043240420520305634, + 0.02229474112391472, + 0.0350869856774807, + 0.0041966512799263, + 0.006515344139188528, + 0.004170550499111414, + 0.00519595155492425, + 0.0032521155662834644, + 0.005141384433954954, + 0.0038195005618035793, + 0.004570373799651861, + 0.009483887813985348, + 0.00846327655017376, + 0.004630736541002989, + 0.007163615897297859, + 0.0071447971276938915, + 0.010001798160374165, + 0.006222719792276621, + 0.011456454172730446, + 0.009068787097930908, + 0.014260368421673775, + 0.004999155644327402, + 0.008481775410473347, + 0.00638818321749568, + 0.0121194738894701, + 0.00892955344170332, + 0.01147040631622076, + 0.004784169606864452, + 0.004636739380657673, + 0.002643114421516657, + 0.004772258456796408, + 0.003684020834043622, + 0.005861535202711821, + 0.005711689591407776, + 0.006258805748075247, + 0.005418017040938139, + 0.006682444829493761, + 0.0047975946217775345, + 0.011956725269556046, + 0.004117240197956562, + 0.008320060558617115, + 0.008829977363348007, + 0.006428126245737076, + 0.0036326937843114138, + 0.003862465498968959, + 0.001859606127254665, + 0.004234107211232185, + 0.0024513942189514637, + 0.0035991331096738577, + 0.004350009374320507, + 0.0037526979576796293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0031977968756109476 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578091025352478 + ], + "opposite_logprob_tokens": [ + -0.6328091025352478 + ], + "logprob_reward": -0.7578091025352478, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578091025352478, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1910327821969986, + 0.17512336373329163, + 0.1007518321275711, + 0.09168645739555359, + 0.08729174733161926, + 0.17111334204673767, + 0.054119907319545746, + 0.03206383436918259, + 0.06289259344339371, + 0.03392411768436432 + ], + "credit_map": [ + 0.021085377782583237, + 0.019057024270296097, + 0.015374270267784595, + 0.033822327852249146, + 0.02178405225276947, + 0.029245194047689438, + 0.03262021392583847, + 0.018044348806142807, + 0.023632103577256203, + 0.035812657326459885, + 0.013832087628543377, + 0.016596976667642593, + 0.017969736829400063, + 0.025104187428951263, + 0.015372673980891705, + 0.026802947744727135, + 0.014100832864642143, + 0.017554527148604393, + 0.01099257729947567, + 0.012388372793793678, + 0.007556952070444822, + 0.014985551126301289, + 0.007884270511567593, + 0.015288753435015678, + 0.014403143897652626, + 0.009192814119160175, + 0.010498122312128544, + 0.011103066615760326, + 0.007953796535730362, + 0.015288403257727623, + 0.012636622413992882, + 0.010610496625304222, + 0.011173236183822155, + 0.011785832233726978, + 0.006350967567414045, + 0.010363194160163403, + 0.007127769291400909, + 0.01790618896484375, + 0.008601120673120022, + 0.013983450829982758, + 0.02123609371483326, + 0.029780583456158638, + 0.010694241151213646, + 0.02305441163480282, + 0.012802871875464916, + 0.029305802658200264, + 0.019112473353743553, + 0.025126880034804344, + 0.00922810472548008, + 0.006170421838760376, + 0.003810200607404113, + 0.008206677623093128, + 0.004791325889527798, + 0.008029179647564888, + 0.007638081908226013, + 0.006245921831578016, + 0.004613831173628569, + 0.0035574876237660646, + 0.0024336387868970633, + 0.0050379084423184395, + 0.002354914089664817, + 0.004502574447542429, + 0.004863806068897247, + 0.004699674900621176, + 0.006123639177531004, + 0.008727002888917923, + 0.005698483902961016, + 0.013939631171524525, + 0.004223035182803869, + 0.009445530362427235, + 0.008349641226232052, + 0.006385638378560543, + 0.004212485160678625, + 0.004055471159517765, + 0.003283299272879958, + 0.006403264589607716, + 0.0033725500106811523, + 0.004717489238828421, + 0.0043398854322731495, + 0.003539676545187831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.7578091025352478 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809947893023491 + ], + "opposite_logprob_tokens": [ + -3.288099527359009 + ], + "logprob_reward": -0.03809947893023491, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809947893023491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08685621619224548, + 0.2013772577047348, + 0.0897672176361084, + 0.22388409078121185, + 0.07865554839372635, + 0.12565413117408752, + 0.08112974464893341, + 0.04671364277601242, + 0.03319133073091507, + 0.0327707938849926 + ], + "credit_map": [ + 0.014871370047330856, + 0.009353192523121834, + 0.01123279519379139, + 0.012755589559674263, + 0.009110762737691402, + 0.014466379769146442, + 0.008149988949298859, + 0.006916135549545288, + 0.02838103286921978, + 0.03276266157627106, + 0.016617603600025177, + 0.02440292201936245, + 0.02065945975482464, + 0.0277099609375, + 0.01742294430732727, + 0.03342065587639809, + 0.010757011361420155, + 0.010953041724860668, + 0.009892874397337437, + 0.011698940768837929, + 0.01332711149007082, + 0.013304013758897781, + 0.009367682039737701, + 0.010466539300978184, + 0.02887771837413311, + 0.023401381447911263, + 0.016455193981528282, + 0.022933928295969963, + 0.019715333357453346, + 0.03790952265262604, + 0.024088706821203232, + 0.05050230026245117, + 0.01191526185721159, + 0.007736860774457455, + 0.0063331774435937405, + 0.011250273324549198, + 0.007453905884176493, + 0.011538409627974033, + 0.01015322282910347, + 0.012274433858692646, + 0.014103436842560768, + 0.022140031680464745, + 0.009242135100066662, + 0.018895337358117104, + 0.010323449037969112, + 0.020199118182063103, + 0.014846418052911758, + 0.015904195606708527, + 0.009266838431358337, + 0.014850711449980736, + 0.005466668866574764, + 0.010087333619594574, + 0.0066230520606040955, + 0.012533764354884624, + 0.009707867167890072, + 0.012593495659530163, + 0.005763114430010319, + 0.007143576163798571, + 0.002662830287590623, + 0.00482947938144207, + 0.005679556634277105, + 0.006939689163118601, + 0.006258366629481316, + 0.007437027059495449, + 0.003155638463795185, + 0.003844411112368107, + 0.0022293590009212494, + 0.005284181330353022, + 0.002763311844319105, + 0.0049738395027816296, + 0.006335061974823475, + 0.004605527967214584, + 0.0028304147999733686, + 0.004144539125263691, + 0.0019207505974918604, + 0.005455045495182276, + 0.002817929955199361, + 0.0042496733367443085, + 0.006983467377722263, + 0.004368968773633242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.03809947893023491 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054947301745414734 + ], + "opposite_logprob_tokens": [ + -2.9299473762512207 + ], + "logprob_reward": -0.054947301745414734, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054947301745414734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27216994762420654, + 0.1696915328502655, + 0.07508383691310883, + 0.16192859411239624, + 0.05575850233435631, + 0.07603239268064499, + 0.05230553448200226, + 0.05312386900186539, + 0.04419085010886192, + 0.03971495479345322 + ], + "credit_map": [ + 0.04335163161158562, + 0.03314818814396858, + 0.02564840018749237, + 0.03534870222210884, + 0.03425443172454834, + 0.0389629602432251, + 0.03456598520278931, + 0.02688964642584324, + 0.020854786038398743, + 0.02376575581729412, + 0.016540778800845146, + 0.024732595309615135, + 0.015556493774056435, + 0.023426320403814316, + 0.018131664022803307, + 0.0266831386834383, + 0.010546442121267319, + 0.006436038762331009, + 0.005718511063605547, + 0.009563214145600796, + 0.006486100144684315, + 0.0142592191696167, + 0.006460930686444044, + 0.015613383613526821, + 0.020821867510676384, + 0.024395128712058067, + 0.01370986457914114, + 0.01765354909002781, + 0.014636389911174774, + 0.025067759677767754, + 0.019273407757282257, + 0.02637062780559063, + 0.007402920164167881, + 0.0069290464743971825, + 0.003925002180039883, + 0.00886299554258585, + 0.00575082004070282, + 0.008433806709945202, + 0.0072343735955655575, + 0.007219537161290646, + 0.01004472840577364, + 0.008709065616130829, + 0.006365420296788216, + 0.00933926273137331, + 0.007023012265563011, + 0.011337180621922016, + 0.010899028740823269, + 0.01231468841433525, + 0.0077036963775753975, + 0.0064062136225402355, + 0.004483391530811787, + 0.006347362417727709, + 0.005912965629249811, + 0.007749818731099367, + 0.006000075023621321, + 0.007702010218054056, + 0.0074070836417376995, + 0.007263071369379759, + 0.003966108895838261, + 0.005724356509745121, + 0.005143826827406883, + 0.007754314690828323, + 0.007964463904500008, + 0.007900643162429333, + 0.004578373860567808, + 0.004831075202673674, + 0.00321219302713871, + 0.007604843005537987, + 0.00404779938980937, + 0.0071104299277067184, + 0.00609960500150919, + 0.006706529762595892, + 0.005805488675832748, + 0.004833087790757418, + 0.00296116154640913, + 0.005036451853811741, + 0.004043871536850929, + 0.0052689663134515285, + 0.00612334581092, + 0.005642578471451998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.054947301745414734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, visceral macro shot captures a moment of intense, natural drama, focusing on a tiny tree frog, its skin a vibrant, almost neon lime green, shimmering with light. The frog desperately clings to the edge of the velvety, ruffled petals of a deep purple iris, its body splayed out in a gesture of pure tension. A single, harsh light source, sharply positioned from the side, pierces the darkness, creating a dramatic black-and-white contrast. This intense light casts deep, impenetrable shadows that obscure much of the flower, yet vividly illuminates the frog's form, capturing bright highlights on its damp skin and sculpting the contours of every muscle in its leg. The camera remains steady, allowing the viewer to see the translucent, perfectly round suction pads on its toes pressing firmly against the soft, royal purple surface of the petals, revealing the immense effort exerted by its precarious grip and creating a powerful, tense atmosphere of utter focus on this primal moment of existence.", + "seed": 8386, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_4/index_8608_seed_8386_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.31232635481283066, + "vlm_logprob": -0.31232635481283066, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.31232635481283066, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0104940515011549 + ], + "opposite_logprob_tokens": [ + -4.635494232177734 + ], + "logprob_reward": -0.0104940515011549, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0104940515011549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1038559228181839, + 0.21401190757751465, + 0.026717493310570717, + 0.042820751667022705, + 0.07743966579437256, + 0.02000448666512966, + 0.06543783843517303, + 0.05937756970524788, + 0.3588724136352539, + 0.031461942940950394 + ], + "credit_map": [ + 0.01647382229566574, + 0.028762472793459892, + 0.010136045515537262, + 0.009395494125783443, + 0.01426434051245451, + 0.006137340795248747, + 0.008070345036685467, + 0.010616065934300423, + 0.03754710033535957, + 0.04488149657845497, + 0.04096696898341179, + 0.01712886430323124, + 0.02760302647948265, + 0.01067997794598341, + 0.01467911060899496, + 0.020525390282273293, + 0.00565749779343605, + 0.0026158634573221207, + 0.007997443899512291, + 0.00257578631862998, + 0.0018564973725005984, + 0.001697661355137825, + 0.0020186451729387045, + 0.002298101084306836, + 0.009330258704721928, + 0.00428756233304739, + 0.012625669129192829, + 0.003738455008715391, + 0.0030719987116754055, + 0.002635581884533167, + 0.0033044505398720503, + 0.003826776286587119, + 0.016063926741480827, + 0.015454928390681744, + 0.00967151578515768, + 0.006983587052673101, + 0.007810261100530624, + 0.0068407608196139336, + 0.003694096114486456, + 0.010920590721070766, + 0.004282487090677023, + 0.0028196838684380054, + 0.0029217065311968327, + 0.002213614759966731, + 0.002483828691765666, + 0.0013803557958453894, + 0.0014954848447814584, + 0.0024073279928416014, + 0.011296763084828854, + 0.004877639934420586, + 0.01278151385486126, + 0.0041149454191327095, + 0.01611287146806717, + 0.0063763451762497425, + 0.003988864365965128, + 0.005888900253921747, + 0.01850666105747223, + 0.005622833967208862, + 0.007950952276587486, + 0.003598170354962349, + 0.01035283599048853, + 0.00494211632758379, + 0.0032951191533356905, + 0.005108886864036322, + 0.06650921702384949, + 0.04930543154478073, + 0.03692202642560005, + 0.04008449241518974, + 0.07464946806430817, + 0.027411000803112984, + 0.02559509687125683, + 0.03839567303657532, + 0.00450115418061614, + 0.004269024822860956, + 0.0035928674042224884, + 0.0033109423238784075, + 0.0050524091348052025, + 0.0033231701236218214, + 0.002660267986357212, + 0.0047521088272333145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0104940515011549 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821038942784071 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.002821038942784071, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821038942784071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0694895088672638, + 0.2290496975183487, + 0.01755031943321228, + 0.07708325237035751, + 0.07192438840866089, + 0.10427543520927429, + 0.1607188731431961, + 0.08230381458997726, + 0.15877753496170044, + 0.02882714755833149 + ], + "credit_map": [ + 0.01045422162860632, + 0.013754470273852348, + 0.006538351997733116, + 0.010465569794178009, + 0.007824895903468132, + 0.006312000099569559, + 0.006373523268848658, + 0.007766477297991514, + 0.04053901880979538, + 0.04796389117836952, + 0.043868985027074814, + 0.01861283928155899, + 0.029698485508561134, + 0.011542183347046375, + 0.015661731362342834, + 0.021162576973438263, + 0.0026799659244716167, + 0.002313082106411457, + 0.0020021088421344757, + 0.0018822994315996766, + 0.0017346658278256655, + 0.002035239478573203, + 0.0017628183122724295, + 0.003140140324831009, + 0.01795060932636261, + 0.006692782044410706, + 0.023861397057771683, + 0.005911046639084816, + 0.005314766429364681, + 0.005087267607450485, + 0.005732665304094553, + 0.006532724481076002, + 0.013982078991830349, + 0.014914507046341896, + 0.011451847851276398, + 0.006201229523867369, + 0.009903485886752605, + 0.0033230073750019073, + 0.0046491557732224464, + 0.00749908247962594, + 0.02598842978477478, + 0.007651414256542921, + 0.036563314497470856, + 0.007541364058852196, + 0.005021438002586365, + 0.007127954158931971, + 0.007616173941642046, + 0.0067653534933924675, + 0.039890747517347336, + 0.013743644580245018, + 0.02223266288638115, + 0.005902962293475866, + 0.04514741152524948, + 0.012238829396665096, + 0.00970527995377779, + 0.011857339181005955, + 0.025445297360420227, + 0.007187048904597759, + 0.0066404882818460464, + 0.0039488691836595535, + 0.019450126215815544, + 0.00668776361271739, + 0.0049097477458417416, + 0.0080344770103693, + 0.0250730998814106, + 0.020572563633322716, + 0.017145799472928047, + 0.016698701307177544, + 0.030903473496437073, + 0.016649171710014343, + 0.012046556919813156, + 0.019688177853822708, + 0.004522726405411959, + 0.004529197700321674, + 0.004081743769347668, + 0.003267659805715084, + 0.002555069513618946, + 0.003648458980023861, + 0.00250655016861856, + 0.003715743077918887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.002821038942784071 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020774602890015 + ], + "opposite_logprob_tokens": [ + -0.25207746028900146 + ], + "logprob_reward": -1.5020774602890015, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020774602890015, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10065868496894836, + 0.12528571486473083, + 0.14862670004367828, + 0.07711850851774216, + 0.10266567021608353, + 0.09674543887376785, + 0.1754082441329956, + 0.0694361999630928, + 0.07241731137037277, + 0.031637560576200485 + ], + "credit_map": [ + 0.014941318891942501, + 0.021413514390587807, + 0.010243874974548817, + 0.011751317419111729, + 0.012828617356717587, + 0.008025683462619781, + 0.008814767934381962, + 0.012639575637876987, + 0.022109219804406166, + 0.02480711042881012, + 0.023356277495622635, + 0.010434544645249844, + 0.01646733097732067, + 0.007579789496958256, + 0.008613199926912785, + 0.011918231844902039, + 0.031676437705755234, + 0.016274873167276382, + 0.041299764066934586, + 0.012092450633645058, + 0.008394903503358364, + 0.009409663267433643, + 0.010762598365545273, + 0.0187159962952137, + 0.0151395658031106, + 0.010330675169825554, + 0.016448885202407837, + 0.006272075232118368, + 0.008234845474362373, + 0.005789119750261307, + 0.006143803708255291, + 0.008759535849094391, + 0.02005956880748272, + 0.01971162110567093, + 0.015712827444076538, + 0.009334208443760872, + 0.014621489681303501, + 0.006074123550206423, + 0.007055193185806274, + 0.010096633806824684, + 0.016757335513830185, + 0.013396856375038624, + 0.021954096853733063, + 0.009497454389929771, + 0.010073901154100895, + 0.009090185165405273, + 0.007069029845297337, + 0.008906571194529533, + 0.05322684347629547, + 0.01416768692433834, + 0.04609988257288933, + 0.012521672062575817, + 0.010827465914189816, + 0.011769769713282585, + 0.011394491419196129, + 0.015400410629808903, + 0.012617095373570919, + 0.007820582948625088, + 0.017131932079792023, + 0.0074323066510260105, + 0.007720470894128084, + 0.005680546630173922, + 0.004141753539443016, + 0.006891503464430571, + 0.01294668111950159, + 0.01262918021529913, + 0.006939773913472891, + 0.010138659738004208, + 0.013063272461295128, + 0.004062679596245289, + 0.005241570062935352, + 0.007395491003990173, + 0.004051991272717714, + 0.0044716051779687405, + 0.004347802139818668, + 0.004557088483124971, + 0.004062083084136248, + 0.002595424186438322, + 0.003385795745998621, + 0.004165768623352051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -1.5020774602890015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337035171687603 + ], + "opposite_logprob_tokens": [ + -3.408703565597534 + ], + "logprob_reward": -0.0337035171687603, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337035171687603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0637316033244133, + 0.27460455894470215, + 0.04907185584306717, + 0.032945044338703156, + 0.2057299017906189, + 0.053178660571575165, + 0.06696133315563202, + 0.0903373509645462, + 0.12595155835151672, + 0.03748824819922447 + ], + "credit_map": [ + 0.009182271547615528, + 0.010964339599013329, + 0.008416940458118916, + 0.0076387678273022175, + 0.0059970379807055, + 0.00554985087364912, + 0.005719692446291447, + 0.010262698866426945, + 0.04863882437348366, + 0.05641864612698555, + 0.0532490499317646, + 0.022601323202252388, + 0.03583243489265442, + 0.014122437685728073, + 0.018780319020152092, + 0.02496151626110077, + 0.009582561440765858, + 0.007426070049405098, + 0.012116361409425735, + 0.003580696415156126, + 0.003949351608753204, + 0.003266924759373069, + 0.0036965885665267706, + 0.005453295074403286, + 0.006660250946879387, + 0.003575787879526615, + 0.009653543122112751, + 0.003404197981581092, + 0.0021860203705728054, + 0.0021379778627306223, + 0.002486562356352806, + 0.0028407021891325712, + 0.0381297804415226, + 0.046284399926662445, + 0.02291368506848812, + 0.017029736191034317, + 0.030597208067774773, + 0.010483705438673496, + 0.013735021464526653, + 0.026556346565485, + 0.00915815494954586, + 0.011226444505155087, + 0.005764367058873177, + 0.005319037940353155, + 0.007924518547952175, + 0.0036281105130910873, + 0.004031691234558821, + 0.006126332096755505, + 0.0187334306538105, + 0.0070152366533875465, + 0.011785528622567654, + 0.005231650546193123, + 0.008637137711048126, + 0.0046629090793430805, + 0.003763561137020588, + 0.007131869439035654, + 0.02804575301706791, + 0.008420087397098541, + 0.01143499556928873, + 0.00487832585349679, + 0.01703176088631153, + 0.007694349158555269, + 0.004978204146027565, + 0.007853859104216099, + 0.02333519048988819, + 0.019218286499381065, + 0.018914436921477318, + 0.014616012573242188, + 0.019205091521143913, + 0.00802083220332861, + 0.008809643797576427, + 0.013832040131092072, + 0.005194894038140774, + 0.006752245128154755, + 0.0044820369221270084, + 0.0036606842186301947, + 0.006098983809351921, + 0.003720656968653202, + 0.0023103568237274885, + 0.005268387030810118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.0337035171687603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0896059200167656, + 0.028552483767271042, + 0.15040217339992523, + 0.0949394553899765, + 0.15364527702331543, + 0.07024655491113663, + 0.2401154488325119, + 0.08392080664634705, + 0.0652577355504036, + 0.02331414632499218 + ], + "credit_map": [ + 0.012713390402495861, + 0.016554398462176323, + 0.010273887775838375, + 0.012148063629865646, + 0.01080471370369196, + 0.0067209345288574696, + 0.006238103844225407, + 0.014152430929243565, + 0.004347487352788448, + 0.004883093759417534, + 0.0030748541466891766, + 0.0037925848737359047, + 0.003778452053666115, + 0.003225581720471382, + 0.002415104303508997, + 0.0030353248585015535, + 0.036526158452034, + 0.010978063568472862, + 0.050190746784210205, + 0.012581312097609043, + 0.008294397033751011, + 0.009038432501256466, + 0.010763642378151417, + 0.012029408477246761, + 0.022653568536043167, + 0.009082947857677937, + 0.0269200187176466, + 0.007796597667038441, + 0.00703822635114193, + 0.006451967637985945, + 0.007052962202578783, + 0.007943161763250828, + 0.028537319973111153, + 0.03200529143214226, + 0.02066959999501705, + 0.014798435382544994, + 0.020437756553292274, + 0.008173133246600628, + 0.0109139084815979, + 0.018109846860170364, + 0.013958454132080078, + 0.008360542356967926, + 0.02076430432498455, + 0.005752844270318747, + 0.006316223181784153, + 0.005169154144823551, + 0.005050045438110828, + 0.004874990321695805, + 0.04204807057976723, + 0.01817486435174942, + 0.0510641410946846, + 0.009972070343792439, + 0.066458560526371, + 0.018787646666169167, + 0.015539614483714104, + 0.018070483580231667, + 0.020829925313591957, + 0.0087016262114048, + 0.010545902885496616, + 0.006412640679627657, + 0.01531887985765934, + 0.007416453678160906, + 0.005640162620693445, + 0.009055214002728462, + 0.010024902410805225, + 0.007495438680052757, + 0.009219188243150711, + 0.005941908340901136, + 0.006785534787923098, + 0.009632832370698452, + 0.006454228423535824, + 0.009703702293336391, + 0.0027768320869654417, + 0.00391630781814456, + 0.0027176961302757263, + 0.0034575886093080044, + 0.003156316466629505, + 0.0025633638724684715, + 0.0017564046429470181, + 0.0029696375131607056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.012535706162452698 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, visceral macro shot captures a moment of intense, natural drama, focusing on a tiny tree frog, its skin a vibrant, almost neon lime green, shimmering with light. The frog desperately clings to the edge of the velvety, ruffled petals of a deep purple iris, its body splayed out in a gesture of pure tension. A single, harsh light source, sharply positioned from the side, pierces the darkness, creating a dramatic black-and-white contrast. This intense light casts deep, impenetrable shadows that obscure much of the flower, yet vividly illuminates the frog's form, capturing bright highlights on its damp skin and sculpting the contours of every muscle in its leg. The camera remains steady, allowing the viewer to see the translucent, perfectly round suction pads on its toes pressing firmly against the soft, royal purple surface of the petals, revealing the immense effort exerted by its precarious grip and creating a powerful, tense atmosphere of utter focus on this primal moment of existence.", + "seed": 8387, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_4/index_8608_seed_8387_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20591931217350065, + "vlm_logprob": -0.20591931217350065, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.20591931217350065, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006987303029745817 + ], + "opposite_logprob_tokens": [ + -5.13198709487915 + ], + "logprob_reward": -0.006987303029745817, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006987303029745817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0725584328174591, + 0.026263779029250145, + 0.3824354112148285, + 0.012440750375390053, + 0.021685665473341942, + 0.025227004662156105, + 0.06759737432003021, + 0.36511486768722534, + 0.018076462671160698, + 0.008600256405770779 + ], + "credit_map": [ + 0.006405092775821686, + 0.006545016076415777, + 0.005748268682509661, + 0.005644236225634813, + 0.007559713441878557, + 0.006829662714153528, + 0.019753582775592804, + 0.014072860591113567, + 0.0027769068256020546, + 0.003432824043557048, + 0.0033539992291480303, + 0.0035874920431524515, + 0.0031813473906368017, + 0.0029717173893004656, + 0.004188765771687031, + 0.0027707249391824007, + 0.022324154153466225, + 0.03721827268600464, + 0.14030936360359192, + 0.019843559712171555, + 0.02405478246510029, + 0.019116707146167755, + 0.0314955972135067, + 0.0880729928612709, + 0.0013706244062632322, + 0.0014845769619569182, + 0.002029130468145013, + 0.001115284627303481, + 0.001643204945139587, + 0.002171141793951392, + 0.000967823201790452, + 0.001658964087255299, + 0.0027153256814926863, + 0.0015181234339252114, + 0.005509181879460812, + 0.002079147147014737, + 0.002364253392443061, + 0.0022709877230226994, + 0.0020340136252343655, + 0.003194633638486266, + 0.0026143596041947603, + 0.0014342564390972257, + 0.004443636164069176, + 0.0019056133460253477, + 0.002683656057342887, + 0.0038995055947452784, + 0.0019967060070484877, + 0.006249270401895046, + 0.009700262919068336, + 0.00717925513163209, + 0.007682916708290577, + 0.009528647176921368, + 0.00846879743039608, + 0.008330607786774635, + 0.007046119775623083, + 0.00966076459735632, + 0.05527951940894127, + 0.01816767267882824, + 0.019436003640294075, + 0.023311685770750046, + 0.028498182073235512, + 0.09680977463722229, + 0.01741361990571022, + 0.10619840770959854, + 0.0030390352476388216, + 0.001808509579859674, + 0.0015870528295636177, + 0.0019047728274017572, + 0.002769333077594638, + 0.0024052876979112625, + 0.0016804590122774243, + 0.0028820137958973646, + 0.0016239701071754098, + 0.0005608256324194372, + 0.0007401502807624638, + 0.0006789885810576379, + 0.0007604663260281086, + 0.0018524836050346494, + 0.0005474290228448808, + 0.0018359425012022257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.006987303029745817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024888513144105673 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.0024888513144105673, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024888513144105673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044473741203546524, + 0.07840791344642639, + 0.3804102838039398, + 0.043841615319252014, + 0.07092128694057465, + 0.05810640752315521, + 0.051063861697912216, + 0.23412738740444183, + 0.030829574912786484, + 0.007817978970706463 + ], + "credit_map": [ + 0.005233331583440304, + 0.005061591975390911, + 0.004897927865386009, + 0.004918170161545277, + 0.0050737932324409485, + 0.006252875551581383, + 0.008541883900761604, + 0.004494168795645237, + 0.005857695825397968, + 0.0073340293020009995, + 0.006071232724934816, + 0.006846277043223381, + 0.00737443333491683, + 0.007289823144674301, + 0.023933175951242447, + 0.013701239600777626, + 0.021898414939641953, + 0.03638729825615883, + 0.1395079344511032, + 0.019825318828225136, + 0.023806773126125336, + 0.019172191619873047, + 0.031316883862018585, + 0.08849547058343887, + 0.006911610718816519, + 0.003711224999278784, + 0.008785195648670197, + 0.003620498813688755, + 0.005267972592264414, + 0.0043398612178862095, + 0.0033424647990614176, + 0.007862784899771214, + 0.004914658144116402, + 0.004747448954731226, + 0.0240319836884737, + 0.0062881819903850555, + 0.0073828441090881824, + 0.0077424789778888226, + 0.005102349910885096, + 0.010711340233683586, + 0.004371423739939928, + 0.003460260573774576, + 0.019673146307468414, + 0.004358339589089155, + 0.005051314830780029, + 0.005307793151587248, + 0.002962652826681733, + 0.012921473942697048, + 0.008958223275840282, + 0.003627733327448368, + 0.005067738238722086, + 0.004687461536377668, + 0.005166579037904739, + 0.012503281235694885, + 0.003329542465507984, + 0.00772330304607749, + 0.03719481825828552, + 0.010022146627306938, + 0.013764589093625546, + 0.012129629962146282, + 0.016781296581029892, + 0.06333489716053009, + 0.011164093390107155, + 0.06973589211702347, + 0.005895218811929226, + 0.0024433021899312735, + 0.0020327104721218348, + 0.0027231997810304165, + 0.0035753236152231693, + 0.006549721583724022, + 0.0024108272045850754, + 0.0051992712542414665, + 0.0014522873098030686, + 0.0006914085824973881, + 0.0007107162964530289, + 0.0007672807550989091, + 0.0009312827023677528, + 0.0015473962994292378, + 0.0005479937535710633, + 0.0011696130968630314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.0024888513144105673 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742662906646729 + ], + "opposite_logprob_tokens": [ + -0.47426626086235046 + ], + "logprob_reward": -0.9742662906646729, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742662906646729, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10654446482658386, + 0.15258574485778809, + 0.27499067783355713, + 0.08632268756628036, + 0.05830681696534157, + 0.0852585956454277, + 0.05685478448867798, + 0.11030346900224686, + 0.046060141175985336, + 0.02277262881398201 + ], + "credit_map": [ + 0.01080319657921791, + 0.013042625039815903, + 0.013859400525689125, + 0.012866882607340813, + 0.012519893236458302, + 0.013103792443871498, + 0.01782853715121746, + 0.012520139105618, + 0.012744266539812088, + 0.010266692377626896, + 0.018333226442337036, + 0.009085194207727909, + 0.014968743547797203, + 0.014295001514256, + 0.04918640851974487, + 0.02370622381567955, + 0.014869599603116512, + 0.02414052188396454, + 0.07678250968456268, + 0.017403366044163704, + 0.020799744874238968, + 0.022287730127573013, + 0.022459257394075394, + 0.07624795287847519, + 0.011963654309511185, + 0.007890455424785614, + 0.017283964902162552, + 0.008449065499007702, + 0.009883699007332325, + 0.010228877887129784, + 0.007034348789602518, + 0.013588628731667995, + 0.0076855383813381195, + 0.0029839444905519485, + 0.018633512780070305, + 0.003801305778324604, + 0.004170455504208803, + 0.00927635096013546, + 0.0028179383371025324, + 0.008937770500779152, + 0.008165594190359116, + 0.004642253741621971, + 0.030572062358260155, + 0.0050667948089540005, + 0.006100266706198454, + 0.013779062777757645, + 0.0032519125379621983, + 0.013680648058652878, + 0.006332906428724527, + 0.0045917402021586895, + 0.005860573146492243, + 0.007252070587128401, + 0.0064203268848359585, + 0.00879721436649561, + 0.00502203730866313, + 0.012577910907566547, + 0.018125848844647408, + 0.0042923204600811005, + 0.006697085220366716, + 0.0049020820297300816, + 0.006886818911880255, + 0.030473293736577034, + 0.005440723616629839, + 0.033485300838947296, + 0.00916352029889822, + 0.0028638618532568216, + 0.003841020632535219, + 0.003890259889885783, + 0.0045795259065926075, + 0.00913252029567957, + 0.003299887990579009, + 0.009289541281759739, + 0.004039144609123468, + 0.0015822987770661712, + 0.0016117882914841175, + 0.0023822628427296877, + 0.0026310200337320566, + 0.004453541710972786, + 0.0017004994442686439, + 0.004372072871774435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.9742662906646729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606273651123047 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.01606273651123047, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606273651123047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07256266474723816, + 0.06976639479398727, + 0.10763079673051834, + 0.06334350258111954, + 0.0733436793088913, + 0.06663380563259125, + 0.04037344828248024, + 0.4499531388282776, + 0.03443284332752228, + 0.021959764882922173 + ], + "credit_map": [ + 0.006489602383226156, + 0.008036143146455288, + 0.008644863963127136, + 0.010387771762907505, + 0.008559049107134342, + 0.007380763068795204, + 0.011587424203753471, + 0.011477045714855194, + 0.006498245522379875, + 0.011992626823484898, + 0.007670280523598194, + 0.009392024949193, + 0.00972159206867218, + 0.007267896085977554, + 0.009530766867101192, + 0.007692970335483551, + 0.008895971812307835, + 0.010199714452028275, + 0.030320584774017334, + 0.007793344557285309, + 0.009458603337407112, + 0.010536978021264076, + 0.014693781733512878, + 0.01573181338608265, + 0.009221713058650494, + 0.005488591734319925, + 0.00898969080299139, + 0.004726626444607973, + 0.00742448540404439, + 0.0132087217643857, + 0.004527705255895853, + 0.009755970910191536, + 0.009708736091852188, + 0.005846024490892887, + 0.01499781385064125, + 0.00857990887016058, + 0.00955363642424345, + 0.006744069512933493, + 0.0065214005298912525, + 0.011392084881663322, + 0.004408176522701979, + 0.003555189585313201, + 0.02592780999839306, + 0.0038449519779533148, + 0.004623969551175833, + 0.0046713268384337425, + 0.0031013728585094213, + 0.01650100387632847, + 0.006709306035190821, + 0.00256660720333457, + 0.0028658583760261536, + 0.004004918038845062, + 0.003653298132121563, + 0.007945048622786999, + 0.003340034745633602, + 0.009288374334573746, + 0.06829967349767685, + 0.022322004660964012, + 0.02356037124991417, + 0.030175834894180298, + 0.035209331661462784, + 0.11829060316085815, + 0.021335281431674957, + 0.13076001405715942, + 0.0041230209171772, + 0.004075791221112013, + 0.0037624556571245193, + 0.0039842016994953156, + 0.006363654974848032, + 0.003971479367464781, + 0.0034154702443629503, + 0.004736769013106823, + 0.0022563557140529156, + 0.0024037351831793785, + 0.002373694907873869, + 0.002623512875288725, + 0.0032150137703865767, + 0.003080422757193446, + 0.002079626079648733, + 0.003927402663975954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.01606273651123047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979137934744358 + ], + "opposite_logprob_tokens": [ + -3.5297913551330566 + ], + "logprob_reward": -0.02979137934744358, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979137934744358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040733639150857925, + 0.041596513241529465, + 0.5832434296607971, + 0.06547952443361282, + 0.04273400828242302, + 0.03513538837432861, + 0.08382584899663925, + 0.08975494652986526, + 0.009323867037892342, + 0.008172799833118916 + ], + "credit_map": [ + 0.0042354026809334755, + 0.004244567360728979, + 0.004424328915774822, + 0.006932498887181282, + 0.005340388044714928, + 0.004825897514820099, + 0.0052208504639565945, + 0.005509708542376757, + 0.002911621704697609, + 0.0036905715242028236, + 0.003885576268658042, + 0.002868383890017867, + 0.004564885515719652, + 0.002954889554530382, + 0.013991501182317734, + 0.006729090586304665, + 0.033951014280319214, + 0.05593044310808182, + 0.21479983627796173, + 0.030656980350613594, + 0.036496952176094055, + 0.02916671335697174, + 0.04542320966720581, + 0.13681833446025848, + 0.009610689245164394, + 0.005391639657318592, + 0.012488335371017456, + 0.005022452212870121, + 0.007434955798089504, + 0.013058367185294628, + 0.004912653472274542, + 0.007560440804809332, + 0.006478347349911928, + 0.002488139783963561, + 0.009103184565901756, + 0.0033235501032322645, + 0.00349214393645525, + 0.007636401802301407, + 0.0030639476608484983, + 0.007148291449993849, + 0.004723978694528341, + 0.0026653213426470757, + 0.008751762099564075, + 0.00345766544342041, + 0.0037804811727255583, + 0.005514305084943771, + 0.0021978849545121193, + 0.004043992143124342, + 0.015307177789509296, + 0.0037559985648840666, + 0.005375706125050783, + 0.005260370206087828, + 0.006244958378374577, + 0.0215936079621315, + 0.004800681956112385, + 0.02148735709488392, + 0.01459286455065012, + 0.0036855959333479404, + 0.004987558349967003, + 0.004435008391737938, + 0.006120199337601662, + 0.024705924093723297, + 0.004308789502829313, + 0.026919016614556313, + 0.0015123351477086544, + 0.0009360593976452947, + 0.0010284394957125187, + 0.001192515599541366, + 0.0015292082680389285, + 0.0011046449653804302, + 0.0008443518308922648, + 0.0011763129150494933, + 0.0016984624089673162, + 0.0007239899132400751, + 0.0008908558520488441, + 0.0008076901431195438, + 0.000920427730306983, + 0.0012422712752595544, + 0.0006862579612061381, + 0.0012028448982164264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.02979137934744358 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, visceral macro shot captures a moment of intense, natural drama, focusing on a tiny tree frog, its skin a vibrant, almost neon lime green, shimmering with light. The frog desperately clings to the edge of the velvety, ruffled petals of a deep purple iris, its body splayed out in a gesture of pure tension. A single, harsh light source, sharply positioned from the side, pierces the darkness, creating a dramatic black-and-white contrast. This intense light casts deep, impenetrable shadows that obscure much of the flower, yet vividly illuminates the frog's form, capturing bright highlights on its damp skin and sculpting the contours of every muscle in its leg. The camera remains steady, allowing the viewer to see the translucent, perfectly round suction pads on its toes pressing firmly against the soft, royal purple surface of the petals, revealing the immense effort exerted by its precarious grip and creating a powerful, tense atmosphere of utter focus on this primal moment of existence.", + "seed": 8388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_4/index_8608_seed_8388_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3363261729944497, + "vlm_logprob": -0.3363261729944497, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.3363261729944497, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008668058551847935 + ], + "opposite_logprob_tokens": [ + -4.883667945861816 + ], + "logprob_reward": -0.008668058551847935, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008668058551847935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19363023340702057, + 0.132612943649292, + 0.040289636701345444, + 0.07317794859409332, + 0.08163806051015854, + 0.07115468382835388, + 0.0859038233757019, + 0.12198466062545776, + 0.12950856983661652, + 0.07009942084550858 + ], + "credit_map": [ + 0.018720846623182297, + 0.02862878143787384, + 0.015100612305104733, + 0.029655158519744873, + 0.032000582665205, + 0.02005479671061039, + 0.027141839265823364, + 0.02232762984931469, + 0.01048373430967331, + 0.021609924733638763, + 0.01469307392835617, + 0.017532672733068466, + 0.028294840827584267, + 0.014442052692174911, + 0.01372575294226408, + 0.011830884963274002, + 0.0034867606591433287, + 0.0036964258179068565, + 0.0038926484994590282, + 0.009723499417304993, + 0.0046669114381074905, + 0.005560888908803463, + 0.004439770709723234, + 0.00482272868975997, + 0.00895762164145708, + 0.013735052198171616, + 0.007848169654607773, + 0.011331255547702312, + 0.00815905537456274, + 0.007750908378511667, + 0.005769755225628614, + 0.009626133367419243, + 0.009964662604033947, + 0.011347781866788864, + 0.00688373064622283, + 0.008385805413126945, + 0.014907613396644592, + 0.011428735218942165, + 0.008740410208702087, + 0.00997932069003582, + 0.005400355905294418, + 0.011417187750339508, + 0.010619224049150944, + 0.009234570898115635, + 0.01178920827805996, + 0.007851445116102695, + 0.0070651862770318985, + 0.00777750089764595, + 0.010318220593035221, + 0.014313512481749058, + 0.0069982693530619144, + 0.011747677810490131, + 0.0158611461520195, + 0.010329590179026127, + 0.008088507689535618, + 0.008246899582445621, + 0.014298021793365479, + 0.02335984818637371, + 0.010838357731699944, + 0.013860462233424187, + 0.023238342255353928, + 0.016325868666172028, + 0.00875752791762352, + 0.011306229047477245, + 0.011803578585386276, + 0.01584223285317421, + 0.015339560806751251, + 0.022554297000169754, + 0.02025672234594822, + 0.01135565247386694, + 0.0101711954921484, + 0.022185327485203743, + 0.007098382338881493, + 0.014348512515425682, + 0.009128917008638382, + 0.005968835204839706, + 0.010748364962637424, + 0.007559242658317089, + 0.004855762235820293, + 0.010391401126980782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.008668058551847935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.438193678855896, + 0.0925481840968132, + 0.06288734823465347, + 0.04667442664504051, + 0.06291797012090683, + 0.044509969651699066, + 0.029851699247956276, + 0.05741642415523529, + 0.12767663598060608, + 0.03732364624738693 + ], + "credit_map": [ + 0.04012247547507286, + 0.050537362694740295, + 0.03122618980705738, + 0.08398623764514923, + 0.08154231309890747, + 0.04595812410116196, + 0.04367320239543915, + 0.061147768050432205, + 0.00973406434059143, + 0.015762435272336006, + 0.009519252926111221, + 0.012697079218924046, + 0.015585548244416714, + 0.01296387892216444, + 0.00756181450560689, + 0.008724106475710869, + 0.006914246361702681, + 0.009292743168771267, + 0.004523183684796095, + 0.007064206525683403, + 0.008873763494193554, + 0.009884640574455261, + 0.005595459137111902, + 0.010739109478890896, + 0.00452263280749321, + 0.007091364823281765, + 0.0037886304780840874, + 0.004899462219327688, + 0.006175355985760689, + 0.006301948335021734, + 0.007772891316562891, + 0.006122138351202011, + 0.006281994748860598, + 0.006050911732017994, + 0.0040614791214466095, + 0.010298903100192547, + 0.007537703961133957, + 0.00697296392172575, + 0.006561357993632555, + 0.015152658335864544, + 0.004782339558005333, + 0.004148997366428375, + 0.0033069157507270575, + 0.005443936679512262, + 0.0043623680248856544, + 0.00545905064791441, + 0.007580440025776625, + 0.009425920434296131, + 0.0037485782522708178, + 0.0023724285420030355, + 0.002930032555013895, + 0.0039569842629134655, + 0.0034131810534745455, + 0.0038078846409916878, + 0.0035245432518422604, + 0.006098067853599787, + 0.006546201650053263, + 0.008711257949471474, + 0.004998919554054737, + 0.007889633998274803, + 0.00826864130795002, + 0.007752906531095505, + 0.007486523129045963, + 0.005762340966612101, + 0.012679378502070904, + 0.014491780661046505, + 0.014131579548120499, + 0.016928646713495255, + 0.02173352986574173, + 0.015140896663069725, + 0.014054491184651852, + 0.018516337499022484, + 0.00312623823992908, + 0.008799592033028603, + 0.0030439060647040606, + 0.0033294223248958588, + 0.0053468928672373295, + 0.0038108297158032656, + 0.004057345446199179, + 0.005809419322758913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0031928059179335833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.600538730621338 + ], + "opposite_logprob_tokens": [ + -0.2255387008190155 + ], + "logprob_reward": -1.600538730621338, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.600538730621338, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.43926283717155457, + 0.2051744908094406, + 0.042323313653469086, + 0.029772911220788956, + 0.04018038511276245, + 0.03240533918142319, + 0.06486596167087555, + 0.02551911398768425, + 0.0680752545595169, + 0.0524204820394516 + ], + "credit_map": [ + 0.04160007834434509, + 0.0552084743976593, + 0.03343408927321434, + 0.07480651140213013, + 0.08082135766744614, + 0.044143784791231155, + 0.047174811363220215, + 0.06207366660237312, + 0.020312897861003876, + 0.03489522635936737, + 0.025426940992474556, + 0.02877720259130001, + 0.03325025364756584, + 0.02759983018040657, + 0.017436349764466286, + 0.017475759610533714, + 0.0039038064423948526, + 0.006224375683814287, + 0.003534463932737708, + 0.006001398898661137, + 0.0043105315417051315, + 0.006749999709427357, + 0.004778457339853048, + 0.006820272654294968, + 0.0033813323825597763, + 0.005032856483012438, + 0.0023973274510353804, + 0.0032223162706941366, + 0.003993502352386713, + 0.004220233298838139, + 0.0030812211334705353, + 0.00444411626085639, + 0.0034276852384209633, + 0.0043803672306239605, + 0.0029270800296217203, + 0.005207621026784182, + 0.006252704653888941, + 0.004920292180031538, + 0.004482637625187635, + 0.008581993170082569, + 0.003470713971182704, + 0.004533429630100727, + 0.003624311415478587, + 0.0036761853843927383, + 0.005325690843164921, + 0.003849982051178813, + 0.0033000109251588583, + 0.0046250117011368275, + 0.006040511187165976, + 0.006620702333748341, + 0.006692150142043829, + 0.009962299838662148, + 0.009112764149904251, + 0.007499240804463625, + 0.007657445035874844, + 0.011280843056738377, + 0.003542599966749549, + 0.002557710511609912, + 0.0027045896276831627, + 0.0034913034178316593, + 0.0026591254863888025, + 0.004001057241111994, + 0.003196798963472247, + 0.0033659273758530617, + 0.005998380482196808, + 0.00780446408316493, + 0.008145194500684738, + 0.010215526446700096, + 0.01337804552167654, + 0.006426895037293434, + 0.0057410672307014465, + 0.010365678928792477, + 0.004499154165387154, + 0.0059612770564854145, + 0.004787395242601633, + 0.004993255715817213, + 0.009592052549123764, + 0.008676659315824509, + 0.006980511825531721, + 0.006930171977728605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -1.600538730621338 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36436209082603455, + 0.21275241672992706, + 0.1621774584054947, + 0.030403247103095055, + 0.03640766441822052, + 0.06419068574905396, + 0.042381107807159424, + 0.03370954096317291, + 0.02790159545838833, + 0.025714218616485596 + ], + "credit_map": [ + 0.034246135503053665, + 0.04430945962667465, + 0.027988798916339874, + 0.06384603679180145, + 0.06943819671869278, + 0.03898349031805992, + 0.0391305610537529, + 0.04641939327120781, + 0.01997814141213894, + 0.0396515317261219, + 0.022517317906022072, + 0.02539927326142788, + 0.039290353655815125, + 0.02830374427139759, + 0.018634755164384842, + 0.01897730864584446, + 0.017612963914871216, + 0.01942916214466095, + 0.010435868054628372, + 0.019016755744814873, + 0.02572062611579895, + 0.028851211071014404, + 0.017997941002249718, + 0.02311292663216591, + 0.0032040083315223455, + 0.0035983375273644924, + 0.0036791800521314144, + 0.003900504671037197, + 0.003509243717417121, + 0.0030555773992091417, + 0.003686940297484398, + 0.0057694558054208755, + 0.0037405062466859818, + 0.004929392132908106, + 0.003180788829922676, + 0.0037929308600723743, + 0.006508339662104845, + 0.004805808886885643, + 0.004785855300724506, + 0.004664043430238962, + 0.006467090919613838, + 0.005475770682096481, + 0.008800234645605087, + 0.01045741606503725, + 0.00648020813241601, + 0.0074156951159238815, + 0.007638872601091862, + 0.011455399915575981, + 0.004700974095612764, + 0.00536790955811739, + 0.00496738962829113, + 0.007558027282357216, + 0.0065170698799192905, + 0.004555427934974432, + 0.0038585944566875696, + 0.004855713807046413, + 0.003823783714324236, + 0.0039360481314361095, + 0.003959116991609335, + 0.005220650229603052, + 0.005394636653363705, + 0.0037171323783695698, + 0.0031777217518538237, + 0.004480453208088875, + 0.002734180772677064, + 0.0034936992451548576, + 0.0025408794172108173, + 0.0031304480507969856, + 0.004175609443336725, + 0.0038684546016156673, + 0.0038173652719706297, + 0.00414095725864172, + 0.002358337864279747, + 0.0044042677618563175, + 0.0026559948455542326, + 0.002093304181471467, + 0.004699452314525843, + 0.0033191542606800795, + 0.0023312282282859087, + 0.003852480323985219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02058546058833599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny tree frog with vibrant neon lime green skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the frog clinging to the edge of a deep purple iris flower with velvety, ruffled petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single harsh light source from the side creating high contrast and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the translucent, perfectly round suction pads on the frog's toes visible pressing against the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and steady throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048645809292793274 + ], + "opposite_logprob_tokens": [ + -3.0486457347869873 + ], + "logprob_reward": -0.048645809292793274, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048645809292793274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3144838213920593, + 0.26139378547668457, + 0.04721330478787422, + 0.04837552085518837, + 0.07530083507299423, + 0.040327288210392, + 0.017134368419647217, + 0.01883815973997116, + 0.12531131505966187, + 0.05162163823843002 + ], + "credit_map": [ + 0.02850269339978695, + 0.03397253900766373, + 0.02246086858212948, + 0.06577738374471664, + 0.05965982750058174, + 0.032863128930330276, + 0.029760198667645454, + 0.04148717597126961, + 0.02744167111814022, + 0.0450664721429348, + 0.02410910837352276, + 0.032636940479278564, + 0.04490271210670471, + 0.04025379568338394, + 0.022984016686677933, + 0.02399904653429985, + 0.005404460243880749, + 0.005667253863066435, + 0.004107342101633549, + 0.007224109023809433, + 0.0061993105337023735, + 0.008266295306384563, + 0.00408921018242836, + 0.006255322135984898, + 0.0058780210092663765, + 0.009655289351940155, + 0.006195774767547846, + 0.005975035019218922, + 0.0065260897390544415, + 0.004563030321151018, + 0.004956233315169811, + 0.004626041743904352, + 0.008728171698749065, + 0.012975283898413181, + 0.006418908946216106, + 0.007917830720543861, + 0.015034199692308903, + 0.009997259825468063, + 0.007321170065551996, + 0.006908010691404343, + 0.003324374556541443, + 0.006950194947421551, + 0.004609629511833191, + 0.004005667753517628, + 0.008022548630833626, + 0.004951226990669966, + 0.004385818261653185, + 0.004077824763953686, + 0.0018861659336835146, + 0.0022691928315907717, + 0.0019090381683781743, + 0.0029459816869348288, + 0.0019536640029400587, + 0.002071187598630786, + 0.00190552300773561, + 0.002193615073338151, + 0.0020377414766699076, + 0.003277574200183153, + 0.0015952493995428085, + 0.0024989224039018154, + 0.0029093357734382153, + 0.002493815263733268, + 0.00208740821108222, + 0.0019381134770810604, + 0.011945897713303566, + 0.013355469331145287, + 0.016319554299116135, + 0.019069071859121323, + 0.020311769098043442, + 0.01288188062608242, + 0.012792490422725677, + 0.018635189160704613, + 0.003929955419152975, + 0.01197419036179781, + 0.004257595632225275, + 0.004682411439716816, + 0.0077633727341890335, + 0.005456277169287205, + 0.0063680242747068405, + 0.007189810276031494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.048645809292793274 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_5/index_tensor([4083, 4083])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_5/index_tensor([4083, 4083])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f34cf7def1181e61087acd65723342d6074b87a0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_5/index_tensor([4083, 4083])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The shot begins with a macro, slightly high-angle shot, the camera slowly and steadily moving forward toward a tiny garden snail resting on a broad, vibrant green leaf. The snail remains perfectly curled and motionless, its translucent amber shell a perfect logarithmic spiral, its delicate concentric growth lines striking. The entire scene is studded with shimmering dewdrops, clinging to the waxy surface of the leaf and the snail's shell like tiny glass spheres. As the camera draws closer, the depth of field becomes incredibly shallow, the lush, out-of-focus garden background transforming into a soft, verdant blur as the shot transitions to extreme close-up. This deliberate movement reveals intricate and shimmering details: a dewdrop magnifying the texture of the leaf's underside, the delicate pearlescent sheen of the shell's surface, and the moist, delicate texture of the snail's visible flesh. The scene is bathed in a soft, golden morning light, creating brilliant highlights on the water droplets and casting gentle shadows that accentuate every curve, creating an atmosphere of tranquility, meditation, and wonder.", + "seed": 16328, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_5/index_4083_seed_16328_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4144616080913693, + "vlm_logprob": -0.4144616080913693, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.4144616080913693, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002865734277293086 + ], + "opposite_logprob_tokens": [ + -6.252865791320801 + ], + "logprob_reward": -0.002865734277293086, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.002865734277293086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0554920919239521, + 0.014926472678780556, + 0.019105631858110428, + 0.17536862194538116, + 0.03165678679943085, + 0.03681597113609314, + 0.5936707854270935, + 0.036931026726961136, + 0.02233271673321724, + 0.013699914328753948 + ], + "credit_map": [ + 0.018373815342783928, + 0.0029096559155732393, + 0.019283803179860115, + 0.0028967619873583317, + 0.00449182465672493, + 0.0021566255018115044, + 0.0024549078661948442, + 0.0029246974736452103, + 0.0019685551524162292, + 0.0023738532327115536, + 0.0015131620457395911, + 0.0019669218454509974, + 0.001548559172078967, + 0.0016277359100058675, + 0.002442206721752882, + 0.0014854779001325369, + 0.0024194824509322643, + 0.0032148631289601326, + 0.002062795916572213, + 0.0030956461559981108, + 0.0021078914869576693, + 0.0016721204156056046, + 0.0031303479336202145, + 0.0014024837873876095, + 0.023577725514769554, + 0.010622000321745872, + 0.011024130508303642, + 0.005769926123321056, + 0.07097723335027695, + 0.018965575844049454, + 0.007747518830001354, + 0.026684507727622986, + 0.0032412721775472164, + 0.0044060139916837215, + 0.004030510783195496, + 0.004189760889858007, + 0.0033824373967945576, + 0.004637429490685463, + 0.004678994417190552, + 0.003090368350967765, + 0.0027845478616654873, + 0.0016678160754963756, + 0.0027025272138416767, + 0.0014440505765378475, + 0.0024440404959023, + 0.01104595884680748, + 0.0019342359155416489, + 0.012792792171239853, + 0.02969457022845745, + 0.012692490592598915, + 0.016480062156915665, + 0.0179995596408844, + 0.050595685839653015, + 0.25484296679496765, + 0.01930897869169712, + 0.19205646216869354, + 0.0047900378704071045, + 0.0056535047478973866, + 0.0028937708120793104, + 0.002206864533945918, + 0.004079869482666254, + 0.0050223288126289845, + 0.004385917913168669, + 0.007898732088506222, + 0.002457271097227931, + 0.003741478081792593, + 0.0018716068007051945, + 0.0016282116994261742, + 0.004110082983970642, + 0.0028249905444681644, + 0.0030728706624358892, + 0.00262620416469872, + 0.002066792920231819, + 0.0024926119949668646, + 0.0012784096179530025, + 0.0010971984593197703, + 0.0015651810681447387, + 0.0013830442912876606, + 0.0016582555836066604, + 0.0021584213245660067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.002865734277293086 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.017791271209717 + ], + "opposite_logprob_tokens": [ + -0.1427912563085556 + ], + "logprob_reward": -2.017791271209717, + "yes_no_margin_reward": -1.875, + "token_credit_reward": -2.017791271209717, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08014838397502899, + 0.010367093607783318, + 0.008044851943850517, + 0.10492684692144394, + 0.02277476154267788, + 0.31692367792129517, + 0.34253084659576416, + 0.03867385908961296, + 0.057045262306928635, + 0.01856444776058197 + ], + "credit_map": [ + 0.02797027677297592, + 0.004094521980732679, + 0.029818430542945862, + 0.003465954214334488, + 0.005273092072457075, + 0.0020946133881807327, + 0.004343267995864153, + 0.00308822700753808, + 0.0013832160038873553, + 0.0011169963981956244, + 0.0012332299957051873, + 0.0014397596241906285, + 0.00104062061291188, + 0.0008687139488756657, + 0.002194029977545142, + 0.001090527162887156, + 0.0014361909125000238, + 0.0009475685656070709, + 0.0010698331752792, + 0.0006854384555481374, + 0.0007430406403727829, + 0.000882549851667136, + 0.0010930299758911133, + 0.001187200890854001, + 0.013264518231153488, + 0.007531380746513605, + 0.00643679965287447, + 0.003927564714103937, + 0.042032912373542786, + 0.015266812406480312, + 0.0049712155014276505, + 0.011495646089315414, + 0.0030621332116425037, + 0.0034366052132099867, + 0.0028245635330677032, + 0.0019707311876118183, + 0.0034664731938391924, + 0.002846132731065154, + 0.002484189346432686, + 0.0026839340571314096, + 0.0198589526116848, + 0.008462890051305294, + 0.012763069942593575, + 0.00848296470940113, + 0.02528562769293785, + 0.12619370222091675, + 0.011857953853905201, + 0.10401850193738937, + 0.017949581146240234, + 0.007219055201858282, + 0.009154344908893108, + 0.01061081513762474, + 0.02839123271405697, + 0.14569693803787231, + 0.011219785548746586, + 0.11228911578655243, + 0.00963180884718895, + 0.006249542348086834, + 0.003603661898523569, + 0.002234874526038766, + 0.0036032653879374266, + 0.004292109049856663, + 0.00400572270154953, + 0.005052875727415085, + 0.006128197535872459, + 0.008280486799776554, + 0.005655137822031975, + 0.004678133875131607, + 0.007312039379030466, + 0.00861888937652111, + 0.008489714935421944, + 0.007882662117481232, + 0.0016062069917097688, + 0.0019913932774215937, + 0.0018556483555585146, + 0.0017312300624325871, + 0.0033433774951845407, + 0.003177089150995016, + 0.0021773867774754763, + 0.0026821158826351166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -2.017791271209717 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020611969754099846 + ], + "opposite_logprob_tokens": [ + -3.8956120014190674 + ], + "logprob_reward": -0.020611969754099846, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020611969754099846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08144883066415787, + 0.008583741262555122, + 0.00840580090880394, + 0.09201322495937347, + 0.02774176560342312, + 0.469038188457489, + 0.19853535294532776, + 0.028123540803790092, + 0.07270089536905289, + 0.013408646918833256 + ], + "credit_map": [ + 0.029959753155708313, + 0.0035397913306951523, + 0.03223272040486336, + 0.0036995268892496824, + 0.00474398210644722, + 0.001837494783103466, + 0.003167333547025919, + 0.002268234733492136, + 0.0014018189394846559, + 0.0010804053163155913, + 0.0009134917636401951, + 0.0009212155709974468, + 0.0010521162766963243, + 0.000746158417314291, + 0.0014255918795242906, + 0.0010429427493363619, + 0.0009599204640835524, + 0.0010497651528567076, + 0.0009870320791378617, + 0.0012357665691524744, + 0.0008995249518193305, + 0.0007967974524945021, + 0.001243162085302174, + 0.0012338324449956417, + 0.010180113837122917, + 0.006942669861018658, + 0.005244262516498566, + 0.003638820955529809, + 0.03643980249762535, + 0.018480289727449417, + 0.004072937648743391, + 0.00701433140784502, + 0.0035351125989109278, + 0.0033746433909982443, + 0.003758667502552271, + 0.0027681903447955847, + 0.0040185945108532906, + 0.003454843768849969, + 0.003276107832789421, + 0.003555605886504054, + 0.02927391603589058, + 0.01144664641469717, + 0.019555076956748962, + 0.012253707274794579, + 0.03678879141807556, + 0.18659530580043793, + 0.017428521066904068, + 0.15569622814655304, + 0.010944614186882973, + 0.005176028236746788, + 0.005704008974134922, + 0.006247362587600946, + 0.014353979378938675, + 0.0819939374923706, + 0.007267037872225046, + 0.06684837490320206, + 0.005307610146701336, + 0.004726596642285585, + 0.0028667973820120096, + 0.0015946851344779134, + 0.0031949984841048717, + 0.0033069979399442673, + 0.004813340958207846, + 0.0023125126026570797, + 0.008994380943477154, + 0.010198486968874931, + 0.006840389687567949, + 0.006611032877117395, + 0.008706109598279, + 0.010036847554147243, + 0.01182206254452467, + 0.009491587989032269, + 0.0009949596133083105, + 0.0018055374966934323, + 0.0013971971347928047, + 0.0008018655353225768, + 0.0022164061665534973, + 0.002208178164437413, + 0.0015834725927561522, + 0.0024010303895920515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.020611969754099846 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026394031941890717 + ], + "opposite_logprob_tokens": [ + -3.6513941287994385 + ], + "logprob_reward": -0.026394031941890717, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026394031941890717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17883175611495972, + 0.012347388081252575, + 0.009017799980938435, + 0.10374995321035385, + 0.012287064455449581, + 0.1971711814403534, + 0.4348643124103546, + 0.028302757069468498, + 0.014476410113275051, + 0.00895135011523962 + ], + "credit_map": [ + 0.06670334935188293, + 0.0073675233870744705, + 0.07086843252182007, + 0.007663572672754526, + 0.010995257645845413, + 0.00486138928681612, + 0.0053120022639632225, + 0.005060241091996431, + 0.0016373153775930405, + 0.0016069018747657537, + 0.0018909646896645427, + 0.0014879251830279827, + 0.0012625533854588866, + 0.0015135694993659854, + 0.0014446614077314734, + 0.0015034977113828063, + 0.0011997459223493934, + 0.0013913451693952084, + 0.0009755129576660693, + 0.0014615465188398957, + 0.0009654280729591846, + 0.0008245323551818728, + 0.0014742565108463168, + 0.0007254341617226601, + 0.011753196828067303, + 0.0062813526019454, + 0.005607263650745153, + 0.0028904902283102274, + 0.0429641455411911, + 0.01900402083992958, + 0.003996421117335558, + 0.011253065429627895, + 0.0015987674705684185, + 0.0014780660858377814, + 0.0011605722829699516, + 0.0012027787743136287, + 0.0012119936291128397, + 0.0015791519545018673, + 0.001972491154447198, + 0.002083244500681758, + 0.012811458669602871, + 0.004912384320050478, + 0.008925913833081722, + 0.0053722006268799305, + 0.01443123072385788, + 0.0749356746673584, + 0.008034516125917435, + 0.06774779409170151, + 0.022483140230178833, + 0.009432334452867508, + 0.011800657957792282, + 0.013257352635264397, + 0.036750428378582, + 0.18583659827709198, + 0.014101083390414715, + 0.14120276272296906, + 0.0051597594283521175, + 0.003414618317037821, + 0.0024669310078024864, + 0.001949014258570969, + 0.0029227519407868385, + 0.0033726731780916452, + 0.0031427594367414713, + 0.00587425148114562, + 0.0018398088868707418, + 0.001674704602919519, + 0.0016383545007556677, + 0.0007646761368960142, + 0.0016736950492486358, + 0.0025873938575387, + 0.0024089436046779156, + 0.0018888351041823626, + 0.0009544608765281737, + 0.001362021081149578, + 0.0010868144454434514, + 0.0009718853980302811, + 0.000809077697340399, + 0.0008741962374188006, + 0.001409422722645104, + 0.0014834717148914933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.026394031941890717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645033273845911 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004645033273845911, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645033273845911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10094678401947021, + 0.008451750501990318, + 0.00915126409381628, + 0.0540877990424633, + 0.019734974950551987, + 0.5437231659889221, + 0.11411601305007935, + 0.03632219135761261, + 0.0941997617483139, + 0.019266344606876373 + ], + "credit_map": [ + 0.03656097874045372, + 0.004729459993541241, + 0.039790328592061996, + 0.004466274753212929, + 0.0061957654543221, + 0.002784748561680317, + 0.0035149208270013332, + 0.002904301043599844, + 0.001171124167740345, + 0.0010510804131627083, + 0.0013804757036268711, + 0.0009839459089562297, + 0.0009068563231267035, + 0.0007292599184438586, + 0.0012908254284411669, + 0.000938182813115418, + 0.0010614312486723065, + 0.0009371355408802629, + 0.001419874606654048, + 0.0009534070268273354, + 0.0008665670757181942, + 0.0010869071120396256, + 0.0010699672857299447, + 0.001755974255502224, + 0.0054001291282474995, + 0.005303253885358572, + 0.003036217764019966, + 0.002588985487818718, + 0.020026827231049538, + 0.012114115990698338, + 0.0026745134964585304, + 0.0029437586199492216, + 0.002033962868154049, + 0.002909110626205802, + 0.002406749641522765, + 0.001985033741220832, + 0.0025687485467642546, + 0.002193523570895195, + 0.0025070852134376764, + 0.0031307612080127, + 0.03386026993393898, + 0.013217924162745476, + 0.022067788988351822, + 0.014133322983980179, + 0.04284844547510147, + 0.2167506366968155, + 0.020001374185085297, + 0.18084336817264557, + 0.006634409539401531, + 0.003558454802259803, + 0.0038487103302031755, + 0.0038302913308143616, + 0.007662431802600622, + 0.045160870999097824, + 0.004508587997406721, + 0.03891225531697273, + 0.00696598133072257, + 0.006878043059259653, + 0.0028207043651491404, + 0.002856417093425989, + 0.004880667198449373, + 0.004013379104435444, + 0.004074721597135067, + 0.0038322790060192347, + 0.01091621071100235, + 0.01227752584964037, + 0.008780336007475853, + 0.007966656237840652, + 0.012419537641108036, + 0.01633160188794136, + 0.013107823207974434, + 0.012400074861943722, + 0.0013019222533330321, + 0.0036423662677407265, + 0.001581539399921894, + 0.0010435056174173951, + 0.003663984825834632, + 0.003617726732045412, + 0.0018985082861036062, + 0.002516791457310319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.004645033273845911 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a macro, slightly high-angle shot, the camera slowly and steadily moving forward toward a tiny garden snail resting on a broad, vibrant green leaf. The snail remains perfectly curled and motionless, its translucent amber shell a perfect logarithmic spiral, its delicate concentric growth lines striking. The entire scene is studded with shimmering dewdrops, clinging to the waxy surface of the leaf and the snail's shell like tiny glass spheres. As the camera draws closer, the depth of field becomes incredibly shallow, the lush, out-of-focus garden background transforming into a soft, verdant blur as the shot transitions to extreme close-up. This deliberate movement reveals intricate and shimmering details: a dewdrop magnifying the texture of the leaf's underside, the delicate pearlescent sheen of the shell's surface, and the moist, delicate texture of the snail's visible flesh. The scene is bathed in a soft, golden morning light, creating brilliant highlights on the water droplets and casting gentle shadows that accentuate every curve, creating an atmosphere of tranquility, meditation, and wonder.", + "seed": 16329, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_5/index_4083_seed_16329_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24042596947401762, + "vlm_logprob": -0.24042596947401762, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.24042596947401762, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002902701497077942 + ], + "opposite_logprob_tokens": [ + -6.377902507781982 + ], + "logprob_reward": -0.002902701497077942, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.002902701497077942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14424657821655273, + 0.06364251673221588, + 0.10172656178474426, + 0.19329528510570526, + 0.07491409778594971, + 0.054902635514736176, + 0.10728809237480164, + 0.09678690135478973, + 0.11668703705072403, + 0.04651029035449028 + ], + "credit_map": [ + 0.017285959795117378, + 0.02773629128932953, + 0.013341863639652729, + 0.017649216577410698, + 0.017446832731366158, + 0.022929983213543892, + 0.01298719272017479, + 0.014869242906570435, + 0.006950900424271822, + 0.017188800498843193, + 0.004544725641608238, + 0.008324085734784603, + 0.007634720765054226, + 0.006482319440692663, + 0.007388427387923002, + 0.005128542892634869, + 0.010047866962850094, + 0.016729576513171196, + 0.010464346967637539, + 0.01611003838479519, + 0.01159137673676014, + 0.011291859671473503, + 0.014026488177478313, + 0.011465009301900864, + 0.00759483827278018, + 0.01643727347254753, + 0.005683063995093107, + 0.010296536609530449, + 0.07120727747678757, + 0.06665924936532974, + 0.008357184939086437, + 0.007059869356453419, + 0.01183466799557209, + 0.011998550966382027, + 0.005505157168954611, + 0.008095884695649147, + 0.009753908030688763, + 0.011822104454040527, + 0.009298212826251984, + 0.006605610251426697, + 0.006799846421927214, + 0.007300988771021366, + 0.004957057535648346, + 0.005753143224865198, + 0.006632624659687281, + 0.007354926783591509, + 0.00950572732836008, + 0.006598322186619043, + 0.010592821054160595, + 0.020857887342572212, + 0.011650355532765388, + 0.013373909518122673, + 0.014984876848757267, + 0.014278240501880646, + 0.011892780661582947, + 0.009657224640250206, + 0.007183133624494076, + 0.015212718397378922, + 0.008043341338634491, + 0.012910529039800167, + 0.015054005198180676, + 0.013593696989119053, + 0.016259774565696716, + 0.008529698476195335, + 0.008741882629692554, + 0.019295532256364822, + 0.008984480053186417, + 0.015214295126497746, + 0.019097154960036278, + 0.01543024368584156, + 0.012597043067216873, + 0.01732640154659748, + 0.006635576486587524, + 0.005619407165795565, + 0.006909257732331753, + 0.004039095249027014, + 0.006159538868814707, + 0.005223928485065699, + 0.0059827533550560474, + 0.005940731149166822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.002902701497077942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370954513549805 + ], + "opposite_logprob_tokens": [ + -0.3870954215526581 + ], + "logprob_reward": -1.1370954513549805, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370954513549805, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11594107747077942, + 0.05144425854086876, + 0.10474862158298492, + 0.09490867704153061, + 0.1038857251405716, + 0.11326637864112854, + 0.14768582582473755, + 0.11783942580223083, + 0.11433086544275284, + 0.03594912961125374 + ], + "credit_map": [ + 0.008757664822041988, + 0.025243960320949554, + 0.015369147062301636, + 0.020523633807897568, + 0.013962049968540668, + 0.009856032207608223, + 0.014428554102778435, + 0.00780003285035491, + 0.0057969484478235245, + 0.012063493952155113, + 0.004504343029111624, + 0.00721612898632884, + 0.006282081361860037, + 0.0055456762202084064, + 0.005724357441067696, + 0.004311230964958668, + 0.009961714036762714, + 0.02166866511106491, + 0.015888722613453865, + 0.013061678037047386, + 0.013124493882060051, + 0.011637028306722641, + 0.01153540425002575, + 0.00787092000246048, + 0.008470729924738407, + 0.022200018167495728, + 0.009324626065790653, + 0.012754549272358418, + 0.01335417851805687, + 0.011704260483384132, + 0.008843670599162579, + 0.008256655186414719, + 0.010122349485754967, + 0.023343142122030258, + 0.008259596303105354, + 0.014626067131757736, + 0.01534467563033104, + 0.013231664896011353, + 0.010201233439147472, + 0.008756999857723713, + 0.01611997000873089, + 0.027324091643095016, + 0.008085138164460659, + 0.013417739421129227, + 0.014714076183736324, + 0.01272821705788374, + 0.010379756800830364, + 0.010497396811842918, + 0.013434317894279957, + 0.031424157321453094, + 0.016229946166276932, + 0.02009100280702114, + 0.01966191455721855, + 0.01945587433874607, + 0.015362851321697235, + 0.012025769799947739, + 0.011021643877029419, + 0.0184734258800745, + 0.006430602166801691, + 0.01746281608939171, + 0.025619909167289734, + 0.01308431290090084, + 0.014091202057898045, + 0.011655529960989952, + 0.012758154422044754, + 0.010869392193853855, + 0.020269135013222694, + 0.011084964498877525, + 0.01897418685257435, + 0.01340037863701582, + 0.013019213452935219, + 0.013955457136034966, + 0.0053868466056883335, + 0.0028509392868727446, + 0.006733792368322611, + 0.003597952425479889, + 0.004622770939022303, + 0.003831038484349847, + 0.004362988285720348, + 0.004562803078442812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -1.1370954513549805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371112421154976 + ], + "opposite_logprob_tokens": [ + -3.4087111949920654 + ], + "logprob_reward": -0.03371112421154976, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371112421154976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08329341560602188, + 0.039267513900995255, + 0.13572260737419128, + 0.07794561982154846, + 0.07619179040193558, + 0.10916256159543991, + 0.18559624254703522, + 0.0855654701590538, + 0.16569387912750244, + 0.04156094044446945 + ], + "credit_map": [ + 0.009065517224371433, + 0.014661399647593498, + 0.012686314061284065, + 0.008560187183320522, + 0.012377890758216381, + 0.009118471294641495, + 0.007585145998746157, + 0.009238488972187042, + 0.004439798183739185, + 0.006185923703014851, + 0.006613649427890778, + 0.004062072839587927, + 0.005121550522744656, + 0.0038870738353580236, + 0.004738780204206705, + 0.004218664951622486, + 0.014226051047444344, + 0.024473749101161957, + 0.021060511469841003, + 0.016501685604453087, + 0.015670171007514, + 0.013810424134135246, + 0.018577180802822113, + 0.01140284352004528, + 0.0068258261308074, + 0.014706596732139587, + 0.00471115205436945, + 0.006989621557295322, + 0.01757863350212574, + 0.014018191955983639, + 0.008166968822479248, + 0.004948626738041639, + 0.01058138720691204, + 0.00820811279118061, + 0.006855749990791082, + 0.007377292029559612, + 0.016439398750662804, + 0.013662891462445259, + 0.007880575023591518, + 0.005186382681131363, + 0.022675208747386932, + 0.014697865583002567, + 0.01096332911401987, + 0.009833437390625477, + 0.014550875872373581, + 0.012330691330134869, + 0.011087399907410145, + 0.013023754581809044, + 0.0219611506909132, + 0.01910502463579178, + 0.024671748280525208, + 0.030221370980143547, + 0.01793537847697735, + 0.02431214600801468, + 0.028178082779049873, + 0.019211331382393837, + 0.006821867544203997, + 0.01760527491569519, + 0.006996224634349346, + 0.011526945047080517, + 0.015145388431847095, + 0.010437550023198128, + 0.00966640841215849, + 0.007365813013166189, + 0.016175875440239906, + 0.018037443980574608, + 0.030918171629309654, + 0.02032488025724888, + 0.01954917423427105, + 0.022349854931235313, + 0.02020207606256008, + 0.018136408179998398, + 0.00616456987336278, + 0.003988981246948242, + 0.006985144689679146, + 0.003766077570617199, + 0.004524227697402239, + 0.005147672723978758, + 0.005926437675952911, + 0.005057833157479763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.03371112421154976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014217865653336048 + ], + "opposite_logprob_tokens": [ + -4.264217853546143 + ], + "logprob_reward": -0.014217865653336048, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014217865653336048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24237848818302155, + 0.06842003762722015, + 0.0434342660009861, + 0.10588522255420685, + 0.09221450984477997, + 0.0641920268535614, + 0.046613238751888275, + 0.1202196404337883, + 0.14250212907791138, + 0.07414039969444275 + ], + "credit_map": [ + 0.026999931782484055, + 0.04618888348340988, + 0.025160102173686028, + 0.028166355565190315, + 0.03098262846469879, + 0.03501475229859352, + 0.023528318852186203, + 0.026337523013353348, + 0.008194653317332268, + 0.016346074640750885, + 0.0063172983936965466, + 0.0073254527524113655, + 0.007808197755366564, + 0.006910786963999271, + 0.008563430979847908, + 0.006954141892492771, + 0.004048244096338749, + 0.007973495870828629, + 0.004682141821831465, + 0.004878542851656675, + 0.005105838645249605, + 0.005235855467617512, + 0.00669034942984581, + 0.004819794557988644, + 0.004637517500668764, + 0.012308642268180847, + 0.005399308633059263, + 0.007048016879707575, + 0.03168007731437683, + 0.0331995002925396, + 0.0067151011899113655, + 0.004897058475762606, + 0.014306957833468914, + 0.01316510047763586, + 0.0066091143526136875, + 0.008821170777082443, + 0.016816699877381325, + 0.015641897916793823, + 0.008380541577935219, + 0.008473023772239685, + 0.009373203851282597, + 0.011509185656905174, + 0.004167964216321707, + 0.008892304264008999, + 0.008446941152215004, + 0.007396358996629715, + 0.008615727536380291, + 0.005790339782834053, + 0.006084450986236334, + 0.004845055285841227, + 0.004385493230074644, + 0.006550677586346865, + 0.0051813931204378605, + 0.005737014580518007, + 0.007376438472419977, + 0.00645271735265851, + 0.011079118587076664, + 0.013677471317350864, + 0.017167888581752777, + 0.012520859017968178, + 0.012484312988817692, + 0.01370447687804699, + 0.023926587775349617, + 0.015658918768167496, + 0.014458643272519112, + 0.014865797013044357, + 0.026998264715075493, + 0.016961432993412018, + 0.01632659323513508, + 0.018298352137207985, + 0.017220165580511093, + 0.017372870817780495, + 0.010381883010268211, + 0.005968767683953047, + 0.014304942451417446, + 0.006781924981623888, + 0.007774305995553732, + 0.007757342886179686, + 0.010138573125004768, + 0.011032660491764545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.014217865653336048 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202704653143883 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202704653143883, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202704653143883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08879625797271729, + 0.04233109951019287, + 0.15089914202690125, + 0.10449942201375961, + 0.11311084777116776, + 0.054139889776706696, + 0.1561424732208252, + 0.09378483146429062, + 0.15190623700618744, + 0.04438981041312218 + ], + "credit_map": [ + 0.012180261313915253, + 0.013131950050592422, + 0.010330972261726856, + 0.008782118558883667, + 0.012781285680830479, + 0.01155229564756155, + 0.010647767223417759, + 0.009389603510499, + 0.005152401514351368, + 0.008319593966007233, + 0.005293803755193949, + 0.004414500202983618, + 0.005438510328531265, + 0.004117006901651621, + 0.005048210266977549, + 0.0045470744371414185, + 0.015690205618739128, + 0.02741466648876667, + 0.02285126782953739, + 0.01847946085035801, + 0.017469806596636772, + 0.01692577451467514, + 0.01868211291730404, + 0.013385837897658348, + 0.008410219103097916, + 0.02622179314494133, + 0.006691796705126762, + 0.010770928114652634, + 0.01993592455983162, + 0.01632777228951454, + 0.009699429385364056, + 0.0064415596425533295, + 0.02056705392897129, + 0.014009655453264713, + 0.01246807910501957, + 0.012446942739188671, + 0.013883227482438087, + 0.011835928075015545, + 0.015118265524506569, + 0.01278169360011816, + 0.012237219139933586, + 0.005895028822124004, + 0.006079256068915129, + 0.0049600847996771336, + 0.006889253854751587, + 0.006602978799492121, + 0.004971647169440985, + 0.006504416931420565, + 0.016718929633498192, + 0.015303236432373524, + 0.02274847775697708, + 0.02404944784939289, + 0.014509258791804314, + 0.021341398358345032, + 0.023493187502026558, + 0.017978528514504433, + 0.007545931730419397, + 0.018788019195199013, + 0.008292945101857185, + 0.010778295807540417, + 0.01778188720345497, + 0.011907821521162987, + 0.011151758953928947, + 0.007538171485066414, + 0.014950914308428764, + 0.014047758653759956, + 0.029429741203784943, + 0.018656298518180847, + 0.0164321381598711, + 0.020596060901880264, + 0.020135371014475822, + 0.01765795797109604, + 0.006450892426073551, + 0.003408183343708515, + 0.008835102431476116, + 0.004097247030586004, + 0.00368337519466877, + 0.0055734701454639435, + 0.006514637731015682, + 0.0058269016444683075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.014202704653143883 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a macro, slightly high-angle shot, the camera slowly and steadily moving forward toward a tiny garden snail resting on a broad, vibrant green leaf. The snail remains perfectly curled and motionless, its translucent amber shell a perfect logarithmic spiral, its delicate concentric growth lines striking. The entire scene is studded with shimmering dewdrops, clinging to the waxy surface of the leaf and the snail's shell like tiny glass spheres. As the camera draws closer, the depth of field becomes incredibly shallow, the lush, out-of-focus garden background transforming into a soft, verdant blur as the shot transitions to extreme close-up. This deliberate movement reveals intricate and shimmering details: a dewdrop magnifying the texture of the leaf's underside, the delicate pearlescent sheen of the shell's surface, and the moist, delicate texture of the snail's visible flesh. The scene is bathed in a soft, golden morning light, creating brilliant highlights on the water droplets and casting gentle shadows that accentuate every curve, creating an atmosphere of tranquility, meditation, and wonder.", + "seed": 16330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_5/index_4083_seed_16330_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3296336279250681, + "vlm_logprob": -0.3296336279250681, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.3296336279250681, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021058782003819942 + ], + "opposite_logprob_tokens": [ + -6.752105712890625 + ], + "logprob_reward": -0.0021058782003819942, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0021058782003819942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06652645021677017, + 0.06610798835754395, + 0.11271427571773529, + 0.08231344074010849, + 0.33692172169685364, + 0.09206881374120712, + 0.05988704040646553, + 0.047174662351608276, + 0.07745695114135742, + 0.058828629553318024 + ], + "credit_map": [ + 0.006991302128881216, + 0.005841233767569065, + 0.019275842234492302, + 0.008624163456261158, + 0.0047106617130339146, + 0.006266765296459198, + 0.007536076474934816, + 0.0072804163210093975, + 0.005177007056772709, + 0.012149467132985592, + 0.009959776885807514, + 0.008162199519574642, + 0.008943351916968822, + 0.004812874365597963, + 0.007382956333458424, + 0.009520361199975014, + 0.010554338805377483, + 0.01650979556143284, + 0.01479591429233551, + 0.010918696410953999, + 0.01510616671293974, + 0.01792825013399124, + 0.012405486777424812, + 0.014495638199150562, + 0.006081804167479277, + 0.007386863697320223, + 0.01259563211351633, + 0.010209457948803902, + 0.007316472474485636, + 0.010585094802081585, + 0.009936108253896236, + 0.01820201240479946, + 0.03867020457983017, + 0.03808354213833809, + 0.04746567830443382, + 0.06278516352176666, + 0.0398075170814991, + 0.04110899940133095, + 0.03630300238728523, + 0.0326976478099823, + 0.006673680618405342, + 0.014627955853939056, + 0.013393771834671497, + 0.01878567598760128, + 0.012769551947712898, + 0.008070377632975578, + 0.009030540473759174, + 0.008717264048755169, + 0.008824115619063377, + 0.005388275254517794, + 0.014450499787926674, + 0.007535713724792004, + 0.007826928980648518, + 0.006264191120862961, + 0.00532376067712903, + 0.004273564554750919, + 0.007905276492238045, + 0.0051183318719267845, + 0.009074687026441097, + 0.005546041298657656, + 0.005574143026024103, + 0.006770831998437643, + 0.003733473364263773, + 0.0034518809989094734, + 0.012492281384766102, + 0.009046711027622223, + 0.017697498202323914, + 0.007444602437317371, + 0.008361350744962692, + 0.01266961544752121, + 0.004413584712892771, + 0.00533131742849946, + 0.009213185869157314, + 0.005652392748743296, + 0.014739444479346275, + 0.010498341172933578, + 0.0047059631906449795, + 0.007130414247512817, + 0.003678187495097518, + 0.0032107033766806126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.0021058782003819942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005723476409912 + ], + "opposite_logprob_tokens": [ + -0.2255723774433136 + ], + "logprob_reward": -1.6005723476409912, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005723476409912, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03036840818822384, + 0.04192344471812248, + 0.08383850753307343, + 0.021249430254101753, + 0.3059825003147125, + 0.17847700417041779, + 0.05400796979665756, + 0.04723113775253296, + 0.19773374497890472, + 0.039187848567962646 + ], + "credit_map": [ + 0.00411996990442276, + 0.002941165119409561, + 0.005479027517139912, + 0.003237626049667597, + 0.0033333832398056984, + 0.00280928798019886, + 0.005011905916035175, + 0.0034360450226813555, + 0.0030105195473879576, + 0.00638015940785408, + 0.006998264696449041, + 0.005388839170336723, + 0.007507021073251963, + 0.0023802253417670727, + 0.006381318438798189, + 0.0038771003019064665, + 0.007864421233534813, + 0.01366643887013197, + 0.011503884568810463, + 0.009559620171785355, + 0.009441295638680458, + 0.01374055165797472, + 0.008339772000908852, + 0.00972252432256937, + 0.0019025596557185054, + 0.001990705030038953, + 0.0025177362840622663, + 0.003948702476918697, + 0.003613533917814493, + 0.0025826990604400635, + 0.0025811067316681147, + 0.002112386515364051, + 0.03810783103108406, + 0.03159351274371147, + 0.043533872812986374, + 0.055623993277549744, + 0.03902600705623627, + 0.03720607981085777, + 0.03339523822069168, + 0.027495985850691795, + 0.01935553178191185, + 0.022217893972992897, + 0.02947716973721981, + 0.023505965247750282, + 0.02478640526533127, + 0.024783285334706306, + 0.013648261316120625, + 0.020702479407191277, + 0.01245358306914568, + 0.00592182157561183, + 0.01329747587442398, + 0.005331226158887148, + 0.00608368543908, + 0.003582586068660021, + 0.0032591901253908873, + 0.0040783993899822235, + 0.007487432099878788, + 0.00488676130771637, + 0.009592260234057903, + 0.005889929831027985, + 0.005514463409781456, + 0.004898366052657366, + 0.0034419302828609943, + 0.0055199977941811085, + 0.027953529730439186, + 0.03135467320680618, + 0.031472835689783096, + 0.03757812827825546, + 0.019088320434093475, + 0.02511330507695675, + 0.01439796481281519, + 0.010774980299174786, + 0.0047652143985033035, + 0.005722127389162779, + 0.005865548271685839, + 0.0046419864520430565, + 0.005527404602617025, + 0.0053975824266672134, + 0.0029922653920948505, + 0.004275716375559568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -1.6005723476409912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023306358605623245 + ], + "opposite_logprob_tokens": [ + -3.773306369781494 + ], + "logprob_reward": -0.023306358605623245, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306358605623245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04671499878168106, + 0.07078590244054794, + 0.1162845641374588, + 0.05334196612238884, + 0.11905675381422043, + 0.25416329503059387, + 0.0580269955098629, + 0.03528710827231407, + 0.2101334035396576, + 0.03620503097772598 + ], + "credit_map": [ + 0.009890060871839523, + 0.004122386686503887, + 0.00616557989269495, + 0.006158117670565844, + 0.0044304318726062775, + 0.003918062429875135, + 0.00832648016512394, + 0.003703877329826355, + 0.0051693543791770935, + 0.012886946089565754, + 0.020434513688087463, + 0.0077972691506147385, + 0.006607030984014273, + 0.0030563771724700928, + 0.009608359076082706, + 0.005226051900535822, + 0.010698662139475346, + 0.01907077617943287, + 0.015748405829072, + 0.012211037799715996, + 0.013799238950014114, + 0.01825747638940811, + 0.012210252694785595, + 0.01428871788084507, + 0.00424278574064374, + 0.0056978873908519745, + 0.007911779917776585, + 0.005741877947002649, + 0.00715248566120863, + 0.007308506406843662, + 0.006653873715549707, + 0.00863276794552803, + 0.01598488911986351, + 0.013254608027637005, + 0.015170877799391747, + 0.01670381985604763, + 0.015575936995446682, + 0.015581540763378143, + 0.01355468574911356, + 0.013230403885245323, + 0.028226690366864204, + 0.032214947044849396, + 0.04003554955124855, + 0.023061446845531464, + 0.03782716393470764, + 0.03451170772314072, + 0.02230427786707878, + 0.03598151355981827, + 0.0165777076035738, + 0.004594557452946901, + 0.015123102813959122, + 0.00511227548122406, + 0.004867523442953825, + 0.004958401899784803, + 0.0036055129021406174, + 0.00318791507743299, + 0.005491935648024082, + 0.0031526931561529636, + 0.007801176514476538, + 0.004584155511111021, + 0.0046727159060537815, + 0.003765344386920333, + 0.002333733718842268, + 0.003485355293378234, + 0.02650955319404602, + 0.028533006086945534, + 0.03893594071269035, + 0.04316588491201401, + 0.01828240416944027, + 0.030974455177783966, + 0.01307774893939495, + 0.010654398240149021, + 0.0048145088367164135, + 0.004449798725545406, + 0.006972724571824074, + 0.005029730498790741, + 0.004060202278196812, + 0.004140888340771198, + 0.002879246836528182, + 0.0038579320535063744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.023306358605623245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01110091432929039 + ], + "opposite_logprob_tokens": [ + -4.511100769042969 + ], + "logprob_reward": -0.01110091432929039, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01110091432929039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036308202892541885, + 0.08597075939178467, + 0.09417423605918884, + 0.055026452988386154, + 0.2884058654308319, + 0.051175475120544434, + 0.07334188371896744, + 0.040775150060653687, + 0.12267664819955826, + 0.15214529633522034 + ], + "credit_map": [ + 0.005392035469412804, + 0.003980966750532389, + 0.006701992824673653, + 0.004975590389221907, + 0.0038558978121727705, + 0.0029910001903772354, + 0.004958414472639561, + 0.003452301723882556, + 0.009736619889736176, + 0.011373485438525677, + 0.02178512141108513, + 0.010130355134606361, + 0.009188558906316757, + 0.006146372761577368, + 0.0085598761215806, + 0.009050370194017887, + 0.008611977100372314, + 0.014311178587377071, + 0.012804191559553146, + 0.00801780354231596, + 0.013368611223995686, + 0.015435588546097279, + 0.01044213492423296, + 0.011182751506567001, + 0.005168858449906111, + 0.0047126044519245625, + 0.007621018216013908, + 0.006487465463578701, + 0.0075380923226475716, + 0.006977636832743883, + 0.008147913962602615, + 0.008372863754630089, + 0.03769645467400551, + 0.031027238816022873, + 0.04086246341466904, + 0.05014107748866081, + 0.034706808626651764, + 0.03716794028878212, + 0.030931036919355392, + 0.025872845202684402, + 0.004562343005090952, + 0.004503143485635519, + 0.009755617938935757, + 0.011252302676439285, + 0.0036286916583776474, + 0.00805206224322319, + 0.004439213778823614, + 0.0049820998683571815, + 0.012710072100162506, + 0.005637711379677057, + 0.02311564050614834, + 0.007816299796104431, + 0.0055607822723686695, + 0.007101360242813826, + 0.004023378249257803, + 0.007376640569418669, + 0.005666271783411503, + 0.003377411514520645, + 0.011413883417844772, + 0.004267050419002771, + 0.004055877216160297, + 0.004915549885481596, + 0.0027212505228817463, + 0.0043578543700277805, + 0.016934705898165703, + 0.020248189568519592, + 0.020489882677793503, + 0.02410353720188141, + 0.01191151887178421, + 0.014592841267585754, + 0.008847086690366268, + 0.0055488841608166695, + 0.026619093492627144, + 0.01162213645875454, + 0.037390489131212234, + 0.03176251798868179, + 0.01098147127777338, + 0.01761764846742153, + 0.008515890687704086, + 0.0076360455714166164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.01110091432929039 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0488789901137352, + 0.06604305654764175, + 0.17827120423316956, + 0.04121672734618187, + 0.09915833920240402, + 0.1721346378326416, + 0.12873654067516327, + 0.0651271864771843, + 0.15933141112327576, + 0.04110189527273178 + ], + "credit_map": [ + 0.009648947045207024, + 0.00405703904107213, + 0.008165860548615456, + 0.006552461534738541, + 0.004992297384887934, + 0.003667642129585147, + 0.007532546762377024, + 0.004262194037437439, + 0.008362885564565659, + 0.0099647156894207, + 0.008663029409945011, + 0.008517188020050526, + 0.0074043432250618935, + 0.005527215544134378, + 0.008919567801058292, + 0.008684113621711731, + 0.01633659191429615, + 0.030160728842020035, + 0.02499670535326004, + 0.018086988478899002, + 0.02103576995432377, + 0.029164627194404602, + 0.017151953652501106, + 0.02133784256875515, + 0.002930091228336096, + 0.006951390765607357, + 0.004923827014863491, + 0.006756386253982782, + 0.005062569864094257, + 0.005236366763710976, + 0.0036865852307528257, + 0.005669509060680866, + 0.007751530967652798, + 0.011187845841050148, + 0.014081737026572227, + 0.023464078083634377, + 0.00934057030826807, + 0.010345540009438992, + 0.011026701889932156, + 0.011960337869822979, + 0.019250525161623955, + 0.02265917882323265, + 0.026432663202285767, + 0.015609249472618103, + 0.02662447653710842, + 0.023516196757555008, + 0.016207415610551834, + 0.021834930405020714, + 0.03731498494744301, + 0.008629031479358673, + 0.03162597119808197, + 0.014843148179352283, + 0.01090683601796627, + 0.011550799012184143, + 0.007575796451419592, + 0.006289979908615351, + 0.008717038668692112, + 0.007267219480127096, + 0.010369524359703064, + 0.010625808499753475, + 0.010576761327683926, + 0.006717070005834103, + 0.004979673773050308, + 0.005874094553291798, + 0.023952670395374298, + 0.027164969593286514, + 0.024477548897266388, + 0.030408618971705437, + 0.01642671972513199, + 0.01575816236436367, + 0.012320038862526417, + 0.00882267951965332, + 0.00585712818428874, + 0.004975530318915844, + 0.00656773429363966, + 0.005847077816724777, + 0.005134186707437038, + 0.004693860188126564, + 0.0032469870056957006, + 0.004779389593750238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.01108264084905386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a macro, slightly high-angle shot, the camera slowly and steadily moving forward toward a tiny garden snail resting on a broad, vibrant green leaf. The snail remains perfectly curled and motionless, its translucent amber shell a perfect logarithmic spiral, its delicate concentric growth lines striking. The entire scene is studded with shimmering dewdrops, clinging to the waxy surface of the leaf and the snail's shell like tiny glass spheres. As the camera draws closer, the depth of field becomes incredibly shallow, the lush, out-of-focus garden background transforming into a soft, verdant blur as the shot transitions to extreme close-up. This deliberate movement reveals intricate and shimmering details: a dewdrop magnifying the texture of the leaf's underside, the delicate pearlescent sheen of the shell's surface, and the moist, delicate texture of the snail's visible flesh. The scene is bathed in a soft, golden morning light, creating brilliant highlights on the water droplets and casting gentle shadows that accentuate every curve, creating an atmosphere of tranquility, meditation, and wonder.", + "seed": 16331, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_5/index_4083_seed_16331_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.286917080800049, + "vlm_logprob": -0.286917080800049, + "vlm_yes_no_margin": 4.124999952316284, + "vlm_token_credit": -0.286917080800049, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001814506365917623 + ], + "opposite_logprob_tokens": [ + -6.751814365386963 + ], + "logprob_reward": -0.001814506365917623, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001814506365917623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07020936161279678, + 0.0397307313978672, + 0.06854302436113358, + 0.1881851851940155, + 0.3691677749156952, + 0.06908728182315826, + 0.05266550928354263, + 0.044462915509939194, + 0.031145526096224785, + 0.06680270284414291 + ], + "credit_map": [ + 0.007591786794364452, + 0.01026358176022768, + 0.007947282865643501, + 0.011106787249445915, + 0.010619457811117172, + 0.008095094934105873, + 0.009838070720434189, + 0.00474729435518384, + 0.003684697672724724, + 0.0044655585661530495, + 0.005775473080575466, + 0.0055744280107319355, + 0.005652240477502346, + 0.004285444971174002, + 0.005234831478446722, + 0.0050580562092363834, + 0.0038003413937985897, + 0.002953865798190236, + 0.022077182307839394, + 0.017334362491965294, + 0.004550027661025524, + 0.008264653384685516, + 0.004218963906168938, + 0.005343633703887463, + 0.02644668146967888, + 0.022241652011871338, + 0.022966839373111725, + 0.019254960119724274, + 0.01921253651380539, + 0.018838606774806976, + 0.02652229182422161, + 0.03270160034298897, + 0.026302674785256386, + 0.017286259680986404, + 0.12849536538124084, + 0.027378158643841743, + 0.016126621514558792, + 0.02190210483968258, + 0.021240506321191788, + 0.11043611168861389, + 0.0075043863616883755, + 0.0056255594827234745, + 0.009890167973935604, + 0.00756470300257206, + 0.005510253831744194, + 0.008320312015712261, + 0.008560405112802982, + 0.016111494973301888, + 0.0069671859964728355, + 0.005070958286523819, + 0.012859542854130268, + 0.004942032508552074, + 0.005541678052395582, + 0.006249308120459318, + 0.004912326578050852, + 0.006122478283941746, + 0.004950633272528648, + 0.00419584009796381, + 0.011062140576541424, + 0.003792989533394575, + 0.0034094369038939476, + 0.007087893784046173, + 0.0036662931088358164, + 0.006297687068581581, + 0.0037940945476293564, + 0.002394293900579214, + 0.007322567980736494, + 0.0032234492246061563, + 0.0032863125670701265, + 0.0038018522318452597, + 0.0030693013686686754, + 0.004253656603395939, + 0.007947737351059914, + 0.0058314623311161995, + 0.0072703915648162365, + 0.006338902283459902, + 0.006151911802589893, + 0.010103822685778141, + 0.0077901664189994335, + 0.015368307009339333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.001814506365917623 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063230752944946 + ], + "opposite_logprob_tokens": [ + -0.28132304549217224 + ], + "logprob_reward": -1.4063230752944946, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063230752944946, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.052674125880002975, + 0.038957931101322174, + 0.211557537317276, + 0.22677873075008392, + 0.18232692778110504, + 0.12170552462339401, + 0.07423284649848938, + 0.024730246514081955, + 0.027934137731790543, + 0.0391019806265831 + ], + "credit_map": [ + 0.0038318659644573927, + 0.005041392985731363, + 0.010191752575337887, + 0.00844903476536274, + 0.006545185577124357, + 0.004077175632119179, + 0.005738419946283102, + 0.008799300529062748, + 0.003856051480397582, + 0.004309445153921843, + 0.007616517599672079, + 0.0052624959498643875, + 0.004896559286862612, + 0.003521264297887683, + 0.004622247535735369, + 0.004873346537351608, + 0.010366206057369709, + 0.010949041694402695, + 0.07188741117715836, + 0.05406203120946884, + 0.015466282144188881, + 0.018115870654582977, + 0.012880907393991947, + 0.017829781398177147, + 0.02942448854446411, + 0.021949054673314095, + 0.03355054184794426, + 0.027125928550958633, + 0.019161399453878403, + 0.02438223548233509, + 0.036044806241989136, + 0.03514028340578079, + 0.01204381138086319, + 0.00951025914400816, + 0.0635080337524414, + 0.011926396749913692, + 0.009492481127381325, + 0.010481195524334908, + 0.009954094886779785, + 0.05541064962744713, + 0.014122399501502514, + 0.012627635151147842, + 0.013157814741134644, + 0.013574262149631977, + 0.013327624648809433, + 0.012200192548334599, + 0.01852206327021122, + 0.02417352981865406, + 0.007664809003472328, + 0.007408415898680687, + 0.009004429914057255, + 0.00911914836615324, + 0.0072357202880084515, + 0.016102854162454605, + 0.00873902253806591, + 0.008958444930613041, + 0.0026772525161504745, + 0.002279298845678568, + 0.004482600372284651, + 0.0024843693245202303, + 0.002668502042070031, + 0.003130832687020302, + 0.003062710864469409, + 0.003944679629057646, + 0.0031481205951422453, + 0.0030120587907731533, + 0.006020082626491785, + 0.003704494098201394, + 0.0034681111574172974, + 0.003003865946084261, + 0.002392612164840102, + 0.0031847916543483734, + 0.006383177358657122, + 0.0033290113788098097, + 0.004206834360957146, + 0.004500030539929867, + 0.003557687858119607, + 0.006225391291081905, + 0.004672722425311804, + 0.00622712355107069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -1.4063230752944946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189646303653717 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189646303653717, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018189646303653717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14160165190696716, + 0.21352705359458923, + 0.12176414579153061, + 0.2664852440357208, + 0.08139251917600632, + 0.05724211037158966, + 0.04232081398367882, + 0.017962416633963585, + 0.021286116912961006, + 0.03641792759299278 + ], + "credit_map": [ + 0.011944079771637917, + 0.012654941529035568, + 0.019966090098023415, + 0.021309977397322655, + 0.020512854680418968, + 0.017136938869953156, + 0.014818587340414524, + 0.023258177563548088, + 0.018377451226115227, + 0.019568653777241707, + 0.041425175964832306, + 0.030330941081047058, + 0.026882385835051537, + 0.026150263845920563, + 0.023227345198392868, + 0.027564851567149162, + 0.00732613867148757, + 0.00850139930844307, + 0.035729698836803436, + 0.02860410138964653, + 0.011726798489689827, + 0.006181578617542982, + 0.01281172875314951, + 0.010882697999477386, + 0.0237631406635046, + 0.03068593144416809, + 0.034285999834537506, + 0.03688867390155792, + 0.029251813888549805, + 0.027205394580960274, + 0.043851178139448166, + 0.04055308550596237, + 0.006478582043200731, + 0.0065590511076152325, + 0.02397310920059681, + 0.006564072798937559, + 0.005271529778838158, + 0.006173795089125633, + 0.0053955912590026855, + 0.02097679115831852, + 0.0054373000748455524, + 0.004483613185584545, + 0.008775411173701286, + 0.007072478998452425, + 0.004359077196568251, + 0.007337603718042374, + 0.005621820222586393, + 0.014154806733131409, + 0.004496750421822071, + 0.004050530958920717, + 0.009743394330143929, + 0.00474040349945426, + 0.004260178189724684, + 0.004796119872480631, + 0.004661699291318655, + 0.005571736954152584, + 0.002141057513654232, + 0.0015131535474210978, + 0.003168097697198391, + 0.0018347406294196844, + 0.001985403010621667, + 0.002378238597884774, + 0.002453259425237775, + 0.002488467376679182, + 0.0023357286117970943, + 0.0011300509795546532, + 0.008021018467843533, + 0.0014725864166393876, + 0.001147241098806262, + 0.0033282728400081396, + 0.0010464012157171965, + 0.002804816933348775, + 0.0038584861904382706, + 0.0025117641780525446, + 0.004138735122978687, + 0.002437908435240388, + 0.0037967346142977476, + 0.004943541716784239, + 0.004367220681160688, + 0.010363535955548286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.018189646303653717 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036202622577548027 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036202622577548027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036202622577548027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10080523788928986, + 0.027517542243003845, + 0.1346609890460968, + 0.06876331567764282, + 0.36861154437065125, + 0.10615021735429764, + 0.06480869650840759, + 0.04116635397076607, + 0.017644090577960014, + 0.06987200677394867 + ], + "credit_map": [ + 0.009847830049693584, + 0.008655812591314316, + 0.018225928768515587, + 0.016140678897500038, + 0.01589539460837841, + 0.011321388185024261, + 0.012084607034921646, + 0.008633588440716267, + 0.0026794206351041794, + 0.003064197488129139, + 0.0035840882919728756, + 0.0037737032398581505, + 0.0038879727944731712, + 0.003527506487444043, + 0.003985777962952852, + 0.0030148737132549286, + 0.007591459434479475, + 0.006033466663211584, + 0.046778012067079544, + 0.03449682518839836, + 0.008969548158347607, + 0.01314452849328518, + 0.007157393731176853, + 0.010489764623343945, + 0.006951453164219856, + 0.0063127330504357815, + 0.00911618210375309, + 0.009982491843402386, + 0.00670247757807374, + 0.008545011281967163, + 0.008535286411643028, + 0.012617682106792927, + 0.025037266314029694, + 0.015565292909741402, + 0.12653635442256927, + 0.02469434216618538, + 0.015583173371851444, + 0.022609028965234756, + 0.018779175356030464, + 0.11980690062046051, + 0.01147979125380516, + 0.010974721983075142, + 0.020070785656571388, + 0.012206105515360832, + 0.010597352869808674, + 0.010790045373141766, + 0.014121679589152336, + 0.015909742563962936, + 0.010030679404735565, + 0.005145109258592129, + 0.005571648478507996, + 0.006504659075289965, + 0.005145583767443895, + 0.01754305325448513, + 0.005415759049355984, + 0.009452198632061481, + 0.005185024347156286, + 0.002621725434437394, + 0.007386976387351751, + 0.004856502171605825, + 0.003299204632639885, + 0.0058884513564407825, + 0.004439687822014093, + 0.0074887825176119804, + 0.002565041184425354, + 0.0013018136378377676, + 0.0038031081203371286, + 0.0016850773245096207, + 0.0018682766240090132, + 0.002095641801133752, + 0.0018095392733812332, + 0.002515592612326145, + 0.010247216559946537, + 0.004809156991541386, + 0.02109414152801037, + 0.005938984919339418, + 0.004581193905323744, + 0.010268881916999817, + 0.004237447865307331, + 0.008694982156157494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.0036202622577548027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tiny garden snail resting on a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail perfectly curled and motionless throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera move slowly and steadily forward toward the snail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, golden morning light with visible dewdrops on the leaf and shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the shot transition to an extreme close-up with a shallow depth of field and a blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16916656494140625, + 0.24278517067432404, + 0.08820750564336777, + 0.3357267677783966, + 0.03540390357375145, + 0.04079850763082504, + 0.04954418167471886, + 0.010446381755173206, + 0.007309490814805031, + 0.020611580461263657 + ], + "credit_map": [ + 0.014756081625819206, + 0.014122297056019306, + 0.026902563869953156, + 0.024599023163318634, + 0.021421875804662704, + 0.01920410804450512, + 0.017997747287154198, + 0.03016287088394165, + 0.020981816574931145, + 0.023145828396081924, + 0.04640435427427292, + 0.03261836618185043, + 0.03154174983501434, + 0.028970129787921906, + 0.027047263458371162, + 0.03207565099000931, + 0.008220290765166283, + 0.0065200007520616055, + 0.020114606246352196, + 0.01388547569513321, + 0.008503278717398643, + 0.013247563503682613, + 0.010230787098407745, + 0.0074855005368590355, + 0.03345542773604393, + 0.038895685225725174, + 0.04235390946269035, + 0.0427040196955204, + 0.036269061267375946, + 0.03694792464375496, + 0.05576123669743538, + 0.04933948442339897, + 0.006933089345693588, + 0.0029222476296126842, + 0.0115611357614398, + 0.002898806007578969, + 0.0023304037749767303, + 0.002676254604011774, + 0.00291086221113801, + 0.0031711070332676172, + 0.004573032259941101, + 0.003196216421201825, + 0.005351263098418713, + 0.0038849124684929848, + 0.0037670491728931665, + 0.004025982692837715, + 0.005293294787406921, + 0.010706755332648754, + 0.003656132612377405, + 0.0027298189233988523, + 0.014994349330663681, + 0.00402050232514739, + 0.0033902283757925034, + 0.013822365552186966, + 0.0032488955184817314, + 0.003681887872517109, + 0.0011670599924400449, + 0.0008645171183161438, + 0.0021498948335647583, + 0.0008777843322604895, + 0.0010146605782210827, + 0.0016031828708946705, + 0.00120951805729419, + 0.0015597640303894877, + 0.0007117725326679647, + 0.0005386128323152661, + 0.0014695933787152171, + 0.0007021863129921257, + 0.0006887066410854459, + 0.0011133953230455518, + 0.0005318388575688004, + 0.0015533845871686935, + 0.0016252569621428847, + 0.0017006248235702515, + 0.002050441922619939, + 0.002431920962408185, + 0.002674429677426815, + 0.0020036674104630947, + 0.0031456139404326677, + 0.004979626275599003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.004637913778424263 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_tensor([9751, 9751])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_tensor([9751, 9751])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a23f1294f8fb2793babf728ffdde9aa796fd598 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_tensor([9751, 9751])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping long shot, moving slowly forward and gently circling from a slightly upward perspective, reveals a white-haired Taoist priest with an ethereal demeanor, standing alone on a flat rock atop a towering sacred mountain. His snow-white hair, held in a simple bun by a wooden hairpin, wisps fluttering gently in the mountain breeze. His face is gaunt, his eyes half-closed, his expression serene and focused. His long white beard falls to his chest. He wears a moon-white, wide-sleeved linen robe, its texture light and billowing gently in the flowing air. A vast, rolling sea of clouds surrounds him. The rising golden sun pierces the mist, casting magnificent beams of light across the mountaintop, casting a soft glow around him and casting a holy aura. He is practicing Tai Chi, his movements smooth and fluid, his body effortlessly executing the \"White Crane Spreads Its Wings\" technique, one arm raised high, the other pressed down, his stance steady and powerful. His movements are as smooth as flowing water; every turn and every arm stroke is silky smooth, without pause or delay. His wide sleeves undulate and drift like ripples, seemingly blending with the surrounding mist. The interweaving of cool clouds and warm morning light imbues the entire scene with a tranquil and solemn martial arts charm, embodying a profound artistic conception of the unity of man and nature.", + "seed": 14504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_9751_seed_14504_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03163498155772686, + "vlm_logprob": -0.03163498155772686, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03163498155772686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06452527642250061 + ], + "opposite_logprob_tokens": [ + -2.8145253658294678 + ], + "logprob_reward": -0.06452527642250061, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06452527642250061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1325267106294632, + 0.08016529679298401, + 0.1268712282180786, + 0.08142131567001343, + 0.05681958422064781, + 0.09114854037761688, + 0.09948194772005081, + 0.0608612522482872, + 0.2395481914281845, + 0.031155923381447792 + ], + "credit_map": [ + 0.017646675929427147, + 0.020615354180336, + 0.026124916970729828, + 0.013471276499330997, + 0.009383673779666424, + 0.01733018457889557, + 0.012162779457867146, + 0.015791848301887512, + 0.004180772230029106, + 0.020628124475479126, + 0.03255414217710495, + 0.0038224586751312017, + 0.0025592332240194082, + 0.012349993921816349, + 0.001904062693938613, + 0.002166514517739415, + 0.013218351639807224, + 0.023003114387392998, + 0.017031289637088776, + 0.022304045036435127, + 0.011915689334273338, + 0.01413509901612997, + 0.014250164851546288, + 0.011013469658792019, + 0.005798876751214266, + 0.01835709437727928, + 0.014948545023798943, + 0.005751363001763821, + 0.00715247355401516, + 0.011002952232956886, + 0.005536367185413837, + 0.012873651459813118, + 0.005509834736585617, + 0.007312594447284937, + 0.008720743469893932, + 0.006310207769274712, + 0.005304196383804083, + 0.010403269901871681, + 0.007053100038319826, + 0.006205642595887184, + 0.00940160546451807, + 0.012149081565439701, + 0.017328863963484764, + 0.009127114899456501, + 0.007297787349671125, + 0.011055619455873966, + 0.008778397925198078, + 0.01601007767021656, + 0.008366541005671024, + 0.010728497989475727, + 0.012920300476253033, + 0.01453907135874033, + 0.011863983236253262, + 0.013553439639508724, + 0.010076234117150307, + 0.017433883622288704, + 0.0051305475644767284, + 0.006019031163305044, + 0.005753280594944954, + 0.006926039233803749, + 0.009849615395069122, + 0.006471490021795034, + 0.013937179930508137, + 0.00677407206967473, + 0.013237162493169308, + 0.024157119914889336, + 0.031806401908397675, + 0.06892483681440353, + 0.023006940260529518, + 0.028334101662039757, + 0.026600882411003113, + 0.023480752483010292, + 0.0037631304003298283, + 0.003176494035869837, + 0.004271512385457754, + 0.003195269498974085, + 0.004624538589268923, + 0.004297190345823765, + 0.00409604050219059, + 0.0037317508831620216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.06452527642250061 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638768292963505 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007638768292963505, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638768292963505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19388341903686523, + 0.07294386625289917, + 0.04127756133675575, + 0.09739461541175842, + 0.09562704712152481, + 0.09094375371932983, + 0.1173199713230133, + 0.07085451483726501, + 0.16998890042304993, + 0.04976632818579674 + ], + "credit_map": [ + 0.01774287223815918, + 0.02458803541958332, + 0.05617721378803253, + 0.019011834636330605, + 0.011182510294020176, + 0.033797044306993484, + 0.014834389090538025, + 0.01654951274394989, + 0.008035562932491302, + 0.017846649512648582, + 0.007144495844841003, + 0.005419979803264141, + 0.009986520744860172, + 0.011245504952967167, + 0.006610561162233353, + 0.00665458757430315, + 0.004840006120502949, + 0.005502316169440746, + 0.005629364401102066, + 0.006632802542299032, + 0.004944767337292433, + 0.004693597089499235, + 0.004663528874516487, + 0.00437117787078023, + 0.005906654521822929, + 0.024590330198407173, + 0.025404224172234535, + 0.009634534828364849, + 0.006516794674098492, + 0.008630722761154175, + 0.0070663997903466225, + 0.0096449488773942, + 0.008334332145750523, + 0.008913524448871613, + 0.022333450615406036, + 0.008741635829210281, + 0.006377689074724913, + 0.022417746484279633, + 0.00857490487396717, + 0.009933763183653355, + 0.008379875682294369, + 0.009932360611855984, + 0.01347472332417965, + 0.012162935920059681, + 0.01478563528507948, + 0.009673533961176872, + 0.009818008169531822, + 0.012716677971184254, + 0.012679701671004295, + 0.01026791613548994, + 0.015549249947071075, + 0.01625765673816204, + 0.01398045476526022, + 0.012148852460086346, + 0.01903081312775612, + 0.017405321821570396, + 0.004621988628059626, + 0.007302869576960802, + 0.00762787414714694, + 0.009145095013082027, + 0.013746661134064198, + 0.008683261461555958, + 0.012053463608026505, + 0.007673298940062523, + 0.011672750115394592, + 0.01928885281085968, + 0.019650612026453018, + 0.051878929138183594, + 0.014566727913916111, + 0.019117234274744987, + 0.01715211756527424, + 0.01666167378425598, + 0.005217623431235552, + 0.0063348496332764626, + 0.007190314587205648, + 0.006054090801626444, + 0.007767044473439455, + 0.005259541794657707, + 0.007146226707845926, + 0.004796636290848255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.007638768292963505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204133853316307 + ], + "opposite_logprob_tokens": [ + -2.8120412826538086 + ], + "logprob_reward": -0.06204133853316307, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204133853316307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10520493984222412, + 0.08055412024259567, + 0.18575569987297058, + 0.06877154111862183, + 0.09919821470975876, + 0.04603569954633713, + 0.06433989107608795, + 0.05880976840853691, + 0.09547366201877594, + 0.1958564817905426 + ], + "credit_map": [ + 0.011790555901825428, + 0.00887226965278387, + 0.016655830666422844, + 0.010308041237294674, + 0.011337846517562866, + 0.023478159680962563, + 0.011030677706003189, + 0.011731559410691261, + 0.004838509950786829, + 0.01716909371316433, + 0.03036215342581272, + 0.004539783578366041, + 0.004198226146399975, + 0.010426810942590237, + 0.004652045201510191, + 0.004367501940578222, + 0.018837517127394676, + 0.0349428690969944, + 0.025035295635461807, + 0.032460376620292664, + 0.01742466166615486, + 0.021170148625969887, + 0.02011936902999878, + 0.015765460208058357, + 0.003796820528805256, + 0.018192969262599945, + 0.018895838409662247, + 0.006489473860710859, + 0.004789132159203291, + 0.005362702067941427, + 0.00499243987724185, + 0.006252167746424675, + 0.0067465705797076225, + 0.015382030978798866, + 0.0187609251588583, + 0.0071104844100773335, + 0.005552820861339569, + 0.03047153726220131, + 0.007380776572972536, + 0.007793079596012831, + 0.00337743922136724, + 0.004056133795529604, + 0.0062115793116390705, + 0.011144345626235008, + 0.00567233981564641, + 0.004680864047259092, + 0.006284001749008894, + 0.0046089994721114635, + 0.005549888592213392, + 0.005307427141815424, + 0.007329958491027355, + 0.011665022000670433, + 0.00828840397298336, + 0.007842634804546833, + 0.010820784606039524, + 0.007535777520388365, + 0.004916021600365639, + 0.007398026064038277, + 0.005837046541273594, + 0.01176702231168747, + 0.0069798813201487064, + 0.006212392821907997, + 0.010084591805934906, + 0.00561479339376092, + 0.007914837449789047, + 0.013261319138109684, + 0.0125468410551548, + 0.017550429329276085, + 0.011725997552275658, + 0.009589215740561485, + 0.014825260266661644, + 0.008059773594141006, + 0.020405316725373268, + 0.029708700254559517, + 0.026623541489243507, + 0.01695263758301735, + 0.03879999741911888, + 0.018746286630630493, + 0.029237501323223114, + 0.015382513403892517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.06204133853316307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778338484466076 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778338484466076, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778338484466076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1639813780784607, + 0.10375817865133286, + 0.2112506628036499, + 0.07364222407341003, + 0.1312732994556427, + 0.04124220460653305, + 0.03593798354268074, + 0.052841391414403915, + 0.07180849462747574, + 0.11426422744989395 + ], + "credit_map": [ + 0.02678057923913002, + 0.015896955505013466, + 0.013672995381057262, + 0.009807702153921127, + 0.01360946986824274, + 0.054725922644138336, + 0.014509286731481552, + 0.01497847493737936, + 0.009398006834089756, + 0.025629673153162003, + 0.023631706833839417, + 0.00853021815419197, + 0.012626799754798412, + 0.006550077814608812, + 0.009262620471417904, + 0.008129074238240719, + 0.02833261340856552, + 0.04175786301493645, + 0.026579411700367928, + 0.033305201679468155, + 0.02016484923660755, + 0.022137947380542755, + 0.018903560936450958, + 0.02006920427083969, + 0.008809446357190609, + 0.008993776515126228, + 0.008542953059077263, + 0.01003080140799284, + 0.008374982513487339, + 0.006318435538560152, + 0.010134694166481495, + 0.012437139637768269, + 0.01700429990887642, + 0.011794233694672585, + 0.025984305888414383, + 0.01673695631325245, + 0.01364993117749691, + 0.01657118648290634, + 0.014306116849184036, + 0.015226263552904129, + 0.004445788916200399, + 0.004943932872265577, + 0.005594240501523018, + 0.006636274978518486, + 0.005318726412951946, + 0.004454114008694887, + 0.00453777052462101, + 0.005311354994773865, + 0.0038302987813949585, + 0.0035411473363637924, + 0.004076804965734482, + 0.0061635891906917095, + 0.005221635103225708, + 0.0043657622300088406, + 0.004622840788215399, + 0.0041159032844007015, + 0.0055868360213935375, + 0.0067261685617268085, + 0.00588048342615366, + 0.008460523560643196, + 0.005549672991037369, + 0.0055069043301045895, + 0.010135102085769176, + 0.004995700437575579, + 0.006584002170711756, + 0.007131162565201521, + 0.011429926380515099, + 0.008547680452466011, + 0.0068544563837349415, + 0.00998761784285307, + 0.008912424556910992, + 0.012361227534711361, + 0.010453382506966591, + 0.015901057049632072, + 0.013804174028337002, + 0.009633270092308521, + 0.024724751710891724, + 0.013116071000695229, + 0.017422132194042206, + 0.009209388867020607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.009778338484466076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191186055541039 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191186055541039, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191186055541039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10341876745223999, + 0.08573325723409653, + 0.6334493160247803, + 0.06006452441215515, + 0.023496903479099274, + 0.019132107496261597, + 0.013598665595054626, + 0.025797564536333084, + 0.017645804211497307, + 0.0176631361246109 + ], + "credit_map": [ + 0.01422672625631094, + 0.0121384859085083, + 0.01931815594434738, + 0.008405258879065514, + 0.004972327500581741, + 0.021187009289860725, + 0.006858842447400093, + 0.01631196215748787, + 0.007358857896178961, + 0.02408958412706852, + 0.03537925332784653, + 0.00451688515022397, + 0.003186117857694626, + 0.005914771463721991, + 0.0024045121390372515, + 0.0028832622338086367, + 0.06935402005910873, + 0.12409593909978867, + 0.08900856226682663, + 0.10910376906394958, + 0.05673689767718315, + 0.06636081635951996, + 0.06582742184400558, + 0.05296188220381737, + 0.004201325122267008, + 0.010405427776277065, + 0.01146190520375967, + 0.007322413846850395, + 0.005627079401165247, + 0.006390987895429134, + 0.006966038607060909, + 0.007689344696700573, + 0.003522590734064579, + 0.002789114834740758, + 0.0042396606877446175, + 0.0024595102295279503, + 0.0017315789591521025, + 0.0037053648848086596, + 0.0021895503159612417, + 0.0028595307376235723, + 0.0022916074376553297, + 0.0028971426654607058, + 0.002321643754839897, + 0.002029644325375557, + 0.002325334120541811, + 0.0020425128750503063, + 0.0027780921664088964, + 0.0024461299180984497, + 0.0012653690064325929, + 0.0014886767603456974, + 0.0019059247570112348, + 0.0015237227780744433, + 0.0019960657227784395, + 0.0017707530641928315, + 0.001720729167573154, + 0.001927424455061555, + 0.0031695114448666573, + 0.002535390667617321, + 0.003549356712028384, + 0.0032361610792577267, + 0.0021741429809480906, + 0.002990879351273179, + 0.004145545419305563, + 0.003996575716882944, + 0.0011026877909898758, + 0.0017130685737356544, + 0.00197354843840003, + 0.005256129894405603, + 0.001354734762571752, + 0.0024078083224594593, + 0.0018573846900835633, + 0.0019804404582828283, + 0.001436316641047597, + 0.0023468327708542347, + 0.001809180248528719, + 0.0017629341455176473, + 0.0035515460185706615, + 0.0021210454870015383, + 0.0028406346682459116, + 0.0017946441657841206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.014191186055541039 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, moving slowly forward and gently circling from a slightly upward perspective, reveals a white-haired Taoist priest with an ethereal demeanor, standing alone on a flat rock atop a towering sacred mountain. His snow-white hair, held in a simple bun by a wooden hairpin, wisps fluttering gently in the mountain breeze. His face is gaunt, his eyes half-closed, his expression serene and focused. His long white beard falls to his chest. He wears a moon-white, wide-sleeved linen robe, its texture light and billowing gently in the flowing air. A vast, rolling sea of clouds surrounds him. The rising golden sun pierces the mist, casting magnificent beams of light across the mountaintop, casting a soft glow around him and casting a holy aura. He is practicing Tai Chi, his movements smooth and fluid, his body effortlessly executing the \"White Crane Spreads Its Wings\" technique, one arm raised high, the other pressed down, his stance steady and powerful. His movements are as smooth as flowing water; every turn and every arm stroke is silky smooth, without pause or delay. His wide sleeves undulate and drift like ripples, seemingly blending with the surrounding mist. The interweaving of cool clouds and warm morning light imbues the entire scene with a tranquil and solemn martial arts charm, embodying a profound artistic conception of the unity of man and nature.", + "seed": 14505, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_9751_seed_14505_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02998903808183968, + "vlm_logprob": -0.02998903808183968, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.02998903808183968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05036301165819168 + ], + "opposite_logprob_tokens": [ + -3.050363063812256 + ], + "logprob_reward": -0.05036301165819168, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05036301165819168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28594493865966797, + 0.05375606566667557, + 0.0601734034717083, + 0.08375895768404007, + 0.08172926306724548, + 0.15111449360847473, + 0.13083909451961517, + 0.04006371274590492, + 0.0895412489771843, + 0.023078901693224907 + ], + "credit_map": [ + 0.024216225370764732, + 0.037892792373895645, + 0.06633362919092178, + 0.03584754094481468, + 0.023138156160712242, + 0.04279118403792381, + 0.026160720735788345, + 0.02956465817987919, + 0.003922736272215843, + 0.005955989938229322, + 0.008621953427791595, + 0.005971433594822884, + 0.006438818294554949, + 0.013082570396363735, + 0.005338216200470924, + 0.004424344748258591, + 0.005337398033589125, + 0.007503402419388294, + 0.008314176462590694, + 0.009995111264288425, + 0.008380780927836895, + 0.00938283372670412, + 0.005909995641559362, + 0.005349695682525635, + 0.008545627817511559, + 0.007060602772980928, + 0.016419237479567528, + 0.01117539219558239, + 0.0060828267596662045, + 0.020136959850788116, + 0.008707210421562195, + 0.005631086882203817, + 0.009116378612816334, + 0.010047005489468575, + 0.01093977503478527, + 0.009775820188224316, + 0.009647144004702568, + 0.014520830474793911, + 0.008058193139731884, + 0.00962410494685173, + 0.015610394068062305, + 0.012884312309324741, + 0.019424771890044212, + 0.014205279760062695, + 0.01106620579957962, + 0.018865667283535004, + 0.02469705045223236, + 0.034360796213150024, + 0.01622890494763851, + 0.011340746656060219, + 0.02301727421581745, + 0.013770554214715958, + 0.011121083050966263, + 0.013119612820446491, + 0.01715061627328396, + 0.025090279057621956, + 0.005555488169193268, + 0.0036631098482757807, + 0.004769838415086269, + 0.0058038500137627125, + 0.004436674993485212, + 0.006426538806408644, + 0.005497056990861893, + 0.003911151550710201, + 0.009966998361051083, + 0.009109354577958584, + 0.010760007426142693, + 0.00904141180217266, + 0.01133242528885603, + 0.012682562693953514, + 0.01122252270579338, + 0.01542594563215971, + 0.002237382112070918, + 0.0020726011134684086, + 0.0029944765847176313, + 0.002245601499453187, + 0.0035005949903279543, + 0.002947306027635932, + 0.004426443483680487, + 0.002654495183378458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.05036301165819168 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743056699633598 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006743056699633598, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743056699633598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1189819797873497, + 0.1307566612958908, + 0.0765109658241272, + 0.0746234804391861, + 0.11774805933237076, + 0.04653453826904297, + 0.12349183857440948, + 0.052431222051382065, + 0.12385740131139755, + 0.13506384193897247 + ], + "credit_map": [ + 0.007018111180514097, + 0.0164236668497324, + 0.017306948080658913, + 0.022213276475667953, + 0.012789630331099033, + 0.018412597477436066, + 0.013095488771796227, + 0.011722262017428875, + 0.006518063135445118, + 0.015644405037164688, + 0.017432469874620438, + 0.03468076139688492, + 0.007920931093394756, + 0.024584896862506866, + 0.014781229197978973, + 0.009193929843604565, + 0.006169822067022324, + 0.008725075982511044, + 0.010444452986121178, + 0.01085134781897068, + 0.007832151837646961, + 0.012393408454954624, + 0.009008875116705894, + 0.01108583714812994, + 0.005595254711806774, + 0.009645284153521061, + 0.012645421549677849, + 0.012721006758511066, + 0.007230550516396761, + 0.013849123381078243, + 0.007516137324273586, + 0.005420713219791651, + 0.01468175183981657, + 0.013340492732822895, + 0.01843104138970375, + 0.010175946168601513, + 0.009233773685991764, + 0.025393549352884293, + 0.010888676159083843, + 0.015602834522724152, + 0.005033709108829498, + 0.00433451309800148, + 0.0073283277451992035, + 0.0045544165186584, + 0.003756657475605607, + 0.005475485697388649, + 0.005257041659206152, + 0.010794388130307198, + 0.016849057748913765, + 0.010703375563025475, + 0.02406027913093567, + 0.010687511414289474, + 0.010125745087862015, + 0.024100949987769127, + 0.01270263735204935, + 0.014262288808822632, + 0.005921242292970419, + 0.005628426559269428, + 0.0061632245779037476, + 0.007299272809177637, + 0.004799216985702515, + 0.006076903082430363, + 0.007745231967419386, + 0.008797709830105305, + 0.014758443459868431, + 0.012085998430848122, + 0.011677955277264118, + 0.014919132925570011, + 0.01565488427877426, + 0.017685892060399055, + 0.02136220969259739, + 0.01571289263665676, + 0.012391353026032448, + 0.013655240647494793, + 0.012990900315344334, + 0.01440118532627821, + 0.023155244067311287, + 0.013366728089749813, + 0.0269895289093256, + 0.018113672733306885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.006743056699633598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897137850522995 + ], + "opposite_logprob_tokens": [ + -2.5789713859558105 + ], + "logprob_reward": -0.07897137850522995, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897137850522995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16875626146793365, + 0.0520024374127388, + 0.05627672001719475, + 0.09420093148946762, + 0.07303955405950546, + 0.04745522886514664, + 0.2072802037000656, + 0.14898112416267395, + 0.07493625581264496, + 0.07707130163908005 + ], + "credit_map": [ + 0.01271994598209858, + 0.018784048035740852, + 0.021749434992671013, + 0.018161019310355186, + 0.03377578780055046, + 0.021745698526501656, + 0.016525041311979294, + 0.025295279920101166, + 0.003862921381369233, + 0.006086901295930147, + 0.009510817006230354, + 0.010416936129331589, + 0.004479140508919954, + 0.006150918547064066, + 0.006908639334142208, + 0.004586157388985157, + 0.005234773736447096, + 0.007873195223510265, + 0.008094354532659054, + 0.005651935003697872, + 0.007533883210271597, + 0.006355639547109604, + 0.006623952649533749, + 0.008908980526030064, + 0.006407680921256542, + 0.012405792251229286, + 0.011656447313725948, + 0.015695732086896896, + 0.0148220369592309, + 0.014821658842265606, + 0.010930468328297138, + 0.007461101748049259, + 0.008823052048683167, + 0.00843722652643919, + 0.01339774951338768, + 0.008224699646234512, + 0.005383058451116085, + 0.012346538715064526, + 0.006718841381371021, + 0.009708378463983536, + 0.005947985220700502, + 0.006971473805606365, + 0.007857115007936954, + 0.004394051153212786, + 0.004185430705547333, + 0.005022516939789057, + 0.00412595272064209, + 0.00895069632679224, + 0.02956276200711727, + 0.01642206870019436, + 0.041203759610652924, + 0.0207443218678236, + 0.016087235882878304, + 0.0362798310816288, + 0.022039735689759254, + 0.024940481409430504, + 0.017908988520503044, + 0.019380273297429085, + 0.012821870855987072, + 0.01891269162297249, + 0.020794382318854332, + 0.014796559698879719, + 0.02361411415040493, + 0.020752234384417534, + 0.007313466165214777, + 0.006239297799766064, + 0.006996466312557459, + 0.010902379639446735, + 0.014253033325076103, + 0.007385632023215294, + 0.011559822596609592, + 0.010286146774888039, + 0.006430582143366337, + 0.008027148433029652, + 0.009178699925541878, + 0.00883500650525093, + 0.0132330022752285, + 0.0073701161891222, + 0.012997029349207878, + 0.010999706573784351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.07897137850522995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531704269349575 + ], + "opposite_logprob_tokens": [ + -4.387531757354736 + ], + "logprob_reward": -0.012531704269349575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531704269349575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12029091268777847, + 0.05935288593173027, + 0.075415700674057, + 0.22625960409641266, + 0.09108644723892212, + 0.05235826224088669, + 0.18443165719509125, + 0.05959617719054222, + 0.08015362173318863, + 0.051054712384939194 + ], + "credit_map": [ + 0.011407418176531792, + 0.014829718507826328, + 0.02906324528157711, + 0.01704336144030094, + 0.01116044819355011, + 0.013856513425707817, + 0.011522169224917889, + 0.011408045887947083, + 0.005268053617328405, + 0.007810636423528194, + 0.00948124099522829, + 0.01054137758910656, + 0.005438908934593201, + 0.007045873440802097, + 0.006613373756408691, + 0.007153417449444532, + 0.006650312338024378, + 0.009975544176995754, + 0.009866164065897465, + 0.010424061678349972, + 0.011578187346458435, + 0.011436006985604763, + 0.00837060809135437, + 0.007114818785339594, + 0.028434962034225464, + 0.03646349161863327, + 0.037757836282253265, + 0.03320784866809845, + 0.01406743936240673, + 0.037377260625362396, + 0.014877576380968094, + 0.02407320775091648, + 0.013470732606947422, + 0.010276515036821365, + 0.01370473112910986, + 0.008544646203517914, + 0.008955865167081356, + 0.011996861547231674, + 0.008470775559544563, + 0.015666315332055092, + 0.005440918263047934, + 0.0057738409377634525, + 0.008540000766515732, + 0.006085721310228109, + 0.0054020993411540985, + 0.0059319185093045235, + 0.005965009797364473, + 0.009218755178153515, + 0.02164633199572563, + 0.019269103184342384, + 0.026523886248469353, + 0.023370523005723953, + 0.019233182072639465, + 0.019157806411385536, + 0.028809692710638046, + 0.026421140879392624, + 0.0077279601246118546, + 0.00876240711659193, + 0.005766850430518389, + 0.005451243836432695, + 0.006192280910909176, + 0.010014139115810394, + 0.006104057654738426, + 0.009577238000929356, + 0.008248688653111458, + 0.007669102866202593, + 0.0070190937258303165, + 0.010903222486376762, + 0.011060928925871849, + 0.007858783937990665, + 0.014179419726133347, + 0.01321437582373619, + 0.004260809160768986, + 0.004254247061908245, + 0.006511340849101543, + 0.004779784940183163, + 0.007651329971849918, + 0.005268377717584372, + 0.006961740553379059, + 0.01136708352714777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.012531704269349575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013360392767935991 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.0013360392767935991, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013360392767935991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17934344708919525, + 0.07870782911777496, + 0.06328339874744415, + 0.04481654614210129, + 0.10318315774202347, + 0.11501511186361313, + 0.22256363928318024, + 0.08878958225250244, + 0.07155635207891464, + 0.032740917056798935 + ], + "credit_map": [ + 0.016356050968170166, + 0.026838622987270355, + 0.04029958322644234, + 0.02082771062850952, + 0.014226067811250687, + 0.022126300260424614, + 0.015861764550209045, + 0.022807355970144272, + 0.00460679130628705, + 0.010360760614275932, + 0.01599293388426304, + 0.00579414889216423, + 0.006405577063560486, + 0.023429332301020622, + 0.006424745079129934, + 0.005693542305380106, + 0.005948106292635202, + 0.007127831690013409, + 0.010027946904301643, + 0.00930519588291645, + 0.008113282732665539, + 0.009280461817979813, + 0.005913855507969856, + 0.007566717453300953, + 0.004562584217637777, + 0.005724999122321606, + 0.007219882216304541, + 0.006622427608817816, + 0.004448463674634695, + 0.007654254790395498, + 0.005095193162560463, + 0.003488740650936961, + 0.01092259306460619, + 0.017070390284061432, + 0.01457651425153017, + 0.013059202581644058, + 0.012694480828940868, + 0.0123282540589571, + 0.008381054736673832, + 0.014150667935609818, + 0.011787576600909233, + 0.01318359375, + 0.01683446206152439, + 0.013262641616165638, + 0.00899800006300211, + 0.01458573155105114, + 0.019250815734267235, + 0.01711229234933853, + 0.028083035722374916, + 0.023393455892801285, + 0.0468214713037014, + 0.022553781047463417, + 0.019806884229183197, + 0.025721663609147072, + 0.02569791115820408, + 0.030485454946756363, + 0.011869588866829872, + 0.016096776351332664, + 0.01089390181005001, + 0.011774303391575813, + 0.009511872194707394, + 0.009976251050829887, + 0.008303607814013958, + 0.01036328449845314, + 0.009248546324670315, + 0.0063104163855314255, + 0.008019538596272469, + 0.0075833480805158615, + 0.009074294939637184, + 0.007306202780455351, + 0.010540909133851528, + 0.013473096303641796, + 0.004649526439607143, + 0.003667892888188362, + 0.005257908254861832, + 0.0028180796653032303, + 0.003452865406870842, + 0.0033281040377914906, + 0.004014065023511648, + 0.0055524748750030994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0013360392767935991 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, moving slowly forward and gently circling from a slightly upward perspective, reveals a white-haired Taoist priest with an ethereal demeanor, standing alone on a flat rock atop a towering sacred mountain. His snow-white hair, held in a simple bun by a wooden hairpin, wisps fluttering gently in the mountain breeze. His face is gaunt, his eyes half-closed, his expression serene and focused. His long white beard falls to his chest. He wears a moon-white, wide-sleeved linen robe, its texture light and billowing gently in the flowing air. A vast, rolling sea of clouds surrounds him. The rising golden sun pierces the mist, casting magnificent beams of light across the mountaintop, casting a soft glow around him and casting a holy aura. He is practicing Tai Chi, his movements smooth and fluid, his body effortlessly executing the \"White Crane Spreads Its Wings\" technique, one arm raised high, the other pressed down, his stance steady and powerful. His movements are as smooth as flowing water; every turn and every arm stroke is silky smooth, without pause or delay. His wide sleeves undulate and drift like ripples, seemingly blending with the surrounding mist. The interweaving of cool clouds and warm morning light imbues the entire scene with a tranquil and solemn martial arts charm, embodying a profound artistic conception of the unity of man and nature.", + "seed": 14506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_9751_seed_14506_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02031831885688007, + "vlm_logprob": -0.02031831885688007, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.02031831885688007, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045865900814533234 + ], + "opposite_logprob_tokens": [ + -3.170866012573242 + ], + "logprob_reward": -0.045865900814533234, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045865900814533234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3809906840324402, + 0.40348297357559204, + 0.03774731233716011, + 0.017572183161973953, + 0.014141540043056011, + 0.030541017651557922, + 0.03865240514278412, + 0.01687999814748764, + 0.05030982196331024, + 0.009682071395218372 + ], + "credit_map": [ + 0.0770387351512909, + 0.023753151297569275, + 0.040815312415361404, + 0.024004297330975533, + 0.02113800123333931, + 0.13522453606128693, + 0.031444307416677475, + 0.02757233940064907, + 0.1285461038351059, + 0.0189268346875906, + 0.17570844292640686, + 0.022643502801656723, + 0.015610624104738235, + 0.017284531146287918, + 0.012212170287966728, + 0.012550726532936096, + 0.004740489646792412, + 0.00461204070597887, + 0.0047497437335550785, + 0.0038947772700339556, + 0.004617721773684025, + 0.006162960547953844, + 0.003636276815086603, + 0.005333303939551115, + 0.0015872193034738302, + 0.0012416228419169784, + 0.0025463260244578123, + 0.0025679876562207937, + 0.0017384892562404275, + 0.0021778964437544346, + 0.002150147221982479, + 0.00356249394826591, + 0.002640659222379327, + 0.001499385223723948, + 0.0023446998093277216, + 0.001535799354314804, + 0.0012732594041153789, + 0.0017735192086547613, + 0.0014729400863870978, + 0.0016012780833989382, + 0.00226507312618196, + 0.003452582750469446, + 0.009072448126971722, + 0.0029078975785523653, + 0.0019128448329865932, + 0.005030236206948757, + 0.002792041515931487, + 0.003107893280684948, + 0.00397943239659071, + 0.004213001579046249, + 0.007762010674923658, + 0.003193845972418785, + 0.003942490555346012, + 0.005751772318035364, + 0.0038107591681182384, + 0.005999093409627676, + 0.0011308544781059027, + 0.0017957740928977728, + 0.004249418620020151, + 0.001692029065452516, + 0.0012183142825961113, + 0.003729019546881318, + 0.0015716392081230879, + 0.0014929476892575622, + 0.002526352182030678, + 0.0025951575953513384, + 0.018466560170054436, + 0.0028499362524598837, + 0.0019199452362954617, + 0.017347289249300957, + 0.0019674545619636774, + 0.002637128345668316, + 0.0007731533842161298, + 0.0008709111716598272, + 0.001998689491301775, + 0.0008978297701105475, + 0.0006434908718802035, + 0.001590621075592935, + 0.0010799446608871222, + 0.0018274304457008839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.045865900814533234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644422516226768 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644422516226768, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644422516226768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14149893820285797, + 0.5028945207595825, + 0.06118214130401611, + 0.052006859332323074, + 0.0484897717833519, + 0.02675820142030716, + 0.043153684586286545, + 0.08038163930177689, + 0.01649058237671852, + 0.027143657207489014 + ], + "credit_map": [ + 0.015815719962120056, + 0.01475021056830883, + 0.03237694501876831, + 0.020064346492290497, + 0.01429098378866911, + 0.010925168171525002, + 0.0181107260286808, + 0.015164821408689022, + 0.16638581454753876, + 0.020826231688261032, + 0.22434014081954956, + 0.02744089998304844, + 0.014638264663517475, + 0.022938115522265434, + 0.013641037046909332, + 0.012684036046266556, + 0.006485183257609606, + 0.004955294542014599, + 0.007808567490428686, + 0.00734569039195776, + 0.005504859611392021, + 0.007934276945888996, + 0.00791651476174593, + 0.013231752440333366, + 0.004309058655053377, + 0.005405658856034279, + 0.006993436720222235, + 0.00699606304988265, + 0.005296686198562384, + 0.006202784366905689, + 0.00863567739725113, + 0.00816749781370163, + 0.005462630186229944, + 0.005408293567597866, + 0.006993340793997049, + 0.0066793435253202915, + 0.0069213202223181725, + 0.004951633047312498, + 0.006190495565533638, + 0.005882713012397289, + 0.00205027824267745, + 0.0024466635659337044, + 0.00454309955239296, + 0.004908959846943617, + 0.002496923552826047, + 0.0026988377794623375, + 0.004241670947521925, + 0.003371767234057188, + 0.00353901251219213, + 0.006755136419087648, + 0.006621804554015398, + 0.006011656951159239, + 0.004106168635189533, + 0.004773491527885199, + 0.004873374011367559, + 0.0064730411395430565, + 0.005778174381703138, + 0.0070648095570504665, + 0.020341770723462105, + 0.007243485189974308, + 0.00579027971252799, + 0.018305830657482147, + 0.008655518293380737, + 0.007201764266937971, + 0.0010086874244734645, + 0.0015837873797863722, + 0.0027682874351739883, + 0.0024694418534636497, + 0.0012430957285687327, + 0.002045797649770975, + 0.002300851745530963, + 0.0030706317629665136, + 0.001833464251831174, + 0.005199553910642862, + 0.004122791346162558, + 0.003459538798779249, + 0.0034903534688055515, + 0.0029121392872184515, + 0.00329692754894495, + 0.0028288881294429302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.008644422516226768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033690378069877625 + ], + "opposite_logprob_tokens": [ + -3.4086904525756836 + ], + "logprob_reward": -0.033690378069877625, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033690378069877625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31869322061538696, + 0.20076628029346466, + 0.16379927098751068, + 0.07935450226068497, + 0.03550102189183235, + 0.09777656942605972, + 0.027843967080116272, + 0.02178829349577427, + 0.040441278368234634, + 0.014035580679774284 + ], + "credit_map": [ + 0.06628112494945526, + 0.01735047996044159, + 0.030522048473358154, + 0.016776686534285545, + 0.018543638288974762, + 0.11909085512161255, + 0.0259109940379858, + 0.0242173969745636, + 0.06413821130990982, + 0.007712631486356258, + 0.0898350179195404, + 0.011986467987298965, + 0.00579031789675355, + 0.00840720534324646, + 0.007039043121039867, + 0.005857387091964483, + 0.014587041921913624, + 0.014341864734888077, + 0.019623294472694397, + 0.02417370304465294, + 0.016412606462836266, + 0.014271354302763939, + 0.02676665410399437, + 0.033622752875089645, + 0.006360452156513929, + 0.00777957821264863, + 0.01049865037202835, + 0.010541848838329315, + 0.007458680309355259, + 0.010067855007946491, + 0.009400310926139355, + 0.01724712736904621, + 0.00570242665708065, + 0.004011871758848429, + 0.005418756511062384, + 0.0045200553722679615, + 0.003752164775505662, + 0.0038926592096686363, + 0.004361420404165983, + 0.0038416702300310135, + 0.01274165790528059, + 0.011307713575661182, + 0.011736650951206684, + 0.010154246352612972, + 0.014121980406343937, + 0.012476895935833454, + 0.012179980054497719, + 0.013057450763881207, + 0.002540973713621497, + 0.004023305140435696, + 0.006879928056150675, + 0.0030405493453145027, + 0.002646777080371976, + 0.004016330000013113, + 0.0024203502107411623, + 0.002275751670822501, + 0.0016197699587792158, + 0.0035058448556810617, + 0.003559516742825508, + 0.00227821571752429, + 0.002125937957316637, + 0.0032768656965345144, + 0.00240132468752563, + 0.0030208167154341936, + 0.0029456450138241053, + 0.005206456873565912, + 0.007761722430586815, + 0.005631564185023308, + 0.0029018050990998745, + 0.006549909245222807, + 0.004630913492292166, + 0.00481326412409544, + 0.0014605335891246796, + 0.0017839532811194658, + 0.0019901078194379807, + 0.0016303288284689188, + 0.0009687198325991631, + 0.0016754784155637026, + 0.0022817819844931364, + 0.0022446769289672375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.033690378069877625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776213206350803 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776213206350803, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776213206350803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2582439184188843, + 0.10366114974021912, + 0.06046893447637558, + 0.05435461178421974, + 0.05200750380754471, + 0.07570809125900269, + 0.09230580180883408, + 0.11343710124492645, + 0.16168829798698425, + 0.028124501928687096 + ], + "credit_map": [ + 0.041196420788764954, + 0.027721351012587547, + 0.06141727417707443, + 0.025342196226119995, + 0.02036888524889946, + 0.04816301912069321, + 0.01864340342581272, + 0.015391383320093155, + 0.02561648003757, + 0.006965679116547108, + 0.03324122354388237, + 0.007577328011393547, + 0.012190400622785091, + 0.0060175214894115925, + 0.006884254515171051, + 0.005168276373296976, + 0.007307983003556728, + 0.005936165805906057, + 0.007792597636580467, + 0.007650510873645544, + 0.00675528310239315, + 0.0074092368595302105, + 0.00648722518235445, + 0.011129938066005707, + 0.005353192798793316, + 0.004653606563806534, + 0.008217592723667622, + 0.006835805252194405, + 0.005903509445488453, + 0.006658894009888172, + 0.0072366297245025635, + 0.00949538592249155, + 0.007945733144879341, + 0.006438703276216984, + 0.007631552871316671, + 0.005907986778765917, + 0.0065386914648115635, + 0.00678598927333951, + 0.006019870284944773, + 0.00473897997289896, + 0.006431443616747856, + 0.008045552298426628, + 0.019307250156998634, + 0.006509100086987019, + 0.007808933034539223, + 0.011398999951779842, + 0.007378657814115286, + 0.008828165009617805, + 0.010129151865839958, + 0.017991727218031883, + 0.014344754628837109, + 0.010887239128351212, + 0.011675696820020676, + 0.010787581093609333, + 0.009612818248569965, + 0.006876837462186813, + 0.0105196014046669, + 0.018891485407948494, + 0.014864029362797737, + 0.01227768324315548, + 0.013673955574631691, + 0.012898492626845837, + 0.013625992462038994, + 0.016685863956809044, + 0.007554405368864536, + 0.008614251390099525, + 0.058086805045604706, + 0.009955620393157005, + 0.007090601604431868, + 0.05509158968925476, + 0.007536032702773809, + 0.007759020198136568, + 0.0018450599163770676, + 0.004357409663498402, + 0.00629224069416523, + 0.0032970921602100134, + 0.0025950837880373, + 0.003470593597739935, + 0.0029546625446528196, + 0.0033123630564659834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.009776213206350803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614679677411914 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003614679677411914, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614679677411914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16682042181491852, + 0.19415709376335144, + 0.06768327951431274, + 0.05851954221725464, + 0.06818448752164841, + 0.04259685426950455, + 0.031354401260614395, + 0.09913389384746552, + 0.07693846523761749, + 0.19461162388324738 + ], + "credit_map": [ + 0.03154991939663887, + 0.010289784520864487, + 0.019989250227808952, + 0.010724121704697609, + 0.010528055019676685, + 0.05671808123588562, + 0.014613952487707138, + 0.012407242320477962, + 0.0642414391040802, + 0.0068532731384038925, + 0.08763771504163742, + 0.01028310228139162, + 0.005135569255799055, + 0.007754128891974688, + 0.005479211453348398, + 0.006772625260055065, + 0.009308606386184692, + 0.008062893524765968, + 0.007881558500230312, + 0.0074925352819263935, + 0.006330690812319517, + 0.013080631382763386, + 0.006710854358971119, + 0.008815506473183632, + 0.010268216952681541, + 0.005452410783618689, + 0.011607550084590912, + 0.005228054244071245, + 0.006347762420773506, + 0.006956298835575581, + 0.005279069766402245, + 0.007380177732557058, + 0.00831943191587925, + 0.00724386190995574, + 0.010362195782363415, + 0.007713287137448788, + 0.008017386309802532, + 0.009933987632393837, + 0.005797279067337513, + 0.010797053575515747, + 0.003481445834040642, + 0.008822817355394363, + 0.008400910533964634, + 0.005134609527885914, + 0.005225908011198044, + 0.00446902122348547, + 0.003771773772314191, + 0.0032903614919632673, + 0.0036518173292279243, + 0.0030886861495673656, + 0.006402958184480667, + 0.003407986368983984, + 0.0020259942393749952, + 0.004648880567401648, + 0.0036245256196707487, + 0.00450355326756835, + 0.0069952006451785564, + 0.015626925975084305, + 0.01905364915728569, + 0.010137064382433891, + 0.010318761691451073, + 0.01640237867832184, + 0.009092533960938454, + 0.011507364921271801, + 0.005162758287042379, + 0.006331458687782288, + 0.023007985204458237, + 0.007180328946560621, + 0.004720696713775396, + 0.019612209871411324, + 0.005627048667520285, + 0.005295971408486366, + 0.016374269500374794, + 0.029844654724001884, + 0.04650221765041351, + 0.018870502710342407, + 0.015252327546477318, + 0.024348294362425804, + 0.01814592443406582, + 0.02527341991662979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.003614679677411914 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, moving slowly forward and gently circling from a slightly upward perspective, reveals a white-haired Taoist priest with an ethereal demeanor, standing alone on a flat rock atop a towering sacred mountain. His snow-white hair, held in a simple bun by a wooden hairpin, wisps fluttering gently in the mountain breeze. His face is gaunt, his eyes half-closed, his expression serene and focused. His long white beard falls to his chest. He wears a moon-white, wide-sleeved linen robe, its texture light and billowing gently in the flowing air. A vast, rolling sea of clouds surrounds him. The rising golden sun pierces the mist, casting magnificent beams of light across the mountaintop, casting a soft glow around him and casting a holy aura. He is practicing Tai Chi, his movements smooth and fluid, his body effortlessly executing the \"White Crane Spreads Its Wings\" technique, one arm raised high, the other pressed down, his stance steady and powerful. His movements are as smooth as flowing water; every turn and every arm stroke is silky smooth, without pause or delay. His wide sleeves undulate and drift like ripples, seemingly blending with the surrounding mist. The interweaving of cool clouds and warm morning light imbues the entire scene with a tranquil and solemn martial arts charm, embodying a profound artistic conception of the unity of man and nature.", + "seed": 14507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_6/index_9751_seed_14507_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02297271927818656, + "vlm_logprob": -0.02297271927818656, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02297271927818656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028854381293058395 + ], + "opposite_logprob_tokens": [ + -3.6538543701171875 + ], + "logprob_reward": -0.028854381293058395, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028854381293058395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24100257456302643, + 0.08375102281570435, + 0.08688533306121826, + 0.09859980642795563, + 0.08172165602445602, + 0.14975692331790924, + 0.05293678864836693, + 0.10252849012613297, + 0.037186384201049805, + 0.06563106924295425 + ], + "credit_map": [ + 0.025373343378305435, + 0.04349712282419205, + 0.05512499064207077, + 0.030115649104118347, + 0.025221316143870354, + 0.022900737822055817, + 0.02150818705558777, + 0.01726122945547104, + 0.012282374314963818, + 0.0167677141726017, + 0.021816661581397057, + 0.006349750328809023, + 0.005944544915109873, + 0.004993709735572338, + 0.008697598241269588, + 0.006898659281432629, + 0.006180410273373127, + 0.005600397940725088, + 0.025967685505747795, + 0.0043725017458200455, + 0.003383501200005412, + 0.008655546233057976, + 0.0046620494686067104, + 0.02806323766708374, + 0.014112033881247044, + 0.007505747489631176, + 0.01866733282804489, + 0.013867066241800785, + 0.007855014875531197, + 0.013708394020795822, + 0.009294211864471436, + 0.01358998753130436, + 0.007116798777133226, + 0.0127733014523983, + 0.00939236395061016, + 0.011557292193174362, + 0.009066928178071976, + 0.011119256727397442, + 0.010912747122347355, + 0.009782960638403893, + 0.022906726226210594, + 0.02584012597799301, + 0.0204900074750185, + 0.021672574803233147, + 0.015138303861021996, + 0.0212300606071949, + 0.01339405681937933, + 0.009085055440664291, + 0.007533404976129532, + 0.008633675053715706, + 0.010870683938264847, + 0.00558020593598485, + 0.004682331345975399, + 0.006963677704334259, + 0.005400242283940315, + 0.0032725606579333544, + 0.00915147177875042, + 0.019859496504068375, + 0.011615713126957417, + 0.017964450642466545, + 0.007951856590807438, + 0.01121200155466795, + 0.014724286273121834, + 0.010049199685454369, + 0.0048651196993887424, + 0.005247981753200293, + 0.0042060003615915775, + 0.005839698947966099, + 0.005811110604554415, + 0.003956155851483345, + 0.003998009953647852, + 0.0032623037695884705, + 0.005697916261851788, + 0.007935997098684311, + 0.013443609699606895, + 0.009708574041724205, + 0.007574799936264753, + 0.009129292331635952, + 0.005229061469435692, + 0.006911814212799072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.028854381293058395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083937250077724 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011083937250077724, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083937250077724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08603320270776749, + 0.08793887495994568, + 0.24161763489246368, + 0.0582931824028492, + 0.03162407502532005, + 0.03821280598640442, + 0.11919864267110825, + 0.11279045790433884, + 0.07417016476392746, + 0.15012095868587494 + ], + "credit_map": [ + 0.007945823483169079, + 0.011099150404334068, + 0.009779704734683037, + 0.012443768791854382, + 0.015481019392609596, + 0.006695025134831667, + 0.01320360042154789, + 0.00938511174172163, + 0.006670011207461357, + 0.020344503223896027, + 0.028070569038391113, + 0.005980628076940775, + 0.0078738359734416, + 0.00565935717895627, + 0.007502921391278505, + 0.005837051197886467, + 0.03510458394885063, + 0.017647147178649902, + 0.05774756893515587, + 0.01763608306646347, + 0.02621951326727867, + 0.03263159841299057, + 0.031662967056035995, + 0.022968187928199768, + 0.006795587483793497, + 0.004069233778864145, + 0.012248932383954525, + 0.008462019264698029, + 0.004143344238400459, + 0.009190261363983154, + 0.005589043255895376, + 0.007794762961566448, + 0.0036309389397501945, + 0.004839026369154453, + 0.003993944730609655, + 0.0038320596795529127, + 0.0033889184705913067, + 0.003747864393517375, + 0.004043784458190203, + 0.004147539846599102, + 0.0036682591307908297, + 0.004355180077254772, + 0.005492230877280235, + 0.004487886093556881, + 0.006160837598145008, + 0.005310724023729563, + 0.004711054731160402, + 0.004026640206575394, + 0.011141249909996986, + 0.010894296690821648, + 0.014549780637025833, + 0.02116713859140873, + 0.014245297759771347, + 0.012985828332602978, + 0.021868392825126648, + 0.012346669100224972, + 0.009852726012468338, + 0.022231165319681168, + 0.014080902561545372, + 0.01859835535287857, + 0.00949605368077755, + 0.012329799123108387, + 0.015191660262644291, + 0.011009798385202885, + 0.0070008947513997555, + 0.0071463435888290405, + 0.0124263446778059, + 0.011691301129758358, + 0.01211603358387947, + 0.009218700230121613, + 0.007738194428384304, + 0.006832357961684465, + 0.015857389196753502, + 0.01664525270462036, + 0.023853447288274765, + 0.022586610168218613, + 0.020423399284482002, + 0.016588836908340454, + 0.014361913315951824, + 0.019804123789072037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.011083937250077724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204738840460777 + ], + "opposite_logprob_tokens": [ + -2.8120474815368652 + ], + "logprob_reward": -0.06204738840460777, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204738840460777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06316369026899338, + 0.06879063695669174, + 0.03079220838844776, + 0.022844992578029633, + 0.08557222038507462, + 0.033022962510585785, + 0.018049033358693123, + 0.028055036440491676, + 0.03341124206781387, + 0.6162980198860168 + ], + "credit_map": [ + 0.00687642116099596, + 0.009067283943295479, + 0.009401623159646988, + 0.008911405690014362, + 0.01056700013577938, + 0.0061415331438183784, + 0.0070708501152694225, + 0.005127567797899246, + 0.007450875826179981, + 0.019006891176104546, + 0.024668868631124496, + 0.004205774050205946, + 0.004313379526138306, + 0.0021785018034279346, + 0.002844313159584999, + 0.004122031386941671, + 0.004103475715965033, + 0.0038442034274339676, + 0.003357062814757228, + 0.0025663799606263638, + 0.002820343244820833, + 0.00610019825398922, + 0.003371485508978367, + 0.004629058297723532, + 0.002552085556089878, + 0.003464421723037958, + 0.0024015719536691904, + 0.0027127540670335293, + 0.003613515058532357, + 0.0031001565512269735, + 0.00268427818082273, + 0.0023162097204476595, + 0.011241555213928223, + 0.010112344287335873, + 0.00830154214054346, + 0.010146625339984894, + 0.014512150548398495, + 0.010501124896109104, + 0.012097890488803387, + 0.00865898746997118, + 0.003459045197814703, + 0.0037018517032265663, + 0.005574877839535475, + 0.005922432988882065, + 0.004074651747941971, + 0.003929114900529385, + 0.0035648567136377096, + 0.002796133514493704, + 0.0011868259171023965, + 0.002059470396488905, + 0.0022508136462420225, + 0.0031778994016349316, + 0.0027966811321675777, + 0.002003247616812587, + 0.0027099058497697115, + 0.0018641890492290258, + 0.002568993018940091, + 0.004501374904066324, + 0.004409352783113718, + 0.004845922347158194, + 0.002821129048243165, + 0.003549627959728241, + 0.003473342163488269, + 0.0018852934008464217, + 0.0035556829534471035, + 0.004654953256249428, + 0.004587233066558838, + 0.00466647045686841, + 0.005045196041464806, + 0.003598084207624197, + 0.0043239169754087925, + 0.002979706972837448, + 0.05172114819288254, + 0.05418407917022705, + 0.14259447157382965, + 0.08663183450698853, + 0.06408523768186569, + 0.11327875405550003, + 0.0472729317843914, + 0.05652955546975136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.06204738840460777 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625873666256666 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625873666256666, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625873666256666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18442685902118683, + 0.1762160211801529, + 0.22150473296642303, + 0.06069811433553696, + 0.14067140221595764, + 0.04137155041098595, + 0.037452686578035355, + 0.03811599314212799, + 0.0501810759305954, + 0.049361538141965866 + ], + "credit_map": [ + 0.02157732844352722, + 0.025285951793193817, + 0.041866593062877655, + 0.016966791823506355, + 0.01908799074590206, + 0.021475711837410927, + 0.015200484544038773, + 0.022966017946600914, + 0.015188055112957954, + 0.0517813116312027, + 0.05919138342142105, + 0.008956849575042725, + 0.009845012798905373, + 0.008276877924799919, + 0.013127629645168781, + 0.00984888058155775, + 0.025586647912859917, + 0.013777032494544983, + 0.07268873602151871, + 0.01091507263481617, + 0.01640504039824009, + 0.022693004459142685, + 0.017898831516504288, + 0.04154036194086075, + 0.009979983791708946, + 0.0072499047964811325, + 0.008902177214622498, + 0.007404142990708351, + 0.00580563023686409, + 0.007335648871958256, + 0.0074654011987149715, + 0.006555226165801287, + 0.016693849116563797, + 0.014173900708556175, + 0.015112541615962982, + 0.015281646512448788, + 0.02622045762836933, + 0.017655890434980392, + 0.02110747992992401, + 0.014425646513700485, + 0.003930004313588142, + 0.004234141670167446, + 0.007026794832199812, + 0.005426491145044565, + 0.0049300664104521275, + 0.005233817733824253, + 0.007579234428703785, + 0.0030109998770058155, + 0.0030437835957854986, + 0.004415414296090603, + 0.005255617666989565, + 0.005342148244380951, + 0.004383872728794813, + 0.004654006566852331, + 0.0066800289787352085, + 0.003677817527204752, + 0.0031552761793136597, + 0.0050329347141087055, + 0.005886526312679052, + 0.005571933928877115, + 0.00530067877843976, + 0.004724429454654455, + 0.004281670320779085, + 0.004162541590631008, + 0.00815503392368555, + 0.008945737965404987, + 0.005357257556170225, + 0.006927476264536381, + 0.007109779864549637, + 0.005207816604524851, + 0.004605815280228853, + 0.0038721575401723385, + 0.004454115871340036, + 0.003755867714062333, + 0.012603002600371838, + 0.008421980775892735, + 0.00363983609713614, + 0.008259659633040428, + 0.004929929506033659, + 0.0032971419859677553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007625873666256666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a flat rock atop a towering sacred mountain with a vast sea of clouds surrounding it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a white-haired Taoist priest wearing a moon-white wide-sleeved linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest performing the 'White Crane Spreads Its Wings' Tai Chi technique with one arm raised and the other pressed down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a sweeping long shot that moves slowly forward and gently circles from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a rising golden sun piercing the mist and casting beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10837516188621521, + 0.6991679072380066, + 0.07905757427215576, + 0.02230018936097622, + 0.010727256536483765, + 0.00818520039319992, + 0.005590158514678478, + 0.025035448372364044, + 0.012136010453104973, + 0.029425079002976418 + ], + "credit_map": [ + 0.01279496680945158, + 0.022639203816652298, + 0.020602745935320854, + 0.01453876681625843, + 0.011649715714156628, + 0.00852868054062128, + 0.009998598136007786, + 0.0076224906370043755, + 0.05668742209672928, + 0.21924683451652527, + 0.27027687430381775, + 0.03553326800465584, + 0.0347876138985157, + 0.023470718413591385, + 0.025939006358385086, + 0.03322621434926987, + 0.010089763440191746, + 0.008267759345471859, + 0.02278049662709236, + 0.005372587125748396, + 0.0050935084000229836, + 0.006920869927853346, + 0.005589169915765524, + 0.014943431131541729, + 0.0035497057251632214, + 0.0029565857257694006, + 0.002574146492406726, + 0.0022986752446740866, + 0.00208127754740417, + 0.003316459944471717, + 0.0024778058286756277, + 0.0030455368105322123, + 0.0012899029534310102, + 0.001720346393994987, + 0.0016525165410712361, + 0.0010919807245954871, + 0.0012769416207447648, + 0.0014269498642534018, + 0.001155251171439886, + 0.0011133685475215316, + 0.0011865821434184909, + 0.001192816998809576, + 0.0010188162559643388, + 0.0009330928442068398, + 0.000860352476593107, + 0.0010064723901450634, + 0.0012805549195036292, + 0.0007065130630508065, + 0.0007037408649921417, + 0.0008323667570948601, + 0.0008778477786108851, + 0.0008731247507967055, + 0.00044605336734093726, + 0.0006405597669072449, + 0.000793265295214951, + 0.00042320042848587036, + 0.0021567456424236298, + 0.004589466378092766, + 0.0033934477251023054, + 0.004379792604595423, + 0.00220939377322793, + 0.0028107869438827038, + 0.003345133038237691, + 0.0021506836637854576, + 0.0014424637192860246, + 0.0017174268141388893, + 0.0014836378395557404, + 0.001529510598629713, + 0.0014313196297734976, + 0.0013198271626606584, + 0.0017285612411797047, + 0.0014832643792033195, + 0.0024425543379038572, + 0.00176659831777215, + 0.007688484154641628, + 0.003881648648530245, + 0.0021594304125756025, + 0.006011940538883209, + 0.0024722949601709843, + 0.0030021299608051777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0052520157769322395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_tensor([3839, 3839])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_tensor([3839, 3839])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1dc3d5939c15bc616edcc6aa1a8278d5e127c523 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_tensor([3839, 3839])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning slow-motion, eye-level medium close-up, the cockpit static, focuses on a strikingly young woman, evoking Lin's ethereal presence at the controls of a sleek hovercraft. Her raven hair is pulled back into a tight, practical bun, highlighting her high cheekbones and the intense, unwavering focus in her dark eyes. She wears a fitted, dark gray aviator suit crafted from a matte, high-tech fabric with subtle, luminous cyan veins along the seams. Through the vast, curving canopy of trees before her, a rain-slicked, neon metropolis unfolds. In exquisite slow motion, individual raindrops leave shimmering, twisting paths across the transparent surface, each one capturing and refracting the city's vibrant light. Towering, multi-layered holographic advertisements for unrecognizable products shift and flicker, bathed in a vibrant wash of red, electric blue, and emerald green light, play across her serene face in the cockpit. Her gloved hands resting steadily on an illuminated control yoke, her fingers making subtle, deliberate adjustments as she expertly maneuvers the aircraft through dense streams of aerial traffic, weaving in and out of other vehicles with a calm, fluid grace that creates an air of tranquil focus amidst the overwhelming chaos of the futuristic city.", + "seed": 16052, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_3839_seed_16052_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.00965487714856863, + "vlm_logprob": -0.00965487714856863, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.00965487714856863, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008738137781620026 + ], + "opposite_logprob_tokens": [ + -5.508738040924072 + ], + "logprob_reward": -0.008738137781620026, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008738137781620026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0867776945233345, + 0.251621812582016, + 0.22066117823123932, + 0.021826347336173058, + 0.023396451026201248, + 0.10581915825605392, + 0.05260638892650604, + 0.09123679250478745, + 0.09481336921453476, + 0.0512407086789608 + ], + "credit_map": [ + 0.015137985348701477, + 0.02344871312379837, + 0.006533466279506683, + 0.005204560700803995, + 0.013135800138115883, + 0.011153348721563816, + 0.006699079647660255, + 0.005464746151119471, + 0.019258175045251846, + 0.08152713626623154, + 0.012963351793587208, + 0.016085678711533546, + 0.014337558299303055, + 0.013900717720389366, + 0.014934075064957142, + 0.07861515134572983, + 0.0189029723405838, + 0.03564947098493576, + 0.010513847693800926, + 0.015446397475898266, + 0.011121256276965141, + 0.011542174965143204, + 0.0888897255063057, + 0.028595328330993652, + 0.0020749473478645086, + 0.0029848639387637377, + 0.0023031397722661495, + 0.0026607296895235777, + 0.0027184283826500177, + 0.003857327625155449, + 0.0026244628243148327, + 0.0026024465914815664, + 0.0026801780331879854, + 0.0022480194456875324, + 0.002204054268077016, + 0.0026419279165565968, + 0.004376953933387995, + 0.003426755778491497, + 0.003211706643924117, + 0.0026068552397191525, + 0.012018034234642982, + 0.012552949599921703, + 0.009428484365344048, + 0.012009340338408947, + 0.023122191429138184, + 0.01658324897289276, + 0.009939709678292274, + 0.010165204294025898, + 0.00778536219149828, + 0.006850801408290863, + 0.0060652801766991615, + 0.005955037660896778, + 0.005487147253006697, + 0.006040872540324926, + 0.004808133002370596, + 0.00961375329643488, + 0.010093438439071178, + 0.014406388625502586, + 0.013323946855962276, + 0.01270727813243866, + 0.011469234712421894, + 0.0087351119145751, + 0.013242824003100395, + 0.007258569821715355, + 0.011401890777051449, + 0.011076484806835651, + 0.015669861808419228, + 0.010810485109686852, + 0.011012591421604156, + 0.010421467013657093, + 0.012687892653048038, + 0.011732696555554867, + 0.006691578775644302, + 0.004959337878972292, + 0.005995775107294321, + 0.006833936553448439, + 0.010500206612050533, + 0.0070678419433534145, + 0.005352571606636047, + 0.0038394618313759565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.008738137781620026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082758195698261 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082758195698261, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082758195698261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05760784074664116, + 0.3675314486026764, + 0.11552631109952927, + 0.011766687966883183, + 0.02800937555730343, + 0.09357165545225143, + 0.05909696966409683, + 0.10743000358343124, + 0.07867161929607391, + 0.08078806847333908 + ], + "credit_map": [ + 0.007426605559885502, + 0.00848948024213314, + 0.005660162772983313, + 0.005495317280292511, + 0.0099057387560606, + 0.009025081992149353, + 0.005352555308490992, + 0.006252904888242483, + 0.02636568620800972, + 0.11735741794109344, + 0.01892125979065895, + 0.023787839338183403, + 0.019460253417491913, + 0.020210908725857735, + 0.021250015124678612, + 0.12017808854579926, + 0.008287187665700912, + 0.020735757425427437, + 0.007788907270878553, + 0.011749149300158024, + 0.007437633350491524, + 0.0074640256352722645, + 0.04146817326545715, + 0.010595488362014294, + 0.0012831533094868064, + 0.002010853961110115, + 0.0010010418482124805, + 0.0012981928884983063, + 0.0013766662450507283, + 0.0018366463482379913, + 0.0012680289801210165, + 0.0016921049682423472, + 0.002024528104811907, + 0.003192516975104809, + 0.0035480514634400606, + 0.004678150173276663, + 0.0025265358854085207, + 0.0032102863769978285, + 0.004425594117492437, + 0.004403713624924421, + 0.007476869970560074, + 0.008615141734480858, + 0.008793625049293041, + 0.01688162051141262, + 0.017885196954011917, + 0.014941133558750153, + 0.008285579271614552, + 0.010692497715353966, + 0.00425774184986949, + 0.009297285228967667, + 0.007937473244965076, + 0.010778389871120453, + 0.006306534633040428, + 0.006481111515313387, + 0.005402064882218838, + 0.008636378683149815, + 0.008031070232391357, + 0.012352382764220238, + 0.01830379106104374, + 0.013002824038267136, + 0.010824928060173988, + 0.01240670308470726, + 0.01615877076983452, + 0.016349541023373604, + 0.007884197868406773, + 0.008381593972444534, + 0.013305370695888996, + 0.008706682361662388, + 0.01012935396283865, + 0.010203930549323559, + 0.010502196848392487, + 0.009558305144309998, + 0.016966305673122406, + 0.010956663638353348, + 0.008668794296681881, + 0.007946047000586987, + 0.00844753347337246, + 0.005774904042482376, + 0.007856780663132668, + 0.014171041548252106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.011082758195698261 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950355436652899 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950355436652899, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950355436652899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05050328001379967, + 0.35039564967155457, + 0.060763243585824966, + 0.041670870035886765, + 0.02338597923517227, + 0.09372968971729279, + 0.15717476606369019, + 0.116790771484375, + 0.06468450278043747, + 0.04090120643377304 + ], + "credit_map": [ + 0.009554686956107616, + 0.01024052407592535, + 0.003919908311218023, + 0.0035214920062571764, + 0.007404186297208071, + 0.00672915717586875, + 0.004090797621756792, + 0.00504252640530467, + 0.02745949476957321, + 0.11731824278831482, + 0.017788464203476906, + 0.020903553813695908, + 0.020201364532113075, + 0.019368737936019897, + 0.01995547115802765, + 0.1074003353714943, + 0.0042226966470479965, + 0.011863404884934425, + 0.004595258738845587, + 0.006005038972944021, + 0.004980660043656826, + 0.00451258011162281, + 0.018288766965270042, + 0.006294843275099993, + 0.00379046774469316, + 0.004959075711667538, + 0.005022600758820772, + 0.00641902070492506, + 0.0050873178988695145, + 0.008995459415018559, + 0.003583110636100173, + 0.0038138150703161955, + 0.0022316998802125454, + 0.0026181319262832403, + 0.0027347065042704344, + 0.003296759445220232, + 0.0021418901160359383, + 0.0028081827331334352, + 0.0032167271710932255, + 0.004337882623076439, + 0.010799924843013287, + 0.012953192926943302, + 0.010603963397443295, + 0.02136012725532055, + 0.009673389606177807, + 0.012759403325617313, + 0.006986487656831741, + 0.008593193255364895, + 0.013199830427765846, + 0.026877036318182945, + 0.024168476462364197, + 0.023854343220591545, + 0.018031854182481766, + 0.018008651211857796, + 0.012112783268094063, + 0.020921774208545685, + 0.017313677817583084, + 0.018913282081484795, + 0.017634235322475433, + 0.016118591651320457, + 0.014316956512629986, + 0.007967331446707249, + 0.01049272995442152, + 0.014033963903784752, + 0.008326269686222076, + 0.006716188043355942, + 0.011162664741277695, + 0.008199658244848251, + 0.007603774778544903, + 0.006969969253987074, + 0.006696039345115423, + 0.009009942412376404, + 0.005720003042370081, + 0.007904191501438618, + 0.006800548173487186, + 0.005435660947114229, + 0.004333017859607935, + 0.0029753795824944973, + 0.0031922892667353153, + 0.004540110472589731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.005950355436652899 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019369428046047688 + ], + "opposite_logprob_tokens": [ + -6.251936912536621 + ], + "logprob_reward": -0.0019369428046047688, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019369428046047688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047175414860248566, + 0.4902541935443878, + 0.2590477168560028, + 0.01729394868016243, + 0.013971666805446148, + 0.03565322607755661, + 0.017099790275096893, + 0.049083855003118515, + 0.038251202553510666, + 0.03216898813843727 + ], + "credit_map": [ + 0.005199809558689594, + 0.009607208892703056, + 0.0050280154682695866, + 0.004575492814183235, + 0.006637423299252987, + 0.006930110044777393, + 0.0047632246278226376, + 0.0044341301545500755, + 0.03653735667467117, + 0.1604945957660675, + 0.024245601147413254, + 0.028811337426304817, + 0.027210842818021774, + 0.02777797169983387, + 0.02851073630154133, + 0.15666575729846954, + 0.015619738027453423, + 0.06274409592151642, + 0.01262906938791275, + 0.018938295543193817, + 0.01838936284184456, + 0.014643133617937565, + 0.09300196915864944, + 0.023082032799720764, + 0.0027411850169301033, + 0.002149979816749692, + 0.0017066318541765213, + 0.0016857613809406757, + 0.002122798003256321, + 0.003823470091447234, + 0.0012766654836013913, + 0.001787457149475813, + 0.002314652083441615, + 0.0017461786046624184, + 0.0009388166363351047, + 0.0017506529111415148, + 0.002360687591135502, + 0.002028014278039336, + 0.0013862292980775237, + 0.0014464345294982195, + 0.003916864283382893, + 0.003760142717510462, + 0.0033885925076901913, + 0.008574098348617554, + 0.0051197814755141735, + 0.005125958472490311, + 0.0024222468491643667, + 0.0033455416560173035, + 0.0013115443289279938, + 0.002172105247154832, + 0.0023745994549244642, + 0.0027713256422430277, + 0.0020814226008951664, + 0.002426156075671315, + 0.0020017537754029036, + 0.0019608838483691216, + 0.00602928688749671, + 0.004387505818158388, + 0.003636300330981612, + 0.004726356361061335, + 0.005689957644790411, + 0.010286766104400158, + 0.006131253205239773, + 0.00819642748683691, + 0.0039007305167615414, + 0.004449082538485527, + 0.005792963784188032, + 0.004040381871163845, + 0.004620562307536602, + 0.006549842189997435, + 0.004547901451587677, + 0.004349742084741592, + 0.00360147375613451, + 0.003516183467581868, + 0.0036072921939194202, + 0.002878618659451604, + 0.00626919511705637, + 0.004097796976566315, + 0.0036253072321414948, + 0.004573119804263115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.0019369428046047688 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020566191524267197 + ], + "opposite_logprob_tokens": [ + -3.89556622505188 + ], + "logprob_reward": -0.020566191524267197, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020566191524267197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04480970278382301, + 0.5143346786499023, + 0.09419354051351547, + 0.029413500800728798, + 0.060471318662166595, + 0.07635568082332611, + 0.0548827201128006, + 0.040234193205833435, + 0.056244999170303345, + 0.029059654101729393 + ], + "credit_map": [ + 0.004577518906444311, + 0.006425050552934408, + 0.004094709642231464, + 0.00688856840133667, + 0.005367991980165243, + 0.004584621638059616, + 0.007228686474263668, + 0.0056425537914037704, + 0.03718622028827667, + 0.16805870831012726, + 0.02596978098154068, + 0.030778953805565834, + 0.02778322994709015, + 0.027280215173959732, + 0.029741855338215828, + 0.1675357073545456, + 0.009579155594110489, + 0.008035524748265743, + 0.0055496571585536, + 0.008713328279554844, + 0.00749750854447484, + 0.005459274165332317, + 0.030408242717385292, + 0.018950842320919037, + 0.00303635373711586, + 0.004307533614337444, + 0.0027766884304583073, + 0.004004159010946751, + 0.003688072320073843, + 0.004937427584081888, + 0.003425115719437599, + 0.003238149220123887, + 0.004313685931265354, + 0.006926917005330324, + 0.007593326736241579, + 0.007547932211309671, + 0.006444793194532394, + 0.008902337402105331, + 0.009058605879545212, + 0.009683718904852867, + 0.010031750425696373, + 0.008544839918613434, + 0.006346695125102997, + 0.015332172624766827, + 0.009725258685648441, + 0.01130784023553133, + 0.00723856408149004, + 0.007828551344573498, + 0.003417785745114088, + 0.007119925692677498, + 0.0052622463554143906, + 0.007114900276064873, + 0.0061809285543859005, + 0.007980702444911003, + 0.005908247549086809, + 0.011897976510226727, + 0.006914899684488773, + 0.003957110457122326, + 0.004295377992093563, + 0.004755586851388216, + 0.004504712764173746, + 0.005068233702331781, + 0.00623773830011487, + 0.004500528797507286, + 0.007699277251958847, + 0.005841236095875502, + 0.0073640719056129456, + 0.00438528461381793, + 0.010263917967677116, + 0.009541895240545273, + 0.005940163508057594, + 0.0052091507241129875, + 0.0034019320737570524, + 0.004774809814989567, + 0.002625132678076625, + 0.002964862622320652, + 0.0051970588974654675, + 0.00253767310641706, + 0.0031004741322249174, + 0.004457707982510328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.020566191524267197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion, eye-level medium close-up, the cockpit static, focuses on a strikingly young woman, evoking Lin's ethereal presence at the controls of a sleek hovercraft. Her raven hair is pulled back into a tight, practical bun, highlighting her high cheekbones and the intense, unwavering focus in her dark eyes. She wears a fitted, dark gray aviator suit crafted from a matte, high-tech fabric with subtle, luminous cyan veins along the seams. Through the vast, curving canopy of trees before her, a rain-slicked, neon metropolis unfolds. In exquisite slow motion, individual raindrops leave shimmering, twisting paths across the transparent surface, each one capturing and refracting the city's vibrant light. Towering, multi-layered holographic advertisements for unrecognizable products shift and flicker, bathed in a vibrant wash of red, electric blue, and emerald green light, play across her serene face in the cockpit. Her gloved hands resting steadily on an illuminated control yoke, her fingers making subtle, deliberate adjustments as she expertly maneuvers the aircraft through dense streams of aerial traffic, weaving in and out of other vehicles with a calm, fluid grace that creates an air of tranquil focus amidst the overwhelming chaos of the futuristic city.", + "seed": 16053, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_3839_seed_16053_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013480723928660154, + "vlm_logprob": -0.013480723928660154, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.013480723928660154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008264627307653427 + ], + "opposite_logprob_tokens": [ + -5.633264541625977 + ], + "logprob_reward": -0.008264627307653427, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.008264627307653427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10241477936506271, + 0.076436348259449, + 0.11065185070037842, + 0.05981504172086716, + 0.1427251100540161, + 0.0680382177233696, + 0.12013890594244003, + 0.08906373381614685, + 0.15194018185138702, + 0.0787758082151413 + ], + "credit_map": [ + 0.02238604985177517, + 0.013501450419425964, + 0.009400364011526108, + 0.0067245797254145145, + 0.015631703659892082, + 0.012367162853479385, + 0.012897715903818607, + 0.009505749680101871, + 0.020518852397799492, + 0.011944490484893322, + 0.010488420724868774, + 0.0069098989479243755, + 0.010082356631755829, + 0.006622848100960255, + 0.005290558096021414, + 0.0045789312571287155, + 0.01600801944732666, + 0.010301332920789719, + 0.016983551904559135, + 0.012167677283287048, + 0.01860935054719448, + 0.01180209033191204, + 0.013380802236497402, + 0.011399030685424805, + 0.008820826187729836, + 0.00896034948527813, + 0.0074154408648610115, + 0.006890097167342901, + 0.009247487410902977, + 0.005090712569653988, + 0.008118306286633015, + 0.005271821282804012, + 0.01806880719959736, + 0.0182145144790411, + 0.014199691824615002, + 0.015447655692696571, + 0.01493607833981514, + 0.010892636142671108, + 0.0321078784763813, + 0.018857838585972786, + 0.01136354822665453, + 0.007305012084543705, + 0.01048758439719677, + 0.005166543181985617, + 0.007214840035885572, + 0.008226307108998299, + 0.010206744074821472, + 0.00806763768196106, + 0.024201909080147743, + 0.012922440655529499, + 0.011780568398535252, + 0.009863872081041336, + 0.015024727210402489, + 0.0109644066542387, + 0.017219822853803635, + 0.018161166459321976, + 0.017869293689727783, + 0.01751120202243328, + 0.012103588320314884, + 0.009851707145571709, + 0.008968101814389229, + 0.006682653911411762, + 0.008353722281754017, + 0.0077234604395926, + 0.024451464414596558, + 0.015814298763871193, + 0.01599830761551857, + 0.022756019607186317, + 0.017364371567964554, + 0.015342853032052517, + 0.021027427166700363, + 0.019185442477464676, + 0.008641118183732033, + 0.010176540352404118, + 0.012037869542837143, + 0.009957083500921726, + 0.006785365287214518, + 0.010601967573165894, + 0.010870743542909622, + 0.009705116041004658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.008264627307653427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087591759860516 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087591759860516, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087591759860516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10567008703947067, + 0.09009943157434464, + 0.0932392105460167, + 0.10276710987091064, + 0.2483632117509842, + 0.04055122658610344, + 0.047429781407117844, + 0.061064526438713074, + 0.07918985933065414, + 0.1316254884004593 + ], + "credit_map": [ + 0.021973293274641037, + 0.015755770727992058, + 0.011117855086922646, + 0.010797940194606781, + 0.01403924822807312, + 0.009085528552532196, + 0.011662276461720467, + 0.011238177306950092, + 0.015133795328438282, + 0.010537194088101387, + 0.012834925204515457, + 0.010241410695016384, + 0.012662312015891075, + 0.008226141333580017, + 0.011508502066135406, + 0.0089551517739892, + 0.015909751877188683, + 0.008814981207251549, + 0.009947234764695168, + 0.010328739881515503, + 0.009719297289848328, + 0.008782830089330673, + 0.012417955324053764, + 0.017318429425358772, + 0.020593516528606415, + 0.014737498015165329, + 0.013994699344038963, + 0.011796144768595695, + 0.008815472014248371, + 0.009079121053218842, + 0.012765235267579556, + 0.010985433124005795, + 0.02253260836005211, + 0.019663618877530098, + 0.027195416390895844, + 0.033220961689949036, + 0.0286887064576149, + 0.026293132454156876, + 0.035778630524873734, + 0.05499015003442764, + 0.004405138548463583, + 0.005036892835050821, + 0.005922448355704546, + 0.004066043999046087, + 0.0036749374121427536, + 0.005085299722850323, + 0.007595991250127554, + 0.004764474928379059, + 0.013662319630384445, + 0.009255664423108101, + 0.0054914215579628944, + 0.0026965057477355003, + 0.0042222896590828896, + 0.003538839053362608, + 0.003933639265596867, + 0.004629103932529688, + 0.007684965152293444, + 0.009804238565266132, + 0.010624287649989128, + 0.007133458741009235, + 0.008868266828358173, + 0.004569411743432283, + 0.0068044220097362995, + 0.005575480870902538, + 0.013621796853840351, + 0.00842504296451807, + 0.010166074149310589, + 0.011613053269684315, + 0.008022603578865528, + 0.007621073629707098, + 0.010748710483312607, + 0.008971510455012321, + 0.009843559004366398, + 0.017011608928442, + 0.018426908180117607, + 0.025928039103746414, + 0.014494312927126884, + 0.011084910482168198, + 0.01471764873713255, + 0.020118512213230133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.011087591759860516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076340362429618835 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.0076340362429618835, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076340362429618835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12182874977588654, + 0.04136532172560692, + 0.20488253235816956, + 0.16772593557834625, + 0.06869103759527206, + 0.08657149225473404, + 0.054134536534547806, + 0.0720318853855133, + 0.08581125736236572, + 0.09695727378129959 + ], + "credit_map": [ + 0.019606370478868484, + 0.014580667950212955, + 0.015595301054418087, + 0.014312393963336945, + 0.020511949434876442, + 0.009812514297664165, + 0.016097337007522583, + 0.011312205344438553, + 0.0061716558411717415, + 0.0058435010723769665, + 0.006331457756459713, + 0.004137525334954262, + 0.005980621557682753, + 0.004231030121445656, + 0.004544970579445362, + 0.004124555271118879, + 0.02531849965453148, + 0.018432488664984703, + 0.024813709780573845, + 0.019711263477802277, + 0.035622164607048035, + 0.026411589235067368, + 0.027091672644019127, + 0.02748112939298153, + 0.02933596633374691, + 0.025335801765322685, + 0.019758759066462517, + 0.024577679112553596, + 0.014184778556227684, + 0.012830586172640324, + 0.024183684960007668, + 0.017518682405352592, + 0.005918137729167938, + 0.009305316023528576, + 0.007544855587184429, + 0.007458479609340429, + 0.007599019445478916, + 0.005905297119170427, + 0.015640996396541595, + 0.009318933822214603, + 0.014138546772301197, + 0.011169202625751495, + 0.010205487720668316, + 0.006567058619111776, + 0.008246282115578651, + 0.01154891774058342, + 0.01436780858784914, + 0.010328191332519054, + 0.014694267883896828, + 0.008643453940749168, + 0.0065620774403214455, + 0.0043168216943740845, + 0.0050711180083453655, + 0.003855867078527808, + 0.004687816835939884, + 0.006303118076175451, + 0.010987847112119198, + 0.010512328706681728, + 0.012572663836181164, + 0.008839290589094162, + 0.00968251284211874, + 0.005837348755449057, + 0.006462577730417252, + 0.007137322332710028, + 0.011249548755586147, + 0.009901938959956169, + 0.0113877197727561, + 0.01482462789863348, + 0.009203736670315266, + 0.006689308676868677, + 0.011385331861674786, + 0.011169042438268661, + 0.008774780668318272, + 0.014770696870982647, + 0.01207300927489996, + 0.01586572267115116, + 0.014126872643828392, + 0.0070673092268407345, + 0.012538239359855652, + 0.01174064353108406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0076340362429618835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732518319040537 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732518319040537, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732518319040537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06592097133398056, + 0.040725383907556534, + 0.08414734154939651, + 0.03853270411491394, + 0.08126793801784515, + 0.04192018136382103, + 0.1271159052848816, + 0.1564301699399948, + 0.2691285312175751, + 0.09481098502874374 + ], + "credit_map": [ + 0.013240689411759377, + 0.009008615277707577, + 0.011011600494384766, + 0.0056238737888634205, + 0.009970568120479584, + 0.0048698196187615395, + 0.007029336411505938, + 0.005166460294276476, + 0.0079916687682271, + 0.005172237288206816, + 0.005015251692384481, + 0.004664885811507702, + 0.007696983404457569, + 0.00419587129727006, + 0.0035035398323088884, + 0.002484940690919757, + 0.01811789534986019, + 0.009841002523899078, + 0.015586821362376213, + 0.009339402429759502, + 0.00923379696905613, + 0.005041858181357384, + 0.009524618275463581, + 0.00746193528175354, + 0.0054728747345507145, + 0.004857936874032021, + 0.005965774413198233, + 0.005185897462069988, + 0.003482033032923937, + 0.0029549424070864916, + 0.006015080027282238, + 0.004598161205649376, + 0.009571311995387077, + 0.012493978254497051, + 0.008754357695579529, + 0.0061919670552015305, + 0.008199420757591724, + 0.006402974016964436, + 0.02147323079407215, + 0.008180689066648483, + 0.004651072900742292, + 0.004973728209733963, + 0.006098998710513115, + 0.003992612939327955, + 0.004703255835920572, + 0.0060208686627447605, + 0.005633431952446699, + 0.005846204701811075, + 0.018771303817629814, + 0.012367638759315014, + 0.011154776439070702, + 0.010427269153296947, + 0.01255582645535469, + 0.011157558299601078, + 0.023109260946512222, + 0.0275722648948431, + 0.02638043835759163, + 0.026577981188893318, + 0.022881977260112762, + 0.01622321642935276, + 0.017252521589398384, + 0.015025272965431213, + 0.014523480087518692, + 0.017565257847309113, + 0.03717601299285889, + 0.02596111223101616, + 0.03801379352807999, + 0.04357070475816727, + 0.030280087143182755, + 0.021987657994031906, + 0.03488900512456894, + 0.0372500941157341, + 0.009605108760297298, + 0.012109634466469288, + 0.012655718252062798, + 0.012955392710864544, + 0.008232461288571358, + 0.01476549357175827, + 0.01473256479948759, + 0.009754602797329426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.006732518319040537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368484601378441 + ], + "opposite_logprob_tokens": [ + -3.408684730529785 + ], + "logprob_reward": -0.03368484601378441, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368484601378441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10177824646234512, + 0.09571409970521927, + 0.12066204845905304, + 0.08184580504894257, + 0.0394405722618103, + 0.0685773566365242, + 0.1498192995786667, + 0.11705850809812546, + 0.12449648231267929, + 0.10060761123895645 + ], + "credit_map": [ + 0.018476583063602448, + 0.014061791822314262, + 0.013919754885137081, + 0.008142167702317238, + 0.01471063680946827, + 0.00915934145450592, + 0.012314075604081154, + 0.010993883945047855, + 0.011507423594594002, + 0.012089591473340988, + 0.013211289420723915, + 0.01634266786277294, + 0.010538534261286259, + 0.01080696564167738, + 0.012476232834160328, + 0.008741387166082859, + 0.017237892374396324, + 0.010871726088225842, + 0.017366189509630203, + 0.015453019179403782, + 0.014601366594433784, + 0.009218166582286358, + 0.016163857653737068, + 0.019749827682971954, + 0.018491677939891815, + 0.011142094619572163, + 0.011032558046281338, + 0.009699215181171894, + 0.007482314016669989, + 0.010879489593207836, + 0.006572566460818052, + 0.00654588220641017, + 0.005769370123744011, + 0.0064346035942435265, + 0.00605790177360177, + 0.004919932223856449, + 0.00426100566983223, + 0.003398702945560217, + 0.004707404877990484, + 0.0038916459307074547, + 0.009919542819261551, + 0.008118151687085629, + 0.010141024366021156, + 0.007118473760783672, + 0.006098337005823851, + 0.00645338324829936, + 0.012432759627699852, + 0.008295678533613682, + 0.043961793184280396, + 0.02089204080402851, + 0.01833314448595047, + 0.010724236257374287, + 0.014456317760050297, + 0.011751535348594189, + 0.013499557040631771, + 0.01620067097246647, + 0.015017849393188953, + 0.019241511821746826, + 0.017117077484726906, + 0.015684502199292183, + 0.014395521953701973, + 0.013561967760324478, + 0.011384274810552597, + 0.010655793361365795, + 0.013576094061136246, + 0.015955591574311256, + 0.016135748475790024, + 0.018596500158309937, + 0.01440084632486105, + 0.011893074959516525, + 0.01756107434630394, + 0.016377544030547142, + 0.011432472616434097, + 0.019853852689266205, + 0.009454804472625256, + 0.013237149454653263, + 0.011589032597839832, + 0.012305454351007938, + 0.013082137331366539, + 0.009652697481215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.03368484601378441 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion, eye-level medium close-up, the cockpit static, focuses on a strikingly young woman, evoking Lin's ethereal presence at the controls of a sleek hovercraft. Her raven hair is pulled back into a tight, practical bun, highlighting her high cheekbones and the intense, unwavering focus in her dark eyes. She wears a fitted, dark gray aviator suit crafted from a matte, high-tech fabric with subtle, luminous cyan veins along the seams. Through the vast, curving canopy of trees before her, a rain-slicked, neon metropolis unfolds. In exquisite slow motion, individual raindrops leave shimmering, twisting paths across the transparent surface, each one capturing and refracting the city's vibrant light. Towering, multi-layered holographic advertisements for unrecognizable products shift and flicker, bathed in a vibrant wash of red, electric blue, and emerald green light, play across her serene face in the cockpit. Her gloved hands resting steadily on an illuminated control yoke, her fingers making subtle, deliberate adjustments as she expertly maneuvers the aircraft through dense streams of aerial traffic, weaving in and out of other vehicles with a calm, fluid grace that creates an air of tranquil focus amidst the overwhelming chaos of the futuristic city.", + "seed": 16054, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_3839_seed_16054_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016088185273110865, + "vlm_logprob": -0.016088185273110865, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.016088185273110865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012697093188762665 + ], + "opposite_logprob_tokens": [ + -5.137697219848633 + ], + "logprob_reward": -0.012697093188762665, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.012697093188762665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14720727503299713, + 0.08787032961845398, + 0.09314600378274918, + 0.08263160288333893, + 0.0640232264995575, + 0.06872691214084625, + 0.16329990327358246, + 0.08109581470489502, + 0.1079309955239296, + 0.10406792908906937 + ], + "credit_map": [ + 0.014851607382297516, + 0.017403870820999146, + 0.014837728813290596, + 0.021665552631020546, + 0.027794834226369858, + 0.024893824011087418, + 0.014419205486774445, + 0.011340665630996227, + 0.009772182442247868, + 0.01176996249705553, + 0.0048209745436906815, + 0.009959753602743149, + 0.01818460039794445, + 0.012766796164214611, + 0.012655656784772873, + 0.007940406911075115, + 0.011912497691810131, + 0.022715287283062935, + 0.00508003169670701, + 0.012923995032906532, + 0.013643804006278515, + 0.011557773686945438, + 0.008155287243425846, + 0.007157338783144951, + 0.009896857663989067, + 0.016145361587405205, + 0.0058784787543118, + 0.006727555301040411, + 0.010962643660604954, + 0.013046728447079659, + 0.012702708132565022, + 0.007271283771842718, + 0.006900451611727476, + 0.010426804423332214, + 0.00525252241641283, + 0.007627862971276045, + 0.008025594986975193, + 0.009109104983508587, + 0.009941884316504002, + 0.00673900218680501, + 0.010852297767996788, + 0.008775089867413044, + 0.005033526569604874, + 0.007302470970898867, + 0.010305003263056278, + 0.012007617391645908, + 0.008496520109474659, + 0.005954389460384846, + 0.020859966054558754, + 0.02460046112537384, + 0.01274221483618021, + 0.025424785912036896, + 0.019775161519646645, + 0.021374624222517014, + 0.020136574283242226, + 0.018386121839284897, + 0.008266419172286987, + 0.007980153895914555, + 0.010409586131572723, + 0.00912930816411972, + 0.01947164721786976, + 0.011017398908734322, + 0.00785953737795353, + 0.006961767096072435, + 0.009700775146484375, + 0.005842298269271851, + 0.009620380587875843, + 0.008896436542272568, + 0.02959470823407173, + 0.028191756457090378, + 0.007681221701204777, + 0.008403422310948372, + 0.014183797873556614, + 0.01575159654021263, + 0.00846922304481268, + 0.009593922644853592, + 0.021220708265900612, + 0.01498881634324789, + 0.013107528910040855, + 0.0067523461766541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.012697093188762665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296575993299484 + ], + "opposite_logprob_tokens": [ + -3.773296594619751 + ], + "logprob_reward": -0.023296575993299484, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296575993299484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13929840922355652, + 0.08321704715490341, + 0.10676180571317673, + 0.1176871508359909, + 0.18173320591449738, + 0.08129388839006424, + 0.05517720431089401, + 0.0717463344335556, + 0.06688856333494186, + 0.09619638323783875 + ], + "credit_map": [ + 0.016881372779607773, + 0.016600564122200012, + 0.011398016475141048, + 0.015793664380908012, + 0.02568618580698967, + 0.02308974415063858, + 0.01920115202665329, + 0.010647705756127834, + 0.008387350477278233, + 0.009513727389276028, + 0.005494088865816593, + 0.009902751073241234, + 0.021489540114998817, + 0.011950166895985603, + 0.00828184001147747, + 0.008197581395506859, + 0.01125954370945692, + 0.01850316673517227, + 0.01103504654020071, + 0.01616828516125679, + 0.01612084172666073, + 0.011319977231323719, + 0.012360782362520695, + 0.009994158521294594, + 0.013944405131042004, + 0.01783696562051773, + 0.00701345968991518, + 0.009243184700608253, + 0.01601477339863777, + 0.019031979143619537, + 0.02357872761785984, + 0.01102366205304861, + 0.01841040886938572, + 0.025634463876485825, + 0.02409503608942032, + 0.030501460656523705, + 0.019294533878564835, + 0.022397803142666817, + 0.021451883018016815, + 0.019947610795497894, + 0.00919061154127121, + 0.008061249740421772, + 0.007470102980732918, + 0.009912095032632351, + 0.01676391251385212, + 0.012403079308569431, + 0.011723184026777744, + 0.00576966255903244, + 0.007169562391936779, + 0.008085375651717186, + 0.004718896001577377, + 0.008187231607735157, + 0.00484249135479331, + 0.007677990943193436, + 0.005997291766107082, + 0.008498365059494972, + 0.006479115691035986, + 0.009914936497807503, + 0.010131136514246464, + 0.00786462053656578, + 0.012124177068471909, + 0.007075716275721788, + 0.01140749640762806, + 0.006749140098690987, + 0.006200409960001707, + 0.007055202964693308, + 0.006131346803158522, + 0.005817190743982792, + 0.014217832125723362, + 0.013134748674929142, + 0.009278558194637299, + 0.005053274799138308, + 0.009712000377476215, + 0.015387404710054398, + 0.0070613110437989235, + 0.009776351973414421, + 0.018937021493911743, + 0.015141298063099384, + 0.012209855020046234, + 0.007971145212650299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.023296575993299484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640403859317303 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640403859317303, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640403859317303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0961800217628479, + 0.10083191841840744, + 0.19898349046707153, + 0.05991848558187485, + 0.21349380910396576, + 0.09107085317373276, + 0.05879715085029602, + 0.061132751405239105, + 0.057434335350990295, + 0.062157291918992996 + ], + "credit_map": [ + 0.013044124469161034, + 0.010490546934306622, + 0.008962644264101982, + 0.011521783657371998, + 0.018411172553896904, + 0.014703516848385334, + 0.013325508683919907, + 0.005720708519220352, + 0.012026372365653515, + 0.012494685128331184, + 0.008209788240492344, + 0.013515165075659752, + 0.0164886936545372, + 0.012730170041322708, + 0.011449303478002548, + 0.013917719013988972, + 0.013271979056298733, + 0.03311808779835701, + 0.01192100066691637, + 0.016537075862288475, + 0.051464714109897614, + 0.04422583431005478, + 0.01684405468404293, + 0.011600698344409466, + 0.0061903526075184345, + 0.011613873764872551, + 0.005342636723071337, + 0.005978374741971493, + 0.007856475189328194, + 0.008515548892319202, + 0.007838078774511814, + 0.00658313138410449, + 0.020037628710269928, + 0.03486338630318642, + 0.026784194633364677, + 0.032069072127342224, + 0.02838967554271221, + 0.023237718269228935, + 0.021836858242750168, + 0.026275241747498512, + 0.01096794568002224, + 0.010317496955394745, + 0.0061411261558532715, + 0.008968785405158997, + 0.016202805563807487, + 0.013886775821447372, + 0.016805171966552734, + 0.007780721876770258, + 0.005455043166875839, + 0.007891034707427025, + 0.006970842834562063, + 0.007664138451218605, + 0.012213122099637985, + 0.0074089509434998035, + 0.005255519412457943, + 0.005938489455729723, + 0.005620953626930714, + 0.006819603964686394, + 0.005637123249471188, + 0.006233644671738148, + 0.016153061762452126, + 0.0067360419780015945, + 0.008688577450811863, + 0.005243734922260046, + 0.005938488990068436, + 0.0052631027065217495, + 0.006245722062885761, + 0.006615049205720425, + 0.010871294885873795, + 0.010329450480639935, + 0.006358722690492868, + 0.005812493152916431, + 0.0074990722350776196, + 0.008459916338324547, + 0.007454498205333948, + 0.008532042615115643, + 0.00914081186056137, + 0.008157946169376373, + 0.007989663630723953, + 0.004923332016915083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.008640403859317303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532057240605354 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532057240605354, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532057240605354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09064464271068573, + 0.18583837151527405, + 0.08496705442667007, + 0.19900697469711304, + 0.037445906549692154, + 0.11532706767320633, + 0.07207087427377701, + 0.12220914661884308, + 0.06082839518785477, + 0.03166155889630318 + ], + "credit_map": [ + 0.0085523696616292, + 0.011612039990723133, + 0.00794281717389822, + 0.011676797643303871, + 0.01665664277970791, + 0.013406312093138695, + 0.01192772202193737, + 0.008869947865605354, + 0.015006308443844318, + 0.016178885474801064, + 0.01369775366038084, + 0.025050751864910126, + 0.040219832211732864, + 0.026976630091667175, + 0.01953716203570366, + 0.029171044006943703, + 0.008576452732086182, + 0.011752753518521786, + 0.005218095611780882, + 0.0072935475036501884, + 0.021333446726202965, + 0.01837823912501335, + 0.006392044480890036, + 0.006022466812282801, + 0.022415801882743835, + 0.025866391137242317, + 0.012281105853617191, + 0.014706207439303398, + 0.02678472362458706, + 0.03546438366174698, + 0.039978452026844025, + 0.021509919315576553, + 0.007347931619733572, + 0.004555391147732735, + 0.00326324300840497, + 0.0040466804057359695, + 0.004630588926374912, + 0.006120974197983742, + 0.004862826783210039, + 0.002618273487314582, + 0.01310689840465784, + 0.007219512481242418, + 0.004659881349653006, + 0.007202861364930868, + 0.034240514039993286, + 0.02386249415576458, + 0.016157949343323708, + 0.00887695886194706, + 0.006731140892952681, + 0.00868304818868637, + 0.006371560040861368, + 0.007434909697622061, + 0.017388971522450447, + 0.011126660741865635, + 0.008321749046444893, + 0.0060128336772322655, + 0.010275879874825478, + 0.013703742995858192, + 0.01426222175359726, + 0.01234881579875946, + 0.025231793522834778, + 0.013005915097892284, + 0.022635478526353836, + 0.010745299980044365, + 0.002438145922496915, + 0.004271027632057667, + 0.0050311628729105, + 0.0048451111651957035, + 0.015845010057091713, + 0.017330404371023178, + 0.006366644520312548, + 0.0047008926048874855, + 0.0032269363291561604, + 0.004966143984347582, + 0.003299195086583495, + 0.003945271950215101, + 0.005299591459333897, + 0.005205220077186823, + 0.0034172229934483767, + 0.002301976550370455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.012532057240605354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023274796083569527 + ], + "opposite_logprob_tokens": [ + -3.7732748985290527 + ], + "logprob_reward": -0.023274796083569527, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023274796083569527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11389918625354767, + 0.11607453227043152, + 0.12505318224430084, + 0.03729121759533882, + 0.1727762520313263, + 0.03531668707728386, + 0.10315434634685516, + 0.12100828438997269, + 0.11189977079629898, + 0.06352651119232178 + ], + "credit_map": [ + 0.012766286730766296, + 0.013714499771595001, + 0.010501417331397533, + 0.013605461455881596, + 0.022505585104227066, + 0.02025914005935192, + 0.01370116975158453, + 0.006845636758953333, + 0.010042380541563034, + 0.010088888928294182, + 0.009723765775561333, + 0.01720895618200302, + 0.02108391560614109, + 0.015547145158052444, + 0.01188535988330841, + 0.0204941239207983, + 0.012422054074704647, + 0.020071113482117653, + 0.0067570749670267105, + 0.010449821129441261, + 0.030644144862890244, + 0.026970557868480682, + 0.010510137304663658, + 0.007228291127830744, + 0.004686783533543348, + 0.0031789811328053474, + 0.0025000255554914474, + 0.0033271661959588528, + 0.005696587730199099, + 0.006642315071076155, + 0.005947810597717762, + 0.005311550106853247, + 0.024780437350273132, + 0.016039546579122543, + 0.018252184614539146, + 0.02356022037565708, + 0.019164785742759705, + 0.025413010269403458, + 0.025453848764300346, + 0.02011222392320633, + 0.004633683245629072, + 0.004058525897562504, + 0.003887239145115018, + 0.004374402575194836, + 0.00599841121584177, + 0.004498202353715897, + 0.004180907737463713, + 0.0036853181663900614, + 0.011433912441134453, + 0.01610451191663742, + 0.009191600605845451, + 0.015571984462440014, + 0.013765309005975723, + 0.012370692566037178, + 0.011510412208735943, + 0.01320592686533928, + 0.011619951575994492, + 0.01593473181128502, + 0.01069696992635727, + 0.011993265710771084, + 0.036215513944625854, + 0.013571960851550102, + 0.013923346996307373, + 0.007052552420645952, + 0.007860134355723858, + 0.010842688381671906, + 0.017252814024686813, + 0.014769070781767368, + 0.011150074191391468, + 0.018927300348877907, + 0.017101053148508072, + 0.01399664580821991, + 0.005798795726150274, + 0.010200591757893562, + 0.006572835613042116, + 0.008690652437508106, + 0.010839005000889301, + 0.007632771506905556, + 0.008992858231067657, + 0.004799004178494215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.023274796083569527 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion, eye-level medium close-up, the cockpit static, focuses on a strikingly young woman, evoking Lin's ethereal presence at the controls of a sleek hovercraft. Her raven hair is pulled back into a tight, practical bun, highlighting her high cheekbones and the intense, unwavering focus in her dark eyes. She wears a fitted, dark gray aviator suit crafted from a matte, high-tech fabric with subtle, luminous cyan veins along the seams. Through the vast, curving canopy of trees before her, a rain-slicked, neon metropolis unfolds. In exquisite slow motion, individual raindrops leave shimmering, twisting paths across the transparent surface, each one capturing and refracting the city's vibrant light. Towering, multi-layered holographic advertisements for unrecognizable products shift and flicker, bathed in a vibrant wash of red, electric blue, and emerald green light, play across her serene face in the cockpit. Her gloved hands resting steadily on an illuminated control yoke, her fingers making subtle, deliberate adjustments as she expertly maneuvers the aircraft through dense streams of aerial traffic, weaving in and out of other vehicles with a calm, fluid grace that creates an air of tranquil focus amidst the overwhelming chaos of the futuristic city.", + "seed": 16055, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_7/index_3839_seed_16055_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014433877635747194, + "vlm_logprob": -0.014433877635747194, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.014433877635747194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013495286926627159 + ], + "opposite_logprob_tokens": [ + -5.013495445251465 + ], + "logprob_reward": -0.013495286926627159, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.013495286926627159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11489394307136536, + 0.09474615007638931, + 0.08541572093963623, + 0.07519283890724182, + 0.1273694932460785, + 0.12596341967582703, + 0.10155432671308517, + 0.09843327105045319, + 0.10303010046482086, + 0.07340069115161896 + ], + "credit_map": [ + 0.010066404938697815, + 0.01588514819741249, + 0.017007747665047646, + 0.02393208257853985, + 0.011811137199401855, + 0.01518916618078947, + 0.012290884740650654, + 0.008711379952728748, + 0.007664177566766739, + 0.006590776611119509, + 0.011064895428717136, + 0.007542329374700785, + 0.030734753236174583, + 0.013537753373384476, + 0.010436591692268848, + 0.0071748788468539715, + 0.009617680683732033, + 0.01236434280872345, + 0.008125055581331253, + 0.015314898453652859, + 0.009647569619119167, + 0.010244342498481274, + 0.009929045103490353, + 0.010172789916396141, + 0.007471832446753979, + 0.007574514020234346, + 0.0071898456662893295, + 0.006784993223845959, + 0.011054343543946743, + 0.02135438099503517, + 0.006558398250490427, + 0.007204531226307154, + 0.00828473549336195, + 0.009661960415542126, + 0.019487081095576286, + 0.010744238272309303, + 0.014914462342858315, + 0.013228204101324081, + 0.040682096034288406, + 0.010366733185946941, + 0.008864453062415123, + 0.013570102863013744, + 0.014425593428313732, + 0.016122786328196526, + 0.02523685432970524, + 0.029904378578066826, + 0.007640309631824493, + 0.010198956355452538, + 0.01095008384436369, + 0.014097681269049644, + 0.011618277989327908, + 0.01937386952340603, + 0.013336461037397385, + 0.00984654575586319, + 0.01094986405223608, + 0.011381548829376698, + 0.006665374618023634, + 0.012306482531130314, + 0.010602282360196114, + 0.013842731714248657, + 0.013083869591355324, + 0.009782684966921806, + 0.008261153474450111, + 0.02388869971036911, + 0.015025660395622253, + 0.010749365203082561, + 0.015156213194131851, + 0.018707558512687683, + 0.008745282888412476, + 0.008856750093400478, + 0.00980475265532732, + 0.015984533354640007, + 0.008578388020396233, + 0.007534128148108721, + 0.00915844738483429, + 0.007330162450671196, + 0.008889677934348583, + 0.008956345729529858, + 0.010224874131381512, + 0.012728673405945301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.013495286926627159 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020604729652404785 + ], + "opposite_logprob_tokens": [ + -3.8956046104431152 + ], + "logprob_reward": -0.020604729652404785, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020604729652404785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13327515125274658, + 0.0504041463136673, + 0.07727367430925369, + 0.06415463984012604, + 0.10631965100765228, + 0.09033729135990143, + 0.06977137923240662, + 0.18012665212154388, + 0.11257733404636383, + 0.11576005816459656 + ], + "credit_map": [ + 0.015535910613834858, + 0.01565931923687458, + 0.020410120487213135, + 0.016894957050681114, + 0.021832531318068504, + 0.01557688508182764, + 0.017156530171632767, + 0.010208901017904282, + 0.004181557800620794, + 0.004420253448188305, + 0.007076228968799114, + 0.004936862271279097, + 0.01368155051022768, + 0.005467015318572521, + 0.005323268473148346, + 0.005317404866218567, + 0.008124669082462788, + 0.006699197459965944, + 0.008136491291224957, + 0.007364163175225258, + 0.013041104190051556, + 0.011686250567436218, + 0.00978165678679943, + 0.01244014035910368, + 0.005184628069400787, + 0.00925799272954464, + 0.004166559781879187, + 0.005330651067197323, + 0.00897408090531826, + 0.017235249280929565, + 0.007650435436517, + 0.00635504862293601, + 0.014281797222793102, + 0.011680368334054947, + 0.012333877384662628, + 0.011503709480166435, + 0.015188138000667095, + 0.014380190521478653, + 0.01822780631482601, + 0.008723761886358261, + 0.006756753660738468, + 0.009518801234662533, + 0.00897317472845316, + 0.015618565492331982, + 0.012524531222879887, + 0.016009127721190453, + 0.008439113385975361, + 0.012497230432927608, + 0.009355425834655762, + 0.008328387513756752, + 0.008600663393735886, + 0.010896597988903522, + 0.01172145176678896, + 0.006099436432123184, + 0.006863070651888847, + 0.007906352169811726, + 0.015760431066155434, + 0.014913024380803108, + 0.015744637697935104, + 0.01581144705414772, + 0.04782431200146675, + 0.02231709100306034, + 0.015588338486850262, + 0.03216736763715744, + 0.015208140015602112, + 0.010052655823528767, + 0.01257113367319107, + 0.018059853464365005, + 0.012454574927687645, + 0.016788695007562637, + 0.01044802088290453, + 0.016994258388876915, + 0.010690730065107346, + 0.010236247442662716, + 0.01679150201380253, + 0.011871042661368847, + 0.011682831682264805, + 0.01472494751214981, + 0.013358650729060173, + 0.02640410326421261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.020604729652404785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637467235326767 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637467235326767, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637467235326767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09673196822404861, + 0.13789348304271698, + 0.13277368247509003, + 0.08032070845365524, + 0.059723857790231705, + 0.17875300347805023, + 0.09186427295207977, + 0.07736855745315552, + 0.07745109498500824, + 0.06711935997009277 + ], + "credit_map": [ + 0.011137374676764011, + 0.01200244389474392, + 0.012226277962327003, + 0.011560935527086258, + 0.008342869579792023, + 0.012467012740671635, + 0.01145100500434637, + 0.01754404976963997, + 0.011774372309446335, + 0.016584062948822975, + 0.015079737640917301, + 0.014464941807091236, + 0.04011983796954155, + 0.015357843600213528, + 0.013765969313681126, + 0.010746706277132034, + 0.01803099922835827, + 0.021558433771133423, + 0.01353657990694046, + 0.026792189106345177, + 0.016828352585434914, + 0.010674688965082169, + 0.013392943888902664, + 0.011959489434957504, + 0.006015501916408539, + 0.010030576027929783, + 0.005597042851150036, + 0.00626583443954587, + 0.012594982050359249, + 0.02582758292555809, + 0.005562314763665199, + 0.008426873944699764, + 0.0064278291538357735, + 0.009115452878177166, + 0.006671305280178785, + 0.009634207934141159, + 0.006659130100160837, + 0.006145073566585779, + 0.009054461494088173, + 0.006016397848725319, + 0.008300209417939186, + 0.010127830319106579, + 0.028238700702786446, + 0.01648375391960144, + 0.05140896886587143, + 0.04698583483695984, + 0.006849976256489754, + 0.010357734747231007, + 0.010898695327341557, + 0.01171059999614954, + 0.011565541848540306, + 0.013178638182580471, + 0.013549981638789177, + 0.009522677399218082, + 0.009091811254620552, + 0.01234633382409811, + 0.008684364147484303, + 0.011362473480403423, + 0.011430935934185982, + 0.009522955864667892, + 0.01147652417421341, + 0.0073833102360367775, + 0.007341179996728897, + 0.010166813619434834, + 0.011318214237689972, + 0.009672533720731735, + 0.01114338357001543, + 0.008912808261811733, + 0.007576545234769583, + 0.008265923708677292, + 0.0074775442481040955, + 0.013084134086966515, + 0.005648077465593815, + 0.008603252470493317, + 0.010128255933523178, + 0.008987998589873314, + 0.008791178464889526, + 0.006323270034044981, + 0.006816767156124115, + 0.01182055938988924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.007637467235326767 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040937201119959354 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040937201119959354, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040937201119959354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08442141115665436, + 0.05237485468387604, + 0.086225725710392, + 0.04991644248366356, + 0.04367239028215408, + 0.17335182428359985, + 0.08341033011674881, + 0.15588705241680145, + 0.2277124673128128, + 0.043027523905038834 + ], + "credit_map": [ + 0.009277785196900368, + 0.010557333938777447, + 0.011313415132462978, + 0.010197469964623451, + 0.008353331126272678, + 0.015408419072628021, + 0.012367076240479946, + 0.006946574430912733, + 0.003915941808372736, + 0.004257633350789547, + 0.004843943286687136, + 0.004518189467489719, + 0.017153728753328323, + 0.007037887815386057, + 0.0068643102422356606, + 0.003783216467127204, + 0.008374000899493694, + 0.008259397000074387, + 0.007871403358876705, + 0.010517627000808716, + 0.010612308979034424, + 0.018209513276815414, + 0.011971453204751015, + 0.010410018265247345, + 0.0051472922787070274, + 0.008638577535748482, + 0.0050955880433321, + 0.005014199297875166, + 0.005902726203203201, + 0.008281581103801727, + 0.006205427460372448, + 0.005631053354591131, + 0.004513318184763193, + 0.004150207620114088, + 0.004029363393783569, + 0.0030243350192904472, + 0.008536876179277897, + 0.00781190674751997, + 0.007444275077432394, + 0.0041621048003435135, + 0.007260594982653856, + 0.007246764376759529, + 0.023414723575115204, + 0.015685638412833214, + 0.05050282180309296, + 0.05227496102452278, + 0.006782742217183113, + 0.010183577425777912, + 0.010396701283752918, + 0.0068760523572564125, + 0.010234666988253593, + 0.01142016053199768, + 0.015599694103002548, + 0.009232738986611366, + 0.010725319385528564, + 0.008924996480345726, + 0.01498444378376007, + 0.01638066954910755, + 0.011291448026895523, + 0.012617725878953934, + 0.04128556698560715, + 0.024783331900835037, + 0.016694804653525352, + 0.017849063500761986, + 0.03391657769680023, + 0.018843870609998703, + 0.01837320439517498, + 0.026760155335068703, + 0.026293959468603134, + 0.05319800600409508, + 0.020263871178030968, + 0.030062824487686157, + 0.003457740182057023, + 0.0037113118451088667, + 0.004749139305204153, + 0.005243545398116112, + 0.007168064825236797, + 0.005369654390960932, + 0.004717835690826178, + 0.008610231801867485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.0040937201119959354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium close-up of a young woman in a cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray aviator suit with luminous cyan veins along the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair pulled back into a tight bun, and are her hands resting on an illuminated control yoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rain-slicked neon metropolis with multi-layered holographic advertisements in red, electric blue, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in slow motion, showing raindrops leaving shimmering paths on the transparent canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338184252381325 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338184252381325, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338184252381325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10444560647010803, + 0.10712923109531403, + 0.06605928391218185, + 0.11456230282783508, + 0.04161515831947327, + 0.385351300239563, + 0.054800525307655334, + 0.04068413004279137, + 0.04731076583266258, + 0.03804166987538338 + ], + "credit_map": [ + 0.012819145806133747, + 0.012606346048414707, + 0.011755846440792084, + 0.01298979390412569, + 0.014655375853180885, + 0.01777353696525097, + 0.014701913110911846, + 0.007143654394894838, + 0.009895546361804008, + 0.012536413036286831, + 0.01490834355354309, + 0.013248773291707039, + 0.021142521873116493, + 0.013565372675657272, + 0.012314737774431705, + 0.009517529048025608, + 0.00888273399323225, + 0.007895680144429207, + 0.007142843212932348, + 0.010054790414869785, + 0.009862037375569344, + 0.008268996141850948, + 0.008166381157934666, + 0.005785827524960041, + 0.01401856541633606, + 0.02079676277935505, + 0.01431073434650898, + 0.012108263559639454, + 0.013839543797075748, + 0.010841124691069126, + 0.01380107644945383, + 0.014846237376332283, + 0.0033781863749027252, + 0.005177963990718126, + 0.004706304520368576, + 0.003783704712986946, + 0.005663004238158464, + 0.004683246836066246, + 0.01008379366248846, + 0.0041389502584934235, + 0.01878235675394535, + 0.02170286700129509, + 0.07627154886722565, + 0.040049392729997635, + 0.10405351966619492, + 0.0908324271440506, + 0.011124949902296066, + 0.022534240037202835, + 0.006711007561534643, + 0.006017592269927263, + 0.007037164177745581, + 0.006376080680638552, + 0.00783470831811428, + 0.006411713548004627, + 0.005588619504123926, + 0.008823643438518047, + 0.0033130256924778223, + 0.0030738788191229105, + 0.004269818775355816, + 0.0033629685640335083, + 0.007444561924785376, + 0.005032961256802082, + 0.004266408272087574, + 0.009920509532094002, + 0.005570678506046534, + 0.005062646698206663, + 0.006566382944583893, + 0.009601349011063576, + 0.004085861146450043, + 0.00543572660535574, + 0.004655859898775816, + 0.0063322605565190315, + 0.0037677993532270193, + 0.0032934544142335653, + 0.005112242419272661, + 0.004920853301882744, + 0.004928461741656065, + 0.003985520452260971, + 0.004234150517731905, + 0.0077991848811507225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.026338184252381325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_8/index_tensor([6800, 6800])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_8/index_tensor([6800, 6800])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8a94b4332c00efc5837445c595be70d836fc8af2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_8/index_tensor([6800, 6800])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tense, fast-cut shot, shot from the first-person point of view (POV) of a surveillance camera, fixed high on a building's facade, captures a deserted alley bathed in the glaring afternoon sun from a slightly bird's-eye perspective. The wide-angle lens's slight fisheye distortion distorts the edges of the image. Suddenly, a rough, red clay brick, its surface covered in weathered cracks and mottled mortar residue, spins and hurtles through the air, whirring directly toward the camera. In the final moments before impact, the brick zooms dramatically in, its angular details crystal clear. Then, with incredible force, it slams into the camera lens, instantly shattering the glass in a dense, radiating spiderweb of cracks. The entire frame, fragmented by the shattered glass prism, fades to black in a violent flash and a blinding white light, creating an abrupt and impactful aesthetic of violence.", + "seed": 18241, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_8/index_6800_seed_18241_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06882236918900161, + "vlm_logprob": -0.06882236918900161, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.06882236918900161, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13432396948337555 + ], + "opposite_logprob_tokens": [ + -2.134324073791504 + ], + "logprob_reward": -0.13432396948337555, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13432396948337555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1855318397283554, + 0.07095398008823395, + 0.16096121072769165, + 0.1308506727218628, + 0.14040791988372803, + 0.11727156490087509, + 0.07906164228916168, + 0.05555995926260948, + 0.0227260310202837, + 0.03667519614100456 + ], + "credit_map": [ + 0.01505586039274931, + 0.011415266431868076, + 0.03064698539674282, + 0.030955027788877487, + 0.022285407409071922, + 0.015263697132468224, + 0.03687895089387894, + 0.023030657321214676, + 0.006389220245182514, + 0.011488625779747963, + 0.009216457605361938, + 0.012689244002103806, + 0.007677879184484482, + 0.008255242370069027, + 0.006807096768170595, + 0.008430211804807186, + 0.022090010344982147, + 0.021033072844147682, + 0.017811771482229233, + 0.02769048884510994, + 0.012338492088019848, + 0.008931082673370838, + 0.017142724245786667, + 0.033923570066690445, + 0.015224041417241096, + 0.015114226378500462, + 0.02195099927484989, + 0.02206091769039631, + 0.013881247490644455, + 0.006398311350494623, + 0.016277367249131203, + 0.019943563267588615, + 0.028274869546294212, + 0.017639055848121643, + 0.025042863562703133, + 0.012299524620175362, + 0.01685228943824768, + 0.010415191762149334, + 0.015758775174617767, + 0.014125349931418896, + 0.021443108096718788, + 0.017575956881046295, + 0.011326358653604984, + 0.01216974388808012, + 0.019096113741397858, + 0.008088678121566772, + 0.01298375055193901, + 0.01458784844726324, + 0.0037499924656003714, + 0.005659709684550762, + 0.0032999140676110983, + 0.020921066403388977, + 0.006146706640720367, + 0.019332611933350563, + 0.009094036184251308, + 0.010857599787414074, + 0.004389193374663591, + 0.0048438566736876965, + 0.003373133484274149, + 0.017027156427502632, + 0.004676244221627712, + 0.007322374265640974, + 0.009486676193773746, + 0.004441325087100267, + 0.001561409910209477, + 0.0023642645683139563, + 0.0023170080967247486, + 0.003220704849809408, + 0.002776801586151123, + 0.002977165160700679, + 0.0031761133577674627, + 0.004332562442868948, + 0.003970461897552013, + 0.002022257074713707, + 0.009652815759181976, + 0.0037000267766416073, + 0.002192664658650756, + 0.002918691374361515, + 0.006952113937586546, + 0.005266162566840649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.13432396948337555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370229199528694 + ], + "opposite_logprob_tokens": [ + -3.6513702869415283 + ], + "logprob_reward": -0.026370229199528694, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370229199528694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18232275545597076, + 0.05433858186006546, + 0.16831006109714508, + 0.14334075152873993, + 0.09296661615371704, + 0.08518089354038239, + 0.08296460658311844, + 0.043329741805791855, + 0.021088633686304092, + 0.1261572688817978 + ], + "credit_map": [ + 0.018127331510186195, + 0.01871751993894577, + 0.030042309314012527, + 0.0299607515335083, + 0.011903049424290657, + 0.019136199727654457, + 0.02796056680381298, + 0.02647504024207592, + 0.004273965954780579, + 0.006816714070737362, + 0.007184992544353008, + 0.008475752547383308, + 0.006641390733420849, + 0.005376975983381271, + 0.007544719614088535, + 0.00802407506853342, + 0.02462160401046276, + 0.021190213039517403, + 0.02296249195933342, + 0.030055779963731766, + 0.011338814161717892, + 0.009625009261071682, + 0.015424638986587524, + 0.033091507852077484, + 0.014538121409714222, + 0.022799013182520866, + 0.017813891172409058, + 0.017513172701001167, + 0.012417455203831196, + 0.009046902880072594, + 0.01900402270257473, + 0.030208170413970947, + 0.019962405785918236, + 0.013697782531380653, + 0.015167553909122944, + 0.00722914282232523, + 0.010438961908221245, + 0.006914528552442789, + 0.010212613269686699, + 0.00934362132102251, + 0.016298580914735794, + 0.011091312393546104, + 0.010382577776908875, + 0.007953841239213943, + 0.013575967401266098, + 0.0062263417057693005, + 0.009350445121526718, + 0.010301828384399414, + 0.0037212101742625237, + 0.00471470458433032, + 0.0021628376562148333, + 0.01942211017012596, + 0.008744839578866959, + 0.015421269461512566, + 0.009592698886990547, + 0.019184939563274384, + 0.003515346674248576, + 0.00363570312038064, + 0.0028994425665587187, + 0.013574735261499882, + 0.004275862127542496, + 0.0054237800650298595, + 0.006938077509403229, + 0.0030667982064187527, + 0.0014501683181151748, + 0.001974300714209676, + 0.002197407418861985, + 0.003718673251569271, + 0.0016132639721035957, + 0.003061401192098856, + 0.0031011956743896008, + 0.003972223028540611, + 0.012735526077449322, + 0.00436779111623764, + 0.035126663744449615, + 0.010795057751238346, + 0.004464427940547466, + 0.005737778265029192, + 0.03941945359110832, + 0.01351056806743145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.026370229199528694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013434203574433923 + ], + "opposite_logprob_tokens": [ + -6.626343250274658 + ], + "logprob_reward": -0.0013434203574433923, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013434203574433923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18854652345180511, + 0.09939703345298767, + 0.03739004209637642, + 0.19797399640083313, + 0.0750698447227478, + 0.08216456323862076, + 0.19382436573505402, + 0.04374100640416145, + 0.03572491556406021, + 0.04616773873567581 + ], + "credit_map": [ + 0.013172990642488003, + 0.011886976659297943, + 0.0428558886051178, + 0.031863149255514145, + 0.01045120321214199, + 0.01670892909169197, + 0.018594099208712578, + 0.04301328957080841, + 0.01064168382436037, + 0.018601516261696815, + 0.012577014975249767, + 0.013460677117109299, + 0.009391732513904572, + 0.011667642742395401, + 0.012770631350576878, + 0.010286130011081696, + 0.005076369270682335, + 0.004328305833041668, + 0.007276336196810007, + 0.004479583818465471, + 0.0035879258066415787, + 0.002412880538031459, + 0.004379050340503454, + 0.00584959052503109, + 0.020723553374409676, + 0.028431838378310204, + 0.02194897271692753, + 0.018050657585263252, + 0.026397863402962685, + 0.011720558628439903, + 0.03442417085170746, + 0.03627638891339302, + 0.01204331498593092, + 0.009849702939391136, + 0.01023296732455492, + 0.007441784255206585, + 0.01103118620812893, + 0.0058441162109375, + 0.010261024348437786, + 0.008365750312805176, + 0.014087089337408543, + 0.010825147852301598, + 0.010168141685426235, + 0.009479031898081303, + 0.011310099624097347, + 0.006248247344046831, + 0.010255811735987663, + 0.009790991432964802, + 0.008260570466518402, + 0.01395462453365326, + 0.00555022619664669, + 0.0755203366279602, + 0.016257504001259804, + 0.039350032806396484, + 0.02584746479988098, + 0.009083610028028488, + 0.0021879258565604687, + 0.0033902646973729134, + 0.002369588939473033, + 0.013719195500016212, + 0.004044923931360245, + 0.002681198762729764, + 0.012010750360786915, + 0.003337153932079673, + 0.002860306529328227, + 0.0028317414689809084, + 0.0035957302898168564, + 0.0060736010782420635, + 0.003152344608679414, + 0.0040501379407942295, + 0.006126411259174347, + 0.007034642621874809, + 0.005433280486613512, + 0.0019649129826575518, + 0.010608701966702938, + 0.00415034219622612, + 0.002576023805886507, + 0.0027205876540392637, + 0.013734436593949795, + 0.004979455843567848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0013434203574433923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021996130235493183 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021996130235493183, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021996130235493183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14338171482086182, + 0.07484735548496246, + 0.11679744720458984, + 0.09271251410245895, + 0.06177332624793053, + 0.09040191769599915, + 0.21875017881393433, + 0.03108079358935356, + 0.04503866285085678, + 0.12521609663963318 + ], + "credit_map": [ + 0.00890995841473341, + 0.009532030671834946, + 0.03496722877025604, + 0.025351254269480705, + 0.007303121034055948, + 0.012011397629976273, + 0.015523461624979973, + 0.029783260077238083, + 0.007445536553859711, + 0.015426469035446644, + 0.008890530094504356, + 0.012176564894616604, + 0.006811769213527441, + 0.009132349863648415, + 0.00746654300019145, + 0.007497589569538832, + 0.0200375784188509, + 0.015905750915408134, + 0.012294663116335869, + 0.019830642268061638, + 0.008658046834170818, + 0.006288869306445122, + 0.013686162419617176, + 0.02009573020040989, + 0.011746003292500973, + 0.01684582233428955, + 0.010402614250779152, + 0.00799997802823782, + 0.009119248948991299, + 0.005592260044068098, + 0.013137809000909328, + 0.017868777737021446, + 0.008367535658180714, + 0.008401438593864441, + 0.008303102105855942, + 0.006880552042275667, + 0.009391829371452332, + 0.0041208392940461636, + 0.009554531425237656, + 0.006753495428711176, + 0.015677327290177345, + 0.011791029013693333, + 0.010697511956095695, + 0.00918099470436573, + 0.012730990536510944, + 0.00727313244715333, + 0.011299427598714828, + 0.011751501820981503, + 0.009783991612493992, + 0.012685324065387249, + 0.0050843264907598495, + 0.09826580435037613, + 0.015878770500421524, + 0.018522903323173523, + 0.035826705396175385, + 0.02270233817398548, + 0.0025156629271805286, + 0.0037171971052885056, + 0.0022377441637218, + 0.004139769822359085, + 0.0046952529810369015, + 0.004817308858036995, + 0.005794093944132328, + 0.003163764951750636, + 0.003044312819838524, + 0.0030802853871136904, + 0.004492726642638445, + 0.007811820600181818, + 0.004018120467662811, + 0.00555703230202198, + 0.008864739909768105, + 0.008169625885784626, + 0.013045038096606731, + 0.003930461127310991, + 0.0343366377055645, + 0.010243179276585579, + 0.004689508583396673, + 0.0057656955905258656, + 0.039812393486499786, + 0.013393172062933445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.0021996130235493183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987461388111115 + ], + "opposite_logprob_tokens": [ + -1.8048746585845947 + ], + "logprob_reward": -0.17987461388111115, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987461388111115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06670702993869781, + 0.05874243378639221, + 0.03685266152024269, + 0.042211443185806274, + 0.0821867510676384, + 0.10799587517976761, + 0.28098902106285095, + 0.07467643916606903, + 0.018659893423318863, + 0.23097842931747437 + ], + "credit_map": [ + 0.005270735826343298, + 0.004126464482396841, + 0.013895758427679539, + 0.012193026952445507, + 0.006099049001932144, + 0.00541490875184536, + 0.009675963781774044, + 0.010031125508248806, + 0.004886072129011154, + 0.010789908468723297, + 0.0070416065864264965, + 0.010645645670592785, + 0.006271447986364365, + 0.0066178166307508945, + 0.0058163548819720745, + 0.006673579569905996, + 0.004722913261502981, + 0.0055061280727386475, + 0.005629899445921183, + 0.005726770963519812, + 0.002751765074208379, + 0.002567090094089508, + 0.004446585197001696, + 0.0055015115067362785, + 0.004775146022439003, + 0.0057553742080926895, + 0.006458235438913107, + 0.006566387601196766, + 0.0045394692569971085, + 0.0020457410719245672, + 0.005570878274738789, + 0.006500213406980038, + 0.011432665400207043, + 0.010861187241971493, + 0.01049097441136837, + 0.008940660394728184, + 0.01174061931669712, + 0.006178764626383781, + 0.013526863418519497, + 0.009015021845698357, + 0.015408666804432869, + 0.014331120997667313, + 0.012639272026717663, + 0.010702785104513168, + 0.0181637741625309, + 0.009005251340568066, + 0.013987074606120586, + 0.013757925480604172, + 0.012232126668095589, + 0.015717018395662308, + 0.00605778768658638, + 0.11755073070526123, + 0.02049361541867256, + 0.04396078735589981, + 0.05247987434267998, + 0.012497086077928543, + 0.004717271775007248, + 0.004476170055568218, + 0.00414311746135354, + 0.025195106863975525, + 0.006011187564581633, + 0.0066517735831439495, + 0.01874535158276558, + 0.004736458417028189, + 0.0012204538797959685, + 0.0016581270610913634, + 0.0016072241123765707, + 0.004039247054606676, + 0.0018212056020274758, + 0.0023133058566600084, + 0.003043206874281168, + 0.00295712286606431, + 0.02321888878941536, + 0.007599635515362024, + 0.0640225037932396, + 0.019565261900424957, + 0.008229896426200867, + 0.009623478166759014, + 0.07233695685863495, + 0.02638179436326027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.17987461388111115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, fast-cut shot, shot from the first-person point of view (POV) of a surveillance camera, fixed high on a building's facade, captures a deserted alley bathed in the glaring afternoon sun from a slightly bird's-eye perspective. The wide-angle lens's slight fisheye distortion distorts the edges of the image. Suddenly, a rough, red clay brick, its surface covered in weathered cracks and mottled mortar residue, spins and hurtles through the air, whirring directly toward the camera. In the final moments before impact, the brick zooms dramatically in, its angular details crystal clear. Then, with incredible force, it slams into the camera lens, instantly shattering the glass in a dense, radiating spiderweb of cracks. The entire frame, fragmented by the shattered glass prism, fades to black in a violent flash and a blinding white light, creating an abrupt and impactful aesthetic of violence.", + "seed": 18242, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_8/index_6800_seed_18242_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21192822763696312, + "vlm_logprob": -0.21192822763696312, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.21192822763696312, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8399919271469116 + ], + "opposite_logprob_tokens": [ + -0.5899919271469116 + ], + "logprob_reward": -0.8399919271469116, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8399919271469116, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10507629066705704, + 0.07692885398864746, + 0.06975772976875305, + 0.05876130238175392, + 0.13642774522304535, + 0.04529373347759247, + 0.08565882593393326, + 0.2891370952129364, + 0.10232734680175781, + 0.030631106346845627 + ], + "credit_map": [ + 0.012773644179105759, + 0.017653200775384903, + 0.019754698500037193, + 0.012750092893838882, + 0.010427393019199371, + 0.0068539902567863464, + 0.009718428365886211, + 0.01514484267681837, + 0.009072220884263515, + 0.007911440916359425, + 0.011692271567881107, + 0.01311381533741951, + 0.008520664647221565, + 0.0061026811599731445, + 0.010190409608185291, + 0.01032535545527935, + 0.009255663491785526, + 0.009957321919500828, + 0.006644054781645536, + 0.010227140970528126, + 0.011696852743625641, + 0.004607846029102802, + 0.006157515570521355, + 0.011211325414478779, + 0.009115606546401978, + 0.008496807888150215, + 0.00768677331507206, + 0.010144074447453022, + 0.007215522695332766, + 0.0044393036514520645, + 0.005547376349568367, + 0.006115831900388002, + 0.011351366527378559, + 0.011570923961699009, + 0.016849707812070847, + 0.01411230769008398, + 0.018054816871881485, + 0.023010320961475372, + 0.016639569774270058, + 0.024838734418153763, + 0.006550049874931574, + 0.005288899876177311, + 0.004591860808432102, + 0.007035143673419952, + 0.00542393047362566, + 0.005337131209671497, + 0.006457408890128136, + 0.004609308205544949, + 0.006427034270018339, + 0.014382393099367619, + 0.012636453844606876, + 0.013114511966705322, + 0.011046973057091236, + 0.010353995487093925, + 0.007341441232711077, + 0.01035601831972599, + 0.02085656300187111, + 0.03125518187880516, + 0.021555669605731964, + 0.03769775480031967, + 0.05831348896026611, + 0.04404584318399429, + 0.04700478911399841, + 0.028407789766788483, + 0.013744261115789413, + 0.008018871769309044, + 0.014952198602259159, + 0.013526685535907745, + 0.018258458003401756, + 0.013805525377392769, + 0.010848135687410831, + 0.009173216298222542, + 0.004004575312137604, + 0.004018232226371765, + 0.0044330027885735035, + 0.005266137886792421, + 0.003587381448596716, + 0.002827213378623128, + 0.0034891448449343443, + 0.003005417762324214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.8399919271469116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005883008241653 + ], + "opposite_logprob_tokens": [ + -2.695058822631836 + ], + "logprob_reward": -0.07005883008241653, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005883008241653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29370781779289246, + 0.10299338400363922, + 0.06656273454427719, + 0.0992475375533104, + 0.0980813056230545, + 0.05091904476284981, + 0.09122081845998764, + 0.11245208233594894, + 0.05065532028675079, + 0.03415984287858009 + ], + "credit_map": [ + 0.040177199989557266, + 0.06544473767280579, + 0.047943104058504105, + 0.04011186957359314, + 0.021388789638876915, + 0.014714803546667099, + 0.02308567613363266, + 0.04084165394306183, + 0.01582461968064308, + 0.010355903767049313, + 0.014048662967979908, + 0.016972357407212257, + 0.011826714500784874, + 0.007759893778711557, + 0.009299691766500473, + 0.01690555363893509, + 0.008607673458755016, + 0.006225409917533398, + 0.00709739001467824, + 0.011782190762460232, + 0.01028294488787651, + 0.005982975009828806, + 0.005342066287994385, + 0.011242092587053776, + 0.012355371378362179, + 0.014044754207134247, + 0.013171160593628883, + 0.014949201606214046, + 0.012910095043480396, + 0.008971932344138622, + 0.012037738226354122, + 0.010807297192513943, + 0.009030431509017944, + 0.009655733592808247, + 0.012029625475406647, + 0.010577240958809853, + 0.01683434471487999, + 0.012288828380405903, + 0.012172004207968712, + 0.015493107959628105, + 0.008811542764306068, + 0.007084977813065052, + 0.004789500962942839, + 0.00618347991257906, + 0.006147845182567835, + 0.005775343626737595, + 0.0062771751545369625, + 0.0058491844683885574, + 0.006820496637374163, + 0.011702518910169601, + 0.008756078779697418, + 0.009999867528676987, + 0.01341646071523428, + 0.01354542188346386, + 0.011762471869587898, + 0.01521751657128334, + 0.008241109549999237, + 0.0070484839379787445, + 0.008710501715540886, + 0.013881640508770943, + 0.024262744933366776, + 0.02015661634504795, + 0.016963956877589226, + 0.01318703405559063, + 0.007100416347384453, + 0.003814384574070573, + 0.009099414572119713, + 0.004069013986736536, + 0.010541572235524654, + 0.0071571688167750835, + 0.005141217261552811, + 0.003732141340151429, + 0.0037954323925077915, + 0.003164054127410054, + 0.004100634716451168, + 0.0045136958360672, + 0.005554894916713238, + 0.0030978878494352102, + 0.004393123090267181, + 0.005540122743695974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.07005883008241653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003626557532697916 + ], + "opposite_logprob_tokens": [ + -5.628626346588135 + ], + "logprob_reward": -0.003626557532697916, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003626557532697916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.087330162525177, + 0.07443492859601974, + 0.14475499093532562, + 0.20146088302135468, + 0.0954238548874855, + 0.06830295920372009, + 0.051552969962358475, + 0.18572677671909332, + 0.07732822000980377, + 0.013684186153113842 + ], + "credit_map": [ + 0.008338706567883492, + 0.013214126229286194, + 0.011858883313834667, + 0.011255181394517422, + 0.011092934757471085, + 0.004608407616615295, + 0.01562691479921341, + 0.011335010640323162, + 0.008556795306503773, + 0.007775969337671995, + 0.010082867927849293, + 0.013041622936725616, + 0.009451879188418388, + 0.004919879138469696, + 0.00921327993273735, + 0.01139263715595007, + 0.017130402848124504, + 0.018610170111060143, + 0.01305284071713686, + 0.024545038118958473, + 0.02838168293237686, + 0.010860047303140163, + 0.012782596983015537, + 0.01939222402870655, + 0.020201826468110085, + 0.04768669977784157, + 0.018955647945404053, + 0.03106655925512314, + 0.033973414450883865, + 0.02049550786614418, + 0.01459686178714037, + 0.01448439247906208, + 0.012340656481683254, + 0.009709731675684452, + 0.013127775862812996, + 0.009342317469418049, + 0.014647216536104679, + 0.014897436834871769, + 0.011535387486219406, + 0.00982334092259407, + 0.006868603173643351, + 0.008172763511538506, + 0.007123785093426704, + 0.00997322890907526, + 0.007714455481618643, + 0.008401524275541306, + 0.011878062039613724, + 0.008170547895133495, + 0.002874774858355522, + 0.00724059296771884, + 0.005427214782685041, + 0.008216270245611668, + 0.005767263937741518, + 0.00696825934574008, + 0.0080103175714612, + 0.007048277650028467, + 0.010387619957327843, + 0.02009156160056591, + 0.015742070972919464, + 0.03147334232926369, + 0.03145541250705719, + 0.027216536924242973, + 0.03200460597872734, + 0.01735563389956951, + 0.00851568952202797, + 0.008831187151372433, + 0.006898610386997461, + 0.011780566535890102, + 0.010245619341731071, + 0.010661423206329346, + 0.010860458016395569, + 0.009534676559269428, + 0.0017521388363093138, + 0.0015998982125893235, + 0.0020427252165973186, + 0.0015296005876734853, + 0.002024197019636631, + 0.0013462918577715755, + 0.0018753150943666697, + 0.0015140202594920993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.003626557532697916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193756565451622 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193756565451622, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193756565451622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05621762201189995, + 0.04685892537236214, + 0.02990850619971752, + 0.08034330606460571, + 0.03646727651357651, + 0.10070262849330902, + 0.08503572642803192, + 0.37541958689689636, + 0.1472897231578827, + 0.04175671190023422 + ], + "credit_map": [ + 0.006325144320726395, + 0.007940571755170822, + 0.009966909885406494, + 0.008958976715803146, + 0.006531805265694857, + 0.003053846303373575, + 0.0064114914275705814, + 0.00702887075021863, + 0.0064999996684491634, + 0.005672144703567028, + 0.006134194787591696, + 0.007954860106110573, + 0.0045842500403523445, + 0.0028309463523328304, + 0.006352392490953207, + 0.006830135360360146, + 0.0032358127646148205, + 0.0048154969699680805, + 0.0027239269111305475, + 0.005005306098610163, + 0.004909902345389128, + 0.002599716652184725, + 0.0023884368129074574, + 0.004229906015098095, + 0.010641603730618954, + 0.016214538365602493, + 0.009902093559503555, + 0.010349460877478123, + 0.012900776229798794, + 0.008369766175746918, + 0.005749620497226715, + 0.0062154438346624374, + 0.004140102304518223, + 0.0057579390704631805, + 0.0041196998208761215, + 0.003487165318801999, + 0.005306673236191273, + 0.004770574159920216, + 0.004372705705463886, + 0.004512417130172253, + 0.008175100199878216, + 0.008992988616228104, + 0.011226764880120754, + 0.015403703786432743, + 0.010142484679818153, + 0.011623114347457886, + 0.023250620812177658, + 0.011887850239872932, + 0.0051308730617165565, + 0.009182928130030632, + 0.010850206017494202, + 0.010865303687751293, + 0.009617927484214306, + 0.012194624170660973, + 0.011669144965708256, + 0.01552471425384283, + 0.01857646182179451, + 0.03375018388032913, + 0.028455808758735657, + 0.055009033530950546, + 0.07378213107585907, + 0.06250786781311035, + 0.06719759851694107, + 0.036140501499176025, + 0.01557100284844637, + 0.01635371707379818, + 0.012216100469231606, + 0.021627457812428474, + 0.02137177623808384, + 0.021245094016194344, + 0.02132907323539257, + 0.01757548376917839, + 0.00396653963252902, + 0.004836365580558777, + 0.0045930081978440285, + 0.00772872194647789, + 0.005439367610961199, + 0.00417478010058403, + 0.004740350879728794, + 0.006277573294937611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.003193756565451622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277006685733795 + ], + "opposite_logprob_tokens": [ + -2.0177700519561768 + ], + "logprob_reward": -0.14277006685733795, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14277006685733795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03664112836122513, + 0.019437046721577644, + 0.027508219704031944, + 0.041767872869968414, + 0.012906819581985474, + 0.0647921934723854, + 0.06039245426654816, + 0.6739658713340759, + 0.050666794180870056, + 0.011921617202460766 + ], + "credit_map": [ + 0.0035879267379641533, + 0.005818255245685577, + 0.003609738312661648, + 0.005778391379863024, + 0.0039906627498567104, + 0.0017315489239990711, + 0.007956165820360184, + 0.004168439656496048, + 0.002227960154414177, + 0.002149442443624139, + 0.0029284439515322447, + 0.0035455513279885054, + 0.0019725405145436525, + 0.0009627117542549968, + 0.002940569771453738, + 0.0027098283171653748, + 0.0027706266846507788, + 0.0031085689552128315, + 0.0027014417573809624, + 0.0059977625496685505, + 0.0046862284652888775, + 0.0018956916173920035, + 0.0021630360279232264, + 0.0041848644614219666, + 0.00434145936742425, + 0.008492011576890945, + 0.003792868461459875, + 0.008273381739854813, + 0.006017548497766256, + 0.0035803872160613537, + 0.0038155214861035347, + 0.0034546926617622375, + 0.0012379615800455213, + 0.0012482431484386325, + 0.0017297022277489305, + 0.0013587435241788626, + 0.0026151593774557114, + 0.002406783401966095, + 0.0012920534936711192, + 0.0010181727120652795, + 0.004831462632864714, + 0.005858116317540407, + 0.008268913254141808, + 0.009787322953343391, + 0.006548131816089153, + 0.007533296011388302, + 0.014601783826947212, + 0.007363168057054281, + 0.0028984439559280872, + 0.005768647417426109, + 0.008034354075789452, + 0.007198127917945385, + 0.008008677512407303, + 0.00790677685290575, + 0.007450374774634838, + 0.01312705036252737, + 0.04092228040099144, + 0.05493824928998947, + 0.04043738543987274, + 0.07333847880363464, + 0.16884967684745789, + 0.12184306234121323, + 0.10328169167041779, + 0.07035498321056366, + 0.006229379214346409, + 0.004666782915592194, + 0.005067865829914808, + 0.00427271006628871, + 0.011517008766531944, + 0.007542324252426624, + 0.006031319033354521, + 0.005339403636753559, + 0.0013402507174760103, + 0.0013747805496677756, + 0.0015672710724174976, + 0.001263022655621171, + 0.002062859246507287, + 0.0012122821062803268, + 0.001564312493428588, + 0.001536837313324213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.14277006685733795 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, fast-cut shot, shot from the first-person point of view (POV) of a surveillance camera, fixed high on a building's facade, captures a deserted alley bathed in the glaring afternoon sun from a slightly bird's-eye perspective. The wide-angle lens's slight fisheye distortion distorts the edges of the image. Suddenly, a rough, red clay brick, its surface covered in weathered cracks and mottled mortar residue, spins and hurtles through the air, whirring directly toward the camera. In the final moments before impact, the brick zooms dramatically in, its angular details crystal clear. Then, with incredible force, it slams into the camera lens, instantly shattering the glass in a dense, radiating spiderweb of cracks. The entire frame, fragmented by the shattered glass prism, fades to black in a violent flash and a blinding white light, creating an abrupt and impactful aesthetic of violence.", + "seed": 18243, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_8/index_6800_seed_18243_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1814601957798004, + "vlm_logprob": -0.1814601957798004, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.1814601957798004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5896638631820679 + ], + "opposite_logprob_tokens": [ + -0.8396638631820679 + ], + "logprob_reward": -0.5896638631820679, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5896638631820679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10014117509126663, + 0.017883095890283585, + 0.02781866490840912, + 0.2664872705936432, + 0.03671211749315262, + 0.06784193217754364, + 0.1535772681236267, + 0.3047966957092285, + 0.0165324117988348, + 0.008209445513784885 + ], + "credit_map": [ + 0.017891135066747665, + 0.010683371685445309, + 0.017352990806102753, + 0.008411305025219917, + 0.009993630461394787, + 0.017247123643755913, + 0.007481554057449102, + 0.011080066673457623, + 0.003220678772777319, + 0.0014869578881189227, + 0.002711073961108923, + 0.0016858780290931463, + 0.0024652520660310984, + 0.0018972631078213453, + 0.002567507792264223, + 0.001848483574576676, + 0.0037631874438375235, + 0.0032991620246320963, + 0.0038382867351174355, + 0.002521587535738945, + 0.004065191838890314, + 0.003599864197894931, + 0.003906624857336283, + 0.0028247609734535217, + 0.10120934247970581, + 0.010892240330576897, + 0.1155950129032135, + 0.009094399400055408, + 0.006386624649167061, + 0.00502729183062911, + 0.00669226236641407, + 0.011590086854994297, + 0.0038518616929650307, + 0.004619687795639038, + 0.004821171052753925, + 0.005007668863981962, + 0.004325821530073881, + 0.004727779887616634, + 0.005560397636145353, + 0.003797727171331644, + 0.004992868285626173, + 0.0115240179002285, + 0.008443037047982216, + 0.012912281788885593, + 0.009565522894263268, + 0.004711408168077469, + 0.009050266817212105, + 0.006642528343945742, + 0.009186065755784512, + 0.008681452833116055, + 0.03574344888329506, + 0.007476571947336197, + 0.009409460239112377, + 0.009423850104212761, + 0.023468513041734695, + 0.05018789693713188, + 0.03816860541701317, + 0.016803059726953506, + 0.07638741284608841, + 0.020067336037755013, + 0.02619626373052597, + 0.019226238131523132, + 0.08558236062526703, + 0.022365406155586243, + 0.0020651656668633223, + 0.0017330540576949716, + 0.002260739915072918, + 0.003209794405847788, + 0.0016172045143321157, + 0.0013325999025255442, + 0.002730343025177717, + 0.0015835098456591368, + 0.0006740028620697558, + 0.0009906821651384234, + 0.0012725412379950285, + 0.0012601037742570043, + 0.000997680239379406, + 0.0008253749692812562, + 0.0014894556952640414, + 0.0006996041629463434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.5896638631820679 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004871219396591 + ], + "opposite_logprob_tokens": [ + -2.6950488090515137 + ], + "logprob_reward": -0.07004871219396591, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004871219396591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06280305236577988, + 0.056064311414957047, + 0.041740722954273224, + 0.5650351047515869, + 0.04574137181043625, + 0.01624131388962269, + 0.14098462462425232, + 0.03786783292889595, + 0.017499208450317383, + 0.016022400930523872 + ], + "credit_map": [ + 0.010039768181741238, + 0.0061330655589699745, + 0.007719789631664753, + 0.007051977328956127, + 0.010011977516114712, + 0.008267179131507874, + 0.007817447185516357, + 0.005761854350566864, + 0.011721811257302761, + 0.004348547197878361, + 0.00874785054475069, + 0.003957552369683981, + 0.006145044695585966, + 0.01117249857634306, + 0.004993901122361422, + 0.004977110773324966, + 0.00501239625737071, + 0.004586211405694485, + 0.006159044802188873, + 0.004123731516301632, + 0.006170148029923439, + 0.005631100852042437, + 0.0060032387264072895, + 0.004054855089634657, + 0.21597205102443695, + 0.023397328332066536, + 0.24730618298053741, + 0.018490932881832123, + 0.012138953432440758, + 0.009957320056855679, + 0.013395784422755241, + 0.024376610293984413, + 0.004223584663122892, + 0.00453189155086875, + 0.0068123904056847095, + 0.005345892161130905, + 0.004485129378736019, + 0.010114967823028564, + 0.00566908298060298, + 0.004558433312922716, + 0.001669999910518527, + 0.00250329845584929, + 0.001995279686525464, + 0.0026448096614331007, + 0.0018348816083744168, + 0.0016862002667039633, + 0.001977623673155904, + 0.0019292214419692755, + 0.01106767263263464, + 0.0071782311424613, + 0.05512933060526848, + 0.007749231532216072, + 0.0073461467400193214, + 0.004937881603837013, + 0.01724582351744175, + 0.030330326408147812, + 0.00410509342327714, + 0.004376071505248547, + 0.006246756762266159, + 0.003600766882300377, + 0.004877905361354351, + 0.0037565885577350855, + 0.006953701842576265, + 0.0039509497582912445, + 0.0018892033258453012, + 0.0019011697731912136, + 0.0026571727357804775, + 0.0021395976655185223, + 0.0018739940132945776, + 0.0017587818438187242, + 0.0037967581301927567, + 0.0014825316611677408, + 0.001554140355437994, + 0.0021044560708105564, + 0.002223128918558359, + 0.002582301152870059, + 0.0017730362014845014, + 0.0018574872519820929, + 0.0022180862724781036, + 0.0017097663367167115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.07004871219396591 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961494520306587 + ], + "opposite_logprob_tokens": [ + -5.1309614181518555 + ], + "logprob_reward": -0.005961494520306587, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961494520306587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015212085098028183, + 0.015686100348830223, + 0.012748250737786293, + 0.18487003445625305, + 0.013035994954407215, + 0.07988078892230988, + 0.12382478266954422, + 0.4694819450378418, + 0.0613432414829731, + 0.023916780948638916 + ], + "credit_map": [ + 0.0024926920887082815, + 0.0017694218549877405, + 0.002722613513469696, + 0.002248360775411129, + 0.001561214798130095, + 0.0016942061483860016, + 0.0014781171921640635, + 0.0012454587267711759, + 0.002918111626058817, + 0.0016116559272632003, + 0.0020016729831695557, + 0.0009562036721035838, + 0.003793543903157115, + 0.0013840910978615284, + 0.0017943974817171693, + 0.0012264238903298974, + 0.001150793395936489, + 0.001653945306316018, + 0.0018462871666997671, + 0.001149181043729186, + 0.0013324751053005457, + 0.00249343179166317, + 0.0015600741608068347, + 0.0015620626509189606, + 0.07016787678003311, + 0.007714309263974428, + 0.08048763126134872, + 0.006104756146669388, + 0.00427192822098732, + 0.0034350028727203608, + 0.0046095941215753555, + 0.008078942075371742, + 0.0013326379703357816, + 0.0011252133408561349, + 0.0013486813986673951, + 0.001177524565719068, + 0.0013918174663558602, + 0.003086918266490102, + 0.0022489780094474554, + 0.0013242242857813835, + 0.005967028439044952, + 0.016828957945108414, + 0.008882412686944008, + 0.015283532440662384, + 0.010271262377500534, + 0.003894550958648324, + 0.010768740437924862, + 0.007984301075339317, + 0.006493192166090012, + 0.004174292553216219, + 0.03325963020324707, + 0.004515743814408779, + 0.006912467069923878, + 0.007377797272056341, + 0.018955135717988014, + 0.042136527597904205, + 0.061462607234716415, + 0.021197594702243805, + 0.1214662715792656, + 0.02533542364835739, + 0.03863772377371788, + 0.02710750512778759, + 0.13785381615161896, + 0.03642099350690842, + 0.006534924264997244, + 0.007946225814521313, + 0.007965597324073315, + 0.012755651958286762, + 0.005979348439723253, + 0.004338547587394714, + 0.009623966179788113, + 0.006198977120220661, + 0.0028540727216750383, + 0.002511872909963131, + 0.0027678697369992733, + 0.0036711215507239103, + 0.0027297118213027716, + 0.002419491531327367, + 0.004866753239184618, + 0.002095888601616025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.005961494520306587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607622765004635 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.01607622765004635, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607622765004635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015199985355138779, + 0.010181710124015808, + 0.007258949335664511, + 0.33567577600479126, + 0.010124005377292633, + 0.01582491770386696, + 0.09184035658836365, + 0.3659726679325104, + 0.10142482072114944, + 0.046496786177158356 + ], + "credit_map": [ + 0.002929792506620288, + 0.002315676538273692, + 0.002395693678408861, + 0.0016199917299672961, + 0.0015088607324287295, + 0.0017136617098003626, + 0.001522684469819069, + 0.0011936243390664458, + 0.0026206730399280787, + 0.0010614868951961398, + 0.0014449587324634194, + 0.0006618050974793732, + 0.0018135621212422848, + 0.001000948017463088, + 0.0009070863015949726, + 0.0006711894529871643, + 0.0008952940115705132, + 0.000795462925452739, + 0.0012168672401458025, + 0.0005948359030298889, + 0.0010095966281369328, + 0.0007060381467454135, + 0.0013400951866060495, + 0.0007007594686001539, + 0.12827269732952118, + 0.013845263049006462, + 0.1467207372188568, + 0.011212804354727268, + 0.007152328733354807, + 0.00603588717058301, + 0.007681034039705992, + 0.014755034819245338, + 0.001111659687012434, + 0.0007122947718016803, + 0.0013539966894313693, + 0.000977508258074522, + 0.0006238519563339651, + 0.0029348209500312805, + 0.0013754887040704489, + 0.0010343845933675766, + 0.001275533577427268, + 0.0034092841669917107, + 0.0014591107610613108, + 0.002322685904800892, + 0.0018227355321869254, + 0.0015605635708197951, + 0.00213253335095942, + 0.0018424707232043147, + 0.007562392856925726, + 0.004701611585915089, + 0.03766210004687309, + 0.004935082513839006, + 0.004628667142242193, + 0.0022944766096770763, + 0.01718088611960411, + 0.012875144369900227, + 0.04745230823755264, + 0.017554420977830887, + 0.09524796158075333, + 0.02020861953496933, + 0.03012501262128353, + 0.02156347967684269, + 0.10627859830856323, + 0.02754225581884384, + 0.011104844510555267, + 0.012997337616980076, + 0.014028256759047508, + 0.020614294335246086, + 0.009943102486431599, + 0.00675237039104104, + 0.015811439603567123, + 0.010173166170716286, + 0.0057712337002158165, + 0.005427305120974779, + 0.005554665811359882, + 0.00776211591437459, + 0.0051560550928115845, + 0.004331901203840971, + 0.008401644416153431, + 0.004091864451766014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.01607622765004635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555068135261536 + ], + "opposite_logprob_tokens": [ + -1.600550651550293 + ], + "logprob_reward": -0.22555068135261536, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555068135261536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05671768635511398, + 0.020520269870758057, + 0.016717948019504547, + 0.14355535805225372, + 0.015850383788347244, + 0.022340096533298492, + 0.12538795173168182, + 0.4471355080604553, + 0.13645149767398834, + 0.015323243103921413 + ], + "credit_map": [ + 0.008376474492251873, + 0.007606081664562225, + 0.006766785867512226, + 0.00541786290705204, + 0.007627594284713268, + 0.009726950898766518, + 0.005303166341036558, + 0.005892773624509573, + 0.005686198826879263, + 0.0014748919056728482, + 0.0038037949707359076, + 0.001288927742280066, + 0.00194354890845716, + 0.003155522746965289, + 0.0016044910298660398, + 0.001562895136885345, + 0.0018028776394203305, + 0.0022728529293090105, + 0.002627420937642455, + 0.0019968077540397644, + 0.0024183017667382956, + 0.001871098531410098, + 0.002332519507035613, + 0.001396069535985589, + 0.05479208007454872, + 0.005979434121400118, + 0.06247631460428238, + 0.004819217137992382, + 0.003035713220015168, + 0.0026222264859825373, + 0.0036790696904063225, + 0.006151304580271244, + 0.0014875541673973203, + 0.001615186338312924, + 0.0019320222781971097, + 0.0018533357651904225, + 0.0018133295234292746, + 0.00355157139711082, + 0.0021592136472463608, + 0.0014381710207089782, + 0.0018750758608803153, + 0.003538672113791108, + 0.0027085498441010714, + 0.004157297778874636, + 0.002965340157970786, + 0.0022102966904640198, + 0.0029056842904537916, + 0.001979181310161948, + 0.009370586834847927, + 0.008476282469928265, + 0.048065438866615295, + 0.00830921158194542, + 0.008184194564819336, + 0.005310419015586376, + 0.017839819192886353, + 0.019832003861665726, + 0.05842681974172592, + 0.019930966198444366, + 0.11691878736019135, + 0.025043172761797905, + 0.03642144426703453, + 0.025638027116656303, + 0.13092811405658722, + 0.03382822498679161, + 0.0143575519323349, + 0.016994576901197433, + 0.019315050914883614, + 0.02432042546570301, + 0.013780124485492706, + 0.011882198974490166, + 0.022610176354646683, + 0.013191415928304195, + 0.00178165128454566, + 0.0013373303227126598, + 0.0028718782123178244, + 0.00217539188452065, + 0.0020334108266979456, + 0.0011601588921621442, + 0.002703816629946232, + 0.0012596054002642632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.22555068135261536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, fast-cut shot, shot from the first-person point of view (POV) of a surveillance camera, fixed high on a building's facade, captures a deserted alley bathed in the glaring afternoon sun from a slightly bird's-eye perspective. The wide-angle lens's slight fisheye distortion distorts the edges of the image. Suddenly, a rough, red clay brick, its surface covered in weathered cracks and mottled mortar residue, spins and hurtles through the air, whirring directly toward the camera. In the final moments before impact, the brick zooms dramatically in, its angular details crystal clear. Then, with incredible force, it slams into the camera lens, instantly shattering the glass in a dense, radiating spiderweb of cracks. The entire frame, fragmented by the shattered glass prism, fades to black in a violent flash and a blinding white light, creating an abrupt and impactful aesthetic of violence.", + "seed": 18244, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_8/index_6800_seed_18244_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24886847352609037, + "vlm_logprob": -0.24886847352609037, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.24886847352609037, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1511528491973877 + ], + "opposite_logprob_tokens": [ + -0.4011528491973877 + ], + "logprob_reward": -1.1511528491973877, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1511528491973877, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10349959880113602, + 0.07017123699188232, + 0.03161175921559334, + 0.43067029118537903, + 0.07548396289348602, + 0.06865467131137848, + 0.05688494071364403, + 0.0379524752497673, + 0.09941533207893372, + 0.02565568871796131 + ], + "credit_map": [ + 0.009542795829474926, + 0.01372904609888792, + 0.012074851430952549, + 0.012723008170723915, + 0.012598531320691109, + 0.008455531671643257, + 0.023503808304667473, + 0.010872027836740017, + 0.00539544178172946, + 0.004728212486952543, + 0.0064813983626663685, + 0.015325777232646942, + 0.013114163652062416, + 0.00791600625962019, + 0.008028705604374409, + 0.009181530214846134, + 0.0023031861055642366, + 0.0029248634818941355, + 0.003975105006247759, + 0.004254205618053675, + 0.0044318800792098045, + 0.003388193203136325, + 0.0057479520328342915, + 0.004586372058838606, + 0.03439442440867424, + 0.057251445949077606, + 0.07400676608085632, + 0.12380967289209366, + 0.0253478791564703, + 0.027708010748028755, + 0.04315841570496559, + 0.04499368742108345, + 0.006914163939654827, + 0.006242493167519569, + 0.014266849495470524, + 0.015718549489974976, + 0.00715760700404644, + 0.007476400583982468, + 0.009866097941994667, + 0.007841800339519978, + 0.007873475551605225, + 0.00758140766993165, + 0.013196990825235844, + 0.007970917038619518, + 0.005495118442922831, + 0.008424454368650913, + 0.0075253769755363464, + 0.010586932301521301, + 0.01018523145467043, + 0.0068232896737754345, + 0.008662325330078602, + 0.006048304960131645, + 0.005206230096518993, + 0.005297847092151642, + 0.006975227501243353, + 0.00768648786470294, + 0.002716539893299341, + 0.005061956588178873, + 0.004654032178223133, + 0.006997126154601574, + 0.004905260633677244, + 0.004390279296785593, + 0.004919291939586401, + 0.004307993222028017, + 0.008664127439260483, + 0.018379902467131615, + 0.0161634162068367, + 0.01516680233180523, + 0.006986293010413647, + 0.009278818964958191, + 0.010406133718788624, + 0.014369846321642399, + 0.0025034251157194376, + 0.0032088076695799828, + 0.0032101054675877094, + 0.004236808978021145, + 0.003656460903584957, + 0.0027088953647762537, + 0.003849776927381754, + 0.0022814078256487846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -1.1511528491973877 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500732734799385 + ], + "opposite_logprob_tokens": [ + -2.930007219314575 + ], + "logprob_reward": -0.05500732734799385, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500732734799385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18503373861312866, + 0.1008070558309555, + 0.06489695608615875, + 0.3424108624458313, + 0.07711605727672577, + 0.05838283151388168, + 0.036414314061403275, + 0.02096957340836525, + 0.09167221188545227, + 0.02229638770222664 + ], + "credit_map": [ + 0.01287601888179779, + 0.010605903342366219, + 0.022662850096821785, + 0.0284018125385046, + 0.03148360177874565, + 0.0201801098883152, + 0.04360681772232056, + 0.015216615051031113, + 0.005673984996974468, + 0.007301452104002237, + 0.009785625152289867, + 0.012378021143376827, + 0.021784748882055283, + 0.013173548504710197, + 0.01727953553199768, + 0.013430137187242508, + 0.0054548196494579315, + 0.005496113561093807, + 0.011607054620981216, + 0.0072411890141665936, + 0.00796573143452406, + 0.006581296678632498, + 0.012029210105538368, + 0.008521538227796555, + 0.027834411710500717, + 0.04384368285536766, + 0.05798253417015076, + 0.0872620940208435, + 0.018521035090088844, + 0.025060610845685005, + 0.03281909227371216, + 0.04908742010593414, + 0.009181882254779339, + 0.004743608646094799, + 0.01660745218396187, + 0.008831603452563286, + 0.010969414375722408, + 0.010378760285675526, + 0.007132180035114288, + 0.009271152317523956, + 0.007146112620830536, + 0.006320674438029528, + 0.009467290714383125, + 0.007585911545902491, + 0.005448201205581427, + 0.00732659175992012, + 0.005592395551502705, + 0.009495653212070465, + 0.0035156384110450745, + 0.0019753368105739355, + 0.004263715352863073, + 0.004956410266458988, + 0.0034380280412733555, + 0.005542888306081295, + 0.009090491570532322, + 0.003631802974268794, + 0.0018960194429382682, + 0.0018942391034215689, + 0.0024554943665862083, + 0.003112510312348604, + 0.0025545358657836914, + 0.0031962827779352665, + 0.003199056489393115, + 0.0026614335365593433, + 0.0076203858479857445, + 0.015916042029857635, + 0.016876904293894768, + 0.012554305605590343, + 0.007366418372839689, + 0.00840520765632391, + 0.00989703368395567, + 0.013035907410085201, + 0.0025610884185880423, + 0.0025658011436462402, + 0.002672083443030715, + 0.0025969056878238916, + 0.0036529896315187216, + 0.003034160239621997, + 0.003195724915713072, + 0.0020176339894533157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.05500732734799385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032055205665528774 + ], + "opposite_logprob_tokens": [ + -5.753205299377441 + ], + "logprob_reward": -0.0032055205665528774, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032055205665528774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05190178006887436, + 0.025330889970064163, + 0.030777210369706154, + 0.16129346191883087, + 0.4596486985683441, + 0.030215099453926086, + 0.02345266379415989, + 0.0324283093214035, + 0.09372837841510773, + 0.09122350066900253 + ], + "credit_map": [ + 0.003961888141930103, + 0.0035023195669054985, + 0.0061466083861887455, + 0.0071207862347364426, + 0.009257281199097633, + 0.007418034598231316, + 0.010055403225123882, + 0.004439456854015589, + 0.0026235184632241726, + 0.0022009408567100763, + 0.003060741815716028, + 0.005865796469151974, + 0.0029233412351459265, + 0.0022383835166692734, + 0.003671081503853202, + 0.0027470840141177177, + 0.004493319895118475, + 0.0037212353199720383, + 0.004215738736093044, + 0.003642069408670068, + 0.003916627261787653, + 0.002973901340737939, + 0.0032196366228163242, + 0.004594677127897739, + 0.010570270009338856, + 0.011625631712377071, + 0.03283020853996277, + 0.01784604601562023, + 0.008656701073050499, + 0.01835307478904724, + 0.01811504364013672, + 0.04329647123813629, + 0.04688763618469238, + 0.0302851852029562, + 0.08206027746200562, + 0.11002594232559204, + 0.04846286401152611, + 0.047149382531642914, + 0.05115643888711929, + 0.04362093657255173, + 0.005439333152025938, + 0.002950466936454177, + 0.005674110725522041, + 0.0033855962101370096, + 0.0032984560821205378, + 0.0030796413775533438, + 0.0035832549910992384, + 0.002804237650707364, + 0.0037817449774593115, + 0.0021139781456440687, + 0.0032742947805672884, + 0.0034553322475403547, + 0.002050718991085887, + 0.002047075890004635, + 0.004581214860081673, + 0.002148302271962166, + 0.002044173190370202, + 0.003558618947863579, + 0.004030121024698019, + 0.005312254652380943, + 0.003556115087121725, + 0.005834994837641716, + 0.004081265535205603, + 0.0040107611566782, + 0.007194543723016977, + 0.017355902120471, + 0.015204943716526031, + 0.012868874706327915, + 0.008363718166947365, + 0.009026695042848587, + 0.010388549417257309, + 0.013325137086212635, + 0.010730593465268612, + 0.012418926693499088, + 0.01055371668189764, + 0.010283960960805416, + 0.013483883813023567, + 0.01466184202581644, + 0.011709819547832012, + 0.007380749098956585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.0032055205665528774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864359550178051 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.00864359550178051, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864359550178051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08102472871541977, + 0.029143542051315308, + 0.03461494296789169, + 0.11223018914461136, + 0.26995882391929626, + 0.03278547152876854, + 0.02200712263584137, + 0.027776356786489487, + 0.07957342267036438, + 0.31088536977767944 + ], + "credit_map": [ + 0.005325764417648315, + 0.006871441379189491, + 0.010074533522129059, + 0.010364538989961147, + 0.01406023558229208, + 0.00710724014788866, + 0.019798265770077705, + 0.007422700524330139, + 0.002950646448880434, + 0.002327647991478443, + 0.003339660121127963, + 0.007030653767287731, + 0.0038194325752556324, + 0.002286153845489025, + 0.004102582577615976, + 0.0032867651898413897, + 0.003949475008994341, + 0.0034791510552167892, + 0.004423927050083876, + 0.004865260794758797, + 0.0054412465542554855, + 0.00394925381988287, + 0.004199350718408823, + 0.004307279363274574, + 0.008881837129592896, + 0.01419448759406805, + 0.018862975761294365, + 0.03168036788702011, + 0.00705260643735528, + 0.0074277473613619804, + 0.012169171124696732, + 0.011961000971496105, + 0.030660508200526237, + 0.018212635070085526, + 0.047325506806373596, + 0.062269438058137894, + 0.02909570001065731, + 0.027337122708559036, + 0.029888106510043144, + 0.025169823318719864, + 0.0031822901219129562, + 0.002538517350330949, + 0.004469230771064758, + 0.004938839003443718, + 0.004923377651721239, + 0.003563143080100417, + 0.006203803699463606, + 0.0029662693850696087, + 0.0024115454871207476, + 0.0017776609165593982, + 0.00372204277664423, + 0.0033476180396974087, + 0.002620302839204669, + 0.002823305083438754, + 0.0030093411915004253, + 0.0022953073494136333, + 0.0020895302295684814, + 0.0034910570830106735, + 0.0033274756278842688, + 0.0058748009614646435, + 0.003773838747292757, + 0.0027721996884793043, + 0.003435140009969473, + 0.0030123135074973106, + 0.00521565368399024, + 0.011553190648555756, + 0.01067084539681673, + 0.010762455873191357, + 0.013114999048411846, + 0.0088738352060318, + 0.009792014956474304, + 0.009590433910489082, + 0.0361630953848362, + 0.04064534232020378, + 0.036189302802085876, + 0.03377458453178406, + 0.044082995504140854, + 0.0530334897339344, + 0.041047871112823486, + 0.025948679074645042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.00864359550178051 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a fixed high first-person point of view (POV) of a surveillance camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a deserted alley bathed in glaring afternoon sun with a slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a rough red clay brick with weathered cracks and mottled mortar residue spin and hurtle toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brick slam into the camera lens, shattering the glass in a dense radiating spiderweb of cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the frame fade to black in a violent flash and blinding white light after the impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026333075016736984 + ], + "opposite_logprob_tokens": [ + -3.6513330936431885 + ], + "logprob_reward": -0.026333075016736984, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026333075016736984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16213026642799377, + 0.0401541031897068, + 0.050352543592453, + 0.10119994729757309, + 0.16313454508781433, + 0.0423450693488121, + 0.05063614249229431, + 0.06998416036367416, + 0.0644620805978775, + 0.255601167678833 + ], + "credit_map": [ + 0.011122802272439003, + 0.013295620679855347, + 0.019810236990451813, + 0.019562264904379845, + 0.028439588844776154, + 0.015501667745411396, + 0.038043007254600525, + 0.01635509915649891, + 0.002839077031239867, + 0.00278448942117393, + 0.003793665673583746, + 0.006873567122966051, + 0.00826159119606018, + 0.004786871373653412, + 0.005778474733233452, + 0.0050363680347800255, + 0.004915294703096151, + 0.005022290162742138, + 0.005691272672265768, + 0.004436538089066744, + 0.008248941041529179, + 0.006443803198635578, + 0.007728108204901218, + 0.007866297848522663, + 0.006400099489837885, + 0.0076927971094846725, + 0.01565527729690075, + 0.02217712625861168, + 0.010964813642203808, + 0.01029315497726202, + 0.012240495532751083, + 0.01577618531882763, + 0.02233094722032547, + 0.014843576587736607, + 0.0242669265717268, + 0.030016766861081123, + 0.01768176257610321, + 0.01781843602657318, + 0.018267324194312096, + 0.01790880598127842, + 0.004281424451619387, + 0.0038054354954510927, + 0.0065009938552975655, + 0.005453509744256735, + 0.004735077265650034, + 0.004939899779856205, + 0.00663763377815485, + 0.00599109660834074, + 0.005111373960971832, + 0.004235594533383846, + 0.006166714243590832, + 0.007062655407935381, + 0.0036485660821199417, + 0.006761259865015745, + 0.013138464652001858, + 0.004511517006903887, + 0.008225181140005589, + 0.013586767949163914, + 0.009566144086420536, + 0.007552689872682095, + 0.008251097984611988, + 0.008509054780006409, + 0.009009137749671936, + 0.0052840933203697205, + 0.0066915820352733135, + 0.010638964362442493, + 0.007885824888944626, + 0.006700495723634958, + 0.005511685740202665, + 0.006001915317028761, + 0.006017180159687996, + 0.015014436095952988, + 0.023938223719596863, + 0.030082905665040016, + 0.03144136816263199, + 0.033540960401296616, + 0.03333758935332298, + 0.04605502262711525, + 0.029997901991009712, + 0.02720719575881958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.026333075016736984 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_9/index_tensor([2116, 2116])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_9/index_tensor([2116, 2116])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..11befe8158e619e521b2ea60336f9b7b656ba266 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_9/index_tensor([2116, 2116])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic slow-motion, low-angle tracking shot follows a figure on quad roller skates as they powerfully approach a concrete skateboard ramp. The camera tilts upward, capturing the moment they launch explosively into the air from the ramp's edge. The skater, with windswept burgundy hair and a focused expression, wears a loose black tank top, baggy denim shorts, and frayed white kneepads. At the peak of their ascent, they execute a perfect 360-degree spin, their body gracefully swirling as one hand decisively grasps the outer edge of their right skate for a stylish rear-end hold. A vibrant, sun-drenched cityscape serves as a stunning backdrop, the golden hour sun casting long shadows and creating brilliant lens flares across the frame. The scooters' polished chrome and translucent bright orange wheels gleam, while every ripple of their clothing and hair is rendered with surreal clarity. The entire scene is bathed in a warm, high-contrast color palette, creating an exhilarating, awe-inspiring atmosphere of athletic prowess and urban freedom.", + "seed": 24427, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_9/index_2116_seed_24427_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025433705374598502, + "vlm_logprob": -0.025433705374598502, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.025433705374598502, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01840594969689846 + ], + "opposite_logprob_tokens": [ + -4.643405914306641 + ], + "logprob_reward": -0.01840594969689846, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01840594969689846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2496708482503891, + 0.08892732113599777, + 0.08370114862918854, + 0.06603160500526428, + 0.0784524530172348, + 0.08926384150981903, + 0.06088976934552193, + 0.06232215464115143, + 0.13518811762332916, + 0.08555284887552261 + ], + "credit_map": [ + 0.03460758179426193, + 0.0322955921292305, + 0.020670246332883835, + 0.02904594875872135, + 0.028175100684165955, + 0.036753904074430466, + 0.03013315238058567, + 0.0379892997443676, + 0.008585934527218342, + 0.009702243842184544, + 0.004930845461785793, + 0.012237440794706345, + 0.013598577119410038, + 0.011529184877872467, + 0.019714297726750374, + 0.008628793992102146, + 0.010867019183933735, + 0.008042462170124054, + 0.010555148124694824, + 0.01225377433001995, + 0.009433104656636715, + 0.009873229078948498, + 0.010998079553246498, + 0.011678331531584263, + 0.007668334525078535, + 0.009142989292740822, + 0.0056095123291015625, + 0.011156604625284672, + 0.008611015044152737, + 0.004847849253565073, + 0.00889584980905056, + 0.010099449194967747, + 0.009002326987683773, + 0.008848140947520733, + 0.008445787243545055, + 0.010064024478197098, + 0.009558137506246567, + 0.016691626980900764, + 0.007703185081481934, + 0.008139223791658878, + 0.017490560188889503, + 0.008267753757536411, + 0.026741614565253258, + 0.008582686074078083, + 0.008864717558026314, + 0.005027896258980036, + 0.008150231093168259, + 0.00613838667050004, + 0.009231407195329666, + 0.005954779218882322, + 0.006716844160109758, + 0.006653065327554941, + 0.005297151394188404, + 0.00870794989168644, + 0.012921514920890331, + 0.00540705444291234, + 0.007302942685782909, + 0.008175509981811047, + 0.01148310024291277, + 0.008520081639289856, + 0.005899808369576931, + 0.006064192857593298, + 0.008360834792256355, + 0.006515680346637964, + 0.018890447914600372, + 0.015056014060974121, + 0.03265168517827988, + 0.015024291351437569, + 0.009675372391939163, + 0.014500333927571774, + 0.015380596742033958, + 0.014009363949298859, + 0.01110741589218378, + 0.00621767109259963, + 0.021026981994509697, + 0.010892423801124096, + 0.010808009654283524, + 0.005321592558175325, + 0.011767175048589706, + 0.008411573246121407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.01840594969689846 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160874892026186 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.0160874892026186, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160874892026186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2551170885562897, + 0.05099794641137123, + 0.09576262533664703, + 0.08211678266525269, + 0.06972721219062805, + 0.1294693648815155, + 0.06651780754327774, + 0.08868639916181564, + 0.12449299544095993, + 0.03711184114217758 + ], + "credit_map": [ + 0.03285316377878189, + 0.02650061808526516, + 0.020144334062933922, + 0.025336388498544693, + 0.0281334538012743, + 0.04287087172269821, + 0.03221394494175911, + 0.047064296901226044, + 0.004853263963013887, + 0.005585446488112211, + 0.0037488294765353203, + 0.005779118277132511, + 0.007689351681619883, + 0.007085804827511311, + 0.009712389670312405, + 0.006543735042214394, + 0.010145602747797966, + 0.007635443937033415, + 0.010610666126012802, + 0.01092391088604927, + 0.017014410346746445, + 0.01065252348780632, + 0.015452316962182522, + 0.013327737338840961, + 0.011376301757991314, + 0.014862310141324997, + 0.006409004796296358, + 0.00933169201016426, + 0.011147762648761272, + 0.0055757747031748295, + 0.010272430256009102, + 0.01314149983227253, + 0.007825563661754131, + 0.007524078246206045, + 0.009568502195179462, + 0.008476439863443375, + 0.010314028710126877, + 0.010841650888323784, + 0.007172491401433945, + 0.008004453964531422, + 0.030521994456648827, + 0.009304585866630077, + 0.04743887484073639, + 0.010854815132915974, + 0.006640409119427204, + 0.005959219299256802, + 0.009695379063487053, + 0.009054080583155155, + 0.009689954109489918, + 0.006955781951546669, + 0.006793143227696419, + 0.005468001123517752, + 0.008005205541849136, + 0.013276124373078346, + 0.010282466188073158, + 0.006047121249139309, + 0.013625581748783588, + 0.009485325776040554, + 0.015753794461488724, + 0.012262862175703049, + 0.008393210358917713, + 0.012966327369213104, + 0.007422605995088816, + 0.008776682429015636, + 0.019769206643104553, + 0.017632242292165756, + 0.024538064375519753, + 0.013541783206164837, + 0.009754898957908154, + 0.014720197767019272, + 0.01313379593193531, + 0.01140279695391655, + 0.004031685646623373, + 0.004441208206117153, + 0.006178789772093296, + 0.005833181086927652, + 0.004768410697579384, + 0.002760165138170123, + 0.005126606207340956, + 0.003971789963543415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0160874892026186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344221085309982 + ], + "opposite_logprob_tokens": [ + -3.6513442993164062 + ], + "logprob_reward": -0.026344221085309982, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344221085309982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19455336034297943, + 0.04297270253300667, + 0.06182863935828209, + 0.04824079945683479, + 0.11098646372556686, + 0.12218831479549408, + 0.06589920073747635, + 0.14723961055278778, + 0.14437173306941986, + 0.06171925365924835 + ], + "credit_map": [ + 0.033305082470178604, + 0.0268412958830595, + 0.017617151141166687, + 0.01796005852520466, + 0.020582720637321472, + 0.030108947306871414, + 0.0185193233191967, + 0.029618777334690094, + 0.0032583530992269516, + 0.0044905696995556355, + 0.0035280198790133, + 0.0048300111666321754, + 0.005781239364296198, + 0.006174310576170683, + 0.007832018658518791, + 0.007078176364302635, + 0.009716032072901726, + 0.006396810058504343, + 0.007842549122869968, + 0.007170174736529589, + 0.009336115792393684, + 0.005900202319025993, + 0.00792234018445015, + 0.007544418331235647, + 0.006030757911503315, + 0.009683126583695412, + 0.004432699643075466, + 0.004267659969627857, + 0.005805442575365305, + 0.0027059975545853376, + 0.003039985429495573, + 0.012275131419301033, + 0.010828431695699692, + 0.00951473694294691, + 0.01781865768134594, + 0.012966649606823921, + 0.01389055885374546, + 0.018109697848558426, + 0.012081621214747429, + 0.015776103362441063, + 0.029632413759827614, + 0.008747315965592861, + 0.0476837083697319, + 0.009419379755854607, + 0.004458851646631956, + 0.004346824251115322, + 0.008918301202356815, + 0.008981522172689438, + 0.010577747598290443, + 0.005519635509699583, + 0.008515847846865654, + 0.006298184394836426, + 0.00606872420758009, + 0.006431770510971546, + 0.015950335189700127, + 0.006536956876516342, + 0.020924562588334084, + 0.014523410238325596, + 0.028318719938397408, + 0.01898934878408909, + 0.014353080652654171, + 0.019244978204369545, + 0.014953305944800377, + 0.015932200476527214, + 0.018653295934200287, + 0.018544858321547508, + 0.02898441068828106, + 0.015528544783592224, + 0.013849704526364803, + 0.01604042761027813, + 0.01791795715689659, + 0.014852525666356087, + 0.009244098328053951, + 0.004608999006450176, + 0.01659819297492504, + 0.007125966250896454, + 0.004227906931191683, + 0.00393097847700119, + 0.005163006950169802, + 0.010820100083947182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.026344221085309982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291101679205894 + ], + "opposite_logprob_tokens": [ + -3.7732911109924316 + ], + "logprob_reward": -0.023291101679205894, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291101679205894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09588008373975754, + 0.05369025468826294, + 0.19884999096393585, + 0.07671158015727997, + 0.043680787086486816, + 0.16293078660964966, + 0.03936343640089035, + 0.025512143969535828, + 0.23366335034370422, + 0.06971753388643265 + ], + "credit_map": [ + 0.0070534381084144115, + 0.010233508422970772, + 0.006113887764513493, + 0.012419267557561398, + 0.012585543096065521, + 0.015604732558131218, + 0.013011281378567219, + 0.01885843649506569, + 0.005526791326701641, + 0.006489970255643129, + 0.004068935755640268, + 0.007394867483526468, + 0.0067328172735869884, + 0.007151027675718069, + 0.01001462060958147, + 0.0063112275674939156, + 0.02217530645430088, + 0.017680294811725616, + 0.0235634483397007, + 0.022138351574540138, + 0.033155329525470734, + 0.021755004301667213, + 0.03156398981809616, + 0.026818281039595604, + 0.013632318936288357, + 0.011122181080281734, + 0.007599071133881807, + 0.01565701887011528, + 0.005473822355270386, + 0.0056734890677034855, + 0.00851812120527029, + 0.009035563096404076, + 0.004902626387774944, + 0.0048430426977574825, + 0.006322953850030899, + 0.00445943046361208, + 0.004990878980606794, + 0.0065827323123812675, + 0.0032152787316590548, + 0.008363844826817513, + 0.041176579892635345, + 0.01039912085980177, + 0.0658416822552681, + 0.011590070091187954, + 0.004281037021428347, + 0.004945233929902315, + 0.012513285502791405, + 0.012183800339698792, + 0.00506970239803195, + 0.003042345168069005, + 0.006237856112420559, + 0.004036263562738895, + 0.003746563568711281, + 0.007338572759181261, + 0.006277772132307291, + 0.0036143632605671883, + 0.0037448599468916655, + 0.0036628544330596924, + 0.004363458137959242, + 0.002793949330225587, + 0.0021458608098328114, + 0.004252491984516382, + 0.002321191830560565, + 0.0022274816874414682, + 0.0311783105134964, + 0.0217240322381258, + 0.05697237327694893, + 0.022614458575844765, + 0.016111891716718674, + 0.027268869802355766, + 0.02817816473543644, + 0.029615283012390137, + 0.007485064212232828, + 0.005387819837778807, + 0.02190079353749752, + 0.008386414498090744, + 0.004804909694939852, + 0.004563033115118742, + 0.0056069460697472095, + 0.011582554318010807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.023291101679205894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303976520895958 + ], + "opposite_logprob_tokens": [ + -3.1680397987365723 + ], + "logprob_reward": -0.04303976520895958, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303976520895958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18835397064685822, + 0.08982600271701813, + 0.05907952040433884, + 0.044252634048461914, + 0.05033960938453674, + 0.18759740889072418, + 0.05938120558857918, + 0.18092650175094604, + 0.09689998626708984, + 0.04334311932325363 + ], + "credit_map": [ + 0.030164547264575958, + 0.030357059091329575, + 0.018693814054131508, + 0.020917529240250587, + 0.020299134775996208, + 0.024115227162837982, + 0.021754883229732513, + 0.022051766514778137, + 0.008404216729104519, + 0.008813604712486267, + 0.005524115636944771, + 0.014015819877386093, + 0.010229906998574734, + 0.012887061573565006, + 0.02243071421980858, + 0.007520565763115883, + 0.006188347470015287, + 0.006647949572652578, + 0.006743578240275383, + 0.007581617683172226, + 0.011296754702925682, + 0.007331987377256155, + 0.008048691786825657, + 0.005240590311586857, + 0.005832629278302193, + 0.005212361458688974, + 0.0044905575923621655, + 0.006871687714010477, + 0.006268010474741459, + 0.003846280975267291, + 0.0069963219575583935, + 0.004734783433377743, + 0.008272322826087475, + 0.007947935722768307, + 0.006625049747526646, + 0.005799781996756792, + 0.0059008426032960415, + 0.006003380753099918, + 0.004616143647581339, + 0.005174152087420225, + 0.04720424488186836, + 0.011840729042887688, + 0.0766647681593895, + 0.01329134963452816, + 0.004706902429461479, + 0.004991896450519562, + 0.014761070720851421, + 0.014136440120637417, + 0.008444496430456638, + 0.0059713684022426605, + 0.008097351528704166, + 0.005922602955251932, + 0.005945716053247452, + 0.006966195069253445, + 0.01244023721665144, + 0.005593240261077881, + 0.02493867836892605, + 0.019434794783592224, + 0.030283333733677864, + 0.0224149227142334, + 0.016121508553624153, + 0.02595842070877552, + 0.019189588725566864, + 0.022585246711969376, + 0.012214476242661476, + 0.00925198569893837, + 0.024678071960806847, + 0.011219354346394539, + 0.006666639354079962, + 0.011538831517100334, + 0.010875638574361801, + 0.01045498251914978, + 0.006655177101492882, + 0.004578760359436274, + 0.007084714248776436, + 0.0059156944043934345, + 0.005832758266478777, + 0.003915132489055395, + 0.004880385473370552, + 0.00448049558326602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.04303976520895958 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic slow-motion, low-angle tracking shot follows a figure on quad roller skates as they powerfully approach a concrete skateboard ramp. The camera tilts upward, capturing the moment they launch explosively into the air from the ramp's edge. The skater, with windswept burgundy hair and a focused expression, wears a loose black tank top, baggy denim shorts, and frayed white kneepads. At the peak of their ascent, they execute a perfect 360-degree spin, their body gracefully swirling as one hand decisively grasps the outer edge of their right skate for a stylish rear-end hold. A vibrant, sun-drenched cityscape serves as a stunning backdrop, the golden hour sun casting long shadows and creating brilliant lens flares across the frame. The scooters' polished chrome and translucent bright orange wheels gleam, while every ripple of their clothing and hair is rendered with surreal clarity. The entire scene is bathed in a warm, high-contrast color palette, creating an exhilarating, awe-inspiring atmosphere of athletic prowess and urban freedom.", + "seed": 24428, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_9/index_2116_seed_24428_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02631458006799221, + "vlm_logprob": -0.02631458006799221, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.02631458006799221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020831836387515068 + ], + "opposite_logprob_tokens": [ + -4.645832061767578 + ], + "logprob_reward": -0.020831836387515068, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.020831836387515068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21336157619953156, + 0.06819632649421692, + 0.0781659334897995, + 0.1482667475938797, + 0.11689577251672745, + 0.08399147540330887, + 0.0719386413693428, + 0.05125245451927185, + 0.05697981268167496, + 0.1109512597322464 + ], + "credit_map": [ + 0.02258882485330105, + 0.032901570200920105, + 0.032122902572155, + 0.033001989126205444, + 0.013110917992889881, + 0.017954161390662193, + 0.02889646589756012, + 0.03278474509716034, + 0.006857156753540039, + 0.0075393058359622955, + 0.011026532389223576, + 0.008911040611565113, + 0.0051387823186814785, + 0.004565301351249218, + 0.011659166775643826, + 0.012499039061367512, + 0.010777700692415237, + 0.007144417613744736, + 0.011606791988015175, + 0.0122707299888134, + 0.007708970457315445, + 0.011096753180027008, + 0.009164441376924515, + 0.008396130055189133, + 0.01708725467324257, + 0.01357613131403923, + 0.026706799864768982, + 0.02091061696410179, + 0.015570802614092827, + 0.01609278842806816, + 0.02389952540397644, + 0.0144228246062994, + 0.012342027388513088, + 0.010283119976520538, + 0.015536155551671982, + 0.024386100471019745, + 0.01584845967590809, + 0.011965987272560596, + 0.014770218171179295, + 0.011763710528612137, + 0.012131002731621265, + 0.010616213083267212, + 0.012827150523662567, + 0.009977925568819046, + 0.011234438978135586, + 0.011090005747973919, + 0.007396514993160963, + 0.008718221448361874, + 0.006744297221302986, + 0.01002187468111515, + 0.01021319255232811, + 0.008745688013732433, + 0.006629958748817444, + 0.01344763208180666, + 0.007086189929395914, + 0.009049812331795692, + 0.0051991501823067665, + 0.0058689177967607975, + 0.009241956286132336, + 0.005513747688382864, + 0.006125819403678179, + 0.0072860438376665115, + 0.0066144890151917934, + 0.005402329843491316, + 0.00609790813177824, + 0.011324924416840076, + 0.006366241257637739, + 0.006673693191260099, + 0.005720684304833412, + 0.0067831422202289104, + 0.0076145511120557785, + 0.0063986703753471375, + 0.010405076667666435, + 0.010725457221269608, + 0.017081085592508316, + 0.02075151540338993, + 0.009284842759370804, + 0.012821810320019722, + 0.011984976008534431, + 0.017896493896842003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.020831836387515068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020610567182302475 + ], + "opposite_logprob_tokens": [ + -3.8956105709075928 + ], + "logprob_reward": -0.020610567182302475, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020610567182302475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17807435989379883, + 0.0380910225212574, + 0.027458110824227333, + 0.09798987954854965, + 0.31069067120552063, + 0.04260918125510216, + 0.05411607772111893, + 0.053302571177482605, + 0.10672207176685333, + 0.09094604849815369 + ], + "credit_map": [ + 0.010668561793863773, + 0.02047811448574066, + 0.024031024426221848, + 0.03521806746721268, + 0.007674186024814844, + 0.013638017699122429, + 0.026304466649889946, + 0.040061913430690765, + 0.0047550383023917675, + 0.007202497683465481, + 0.008829803206026554, + 0.004302911926060915, + 0.001593144959770143, + 0.003159809857606888, + 0.004622104577720165, + 0.003625712124630809, + 0.0051321242935955524, + 0.0021703268866986036, + 0.003322611330077052, + 0.005119992420077324, + 0.002259586937725544, + 0.003181475680321455, + 0.0035742754116654396, + 0.0026977199595421553, + 0.011450622230768204, + 0.01864008419215679, + 0.009160910733044147, + 0.010390362702310085, + 0.016256574541330338, + 0.012196225114166737, + 0.012415789067745209, + 0.007479314226657152, + 0.041610561311244965, + 0.025764234364032745, + 0.042806778103113174, + 0.04741350933909416, + 0.041833795607089996, + 0.037775684148073196, + 0.0227190013974905, + 0.050767093896865845, + 0.008543086238205433, + 0.005646981298923492, + 0.005173867102712393, + 0.0038906806148588657, + 0.004641308914870024, + 0.006677902303636074, + 0.00327102979645133, + 0.004764324054121971, + 0.005829717963933945, + 0.0054001943208277225, + 0.006808983162045479, + 0.00870504416525364, + 0.004924742504954338, + 0.00945156067609787, + 0.005555382464081049, + 0.007440449669957161, + 0.004912637174129486, + 0.006531768944114447, + 0.00872358214110136, + 0.004785272758454084, + 0.010214177891612053, + 0.007863829843699932, + 0.006478689610958099, + 0.0037926118820905685, + 0.01236663106828928, + 0.030619388446211815, + 0.008486289530992508, + 0.01345356460660696, + 0.008845050819218159, + 0.013327552005648613, + 0.009207217954099178, + 0.010416370816528797, + 0.008243903517723083, + 0.0070808399468660355, + 0.01319869700819254, + 0.020051712170243263, + 0.007370883133262396, + 0.010056397877633572, + 0.009607715532183647, + 0.015335896983742714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.020610567182302475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287957534193993 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023287957534193993, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287957534193993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38621315360069275, + 0.047879137098789215, + 0.037766702473163605, + 0.04758864641189575, + 0.1329786777496338, + 0.09498754143714905, + 0.04132963716983795, + 0.05249559506773949, + 0.046367861330509186, + 0.11239296942949295 + ], + "credit_map": [ + 0.02280164323747158, + 0.04640898481011391, + 0.04694254696369171, + 0.08111002296209335, + 0.015674885362386703, + 0.029271308332681656, + 0.06022118031978607, + 0.08378259837627411, + 0.005687689874321222, + 0.007072650361806154, + 0.007757244631648064, + 0.0060945069417357445, + 0.0029588646721094847, + 0.005758419632911682, + 0.007402616087347269, + 0.005147148855030537, + 0.007375750690698624, + 0.0030426804441958666, + 0.005484274588525295, + 0.005762494169175625, + 0.0030650468543171883, + 0.0036055054515600204, + 0.004561629146337509, + 0.004869320895522833, + 0.007444141898304224, + 0.006375525612384081, + 0.005833801813423634, + 0.004827405791729689, + 0.00790749117732048, + 0.007013467140495777, + 0.004209728445857763, + 0.003977085929363966, + 0.011143428273499012, + 0.009804567322134972, + 0.024139702320098877, + 0.019330991432070732, + 0.018877945840358734, + 0.017366081476211548, + 0.014942005276679993, + 0.017373953014612198, + 0.010287484154105186, + 0.010886048898100853, + 0.015667377039790154, + 0.014565957710146904, + 0.014963269233703613, + 0.01017250120639801, + 0.00862574391067028, + 0.009819152764976025, + 0.005231465213000774, + 0.005559428595006466, + 0.006089082919061184, + 0.005194447003304958, + 0.0062685515731573105, + 0.004893890582025051, + 0.003997573163360357, + 0.0040951985865831375, + 0.005459912586957216, + 0.0073659177869558334, + 0.004782571457326412, + 0.004888869356364012, + 0.01190326176583767, + 0.007024870254099369, + 0.006888631731271744, + 0.00418156199157238, + 0.004716560244560242, + 0.010390400886535645, + 0.004300444852560759, + 0.005026084836572409, + 0.004671864677220583, + 0.0055826641619205475, + 0.005305611994117498, + 0.0063742296770215034, + 0.010526061058044434, + 0.009589673951268196, + 0.016006793826818466, + 0.025366785004734993, + 0.009010693989694118, + 0.01313991378992796, + 0.011023838073015213, + 0.017729200422763824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.023287957534193993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048666588962078094 + ], + "opposite_logprob_tokens": [ + -3.048666477203369 + ], + "logprob_reward": -0.048666588962078094, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048666588962078094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4646386504173279, + 0.05355684086680412, + 0.051507025957107544, + 0.09021203219890594, + 0.08774302154779434, + 0.06570902466773987, + 0.030144982039928436, + 0.053089629858732224, + 0.03577740862965584, + 0.0676213875412941 + ], + "credit_map": [ + 0.024204136803746223, + 0.05312677100300789, + 0.07093755900859833, + 0.10931940376758575, + 0.020932208746671677, + 0.0318940132856369, + 0.06609713286161423, + 0.08812736719846725, + 0.006627361290156841, + 0.008334161713719368, + 0.012212809175252914, + 0.005883685313165188, + 0.0021789551246911287, + 0.004088469780981541, + 0.007127175107598305, + 0.007104223594069481, + 0.006100012455135584, + 0.005412642378360033, + 0.007191166281700134, + 0.011411622166633606, + 0.005125152412801981, + 0.006579007487744093, + 0.005225463770329952, + 0.004461953882128, + 0.009992255829274654, + 0.009791059419512749, + 0.016571111977100372, + 0.011286484077572823, + 0.010746398009359837, + 0.0108407037332654, + 0.013600095175206661, + 0.007383920252323151, + 0.007686277385801077, + 0.00797833502292633, + 0.014535157941281796, + 0.011527666822075844, + 0.012914206832647324, + 0.011192495934665203, + 0.010816450230777264, + 0.011092426255345345, + 0.011951928958296776, + 0.008745293132960796, + 0.008279716596007347, + 0.00674132676795125, + 0.0076297312043607235, + 0.009772797115147114, + 0.004981133621186018, + 0.007607089355587959, + 0.003020013915374875, + 0.0037612230516970158, + 0.004125027451664209, + 0.003951443824917078, + 0.003666120348498225, + 0.00536734564229846, + 0.003347852034494281, + 0.00290595181286335, + 0.0047880783677101135, + 0.005732642486691475, + 0.009601502679288387, + 0.005950879771262407, + 0.007967204786837101, + 0.007113854866474867, + 0.007790892850607634, + 0.004144574515521526, + 0.003993474878370762, + 0.003998918924480677, + 0.004769671708345413, + 0.004714653827250004, + 0.0036663373466581106, + 0.004206581506878138, + 0.006110915448516607, + 0.004316851496696472, + 0.00652235746383667, + 0.005713800899684429, + 0.010446728207170963, + 0.01487382035702467, + 0.004788442514836788, + 0.0072391098365187645, + 0.007483620196580887, + 0.010553502477705479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.048666588962078094 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175950273871422 + ], + "opposite_logprob_tokens": [ + -4.018176078796387 + ], + "logprob_reward": -0.018175950273871422, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018175950273871422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10561975091695786, + 0.06444745510816574, + 0.08502698689699173, + 0.09069021046161652, + 0.1684274673461914, + 0.11134634166955948, + 0.058292463421821594, + 0.05201513320207596, + 0.1394360363483429, + 0.12469814717769623 + ], + "credit_map": [ + 0.013388043269515038, + 0.013708606362342834, + 0.016567904502153397, + 0.01604427397251129, + 0.006433503236621618, + 0.005542448256164789, + 0.018435567617416382, + 0.015499399043619633, + 0.009157905355095863, + 0.013313992880284786, + 0.0082305409014225, + 0.008197558112442493, + 0.003958715125918388, + 0.0063807163387537, + 0.009808670729398727, + 0.005399356596171856, + 0.01320281345397234, + 0.007149382494390011, + 0.014962468296289444, + 0.01171237975358963, + 0.010680069215595722, + 0.008715815842151642, + 0.010090215131640434, + 0.008513842709362507, + 0.009132087230682373, + 0.011866634711623192, + 0.012462196871638298, + 0.010106011293828487, + 0.011821639724075794, + 0.011003004387021065, + 0.01434510201215744, + 0.009953534230589867, + 0.024379733949899673, + 0.012206606566905975, + 0.024540258571505547, + 0.021642765030264854, + 0.01675739511847496, + 0.024852199479937553, + 0.013955565169453621, + 0.030092937871813774, + 0.012315751053392887, + 0.00982583686709404, + 0.01702403649687767, + 0.015064766630530357, + 0.01506161130964756, + 0.016613639891147614, + 0.010581056587398052, + 0.014859646558761597, + 0.0067500947043299675, + 0.008397591300308704, + 0.004217643290758133, + 0.010418389923870564, + 0.007026515435427427, + 0.00848887488245964, + 0.006613590754568577, + 0.006379762664437294, + 0.005155231803655624, + 0.0038493014872074127, + 0.010566330514848232, + 0.004391007591038942, + 0.0091360779479146, + 0.0076234652660787106, + 0.0045582614839077, + 0.006735454313457012, + 0.017355814576148987, + 0.035152483731508255, + 0.013363218866288662, + 0.015288585796952248, + 0.01271701417863369, + 0.01697542332112789, + 0.011963115073740482, + 0.01662038080394268, + 0.011197429150342941, + 0.011732586659491062, + 0.017668064683675766, + 0.024609042331576347, + 0.012019727379083633, + 0.014149384573101997, + 0.012086833827197552, + 0.02123507298529148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.018175950273871422 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic slow-motion, low-angle tracking shot follows a figure on quad roller skates as they powerfully approach a concrete skateboard ramp. The camera tilts upward, capturing the moment they launch explosively into the air from the ramp's edge. The skater, with windswept burgundy hair and a focused expression, wears a loose black tank top, baggy denim shorts, and frayed white kneepads. At the peak of their ascent, they execute a perfect 360-degree spin, their body gracefully swirling as one hand decisively grasps the outer edge of their right skate for a stylish rear-end hold. A vibrant, sun-drenched cityscape serves as a stunning backdrop, the golden hour sun casting long shadows and creating brilliant lens flares across the frame. The scooters' polished chrome and translucent bright orange wheels gleam, while every ripple of their clothing and hair is rendered with surreal clarity. The entire scene is bathed in a warm, high-contrast color palette, creating an exhilarating, awe-inspiring atmosphere of athletic prowess and urban freedom.", + "seed": 24429, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_9/index_2116_seed_24429_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02419322058558464, + "vlm_logprob": -0.02419322058558464, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02419322058558464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0184225682169199 + ], + "opposite_logprob_tokens": [ + -4.768422603607178 + ], + "logprob_reward": -0.0184225682169199, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0184225682169199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4882444143295288, + 0.13051745295524597, + 0.04863330349326134, + 0.03621170297265053, + 0.031197674572467804, + 0.061047833412885666, + 0.035237111151218414, + 0.05495933070778847, + 0.038759637624025345, + 0.07519159466028214 + ], + "credit_map": [ + 0.025088269263505936, + 0.1663600653409958, + 0.15456749498844147, + 0.038155555725097656, + 0.031214335933327675, + 0.03357097879052162, + 0.021399302408099174, + 0.017888391390442848, + 0.015056289732456207, + 0.016104063019156456, + 0.01844531111419201, + 0.011259561404585838, + 0.026698192581534386, + 0.01710362359881401, + 0.010463053360581398, + 0.015387343242764473, + 0.004119225777685642, + 0.006352545693516731, + 0.005385509692132473, + 0.006872947793453932, + 0.00519539276137948, + 0.007573965471237898, + 0.0064501105807721615, + 0.0066836001351475716, + 0.004810212645679712, + 0.00410108920186758, + 0.004630636423826218, + 0.003938903100788593, + 0.004574991762638092, + 0.003909531515091658, + 0.004788955207914114, + 0.0054573798552155495, + 0.005771669093519449, + 0.004352125804871321, + 0.002819003537297249, + 0.0035573483910411596, + 0.005493638571351767, + 0.003938814625144005, + 0.0023085209541022778, + 0.0029565468430519104, + 0.0080659668892622, + 0.00799123290926218, + 0.010816949419677258, + 0.008016914129257202, + 0.00871356949210167, + 0.006128635723143816, + 0.006688856054097414, + 0.004625702276825905, + 0.0034200772643089294, + 0.003595654619857669, + 0.005045002792030573, + 0.005226758774369955, + 0.0045792884193360806, + 0.0032528734300285578, + 0.006454110611230135, + 0.003663343843072653, + 0.005873930640518665, + 0.006707490887492895, + 0.006254760082811117, + 0.0050965421833097935, + 0.007313980255275965, + 0.0042683761566877365, + 0.013083902187645435, + 0.0063603403978049755, + 0.007860478945076466, + 0.00360234547406435, + 0.005090284161269665, + 0.004768141079694033, + 0.0029227205086499453, + 0.005417707841843367, + 0.0035544501151889563, + 0.005543501581996679, + 0.006467511411756277, + 0.007581224199384451, + 0.01201760582625866, + 0.009435173124074936, + 0.010053794831037521, + 0.008983409963548183, + 0.0099398884922266, + 0.010712970979511738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0184225682169199 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558897957205772 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.012558897957205772, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558897957205772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7474789619445801, + 0.06784143298864365, + 0.02439173310995102, + 0.006981076672673225, + 0.03677435591816902, + 0.010912981815636158, + 0.009709588252007961, + 0.01650458201766014, + 0.022292042151093483, + 0.05711321160197258 + ], + "credit_map": [ + 0.03246733918786049, + 0.30014437437057495, + 0.2820620536804199, + 0.04938410222530365, + 0.04426206275820732, + 0.015787746757268906, + 0.01197206974029541, + 0.011399268172681332, + 0.01065770909190178, + 0.007045425940304995, + 0.008163901045918465, + 0.00676969438791275, + 0.01303904503583908, + 0.007974503561854362, + 0.005755315069109201, + 0.008435847237706184, + 0.0024810864124447107, + 0.0021941643208265305, + 0.0018903663149103522, + 0.0024968297220766544, + 0.0031201117672026157, + 0.004005743190646172, + 0.003964466042816639, + 0.0042389677837491035, + 0.0013058624463155866, + 0.0010782493045553565, + 0.0009365968289785087, + 0.0008914382196962833, + 0.0007964897085912526, + 0.0007552088354714215, + 0.000601775711402297, + 0.0006154562579467893, + 0.007456151302903891, + 0.0040671019814908504, + 0.004565106704831123, + 0.003160981461405754, + 0.004598402883857489, + 0.004866486415266991, + 0.003560685785487294, + 0.004499444272369146, + 0.0014307288220152259, + 0.0018285911064594984, + 0.0014806739054620266, + 0.0013770583318546414, + 0.0015217299805954099, + 0.0011786327231675386, + 0.0011688444064930081, + 0.0009267234709113836, + 0.0013519697822630405, + 0.0010339852888137102, + 0.0011106636375188828, + 0.0012266230769455433, + 0.001360709429718554, + 0.0010069477139040828, + 0.0018110191449522972, + 0.0008076708181761205, + 0.0016332367667928338, + 0.0016010381514206529, + 0.0015283898683264852, + 0.0014674336416646838, + 0.0023162292782217264, + 0.0011709064710885286, + 0.004801989533007145, + 0.0019853597041219473, + 0.0023225124459713697, + 0.003695013467222452, + 0.0021028558257967234, + 0.004394535440951586, + 0.0025786461774259806, + 0.0021089992951601744, + 0.0028947305399924517, + 0.00219474988989532, + 0.003834766335785389, + 0.008493438363075256, + 0.0052507733926177025, + 0.007815366610884666, + 0.008956380188465118, + 0.006748018320649862, + 0.008179698139429092, + 0.007834777235984802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012558897957205772 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30863815546035767, + 0.22616852819919586, + 0.053819116204977036, + 0.03549408167600632, + 0.07278923690319061, + 0.09806998074054718, + 0.058718983083963394, + 0.030693786218762398, + 0.07631605863571167, + 0.03929201513528824 + ], + "credit_map": [ + 0.025788050144910812, + 0.08322864770889282, + 0.07451124489307404, + 0.04022357985377312, + 0.02416682243347168, + 0.02607487514615059, + 0.019610418006777763, + 0.015034526586532593, + 0.03497931733727455, + 0.023266717791557312, + 0.018786681815981865, + 0.02157295122742653, + 0.050270676612854004, + 0.025862257927656174, + 0.020732760429382324, + 0.03069717437028885, + 0.005597169045358896, + 0.004910754971206188, + 0.0067008319310843945, + 0.005542354192584753, + 0.009357059374451637, + 0.008786514401435852, + 0.007126296870410442, + 0.00579813402146101, + 0.0043086581863462925, + 0.003896870417520404, + 0.006244544871151447, + 0.006315844599157572, + 0.004272950813174248, + 0.0033430466428399086, + 0.0038332308176904917, + 0.003278936492279172, + 0.008372794836759567, + 0.011709355749189854, + 0.005396557971835136, + 0.010136906988918781, + 0.013185334391891956, + 0.008633355610072613, + 0.010492505505681038, + 0.004862423986196518, + 0.011426553130149841, + 0.013647213578224182, + 0.016741925850510597, + 0.010066147893667221, + 0.015789847820997238, + 0.00983953196555376, + 0.010772136971354485, + 0.00978662259876728, + 0.004910934250801802, + 0.0055901953019201756, + 0.0077602132223546505, + 0.007153579965233803, + 0.00906308088451624, + 0.004739310592412949, + 0.013205345720052719, + 0.006296320352703333, + 0.0032424810342490673, + 0.00414813868701458, + 0.004069335293024778, + 0.003325797151774168, + 0.0040984321385622025, + 0.0025749648921191692, + 0.005529959686100483, + 0.0037046787329018116, + 0.008467255160212517, + 0.009340522810816765, + 0.012849930673837662, + 0.011275780387222767, + 0.008017676882445812, + 0.00974662322551012, + 0.008389323018491268, + 0.008228945545852184, + 0.0032440030481666327, + 0.005719915498048067, + 0.006226138677448034, + 0.006902788765728474, + 0.003060710383579135, + 0.004854254890233278, + 0.005224335938692093, + 0.00405986700206995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02058347500860691 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306785389780998 + ], + "opposite_logprob_tokens": [ + -3.1680679321289062 + ], + "logprob_reward": -0.04306785389780998, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306785389780998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5212676525115967, + 0.15288019180297852, + 0.0514826662838459, + 0.026434043422341347, + 0.05884571745991707, + 0.03970301151275635, + 0.01889694295823574, + 0.05034728720784187, + 0.05071888491511345, + 0.029423583298921585 + ], + "credit_map": [ + 0.02496377006173134, + 0.19380587339401245, + 0.18359984457492828, + 0.0381888709962368, + 0.03068409487605095, + 0.019250810146331787, + 0.016540206968784332, + 0.014234175905585289, + 0.024993734434247017, + 0.01431509479880333, + 0.02936253324151039, + 0.014187186025083065, + 0.02366967871785164, + 0.020340044051408768, + 0.009208190254867077, + 0.01680372841656208, + 0.004492142237722874, + 0.0049542998895049095, + 0.007642155513167381, + 0.005638725124299526, + 0.006155874580144882, + 0.007287926040589809, + 0.007155694533139467, + 0.00815584883093834, + 0.004064240492880344, + 0.004455562215298414, + 0.004259495995938778, + 0.0037502374034374952, + 0.0027012459468096495, + 0.002402092097327113, + 0.0024520617444068193, + 0.0023491065949201584, + 0.010713853873312473, + 0.008358968421816826, + 0.0062989601865410805, + 0.005693444050848484, + 0.009093091823160648, + 0.0074747842736542225, + 0.005321027711033821, + 0.0058915880508720875, + 0.004786199890077114, + 0.004290977027267218, + 0.01152096502482891, + 0.004018179140985012, + 0.003230166155844927, + 0.004082628525793552, + 0.003847025567665696, + 0.003926871344447136, + 0.002416773932054639, + 0.0018386851297691464, + 0.002464398043230176, + 0.0022719588596373796, + 0.0024569430388510227, + 0.0019039935432374477, + 0.003808060195297003, + 0.001736131845973432, + 0.005209123250097036, + 0.00467001274228096, + 0.005317205563187599, + 0.004425561521202326, + 0.0077131702564656734, + 0.003769293427467346, + 0.014685401692986488, + 0.0045575168915092945, + 0.008199945092201233, + 0.005462304688990116, + 0.005928593222051859, + 0.007469454314559698, + 0.004888236988335848, + 0.00564582971855998, + 0.008056922815740108, + 0.005067597609013319, + 0.003564795246347785, + 0.0034893096890300512, + 0.005110876634716988, + 0.0036514855455607176, + 0.002783410483971238, + 0.0028751781210303307, + 0.004117995034903288, + 0.0038305334746837616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04306785389780998 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026333307847380638 + ], + "opposite_logprob_tokens": [ + -3.6513333320617676 + ], + "logprob_reward": -0.026333307847380638, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026333307847380638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7432892322540283, + 0.06317470222711563, + 0.030384063720703125, + 0.010187312960624695, + 0.03982890769839287, + 0.06970951706171036, + 0.013334522023797035, + 0.018492668867111206, + 0.005346864461898804, + 0.006252204533666372 + ], + "credit_map": [ + 0.031223613768815994, + 0.3036814332008362, + 0.2847716212272644, + 0.0459514856338501, + 0.044093649834394455, + 0.013344951905310154, + 0.00996884610503912, + 0.01025363802909851, + 0.01186006423085928, + 0.006963218096643686, + 0.011319665238261223, + 0.006158468313515186, + 0.008525705896317959, + 0.007543390616774559, + 0.004713996313512325, + 0.006090191658586264, + 0.0020972273778170347, + 0.003577139461413026, + 0.002523403149098158, + 0.003656388260424137, + 0.0027280794456601143, + 0.0032693403773009777, + 0.005895029287785292, + 0.006637456826865673, + 0.001045350800268352, + 0.0013559681829065084, + 0.0016419434687122703, + 0.0011007206048816442, + 0.001026785932481289, + 0.0009535459685139358, + 0.0016784695908427238, + 0.0013845288194715977, + 0.006843904964625835, + 0.005196407902985811, + 0.0037899247836321592, + 0.004322508350014687, + 0.0075430735014379025, + 0.005062333308160305, + 0.00396711565554142, + 0.0031036369036883116, + 0.00934501551091671, + 0.008251402527093887, + 0.015678783878684044, + 0.007976466789841652, + 0.008984419517219067, + 0.007265741005539894, + 0.007021671626716852, + 0.005186018999665976, + 0.0014709490351378918, + 0.0012417505495250225, + 0.002083919709548354, + 0.001652401639148593, + 0.0018573412671685219, + 0.001335833570919931, + 0.0022074843291193247, + 0.0014848419232293963, + 0.0020723878405988216, + 0.00267436308786273, + 0.001592913642525673, + 0.0013364461483433843, + 0.002601229352876544, + 0.0018667641561478376, + 0.003811499336734414, + 0.0025370665825903416, + 0.000903763750102371, + 0.0005962132127024233, + 0.0008697831071913242, + 0.000533413840457797, + 0.00045865497668273747, + 0.0006065124762244523, + 0.0005190619267523289, + 0.0008594607934355736, + 0.00048821576638147235, + 0.0009425932657904923, + 0.0011206208728253841, + 0.0011291112750768661, + 0.0004741839657071978, + 0.0007793506374582648, + 0.0006467506173066795, + 0.0006713784532621503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.026333307847380638 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic slow-motion, low-angle tracking shot follows a figure on quad roller skates as they powerfully approach a concrete skateboard ramp. The camera tilts upward, capturing the moment they launch explosively into the air from the ramp's edge. The skater, with windswept burgundy hair and a focused expression, wears a loose black tank top, baggy denim shorts, and frayed white kneepads. At the peak of their ascent, they execute a perfect 360-degree spin, their body gracefully swirling as one hand decisively grasps the outer edge of their right skate for a stylish rear-end hold. A vibrant, sun-drenched cityscape serves as a stunning backdrop, the golden hour sun casting long shadows and creating brilliant lens flares across the frame. The scooters' polished chrome and translucent bright orange wheels gleam, while every ripple of their clothing and hair is rendered with surreal clarity. The entire scene is bathed in a warm, high-contrast color palette, creating an exhilarating, awe-inspiring atmosphere of athletic prowess and urban freedom.", + "seed": 24430, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000011/group_9/index_2116_seed_24430_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02864092756062746, + "vlm_logprob": -0.02864092756062746, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02864092756062746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029191240668296814 + ], + "opposite_logprob_tokens": [ + -4.779191017150879 + ], + "logprob_reward": -0.029191240668296814, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.029191240668296814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04606550186872482, + 0.02610872872173786, + 0.04479025676846504, + 0.4619494378566742, + 0.02466641366481781, + 0.09464254975318909, + 0.11849993467330933, + 0.09395888447761536, + 0.03255114704370499, + 0.05676720663905144 + ], + "credit_map": [ + 0.0038421754725277424, + 0.006566761527210474, + 0.005337852519005537, + 0.007273200433701277, + 0.003758043050765991, + 0.003324355697259307, + 0.010158112272620201, + 0.00580500066280365, + 0.0034156786277890205, + 0.002961576683446765, + 0.004031820222735405, + 0.0017653501126915216, + 0.00410292437300086, + 0.002968729240819812, + 0.002726504812017083, + 0.004136145580559969, + 0.006311351899057627, + 0.00896711740642786, + 0.006365692242980003, + 0.004003798123449087, + 0.003800705075263977, + 0.0037281776312738657, + 0.00515176635235548, + 0.006461647339165211, + 0.11502526700496674, + 0.024173546582460403, + 0.021840259432792664, + 0.04013805463910103, + 0.016879431903362274, + 0.030177325010299683, + 0.035883575677871704, + 0.1778320074081421, + 0.002877794438973069, + 0.0032896294724196196, + 0.0028624178376048803, + 0.0031067533418536186, + 0.002864233450964093, + 0.0033666726667433977, + 0.0033880043774843216, + 0.00291090807877481, + 0.01822165958583355, + 0.007883520796895027, + 0.014942445792257786, + 0.009627707302570343, + 0.006635080091655254, + 0.01858837902545929, + 0.007125246338546276, + 0.011618515476584435, + 0.010472019203007221, + 0.013407043181359768, + 0.019445136189460754, + 0.01712494157254696, + 0.012341751717031002, + 0.017522329464554787, + 0.010345934890210629, + 0.017840780317783356, + 0.014213485643267632, + 0.010133191011846066, + 0.012002736330032349, + 0.013914184644818306, + 0.006870166398584843, + 0.012346157804131508, + 0.010934634134173393, + 0.013544321991503239, + 0.0036553761456161737, + 0.0029347967356443405, + 0.005486570298671722, + 0.003063254989683628, + 0.007445099763572216, + 0.0027260296046733856, + 0.003221364226192236, + 0.0040186564438045025, + 0.005105605814605951, + 0.00598466070368886, + 0.009574227966368198, + 0.008347814902663231, + 0.008235054090619087, + 0.007385898381471634, + 0.006584944669157267, + 0.005549000576138496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.029191240668296814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607998594641685 + ], + "opposite_logprob_tokens": [ + -3.8956079483032227 + ], + "logprob_reward": -0.020607998594641685, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607998594641685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03932976350188255, + 0.02090904489159584, + 0.051829881966114044, + 0.054729897528886795, + 0.08220785856246948, + 0.14547334611415863, + 0.10673227161169052, + 0.15764673054218292, + 0.23023968935012817, + 0.11090147495269775 + ], + "credit_map": [ + 0.0038626627065241337, + 0.004680183250457048, + 0.004448643419891596, + 0.005471866112202406, + 0.0033317275810986757, + 0.0038384608924388885, + 0.007682000286877155, + 0.006014218553900719, + 0.0019296816317364573, + 0.003937436733394861, + 0.002718596952036023, + 0.002448233775794506, + 0.002374870702624321, + 0.0015827497700229287, + 0.003112365724518895, + 0.002805108204483986, + 0.004953473806381226, + 0.011837688274681568, + 0.006452375091612339, + 0.003458860097452998, + 0.005305924452841282, + 0.0061814612708985806, + 0.006411740090698004, + 0.0072283572517335415, + 0.00401466665789485, + 0.005889615975320339, + 0.010575201362371445, + 0.007392032537609339, + 0.0045455582439899445, + 0.006024597678333521, + 0.009102702140808105, + 0.007185520604252815, + 0.0093932393938303, + 0.007513970602303743, + 0.008766215294599533, + 0.007276535499840975, + 0.019750723615288734, + 0.012075317092239857, + 0.009585394524037838, + 0.007846459746360779, + 0.02060152404010296, + 0.01269155740737915, + 0.024447083473205566, + 0.016753239557147026, + 0.011496145278215408, + 0.03457248955965042, + 0.009701520204544067, + 0.015209795907139778, + 0.007550303358584642, + 0.015207820571959019, + 0.012031728401780128, + 0.012780198827385902, + 0.011624526232481003, + 0.012244031764566898, + 0.011961839161813259, + 0.02333182282745838, + 0.014112070202827454, + 0.033452607691287994, + 0.014971098862588406, + 0.027930889278650284, + 0.014721101149916649, + 0.01631719246506691, + 0.019343895837664604, + 0.016797881573438644, + 0.019437754526734352, + 0.02131824567914009, + 0.023409618064761162, + 0.02619754523038864, + 0.07469426095485687, + 0.014155872166156769, + 0.03006884455680847, + 0.020957551896572113, + 0.009078494273126125, + 0.01474430225789547, + 0.015116546303033829, + 0.023512352257966995, + 0.012113234028220177, + 0.010357990860939026, + 0.01566201075911522, + 0.010316544212400913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.020607998594641685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802486300468445 + ], + "opposite_logprob_tokens": [ + -3.5298025608062744 + ], + "logprob_reward": -0.029802486300468445, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802486300468445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04519963264465332, + 0.025113383308053017, + 0.07552039623260498, + 0.37613245844841003, + 0.0422806441783905, + 0.10221745073795319, + 0.18314100801944733, + 0.04118794947862625, + 0.05140906572341919, + 0.05779802426695824 + ], + "credit_map": [ + 0.006879910826683044, + 0.007762260735034943, + 0.005280466750264168, + 0.005185154732316732, + 0.003779591526836157, + 0.005128417629748583, + 0.0053504109382629395, + 0.005833424162119627, + 0.002293577417731285, + 0.0028302741702646017, + 0.004534066189080477, + 0.0023503887932747602, + 0.003221367485821247, + 0.003006091108545661, + 0.003516249358654022, + 0.0033613701816648245, + 0.006049366667866707, + 0.013851313851773739, + 0.009155440144240856, + 0.004661745857447386, + 0.013701722957193851, + 0.009231929667294025, + 0.00946684181690216, + 0.009402042254805565, + 0.0982244536280632, + 0.019981371238827705, + 0.015529045835137367, + 0.031171156093478203, + 0.00761070242151618, + 0.025134379044175148, + 0.030363351106643677, + 0.14811798930168152, + 0.004388430621474981, + 0.004593519493937492, + 0.004127201158553362, + 0.00460819760337472, + 0.007538687903434038, + 0.004396637436002493, + 0.007596530951559544, + 0.0050314441323280334, + 0.017242414876818657, + 0.013760297559201717, + 0.014546764083206654, + 0.012296069413423538, + 0.009648923762142658, + 0.01687583699822426, + 0.0067827594466507435, + 0.011064385063946247, + 0.011317433789372444, + 0.022657649591565132, + 0.029506109654903412, + 0.018100712448358536, + 0.026952918618917465, + 0.03252702206373215, + 0.018691251054406166, + 0.023387931287288666, + 0.005610325373709202, + 0.00676809623837471, + 0.004949804395437241, + 0.0058863721787929535, + 0.004926979076117277, + 0.004735835827887058, + 0.00353902205824852, + 0.004771517124027014, + 0.005546047817915678, + 0.005275160539895296, + 0.006660272367298603, + 0.004853862803429365, + 0.013608968816697598, + 0.004336864221841097, + 0.005599264521151781, + 0.005528626963496208, + 0.0057416982017457485, + 0.007208799012005329, + 0.009067495353519917, + 0.009704940021038055, + 0.007153918966650963, + 0.006611605640500784, + 0.006347839720547199, + 0.005961730144917965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.029802486300468445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496806278824806 + ], + "opposite_logprob_tokens": [ + -2.9299681186676025 + ], + "logprob_reward": -0.05496806278824806, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496806278824806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07615172863006592, + 0.05136151984333992, + 0.11770085245370865, + 0.16684196889400482, + 0.045051176100969315, + 0.1523488163948059, + 0.13637159764766693, + 0.03798625245690346, + 0.10900036245584488, + 0.10718569904565811 + ], + "credit_map": [ + 0.007649767212569714, + 0.007800341118127108, + 0.008643793873488903, + 0.011957147158682346, + 0.005473573692142963, + 0.006047213915735483, + 0.018993934616446495, + 0.009585960768163204, + 0.0054393671452999115, + 0.006020577158778906, + 0.009428710676729679, + 0.00367330783046782, + 0.007035623304545879, + 0.005947763100266457, + 0.005530999973416328, + 0.00828517135232687, + 0.01249605044722557, + 0.015280048362910748, + 0.014161224476993084, + 0.007492966018617153, + 0.02428445592522621, + 0.01721690036356449, + 0.012554948218166828, + 0.01421426236629486, + 0.02630343660712242, + 0.015792230144143105, + 0.014241762459278107, + 0.01383183803409338, + 0.020817754790186882, + 0.017642783001065254, + 0.015099692158401012, + 0.04311247169971466, + 0.004427039530128241, + 0.0051796892657876015, + 0.005052794236689806, + 0.004018814768642187, + 0.00768241984769702, + 0.005756087601184845, + 0.00781312771141529, + 0.005121203605085611, + 0.012202735058963299, + 0.010780616663396358, + 0.028639260679483414, + 0.018239682540297508, + 0.010386676527559757, + 0.03932184353470802, + 0.014295264147222042, + 0.018482737243175507, + 0.009509934112429619, + 0.021327465772628784, + 0.012405372224748135, + 0.0170911718159914, + 0.01744382455945015, + 0.015524798072874546, + 0.015059989877045155, + 0.028009045869112015, + 0.0037470946554094553, + 0.0037604598328471184, + 0.005509967915713787, + 0.00580572010949254, + 0.0031930734403431416, + 0.005504957400262356, + 0.005915413610637188, + 0.0045495666563510895, + 0.011468609794974327, + 0.00880530197173357, + 0.011101294308900833, + 0.010311927646398544, + 0.03300489857792854, + 0.006949890870600939, + 0.014358436688780785, + 0.0130000039935112, + 0.011607523076236248, + 0.015415586531162262, + 0.013849755749106407, + 0.014165990985929966, + 0.014171709306538105, + 0.013907359912991524, + 0.013572379015386105, + 0.010495396330952644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.05496806278824806 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a figure on quad roller skates approaching a concrete skateboard ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the skater have burgundy hair, wear a loose black tank top, baggy denim shorts, and frayed white kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skater execute a 360-degree spin in the air while holding the outer edge of their right skate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the quad roller skates polished chrome with translucent bright orange wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot that tilts upward to capture the launch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634849451482296 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634849451482296, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634849451482296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039194997400045395, + 0.022844601422548294, + 0.028622159734368324, + 0.5480554699897766, + 0.017819805070757866, + 0.07736697047948837, + 0.163788303732872, + 0.047738216817379, + 0.008394109085202217, + 0.046175405383110046 + ], + "credit_map": [ + 0.0036162985488772392, + 0.005344636272639036, + 0.004439931828528643, + 0.004841316025704145, + 0.003752912161871791, + 0.002653445117175579, + 0.0086751040071249, + 0.005871355999261141, + 0.002255004597827792, + 0.0038268573116511106, + 0.0026963406708091497, + 0.0023868740536272526, + 0.0021698223426938057, + 0.0019363428000360727, + 0.0038405919913202524, + 0.0037327681202441454, + 0.004371085669845343, + 0.0059516169130802155, + 0.003497611964121461, + 0.002226266311481595, + 0.002534962259232998, + 0.002242028247565031, + 0.0040628560818731785, + 0.0037357318215072155, + 0.1440265029668808, + 0.025539232417941093, + 0.024533648043870926, + 0.046498313546180725, + 0.011257298290729523, + 0.03330010548233986, + 0.04464256390929222, + 0.21825778484344482, + 0.00248707109130919, + 0.002051210729405284, + 0.0023920945823192596, + 0.0018046826589852571, + 0.003027464495971799, + 0.002335268771275878, + 0.0018234794260933995, + 0.0018985331989824772, + 0.010524846613407135, + 0.005604899488389492, + 0.01377467904239893, + 0.00710862223058939, + 0.004986753221601248, + 0.017479239031672478, + 0.006054040975868702, + 0.011833888478577137, + 0.012534119188785553, + 0.02169700153172016, + 0.020017679780721664, + 0.020205970853567123, + 0.02135504223406315, + 0.02243817411363125, + 0.02331184595823288, + 0.02222847007215023, + 0.004671418108046055, + 0.009520508348941803, + 0.005485206842422485, + 0.0072901248931884766, + 0.004569974262267351, + 0.005618297960609198, + 0.006051392760127783, + 0.004531290847808123, + 0.0010241607669740915, + 0.0007036174647510052, + 0.0010116836056113243, + 0.0008175244438461959, + 0.0018568020313978195, + 0.000748948659747839, + 0.0009796821977943182, + 0.001251690904609859, + 0.005044728051871061, + 0.004802870098501444, + 0.00878043845295906, + 0.009128356352448463, + 0.0045593176037073135, + 0.004373427014797926, + 0.004055006895214319, + 0.005431263241916895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.008634849451482296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_0/index_tensor([3814, 3814])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_0/index_tensor([3814, 3814])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0a30fa37121aa99e1902f4f41585895a8c5d7e62 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_0/index_tensor([3814, 3814])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level extreme close-up captures an imperfectly shaped cube of ice resting on the polished, dark surface of a speckled granite kitchen counter. Crisp morning light from an unseen window slants across the frame, creating sharp, crystalline highlights along the cube's edges and refracting through its translucent body, illuminating the murky core of trapped air bubbles and a network of subtle internal cracks. For a moment, the cube remains perfectly still, a tiny pool of meltwater forming a shimmering halo around its base. Then, a microscopic fracture suddenly appears, a silent act of violence, decisively darting from side to side in a fraction of a second. The internal tension releases, and the cube splits cleanly in two. The two halves slide a mere millimeter across the damp surface, exposing the newly exposed, rough, crystalline interior to the light, creating a serene yet dramatic atmosphere focused on a simple, beautiful moment of physics.", + "seed": 28667, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_0/index_3814_seed_28667_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12149576433002948, + "vlm_logprob": -0.12149576433002948, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.12149576433002948, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033996693789958954 + ], + "opposite_logprob_tokens": [ + -3.65899658203125 + ], + "logprob_reward": -0.033996693789958954, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.033996693789958954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17223769426345825, + 0.047867439687252045, + 0.04828863590955734, + 0.07501769810914993, + 0.15699049830436707, + 0.11598172783851624, + 0.13372652232646942, + 0.09122157841920853, + 0.08202449977397919, + 0.07664359360933304 + ], + "credit_map": [ + 0.028128720819950104, + 0.01943296194076538, + 0.02223609760403633, + 0.0211936142295599, + 0.021122708916664124, + 0.016034163534641266, + 0.02320341020822525, + 0.020886041224002838, + 0.006791866850107908, + 0.006893705110996962, + 0.007363298442214727, + 0.006295915227383375, + 0.005285811610519886, + 0.004749059211462736, + 0.0049969651736319065, + 0.005490824580192566, + 0.007194626610726118, + 0.005963523406535387, + 0.008054868318140507, + 0.005215159617364407, + 0.006008815485984087, + 0.004915896337479353, + 0.004801476374268532, + 0.0061342730186879635, + 0.011191971600055695, + 0.011130981147289276, + 0.00915305595844984, + 0.0075310515239834785, + 0.009675085544586182, + 0.006816172040998936, + 0.011991903185844421, + 0.007527488283813, + 0.02346242219209671, + 0.020727572962641716, + 0.02856353111565113, + 0.013889213092625141, + 0.017277434468269348, + 0.014986922964453697, + 0.02216525189578533, + 0.015918172895908356, + 0.01633101888000965, + 0.019415587186813354, + 0.01446908712387085, + 0.013260449282824993, + 0.01752890832722187, + 0.009231219999492168, + 0.014665617607533932, + 0.011079853400588036, + 0.015850989148020744, + 0.01501399651169777, + 0.01759694702923298, + 0.01575915515422821, + 0.019518820568919182, + 0.015468955039978027, + 0.017933974042534828, + 0.016583703458309174, + 0.009820416569709778, + 0.016063641756772995, + 0.00899600237607956, + 0.010380689054727554, + 0.012190695852041245, + 0.007981931790709496, + 0.015807224437594414, + 0.00998099148273468, + 0.014202256686985493, + 0.01338616106659174, + 0.012359040789306164, + 0.010290662758052349, + 0.009426211006939411, + 0.006872526369988918, + 0.006917296908795834, + 0.008570350706577301, + 0.012948784045875072, + 0.014173016883432865, + 0.009947750717401505, + 0.009752881713211536, + 0.007720707915723324, + 0.00871091615408659, + 0.006949388422071934, + 0.0064401570707559586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.033996693789958954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331940270960331 + ], + "opposite_logprob_tokens": [ + -3.7733194828033447 + ], + "logprob_reward": -0.02331940270960331, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331940270960331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21407556533813477, + 0.10771545767784119, + 0.05062452703714371, + 0.030629904940724373, + 0.05868597328662872, + 0.22875645756721497, + 0.09936502575874329, + 0.06085591018199921, + 0.09009399265050888, + 0.05919710174202919 + ], + "credit_map": [ + 0.030324069783091545, + 0.021701153367757797, + 0.034597594290971756, + 0.02177242562174797, + 0.02013062685728073, + 0.020456964150071144, + 0.03635526821017265, + 0.02873748540878296, + 0.014564735814929008, + 0.012253784574568272, + 0.01671585626900196, + 0.01125671248883009, + 0.010087097063660622, + 0.00976550206542015, + 0.018584411591291428, + 0.014487365260720253, + 0.006874567363411188, + 0.005974743980914354, + 0.00937466137111187, + 0.004666609223932028, + 0.004159144591540098, + 0.00423011090606451, + 0.007576226722449064, + 0.007768469862639904, + 0.004364326596260071, + 0.002497404348105192, + 0.00671018660068512, + 0.0027152167167514563, + 0.002373647643253207, + 0.0024236247409135103, + 0.004110825248062611, + 0.005434674210846424, + 0.009222397580742836, + 0.007584399078041315, + 0.010162039659917355, + 0.005225439090281725, + 0.00487402081489563, + 0.004958242643624544, + 0.008659658022224903, + 0.007999780587852001, + 0.024737199768424034, + 0.023105071857571602, + 0.04529320448637009, + 0.029047058895230293, + 0.0268478412181139, + 0.018901774659752846, + 0.03234117105603218, + 0.028483156114816666, + 0.012245191261172295, + 0.012609357945621014, + 0.012116248719394207, + 0.012843068689107895, + 0.012568887323141098, + 0.00978037528693676, + 0.016206422820687294, + 0.010995480231940746, + 0.004876450169831514, + 0.0127119030803442, + 0.005869884509593248, + 0.014004030264914036, + 0.006364102475345135, + 0.006510543636977673, + 0.00627258513122797, + 0.004246422555297613, + 0.014580784365534782, + 0.0132320336997509, + 0.009049637243151665, + 0.008898606523871422, + 0.01697099395096302, + 0.00928440224379301, + 0.006877544801682234, + 0.01119999773800373, + 0.009296522475779057, + 0.00909580197185278, + 0.0065839337185025215, + 0.007862626574933529, + 0.0051356214098632336, + 0.008158725686371326, + 0.006552285049110651, + 0.006511590909212828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.02331940270960331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011101268231868744 + ], + "opposite_logprob_tokens": [ + -4.511101245880127 + ], + "logprob_reward": -0.011101268231868744, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011101268231868744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2068331390619278, + 0.05023253709077835, + 0.04198869317770004, + 0.0568230003118515, + 0.048092570155858994, + 0.10734203457832336, + 0.10795846581459045, + 0.1146726980805397, + 0.13231506943702698, + 0.13374175131320953 + ], + "credit_map": [ + 0.04133598878979683, + 0.027158288285136223, + 0.031773123890161514, + 0.017379701137542725, + 0.019949153065681458, + 0.019198985770344734, + 0.02067457139492035, + 0.029363339766860008, + 0.009103590622544289, + 0.006253122352063656, + 0.008176249451935291, + 0.00452762795612216, + 0.004319494124501944, + 0.0052311173640191555, + 0.005034704692661762, + 0.007586634252220392, + 0.009035852737724781, + 0.005299089942127466, + 0.004962381906807423, + 0.003914779517799616, + 0.004145406186580658, + 0.0035854510497301817, + 0.005446148570626974, + 0.005599586758762598, + 0.01165523286908865, + 0.005731853190809488, + 0.008622273802757263, + 0.0052913157269358635, + 0.005656130146235228, + 0.004988572094589472, + 0.007794653996825218, + 0.007082974538207054, + 0.007819878868758678, + 0.004240161739289761, + 0.007642767392098904, + 0.004784867633134127, + 0.004528766497969627, + 0.006137988064438105, + 0.007272196002304554, + 0.005665946751832962, + 0.017186805605888367, + 0.011396621353924274, + 0.01904461905360222, + 0.010578651912510395, + 0.013001457788050175, + 0.011365456506609917, + 0.010576759465038776, + 0.014191671274602413, + 0.010930326767265797, + 0.01666235364973545, + 0.00975243840366602, + 0.013813111931085587, + 0.018621817231178284, + 0.011376449838280678, + 0.014560776762664318, + 0.012241197749972343, + 0.010687164030969143, + 0.027371404692530632, + 0.00953640416264534, + 0.01847853697836399, + 0.011117635294795036, + 0.012824293226003647, + 0.014019042253494263, + 0.010638216510415077, + 0.017411639913916588, + 0.01759694702923298, + 0.018662232905626297, + 0.019985299557447433, + 0.015764692798256874, + 0.01642347127199173, + 0.012227645143866539, + 0.014243162237107754, + 0.019714126363396645, + 0.020619161427021027, + 0.028807301074266434, + 0.012925835326313972, + 0.012756803072988987, + 0.010485106147825718, + 0.01771412044763565, + 0.010719308629631996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.011101268231868744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255856990814209 + ], + "opposite_logprob_tokens": [ + -1.600585699081421 + ], + "logprob_reward": -0.2255856990814209, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255856990814209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20591729879379272, + 0.03645375743508339, + 0.029407063499093056, + 0.043114930391311646, + 0.08014018088579178, + 0.045373667031526566, + 0.177590012550354, + 0.11295285820960999, + 0.17798112332820892, + 0.0910690501332283 + ], + "credit_map": [ + 0.032229162752628326, + 0.024747367948293686, + 0.02853342704474926, + 0.0215078704059124, + 0.022009054198861122, + 0.01935187354683876, + 0.03283242881298065, + 0.02470611408352852, + 0.005297838244587183, + 0.005358721129596233, + 0.005214963108301163, + 0.0039986069314181805, + 0.003910279832780361, + 0.004206306301057339, + 0.00407114764675498, + 0.004395895637571812, + 0.00415897648781538, + 0.0038913777098059654, + 0.004903281573206186, + 0.0032135024666786194, + 0.003136260202154517, + 0.0034334545489400625, + 0.003346857614815235, + 0.0033233538269996643, + 0.0045265499502420425, + 0.00496015977114439, + 0.006907844450324774, + 0.005121068563312292, + 0.005195040721446276, + 0.004437524359673262, + 0.007943088188767433, + 0.004023651126772165, + 0.012088960967957973, + 0.009097247384488583, + 0.01321234181523323, + 0.007952597923576832, + 0.008803442120552063, + 0.008567238226532936, + 0.011247966438531876, + 0.00917038694024086, + 0.007011094130575657, + 0.006590902805328369, + 0.004204912111163139, + 0.0049350680783391, + 0.008271312341094017, + 0.0038529883604496717, + 0.005708653014153242, + 0.004798737354576588, + 0.01959048956632614, + 0.0222882442176342, + 0.024485638365149498, + 0.019205108284950256, + 0.030467715114355087, + 0.01852218247950077, + 0.02369639277458191, + 0.01933424547314644, + 0.010564560070633888, + 0.021760528907179832, + 0.011527035385370255, + 0.01590045355260372, + 0.01059099193662405, + 0.015136673115193844, + 0.01673174649477005, + 0.010740871541202068, + 0.023499248549342155, + 0.02308676578104496, + 0.019527088850736618, + 0.02438480220735073, + 0.02462327480316162, + 0.015472104772925377, + 0.028364239260554314, + 0.019023584201931953, + 0.011746667325496674, + 0.014185765758156776, + 0.006947352085262537, + 0.008382013998925686, + 0.009128941223025322, + 0.016841839998960495, + 0.015362768433988094, + 0.008473698049783707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.2255856990814209 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31347575783729553 + ], + "opposite_logprob_tokens": [ + -1.3134757280349731 + ], + "logprob_reward": -0.31347575783729553, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31347575783729553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1451188623905182, + 0.059642769396305084, + 0.044933248311281204, + 0.09070676565170288, + 0.03693427890539169, + 0.23113663494586945, + 0.10820925980806351, + 0.06282483041286469, + 0.1540331244468689, + 0.06646019220352173 + ], + "credit_map": [ + 0.026505105197429657, + 0.02153765968978405, + 0.018085302785038948, + 0.016676926985383034, + 0.01378907822072506, + 0.012795223854482174, + 0.01624494418501854, + 0.019484631717205048, + 0.010930404998362064, + 0.007419896312057972, + 0.007798740174621344, + 0.00588735518977046, + 0.0064888689666986465, + 0.00568582396954298, + 0.007742961868643761, + 0.007688723038882017, + 0.007883677259087563, + 0.0062818871811032295, + 0.005139742977917194, + 0.004655993077903986, + 0.005806227680295706, + 0.004126694984734058, + 0.005278720986098051, + 0.005760301370173693, + 0.012288729660212994, + 0.011194892227649689, + 0.011259319260716438, + 0.010748346336185932, + 0.012096960097551346, + 0.009530645795166492, + 0.012567412108182907, + 0.011020461097359657, + 0.005337254144251347, + 0.003294274676591158, + 0.008756618946790695, + 0.0034310640767216682, + 0.00406612129881978, + 0.0033144564367830753, + 0.004533693194389343, + 0.004200794268399477, + 0.03778454661369324, + 0.03795406594872475, + 0.023834317922592163, + 0.02519238367676735, + 0.03480920568108559, + 0.017883703112602234, + 0.03480453044176102, + 0.01887388341128826, + 0.012538452632725239, + 0.017897561192512512, + 0.012033192440867424, + 0.011715106666088104, + 0.01659630611538887, + 0.012499729171395302, + 0.014346933923661709, + 0.010581977665424347, + 0.005060215014964342, + 0.015148640610277653, + 0.005554812029004097, + 0.008160962723195553, + 0.007982533425092697, + 0.006648985203355551, + 0.008451046422123909, + 0.005817635916173458, + 0.023917390033602715, + 0.024345068261027336, + 0.013589209876954556, + 0.021682502701878548, + 0.018595073372125626, + 0.012137857265770435, + 0.02403399907052517, + 0.015732020139694214, + 0.010206959210336208, + 0.008088575676083565, + 0.01118408516049385, + 0.006431164685636759, + 0.006523828953504562, + 0.007957136258482933, + 0.008051269687712193, + 0.00801717396825552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.31347575783729553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level extreme close-up captures an imperfectly shaped cube of ice resting on the polished, dark surface of a speckled granite kitchen counter. Crisp morning light from an unseen window slants across the frame, creating sharp, crystalline highlights along the cube's edges and refracting through its translucent body, illuminating the murky core of trapped air bubbles and a network of subtle internal cracks. For a moment, the cube remains perfectly still, a tiny pool of meltwater forming a shimmering halo around its base. Then, a microscopic fracture suddenly appears, a silent act of violence, decisively darting from side to side in a fraction of a second. The internal tension releases, and the cube splits cleanly in two. The two halves slide a mere millimeter across the damp surface, exposing the newly exposed, rough, crystalline interior to the light, creating a serene yet dramatic atmosphere focused on a simple, beautiful moment of physics.", + "seed": 28668, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_0/index_3814_seed_28668_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2800192467868328, + "vlm_logprob": -0.2800192467868328, + "vlm_yes_no_margin": 1.775, + "vlm_token_credit": -0.2800192467868328, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4773585796356201 + ], + "opposite_logprob_tokens": [ + -0.9773585796356201 + ], + "logprob_reward": -0.4773585796356201, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4773585796356201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37472212314605713, + 0.07425429672002792, + 0.041161999106407166, + 0.041523441672325134, + 0.06809982657432556, + 0.13186702132225037, + 0.0620323047041893, + 0.054459139704704285, + 0.05990320071578026, + 0.09197665750980377 + ], + "credit_map": [ + 0.057366978377103806, + 0.04932882636785507, + 0.07217717915773392, + 0.03224420174956322, + 0.04199131578207016, + 0.03558490052819252, + 0.03978467360138893, + 0.04624403268098831, + 0.014009214006364346, + 0.008406001143157482, + 0.017807217314839363, + 0.005110427737236023, + 0.006317636929452419, + 0.010454985313117504, + 0.005020395386964083, + 0.007128414697945118, + 0.005615721456706524, + 0.006299485452473164, + 0.007097553461790085, + 0.003995856270194054, + 0.0038796765729784966, + 0.005384895950555801, + 0.004400896839797497, + 0.004487913567572832, + 0.006370645482093096, + 0.004584421869367361, + 0.008543749339878559, + 0.0035075079649686813, + 0.004302531480789185, + 0.007010300178080797, + 0.0035978956148028374, + 0.0036063888110220432, + 0.010952964425086975, + 0.006736482959240675, + 0.01804598607122898, + 0.004529299680143595, + 0.0066362787038087845, + 0.005900254473090172, + 0.0062768710777163506, + 0.009021684527397156, + 0.024949759244918823, + 0.014338534325361252, + 0.036233242601156235, + 0.008089407347142696, + 0.00893587525933981, + 0.011172894388437271, + 0.01031631138175726, + 0.017830999568104744, + 0.0074110557325184345, + 0.012555566616356373, + 0.008603973314166069, + 0.005465253721922636, + 0.008665278553962708, + 0.00589782977476716, + 0.006212185602635145, + 0.007221166510134935, + 0.014027077704668045, + 0.005431884899735451, + 0.0061590890400111675, + 0.003779541002586484, + 0.005729869939386845, + 0.0062930104322731495, + 0.005562038626521826, + 0.007476635277271271, + 0.008292589336633682, + 0.009235771372914314, + 0.004885551054030657, + 0.007360939402133226, + 0.007807131856679916, + 0.007291371002793312, + 0.009077988564968109, + 0.005951856728643179, + 0.007057830225676298, + 0.011772851459681988, + 0.01305135153234005, + 0.011089373379945755, + 0.010141595266759396, + 0.018870633095502853, + 0.00997739378362894, + 0.010015632025897503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.4773585796356201 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043094001710414886 + ], + "opposite_logprob_tokens": [ + -3.1680939197540283 + ], + "logprob_reward": -0.043094001710414886, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043094001710414886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19903799891471863, + 0.10896195471286774, + 0.12935373187065125, + 0.03524480387568474, + 0.15138652920722961, + 0.07373934984207153, + 0.06966157257556915, + 0.048764053732156754, + 0.09752321988344193, + 0.08632688969373703 + ], + "credit_map": [ + 0.03093246929347515, + 0.03396192193031311, + 0.03942188620567322, + 0.016141241416335106, + 0.017177777364850044, + 0.019875548779964447, + 0.01723339967429638, + 0.02429371140897274, + 0.01615254580974579, + 0.013104170560836792, + 0.0194703359156847, + 0.007716418709605932, + 0.007971569895744324, + 0.019083572551608086, + 0.00732897873967886, + 0.01813434436917305, + 0.02368987165391445, + 0.014599147252738476, + 0.024224689230322838, + 0.008295050822198391, + 0.006961222738027573, + 0.023530831560492516, + 0.008400323800742626, + 0.01965256966650486, + 0.006992506794631481, + 0.0029800888150930405, + 0.007583056576550007, + 0.0026211768854409456, + 0.002642048057168722, + 0.005239011254161596, + 0.0026836309116333723, + 0.004503281321376562, + 0.03141402453184128, + 0.010378125123679638, + 0.037533946335315704, + 0.01142327394336462, + 0.008737687952816486, + 0.022434087470173836, + 0.01139422319829464, + 0.018071137368679047, + 0.009874393232166767, + 0.011333728209137917, + 0.012254797853529453, + 0.007154499180614948, + 0.006256554741412401, + 0.00960029661655426, + 0.007549661677330732, + 0.00971540529280901, + 0.010972763411700726, + 0.008479394018650055, + 0.014446783810853958, + 0.00485307676717639, + 0.007412212900817394, + 0.009262342005968094, + 0.005094394087791443, + 0.009140594862401485, + 0.0067860460840165615, + 0.005464527755975723, + 0.006073757540434599, + 0.005662098992615938, + 0.006162700243294239, + 0.006211091764271259, + 0.0053519573993980885, + 0.007051866035908461, + 0.013893963769078255, + 0.020109873265028, + 0.007158295251429081, + 0.010188371874392033, + 0.010685892775654793, + 0.012388358823955059, + 0.014478911645710468, + 0.008619533851742744, + 0.006298876367509365, + 0.011911403387784958, + 0.008356052450835705, + 0.008374849334359169, + 0.01030562724918127, + 0.01647929660975933, + 0.010669411160051823, + 0.01393135916441679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.043094001710414886 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209312111139297 + ], + "opposite_logprob_tokens": [ + -4.018209457397461 + ], + "logprob_reward": -0.018209312111139297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209312111139297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2871828079223633, + 0.10635482519865036, + 0.06701525300741196, + 0.03418101742863655, + 0.04812036454677582, + 0.10911494493484497, + 0.13774049282073975, + 0.06007910519838333, + 0.08073971420526505, + 0.06947146356105804 + ], + "credit_map": [ + 0.040811687707901, + 0.03497583419084549, + 0.0487496554851532, + 0.021250881254673004, + 0.02701842226088047, + 0.05295790731906891, + 0.020808471366763115, + 0.04060995578765869, + 0.01665368676185608, + 0.010782611556351185, + 0.018351810052990913, + 0.006054782774299383, + 0.008702985942363739, + 0.02543334662914276, + 0.005839372985064983, + 0.014536227099597454, + 0.009814529679715633, + 0.005820609629154205, + 0.01143427100032568, + 0.004418366122990847, + 0.004783493932336569, + 0.016754334792494774, + 0.0049275062046945095, + 0.009062139317393303, + 0.006472359411418438, + 0.0026563801802694798, + 0.009966378100216389, + 0.0024353202898055315, + 0.002913258969783783, + 0.0026569629553705454, + 0.0025841586757451296, + 0.004496199078857899, + 0.008091216906905174, + 0.00361130153760314, + 0.009874442592263222, + 0.002981767291203141, + 0.0029088654555380344, + 0.01014404371380806, + 0.003607556689530611, + 0.00690117385238409, + 0.016001593321561813, + 0.015351170673966408, + 0.02173304185271263, + 0.009630080312490463, + 0.007519014645367861, + 0.013494440354406834, + 0.010048631578683853, + 0.015336978249251842, + 0.018718337640166283, + 0.025437403470277786, + 0.0203167162835598, + 0.011160760186612606, + 0.021942563354969025, + 0.01451167743653059, + 0.011473581194877625, + 0.014179453253746033, + 0.012620851397514343, + 0.0072455876506865025, + 0.008766885846853256, + 0.005135940387845039, + 0.005488234106451273, + 0.00935350637882948, + 0.004962898790836334, + 0.006505201570689678, + 0.012467763386666775, + 0.014624279923737049, + 0.005604051984846592, + 0.009586594067513943, + 0.007275525946170092, + 0.011127477511763573, + 0.013152580708265305, + 0.006901436951011419, + 0.006310416851192713, + 0.009757520630955696, + 0.008023259229958057, + 0.007715174928307533, + 0.009852016344666481, + 0.012306987307965755, + 0.00787875521928072, + 0.007627339102327824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.018209312111139297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47433042526245117 + ], + "opposite_logprob_tokens": [ + -0.9743304252624512 + ], + "logprob_reward": -0.47433042526245117, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47433042526245117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23228144645690918, + 0.10730556398630142, + 0.03969896212220192, + 0.029411939904093742, + 0.04206263646483421, + 0.1208067461848259, + 0.1892641931772232, + 0.057531971484422684, + 0.11712571233510971, + 0.06451082229614258 + ], + "credit_map": [ + 0.03154021501541138, + 0.031864456832408905, + 0.04456024244427681, + 0.02121714875102043, + 0.019770530983805656, + 0.030866900458931923, + 0.01974617689847946, + 0.032715801149606705, + 0.015959545969963074, + 0.010154884308576584, + 0.017743047326803207, + 0.008164369501173496, + 0.006751293316483498, + 0.021661721169948578, + 0.009236597456037998, + 0.017634104937314987, + 0.006133486516773701, + 0.004085270222276449, + 0.009059893898665905, + 0.0028883207123726606, + 0.002106177154928446, + 0.006085167173296213, + 0.0025869717355817556, + 0.006753676570951939, + 0.005443889647722244, + 0.0028468184173107147, + 0.00842560175806284, + 0.0019185565179213881, + 0.0024672148283571005, + 0.0018586365040391684, + 0.0023189333733171225, + 0.00413229176774621, + 0.008315619081258774, + 0.003793821670114994, + 0.0081273028627038, + 0.0038185222074389458, + 0.003063713200390339, + 0.006788887083530426, + 0.003795810742303729, + 0.00435896310955286, + 0.020363356918096542, + 0.011701985262334347, + 0.02929139882326126, + 0.010152310132980347, + 0.009977344423532486, + 0.01421874575316906, + 0.009525561705231667, + 0.015576046891510487, + 0.03992704302072525, + 0.022494448348879814, + 0.05348002165555954, + 0.013659178279340267, + 0.01157994195818901, + 0.010875631123781204, + 0.016455769538879395, + 0.020792169496417046, + 0.009372537024319172, + 0.0070548513904213905, + 0.007633212022483349, + 0.006616176106035709, + 0.0062971836887300014, + 0.007713312283158302, + 0.006165696308016777, + 0.006679007317870855, + 0.011474641039967537, + 0.015861058607697487, + 0.007673530839383602, + 0.016901953145861626, + 0.014302179217338562, + 0.020466655492782593, + 0.018197569996118546, + 0.01224813237786293, + 0.006483233068138361, + 0.007124355062842369, + 0.006071510259062052, + 0.006487032864242792, + 0.00902144517749548, + 0.013324336148798466, + 0.007672937586903572, + 0.008325979113578796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.47433042526245117 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871039152145386 + ], + "opposite_logprob_tokens": [ + -1.1371039152145386 + ], + "logprob_reward": -0.3871039152145386, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871039152145386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15603980422019958, + 0.03850499913096428, + 0.05304553359746933, + 0.04221266135573387, + 0.05053824558854103, + 0.1071510910987854, + 0.1577262580394745, + 0.14445416629314423, + 0.18167530000209808, + 0.06865199655294418 + ], + "credit_map": [ + 0.02387896738946438, + 0.02014664001762867, + 0.025867335498332977, + 0.015585113316774368, + 0.016532210633158684, + 0.017291750758886337, + 0.015145727433264256, + 0.021592069417238235, + 0.006936413235962391, + 0.004874844569712877, + 0.008681812323629856, + 0.0030195116996765137, + 0.0036270450800657272, + 0.00405991030856967, + 0.002897818572819233, + 0.004407639615237713, + 0.007312188856303692, + 0.004544325638562441, + 0.00816821213811636, + 0.003182977670803666, + 0.004024123772978783, + 0.012128946371376514, + 0.004617923870682716, + 0.00906683225184679, + 0.005498922895640135, + 0.004714902490377426, + 0.008604256436228752, + 0.0026554421056061983, + 0.0033270176500082016, + 0.00858727190643549, + 0.0031646862626075745, + 0.005660158582031727, + 0.008314876817166805, + 0.0043289149180054665, + 0.014042052440345287, + 0.0024735438637435436, + 0.0035799462348222733, + 0.00755664985626936, + 0.0031046043150126934, + 0.007137653883546591, + 0.02008051797747612, + 0.009581761434674263, + 0.031704433262348175, + 0.006639161147177219, + 0.009162546135485172, + 0.008526868186891079, + 0.007988273166120052, + 0.013467523269355297, + 0.03149271383881569, + 0.01388260256499052, + 0.04096350818872452, + 0.011110431514680386, + 0.015431219711899757, + 0.016710788011550903, + 0.012898091226816177, + 0.015236906707286835, + 0.04030326381325722, + 0.013819490559399128, + 0.011217283084988594, + 0.010455289855599403, + 0.012075124308466911, + 0.013940108940005302, + 0.012530510313808918, + 0.030113087967038155, + 0.01831546239554882, + 0.04048382118344307, + 0.013067035004496574, + 0.020675038918852806, + 0.022377682849764824, + 0.02672519162297249, + 0.021768352016806602, + 0.0182627085596323, + 0.005680468399077654, + 0.009539560414850712, + 0.006790583953261375, + 0.007433857303112745, + 0.007756439503282309, + 0.012200610712170601, + 0.008221804164350033, + 0.01102867815643549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.3871039152145386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level extreme close-up captures an imperfectly shaped cube of ice resting on the polished, dark surface of a speckled granite kitchen counter. Crisp morning light from an unseen window slants across the frame, creating sharp, crystalline highlights along the cube's edges and refracting through its translucent body, illuminating the murky core of trapped air bubbles and a network of subtle internal cracks. For a moment, the cube remains perfectly still, a tiny pool of meltwater forming a shimmering halo around its base. Then, a microscopic fracture suddenly appears, a silent act of violence, decisively darting from side to side in a fraction of a second. The internal tension releases, and the cube splits cleanly in two. The two halves slide a mere millimeter across the damp surface, exposing the newly exposed, rough, crystalline interior to the light, creating a serene yet dramatic atmosphere focused on a simple, beautiful moment of physics.", + "seed": 28669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_0/index_3814_seed_28669_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.26528267152607443, + "vlm_logprob": -0.26528267152607443, + "vlm_yes_no_margin": 1.85, + "vlm_token_credit": -0.26528267152607443, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18641351163387299 + ], + "opposite_logprob_tokens": [ + -1.8114135265350342 + ], + "logprob_reward": -0.18641351163387299, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18641351163387299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0709727481007576, + 0.058849819004535675, + 0.03300204873085022, + 0.044868648052215576, + 0.053506508469581604, + 0.1882917582988739, + 0.2085205614566803, + 0.2286280393600464, + 0.07136073708534241, + 0.041999105364084244 + ], + "credit_map": [ + 0.007394568994641304, + 0.007933604530990124, + 0.011081703938543797, + 0.007769586052745581, + 0.005666409619152546, + 0.011387431062757969, + 0.005652717314660549, + 0.014086729846894741, + 0.005861188750714064, + 0.007102598901838064, + 0.008979903534054756, + 0.0072971489280462265, + 0.007868261076509953, + 0.007521206978708506, + 0.006405483931303024, + 0.007814033888280392, + 0.0031412853859364986, + 0.003413404803723097, + 0.006743165198713541, + 0.003151682671159506, + 0.002186123514547944, + 0.00838149432092905, + 0.0017274102428928018, + 0.004257487133145332, + 0.005225084256380796, + 0.004823413677513599, + 0.007135195657610893, + 0.004075889475643635, + 0.004588739480823278, + 0.009647059254348278, + 0.0034681730903685093, + 0.0059050945565104485, + 0.0054503451101481915, + 0.007607961073517799, + 0.00743096275255084, + 0.006808655336499214, + 0.007254564668983221, + 0.0068047912791371346, + 0.005433781072497368, + 0.0067154522985219955, + 0.013974644243717194, + 0.010286192409694195, + 0.06522735208272934, + 0.0076232957653701305, + 0.006487162783741951, + 0.06531265377998352, + 0.005826238542795181, + 0.01355423592031002, + 0.016185445711016655, + 0.03379763662815094, + 0.028271645307540894, + 0.02029259316623211, + 0.04126613214612007, + 0.023373505100607872, + 0.023350998759269714, + 0.02198260836303234, + 0.01143953949213028, + 0.0899834856390953, + 0.06308727711439133, + 0.012125231325626373, + 0.0190227422863245, + 0.013176441192626953, + 0.01158600952476263, + 0.008207324892282486, + 0.007449421100318432, + 0.009997334331274033, + 0.0104474276304245, + 0.008733964525163174, + 0.008407313376665115, + 0.010163434781134129, + 0.008648263290524483, + 0.0075135803781449795, + 0.004046950489282608, + 0.00513869384303689, + 0.0073972768150269985, + 0.00621867086738348, + 0.004195844754576683, + 0.006504210643470287, + 0.0038207124453037977, + 0.004676745273172855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.18641351163387299 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209287419915199 + ], + "opposite_logprob_tokens": [ + -2.8120927810668945 + ], + "logprob_reward": -0.06209287419915199, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209287419915199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03519381582736969, + 0.051102157682180405, + 0.028416601940989494, + 0.01735137589275837, + 0.0118436673656106, + 0.04417521506547928, + 0.10427894443273544, + 0.6645321846008301, + 0.021964827552437782, + 0.021141162142157555 + ], + "credit_map": [ + 0.004081296268850565, + 0.004674273077398539, + 0.005791677627712488, + 0.0033646102529019117, + 0.0040274313651025295, + 0.006044705864042044, + 0.0028784095775336027, + 0.00433141365647316, + 0.004814777057617903, + 0.00430762255564332, + 0.010238820686936378, + 0.005313478875905275, + 0.003745004767552018, + 0.012753642164170742, + 0.003312811255455017, + 0.006615996826440096, + 0.0031582119408994913, + 0.002291944343596697, + 0.004323194734752178, + 0.0032023724634200335, + 0.0027282547671347857, + 0.0060836272314190865, + 0.002461361465975642, + 0.00416763499379158, + 0.002255318220704794, + 0.001978576648980379, + 0.00324814161285758, + 0.0015213439473882318, + 0.001528889057226479, + 0.003664060030132532, + 0.001088530756533146, + 0.002066515851765871, + 0.0014121999265626073, + 0.0012805837905034423, + 0.0015265403781086206, + 0.0018156194128096104, + 0.0015729459701105952, + 0.0014345397939905524, + 0.001258539967238903, + 0.0015426983591169119, + 0.0038241297006607056, + 0.0018565020291134715, + 0.014152624644339085, + 0.0021236047614365816, + 0.0016489868285134435, + 0.015576091594994068, + 0.001651783473789692, + 0.003341489937156439, + 0.014112942852079868, + 0.01382363960146904, + 0.01654810830950737, + 0.010380532592535019, + 0.01640576682984829, + 0.010898941196501255, + 0.010125345550477505, + 0.011983660981059074, + 0.03453843668103218, + 0.27697455883026123, + 0.20449140667915344, + 0.026206916198134422, + 0.0474536269903183, + 0.022415297105908394, + 0.029846668243408203, + 0.022605257108807564, + 0.0021803616546094418, + 0.0038098024670034647, + 0.0028181809466332197, + 0.002645844826474786, + 0.0026735246647149324, + 0.002614147262647748, + 0.0031065999064594507, + 0.002116366056725383, + 0.0019346436019986868, + 0.005946365650743246, + 0.003759237704798579, + 0.0019196077482774854, + 0.00219904026016593, + 0.0020903011318296194, + 0.0017078210366889834, + 0.0015841452404856682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.06209287419915199 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610473357141018 + ], + "opposite_logprob_tokens": [ + -4.141104698181152 + ], + "logprob_reward": -0.01610473357141018, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610473357141018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06623709946870804, + 0.016877302899956703, + 0.01146961934864521, + 0.017923936247825623, + 0.022613154724240303, + 0.43165352940559387, + 0.06922212988138199, + 0.2681295871734619, + 0.04046069458127022, + 0.05541297793388367 + ], + "credit_map": [ + 0.0070703765377402306, + 0.006212026812136173, + 0.012896709144115448, + 0.004448284860700369, + 0.004587838891893625, + 0.01934424415230751, + 0.003388453973457217, + 0.00828916672617197, + 0.001733540091663599, + 0.0020706993527710438, + 0.002654433948919177, + 0.0016801876481622458, + 0.0021631766576319933, + 0.00215913075953722, + 0.0018955250270664692, + 0.0025206096470355988, + 0.0011830230941995978, + 0.0013474231818690896, + 0.00219345698133111, + 0.001430802047252655, + 0.00129793887026608, + 0.0013712011277675629, + 0.0009822448482736945, + 0.0016635287320241332, + 0.00183901097625494, + 0.0016540943179279566, + 0.00401652418076992, + 0.0017124456353485584, + 0.0015499894507229328, + 0.00391590129584074, + 0.0013065782841295004, + 0.0019293901277706027, + 0.002087353030219674, + 0.002865305170416832, + 0.004535941872745752, + 0.0020253604743629694, + 0.0027322859968990088, + 0.0038203049916774035, + 0.002081900369375944, + 0.002464703284204006, + 0.02981138415634632, + 0.01423730794340372, + 0.15989726781845093, + 0.012870485894382, + 0.012506145983934402, + 0.1621166616678238, + 0.011262839660048485, + 0.028951425105333328, + 0.009473244659602642, + 0.012729626148939133, + 0.010471230372786522, + 0.007181853987276554, + 0.006265919655561447, + 0.00835308339446783, + 0.006415337324142456, + 0.00833183340728283, + 0.015050925314426422, + 0.11407697945833206, + 0.08093477785587311, + 0.010283113457262516, + 0.017957443371415138, + 0.009295519441366196, + 0.011087772436439991, + 0.009443055838346481, + 0.003088988596573472, + 0.006401797756552696, + 0.0057802340015769005, + 0.00522576691582799, + 0.004681686870753765, + 0.005636353045701981, + 0.006131463218480349, + 0.0035144074354320765, + 0.005323023535311222, + 0.021091561764478683, + 0.010630672797560692, + 0.0034706739243119955, + 0.0064006648026406765, + 0.0032884476240724325, + 0.0026263457257300615, + 0.002581588225439191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.01610473357141018 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.63286292552948 + ], + "opposite_logprob_tokens": [ + -0.75786292552948 + ], + "logprob_reward": -0.63286292552948, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.63286292552948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041395898908376694, + 0.027401506900787354, + 0.02308541163802147, + 0.01915971376001835, + 0.017421716824173927, + 0.7470293045043945, + 0.04075627028942108, + 0.03651387244462967, + 0.01898595131933689, + 0.028250372037291527 + ], + "credit_map": [ + 0.004109052009880543, + 0.00573858292773366, + 0.0055816397070884705, + 0.00542633980512619, + 0.005772402975708246, + 0.004777683410793543, + 0.004821053706109524, + 0.005169144831597805, + 0.002375323325395584, + 0.0036853880155831575, + 0.004037422128021717, + 0.003227824345231056, + 0.003980881534516811, + 0.0034456958528608084, + 0.003194343764334917, + 0.0034546307288110256, + 0.002084272215142846, + 0.0032386058010160923, + 0.0035420835483819246, + 0.002492650644853711, + 0.0031230791937559843, + 0.0032723320182412863, + 0.002894959645345807, + 0.0024374278727918863, + 0.0017504355637356639, + 0.002384408377110958, + 0.003253799630329013, + 0.001878396375104785, + 0.0022794948890805244, + 0.0036669711116701365, + 0.0017980595584958792, + 0.002148147439584136, + 0.0014896757202222943, + 0.0020414688624441624, + 0.0036028020549565554, + 0.001631156075745821, + 0.001991022378206253, + 0.0034409251529723406, + 0.001696548773907125, + 0.001528118271380663, + 0.05184789001941681, + 0.023243805393576622, + 0.27669188380241394, + 0.02102479338645935, + 0.0216019619256258, + 0.28199782967567444, + 0.019504478201270103, + 0.05111665651202202, + 0.0044619375839829445, + 0.006292399019002914, + 0.005798963364213705, + 0.003888710169121623, + 0.0075044455006718636, + 0.004434075206518173, + 0.004271307960152626, + 0.004104433581233025, + 0.0030399190727621317, + 0.008685842156410217, + 0.006309965159744024, + 0.003883522469550371, + 0.003491607727482915, + 0.005095027387142181, + 0.0026748268865048885, + 0.0033331613522022963, + 0.0015535461716353893, + 0.0023360454943031073, + 0.003429015865549445, + 0.0018915040418505669, + 0.0025258490350097418, + 0.002758355112746358, + 0.0027883907314389944, + 0.0017032441683113575, + 0.003332073101773858, + 0.011228666640818119, + 0.0038295963313430548, + 0.0016536422772333026, + 0.003534061834216118, + 0.002069211332127452, + 0.0012579677859321237, + 0.001345152617432177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.63286292552948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289393126964569 + ], + "opposite_logprob_tokens": [ + -1.0539393424987793 + ], + "logprob_reward": -0.4289393126964569, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289393126964569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05142142251133919, + 0.01413531880825758, + 0.018277892842888832, + 0.025984128937125206, + 0.026288049295544624, + 0.4815479516983032, + 0.08913765102624893, + 0.20566242933273315, + 0.028254199773073196, + 0.05929095298051834 + ], + "credit_map": [ + 0.004889723379164934, + 0.003771123243495822, + 0.013513704761862755, + 0.002496837405487895, + 0.0027331714518368244, + 0.01595420576632023, + 0.002358909696340561, + 0.005703749135136604, + 0.001900182105600834, + 0.0017570399213582277, + 0.0030794080812484026, + 0.0013644136488437653, + 0.0011629631044343114, + 0.0022234604693949223, + 0.0007394711137749255, + 0.0019083806546404958, + 0.0016782135935500264, + 0.0022069320548325777, + 0.0030082555022090673, + 0.001452616648748517, + 0.00180139415897429, + 0.004255635663866997, + 0.001382110407575965, + 0.00249273469671607, + 0.0018269449938088655, + 0.00316270487383008, + 0.004718539770692587, + 0.0025857118889689445, + 0.0028952383436262608, + 0.005748968105763197, + 0.0023200823925435543, + 0.0027259388007223606, + 0.002280203625559807, + 0.0022007387597113848, + 0.006952264346182346, + 0.0016315745888277888, + 0.0013252891367301345, + 0.007543607614934444, + 0.0011941225966438651, + 0.0031602485105395317, + 0.03334398195147514, + 0.015132474713027477, + 0.1784292757511139, + 0.01316738035529852, + 0.014054813422262669, + 0.1813727617263794, + 0.012714150361716747, + 0.03333308547735214, + 0.010676724836230278, + 0.011589152738451958, + 0.011618864722549915, + 0.009158024564385414, + 0.013749576173722744, + 0.008862304501235485, + 0.009241183288395405, + 0.014241812750697136, + 0.010632980614900589, + 0.08540324121713638, + 0.0646226778626442, + 0.008483207784593105, + 0.014586133882403374, + 0.006642173510044813, + 0.009016755037009716, + 0.006275249645113945, + 0.0029993713833391666, + 0.0037247806321829557, + 0.003844365244731307, + 0.0035912038292735815, + 0.003693279577419162, + 0.0031390474177896976, + 0.003930593840777874, + 0.0033315573818981647, + 0.006465098354965448, + 0.022721927613019943, + 0.007934956811368465, + 0.0033818441443145275, + 0.007269275840371847, + 0.004501708317548037, + 0.002945102984085679, + 0.004071041475981474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.4289393126964569 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level extreme close-up captures an imperfectly shaped cube of ice resting on the polished, dark surface of a speckled granite kitchen counter. Crisp morning light from an unseen window slants across the frame, creating sharp, crystalline highlights along the cube's edges and refracting through its translucent body, illuminating the murky core of trapped air bubbles and a network of subtle internal cracks. For a moment, the cube remains perfectly still, a tiny pool of meltwater forming a shimmering halo around its base. Then, a microscopic fracture suddenly appears, a silent act of violence, decisively darting from side to side in a fraction of a second. The internal tension releases, and the cube splits cleanly in two. The two halves slide a mere millimeter across the damp surface, exposing the newly exposed, rough, crystalline interior to the light, creating a serene yet dramatic atmosphere focused on a simple, beautiful moment of physics.", + "seed": 28670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_0/index_3814_seed_28670_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.22633017860352994, + "vlm_logprob": -0.22633017860352994, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.22633017860352994, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03170374408364296 + ], + "opposite_logprob_tokens": [ + -3.9067037105560303 + ], + "logprob_reward": -0.03170374408364296, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03170374408364296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2864699363708496, + 0.05709965527057648, + 0.21769969165325165, + 0.058871135115623474, + 0.021000249311327934, + 0.11233943700790405, + 0.06547949463129044, + 0.029884636402130127, + 0.09974513202905655, + 0.05141056701540947 + ], + "credit_map": [ + 0.021401727572083473, + 0.01937585137784481, + 0.021756306290626526, + 0.014161120168864727, + 0.020370302721858025, + 0.08817614614963531, + 0.025986555963754654, + 0.07524192333221436, + 0.004782204981893301, + 0.008418472483754158, + 0.005829497706145048, + 0.005261973012238741, + 0.008680512197315693, + 0.007795742712914944, + 0.008193346671760082, + 0.008137905970215797, + 0.01766018383204937, + 0.008876217529177666, + 0.011016475968062878, + 0.006322542205452919, + 0.010558240115642548, + 0.0840873196721077, + 0.010723872110247612, + 0.06845485419034958, + 0.004884717520326376, + 0.008598090149462223, + 0.004971405956894159, + 0.00500126089900732, + 0.007119475398212671, + 0.011599399149417877, + 0.005698359105736017, + 0.010998429730534554, + 0.0013524091336876154, + 0.0030740711372345686, + 0.0016606285935267806, + 0.0019062525825574994, + 0.0026136061642318964, + 0.0038278137799352407, + 0.002184207085520029, + 0.004381262697279453, + 0.009868299588561058, + 0.015252592042088509, + 0.010400358587503433, + 0.010681916028261185, + 0.013969657011330128, + 0.01987011916935444, + 0.01320724282413721, + 0.019089261069893837, + 0.004452499561011791, + 0.004191252402961254, + 0.003967540338635445, + 0.0035358737222850323, + 0.0053732856176793575, + 0.019659247249364853, + 0.0050137704238295555, + 0.019286032766103745, + 0.0024132656399160624, + 0.003897018264979124, + 0.002694548573344946, + 0.0041303192265331745, + 0.0037859557196497917, + 0.003370271297171712, + 0.004524702206254005, + 0.005068555008620024, + 0.008055600337684155, + 0.011527937836945057, + 0.01197391003370285, + 0.010461339727044106, + 0.02127567119896412, + 0.01074553094804287, + 0.01279508601874113, + 0.012910062447190285, + 0.0034217997454106808, + 0.00403547240421176, + 0.002433891873806715, + 0.002649783156812191, + 0.016286052763462067, + 0.0035025635734200478, + 0.003368481295183301, + 0.015712525695562363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.03170374408364296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07009617239236832 + ], + "opposite_logprob_tokens": [ + -2.695096254348755 + ], + "logprob_reward": -0.07009617239236832, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07009617239236832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19722411036491394, + 0.14172916114330292, + 0.1457945555448532, + 0.13824597001075745, + 0.07718522101640701, + 0.07790964841842651, + 0.04173488914966583, + 0.048947934061288834, + 0.02627730183303356, + 0.10495129972696304 + ], + "credit_map": [ + 0.015997737646102905, + 0.006959773600101471, + 0.012166504748165607, + 0.00726367486640811, + 0.010276383720338345, + 0.07489526271820068, + 0.009352048859000206, + 0.06031268835067749, + 0.012979238294064999, + 0.010823356918990612, + 0.01267976500093937, + 0.007104301825165749, + 0.010137692093849182, + 0.04330674558877945, + 0.009842893108725548, + 0.034855153411626816, + 0.011761676520109177, + 0.0033564509358257055, + 0.006728179752826691, + 0.0030246505048125982, + 0.004877045284956694, + 0.060586877167224884, + 0.0051080528646707535, + 0.050351619720458984, + 0.010559795424342155, + 0.010262725874781609, + 0.01482323370873928, + 0.00621795142069459, + 0.010390054434537888, + 0.03969522565603256, + 0.009492462500929832, + 0.03680449724197388, + 0.006680366117507219, + 0.003123178845271468, + 0.008221342228353024, + 0.0031663035042583942, + 0.0037946486845612526, + 0.025065073743462563, + 0.004265340976417065, + 0.022868957370519638, + 0.0059967199340462685, + 0.010974847711622715, + 0.008578109554946423, + 0.007158044259995222, + 0.010390901006758213, + 0.011389168910682201, + 0.009709395468235016, + 0.013712451793253422, + 0.002713299123570323, + 0.0019884260836988688, + 0.0015690975124016404, + 0.0015521157765761018, + 0.004818549379706383, + 0.01345803588628769, + 0.0035873898304998875, + 0.012047974392771721, + 0.00419688131660223, + 0.00681539811193943, + 0.0069649117067456245, + 0.007529761176556349, + 0.0050276839174330235, + 0.005970377940684557, + 0.005095369648188353, + 0.007347546052187681, + 0.0019747340120375156, + 0.0023482288233935833, + 0.003269260749220848, + 0.0033129604998975992, + 0.004537602886557579, + 0.003589172614738345, + 0.003507712623104453, + 0.0037376265972852707, + 0.009301085956394672, + 0.0080041429027915, + 0.0067533086985349655, + 0.006986191961914301, + 0.028379814699292183, + 0.008171484805643559, + 0.009857965633273125, + 0.027497297152876854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07009617239236832 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009811745956540108 + ], + "opposite_logprob_tokens": [ + -4.634811878204346 + ], + "logprob_reward": -0.009811745956540108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009811745956540108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1114552691578865, + 0.19668009877204895, + 0.07543573528528214, + 0.07764443755149841, + 0.05627316236495972, + 0.020618820562958717, + 0.03643041104078293, + 0.03748702257871628, + 0.30355334281921387, + 0.08442176133394241 + ], + "credit_map": [ + 0.010595280677080154, + 0.008511047810316086, + 0.012277425266802311, + 0.007630628999322653, + 0.013118065893650055, + 0.027007756754755974, + 0.009145739488303661, + 0.023169323801994324, + 0.018210839480161667, + 0.023261889815330505, + 0.017027538269758224, + 0.017544128000736237, + 0.026746224611997604, + 0.036086563020944595, + 0.024343246594071388, + 0.03345966711640358, + 0.006689788773655891, + 0.0043447487987577915, + 0.005610965192317963, + 0.0032256287522614002, + 0.004494502209126949, + 0.02575122006237507, + 0.004537846427410841, + 0.02078104391694069, + 0.00510438671335578, + 0.012841678224503994, + 0.006325222086161375, + 0.006023820489645004, + 0.010473965667188168, + 0.015647174790501595, + 0.005415779072791338, + 0.015812410041689873, + 0.004259012173861265, + 0.004714353941380978, + 0.0041013783775269985, + 0.004281642846763134, + 0.004895319230854511, + 0.015755489468574524, + 0.004685672000050545, + 0.013580293394625187, + 0.002340927952900529, + 0.0018776266369968653, + 0.002793773775920272, + 0.002111115027219057, + 0.003072794061154127, + 0.0028076923917979, + 0.002158448798581958, + 0.003456443315371871, + 0.002499327063560486, + 0.0029762357007712126, + 0.002577085979282856, + 0.0037601750809699297, + 0.00694150198251009, + 0.0054580108262598515, + 0.0054298886097967625, + 0.006788184866309166, + 0.0031886573415249586, + 0.003398086642846465, + 0.005281374324113131, + 0.0040159630589187145, + 0.003602398559451103, + 0.006921351887285709, + 0.004869513213634491, + 0.006209678947925568, + 0.02396429516375065, + 0.03673933446407318, + 0.033578045666217804, + 0.030979584902524948, + 0.06497050076723099, + 0.03453240543603897, + 0.03854105621576309, + 0.040248095989227295, + 0.005471927113831043, + 0.005610054824501276, + 0.005716082639992237, + 0.006075562443584204, + 0.02630968950688839, + 0.007295450661331415, + 0.005404104478657246, + 0.022538885474205017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.009811745956540108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328848600387573 + ], + "opposite_logprob_tokens": [ + -0.7578848600387573 + ], + "logprob_reward": -0.6328848600387573, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328848600387573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06304328143596649, + 0.12865357100963593, + 0.17088273167610168, + 0.19984687864780426, + 0.03999607637524605, + 0.05177353322505951, + 0.0370117649435997, + 0.09319136291742325, + 0.12383068352937698, + 0.09177009016275406 + ], + "credit_map": [ + 0.0072379689663648605, + 0.00925226416438818, + 0.005784482695162296, + 0.006651290692389011, + 0.010224337689578533, + 0.009257912635803223, + 0.0065485951490700245, + 0.00808642990887165, + 0.012151414528489113, + 0.012785432860255241, + 0.013448089361190796, + 0.01127929612994194, + 0.01789947599172592, + 0.023167820647358894, + 0.016327228397130966, + 0.021594811230897903, + 0.014154586009681225, + 0.004438788164407015, + 0.008270636200904846, + 0.0036500901915133, + 0.006390870548784733, + 0.07089457660913467, + 0.005244285333901644, + 0.05783889442682266, + 0.016825461760163307, + 0.013872913084924221, + 0.018748639151453972, + 0.016421813517808914, + 0.01584213599562645, + 0.053780920803546906, + 0.01769285276532173, + 0.04666215181350708, + 0.0036866264417767525, + 0.005845184437930584, + 0.0043878029100596905, + 0.0038596822414547205, + 0.004101203288882971, + 0.006571544799953699, + 0.004356211982667446, + 0.007187817711383104, + 0.003770491573959589, + 0.00702943280339241, + 0.0035839544143527746, + 0.005242061801254749, + 0.006849698256701231, + 0.01078958809375763, + 0.004864194430410862, + 0.009644110687077045, + 0.002676117466762662, + 0.003910164814442396, + 0.004047758877277374, + 0.0034465764183551073, + 0.005797416437417269, + 0.00608735391870141, + 0.005473346449434757, + 0.005573029164224863, + 0.009215271100401878, + 0.012196872383356094, + 0.00963526125997305, + 0.010983237065374851, + 0.011150258593261242, + 0.010162155143916607, + 0.012526012025773525, + 0.017322296276688576, + 0.011339276097714901, + 0.013667363673448563, + 0.013858038932085037, + 0.01448020339012146, + 0.023474683985114098, + 0.015035734511911869, + 0.015769485384225845, + 0.01620589755475521, + 0.00681088725104928, + 0.006428224965929985, + 0.006399837322533131, + 0.0069748833775520325, + 0.02825656160712242, + 0.006643515080213547, + 0.005652028601616621, + 0.024604152888059616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.6328848600387573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an imperfectly shaped cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice cube resting on a polished, dark, speckled granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature crisp morning light creating sharp highlights and revealing internal air bubbles and cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cube split cleanly in two after a microscopic fracture appears and slides a millimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38715437054634094 + ], + "opposite_logprob_tokens": [ + -1.1371543407440186 + ], + "logprob_reward": -0.38715437054634094, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38715437054634094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09265279769897461, + 0.06870504468679428, + 0.1117059662938118, + 0.12788987159729004, + 0.04247337952256203, + 0.048993565142154694, + 0.17201484739780426, + 0.06969071924686432, + 0.04196352884173393, + 0.22391028702259064 + ], + "credit_map": [ + 0.010337451472878456, + 0.013342232443392277, + 0.012563345022499561, + 0.007352790795266628, + 0.012965086847543716, + 0.011624399572610855, + 0.012250430881977081, + 0.012217056937515736, + 0.004881347995251417, + 0.012857349589467049, + 0.005343026015907526, + 0.008337737061083317, + 0.01094964798539877, + 0.007634519599378109, + 0.009071839042007923, + 0.009629583917558193, + 0.008820457383990288, + 0.0031366480980068445, + 0.004900382831692696, + 0.0023979567922651768, + 0.004073567222803831, + 0.04641233757138252, + 0.0033137656282633543, + 0.03865085914731026, + 0.011852381750941277, + 0.006937640253454447, + 0.014069094322621822, + 0.008408921770751476, + 0.008177596144378185, + 0.03746382147073746, + 0.007736338768154383, + 0.03324408456683159, + 0.003654377069324255, + 0.004321915563195944, + 0.002322955522686243, + 0.003125872230157256, + 0.0034092010464519262, + 0.012209858745336533, + 0.003194199874997139, + 0.010234998539090157, + 0.004945012275129557, + 0.006530775688588619, + 0.00409332662820816, + 0.004513964522629976, + 0.008726751431822777, + 0.007776563987135887, + 0.0054840450175106525, + 0.006923127453774214, + 0.011200032196938992, + 0.011579535901546478, + 0.010333506390452385, + 0.012861551716923714, + 0.02260301634669304, + 0.04570598527789116, + 0.014440743252635002, + 0.043290477246046066, + 0.005464023444801569, + 0.009490320459008217, + 0.005483862943947315, + 0.010270696133375168, + 0.009293235838413239, + 0.006857337895780802, + 0.011000514961779118, + 0.011830728501081467, + 0.003850502660498023, + 0.00451787980273366, + 0.004551074001938105, + 0.005103617440909147, + 0.007119850721210241, + 0.006194923538714647, + 0.004625316709280014, + 0.0060003637336194515, + 0.01405332051217556, + 0.01065684575587511, + 0.00844937190413475, + 0.011177381500601768, + 0.07977048307657242, + 0.014278463087975979, + 0.01254272647202015, + 0.07298169285058975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.38715437054634094 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_tensor([5838, 5838])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_tensor([5838, 5838])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2785c820864b3101f86badab0559e8eebb404144 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_tensor([5838, 5838])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Using a striking, upward-sweeping, ultra-wide-angle lens, the camera remains stationary, capturing the entirety of a vast, circular stained-glass window embedded within the stone walls of a magnificent building. The window, composed of countless panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green, is framed by thin, sturdy strips of black lead, creating intricate geometric and mythological patterns. Filtered by the intense, holy white light from outside, the entire window exudes a majestic, psychedelic brilliance. Suddenly, in absolute silence, a star-shaped crack appears without warning in the very center of the window. Then, as if infused with life, like white lightning, it spreads wildly in all directions at a breathtaking speed visible to the naked eye, emitting a crisp, high-pitched, and continuous crackling sound. Countless forking cracks race across the glass, shattering the once harmonious pattern. Light pierces through the newly created crevices, casting a chaotic, ever-changing pattern across the dimly lit room. Finally, when the crack touched the edge of the window, the entire huge stained glass window suddenly exploded inward with a dull bang, and thousands of colorful glass fragments turned into a deadly and gorgeous gem storm, carrying dust, flying, rolling, and flashing in the air, and finally pouring down like raindrops, creating an epic visual spectacle full of destructive beauty and a suffocating atmosphere of tension.", + "seed": 31265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_5838_seed_31265_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2982493128161877, + "vlm_logprob": -0.2982493128161877, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.2982493128161877, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4172537326812744 + ], + "opposite_logprob_tokens": [ + -0.2922537326812744 + ], + "logprob_reward": -1.4172537326812744, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4172537326812744, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09051012992858887, + 0.03173965588212013, + 0.02684338018298149, + 0.4732695519924164, + 0.02779088355600834, + 0.03051755391061306, + 0.06220609322190285, + 0.12151575088500977, + 0.068177230656147, + 0.06742982566356659 + ], + "credit_map": [ + 0.0075525883585214615, + 0.007411320693790913, + 0.01223191898316145, + 0.02394246682524681, + 0.0070470767095685005, + 0.005512351635843515, + 0.02102632075548172, + 0.0057860794477164745, + 0.0028734670486301184, + 0.0034193487372249365, + 0.004689852241426706, + 0.007314424030482769, + 0.0031590776052325964, + 0.002564407652243972, + 0.005451326258480549, + 0.0022677513770759106, + 0.002568204887211323, + 0.002857098588719964, + 0.0035295234993100166, + 0.006437243893742561, + 0.0024318918585777283, + 0.0023763945791870356, + 0.004664203617721796, + 0.0019788185600191355, + 0.029554324224591255, + 0.030308086425065994, + 0.05649483576416969, + 0.1594068557024002, + 0.02594623528420925, + 0.023636899888515472, + 0.1285477727651596, + 0.019374560564756393, + 0.0024976821150630713, + 0.002763603813946247, + 0.0030051900539547205, + 0.008568228222429752, + 0.0018528796499595046, + 0.0027886591851711273, + 0.004405096173286438, + 0.0019095444586127996, + 0.0034316773526370525, + 0.003755054669454694, + 0.002736163092777133, + 0.004091913811862469, + 0.00594155490398407, + 0.002294780919328332, + 0.0046106064692139626, + 0.003655805252492428, + 0.004402694758027792, + 0.0038288445211946964, + 0.0030069442000240088, + 0.023617373779416084, + 0.015611598268151283, + 0.0029934158083051443, + 0.005883466452360153, + 0.0028617577627301216, + 0.009364086203277111, + 0.013365240767598152, + 0.015538248233497143, + 0.02277771383523941, + 0.013723124749958515, + 0.008856486529111862, + 0.027591275051236153, + 0.010299586690962315, + 0.005565033759921789, + 0.00962006114423275, + 0.008593576028943062, + 0.009212708100676537, + 0.006288764998316765, + 0.006014601327478886, + 0.014945474453270435, + 0.007937014102935791, + 0.005623928736895323, + 0.008260950446128845, + 0.006618105806410313, + 0.009643018245697021, + 0.008660271763801575, + 0.00811897125095129, + 0.015790477395057678, + 0.004714100621640682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -1.4172537326812744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014222566038370132 + ], + "opposite_logprob_tokens": [ + -4.264222621917725 + ], + "logprob_reward": -0.014222566038370132, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014222566038370132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07039406895637512, + 0.025705337524414062, + 0.029677052050828934, + 0.3749566078186035, + 0.047207802534103394, + 0.0589936301112175, + 0.2167288064956665, + 0.05846044048666954, + 0.040426209568977356, + 0.07745000720024109 + ], + "credit_map": [ + 0.0062393001280725, + 0.0076622371561825275, + 0.007658469956368208, + 0.016604529693722725, + 0.006660629063844681, + 0.005783866159617901, + 0.015123575925827026, + 0.004661457613110542, + 0.0024731995072215796, + 0.0025443206541240215, + 0.003980614710599184, + 0.004859299398958683, + 0.0022623257245868444, + 0.0027237702161073685, + 0.004189446568489075, + 0.002672360045835376, + 0.002785720629617572, + 0.003239062149077654, + 0.004398434888571501, + 0.005344975274056196, + 0.0031319228000938892, + 0.0030010065529495478, + 0.005233545321971178, + 0.0025423825718462467, + 0.022963732481002808, + 0.022486157715320587, + 0.04800179600715637, + 0.13066302239894867, + 0.018376052379608154, + 0.01844828948378563, + 0.0984288677573204, + 0.015588711015880108, + 0.004143489990383387, + 0.006836798973381519, + 0.003542480757459998, + 0.008650464937090874, + 0.00408027321100235, + 0.00965805258601904, + 0.007044461090117693, + 0.0032517840154469013, + 0.0052504329942166805, + 0.007184954360127449, + 0.0077851563692092896, + 0.006990743800997734, + 0.00820424035191536, + 0.004953617695719004, + 0.007813230156898499, + 0.010811255313456059, + 0.01300471555441618, + 0.012697553262114525, + 0.009507980197668076, + 0.08537379652261734, + 0.055156294256448746, + 0.011394563131034374, + 0.021114449948072433, + 0.008479461073875427, + 0.0058451141230762005, + 0.00511600635945797, + 0.012445464730262756, + 0.00915283802896738, + 0.00659994687885046, + 0.005447038449347019, + 0.006569845136255026, + 0.00728418817743659, + 0.0028628231957554817, + 0.0063490536995232105, + 0.004561333451420069, + 0.0055871084332466125, + 0.0038444786332547665, + 0.0032656642142683268, + 0.009533693082630634, + 0.004422052297741175, + 0.005517886020243168, + 0.009503395296633244, + 0.0071906428784132, + 0.011743379756808281, + 0.01057781558483839, + 0.007794007658958435, + 0.019363777711987495, + 0.005759099498391151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.014222566038370132 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647141046822071 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008647141046822071, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647141046822071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0695660263299942, + 0.09505368769168854, + 0.12941184639930725, + 0.24548928439617157, + 0.08804421126842499, + 0.07616186887025833, + 0.13709931075572968, + 0.0738314762711525, + 0.06105389446020126, + 0.024288363754749298 + ], + "credit_map": [ + 0.006585749331861734, + 0.006571880541741848, + 0.009970920160412788, + 0.01738358475267887, + 0.005680916365236044, + 0.005372173618525267, + 0.011788664385676384, + 0.006212148349732161, + 0.0073215845040977, + 0.007518957369029522, + 0.012563642114400864, + 0.024190425872802734, + 0.008325621485710144, + 0.005133248399943113, + 0.024095894768834114, + 0.0059043290093541145, + 0.009491357952356339, + 0.008867028169333935, + 0.008153753355145454, + 0.036594241857528687, + 0.010856456123292446, + 0.006908348295837641, + 0.04133283346891403, + 0.007207823917269707, + 0.018233783543109894, + 0.01801343262195587, + 0.03208572044968605, + 0.07266213744878769, + 0.0161290243268013, + 0.012916109524667263, + 0.06367400288581848, + 0.011775086633861065, + 0.006855495274066925, + 0.008750529028475285, + 0.003941492177546024, + 0.03195754438638687, + 0.00401341961696744, + 0.01580161042511463, + 0.012261072173714638, + 0.004463053308427334, + 0.007174371276050806, + 0.013421480543911457, + 0.009060001000761986, + 0.00822235643863678, + 0.01607624441385269, + 0.007094906643033028, + 0.007924416102468967, + 0.007188099902123213, + 0.009721461683511734, + 0.017231693491339684, + 0.011841071769595146, + 0.03672204911708832, + 0.03491644561290741, + 0.0070769302546978, + 0.010611901991069317, + 0.008977774530649185, + 0.00659123295918107, + 0.007636815309524536, + 0.013194563798606396, + 0.012541349045932293, + 0.008850917220115662, + 0.005420250352472067, + 0.011527188122272491, + 0.008069164119660854, + 0.00585072161629796, + 0.011087113991379738, + 0.005360743030905724, + 0.007902508601546288, + 0.00808533001691103, + 0.004968542605638504, + 0.012783688493072987, + 0.005015247967094183, + 0.002782762749120593, + 0.0038938841316848993, + 0.0030063872691243887, + 0.002718260744586587, + 0.0033499847631901503, + 0.002604366047307849, + 0.004025963135063648, + 0.0019067557295784354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.008647141046822071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024899214040488005 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.0024899214040488005, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024899214040488005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10738644003868103, + 0.04663873463869095, + 0.027501266449689865, + 0.4761410653591156, + 0.0329681821167469, + 0.0562748908996582, + 0.05944807082414627, + 0.10455375909805298, + 0.036349810659885406, + 0.052737731486558914 + ], + "credit_map": [ + 0.009177474305033684, + 0.008266321383416653, + 0.013957219198346138, + 0.027730725705623627, + 0.00881114974617958, + 0.00698617659509182, + 0.02543284371495247, + 0.007024534046649933, + 0.00450846878811717, + 0.0044428640976548195, + 0.007849812507629395, + 0.01039524469524622, + 0.0048113251104950905, + 0.0035518500953912735, + 0.00752573786303401, + 0.0035534382332116365, + 0.0027206267695873976, + 0.0025152929592877626, + 0.0041044638492167, + 0.006271079648286104, + 0.0024804812856018543, + 0.0021571600809693336, + 0.005064794328063726, + 0.002187372650951147, + 0.029985375702381134, + 0.030609624460339546, + 0.056822557002305984, + 0.15956702828407288, + 0.026406243443489075, + 0.023936428129673004, + 0.12915021181106567, + 0.019663676619529724, + 0.0031814626418054104, + 0.0037029467057436705, + 0.0039175874553620815, + 0.005283791571855545, + 0.002654844429343939, + 0.008096149191260338, + 0.003756333841010928, + 0.002375067211687565, + 0.005602309480309486, + 0.008456972427666187, + 0.004338196944445372, + 0.009405532851815224, + 0.006925972644239664, + 0.0043981303460896015, + 0.01261854451149702, + 0.004529241472482681, + 0.005786489229649305, + 0.007231590338051319, + 0.00496731698513031, + 0.01010307390242815, + 0.007582452148199081, + 0.00450406176969409, + 0.014437546022236347, + 0.004835549276322126, + 0.007844851352274418, + 0.011222991161048412, + 0.012607268989086151, + 0.020649544894695282, + 0.011595071293413639, + 0.00728246383368969, + 0.024756701663136482, + 0.008594873361289501, + 0.002417423529550433, + 0.005821432452648878, + 0.003339733462780714, + 0.004906227346509695, + 0.00374647811986506, + 0.0028340930584818125, + 0.009207721799612045, + 0.004076705314218998, + 0.004200777970254421, + 0.0056956191547214985, + 0.005374494008719921, + 0.007907738909125328, + 0.006991636473685503, + 0.006340536754578352, + 0.01256597600877285, + 0.0036609552334994078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.0024899214040488005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863320291042328 + ], + "opposite_logprob_tokens": [ + -3.048633098602295 + ], + "logprob_reward": -0.04863320291042328, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863320291042328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06891439855098724, + 0.04485907405614853, + 0.02837102860212326, + 0.25324779748916626, + 0.1997016966342926, + 0.05052759125828743, + 0.06407685577869415, + 0.18827946484088898, + 0.06290755420923233, + 0.0391145795583725 + ], + "credit_map": [ + 0.006349800620228052, + 0.006748383864760399, + 0.004913795739412308, + 0.016504641622304916, + 0.005538681522011757, + 0.0057581220753490925, + 0.018404189497232437, + 0.004696778953075409, + 0.004176913294941187, + 0.004448976833373308, + 0.00826206337660551, + 0.008251616731286049, + 0.004808209370821714, + 0.003398273838683963, + 0.007654192857444286, + 0.0038588286843150854, + 0.00226385029964149, + 0.0026743165217339993, + 0.002626305678859353, + 0.0066725206561386585, + 0.0025014164857566357, + 0.002074379473924637, + 0.0077445171773433685, + 0.001813719398342073, + 0.01744256168603897, + 0.016016574576497078, + 0.03590429946780205, + 0.08133718371391296, + 0.013818953186273575, + 0.013044879771769047, + 0.06385841220617294, + 0.01182490959763527, + 0.015769531950354576, + 0.020796731114387512, + 0.014568707905709743, + 0.08075257390737534, + 0.00909758172929287, + 0.016843348741531372, + 0.03310464695096016, + 0.008768553845584393, + 0.00600087083876133, + 0.005790695082396269, + 0.003234483301639557, + 0.009396330453455448, + 0.007103812415152788, + 0.006170155014842749, + 0.008844677358865738, + 0.003986566327512264, + 0.004024128895252943, + 0.0050740912556648254, + 0.004345833323895931, + 0.018093498423695564, + 0.015141868032515049, + 0.004354118835180998, + 0.008922473527491093, + 0.004120846278965473, + 0.013197866268455982, + 0.027754414826631546, + 0.016326718032360077, + 0.03958417475223541, + 0.019095540046691895, + 0.012501831166446209, + 0.04551686346530914, + 0.01430206187069416, + 0.0053571173921227455, + 0.007061472162604332, + 0.004828429315239191, + 0.0117914704605937, + 0.009096404537558556, + 0.004207382909953594, + 0.014222811907529831, + 0.006342459470033646, + 0.003729140618816018, + 0.005352536682039499, + 0.0033578036818653345, + 0.007863597013056278, + 0.004554004408419132, + 0.003019101219251752, + 0.00886907521635294, + 0.002369314432144165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.04863320291042328 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Using a striking, upward-sweeping, ultra-wide-angle lens, the camera remains stationary, capturing the entirety of a vast, circular stained-glass window embedded within the stone walls of a magnificent building. The window, composed of countless panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green, is framed by thin, sturdy strips of black lead, creating intricate geometric and mythological patterns. Filtered by the intense, holy white light from outside, the entire window exudes a majestic, psychedelic brilliance. Suddenly, in absolute silence, a star-shaped crack appears without warning in the very center of the window. Then, as if infused with life, like white lightning, it spreads wildly in all directions at a breathtaking speed visible to the naked eye, emitting a crisp, high-pitched, and continuous crackling sound. Countless forking cracks race across the glass, shattering the once harmonious pattern. Light pierces through the newly created crevices, casting a chaotic, ever-changing pattern across the dimly lit room. Finally, when the crack touched the edge of the window, the entire huge stained glass window suddenly exploded inward with a dull bang, and thousands of colorful glass fragments turned into a deadly and gorgeous gem storm, carrying dust, flying, rolling, and flashing in the air, and finally pouring down like raindrops, creating an epic visual spectacle full of destructive beauty and a suffocating atmosphere of tension.", + "seed": 31266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_5838_seed_31266_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1890915178693831, + "vlm_logprob": -0.1890915178693831, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.1890915178693831, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9085713624954224 + ], + "opposite_logprob_tokens": [ + -0.5335713624954224 + ], + "logprob_reward": -0.9085713624954224, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9085713624954224, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1873045414686203, + 0.06806237250566483, + 0.04715506359934807, + 0.07939442247152328, + 0.09533824026584625, + 0.150795117020607, + 0.093169666826725, + 0.08203443139791489, + 0.11823432147502899, + 0.07851183414459229 + ], + "credit_map": [ + 0.026925474405288696, + 0.03319895267486572, + 0.0139356954023242, + 0.013435006141662598, + 0.03326881676912308, + 0.02423258312046528, + 0.027269717305898666, + 0.01503828540444374, + 0.009469154290854931, + 0.011435036547482014, + 0.005775861442089081, + 0.005386300850659609, + 0.010254870168864727, + 0.008781296201050282, + 0.0098378024995327, + 0.00712205097079277, + 0.0059262970462441444, + 0.006722318008542061, + 0.0049005248583853245, + 0.007606602273881435, + 0.004460087977349758, + 0.00630019698292017, + 0.007243121974170208, + 0.003995917271822691, + 0.026116766035556793, + 0.005548039451241493, + 0.006333629600703716, + 0.006437873933464289, + 0.0053396448493003845, + 0.008916499093174934, + 0.01615794561803341, + 0.00454401969909668, + 0.022345853969454765, + 0.011607097461819649, + 0.008561973460018635, + 0.00782441534101963, + 0.010913023725152016, + 0.01260499469935894, + 0.015972279012203217, + 0.005508603062480688, + 0.020853834226727486, + 0.021091386675834656, + 0.01165135856717825, + 0.01735015958547592, + 0.014612688682973385, + 0.028670212253928185, + 0.018142184242606163, + 0.018423281610012054, + 0.00942915491759777, + 0.01331267599016428, + 0.008997423574328423, + 0.009841310791671276, + 0.022601168602705002, + 0.009113493375480175, + 0.01399318128824234, + 0.005881250835955143, + 0.006872951984405518, + 0.010238387621939182, + 0.006126668304204941, + 0.010905099101364613, + 0.018269479274749756, + 0.01068494375795126, + 0.011880764737725258, + 0.0070561375468969345, + 0.020933369174599648, + 0.01166724693030119, + 0.01577974110841751, + 0.012903385795652866, + 0.013998729176819324, + 0.014090141281485558, + 0.016753489151597023, + 0.01210822258144617, + 0.009797130711376667, + 0.012557534500956535, + 0.00769426953047514, + 0.0073266527615487576, + 0.016011489555239677, + 0.009741537272930145, + 0.009217462502419949, + 0.006165753584355116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.9085713624954224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659431710839272 + ], + "opposite_logprob_tokens": [ + -4.758659362792969 + ], + "logprob_reward": -0.008659431710839272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008659431710839272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09931861609220505, + 0.09708637744188309, + 0.04115024581551552, + 0.10617543756961823, + 0.30102282762527466, + 0.14170150458812714, + 0.04067527875304222, + 0.054764240980148315, + 0.06453248113393784, + 0.053573016077280045 + ], + "credit_map": [ + 0.012634380720555782, + 0.01614946313202381, + 0.008849899284541607, + 0.010308428667485714, + 0.017322853207588196, + 0.012014496140182018, + 0.012891760095953941, + 0.00914732925593853, + 0.011460613459348679, + 0.017004648223519325, + 0.007587290368974209, + 0.007469017058610916, + 0.01784057542681694, + 0.012068284675478935, + 0.014125692658126354, + 0.009530250914394855, + 0.006101024337112904, + 0.005184410139918327, + 0.004215739201754332, + 0.0057723987847566605, + 0.004601446911692619, + 0.005398697219789028, + 0.005165698006749153, + 0.004710828885436058, + 0.030504627153277397, + 0.008529523387551308, + 0.009905576705932617, + 0.009950491599738598, + 0.007804345339536667, + 0.01069826353341341, + 0.020469482988119125, + 0.008313125930726528, + 0.04113772138953209, + 0.033985137939453125, + 0.04258114472031593, + 0.0366441048681736, + 0.03925531357526779, + 0.033111099153757095, + 0.03447413071990013, + 0.03983419016003609, + 0.01810290478169918, + 0.015526853501796722, + 0.011329038999974728, + 0.021081820130348206, + 0.012770761735737324, + 0.025167711079120636, + 0.02063891850411892, + 0.01708349399268627, + 0.003172273514792323, + 0.004875000566244125, + 0.0036510012578219175, + 0.006122383289039135, + 0.006389081012457609, + 0.0038527173455804586, + 0.007681149989366531, + 0.004931669216603041, + 0.006095981225371361, + 0.007907046005129814, + 0.007329491898417473, + 0.007017003837972879, + 0.010317638516426086, + 0.005607889965176582, + 0.005779479164630175, + 0.00470970943570137, + 0.006781361065804958, + 0.005619561765342951, + 0.008910049684345722, + 0.009846264496445656, + 0.007381538860499859, + 0.006491299252957106, + 0.014651472680270672, + 0.004850930068641901, + 0.005704110953956842, + 0.0062605408020317554, + 0.008936953730881214, + 0.005515678320080042, + 0.008729900233447552, + 0.006254342384636402, + 0.007231763564050198, + 0.004939729813486338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.008659431710839272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956991575658321 + ], + "opposite_logprob_tokens": [ + -5.130957126617432 + ], + "logprob_reward": -0.005956991575658321, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956991575658321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10871972888708115, + 0.0663953572511673, + 0.13195808231830597, + 0.10700906068086624, + 0.10411381721496582, + 0.1633387804031372, + 0.09795120358467102, + 0.05705796927213669, + 0.11418730765581131, + 0.04926864057779312 + ], + "credit_map": [ + 0.019279299303889275, + 0.013937902636826038, + 0.0110530536621809, + 0.012012368999421597, + 0.010663856752216816, + 0.013116435147821903, + 0.015148423612117767, + 0.013508398085832596, + 0.009632634930312634, + 0.011749944649636745, + 0.005581972189247608, + 0.006329454015940428, + 0.00789246428757906, + 0.006990337744355202, + 0.007837960496544838, + 0.010380587540566921, + 0.018099971115589142, + 0.015622646547853947, + 0.012214791029691696, + 0.017349088564515114, + 0.016836363822221756, + 0.017842138186097145, + 0.020381063222885132, + 0.01361202821135521, + 0.02403685823082924, + 0.01131102629005909, + 0.009524526074528694, + 0.011730387806892395, + 0.013946998864412308, + 0.008660971187055111, + 0.019264811649918556, + 0.008533492684364319, + 0.024287940934300423, + 0.009912217035889626, + 0.012635807506740093, + 0.011344955302774906, + 0.01002743374556303, + 0.013413084670901299, + 0.011692265048623085, + 0.010800124146044254, + 0.025700023397803307, + 0.021616406738758087, + 0.013280869461596012, + 0.016976898536086082, + 0.018728477880358696, + 0.035438161343336105, + 0.012687978334724903, + 0.01890997774899006, + 0.010054654441773891, + 0.013664430007338524, + 0.010221095755696297, + 0.010182755999267101, + 0.022685488685965538, + 0.010120726190507412, + 0.014028765261173248, + 0.006993305403739214, + 0.00528676575049758, + 0.00579657219350338, + 0.0065531060099601746, + 0.010303123854100704, + 0.0069701289758086205, + 0.006142654921859503, + 0.006658907979726791, + 0.009346714243292809, + 0.009499035775661469, + 0.012055858969688416, + 0.01390672568231821, + 0.015316703356802464, + 0.01614040695130825, + 0.01290903426706791, + 0.026961419731378555, + 0.007398136425763369, + 0.007228579837828875, + 0.00504465913400054, + 0.005789034999907017, + 0.005228834692388773, + 0.0068044341169297695, + 0.005589576438069344, + 0.007608171086758375, + 0.005975356791168451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.005956991575658321 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409490754827857 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.00409490754827857, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409490754827857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07743239402770996, + 0.11224262416362762, + 0.04733184725046158, + 0.2673667073249817, + 0.09145643562078476, + 0.22211606800556183, + 0.07736983150243759, + 0.04110454022884369, + 0.04211633652448654, + 0.02146318182349205 + ], + "credit_map": [ + 0.010680398903787136, + 0.01370936632156372, + 0.007094520144164562, + 0.006663339212536812, + 0.014056524261832237, + 0.009951182641088963, + 0.008781497366726398, + 0.006495580542832613, + 0.013904157094657421, + 0.019070912152528763, + 0.008668787777423859, + 0.009604563936591148, + 0.017820268869400024, + 0.014385949820280075, + 0.017924875020980835, + 0.010863121598958969, + 0.00617786543443799, + 0.00618891092017293, + 0.00452471524477005, + 0.007855589501559734, + 0.004109696950763464, + 0.006115293130278587, + 0.008397933095693588, + 0.003961846232414246, + 0.05722571909427643, + 0.026265624910593033, + 0.030481209978461266, + 0.034899886697530746, + 0.028051849454641342, + 0.019593557342886925, + 0.04258653149008751, + 0.02826237678527832, + 0.010722039267420769, + 0.012374134734272957, + 0.009955286979675293, + 0.011692349798977375, + 0.01376975979655981, + 0.007631220389157534, + 0.014628570526838303, + 0.010683083906769753, + 0.032543983310461044, + 0.019031833857297897, + 0.018433963879942894, + 0.022226613014936447, + 0.034585997462272644, + 0.054734036326408386, + 0.018212322145700455, + 0.022347349673509598, + 0.010383479297161102, + 0.010922533459961414, + 0.009280508384108543, + 0.008477970026433468, + 0.014989055693149567, + 0.007795901969075203, + 0.009522837586700916, + 0.005997557193040848, + 0.0043630450963974, + 0.004696371033787727, + 0.004415973089635372, + 0.006786068435758352, + 0.007030767388641834, + 0.004154509864747524, + 0.004126723390072584, + 0.0055310847237706184, + 0.007378465961664915, + 0.0036694577429443598, + 0.0052822185680270195, + 0.004441168159246445, + 0.00538243493065238, + 0.005007030908018351, + 0.00645048450678587, + 0.004505080636590719, + 0.003028204431757331, + 0.003031810512766242, + 0.0024040613789111376, + 0.002348627895116806, + 0.003572239773347974, + 0.0027626212686300278, + 0.0025087380781769753, + 0.0018068791832774878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.00409490754827857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018174896016716957 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.018174896016716957, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018174896016716957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08081873506307602, + 0.13175569474697113, + 0.09368617087602615, + 0.22108681499958038, + 0.05502711981534958, + 0.22558699548244476, + 0.05432639643549919, + 0.04361381381750107, + 0.03530086204409599, + 0.058797407895326614 + ], + "credit_map": [ + 0.013456531800329685, + 0.012827688828110695, + 0.008198323659598827, + 0.005700157955288887, + 0.013211957179009914, + 0.010813593864440918, + 0.008498476818203926, + 0.008112004958093166, + 0.016506033018231392, + 0.02284657210111618, + 0.010641472414135933, + 0.012694831937551498, + 0.021192185580730438, + 0.016768399626016617, + 0.019686056300997734, + 0.011420139111578465, + 0.014522755518555641, + 0.01116021815687418, + 0.007749005686491728, + 0.01403825543820858, + 0.01044608373194933, + 0.01346136350184679, + 0.015452894382178783, + 0.0068555898033082485, + 0.06082192063331604, + 0.021370701491832733, + 0.016176827251911163, + 0.02301720716059208, + 0.0261877179145813, + 0.021309852600097656, + 0.04087948799133301, + 0.011323105543851852, + 0.015229452401399612, + 0.004741017706692219, + 0.005392895545810461, + 0.005908900871872902, + 0.0033999003935605288, + 0.008841006085276604, + 0.006979501340538263, + 0.004534446634352207, + 0.030182290822267532, + 0.023165419697761536, + 0.02680834010243416, + 0.02618255279958248, + 0.022342154756188393, + 0.03244033455848694, + 0.027140839025378227, + 0.03732505068182945, + 0.007250227499753237, + 0.006942701991647482, + 0.006881420500576496, + 0.005129423923790455, + 0.01073440071195364, + 0.006318215746432543, + 0.005055241286754608, + 0.006014768034219742, + 0.005109545309096575, + 0.005280428100377321, + 0.005489697214215994, + 0.0061240773648023605, + 0.007379660848528147, + 0.004738530144095421, + 0.004989864304661751, + 0.004502010531723499, + 0.004480157978832722, + 0.0028570073191076517, + 0.005225806962698698, + 0.004341539461165667, + 0.004765950608998537, + 0.003912748768925667, + 0.006045313086360693, + 0.0036723376251757145, + 0.007482222281396389, + 0.006381161976605654, + 0.01202969066798687, + 0.007249135989695787, + 0.00673192972317338, + 0.00666892621666193, + 0.008185133337974548, + 0.0040692053735256195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.018174896016716957 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Using a striking, upward-sweeping, ultra-wide-angle lens, the camera remains stationary, capturing the entirety of a vast, circular stained-glass window embedded within the stone walls of a magnificent building. The window, composed of countless panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green, is framed by thin, sturdy strips of black lead, creating intricate geometric and mythological patterns. Filtered by the intense, holy white light from outside, the entire window exudes a majestic, psychedelic brilliance. Suddenly, in absolute silence, a star-shaped crack appears without warning in the very center of the window. Then, as if infused with life, like white lightning, it spreads wildly in all directions at a breathtaking speed visible to the naked eye, emitting a crisp, high-pitched, and continuous crackling sound. Countless forking cracks race across the glass, shattering the once harmonious pattern. Light pierces through the newly created crevices, casting a chaotic, ever-changing pattern across the dimly lit room. Finally, when the crack touched the edge of the window, the entire huge stained glass window suddenly exploded inward with a dull bang, and thousands of colorful glass fragments turned into a deadly and gorgeous gem storm, carrying dust, flying, rolling, and flashing in the air, and finally pouring down like raindrops, creating an epic visual spectacle full of destructive beauty and a suffocating atmosphere of tension.", + "seed": 31267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_5838_seed_31267_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18522319714538754, + "vlm_logprob": -0.18522319714538754, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.18522319714538754, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8402788639068604 + ], + "opposite_logprob_tokens": [ + -0.5902788639068604 + ], + "logprob_reward": -0.8402788639068604, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8402788639068604, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.018236326053738594, + 0.009203799068927765, + 0.008525341749191284, + 0.013061405159533024, + 0.03168102353811264, + 0.014693750999867916, + 0.025286493822932243, + 0.5004150867462158, + 0.02211778424680233, + 0.35677891969680786 + ], + "credit_map": [ + 0.0022023110650479794, + 0.0029727222863584757, + 0.0019377578282728791, + 0.0025170836597681046, + 0.002370369853451848, + 0.001999961445108056, + 0.0026818918995559216, + 0.001554229762405157, + 0.0008991176728159189, + 0.0027277495246380568, + 0.0006862442824058235, + 0.0007685258751735091, + 0.0009742170223034918, + 0.0008557472610846162, + 0.0015195956220850348, + 0.0007726024487055838, + 0.0008323623915202916, + 0.0017672982066869736, + 0.0008297119638882577, + 0.0010621841065585613, + 0.0013195107458159328, + 0.0007070412393659353, + 0.0011333681177347898, + 0.0008738654432818294, + 0.0016997276106849313, + 0.002137721050530672, + 0.0024033854715526104, + 0.001271125627681613, + 0.0012788072926923633, + 0.0011543863220140338, + 0.0019045289373025298, + 0.001211723661981523, + 0.002528292592614889, + 0.004707984160631895, + 0.003811266738921404, + 0.004888289608061314, + 0.005818034987896681, + 0.0034235704224556684, + 0.004506231751292944, + 0.0019973565358668566, + 0.0012233866145834327, + 0.0022981136571615934, + 0.0010073662269860506, + 0.0017864605179056525, + 0.0024893744848668575, + 0.0014662932371720672, + 0.003247333923354745, + 0.0011754237348213792, + 0.0023388650733977556, + 0.0032848475966602564, + 0.002872808836400509, + 0.0027099803555756807, + 0.006816159468144178, + 0.0020282703917473555, + 0.0029557901434600353, + 0.002279773121699691, + 0.021115317940711975, + 0.2113017737865448, + 0.00868641585111618, + 0.08718089759349823, + 0.05408918485045433, + 0.02196185104548931, + 0.023928653448820114, + 0.07215102016925812, + 0.0017125036101788282, + 0.006034607999026775, + 0.0017824436072260141, + 0.0028959321789443493, + 0.003521726932376623, + 0.0018248429987579584, + 0.002722611650824547, + 0.0016231162007898092, + 0.017661770805716515, + 0.15284907817840576, + 0.008373042568564415, + 0.10430675745010376, + 0.027991198003292084, + 0.013126885518431664, + 0.019624853506684303, + 0.01284535601735115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.8402788639068604 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565725483000278 + ], + "opposite_logprob_tokens": [ + -4.387565612792969 + ], + "logprob_reward": -0.012565725483000278, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565725483000278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05043330788612366, + 0.011093858629465103, + 0.020151201635599136, + 0.014036615379154682, + 0.05746691673994064, + 0.034953683614730835, + 0.0945751741528511, + 0.2455035299062729, + 0.03342197462916374, + 0.43836385011672974 + ], + "credit_map": [ + 0.005729839205741882, + 0.011596189811825752, + 0.005396849941462278, + 0.005501021631062031, + 0.004990665707737207, + 0.005502423737198114, + 0.006501379888504744, + 0.005214932840317488, + 0.0014619616558775306, + 0.001487800502218306, + 0.0011958831455558538, + 0.0011452491162344813, + 0.001450254931114614, + 0.0012895901454612613, + 0.0016541560180485249, + 0.0014089628821238875, + 0.002729892497882247, + 0.002705275546759367, + 0.0036210250109434128, + 0.002011269563809037, + 0.0019905802328139544, + 0.002503790659829974, + 0.002000603824853897, + 0.0025887624360620975, + 0.001622380455955863, + 0.0017894105985760689, + 0.0018290033331140876, + 0.001567882252857089, + 0.001847078907303512, + 0.0018413284560665488, + 0.0020353663712739944, + 0.0015041633741930127, + 0.007394703570753336, + 0.007391324266791344, + 0.004825452342629433, + 0.005755031015723944, + 0.011897607706487179, + 0.008689086884260178, + 0.007028340827673674, + 0.004485360812395811, + 0.0035153008066117764, + 0.0038493196479976177, + 0.005079900845885277, + 0.0029940966051071882, + 0.0064316908828914165, + 0.004546265117824078, + 0.004636615049093962, + 0.0039004930295050144, + 0.005709929391741753, + 0.02108362503349781, + 0.004553502891212702, + 0.006394816562533379, + 0.03385497257113457, + 0.00927360076457262, + 0.0069847796112298965, + 0.006719931494444609, + 0.010457849130034447, + 0.10345848649740219, + 0.004316622391343117, + 0.04644770175218582, + 0.02571329101920128, + 0.010898659005761147, + 0.012037480250000954, + 0.03217342868447304, + 0.0027081689331680536, + 0.007248392328619957, + 0.0031001584138721228, + 0.00984129123389721, + 0.002712563844397664, + 0.0020791166462004185, + 0.0034953122958540916, + 0.002236970467492938, + 0.021179554983973503, + 0.1870339810848236, + 0.010059447959065437, + 0.13051766157150269, + 0.03344883397221565, + 0.0160616897046566, + 0.024196548387408257, + 0.015866098925471306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.012565725483000278 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076382951810956 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.0076382951810956, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076382951810956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027363887056708336, + 0.012614964507520199, + 0.03731011226773262, + 0.027001509442925453, + 0.08913230150938034, + 0.016956381499767303, + 0.1202024519443512, + 0.14824171364307404, + 0.016102643683552742, + 0.5050740242004395 + ], + "credit_map": [ + 0.0028402768075466156, + 0.0033297608606517315, + 0.0026193251833319664, + 0.0036463653668761253, + 0.003824029816314578, + 0.0037064109928905964, + 0.004750960506498814, + 0.0026467598509043455, + 0.0010419170139357448, + 0.0019030104158446193, + 0.0015258610947057605, + 0.001465765992179513, + 0.002284369897097349, + 0.0010097339982166886, + 0.0023906256537884474, + 0.0009936800925061107, + 0.004937564022839069, + 0.004458440933376551, + 0.005571343004703522, + 0.0035723710898309946, + 0.004000373650342226, + 0.004688003100454807, + 0.004906510468572378, + 0.005175508093088865, + 0.002869328251108527, + 0.0041953809559345245, + 0.003095425432547927, + 0.0038270163349807262, + 0.004070363938808441, + 0.0026133013889193535, + 0.004407023079693317, + 0.0019236706430092454, + 0.012205694802105427, + 0.008735325187444687, + 0.013067628256976604, + 0.00817673746496439, + 0.013553328812122345, + 0.01671025902032852, + 0.008072980679571629, + 0.008610351011157036, + 0.0016277648974210024, + 0.0019505027448758483, + 0.0015125523786991835, + 0.002238344168290496, + 0.0026156750973314047, + 0.0024065752513706684, + 0.0030872896313667297, + 0.0015176788438111544, + 0.011118701659142971, + 0.01603112742304802, + 0.014592441730201244, + 0.009526025503873825, + 0.021563367918133736, + 0.023363132029771805, + 0.012180227786302567, + 0.01182742603123188, + 0.006497043184936047, + 0.062429528683423996, + 0.0027189545799046755, + 0.02532971277832985, + 0.016029682010412216, + 0.006748275365680456, + 0.0072084651328623295, + 0.021280059590935707, + 0.0014398222556337714, + 0.00522278668358922, + 0.0010269643971696496, + 0.0025415162090212107, + 0.0022426636423915625, + 0.0011107797035947442, + 0.0015117097645998001, + 0.001006401376798749, + 0.03141835331916809, + 0.22298306226730347, + 0.01149903703480959, + 0.129353865981102, + 0.04541609808802605, + 0.01955004222691059, + 0.027511345222592354, + 0.017342224717140198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.0076382951810956 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617530455812812 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.003617530455812812, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617530455812812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028309641405940056, + 0.017629427835345268, + 0.029718967154622078, + 0.012739940546452999, + 0.09542009979486465, + 0.026519838720560074, + 0.03425883501768112, + 0.35810109972953796, + 0.03036206215620041, + 0.36694011092185974 + ], + "credit_map": [ + 0.0030142818577587605, + 0.0068788169883191586, + 0.003001227742061019, + 0.0034996680915355682, + 0.004587940871715546, + 0.0021476843394339085, + 0.003372467588633299, + 0.0018075513653457165, + 0.0015307727735489607, + 0.0045596519485116005, + 0.0014130087802186608, + 0.002087229397147894, + 0.0035024990793317556, + 0.0012476934352889657, + 0.0022752052173018456, + 0.001013367436826229, + 0.0030430180486291647, + 0.008528213016688824, + 0.0026123581919819117, + 0.002508334582671523, + 0.004567483905702829, + 0.0022816676646471024, + 0.003827258013188839, + 0.002350633731111884, + 0.0014890528982505202, + 0.0018256878247484565, + 0.0018864184385165572, + 0.0016105977119877934, + 0.001872597960755229, + 0.0013025617226958275, + 0.0018068525241687894, + 0.0009461722802370787, + 0.008251722902059555, + 0.01684127189218998, + 0.007186464499682188, + 0.01465766690671444, + 0.01857084594666958, + 0.008113304153084755, + 0.014689082279801369, + 0.007109747268259525, + 0.0024566580541431904, + 0.004210980609059334, + 0.002002651570364833, + 0.003073571017012, + 0.007157527841627598, + 0.0018541020108386874, + 0.0038252195809036493, + 0.0019391269888728857, + 0.0024596559815108776, + 0.006905744317919016, + 0.0030763621907681227, + 0.0026001809164881706, + 0.010129774920642376, + 0.002557117957621813, + 0.003547598607838154, + 0.0029823994264006615, + 0.01519850268959999, + 0.15115639567375183, + 0.006230596918612719, + 0.06445161998271942, + 0.03827857971191406, + 0.015488334000110626, + 0.017340485006570816, + 0.04995659366250038, + 0.0026625320315361023, + 0.005815827287733555, + 0.001481764018535614, + 0.010582773014903069, + 0.0036723162047564983, + 0.00179697060957551, + 0.0023865995462983847, + 0.001963277580216527, + 0.021410586312413216, + 0.16163325309753418, + 0.008198230527341366, + 0.09577054530382156, + 0.03264530748128891, + 0.014375999569892883, + 0.02012004889547825, + 0.012786127626895905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.003617530455812812 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201557070016861 + ], + "opposite_logprob_tokens": [ + -2.8120155334472656 + ], + "logprob_reward": -0.06201557070016861, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201557070016861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09923562407493591, + 0.020428430289030075, + 0.01725904271006584, + 0.026964867487549782, + 0.15707795321941376, + 0.06238097697496414, + 0.04837996885180473, + 0.30276763439178467, + 0.06316075474023819, + 0.20234468579292297 + ], + "credit_map": [ + 0.011480866931378841, + 0.01775343343615532, + 0.012627820484340191, + 0.008634855970740318, + 0.01023245882242918, + 0.014709779992699623, + 0.01189848966896534, + 0.011897938326001167, + 0.0019465053919702768, + 0.0028789935167878866, + 0.002270336030051112, + 0.0023400753270834684, + 0.0026977125089615583, + 0.0024601586628705263, + 0.0033939036075025797, + 0.0024407454766333103, + 0.0014594296226277947, + 0.0023309977259486914, + 0.0016888848040252924, + 0.002520097652450204, + 0.002585685346275568, + 0.0016740792198106647, + 0.003337146947160363, + 0.0016627219738438725, + 0.0035815155133605003, + 0.003796547884121537, + 0.0035977736115455627, + 0.002734811743721366, + 0.004639008082449436, + 0.003098562126979232, + 0.0030383991543203592, + 0.0024782507680356503, + 0.014318284578621387, + 0.019638119265437126, + 0.014645264483988285, + 0.016861429437994957, + 0.04013705626130104, + 0.018939394503831863, + 0.02145262621343136, + 0.011085785925388336, + 0.007057068403810263, + 0.006706452462822199, + 0.005161354783922434, + 0.007850530557334423, + 0.009676875546574593, + 0.008722314611077309, + 0.011055713519454002, + 0.006150675471872091, + 0.004229249898344278, + 0.0045618414878845215, + 0.005678968504071236, + 0.0066122030839324, + 0.006965850945562124, + 0.006913724355399609, + 0.007359608542174101, + 0.006058523431420326, + 0.012990252114832401, + 0.12394309788942337, + 0.007731378544121981, + 0.05740949511528015, + 0.02992127276957035, + 0.014296961948275566, + 0.015376351773738861, + 0.04109886288642883, + 0.004372235853224993, + 0.01816515251994133, + 0.0037665965501219034, + 0.01424957625567913, + 0.0076797278597950935, + 0.004335636273026466, + 0.0061628688126802444, + 0.004428968299180269, + 0.04555737227201462, + 0.024791138246655464, + 0.010424950160086155, + 0.0602794773876667, + 0.021914983168244362, + 0.011570420116186142, + 0.01920720562338829, + 0.008599149994552135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.06201557070016861 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Using a striking, upward-sweeping, ultra-wide-angle lens, the camera remains stationary, capturing the entirety of a vast, circular stained-glass window embedded within the stone walls of a magnificent building. The window, composed of countless panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green, is framed by thin, sturdy strips of black lead, creating intricate geometric and mythological patterns. Filtered by the intense, holy white light from outside, the entire window exudes a majestic, psychedelic brilliance. Suddenly, in absolute silence, a star-shaped crack appears without warning in the very center of the window. Then, as if infused with life, like white lightning, it spreads wildly in all directions at a breathtaking speed visible to the naked eye, emitting a crisp, high-pitched, and continuous crackling sound. Countless forking cracks race across the glass, shattering the once harmonious pattern. Light pierces through the newly created crevices, casting a chaotic, ever-changing pattern across the dimly lit room. Finally, when the crack touched the edge of the window, the entire huge stained glass window suddenly exploded inward with a dull bang, and thousands of colorful glass fragments turned into a deadly and gorgeous gem storm, carrying dust, flying, rolling, and flashing in the air, and finally pouring down like raindrops, creating an epic visual spectacle full of destructive beauty and a suffocating atmosphere of tension.", + "seed": 31268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_1/index_5838_seed_31268_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2656326306052506, + "vlm_logprob": -0.2656326306052506, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.2656326306052506, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1529898643493652 + ], + "opposite_logprob_tokens": [ + -0.4029898941516876 + ], + "logprob_reward": -1.1529898643493652, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1529898643493652, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20640335977077484, + 0.030943123623728752, + 0.05825318023562431, + 0.2243940234184265, + 0.10463616997003555, + 0.084677554666996, + 0.10215488076210022, + 0.03650970757007599, + 0.06900838762521744, + 0.08301956206560135 + ], + "credit_map": [ + 0.02356361597776413, + 0.02665037102997303, + 0.01959073171019554, + 0.03113248571753502, + 0.020710891112685204, + 0.04118308797478676, + 0.021644122898578644, + 0.02192806825041771, + 0.004675579257309437, + 0.003002183511853218, + 0.0032574974466115236, + 0.0040647112764418125, + 0.0024945756886154413, + 0.00471739936619997, + 0.004532339982688427, + 0.004198837094008923, + 0.007094119675457478, + 0.007940446957945824, + 0.004759412724524736, + 0.008327065035700798, + 0.006368921138346195, + 0.008195781148970127, + 0.006599446292966604, + 0.008967985399067402, + 0.03918536752462387, + 0.01692534051835537, + 0.018832309171557426, + 0.024414993822574615, + 0.01368532981723547, + 0.0660824179649353, + 0.015582310035824776, + 0.02968594618141651, + 0.0128936767578125, + 0.013201919384300709, + 0.010224232450127602, + 0.012404323555529118, + 0.009419855661690235, + 0.022300777956843376, + 0.009289834648370743, + 0.014901545830070972, + 0.012141471728682518, + 0.01094387099146843, + 0.007214381359517574, + 0.015537956729531288, + 0.0062701515853405, + 0.016057124361395836, + 0.009315838105976582, + 0.007196763996034861, + 0.01576094888150692, + 0.010436924174427986, + 0.012469111010432243, + 0.012946601957082748, + 0.011684957891702652, + 0.01136265229433775, + 0.013355591334402561, + 0.014138087630271912, + 0.006820029579102993, + 0.0038899239152669907, + 0.003882773220539093, + 0.0043982844799757, + 0.003241105005145073, + 0.004094504751265049, + 0.0049523478373885155, + 0.005230739247053862, + 0.009162990376353264, + 0.00861426442861557, + 0.0037961830385029316, + 0.01679912582039833, + 0.006180525757372379, + 0.011503768153488636, + 0.00686300452798605, + 0.006088524591177702, + 0.016346067190170288, + 0.009645695798099041, + 0.006701664533466101, + 0.00935231801122427, + 0.00805388018488884, + 0.018558740615844727, + 0.006786387413740158, + 0.007574812043458223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -1.1529898643493652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373682498931885 + ], + "opposite_logprob_tokens": [ + -3.4087367057800293 + ], + "logprob_reward": -0.03373682498931885, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373682498931885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09534871578216553, + 0.016245005652308464, + 0.03619271144270897, + 0.12230871617794037, + 0.06086666136980057, + 0.12276472896337509, + 0.16675658524036407, + 0.02330486848950386, + 0.3350082039833069, + 0.021203795447945595 + ], + "credit_map": [ + 0.012627173215150833, + 0.00984939094632864, + 0.007222154643386602, + 0.012798338197171688, + 0.010074337013065815, + 0.026764193549752235, + 0.008592871949076653, + 0.007420250680297613, + 0.0023002438247203827, + 0.0019551950972527266, + 0.002024702960625291, + 0.0019502189243212342, + 0.0016045474912971258, + 0.001960969064384699, + 0.0016681818524375558, + 0.0027809464372694492, + 0.005149837117642164, + 0.004056127741932869, + 0.003854643553495407, + 0.004505346063524485, + 0.0032352120615541935, + 0.0061761424876749516, + 0.0032509162556380033, + 0.005964485928416252, + 0.02007170580327511, + 0.010029124096035957, + 0.009103348478674889, + 0.015822378918528557, + 0.007362253498286009, + 0.036446671932935715, + 0.009320405311882496, + 0.014152833260595798, + 0.00878230482339859, + 0.006029694806784391, + 0.008695986121892929, + 0.0053961691446602345, + 0.005089935846626759, + 0.008255377411842346, + 0.0037480085156857967, + 0.014869186095893383, + 0.015976682305336, + 0.014291323721408844, + 0.007786547765135765, + 0.022766511887311935, + 0.009974020533263683, + 0.029787782579660416, + 0.01286543533205986, + 0.009316429495811462, + 0.020595338195562363, + 0.02546774223446846, + 0.020618442445993423, + 0.027059003710746765, + 0.01482380460947752, + 0.020327087491750717, + 0.02321462146937847, + 0.014650549739599228, + 0.004850845783948898, + 0.0024118635337799788, + 0.002864932408556342, + 0.00199432997033, + 0.002256357576698065, + 0.0026920444797724485, + 0.00240431004203856, + 0.0038301858585327864, + 0.050681404769420624, + 0.02157777175307274, + 0.014584176242351532, + 0.07915205508470535, + 0.018228847533464432, + 0.10517904162406921, + 0.02970137633383274, + 0.01590351015329361, + 0.0026070107705891132, + 0.004677761811763048, + 0.0017510895850136876, + 0.002104074927046895, + 0.0028226638678461313, + 0.0026369462721049786, + 0.0019045720109716058, + 0.0026996771339327097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.03373682498931885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646813031286001 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646813031286001, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646813031286001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24451930820941925, + 0.09154088795185089, + 0.06147555261850357, + 0.10496407747268677, + 0.04382127895951271, + 0.05990314483642578, + 0.0641743540763855, + 0.02517896704375744, + 0.2583206295967102, + 0.04610176384449005 + ], + "credit_map": [ + 0.031576648354530334, + 0.027202807366847992, + 0.02103683352470398, + 0.03459964320063591, + 0.0229681134223938, + 0.06067504361271858, + 0.024776378646492958, + 0.021683834493160248, + 0.00970422849059105, + 0.012449181638658047, + 0.008970319293439388, + 0.014136319980025291, + 0.011198829859495163, + 0.01378118246793747, + 0.010040716268122196, + 0.011260111816227436, + 0.00701000215485692, + 0.006390670780092478, + 0.006235333159565926, + 0.009293535724282265, + 0.005734086502343416, + 0.012812200002372265, + 0.005376213230192661, + 0.008623515255749226, + 0.01886069029569626, + 0.008165705017745495, + 0.008900382556021214, + 0.011251084506511688, + 0.006363246124237776, + 0.03029204159975052, + 0.006897346582263708, + 0.014233579859137535, + 0.007317963056266308, + 0.005027000326663256, + 0.004141231533139944, + 0.00436757504940033, + 0.004132275469601154, + 0.007914043962955475, + 0.004533855244517326, + 0.006387334782630205, + 0.009816441684961319, + 0.006521921139210463, + 0.005419379100203514, + 0.005409287754446268, + 0.005495144985616207, + 0.01591861993074417, + 0.004804209340363741, + 0.006518136709928513, + 0.00725352019071579, + 0.006126935128122568, + 0.007997742854058743, + 0.006499292328953743, + 0.006058470346033573, + 0.006225333083420992, + 0.010903521440923214, + 0.013109538704156876, + 0.004352555610239506, + 0.0026921986136585474, + 0.0030002743005752563, + 0.002701371442526579, + 0.0024497241247445345, + 0.002510249149054289, + 0.0033198774326592684, + 0.004152714274823666, + 0.03843767195940018, + 0.01785426214337349, + 0.010942312888801098, + 0.06382051855325699, + 0.014521487057209015, + 0.0770958960056305, + 0.02340022474527359, + 0.012248270213603973, + 0.00804606918245554, + 0.006354168988764286, + 0.004034341312944889, + 0.004420812241733074, + 0.004116229247301817, + 0.009139172732830048, + 0.004150952212512493, + 0.005840017460286617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.004646813031286001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783060289919376 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783060289919376, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783060289919376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10991987586021423, + 0.018296660855412483, + 0.0325576588511467, + 0.20450834929943085, + 0.24952469766139984, + 0.07532498985528946, + 0.05479171872138977, + 0.015493028797209263, + 0.20202574133872986, + 0.03755723685026169 + ], + "credit_map": [ + 0.015236711129546165, + 0.010633954778313637, + 0.008771488443017006, + 0.014428071677684784, + 0.009981672279536724, + 0.029900625348091125, + 0.01130484975874424, + 0.009662497788667679, + 0.002396966563537717, + 0.0024441287387162447, + 0.0014236600836738944, + 0.003106586402282119, + 0.0017814583843573928, + 0.0029257088899612427, + 0.0023111137561500072, + 0.0019070381531491876, + 0.0046980492770671844, + 0.0028413215186446905, + 0.0036197942681610584, + 0.0034215303603559732, + 0.0027569634839892387, + 0.00726708397269249, + 0.0025746668688952923, + 0.005378248635679483, + 0.03096381574869156, + 0.018120964989066124, + 0.015580127947032452, + 0.02846132591366768, + 0.013492065481841564, + 0.05725963041186333, + 0.018094012513756752, + 0.022536413744091988, + 0.035321399569511414, + 0.03338297829031944, + 0.023399431258440018, + 0.03430407866835594, + 0.02193826623260975, + 0.048906225711107254, + 0.019963717088103294, + 0.03230861574411392, + 0.011732302606105804, + 0.011542853899300098, + 0.005917889531701803, + 0.009197012521326542, + 0.007188106887042522, + 0.015254606492817402, + 0.007203953806310892, + 0.00728825805708766, + 0.004792951513081789, + 0.008014829829335213, + 0.006471299100667238, + 0.005380643531680107, + 0.005737403407692909, + 0.005376382265239954, + 0.0077056558802723885, + 0.01131254993379116, + 0.0024563760962337255, + 0.0017163475276902318, + 0.0021126989740878344, + 0.0015059199649840593, + 0.0014575404347851872, + 0.0015125677455216646, + 0.0019641576800495386, + 0.002767421305179596, + 0.030449572950601578, + 0.01261103805154562, + 0.008084501139819622, + 0.04945868253707886, + 0.011079763062298298, + 0.06228740140795708, + 0.018010739237070084, + 0.010044057853519917, + 0.0072238268330693245, + 0.005117929540574551, + 0.0029927720315754414, + 0.0037581571377813816, + 0.0037705982103943825, + 0.00837307795882225, + 0.0031293542124330997, + 0.003191523952409625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.009783060289919376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast, circular stained-glass window embedded in stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the window contain panels of glass in shades of deep sapphire blue, fiery scarlet, and emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a star-shaped crack appear in the center of the window and spread rapidly in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the window eventually explode inward, showering the room with thousands of flying and rolling glass fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12700659036636353 + ], + "opposite_logprob_tokens": [ + -2.1270065307617188 + ], + "logprob_reward": -0.12700659036636353, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12700659036636353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1691594272851944, + 0.040992531925439835, + 0.039089929312467575, + 0.12069841474294662, + 0.1392063945531845, + 0.1896592527627945, + 0.08460251986980438, + 0.02633548714220524, + 0.16356231272220612, + 0.02669372782111168 + ], + "credit_map": [ + 0.0189970713108778, + 0.022520115599036217, + 0.014628972858190536, + 0.030202731490135193, + 0.020146533846855164, + 0.03285915404558182, + 0.017130058258771896, + 0.012674786150455475, + 0.00663705775514245, + 0.004994921386241913, + 0.004313729703426361, + 0.004399449564516544, + 0.004119550343602896, + 0.008453907445073128, + 0.0039626662619411945, + 0.004111249465495348, + 0.00499656843021512, + 0.0037437661085277796, + 0.003456954611465335, + 0.005468681454658508, + 0.0040953797288239, + 0.008510255254805088, + 0.00396195612847805, + 0.004856365732848644, + 0.019637085497379303, + 0.010441523045301437, + 0.009701531380414963, + 0.016730738803744316, + 0.00814564898610115, + 0.028624024242162704, + 0.009854093194007874, + 0.017563773319125175, + 0.017078209668397903, + 0.020888926461338997, + 0.011849903501570225, + 0.02209199219942093, + 0.01699148677289486, + 0.018794937059283257, + 0.018031196668744087, + 0.013479754328727722, + 0.027413947507739067, + 0.01966380886733532, + 0.01263970136642456, + 0.03087647259235382, + 0.01404014602303505, + 0.0514240637421608, + 0.01899750530719757, + 0.014603597111999989, + 0.011463656090199947, + 0.009074369445443153, + 0.014417084865272045, + 0.008497742004692554, + 0.008275854401290417, + 0.011085662990808487, + 0.010533173568546772, + 0.011254975572228432, + 0.003331041894853115, + 0.0029479668010026217, + 0.0030160422902554274, + 0.0033953844103962183, + 0.003385919379070401, + 0.0034148646518588066, + 0.0030091216322034597, + 0.0038351465482264757, + 0.024118250235915184, + 0.010565505363047123, + 0.00667271763086319, + 0.04006260633468628, + 0.00932181254029274, + 0.049155548214912415, + 0.014578603208065033, + 0.009087261743843555, + 0.0042543415911495686, + 0.004005552735179663, + 0.003189726732671261, + 0.002452730666846037, + 0.0038531979080289602, + 0.003344997065141797, + 0.002546529984101653, + 0.003046649042516947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.12700659036636353 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_tensor([1611, 1611])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_tensor([1611, 1611])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b18ed0bfe8c0b9807c23b490e5a98aa1ddc81478 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_tensor([1611, 1611])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "handheld, slightly high-angle close-up, subtly trembling, intimately tracks the dynamic movement of a smooth paintbrush across paper. The pen, featuring a sleek black plastic tip filled with deep black ink, gracefully meets a sheet of thick, cream-colored, textured paper, held firmly in place. With fluid, confident swirls, the pen glides across the surface, its bristles flexing under deliberate pressure, creating single, bold, and expressive strokes that vary in thickness from delicate hairlines to powerful, broad lines. The damp, inky black ink momentarily shimmers before settling into the paper fibers. The entire scene is rendered in a classic '80s anime style, defined by rich cel shading that casts hard-edged shadows and a hint of film grain across the barrel, lending the image a nostalgic quality and evoking an atmosphere of intimate and focused artistic creation.", + "seed": 9656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_1611_seed_9656_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2529161375015974, + "vlm_logprob": -0.2529161375015974, + "vlm_yes_no_margin": 2.200000023841858, + "vlm_token_credit": -0.2529161375015974, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03661232441663742 + ], + "opposite_logprob_tokens": [ + -4.161612510681152 + ], + "logprob_reward": -0.03661232441663742, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03661232441663742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1623496562242508, + 0.07251417636871338, + 0.13632027804851532, + 0.07067649811506271, + 0.04507435858249664, + 0.08014623075723648, + 0.0621260404586792, + 0.11787309497594833, + 0.20269298553466797, + 0.05022676661610603 + ], + "credit_map": [ + 0.014516549184918404, + 0.01058069709688425, + 0.03067188709974289, + 0.013061813078820705, + 0.012502077035605907, + 0.04245917871594429, + 0.016507113352417946, + 0.02205033041536808, + 0.009702849201858044, + 0.008135136216878891, + 0.012620383873581886, + 0.007269166875630617, + 0.00783994048833847, + 0.010225038975477219, + 0.007351988926529884, + 0.009369662962853909, + 0.014504346996545792, + 0.012270272709429264, + 0.023852556943893433, + 0.013222889974713326, + 0.023788129910826683, + 0.014676141552627087, + 0.019543098285794258, + 0.014462839812040329, + 0.008228612132370472, + 0.006884070113301277, + 0.010306691750884056, + 0.005079370457679033, + 0.011900665238499641, + 0.009520651772618294, + 0.007506620138883591, + 0.011249805800616741, + 0.007798262406140566, + 0.004816513042896986, + 0.00817541591823101, + 0.003562414785847068, + 0.005606840364634991, + 0.004361937753856182, + 0.004223335534334183, + 0.006529633887112141, + 0.004744387697428465, + 0.0040497081354260445, + 0.010088725946843624, + 0.0026520199608057737, + 0.005304777529090643, + 0.0303892083466053, + 0.005887599196285009, + 0.01702979765832424, + 0.007162534631788731, + 0.005522394087165594, + 0.0068463487550616264, + 0.004191715270280838, + 0.011853985488414764, + 0.012425481341779232, + 0.003913882188498974, + 0.010209692642092705, + 0.01701705902814865, + 0.014172698371112347, + 0.016717862337827682, + 0.009263839572668076, + 0.013183754868805408, + 0.01659390516579151, + 0.00819640327244997, + 0.022727562114596367, + 0.032822318375110626, + 0.028619928285479546, + 0.023617148399353027, + 0.01639632321894169, + 0.02803405560553074, + 0.022226840257644653, + 0.021010098978877068, + 0.02996625192463398, + 0.006506501697003841, + 0.0051583582535386086, + 0.006203853525221348, + 0.003312518587335944, + 0.006629240233451128, + 0.009924793615937233, + 0.005199001636356115, + 0.007292494177818298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.03661232441663742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283673465251923 + ], + "opposite_logprob_tokens": [ + -2.017836809158325 + ], + "logprob_reward": -0.14283673465251923, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14283673465251923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10084333270788193, + 0.05136354640126228, + 0.07836131006479263, + 0.062220484018325806, + 0.10558054596185684, + 0.17830240726470947, + 0.054830681532621384, + 0.2587725520133972, + 0.07201026380062103, + 0.03771485015749931 + ], + "credit_map": [ + 0.013120083138346672, + 0.006606813985854387, + 0.019593561068177223, + 0.010938167572021484, + 0.007925170473754406, + 0.019315257668495178, + 0.009434779174625874, + 0.013909505680203438, + 0.0062185004353523254, + 0.006117196287959814, + 0.006995587144047022, + 0.006646382622420788, + 0.0054034581407904625, + 0.007078581023961306, + 0.006030208431184292, + 0.0068736327812075615, + 0.009427101351320744, + 0.006379144266247749, + 0.012600127607584, + 0.006152140907943249, + 0.014556022360920906, + 0.010219926945865154, + 0.009899325668811798, + 0.009127519093453884, + 0.011643701232969761, + 0.004028883296996355, + 0.01364988274872303, + 0.003290631575509906, + 0.008251629769802094, + 0.006669323425740004, + 0.005553937051445246, + 0.009132497012615204, + 0.0108531154692173, + 0.011368053033947945, + 0.013243204914033413, + 0.009749185293912888, + 0.014223725534975529, + 0.012914407067000866, + 0.017625579610466957, + 0.01560327410697937, + 0.014735418371856213, + 0.0137562807649374, + 0.04965247958898544, + 0.009142155759036541, + 0.016500959172844887, + 0.04093162715435028, + 0.011669334955513477, + 0.021914144977927208, + 0.007453096564859152, + 0.00788489356637001, + 0.006339413579553366, + 0.004953401628881693, + 0.007191856857389212, + 0.008229976519942284, + 0.004700989928096533, + 0.00807705707848072, + 0.022612186148762703, + 0.017992286011576653, + 0.02256285585463047, + 0.01385476067662239, + 0.026019543409347534, + 0.07907505333423615, + 0.016557132825255394, + 0.06009873375296593, + 0.011795135214924812, + 0.00839268323034048, + 0.008142706006765366, + 0.005498390179127455, + 0.006669446360319853, + 0.012223797850310802, + 0.007272685877978802, + 0.012015425600111485, + 0.005411925725638866, + 0.0031774307135492563, + 0.004576907027512789, + 0.0024514710530638695, + 0.004771092906594276, + 0.008315169252455235, + 0.0028345882892608643, + 0.006176264025270939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.14283673465251923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604110598564148 + ], + "opposite_logprob_tokens": [ + -1.9104111194610596 + ], + "logprob_reward": -0.1604110598564148, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604110598564148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07998758554458618, + 0.04281982034444809, + 0.12560872733592987, + 0.031775396317243576, + 0.1010666936635971, + 0.27308031916618347, + 0.08122671395540237, + 0.04684516414999962, + 0.17693091928958893, + 0.04065869376063347 + ], + "credit_map": [ + 0.00875858124345541, + 0.008322225883603096, + 0.014492204412817955, + 0.011289514601230621, + 0.008369920775294304, + 0.009024023078382015, + 0.00948441680520773, + 0.010246701538562775, + 0.004745930433273315, + 0.003997165709733963, + 0.006561089772731066, + 0.005291043780744076, + 0.005549510475248098, + 0.00573776988312602, + 0.005609865300357342, + 0.005327445920556784, + 0.014722076244652271, + 0.011932400986552238, + 0.026829760521650314, + 0.011233675293624401, + 0.018474509939551353, + 0.012983908876776695, + 0.015081565827131271, + 0.014350834302604198, + 0.004633869975805283, + 0.003178203012794256, + 0.00422111339867115, + 0.0020274417474865913, + 0.004393768031150103, + 0.003673965809866786, + 0.004003860522061586, + 0.0056431712582707405, + 0.011846194043755531, + 0.009877625852823257, + 0.013031180948019028, + 0.008133646100759506, + 0.013065849430859089, + 0.014608628116548061, + 0.013323134742677212, + 0.017180433496832848, + 0.015266001224517822, + 0.007480393163859844, + 0.07937023043632507, + 0.007714318111538887, + 0.011659473180770874, + 0.12034271657466888, + 0.006563903298228979, + 0.024683302268385887, + 0.0077812206000089645, + 0.01310623250901699, + 0.008369609713554382, + 0.008266336284577847, + 0.012891286984086037, + 0.009968026541173458, + 0.007263726554811001, + 0.013580276630818844, + 0.003872558707371354, + 0.0033617219887673855, + 0.003846431151032448, + 0.0018823861610144377, + 0.01297685969620943, + 0.01391076110303402, + 0.001921687158755958, + 0.005072759930044413, + 0.023910479620099068, + 0.018729684874415398, + 0.022047005593776703, + 0.010979135520756245, + 0.018723534420132637, + 0.024129139259457588, + 0.018638668581843376, + 0.03977326676249504, + 0.005202528089284897, + 0.0035990646574646235, + 0.005811503157019615, + 0.0035976700019091368, + 0.004910668358206749, + 0.007308001164346933, + 0.003921664785593748, + 0.006307595409452915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.1604110598564148 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634863369166851 + ], + "opposite_logprob_tokens": [ + -3.65134859085083 + ], + "logprob_reward": -0.02634863369166851, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634863369166851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11311670392751694, + 0.05101533979177475, + 0.10018811374902725, + 0.1206490695476532, + 0.19390645623207092, + 0.1311248540878296, + 0.03126553073525429, + 0.07927745580673218, + 0.11241229623556137, + 0.06704419106245041 + ], + "credit_map": [ + 0.024773836135864258, + 0.008770129643380642, + 0.014204266481101513, + 0.011844862252473831, + 0.011736394837498665, + 0.015522640198469162, + 0.01167290098965168, + 0.014591671526432037, + 0.007797015365213156, + 0.005166532937437296, + 0.006310055032372475, + 0.005736414343118668, + 0.006050651893019676, + 0.007238269317895174, + 0.005807730834931135, + 0.0069086686708033085, + 0.00913377944380045, + 0.007544648367911577, + 0.019199445843696594, + 0.0088098905980587, + 0.018595566973090172, + 0.011290205642580986, + 0.01505464781075716, + 0.010559920221567154, + 0.01563769020140171, + 0.013899696059525013, + 0.023803049698472023, + 0.008417694829404354, + 0.014006059616804123, + 0.015279115177690983, + 0.008129332214593887, + 0.021476423367857933, + 0.03451674059033394, + 0.020383380353450775, + 0.022761564701795578, + 0.021577468141913414, + 0.020207688212394714, + 0.020856516435742378, + 0.024901149794459343, + 0.028701944276690483, + 0.008296960964798927, + 0.00530165433883667, + 0.020025968551635742, + 0.003068994963541627, + 0.006721179001033306, + 0.05572360381484032, + 0.004437038209289312, + 0.02754945494234562, + 0.0031419587321579456, + 0.004267537035048008, + 0.003936822060495615, + 0.0030947038903832436, + 0.0038863751105964184, + 0.004185320343822241, + 0.0027576209977269173, + 0.005995193962007761, + 0.007608299609273672, + 0.004391631111502647, + 0.008318950422108173, + 0.004437013529241085, + 0.010383362881839275, + 0.023359524086117744, + 0.004600252956151962, + 0.016178414225578308, + 0.01669865846633911, + 0.014741242863237858, + 0.012516122311353683, + 0.008372277021408081, + 0.013873251155018806, + 0.012834472581744194, + 0.013058638200163841, + 0.020317625254392624, + 0.007593098096549511, + 0.006472873501479626, + 0.007569781504571438, + 0.005316103342920542, + 0.009405226446688175, + 0.012166058644652367, + 0.008053993806242943, + 0.010467049665749073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.02634863369166851 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983719348907471 + ], + "opposite_logprob_tokens": [ + -0.5233719348907471 + ], + "logprob_reward": -0.8983719348907471, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983719348907471, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13303673267364502, + 0.07998453825712204, + 0.11803965270519257, + 0.04149116203188896, + 0.13180312514305115, + 0.3311288058757782, + 0.038765642791986465, + 0.04922322556376457, + 0.039276134222745895, + 0.03725101053714752 + ], + "credit_map": [ + 0.020450269803404808, + 0.012367635034024715, + 0.02365626022219658, + 0.01309780403971672, + 0.011688767932355404, + 0.020860465243458748, + 0.014262541197240353, + 0.016652986407279968, + 0.013158596120774746, + 0.007763829547911882, + 0.011119981296360493, + 0.010175940580666065, + 0.008175879716873169, + 0.009837714955210686, + 0.009663187898695469, + 0.010089395567774773, + 0.019911356270313263, + 0.014038609340786934, + 0.013866240158677101, + 0.008479793556034565, + 0.013374016620218754, + 0.016289500519633293, + 0.018086183816194534, + 0.013993932865560055, + 0.005422057583928108, + 0.0032410111743956804, + 0.008409333415329456, + 0.00272062374278903, + 0.005925733596086502, + 0.0054092626087367535, + 0.0027434539515525103, + 0.007619683630764484, + 0.011514231562614441, + 0.013209523633122444, + 0.018540991470217705, + 0.01642012968659401, + 0.012600881978869438, + 0.016842463985085487, + 0.021657390519976616, + 0.021017488092184067, + 0.019511042162775993, + 0.008452006615698338, + 0.09014852344989777, + 0.008309509605169296, + 0.013087578117847443, + 0.14835508167743683, + 0.006962867919355631, + 0.03630217909812927, + 0.003948131110519171, + 0.003797709010541439, + 0.004429370164871216, + 0.002935250522568822, + 0.00843694806098938, + 0.005926339887082577, + 0.003140337998047471, + 0.006151554640382528, + 0.004935048520565033, + 0.003030624007806182, + 0.00519641675055027, + 0.0028533274307847023, + 0.015407027676701546, + 0.005995044484734535, + 0.0029608148615807295, + 0.008844917640089989, + 0.005236343946307898, + 0.0034176926128566265, + 0.005169885698705912, + 0.002291459124535322, + 0.004038048442453146, + 0.003946544136852026, + 0.004233948886394501, + 0.010942205786705017, + 0.005102433264255524, + 0.0030033281072974205, + 0.0046918378211557865, + 0.0021863714791834354, + 0.005404013674706221, + 0.007881104946136475, + 0.003359805326908827, + 0.005622114986181259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.8983719348907471 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "handheld, slightly high-angle close-up, subtly trembling, intimately tracks the dynamic movement of a smooth paintbrush across paper. The pen, featuring a sleek black plastic tip filled with deep black ink, gracefully meets a sheet of thick, cream-colored, textured paper, held firmly in place. With fluid, confident swirls, the pen glides across the surface, its bristles flexing under deliberate pressure, creating single, bold, and expressive strokes that vary in thickness from delicate hairlines to powerful, broad lines. The damp, inky black ink momentarily shimmers before settling into the paper fibers. The entire scene is rendered in a classic '80s anime style, defined by rich cel shading that casts hard-edged shadows and a hint of film grain across the barrel, lending the image a nostalgic quality and evoking an atmosphere of intimate and focused artistic creation.", + "seed": 9657, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_1611_seed_9657_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.507035881280899, + "vlm_logprob": -0.507035881280899, + "vlm_yes_no_margin": 1.675, + "vlm_token_credit": -0.507035881280899, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04141584411263466 + ], + "opposite_logprob_tokens": [ + -3.7914159297943115 + ], + "logprob_reward": -0.04141584411263466, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04141584411263466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11100029200315475, + 0.14864055812358856, + 0.0839204490184784, + 0.0425078310072422, + 0.1533915251493454, + 0.25441989302635193, + 0.03871066868305206, + 0.025321438908576965, + 0.03501346334815025, + 0.10707393288612366 + ], + "credit_map": [ + 0.020970379933714867, + 0.011527477763593197, + 0.011187003925442696, + 0.014529450796544552, + 0.009599057026207447, + 0.011549487709999084, + 0.016429059207439423, + 0.015208359807729721, + 0.020519627258181572, + 0.020117176696658134, + 0.012454461306333542, + 0.015488028526306152, + 0.0120344627648592, + 0.01855584606528282, + 0.020479217171669006, + 0.02899172157049179, + 0.01349928230047226, + 0.00692611513659358, + 0.017465272918343544, + 0.015892883762717247, + 0.00595855712890625, + 0.007694676984101534, + 0.008165571838617325, + 0.008318081498146057, + 0.007552920840680599, + 0.003469105577096343, + 0.004228442907333374, + 0.0033875340595841408, + 0.0035707575734704733, + 0.007196288090199232, + 0.0047826459631323814, + 0.008320128545165062, + 0.02640947513282299, + 0.011025757528841496, + 0.016860881820321083, + 0.010531178675591946, + 0.010537693277001381, + 0.02825983427464962, + 0.015668321400880814, + 0.03409836068749428, + 0.038746368139982224, + 0.024308184161782265, + 0.03550238907337189, + 0.026052284985780716, + 0.023005004972219467, + 0.02553465962409973, + 0.02644832246005535, + 0.05482269078493118, + 0.006179484073072672, + 0.004382237792015076, + 0.004251484293490648, + 0.003390921512618661, + 0.0034773945808410645, + 0.0050807218067348, + 0.003608384635299444, + 0.008340034633874893, + 0.004279551096260548, + 0.0017852988094091415, + 0.005412477534264326, + 0.0016982483211904764, + 0.0019245595904067159, + 0.003401577239856124, + 0.0025935883168131113, + 0.004226133693009615, + 0.0053243073634803295, + 0.004348853137344122, + 0.0035178866237401962, + 0.005219753831624985, + 0.003605191595852375, + 0.003785734763368964, + 0.004494861699640751, + 0.0047168731689453125, + 0.016877803951501846, + 0.013163641095161438, + 0.013711425475776196, + 0.009929780848324299, + 0.012741999700665474, + 0.012346673756837845, + 0.013797303661704063, + 0.0145052969455719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.04141584411263466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22564344108104706 + ], + "opposite_logprob_tokens": [ + -1.6006433963775635 + ], + "logprob_reward": -0.22564344108104706, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22564344108104706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08105345070362091, + 0.17471368610858917, + 0.04227493703365326, + 0.08512367308139801, + 0.1938435286283493, + 0.20053088665008545, + 0.07892546057701111, + 0.04730896279215813, + 0.03497067466378212, + 0.06125471368432045 + ], + "credit_map": [ + 0.014562085270881653, + 0.007208758499473333, + 0.007222411688417196, + 0.007970910519361496, + 0.007572106551378965, + 0.009159781970083714, + 0.01331984531134367, + 0.01403755508363247, + 0.02482566237449646, + 0.02310202270746231, + 0.013095922768115997, + 0.016458990052342415, + 0.015567077323794365, + 0.030083067715168, + 0.02268475480377674, + 0.028896188363432884, + 0.006804564502090216, + 0.003643230302259326, + 0.008651359006762505, + 0.007591540925204754, + 0.0025192690081894398, + 0.00416459608823061, + 0.0035674090031534433, + 0.005332966800779104, + 0.00936132576316595, + 0.007797703146934509, + 0.0072863404639065266, + 0.009608463384211063, + 0.008143482729792595, + 0.013517430052161217, + 0.011339048855006695, + 0.018069878220558167, + 0.058378927409648895, + 0.00785973109304905, + 0.012474775314331055, + 0.00684077525511384, + 0.008083603344857693, + 0.05862964317202568, + 0.0124572878703475, + 0.029118791222572327, + 0.028996342793107033, + 0.016504231840372086, + 0.050324734300374985, + 0.019623449072241783, + 0.014430146664381027, + 0.017900684848427773, + 0.01620396599173546, + 0.036547333002090454, + 0.013478335924446583, + 0.005898801144212484, + 0.008844126015901566, + 0.004314460325986147, + 0.006403267849236727, + 0.011091327294707298, + 0.006783665157854557, + 0.022111477330327034, + 0.006879036780446768, + 0.004529679659754038, + 0.007700567599385977, + 0.004217389039695263, + 0.005431154277175665, + 0.0071523538790643215, + 0.005025103222578764, + 0.006373675540089607, + 0.005709456745535135, + 0.003865121863782406, + 0.0038807913661003113, + 0.0037645737174898386, + 0.003916568122804165, + 0.003014791989699006, + 0.0046238973736763, + 0.006195475347340107, + 0.009598695673048496, + 0.006252687890082598, + 0.00829209852963686, + 0.005655227694660425, + 0.0084034763276577, + 0.005398939363658428, + 0.007690317463129759, + 0.009963269345462322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.22564344108104706 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790247991681099 + ], + "opposite_logprob_tokens": [ + -2.5790247917175293 + ], + "logprob_reward": -0.0790247991681099, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790247991681099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10232587903738022, + 0.14640285074710846, + 0.06209127604961395, + 0.030805112794041634, + 0.26781877875328064, + 0.12775474786758423, + 0.055518146604299545, + 0.11845999211072922, + 0.05874580517411232, + 0.03007739968597889 + ], + "credit_map": [ + 0.01604057103395462, + 0.009913680143654346, + 0.008288318291306496, + 0.012090845964848995, + 0.008919413201510906, + 0.014733455143868923, + 0.017795555293560028, + 0.01454403717070818, + 0.01811334304511547, + 0.02309274859726429, + 0.012783646583557129, + 0.01801193505525589, + 0.015561243519186974, + 0.016243072226643562, + 0.020522495731711388, + 0.022074371576309204, + 0.010114006698131561, + 0.0038410662673413754, + 0.009535730816423893, + 0.007673123385757208, + 0.005424708593636751, + 0.008287651464343071, + 0.009461401961743832, + 0.0077535901218652725, + 0.00468906806781888, + 0.0026641099248081446, + 0.00327126681804657, + 0.0030402184929698706, + 0.004794477950781584, + 0.004768904764205217, + 0.0035825499799102545, + 0.003994514234364033, + 0.055273592472076416, + 0.019327541813254356, + 0.02496279776096344, + 0.016746005043387413, + 0.02600705996155739, + 0.0561952106654644, + 0.027014976367354393, + 0.04229160025715828, + 0.018635569140315056, + 0.01079903356730938, + 0.0217348113656044, + 0.008227569982409477, + 0.00967380590736866, + 0.010935272090137005, + 0.014152525924146175, + 0.03359614685177803, + 0.00908056739717722, + 0.006394767668098211, + 0.005405786447227001, + 0.005054873879998922, + 0.004958444274961948, + 0.008696426637470722, + 0.00402984069660306, + 0.011897438205778599, + 0.018705612048506737, + 0.009847387671470642, + 0.019788241013884544, + 0.011989193968474865, + 0.015131473541259766, + 0.014288983307778835, + 0.013156944885849953, + 0.015552155673503876, + 0.011147994548082352, + 0.006697202567011118, + 0.006402686238288879, + 0.007706309203058481, + 0.007363683078438044, + 0.005983022972941399, + 0.0060268063098192215, + 0.00741810305044055, + 0.004877734929323196, + 0.003517411183565855, + 0.0035961519461125135, + 0.0037161759100854397, + 0.0038267623167485, + 0.002391793765127659, + 0.003701611654832959, + 0.004449759144335985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.0790247991681099 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062056127935647964 + ], + "opposite_logprob_tokens": [ + -2.812056064605713 + ], + "logprob_reward": -0.062056127935647964, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062056127935647964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11141843348741531, + 0.26712456345558167, + 0.026117080822587013, + 0.08021241426467896, + 0.08991852402687073, + 0.044723521918058395, + 0.29324954748153687, + 0.03129906579852104, + 0.024660710245370865, + 0.03127610310912132 + ], + "credit_map": [ + 0.018434949219226837, + 0.011355804279446602, + 0.010601547546684742, + 0.012388529255986214, + 0.010029078461229801, + 0.014699974097311497, + 0.018912194296717644, + 0.014996358193457127, + 0.034727830439805984, + 0.03585892915725708, + 0.020663417875766754, + 0.02974313125014305, + 0.02179483138024807, + 0.039667386561632156, + 0.03795322775840759, + 0.046715814620256424, + 0.003802510444074869, + 0.001810713903978467, + 0.0028523574583232403, + 0.0026466045528650284, + 0.0019583478569984436, + 0.004109847825020552, + 0.0032547193113714457, + 0.005681976210325956, + 0.013147883117198944, + 0.0069274017587304115, + 0.007073584944009781, + 0.007703250274062157, + 0.008207336068153381, + 0.01375077199190855, + 0.00761133199557662, + 0.01579085737466812, + 0.014134354889392853, + 0.00518042640760541, + 0.010806928388774395, + 0.004874334204941988, + 0.004863818641752005, + 0.01795368827879429, + 0.009358123876154423, + 0.022746847942471504, + 0.006450459361076355, + 0.005608749110251665, + 0.006671482231467962, + 0.005136285908520222, + 0.004138692747801542, + 0.005066821351647377, + 0.004008844029158354, + 0.007642185315489769, + 0.048411909490823746, + 0.023702925071120262, + 0.034911200404167175, + 0.01601741462945938, + 0.02514406107366085, + 0.04721692577004433, + 0.024319300428032875, + 0.0735258162021637, + 0.004839759785681963, + 0.002714536851271987, + 0.004715601913630962, + 0.0032824587542563677, + 0.003901241347193718, + 0.003854648908600211, + 0.0035494640469551086, + 0.004441356286406517, + 0.004528391640633345, + 0.002842486836016178, + 0.002626683097332716, + 0.00295250047929585, + 0.003028439823538065, + 0.0027760399971157312, + 0.002692361129447818, + 0.0032138077076524496, + 0.005257474724203348, + 0.003946318756788969, + 0.004093462135642767, + 0.0036891577765345573, + 0.003234707284718752, + 0.003257456934079528, + 0.003604718018323183, + 0.004192806314677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.062056127935647964 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.1270391941070557 + ], + "opposite_logprob_tokens": [ + -0.12703923881053925 + ], + "logprob_reward": -2.1270391941070557, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.1270391941070557, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17334915697574615, + 0.06181900575757027, + 0.038920946419239044, + 0.10231883078813553, + 0.1885426640510559, + 0.2468753308057785, + 0.04393636807799339, + 0.0571846142411232, + 0.0205888319760561, + 0.06646419316530228 + ], + "credit_map": [ + 0.026073751971125603, + 0.018242523074150085, + 0.015283489599823952, + 0.019213512539863586, + 0.015061327256262302, + 0.02694752812385559, + 0.026683153584599495, + 0.025843895971775055, + 0.007853027433156967, + 0.006629451643675566, + 0.006926076486706734, + 0.009401190094649792, + 0.004694747272878885, + 0.009203935042023659, + 0.00836906861513853, + 0.008741510100662708, + 0.006979487370699644, + 0.0037607683334499598, + 0.00587255135178566, + 0.004805988632142544, + 0.0029091094620525837, + 0.004485838580876589, + 0.0040350500494241714, + 0.006072159390896559, + 0.011284006759524345, + 0.006233806721866131, + 0.009699778631329536, + 0.0072829523123800755, + 0.009854850359261036, + 0.0167525801807642, + 0.01797615736722946, + 0.023234719410538673, + 0.06472300738096237, + 0.006971408613026142, + 0.011994380503892899, + 0.006231503561139107, + 0.008284609764814377, + 0.07069557905197144, + 0.009569997899234295, + 0.01007218286395073, + 0.03721990808844566, + 0.02374281920492649, + 0.041173409670591354, + 0.025676865130662918, + 0.022422153502702713, + 0.027279697358608246, + 0.02361421287059784, + 0.04574628546833992, + 0.006435831543058157, + 0.004793877247720957, + 0.00475592166185379, + 0.003909387160092592, + 0.00514213927090168, + 0.006721123121678829, + 0.0032215802930295467, + 0.008956514298915863, + 0.009142177179455757, + 0.005106422584503889, + 0.008677995763719082, + 0.004624000750482082, + 0.006687427405267954, + 0.006482829339802265, + 0.006538752466440201, + 0.009925012476742268, + 0.002990742214024067, + 0.002491164021193981, + 0.0027880012057721615, + 0.0021986509673297405, + 0.0018884858582168818, + 0.002575002145022154, + 0.0024787774309515953, + 0.003178010694682598, + 0.009353764355182648, + 0.008995255455374718, + 0.009613176807761192, + 0.008047081530094147, + 0.007399815134704113, + 0.008379618637263775, + 0.007326715160161257, + 0.007348775863647461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -2.1270391941070557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "handheld, slightly high-angle close-up, subtly trembling, intimately tracks the dynamic movement of a smooth paintbrush across paper. The pen, featuring a sleek black plastic tip filled with deep black ink, gracefully meets a sheet of thick, cream-colored, textured paper, held firmly in place. With fluid, confident swirls, the pen glides across the surface, its bristles flexing under deliberate pressure, creating single, bold, and expressive strokes that vary in thickness from delicate hairlines to powerful, broad lines. The damp, inky black ink momentarily shimmers before settling into the paper fibers. The entire scene is rendered in a classic '80s anime style, defined by rich cel shading that casts hard-edged shadows and a hint of film grain across the barrel, lending the image a nostalgic quality and evoking an atmosphere of intimate and focused artistic creation.", + "seed": 9658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_1611_seed_9658_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.33464580625295637, + "vlm_logprob": -0.33464580625295637, + "vlm_yes_no_margin": 1.6749999880790711, + "vlm_token_credit": -0.33464580625295637, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0989278256893158 + ], + "opposite_logprob_tokens": [ + -2.7239277362823486 + ], + "logprob_reward": -0.0989278256893158, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0989278256893158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10291441529989243, + 0.08227498084306717, + 0.0978274792432785, + 0.12163110077381134, + 0.03470231592655182, + 0.0586216039955616, + 0.05790585279464722, + 0.08056937158107758, + 0.25657492876052856, + 0.10697789490222931 + ], + "credit_map": [ + 0.01962914690375328, + 0.017891718074679375, + 0.01625208929181099, + 0.010142529383301735, + 0.004399497993290424, + 0.0061351824551820755, + 0.00929334294050932, + 0.019170906394720078, + 0.011555476114153862, + 0.011053061112761497, + 0.010600434616208076, + 0.005668530706316233, + 0.006630739197134972, + 0.018164856359362602, + 0.006987749598920345, + 0.011614132672548294, + 0.011365324258804321, + 0.008201220072805882, + 0.00932142324745655, + 0.007245911750942469, + 0.015204832889139652, + 0.015297883190214634, + 0.009686622768640518, + 0.02150425873696804, + 0.02088897116482258, + 0.009958596900105476, + 0.009214509278535843, + 0.01339760608971119, + 0.016263727098703384, + 0.021240008994936943, + 0.01322244293987751, + 0.01744524948298931, + 0.005430619698017836, + 0.0037361797876656055, + 0.003042964730411768, + 0.003513423725962639, + 0.003201274899765849, + 0.004723094869405031, + 0.004385646898299456, + 0.006669111084192991, + 0.008273432031273842, + 0.006293746642768383, + 0.006060827523469925, + 0.005367110949009657, + 0.006128696724772453, + 0.008448922075331211, + 0.006815330125391483, + 0.011233535595238209, + 0.0067391046322882175, + 0.0030421307310462, + 0.0065297470428049564, + 0.008129309862852097, + 0.004179217852652073, + 0.008491483516991138, + 0.005894920323044062, + 0.01489993929862976, + 0.015447244048118591, + 0.012815504334867, + 0.017523879185318947, + 0.008396345190703869, + 0.0052490695379674435, + 0.0061108083464205265, + 0.006559406872838736, + 0.008467116393148899, + 0.04378783330321312, + 0.042242761701345444, + 0.033607516437768936, + 0.03182089328765869, + 0.01945582590997219, + 0.026835426688194275, + 0.02010994218289852, + 0.03871474787592888, + 0.0111238369718194, + 0.010865490883588791, + 0.01142637524753809, + 0.01413857750594616, + 0.01866428554058075, + 0.013613518327474594, + 0.014112149365246296, + 0.013033662922680378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.0989278256893158 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161642134189606 + ], + "opposite_logprob_tokens": [ + -1.7016164064407349 + ], + "logprob_reward": -0.20161642134189606, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161642134189606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.212345153093338, + 0.08408396691083908, + 0.20570555329322815, + 0.03338915482163429, + 0.06040431931614876, + 0.04954967275261879, + 0.11014009267091751, + 0.05020293593406677, + 0.14335046708583832, + 0.05082867294549942 + ], + "credit_map": [ + 0.03918161988258362, + 0.035775378346443176, + 0.02798108384013176, + 0.014001229777932167, + 0.0090898172929883, + 0.01811574399471283, + 0.019372757524251938, + 0.04882752522826195, + 0.013855797238647938, + 0.008446578867733479, + 0.011135161854326725, + 0.0064965360797941685, + 0.006076194811612368, + 0.0136114452034235, + 0.009128890931606293, + 0.015333360061049461, + 0.024171747267246246, + 0.01951385661959648, + 0.018851257860660553, + 0.01643637754023075, + 0.024667147547006607, + 0.03186541050672531, + 0.021387172862887383, + 0.048812586814165115, + 0.005390371661633253, + 0.0026332028210163116, + 0.004410926252603531, + 0.0029556946828961372, + 0.003700822126120329, + 0.0050566112622618675, + 0.003360205562785268, + 0.005881322082132101, + 0.010187954641878605, + 0.0060273450799286366, + 0.007074735127389431, + 0.004721523728221655, + 0.00414615124464035, + 0.006925549823790789, + 0.007906446233391762, + 0.013414614833891392, + 0.00956553965806961, + 0.004004351329058409, + 0.00386316841468215, + 0.003123451955616474, + 0.0038741584867239, + 0.00770516786724329, + 0.005286097526550293, + 0.012127736583352089, + 0.012981933541595936, + 0.006534733343869448, + 0.0117149967700243, + 0.014376985840499401, + 0.009821387939155102, + 0.018632784485816956, + 0.012080630287528038, + 0.023996638134121895, + 0.008989905007183552, + 0.0062895105220377445, + 0.009219679981470108, + 0.005761777050793171, + 0.002288456540554762, + 0.005119671113789082, + 0.004930412396788597, + 0.007603525184094906, + 0.024194393306970596, + 0.023504560813307762, + 0.017055464908480644, + 0.014675487764179707, + 0.012442458420991898, + 0.017683643847703934, + 0.014415664598345757, + 0.01937880739569664, + 0.005705500487238169, + 0.004617620259523392, + 0.0047217123210430145, + 0.004829462617635727, + 0.007833048701286316, + 0.007052486762404442, + 0.005665312521159649, + 0.010403530672192574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.20161642134189606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005671411752701 + ], + "opposite_logprob_tokens": [ + -2.695056676864624 + ], + "logprob_reward": -0.07005671411752701, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005671411752701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0901254341006279, + 0.14018192887306213, + 0.07740096002817154, + 0.0392451211810112, + 0.05599342659115791, + 0.0658935010433197, + 0.15284273028373718, + 0.08967450261116028, + 0.17845682799816132, + 0.11018556356430054 + ], + "credit_map": [ + 0.013921573758125305, + 0.015592982992529869, + 0.010552906431257725, + 0.006639237515628338, + 0.005590932443737984, + 0.008155517280101776, + 0.010619904845952988, + 0.019052380695939064, + 0.02149520255625248, + 0.01837964728474617, + 0.017026452347636223, + 0.01146551501005888, + 0.009116275236010551, + 0.018593646585941315, + 0.015425538644194603, + 0.028679650276899338, + 0.010362779721617699, + 0.005287390202283859, + 0.009518617764115334, + 0.007329918909817934, + 0.009892220608890057, + 0.010871919803321362, + 0.008392415009438992, + 0.015745697543025017, + 0.0057855877093970776, + 0.0029076070059090853, + 0.002837288659065962, + 0.003461836138740182, + 0.0033130983356386423, + 0.007379889488220215, + 0.0048134163953363895, + 0.008746398612856865, + 0.014415901154279709, + 0.005319484043866396, + 0.0046925717033445835, + 0.005134917329996824, + 0.00436039175838232, + 0.006773087196052074, + 0.005162031855434179, + 0.010135041549801826, + 0.010233715176582336, + 0.005940836388617754, + 0.005868874955922365, + 0.004604128655046225, + 0.006660329643636942, + 0.00958393793553114, + 0.0068317679688334465, + 0.01616991125047207, + 0.01924728974699974, + 0.007417153101414442, + 0.020675450563430786, + 0.020423468202352524, + 0.009445461444556713, + 0.023559080436825752, + 0.01832265593111515, + 0.03375217691063881, + 0.015366307459771633, + 0.013275053352117538, + 0.0152925755828619, + 0.009856615215539932, + 0.007569050416350365, + 0.008933471515774727, + 0.007583617232739925, + 0.011797807179391384, + 0.03718794509768486, + 0.03503067418932915, + 0.020789580419659615, + 0.021811451762914658, + 0.011352808214724064, + 0.016707485541701317, + 0.01279342733323574, + 0.02278345637023449, + 0.013710864819586277, + 0.009333875961601734, + 0.010654148645699024, + 0.011831832118332386, + 0.01603643223643303, + 0.014642694965004921, + 0.011802525259554386, + 0.022173188626766205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.07005671411752701 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899120450019836 + ], + "opposite_logprob_tokens": [ + -2.578991174697876 + ], + "logprob_reward": -0.07899120450019836, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899120450019836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056148119270801544, + 0.02348916232585907, + 0.09432626515626907, + 0.03444589674472809, + 0.026490172371268272, + 0.29146048426628113, + 0.05904087424278259, + 0.04837484657764435, + 0.15224693715572357, + 0.21397726237773895 + ], + "credit_map": [ + 0.010180599056184292, + 0.00916287861764431, + 0.009737652726471424, + 0.003803085768595338, + 0.0023794793523848057, + 0.0044739581644535065, + 0.006555503234267235, + 0.00985496211796999, + 0.003642178373411298, + 0.0023445403203368187, + 0.00347731146030128, + 0.0019074021838605404, + 0.0015661180950701237, + 0.0031684760469943285, + 0.0030647467356175184, + 0.004318387247622013, + 0.009862501174211502, + 0.0096375597640872, + 0.008940565399825573, + 0.010181997902691364, + 0.009926646947860718, + 0.014065698720514774, + 0.010220355354249477, + 0.021490948274731636, + 0.006517880596220493, + 0.0029495088383555412, + 0.003958086483180523, + 0.002996588358655572, + 0.003336207242682576, + 0.005068507045507431, + 0.004276196006685495, + 0.005342923104763031, + 0.005078637041151524, + 0.0033183565828949213, + 0.002302565611898899, + 0.0026192767545580864, + 0.0020980951376259327, + 0.0027542978059500456, + 0.0026012337766587734, + 0.005717710591852665, + 0.04457709565758705, + 0.024371089413762093, + 0.027448732405900955, + 0.02102523110806942, + 0.020880253985524178, + 0.040706515312194824, + 0.032173968851566315, + 0.08027758449316025, + 0.008732999674975872, + 0.003665631404146552, + 0.006760654039680958, + 0.006625101435929537, + 0.0060959202237427235, + 0.010069021955132484, + 0.007368316873908043, + 0.00972322840243578, + 0.00878907646983862, + 0.007155164144933224, + 0.00883102510124445, + 0.006064033601433039, + 0.003228995017707348, + 0.004420348908752203, + 0.004383035935461521, + 0.005503168795257807, + 0.029976723715662956, + 0.03013010323047638, + 0.01840663142502308, + 0.018244603648781776, + 0.00977017730474472, + 0.012254836969077587, + 0.011133784428238869, + 0.022330068051815033, + 0.02805108018219471, + 0.015371603891253471, + 0.019192686304450035, + 0.018410973250865936, + 0.02711309865117073, + 0.032333556562662125, + 0.02037297748029232, + 0.053131282329559326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.07899120450019836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236368656158447 + ], + "opposite_logprob_tokens": [ + -0.34863680601119995 + ], + "logprob_reward": -1.2236368656158447, + "yes_no_margin_reward": -0.8750000596046448, + "token_credit_reward": -1.2236368656158447, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1501247137784958, + 0.07820815593004227, + 0.09248406440019608, + 0.07270076870918274, + 0.06110744923353195, + 0.07745656371116638, + 0.13069652020931244, + 0.10835083574056625, + 0.13934984803199768, + 0.08952099829912186 + ], + "credit_map": [ + 0.02855352871119976, + 0.026494942605495453, + 0.01967485435307026, + 0.007648157421499491, + 0.006012474652379751, + 0.01154918409883976, + 0.009032501839101315, + 0.041159071028232574, + 0.01208103634417057, + 0.010431100614368916, + 0.011365332640707493, + 0.005317650735378265, + 0.006743330974131823, + 0.014121686108410358, + 0.008406062610447407, + 0.009741955436766148, + 0.012585734017193317, + 0.007695428561419249, + 0.011697961017489433, + 0.008042803965508938, + 0.015219983644783497, + 0.013664963655173779, + 0.008533719927072525, + 0.015043474733829498, + 0.012546797282993793, + 0.006561968009918928, + 0.00891673844307661, + 0.006006590090692043, + 0.006581026595085859, + 0.010744619183242321, + 0.009211638011038303, + 0.01213139295578003, + 0.01052987389266491, + 0.007339706178754568, + 0.007274573668837547, + 0.00494384765625, + 0.004911030177026987, + 0.007282319478690624, + 0.007608951535075903, + 0.011217148043215275, + 0.012380837462842464, + 0.006417381577193737, + 0.00883063580840826, + 0.007644422817975283, + 0.004998546093702316, + 0.011371448636054993, + 0.008493523113429546, + 0.01731977052986622, + 0.014445303939282894, + 0.006578829605132341, + 0.014381677843630314, + 0.016096357256174088, + 0.010967344976961613, + 0.02294774353504181, + 0.01471665222197771, + 0.030562611296772957, + 0.019562412053346634, + 0.015288055874407291, + 0.022465968504548073, + 0.010248017497360706, + 0.007080729119479656, + 0.00873942393809557, + 0.010001355782151222, + 0.014964877627789974, + 0.025393668562173843, + 0.025843853130936623, + 0.016742143779993057, + 0.0157620869576931, + 0.010213281959295273, + 0.01602749526500702, + 0.010815232992172241, + 0.018552079796791077, + 0.00911689456552267, + 0.007417185232043266, + 0.00935399066656828, + 0.013224370777606964, + 0.013072562403976917, + 0.010450063273310661, + 0.010292893275618553, + 0.016593042761087418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -1.2236368656158447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "handheld, slightly high-angle close-up, subtly trembling, intimately tracks the dynamic movement of a smooth paintbrush across paper. The pen, featuring a sleek black plastic tip filled with deep black ink, gracefully meets a sheet of thick, cream-colored, textured paper, held firmly in place. With fluid, confident swirls, the pen glides across the surface, its bristles flexing under deliberate pressure, creating single, bold, and expressive strokes that vary in thickness from delicate hairlines to powerful, broad lines. The damp, inky black ink momentarily shimmers before settling into the paper fibers. The entire scene is rendered in a classic '80s anime style, defined by rich cel shading that casts hard-edged shadows and a hint of film grain across the barrel, lending the image a nostalgic quality and evoking an atmosphere of intimate and focused artistic creation.", + "seed": 9659, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_10/index_1611_seed_9659_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4909399226307869, + "vlm_logprob": -0.4909399226307869, + "vlm_yes_no_margin": 1.4, + "vlm_token_credit": -0.4909399226307869, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07445720583200455 + ], + "opposite_logprob_tokens": [ + -3.0744571685791016 + ], + "logprob_reward": -0.07445720583200455, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.07445720583200455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15128564834594727, + 0.11409121006727219, + 0.04038145765662193, + 0.11092408001422882, + 0.1055033802986145, + 0.03795124590396881, + 0.12754230201244354, + 0.18306764960289001, + 0.10500296950340271, + 0.024250049144029617 + ], + "credit_map": [ + 0.017793651670217514, + 0.018472837284207344, + 0.025351079180836678, + 0.01332239992916584, + 0.01982955075800419, + 0.024646226316690445, + 0.015558074228465557, + 0.016311826184391975, + 0.018664948642253876, + 0.012155533768236637, + 0.011995743960142136, + 0.010755750350654125, + 0.012599291279911995, + 0.018044058233499527, + 0.007977106608450413, + 0.02189878188073635, + 0.005338332150131464, + 0.006145828869193792, + 0.003351068589836359, + 0.0025912912096828222, + 0.00614173524081707, + 0.005740813910961151, + 0.0038315465208142996, + 0.007240841165184975, + 0.019161326810717583, + 0.018537282943725586, + 0.0068671382032334805, + 0.011245244182646275, + 0.010727142915129662, + 0.0107883857563138, + 0.016555311158299446, + 0.017042251303792, + 0.00912214070558548, + 0.038598328828811646, + 0.0026568181347101927, + 0.002223279559984803, + 0.004828684497624636, + 0.007640344090759754, + 0.0021094800904393196, + 0.03832430765032768, + 0.007879874669015408, + 0.009256874211132526, + 0.003702440531924367, + 0.0029090968891978264, + 0.002913727192208171, + 0.003400526475161314, + 0.002337868558242917, + 0.0055508362129330635, + 0.029096586629748344, + 0.011006967164576054, + 0.007966497913002968, + 0.007579166907817125, + 0.010593553073704243, + 0.010603954084217548, + 0.009810629300773144, + 0.04088494926691055, + 0.02542339451611042, + 0.022657698020339012, + 0.0180892925709486, + 0.020006638020277023, + 0.028836706653237343, + 0.027662530541419983, + 0.021518167108297348, + 0.018873224034905434, + 0.017318157479166985, + 0.013838669285178185, + 0.010340402834117413, + 0.008910050615668297, + 0.01804952695965767, + 0.01455752644687891, + 0.009241649881005287, + 0.012746992520987988, + 0.0038019386120140553, + 0.0026545305736362934, + 0.0019427555380389094, + 0.001715477672405541, + 0.0032260091975331306, + 0.004759674891829491, + 0.002756278496235609, + 0.003393383463844657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.07445720583200455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233854651451111 + ], + "opposite_logprob_tokens": [ + -0.8983854651451111 + ], + "logprob_reward": -0.5233854651451111, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233854651451111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25835010409355164, + 0.12043363600969315, + 0.03804580494761467, + 0.06665512174367905, + 0.10931307822465897, + 0.060205619782209396, + 0.08643897622823715, + 0.13757139444351196, + 0.09522894769906998, + 0.02775725908577442 + ], + "credit_map": [ + 0.023696716874837875, + 0.014913169667124748, + 0.019708510488271713, + 0.017454320564866066, + 0.018742989748716354, + 0.04284132644534111, + 0.048729196190834045, + 0.07226389646530151, + 0.015943795442581177, + 0.011987550184130669, + 0.013737346045672894, + 0.0106740053743124, + 0.010236319154500961, + 0.0185368824750185, + 0.00973844900727272, + 0.029579294845461845, + 0.005024795886129141, + 0.004428438376635313, + 0.0035662802401930094, + 0.0019415855640545487, + 0.006364746950566769, + 0.006074210628867149, + 0.0037122643552720547, + 0.0069334851577878, + 0.01165995467454195, + 0.011404241435229778, + 0.004297139588743448, + 0.0061492170207202435, + 0.009500699117779732, + 0.008223667740821838, + 0.00784231536090374, + 0.007577901240438223, + 0.00709014805033803, + 0.04433257505297661, + 0.0028445429634302855, + 0.0019343795720487833, + 0.0030463009607046843, + 0.0038065968547016382, + 0.001523591810837388, + 0.04473495855927467, + 0.009874831885099411, + 0.00981845147907734, + 0.005926360841840506, + 0.00385111547075212, + 0.006568034179508686, + 0.009769846685230732, + 0.005516419652849436, + 0.008880560286343098, + 0.016493359580636024, + 0.0084830978885293, + 0.00632892269641161, + 0.005903941113501787, + 0.012055451981723309, + 0.006838781759142876, + 0.005278892815113068, + 0.025056537240743637, + 0.027012016624212265, + 0.016502227634191513, + 0.010055833496153355, + 0.009507313370704651, + 0.010220697149634361, + 0.015813389793038368, + 0.012851616367697716, + 0.03560829535126686, + 0.01621830463409424, + 0.012092911638319492, + 0.008647087961435318, + 0.007213479373604059, + 0.0198195930570364, + 0.012304309755563736, + 0.008287747390568256, + 0.01064552552998066, + 0.00439565209671855, + 0.002247921656817198, + 0.0030882374849170446, + 0.001883722492493689, + 0.003573710797354579, + 0.005317325238138437, + 0.0030122397001832724, + 0.004238454159349203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.5233854651451111 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003437489271164 + ], + "opposite_logprob_tokens": [ + -2.350343704223633 + ], + "logprob_reward": -0.1003437489271164, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003437489271164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30788666009902954, + 0.09448931366205215, + 0.08101870864629745, + 0.07152221351861954, + 0.14560723304748535, + 0.024018114432692528, + 0.030763015151023865, + 0.03980115056037903, + 0.19740189611911774, + 0.007491692900657654 + ], + "credit_map": [ + 0.036276038736104965, + 0.039092760533094406, + 0.035724177956581116, + 0.025990718975663185, + 0.04043453559279442, + 0.04270495846867561, + 0.04155338183045387, + 0.046110108494758606, + 0.011922611854970455, + 0.016380436718463898, + 0.010922275483608246, + 0.010809969156980515, + 0.011796955950558186, + 0.01345974300056696, + 0.0071080694906413555, + 0.012089253403246403, + 0.012399671599268913, + 0.00916330236941576, + 0.00546222971752286, + 0.005460688844323158, + 0.017573527991771698, + 0.011453179642558098, + 0.007831837050616741, + 0.011674272827804089, + 0.021805258467793465, + 0.01938030868768692, + 0.0042146118357777596, + 0.00431029824540019, + 0.0046897148713469505, + 0.0059876576997339725, + 0.006054461933672428, + 0.005079900845885277, + 0.015623034909367561, + 0.06108976900577545, + 0.002727814484387636, + 0.0022979446221143007, + 0.003861156525090337, + 0.005235093180090189, + 0.0020656301639974117, + 0.05270678922533989, + 0.0042756227776408195, + 0.004656525794416666, + 0.0029821658972650766, + 0.001971860881894827, + 0.002330941380932927, + 0.0035364259965717793, + 0.0013020405313000083, + 0.002962531289085746, + 0.005469139199703932, + 0.0029292444232851267, + 0.002091397764161229, + 0.002739696530625224, + 0.0046807159669697285, + 0.0034322685096412897, + 0.0021652572322636843, + 0.007255294360220432, + 0.0053594824858009815, + 0.004765976686030626, + 0.0028254559729248285, + 0.0038872682489454746, + 0.005924738012254238, + 0.005871241446584463, + 0.005375937558710575, + 0.005791045725345612, + 0.030287377536296844, + 0.026524381712079048, + 0.0201468113809824, + 0.018887760117650032, + 0.023582875728607178, + 0.030141497030854225, + 0.017754508182406425, + 0.030076688155531883, + 0.001077120192348957, + 0.0007502283551730216, + 0.0007702481234446168, + 0.0005735703161917627, + 0.0011058232048526406, + 0.0013230693293735385, + 0.0007167893345467746, + 0.001174843986518681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.1003437489271164 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496366322040558 + ], + "opposite_logprob_tokens": [ + -2.9299635887145996 + ], + "logprob_reward": -0.05496366322040558, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496366322040558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08909288048744202, + 0.13167859613895416, + 0.026506327092647552, + 0.03671681508421898, + 0.5865127444267273, + 0.008825227618217468, + 0.016982071101665497, + 0.05201450735330582, + 0.03908034414052963, + 0.012590501457452774 + ], + "credit_map": [ + 0.013166802003979683, + 0.010320971719920635, + 0.01474751252681017, + 0.006379629950970411, + 0.008004426956176758, + 0.013572393916547298, + 0.010325942188501358, + 0.012575196102261543, + 0.02159983292222023, + 0.01653035543859005, + 0.014836736023426056, + 0.012684371322393417, + 0.01355860847979784, + 0.020366474986076355, + 0.009123955853283405, + 0.022978264838457108, + 0.0038215278182178736, + 0.003199326805770397, + 0.0023850142024457455, + 0.0013535298639908433, + 0.004140639677643776, + 0.004049659240990877, + 0.002869005547836423, + 0.004687622655183077, + 0.008168446831405163, + 0.008555492386221886, + 0.002570588607341051, + 0.003100569825619459, + 0.0030738417990505695, + 0.0037251212634146214, + 0.003899951232597232, + 0.003622798714786768, + 0.05697299912571907, + 0.25353726744651794, + 0.007774162571877241, + 0.006409839261323214, + 0.013087396509945393, + 0.014713709242641926, + 0.00622188625857234, + 0.22779546678066254, + 0.001641731709241867, + 0.00176808622200042, + 0.0006920128944329917, + 0.00047906546387821436, + 0.0009601137135177851, + 0.0014196736738085747, + 0.0005241294857114553, + 0.001340414397418499, + 0.0027134264819324017, + 0.0022157884668558836, + 0.0010870337719097733, + 0.0013174869818612933, + 0.0034040880855172873, + 0.001356259104795754, + 0.0009029608336277306, + 0.003985025454312563, + 0.010447115637362003, + 0.004906525369733572, + 0.00399384880438447, + 0.0030362759716808796, + 0.004795188084244728, + 0.0066928560845553875, + 0.004386849235743284, + 0.01375584490597248, + 0.005994162056595087, + 0.005291458684951067, + 0.003780212253332138, + 0.0037582763470709324, + 0.004832056816667318, + 0.0057999007403850555, + 0.0035722856409847736, + 0.006051989737898111, + 0.0018279929645359516, + 0.0016078794142231345, + 0.0018857276299968362, + 0.0010896603344008327, + 0.0015458359848707914, + 0.0018096620915457606, + 0.0013256677193567157, + 0.0014980752021074295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.05496366322040558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a handheld, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement subtly trembling while tracking the paintbrush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pen tip black plastic and filled with deep black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the paper thick, cream-colored, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a classic '80s anime style with rich cel shading and film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015495300292969 + ], + "opposite_logprob_tokens": [ + -0.2015494853258133 + ], + "logprob_reward": -1.7015495300292969, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015495300292969, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.28620073199272156, + 0.07668701559305191, + 0.09041763097047806, + 0.05899167060852051, + 0.07017703354358673, + 0.024720363318920135, + 0.03922049701213837, + 0.029015198349952698, + 0.31320828199386597, + 0.011361571960151196 + ], + "credit_map": [ + 0.02118431217968464, + 0.02339319698512554, + 0.03435198962688446, + 0.02216058224439621, + 0.026423053815960884, + 0.04344545677304268, + 0.047714993357658386, + 0.0675271674990654, + 0.010512890294194221, + 0.006659302394837141, + 0.008233897387981415, + 0.005688722711056471, + 0.005610674619674683, + 0.012779523618519306, + 0.007827096618711948, + 0.01937491074204445, + 0.010900493711233139, + 0.011301136575639248, + 0.007708271499723196, + 0.005617013666778803, + 0.017374863848090172, + 0.013214431703090668, + 0.008720005862414837, + 0.015581417828798294, + 0.012245315127074718, + 0.011518503539264202, + 0.003750000149011612, + 0.0035858568735420704, + 0.005473378114402294, + 0.00636761449277401, + 0.007779273670166731, + 0.008271729573607445, + 0.012040098197758198, + 0.024844015017151833, + 0.003088974393904209, + 0.002533531980589032, + 0.0035145131405442953, + 0.004849068820476532, + 0.0021607736125588417, + 0.01714606024324894, + 0.005029725376516581, + 0.0053224400617182255, + 0.0016446516383439302, + 0.0012267747661098838, + 0.0026479363441467285, + 0.003794497111812234, + 0.0010161479003727436, + 0.0040381900034844875, + 0.0072104549035429955, + 0.003336456371471286, + 0.0029121541883796453, + 0.0025537745095789433, + 0.006465574726462364, + 0.003760618157684803, + 0.0034397123381495476, + 0.009541750885546207, + 0.004295811522752047, + 0.004501985851675272, + 0.0027080904692411423, + 0.0025029140524566174, + 0.003114241873845458, + 0.0031560882925987244, + 0.0036820420064032078, + 0.005054025445133448, + 0.05317704379558563, + 0.04118098318576813, + 0.028375638648867607, + 0.02196279540657997, + 0.07041836529970169, + 0.038868702948093414, + 0.025721004232764244, + 0.03350372239947319, + 0.0016616653883829713, + 0.0011626852210611105, + 0.0013547559501603246, + 0.0008724775398150086, + 0.0014087812742218375, + 0.002508374396711588, + 0.0010648836614564061, + 0.0013279482955113053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -1.7015495300292969 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_11/index_tensor([4503, 4503])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_11/index_tensor([4503, 4503])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..86ec31553bb03275841e3561ea727b308c7ef9d4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_11/index_tensor([4503, 4503])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This cinematic close-up, with a remarkably shallow depth of field, captures a shimmering soap bubble, suspended almost motionless in the air of a quiet, sunlit room. The camera remains still for a moment, focused on the delicate iridescent film on the bubble's surface, where swirling, hypnotic patterns of mauve, cyan, and liquid gold catch the light. The distortions of the large windows, fisheye reflections, and the room's out-of-focus interior twist and glide across its nearly perfectly spherical shape. Then, the camera begins a steady upward tilt, following the bubble, gently propelled by an invisible column of hot air emanating from a vent in the black metal floor below. The bubble sways unsteadily, its shape momentarily distorting as it rises on the gentle current. Then, it slowly drifts out of the column of air, and the camera follows its graceful downward arc toward the polished, dark hardwood floor. Just before contact, it is captured again by the rising air, and the camera tilts again, tracking its final, brief ascent. Finally, the bubble drifts downwards for the last time; the camera tilts down to capture its descent, and upon contact with the floor, the bubble vanishes instantly, leaving only a tiny, shimmering, wet ring on the wood that quickly evaporates in the soft ambient light. The entire scene is bathed in the warm golden glow of the afternoon sun, illuminating the dancing dust particles and creating an atmosphere of tranquility, contemplation, and transience.", + "seed": 14772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_11/index_4503_seed_14772_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.031355947349220516, + "vlm_logprob": -0.031355947349220516, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.031355947349220516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007459642831236124 + ], + "opposite_logprob_tokens": [ + -6.13245964050293 + ], + "logprob_reward": -0.007459642831236124, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.007459642831236124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02383681759238243, + 0.033711470663547516, + 0.024327868595719337, + 0.03906739503145218, + 0.06067073345184326, + 0.02353256195783615, + 0.02768169716000557, + 0.03081895038485527, + 0.032029181718826294, + 0.704323410987854 + ], + "credit_map": [ + 0.002247412921860814, + 0.0021675170864909887, + 0.003403366543352604, + 0.0035518177319318056, + 0.002913861069828272, + 0.002772076753899455, + 0.0034989018458873034, + 0.003281859215348959, + 0.0025878124870359898, + 0.0017561963759362698, + 0.0026448138523846865, + 0.0027629800606518984, + 0.002144003054127097, + 0.010565578937530518, + 0.0024501339066773653, + 0.008799945935606956, + 0.0019095296738669276, + 0.002399065066128969, + 0.003063959302380681, + 0.0023694911506026983, + 0.002610960276797414, + 0.0052048987708985806, + 0.002458168426528573, + 0.004311789758503437, + 0.004918298684060574, + 0.0030933215748518705, + 0.006620144471526146, + 0.005444798152893782, + 0.0063553787767887115, + 0.004164989572018385, + 0.004659532569348812, + 0.0038109251763671637, + 0.006476927548646927, + 0.008017098531126976, + 0.007575920782983303, + 0.010270634666085243, + 0.005217472556978464, + 0.009234345518052578, + 0.009248416870832443, + 0.0046299053356051445, + 0.00281754438765347, + 0.0032492068130522966, + 0.003267016727477312, + 0.0036693166475743055, + 0.0037734138313680887, + 0.0026919005904346704, + 0.002000657143071294, + 0.0020635020919144154, + 0.0052853734232485294, + 0.003945319447666407, + 0.003583719488233328, + 0.0033791111782193184, + 0.0035779208410531282, + 0.002412219299003482, + 0.0022801323793828487, + 0.003217898542061448, + 0.0019682347774505615, + 0.0032269135117530823, + 0.0031187974382191896, + 0.003475950798019767, + 0.0065952870063483715, + 0.0027473748195916414, + 0.005723238922655582, + 0.003963149152696133, + 0.0032071706373244524, + 0.004773919004946947, + 0.00372017128393054, + 0.002825098345056176, + 0.004674377851188183, + 0.00315085519105196, + 0.004495706409215927, + 0.0051818788051605225, + 0.030363047495484352, + 0.01678970456123352, + 0.014073599129915237, + 0.07158440351486206, + 0.026355277746915817, + 0.032838765531778336, + 0.28890591859817505, + 0.22341255843639374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.007459642831236124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026371391490101814 + ], + "opposite_logprob_tokens": [ + -3.651371479034424 + ], + "logprob_reward": -0.026371391490101814, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371391490101814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028776604682207108, + 0.0217506792396307, + 0.013696568086743355, + 0.03008854016661644, + 0.030240938067436218, + 0.04486585780978203, + 0.030931387096643448, + 0.013922519981861115, + 0.029729017987847328, + 0.7559979557991028 + ], + "credit_map": [ + 0.002735845511779189, + 0.002507568337023258, + 0.006092015653848648, + 0.0031719510443508625, + 0.003774875309318304, + 0.0036491407081484795, + 0.003372813807800412, + 0.0034723959397524595, + 0.0013425914803519845, + 0.0009766406146809459, + 0.0016533714951947331, + 0.0011022082762792706, + 0.0012035074178129435, + 0.007476188708096743, + 0.0014491595793515444, + 0.0065470109693706036, + 0.0014891669852659106, + 0.0017514893552288413, + 0.002340560080483556, + 0.0015449883649125695, + 0.0014003856340423226, + 0.001959056593477726, + 0.00120417436119169, + 0.00200674613006413, + 0.0035145627334713936, + 0.0019243424758315086, + 0.007588659878820181, + 0.003030510153621435, + 0.003435260383412242, + 0.002863172208890319, + 0.0034279015380889177, + 0.004304132889956236, + 0.002680059289559722, + 0.005126560106873512, + 0.004351084586232901, + 0.004492301028221846, + 0.002743918215855956, + 0.003318025264889002, + 0.005023977719247341, + 0.0025050127878785133, + 0.0031064494978636503, + 0.003722422057762742, + 0.011042228899896145, + 0.010666006244719028, + 0.0036100964061915874, + 0.005183023866266012, + 0.0028516578022390604, + 0.00468397093936801, + 0.00511206453666091, + 0.004894552752375603, + 0.0035449694842100143, + 0.00381528795696795, + 0.005010935943573713, + 0.002858330961316824, + 0.0026722357142716646, + 0.003023008815944195, + 0.0011351186549291015, + 0.0011107944883406162, + 0.0014752424322068691, + 0.0021730451844632626, + 0.0026049045845866203, + 0.0014778263866901398, + 0.0024919207207858562, + 0.0014536671806126833, + 0.0023461084347218275, + 0.004563117399811745, + 0.0029315974097698927, + 0.003566925646737218, + 0.004667314235121012, + 0.005123232491314411, + 0.0026231477968394756, + 0.003907575272023678, + 0.03410656377673149, + 0.01947842352092266, + 0.020164737477898598, + 0.07531076669692993, + 0.02927974797785282, + 0.035306040197610855, + 0.30664747953414917, + 0.2357041984796524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.026371391490101814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950592458248138 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950592458248138, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950592458248138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07819856703281403, + 0.03230373188853264, + 0.021219108253717422, + 0.03830605372786522, + 0.034458380192518234, + 0.03305315598845482, + 0.06310126930475235, + 0.018915509805083275, + 0.09209804981946945, + 0.5883461833000183 + ], + "credit_map": [ + 0.005796533077955246, + 0.006376086268573999, + 0.00871699396520853, + 0.009310145862400532, + 0.00986173190176487, + 0.015002735890448093, + 0.0067475163377821445, + 0.016386816278100014, + 0.0020686020143330097, + 0.0017384475795552135, + 0.0024078069254755974, + 0.0022155344486236572, + 0.0017494501080363989, + 0.009776845574378967, + 0.0020107245072722435, + 0.010336320847272873, + 0.001423858804628253, + 0.0020983952563256025, + 0.001878999755717814, + 0.0024357810616493225, + 0.004048548638820648, + 0.003580976277589798, + 0.0026030943263322115, + 0.0031494523864239454, + 0.004452175460755825, + 0.0019039385952055454, + 0.010287784039974213, + 0.003464115085080266, + 0.003835786832496524, + 0.004986090119928122, + 0.0029986489098519087, + 0.0063775088638067245, + 0.0031872191466391087, + 0.005194308236241341, + 0.004057826939970255, + 0.007265984546393156, + 0.0034216942731291056, + 0.0032744850032031536, + 0.004956134129315615, + 0.00310072535648942, + 0.00315209967084229, + 0.003022785997018218, + 0.004496457055211067, + 0.006030040793120861, + 0.0041346196085214615, + 0.006046702619642019, + 0.0025625494308769703, + 0.0036078975535929203, + 0.006476967129856348, + 0.008279137313365936, + 0.00692682433873415, + 0.00809667631983757, + 0.012949211522936821, + 0.00785466656088829, + 0.005359997041523457, + 0.007157785817980766, + 0.001313932123593986, + 0.0016917516477406025, + 0.0016933254664763808, + 0.0023036906495690346, + 0.003006124868988991, + 0.001339962356723845, + 0.0049820044077932835, + 0.002584716770797968, + 0.00793893076479435, + 0.0113345542922616, + 0.013089354149997234, + 0.011764390394091606, + 0.014975957572460175, + 0.01243552751839161, + 0.012053399346768856, + 0.008505934849381447, + 0.025918805971741676, + 0.01521393097937107, + 0.013988018967211246, + 0.06072932854294777, + 0.023304559290409088, + 0.027019159868359566, + 0.23967383801937103, + 0.182498499751091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.005950592458248138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205814331769943 + ], + "opposite_logprob_tokens": [ + -2.812058210372925 + ], + "logprob_reward": -0.06205814331769943, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205814331769943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02488280087709427, + 0.030068229883909225, + 0.016274534165859222, + 0.013070871122181416, + 0.03322116658091545, + 0.027609826996922493, + 0.027626514434814453, + 0.026460319757461548, + 0.07735495269298553, + 0.7234308123588562 + ], + "credit_map": [ + 0.0028808251954615116, + 0.001939129433594644, + 0.0036572765093296766, + 0.0022337627597153187, + 0.0019154915353283286, + 0.0053166598081588745, + 0.0019357178825885057, + 0.005003937985748053, + 0.001852600951679051, + 0.0009584050858393312, + 0.00206691212952137, + 0.001268543303012848, + 0.0012498588766902685, + 0.010857955552637577, + 0.0015401641139760613, + 0.01027379184961319, + 0.001328733516857028, + 0.002416177187114954, + 0.0017180567374452949, + 0.0016130773583427072, + 0.002388308523222804, + 0.0021543996408581734, + 0.0017892596079036593, + 0.0028665231075137854, + 0.002025296213105321, + 0.0014894900377839804, + 0.001673458726145327, + 0.002195768989622593, + 0.001783608691766858, + 0.001150350784882903, + 0.0014727059751749039, + 0.0012801929842680693, + 0.002913119737058878, + 0.004953396040946245, + 0.004894074518233538, + 0.0054995156824588776, + 0.0027295961044728756, + 0.0035262014716863632, + 0.005278227850794792, + 0.003427039133384824, + 0.002257663756608963, + 0.0038050301373004913, + 0.006149810738861561, + 0.005262892693281174, + 0.0023700729943811893, + 0.002574389334768057, + 0.002196290995925665, + 0.0029936786741018295, + 0.0030334736220538616, + 0.003153753001242876, + 0.003180168801918626, + 0.003264154540374875, + 0.00571428844705224, + 0.002712426707148552, + 0.003527382155880332, + 0.003040869254618883, + 0.0017401425866410136, + 0.0026855873875319958, + 0.0027871960774064064, + 0.003173345699906349, + 0.00475602550432086, + 0.002265584422275424, + 0.0057542044669389725, + 0.0032982341945171356, + 0.007981843315064907, + 0.010267731733620167, + 0.010062189772725105, + 0.008457576856017113, + 0.010949654504656792, + 0.012812464497983456, + 0.009316513314843178, + 0.007506984751671553, + 0.03190426155924797, + 0.01829010620713234, + 0.017523787915706635, + 0.07056381553411484, + 0.0283204335719347, + 0.03384365886449814, + 0.2946370542049408, + 0.22834773361682892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.06205814331769943 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493996664881706 + ], + "opposite_logprob_tokens": [ + -2.9299399852752686 + ], + "logprob_reward": -0.05493996664881706, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493996664881706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08096063137054443, + 0.03826066479086876, + 0.08137897402048111, + 0.03156924992799759, + 0.03391685709357262, + 0.030716808512806892, + 0.06410080194473267, + 0.10890951007604599, + 0.0728539377450943, + 0.4573326110839844 + ], + "credit_map": [ + 0.0044119758531451225, + 0.007345004007220268, + 0.006827214267104864, + 0.009982701390981674, + 0.011279236525297165, + 0.013283373787999153, + 0.01251918077468872, + 0.01531194057315588, + 0.001795287593267858, + 0.0011934940703213215, + 0.0027248242404311895, + 0.002331274328753352, + 0.0022978640627115965, + 0.012762871570885181, + 0.002397039206698537, + 0.012758007273077965, + 0.007962465286254883, + 0.011870711110532284, + 0.009887998923659325, + 0.00923116598278284, + 0.00857249740511179, + 0.015871254727244377, + 0.006794809829443693, + 0.011188066564500332, + 0.0022991651203483343, + 0.0025645513087511063, + 0.004969117697328329, + 0.005092991050332785, + 0.005880720913410187, + 0.002931132446974516, + 0.005155407823622227, + 0.002676161006093025, + 0.004289886448532343, + 0.004347511101514101, + 0.006830384023487568, + 0.005846384912729263, + 0.0027348531875759363, + 0.002969980239868164, + 0.003469526069238782, + 0.003428332507610321, + 0.0023443077225238085, + 0.0029742936603724957, + 0.003715073922649026, + 0.004031472373753786, + 0.004166634753346443, + 0.007411760278046131, + 0.002431504661217332, + 0.0036417595110833645, + 0.0074856881983578205, + 0.007024078164249659, + 0.007537902798503637, + 0.0052925399504601955, + 0.012548903934657574, + 0.010634100995957851, + 0.005560729186981916, + 0.008016861043870449, + 0.006959539372473955, + 0.011633828282356262, + 0.008816401474177837, + 0.01454668678343296, + 0.020467659458518028, + 0.00866742990911007, + 0.025633128359913826, + 0.012184834107756615, + 0.006581210996955633, + 0.011063636280596256, + 0.009486347436904907, + 0.007163628004491329, + 0.010501895099878311, + 0.013981406576931477, + 0.00741841085255146, + 0.006657402962446213, + 0.020655488595366478, + 0.013240968808531761, + 0.014562474563717842, + 0.04747689887881279, + 0.021350989118218422, + 0.02196793630719185, + 0.18182915449142456, + 0.13624869287014008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.05493996664881706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This cinematic close-up, with a remarkably shallow depth of field, captures a shimmering soap bubble, suspended almost motionless in the air of a quiet, sunlit room. The camera remains still for a moment, focused on the delicate iridescent film on the bubble's surface, where swirling, hypnotic patterns of mauve, cyan, and liquid gold catch the light. The distortions of the large windows, fisheye reflections, and the room's out-of-focus interior twist and glide across its nearly perfectly spherical shape. Then, the camera begins a steady upward tilt, following the bubble, gently propelled by an invisible column of hot air emanating from a vent in the black metal floor below. The bubble sways unsteadily, its shape momentarily distorting as it rises on the gentle current. Then, it slowly drifts out of the column of air, and the camera follows its graceful downward arc toward the polished, dark hardwood floor. Just before contact, it is captured again by the rising air, and the camera tilts again, tracking its final, brief ascent. Finally, the bubble drifts downwards for the last time; the camera tilts down to capture its descent, and upon contact with the floor, the bubble vanishes instantly, leaving only a tiny, shimmering, wet ring on the wood that quickly evaporates in the soft ambient light. The entire scene is bathed in the warm golden glow of the afternoon sun, illuminating the dancing dust particles and creating an atmosphere of tranquility, contemplation, and transience.", + "seed": 14773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_11/index_4503_seed_14773_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.031006095837801693, + "vlm_logprob": -0.031006095837801693, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.031006095837801693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007898171432316303 + ], + "opposite_logprob_tokens": [ + -6.257898330688477 + ], + "logprob_reward": -0.007898171432316303, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.007898171432316303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11516907811164856, + 0.22849588096141815, + 0.1430531144142151, + 0.06050426885485649, + 0.06619357317686081, + 0.04506176337599754, + 0.1340959370136261, + 0.09157343208789825, + 0.051005374640226364, + 0.06484757363796234 + ], + "credit_map": [ + 0.021189963445067406, + 0.021004317328333855, + 0.010137809440493584, + 0.01129485946148634, + 0.014019536785781384, + 0.01375920232385397, + 0.011787527240812778, + 0.011975865811109543, + 0.033772241324186325, + 0.036982014775276184, + 0.019637631252408028, + 0.020613674074411392, + 0.026277005672454834, + 0.025571703910827637, + 0.03770829737186432, + 0.02793331816792488, + 0.015320691280066967, + 0.031944453716278076, + 0.01391529105603695, + 0.012213087640702724, + 0.02066742815077305, + 0.016188127920031548, + 0.012103116139769554, + 0.020700912922620773, + 0.009526646696031094, + 0.009234084747731686, + 0.003587088780477643, + 0.005308537743985653, + 0.011635024100542068, + 0.007408041972666979, + 0.006628307513892651, + 0.007176538929343224, + 0.008423221297562122, + 0.015745030716061592, + 0.004528604447841644, + 0.005646835081279278, + 0.008223547600209713, + 0.00794295035302639, + 0.0071779112331569195, + 0.008505469188094139, + 0.004941804800182581, + 0.008059092797338963, + 0.005292815156280994, + 0.004072861280292273, + 0.005300214514136314, + 0.005229030270129442, + 0.004190255422145128, + 0.007975690066814423, + 0.009100154042243958, + 0.012144590727984905, + 0.019641626626253128, + 0.020795054733753204, + 0.010561884380877018, + 0.025257844477891922, + 0.021034181118011475, + 0.01556059904396534, + 0.006706311833113432, + 0.0076903230510652065, + 0.012233552522957325, + 0.013676844537258148, + 0.007957224734127522, + 0.007172602694481611, + 0.024331912398338318, + 0.011804667301476002, + 0.0034463610500097275, + 0.004658942576497793, + 0.00528369564563036, + 0.006181037053465843, + 0.004227703437209129, + 0.010451175272464752, + 0.010988063178956509, + 0.0057683950290083885, + 0.008428825996816158, + 0.007891529239714146, + 0.012833723798394203, + 0.0074618165381252766, + 0.005012001376599073, + 0.007153133396059275, + 0.007526215631514788, + 0.008540328592061996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.007898171432316303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038119908422231674 + ], + "opposite_logprob_tokens": [ + -3.2881197929382324 + ], + "logprob_reward": -0.038119908422231674, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038119908422231674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045364752411842346, + 0.09496241807937622, + 0.250924676656723, + 0.04938361793756485, + 0.06357601284980774, + 0.04618220776319504, + 0.08945338428020477, + 0.09232717752456665, + 0.16586604714393616, + 0.10195973515510559 + ], + "credit_map": [ + 0.0055762724950909615, + 0.008308864198625088, + 0.003930380102247, + 0.0046807341277599335, + 0.006432370748370886, + 0.005714551545679569, + 0.0043259188532829285, + 0.00639566034078598, + 0.012959717772901058, + 0.015620063990354538, + 0.008073095232248306, + 0.008772069588303566, + 0.01070934347808361, + 0.011126479133963585, + 0.014134673401713371, + 0.013566968962550163, + 0.04055259749293327, + 0.03284720331430435, + 0.027679281309247017, + 0.01665830798447132, + 0.027209285646677017, + 0.03482268750667572, + 0.027995111420750618, + 0.04316018149256706, + 0.0071883369237184525, + 0.00848076306283474, + 0.004851067438721657, + 0.003991142846643925, + 0.007222131360322237, + 0.005816213320940733, + 0.0050729308277368546, + 0.006761032156646252, + 0.007173316553235054, + 0.017629235982894897, + 0.004401911981403828, + 0.005225063301622868, + 0.006883726920932531, + 0.007963650859892368, + 0.006822253577411175, + 0.007476859726011753, + 0.005137376021593809, + 0.007111393380910158, + 0.005575258284807205, + 0.004443705081939697, + 0.005849161650985479, + 0.004884003661572933, + 0.0036644721403717995, + 0.009516836144030094, + 0.00562278600409627, + 0.007390922401100397, + 0.01638609915971756, + 0.009989550337195396, + 0.007700337562710047, + 0.012923922389745712, + 0.015179822221398354, + 0.014259940944612026, + 0.007040714379400015, + 0.009019333869218826, + 0.012229927815496922, + 0.014980066567659378, + 0.009352772496640682, + 0.009760643355548382, + 0.020053256303071976, + 0.009890468791127205, + 0.011224658228456974, + 0.015567108988761902, + 0.015308785252273083, + 0.0210023894906044, + 0.014032531529664993, + 0.025799788534641266, + 0.043767187744379044, + 0.019163595512509346, + 0.008763214573264122, + 0.009750154800713062, + 0.019934533163905144, + 0.010012333281338215, + 0.008052637800574303, + 0.00964142382144928, + 0.01270210649818182, + 0.02310332842171192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.038119908422231674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740807089954615 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740807089954615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740807089954615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.103885218501091, + 0.13178275525569916, + 0.27206701040267944, + 0.030358925461769104, + 0.08751633763313293, + 0.05468887835741043, + 0.14369994401931763, + 0.08314482867717743, + 0.05751698836684227, + 0.03533913567662239 + ], + "credit_map": [ + 0.015574902296066284, + 0.019964054226875305, + 0.0076042297296226025, + 0.010937216691672802, + 0.01527631189674139, + 0.012655103579163551, + 0.010902024805545807, + 0.010971371084451675, + 0.020309899002313614, + 0.02305421605706215, + 0.01131517719477415, + 0.01144325453788042, + 0.015306231565773487, + 0.014525353908538818, + 0.020016591995954514, + 0.01581202819943428, + 0.03555646538734436, + 0.051647331565618515, + 0.02821800298988819, + 0.021024487912654877, + 0.03554760664701462, + 0.03319486603140831, + 0.02692531980574131, + 0.03995291516184807, + 0.005969599820673466, + 0.004904787056148052, + 0.002854602411389351, + 0.002643572399392724, + 0.0034424755722284317, + 0.003709289478138089, + 0.0023666510824114084, + 0.004467947408556938, + 0.009376992471516132, + 0.02206115983426571, + 0.0059415074065327644, + 0.007955213077366352, + 0.00996247585862875, + 0.010917848907411098, + 0.010760373435914516, + 0.010540762916207314, + 0.00784122385084629, + 0.009746714495122433, + 0.006716899108141661, + 0.005090347025543451, + 0.005469433963298798, + 0.007010177243500948, + 0.0037503992207348347, + 0.009063683450222015, + 0.006400078535079956, + 0.004151538945734501, + 0.05430375412106514, + 0.008927196264266968, + 0.007695834152400494, + 0.04602176323533058, + 0.007480451837182045, + 0.008719314821064472, + 0.006383726838976145, + 0.008427275344729424, + 0.010729781351983547, + 0.012664394453167915, + 0.006797738838940859, + 0.005452258978039026, + 0.021188518032431602, + 0.011501140892505646, + 0.0029288805089890957, + 0.004377713892608881, + 0.006069611292332411, + 0.008323577232658863, + 0.004025258123874664, + 0.0062362197786569595, + 0.018099969252943993, + 0.007455761078745127, + 0.003415292827412486, + 0.0034113836009055376, + 0.00916081853210926, + 0.004436577670276165, + 0.0024286953266710043, + 0.003028406761586666, + 0.0027510214131325483, + 0.006706940475851297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.006740807089954615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899120450019836 + ], + "opposite_logprob_tokens": [ + -2.578991174697876 + ], + "logprob_reward": -0.07899120450019836, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899120450019836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030280090868473053, + 0.30788692831993103, + 0.15202520787715912, + 0.02999257482588291, + 0.02985497936606407, + 0.04351216182112694, + 0.08625523746013641, + 0.10245121270418167, + 0.10632310807704926, + 0.1114184632897377 + ], + "credit_map": [ + 0.0036684032529592514, + 0.005232642404735088, + 0.002268123207613826, + 0.003858046606183052, + 0.004283037502318621, + 0.00381219619885087, + 0.003522100392729044, + 0.003635542234405875, + 0.04694947227835655, + 0.05293445289134979, + 0.02677977830171585, + 0.026781102642416954, + 0.03519667685031891, + 0.03388770669698715, + 0.048420436680316925, + 0.03693730756640434, + 0.019838135689496994, + 0.02888498827815056, + 0.015559112653136253, + 0.012005185708403587, + 0.020341219380497932, + 0.01836455799639225, + 0.014345725066959858, + 0.022686298936605453, + 0.0036106950137764215, + 0.004875923506915569, + 0.0025851670652627945, + 0.0022816755808889866, + 0.0056486790999770164, + 0.003926766570657492, + 0.0033611163962632418, + 0.003702553454786539, + 0.004663192201405764, + 0.0066726370714604855, + 0.0022780424915254116, + 0.0029510974418371916, + 0.004050331190228462, + 0.003634585766121745, + 0.0032448931597173214, + 0.00236020190641284, + 0.006394240539520979, + 0.005169745534658432, + 0.006014523562043905, + 0.004020777065306902, + 0.0040494175627827644, + 0.005351553205400705, + 0.002606556285172701, + 0.00990535132586956, + 0.006574027240276337, + 0.008804342709481716, + 0.011211100034415722, + 0.015599368140101433, + 0.005911034066230059, + 0.01130147185176611, + 0.01607549376785755, + 0.010778410360217094, + 0.008395906537771225, + 0.01230619940906763, + 0.010926532559096813, + 0.015600407496094704, + 0.009460554458200932, + 0.010648851282894611, + 0.022962410002946854, + 0.012150359340012074, + 0.0075493487529456615, + 0.010882367379963398, + 0.009763757698237896, + 0.013802969828248024, + 0.007980437949299812, + 0.018548469990491867, + 0.026296526193618774, + 0.011499240063130856, + 0.009296770207583904, + 0.009731333702802658, + 0.02711137756705284, + 0.013916890136897564, + 0.006833105813711882, + 0.01053085457533598, + 0.011972135864198208, + 0.022026000544428825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.07899120450019836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280387744307518 + ], + "opposite_logprob_tokens": [ + -3.773280382156372 + ], + "logprob_reward": -0.023280387744307518, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280387744307518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04093490168452263, + 0.216984361410141, + 0.15610498189926147, + 0.04413692653179169, + 0.04254911094903946, + 0.04382026195526123, + 0.15181073546409607, + 0.1698082685470581, + 0.06335742026567459, + 0.07049302756786346 + ], + "credit_map": [ + 0.005594491492956877, + 0.005167492665350437, + 0.004604467656463385, + 0.00369356875307858, + 0.004245112184435129, + 0.006015373393893242, + 0.0034624170511960983, + 0.008151976391673088, + 0.03386344760656357, + 0.03531424328684807, + 0.01913699321448803, + 0.01877519302070141, + 0.024799523875117302, + 0.023869607597589493, + 0.034920427948236465, + 0.026304904371500015, + 0.017829230055212975, + 0.0338473841547966, + 0.015282013453543186, + 0.013084852136671543, + 0.021348340436816216, + 0.017973650246858597, + 0.013719756156206131, + 0.023019755259156227, + 0.005216538440436125, + 0.007260454818606377, + 0.0024017314426600933, + 0.003751002950593829, + 0.010751917958259583, + 0.00557007035240531, + 0.005131442099809647, + 0.0040537649765610695, + 0.005090947262942791, + 0.011093324050307274, + 0.0033149125520139933, + 0.0038489913567900658, + 0.005430128891021013, + 0.005494846496731043, + 0.004069188609719276, + 0.00420676963403821, + 0.006978955585509539, + 0.004643885884433985, + 0.0062991599552333355, + 0.004196302033960819, + 0.003502628533169627, + 0.005726576317101717, + 0.003042664611712098, + 0.00943008903414011, + 0.006744033191353083, + 0.011475329287350178, + 0.0389409065246582, + 0.013126474805176258, + 0.008866360411047935, + 0.043233200907707214, + 0.013072595000267029, + 0.016351832076907158, + 0.011800152249634266, + 0.018088700249791145, + 0.02081306092441082, + 0.02671307697892189, + 0.018124433234333992, + 0.01575344242155552, + 0.03517869487404823, + 0.023336704820394516, + 0.0058562010526657104, + 0.007273489609360695, + 0.006335785146802664, + 0.006687846966087818, + 0.005832195281982422, + 0.012809772044420242, + 0.012287762016057968, + 0.006274370476603508, + 0.006723636761307716, + 0.004326966125518084, + 0.019227923825383186, + 0.009529056958854198, + 0.0036749448627233505, + 0.005891719367355108, + 0.005764237605035305, + 0.015354542061686516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.023280387744307518 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This cinematic close-up, with a remarkably shallow depth of field, captures a shimmering soap bubble, suspended almost motionless in the air of a quiet, sunlit room. The camera remains still for a moment, focused on the delicate iridescent film on the bubble's surface, where swirling, hypnotic patterns of mauve, cyan, and liquid gold catch the light. The distortions of the large windows, fisheye reflections, and the room's out-of-focus interior twist and glide across its nearly perfectly spherical shape. Then, the camera begins a steady upward tilt, following the bubble, gently propelled by an invisible column of hot air emanating from a vent in the black metal floor below. The bubble sways unsteadily, its shape momentarily distorting as it rises on the gentle current. Then, it slowly drifts out of the column of air, and the camera follows its graceful downward arc toward the polished, dark hardwood floor. Just before contact, it is captured again by the rising air, and the camera tilts again, tracking its final, brief ascent. Finally, the bubble drifts downwards for the last time; the camera tilts down to capture its descent, and upon contact with the floor, the bubble vanishes instantly, leaving only a tiny, shimmering, wet ring on the wood that quickly evaporates in the soft ambient light. The entire scene is bathed in the warm golden glow of the afternoon sun, illuminating the dancing dust particles and creating an atmosphere of tranquility, contemplation, and transience.", + "seed": 14774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_11/index_4503_seed_14774_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.041930786520242694, + "vlm_logprob": -0.041930786520242694, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.041930786520242694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006998193915933371 + ], + "opposite_logprob_tokens": [ + -6.881998062133789 + ], + "logprob_reward": -0.006998193915933371, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.006998193915933371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09315958619117737, + 0.0512002594769001, + 0.07649870961904526, + 0.12366814911365509, + 0.05461890250444412, + 0.10961175709962845, + 0.21192649006843567, + 0.0777750238776207, + 0.08358678221702576, + 0.11795435100793839 + ], + "credit_map": [ + 0.012520167045295238, + 0.01070176251232624, + 0.013796060346066952, + 0.013261121697723866, + 0.008412552997469902, + 0.009560937993228436, + 0.009140850976109505, + 0.01576613448560238, + 0.005673260428011417, + 0.005810739006847143, + 0.006628342904150486, + 0.006778656970709562, + 0.0054730018600821495, + 0.006640901789069176, + 0.0054674819111824036, + 0.008727875538170338, + 0.009206768125295639, + 0.012982633896172047, + 0.00829868484288454, + 0.009205671958625317, + 0.007251187227666378, + 0.007168157957494259, + 0.0059961280785501, + 0.016389479860663414, + 0.012441486120223999, + 0.018402988091111183, + 0.015628617256879807, + 0.019984643906354904, + 0.01058658491820097, + 0.010302674025297165, + 0.02001132071018219, + 0.016309835016727448, + 0.006168840453028679, + 0.007629169151186943, + 0.008107314817607403, + 0.005788844078779221, + 0.004954280331730843, + 0.005740657448768616, + 0.008011543191969395, + 0.008218251168727875, + 0.01863524317741394, + 0.007699288427829742, + 0.02926364354789257, + 0.013399980030953884, + 0.010439914651215076, + 0.014741259627044201, + 0.007441632449626923, + 0.007990792393684387, + 0.03186584264039993, + 0.022015579044818878, + 0.016867414116859436, + 0.012799717485904694, + 0.021732280030846596, + 0.060758668929338455, + 0.012041754089295864, + 0.03384523093700409, + 0.0058195702731609344, + 0.012125635519623756, + 0.010062817484140396, + 0.01500216219574213, + 0.009264069609344006, + 0.004728708416223526, + 0.013941041193902493, + 0.006831016391515732, + 0.007868233136832714, + 0.014896748587489128, + 0.009349874220788479, + 0.012123143300414085, + 0.013496160507202148, + 0.007365641184151173, + 0.008969702757894993, + 0.009517286904156208, + 0.010659734718501568, + 0.03070230968296528, + 0.009783701039850712, + 0.010661729611456394, + 0.02618538774549961, + 0.008192254230380058, + 0.013065162114799023, + 0.008704067207872868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.006998193915933371 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005749642848969 + ], + "opposite_logprob_tokens": [ + -2.6950573921203613 + ], + "logprob_reward": -0.07005749642848969, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005749642848969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06790357083082199, + 0.039584364742040634, + 0.06486383825540543, + 0.10338717699050903, + 0.08154749870300293, + 0.040231604129076004, + 0.09117287397384644, + 0.0541447289288044, + 0.2703409492969513, + 0.18682344257831573 + ], + "credit_map": [ + 0.008385192602872849, + 0.010453316383063793, + 0.013056221418082714, + 0.0066031524911522865, + 0.006214729510247707, + 0.007059972267597914, + 0.0066917152144014835, + 0.009439271874725819, + 0.004492391366511583, + 0.004106412176042795, + 0.005747940391302109, + 0.0029263414908200502, + 0.003989167511463165, + 0.004692108370363712, + 0.004470885265618563, + 0.009159116074442863, + 0.007899215444922447, + 0.010304446332156658, + 0.007199107203632593, + 0.010635435581207275, + 0.00900628138333559, + 0.006025757174938917, + 0.006438450887799263, + 0.007355145178735256, + 0.011906414292752743, + 0.007357226684689522, + 0.013684045523405075, + 0.010874330997467041, + 0.006004103925079107, + 0.015105296857655048, + 0.01047529373317957, + 0.027980463579297066, + 0.009970702230930328, + 0.006838998291641474, + 0.011414334177970886, + 0.007332474924623966, + 0.0068540871143341064, + 0.017501601949334145, + 0.008003445342183113, + 0.013631854206323624, + 0.004974074196070433, + 0.002515943953767419, + 0.0069528743624687195, + 0.0035868515260517597, + 0.003872203640639782, + 0.010155076161026955, + 0.00355101702734828, + 0.004623561631888151, + 0.012632952071726322, + 0.011126574128866196, + 0.009265178814530373, + 0.007134073413908482, + 0.009625767357647419, + 0.022787507623434067, + 0.007447545416653156, + 0.011153274215757847, + 0.00364688946865499, + 0.00887261237949133, + 0.006516559515148401, + 0.007702402770519257, + 0.00797935202717781, + 0.004803732968866825, + 0.009006168693304062, + 0.005617010407149792, + 0.027261173352599144, + 0.03284212574362755, + 0.02639538049697876, + 0.031616028398275375, + 0.045058656483888626, + 0.026925979182124138, + 0.057180676609277725, + 0.02306092157959938, + 0.02015685848891735, + 0.028763042762875557, + 0.02148822322487831, + 0.023588480427861214, + 0.02720785327255726, + 0.02094099298119545, + 0.022264689207077026, + 0.022413307800889015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.07005749642848969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634746376425028 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634746376425028, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634746376425028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09825994819402695, + 0.021783677861094475, + 0.08839038014411926, + 0.1046910285949707, + 0.03941020369529724, + 0.10559011995792389, + 0.12263233214616776, + 0.10408636182546616, + 0.11513219773769379, + 0.20002374053001404 + ], + "credit_map": [ + 0.01309593953192234, + 0.009987279772758484, + 0.016616567969322205, + 0.009060762822628021, + 0.009403114207088947, + 0.011860036291182041, + 0.013130548410117626, + 0.015105696395039558, + 0.0025902357883751392, + 0.0024731303565204144, + 0.003521451260894537, + 0.002671021968126297, + 0.0017244588816538453, + 0.0021099222358316183, + 0.002319130813702941, + 0.004374327138066292, + 0.011938407085835934, + 0.017164109274744987, + 0.009497757069766521, + 0.011373840272426605, + 0.011761709116399288, + 0.00714713241904974, + 0.008183426223695278, + 0.011323994025588036, + 0.01020533312112093, + 0.012884652242064476, + 0.015379720367491245, + 0.01409005094319582, + 0.009351283311843872, + 0.016470743343234062, + 0.012645835988223553, + 0.013663407415151596, + 0.006243582349270582, + 0.003647674573585391, + 0.006055703852325678, + 0.004839994013309479, + 0.0037342305295169353, + 0.006111446302384138, + 0.003221409860998392, + 0.005556166172027588, + 0.009410419501364231, + 0.003459247527644038, + 0.02448675036430359, + 0.005277626682072878, + 0.0051552061922848225, + 0.03905206173658371, + 0.0036522855516523123, + 0.015096518211066723, + 0.016105128452181816, + 0.010864770039916039, + 0.013387699611485004, + 0.010237104259431362, + 0.011972813867032528, + 0.034023646265268326, + 0.007573654875159264, + 0.01846751756966114, + 0.007759072817862034, + 0.014265579171478748, + 0.017642207443714142, + 0.015393231064081192, + 0.014346899464726448, + 0.009781768545508385, + 0.015026208013296127, + 0.00987139530479908, + 0.012551043182611465, + 0.012968593277037144, + 0.00995660200715065, + 0.016912085935473442, + 0.020054752007126808, + 0.010277099907398224, + 0.02111833356320858, + 0.011293691582977772, + 0.018103327602148056, + 0.05811136215925217, + 0.016677161678671837, + 0.01451227068901062, + 0.04563290253281593, + 0.012586232274770737, + 0.018337968736886978, + 0.016062505543231964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.007634746376425028 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002804428339005 + ], + "opposite_logprob_tokens": [ + -2.695028066635132 + ], + "logprob_reward": -0.07002804428339005, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002804428339005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048593103885650635, + 0.018714258447289467, + 0.048924025148153305, + 0.05412382632493973, + 0.13879473507404327, + 0.21457916498184204, + 0.06358844041824341, + 0.08211709558963776, + 0.1479136347770691, + 0.18265168368816376 + ], + "credit_map": [ + 0.006436608731746674, + 0.005650282371789217, + 0.009692235849797726, + 0.0040862346068024635, + 0.004383708816021681, + 0.005297303199768066, + 0.004471524618566036, + 0.008575208485126495, + 0.0021915931720286608, + 0.0017244754126295447, + 0.0025237351655960083, + 0.0017808577977120876, + 0.002344654407352209, + 0.003008540952578187, + 0.0018365047872066498, + 0.0033038975670933723, + 0.006709325592964888, + 0.009643346071243286, + 0.0059430296532809734, + 0.005498635116964579, + 0.006307343486696482, + 0.004445445258170366, + 0.003861001692712307, + 0.006515896413475275, + 0.007162358611822128, + 0.004206743091344833, + 0.005739759653806686, + 0.004229346290230751, + 0.002184218028560281, + 0.011240593157708645, + 0.003943329676985741, + 0.015417478047311306, + 0.01890309527516365, + 0.01633414998650551, + 0.019760075956583023, + 0.015571900643408298, + 0.012095194309949875, + 0.020393721759319305, + 0.013899924233555794, + 0.02183668501675129, + 0.01700005494058132, + 0.008016452193260193, + 0.044762980192899704, + 0.014747713692486286, + 0.01247989758849144, + 0.07596522569656372, + 0.009170988574624062, + 0.032435834407806396, + 0.009762481786310673, + 0.010943860746920109, + 0.004883408080786467, + 0.00620210962370038, + 0.007157383486628532, + 0.010896616615355015, + 0.006590917240828276, + 0.0071516684256494045, + 0.00679583428427577, + 0.012075996026396751, + 0.01330478023737669, + 0.010868728160858154, + 0.011643728241324425, + 0.0073611740954220295, + 0.011972603388130665, + 0.008094245567917824, + 0.01585800014436245, + 0.01788824424147606, + 0.013406672514975071, + 0.02035224810242653, + 0.024970388039946556, + 0.013351275585591793, + 0.027464857324957848, + 0.014621945098042488, + 0.015507428906857967, + 0.05494656786322594, + 0.01355072483420372, + 0.01525209192186594, + 0.04251747205853462, + 0.01035654079169035, + 0.012645548209547997, + 0.017875319346785545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.07002804428339005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493545159697533 + ], + "opposite_logprob_tokens": [ + -2.9299354553222656 + ], + "logprob_reward": -0.05493545159697533, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493545159697533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03726474940776825, + 0.04225558042526245, + 0.03385554626584053, + 0.05659578740596771, + 0.1681840568780899, + 0.26455995440483093, + 0.09359915554523468, + 0.10492599755525589, + 0.11068477481603622, + 0.0880744531750679 + ], + "credit_map": [ + 0.005269686691462994, + 0.004522836767137051, + 0.005781992804259062, + 0.0038822428323328495, + 0.0032197777181863785, + 0.004240004345774651, + 0.0033038256224244833, + 0.0070443847216665745, + 0.005126467440277338, + 0.004409019835293293, + 0.006458350922912359, + 0.0036278970073908567, + 0.0047278208658099174, + 0.007241544313728809, + 0.0036882171407341957, + 0.006976263597607613, + 0.004206846002489328, + 0.004574266262352467, + 0.005308863706886768, + 0.0028730707708746195, + 0.003894209861755371, + 0.004461623728275299, + 0.002913442673161626, + 0.005623220931738615, + 0.005350016988813877, + 0.007897674106061459, + 0.010289999656379223, + 0.00953962653875351, + 0.0036641121841967106, + 0.005693807732313871, + 0.008772743865847588, + 0.005387802142649889, + 0.024832336232066154, + 0.019997114315629005, + 0.025543974712491035, + 0.017478296533226967, + 0.013776160776615143, + 0.02421044558286667, + 0.014101880602538586, + 0.028243834152817726, + 0.025450002402067184, + 0.009929325431585312, + 0.04722802713513374, + 0.018724555149674416, + 0.015280340798199177, + 0.09434542059898376, + 0.010551304556429386, + 0.04305097833275795, + 0.013370699249207973, + 0.010688304901123047, + 0.008135901764035225, + 0.007436010520905256, + 0.009796461090445518, + 0.024050604552030563, + 0.005776539910584688, + 0.014344630762934685, + 0.007952268235385418, + 0.013000334613025188, + 0.012469690293073654, + 0.02129560522735119, + 0.012404529377818108, + 0.006087584421038628, + 0.02378036454319954, + 0.007935620844364166, + 0.01184234581887722, + 0.01753651723265648, + 0.01216317992657423, + 0.016836902126669884, + 0.015561048872768879, + 0.010849137790501118, + 0.014145485125482082, + 0.011750164441764355, + 0.006471844855695963, + 0.02225845493376255, + 0.006801533978432417, + 0.014593146741390228, + 0.016373520717024803, + 0.004685644526034594, + 0.010228168219327927, + 0.006662136409431696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.05493545159697533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This cinematic close-up, with a remarkably shallow depth of field, captures a shimmering soap bubble, suspended almost motionless in the air of a quiet, sunlit room. The camera remains still for a moment, focused on the delicate iridescent film on the bubble's surface, where swirling, hypnotic patterns of mauve, cyan, and liquid gold catch the light. The distortions of the large windows, fisheye reflections, and the room's out-of-focus interior twist and glide across its nearly perfectly spherical shape. Then, the camera begins a steady upward tilt, following the bubble, gently propelled by an invisible column of hot air emanating from a vent in the black metal floor below. The bubble sways unsteadily, its shape momentarily distorting as it rises on the gentle current. Then, it slowly drifts out of the column of air, and the camera follows its graceful downward arc toward the polished, dark hardwood floor. Just before contact, it is captured again by the rising air, and the camera tilts again, tracking its final, brief ascent. Finally, the bubble drifts downwards for the last time; the camera tilts down to capture its descent, and upon contact with the floor, the bubble vanishes instantly, leaving only a tiny, shimmering, wet ring on the wood that quickly evaporates in the soft ambient light. The entire scene is bathed in the warm golden glow of the afternoon sun, illuminating the dancing dust particles and creating an atmosphere of tranquility, contemplation, and transience.", + "seed": 14775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_11/index_4503_seed_14775_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15352321192622184, + "vlm_logprob": -0.15352321192622184, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.15352321192622184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008751136250793934 + ], + "opposite_logprob_tokens": [ + -5.508750915527344 + ], + "logprob_reward": -0.008751136250793934, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008751136250793934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15028555691242218, + 0.08304524421691895, + 0.17573301494121552, + 0.03613441437482834, + 0.09360217303037643, + 0.07827841490507126, + 0.08374273031949997, + 0.13615265488624573, + 0.10569557547569275, + 0.05733015388250351 + ], + "credit_map": [ + 0.032256629317998886, + 0.011465643532574177, + 0.019274219870567322, + 0.015701038762927055, + 0.02478799968957901, + 0.01194743812084198, + 0.020417166873812675, + 0.01443543005734682, + 0.0144593371078372, + 0.008463706821203232, + 0.0180169977247715, + 0.012025895528495312, + 0.007142883259803057, + 0.005349996034055948, + 0.009564507752656937, + 0.008021915331482887, + 0.016562432050704956, + 0.007723608054220676, + 0.06091760843992233, + 0.03821777179837227, + 0.0077574728056788445, + 0.00543044600635767, + 0.015832364559173584, + 0.023291336372494698, + 0.004171506501734257, + 0.002467377344146371, + 0.004273276776075363, + 0.005504417698830366, + 0.004031384829431772, + 0.003922150004655123, + 0.0049951146356761456, + 0.006769183557480574, + 0.009913787245750427, + 0.009271059185266495, + 0.013621410354971886, + 0.015191419050097466, + 0.013041911646723747, + 0.010679872706532478, + 0.010694220662117004, + 0.011188494972884655, + 0.011713752523064613, + 0.010488945059478283, + 0.011484992690384388, + 0.010165414772927761, + 0.008463363163173199, + 0.009096518158912659, + 0.0071630533784627914, + 0.00970237422734499, + 0.01015008520334959, + 0.00612092949450016, + 0.005786010529845953, + 0.008069982752203941, + 0.02476794458925724, + 0.014423240907490253, + 0.008211802691221237, + 0.0062127429991960526, + 0.013850784860551357, + 0.01670251227915287, + 0.017964856699109077, + 0.01394321583211422, + 0.01587783172726631, + 0.01883659139275551, + 0.02058657445013523, + 0.018390284851193428, + 0.01060203555971384, + 0.014733770862221718, + 0.017995165660977364, + 0.01615232415497303, + 0.015329500660300255, + 0.009450726211071014, + 0.011979327537119389, + 0.00945273321121931, + 0.0042441654950380325, + 0.007308015134185553, + 0.0066783190704882145, + 0.007673972751945257, + 0.003949280828237534, + 0.006736918818205595, + 0.013100899755954742, + 0.007638583891093731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.008751136250793934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709277868270874 + ], + "opposite_logprob_tokens": [ + -2.1270928382873535 + ], + "logprob_reward": -0.12709277868270874, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709277868270874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07830348610877991, + 0.17622064054012299, + 0.09901430457830429, + 0.10101182013750076, + 0.05196508392691612, + 0.0477839931845665, + 0.0984707623720169, + 0.13150854408740997, + 0.16412003338336945, + 0.05160132795572281 + ], + "credit_map": [ + 0.014428956434130669, + 0.007276562042534351, + 0.010853508487343788, + 0.010990983806550503, + 0.007150891702622175, + 0.006367549765855074, + 0.011683765798807144, + 0.00955126527696848, + 0.026360025629401207, + 0.016563264653086662, + 0.03162238746881485, + 0.02437036857008934, + 0.02202264964580536, + 0.011814060620963573, + 0.02239946462213993, + 0.021068407222628593, + 0.011578086763620377, + 0.01212098728865385, + 0.015445545315742493, + 0.023403283208608627, + 0.009948879480361938, + 0.005150417797267437, + 0.010222741402685642, + 0.011144354939460754, + 0.010238486342132092, + 0.009563252329826355, + 0.018471039831638336, + 0.013735401444137096, + 0.014881795272231102, + 0.010101726278662682, + 0.012902900576591492, + 0.011117208749055862, + 0.006164684891700745, + 0.005687336903065443, + 0.0076429410837590694, + 0.009655502624809742, + 0.004235971253365278, + 0.005455236881971359, + 0.005910119507461786, + 0.007213287521153688, + 0.005392343271523714, + 0.00557703198865056, + 0.005718613509088755, + 0.006977323908358812, + 0.006076193414628506, + 0.0050896769389510155, + 0.006137502379715443, + 0.006815304514020681, + 0.014048117212951183, + 0.017384713515639305, + 0.00981118530035019, + 0.011080386117100716, + 0.014941730536520481, + 0.011887994594871998, + 0.011281326413154602, + 0.008035300299525261, + 0.013167220167815685, + 0.01767539419233799, + 0.018954632803797722, + 0.012976597063243389, + 0.015212656930088997, + 0.021916810423135757, + 0.016714047640562057, + 0.014891188591718674, + 0.017863549292087555, + 0.022292640060186386, + 0.025823822245001793, + 0.024965602904558182, + 0.020326193422079086, + 0.016866911202669144, + 0.02197413332760334, + 0.014007160440087318, + 0.004222020506858826, + 0.00358587852679193, + 0.006412276066839695, + 0.008516870439052582, + 0.0036142882890999317, + 0.004281133413314819, + 0.014668858610093594, + 0.006299994885921478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.12709277868270874 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1789625883102417, + 0.13789449632167816, + 0.11344990879297256, + 0.05125926062464714, + 0.030494006350636482, + 0.18563342094421387, + 0.10638786852359772, + 0.08489476889371872, + 0.07087425887584686, + 0.04014937952160835 + ], + "credit_map": [ + 0.02558634616434574, + 0.011684800498187542, + 0.039605215191841125, + 0.021638529375195503, + 0.01388691458851099, + 0.01267644390463829, + 0.021096164360642433, + 0.03278817608952522, + 0.022031014785170555, + 0.009158957749605179, + 0.03968390077352524, + 0.02758640982210636, + 0.009459873661398888, + 0.004808115307241678, + 0.011788460426032543, + 0.013377762399613857, + 0.013228575699031353, + 0.011070127598941326, + 0.028592579066753387, + 0.02758127823472023, + 0.006025038193911314, + 0.003949240781366825, + 0.012394784018397331, + 0.010608285665512085, + 0.005806191358715296, + 0.004081851337105036, + 0.008495332673192024, + 0.00783518236130476, + 0.00648756418377161, + 0.004411890171468258, + 0.006974060088396072, + 0.007167187985032797, + 0.002694870810955763, + 0.002733126748353243, + 0.004754077177494764, + 0.0049811601638793945, + 0.004235780797898769, + 0.0036918949335813522, + 0.003866551211103797, + 0.0035365456715226173, + 0.02668454684317112, + 0.021972620859742165, + 0.01618548110127449, + 0.019695116207003593, + 0.026638083159923553, + 0.027444999665021896, + 0.025463629513978958, + 0.021548951044678688, + 0.012454591691493988, + 0.008910787291824818, + 0.007136805448681116, + 0.00444293487817049, + 0.036980465054512024, + 0.022666731849312782, + 0.007172542158514261, + 0.006623005028814077, + 0.008269209414720535, + 0.011178338900208473, + 0.00922261830419302, + 0.010395649820566177, + 0.011447626166045666, + 0.011195782572031021, + 0.013718747533857822, + 0.009466797113418579, + 0.006484385579824448, + 0.008059676736593246, + 0.009221656247973442, + 0.009329814463853836, + 0.011478443630039692, + 0.006272852886468172, + 0.012918398715555668, + 0.007109023630619049, + 0.0033552111126482487, + 0.004238966386765242, + 0.003697566222399473, + 0.005327744875103235, + 0.00347141083329916, + 0.004483757074922323, + 0.010487865656614304, + 0.005086858756840229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.012543476186692715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761763453483582 + ], + "opposite_logprob_tokens": [ + -0.8261763453483582 + ], + "logprob_reward": -0.5761763453483582, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761763453483582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07256672531366348, + 0.0277999434620142, + 0.04059552028775215, + 0.04948754981160164, + 0.06823370605707169, + 0.07037857174873352, + 0.09175120294094086, + 0.22761844098567963, + 0.28344595432281494, + 0.06812231242656708 + ], + "credit_map": [ + 0.008303377777338028, + 0.005435132421553135, + 0.01186011079698801, + 0.008424747735261917, + 0.010567409917712212, + 0.00620888639241457, + 0.012396782636642456, + 0.009370289742946625, + 0.003797205165028572, + 0.0032870990689843893, + 0.004252701532095671, + 0.003930026665329933, + 0.002573868492618203, + 0.002023606328293681, + 0.0031466553919017315, + 0.0047887819819152355, + 0.00342772644944489, + 0.003554822178557515, + 0.005411257967352867, + 0.006142060272395611, + 0.004993922542780638, + 0.0042938231490552425, + 0.008069595322012901, + 0.004702315665781498, + 0.0060384091921150684, + 0.005021716933697462, + 0.007134563289582729, + 0.008657444268465042, + 0.0033888262696564198, + 0.004777437541633844, + 0.005103357136249542, + 0.009365799836814404, + 0.00794015359133482, + 0.008099528960883617, + 0.00875266082584858, + 0.011719726026058197, + 0.009118055924773216, + 0.007018836215138435, + 0.006560199894011021, + 0.009024549275636673, + 0.007144785486161709, + 0.007215701974928379, + 0.005670835264027119, + 0.0097957169637084, + 0.011504064314067364, + 0.007884991355240345, + 0.014047560282051563, + 0.007114925421774387, + 0.008847380988299847, + 0.009170029312372208, + 0.014960432425141335, + 0.01278410293161869, + 0.00811379961669445, + 0.010307395830750465, + 0.017860047519207, + 0.009708023630082607, + 0.02168658748269081, + 0.030325807631015778, + 0.02529751881957054, + 0.02465110272169113, + 0.027526548132300377, + 0.033746588975191116, + 0.03644757717847824, + 0.027936730533838272, + 0.030896907672286034, + 0.04011038318276405, + 0.03906720504164696, + 0.035032112151384354, + 0.0402226448059082, + 0.02872307598590851, + 0.035759855061769485, + 0.03363380208611488, + 0.004690750036388636, + 0.00454807560890913, + 0.009005660191178322, + 0.012385365553200245, + 0.0046284799464046955, + 0.0076741105876863, + 0.015296101570129395, + 0.009893776848912239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.5761763453483582 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering soap bubble with iridescent patterns of mauve, cyan, and liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet, sunlit room with large windows and a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a cinematic close-up featuring a remarkably shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a steady upward tilt to follow the bubble rising from a vent in the black metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bubble vanish instantly upon contact with the floor, leaving only a tiny shimmering wet ring?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043052323162555695 + ], + "opposite_logprob_tokens": [ + -3.1680524349212646 + ], + "logprob_reward": -0.043052323162555695, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043052323162555695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038987621665000916, + 0.04127723351120949, + 0.08276902139186859, + 0.01816241256892681, + 0.021849680691957474, + 0.07769658416509628, + 0.3765295147895813, + 0.1785614937543869, + 0.06574258208274841, + 0.09842383116483688 + ], + "credit_map": [ + 0.0040946840308606625, + 0.0038178819231688976, + 0.007122835610061884, + 0.005565468221902847, + 0.00452310498803854, + 0.004079031758010387, + 0.00489284610375762, + 0.004891769029200077, + 0.004779726732522249, + 0.0037113509606570005, + 0.004097032826393843, + 0.005342438351362944, + 0.008227623999118805, + 0.0035106001887470484, + 0.006476363632827997, + 0.005132097750902176, + 0.010047489777207375, + 0.004447959829121828, + 0.024113209918141365, + 0.016350969672203064, + 0.005019919481128454, + 0.00487548066303134, + 0.00940963625907898, + 0.00850436557084322, + 0.002003907226026058, + 0.0017205281183123589, + 0.0021001712884753942, + 0.00279285223223269, + 0.0026084943674504757, + 0.0018723851535469294, + 0.0023914037737995386, + 0.0026726704090833664, + 0.002636529039591551, + 0.002579089254140854, + 0.0029704265762120485, + 0.003665814409032464, + 0.0024574813432991505, + 0.0022690712939947844, + 0.002401622710749507, + 0.0028696460649371147, + 0.006499359384179115, + 0.0065361312590539455, + 0.006605190224945545, + 0.00910093542188406, + 0.01581159047782421, + 0.011823522858321667, + 0.012393394485116005, + 0.008926458656787872, + 0.029626015573740005, + 0.025958111509680748, + 0.05425466597080231, + 0.04748745635151863, + 0.07022468745708466, + 0.03881144896149635, + 0.07346353679895401, + 0.03670358285307884, + 0.015070335008203983, + 0.023798787966370583, + 0.024369673803448677, + 0.01955421455204487, + 0.01714242808520794, + 0.022515393793582916, + 0.028643188998103142, + 0.02746746689081192, + 0.006958842743188143, + 0.0062767514027655125, + 0.00863905530422926, + 0.00908964779227972, + 0.009269779548048973, + 0.006476364564150572, + 0.012129268608987331, + 0.0069028763100504875, + 0.007485073991119862, + 0.006582808215171099, + 0.011735213920474052, + 0.01620102860033512, + 0.00678141787648201, + 0.010226988233625889, + 0.02585296519100666, + 0.013558330945670605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.043052323162555695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_12/index_tensor([8922, 8922])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_12/index_tensor([8922, 8922])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..58820c699b5c9497920609974a554ec3a760067b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_12/index_tensor([8922, 8922])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, high-angle, static shot is fixed on a parcel of dry earth, resting on a flat patch of sunlight. The earth, a brittle substance composed of ochre and sienna, already bears a prominent, jagged crack. The surrounding ground is a mosaic of dried cracks, forming a complex network of polygonal patterns in the harsh, direct sunlight, casting sharp, black shadows. A faint, almost imperceptible vibration travels through the frame, causing the existing cracks in the earth to propagate rapidly and decisively; the cracks spider outward for a moment before the earth cleanly splits in two, releasing a cloud of thin dust. Pulled by gravity, the two newly separated halves settle with a last-minute shift, coming to rest a short distance apart, revealing the dark, shadowed interior of the resting area. The entire event unfolds in complete silence, creating an atmosphere of intense concentration that amplifies the profound fragility of the arid landscape.", + "seed": 20819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_12/index_8922_seed_20819_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11357816723175347, + "vlm_logprob": -0.11357816723175347, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.11357816723175347, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007298000156879425 + ], + "opposite_logprob_tokens": [ + -5.507297992706299 + ], + "logprob_reward": -0.007298000156879425, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007298000156879425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062713623046875, + 0.05687595158815384, + 0.08957070857286453, + 0.06203863397240639, + 0.07186567783355713, + 0.08525486290454865, + 0.028234994038939476, + 0.03584028035402298, + 0.13811062276363373, + 0.36949461698532104 + ], + "credit_map": [ + 0.0070154243148863316, + 0.005724778864532709, + 0.006921972148120403, + 0.0062281000427901745, + 0.010294110514223576, + 0.006151437759399414, + 0.008228790014982224, + 0.012149007990956306, + 0.011166205629706383, + 0.004153911489993334, + 0.006790628656744957, + 0.0027751103043556213, + 0.008514446206390858, + 0.004972528666257858, + 0.003981532528996468, + 0.014521585777401924, + 0.012145889922976494, + 0.011455199681222439, + 0.009604918770492077, + 0.01517218817025423, + 0.014387214556336403, + 0.008283764123916626, + 0.0077931880950927734, + 0.010728340595960617, + 0.0071720462292432785, + 0.00683427881449461, + 0.006707632448524237, + 0.00779668428003788, + 0.01029997505247593, + 0.009482587687671185, + 0.007628066465258598, + 0.006117359269410372, + 0.0105280177667737, + 0.010765265673398972, + 0.007878395728766918, + 0.007351035717874765, + 0.01614317111670971, + 0.006198136135935783, + 0.007790462113916874, + 0.00521119823679328, + 0.008677884005010128, + 0.020588384941220284, + 0.0071613420732319355, + 0.009897490963339806, + 0.011474138125777245, + 0.009496072307229042, + 0.009390519931912422, + 0.008569029159843922, + 0.004275789484381676, + 0.0029687308706343174, + 0.0030030112247914076, + 0.0025262711569666862, + 0.004800822120159864, + 0.005023863632231951, + 0.0029451334848999977, + 0.0026913736946880817, + 0.007685494143515825, + 0.005921501200646162, + 0.006305614486336708, + 0.0018787164008244872, + 0.004520217422395945, + 0.005426973104476929, + 0.0019036598969250917, + 0.002198106376454234, + 0.032672904431819916, + 0.016542993485927582, + 0.012265174649655819, + 0.007780823390930891, + 0.034318722784519196, + 0.018867434933781624, + 0.0072457375936210155, + 0.008416843600571156, + 0.055737171322107315, + 0.03822225332260132, + 0.017262209206819534, + 0.02391650713980198, + 0.03459971770644188, + 0.09879498183727264, + 0.02449144795536995, + 0.07647032290697098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.007298000156879425 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008677395060658455 + ], + "opposite_logprob_tokens": [ + -4.7586774826049805 + ], + "logprob_reward": -0.008677395060658455, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008677395060658455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.186349019408226, + 0.08230630308389664, + 0.06625976413488388, + 0.016388725489377975, + 0.032774023711681366, + 0.02429221011698246, + 0.040681660175323486, + 0.03560563176870346, + 0.3285517394542694, + 0.18679095804691315 + ], + "credit_map": [ + 0.018654771149158478, + 0.015014209784567356, + 0.021062176674604416, + 0.030271178111433983, + 0.022365039214491844, + 0.021012313663959503, + 0.03696811571717262, + 0.021001216024160385, + 0.014677602797746658, + 0.010224075987935066, + 0.008959038183093071, + 0.0055063446052372456, + 0.009892183355987072, + 0.01333568710833788, + 0.006973781157284975, + 0.012737590819597244, + 0.008241584524512291, + 0.008168783970177174, + 0.007255811709910631, + 0.011994525790214539, + 0.010620818473398685, + 0.0053573353216052055, + 0.006052778568118811, + 0.00856812298297882, + 0.0019732117652893066, + 0.0018457402475178242, + 0.0022447705268859863, + 0.001619798829779029, + 0.0022729886695742607, + 0.0024703368544578552, + 0.0019103324739262462, + 0.002051546238362789, + 0.00572547921910882, + 0.003772842464968562, + 0.003493847558274865, + 0.0022603331599384546, + 0.009389049373567104, + 0.003373420564457774, + 0.0025113492738455534, + 0.002247704192996025, + 0.0029152429196983576, + 0.0035634804517030716, + 0.002672661328688264, + 0.0032167586032301188, + 0.002780364127829671, + 0.0029825579840689898, + 0.0031762574799358845, + 0.002984886523336172, + 0.006996532436460257, + 0.005618267692625523, + 0.006257268600165844, + 0.003450753865763545, + 0.004692831076681614, + 0.006482055876404047, + 0.0038403701037168503, + 0.0033435793593525887, + 0.007740228436887264, + 0.005371925886720419, + 0.007105187978595495, + 0.0022057287860661745, + 0.0029618251137435436, + 0.006473497021943331, + 0.001524418476037681, + 0.002222821582108736, + 0.10296013951301575, + 0.016226662322878838, + 0.011548163369297981, + 0.012670768424868584, + 0.13131219148635864, + 0.026406724005937576, + 0.012869956903159618, + 0.014557115733623505, + 0.027116889134049416, + 0.02200152352452278, + 0.010121180675923824, + 0.012646734714508057, + 0.019223935902118683, + 0.04606404900550842, + 0.01294217724353075, + 0.03667446970939636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.008677395060658455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636326290667057 + ], + "opposite_logprob_tokens": [ + -3.6513633728027344 + ], + "logprob_reward": -0.02636326290667057, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636326290667057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05101301148533821, + 0.03781893476843834, + 0.06202763691544533, + 0.019987966865301132, + 0.0666365996003151, + 0.06077275052666664, + 0.05263591185212135, + 0.024330459535121918, + 0.49731627106666565, + 0.12746043503284454 + ], + "credit_map": [ + 0.004942538216710091, + 0.005484155844897032, + 0.005382619332522154, + 0.007999003864824772, + 0.005827381741255522, + 0.008816306479275227, + 0.0059049311093986034, + 0.006656077690422535, + 0.0065336874686181545, + 0.0042701298370957375, + 0.004517022520303726, + 0.002650628099218011, + 0.004881212953478098, + 0.0032109234016388655, + 0.002945530228316784, + 0.008809799328446388, + 0.008304283022880554, + 0.005680754780769348, + 0.006102170795202255, + 0.0070736706256866455, + 0.009244503453373909, + 0.007839180529117584, + 0.006363257300108671, + 0.011419817805290222, + 0.002135928487405181, + 0.0025988114066421986, + 0.0017957179807126522, + 0.0021788596641272306, + 0.0031187476124614477, + 0.0028108512051403522, + 0.0032069003209471703, + 0.002142150653526187, + 0.007677393965423107, + 0.008795724250376225, + 0.008018954657018185, + 0.009756471961736679, + 0.00802722480148077, + 0.00992911122739315, + 0.00738551327958703, + 0.007046206388622522, + 0.007882640697062016, + 0.013671415857970715, + 0.00410492392256856, + 0.005587979219853878, + 0.009198358282446861, + 0.005707687698304653, + 0.009122666902840137, + 0.0054970732890069485, + 0.008878479711711407, + 0.008071256801486015, + 0.005057597998529673, + 0.004524528048932552, + 0.00727869663387537, + 0.008112994953989983, + 0.006399285513907671, + 0.004313073121011257, + 0.0031623374670743942, + 0.003088097320869565, + 0.0028040450997650623, + 0.0034334303345531225, + 0.0025239577516913414, + 0.0036340418737381697, + 0.003361258888617158, + 0.0023232917301356792, + 0.15874463319778442, + 0.026481933891773224, + 0.01884385757148266, + 0.019960280507802963, + 0.18930234014987946, + 0.04321293160319328, + 0.019553735852241516, + 0.02121657133102417, + 0.01900741085410118, + 0.014387545175850391, + 0.006416182033717632, + 0.008168220520019531, + 0.011337785981595516, + 0.03412333503365517, + 0.008544388227164745, + 0.025475576519966125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.02636326290667057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198899397626519 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198899397626519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198899397626519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07567770034074783, + 0.018611308187246323, + 0.04644808918237686, + 0.027290094643831253, + 0.05283033102750778, + 0.07645761966705322, + 0.08402719348669052, + 0.04802357032895088, + 0.0892811268568039, + 0.4813530147075653 + ], + "credit_map": [ + 0.007501530461013317, + 0.00674872612580657, + 0.008848308585584164, + 0.01332840509712696, + 0.008487746119499207, + 0.009448514319956303, + 0.012035466730594635, + 0.009278998710215092, + 0.002109171124175191, + 0.003554609138518572, + 0.0018158832099288702, + 0.0018297485075891018, + 0.0032710114028304815, + 0.002500917762517929, + 0.001999014290049672, + 0.0015309520531445742, + 0.007102121599018574, + 0.00418649660423398, + 0.004188716411590576, + 0.004407454282045364, + 0.006828970741480589, + 0.005090415012091398, + 0.004919434431940317, + 0.009724474512040615, + 0.0030591916292905807, + 0.0055309864692389965, + 0.0027731596492230892, + 0.002818565582856536, + 0.004739596974104643, + 0.002836196217685938, + 0.00305492477491498, + 0.0024774684570729733, + 0.00885965395718813, + 0.007449877914041281, + 0.0037165768444538116, + 0.003371899714693427, + 0.014351232908666134, + 0.005924154072999954, + 0.005940708331763744, + 0.00321621703915298, + 0.010375367477536201, + 0.01058515626937151, + 0.008018393069505692, + 0.008217308670282364, + 0.008430318906903267, + 0.009655032306909561, + 0.009912536479532719, + 0.011263500899076462, + 0.01437949389219284, + 0.012126456014811993, + 0.00858315546065569, + 0.007126732263714075, + 0.012480667792260647, + 0.01322106271982193, + 0.009097599424421787, + 0.007012016139924526, + 0.006838604807853699, + 0.006986143067479134, + 0.0071581569500267506, + 0.004549017176032066, + 0.005971858277916908, + 0.006976662203669548, + 0.004871865268796682, + 0.004671256057918072, + 0.014512885361909866, + 0.01088316086679697, + 0.00849036406725645, + 0.005437546409666538, + 0.014167483896017075, + 0.020755471661686897, + 0.00596867548301816, + 0.009065525606274605, + 0.0726526752114296, + 0.06318751722574234, + 0.024110177531838417, + 0.03178126737475395, + 0.04762071743607521, + 0.12452466785907745, + 0.03190361335873604, + 0.08557232469320297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.002198899397626519 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233532786369324 + ], + "opposite_logprob_tokens": [ + -0.8983532786369324 + ], + "logprob_reward": -0.5233532786369324, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233532786369324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051349788904190063, + 0.027081886306405067, + 0.03456220030784607, + 0.020142937079072, + 0.04742197319865227, + 0.0765114352107048, + 0.026474112644791603, + 0.063131183385849, + 0.5902581214904785, + 0.06306640058755875 + ], + "credit_map": [ + 0.005725004710257053, + 0.004777264315634966, + 0.007335733622312546, + 0.00744682177901268, + 0.005752975586801767, + 0.005279960576444864, + 0.009785274975001812, + 0.005246753804385662, + 0.0030367921572178602, + 0.004337601829320192, + 0.0031289937905967236, + 0.0025285123847424984, + 0.004292542580515146, + 0.003920627757906914, + 0.0036168540827929974, + 0.0022199612576514482, + 0.003877851413562894, + 0.0036244168877601624, + 0.003777001053094864, + 0.006436365656554699, + 0.005318725481629372, + 0.0032800056505948305, + 0.003803387051448226, + 0.004444445017725229, + 0.003484658896923065, + 0.0023875318001955748, + 0.0017449143342673779, + 0.002997249597683549, + 0.0029845992103219032, + 0.0017824554815888405, + 0.002251994563266635, + 0.002509534824639559, + 0.006537039298564196, + 0.0061080870218575, + 0.005135408602654934, + 0.004358262754976749, + 0.010933579877018929, + 0.005324178840965033, + 0.005780206993222237, + 0.0032452105078846216, + 0.009090584702789783, + 0.02079703100025654, + 0.005375965498387814, + 0.007487439084798098, + 0.011452444829046726, + 0.00856325589120388, + 0.00842826534062624, + 0.005316451657563448, + 0.004793514031916857, + 0.003882240504026413, + 0.002167206024751067, + 0.002755281515419483, + 0.0037153949961066246, + 0.0036921033170074224, + 0.0030961402226239443, + 0.0023722329642623663, + 0.010909786447882652, + 0.009835938923060894, + 0.011663571931421757, + 0.00513563072308898, + 0.00775611586868763, + 0.008554767817258835, + 0.004504688084125519, + 0.004770676605403423, + 0.18560385704040527, + 0.027921847999095917, + 0.01893072947859764, + 0.022770516574382782, + 0.23748844861984253, + 0.047922879457473755, + 0.023830438032746315, + 0.025789372622966766, + 0.009731748141348362, + 0.006667701527476311, + 0.0030483787413686514, + 0.004130223300307989, + 0.005841801408678293, + 0.017254242673516273, + 0.003928261809051037, + 0.012464040890336037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.5233532786369324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle, static shot is fixed on a parcel of dry earth, resting on a flat patch of sunlight. The earth, a brittle substance composed of ochre and sienna, already bears a prominent, jagged crack. The surrounding ground is a mosaic of dried cracks, forming a complex network of polygonal patterns in the harsh, direct sunlight, casting sharp, black shadows. A faint, almost imperceptible vibration travels through the frame, causing the existing cracks in the earth to propagate rapidly and decisively; the cracks spider outward for a moment before the earth cleanly splits in two, releasing a cloud of thin dust. Pulled by gravity, the two newly separated halves settle with a last-minute shift, coming to rest a short distance apart, revealing the dark, shadowed interior of the resting area. The entire event unfolds in complete silence, creating an atmosphere of intense concentration that amplifies the profound fragility of the arid landscape.", + "seed": 20820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_12/index_8922_seed_20820_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12195708230137825, + "vlm_logprob": -0.12195708230137825, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.12195708230137825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00400099391117692 + ], + "opposite_logprob_tokens": [ + -6.754001140594482 + ], + "logprob_reward": -0.00400099391117692, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.00400099391117692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0770607441663742, + 0.01376771554350853, + 0.057056356221437454, + 0.05643920227885246, + 0.1984764188528061, + 0.19703838229179382, + 0.018144825473427773, + 0.033099688589572906, + 0.2258915901184082, + 0.12302510440349579 + ], + "credit_map": [ + 0.006925130728632212, + 0.011756833642721176, + 0.007243258412927389, + 0.0076273782178759575, + 0.009246356785297394, + 0.010625873692333698, + 0.017408236861228943, + 0.00622766837477684, + 0.0018797378288581967, + 0.001777262194082141, + 0.0011800588108599186, + 0.0012739583617076278, + 0.002448768587782979, + 0.0017759102629497647, + 0.002301436383277178, + 0.0011305822990834713, + 0.005402457434684038, + 0.007097360212355852, + 0.006659668870270252, + 0.00829019583761692, + 0.01060456968843937, + 0.008937505073845387, + 0.006047108210623264, + 0.004017486236989498, + 0.013296433724462986, + 0.013537337072193623, + 0.0045776995830237865, + 0.004437904804944992, + 0.005876568146049976, + 0.005392180755734444, + 0.005924378987401724, + 0.0033966987393796444, + 0.0254247784614563, + 0.018667979165911674, + 0.015351039357483387, + 0.01714414544403553, + 0.03058265522122383, + 0.03718305751681328, + 0.03458831086754799, + 0.01953442394733429, + 0.02838117629289627, + 0.019779888913035393, + 0.022422119975090027, + 0.017708666622638702, + 0.04437725618481636, + 0.03644488751888275, + 0.01799214631319046, + 0.00993224699050188, + 0.001919646980240941, + 0.003469723742455244, + 0.0022100943606346846, + 0.002202757401391864, + 0.0029947676230221987, + 0.001810489920899272, + 0.0019741852302104235, + 0.0015631580026820302, + 0.0036380556412041187, + 0.005491156131029129, + 0.002702438971027732, + 0.0028798035345971584, + 0.006116944365203381, + 0.004482486750930548, + 0.004542796406894922, + 0.0032460058573633432, + 0.02900140918791294, + 0.032952550798654556, + 0.015800803899765015, + 0.019258202984929085, + 0.06520869582891464, + 0.02377478778362274, + 0.01806039921939373, + 0.02183471992611885, + 0.030464837327599525, + 0.014979714527726173, + 0.010681692510843277, + 0.009717043489217758, + 0.016815878450870514, + 0.020945394411683083, + 0.009901088662445545, + 0.009519441053271294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.00400099391117692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004668171051889658 + ], + "opposite_logprob_tokens": [ + -5.379668235778809 + ], + "logprob_reward": -0.004668171051889658, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004668171051889658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08293340355157852, + 0.1013268455862999, + 0.050109777599573135, + 0.1703236997127533, + 0.04743625968694687, + 0.19959713518619537, + 0.028150497004389763, + 0.041099391877651215, + 0.22086511552333832, + 0.05815790221095085 + ], + "credit_map": [ + 0.011935102753341198, + 0.009628329426050186, + 0.00844823569059372, + 0.011140518821775913, + 0.007910830900073051, + 0.009899040684103966, + 0.015794949606060982, + 0.008176390081644058, + 0.011353913694620132, + 0.022516611963510513, + 0.008309299126267433, + 0.009163541719317436, + 0.015643509104847908, + 0.015717048197984695, + 0.01123792864382267, + 0.007384980097413063, + 0.004460999276489019, + 0.006286096293479204, + 0.006402392406016588, + 0.007751119788736105, + 0.008562883362174034, + 0.007408779580146074, + 0.005355330649763346, + 0.003882171818986535, + 0.03686818853020668, + 0.06334298849105835, + 0.007351371459662914, + 0.008517211303114891, + 0.0129658542573452, + 0.015300743281841278, + 0.01687215454876423, + 0.009105173870921135, + 0.005843809340149164, + 0.006006763316690922, + 0.006076692137867212, + 0.0047994754277169704, + 0.007061478216201067, + 0.006289761513471603, + 0.005473283119499683, + 0.005884993355721235, + 0.036545995622873306, + 0.019127648323774338, + 0.020527854561805725, + 0.016249125823378563, + 0.05009728670120239, + 0.029102087020874023, + 0.01765756867825985, + 0.010289554484188557, + 0.002792220562696457, + 0.004105277359485626, + 0.0029578653629869223, + 0.00321120023727417, + 0.00535658560693264, + 0.0027256144676357508, + 0.003902735421434045, + 0.0030989940278232098, + 0.00576379569247365, + 0.00626086862757802, + 0.003192235017195344, + 0.003621394280344248, + 0.011294330470263958, + 0.004696656949818134, + 0.0032642849255353212, + 0.0030058196280151606, + 0.027815287932753563, + 0.031058331951498985, + 0.015870781615376472, + 0.01799234002828598, + 0.06428690254688263, + 0.024519799277186394, + 0.01722201332449913, + 0.022099632769823074, + 0.0094464011490345, + 0.005638994742184877, + 0.004635212477296591, + 0.003603330347687006, + 0.004755753558129072, + 0.018151437863707542, + 0.004312685690820217, + 0.007614076137542725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.004668171051889658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201936036348343 + ], + "opposite_logprob_tokens": [ + -4.01820182800293 + ], + "logprob_reward": -0.018201936036348343, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201936036348343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05234457924962044, + 0.06512647122144699, + 0.06481193006038666, + 0.26240891218185425, + 0.3272387981414795, + 0.08934051543474197, + 0.029016414657235146, + 0.04724859073758125, + 0.026147155091166496, + 0.036316659301519394 + ], + "credit_map": [ + 0.006409395486116409, + 0.011507002636790276, + 0.005466141737997532, + 0.005104666575789452, + 0.005695100408047438, + 0.007082455791532993, + 0.005532160867005587, + 0.00554765434935689, + 0.008106713183224201, + 0.015193236991763115, + 0.00561217125505209, + 0.005959803704172373, + 0.009984694421291351, + 0.006861926522105932, + 0.008914166130125523, + 0.00449376180768013, + 0.007163606118410826, + 0.009161580353975296, + 0.007198304869234562, + 0.011524584144353867, + 0.007134349085390568, + 0.00599298719316721, + 0.01092124916613102, + 0.005715268198400736, + 0.06965308636426926, + 0.08716125786304474, + 0.016060123220086098, + 0.01463728491216898, + 0.021700425073504448, + 0.01971948705613613, + 0.019121810793876648, + 0.01435541920363903, + 0.04207761585712433, + 0.03604556992650032, + 0.027813376858830452, + 0.030016034841537476, + 0.046604376286268234, + 0.04605097323656082, + 0.05403018370270729, + 0.044600650668144226, + 0.012314967811107635, + 0.010677567683160305, + 0.010136870667338371, + 0.004894543904811144, + 0.01886705495417118, + 0.019401682540774345, + 0.006289605051279068, + 0.006758207455277443, + 0.002676905132830143, + 0.00625744741410017, + 0.0024952825624495745, + 0.0024580310564488173, + 0.00486534321680665, + 0.0036665804218500853, + 0.003753425320610404, + 0.002843395806849003, + 0.008837322704494, + 0.0061413743533194065, + 0.003086753422394395, + 0.0033238655887544155, + 0.01405420247465372, + 0.004957806318998337, + 0.0034466227516531944, + 0.0034006412606686354, + 0.002827864373102784, + 0.0037775582168251276, + 0.0022196522913873196, + 0.0021715201437473297, + 0.006696104537695646, + 0.003211908508092165, + 0.002971099689602852, + 0.0022714450024068356, + 0.008050195872783661, + 0.0040079629980027676, + 0.0035586163867264986, + 0.002427369123324752, + 0.003462297609075904, + 0.005935448221862316, + 0.0037918949965387583, + 0.005082870367914438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.018201936036348343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067379651591181755 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.0067379651591181755, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067379651591181755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04580288380384445, + 0.0400085523724556, + 0.11427783966064453, + 0.24077120423316956, + 0.313255250453949, + 0.12521035969257355, + 0.02455367147922516, + 0.04117991402745247, + 0.034156426787376404, + 0.020783910527825356 + ], + "credit_map": [ + 0.006108848378062248, + 0.007543210405856371, + 0.0047797514125704765, + 0.0035439101047813892, + 0.005180490668863058, + 0.007800935301929712, + 0.007633069064468145, + 0.0032126735895872116, + 0.004961983300745487, + 0.00869651697576046, + 0.0037395397666841745, + 0.0031920436304062605, + 0.00600161449983716, + 0.003883928759023547, + 0.006657595746219158, + 0.002875333884730935, + 0.013232686556875706, + 0.02333233319222927, + 0.01125531829893589, + 0.015934158116579056, + 0.017467409372329712, + 0.01155800186097622, + 0.01445639505982399, + 0.007041535805910826, + 0.06013008952140808, + 0.039038579910993576, + 0.020106058567762375, + 0.025156652554869652, + 0.03400040417909622, + 0.026150472462177277, + 0.02540593221783638, + 0.01078302413225174, + 0.042856164276599884, + 0.03241831809282303, + 0.02399905025959015, + 0.029437918215990067, + 0.039507366716861725, + 0.054987575858831406, + 0.055559854954481125, + 0.03448902815580368, + 0.018022559583187103, + 0.01383516751229763, + 0.014866826124489307, + 0.007828726433217525, + 0.025533774867653847, + 0.025999723002314568, + 0.010351675562560558, + 0.008771910332143307, + 0.0025658910162746906, + 0.0033503822050988674, + 0.0022115863393992186, + 0.002799428766593337, + 0.0030414739158004522, + 0.0029669711366295815, + 0.004817845765501261, + 0.002800095360726118, + 0.005108359269797802, + 0.006112588569521904, + 0.0037085474468767643, + 0.004011671990156174, + 0.006506850942969322, + 0.007128922268748283, + 0.004693410359323025, + 0.003909567836672068, + 0.004643656313419342, + 0.005308154970407486, + 0.0029194909147918224, + 0.0029309927485883236, + 0.005918098147958517, + 0.00632624514400959, + 0.0035012681037187576, + 0.0026085213758051395, + 0.003925519064068794, + 0.0025776331312954426, + 0.0018237567273899913, + 0.0018623725045472383, + 0.0021244175732135773, + 0.0038656871765851974, + 0.002284621587023139, + 0.0023199052084237337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.0067379651591181755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761763453483582 + ], + "opposite_logprob_tokens": [ + -0.8261763453483582 + ], + "logprob_reward": -0.5761763453483582, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761763453483582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053497765213251114, + 0.028013648465275764, + 0.11909853667020798, + 0.3495960533618927, + 0.21369951963424683, + 0.10405538231134415, + 0.025902563706040382, + 0.01580481044948101, + 0.03125857934355736, + 0.05907318368554115 + ], + "credit_map": [ + 0.009154831990599632, + 0.008296127431094646, + 0.004847807809710503, + 0.0044677359983325005, + 0.008462370373308659, + 0.006314289756119251, + 0.0077212355099618435, + 0.004233366809785366, + 0.004260581452399492, + 0.005227210931479931, + 0.003046173369511962, + 0.002172048669308424, + 0.0036900027189403772, + 0.003202602732926607, + 0.004606391303241253, + 0.0018086390336975455, + 0.011914417147636414, + 0.02061009407043457, + 0.011484281159937382, + 0.01795787177979946, + 0.018624845892190933, + 0.011251281015574932, + 0.01836255006492138, + 0.008893189951777458, + 0.08266136050224304, + 0.09414119273424149, + 0.018756264820694923, + 0.027382031083106995, + 0.02404165267944336, + 0.027941470965743065, + 0.05854678526520729, + 0.016125304624438286, + 0.031212562695145607, + 0.022147666662931442, + 0.018120232969522476, + 0.017898842692375183, + 0.02881200611591339, + 0.035210851579904556, + 0.032279349863529205, + 0.02801799215376377, + 0.027694620192050934, + 0.009687844663858414, + 0.009660206735134125, + 0.00709754740819335, + 0.02228672057390213, + 0.012314887717366219, + 0.008333447389304638, + 0.006980111822485924, + 0.0021300590597093105, + 0.00431052315980196, + 0.002590833930298686, + 0.0024968753568828106, + 0.0037961641792207956, + 0.0027725519612431526, + 0.004925420042127371, + 0.0028801350854337215, + 0.0023707440122962, + 0.0026484213303774595, + 0.0011800049105659127, + 0.0013623526319861412, + 0.002728331135585904, + 0.0017843457171693444, + 0.002144911792129278, + 0.0015856986865401268, + 0.005753867328166962, + 0.004657227545976639, + 0.002516811480745673, + 0.002459258073940873, + 0.004067384172230959, + 0.003861811477690935, + 0.003818598808720708, + 0.004123619291931391, + 0.0157658439129591, + 0.005976208485662937, + 0.006206093356013298, + 0.00426102289929986, + 0.005464242305606604, + 0.0071137212216854095, + 0.005862376652657986, + 0.008423671126365662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.5761763453483582 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle, static shot is fixed on a parcel of dry earth, resting on a flat patch of sunlight. The earth, a brittle substance composed of ochre and sienna, already bears a prominent, jagged crack. The surrounding ground is a mosaic of dried cracks, forming a complex network of polygonal patterns in the harsh, direct sunlight, casting sharp, black shadows. A faint, almost imperceptible vibration travels through the frame, causing the existing cracks in the earth to propagate rapidly and decisively; the cracks spider outward for a moment before the earth cleanly splits in two, releasing a cloud of thin dust. Pulled by gravity, the two newly separated halves settle with a last-minute shift, coming to rest a short distance apart, revealing the dark, shadowed interior of the resting area. The entire event unfolds in complete silence, creating an atmosphere of intense concentration that amplifies the profound fragility of the arid landscape.", + "seed": 20821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_12/index_8922_seed_20821_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13565588844940066, + "vlm_logprob": -0.13565588844940066, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.13565588844940066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004666865803301334 + ], + "opposite_logprob_tokens": [ + -6.879666805267334 + ], + "logprob_reward": -0.004666865803301334, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.004666865803301334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07463232427835464, + 0.49608051776885986, + 0.0392717570066452, + 0.06126471981406212, + 0.023123621940612793, + 0.06563466042280197, + 0.14467252790927887, + 0.03358880430459976, + 0.02597707137465477, + 0.035753991454839706 + ], + "credit_map": [ + 0.009922709316015244, + 0.007345403544604778, + 0.009098111651837826, + 0.011992868967354298, + 0.006685084663331509, + 0.014727568253874779, + 0.0069787162356078625, + 0.007881859317421913, + 0.17034108936786652, + 0.2007938176393509, + 0.01761235110461712, + 0.014432115480303764, + 0.0298224575817585, + 0.031497228890657425, + 0.016606882214546204, + 0.0149745624512434, + 0.007397372741252184, + 0.005532498471438885, + 0.00314745563082397, + 0.002580845495685935, + 0.005621683783829212, + 0.006784395780414343, + 0.0035090839955955744, + 0.0046984232030808926, + 0.015371983870863914, + 0.010457070544362068, + 0.003721756162121892, + 0.0028178109787404537, + 0.003918695729225874, + 0.006202800665050745, + 0.004130241461098194, + 0.014644361101090908, + 0.004961218684911728, + 0.003455517115071416, + 0.0023397235199809074, + 0.0018567348597571254, + 0.0032038437202572823, + 0.0032866813708096743, + 0.0019848993979394436, + 0.0020350031554698944, + 0.014202133752405643, + 0.015432815998792648, + 0.00503884069621563, + 0.004142181482166052, + 0.011836392804980278, + 0.004916768055409193, + 0.004500655457377434, + 0.005564873106777668, + 0.05035727471113205, + 0.058992207050323486, + 0.0055242665112018585, + 0.005270285066217184, + 0.008195312693715096, + 0.005687989294528961, + 0.005172970704734325, + 0.005472214892506599, + 0.004712691996246576, + 0.005208935122936964, + 0.00271120248362422, + 0.002401797566562891, + 0.004748293664306402, + 0.004653113428503275, + 0.0032196829561144114, + 0.005933087319135666, + 0.0033784674014896154, + 0.0045575955882668495, + 0.0019782984163612127, + 0.00307145481929183, + 0.004991571418941021, + 0.0031858542934060097, + 0.0028678898233920336, + 0.0019459374016150832, + 0.005053242202848196, + 0.00433563394472003, + 0.0025800010189414024, + 0.002770442282781005, + 0.0034541073255240917, + 0.004947981331497431, + 0.004308756906539202, + 0.008303828537464142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.004666865803301334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009830396622419357 + ], + "opposite_logprob_tokens": [ + -4.634830474853516 + ], + "logprob_reward": -0.009830396622419357, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009830396622419357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.146124467253685, + 0.22822540998458862, + 0.10381368547677994, + 0.057940516620874405, + 0.03522526100277901, + 0.11458149552345276, + 0.09726174920797348, + 0.0820726752281189, + 0.06512930244207382, + 0.06962541490793228 + ], + "credit_map": [ + 0.017560863867402077, + 0.02392253465950489, + 0.012032100930809975, + 0.02265075035393238, + 0.0166550874710083, + 0.021791821345686913, + 0.01518572773784399, + 0.016325583681464195, + 0.07448989897966385, + 0.08595677465200424, + 0.009404665790498257, + 0.012088392861187458, + 0.015684984624385834, + 0.014944905415177345, + 0.00874602235853672, + 0.006909757386893034, + 0.019852211698889732, + 0.014681223779916763, + 0.008066893555223942, + 0.009738867171108723, + 0.014849793165922165, + 0.012276857160031796, + 0.013818969950079918, + 0.010528874583542347, + 0.014085644856095314, + 0.010038599371910095, + 0.0034183007664978504, + 0.003852435853332281, + 0.004357927944511175, + 0.006839227396994829, + 0.0037250753957778215, + 0.011623302474617958, + 0.008614156395196915, + 0.00575863104313612, + 0.0021522792521864176, + 0.0020601225551217794, + 0.0028838119469583035, + 0.0028719406109303236, + 0.003109204350039363, + 0.007775113917887211, + 0.023380430415272713, + 0.02758103981614113, + 0.005272942595183849, + 0.006206056568771601, + 0.012993126176297665, + 0.008545712567865849, + 0.007257398217916489, + 0.023344798013567924, + 0.02580920420587063, + 0.02747616544365883, + 0.006319896318018436, + 0.0052588870748877525, + 0.008036885410547256, + 0.007790022529661655, + 0.004699313081800938, + 0.011871377006173134, + 0.010627223178744316, + 0.014764996245503426, + 0.005651097744703293, + 0.008109097369015217, + 0.016744110733270645, + 0.011135255917906761, + 0.008782099932432175, + 0.006258794572204351, + 0.008168946020305157, + 0.011818686500191689, + 0.0051918188109993935, + 0.007520357612520456, + 0.011847454123198986, + 0.008391371928155422, + 0.005790950730443001, + 0.006399716250598431, + 0.009811751544475555, + 0.009577159769833088, + 0.005315207876265049, + 0.005316309165209532, + 0.007209283299744129, + 0.01124266441911459, + 0.008061598055064678, + 0.013091442175209522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.009830396622419357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023306243121623993 + ], + "opposite_logprob_tokens": [ + -3.773306131362915 + ], + "logprob_reward": -0.023306243121623993, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306243121623993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037772294133901596, + 0.24246655404567719, + 0.032240379601716995, + 0.04051081836223602, + 0.034100115299224854, + 0.21071575582027435, + 0.09270394593477249, + 0.14192496240139008, + 0.1348659098148346, + 0.032699279487133026 + ], + "credit_map": [ + 0.005881993565708399, + 0.004314109217375517, + 0.004405061714351177, + 0.006332028657197952, + 0.00420871889218688, + 0.0052603064104914665, + 0.0036066491156816483, + 0.0037634249310940504, + 0.08282968401908875, + 0.09756089746952057, + 0.00857690628618002, + 0.007144921459257603, + 0.014731073752045631, + 0.01649782620370388, + 0.007995670661330223, + 0.007129574194550514, + 0.007383666001260281, + 0.005817849189043045, + 0.0027180819306522608, + 0.002982968231663108, + 0.003566933097317815, + 0.004088046960532665, + 0.0028547903057187796, + 0.002828046679496765, + 0.010154255665838718, + 0.007887061685323715, + 0.0022008924279361963, + 0.0020874724723398685, + 0.0027704723179340363, + 0.005708068143576384, + 0.003354239510372281, + 0.006348357070237398, + 0.0063549247570335865, + 0.004048521164804697, + 0.0026200232096016407, + 0.003808006877079606, + 0.0047028386034071445, + 0.0027572778053581715, + 0.003849918721243739, + 0.005958601366728544, + 0.04377135634422302, + 0.05265441909432411, + 0.010745961219072342, + 0.014520148746669292, + 0.03955583646893501, + 0.01675860770046711, + 0.016188107430934906, + 0.016521327197551727, + 0.027310144156217575, + 0.03228464350104332, + 0.004077318590134382, + 0.0047910758294165134, + 0.007818426936864853, + 0.005807476583868265, + 0.0045775058679282665, + 0.006037354469299316, + 0.018987536430358887, + 0.025527194142341614, + 0.008579837158322334, + 0.012759082019329071, + 0.029506968334317207, + 0.018472466617822647, + 0.015426994301378727, + 0.012664892710745335, + 0.016828596591949463, + 0.024034300819039345, + 0.009583763778209686, + 0.015318715013563633, + 0.024205172434449196, + 0.016885312274098396, + 0.011939027346670628, + 0.016071012243628502, + 0.004426976200193167, + 0.005347986239939928, + 0.002804497256875038, + 0.0030144276097416878, + 0.003898024093359709, + 0.005219689570367336, + 0.0035140872932970524, + 0.0044735935516655445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.023306243121623993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632025051862001 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632025051862001, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632025051862001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0610506571829319, + 0.1953275203704834, + 0.018490934744477272, + 0.031120024621486664, + 0.04234464094042778, + 0.08427784591913223, + 0.0613853819668293, + 0.2127297967672348, + 0.19635823369026184, + 0.09691499173641205 + ], + "credit_map": [ + 0.010644947178661823, + 0.0067801098339259624, + 0.006595737766474485, + 0.006946199107915163, + 0.006254288367927074, + 0.008156836032867432, + 0.006943928077816963, + 0.008728609420359135, + 0.06491056084632874, + 0.07662414759397507, + 0.006839852314442396, + 0.0059426468797028065, + 0.013375934213399887, + 0.014576155692338943, + 0.0068042767234146595, + 0.006253945175558329, + 0.0041521773673594, + 0.003160102991387248, + 0.0014257367001846433, + 0.0018015098758041859, + 0.002368312096223235, + 0.002469089115038514, + 0.0014796342002227902, + 0.001634372165426612, + 0.007702023722231388, + 0.006150761619210243, + 0.001611695159226656, + 0.0016947120893746614, + 0.0021035661920905113, + 0.004077019635587931, + 0.0025496631860733032, + 0.0052305832505226135, + 0.007130944170057774, + 0.005440850742161274, + 0.0025617789942771196, + 0.0031011225655674934, + 0.004642052110284567, + 0.0031289062462747097, + 0.004751444328576326, + 0.011587537825107574, + 0.017067739740014076, + 0.017761869356036186, + 0.0059194364584982395, + 0.006064974702894688, + 0.016467701643705368, + 0.006440403405576944, + 0.006833811290562153, + 0.007721912115812302, + 0.013650113716721535, + 0.014224715530872345, + 0.004411548841744661, + 0.0048106457106769085, + 0.007465161848813295, + 0.004782605916261673, + 0.004605912137776613, + 0.007434682454913855, + 0.02747800014913082, + 0.036608923226594925, + 0.013226447626948357, + 0.01948656514286995, + 0.04361504688858986, + 0.026049189269542694, + 0.023489132523536682, + 0.02277650311589241, + 0.025633955374360085, + 0.032421015202999115, + 0.014017997309565544, + 0.023237090557813644, + 0.037098478525877, + 0.0232671070843935, + 0.017937464639544487, + 0.02274511009454727, + 0.012553785927593708, + 0.015919825062155724, + 0.007203633431345224, + 0.008923095650970936, + 0.011863700114190578, + 0.015306945890188217, + 0.011596045456826687, + 0.013547965325415134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.007632025051862001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328439116477966 + ], + "opposite_logprob_tokens": [ + -0.7578439116477966 + ], + "logprob_reward": -0.6328439116477966, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328439116477966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10588232427835464, + 0.2341395616531372, + 0.02841154672205448, + 0.056977443397045135, + 0.10382402688264847, + 0.08634507656097412, + 0.14630909264087677, + 0.12149333208799362, + 0.08146034926176071, + 0.03515723720192909 + ], + "credit_map": [ + 0.013896319083869457, + 0.015387301333248615, + 0.006227047182619572, + 0.015130054205656052, + 0.00957099162042141, + 0.01799951307475567, + 0.012035242281854153, + 0.015635859221220016, + 0.07880101352930069, + 0.09291300922632217, + 0.008835279382765293, + 0.00697086239233613, + 0.015180126763880253, + 0.015886006876826286, + 0.007652296684682369, + 0.007900968194007874, + 0.004421548452228308, + 0.0037483491469174623, + 0.0024778535589575768, + 0.003215381409972906, + 0.003599828574806452, + 0.004806839395314455, + 0.0029189870692789555, + 0.0032227581832557917, + 0.010458591394126415, + 0.007144824601709843, + 0.006915414240211248, + 0.004098416771739721, + 0.007049690466374159, + 0.006972676608711481, + 0.0069494121707975864, + 0.007388414815068245, + 0.025682587176561356, + 0.01713629998266697, + 0.005509030073881149, + 0.004931970499455929, + 0.009669686667621136, + 0.0069033559411764145, + 0.007020008284598589, + 0.02697109803557396, + 0.015553616918623447, + 0.015354998409748077, + 0.008813244290649891, + 0.006151057314127684, + 0.016888722777366638, + 0.007105303928256035, + 0.00692970072850585, + 0.009548429399728775, + 0.04919951781630516, + 0.05916200205683708, + 0.006540923845022917, + 0.004949858877807856, + 0.008117376826703548, + 0.00656494963914156, + 0.005121983587741852, + 0.00665248092263937, + 0.016072718426585197, + 0.022688452154397964, + 0.00777883967384696, + 0.010364016517996788, + 0.022736618295311928, + 0.016656069085001945, + 0.012812254019081593, + 0.012384366244077682, + 0.010345974937081337, + 0.01445242203772068, + 0.005118092522025108, + 0.009409187361598015, + 0.017843298614025116, + 0.009288850240409374, + 0.007790897507220507, + 0.007211622782051563, + 0.004332818556576967, + 0.005674118176102638, + 0.0026114957872778177, + 0.003308802144601941, + 0.0053278012201189995, + 0.005063680466264486, + 0.004550160840153694, + 0.004288360010832548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.6328439116477966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle, static shot is fixed on a parcel of dry earth, resting on a flat patch of sunlight. The earth, a brittle substance composed of ochre and sienna, already bears a prominent, jagged crack. The surrounding ground is a mosaic of dried cracks, forming a complex network of polygonal patterns in the harsh, direct sunlight, casting sharp, black shadows. A faint, almost imperceptible vibration travels through the frame, causing the existing cracks in the earth to propagate rapidly and decisively; the cracks spider outward for a moment before the earth cleanly splits in two, releasing a cloud of thin dust. Pulled by gravity, the two newly separated halves settle with a last-minute shift, coming to rest a short distance apart, revealing the dark, shadowed interior of the resting area. The entire event unfolds in complete silence, creating an atmosphere of intense concentration that amplifies the profound fragility of the arid landscape.", + "seed": 20822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_12/index_8922_seed_20822_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14375904675107448, + "vlm_logprob": -0.14375904675107448, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.14375904675107448, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005692580249160528 + ], + "opposite_logprob_tokens": [ + -6.005692481994629 + ], + "logprob_reward": -0.005692580249160528, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005692580249160528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0962352380156517, + 0.04943876340985298, + 0.24179595708847046, + 0.17281657457351685, + 0.06842074543237686, + 0.03753257542848587, + 0.031954649835824966, + 0.09488187730312347, + 0.17065495252609253, + 0.03626871481537819 + ], + "credit_map": [ + 0.009956663474440575, + 0.014807560481131077, + 0.012861075811088085, + 0.014494079165160656, + 0.010644392110407352, + 0.008563296869397163, + 0.014071444049477577, + 0.010836723260581493, + 0.008466944098472595, + 0.010438649915158749, + 0.007714354898780584, + 0.0038457494229078293, + 0.006027146242558956, + 0.005121988710016012, + 0.0031837327405810356, + 0.004640193656086922, + 0.016285309568047523, + 0.010234443470835686, + 0.0835573822259903, + 0.010559451766312122, + 0.08334827423095703, + 0.021053096279501915, + 0.007615074981004, + 0.0091429203748703, + 0.013645868748426437, + 0.015138315036892891, + 0.01689976267516613, + 0.016191594302654266, + 0.03693732991814613, + 0.03807473182678223, + 0.015361577272415161, + 0.020567376166582108, + 0.007757210172712803, + 0.003964761272072792, + 0.01922452077269554, + 0.004721669014543295, + 0.005756977945566177, + 0.019121702760457993, + 0.003002643119543791, + 0.004871253855526447, + 0.005847310647368431, + 0.0035447555128484964, + 0.006511316634714603, + 0.002635351615026593, + 0.0034105319064110518, + 0.009240594692528248, + 0.0030385253485292196, + 0.0033041846472769976, + 0.005048474762588739, + 0.003299958538264036, + 0.004874437116086483, + 0.003321426222100854, + 0.004037711303681135, + 0.004243341274559498, + 0.0034886428620666265, + 0.0036406584549695253, + 0.014166138134896755, + 0.010758524760603905, + 0.01015517022460699, + 0.009726438671350479, + 0.015853112563490868, + 0.012578590773046017, + 0.009234371595084667, + 0.012409521266818047, + 0.021634357050061226, + 0.06457804143428802, + 0.0051893615163862705, + 0.004100348334759474, + 0.00914095900952816, + 0.054818395525217056, + 0.004819185938686132, + 0.006374279037117958, + 0.00424125837162137, + 0.009228063747286797, + 0.0026448690332472324, + 0.0028050688561052084, + 0.0036344588734209538, + 0.008998595178127289, + 0.0018795704236254096, + 0.002836825791746378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.005692580249160528 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041281492449343204 + ], + "opposite_logprob_tokens": [ + -5.5041279792785645 + ], + "logprob_reward": -0.0041281492449343204, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041281492449343204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12507852911949158, + 0.0931900218129158, + 0.3572397828102112, + 0.03218987584114075, + 0.014718340709805489, + 0.03140067309141159, + 0.04664617031812668, + 0.0205618254840374, + 0.11757112294435501, + 0.16140364110469818 + ], + "credit_map": [ + 0.013322195038199425, + 0.023320667445659637, + 0.01809210516512394, + 0.016294226050376892, + 0.022715231403708458, + 0.010868865065276623, + 0.01067735068500042, + 0.009787879884243011, + 0.008644627407193184, + 0.018231647089123726, + 0.030284574255347252, + 0.007817291654646397, + 0.012936380691826344, + 0.006733875256031752, + 0.0038276046980172396, + 0.004714015871286392, + 0.029233774170279503, + 0.021023714914917946, + 0.1231338307261467, + 0.01744510978460312, + 0.11827684938907623, + 0.02693759649991989, + 0.011721992865204811, + 0.0094669284299016, + 0.002671449910849333, + 0.00394894415512681, + 0.0028539435006678104, + 0.002850322751328349, + 0.0071625723503530025, + 0.005691585596650839, + 0.003527724649757147, + 0.003483333857730031, + 0.0026949921157211065, + 0.0021624674554914236, + 0.0020326804369688034, + 0.001330378232523799, + 0.0017672956455498934, + 0.001689633703790605, + 0.00124122085981071, + 0.0017996728420257568, + 0.0035273381508886814, + 0.003325544996187091, + 0.004428542219102383, + 0.0029898176435381174, + 0.0036401760298758745, + 0.00538678839802742, + 0.00416765408590436, + 0.003934809938073158, + 0.006030983757227659, + 0.006543179973959923, + 0.007091998588293791, + 0.005057553760707378, + 0.006239716429263353, + 0.005104803945869207, + 0.0047150868922472, + 0.005862847901880741, + 0.0027919600252062082, + 0.0021691627334803343, + 0.002261532237753272, + 0.0020154607482254505, + 0.004309275187551975, + 0.002131539396941662, + 0.0024580347817391157, + 0.002424859208986163, + 0.015082130208611488, + 0.04469101130962372, + 0.003249255008995533, + 0.002780401613563299, + 0.005218375474214554, + 0.039931152015924454, + 0.002925517037510872, + 0.0036932840012013912, + 0.018412083387374878, + 0.05891525000333786, + 0.0052147964015603065, + 0.006631160154938698, + 0.006978416349738836, + 0.05623673275113106, + 0.003915088716894388, + 0.00510012311860919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0041281492449343204 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208580367267132 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208580367267132, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208580367267132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02771644853055477, + 0.02489830181002617, + 0.07065386325120926, + 0.1405421495437622, + 0.10298927873373032, + 0.0924697294831276, + 0.13133473694324493, + 0.15247447788715363, + 0.2127433568239212, + 0.04417763650417328 + ], + "credit_map": [ + 0.0026837028563022614, + 0.004303182475268841, + 0.004301862325519323, + 0.004040820989757776, + 0.003031742060557008, + 0.0029309350065886974, + 0.003120392793789506, + 0.0033038118854165077, + 0.003659740323200822, + 0.0022854784037917852, + 0.006210801191627979, + 0.0012981158215552568, + 0.0060306740924716, + 0.0020690071396529675, + 0.0011900716926902533, + 0.0021544164046645164, + 0.007089610211551189, + 0.004418277181684971, + 0.022028543055057526, + 0.0031495026778429747, + 0.02229275181889534, + 0.006574259139597416, + 0.0027371260803192854, + 0.002363790525123477, + 0.011743111535906792, + 0.01585523411631584, + 0.012077068910002708, + 0.013558055274188519, + 0.03992638364434242, + 0.024427490308880806, + 0.013818352483212948, + 0.009136464446783066, + 0.008436574600636959, + 0.006302583962678909, + 0.01253857184201479, + 0.005378777626901865, + 0.0238778218626976, + 0.0365384966135025, + 0.005008271429687738, + 0.00490818964317441, + 0.010581275448203087, + 0.011886844411492348, + 0.01254209689795971, + 0.008877789601683617, + 0.012559971772134304, + 0.017444148659706116, + 0.010149447247385979, + 0.00842816662043333, + 0.017169032245874405, + 0.019440073519945145, + 0.022956667467951775, + 0.011909817345440388, + 0.015397394075989723, + 0.015910068526864052, + 0.013917582109570503, + 0.014634122140705585, + 0.02131027728319168, + 0.01724022999405861, + 0.016589580103754997, + 0.015013658441603184, + 0.02618921734392643, + 0.01995130628347397, + 0.01584634743630886, + 0.020333873108029366, + 0.027026113122701645, + 0.07597602903842926, + 0.00606517493724823, + 0.005009850487112999, + 0.010356808081269264, + 0.07422157377004623, + 0.006984864827245474, + 0.007102953270077705, + 0.0044405036605894566, + 0.013655392453074455, + 0.002050258917734027, + 0.002481393050402403, + 0.005249843467026949, + 0.012389267794787884, + 0.0018977007130160928, + 0.00201327889226377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.014208580367267132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013380631571635604 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013380631571635604, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013380631571635604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03567136079072952, + 0.02675911784172058, + 0.061210162937641144, + 0.09415692836046219, + 0.03358594328165054, + 0.13088485598564148, + 0.02774304710328579, + 0.06886875629425049, + 0.1825016587972641, + 0.3386181592941284 + ], + "credit_map": [ + 0.0036148037761449814, + 0.004634103737771511, + 0.005200231913477182, + 0.006963289342820644, + 0.004249367862939835, + 0.0035748975351452827, + 0.0039834207855165005, + 0.0034512437414377928, + 0.0036465111188590527, + 0.0053271218203008175, + 0.002881026128306985, + 0.0030661066994071007, + 0.005040586460381746, + 0.002356036566197872, + 0.0024128009099513292, + 0.0020289281383156776, + 0.006494815461337566, + 0.003751551965251565, + 0.019501160830259323, + 0.0026783980429172516, + 0.019372161477804184, + 0.005174429155886173, + 0.002136301016435027, + 0.0021013468503952026, + 0.009370502084493637, + 0.01393149234354496, + 0.008211303502321243, + 0.011073031462728977, + 0.02364237792789936, + 0.01169257890433073, + 0.011058586649596691, + 0.005177051294595003, + 0.005999935325235128, + 0.003017744980752468, + 0.004748529754579067, + 0.002282574540004134, + 0.004461775999516249, + 0.007298552896827459, + 0.0023399819619953632, + 0.003436849219724536, + 0.015299179591238499, + 0.016873950138688087, + 0.01615164428949356, + 0.012175662443041801, + 0.01506650261580944, + 0.020314235240221024, + 0.016553569585084915, + 0.018450116738677025, + 0.003806602442637086, + 0.0022481682244688272, + 0.0025860227178782225, + 0.003977352287620306, + 0.003623224562034011, + 0.004490273538976908, + 0.0035375694278627634, + 0.0034738352987915277, + 0.009485799819231033, + 0.008804299868643284, + 0.009450206533074379, + 0.008056561462581158, + 0.0104465801268816, + 0.008033785969018936, + 0.006584984716027975, + 0.00800654012709856, + 0.02306724712252617, + 0.06733198463916779, + 0.0051664672791957855, + 0.004591313656419516, + 0.008273448795080185, + 0.0626353919506073, + 0.005612767767161131, + 0.005823030136525631, + 0.03871084004640579, + 0.12805287539958954, + 0.010606610216200352, + 0.010786735452711582, + 0.012456150725483894, + 0.12029003351926804, + 0.007609042804688215, + 0.01010587066411972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.0013380631571635604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on a parcel of dry earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the earth colored with ochre and sienna tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the earth contain a prominent, jagged crack that propagates and causes the ground to split in two?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the two separated halves of the earth settle apart revealing a dark, shadowed interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934278607368469 + ], + "opposite_logprob_tokens": [ + -0.6934278607368469 + ], + "logprob_reward": -0.6934278607368469, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934278607368469, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12420842051506042, + 0.035793811082839966, + 0.1976643204689026, + 0.04234855994582176, + 0.08110290765762329, + 0.0657169297337532, + 0.04691791534423828, + 0.12617018818855286, + 0.18184331059455872, + 0.0982336476445198 + ], + "credit_map": [ + 0.012251326814293861, + 0.0147660281509161, + 0.01314661093056202, + 0.01864468865096569, + 0.019673671573400497, + 0.017349086701869965, + 0.01388899888843298, + 0.014488018117845058, + 0.004430580418556929, + 0.007096241693943739, + 0.005042848642915487, + 0.0028271926566958427, + 0.005282682832330465, + 0.00532478466629982, + 0.0022669874597340822, + 0.003522490616887808, + 0.01582822948694229, + 0.010692482814192772, + 0.06601112335920334, + 0.010504410602152348, + 0.06425240635871887, + 0.01754653826355934, + 0.007170804776251316, + 0.005658325273543596, + 0.004774727858603001, + 0.0072969612665474415, + 0.004495659843087196, + 0.004886520095169544, + 0.005836457945406437, + 0.005085055250674486, + 0.005271788686513901, + 0.004701386671513319, + 0.007239512167870998, + 0.006275725085288286, + 0.012001270428299904, + 0.005252850707620382, + 0.01632089912891388, + 0.027193618938326836, + 0.003148061456158757, + 0.0036709634587168694, + 0.006294816732406616, + 0.008839236572384834, + 0.008853654377162457, + 0.0062754107639193535, + 0.009887666441500187, + 0.011761723086237907, + 0.007459299638867378, + 0.006345120724290609, + 0.006035080179572105, + 0.004738062620162964, + 0.004580318462103605, + 0.00673519354313612, + 0.0067912619560956955, + 0.007896061055362225, + 0.005887556821107864, + 0.004254381638020277, + 0.017874451354146004, + 0.01582673378288746, + 0.015160489827394485, + 0.015073473565280437, + 0.02023332566022873, + 0.012888885103166103, + 0.016163179650902748, + 0.012949648313224316, + 0.022814473137259483, + 0.06792586296796799, + 0.0061532314866781235, + 0.004535599146038294, + 0.009985567070543766, + 0.059644926339387894, + 0.004706292878836393, + 0.00607735849916935, + 0.01108777616173029, + 0.027931010350584984, + 0.00518382852897048, + 0.0075998189859092236, + 0.009220151230692863, + 0.02757842466235161, + 0.004867215640842915, + 0.004765419755131006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.6934278607368469 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_tensor([5589, 5589])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_tensor([5589, 5589])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b6ddfd430a247ef675b0d5d5493ac242565f90e6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_tensor([5589, 5589])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up shot from a level angle, using high frame rate cinematography for breathtaking super-slow motion, focuses on a large fruit bat, hanging upside down from a gnarled, moss-covered branch. In the soft golden glow of dusk, two large, moist black eyes gleam alertly in its fox-like face. A dense, dark brown down coats its body, and a halo of golden hair at its shoulders and neck creates a soft halo in the backlight. The action begins: its tiny paws delicately, almost weightlessly, release from the bark, and its body begins a graceful, controlled descent. In that instant, its wings, once wrapped around its body like a black leather cape, unfold in a breathtaking sequence with the precision of origami. Slender, bony phalanges stretched outward, fully expanding the wrinkled, translucent wing membrane. The setting sun's rays penetrated the membrane, illuminating the complex network of blood vessels within, like a stained-glass window in a Gothic cathedral. With its wings fully extended, it used a long-accumulated force to complete a powerful yet gentle downward push. The wing membrane bulged in the air, clearly revealing the ripples of muscle beneath the fur. This movement instantly stopped its descent and propelled its body forward steadily and calmly, blending into the twilight forest backdrop. The entire scene was filled with primal vitality and a solemn, almost sacred tranquility.", + "seed": 23268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_5589_seed_23268_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028381537273526193, + "vlm_logprob": -0.028381537273526193, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.028381537273526193, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019704781472682953 + ], + "opposite_logprob_tokens": [ + -4.144704818725586 + ], + "logprob_reward": -0.019704781472682953, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019704781472682953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09536030888557434, + 0.06421579420566559, + 0.0445755310356617, + 0.10214853286743164, + 0.06763595342636108, + 0.12204592674970627, + 0.18740785121917725, + 0.11341937631368637, + 0.10938803106546402, + 0.09380271285772324 + ], + "credit_map": [ + 0.0100401621311903, + 0.009335274808108807, + 0.026414044201374054, + 0.009561021812260151, + 0.0064780451357364655, + 0.007437144406139851, + 0.008853786624968052, + 0.017240827903151512, + 0.006008184514939785, + 0.011099175550043583, + 0.00932659488171339, + 0.008624075911939144, + 0.006353067699819803, + 0.008926005102694035, + 0.005634558852761984, + 0.008244127035140991, + 0.006795684807002544, + 0.00829450786113739, + 0.006430208217352629, + 0.005598760209977627, + 0.004042027052491903, + 0.004024678375571966, + 0.003025102661922574, + 0.006364564411342144, + 0.01460073608905077, + 0.01997927762567997, + 0.014253119938075542, + 0.016106897965073586, + 0.007738889195024967, + 0.01085385400801897, + 0.007446281146258116, + 0.011169465258717537, + 0.011114108376204967, + 0.009977640584111214, + 0.0092332623898983, + 0.011694526299834251, + 0.004149111919105053, + 0.006768946070224047, + 0.00938398577272892, + 0.005314366891980171, + 0.016432518139481544, + 0.013588075526058674, + 0.01676412858068943, + 0.01715044677257538, + 0.011854061856865883, + 0.013874230906367302, + 0.018679101020097733, + 0.01370336301624775, + 0.019043726846575737, + 0.034992996603250504, + 0.020342811942100525, + 0.026635006070137024, + 0.015795279294252396, + 0.014686701819300652, + 0.031028987839818, + 0.024882344529032707, + 0.011738983914256096, + 0.010738288052380085, + 0.016263972967863083, + 0.011093571782112122, + 0.01873028464615345, + 0.022036073729395866, + 0.01152830757200718, + 0.011289894580841064, + 0.011347249150276184, + 0.018219733610749245, + 0.016827847808599472, + 0.01785297691822052, + 0.009797834791243076, + 0.008472205139696598, + 0.01255069486796856, + 0.014319492504000664, + 0.013620969839394093, + 0.012762660160660744, + 0.011195291765034199, + 0.012993128970265388, + 0.013710390776395798, + 0.00782579556107521, + 0.006347951479256153, + 0.015346516855061054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.019704781472682953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499040335416794 + ], + "opposite_logprob_tokens": [ + -2.929990291595459 + ], + "logprob_reward": -0.05499040335416794, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499040335416794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15791085362434387, + 0.10575947165489197, + 0.02914418652653694, + 0.07244814932346344, + 0.09555239975452423, + 0.17605407536029816, + 0.1428520828485489, + 0.08220401406288147, + 0.07928929477930069, + 0.05878552049398422 + ], + "credit_map": [ + 0.01493163127452135, + 0.018701016902923584, + 0.02896571159362793, + 0.030231203883886337, + 0.01085707638412714, + 0.0083836130797863, + 0.024036606773734093, + 0.021803993731737137, + 0.012259633280336857, + 0.016746418550610542, + 0.012772032991051674, + 0.019345421344041824, + 0.008325209841132164, + 0.008971665054559708, + 0.011274178512394428, + 0.01606491021811962, + 0.004698372446000576, + 0.0034529485274106264, + 0.004288696683943272, + 0.003517228877171874, + 0.0016813442343845963, + 0.0034178004134446383, + 0.002473047934472561, + 0.005614748224616051, + 0.009654592722654343, + 0.009764058515429497, + 0.0098344087600708, + 0.010779418051242828, + 0.006272437516599894, + 0.011309086345136166, + 0.0061009665951132774, + 0.00873318687081337, + 0.016809502616524696, + 0.01497690100222826, + 0.014231587760150433, + 0.018317988142371178, + 0.006964348256587982, + 0.005765441805124283, + 0.01227596402168274, + 0.006210668943822384, + 0.017366774380207062, + 0.03378085792064667, + 0.02190660685300827, + 0.02985021471977234, + 0.0176674947142601, + 0.015701211988925934, + 0.029399391263723373, + 0.010381524451076984, + 0.015159047208726406, + 0.025509269908070564, + 0.020026493817567825, + 0.029809940606355667, + 0.011714514344930649, + 0.016048546880483627, + 0.014946076087653637, + 0.009638208895921707, + 0.008451433852314949, + 0.007984420284628868, + 0.011664766818284988, + 0.010121423751115799, + 0.010622542351484299, + 0.015499151311814785, + 0.008234470151364803, + 0.009625805541872978, + 0.007498371414840221, + 0.008090442046523094, + 0.015922196209430695, + 0.008663571439683437, + 0.006563277915120125, + 0.008417578414082527, + 0.008655928075313568, + 0.015477922745049, + 0.005353868007659912, + 0.008964823558926582, + 0.007834102027118206, + 0.010324602946639061, + 0.008035615086555481, + 0.005313219036906958, + 0.004883126821368933, + 0.008076165802776814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.05499040335416794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023312531411647797 + ], + "opposite_logprob_tokens": [ + -3.773312568664551 + ], + "logprob_reward": -0.023312531411647797, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023312531411647797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07151326537132263, + 0.1667434573173523, + 0.026762740686535835, + 0.059179503470659256, + 0.06832947582006454, + 0.06810428202152252, + 0.21496136486530304, + 0.02533108741044998, + 0.13290518522262573, + 0.1661696434020996 + ], + "credit_map": [ + 0.010518482886254787, + 0.008642910979688168, + 0.009104524739086628, + 0.009784169495105743, + 0.009274936281144619, + 0.00684988871216774, + 0.00961916521191597, + 0.007719182409346104, + 0.01602742075920105, + 0.010140424594283104, + 0.06437835097312927, + 0.013409710489213467, + 0.005656723398715258, + 0.00947780441492796, + 0.006722352001816034, + 0.04093065485358238, + 0.003540759440511465, + 0.006136102136224508, + 0.003169976407662034, + 0.004186591133475304, + 0.00286203739233315, + 0.0020837292540818453, + 0.002540396759286523, + 0.002243146300315857, + 0.0062356614507734776, + 0.0143130449578166, + 0.008775775320827961, + 0.0072232382372021675, + 0.0045238821767270565, + 0.0063676610589027405, + 0.005678856745362282, + 0.006061383057385683, + 0.011611736379563808, + 0.0126146599650383, + 0.007977535016834736, + 0.009422168135643005, + 0.005815078038722277, + 0.007203714922070503, + 0.009051612578332424, + 0.004632969386875629, + 0.007166706956923008, + 0.006695304531604052, + 0.00962381437420845, + 0.009251173585653305, + 0.006779405754059553, + 0.007766983471810818, + 0.011535724624991417, + 0.00928516499698162, + 0.02151738665997982, + 0.03735920414328575, + 0.027045579627156258, + 0.03303608298301697, + 0.01661563292145729, + 0.01715112291276455, + 0.03565770015120506, + 0.026578659191727638, + 0.002515424508601427, + 0.0024798293597996235, + 0.0037718075327575207, + 0.0050351847894489765, + 0.0034292128402739763, + 0.002648349618539214, + 0.003332757158204913, + 0.0021185213699936867, + 0.014752199873328209, + 0.030144263058900833, + 0.017568999901413918, + 0.017542723566293716, + 0.011327467858791351, + 0.008412404917180538, + 0.015005898661911488, + 0.018151240423321724, + 0.01602836512029171, + 0.024056391790509224, + 0.02016741782426834, + 0.02836672216653824, + 0.02791910618543625, + 0.018043842166662216, + 0.01587524637579918, + 0.015712551772594452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.023312531411647797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058476023375988 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.02058476023375988, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058476023375988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07963351905345917, + 0.044972989708185196, + 0.03641074523329735, + 0.04837167635560036, + 0.055398810654878616, + 0.13490724563598633, + 0.0857660174369812, + 0.05498002469539642, + 0.19066911935806274, + 0.268889844417572 + ], + "credit_map": [ + 0.009265038184821606, + 0.010801099240779877, + 0.012364473193883896, + 0.013238020241260529, + 0.0072614289820194244, + 0.006474766414612532, + 0.010945549234747887, + 0.009283141233026981, + 0.005296623334288597, + 0.007103934418410063, + 0.00591687299311161, + 0.004520230460911989, + 0.0038189911283552647, + 0.005029234569519758, + 0.0039637647569179535, + 0.009323340840637684, + 0.004690451547503471, + 0.004682213068008423, + 0.00540382694453001, + 0.004341273568570614, + 0.0035598506219685078, + 0.004355283919721842, + 0.0033233878202736378, + 0.006054458674043417, + 0.006343552377074957, + 0.008907740004360676, + 0.007333202753216028, + 0.005897994618862867, + 0.00503876805305481, + 0.004751688800752163, + 0.004882830660790205, + 0.00521590281277895, + 0.008000391535460949, + 0.007025972940027714, + 0.007646204903721809, + 0.009969502687454224, + 0.003238917561247945, + 0.005646927282214165, + 0.008866040036082268, + 0.005004849284887314, + 0.018083415925502777, + 0.020278185606002808, + 0.018628107383847237, + 0.02275964431464672, + 0.009897178038954735, + 0.01350084226578474, + 0.019355852156877518, + 0.012404024600982666, + 0.009866980835795403, + 0.0126701220870018, + 0.013651003129780293, + 0.012285558506846428, + 0.006954653188586235, + 0.007051803637295961, + 0.012643516063690186, + 0.010642385110259056, + 0.0058221518993377686, + 0.006018179003149271, + 0.006633264012634754, + 0.006254333537071943, + 0.007058885879814625, + 0.00999181903898716, + 0.006018759682774544, + 0.007182630710303783, + 0.017476607114076614, + 0.02533508464694023, + 0.025915151461958885, + 0.028864149004220963, + 0.024701397866010666, + 0.014362053014338017, + 0.024764418601989746, + 0.02925024926662445, + 0.0267656110227108, + 0.0347827710211277, + 0.03454979881644249, + 0.04718879982829094, + 0.0449163094162941, + 0.026840059086680412, + 0.02610127255320549, + 0.027745233848690987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.02058476023375988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331520989537239 + ], + "opposite_logprob_tokens": [ + -3.773315191268921 + ], + "logprob_reward": -0.02331520989537239, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331520989537239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09381623566150665, + 0.04823929816484451, + 0.05050146207213402, + 0.07644426822662354, + 0.08284260332584381, + 0.0862937718629837, + 0.19585174322128296, + 0.05245671421289444, + 0.07950855791568756, + 0.23404531180858612 + ], + "credit_map": [ + 0.014288930222392082, + 0.013518241234123707, + 0.009205643087625504, + 0.017669139429926872, + 0.009060119278728962, + 0.007080249488353729, + 0.0146024189889431, + 0.008391502313315868, + 0.0049601527862250805, + 0.007861834019422531, + 0.00686892494559288, + 0.0072656371630728245, + 0.0042574722319841385, + 0.004472485743463039, + 0.0053009092807769775, + 0.0072518824599683285, + 0.0064708576537668705, + 0.01204881351441145, + 0.005754887126386166, + 0.006240526679903269, + 0.004811449442058802, + 0.004910611547529697, + 0.0043289512395858765, + 0.005935365334153175, + 0.009788976050913334, + 0.008298123255372047, + 0.012013668194413185, + 0.008391715586185455, + 0.007254281546920538, + 0.011470927856862545, + 0.008412015624344349, + 0.010814555920660496, + 0.013417236506938934, + 0.0126968864351511, + 0.011858928948640823, + 0.019750447943806648, + 0.004051098134368658, + 0.004434439819306135, + 0.01103556901216507, + 0.005598000250756741, + 0.011011678725481033, + 0.012186678126454353, + 0.011403265409171581, + 0.014877749606966972, + 0.005871438886970282, + 0.010133168660104275, + 0.012090123258531094, + 0.00871967151761055, + 0.020989196375012398, + 0.026118382811546326, + 0.02248539961874485, + 0.03320779278874397, + 0.015491638332605362, + 0.014035429805517197, + 0.03716537728905678, + 0.02635852061212063, + 0.005393668543547392, + 0.0046317083761096, + 0.007869016379117966, + 0.00658907275646925, + 0.006055091507732868, + 0.010363689623773098, + 0.0046360548585653305, + 0.00691841309890151, + 0.009063354693353176, + 0.01304224319756031, + 0.00869322195649147, + 0.011151805520057678, + 0.008684764616191387, + 0.004869960248470306, + 0.010861716233193874, + 0.013141492381691933, + 0.0276222825050354, + 0.03626907616853714, + 0.030627010390162468, + 0.03235800936818123, + 0.038037996739149094, + 0.026646049693226814, + 0.020340438932180405, + 0.02214444801211357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.02331520989537239 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a level angle, using high frame rate cinematography for breathtaking super-slow motion, focuses on a large fruit bat, hanging upside down from a gnarled, moss-covered branch. In the soft golden glow of dusk, two large, moist black eyes gleam alertly in its fox-like face. A dense, dark brown down coats its body, and a halo of golden hair at its shoulders and neck creates a soft halo in the backlight. The action begins: its tiny paws delicately, almost weightlessly, release from the bark, and its body begins a graceful, controlled descent. In that instant, its wings, once wrapped around its body like a black leather cape, unfold in a breathtaking sequence with the precision of origami. Slender, bony phalanges stretched outward, fully expanding the wrinkled, translucent wing membrane. The setting sun's rays penetrated the membrane, illuminating the complex network of blood vessels within, like a stained-glass window in a Gothic cathedral. With its wings fully extended, it used a long-accumulated force to complete a powerful yet gentle downward push. The wing membrane bulged in the air, clearly revealing the ripples of muscle beneath the fur. This movement instantly stopped its descent and propelled its body forward steadily and calmly, blending into the twilight forest backdrop. The entire scene was filled with primal vitality and a solemn, almost sacred tranquility.", + "seed": 23269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_5589_seed_23269_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029822661727666854, + "vlm_logprob": -0.029822661727666854, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.029822661727666854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022794954478740692 + ], + "opposite_logprob_tokens": [ + -3.8977949619293213 + ], + "logprob_reward": -0.022794954478740692, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022794954478740692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07890020310878754, + 0.09037970751523972, + 0.09154939651489258, + 0.08073888719081879, + 0.04670745134353638, + 0.06348632276058197, + 0.06150459498167038, + 0.03787529841065407, + 0.28795692324638367, + 0.16090121865272522 + ], + "credit_map": [ + 0.012141450308263302, + 0.008381500840187073, + 0.010605055838823318, + 0.011009326204657555, + 0.010189391672611237, + 0.010075210593640804, + 0.00961453840136528, + 0.00688373064622283, + 0.021941130980849266, + 0.008550516329705715, + 0.01259220577776432, + 0.010361230000853539, + 0.007419754285365343, + 0.0069697848521173, + 0.010040109045803547, + 0.012504976242780685, + 0.008898809552192688, + 0.00900279637426138, + 0.011443271301686764, + 0.014515984803438187, + 0.010106776840984821, + 0.015072567388415337, + 0.014603993855416775, + 0.007905200123786926, + 0.008467631414532661, + 0.008537338115274906, + 0.014966911636292934, + 0.011683852411806583, + 0.014749403111636639, + 0.005622896831482649, + 0.009255798533558846, + 0.007455059792846441, + 0.004617762286216021, + 0.004100169520825148, + 0.006554220337420702, + 0.006178746465593576, + 0.006996287032961845, + 0.004245641175657511, + 0.008438633754849434, + 0.005575994029641151, + 0.0115456348285079, + 0.0038556831423193216, + 0.012023987248539925, + 0.006435031536966562, + 0.004142077174037695, + 0.00823833979666233, + 0.005417095497250557, + 0.011828478425741196, + 0.005954972468316555, + 0.00501726008951664, + 0.01361342053860426, + 0.007996510714292526, + 0.003912485670298338, + 0.011622454971075058, + 0.0051040914840996265, + 0.008283395320177078, + 0.006728701759129763, + 0.003699018619954586, + 0.005676743574440479, + 0.003667915938422084, + 0.001697259140200913, + 0.00654270825907588, + 0.004512013867497444, + 0.005350933875888586, + 0.06090179458260536, + 0.023093348369002342, + 0.034912846982479095, + 0.03502132371068001, + 0.01880526728928089, + 0.05014263093471527, + 0.03363800048828125, + 0.03144170716404915, + 0.0257865022867918, + 0.01391868107020855, + 0.014936334453523159, + 0.03551016375422478, + 0.012809962034225464, + 0.015136295929551125, + 0.029082100838422775, + 0.013721168972551823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.022794954478740692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499401316046715 + ], + "opposite_logprob_tokens": [ + -2.9299941062927246 + ], + "logprob_reward": -0.05499401316046715, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499401316046715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10276392102241516, + 0.051834698766469955, + 0.0973002165555954, + 0.03762020543217659, + 0.04915737360715866, + 0.18677377700805664, + 0.07294163107872009, + 0.04052357375621796, + 0.3377678394317627, + 0.023316720500588417 + ], + "credit_map": [ + 0.01414133608341217, + 0.010309613309800625, + 0.011701663956046104, + 0.014969922602176666, + 0.013334836810827255, + 0.013967151753604412, + 0.011410422623157501, + 0.012928973883390427, + 0.009745980612933636, + 0.00496439216658473, + 0.0077029038220644, + 0.007469566073268652, + 0.00588285643607378, + 0.005027043167501688, + 0.005389368627220392, + 0.005652588326483965, + 0.009260348044335842, + 0.006281369365751743, + 0.014414973556995392, + 0.01603373885154724, + 0.013780487701296806, + 0.006771518383175135, + 0.024493174627423286, + 0.00626459950581193, + 0.004433405119925737, + 0.003722238354384899, + 0.004452346358448267, + 0.007041519042104483, + 0.006052938289940357, + 0.002376945223659277, + 0.0047365170903503895, + 0.0048042950220406055, + 0.011549859307706356, + 0.004785896744579077, + 0.011324130930006504, + 0.0044913385063409805, + 0.002743195276707411, + 0.004173724912106991, + 0.0038314464036375284, + 0.006257776636630297, + 0.021579684689641, + 0.008639058098196983, + 0.03744000941514969, + 0.03875771164894104, + 0.00635777972638607, + 0.0238778255879879, + 0.008388394489884377, + 0.04173331707715988, + 0.006686248350888491, + 0.004879993852227926, + 0.014747011475265026, + 0.01289325300604105, + 0.004680198151618242, + 0.01602627895772457, + 0.005324130412191153, + 0.007704512216150761, + 0.007574899587780237, + 0.00380669510923326, + 0.005681030917912722, + 0.0038186144083738327, + 0.0017825528047978878, + 0.008012397214770317, + 0.0036635324358940125, + 0.006183851044625044, + 0.06482326239347458, + 0.029396329075098038, + 0.0427587516605854, + 0.0451706238090992, + 0.021001843735575676, + 0.055148713290691376, + 0.04229278117418289, + 0.037175532430410385, + 0.0027161594480276108, + 0.0022727048490196466, + 0.0028098784387111664, + 0.0041643111035227776, + 0.002339784987270832, + 0.0038199606351554394, + 0.003112258855253458, + 0.002081663114950061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.05499401316046715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642790675163269 + ], + "opposite_logprob_tokens": [ + -4.88264274597168 + ], + "logprob_reward": -0.007642790675163269, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642790675163269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24184909462928772, + 0.03596245124936104, + 0.04620261490345001, + 0.08101482689380646, + 0.0908045619726181, + 0.05771467462182045, + 0.04748426377773285, + 0.05628601834177971, + 0.20053119957447052, + 0.14215026795864105 + ], + "credit_map": [ + 0.05163753405213356, + 0.021080732345581055, + 0.02956414595246315, + 0.033901672810316086, + 0.023269589990377426, + 0.02377534843981266, + 0.03311872109770775, + 0.025501348078250885, + 0.006540862377732992, + 0.0027809725143015385, + 0.004120725207030773, + 0.004874238278716803, + 0.004433851223438978, + 0.00345258554443717, + 0.00455464469268918, + 0.005204574670642614, + 0.003733010496944189, + 0.0028118619229644537, + 0.006752586457878351, + 0.007210837677121162, + 0.005981528665870428, + 0.002642933279275894, + 0.01316632516682148, + 0.0039035300724208355, + 0.008420704863965511, + 0.009887848980724812, + 0.010778501629829407, + 0.012345733121037483, + 0.014500897377729416, + 0.005746053997427225, + 0.01151329092681408, + 0.007821796461939812, + 0.011367106810212135, + 0.007316388189792633, + 0.02385447919368744, + 0.008453245274722576, + 0.011499527841806412, + 0.006742405705153942, + 0.012378147803246975, + 0.009193258360028267, + 0.0076330373995006084, + 0.0038960876408964396, + 0.01265734527260065, + 0.005667202174663544, + 0.0031827655620872974, + 0.008096776902675629, + 0.004948235582560301, + 0.011633223854005337, + 0.005048174411058426, + 0.00387748540379107, + 0.007826373912394047, + 0.00589141296222806, + 0.0031789441127330065, + 0.010311520658433437, + 0.003278107848018408, + 0.008072244003415108, + 0.008558941073715687, + 0.006898723542690277, + 0.00917520746588707, + 0.005755648482590914, + 0.0027929863426834345, + 0.007864593528211117, + 0.00854511745274067, + 0.006694803014397621, + 0.040045518428087234, + 0.016974689438939095, + 0.03205089271068573, + 0.01923193782567978, + 0.016286708414554596, + 0.03456442803144455, + 0.02242649905383587, + 0.01895054616034031, + 0.01785501092672348, + 0.010312455706298351, + 0.013086763210594654, + 0.02259446308016777, + 0.015430953353643417, + 0.020276082679629326, + 0.029390938580036163, + 0.01320360042154789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.007642790675163269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307413473725319 + ], + "opposite_logprob_tokens": [ + -3.168074131011963 + ], + "logprob_reward": -0.04307413473725319, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307413473725319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11929665505886078, + 0.05809742212295532, + 0.06094840541481972, + 0.09573061764240265, + 0.16916680335998535, + 0.10243172943592072, + 0.05009729415178299, + 0.04208950325846672, + 0.0905950665473938, + 0.21154643595218658 + ], + "credit_map": [ + 0.025737833231687546, + 0.009594597853720188, + 0.017923636361956596, + 0.015110875479876995, + 0.011536539532244205, + 0.010847597382962704, + 0.01836865209043026, + 0.010176925919950008, + 0.009918284602463245, + 0.005384939722716808, + 0.01040701661258936, + 0.006827260367572308, + 0.005995008163154125, + 0.005024606827646494, + 0.005314321257174015, + 0.00922598410397768, + 0.0074251312762498856, + 0.005819390993565321, + 0.005342843011021614, + 0.007394846994429827, + 0.010945005342364311, + 0.010674573481082916, + 0.007876764982938766, + 0.005469853058457375, + 0.011073306202888489, + 0.010593977756798267, + 0.015406804159283638, + 0.013434242457151413, + 0.015214327722787857, + 0.007395300082862377, + 0.010001408867537975, + 0.01261125598102808, + 0.02878771722316742, + 0.01517048105597496, + 0.038754940032958984, + 0.014779594726860523, + 0.016751093789935112, + 0.012676491402089596, + 0.022780366241931915, + 0.019466131925582886, + 0.03064778260886669, + 0.0045991200022399426, + 0.010686243884265423, + 0.01188001036643982, + 0.0028075098525732756, + 0.00789640098810196, + 0.00517761055380106, + 0.02873704768717289, + 0.0052854148671031, + 0.004187347367405891, + 0.009720562025904655, + 0.005642416421324015, + 0.003746291156858206, + 0.009569643996655941, + 0.004828010685741901, + 0.007117606699466705, + 0.006823652423918247, + 0.004327842965722084, + 0.006661924999207258, + 0.004954876843839884, + 0.0022236742079257965, + 0.0071061053313314915, + 0.004892106633633375, + 0.005099318455904722, + 0.018491756170988083, + 0.0055852290242910385, + 0.01622467301785946, + 0.009082982316613197, + 0.005140164401382208, + 0.015266995877027512, + 0.010873174294829369, + 0.009930094704031944, + 0.030131326988339424, + 0.018964946269989014, + 0.019956612959504128, + 0.041861098259687424, + 0.0187770314514637, + 0.02285894937813282, + 0.041787005960941315, + 0.01720946468412876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.04307413473725319 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607415586709976 + ], + "opposite_logprob_tokens": [ + -3.8956074714660645 + ], + "logprob_reward": -0.020607415586709976, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607415586709976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19941705465316772, + 0.10273202508687973, + 0.06249569356441498, + 0.07950079441070557, + 0.0767136663198471, + 0.08164390921592712, + 0.07989192754030228, + 0.04789002239704132, + 0.1846216320991516, + 0.08509320765733719 + ], + "credit_map": [ + 0.033600978553295135, + 0.018644563853740692, + 0.026236306875944138, + 0.03301970288157463, + 0.020229404792189598, + 0.018446210771799088, + 0.025486674159765244, + 0.02375325933098793, + 0.01748286373913288, + 0.010978597216308117, + 0.013755911961197853, + 0.01483512669801712, + 0.01205400750041008, + 0.007283429615199566, + 0.0129231046885252, + 0.01341900322586298, + 0.006555244792252779, + 0.005410932470113039, + 0.008896923623979092, + 0.00858892872929573, + 0.012573590502142906, + 0.006881198845803738, + 0.00896396767348051, + 0.004624912980943918, + 0.012234579771757126, + 0.005884281825274229, + 0.021264439448714256, + 0.012485845014452934, + 0.010463980957865715, + 0.0038260777946561575, + 0.0072480193339288235, + 0.006093581672757864, + 0.016348207369446754, + 0.006588988937437534, + 0.022846264764666557, + 0.005481929983943701, + 0.004837402608245611, + 0.005134143866598606, + 0.006646406836807728, + 0.008830330334603786, + 0.011852560564875603, + 0.0041941809467971325, + 0.01867060363292694, + 0.011245113797485828, + 0.0039152768440544605, + 0.012327589094638824, + 0.00525543000549078, + 0.0141831673681736, + 0.007738929241895676, + 0.006407666020095348, + 0.016434865072369576, + 0.009854204021394253, + 0.0048941937275230885, + 0.01887509599328041, + 0.005705124698579311, + 0.009981859475374222, + 0.010403832420706749, + 0.004108653403818607, + 0.008328111842274666, + 0.005477372091263533, + 0.002215026179328561, + 0.008060176856815815, + 0.0044375513680279255, + 0.004859304521232843, + 0.04137059301137924, + 0.01401425339281559, + 0.022195929661393166, + 0.021003618836402893, + 0.013542247004806995, + 0.029058661311864853, + 0.023235928267240524, + 0.02020043134689331, + 0.014638533815741539, + 0.006072981283068657, + 0.008427788503468037, + 0.012705571949481964, + 0.01180990133434534, + 0.012310131452977657, + 0.012791633605957031, + 0.006336684338748455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.020607415586709976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a level angle, using high frame rate cinematography for breathtaking super-slow motion, focuses on a large fruit bat, hanging upside down from a gnarled, moss-covered branch. In the soft golden glow of dusk, two large, moist black eyes gleam alertly in its fox-like face. A dense, dark brown down coats its body, and a halo of golden hair at its shoulders and neck creates a soft halo in the backlight. The action begins: its tiny paws delicately, almost weightlessly, release from the bark, and its body begins a graceful, controlled descent. In that instant, its wings, once wrapped around its body like a black leather cape, unfold in a breathtaking sequence with the precision of origami. Slender, bony phalanges stretched outward, fully expanding the wrinkled, translucent wing membrane. The setting sun's rays penetrated the membrane, illuminating the complex network of blood vessels within, like a stained-glass window in a Gothic cathedral. With its wings fully extended, it used a long-accumulated force to complete a powerful yet gentle downward push. The wing membrane bulged in the air, clearly revealing the ripples of muscle beneath the fur. This movement instantly stopped its descent and propelled its body forward steadily and calmly, blending into the twilight forest backdrop. The entire scene was filled with primal vitality and a solemn, almost sacred tranquility.", + "seed": 23270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_5589_seed_23270_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.032218044996261595, + "vlm_logprob": -0.032218044996261595, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.032218044996261595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01970735378563404 + ], + "opposite_logprob_tokens": [ + -4.144707202911377 + ], + "logprob_reward": -0.01970735378563404, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01970735378563404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12790447473526, + 0.08029290288686752, + 0.08057728409767151, + 0.07511956244707108, + 0.06228584051132202, + 0.10945440083742142, + 0.08224380761384964, + 0.21505960822105408, + 0.08965659141540527, + 0.07740549743175507 + ], + "credit_map": [ + 0.014361316338181496, + 0.014742824248969555, + 0.01663029193878174, + 0.020262354984879494, + 0.015468822792172432, + 0.0202640388160944, + 0.012195263057947159, + 0.013979571871459484, + 0.011366094462573528, + 0.007957055233418941, + 0.011430121958255768, + 0.007770117372274399, + 0.0098239965736866, + 0.014745611697435379, + 0.008685750886797905, + 0.00851416401565075, + 0.01263977400958538, + 0.013388578779995441, + 0.01281822845339775, + 0.011046404018998146, + 0.009244847111403942, + 0.005250673275440931, + 0.007678445894271135, + 0.008510341867804527, + 0.007097759749740362, + 0.005581798497587442, + 0.00847354345023632, + 0.016321174800395966, + 0.004324993118643761, + 0.00889771431684494, + 0.007104008924216032, + 0.017318572849035263, + 0.012359706684947014, + 0.005405391566455364, + 0.013591350987553596, + 0.006852878723293543, + 0.005497342441231012, + 0.005603685509413481, + 0.005289889872074127, + 0.007685603573918343, + 0.01653904654085636, + 0.011501394212245941, + 0.021160969510674477, + 0.01367933675646782, + 0.012561104260385036, + 0.012456119060516357, + 0.01121070422232151, + 0.01034573744982481, + 0.011613554321229458, + 0.004364138934761286, + 0.011944344267249107, + 0.009004558436572552, + 0.005337726324796677, + 0.007041218690574169, + 0.023495495319366455, + 0.009442775510251522, + 0.03211653232574463, + 0.012805194593966007, + 0.022286366671323776, + 0.03909032791852951, + 0.019572876393795013, + 0.01645878329873085, + 0.03793987259268761, + 0.03478967398405075, + 0.011615234427154064, + 0.007827633060514927, + 0.01106058806180954, + 0.01798967458307743, + 0.007363314274698496, + 0.010201571509242058, + 0.013329343870282173, + 0.010269246064126492, + 0.01038958691060543, + 0.010224444791674614, + 0.00958263874053955, + 0.008267665281891823, + 0.010483570396900177, + 0.010804841294884682, + 0.008223338052630424, + 0.009429422207176685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.01970735378563404 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005004584789276 + ], + "opposite_logprob_tokens": [ + -2.695050001144409 + ], + "logprob_reward": -0.07005004584789276, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005004584789276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04185822233557701, + 0.13903364539146423, + 0.035587068647146225, + 0.08089730888605118, + 0.12834955751895905, + 0.3286653757095337, + 0.07321599125862122, + 0.05214821547269821, + 0.02895769104361534, + 0.09128693491220474 + ], + "credit_map": [ + 0.00539069389924407, + 0.00589363370090723, + 0.005490147974342108, + 0.005848804488778114, + 0.006021250505000353, + 0.0050626350566744804, + 0.004077033139765263, + 0.004074024967849255, + 0.017051460221409798, + 0.0125243766233325, + 0.01553589478135109, + 0.01939556747674942, + 0.019596802070736885, + 0.023462088778614998, + 0.013220580294728279, + 0.018246887251734734, + 0.003970906138420105, + 0.006479312665760517, + 0.00535171153023839, + 0.006065357476472855, + 0.0044899312779307365, + 0.0026682408060878515, + 0.002972852438688278, + 0.0035887551493942738, + 0.0066529023461043835, + 0.0054333205334842205, + 0.007908845320343971, + 0.02474602870643139, + 0.0036369666922837496, + 0.004513927269726992, + 0.007165160030126572, + 0.02084016054868698, + 0.037122439593076706, + 0.004862260073423386, + 0.054249364882707596, + 0.01159052737057209, + 0.00481357891112566, + 0.0037678040098398924, + 0.0033759025391191244, + 0.008567684330046177, + 0.03133619949221611, + 0.02824799343943596, + 0.036211829632520676, + 0.07396116107702255, + 0.026645304635167122, + 0.022628339007496834, + 0.02353844605386257, + 0.08609609305858612, + 0.01020260900259018, + 0.005685943644493818, + 0.006811453029513359, + 0.01261044293642044, + 0.007178663741797209, + 0.012126380577683449, + 0.011182560585439205, + 0.007417930290102959, + 0.006237393710762262, + 0.0042746528051793575, + 0.007174666039645672, + 0.008675549179315567, + 0.004648798145353794, + 0.0049688853323459625, + 0.008558922447264194, + 0.007609345484524965, + 0.003772763069719076, + 0.0019163384567946196, + 0.004187829792499542, + 0.005315092392265797, + 0.0016872818814590573, + 0.003016394330188632, + 0.004895245656371117, + 0.004166744649410248, + 0.010955206118524075, + 0.012230471707880497, + 0.00941947940737009, + 0.008136158809065819, + 0.019147032871842384, + 0.013234139420092106, + 0.007009089924395084, + 0.011155360378324986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.07005004584789276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314977064728737 + ], + "opposite_logprob_tokens": [ + -3.773314952850342 + ], + "logprob_reward": -0.023314977064728737, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314977064728737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053284578025341034, + 0.08182749152183533, + 0.14713633060455322, + 0.034659091383218765, + 0.16687816381454468, + 0.12146583944559097, + 0.07730801403522491, + 0.1066368967294693, + 0.11265143752098083, + 0.09815213084220886 + ], + "credit_map": [ + 0.009496360085904598, + 0.004787403624504805, + 0.00865235086530447, + 0.008534089662134647, + 0.0052896300330758095, + 0.004080365411937237, + 0.005639787297695875, + 0.006804593838751316, + 0.00833396427333355, + 0.008019625209271908, + 0.008393310941755772, + 0.014161641709506512, + 0.011805826798081398, + 0.012295353226363659, + 0.00806112214922905, + 0.010756656527519226, + 0.01961689442396164, + 0.019116822630167007, + 0.02243570238351822, + 0.01642804779112339, + 0.01741616427898407, + 0.020301295444369316, + 0.015981294214725494, + 0.01584012061357498, + 0.0028991769067943096, + 0.003518089186400175, + 0.004238400142639875, + 0.00655187014490366, + 0.0028409312944859266, + 0.004375537857413292, + 0.00333643751218915, + 0.0068986546248197556, + 0.05260663852095604, + 0.006053064949810505, + 0.07123841345310211, + 0.013906414620578289, + 0.004714014008641243, + 0.004674876108765602, + 0.004600639455020428, + 0.00908411480486393, + 0.014029954560101032, + 0.013383862562477589, + 0.015905756503343582, + 0.017102070152759552, + 0.015201826579868793, + 0.015134957619011402, + 0.010415836237370968, + 0.020291581749916077, + 0.01165250688791275, + 0.004912524949759245, + 0.010773559100925922, + 0.015780629590153694, + 0.004969487432390451, + 0.006071171723306179, + 0.010340369306504726, + 0.012807771563529968, + 0.013317158445715904, + 0.009570584632456303, + 0.013257359154522419, + 0.019157644361257553, + 0.008362805470824242, + 0.013497677631676197, + 0.013992387801408768, + 0.015481282025575638, + 0.016168970614671707, + 0.010819726623594761, + 0.01392355840653181, + 0.021801700815558434, + 0.008565972559154034, + 0.012382064945995808, + 0.01732983998954296, + 0.011659613810479641, + 0.01020963117480278, + 0.010529224760830402, + 0.013845337554812431, + 0.010430919006466866, + 0.019921893253922462, + 0.011195112951099873, + 0.011287667788565159, + 0.010732351802289486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.023314977064728737 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805727303028107 + ], + "opposite_logprob_tokens": [ + -3.5298056602478027 + ], + "logprob_reward": -0.029805727303028107, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805727303028107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18935376405715942, + 0.07086832076311111, + 0.0738108828663826, + 0.030009541660547256, + 0.054120391607284546, + 0.21216131746768951, + 0.0667070597410202, + 0.09253054112195969, + 0.10716356337070465, + 0.10327461361885071 + ], + "credit_map": [ + 0.01888182945549488, + 0.023134339600801468, + 0.015209249220788479, + 0.025531766936182976, + 0.030148472636938095, + 0.03718898817896843, + 0.019864819943904877, + 0.01939430460333824, + 0.01701292395591736, + 0.007202850189059973, + 0.012116365134716034, + 0.009396960958838463, + 0.00838849600404501, + 0.00515563553199172, + 0.005569816567003727, + 0.006025276612490416, + 0.012153788469731808, + 0.007902358658611774, + 0.011467227712273598, + 0.007672111503779888, + 0.007675715256482363, + 0.010587941855192184, + 0.008307347074151039, + 0.008044383488595486, + 0.0031719403341412544, + 0.003180063795298338, + 0.004610337782651186, + 0.004345591180026531, + 0.002919438760727644, + 0.004609571769833565, + 0.004756529349833727, + 0.0024160693865269423, + 0.006922597531229258, + 0.004170360043644905, + 0.014930945821106434, + 0.005856562405824661, + 0.00421320041641593, + 0.004003662616014481, + 0.00393314566463232, + 0.010089912451803684, + 0.019909052178263664, + 0.01961841993033886, + 0.025936873629689217, + 0.04651405289769173, + 0.021407729014754295, + 0.01651182398200035, + 0.015660390257835388, + 0.04660297930240631, + 0.010637745261192322, + 0.003473534481599927, + 0.007607251405715942, + 0.012535388581454754, + 0.003506380133330822, + 0.00641631567850709, + 0.016804827377200127, + 0.00572562450543046, + 0.014065570197999477, + 0.00591380475088954, + 0.015242993831634521, + 0.019388217478990555, + 0.007112841587513685, + 0.008220857009291649, + 0.012563138268887997, + 0.010023125447332859, + 0.008842348121106625, + 0.009157732129096985, + 0.010186928324401379, + 0.020183365792036057, + 0.008278763853013515, + 0.006108547560870647, + 0.028735380619764328, + 0.015670495107769966, + 0.010876079089939594, + 0.012781678698956966, + 0.01510645542293787, + 0.014060163870453835, + 0.016896642744541168, + 0.011276006698608398, + 0.008985280990600586, + 0.013292303308844566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.029805727303028107 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018212120980024338 + ], + "opposite_logprob_tokens": [ + -4.01821231842041 + ], + "logprob_reward": -0.018212120980024338, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018212120980024338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16557615995407104, + 0.0566759891808033, + 0.07776732742786407, + 0.09114202111959457, + 0.07334794849157333, + 0.0985565111041069, + 0.103346087038517, + 0.1955348700284958, + 0.05604168772697449, + 0.08201133459806442 + ], + "credit_map": [ + 0.018514681607484818, + 0.018432490527629852, + 0.0143372667953372, + 0.01919451914727688, + 0.023872170597314835, + 0.03963473066687584, + 0.015006915666162968, + 0.016583403572440147, + 0.007918043993413448, + 0.006233166437596083, + 0.007133577484637499, + 0.009186336770653725, + 0.007708204910159111, + 0.0077273426577448845, + 0.005095252301543951, + 0.005674069281667471, + 0.012408779934048653, + 0.009807596914470196, + 0.011929077096283436, + 0.009161864407360554, + 0.007953762076795101, + 0.0092161251232028, + 0.006706473417580128, + 0.010583661496639252, + 0.01075577549636364, + 0.0072950455360114574, + 0.015934284776449203, + 0.01674768142402172, + 0.006350483745336533, + 0.012078849598765373, + 0.011419439688324928, + 0.010560469701886177, + 0.009588969871401787, + 0.008351675234735012, + 0.01104801706969738, + 0.009493032470345497, + 0.009253686293959618, + 0.009309696964919567, + 0.006612397730350494, + 0.009690475650131702, + 0.013226534239947796, + 0.01085138600319624, + 0.018228068947792053, + 0.014649111777544022, + 0.011971849948167801, + 0.009122620336711407, + 0.010105889290571213, + 0.010401054285466671, + 0.029777754098176956, + 0.006510415114462376, + 0.008506500162184238, + 0.010295872576534748, + 0.0065666548907756805, + 0.005311938934028149, + 0.028965316712856293, + 0.007411645725369453, + 0.029683398082852364, + 0.010789659805595875, + 0.02030721865594387, + 0.03543539717793465, + 0.019495896995067596, + 0.01789136417210102, + 0.029955318197607994, + 0.031976643949747086, + 0.007566280663013458, + 0.006699420511722565, + 0.0069327084347605705, + 0.0075238230638206005, + 0.00388105190359056, + 0.006396017968654633, + 0.007983176037669182, + 0.009059218689799309, + 0.00875828042626381, + 0.010411921888589859, + 0.010784665122628212, + 0.007884158752858639, + 0.014699188992381096, + 0.013027905486524105, + 0.008018752560019493, + 0.008426465094089508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.018212120980024338 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a level angle, using high frame rate cinematography for breathtaking super-slow motion, focuses on a large fruit bat, hanging upside down from a gnarled, moss-covered branch. In the soft golden glow of dusk, two large, moist black eyes gleam alertly in its fox-like face. A dense, dark brown down coats its body, and a halo of golden hair at its shoulders and neck creates a soft halo in the backlight. The action begins: its tiny paws delicately, almost weightlessly, release from the bark, and its body begins a graceful, controlled descent. In that instant, its wings, once wrapped around its body like a black leather cape, unfold in a breathtaking sequence with the precision of origami. Slender, bony phalanges stretched outward, fully expanding the wrinkled, translucent wing membrane. The setting sun's rays penetrated the membrane, illuminating the complex network of blood vessels within, like a stained-glass window in a Gothic cathedral. With its wings fully extended, it used a long-accumulated force to complete a powerful yet gentle downward push. The wing membrane bulged in the air, clearly revealing the ripples of muscle beneath the fur. This movement instantly stopped its descent and propelled its body forward steadily and calmly, blending into the twilight forest backdrop. The entire scene was filled with primal vitality and a solemn, almost sacred tranquility.", + "seed": 23271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_13/index_5589_seed_23271_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025310812704265117, + "vlm_logprob": -0.025310812704265117, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.025310812704265117, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019691457971930504 + ], + "opposite_logprob_tokens": [ + -4.144691467285156 + ], + "logprob_reward": -0.019691457971930504, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019691457971930504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08339089155197144, + 0.0509454607963562, + 0.5446979999542236, + 0.02176680974662304, + 0.051476266235113144, + 0.0968705415725708, + 0.024070605635643005, + 0.03739238530397415, + 0.033978477120399475, + 0.055410563945770264 + ], + "credit_map": [ + 0.012812763452529907, + 0.006623826455324888, + 0.011463256552815437, + 0.005915514659136534, + 0.006891120690852404, + 0.009849411435425282, + 0.0069535523653030396, + 0.022881435230374336, + 0.008592606522142887, + 0.004673824645578861, + 0.006938219536095858, + 0.005622728727757931, + 0.0045764935202896595, + 0.004280871246010065, + 0.008424384519457817, + 0.007836326956748962, + 0.12031896412372589, + 0.02482738345861435, + 0.10392753034830093, + 0.01941067725419998, + 0.016914885491132736, + 0.03120156191289425, + 0.022877508774399757, + 0.20521949231624603, + 0.0036886692978441715, + 0.0016376341227442026, + 0.0026424399111419916, + 0.0024190829135477543, + 0.0016067964024841785, + 0.002499217865988612, + 0.001885423669591546, + 0.005387545097619295, + 0.0066426401026546955, + 0.006441241130232811, + 0.008287549018859863, + 0.004326908383518457, + 0.004502329975366592, + 0.009271723218262196, + 0.005130405072122812, + 0.006873462349176407, + 0.008558390662074089, + 0.015264494344592094, + 0.012316721491515636, + 0.017176808789372444, + 0.01235902588814497, + 0.0071070329286158085, + 0.012488780543208122, + 0.011599280871450901, + 0.002321364125236869, + 0.0033114321995526552, + 0.003537750104442239, + 0.0040864418260753155, + 0.0030097372364252806, + 0.001624660100787878, + 0.00409844983369112, + 0.0020807671826332808, + 0.003631792962551117, + 0.0042084879241883755, + 0.006290534045547247, + 0.005115213338285685, + 0.003189487149938941, + 0.00403174851089716, + 0.005049274768680334, + 0.005875840783119202, + 0.003370808670297265, + 0.002355902921408415, + 0.008196992799639702, + 0.004216205794364214, + 0.003026563674211502, + 0.0024962215684354305, + 0.004492395091801882, + 0.005823384504765272, + 0.010595123283565044, + 0.0022814020048826933, + 0.008162405341863632, + 0.00772207323461771, + 0.002832732629030943, + 0.003083376679569483, + 0.015524548478424549, + 0.005208897870033979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.019691457971930504 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812587261199951 + ], + "opposite_logprob_tokens": [ + -3.288125991821289 + ], + "logprob_reward": -0.03812587261199951, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812587261199951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10555516928434372, + 0.029443563893437386, + 0.14383389055728912, + 0.02691739983856678, + 0.09789851307868958, + 0.21042993664741516, + 0.05088035762310028, + 0.13119451701641083, + 0.13893717527389526, + 0.06490946561098099 + ], + "credit_map": [ + 0.012895752675831318, + 0.01256569754332304, + 0.013533147983253002, + 0.014637179672718048, + 0.01046186313033104, + 0.00920960959047079, + 0.012285229749977589, + 0.019966691732406616, + 0.0056265415623784065, + 0.0024801387917250395, + 0.003022274700924754, + 0.0035588627215474844, + 0.002444581827148795, + 0.0020300885662436485, + 0.004316890146583319, + 0.005964186508208513, + 0.022593950852751732, + 0.007369049824774265, + 0.02031867206096649, + 0.014623523689806461, + 0.0080217020586133, + 0.010845761746168137, + 0.019006649032235146, + 0.041054584085941315, + 0.005199046805500984, + 0.0021445685997605324, + 0.0032381187193095684, + 0.00388826965354383, + 0.0015601705526933074, + 0.002493330743163824, + 0.002429329790174961, + 0.005964567419141531, + 0.012762206606566906, + 0.012320280075073242, + 0.012388436123728752, + 0.013241438195109367, + 0.0074980841018259525, + 0.013041842728853226, + 0.010102267377078533, + 0.016543958336114883, + 0.021395765244960785, + 0.03359757736325264, + 0.028144828975200653, + 0.030212994664907455, + 0.023965517058968544, + 0.017369607463479042, + 0.028287095949053764, + 0.02745654247701168, + 0.0054100388661026955, + 0.0057159364223480225, + 0.009710012935101986, + 0.01291200052946806, + 0.0036032539792358875, + 0.0026541263796389103, + 0.005760824307799339, + 0.005114166997373104, + 0.009980319067835808, + 0.011926015838980675, + 0.021667178720235825, + 0.024506594985723495, + 0.010269843973219395, + 0.009971512481570244, + 0.02470671944320202, + 0.018166346475481987, + 0.01826455071568489, + 0.008059442974627018, + 0.023792728781700134, + 0.044290799647569656, + 0.008972354233264923, + 0.0068493615835905075, + 0.01252843253314495, + 0.01617949828505516, + 0.013909331522881985, + 0.0030834609642624855, + 0.007983089424669743, + 0.010921507142484188, + 0.0025614863261580467, + 0.003252392401918769, + 0.016738470643758774, + 0.006459726020693779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.03812587261199951 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559722177684307 + ], + "opposite_logprob_tokens": [ + -4.38755989074707 + ], + "logprob_reward": -0.012559722177684307, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559722177684307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07304756343364716, + 0.07724090665578842, + 0.25303465127944946, + 0.08170590549707413, + 0.05525488033890724, + 0.03755216300487518, + 0.048514071851968765, + 0.1540491282939911, + 0.10891060531139374, + 0.1106901615858078 + ], + "credit_map": [ + 0.009976130910217762, + 0.005061163567006588, + 0.009417288936674595, + 0.007560598663985729, + 0.0059301769360899925, + 0.007837595418095589, + 0.011325900442898273, + 0.01593870297074318, + 0.010542123578488827, + 0.004598511382937431, + 0.02424595318734646, + 0.009312986396253109, + 0.004729871172457933, + 0.00497675035148859, + 0.009078841656446457, + 0.009755859151482582, + 0.032191045582294464, + 0.009801037609577179, + 0.07043546438217163, + 0.013968236744403839, + 0.010697088204324245, + 0.012722023762762547, + 0.012154009193181992, + 0.09106572717428207, + 0.013443777337670326, + 0.005688782315701246, + 0.009431470185518265, + 0.010705681517720222, + 0.0059362854808568954, + 0.00868228916078806, + 0.011453303508460522, + 0.016364308074116707, + 0.007940510287880898, + 0.006604285445064306, + 0.006551451049745083, + 0.007608545478433371, + 0.005089792422950268, + 0.008198341354727745, + 0.005314034875482321, + 0.007947912439703941, + 0.005309185478836298, + 0.005654873326420784, + 0.0046462565660476685, + 0.004576768726110458, + 0.004783087410032749, + 0.0030929751228541136, + 0.004898505751043558, + 0.004590508062392473, + 0.0046601854264736176, + 0.00644534919410944, + 0.007029896602034569, + 0.006985654588788748, + 0.005757885053753853, + 0.0039007961750030518, + 0.009289108216762543, + 0.0044451928697526455, + 0.012939962558448315, + 0.013469657860696316, + 0.02480621077120304, + 0.02708587422966957, + 0.012256236746907234, + 0.011189760640263557, + 0.027398135513067245, + 0.02490326762199402, + 0.012966256588697433, + 0.005929498467594385, + 0.019103771075606346, + 0.03253631293773651, + 0.0070485942997038364, + 0.005638509523123503, + 0.011180237866938114, + 0.014507416635751724, + 0.01695159636437893, + 0.00565843190997839, + 0.019722294062376022, + 0.014237220399081707, + 0.007153088692575693, + 0.013903179205954075, + 0.01604490540921688, + 0.017019441351294518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.012559722177684307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803529381752014 + ], + "opposite_logprob_tokens": [ + -3.529803514480591 + ], + "logprob_reward": -0.029803529381752014, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803529381752014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07686954736709595, + 0.031845949590206146, + 0.36799877882003784, + 0.09500015527009964, + 0.16442936658859253, + 0.07423041760921478, + 0.029504168778657913, + 0.037942033261060715, + 0.06269528716802597, + 0.059484247118234634 + ], + "credit_map": [ + 0.012577265501022339, + 0.006336950697004795, + 0.010965606197714806, + 0.009824550710618496, + 0.006389491725713015, + 0.005478391889482737, + 0.008242865093052387, + 0.017054425552487373, + 0.0037792797666043043, + 0.002408746164292097, + 0.005173375364392996, + 0.002898648614063859, + 0.003196826670318842, + 0.002729976549744606, + 0.0045941416174173355, + 0.0070649548433721066, + 0.04659662023186684, + 0.014986501075327396, + 0.09985148161649704, + 0.029158465564250946, + 0.013426309451460838, + 0.018043091520667076, + 0.01789836585521698, + 0.1280379593372345, + 0.022817404940724373, + 0.006082182750105858, + 0.013711785897612572, + 0.005854969844222069, + 0.003730891738086939, + 0.007539927959442139, + 0.006222802679985762, + 0.029040183871984482, + 0.02104070782661438, + 0.023266633972525597, + 0.022361526265740395, + 0.0210666973143816, + 0.014341825619339943, + 0.024702012538909912, + 0.017204102128744125, + 0.020445851609110832, + 0.006642238236963749, + 0.00972486101090908, + 0.01007159985601902, + 0.015004110522568226, + 0.008657973259687424, + 0.0075039733201265335, + 0.009391817264258862, + 0.007233841344714165, + 0.003894588677212596, + 0.0026966508012264967, + 0.004701044410467148, + 0.003611799329519272, + 0.002791109960526228, + 0.0024417147506028414, + 0.005574106238782406, + 0.0037931560073047876, + 0.003952970262616873, + 0.004503319039940834, + 0.005477906204760075, + 0.004909218288958073, + 0.003169348230585456, + 0.0034922657068818808, + 0.006214021239429712, + 0.006222984287887812, + 0.004870415199548006, + 0.0033766364213079214, + 0.010626537725329399, + 0.01133895292878151, + 0.0048195612616837025, + 0.004328574053943157, + 0.008637329563498497, + 0.014697271399199963, + 0.008228402584791183, + 0.0035758698359131813, + 0.01410745270550251, + 0.01051564421504736, + 0.003203485393896699, + 0.006187709514051676, + 0.006083373446017504, + 0.007582311984151602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.029803529381752014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a large fruit bat hanging upside down from a gnarled, moss-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bat have a fox-like face with two large, moist black eyes and a dense dark brown down coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bat's wings unfold from a wrapped position into a fully extended state, revealing a translucent membrane with visible blood vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bat perform a controlled descent followed by a powerful downward push that propels it forward into the twilight forest backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot as a stunning close-up from a level angle using high frame rate super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637348137795925 + ], + "opposite_logprob_tokens": [ + -3.6513733863830566 + ], + "logprob_reward": -0.02637348137795925, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637348137795925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03556351363658905, + 0.038824714720249176, + 0.8030663132667542, + 0.010622798465192318, + 0.02409127913415432, + 0.012539842166006565, + 0.013408762402832508, + 0.010512415319681168, + 0.01708495058119297, + 0.03428537771105766 + ], + "credit_map": [ + 0.005382381845265627, + 0.004371373914182186, + 0.007129077333956957, + 0.002997450064867735, + 0.0024174468126147985, + 0.0023604375310242176, + 0.002774182241410017, + 0.008131164126098156, + 0.0071254754438996315, + 0.0033123812172561884, + 0.0091700479388237, + 0.003978123422712088, + 0.0030384997371584177, + 0.003959173336625099, + 0.004468308761715889, + 0.003772701835259795, + 0.16497434675693512, + 0.033646948635578156, + 0.17262279987335205, + 0.02534296363592148, + 0.02256166748702526, + 0.04558093100786209, + 0.028392573818564415, + 0.3099440634250641, + 0.0015607140958309174, + 0.0007712570950388908, + 0.0016470963601022959, + 0.0012532732216641307, + 0.0008893359918147326, + 0.00150225218385458, + 0.001180108985863626, + 0.001818759716115892, + 0.003150236327201128, + 0.0028728567995131016, + 0.002734934212639928, + 0.004043126478791237, + 0.0023956571239978075, + 0.0032042544335126877, + 0.002380315214395523, + 0.0033098997082561255, + 0.0018115820130333304, + 0.0018003490986302495, + 0.0016959081403911114, + 0.0014359354972839355, + 0.0014347346732392907, + 0.001146372640505433, + 0.0013405181234702468, + 0.0018744422122836113, + 0.001301420503295958, + 0.0014741681516170502, + 0.0019388834480196238, + 0.0023576184175908566, + 0.0011451561003923416, + 0.0009545397479087114, + 0.0024817148223519325, + 0.0017552608624100685, + 0.0011217055143788457, + 0.0009655827889218926, + 0.001666454947553575, + 0.0014072251506149769, + 0.0007716023828834295, + 0.0009308129083365202, + 0.0021480075083673, + 0.0015010233037173748, + 0.0019422121113166213, + 0.0011437473585829139, + 0.003350463230162859, + 0.004829255864024162, + 0.0010144629050046206, + 0.0008500308613292873, + 0.0020746763329952955, + 0.0018800991820171475, + 0.007973366416990757, + 0.001204225467517972, + 0.008805880323052406, + 0.005230287089943886, + 0.0011875225463882089, + 0.0020361694041639566, + 0.004904198460280895, + 0.002943724859505892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.02637348137795925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_14/index_tensor([6127, 6127])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_14/index_tensor([6127, 6127])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f31f7ad34b12c120ab3718511f7739a45227fed1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_14/index_tensor([6127, 6127])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful first-person perspective shot places the viewer directly on the back of a high-speed motorcycle. Your hands, clad in black, full-fingered cycling gloves, grip the metal handlebars tightly as you gaze over the motorcycle's crimson windshield and dashboard as you speed along Kuala Lumpur's vast, multi-lane arterial. Directly ahead, the brake lights of a white car suddenly flash. Responding in a split second, you force your body to the left, and the world tilts dramatically. To avoid the car, the car's body swerves into the ground at a jarring angle, your left knee nearly scraping the rough tarmac. During this breathtaking tilt, the iconic Kuala Lumpur buildings in the background—the majestic Petronas Twin Towers and the towering Merdeka 118—are rendered in intense motion blur and wide-angle distortion, as if stretched to the edges of your vision by an invisible force. Their contours become distorted and elongated, transforming into a dazzling, flowing line alongside the surrounding cityscape. The entire picture is rendered through the unique imaging style of Leica lenses, presenting a signature oily texture and extremely rich and saturated colors. The soft evening light gives the metal body and glass curtain wall a gorgeous luster, perfectly capturing this adrenaline-rushing moment in a cinematic way that is full of extreme dynamic beauty and tense atmosphere.", + "seed": 2482, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_14/index_6127_seed_2482_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15423302464187144, + "vlm_logprob": -0.15423302464187144, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.15423302464187144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2257360965013504 + ], + "opposite_logprob_tokens": [ + -1.725736141204834 + ], + "logprob_reward": -0.2257360965013504, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2257360965013504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07827570289373398, + 0.11774107068777084, + 0.0213677566498518, + 0.049798231571912766, + 0.05048681050539017, + 0.3497704863548279, + 0.0939524695277214, + 0.09535746276378632, + 0.07335741072893143, + 0.06989259272813797 + ], + "credit_map": [ + 0.014094997197389603, + 0.013327606953680515, + 0.00711342366412282, + 0.006998325698077679, + 0.010427167639136314, + 0.011939644813537598, + 0.006387122441083193, + 0.007987406104803085, + 0.02609442174434662, + 0.023618727922439575, + 0.0100050438195467, + 0.006725161336362362, + 0.014021508395671844, + 0.013798728585243225, + 0.008233788423240185, + 0.01524369791150093, + 0.002847526455298066, + 0.0031271642073988914, + 0.0020467734429985285, + 0.002708862069994211, + 0.0025860299356281757, + 0.0028199658263474703, + 0.0021719816140830517, + 0.0030594554264098406, + 0.007185902912169695, + 0.007704201620072126, + 0.004369488917291164, + 0.0041158078238368034, + 0.007634816225618124, + 0.00712808221578598, + 0.004825884010642767, + 0.006834049243479967, + 0.0034399759024381638, + 0.013163931667804718, + 0.004184999968856573, + 0.0035120912361890078, + 0.005988202523440123, + 0.0027489999774843454, + 0.004212846048176289, + 0.013235759921371937, + 0.05984235927462578, + 0.027614496648311615, + 0.08147634565830231, + 0.04861615225672722, + 0.024501414969563484, + 0.04048797860741615, + 0.032254479825496674, + 0.03497724607586861, + 0.008981176652014256, + 0.007097104098647833, + 0.013500512577593327, + 0.01752367615699768, + 0.00956745445728302, + 0.01192412618547678, + 0.01786840707063675, + 0.007490010466426611, + 0.014809790067374706, + 0.007295260671526194, + 0.02458629570901394, + 0.008924472145736217, + 0.007338099181652069, + 0.01729394681751728, + 0.008341571316123009, + 0.0067680287174880505, + 0.010518673807382584, + 0.01112365536391735, + 0.00961256306618452, + 0.009975194931030273, + 0.006858613807708025, + 0.008689802139997482, + 0.010185371153056622, + 0.00639354158192873, + 0.010825276374816895, + 0.007743555121123791, + 0.016852959990501404, + 0.009797886945307255, + 0.005899128504097462, + 0.005928319413214922, + 0.006971708498895168, + 0.005873761139810085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.2257360965013504 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869803786277771 + ], + "opposite_logprob_tokens": [ + -3.0486979484558105 + ], + "logprob_reward": -0.04869803786277771, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869803786277771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1423788070678711, + 0.07679879665374756, + 0.05196553096175194, + 0.08875666558742523, + 0.08401180803775787, + 0.02653549611568451, + 0.2177613228559494, + 0.10387668758630753, + 0.08229478448629379, + 0.1256200522184372 + ], + "credit_map": [ + 0.019798750057816505, + 0.024950679391622543, + 0.015357867814600468, + 0.013249524869024754, + 0.015968840569257736, + 0.01618623547255993, + 0.017618797719478607, + 0.019248122349381447, + 0.013114584609866142, + 0.013798819854855537, + 0.009423775598406792, + 0.006273925304412842, + 0.008296665735542774, + 0.009563609026372433, + 0.005913682747632265, + 0.010413742624223232, + 0.008078427985310555, + 0.007014925125986338, + 0.004753968678414822, + 0.007341468706727028, + 0.0059427861124277115, + 0.00496641406789422, + 0.0070364284329116344, + 0.006831117440015078, + 0.006300435867160559, + 0.007395946886390448, + 0.008456983603537083, + 0.016032880172133446, + 0.008984331972897053, + 0.005971889942884445, + 0.020734531804919243, + 0.014879679307341576, + 0.013406017795205116, + 0.019327791407704353, + 0.006601477973163128, + 0.005746122449636459, + 0.010335886850953102, + 0.011396372690796852, + 0.005080465227365494, + 0.012117681093513966, + 0.00233498215675354, + 0.00343875284306705, + 0.003707560244947672, + 0.0027198195457458496, + 0.005845261737704277, + 0.003260145429521799, + 0.0027403857093304396, + 0.002488593105226755, + 0.014774322509765625, + 0.01486427616328001, + 0.03552604094147682, + 0.042394720017910004, + 0.02570083551108837, + 0.01951632648706436, + 0.045750971883535385, + 0.019233837723731995, + 0.014092334546148777, + 0.007935576140880585, + 0.020114783197641373, + 0.013472932390868664, + 0.008218353614211082, + 0.01459980383515358, + 0.013940563425421715, + 0.011502353474497795, + 0.012342661619186401, + 0.012926860712468624, + 0.010252184234559536, + 0.012746608816087246, + 0.0072714583948254585, + 0.007409777492284775, + 0.013115877285599709, + 0.006229363847523928, + 0.017039351165294647, + 0.014183053746819496, + 0.018174804747104645, + 0.019513854756951332, + 0.010220525786280632, + 0.010092060081660748, + 0.01642713136970997, + 0.019969284534454346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.04869803786277771 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38711631298065186 + ], + "opposite_logprob_tokens": [ + -1.1371163129806519 + ], + "logprob_reward": -0.38711631298065186, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38711631298065186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16510377824306488, + 0.06640095263719559, + 0.025239650160074234, + 0.04947225749492645, + 0.13232626020908356, + 0.1303628534078598, + 0.1111782044172287, + 0.11929886043071747, + 0.07413092255592346, + 0.12648624181747437 + ], + "credit_map": [ + 0.02144479751586914, + 0.020032765343785286, + 0.01573670655488968, + 0.02032747119665146, + 0.01807275228202343, + 0.0235647764056921, + 0.019181860610842705, + 0.026742644608020782, + 0.011394072324037552, + 0.009412224404513836, + 0.007077978923916817, + 0.0066504389978945255, + 0.005150527693331242, + 0.006435915362089872, + 0.008958171121776104, + 0.011321624740958214, + 0.0049109249375760555, + 0.004175898618996143, + 0.0024617102462798357, + 0.0028236745856702328, + 0.0025699185207486153, + 0.002949954941868782, + 0.002548852236941457, + 0.0027987181674689054, + 0.004372046794742346, + 0.004643700551241636, + 0.0047775814309716225, + 0.008591340854763985, + 0.004766492638736963, + 0.00400313176214695, + 0.009280326776206493, + 0.009037638083100319, + 0.009527034126222134, + 0.033995628356933594, + 0.012831351719796658, + 0.010121176019310951, + 0.016286566853523254, + 0.009425049647688866, + 0.010183949023485184, + 0.02995549514889717, + 0.020736880600452423, + 0.01100330613553524, + 0.03294875845313072, + 0.017331324517726898, + 0.01006658747792244, + 0.014190992340445518, + 0.012124616652727127, + 0.011960381641983986, + 0.009792458266019821, + 0.007856319658458233, + 0.015237340703606606, + 0.019146105274558067, + 0.014042642898857594, + 0.016379211097955704, + 0.02011554315686226, + 0.008608588948845863, + 0.022335711866617203, + 0.012242172844707966, + 0.01980980858206749, + 0.009382530115544796, + 0.012487316504120827, + 0.024342520162463188, + 0.007525898981839418, + 0.011172906495630741, + 0.01108317356556654, + 0.008860422298312187, + 0.009881733916699886, + 0.007480226922780275, + 0.008587686344981194, + 0.017688900232315063, + 0.004365678410977125, + 0.006183099467307329, + 0.019430972635746002, + 0.013352196663618088, + 0.032150425016880035, + 0.01880556158721447, + 0.011332841590046883, + 0.011571640148758888, + 0.010618266649544239, + 0.00922433752566576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.38711631298065186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058218978345394 + ], + "opposite_logprob_tokens": [ + -3.8955821990966797 + ], + "logprob_reward": -0.02058218978345394, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058218978345394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15680573880672455, + 0.05606387183070183, + 0.11560751497745514, + 0.16907504200935364, + 0.10904659330844879, + 0.07183738052845001, + 0.12144892662763596, + 0.09829452633857727, + 0.027785342186689377, + 0.07403508573770523 + ], + "credit_map": [ + 0.028661619871854782, + 0.02815035730600357, + 0.01708291657269001, + 0.009673409163951874, + 0.019830744713544846, + 0.024894749745726585, + 0.012444383464753628, + 0.016067564487457275, + 0.013952070847153664, + 0.010125151835381985, + 0.004101335536688566, + 0.004274329636245966, + 0.005937333684414625, + 0.006540882401168346, + 0.005279983393847942, + 0.005852787755429745, + 0.01564130373299122, + 0.021606147289276123, + 0.009959892369806767, + 0.008109373040497303, + 0.01762261614203453, + 0.02150340937077999, + 0.007563023362308741, + 0.013601750135421753, + 0.024879077449440956, + 0.03166330233216286, + 0.014296851120889187, + 0.01135456096380949, + 0.026915444061160088, + 0.030612777918577194, + 0.011329022236168385, + 0.018024006858468056, + 0.018841387704014778, + 0.011548012495040894, + 0.010516861453652382, + 0.0065234205685555935, + 0.011578962206840515, + 0.0176693145185709, + 0.009060381911695004, + 0.023308245465159416, + 0.005400697235018015, + 0.005250813905149698, + 0.01998407207429409, + 0.0139160780236125, + 0.004662139341235161, + 0.0062681506387889385, + 0.008819193579256535, + 0.007536240853369236, + 0.014643313363194466, + 0.009245045483112335, + 0.017365125939249992, + 0.01988261379301548, + 0.014520544558763504, + 0.012354710139334202, + 0.02397226355969906, + 0.009465319104492664, + 0.020023947581648827, + 0.010190754197537899, + 0.01982036605477333, + 0.008360501378774643, + 0.007200961466878653, + 0.018460797145962715, + 0.00660566845908761, + 0.007631535641849041, + 0.005722793284803629, + 0.005088640842586756, + 0.003225193591788411, + 0.003030430292710662, + 0.0028494831640273333, + 0.003441928420215845, + 0.002170420018956065, + 0.0022564532700926065, + 0.009894713759422302, + 0.009734398685395718, + 0.016204899176955223, + 0.011142648756504059, + 0.006518187001347542, + 0.005762028973549604, + 0.008715249598026276, + 0.006062966305762529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.02058218978345394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903248608112335 + ], + "opposite_logprob_tokens": [ + -2.4640324115753174 + ], + "logprob_reward": -0.08903248608112335, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903248608112335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19793638586997986, + 0.1213199645280838, + 0.028924815356731415, + 0.03212400898337364, + 0.11593244224786758, + 0.08663935959339142, + 0.047739025205373764, + 0.12567591667175293, + 0.0534142404794693, + 0.19029377400875092 + ], + "credit_map": [ + 0.03232625499367714, + 0.031155860051512718, + 0.025614308193325996, + 0.027227509766817093, + 0.02140081115067005, + 0.019073927775025368, + 0.02006230689585209, + 0.021075427532196045, + 0.026381341740489006, + 0.020937113091349602, + 0.007200171705335379, + 0.010801626369357109, + 0.01604270003736019, + 0.01686827279627323, + 0.010415942408144474, + 0.012672808952629566, + 0.006537583190947771, + 0.003977753221988678, + 0.002457277849316597, + 0.002492732834070921, + 0.0033588053192943335, + 0.0036558723077178, + 0.0029755409341305494, + 0.0034692503977566957, + 0.0037547829560935497, + 0.006006565410643816, + 0.0019893820863217115, + 0.005183483939617872, + 0.0036163697950541973, + 0.0038775524590164423, + 0.003497390542179346, + 0.0041984859853982925, + 0.01817694678902626, + 0.012510713189840317, + 0.011989026330411434, + 0.009381012991070747, + 0.01787978783249855, + 0.015143249183893204, + 0.00907913688570261, + 0.021772578358650208, + 0.010318403132259846, + 0.0056466455571353436, + 0.022495241835713387, + 0.014353453181684017, + 0.00721823051571846, + 0.007873179391026497, + 0.009710888378322124, + 0.009023326449096203, + 0.006145817693322897, + 0.005047677084803581, + 0.0060429261066019535, + 0.005193926859647036, + 0.0063080559484660625, + 0.008509375154972076, + 0.005303660873323679, + 0.005187586881220341, + 0.018745338544249535, + 0.013425862416625023, + 0.026576455682516098, + 0.012635592371225357, + 0.008979381062090397, + 0.021424125880002975, + 0.011818904429674149, + 0.012070263735949993, + 0.0064452472142875195, + 0.009369159117341042, + 0.007880454882979393, + 0.007698375731706619, + 0.005589487962424755, + 0.006735672242939472, + 0.003810713766142726, + 0.005885131191462278, + 0.02623096853494644, + 0.023632382974028587, + 0.04219263046979904, + 0.0294181015342474, + 0.01485636830329895, + 0.014802778139710426, + 0.020533133298158646, + 0.018627434968948364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.08903248608112335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person perspective shot places the viewer directly on the back of a high-speed motorcycle. Your hands, clad in black, full-fingered cycling gloves, grip the metal handlebars tightly as you gaze over the motorcycle's crimson windshield and dashboard as you speed along Kuala Lumpur's vast, multi-lane arterial. Directly ahead, the brake lights of a white car suddenly flash. Responding in a split second, you force your body to the left, and the world tilts dramatically. To avoid the car, the car's body swerves into the ground at a jarring angle, your left knee nearly scraping the rough tarmac. During this breathtaking tilt, the iconic Kuala Lumpur buildings in the background—the majestic Petronas Twin Towers and the towering Merdeka 118—are rendered in intense motion blur and wide-angle distortion, as if stretched to the edges of your vision by an invisible force. Their contours become distorted and elongated, transforming into a dazzling, flowing line alongside the surrounding cityscape. The entire picture is rendered through the unique imaging style of Leica lenses, presenting a signature oily texture and extremely rich and saturated colors. The soft evening light gives the metal body and glass curtain wall a gorgeous luster, perfectly capturing this adrenaline-rushing moment in a cinematic way that is full of extreme dynamic beauty and tense atmosphere.", + "seed": 2483, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_14/index_6127_seed_2483_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.055988306552171706, + "vlm_logprob": -0.055988306552171706, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.055988306552171706, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06462786346673965 + ], + "opposite_logprob_tokens": [ + -2.9396278858184814 + ], + "logprob_reward": -0.06462786346673965, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06462786346673965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18699102103710175, + 0.060208145529031754, + 0.0791267678141594, + 0.16905240714550018, + 0.06981466710567474, + 0.058238010853528976, + 0.08604688942432404, + 0.16294234991073608, + 0.04994409903883934, + 0.07763566821813583 + ], + "credit_map": [ + 0.019687224179506302, + 0.018923519179224968, + 0.02042035199701786, + 0.030578797683119774, + 0.014766111969947815, + 0.041599880903959274, + 0.026350688189268112, + 0.01466443482786417, + 0.007026124279946089, + 0.006145221181213856, + 0.006217509973794222, + 0.008547874167561531, + 0.00782831571996212, + 0.006981017533689737, + 0.009205201640725136, + 0.008256879635155201, + 0.006022993009537458, + 0.007117633707821369, + 0.0250746738165617, + 0.01046412531286478, + 0.01671331748366356, + 0.003979165107011795, + 0.005380331538617611, + 0.004374522715806961, + 0.014932626858353615, + 0.04750174283981323, + 0.0172253604978323, + 0.01815691590309143, + 0.021241532638669014, + 0.014380870386958122, + 0.014786371029913425, + 0.02082696743309498, + 0.006190821994096041, + 0.008975930511951447, + 0.00830097496509552, + 0.0064757829532027245, + 0.017002327367663383, + 0.010231838561594486, + 0.00747816264629364, + 0.005158820189535618, + 0.007890661247074604, + 0.007073032669723034, + 0.008869740180671215, + 0.005942181218415499, + 0.012387359514832497, + 0.005714961793273687, + 0.0057396916672587395, + 0.004620380699634552, + 0.007247440982609987, + 0.01092034112662077, + 0.011208802461624146, + 0.008699622005224228, + 0.013700692914426327, + 0.012819982133805752, + 0.012118094600737095, + 0.009331898763775826, + 0.025044819340109825, + 0.017160197719931602, + 0.02162741683423519, + 0.017997082322835922, + 0.01250901073217392, + 0.02584940940141678, + 0.021166322752833366, + 0.02158808708190918, + 0.005379727575927973, + 0.005433026701211929, + 0.011236649006605148, + 0.006431858986616135, + 0.005167689174413681, + 0.005018334835767746, + 0.005278218071907759, + 0.005998590961098671, + 0.011463853530585766, + 0.005738632287830114, + 0.012845557183027267, + 0.010584961622953415, + 0.0071237715892493725, + 0.01219217386096716, + 0.007592097390443087, + 0.01009460911154747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.06462786346673965 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982308343052864 + ], + "opposite_logprob_tokens": [ + -3.529823064804077 + ], + "logprob_reward": -0.02982308343052864, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982308343052864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2886367440223694, + 0.1475609540939331, + 0.08247032761573792, + 0.09110821038484573, + 0.04647662118077278, + 0.0634586289525032, + 0.05077320337295532, + 0.08726371824741364, + 0.04589904844760895, + 0.09635254740715027 + ], + "credit_map": [ + 0.02560795657336712, + 0.03929166495800018, + 0.029059723019599915, + 0.033872440457344055, + 0.028341278433799744, + 0.06895051896572113, + 0.03552641719579697, + 0.027986722066998482, + 0.014615326188504696, + 0.020784562453627586, + 0.019084813073277473, + 0.023051591590046883, + 0.018640844151377678, + 0.015387621708214283, + 0.0181476678699255, + 0.01784851774573326, + 0.008599817752838135, + 0.005358738359063864, + 0.029313530772924423, + 0.006992224138230085, + 0.010557396337389946, + 0.004839475266635418, + 0.008714991621673107, + 0.00809414405375719, + 0.008649066090583801, + 0.034337736666202545, + 0.006306901574134827, + 0.005860903766006231, + 0.014060841873288155, + 0.009833402931690216, + 0.005440790671855211, + 0.0066185519099235535, + 0.006349639035761356, + 0.005790475755929947, + 0.0073732128366827965, + 0.004106666427105665, + 0.007060167845338583, + 0.003988574258983135, + 0.005931643303483725, + 0.005876239389181137, + 0.010005868971347809, + 0.008411251939833164, + 0.006888625677675009, + 0.004125879146158695, + 0.013501864857971668, + 0.008299507200717926, + 0.005023238714784384, + 0.007202382665127516, + 0.004138648975640535, + 0.009148893877863884, + 0.005728460382670164, + 0.007166686467826366, + 0.006574142724275589, + 0.004554571583867073, + 0.007674625143408775, + 0.005787170957773924, + 0.008543563075363636, + 0.00647697551175952, + 0.016260921955108643, + 0.017345625907182693, + 0.0094972038641572, + 0.008641771972179413, + 0.008333624340593815, + 0.012164019048213959, + 0.004432687535881996, + 0.0036909745540469885, + 0.007404915522783995, + 0.008206416852772236, + 0.005251547787338495, + 0.006906926166266203, + 0.005131103098392487, + 0.004874472972005606, + 0.0106542082503438, + 0.013523981906473637, + 0.01151314377784729, + 0.008869899436831474, + 0.01647394709289074, + 0.0128969457000494, + 0.010812385007739067, + 0.011608024127781391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02982308343052864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270582377910614 + ], + "opposite_logprob_tokens": [ + -2.127058267593384 + ], + "logprob_reward": -0.1270582377910614, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270582377910614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1446022242307663, + 0.19889864325523376, + 0.24000777304172516, + 0.07486449182033539, + 0.09424632787704468, + 0.03022170066833496, + 0.0199983362108469, + 0.07664237171411514, + 0.0493374802172184, + 0.07118058204650879 + ], + "credit_map": [ + 0.020635206252336502, + 0.014039838686585426, + 0.015697898343205452, + 0.024052424356341362, + 0.01353267952799797, + 0.029911302030086517, + 0.015723077580332756, + 0.011009813286364079, + 0.023771772161126137, + 0.01665649004280567, + 0.01753379963338375, + 0.03213510289788246, + 0.021571962162852287, + 0.045356374233961105, + 0.02057805471122265, + 0.0212950948625803, + 0.023875681683421135, + 0.02285854145884514, + 0.029655952006578445, + 0.031136764213442802, + 0.07506085932254791, + 0.02368924766778946, + 0.018675874918699265, + 0.015054859220981598, + 0.005786991212517023, + 0.018205653876066208, + 0.005796712823212147, + 0.009324527345597744, + 0.01101353857666254, + 0.009817993268370628, + 0.008177866227924824, + 0.006741211283951998, + 0.005826137028634548, + 0.019843971356749535, + 0.018142353743314743, + 0.008425061590969563, + 0.011665177531540394, + 0.005481116473674774, + 0.007618515752255917, + 0.017243999987840652, + 0.004566822666674852, + 0.003616317408159375, + 0.005323790945112705, + 0.0026822583749890327, + 0.004432557616382837, + 0.00336700351908803, + 0.0031913442071527243, + 0.0030416075605899096, + 0.0018974050180986524, + 0.0035822943318635225, + 0.0018736497731879354, + 0.0018533786060288548, + 0.0025391452945768833, + 0.0018879193812608719, + 0.0042449478060007095, + 0.002119597513228655, + 0.010865723714232445, + 0.007872064597904682, + 0.011543012224137783, + 0.010623888112604618, + 0.0055398535914719105, + 0.00853492971509695, + 0.010582988150417805, + 0.011079917661845684, + 0.004067696165293455, + 0.0057167853228747845, + 0.006849127821624279, + 0.006155508104711771, + 0.005392007064074278, + 0.006887800991535187, + 0.006511022802442312, + 0.007757536601275206, + 0.00593948969617486, + 0.005798266269266605, + 0.005682765040546656, + 0.005706181284040213, + 0.00893408339470625, + 0.014244559220969677, + 0.009853454306721687, + 0.01502179354429245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.1270582377910614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978199765086174 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.00978199765086174, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978199765086174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2079365849494934, + 0.18710333108901978, + 0.13701660931110382, + 0.10303153097629547, + 0.06527015566825867, + 0.0325116403400898, + 0.03906330093741417, + 0.1279856115579605, + 0.041919972747564316, + 0.05816136300563812 + ], + "credit_map": [ + 0.026047123596072197, + 0.027198517695069313, + 0.01801665499806404, + 0.02762208692729473, + 0.02634037472307682, + 0.04036715254187584, + 0.024543561041355133, + 0.017801091074943542, + 0.0237263236194849, + 0.0193388219922781, + 0.016196347773075104, + 0.02830003760755062, + 0.02236098237335682, + 0.03121848776936531, + 0.01847497560083866, + 0.02748733013868332, + 0.01402353122830391, + 0.011077477596700191, + 0.027529342100024223, + 0.007585127372294664, + 0.04314874857664108, + 0.011800434440374374, + 0.008869715966284275, + 0.012982207350432873, + 0.01023806817829609, + 0.03715706616640091, + 0.005917402915656567, + 0.009731242433190346, + 0.014528493396937847, + 0.0081772580742836, + 0.0075803338550031185, + 0.009701657108962536, + 0.0053497180342674255, + 0.020802123472094536, + 0.004355688113719225, + 0.004394636023789644, + 0.017558176070451736, + 0.005647525191307068, + 0.003215592121705413, + 0.003946688026189804, + 0.003485071239992976, + 0.004128004889935255, + 0.0055051203817129135, + 0.0031629218719899654, + 0.004792800638824701, + 0.005325718317180872, + 0.002529021818190813, + 0.0035829776898026466, + 0.004426822066307068, + 0.007140444591641426, + 0.003958410583436489, + 0.007339953910559416, + 0.00535886874422431, + 0.0035182207357138395, + 0.0035990807227790356, + 0.0037214939948171377, + 0.01296780351549387, + 0.010410968214273453, + 0.024831905961036682, + 0.01693342626094818, + 0.011598004028201103, + 0.013664028607308865, + 0.01818634197115898, + 0.01939311809837818, + 0.0039398870430886745, + 0.0040053012780845165, + 0.004714129958301783, + 0.008149630390107632, + 0.004542123526334763, + 0.006719193886965513, + 0.005517000798135996, + 0.004332700278609991, + 0.005484873428940773, + 0.005178054329007864, + 0.008478845469653606, + 0.006124348845332861, + 0.006093861069530249, + 0.010196390561759472, + 0.006469149142503738, + 0.010135835967957973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00978199765086174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486503504216671 + ], + "opposite_logprob_tokens": [ + -3.0486502647399902 + ], + "logprob_reward": -0.0486503504216671, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486503504216671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11243075877428055, + 0.0930219441652298, + 0.044390834867954254, + 0.12616100907325745, + 0.07617435604333878, + 0.02680809423327446, + 0.03161565586924553, + 0.06640166789293289, + 0.10742391645908356, + 0.3155718147754669 + ], + "credit_map": [ + 0.015825849026441574, + 0.011045749299228191, + 0.012222898192703724, + 0.01597132906317711, + 0.01087690144777298, + 0.02007521688938141, + 0.017128318548202515, + 0.009284496307373047, + 0.011059808544814587, + 0.008971106261014938, + 0.007795051671564579, + 0.013450562953948975, + 0.011155798099935055, + 0.02203177474439144, + 0.009241112507879734, + 0.009316731244325638, + 0.0034223967231810093, + 0.003893463406711817, + 0.011094216257333755, + 0.0036648535169661045, + 0.008105103857815266, + 0.0067656575702130795, + 0.0027111442759633064, + 0.004733996000140905, + 0.010576420463621616, + 0.04159140959382057, + 0.007313634268939495, + 0.01089393999427557, + 0.01870942860841751, + 0.01602390967309475, + 0.00955778919160366, + 0.011494479142129421, + 0.006254388950765133, + 0.00976775772869587, + 0.012830561958253384, + 0.00919574499130249, + 0.010784785263240337, + 0.007173195946961641, + 0.006878489628434181, + 0.013289430178701878, + 0.0038986920844763517, + 0.002464800141751766, + 0.00554479006677866, + 0.0028012555558234453, + 0.0035078264772892, + 0.0033322065137326717, + 0.002821523230522871, + 0.0024369992315769196, + 0.0030568677466362715, + 0.004298016428947449, + 0.0029655529651790857, + 0.003715361701324582, + 0.004250430501997471, + 0.005819289479404688, + 0.0035433147568255663, + 0.003966820891946554, + 0.008082269690930843, + 0.006410976406186819, + 0.011433879844844341, + 0.009828555397689342, + 0.005602909717708826, + 0.006136888638138771, + 0.009956994093954563, + 0.008949197828769684, + 0.01589171588420868, + 0.007475102320313454, + 0.03004744090139866, + 0.010936892591416836, + 0.009192259982228279, + 0.011176885105669498, + 0.011669573374092579, + 0.011034039780497551, + 0.022353699430823326, + 0.02818256802856922, + 0.025042731314897537, + 0.0236141849309206, + 0.04631591960787773, + 0.0654526874423027, + 0.041262682527303696, + 0.06334733217954636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0486503504216671 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person perspective shot places the viewer directly on the back of a high-speed motorcycle. Your hands, clad in black, full-fingered cycling gloves, grip the metal handlebars tightly as you gaze over the motorcycle's crimson windshield and dashboard as you speed along Kuala Lumpur's vast, multi-lane arterial. Directly ahead, the brake lights of a white car suddenly flash. Responding in a split second, you force your body to the left, and the world tilts dramatically. To avoid the car, the car's body swerves into the ground at a jarring angle, your left knee nearly scraping the rough tarmac. During this breathtaking tilt, the iconic Kuala Lumpur buildings in the background—the majestic Petronas Twin Towers and the towering Merdeka 118—are rendered in intense motion blur and wide-angle distortion, as if stretched to the edges of your vision by an invisible force. Their contours become distorted and elongated, transforming into a dazzling, flowing line alongside the surrounding cityscape. The entire picture is rendered through the unique imaging style of Leica lenses, presenting a signature oily texture and extremely rich and saturated colors. The soft evening light gives the metal body and glass curtain wall a gorgeous luster, perfectly capturing this adrenaline-rushing moment in a cinematic way that is full of extreme dynamic beauty and tense atmosphere.", + "seed": 2484, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_14/index_6127_seed_2484_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.050793860107660294, + "vlm_logprob": -0.050793860107660294, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.050793860107660294, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052725356072187424 + ], + "opposite_logprob_tokens": [ + -3.177725315093994 + ], + "logprob_reward": -0.052725356072187424, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.052725356072187424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11614321172237396, + 0.05262293666601181, + 0.0661039650440216, + 0.42457839846611023, + 0.055606089532375336, + 0.021790336817502975, + 0.09567227959632874, + 0.03609824925661087, + 0.06213084235787392, + 0.06925363838672638 + ], + "credit_map": [ + 0.01447248738259077, + 0.015457097440958023, + 0.012438776902854443, + 0.019461262971162796, + 0.02014351077377796, + 0.014018664136528969, + 0.01000499352812767, + 0.010146435350179672, + 0.005380339454859495, + 0.00701835798099637, + 0.009135845117270947, + 0.01088913157582283, + 0.0060941120609641075, + 0.003605695441365242, + 0.005526394117623568, + 0.0049730646423995495, + 0.009206927381455898, + 0.01681288331747055, + 0.005566184408962727, + 0.010401350446045399, + 0.0052297175861895084, + 0.00390963489189744, + 0.009873569943010807, + 0.00510370684787631, + 0.046694349497556686, + 0.04294368624687195, + 0.06267619878053665, + 0.06393882632255554, + 0.03740730881690979, + 0.022867431864142418, + 0.10635381937026978, + 0.041696783155202866, + 0.0065910532139241695, + 0.007759826257824898, + 0.0070634144358336926, + 0.007582233287394047, + 0.004468856379389763, + 0.004521842580288649, + 0.010596323758363724, + 0.007022540550678968, + 0.0028189318254590034, + 0.003230424365028739, + 0.0022231752518564463, + 0.0032133623026311398, + 0.0028305770829319954, + 0.0026325995568186045, + 0.003051064210012555, + 0.0017902035033330321, + 0.011763070710003376, + 0.018716448917984962, + 0.006443219259381294, + 0.00513997208327055, + 0.015618461184203625, + 0.007765074260532856, + 0.008647957816720009, + 0.021578079089522362, + 0.002455413108691573, + 0.00296240346506238, + 0.0029923098627477884, + 0.005123971030116081, + 0.002900637686252594, + 0.004589524585753679, + 0.009370969608426094, + 0.0057030245661735535, + 0.004766157828271389, + 0.005948423407971859, + 0.011694869957864285, + 0.012182465754449368, + 0.006306144408881664, + 0.0049286591820418835, + 0.007679474540054798, + 0.008624652400612831, + 0.004551223013550043, + 0.00767984613776207, + 0.013054467737674713, + 0.012134680524468422, + 0.00466261338442564, + 0.009180355817079544, + 0.007760317996144295, + 0.010230143554508686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.052725356072187424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054974157363176346 + ], + "opposite_logprob_tokens": [ + -2.92997407913208 + ], + "logprob_reward": -0.054974157363176346, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054974157363176346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1471264809370041, + 0.03414466604590416, + 0.04943880811333656, + 0.09724988043308258, + 0.05872543528676033, + 0.027744123712182045, + 0.04576722905039787, + 0.3277638554573059, + 0.06473518908023834, + 0.14730434119701385 + ], + "credit_map": [ + 0.01887403428554535, + 0.025852719321846962, + 0.014162044040858746, + 0.016589218750596046, + 0.0313689187169075, + 0.019448379054665565, + 0.011970573104918003, + 0.008860604837536812, + 0.004399486817419529, + 0.005381944589316845, + 0.005427921190857887, + 0.0046579260379076, + 0.0033204061910510063, + 0.0034770669881254435, + 0.00374068901874125, + 0.003739224513992667, + 0.006977444514632225, + 0.005271708592772484, + 0.006009043660014868, + 0.008401316590607166, + 0.004104136489331722, + 0.0051605720072984695, + 0.00767902797088027, + 0.005835558287799358, + 0.01117045059800148, + 0.0100051024928689, + 0.014071074314415455, + 0.013172621838748455, + 0.009909440763294697, + 0.005426331888884306, + 0.024292055517435074, + 0.009202801622450352, + 0.008964752778410912, + 0.008156303316354752, + 0.007605402730405331, + 0.006836999673396349, + 0.0067922696471214294, + 0.00495157903060317, + 0.008270525373518467, + 0.0071476055309176445, + 0.003450611140578985, + 0.0034846027847379446, + 0.0024309440050274134, + 0.0038605083245784044, + 0.0032838827464729548, + 0.003101736307144165, + 0.005371004343032837, + 0.0027608342934399843, + 0.004092869348824024, + 0.0040773628279566765, + 0.0051492913626134396, + 0.005088525824248791, + 0.005315762013196945, + 0.006325376685708761, + 0.006379485595971346, + 0.009338554926216602, + 0.016306135803461075, + 0.02098090946674347, + 0.02933703549206257, + 0.03401665389537811, + 0.023468144237995148, + 0.041023045778274536, + 0.11157240718603134, + 0.051059555262327194, + 0.005097493063658476, + 0.008533230051398277, + 0.010434879921376705, + 0.011030638590455055, + 0.007164532784372568, + 0.005797587335109711, + 0.00674567511305213, + 0.009931149892508984, + 0.01062476821243763, + 0.01536084059625864, + 0.012374348938465118, + 0.009048004634678364, + 0.01474328339099884, + 0.03224280849099159, + 0.01692027598619461, + 0.03599001467227936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.054974157363176346 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206105649471283 + ], + "opposite_logprob_tokens": [ + -2.812061071395874 + ], + "logprob_reward": -0.06206105649471283, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206105649471283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07697528600692749, + 0.035152725875377655, + 0.056319840252399445, + 0.07373091578483582, + 0.04792933166027069, + 0.03151187300682068, + 0.2151113897562027, + 0.058618608862161636, + 0.1569005697965622, + 0.24774940311908722 + ], + "credit_map": [ + 0.009324686601758003, + 0.009815387427806854, + 0.012417507357895374, + 0.013968119397759438, + 0.0066199223510921, + 0.008506430312991142, + 0.008006803691387177, + 0.008316430263221264, + 0.0035316848661750555, + 0.003674073377624154, + 0.004669831600040197, + 0.007417858578264713, + 0.005656467285007238, + 0.0036850273609161377, + 0.004004866350442171, + 0.002512921579182148, + 0.00910139735788107, + 0.01394414622336626, + 0.005060918163508177, + 0.009435780346393585, + 0.004936929792165756, + 0.004408829379826784, + 0.0051401834934949875, + 0.004291662946343422, + 0.009493504650890827, + 0.01106609869748354, + 0.012796124443411827, + 0.010185059159994125, + 0.0053251516073942184, + 0.00698971189558506, + 0.009123784489929676, + 0.00875148456543684, + 0.006360739003866911, + 0.006766812410205603, + 0.004794469568878412, + 0.004148585256189108, + 0.0096870893612504, + 0.004097629338502884, + 0.00568372942507267, + 0.006390280555933714, + 0.004576137755066156, + 0.003591600339859724, + 0.003870496992021799, + 0.0042994278483092785, + 0.003950939513742924, + 0.0025726717431098223, + 0.005032796878367662, + 0.003617804031819105, + 0.032645389437675476, + 0.05533715337514877, + 0.01506083831191063, + 0.010477038100361824, + 0.011740089394152164, + 0.016997354105114937, + 0.017483806237578392, + 0.05536973848938942, + 0.0037838073913007975, + 0.004184863995760679, + 0.00447813281789422, + 0.006015402264893055, + 0.00375964492559433, + 0.009559303522109985, + 0.013152782805263996, + 0.01368467602878809, + 0.011865130625665188, + 0.017749251797795296, + 0.029501210898160934, + 0.027758948504924774, + 0.01444320846349001, + 0.014510026201605797, + 0.015100120566785336, + 0.025972681120038033, + 0.015523761510848999, + 0.018521545454859734, + 0.05457363277673721, + 0.04638843983411789, + 0.014820883981883526, + 0.04173656925559044, + 0.02647448144853115, + 0.029710108414292336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.06206105649471283 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192361384630203 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192361384630203, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192361384630203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1817326843738556, + 0.08212180435657501, + 0.03524020314216614, + 0.26027587056159973, + 0.058183226734399796, + 0.03281690925359726, + 0.11843180656433105, + 0.09008561074733734, + 0.03798554465174675, + 0.10312636196613312 + ], + "credit_map": [ + 0.021318519487977028, + 0.02990669570863247, + 0.016717983409762383, + 0.017605112865567207, + 0.04187319055199623, + 0.030463092029094696, + 0.012722357176244259, + 0.01112572755664587, + 0.01116687897592783, + 0.010576185770332813, + 0.014055363833904266, + 0.010928955860435963, + 0.0062116459012031555, + 0.01024406123906374, + 0.010573090054094791, + 0.008365618996322155, + 0.005876168143004179, + 0.009180051274597645, + 0.0030284880194813013, + 0.004192352294921875, + 0.0036435562651604414, + 0.00305726146325469, + 0.003387195523828268, + 0.0028751289937645197, + 0.029394621029496193, + 0.025632459670305252, + 0.03444475308060646, + 0.04232466593384743, + 0.02306527830660343, + 0.0135103864595294, + 0.06613121926784515, + 0.02577248215675354, + 0.008252357132732868, + 0.005278690718114376, + 0.0087870629504323, + 0.007429048418998718, + 0.0050850980915129185, + 0.004946501459926367, + 0.010118664242327213, + 0.008285805583000183, + 0.004832626320421696, + 0.005127469077706337, + 0.00298672984354198, + 0.004392027854919434, + 0.005555250681936741, + 0.003269345499575138, + 0.004108896479010582, + 0.002544563263654709, + 0.015521054156124592, + 0.027539066970348358, + 0.006654511205852032, + 0.007213952951133251, + 0.008126425556838512, + 0.01189806405454874, + 0.01003746222704649, + 0.03144126757979393, + 0.005306621547788382, + 0.003528396598994732, + 0.00783543847501278, + 0.008611055091023445, + 0.005879556760191917, + 0.015288284979760647, + 0.025782190263271332, + 0.01785406656563282, + 0.003263116115704179, + 0.004318214952945709, + 0.0076891272328794, + 0.006766196805983782, + 0.0032782386988401413, + 0.0035209553316235542, + 0.0037786851171404123, + 0.005371008068323135, + 0.00655008340254426, + 0.01366940513253212, + 0.01853257603943348, + 0.018678832799196243, + 0.006966966670006514, + 0.014846397563815117, + 0.01258754450827837, + 0.011294547468423843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.014192361384630203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001636922359467 + ], + "opposite_logprob_tokens": [ + -2.695016384124756 + ], + "logprob_reward": -0.07001636922359467, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001636922359467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07738619297742844, + 0.058153364807367325, + 0.05486694723367691, + 0.24672409892082214, + 0.03811392933130264, + 0.034991614520549774, + 0.15510304272174835, + 0.049520086497068405, + 0.10777624696493149, + 0.1773644983768463 + ], + "credit_map": [ + 0.01120287086814642, + 0.010052499361336231, + 0.012184441089630127, + 0.009208559058606625, + 0.009868857450783253, + 0.008774375542998314, + 0.009694632142782211, + 0.006399951875209808, + 0.007050928194075823, + 0.0074867201037704945, + 0.009100508876144886, + 0.014210514724254608, + 0.003822371596470475, + 0.00358477677218616, + 0.007914021611213684, + 0.0049835266545414925, + 0.00723642110824585, + 0.0133132329210639, + 0.004812931641936302, + 0.00835260096937418, + 0.004797958768904209, + 0.004252907820045948, + 0.008350322023034096, + 0.003750570584088564, + 0.02855885587632656, + 0.017279978841543198, + 0.03581700474023819, + 0.03206615895032883, + 0.02373863384127617, + 0.017903991043567657, + 0.06602763384580612, + 0.025331862270832062, + 0.004242478869855404, + 0.004677991848438978, + 0.005289727821946144, + 0.006611432880163193, + 0.003860306227579713, + 0.0030665970407426357, + 0.0057240696623921394, + 0.004641323816031218, + 0.004202291369438171, + 0.003122220514342189, + 0.004404795356094837, + 0.005908323917537928, + 0.003662376431748271, + 0.0027221215423196554, + 0.006471371278166771, + 0.004498114809393883, + 0.022246111184358597, + 0.03521614521741867, + 0.012781310826539993, + 0.008575818501412868, + 0.009504774585366249, + 0.013543170876801014, + 0.013754479587078094, + 0.039481230080127716, + 0.0035066725686192513, + 0.0028642748948186636, + 0.004181104712188244, + 0.0042406232096254826, + 0.0055196331813931465, + 0.005873865447938442, + 0.01617092452943325, + 0.007162986323237419, + 0.007677226327359676, + 0.020236024633049965, + 0.015715401619672775, + 0.014333435334265232, + 0.009458223357796669, + 0.009147393517196178, + 0.008578294888138771, + 0.02263023890554905, + 0.013107303529977798, + 0.01576555147767067, + 0.022020485252141953, + 0.025595175102353096, + 0.01757347211241722, + 0.0402606725692749, + 0.021774042397737503, + 0.02126779407262802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.07001636922359467 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person perspective shot places the viewer directly on the back of a high-speed motorcycle. Your hands, clad in black, full-fingered cycling gloves, grip the metal handlebars tightly as you gaze over the motorcycle's crimson windshield and dashboard as you speed along Kuala Lumpur's vast, multi-lane arterial. Directly ahead, the brake lights of a white car suddenly flash. Responding in a split second, you force your body to the left, and the world tilts dramatically. To avoid the car, the car's body swerves into the ground at a jarring angle, your left knee nearly scraping the rough tarmac. During this breathtaking tilt, the iconic Kuala Lumpur buildings in the background—the majestic Petronas Twin Towers and the towering Merdeka 118—are rendered in intense motion blur and wide-angle distortion, as if stretched to the edges of your vision by an invisible force. Their contours become distorted and elongated, transforming into a dazzling, flowing line alongside the surrounding cityscape. The entire picture is rendered through the unique imaging style of Leica lenses, presenting a signature oily texture and extremely rich and saturated colors. The soft evening light gives the metal body and glass curtain wall a gorgeous luster, perfectly capturing this adrenaline-rushing moment in a cinematic way that is full of extreme dynamic beauty and tense atmosphere.", + "seed": 2485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_14/index_6127_seed_2485_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09904842488467694, + "vlm_logprob": -0.09904842488467694, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.09904842488467694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1795729249715805 + ], + "opposite_logprob_tokens": [ + -1.9295729398727417 + ], + "logprob_reward": -0.1795729249715805, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1795729249715805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4520741403102875, + 0.06897483766078949, + 0.09142819046974182, + 0.04561750963330269, + 0.08870622515678406, + 0.08189673721790314, + 0.0403381772339344, + 0.023809775710105896, + 0.0519961379468441, + 0.05515824630856514 + ], + "credit_map": [ + 0.034717243164777756, + 0.04953084886074066, + 0.06312794983386993, + 0.042636945843696594, + 0.0668131485581398, + 0.05850139260292053, + 0.07702898979187012, + 0.05971763655543327, + 0.008375126868486404, + 0.010080563835799694, + 0.0053644184954464436, + 0.005803493782877922, + 0.013013551011681557, + 0.01134994626045227, + 0.008297182619571686, + 0.006690559443086386, + 0.014017942361533642, + 0.014479836449027061, + 0.007738699670881033, + 0.011243230663239956, + 0.013482731766998768, + 0.011173921637237072, + 0.008769551292061806, + 0.010522276163101196, + 0.006309139542281628, + 0.004601383116096258, + 0.006429434288293123, + 0.004343495238572359, + 0.0055532315745949745, + 0.006882709451019764, + 0.005812852643430233, + 0.0056852675043046474, + 0.01134722214192152, + 0.010357762686908245, + 0.010281214490532875, + 0.009939756244421005, + 0.008839904330670834, + 0.021811392158269882, + 0.009436242282390594, + 0.006692732218652964, + 0.009967639110982418, + 0.0091236662119627, + 0.012443716637790203, + 0.008162159472703934, + 0.00820175651460886, + 0.023242173716425896, + 0.004931838251650333, + 0.0058237845078110695, + 0.004406558815389872, + 0.0051691485568881035, + 0.004369713831692934, + 0.005204958841204643, + 0.004143664613366127, + 0.007512663025408983, + 0.004371777176856995, + 0.005159697961062193, + 0.0031464502681046724, + 0.00407937727868557, + 0.002220441587269306, + 0.00298492843285203, + 0.0025860543828457594, + 0.0036112945526838303, + 0.0024892028886824846, + 0.0026920277159661055, + 0.006812361069023609, + 0.005011280998587608, + 0.004673643968999386, + 0.007160732988268137, + 0.005259426310658455, + 0.01002564374357462, + 0.007791453041136265, + 0.005261599086225033, + 0.006876063067466021, + 0.008508200757205486, + 0.007167609874159098, + 0.007056365720927715, + 0.006603955291211605, + 0.007987814024090767, + 0.006919588427990675, + 0.004038652870804071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.1795729249715805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209690868854523 + ], + "opposite_logprob_tokens": [ + -2.8120968341827393 + ], + "logprob_reward": -0.06209690868854523, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209690868854523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2544432282447815, + 0.15836666524410248, + 0.12332666665315628, + 0.11224145442247391, + 0.04091491550207138, + 0.06664790958166122, + 0.03316202014684677, + 0.04236287996172905, + 0.09069906175136566, + 0.07783520221710205 + ], + "credit_map": [ + 0.04055115953087807, + 0.03822115436196327, + 0.03681325539946556, + 0.021723998710513115, + 0.03455135226249695, + 0.030086016282439232, + 0.024481870234012604, + 0.028014441952109337, + 0.018825693055987358, + 0.02122003212571144, + 0.012533465400338173, + 0.01473306119441986, + 0.028242461383342743, + 0.026883311569690704, + 0.0206015482544899, + 0.015327103435993195, + 0.01356523111462593, + 0.01447243057191372, + 0.015808619558811188, + 0.014821110293269157, + 0.017672685906291008, + 0.01865035854279995, + 0.01510054711252451, + 0.013235685415565968, + 0.01668611727654934, + 0.014385715126991272, + 0.011424855329096317, + 0.007494975812733173, + 0.014547513797879219, + 0.029976462945342064, + 0.00868123397231102, + 0.009044581092894077, + 0.006769502069801092, + 0.006787485908716917, + 0.005942611489444971, + 0.004055216442793608, + 0.004953056573867798, + 0.004489026963710785, + 0.004401622340083122, + 0.003516396041959524, + 0.008383464068174362, + 0.008896506391465664, + 0.009014001116156578, + 0.008275331929326057, + 0.0065831998363137245, + 0.01712133176624775, + 0.003907324280589819, + 0.004466751590371132, + 0.005120833404362202, + 0.003633305663242936, + 0.0040269955061376095, + 0.0052558728493750095, + 0.0031412746757268906, + 0.0054762354120612144, + 0.003827652195468545, + 0.002679852768778801, + 0.005353302229195833, + 0.004913579672574997, + 0.005310472100973129, + 0.00507687171921134, + 0.004131498280912638, + 0.006395328789949417, + 0.004928676411509514, + 0.006253149826079607, + 0.01528309565037489, + 0.01596391759812832, + 0.010167863219976425, + 0.007533168885856867, + 0.013294730335474014, + 0.01604594849050045, + 0.005690222606062889, + 0.006720124743878841, + 0.010441343300044537, + 0.014177209697663784, + 0.009589500725269318, + 0.009332054294645786, + 0.009888140484690666, + 0.011980931274592876, + 0.007782330270856619, + 0.0046436963602900505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06209690868854523 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038444638252258 + ], + "opposite_logprob_tokens": [ + -1.9103844165802002 + ], + "logprob_reward": -0.16038444638252258, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038444638252258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22151756286621094, + 0.11452478915452957, + 0.05099835991859436, + 0.05488187074661255, + 0.0805804580450058, + 0.04478050023317337, + 0.03599834442138672, + 0.049005474895238876, + 0.25998860597610474, + 0.08772404491901398 + ], + "credit_map": [ + 0.020990340039134026, + 0.026440056040883064, + 0.02130941115319729, + 0.019864819943904877, + 0.03436720371246338, + 0.031104642897844315, + 0.03236529231071472, + 0.03507579118013382, + 0.012779762037098408, + 0.016315095126628876, + 0.013483059592545033, + 0.008508940227329731, + 0.021227924153208733, + 0.02069147303700447, + 0.011610510759055614, + 0.009908019565045834, + 0.00629496481269598, + 0.007248659618198872, + 0.004024472087621689, + 0.003899709554389119, + 0.01001153327524662, + 0.011560882441699505, + 0.004058621823787689, + 0.003899513278156519, + 0.008592558093369007, + 0.005752838682383299, + 0.00618766387924552, + 0.006552545353770256, + 0.006677040830254555, + 0.008343263529241085, + 0.0060889218002557755, + 0.0066870409063994884, + 0.00929477158933878, + 0.008093776181340218, + 0.004893356468528509, + 0.007193665485829115, + 0.00822585541754961, + 0.02576686441898346, + 0.009726373478770256, + 0.007385792210698128, + 0.005101148504763842, + 0.005596887785941362, + 0.008622282184660435, + 0.008128497749567032, + 0.003969521727412939, + 0.006329937372356653, + 0.0037944635841995478, + 0.003237761091440916, + 0.004481416195631027, + 0.00492296926677227, + 0.005336083937436342, + 0.0033907683100551367, + 0.00363959651440382, + 0.009451518766582012, + 0.0021897051483392715, + 0.0025862855836749077, + 0.008155493065714836, + 0.008045801892876625, + 0.004680747166275978, + 0.0047930567525327206, + 0.005438602063804865, + 0.008314083330333233, + 0.0045457943342626095, + 0.005031896755099297, + 0.06067642569541931, + 0.04207218065857887, + 0.03687795251607895, + 0.02596474066376686, + 0.03274806588888168, + 0.023475617170333862, + 0.017586177214980125, + 0.02058744989335537, + 0.007975547574460506, + 0.019048651680350304, + 0.008856981992721558, + 0.008961958810687065, + 0.013297544792294502, + 0.016028698533773422, + 0.007305198349058628, + 0.006249465513974428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.16038444638252258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199061319231987 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199061319231987, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199061319231987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20261119306087494, + 0.07765310257673264, + 0.07668719440698624, + 0.055201634764671326, + 0.11267193406820297, + 0.09510652720928192, + 0.055379949510097504, + 0.05006759241223335, + 0.16757524013519287, + 0.10704557597637177 + ], + "credit_map": [ + 0.026722658425569534, + 0.028670450672507286, + 0.030717244371771812, + 0.015727974474430084, + 0.035615552216768265, + 0.03262599930167198, + 0.01863907091319561, + 0.01389225758612156, + 0.008501020260155201, + 0.011828286573290825, + 0.007073362357914448, + 0.006175236776471138, + 0.016248976811766624, + 0.012316334992647171, + 0.007579685654491186, + 0.007930208928883076, + 0.008734597824513912, + 0.010728736408054829, + 0.009051892906427383, + 0.009732972830533981, + 0.010299671441316605, + 0.00783105194568634, + 0.010380806401371956, + 0.009927468374371529, + 0.01068895123898983, + 0.006518857553601265, + 0.005387575831264257, + 0.005030024331063032, + 0.007951351813971996, + 0.011122534982860088, + 0.003906898200511932, + 0.004595448728650808, + 0.015807032585144043, + 0.009654643014073372, + 0.009472720324993134, + 0.007340188603848219, + 0.01256730780005455, + 0.03861396387219429, + 0.00997896771878004, + 0.009237129241228104, + 0.012804451398551464, + 0.012001829221844673, + 0.013933556154370308, + 0.01203826256096363, + 0.008564037270843983, + 0.018918126821517944, + 0.009260733611881733, + 0.007585544604808092, + 0.006635093595832586, + 0.010869824327528477, + 0.006159374490380287, + 0.0038141063414514065, + 0.006717471405863762, + 0.013994693756103516, + 0.0036456547677516937, + 0.0035437336191534996, + 0.008441210724413395, + 0.008375869132578373, + 0.004380823113024235, + 0.004377860110253096, + 0.005861805751919746, + 0.009968559257686138, + 0.004768743645399809, + 0.003892728826031089, + 0.030875304713845253, + 0.032243646681308746, + 0.01750238984823227, + 0.02037346363067627, + 0.021770833060145378, + 0.018626302480697632, + 0.012627127580344677, + 0.01355619728565216, + 0.012255637906491756, + 0.017473220825195312, + 0.011360730975866318, + 0.012638104148209095, + 0.01290943194180727, + 0.01621585339307785, + 0.014341876842081547, + 0.009850730188190937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.014199061319231987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a first-person perspective from behind a high-speed motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rider's hands visible wearing black full-fingered cycling gloves gripping metal handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white car ahead brake and swerve into the ground at a jarring angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Petronas Twin Towers and Merdeka 118 visible in the background with intense motion blur and wide-angle distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a Leica lens style featuring an oily texture and extremely rich, saturated colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898878306150436 + ], + "opposite_logprob_tokens": [ + -2.578988790512085 + ], + "logprob_reward": -0.07898878306150436, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898878306150436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20574095845222473, + 0.07817590236663818, + 0.05361903831362724, + 0.05582062155008316, + 0.031283993273973465, + 0.025084763765335083, + 0.03493470326066017, + 0.026007112115621567, + 0.2713143229484558, + 0.2180185467004776 + ], + "credit_map": [ + 0.02930404432117939, + 0.02648823708295822, + 0.04428147152066231, + 0.014350448735058308, + 0.026845300570130348, + 0.027750765904784203, + 0.018111109733581543, + 0.018609587103128433, + 0.008356492035090923, + 0.010079307481646538, + 0.008274667896330357, + 0.007940120995044708, + 0.010360419750213623, + 0.006782631855458021, + 0.011951354332268238, + 0.014430910348892212, + 0.007896604016423225, + 0.007238505873829126, + 0.004804376047104597, + 0.005936977453529835, + 0.008233715780079365, + 0.00950995460152626, + 0.005098354537039995, + 0.004900551401078701, + 0.012035096064209938, + 0.007446133065968752, + 0.004988892935216427, + 0.004277472849935293, + 0.0076488954946398735, + 0.008151145651936531, + 0.006453805603086948, + 0.0048191845417022705, + 0.006667288485914469, + 0.003212104784324765, + 0.002985473023727536, + 0.0018499580910429358, + 0.004088450223207474, + 0.0072637214325368404, + 0.0023533122148364782, + 0.002863686764612794, + 0.003799582365900278, + 0.0029631105717271566, + 0.004750950261950493, + 0.0037891899701207876, + 0.0020810426212847233, + 0.003266352228820324, + 0.0027730371803045273, + 0.0016615006607025862, + 0.004385435953736305, + 0.0042439233511686325, + 0.003877433016896248, + 0.00419453764334321, + 0.0035961708053946495, + 0.00936761312186718, + 0.0028321559075266123, + 0.002437434857711196, + 0.003921143710613251, + 0.005009982269257307, + 0.002776080509647727, + 0.0027898713015019894, + 0.00301731051877141, + 0.003961370326578617, + 0.002253024373203516, + 0.002278328640386462, + 0.0643000602722168, + 0.03877021744847298, + 0.03402961418032646, + 0.026382356882095337, + 0.03558230400085449, + 0.031523868441581726, + 0.019586702808737755, + 0.02113921009004116, + 0.02221483178436756, + 0.04157218709588051, + 0.02685410901904106, + 0.024743758141994476, + 0.030607057735323906, + 0.038783274590969086, + 0.016089456155896187, + 0.0171539057046175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.07898878306150436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_15/index_tensor([7279, 7279])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_15/index_tensor([7279, 7279])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..60f25f77ea02af79f7a2c05a89d647abfd7f2ea4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_15/index_tensor([7279, 7279])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning wide, static, high-angle shot capturing Earth from a near-orbital perspective, a vast, ethereal atmosphere drapes across the planet's curve, emitting an inner glow. This translucent veil visualizes global temperature gradients in Makoto Shinkai's ultra-detailed and emotionally resonant style; vast expanses of fiery crimson and vibrant orange flow across the equator, seamlessly blending into gentle greens and yellows before dissolving into deep sapphire and violet blues at the polar caps. The entire layer undulates with a slow, hypnotic rhythm, like silk drifting in a cosmic current. As the Sun sinks below the planet's horizon, its bright, warm light casts vast, softly-edged shadows across the visualized surface, with terminal lines sweeping majestically across the continents. The scene is pierced by Shinkai's signature anamorphic lens flare, rendering the horizons in brilliant blues and oranges, creating a painterly, nostalgic effect. The atmosphere is one of profound, quiet awe, capturing a fleeting moment of cosmic beauty on a planetary scale, imbued with the melancholic yet hopeful emotionality characteristic of the artist's work.", + "seed": 10952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_15/index_7279_seed_10952_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4764574373140931, + "vlm_logprob": -0.4764574373140931, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.4764574373140931, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11806698143482208 + ], + "opposite_logprob_tokens": [ + -2.2430670261383057 + ], + "logprob_reward": -0.11806698143482208, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11806698143482208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07138759642839432, + 0.022792281582951546, + 0.07666762173175812, + 0.0820244625210762, + 0.3925107717514038, + 0.07243212312459946, + 0.05648370087146759, + 0.16246448457241058, + 0.02089070901274681, + 0.0423462875187397 + ], + "credit_map": [ + 0.0059501552022993565, + 0.00689759012311697, + 0.005264547187834978, + 0.013578295707702637, + 0.0056938910856842995, + 0.01476741861552, + 0.008603637106716633, + 0.010632057674229145, + 0.0018728983122855425, + 0.002305970061570406, + 0.0037707698065787554, + 0.0024212696589529514, + 0.0023588319309055805, + 0.004926860332489014, + 0.0026661348529160023, + 0.0024695477914065123, + 0.008550289086997509, + 0.008126200176775455, + 0.008310029283165932, + 0.011046557687222958, + 0.00664337957277894, + 0.012763799168169498, + 0.010549872182309628, + 0.010677486658096313, + 0.009218030609190464, + 0.008446433581411839, + 0.00992622785270214, + 0.013591059483587742, + 0.008250909857451916, + 0.01005049142986536, + 0.010677995160222054, + 0.011863312683999538, + 0.024891480803489685, + 0.04340774565935135, + 0.037744149565696716, + 0.07309894263744354, + 0.0352310985326767, + 0.04108467325568199, + 0.07480226457118988, + 0.06225040927529335, + 0.008789276704192162, + 0.01087124552577734, + 0.006943513639271259, + 0.011979887261986732, + 0.00781451165676117, + 0.008227093145251274, + 0.009105387143790722, + 0.008701215498149395, + 0.005123083014041185, + 0.007930787280201912, + 0.004926484078168869, + 0.007596462499350309, + 0.005841537844389677, + 0.005714701488614082, + 0.009227694012224674, + 0.010122948326170444, + 0.01856352388858795, + 0.019733518362045288, + 0.011429937556385994, + 0.027586501091718674, + 0.017004653811454773, + 0.020604077726602554, + 0.02246025763452053, + 0.025082021951675415, + 0.0018686325056478381, + 0.0020131398923695087, + 0.0014409482246264815, + 0.002839474007487297, + 0.0027950992807745934, + 0.0033284155651926994, + 0.003207877743989229, + 0.003397122723981738, + 0.004092348739504814, + 0.0051516396924853325, + 0.002478937851265073, + 0.007356738205999136, + 0.007356819696724415, + 0.006330717355012894, + 0.00443994952365756, + 0.005139135755598545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.11806698143482208 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026376964524388313 + ], + "opposite_logprob_tokens": [ + -3.651376962661743 + ], + "logprob_reward": -0.026376964524388313, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026376964524388313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06377632170915604, + 0.023948626592755318, + 0.09462974220514297, + 0.03360825404524803, + 0.4182160794734955, + 0.09690594673156738, + 0.0873769223690033, + 0.11665236949920654, + 0.0328480489552021, + 0.03203766793012619 + ], + "credit_map": [ + 0.005415352061390877, + 0.005892574787139893, + 0.007322162855416536, + 0.010454428382217884, + 0.007019120268523693, + 0.0059736003167927265, + 0.011357114650309086, + 0.010341964662075043, + 0.0026787708047777414, + 0.002558163832873106, + 0.0027784635312855244, + 0.003039516741409898, + 0.0027203198987990618, + 0.002840452827513218, + 0.0032042749226093292, + 0.004128662403672934, + 0.008797422051429749, + 0.009587850421667099, + 0.016433266922831535, + 0.01002024207264185, + 0.009616930969059467, + 0.025301838293671608, + 0.00644289143383503, + 0.008429302833974361, + 0.00451625045388937, + 0.0037107004318386316, + 0.003782661398872733, + 0.004826187621802092, + 0.0034514807630330324, + 0.004749920219182968, + 0.004222573712468147, + 0.004348477814346552, + 0.024646325036883354, + 0.052000075578689575, + 0.040109459310770035, + 0.08632636815309525, + 0.04385288059711456, + 0.04657939448952675, + 0.06457209587097168, + 0.06012946739792824, + 0.0094342902302742, + 0.013208013027906418, + 0.009130298160016537, + 0.01844024658203125, + 0.00976176280528307, + 0.010828505270183086, + 0.013226310722529888, + 0.01287651527673006, + 0.007822312414646149, + 0.010650141164660454, + 0.007341261021792889, + 0.01083868183195591, + 0.00906426552683115, + 0.01062619686126709, + 0.017383774742484093, + 0.013650286942720413, + 0.01408607792109251, + 0.013526008464396, + 0.008630428463220596, + 0.02142111398279667, + 0.012340313754975796, + 0.015222964808344841, + 0.017517851665616035, + 0.013907603919506073, + 0.003144301241263747, + 0.004535542335361242, + 0.0025020490866154432, + 0.005174551624804735, + 0.004711278714239597, + 0.004541667178273201, + 0.003832984948530793, + 0.0044056735932827, + 0.002783354138955474, + 0.004343127366155386, + 0.002532687271013856, + 0.005151763558387756, + 0.0049604312516748905, + 0.005135056562721729, + 0.002458496717736125, + 0.00467274896800518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.026376964524388313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260909907519817 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005260909907519817, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260909907519817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12956617772579193, + 0.04307983070611954, + 0.10578621178865433, + 0.04503714293241501, + 0.1425796002149582, + 0.12704555690288544, + 0.15269052982330322, + 0.1286982297897339, + 0.029265444725751877, + 0.09625127166509628 + ], + "credit_map": [ + 0.009037330746650696, + 0.017407676205039024, + 0.011135519482195377, + 0.024257207289338112, + 0.014246598817408085, + 0.01876050792634487, + 0.018636733293533325, + 0.016084598377346992, + 0.0035385333467274904, + 0.0030863916035741568, + 0.006923299748450518, + 0.004619306419044733, + 0.005313777830451727, + 0.007558181416243315, + 0.005491945892572403, + 0.00654839351773262, + 0.011618501506745815, + 0.009890764951705933, + 0.012779904529452324, + 0.016739333048462868, + 0.00813464168459177, + 0.016303420066833496, + 0.016993148252367973, + 0.013326501473784447, + 0.003873388282954693, + 0.005880485754460096, + 0.005502409767359495, + 0.0059688882902264595, + 0.007426498457789421, + 0.006712396163493395, + 0.0052601005882024765, + 0.004412971902638674, + 0.010875103063881397, + 0.019334113225340843, + 0.012132100760936737, + 0.024976173415780067, + 0.01460626907646656, + 0.018167652189731598, + 0.020590147003531456, + 0.021898038685321808, + 0.0096726780757308, + 0.01381615363061428, + 0.011706993915140629, + 0.023408664390444756, + 0.012926096096634865, + 0.01277363020926714, + 0.023839561268687248, + 0.018901778385043144, + 0.008161904290318489, + 0.01621764898300171, + 0.010782349854707718, + 0.020748522132635117, + 0.016217030584812164, + 0.01863834261894226, + 0.03209743648767471, + 0.029827294871211052, + 0.010332738049328327, + 0.013878460973501205, + 0.010499192401766777, + 0.022184111177921295, + 0.012552034109830856, + 0.014773353934288025, + 0.026738757267594337, + 0.01773957721889019, + 0.003435698803514242, + 0.004397402051836252, + 0.0024359910748898983, + 0.004969662055373192, + 0.004206744953989983, + 0.0033810760360211134, + 0.003534659743309021, + 0.0029042111709713936, + 0.007939854636788368, + 0.008787769824266434, + 0.005508891306817532, + 0.022019769996404648, + 0.016088644042611122, + 0.01363617554306984, + 0.011136777698993683, + 0.011133396066725254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.005260909907519817 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063634872436523 + ], + "opposite_logprob_tokens": [ + -0.2813635468482971 + ], + "logprob_reward": -1.4063634872436523, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063634872436523, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.057646993547677994, + 0.03687988966703415, + 0.037018824368715286, + 0.08845122158527374, + 0.21056067943572998, + 0.08614729344844818, + 0.09362197667360306, + 0.13300442695617676, + 0.08066600561141968, + 0.17600268125534058 + ], + "credit_map": [ + 0.006917618680745363, + 0.0065247295424342155, + 0.005673006176948547, + 0.008331555873155594, + 0.0054225497879087925, + 0.011246273294091225, + 0.0067505971528589725, + 0.006780659314244986, + 0.0022950356360524893, + 0.003648014273494482, + 0.00702533358708024, + 0.0026828197296708822, + 0.0032603112049400806, + 0.010126077570021152, + 0.003875328227877617, + 0.003966968972235918, + 0.0036468505859375, + 0.003704905742779374, + 0.004626998212188482, + 0.00533319590613246, + 0.0033663378562778234, + 0.005785820074379444, + 0.005536019802093506, + 0.005018695257604122, + 0.008851934224367142, + 0.006930698174983263, + 0.0107524199411273, + 0.013140866532921791, + 0.00855716597288847, + 0.01141126174479723, + 0.014331280253827572, + 0.01447559054940939, + 0.0167071633040905, + 0.027369016781449318, + 0.018441922962665558, + 0.03551751375198364, + 0.02246166579425335, + 0.024145852774381638, + 0.030568771064281464, + 0.03534875810146332, + 0.007672949228435755, + 0.012200310826301575, + 0.008326070383191109, + 0.013208921067416668, + 0.010447659529745579, + 0.010446429252624512, + 0.010128349997103214, + 0.013716605491936207, + 0.008622439578175545, + 0.012078200466930866, + 0.008328117430210114, + 0.014523521065711975, + 0.010770690627396107, + 0.009709270671010017, + 0.015222582034766674, + 0.014367159456014633, + 0.013768192380666733, + 0.017057714983820915, + 0.008245725184679031, + 0.026112578809261322, + 0.014822685159742832, + 0.018137453123927116, + 0.019490353763103485, + 0.015369724482297897, + 0.008989550173282623, + 0.011293799616396427, + 0.00654206145554781, + 0.011756920255720615, + 0.01040890533477068, + 0.012189190834760666, + 0.008640124462544918, + 0.010845459066331387, + 0.018086960539221764, + 0.02046586200594902, + 0.011307182721793652, + 0.028845148161053658, + 0.025415992364287376, + 0.015787916257977486, + 0.03222208470106125, + 0.023871535435318947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -1.4063634872436523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.826218843460083 + ], + "opposite_logprob_tokens": [ + -0.576218843460083 + ], + "logprob_reward": -0.826218843460083, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.826218843460083, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0659346804022789, + 0.03640415146946907, + 0.09026958793401718, + 0.03547964245080948, + 0.19187387824058533, + 0.10764593631029129, + 0.13802242279052734, + 0.11248219013214111, + 0.08479291200637817, + 0.13709460198879242 + ], + "credit_map": [ + 0.006865900941193104, + 0.007273525930941105, + 0.007164218928664923, + 0.00965430960059166, + 0.0056927986443042755, + 0.009677983820438385, + 0.011021735146641731, + 0.008584209717810154, + 0.003663363866508007, + 0.003006849205121398, + 0.007268643472343683, + 0.003427013987675309, + 0.0031266508158296347, + 0.006959801539778709, + 0.004702186677604914, + 0.00424963841214776, + 0.007545854896306992, + 0.008676549419760704, + 0.01796286180615425, + 0.008529962040483952, + 0.010618026368319988, + 0.023690328001976013, + 0.006314931903034449, + 0.006931071635335684, + 0.003920805174857378, + 0.004512271843850613, + 0.003903251141309738, + 0.00612526573240757, + 0.004816386848688126, + 0.00416809506714344, + 0.004143730271607637, + 0.003889836138114333, + 0.021912233904004097, + 0.023075347766280174, + 0.01847437396645546, + 0.030007848516106606, + 0.020787091925740242, + 0.020751401782035828, + 0.024252120405435562, + 0.03261345624923706, + 0.008202140219509602, + 0.015526840463280678, + 0.010517548769712448, + 0.014937590807676315, + 0.013154536485671997, + 0.011044626124203205, + 0.016377687454223633, + 0.017884964123368263, + 0.0140256118029356, + 0.021587589755654335, + 0.009611767716705799, + 0.015838341787457466, + 0.017347248271107674, + 0.01984519511461258, + 0.019733794033527374, + 0.02003287523984909, + 0.013370650820434093, + 0.01418957021087408, + 0.008763223886489868, + 0.01769765466451645, + 0.012072492390871048, + 0.014396091923117638, + 0.015971388667821884, + 0.0160211268812418, + 0.00962562020868063, + 0.010730101726949215, + 0.007814362645149231, + 0.011366207152605057, + 0.011661491356790066, + 0.008979501202702522, + 0.01294697355479002, + 0.011668655090034008, + 0.011838164180517197, + 0.01877068169414997, + 0.01135424617677927, + 0.019618691876530647, + 0.021175285801291466, + 0.02815915271639824, + 0.008256115019321442, + 0.01792227104306221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.826218843460083 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + }, + { + "prompt": "In a stunning wide, static, high-angle shot capturing Earth from a near-orbital perspective, a vast, ethereal atmosphere drapes across the planet's curve, emitting an inner glow. This translucent veil visualizes global temperature gradients in Makoto Shinkai's ultra-detailed and emotionally resonant style; vast expanses of fiery crimson and vibrant orange flow across the equator, seamlessly blending into gentle greens and yellows before dissolving into deep sapphire and violet blues at the polar caps. The entire layer undulates with a slow, hypnotic rhythm, like silk drifting in a cosmic current. As the Sun sinks below the planet's horizon, its bright, warm light casts vast, softly-edged shadows across the visualized surface, with terminal lines sweeping majestically across the continents. The scene is pierced by Shinkai's signature anamorphic lens flare, rendering the horizons in brilliant blues and oranges, creating a painterly, nostalgic effect. The atmosphere is one of profound, quiet awe, capturing a fleeting moment of cosmic beauty on a planetary scale, imbued with the melancholic yet hopeful emotionality characteristic of the artist's work.", + "seed": 10953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_15/index_7279_seed_10953_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.37230877596884965, + "vlm_logprob": -0.37230877596884965, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.37230877596884965, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016391737386584282 + ], + "opposite_logprob_tokens": [ + -4.516391754150391 + ], + "logprob_reward": -0.016391737386584282, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.016391737386584282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10897047817707062, + 0.06545614451169968, + 0.10778554528951645, + 0.08586006611585617, + 0.09609644114971161, + 0.14201456308364868, + 0.058593545109033585, + 0.12205943465232849, + 0.08686058968305588, + 0.12630322575569153 + ], + "credit_map": [ + 0.012943725101649761, + 0.02354360744357109, + 0.013871416449546814, + 0.010414251126348972, + 0.01647690124809742, + 0.009039399214088917, + 0.012083093635737896, + 0.010598072782158852, + 0.006036026868969202, + 0.011701483279466629, + 0.007343695033341646, + 0.007467167917639017, + 0.015826057642698288, + 0.00484185665845871, + 0.0069634513929486275, + 0.005276397801935673, + 0.011958161368966103, + 0.010206888429820538, + 0.016206789761781693, + 0.015494560822844505, + 0.01201121136546135, + 0.012537242844700813, + 0.012386969290673733, + 0.01698371022939682, + 0.007306132465600967, + 0.013957729563117027, + 0.009514990262687206, + 0.01301202829927206, + 0.012954686768352985, + 0.009471192955970764, + 0.008025440387427807, + 0.011617864482104778, + 0.007085799239575863, + 0.01958978734910488, + 0.012654483318328857, + 0.013047625310719013, + 0.012435088865458965, + 0.010931966826319695, + 0.01049666665494442, + 0.009855016134679317, + 0.007899606600403786, + 0.01724034547805786, + 0.025709904730319977, + 0.012196841649711132, + 0.0382874421775341, + 0.008261571638286114, + 0.02139803022146225, + 0.011020808480679989, + 0.00833235215395689, + 0.00801775511354208, + 0.006204952020198107, + 0.00795538630336523, + 0.00814779382199049, + 0.006290733348578215, + 0.007612594868987799, + 0.006031975615769625, + 0.015755847096443176, + 0.0133430864661932, + 0.012741764076054096, + 0.012142480351030827, + 0.013696110807359219, + 0.012804781086742878, + 0.018036944791674614, + 0.023538412526249886, + 0.00874103233218193, + 0.0061861611902713776, + 0.007055981550365686, + 0.01027283538132906, + 0.008665218949317932, + 0.022838864475488663, + 0.010706460103392601, + 0.012394028715789318, + 0.01613524928689003, + 0.012771005742251873, + 0.012840653769671917, + 0.014418291859328747, + 0.013111976906657219, + 0.022511983290314674, + 0.02073138765990734, + 0.013782665133476257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.016391737386584282 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034946352243423 + ], + "opposite_logprob_tokens": [ + -2.3503494262695312 + ], + "logprob_reward": -0.10034946352243423, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034946352243423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10161591321229935, + 0.08470078557729721, + 0.11175063997507095, + 0.05087839439511299, + 0.16059242188930511, + 0.2213439643383026, + 0.03956218063831329, + 0.038571760058403015, + 0.14018386602401733, + 0.05080007016658783 + ], + "credit_map": [ + 0.008484785445034504, + 0.015941806137561798, + 0.017765499651432037, + 0.009873280301690102, + 0.011562284082174301, + 0.008547797799110413, + 0.01980985887348652, + 0.009630600921809673, + 0.008000495843589306, + 0.011726375669240952, + 0.016154218465089798, + 0.012319693341851234, + 0.013593155890703201, + 0.0064839827828109264, + 0.00941805075854063, + 0.007004817947745323, + 0.014363856986165047, + 0.011908920481801033, + 0.013848857954144478, + 0.014675706624984741, + 0.013696187175810337, + 0.01379158440977335, + 0.01213783584535122, + 0.0173276886343956, + 0.0037373045925050974, + 0.004076467361301184, + 0.00822172500193119, + 0.0088987797498703, + 0.006138788070529699, + 0.004904801957309246, + 0.007693881168961525, + 0.007206643931567669, + 0.00828612595796585, + 0.02468760870397091, + 0.018398655578494072, + 0.026862798258662224, + 0.016765901818871498, + 0.013227659277617931, + 0.03559663146734238, + 0.016767039895057678, + 0.006744181737303734, + 0.022852106019854546, + 0.00933557003736496, + 0.010473936796188354, + 0.06328582763671875, + 0.012983341701328754, + 0.06773167103528976, + 0.02793734148144722, + 0.004858218599110842, + 0.003675202140584588, + 0.005928785540163517, + 0.0054499125108122826, + 0.003907620906829834, + 0.004312081728130579, + 0.005411062389612198, + 0.006019295658916235, + 0.005282829515635967, + 0.0035879455972462893, + 0.004462168086320162, + 0.005809833761304617, + 0.0032581675332039595, + 0.004039348568767309, + 0.00535109406337142, + 0.006780373398214579, + 0.01279057189822197, + 0.010765050537884235, + 0.011781190522015095, + 0.012072442099452019, + 0.01779259741306305, + 0.042238544672727585, + 0.014812388457357883, + 0.017931073904037476, + 0.005009587854146957, + 0.005064962897449732, + 0.0061806910671293736, + 0.006497928407043219, + 0.007976637221872807, + 0.006811541970819235, + 0.0063100275583565235, + 0.006948695983737707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.10034946352243423 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553836219012737 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553836219012737, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553836219012737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08163942396640778, + 0.06848292797803879, + 0.08830659091472626, + 0.06884969025850296, + 0.08353202790021896, + 0.4303651750087738, + 0.04654891788959503, + 0.030993202701210976, + 0.05208253115415573, + 0.04919953644275665 + ], + "credit_map": [ + 0.008365926332771778, + 0.015194263309240341, + 0.013092146255075932, + 0.009161490015685558, + 0.01079823262989521, + 0.007376333698630333, + 0.00964843388646841, + 0.00800259504467249, + 0.005068118218332529, + 0.012109548784792423, + 0.00890731904655695, + 0.008938908576965332, + 0.01617937907576561, + 0.004379601217806339, + 0.007508035749197006, + 0.005392017308622599, + 0.009547729045152664, + 0.009610386565327644, + 0.018180131912231445, + 0.009052149020135403, + 0.008583229035139084, + 0.008911676704883575, + 0.008958376944065094, + 0.015462917275726795, + 0.005423080176115036, + 0.00885955709964037, + 0.006715267896652222, + 0.008544404990971088, + 0.012957406230270863, + 0.009089468978345394, + 0.007482015993446112, + 0.009778491221368313, + 0.004926084540784359, + 0.026204707100987434, + 0.006305989809334278, + 0.008398741483688354, + 0.014554129913449287, + 0.005583088379353285, + 0.005918240174651146, + 0.01164104975759983, + 0.013440994545817375, + 0.04702313244342804, + 0.015869369730353355, + 0.01938937045633793, + 0.13215519487857819, + 0.025465015321969986, + 0.12946921586990356, + 0.04755287989974022, + 0.00600070646032691, + 0.005594590678811073, + 0.008474322035908699, + 0.007437449414283037, + 0.003972683101892471, + 0.0038279362488538027, + 0.00668551167473197, + 0.004555716644972563, + 0.0031853795517235994, + 0.0029882777016609907, + 0.0035760218743234873, + 0.005087301600724459, + 0.00418162951245904, + 0.003468871582299471, + 0.004794664215296507, + 0.003711054800078273, + 0.008719987235963345, + 0.0035864352248609066, + 0.005729012656956911, + 0.008932022377848625, + 0.005440187640488148, + 0.006718933116644621, + 0.006071353331208229, + 0.006884599104523659, + 0.005153516307473183, + 0.005748094525188208, + 0.00428807595744729, + 0.004669297020882368, + 0.009439593181014061, + 0.006461889483034611, + 0.007331992499530315, + 0.0061070737428963184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.012553836219012737 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743750691413879 + ], + "opposite_logprob_tokens": [ + -0.47437506914138794 + ], + "logprob_reward": -0.9743750691413879, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743750691413879, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17521022260189056, + 0.04561980068683624, + 0.16959136724472046, + 0.07819394022226334, + 0.12369658052921295, + 0.21297357976436615, + 0.06789832562208176, + 0.037942588329315186, + 0.06180798262357712, + 0.02706565335392952 + ], + "credit_map": [ + 0.017930712550878525, + 0.03661974146962166, + 0.018973762169480324, + 0.019562626257538795, + 0.030436884611845016, + 0.009089328348636627, + 0.026208287104964256, + 0.016388889402151108, + 0.005329975858330727, + 0.006209212820976973, + 0.008733660914003849, + 0.005753221455961466, + 0.00675911083817482, + 0.004137173295021057, + 0.0051391227170825005, + 0.003558320226147771, + 0.015495335683226585, + 0.017157379537820816, + 0.030765805393457413, + 0.020121054723858833, + 0.01886570081114769, + 0.021430425345897675, + 0.017527293413877487, + 0.02822836861014366, + 0.006484030745923519, + 0.0083799809217453, + 0.010921292006969452, + 0.013108564540743828, + 0.008142183534801006, + 0.011580289341509342, + 0.008970558643341064, + 0.010607040487229824, + 0.009931804612278938, + 0.04081524536013603, + 0.014816326089203358, + 0.010176535695791245, + 0.017614087089896202, + 0.006818658672273159, + 0.0075774542056024075, + 0.015946468338370323, + 0.007797902915626764, + 0.027246737852692604, + 0.01639341562986374, + 0.008813654072582722, + 0.06703269481658936, + 0.017063923180103302, + 0.020226584747433662, + 0.04839865118265152, + 0.007767209783196449, + 0.00529443146660924, + 0.008974413387477398, + 0.010124250315129757, + 0.006997715216130018, + 0.0055834101513028145, + 0.011384562589228153, + 0.011772338300943375, + 0.003986044321209192, + 0.003833261551335454, + 0.004323308821767569, + 0.0053729950450360775, + 0.0028803381137549877, + 0.0032057061325758696, + 0.004524353425949812, + 0.009816582314670086, + 0.005364210810512304, + 0.006151924375444651, + 0.004966633394360542, + 0.005866168066859245, + 0.007984892465174198, + 0.018702881410717964, + 0.006134382914751768, + 0.006636888720095158, + 0.002543815178796649, + 0.002779699396342039, + 0.002895461395382881, + 0.0026569634210318327, + 0.004135381430387497, + 0.00311101833358407, + 0.003691798774525523, + 0.005251515656709671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.9743750691413879 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578737735748291 + ], + "opposite_logprob_tokens": [ + -0.6328737735748291 + ], + "logprob_reward": -0.7578737735748291, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578737735748291, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19850265979766846, + 0.04010557010769844, + 0.05566895753145218, + 0.0873427465558052, + 0.12254354357719421, + 0.253385990858078, + 0.06175290048122406, + 0.03878083825111389, + 0.10844744741916656, + 0.03346936032176018 + ], + "credit_map": [ + 0.01848609559237957, + 0.0443718284368515, + 0.02482929825782776, + 0.024031441658735275, + 0.03541113808751106, + 0.010228087194263935, + 0.0246278028935194, + 0.016516968607902527, + 0.003276497358456254, + 0.007196452002972364, + 0.005589252803474665, + 0.005523900967091322, + 0.006713040638715029, + 0.003085123607888818, + 0.005734252743422985, + 0.002987048588693142, + 0.005761881358921528, + 0.00632175849750638, + 0.00835595466196537, + 0.007416823413223028, + 0.006610504351556301, + 0.005673789419233799, + 0.007258231285959482, + 0.008270017802715302, + 0.008480026386678219, + 0.009893476963043213, + 0.011038324795663357, + 0.011708751320838928, + 0.013574942946434021, + 0.01286336313933134, + 0.00967092253267765, + 0.010112945921719074, + 0.00807760376483202, + 0.028598008677363396, + 0.01819239929318428, + 0.01372613850980997, + 0.016227340325713158, + 0.010695516131818295, + 0.01721528172492981, + 0.00981125608086586, + 0.007551365997642279, + 0.021573461592197418, + 0.01171646174043417, + 0.012135002762079239, + 0.057242151349782944, + 0.015840832144021988, + 0.08198823034763336, + 0.045338474214076996, + 0.007963702082633972, + 0.006703777238726616, + 0.008534780703485012, + 0.008846452459692955, + 0.007772543467581272, + 0.006539401598274708, + 0.008203870616853237, + 0.007188370916992426, + 0.004941293504089117, + 0.00301549956202507, + 0.005205231253057718, + 0.0065922727808356285, + 0.004297942388802767, + 0.004372023977339268, + 0.005091285798698664, + 0.005265288520604372, + 0.008544858545064926, + 0.009339617565274239, + 0.013506502844393253, + 0.012349399738013744, + 0.017344864085316658, + 0.01389552652835846, + 0.013186892494559288, + 0.0202797818928957, + 0.003438622923567891, + 0.002457564463838935, + 0.003799203084781766, + 0.004129454027861357, + 0.004400165751576424, + 0.0034890095703303814, + 0.005417891312390566, + 0.006337449420243502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.7578737735748291 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + }, + { + "prompt": "In a stunning wide, static, high-angle shot capturing Earth from a near-orbital perspective, a vast, ethereal atmosphere drapes across the planet's curve, emitting an inner glow. This translucent veil visualizes global temperature gradients in Makoto Shinkai's ultra-detailed and emotionally resonant style; vast expanses of fiery crimson and vibrant orange flow across the equator, seamlessly blending into gentle greens and yellows before dissolving into deep sapphire and violet blues at the polar caps. The entire layer undulates with a slow, hypnotic rhythm, like silk drifting in a cosmic current. As the Sun sinks below the planet's horizon, its bright, warm light casts vast, softly-edged shadows across the visualized surface, with terminal lines sweeping majestically across the continents. The scene is pierced by Shinkai's signature anamorphic lens flare, rendering the horizons in brilliant blues and oranges, creating a painterly, nostalgic effect. The atmosphere is one of profound, quiet awe, capturing a fleeting moment of cosmic beauty on a planetary scale, imbued with the melancholic yet hopeful emotionality characteristic of the artist's work.", + "seed": 10954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_15/index_7279_seed_10954_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.6685750854201615, + "vlm_logprob": -0.6685750854201615, + "vlm_yes_no_margin": 2.0, + "vlm_token_credit": -0.6685750854201615, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024237047880887985 + ], + "opposite_logprob_tokens": [ + -3.8992371559143066 + ], + "logprob_reward": -0.024237047880887985, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024237047880887985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10022240877151489, + 0.14076541364192963, + 0.10113517194986343, + 0.0707576647400856, + 0.04431662708520889, + 0.062054891139268875, + 0.22722935676574707, + 0.1303836852312088, + 0.0875512957572937, + 0.03558347746729851 + ], + "credit_map": [ + 0.008711813017725945, + 0.011984148062765598, + 0.011523066088557243, + 0.016719648614525795, + 0.009367826394736767, + 0.009846558794379234, + 0.013237321749329567, + 0.018832024186849594, + 0.008746180683374405, + 0.0114193856716156, + 0.010449239984154701, + 0.039584044367074966, + 0.011502391658723354, + 0.007861797697842121, + 0.03395466133952141, + 0.017247704789042473, + 0.010310184210538864, + 0.011169441975653172, + 0.010157291777431965, + 0.012777823023498058, + 0.011944019235670567, + 0.018035920336842537, + 0.01335117407143116, + 0.013389312662184238, + 0.008368990384042263, + 0.008487485349178314, + 0.007656923495233059, + 0.007624645251780748, + 0.007878272794187069, + 0.01118729542940855, + 0.0071986159309744835, + 0.012355431914329529, + 0.006271115504205227, + 0.006502002943307161, + 0.0030979143921285868, + 0.005379310809075832, + 0.006141379941254854, + 0.006029374897480011, + 0.00423874706029892, + 0.0066567822359502316, + 0.008907927200198174, + 0.009439329616725445, + 0.005337078589946032, + 0.007137381937354803, + 0.008518972434103489, + 0.009177453815937042, + 0.0073614055290818214, + 0.006175335496664047, + 0.013809412717819214, + 0.022761007770895958, + 0.015564445406198502, + 0.022254982963204384, + 0.02672441676259041, + 0.031149601563811302, + 0.05197228491306305, + 0.04299319162964821, + 0.01173820998519659, + 0.015311193652451038, + 0.013562200590968132, + 0.01374233327805996, + 0.024551890790462494, + 0.020364807918667793, + 0.01523816678673029, + 0.01587488129734993, + 0.00802780780941248, + 0.008609007112681866, + 0.009457163512706757, + 0.014682787470519543, + 0.010763765312731266, + 0.011235371232032776, + 0.012754416093230247, + 0.012020979076623917, + 0.0033964598551392555, + 0.004854454193264246, + 0.004501383285969496, + 0.0032233644742518663, + 0.0059881992638111115, + 0.00473535293713212, + 0.004921024665236473, + 0.003963237628340721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.024237047880887985 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603911951184273 + ], + "opposite_logprob_tokens": [ + -3.895603895187378 + ], + "logprob_reward": -0.020603911951184273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603911951184273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22969740629196167, + 0.08429230749607086, + 0.1452818512916565, + 0.09392672777175903, + 0.0653800368309021, + 0.04588688164949417, + 0.035347361117601395, + 0.0677596777677536, + 0.17469948530197144, + 0.05772823467850685 + ], + "credit_map": [ + 0.020336639136075974, + 0.023508209735155106, + 0.029318736866116524, + 0.03313155844807625, + 0.024059420451521873, + 0.02241106703877449, + 0.030254902318120003, + 0.046676892787218094, + 0.006468883249908686, + 0.007883338257670403, + 0.007004403043538332, + 0.02321912720799446, + 0.0065641929395496845, + 0.005751121789216995, + 0.018094895407557487, + 0.009306344203650951, + 0.01524090114980936, + 0.015354400500655174, + 0.016796711832284927, + 0.021562181413173676, + 0.016770360991358757, + 0.012913175858557224, + 0.019860627129673958, + 0.026783499866724014, + 0.009938610717654228, + 0.0105953523889184, + 0.008241299539804459, + 0.013978464528918266, + 0.013094061054289341, + 0.011185579001903534, + 0.012912142090499401, + 0.013981221243739128, + 0.008933928795158863, + 0.008251640945672989, + 0.005232870113104582, + 0.006805852521210909, + 0.009003803133964539, + 0.011597341857850552, + 0.0065981303341686726, + 0.008956470526754856, + 0.004493287764489651, + 0.005897863768041134, + 0.004659414757043123, + 0.004957936238497496, + 0.007199200801551342, + 0.005535497795790434, + 0.005925123579800129, + 0.007218553684651852, + 0.0030924498569220304, + 0.004384732339531183, + 0.002212060848250985, + 0.004245130345225334, + 0.004617258440703154, + 0.0033790413290262222, + 0.007542415522038937, + 0.005874270107597113, + 0.005882703233510256, + 0.008819732815027237, + 0.006824997253715992, + 0.006453271489590406, + 0.012767349369823933, + 0.006868012715131044, + 0.009285809472203255, + 0.010857808403670788, + 0.013887198641896248, + 0.016930749639868736, + 0.02170645259320736, + 0.02144559659063816, + 0.02531798928976059, + 0.019084377214312553, + 0.02528993971645832, + 0.03103717416524887, + 0.0046321251429617405, + 0.006070469040423632, + 0.009080381132662296, + 0.00817691721022129, + 0.0069437213242053986, + 0.006247207522392273, + 0.00938408076763153, + 0.00719333253800869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.020603911951184273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076359291560947895 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.0076359291560947895, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076359291560947895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06166218966245651, + 0.027425428852438927, + 0.04093703255057335, + 0.08080189675092697, + 0.08792979270219803, + 0.08822932839393616, + 0.24214833974838257, + 0.10774888843297958, + 0.21432814002037048, + 0.048788949847221375 + ], + "credit_map": [ + 0.0066752382554113865, + 0.00904932152479887, + 0.009247801266610622, + 0.006736990064382553, + 0.008277234621345997, + 0.00566874910145998, + 0.007953234016895294, + 0.008053623139858246, + 0.0017906292341649532, + 0.0037832234520465136, + 0.0032392144203186035, + 0.004349700175225735, + 0.00355801684781909, + 0.002348248613998294, + 0.004225139506161213, + 0.004131256137043238, + 0.003095868742093444, + 0.005820067133754492, + 0.004184285178780556, + 0.006121950689703226, + 0.0054440065287053585, + 0.0043710065074265, + 0.006087249610573053, + 0.005812600255012512, + 0.00999168585985899, + 0.009785630740225315, + 0.007858113385736942, + 0.009402568452060223, + 0.011895149014890194, + 0.013190848752856255, + 0.007767752278596163, + 0.010910152457654476, + 0.013872502371668816, + 0.010974563658237457, + 0.007316937670111656, + 0.010143168270587921, + 0.014643446542322636, + 0.011520820669829845, + 0.010533982887864113, + 0.008924370631575584, + 0.010848842561244965, + 0.010511563159525394, + 0.007482170592993498, + 0.010815555229783058, + 0.01610534079372883, + 0.010306871496140957, + 0.010778767988085747, + 0.011380217038094997, + 0.015021881088614464, + 0.023457178846001625, + 0.01604773849248886, + 0.0247122123837471, + 0.02696102485060692, + 0.02883205935359001, + 0.05916505306959152, + 0.04795120656490326, + 0.010111702606081963, + 0.012087224051356316, + 0.009503919631242752, + 0.011690116487443447, + 0.018845222890377045, + 0.016701512038707733, + 0.013456336222589016, + 0.015352851711213589, + 0.017803041264414787, + 0.021104982122778893, + 0.024196326732635498, + 0.02437175065279007, + 0.03303822875022888, + 0.022797895595431328, + 0.033770207315683365, + 0.03724570944905281, + 0.0055752163752913475, + 0.005350484512746334, + 0.006302888039499521, + 0.0051819030195474625, + 0.007030573207885027, + 0.006235672160983086, + 0.006469074636697769, + 0.006643133237957954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.0076359291560947895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.4640395641326904 + ], + "opposite_logprob_tokens": [ + -0.08903957158327103 + ], + "logprob_reward": -2.4640395641326904, + "yes_no_margin_reward": -2.375, + "token_credit_reward": -2.4640395641326904, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11938732117414474, + 0.038271497935056686, + 0.054103925824165344, + 0.12759803235530853, + 0.10159996896982193, + 0.0950302705168724, + 0.19494937360286713, + 0.07194097340106964, + 0.1270238757133484, + 0.0700947716832161 + ], + "credit_map": [ + 0.010373545810580254, + 0.016291992738842964, + 0.013180173002183437, + 0.017794519662857056, + 0.015391496010124683, + 0.009562897495925426, + 0.016590768471360207, + 0.020201927050948143, + 0.004271665122359991, + 0.006115475669503212, + 0.004985895939171314, + 0.00407974561676383, + 0.005678135436028242, + 0.003599227638915181, + 0.0043169111013412476, + 0.00522444024682045, + 0.005841765087097883, + 0.006267216056585312, + 0.006339015439152718, + 0.009489973075687885, + 0.005057741887867451, + 0.004492986015975475, + 0.008235190995037556, + 0.008380039595067501, + 0.013861595652997494, + 0.017066337168216705, + 0.013809824362397194, + 0.01236476469784975, + 0.012496914714574814, + 0.018780862912535667, + 0.0168621726334095, + 0.02235555648803711, + 0.010082240216434002, + 0.011329864151775837, + 0.014219401404261589, + 0.012945929542183876, + 0.010787589475512505, + 0.014968697912991047, + 0.01418662816286087, + 0.013079611584544182, + 0.010551008395850658, + 0.012905094772577286, + 0.00961543433368206, + 0.01096161175519228, + 0.01533501222729683, + 0.01414006669074297, + 0.009464273229241371, + 0.012057779356837273, + 0.011670076288282871, + 0.01939442940056324, + 0.012519601732492447, + 0.0223090797662735, + 0.019294198602437973, + 0.017382536083459854, + 0.052616771310567856, + 0.03976266831159592, + 0.008642870001494884, + 0.007560608442872763, + 0.007901768200099468, + 0.007992967031896114, + 0.010244769044220448, + 0.008471853099763393, + 0.009353636763989925, + 0.011772507801651955, + 0.009969324804842472, + 0.011668158695101738, + 0.01591484062373638, + 0.017569907009601593, + 0.01852240599691868, + 0.014639788307249546, + 0.018799914047122, + 0.019939538091421127, + 0.008151578716933727, + 0.00799171719700098, + 0.010890169069170952, + 0.006720191799104214, + 0.008822495117783546, + 0.009874830022454262, + 0.009158614091575146, + 0.008485173806548119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -2.4640395641326904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.82635897397995 + ], + "opposite_logprob_tokens": [ + -0.57635897397995 + ], + "logprob_reward": -0.82635897397995, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.82635897397995, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08531029522418976, + 0.06371355801820755, + 0.09479043632745743, + 0.15487122535705566, + 0.09553409367799759, + 0.09464769810438156, + 0.07654853165149689, + 0.11792466789484024, + 0.12391119450330734, + 0.09274827688932419 + ], + "credit_map": [ + 0.007566072978079319, + 0.011656595394015312, + 0.0104549964889884, + 0.01065991073846817, + 0.010096026584506035, + 0.009624666534364223, + 0.011805761605501175, + 0.013446272350847721, + 0.0063364142552018166, + 0.006961353588849306, + 0.008217085152864456, + 0.010203630663454533, + 0.006707699038088322, + 0.0062539842911064625, + 0.008381317369639874, + 0.01065207552164793, + 0.008876137435436249, + 0.009779226034879684, + 0.010704204440116882, + 0.014113552868366241, + 0.008182044140994549, + 0.00948289129883051, + 0.013966930098831654, + 0.019685450941324234, + 0.018524380400776863, + 0.017239829525351524, + 0.017450720071792603, + 0.018192386254668236, + 0.018729856237769127, + 0.02566567063331604, + 0.01674339734017849, + 0.022324981167912483, + 0.010028647258877754, + 0.009704875759780407, + 0.011652303859591484, + 0.01182053703814745, + 0.01047887746244669, + 0.011411362327635288, + 0.015316199511289597, + 0.01512129046022892, + 0.01158246397972107, + 0.012313048355281353, + 0.010097462683916092, + 0.010722063481807709, + 0.012236760929226875, + 0.011055328883230686, + 0.009073208086192608, + 0.017567364498972893, + 0.007884194143116474, + 0.00913862232118845, + 0.006448227446526289, + 0.0093606011942029, + 0.011112234555184841, + 0.00886518508195877, + 0.012784749269485474, + 0.010954712517559528, + 0.010589961893856525, + 0.013232492841780186, + 0.00879408698529005, + 0.010582564398646355, + 0.022619783878326416, + 0.014201807789504528, + 0.020352497696876526, + 0.017551474273204803, + 0.01004402432590723, + 0.013078941032290459, + 0.014757479541003704, + 0.021729109808802605, + 0.016644014045596123, + 0.009583285078406334, + 0.016596399247646332, + 0.02147793583571911, + 0.00923121441155672, + 0.009948473423719406, + 0.012810018844902515, + 0.010986885987222195, + 0.011251370422542095, + 0.008624228648841381, + 0.014823096804320812, + 0.015072985552251339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.82635897397995 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + }, + { + "prompt": "In a stunning wide, static, high-angle shot capturing Earth from a near-orbital perspective, a vast, ethereal atmosphere drapes across the planet's curve, emitting an inner glow. This translucent veil visualizes global temperature gradients in Makoto Shinkai's ultra-detailed and emotionally resonant style; vast expanses of fiery crimson and vibrant orange flow across the equator, seamlessly blending into gentle greens and yellows before dissolving into deep sapphire and violet blues at the polar caps. The entire layer undulates with a slow, hypnotic rhythm, like silk drifting in a cosmic current. As the Sun sinks below the planet's horizon, its bright, warm light casts vast, softly-edged shadows across the visualized surface, with terminal lines sweeping majestically across the continents. The scene is pierced by Shinkai's signature anamorphic lens flare, rendering the horizons in brilliant blues and oranges, creating a painterly, nostalgic effect. The atmosphere is one of profound, quiet awe, capturing a fleeting moment of cosmic beauty on a planetary scale, imbued with the melancholic yet hopeful emotionality characteristic of the artist's work.", + "seed": 10955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_15/index_7279_seed_10955_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.6471476590260863, + "vlm_logprob": -0.6471476590260863, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.6471476590260863, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033924441784620285 + ], + "opposite_logprob_tokens": [ + -3.5339243412017822 + ], + "logprob_reward": -0.033924441784620285, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.033924441784620285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20656228065490723, + 0.07719022035598755, + 0.048266906291246414, + 0.054930515587329865, + 0.07939266413450241, + 0.15369877219200134, + 0.2329237014055252, + 0.0596311017870903, + 0.0521412156522274, + 0.035262662917375565 + ], + "credit_map": [ + 0.012266388162970543, + 0.021874815225601196, + 0.02855563350021839, + 0.052342284470796585, + 0.015694066882133484, + 0.01520761288702488, + 0.029057854786515236, + 0.03156362846493721, + 0.004524257034063339, + 0.009973661974072456, + 0.00987907126545906, + 0.016013862565159798, + 0.010092084296047688, + 0.004702168982475996, + 0.014281323179602623, + 0.007723791990429163, + 0.003807202447205782, + 0.006784194149076939, + 0.004613319877535105, + 0.0058800107799470425, + 0.007508048787713051, + 0.005314317531883717, + 0.008684775792062283, + 0.005675032734870911, + 0.0052414750680327415, + 0.007744334172457457, + 0.008469757623970509, + 0.007958109490573406, + 0.005542720668017864, + 0.0042547122575342655, + 0.01012036669999361, + 0.005599040538072586, + 0.007738081738352776, + 0.007644347380846739, + 0.01627480611205101, + 0.009639676660299301, + 0.008590041659772396, + 0.0038372559938579798, + 0.013050161302089691, + 0.012618289329111576, + 0.014566415920853615, + 0.01305718906223774, + 0.023282328620553017, + 0.018781006336212158, + 0.012014246545732021, + 0.013993695378303528, + 0.03376038372516632, + 0.02424350008368492, + 0.02540668286383152, + 0.023628393188118935, + 0.03830038011074066, + 0.03234455734491348, + 0.023753060027956963, + 0.02065802924335003, + 0.04712393879890442, + 0.021708661690354347, + 0.005685722921043634, + 0.004050928633660078, + 0.009119649417698383, + 0.011094851419329643, + 0.0045765419490635395, + 0.004353499040007591, + 0.013398904353380203, + 0.0073510101065039635, + 0.005196189973503351, + 0.005913118831813335, + 0.006860443856567144, + 0.008095525205135345, + 0.005238417070358992, + 0.00332423928193748, + 0.011162442155182362, + 0.006350839510560036, + 0.003886996302753687, + 0.0031106409151107073, + 0.005761697888374329, + 0.00495509710162878, + 0.003136125858873129, + 0.0022626875434070826, + 0.0070074438117444515, + 0.0051419720984995365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.033924441784620285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556190602481365 + ], + "opposite_logprob_tokens": [ + -4.387556076049805 + ], + "logprob_reward": -0.012556190602481365, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556190602481365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07647263258695602, + 0.1407269388437271, + 0.061952851712703705, + 0.10264954715967178, + 0.11637403815984726, + 0.045611731708049774, + 0.18934297561645508, + 0.1412305235862732, + 0.05859233811497688, + 0.06704642623662949 + ], + "credit_map": [ + 0.005093927029520273, + 0.008056733757257462, + 0.009392725303769112, + 0.01357137132436037, + 0.007445191964507103, + 0.0052233911119401455, + 0.01512786652892828, + 0.01256142370402813, + 0.012416082434356213, + 0.015585235320031643, + 0.017009086906909943, + 0.032302942126989365, + 0.01628972962498665, + 0.009839602746069431, + 0.02332605980336666, + 0.013958191499114037, + 0.007220539730042219, + 0.008352302014827728, + 0.006857586558908224, + 0.00779297761619091, + 0.00617307098582387, + 0.004743696190416813, + 0.011764302849769592, + 0.00904837716370821, + 0.009343358688056469, + 0.017315927892923355, + 0.016936343163251877, + 0.014311712235212326, + 0.012005945667624474, + 0.0067541589960455894, + 0.015114924870431423, + 0.0108671635389328, + 0.008667605929076672, + 0.011805618181824684, + 0.01579607091844082, + 0.016302727162837982, + 0.011073188856244087, + 0.00886471290141344, + 0.02763882465660572, + 0.016225293278694153, + 0.0054061501286923885, + 0.005688659381121397, + 0.006548818666487932, + 0.005705401301383972, + 0.003819864708930254, + 0.0038708425126969814, + 0.007525154855102301, + 0.007046838756650686, + 0.02010318636894226, + 0.01693595200777054, + 0.03929532691836357, + 0.031002260744571686, + 0.01805516704916954, + 0.01748514547944069, + 0.0300665944814682, + 0.016399351879954338, + 0.01851659081876278, + 0.017231548205018044, + 0.025192653760313988, + 0.020799314603209496, + 0.013189869932830334, + 0.011551879346370697, + 0.017837516963481903, + 0.016911156475543976, + 0.005658372305333614, + 0.007773215416818857, + 0.00567246787250042, + 0.007062130142003298, + 0.011342375539243221, + 0.006101980339735746, + 0.0076553914695978165, + 0.00732641015201807, + 0.008746711537241936, + 0.008933531120419502, + 0.009510350413620472, + 0.00745282182469964, + 0.010292956605553627, + 0.007497922517359257, + 0.007106009870767593, + 0.007506120018661022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.012556190602481365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547831982374191 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547831982374191, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547831982374191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18792761862277985, + 0.06955745071172714, + 0.0790441483259201, + 0.06025298312306404, + 0.10116112232208252, + 0.08074454218149185, + 0.2754670977592468, + 0.041719865053892136, + 0.05780282989144325, + 0.046322375535964966 + ], + "credit_map": [ + 0.014228617772459984, + 0.03979315236210823, + 0.024024128913879395, + 0.03151247650384903, + 0.018677562475204468, + 0.013686622492969036, + 0.021237438544631004, + 0.024767613038420677, + 0.005644123535603285, + 0.008252980187535286, + 0.00690359016880393, + 0.010568346828222275, + 0.007240699604153633, + 0.005977842025458813, + 0.014057907275855541, + 0.010911956429481506, + 0.005219195503741503, + 0.007855667732656002, + 0.010533096268773079, + 0.015468185767531395, + 0.007332544308155775, + 0.005789912771433592, + 0.016841745004057884, + 0.010003799572587013, + 0.0042056674137711525, + 0.007345365826040506, + 0.004660727921873331, + 0.006906726397573948, + 0.005787987262010574, + 0.00559866800904274, + 0.015274613164365292, + 0.010473227128386497, + 0.008087029680609703, + 0.00968808401376009, + 0.013712049461901188, + 0.013320782221853733, + 0.00898185558617115, + 0.009497947059571743, + 0.024366814643144608, + 0.01350656058639288, + 0.007570009212940931, + 0.008244995027780533, + 0.008280358277261257, + 0.009463044814765453, + 0.00793719757348299, + 0.009594544768333435, + 0.016900615766644478, + 0.012753774411976337, + 0.02896871045231819, + 0.029298441484570503, + 0.052609581500291824, + 0.04431750252842903, + 0.02464786358177662, + 0.022901782765984535, + 0.045782703906297684, + 0.026940518990159035, + 0.004828642588108778, + 0.006126608233898878, + 0.004657917656004429, + 0.004859321750700474, + 0.004003321286290884, + 0.003924322780221701, + 0.008085165172815323, + 0.0052345674484968185, + 0.00493165198713541, + 0.006085682660341263, + 0.006649625487625599, + 0.00897924043238163, + 0.008475324138998985, + 0.003866789164021611, + 0.01022434514015913, + 0.008590172044932842, + 0.005264956038445234, + 0.004647369030863047, + 0.00780416140332818, + 0.006054188124835491, + 0.005696460139006376, + 0.0035569514147937298, + 0.00831595342606306, + 0.004982335958629847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.012547831982374191 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.350304365158081 + ], + "opposite_logprob_tokens": [ + -0.10030438750982285 + ], + "logprob_reward": -2.350304365158081, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.350304365158081, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1387249231338501, + 0.09446737915277481, + 0.05775691568851471, + 0.06381269544363022, + 0.11844701319932938, + 0.11606096476316452, + 0.1994808316230774, + 0.09642963111400604, + 0.0791570395231247, + 0.03566263988614082 + ], + "credit_map": [ + 0.009314000606536865, + 0.029250986874103546, + 0.016169190406799316, + 0.024158352985978127, + 0.009114986285567284, + 0.008283460512757301, + 0.024645164608955383, + 0.017788778990507126, + 0.00854459498077631, + 0.014185885898768902, + 0.009613328613340855, + 0.019233543425798416, + 0.00932470802217722, + 0.006014724727720022, + 0.016884712502360344, + 0.01066588144749403, + 0.005407993216067553, + 0.007009411230683327, + 0.008650540374219418, + 0.009365263395011425, + 0.006711733061820269, + 0.003983150236308575, + 0.010626726783812046, + 0.006002094130963087, + 0.007233428303152323, + 0.00770717253908515, + 0.00584373390302062, + 0.006171703338623047, + 0.009222632274031639, + 0.006788077764213085, + 0.009848952293395996, + 0.01099699642509222, + 0.009622338227927685, + 0.011739461682736874, + 0.016035374253988266, + 0.018653007224202156, + 0.010231264866888523, + 0.007478369399905205, + 0.02961041033267975, + 0.01507678534835577, + 0.01109983678907156, + 0.01071951910853386, + 0.021123608574271202, + 0.015657762065529823, + 0.008494390174746513, + 0.006417635828256607, + 0.028412379324436188, + 0.014135845005512238, + 0.021210558712482452, + 0.020521560683846474, + 0.034753888845443726, + 0.031024301424622536, + 0.019381707534193993, + 0.01669798418879509, + 0.03563247248530388, + 0.020258337259292603, + 0.010611845180392265, + 0.01000180747359991, + 0.015455557964742184, + 0.01554569136351347, + 0.008556813932955265, + 0.009109534323215485, + 0.013158934190869331, + 0.013989439234137535, + 0.008046177215874195, + 0.00926864705979824, + 0.010615005157887936, + 0.011086221784353256, + 0.007801990956068039, + 0.007681630086153746, + 0.013892912305891514, + 0.010764451697468758, + 0.004130341578274965, + 0.0037192762829363346, + 0.0047503262758255005, + 0.004600207787007093, + 0.004568104166537523, + 0.0038608249742537737, + 0.005518527235835791, + 0.004515029489994049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -2.350304365158081 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide, static, high-angle view of Earth from a near-orbital perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the atmosphere display a gradient of fiery crimson and orange at the equator transitioning to sapphire and violet blues at the poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the atmosphere undulating with a slow, hypnotic rhythm like silk drifting in a cosmic current?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in Makoto Shinkai's ultra-detailed style featuring anamorphic lens flare and painterly horizons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the Sun appear to sink below the planet's horizon casting softly-edged shadows across the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8264054656028748 + ], + "opposite_logprob_tokens": [ + -0.5764054656028748 + ], + "logprob_reward": -0.8264054656028748, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8264054656028748, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15638315677642822, + 0.13724735379219055, + 0.0685746818780899, + 0.06411455571651459, + 0.10221768170595169, + 0.12553836405277252, + 0.12566006183624268, + 0.0736183449625969, + 0.07926430553197861, + 0.06738152354955673 + ], + "credit_map": [ + 0.01119218859821558, + 0.027389265596866608, + 0.021026331931352615, + 0.0337974913418293, + 0.011137178167700768, + 0.00643160892650485, + 0.02821795828640461, + 0.017191139981150627, + 0.010206509381532669, + 0.015614706091582775, + 0.014785442501306534, + 0.030684063211083412, + 0.014995069243013859, + 0.009521275758743286, + 0.024476079270243645, + 0.016964197158813477, + 0.00623253732919693, + 0.00848506111651659, + 0.007364187389612198, + 0.01069309376180172, + 0.007831006310880184, + 0.0057398490607738495, + 0.013891052454710007, + 0.008337890729308128, + 0.006432098802179098, + 0.010357801802456379, + 0.008386378176510334, + 0.008452107198536396, + 0.008329428732395172, + 0.0052398634143173695, + 0.010546013712882996, + 0.006370860617607832, + 0.010326947085559368, + 0.010286846198141575, + 0.01838560402393341, + 0.014488919638097286, + 0.00819095317274332, + 0.006593882106244564, + 0.02107083424925804, + 0.012873691506683826, + 0.012703743763267994, + 0.011944970116019249, + 0.01520854514092207, + 0.014712643809616566, + 0.014659582637250423, + 0.017944123595952988, + 0.020403753966093063, + 0.017960989847779274, + 0.013575321063399315, + 0.011122955940663815, + 0.023388266563415527, + 0.018398981541395187, + 0.014677661471068859, + 0.013411709107458591, + 0.020511813461780548, + 0.010573350824415684, + 0.011619065888226032, + 0.007810567505657673, + 0.009785531088709831, + 0.010709388181567192, + 0.006726504303514957, + 0.006163578014820814, + 0.012050576508045197, + 0.008753137663006783, + 0.007703253999352455, + 0.00891045294702053, + 0.010029971599578857, + 0.007926025427877903, + 0.008343257941305637, + 0.008231403306126595, + 0.011650598607957363, + 0.01646934635937214, + 0.006938978563994169, + 0.007340764626860619, + 0.008939243853092194, + 0.006403985433280468, + 0.0074456785805523396, + 0.011436420492827892, + 0.007131294347345829, + 0.01174516137689352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.8264054656028748 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_tensor([232, 232])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_tensor([232, 232])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ed9fd19f4b0f2ca2b5cd664a3f050985ec5fa3e2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_tensor([232, 232])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extremely low-angle shot, positioned at ground level, perfectly captures the scene through the shimmering, distorted reflections of a rain-slicked outdoor basketball court. A skilled African-American basketball player, a confident, boyish smile radiating from his face, is depicted with soft, impressionistic brushstrokes. He wears a vibrant royal blue jersey with gold trim, the fabric's texture suggested by the rich, painterly color, while his short, neatly braided hair takes on a dark swirl against the bright, fragmented glare of the stadium overhead. With a low, athletic stance, he rhythmically dribbles the ball, each bounce creating concentric ripples and tiny, shimmering splashes on the water's surface, further abstracting the reflected image. His form and the bright orange of the ball appear to merge and swirl together in a fluid dance of flecked light and vibrant, textured colors, while the background dissolves into a mosaic of cool blues and warm yellows. The entire scene is a painterly vision, transforming moments of athletic prowess into beautiful, fleeting impressions, emphasizing the emotion of joyful mastery and the artistry of movement over literal details.", + "seed": 28034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_232_seed_28034_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17002022699452937, + "vlm_logprob": -0.17002022699452937, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.17002022699452937, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8379784822463989 + ], + "opposite_logprob_tokens": [ + -0.5879784822463989 + ], + "logprob_reward": -0.8379784822463989, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8379784822463989, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1305970698595047, + 0.15013253688812256, + 0.05419380962848663, + 0.1444001942873001, + 0.10508808493614197, + 0.08240818977355957, + 0.1719689965248108, + 0.026739152148365974, + 0.09600142389535904, + 0.03847051411867142 + ], + "credit_map": [ + 0.013146461918950081, + 0.021922240033745766, + 0.024468757212162018, + 0.022995179519057274, + 0.014624838717281818, + 0.01310709211975336, + 0.010096370242536068, + 0.010236143134534359, + 0.02087748982012272, + 0.01928463578224182, + 0.025816839188337326, + 0.018443111330270767, + 0.025752969086170197, + 0.013166069984436035, + 0.014387930743396282, + 0.01240349467843771, + 0.009670646861195564, + 0.010540119372308254, + 0.009024842642247677, + 0.0052711558528244495, + 0.00535343773663044, + 0.002927046734839678, + 0.005155658349394798, + 0.0062509034760296345, + 0.02203245833516121, + 0.02310691401362419, + 0.03528951480984688, + 0.01463804766535759, + 0.011872182600200176, + 0.024001426994800568, + 0.007125489413738251, + 0.006334157194942236, + 0.013958248309791088, + 0.01766066625714302, + 0.01856212317943573, + 0.010484262369573116, + 0.014328657649457455, + 0.012023210525512695, + 0.007153838872909546, + 0.010917074047029018, + 0.013316481374204159, + 0.011367300525307655, + 0.008167530409991741, + 0.007880129851400852, + 0.01185290701687336, + 0.010129149071872234, + 0.009965475648641586, + 0.009729218669235706, + 0.024671468883752823, + 0.022078370675444603, + 0.019445138052105904, + 0.02515518106520176, + 0.01695476844906807, + 0.01617957279086113, + 0.032634805887937546, + 0.014849698171019554, + 0.006976032629609108, + 0.002806018339470029, + 0.0047439029440283775, + 0.002166487043723464, + 0.0023489377927035093, + 0.0036832857877016068, + 0.0019684608560055494, + 0.002046027686446905, + 0.019453061744570732, + 0.012502248398959637, + 0.03143883869051933, + 0.009596419520676136, + 0.004971384070813656, + 0.006966761779040098, + 0.004831638187170029, + 0.00624106964096427, + 0.004877180326730013, + 0.007997525855898857, + 0.0046632480807602406, + 0.0031329323537647724, + 0.005660323891788721, + 0.00318768504075706, + 0.0041728210635483265, + 0.004778796806931496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.8379784822463989 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410511763766408 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.00410511763766408, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410511763766408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2747463881969452, + 0.1844066083431244, + 0.10151262581348419, + 0.1318598985671997, + 0.057672061026096344, + 0.03446904942393303, + 0.05402639880776405, + 0.024259598925709724, + 0.053319334983825684, + 0.08372806757688522 + ], + "credit_map": [ + 0.07430268079042435, + 0.018527483567595482, + 0.09214824438095093, + 0.030739067122340202, + 0.01823165826499462, + 0.015013175085186958, + 0.013029352761805058, + 0.012754750438034534, + 0.05375824496150017, + 0.06755299866199493, + 0.012172224931418896, + 0.011419675312936306, + 0.01322447694838047, + 0.009025455452501774, + 0.009886901825666428, + 0.007366610690951347, + 0.02213892713189125, + 0.019441133365035057, + 0.017385095357894897, + 0.008994534611701965, + 0.010449416004121304, + 0.005532701034098864, + 0.00909233745187521, + 0.008478478528559208, + 0.028108205646276474, + 0.015676293522119522, + 0.020083686336874962, + 0.009326969273388386, + 0.012203820049762726, + 0.03040965273976326, + 0.007967593148350716, + 0.008083664812147617, + 0.007948636077344418, + 0.008195432834327221, + 0.008437754586338997, + 0.0045702820643782616, + 0.010046619921922684, + 0.007316389586776495, + 0.004489733371883631, + 0.006667209789156914, + 0.006245885509997606, + 0.003871020395308733, + 0.0037051860708743334, + 0.0036831824108958244, + 0.0051512825302779675, + 0.003860079450532794, + 0.004094933625310659, + 0.0038574771024286747, + 0.008981175720691681, + 0.0070904940366744995, + 0.007399979513138533, + 0.009129739366471767, + 0.004665909800678492, + 0.005489657167345285, + 0.0063583701848983765, + 0.004911073483526707, + 0.005857267417013645, + 0.0030945204198360443, + 0.004253816790878773, + 0.0016384882619604468, + 0.0025161218363791704, + 0.0033174236305058002, + 0.001638147048652172, + 0.0019438115414232016, + 0.011465181596577168, + 0.006100737489759922, + 0.01689666509628296, + 0.004535342566668987, + 0.003105305368080735, + 0.0042023491114377975, + 0.00380195421166718, + 0.003211800241842866, + 0.00972367636859417, + 0.019269432872533798, + 0.010771173052489758, + 0.0077725062146782875, + 0.01132267713546753, + 0.007149782497435808, + 0.008215362206101418, + 0.009503454901278019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.00410511763766408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3952200710773468, + 0.09215882420539856, + 0.11076179146766663, + 0.07031066715717316, + 0.021473603323101997, + 0.0641212984919548, + 0.09584617614746094, + 0.028047921136021614, + 0.07489362359046936, + 0.04716598242521286 + ], + "credit_map": [ + 0.12021073698997498, + 0.02126958593726158, + 0.1448524296283722, + 0.04528375715017319, + 0.019502008333802223, + 0.0172671377658844, + 0.012442861683666706, + 0.014391559176146984, + 0.022279033437371254, + 0.02631247416138649, + 0.008835826069116592, + 0.005502229556441307, + 0.008671557530760765, + 0.007837674580514431, + 0.0056861150078475475, + 0.007033919915556908, + 0.02220095880329609, + 0.021336903795599937, + 0.018529873341321945, + 0.010001224465668201, + 0.010840283706784248, + 0.008523019030690193, + 0.010260875336825848, + 0.009068657644093037, + 0.01045684702694416, + 0.010077141225337982, + 0.016137033700942993, + 0.006294265389442444, + 0.0070375436916947365, + 0.010915949940681458, + 0.004840969108045101, + 0.004550917539745569, + 0.002948372159153223, + 0.002553505590185523, + 0.002446557395160198, + 0.0024974937550723553, + 0.0028605302795767784, + 0.0034533401485532522, + 0.00186133140232414, + 0.002852473873645067, + 0.008238127455115318, + 0.008326636627316475, + 0.00891101360321045, + 0.004973716102540493, + 0.01029220037162304, + 0.008425199426710606, + 0.008813816122710705, + 0.006140589248389006, + 0.016126461327075958, + 0.01746238023042679, + 0.010408391244709492, + 0.01588713936507702, + 0.008235753513872623, + 0.010325306095182896, + 0.009754162281751633, + 0.007646587211638689, + 0.0073870704509317875, + 0.0036127925850450993, + 0.003674536244943738, + 0.0019022793276235461, + 0.0026959737297147512, + 0.0037916835863143206, + 0.0018202053615823388, + 0.0031633798498660326, + 0.01656481623649597, + 0.01158644538372755, + 0.022662213072180748, + 0.006147765554487705, + 0.004443701822310686, + 0.005176586098968983, + 0.004060361534357071, + 0.004251733887940645, + 0.008638477884232998, + 0.009357631206512451, + 0.006577066145837307, + 0.003479565726593137, + 0.006025183014571667, + 0.003535365452989936, + 0.004683574195951223, + 0.004869118332862854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.004637913778424263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028190179727971554 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028190179727971554, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028190179727971554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1919071227312088, + 0.14547353982925415, + 0.1766369789838791, + 0.05520530045032501, + 0.04217250272631645, + 0.051806118339300156, + 0.0545387864112854, + 0.03673212230205536, + 0.20645737648010254, + 0.039070166647434235 + ], + "credit_map": [ + 0.03138982132077217, + 0.032042428851127625, + 0.04473373666405678, + 0.029415132477879524, + 0.019823860377073288, + 0.012001581490039825, + 0.01138579286634922, + 0.011114760302007198, + 0.041386984288692474, + 0.05208124592900276, + 0.013529079966247082, + 0.009271051734685898, + 0.008937753736972809, + 0.006826657336205244, + 0.007793427910655737, + 0.005647328682243824, + 0.04741291329264641, + 0.0427107997238636, + 0.03609256446361542, + 0.014446718618273735, + 0.009828872978687286, + 0.006787602324038744, + 0.0075194621458649635, + 0.011838041245937347, + 0.010412821546196938, + 0.008523532189428806, + 0.009015941992402077, + 0.00562329450622201, + 0.005338976625353098, + 0.007824811153113842, + 0.005048428662121296, + 0.0034174940083175898, + 0.005865080282092094, + 0.006942796986550093, + 0.006946769077330828, + 0.004868421703577042, + 0.005703490227460861, + 0.0051545253954827785, + 0.0028745061717927456, + 0.003816913580521941, + 0.009525875560939312, + 0.006225109100341797, + 0.006685453001409769, + 0.005513913929462433, + 0.00706406868994236, + 0.00647978950291872, + 0.004269373137503862, + 0.00604253588244319, + 0.008355456404387951, + 0.0075099878013134, + 0.006674336269497871, + 0.006896576378494501, + 0.005184714682400227, + 0.005154078360646963, + 0.01004748698323965, + 0.0047161513939499855, + 0.010429345071315765, + 0.004004091490060091, + 0.0039766705594956875, + 0.0026237796992063522, + 0.0036026195157319307, + 0.005050027277320623, + 0.0028271556366235018, + 0.004218431189656258, + 0.04574770852923393, + 0.02011255733668804, + 0.07665932178497314, + 0.018538838252425194, + 0.009201682172715664, + 0.013696744106709957, + 0.010973461903631687, + 0.011527078226208687, + 0.00892821978777647, + 0.0050632404163479805, + 0.006279973778873682, + 0.003541838377714157, + 0.004048331640660763, + 0.003085911273956299, + 0.003762943670153618, + 0.00435970863327384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0028190179727971554 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005606033373624086 + ], + "opposite_logprob_tokens": [ + -7.500560760498047 + ], + "logprob_reward": -0.0005606033373624086, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005606033373624086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1715586930513382, + 0.25373995304107666, + 0.07644259929656982, + 0.0857405960559845, + 0.05117372050881386, + 0.04794473946094513, + 0.08833774924278259, + 0.021419771015644073, + 0.16629546880722046, + 0.03734666109085083 + ], + "credit_map": [ + 0.04075079411268234, + 0.015833595767617226, + 0.042360104620456696, + 0.02308313176035881, + 0.014271488413214684, + 0.013767088763415813, + 0.010794978588819504, + 0.010697532445192337, + 0.0703233852982521, + 0.08195044100284576, + 0.025929277762770653, + 0.015551753342151642, + 0.02714700996875763, + 0.01139345858246088, + 0.010494817048311234, + 0.010949842631816864, + 0.01948740892112255, + 0.014076941646635532, + 0.013742152601480484, + 0.008662713691592216, + 0.006247255019843578, + 0.004113055299967527, + 0.004395602270960808, + 0.00571747962385416, + 0.014482002705335617, + 0.010345585644245148, + 0.01818164438009262, + 0.007020797114819288, + 0.00968500692397356, + 0.016095343977212906, + 0.005153656471520662, + 0.00477656489238143, + 0.006390386261045933, + 0.008060001768171787, + 0.007953292690217495, + 0.00443085515871644, + 0.007703580893576145, + 0.007783099543303251, + 0.003988346084952354, + 0.004864165093749762, + 0.0058394125662744045, + 0.005705554503947496, + 0.003962590824812651, + 0.006145673803985119, + 0.008582714945077896, + 0.00632860604673624, + 0.005013674963265657, + 0.0063665215857326984, + 0.013849565759301186, + 0.011080235242843628, + 0.011098244227468967, + 0.012825831770896912, + 0.007200995925813913, + 0.008422446437180042, + 0.0149284852668643, + 0.008931951597332954, + 0.0043946788646280766, + 0.0024640162009745836, + 0.004698937758803368, + 0.0015206815442070365, + 0.0015400314005091786, + 0.0028469564858824015, + 0.0016984809190034866, + 0.0022559897042810917, + 0.0360710471868515, + 0.01865643449127674, + 0.05656133592128754, + 0.015406332910060883, + 0.008209513500332832, + 0.012630611658096313, + 0.008776845410466194, + 0.009983359836041927, + 0.010127737186849117, + 0.003711723955348134, + 0.008878528140485287, + 0.003442792920395732, + 0.003503612009808421, + 0.0023137058597058058, + 0.0025829016231000423, + 0.0027856628876179457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0005606033373624086 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle shot, positioned at ground level, perfectly captures the scene through the shimmering, distorted reflections of a rain-slicked outdoor basketball court. A skilled African-American basketball player, a confident, boyish smile radiating from his face, is depicted with soft, impressionistic brushstrokes. He wears a vibrant royal blue jersey with gold trim, the fabric's texture suggested by the rich, painterly color, while his short, neatly braided hair takes on a dark swirl against the bright, fragmented glare of the stadium overhead. With a low, athletic stance, he rhythmically dribbles the ball, each bounce creating concentric ripples and tiny, shimmering splashes on the water's surface, further abstracting the reflected image. His form and the bright orange of the ball appear to merge and swirl together in a fluid dance of flecked light and vibrant, textured colors, while the background dissolves into a mosaic of cool blues and warm yellows. The entire scene is a painterly vision, transforming moments of athletic prowess into beautiful, fleeting impressions, emphasizing the emotion of joyful mastery and the artistry of movement over literal details.", + "seed": 28035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_232_seed_28035_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2860781975090504, + "vlm_logprob": -0.2860781975090504, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.2860781975090504, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4118502140045166 + ], + "opposite_logprob_tokens": [ + -0.2868501842021942 + ], + "logprob_reward": -1.4118502140045166, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4118502140045166, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.40376272797584534, + 0.12032332271337509, + 0.06375707685947418, + 0.061830710619688034, + 0.13201415538787842, + 0.03297071158885956, + 0.046368230134248734, + 0.06766018271446228, + 0.04430896043777466, + 0.027003902941942215 + ], + "credit_map": [ + 0.04900647699832916, + 0.026153849437832832, + 0.015018061734735966, + 0.011589177884161472, + 0.027389058843255043, + 0.14908578991889954, + 0.017265325412154198, + 0.10825499892234802, + 0.01725524291396141, + 0.012004921212792397, + 0.026930565014481544, + 0.007239342667162418, + 0.015527039766311646, + 0.01913335546851158, + 0.009280486963689327, + 0.012952355667948723, + 0.008134200237691402, + 0.007151834666728973, + 0.0062871165573596954, + 0.006135564297437668, + 0.007941152900457382, + 0.006580227054655552, + 0.007091264706104994, + 0.014435715973377228, + 0.007235594093799591, + 0.014847774058580399, + 0.006952614989131689, + 0.004551492165774107, + 0.005201236344873905, + 0.006062667816877365, + 0.0060913292691111565, + 0.010888001881539822, + 0.02332819439470768, + 0.008999369107186794, + 0.009758188389241695, + 0.014294023625552654, + 0.009087033569812775, + 0.044134099036455154, + 0.012392627075314522, + 0.010020622983574867, + 0.0060528842732310295, + 0.005956337787210941, + 0.0045455703511834145, + 0.0032887416891753674, + 0.0035464453976601362, + 0.0033187640365213156, + 0.002875532489269972, + 0.003386435331776738, + 0.007226345594972372, + 0.007867453619837761, + 0.0060289534740149975, + 0.0038501927629113197, + 0.003970345947891474, + 0.005377986468374729, + 0.0034815981052815914, + 0.008565355092287064, + 0.011081340722739697, + 0.005837047006934881, + 0.006628567818552256, + 0.004389477893710136, + 0.00547215947881341, + 0.012523185461759567, + 0.011233413591980934, + 0.010494991205632687, + 0.0067304871045053005, + 0.006146502681076527, + 0.0035372155252844095, + 0.003500122344121337, + 0.005761436186730862, + 0.009179356507956982, + 0.0033116729464381933, + 0.006142167374491692, + 0.002895896788686514, + 0.004048004746437073, + 0.0023766118101775646, + 0.002671458525583148, + 0.007033087313175201, + 0.002687155269086361, + 0.002255417639389634, + 0.003036272246390581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -1.4118502140045166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104642663151026 + ], + "opposite_logprob_tokens": [ + -5.5041046142578125 + ], + "logprob_reward": -0.004104642663151026, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104642663151026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27977195382118225, + 0.12272608280181885, + 0.05493532121181488, + 0.07056088000535965, + 0.13098347187042236, + 0.033441212028265, + 0.04563068971037865, + 0.08222825825214386, + 0.04003741964697838, + 0.13968469202518463 + ], + "credit_map": [ + 0.04111017286777496, + 0.01526745967566967, + 0.013747398741543293, + 0.008601987734436989, + 0.026662621647119522, + 0.09845665097236633, + 0.014207979664206505, + 0.06171770393848419, + 0.018080007284879684, + 0.014153126627206802, + 0.02306494116783142, + 0.007488651666790247, + 0.016387393698096275, + 0.02074105106294155, + 0.008768727071583271, + 0.014042196795344353, + 0.006760559976100922, + 0.007637074217200279, + 0.006014540325850248, + 0.004769600927829742, + 0.0063865212723612785, + 0.006178881041705608, + 0.006314452737569809, + 0.010873699560761452, + 0.008385197259485722, + 0.020669318735599518, + 0.005628866609185934, + 0.005068219266831875, + 0.006391170900315046, + 0.005622915457934141, + 0.007521930616348982, + 0.011273261159658432, + 0.019146407023072243, + 0.0148270633071661, + 0.013925463892519474, + 0.016339357942342758, + 0.00915901642292738, + 0.02617006003856659, + 0.018471835181117058, + 0.012944280169904232, + 0.0065484968945384026, + 0.005354227032512426, + 0.0050573875196278095, + 0.004147871397435665, + 0.003275902010500431, + 0.002952530048787594, + 0.0031654781196266413, + 0.0029393217992037535, + 0.0066145723685622215, + 0.008348677307367325, + 0.005073653068393469, + 0.003931202460080385, + 0.004702421836555004, + 0.004650217015296221, + 0.003330943174660206, + 0.008979005739092827, + 0.0182713083922863, + 0.0072832792066037655, + 0.004703542683273554, + 0.005088556557893753, + 0.005583597347140312, + 0.026366878300905228, + 0.0064050727523863316, + 0.008526023477315903, + 0.005777142010629177, + 0.0051896460354328156, + 0.005494990386068821, + 0.004623902030289173, + 0.0045821769163012505, + 0.006953476928174496, + 0.0030364496633410454, + 0.0043796394020318985, + 0.015834994614124298, + 0.01953628659248352, + 0.015249963849782944, + 0.011507701128721237, + 0.033744167536497116, + 0.02049192413687706, + 0.010515705682337284, + 0.012803958728909492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.004104642663151026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780462831258774 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780462831258774, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780462831258774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1866505891084671, + 0.06498287618160248, + 0.05104771628975868, + 0.0399155393242836, + 0.2333318293094635, + 0.051396485418081284, + 0.03813772276043892, + 0.1633155643939972, + 0.03182432800531387, + 0.13939735293388367 + ], + "credit_map": [ + 0.01568109355866909, + 0.00912032276391983, + 0.006704252678900957, + 0.004664111882448196, + 0.012169222347438335, + 0.07207510620355606, + 0.006992594804614782, + 0.05924386903643608, + 0.00959648098796606, + 0.005104535724967718, + 0.012188062071800232, + 0.005067151505500078, + 0.008572466671466827, + 0.011003786697983742, + 0.004484392702579498, + 0.00896600354462862, + 0.00456780893728137, + 0.0042241839691996574, + 0.003176421392709017, + 0.004613784141838551, + 0.005012682639062405, + 0.011317772790789604, + 0.004141583573073149, + 0.013993476517498493, + 0.005388585384935141, + 0.009218317456543446, + 0.003791866125538945, + 0.0031071295961737633, + 0.00342489848844707, + 0.0033247587271034718, + 0.004495227709412575, + 0.007164756301790476, + 0.027127167209982872, + 0.030581846833229065, + 0.036041855812072754, + 0.026837164536118507, + 0.018685167655348778, + 0.03929895907640457, + 0.031806185841560364, + 0.02295348420739174, + 0.009902403689920902, + 0.008190841414034367, + 0.006583483424037695, + 0.006340429186820984, + 0.006535611115396023, + 0.005106424912810326, + 0.00403256993740797, + 0.004704722203314304, + 0.00573928514495492, + 0.007753278128802776, + 0.005212747026234865, + 0.004214117769151926, + 0.003562136320397258, + 0.00451325299218297, + 0.002888939343392849, + 0.004253964405506849, + 0.039746303111314774, + 0.012065800838172436, + 0.008423675782978535, + 0.009511175565421581, + 0.010430685244500637, + 0.05549054965376854, + 0.013204850256443024, + 0.014442524872720242, + 0.004638809245079756, + 0.003222248749807477, + 0.004124263767153025, + 0.0036912953946739435, + 0.0030442923307418823, + 0.00567166181281209, + 0.002266278024762869, + 0.005165478680282831, + 0.014468495734035969, + 0.015739498659968376, + 0.013265586458146572, + 0.013870110735297203, + 0.04023692011833191, + 0.019572393968701363, + 0.010698405094444752, + 0.011545935645699501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.009780462831258774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036151548847556114 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036151548847556114, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036151548847556114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6112529635429382, + 0.023905059322714806, + 0.05473152920603752, + 0.02596335858106613, + 0.07847108691930771, + 0.031224800273776054, + 0.02302335575222969, + 0.08690405637025833, + 0.027157923206686974, + 0.03736591339111328 + ], + "credit_map": [ + 0.052859533578157425, + 0.025761887431144714, + 0.014255767688155174, + 0.011570391245186329, + 0.022356892004609108, + 0.2529360353946686, + 0.02020222321152687, + 0.21131020784378052, + 0.0035633998923003674, + 0.002708140527829528, + 0.004122158046811819, + 0.0017021276289597154, + 0.002740213181823492, + 0.0037237792275846004, + 0.0015934673137962818, + 0.00375177338719368, + 0.007745977956801653, + 0.0043686917051672935, + 0.00454852357506752, + 0.00340693024918437, + 0.00548823457211256, + 0.007927719503641129, + 0.004366434644907713, + 0.016879020258784294, + 0.0031152560841292143, + 0.005467033479362726, + 0.0029065010603517294, + 0.0019234101055189967, + 0.0033892402425408363, + 0.0024964523036032915, + 0.0024420758709311485, + 0.004223389085382223, + 0.012736205942928791, + 0.007349010091274977, + 0.007058377377688885, + 0.007744635920971632, + 0.004229124169796705, + 0.027135826647281647, + 0.004797861445695162, + 0.007420049048960209, + 0.005915686022490263, + 0.004947641398757696, + 0.004107390530407429, + 0.0038936296477913857, + 0.0038899767678231, + 0.002759279916062951, + 0.0029102431144565344, + 0.0028009514790028334, + 0.0034901523031294346, + 0.0036496403627097607, + 0.002705968916416168, + 0.0020848968997597694, + 0.002354253316298127, + 0.002475402317941189, + 0.001773999072611332, + 0.004489042330533266, + 0.01604825258255005, + 0.004891206976026297, + 0.007056902628391981, + 0.004988306667655706, + 0.006504864431917667, + 0.018523376435041428, + 0.015929952263832092, + 0.012961197644472122, + 0.003489034716039896, + 0.003336805384606123, + 0.0034269094467163086, + 0.0033734214957803488, + 0.0032150682527571917, + 0.0031055076979100704, + 0.002378821140155196, + 0.004832354374229908, + 0.0037399043794721365, + 0.004018736537545919, + 0.0030094694811850786, + 0.0041204593144357204, + 0.010561511851847172, + 0.005451375152915716, + 0.003346080891788006, + 0.003118377411738038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0036151548847556114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010405131615698338 + ], + "opposite_logprob_tokens": [ + -6.876040458679199 + ], + "logprob_reward": -0.0010405131615698338, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010405131615698338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22134381532669067, + 0.11387230455875397, + 0.1197337880730629, + 0.06851723790168762, + 0.18671831488609314, + 0.05011645704507828, + 0.061365678906440735, + 0.061274878680706024, + 0.04628889262676239, + 0.07076866179704666 + ], + "credit_map": [ + 0.04835725203156471, + 0.024890651926398277, + 0.019085844978690147, + 0.01750846952199936, + 0.02543894201517105, + 0.03566817566752434, + 0.017854295670986176, + 0.03254018723964691, + 0.01548159308731556, + 0.00986838061362505, + 0.02220739983022213, + 0.00782580766826868, + 0.015785707160830498, + 0.019966576248407364, + 0.00819212943315506, + 0.014544708654284477, + 0.02126891352236271, + 0.008352117612957954, + 0.008900627493858337, + 0.008132806979119778, + 0.00851707998663187, + 0.022211143746972084, + 0.012833386659622192, + 0.02951771952211857, + 0.00696288887411356, + 0.014153443276882172, + 0.008353332057595253, + 0.004574045073240995, + 0.006807688158005476, + 0.005707946140319109, + 0.007110936567187309, + 0.014846955426037312, + 0.030680259689688683, + 0.011338997632265091, + 0.01467642281204462, + 0.015892161056399345, + 0.012002654373645782, + 0.0742119550704956, + 0.012324938550591469, + 0.015590920113027096, + 0.009037968702614307, + 0.007589498069137335, + 0.007891274988651276, + 0.006997691001743078, + 0.00587769178673625, + 0.004587897099554539, + 0.004063171800225973, + 0.004071264527738094, + 0.011950911022722721, + 0.011032846756279469, + 0.009162812493741512, + 0.005634367000311613, + 0.005424544680863619, + 0.006559907458722591, + 0.004601753316819668, + 0.006998539436608553, + 0.008953209035098553, + 0.005737728904932737, + 0.006722288206219673, + 0.0038814411964267492, + 0.004874770529568195, + 0.009221774525940418, + 0.010947070084512234, + 0.010936598293483257, + 0.006745274178683758, + 0.005502553191035986, + 0.004475579597055912, + 0.00407263683155179, + 0.005790600087493658, + 0.010775675065815449, + 0.0033624100033193827, + 0.005564166232943535, + 0.00901001412421465, + 0.0066643464379012585, + 0.008067000657320023, + 0.005932180676609278, + 0.016626140102744102, + 0.013473891653120518, + 0.004515107721090317, + 0.006479976233094931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0010405131615698338 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle shot, positioned at ground level, perfectly captures the scene through the shimmering, distorted reflections of a rain-slicked outdoor basketball court. A skilled African-American basketball player, a confident, boyish smile radiating from his face, is depicted with soft, impressionistic brushstrokes. He wears a vibrant royal blue jersey with gold trim, the fabric's texture suggested by the rich, painterly color, while his short, neatly braided hair takes on a dark swirl against the bright, fragmented glare of the stadium overhead. With a low, athletic stance, he rhythmically dribbles the ball, each bounce creating concentric ripples and tiny, shimmering splashes on the water's surface, further abstracting the reflected image. His form and the bright orange of the ball appear to merge and swirl together in a fluid dance of flecked light and vibrant, textured colors, while the background dissolves into a mosaic of cool blues and warm yellows. The entire scene is a painterly vision, transforming moments of athletic prowess into beautiful, fleeting impressions, emphasizing the emotion of joyful mastery and the artistry of movement over literal details.", + "seed": 28036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_232_seed_28036_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23195655765011908, + "vlm_logprob": -0.23195655765011908, + "vlm_yes_no_margin": 4.624999988079071, + "vlm_token_credit": -0.23195655765011908, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1451010704040527 + ], + "opposite_logprob_tokens": [ + -0.39510101079940796 + ], + "logprob_reward": -1.1451010704040527, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.1451010704040527, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1809869259595871, + 0.06236700341105461, + 0.1600944995880127, + 0.1457371860742569, + 0.04291961342096329, + 0.049166541546583176, + 0.032944757491350174, + 0.036049842834472656, + 0.26671937108039856, + 0.02301422506570816 + ], + "credit_map": [ + 0.0310507845133543, + 0.010983683168888092, + 0.011123480275273323, + 0.006776090711355209, + 0.01135066244751215, + 0.021884078159928322, + 0.020655959844589233, + 0.0671621784567833, + 0.007563599850982428, + 0.009069755673408508, + 0.009089341387152672, + 0.005647139623761177, + 0.008377590216696262, + 0.009754698723554611, + 0.006897255778312683, + 0.0059676216915249825, + 0.019627252593636513, + 0.008338510058820248, + 0.011763030663132668, + 0.009671314619481564, + 0.008277054876089096, + 0.028632046654820442, + 0.036291055381298065, + 0.0374942421913147, + 0.017407650128006935, + 0.015841713175177574, + 0.013818632811307907, + 0.013789579272270203, + 0.013803922571241856, + 0.043650299310684204, + 0.01292307861149311, + 0.014502311125397682, + 0.005370066501200199, + 0.005317794159054756, + 0.004742774646729231, + 0.005692893173545599, + 0.005329219624400139, + 0.006557477172464132, + 0.0048557184636592865, + 0.00505366874858737, + 0.012252811342477798, + 0.012405059300363064, + 0.00825436320155859, + 0.004172933753579855, + 0.003664050018414855, + 0.002663103863596916, + 0.00285924319177866, + 0.002894977107644081, + 0.004676187410950661, + 0.0032319382298737764, + 0.0038254866376519203, + 0.005794401280581951, + 0.0039194319397211075, + 0.0033859226386994123, + 0.004708595108240843, + 0.003402795409783721, + 0.005804054904729128, + 0.0024205732624977827, + 0.0025187069550156593, + 0.003040290204808116, + 0.004356949590146542, + 0.009632849134504795, + 0.002760860836133361, + 0.0055155567824840546, + 0.09275505691766739, + 0.009162656031548977, + 0.01258857548236847, + 0.009591788053512573, + 0.008536535315215588, + 0.024609962478280067, + 0.012627567164599895, + 0.09684722125530243, + 0.00394978653639555, + 0.002181651769205928, + 0.002472867025062442, + 0.002819109009578824, + 0.002020517596974969, + 0.0030947017949074507, + 0.0021170226391404867, + 0.004358566831797361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -1.1451010704040527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025022882036864758 + ], + "opposite_logprob_tokens": [ + -6.00250244140625 + ], + "logprob_reward": -0.0025022882036864758, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025022882036864758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41136425733566284, + 0.04796214774250984, + 0.1561189889907837, + 0.08629482984542847, + 0.0327029787003994, + 0.07866600900888443, + 0.022868599742650986, + 0.02244452014565468, + 0.11065609008073807, + 0.030921567231416702 + ], + "credit_map": [ + 0.05894935503602028, + 0.0326710119843483, + 0.021372707560658455, + 0.016240794211626053, + 0.02790028788149357, + 0.060051653534173965, + 0.05136952921748161, + 0.14280889928340912, + 0.005227129906415939, + 0.005916604306548834, + 0.010092281736433506, + 0.004008361138403416, + 0.004128425382077694, + 0.008931459859013557, + 0.004128407686948776, + 0.005529476795345545, + 0.011188307777047157, + 0.00548213766887784, + 0.007645434234291315, + 0.007747798226773739, + 0.006991841364651918, + 0.013294398784637451, + 0.0521349236369133, + 0.05163414776325226, + 0.008244405500590801, + 0.00687023950740695, + 0.0065981014631688595, + 0.006158837117254734, + 0.0070201801136136055, + 0.032593049108982086, + 0.00857932772487402, + 0.010230697691440582, + 0.0038700366858392954, + 0.0036553428508341312, + 0.003932065796107054, + 0.005707246251404285, + 0.0036626302171498537, + 0.005104684736579657, + 0.0032618611585348845, + 0.003509110538288951, + 0.02944118157029152, + 0.026955820620059967, + 0.006136874668300152, + 0.004124345723539591, + 0.004595532547682524, + 0.0022939194459468126, + 0.0024742200039327145, + 0.0026441102381795645, + 0.0028224079869687557, + 0.0030794551130384207, + 0.0022843859624117613, + 0.002718408592045307, + 0.003960986156016588, + 0.003045769641175866, + 0.0026508797891438007, + 0.002306306501850486, + 0.0034767386969178915, + 0.0019124625250697136, + 0.001976562896743417, + 0.0021700169891119003, + 0.0027350005693733692, + 0.003798373509198427, + 0.0020177732221782207, + 0.004357592668384314, + 0.039413586258888245, + 0.00364190386608243, + 0.00488487770780921, + 0.003621953073889017, + 0.0032390912529081106, + 0.009390371851623058, + 0.00521348649635911, + 0.041250813752412796, + 0.004867502488195896, + 0.0030144357588142157, + 0.002155066467821598, + 0.0034356953110545874, + 0.003542356425896287, + 0.0053148758597671986, + 0.0026313711423426867, + 0.005960263777524233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0025022882036864758 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639340288937092 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639340288937092, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639340288937092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3200834095478058, + 0.05455353483557701, + 0.22358796000480652, + 0.14549393951892853, + 0.027149181813001633, + 0.028425274416804314, + 0.09841015934944153, + 0.02680910751223564, + 0.050276726484298706, + 0.025210676714777946 + ], + "credit_map": [ + 0.03948451951146126, + 0.03871743008494377, + 0.024501636624336243, + 0.017036689445376396, + 0.041653458029031754, + 0.044266898185014725, + 0.04208045080304146, + 0.07234232872724533, + 0.005256981123238802, + 0.007569271605461836, + 0.006723529193550348, + 0.006730902008712292, + 0.005352416075766087, + 0.008757037110626698, + 0.00761039974167943, + 0.006552995182573795, + 0.018437976017594337, + 0.008030234836041927, + 0.00695396214723587, + 0.006843288894742727, + 0.010352712124586105, + 0.018280601128935814, + 0.08064200729131699, + 0.07404719293117523, + 0.010951507836580276, + 0.009432929567992687, + 0.014327850192785263, + 0.009506111033260822, + 0.0074632177129387856, + 0.035557929426431656, + 0.029313305392861366, + 0.028941089287400246, + 0.002603687345981598, + 0.003613912034779787, + 0.0041327523067593575, + 0.003596925875172019, + 0.0023143948055803776, + 0.005357589107006788, + 0.002356189303100109, + 0.00317373126745224, + 0.003986409865319729, + 0.0044855051673948765, + 0.00403427891433239, + 0.0026869294233620167, + 0.003457819577306509, + 0.002957255346700549, + 0.003287028521299362, + 0.0035300476010888815, + 0.009342087432742119, + 0.017235834151506424, + 0.009422983974218369, + 0.011220582760870457, + 0.017037080600857735, + 0.012798367068171501, + 0.01164577342569828, + 0.009707458317279816, + 0.004326865077018738, + 0.0015874664532020688, + 0.002829021541401744, + 0.0021597864106297493, + 0.002826808486133814, + 0.006258942186832428, + 0.002150612184777856, + 0.004669606685638428, + 0.017559746280312538, + 0.0018280060030519962, + 0.0021584422793239355, + 0.0018990153912454844, + 0.001811780035495758, + 0.0044536893256008625, + 0.0022647040896117687, + 0.01830134354531765, + 0.0033990449737757444, + 0.0028807073831558228, + 0.0019332226365804672, + 0.002383825136348605, + 0.0029610085766762495, + 0.00437054643407464, + 0.002937604673206806, + 0.004344716668128967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.008639340288937092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820919966325164 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.002820919966325164, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820919966325164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14227253198623657, + 0.04667364060878754, + 0.16447114944458008, + 0.1265626698732376, + 0.03223464637994766, + 0.20931679010391235, + 0.025569921359419823, + 0.03158991411328316, + 0.19497668743133545, + 0.02633204683661461 + ], + "credit_map": [ + 0.019863203167915344, + 0.006462481804192066, + 0.008196884766221046, + 0.004617861472070217, + 0.007936988957226276, + 0.022800473496317863, + 0.018928242847323418, + 0.053466398268938065, + 0.005420878529548645, + 0.0026964133139699697, + 0.004495273809880018, + 0.0027285588439553976, + 0.0038445708341896534, + 0.013105795718729496, + 0.006142673548310995, + 0.008239476010203362, + 0.012790820561349392, + 0.006530563812702894, + 0.0076390039175748825, + 0.005912353750318289, + 0.006834012921899557, + 0.013421989046037197, + 0.05632825940847397, + 0.055014144629240036, + 0.01100290659815073, + 0.007019291166216135, + 0.008938612416386604, + 0.007666813675314188, + 0.008144333958625793, + 0.03061630018055439, + 0.02777070179581642, + 0.02540370635688305, + 0.0037025550846010447, + 0.005014720372855663, + 0.00362352654337883, + 0.0038759643211960793, + 0.00394525146111846, + 0.004937097430229187, + 0.0030507673509418964, + 0.004084762651473284, + 0.0870075672864914, + 0.08003389835357666, + 0.007583297323435545, + 0.006894601974636316, + 0.01240271795541048, + 0.0043139467015862465, + 0.005290484521538019, + 0.005790275521576405, + 0.0034936140291392803, + 0.0031095927115529776, + 0.002813209779560566, + 0.004028754308819771, + 0.00302209984511137, + 0.0031046713702380657, + 0.003153436351567507, + 0.0028445443604141474, + 0.004576721228659153, + 0.002288431627675891, + 0.002357977442443371, + 0.0026520066894590855, + 0.0038604377768933773, + 0.009210694581270218, + 0.002168942242860794, + 0.004474700894206762, + 0.06931418180465698, + 0.0063905268907547, + 0.008852546103298664, + 0.006574865430593491, + 0.005628669634461403, + 0.017042048275470734, + 0.008921711705625057, + 0.07225213199853897, + 0.004541538190096617, + 0.0029360821936279535, + 0.0017626737244427204, + 0.0017600500723347068, + 0.0026050249580293894, + 0.005282852333039045, + 0.0026999807450920343, + 0.004743845202028751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.002820919966325164 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007191693875938654 + ], + "opposite_logprob_tokens": [ + -7.25071907043457 + ], + "logprob_reward": -0.0007191693875938654, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007191693875938654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11258859932422638, + 0.04405555501580238, + 0.19867880642414093, + 0.24218909442424774, + 0.040906500071287155, + 0.08994200825691223, + 0.02685392275452614, + 0.03823684900999069, + 0.17853416502475739, + 0.02801447920501232 + ], + "credit_map": [ + 0.018037009984254837, + 0.009370344690978527, + 0.009481680579483509, + 0.007642174605280161, + 0.01376674696803093, + 0.013260072097182274, + 0.01932297646999359, + 0.02170759066939354, + 0.005042892415076494, + 0.005704307928681374, + 0.005917660426348448, + 0.0044747330248355865, + 0.006070132367312908, + 0.006086663343012333, + 0.005705398507416248, + 0.005053771194070578, + 0.02275676652789116, + 0.009912943467497826, + 0.013152742758393288, + 0.011470775119960308, + 0.00914914719760418, + 0.03215755894780159, + 0.051312826573848724, + 0.04876605421304703, + 0.020394427701830864, + 0.011638907715678215, + 0.013818388804793358, + 0.010816778987646103, + 0.010037415660917759, + 0.03383879363536835, + 0.07414194196462631, + 0.06750243902206421, + 0.005329465959221125, + 0.006514719687402248, + 0.004326361697167158, + 0.005940933711826801, + 0.004592919256538153, + 0.005214557982981205, + 0.004485783167183399, + 0.0045017581433057785, + 0.02890234813094139, + 0.026716791093349457, + 0.009249422699213028, + 0.007100197486579418, + 0.00557073624804616, + 0.003199906786903739, + 0.003605553414672613, + 0.005597052630037069, + 0.004276858642697334, + 0.0036976283881813288, + 0.002317073754966259, + 0.003554225666448474, + 0.003081102157011628, + 0.0031108821276575327, + 0.0035118660889565945, + 0.0033042856957763433, + 0.006509884726256132, + 0.002462463453412056, + 0.0035507301799952984, + 0.0030093900859355927, + 0.0045992969535291195, + 0.007544308435171843, + 0.0034441440366208553, + 0.007116629742085934, + 0.06493046134710312, + 0.005857271607965231, + 0.007391428109258413, + 0.00533802155405283, + 0.004673199728131294, + 0.013907519169151783, + 0.008329122327268124, + 0.06810712814331055, + 0.0044908649288117886, + 0.003163158195093274, + 0.002114512724801898, + 0.0027357377111911774, + 0.002853282494470477, + 0.004261246416717768, + 0.0026275001000612974, + 0.005768178030848503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.0007191693875938654 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle shot, positioned at ground level, perfectly captures the scene through the shimmering, distorted reflections of a rain-slicked outdoor basketball court. A skilled African-American basketball player, a confident, boyish smile radiating from his face, is depicted with soft, impressionistic brushstrokes. He wears a vibrant royal blue jersey with gold trim, the fabric's texture suggested by the rich, painterly color, while his short, neatly braided hair takes on a dark swirl against the bright, fragmented glare of the stadium overhead. With a low, athletic stance, he rhythmically dribbles the ball, each bounce creating concentric ripples and tiny, shimmering splashes on the water's surface, further abstracting the reflected image. His form and the bright orange of the ball appear to merge and swirl together in a fluid dance of flecked light and vibrant, textured colors, while the background dissolves into a mosaic of cool blues and warm yellows. The entire scene is a painterly vision, transforming moments of athletic prowess into beautiful, fleeting impressions, emphasizing the emotion of joyful mastery and the artistry of movement over literal details.", + "seed": 28037, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_2/index_232_seed_28037_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.28564669471234083, + "vlm_logprob": -0.28564669471234083, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.28564669471234083, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4154868125915527 + ], + "opposite_logprob_tokens": [ + -0.29048678278923035 + ], + "logprob_reward": -1.4154868125915527, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4154868125915527, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3372913599014282, + 0.0339694507420063, + 0.05566118657588959, + 0.03070123866200447, + 0.13976819813251495, + 0.12917345762252808, + 0.07615149766206741, + 0.06706304103136063, + 0.10508725792169571, + 0.025133315473794937 + ], + "credit_map": [ + 0.0760863795876503, + 0.03436170890927315, + 0.11616300791501999, + 0.018350783735513687, + 0.024458011612296104, + 0.02655528485774994, + 0.012298183515667915, + 0.029018009081482887, + 0.004429181106388569, + 0.00398925319314003, + 0.007079177536070347, + 0.0024914895184338093, + 0.0032622194848954678, + 0.004698519129306078, + 0.0030569657683372498, + 0.004962645471096039, + 0.008000054396688938, + 0.005759159568697214, + 0.011743688955903053, + 0.0037595767062157393, + 0.006513358559459448, + 0.00791194662451744, + 0.0041565359570086, + 0.007816866971552372, + 0.007981155067682266, + 0.005198989994823933, + 0.004552153870463371, + 0.0025241728872060776, + 0.0032863998785614967, + 0.0032319058664143085, + 0.002286807168275118, + 0.0016396547434851527, + 0.014993744902312756, + 0.046610888093709946, + 0.0457378588616848, + 0.007203760091215372, + 0.007736545521765947, + 0.006426665931940079, + 0.0045072948560118675, + 0.006551431026309729, + 0.018714632838964462, + 0.01885785162448883, + 0.024272700771689415, + 0.015035482123494148, + 0.011258207261562347, + 0.0157936979085207, + 0.010971296578645706, + 0.014269591309130192, + 0.0189730916172266, + 0.014169053174555302, + 0.01321644801646471, + 0.006530563347041607, + 0.00645099300891161, + 0.005877898074686527, + 0.006227343808859587, + 0.004706105217337608, + 0.009634179063141346, + 0.009428508579730988, + 0.006720901466906071, + 0.005409717094153166, + 0.012383866123855114, + 0.008702174760401249, + 0.008816445246338844, + 0.005967244505882263, + 0.013989442028105259, + 0.010158111341297626, + 0.019099853932857513, + 0.00917593203485012, + 0.010251468978822231, + 0.01840420812368393, + 0.009278728626668453, + 0.014729510992765427, + 0.003557688556611538, + 0.0026039760559797287, + 0.0027937868144363165, + 0.00283527048304677, + 0.003840662771835923, + 0.004756965674459934, + 0.002514343708753586, + 0.0022306228056550026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -1.4154868125915527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036309524439275265 + ], + "opposite_logprob_tokens": [ + -5.628631114959717 + ], + "logprob_reward": -0.0036309524439275265, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036309524439275265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4463047981262207, + 0.04785969853401184, + 0.03879043832421303, + 0.04058528691530228, + 0.03088962659239769, + 0.11034280806779861, + 0.08565641939640045, + 0.06563065201044083, + 0.07170476019382477, + 0.06223554164171219 + ], + "credit_map": [ + 0.12983572483062744, + 0.03841375559568405, + 0.16734693944454193, + 0.017620010301470757, + 0.028183629736304283, + 0.026506299152970314, + 0.01209067553281784, + 0.02630775421857834, + 0.009468776173889637, + 0.006594865582883358, + 0.006306252907961607, + 0.004805441480129957, + 0.007351398468017578, + 0.003963746130466461, + 0.004218798130750656, + 0.005150419659912586, + 0.005640739109367132, + 0.004321776796132326, + 0.006890288554131985, + 0.003471484873443842, + 0.004746085498481989, + 0.003905395045876503, + 0.004249912686645985, + 0.005564756691455841, + 0.011759936809539795, + 0.006660655606538057, + 0.0061455387622118, + 0.004153045359998941, + 0.0025560217909514904, + 0.0031696436926722527, + 0.004124998115003109, + 0.002015446312725544, + 0.0029965469148010015, + 0.00657317228615284, + 0.008043026551604271, + 0.0030242789071053267, + 0.003248864784836769, + 0.0028925526421517134, + 0.002043948508799076, + 0.0020672352984547615, + 0.02521139569580555, + 0.024512410163879395, + 0.01571226678788662, + 0.009895994327962399, + 0.007469241041690111, + 0.010869486257433891, + 0.0071836793795228004, + 0.009488333016633987, + 0.022763825953006744, + 0.014992610551416874, + 0.016166195273399353, + 0.006364869885146618, + 0.005963965784758329, + 0.005407408811151981, + 0.007200926076620817, + 0.00679661612957716, + 0.008600248023867607, + 0.007681339979171753, + 0.00562316132709384, + 0.006021171808242798, + 0.010422379709780216, + 0.010151002556085587, + 0.007373845670372248, + 0.00975750107318163, + 0.005856392439454794, + 0.007858077995479107, + 0.00986942183226347, + 0.0070960866287350655, + 0.004801189061254263, + 0.01596825383603573, + 0.0049913013353943825, + 0.015264037996530533, + 0.008971082046627998, + 0.00841420330107212, + 0.007231942843645811, + 0.005867931991815567, + 0.008911876007914543, + 0.011813833378255367, + 0.0055146547965705395, + 0.005510019604116678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0036309524439275265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252727307379246 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252727307379246, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252727307379246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6138080954551697, + 0.048991043120622635, + 0.023896334692835808, + 0.019310330972075462, + 0.038306824862957, + 0.08448376506567001, + 0.051022473722696304, + 0.06315447390079498, + 0.03763433173298836, + 0.01939234510064125 + ], + "credit_map": [ + 0.1841699779033661, + 0.05535060539841652, + 0.24782416224479675, + 0.021447064355015755, + 0.030062556266784668, + 0.02562367543578148, + 0.016988029703497887, + 0.03234196826815605, + 0.008198471739888191, + 0.006488199811428785, + 0.009318689815700054, + 0.0037515447475016117, + 0.0040246727876365185, + 0.007690845988690853, + 0.004263805225491524, + 0.005254809744656086, + 0.0039312406443059444, + 0.0027758306823670864, + 0.004252541810274124, + 0.002146703889593482, + 0.0028440533205866814, + 0.0028361710719764233, + 0.0016415799036622047, + 0.0034682077821344137, + 0.002833676990121603, + 0.002759344642981887, + 0.0045380257070064545, + 0.002214306965470314, + 0.0016777722630649805, + 0.0023039940278977156, + 0.0018005608581006527, + 0.0011826461413875222, + 0.00426055770367384, + 0.011004212312400341, + 0.010669456794857979, + 0.002783105941489339, + 0.0030861548148095608, + 0.0022359720896929502, + 0.0022034388966858387, + 0.0020639197900891304, + 0.025033148005604744, + 0.024788748472929, + 0.008264082483947277, + 0.004231791477650404, + 0.004457342438399792, + 0.008548670448362827, + 0.004164915531873703, + 0.0049950554966926575, + 0.015095453709363937, + 0.007879437878727913, + 0.011389176361262798, + 0.003715853439643979, + 0.0034855196718126535, + 0.0035366425290703773, + 0.0036526257172226906, + 0.0022677576635032892, + 0.007211674004793167, + 0.006202087737619877, + 0.005181408021599054, + 0.006364874541759491, + 0.010572650469839573, + 0.009529536589980125, + 0.011085323058068752, + 0.007006905507296324, + 0.005269248504191637, + 0.005181552842259407, + 0.005330756772309542, + 0.0038921895902603865, + 0.0036314097233116627, + 0.0049201794900000095, + 0.004556186031550169, + 0.004852802958339453, + 0.003181676147505641, + 0.002746431389823556, + 0.002426489721983671, + 0.0022832131944596767, + 0.002220677211880684, + 0.00387359457090497, + 0.0016661961562931538, + 0.000994063913822174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.005252727307379246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821038942784071 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.002821038942784071, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821038942784071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22383680939674377, + 0.043667666614055634, + 0.03164517879486084, + 0.0393030121922493, + 0.1441093385219574, + 0.2650313675403595, + 0.07988851517438889, + 0.0895625427365303, + 0.05812856927514076, + 0.024827077984809875 + ], + "credit_map": [ + 0.06226975470781326, + 0.02097560279071331, + 0.07279502600431442, + 0.012927975505590439, + 0.014463846571743488, + 0.016804704442620277, + 0.007731687277555466, + 0.015868205577135086, + 0.0056573073379695415, + 0.006469783838838339, + 0.008728104643523693, + 0.0037381956353783607, + 0.003807488363236189, + 0.007882488891482353, + 0.003350675106048584, + 0.004033621400594711, + 0.006278451066464186, + 0.004818364977836609, + 0.00409849314019084, + 0.002584184054285288, + 0.0038865646347403526, + 0.005367547273635864, + 0.0016840457683429122, + 0.002927531022578478, + 0.010524080134928226, + 0.007472821976989508, + 0.008152552880346775, + 0.0034497929736971855, + 0.0021066502667963505, + 0.001973081612959504, + 0.0035075123887509108, + 0.002116518560796976, + 0.012696323916316032, + 0.04906044527888298, + 0.05092468112707138, + 0.006926028057932854, + 0.007943736389279366, + 0.006340367253869772, + 0.003862877609208226, + 0.006354866083711386, + 0.10432113707065582, + 0.09351707994937897, + 0.014442547224462032, + 0.011431834660470486, + 0.011452807113528252, + 0.013583647087216377, + 0.008174490183591843, + 0.008107801899313927, + 0.029627392068505287, + 0.009551742114126682, + 0.02507787011563778, + 0.003646034048870206, + 0.0025164007674902678, + 0.0035706376656889915, + 0.0032949859742075205, + 0.0026034521870315075, + 0.012450076639652252, + 0.007522364147007465, + 0.01235436461865902, + 0.008411003276705742, + 0.009132718667387962, + 0.012594370171427727, + 0.015413885936141014, + 0.011683757416903973, + 0.0052583422511816025, + 0.004128008149564266, + 0.008776258677244186, + 0.005130015313625336, + 0.003760780207812786, + 0.014060057699680328, + 0.0033670731354504824, + 0.013648031279444695, + 0.0037462941836565733, + 0.0028527243994176388, + 0.00257948599755764, + 0.0028583137318491936, + 0.0029449264984577894, + 0.004683434963226318, + 0.0031128597911447287, + 0.0020490374881774187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.002821038942784071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, extremely low-angle shot at ground level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the primary subject an African-American basketball player wearing a royal blue jersey with gold trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the player depicted with a confident, boyish smile and short, neatly braided hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style impressionistic with soft brushstrokes, showing the player dribbling a bright orange ball that creates ripples and splashes on a water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered as a mosaic of cool blues and warm yellows with a painterly, abstract quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010419422760605812 + ], + "opposite_logprob_tokens": [ + -6.876041889190674 + ], + "logprob_reward": -0.0010419422760605812, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010419422760605812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.501339852809906, + 0.029922664165496826, + 0.03794550895690918, + 0.028714656829833984, + 0.03681624308228493, + 0.10778729617595673, + 0.08395542204380035, + 0.05971962958574295, + 0.06590600311756134, + 0.047892749309539795 + ], + "credit_map": [ + 0.15172970294952393, + 0.04760252311825752, + 0.1741890013217926, + 0.018635321408510208, + 0.023260729387402534, + 0.04018515348434448, + 0.014937566593289375, + 0.030799858272075653, + 0.003908480517566204, + 0.0035692735109478235, + 0.0073979985900223255, + 0.0027863546274602413, + 0.002598542021587491, + 0.0025815763510763645, + 0.0030383244156837463, + 0.004042115528136492, + 0.006345324218273163, + 0.0046518221497535706, + 0.006860814057290554, + 0.002872088924050331, + 0.004742266144603491, + 0.005390426609665155, + 0.0025051895063370466, + 0.0045775785110890865, + 0.006608553230762482, + 0.004849218763411045, + 0.005505233537405729, + 0.002580262254923582, + 0.002517402172088623, + 0.002712382934987545, + 0.002395488088950515, + 0.001546115498058498, + 0.004497006069868803, + 0.00818345881998539, + 0.00885034166276455, + 0.0030279711354523897, + 0.0035989328753203154, + 0.003606380894780159, + 0.0024539825972169638, + 0.0025981704238802195, + 0.01869647204875946, + 0.01766592636704445, + 0.01747419498860836, + 0.009419244714081287, + 0.008745378814637661, + 0.017650166526436806, + 0.007269721012562513, + 0.01086619682610035, + 0.028425607830286026, + 0.012132533825933933, + 0.023291712626814842, + 0.004098049830645323, + 0.004732530098408461, + 0.004067875444889069, + 0.0036311650183051825, + 0.003575948067009449, + 0.00873052328824997, + 0.0077704982832074165, + 0.007206653244793415, + 0.004395938478410244, + 0.011701913550496101, + 0.007359419018030167, + 0.006515396758913994, + 0.006039286032319069, + 0.008939789608120918, + 0.007092402782291174, + 0.01056674588471651, + 0.006034888327121735, + 0.006321537308394909, + 0.011001716367900372, + 0.005219538230448961, + 0.010729379951953888, + 0.012345540337264538, + 0.005652545019984245, + 0.004129743669182062, + 0.002559005981311202, + 0.003215156262740493, + 0.01506931520998478, + 0.0023579150438308716, + 0.0025635287165641785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0010419422760605812 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_tensor([8253, 8253])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_tensor([8253, 8253])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d3c029645fb348eb323f7c0f96a0d37b0de02063 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_tensor([8253, 8253])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot, captured from a low angle close to the water, slowly pans right, following an ancient wooden boat. Its hull, eroded by age, has taken on a deep brown hue, its wood grain clearly visible. On the mirror-smooth surface of the lake, an elderly man in a dark blue linen double-breasted jacket and a conical hat, his back to the camera, steadily and deliberately rowed his oar. The soft, diffused light of dawn penetrates the morning mist, casting a hazy glow over the scene. The distant mountains, blurred by the haze, resemble a flowing ink painting. The boat's bow slices sharply through the water, making almost no sound. Behind the boat's hull, a V-shaped ripple gracefully unfolds, spreading outward into gentle, regular waves that gradually dissipate across the vast expanse. The main colors of the entire picture are cool blue-gray and ink-like black and white, with the warm brown of the wooden boat as the only warm color embellishment, together creating an extremely tranquil, lonely and Zen-like oriental aesthetic atmosphere.", + "seed": 26393, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_8253_seed_26393_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12868080139160157, + "vlm_logprob": -0.12868080139160157, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.12868080139160157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1560322493314743 + ], + "opposite_logprob_tokens": [ + -2.5310323238372803 + ], + "logprob_reward": -0.1560322493314743, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1560322493314743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2553417384624481, + 0.21570467948913574, + 0.05665314570069313, + 0.05974861979484558, + 0.10611367225646973, + 0.05104151740670204, + 0.06363081187009811, + 0.0811251699924469, + 0.040191058069467545, + 0.07044956833124161 + ], + "credit_map": [ + 0.02498742938041687, + 0.027743102982640266, + 0.02960655279457569, + 0.03861154243350029, + 0.03197318688035011, + 0.04171191528439522, + 0.03859621658921242, + 0.02211180329322815, + 0.01749821938574314, + 0.01918231137096882, + 0.015390858054161072, + 0.0194748155772686, + 0.031948771327733994, + 0.038249287754297256, + 0.028585346415638924, + 0.04537507891654968, + 0.004479547962546349, + 0.0037063772324472666, + 0.00522096361964941, + 0.006028652656823397, + 0.008161515928804874, + 0.011461734771728516, + 0.00891950260847807, + 0.008674858137965202, + 0.006150693167001009, + 0.0049574063159525394, + 0.005225065164268017, + 0.0064101433381438255, + 0.007792201358824968, + 0.011913429014384747, + 0.007961595430970192, + 0.009338092990219593, + 0.00876192282885313, + 0.011164288967847824, + 0.009801812469959259, + 0.010257875546813011, + 0.014034304767847061, + 0.020215054973959923, + 0.016771357506513596, + 0.01510705891996622, + 0.004763845819979906, + 0.004029240924865007, + 0.006326362956315279, + 0.005911444313824177, + 0.007657546084374189, + 0.011451579630374908, + 0.0054703205823898315, + 0.005431185010820627, + 0.00472022220492363, + 0.006076161749660969, + 0.007982787676155567, + 0.007857175543904305, + 0.008540485054254532, + 0.012689785100519657, + 0.007603905629366636, + 0.00816029217094183, + 0.00490990374237299, + 0.004394453018903732, + 0.006215584930032492, + 0.006284049712121487, + 0.015045822598040104, + 0.024383261799812317, + 0.00986398570239544, + 0.010028112679719925, + 0.005112179554998875, + 0.0050628394819796085, + 0.0038222309667617083, + 0.004475816618651152, + 0.005044844467192888, + 0.005527263041585684, + 0.004941877909004688, + 0.006204008590430021, + 0.003861210076138377, + 0.006536387372761965, + 0.005927238613367081, + 0.008665072731673717, + 0.011293132789433002, + 0.013697830960154533, + 0.01220091711729765, + 0.008267782628536224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.1560322493314743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42894598841667175 + ], + "opposite_logprob_tokens": [ + -1.0539460182189941 + ], + "logprob_reward": -0.42894598841667175, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42894598841667175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1862204670906067, + 0.13275544345378876, + 0.062180083245038986, + 0.0805911272764206, + 0.04234028235077858, + 0.08981609344482422, + 0.054046809673309326, + 0.15560540556907654, + 0.06897956132888794, + 0.12746471166610718 + ], + "credit_map": [ + 0.01962384395301342, + 0.028420040383934975, + 0.024743635207414627, + 0.03218750283122063, + 0.022483112290501595, + 0.018629251047968864, + 0.021868104115128517, + 0.018264980986714363, + 0.012872218154370785, + 0.023580748587846756, + 0.01038628164678812, + 0.015843819826841354, + 0.013934668153524399, + 0.012555534020066261, + 0.016979413107037544, + 0.02660275250673294, + 0.00544555252417922, + 0.011601656675338745, + 0.006199710536748171, + 0.008369001559913158, + 0.00630510039627552, + 0.006914426572620869, + 0.008126171305775642, + 0.00921846367418766, + 0.005170778837054968, + 0.010170433670282364, + 0.007473586592823267, + 0.010887320153415203, + 0.01072933804243803, + 0.01557251624763012, + 0.0105067677795887, + 0.01008038129657507, + 0.004124423488974571, + 0.0070630330592393875, + 0.00457283528521657, + 0.0037996608298271894, + 0.005486154463142157, + 0.006249850615859032, + 0.005266061518341303, + 0.005778264254331589, + 0.0068668159656226635, + 0.0112950149923563, + 0.00751187838613987, + 0.008705402724444866, + 0.011351257562637329, + 0.012575953267514706, + 0.014646497555077076, + 0.016863269731402397, + 0.003621647832915187, + 0.005020332522690296, + 0.006573993247002363, + 0.007554186042398214, + 0.006281814072281122, + 0.008203445002436638, + 0.007637962698936462, + 0.009153430350124836, + 0.010210961103439331, + 0.008031497709453106, + 0.011431162245571613, + 0.010379158891737461, + 0.02693256549537182, + 0.04267645999789238, + 0.019001012668013573, + 0.026942595839500427, + 0.00889682024717331, + 0.005893206689506769, + 0.00685218907892704, + 0.007597785443067551, + 0.007981415838003159, + 0.009350432083010674, + 0.008884841576218605, + 0.013522871769964695, + 0.009420990943908691, + 0.01806679181754589, + 0.012053132057189941, + 0.015104544349014759, + 0.016703836619853973, + 0.015392824076116085, + 0.021560726687312126, + 0.019161857664585114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.42894598841667175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082914546132088 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082914546132088, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082914546132088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21925751864910126, + 0.12717674672603607, + 0.06478898972272873, + 0.07737994939088821, + 0.24488306045532227, + 0.0739407166838646, + 0.05220719799399376, + 0.052897196263074875, + 0.048355843871831894, + 0.03911278024315834 + ], + "credit_map": [ + 0.027622554451227188, + 0.03855029121041298, + 0.025847844779491425, + 0.0328698456287384, + 0.0253335889428854, + 0.024219797924160957, + 0.023833556100726128, + 0.02098003402352333, + 0.01625443994998932, + 0.028810201212763786, + 0.01149747520685196, + 0.01574302837252617, + 0.013932212255895138, + 0.0128197455778718, + 0.014081485569477081, + 0.014038165099918842, + 0.006015208084136248, + 0.010906249284744263, + 0.010401541367173195, + 0.011591068468987942, + 0.005978416185826063, + 0.005487743299454451, + 0.006222781725227833, + 0.008185984566807747, + 0.00692873727530241, + 0.00936644896864891, + 0.014054614119231701, + 0.013524239882826805, + 0.007588768377900124, + 0.007183942012488842, + 0.010435736738145351, + 0.008297467604279518, + 0.0204385407269001, + 0.03532871603965759, + 0.03502589091658592, + 0.04763834550976753, + 0.024603111669421196, + 0.024232393130660057, + 0.03116316720843315, + 0.02645292319357395, + 0.005609005689620972, + 0.00957421213388443, + 0.008895227685570717, + 0.01024074200540781, + 0.00886482559144497, + 0.01132502406835556, + 0.009735614992678165, + 0.009696072898805141, + 0.004624002147465944, + 0.007589803542941809, + 0.005163926165550947, + 0.006990774068981409, + 0.00615227734670043, + 0.007523071486502886, + 0.007713352795690298, + 0.006449991371482611, + 0.004322050139307976, + 0.007474836893379688, + 0.005555993411689997, + 0.004270023200660944, + 0.007545363157987595, + 0.007304623257368803, + 0.00772776547819376, + 0.008696544915437698, + 0.00422182772308588, + 0.006642011925578117, + 0.006087653338909149, + 0.006419409066438675, + 0.0061732204630970955, + 0.0059618838131427765, + 0.006756292190402746, + 0.006093552801758051, + 0.002975413342937827, + 0.006950559094548225, + 0.003405123483389616, + 0.004448645748198032, + 0.005016830749809742, + 0.005589168984442949, + 0.006519283168017864, + 0.004207759164273739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.016082914546132088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798436909914017 + ], + "opposite_logprob_tokens": [ + -3.5297985076904297 + ], + "logprob_reward": -0.029798436909914017, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798436909914017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13375994563102722, + 0.29215312004089355, + 0.10909134149551392, + 0.08719619363546371, + 0.08953980356454849, + 0.11151839047670364, + 0.05653421953320503, + 0.05842607840895653, + 0.03261759504675865, + 0.029163308441638947 + ], + "credit_map": [ + 0.01460283249616623, + 0.015302412211894989, + 0.015039563179016113, + 0.023964831605553627, + 0.014433076605200768, + 0.015707718208432198, + 0.017119642347097397, + 0.017589876428246498, + 0.024426082149147987, + 0.021652529016137123, + 0.021127022802829742, + 0.03033355064690113, + 0.04337944835424423, + 0.054885607212781906, + 0.03961837664246559, + 0.05673050135374069, + 0.008342647925019264, + 0.011569727212190628, + 0.007038102950900793, + 0.010911510325968266, + 0.014858387410640717, + 0.017927013337612152, + 0.01652628183364868, + 0.02191767282783985, + 0.008489874191582203, + 0.009876808151602745, + 0.011738364584743977, + 0.012559636496007442, + 0.009131420403718948, + 0.009425390511751175, + 0.0129924351349473, + 0.012982265092432499, + 0.007036247290670872, + 0.01101948507130146, + 0.011708492413163185, + 0.012599784880876541, + 0.009953780099749565, + 0.010976001620292664, + 0.015446211211383343, + 0.010799800045788288, + 0.009775879792869091, + 0.012396687641739845, + 0.010464800521731377, + 0.011304805055260658, + 0.015809446573257446, + 0.018248723819851875, + 0.016525860875844955, + 0.016992192715406418, + 0.003089776961132884, + 0.005929968785494566, + 0.004973256029188633, + 0.005091308616101742, + 0.007793258409947157, + 0.013490699231624603, + 0.007274118717759848, + 0.008891835808753967, + 0.005683999974280596, + 0.00920773483812809, + 0.006313589867204428, + 0.007051922846585512, + 0.007799431681632996, + 0.0075912391766905785, + 0.008178635500371456, + 0.006599524989724159, + 0.0029690449591726065, + 0.0034182616509497166, + 0.003616043832153082, + 0.003907016012817621, + 0.0038841748610138893, + 0.0046752276830375195, + 0.005172395147383213, + 0.004975430201739073, + 0.0022215733770281076, + 0.005110722966492176, + 0.003202360589057207, + 0.003673494327813387, + 0.0035752151161432266, + 0.004542117472738028, + 0.003902536816895008, + 0.0029352884739637375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.029798436909914017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544417753815651 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544417753815651, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544417753815651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21998755633831024, + 0.17073741555213928, + 0.11096685379743576, + 0.10499504208564758, + 0.06589354574680328, + 0.08995558321475983, + 0.04428194835782051, + 0.05752912908792496, + 0.07491040974855423, + 0.06074252724647522 + ], + "credit_map": [ + 0.0243659820407629, + 0.03163694217801094, + 0.026285281404852867, + 0.03581590950489044, + 0.025378037244081497, + 0.023667706176638603, + 0.020883172750473022, + 0.03195452690124512, + 0.014305680990219116, + 0.01666109263896942, + 0.022429458796977997, + 0.030542727559804916, + 0.020648635923862457, + 0.02259061671793461, + 0.02139970473945141, + 0.022159500047564507, + 0.008202706463634968, + 0.012982051819562912, + 0.010764761827886105, + 0.012607534416019917, + 0.014424045570194721, + 0.01682412065565586, + 0.019088098779320717, + 0.016073541715741158, + 0.010859356261789799, + 0.014611001126468182, + 0.01047952938824892, + 0.0121536236256361, + 0.011925460770726204, + 0.013140625320374966, + 0.01804475300014019, + 0.013780691660940647, + 0.004728018771857023, + 0.005628928076475859, + 0.006610837299376726, + 0.007343309931457043, + 0.009361027739942074, + 0.014107441529631615, + 0.009270243346691132, + 0.008843742311000824, + 0.0062169525772333145, + 0.009060310199856758, + 0.00928044505417347, + 0.012091475538909435, + 0.011995694600045681, + 0.01262169424444437, + 0.01325503271073103, + 0.015433977358043194, + 0.0033300668001174927, + 0.004965619649738073, + 0.0032315575517714024, + 0.0036362071987241507, + 0.0068635172210633755, + 0.010371643118560314, + 0.006391364149749279, + 0.005491974763572216, + 0.003942529670894146, + 0.0054314471781253815, + 0.0059204380959272385, + 0.0049400776624679565, + 0.007515035104006529, + 0.010105472058057785, + 0.009158643893897533, + 0.01051548682153225, + 0.0075548007152974606, + 0.008340253494679928, + 0.00758273946121335, + 0.007171232253313065, + 0.010481900535523891, + 0.0130684245377779, + 0.011072597466409206, + 0.009638462215662003, + 0.00445198267698288, + 0.013122576288878918, + 0.005454571917653084, + 0.005672716535627842, + 0.007973704487085342, + 0.008938862010836601, + 0.007690111640840769, + 0.007438003085553646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012544417753815651 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, captured from a low angle close to the water, slowly pans right, following an ancient wooden boat. Its hull, eroded by age, has taken on a deep brown hue, its wood grain clearly visible. On the mirror-smooth surface of the lake, an elderly man in a dark blue linen double-breasted jacket and a conical hat, his back to the camera, steadily and deliberately rowed his oar. The soft, diffused light of dawn penetrates the morning mist, casting a hazy glow over the scene. The distant mountains, blurred by the haze, resemble a flowing ink painting. The boat's bow slices sharply through the water, making almost no sound. Behind the boat's hull, a V-shaped ripple gracefully unfolds, spreading outward into gentle, regular waves that gradually dissipate across the vast expanse. The main colors of the entire picture are cool blue-gray and ink-like black and white, with the warm brown of the wooden boat as the only warm color embellishment, together creating an extremely tranquil, lonely and Zen-like oriental aesthetic atmosphere.", + "seed": 26394, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_8253_seed_26394_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11062709130346775, + "vlm_logprob": -0.11062709130346775, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.11062709130346775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10959024727344513 + ], + "opposite_logprob_tokens": [ + -3.1095902919769287 + ], + "logprob_reward": -0.10959024727344513, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.10959024727344513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09407737106084824, + 0.12411779910326004, + 0.2888202667236328, + 0.04432093724608421, + 0.05329485610127449, + 0.15110422670841217, + 0.03687342628836632, + 0.049634601920843124, + 0.08110176771879196, + 0.07665478438138962 + ], + "credit_map": [ + 0.014972159639000893, + 0.011316927149891853, + 0.009719847701489925, + 0.009525774978101254, + 0.013688500970602036, + 0.012851744890213013, + 0.00933178048580885, + 0.012670631520450115, + 0.00874343328177929, + 0.008946706540882587, + 0.008160633035004139, + 0.010026545263826847, + 0.026719991117715836, + 0.017955532297492027, + 0.015086698345839977, + 0.02847825549542904, + 0.021735750138759613, + 0.015573008917272091, + 0.012531425803899765, + 0.012106357142329216, + 0.08746001869440079, + 0.08362995088100433, + 0.02516300044953823, + 0.0306207537651062, + 0.0031459368765354156, + 0.00643385061994195, + 0.0036218520253896713, + 0.0053942217491567135, + 0.010383591055870056, + 0.006697366014122963, + 0.004421854391694069, + 0.004222262650728226, + 0.0032626416068524122, + 0.004234857391566038, + 0.003452262142673135, + 0.00525831151753664, + 0.008672532625496387, + 0.007545728236436844, + 0.01316472701728344, + 0.007703794166445732, + 0.004406201187521219, + 0.005315809976309538, + 0.0061544934287667274, + 0.004692421294748783, + 0.05139641836285591, + 0.0075158001855015755, + 0.05846572667360306, + 0.013157341629266739, + 0.004348532296717167, + 0.006380683276802301, + 0.003448914270848036, + 0.006805482320487499, + 0.003797167679294944, + 0.0035234547685831785, + 0.004731147084385157, + 0.003838048316538334, + 0.006958469748497009, + 0.011013212613761425, + 0.004334697965532541, + 0.006433870177716017, + 0.0048500713892281055, + 0.0051791369915008545, + 0.00474162120372057, + 0.0061235250905156136, + 0.007612519431859255, + 0.012300010770559311, + 0.007149504031985998, + 0.010258490219712257, + 0.012840598821640015, + 0.009486586786806583, + 0.013390861451625824, + 0.008063203655183315, + 0.008078141137957573, + 0.00764063186943531, + 0.004105763044208288, + 0.005058461334556341, + 0.02510669268667698, + 0.004985123872756958, + 0.009946219623088837, + 0.011733745224773884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.10959024727344513 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38711389899253845 + ], + "opposite_logprob_tokens": [ + -1.1371139287948608 + ], + "logprob_reward": -0.38711389899253845, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38711389899253845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07977327704429626, + 0.04865892231464386, + 0.04693630710244179, + 0.06462827324867249, + 0.04111919924616814, + 0.5454959273338318, + 0.023335523903369904, + 0.022534361109137535, + 0.07420127838850021, + 0.0533168651163578 + ], + "credit_map": [ + 0.01344720833003521, + 0.0075220633298158646, + 0.008513495326042175, + 0.00810966081917286, + 0.018272174522280693, + 0.0106123648583889, + 0.00606717262417078, + 0.007229140028357506, + 0.004966088104993105, + 0.006354889366775751, + 0.0066187395714223385, + 0.005791083909571171, + 0.006197237875312567, + 0.005561806261539459, + 0.006321644876152277, + 0.006847431883215904, + 0.0058114174753427505, + 0.006596752442419529, + 0.0053356485441327095, + 0.004700492601841688, + 0.0075124348513782024, + 0.006238106172531843, + 0.005703985691070557, + 0.0050374711863696575, + 0.00577462138608098, + 0.0069212717935442924, + 0.004781010560691357, + 0.006827800069004297, + 0.01125908549875021, + 0.016261015087366104, + 0.006013498175889254, + 0.006789968349039555, + 0.0033453612122684717, + 0.005676330998539925, + 0.0046476018615067005, + 0.0066360607743263245, + 0.00643288716673851, + 0.003863602876663208, + 0.004427003208547831, + 0.006090353708714247, + 0.014502080157399178, + 0.02012798562645912, + 0.01755031757056713, + 0.015559284947812557, + 0.14990738034248352, + 0.025274645537137985, + 0.2161296159029007, + 0.08644464612007141, + 0.001583878300152719, + 0.0037941087502986193, + 0.0017606583423912525, + 0.002193148247897625, + 0.0037962018977850676, + 0.00343060540035367, + 0.003909955732524395, + 0.0028669678140431643, + 0.0021883926820009947, + 0.002811660524457693, + 0.0025874096900224686, + 0.0027530742809176445, + 0.002449351130053401, + 0.002519519068300724, + 0.002877471037209034, + 0.00434748362749815, + 0.008911228738725185, + 0.013607609085738659, + 0.007837115786969662, + 0.009927763603627682, + 0.008190559223294258, + 0.008039424195885658, + 0.008816919289529324, + 0.008870655670762062, + 0.005365333054214716, + 0.0045657530426979065, + 0.004207950085401535, + 0.0037066510412842035, + 0.01288679987192154, + 0.0031669982708990574, + 0.0066764019429683685, + 0.012740975245833397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.38711389899253845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550069019198418 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.012550069019198418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550069019198418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19431567192077637, + 0.13599708676338196, + 0.04030296951532364, + 0.04649445787072182, + 0.05663103610277176, + 0.4321073293685913, + 0.02976519614458084, + 0.013852489180862904, + 0.02068132720887661, + 0.029852475970983505 + ], + "credit_map": [ + 0.031230369582772255, + 0.033972762525081635, + 0.021696971729397774, + 0.020739329978823662, + 0.03182728961110115, + 0.021611230447888374, + 0.01607602834701538, + 0.017161695286631584, + 0.012534632347524166, + 0.016405057162046432, + 0.019784290343523026, + 0.014478110708296299, + 0.018312759697437286, + 0.01676216907799244, + 0.013773076236248016, + 0.02394699864089489, + 0.006225440185517073, + 0.004490318242460489, + 0.003948782104998827, + 0.0038628303445875645, + 0.0064262268133461475, + 0.0059084901586174965, + 0.003974343184381723, + 0.005466539412736893, + 0.0034595143515616655, + 0.003457310376688838, + 0.0024655668530613184, + 0.0032836294267326593, + 0.009357413277029991, + 0.014128034934401512, + 0.0037647937424480915, + 0.00657819677144289, + 0.0033413947094231844, + 0.007634138222783804, + 0.005099875386804342, + 0.0080489506945014, + 0.00781614612787962, + 0.006019280292093754, + 0.009001984260976315, + 0.00966926570981741, + 0.011184867471456528, + 0.01681944541633129, + 0.012931833043694496, + 0.011816502548754215, + 0.1072205975651741, + 0.018492206931114197, + 0.17479988932609558, + 0.07884200662374496, + 0.0014204526087269187, + 0.0026818085461854935, + 0.0017283209599554539, + 0.0024256939068436623, + 0.007132888305932283, + 0.002634104574099183, + 0.00843676459044218, + 0.003305160440504551, + 0.0013039568439126015, + 0.0019112351583316922, + 0.001488331239670515, + 0.0019195249769836664, + 0.0012956192949786782, + 0.0017783514922484756, + 0.0017318740719929338, + 0.002423596568405628, + 0.002401047619059682, + 0.0026485659182071686, + 0.002092901850119233, + 0.00373276649042964, + 0.0029320139437913895, + 0.0023226335179060698, + 0.002446885919198394, + 0.002104512182995677, + 0.002970631467178464, + 0.00491079268977046, + 0.0012947040377184749, + 0.0015783506678417325, + 0.003127348143607378, + 0.0017845447873696685, + 0.007562851067632437, + 0.0066232536919415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.012550069019198418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288307711482048 + ], + "opposite_logprob_tokens": [ + -3.7732882499694824 + ], + "logprob_reward": -0.023288307711482048, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288307711482048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07572122663259506, + 0.11696269363164902, + 0.061299070715904236, + 0.026950621977448463, + 0.07127048075199127, + 0.5311024188995361, + 0.02889915555715561, + 0.019404342398047447, + 0.03813653066754341, + 0.030253466218709946 + ], + "credit_map": [ + 0.008735619485378265, + 0.010591207072138786, + 0.008744850754737854, + 0.008314118720591068, + 0.015867942944169044, + 0.010203568264842033, + 0.005507852416485548, + 0.00775606045499444, + 0.013073621317744255, + 0.011035007424652576, + 0.015649141743779182, + 0.013589716516435146, + 0.015598067082464695, + 0.015091945417225361, + 0.015138021670281887, + 0.017787162214517593, + 0.007924254983663559, + 0.005315624643117189, + 0.006036653649061918, + 0.005232863128185272, + 0.012840896844863892, + 0.01197303831577301, + 0.0060075982473790646, + 0.005968139972537756, + 0.002793233608826995, + 0.003677484579384327, + 0.0021852587815374136, + 0.0028881807811558247, + 0.00414939783513546, + 0.005031075794249773, + 0.0030374638736248016, + 0.003188524628058076, + 0.004111822694540024, + 0.006993776652961969, + 0.006322068627923727, + 0.0097178490832448, + 0.008129887282848358, + 0.009051457978785038, + 0.018076423555612564, + 0.008867190219461918, + 0.014272944070398808, + 0.015191267244517803, + 0.018765348941087723, + 0.011771700344979763, + 0.16507191956043243, + 0.02578994259238243, + 0.2112817019224167, + 0.06895756721496582, + 0.0020869437139481306, + 0.004722326993942261, + 0.0033551291562616825, + 0.0047034998424351215, + 0.0031584773678332567, + 0.003313046647235751, + 0.004719600547105074, + 0.002840129192918539, + 0.001748987240716815, + 0.003346575889736414, + 0.0024489243514835835, + 0.0029523710254579782, + 0.0016648032469674945, + 0.0016994725447148085, + 0.002323554363101721, + 0.0032196526881307364, + 0.003854820504784584, + 0.005255266558378935, + 0.003486264031380415, + 0.004329392220824957, + 0.004829619079828262, + 0.003430341836065054, + 0.007366189267486334, + 0.005584633909165859, + 0.003996478859335184, + 0.00474128220230341, + 0.002921953098848462, + 0.004095669835805893, + 0.003587310202419758, + 0.0029900367371737957, + 0.0049447230994701385, + 0.0029760103207081556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.023288307711482048 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592933520674706 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592933520674706, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592933520674706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10934346914291382, + 0.06583354622125626, + 0.08148662745952606, + 0.13589240610599518, + 0.057598840445280075, + 0.33358678221702576, + 0.03546698018908501, + 0.03769662603735924, + 0.04478348791599274, + 0.09831120818853378 + ], + "credit_map": [ + 0.021620651707053185, + 0.010925949551165104, + 0.010568197816610336, + 0.009504656307399273, + 0.022813014686107635, + 0.014719835482537746, + 0.008518737740814686, + 0.010672440752387047, + 0.007494855206459761, + 0.006869868375360966, + 0.01005132868885994, + 0.008421356789767742, + 0.009052365086972713, + 0.00777471112087369, + 0.007976744323968887, + 0.008192327804863453, + 0.01401480846107006, + 0.009821788407862186, + 0.009400196373462677, + 0.008907063864171505, + 0.011610490269958973, + 0.009729057550430298, + 0.009101777337491512, + 0.00890145730227232, + 0.00829558726400137, + 0.007452777121216059, + 0.011095499619841576, + 0.0181049183011055, + 0.017275122925639153, + 0.027908341959118843, + 0.012937948107719421, + 0.03282221034169197, + 0.004295635037124157, + 0.006644085980951786, + 0.003850985784083605, + 0.00625819293782115, + 0.013162174262106419, + 0.00826060026884079, + 0.007968035526573658, + 0.007159133907407522, + 0.008030219003558159, + 0.00934498943388462, + 0.010641699656844139, + 0.007669354323297739, + 0.08384127169847488, + 0.013149972073733807, + 0.13877280056476593, + 0.06213650107383728, + 0.00373206939548254, + 0.005182198714464903, + 0.0023837375920265913, + 0.004017834551632404, + 0.006419772747904062, + 0.005474606528878212, + 0.003837315132841468, + 0.004419450648128986, + 0.0032582369167357683, + 0.005753424949944019, + 0.0025251880288124084, + 0.004686439875513315, + 0.004492034204304218, + 0.005139388609677553, + 0.005209081340581179, + 0.006632838398218155, + 0.004616567399352789, + 0.006453839596360922, + 0.0038132306654006243, + 0.005006719846278429, + 0.006012066267430782, + 0.007960737682878971, + 0.0056989663280546665, + 0.005221367347985506, + 0.010159911587834358, + 0.007616764400154352, + 0.004345353692770004, + 0.005434765014797449, + 0.03393122926354408, + 0.006184570491313934, + 0.014714880846440792, + 0.01592375338077545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.020592933520674706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, captured from a low angle close to the water, slowly pans right, following an ancient wooden boat. Its hull, eroded by age, has taken on a deep brown hue, its wood grain clearly visible. On the mirror-smooth surface of the lake, an elderly man in a dark blue linen double-breasted jacket and a conical hat, his back to the camera, steadily and deliberately rowed his oar. The soft, diffused light of dawn penetrates the morning mist, casting a hazy glow over the scene. The distant mountains, blurred by the haze, resemble a flowing ink painting. The boat's bow slices sharply through the water, making almost no sound. Behind the boat's hull, a V-shaped ripple gracefully unfolds, spreading outward into gentle, regular waves that gradually dissipate across the vast expanse. The main colors of the entire picture are cool blue-gray and ink-like black and white, with the warm brown of the wooden boat as the only warm color embellishment, together creating an extremely tranquil, lonely and Zen-like oriental aesthetic atmosphere.", + "seed": 26395, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_8253_seed_26395_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11277328319847583, + "vlm_logprob": -0.11277328319847583, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.11277328319847583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06696637719869614 + ], + "opposite_logprob_tokens": [ + -3.4419662952423096 + ], + "logprob_reward": -0.06696637719869614, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.06696637719869614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1955415904521942, + 0.06448950618505478, + 0.07674441486597061, + 0.30551713705062866, + 0.032297082245349884, + 0.05535147339105606, + 0.1428680568933487, + 0.039974261075258255, + 0.03284638375043869, + 0.054370149970054626 + ], + "credit_map": [ + 0.019795391708612442, + 0.025930678471922874, + 0.020550139248371124, + 0.01444643922150135, + 0.03778669983148575, + 0.05128483101725578, + 0.01415445189923048, + 0.011592952534556389, + 0.007668742444366217, + 0.008570956997573376, + 0.00641897227615118, + 0.007092332001775503, + 0.011076709255576134, + 0.009922090917825699, + 0.007973724976181984, + 0.005765974055975676, + 0.012942206114530563, + 0.009249000810086727, + 0.00905423890799284, + 0.004931132309138775, + 0.011849058791995049, + 0.008137129247188568, + 0.008012684062123299, + 0.012568959034979343, + 0.025350889191031456, + 0.019829289987683296, + 0.03265921398997307, + 0.015970470383763313, + 0.04871375486254692, + 0.030716102570295334, + 0.054182324558496475, + 0.07809504866600037, + 0.003318173112347722, + 0.003583968384191394, + 0.0026716317515820265, + 0.0025424628984183073, + 0.010855033062398434, + 0.005509458016604185, + 0.0019565464463084936, + 0.0018598035676404834, + 0.008938271552324295, + 0.006699971854686737, + 0.00610260060057044, + 0.004421434830874205, + 0.01066112332046032, + 0.007049901410937309, + 0.005856877658516169, + 0.005621286109089851, + 0.02982344664633274, + 0.007331062573939562, + 0.043617140501737595, + 0.011585350148379803, + 0.006753590423613787, + 0.008530477993190289, + 0.015575802884995937, + 0.019651174545288086, + 0.005122006870806217, + 0.005873358342796564, + 0.0050487760454416275, + 0.005667753051966429, + 0.005405507981777191, + 0.0036756261251866817, + 0.005399466026574373, + 0.003781761508435011, + 0.0052219959907233715, + 0.0036426910664886236, + 0.004105996340513229, + 0.0042539625428617, + 0.005219863262027502, + 0.003723423695191741, + 0.003942458890378475, + 0.0027359912637621164, + 0.00913530495017767, + 0.007526729721575975, + 0.014425782486796379, + 0.005847683176398277, + 0.004742491990327835, + 0.0042915912345051765, + 0.0044252388179302216, + 0.003975324798375368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.06696637719869614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892611026763916 + ], + "opposite_logprob_tokens": [ + -1.0539261102676392 + ], + "logprob_reward": -0.42892611026763916, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42892611026763916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12308716028928757, + 0.07044313848018646, + 0.046697162091732025, + 0.18511907756328583, + 0.04537183418869972, + 0.0732491984963417, + 0.22472630441188812, + 0.05229217931628227, + 0.06454707682132721, + 0.11446686834096909 + ], + "credit_map": [ + 0.017037760466337204, + 0.015141178853809834, + 0.016728604212403297, + 0.01295129582285881, + 0.022806599736213684, + 0.01504824310541153, + 0.011217598803341389, + 0.01215587742626667, + 0.007329142186790705, + 0.009719894267618656, + 0.007141039241105318, + 0.007964682765305042, + 0.01347495336085558, + 0.01288150716573, + 0.006478745024651289, + 0.005453174468129873, + 0.006409434136003256, + 0.004978100303560495, + 0.005699982866644859, + 0.0042494828812778, + 0.006267119199037552, + 0.004749515559524298, + 0.00597061263397336, + 0.008372918702661991, + 0.012300043366849422, + 0.01480596512556076, + 0.014154408127069473, + 0.010662135668098927, + 0.026501690968871117, + 0.019545406103134155, + 0.03576037287712097, + 0.05138906091451645, + 0.005909455008804798, + 0.0052558560855686665, + 0.005274427589029074, + 0.00428927643224597, + 0.01156255230307579, + 0.005650665145367384, + 0.004115990363061428, + 0.00331360986456275, + 0.009388653561472893, + 0.010597558692097664, + 0.008076157420873642, + 0.006043314002454281, + 0.01486289408057928, + 0.009492916986346245, + 0.007338473107665777, + 0.007449227850884199, + 0.04609426483511925, + 0.010625520721077919, + 0.0867786705493927, + 0.021567652001976967, + 0.013470852747559547, + 0.010600554756820202, + 0.013557396829128265, + 0.022031407803297043, + 0.005459200125187635, + 0.005597799550741911, + 0.005509922280907631, + 0.006199943367391825, + 0.011047900654375553, + 0.005625659134238958, + 0.007294386625289917, + 0.005557362921535969, + 0.008176920004189014, + 0.011923879384994507, + 0.007957058027386665, + 0.005894827656447887, + 0.012504772283136845, + 0.005836492404341698, + 0.005585296545177698, + 0.006667831912636757, + 0.012766965664923191, + 0.011086983606219292, + 0.022495046257972717, + 0.0121169313788414, + 0.015352493152022362, + 0.011644534766674042, + 0.015101468190550804, + 0.013902442529797554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.42892611026763916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018203342333436012 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.018203342333436012, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018203342333436012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08789197355508804, + 0.09984388202428818, + 0.13172073662281036, + 0.31912142038345337, + 0.050344545394182205, + 0.1287347674369812, + 0.10063885897397995, + 0.025739189237356186, + 0.0375547781586647, + 0.01840985007584095 + ], + "credit_map": [ + 0.012466318905353546, + 0.010288127698004246, + 0.011666137725114822, + 0.007077320944517851, + 0.013840874657034874, + 0.015013501979410648, + 0.009206537157297134, + 0.00833315122872591, + 0.008820771239697933, + 0.008103233762085438, + 0.010128798894584179, + 0.006832169834524393, + 0.021409379318356514, + 0.02164478413760662, + 0.012172394432127476, + 0.010732358321547508, + 0.011553962714970112, + 0.009787309914827347, + 0.013819227926433086, + 0.007871635258197784, + 0.024310000240802765, + 0.01685306802392006, + 0.017161449417471886, + 0.030364084988832474, + 0.021618381142616272, + 0.019579997286200523, + 0.02458455227315426, + 0.020396996289491653, + 0.03499101847410202, + 0.030684374272823334, + 0.0644773542881012, + 0.1027887761592865, + 0.00566656282171607, + 0.007183639798313379, + 0.005625271704047918, + 0.005786862224340439, + 0.009895208291709423, + 0.006478214170783758, + 0.005039636045694351, + 0.00466914800927043, + 0.010302704758942127, + 0.009617861360311508, + 0.011693078093230724, + 0.00830724649131298, + 0.01507275365293026, + 0.012126423418521881, + 0.027951555326581, + 0.03366314247250557, + 0.01525872852653265, + 0.007522992789745331, + 0.020138517022132874, + 0.007072219159454107, + 0.01759221963584423, + 0.01061386801302433, + 0.009804550558328629, + 0.012635761871933937, + 0.0032207672484219074, + 0.002634732984006405, + 0.003184488508850336, + 0.0040610493160784245, + 0.0033221314661204815, + 0.0026530432514846325, + 0.0035938338842242956, + 0.0030691437423229218, + 0.005653156898915768, + 0.0068291169591248035, + 0.0044720359146595, + 0.004269047174602747, + 0.0043274336494505405, + 0.003515876829624176, + 0.004177235532552004, + 0.004310873802751303, + 0.002245224779471755, + 0.002224499825388193, + 0.0018726057605817914, + 0.0019910286646336317, + 0.0023163959849625826, + 0.002054167678579688, + 0.0025420086458325386, + 0.003163919784128666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.018203342333436012 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369729220867157 + ], + "opposite_logprob_tokens": [ + -3.4086973667144775 + ], + "logprob_reward": -0.03369729220867157, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369729220867157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.096773661673069, + 0.05041106790304184, + 0.11957389861345291, + 0.202235609292984, + 0.1504141390323639, + 0.12785591185092926, + 0.14209088683128357, + 0.06145958602428436, + 0.028726011514663696, + 0.02045922912657261 + ], + "credit_map": [ + 0.015476740896701813, + 0.011268407106399536, + 0.012237342074513435, + 0.00886119157075882, + 0.015137270092964172, + 0.013839811086654663, + 0.009730279445648193, + 0.010222616605460644, + 0.006775757297873497, + 0.00606708275154233, + 0.005134995095431805, + 0.005562521051615477, + 0.0076478952541947365, + 0.008846545591950417, + 0.00550429942086339, + 0.004871968645602465, + 0.01442967914044857, + 0.0076420605182647705, + 0.014102237299084663, + 0.008654060773551464, + 0.020050855353474617, + 0.0129344891756773, + 0.014679338783025742, + 0.02708118036389351, + 0.020380737259984016, + 0.01263575442135334, + 0.018131842836737633, + 0.010958848521113396, + 0.028660546988248825, + 0.01440392341464758, + 0.04348486289381981, + 0.053579073399305344, + 0.042025282979011536, + 0.014876759611070156, + 0.0238985363394022, + 0.01194815244525671, + 0.014259114861488342, + 0.012190285138785839, + 0.013823991641402245, + 0.01739201880991459, + 0.01344151608645916, + 0.014598121866583824, + 0.011002011597156525, + 0.007680908776819706, + 0.03418897092342377, + 0.019288195297122, + 0.013437192887067795, + 0.014218996278941631, + 0.020066894590854645, + 0.008056185208261013, + 0.035694535821676254, + 0.009511590003967285, + 0.009186902083456516, + 0.010384010151028633, + 0.02368549071252346, + 0.02550528012216091, + 0.00885308813303709, + 0.010779162868857384, + 0.006105656735599041, + 0.006682612467557192, + 0.006627533584833145, + 0.005452015902847052, + 0.008122910745441914, + 0.008836605586111546, + 0.004073026590049267, + 0.004653409123420715, + 0.00339679978787899, + 0.003660328686237335, + 0.0032127508893609047, + 0.0027848503086715937, + 0.003868219442665577, + 0.0030766241252422333, + 0.0029295010026544333, + 0.0029616612009704113, + 0.002715508220717311, + 0.0024027074687182903, + 0.001422167057171464, + 0.0018929553916677833, + 0.0031470267567783594, + 0.0029877019114792347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.03369729220867157 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607329398393631 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607329398393631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607329398393631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15174593031406403, + 0.21426334977149963, + 0.05903344601392746, + 0.11153092980384827, + 0.06276597082614899, + 0.07381276041269302, + 0.21219828724861145, + 0.04620806872844696, + 0.03711872175335884, + 0.03132258355617523 + ], + "credit_map": [ + 0.016888994723558426, + 0.016824867576360703, + 0.016999702900648117, + 0.01220825407654047, + 0.016403421759605408, + 0.01689300686120987, + 0.019713489338755608, + 0.03581418842077255, + 0.02221132069826126, + 0.05526198446750641, + 0.025930603966116905, + 0.022363074123859406, + 0.03537571430206299, + 0.02203940786421299, + 0.01944103091955185, + 0.011640213429927826, + 0.009560196660459042, + 0.0060004726983606815, + 0.007649868726730347, + 0.005615605041384697, + 0.008162876591086388, + 0.007628439925611019, + 0.006808080710470676, + 0.007607906125485897, + 0.011708595789968967, + 0.01301996223628521, + 0.013978907838463783, + 0.006445867940783501, + 0.023229345679283142, + 0.01578369364142418, + 0.011550331488251686, + 0.015814227983355522, + 0.008473705500364304, + 0.006857761647552252, + 0.007275895215570927, + 0.004721152596175671, + 0.015352770686149597, + 0.00793514121323824, + 0.006229173857718706, + 0.005920371972024441, + 0.009644441306591034, + 0.008546962402760983, + 0.010043520480394363, + 0.006649109069257975, + 0.011493229307234287, + 0.008605016395449638, + 0.009593631140887737, + 0.009236845187842846, + 0.045058440417051315, + 0.013794710859656334, + 0.0733441710472107, + 0.01911785639822483, + 0.01551499217748642, + 0.011304894462227821, + 0.01444033533334732, + 0.019622892141342163, + 0.005189769435673952, + 0.005363864358514547, + 0.0044576264917850494, + 0.0047027114778757095, + 0.008978467434644699, + 0.004003029782325029, + 0.007468696217983961, + 0.006043902598321438, + 0.0066680884920060635, + 0.005548643879592419, + 0.0039026266895234585, + 0.004119012504816055, + 0.004947125446051359, + 0.002999149030074477, + 0.004791893530637026, + 0.004142183344811201, + 0.004053544253110886, + 0.0030385854188352823, + 0.003673853585496545, + 0.00312704686075449, + 0.003773141885176301, + 0.00359866744838655, + 0.004350318107753992, + 0.0057074264623224735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.01607329398393631 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, captured from a low angle close to the water, slowly pans right, following an ancient wooden boat. Its hull, eroded by age, has taken on a deep brown hue, its wood grain clearly visible. On the mirror-smooth surface of the lake, an elderly man in a dark blue linen double-breasted jacket and a conical hat, his back to the camera, steadily and deliberately rowed his oar. The soft, diffused light of dawn penetrates the morning mist, casting a hazy glow over the scene. The distant mountains, blurred by the haze, resemble a flowing ink painting. The boat's bow slices sharply through the water, making almost no sound. Behind the boat's hull, a V-shaped ripple gracefully unfolds, spreading outward into gentle, regular waves that gradually dissipate across the vast expanse. The main colors of the entire picture are cool blue-gray and ink-like black and white, with the warm brown of the wooden boat as the only warm color embellishment, together creating an extremely tranquil, lonely and Zen-like oriental aesthetic atmosphere.", + "seed": 26396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_3/index_8253_seed_26396_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06155738867819309, + "vlm_logprob": -0.06155738867819309, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.06155738867819309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08206311613321304 + ], + "opposite_logprob_tokens": [ + -3.9570631980895996 + ], + "logprob_reward": -0.08206311613321304, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.08206311613321304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10483259707689285, + 0.14859795570373535, + 0.06177666410803795, + 0.32522690296173096, + 0.13900378346443176, + 0.018546951934695244, + 0.06157898157835007, + 0.028571005910634995, + 0.05181008577346802, + 0.060055091977119446 + ], + "credit_map": [ + 0.02746925689280033, + 0.009455968625843525, + 0.009490761905908585, + 0.007600798737257719, + 0.01729300245642662, + 0.014165209606289864, + 0.005549105815589428, + 0.013808481395244598, + 0.023548590019345284, + 0.010740183293819427, + 0.01010109018534422, + 0.007913700304925442, + 0.028229963034391403, + 0.04246719926595688, + 0.0078023504465818405, + 0.01779487356543541, + 0.006731012370437384, + 0.005752106662839651, + 0.0043734400533139706, + 0.006734424736350775, + 0.011156941764056683, + 0.015438265167176723, + 0.0047161150723695755, + 0.006874351762235165, + 0.05844379588961601, + 0.03455585613846779, + 0.018223553895950317, + 0.019288325682282448, + 0.06325334310531616, + 0.08443593233823776, + 0.017483053728938103, + 0.029543034732341766, + 0.028550000861287117, + 0.012689664959907532, + 0.013164020143449306, + 0.015444828197360039, + 0.02002444863319397, + 0.028928490355610847, + 0.010098415426909924, + 0.010103903710842133, + 0.002942737890407443, + 0.00256395828910172, + 0.001759529346600175, + 0.0018639623885974288, + 0.0020844489336013794, + 0.0035592433996498585, + 0.001641131122596562, + 0.002131939399987459, + 0.012868440710008144, + 0.009237542748451233, + 0.0031606918200850487, + 0.004999071359634399, + 0.004476980771869421, + 0.020944098010659218, + 0.0032516554929316044, + 0.0026404999662190676, + 0.0037373187951743603, + 0.006801219191402197, + 0.0020771780982613564, + 0.0020208091009408236, + 0.004402198363095522, + 0.006140340585261583, + 0.002287547569721937, + 0.0011043914128094912, + 0.008285008370876312, + 0.007034089416265488, + 0.006023827008903027, + 0.005880625452846289, + 0.006314639933407307, + 0.010766550898551941, + 0.0036222978960722685, + 0.003883042372763157, + 0.007448548451066017, + 0.007371044717729092, + 0.0077375550754368305, + 0.008852296508848667, + 0.007092054467648268, + 0.010716772638261318, + 0.006058854050934315, + 0.004777961876243353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.08206311613321304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282917976379395 + ], + "opposite_logprob_tokens": [ + -2.017829179763794 + ], + "logprob_reward": -0.14282917976379395, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282917976379395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09863842278718948, + 0.16485537588596344, + 0.06084074825048447, + 0.1478191316127777, + 0.1320977807044983, + 0.06174010783433914, + 0.11617778986692429, + 0.03747820109128952, + 0.10152420401573181, + 0.07882823795080185 + ], + "credit_map": [ + 0.005352465435862541, + 0.00928399059921503, + 0.00590891856700182, + 0.0052666980773210526, + 0.03111698292195797, + 0.032333411276340485, + 0.0032947417348623276, + 0.006081215105950832, + 0.025017745792865753, + 0.013079216703772545, + 0.0113524179905653, + 0.009702011942863464, + 0.029839904978871346, + 0.04547061026096344, + 0.011589295230805874, + 0.018804175779223442, + 0.010524546727538109, + 0.009219732135534286, + 0.005092247389256954, + 0.006445565260946751, + 0.007956086657941341, + 0.008863205090165138, + 0.006292204838246107, + 0.006447163876146078, + 0.02379940263926983, + 0.014690498821437359, + 0.0065254028886556625, + 0.007909897714853287, + 0.030442945659160614, + 0.04421267285943031, + 0.007272840477526188, + 0.012965471483767033, + 0.02069948799908161, + 0.008421617560088634, + 0.015985455363988876, + 0.01798505149781704, + 0.01691516488790512, + 0.02376760169863701, + 0.017727410420775414, + 0.010595982894301414, + 0.00882033072412014, + 0.010496055707335472, + 0.007644576486200094, + 0.00924964714795351, + 0.005700880661606789, + 0.008493883535265923, + 0.006379075814038515, + 0.004955661948770285, + 0.025818869471549988, + 0.016003813594579697, + 0.008831650018692017, + 0.007923183031380177, + 0.010020171292126179, + 0.036673836410045624, + 0.006223019678145647, + 0.004683241248130798, + 0.004402798600494862, + 0.008357508108019829, + 0.0033026840537786484, + 0.004620078485459089, + 0.0042846049182116985, + 0.007789480499923229, + 0.0029341585468500853, + 0.0017868910217657685, + 0.012562093324959278, + 0.01354629173874855, + 0.005787290167063475, + 0.013003693893551826, + 0.02133096009492874, + 0.013931350782513618, + 0.016330914571881294, + 0.005031603388488293, + 0.008521533571183681, + 0.012365632690489292, + 0.006165916100144386, + 0.011927922256290913, + 0.012571076862514019, + 0.007859566248953342, + 0.015038385055959225, + 0.004378203768283129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.14282917976379395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980387583374977 + ], + "opposite_logprob_tokens": [ + -3.529803991317749 + ], + "logprob_reward": -0.02980387583374977, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980387583374977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3317035734653473, + 0.13673827052116394, + 0.045937679708004, + 0.11152340471744537, + 0.06626799702644348, + 0.016696399077773094, + 0.026382021605968475, + 0.011801122687757015, + 0.13788020610809326, + 0.11506930738687515 + ], + "credit_map": [ + 0.022665390744805336, + 0.022591685876250267, + 0.016544606536626816, + 0.01875394582748413, + 0.10660097002983093, + 0.1171371340751648, + 0.011076346971094608, + 0.016333511099219322, + 0.018012966960668564, + 0.01567271165549755, + 0.011148861609399319, + 0.009800270199775696, + 0.022305533289909363, + 0.03255101293325424, + 0.010509870015084743, + 0.016737040132284164, + 0.0071020787581801414, + 0.005233635660260916, + 0.0034271851181983948, + 0.004406267311424017, + 0.007201774045825005, + 0.01180975791066885, + 0.0027973910328000784, + 0.0039595854468643665, + 0.020383046939969063, + 0.010025642812252045, + 0.00453111482784152, + 0.005403340328484774, + 0.022905590012669563, + 0.031813669949769974, + 0.00541523564606905, + 0.011045760475099087, + 0.008639737032353878, + 0.009937605820596218, + 0.0074251191690564156, + 0.007917621172964573, + 0.00836280919611454, + 0.009591840207576752, + 0.008133461698889732, + 0.0062597994692623615, + 0.0026197831612080336, + 0.002527373842895031, + 0.001643455820158124, + 0.0016917899483814836, + 0.0029361164197325706, + 0.002797919325530529, + 0.0013182490365579724, + 0.00116171152330935, + 0.004097516182810068, + 0.0042841192334890366, + 0.002253391081467271, + 0.0023154665250331163, + 0.003529944224283099, + 0.0065574622713029385, + 0.001784027903340757, + 0.001560091390274465, + 0.0014464177656918764, + 0.0021651512943208218, + 0.001005368772894144, + 0.0015878347912803292, + 0.0013322369195520878, + 0.0022971490398049355, + 0.001385594834573567, + 0.0005813685129396617, + 0.013914335519075394, + 0.015117271803319454, + 0.011164060793817043, + 0.01986543834209442, + 0.027222895994782448, + 0.0161028653383255, + 0.02511836588382721, + 0.009374969638884068, + 0.012457350268959999, + 0.01534140296280384, + 0.008209518156945705, + 0.013040468096733093, + 0.027934949845075607, + 0.015144581906497478, + 0.018014904111623764, + 0.004926143679767847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.02980387583374977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284928873181343 + ], + "opposite_logprob_tokens": [ + -3.773284912109375 + ], + "logprob_reward": -0.023284928873181343, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284928873181343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18462589383125305, + 0.28464022278785706, + 0.0557534359395504, + 0.12782302498817444, + 0.16168531775474548, + 0.02980009652674198, + 0.024332119151949883, + 0.02577427215874195, + 0.03889545425772667, + 0.06667015701532364 + ], + "credit_map": [ + 0.01986825279891491, + 0.012050171382725239, + 0.008388511836528778, + 0.01171933114528656, + 0.05722331628203392, + 0.06050567328929901, + 0.006360108032822609, + 0.008510539308190346, + 0.042717888951301575, + 0.024181537330150604, + 0.01882409118115902, + 0.015652276575565338, + 0.05473757162690163, + 0.08170804381370544, + 0.010742413811385632, + 0.036076370626688004, + 0.005118940956890583, + 0.011043259873986244, + 0.006041240878403187, + 0.005852899048477411, + 0.009206059388816357, + 0.011637736111879349, + 0.0030121486634016037, + 0.0038411449640989304, + 0.017862627282738686, + 0.011182762682437897, + 0.00857890211045742, + 0.011519886553287506, + 0.023764654994010925, + 0.034466005861759186, + 0.007729607634246349, + 0.012718583457171917, + 0.029831537976861, + 0.010789762251079082, + 0.015740470960736275, + 0.018463967368006706, + 0.025147661566734314, + 0.04007919132709503, + 0.009859309531748295, + 0.011773431673645973, + 0.004596095997840166, + 0.003971803933382034, + 0.002854405902326107, + 0.0036258178297430277, + 0.0030877599492669106, + 0.00523832393810153, + 0.002933652140200138, + 0.003492240095511079, + 0.0029942637775093317, + 0.0032985720317810774, + 0.003079358022660017, + 0.0032236818224191666, + 0.0027260128408670425, + 0.004142018035054207, + 0.002465985482558608, + 0.0024022269062697887, + 0.003096085973083973, + 0.004119144752621651, + 0.0022189226001501083, + 0.0023439964279532433, + 0.0047537097707390785, + 0.005052807275205851, + 0.002866534749045968, + 0.0013230706099420786, + 0.005264162085950375, + 0.004706427920609713, + 0.0032769115641713142, + 0.004219961352646351, + 0.007457879837602377, + 0.0055877757258713245, + 0.006407224573194981, + 0.0019751114305108786, + 0.008175195194780827, + 0.007954939268529415, + 0.005937603302299976, + 0.009171693585813046, + 0.012939107604324818, + 0.008705664426088333, + 0.009730384685099125, + 0.004055570811033249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.023284928873181343 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle close to the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan right while following an ancient wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly man wearing a dark blue linen double-breasted jacket and a conical hat visible with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man rowing steadily and deliberately, creating a V-shaped ripple behind the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cool blue-gray and ink-like black and white tones with the wooden boat as the only warm color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980584278702736 + ], + "opposite_logprob_tokens": [ + -3.529805898666382 + ], + "logprob_reward": -0.02980584278702736, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980584278702736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4227065443992615, + 0.0274910107254982, + 0.07135040313005447, + 0.17306584119796753, + 0.0648493841290474, + 0.06361114978790283, + 0.044500816613435745, + 0.04173842445015907, + 0.05267901346087456, + 0.03800748288631439 + ], + "credit_map": [ + 0.02255868911743164, + 0.020097877830266953, + 0.01256473083049059, + 0.010808875784277916, + 0.1646137833595276, + 0.17109669744968414, + 0.008308973163366318, + 0.012656849808990955, + 0.003269411623477936, + 0.005915987305343151, + 0.002827737247571349, + 0.003397933440282941, + 0.003578724805265665, + 0.003633142914623022, + 0.0021713038440793753, + 0.002696764888241887, + 0.011056894436478615, + 0.0071928189136087894, + 0.006677725352346897, + 0.005938715301454067, + 0.011844970285892487, + 0.015808770433068275, + 0.004575689788907766, + 0.00825481303036213, + 0.031654827296733856, + 0.016645990312099457, + 0.00830178614705801, + 0.008876965381205082, + 0.03510991856455803, + 0.048075366765260696, + 0.008077049627900124, + 0.01632392592728138, + 0.011580737307667732, + 0.0045100012794137, + 0.0060612503439188, + 0.0066490573808550835, + 0.009578178636729717, + 0.013646104373037815, + 0.007329274900257587, + 0.005494772456586361, + 0.013588879257440567, + 0.008222981356084347, + 0.0041578058153390884, + 0.005436671897768974, + 0.009854784235358238, + 0.013180463574826717, + 0.005232610274106264, + 0.003936941269785166, + 0.007300011347979307, + 0.008064590394496918, + 0.00503529142588377, + 0.004643382970243692, + 0.003504624357447028, + 0.010123327374458313, + 0.003100781235843897, + 0.0027288051787763834, + 0.004536664579063654, + 0.009799957275390625, + 0.00313260848633945, + 0.003943471238017082, + 0.0072858259081840515, + 0.007132701575756073, + 0.003637698944658041, + 0.0022694929502904415, + 0.004648595117032528, + 0.005396400578320026, + 0.003275127848610282, + 0.007886727340519428, + 0.012423778884112835, + 0.00594362011179328, + 0.009931132197380066, + 0.003173628356307745, + 0.0052115763537585735, + 0.004553128499537706, + 0.0029777854215353727, + 0.00411522202193737, + 0.0075547886081039906, + 0.005389770958572626, + 0.004829028155654669, + 0.0033761803060770035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.02980584278702736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_4/index_tensor([2980, 2980])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_4/index_tensor([2980, 2980])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f80a33e700ace6b98474c80ee6508fefaa788c42 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_4/index_tensor([2980, 2980])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Embracing the meticulously hand-painted aesthetic of Ghibli animation, a striking overhead fisheye lens places the viewer on the ground, facing a vibrant, multicolored, segmented giant centipede from Madagascar. It nestles in a humid, vibrant, fantasy forest floor, nestled against a backdrop of giant, mossy ferns and shimmering, exotic mushrooms. The exaggerated distortion of the fisheye lens causes these plants to curve and stretch toward the edge of the frame, forming an enveloping, dizzying dome of green. The centipede's carapace displays an intricate pattern of deep russet and bright orange. Each smooth segment reflects dappled light filtering through the gaps in the forest canopy, and countless pairs of yellow legs connect the segments, each in a frenzy of activity. First, its long, segmented body arches back sharply into a tense \"C\" shape, its massive, venom-gleaming black mandibules poised at the front. Then, in a flash, it launches its entire body forward with explosive speed, hurtling towards the camera. During this rapid movement, its head and maxillipedia are disproportionately magnified in front of the camera, as if they are about to break through the screen. Countless pairs of legs are frantically paddling in the air, forming a blurred afterimage. The entire picture is full of dynamic tension and a suffocating sense of oppression from the huge creature, perfectly reproducing the classic depiction of the awe and fear of natural forces in Ghibli animation.", + "seed": 5084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_4/index_2980_seed_5084_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.032050031097605826, + "vlm_logprob": -0.032050031097605826, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.032050031097605826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11408942192792892 + ], + "opposite_logprob_tokens": [ + -2.3640894889831543 + ], + "logprob_reward": -0.11408942192792892, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11408942192792892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20534098148345947, + 0.056986045092344284, + 0.04391660913825035, + 0.09066414088010788, + 0.04787512868642807, + 0.047991715371608734, + 0.07770014554262161, + 0.036089520901441574, + 0.056650832295417786, + 0.3367849588394165 + ], + "credit_map": [ + 0.024221336469054222, + 0.020778793841600418, + 0.022769348695874214, + 0.027262859046459198, + 0.01998385228216648, + 0.03142346441745758, + 0.023516133427619934, + 0.03538518399000168, + 0.006771221291273832, + 0.006452034693211317, + 0.008710930123925209, + 0.008505292236804962, + 0.00682627921923995, + 0.007654774468392134, + 0.004395550116896629, + 0.007669955492019653, + 0.0033743928652256727, + 0.003433234291151166, + 0.007017100229859352, + 0.00366150657646358, + 0.0036280592903494835, + 0.014059862121939659, + 0.004009302239865065, + 0.00473314942792058, + 0.013774131424725056, + 0.012077112682163715, + 0.01342326495796442, + 0.007912147790193558, + 0.012167152017354965, + 0.015486341901123524, + 0.006348692812025547, + 0.00947529636323452, + 0.005914431996643543, + 0.007938560098409653, + 0.0063990820199251175, + 0.004109203815460205, + 0.004106131847947836, + 0.011130720376968384, + 0.0045393179170787334, + 0.003737674094736576, + 0.010605314746499062, + 0.005484116729348898, + 0.0041583264246582985, + 0.004641607869416475, + 0.008317941799759865, + 0.005548697430640459, + 0.0058295889757573605, + 0.003406114876270294, + 0.012303262017667294, + 0.016531771048903465, + 0.006166534498333931, + 0.007263700012117624, + 0.008398965001106262, + 0.012599298730492592, + 0.0075461785309016705, + 0.006890422664582729, + 0.0060998788103461266, + 0.0032130805775523186, + 0.004315638914704323, + 0.004779582843184471, + 0.004619455896317959, + 0.004406959284096956, + 0.0039855726063251495, + 0.0046693505719304085, + 0.0032434077002108097, + 0.002177078276872635, + 0.0037449155934154987, + 0.004610203206539154, + 0.006584499962627888, + 0.018393851816654205, + 0.003532643662765622, + 0.01436422299593687, + 0.02319621853530407, + 0.11970949172973633, + 0.011663499288260937, + 0.12458266317844391, + 0.01379692368209362, + 0.013728258199989796, + 0.015328380279242992, + 0.014779441058635712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.11408942192792892 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549244798719883 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549244798719883, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549244798719883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12255264073610306, + 0.03833015635609627, + 0.0721808671951294, + 0.1608540564775467, + 0.05883048474788666, + 0.2287282645702362, + 0.11388801038265228, + 0.05581844225525856, + 0.08311503380537033, + 0.06570205092430115 + ], + "credit_map": [ + 0.01234806701540947, + 0.01228990126401186, + 0.012342848815023899, + 0.015380068682134151, + 0.016677312552928925, + 0.015104205347597599, + 0.017143717035651207, + 0.021266505122184753, + 0.004129635635763407, + 0.0032515490893274546, + 0.004726582206785679, + 0.005839917343109846, + 0.005635865963995457, + 0.004984014201909304, + 0.004289146047085524, + 0.005473443306982517, + 0.004884306341409683, + 0.004869697615504265, + 0.012169393710792065, + 0.005029592663049698, + 0.004518413450568914, + 0.023776967078447342, + 0.008141212165355682, + 0.008791273459792137, + 0.019717805087566376, + 0.02157633565366268, + 0.024921350181102753, + 0.01683589443564415, + 0.021663378924131393, + 0.02498752810060978, + 0.015241941437125206, + 0.015909818932414055, + 0.0070674787275493145, + 0.005446611903607845, + 0.011057519353926182, + 0.0051469961181283, + 0.0054174866527318954, + 0.012089521624147892, + 0.0066888281144201756, + 0.005916034337133169, + 0.058875732123851776, + 0.023484056815505028, + 0.02202751301229, + 0.020025646314024925, + 0.03861038386821747, + 0.026613468304276466, + 0.02287106402218342, + 0.016220377758145332, + 0.014556077308952808, + 0.020584408193826675, + 0.012158378958702087, + 0.011949797160923481, + 0.01290261372923851, + 0.014493483118712902, + 0.012757401913404465, + 0.014485836960375309, + 0.007952864281833172, + 0.008059526793658733, + 0.005228338297456503, + 0.00637452257797122, + 0.008384525775909424, + 0.007435675244778395, + 0.006279428023844957, + 0.006103555206209421, + 0.006591116078197956, + 0.00811318214982748, + 0.016700739040970802, + 0.010214343667030334, + 0.014089616015553474, + 0.013044104911386967, + 0.006130751688033342, + 0.008231173269450665, + 0.006184947211295366, + 0.012034806422889233, + 0.013036064803600311, + 0.008486578240990639, + 0.006065614987164736, + 0.005011115223169327, + 0.007576644420623779, + 0.007306272629648447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.012549244798719883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024901593569666147 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.0024901593569666147, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024901593569666147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10330003499984741, + 0.04841035604476929, + 0.02709786221385002, + 0.08052024990320206, + 0.09691523015499115, + 0.09127519279718399, + 0.0838533490896225, + 0.07410305738449097, + 0.049681905657052994, + 0.3448427617549896 + ], + "credit_map": [ + 0.015100739896297455, + 0.016935158520936966, + 0.010387406684458256, + 0.011615539900958538, + 0.013182315044105053, + 0.01319863460958004, + 0.008792361244559288, + 0.01408787164837122, + 0.006852045655250549, + 0.006709451787173748, + 0.006231103092432022, + 0.005931555759161711, + 0.00654801819473505, + 0.0063081406988203526, + 0.004527983255684376, + 0.005302050616592169, + 0.0035571700427681208, + 0.0033525191247463226, + 0.0028410672675818205, + 0.002861808519810438, + 0.004758534487336874, + 0.0039206333458423615, + 0.0021965799387544394, + 0.003609549021348357, + 0.007250074762851, + 0.007857116870582104, + 0.012844699434936047, + 0.008422241546213627, + 0.009257304482161999, + 0.017902767285704613, + 0.007521233055740595, + 0.00946480967104435, + 0.00815322995185852, + 0.009710238315165043, + 0.021925631910562515, + 0.00888012070208788, + 0.007147775497287512, + 0.02340017817914486, + 0.009878003969788551, + 0.00782004650682211, + 0.020093737170100212, + 0.009535228833556175, + 0.009285945445299149, + 0.008913214318454266, + 0.015259642153978348, + 0.009190451353788376, + 0.01107875257730484, + 0.007918215356767178, + 0.011534828692674637, + 0.015428818762302399, + 0.008206152357161045, + 0.00798103678971529, + 0.008005457930266857, + 0.01068057306110859, + 0.01195305772125721, + 0.010063420049846172, + 0.011985491961240768, + 0.011725714430212975, + 0.007380492519587278, + 0.007736119441688061, + 0.008472041226923466, + 0.009577697142958641, + 0.009285402484238148, + 0.007940093986690044, + 0.004254850093275309, + 0.00417083827778697, + 0.0066848029382526875, + 0.007643586024641991, + 0.008260219357907772, + 0.008601674810051918, + 0.004886763170361519, + 0.0051791672594845295, + 0.02256268449127674, + 0.11752115190029144, + 0.014058174565434456, + 0.13087017834186554, + 0.013007041998207569, + 0.015798797830939293, + 0.016592558473348618, + 0.014432153664529324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0024901593569666147 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783742502331734 + ], + "opposite_logprob_tokens": [ + -3.5297837257385254 + ], + "logprob_reward": -0.029783742502331734, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783742502331734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08944954723119736, + 0.0414087139070034, + 0.02583216316998005, + 0.03955195099115372, + 0.0668909028172493, + 0.2035461962223053, + 0.02103709988296032, + 0.034432582557201385, + 0.04143979400396347, + 0.4364110827445984 + ], + "credit_map": [ + 0.012303897179663181, + 0.013982854783535004, + 0.009543725289404392, + 0.0133597943931818, + 0.011267692781984806, + 0.011286287568509579, + 0.007678950671106577, + 0.01002634596079588, + 0.004985544364899397, + 0.004559079650789499, + 0.004683160223066807, + 0.004587154369801283, + 0.0060429647564888, + 0.0062943450175225735, + 0.0044784038327634335, + 0.005778060294687748, + 0.003217459423467517, + 0.002881936728954315, + 0.002672486240044236, + 0.0036140463780611753, + 0.0033529680222272873, + 0.0032616571988910437, + 0.0030640275217592716, + 0.003767581656575203, + 0.005444140173494816, + 0.00472638662904501, + 0.004755302798002958, + 0.0037879999727010727, + 0.004979962948709726, + 0.005678020883351564, + 0.004309523850679398, + 0.005870616529136896, + 0.007937169633805752, + 0.007428267039358616, + 0.0069430856965482235, + 0.009505434893071651, + 0.008767970837652683, + 0.007012084126472473, + 0.0089674461632967, + 0.010329442098736763, + 0.04696931689977646, + 0.02058662660419941, + 0.020821724086999893, + 0.02098924107849598, + 0.032237451523542404, + 0.021826380863785744, + 0.02263115532696247, + 0.01748429611325264, + 0.002441415097564459, + 0.003309008665382862, + 0.0021476156543940306, + 0.003772876225411892, + 0.0018995939753949642, + 0.0027507238555699587, + 0.002895960584282875, + 0.0018199071055278182, + 0.003984689712524414, + 0.0036657387390732765, + 0.0044756862334907055, + 0.004060317762196064, + 0.0044698044657707214, + 0.005878719966858625, + 0.0037427961360663176, + 0.0041548279114067554, + 0.004521544557064772, + 0.0034145028330385685, + 0.007636807393282652, + 0.005412223283201456, + 0.0062325396575033665, + 0.005242951679974794, + 0.004440672229975462, + 0.004538552835583687, + 0.034462861716747284, + 0.13104034960269928, + 0.02665037102997303, + 0.15262094140052795, + 0.02388341911137104, + 0.02257680892944336, + 0.022192658856511116, + 0.022983653470873833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.029783742502331734 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013375869020819664 + ], + "opposite_logprob_tokens": [ + -6.62633752822876 + ], + "logprob_reward": -0.0013375869020819664, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013375869020819664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03113606572151184, + 0.011770878918468952, + 0.007597567047923803, + 0.024178149178624153, + 0.014268076978623867, + 0.0497831329703331, + 0.01676628738641739, + 0.01255924440920353, + 0.07560787349939346, + 0.7563326954841614 + ], + "credit_map": [ + 0.0030182322952896357, + 0.0034450467210263014, + 0.003977807704359293, + 0.0051570977084338665, + 0.0032016427721828222, + 0.003869910491630435, + 0.0026767291128635406, + 0.005789598915725946, + 0.0014366498216986656, + 0.001176798832602799, + 0.0014193784445524216, + 0.0017490704776719213, + 0.0011935612419620156, + 0.0017553187208250165, + 0.0010893400758504868, + 0.0019507616525515914, + 0.0006090861279517412, + 0.000828722957521677, + 0.0007304089376702905, + 0.0010603194823488593, + 0.0012908506905660033, + 0.0010183937847614288, + 0.0009351039188914001, + 0.0011246810900047421, + 0.0023304917849600315, + 0.0020896941423416138, + 0.004626665264368057, + 0.002004216890782118, + 0.003637902671471238, + 0.0046631586737930775, + 0.0018977271392941475, + 0.002928293775767088, + 0.0019829608500003815, + 0.0019010442774742842, + 0.0023923299740999937, + 0.0012051622616127133, + 0.0011503187706694007, + 0.002765276236459613, + 0.0017883209511637688, + 0.0010826631914824247, + 0.010989755392074585, + 0.005804971791803837, + 0.0046157329343259335, + 0.0052125463262200356, + 0.007404448464512825, + 0.005853019189089537, + 0.005081694107502699, + 0.004820966627448797, + 0.0023458143696188927, + 0.0037123386282473803, + 0.0011387759586796165, + 0.0015612517017871141, + 0.002306760987266898, + 0.0016592353349551558, + 0.001441103988327086, + 0.002601007232442498, + 0.0015038667479529977, + 0.001212662784382701, + 0.001609714119695127, + 0.0021041196305304766, + 0.0020658927969634533, + 0.00141858437564224, + 0.001221981248818338, + 0.001422422705218196, + 0.005851939786225557, + 0.006947936490178108, + 0.016582468524575233, + 0.009470679797232151, + 0.013057123869657516, + 0.011642638593912125, + 0.00473646679893136, + 0.0073186238296329975, + 0.051771827042102814, + 0.2478249967098236, + 0.039215244352817535, + 0.2738089859485626, + 0.03751891106367111, + 0.03263732045888901, + 0.03607238456606865, + 0.03748302906751633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.0013375869020819664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Embracing the meticulously hand-painted aesthetic of Ghibli animation, a striking overhead fisheye lens places the viewer on the ground, facing a vibrant, multicolored, segmented giant centipede from Madagascar. It nestles in a humid, vibrant, fantasy forest floor, nestled against a backdrop of giant, mossy ferns and shimmering, exotic mushrooms. The exaggerated distortion of the fisheye lens causes these plants to curve and stretch toward the edge of the frame, forming an enveloping, dizzying dome of green. The centipede's carapace displays an intricate pattern of deep russet and bright orange. Each smooth segment reflects dappled light filtering through the gaps in the forest canopy, and countless pairs of yellow legs connect the segments, each in a frenzy of activity. First, its long, segmented body arches back sharply into a tense \"C\" shape, its massive, venom-gleaming black mandibules poised at the front. Then, in a flash, it launches its entire body forward with explosive speed, hurtling towards the camera. During this rapid movement, its head and maxillipedia are disproportionately magnified in front of the camera, as if they are about to break through the screen. Countless pairs of legs are frantically paddling in the air, forming a blurred afterimage. The entire picture is full of dynamic tension and a suffocating sense of oppression from the huge creature, perfectly reproducing the classic depiction of the awe and fear of natural forces in Ghibli animation.", + "seed": 5085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_4/index_2980_seed_5085_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019196216273121534, + "vlm_logprob": -0.019196216273121534, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.019196216273121534, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03067108429968357 + ], + "opposite_logprob_tokens": [ + -4.030671119689941 + ], + "logprob_reward": -0.03067108429968357, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03067108429968357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022775933146476746, + 0.1064450666308403, + 0.00884532742202282, + 0.6721767783164978, + 0.004885457921773195, + 0.0032541232649236917, + 0.008915736339986324, + 0.03565709665417671, + 0.12447304278612137, + 0.012571372091770172 + ], + "credit_map": [ + 0.002064279979094863, + 0.0033114508260041475, + 0.0023454700130969286, + 0.0018858159892261028, + 0.002829229459166527, + 0.0021245379466563463, + 0.003082041861489415, + 0.005133108235895634, + 0.005331294145435095, + 0.03847089037299156, + 0.029724277555942535, + 0.003745516762137413, + 0.012034677900373936, + 0.007067653816193342, + 0.005862777587026358, + 0.004207978490740061, + 0.0009530974202789366, + 0.0015945152845233679, + 0.0012920646695420146, + 0.00100643839687109, + 0.0011138728586956859, + 0.0011472046608105302, + 0.0009068406070582569, + 0.000831293931696564, + 0.03587779402732849, + 0.24665966629981995, + 0.20281526446342468, + 0.02958461456000805, + 0.08230802416801453, + 0.029388166964054108, + 0.018191438168287277, + 0.02735181339085102, + 0.0006802313728258014, + 0.0006534337881021202, + 0.0007275792304426432, + 0.0005518334219232202, + 0.0005059214308857918, + 0.0007120273075997829, + 0.0006114665302447975, + 0.00044296489795669913, + 0.0002648947120178491, + 0.0006170468404889107, + 0.00032492546597495675, + 0.0005488934111781418, + 0.00037683287519030273, + 0.0003481825115159154, + 0.0003571920678950846, + 0.00041615532245486975, + 0.0006393095245584846, + 0.0013102389639243484, + 0.0008822345989756286, + 0.0028011768590658903, + 0.0008129518246278167, + 0.0009262983221560717, + 0.0008283024653792381, + 0.0007152238977141678, + 0.0022976200561970472, + 0.0033421404659748077, + 0.01215414423495531, + 0.001259406446479261, + 0.0017037905054166913, + 0.010312964208424091, + 0.0031463338527828455, + 0.0014406967675313354, + 0.0048180739395320415, + 0.009357023984193802, + 0.002027750015258789, + 0.0020804747473448515, + 0.053323253989219666, + 0.006861384026706219, + 0.040995411574840546, + 0.00500967912375927, + 0.0013775891857221723, + 0.0019368380308151245, + 0.0008880154346115887, + 0.001480124075897038, + 0.002358413301408291, + 0.001467924565076828, + 0.0016948975389823318, + 0.0013675704831257463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.03067108429968357 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545360252261162 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545360252261162, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545360252261162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04299938306212425, + 0.2197301983833313, + 0.01981673762202263, + 0.49447399377822876, + 0.010447350330650806, + 0.004201204515993595, + 0.013648728840053082, + 0.06181452050805092, + 0.11099472641944885, + 0.021873122081160545 + ], + "credit_map": [ + 0.002492620376870036, + 0.01107142772525549, + 0.00400895718485117, + 0.0037216937635093927, + 0.00627026055008173, + 0.006927071139216423, + 0.003942530602216721, + 0.004564828239381313, + 0.010044824331998825, + 0.08352914452552795, + 0.05859676003456116, + 0.00798638816922903, + 0.028200309723615646, + 0.011685425415635109, + 0.012135922908782959, + 0.007551451213657856, + 0.0015216447645798326, + 0.002820656867697835, + 0.0034611786250025034, + 0.003082923125475645, + 0.002633251715451479, + 0.002040426479652524, + 0.002573343925178051, + 0.0016833156114444137, + 0.026578141376376152, + 0.18017636239528656, + 0.15534746646881104, + 0.022089768201112747, + 0.0563335083425045, + 0.0218781940639019, + 0.012095988728106022, + 0.01997455768287182, + 0.0009234508615918458, + 0.0015497872373089194, + 0.0011194997932761908, + 0.001286738202907145, + 0.0019996813498437405, + 0.0012006292818114161, + 0.0012510925298556685, + 0.0011164721800014377, + 0.0003286225546617061, + 0.0008380793733522296, + 0.0003390515339560807, + 0.000673448434099555, + 0.00046861235750839114, + 0.0004032702709082514, + 0.0006776362424716353, + 0.0004724841273855418, + 0.0009054086985997856, + 0.0022938470356166363, + 0.001478889724239707, + 0.0031590431462973356, + 0.0015446769539266825, + 0.0019589681178331375, + 0.0012038894928991795, + 0.001104007358662784, + 0.003468257375061512, + 0.006027230992913246, + 0.019543804228305817, + 0.002137246076017618, + 0.0035413713194429874, + 0.019012194126844406, + 0.005465501453727484, + 0.002618928672745824, + 0.004469676408916712, + 0.008181808516383171, + 0.002121339552104473, + 0.00204627332277596, + 0.046914003789424896, + 0.006064038258045912, + 0.03662673383951187, + 0.004570866469293833, + 0.002259464468806982, + 0.0035462474916130304, + 0.0015766751021146774, + 0.003344497177749872, + 0.0031336869578808546, + 0.002215603133663535, + 0.003227123524993658, + 0.0025698253884911537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.012545360252261162 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019418209558352828 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019418209558352828, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019418209558352828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11498577892780304, + 0.11855682730674744, + 0.03775128722190857, + 0.13078105449676514, + 0.02122066356241703, + 0.023563258349895477, + 0.06259658932685852, + 0.18998396396636963, + 0.2123996466398239, + 0.08816094696521759 + ], + "credit_map": [ + 0.011217637918889523, + 0.009651141241192818, + 0.015446027740836143, + 0.008998801931738853, + 0.016720108687877655, + 0.02471982128918171, + 0.012010172940790653, + 0.016222061589360237, + 0.0070658354088664055, + 0.04099505394697189, + 0.02433379925787449, + 0.006131331901997328, + 0.017836786806583405, + 0.008247194811701775, + 0.007820384576916695, + 0.006126427557319403, + 0.004025617148727179, + 0.004255292471498251, + 0.0042895046062767506, + 0.002825383795425296, + 0.0057763829827308655, + 0.008541013114154339, + 0.003613668028265238, + 0.0044244201853871346, + 0.008160099387168884, + 0.029541796073317528, + 0.04521148279309273, + 0.008506138809025288, + 0.012510541826486588, + 0.010260732844471931, + 0.009042651392519474, + 0.007547606248408556, + 0.00182898691855371, + 0.0032171260099858046, + 0.0023199657443910837, + 0.0023391598369926214, + 0.003221857128664851, + 0.003821702441200614, + 0.0019795545376837254, + 0.0024923088494688272, + 0.002000539330765605, + 0.003855932503938675, + 0.0024352059699594975, + 0.002431410364806652, + 0.0028687736485153437, + 0.0029076705686748028, + 0.0031831562519073486, + 0.003880569012835622, + 0.003929671365767717, + 0.008749156259000301, + 0.0050209215842187405, + 0.016499437391757965, + 0.009452905505895615, + 0.007377295289188623, + 0.006632986944168806, + 0.004934208933264017, + 0.011547829024493694, + 0.019050272181630135, + 0.0634797066450119, + 0.006993592716753483, + 0.009381347335875034, + 0.05431440845131874, + 0.016790062189102173, + 0.00842673797160387, + 0.009777432307600975, + 0.015847504138946533, + 0.004784197546541691, + 0.0050117545761168, + 0.08567822724580765, + 0.01294679008424282, + 0.06855034828186035, + 0.009803383611142635, + 0.007004166021943092, + 0.013673611916601658, + 0.005396883003413677, + 0.01650105230510235, + 0.013240410014986992, + 0.008944878354668617, + 0.012850111350417137, + 0.010549835860729218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.0019418209558352828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862593859434128 + ], + "opposite_logprob_tokens": [ + -3.048625946044922 + ], + "logprob_reward": -0.04862593859434128, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862593859434128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027258532121777534, + 0.025344539433717728, + 0.006544928532093763, + 0.8156375288963318, + 0.024640215560793877, + 0.007070624735206366, + 0.02301851660013199, + 0.0127538051456213, + 0.023381877690553665, + 0.034349460154771805 + ], + "credit_map": [ + 0.003121140645816922, + 0.0027766525745391846, + 0.0025119297206401825, + 0.003760712221264839, + 0.004398503340780735, + 0.003131822682917118, + 0.0029432792216539383, + 0.00461448822170496, + 0.0016436493024230003, + 0.007958192378282547, + 0.00435445224866271, + 0.0023403149098157883, + 0.0034379144199192524, + 0.0017920994432643056, + 0.002116319490596652, + 0.0017015963094308972, + 0.0006999374600127339, + 0.0013072615256533027, + 0.0013192816404625773, + 0.0005322445067577064, + 0.0009592618444003165, + 0.0006400113343261182, + 0.0005181245505809784, + 0.0005688052042387426, + 0.04266808554530144, + 0.3008539378643036, + 0.2411256730556488, + 0.03466583043336868, + 0.10507041215896606, + 0.03455168008804321, + 0.02372400090098381, + 0.03297785297036171, + 0.002551653189584613, + 0.004467394668608904, + 0.002943106461316347, + 0.0036545605398714542, + 0.00395119609311223, + 0.0021400591358542442, + 0.0025363483000546694, + 0.00239589624106884, + 0.0007049051928333938, + 0.0012043551541864872, + 0.0006082481704652309, + 0.0011043880367651582, + 0.0009040864533744752, + 0.0006300631212070584, + 0.0011384624522179365, + 0.0007761158631183207, + 0.002177405869588256, + 0.002420298755168915, + 0.002412969246506691, + 0.0024314785841852427, + 0.0027242349460721016, + 0.0037119558546692133, + 0.0030052419751882553, + 0.00413493113592267, + 0.0012804910074919462, + 0.0015670537250116467, + 0.0009519842569716275, + 0.0007045611855573952, + 0.00148602609988302, + 0.003564165672287345, + 0.002165119629353285, + 0.0010344035690650344, + 0.0018230489222332835, + 0.002161551034078002, + 0.0011835653567686677, + 0.0011395281180739403, + 0.008076237514615059, + 0.0020978024695068598, + 0.005434378050267696, + 0.001465764013119042, + 0.003428003052249551, + 0.004694153089076281, + 0.0023621884174644947, + 0.006411819718778133, + 0.005078496411442757, + 0.0034012182150036097, + 0.004866111092269421, + 0.004107465967535973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.04862593859434128 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021968772634863853 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021968772634863853, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021968772634863853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040750499814748764, + 0.022274304181337357, + 0.016609512269496918, + 0.5106484889984131, + 0.012203697115182877, + 0.009307284839451313, + 0.047503821551799774, + 0.02389450930058956, + 0.26453810930252075, + 0.05226990953087807 + ], + "credit_map": [ + 0.0025688158348202705, + 0.003574554342776537, + 0.00438540056347847, + 0.0035169613547623158, + 0.008031247183680534, + 0.009846084751188755, + 0.00244716159068048, + 0.0063802702352404594, + 0.002582348184660077, + 0.0030444758012890816, + 0.00314993760548532, + 0.002088243141770363, + 0.004615393932908773, + 0.002486891346052289, + 0.0014397434424608946, + 0.0028672676999121904, + 0.001505803200416267, + 0.0016829243395477533, + 0.0023086483124643564, + 0.0021156524308025837, + 0.0026567287277430296, + 0.0028465951327234507, + 0.0017318001482635736, + 0.0017613588133826852, + 0.027039723470807076, + 0.1882045418024063, + 0.1524009257555008, + 0.021213820204138756, + 0.06544295698404312, + 0.021566618233919144, + 0.014117147773504257, + 0.020662697032094002, + 0.0011472963960841298, + 0.0014996824320405722, + 0.00144639250356704, + 0.001820847624912858, + 0.0018870661733672023, + 0.0013087025145068765, + 0.0017312202835455537, + 0.001362488605082035, + 0.0009312361944466829, + 0.0017189986538141966, + 0.0008248585509136319, + 0.0011606429470703006, + 0.0011068495223298669, + 0.0011957189999520779, + 0.0011251219548285007, + 0.001243857084773481, + 0.003447800874710083, + 0.007732007186859846, + 0.005652260966598988, + 0.008839390240609646, + 0.0037448140792548656, + 0.008826853707432747, + 0.0044109392911195755, + 0.004849751479923725, + 0.002109294291585684, + 0.003153641941025853, + 0.003234581323340535, + 0.0016112066805362701, + 0.0047312271781265736, + 0.0031984702218323946, + 0.0031693698838353157, + 0.0026867142878472805, + 0.009329095482826233, + 0.020618071779608727, + 0.004900740925222635, + 0.004395763855427504, + 0.11415857076644897, + 0.014603612944483757, + 0.08583584427833557, + 0.010696358047425747, + 0.005739783402532339, + 0.008320904336869717, + 0.004388229455798864, + 0.005082902032881975, + 0.008833205327391624, + 0.0070876204408705235, + 0.005949437152594328, + 0.006867821328341961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.0021968772634863853 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Embracing the meticulously hand-painted aesthetic of Ghibli animation, a striking overhead fisheye lens places the viewer on the ground, facing a vibrant, multicolored, segmented giant centipede from Madagascar. It nestles in a humid, vibrant, fantasy forest floor, nestled against a backdrop of giant, mossy ferns and shimmering, exotic mushrooms. The exaggerated distortion of the fisheye lens causes these plants to curve and stretch toward the edge of the frame, forming an enveloping, dizzying dome of green. The centipede's carapace displays an intricate pattern of deep russet and bright orange. Each smooth segment reflects dappled light filtering through the gaps in the forest canopy, and countless pairs of yellow legs connect the segments, each in a frenzy of activity. First, its long, segmented body arches back sharply into a tense \"C\" shape, its massive, venom-gleaming black mandibules poised at the front. Then, in a flash, it launches its entire body forward with explosive speed, hurtling towards the camera. During this rapid movement, its head and maxillipedia are disproportionately magnified in front of the camera, as if they are about to break through the screen. Countless pairs of legs are frantically paddling in the air, forming a blurred afterimage. The entire picture is full of dynamic tension and a suffocating sense of oppression from the huge creature, perfectly reproducing the classic depiction of the awe and fear of natural forces in Ghibli animation.", + "seed": 5086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_4/index_2980_seed_5086_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031145892478525637, + "vlm_logprob": -0.031145892478525637, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.031145892478525637, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09276911616325378 + ], + "opposite_logprob_tokens": [ + -2.592769145965576 + ], + "logprob_reward": -0.09276911616325378, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09276911616325378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10979479551315308, + 0.02112811617553234, + 0.03716207668185234, + 0.14106078445911407, + 0.18826565146446228, + 0.22211793065071106, + 0.06282474100589752, + 0.01755594089627266, + 0.04913758859038353, + 0.15095242857933044 + ], + "credit_map": [ + 0.01185647677630186, + 0.014628912322223186, + 0.014417407102882862, + 0.012034201994538307, + 0.011755176819860935, + 0.011132791638374329, + 0.01615818589925766, + 0.01781163178384304, + 0.0026689167134463787, + 0.0030972796957939863, + 0.0026181968860328197, + 0.0021022488363087177, + 0.0025195004418492317, + 0.002915713470429182, + 0.0022939657792448997, + 0.0029122952837496996, + 0.004921835847198963, + 0.004278880078345537, + 0.005121688824146986, + 0.004171021282672882, + 0.004254787694662809, + 0.0063235219568014145, + 0.003750098869204521, + 0.004340238869190216, + 0.00949892308562994, + 0.025071924552321434, + 0.01662910357117653, + 0.015263800509274006, + 0.00972780678421259, + 0.013570631854236126, + 0.009856957010924816, + 0.04144163802266121, + 0.020361941307783127, + 0.025619998574256897, + 0.024565940722823143, + 0.03265363350510597, + 0.02132338657975197, + 0.01730310171842575, + 0.02404163032770157, + 0.02239600382745266, + 0.036999888718128204, + 0.017841078341007233, + 0.06375942379236221, + 0.01980854943394661, + 0.015081891790032387, + 0.018813081085681915, + 0.01850976049900055, + 0.031304240226745605, + 0.0061015901155769825, + 0.007085362449288368, + 0.0067952899262309074, + 0.008861935697495937, + 0.007579157128930092, + 0.008164700120687485, + 0.012134830467402935, + 0.006101870909333229, + 0.00241849129088223, + 0.0015035693068057299, + 0.0021892439108341932, + 0.0021600238978862762, + 0.0014963960275053978, + 0.002393237082287669, + 0.0022112170699983835, + 0.003183761378750205, + 0.0029922707471996546, + 0.004525279626250267, + 0.004414428491145372, + 0.004655472002923489, + 0.005371433682739735, + 0.006767180748283863, + 0.010975711978971958, + 0.009435810148715973, + 0.01191822811961174, + 0.017174480482935905, + 0.014527210965752602, + 0.013965096324682236, + 0.018707692623138428, + 0.03514283522963524, + 0.015000080689787865, + 0.024516794830560684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.09276911616325378 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864844024181366 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.00864844024181366, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864844024181366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06534481793642044, + 0.048300404101610184, + 0.05653712525963783, + 0.19114908576011658, + 0.1638956218957901, + 0.1795129030942917, + 0.0867801308631897, + 0.03463797643780708, + 0.07923620194196701, + 0.09460575878620148 + ], + "credit_map": [ + 0.006251325365155935, + 0.011428008787333965, + 0.007282746024429798, + 0.006495993584394455, + 0.008867121301591396, + 0.008136165328323841, + 0.007745806593447924, + 0.009137652814388275, + 0.0050876559689641, + 0.007880485616624355, + 0.006446664687246084, + 0.0048874299973249435, + 0.006647840142250061, + 0.0056409696117043495, + 0.00624168710783124, + 0.005467669572681189, + 0.004967851564288139, + 0.006400929763913155, + 0.007578152697533369, + 0.008591307327151299, + 0.005699094384908676, + 0.007186685223132372, + 0.0078649390488863, + 0.008248161524534225, + 0.015628241002559662, + 0.016056867316365242, + 0.019336797297000885, + 0.023267380893230438, + 0.015058102086186409, + 0.04244058206677437, + 0.01735503226518631, + 0.042006079107522964, + 0.017626266926527023, + 0.022393086925148964, + 0.019664006307721138, + 0.028137382119894028, + 0.018354516476392746, + 0.01474501471966505, + 0.022516712546348572, + 0.02045864798128605, + 0.02257411740720272, + 0.01158740371465683, + 0.04598654806613922, + 0.014070817269384861, + 0.010131248272955418, + 0.03526849299669266, + 0.01302715577185154, + 0.02686711587011814, + 0.009341600351035595, + 0.009116817265748978, + 0.009069742634892464, + 0.01402404997497797, + 0.006658035796135664, + 0.013392167165875435, + 0.017179353162646294, + 0.00799836590886116, + 0.003556119743734598, + 0.004247468430548906, + 0.003521042177453637, + 0.003842558478936553, + 0.004867000505328178, + 0.005153990816324949, + 0.005803714040666819, + 0.0036460792180150747, + 0.005689824000000954, + 0.0065919565968215466, + 0.008903506211936474, + 0.007408556062728167, + 0.0061578392051160336, + 0.009923113510012627, + 0.01622910052537918, + 0.018332311883568764, + 0.006592875812202692, + 0.008375391364097595, + 0.008701375685632229, + 0.008185353130102158, + 0.012188286520540714, + 0.019983310252428055, + 0.011953037232160568, + 0.018626123666763306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.00864844024181366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031943507492542267 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.0031943507492542267, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031943507492542267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10322855412960052, + 0.05790666863322258, + 0.03301825746893883, + 0.04950717091560364, + 0.32112830877304077, + 0.07449443638324738, + 0.16879092156887054, + 0.06453558057546616, + 0.05038389191031456, + 0.07700616866350174 + ], + "credit_map": [ + 0.011338462121784687, + 0.015577636659145355, + 0.008725852705538273, + 0.010208393447101116, + 0.013922818936407566, + 0.021728508174419403, + 0.009898853488266468, + 0.011828036978840828, + 0.007340859156101942, + 0.009316532872617245, + 0.005039753392338753, + 0.004474000073969364, + 0.0076624490320682526, + 0.01535178441554308, + 0.004196929279714823, + 0.004524364601820707, + 0.004034083802253008, + 0.007089660502970219, + 0.0037710487376898527, + 0.003429178614169359, + 0.0047067576088011265, + 0.002896364312618971, + 0.0036779604852199554, + 0.003413204336538911, + 0.004605340305715799, + 0.008905673399567604, + 0.006790523882955313, + 0.007162060122936964, + 0.004825529642403126, + 0.007612941320985556, + 0.004432246088981628, + 0.005172856617718935, + 0.03456145524978638, + 0.04359040781855583, + 0.04979144036769867, + 0.0563432052731514, + 0.03459443524479866, + 0.02789881080389023, + 0.03978811204433441, + 0.034560445696115494, + 0.010508595034480095, + 0.00559232709929347, + 0.01574578322470188, + 0.006061329506337643, + 0.005243147257715464, + 0.01257055252790451, + 0.006318456027656794, + 0.012454250827431679, + 0.019403597339987755, + 0.020710717886686325, + 0.016824163496494293, + 0.02412896603345871, + 0.01533739548176527, + 0.026187684386968613, + 0.0307843629270792, + 0.015414060093462467, + 0.006692866329103708, + 0.006884112022817135, + 0.006844710558652878, + 0.007183598820120096, + 0.009061711840331554, + 0.009735312312841415, + 0.01131577417254448, + 0.006817500572651625, + 0.00350463530048728, + 0.0061630685813724995, + 0.0051861172541975975, + 0.004368798807263374, + 0.005523188970983028, + 0.005902508273720741, + 0.009386519901454449, + 0.010349060408771038, + 0.006846974138170481, + 0.012656969018280506, + 0.00842593889683485, + 0.009091612882912159, + 0.010441062971949577, + 0.010805516503751278, + 0.009146780706942081, + 0.00959132518619299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.0031943507492542267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862775281071663 + ], + "opposite_logprob_tokens": [ + -3.0486278533935547 + ], + "logprob_reward": -0.04862775281071663, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862775281071663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1811572164297104, + 0.11132485419511795, + 0.02736249752342701, + 0.09117947518825531, + 0.08947374671697617, + 0.1400952935218811, + 0.08982296288013458, + 0.11391071230173111, + 0.07075407356023788, + 0.08491913974285126 + ], + "credit_map": [ + 0.017803465947508812, + 0.0247388556599617, + 0.021288232877850533, + 0.02173502743244171, + 0.022673025727272034, + 0.015737364068627357, + 0.027289440855383873, + 0.029891811311244965, + 0.009037909097969532, + 0.012454615905880928, + 0.013710321858525276, + 0.016091974452137947, + 0.011928223073482513, + 0.011378979310393333, + 0.020700087770819664, + 0.016022752970457077, + 0.0028831567615270615, + 0.0032997983507812023, + 0.004045443143695593, + 0.003946879878640175, + 0.00261680013500154, + 0.0030429218895733356, + 0.0038864496164023876, + 0.0036410472821444273, + 0.009805404581129551, + 0.009261331520974636, + 0.014306757599115372, + 0.0122850202023983, + 0.010041064582765102, + 0.00930564571171999, + 0.015188303776085377, + 0.010985943488776684, + 0.009291820228099823, + 0.010756752453744411, + 0.010552693158388138, + 0.01570417732000351, + 0.011338525451719761, + 0.007270559668540955, + 0.014200887642800808, + 0.010358331725001335, + 0.01373150572180748, + 0.009053604677319527, + 0.043172676116228104, + 0.01019904762506485, + 0.008402473293244839, + 0.007365310564637184, + 0.008002826943993568, + 0.04016784206032753, + 0.008614735677838326, + 0.009049400687217712, + 0.009601683355867863, + 0.015512281097471714, + 0.00786000769585371, + 0.011100494302809238, + 0.019414007663726807, + 0.00867035798728466, + 0.01479327492415905, + 0.009553344920277596, + 0.010999827645719051, + 0.011961097829043865, + 0.013847325928509235, + 0.012609927915036678, + 0.025413010269403458, + 0.014732914976775646, + 0.0054691811092197895, + 0.005701587535440922, + 0.009845860302448273, + 0.00853271596133709, + 0.006326376926153898, + 0.011939355172216892, + 0.011431745253503323, + 0.01150725781917572, + 0.008113541640341282, + 0.00742469634860754, + 0.012752854265272617, + 0.010722207836806774, + 0.010060411877930164, + 0.011818476021289825, + 0.011708906851708889, + 0.012318040244281292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.04862775281071663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024898024275898933 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024898024275898933, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024898024275898933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13741886615753174, + 0.06930750608444214, + 0.04471509903669357, + 0.05435528606176376, + 0.21723148226737976, + 0.13489997386932373, + 0.07794555276632309, + 0.08263099938631058, + 0.0719742700457573, + 0.10952088981866837 + ], + "credit_map": [ + 0.012708554044365883, + 0.018346106633543968, + 0.015488257631659508, + 0.01721835881471634, + 0.016635742038488388, + 0.011967484839260578, + 0.021258078515529633, + 0.023796284571290016, + 0.005881070625036955, + 0.00832457561045885, + 0.008672450669109821, + 0.010265536606311798, + 0.007227340713143349, + 0.0070213149301707745, + 0.011549749411642551, + 0.010365470312535763, + 0.004155379720032215, + 0.004830938298255205, + 0.005758573766797781, + 0.006770424544811249, + 0.004428883548825979, + 0.005139708984643221, + 0.007024690508842468, + 0.006606501527130604, + 0.0047110929153859615, + 0.005778249818831682, + 0.006278111133724451, + 0.005888639949262142, + 0.0034694198984652758, + 0.010953287594020367, + 0.0038786232471466064, + 0.013397864066064358, + 0.024520857259631157, + 0.028274718672037125, + 0.033310890197753906, + 0.03823195770382881, + 0.02243703603744507, + 0.018914228305220604, + 0.0263915304094553, + 0.025150267407298088, + 0.013709546998143196, + 0.012741109356284142, + 0.0173520278185606, + 0.015121092088520527, + 0.010777730494737625, + 0.031630903482437134, + 0.014584969729185104, + 0.018982592970132828, + 0.006547277793288231, + 0.006290613207966089, + 0.009767170064151287, + 0.014852318912744522, + 0.00724318390712142, + 0.011961717158555984, + 0.015718501061201096, + 0.005564769264310598, + 0.00846844632178545, + 0.010044524446129799, + 0.008175627328455448, + 0.009875585325062275, + 0.01129878219217062, + 0.009732205420732498, + 0.011087191291153431, + 0.013948645442724228, + 0.004773654509335756, + 0.007358043920248747, + 0.0071976035833358765, + 0.007362902630120516, + 0.006983635481446981, + 0.011734656058251858, + 0.014889896847307682, + 0.011673877015709877, + 0.010944005101919174, + 0.01427558995783329, + 0.018913831561803818, + 0.014315742999315262, + 0.011149120517075062, + 0.011986112222075462, + 0.013286546804010868, + 0.014649948105216026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0024898024275898933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Embracing the meticulously hand-painted aesthetic of Ghibli animation, a striking overhead fisheye lens places the viewer on the ground, facing a vibrant, multicolored, segmented giant centipede from Madagascar. It nestles in a humid, vibrant, fantasy forest floor, nestled against a backdrop of giant, mossy ferns and shimmering, exotic mushrooms. The exaggerated distortion of the fisheye lens causes these plants to curve and stretch toward the edge of the frame, forming an enveloping, dizzying dome of green. The centipede's carapace displays an intricate pattern of deep russet and bright orange. Each smooth segment reflects dappled light filtering through the gaps in the forest canopy, and countless pairs of yellow legs connect the segments, each in a frenzy of activity. First, its long, segmented body arches back sharply into a tense \"C\" shape, its massive, venom-gleaming black mandibules poised at the front. Then, in a flash, it launches its entire body forward with explosive speed, hurtling towards the camera. During this rapid movement, its head and maxillipedia are disproportionately magnified in front of the camera, as if they are about to break through the screen. Countless pairs of legs are frantically paddling in the air, forming a blurred afterimage. The entire picture is full of dynamic tension and a suffocating sense of oppression from the huge creature, perfectly reproducing the classic depiction of the awe and fear of natural forces in Ghibli animation.", + "seed": 5087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_4/index_2980_seed_5087_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05043019945733249, + "vlm_logprob": -0.05043019945733249, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.05043019945733249, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21822214126586914 + ], + "opposite_logprob_tokens": [ + -1.7182221412658691 + ], + "logprob_reward": -0.21822214126586914, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21822214126586914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17065395414829254, + 0.06321282684803009, + 0.03773493319749832, + 0.02802594006061554, + 0.035008158534765244, + 0.03851709142327309, + 0.21977151930332184, + 0.09182105958461761, + 0.13756166398525238, + 0.17769289016723633 + ], + "credit_map": [ + 0.018463606014847755, + 0.019373420625925064, + 0.0204608291387558, + 0.025340190157294273, + 0.02188810333609581, + 0.018640298396348953, + 0.020523596554994583, + 0.025963913649320602, + 0.007943636737763882, + 0.007462528068572283, + 0.00571751082316041, + 0.005825118161737919, + 0.014781271107494831, + 0.008128922432661057, + 0.007480420637875795, + 0.005873422138392925, + 0.00485713267698884, + 0.00638442812487483, + 0.005723708774894476, + 0.003938932437449694, + 0.005054705310612917, + 0.004668478854000568, + 0.0029289189260452986, + 0.004178629722446203, + 0.0027119116857647896, + 0.003193836659193039, + 0.0037968435790389776, + 0.004060729872435331, + 0.0037343057338148355, + 0.0035366625525057316, + 0.002837300067767501, + 0.004154350142925978, + 0.0031851313542574644, + 0.0029265827033668756, + 0.003664652816951275, + 0.006834174040704966, + 0.003759512212127447, + 0.004144614562392235, + 0.006145613268017769, + 0.0043478780426084995, + 0.0035306732170283794, + 0.0033686619717627764, + 0.005184763111174107, + 0.007529516238719225, + 0.004790369421243668, + 0.00507451081648469, + 0.0050051030702888966, + 0.0040334961377084255, + 0.03748636692762375, + 0.016240142285823822, + 0.027715807780623436, + 0.017613854259252548, + 0.03352941572666168, + 0.04853426665067673, + 0.020215939730405807, + 0.01843574270606041, + 0.00929785706102848, + 0.013236056081950665, + 0.01311726588755846, + 0.010619268752634525, + 0.012969430536031723, + 0.011482148431241512, + 0.012859189882874489, + 0.00823984108865261, + 0.010640396736562252, + 0.03636464104056358, + 0.012230473570525646, + 0.014400927349925041, + 0.017953108996152878, + 0.011842051520943642, + 0.020392615348100662, + 0.013737461529672146, + 0.03662826865911484, + 0.02744722180068493, + 0.018896430730819702, + 0.01548969279974699, + 0.024493280798196793, + 0.023126322776079178, + 0.014332973398268223, + 0.01727868989109993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.21822214126586914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089478619396687 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.011089478619396687, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089478619396687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12625478208065033, + 0.26570644974708557, + 0.0906451940536499, + 0.06487440317869186, + 0.044009093195199966, + 0.0508895106613636, + 0.032145097851753235, + 0.08568808436393738, + 0.09294699132442474, + 0.1468404233455658 + ], + "credit_map": [ + 0.01347809936851263, + 0.010786147788167, + 0.016349993646144867, + 0.02341718226671219, + 0.016695043072104454, + 0.01446131058037281, + 0.01592782512307167, + 0.015139169991016388, + 0.028144223615527153, + 0.02337009832262993, + 0.04213017225265503, + 0.03916303440928459, + 0.055584896355867386, + 0.033954210579395294, + 0.025029277428984642, + 0.018330510705709457, + 0.008747415617108345, + 0.010408644564449787, + 0.011791066266596317, + 0.015842845663428307, + 0.015680262818932533, + 0.010000920854508877, + 0.009048030711710453, + 0.00912600476294756, + 0.006535724736750126, + 0.005291127134114504, + 0.0077690184116363525, + 0.008851515129208565, + 0.008500207215547562, + 0.0071043712086975574, + 0.006717392709106207, + 0.014105040580034256, + 0.0034947379026561975, + 0.0034098660107702017, + 0.005059004295617342, + 0.011747227050364017, + 0.0044664680026471615, + 0.004780159797519445, + 0.005342193879187107, + 0.005709430668503046, + 0.005454009398818016, + 0.004963201936334372, + 0.007128894794732332, + 0.007830098271369934, + 0.006368866190314293, + 0.006635688245296478, + 0.005982979200780392, + 0.006525769829750061, + 0.0038070594891905785, + 0.002623955952003598, + 0.005012485198676586, + 0.00366022321395576, + 0.0036645822692662477, + 0.004916402976959944, + 0.004474966321140528, + 0.003985419403761625, + 0.008399262093007565, + 0.011710833758115768, + 0.009750576689839363, + 0.008427996188402176, + 0.011739562265574932, + 0.011278623715043068, + 0.015611819922924042, + 0.00876940693706274, + 0.00914262980222702, + 0.02699190005660057, + 0.007177692838013172, + 0.00926020834594965, + 0.017430849373340607, + 0.00621409434825182, + 0.01051307376474142, + 0.006216535344719887, + 0.026063570752739906, + 0.022931726649403572, + 0.01558228861540556, + 0.012085044756531715, + 0.02168559469282627, + 0.021651823073625565, + 0.0076202270574867725, + 0.019220151007175446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.011089478619396687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619549795985222 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619549795985222, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619549795985222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19010473787784576, + 0.16054750978946686, + 0.06788598746061325, + 0.05233273655176163, + 0.0263442974537611, + 0.0408041775226593, + 0.0882822796702385, + 0.09840395301580429, + 0.1316460520029068, + 0.14364825189113617 + ], + "credit_map": [ + 0.02328505739569664, + 0.0199336688965559, + 0.019727619364857674, + 0.025026332587003708, + 0.0363311804831028, + 0.022509988397359848, + 0.01801496185362339, + 0.025275934487581253, + 0.021661898121237755, + 0.017900453880429268, + 0.017645668238401413, + 0.014210665598511696, + 0.04248272255063057, + 0.020938852801918983, + 0.016505053266882896, + 0.00920218974351883, + 0.00855601578950882, + 0.010843697935342789, + 0.00787079893052578, + 0.0067383842542767525, + 0.011548342183232307, + 0.009171445854008198, + 0.006080773193389177, + 0.007076531182974577, + 0.007131971884518862, + 0.006629373412579298, + 0.006118111778050661, + 0.0047437092289328575, + 0.01144335325807333, + 0.006997310556471348, + 0.003729062620550394, + 0.00553984334692359, + 0.0032693862449377775, + 0.0030306370463222265, + 0.0030313432216644287, + 0.00378889637067914, + 0.004147629253566265, + 0.0031218763906508684, + 0.0034875795245170593, + 0.0024669517297297716, + 0.005485808942466974, + 0.004792698193341494, + 0.004981177393347025, + 0.004098153207451105, + 0.007619139738380909, + 0.004994288086891174, + 0.004040122963488102, + 0.0047927857376635075, + 0.016708331182599068, + 0.0078015378676354885, + 0.015272099524736404, + 0.008368130773305893, + 0.01139562577009201, + 0.010671565309166908, + 0.009448623284697533, + 0.008616369217634201, + 0.0076433164067566395, + 0.013119268231093884, + 0.0076347822323441505, + 0.010361584834754467, + 0.015179109759628773, + 0.019171958789229393, + 0.0175616517663002, + 0.0077322861179709435, + 0.012518619187176228, + 0.02848699875175953, + 0.013180457055568695, + 0.010112779214978218, + 0.03184572234749794, + 0.008354518562555313, + 0.01803113892674446, + 0.009115828201174736, + 0.02109915018081665, + 0.020519189536571503, + 0.01531240250915289, + 0.01440711785107851, + 0.01720505766570568, + 0.025272533297538757, + 0.00936911255121231, + 0.020463690161705017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.003619549795985222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018174193799495697 + ], + "opposite_logprob_tokens": [ + -4.018174171447754 + ], + "logprob_reward": -0.018174193799495697, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018174193799495697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05346762388944626, + 0.03764067590236664, + 0.05724433436989784, + 0.036764126271009445, + 0.034351617097854614, + 0.03818716108798981, + 0.06571923196315765, + 0.07149331271648407, + 0.28466516733169556, + 0.3204667866230011 + ], + "credit_map": [ + 0.00792953185737133, + 0.006097607780247927, + 0.007077464368194342, + 0.007609648164361715, + 0.009331116452813148, + 0.006387469358742237, + 0.00445640878751874, + 0.004578379448503256, + 0.003781293984502554, + 0.004294275306165218, + 0.003645501099526882, + 0.006317495368421078, + 0.005781005136668682, + 0.004514357075095177, + 0.004516684915870428, + 0.00479006115347147, + 0.004911487922072411, + 0.006835408043116331, + 0.005050642881542444, + 0.006977208890020847, + 0.011891359463334084, + 0.008118645288050175, + 0.007495494559407234, + 0.005964083597064018, + 0.004181023687124252, + 0.004717853385955095, + 0.004054037854075432, + 0.005904650315642357, + 0.004292583093047142, + 0.004348379094153643, + 0.004005286376923323, + 0.0052603124640882015, + 0.0036244522780179977, + 0.003300498006865382, + 0.003920872695744038, + 0.006838650908321142, + 0.004998310469090939, + 0.004202915821224451, + 0.003542689373716712, + 0.003923225682228804, + 0.005495084915310144, + 0.0033232485875487328, + 0.0055630505084991455, + 0.005775986239314079, + 0.0033766855485737324, + 0.005027663428336382, + 0.0053718555718660355, + 0.0042535895481705666, + 0.012674950063228607, + 0.005931837484240532, + 0.00973596516996622, + 0.006432178895920515, + 0.006758610252290964, + 0.010603945702314377, + 0.008901887573301792, + 0.004679854493588209, + 0.005358194001019001, + 0.007404003757983446, + 0.010404103435575962, + 0.010873922146856785, + 0.006891380529850721, + 0.009521066211163998, + 0.013181338086724281, + 0.00785930547863245, + 0.025440193712711334, + 0.06961829960346222, + 0.018934985622763634, + 0.032422617077827454, + 0.05625806748867035, + 0.020030677318572998, + 0.04173295572400093, + 0.020227359607815742, + 0.056450456380844116, + 0.0471162348985672, + 0.0336996428668499, + 0.03400380536913872, + 0.04180676117539406, + 0.0416991151869297, + 0.025110330432653427, + 0.040580421686172485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.018174193799495697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot with an overhead fisheye lens perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a giant centipede with a carapace patterned in deep russet and bright orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a humid fantasy forest floor featuring giant mossy ferns and shimmering exotic mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede start in a tense 'C' shape before launching its body forward with explosive speed towards the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"During the forward launch, do the centipede's head and legs appear disproportionately magnified and blurred due to rapid movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010456338059157133 + ], + "opposite_logprob_tokens": [ + -6.8760457038879395 + ], + "logprob_reward": -0.0010456338059157133, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010456338059157133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022837771102786064, + 0.020544512197375298, + 0.048121917992830276, + 0.02411598153412342, + 0.04456678405404091, + 0.027374345809221268, + 0.02567887306213379, + 0.07173143327236176, + 0.22439928352832794, + 0.49062907695770264 + ], + "credit_map": [ + 0.003491911105811596, + 0.0027303656097501516, + 0.003211540635675192, + 0.00264324271120131, + 0.0032082353718578815, + 0.00248019490391016, + 0.0023837268818169832, + 0.002688555745407939, + 0.002277859253808856, + 0.001617527799680829, + 0.0037540337070822716, + 0.0034862591419368982, + 0.002460458315908909, + 0.002509837504476309, + 0.0026032866444438696, + 0.001835250062867999, + 0.004689056426286697, + 0.006088009104132652, + 0.003949746489524841, + 0.004891496617347002, + 0.012950251810252666, + 0.005576706025749445, + 0.005698156077414751, + 0.004278496373444796, + 0.0037037015426903963, + 0.0023348242975771427, + 0.002771934727206826, + 0.002781893592327833, + 0.003761579282581806, + 0.0029556502122431993, + 0.0021992488764226437, + 0.0036071480717509985, + 0.004677407909184694, + 0.004292187746614218, + 0.004869729280471802, + 0.007691825274378061, + 0.005949490703642368, + 0.0051209148950874805, + 0.006387764122337103, + 0.0055774603970348835, + 0.002720245625823736, + 0.0020445433910936117, + 0.003311304608359933, + 0.00379516021348536, + 0.004135714843869209, + 0.003138050204142928, + 0.003588516730815172, + 0.0046408106572926044, + 0.0020514887291938066, + 0.003284508129581809, + 0.0035738826263695955, + 0.00411861389875412, + 0.002775550354272127, + 0.0033548439387232065, + 0.0037636097986251116, + 0.0027563755866140127, + 0.005738027393817902, + 0.007933540269732475, + 0.005468348041176796, + 0.005853413604199886, + 0.011136326938867569, + 0.015164399519562721, + 0.015669509768486023, + 0.0047678700648248196, + 0.018844060599803925, + 0.05474410206079483, + 0.016626298427581787, + 0.024619823321700096, + 0.0537126362323761, + 0.012210485525429249, + 0.029828432947397232, + 0.013813450932502747, + 0.10382772237062454, + 0.06811194121837616, + 0.05257641524076462, + 0.04651905596256256, + 0.06059003621339798, + 0.06654872000217438, + 0.04016383737325668, + 0.05229134485125542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0010456338059157133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_5/index_tensor([4347, 4347])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_5/index_tensor([4347, 4347])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..11de4fb128f20ad82238ffc5fa0dbfd978177c8b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_5/index_tensor([4347, 4347])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from an extremely low angle, the camera held still, its focus fixed squarely on the hem of a deep emerald green silk cheongsam. Crafted from the finest mulberry silk, the cheongsam takes on a fluid, pearly sheen in the soft afternoon light filtering through the carved window lattices. Embroidered with a magnificent pattern of phoenix tail feathers, delicately embroidered in golden thread, each feather is rendered vividly. A gentle breeze caresses the flowing skirt, gracefully billowing. High side slits allow the silky fabric to rhythmically undulate like waves, sometimes brushing gently against the polished dark wood flooring below, sometimes rising slightly into the air to reveal a hint of the equally textured lining. Light and shadow play dappled across the dancing skirt, highlighting the delicate texture of the silk and the three-dimensional quality of the embroidery. In the background, a blurred, out-of-focus scene of a classical garden, faintly visible are the shadows of swaying bamboo. The entire scene is imbued with a tranquil, classical, and poetic beauty, creating an elegant atmosphere where time stands still.", + "seed": 28619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_5/index_4347_seed_28619_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17962721586227418, + "vlm_logprob": -0.17962721586227418, + "vlm_yes_no_margin": 2.0, + "vlm_token_credit": -0.17962721586227418, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.44896677136421204 + ], + "opposite_logprob_tokens": [ + -1.0739667415618896 + ], + "logprob_reward": -0.44896677136421204, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.44896677136421204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1362687051296234, + 0.1959722340106964, + 0.2647480368614197, + 0.10180298238992691, + 0.06090671941637993, + 0.08145744353532791, + 0.04938254877924919, + 0.025248030200600624, + 0.060270145535469055, + 0.023943135514855385 + ], + "credit_map": [ + 0.00945920031517744, + 0.014593828469514847, + 0.0184404868632555, + 0.018551647663116455, + 0.024567708373069763, + 0.012393695302307606, + 0.025586295872926712, + 0.012675845064222813, + 0.01955912448465824, + 0.013046436943113804, + 0.023774396628141403, + 0.02091604471206665, + 0.024528006091713905, + 0.054838474839925766, + 0.020981663838028908, + 0.0183281060308218, + 0.020865323022007942, + 0.02299029938876629, + 0.021077213808894157, + 0.021544935181736946, + 0.04649633541703224, + 0.06033959612250328, + 0.02765391767024994, + 0.043780453503131866, + 0.008309279568493366, + 0.017217967659235, + 0.011002786457538605, + 0.01578712835907936, + 0.016696691513061523, + 0.007872791029512882, + 0.012157066725194454, + 0.012759270146489143, + 0.00568778021261096, + 0.01029864139854908, + 0.0074196672067046165, + 0.009667158126831055, + 0.007855584844946861, + 0.006161693949252367, + 0.00709144864231348, + 0.006724752020090818, + 0.006000697612762451, + 0.009246202185750008, + 0.008433674462139606, + 0.008571811951696873, + 0.01342360395938158, + 0.013239001855254173, + 0.008020121604204178, + 0.014522336423397064, + 0.006385475862771273, + 0.006926553789526224, + 0.007957551628351212, + 0.0041502490639686584, + 0.005339653231203556, + 0.007191394921392202, + 0.004435032140463591, + 0.006996642332524061, + 0.002355749486014247, + 0.0029807889368385077, + 0.0024580080062150955, + 0.003021036274731159, + 0.004153268411755562, + 0.0038989277090877295, + 0.003232037415727973, + 0.003148216288536787, + 0.003703654045239091, + 0.0066221109591424465, + 0.008370486088097095, + 0.0106166061013937, + 0.009400319308042526, + 0.003892452921718359, + 0.009029088541865349, + 0.008635438978672028, + 0.0018394582439213991, + 0.002843896159902215, + 0.0025770983193069696, + 0.002994958544149995, + 0.003091607941314578, + 0.0026927487924695015, + 0.005359189584851265, + 0.002544180955737829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.44896677136421204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02985270321369171 + ], + "opposite_logprob_tokens": [ + -3.5298526287078857 + ], + "logprob_reward": -0.02985270321369171, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02985270321369171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1115819439291954, + 0.0869482234120369, + 0.23244059085845947, + 0.12864279747009277, + 0.11333472281694412, + 0.08740591257810593, + 0.08023123443126678, + 0.10168015211820602, + 0.03736678510904312, + 0.0203676987439394 + ], + "credit_map": [ + 0.01043968927115202, + 0.01116003468632698, + 0.016869042068719864, + 0.0132509870454669, + 0.014090705662965775, + 0.019558532163500786, + 0.015314974822103977, + 0.010897979140281677, + 0.006168758030980825, + 0.004413407761603594, + 0.0242522731423378, + 0.006585281807929277, + 0.008545469492673874, + 0.023328620940446854, + 0.005841463804244995, + 0.007812951691448689, + 0.021958284080028534, + 0.0186197217553854, + 0.02181919850409031, + 0.023656535893678665, + 0.04238506779074669, + 0.05218289792537689, + 0.027976462617516518, + 0.02384241297841072, + 0.017566340044140816, + 0.008958726190030575, + 0.02006017230451107, + 0.010138444602489471, + 0.02434057928621769, + 0.02001447044312954, + 0.010253711603581905, + 0.017310352995991707, + 0.013915253803133965, + 0.010432236827909946, + 0.0154577000066638, + 0.007271577604115009, + 0.02398580312728882, + 0.022072823718190193, + 0.005431632045656443, + 0.014767690561711788, + 0.008352559991180897, + 0.00714137451723218, + 0.009669668972492218, + 0.006085193250328302, + 0.010853322222828865, + 0.018392765894532204, + 0.006227060686796904, + 0.02068396471440792, + 0.005180624779313803, + 0.008677356876432896, + 0.01084115169942379, + 0.009386368095874786, + 0.010671555064618587, + 0.014054302126169205, + 0.008368576876819134, + 0.013051299378275871, + 0.007412334438413382, + 0.020007198676466942, + 0.005901985336095095, + 0.010581334121525288, + 0.024257294833660126, + 0.008731742389500141, + 0.013745839707553387, + 0.011042419821023941, + 0.0024877542164176702, + 0.003400332061573863, + 0.003838188713416457, + 0.005644747521728277, + 0.005720373243093491, + 0.005119694396853447, + 0.006844210438430309, + 0.0043114866130054, + 0.0016880331095308065, + 0.0019376573618501425, + 0.003214032156392932, + 0.0024757524952292442, + 0.0026118140667676926, + 0.0013143511023372412, + 0.005105276126414537, + 0.002020782558247447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.02985270321369171 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790029913187027 + ], + "opposite_logprob_tokens": [ + -2.579003095626831 + ], + "logprob_reward": -0.0790029913187027, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790029913187027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15675683319568634, + 0.062402911484241486, + 0.1049494817852974, + 0.15996645390987396, + 0.0724112018942833, + 0.16677609086036682, + 0.13515089452266693, + 0.05083801969885826, + 0.06253595650196075, + 0.028212133795022964 + ], + "credit_map": [ + 0.013680324889719486, + 0.017595157027244568, + 0.01798429898917675, + 0.023492377251386642, + 0.02862369827926159, + 0.013849752955138683, + 0.02401009574532509, + 0.017521120607852936, + 0.006618478801101446, + 0.004437499679625034, + 0.00591715844348073, + 0.004649268928915262, + 0.008020470850169659, + 0.015271704643964767, + 0.006546164397150278, + 0.010942168533802032, + 0.00824581179767847, + 0.0090923011302948, + 0.010623103007674217, + 0.012047160416841507, + 0.01614180952310562, + 0.018274419009685516, + 0.016914842650294304, + 0.013610037975013256, + 0.024305855855345726, + 0.009191765449941158, + 0.024822557345032692, + 0.00955392699688673, + 0.031932275742292404, + 0.028684303164482117, + 0.010631762444972992, + 0.020844001322984695, + 0.008259939029812813, + 0.00846797227859497, + 0.010306301526725292, + 0.007660130970180035, + 0.012401450425386429, + 0.012373133562505245, + 0.0049329036846756935, + 0.008009364828467369, + 0.011918582022190094, + 0.01995217055082321, + 0.014802665449678898, + 0.01868400163948536, + 0.03308500722050667, + 0.017258673906326294, + 0.016958223655819893, + 0.03411676734685898, + 0.0174235999584198, + 0.013507804833352566, + 0.030262257903814316, + 0.012383716180920601, + 0.01348136831074953, + 0.013756143860518932, + 0.015231085941195488, + 0.019104909151792526, + 0.0037504269275814295, + 0.007225789129734039, + 0.004223308060318232, + 0.006963634397834539, + 0.008545695804059505, + 0.006583987269550562, + 0.005637380760163069, + 0.00790779571980238, + 0.004840172827243805, + 0.00515385577455163, + 0.011035232804715633, + 0.007841571234166622, + 0.006313961930572987, + 0.006937545724213123, + 0.012101379223167896, + 0.008312240242958069, + 0.0024385335855185986, + 0.0020039768423885107, + 0.006364412140101194, + 0.0027398737147450447, + 0.0023208847269415855, + 0.003477756166830659, + 0.0024162675254046917, + 0.006450430490076542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.0790029913187027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992570996284485 + ], + "opposite_logprob_tokens": [ + -1.8049256801605225 + ], + "logprob_reward": -0.17992570996284485, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992570996284485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13910746574401855, + 0.0844665989279747, + 0.04740617051720619, + 0.29411545395851135, + 0.14002233743667603, + 0.11956655979156494, + 0.09932611137628555, + 0.021339550614356995, + 0.0263822041451931, + 0.028267526999115944 + ], + "credit_map": [ + 0.014271770603954792, + 0.012745152227580547, + 0.024578692391514778, + 0.01709362491965294, + 0.021564481779932976, + 0.01797059178352356, + 0.01585458591580391, + 0.015028578229248524, + 0.007363470736891031, + 0.00594389159232378, + 0.020235858857631683, + 0.00863786693662405, + 0.010860961861908436, + 0.008831856772303581, + 0.00907412450760603, + 0.01351857278496027, + 0.007415860425680876, + 0.004082649014890194, + 0.0063866497948765755, + 0.0038433449808508158, + 0.009534383192658424, + 0.005556477699428797, + 0.004735125228762627, + 0.005851683672517538, + 0.04391755908727646, + 0.016765641048550606, + 0.046867817640304565, + 0.02320403791964054, + 0.05222921073436737, + 0.05413689464330673, + 0.017369916662573814, + 0.03962438926100731, + 0.01578681915998459, + 0.011715400032699108, + 0.027223963290452957, + 0.014362230896949768, + 0.01990964263677597, + 0.02284429781138897, + 0.009434126317501068, + 0.018745847046375275, + 0.013667689636349678, + 0.012498214840888977, + 0.01766974665224552, + 0.008486025966703892, + 0.021389195695519447, + 0.014480805955827236, + 0.008494094014167786, + 0.022880781441926956, + 0.0193171463906765, + 0.008582324720919132, + 0.017985539510846138, + 0.006857286673039198, + 0.009671569801867008, + 0.012998197227716446, + 0.008918816223740578, + 0.014995223842561245, + 0.0017100920667871833, + 0.002892305376008153, + 0.002759654074907303, + 0.0026918293442577124, + 0.003913408145308495, + 0.0021583440247923136, + 0.0023987507447600365, + 0.002815165091305971, + 0.0023016249760985374, + 0.005333031993359327, + 0.0034225720446556807, + 0.0030581667087972164, + 0.0035229597706347704, + 0.0024224475491791964, + 0.0034287648741155863, + 0.0028926380909979343, + 0.0038327593356370926, + 0.0017337911995127797, + 0.006925896275788546, + 0.0036262006033211946, + 0.0025738077238202095, + 0.0022667364683002234, + 0.004002278205007315, + 0.003306058468297124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.17992570996284485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038790345191956 + ], + "opposite_logprob_tokens": [ + -1.9103878736495972 + ], + "logprob_reward": -0.16038790345191956, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038790345191956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13267481327056885, + 0.10407985001802444, + 0.1508169174194336, + 0.05757554620504379, + 0.09116848558187485, + 0.1277751326560974, + 0.05910411849617958, + 0.06768544763326645, + 0.17318804562091827, + 0.035931605845689774 + ], + "credit_map": [ + 0.011317248456180096, + 0.0116658303886652, + 0.02028021030128002, + 0.01673603616654873, + 0.022603316232562065, + 0.011403449811041355, + 0.02405535615980625, + 0.014613369479775429, + 0.0104390112683177, + 0.006425010040402412, + 0.01276098471134901, + 0.010203543119132519, + 0.011635330505669117, + 0.029317202046513557, + 0.009073804132640362, + 0.01422495860606432, + 0.017042793333530426, + 0.012347287498414516, + 0.024197330698370934, + 0.01268432941287756, + 0.01920909620821476, + 0.039124660193920135, + 0.014925725758075714, + 0.011285698041319847, + 0.006094061769545078, + 0.006684368010610342, + 0.006627767812460661, + 0.0066499062813818455, + 0.008919108659029007, + 0.00834941491484642, + 0.006545593496412039, + 0.007705325726419687, + 0.008832202292978764, + 0.008488267660140991, + 0.008457298390567303, + 0.0107542984187603, + 0.020331820473074913, + 0.01436726376414299, + 0.008086388930678368, + 0.011850947514176369, + 0.008601520210504532, + 0.012710701674222946, + 0.016081668436527252, + 0.015129794366657734, + 0.017902899533510208, + 0.018557202070951462, + 0.016770297661423683, + 0.02202104777097702, + 0.005823113955557346, + 0.00601763604208827, + 0.0068174502812325954, + 0.005969174671918154, + 0.009181817062199116, + 0.007224722299724817, + 0.00961062591522932, + 0.008459582924842834, + 0.005985158029943705, + 0.012303601950407028, + 0.006096417084336281, + 0.007276007439941168, + 0.01303862500935793, + 0.007343607023358345, + 0.00788532942533493, + 0.007756706792861223, + 0.011092864908277988, + 0.02106628566980362, + 0.025982465595006943, + 0.022535698488354683, + 0.017070334404706955, + 0.02072676457464695, + 0.025012867525219917, + 0.029700759798288345, + 0.0026692610699683428, + 0.003979444969445467, + 0.006067728158086538, + 0.004946387838572264, + 0.0047152079641819, + 0.0027516523841768503, + 0.006013409700244665, + 0.004788513761013746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.16038790345191956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera held still, its focus fixed squarely on the hem of a deep emerald green silk cheongsam. Crafted from the finest mulberry silk, the cheongsam takes on a fluid, pearly sheen in the soft afternoon light filtering through the carved window lattices. Embroidered with a magnificent pattern of phoenix tail feathers, delicately embroidered in golden thread, each feather is rendered vividly. A gentle breeze caresses the flowing skirt, gracefully billowing. High side slits allow the silky fabric to rhythmically undulate like waves, sometimes brushing gently against the polished dark wood flooring below, sometimes rising slightly into the air to reveal a hint of the equally textured lining. Light and shadow play dappled across the dancing skirt, highlighting the delicate texture of the silk and the three-dimensional quality of the embroidery. In the background, a blurred, out-of-focus scene of a classical garden, faintly visible are the shadows of swaying bamboo. The entire scene is imbued with a tranquil, classical, and poetic beauty, creating an elegant atmosphere where time stands still.", + "seed": 28620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_5/index_4347_seed_28620_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.29835096225142477, + "vlm_logprob": -0.29835096225142477, + "vlm_yes_no_margin": 1.625, + "vlm_token_credit": -0.29835096225142477, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8433437943458557 + ], + "opposite_logprob_tokens": [ + -0.5933437943458557 + ], + "logprob_reward": -0.8433437943458557, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8433437943458557, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07864838093519211, + 0.1144922524690628, + 0.24324165284633636, + 0.0652017816901207, + 0.05877432972192764, + 0.24796776473522186, + 0.08394032716751099, + 0.06518042832612991, + 0.015390847809612751, + 0.027162279933691025 + ], + "credit_map": [ + 0.004898631013929844, + 0.009623300284147263, + 0.009190794080495834, + 0.010613157413899899, + 0.012066282331943512, + 0.0052614896558225155, + 0.015528375282883644, + 0.011466347612440586, + 0.008685438893735409, + 0.010698062367737293, + 0.016896355897188187, + 0.021959641948342323, + 0.0180217232555151, + 0.007290356792509556, + 0.013039369136095047, + 0.017901310697197914, + 0.014251098036766052, + 0.014223643578588963, + 0.008197160437703133, + 0.09083632379770279, + 0.08176479488611221, + 0.00874217227101326, + 0.012890608981251717, + 0.012335848063230515, + 0.00584323238581419, + 0.01139847468584776, + 0.007589535787701607, + 0.009737169370055199, + 0.010566300712525845, + 0.0046860272996127605, + 0.008401203900575638, + 0.006979837082326412, + 0.0037143914960324764, + 0.007404332049190998, + 0.008459028787910938, + 0.010423561558127403, + 0.009144715033471584, + 0.003616548841819167, + 0.009760315530002117, + 0.0062514403834939, + 0.011936726048588753, + 0.040976058691740036, + 0.01463329792022705, + 0.029249966144561768, + 0.07894151657819748, + 0.012599607929587364, + 0.0352935828268528, + 0.02433699183166027, + 0.0043951054103672504, + 0.01123702060431242, + 0.009871667250990868, + 0.01582983508706093, + 0.013564792461693287, + 0.0033085583709180355, + 0.015513516031205654, + 0.01021982729434967, + 0.00249969819560647, + 0.00716719776391983, + 0.004148542415350676, + 0.009675622917711735, + 0.020282937213778496, + 0.0028701825067400932, + 0.013935884460806847, + 0.0046003637835383415, + 0.0008122730068862438, + 0.0022325848694890738, + 0.0017661628080531955, + 0.002965347608551383, + 0.0026916777715086937, + 0.0012779353419318795, + 0.002007793402299285, + 0.0016370731173083186, + 0.0013626809231936932, + 0.002160955686122179, + 0.00261997664347291, + 0.002516144886612892, + 0.009245442226529121, + 0.0018523054895922542, + 0.005397330969572067, + 0.002007443457841873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.8433437943458557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043125856667757034 + ], + "opposite_logprob_tokens": [ + -3.168125867843628 + ], + "logprob_reward": -0.043125856667757034, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043125856667757034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15211330354213715, + 0.060728706419467926, + 0.24146316945552826, + 0.16435331106185913, + 0.046070072799921036, + 0.09599423408508301, + 0.04055960103869438, + 0.046569276601076126, + 0.09898187965154648, + 0.053166475147008896 + ], + "credit_map": [ + 0.010952455922961235, + 0.01929900236427784, + 0.0193118117749691, + 0.016765490174293518, + 0.03456488996744156, + 0.008513973094522953, + 0.027928581461310387, + 0.014777103438973427, + 0.004282726440578699, + 0.006142506375908852, + 0.007438294123858213, + 0.013953158631920815, + 0.011880075559020042, + 0.003481585066765547, + 0.006663983687758446, + 0.006886373274028301, + 0.03443434089422226, + 0.02901129424571991, + 0.018319396302103996, + 0.04610248655080795, + 0.043307963758707047, + 0.021409587934613228, + 0.019191598519682884, + 0.029686499387025833, + 0.012465412728488445, + 0.02215825580060482, + 0.02195751667022705, + 0.02446276694536209, + 0.02807830087840557, + 0.009697606787085533, + 0.021870560944080353, + 0.02366289682686329, + 0.004231319297105074, + 0.010362210683524609, + 0.0033466049935668707, + 0.005873823072761297, + 0.00641475897282362, + 0.0032907004933804274, + 0.006774283479899168, + 0.005776371341198683, + 0.005276774056255817, + 0.014683927409350872, + 0.0062206806614995, + 0.013050203211605549, + 0.027323894202709198, + 0.006383877247571945, + 0.0130009138956666, + 0.010053970851004124, + 0.0030687639955431223, + 0.006267650052905083, + 0.005894763860851526, + 0.005951218772679567, + 0.004442177247256041, + 0.002246176591143012, + 0.006813128013163805, + 0.005875722970813513, + 0.002180175157263875, + 0.004847257863730192, + 0.003611613530665636, + 0.006969874259084463, + 0.012172648683190346, + 0.003017489332705736, + 0.007878906093537807, + 0.005891310516744852, + 0.005011825822293758, + 0.01124467421323061, + 0.009158993139863014, + 0.023119863122701645, + 0.017887121066451073, + 0.0045571173541247845, + 0.01788146421313286, + 0.010120823048055172, + 0.002571034710854292, + 0.004449337255209684, + 0.008270515128970146, + 0.006012967322021723, + 0.012674698606133461, + 0.003665059106424451, + 0.009528220631182194, + 0.0059946440160274506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.043125856667757034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054971449077129364 + ], + "opposite_logprob_tokens": [ + -2.92997145652771 + ], + "logprob_reward": -0.054971449077129364, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054971449077129364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11027844995260239, + 0.09186656773090363, + 0.27992331981658936, + 0.09727372974157333, + 0.12529735267162323, + 0.08829857409000397, + 0.09954170882701874, + 0.0506574921309948, + 0.028723783791065216, + 0.028139065951108932 + ], + "credit_map": [ + 0.00759557681158185, + 0.010015981271862984, + 0.0163218192756176, + 0.02196212112903595, + 0.016723832115530968, + 0.006100683007389307, + 0.01590554602444172, + 0.015652891248464584, + 0.005786430090665817, + 0.00856848619878292, + 0.011893685907125473, + 0.021996039897203445, + 0.01490295585244894, + 0.005945245269685984, + 0.01043575257062912, + 0.01233796589076519, + 0.011973923072218895, + 0.016275862231850624, + 0.008106506429612637, + 0.10808733850717545, + 0.09533172845840454, + 0.009274045005440712, + 0.015822581946849823, + 0.015051336027681828, + 0.004762381315231323, + 0.014426476322114468, + 0.010473930276930332, + 0.02181815728545189, + 0.015106976963579655, + 0.004887067247182131, + 0.014950299635529518, + 0.010848444886505604, + 0.009657374583184719, + 0.012127275578677654, + 0.02992560900747776, + 0.01596495695412159, + 0.020890885964035988, + 0.005725172348320484, + 0.016890693455934525, + 0.01411539874970913, + 0.007546884007751942, + 0.011546437628567219, + 0.011412515304982662, + 0.01090557686984539, + 0.016274545341730118, + 0.00649176025763154, + 0.01101542729884386, + 0.013105425052344799, + 0.005127562675625086, + 0.013564331457018852, + 0.009768348187208176, + 0.022588839754462242, + 0.010032995603978634, + 0.0032426153775304556, + 0.017079751938581467, + 0.018137263134121895, + 0.0014875372871756554, + 0.006509586237370968, + 0.004032573662698269, + 0.008728228509426117, + 0.01272091269493103, + 0.0020402136724442244, + 0.01114761084318161, + 0.003990828990936279, + 0.0014420540537685156, + 0.002667617052793503, + 0.0030972470995038748, + 0.005528315901756287, + 0.005822290200740099, + 0.0017717235023155808, + 0.005102045368403196, + 0.00329248933121562, + 0.001477814163081348, + 0.0018887887708842754, + 0.003154244041070342, + 0.003169392701238394, + 0.00813782773911953, + 0.0016152827301993966, + 0.004322821740061045, + 0.004372893832623959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.054971449077129364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016158401966095 + ], + "opposite_logprob_tokens": [ + -1.701615810394287 + ], + "logprob_reward": -0.2016158401966095, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016158401966095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08372791111469269, + 0.15956883132457733, + 0.11452822387218475, + 0.1769886612892151, + 0.051020003855228424, + 0.2601219713687897, + 0.05377206578850746, + 0.07013741135597229, + 0.02136351726949215, + 0.008771401830017567 + ], + "credit_map": [ + 0.008974823169410229, + 0.006814930588006973, + 0.013638324104249477, + 0.013428493402898312, + 0.01396183855831623, + 0.00816273968666792, + 0.010952609591186047, + 0.007794148288667202, + 0.009713800624012947, + 0.014337821863591671, + 0.023067167028784752, + 0.03678556904196739, + 0.029181338846683502, + 0.008121352642774582, + 0.01834949105978012, + 0.020012281835079193, + 0.010934244841337204, + 0.009145867079496384, + 0.010802258737385273, + 0.02156263403594494, + 0.024001887068152428, + 0.009411318227648735, + 0.014432646334171295, + 0.014237361960113049, + 0.014839389361441135, + 0.02113470807671547, + 0.024869345128536224, + 0.02881317399442196, + 0.03192708268761635, + 0.009365181438624859, + 0.024904873222112656, + 0.021134890615940094, + 0.003384550102055073, + 0.007266639266163111, + 0.006136317737400532, + 0.010558821260929108, + 0.0077342623844742775, + 0.0027114958502352238, + 0.008003350347280502, + 0.0052245622500777245, + 0.02299877069890499, + 0.020668035373091698, + 0.030268078669905663, + 0.031351618468761444, + 0.03723280876874924, + 0.021220866590738297, + 0.039584383368492126, + 0.05679740011692047, + 0.0035811245907098055, + 0.006667857524007559, + 0.006184312980622053, + 0.008073999546468258, + 0.009288148023188114, + 0.0031027859076857567, + 0.007778366561979055, + 0.009095467627048492, + 0.0032825730741024017, + 0.008298703469336033, + 0.0073966775089502335, + 0.01402218546718359, + 0.013965015299618244, + 0.004044777248054743, + 0.012348392978310585, + 0.006779077462852001, + 0.001355014625005424, + 0.0024610962718725204, + 0.00330729759298265, + 0.003409704891964793, + 0.003518462646752596, + 0.0015407279133796692, + 0.0034258319064974785, + 0.0023453794419765472, + 0.0007494091987609863, + 0.0010096961632370949, + 0.0009077925351448357, + 0.0009954163106158376, + 0.0017380889039486647, + 0.000649369671009481, + 0.0011813470628112555, + 0.001540281460620463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.2016158401966095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34869787096977234 + ], + "opposite_logprob_tokens": [ + -1.2236979007720947 + ], + "logprob_reward": -0.34869787096977234, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34869787096977234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1194036528468132, + 0.06987175345420837, + 0.18940047919750214, + 0.08619246631860733, + 0.07688870280981064, + 0.24271023273468018, + 0.08470281958580017, + 0.0610671266913414, + 0.036173269152641296, + 0.033589500933885574 + ], + "credit_map": [ + 0.007559735793620348, + 0.011714677326381207, + 0.014353961683809757, + 0.022270970046520233, + 0.019267406314611435, + 0.00616278825327754, + 0.019083172082901, + 0.01899094507098198, + 0.004425655584782362, + 0.008851997554302216, + 0.008090556599199772, + 0.011285046115517616, + 0.014023831114172935, + 0.003826100379228592, + 0.01205087173730135, + 0.007317685056477785, + 0.013564929366111755, + 0.01593102142214775, + 0.011888174340128899, + 0.05885986611247063, + 0.04827951639890671, + 0.01027870737016201, + 0.01121586188673973, + 0.019382404163479805, + 0.0067691924050450325, + 0.01140793226659298, + 0.009251994080841541, + 0.009835069999098778, + 0.01694481261074543, + 0.0051668379455804825, + 0.010439795441925526, + 0.01637682318687439, + 0.006936202757060528, + 0.007097579073160887, + 0.014986288733780384, + 0.0072885616682469845, + 0.014637527987360954, + 0.003998580388724804, + 0.011441742070019245, + 0.010502221062779427, + 0.0111045828089118, + 0.043862707912921906, + 0.01643543131649494, + 0.026985732838511467, + 0.0748181939125061, + 0.011157387867569923, + 0.03589572384953499, + 0.02245047502219677, + 0.0036394414491951466, + 0.013232487253844738, + 0.00736185722053051, + 0.012953060679137707, + 0.020279528573155403, + 0.005318203940987587, + 0.013720463961362839, + 0.008197779767215252, + 0.002393791452050209, + 0.006473516579717398, + 0.005700071342289448, + 0.011170837096869946, + 0.015683794394135475, + 0.00253780628554523, + 0.010840971022844315, + 0.006266339682042599, + 0.001627163728699088, + 0.005823081359267235, + 0.00428737560287118, + 0.00678941560909152, + 0.006048545707017183, + 0.0018063272582367063, + 0.005846026819199324, + 0.003945331089198589, + 0.0022422836627811193, + 0.0041199373081326485, + 0.003670812351629138, + 0.0034809925127774477, + 0.0072410800494253635, + 0.0034226977732032537, + 0.0046582757495343685, + 0.004753418732434511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.34869787096977234 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera held still, its focus fixed squarely on the hem of a deep emerald green silk cheongsam. Crafted from the finest mulberry silk, the cheongsam takes on a fluid, pearly sheen in the soft afternoon light filtering through the carved window lattices. Embroidered with a magnificent pattern of phoenix tail feathers, delicately embroidered in golden thread, each feather is rendered vividly. A gentle breeze caresses the flowing skirt, gracefully billowing. High side slits allow the silky fabric to rhythmically undulate like waves, sometimes brushing gently against the polished dark wood flooring below, sometimes rising slightly into the air to reveal a hint of the equally textured lining. Light and shadow play dappled across the dancing skirt, highlighting the delicate texture of the silk and the three-dimensional quality of the embroidery. In the background, a blurred, out-of-focus scene of a classical garden, faintly visible are the shadows of swaying bamboo. The entire scene is imbued with a tranquil, classical, and poetic beauty, creating an elegant atmosphere where time stands still.", + "seed": 28621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_5/index_4347_seed_28621_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09764640629291535, + "vlm_logprob": -0.09764640629291535, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.09764640629291535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22923344373703003 + ], + "opposite_logprob_tokens": [ + -1.7292335033416748 + ], + "logprob_reward": -0.22923344373703003, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22923344373703003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14800480008125305, + 0.036812976002693176, + 0.025973008945584297, + 0.01703973300755024, + 0.022198809310793877, + 0.03149982541799545, + 0.031013067811727524, + 0.5272103548049927, + 0.059278421103954315, + 0.10096895694732666 + ], + "credit_map": [ + 0.006506717763841152, + 0.006839558947831392, + 0.022650940343737602, + 0.0070630330592393875, + 0.013290125876665115, + 0.056140974164009094, + 0.020989445969462395, + 0.014524013735353947, + 0.001372586004436016, + 0.008927352726459503, + 0.0017388054402545094, + 0.004617076367139816, + 0.0029252036474645138, + 0.00536706717684865, + 0.002042069798335433, + 0.009822813794016838, + 0.0016068777767941356, + 0.004511554259806871, + 0.002175969537347555, + 0.004164557438343763, + 0.005422818940132856, + 0.002471338491886854, + 0.003463181434199214, + 0.0021567100193351507, + 0.001221289043314755, + 0.0024052662774920464, + 0.0016523683443665504, + 0.0026160809211432934, + 0.0021663953084498644, + 0.0015557942679151893, + 0.003289440181106329, + 0.002133098663762212, + 0.0011426109122112393, + 0.00658683106303215, + 0.0013055149465799332, + 0.0022540236823260784, + 0.0025999355129897594, + 0.0013089668937027454, + 0.005313452333211899, + 0.0016874739667400718, + 0.0015646576648578048, + 0.009676264598965645, + 0.0016055380692705512, + 0.0022817710414528847, + 0.004997008014470339, + 0.0016204087296500802, + 0.007734139449894428, + 0.0020200361032038927, + 0.002294697565957904, + 0.002547862008213997, + 0.006597721949219704, + 0.0037654272746294737, + 0.0047615584917366505, + 0.0021795593202114105, + 0.006100586149841547, + 0.002765653422102332, + 0.05522029846906662, + 0.04693074896931648, + 0.13620831072330475, + 0.04798324406147003, + 0.04964546486735344, + 0.01759171113371849, + 0.13414466381072998, + 0.03948591277003288, + 0.006421335507184267, + 0.014898312278091908, + 0.0033215240109711885, + 0.006541228387504816, + 0.006746811792254448, + 0.0032686730846762657, + 0.012500866316258907, + 0.005579669028520584, + 0.004434272181242704, + 0.010026535019278526, + 0.006838980130851269, + 0.018683206290006638, + 0.015553143806755543, + 0.008049944415688515, + 0.01890191249549389, + 0.018480967730283737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.22923344373703003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05502470210194588 + ], + "opposite_logprob_tokens": [ + -2.9300246238708496 + ], + "logprob_reward": -0.05502470210194588, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05502470210194588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10459594428539276, + 0.07249710708856583, + 0.18749302625656128, + 0.03643004596233368, + 0.02302195318043232, + 0.029368175193667412, + 0.1327754259109497, + 0.316280722618103, + 0.038394197821617126, + 0.059143420308828354 + ], + "credit_map": [ + 0.004056550562381744, + 0.009109152480959892, + 0.015439378097653389, + 0.008776997216045856, + 0.015040389262139797, + 0.030284589156508446, + 0.012658478692173958, + 0.009230414405465126, + 0.002704403130337596, + 0.01771208457648754, + 0.003129668766632676, + 0.008525061421096325, + 0.005899627227336168, + 0.01045237760990858, + 0.0032481413800269365, + 0.020825739949941635, + 0.006662675645202398, + 0.0352415069937706, + 0.01085435040295124, + 0.02456957846879959, + 0.0516594797372818, + 0.021252743899822235, + 0.021385440602898598, + 0.01586724817752838, + 0.0031457801815122366, + 0.006701071746647358, + 0.003935794811695814, + 0.003982881549745798, + 0.004995221737772226, + 0.003543964819982648, + 0.005918425507843494, + 0.004206905607134104, + 0.001179565442726016, + 0.005097659304738045, + 0.0015947729116305709, + 0.004171458538621664, + 0.003672423306852579, + 0.0014490607427433133, + 0.003926405683159828, + 0.0019306077156215906, + 0.0013171761529520154, + 0.0061462074518203735, + 0.0019660883117467165, + 0.003739836160093546, + 0.006773496046662331, + 0.001278979703783989, + 0.005736866034567356, + 0.002409527078270912, + 0.005720240529626608, + 0.010383093729615211, + 0.016474751755595207, + 0.019848983734846115, + 0.017133427783846855, + 0.0050262706354260445, + 0.04868190363049507, + 0.009506766684353352, + 0.013715174049139023, + 0.08342717587947845, + 0.019841628149151802, + 0.035447802394628525, + 0.034014035016298294, + 0.007808796130120754, + 0.09564030170440674, + 0.026385808363556862, + 0.004992841277271509, + 0.009337387979030609, + 0.0026865957770496607, + 0.004488191567361355, + 0.004490358754992485, + 0.003181301988661289, + 0.0061924392357468605, + 0.0030250833369791508, + 0.0029543789569288492, + 0.0049250321462750435, + 0.004580373410135508, + 0.00989740900695324, + 0.01417396031320095, + 0.004913524724543095, + 0.009986840188503265, + 0.00771190645173192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.05502470210194588 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496400222182274 + ], + "opposite_logprob_tokens": [ + -2.929964065551758 + ], + "logprob_reward": -0.05496400222182274, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496400222182274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11877443641424179, + 0.04995180293917656, + 0.0703556165099144, + 0.030378596857190132, + 0.0977749228477478, + 0.07609136402606964, + 0.19622045755386353, + 0.17770977318286896, + 0.05449846014380455, + 0.12824459373950958 + ], + "credit_map": [ + 0.004768923856317997, + 0.007793173659592867, + 0.016789628192782402, + 0.00963021069765091, + 0.01611313782632351, + 0.036928918212652206, + 0.015269441530108452, + 0.011480999179184437, + 0.0026888339780271053, + 0.008830158039927483, + 0.003492735791951418, + 0.004966805689036846, + 0.003310401691123843, + 0.008966491557657719, + 0.002220599912106991, + 0.015475772321224213, + 0.0030288677662611008, + 0.011405288241803646, + 0.006699644960463047, + 0.00928075797855854, + 0.017697183415293694, + 0.006598709616810083, + 0.008558203466236591, + 0.007086954545229673, + 0.003952395636588335, + 0.005140281748026609, + 0.004181945230811834, + 0.004025152884423733, + 0.003512885421514511, + 0.002251886995509267, + 0.0038525694981217384, + 0.003461478278040886, + 0.0048520928248763084, + 0.023959333077073097, + 0.005455471575260162, + 0.01753426529467106, + 0.012280996888875961, + 0.00833425484597683, + 0.01874569058418274, + 0.0066128140315413475, + 0.003096968401223421, + 0.018916746601462364, + 0.005018376745283604, + 0.007688924670219421, + 0.01662924699485302, + 0.003527614288032055, + 0.015667883679270744, + 0.0055455975234508514, + 0.009263824671506882, + 0.011479299515485764, + 0.03800865635275841, + 0.026744235306978226, + 0.020194046199321747, + 0.006462030112743378, + 0.07103011012077332, + 0.013038254342973232, + 0.012414609082043171, + 0.04266275465488434, + 0.018429609015583992, + 0.019006287679076195, + 0.021461255848407745, + 0.005323381628841162, + 0.04375198855996132, + 0.014659852720797062, + 0.008541576564311981, + 0.0045127649791538715, + 0.006300109438598156, + 0.0033481218852102757, + 0.0067312768660485744, + 0.002957608550786972, + 0.008061500266194344, + 0.014045498333871365, + 0.004304491449147463, + 0.010842674411833286, + 0.007439250126481056, + 0.03187941387295723, + 0.02377655915915966, + 0.010451783426105976, + 0.02175711840391159, + 0.01779329590499401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.05496400222182274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048674650490283966 + ], + "opposite_logprob_tokens": [ + -3.0486745834350586 + ], + "logprob_reward": -0.048674650490283966, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048674650490283966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07378801703453064, + 0.32370737195014954, + 0.05569072812795639, + 0.1200084239244461, + 0.0628216564655304, + 0.10845236480236053, + 0.029372360557317734, + 0.0873480886220932, + 0.07485061138868332, + 0.06396037340164185 + ], + "credit_map": [ + 0.005025792866945267, + 0.008329002186655998, + 0.010324577800929546, + 0.005538749508559704, + 0.008929249830543995, + 0.019738556817173958, + 0.008183054625988007, + 0.007719033397734165, + 0.011936364695429802, + 0.08104805648326874, + 0.012808065861463547, + 0.039129674434661865, + 0.018206000328063965, + 0.04891887307167053, + 0.014731298200786114, + 0.09692903608083725, + 0.006850132253021002, + 0.008657213300466537, + 0.008097045123577118, + 0.006283458322286606, + 0.011247505433857441, + 0.004773318767547607, + 0.005440443754196167, + 0.00434160977602005, + 0.017601678147912025, + 0.022642778232693672, + 0.020356634631752968, + 0.01126156933605671, + 0.009687595069408417, + 0.01043617818504572, + 0.01362521667033434, + 0.01439677830785513, + 0.0031458507291972637, + 0.019361745566129684, + 0.0036887237802147865, + 0.006829371675848961, + 0.005535941105335951, + 0.004802855663001537, + 0.014722343534231186, + 0.004734822548925877, + 0.005372983403503895, + 0.034039534628391266, + 0.007428618613630533, + 0.009229549206793308, + 0.011178544722497463, + 0.004841533023864031, + 0.027486810460686684, + 0.00887480191886425, + 0.003045120509341359, + 0.002234404906630516, + 0.0057021742686629295, + 0.0031563357915729284, + 0.003950701095163822, + 0.002727973973378539, + 0.0038082171231508255, + 0.004747434984892607, + 0.004517111927270889, + 0.02157275751233101, + 0.008203786797821522, + 0.009461401961743832, + 0.008161581121385098, + 0.002155157970264554, + 0.027042804285883904, + 0.006233483087271452, + 0.00712579395622015, + 0.024340925738215446, + 0.0038499408401548862, + 0.010404692962765694, + 0.007882892154157162, + 0.0023796309251338243, + 0.015173089690506458, + 0.0036936451215296984, + 0.0032524941489100456, + 0.005012665409594774, + 0.003860549768432975, + 0.010141515173017979, + 0.0165640190243721, + 0.004543388728052378, + 0.01309153251349926, + 0.0074942102655768394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.048674650490283966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033523291349411 + ], + "opposite_logprob_tokens": [ + -2.350335121154785 + ], + "logprob_reward": -0.10033523291349411, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033523291349411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1573960781097412, + 0.027056120336055756, + 0.04813407361507416, + 0.04075023531913757, + 0.05754609405994415, + 0.08551779389381409, + 0.2925383448600769, + 0.1406731754541397, + 0.07320021837949753, + 0.07718781381845474 + ], + "credit_map": [ + 0.008126833476126194, + 0.00477606849744916, + 0.02625195123255253, + 0.007782405242323875, + 0.011754698120057583, + 0.06026073917746544, + 0.020435377955436707, + 0.01800801046192646, + 0.0009807465830817819, + 0.00604632031172514, + 0.0018392970087006688, + 0.0031305744778364897, + 0.0039387973956763744, + 0.0034884687047451735, + 0.002383387880399823, + 0.005248528439551592, + 0.0024758123327046633, + 0.008246992714703083, + 0.005592419300228357, + 0.005968084558844566, + 0.007363323122262955, + 0.00559402396902442, + 0.007847864180803299, + 0.005045553669333458, + 0.0024420018307864666, + 0.0114295007660985, + 0.0026346503291279078, + 0.004216791596263647, + 0.004558928310871124, + 0.0033634030260145664, + 0.007223965600132942, + 0.004880993161350489, + 0.002898662118241191, + 0.013392497785389423, + 0.0034988855477422476, + 0.012095142155885696, + 0.005970660597085953, + 0.0036716703325510025, + 0.010924253612756729, + 0.005094324704259634, + 0.0035111939068883657, + 0.026025494560599327, + 0.00570138543844223, + 0.007748282514512539, + 0.010887674987316132, + 0.0035078509245067835, + 0.02096010372042656, + 0.007175802253186703, + 0.013926601968705654, + 0.027587536722421646, + 0.023617230355739594, + 0.045059483498334885, + 0.04448597878217697, + 0.013329682871699333, + 0.10177262872457504, + 0.022759195417165756, + 0.010837859474122524, + 0.035708650946617126, + 0.029452865943312645, + 0.018268708139657974, + 0.01339301373809576, + 0.003906509838998318, + 0.01873743161559105, + 0.01036812737584114, + 0.007867107167840004, + 0.005434972234070301, + 0.008056183345615864, + 0.005829506553709507, + 0.011865957640111446, + 0.0052530355751514435, + 0.010579830966889858, + 0.018313629552721977, + 0.003385320771485567, + 0.00587531179189682, + 0.005239301361143589, + 0.018062086775898933, + 0.011767195537686348, + 0.008511340245604515, + 0.013205966912209988, + 0.011141288094222546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.10033523291349411 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera held still, its focus fixed squarely on the hem of a deep emerald green silk cheongsam. Crafted from the finest mulberry silk, the cheongsam takes on a fluid, pearly sheen in the soft afternoon light filtering through the carved window lattices. Embroidered with a magnificent pattern of phoenix tail feathers, delicately embroidered in golden thread, each feather is rendered vividly. A gentle breeze caresses the flowing skirt, gracefully billowing. High side slits allow the silky fabric to rhythmically undulate like waves, sometimes brushing gently against the polished dark wood flooring below, sometimes rising slightly into the air to reveal a hint of the equally textured lining. Light and shadow play dappled across the dancing skirt, highlighting the delicate texture of the silk and the three-dimensional quality of the embroidery. In the background, a blurred, out-of-focus scene of a classical garden, faintly visible are the shadows of swaying bamboo. The entire scene is imbued with a tranquil, classical, and poetic beauty, creating an elegant atmosphere where time stands still.", + "seed": 28622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_5/index_4347_seed_28622_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12499641254544258, + "vlm_logprob": -0.12499641254544258, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.12499641254544258, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1521739810705185 + ], + "opposite_logprob_tokens": [ + -2.1521739959716797 + ], + "logprob_reward": -0.1521739810705185, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1521739810705185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08104343712329865, + 0.14984822273254395, + 0.05120793730020523, + 0.1880284547805786, + 0.10572019964456558, + 0.06131596863269806, + 0.2203335165977478, + 0.048374906182289124, + 0.051079925149679184, + 0.04304736107587814 + ], + "credit_map": [ + 0.0060783978551626205, + 0.007967314682900906, + 0.010116365738213062, + 0.009143968112766743, + 0.010467179119586945, + 0.009606469422578812, + 0.015777261927723885, + 0.01188648771494627, + 0.010408543050289154, + 0.016951080411672592, + 0.01766018569469452, + 0.008966445922851562, + 0.04429638385772705, + 0.012296095490455627, + 0.025317810475826263, + 0.013951687142252922, + 0.003907829523086548, + 0.006999250967055559, + 0.003673239378258586, + 0.003919779788702726, + 0.008199487812817097, + 0.007267908193171024, + 0.008291326463222504, + 0.008949119597673416, + 0.01648353599011898, + 0.01628216728568077, + 0.019815368577837944, + 0.016242947429418564, + 0.026530401781201363, + 0.026820948347449303, + 0.029698383063077927, + 0.03615471348166466, + 0.008600627072155476, + 0.018074776977300644, + 0.009428730234503746, + 0.013283095322549343, + 0.016188062727451324, + 0.015333844348788261, + 0.01306015346199274, + 0.01175091415643692, + 0.006848048884421587, + 0.004728763364255428, + 0.0070144617930054665, + 0.0039023577701300383, + 0.006319891661405563, + 0.017097709700465202, + 0.009010856039822102, + 0.006393889430910349, + 0.00926123559474945, + 0.017174459993839264, + 0.010176208801567554, + 0.015416770242154598, + 0.06507149338722229, + 0.033178746700286865, + 0.04434669390320778, + 0.02570793218910694, + 0.0035896694753319025, + 0.005504518747329712, + 0.00510379346087575, + 0.005890108644962311, + 0.006743216421455145, + 0.00578640541061759, + 0.007995259948074818, + 0.0077619426883757114, + 0.0029294383712112904, + 0.007533619645982981, + 0.004661168437451124, + 0.005916583351790905, + 0.00862505566328764, + 0.005606828723102808, + 0.009454920887947083, + 0.006352316588163376, + 0.0037507344968616962, + 0.008209229446947575, + 0.005094991065561771, + 0.0053438288159668446, + 0.006094943266361952, + 0.0033006411977112293, + 0.007088968995958567, + 0.004164027515798807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.1521739810705185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0891265943646431 + ], + "opposite_logprob_tokens": [ + -2.4641265869140625 + ], + "logprob_reward": -0.0891265943646431, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0891265943646431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08503337949514389, + 0.05839911103248596, + 0.0393945649266243, + 0.403587281703949, + 0.06554566323757172, + 0.03275885432958603, + 0.07511033117771149, + 0.10282941907644272, + 0.05297267809510231, + 0.08436870574951172 + ], + "credit_map": [ + 0.006687596440315247, + 0.009208152070641518, + 0.009783363901078701, + 0.00995618849992752, + 0.011250404641032219, + 0.01017703302204609, + 0.01582871377468109, + 0.012141936458647251, + 0.0029153451323509216, + 0.010107919573783875, + 0.004615071229636669, + 0.007918063551187515, + 0.012605960480868816, + 0.007827987894415855, + 0.005546774249523878, + 0.006861991249024868, + 0.003721719840541482, + 0.005321641452610493, + 0.0040983171202242374, + 0.0034432534594088793, + 0.006768663879483938, + 0.004909314215183258, + 0.005335142370313406, + 0.00579651165753603, + 0.024408698081970215, + 0.03907310962677002, + 0.03111501969397068, + 0.04407753422856331, + 0.06543492525815964, + 0.07914526760578156, + 0.05369093641638756, + 0.06664178520441055, + 0.007924061268568039, + 0.006163036450743675, + 0.006476943846791983, + 0.005365401040762663, + 0.00835912860929966, + 0.009914882481098175, + 0.013245359063148499, + 0.008096852339804173, + 0.004574977792799473, + 0.0029816071037203074, + 0.004100066609680653, + 0.003053822088986635, + 0.004127830266952515, + 0.006165948696434498, + 0.0040086726658046246, + 0.0037459309678524733, + 0.0034411833621561527, + 0.0047754147090017796, + 0.004153802525252104, + 0.0053278920240700245, + 0.0234333798289299, + 0.012681297026574612, + 0.013471333310008049, + 0.007826026529073715, + 0.006865423172712326, + 0.010204884223639965, + 0.011465057730674744, + 0.012543775141239166, + 0.013885566033422947, + 0.014064555056393147, + 0.01730097457766533, + 0.01649918034672737, + 0.002671603113412857, + 0.0057118134573102, + 0.004495060537010431, + 0.006035131867974997, + 0.013577231205999851, + 0.005521354265511036, + 0.009742415510118008, + 0.005218065343797207, + 0.0038595532532781363, + 0.023301932960748672, + 0.0070725856348872185, + 0.01068385411053896, + 0.011636607348918915, + 0.0075380876660346985, + 0.013015844859182835, + 0.007260237354785204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.0891265943646431 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707965075969696 + ], + "opposite_logprob_tokens": [ + -2.127079725265503 + ], + "logprob_reward": -0.12707965075969696, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707965075969696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19537153840065002, + 0.08739583939313889, + 0.038205817341804504, + 0.16446033120155334, + 0.14835040271282196, + 0.11747876554727554, + 0.14948996901512146, + 0.03158724308013916, + 0.038849879056215286, + 0.028810232877731323 + ], + "credit_map": [ + 0.011225792579352856, + 0.02226000837981701, + 0.01981431432068348, + 0.0269086305052042, + 0.02535669319331646, + 0.02375028282403946, + 0.03773951902985573, + 0.028316287323832512, + 0.004687275737524033, + 0.01223305705934763, + 0.006803096737712622, + 0.009414957836270332, + 0.01903250627219677, + 0.010602185502648354, + 0.011038362048566341, + 0.013584395870566368, + 0.0024873942602425814, + 0.005490394309163094, + 0.0032485038973391056, + 0.0033021809067577124, + 0.006536754313856363, + 0.006365001667290926, + 0.006063138134777546, + 0.004712447989732027, + 0.01025230810046196, + 0.019478043541312218, + 0.011572551913559437, + 0.018151460215449333, + 0.036932654678821564, + 0.025218011811375618, + 0.020809298381209373, + 0.022045979276299477, + 0.019166268408298492, + 0.019355179741978645, + 0.016292432323098183, + 0.015147404745221138, + 0.021389132365584373, + 0.01463306788355112, + 0.025906316936016083, + 0.016460593789815903, + 0.016099318861961365, + 0.009446900337934494, + 0.016954734921455383, + 0.010832601226866245, + 0.013520695269107819, + 0.023829573765397072, + 0.013929612003266811, + 0.012865322642028332, + 0.00840513501316309, + 0.011603650636970997, + 0.012939265929162502, + 0.014807782135903835, + 0.03738389536738396, + 0.024431537836790085, + 0.022569335997104645, + 0.01734936237335205, + 0.0018040371360257268, + 0.0036636327859014273, + 0.0029322830960154533, + 0.003428261261433363, + 0.005168610718101263, + 0.005053772125393152, + 0.004768776707351208, + 0.004767867736518383, + 0.001811457797884941, + 0.005183027125895023, + 0.003399818204343319, + 0.008150194771587849, + 0.006455590482801199, + 0.0038463345263153315, + 0.006500404793769121, + 0.0035030499566346407, + 0.0024922327138483524, + 0.004096002317965031, + 0.0036425997968763113, + 0.003450315911322832, + 0.004805612377822399, + 0.0025952861178666353, + 0.004147226456552744, + 0.003580957418307662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.12707965075969696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497528240084648 + ], + "opposite_logprob_tokens": [ + -2.9299752712249756 + ], + "logprob_reward": -0.05497528240084648, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497528240084648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035551927983760834, + 0.057987671345472336, + 0.04010402038693428, + 0.6798682808876038, + 0.0969734936952591, + 0.04319639131426811, + 0.014699611812829971, + 0.008726575411856174, + 0.005927822552621365, + 0.016964280977845192 + ], + "credit_map": [ + 0.0028287656605243683, + 0.004449417814612389, + 0.0045202672481536865, + 0.004243913572281599, + 0.005926206707954407, + 0.00443474343046546, + 0.004562588408589363, + 0.004586020018905401, + 0.0035583300050348043, + 0.009948746301233768, + 0.005387051962316036, + 0.006066204980015755, + 0.013714747503399849, + 0.005080849397927523, + 0.008048242889344692, + 0.0061834887601435184, + 0.0031294149812310934, + 0.005640899296849966, + 0.0028286809101700783, + 0.0034450243692845106, + 0.006484069395810366, + 0.008301792666316032, + 0.00495514739304781, + 0.005318985320627689, + 0.06371802091598511, + 0.060427457094192505, + 0.07576915621757507, + 0.05678762122988701, + 0.09439805150032043, + 0.09315494447946548, + 0.10096092522144318, + 0.1346520334482193, + 0.015969865024089813, + 0.006346318870782852, + 0.014755544252693653, + 0.005789433140307665, + 0.01050481479614973, + 0.014567641541361809, + 0.019468899816274643, + 0.009570959955453873, + 0.007207117974758148, + 0.004665824584662914, + 0.006664586253464222, + 0.0038226647302508354, + 0.004351069685071707, + 0.006353799253702164, + 0.004162528086453676, + 0.005968799348920584, + 0.0014523466816172004, + 0.001941544353030622, + 0.001626331708393991, + 0.00132315035443753, + 0.002291487529873848, + 0.0017897062934935093, + 0.0027095824480056763, + 0.0015654605813324451, + 0.0006438668351620436, + 0.000862828572280705, + 0.0009690718143247068, + 0.0009384233271703124, + 0.0011243370827287436, + 0.0011672445107251406, + 0.0015152813866734505, + 0.0015055208932608366, + 0.0003657547349575907, + 0.0007536777993664145, + 0.0006370196933858097, + 0.0005129881901666522, + 0.0010830702958628535, + 0.0005712569691240788, + 0.0011315954616293311, + 0.00087245914619416, + 0.0007788970833644271, + 0.005023387260735035, + 0.0013450384140014648, + 0.00201111170463264, + 0.002454773522913456, + 0.001855276757851243, + 0.002126960549503565, + 0.0013688354520127177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.05497528240084648 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the garment a deep emerald green silk cheongsam with golden phoenix tail feather embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skirt billow and undulate due to a gentle breeze, brushing against the dark wood flooring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred classical garden with faint shadows of swaying bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still with its focus fixed squarely on the hem of the cheongsam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20162655413150787 + ], + "opposite_logprob_tokens": [ + -1.7016265392303467 + ], + "logprob_reward": -0.20162655413150787, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20162655413150787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10122194141149521, + 0.1269180178642273, + 0.025300435721874237, + 0.22064898908138275, + 0.1096872091293335, + 0.04581822082400322, + 0.16429318487644196, + 0.027541611343622208, + 0.15465253591537476, + 0.023917904123663902 + ], + "credit_map": [ + 0.0060553597286343575, + 0.01421752292662859, + 0.006955110002309084, + 0.012805343605577946, + 0.013263615779578686, + 0.01804295927286148, + 0.014140913262963295, + 0.01574111543595791, + 0.004689974244683981, + 0.024814730510115623, + 0.00852394849061966, + 0.015433095395565033, + 0.030121970921754837, + 0.012522263452410698, + 0.017709963023662567, + 0.013102065771818161, + 0.0025581142399460077, + 0.0034555173479020596, + 0.0025925429072231054, + 0.002037873724475503, + 0.004678589291870594, + 0.0027758581563830376, + 0.0033583457116037607, + 0.0038435959722846746, + 0.011020392179489136, + 0.0342310406267643, + 0.014211667701601982, + 0.023624485358595848, + 0.039352238178253174, + 0.03656844422221184, + 0.028528599068522453, + 0.03311212360858917, + 0.011074647307395935, + 0.009756599552929401, + 0.009246749803423882, + 0.010731271468102932, + 0.015991175547242165, + 0.01844981126487255, + 0.02158922515809536, + 0.012847729958593845, + 0.0077020712196826935, + 0.004529197700321674, + 0.006863811053335667, + 0.0033199689351022243, + 0.004230218473821878, + 0.008275764063000679, + 0.005874429829418659, + 0.005022759549319744, + 0.007779363542795181, + 0.010807666927576065, + 0.0070482343435287476, + 0.010514115914702415, + 0.05316217243671417, + 0.028660176321864128, + 0.02901245653629303, + 0.017308993265032768, + 0.0027966261841356754, + 0.0025208645965903997, + 0.002849596319720149, + 0.003863225458189845, + 0.0043304795399308205, + 0.003474646247923374, + 0.0040510306134819984, + 0.0036551421508193016, + 0.006547512486577034, + 0.017760977149009705, + 0.010630613192915916, + 0.023745033890008926, + 0.037804439663887024, + 0.011231068521738052, + 0.031217793002724648, + 0.0157150961458683, + 0.0016694748774170876, + 0.00433113519102335, + 0.0024643237702548504, + 0.004199271090328693, + 0.002982919802889228, + 0.0025614099577069283, + 0.003429755801334977, + 0.0022796127013862133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.20162655413150787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_tensor([1808, 1808])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_tensor([1808, 1808])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..449a2510b379b91c80f58535ee02dc6881a8f4b5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_tensor([1808, 1808])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, upward tracking shot follows, revealing a massive, cyan dragon, as vast as a mountain. Its sinuous, serpentine form is covered in countless shield-sized scales, each shimmering in the light with iridescent hues ranging from jade to emerald. With a rhythm imbued with power and unparalleled grace, the dragon weaves its way through the swirling sea of clouds below. Like boiling milk, the sea of clouds forms countless towering peaks and deep canyons, through which the dragon deftly weaves, its flowing mane like white silk and long whiskers dancing in the currents. Its majestic head, crowned by a pair of towering horns reaching toward the sky, gazes forward from amber eyes blazing with ancient wisdom. Hidden beneath the clouds, the sun casts thousands of golden rays (god rays) upward, piercing through the gaps and casting a dazzling outline of light around the dragon's belly and the edge of the sea of clouds. With each leisurely swing of the animal, some of its scales instantly reflect dazzling light. The entire picture is filled with mythical grandeur and a quiet and powerful sense of power, creating an awe-inspiring epic atmosphere.", + "seed": 21464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_1808_seed_21464_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06626749522984028, + "vlm_logprob": -0.06626749522984028, + "vlm_yes_no_margin": 3.3999999523162843, + "vlm_token_credit": -0.06626749522984028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012792782858014107 + ], + "opposite_logprob_tokens": [ + -4.512792587280273 + ], + "logprob_reward": -0.012792782858014107, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012792782858014107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0889880508184433, + 0.09880255162715912, + 0.01439203042536974, + 0.023262115195393562, + 0.03302425146102905, + 0.019049616530537605, + 0.02303018979728222, + 0.6210644841194153, + 0.03963500261306763, + 0.03875168412923813 + ], + "credit_map": [ + 0.019021499902009964, + 0.011067424900829792, + 0.012914841063320637, + 0.011909297667443752, + 0.01065034605562687, + 0.009251796640455723, + 0.008249902166426182, + 0.00592294055968523, + 0.008939354680478573, + 0.010908843949437141, + 0.01259568054229021, + 0.013652334921061993, + 0.01711900345981121, + 0.013353470712900162, + 0.012033683247864246, + 0.01020018383860588, + 0.0017629521898925304, + 0.0020398651249706745, + 0.001372437458485365, + 0.0013717968249693513, + 0.0026849438436329365, + 0.0018702190136536956, + 0.00203332188539207, + 0.0012564950156956911, + 0.0021639103069901466, + 0.003097856417298317, + 0.0026207193732261658, + 0.003154983976855874, + 0.0033126831986010075, + 0.002752187428995967, + 0.0028357324190437794, + 0.003324044868350029, + 0.003572171088308096, + 0.004349576774984598, + 0.00551282474771142, + 0.005354663822799921, + 0.0035362811759114265, + 0.003489146940410137, + 0.0034387714695185423, + 0.003770817071199417, + 0.0027380618266761303, + 0.002326373476535082, + 0.0022375606931746006, + 0.0018877654802054167, + 0.002161224139854312, + 0.0026480017695575953, + 0.0028291605412960052, + 0.0022214686032384634, + 0.003438284620642662, + 0.006316262762993574, + 0.0022527535911649466, + 0.0021814804058521986, + 0.003458841471001506, + 0.0021933806128799915, + 0.0013831733958795667, + 0.0018060134025290608, + 0.22587481141090393, + 0.036108098924160004, + 0.016692452132701874, + 0.015361660160124302, + 0.24416318535804749, + 0.033954739570617676, + 0.023096688091754913, + 0.02581288293004036, + 0.004744280129671097, + 0.007137894164770842, + 0.0037390817888081074, + 0.006664179265499115, + 0.0050136372447013855, + 0.002869573887437582, + 0.004496595822274685, + 0.004969761241227388, + 0.0045821527019143105, + 0.0061733303591609, + 0.0040320781990885735, + 0.005127916112542152, + 0.00653507374227047, + 0.0020886934362351894, + 0.006392874289304018, + 0.0038195671513676643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.012792782858014107 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903945982456207 + ], + "opposite_logprob_tokens": [ + -2.4640395641326904 + ], + "logprob_reward": -0.08903945982456207, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903945982456207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1029742956161499, + 0.18856513500213623, + 0.069450743496418, + 0.029480695724487305, + 0.06588766723871231, + 0.07025019824504852, + 0.04166174307465553, + 0.26424264907836914, + 0.10074210166931152, + 0.06674475222826004 + ], + "credit_map": [ + 0.017669130116701126, + 0.011435456573963165, + 0.012069006450474262, + 0.012288517318665981, + 0.01777898706495762, + 0.009107914753258228, + 0.011720575392246246, + 0.010904711671173573, + 0.022469647228717804, + 0.024222826585173607, + 0.027046190574765205, + 0.02700703591108322, + 0.020022336393594742, + 0.031262513250112534, + 0.02125663310289383, + 0.015277951955795288, + 0.005929003935307264, + 0.009379726834595203, + 0.008061393164098263, + 0.006331293378025293, + 0.014661600813269615, + 0.00813721027225256, + 0.010947265662252903, + 0.006003251299262047, + 0.0034667726140469313, + 0.004321663174778223, + 0.004476385656744242, + 0.003604137571528554, + 0.002653936855494976, + 0.0037233324255794287, + 0.0029742829501628876, + 0.0042601823806762695, + 0.007077243644744158, + 0.0077809300273656845, + 0.008309761993587017, + 0.009113380685448647, + 0.007767241448163986, + 0.0073668817058205605, + 0.008887339383363724, + 0.009584891609847546, + 0.011263399384915829, + 0.0077567873522639275, + 0.006641780026257038, + 0.0077172196470201015, + 0.008912731893360615, + 0.0107311075553298, + 0.009042990393936634, + 0.00818418338894844, + 0.005589935928583145, + 0.0074830870144069195, + 0.00526050291955471, + 0.004577389918267727, + 0.005118691828101873, + 0.004560390021651983, + 0.0035797222517430782, + 0.005492024123668671, + 0.08970805257558823, + 0.020102716982364655, + 0.009024597704410553, + 0.01076615508645773, + 0.09669356793165207, + 0.014468088746070862, + 0.01120439451187849, + 0.012275045737624168, + 0.012984156608581543, + 0.01814272254705429, + 0.009813264012336731, + 0.01779947429895401, + 0.012856039218604565, + 0.0072005148977041245, + 0.011299517937004566, + 0.010646418668329716, + 0.01008338201791048, + 0.012498564086854458, + 0.006306412164121866, + 0.008112253621220589, + 0.009833230637013912, + 0.0037991609424352646, + 0.009708818048238754, + 0.006402930244803429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.08903945982456207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20212092995643616, + 0.16655533015727997, + 0.10673051327466965, + 0.06989452242851257, + 0.0565849207341671, + 0.06702485680580139, + 0.04402575269341469, + 0.10356016457080841, + 0.09636923670768738, + 0.0871337279677391 + ], + "credit_map": [ + 0.019376546144485474, + 0.017124107107520103, + 0.025099696591496468, + 0.02337280847132206, + 0.028001612052321434, + 0.020257074385881424, + 0.04424380883574486, + 0.024645278230309486, + 0.015148746781051159, + 0.015215011313557625, + 0.01625530607998371, + 0.017138192430138588, + 0.04446244612336159, + 0.021073857322335243, + 0.015642564743757248, + 0.021619202569127083, + 0.009586978703737259, + 0.014311939477920532, + 0.01270012091845274, + 0.010581456124782562, + 0.013730662874877453, + 0.015736112371087074, + 0.019757257774472237, + 0.010325982235372066, + 0.008180052973330021, + 0.008465570397675037, + 0.007451918441802263, + 0.007760075386613607, + 0.008597783744335175, + 0.009135758504271507, + 0.011281565763056278, + 0.009021797217428684, + 0.006490899249911308, + 0.00728872325271368, + 0.00695682130753994, + 0.005589044652879238, + 0.006672502961009741, + 0.010643655434250832, + 0.005747759249061346, + 0.007195513229817152, + 0.007442592177540064, + 0.007754276506602764, + 0.005607491824775934, + 0.00957611482590437, + 0.007088324520736933, + 0.012717545963823795, + 0.009375040419399738, + 0.007463469635695219, + 0.006064889021217823, + 0.009896254166960716, + 0.0036677205935120583, + 0.003920860588550568, + 0.0058814589865505695, + 0.006506997160613537, + 0.0023592037614434958, + 0.005728370044380426, + 0.014193100854754448, + 0.02399560436606407, + 0.007885761559009552, + 0.010996614582836628, + 0.01661440171301365, + 0.011597719974815845, + 0.009422252885997295, + 0.008854716084897518, + 0.013252361677587032, + 0.02240157686173916, + 0.008264763280749321, + 0.014145498163998127, + 0.011520517989993095, + 0.006232272833585739, + 0.009982137940824032, + 0.010570110753178596, + 0.01189116109162569, + 0.01196877658367157, + 0.009915318340063095, + 0.006190756801515818, + 0.014925284311175346, + 0.0038849245756864548, + 0.017575636506080627, + 0.010781871154904366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00978010892868042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154929161071777 + ], + "opposite_logprob_tokens": [ + -1.7015492916107178 + ], + "logprob_reward": -0.20154929161071777, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154929161071777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15367542207241058, + 0.04133369401097298, + 0.02537546306848526, + 0.03413863852620125, + 0.06738679111003876, + 0.024262575432658195, + 0.03641250357031822, + 0.36396411061286926, + 0.14186809957027435, + 0.11158275604248047 + ], + "credit_map": [ + 0.012718671932816505, + 0.015313733369112015, + 0.010398855432868004, + 0.017347216606140137, + 0.023251645267009735, + 0.016806330531835556, + 0.03787587210536003, + 0.019963089376688004, + 0.004186792299151421, + 0.005166205111891031, + 0.0048374272882938385, + 0.005530558992177248, + 0.006766144651919603, + 0.005072584375739098, + 0.005018690600991249, + 0.004755286034196615, + 0.002662032376974821, + 0.003259674645960331, + 0.002800842747092247, + 0.00232810340821743, + 0.0044105867855250835, + 0.003696439089253545, + 0.003693193895742297, + 0.002524586860090494, + 0.004155323840677738, + 0.0050596739165484905, + 0.004744995851069689, + 0.004147043451666832, + 0.004482461139559746, + 0.003977058921009302, + 0.003957450855523348, + 0.0036146275233477354, + 0.0068840691819787025, + 0.008680727332830429, + 0.009250032715499401, + 0.00934840738773346, + 0.007591546047478914, + 0.00595231493934989, + 0.010350503027439117, + 0.009329189546406269, + 0.003740851068869233, + 0.0028996530454605818, + 0.002046223497018218, + 0.002926543587818742, + 0.0031345789320766926, + 0.00481654517352581, + 0.0023034438490867615, + 0.0023947330191731453, + 0.004252293612807989, + 0.0072207399643957615, + 0.004367330111563206, + 0.004503405187278986, + 0.004251809325069189, + 0.003792740171775222, + 0.003358452348038554, + 0.004665727727115154, + 0.12387476861476898, + 0.027091458439826965, + 0.012289565987884998, + 0.014756903983652592, + 0.13341617584228516, + 0.021602584049105644, + 0.014980311505496502, + 0.015952300280332565, + 0.018437618389725685, + 0.02565082721412182, + 0.014840848743915558, + 0.023464776575565338, + 0.018569884821772575, + 0.010630590841174126, + 0.016674214974045753, + 0.013599332422018051, + 0.014293413609266281, + 0.017304478213191032, + 0.010923173278570175, + 0.015364307910203934, + 0.018076667562127113, + 0.006876847241073847, + 0.015920352190732956, + 0.012823502533137798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.20154929161071777 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817583292722702 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.01817583292722702, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01817583292722702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06970509886741638, + 0.06697764992713928, + 0.04406357556581497, + 0.032985009253025055, + 0.0237114280462265, + 0.03723405301570892, + 0.046331897377967834, + 0.4934697449207306, + 0.12377195805311203, + 0.06174954026937485 + ], + "credit_map": [ + 0.011640191078186035, + 0.007840141654014587, + 0.008153504692018032, + 0.009657212533056736, + 0.011772357858717442, + 0.0055653806775808334, + 0.008703126572072506, + 0.00637319078668952, + 0.005621721036732197, + 0.004788090009242296, + 0.009278799407184124, + 0.005786147899925709, + 0.017303505912423134, + 0.007178404834121466, + 0.0062484429217875, + 0.01077253744006157, + 0.004807945340871811, + 0.005766524933278561, + 0.006569729186594486, + 0.0036660931073129177, + 0.005168164614588022, + 0.006784338504076004, + 0.00711231492459774, + 0.0041884686797857285, + 0.0033344116527587175, + 0.004875548183917999, + 0.0038766595534980297, + 0.003210043767467141, + 0.005478410981595516, + 0.004158121533691883, + 0.003957373555749655, + 0.004094438627362251, + 0.0026863880921155214, + 0.0027784849517047405, + 0.0027556028217077255, + 0.00227416236884892, + 0.002518867375329137, + 0.00435797031968832, + 0.0029258111026138067, + 0.0034141431096941233, + 0.0059132883325219154, + 0.004551635589450598, + 0.003932633902877569, + 0.003459782339632511, + 0.0039778705686330795, + 0.007490493357181549, + 0.0039421189576387405, + 0.00396623183041811, + 0.005653509870171547, + 0.007387513294816017, + 0.0032256480772048235, + 0.005143245682120323, + 0.006534208543598652, + 0.007109363563358784, + 0.004130031447857618, + 0.0071483817882835865, + 0.17220614850521088, + 0.02517125755548477, + 0.017257649451494217, + 0.011212160810828209, + 0.18451473116874695, + 0.03839171305298805, + 0.01888980157673359, + 0.02582632750272751, + 0.01926649734377861, + 0.017563920468091965, + 0.011349902488291264, + 0.017235420644283295, + 0.012865670025348663, + 0.01595356874167919, + 0.012035034596920013, + 0.017501946538686752, + 0.010660378262400627, + 0.01741897687315941, + 0.006577524822205305, + 0.005417034029960632, + 0.0054718744941055775, + 0.0027985305059701204, + 0.005510539282113314, + 0.007894685491919518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.01817583292722702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, upward tracking shot follows, revealing a massive, cyan dragon, as vast as a mountain. Its sinuous, serpentine form is covered in countless shield-sized scales, each shimmering in the light with iridescent hues ranging from jade to emerald. With a rhythm imbued with power and unparalleled grace, the dragon weaves its way through the swirling sea of clouds below. Like boiling milk, the sea of clouds forms countless towering peaks and deep canyons, through which the dragon deftly weaves, its flowing mane like white silk and long whiskers dancing in the currents. Its majestic head, crowned by a pair of towering horns reaching toward the sky, gazes forward from amber eyes blazing with ancient wisdom. Hidden beneath the clouds, the sun casts thousands of golden rays (god rays) upward, piercing through the gaps and casting a dazzling outline of light around the dragon's belly and the edge of the sea of clouds. With each leisurely swing of the animal, some of its scales instantly reflect dazzling light. The entire picture is filled with mythical grandeur and a quiet and powerful sense of power, creating an awe-inspiring epic atmosphere.", + "seed": 21465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_1808_seed_21465_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06653725560754538, + "vlm_logprob": -0.06653725560754538, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06653725560754538, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01083078421652317 + ], + "opposite_logprob_tokens": [ + -4.635830879211426 + ], + "logprob_reward": -0.01083078421652317, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01083078421652317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11165390908718109, + 0.18230217695236206, + 0.020155947655439377, + 0.043148841708898544, + 0.030896082520484924, + 0.16805003583431244, + 0.03977268189191818, + 0.14265601336956024, + 0.11094660311937332, + 0.15041768550872803 + ], + "credit_map": [ + 0.011206043884158134, + 0.017524268478155136, + 0.01398997101932764, + 0.01365953404456377, + 0.01229859422892332, + 0.014272449538111687, + 0.014373727142810822, + 0.014329330995678902, + 0.0082569420337677, + 0.01349748857319355, + 0.009867308661341667, + 0.011949125677347183, + 0.017922915518283844, + 0.015139954164624214, + 0.050266750156879425, + 0.055401697754859924, + 0.0013766176998615265, + 0.0020398490596562624, + 0.002413997193798423, + 0.003253209637477994, + 0.003235433716326952, + 0.0019507744582369924, + 0.0032539451494812965, + 0.0026321234181523323, + 0.004468202590942383, + 0.003771733958274126, + 0.0034924352075904608, + 0.003939903806895018, + 0.004847354255616665, + 0.010296097956597805, + 0.003879538271576166, + 0.008453582413494587, + 0.003154605394229293, + 0.0055108689703047276, + 0.0031797196716070175, + 0.0027599160093814135, + 0.004899648949503899, + 0.0040718140080571175, + 0.003569650463759899, + 0.0037498618476092815, + 0.007643681485205889, + 0.00914154015481472, + 0.010792034678161144, + 0.03947269544005394, + 0.03727858513593674, + 0.03195597231388092, + 0.01586940325796604, + 0.015896135941147804, + 0.004841587971895933, + 0.005259078461676836, + 0.0037152583245187998, + 0.005051256623119116, + 0.004909731913357973, + 0.004913589917123318, + 0.0056306435726583, + 0.0054515390656888485, + 0.011396635323762894, + 0.028442328795790672, + 0.008996852673590183, + 0.01574861817061901, + 0.03218723088502884, + 0.01965082809329033, + 0.014021345414221287, + 0.012212185189127922, + 0.01400409359484911, + 0.017095021903514862, + 0.008887074887752533, + 0.01393873617053032, + 0.018983062356710434, + 0.015009908005595207, + 0.01199527271091938, + 0.01103343814611435, + 0.019083499908447266, + 0.023593775928020477, + 0.01241513155400753, + 0.01601262390613556, + 0.034236736595630646, + 0.020665258169174194, + 0.009479577653110027, + 0.01493110228329897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.01083078421652317 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037937998771667 + ], + "opposite_logprob_tokens": [ + -1.910379409790039 + ], + "logprob_reward": -0.16037937998771667, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037937998771667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1054440513253212, + 0.3205905258655548, + 0.020528553053736687, + 0.035220127552747726, + 0.026641875505447388, + 0.2052658051252365, + 0.040504340082407, + 0.038567934185266495, + 0.14878523349761963, + 0.058451615273952484 + ], + "credit_map": [ + 0.010946596972644329, + 0.01593237742781639, + 0.014440014027059078, + 0.012073338963091373, + 0.011438773013651371, + 0.015224132686853409, + 0.013596814125776291, + 0.011791999451816082, + 0.01700161024928093, + 0.02693963423371315, + 0.022027555853128433, + 0.02171139046549797, + 0.03517574816942215, + 0.018197473138570786, + 0.08779030293226242, + 0.09174678474664688, + 0.001579665462486446, + 0.0023619341664016247, + 0.0021447816397994757, + 0.0036356542259454727, + 0.002479210728779435, + 0.0017500948160886765, + 0.00434182258322835, + 0.002235389081761241, + 0.0033942246809601784, + 0.004340318031609058, + 0.003813367336988449, + 0.002616217592731118, + 0.003626315388828516, + 0.00722659332677722, + 0.0034285353031009436, + 0.006774554960429668, + 0.002524167997762561, + 0.0038489121943712234, + 0.0029305145144462585, + 0.002168883802369237, + 0.004395138006657362, + 0.004653254523873329, + 0.003372494364157319, + 0.002748510567471385, + 0.010585318319499493, + 0.014495152980089188, + 0.016180165112018585, + 0.03869755566120148, + 0.03549950569868088, + 0.041772931814193726, + 0.021413259208202362, + 0.02662191353738308, + 0.004899141378700733, + 0.005469701252877712, + 0.004393687471747398, + 0.0051966095343232155, + 0.0034754762891680002, + 0.005235862918198109, + 0.005093410145491362, + 0.006740449462085962, + 0.0029041513334959745, + 0.004836916457861662, + 0.0036642667837440968, + 0.0030618691816926003, + 0.00781876128166914, + 0.005000812467187643, + 0.006194251123815775, + 0.005086907651275396, + 0.0212229173630476, + 0.015496335923671722, + 0.01693429984152317, + 0.019176309928297997, + 0.020319882780313492, + 0.022198330610990524, + 0.014988427050411701, + 0.0184487234801054, + 0.007099578157067299, + 0.008336097933351994, + 0.005102219060063362, + 0.0058355554938316345, + 0.011357506737112999, + 0.008347618393599987, + 0.004445492755621672, + 0.00792754627764225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.16037937998771667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637320525944233 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637320525944233, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637320525944233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2139544039964676, + 0.17364127933979034, + 0.022868525236845016, + 0.05387173220515251, + 0.030438099056482315, + 0.11567661911249161, + 0.04501613602042198, + 0.11690407991409302, + 0.11469532549381256, + 0.11293382197618484 + ], + "credit_map": [ + 0.01904858835041523, + 0.030124755576252937, + 0.028100280091166496, + 0.03431951254606247, + 0.02306356653571129, + 0.03041841834783554, + 0.022471031174063683, + 0.026408260688185692, + 0.007648160215467215, + 0.01599900610744953, + 0.015205246396362782, + 0.01589035987854004, + 0.02124592661857605, + 0.010683523491024971, + 0.04331114888191223, + 0.04365790635347366, + 0.002334393560886383, + 0.002528670709580183, + 0.0031869641970843077, + 0.0029300465248525143, + 0.002623426727950573, + 0.0020797636825591326, + 0.00464177131652832, + 0.0025434878189116716, + 0.003885323414579034, + 0.004801169969141483, + 0.005933928769081831, + 0.0053130341693758965, + 0.007563555613160133, + 0.010819416493177414, + 0.005954849999397993, + 0.009600452147424221, + 0.003128595417365432, + 0.002777520101517439, + 0.002499758033081889, + 0.003847994841635227, + 0.004098109435290098, + 0.006037282757461071, + 0.003785755019634962, + 0.004263084381818771, + 0.005237907636910677, + 0.006495453417301178, + 0.007081035058945417, + 0.030256908386945724, + 0.029853900894522667, + 0.018216384574770927, + 0.009684537537395954, + 0.008850486017763615, + 0.0035934008192270994, + 0.0038593586068600416, + 0.005929690785706043, + 0.004767162259668112, + 0.005026053637266159, + 0.005577078554779291, + 0.008939383551478386, + 0.007324006408452988, + 0.00892089493572712, + 0.023957261815667152, + 0.00744288694113493, + 0.01211952231824398, + 0.025981266051530838, + 0.014147690497338772, + 0.01293649896979332, + 0.011398054659366608, + 0.01740252412855625, + 0.014600472524762154, + 0.01001282874494791, + 0.009869523346424103, + 0.012332255020737648, + 0.022432787343859673, + 0.014587107114493847, + 0.013457836583256721, + 0.013384328223764896, + 0.017060091719031334, + 0.01163568627089262, + 0.011757905595004559, + 0.02283264696598053, + 0.016336163505911827, + 0.007388909813016653, + 0.012538079172372818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004637320525944233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704972922801971 + ], + "opposite_logprob_tokens": [ + -2.127049684524536 + ], + "logprob_reward": -0.12704972922801971, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704972922801971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04487275704741478, + 0.1656462401151657, + 0.02997315302491188, + 0.037778206169605255, + 0.028994828462600708, + 0.15529079735279083, + 0.06727855652570724, + 0.15611955523490906, + 0.12767548859119415, + 0.18637043237686157 + ], + "credit_map": [ + 0.0037457363214343786, + 0.006483479402959347, + 0.0061704786494374275, + 0.006931856740266085, + 0.005170530639588833, + 0.006112905219197273, + 0.004927828907966614, + 0.005329941399395466, + 0.016545554623007774, + 0.030760712921619415, + 0.01908908598124981, + 0.02022923342883587, + 0.02685634419322014, + 0.022400468587875366, + 0.01235022209584713, + 0.017414618283510208, + 0.0019577261991798878, + 0.0028223239351063967, + 0.003602905198931694, + 0.0037385234609246254, + 0.0038578102830797434, + 0.003014832502231002, + 0.0077285487204790115, + 0.0032504817936569452, + 0.004383948165923357, + 0.003568634856492281, + 0.0026643117889761925, + 0.0037839505821466446, + 0.003910736180841923, + 0.008640710264444351, + 0.0029545605648308992, + 0.007871352136135101, + 0.0030762110836803913, + 0.004048568196594715, + 0.0030252321157604456, + 0.0023836533073335886, + 0.005095168948173523, + 0.005421098321676254, + 0.002621196210384369, + 0.0033237014431506395, + 0.008081505075097084, + 0.010665420442819595, + 0.01106894202530384, + 0.03544948622584343, + 0.03462137654423714, + 0.025712482631206512, + 0.01479037944227457, + 0.014901206828653812, + 0.005574357695877552, + 0.009304440580308437, + 0.008582981303334236, + 0.007469198666512966, + 0.00736108236014843, + 0.007447042036801577, + 0.011211029253900051, + 0.01032841857522726, + 0.011921563185751438, + 0.027159765362739563, + 0.010332686826586723, + 0.01439792849123478, + 0.03592180833220482, + 0.020076844841241837, + 0.019684189930558205, + 0.01662476547062397, + 0.018048886209726334, + 0.017511121928691864, + 0.011289930902421474, + 0.01407241728156805, + 0.022356444969773293, + 0.019454995170235634, + 0.014302297495305538, + 0.010639406740665436, + 0.022592028602957726, + 0.03152485936880112, + 0.02351928874850273, + 0.023507922887802124, + 0.030880916863679886, + 0.02533910609781742, + 0.009930749423801899, + 0.019075559452176094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.12704972922801971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789064079523087 + ], + "opposite_logprob_tokens": [ + -3.5297889709472656 + ], + "logprob_reward": -0.029789064079523087, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789064079523087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04706069827079773, + 0.37830686569213867, + 0.028818858787417412, + 0.0335116907954216, + 0.03415890038013458, + 0.04283903166651726, + 0.031516727060079575, + 0.038422588258981705, + 0.31059548258781433, + 0.05476915091276169 + ], + "credit_map": [ + 0.0044579096138477325, + 0.006841122172772884, + 0.006055925972759724, + 0.005228128284215927, + 0.0049768127501010895, + 0.005999458953738213, + 0.007097492925822735, + 0.006403848994523287, + 0.01740312948822975, + 0.027787426486611366, + 0.024294966831803322, + 0.024895548820495605, + 0.034928180277347565, + 0.03150125592947006, + 0.10491903126239777, + 0.11257731914520264, + 0.0024587581865489483, + 0.0026437335181981325, + 0.003907233010977507, + 0.00411609560251236, + 0.003153362311422825, + 0.0024336257483810186, + 0.0073254723101854324, + 0.002780575305223465, + 0.0043558343313634396, + 0.004224523436278105, + 0.00294732884503901, + 0.0025033385027199984, + 0.0027331295423209667, + 0.006326580420136452, + 0.0029416335746645927, + 0.007479317486286163, + 0.0022506604436784983, + 0.0029366181697696447, + 0.004395117051899433, + 0.005048686172813177, + 0.0029373490251600742, + 0.0058732046745717525, + 0.006297542713582516, + 0.004419720731675625, + 0.0023575189989060163, + 0.00283029000274837, + 0.0032875407487154007, + 0.008815436623990536, + 0.007557184435427189, + 0.009159012697637081, + 0.004529175814241171, + 0.004302872344851494, + 0.0024891337379813194, + 0.004070321563631296, + 0.0036756277550011873, + 0.004301362670958042, + 0.003920040093362331, + 0.005053351167589426, + 0.004882665351033211, + 0.00312422146089375, + 0.002783329924568534, + 0.004430901724845171, + 0.003402718808501959, + 0.004153385292738676, + 0.006962561048567295, + 0.005731855984777212, + 0.00477901054546237, + 0.006178822368383408, + 0.05719098076224327, + 0.031387124210596085, + 0.03177615627646446, + 0.02746717818081379, + 0.0342339463531971, + 0.05686794966459274, + 0.02816886454820633, + 0.04350326582789421, + 0.005862497724592686, + 0.0061026946641504765, + 0.0072013502940535545, + 0.008203316479921341, + 0.007811118848621845, + 0.008162322454154491, + 0.0049970317631959915, + 0.006428814493119717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.029789064079523087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, upward tracking shot follows, revealing a massive, cyan dragon, as vast as a mountain. Its sinuous, serpentine form is covered in countless shield-sized scales, each shimmering in the light with iridescent hues ranging from jade to emerald. With a rhythm imbued with power and unparalleled grace, the dragon weaves its way through the swirling sea of clouds below. Like boiling milk, the sea of clouds forms countless towering peaks and deep canyons, through which the dragon deftly weaves, its flowing mane like white silk and long whiskers dancing in the currents. Its majestic head, crowned by a pair of towering horns reaching toward the sky, gazes forward from amber eyes blazing with ancient wisdom. Hidden beneath the clouds, the sun casts thousands of golden rays (god rays) upward, piercing through the gaps and casting a dazzling outline of light around the dragon's belly and the edge of the sea of clouds. With each leisurely swing of the animal, some of its scales instantly reflect dazzling light. The entire picture is filled with mythical grandeur and a quiet and powerful sense of power, creating an awe-inspiring epic atmosphere.", + "seed": 21466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_1808_seed_21466_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06630246993154287, + "vlm_logprob": -0.06630246993154287, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.06630246993154287, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015248040668666363 + ], + "opposite_logprob_tokens": [ + -4.265247821807861 + ], + "logprob_reward": -0.015248040668666363, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015248040668666363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1413659006357193, + 0.06920446455478668, + 0.14246834814548492, + 0.06337016820907593, + 0.045803021639585495, + 0.18814009428024292, + 0.08025802671909332, + 0.12669071555137634, + 0.08826043456792831, + 0.054438892751932144 + ], + "credit_map": [ + 0.018729794770479202, + 0.013536712154746056, + 0.015785012394189835, + 0.021280674263834953, + 0.016517847776412964, + 0.024177610874176025, + 0.010228476487100124, + 0.021109767258167267, + 0.007906273007392883, + 0.00971144437789917, + 0.0073523083701729774, + 0.006844308692961931, + 0.010778706520795822, + 0.008613546378910542, + 0.007729574106633663, + 0.010268304497003555, + 0.011355203576385975, + 0.01597331278026104, + 0.011288504116237164, + 0.011873941868543625, + 0.02472108043730259, + 0.02838139422237873, + 0.016717124730348587, + 0.022157777100801468, + 0.005491101182997227, + 0.005211944226175547, + 0.006156847346574068, + 0.005362283438444138, + 0.006488938350230455, + 0.01340414211153984, + 0.01080033928155899, + 0.010454573668539524, + 0.004045644775032997, + 0.006309645250439644, + 0.0058316574431955814, + 0.00630456767976284, + 0.004716064780950546, + 0.006056005600839853, + 0.005211696028709412, + 0.0073277391493320465, + 0.017574584111571312, + 0.034237634390592575, + 0.0148932421579957, + 0.018880868330597878, + 0.020085547119379044, + 0.02082451619207859, + 0.030299924314022064, + 0.031343765556812286, + 0.007277366705238819, + 0.013144063763320446, + 0.00616235937923193, + 0.006458314601331949, + 0.021028708666563034, + 0.008903955109417439, + 0.01015082374215126, + 0.007132439408451319, + 0.013763130642473698, + 0.015703847631812096, + 0.014086544513702393, + 0.012591835111379623, + 0.018631653860211372, + 0.023858683183789253, + 0.013798056170344353, + 0.01425696536898613, + 0.009379924274981022, + 0.011819986626505852, + 0.00604447303339839, + 0.006489051505923271, + 0.014704173430800438, + 0.022634556517004967, + 0.006845016963779926, + 0.010343254543840885, + 0.008255607448518276, + 0.013677123934030533, + 0.005568881519138813, + 0.0037804057355970144, + 0.006335543934255838, + 0.008039586246013641, + 0.004673157352954149, + 0.004108589608222246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.015248040668666363 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560948133468628 + ], + "opposite_logprob_tokens": [ + -1.600609540939331 + ], + "logprob_reward": -0.22560948133468628, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560948133468628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1278676986694336, + 0.1088385209441185, + 0.1645175963640213, + 0.14466552436351776, + 0.07095454633235931, + 0.0550592802464962, + 0.050904277712106705, + 0.07962916791439056, + 0.04242799058556557, + 0.1551353931427002 + ], + "credit_map": [ + 0.016982577741146088, + 0.015329359099268913, + 0.013610605150461197, + 0.01981930434703827, + 0.01483211014419794, + 0.020664626732468605, + 0.009453793056309223, + 0.017175322398543358, + 0.011812213808298111, + 0.014683960005640984, + 0.014020304195582867, + 0.010767498053610325, + 0.01557687483727932, + 0.01382709015160799, + 0.01207807194441557, + 0.01607251539826393, + 0.010823976248502731, + 0.019361980259418488, + 0.01340482011437416, + 0.014232169836759567, + 0.030979659408330917, + 0.027009649202227592, + 0.01956438273191452, + 0.02914096787571907, + 0.010280157439410686, + 0.010508924722671509, + 0.012304559350013733, + 0.010067496448755264, + 0.013556456193327904, + 0.03445027396082878, + 0.02684698812663555, + 0.026650678366422653, + 0.005429331678897142, + 0.006139552686363459, + 0.006504415534436703, + 0.0073728845454752445, + 0.006300953682512045, + 0.010097368620336056, + 0.010137993842363358, + 0.01897205226123333, + 0.006168186664581299, + 0.009928521700203419, + 0.003839402226731181, + 0.0055951620452106, + 0.006623655091971159, + 0.006155007053166628, + 0.008926006965339184, + 0.007823344320058823, + 0.005227102432399988, + 0.007021029945462942, + 0.005650906357914209, + 0.0051648905500769615, + 0.011440912261605263, + 0.00691424822434783, + 0.0050942194648087025, + 0.004390973597764969, + 0.008234577253460884, + 0.01016802154481411, + 0.008878570050001144, + 0.0081631513312459, + 0.011778527870774269, + 0.014766289852559566, + 0.00857428926974535, + 0.009065745398402214, + 0.004770047962665558, + 0.005813506431877613, + 0.004755283705890179, + 0.003948845434933901, + 0.005353162530809641, + 0.007665104698389769, + 0.003784611588343978, + 0.006337427534162998, + 0.012478295713663101, + 0.059479907155036926, + 0.006338008213788271, + 0.005206836387515068, + 0.04798586294054985, + 0.011723009869456291, + 0.005545909516513348, + 0.006377561017870903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.22560948133468628 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539591640233994 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539591640233994, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539591640233994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09499506652355194, + 0.08911361545324326, + 0.12095934897661209, + 0.07809194922447205, + 0.040294233709573746, + 0.20971347391605377, + 0.07521717995405197, + 0.09336208552122116, + 0.06781631708145142, + 0.1304367631673813 + ], + "credit_map": [ + 0.011440041474997997, + 0.007005327381193638, + 0.015146180987358093, + 0.014573152177035809, + 0.006584076210856438, + 0.013288463465869427, + 0.007848513312637806, + 0.01910931058228016, + 0.011868434026837349, + 0.010901959612965584, + 0.0078015257604420185, + 0.010923118330538273, + 0.01221717894077301, + 0.009233388118445873, + 0.013421346433460712, + 0.012746665626764297, + 0.007994269020855427, + 0.013388088904321194, + 0.00879664160311222, + 0.011268364265561104, + 0.02640928328037262, + 0.022075941786170006, + 0.014599102549254894, + 0.016427669674158096, + 0.006441854406148195, + 0.007294543087482452, + 0.0077591934241354465, + 0.006223756354302168, + 0.009425381198525429, + 0.014694374985992908, + 0.01371982041746378, + 0.012533023953437805, + 0.0034411840606480837, + 0.0041688731871545315, + 0.004050237126648426, + 0.004613552708178759, + 0.003914497327059507, + 0.004587220493704081, + 0.0071878633461892605, + 0.008330806158483028, + 0.02483944036066532, + 0.025128070265054703, + 0.012744506821036339, + 0.0212580356746912, + 0.023906687274575233, + 0.0333063118159771, + 0.030633771792054176, + 0.03789665177464485, + 0.008694303222000599, + 0.010534518398344517, + 0.007488981820642948, + 0.007784344255924225, + 0.01213316060602665, + 0.0075505580753088, + 0.012653300538659096, + 0.008378013037145138, + 0.010144546627998352, + 0.010906530544161797, + 0.010199807584285736, + 0.010250506922602654, + 0.013967511244118214, + 0.017183156684041023, + 0.010965251363813877, + 0.00974477268755436, + 0.006848864257335663, + 0.009163599461317062, + 0.005211314652115107, + 0.005887113511562347, + 0.010650829412043095, + 0.015262416563928127, + 0.005574109964072704, + 0.009218075312674046, + 0.017925001680850983, + 0.0442507341504097, + 0.008077471517026424, + 0.00786006823182106, + 0.02483716607093811, + 0.012045801617205143, + 0.005841987207531929, + 0.009598538279533386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.012539591640233994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062053438276052475 + ], + "opposite_logprob_tokens": [ + -2.8120534420013428 + ], + "logprob_reward": -0.062053438276052475, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062053438276052475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09443401545286179, + 0.04381166025996208, + 0.12655527889728546, + 0.0512901172041893, + 0.06598293781280518, + 0.1210540309548378, + 0.14261890947818756, + 0.16485247015953064, + 0.14017479121685028, + 0.049225807189941406 + ], + "credit_map": [ + 0.00932222232222557, + 0.009211313910782337, + 0.01361282728612423, + 0.019899485632777214, + 0.008371387608349323, + 0.011640136130154133, + 0.009338217787444592, + 0.01303842756897211, + 0.0046706427820026875, + 0.006847668904811144, + 0.003980920650064945, + 0.004529469646513462, + 0.008155998773872852, + 0.004774917848408222, + 0.004862810019403696, + 0.005989231169223785, + 0.009167972020804882, + 0.014092200435698032, + 0.009735672734677792, + 0.01345113292336464, + 0.027086762711405754, + 0.01826060377061367, + 0.0157844889909029, + 0.01897646114230156, + 0.0043468596413731575, + 0.006311831530183554, + 0.006190930027514696, + 0.006200722418725491, + 0.008282343856990337, + 0.007876478135585785, + 0.006430502515286207, + 0.005650447681546211, + 0.005192067939788103, + 0.005732374731451273, + 0.006144125945866108, + 0.007688035257160664, + 0.004324715584516525, + 0.00894958060234785, + 0.008761820383369923, + 0.019190214574337006, + 0.011499375104904175, + 0.021998651325702667, + 0.009134531952440739, + 0.012482431717216969, + 0.013063193298876286, + 0.012987226247787476, + 0.019813179969787598, + 0.02007543481886387, + 0.012918705120682716, + 0.0240386500954628, + 0.009935995563864708, + 0.01117271650582552, + 0.04320841655135155, + 0.01248315442353487, + 0.01826118491590023, + 0.010600093752145767, + 0.016379499807953835, + 0.021735433489084244, + 0.015620701014995575, + 0.017548032104969025, + 0.030704829841852188, + 0.0281984843313694, + 0.017534837126731873, + 0.0171306524425745, + 0.017842784523963928, + 0.019768372178077698, + 0.01024382933974266, + 0.007290130481123924, + 0.024384895339608192, + 0.03668368607759476, + 0.01082658302038908, + 0.01313451025635004, + 0.0050489590503275394, + 0.011187423020601273, + 0.0034389265347272158, + 0.0029557517264038324, + 0.008710558526217937, + 0.01171772088855505, + 0.002837101463228464, + 0.0033293645828962326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.062053438276052475 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061797738075256 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061797738075256, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061797738075256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040134891867637634, + 0.02956615947186947, + 0.07953798025846481, + 0.02936278097331524, + 0.05447082966566086, + 0.03235522657632828, + 0.03549866005778313, + 0.12778237462043762, + 0.17219418287277222, + 0.3990969657897949 + ], + "credit_map": [ + 0.003466368652880192, + 0.006518452428281307, + 0.004992946516722441, + 0.005582335405051708, + 0.005090156104415655, + 0.005103365983814001, + 0.004020814783871174, + 0.005360451061278582, + 0.003651684382930398, + 0.0033064119052141905, + 0.003113610902801156, + 0.003646811470389366, + 0.0054124267771840096, + 0.0031496891751885414, + 0.0038406553212553263, + 0.0034448690712451935, + 0.005406153853982687, + 0.009081666357815266, + 0.005513779819011688, + 0.007806527428328991, + 0.017051469534635544, + 0.014907908625900745, + 0.00988681335002184, + 0.009883659891784191, + 0.0022738720290362835, + 0.0032152377534657717, + 0.0034107316751033068, + 0.0029323361814022064, + 0.004862344358116388, + 0.004877799656242132, + 0.0034529825206846, + 0.0043374765664339066, + 0.004364720545709133, + 0.0036647776141762733, + 0.003985205199569464, + 0.005688689649105072, + 0.003496772376820445, + 0.008822867646813393, + 0.007851696573197842, + 0.016596099361777306, + 0.003174081677570939, + 0.003455136204138398, + 0.002274438738822937, + 0.0034820714499801397, + 0.00662417896091938, + 0.003142161527648568, + 0.0069657485000789165, + 0.0032374111469835043, + 0.0037453046534210443, + 0.005881079006940126, + 0.004465440288186073, + 0.0041006943210959435, + 0.005544294603168964, + 0.003878838149830699, + 0.00435106735676527, + 0.003531942144036293, + 0.0128558911383152, + 0.015836244449019432, + 0.012985922396183014, + 0.013351889327168465, + 0.024073125794529915, + 0.02111072465777397, + 0.015112041495740414, + 0.012456539086997509, + 0.018430907279253006, + 0.023237308487296104, + 0.018792908638715744, + 0.010291820392012596, + 0.025802161544561386, + 0.04439027979969978, + 0.0139799565076828, + 0.017268836498260498, + 0.0377209447324276, + 0.1702113300561905, + 0.009699154645204544, + 0.008803784847259521, + 0.13028746843338013, + 0.022086381912231445, + 0.012164059095084667, + 0.00812383834272623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.016061797738075256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, upward tracking shot follows, revealing a massive, cyan dragon, as vast as a mountain. Its sinuous, serpentine form is covered in countless shield-sized scales, each shimmering in the light with iridescent hues ranging from jade to emerald. With a rhythm imbued with power and unparalleled grace, the dragon weaves its way through the swirling sea of clouds below. Like boiling milk, the sea of clouds forms countless towering peaks and deep canyons, through which the dragon deftly weaves, its flowing mane like white silk and long whiskers dancing in the currents. Its majestic head, crowned by a pair of towering horns reaching toward the sky, gazes forward from amber eyes blazing with ancient wisdom. Hidden beneath the clouds, the sun casts thousands of golden rays (god rays) upward, piercing through the gaps and casting a dazzling outline of light around the dragon's belly and the edge of the sea of clouds. With each leisurely swing of the animal, some of its scales instantly reflect dazzling light. The entire picture is filled with mythical grandeur and a quiet and powerful sense of power, creating an awe-inspiring epic atmosphere.", + "seed": 21467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_6/index_1808_seed_21467_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08626641035079956, + "vlm_logprob": -0.08626641035079956, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.08626641035079956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03449733182787895 + ], + "opposite_logprob_tokens": [ + -3.4094972610473633 + ], + "logprob_reward": -0.03449733182787895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03449733182787895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07406067848205566, + 0.1277846097946167, + 0.23830704391002655, + 0.14370909333229065, + 0.020998479798436165, + 0.0842755138874054, + 0.1111549586057663, + 0.08795463293790817, + 0.03960385546088219, + 0.07215115427970886 + ], + "credit_map": [ + 0.011157725006341934, + 0.007564651779830456, + 0.007938999682664871, + 0.010084560140967369, + 0.011000070720911026, + 0.009029606357216835, + 0.012034119106829166, + 0.0052509428933262825, + 0.010467431508004665, + 0.013911476358771324, + 0.036622241139411926, + 0.008088292554020882, + 0.0076316120103001595, + 0.03833328187465668, + 0.00840968918055296, + 0.00432058610022068, + 0.011461677961051464, + 0.013302857056260109, + 0.01311380323022604, + 0.10189860314130783, + 0.013741358183324337, + 0.014535227790474892, + 0.018040731549263, + 0.05221278592944145, + 0.008678426966071129, + 0.006658970378339291, + 0.013403592631220818, + 0.05478785187005997, + 0.03045094944536686, + 0.01165539026260376, + 0.01080857403576374, + 0.007265336811542511, + 0.0021692768204957247, + 0.001646404038183391, + 0.0024509329814463854, + 0.004910420160740614, + 0.0017327464884147048, + 0.003587782848626375, + 0.002013992052525282, + 0.0024869232438504696, + 0.006474560126662254, + 0.008009888231754303, + 0.009230237454175949, + 0.02167387679219246, + 0.00894870888441801, + 0.012643847614526749, + 0.010832700878381729, + 0.006461694836616516, + 0.016229158267378807, + 0.009793156757950783, + 0.018601270392537117, + 0.01952933706343174, + 0.008663851767778397, + 0.019774921238422394, + 0.010869832709431648, + 0.007693429011851549, + 0.013242210261523724, + 0.011553824879229069, + 0.015433288179337978, + 0.010983925312757492, + 0.00975166354328394, + 0.009867380373179913, + 0.011457336135208607, + 0.00566499587148428, + 0.006275363732129335, + 0.00465696444734931, + 0.005436017643660307, + 0.0037458823062479496, + 0.005306144244968891, + 0.004448632709681988, + 0.006732895039021969, + 0.0030019504483789206, + 0.007317543029785156, + 0.008729202672839165, + 0.007537438068538904, + 0.009480878710746765, + 0.016673985868692398, + 0.010071895085275173, + 0.005832113791257143, + 0.006508094258606434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.03449733182787895 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129465103149414 + ], + "opposite_logprob_tokens": [ + -2.2379465103149414 + ], + "logprob_reward": -0.1129465103149414, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129465103149414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0737697184085846, + 0.15242980420589447, + 0.31939390301704407, + 0.04682549834251404, + 0.029284313321113586, + 0.04436350613832474, + 0.09696970880031586, + 0.13973815739154816, + 0.03501899540424347, + 0.06220640614628792 + ], + "credit_map": [ + 0.019284559413790703, + 0.007520556449890137, + 0.00743855582550168, + 0.00639043003320694, + 0.010005367919802666, + 0.008263736963272095, + 0.00956815481185913, + 0.0052983579225838184, + 0.010774904862046242, + 0.01221618615090847, + 0.04789797216653824, + 0.01140979677438736, + 0.008421246893703938, + 0.0493258498609066, + 0.00645184563472867, + 0.005932000931352377, + 0.014752939343452454, + 0.02373592182993889, + 0.010615761391818523, + 0.08776570111513138, + 0.125176340341568, + 0.021888654679059982, + 0.012131637893617153, + 0.023326970636844635, + 0.0032215884421020746, + 0.0030660254415124655, + 0.013468200340867043, + 0.015843158587813377, + 0.002923604566603899, + 0.002586694899946451, + 0.0029540855903178453, + 0.002762144198641181, + 0.0028785625472664833, + 0.002748137107118964, + 0.0037943206261843443, + 0.00831305980682373, + 0.002569260075688362, + 0.0026311285328119993, + 0.003980455920100212, + 0.0023693896364420652, + 0.005171606782823801, + 0.0044884453527629375, + 0.00521962670609355, + 0.004815379157662392, + 0.00492617953568697, + 0.010568034835159779, + 0.0051640672609210014, + 0.004010168369859457, + 0.012412824667990208, + 0.008680724538862705, + 0.01594921387732029, + 0.019547078758478165, + 0.007187634706497192, + 0.013094726949930191, + 0.01311564538627863, + 0.006981861777603626, + 0.021650904789566994, + 0.01626025140285492, + 0.019879236817359924, + 0.021326659247279167, + 0.012956362217664719, + 0.012930327095091343, + 0.024144215509295464, + 0.01059019286185503, + 0.005461372900754213, + 0.005682399496436119, + 0.003456168109551072, + 0.0027506579644978046, + 0.005200252868235111, + 0.0050929333083331585, + 0.003949054982513189, + 0.0034261515829712152, + 0.0052777249366045, + 0.007576408330351114, + 0.005780160892754793, + 0.009465409442782402, + 0.01391939539462328, + 0.00712546007707715, + 0.007040322292596102, + 0.006021527573466301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.1129465103149414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693257719278336 + ], + "opposite_logprob_tokens": [ + -3.408693313598633 + ], + "logprob_reward": -0.033693257719278336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693257719278336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09953764826059341, + 0.17337366938591003, + 0.10150307416915894, + 0.1440727412700653, + 0.043830323964357376, + 0.09975547343492508, + 0.0563385933637619, + 0.06921566277742386, + 0.11476164311170578, + 0.0976111963391304 + ], + "credit_map": [ + 0.01927223987877369, + 0.012035802938044071, + 0.010346034541726112, + 0.010565957054495811, + 0.01567195914685726, + 0.013490025885403156, + 0.010963925160467625, + 0.007191699463874102, + 0.01659863442182541, + 0.020967716351151466, + 0.04065917059779167, + 0.013459237292408943, + 0.016139574348926544, + 0.035754043608903885, + 0.02159690670669079, + 0.008198375813663006, + 0.004804648924618959, + 0.007217875681817532, + 0.003862605895847082, + 0.02979540079832077, + 0.037178315222263336, + 0.006199141964316368, + 0.005347427446395159, + 0.007097646594047546, + 0.009599791839718819, + 0.007044848520308733, + 0.01345237996429205, + 0.052596647292375565, + 0.025559408590197563, + 0.0160604827105999, + 0.011048149317502975, + 0.008711026050150394, + 0.0042192088440060616, + 0.004030564334243536, + 0.005927781108766794, + 0.012324610725045204, + 0.0039946842007339, + 0.00438487995415926, + 0.004950117319822311, + 0.003998477477580309, + 0.009554458782076836, + 0.007323551923036575, + 0.012517931871116161, + 0.020009605213999748, + 0.007925302721560001, + 0.02403196319937706, + 0.010802332311868668, + 0.007590329274535179, + 0.010036184452474117, + 0.004788230173289776, + 0.011040245182812214, + 0.006520195864140987, + 0.006364226341247559, + 0.006704750936478376, + 0.006298350170254707, + 0.004586412571370602, + 0.010109182447195053, + 0.00632561044767499, + 0.01139748003333807, + 0.009194801561534405, + 0.005832147318869829, + 0.00700657581910491, + 0.012483411468565464, + 0.006866447161883116, + 0.017084188759326935, + 0.0219081100076437, + 0.02368268556892872, + 0.011033241637051105, + 0.009714828804135323, + 0.014179138466715813, + 0.00903052743524313, + 0.008128920570015907, + 0.009040109813213348, + 0.011896240524947643, + 0.01137792319059372, + 0.012527351267635822, + 0.01869458146393299, + 0.015863588079810143, + 0.01056497823446989, + 0.007646424695849419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.033693257719278336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156049728393555 + ], + "opposite_logprob_tokens": [ + -1.7015604972839355 + ], + "logprob_reward": -0.20156049728393555, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156049728393555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046241819858551025, + 0.04691951721906662, + 0.2839306592941284, + 0.04500748962163925, + 0.022195009514689445, + 0.0628519132733345, + 0.14087297022342682, + 0.17360661923885345, + 0.0977293998003006, + 0.0806446373462677 + ], + "credit_map": [ + 0.0084588173776865, + 0.005161834415048361, + 0.004416339099407196, + 0.004570384975522757, + 0.006730642169713974, + 0.00482945004478097, + 0.008674200624227524, + 0.003400151850655675, + 0.005179993808269501, + 0.005887281149625778, + 0.010787887498736382, + 0.0036645284853875637, + 0.003250121371820569, + 0.01181777473539114, + 0.0038926820270717144, + 0.002439247677102685, + 0.01282505039125681, + 0.01378485094755888, + 0.014971488155424595, + 0.12214586138725281, + 0.04252595081925392, + 0.012818352319300175, + 0.020563405007123947, + 0.04429570585489273, + 0.003392312442883849, + 0.004088099580258131, + 0.0053804595954716206, + 0.011316114105284214, + 0.0048675378784537315, + 0.006531310733407736, + 0.005344942677766085, + 0.004086711909621954, + 0.0022545948158949614, + 0.0024447471369057894, + 0.003145703114569187, + 0.0041275122202932835, + 0.0022638305090367794, + 0.0023794264998286963, + 0.0030667188111692667, + 0.0025124764069914818, + 0.006198601331561804, + 0.004788106773048639, + 0.008236444555222988, + 0.013445197604596615, + 0.005074773449450731, + 0.01270751841366291, + 0.007532884832471609, + 0.004868391901254654, + 0.02019057422876358, + 0.011576609686017036, + 0.023300494998693466, + 0.02600584737956524, + 0.012643606401979923, + 0.02290290966629982, + 0.014034031890332699, + 0.010218892246484756, + 0.02391001023352146, + 0.02863646298646927, + 0.02356315590441227, + 0.022258779034018517, + 0.020984504371881485, + 0.017785394564270973, + 0.025129584595561028, + 0.011338715441524982, + 0.010477269068360329, + 0.012999688275158405, + 0.012936270795762539, + 0.007520787883549929, + 0.023146990686655045, + 0.01005486398935318, + 0.01433216780424118, + 0.006261365953832865, + 0.008007668890058994, + 0.009859493002295494, + 0.008574074134230614, + 0.009953103959560394, + 0.0202738456428051, + 0.009810378775000572, + 0.0071785589680075645, + 0.006987506989389658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.20156049728393555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive cyan dragon with a sinuous, serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales visible as shield-sized and shimmering with iridescent hues from jade to emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon weaving through a sea of clouds that forms towering peaks and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have a flowing white silk mane, long whiskers, and a head with towering horns and amber eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden god rays visible piercing through the clouds to cast a dazzling outline of light around the dragon's belly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863445460796356 + ], + "opposite_logprob_tokens": [ + -3.0486345291137695 + ], + "logprob_reward": -0.04863445460796356, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863445460796356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1077772006392479, + 0.03016107901930809, + 0.2234240472316742, + 0.14745578169822693, + 0.06636939197778702, + 0.04481608420610428, + 0.08588074892759323, + 0.10012329369783401, + 0.04888858273625374, + 0.1451037973165512 + ], + "credit_map": [ + 0.013206938281655312, + 0.011377159506082535, + 0.010386602021753788, + 0.016503971070051193, + 0.01281040534377098, + 0.012228075414896011, + 0.02368842624127865, + 0.007575637195259333, + 0.003374130232259631, + 0.002970310626551509, + 0.007551959250122309, + 0.0030682566575706005, + 0.0026747025549411774, + 0.0057080271653831005, + 0.0028511674609035254, + 0.001962528796866536, + 0.01040581427514553, + 0.01670181006193161, + 0.01104852743446827, + 0.08846510946750641, + 0.06185551732778549, + 0.009626111947000027, + 0.011451087892055511, + 0.013870076276361942, + 0.008635207079350948, + 0.006739167030900717, + 0.012391380965709686, + 0.052879106253385544, + 0.03448174521327019, + 0.0129066938534379, + 0.011270073242485523, + 0.008152401074767113, + 0.006107292138040066, + 0.0057645863853394985, + 0.008506692945957184, + 0.01760081760585308, + 0.005395465064793825, + 0.00731729855760932, + 0.008657692931592464, + 0.0070195444859564304, + 0.004758501425385475, + 0.00478471489623189, + 0.005200654733926058, + 0.006753469351679087, + 0.0036195158027112484, + 0.010473719798028469, + 0.005307511426508427, + 0.003917999565601349, + 0.013793371617794037, + 0.007574254646897316, + 0.013041696511209011, + 0.014882714487612247, + 0.006436418741941452, + 0.017061153426766396, + 0.007271008566021919, + 0.005820134654641151, + 0.015225891023874283, + 0.009993307292461395, + 0.01154466811567545, + 0.016050070524215698, + 0.008919903077185154, + 0.008957565762102604, + 0.018340835347771645, + 0.011091059073805809, + 0.00777950044721365, + 0.009088038466870785, + 0.005963433999568224, + 0.0043409415520727634, + 0.00481981597840786, + 0.007901385426521301, + 0.004689461085945368, + 0.0043060071766376495, + 0.013896671123802662, + 0.018356414511799812, + 0.014348347671329975, + 0.01620766706764698, + 0.03250499442219734, + 0.020912526175379753, + 0.017144320532679558, + 0.01173285860568285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.04863445460796356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_tensor([8728, 8728])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_tensor([8728, 8728])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3d6c023f536d350dfae53dff3529fd1a082f97fa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_tensor([8728, 8728])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static shot establishes a scene of immense and terrifying pressure, focusing on a massive, hollow, rectangular steel beam that dominates the scene. The entire reality is drenched in a dark, sleek, and oppressive atmosphere, bathed in an unnatural, monochromatic green emanating from the hollow-like environment. The beam's gunmetal gray surface, with faint traces of industrial wear and tear, gleams with a damp, oily sheen, catching distorted reflections of an eerie green light. Under the force of an unseen crushing force, the beam begins to collapse; a low, resonant groan vibrates through the space as its once-perfect geometric form begins to visibly bend. The movement is depicted with terrifying slowness, forming a deep crease along its length as the flattened steel plates carefully fold inward. As the structural integrity is compromised, the entire view is severely marred—the image flickers with glaring white and damaged green, the scene fractured by digital glitches, screen tearing, and momentary freezes, as if the recording medium, or reality itself, is collapsing. Through this chaotic visual static, the beam’s collapse accelerates, its hollow form decisively flattening as metal screams and tears, catastrophically, before finally being crushed into a mangled, unrecognizable ruin in a shimmering green darkness.", + "seed": 3853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_8728_seed_3853_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4071218200959265, + "vlm_logprob": -0.4071218200959265, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.4071218200959265, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1407260149717331 + ], + "opposite_logprob_tokens": [ + -2.140726089477539 + ], + "logprob_reward": -0.1407260149717331, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1407260149717331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031525589525699615, + 0.008192995563149452, + 0.010846640914678574, + 0.05402025580406189, + 0.6495088934898376, + 0.11860257387161255, + 0.04117540270090103, + 0.02637353539466858, + 0.040839120745658875, + 0.018914928659796715 + ], + "credit_map": [ + 0.00340178981423378, + 0.004145469982177019, + 0.0055511812679469585, + 0.0041555967181921005, + 0.005577080883085728, + 0.0035554382484406233, + 0.0028920050244778395, + 0.00224702968262136, + 0.0009681404917500913, + 0.0010370892705395818, + 0.0014086633455008268, + 0.0008105225278995931, + 0.0015313309850171208, + 0.0009379390394315124, + 0.0009457430569455028, + 0.0005535682430490851, + 0.0010242856806144118, + 0.0010366070782765746, + 0.0017304345965385437, + 0.0011590728536248207, + 0.0023844020906835794, + 0.0013562286039814353, + 0.0010455215815454721, + 0.0011100891279056668, + 0.003310321131721139, + 0.016308775171637535, + 0.00689483480527997, + 0.0059162587858736515, + 0.0061091105453670025, + 0.005406759213656187, + 0.0038764982018619776, + 0.0061977035365998745, + 0.030230440199375153, + 0.02066168189048767, + 0.20672538876533508, + 0.023023050278425217, + 0.077557273209095, + 0.26056453585624695, + 0.00959234219044447, + 0.0211542546749115, + 0.00825272873044014, + 0.004951981361955404, + 0.031077897176146507, + 0.010530514642596245, + 0.010194813832640648, + 0.027748992666602135, + 0.011696741916239262, + 0.014148913323879242, + 0.0035754370037466288, + 0.0033239778131246567, + 0.006630412302911282, + 0.007552384398877621, + 0.0039032704662531614, + 0.006152533460408449, + 0.0049753496423363686, + 0.0050620404072105885, + 0.002627253532409668, + 0.00364074413664639, + 0.004027716815471649, + 0.003055833512917161, + 0.002785410964861512, + 0.004217839799821377, + 0.0029975567013025284, + 0.00302118225954473, + 0.002617263002321124, + 0.00217276974581182, + 0.013948735781013966, + 0.002172730164602399, + 0.0016388432122766972, + 0.01381959579885006, + 0.001731106429360807, + 0.0027380825486034155, + 0.0015036296099424362, + 0.002172773703932762, + 0.0023389840498566628, + 0.001806645537726581, + 0.002319978317245841, + 0.0030325374100357294, + 0.0033798234071582556, + 0.002360557671636343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.1407260149717331 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048622608184814 + ], + "opposite_logprob_tokens": [ + -0.17986227571964264 + ], + "logprob_reward": -1.8048622608184814, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048622608184814, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.020968398079276085, + 0.01010785810649395, + 0.006917260121554136, + 0.029185794293880463, + 0.7535404562950134, + 0.05346321687102318, + 0.02368951216340065, + 0.01704666204750538, + 0.06660643965005875, + 0.018474416807293892 + ], + "credit_map": [ + 0.003193903947249055, + 0.003484338754788041, + 0.0023905313573777676, + 0.0022272192873060703, + 0.004352509509772062, + 0.001860560616478324, + 0.002014661906287074, + 0.001444673747755587, + 0.0012555704452097416, + 0.0013379399897530675, + 0.0016671824268996716, + 0.0011635172413662076, + 0.001761854044161737, + 0.0010961246443912387, + 0.0008447804138995707, + 0.000980889773927629, + 0.0005109026096761227, + 0.0006853267550468445, + 0.0009797578677535057, + 0.0009466407354921103, + 0.0014389752177521586, + 0.0009149815887212753, + 0.0006040862645022571, + 0.0008365895482711494, + 0.0018682830268517137, + 0.009516457095742226, + 0.0029020216315984726, + 0.002283417386934161, + 0.0032297137659043074, + 0.002627864945679903, + 0.002213711617514491, + 0.004544327035546303, + 0.046389561146497726, + 0.025088202208280563, + 0.31664061546325684, + 0.030049579218029976, + 0.07153469324111938, + 0.2116888016462326, + 0.013509102165699005, + 0.03863992169499397, + 0.0049936650320887566, + 0.0036690356209874153, + 0.009912137873470783, + 0.008106546476483345, + 0.0058664982207119465, + 0.00883710291236639, + 0.007432322017848492, + 0.0046459101140499115, + 0.0029950402677059174, + 0.0021651005372405052, + 0.0034361272118985653, + 0.003643058007583022, + 0.002160297241061926, + 0.0038415666203945875, + 0.002137382747605443, + 0.00331094185821712, + 0.00143235397990793, + 0.003353514475747943, + 0.0022485831286758184, + 0.001951431157067418, + 0.002089116023853421, + 0.0026318177115172148, + 0.0015960633754730225, + 0.0017437838250771165, + 0.00469345785677433, + 0.002218968467786908, + 0.023146262392401695, + 0.004129193257540464, + 0.0029775388538837433, + 0.02331116795539856, + 0.0032565509900450706, + 0.0028733056969940662, + 0.0014404545072466135, + 0.0038579159881919622, + 0.0021011645440012217, + 0.002817669417709112, + 0.0024849469773471355, + 0.0016703061992302537, + 0.002324391156435013, + 0.0017775708111003041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -1.8048622608184814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048672039061784744 + ], + "opposite_logprob_tokens": [ + -3.0486719608306885 + ], + "logprob_reward": -0.048672039061784744, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048672039061784744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039933860301971436, + 0.017943117767572403, + 0.014745491556823254, + 0.05214470997452736, + 0.5356585383415222, + 0.15400123596191406, + 0.05298060178756714, + 0.024312570691108704, + 0.05608747899532318, + 0.052192434668540955 + ], + "credit_map": [ + 0.0030492511577904224, + 0.007169481832534075, + 0.0052168406546115875, + 0.003159580985084176, + 0.01112818531692028, + 0.0036723497323691845, + 0.003892368869856, + 0.0026458059437572956, + 0.001876550610177219, + 0.002669883891940117, + 0.0023412734735757113, + 0.0020663950126618147, + 0.004050917457789183, + 0.0019508703844621778, + 0.0018386595183983445, + 0.0011485692812129855, + 0.001932969898916781, + 0.0016475796001031995, + 0.002086435444653034, + 0.0012443573214113712, + 0.0037370421923696995, + 0.00153399130795151, + 0.001430205418728292, + 0.0011329104891046882, + 0.004101259633898735, + 0.007308466359972954, + 0.007021784316748381, + 0.0073845055885612965, + 0.009098338894546032, + 0.006119646597653627, + 0.006171570159494877, + 0.004939141683280468, + 0.01723574846982956, + 0.02944295108318329, + 0.05932697281241417, + 0.03285617381334305, + 0.1877843588590622, + 0.15090487897396088, + 0.030922984704375267, + 0.027184441685676575, + 0.01092613022774458, + 0.007850829511880875, + 0.042430657893419266, + 0.01612015813589096, + 0.012536682188510895, + 0.0353546142578125, + 0.01334515493363142, + 0.015437012538313866, + 0.0036600069142878056, + 0.003474364522844553, + 0.00980497244745493, + 0.008120634593069553, + 0.004239221569150686, + 0.011454172432422638, + 0.005223752930760384, + 0.007003475446254015, + 0.0022963471710681915, + 0.004064641892910004, + 0.0034363255836069584, + 0.003807518631219864, + 0.002303807297721505, + 0.0025009869132190943, + 0.00325774191878736, + 0.0026452031452208757, + 0.0029549922328442335, + 0.001924582407809794, + 0.01872018724679947, + 0.003994670230895281, + 0.00322496029548347, + 0.018822548910975456, + 0.002589157782495022, + 0.0038563820999115705, + 0.0039887866005301476, + 0.0071312543004751205, + 0.007011872716248035, + 0.0062435828149318695, + 0.008556789718568325, + 0.0065844422206282616, + 0.008122066035866737, + 0.004553643055260181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.048672039061784744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033717233687639236 + ], + "opposite_logprob_tokens": [ + -3.408717155456543 + ], + "logprob_reward": -0.033717233687639236, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033717233687639236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02677454613149166, + 0.020807886496186256, + 0.011373911052942276, + 0.17035531997680664, + 0.38798245787620544, + 0.15033367276191711, + 0.030091287568211555, + 0.02970118075609207, + 0.14612331986427307, + 0.026456383988261223 + ], + "credit_map": [ + 0.004472091794013977, + 0.0041909958235919476, + 0.002586382906883955, + 0.003311548847705126, + 0.005257775075733662, + 0.002597372280433774, + 0.0019612833857536316, + 0.002397099742665887, + 0.003846141742542386, + 0.0031809608917683363, + 0.003052742686122656, + 0.0016811270033940673, + 0.0030072627123445272, + 0.002211072714999318, + 0.0013682761928066611, + 0.002460302785038948, + 0.0009048603242263198, + 0.0013516555773094296, + 0.0015921617159619927, + 0.0014964502770453691, + 0.0029841859359294176, + 0.0010358403669670224, + 0.001110042561776936, + 0.0008987142355181277, + 0.009428531862795353, + 0.06241322308778763, + 0.020918671041727066, + 0.009926174767315388, + 0.018581433221697807, + 0.013387682847678661, + 0.010295247659087181, + 0.025404350832104683, + 0.020839780569076538, + 0.013488278724253178, + 0.15586574375629425, + 0.015612272545695305, + 0.1293240189552307, + 0.029858671128749847, + 0.008510426618158817, + 0.014483261853456497, + 0.011880051344633102, + 0.006290568504482508, + 0.03796057030558586, + 0.01609652489423752, + 0.012623896822333336, + 0.03367225453257561, + 0.016637258231639862, + 0.01517253927886486, + 0.0022903087083250284, + 0.0023568463511765003, + 0.003709198208525777, + 0.004773737862706184, + 0.002901076339185238, + 0.004886386450380087, + 0.0035183525178581476, + 0.005655381828546524, + 0.00264702457934618, + 0.002215656219050288, + 0.0034691800829023123, + 0.004762285389006138, + 0.0040320721454918385, + 0.0039526475593447685, + 0.005357581190764904, + 0.003264732426032424, + 0.010698731988668442, + 0.0038508055731654167, + 0.05481526628136635, + 0.008874930441379547, + 0.005162556190043688, + 0.05179040506482124, + 0.005298161879181862, + 0.005632453598082066, + 0.0019834430422633886, + 0.0033981630112975836, + 0.003957546781748533, + 0.0038973873015493155, + 0.002876447979360819, + 0.003453508485108614, + 0.004126178566366434, + 0.002763708820566535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.033717233687639236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631551939994097 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631551939994097, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631551939994097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031827785074710846, + 0.013443614356219769, + 0.012026246637105942, + 0.08281372487545013, + 0.45098012685775757, + 0.11524631083011627, + 0.03968062624335289, + 0.028601745143532753, + 0.19330550730228424, + 0.03207433968782425 + ], + "credit_map": [ + 0.003923959098756313, + 0.0051534902304410934, + 0.004869373980909586, + 0.0035456717014312744, + 0.006478216499090195, + 0.0033033224754035473, + 0.0021624004002660513, + 0.0023913471959531307, + 0.0011969127226620913, + 0.002225899137556553, + 0.0015102040488272905, + 0.0013467054814100266, + 0.0035849104169756174, + 0.001222800463438034, + 0.001293360604904592, + 0.0010628201998770237, + 0.001036189030855894, + 0.0012347052106633782, + 0.00196448084898293, + 0.001843116944655776, + 0.0024084881879389286, + 0.0014407184207811952, + 0.0009891215013340116, + 0.0011094261426478624, + 0.005714133381843567, + 0.022846227511763573, + 0.010804087854921818, + 0.008208131417632103, + 0.010637970641255379, + 0.007379075977951288, + 0.007498681079596281, + 0.009725409559905529, + 0.02926623262465, + 0.017251498997211456, + 0.1846162974834442, + 0.019634542986750603, + 0.044445086270570755, + 0.1180453673005104, + 0.01123605016618967, + 0.02648502215743065, + 0.008207062259316444, + 0.006431364454329014, + 0.027534151449799538, + 0.01515547651797533, + 0.010726372711360455, + 0.02627355046570301, + 0.012565878219902515, + 0.008352451026439667, + 0.002935181139037013, + 0.00255439430475235, + 0.008092528209090233, + 0.007612426765263081, + 0.0031648764852434397, + 0.008278414607048035, + 0.003682437352836132, + 0.0033603650517761707, + 0.00247353152371943, + 0.0031700583640486, + 0.004458946641534567, + 0.0039125652983784676, + 0.0032187781762331724, + 0.004946960601955652, + 0.0032425532117486, + 0.003178349928930402, + 0.014576835557818413, + 0.0055915722623467445, + 0.0733083188533783, + 0.01040179468691349, + 0.006677217315882444, + 0.06797251850366592, + 0.006137192249298096, + 0.008640085346996784, + 0.0022326214239001274, + 0.0034266668371856213, + 0.00513047631829977, + 0.003724557114765048, + 0.006158391013741493, + 0.003518796293064952, + 0.004410186316817999, + 0.003472643904387951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.007631551939994097 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot establishes a scene of immense and terrifying pressure, focusing on a massive, hollow, rectangular steel beam that dominates the scene. The entire reality is drenched in a dark, sleek, and oppressive atmosphere, bathed in an unnatural, monochromatic green emanating from the hollow-like environment. The beam's gunmetal gray surface, with faint traces of industrial wear and tear, gleams with a damp, oily sheen, catching distorted reflections of an eerie green light. Under the force of an unseen crushing force, the beam begins to collapse; a low, resonant groan vibrates through the space as its once-perfect geometric form begins to visibly bend. The movement is depicted with terrifying slowness, forming a deep crease along its length as the flattened steel plates carefully fold inward. As the structural integrity is compromised, the entire view is severely marred—the image flickers with glaring white and damaged green, the scene fractured by digital glitches, screen tearing, and momentary freezes, as if the recording medium, or reality itself, is collapsing. Through this chaotic visual static, the beam’s collapse accelerates, its hollow form decisively flattening as metal screams and tears, catastrophically, before finally being crushed into a mangled, unrecognizable ruin in a shimmering green darkness.", + "seed": 3854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_8728_seed_3854_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.34512264002114534, + "vlm_logprob": -0.34512264002114534, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.34512264002114534, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14688624441623688 + ], + "opposite_logprob_tokens": [ + -2.1468863487243652 + ], + "logprob_reward": -0.14688624441623688, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14688624441623688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1472838819026947, + 0.0936606302857399, + 0.046401266008615494, + 0.029700089246034622, + 0.06456815451383591, + 0.06668095290660858, + 0.37306663393974304, + 0.057784147560596466, + 0.060654737055301666, + 0.06019951403141022 + ], + "credit_map": [ + 0.014773350208997726, + 0.013507655821740627, + 0.013482382521033287, + 0.014774318784475327, + 0.02208719030022621, + 0.01983107626438141, + 0.027278702706098557, + 0.02154920995235443, + 0.010781338438391685, + 0.0075545175932347775, + 0.008971436880528927, + 0.008526374585926533, + 0.01564963161945343, + 0.013724280521273613, + 0.013834293931722641, + 0.014618759974837303, + 0.0061148484237492085, + 0.00354126188904047, + 0.003033145098015666, + 0.005287423264235258, + 0.006911790929734707, + 0.0053672813810408115, + 0.007786863017827272, + 0.008358649909496307, + 0.0034464423079043627, + 0.0027738134376704693, + 0.0037204117979854345, + 0.004264073446393013, + 0.003993230406194925, + 0.0041411761194467545, + 0.0032329848036170006, + 0.00412795739248395, + 0.006991778500378132, + 0.006087522488087416, + 0.006149674765765667, + 0.006956492085009813, + 0.007756838109344244, + 0.010896416381001472, + 0.011023853905498981, + 0.008705575950443745, + 0.008830892853438854, + 0.008887108415365219, + 0.0065932306461036205, + 0.0072737750597298145, + 0.006612664088606834, + 0.007869633845984936, + 0.009619190357625484, + 0.010994456708431244, + 0.029362350702285767, + 0.05035283789038658, + 0.041376993060112, + 0.037272363901138306, + 0.045665621757507324, + 0.045263949781656265, + 0.0764165148139, + 0.04735599085688591, + 0.006054156459867954, + 0.007583719678223133, + 0.008845522068440914, + 0.007182780187577009, + 0.007537598256021738, + 0.008128825575113297, + 0.006455747410655022, + 0.005995795596390963, + 0.004577886313199997, + 0.006942741107195616, + 0.007208588533103466, + 0.00763001199811697, + 0.005823304411023855, + 0.011995394714176655, + 0.005622443277388811, + 0.010854370892047882, + 0.0063408068381249905, + 0.010562841780483723, + 0.006212145090103149, + 0.009884069673717022, + 0.006095847114920616, + 0.008245968259871006, + 0.005581499543040991, + 0.007276334799826145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.14688624441623688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5021034479141235 + ], + "opposite_logprob_tokens": [ + -0.25210341811180115 + ], + "logprob_reward": -1.5021034479141235, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5021034479141235, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06787361204624176, + 0.04589157924056053, + 0.04765777289867401, + 0.045899294316768646, + 0.06498203426599503, + 0.06236639618873596, + 0.13230684399604797, + 0.2522352635860443, + 0.15323099493980408, + 0.12755624949932098 + ], + "credit_map": [ + 0.00770537881180644, + 0.006365151144564152, + 0.0043733976781368256, + 0.0048685818910598755, + 0.015061765909194946, + 0.007702796254307032, + 0.01277665514498949, + 0.009019884280860424, + 0.005405164789408445, + 0.003284297650679946, + 0.0037321255076676607, + 0.0033377143554389477, + 0.0072529553435742855, + 0.006430961191654205, + 0.007506907917559147, + 0.00894145667552948, + 0.004814725834876299, + 0.004999147262424231, + 0.0047789933159947395, + 0.00392630510032177, + 0.010374464094638824, + 0.006512368563562632, + 0.005248401314020157, + 0.007003366481512785, + 0.005517913959920406, + 0.0042411694303154945, + 0.005732010118663311, + 0.005659515503793955, + 0.00499571580439806, + 0.0075915781781077385, + 0.004599370993673801, + 0.007562019396573305, + 0.0075889816507697105, + 0.007570366840809584, + 0.010597260668873787, + 0.007386042736470699, + 0.007912587374448776, + 0.008160240016877651, + 0.008016281761229038, + 0.007750269956886768, + 0.006858821492642164, + 0.008946891874074936, + 0.008167413994669914, + 0.007257505785673857, + 0.0065036858431994915, + 0.008480675518512726, + 0.00739321717992425, + 0.008758182637393475, + 0.011659052222967148, + 0.013476611115038395, + 0.014629747718572617, + 0.012243730947375298, + 0.018475135788321495, + 0.01574193499982357, + 0.024854477494955063, + 0.021226156502962112, + 0.030390657484531403, + 0.02107928693294525, + 0.028814978897571564, + 0.026996679604053497, + 0.026858944445848465, + 0.035704419016838074, + 0.02697994001209736, + 0.05541034787893295, + 0.01621549017727375, + 0.018275529146194458, + 0.020313089713454247, + 0.016384154558181763, + 0.01517891138792038, + 0.02123488113284111, + 0.030163662508130074, + 0.015465283766388893, + 0.015998579561710358, + 0.020359912887215614, + 0.018051475286483765, + 0.014555681496858597, + 0.012996195815503597, + 0.01639462634921074, + 0.01151533517986536, + 0.017684435471892357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -1.5021034479141235 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307522758841515 + ], + "opposite_logprob_tokens": [ + -3.7733075618743896 + ], + "logprob_reward": -0.023307522758841515, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307522758841515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06855635344982147, + 0.03378230333328247, + 0.06507645547389984, + 0.05680234730243683, + 0.5016506910324097, + 0.046792298555374146, + 0.033817462623119354, + 0.10845725983381271, + 0.04877789691090584, + 0.0362868532538414 + ], + "credit_map": [ + 0.007506412919610739, + 0.006938282400369644, + 0.005672292318195105, + 0.004519371781498194, + 0.01187059935182333, + 0.009601396508514881, + 0.014246374368667603, + 0.008201631717383862, + 0.003536433447152376, + 0.00248347339220345, + 0.002581241074949503, + 0.00258412747643888, + 0.0056842295452952385, + 0.005511621944606304, + 0.0056274570524692535, + 0.005773724522441626, + 0.006322033237665892, + 0.005599782336503267, + 0.006330885924398899, + 0.008468232117593288, + 0.011222859844565392, + 0.008609418757259846, + 0.009077229537069798, + 0.009446021169424057, + 0.008297791704535484, + 0.005103716626763344, + 0.005716749466955662, + 0.004906839225441217, + 0.007866811007261276, + 0.008170761168003082, + 0.006300423759967089, + 0.010439259931445122, + 0.049457039684057236, + 0.051294680684804916, + 0.04450375214219093, + 0.051662229001522064, + 0.06899555772542953, + 0.0665014386177063, + 0.10792265087366104, + 0.061313387006521225, + 0.00440718000754714, + 0.003544741775840521, + 0.005276866722851992, + 0.0061227986589074135, + 0.0050314851105213165, + 0.006772697903215885, + 0.008992076851427555, + 0.00664445711299777, + 0.002777434652671218, + 0.004813326522707939, + 0.0038125202991068363, + 0.003212522016838193, + 0.00483881076797843, + 0.003792682895436883, + 0.005937365349382162, + 0.004632806871086359, + 0.014912018552422523, + 0.006371690891683102, + 0.013163910247385502, + 0.01026620902121067, + 0.009019790217280388, + 0.01617586612701416, + 0.009728060103952885, + 0.028819728642702103, + 0.004083724692463875, + 0.00484824925661087, + 0.0036678139586001635, + 0.005010808352380991, + 0.005559003446251154, + 0.009666688740253448, + 0.008386081084609032, + 0.007555536460131407, + 0.004177902825176716, + 0.004199597053229809, + 0.0044749947264790535, + 0.004587641917169094, + 0.004606593865901232, + 0.0068765850737690926, + 0.0025270574260503054, + 0.0048364861868321896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.023307522758841515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867510497570038 + ], + "opposite_logprob_tokens": [ + -3.048675060272217 + ], + "logprob_reward": -0.04867510497570038, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867510497570038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045145466923713684, + 0.03067677468061447, + 0.06007234752178192, + 0.09305454045534134, + 0.11595802754163742, + 0.1543406844139099, + 0.13003529608249664, + 0.06227915734052658, + 0.19318999350070953, + 0.1152477040886879 + ], + "credit_map": [ + 0.005665068048983812, + 0.004688818007707596, + 0.004035014193505049, + 0.0029773851856589317, + 0.007458342704921961, + 0.005618913099169731, + 0.006620938889682293, + 0.0080809835344553, + 0.0039775739423930645, + 0.002306017791852355, + 0.002314372221007943, + 0.004254749044775963, + 0.004424193408340216, + 0.003737977473065257, + 0.003990384750068188, + 0.005671507213264704, + 0.005991377867758274, + 0.005069543607532978, + 0.00602514436468482, + 0.010134201496839523, + 0.010480238124728203, + 0.007304191589355469, + 0.007033776957541704, + 0.008033869788050652, + 0.010667338036000729, + 0.009464714676141739, + 0.011230346746742725, + 0.010219849646091461, + 0.013839539140462875, + 0.01540098711848259, + 0.01174355298280716, + 0.010488214902579784, + 0.010894435457885265, + 0.012159131467342377, + 0.007176936604082584, + 0.009304730221629143, + 0.01362184900790453, + 0.01321353018283844, + 0.030815692618489265, + 0.018771713599562645, + 0.013476201333105564, + 0.018054647371172905, + 0.019287681207060814, + 0.019754130393266678, + 0.01751163974404335, + 0.02112911082804203, + 0.02423824556171894, + 0.020889027044177055, + 0.011317328549921513, + 0.015986602753400803, + 0.014078916050493717, + 0.015613975934684277, + 0.01772066205739975, + 0.016300143674016, + 0.022160161286592484, + 0.016857504844665527, + 0.008254270069301128, + 0.004933610092848539, + 0.006251391023397446, + 0.006949725095182657, + 0.005204867105931044, + 0.00886773131787777, + 0.006832285784184933, + 0.014985277317464352, + 0.029968248680233955, + 0.022981785237789154, + 0.019297368824481964, + 0.015978064388036728, + 0.01930287666618824, + 0.033926863223314285, + 0.02538408897817135, + 0.026350708678364754, + 0.017958926036953926, + 0.022001860663294792, + 0.014865916222333908, + 0.008297644555568695, + 0.009056661278009415, + 0.014478919096291065, + 0.008554890751838684, + 0.020032890141010284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.04867510497570038 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640880040824413 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640880040824413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640880040824413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041244830936193466, + 0.023696545511484146, + 0.051399704068899155, + 0.03378372639417648, + 0.12296445667743683, + 0.07745566964149475, + 0.0772431418299675, + 0.29574719071388245, + 0.19986170530319214, + 0.07660304009914398 + ], + "credit_map": [ + 0.005777004174888134, + 0.0035316217690706253, + 0.0036673746071755886, + 0.003868748666718602, + 0.005971583537757397, + 0.004550539422780275, + 0.007742879446595907, + 0.006135075818747282, + 0.0035738872829824686, + 0.0017007570713758469, + 0.0018037190893664956, + 0.0021444670855998993, + 0.0027698553167283535, + 0.0030294395983219147, + 0.005383890587836504, + 0.003290528431534767, + 0.006195198744535446, + 0.0036490203347057104, + 0.004819735884666443, + 0.007886538282036781, + 0.007940650917589664, + 0.006988649722188711, + 0.006674620322883129, + 0.007245288696140051, + 0.0035423405934125185, + 0.003343984019011259, + 0.00426245154812932, + 0.0036818278022110462, + 0.005374981090426445, + 0.005682368762791157, + 0.0037796811666339636, + 0.004116093274205923, + 0.011802002787590027, + 0.013591590337455273, + 0.007787160575389862, + 0.008361211977899075, + 0.015539376065135002, + 0.01427437923848629, + 0.03220326453447342, + 0.019405484199523926, + 0.006977292243391275, + 0.008641520515084267, + 0.00965291727334261, + 0.010576682165265083, + 0.008931067772209644, + 0.010148607194423676, + 0.013449948281049728, + 0.009077641181647778, + 0.009776412509381771, + 0.00913784559816122, + 0.009743851609528065, + 0.0063029187731444836, + 0.007257553283125162, + 0.009359996765851974, + 0.011038538068532944, + 0.014626018702983856, + 0.04390135407447815, + 0.02217445708811283, + 0.03884446993470192, + 0.027341576293110847, + 0.02052335813641548, + 0.04364045709371567, + 0.024422496557235718, + 0.07489901036024094, + 0.02037498727440834, + 0.02202954888343811, + 0.02201504074037075, + 0.02235311083495617, + 0.025220666080713272, + 0.03136221691966057, + 0.026516083627939224, + 0.029990050941705704, + 0.009775001555681229, + 0.01141148991882801, + 0.005445375107228756, + 0.008634843863546848, + 0.008747722022235394, + 0.010663759894669056, + 0.012646310031414032, + 0.009278540499508381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.004640880040824413 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot establishes a scene of immense and terrifying pressure, focusing on a massive, hollow, rectangular steel beam that dominates the scene. The entire reality is drenched in a dark, sleek, and oppressive atmosphere, bathed in an unnatural, monochromatic green emanating from the hollow-like environment. The beam's gunmetal gray surface, with faint traces of industrial wear and tear, gleams with a damp, oily sheen, catching distorted reflections of an eerie green light. Under the force of an unseen crushing force, the beam begins to collapse; a low, resonant groan vibrates through the space as its once-perfect geometric form begins to visibly bend. The movement is depicted with terrifying slowness, forming a deep crease along its length as the flattened steel plates carefully fold inward. As the structural integrity is compromised, the entire view is severely marred—the image flickers with glaring white and damaged green, the scene fractured by digital glitches, screen tearing, and momentary freezes, as if the recording medium, or reality itself, is collapsing. Through this chaotic visual static, the beam’s collapse accelerates, its hollow form decisively flattening as metal screams and tears, catastrophically, before finally being crushed into a mangled, unrecognizable ruin in a shimmering green darkness.", + "seed": 3855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_8728_seed_3855_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3903554965741932, + "vlm_logprob": -0.3903554965741932, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.3903554965741932, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31123530864715576 + ], + "opposite_logprob_tokens": [ + -1.4362353086471558 + ], + "logprob_reward": -0.31123530864715576, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.31123530864715576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6283991932868958, + 0.012256980873644352, + 0.009999791160225868, + 0.1278434693813324, + 0.00455468101426959, + 0.06987378001213074, + 0.11224058270454407, + 0.025758834555745125, + 0.006036922335624695, + 0.0030356887727975845 + ], + "credit_map": [ + 0.030788417905569077, + 0.009518777951598167, + 0.010644313879311085, + 0.008862665854394436, + 0.019624024629592896, + 0.24555081129074097, + 0.052330050617456436, + 0.2510802447795868, + 0.001076486543752253, + 0.0008209675434045494, + 0.0009731299360282719, + 0.0009867892367765307, + 0.0012242571683600545, + 0.0024666949175298214, + 0.0019689002074301243, + 0.0027397582307457924, + 0.0010795647976920009, + 0.0004924155655317008, + 0.0005109603516757488, + 0.0004369543748907745, + 0.0007405572105199099, + 0.002342767082154751, + 0.0015027555637061596, + 0.002893817611038685, + 0.006067880429327488, + 0.0011362620862200856, + 0.0016502708895131946, + 0.0015495436964556575, + 0.002610059455037117, + 0.05418691039085388, + 0.006795977242290974, + 0.0538465790450573, + 0.0003278454823885113, + 0.00013507006224244833, + 0.00014156838005874306, + 0.00012621820496860892, + 0.00023331903503276408, + 0.001420258660800755, + 0.0005469200550578535, + 0.0016234820941463113, + 0.00338809285312891, + 0.0012727276189252734, + 0.000940561352763325, + 0.0008639892912469804, + 0.0012544253841042519, + 0.029579030349850655, + 0.0033648894168436527, + 0.029210075736045837, + 0.005096672102808952, + 0.003661177586764097, + 0.0017793758306652308, + 0.002327319933101535, + 0.004322849214076996, + 0.04356285184621811, + 0.0076819537207484245, + 0.04380841553211212, + 0.0018884609453380108, + 0.0020606699399650097, + 0.0009301200043410063, + 0.0013788316864520311, + 0.0047032092697918415, + 0.0039510466158390045, + 0.006808997131884098, + 0.004037503153085709, + 0.0008597620762884617, + 0.0007812886033207178, + 0.00032059947261586785, + 0.0007445921073667705, + 0.0005948367179371417, + 0.0010445811785757542, + 0.0005586273036897182, + 0.0011326357489451766, + 0.0003360239206813276, + 0.0004556782077997923, + 0.0001916625042213127, + 0.00018721030210144818, + 0.0007649323088116944, + 0.00046200028737075627, + 0.0002705264778342098, + 0.00036765559343621135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.31123530864715576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005825996398926 + ], + "opposite_logprob_tokens": [ + -0.225582554936409 + ], + "logprob_reward": -1.6005825996398926, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005825996398926, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2587798833847046, + 0.01432123128324747, + 0.03981152921915054, + 0.11623434722423553, + 0.02040097303688526, + 0.2825937271118164, + 0.1463550329208374, + 0.0705079585313797, + 0.028828993439674377, + 0.02216637134552002 + ], + "credit_map": [ + 0.012671079486608505, + 0.004442769568413496, + 0.004499970003962517, + 0.005133945029228926, + 0.009064925834536552, + 0.0954701155424118, + 0.028490928933024406, + 0.09900613874197006, + 0.0011634507682174444, + 0.0005422402755357325, + 0.0006759834941476583, + 0.0009206950780935585, + 0.001528294524177909, + 0.0032147096935659647, + 0.0029961259569972754, + 0.0032797299791127443, + 0.004230357240885496, + 0.0022817105054855347, + 0.002298657549545169, + 0.0022425325587391853, + 0.003174817655235529, + 0.008945903740823269, + 0.007615978829562664, + 0.009021570906043053, + 0.005556708201766014, + 0.0011921199038624763, + 0.0016453127609565854, + 0.0014347434043884277, + 0.003016563830897212, + 0.04817003384232521, + 0.007219931576400995, + 0.04799893498420715, + 0.0018823809223249555, + 0.0006307908915914595, + 0.0008556850953027606, + 0.0006891572847962379, + 0.0010464148363098502, + 0.006789483595639467, + 0.002516952808946371, + 0.005990108009427786, + 0.0144192511215806, + 0.011618669144809246, + 0.004616559483110905, + 0.004968462511897087, + 0.006121361628174782, + 0.1174241453409195, + 0.01072898879647255, + 0.11269629746675491, + 0.006439797580242157, + 0.00337564991787076, + 0.0025023818016052246, + 0.0026465565897524357, + 0.005644843447953463, + 0.059912871569395065, + 0.006804403383284807, + 0.05902852863073349, + 0.0056261527352035046, + 0.005330762825906277, + 0.0029081290122121572, + 0.005924371536821127, + 0.007667398080229759, + 0.01418448518961668, + 0.017385762184858322, + 0.011480900458991528, + 0.004046735353767872, + 0.003035311121493578, + 0.002148014260455966, + 0.004522422794252634, + 0.0030255287420004606, + 0.004880259279161692, + 0.0028789672069251537, + 0.004291752818971872, + 0.004963348153978586, + 0.003309760708361864, + 0.0016893631545826793, + 0.00148171023465693, + 0.003024631878361106, + 0.0034324333537369967, + 0.0013350486988201737, + 0.0029300758615136147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -1.6005825996398926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308921605348587 + ], + "opposite_logprob_tokens": [ + -3.7733089923858643 + ], + "logprob_reward": -0.023308921605348587, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308921605348587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4868395924568176, + 0.03405407443642616, + 0.004742662888020277, + 0.12109865993261337, + 0.008908044546842575, + 0.03195388615131378, + 0.2738553285598755, + 0.020394980907440186, + 0.008134597912430763, + 0.01001814380288124 + ], + "credit_map": [ + 0.0244483333081007, + 0.005572118796408176, + 0.007946998812258244, + 0.006627271417528391, + 0.013210962526500225, + 0.1958150565624237, + 0.036224495619535446, + 0.1969943642616272, + 0.0022078349720686674, + 0.0008601154549978673, + 0.0014239666052162647, + 0.0008995163370855153, + 0.0017218315042555332, + 0.012447197921574116, + 0.0025262117851525545, + 0.011967402882874012, + 0.0005264355568215251, + 0.00030533128301613033, + 0.00022611003078054637, + 0.0002233614941360429, + 0.00030594016425311565, + 0.0014005895936861634, + 0.0007658441318199039, + 0.0009890507208183408, + 0.005877327173948288, + 0.0012099953601136804, + 0.001661836402490735, + 0.0014261434553191066, + 0.0027197941672056913, + 0.051035135984420776, + 0.0067414576187729836, + 0.05042697489261627, + 0.000546887400560081, + 0.00019683153368532658, + 0.00019432397675700486, + 0.0002239389141323045, + 0.00034618060453794897, + 0.0033187405206263065, + 0.000758047797717154, + 0.003323094453662634, + 0.0017474440392106771, + 0.0016571242595091462, + 0.0006622282089665532, + 0.0013348362408578396, + 0.0008304758812300861, + 0.012698099948465824, + 0.0014520443510264158, + 0.011571632698178291, + 0.011936291120946407, + 0.004023666027933359, + 0.00374444667249918, + 0.002978404052555561, + 0.006879530847072601, + 0.11821931600570679, + 0.008573674596846104, + 0.11749999970197678, + 0.0018443805165588856, + 0.001750453608110547, + 0.0009599432814866304, + 0.0014226631028577685, + 0.002991008572280407, + 0.0036987510975450277, + 0.004052582196891308, + 0.003675199346616864, + 0.0011989688500761986, + 0.0009778729872778058, + 0.0006781626143492758, + 0.0009307971340604126, + 0.001236574724316597, + 0.0009975607972592115, + 0.0010095116449519992, + 0.0011051488108932972, + 0.0026230383664369583, + 0.0014526231680065393, + 0.0007483077934011817, + 0.0006697587668895721, + 0.001429717056453228, + 0.0011739388573914766, + 0.0006954812561161816, + 0.0012252775486558676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.023308921605348587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552659027278423 + ], + "opposite_logprob_tokens": [ + -4.387552738189697 + ], + "logprob_reward": -0.012552659027278423, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552659027278423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07185330986976624, + 0.025059277191758156, + 0.06838886439800262, + 0.1336502581834793, + 0.019398201256990433, + 0.08649970591068268, + 0.4573209285736084, + 0.04259152337908745, + 0.04132736101746559, + 0.05391060188412666 + ], + "credit_map": [ + 0.0055899848230183125, + 0.002792914630845189, + 0.004051355645060539, + 0.004228703211992979, + 0.006671916227787733, + 0.01585288718342781, + 0.011179393157362938, + 0.021486148238182068, + 0.0018077354179695249, + 0.0008654603152535856, + 0.0010988616850227118, + 0.0011955375084653497, + 0.0023387812543660402, + 0.006916390266269445, + 0.0029886486008763313, + 0.007847863249480724, + 0.006435287185013294, + 0.0021858704276382923, + 0.0019466375233605504, + 0.0026214660611003637, + 0.0032029638532549143, + 0.020686769858002663, + 0.009439922869205475, + 0.02186994068324566, + 0.0063272640109062195, + 0.0014527463354170322, + 0.0019414887065067887, + 0.0019394181435927749, + 0.003613534616306424, + 0.05413113534450531, + 0.010502515360713005, + 0.05374214053153992, + 0.001654385938309133, + 0.0006547338562086225, + 0.000855035672429949, + 0.0007691916543990374, + 0.0009201230714097619, + 0.006564219016581774, + 0.002219733316451311, + 0.005760779604315758, + 0.00836122315376997, + 0.009540633298456669, + 0.003946274518966675, + 0.00761946989223361, + 0.004643604159355164, + 0.019648054614663124, + 0.01538463681936264, + 0.01735580898821354, + 0.020173192024230957, + 0.008279530331492424, + 0.006411905400454998, + 0.00526086101308465, + 0.011260947212576866, + 0.19440600275993347, + 0.017094962298870087, + 0.19443351030349731, + 0.002542714821174741, + 0.0024750123266130686, + 0.0016491904389113188, + 0.003293644404038787, + 0.004062653984874487, + 0.010935193859040737, + 0.006596001330763102, + 0.011037110351026058, + 0.004585033282637596, + 0.00533002894371748, + 0.0038602205459028482, + 0.006745887920260429, + 0.004300318192690611, + 0.005109150428324938, + 0.004037871956825256, + 0.007358844392001629, + 0.01634751632809639, + 0.008351503871381283, + 0.0037850113585591316, + 0.003170269075781107, + 0.006966273766011, + 0.005914864130318165, + 0.0030350538436323404, + 0.006340107414871454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.012552659027278423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097993951290846 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097993951290846, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097993951290846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6268818974494934, + 0.02868688851594925, + 0.0225081704556942, + 0.0059043834917247295, + 0.007426750380545855, + 0.06693021953105927, + 0.16254094243049622, + 0.028848087415099144, + 0.02159687504172325, + 0.028675740584731102 + ], + "credit_map": [ + 0.03016429953277111, + 0.009588639251887798, + 0.010143483988940716, + 0.010085172951221466, + 0.018924450501799583, + 0.24184055626392365, + 0.05689400061964989, + 0.24924132227897644, + 0.001528588472865522, + 0.0004931132425554097, + 0.0005622658063657582, + 0.0005714674480259418, + 0.0008017960353754461, + 0.011252019554376602, + 0.002544342540204525, + 0.010933294892311096, + 0.001268537133000791, + 0.0005661436589434743, + 0.0005303862853907049, + 0.00047975676716305315, + 0.0007922776858322322, + 0.008809437975287437, + 0.0015594010474160314, + 0.008502230979502201, + 0.0005795681499876082, + 0.0003094608255196363, + 0.0003605425590649247, + 0.0003692805184982717, + 0.0005385735421441495, + 0.0010068531846627593, + 0.0014813020825386047, + 0.0012588024837896228, + 0.0006659094360657036, + 0.00040486373472958803, + 0.000473644002340734, + 0.000412214343668893, + 0.0006980161997489631, + 0.001784247113391757, + 0.0014937641099095345, + 0.00149409135337919, + 0.0036904937587678432, + 0.003062367672100663, + 0.001374163432046771, + 0.003029798623174429, + 0.0014272044645622373, + 0.02619432844221592, + 0.003374646184965968, + 0.02477722056210041, + 0.006884355563670397, + 0.0026845894753932953, + 0.0022407311480492353, + 0.0018101709429174662, + 0.003740488551557064, + 0.06955558806657791, + 0.005672899074852467, + 0.0699521154165268, + 0.002649253234267235, + 0.0018603921635076404, + 0.0010153900366276503, + 0.0015376628143712878, + 0.0029678416904062033, + 0.0062543414533138275, + 0.005681448616087437, + 0.0068817585706710815, + 0.003153144381940365, + 0.0025479942560195923, + 0.0018640134949237108, + 0.0035822365898638964, + 0.0021914509125053883, + 0.003053266555070877, + 0.0019768786150962114, + 0.0032278909347951412, + 0.00720252376049757, + 0.004471330903470516, + 0.0018982243491336703, + 0.0016696691745892167, + 0.0032286997884511948, + 0.003997249994426966, + 0.002070147078484297, + 0.004137894604355097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.004097993951290846 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot establishes a scene of immense and terrifying pressure, focusing on a massive, hollow, rectangular steel beam that dominates the scene. The entire reality is drenched in a dark, sleek, and oppressive atmosphere, bathed in an unnatural, monochromatic green emanating from the hollow-like environment. The beam's gunmetal gray surface, with faint traces of industrial wear and tear, gleams with a damp, oily sheen, catching distorted reflections of an eerie green light. Under the force of an unseen crushing force, the beam begins to collapse; a low, resonant groan vibrates through the space as its once-perfect geometric form begins to visibly bend. The movement is depicted with terrifying slowness, forming a deep crease along its length as the flattened steel plates carefully fold inward. As the structural integrity is compromised, the entire view is severely marred—the image flickers with glaring white and damaged green, the scene fractured by digital glitches, screen tearing, and momentary freezes, as if the recording medium, or reality itself, is collapsing. Through this chaotic visual static, the beam’s collapse accelerates, its hollow form decisively flattening as metal screams and tears, catastrophically, before finally being crushed into a mangled, unrecognizable ruin in a shimmering green darkness.", + "seed": 3856, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_7/index_8728_seed_3856_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.39529218915849923, + "vlm_logprob": -0.39529218915849923, + "vlm_yes_no_margin": 1.775, + "vlm_token_credit": -0.39529218915849923, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16466885805130005 + ], + "opposite_logprob_tokens": [ + -2.0396687984466553 + ], + "logprob_reward": -0.16466885805130005, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.16466885805130005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043385520577430725, + 0.02667068876326084, + 0.018644161522388458, + 0.020095810294151306, + 0.03455203026533127, + 0.2631012499332428, + 0.02537887543439865, + 0.04276209697127342, + 0.4893840551376343, + 0.03602549061179161 + ], + "credit_map": [ + 0.005263712257146835, + 0.008132031187415123, + 0.005016640294343233, + 0.003359798341989517, + 0.005812197923660278, + 0.007134007290005684, + 0.0031058378517627716, + 0.005561295431107283, + 0.002816723892465234, + 0.00478954566642642, + 0.0039003570564091206, + 0.003959507681429386, + 0.0030077972915023565, + 0.0031949146650731564, + 0.0021952337119728327, + 0.0028066120576113462, + 0.00171603262424469, + 0.0032406102400273085, + 0.0024949521757662296, + 0.002599813509732485, + 0.0025295510422438383, + 0.002454464789479971, + 0.0014724720967933536, + 0.002136264694854617, + 0.0025342402514070272, + 0.00283099920488894, + 0.0030603541526943445, + 0.0019992461893707514, + 0.003423114074394107, + 0.0025109529960900545, + 0.001606137491762638, + 0.0021307654678821564, + 0.002490698592737317, + 0.00345181068405509, + 0.003581459866836667, + 0.0048992326483130455, + 0.0024634378496557474, + 0.007585464045405388, + 0.002247428521513939, + 0.007832498289644718, + 0.032028310000896454, + 0.01594497635960579, + 0.043752171099185944, + 0.015577585436403751, + 0.00919315218925476, + 0.06537073105573654, + 0.008650660514831543, + 0.07258366793394089, + 0.003639515256509185, + 0.00240768282674253, + 0.0042101494036614895, + 0.0038736567366868258, + 0.001433997298590839, + 0.0036003473214805126, + 0.0021158545278012753, + 0.004097672179341316, + 0.0017547785537317395, + 0.0014201708836480975, + 0.004483459517359734, + 0.0031944357324391603, + 0.001250601839274168, + 0.0145973339676857, + 0.0021825607400387526, + 0.013878753408789635, + 0.02511175535619259, + 0.009074772708117962, + 0.07552796602249146, + 0.011975164525210857, + 0.013394864276051521, + 0.19697533547878265, + 0.015346436761319637, + 0.14197774231433868, + 0.0030736345797777176, + 0.006303311325609684, + 0.004831509664654732, + 0.005024867132306099, + 0.003995729144662619, + 0.00417895894497633, + 0.003594743087887764, + 0.005022735334932804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.16466885805130005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236707210540771 + ], + "opposite_logprob_tokens": [ + -0.34867069125175476 + ], + "logprob_reward": -1.2236707210540771, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236707210540771, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03998608514666557, + 0.029118899255990982, + 0.046434078365564346, + 0.030094292014837265, + 0.16276653110980988, + 0.353884220123291, + 0.032571904361248016, + 0.045455336570739746, + 0.14067357778549194, + 0.11901506036520004 + ], + "credit_map": [ + 0.0044045173563063145, + 0.00563351484015584, + 0.004733219742774963, + 0.003493921598419547, + 0.004883408546447754, + 0.006596498657017946, + 0.0028348276391625404, + 0.007406176067888737, + 0.0022543352097272873, + 0.0069123865105211735, + 0.003985540941357613, + 0.003334400011226535, + 0.0028335631359368563, + 0.004125065635889769, + 0.0022891713306307793, + 0.0033844346180558205, + 0.0034809312783181667, + 0.008069731295108795, + 0.007308248430490494, + 0.004930894821882248, + 0.0074431453831493855, + 0.006265158299356699, + 0.003943949472159147, + 0.004992017988115549, + 0.003011556575074792, + 0.005028590559959412, + 0.0036825144197791815, + 0.003621965181082487, + 0.004623919725418091, + 0.003801485989242792, + 0.0028055645525455475, + 0.0035186957102268934, + 0.006787531077861786, + 0.015843834728002548, + 0.011165529489517212, + 0.019229408353567123, + 0.008359777741134167, + 0.044254954904317856, + 0.008644933812320232, + 0.048480551689863205, + 0.03962520509958267, + 0.022713683545589447, + 0.05217691883444786, + 0.02223106473684311, + 0.011839947663247585, + 0.09204816818237305, + 0.010992649011313915, + 0.10225659608840942, + 0.002477631438523531, + 0.003881763434037566, + 0.00444481847807765, + 0.0036273696459829807, + 0.0013461626367643476, + 0.007648232392966747, + 0.0020463750697672367, + 0.007099551148712635, + 0.002182439435273409, + 0.001842209487222135, + 0.004445191007107496, + 0.002237536944448948, + 0.001287015969865024, + 0.016327764838933945, + 0.0016417286824434996, + 0.01549145020544529, + 0.008312409743666649, + 0.004096697084605694, + 0.015982890501618385, + 0.006091831251978874, + 0.003990615252405405, + 0.04554964229464531, + 0.006338891107589006, + 0.050310589373111725, + 0.009728881530463696, + 0.026479046791791916, + 0.01821299083530903, + 0.01809169352054596, + 0.01468015369027853, + 0.010012485086917877, + 0.013678256422281265, + 0.008131553418934345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -1.2236707210540771 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233443975448608 + ], + "opposite_logprob_tokens": [ + -0.8983443975448608 + ], + "logprob_reward": -0.5233443975448608, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233443975448608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05641515180468559, + 0.03898301348090172, + 0.024727294221520424, + 0.02561587281525135, + 0.0361858531832695, + 0.4376896619796753, + 0.12012378126382828, + 0.16307960450649261, + 0.07215210795402527, + 0.025027649477124214 + ], + "credit_map": [ + 0.005469897296279669, + 0.010491810739040375, + 0.007502230815589428, + 0.004761890508234501, + 0.0067145731300115585, + 0.008737156167626381, + 0.0037942430935800076, + 0.008943353779613972, + 0.004601405467838049, + 0.004920839797705412, + 0.005272872280329466, + 0.0036020451225340366, + 0.0051263622008264065, + 0.005312041845172644, + 0.0030493130907416344, + 0.0070981369353830814, + 0.0030436080414801836, + 0.0035818549804389477, + 0.003700418397784233, + 0.002718414179980755, + 0.003807842032983899, + 0.003121779067441821, + 0.002074951771646738, + 0.002678429242223501, + 0.002178141614422202, + 0.0037834455724805593, + 0.004719379357993603, + 0.002805486088618636, + 0.003166978480294347, + 0.0032756072469055653, + 0.0027497897390276194, + 0.002937048440799117, + 0.0030541967134922743, + 0.005898261442780495, + 0.003646386554464698, + 0.005968836136162281, + 0.004255527630448341, + 0.004731619730591774, + 0.0033792154863476753, + 0.005251813214272261, + 0.05007931962609291, + 0.026454035192728043, + 0.0629635825753212, + 0.029001152142882347, + 0.015279674902558327, + 0.1133495569229126, + 0.014563019387423992, + 0.12599939107894897, + 0.01026851125061512, + 0.011037930846214294, + 0.024700088426470757, + 0.012693020515143871, + 0.005108174402266741, + 0.023208187893033028, + 0.00748799042776227, + 0.02561989054083824, + 0.005335192196071148, + 0.0036348223220556974, + 0.01615767553448677, + 0.005430888384580612, + 0.003118004882708192, + 0.06565552204847336, + 0.005428641568869352, + 0.058318864554166794, + 0.004388293717056513, + 0.0022860749159008265, + 0.009931501932442188, + 0.0025352612137794495, + 0.0032696074340492487, + 0.02621600218117237, + 0.0037065402138978243, + 0.01981883868575096, + 0.0017848361749202013, + 0.003586573526263237, + 0.004015753977000713, + 0.0034938787575811148, + 0.0043304674327373505, + 0.0023802253417670727, + 0.003463207744061947, + 0.0019727081526070833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.5233443975448608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054988596588373184 + ], + "opposite_logprob_tokens": [ + -2.9299886226654053 + ], + "logprob_reward": -0.054988596588373184, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054988596588373184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030399875715374947, + 0.031230228021740913, + 0.030614010989665985, + 0.020623093470931053, + 0.037652455270290375, + 0.40966254472732544, + 0.028904013335704803, + 0.1896170675754547, + 0.1235957220196724, + 0.0977010503411293 + ], + "credit_map": [ + 0.003627066034823656, + 0.006423073820769787, + 0.004504113923758268, + 0.0023310391698032618, + 0.0033006840385496616, + 0.004860764369368553, + 0.0019029012182727456, + 0.0034502320922911167, + 0.003846302628517151, + 0.005638824310153723, + 0.005079625640064478, + 0.003328463761135936, + 0.0033416401129215956, + 0.004059618338942528, + 0.001872451277449727, + 0.004063303116708994, + 0.003199779661372304, + 0.005193253979086876, + 0.004137316718697548, + 0.003202232299372554, + 0.004309909418225288, + 0.004558416549116373, + 0.0024412593338638544, + 0.0035718439612537622, + 0.001926373690366745, + 0.002503200201317668, + 0.0031424853950738907, + 0.002229514764621854, + 0.0024011721834540367, + 0.0034217792563140392, + 0.0021152349654585123, + 0.00288333254866302, + 0.003128885291516781, + 0.007395529188215733, + 0.004156202077865601, + 0.006910133641213179, + 0.0038239676505327225, + 0.005591523367911577, + 0.0027146756183356047, + 0.003931539133191109, + 0.05037541314959526, + 0.025647424161434174, + 0.06814075261354446, + 0.024756161496043205, + 0.01414760947227478, + 0.1011529415845871, + 0.012775537557899952, + 0.11266671121120453, + 0.00451572285965085, + 0.00540729146450758, + 0.004712691996246576, + 0.0036641242913901806, + 0.0014582574367523193, + 0.003894918831065297, + 0.0018752736505120993, + 0.0033757321070879698, + 0.006007741671055555, + 0.00562999676913023, + 0.017080042511224747, + 0.009337960742413998, + 0.004829553421586752, + 0.07397878170013428, + 0.005134988576173782, + 0.0676180049777031, + 0.004740449134260416, + 0.0035782549530267715, + 0.025326557457447052, + 0.005555630661547184, + 0.003108308184891939, + 0.048003170639276505, + 0.005403149873018265, + 0.027880188077688217, + 0.005923260934650898, + 0.013305672444403172, + 0.017234789207577705, + 0.01733579859137535, + 0.013940878212451935, + 0.010143878869712353, + 0.011622940190136433, + 0.008193830959498882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.054988596588373184 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive, hollow, rectangular steel beam with a gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the beam visibly bend, form a deep crease, and flatten under an unseen crushing force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in an unnatural, monochromatic green light with digital glitches, screen tearing, and visual static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788372553884983 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.009788372553884983, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788372553884983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08694572746753693, + 0.04707002639770508, + 0.021988503634929657, + 0.032482121139764786, + 0.10418195277452469, + 0.11597785353660583, + 0.08296888321638107, + 0.033269673585891724, + 0.41136226058006287, + 0.06375300884246826 + ], + "credit_map": [ + 0.009466719813644886, + 0.017846597358584404, + 0.009253463707864285, + 0.008577846921980381, + 0.008660289458930492, + 0.012916874140501022, + 0.006583943963050842, + 0.013639985583722591, + 0.004945872817188501, + 0.012031211517751217, + 0.004272863268852234, + 0.00370975723490119, + 0.005032021552324295, + 0.007971036247909069, + 0.0031909928657114506, + 0.00591626949608326, + 0.002272237092256546, + 0.0036988414358347654, + 0.0046800849959254265, + 0.0022164147812873125, + 0.0025156487245112658, + 0.0027819371316581964, + 0.0015039785066619515, + 0.002319362945854664, + 0.00276811677031219, + 0.004220522474497557, + 0.007238446269184351, + 0.0038187559694051743, + 0.0035031214356422424, + 0.004058947321027517, + 0.003176288679242134, + 0.0036979226861149073, + 0.005626826547086239, + 0.007313738111406565, + 0.009640148840844631, + 0.015124852769076824, + 0.006175324786454439, + 0.0251065231859684, + 0.007526766043156385, + 0.02766776643693447, + 0.0066230506636202335, + 0.01916193589568138, + 0.010723079554736614, + 0.01069306954741478, + 0.0046803331933915615, + 0.028779761865735054, + 0.0051889847964048386, + 0.030127648264169693, + 0.007459370885044336, + 0.006241116207093, + 0.012784521095454693, + 0.008192192763090134, + 0.0029111651238054037, + 0.019583191722631454, + 0.004919824656099081, + 0.020877499133348465, + 0.001910920487716794, + 0.0023625781759619713, + 0.0036105962935835123, + 0.003895111382007599, + 0.0017493744380772114, + 0.0077132973819971085, + 0.003127295756712556, + 0.008900498040020466, + 0.01948002725839615, + 0.011471669189631939, + 0.06812722235918045, + 0.0107595669105649, + 0.015038775280117989, + 0.16072353720664978, + 0.016223426908254623, + 0.10953804850578308, + 0.006204763427376747, + 0.006970255635678768, + 0.010630390606820583, + 0.010606905445456505, + 0.006860665045678616, + 0.007222429849207401, + 0.008913653902709484, + 0.0063439421355724335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.009788372553884983 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_8/index_tensor([7634, 7634])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_8/index_tensor([7634, 7634])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0ba88f34e7133104d39912ed69df0f106e9c34c3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_8/index_tensor([7634, 7634])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot, shot from a slightly elevated angle, captures a perfect, pale yellow rectangle of cold butter, placed in the center of a hot, black, seasoned cast iron skillet. Warm, direct overhead light casts a soft sheen on the butter's smooth surface and creates subtle highlights on the pan's textured matte finish. Almost immediately, the butter's sharp edges begin to soften and turn translucent, as a shimmering golden liquid rapidly pools at its base and spreads outward. The solid structure gracefully collapses and shrinks, its shape dissolving as it melts from below, becoming a tiny island within its own expanding pool. Tiny, hissing bubbles emerge and dance on the liquid's surface, which steadily spreads across the dark pan, reflecting light like a golden mirror, creating a captivating, satisfying atmosphere completely focused on physical transformation.", + "seed": 15956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_8/index_7634_seed_15956_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01367408111691475, + "vlm_logprob": -0.01367408111691475, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.01367408111691475, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011749823577702045 + ], + "opposite_logprob_tokens": [ + -5.511749744415283 + ], + "logprob_reward": -0.011749823577702045, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.011749823577702045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08655082434415817, + 0.07649863511323929, + 0.05661727488040924, + 0.09551677852869034, + 0.12045755237340927, + 0.2079235166311264, + 0.06608326733112335, + 0.06374765932559967, + 0.08969251811504364, + 0.136912003159523 + ], + "credit_map": [ + 0.012723889201879501, + 0.009832551702857018, + 0.008930224925279617, + 0.009852774441242218, + 0.011125678196549416, + 0.012002117931842804, + 0.00842012744396925, + 0.013663451187312603, + 0.00697499793022871, + 0.009686056524515152, + 0.007964696735143661, + 0.008123326115310192, + 0.010355504229664803, + 0.011005728505551815, + 0.013079311698675156, + 0.009309007786214352, + 0.005179555155336857, + 0.011222032830119133, + 0.008324116468429565, + 0.005938860587775707, + 0.006656017620116472, + 0.006283446215093136, + 0.005853675771504641, + 0.007159566972404718, + 0.0077276648953557014, + 0.0133463554084301, + 0.012475317344069481, + 0.017390485852956772, + 0.014703793451189995, + 0.00827538687735796, + 0.008619404397904873, + 0.012978359125554562, + 0.007677782326936722, + 0.015416587702929974, + 0.01795404963195324, + 0.022060757502913475, + 0.01635170169174671, + 0.009098256006836891, + 0.015356087125837803, + 0.01654232107102871, + 0.019088461995124817, + 0.0784398540854454, + 0.05824260786175728, + 0.013931064866483212, + 0.01604822650551796, + 0.008479888550937176, + 0.00790998712182045, + 0.005783397704362869, + 0.005307196639478207, + 0.01569579355418682, + 0.01034728717058897, + 0.00796353816986084, + 0.009341849945485592, + 0.005560059566050768, + 0.007577797397971153, + 0.004289742559194565, + 0.00765440845862031, + 0.008765177801251411, + 0.013106214813888073, + 0.009523936547338963, + 0.008374687284231186, + 0.005463060922920704, + 0.006578755099326372, + 0.0042814151383936405, + 0.009217637591063976, + 0.012845002114772797, + 0.015846870839595795, + 0.013564759865403175, + 0.010579613037407398, + 0.011756167747080326, + 0.00877000205218792, + 0.007112450432032347, + 0.014713358134031296, + 0.018868062645196915, + 0.024042878299951553, + 0.019143568351864815, + 0.01575562171638012, + 0.013015909120440483, + 0.018881026655435562, + 0.012491567991673946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.011749823577702045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02333465963602066 + ], + "opposite_logprob_tokens": [ + -3.7733347415924072 + ], + "logprob_reward": -0.02333465963602066, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02333465963602066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03266420215368271, + 0.02945680171251297, + 0.016780951991677284, + 0.05516741797327995, + 0.03775782510638237, + 0.14113667607307434, + 0.08724237233400345, + 0.07696060836315155, + 0.04397911950945854, + 0.4788539707660675 + ], + "credit_map": [ + 0.0033198061864823103, + 0.004986926447600126, + 0.0035627891775220633, + 0.003338103648275137, + 0.005307960323989391, + 0.00455015292391181, + 0.0033534402027726173, + 0.004245022777467966, + 0.00336090219207108, + 0.003909612074494362, + 0.00408876221626997, + 0.003482194384559989, + 0.00338739063590765, + 0.004071942530572414, + 0.0028084437362849712, + 0.004347554408013821, + 0.0016287803882732987, + 0.0027799324598163366, + 0.002150461543351412, + 0.0017859400250017643, + 0.0024010862689465284, + 0.001859001349657774, + 0.0017859479412436485, + 0.002389802597463131, + 0.00450781499966979, + 0.018410800024867058, + 0.012966101057827473, + 0.0043780808337032795, + 0.004480196628719568, + 0.004263092298060656, + 0.0027861492708325386, + 0.003375181695446372, + 0.0027279453352093697, + 0.0046978965401649475, + 0.004462235141545534, + 0.008046416565775871, + 0.005085578188300133, + 0.002499031601473689, + 0.006107134744524956, + 0.0041315858252346516, + 0.01348947361111641, + 0.05322062969207764, + 0.034226641058921814, + 0.010904379189014435, + 0.01272560190409422, + 0.006471770349889994, + 0.005821788217872381, + 0.004276387393474579, + 0.006477704271674156, + 0.021577440202236176, + 0.013879986479878426, + 0.011484747752547264, + 0.012682241387665272, + 0.0062306723557412624, + 0.00932412501424551, + 0.00558545533567667, + 0.008657250553369522, + 0.011400675401091576, + 0.013969105668365955, + 0.01200597919523716, + 0.009523770771920681, + 0.008027581498026848, + 0.008207101374864578, + 0.0051691411063075066, + 0.00486582750454545, + 0.007283579558134079, + 0.00802538637071848, + 0.006899408996105194, + 0.005298899952322245, + 0.003486440982669592, + 0.004577350337058306, + 0.003542226506397128, + 0.04643069580197334, + 0.13316239416599274, + 0.059104979038238525, + 0.07100611180067062, + 0.052860863506793976, + 0.04321559891104698, + 0.03784122318029404, + 0.03523210436105728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.02333465963602066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111884627491236 + ], + "opposite_logprob_tokens": [ + -5.5041117668151855 + ], + "logprob_reward": -0.004111884627491236, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111884627491236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09201303124427795, + 0.04849272966384888, + 0.049493540078401566, + 0.03589072823524475, + 0.027021216228604317, + 0.41426655650138855, + 0.04477648809552193, + 0.10403314232826233, + 0.07666223496198654, + 0.10735033452510834 + ], + "credit_map": [ + 0.014458796940743923, + 0.013763400726020336, + 0.00902267824858427, + 0.00903143361210823, + 0.00976825226098299, + 0.015083950012922287, + 0.007914120331406593, + 0.012970397248864174, + 0.00514608109369874, + 0.006362710148096085, + 0.0050329966470599174, + 0.005920806899666786, + 0.005131922662258148, + 0.008049826137721539, + 0.006706558167934418, + 0.00614182697609067, + 0.0039758444763720036, + 0.008325564675033092, + 0.005289802793413401, + 0.0066307843662798405, + 0.00852581113576889, + 0.005046555772423744, + 0.006096929311752319, + 0.005602243822067976, + 0.002843721304088831, + 0.005468817427754402, + 0.003905004821717739, + 0.006175965070724487, + 0.004842177964746952, + 0.004460539203137159, + 0.0027791797183454037, + 0.005415323656052351, + 0.0026812402065843344, + 0.004729820881038904, + 0.0029947739094495773, + 0.004295869264751673, + 0.002311288844794035, + 0.0032635647803545, + 0.002425501123070717, + 0.004319158382713795, + 0.037015967071056366, + 0.15866132080554962, + 0.12068486958742142, + 0.027347499504685402, + 0.029542723670601845, + 0.017107782885432243, + 0.013498451560735703, + 0.010407893918454647, + 0.0029948465526103973, + 0.0066692219115793705, + 0.008465480990707874, + 0.008112537674605846, + 0.004939834587275982, + 0.0026678640861064196, + 0.007955018430948257, + 0.002971684094518423, + 0.00840726774185896, + 0.014955156482756138, + 0.01285428274422884, + 0.017190957441926003, + 0.018507622182369232, + 0.010803821496665478, + 0.012739903293550014, + 0.008574124425649643, + 0.006172486115247011, + 0.015919966623187065, + 0.01639135368168354, + 0.009437124244868755, + 0.008319445885717869, + 0.008935028687119484, + 0.006034964695572853, + 0.00545186921954155, + 0.011171522550284863, + 0.019457299262285233, + 0.014781806617975235, + 0.016957145184278488, + 0.012829208746552467, + 0.010465764440596104, + 0.013975299894809723, + 0.0077122836373746395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.004111884627491236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002824010793119669 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.002824010793119669, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002824010793119669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04146946966648102, + 0.021405894309282303, + 0.06724898517131805, + 0.1295473873615265, + 0.0396285243332386, + 0.20504894852638245, + 0.18012522161006927, + 0.0961180329322815, + 0.0534176267683506, + 0.1659899204969406 + ], + "credit_map": [ + 0.004745069425553083, + 0.005220222752541304, + 0.00593947246670723, + 0.005084541160613298, + 0.00547519838437438, + 0.0038173196371644735, + 0.004896339029073715, + 0.006291309837251902, + 0.0030514022801071405, + 0.0035217883996665478, + 0.0019366929773241282, + 0.00250190868973732, + 0.003157728584483266, + 0.00237890612334013, + 0.001827344298362732, + 0.0030301229562610388, + 0.006886124610900879, + 0.013580212369561195, + 0.007969193160533905, + 0.0066562616266310215, + 0.00842552911490202, + 0.008816097863018513, + 0.005108827259391546, + 0.009806746616959572, + 0.008429104462265968, + 0.018622372299432755, + 0.019258642569184303, + 0.02438032440841198, + 0.019655777141451836, + 0.008488603867590427, + 0.016299614682793617, + 0.014412947930395603, + 0.003016640665009618, + 0.007980834692716599, + 0.006757583934813738, + 0.007036714814603329, + 0.004391687922179699, + 0.0028019919991493225, + 0.00384608656167984, + 0.0037969830445945263, + 0.018029330298304558, + 0.0713064968585968, + 0.06907954066991806, + 0.015894291922450066, + 0.011127734556794167, + 0.0063096159137785435, + 0.007107736077159643, + 0.00619420874863863, + 0.012617701664566994, + 0.03285142406821251, + 0.019778383895754814, + 0.02454194612801075, + 0.03785127401351929, + 0.013474566861987114, + 0.020867595449090004, + 0.018142325803637505, + 0.006320721469819546, + 0.017126653343439102, + 0.013630018569529057, + 0.014470533467829227, + 0.01607751101255417, + 0.009889115579426289, + 0.010683171451091766, + 0.007920312695205212, + 0.005087381694465876, + 0.008463037200272083, + 0.007253157906234264, + 0.007213098928332329, + 0.008801272138953209, + 0.0047064577229321, + 0.006691980175673962, + 0.005201239604502916, + 0.014728548005223274, + 0.024244878441095352, + 0.028983034193515778, + 0.023086057975888252, + 0.024649418890476227, + 0.014343122020363808, + 0.02241203747689724, + 0.013542814180254936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.002824010793119669 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350026950240135 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026350026950240135, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350026950240135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09982993453741074, + 0.04314656928181648, + 0.028141651302576065, + 0.054109904915094376, + 0.05147213116288185, + 0.10738500952720642, + 0.1639997363090515, + 0.22871704399585724, + 0.05167055502533913, + 0.17152740061283112 + ], + "credit_map": [ + 0.013217530213296413, + 0.013210366480052471, + 0.010114029049873352, + 0.011617995798587799, + 0.013778047636151314, + 0.011350013315677643, + 0.011601765640079975, + 0.014940185472369194, + 0.0037660780362784863, + 0.005800623446702957, + 0.004100167658179998, + 0.006759636104106903, + 0.008115875534713268, + 0.003968413453549147, + 0.005336910020560026, + 0.005298861302435398, + 0.0029070188757032156, + 0.005105533637106419, + 0.003123484319075942, + 0.0030758955981582403, + 0.005720988847315311, + 0.0026969655882567167, + 0.0023797040339559317, + 0.003132059471681714, + 0.003823446109890938, + 0.007408349309116602, + 0.006390378810465336, + 0.008966303430497646, + 0.009670483879745007, + 0.006010976620018482, + 0.006080168299376965, + 0.005759797524660826, + 0.0036554753314703703, + 0.007371797226369381, + 0.006932164076715708, + 0.008906076662242413, + 0.006046521011739969, + 0.005136324092745781, + 0.006392987444996834, + 0.007030784618109465, + 0.009202338755130768, + 0.03709397092461586, + 0.02818799950182438, + 0.00800678227096796, + 0.008215607143938541, + 0.00754697946831584, + 0.004773847293108702, + 0.004357478115707636, + 0.013711346313357353, + 0.03988059237599373, + 0.021436549723148346, + 0.01844775676727295, + 0.028232814744114876, + 0.010835431516170502, + 0.017857158556580544, + 0.013598103076219559, + 0.020718393847346306, + 0.03692547604441643, + 0.03614511340856552, + 0.034948792308568954, + 0.033237189054489136, + 0.024398496374487877, + 0.02462664619088173, + 0.017716936767101288, + 0.008418161422014236, + 0.007187018170952797, + 0.010506470687687397, + 0.006131970323622227, + 0.0053775981068611145, + 0.005810742266476154, + 0.004294462036341429, + 0.003944129683077335, + 0.015661925077438354, + 0.02444854937493801, + 0.02584155648946762, + 0.027074091136455536, + 0.022756921127438545, + 0.016455447301268578, + 0.026929117739200592, + 0.0123598026111722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.026350026950240135 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, shot from a slightly elevated angle, captures a perfect, pale yellow rectangle of cold butter, placed in the center of a hot, black, seasoned cast iron skillet. Warm, direct overhead light casts a soft sheen on the butter's smooth surface and creates subtle highlights on the pan's textured matte finish. Almost immediately, the butter's sharp edges begin to soften and turn translucent, as a shimmering golden liquid rapidly pools at its base and spreads outward. The solid structure gracefully collapses and shrinks, its shape dissolving as it melts from below, becoming a tiny island within its own expanding pool. Tiny, hissing bubbles emerge and dance on the liquid's surface, which steadily spreads across the dark pan, reflecting light like a golden mirror, creating a captivating, satisfying atmosphere completely focused on physical transformation.", + "seed": 15957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_8/index_7634_seed_15957_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02044794587418437, + "vlm_logprob": -0.02044794587418437, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02044794587418437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011284103617072105 + ], + "opposite_logprob_tokens": [ + -5.636283874511719 + ], + "logprob_reward": -0.011284103617072105, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.011284103617072105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10590613633394241, + 0.0395234040915966, + 0.02035437896847725, + 0.021054707467556, + 0.016338832676410675, + 0.02782633900642395, + 0.10767710208892822, + 0.42759859561920166, + 0.17480045557022095, + 0.05892005190253258 + ], + "credit_map": [ + 0.014647630043327808, + 0.006733262445777655, + 0.015090582892298698, + 0.013939297758042812, + 0.006220097187906504, + 0.012739066034555435, + 0.009769970551133156, + 0.0267662163823843, + 0.004975044168531895, + 0.0032719201408326626, + 0.00450837891548872, + 0.00622568279504776, + 0.003093551844358444, + 0.005039796698838472, + 0.005781858693808317, + 0.00662717130035162, + 0.003027889411896467, + 0.0014198196586221457, + 0.0028059568721801043, + 0.0018712917808443308, + 0.001354923821054399, + 0.003347203368321061, + 0.003026266349479556, + 0.003501027589663863, + 0.00271357549354434, + 0.0017274253768846393, + 0.0029158182442188263, + 0.002197916153818369, + 0.00125921331346035, + 0.0038511536549776793, + 0.002644012216478586, + 0.003745592897757888, + 0.0018556405557319522, + 0.0011806886177510023, + 0.0022943958174437284, + 0.0015000718412920833, + 0.0016617189394310117, + 0.0032602890860289335, + 0.0013758878922089934, + 0.00321013992652297, + 0.005114070605486631, + 0.0028219108935445547, + 0.0030878023244440556, + 0.0017692577093839645, + 0.002499877940863371, + 0.004613101482391357, + 0.0018799506360664964, + 0.0060403659008443356, + 0.019336124882102013, + 0.01397782564163208, + 0.011428347788751125, + 0.005652199033647776, + 0.012163359671831131, + 0.016577396541833878, + 0.006048283074051142, + 0.022493571043014526, + 0.061147961765527725, + 0.06057201325893402, + 0.0465800017118454, + 0.02740943804383278, + 0.051880575716495514, + 0.06289947777986526, + 0.02892308123409748, + 0.08818605542182922, + 0.023551028221845627, + 0.012568770907819271, + 0.025102250277996063, + 0.026120241731405258, + 0.013592017814517021, + 0.024959983304142952, + 0.015474664978682995, + 0.03343150392174721, + 0.009766549803316593, + 0.005640209186822176, + 0.007357414811849594, + 0.005433779675513506, + 0.0037659350782632828, + 0.009521995671093464, + 0.006399217993021011, + 0.011034945957362652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.011284103617072105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018245598301291466 + ], + "opposite_logprob_tokens": [ + -4.018245697021484 + ], + "logprob_reward": -0.018245598301291466, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018245598301291466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018723100423812866, + 0.00920799933373928, + 0.007541805040091276, + 0.014338795095682144, + 0.02168070711195469, + 0.00942976400256157, + 0.6740953922271729, + 0.07036374509334564, + 0.1378719061613083, + 0.03674677386879921 + ], + "credit_map": [ + 0.002912051510065794, + 0.001025291858240962, + 0.0025765355676412582, + 0.0017837242921814322, + 0.0010094837052747607, + 0.003237525001168251, + 0.0014235482085496187, + 0.004754941910505295, + 0.0014740194892510772, + 0.0006918614963069558, + 0.001207814202643931, + 0.0008044886635616422, + 0.0004765935300383717, + 0.0017366159008815885, + 0.0007293978706002235, + 0.0020872089080512524, + 0.0009555922588333488, + 0.0008953001815825701, + 0.0010171807371079922, + 0.0008351209689863026, + 0.0006511473329737782, + 0.0008627436473034322, + 0.0008200198644772172, + 0.0015047008637338877, + 0.0028173865284770727, + 0.0007396845030598342, + 0.0019376382697373629, + 0.0008536434033885598, + 0.0006293323240242898, + 0.0029219151474535465, + 0.0009287425200454891, + 0.0035104539711028337, + 0.004241196904331446, + 0.0020864924881607294, + 0.002534877508878708, + 0.0014420700026676059, + 0.002109701745212078, + 0.003463095286861062, + 0.001503722625784576, + 0.004299554508179426, + 0.001028655911795795, + 0.0008970710914582014, + 0.001185322762466967, + 0.0013245029840618372, + 0.0009537197183817625, + 0.00120087293908, + 0.0008276121225208044, + 0.0020120074041187763, + 0.12767761945724487, + 0.07272694259881973, + 0.07292037457227707, + 0.030592015013098717, + 0.068423792719841, + 0.1124969944357872, + 0.03610299155116081, + 0.15315468609333038, + 0.010444076731801033, + 0.008552229963243008, + 0.007858498953282833, + 0.004079318605363369, + 0.0072895316407084465, + 0.010552984662353992, + 0.005072688218206167, + 0.016514424234628677, + 0.018742041662335396, + 0.010336765088140965, + 0.0187168437987566, + 0.020672932267189026, + 0.01141777541488409, + 0.020520927384495735, + 0.012398363091051579, + 0.025066262111067772, + 0.005730327684432268, + 0.0024211807176470757, + 0.005862021818757057, + 0.0035540589597076178, + 0.00167184229940176, + 0.004510125610977411, + 0.005357217974960804, + 0.007640000432729721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.018245598301291466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005273835267871618 + ], + "opposite_logprob_tokens": [ + -5.255273818969727 + ], + "logprob_reward": -0.005273835267871618, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005273835267871618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0422331765294075, + 0.02044512704014778, + 0.009867224842309952, + 0.010624350048601627, + 0.013413896784186363, + 0.017807425931096077, + 0.3044736087322235, + 0.4647572934627533, + 0.04341702163219452, + 0.07296092063188553 + ], + "credit_map": [ + 0.006036146078258753, + 0.002294989302754402, + 0.00653511518612504, + 0.0036684630904346704, + 0.00220264820381999, + 0.008109375834465027, + 0.0034643581602722406, + 0.009922082535922527, + 0.0026322200428694487, + 0.0015576182631775737, + 0.002752457745373249, + 0.0032534035854041576, + 0.0016400020103901625, + 0.0035891495645046234, + 0.0022628088481724262, + 0.0027574680279940367, + 0.0014056682121008635, + 0.000753865868318826, + 0.0017377760959789157, + 0.0008501323754899204, + 0.0005860453238710761, + 0.0015648555709049106, + 0.000813179649412632, + 0.0021557018626481295, + 0.0014234959380701184, + 0.0008174092508852482, + 0.0019094268791377544, + 0.0008421365055255592, + 0.0006359575199894607, + 0.002587568014860153, + 0.0008594578830525279, + 0.0015488985227420926, + 0.002168151782825589, + 0.001081614987924695, + 0.0016092393780127168, + 0.001041391515173018, + 0.0012190884444862604, + 0.0028271616902202368, + 0.0009321194957010448, + 0.0025351298972964287, + 0.00381724932231009, + 0.000988071202300489, + 0.002508848439902067, + 0.0009279063669964671, + 0.0009336293442174792, + 0.003599173855036497, + 0.0010847258381545544, + 0.003947821911424398, + 0.05783648416399956, + 0.03301318734884262, + 0.03299941122531891, + 0.013940190896391869, + 0.03087714873254299, + 0.050464414060115814, + 0.016109555959701538, + 0.06923321634531021, + 0.06580061465501785, + 0.06641334295272827, + 0.050698596984148026, + 0.03051193244755268, + 0.055877573788166046, + 0.06868042051792145, + 0.03134671226143837, + 0.09542809426784515, + 0.006137863267213106, + 0.0038100501988083124, + 0.00537148816511035, + 0.005450281780213118, + 0.0037944260984659195, + 0.0074781752191483974, + 0.004923681728541851, + 0.006451054476201534, + 0.007048307918012142, + 0.006978942081332207, + 0.01072053425014019, + 0.009340209886431694, + 0.005959169939160347, + 0.007942216470837593, + 0.012481465935707092, + 0.012490078806877136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.005273835267871618 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372357040643692 + ], + "opposite_logprob_tokens": [ + -3.4087235927581787 + ], + "logprob_reward": -0.03372357040643692, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372357040643692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01774127036333084, + 0.01315685361623764, + 0.006120521109551191, + 0.005809110589325428, + 0.007787507958710194, + 0.009850461967289448, + 0.37459349632263184, + 0.45537859201431274, + 0.030790647491812706, + 0.07877150923013687 + ], + "credit_map": [ + 0.0030385733116418123, + 0.0011041904799640179, + 0.002474068198353052, + 0.001531468820758164, + 0.0011390475556254387, + 0.0025319471023976803, + 0.0013287147739902139, + 0.004593259654939175, + 0.0015842864522710443, + 0.00101918401196599, + 0.0018104512710124254, + 0.0019329923670738935, + 0.0010848372476175427, + 0.0024605412036180496, + 0.001426178845576942, + 0.0018383811693638563, + 0.001144333160482347, + 0.0004175168287474662, + 0.0008898145752027631, + 0.0003868571075145155, + 0.00041207423782907426, + 0.0009461427107453346, + 0.0004836635198444128, + 0.00144011911470443, + 0.000844852183945477, + 0.00038202019641175866, + 0.0007767100469209254, + 0.0004900682251900434, + 0.0003777222300413996, + 0.0013531798031181097, + 0.0004944963729940355, + 0.0010900618508458138, + 0.0017008153954520822, + 0.0005215888959355652, + 0.0009412106592208147, + 0.0004866820527240634, + 0.0004758248687721789, + 0.0015984154306352139, + 0.0004796718421857804, + 0.0015832986682653427, + 0.00192751525901258, + 0.0006898259161971509, + 0.0013247147435322404, + 0.0006430511130020022, + 0.0006026047631166875, + 0.0016518178163096309, + 0.0007986864075064659, + 0.00221224594861269, + 0.07071629166603088, + 0.040628451853990555, + 0.0407935194671154, + 0.016920728608965874, + 0.03811822459101677, + 0.06218944489955902, + 0.020160896703600883, + 0.08506591618061066, + 0.06493674218654633, + 0.06467849761247635, + 0.04888841137290001, + 0.029387522488832474, + 0.054397448897361755, + 0.0681268721818924, + 0.030168434605002403, + 0.09479467570781708, + 0.004265987779945135, + 0.00297728693112731, + 0.0037791382055729628, + 0.0038346112705767155, + 0.002670260611921549, + 0.005405706353485584, + 0.0036316660698503256, + 0.004225988872349262, + 0.008321965113282204, + 0.0070271119475364685, + 0.012643372640013695, + 0.010712538845837116, + 0.006290499120950699, + 0.008461417630314827, + 0.014751343056559563, + 0.010563257150352001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.03372357040643692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371262177824974 + ], + "opposite_logprob_tokens": [ + -3.40871262550354 + ], + "logprob_reward": -0.03371262177824974, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371262177824974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01633119210600853, + 0.012468834407627583, + 0.005179946310818195, + 0.0036510194186121225, + 0.006656355690211058, + 0.006102260202169418, + 0.4190601408481598, + 0.4336821436882019, + 0.03587668761610985, + 0.06099141761660576 + ], + "credit_map": [ + 0.0018686080584302545, + 0.001200063736177981, + 0.0026011725421994925, + 0.0019421990728005767, + 0.001360382535494864, + 0.002070532413199544, + 0.002002415945753455, + 0.003285816637799144, + 0.0014840770745649934, + 0.0009839726844802499, + 0.00173051783349365, + 0.0017731830012053251, + 0.0010803258046507835, + 0.0024970017839223146, + 0.0012368718162178993, + 0.0016828830121085048, + 0.0009339993703179061, + 0.00035408325493335724, + 0.0007295896648429334, + 0.00036362343234941363, + 0.0004029714036732912, + 0.0008004018454812467, + 0.00046316723455674946, + 0.0011321095516905189, + 0.0004933523596264422, + 0.0002791265142150223, + 0.0004904115339741111, + 0.0005362344672903419, + 0.0003418104024603963, + 0.0005469014868140221, + 0.0004614500794559717, + 0.0005017320509068668, + 0.001289975829422474, + 0.0004636499797925353, + 0.0008401671657338738, + 0.0006652479059994221, + 0.00047859971527941525, + 0.0011892602778971195, + 0.00046484824270009995, + 0.0012646059039980173, + 0.001012571738101542, + 0.0004164772399235517, + 0.0008751895511522889, + 0.0005708753014914691, + 0.00043245829874649644, + 0.0009186563547700644, + 0.0005310814012773335, + 0.001344949472695589, + 0.07919694483280182, + 0.04551580548286438, + 0.045643970370292664, + 0.019058799371123314, + 0.042564790695905685, + 0.06957843154668808, + 0.022416846826672554, + 0.0950845405459404, + 0.06187606230378151, + 0.061642155051231384, + 0.04668938368558884, + 0.02804727852344513, + 0.05196867138147354, + 0.06489622592926025, + 0.02886122837662697, + 0.08970112353563309, + 0.005006363615393639, + 0.003200407139956951, + 0.00452288705855608, + 0.0047986856661736965, + 0.0030772057361900806, + 0.006008475553244352, + 0.003997542895376682, + 0.0052651166915893555, + 0.006587544456124306, + 0.005473831668496132, + 0.009873298928141594, + 0.007888083346188068, + 0.004594291094690561, + 0.006549306213855743, + 0.011208640411496162, + 0.008816416375339031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.03371262177824974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, shot from a slightly elevated angle, captures a perfect, pale yellow rectangle of cold butter, placed in the center of a hot, black, seasoned cast iron skillet. Warm, direct overhead light casts a soft sheen on the butter's smooth surface and creates subtle highlights on the pan's textured matte finish. Almost immediately, the butter's sharp edges begin to soften and turn translucent, as a shimmering golden liquid rapidly pools at its base and spreads outward. The solid structure gracefully collapses and shrinks, its shape dissolving as it melts from below, becoming a tiny island within its own expanding pool. Tiny, hissing bubbles emerge and dance on the liquid's surface, which steadily spreads across the dark pan, reflecting light like a golden mirror, creating a captivating, satisfying atmosphere completely focused on physical transformation.", + "seed": 15958, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_8/index_7634_seed_15958_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011732615623623132, + "vlm_logprob": -0.011732615623623132, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.011732615623623132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007906214334070683 + ], + "opposite_logprob_tokens": [ + -6.257906436920166 + ], + "logprob_reward": -0.007906214334070683, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.007906214334070683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09131336957216263, + 0.031199069693684578, + 0.017538873478770256, + 0.024092093110084534, + 0.05441580340266228, + 0.027373304590582848, + 0.34313783049583435, + 0.05266249552369118, + 0.10791081190109253, + 0.2503563165664673 + ], + "credit_map": [ + 0.008855951949954033, + 0.007750635966658592, + 0.022321276366710663, + 0.01040299329906702, + 0.007621572352945805, + 0.009917943738400936, + 0.009829116053879261, + 0.01461388636380434, + 0.0030452923383563757, + 0.00296516390517354, + 0.0037953597493469715, + 0.00720964977517724, + 0.0027488640043884516, + 0.0030321795493364334, + 0.0050120302475988865, + 0.0033905315212905407, + 0.002090942580252886, + 0.0014037933433428407, + 0.0026079181116074324, + 0.00286274217069149, + 0.0014738718746230006, + 0.0020604629535228014, + 0.0030032016802579165, + 0.0020359433256089687, + 0.0037989711854606867, + 0.0021884809248149395, + 0.002835669321939349, + 0.00371026829816401, + 0.0016694237710908055, + 0.0027773305773735046, + 0.0032523174304515123, + 0.0038596345111727715, + 0.006794667802751064, + 0.0045233177952468395, + 0.008965050801634789, + 0.008589284494519234, + 0.003909311257302761, + 0.0070976079441607, + 0.007340320385992527, + 0.0071962494403123856, + 0.0036579500883817673, + 0.0031658969819545746, + 0.0031219255179166794, + 0.0035460512153804302, + 0.0026150690391659737, + 0.002688761567696929, + 0.0035957363434135914, + 0.004981917329132557, + 0.04878599941730499, + 0.06050853803753853, + 0.036828964948654175, + 0.03378598764538765, + 0.05352684482932091, + 0.03167309984564781, + 0.026961732655763626, + 0.05106668919324875, + 0.0073581659235060215, + 0.005389940459281206, + 0.011079000309109688, + 0.0059007927775382996, + 0.00469575822353363, + 0.004316958133131266, + 0.005446193739771843, + 0.008475691080093384, + 0.010119527578353882, + 0.016419196501374245, + 0.02936323545873165, + 0.012102268636226654, + 0.009767442010343075, + 0.006685899570584297, + 0.013755513355135918, + 0.009697738103568554, + 0.026500631123781204, + 0.06706435978412628, + 0.026001567021012306, + 0.03865932300686836, + 0.035195592790842056, + 0.013777540065348148, + 0.024400891736149788, + 0.01875643990933895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.007906214334070683 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01110810600221157 + ], + "opposite_logprob_tokens": [ + -4.511107921600342 + ], + "logprob_reward": -0.01110810600221157, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01110810600221157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07507698237895966, + 0.06556171178817749, + 0.047568656504154205, + 0.0849384143948555, + 0.06876534223556519, + 0.03459024801850319, + 0.164993017911911, + 0.07877177000045776, + 0.12878058850765228, + 0.25095322728157043 + ], + "credit_map": [ + 0.01027707289904356, + 0.005569204688072205, + 0.011218228377401829, + 0.014385731890797615, + 0.005657399073243141, + 0.006706234999001026, + 0.011233219876885414, + 0.010029885917901993, + 0.005889159627258778, + 0.005786030553281307, + 0.00915740430355072, + 0.014948784373700619, + 0.006174280773848295, + 0.006234743632376194, + 0.010233919136226177, + 0.007137392181903124, + 0.004843292757868767, + 0.004001246299594641, + 0.009535624645650387, + 0.0060104201547801495, + 0.003707040101289749, + 0.004074646160006523, + 0.008588834665715694, + 0.0068075512535870075, + 0.008705546148121357, + 0.005751905031502247, + 0.01138046570122242, + 0.010975331999361515, + 0.005353956948965788, + 0.012806423008441925, + 0.016473770141601562, + 0.01349102146923542, + 0.006178842857480049, + 0.0047903587110340595, + 0.009862965904176235, + 0.01103876531124115, + 0.004934131633490324, + 0.008405868895351887, + 0.013736049644649029, + 0.009818358346819878, + 0.0030635055154561996, + 0.003140218323096633, + 0.0039020502008497715, + 0.006163503509014845, + 0.003167039481922984, + 0.0035521413665264845, + 0.006198478396981955, + 0.005403308197855949, + 0.026129357516765594, + 0.01802925579249859, + 0.024742236360907555, + 0.0181020088493824, + 0.019930033013224602, + 0.01642175391316414, + 0.015624689869582653, + 0.026013681665062904, + 0.009797538630664349, + 0.008700543083250523, + 0.01874326728284359, + 0.009782173670828342, + 0.009807096794247627, + 0.006091092247515917, + 0.0068290275521576405, + 0.009021028876304626, + 0.013627247884869576, + 0.0205873753875494, + 0.014583567157387733, + 0.022705163806676865, + 0.013543879613280296, + 0.008511226624250412, + 0.023650920018553734, + 0.011571217328310013, + 0.03228062763810158, + 0.025727400556206703, + 0.06622684746980667, + 0.032065678387880325, + 0.02229376509785652, + 0.01531144231557846, + 0.022651659324765205, + 0.03439580276608467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.01110810600221157 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004655356053262949 + ], + "opposite_logprob_tokens": [ + -5.379655361175537 + ], + "logprob_reward": -0.004655356053262949, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004655356053262949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06053374707698822, + 0.03450575843453407, + 0.02067701332271099, + 0.040372371673583984, + 0.03638531267642975, + 0.034282341599464417, + 0.33282044529914856, + 0.042621318250894547, + 0.13067837059497833, + 0.2671233117580414 + ], + "credit_map": [ + 0.007075842469930649, + 0.004823089577257633, + 0.012072328478097916, + 0.007505675777792931, + 0.005104187875986099, + 0.007227120455354452, + 0.0075971754267811775, + 0.009128326550126076, + 0.0034241836983710527, + 0.003004273399710655, + 0.005808659363538027, + 0.007101962808519602, + 0.00289733218960464, + 0.0036209237296134233, + 0.004422853235155344, + 0.004225568845868111, + 0.0019124048994854093, + 0.0017805491806939244, + 0.002748786471784115, + 0.0034141766373068094, + 0.0019291606731712818, + 0.0015897005796432495, + 0.00535338232293725, + 0.0019488512771204114, + 0.005005786195397377, + 0.0034589164424687624, + 0.006250145845115185, + 0.005999865476042032, + 0.0029334495775401592, + 0.0035056984052062035, + 0.007635077927261591, + 0.005583430640399456, + 0.003912594169378281, + 0.0037442687898874283, + 0.006652034353464842, + 0.0065343971364200115, + 0.002757916459813714, + 0.0033355001360177994, + 0.0048270318657159805, + 0.004621568135917187, + 0.0032354064751416445, + 0.004376998636871576, + 0.004373563919216394, + 0.0050283391028642654, + 0.0038106937427073717, + 0.003383347997441888, + 0.004817422479391098, + 0.005256569012999535, + 0.054224416613578796, + 0.048460133373737335, + 0.049316976219415665, + 0.03159483149647713, + 0.043904319405555725, + 0.030697330832481384, + 0.02220572531223297, + 0.05241672694683075, + 0.007079108152538538, + 0.0034756904933601618, + 0.009507078677415848, + 0.004884233698248863, + 0.004145655781030655, + 0.003486515488475561, + 0.004465755075216293, + 0.005577282980084419, + 0.010094662196934223, + 0.023454738780856133, + 0.013175074942409992, + 0.025423455983400345, + 0.01694517210125923, + 0.008941364474594593, + 0.02283642068505287, + 0.009807487949728966, + 0.027284419164061546, + 0.07069065421819687, + 0.03114086203277111, + 0.024707434698939323, + 0.044703394174575806, + 0.012519690208137035, + 0.021344242617487907, + 0.03473260626196861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.004655356053262949 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644304238259792 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644304238259792, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644304238259792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06195290386676788, + 0.018338311463594437, + 0.027001330628991127, + 0.04474300146102905, + 0.07244033366441727, + 0.047699589282274246, + 0.2207842320203781, + 0.038209687918424606, + 0.2548161745071411, + 0.21401439607143402 + ], + "credit_map": [ + 0.00763090280815959, + 0.00405333423987031, + 0.011451770551502705, + 0.00881011039018631, + 0.004199552349746227, + 0.006682112812995911, + 0.009571218863129616, + 0.009553899057209492, + 0.0021540916059166193, + 0.0017390730790793896, + 0.0029676330741494894, + 0.0027140500023961067, + 0.001528802327811718, + 0.0020707882940769196, + 0.0026909117586910725, + 0.00247296248562634, + 0.003487061243504286, + 0.0027801361866295338, + 0.0029502478428184986, + 0.004908855073153973, + 0.0025475004222244024, + 0.0019478691974654794, + 0.005136677995324135, + 0.0032429846469312906, + 0.004767091944813728, + 0.0037179458886384964, + 0.00593365915119648, + 0.007773314602673054, + 0.0028783371672034264, + 0.004935786593705416, + 0.009080503135919571, + 0.005656362511217594, + 0.007250586524605751, + 0.00543465930968523, + 0.011412518098950386, + 0.013618255965411663, + 0.004683786071836948, + 0.007471777964383364, + 0.012785467319190502, + 0.009783281944692135, + 0.0043490249663591385, + 0.004590767901390791, + 0.007370630744844675, + 0.007159598637372255, + 0.004598961677402258, + 0.005731917917728424, + 0.006088797468692064, + 0.00780988996848464, + 0.037047840654850006, + 0.02820131927728653, + 0.034888576716184616, + 0.022426122799515724, + 0.025267090648412704, + 0.020587405189871788, + 0.016980549320578575, + 0.03538532555103302, + 0.007274404168128967, + 0.0030241587664932013, + 0.007776731159538031, + 0.004411650355905294, + 0.0037334237713366747, + 0.0032532992772758007, + 0.003451545722782612, + 0.005284472834318876, + 0.018925096839666367, + 0.045646458864212036, + 0.028302473947405815, + 0.048680443316698074, + 0.03429235517978668, + 0.016056228429079056, + 0.0432182252407074, + 0.019694916903972626, + 0.01422678492963314, + 0.058673031628131866, + 0.06418339908123016, + 0.01722448132932186, + 0.022725045680999756, + 0.008230484090745449, + 0.015654733404517174, + 0.013096444308757782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.008644304238259792 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634909749031067 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.02634909749031067, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634909749031067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07014236599206924, + 0.024449048563838005, + 0.028218897059559822, + 0.030098069459199905, + 0.057698171585798264, + 0.039476510137319565, + 0.2610276937484741, + 0.07286731898784637, + 0.19113385677337646, + 0.22488802671432495 + ], + "credit_map": [ + 0.00734522333368659, + 0.005282520316541195, + 0.016078762710094452, + 0.008187132887542248, + 0.005348688922822475, + 0.008768389001488686, + 0.00813987199217081, + 0.010991785675287247, + 0.002429765183478594, + 0.0027931055519729853, + 0.0032585046719759703, + 0.0048020160757005215, + 0.002364915329962969, + 0.0025835351552814245, + 0.003018361981958151, + 0.0031988469418138266, + 0.0030927653424441814, + 0.0028832119423896074, + 0.0033190285321325064, + 0.005506353452801704, + 0.002802976407110691, + 0.0020014820620417595, + 0.005587225779891014, + 0.0030258558690547943, + 0.0034663856495171785, + 0.0026609308551996946, + 0.0041167885065078735, + 0.005194602068513632, + 0.0020035284105688334, + 0.0026856970507651567, + 0.006277844309806824, + 0.0036922944709658623, + 0.00625908188521862, + 0.004313831217586994, + 0.009332723915576935, + 0.01113935187458992, + 0.0036735059693455696, + 0.006174277514219284, + 0.009068111889064312, + 0.007737292442470789, + 0.0033305382821708918, + 0.0036237938329577446, + 0.005733652971684933, + 0.00667658681049943, + 0.003442624118179083, + 0.004270261153578758, + 0.005789636634290218, + 0.006609419826418161, + 0.04678875580430031, + 0.030204351991415024, + 0.04492528736591339, + 0.024475207552313805, + 0.027101172134280205, + 0.024621615186333656, + 0.019069204106926918, + 0.043842118233442307, + 0.014202311635017395, + 0.0073173763230443, + 0.013125027529895306, + 0.007754280231893063, + 0.006569032557308674, + 0.005512700881808996, + 0.006700143218040466, + 0.01168645266443491, + 0.015156206674873829, + 0.03315811604261398, + 0.023295322433114052, + 0.03528926149010658, + 0.02340974658727646, + 0.013098012655973434, + 0.03297882527112961, + 0.014748399145901203, + 0.017497949302196503, + 0.04914170131087303, + 0.07273206114768982, + 0.020197806879878044, + 0.01604139246046543, + 0.010279460810124874, + 0.017682332545518875, + 0.021315349265933037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.02634909749031067 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, shot from a slightly elevated angle, captures a perfect, pale yellow rectangle of cold butter, placed in the center of a hot, black, seasoned cast iron skillet. Warm, direct overhead light casts a soft sheen on the butter's smooth surface and creates subtle highlights on the pan's textured matte finish. Almost immediately, the butter's sharp edges begin to soften and turn translucent, as a shimmering golden liquid rapidly pools at its base and spreads outward. The solid structure gracefully collapses and shrinks, its shape dissolving as it melts from below, becoming a tiny island within its own expanding pool. Tiny, hissing bubbles emerge and dance on the liquid's surface, which steadily spreads across the dark pan, reflecting light like a golden mirror, creating a captivating, satisfying atmosphere completely focused on physical transformation.", + "seed": 15959, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_8/index_7634_seed_15959_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012304328568279743, + "vlm_logprob": -0.012304328568279743, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.012304328568279743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012763594277203083 + ], + "opposite_logprob_tokens": [ + -5.512763500213623 + ], + "logprob_reward": -0.012763594277203083, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.012763594277203083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0712881088256836, + 0.33509883284568787, + 0.020687555894255638, + 0.0844675824046135, + 0.2975117862224579, + 0.06891132891178131, + 0.02359985001385212, + 0.011837754398584366, + 0.06271650642156601, + 0.023880677297711372 + ], + "credit_map": [ + 0.02270611748099327, + 0.004617113154381514, + 0.008369476534426212, + 0.0037424673791974783, + 0.005049295723438263, + 0.009994708932936192, + 0.0051321592181921005, + 0.011676767840981483, + 0.010691117495298386, + 0.004853904712945223, + 0.010847006924450397, + 0.004226377699524164, + 0.02307143807411194, + 0.14464561641216278, + 0.010493558831512928, + 0.12626983225345612, + 0.002500600414350629, + 0.0015666200779378414, + 0.002451502252370119, + 0.0009898249991238117, + 0.006454148795455694, + 0.0016902080969884992, + 0.001754396827891469, + 0.003280254080891609, + 0.0030073102097958326, + 0.0037114284932613373, + 0.004611932206898928, + 0.0016107108676806092, + 0.021391039714217186, + 0.031175771728157997, + 0.0026691234670579433, + 0.016290271654725075, + 0.012583744712173939, + 0.009530984796583652, + 0.00980477873235941, + 0.006308555603027344, + 0.01072920672595501, + 0.13622607290744781, + 0.0066731153056025505, + 0.10565534234046936, + 0.0037584356032311916, + 0.003951405640691519, + 0.00409254664555192, + 0.0016965410904958844, + 0.01866866834461689, + 0.023049985989928246, + 0.003273383714258671, + 0.01042036060243845, + 0.0021599349565804005, + 0.0030388610903173685, + 0.0027010205667465925, + 0.0027240244671702385, + 0.003852267749607563, + 0.0036612050607800484, + 0.0026168955955654383, + 0.0028456407599151134, + 0.000777247769292444, + 0.0015298069920390844, + 0.0012797184754163027, + 0.0009505048510618508, + 0.0022602961398661137, + 0.0016113080782815814, + 0.0016025479417294264, + 0.0018263239180669188, + 0.00493675097823143, + 0.006018012762069702, + 0.003517559729516506, + 0.0019772755913436413, + 0.01920158416032791, + 0.005744176916778088, + 0.004282298032194376, + 0.017038848251104355, + 0.0016400692984461784, + 0.003004183294251561, + 0.0027656524907797575, + 0.003412491176277399, + 0.005421565379947424, + 0.0025261470582336187, + 0.0035782880149781704, + 0.0015322829131036997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.012763594277203083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016100158914923668 + ], + "opposite_logprob_tokens": [ + -4.14109992980957 + ], + "logprob_reward": -0.016100158914923668, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016100158914923668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031443219631910324, + 0.22542165219783783, + 0.04980811849236488, + 0.08819513022899628, + 0.4156244695186615, + 0.051973406225442886, + 0.021982470527291298, + 0.015468467958271503, + 0.07852806150913239, + 0.021554961800575256 + ], + "credit_map": [ + 0.004519117530435324, + 0.0022608833387494087, + 0.004995040129870176, + 0.001740420120768249, + 0.003228840883821249, + 0.00525125116109848, + 0.002702992409467697, + 0.00674467533826828, + 0.007837600074708462, + 0.0035312394611537457, + 0.006759442389011383, + 0.0029879773501306772, + 0.005166397895663977, + 0.09931274503469467, + 0.0072217658162117004, + 0.09260449558496475, + 0.004813758190721273, + 0.0030696147587150335, + 0.004811062943190336, + 0.00273085362277925, + 0.011686859652400017, + 0.012067517265677452, + 0.0036493809893727303, + 0.006979074329137802, + 0.006693329196423292, + 0.0051580676808953285, + 0.005791176110506058, + 0.003257468342781067, + 0.017519013956189156, + 0.03025575913488865, + 0.0034405929036438465, + 0.016079725697636604, + 0.014943704940378666, + 0.017865344882011414, + 0.01878342777490616, + 0.011558703146874905, + 0.09675689786672592, + 0.15935465693473816, + 0.012064586393535137, + 0.08429720252752304, + 0.0033068519551306963, + 0.0039049095939844847, + 0.002681959420442581, + 0.0023543918505311012, + 0.006879557389765978, + 0.01786835677921772, + 0.00378861790522933, + 0.011188765056431293, + 0.002282623201608658, + 0.0023995598312467337, + 0.0018565811915323138, + 0.0018637661123648286, + 0.0034266228321939707, + 0.004877517931163311, + 0.002638711594045162, + 0.0026370896957814693, + 0.001943316776305437, + 0.0017967133317142725, + 0.002612142823636532, + 0.0012783349957317114, + 0.0019945797976106405, + 0.0014375860337167978, + 0.002737650880590081, + 0.0016681451816111803, + 0.006033819168806076, + 0.00726944487541914, + 0.0035529183223843575, + 0.002428994048386812, + 0.02350044623017311, + 0.008187639527022839, + 0.005045207217335701, + 0.02250959724187851, + 0.0016603483818471432, + 0.0016430364921689034, + 0.0021082486491650343, + 0.0023433207534253597, + 0.0020963081624358892, + 0.0054818554781377316, + 0.0021669829729944468, + 0.004054861608892679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.016100158914923668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110459703952074 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110459703952074, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110459703952074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055910736322402954, + 0.1587044596672058, + 0.046355653554201126, + 0.06143811345100403, + 0.40498048067092896, + 0.06146658957004547, + 0.039150092750787735, + 0.016351981088519096, + 0.09016880393028259, + 0.06547307223081589 + ], + "credit_map": [ + 0.012059868313372135, + 0.004936033394187689, + 0.00683100800961256, + 0.005772776436060667, + 0.00438934238627553, + 0.007733470294624567, + 0.005024739075452089, + 0.009163502603769302, + 0.009154723025858402, + 0.01841410994529724, + 0.012927595525979996, + 0.004635925404727459, + 0.05782240256667137, + 0.022837989032268524, + 0.015431215986609459, + 0.017480507493019104, + 0.005282399244606495, + 0.0036152314860373735, + 0.004802412819117308, + 0.0041445051319897175, + 0.006814921274781227, + 0.009890828281641006, + 0.005576379131525755, + 0.006228982005268335, + 0.004738929215818644, + 0.004031378775835037, + 0.003747747978195548, + 0.0024349805898964405, + 0.014703507535159588, + 0.018912237137556076, + 0.003272430505603552, + 0.009596905671060085, + 0.018337907269597054, + 0.012189391069114208, + 0.014357254840433598, + 0.007653817534446716, + 0.018250517547130585, + 0.18530502915382385, + 0.009050743654370308, + 0.13983581960201263, + 0.004317384213209152, + 0.005558077711611986, + 0.004207964986562729, + 0.004854483529925346, + 0.019101310521364212, + 0.013595642521977425, + 0.004926588851958513, + 0.004905139096081257, + 0.003580619115382433, + 0.003910092171281576, + 0.002956486539915204, + 0.0037889203522354364, + 0.006942200008779764, + 0.009500067681074142, + 0.004449451807886362, + 0.004022256005555391, + 0.0011645713821053505, + 0.0017849653959274292, + 0.002003644360229373, + 0.0013157454086467624, + 0.003167801769450307, + 0.0021278816275298595, + 0.0019518961198627949, + 0.00283547630533576, + 0.007598059717565775, + 0.005717444233596325, + 0.005532551556825638, + 0.005451884586364031, + 0.016142461448907852, + 0.013349312357604504, + 0.007629368454217911, + 0.028747739270329475, + 0.0043127164244651794, + 0.0029361664783209562, + 0.0029573077335953712, + 0.004164986778050661, + 0.0029063920956104994, + 0.023181956261396408, + 0.003351123072206974, + 0.021662427112460136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.004110459703952074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258656572550535 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258656572550535, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258656572550535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04086712375283241, + 0.18488267064094543, + 0.024978289380669594, + 0.10049302130937576, + 0.3833046853542328, + 0.042542606592178345, + 0.017803290858864784, + 0.04294845089316368, + 0.1288709193468094, + 0.03330889344215393 + ], + "credit_map": [ + 0.0039801462553441525, + 0.004813049919903278, + 0.0039911698549985886, + 0.00448403786867857, + 0.006732301786541939, + 0.005149023607373238, + 0.006160527467727661, + 0.005556870251893997, + 0.00783579982817173, + 0.016300147399306297, + 0.012214823625981808, + 0.003761552507057786, + 0.054656997323036194, + 0.052077822387218475, + 0.012131880037486553, + 0.02590365521609783, + 0.0017574153607711196, + 0.0015299965161830187, + 0.0020697324071079493, + 0.0014341858914121985, + 0.006190718617290258, + 0.006634108256548643, + 0.0020119575783610344, + 0.0033501731231808662, + 0.006019541062414646, + 0.005124888848513365, + 0.005954289343208075, + 0.0024701671209186316, + 0.024637624621391296, + 0.035481199622154236, + 0.003459359286352992, + 0.01734594814479351, + 0.016953714191913605, + 0.014016657136380672, + 0.012762065045535564, + 0.007372237741947174, + 0.011205693706870079, + 0.17456698417663574, + 0.008469893597066402, + 0.13795745372772217, + 0.002728165127336979, + 0.004029048141092062, + 0.002900694729760289, + 0.0036721352953463793, + 0.01334513071924448, + 0.008264826610684395, + 0.004370298236608505, + 0.003232308430597186, + 0.0014420906081795692, + 0.002768272999674082, + 0.002037789672613144, + 0.0029300875030457973, + 0.0023443158715963364, + 0.0015674543101340532, + 0.0030678058974444866, + 0.0016454716678708792, + 0.003794151358306408, + 0.0038056999910622835, + 0.003956574015319347, + 0.0044612097553908825, + 0.007991316728293896, + 0.0034951434936374426, + 0.007730293087661266, + 0.007714061066508293, + 0.009715731255710125, + 0.012683187611401081, + 0.0061552212573587894, + 0.0034720213152468204, + 0.041353289037942886, + 0.011546382680535316, + 0.008443235419690609, + 0.03550185635685921, + 0.0019505879608914256, + 0.0016732360236346722, + 0.0013640362303704023, + 0.0023491091560572386, + 0.0013206364819779992, + 0.011499264277517796, + 0.002038896782323718, + 0.011113127693533897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.005258656572550535 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pale yellow rectangle of butter placed in the center of a black cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm, direct overhead light create a soft sheen on the butter and highlights on the pan's textured matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter's solid structure collapse and shrink as it melts from below, forming a tiny island within an expanding pool of golden liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny hissing bubbles emerge and dance on the surface of the spreading liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288773372769356 + ], + "opposite_logprob_tokens": [ + -3.7732887268066406 + ], + "logprob_reward": -0.023288773372769356, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288773372769356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019984783604741096, + 0.2127901166677475, + 0.04496210440993309, + 0.05159073323011398, + 0.3128405809402466, + 0.058127742260694504, + 0.011016535572707653, + 0.02441265434026718, + 0.2246142327785492, + 0.03966052085161209 + ], + "credit_map": [ + 0.0033082072623074055, + 0.0017344271764159203, + 0.0032494983170181513, + 0.0019010923570021987, + 0.0016888220561668277, + 0.002440241165459156, + 0.0027190372347831726, + 0.0029434561729431152, + 0.008073638193309307, + 0.016890037804841995, + 0.012906219810247421, + 0.003630660939961672, + 0.05740862339735031, + 0.06538955122232437, + 0.012138606980443, + 0.03635276481509209, + 0.0017964281141757965, + 0.0019105446990579367, + 0.0026855466421693563, + 0.0015459473943337798, + 0.009023739956319332, + 0.01591913029551506, + 0.0021027789916843176, + 0.009977985173463821, + 0.0035795376170426607, + 0.002811296144500375, + 0.0031804083846509457, + 0.001552152680233121, + 0.01271826308220625, + 0.017210977151989937, + 0.002379313576966524, + 0.008158777840435505, + 0.013756812550127506, + 0.011457976885139942, + 0.010251623578369617, + 0.005859298165887594, + 0.00902643520385027, + 0.14278164505958557, + 0.006759640760719776, + 0.11294712126255035, + 0.0032751918770372868, + 0.0049171485006809235, + 0.003804447827860713, + 0.0037237992510199547, + 0.01893436536192894, + 0.014508065767586231, + 0.004087952431291342, + 0.004876765888184309, + 0.0012458896962925792, + 0.0014141267165541649, + 0.001528994762338698, + 0.0013216539518907666, + 0.0015316435601562262, + 0.0011858271900564432, + 0.0018216652097180486, + 0.0009667343692854047, + 0.0016316946130245924, + 0.0032969231251627207, + 0.0019121926743537188, + 0.001616205321624875, + 0.0071625555865466595, + 0.0019021591870114207, + 0.0026206050533801317, + 0.004270317032933235, + 0.016568494960665703, + 0.018830792978405952, + 0.010627766139805317, + 0.0077479807659983635, + 0.06455783545970917, + 0.02291151136159897, + 0.01961839199066162, + 0.0637514665722847, + 0.003144963877275586, + 0.0013906739186495543, + 0.0017956231022253633, + 0.0020121971610933542, + 0.004825175274163485, + 0.013898972421884537, + 0.002553466474637389, + 0.01003944594413042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.023288773372769356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_9/index_tensor([5591, 5591])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_9/index_tensor([5591, 5591])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..744ce03f1d00c6a3df1bc0c479e88dcf552e6bac --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_9/index_tensor([5591, 5591])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot centers on a bicycle wheel, suspended and isolated against a seamless dark gray background. Under the stark, cool illumination of a single overhead keylight, the polished chrome rim gleams with crisp, defined highlights, while the matte black rubber tire reveals a detailed, unworn tread pattern. In a surreal and mesmerizing slow-motion sequence, the wheel begins to uncannily deform. The rigid metal rim begins to fluidly stretch and bend, breaking away from its perfectly round shape as if it were molten silver. Simultaneously, the rubber tire deforms, its surface becoming smooth and taut in the stretched areas. Initially rigid and straight, the thin metal spokes unnaturally elongate like shimmering elastic, stretching to several times their original length and vibrating with a silent, otherworldly tension. The play of light on the distorted surface creates a liquid, flowing effect, emphasizing the physically impossible deformation and imbuing the surreal scene with a deeply dreamlike and unsettling atmosphere.", + "seed": 14515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_9/index_5591_seed_14515_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.195345772895962, + "vlm_logprob": -0.195345772895962, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.195345772895962, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0366092212498188 + ], + "opposite_logprob_tokens": [ + -3.911609172821045 + ], + "logprob_reward": -0.0366092212498188, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0366092212498188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13488784432411194, + 0.03878193721175194, + 0.4673299491405487, + 0.10807152837514877, + 0.06433180719614029, + 0.08737856894731522, + 0.0256046075373888, + 0.031285520642995834, + 0.023327194154262543, + 0.019000958651304245 + ], + "credit_map": [ + 0.009913557209074497, + 0.02059393934905529, + 0.007692789193242788, + 0.02271328680217266, + 0.013429686427116394, + 0.020756904035806656, + 0.016794633120298386, + 0.0229930579662323, + 0.005639728624373674, + 0.004482796415686607, + 0.0024258538614958525, + 0.005295221693813801, + 0.0051933699287474155, + 0.005758911836892366, + 0.0038221285212785006, + 0.00616392632946372, + 0.053170789033174515, + 0.054229166358709335, + 0.03481427952647209, + 0.04925122112035751, + 0.05419975519180298, + 0.10368746519088745, + 0.03830423578619957, + 0.07967308908700943, + 0.007919498719274998, + 0.011790827848017216, + 0.007939403876662254, + 0.00570470979437232, + 0.020692128688097, + 0.024610791355371475, + 0.006126851309090853, + 0.023287326097488403, + 0.006957530975341797, + 0.014182254672050476, + 0.0066695669665932655, + 0.005089432466775179, + 0.009279010817408562, + 0.007502938155084848, + 0.005118608009070158, + 0.009532474912703037, + 0.010766640305519104, + 0.014335934072732925, + 0.009060126729309559, + 0.008451778441667557, + 0.012218933552503586, + 0.01319315005093813, + 0.009114019572734833, + 0.010237990878522396, + 0.0029578129760921, + 0.004689217545092106, + 0.0031547301914542913, + 0.0022949553094804287, + 0.0034540516790002584, + 0.0038007907569408417, + 0.0018823397113010287, + 0.0033707113470882177, + 0.0033177663572132587, + 0.005781090818345547, + 0.0047691455110907555, + 0.002894113538786769, + 0.004259971436113119, + 0.002969239139929414, + 0.0032357200980186462, + 0.004058480262756348, + 0.002079535974189639, + 0.0034194509498775005, + 0.0015159712638705969, + 0.0019743291195482016, + 0.00325295003131032, + 0.004640343599021435, + 0.003186769550666213, + 0.0032578445971012115, + 0.0019445248181000352, + 0.002956080250442028, + 0.002339818747714162, + 0.0016517662443220615, + 0.00234678341075778, + 0.003211241913959384, + 0.001907975529320538, + 0.0026427677366882563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0366092212498188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361636817455292 + ], + "opposite_logprob_tokens": [ + -3.6513617038726807 + ], + "logprob_reward": -0.026361636817455292, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361636817455292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28338730335235596, + 0.13361413776874542, + 0.11859144270420074, + 0.08781391382217407, + 0.11836327612400055, + 0.06879236549139023, + 0.09307961910963058, + 0.05238613858819008, + 0.022664355114102364, + 0.02130744233727455 + ], + "credit_map": [ + 0.03825332969427109, + 0.021029585972428322, + 0.022588811814785004, + 0.018725169822573662, + 0.022371752187609673, + 0.08440020680427551, + 0.015507685951888561, + 0.06051074340939522, + 0.012257593683898449, + 0.010993266478180885, + 0.007877741008996964, + 0.011045009829103947, + 0.011911970563232899, + 0.03506188094615936, + 0.01253734901547432, + 0.03192932531237602, + 0.009612205438315868, + 0.015060867182910442, + 0.008621703833341599, + 0.010565121658146381, + 0.01632525399327278, + 0.02639308199286461, + 0.008091039955615997, + 0.023922177031636238, + 0.008866672404110432, + 0.00928359292447567, + 0.006069896277040243, + 0.005045690108090639, + 0.016013765707612038, + 0.019622599706053734, + 0.005889641121029854, + 0.017022058367729187, + 0.015832530334591866, + 0.023325685411691666, + 0.014614889398217201, + 0.008300102315843105, + 0.020611071959137917, + 0.014260449446737766, + 0.006529279984533787, + 0.014889262616634369, + 0.006138923577964306, + 0.013593425042927265, + 0.0062535773031413555, + 0.008106343448162079, + 0.010479722172021866, + 0.007843906059861183, + 0.006556518375873566, + 0.009819946251809597, + 0.009265761822462082, + 0.01744811423122883, + 0.009337825700640678, + 0.008091598749160767, + 0.012797708623111248, + 0.015693824738264084, + 0.0066551826894283295, + 0.013789603486657143, + 0.005205404479056597, + 0.007809123490005732, + 0.0037294882349669933, + 0.0051948921754956245, + 0.009004116989672184, + 0.007848287932574749, + 0.007442779839038849, + 0.006152044050395489, + 0.001991459634155035, + 0.0031495033763349056, + 0.0022398384753614664, + 0.0028788247145712376, + 0.003063683630898595, + 0.003905793186277151, + 0.0025527384132146835, + 0.0028825143817812204, + 0.0029181933496147394, + 0.0038701023440808058, + 0.0018554815324023366, + 0.0024936851114034653, + 0.0030365081038326025, + 0.002563966903835535, + 0.002299017971381545, + 0.002270486205816269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026361636817455292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983432054519653 + ], + "opposite_logprob_tokens": [ + -0.5233432054519653 + ], + "logprob_reward": -0.8983432054519653, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983432054519653, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0661749392747879, + 0.024234876036643982, + 0.29873788356781006, + 0.32585763931274414, + 0.060394782572984695, + 0.08684008568525314, + 0.06814023107290268, + 0.02305702492594719, + 0.028995107859373093, + 0.017567504197359085 + ], + "credit_map": [ + 0.004054500255733728, + 0.010051541030406952, + 0.0030673802830278873, + 0.010562997311353683, + 0.0066940151154994965, + 0.011248758994042873, + 0.005367102101445198, + 0.015128644183278084, + 0.001751043600961566, + 0.004013837315142155, + 0.0017225872725248337, + 0.005002949386835098, + 0.0037790711503475904, + 0.002408282831311226, + 0.0026110108010470867, + 0.0029460936784744263, + 0.035963062196969986, + 0.021929481998085976, + 0.022178977727890015, + 0.016879120841622353, + 0.028884999454021454, + 0.08753033727407455, + 0.020325521007180214, + 0.06504637002944946, + 0.050848253071308136, + 0.021170709282159805, + 0.03420630469918251, + 0.010094680823385715, + 0.046883512288331985, + 0.07904985547065735, + 0.013767109252512455, + 0.06983721256256104, + 0.005194468889385462, + 0.009812639094889164, + 0.004966819658875465, + 0.005882467143237591, + 0.01079905591905117, + 0.009486859664320946, + 0.0056763216853141785, + 0.008576150052249432, + 0.013694020919501781, + 0.012486668303608894, + 0.010627500712871552, + 0.010028475895524025, + 0.011548277921974659, + 0.008453138172626495, + 0.011005480773746967, + 0.008996522054076195, + 0.006080055143684149, + 0.019504228606820107, + 0.004973892122507095, + 0.005745524540543556, + 0.008177947252988815, + 0.008480159565806389, + 0.006894957739859819, + 0.008283465169370174, + 0.0019100381759926677, + 0.004246123600751162, + 0.00169267226010561, + 0.002292447257786989, + 0.0028250066097825766, + 0.0038962699472904205, + 0.0033770527224987745, + 0.002817415865138173, + 0.00306572369299829, + 0.004182582255452871, + 0.002803121693432331, + 0.0037009490188211203, + 0.00375921418890357, + 0.003619004972279072, + 0.004586749244481325, + 0.0032777637243270874, + 0.0020107056479901075, + 0.003260623198002577, + 0.0015964884078130126, + 0.00143075839150697, + 0.0024552077520638704, + 0.0028660392854362726, + 0.0015695532783865929, + 0.002378128468990326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.8983432054519653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006754305679351091 + ], + "opposite_logprob_tokens": [ + -5.006754398345947 + ], + "logprob_reward": -0.006754305679351091, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006754305679351091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06025998666882515, + 0.027930833399295807, + 0.22534555196762085, + 0.1968153417110443, + 0.1992877721786499, + 0.1421082317829132, + 0.04994581639766693, + 0.052086975425481796, + 0.023479348048567772, + 0.022740095853805542 + ], + "credit_map": [ + 0.0053246887400746346, + 0.007094571366906166, + 0.0034725936129689217, + 0.006007807329297066, + 0.007958038710057735, + 0.01480479259043932, + 0.004636563826352358, + 0.010960938408970833, + 0.0021346688736230135, + 0.0042639258317649364, + 0.0015052750241011381, + 0.0036211491096764803, + 0.005270930472761393, + 0.004444981925189495, + 0.0022478089667856693, + 0.004442096222192049, + 0.02714686281979084, + 0.017238810658454895, + 0.015414729714393616, + 0.017330268397927284, + 0.028033582493662834, + 0.059011731296777725, + 0.01446456741541624, + 0.04670504480600357, + 0.033279623836278915, + 0.015223533846437931, + 0.020553123205900192, + 0.006570302881300449, + 0.027540622279047966, + 0.04500140994787216, + 0.009046417661011219, + 0.03960033506155014, + 0.023286620154976845, + 0.04553186893463135, + 0.02288895845413208, + 0.017150476574897766, + 0.03205183893442154, + 0.020228590816259384, + 0.012314882129430771, + 0.025834545493125916, + 0.016093725338578224, + 0.022187314927577972, + 0.014402268454432487, + 0.01312530878931284, + 0.0224006250500679, + 0.024277379736304283, + 0.01196814700961113, + 0.017653485760092735, + 0.004641054663807154, + 0.012042065151035786, + 0.0037466958165168762, + 0.004620835650712252, + 0.008677143603563309, + 0.005826294422149658, + 0.004484841134399176, + 0.005906888283789158, + 0.0049463496543467045, + 0.007980002090334892, + 0.005162052810192108, + 0.004508418496698141, + 0.007811214774847031, + 0.008040014654397964, + 0.006785853300243616, + 0.006853074301034212, + 0.0024718025233596563, + 0.0039373040199279785, + 0.0015832860954105854, + 0.0027676608879119158, + 0.0030240549240261316, + 0.00292767072096467, + 0.003931642975658178, + 0.0028359282296150923, + 0.002126063220202923, + 0.003644240088760853, + 0.0018040945287793875, + 0.0024929221253842115, + 0.004057590384036303, + 0.0034346741158515215, + 0.002346603898331523, + 0.0028339114505797625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.006754305679351091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008660495281219482 + ], + "opposite_logprob_tokens": [ + -4.758660316467285 + ], + "logprob_reward": -0.008660495281219482, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008660495281219482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18691515922546387, + 0.03545178100466728, + 0.20630498230457306, + 0.1152268573641777, + 0.14658518135547638, + 0.1269177496433258, + 0.06740369647741318, + 0.038816872984170914, + 0.04023143649101257, + 0.03614625334739685 + ], + "credit_map": [ + 0.016335757449269295, + 0.02166004106402397, + 0.009990948252379894, + 0.028920620679855347, + 0.014202759601175785, + 0.03749784454703331, + 0.013433045707643032, + 0.044874157756567, + 0.0042264750227332115, + 0.005105155054479837, + 0.0020792772993445396, + 0.0060160402208566666, + 0.005330208223313093, + 0.0046389964409172535, + 0.0029846385587006807, + 0.005070994142442942, + 0.018384043127298355, + 0.03011641465127468, + 0.013362252153456211, + 0.0348389632999897, + 0.030232129618525505, + 0.026233501732349396, + 0.022074425593018532, + 0.031063279137015343, + 0.016375860199332237, + 0.015754567459225655, + 0.01921083591878414, + 0.009151618927717209, + 0.018679052591323853, + 0.012573052197694778, + 0.009869908913969994, + 0.013611975125968456, + 0.012152167037129402, + 0.026534883305430412, + 0.013080036267638206, + 0.011283975094556808, + 0.028711935505270958, + 0.01917024329304695, + 0.014864588156342506, + 0.02078736573457718, + 0.02048678509891033, + 0.01585848443210125, + 0.016169922426342964, + 0.013560688123106956, + 0.01918359473347664, + 0.015582060441374779, + 0.011603455059230328, + 0.014472764916718006, + 0.004647745285183191, + 0.019223719835281372, + 0.0041933441534638405, + 0.007207932882010937, + 0.010649526491761208, + 0.007315147202461958, + 0.0062247891910374165, + 0.007941492833197117, + 0.0040068295784294605, + 0.007073726039379835, + 0.0043542892672121525, + 0.0032278166618198156, + 0.004051923751831055, + 0.007385144010186195, + 0.0032507642172276974, + 0.00546638248488307, + 0.003653605468571186, + 0.006563431583344936, + 0.0024762945249676704, + 0.0038821715861558914, + 0.006612627301365137, + 0.006471908651292324, + 0.005548889748752117, + 0.005022509954869747, + 0.0034652433823794127, + 0.005662874784320593, + 0.0027863988652825356, + 0.002782306866720319, + 0.004648054949939251, + 0.008484154008328915, + 0.002938974415883422, + 0.005378252360969782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.008660495281219482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot centers on a bicycle wheel, suspended and isolated against a seamless dark gray background. Under the stark, cool illumination of a single overhead keylight, the polished chrome rim gleams with crisp, defined highlights, while the matte black rubber tire reveals a detailed, unworn tread pattern. In a surreal and mesmerizing slow-motion sequence, the wheel begins to uncannily deform. The rigid metal rim begins to fluidly stretch and bend, breaking away from its perfectly round shape as if it were molten silver. Simultaneously, the rubber tire deforms, its surface becoming smooth and taut in the stretched areas. Initially rigid and straight, the thin metal spokes unnaturally elongate like shimmering elastic, stretching to several times their original length and vibrating with a silent, otherworldly tension. The play of light on the distorted surface creates a liquid, flowing effect, emphasizing the physically impossible deformation and imbuing the surreal scene with a deeply dreamlike and unsettling atmosphere.", + "seed": 14516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_9/index_5591_seed_14516_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2861393418163061, + "vlm_logprob": -0.2861393418163061, + "vlm_yes_no_margin": 2.500000023841858, + "vlm_token_credit": -0.2861393418163061, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01719694770872593 + ], + "opposite_logprob_tokens": [ + -5.392197132110596 + ], + "logprob_reward": -0.01719694770872593, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01719694770872593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17159529030323029, + 0.05148245021700859, + 0.09430091083049774, + 0.04510956257581711, + 0.0715852826833725, + 0.15952055156230927, + 0.14113271236419678, + 0.07285322993993759, + 0.09133048355579376, + 0.1010894924402237 + ], + "credit_map": [ + 0.014776984229683876, + 0.048173870891332626, + 0.01674528792500496, + 0.01102787908166647, + 0.03085174784064293, + 0.009812826290726662, + 0.01283326093107462, + 0.027373427525162697, + 0.004096912685781717, + 0.009769496507942677, + 0.007888874039053917, + 0.004137228708714247, + 0.008552479557693005, + 0.003209827234968543, + 0.007035254966467619, + 0.00679237674921751, + 0.00816577672958374, + 0.01612357795238495, + 0.016450293362140656, + 0.009524029679596424, + 0.01415836252272129, + 0.004801560193300247, + 0.01149208378046751, + 0.01358522567898035, + 0.0028183837421238422, + 0.006258971057832241, + 0.006663241423666477, + 0.004934491589665413, + 0.008395115844905376, + 0.004406999796628952, + 0.004005025140941143, + 0.007627331651747227, + 0.006973362527787685, + 0.008613641373813152, + 0.011262234300374985, + 0.011891216039657593, + 0.009615647606551647, + 0.004179751966148615, + 0.009064323268830776, + 0.009985108859837055, + 0.010009575635194778, + 0.020616957917809486, + 0.01972377859055996, + 0.02310515008866787, + 0.02611302211880684, + 0.017688676714897156, + 0.023463111370801926, + 0.0188002809882164, + 0.011269324459135532, + 0.016363611444830894, + 0.03214658051729202, + 0.011227996088564396, + 0.021780652925372124, + 0.006144453771412373, + 0.024248750880360603, + 0.01795133762061596, + 0.0062921722419559956, + 0.014834162779152393, + 0.011352427303791046, + 0.007652666885405779, + 0.00950936134904623, + 0.0038227257318794727, + 0.007356706541031599, + 0.01203300803899765, + 0.007737939711660147, + 0.015117892064154148, + 0.013368654064834118, + 0.007260695565491915, + 0.015238129533827305, + 0.009054126217961311, + 0.011603857390582561, + 0.011949190869927406, + 0.008726414293050766, + 0.01258796639740467, + 0.010131285525858402, + 0.0249356497079134, + 0.011091954074800014, + 0.00639125844463706, + 0.015122300013899803, + 0.012102668173611164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.01719694770872593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295502632856369 + ], + "opposite_logprob_tokens": [ + -2.237955093383789 + ], + "logprob_reward": -0.11295502632856369, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295502632856369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20445415377616882, + 0.03352442383766174, + 0.14064861834049225, + 0.10310758650302887, + 0.08445541560649872, + 0.1044001430273056, + 0.12876677513122559, + 0.030121110379695892, + 0.07140231877565384, + 0.09911941736936569 + ], + "credit_map": [ + 0.01980680227279663, + 0.03088018111884594, + 0.029369322583079338, + 0.024271832779049873, + 0.029201650992035866, + 0.01632271334528923, + 0.029992086812853813, + 0.024609580636024475, + 0.0037462408654391766, + 0.005442382767796516, + 0.003795065451413393, + 0.004230196587741375, + 0.005536175798624754, + 0.002457102295011282, + 0.004063577391207218, + 0.004253679886460304, + 0.013365660794079304, + 0.0141997579485178, + 0.03686584159731865, + 0.011151419021189213, + 0.014033133164048195, + 0.0068551828153431416, + 0.020857691764831543, + 0.02331993356347084, + 0.007795160636305809, + 0.014864454977214336, + 0.018044363707304, + 0.00820087268948555, + 0.01861857995390892, + 0.006363143678754568, + 0.011860568076372147, + 0.017360445111989975, + 0.005725156050175428, + 0.016503337770700455, + 0.00968586839735508, + 0.007745693903416395, + 0.015323913656175137, + 0.007104064337909222, + 0.009845735505223274, + 0.012521658092737198, + 0.006888759788125753, + 0.01086923386901617, + 0.028346264734864235, + 0.009246614761650562, + 0.014335189945995808, + 0.01033856999129057, + 0.009972959756851196, + 0.014402554370462894, + 0.01525060087442398, + 0.011897998861968517, + 0.02527564764022827, + 0.016929717734456062, + 0.016386287286877632, + 0.006279276683926582, + 0.019429322332143784, + 0.01731792651116848, + 0.0029240092262625694, + 0.004020057152956724, + 0.004526566714048386, + 0.003246213309466839, + 0.003730260767042637, + 0.0036593282129615545, + 0.004067135043442249, + 0.003947542980313301, + 0.006078744772821665, + 0.01231442578136921, + 0.010540641844272614, + 0.0069371373392641544, + 0.010406488552689552, + 0.00620288634672761, + 0.008258648216724396, + 0.0106633510440588, + 0.008715550415217876, + 0.010489587672054768, + 0.011988489888608456, + 0.029957404360175133, + 0.01034184917807579, + 0.004700643941760063, + 0.012325885705649853, + 0.010600006207823753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.11295502632856369 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371139287948608 + ], + "opposite_logprob_tokens": [ + -0.3871139585971832 + ], + "logprob_reward": -1.1371139287948608, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371139287948608, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09465959668159485, + 0.06260868161916733, + 0.09123394638299942, + 0.05320626497268677, + 0.06881533563137054, + 0.23607254028320312, + 0.1866176873445511, + 0.08352046459913254, + 0.04967931658029556, + 0.07358615845441818 + ], + "credit_map": [ + 0.008068879134953022, + 0.015620995312929153, + 0.015449450351297855, + 0.011786375194787979, + 0.014549707062542439, + 0.005841855425387621, + 0.00811691116541624, + 0.015225423499941826, + 0.006076320540159941, + 0.014626992866396904, + 0.007114709820598364, + 0.0063264756463468075, + 0.009250637143850327, + 0.0038219064008444548, + 0.006080362014472485, + 0.009311276488006115, + 0.007856868207454681, + 0.016608046367764473, + 0.01441612746566534, + 0.011800957843661308, + 0.010665876790881157, + 0.004760884679853916, + 0.011248212307691574, + 0.013876968063414097, + 0.003531423397362232, + 0.00797056034207344, + 0.010115214623510838, + 0.0035145722795277834, + 0.009674816392362118, + 0.003364972537383437, + 0.007258458063006401, + 0.007776242680847645, + 0.006111022550612688, + 0.010051911696791649, + 0.007537713274359703, + 0.010350697673857212, + 0.014111028052866459, + 0.004013399593532085, + 0.007311342284083366, + 0.009328221902251244, + 0.015270652249455452, + 0.02812262624502182, + 0.032968975603580475, + 0.031242143362760544, + 0.04159941151738167, + 0.025335140526294708, + 0.0311282929033041, + 0.030405286699533463, + 0.023927774280309677, + 0.020969778299331665, + 0.04027380794286728, + 0.0151806166395545, + 0.03081868775188923, + 0.00970594584941864, + 0.01932976394891739, + 0.026411304250359535, + 0.009594074450433254, + 0.01406505424529314, + 0.010976490564644337, + 0.010159743949770927, + 0.009425229392945766, + 0.007304208818823099, + 0.01005014218389988, + 0.011945522390305996, + 0.003954306710511446, + 0.007983448915183544, + 0.005454976577311754, + 0.011159236542880535, + 0.0056386380456388, + 0.003680469235405326, + 0.006221776828169823, + 0.0055864593014121056, + 0.007179916370660067, + 0.010880791582167149, + 0.008840585127472878, + 0.011194503866136074, + 0.01223231852054596, + 0.004380915313959122, + 0.008988376706838608, + 0.00988875050097704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -1.1371139287948608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282526075839996 + ], + "opposite_logprob_tokens": [ + -2.0178253650665283 + ], + "logprob_reward": -0.14282526075839996, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14282526075839996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11915315687656403, + 0.062184009701013565, + 0.061610907316207886, + 0.03200951963663101, + 0.03747205436229706, + 0.23953144252300262, + 0.031056644394993782, + 0.24759890139102936, + 0.039553701877593994, + 0.12982963025569916 + ], + "credit_map": [ + 0.009587526321411133, + 0.0335744172334671, + 0.014873730018734932, + 0.008404127322137356, + 0.019696129485964775, + 0.005871950648725033, + 0.00912667065858841, + 0.018018608912825584, + 0.006629831623286009, + 0.013128391467034817, + 0.0074842250905931, + 0.006564720533788204, + 0.007828151807188988, + 0.0039160787127912045, + 0.00752123212441802, + 0.009111382067203522, + 0.0041963024996221066, + 0.01734195649623871, + 0.00596648408100009, + 0.00503581203520298, + 0.011939204297959805, + 0.003090131562203169, + 0.004568732809275389, + 0.00947228167206049, + 0.0023893029429018497, + 0.005283867474645376, + 0.004268341232091188, + 0.003196533303707838, + 0.006186679471284151, + 0.0027261998038738966, + 0.002774142427369952, + 0.005184453912079334, + 0.0038603420834988356, + 0.005785180721431971, + 0.006712459493428469, + 0.0038273874670267105, + 0.00505355978384614, + 0.0029993811622262, + 0.0038013120647519827, + 0.005432430189102888, + 0.013738201931118965, + 0.02223743125796318, + 0.028131945058703423, + 0.04251174256205559, + 0.030140144750475883, + 0.024100620299577713, + 0.05148229002952576, + 0.027189068496227264, + 0.0027682208456099033, + 0.004941197577863932, + 0.0034809818025678396, + 0.003521982114762068, + 0.0041192760691046715, + 0.002188537735491991, + 0.005150421056896448, + 0.004886027425527573, + 0.020242825150489807, + 0.05167878791689873, + 0.02713216468691826, + 0.02211127057671547, + 0.04491271823644638, + 0.018664324656128883, + 0.025854723528027534, + 0.0370020866394043, + 0.004257973283529282, + 0.005997445899993181, + 0.006117581389844418, + 0.005110628437250853, + 0.0044353799894452095, + 0.003736424958333373, + 0.003913049586117268, + 0.005985215771943331, + 0.012279722839593887, + 0.019619867205619812, + 0.011486872099339962, + 0.025822654366493225, + 0.016318704932928085, + 0.006766714621335268, + 0.019641563296318054, + 0.017893534153699875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.14282526075839996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060554549098015 + ], + "opposite_logprob_tokens": [ + -3.8956055641174316 + ], + "logprob_reward": -0.02060554549098015, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060554549098015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09291712939739227, + 0.08542317897081375, + 0.048224274069070816, + 0.04230136424303055, + 0.04019806906580925, + 0.17858567833900452, + 0.09831072390079498, + 0.19877730309963226, + 0.09759311378002167, + 0.11766917258501053 + ], + "credit_map": [ + 0.007828766480088234, + 0.02121596783399582, + 0.011250194162130356, + 0.007103498559445143, + 0.014376035891473293, + 0.005459586624056101, + 0.01157358754426241, + 0.014109487645328045, + 0.008065319620072842, + 0.018732206895947456, + 0.008542928844690323, + 0.006706468295305967, + 0.012284571304917336, + 0.005441424902528524, + 0.014181338250637054, + 0.011468921788036823, + 0.004018035717308521, + 0.008016519248485565, + 0.009486171416938305, + 0.004046963062137365, + 0.007055440451949835, + 0.0029536320362240076, + 0.0057555087842047215, + 0.006892003584653139, + 0.004246789030730724, + 0.006234986241906881, + 0.00886707566678524, + 0.004540173336863518, + 0.004898298531770706, + 0.003127268748357892, + 0.004020375665277243, + 0.006366397719830275, + 0.002818768611177802, + 0.008753473870456219, + 0.0049598501063883305, + 0.00299312942661345, + 0.008347047492861748, + 0.003410929813981056, + 0.0033887445461004972, + 0.005526129622012377, + 0.011009901762008667, + 0.022036058828234673, + 0.029565921053290367, + 0.02319645695388317, + 0.02845497615635395, + 0.018998030573129654, + 0.02427094802260399, + 0.021053388714790344, + 0.012921130284667015, + 0.008525454439222813, + 0.02043667994439602, + 0.01062021218240261, + 0.015157566405832767, + 0.004869085270911455, + 0.01255801785737276, + 0.013222577050328255, + 0.020164266228675842, + 0.04202607646584511, + 0.025303712114691734, + 0.01699541136622429, + 0.032051973044872284, + 0.013941673561930656, + 0.019880060106515884, + 0.02841414511203766, + 0.01086545828729868, + 0.014205477200448513, + 0.016239134594798088, + 0.007802289444953203, + 0.013250461779534817, + 0.008869420737028122, + 0.009872907772660255, + 0.016487967222929, + 0.011715061031281948, + 0.01722271926701069, + 0.011064721271395683, + 0.028527241200208664, + 0.013319398276507854, + 0.007782442960888147, + 0.013618546538054943, + 0.014419042505323887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.02060554549098015 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot centers on a bicycle wheel, suspended and isolated against a seamless dark gray background. Under the stark, cool illumination of a single overhead keylight, the polished chrome rim gleams with crisp, defined highlights, while the matte black rubber tire reveals a detailed, unworn tread pattern. In a surreal and mesmerizing slow-motion sequence, the wheel begins to uncannily deform. The rigid metal rim begins to fluidly stretch and bend, breaking away from its perfectly round shape as if it were molten silver. Simultaneously, the rubber tire deforms, its surface becoming smooth and taut in the stretched areas. Initially rigid and straight, the thin metal spokes unnaturally elongate like shimmering elastic, stretching to several times their original length and vibrating with a silent, otherworldly tension. The play of light on the distorted surface creates a liquid, flowing effect, emphasizing the physically impossible deformation and imbuing the surreal scene with a deeply dreamlike and unsettling atmosphere.", + "seed": 14517, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_9/index_5591_seed_14517_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.22567790951579808, + "vlm_logprob": -0.22567790951579808, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.22567790951579808, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028376052156090736 + ], + "opposite_logprob_tokens": [ + -4.27837610244751 + ], + "logprob_reward": -0.028376052156090736, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.028376052156090736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1947554051876068, + 0.19641564786434174, + 0.3490080237388611, + 0.10218110680580139, + 0.019150955602526665, + 0.02162299118936062, + 0.019570961594581604, + 0.01683497428894043, + 0.06621930748224258, + 0.014240654185414314 + ], + "credit_map": [ + 0.011593438684940338, + 0.0062073394656181335, + 0.0450575016438961, + 0.0061165993101894855, + 0.00806864257901907, + 0.07443059235811234, + 0.007328496314585209, + 0.035952791571617126, + 0.01748598739504814, + 0.007637581322342157, + 0.0554451122879982, + 0.006155870389193296, + 0.007377496920526028, + 0.028182653710246086, + 0.0061095524579286575, + 0.06802140176296234, + 0.0317615307867527, + 0.0068410406820476055, + 0.09380590170621872, + 0.008279798552393913, + 0.009992454200983047, + 0.06150052323937416, + 0.00811760313808918, + 0.12870918214321136, + 0.006820250768214464, + 0.00253746728412807, + 0.0351177453994751, + 0.0028772172518074512, + 0.004087056033313274, + 0.038492601364851, + 0.002495403401553631, + 0.009753364138305187, + 0.002453426131978631, + 0.0026314440183341503, + 0.002586553106084466, + 0.002222337294369936, + 0.0029115963261574507, + 0.0020658785942941904, + 0.002530452562496066, + 0.0017492679180577397, + 0.0025267137680202723, + 0.0025574390310794115, + 0.0036657904274761677, + 0.0023885448463261127, + 0.003861903678625822, + 0.0015495114494115114, + 0.0027174523565918207, + 0.002355635864660144, + 0.0024576899595558643, + 0.0020016361959278584, + 0.005177535116672516, + 0.001683474867604673, + 0.0017992316279560328, + 0.0013468468096107244, + 0.0013430622639134526, + 0.0037614840548485518, + 0.001675378647632897, + 0.002618233673274517, + 0.002382643288001418, + 0.0019658382516354322, + 0.002453005872666836, + 0.0013976520858705044, + 0.002500734059140086, + 0.0018414887599647045, + 0.007748978212475777, + 0.0034466437064111233, + 0.02251351624727249, + 0.0028431955724954605, + 0.0027748362626880407, + 0.0038446574471890926, + 0.0024829001631587744, + 0.020564576610922813, + 0.0013340685982257128, + 0.0010490124113857746, + 0.00418433640152216, + 0.0007798398728482425, + 0.00069087790325284, + 0.0010242948774248362, + 0.0009168670512735844, + 0.00426135677844286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.028376052156090736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307539016008377 + ], + "opposite_logprob_tokens": [ + -3.1680753231048584 + ], + "logprob_reward": -0.04307539016008377, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307539016008377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46988338232040405, + 0.02962055429816246, + 0.21421445906162262, + 0.08129696547985077, + 0.06867282837629318, + 0.04584231227636337, + 0.023826396092772484, + 0.016864974051713943, + 0.006775440648198128, + 0.043002668768167496 + ], + "credit_map": [ + 0.026483731344342232, + 0.01252754032611847, + 0.09279463440179825, + 0.011175441555678844, + 0.018172023817896843, + 0.18841835856437683, + 0.015350141562521458, + 0.10496154427528381, + 0.0031689596362411976, + 0.0020543953869491816, + 0.007638758048415184, + 0.0011922895209863782, + 0.0016689751064404845, + 0.004364935215562582, + 0.0014241712633520365, + 0.00810807291418314, + 0.017012150958180428, + 0.005424131639301777, + 0.043663427233695984, + 0.004629327449947596, + 0.007771048694849014, + 0.05028320848941803, + 0.004838084802031517, + 0.08059309422969818, + 0.0034418508876115084, + 0.002570847515016794, + 0.018657630309462547, + 0.0019734366796910763, + 0.0033867417369037867, + 0.03261641785502434, + 0.0023939639795571566, + 0.01625608094036579, + 0.00862063653767109, + 0.0033292348962277174, + 0.023016465827822685, + 0.003558262251317501, + 0.003354814602062106, + 0.004860787186771631, + 0.0036653412971645594, + 0.01826728694140911, + 0.006888180505484343, + 0.004707284737378359, + 0.007463931106030941, + 0.0040657371282577515, + 0.00596549641340971, + 0.004561937879770994, + 0.007928663864731789, + 0.004261085297912359, + 0.0025143278762698174, + 0.0018293943721801043, + 0.0068189967423677444, + 0.001394439721480012, + 0.002070074900984764, + 0.0015119322342798114, + 0.001242928788997233, + 0.006444302387535572, + 0.0015919270226731896, + 0.0011042873375117779, + 0.00512361666187644, + 0.000965969345998019, + 0.0008926170994527638, + 0.0011422944953665137, + 0.0008878424996510148, + 0.005156420636922121, + 0.0008908231393434107, + 0.0007444507209584117, + 0.0009973827982321382, + 0.0006779272807762027, + 0.0007457197061739862, + 0.0006850673235021532, + 0.0008748945547267795, + 0.0011591754155233502, + 0.004559613764286041, + 0.0012210672721266747, + 0.015454352833330631, + 0.0019497303292155266, + 0.0012411788338795304, + 0.002692848676815629, + 0.0015897256089374423, + 0.014294157736003399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.04307539016008377 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743255972862244 + ], + "opposite_logprob_tokens": [ + -0.47432559728622437 + ], + "logprob_reward": -0.9743255972862244, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743255972862244, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3030289113521576, + 0.06250172853469849, + 0.27685511112213135, + 0.19196723401546478, + 0.04051446169614792, + 0.050915345549583435, + 0.0216783806681633, + 0.01768491417169571, + 0.030177997425198555, + 0.004675938282161951 + ], + "credit_map": [ + 0.0226471945643425, + 0.006062103435397148, + 0.09085021167993546, + 0.008105580694973469, + 0.009677205234766006, + 0.11353997886180878, + 0.00959914829581976, + 0.04254748672246933, + 0.0052674622274935246, + 0.004586157854646444, + 0.005010749213397503, + 0.002495216904208064, + 0.004192629363387823, + 0.02100808173418045, + 0.003451873082667589, + 0.016489559784531593, + 0.027236053720116615, + 0.0056429957039654255, + 0.11738067120313644, + 0.008766709826886654, + 0.0070319147780537605, + 0.08016914874315262, + 0.006626382004469633, + 0.024001220241189003, + 0.010044727474451065, + 0.006050779018551111, + 0.009339463897049427, + 0.003305766498669982, + 0.006989988032728434, + 0.07635144889354706, + 0.004548773169517517, + 0.07533629238605499, + 0.004165297839790583, + 0.00411275215446949, + 0.00898702535778284, + 0.003466653870418668, + 0.00563420495018363, + 0.0033692687284201384, + 0.0032317114528268576, + 0.007547548972070217, + 0.004707595333456993, + 0.007746649440377951, + 0.0060392688028514385, + 0.0062138596549630165, + 0.009594898670911789, + 0.0040896497666835785, + 0.00779589731246233, + 0.0047275288961827755, + 0.0028342369478195906, + 0.00217238487675786, + 0.005055185407400131, + 0.001981587614864111, + 0.002646177774295211, + 0.001528303138911724, + 0.0014396699843928218, + 0.00402083620429039, + 0.0023412953596562147, + 0.000495891144964844, + 0.006644146982580423, + 0.0006724097765982151, + 0.0005167670897208154, + 0.0009651820291765034, + 0.0005200734012760222, + 0.0055291480384767056, + 0.003834074828773737, + 0.0013766475021839142, + 0.009809361770749092, + 0.001801951788365841, + 0.0011504888534545898, + 0.001932067214511335, + 0.0012476103147491813, + 0.009025796316564083, + 0.0005411700694821775, + 0.0003770580515265465, + 0.0012780273100361228, + 0.00041384249925613403, + 0.0003417897969484329, + 0.00036262793582864106, + 0.0002767896221484989, + 0.001084633287973702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.9743255972862244 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005149126052856 + ], + "opposite_logprob_tokens": [ + -2.695051431655884 + ], + "logprob_reward": -0.07005149126052856, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005149126052856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22854959964752197, + 0.07321378588676453, + 0.14994339644908905, + 0.46340271830558777, + 0.02150735817849636, + 0.029134228825569153, + 0.010790949687361717, + 0.004970767069607973, + 0.01200803555548191, + 0.006479183211922646 + ], + "credit_map": [ + 0.012663880363106728, + 0.006288712844252586, + 0.04699648544192314, + 0.005613015033304691, + 0.00906190276145935, + 0.09151937067508698, + 0.00704036932438612, + 0.04936586692929268, + 0.006706530228257179, + 0.006357660051435232, + 0.010706293396651745, + 0.0035819089971482754, + 0.006213459651917219, + 0.016159076243638992, + 0.00446698721498251, + 0.01902187056839466, + 0.01612219773232937, + 0.006053574848920107, + 0.058889240026474, + 0.005654636304825544, + 0.00767243979498744, + 0.01698414981365204, + 0.0048085604794323444, + 0.03375859558582306, + 0.01557716354727745, + 0.008851082995533943, + 0.06133681535720825, + 0.006708524189889431, + 0.010108279064297676, + 0.20379705727100372, + 0.008354670368134975, + 0.14866912364959717, + 0.0022448136005550623, + 0.002923958469182253, + 0.003808199893683195, + 0.001918131485581398, + 0.003247699700295925, + 0.00198146584443748, + 0.0023380641359835863, + 0.003045027144253254, + 0.0028738267719745636, + 0.0022112317383289337, + 0.008899797685444355, + 0.0018051649676635861, + 0.0028210405725985765, + 0.0017289285315200686, + 0.0014472720213234425, + 0.007346967700868845, + 0.0011010731104761362, + 0.0008615205297246575, + 0.002776901237666607, + 0.0008700865437276661, + 0.0009015480754896998, + 0.000843298330437392, + 0.0006990222609601915, + 0.002737500239163637, + 0.0007053497829474509, + 0.0008518386166542768, + 0.000619919563177973, + 0.0005115671083331108, + 0.0007145926938392222, + 0.00043412766535766423, + 0.0005502837593667209, + 0.000583087676204741, + 0.0014164639869704843, + 0.002156745409592986, + 0.0018906454788520932, + 0.0011197568383067846, + 0.0021397885866463184, + 0.0009378946269862354, + 0.0010328491916880012, + 0.0013138916110619903, + 0.0007689860067330301, + 0.0008005847339518368, + 0.0015422331634908915, + 0.0006153223221190274, + 0.0006438382552005351, + 0.0004604536516126245, + 0.0005401074304245412, + 0.0011076575610786676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.07005149126052856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561016716063023 + ], + "opposite_logprob_tokens": [ + -4.387560844421387 + ], + "logprob_reward": -0.012561016716063023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561016716063023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06013449653983116, + 0.07067607343196869, + 0.4528280198574066, + 0.23478826880455017, + 0.046511031687259674, + 0.04356449469923973, + 0.034816719591617584, + 0.034141723066568375, + 0.011472932063043118, + 0.011066293343901634 + ], + "credit_map": [ + 0.004301862791180611, + 0.006154180504381657, + 0.006974845193326473, + 0.004890555050224066, + 0.006162414327263832, + 0.010077383369207382, + 0.005257720127701759, + 0.016315538436174393, + 0.006334294565021992, + 0.006113077979534864, + 0.014940746128559113, + 0.004833146929740906, + 0.006532271858304739, + 0.008313014172017574, + 0.005932067986577749, + 0.01767745241522789, + 0.04850640147924423, + 0.007467248011380434, + 0.16062049567699432, + 0.012494146823883057, + 0.008030351251363754, + 0.046109091490507126, + 0.01100781187415123, + 0.15859249234199524, + 0.008216837421059608, + 0.006248273886740208, + 0.04273693636059761, + 0.0043080770410597324, + 0.006876428611576557, + 0.09900762885808945, + 0.004692887421697378, + 0.06270118057727814, + 0.004834468476474285, + 0.0068251281045377254, + 0.00623455923050642, + 0.004615667276084423, + 0.00851958803832531, + 0.004502436146140099, + 0.00553286075592041, + 0.005446325987577438, + 0.004901466891169548, + 0.00637302128598094, + 0.005963328294456005, + 0.005358334630727768, + 0.00675695575773716, + 0.004292708355933428, + 0.0051015280187129974, + 0.004817152861505747, + 0.003647811245173216, + 0.004349338356405497, + 0.007445329334586859, + 0.0028491741977632046, + 0.005179786588996649, + 0.002405829494819045, + 0.0028629659209400415, + 0.00607648491859436, + 0.004119118209928274, + 0.0017997075337916613, + 0.012015962973237038, + 0.001629372127354145, + 0.0013625308638438582, + 0.0018169990507885814, + 0.0014387413393706083, + 0.00995929166674614, + 0.0012866966426372528, + 0.0014724506763741374, + 0.0017547659808769822, + 0.0011539121624082327, + 0.0014070760225877166, + 0.0010276499669998884, + 0.0014467282453551888, + 0.001923652016557753, + 0.001456870581023395, + 0.002025357447564602, + 0.0014464081032201648, + 0.0014479634119197726, + 0.0012199260527268052, + 0.001137304469011724, + 0.0011898253578692675, + 0.0011426379205659032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.012561016716063023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot centers on a bicycle wheel, suspended and isolated against a seamless dark gray background. Under the stark, cool illumination of a single overhead keylight, the polished chrome rim gleams with crisp, defined highlights, while the matte black rubber tire reveals a detailed, unworn tread pattern. In a surreal and mesmerizing slow-motion sequence, the wheel begins to uncannily deform. The rigid metal rim begins to fluidly stretch and bend, breaking away from its perfectly round shape as if it were molten silver. Simultaneously, the rubber tire deforms, its surface becoming smooth and taut in the stretched areas. Initially rigid and straight, the thin metal spokes unnaturally elongate like shimmering elastic, stretching to several times their original length and vibrating with a silent, otherworldly tension. The play of light on the distorted surface creates a liquid, flowing effect, emphasizing the physically impossible deformation and imbuing the surreal scene with a deeply dreamlike and unsettling atmosphere.", + "seed": 14518, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000016/group_9/index_5591_seed_14518_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.29318551914766433, + "vlm_logprob": -0.29318551914766433, + "vlm_yes_no_margin": 2.8999999523162843, + "vlm_token_credit": -0.29318551914766433, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03752141818404198 + ], + "opposite_logprob_tokens": [ + -4.2875213623046875 + ], + "logprob_reward": -0.03752141818404198, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03752141818404198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12822741270065308, + 0.13235121965408325, + 0.047544676810503006, + 0.08085321635007858, + 0.04541748762130737, + 0.12211688607931137, + 0.1486286073923111, + 0.14957253634929657, + 0.062163516879081726, + 0.08312437683343887 + ], + "credit_map": [ + 0.015688922256231308, + 0.010188093408942223, + 0.032107505947351456, + 0.008900435641407967, + 0.00993706937879324, + 0.021774953231215477, + 0.007124222815036774, + 0.022506223991513252, + 0.01193839032202959, + 0.015434865839779377, + 0.017514070495963097, + 0.005825426895171404, + 0.00871478021144867, + 0.04221352934837341, + 0.006844640243798494, + 0.02386552467942238, + 0.004556374624371529, + 0.011904299259185791, + 0.005556093528866768, + 0.0033732138108462095, + 0.006230161990970373, + 0.00740407407283783, + 0.004052531905472279, + 0.00446793157607317, + 0.0085352323949337, + 0.009320144541561604, + 0.010625627823174, + 0.0062202345579862595, + 0.006375430151820183, + 0.02175971493124962, + 0.005532508250325918, + 0.012484326958656311, + 0.004218931309878826, + 0.0047751073725521564, + 0.007053783629089594, + 0.0038707128260284662, + 0.007614110596477985, + 0.009551357477903366, + 0.0032868278212845325, + 0.0050466591492295265, + 0.010016480460762978, + 0.006872518919408321, + 0.02370453253388405, + 0.004059667233377695, + 0.007056220900267363, + 0.045944489538669586, + 0.004057750105857849, + 0.020405234768986702, + 0.009324144572019577, + 0.006314808037132025, + 0.03874371573328972, + 0.004192004445940256, + 0.009561335667967796, + 0.058305464684963226, + 0.004360733088105917, + 0.01782641001045704, + 0.009082437492907047, + 0.011681654490530491, + 0.029021991416811943, + 0.006602162960916758, + 0.014439212158322334, + 0.04963739961385727, + 0.006539381109178066, + 0.022568311542272568, + 0.005004127509891987, + 0.0050036427564918995, + 0.012578803114593029, + 0.00445219362154603, + 0.006126424763351679, + 0.016871415078639984, + 0.00467597134411335, + 0.007450943812727928, + 0.007623137440532446, + 0.006906156428158283, + 0.011081594973802567, + 0.005784986075013876, + 0.006862193811684847, + 0.0224568210542202, + 0.006078822538256645, + 0.016330672428011894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.03752141818404198 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.089073047041893 + ], + "opposite_logprob_tokens": [ + -2.4640729427337646 + ], + "logprob_reward": -0.089073047041893, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.089073047041893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3827713131904602, + 0.06987511366605759, + 0.14149872958660126, + 0.035293515771627426, + 0.023706670850515366, + 0.057959314435720444, + 0.06300763785839081, + 0.10793064534664154, + 0.08662024140357971, + 0.03133687749505043 + ], + "credit_map": [ + 0.039207205176353455, + 0.01845773123204708, + 0.1048152893781662, + 0.017564650624990463, + 0.020349986851215363, + 0.11605864763259888, + 0.01953543908894062, + 0.04678232967853546, + 0.006288568023592234, + 0.00939960591495037, + 0.012846444733440876, + 0.005127879790961742, + 0.007555338088423014, + 0.013245581649243832, + 0.005002244375646114, + 0.010409441776573658, + 0.01556218508630991, + 0.021402157843112946, + 0.024717947468161583, + 0.00824157427996397, + 0.009534863755106926, + 0.03085203841328621, + 0.009635575115680695, + 0.021552378311753273, + 0.004733504727482796, + 0.003951116930693388, + 0.0062638153322041035, + 0.002910443814471364, + 0.00305404094979167, + 0.005716557148844004, + 0.0021530084777623415, + 0.006511026062071323, + 0.002825750270858407, + 0.002117984229698777, + 0.0037244863342493773, + 0.0018594228895381093, + 0.004702854435890913, + 0.0028341792058199644, + 0.0017454363405704498, + 0.003896552138030529, + 0.0049597579054534435, + 0.004327261820435524, + 0.01174931600689888, + 0.002984400372952223, + 0.004913464654237032, + 0.01767089031636715, + 0.0030741773080080748, + 0.008280036970973015, + 0.005069209262728691, + 0.0047071571461856365, + 0.015019579790532589, + 0.004362569190561771, + 0.005350434686988592, + 0.01899135485291481, + 0.0033017522655427456, + 0.0062055704183876514, + 0.008750533685088158, + 0.013013838790357113, + 0.010579404421150684, + 0.008295627310872078, + 0.014446725137531757, + 0.02694023586809635, + 0.008531088940799236, + 0.01737317629158497, + 0.007955193519592285, + 0.009521755389869213, + 0.011460874229669571, + 0.006740503944456577, + 0.010530026629567146, + 0.019349422305822372, + 0.00872635655105114, + 0.012336098589003086, + 0.004055974539369345, + 0.003514450741931796, + 0.006305358372628689, + 0.0019711621571332216, + 0.00413757236674428, + 0.0054276855662465096, + 0.0029204017482697964, + 0.0030042692087590694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.089073047041893 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134675025939941 + ], + "opposite_logprob_tokens": [ + -0.31346747279167175 + ], + "logprob_reward": -1.3134675025939941, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134675025939941, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09535543620586395, + 0.07182014733552933, + 0.034002549946308136, + 0.08533082902431488, + 0.04173387214541435, + 0.07285279780626297, + 0.08489904552698135, + 0.2738366723060608, + 0.059335097670555115, + 0.18083347380161285 + ], + "credit_map": [ + 0.008722581900656223, + 0.012771446257829666, + 0.01812811754643917, + 0.008314640261232853, + 0.010844830423593521, + 0.021675655618309975, + 0.0053423382341861725, + 0.009555840864777565, + 0.007469623349606991, + 0.012758643366396427, + 0.012497982941567898, + 0.007026969455182552, + 0.00780109316110611, + 0.007922750897705555, + 0.005493343807756901, + 0.010849750600755215, + 0.004184175748378038, + 0.006473538000136614, + 0.004654234740883112, + 0.0036935799289494753, + 0.004710397683084011, + 0.0034781056456267834, + 0.002661513863131404, + 0.004147008992731571, + 0.007978367619216442, + 0.01539409626275301, + 0.010959966108202934, + 0.005955505650490522, + 0.008362584747374058, + 0.020131949335336685, + 0.004678085446357727, + 0.011870281770825386, + 0.0037260418757796288, + 0.00430018175393343, + 0.005363313481211662, + 0.0019724483136087656, + 0.008194120600819588, + 0.01055938471108675, + 0.0021846650633960962, + 0.00543372193351388, + 0.007699454203248024, + 0.008413894101977348, + 0.008063971996307373, + 0.006349597591906786, + 0.011336118914186954, + 0.0150928795337677, + 0.005149526987224817, + 0.010747364722192287, + 0.006478216033428907, + 0.01091628149151802, + 0.014617959968745708, + 0.0068682339042425156, + 0.012983079068362713, + 0.019341053441166878, + 0.006065387278795242, + 0.007628850173205137, + 0.028634509071707726, + 0.014732920564711094, + 0.09501539170742035, + 0.01295145507901907, + 0.015415082685649395, + 0.07491980493068695, + 0.01013651117682457, + 0.0220310278236866, + 0.0051107159815728664, + 0.006179350428283215, + 0.007591322995722294, + 0.003944420255720615, + 0.007740124594420195, + 0.013806600123643875, + 0.006599045358598232, + 0.008363528177142143, + 0.01901421509683132, + 0.01394664403051138, + 0.018883438780903816, + 0.009978660382330418, + 0.014301372691988945, + 0.04769519343972206, + 0.011640914715826511, + 0.04537304490804672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -1.3134675025939941 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018216803669929504 + ], + "opposite_logprob_tokens": [ + -4.018216609954834 + ], + "logprob_reward": -0.018216803669929504, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018216803669929504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11369198560714722, + 0.10050734877586365, + 0.03661840409040451, + 0.03398201987147331, + 0.07437222450971603, + 0.16795720160007477, + 0.0811254158616066, + 0.2008800506591797, + 0.03288332372903824, + 0.1579819917678833 + ], + "credit_map": [ + 0.010421030223369598, + 0.01067332923412323, + 0.03247730806469917, + 0.00526347616687417, + 0.006367545109242201, + 0.03131462261080742, + 0.004870960488915443, + 0.012303709052503109, + 0.007917126640677452, + 0.016168557107448578, + 0.017155170440673828, + 0.006812088657170534, + 0.010348787531256676, + 0.023473430424928665, + 0.007519535254687071, + 0.011112657375633717, + 0.004343112464994192, + 0.004869705531746149, + 0.0054603395983576775, + 0.004149284213781357, + 0.0049477024003863335, + 0.0047307927161455154, + 0.0039559826254844666, + 0.004161484073847532, + 0.0030006607994437218, + 0.005758154205977917, + 0.0040372563526034355, + 0.003098380519077182, + 0.00425978796556592, + 0.006296776235103607, + 0.0032109245657920837, + 0.004320081323385239, + 0.0066083874553442, + 0.00912638008594513, + 0.006985417101532221, + 0.006109581794589758, + 0.014732124283909798, + 0.013224913738667965, + 0.006708779372274876, + 0.01087663508951664, + 0.014778700657188892, + 0.016425788402557373, + 0.030131183564662933, + 0.008592525497078896, + 0.019544893875718117, + 0.047893427312374115, + 0.010910389944911003, + 0.019680293276906013, + 0.008060364983975887, + 0.011243141256272793, + 0.013118002563714981, + 0.00735187204554677, + 0.010106684640049934, + 0.01589239202439785, + 0.005993430037051439, + 0.009359532035887241, + 0.017146630212664604, + 0.01936461590230465, + 0.02225073240697384, + 0.011231483891606331, + 0.01847890391945839, + 0.06346188485622406, + 0.010707336477935314, + 0.038238488137722015, + 0.0034021507017314434, + 0.005619373172521591, + 0.003961971960961819, + 0.0035940574016422033, + 0.0052596190944314, + 0.004307258408516645, + 0.0031752819195389748, + 0.0035636124666780233, + 0.014808619394898415, + 0.013410981744527817, + 0.02622118592262268, + 0.007211643271148205, + 0.014997017569839954, + 0.046180300414562225, + 0.009904765523970127, + 0.02524748630821705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.018216803669929504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up centered on a bicycle wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a seamless dark gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the metal rim deform by fluidly stretching and bending into a non-round shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal spokes elongate like shimmering elastic and vibrate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007648824248462915 + ], + "opposite_logprob_tokens": [ + -4.882648944854736 + ], + "logprob_reward": -0.007648824248462915, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007648824248462915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14760936796665192, + 0.11281251907348633, + 0.035282377153635025, + 0.05564596876502037, + 0.07336409389972687, + 0.11330855637788773, + 0.05158795416355133, + 0.22378669679164886, + 0.07424838095903397, + 0.11235405504703522 + ], + "credit_map": [ + 0.014605056494474411, + 0.009257091209292412, + 0.039347074925899506, + 0.007247962057590485, + 0.011501561850309372, + 0.045483000576496124, + 0.0062121981754899025, + 0.01395542360842228, + 0.007311890367418528, + 0.014745417051017284, + 0.020486751571297646, + 0.006557312794029713, + 0.010802296921610832, + 0.03132188692688942, + 0.006878835614770651, + 0.014708137139678001, + 0.003819034667685628, + 0.005306474398821592, + 0.004834059625864029, + 0.003132025944069028, + 0.003703389083966613, + 0.006054280325770378, + 0.0031881299801170826, + 0.005244986619800329, + 0.00602450966835022, + 0.01109327282756567, + 0.006045382935553789, + 0.0052219717763364315, + 0.007627388462424278, + 0.007753764744848013, + 0.005239754915237427, + 0.006639930419623852, + 0.005523665342479944, + 0.0077162557281553745, + 0.007257732097059488, + 0.004306409042328596, + 0.01386385876685381, + 0.016810888424515724, + 0.005809348542243242, + 0.012075943872332573, + 0.009563860483467579, + 0.009972678497433662, + 0.020632341504096985, + 0.005447490606456995, + 0.011579000391066074, + 0.03469604626297951, + 0.006628204137086868, + 0.01478894054889679, + 0.005070423241704702, + 0.005504803732037544, + 0.010407941415905952, + 0.0034126844257116318, + 0.006190687883645296, + 0.011750450357794762, + 0.003737060818821192, + 0.0055139074102044106, + 0.014644451439380646, + 0.016854403540492058, + 0.03642648085951805, + 0.012702791020274162, + 0.01750270090997219, + 0.07709689438343048, + 0.01103330496698618, + 0.037525683641433716, + 0.007455546874552965, + 0.007307922467589378, + 0.008992712944746017, + 0.005725584924221039, + 0.007023929618299007, + 0.018315138295292854, + 0.006290895864367485, + 0.013136658817529678, + 0.020484866574406624, + 0.012455775402486324, + 0.02078564651310444, + 0.006080118007957935, + 0.014260069467127323, + 0.010860328562557697, + 0.010414821095764637, + 0.01701243966817856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.007648824248462915 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_0/index_tensor([6763, 6763])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_0/index_tensor([6763, 6763])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..35a7e68255c7a494bac9f543e86d4f4e8789869c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_0/index_tensor([6763, 6763])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a high close-up, the camera slowly and gently panning right, focusing on the beautiful legs and delicate feet of a woman soaking in a bathtub. Her skin gleams in the soft, warm light, appearing smooth and flawless. Her two delicate feet, with their manicured toenails painted a subtle pearlescent hue, are initially propped up on the gleaming white porcelain edge of the tub, which is filled with steaming, slightly milky water. Then, in one fluid motion, she slowly withdraws her left foot from the edge, water cascading from her heel in a shimmering stream. With graceful control, she gently and sensually rubs her left big toe along the length of her wet right leg, from ankle to knee. The scene is lit from the side by a soft, diffuse light source, casting gentle highlights on the water's surface and the contours of her skin, creating an atmosphere of tranquility, intimacy, and luxurious relaxation.", + "seed": 10134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_0/index_6763_seed_10134_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20270268768072128, + "vlm_logprob": -0.20270268768072128, + "vlm_yes_no_margin": 1.800000011920929, + "vlm_token_credit": -0.20270268768072128, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10268185287714005 + ], + "opposite_logprob_tokens": [ + -2.977681875228882 + ], + "logprob_reward": -0.10268185287714005, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.10268185287714005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0719270259141922, + 0.34401485323905945, + 0.08096390217542648, + 0.15032383799552917, + 0.06078728288412094, + 0.10706482082605362, + 0.04936979338526726, + 0.05185934156179428, + 0.04868830740451813, + 0.035000838339328766 + ], + "credit_map": [ + 0.0043659452348947525, + 0.010163625702261925, + 0.009277751669287682, + 0.011994742788374424, + 0.006033616606146097, + 0.007990948855876923, + 0.00924970768392086, + 0.012850684113800526, + 0.02220631204545498, + 0.06778472661972046, + 0.08588128536939621, + 0.046798937022686005, + 0.018826695159077644, + 0.02783789113163948, + 0.044443149119615555, + 0.030235856771469116, + 0.008611414581537247, + 0.012201403267681599, + 0.013353627175092697, + 0.00991042796522379, + 0.013723878189921379, + 0.008624770678579807, + 0.006543821655213833, + 0.007994561456143856, + 0.008413184434175491, + 0.050045061856508255, + 0.008943079970777035, + 0.006574895698577166, + 0.056019559502601624, + 0.007986235432326794, + 0.006214722990989685, + 0.006127103231847286, + 0.004821293987333775, + 0.009879888966679573, + 0.007185273338109255, + 0.006692723371088505, + 0.007708416786044836, + 0.012133758515119553, + 0.007637214381247759, + 0.004728713538497686, + 0.008756941184401512, + 0.02021852694451809, + 0.009426360949873924, + 0.005972688551992178, + 0.0268412996083498, + 0.0228181853890419, + 0.005757658742368221, + 0.007273152470588684, + 0.006466670427471399, + 0.005892477463930845, + 0.006112143862992525, + 0.0048041208647191525, + 0.006689667236059904, + 0.005967233330011368, + 0.005416310857981443, + 0.00802116934210062, + 0.004778706002980471, + 0.005644690245389938, + 0.005215278826653957, + 0.002557082800194621, + 0.0062150899320840836, + 0.007759532425552607, + 0.0056965891271829605, + 0.013992373831570148, + 0.006035009864717722, + 0.003996717277914286, + 0.004847272764891386, + 0.007816607132554054, + 0.006857165135443211, + 0.005761654581874609, + 0.007333834655582905, + 0.0060400464572012424, + 0.002496592700481415, + 0.00443426426500082, + 0.003649897174909711, + 0.00553206168115139, + 0.0055544753558933735, + 0.004510205704718828, + 0.005205987486988306, + 0.003617353504523635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.10268185287714005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25212517380714417 + ], + "opposite_logprob_tokens": [ + -1.5021251440048218 + ], + "logprob_reward": -0.25212517380714417, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25212517380714417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08772100508213043, + 0.11825714260339737, + 0.05889970064163208, + 0.2455425262451172, + 0.08904954046010971, + 0.07278177887201309, + 0.03985205665230751, + 0.08719474822282791, + 0.13501852750778198, + 0.06568293273448944 + ], + "credit_map": [ + 0.006364194676280022, + 0.009371854364871979, + 0.010171747766435146, + 0.015702590346336365, + 0.006896317936480045, + 0.011800728738307953, + 0.012717901729047298, + 0.014695663005113602, + 0.006886468734592199, + 0.023054221644997597, + 0.028260743245482445, + 0.013545854948461056, + 0.006970533169806004, + 0.011704514734447002, + 0.01811230182647705, + 0.00972250010818243, + 0.005049973260611296, + 0.009129759855568409, + 0.01065695658326149, + 0.007511164527386427, + 0.006036230828613043, + 0.004915494006127119, + 0.005729993339627981, + 0.009870131500065327, + 0.014345711097121239, + 0.06980575621128082, + 0.024252137169241905, + 0.014197858050465584, + 0.06732006371021271, + 0.014443700201809406, + 0.011791752651333809, + 0.029385559260845184, + 0.009139605797827244, + 0.01068867277354002, + 0.007994181476533413, + 0.009519103914499283, + 0.011766871437430382, + 0.02446143701672554, + 0.007311046589165926, + 0.00816862192004919, + 0.006961649749428034, + 0.0131006371229887, + 0.010993825271725655, + 0.007016351912170649, + 0.011526008136570454, + 0.007856489159166813, + 0.008463868871331215, + 0.006862948648631573, + 0.006275853607803583, + 0.004683355800807476, + 0.006124924402683973, + 0.003430622396990657, + 0.006252078339457512, + 0.0049459016881883144, + 0.004515869542956352, + 0.003623452503234148, + 0.008448834531009197, + 0.008411080576479435, + 0.007610277272760868, + 0.008345860056579113, + 0.006294952239841223, + 0.013157723471522331, + 0.010610860772430897, + 0.024315163493156433, + 0.012772006914019585, + 0.011759954504668713, + 0.02157711237668991, + 0.02433115616440773, + 0.017537778243422508, + 0.016091223806142807, + 0.0192151740193367, + 0.011734109371900558, + 0.007590294815599918, + 0.006348308175802231, + 0.008440402336418629, + 0.008620359934866428, + 0.007680424489080906, + 0.008760408498346806, + 0.007319687400013208, + 0.010923041962087154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.25212517380714417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156750082969666 + ], + "opposite_logprob_tokens": [ + -1.701567530632019 + ], + "logprob_reward": -0.20156750082969666, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156750082969666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03793594241142273, + 0.4733082354068756, + 0.04085389897227287, + 0.18251493573188782, + 0.13125450909137726, + 0.01641373336315155, + 0.008471590466797352, + 0.05756384879350662, + 0.03334053233265877, + 0.01834276132285595 + ], + "credit_map": [ + 0.002998913871124387, + 0.003574908245354891, + 0.004222814925014973, + 0.006979240104556084, + 0.0032499132212251425, + 0.004392093513160944, + 0.004593532532453537, + 0.00792453158646822, + 0.025654932484030724, + 0.11995445191860199, + 0.14075785875320435, + 0.0493650957942009, + 0.02626163326203823, + 0.029923079535365105, + 0.05034736171364784, + 0.03104386292397976, + 0.00343256164342165, + 0.006608285941183567, + 0.0058449492789804935, + 0.005104023031890392, + 0.0068522486835718155, + 0.00407273368909955, + 0.004554350860416889, + 0.0043847449123859406, + 0.006335222162306309, + 0.07015808671712875, + 0.008202915079891682, + 0.007207938004285097, + 0.07231537252664566, + 0.007232470903545618, + 0.004415297415107489, + 0.006647658534348011, + 0.011213547550141811, + 0.023161211982369423, + 0.016612183302640915, + 0.012827661819756031, + 0.029468238353729248, + 0.01455710083246231, + 0.013058983720839024, + 0.010355588048696518, + 0.0015783945564180613, + 0.002349244663491845, + 0.002395904390141368, + 0.0016795041738077998, + 0.002401829231530428, + 0.00224949698895216, + 0.002219818066805601, + 0.00153954210691154, + 0.0007646162412129343, + 0.0009541658801026642, + 0.0009915599366649985, + 0.0007415192667394876, + 0.0009205364622175694, + 0.001026442158035934, + 0.0016006820369511843, + 0.001472068834118545, + 0.004149190150201321, + 0.00461565051227808, + 0.0051217819564044476, + 0.0035773294512182474, + 0.004743870813399553, + 0.009435923770070076, + 0.0074263266287744045, + 0.018493780866265297, + 0.002382769947871566, + 0.002055459190160036, + 0.0030991167295724154, + 0.004486150573939085, + 0.0038287572097033262, + 0.004820159170776606, + 0.007602455094456673, + 0.0050656674429774284, + 0.0014271339168772101, + 0.0017832741141319275, + 0.0021942679304629564, + 0.002020239597186446, + 0.0019093214068561792, + 0.002754268702119589, + 0.0024811956100165844, + 0.0037730622570961714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.20156750082969666 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871106505393982 + ], + "opposite_logprob_tokens": [ + -1.137110710144043 + ], + "logprob_reward": -0.3871106505393982, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3871106505393982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0476645901799202, + 0.3916056156158447, + 0.053661566227674484, + 0.11034922301769257, + 0.1071525365114212, + 0.02264881134033203, + 0.026195209473371506, + 0.11071937531232834, + 0.0600249357521534, + 0.06997817754745483 + ], + "credit_map": [ + 0.004222055431455374, + 0.004127140156924725, + 0.005940334405750036, + 0.008036530576646328, + 0.00430626654997468, + 0.007162156049162149, + 0.006105468608438969, + 0.007764643058180809, + 0.023399651050567627, + 0.08952657878398895, + 0.1074444055557251, + 0.044441550970077515, + 0.023488808423280716, + 0.0291508287191391, + 0.045336611568927765, + 0.028817174956202507, + 0.005353055894374847, + 0.007357540540397167, + 0.008224613964557648, + 0.007160319481045008, + 0.00888970959931612, + 0.005382819566875696, + 0.005946539342403412, + 0.005346967838704586, + 0.006352546159178019, + 0.027768176048994064, + 0.011315228417515755, + 0.011680159717798233, + 0.023253627121448517, + 0.010016794316470623, + 0.007782625965774059, + 0.012180078774690628, + 0.010042700916528702, + 0.012553402222692966, + 0.01303879451006651, + 0.008100275881588459, + 0.014537512324750423, + 0.023181410506367683, + 0.010558981448411942, + 0.015139457769691944, + 0.002379460260272026, + 0.003479711478576064, + 0.003266196232289076, + 0.0018484622705727816, + 0.0035507469438016415, + 0.003068291349336505, + 0.001986699877306819, + 0.0030692417640239, + 0.0021715385373681784, + 0.004183589480817318, + 0.003363414201885462, + 0.0017639031866565347, + 0.0037868304643779993, + 0.0031627339776605368, + 0.003904090728610754, + 0.0038591090124100447, + 0.007582507096230984, + 0.00673663942143321, + 0.00895690731704235, + 0.00804510060697794, + 0.0091998977586627, + 0.017743898555636406, + 0.01599835231900215, + 0.036456067115068436, + 0.00740795535966754, + 0.004818630870431662, + 0.00859811156988144, + 0.009855019859969616, + 0.007208611816167831, + 0.006642287131398916, + 0.008720560930669308, + 0.006773760542273521, + 0.005497935228049755, + 0.00906691886484623, + 0.008058683015406132, + 0.011193249374628067, + 0.00812525674700737, + 0.009093265980482101, + 0.010217830538749695, + 0.008725043386220932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.3871106505393982 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002826035022736 + ], + "opposite_logprob_tokens": [ + -2.695028305053711 + ], + "logprob_reward": -0.07002826035022736, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002826035022736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05316014215350151, + 0.050664108246564865, + 0.03133728355169296, + 0.3001891076564789, + 0.0993056520819664, + 0.054625000804662704, + 0.03264438360929489, + 0.0649087131023407, + 0.10517039149999619, + 0.2079952359199524 + ], + "credit_map": [ + 0.0047521633096039295, + 0.005129320081323385, + 0.006150072440505028, + 0.00909959152340889, + 0.004449774976819754, + 0.006743441801518202, + 0.009627929888665676, + 0.007207842543721199, + 0.00440418254584074, + 0.005642793606966734, + 0.006657624617218971, + 0.007643799763172865, + 0.004384104628115892, + 0.007621249184012413, + 0.008267867378890514, + 0.006042480934411287, + 0.0026418701745569706, + 0.0037963050417602062, + 0.004518234170973301, + 0.004502126481384039, + 0.004819648806005716, + 0.003424298018217087, + 0.0033282863441854715, + 0.004306511487811804, + 0.012289880774915218, + 0.10972680151462555, + 0.015934685245156288, + 0.012760566547513008, + 0.11338740587234497, + 0.01454195473343134, + 0.007607770152390003, + 0.013940022327005863, + 0.0105460649356246, + 0.014722298830747604, + 0.007491302210837603, + 0.007989234291017056, + 0.015421450138092041, + 0.02108086086809635, + 0.009161652997136116, + 0.012892776168882847, + 0.005677447654306889, + 0.00993807427585125, + 0.005728388670831919, + 0.005746449343860149, + 0.00807972066104412, + 0.006148198153823614, + 0.007895035669207573, + 0.005411677528172731, + 0.004935231059789658, + 0.004964762832969427, + 0.005198011174798012, + 0.0029220846481621265, + 0.004746168851852417, + 0.003824494080618024, + 0.0030767321586608887, + 0.0029768969397991896, + 0.008076267316937447, + 0.006521393079310656, + 0.007377512287348509, + 0.004556058906018734, + 0.009230881929397583, + 0.008766035549342632, + 0.008525731973350048, + 0.011854824610054493, + 0.006489640101790428, + 0.006516463588923216, + 0.008044049143791199, + 0.012410786002874374, + 0.013630980625748634, + 0.014411120675504208, + 0.03147148713469505, + 0.012195855379104614, + 0.01736132614314556, + 0.028126006945967674, + 0.020602071657776833, + 0.024613376706838608, + 0.026348315179347992, + 0.029181255027651787, + 0.030411235988140106, + 0.03135163336992264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.07002826035022736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a high close-up, the camera slowly and gently panning right, focusing on the beautiful legs and delicate feet of a woman soaking in a bathtub. Her skin gleams in the soft, warm light, appearing smooth and flawless. Her two delicate feet, with their manicured toenails painted a subtle pearlescent hue, are initially propped up on the gleaming white porcelain edge of the tub, which is filled with steaming, slightly milky water. Then, in one fluid motion, she slowly withdraws her left foot from the edge, water cascading from her heel in a shimmering stream. With graceful control, she gently and sensually rubs her left big toe along the length of her wet right leg, from ankle to knee. The scene is lit from the side by a soft, diffuse light source, casting gentle highlights on the water's surface and the contours of her skin, creating an atmosphere of tranquility, intimacy, and luxurious relaxation.", + "seed": 10135, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_0/index_6763_seed_10135_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15399108380079268, + "vlm_logprob": -0.15399108380079268, + "vlm_yes_no_margin": 1.974999976158142, + "vlm_token_credit": -0.15399108380079268, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08433232456445694 + ], + "opposite_logprob_tokens": [ + -2.959332227706909 + ], + "logprob_reward": -0.08433232456445694, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08433232456445694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21202492713928223, + 0.06970883160829544, + 0.045029692351818085, + 0.07561156898736954, + 0.07667583972215652, + 0.06647072732448578, + 0.12889792025089264, + 0.06591395288705826, + 0.12919384241104126, + 0.13047274947166443 + ], + "credit_map": [ + 0.018970714882016182, + 0.019431745633482933, + 0.019693994894623756, + 0.017284097149968147, + 0.027107954025268555, + 0.0479254387319088, + 0.025650525465607643, + 0.03596045449376106, + 0.0070759691298007965, + 0.006028326228260994, + 0.008591575548052788, + 0.005583437625318766, + 0.008793388493359089, + 0.017841771245002747, + 0.007437751162797213, + 0.008356613107025623, + 0.004455489106476307, + 0.005616181064397097, + 0.005040864925831556, + 0.0036280991043895483, + 0.006445356644690037, + 0.00924984086304903, + 0.004521378315985203, + 0.006072483025491238, + 0.006479689385741949, + 0.015950975939631462, + 0.005614090245217085, + 0.005541803780943155, + 0.014665796421468258, + 0.017319945618510246, + 0.004793435335159302, + 0.005245829001069069, + 0.0116988904774189, + 0.010100179351866245, + 0.006199451629072428, + 0.007333371322602034, + 0.009494303725659847, + 0.01348672155290842, + 0.0070696426555514336, + 0.011293274350464344, + 0.007794365752488375, + 0.01069407444447279, + 0.004967003595083952, + 0.00835192110389471, + 0.016349423676729202, + 0.007367223501205444, + 0.004400293808430433, + 0.006546413991600275, + 0.013044469058513641, + 0.014390104450285435, + 0.012209263630211353, + 0.01374016236513853, + 0.019034121185541153, + 0.03173988312482834, + 0.013211270794272423, + 0.011528645642101765, + 0.005853666923940182, + 0.013200332410633564, + 0.009114408865571022, + 0.010024311020970345, + 0.006435886491090059, + 0.009896533563733101, + 0.0051573170349001884, + 0.006231495179235935, + 0.009058118797838688, + 0.024170108139514923, + 0.00929170846939087, + 0.009741835296154022, + 0.01561592798680067, + 0.04507719352841377, + 0.007580198813229799, + 0.008658750914037228, + 0.007562368642538786, + 0.04751520976424217, + 0.012436624616384506, + 0.014962779358029366, + 0.010552696883678436, + 0.016258062794804573, + 0.007020303979516029, + 0.014164709486067295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.08433232456445694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282019436359406 + ], + "opposite_logprob_tokens": [ + -2.017820119857788 + ], + "logprob_reward": -0.14282019436359406, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14282019436359406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12695637345314026, + 0.03228907659649849, + 0.07054971158504486, + 0.07411128282546997, + 0.03322098031640053, + 0.049119822680950165, + 0.13005009293556213, + 0.044860776513814926, + 0.1426500380039215, + 0.2961918115615845 + ], + "credit_map": [ + 0.010235056281089783, + 0.014766469597816467, + 0.023333925753831863, + 0.011915333569049835, + 0.019103355705738068, + 0.024209365248680115, + 0.010511133819818497, + 0.012881720438599586, + 0.004920653998851776, + 0.004212561063468456, + 0.0034856528509408236, + 0.003279317868873477, + 0.0037941518239676952, + 0.004559523891657591, + 0.003763023065403104, + 0.004274188540875912, + 0.006063421722501516, + 0.012994350865483284, + 0.00489806616678834, + 0.003476896323263645, + 0.017955584451556206, + 0.014363780617713928, + 0.0047884369269013405, + 0.006009174045175314, + 0.0081970589235425, + 0.011561461724340916, + 0.007312897592782974, + 0.008001280017197132, + 0.011002391576766968, + 0.01383958663791418, + 0.008195821195840836, + 0.00600078608840704, + 0.002905180212110281, + 0.005208767019212246, + 0.0043984525837004185, + 0.0030156320426613092, + 0.00648035341873765, + 0.004865073598921299, + 0.003264541504904628, + 0.003082981565967202, + 0.004863102920353413, + 0.008087502792477608, + 0.0030707477126270533, + 0.005889597814530134, + 0.014081885106861591, + 0.005270862951874733, + 0.0033813531044870615, + 0.00447477214038372, + 0.016138507053256035, + 0.019094910472631454, + 0.01644713059067726, + 0.014543404802680016, + 0.020805329084396362, + 0.018922092393040657, + 0.013239887543022633, + 0.01085883192718029, + 0.002878301776945591, + 0.00594966858625412, + 0.007456780876964331, + 0.006591583602130413, + 0.005213720723986626, + 0.006443685851991177, + 0.0055612423457205296, + 0.004765791352838278, + 0.009038129821419716, + 0.021114671602845192, + 0.006909278221428394, + 0.010992193594574928, + 0.019833600148558617, + 0.05057148635387421, + 0.009012616239488125, + 0.015178048983216286, + 0.018115459010004997, + 0.10046304762363434, + 0.016063611954450607, + 0.02317960374057293, + 0.019479505717754364, + 0.02214650623500347, + 0.011057483032345772, + 0.0856865867972374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.14282019436359406 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016179859638214 + ], + "opposite_logprob_tokens": [ + -1.701617956161499 + ], + "logprob_reward": -0.2016179859638214, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016179859638214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4484104514122009, + 0.04177866876125336, + 0.06027275696396828, + 0.046507686376571655, + 0.12096942216157913, + 0.062220972031354904, + 0.04735998809337616, + 0.04394717514514923, + 0.09091410040855408, + 0.03761876001954079 + ], + "credit_map": [ + 0.07577057927846909, + 0.05807221680879593, + 0.047980230301618576, + 0.05362910032272339, + 0.0609806589782238, + 0.057251688092947006, + 0.0475422628223896, + 0.04718371108174324, + 0.0034525583032518625, + 0.004246545024216175, + 0.0060245562344789505, + 0.0027972599491477013, + 0.0067444914020597935, + 0.009695145301520824, + 0.004061670508235693, + 0.004756442271173, + 0.006501931697130203, + 0.007535826880484819, + 0.006633917335420847, + 0.004483688622713089, + 0.007507792674005032, + 0.011415164917707443, + 0.0075425258837640285, + 0.008651906624436378, + 0.006056045647710562, + 0.007157878950238228, + 0.005833473522216082, + 0.003579476149752736, + 0.006968394853174686, + 0.007746801245957613, + 0.003371061524376273, + 0.0057945530861616135, + 0.019747313112020493, + 0.01860726810991764, + 0.010707682929933071, + 0.011730030179023743, + 0.015368912369012833, + 0.019793767482042313, + 0.011440873146057129, + 0.013573584146797657, + 0.0064135645516216755, + 0.009927633218467236, + 0.003575638635084033, + 0.008413946256041527, + 0.018523935228586197, + 0.0055992784909904, + 0.0038353477139025927, + 0.005931631196290255, + 0.004745959769934416, + 0.007135169114917517, + 0.006825235206633806, + 0.005192094016820192, + 0.006878596264868975, + 0.008297820575535297, + 0.004080221522599459, + 0.004204892087727785, + 0.0032500929664820433, + 0.006123187951743603, + 0.0068525513634085655, + 0.005782763008028269, + 0.004010891541838646, + 0.00623407494276762, + 0.007940187118947506, + 0.003753423923626542, + 0.005999858491122723, + 0.016114717349410057, + 0.0044087814167141914, + 0.006058318540453911, + 0.011386592872440815, + 0.03366835042834282, + 0.005323399789631367, + 0.007954072207212448, + 0.0028485949151217937, + 0.011715972796082497, + 0.0038894829340279102, + 0.005596606992185116, + 0.0031290727201849222, + 0.005741069093346596, + 0.0021908606868237257, + 0.0025071001145988703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.2016179859638214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25213825702667236 + ], + "opposite_logprob_tokens": [ + -1.5021382570266724 + ], + "logprob_reward": -0.25213825702667236, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25213825702667236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04640341177582741, + 0.039883922785520554, + 0.05419247969985008, + 0.3204382359981537, + 0.023941989988088608, + 0.03658639267086983, + 0.07190485298633575, + 0.06173638254404068, + 0.20859786868095398, + 0.13631445169448853 + ], + "credit_map": [ + 0.005497144069522619, + 0.004901512060314417, + 0.004595370963215828, + 0.003967495635151863, + 0.006580374203622341, + 0.008434622548520565, + 0.00504471966996789, + 0.007382174022495747, + 0.005055979359894991, + 0.00421725120395422, + 0.004180693533271551, + 0.0033324926625937223, + 0.0051122624427080154, + 0.008836083114147186, + 0.004217876587063074, + 0.004931282717734575, + 0.007814232259988785, + 0.0075492835603654385, + 0.005007361993193626, + 0.003290456021204591, + 0.008298300206661224, + 0.006425742991268635, + 0.006430089473724365, + 0.009377016685903072, + 0.022058313712477684, + 0.084274061024189, + 0.01810537278652191, + 0.013991158455610275, + 0.08945905417203903, + 0.05251706764101982, + 0.01968206837773323, + 0.020351145416498184, + 0.002979051088914275, + 0.003494343953207135, + 0.003302927827462554, + 0.002846967661753297, + 0.003238550154492259, + 0.0031832102686166763, + 0.0022850711829960346, + 0.002611867617815733, + 0.004344306420534849, + 0.004349955357611179, + 0.0037259545642882586, + 0.003218988887965679, + 0.009655777364969254, + 0.004068692680448294, + 0.0030156124848872423, + 0.004207106307148933, + 0.00836560595780611, + 0.009609105996787548, + 0.008793273940682411, + 0.007485547102987766, + 0.011641094461083412, + 0.014451666735112667, + 0.005356712266802788, + 0.0062018465250730515, + 0.005924974102526903, + 0.009848423302173615, + 0.006145657040178776, + 0.0057219271548092365, + 0.009080247953534126, + 0.00863718893378973, + 0.008610107935965061, + 0.0077678607776761055, + 0.016616102308034897, + 0.023447543382644653, + 0.017926830798387527, + 0.030730146914720535, + 0.02740614116191864, + 0.0583747923374176, + 0.019621046259999275, + 0.01447526179254055, + 0.012594959698617458, + 0.0289132222533226, + 0.015847494825720787, + 0.017195792868733406, + 0.010486481711268425, + 0.02180306985974312, + 0.008189279586076736, + 0.02128414250910282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.25213825702667236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890466570854187 + ], + "opposite_logprob_tokens": [ + -2.4640467166900635 + ], + "logprob_reward": -0.0890466570854187, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890466570854187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0986570194363594, + 0.051020704209804535, + 0.08583738654851913, + 0.03638524189591408, + 0.06620439141988754, + 0.08463925868272781, + 0.15316052734851837, + 0.06508120149374008, + 0.26927638053894043, + 0.08973795175552368 + ], + "credit_map": [ + 0.013863458298146725, + 0.011829609051346779, + 0.012912603095173836, + 0.01303178071975708, + 0.01365625485777855, + 0.013495029881596565, + 0.009160339832305908, + 0.01070795115083456, + 0.006066561210900545, + 0.005985376425087452, + 0.006333353463560343, + 0.003999305423349142, + 0.007159592118114233, + 0.011535034514963627, + 0.004447457380592823, + 0.005494020879268646, + 0.006205588113516569, + 0.012376101687550545, + 0.005908203311264515, + 0.00477573974058032, + 0.015310961753129959, + 0.013697052374482155, + 0.010407316498458385, + 0.017156418412923813, + 0.003202598774805665, + 0.007511555217206478, + 0.0031785022001713514, + 0.002175872214138508, + 0.009085009805858135, + 0.00597952026873827, + 0.0023791396524757147, + 0.0028730472549796104, + 0.00867366697639227, + 0.008596262894570827, + 0.0068609584122896194, + 0.006620689760893583, + 0.010272646322846413, + 0.012399756349623203, + 0.005765700247138739, + 0.007014710456132889, + 0.012202016077935696, + 0.01119507197290659, + 0.01008293591439724, + 0.00996465515345335, + 0.013162958435714245, + 0.00903556402772665, + 0.006202697288244963, + 0.01279336679726839, + 0.01628653146326542, + 0.020077120512723923, + 0.02020009607076645, + 0.01896565407514572, + 0.023724351078271866, + 0.026711193844676018, + 0.011302979663014412, + 0.015892595052719116, + 0.004868907853960991, + 0.006766132544726133, + 0.006343881133943796, + 0.010389449074864388, + 0.010073600336909294, + 0.009752104990184307, + 0.008677428588271141, + 0.008209695108234882, + 0.030661918222904205, + 0.026361331343650818, + 0.029884472489356995, + 0.04027267172932625, + 0.031386032700538635, + 0.05294027552008629, + 0.03399853780865669, + 0.023771138861775398, + 0.00954238511621952, + 0.009201369248330593, + 0.006973347160965204, + 0.0074464003555476665, + 0.006667285226285458, + 0.011668494902551174, + 0.006478197872638702, + 0.031760476529598236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.0890466570854187 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a high close-up, the camera slowly and gently panning right, focusing on the beautiful legs and delicate feet of a woman soaking in a bathtub. Her skin gleams in the soft, warm light, appearing smooth and flawless. Her two delicate feet, with their manicured toenails painted a subtle pearlescent hue, are initially propped up on the gleaming white porcelain edge of the tub, which is filled with steaming, slightly milky water. Then, in one fluid motion, she slowly withdraws her left foot from the edge, water cascading from her heel in a shimmering stream. With graceful control, she gently and sensually rubs her left big toe along the length of her wet right leg, from ankle to knee. The scene is lit from the side by a soft, diffuse light source, casting gentle highlights on the water's surface and the contours of her skin, creating an atmosphere of tranquility, intimacy, and luxurious relaxation.", + "seed": 10136, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_0/index_6763_seed_10136_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20315168425440788, + "vlm_logprob": -0.20315168425440788, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.20315168425440788, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0633777305483818 + ], + "opposite_logprob_tokens": [ + -3.438377618789673 + ], + "logprob_reward": -0.0633777305483818, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0633777305483818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05964192748069763, + 0.10075237601995468, + 0.2858874499797821, + 0.08682837337255478, + 0.20574654638767242, + 0.10355091094970703, + 0.03355763480067253, + 0.05092791095376015, + 0.03927066922187805, + 0.033836256712675095 + ], + "credit_map": [ + 0.00378756457939744, + 0.006009493954479694, + 0.005045630037784576, + 0.010859057307243347, + 0.0060964059084653854, + 0.009788230992853642, + 0.008676535449922085, + 0.009379011578857899, + 0.006276573054492474, + 0.009578355588018894, + 0.008130965754389763, + 0.015592392534017563, + 0.014048799872398376, + 0.013031023554503918, + 0.014843225479125977, + 0.019251041114330292, + 0.0872427299618721, + 0.008767449297010899, + 0.017488934099674225, + 0.016744138672947884, + 0.023162852972745895, + 0.11027702689170837, + 0.008454388938844204, + 0.013749925419688225, + 0.012919287197291851, + 0.012063340283930302, + 0.011941230855882168, + 0.007718253415077925, + 0.014425293542444706, + 0.010225946083664894, + 0.00878724455833435, + 0.008747772313654423, + 0.02112683467566967, + 0.06355148553848267, + 0.06494718044996262, + 0.011719402857124805, + 0.016173023730516434, + 0.007072123698890209, + 0.009768086485564709, + 0.011388405226171017, + 0.007511561270803213, + 0.016906870529055595, + 0.0119031285867095, + 0.012405332177877426, + 0.012806486338376999, + 0.009029479697346687, + 0.008214147761464119, + 0.02477389946579933, + 0.004368223715573549, + 0.00510576693341136, + 0.004634577315300703, + 0.002755014458671212, + 0.004607546143233776, + 0.004558951128274202, + 0.0034805373288691044, + 0.0040470147505402565, + 0.004066294524818659, + 0.007068426348268986, + 0.004431616049259901, + 0.004725052043795586, + 0.012363310903310776, + 0.009004414081573486, + 0.003814802737906575, + 0.005453996825963259, + 0.0022100750356912613, + 0.004586688708513975, + 0.0020834659226238728, + 0.0020607139449566603, + 0.00598245020955801, + 0.011668729595839977, + 0.003121856367215514, + 0.007556689903140068, + 0.0023567546159029007, + 0.0028525632806122303, + 0.0023982429411262274, + 0.0024906815961003304, + 0.008070239797234535, + 0.00728951022028923, + 0.004068526905030012, + 0.004309738986194134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.0633777305483818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521221339702606 + ], + "opposite_logprob_tokens": [ + -1.502122163772583 + ], + "logprob_reward": -0.2521221339702606, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521221339702606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20128214359283447, + 0.05079992860555649, + 0.060826316475868225, + 0.12118297815322876, + 0.07937320321798325, + 0.2929147779941559, + 0.0642499104142189, + 0.038584232330322266, + 0.06393171101808548, + 0.02685476467013359 + ], + "credit_map": [ + 0.004784185905009508, + 0.011622549034655094, + 0.006068079732358456, + 0.012129811570048332, + 0.07549037784337997, + 0.04535714536905289, + 0.015319661237299442, + 0.03051033243536949, + 0.005689677316695452, + 0.005104824434965849, + 0.004486788529902697, + 0.0057823373936116695, + 0.006482258904725313, + 0.005511906463652849, + 0.008969970047473907, + 0.008772166445851326, + 0.013496271334588528, + 0.002787545323371887, + 0.004095365758985281, + 0.009184037335216999, + 0.005168016999959946, + 0.018693355843424797, + 0.004061395768076181, + 0.0033403278794139624, + 0.014834810048341751, + 0.014349556528031826, + 0.027971362695097923, + 0.012834263034164906, + 0.013576834462583065, + 0.010863873176276684, + 0.011305010877549648, + 0.01544725988060236, + 0.01403578370809555, + 0.011293094605207443, + 0.011719750240445137, + 0.007699870970100164, + 0.012308607809245586, + 0.006552661303430796, + 0.008538470603525639, + 0.007224964443594217, + 0.02322859689593315, + 0.04451930150389671, + 0.035120852291584015, + 0.029737669974565506, + 0.04251253977417946, + 0.02749626152217388, + 0.02481185272336006, + 0.06548769772052765, + 0.011089358478784561, + 0.00936070829629898, + 0.009597094729542732, + 0.005661422852426767, + 0.009432094171643257, + 0.006941355299204588, + 0.0063897958025336266, + 0.005778083112090826, + 0.0036965669132769108, + 0.004958711098879576, + 0.0037505459040403366, + 0.004066163673996925, + 0.005359793081879616, + 0.004604265559464693, + 0.005434190854430199, + 0.006713995710015297, + 0.005226346664130688, + 0.004881649743765593, + 0.004483998753130436, + 0.004012545105069876, + 0.012305167503654957, + 0.01930432766675949, + 0.005178125109523535, + 0.008539552800357342, + 0.00316033698618412, + 0.0027747233398258686, + 0.0025435860734432936, + 0.0033043515868484974, + 0.003916646353900433, + 0.0037297760136425495, + 0.0034621928352862597, + 0.003963152877986431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.2521221339702606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31349658966064453 + ], + "opposite_logprob_tokens": [ + -1.3134965896606445 + ], + "logprob_reward": -0.31349658966064453, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31349658966064453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671604961156845, + 0.3696325123310089, + 0.2512053847312927, + 0.014361586421728134, + 0.053605612367391586, + 0.016036778688430786, + 0.01881617307662964, + 0.05149202421307564, + 0.02191009372472763, + 0.03577933460474014 + ], + "credit_map": [ + 0.002518258523195982, + 0.005413762293756008, + 0.0029005662072449923, + 0.006059923674911261, + 0.06981181353330612, + 0.0641770288348198, + 0.009121104143559933, + 0.0071580251678824425, + 0.022053031250834465, + 0.03971540182828903, + 0.03723279386758804, + 0.06022639572620392, + 0.049262192100286484, + 0.04383902996778488, + 0.05181901901960373, + 0.06548464298248291, + 0.07754404097795486, + 0.008576525375247002, + 0.013111975975334644, + 0.01564832776784897, + 0.021474570035934448, + 0.09461385756731033, + 0.006529451347887516, + 0.013706648722290993, + 0.0018472835654392838, + 0.0018591430271044374, + 0.0024900955613702536, + 0.0012883646413683891, + 0.002012078883126378, + 0.0016334477113559842, + 0.0012896646512672305, + 0.0019415081478655338, + 0.003038560738787055, + 0.021085353568196297, + 0.01706024259328842, + 0.002224849769845605, + 0.005780644249171019, + 0.0013271789066493511, + 0.001211537979543209, + 0.001877243397757411, + 0.0021678206976503134, + 0.003347712103277445, + 0.001806878950446844, + 0.0013575957855209708, + 0.0027106807101517916, + 0.0014339513145387173, + 0.0011229641968384385, + 0.002089174697175622, + 0.0036532373633235693, + 0.002914600307121873, + 0.002738062757998705, + 0.0021188058890402317, + 0.002629681257531047, + 0.001799749443307519, + 0.0017432630993425846, + 0.001218771911226213, + 0.0037741598207503557, + 0.008313867263495922, + 0.003297950839623809, + 0.0027063970919698477, + 0.012985285371541977, + 0.010384649969637394, + 0.0037447605282068253, + 0.006284953095018864, + 0.0020426656119525433, + 0.0025661184918135405, + 0.001920477137900889, + 0.0019396606367081404, + 0.002051066840067506, + 0.004341703839600086, + 0.0024208850227296352, + 0.00462751742452383, + 0.0024465727619826794, + 0.002187450649216771, + 0.001384402858093381, + 0.0030039006378501654, + 0.011339589953422546, + 0.010700102895498276, + 0.002608517650514841, + 0.0021087981294840574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.31349658966064453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486507833003998 + ], + "opposite_logprob_tokens": [ + -1.2236508131027222 + ], + "logprob_reward": -0.3486507833003998, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486507833003998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14163245260715485, + 0.06507320702075958, + 0.030484717339277267, + 0.042662546038627625, + 0.05059099197387695, + 0.45151880383491516, + 0.034298595041036606, + 0.04527577757835388, + 0.10394682735204697, + 0.03451601415872574 + ], + "credit_map": [ + 0.0026621597353368998, + 0.005193246994167566, + 0.0028983638621866703, + 0.005340771749615669, + 0.058794498443603516, + 0.039712775498628616, + 0.0069177886471152306, + 0.02011285349726677, + 0.0043264091946184635, + 0.007101401221007109, + 0.005958691239356995, + 0.011084499768912792, + 0.009111563675105572, + 0.007589408662170172, + 0.009700631722807884, + 0.010200602933764458, + 0.005319217219948769, + 0.001802991726435721, + 0.0028020672034472227, + 0.005166459362953901, + 0.0033397586084902287, + 0.006056748330593109, + 0.0035844866652041674, + 0.0024129878729581833, + 0.005658338777720928, + 0.006925026420503855, + 0.005942599382251501, + 0.0041805449873209, + 0.005532621871680021, + 0.0043033757247030735, + 0.0038828712422400713, + 0.00623716413974762, + 0.002875318517908454, + 0.018291587010025978, + 0.013850977644324303, + 0.0023807191755622625, + 0.00611360976472497, + 0.0018955993000417948, + 0.0027060515712946653, + 0.0024771266616880894, + 0.034820470958948135, + 0.06477947533130646, + 0.0557156503200531, + 0.04879220202565193, + 0.06113997846841812, + 0.03964205086231232, + 0.039314284920692444, + 0.10731472074985504, + 0.00687029305845499, + 0.006603705231100321, + 0.0038401761557906866, + 0.0034725749865174294, + 0.00429899338632822, + 0.0032538319937884808, + 0.0033988545183092356, + 0.0025601647794246674, + 0.0038091798778623343, + 0.005793032702058554, + 0.004454805515706539, + 0.0037584800738841295, + 0.0060325851663947105, + 0.004407672211527824, + 0.007017005700618029, + 0.010003017261624336, + 0.005790574476122856, + 0.009240183979272842, + 0.005728228949010372, + 0.005784795619547367, + 0.01666487194597721, + 0.032418619841337204, + 0.007372359279543161, + 0.020947203040122986, + 0.005097313318401575, + 0.0039411443285644054, + 0.0035912799648940563, + 0.005513363052159548, + 0.004109147936105728, + 0.004815630614757538, + 0.0043425592593848705, + 0.0031055770814418793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.3486507833003998 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038111183792352676 + ], + "opposite_logprob_tokens": [ + -3.2881112098693848 + ], + "logprob_reward": -0.038111183792352676, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038111183792352676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1158718541264534, + 0.05770312622189522, + 0.05903693288564682, + 0.05855235829949379, + 0.03934354707598686, + 0.05124574527144432, + 0.08356164395809174, + 0.2183527797460556, + 0.21525681018829346, + 0.10107516497373581 + ], + "credit_map": [ + 0.0038148167077451944, + 0.008996350690722466, + 0.004572744946926832, + 0.0078690554946661, + 0.03911547362804413, + 0.024149645119905472, + 0.010207287967205048, + 0.017146488651633263, + 0.008902110159397125, + 0.005992500577121973, + 0.006298731546849012, + 0.005476053338497877, + 0.006964452099055052, + 0.005678166169673204, + 0.010732469148933887, + 0.007658644579350948, + 0.016966192051768303, + 0.00237565697170794, + 0.003126192605122924, + 0.004495892208069563, + 0.004805341362953186, + 0.021394431591033936, + 0.002245705109089613, + 0.00362751679494977, + 0.008874751627445221, + 0.00922056008130312, + 0.012207542546093464, + 0.005036195740103722, + 0.0073928008787333965, + 0.004692648071795702, + 0.005762706510722637, + 0.005365155171602964, + 0.005121730268001556, + 0.006157987751066685, + 0.005207763519138098, + 0.004963184706866741, + 0.006255522835999727, + 0.0033301387447863817, + 0.003936435095965862, + 0.0043707843869924545, + 0.005766973830759525, + 0.007306838873773813, + 0.006348130293190479, + 0.008056333288550377, + 0.007257871795445681, + 0.004604381043463945, + 0.005666072480380535, + 0.0062391432002186775, + 0.016482708975672722, + 0.014458914287388325, + 0.010337171144783497, + 0.006715077441185713, + 0.011241985484957695, + 0.007492941338568926, + 0.008756114169955254, + 0.008076728321611881, + 0.014318265952169895, + 0.02146773226559162, + 0.014439661987125874, + 0.013474604114890099, + 0.05871535837650299, + 0.042031481862068176, + 0.015871280804276466, + 0.03803440183401108, + 0.018428126350045204, + 0.026720182970166206, + 0.018919849768280983, + 0.01763090491294861, + 0.02277394011616707, + 0.044077835977077484, + 0.021660739555954933, + 0.04504523053765297, + 0.007215358782559633, + 0.007218439597636461, + 0.006456661503762007, + 0.00615424569696188, + 0.03293566033244133, + 0.015330588445067406, + 0.007008037995547056, + 0.018756171688437462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.038111183792352676 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a high close-up, the camera slowly and gently panning right, focusing on the beautiful legs and delicate feet of a woman soaking in a bathtub. Her skin gleams in the soft, warm light, appearing smooth and flawless. Her two delicate feet, with their manicured toenails painted a subtle pearlescent hue, are initially propped up on the gleaming white porcelain edge of the tub, which is filled with steaming, slightly milky water. Then, in one fluid motion, she slowly withdraws her left foot from the edge, water cascading from her heel in a shimmering stream. With graceful control, she gently and sensually rubs her left big toe along the length of her wet right leg, from ankle to knee. The scene is lit from the side by a soft, diffuse light source, casting gentle highlights on the water's surface and the contours of her skin, creating an atmosphere of tranquility, intimacy, and luxurious relaxation.", + "seed": 10137, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_0/index_6763_seed_10137_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13030622228980066, + "vlm_logprob": -0.13030622228980066, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.13030622228980066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059838708490133286 + ], + "opposite_logprob_tokens": [ + -3.6848387718200684 + ], + "logprob_reward": -0.059838708490133286, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.059838708490133286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07712192088365555, + 0.03396780788898468, + 0.09561583399772644, + 0.023167692124843597, + 0.020519917830824852, + 0.03897802159190178, + 0.12071168422698975, + 0.03854388743638992, + 0.00904255174100399, + 0.5423306822776794 + ], + "credit_map": [ + 0.011951476335525513, + 0.009500066749751568, + 0.005733109079301357, + 0.004754563793540001, + 0.012442258186638355, + 0.016185805201530457, + 0.009011528454720974, + 0.007543118670582771, + 0.0033415942452847958, + 0.003624600125476718, + 0.004297669045627117, + 0.004505359102040529, + 0.005880819633603096, + 0.003805710468441248, + 0.003154907375574112, + 0.005357151385396719, + 0.011387648060917854, + 0.014436248689889908, + 0.015193878673017025, + 0.00777523685246706, + 0.015496381558477879, + 0.013125067576766014, + 0.00836129393428564, + 0.009840082377195358, + 0.0034165000542998314, + 0.0023964294232428074, + 0.0032476119231432676, + 0.001734351273626089, + 0.002323400927707553, + 0.0027319754008203745, + 0.0035927407443523407, + 0.003724681679159403, + 0.002503823721781373, + 0.002540695248171687, + 0.00312441261485219, + 0.0030986969359219074, + 0.0025577808264642954, + 0.0020479182712733746, + 0.001696274965070188, + 0.0029503165278583765, + 0.005212536547333002, + 0.004581037443131208, + 0.004661042243242264, + 0.0031610424630343914, + 0.009894372895359993, + 0.005376853048801422, + 0.0035903151147067547, + 0.002500820904970169, + 0.011122055351734161, + 0.009142691269516945, + 0.01068367063999176, + 0.0077417148277163506, + 0.012311290949583054, + 0.01532833930104971, + 0.029888154938817024, + 0.024493759498000145, + 0.004959600046277046, + 0.006544088013470173, + 0.0063546630553901196, + 0.004995402414351702, + 0.0048027122393250465, + 0.0050589251331985, + 0.0028416337445378304, + 0.002986862091347575, + 0.0009475582046434283, + 0.0014170149806886911, + 0.0011761486530303955, + 0.0008749988628551364, + 0.0010625949362292886, + 0.001906160730868578, + 0.0005496539524756372, + 0.0011084217112511396, + 0.05719134584069252, + 0.08102504163980484, + 0.029410777613520622, + 0.03090108186006546, + 0.1883334368467331, + 0.055773425847291946, + 0.023407848551869392, + 0.07628773897886276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.059838708490133286 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428227722644806 + ], + "opposite_logprob_tokens": [ + -2.017822742462158 + ], + "logprob_reward": -0.1428227722644806, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428227722644806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13344493508338928, + 0.1324245035648346, + 0.3217901885509491, + 0.05207427591085434, + 0.0636613667011261, + 0.10865096002817154, + 0.03309626132249832, + 0.052608292549848557, + 0.03656671568751335, + 0.06568249315023422 + ], + "credit_map": [ + 0.016239160671830177, + 0.013684953562915325, + 0.010702603496611118, + 0.014808379113674164, + 0.024873584508895874, + 0.020353740081191063, + 0.011334738694131374, + 0.021447770297527313, + 0.011877881363034248, + 0.01545105129480362, + 0.02542438916862011, + 0.014714171178638935, + 0.020937029272317886, + 0.017680654302239418, + 0.014792359434068203, + 0.011546973139047623, + 0.03474040329456329, + 0.0455913171172142, + 0.0549541637301445, + 0.027082405984401703, + 0.04362022876739502, + 0.04899469390511513, + 0.031573306769132614, + 0.035233668982982635, + 0.007915716618299484, + 0.004624123685061932, + 0.00692817522212863, + 0.006565520074218512, + 0.0061764903366565704, + 0.0065411897376179695, + 0.005721439141780138, + 0.007601622026413679, + 0.00897146388888359, + 0.0077024479396641254, + 0.009501639753580093, + 0.0077084703370928764, + 0.006854641251266003, + 0.005445437505841255, + 0.006241702940315008, + 0.011235561221837997, + 0.016773896291851997, + 0.012547533959150314, + 0.012942677363753319, + 0.008889340795576572, + 0.0224300529807806, + 0.013473352417349815, + 0.011605720035731792, + 0.009988395497202873, + 0.003066424513235688, + 0.003854981390759349, + 0.007012825924903154, + 0.0038288827054202557, + 0.0028836512938141823, + 0.0029227256309241056, + 0.0051282355561852455, + 0.004398533143103123, + 0.00743609294295311, + 0.007579749915748835, + 0.009090276435017586, + 0.008321125991642475, + 0.0070607587695121765, + 0.006531999912112951, + 0.0031832188833504915, + 0.0034050673712044954, + 0.0028752493672072887, + 0.0059219347313046455, + 0.0043268101289868355, + 0.0030476415995508432, + 0.004208211787045002, + 0.011181668378412724, + 0.0021610187832266092, + 0.002844181377440691, + 0.00651521235704422, + 0.009569384157657623, + 0.006351301446557045, + 0.007230957970023155, + 0.014337372034788132, + 0.010644149035215378, + 0.004841946996748447, + 0.0061921668238937855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.1428227722644806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991046607494354 + ], + "opposite_logprob_tokens": [ + -1.80491042137146 + ], + "logprob_reward": -0.17991046607494354, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991046607494354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13515335321426392, + 0.07526665180921555, + 0.1710806041955948, + 0.03437254577875137, + 0.0780012384057045, + 0.0869346335530281, + 0.12838660180568695, + 0.14684125781059265, + 0.0651252493262291, + 0.07883782684803009 + ], + "credit_map": [ + 0.01895804889500141, + 0.01822044886648655, + 0.011538264341652393, + 0.012548262253403664, + 0.024363191798329353, + 0.020453140139579773, + 0.01574471779167652, + 0.013327283784747124, + 0.007374269887804985, + 0.011317149735987186, + 0.007942468859255314, + 0.008801713585853577, + 0.0124960383400321, + 0.01088115107268095, + 0.008947095833718777, + 0.007506763096898794, + 0.0192385483533144, + 0.022759564220905304, + 0.027863774448633194, + 0.012369222939014435, + 0.029392650350928307, + 0.027605624869465828, + 0.018040230497717857, + 0.013811005279421806, + 0.004809624049812555, + 0.003384884912520647, + 0.0036523628514260054, + 0.002860086038708687, + 0.004207063931971788, + 0.007214176934212446, + 0.004501840099692345, + 0.003742509987205267, + 0.012301480397582054, + 0.0097133694216609, + 0.015557597391307354, + 0.009232774376869202, + 0.006817538756877184, + 0.005943235941231251, + 0.006635949481278658, + 0.011799297295510769, + 0.01415928453207016, + 0.008410925045609474, + 0.013513649813830853, + 0.006792501546442509, + 0.016656087711453438, + 0.010334325954318047, + 0.008686996065080166, + 0.008380872197449207, + 0.009364176541566849, + 0.015893438830971718, + 0.01799388788640499, + 0.009703624062240124, + 0.011752240359783173, + 0.014554468914866447, + 0.027816424146294594, + 0.021308336406946182, + 0.011748097836971283, + 0.02860930934548378, + 0.026839235797524452, + 0.021557608619332314, + 0.01907753385603428, + 0.020515164360404015, + 0.010766643099486828, + 0.007727679330855608, + 0.004554561339318752, + 0.011296296492218971, + 0.0058952998369932175, + 0.004256277810782194, + 0.008074330165982246, + 0.024438414722681046, + 0.0028354236856102943, + 0.003774654818698764, + 0.007593510672450066, + 0.01238620188087225, + 0.008206297643482685, + 0.006850076373666525, + 0.019267847761511803, + 0.010935043916106224, + 0.005121779628098011, + 0.008477073162794113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.17991046607494354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991773784160614 + ], + "opposite_logprob_tokens": [ + -1.8049176931381226 + ], + "logprob_reward": -0.17991773784160614, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991773784160614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0978139266371727, + 0.05233048275113106, + 0.287393182516098, + 0.03722240775823593, + 0.07841456681489944, + 0.05548442527651787, + 0.13610604405403137, + 0.06690938770771027, + 0.036163344979286194, + 0.15216225385665894 + ], + "credit_map": [ + 0.01279505155980587, + 0.009931374341249466, + 0.013621971011161804, + 0.008549364283680916, + 0.014133049175143242, + 0.017131192609667778, + 0.009703495539724827, + 0.011948422528803349, + 0.005784212611615658, + 0.006102034822106361, + 0.0060970960184931755, + 0.00690811313688755, + 0.009511941112577915, + 0.007204627152532339, + 0.00489121675491333, + 0.005831239279359579, + 0.029040856286883354, + 0.04061160981655121, + 0.04670875892043114, + 0.02361082285642624, + 0.04449581354856491, + 0.04159829765558243, + 0.02892295829951763, + 0.03240405023097992, + 0.005316720809787512, + 0.0027680257335305214, + 0.005846863612532616, + 0.0027407456655055285, + 0.005635984241962433, + 0.003975782543420792, + 0.003647058503702283, + 0.007291228044778109, + 0.011731158010661602, + 0.009149857796728611, + 0.016865240409970284, + 0.00934626441448927, + 0.006129419896751642, + 0.005920463241636753, + 0.006037450861185789, + 0.013234705664217472, + 0.006478878669440746, + 0.006136229261755943, + 0.009581513702869415, + 0.006404627114534378, + 0.008897680789232254, + 0.005859864875674248, + 0.00641100388020277, + 0.005714627914130688, + 0.011724688112735748, + 0.01188476663082838, + 0.014857267960906029, + 0.009630168788135052, + 0.01340752374380827, + 0.014633440412580967, + 0.03119659423828125, + 0.028771599754691124, + 0.005607170984148979, + 0.010109609924256802, + 0.010673761367797852, + 0.01058234740048647, + 0.007544290740042925, + 0.012060984037816525, + 0.004590270109474659, + 0.005740952678024769, + 0.003907885402441025, + 0.0059142205864191055, + 0.005253207869827747, + 0.0027862791903316975, + 0.0044296253472566605, + 0.00852685421705246, + 0.0024202847853302956, + 0.0029249866493046284, + 0.016334444284439087, + 0.02242368645966053, + 0.008821487426757812, + 0.008124141953885555, + 0.056017473340034485, + 0.014417586848139763, + 0.006684567779302597, + 0.01933886855840683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.17991773784160614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera starting with a high close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gently pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's feet painted with a subtle pearlescent hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman withdraw her left foot from the tub edge while water cascades from her heel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman rub her left big toe along her wet right leg from ankle to knee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904142677783966 + ], + "opposite_logprob_tokens": [ + -2.4640414714813232 + ], + "logprob_reward": -0.08904142677783966, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904142677783966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10153336077928543, + 0.05808271840214729, + 0.15387839078903198, + 0.03246947005391121, + 0.05037379637360573, + 0.050406213849782944, + 0.056045398116111755, + 0.06112854555249214, + 0.27158957719802856, + 0.16449253261089325 + ], + "credit_map": [ + 0.010611769743263721, + 0.012949668802320957, + 0.011135779321193695, + 0.011420855298638344, + 0.014418809674680233, + 0.014620203524827957, + 0.010485249571502209, + 0.01589101552963257, + 0.0077754901722073555, + 0.008672699332237244, + 0.00856705941259861, + 0.0057358392514288425, + 0.006339150946587324, + 0.0065210130997002125, + 0.007838915102183819, + 0.006632552482187748, + 0.016458168625831604, + 0.01922430470585823, + 0.026278452947735786, + 0.013049539178609848, + 0.019130777567625046, + 0.02439388446509838, + 0.017596520483493805, + 0.01774674654006958, + 0.004060743376612663, + 0.0037588668055832386, + 0.004335467703640461, + 0.0029998531099408865, + 0.0027404690627008677, + 0.004816295579075813, + 0.00477197952568531, + 0.004985797218978405, + 0.007655436638742685, + 0.0065247295424342155, + 0.009987736120820045, + 0.006479920353740454, + 0.0035378243774175644, + 0.0041128890588879585, + 0.003440150059759617, + 0.008635108359158039, + 0.008065765723586082, + 0.00511713745072484, + 0.004079802427440882, + 0.004602683708071709, + 0.013743584044277668, + 0.006972093600779772, + 0.00479153310880065, + 0.0030336121562868357, + 0.004400013480335474, + 0.007131447549909353, + 0.008847315795719624, + 0.004355175420641899, + 0.005283202975988388, + 0.005347290076315403, + 0.011482138186693192, + 0.009198816493153572, + 0.0065185860730707645, + 0.0100952023640275, + 0.011711440049111843, + 0.009913991205394268, + 0.005645502358675003, + 0.008685732260346413, + 0.004152263514697552, + 0.004405831918120384, + 0.018147042021155357, + 0.04771893471479416, + 0.027834434062242508, + 0.014338028617203236, + 0.03369400277733803, + 0.10565510392189026, + 0.009581153281033039, + 0.014620878733694553, + 0.01703435182571411, + 0.025302162393927574, + 0.008501235395669937, + 0.009223055094480515, + 0.05767463892698288, + 0.015320715494453907, + 0.008996500633656979, + 0.0224398672580719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.08904142677783966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_tensor([7038, 7038])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_tensor([7038, 7038])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f1bb1ea1c049475d02b79b4137c3da8ea36be875 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_tensor([7038, 7038])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a level angle, focuses on a young donkey, comfortably reclining on its side on a vast, soft, golden bale of hay. Warm afternoon sunlight filters through the fine dust in the air, tinting the entire frame with a golden hue and casting a bright halo around the edges of its fluffy, light-gray fur. Its long eyelashes flutter slightly in the soft light, and its tightly closed eyelids cautiously pry open a crack. Then, slowly, with a hint of sleepiness, they open fully, revealing a deep, moist black eye that reflects the golden sunlight. At this moment, the camera steadily and slowly pulls back, gradually revealing a wider, pastoral landscape: the bale of hay nestles in the center of a lush pasture. In the distance, a quaint wooden fence and a few scattered oak trees stand out. The azure sky is dotted with cotton-candy clouds, and the entire scene is permeated with a tranquil, lazy, and soothing atmosphere of the afternoon.", + "seed": 18277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_7038_seed_18277_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026499589998275043, + "vlm_logprob": -0.026499589998275043, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.026499589998275043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08298471570014954 + ], + "opposite_logprob_tokens": [ + -2.582984685897827 + ], + "logprob_reward": -0.08298471570014954, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08298471570014954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14976678788661957, + 0.0460379533469677, + 0.03359478712081909, + 0.01792001537978649, + 0.018528897315263748, + 0.022875554859638214, + 0.05016948655247688, + 0.09520107507705688, + 0.4657580554485321, + 0.10014736652374268 + ], + "credit_map": [ + 0.016283273696899414, + 0.02077498473227024, + 0.024420421570539474, + 0.024569153785705566, + 0.014418763108551502, + 0.009381781332194805, + 0.02160504087805748, + 0.018313361331820488, + 0.005425205919891596, + 0.006023746449500322, + 0.0051433974876999855, + 0.008033724501729012, + 0.005855012219399214, + 0.003995710052549839, + 0.005529697984457016, + 0.006031459663063288, + 0.0038699752185493708, + 0.005025795660912991, + 0.005765208508819342, + 0.005712069571018219, + 0.004051942843943834, + 0.0018319319933652878, + 0.003918483387678862, + 0.0034193776082247496, + 0.0021308937575668097, + 0.0024196167942136526, + 0.002179098315536976, + 0.0031365042086690664, + 0.001803528401069343, + 0.0013728940393775702, + 0.002002466470003128, + 0.0028750135097652674, + 0.002294254256412387, + 0.0025068027898669243, + 0.002328192349523306, + 0.0026612095534801483, + 0.0017470192397013307, + 0.0016894815489649773, + 0.0021559211891144514, + 0.0031460169702768326, + 0.0022599808871746063, + 0.003261679783463478, + 0.0026036263443529606, + 0.003982917871326208, + 0.003163015702739358, + 0.0019566533155739307, + 0.002809930359944701, + 0.0028377515263855457, + 0.0049468339420855045, + 0.006130983121693134, + 0.005960794165730476, + 0.008547981269657612, + 0.006555948872119188, + 0.00587829016149044, + 0.006176953669637442, + 0.005971703212708235, + 0.01204296201467514, + 0.014431817457079887, + 0.008479814045131207, + 0.011986643075942993, + 0.017332879826426506, + 0.009069024585187435, + 0.012165175750851631, + 0.009692751802504063, + 0.0618039071559906, + 0.04484569653868675, + 0.04204430431127548, + 0.06549443304538727, + 0.08291900902986526, + 0.05455820634961128, + 0.06297457963228226, + 0.051117949187755585, + 0.008836696855723858, + 0.006740453653037548, + 0.009356173686683178, + 0.01734038069844246, + 0.011614086106419563, + 0.01114376075565815, + 0.017433786764740944, + 0.01768202893435955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.08298471570014954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800649248063564 + ], + "opposite_logprob_tokens": [ + -4.634800434112549 + ], + "logprob_reward": -0.009800649248063564, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800649248063564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07408086210489273, + 0.04135172814130783, + 0.032747533172369, + 0.0278141051530838, + 0.02106354385614395, + 0.02751837484538555, + 0.05139518529176712, + 0.12026957422494888, + 0.5027848482131958, + 0.10097426176071167 + ], + "credit_map": [ + 0.00884552113711834, + 0.011077133938670158, + 0.012540250085294247, + 0.012020295485854149, + 0.006654039956629276, + 0.006172426976263523, + 0.008777801878750324, + 0.007993398234248161, + 0.005333494860678911, + 0.005303768441081047, + 0.009109445847570896, + 0.005166700109839439, + 0.004711276851594448, + 0.004031029995530844, + 0.003850179258733988, + 0.003845831612125039, + 0.0034595655743032694, + 0.0037283143028616905, + 0.00449084397405386, + 0.007380700204521418, + 0.0037644116673618555, + 0.0021854136139154434, + 0.0038253809325397015, + 0.00391290383413434, + 0.0038335458375513554, + 0.004202593583613634, + 0.0033238132018595934, + 0.0034869173541665077, + 0.0032230971846729517, + 0.0026244737673550844, + 0.002541969297453761, + 0.004577693063765764, + 0.002477200236171484, + 0.0027990182861685753, + 0.00335053913295269, + 0.004291629418730736, + 0.002463649958372116, + 0.0018786318833008409, + 0.001843583071604371, + 0.001959290122613311, + 0.0027254584711045027, + 0.00322136003524065, + 0.004736797418445349, + 0.006415386218577623, + 0.002566373674198985, + 0.0019613560289144516, + 0.0028535895980894566, + 0.0030380538664758205, + 0.006347748916596174, + 0.009182991459965706, + 0.004516470246016979, + 0.010102388449013233, + 0.005752270575612783, + 0.004330775234848261, + 0.004566582851111889, + 0.00659595662727952, + 0.013476690277457237, + 0.022903213277459145, + 0.0113705825060606, + 0.017575502395629883, + 0.016272157430648804, + 0.012806929647922516, + 0.0136647904291749, + 0.012199712917208672, + 0.05699724704027176, + 0.04303599148988724, + 0.06204900145530701, + 0.08930278569459915, + 0.074004627764225, + 0.06364918500185013, + 0.061236049979925156, + 0.05250994488596916, + 0.011220352724194527, + 0.006673928815871477, + 0.009581726975739002, + 0.020057685673236847, + 0.012345666065812111, + 0.010842317715287209, + 0.01660076156258583, + 0.013651824556291103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.009800649248063564 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821231961250305 + ], + "opposite_logprob_tokens": [ + -3.5298211574554443 + ], + "logprob_reward": -0.029821231961250305, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821231961250305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08775331825017929, + 0.08420508354902267, + 0.06055009737610817, + 0.02900327928364277, + 0.03814345598220825, + 0.03327775374054909, + 0.06041795760393143, + 0.2324836254119873, + 0.2702397108078003, + 0.10392570495605469 + ], + "credit_map": [ + 0.011370670981705189, + 0.010179356671869755, + 0.008303054608404636, + 0.010172341018915176, + 0.01134034339338541, + 0.009643652476370335, + 0.011334811337292194, + 0.015409092418849468, + 0.012008431367576122, + 0.013075816445052624, + 0.012844429351389408, + 0.01269836537539959, + 0.007959442213177681, + 0.005025975871831179, + 0.006994782947003841, + 0.013597839511930943, + 0.009152076207101345, + 0.008146568201482296, + 0.009017054922878742, + 0.009364396333694458, + 0.005990032106637955, + 0.004005704540759325, + 0.005238128826022148, + 0.009636137634515762, + 0.00400423863902688, + 0.003688274649903178, + 0.0038534642662853003, + 0.005135891959071159, + 0.0021869679912924767, + 0.0024404588621109724, + 0.003039955860003829, + 0.00465402752161026, + 0.005128372926265001, + 0.005316206719726324, + 0.0045153116807341576, + 0.007625283673405647, + 0.0029569165781140327, + 0.0028570760041475296, + 0.003845119383186102, + 0.005899169947952032, + 0.004993526730686426, + 0.0046207718551158905, + 0.0043746172450482845, + 0.0060266233049333096, + 0.002929206006228924, + 0.0026694757398217916, + 0.003077099798247218, + 0.004586431663483381, + 0.006778957787901163, + 0.007402657065540552, + 0.004671348724514246, + 0.009735919535160065, + 0.006876835133880377, + 0.006849435158073902, + 0.007032609544694424, + 0.011070194654166698, + 0.028421377763152122, + 0.03879255801439285, + 0.019722407683730125, + 0.02603081241250038, + 0.035419996827840805, + 0.0185263492166996, + 0.030318954959511757, + 0.03525116294622421, + 0.027587158605456352, + 0.020549573004245758, + 0.02207367680966854, + 0.06056194007396698, + 0.03679186850786209, + 0.018738234415650368, + 0.04518498480319977, + 0.038752276450395584, + 0.012153864838182926, + 0.010268368758261204, + 0.010186738334596157, + 0.013468818739056587, + 0.013607272878289223, + 0.019144609570503235, + 0.013747395016252995, + 0.011348641477525234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.029821231961250305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252846051007509 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252846051007509, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252846051007509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10028759390115738, + 0.037978533655405045, + 0.02728043496608734, + 0.022819293662905693, + 0.015084675513207912, + 0.03025035560131073, + 0.06438376009464264, + 0.14357317984104156, + 0.4076188802719116, + 0.15072324872016907 + ], + "credit_map": [ + 0.010044450871646404, + 0.008599966764450073, + 0.01383594237267971, + 0.015934476628899574, + 0.008172629401087761, + 0.016425473615527153, + 0.010700496844947338, + 0.016574157401919365, + 0.005860507022589445, + 0.003778718411922455, + 0.006991437170654535, + 0.006209973245859146, + 0.0031507096718996763, + 0.0029287897050380707, + 0.0034355183597654104, + 0.005622881930321455, + 0.0038007497787475586, + 0.0035870675928890705, + 0.003413519822061062, + 0.004620332270860672, + 0.0026203140150755644, + 0.002232162281870842, + 0.0027238246984779835, + 0.0042824638076126575, + 0.003599051386117935, + 0.0021413657814264297, + 0.005014253780245781, + 0.003360903123393655, + 0.0020672252867370844, + 0.00134718744084239, + 0.00215149181894958, + 0.003137813415378332, + 0.0017950719920918345, + 0.0015269608702510595, + 0.0018160662148147821, + 0.0020461378153413534, + 0.0017109180334955454, + 0.0022676526568830013, + 0.0013890000991523266, + 0.0025328691117465496, + 0.003925621975213289, + 0.0038528735749423504, + 0.0057646771892905235, + 0.003994705155491829, + 0.003972318489104509, + 0.003261780133470893, + 0.0029352374840527773, + 0.002543141134083271, + 0.006912936456501484, + 0.012118342332541943, + 0.006390225142240524, + 0.009022722952067852, + 0.007557924371212721, + 0.006448725238442421, + 0.0062980349175632, + 0.009634844958782196, + 0.020876802504062653, + 0.022350799292325974, + 0.014305656775832176, + 0.015248163603246212, + 0.022582005709409714, + 0.019343342632055283, + 0.012699085287749767, + 0.016167327761650085, + 0.045935992151498795, + 0.03714267909526825, + 0.05195421725511551, + 0.07512211054563522, + 0.05022239312529564, + 0.054275911301374435, + 0.049738891422748566, + 0.04322667047381401, + 0.01724916137754917, + 0.01515769399702549, + 0.010444794781506062, + 0.027233010157942772, + 0.02089575305581093, + 0.01620975323021412, + 0.020800631493330002, + 0.022732459008693695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.005252846051007509 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638507030904293 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638507030904293, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638507030904293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11817321181297302, + 0.05160526931285858, + 0.02752685733139515, + 0.01667187549173832, + 0.043185848742723465, + 0.06340634077787399, + 0.03801194578409195, + 0.14463365077972412, + 0.30275407433509827, + 0.1940309703350067 + ], + "credit_map": [ + 0.014429674483835697, + 0.016206666827201843, + 0.012453224509954453, + 0.014929932542145252, + 0.012047715485095978, + 0.02166096493601799, + 0.012784816324710846, + 0.013660205528140068, + 0.006496604066342115, + 0.006362416315823793, + 0.008007143624126911, + 0.00963300559669733, + 0.003965169191360474, + 0.005346491001546383, + 0.004887736868113279, + 0.006906701717525721, + 0.0036273282021284103, + 0.003543339902535081, + 0.003655140521004796, + 0.0059546916745603085, + 0.0024762293323874474, + 0.0016086101531982422, + 0.003089963225647807, + 0.0035715552512556314, + 0.0022544546518474817, + 0.001881902338936925, + 0.0025505712255835533, + 0.0030775154009461403, + 0.0014060026733204722, + 0.0012364148860797286, + 0.0018855532398447394, + 0.0023794607259333134, + 0.005585463251918554, + 0.005353953689336777, + 0.009138517081737518, + 0.007456788327544928, + 0.0044689178466796875, + 0.0031617647036910057, + 0.004030662588775158, + 0.003989781253039837, + 0.009129711426794529, + 0.007501570507884026, + 0.01659633405506611, + 0.009371922351419926, + 0.005619778763502836, + 0.0035278720315545797, + 0.005929103586822748, + 0.00573004549369216, + 0.0037797060795128345, + 0.005095093511044979, + 0.0034056787844747305, + 0.007974564097821712, + 0.005175422877073288, + 0.003433971432968974, + 0.004392395261675119, + 0.004755112808197737, + 0.016909394413232803, + 0.022140854969620705, + 0.010827519930899143, + 0.017150091007351875, + 0.018613530322909355, + 0.02243569679558277, + 0.018593987450003624, + 0.017962582409381866, + 0.030069101601839066, + 0.02789616957306862, + 0.041411906480789185, + 0.06837403029203415, + 0.03281453251838684, + 0.03328781574964523, + 0.0426289401948452, + 0.026271570473909378, + 0.031755100935697556, + 0.032985787838697433, + 0.023663010448217392, + 0.02285698615014553, + 0.01686367206275463, + 0.026440884917974472, + 0.017099641263484955, + 0.022365890443325043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.004638507030904293 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a level angle, focuses on a young donkey, comfortably reclining on its side on a vast, soft, golden bale of hay. Warm afternoon sunlight filters through the fine dust in the air, tinting the entire frame with a golden hue and casting a bright halo around the edges of its fluffy, light-gray fur. Its long eyelashes flutter slightly in the soft light, and its tightly closed eyelids cautiously pry open a crack. Then, slowly, with a hint of sleepiness, they open fully, revealing a deep, moist black eye that reflects the golden sunlight. At this moment, the camera steadily and slowly pulls back, gradually revealing a wider, pastoral landscape: the bale of hay nestles in the center of a lush pasture. In the distance, a quaint wooden fence and a few scattered oak trees stand out. The azure sky is dotted with cotton-candy clouds, and the entire scene is permeated with a tranquil, lazy, and soothing atmosphere of the afternoon.", + "seed": 18278, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_7038_seed_18278_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020842136070132255, + "vlm_logprob": -0.020842136070132255, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.020842136070132255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025841571390628815 + ], + "opposite_logprob_tokens": [ + -3.900841474533081 + ], + "logprob_reward": -0.025841571390628815, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025841571390628815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06585531681776047, + 0.07446609437465668, + 0.02424708567559719, + 0.6062862277030945, + 0.029883794486522675, + 0.013440919108688831, + 0.025020718574523926, + 0.02675202302634716, + 0.11691640317440033, + 0.017131425440311432 + ], + "credit_map": [ + 0.0069060069508850574, + 0.009609758853912354, + 0.013941007666289806, + 0.010169806890189648, + 0.0062066055834293365, + 0.004211066290736198, + 0.007038361858576536, + 0.007772705052047968, + 0.008996003307402134, + 0.00930989719927311, + 0.011148004792630672, + 0.011076846159994602, + 0.00994037464261055, + 0.003830419387668371, + 0.01239057071506977, + 0.007773978635668755, + 0.0026571208145469427, + 0.003052766900509596, + 0.0035197902470827103, + 0.005389058031141758, + 0.002545332070440054, + 0.0018576845759525895, + 0.0030998988077044487, + 0.0021254357416182756, + 0.09903287887573242, + 0.03267544507980347, + 0.06725329160690308, + 0.25718408823013306, + 0.02771124057471752, + 0.020017465576529503, + 0.05078291520476341, + 0.05162894353270531, + 0.005199186969548464, + 0.004811557941138744, + 0.003617693670094013, + 0.004666203632950783, + 0.0028879670426249504, + 0.002141231670975685, + 0.0028809248469769955, + 0.0036790308076888323, + 0.0020898887887597084, + 0.0017561920685693622, + 0.0017854261677712202, + 0.0018684291280806065, + 0.00145338149741292, + 0.001775493728928268, + 0.0012941710883751512, + 0.0014179374556988478, + 0.0032027377746999264, + 0.003347197314724326, + 0.003066999139264226, + 0.004845390561968088, + 0.0037023036275058985, + 0.0018630784470587969, + 0.00313584809191525, + 0.0018571639666333795, + 0.004462245386093855, + 0.0025555442553013563, + 0.002858635736629367, + 0.004618756473064423, + 0.003240369725972414, + 0.002282529603689909, + 0.003009310457855463, + 0.003724633250385523, + 0.00905455369502306, + 0.029386261478066444, + 0.041329801082611084, + 0.01226365938782692, + 0.005559974815696478, + 0.003994631115347147, + 0.005268079228699207, + 0.01005945447832346, + 0.0026357986498624086, + 0.0027364916168153286, + 0.002523466944694519, + 0.0025432067923247814, + 0.0013591505121439695, + 0.0013817563885822892, + 0.0018739651422947645, + 0.00207758997566998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.025841571390628815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263650044798851 + ], + "opposite_logprob_tokens": [ + -3.651365041732788 + ], + "logprob_reward": -0.0263650044798851, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263650044798851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08369994163513184, + 0.07383696734905243, + 0.06591616570949554, + 0.27296024560928345, + 0.04966788738965988, + 0.0629056841135025, + 0.1382455676794052, + 0.051005054265260696, + 0.16969092190265656, + 0.03207157924771309 + ], + "credit_map": [ + 0.013077073730528355, + 0.007955926470458508, + 0.006494759116321802, + 0.01479897741228342, + 0.010009020566940308, + 0.004601433407515287, + 0.012571332044899464, + 0.014191417023539543, + 0.00846156757324934, + 0.009606637060642242, + 0.014716320671141148, + 0.010130668990314007, + 0.009124099276959896, + 0.00383983482606709, + 0.010164986364543438, + 0.007792846765369177, + 0.00615787273272872, + 0.010361066088080406, + 0.010260033421218395, + 0.01280104648321867, + 0.0062635899521410465, + 0.004438107367604971, + 0.008450571447610855, + 0.00718387495726347, + 0.04559849575161934, + 0.015860186889767647, + 0.029568787664175034, + 0.1098242849111557, + 0.01572699099779129, + 0.010121661238372326, + 0.024110179394483566, + 0.022149652242660522, + 0.0050185914151370525, + 0.006692646071314812, + 0.004737526644021273, + 0.00907783955335617, + 0.006646761205047369, + 0.003595571266487241, + 0.00823583547025919, + 0.005663114134222269, + 0.010984888300299644, + 0.007419354747980833, + 0.012392287142574787, + 0.010245803743600845, + 0.0051715802401304245, + 0.005168930161744356, + 0.006332347169518471, + 0.00519048934802413, + 0.011575277894735336, + 0.014283106662333012, + 0.015563054010272026, + 0.031059708446264267, + 0.018686888739466667, + 0.010295658372342587, + 0.021640203893184662, + 0.015141662210226059, + 0.007193099707365036, + 0.005102438852190971, + 0.005176892504096031, + 0.00815576035529375, + 0.006151990033686161, + 0.006739201955497265, + 0.005386212840676308, + 0.007099458482116461, + 0.013134140521287918, + 0.041872136294841766, + 0.05840526893734932, + 0.018680531531572342, + 0.007786586415022612, + 0.007671646773815155, + 0.007644125260412693, + 0.014496502466499805, + 0.002267014468088746, + 0.004899173509329557, + 0.007395720109343529, + 0.004015246406197548, + 0.003217048943042755, + 0.0029873973689973354, + 0.003285276470705867, + 0.004004701506346464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.0263650044798851 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812897205352783 + ], + "opposite_logprob_tokens": [ + -3.2881288528442383 + ], + "logprob_reward": -0.03812897205352783, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812897205352783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04565131664276123, + 0.019372761249542236, + 0.03202694281935692, + 0.5215365290641785, + 0.0493183434009552, + 0.14931590855121613, + 0.02744228206574917, + 0.04727041348814964, + 0.08464366942644119, + 0.023421769961714745 + ], + "credit_map": [ + 0.005598474759608507, + 0.004515770822763443, + 0.00623064860701561, + 0.008798274211585522, + 0.0035718774888664484, + 0.0024951589293777943, + 0.00658143125474453, + 0.00785968266427517, + 0.0020787320099771023, + 0.002443918026983738, + 0.0026464303955435753, + 0.0028059782925993204, + 0.0027356951031833887, + 0.0012711349409073591, + 0.0032588194590061903, + 0.0021320537198334932, + 0.003427902003750205, + 0.003643495263531804, + 0.00563869858160615, + 0.006604406516999006, + 0.002604263834655285, + 0.00235706171952188, + 0.004749406594783068, + 0.003001706674695015, + 0.08770326524972916, + 0.028072526678442955, + 0.057778917253017426, + 0.2192171961069107, + 0.024261711165308952, + 0.018008898943662643, + 0.044079091399908066, + 0.042414963245391846, + 0.007139620371162891, + 0.0061289784498512745, + 0.004689307883381844, + 0.009841516613960266, + 0.006464868318289518, + 0.003884651465341449, + 0.006196645088493824, + 0.004972751252353191, + 0.013676389120519161, + 0.017824791371822357, + 0.024419937282800674, + 0.018217699602246284, + 0.02669585309922695, + 0.013273714110255241, + 0.018010858446359634, + 0.017196660861372948, + 0.0030224286019802094, + 0.004964327905327082, + 0.0027074243407696486, + 0.004604061599820852, + 0.0030260453931987286, + 0.0024674527812749147, + 0.003169544041156769, + 0.0034809978678822517, + 0.0035264051984995604, + 0.005326415877789259, + 0.004699560813605785, + 0.010274211876094341, + 0.008034718222916126, + 0.004113145638257265, + 0.006605807226151228, + 0.004690148402005434, + 0.006399031262844801, + 0.0199718177318573, + 0.028326809406280518, + 0.008217129856348038, + 0.0055206529796123505, + 0.0032421289943158627, + 0.005141063127666712, + 0.007825038395822048, + 0.004153937567025423, + 0.0033983527682721615, + 0.005126933101564646, + 0.002301278756931424, + 0.001902456977404654, + 0.002056099707260728, + 0.002144555328413844, + 0.002338153775781393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.03812897205352783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978058110922575 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.00978058110922575, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978058110922575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06204577907919884, + 0.042818691581487656, + 0.08329951018095016, + 0.29563236236572266, + 0.07918909937143326, + 0.10612241923809052, + 0.17485187947750092, + 0.05484959855675697, + 0.023788008838891983, + 0.07740261405706406 + ], + "credit_map": [ + 0.00741187110543251, + 0.007237887475639582, + 0.005825071129947901, + 0.010675681754946709, + 0.0073814187198877335, + 0.004647784400731325, + 0.011469652876257896, + 0.007396409288048744, + 0.005860842298716307, + 0.005441001150757074, + 0.005656953901052475, + 0.0061287786811590195, + 0.005385404918342829, + 0.0028562231454998255, + 0.00696374848484993, + 0.004525742493569851, + 0.009117001667618752, + 0.011833555996418, + 0.011430216953158379, + 0.015159073285758495, + 0.006887204013764858, + 0.004573351703584194, + 0.011247532442212105, + 0.013051575981080532, + 0.04700212553143501, + 0.018681887537240982, + 0.03293496370315552, + 0.12129645049571991, + 0.013480167835950851, + 0.01133425347507, + 0.024470679461956024, + 0.026431851089000702, + 0.008698783814907074, + 0.011011652648448944, + 0.010680248960852623, + 0.01116721797734499, + 0.009609838016331196, + 0.006555536761879921, + 0.011349188163876534, + 0.010116633027791977, + 0.010043195448815823, + 0.014584353193640709, + 0.013258442282676697, + 0.010565799660980701, + 0.019680211320519447, + 0.012711695395410061, + 0.012343760579824448, + 0.012934958562254906, + 0.02120073139667511, + 0.019104141741991043, + 0.021087665110826492, + 0.026280036196112633, + 0.02945152297616005, + 0.01412904728204012, + 0.024943774566054344, + 0.018654972314834595, + 0.005357455927878618, + 0.005448858253657818, + 0.006387689150869846, + 0.010471007786691189, + 0.00835680216550827, + 0.004432337358593941, + 0.007059731520712376, + 0.007335719186812639, + 0.002118482952937484, + 0.004862474277615547, + 0.005282083060592413, + 0.002512806560844183, + 0.0019455974688753486, + 0.0023476744536310434, + 0.002297902712598443, + 0.002420988166704774, + 0.01884206011891365, + 0.014564055018126965, + 0.01686796545982361, + 0.00789235346019268, + 0.006258493755012751, + 0.00399310328066349, + 0.005247573368251324, + 0.0037370063364505768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.00978058110922575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409455131739378 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.00409455131739378, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409455131739378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11531402915716171, + 0.09068240225315094, + 0.06334659457206726, + 0.4379907548427582, + 0.04448365420103073, + 0.03120262548327446, + 0.035240937024354935, + 0.046713292598724365, + 0.0960317924618721, + 0.038993921130895615 + ], + "credit_map": [ + 0.018096668645739555, + 0.007030945736914873, + 0.015404216013848782, + 0.01157627534121275, + 0.008731657639145851, + 0.010212103836238384, + 0.017608771100640297, + 0.026653390377759933, + 0.010495833121240139, + 0.009988279081881046, + 0.010943564586341381, + 0.013321269303560257, + 0.01048651710152626, + 0.007056834641844034, + 0.014507194049656391, + 0.013882912695407867, + 0.00522108469158411, + 0.007779704872518778, + 0.008108226582407951, + 0.013666721060872078, + 0.004401957616209984, + 0.0054412540048360825, + 0.009185182861983776, + 0.009542456828057766, + 0.0710151195526123, + 0.02357364632189274, + 0.04691910371184349, + 0.1802220642566681, + 0.02037370391190052, + 0.018678419291973114, + 0.0358453243970871, + 0.04136337339878082, + 0.006647525355219841, + 0.009654625318944454, + 0.003552147187292576, + 0.005233697593212128, + 0.005543075501918793, + 0.0036078717093914747, + 0.005829763598740101, + 0.004414945840835571, + 0.003778208279982209, + 0.004601187538355589, + 0.003905757563188672, + 0.004882975947111845, + 0.0039779916405677795, + 0.003835383802652359, + 0.0032974628265947104, + 0.0029236581176519394, + 0.004496010020375252, + 0.005084401462227106, + 0.004391447175294161, + 0.007019378710538149, + 0.004111991263926029, + 0.003956219647079706, + 0.003162675304338336, + 0.0030188127420842648, + 0.005258582532405853, + 0.005554696079343557, + 0.005652517080307007, + 0.00961159635335207, + 0.00453175837174058, + 0.005615819711238146, + 0.004220588132739067, + 0.006267734803259373, + 0.007364468649029732, + 0.021877191960811615, + 0.030275166034698486, + 0.010303886607289314, + 0.006517366971820593, + 0.0033301226794719696, + 0.007750227116048336, + 0.008613363839685917, + 0.0037273592315614223, + 0.004737379960715771, + 0.004021044354885817, + 0.004426649771630764, + 0.005651605315506458, + 0.004144842270761728, + 0.007048997096717358, + 0.005236043594777584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.00409455131739378 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a level angle, focuses on a young donkey, comfortably reclining on its side on a vast, soft, golden bale of hay. Warm afternoon sunlight filters through the fine dust in the air, tinting the entire frame with a golden hue and casting a bright halo around the edges of its fluffy, light-gray fur. Its long eyelashes flutter slightly in the soft light, and its tightly closed eyelids cautiously pry open a crack. Then, slowly, with a hint of sleepiness, they open fully, revealing a deep, moist black eye that reflects the golden sunlight. At this moment, the camera steadily and slowly pulls back, gradually revealing a wider, pastoral landscape: the bale of hay nestles in the center of a lush pasture. In the distance, a quaint wooden fence and a few scattered oak trees stand out. The azure sky is dotted with cotton-candy clouds, and the entire scene is permeated with a tranquil, lazy, and soothing atmosphere of the afternoon.", + "seed": 18279, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_7038_seed_18279_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01882281294092536, + "vlm_logprob": -0.01882281294092536, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01882281294092536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03161261975765228 + ], + "opposite_logprob_tokens": [ + -3.6566126346588135 + ], + "logprob_reward": -0.03161261975765228, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03161261975765228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2948123812675476, + 0.10907570272684097, + 0.08614803105592728, + 0.03936072066426277, + 0.04873620346188545, + 0.04981349781155586, + 0.07489429414272308, + 0.10544613748788834, + 0.04350028187036514, + 0.1482127159833908 + ], + "credit_map": [ + 0.056742992252111435, + 0.048475950956344604, + 0.04813332483172417, + 0.038328442722558975, + 0.02023501507937908, + 0.03520103543996811, + 0.019140608608722687, + 0.028555020689964294, + 0.015215580351650715, + 0.019941680133342743, + 0.016960211098194122, + 0.014236461371183395, + 0.012141418643295765, + 0.010979588143527508, + 0.008469346910715103, + 0.011131415143609047, + 0.01229716744273901, + 0.012946713715791702, + 0.013621850870549679, + 0.008250381797552109, + 0.009038629941642284, + 0.009045717306435108, + 0.008648250252008438, + 0.012299317866563797, + 0.00497036287561059, + 0.006158525589853525, + 0.006406925152987242, + 0.004497044254094362, + 0.004215519409626722, + 0.004521772265434265, + 0.0047266255132853985, + 0.003863943973556161, + 0.007985472679138184, + 0.00832909531891346, + 0.005513000302016735, + 0.004951667971909046, + 0.00360352941788733, + 0.009301947429776192, + 0.005597894079983234, + 0.0034535927698016167, + 0.007811506278812885, + 0.01180806290358305, + 0.00526454346254468, + 0.005210204049944878, + 0.005386100150644779, + 0.006171763874590397, + 0.0041680242866277695, + 0.003993290476500988, + 0.015723302960395813, + 0.009384257718920708, + 0.006034904159605503, + 0.006554583553224802, + 0.005513319279998541, + 0.01296949852257967, + 0.006688766181468964, + 0.012025664560496807, + 0.010306689888238907, + 0.019329600036144257, + 0.009043094702064991, + 0.01943744346499443, + 0.009681597352027893, + 0.012975159101188183, + 0.009383069351315498, + 0.0152894901111722, + 0.005980712361633778, + 0.00657557463273406, + 0.003792964154854417, + 0.0035652159713208675, + 0.00820021703839302, + 0.0062775202095508575, + 0.0037012684624642134, + 0.005406809505075216, + 0.023628734052181244, + 0.011344391852617264, + 0.01739637367427349, + 0.015160541981458664, + 0.008635065518319607, + 0.007346678525209427, + 0.048061124980449677, + 0.016639795154333115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.03161261975765228 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658131584525108 + ], + "opposite_logprob_tokens": [ + -4.758657932281494 + ], + "logprob_reward": -0.008658131584525108, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658131584525108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13332419097423553, + 0.10269927233457565, + 0.08178460597991943, + 0.19358617067337036, + 0.04559359326958656, + 0.03314489126205444, + 0.1014469638466835, + 0.11895176023244858, + 0.045242201536893845, + 0.1442263126373291 + ], + "credit_map": [ + 0.0154403792694211, + 0.015682879835367203, + 0.028007805347442627, + 0.022232668474316597, + 0.00821226928383112, + 0.014170560985803604, + 0.016346439719200134, + 0.013231189921498299, + 0.016924237832427025, + 0.011478058062493801, + 0.025073181837797165, + 0.011850070208311081, + 0.00864412635564804, + 0.011372181586921215, + 0.009094569832086563, + 0.008262843824923038, + 0.014673225581645966, + 0.010689733549952507, + 0.014316452667117119, + 0.007287937216460705, + 0.008716760203242302, + 0.008487947285175323, + 0.008604966104030609, + 0.009007584303617477, + 0.028803914785385132, + 0.021757297217845917, + 0.05300522595643997, + 0.017254551872611046, + 0.014747503213584423, + 0.021345213055610657, + 0.018339339643716812, + 0.018333133310079575, + 0.006566055119037628, + 0.0041322894394397736, + 0.0074408287182450294, + 0.0053582582622766495, + 0.002864938462153077, + 0.007335790898650885, + 0.005260570906102657, + 0.006634862162172794, + 0.00473340880125761, + 0.004898907151073217, + 0.0036291354335844517, + 0.0050882697105407715, + 0.004250509664416313, + 0.00359183712862432, + 0.0034208649303764105, + 0.003531956346705556, + 0.01984134502708912, + 0.01544740330427885, + 0.009654759429395199, + 0.010140585713088512, + 0.008993549272418022, + 0.016460780054330826, + 0.008197219111025333, + 0.012711318209767342, + 0.011954854242503643, + 0.02438153140246868, + 0.010221024043858051, + 0.021293235942721367, + 0.011421993374824524, + 0.01379643939435482, + 0.009992324747145176, + 0.01589035987854004, + 0.005924994125962257, + 0.00688955606892705, + 0.006173272617161274, + 0.004090290982276201, + 0.005584161728620529, + 0.0064004408195614815, + 0.003957697656005621, + 0.00622178940102458, + 0.021769914776086807, + 0.009826707653701305, + 0.01634102687239647, + 0.013959147967398167, + 0.00854229275137186, + 0.006670914590358734, + 0.05024096742272377, + 0.01687534712255001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.008658131584525108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812805563211441 + ], + "opposite_logprob_tokens": [ + -3.288128137588501 + ], + "logprob_reward": -0.03812805563211441, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812805563211441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18501754105091095, + 0.11196217685937881, + 0.05515296384692192, + 0.09676426649093628, + 0.10388970375061035, + 0.057993654161691666, + 0.09783952683210373, + 0.16065187752246857, + 0.06925571709871292, + 0.061472583562135696 + ], + "credit_map": [ + 0.02633184939622879, + 0.022690879181027412, + 0.03574099391698837, + 0.023995554074645042, + 0.016910923644900322, + 0.015847433358430862, + 0.019666770473122597, + 0.023833131417632103, + 0.0244531761854887, + 0.01405243668705225, + 0.01892659440636635, + 0.014142327941954136, + 0.009204963222146034, + 0.014019952155649662, + 0.007560295518487692, + 0.009602438658475876, + 0.010085837915539742, + 0.00633428618311882, + 0.007417353801429272, + 0.0050817313604056835, + 0.006858689244836569, + 0.006579706445336342, + 0.00587536720559001, + 0.00691999401897192, + 0.011748644523322582, + 0.014429563656449318, + 0.021239779889583588, + 0.009756846353411674, + 0.008778251707553864, + 0.012081320397555828, + 0.009134043008089066, + 0.009595816023647785, + 0.01454444881528616, + 0.01870281808078289, + 0.014660910703241825, + 0.013873887248337269, + 0.009359006769955158, + 0.012761600315570831, + 0.010135534219443798, + 0.009851502254605293, + 0.009051308035850525, + 0.01136991661041975, + 0.006947193760424852, + 0.007208626251667738, + 0.006276313681155443, + 0.007571007125079632, + 0.005534880794584751, + 0.00403441209346056, + 0.019868597388267517, + 0.013290968723595142, + 0.007533079478889704, + 0.009273785166442394, + 0.007805934641510248, + 0.014665987342596054, + 0.00873930286616087, + 0.016661865636706352, + 0.015556498430669308, + 0.03053588978946209, + 0.013893991708755493, + 0.030194560065865517, + 0.014752221293747425, + 0.01856580190360546, + 0.014408481307327747, + 0.0227444376796484, + 0.0077573927119374275, + 0.009189045988023281, + 0.009238344617187977, + 0.007884033024311066, + 0.008054609410464764, + 0.01120750978589058, + 0.006367297377437353, + 0.00955748651176691, + 0.0066217053681612015, + 0.006618767976760864, + 0.007795257959514856, + 0.005193955730646849, + 0.0040163262747228146, + 0.004548597615212202, + 0.017178276553750038, + 0.009499696083366871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03812805563211441 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107603870332241 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.01107603870332241, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107603870332241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10517182946205139, + 0.06059962883591652, + 0.03099740296602249, + 0.0698636844754219, + 0.07301175594329834, + 0.11620792746543884, + 0.1172899454832077, + 0.0878881961107254, + 0.05907739698886871, + 0.27989229559898376 + ], + "credit_map": [ + 0.015088397078216076, + 0.013048998080193996, + 0.020058756694197655, + 0.014741181395947933, + 0.00947677530348301, + 0.00973616261035204, + 0.013275554403662682, + 0.009745991788804531, + 0.013322139158844948, + 0.011178912594914436, + 0.0060164532624185085, + 0.005840504076331854, + 0.008282318711280823, + 0.007755258120596409, + 0.0040767258033156395, + 0.0041273110546171665, + 0.005694305058568716, + 0.005303552374243736, + 0.0034239389933645725, + 0.002504791598767042, + 0.005256672389805317, + 0.0039012786000967026, + 0.002460069488734007, + 0.0024527886416763067, + 0.008160104043781757, + 0.007994051091372967, + 0.012245161458849907, + 0.007617090828716755, + 0.00883009284734726, + 0.008703592233359814, + 0.009108883328735828, + 0.007204702589660883, + 0.012192577123641968, + 0.006831605918705463, + 0.008904668502509594, + 0.00819987989962101, + 0.007111514452844858, + 0.012450779788196087, + 0.007729183416813612, + 0.0095915412530303, + 0.0189489908516407, + 0.027872905135154724, + 0.010102768428623676, + 0.012756827287375927, + 0.011653689667582512, + 0.015366923995316029, + 0.008746781386435032, + 0.01075903419405222, + 0.023658376187086105, + 0.014902926981449127, + 0.009053032845258713, + 0.014916575513780117, + 0.011329779401421547, + 0.02009478770196438, + 0.00968724861741066, + 0.013647201471030712, + 0.010667432099580765, + 0.01894661970436573, + 0.008412021212279797, + 0.012596605345606804, + 0.008838760666549206, + 0.011449091136455536, + 0.0065165129490196705, + 0.01046114694327116, + 0.007461901754140854, + 0.007117202039808035, + 0.005882156081497669, + 0.004969361238181591, + 0.00828187633305788, + 0.007468160707503557, + 0.007064213044941425, + 0.010832519270479679, + 0.04232778027653694, + 0.014316700398921967, + 0.02974831685423851, + 0.0247996523976326, + 0.018753184005618095, + 0.02188812755048275, + 0.09489353001117706, + 0.03316497057676315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.01107603870332241 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639219027012587 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639219027012587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639219027012587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13055799901485443, + 0.11817550659179688, + 0.04449157416820526, + 0.08380778878927231, + 0.065403051674366, + 0.11536532640457153, + 0.13198809325695038, + 0.19101305305957794, + 0.06267586350440979, + 0.05652179196476936 + ], + "credit_map": [ + 0.01495963241904974, + 0.012767985463142395, + 0.017038118094205856, + 0.015728162601590157, + 0.011712413281202316, + 0.01289025042206049, + 0.020171036943793297, + 0.025290396064519882, + 0.03139913082122803, + 0.019198592752218246, + 0.01484981644898653, + 0.010842829942703247, + 0.009926047176122665, + 0.016361670568585396, + 0.006962310988456011, + 0.008635095320641994, + 0.005309624597430229, + 0.006756376475095749, + 0.00706947548314929, + 0.004204868339002132, + 0.005271746776998043, + 0.004827861674129963, + 0.005164329428225756, + 0.005887291859835386, + 0.012186693027615547, + 0.007952099665999413, + 0.013455689884722233, + 0.006990467198193073, + 0.012245943769812584, + 0.010439223609864712, + 0.006570871453732252, + 0.013966796919703484, + 0.009583565406501293, + 0.008001083508133888, + 0.009133676066994667, + 0.009447401389479637, + 0.0056657991372048855, + 0.010396834462881088, + 0.007172891870141029, + 0.006001797039061785, + 0.017534764483571053, + 0.01708371564745903, + 0.016836347058415413, + 0.010587840341031551, + 0.011975769884884357, + 0.019276447594165802, + 0.008142326958477497, + 0.013928097672760487, + 0.027787579223513603, + 0.0192344319075346, + 0.01115890871733427, + 0.010756046511232853, + 0.01155784446746111, + 0.02372175268828869, + 0.010469566099345684, + 0.01730196736752987, + 0.018232664093375206, + 0.03548577055335045, + 0.016082432121038437, + 0.03729579225182533, + 0.01710273139178753, + 0.021843403577804565, + 0.016607334837317467, + 0.02836291491985321, + 0.008394625969231129, + 0.009219832718372345, + 0.004665874410420656, + 0.005068376660346985, + 0.012556041590869427, + 0.010173842310905457, + 0.00606260122731328, + 0.00653466721996665, + 0.006979098543524742, + 0.005680331960320473, + 0.006387993693351746, + 0.005024496931582689, + 0.006134629249572754, + 0.009316046722233295, + 0.009492692537605762, + 0.007506500463932753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.004639219027012587 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a level angle, focuses on a young donkey, comfortably reclining on its side on a vast, soft, golden bale of hay. Warm afternoon sunlight filters through the fine dust in the air, tinting the entire frame with a golden hue and casting a bright halo around the edges of its fluffy, light-gray fur. Its long eyelashes flutter slightly in the soft light, and its tightly closed eyelids cautiously pry open a crack. Then, slowly, with a hint of sleepiness, they open fully, revealing a deep, moist black eye that reflects the golden sunlight. At this moment, the camera steadily and slowly pulls back, gradually revealing a wider, pastoral landscape: the bale of hay nestles in the center of a lush pasture. In the distance, a quaint wooden fence and a few scattered oak trees stand out. The azure sky is dotted with cotton-candy clouds, and the entire scene is permeated with a tranquil, lazy, and soothing atmosphere of the afternoon.", + "seed": 18280, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_1/index_7038_seed_18280_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013474986422806978, + "vlm_logprob": -0.013474986422806978, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.013474986422806978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02139698900282383 + ], + "opposite_logprob_tokens": [ + -4.021397113800049 + ], + "logprob_reward": -0.02139698900282383, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02139698900282383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2320283055305481, + 0.1396280825138092, + 0.12837500870227814, + 0.09012870490550995, + 0.1442251056432724, + 0.031210489571094513, + 0.1325363963842392, + 0.0327753983438015, + 0.023311298340559006, + 0.04578119516372681 + ], + "credit_map": [ + 0.01864997111260891, + 0.019858334213495255, + 0.03728416562080383, + 0.0257946215569973, + 0.02567959949374199, + 0.04368549957871437, + 0.034950438886880875, + 0.026125671342015266, + 0.009370105341076851, + 0.01753336377441883, + 0.02082068845629692, + 0.012001626193523407, + 0.022891350090503693, + 0.02653695084154606, + 0.017044339329004288, + 0.01342964917421341, + 0.01728852652013302, + 0.022843308746814728, + 0.0177466981112957, + 0.014062104746699333, + 0.01384308934211731, + 0.01205538958311081, + 0.018986638635396957, + 0.011549255810678005, + 0.007679103873670101, + 0.008188171312212944, + 0.014473920688033104, + 0.010259909555315971, + 0.01514789741486311, + 0.013270722702145576, + 0.012723161838948727, + 0.008385821245610714, + 0.02061593160033226, + 0.02785705216228962, + 0.01896166428923607, + 0.019550204277038574, + 0.013007139787077904, + 0.019199565052986145, + 0.014590173959732056, + 0.0104433698579669, + 0.003140031360089779, + 0.003978652413934469, + 0.0041951085440814495, + 0.0050468142144382, + 0.003532358445227146, + 0.00313408556394279, + 0.004565250128507614, + 0.0036181891337037086, + 0.010040289722383022, + 0.022645438089966774, + 0.010311480611562729, + 0.02381778135895729, + 0.019626479595899582, + 0.016939686611294746, + 0.018921788781881332, + 0.010233446024358273, + 0.0035285460762679577, + 0.00440731318667531, + 0.0033747118432074785, + 0.007231276482343674, + 0.0034746965393424034, + 0.002668539062142372, + 0.005212507676333189, + 0.0028778084088116884, + 0.0019611010793596506, + 0.003498989390209317, + 0.002242457587271929, + 0.0034305450972169638, + 0.003819720121100545, + 0.003098547924309969, + 0.003049379913136363, + 0.002210556762292981, + 0.0030850295443087816, + 0.00471486197784543, + 0.0039061284624040127, + 0.007294636219739914, + 0.009515820071101189, + 0.006693733390420675, + 0.006315363105386496, + 0.004255620297044516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02139698900282383 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792386554181576 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792386554181576, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792386554181576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22953291237354279, + 0.17565132677555084, + 0.1377570629119873, + 0.1175331324338913, + 0.07551966607570648, + 0.0379493348300457, + 0.1339915543794632, + 0.034607741981744766, + 0.019497910514473915, + 0.03795938566327095 + ], + "credit_map": [ + 0.01057868916541338, + 0.016843581572175026, + 0.030047765001654625, + 0.043224629014730453, + 0.02951095439493656, + 0.04236646369099617, + 0.027891110628843307, + 0.02906971424818039, + 0.008982918225228786, + 0.02072776108980179, + 0.044438436627388, + 0.01615207828581333, + 0.022401534020900726, + 0.026837771758437157, + 0.021172504872083664, + 0.014938311651349068, + 0.016411133110523224, + 0.019514253363013268, + 0.021643446758389473, + 0.017697850242257118, + 0.014203198254108429, + 0.013771876692771912, + 0.020959073677659035, + 0.013556207530200481, + 0.008258308283984661, + 0.01151417288929224, + 0.018456364050507545, + 0.012779767625033855, + 0.019034424796700478, + 0.01217041164636612, + 0.02059789188206196, + 0.014721768908202648, + 0.009033925831317902, + 0.00975863914936781, + 0.010838055051863194, + 0.01388311292976141, + 0.009340818971395493, + 0.00848302710801363, + 0.007852757349610329, + 0.00632932223379612, + 0.0025390186347067356, + 0.004460230004042387, + 0.004482900723814964, + 0.0064262025989592075, + 0.005237669218331575, + 0.003960977308452129, + 0.00627930648624897, + 0.004563024267554283, + 0.010218067094683647, + 0.021468129009008408, + 0.009926212020218372, + 0.025031229481101036, + 0.02124566212296486, + 0.015753403306007385, + 0.020198648795485497, + 0.010150191374123096, + 0.0035263390745967627, + 0.005775650963187218, + 0.002852972596883774, + 0.00601141108199954, + 0.004416730720549822, + 0.003261334029957652, + 0.005419835448265076, + 0.0033434678334742785, + 0.001581554184667766, + 0.0027003951836377382, + 0.002109057968482375, + 0.00231380108743906, + 0.0033508457709103823, + 0.0028763888403773308, + 0.001838011434301734, + 0.0027278531342744827, + 0.0026076745707541704, + 0.004787827376276255, + 0.003276323201134801, + 0.0068483734503388405, + 0.005932938773185015, + 0.005013440269976854, + 0.004328025970607996, + 0.005164780654013157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009792386554181576 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305892944335938 + ], + "opposite_logprob_tokens": [ + -3.773305892944336 + ], + "logprob_reward": -0.023305892944335938, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305892944335938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1900690197944641, + 0.2641489505767822, + 0.11664438992738724, + 0.05699197202920914, + 0.0994330570101738, + 0.07961888611316681, + 0.10295925289392471, + 0.03488351032137871, + 0.024212587624788284, + 0.031038422137498856 + ], + "credit_map": [ + 0.012572987005114555, + 0.013245202600955963, + 0.028722846880555153, + 0.0291377454996109, + 0.02167530357837677, + 0.03549684211611748, + 0.027398018166422844, + 0.021820062771439552, + 0.016146071255207062, + 0.026173392310738564, + 0.06336397677659988, + 0.028775831684470177, + 0.030790913850069046, + 0.05332063511013985, + 0.023598384112119675, + 0.021979717537760735, + 0.015142678283154964, + 0.018367236480116844, + 0.01753358729183674, + 0.014390924014151096, + 0.012460080906748772, + 0.011220348067581654, + 0.01705518551170826, + 0.01047433726489544, + 0.005405906587839127, + 0.006588164251297712, + 0.008080312050879002, + 0.005378649570047855, + 0.012366846203804016, + 0.006575521547347307, + 0.008188960142433643, + 0.004407606087625027, + 0.009510193951427937, + 0.012765156105160713, + 0.01706777513027191, + 0.015186582691967487, + 0.009863137267529964, + 0.010592343285679817, + 0.015054970048367977, + 0.009392892010509968, + 0.007635797839611769, + 0.011469721794128418, + 0.00991844478994608, + 0.012372109107673168, + 0.009982720948755741, + 0.008276738226413727, + 0.011993655934929848, + 0.007969694212079048, + 0.00916976947337389, + 0.018870387226343155, + 0.007579802069813013, + 0.01806831732392311, + 0.01472530234605074, + 0.01287409383803606, + 0.014080189168453217, + 0.007591383997350931, + 0.003559127217158675, + 0.005544682033360004, + 0.00416432972997427, + 0.004786282777786255, + 0.004805251955986023, + 0.003152532037347555, + 0.004427270498126745, + 0.004444029647856951, + 0.0021542648319154978, + 0.002845901995897293, + 0.0023860943038016558, + 0.0045829713344573975, + 0.002961815567687154, + 0.0030137200374156237, + 0.00456633698195219, + 0.0017014782642945647, + 0.002256627893075347, + 0.00446075526997447, + 0.0028797152917832136, + 0.004618950188159943, + 0.005195731297135353, + 0.005088619887828827, + 0.004133788403123617, + 0.002404232043772936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.023305892944335938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627647835761309 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627647835761309, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627647835761309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2079203575849533, + 0.10732400417327881, + 0.05917828902602196, + 0.07027073949575424, + 0.16807159781455994, + 0.08221732825040817, + 0.10809092223644257, + 0.031065162271261215, + 0.06410390138626099, + 0.10175769031047821 + ], + "credit_map": [ + 0.01952301897108555, + 0.01903247833251953, + 0.04201267659664154, + 0.02608686313033104, + 0.024226075038313866, + 0.029327504336833954, + 0.026135273277759552, + 0.021576479077339172, + 0.005569534841924906, + 0.009114107117056847, + 0.013726883567869663, + 0.016228897497057915, + 0.01165688969194889, + 0.015726039186120033, + 0.019508104771375656, + 0.015793543308973312, + 0.008569950237870216, + 0.009574320167303085, + 0.0052728489972651005, + 0.00889767985790968, + 0.006695153657346964, + 0.006636325735598803, + 0.0068333945237100124, + 0.006698617711663246, + 0.008421842940151691, + 0.008321508765220642, + 0.01455994788557291, + 0.00704285828396678, + 0.0079147694632411, + 0.008891248144209385, + 0.00797644630074501, + 0.007142119109630585, + 0.014422675594687462, + 0.018449103459715843, + 0.026643848046660423, + 0.029570233076810837, + 0.018222929909825325, + 0.017018476501107216, + 0.024011455476284027, + 0.019732877612113953, + 0.007097945548593998, + 0.01162087544798851, + 0.013850812800228596, + 0.014592139981687069, + 0.007988018915057182, + 0.008270056918263435, + 0.011839614249765873, + 0.006957858335226774, + 0.00929991528391838, + 0.020261751487851143, + 0.007440401241183281, + 0.01961541548371315, + 0.015657993033528328, + 0.013219698332250118, + 0.014036191627383232, + 0.00855955109000206, + 0.0026520751416683197, + 0.004871653392910957, + 0.004186120815575123, + 0.0037818264681845903, + 0.003891725093126297, + 0.003277671756222844, + 0.0037297517992556095, + 0.004674337804317474, + 0.0054108514450490475, + 0.008908859454095364, + 0.005596071016043425, + 0.010140144266188145, + 0.007939095608890057, + 0.00644659111276269, + 0.010374381206929684, + 0.009287904016673565, + 0.006288041826337576, + 0.008341215550899506, + 0.004715308081358671, + 0.015891635790467262, + 0.016184872016310692, + 0.010780326090753078, + 0.023040838539600372, + 0.016515452414751053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.007627647835761309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young donkey with fluffy, light-gray fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the donkey initially shown comfortably reclining on its side on a golden bale of hay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the donkey's eye open slowly from a crack to reveal a deep, moist black eye reflecting sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera steadily pull back to reveal a wider pastoral landscape with a wooden fence and oak trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm afternoon sunlight with a golden hue and visible dust particles in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0636562705039978, + 0.053857479244470596, + 0.06560621410608292, + 0.03396470099687576, + 0.18908444046974182, + 0.07038597017526627, + 0.41206666827201843, + 0.016505606472492218, + 0.038887202739715576, + 0.05598544329404831 + ], + "credit_map": [ + 0.005868855863809586, + 0.006778703071177006, + 0.007324986159801483, + 0.006890084594488144, + 0.010325131937861443, + 0.010288494639098644, + 0.008992408402264118, + 0.0071876076981425285, + 0.00437372550368309, + 0.0037867522332817316, + 0.008657801896333694, + 0.006078334990888834, + 0.007910606451332569, + 0.011043756268918514, + 0.005733771715313196, + 0.006272732280194759, + 0.011348169296979904, + 0.011208020150661469, + 0.008794188499450684, + 0.00743343960493803, + 0.007115100510418415, + 0.005827505141496658, + 0.008244037628173828, + 0.00563575467094779, + 0.005223372019827366, + 0.003920406103134155, + 0.004905019886791706, + 0.0030383712146431208, + 0.005237626377493143, + 0.0038294459227472544, + 0.0031999831553548574, + 0.004610476084053516, + 0.02125038206577301, + 0.029887720942497253, + 0.02370636723935604, + 0.03041440062224865, + 0.018695605918765068, + 0.02155247889459133, + 0.026735924184322357, + 0.016841545701026917, + 0.006739332340657711, + 0.01259650569409132, + 0.007883662357926369, + 0.01051507331430912, + 0.008536603301763535, + 0.008075369521975517, + 0.009370679035782814, + 0.006668741814792156, + 0.0337580144405365, + 0.06948014348745346, + 0.03096049465239048, + 0.07270370423793793, + 0.0632735937833786, + 0.05136643350124359, + 0.059951961040496826, + 0.030572332441806793, + 0.001980804605409503, + 0.002515241736546159, + 0.0020120886620134115, + 0.002519376575946808, + 0.001547835418023169, + 0.0016471792478114367, + 0.002471294952556491, + 0.0018117857398465276, + 0.003241871716454625, + 0.006216845009475946, + 0.004107923246920109, + 0.004254464525729418, + 0.006412912160158157, + 0.0045578558929264545, + 0.005282409023493528, + 0.004812923725694418, + 0.005945675540715456, + 0.010422307066619396, + 0.0035566971637308598, + 0.00759913632646203, + 0.006626144517213106, + 0.009965683333575726, + 0.006970137357711792, + 0.004899661988019943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.0052520157769322395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_tensor([3194, 3194])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_tensor([3194, 3194])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ea90f9db6b642610979218c1b519d11c76a870ba --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_tensor([3194, 3194])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An unstable, handheld, shaky camera closely follows a relentlessly moving stainless steel conveyor belt, its cold surface coated in a thin layer of condensation as it glides forward in a steady, mechanical rhythm. Under the harsh, cold white light cast by rows of overhead industrial LED tubes, precisely cut, perfectly cubed chunks of raw, bright red meat are neatly arranged on the conveyor belt. Their wet surfaces reflect sharp highlights, and the distinct white fat veins contrast starkly with the deep red muscle tissue. The camera follows the movement of the meat, gently panning up and down and left and right, capturing this world of cold metal and lingering steam. In the background, a complex web of mechanical pipes crisscrosses, with thick white steam occasionally emanating from the seams, obscuring the distant landscape and enhancing the scene's depth and mystery. The entire picture presents a surreal Cinema 4D style, with highly rendered smooth textures and clear geometric shapes. The reflections of metal, the haziness of steam and the vivid colors of meat together construct a cold, alienated yet visually impactful inhuman industrial spectacle.", + "seed": 8828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_3194_seed_8828_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.272841127961874, + "vlm_logprob": -0.272841127961874, + "vlm_yes_no_margin": 1.875, + "vlm_token_credit": -0.272841127961874, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034277915954589844 + ], + "opposite_logprob_tokens": [ + -4.03427791595459 + ], + "logprob_reward": -0.034277915954589844, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.034277915954589844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14646132290363312, + 0.08663896471261978, + 0.059754155576229095, + 0.14164897799491882, + 0.0693267285823822, + 0.07107656449079514, + 0.054817479103803635, + 0.08532728999853134, + 0.04405343905091286, + 0.2408950924873352 + ], + "credit_map": [ + 0.01615549810230732, + 0.016728367656469345, + 0.025378922000527382, + 0.026826152577996254, + 0.014131637290120125, + 0.011157981120049953, + 0.02080279029905796, + 0.01527996826916933, + 0.007941207848489285, + 0.0070155709981918335, + 0.024487629532814026, + 0.012250247411429882, + 0.007951874285936356, + 0.005371931474655867, + 0.008558709174394608, + 0.013061794452369213, + 0.013524921610951424, + 0.006440578959882259, + 0.008577793836593628, + 0.006874806713312864, + 0.004361817613244057, + 0.003278580028563738, + 0.008368109352886677, + 0.008327549323439598, + 0.037910256534814835, + 0.008840907365083694, + 0.054050348699092865, + 0.010973872616887093, + 0.008965445682406425, + 0.006253822706639767, + 0.008908822201192379, + 0.0057455021888017654, + 0.008983750827610493, + 0.0112002519890666, + 0.011522208340466022, + 0.014238415285944939, + 0.008451191708445549, + 0.0036749807186424732, + 0.007949309423565865, + 0.0033066188916563988, + 0.0078818928450346, + 0.0053957318887114525, + 0.01079543773084879, + 0.012075865641236305, + 0.006767477374523878, + 0.009492686949670315, + 0.010716011747717857, + 0.007951461710035801, + 0.005332118831574917, + 0.008973275311291218, + 0.007892260327935219, + 0.008756753988564014, + 0.008539177477359772, + 0.004717343952506781, + 0.0063844420947134495, + 0.004222104791551828, + 0.009917517192661762, + 0.014998460188508034, + 0.010698096826672554, + 0.008320868015289307, + 0.009596411138772964, + 0.009126984514296055, + 0.012474849820137024, + 0.010194101370871067, + 0.005718606058508158, + 0.007655980996787548, + 0.004285383969545364, + 0.004814606625586748, + 0.007753534242510796, + 0.004549914970993996, + 0.004679701291024685, + 0.0045957076363265514, + 0.026166167110204697, + 0.03360884264111519, + 0.033637646585702896, + 0.030866283923387527, + 0.03403822332620621, + 0.031092612072825432, + 0.0350082665681839, + 0.016477053984999657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.034277915954589844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206228956580162 + ], + "opposite_logprob_tokens": [ + -2.8120622634887695 + ], + "logprob_reward": -0.06206228956580162, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206228956580162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16321058571338654, + 0.04220730438828468, + 0.059462014585733414, + 0.14710941910743713, + 0.19084125757217407, + 0.08219937235116959, + 0.06277425587177277, + 0.047640204429626465, + 0.06288745254278183, + 0.1416681408882141 + ], + "credit_map": [ + 0.013935886323451996, + 0.01588224247097969, + 0.02290186658501625, + 0.030444469302892685, + 0.022094473242759705, + 0.011281831189990044, + 0.0284444447606802, + 0.018225356936454773, + 0.0043675764463841915, + 0.004850088618695736, + 0.006267345044761896, + 0.0078008039854466915, + 0.004482466261833906, + 0.0036272252909839153, + 0.0038204723969101906, + 0.006991324946284294, + 0.010936813428997993, + 0.005993531551212072, + 0.006691219285130501, + 0.006887695286422968, + 0.005075243767350912, + 0.00400905217975378, + 0.012634463608264923, + 0.007233994547277689, + 0.028773333877325058, + 0.01950276829302311, + 0.04379153251647949, + 0.010661521926522255, + 0.018051058053970337, + 0.008627399802207947, + 0.011247738264501095, + 0.006454085931181908, + 0.03774121776223183, + 0.03774907439947128, + 0.025344157591462135, + 0.027710609138011932, + 0.020681820809841156, + 0.013428919948637486, + 0.018599050119519234, + 0.009586417116224766, + 0.009997679851949215, + 0.00927860289812088, + 0.010115191340446472, + 0.01119166798889637, + 0.012145360931754112, + 0.00799600500613451, + 0.015071784146130085, + 0.006403083447366953, + 0.005722250789403915, + 0.0100104296579957, + 0.007036749739199877, + 0.012009925208985806, + 0.010310954414308071, + 0.0052246167324483395, + 0.008102008141577244, + 0.004357322584837675, + 0.005233239382505417, + 0.009794934652745724, + 0.005806432571262121, + 0.004868694581091404, + 0.007180329877883196, + 0.004628489259630442, + 0.005688873119652271, + 0.004439210519194603, + 0.009649637155234814, + 0.011790402233600616, + 0.008177368901669979, + 0.005379347130656242, + 0.009967714548110962, + 0.0058279382064938545, + 0.005129475612193346, + 0.006965567823499441, + 0.019219281151890755, + 0.021665453910827637, + 0.02491878531873226, + 0.018227607011795044, + 0.01798739843070507, + 0.012948363088071346, + 0.018091002479195595, + 0.008610248565673828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.06206228956580162 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261364102363586 + ], + "opposite_logprob_tokens": [ + -0.5761364102363586 + ], + "logprob_reward": -0.8261364102363586, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261364102363586, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09613631665706635, + 0.06750263273715973, + 0.12822379171848297, + 0.2012220323085785, + 0.10979790985584259, + 0.05511169880628586, + 0.053504474461078644, + 0.061775557696819305, + 0.11692702025175095, + 0.10979864746332169 + ], + "credit_map": [ + 0.01125217042863369, + 0.012144467793405056, + 0.014696670696139336, + 0.014177498407661915, + 0.01242661289870739, + 0.006745621562004089, + 0.01292343158274889, + 0.011769835837185383, + 0.004818834364414215, + 0.005163143388926983, + 0.021725958213210106, + 0.009017324075102806, + 0.005520672537386417, + 0.004978852346539497, + 0.004886539652943611, + 0.011391306295990944, + 0.021157557144761086, + 0.014521128498017788, + 0.030519599094986916, + 0.018772806972265244, + 0.014262956567108631, + 0.010447192005813122, + 0.010676673613488674, + 0.007865863852202892, + 0.051333483308553696, + 0.012721172533929348, + 0.07586213201284409, + 0.015263495966792107, + 0.01544084120541811, + 0.010471928864717484, + 0.012629122473299503, + 0.007499820552766323, + 0.017730586230754852, + 0.016233248636126518, + 0.013202643021941185, + 0.020923888310790062, + 0.01373457070440054, + 0.008336878381669521, + 0.013769908808171749, + 0.0058661773800849915, + 0.007225297391414642, + 0.0059988973662257195, + 0.005291822832077742, + 0.00874033197760582, + 0.006159671116620302, + 0.0047372193075716496, + 0.00891163107007742, + 0.008046824485063553, + 0.005202786531299353, + 0.007937907241284847, + 0.006246213335543871, + 0.008193671703338623, + 0.011022903956472874, + 0.005401355680078268, + 0.00616358732804656, + 0.003336044028401375, + 0.00729331374168396, + 0.01075141504406929, + 0.00784269068390131, + 0.008147042244672775, + 0.0061876666732132435, + 0.005897809751331806, + 0.008707997389137745, + 0.0069476221688091755, + 0.01639167033135891, + 0.017454301938414574, + 0.022265885025262833, + 0.009934715926647186, + 0.014568700455129147, + 0.010860089212656021, + 0.01376035250723362, + 0.011691294610500336, + 0.013330458663403988, + 0.014213189482688904, + 0.017703896388411522, + 0.015405583195388317, + 0.01154786255210638, + 0.011529042385518551, + 0.015673518180847168, + 0.01039508730173111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.8261364102363586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035763919353485 + ], + "opposite_logprob_tokens": [ + -1.9103575944900513 + ], + "logprob_reward": -0.16035763919353485, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035763919353485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06746166944503784, + 0.027672644704580307, + 0.031006023287773132, + 0.20144307613372803, + 0.49138471484184265, + 0.031559526920318604, + 0.026882817968726158, + 0.01716548763215542, + 0.04003462940454483, + 0.06538936495780945 + ], + "credit_map": [ + 0.009556833654642105, + 0.006379878148436546, + 0.010749880224466324, + 0.011127025820314884, + 0.006622273474931717, + 0.0045440467074513435, + 0.008470929227769375, + 0.010010804049670696, + 0.0026459142100065947, + 0.002600952982902527, + 0.004983060993254185, + 0.004057067446410656, + 0.003729818155989051, + 0.0033190122339874506, + 0.0030714706517755985, + 0.0032653503585606813, + 0.006081998348236084, + 0.002873368561267853, + 0.0065396446734666824, + 0.003777323989197612, + 0.002539196750149131, + 0.002469379222020507, + 0.004029070492833853, + 0.002696043811738491, + 0.055430300533771515, + 0.01432635635137558, + 0.07427532225847244, + 0.015735801309347153, + 0.01244345586746931, + 0.008705925196409225, + 0.013097374700009823, + 0.0074285538867115974, + 0.10501226782798767, + 0.11286009103059769, + 0.07674489915370941, + 0.04422423616051674, + 0.05288873612880707, + 0.029483044520020485, + 0.044038139283657074, + 0.0261333379894495, + 0.003303617937490344, + 0.002495614578947425, + 0.0047277528792619705, + 0.005197769496589899, + 0.003372528590261936, + 0.004269717261195183, + 0.005356914363801479, + 0.002835615538060665, + 0.002078848425298929, + 0.0022087323013693094, + 0.003691714257001877, + 0.006537484470754862, + 0.0026891829911619425, + 0.0027833746280521154, + 0.0050847334787249565, + 0.001808749744668603, + 0.0027942636515945196, + 0.003250398673117161, + 0.0017061159014701843, + 0.001626283978112042, + 0.0017361140344291925, + 0.0021160771138966084, + 0.00192220916505903, + 0.002014026278629899, + 0.006390582770109177, + 0.00816251803189516, + 0.004394959192723036, + 0.002863676752895117, + 0.006840340793132782, + 0.0033114359248429537, + 0.003544815583154559, + 0.004526300355792046, + 0.007804499939084053, + 0.009357897564768791, + 0.013433890417218208, + 0.008597129955887794, + 0.008003312163054943, + 0.007258553523570299, + 0.0065106721594929695, + 0.004423408303409815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.16035763919353485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813713848590851 + ], + "opposite_logprob_tokens": [ + -1.4063713550567627 + ], + "logprob_reward": -0.2813713848590851, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813713848590851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18020310997962952, + 0.06303049623966217, + 0.05043167993426323, + 0.09714008867740631, + 0.10851085931062698, + 0.08940364420413971, + 0.07344693690538406, + 0.038407277315855026, + 0.08022706210613251, + 0.21919874846935272 + ], + "credit_map": [ + 0.02158351242542267, + 0.01969064027070999, + 0.033165354281663895, + 0.03232695162296295, + 0.013052250258624554, + 0.011607569642364979, + 0.027910815551877022, + 0.020866038277745247, + 0.005581666249781847, + 0.005543456878513098, + 0.013917728327214718, + 0.011343426071107388, + 0.005073051434010267, + 0.004063868895173073, + 0.007880049757659435, + 0.009627250954508781, + 0.01161875855177641, + 0.005074394401162863, + 0.008502969518303871, + 0.0051581840962171555, + 0.0031937381718307734, + 0.0028548266272991896, + 0.007913419045507908, + 0.006115397438406944, + 0.02253371849656105, + 0.008963791653513908, + 0.03278732672333717, + 0.008339177817106247, + 0.008645579218864441, + 0.005804363638162613, + 0.006467780563980341, + 0.003598368028178811, + 0.014840414747595787, + 0.024608587846159935, + 0.019349919632077217, + 0.015909666195511818, + 0.010493636131286621, + 0.006343018729239702, + 0.012699916958808899, + 0.004265711177140474, + 0.009591739624738693, + 0.010056223720312119, + 0.010362304747104645, + 0.013002428226172924, + 0.011088293977081776, + 0.010636847466230392, + 0.017764167860150337, + 0.006901656277477741, + 0.006102375220507383, + 0.010579189285635948, + 0.010232817381620407, + 0.012438242323696613, + 0.015083190985023975, + 0.006647452712059021, + 0.00851687416434288, + 0.003846805077046156, + 0.00572598073631525, + 0.005069938022643328, + 0.004439669195562601, + 0.004050421994179487, + 0.003929388243705034, + 0.004008361138403416, + 0.006631595082581043, + 0.004551925230771303, + 0.01013582106679678, + 0.012397784739732742, + 0.009038668125867844, + 0.007410939782857895, + 0.012377523817121983, + 0.008503418415784836, + 0.008199965581297874, + 0.012162950821220875, + 0.029103243723511696, + 0.03325223922729492, + 0.044213514775037766, + 0.02775004133582115, + 0.028937479481101036, + 0.016913406550884247, + 0.024751275777816772, + 0.014277568086981773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.2813713848590851 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unstable, handheld, shaky camera closely follows a relentlessly moving stainless steel conveyor belt, its cold surface coated in a thin layer of condensation as it glides forward in a steady, mechanical rhythm. Under the harsh, cold white light cast by rows of overhead industrial LED tubes, precisely cut, perfectly cubed chunks of raw, bright red meat are neatly arranged on the conveyor belt. Their wet surfaces reflect sharp highlights, and the distinct white fat veins contrast starkly with the deep red muscle tissue. The camera follows the movement of the meat, gently panning up and down and left and right, capturing this world of cold metal and lingering steam. In the background, a complex web of mechanical pipes crisscrosses, with thick white steam occasionally emanating from the seams, obscuring the distant landscape and enhancing the scene's depth and mystery. The entire picture presents a surreal Cinema 4D style, with highly rendered smooth textures and clear geometric shapes. The reflections of metal, the haziness of steam and the vivid colors of meat together construct a cold, alienated yet visually impactful inhuman industrial spectacle.", + "seed": 8829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_3194_seed_8829_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12276985570788383, + "vlm_logprob": -0.12276985570788383, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.12276985570788383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04134298488497734 + ], + "opposite_logprob_tokens": [ + -3.4163429737091064 + ], + "logprob_reward": -0.04134298488497734, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04134298488497734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08879055827856064, + 0.05585440993309021, + 0.11145222932100296, + 0.08036567270755768, + 0.1351316273212433, + 0.03743629902601242, + 0.058171290904283524, + 0.09779819846153259, + 0.2738092541694641, + 0.061190444976091385 + ], + "credit_map": [ + 0.012957636266946793, + 0.011190318502485752, + 0.013088909909129143, + 0.014158430509269238, + 0.007142007350921631, + 0.007166638039052486, + 0.010405981913208961, + 0.012680633924901485, + 0.005796329118311405, + 0.0054540918208658695, + 0.006071135867387056, + 0.008036703802645206, + 0.008212069980800152, + 0.00675919046625495, + 0.009001486003398895, + 0.006523405201733112, + 0.009783582761883736, + 0.01091833971440792, + 0.009472619742155075, + 0.011429146863520145, + 0.021023912355303764, + 0.015750546008348465, + 0.015225089155137539, + 0.017848985269665718, + 0.007704500574618578, + 0.009135629050433636, + 0.005702746566385031, + 0.006666873581707478, + 0.017635652795433998, + 0.016650043427944183, + 0.006652774754911661, + 0.01021745428442955, + 0.024103771895170212, + 0.01752549409866333, + 0.020912982523441315, + 0.014101426117122173, + 0.011979175731539726, + 0.011970195919275284, + 0.01721341349184513, + 0.017325157299637794, + 0.006969948764890432, + 0.007449816912412643, + 0.0026495177298784256, + 0.0029806585516780615, + 0.0038884917739778757, + 0.004260259680449963, + 0.004348096437752247, + 0.00488950964063406, + 0.0074470373801887035, + 0.007927422411739826, + 0.003612842410802841, + 0.004963343963027, + 0.008377325721085072, + 0.01111592911183834, + 0.00885864533483982, + 0.005868741311132908, + 0.01008676178753376, + 0.009881969541311264, + 0.004524779971688986, + 0.00789970438927412, + 0.014104271307587624, + 0.014371853321790695, + 0.01427037175744772, + 0.02265848219394684, + 0.02778645232319832, + 0.03345511853694916, + 0.01628846302628517, + 0.0326109305024147, + 0.05563564598560333, + 0.03167438134551048, + 0.05214625224471092, + 0.024212002754211426, + 0.008592980913817883, + 0.007509809453040361, + 0.008128160610795021, + 0.006395116914063692, + 0.0065601421520113945, + 0.009854458272457123, + 0.007032614666968584, + 0.007117156405001879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.04134298488497734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816951602697372 + ], + "opposite_logprob_tokens": [ + -3.5298168659210205 + ], + "logprob_reward": -0.029816951602697372, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816951602697372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19366171956062317, + 0.08185900747776031, + 0.0662316381931305, + 0.05616484954953194, + 0.05424310639500618, + 0.05418390408158302, + 0.042295388877391815, + 0.2874409258365631, + 0.09715112298727036, + 0.06676831841468811 + ], + "credit_map": [ + 0.022961769253015518, + 0.030664073303341866, + 0.015246325172483921, + 0.03868347778916359, + 0.017897119745612144, + 0.019590726122260094, + 0.02571174129843712, + 0.022906476631760597, + 0.011514104902744293, + 0.010403420776128769, + 0.009697305969893932, + 0.010803471319377422, + 0.011284210719168186, + 0.011907467618584633, + 0.00738540617749095, + 0.00886362325400114, + 0.005972661077976227, + 0.006861452478915453, + 0.008430367335677147, + 0.007945074699819088, + 0.011417926289141178, + 0.00797947309911251, + 0.008838620968163013, + 0.00878605991601944, + 0.005744320340454578, + 0.0073350565508008, + 0.0041539957746863365, + 0.0042523471638560295, + 0.012406555935740471, + 0.011259307153522968, + 0.0035733389668166637, + 0.007439928129315376, + 0.011029069311916828, + 0.007558876182883978, + 0.008123408071696758, + 0.0063604917377233505, + 0.002501091454178095, + 0.0045029036700725555, + 0.007157865911722183, + 0.007009400054812431, + 0.0062066251412034035, + 0.005548607092350721, + 0.003915751352906227, + 0.005794926546514034, + 0.006132840644568205, + 0.009976114146411419, + 0.00675527099519968, + 0.009853766299784184, + 0.0049728816375136375, + 0.006085716187953949, + 0.0029492811299860477, + 0.005083612632006407, + 0.007229974493384361, + 0.00593390641734004, + 0.005989251658320427, + 0.004050767049193382, + 0.031384728848934174, + 0.030917465686798096, + 0.018897583708167076, + 0.05762897804379463, + 0.029394257813692093, + 0.021588711068034172, + 0.03803088143467903, + 0.05959830805659294, + 0.010999104008078575, + 0.013014801777899265, + 0.005532925482839346, + 0.009332146495580673, + 0.01833971031010151, + 0.010507535189390182, + 0.020446570590138435, + 0.00897833239287138, + 0.008301334455609322, + 0.008020400069653988, + 0.008699996396899223, + 0.006268845405429602, + 0.010336445644497871, + 0.009193322621285915, + 0.008177166804671288, + 0.007770805153995752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.029816951602697372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486223518848419 + ], + "opposite_logprob_tokens": [ + -1.2236223220825195 + ], + "logprob_reward": -0.3486223518848419, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486223518848419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15544287860393524, + 0.08920733630657196, + 0.0705162063241005, + 0.30329278111457825, + 0.0826406180858612, + 0.05400565639138222, + 0.07062231749296188, + 0.0952661857008934, + 0.04373529180884361, + 0.03527076542377472 + ], + "credit_map": [ + 0.018385525792837143, + 0.01424473337829113, + 0.010468919761478901, + 0.028789019212126732, + 0.01741054095327854, + 0.015035584568977356, + 0.020194509997963905, + 0.030914053320884705, + 0.01374923624098301, + 0.01348157599568367, + 0.007389232981950045, + 0.008455019444227219, + 0.01344828587025404, + 0.015951288864016533, + 0.00878762174397707, + 0.00794508121907711, + 0.006977448705583811, + 0.006690672133117914, + 0.006941618397831917, + 0.006964976899325848, + 0.011910978704690933, + 0.00882619246840477, + 0.00962875410914421, + 0.01257556863129139, + 0.02986016310751438, + 0.03283131122589111, + 0.0215607937425375, + 0.028581909835338593, + 0.06276378035545349, + 0.05674492195248604, + 0.028873635455965996, + 0.04207627847790718, + 0.01879122667014599, + 0.014299452304840088, + 0.009071903303265572, + 0.006650369148701429, + 0.007005344610661268, + 0.013024802319705486, + 0.007502465043216944, + 0.006295063998550177, + 0.008652659133076668, + 0.008117842487990856, + 0.006451524328440428, + 0.011528920382261276, + 0.003962973598390818, + 0.004321725573390722, + 0.0060207424685359, + 0.004949270281940699, + 0.009366513229906559, + 0.012918860651552677, + 0.005390341393649578, + 0.008617066778242588, + 0.007736563682556152, + 0.009706609882414341, + 0.009374488145112991, + 0.007511880248785019, + 0.011294259689748287, + 0.01371738687157631, + 0.004755441099405289, + 0.010283275507390499, + 0.013015644624829292, + 0.012070967815816402, + 0.013205387629568577, + 0.016923826187849045, + 0.007118953391909599, + 0.007055940572172403, + 0.002575257793068886, + 0.003959345165640116, + 0.006476199254393578, + 0.005720395129173994, + 0.006027508061379194, + 0.004801700357347727, + 0.003941294737160206, + 0.003046097932383418, + 0.0062392582185566425, + 0.002401003148406744, + 0.005077838897705078, + 0.008607367053627968, + 0.0030569997616112232, + 0.002900907304137945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.3486223518848419 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035713255405426 + ], + "opposite_logprob_tokens": [ + -1.910357117652893 + ], + "logprob_reward": -0.16035713255405426, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035713255405426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11292675882577896, + 0.1526620090007782, + 0.10128385573625565, + 0.0967518538236618, + 0.09867268055677414, + 0.05095827206969261, + 0.10902394354343414, + 0.04415207728743553, + 0.04638702794909477, + 0.18718159198760986 + ], + "credit_map": [ + 0.011986133642494678, + 0.019014371559023857, + 0.007411078084260225, + 0.01391712948679924, + 0.010082554072141647, + 0.013781437650322914, + 0.017043359577655792, + 0.01969068869948387, + 0.01827925071120262, + 0.01844107359647751, + 0.015420077368617058, + 0.021807920187711716, + 0.01887446828186512, + 0.021128041669726372, + 0.01926334574818611, + 0.019447816535830498, + 0.010751104913651943, + 0.009283577091991901, + 0.010218861512839794, + 0.008886652998626232, + 0.01754751428961754, + 0.013948744162917137, + 0.013802697882056236, + 0.016844695433974266, + 0.009865579195320606, + 0.012935156933963299, + 0.006740323733538389, + 0.011387072503566742, + 0.01708305813372135, + 0.015433688648045063, + 0.010562780313193798, + 0.012744189240038395, + 0.01371262688189745, + 0.011131489649415016, + 0.01688378117978573, + 0.011156588792800903, + 0.00986493844538927, + 0.008966803550720215, + 0.013018066063523293, + 0.013938374817371368, + 0.004860226064920425, + 0.0054200636222958565, + 0.0040106140077114105, + 0.00629542488604784, + 0.006272885482758284, + 0.008292836137115955, + 0.007513851393014193, + 0.00829236675053835, + 0.01327885314822197, + 0.015306344255805016, + 0.009319203905761242, + 0.022335344925522804, + 0.015092913992702961, + 0.01075681485235691, + 0.01444605179131031, + 0.008488401770591736, + 0.006255608517676592, + 0.004634039010852575, + 0.002941354177892208, + 0.007364010438323021, + 0.0057222479954361916, + 0.005283223930746317, + 0.006047065835446119, + 0.005904519930481911, + 0.004557925276458263, + 0.005584770813584328, + 0.0029183814767748117, + 0.00516556529328227, + 0.008892783895134926, + 0.004773568827658892, + 0.009955087676644325, + 0.004538942128419876, + 0.020340394228696823, + 0.02105862647294998, + 0.02258901670575142, + 0.01659831404685974, + 0.030367767438292503, + 0.0356936901807785, + 0.02080465666949749, + 0.019729118794202805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.16035713255405426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370985761284828 + ], + "opposite_logprob_tokens": [ + -3.408709764480591 + ], + "logprob_reward": -0.03370985761284828, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370985761284828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1729447990655899, + 0.18539974093437195, + 0.12239409983158112, + 0.19579626619815826, + 0.05166454240679741, + 0.06025182083249092, + 0.05813773721456528, + 0.052006833255290985, + 0.036041490733623505, + 0.0653626099228859 + ], + "credit_map": [ + 0.02044791728258133, + 0.020605821162462234, + 0.011209238320589066, + 0.015587861649692059, + 0.016564171761274338, + 0.020658347755670547, + 0.02392587810754776, + 0.04394556209445, + 0.025543522089719772, + 0.02543022856116295, + 0.014769366942346096, + 0.01789192296564579, + 0.030773887410759926, + 0.033013030886650085, + 0.02070491574704647, + 0.01727287657558918, + 0.01196634117513895, + 0.013145958073437214, + 0.013406089507043362, + 0.012512479908764362, + 0.025086339563131332, + 0.015831902623176575, + 0.01684553362429142, + 0.01359946746379137, + 0.01922415941953659, + 0.019684722647070885, + 0.015403620898723602, + 0.022534634917974472, + 0.035760991275310516, + 0.032499656081199646, + 0.023289697244763374, + 0.02739879861474037, + 0.00941099040210247, + 0.008461560122668743, + 0.0062139760702848434, + 0.004884790163487196, + 0.006685289088636637, + 0.0068672639317810535, + 0.004966659937053919, + 0.004174017813056707, + 0.009347292594611645, + 0.01318397093564272, + 0.0033734843600541353, + 0.004056654404848814, + 0.007097338326275349, + 0.00765986880287528, + 0.007537465076893568, + 0.007995749823749065, + 0.008027482777833939, + 0.011249084956943989, + 0.004725859500467777, + 0.0049826414324343204, + 0.00671728840097785, + 0.01004551537334919, + 0.005380092654377222, + 0.007009776309132576, + 0.006245127879083157, + 0.006857122294604778, + 0.0038094166666269302, + 0.006848656572401524, + 0.007235980127006769, + 0.006897329818457365, + 0.007837114855647087, + 0.006276089232414961, + 0.003870139829814434, + 0.004984906874597073, + 0.0021118291188031435, + 0.00420339684933424, + 0.0061734444461762905, + 0.004736768081784248, + 0.005835035350173712, + 0.004125974606722593, + 0.006950231269001961, + 0.0073095220141112804, + 0.0056391325779259205, + 0.004243091680109501, + 0.012769044376909733, + 0.014599650166928768, + 0.00772780179977417, + 0.006124138366430998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.03370985761284828 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unstable, handheld, shaky camera closely follows a relentlessly moving stainless steel conveyor belt, its cold surface coated in a thin layer of condensation as it glides forward in a steady, mechanical rhythm. Under the harsh, cold white light cast by rows of overhead industrial LED tubes, precisely cut, perfectly cubed chunks of raw, bright red meat are neatly arranged on the conveyor belt. Their wet surfaces reflect sharp highlights, and the distinct white fat veins contrast starkly with the deep red muscle tissue. The camera follows the movement of the meat, gently panning up and down and left and right, capturing this world of cold metal and lingering steam. In the background, a complex web of mechanical pipes crisscrosses, with thick white steam occasionally emanating from the seams, obscuring the distant landscape and enhancing the scene's depth and mystery. The entire picture presents a surreal Cinema 4D style, with highly rendered smooth textures and clear geometric shapes. The reflections of metal, the haziness of steam and the vivid colors of meat together construct a cold, alienated yet visually impactful inhuman industrial spectacle.", + "seed": 8830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_3194_seed_8830_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2507485806941986, + "vlm_logprob": -0.2507485806941986, + "vlm_yes_no_margin": 2.3749999523162844, + "vlm_token_credit": -0.2507485806941986, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030983062461018562 + ], + "opposite_logprob_tokens": [ + -3.7809829711914062 + ], + "logprob_reward": -0.030983062461018562, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.030983062461018562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3186306059360504, + 0.09528223425149918, + 0.11575982719659805, + 0.05063294246792793, + 0.13725817203521729, + 0.026615265756845474, + 0.04622616618871689, + 0.056647103279829025, + 0.10634510964155197, + 0.04660259187221527 + ], + "credit_map": [ + 0.0203769002109766, + 0.01834762841463089, + 0.02234751358628273, + 0.09120190888643265, + 0.019043631851673126, + 0.011245554313063622, + 0.1146976575255394, + 0.021369798108935356, + 0.005342200864106417, + 0.007665401324629784, + 0.007459846790879965, + 0.008491121232509613, + 0.008563149720430374, + 0.011904596351087093, + 0.02493097260594368, + 0.020924942567944527, + 0.013493123464286327, + 0.017033107578754425, + 0.011743930168449879, + 0.012976707890629768, + 0.011261878535151482, + 0.008680064231157303, + 0.019973991438746452, + 0.020597035065293312, + 0.006027191877365112, + 0.005922276992350817, + 0.0048708925023674965, + 0.00439046323299408, + 0.008808858692646027, + 0.007845323532819748, + 0.00866029690951109, + 0.004107637330889702, + 0.006176005117595196, + 0.006303706206381321, + 0.005609296727925539, + 0.04922236502170563, + 0.00742635503411293, + 0.015200800262391567, + 0.009099969640374184, + 0.03821966424584389, + 0.004482585936784744, + 0.0022648025769740343, + 0.002446470083668828, + 0.002860349602997303, + 0.004199938848614693, + 0.00332642812281847, + 0.004832285922020674, + 0.0022024051286280155, + 0.004829396028071642, + 0.005356395151466131, + 0.005767364054918289, + 0.0035356238950043917, + 0.007670581806451082, + 0.011610161513090134, + 0.0035514147020876408, + 0.003905230900272727, + 0.005634654778987169, + 0.006930067203938961, + 0.010354958474636078, + 0.006849226541817188, + 0.009136797860264778, + 0.006038160528987646, + 0.006521310191601515, + 0.005181929562240839, + 0.028657374903559685, + 0.013955334201455116, + 0.0022387895733118057, + 0.0034828479401767254, + 0.03655632585287094, + 0.01567404344677925, + 0.002930096350610256, + 0.0028502950444817543, + 0.005347313825041056, + 0.007072167005389929, + 0.0064484174363315105, + 0.004834044724702835, + 0.008435019291937351, + 0.004950123839080334, + 0.005237795412540436, + 0.00427771033719182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.030983062461018562 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981521561741829 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.02981521561741829, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981521561741829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5187488794326782, + 0.08100380003452301, + 0.027336485683918, + 0.01485844049602747, + 0.10679560154676437, + 0.018753664568066597, + 0.01640252396464348, + 0.012960339896380901, + 0.11403679847717285, + 0.08910345286130905 + ], + "credit_map": [ + 0.03160623833537102, + 0.020116671919822693, + 0.02415817230939865, + 0.1560048907995224, + 0.033590301871299744, + 0.015812121331691742, + 0.20561185479164124, + 0.03184859827160835, + 0.00345043302513659, + 0.004088201094418764, + 0.004193353001028299, + 0.005724132061004639, + 0.005199876148253679, + 0.0076496838591992855, + 0.028510766103863716, + 0.022187357768416405, + 0.0028159115463495255, + 0.0032759392634034157, + 0.0031826391350477934, + 0.0038116667419672012, + 0.0029725022614002228, + 0.0042750476859509945, + 0.0037480394821614027, + 0.0032547391019761562, + 0.001404627226293087, + 0.0012564215576276183, + 0.0018491075607016683, + 0.001954692415893078, + 0.002199649577960372, + 0.0026860034558922052, + 0.0022307129111140966, + 0.0012772259069606662, + 0.004204215481877327, + 0.0042921509593725204, + 0.0031958131585270166, + 0.04068780317902565, + 0.002992864465340972, + 0.008271276950836182, + 0.0071329050697386265, + 0.036018576472997665, + 0.0037303788121789694, + 0.0019452684791758657, + 0.0016869819955900311, + 0.0012499479344114661, + 0.0025767257902771235, + 0.0023449514992535114, + 0.0036740207578986883, + 0.0015453906962648034, + 0.001842938712798059, + 0.0020288550294935703, + 0.0019614319317042828, + 0.0014513724017888308, + 0.0027209725230932236, + 0.0034140185452997684, + 0.0018378475215286016, + 0.0011450857855379581, + 0.0013693199725821614, + 0.0013085838872939348, + 0.0015540783060714602, + 0.00154939538333565, + 0.0029112761840224266, + 0.0014586296165362, + 0.0016135864425450563, + 0.001195470686070621, + 0.027487149462103844, + 0.017412986606359482, + 0.0022238923702389, + 0.0035366553347557783, + 0.039809923619031906, + 0.017874889075756073, + 0.002884699497371912, + 0.0028066071681678295, + 0.00992228090763092, + 0.01273698266595602, + 0.008380887098610401, + 0.008777303621172905, + 0.023417284712195396, + 0.0103056151419878, + 0.008598698303103447, + 0.006964396685361862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.02981521561741829 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261285424232483 + ], + "opposite_logprob_tokens": [ + -0.5761285424232483 + ], + "logprob_reward": -0.8261285424232483, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261285424232483, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3763848543167114, + 0.06588274985551834, + 0.09958882629871368, + 0.06187715753912926, + 0.06684473156929016, + 0.06583670526742935, + 0.0716744214296341, + 0.0417436882853508, + 0.03924063965678215, + 0.11092615872621536 + ], + "credit_map": [ + 0.022774456068873405, + 0.025960156694054604, + 0.0301199983805418, + 0.11047109216451645, + 0.024548640474677086, + 0.015597857534885406, + 0.12425798922777176, + 0.022654641419649124, + 0.005954628810286522, + 0.00802586693316698, + 0.007931104861199856, + 0.006150516681373119, + 0.008521328680217266, + 0.007775692269206047, + 0.011274642311036587, + 0.010248975828289986, + 0.010445198975503445, + 0.015813974663615227, + 0.008690816350281239, + 0.011707407422363758, + 0.00953738484531641, + 0.010907605290412903, + 0.011824856512248516, + 0.020661581307649612, + 0.009945319965481758, + 0.004338674247264862, + 0.00627526268362999, + 0.00614975206553936, + 0.006075268145650625, + 0.007162465713918209, + 0.008240137249231339, + 0.013690275140106678, + 0.004415494855493307, + 0.004732091911137104, + 0.006553583778440952, + 0.01328697893768549, + 0.008686873130500317, + 0.01852305233478546, + 0.004874657839536667, + 0.005771998781710863, + 0.011008421890437603, + 0.005889948457479477, + 0.005490713287144899, + 0.004773283842951059, + 0.010306291282176971, + 0.009125112555921078, + 0.013639582321047783, + 0.00560334837064147, + 0.006902485154569149, + 0.00821518711745739, + 0.00996229238808155, + 0.004776401445269585, + 0.011168493889272213, + 0.01934817060828209, + 0.005065773613750935, + 0.006235615350306034, + 0.004665021318942308, + 0.006096653640270233, + 0.007966754958033562, + 0.005118523258715868, + 0.005435526836663485, + 0.003919923212379217, + 0.003882704069837928, + 0.004658577963709831, + 0.005367367994040251, + 0.007523661945015192, + 0.0023103421553969383, + 0.0031831294763833284, + 0.007886447012424469, + 0.007740745320916176, + 0.0020691286772489548, + 0.003159817773848772, + 0.008225162513554096, + 0.014233162626624107, + 0.01565738394856453, + 0.011546164751052856, + 0.022964971140027046, + 0.017353542149066925, + 0.011407379992306232, + 0.009538387879729271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.8261285424232483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862175583839417 + ], + "opposite_logprob_tokens": [ + -1.2236217260360718 + ], + "logprob_reward": -0.34862175583839417, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34862175583839417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2633958160877228, + 0.05532417073845863, + 0.02130044624209404, + 0.02362695150077343, + 0.10184718668460846, + 0.05317775160074234, + 0.018268676474690437, + 0.027066966518759727, + 0.22731110453605652, + 0.20868094265460968 + ], + "credit_map": [ + 0.020822573453187943, + 0.024848362430930138, + 0.0338970385491848, + 0.03970334678888321, + 0.022813519462943077, + 0.01837126724421978, + 0.06910298764705658, + 0.03383670747280121, + 0.002278328174725175, + 0.002486519981175661, + 0.003338922979310155, + 0.003557936754077673, + 0.0040678405202925205, + 0.007092468906193972, + 0.018500640988349915, + 0.014001508243381977, + 0.002113758120685816, + 0.002715412061661482, + 0.0026829997077584267, + 0.003193204291164875, + 0.0018622002098709345, + 0.0024723345413804054, + 0.0037236579228192568, + 0.002536877989768982, + 0.0019475942244753242, + 0.0027123061008751392, + 0.002065467881038785, + 0.0021473756060004234, + 0.004951522219926119, + 0.004009544849395752, + 0.0037241324316710234, + 0.0020690057426691055, + 0.004576022271066904, + 0.004920520354062319, + 0.003728233277797699, + 0.034441567957401276, + 0.004788509104400873, + 0.02044971100986004, + 0.006626702379435301, + 0.02231590822339058, + 0.009057079441845417, + 0.005273603834211826, + 0.00498606264591217, + 0.0033811554312705994, + 0.008515352383255959, + 0.0076427520252764225, + 0.009476344101130962, + 0.004845397546887398, + 0.0025136203039437532, + 0.0017314184224233031, + 0.002463046694174409, + 0.0013347020139917731, + 0.0026511666364967823, + 0.00465011689811945, + 0.0015209546545520425, + 0.0014036484062671661, + 0.003046168480068445, + 0.0036722298245877028, + 0.00356672378256917, + 0.0042763808742165565, + 0.004460456315428019, + 0.002661481499671936, + 0.003248442430049181, + 0.0021350793540477753, + 0.05400130897760391, + 0.036562804132699966, + 0.005467303097248077, + 0.007478718180209398, + 0.08018714934587479, + 0.03186037018895149, + 0.005668241530656815, + 0.006085190922021866, + 0.014830322004854679, + 0.03961366042494774, + 0.022289853543043137, + 0.012324429117143154, + 0.07017363607883453, + 0.026287896558642387, + 0.012029876932501793, + 0.011131247505545616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.34862175583839417 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194327130913734 + ], + "opposite_logprob_tokens": [ + -4.018194198608398 + ], + "logprob_reward": -0.018194327130913734, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018194327130913734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3248530924320221, + 0.09373367577791214, + 0.0367598831653595, + 0.04023657739162445, + 0.22721415758132935, + 0.014657987281680107, + 0.02401532046496868, + 0.02582688443362713, + 0.13942722976207733, + 0.07327516376972198 + ], + "credit_map": [ + 0.01344582810997963, + 0.01887655444443226, + 0.03558938205242157, + 0.10006118565797806, + 0.021182620897889137, + 0.014494327828288078, + 0.09457399696111679, + 0.026629241183400154, + 0.007274313364177942, + 0.009273281320929527, + 0.00960482843220234, + 0.009976205416023731, + 0.009930351749062538, + 0.009378167800605297, + 0.02074572443962097, + 0.01755080558359623, + 0.003334119450300932, + 0.0035986253060400486, + 0.004470554646104574, + 0.005246568471193314, + 0.0022553622256964445, + 0.0036904606968164444, + 0.009237753227353096, + 0.004926439840346575, + 0.00511245708912611, + 0.004171506501734257, + 0.0033009317703545094, + 0.0042226905934512615, + 0.0064759342931210995, + 0.005549605004489422, + 0.0058431862853467464, + 0.005560270044952631, + 0.009717730805277824, + 0.010599064640700817, + 0.007952737621963024, + 0.07869450002908707, + 0.01040879637002945, + 0.04287521913647652, + 0.014905010350048542, + 0.052061110734939575, + 0.002246318617835641, + 0.0012370080221444368, + 0.001161106745712459, + 0.0020664792973548174, + 0.002149637322872877, + 0.0019341447623446584, + 0.002113910159096122, + 0.001749382819980383, + 0.0028431883547455072, + 0.0025419252924621105, + 0.0032622336875647306, + 0.0017013895558193326, + 0.003713600104674697, + 0.006276288535445929, + 0.0016878320602700114, + 0.0019888635724782944, + 0.002467603189870715, + 0.0032741203904151917, + 0.004520777612924576, + 0.003936127293854952, + 0.003463584929704666, + 0.002862561959773302, + 0.0031901283655315638, + 0.0021119844168424606, + 0.03317866101861, + 0.021693764254450798, + 0.0038046687841415405, + 0.0050407713279128075, + 0.04971943795681, + 0.017105944454669952, + 0.003938260488212109, + 0.004945727996528149, + 0.004701895173639059, + 0.0119717912748456, + 0.007460416294634342, + 0.005283967591822147, + 0.023295897990465164, + 0.01024326030164957, + 0.005483798682689667, + 0.0048341392539441586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.018194327130913734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unstable, handheld, shaky camera closely follows a relentlessly moving stainless steel conveyor belt, its cold surface coated in a thin layer of condensation as it glides forward in a steady, mechanical rhythm. Under the harsh, cold white light cast by rows of overhead industrial LED tubes, precisely cut, perfectly cubed chunks of raw, bright red meat are neatly arranged on the conveyor belt. Their wet surfaces reflect sharp highlights, and the distinct white fat veins contrast starkly with the deep red muscle tissue. The camera follows the movement of the meat, gently panning up and down and left and right, capturing this world of cold metal and lingering steam. In the background, a complex web of mechanical pipes crisscrosses, with thick white steam occasionally emanating from the seams, obscuring the distant landscape and enhancing the scene's depth and mystery. The entire picture presents a surreal Cinema 4D style, with highly rendered smooth textures and clear geometric shapes. The reflections of metal, the haziness of steam and the vivid colors of meat together construct a cold, alienated yet visually impactful inhuman industrial spectacle.", + "seed": 8831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_10/index_3194_seed_8831_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14476363435387612, + "vlm_logprob": -0.14476363435387612, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.14476363435387612, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041395027190446854 + ], + "opposite_logprob_tokens": [ + -4.04139518737793 + ], + "logprob_reward": -0.041395027190446854, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.041395027190446854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3753311038017273, + 0.059188857674598694, + 0.11108077317476273, + 0.06739667057991028, + 0.07667254656553268, + 0.04576411098241806, + 0.07360880821943283, + 0.044571276754140854, + 0.02707352116703987, + 0.11931236833333969 + ], + "credit_map": [ + 0.02930043451488018, + 0.09460795670747757, + 0.02087770216166973, + 0.039413392543792725, + 0.074535071849823, + 0.05755077674984932, + 0.032692864537239075, + 0.02635289542376995, + 0.006874979939311743, + 0.008065272122621536, + 0.006140814628452063, + 0.006417362950742245, + 0.012049761600792408, + 0.009467900730669498, + 0.0057326690293848515, + 0.004440095741301775, + 0.011074290610849857, + 0.018318770453333855, + 0.011186322197318077, + 0.013706250116229057, + 0.022553903982043266, + 0.015338830649852753, + 0.008186114951968193, + 0.010716278105974197, + 0.005884954240173101, + 0.008336500264704227, + 0.005189245566725731, + 0.00864468701183796, + 0.009186171926558018, + 0.010764791630208492, + 0.006626154761761427, + 0.012764165177941322, + 0.008172198198735714, + 0.013134895823895931, + 0.00548434816300869, + 0.011150541715323925, + 0.010506768710911274, + 0.009819276630878448, + 0.008228395134210587, + 0.010176117531955242, + 0.0056344326585531235, + 0.009020925499498844, + 0.0038463566452264786, + 0.004104959778487682, + 0.006838803179562092, + 0.006411828100681305, + 0.005958477966487408, + 0.003948328550904989, + 0.005870187189429998, + 0.010370486415922642, + 0.005661674775183201, + 0.010745617561042309, + 0.00774902431294322, + 0.01103166863322258, + 0.011783160269260406, + 0.010396990925073624, + 0.004332142416387796, + 0.004200807772576809, + 0.003374163992702961, + 0.006026336923241615, + 0.0032880359794944525, + 0.005491766147315502, + 0.005436714272946119, + 0.012421310879290104, + 0.0026172022335231304, + 0.003669114550575614, + 0.0033521733712404966, + 0.003964649513363838, + 0.0033871124032884836, + 0.0032097341027110815, + 0.003779266495257616, + 0.0030942698940634727, + 0.007566920015960932, + 0.009002248756587505, + 0.006172905210405588, + 0.02607404813170433, + 0.011799115687608719, + 0.012047897092998028, + 0.03380300849676132, + 0.01284622959792614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.041395027190446854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003237307071686 + ], + "opposite_logprob_tokens": [ + -2.6950323581695557 + ], + "logprob_reward": -0.07003237307071686, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003237307071686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32323917746543884, + 0.037423182278871536, + 0.12652809917926788, + 0.03800911456346512, + 0.15603306889533997, + 0.11470871418714523, + 0.09871863573789597, + 0.046652473509311676, + 0.016304349526762962, + 0.042383164167404175 + ], + "credit_map": [ + 0.02775389887392521, + 0.08739941567182541, + 0.020195264369249344, + 0.03170819580554962, + 0.0569903664290905, + 0.0530061274766922, + 0.02620571479201317, + 0.019980184733867645, + 0.0046447766944766045, + 0.004611427895724773, + 0.004657619632780552, + 0.00486336974427104, + 0.006314603611826897, + 0.004423833917826414, + 0.003977870102971792, + 0.003929682541638613, + 0.014400343410670757, + 0.024520885199308395, + 0.009098010137677193, + 0.013700153678655624, + 0.028886770829558372, + 0.01323722768574953, + 0.00987014640122652, + 0.012814576737582684, + 0.0038332773838192225, + 0.004776333924382925, + 0.006780677940696478, + 0.005048425868153572, + 0.00547567754983902, + 0.003413696074858308, + 0.0034575914032757282, + 0.005223434418439865, + 0.017855336889624596, + 0.022416727617383003, + 0.012914442457258701, + 0.02064741589128971, + 0.02295919880270958, + 0.020328138023614883, + 0.016822028905153275, + 0.022089781239628792, + 0.008574566803872585, + 0.009779855608940125, + 0.009665698744356632, + 0.015388952568173409, + 0.022922171279788017, + 0.0259365513920784, + 0.012170561589300632, + 0.010270358994603157, + 0.0057310499250888824, + 0.00765572115778923, + 0.005951056256890297, + 0.017667273059487343, + 0.00991755723953247, + 0.015068343840539455, + 0.02235626056790352, + 0.014371375553309917, + 0.004324608016759157, + 0.004120262805372477, + 0.0034666399005800486, + 0.0064262100495398045, + 0.0027476230170577765, + 0.005518319085240364, + 0.005955399479717016, + 0.014093410223722458, + 0.0015676948241889477, + 0.002672168193385005, + 0.0015573984710499644, + 0.002151248510926962, + 0.002639057580381632, + 0.0022244371939450502, + 0.001798716257326305, + 0.0016936283791437745, + 0.0034181640949100256, + 0.004231995902955532, + 0.002822393085807562, + 0.007191663142293692, + 0.004969156347215176, + 0.00616555567830801, + 0.00907130166888237, + 0.004512933548539877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.07003237307071686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208580493927 + ], + "opposite_logprob_tokens": [ + -1.50208580493927 + ], + "logprob_reward": -0.25208580493927, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208580493927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27488183975219727, + 0.0699419304728508, + 0.07101115584373474, + 0.14235393702983856, + 0.16499219834804535, + 0.07878680527210236, + 0.040219467133283615, + 0.07219923287630081, + 0.0597517266869545, + 0.025861706584692 + ], + "credit_map": [ + 0.02528485096991062, + 0.07610329240560532, + 0.018717285245656967, + 0.025354688987135887, + 0.0477314256131649, + 0.04411916434764862, + 0.020951882004737854, + 0.016619231551885605, + 0.006872270721942186, + 0.016124054789543152, + 0.004888028372079134, + 0.006606025621294975, + 0.01687653549015522, + 0.010880019515752792, + 0.0039271763525903225, + 0.003767813090234995, + 0.006832171231508255, + 0.012047920376062393, + 0.007029668428003788, + 0.008022647351026535, + 0.00932899210602045, + 0.00852167047560215, + 0.0075759487226605415, + 0.011652129702270031, + 0.010470667853951454, + 0.01794685423374176, + 0.011293413117527962, + 0.01998169906437397, + 0.01974930427968502, + 0.022848978638648987, + 0.013905754312872887, + 0.026157258078455925, + 0.016263917088508606, + 0.020313480868935585, + 0.012757916003465652, + 0.019988492131233215, + 0.021722549572587013, + 0.02378278598189354, + 0.025800680741667747, + 0.0243623536080122, + 0.009110216051340103, + 0.008229516446590424, + 0.00678399158641696, + 0.011884242296218872, + 0.012009202502667904, + 0.012589720077812672, + 0.009775396436452866, + 0.008404512889683247, + 0.003653117921203375, + 0.005783591885119677, + 0.0033594444394111633, + 0.0060613155364990234, + 0.006479840259999037, + 0.0058683669194579124, + 0.004237562883645296, + 0.0047762226313352585, + 0.0071932123973965645, + 0.006531742867082357, + 0.005307083483785391, + 0.010900813154876232, + 0.005061998963356018, + 0.008842353709042072, + 0.008984911255538464, + 0.01937711238861084, + 0.0043878997676074505, + 0.00518766138702631, + 0.006376474630087614, + 0.011958845891058445, + 0.005194180179387331, + 0.007766474038362503, + 0.006986523978412151, + 0.011893657967448235, + 0.0020075985230505466, + 0.0024884017184376717, + 0.0015734077896922827, + 0.004875815007835627, + 0.0025653918273746967, + 0.002817927859723568, + 0.006672012619674206, + 0.0028611498419195414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.25208580493927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28131353855133057 + ], + "opposite_logprob_tokens": [ + -1.4063135385513306 + ], + "logprob_reward": -0.28131353855133057, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28131353855133057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1331929713487625, + 0.08929596841335297, + 0.11902261525392532, + 0.04021364450454712, + 0.1347673386335373, + 0.0529051348567009, + 0.15969863533973694, + 0.08245331794023514, + 0.0871686339378357, + 0.1012817770242691 + ], + "credit_map": [ + 0.01249261200428009, + 0.025387855246663094, + 0.016676271334290504, + 0.011359273456037045, + 0.024553848430514336, + 0.01916177198290825, + 0.013299507088959217, + 0.010261835530400276, + 0.010045618750154972, + 0.011449496261775494, + 0.008866889402270317, + 0.009612802416086197, + 0.015009465627372265, + 0.013398366048932076, + 0.010720745660364628, + 0.010192584246397018, + 0.0133549515157938, + 0.01942109875380993, + 0.009219245985150337, + 0.012497365474700928, + 0.02495298720896244, + 0.013385696336627007, + 0.010953287594020367, + 0.015237982384860516, + 0.003983485512435436, + 0.006737884599715471, + 0.004691618494689465, + 0.005133951082825661, + 0.0049518742598593235, + 0.004500413313508034, + 0.0040300809778273106, + 0.006184339988976717, + 0.014635376632213593, + 0.019081667065620422, + 0.012519474141299725, + 0.01665298268198967, + 0.019604569301009178, + 0.02202906832098961, + 0.011951620690524578, + 0.01829257607460022, + 0.004914653021842241, + 0.007513372227549553, + 0.005442009307444096, + 0.007129882927983999, + 0.008424745872616768, + 0.009592564776539803, + 0.005240495316684246, + 0.004647413734346628, + 0.010122308507561684, + 0.017086749896407127, + 0.012218039482831955, + 0.02159995026886463, + 0.022480433806777, + 0.02433149516582489, + 0.02858388051390648, + 0.023275764659047127, + 0.011285376735031605, + 0.008744724094867706, + 0.005599235184490681, + 0.010010483674705029, + 0.005000836681574583, + 0.009354005567729473, + 0.009393727406859398, + 0.02306492254137993, + 0.0067867496982216835, + 0.013190609402954578, + 0.009391949512064457, + 0.013369441032409668, + 0.010622595436871052, + 0.014092197641730309, + 0.008692390285432339, + 0.011022700928151608, + 0.005009416490793228, + 0.007403427269309759, + 0.004541198257356882, + 0.019496681168675423, + 0.014276809990406036, + 0.015647465363144875, + 0.025461234152317047, + 0.009445542469620705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.28131353855133057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a stainless steel conveyor belt carrying cubed chunks of raw red meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the conveyor belt surface coated in a thin layer of condensation with wet surfaces reflecting sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot handheld and shaky, following the movement of the meat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with a complex web of mechanical pipes emitting thick white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a surreal Cinema 4D rendering with smooth textures and clear geometric shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899142801761627 + ], + "opposite_logprob_tokens": [ + -2.578991413116455 + ], + "logprob_reward": -0.07899142801761627, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899142801761627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2148328274488449, + 0.046007100492715836, + 0.17577128112316132, + 0.11779427528381348, + 0.15721502900123596, + 0.11421097815036774, + 0.06989278644323349, + 0.04311705753207207, + 0.022636741399765015, + 0.0385218970477581 + ], + "credit_map": [ + 0.018783094361424446, + 0.05290861427783966, + 0.0146385682746768, + 0.021127061918377876, + 0.03364364802837372, + 0.03186457231640816, + 0.024630937725305557, + 0.017236333340406418, + 0.004783890210092068, + 0.007670186925679445, + 0.004072327166795731, + 0.00548077467828989, + 0.010509195737540722, + 0.006509929429739714, + 0.0035711603704839945, + 0.0034096373710781336, + 0.017369097098708153, + 0.03502655401825905, + 0.012305312789976597, + 0.019529083743691444, + 0.03825802728533745, + 0.01839265786111355, + 0.013898293487727642, + 0.020992273464798927, + 0.008445474319159985, + 0.01643851399421692, + 0.00999835878610611, + 0.013939233496785164, + 0.018863335251808167, + 0.019830798730254173, + 0.010990827344357967, + 0.019287731498479843, + 0.020019493997097015, + 0.026158759370446205, + 0.010397937148809433, + 0.01920025423169136, + 0.02161237597465515, + 0.021669693291187286, + 0.017926819622516632, + 0.020229702815413475, + 0.008637077175080776, + 0.01024242490530014, + 0.010071463882923126, + 0.014215104281902313, + 0.02440032735466957, + 0.026232698932290077, + 0.01080260518938303, + 0.009609273634850979, + 0.005896070972084999, + 0.010921013541519642, + 0.0058234077878296375, + 0.006735139060765505, + 0.010454080067574978, + 0.010141344740986824, + 0.01082637533545494, + 0.009095358662307262, + 0.005879465956240892, + 0.0049374462105333805, + 0.002884025452658534, + 0.006312455050647259, + 0.0029863405507057905, + 0.0053391968831419945, + 0.004868547897785902, + 0.009909579530358315, + 0.002070437651127577, + 0.002763336058706045, + 0.002281531924381852, + 0.0036070761270821095, + 0.002551604062318802, + 0.0027327402494847775, + 0.003260764991864562, + 0.0033692498691380024, + 0.00326615315862, + 0.0032747911754995584, + 0.002415986265987158, + 0.0026814723387360573, + 0.010380138643085957, + 0.009881723672151566, + 0.004038899205625057, + 0.0025827325880527496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.07899142801761627 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_11/index_tensor([7042, 7042])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_11/index_tensor([7042, 7042])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b35dc864ccd24be1f65e3e6e1beecdc32dec3224 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_11/index_tensor([7042, 7042])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from an extremely low angle, the camera slowly and steadily tilts upward, following a pink lotus bud about to bloom. With a lifelike, slow, and resolute gesture, it pierces the nearly opaque surface of deep, dark green water and rises. The bud is compact, spindle-shaped, and its outer petals display a delicate gradation from soft ivory white at the base to delicate pink at the tip. Coated with a nearly invisible waxy layer, the surface allows the crystal-clear water droplets to tumble down the graceful curves of the petals like tiny diamonds, ultimately dripping back into the murky water below. A soft yet focused overhead light casts precisely on the lotus, making it the sole focal point against the bottomless, pure darkness. The light not only outlines the full contours of the petals but also reflects a dazzling radiance from each falling droplet. The entire image is imbued with a sense of Zen and the power of life, conveying a profound atmosphere of purity, resilience, and tranquility born from darkness.", + "seed": 28224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_11/index_7042_seed_28224_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.037298434227705, + "vlm_logprob": -0.037298434227705, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.037298434227705, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07814783602952957 + ], + "opposite_logprob_tokens": [ + -3.0781478881835938 + ], + "logprob_reward": -0.07814783602952957, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.07814783602952957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2602567672729492, + 0.059825386852025986, + 0.0760231465101242, + 0.07042296230792999, + 0.11286940425634384, + 0.14478790760040283, + 0.05018298700451851, + 0.15337646007537842, + 0.03867106884717941, + 0.03358389437198639 + ], + "credit_map": [ + 0.04770709574222565, + 0.01631111279129982, + 0.03732854872941971, + 0.034835923463106155, + 0.023898176848888397, + 0.017834879457950592, + 0.050388846546411514, + 0.031952179968357086, + 0.008269236423075199, + 0.005631787236779928, + 0.009793093428015709, + 0.008272706530988216, + 0.006378475576639175, + 0.004102450795471668, + 0.010347248986363411, + 0.007030387409031391, + 0.007990214042365551, + 0.006074822973459959, + 0.0115147540345788, + 0.01471063494682312, + 0.006084495689719915, + 0.006108697038143873, + 0.014260183088481426, + 0.009279346093535423, + 0.010281586088240147, + 0.005106670781970024, + 0.01176169328391552, + 0.0087382597848773, + 0.006433788221329451, + 0.005059479735791683, + 0.012510999105870724, + 0.010530478321015835, + 0.013980588875710964, + 0.006431274581700563, + 0.018399527296423912, + 0.013845378533005714, + 0.009461520239710808, + 0.010508468374609947, + 0.023502949625253677, + 0.016739701852202415, + 0.02096540294587612, + 0.01085225585848093, + 0.02401181310415268, + 0.015248163603246212, + 0.01429100800305605, + 0.013200250454246998, + 0.024571755900979042, + 0.0216472577303648, + 0.008679650723934174, + 0.004408018197864294, + 0.008295518346130848, + 0.0038608673494309187, + 0.0035449338611215353, + 0.004270216450095177, + 0.006788359489291906, + 0.010335427708923817, + 0.02402680739760399, + 0.014225582592189312, + 0.02080991305410862, + 0.01547288242727518, + 0.018816135823726654, + 0.012756367214024067, + 0.022211389616131783, + 0.025057386606931686, + 0.007013776805251837, + 0.003925792407244444, + 0.005458927247673273, + 0.0029106128495186567, + 0.0048566702753305435, + 0.004332255106419325, + 0.005396980792284012, + 0.004776054993271828, + 0.006010173354297876, + 0.003278734628111124, + 0.00401205662637949, + 0.0029930430464446545, + 0.0035357021261006594, + 0.005094051361083984, + 0.0048387316055595875, + 0.003821403719484806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.07814783602952957 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813838213682175 + ], + "opposite_logprob_tokens": [ + -3.2881383895874023 + ], + "logprob_reward": -0.03813838213682175, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813838213682175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08133003115653992, + 0.061838943511247635, + 0.026751592755317688, + 0.04844934120774269, + 0.088679738342762, + 0.17677220702171326, + 0.09768502414226532, + 0.37808024883270264, + 0.026950426399707794, + 0.013462433591485023 + ], + "credit_map": [ + 0.010631081648170948, + 0.006090770475566387, + 0.010645372793078423, + 0.01609320566058159, + 0.0062238010577857494, + 0.006789724342525005, + 0.0155652966350317, + 0.009290783666074276, + 0.007602249272167683, + 0.004147399216890335, + 0.008824314922094345, + 0.009430412203073502, + 0.0060857427306473255, + 0.00510217621922493, + 0.014267547987401485, + 0.006379101891070604, + 0.003244909690693021, + 0.0021557030268013477, + 0.003912326414138079, + 0.0045790602453053, + 0.0023669740185141563, + 0.0023219517897814512, + 0.0044556730426847935, + 0.0037149935960769653, + 0.0061777993105351925, + 0.0042402357794344425, + 0.007197607774287462, + 0.006170765031129122, + 0.00377566902898252, + 0.0046683805994689465, + 0.010464816354215145, + 0.005754066631197929, + 0.008674276992678642, + 0.006112050265073776, + 0.014067876152694225, + 0.01144501008093357, + 0.006628390867263079, + 0.00854535587131977, + 0.018971970304846764, + 0.014234802685678005, + 0.02678021602332592, + 0.011671289801597595, + 0.029055003076791763, + 0.01764897257089615, + 0.016862986609339714, + 0.013121889904141426, + 0.03218240290880203, + 0.02944943867623806, + 0.01849731244146824, + 0.00927302148193121, + 0.016089104115962982, + 0.009658818133175373, + 0.008287213742733002, + 0.007033847272396088, + 0.012646856717765331, + 0.016198845580220222, + 0.04879629611968994, + 0.038825515657663345, + 0.0495862141251564, + 0.03716497868299484, + 0.05499668046832085, + 0.04398958012461662, + 0.05745459720492363, + 0.0472663976252079, + 0.0025456843432039022, + 0.0023898526560515165, + 0.0034856556449085474, + 0.0026188010815531015, + 0.0044196187518537045, + 0.003444908419623971, + 0.004015270154923201, + 0.004030636511743069, + 0.0018474487587809563, + 0.0009357893723063171, + 0.0019825559575110674, + 0.0014319909969344735, + 0.00156917585991323, + 0.001855295617133379, + 0.001959583954885602, + 0.0018805930158123374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.03813838213682175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.162051260471344, + 0.0835409089922905, + 0.06627270579338074, + 0.07686205208301544, + 0.19678430259227753, + 0.03689082711935043, + 0.0681348368525505, + 0.1641395390033722, + 0.08231028914451599, + 0.06301335245370865 + ], + "credit_map": [ + 0.01927240379154682, + 0.013441750779747963, + 0.021493075415492058, + 0.025334391742944717, + 0.01663779653608799, + 0.011253144592046738, + 0.03400101512670517, + 0.020617669448256493, + 0.00912535097450018, + 0.006549757439643145, + 0.011462604627013206, + 0.010023844428360462, + 0.010551748797297478, + 0.006368353497236967, + 0.017435088753700256, + 0.012024154886603355, + 0.007345883641391993, + 0.005827953573316336, + 0.010483142919838428, + 0.009756702929735184, + 0.006676633842289448, + 0.005206242669373751, + 0.011674739420413971, + 0.009301401674747467, + 0.007247717119753361, + 0.006601544562727213, + 0.010744523257017136, + 0.009046817198395729, + 0.007717425003647804, + 0.004991607740521431, + 0.01838725619018078, + 0.012125147506594658, + 0.018828513100743294, + 0.014675119891762733, + 0.028916431590914726, + 0.022660719230771065, + 0.01765833981335163, + 0.026031063869595528, + 0.035068709403276443, + 0.032945387065410614, + 0.0057577150873839855, + 0.002854526275768876, + 0.0053977202624082565, + 0.0032957864459604025, + 0.002994801616296172, + 0.00321283214725554, + 0.006380528211593628, + 0.006996912881731987, + 0.011310886591672897, + 0.005497328005731106, + 0.01007487066090107, + 0.009417484514415264, + 0.0051463996060192585, + 0.007166900206357241, + 0.010481566190719604, + 0.009039398282766342, + 0.020904721692204475, + 0.01772313378751278, + 0.02243918739259243, + 0.017207374796271324, + 0.025300845503807068, + 0.019034981727600098, + 0.023958414793014526, + 0.017570864409208298, + 0.007855471223592758, + 0.005417353007942438, + 0.010784131474792957, + 0.007555547170341015, + 0.010007651522755623, + 0.009064864367246628, + 0.01420085784047842, + 0.017424406483769417, + 0.01038793008774519, + 0.005028679035604, + 0.006971498019993305, + 0.004712570924311876, + 0.006886085495352745, + 0.010767048224806786, + 0.009371410124003887, + 0.008888130076229572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.01107780635356903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306282848119736 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306282848119736, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306282848119736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5891426205635071, + 0.07207685708999634, + 0.03328939899802208, + 0.05828550085425377, + 0.08989612758159637, + 0.035569023340940475, + 0.032464392483234406, + 0.023986654356122017, + 0.04755145311355591, + 0.01773793064057827 + ], + "credit_map": [ + 0.09322492778301239, + 0.04005032032728195, + 0.07770474255084991, + 0.10202834755182266, + 0.04564293473958969, + 0.03955571725964546, + 0.12264438718557358, + 0.06829128414392471, + 0.012563291937112808, + 0.005910025909543037, + 0.011183769442141056, + 0.009165063500404358, + 0.005845468956977129, + 0.004489663522690535, + 0.013644604943692684, + 0.00927497074007988, + 0.0029250297229737043, + 0.0035597216337919235, + 0.005355833563953638, + 0.004716162104159594, + 0.0035192330833524466, + 0.0035436442121863365, + 0.005944965872913599, + 0.0037248115986585617, + 0.0078537343069911, + 0.004198324400931597, + 0.010060987435281277, + 0.0073723746463656425, + 0.004947816953063011, + 0.004442918114364147, + 0.011403447017073631, + 0.008005902171134949, + 0.00991602148860693, + 0.005656806752085686, + 0.014453906565904617, + 0.01057014986872673, + 0.007289101369678974, + 0.010105490684509277, + 0.01687297783792019, + 0.015031677670776844, + 0.005440318025648594, + 0.0034459687303751707, + 0.00499798683449626, + 0.004004618618637323, + 0.0028125676326453686, + 0.0037257387302815914, + 0.005669554229825735, + 0.005472270771861076, + 0.005048220977187157, + 0.002228663070127368, + 0.004429598338901997, + 0.004594528116285801, + 0.002654677489772439, + 0.0030192977283149958, + 0.0070469263009727, + 0.0034424844197928905, + 0.003268619067966938, + 0.002069203881546855, + 0.003416047664359212, + 0.0024560142774134874, + 0.003395984647795558, + 0.002076603239402175, + 0.004268177319318056, + 0.0030360056553035975, + 0.0052063786424696445, + 0.003360422793775797, + 0.006395268719643354, + 0.0038657786790281534, + 0.0060200863517820835, + 0.007231945637613535, + 0.006909002549946308, + 0.008562575094401836, + 0.003240311285480857, + 0.0013635300565510988, + 0.002634746953845024, + 0.0015419989358633757, + 0.0016193316550925374, + 0.002993559231981635, + 0.002263957867398858, + 0.0020804950036108494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04306282848119736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3236844837665558, + 0.11683753877878189, + 0.04844101890921593, + 0.05820530652999878, + 0.11013162136077881, + 0.044381313025951385, + 0.06134272739291191, + 0.11478792876005173, + 0.08251994848251343, + 0.039668042212724686 + ], + "credit_map": [ + 0.03949017822742462, + 0.031709007918834686, + 0.04497230052947998, + 0.053108666092157364, + 0.03255006670951843, + 0.019048083573579788, + 0.06068317964673042, + 0.04212303087115288, + 0.025142397731542587, + 0.007210570853203535, + 0.0246741846203804, + 0.010342157445847988, + 0.007129153702408075, + 0.006579637993127108, + 0.017317816615104675, + 0.018441636115312576, + 0.0066397967748343945, + 0.003523002378642559, + 0.006711509078741074, + 0.00813710130751133, + 0.0036489772610366344, + 0.0030880137346684933, + 0.012133518233895302, + 0.004559108056128025, + 0.006850809324532747, + 0.004061894956976175, + 0.008773698471486568, + 0.007600891869515181, + 0.00563044473528862, + 0.00340079003944993, + 0.015274889767169952, + 0.006611895281821489, + 0.014394096098840237, + 0.00978452991694212, + 0.015503580681979656, + 0.011268298141658306, + 0.008672835305333138, + 0.014550382271409035, + 0.019195038825273514, + 0.0167628712952137, + 0.009096208028495312, + 0.003222154453396797, + 0.007938816212117672, + 0.003556109732016921, + 0.003784103086218238, + 0.004474056884646416, + 0.006825948134064674, + 0.005483920220285654, + 0.007767211180180311, + 0.005032849032431841, + 0.009197086095809937, + 0.00831855833530426, + 0.0048189181834459305, + 0.007135489024221897, + 0.009300359524786472, + 0.009772264398634434, + 0.018881220370531082, + 0.011700429953634739, + 0.015529895201325417, + 0.010964155197143555, + 0.015032761730253696, + 0.010284577496349812, + 0.01654007099568844, + 0.015854833647608757, + 0.009367190301418304, + 0.005916606169193983, + 0.010522356256842613, + 0.0058152624405920506, + 0.01186978816986084, + 0.013429801911115646, + 0.011398667469620705, + 0.01420027855783701, + 0.009169249795377254, + 0.002967861481010914, + 0.006090906914323568, + 0.0027260470669716597, + 0.0038012173026800156, + 0.004721034783869982, + 0.004514999222010374, + 0.005676730535924435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016065318137407303 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera slowly and steadily tilts upward, following a pink lotus bud about to bloom. With a lifelike, slow, and resolute gesture, it pierces the nearly opaque surface of deep, dark green water and rises. The bud is compact, spindle-shaped, and its outer petals display a delicate gradation from soft ivory white at the base to delicate pink at the tip. Coated with a nearly invisible waxy layer, the surface allows the crystal-clear water droplets to tumble down the graceful curves of the petals like tiny diamonds, ultimately dripping back into the murky water below. A soft yet focused overhead light casts precisely on the lotus, making it the sole focal point against the bottomless, pure darkness. The light not only outlines the full contours of the petals but also reflects a dazzling radiance from each falling droplet. The entire image is imbued with a sense of Zen and the power of life, conveying a profound atmosphere of purity, resilience, and tranquility born from darkness.", + "seed": 28225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_11/index_7042_seed_28225_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03564912797883153, + "vlm_logprob": -0.03564912797883153, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03564912797883153, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058444391936063766 + ], + "opposite_logprob_tokens": [ + -3.9334444999694824 + ], + "logprob_reward": -0.058444391936063766, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.058444391936063766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3165026009082794, + 0.08958160877227783, + 0.0391186885535717, + 0.12665356695652008, + 0.03966737538576126, + 0.08696518838405609, + 0.0710766538977623, + 0.11178980767726898, + 0.04672184959053993, + 0.07192272692918777 + ], + "credit_map": [ + 0.07707478106021881, + 0.02516130730509758, + 0.039287421852350235, + 0.025386160239577293, + 0.02479972504079342, + 0.03204161673784256, + 0.02603764459490776, + 0.06671395152807236, + 0.014093362726271152, + 0.009314634837210178, + 0.01914680376648903, + 0.007811515126377344, + 0.009852904826402664, + 0.007576958741992712, + 0.012141130864620209, + 0.009644298814237118, + 0.0042059654369950294, + 0.0029058903455734253, + 0.005964951124042273, + 0.005273858550935984, + 0.00404246523976326, + 0.007267683278769255, + 0.004650522023439407, + 0.004807351157069206, + 0.015045040287077427, + 0.005599202588200569, + 0.025100165978074074, + 0.00924626924097538, + 0.005745346192270517, + 0.02132864110171795, + 0.019530039280653, + 0.025058867409825325, + 0.00549201900139451, + 0.002177257090806961, + 0.006328717805445194, + 0.003000442637130618, + 0.0026137453969568014, + 0.004272959195077419, + 0.005770043004304171, + 0.010012191720306873, + 0.007927453145384789, + 0.0052577597089111805, + 0.016149651259183884, + 0.007346840109676123, + 0.005176187958568335, + 0.017139114439487457, + 0.013890692964196205, + 0.014077489264309406, + 0.006201551295816898, + 0.005052326712757349, + 0.01595127210021019, + 0.005166900344192982, + 0.00676703080534935, + 0.015050987713038921, + 0.008046521805226803, + 0.008840065449476242, + 0.018600530922412872, + 0.006935769692063332, + 0.017340991646051407, + 0.0052970293909311295, + 0.0074754562228918076, + 0.013740427792072296, + 0.010197822004556656, + 0.03220178186893463, + 0.008562727831304073, + 0.004037114791572094, + 0.009880534373223782, + 0.004042238928377628, + 0.0043202731758356094, + 0.0035683002788573503, + 0.006544451229274273, + 0.005766208283603191, + 0.010858959518373013, + 0.007918731309473515, + 0.012619503773748875, + 0.004443588200956583, + 0.004868188872933388, + 0.0057618203572928905, + 0.012722899205982685, + 0.01272903848439455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.058444391936063766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983187697827816 + ], + "opposite_logprob_tokens": [ + -3.529831886291504 + ], + "logprob_reward": -0.02983187697827816, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983187697827816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30280131101608276, + 0.11021222174167633, + 0.07896909862756729, + 0.05048249289393425, + 0.15645140409469604, + 0.08313233405351639, + 0.030152766034007072, + 0.07019665092229843, + 0.055778659880161285, + 0.0618230402469635 + ], + "credit_map": [ + 0.10943441838026047, + 0.015648776665329933, + 0.019548814743757248, + 0.01222524605691433, + 0.013668126426637173, + 0.024771543219685555, + 0.01483366172760725, + 0.0926707461476326, + 0.016446741297841072, + 0.00612616678699851, + 0.02295217290520668, + 0.012284447439014912, + 0.010001454502344131, + 0.010956751182675362, + 0.019737692549824715, + 0.011706791818141937, + 0.011152984574437141, + 0.00434991205111146, + 0.012424941174685955, + 0.005274456460028887, + 0.00464903749525547, + 0.024065807461738586, + 0.00807814672589302, + 0.00897381454706192, + 0.005437261890619993, + 0.0024142302572727203, + 0.010549412108957767, + 0.002682223916053772, + 0.003972862847149372, + 0.01232299953699112, + 0.00563126290217042, + 0.007472238503396511, + 0.017771434038877487, + 0.008786649443209171, + 0.024256553500890732, + 0.01387198455631733, + 0.01013126689940691, + 0.01938064768910408, + 0.02716979756951332, + 0.035083066672086716, + 0.007945251651108265, + 0.004010203294456005, + 0.015058073215186596, + 0.008650548756122589, + 0.004449489992111921, + 0.013965483754873276, + 0.012782271020114422, + 0.016271013766527176, + 0.0034849639050662518, + 0.0021108018700033426, + 0.006129139568656683, + 0.002210620092228055, + 0.0022361872252076864, + 0.005456001963466406, + 0.0032020765356719494, + 0.005322975106537342, + 0.009235710836946964, + 0.006994619965553284, + 0.009773876518011093, + 0.0070459577254951, + 0.007319429889321327, + 0.007505988236516714, + 0.012470569461584091, + 0.009850498288869858, + 0.0071260277181863785, + 0.004817152861505747, + 0.007987180724740028, + 0.005824056454002857, + 0.005595950409770012, + 0.006172921974211931, + 0.009915761649608612, + 0.008339609950780869, + 0.00808501336723566, + 0.005884782876819372, + 0.010696297511458397, + 0.005052336025983095, + 0.004145240876823664, + 0.006563730072230101, + 0.009564842097461224, + 0.011830795556306839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02983187697827816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693719655275345 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.033693719655275345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693719655275345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14187172055244446, + 0.12043498456478119, + 0.07565648853778839, + 0.17964155972003937, + 0.05758943036198616, + 0.08607029914855957, + 0.1342361867427826, + 0.12735454738140106, + 0.0447082482278347, + 0.032436538487672806 + ], + "credit_map": [ + 0.032047584652900696, + 0.013323219493031502, + 0.01592094823718071, + 0.013028539717197418, + 0.013730639591813087, + 0.014831791631877422, + 0.014850052073597908, + 0.024138938635587692, + 0.02075020968914032, + 0.017037170007824898, + 0.02495693787932396, + 0.009022679179906845, + 0.013319405727088451, + 0.010872211307287216, + 0.011088907718658447, + 0.013387471437454224, + 0.008237694390118122, + 0.004658185876905918, + 0.012029150500893593, + 0.0063008880242705345, + 0.0073165069334208965, + 0.018233394250273705, + 0.01045134011656046, + 0.008429327048361301, + 0.021076321601867676, + 0.00631865207105875, + 0.035560257732868195, + 0.012284045107662678, + 0.007154533173888922, + 0.05298672243952751, + 0.0140847684815526, + 0.030176270753145218, + 0.009148810990154743, + 0.003406776115298271, + 0.008439131081104279, + 0.004331372212618589, + 0.004411469679325819, + 0.006791915278881788, + 0.009984898380935192, + 0.011075057089328766, + 0.008340129628777504, + 0.003747044363990426, + 0.01899235136806965, + 0.007811196148395538, + 0.003933298401534557, + 0.016024084761738777, + 0.010330547578632832, + 0.016891639679670334, + 0.014178629964590073, + 0.009741691872477531, + 0.030086509883403778, + 0.01090890634804964, + 0.011431886814534664, + 0.02180504985153675, + 0.013170053251087666, + 0.022913455963134766, + 0.022635862231254578, + 0.006335660349577665, + 0.019710201770067215, + 0.006053359247744083, + 0.007976213470101357, + 0.015433024615049362, + 0.009502720087766647, + 0.039707500487565994, + 0.005782333202660084, + 0.004696368705481291, + 0.006877306383103132, + 0.005428716540336609, + 0.004156702198088169, + 0.004080199170857668, + 0.00784155074506998, + 0.0058450717478990555, + 0.004939481150358915, + 0.004631670657545328, + 0.00580259133130312, + 0.002303509972989559, + 0.0023892216850072145, + 0.002920265309512615, + 0.0046700178645551205, + 0.004779782146215439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.033693719655275345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762575538828969 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.00762575538828969, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762575538828969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28024575114250183, + 0.06587446480989456, + 0.0771779865026474, + 0.15745970606803894, + 0.08578793704509735, + 0.1109648272395134, + 0.04637596756219864, + 0.12549832463264465, + 0.01684783585369587, + 0.03376713767647743 + ], + "credit_map": [ + 0.10105369240045547, + 0.009973809123039246, + 0.0240229070186615, + 0.014090105891227722, + 0.013140366412699223, + 0.02243686281144619, + 0.013042468577623367, + 0.08248554915189743, + 0.009091883897781372, + 0.004635465331375599, + 0.016760246828198433, + 0.00663806963711977, + 0.005475819576531649, + 0.00822294782847166, + 0.008718917146325111, + 0.006331109907478094, + 0.011833465658128262, + 0.0031919614411890507, + 0.012118064798414707, + 0.0067108492366969585, + 0.0046872422099113464, + 0.020435458049178123, + 0.009562413208186626, + 0.008638529106974602, + 0.016141094267368317, + 0.006192795932292938, + 0.03821056708693504, + 0.015218229033052921, + 0.005699882749468088, + 0.04978906363248825, + 0.009447840042412281, + 0.016760241240262985, + 0.013116663321852684, + 0.0033184930216521025, + 0.01262718252837658, + 0.005059985909610987, + 0.0049713351763784885, + 0.010802612639963627, + 0.009579477831721306, + 0.02631218358874321, + 0.011055120266973972, + 0.006036666221916676, + 0.02045946940779686, + 0.012044019065797329, + 0.005976504646241665, + 0.019226087257266045, + 0.013024545274674892, + 0.023142419755458832, + 0.004021796863526106, + 0.0025791116058826447, + 0.010730236768722534, + 0.0027070867363363504, + 0.0038593546487390995, + 0.010205132886767387, + 0.0057390909641981125, + 0.006534155458211899, + 0.022472582757472992, + 0.007225849665701389, + 0.019298573955893517, + 0.005317711271345615, + 0.006250637583434582, + 0.01616082340478897, + 0.009836006909608841, + 0.038936130702495575, + 0.0029289610683918, + 0.001553142094053328, + 0.0027102224994450808, + 0.0018326283898204565, + 0.001799162244424224, + 0.001571840955875814, + 0.002026224974542856, + 0.0024256552569568157, + 0.005938216578215361, + 0.006442431826144457, + 0.005384096410125494, + 0.003453264245763421, + 0.002156158909201622, + 0.002825547708198428, + 0.003529437817633152, + 0.004037986975163221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.00762575538828969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864989593625069 + ], + "opposite_logprob_tokens": [ + -3.048649787902832 + ], + "logprob_reward": -0.04864989593625069, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864989593625069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20936943590641022, + 0.10190903395414352, + 0.03312156721949577, + 0.14883965253829956, + 0.15593819320201874, + 0.04327880218625069, + 0.023878030478954315, + 0.20296990871429443, + 0.013901001773774624, + 0.06679445505142212 + ], + "credit_map": [ + 0.06635264307260513, + 0.008841694332659245, + 0.02993089333176613, + 0.010814985260367393, + 0.010586677119135857, + 0.018990302458405495, + 0.011286141350865364, + 0.052566077560186386, + 0.014432105235755444, + 0.006237145513296127, + 0.03102000057697296, + 0.006409924011677504, + 0.00788908265531063, + 0.016085274517536163, + 0.007623034063726664, + 0.012212454341351986, + 0.00576702831313014, + 0.0016685706796124578, + 0.00529084587469697, + 0.0030016738455742598, + 0.0019648862071335316, + 0.008734386414289474, + 0.003086572978645563, + 0.0036075981333851814, + 0.014765172265470028, + 0.005614944268018007, + 0.03202178329229355, + 0.018676355481147766, + 0.0062236664816737175, + 0.03341599553823471, + 0.0167594775557518, + 0.02136225439608097, + 0.03148671239614487, + 0.008871571160852909, + 0.026986004784703255, + 0.009834034368395805, + 0.007827535271644592, + 0.01875830814242363, + 0.018900170922279358, + 0.033273845911026, + 0.0044162399135529995, + 0.0021864844020456076, + 0.008895596489310265, + 0.003097366541624069, + 0.0021019333507865667, + 0.008128102868795395, + 0.004477380774915218, + 0.009975694119930267, + 0.0023775119334459305, + 0.0012765750288963318, + 0.00555815827101469, + 0.0013074256712570786, + 0.0016901325434446335, + 0.004946528933942318, + 0.002138708485290408, + 0.004582987632602453, + 0.03312408924102783, + 0.01187916286289692, + 0.03215362876653671, + 0.007976643741130829, + 0.010370556265115738, + 0.02687334455549717, + 0.01643977127969265, + 0.06415271759033203, + 0.002777679357677698, + 0.0011099865660071373, + 0.0023757449816912413, + 0.0013257922837510705, + 0.0011591532966122031, + 0.0015114943962544203, + 0.0015704490942880511, + 0.0020707007497549057, + 0.012828867882490158, + 0.012884999625384808, + 0.012380409985780716, + 0.004447196610271931, + 0.003400401445105672, + 0.005549142137169838, + 0.0077650416642427444, + 0.007538389414548874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.04864989593625069 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera slowly and steadily tilts upward, following a pink lotus bud about to bloom. With a lifelike, slow, and resolute gesture, it pierces the nearly opaque surface of deep, dark green water and rises. The bud is compact, spindle-shaped, and its outer petals display a delicate gradation from soft ivory white at the base to delicate pink at the tip. Coated with a nearly invisible waxy layer, the surface allows the crystal-clear water droplets to tumble down the graceful curves of the petals like tiny diamonds, ultimately dripping back into the murky water below. A soft yet focused overhead light casts precisely on the lotus, making it the sole focal point against the bottomless, pure darkness. The light not only outlines the full contours of the petals but also reflects a dazzling radiance from each falling droplet. The entire image is imbued with a sense of Zen and the power of life, conveying a profound atmosphere of purity, resilience, and tranquility born from darkness.", + "seed": 28226, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_11/index_7042_seed_28226_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027788232453167437, + "vlm_logprob": -0.027788232453167437, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.027788232453167437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04952678829431534 + ], + "opposite_logprob_tokens": [ + -3.7995266914367676 + ], + "logprob_reward": -0.04952678829431534, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04952678829431534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12207649648189545, + 0.04223250970244408, + 0.04035978019237518, + 0.06636179238557816, + 0.04958995059132576, + 0.06362853944301605, + 0.1810085028409958, + 0.1053798496723175, + 0.1934722512960434, + 0.13589034974575043 + ], + "credit_map": [ + 0.016654081642627716, + 0.005833607632666826, + 0.032829105854034424, + 0.006453432142734528, + 0.004415616393089294, + 0.03625861555337906, + 0.008891168050467968, + 0.010740868747234344, + 0.005132769700139761, + 0.003244229592382908, + 0.007447962183505297, + 0.0025327028706669807, + 0.0028964478988200426, + 0.009308961220085621, + 0.004677894525229931, + 0.006991541478782892, + 0.005313396919518709, + 0.0030889753252267838, + 0.008923972956836224, + 0.0030530784279108047, + 0.0032536170911043882, + 0.00499059958383441, + 0.0058927289210259914, + 0.0058434102684259415, + 0.01335210632532835, + 0.005376577842980623, + 0.011857188306748867, + 0.004373272880911827, + 0.0052223969250917435, + 0.009480627253651619, + 0.0060927653685212135, + 0.010606855154037476, + 0.006754051893949509, + 0.005897579248994589, + 0.007965617813169956, + 0.005873081739991903, + 0.005041651427745819, + 0.004354602657258511, + 0.005661074072122574, + 0.008042292669415474, + 0.009981727227568626, + 0.00836569257080555, + 0.00870412029325962, + 0.006582917179912329, + 0.008254715241491795, + 0.006117149256169796, + 0.007745715323835611, + 0.007876497693359852, + 0.03727353736758232, + 0.011860376223921776, + 0.04835492745041847, + 0.017403941601514816, + 0.01602594368159771, + 0.013571442104876041, + 0.014424343593418598, + 0.022093988955020905, + 0.017892079427838326, + 0.008191453292965889, + 0.0190594345331192, + 0.008787641301751137, + 0.012814432382583618, + 0.018637854605913162, + 0.009841855615377426, + 0.010155091993510723, + 0.023916829377412796, + 0.01902075670659542, + 0.028785280883312225, + 0.02074228599667549, + 0.01895981840789318, + 0.020675627514719963, + 0.040251247584819794, + 0.02112039551138878, + 0.006847891490906477, + 0.006905082613229752, + 0.005113131832331419, + 0.005560425575822592, + 0.055490218102931976, + 0.04203947260975838, + 0.006323609035462141, + 0.007610517553985119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.04952678829431534 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016091125085949898 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.016091125085949898, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016091125085949898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0881919264793396, + 0.06566968560218811, + 0.03194058686494827, + 0.03328634798526764, + 0.026802344247698784, + 0.03680989518761635, + 0.2642681300640106, + 0.1175178661942482, + 0.3130471110343933, + 0.022466139867901802 + ], + "credit_map": [ + 0.010246969759464264, + 0.004498397000133991, + 0.026386205106973648, + 0.0056953225284814835, + 0.0034944138024002314, + 0.021952614188194275, + 0.007777208928018808, + 0.008140794932842255, + 0.008317921310663223, + 0.006110372021794319, + 0.009799107909202576, + 0.006239545997232199, + 0.004871138837188482, + 0.012527459301054478, + 0.008500845171511173, + 0.00930329505354166, + 0.004684084560722113, + 0.004220515955239534, + 0.0063990456983447075, + 0.002261519432067871, + 0.0034810483921319246, + 0.0028818300925195217, + 0.0026167461182922125, + 0.00539579801261425, + 0.005225189495831728, + 0.004859743174165487, + 0.003945653326809406, + 0.0026402827352285385, + 0.005109779536724091, + 0.003392871469259262, + 0.004321397747844458, + 0.0037914325948804617, + 0.003801220329478383, + 0.003386087715625763, + 0.003947149496525526, + 0.0032286106143146753, + 0.002742405515164137, + 0.00197361852042377, + 0.00407926831394434, + 0.0036439825780689716, + 0.006007918156683445, + 0.005280470475554466, + 0.005426271352916956, + 0.0035655603278428316, + 0.004149779677391052, + 0.0026313967537134886, + 0.004835626110434532, + 0.004912872798740864, + 0.05230754613876343, + 0.017292747274041176, + 0.06488213688135147, + 0.025495629757642746, + 0.03237795829772949, + 0.02184920944273472, + 0.020418405532836914, + 0.02964450605213642, + 0.01800857111811638, + 0.009081030264496803, + 0.02020053006708622, + 0.010245799086987972, + 0.01606660895049572, + 0.01606304757297039, + 0.01426813006401062, + 0.013584143482148647, + 0.040070343762636185, + 0.030208846554160118, + 0.049418799579143524, + 0.03311731293797493, + 0.030543044209480286, + 0.03383190557360649, + 0.0624057911336422, + 0.03345106914639473, + 0.0016064532101154327, + 0.0019988215062767267, + 0.0018374925712123513, + 0.0012895036488771439, + 0.005406328476965427, + 0.006632883567363024, + 0.0022159148938953876, + 0.0014787415275350213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.016091125085949898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16768114268779755, + 0.040136370807886124, + 0.09819779545068741, + 0.12297148257493973, + 0.04632999002933502, + 0.028592174872756004, + 0.05175295099616051, + 0.05592458322644234, + 0.06420925259590149, + 0.3242041766643524 + ], + "credit_map": [ + 0.019990693777799606, + 0.006833880208432674, + 0.06174325942993164, + 0.006378005724400282, + 0.006097110453993082, + 0.038847509771585464, + 0.008715283125638962, + 0.019075408577919006, + 0.004708827938884497, + 0.005116557702422142, + 0.005498664453625679, + 0.00317283789627254, + 0.004447673913091421, + 0.005754035897552967, + 0.005288021173328161, + 0.006149753462523222, + 0.010116402991116047, + 0.008793825283646584, + 0.02154676243662834, + 0.007653241977095604, + 0.008861626498401165, + 0.013873196206986904, + 0.008390259929001331, + 0.01896248757839203, + 0.023344213142991066, + 0.010920165106654167, + 0.020162085071206093, + 0.008942481130361557, + 0.01148197427392006, + 0.014935147017240524, + 0.014509670436382294, + 0.018675759434700012, + 0.006244173273444176, + 0.005914425943046808, + 0.006199411116540432, + 0.005108054727315903, + 0.0053147319704294205, + 0.003110891208052635, + 0.007945746183395386, + 0.006492560729384422, + 0.004522437229752541, + 0.0033641932532191277, + 0.0037540090270340443, + 0.004153004381805658, + 0.0030008864123374224, + 0.0022578625939786434, + 0.0036408216692507267, + 0.003898962866514921, + 0.008283940143883228, + 0.003203673753887415, + 0.009114270098507404, + 0.004581316374242306, + 0.009264889173209667, + 0.008055747486650944, + 0.004150764085352421, + 0.005098351743072271, + 0.00702812522649765, + 0.003660069080069661, + 0.009677872061729431, + 0.0036381129175424576, + 0.008978667668998241, + 0.013469230383634567, + 0.004273067694157362, + 0.0051994421519339085, + 0.005298295523971319, + 0.004851701203733683, + 0.00807364471256733, + 0.007472777273505926, + 0.007849588058888912, + 0.006308822426944971, + 0.01614706590771675, + 0.008207368664443493, + 0.011418556794524193, + 0.012930289842188358, + 0.008503624238073826, + 0.008914630860090256, + 0.13685420155525208, + 0.120692677795887, + 0.010137655772268772, + 0.01475254725664854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.014192949049174786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306134581565857 + ], + "opposite_logprob_tokens": [ + -3.1680612564086914 + ], + "logprob_reward": -0.04306134581565857, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306134581565857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39836978912353516, + 0.18161603808403015, + 0.05337067320942879, + 0.0492180734872818, + 0.020540639758110046, + 0.030972378328442574, + 0.047798704355955124, + 0.04068832844495773, + 0.13215596973896027, + 0.0452694408595562 + ], + "credit_map": [ + 0.04636447876691818, + 0.010933789424598217, + 0.14891749620437622, + 0.01014301273971796, + 0.010102900676429272, + 0.1252279281616211, + 0.013818944804370403, + 0.03286121040582657, + 0.024760445579886436, + 0.014634198509156704, + 0.02736334316432476, + 0.009902330115437508, + 0.015994105488061905, + 0.03764435276389122, + 0.017645176500082016, + 0.03367207944393158, + 0.005627060774713755, + 0.0028119944036006927, + 0.011357035487890244, + 0.005236722994595766, + 0.0032820366322994232, + 0.007874669507145882, + 0.0073339673690497875, + 0.009847176261246204, + 0.008223416283726692, + 0.0048208958469331264, + 0.007330453023314476, + 0.0039986432529985905, + 0.004857024643570185, + 0.0059778159484267235, + 0.005448054987937212, + 0.008561767637729645, + 0.0032171562779694796, + 0.0027052818331867456, + 0.003397956257686019, + 0.0024330555461347103, + 0.002006881171837449, + 0.001622432959266007, + 0.0024242838844656944, + 0.0027335912454873323, + 0.005133980419486761, + 0.004360314458608627, + 0.004681413061916828, + 0.0033202366903424263, + 0.003713886486366391, + 0.002222721930593252, + 0.0037169845309108496, + 0.0038228367920964956, + 0.008504794910550117, + 0.003969245590269566, + 0.008341426961123943, + 0.004237911198288202, + 0.006792376283556223, + 0.005628662183880806, + 0.003924619872123003, + 0.006399661768227816, + 0.005941859446465969, + 0.002667711116373539, + 0.007747552823275328, + 0.003295107977464795, + 0.0053524356335401535, + 0.006492690648883581, + 0.004172301851212978, + 0.0050186640582978725, + 0.016223162412643433, + 0.01260930486023426, + 0.019028058275580406, + 0.01396177988499403, + 0.013896419666707516, + 0.015012661926448345, + 0.027420273050665855, + 0.014004302211105824, + 0.0017895445926114917, + 0.0023199873976409435, + 0.0018719907384365797, + 0.0014153459342196584, + 0.017225828021764755, + 0.017065713182091713, + 0.0017109235050156713, + 0.001870106440037489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.04306134581565857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160689540207386 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160689540207386, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160689540207386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2214105874300003, + 0.08171522617340088, + 0.041283607482910156, + 0.017985066398978233, + 0.033846937119960785, + 0.028572509065270424, + 0.2354329377412796, + 0.12389755994081497, + 0.02710532583296299, + 0.1887502372264862 + ], + "credit_map": [ + 0.028085893020033836, + 0.007890090346336365, + 0.03958883881568909, + 0.00921153649687767, + 0.009451467543840408, + 0.07855232059955597, + 0.015270480886101723, + 0.03335994854569435, + 0.011035195551812649, + 0.005423981696367264, + 0.015073033981025219, + 0.003470025025308132, + 0.0038032697048038244, + 0.01793045550584793, + 0.007827536202967167, + 0.0171517226845026, + 0.005503060296177864, + 0.0030336726922541857, + 0.006749866064637899, + 0.004668767563998699, + 0.003271280089393258, + 0.006824671756476164, + 0.004100421443581581, + 0.007131861988455057, + 0.002623197855427861, + 0.0025842878967523575, + 0.0026800858322530985, + 0.001219483558088541, + 0.0020006385166198015, + 0.0024268117267638445, + 0.0017311624251306057, + 0.002719396958127618, + 0.004964836407452822, + 0.0035765739157795906, + 0.004852894693613052, + 0.003964276053011417, + 0.0028328283224254847, + 0.0030237268656492233, + 0.004473178647458553, + 0.00615861639380455, + 0.0051819863729178905, + 0.0025756005197763443, + 0.004556580912321806, + 0.00321844220161438, + 0.0027806381694972515, + 0.002655486110597849, + 0.003312043845653534, + 0.0042917258106172085, + 0.045564983040094376, + 0.016852345317602158, + 0.06009209528565407, + 0.023483561351895332, + 0.024599744006991386, + 0.017429102212190628, + 0.019031338393688202, + 0.028379766270518303, + 0.02087651751935482, + 0.005196217447519302, + 0.039656415581703186, + 0.007488403003662825, + 0.012639804743230343, + 0.019919849932193756, + 0.005540423560887575, + 0.012579911388456821, + 0.002914111129939556, + 0.0024350672028958797, + 0.0033703320659697056, + 0.003253401955589652, + 0.0035031414590775967, + 0.003733753226697445, + 0.004771709907799959, + 0.003123807953670621, + 0.008019124157726765, + 0.007910283282399178, + 0.004936514422297478, + 0.006529612001031637, + 0.08035141974687576, + 0.06437298655509949, + 0.006169355474412441, + 0.010460927151143551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.0160689540207386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera slowly and steadily tilts upward, following a pink lotus bud about to bloom. With a lifelike, slow, and resolute gesture, it pierces the nearly opaque surface of deep, dark green water and rises. The bud is compact, spindle-shaped, and its outer petals display a delicate gradation from soft ivory white at the base to delicate pink at the tip. Coated with a nearly invisible waxy layer, the surface allows the crystal-clear water droplets to tumble down the graceful curves of the petals like tiny diamonds, ultimately dripping back into the murky water below. A soft yet focused overhead light casts precisely on the lotus, making it the sole focal point against the bottomless, pure darkness. The light not only outlines the full contours of the petals but also reflects a dazzling radiance from each falling droplet. The entire image is imbued with a sense of Zen and the power of life, conveying a profound atmosphere of purity, resilience, and tranquility born from darkness.", + "seed": 28227, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_11/index_7042_seed_28227_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1670391395688057, + "vlm_logprob": -0.1670391395688057, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.1670391395688057, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11785286664962769 + ], + "opposite_logprob_tokens": [ + -2.4928529262542725 + ], + "logprob_reward": -0.11785286664962769, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11785286664962769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19561532139778137, + 0.5796241164207458, + 0.030304329469799995, + 0.051922980695962906, + 0.039246413856744766, + 0.011821245774626732, + 0.06602936238050461, + 0.011766590178012848, + 0.007921155542135239, + 0.005748533643782139 + ], + "credit_map": [ + 0.043289732187986374, + 0.016161784529685974, + 0.028307843953371048, + 0.00968033168464899, + 0.017118383198976517, + 0.019023356959223747, + 0.014954577200114727, + 0.04707931727170944, + 0.07561566680669785, + 0.022569408640265465, + 0.16058503091335297, + 0.046350713819265366, + 0.03162907063961029, + 0.10004331916570663, + 0.039703503251075745, + 0.1031273752450943, + 0.002850410994142294, + 0.0038104387931525707, + 0.003836537478491664, + 0.0036957485135644674, + 0.0027072764933109283, + 0.005223121028393507, + 0.004108380991965532, + 0.004072417505085468, + 0.005406443495303392, + 0.00612405501306057, + 0.007507174275815487, + 0.006367482710629702, + 0.004282528068870306, + 0.005262032151222229, + 0.008417260833084583, + 0.008556008338928223, + 0.0031478109303861856, + 0.0034503706265240908, + 0.003910761326551437, + 0.0033059457782655954, + 0.0038162516430020332, + 0.006241834256798029, + 0.008964449167251587, + 0.006408989895135164, + 0.0010174672352150083, + 0.0014840398216620088, + 0.0011240432504564524, + 0.0019698392134159803, + 0.0021076612174510956, + 0.0009341391269117594, + 0.002286288421601057, + 0.0008977691759355366, + 0.0026413300074636936, + 0.0026741353794932365, + 0.009737763553857803, + 0.003470492549240589, + 0.0023951022885739803, + 0.024909894913434982, + 0.003570560133084655, + 0.016630088910460472, + 0.0014558059629052877, + 0.0008869442390277982, + 0.0019372079987078905, + 0.0012331852922216058, + 0.0010324612958356738, + 0.00212520151399076, + 0.0009496906423009932, + 0.002146093873307109, + 0.0011272167321294546, + 0.0006332676275633276, + 0.0010937172919511795, + 0.001002512639388442, + 0.0008076532394625247, + 0.0011208101641386747, + 0.0009781323606148362, + 0.001157846418209374, + 0.0009497713181190193, + 0.0004311569791752845, + 0.0008450837340205908, + 0.000544618524145335, + 0.0003693151520565152, + 0.0010809425730258226, + 0.0005551238427869976, + 0.0009725219570100307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.11785286664962769 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020629718899726868 + ], + "opposite_logprob_tokens": [ + -3.895629644393921 + ], + "logprob_reward": -0.020629718899726868, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020629718899726868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058108288794755936, + 0.4467945098876953, + 0.08303292840719223, + 0.15209358930587769, + 0.043587882071733475, + 0.036718737334012985, + 0.11261595785617828, + 0.02612420730292797, + 0.03258811682462692, + 0.008335758000612259 + ], + "credit_map": [ + 0.008803628385066986, + 0.003477838821709156, + 0.008342250250279903, + 0.0033819794189184904, + 0.0063077835366129875, + 0.010810776613652706, + 0.004562521353363991, + 0.012421508319675922, + 0.057680919766426086, + 0.018864857032895088, + 0.12268929928541183, + 0.03726698458194733, + 0.02493932470679283, + 0.07625196874141693, + 0.0303183626383543, + 0.07878279685974121, + 0.008024671114981174, + 0.009347486309707165, + 0.010933824814856052, + 0.010836726054549217, + 0.008610195480287075, + 0.009296238422393799, + 0.012507079169154167, + 0.01347670704126358, + 0.012564004398882389, + 0.019004754722118378, + 0.02508055791258812, + 0.016655515879392624, + 0.01757722906768322, + 0.01586747169494629, + 0.02073029987514019, + 0.02461376041173935, + 0.0028106146492064, + 0.0029993592761456966, + 0.004482207354158163, + 0.007456487510353327, + 0.004206349141895771, + 0.004173120483756065, + 0.010915042832493782, + 0.0065447003580629826, + 0.003312879242002964, + 0.00401719007641077, + 0.003785486565902829, + 0.0045495112426579, + 0.005562092177569866, + 0.0047866711392998695, + 0.00790781993418932, + 0.00279708718881011, + 0.00992439966648817, + 0.00689250323921442, + 0.026346445083618164, + 0.01310538500547409, + 0.006375117693096399, + 0.013336499221622944, + 0.013332403264939785, + 0.023303208872675896, + 0.0034075824078172445, + 0.001871804241091013, + 0.005139535292983055, + 0.0026749023236334324, + 0.0020584322046488523, + 0.004560297355055809, + 0.002529775956645608, + 0.0038818770553916693, + 0.0036329273134469986, + 0.001787079032510519, + 0.009692225605249405, + 0.0024724819231778383, + 0.001451070886105299, + 0.007799120154231787, + 0.0023921597748994827, + 0.0033610505051910877, + 0.0006756290094926953, + 0.0009970450773835182, + 0.0010913880541920662, + 0.001096375286579132, + 0.0004688457993324846, + 0.0013145121047273278, + 0.0009696622146293521, + 0.0017223006580024958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.020629718899726868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204929202795029 + ], + "opposite_logprob_tokens": [ + -2.812049388885498 + ], + "logprob_reward": -0.06204929202795029, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204929202795029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2100445181131363, + 0.3242669999599457, + 0.26531538367271423, + 0.04139119014143944, + 0.05317576602101326, + 0.01926136575639248, + 0.04588101804256439, + 0.01232485193759203, + 0.0169591773301363, + 0.01137969084084034 + ], + "credit_map": [ + 0.045683588832616806, + 0.016976460814476013, + 0.03152576833963394, + 0.010769261978566647, + 0.016605623066425323, + 0.02660256065428257, + 0.01681206375360489, + 0.04506919905543327, + 0.0420522578060627, + 0.01280250959098339, + 0.08976385742425919, + 0.02740946225821972, + 0.01768285408616066, + 0.056350190192461014, + 0.021512411534786224, + 0.05669345334172249, + 0.028691548854112625, + 0.03567543253302574, + 0.03329579532146454, + 0.02689943090081215, + 0.028471071273088455, + 0.03754769638180733, + 0.0353795550763607, + 0.03935486078262329, + 0.004966842941939831, + 0.006788982544094324, + 0.00462335254997015, + 0.003907679580152035, + 0.006838537752628326, + 0.003618928138166666, + 0.004211382009088993, + 0.0064354850910604, + 0.003997196909040213, + 0.003857256844639778, + 0.00394138228148222, + 0.005394088104367256, + 0.004321760963648558, + 0.005886122118681669, + 0.015819890424609184, + 0.009958069771528244, + 0.0017118561081588268, + 0.002240826841443777, + 0.002212671795859933, + 0.002286147326231003, + 0.0021633810829371214, + 0.0036158226430416107, + 0.003112457925453782, + 0.0019182016840204597, + 0.003931982908397913, + 0.0032298965379595757, + 0.007341037038713694, + 0.005981040187180042, + 0.003330558305606246, + 0.0101093128323555, + 0.006040552165359259, + 0.005916643887758255, + 0.0008281425689347088, + 0.0013150812592357397, + 0.0026442327070981264, + 0.0014368828851729631, + 0.0009804303990677, + 0.002093595452606678, + 0.0016220854595303535, + 0.001404401147738099, + 0.0012142980704084039, + 0.0009029264329001307, + 0.0045972298830747604, + 0.0017743301577866077, + 0.0007571999449282885, + 0.003968036267906427, + 0.0014834018656983972, + 0.0022617545910179615, + 0.002457451308146119, + 0.0009783526184037328, + 0.0012491345405578613, + 0.001250239321961999, + 0.0009761243127286434, + 0.0013977730413898826, + 0.001224614679813385, + 0.0018460005521774292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.06204929202795029 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038739681243896 + ], + "opposite_logprob_tokens": [ + -1.910387396812439 + ], + "logprob_reward": -0.16038739681243896, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038739681243896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16334006190299988, + 0.541561484336853, + 0.035782769322395325, + 0.011466119438409805, + 0.050540078431367874, + 0.0055240606889128685, + 0.17488810420036316, + 0.0059693376533687115, + 0.008089948445558548, + 0.0028380188159644604 + ], + "credit_map": [ + 0.03617987409234047, + 0.013600186444818974, + 0.024965079501271248, + 0.008089345879852772, + 0.01120302826166153, + 0.019543344154953957, + 0.012749395333230495, + 0.03700981289148331, + 0.07006677240133286, + 0.02045280858874321, + 0.150807723402977, + 0.04306511953473091, + 0.02937774546444416, + 0.09460799396038055, + 0.03675110265612602, + 0.09643220156431198, + 0.003248263383284211, + 0.0038262957241386175, + 0.005568415392190218, + 0.004836029838770628, + 0.0033566246274858713, + 0.0028199488297104836, + 0.006323713343590498, + 0.005803478416055441, + 0.001255079172551632, + 0.001026948681101203, + 0.0022975034080445766, + 0.0010453949216753244, + 0.0009127127123065293, + 0.0015468571800738573, + 0.0013406788930296898, + 0.0020409440621733665, + 0.003701776033267379, + 0.0039992439560592175, + 0.004116775933653116, + 0.0035821448545902967, + 0.004691933281719685, + 0.008049052208662033, + 0.013825508765876293, + 0.008573643863201141, + 0.000484298390801996, + 0.000576057645957917, + 0.0005758059560321271, + 0.0007030683918856084, + 0.0006272273603826761, + 0.0009865903994068503, + 0.0012036465341225266, + 0.0003673657774925232, + 0.0074490755796432495, + 0.005792989861220121, + 0.03762201964855194, + 0.006872666534036398, + 0.00661788834258914, + 0.06741786748170853, + 0.010379211977124214, + 0.032736390829086304, + 0.0003472974640317261, + 0.0002996901166625321, + 0.0008401416125707328, + 0.0003337398520670831, + 0.00031165758264251053, + 0.001663188566453755, + 0.0004118691140320152, + 0.0017617539269849658, + 0.0004399414174258709, + 0.0002632412360981107, + 0.002790614729747176, + 0.0007228132453747094, + 0.00034999215858988464, + 0.0020363503135740757, + 0.00048799155047163367, + 0.0009990037651732564, + 0.0002752458385657519, + 0.00026192329823970795, + 0.0003350706538185477, + 0.000324360269587487, + 0.00021563642076216638, + 0.00045750057324767113, + 0.00030483101727440953, + 0.0006634507444687188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.16038739681243896 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward while following a pink lotus bud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus bud spindle-shaped with petals graded from soft ivory white at the base to delicate pink at the tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bud pierce the nearly opaque surface of deep, dark green water and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft yet focused overhead light that illuminates the lotus against a bottomless, pure darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47427642345428467 + ], + "opposite_logprob_tokens": [ + -0.9742764234542847 + ], + "logprob_reward": -0.47427642345428467, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47427642345428467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19641953706741333, + 0.5099214911460876, + 0.06313224881887436, + 0.031238798052072525, + 0.06738246977329254, + 0.006959361024200916, + 0.11837347596883774, + 0.0024554836563766003, + 0.0012464242754504085, + 0.0028707103338092566 + ], + "credit_map": [ + 0.0434175468981266, + 0.015784716233611107, + 0.0282023586332798, + 0.009988811798393726, + 0.017338864505290985, + 0.01896749436855316, + 0.015703333541750908, + 0.04701639339327812, + 0.06660538911819458, + 0.01945592276751995, + 0.13850152492523193, + 0.041008900851011276, + 0.028007248416543007, + 0.0888645350933075, + 0.03761192783713341, + 0.0898660197854042, + 0.0059469533152878284, + 0.007443098351359367, + 0.009903701953589916, + 0.008208389393985271, + 0.006297179497778416, + 0.005895647685974836, + 0.010019630193710327, + 0.009417642839252949, + 0.0035082276444882154, + 0.0022257522214204073, + 0.007016914896667004, + 0.0028849299997091293, + 0.00253696390427649, + 0.00402276823297143, + 0.004318021237850189, + 0.004725216887891293, + 0.005824102554470301, + 0.004227179568260908, + 0.006159848067909479, + 0.009026355110108852, + 0.006520797032862902, + 0.010084595531225204, + 0.015481270849704742, + 0.010058321990072727, + 0.0007648623432032764, + 0.0007326213526539505, + 0.0010287798941135406, + 0.0007304179598577321, + 0.000641954189632088, + 0.0015537478029727936, + 0.0009311140165664256, + 0.0005758634652011096, + 0.006004549562931061, + 0.0042935386300086975, + 0.02090206742286682, + 0.005426986142992973, + 0.004936700686812401, + 0.043650925159454346, + 0.008165856823325157, + 0.024992842227220535, + 0.0001950712176039815, + 0.00014148994523566216, + 0.0003625450190156698, + 0.0001592914341017604, + 0.0001461404317524284, + 0.0006138115422800183, + 0.0001904308592202142, + 0.0006467031198553741, + 9.989809041144326e-05, + 0.00011565858585527167, + 0.00016699018306098878, + 0.00015409129264298826, + 0.00010102843953063712, + 0.00015012516814749688, + 0.00014805127284489572, + 0.0003105810610577464, + 0.0003754599310923368, + 0.00021258012566249818, + 0.0005561529542319477, + 0.00030789856100454926, + 0.00018394694779999554, + 0.0005267990054562688, + 0.00029003751114942133, + 0.00041783496271818876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.47427642345428467 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_12/index_tensor([8400, 8400])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_12/index_tensor([8400, 8400])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..95d1db15e220c4135ae96d336bd1da36141e58fc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_12/index_tensor([8400, 8400])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, slightly low-angled medium shot slowly settles in, revealing a rich fantasy concept art style. The camera focuses on a tall, slender figure seated at a potter's wheel in a dimly lit ancient stone workshop. This entity is wrapped in a living fabric woven from shifting purple shadows and deep blue dusk, with countless tiny, star-like points of light glittering within its undulating folds. In place of a head, a perfectly smooth, featureless sphere of swirling cosmic energy—an enveloping cloud of vibrant hues of mauve, cyan, and gold—serves as the scene's primary dynamic light source. Its slender, ethereal fingers, translucent and trailing a faint purplish energy, gracefully swing over the wheel, crafted from a single block of polished dark obsidian. Moving with steady, deliberate tenderness, the fingers deftly shape a lump of damp, dark clay, flecks of stardust shimmering in the cosmic glow. As the wheel rotates almost silently, the figure's hands delicately grasp and guide the rising mass, leaving faint, shimmering trails of light on its damp surface. The background is softly illuminated, revealing shelves housing previous creations: ceramic containers containing miniature rotating galaxies and perfectly shaped planets. The light from the figures’ heads casts dramatic, shifting shadows on the stone walls, creating an atmosphere of tranquility, mystery, and focused concentration.", + "seed": 28203, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_12/index_8400_seed_28203_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.27017860654741527, + "vlm_logprob": -0.27017860654741527, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.27017860654741527, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2252007722854614 + ], + "opposite_logprob_tokens": [ + -0.3502008020877838 + ], + "logprob_reward": -1.2252007722854614, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2252007722854614, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3840421438217163, + 0.1774444431066513, + 0.0384502075612545, + 0.022703975439071655, + 0.03842652216553688, + 0.04229479283094406, + 0.02972828410565853, + 0.2193080335855484, + 0.019319122657179832, + 0.028282405808568 + ], + "credit_map": [ + 0.08575994521379471, + 0.03867747634649277, + 0.03392524644732475, + 0.033929139375686646, + 0.030748547986149788, + 0.08574704825878143, + 0.026645343750715256, + 0.04860939085483551, + 0.019435204565525055, + 0.022394035011529922, + 0.01901179738342762, + 0.015446688048541546, + 0.013854419812560081, + 0.04017900675535202, + 0.017761345952749252, + 0.029361959546804428, + 0.00443196389824152, + 0.004517626017332077, + 0.006020489614456892, + 0.004455359652638435, + 0.0036092440132051706, + 0.005992693826556206, + 0.0034844730980694294, + 0.005938355345278978, + 0.002365667140111327, + 0.0029608188197016716, + 0.0023373053409159184, + 0.0021734051406383514, + 0.002949001034721732, + 0.003697276348248124, + 0.0023551983758807182, + 0.003865303937345743, + 0.004882691893726587, + 0.00430519413203001, + 0.00392555445432663, + 0.0042299493215978146, + 0.005121714435517788, + 0.006779392249882221, + 0.00434102164581418, + 0.004841006826609373, + 0.0047423518262803555, + 0.005955400876700878, + 0.0047759092412889, + 0.004018192179501057, + 0.003752850228920579, + 0.00894353911280632, + 0.003658002708107233, + 0.006448544096201658, + 0.0032883251551538706, + 0.0030024482402950525, + 0.004474976100027561, + 0.0037067013327032328, + 0.003474712371826172, + 0.005518786609172821, + 0.003135902341455221, + 0.0031264317221939564, + 0.02632053568959236, + 0.02191757597029209, + 0.018909400328993797, + 0.020903058350086212, + 0.01614830642938614, + 0.05193913355469704, + 0.02165795862674713, + 0.04151206091046333, + 0.0018082557944580913, + 0.002295977436006069, + 0.0018302691169083118, + 0.002675719326362014, + 0.0019337389385327697, + 0.0031913151033222675, + 0.001640813541598618, + 0.0039430344477295876, + 0.002718860749155283, + 0.00176893116440624, + 0.003310414729639888, + 0.0029408095870167017, + 0.002259886125102639, + 0.004313374869525433, + 0.003122851951047778, + 0.007847278378903866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -1.2252007722854614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821578413248062 + ], + "opposite_logprob_tokens": [ + -3.5298216342926025 + ], + "logprob_reward": -0.029821578413248062, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821578413248062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3393385410308838, + 0.19625331461429596, + 0.0339689627289772, + 0.08674293011426926, + 0.04781675711274147, + 0.03244776651263237, + 0.060968805104494095, + 0.14562924206256866, + 0.014609945937991142, + 0.04222371429204941 + ], + "credit_map": [ + 0.07144301384687424, + 0.03810925409197807, + 0.035363804548978806, + 0.0312715508043766, + 0.025639327242970467, + 0.0735381469130516, + 0.025734277442097664, + 0.038239169865846634, + 0.0399353913962841, + 0.01759514957666397, + 0.015317317098379135, + 0.013096291571855545, + 0.012082956731319427, + 0.06393813341856003, + 0.014157382771372795, + 0.0201307013630867, + 0.004023137502372265, + 0.003595892572775483, + 0.004528773948550224, + 0.0038325334899127483, + 0.003878215327858925, + 0.00616583414375782, + 0.0029729793313890696, + 0.0049715968780219555, + 0.011348467320203781, + 0.006377687212079763, + 0.005689859855920076, + 0.006364222150295973, + 0.004352112766355276, + 0.02791786752641201, + 0.004897305741906166, + 0.019795410335063934, + 0.005270908586680889, + 0.004996698349714279, + 0.005335496738553047, + 0.005649402737617493, + 0.005651865154504776, + 0.00830964744091034, + 0.006506710313260555, + 0.006096028722822666, + 0.0032577046658843756, + 0.004431809764355421, + 0.003987652715295553, + 0.003263590857386589, + 0.003128126496449113, + 0.005619263742119074, + 0.0032805295195430517, + 0.005479089450091124, + 0.007748208940029144, + 0.005944485776126385, + 0.006938131991773844, + 0.00599108450114727, + 0.005393367260694504, + 0.010869953781366348, + 0.007159562315791845, + 0.010924005880951881, + 0.018877625465393066, + 0.014002934098243713, + 0.012209845706820488, + 0.013438072055578232, + 0.010185396298766136, + 0.03712746128439903, + 0.01392526738345623, + 0.02586262859404087, + 0.0015168226091191173, + 0.0017353208968415856, + 0.0014753936557099223, + 0.001611070241779089, + 0.0016640982357785106, + 0.002475338988006115, + 0.001577565330080688, + 0.0025543374940752983, + 0.004556132480502129, + 0.002614897908642888, + 0.00452013686299324, + 0.004651839379221201, + 0.003481879597529769, + 0.005956233013421297, + 0.0041621034033596516, + 0.01228049024939537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029821578413248062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01109513733536005 + ], + "opposite_logprob_tokens": [ + -4.51109504699707 + ], + "logprob_reward": -0.01109513733536005, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01109513733536005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3910846710205078, + 0.1503380984067917, + 0.11098436266183853, + 0.03009212389588356, + 0.06653910130262375, + 0.039912864565849304, + 0.027585726231336594, + 0.1418963074684143, + 0.01836223155260086, + 0.02320450358092785 + ], + "credit_map": [ + 0.08955298364162445, + 0.03913506865501404, + 0.03074043057858944, + 0.033665984869003296, + 0.030997304245829582, + 0.08839522302150726, + 0.025997158139944077, + 0.05260053277015686, + 0.028807297348976135, + 0.017420582473278046, + 0.016556572169065475, + 0.012195699848234653, + 0.010560275055468082, + 0.0263086948543787, + 0.012789608910679817, + 0.025699365884065628, + 0.020430315285921097, + 0.01250633504241705, + 0.01432986743748188, + 0.008769025094807148, + 0.01048241462558508, + 0.020802091807127, + 0.00691124377772212, + 0.016753070056438446, + 0.004275180399417877, + 0.0027726825792342424, + 0.0029502524994313717, + 0.002608607057482004, + 0.0025768813211470842, + 0.007250464055687189, + 0.002240760251879692, + 0.005417295265942812, + 0.006564325653016567, + 0.010600165463984013, + 0.007932673208415508, + 0.006686560809612274, + 0.008089330047369003, + 0.012264791876077652, + 0.006534234620630741, + 0.007867014966905117, + 0.004271980840712786, + 0.003888010047376156, + 0.0035459809005260468, + 0.0037736992817372084, + 0.003500595921650529, + 0.008017567917704582, + 0.003426716197282076, + 0.009488311596214771, + 0.0046081943437457085, + 0.0030758113134652376, + 0.003180329455062747, + 0.0026885580737143755, + 0.0036704412195831537, + 0.005769982933998108, + 0.0023700136225670576, + 0.002222395734861493, + 0.018322763964533806, + 0.01476918812841177, + 0.011692048981785774, + 0.013626220636069775, + 0.010756666772067547, + 0.0346553772687912, + 0.013624834828078747, + 0.02444920688867569, + 0.0022031187545508146, + 0.002263409784063697, + 0.0018942486494779587, + 0.0022737043909728527, + 0.0024248219560831785, + 0.002914917189627886, + 0.001392392790876329, + 0.0029956179205328226, + 0.0028316383250057697, + 0.0016120115760713816, + 0.002781200222671032, + 0.002306877402588725, + 0.0019270684570074081, + 0.004796877037733793, + 0.002960497047752142, + 0.003988333512097597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01109513733536005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980376034975052 + ], + "opposite_logprob_tokens": [ + -3.52980375289917 + ], + "logprob_reward": -0.02980376034975052, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980376034975052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08724461495876312, + 0.26994550228118896, + 0.1969185322523117, + 0.04660217836499214, + 0.03942660614848137, + 0.18742401897907257, + 0.027664577588438988, + 0.0572107769548893, + 0.029473649337887764, + 0.05808960273861885 + ], + "credit_map": [ + 0.009631997905671597, + 0.013832597061991692, + 0.008274005725979805, + 0.012637872248888016, + 0.010422168299555779, + 0.015603417530655861, + 0.007682291325181723, + 0.009160259738564491, + 0.08405005186796188, + 0.013544811867177486, + 0.017568165436387062, + 0.012490704655647278, + 0.014314176514744759, + 0.08470568060874939, + 0.013804350979626179, + 0.029467549175024033, + 0.04364435747265816, + 0.016165567561984062, + 0.015577291138470173, + 0.00938697624951601, + 0.012617118656635284, + 0.057063598185777664, + 0.009897843934595585, + 0.03256576508283615, + 0.007582361809909344, + 0.0027898678090423346, + 0.0032181551214307547, + 0.003133156104013324, + 0.0035757662262767553, + 0.01622018963098526, + 0.002988888183608651, + 0.007093789055943489, + 0.00385261420160532, + 0.005788426846265793, + 0.005115479696542025, + 0.004329554736614227, + 0.004480575677007437, + 0.006259775720536709, + 0.004168885760009289, + 0.005431286990642548, + 0.024266306310892105, + 0.024046139791607857, + 0.01848684623837471, + 0.013187414966523647, + 0.01834889128804207, + 0.04981061816215515, + 0.011329486034810543, + 0.027948293834924698, + 0.0029569740872830153, + 0.0027710963040590286, + 0.0035507131833583117, + 0.003060111775994301, + 0.0025394349358975887, + 0.0055231195874512196, + 0.003157877130433917, + 0.004105247091501951, + 0.009324263781309128, + 0.00754863116890192, + 0.004444403108209372, + 0.0057909018360078335, + 0.005674606189131737, + 0.014640687964856625, + 0.0037312398198992014, + 0.006056040059775114, + 0.0030320340301841497, + 0.003614230314269662, + 0.0029678449500352144, + 0.002827022224664688, + 0.0034411954693496227, + 0.004891794174909592, + 0.002356879645958543, + 0.006342642940580845, + 0.006191224325448275, + 0.007493093144148588, + 0.007507408037781715, + 0.005186456721276045, + 0.0064134858548641205, + 0.008669331669807434, + 0.00427946588024497, + 0.01234913058578968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.02980376034975052 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054971784353256226 + ], + "opposite_logprob_tokens": [ + -2.929971694946289 + ], + "logprob_reward": -0.054971784353256226, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054971784353256226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.230188250541687, + 0.09810969233512878, + 0.21945174038410187, + 0.11431686580181122, + 0.048375215381383896, + 0.07103714346885681, + 0.03482880815863609, + 0.0825052261352539, + 0.046714745461940765, + 0.05447232723236084 + ], + "credit_map": [ + 0.05443238839507103, + 0.02027464471757412, + 0.0161348395049572, + 0.0192548967897892, + 0.017612101510167122, + 0.0553399883210659, + 0.014921177178621292, + 0.03221820294857025, + 0.017824457958340645, + 0.010945077985525131, + 0.00837021041661501, + 0.009583903476595879, + 0.009005802683532238, + 0.02657199464738369, + 0.008258442394435406, + 0.007549799047410488, + 0.0247064009308815, + 0.02674887701869011, + 0.0246062483638525, + 0.024492548778653145, + 0.028853336349129677, + 0.029944123700261116, + 0.02501794509589672, + 0.035082243382930756, + 0.015251749195158482, + 0.006977441255003214, + 0.010003914125263691, + 0.009528802707791328, + 0.010093755088746548, + 0.03321046009659767, + 0.010126993060112, + 0.01912374049425125, + 0.004644372966140509, + 0.006259263027459383, + 0.006936228834092617, + 0.005383494310081005, + 0.006116689648479223, + 0.00867614708840847, + 0.005078642629086971, + 0.005280372221022844, + 0.008261495269834995, + 0.005790675990283489, + 0.00570685463026166, + 0.006247159093618393, + 0.0065046632662415504, + 0.0179519671946764, + 0.005312406457960606, + 0.015261917375028133, + 0.004107248038053513, + 0.0026825282257050276, + 0.004104126710444689, + 0.00400557741522789, + 0.0031145820394158363, + 0.007621459197252989, + 0.00454464927315712, + 0.0046486328355968, + 0.021303700283169746, + 0.008338754065334797, + 0.004920864012092352, + 0.006462207995355129, + 0.006478680297732353, + 0.0241951122879982, + 0.004376234952360392, + 0.006429671309888363, + 0.004348287358880043, + 0.005227147601544857, + 0.004604427609592676, + 0.004665367770940065, + 0.004871262703090906, + 0.009045438840985298, + 0.004757242742925882, + 0.009195567108690739, + 0.006325673311948776, + 0.00433528795838356, + 0.007173709571361542, + 0.005839071236550808, + 0.005591315682977438, + 0.007470954675227404, + 0.0068314154632389545, + 0.010904894210398197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.054971784353256226 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, slightly low-angled medium shot slowly settles in, revealing a rich fantasy concept art style. The camera focuses on a tall, slender figure seated at a potter's wheel in a dimly lit ancient stone workshop. This entity is wrapped in a living fabric woven from shifting purple shadows and deep blue dusk, with countless tiny, star-like points of light glittering within its undulating folds. In place of a head, a perfectly smooth, featureless sphere of swirling cosmic energy—an enveloping cloud of vibrant hues of mauve, cyan, and gold—serves as the scene's primary dynamic light source. Its slender, ethereal fingers, translucent and trailing a faint purplish energy, gracefully swing over the wheel, crafted from a single block of polished dark obsidian. Moving with steady, deliberate tenderness, the fingers deftly shape a lump of damp, dark clay, flecks of stardust shimmering in the cosmic glow. As the wheel rotates almost silently, the figure's hands delicately grasp and guide the rising mass, leaving faint, shimmering trails of light on its damp surface. The background is softly illuminated, revealing shelves housing previous creations: ceramic containers containing miniature rotating galaxies and perfectly shaped planets. The light from the figures’ heads casts dramatic, shifting shadows on the stone walls, creating an atmosphere of tranquility, mystery, and focused concentration.", + "seed": 28204, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_12/index_8400_seed_28204_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1331307042390108, + "vlm_logprob": -0.1331307042390108, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.1331307042390108, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5292794704437256 + ], + "opposite_logprob_tokens": [ + -0.9042794704437256 + ], + "logprob_reward": -0.5292794704437256, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5292794704437256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4209519326686859, + 0.16414999961853027, + 0.04834417998790741, + 0.06667964905500412, + 0.07044149190187454, + 0.08978487551212311, + 0.04085320606827736, + 0.0354434959590435, + 0.044949743896722794, + 0.018401486799120903 + ], + "credit_map": [ + 0.06012090668082237, + 0.040070515125989914, + 0.05803990364074707, + 0.029421504586935043, + 0.046559154987335205, + 0.07811315357685089, + 0.03147120401263237, + 0.07715556770563126, + 0.02041509561240673, + 0.01594027690589428, + 0.023025838658213615, + 0.010844647884368896, + 0.016851048916578293, + 0.0309657771140337, + 0.01628369837999344, + 0.02982359193265438, + 0.006487635895609856, + 0.004194803070276976, + 0.006724975071847439, + 0.003293382003903389, + 0.005101985298097134, + 0.009247608482837677, + 0.004715327639132738, + 0.00857845600694418, + 0.009589096531271935, + 0.005389578174799681, + 0.009614669717848301, + 0.004241902846843004, + 0.005893634166568518, + 0.014264918863773346, + 0.0047248839400708675, + 0.012960956431925297, + 0.008076653815805912, + 0.005317630246281624, + 0.011519599705934525, + 0.003514674259349704, + 0.0070251463912427425, + 0.0189986415207386, + 0.004166949540376663, + 0.011822191998362541, + 0.012098337523639202, + 0.005752525757998228, + 0.014653456397354603, + 0.0045376988127827644, + 0.008193885907530785, + 0.025318995118141174, + 0.00473950756713748, + 0.014490454457700253, + 0.006679203361272812, + 0.004312080796808004, + 0.006423150654882193, + 0.0040447828359901905, + 0.004773403517901897, + 0.005134277977049351, + 0.003650922793895006, + 0.005835382733494043, + 0.005080698058009148, + 0.0034146225079894066, + 0.0058785174041986465, + 0.002713482128456235, + 0.004322712309658527, + 0.006506455596536398, + 0.003268449567258358, + 0.004258556291460991, + 0.005502891726791859, + 0.0042357551865279675, + 0.005520978011190891, + 0.0066291699185967445, + 0.005043720826506615, + 0.006109035108238459, + 0.003913295455276966, + 0.007994893938302994, + 0.0024142074398696423, + 0.0025305747985839844, + 0.0025464454665780067, + 0.0017536183586344123, + 0.002244840143248439, + 0.0024669100530445576, + 0.0015814736252650619, + 0.0028634159825742245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.5292794704437256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314859718084335 + ], + "opposite_logprob_tokens": [ + -3.773314952850342 + ], + "logprob_reward": -0.023314859718084335, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314859718084335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35796019434928894, + 0.17547766864299774, + 0.05526444688439369, + 0.06425359100103378, + 0.06137262284755707, + 0.06418628990650177, + 0.04864279925823212, + 0.05618063360452652, + 0.05524134635925293, + 0.061420366168022156 + ], + "credit_map": [ + 0.05436763912439346, + 0.03615051507949829, + 0.04838479310274124, + 0.02943367138504982, + 0.037478018552064896, + 0.06307291239500046, + 0.03339574113488197, + 0.0556769073009491, + 0.02690652385354042, + 0.016038551926612854, + 0.024763455614447594, + 0.014359673485159874, + 0.017373837530612946, + 0.0311278086155653, + 0.01868196204304695, + 0.026225870475172997, + 0.006771677173674107, + 0.008324016816914082, + 0.00702366279438138, + 0.004737220238894224, + 0.008651933632791042, + 0.006468987558037043, + 0.006800545379519463, + 0.006486410275101662, + 0.009413684718310833, + 0.006595544517040253, + 0.009786221198737621, + 0.004069838207215071, + 0.006969870533794165, + 0.009410138241946697, + 0.007072321139276028, + 0.01093597337603569, + 0.0077140131033957005, + 0.005993404891341925, + 0.008326494134962559, + 0.005176824983209372, + 0.006669938564300537, + 0.012300563976168633, + 0.00505287479609251, + 0.010138512589037418, + 0.007843839935958385, + 0.00672141695395112, + 0.0086728660389781, + 0.0027853345964103937, + 0.008246666751801968, + 0.016985630616545677, + 0.002984861843287945, + 0.00994567759335041, + 0.007442974019795656, + 0.00874983612447977, + 0.007964336313307285, + 0.0048991101793944836, + 0.005699191242456436, + 0.005362388677895069, + 0.004397887270897627, + 0.004127080086618662, + 0.0082015972584486, + 0.0053346725180745125, + 0.009892640635371208, + 0.003701859852299094, + 0.0058801001869142056, + 0.01147861685603857, + 0.004663946572691202, + 0.007027204614132643, + 0.007429470308125019, + 0.004849148448556662, + 0.00818494614213705, + 0.007599318400025368, + 0.007227038498967886, + 0.008488492108881474, + 0.006186000537127256, + 0.005276932381093502, + 0.00850778166204691, + 0.007722608279436827, + 0.00932918768376112, + 0.005659292917698622, + 0.007883075624704361, + 0.01004745066165924, + 0.005259179510176182, + 0.007011794950813055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023314859718084335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009803364053368568 + ], + "opposite_logprob_tokens": [ + -4.634803295135498 + ], + "logprob_reward": -0.009803364053368568, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009803364053368568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30578991770744324, + 0.2155718207359314, + 0.06834318488836288, + 0.0867803692817688, + 0.05928841605782509, + 0.07572367042303085, + 0.062145430594682693, + 0.03622891753911972, + 0.045675113797187805, + 0.044453199952840805 + ], + "credit_map": [ + 0.03789283707737923, + 0.03163916617631912, + 0.04259224608540535, + 0.02081225998699665, + 0.044639043509960175, + 0.05166054144501686, + 0.020465638488531113, + 0.056088194251060486, + 0.028619782999157906, + 0.017644057050347328, + 0.030431704595685005, + 0.014670481905341148, + 0.018935389816761017, + 0.04352327063679695, + 0.022143809124827385, + 0.03960331901907921, + 0.008878138847649097, + 0.0074698710814118385, + 0.00878920592367649, + 0.005896223708987236, + 0.008380388841032982, + 0.00946563296020031, + 0.007226875517517328, + 0.012236849404871464, + 0.012164830230176449, + 0.007896112278103828, + 0.011986141093075275, + 0.0065822782926261425, + 0.010073265992105007, + 0.013383080251514912, + 0.007947606965899467, + 0.01674704998731613, + 0.00657157925888896, + 0.005841277074068785, + 0.008973697200417519, + 0.0037402037996798754, + 0.007996671833097935, + 0.011501289904117584, + 0.005619421135634184, + 0.009044276550412178, + 0.008920745924115181, + 0.004935311619192362, + 0.010982786305248737, + 0.0029880227521061897, + 0.007120026741176844, + 0.02321249432861805, + 0.004051823168992996, + 0.013512460514903069, + 0.008397961035370827, + 0.009343136101961136, + 0.010117091238498688, + 0.005033867433667183, + 0.008614882826805115, + 0.009438558481633663, + 0.0048025972209870815, + 0.006397330202162266, + 0.005249887239187956, + 0.0030125363264232874, + 0.0062748584896326065, + 0.0021377489902079105, + 0.0034879951272159815, + 0.008206909522414207, + 0.0030878609977662563, + 0.004771119449287653, + 0.0062154559418559074, + 0.004884397145360708, + 0.005393386352807283, + 0.007727798540145159, + 0.00593869574368, + 0.004785733763128519, + 0.00481965159997344, + 0.005909996572881937, + 0.006492044776678085, + 0.0050702993758022785, + 0.0071801682934165, + 0.0037599571514874697, + 0.005221088882535696, + 0.007578834891319275, + 0.0036627098452299833, + 0.0054880958050489426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009803364053368568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195535331964493 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195535331964493, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195535331964493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2046223282814026, + 0.12098385393619537, + 0.0806988999247551, + 0.09052382409572601, + 0.13370104134082794, + 0.14128290116786957, + 0.07385718077421188, + 0.053684212267398834, + 0.056011974811553955, + 0.044633783400058746 + ], + "credit_map": [ + 0.025644522160291672, + 0.017917219549417496, + 0.03055272065103054, + 0.01450363453477621, + 0.016185345128178596, + 0.041685108095407486, + 0.013969223946332932, + 0.04416454955935478, + 0.016373418271541595, + 0.010825470089912415, + 0.015879416838288307, + 0.009212234057486057, + 0.01527670118957758, + 0.021000072360038757, + 0.012572678737342358, + 0.01984386518597603, + 0.009022132493555546, + 0.010972012765705585, + 0.01108626089990139, + 0.00563898915424943, + 0.014167677611112595, + 0.012225105427205563, + 0.006433952134102583, + 0.011152763850986958, + 0.011167047545313835, + 0.011017728596925735, + 0.014623617753386497, + 0.006627585273236036, + 0.011137071996927261, + 0.016427958384156227, + 0.006757078226655722, + 0.012765728868544102, + 0.016255449503660202, + 0.019296271726489067, + 0.017937909811735153, + 0.011022833175957203, + 0.024068377912044525, + 0.016346119344234467, + 0.013199587352573872, + 0.015574480406939983, + 0.02071230858564377, + 0.022198714315891266, + 0.017277702689170837, + 0.009453162550926208, + 0.025586050003767014, + 0.02336837351322174, + 0.008526314981281757, + 0.014160273596644402, + 0.010525016114115715, + 0.0076122283935546875, + 0.012671787291765213, + 0.005191472824662924, + 0.01072280015796423, + 0.011928403750061989, + 0.006199614144861698, + 0.009005855768918991, + 0.007086825557053089, + 0.005659686867147684, + 0.0079341446980834, + 0.003915184643119574, + 0.006026780232787132, + 0.011087533086538315, + 0.004794008564203978, + 0.007180050481110811, + 0.006871342658996582, + 0.005765991285443306, + 0.0070608570240437984, + 0.01005689799785614, + 0.007075215224176645, + 0.00631287507712841, + 0.0069638704881072044, + 0.00590492133051157, + 0.006321254186332226, + 0.005648619029670954, + 0.006367978174239397, + 0.0046074059791862965, + 0.006211889907717705, + 0.006181948818266392, + 0.00407195882871747, + 0.005222727544605732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.014195535331964493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906029164791107 + ], + "opposite_logprob_tokens": [ + -2.4640603065490723 + ], + "logprob_reward": -0.08906029164791107, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906029164791107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23501095175743103, + 0.1739240288734436, + 0.12497574836015701, + 0.08865803480148315, + 0.08074816316366196, + 0.04507087543606758, + 0.03128211945295334, + 0.09040284156799316, + 0.040768448263406754, + 0.08915877342224121 + ], + "credit_map": [ + 0.03318892791867256, + 0.03477003052830696, + 0.037226103246212006, + 0.020773788914084435, + 0.035806216299533844, + 0.02595440112054348, + 0.022336797788739204, + 0.024954665452241898, + 0.023843904957175255, + 0.016131971031427383, + 0.02262832038104534, + 0.012782425619661808, + 0.01632167585194111, + 0.03308119252324104, + 0.01989211142063141, + 0.02924242801964283, + 0.015202627517282963, + 0.013111455366015434, + 0.016352588310837746, + 0.0076561132445931435, + 0.015157279558479786, + 0.024639245122671127, + 0.010743951424956322, + 0.022112492471933365, + 0.010804379358887672, + 0.009466438554227352, + 0.012492149136960506, + 0.005568261723965406, + 0.009796866215765476, + 0.01963840425014496, + 0.006240117829293013, + 0.014651416800916195, + 0.01093495823442936, + 0.011582800187170506, + 0.013571595773100853, + 0.005736738443374634, + 0.016731439158320427, + 0.007232458796352148, + 0.007678720634430647, + 0.0072794570587575436, + 0.006676895078271627, + 0.00606542406603694, + 0.0069664837792515755, + 0.003580724820494652, + 0.0063504003919661045, + 0.006781656760722399, + 0.003586642909795046, + 0.005062648095190525, + 0.004444051533937454, + 0.003842072794213891, + 0.00451056519523263, + 0.002963833510875702, + 0.004469104576855898, + 0.00478703947737813, + 0.003242786508053541, + 0.003022665623575449, + 0.01227869838476181, + 0.007926402613520622, + 0.01474134810268879, + 0.006350807845592499, + 0.008842616342008114, + 0.019635945558547974, + 0.00813283771276474, + 0.01249418780207634, + 0.005683643743395805, + 0.0040471660904586315, + 0.005201971624046564, + 0.00498098274692893, + 0.006932713091373444, + 0.0060707018710672855, + 0.0038415533490478992, + 0.004009717144072056, + 0.013117964379489422, + 0.007120862603187561, + 0.015312552452087402, + 0.007691219449043274, + 0.010160773992538452, + 0.016395259648561478, + 0.008133608847856522, + 0.01122653391212225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08906029164791107 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, slightly low-angled medium shot slowly settles in, revealing a rich fantasy concept art style. The camera focuses on a tall, slender figure seated at a potter's wheel in a dimly lit ancient stone workshop. This entity is wrapped in a living fabric woven from shifting purple shadows and deep blue dusk, with countless tiny, star-like points of light glittering within its undulating folds. In place of a head, a perfectly smooth, featureless sphere of swirling cosmic energy—an enveloping cloud of vibrant hues of mauve, cyan, and gold—serves as the scene's primary dynamic light source. Its slender, ethereal fingers, translucent and trailing a faint purplish energy, gracefully swing over the wheel, crafted from a single block of polished dark obsidian. Moving with steady, deliberate tenderness, the fingers deftly shape a lump of damp, dark clay, flecks of stardust shimmering in the cosmic glow. As the wheel rotates almost silently, the figure's hands delicately grasp and guide the rising mass, leaving faint, shimmering trails of light on its damp surface. The background is softly illuminated, revealing shelves housing previous creations: ceramic containers containing miniature rotating galaxies and perfectly shaped planets. The light from the figures’ heads casts dramatic, shifting shadows on the stone walls, creating an atmosphere of tranquility, mystery, and focused concentration.", + "seed": 28205, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_12/index_8400_seed_28205_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09692241745069623, + "vlm_logprob": -0.09692241745069623, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.09692241745069623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.389062762260437 + ], + "opposite_logprob_tokens": [ + -1.139062762260437 + ], + "logprob_reward": -0.389062762260437, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.389062762260437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.184846431016922, + 0.07284871488809586, + 0.01909797452390194, + 0.08500958234071732, + 0.019267316907644272, + 0.10187133401632309, + 0.1684485524892807, + 0.017784759402275085, + 0.31868356466293335, + 0.012141755782067776 + ], + "credit_map": [ + 0.03368464484810829, + 0.015920430421829224, + 0.039561979472637177, + 0.01764635555446148, + 0.015948835760354996, + 0.02145557291805744, + 0.017804879695177078, + 0.022823719307780266, + 0.01725092902779579, + 0.0047140647657215595, + 0.013111086562275887, + 0.00521128112450242, + 0.004210660699754953, + 0.008493347093462944, + 0.007089988328516483, + 0.012767351232469082, + 0.003671521320939064, + 0.0017294826684519649, + 0.0030519880820065737, + 0.001995209138840437, + 0.0015558336162939668, + 0.0029360384214669466, + 0.0015076532727107406, + 0.0026502495165914297, + 0.022937944158911705, + 0.005611629690974951, + 0.011349301785230637, + 0.005312408786267042, + 0.00578320911154151, + 0.01188588049262762, + 0.0057543502189219, + 0.01637485809624195, + 0.003414779668673873, + 0.001623104210011661, + 0.0034307895693928003, + 0.0013396712020039558, + 0.001537711825221777, + 0.0035597370006144047, + 0.0021041976287961006, + 0.0022573249880224466, + 0.04157068952918053, + 0.003766146954149008, + 0.009332290850579739, + 0.005301347468048334, + 0.00326373428106308, + 0.007601557299494743, + 0.0031026715878397226, + 0.027932889759540558, + 0.06467825174331665, + 0.008016148582100868, + 0.017359724268317223, + 0.006740270648151636, + 0.005709667690098286, + 0.011455127038061619, + 0.007113313302397728, + 0.04737605154514313, + 0.0028380725998431444, + 0.002130642533302307, + 0.0030324459075927734, + 0.001907629775814712, + 0.001513066585175693, + 0.0025031522382050753, + 0.0015178913017734885, + 0.002341858111321926, + 0.13016434013843536, + 0.014576872810721397, + 0.030236054211854935, + 0.012776480987668037, + 0.011906088329851627, + 0.025466356426477432, + 0.012079067528247833, + 0.08147832751274109, + 0.002081652171909809, + 0.0012765169376507401, + 0.0018297062488272786, + 0.0008868296281434596, + 0.0013228413881734014, + 0.002208966761827469, + 0.0011758912587538362, + 0.0013593509793281555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.389062762260437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608232781291008 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608232781291008, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608232781291008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05879734829068184, + 0.023981161415576935, + 0.007114697713404894, + 0.012784687802195549, + 0.006582253146916628, + 0.07427645474672318, + 0.730488121509552, + 0.00547605799511075, + 0.07351325452327728, + 0.006985990330576897 + ], + "credit_map": [ + 0.015733089298009872, + 0.004322128836065531, + 0.00809414591640234, + 0.00437709828838706, + 0.004355538170784712, + 0.007592129521071911, + 0.004086395259946585, + 0.010236823931336403, + 0.004047916736453772, + 0.0025444224011152983, + 0.0037939781323075294, + 0.0018688186537474394, + 0.0030474697705358267, + 0.003296289825811982, + 0.0027648054528981447, + 0.0026174630038440228, + 0.0016011992702260613, + 0.0005943800788372755, + 0.001052382867783308, + 0.0007593870395794511, + 0.0006019534775987267, + 0.0008463104022666812, + 0.0004661905695684254, + 0.00119289371650666, + 0.003707681316882372, + 0.0007585336570627987, + 0.0014627458294853568, + 0.0009121769107878208, + 0.000873786339070648, + 0.001469418522901833, + 0.000973427842836827, + 0.0026269175577908754, + 0.0012350559700280428, + 0.0005452334880828857, + 0.0011129494523629546, + 0.0004947484703734517, + 0.0006540536414831877, + 0.0011546746827661991, + 0.0005663545452989638, + 0.0008191828965209424, + 0.03179832920432091, + 0.0023907574359327555, + 0.006320721935480833, + 0.0034690049942582846, + 0.0017593462252989411, + 0.004958410281687975, + 0.002027978887781501, + 0.02155190147459507, + 0.30714675784111023, + 0.025974800810217857, + 0.06935422867536545, + 0.02014782652258873, + 0.01954183354973793, + 0.04581131413578987, + 0.024453934282064438, + 0.21805742383003235, + 0.0010225368896499276, + 0.0007711686776019633, + 0.0008655316778458655, + 0.0006529256934300065, + 0.0004952571471221745, + 0.0005048304446972907, + 0.0005615408881567419, + 0.0006022668676450849, + 0.02969704382121563, + 0.0033845247235149145, + 0.006820530164986849, + 0.003101849230006337, + 0.00277662742882967, + 0.006478787865489721, + 0.002783079631626606, + 0.01847081072628498, + 0.001439978601410985, + 0.0007298147538676858, + 0.0010647432645782828, + 0.0006380905397236347, + 0.0006704719853587449, + 0.0008319301996380091, + 0.0007602018304169178, + 0.0008507590391673148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.01608232781291008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643500808626413 + ], + "opposite_logprob_tokens": [ + -4.882643699645996 + ], + "logprob_reward": -0.007643500808626413, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643500808626413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16764724254608154, + 0.11840825527906418, + 0.03786766156554222, + 0.0628795474767685, + 0.011215302161872387, + 0.3119264543056488, + 0.17213791608810425, + 0.027078449726104736, + 0.07150004059076309, + 0.019339125603437424 + ], + "credit_map": [ + 0.028862889856100082, + 0.015057596378028393, + 0.0373770073056221, + 0.018970666453242302, + 0.015110553242266178, + 0.01859130710363388, + 0.016193130984902382, + 0.017484091222286224, + 0.02947300858795643, + 0.009531461633741856, + 0.016525056213140488, + 0.007761162705719471, + 0.008479920215904713, + 0.014646701514720917, + 0.010946165770292282, + 0.021044781431555748, + 0.007456005550920963, + 0.003099614754319191, + 0.006475362461060286, + 0.003495982848107815, + 0.0029691862873733044, + 0.00727806193754077, + 0.002390845213085413, + 0.004702604375779629, + 0.01823754794895649, + 0.004520858637988567, + 0.00762259541079402, + 0.004113505128771067, + 0.004439204465597868, + 0.006498026195913553, + 0.0040866597555577755, + 0.013361155055463314, + 0.0020117221865803003, + 0.0008872231119312346, + 0.0021937054116278887, + 0.0009406147291883826, + 0.0007697642431594431, + 0.0020184963941574097, + 0.0011082683922722936, + 0.0012855078093707561, + 0.13062423467636108, + 0.010879769921302795, + 0.027759263291954994, + 0.014914046041667461, + 0.008998512290418148, + 0.023331211879849434, + 0.008685850538313389, + 0.08673356473445892, + 0.0717962384223938, + 0.006548067554831505, + 0.016218073666095734, + 0.004351987969130278, + 0.0047350660897791386, + 0.011596004478633404, + 0.005585850216448307, + 0.05130661278963089, + 0.004345287103205919, + 0.0031267802696675062, + 0.006061327178031206, + 0.0028068707324564457, + 0.002456930698826909, + 0.0027056371327489614, + 0.0019536279141902924, + 0.0036219866015017033, + 0.026587648317217827, + 0.003595740767195821, + 0.008142081089317799, + 0.0035136656370013952, + 0.0037617178168147802, + 0.004615230951458216, + 0.0036955426912754774, + 0.01758841797709465, + 0.0030243319924920797, + 0.002052085706964135, + 0.0035870145075023174, + 0.0016138345235958695, + 0.0015877325786277652, + 0.003914133179932833, + 0.001588787417858839, + 0.001971206860616803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.007643500808626413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778574109077454 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778574109077454, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778574109077454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24212408065795898, + 0.08251488208770752, + 0.04509810358285904, + 0.04684815928339958, + 0.023106709122657776, + 0.18939581513404846, + 0.20190003514289856, + 0.030402328819036484, + 0.11215713620185852, + 0.026452692225575447 + ], + "credit_map": [ + 0.0700133889913559, + 0.024698467925190926, + 0.02652004361152649, + 0.014856209047138691, + 0.021252410486340523, + 0.0215587355196476, + 0.01883459836244583, + 0.04439025744795799, + 0.013154953718185425, + 0.011472171172499657, + 0.011302012950181961, + 0.007659933064132929, + 0.00990832969546318, + 0.011644518934190273, + 0.010478281415998936, + 0.006894687656313181, + 0.010259049944579601, + 0.0060709151439368725, + 0.0049622152000665665, + 0.004732911475002766, + 0.004946723580360413, + 0.004688201937824488, + 0.0038494463078677654, + 0.005588643718510866, + 0.01122067216783762, + 0.0043576303869485855, + 0.00574085908010602, + 0.0033742962405085564, + 0.003985109273344278, + 0.005756665486842394, + 0.003557171206921339, + 0.008855757303535938, + 0.004120438359677792, + 0.0028754661325365305, + 0.0033914255909621716, + 0.001825491082854569, + 0.002028458984568715, + 0.0034824141766875982, + 0.002222553128376603, + 0.0031604627147316933, + 0.07934999465942383, + 0.006820602808147669, + 0.016967542469501495, + 0.009020503610372543, + 0.00567304901778698, + 0.013578858226537704, + 0.005354530178010464, + 0.052630744874477386, + 0.08296577632427216, + 0.007492503616958857, + 0.019608886912465096, + 0.006658392027020454, + 0.006197926588356495, + 0.01284042838960886, + 0.007615314330905676, + 0.0585208386182785, + 0.004499754868447781, + 0.0044341799803078175, + 0.00415852339938283, + 0.0035841206554323435, + 0.0033700251951813698, + 0.004116034135222435, + 0.0023998727556318045, + 0.0038398192264139652, + 0.04538343474268913, + 0.0050683291628956795, + 0.010996558703482151, + 0.004866795148700476, + 0.004531350918114185, + 0.0076513513922691345, + 0.005113217514008284, + 0.028546102344989777, + 0.00437858235090971, + 0.004947188775986433, + 0.003093559527769685, + 0.0023516854271292686, + 0.002879486419260502, + 0.002848547650501132, + 0.0027881250716745853, + 0.003165519330650568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.009778574109077454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204492226243019 + ], + "opposite_logprob_tokens": [ + -2.812044858932495 + ], + "logprob_reward": -0.06204492226243019, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204492226243019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11037136614322662, + 0.11089419573545456, + 0.04311337321996689, + 0.02176518552005291, + 0.035249751061201096, + 0.286467045545578, + 0.23863832652568817, + 0.032548245042562485, + 0.09240708500146866, + 0.02854541689157486 + ], + "credit_map": [ + 0.025182293727993965, + 0.010254045017063618, + 0.01941194199025631, + 0.009359097108244896, + 0.010445527732372284, + 0.01433537807315588, + 0.00829396490007639, + 0.013089114800095558, + 0.02823815308511257, + 0.006707245018333197, + 0.019920222461223602, + 0.007216997444629669, + 0.006379575468599796, + 0.013858488760888577, + 0.008322133682668209, + 0.020251376554369926, + 0.006177208386361599, + 0.0038715230766683817, + 0.008072257973253727, + 0.004547944292426109, + 0.004972527269273996, + 0.007124040275812149, + 0.003106118878349662, + 0.005241753533482552, + 0.006487243343144655, + 0.0015924471663311124, + 0.0026610130444169044, + 0.001090974546968937, + 0.0016823416808620095, + 0.0027953239623457193, + 0.001385923707857728, + 0.004069918766617775, + 0.006787019781768322, + 0.0030427947640419006, + 0.00863012857735157, + 0.0024177038576453924, + 0.0026524255517870188, + 0.004457526374608278, + 0.0025657685473561287, + 0.004696386866271496, + 0.12232263386249542, + 0.009726147167384624, + 0.024710657075047493, + 0.01301119476556778, + 0.007146970368921757, + 0.0199295561760664, + 0.007853735238313675, + 0.08176616579294205, + 0.09733196347951889, + 0.008776938542723656, + 0.023146195337176323, + 0.008944196626543999, + 0.007124338299036026, + 0.014262319542467594, + 0.009969279170036316, + 0.06908309459686279, + 0.005516229663044214, + 0.004102778155356646, + 0.004675919655710459, + 0.003475550562143326, + 0.004106310661882162, + 0.0036362765822559595, + 0.0026548048481345177, + 0.004380377475172281, + 0.03824356570839882, + 0.004103694576770067, + 0.00847090594470501, + 0.0037934465799480677, + 0.003211923409253359, + 0.00650437967851758, + 0.004382326267659664, + 0.023696843534708023, + 0.005399756133556366, + 0.0033316926565021276, + 0.0040864404290914536, + 0.0025739928241819143, + 0.0031226796563714743, + 0.0037224451079964638, + 0.0028143052477389574, + 0.0034941041376441717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.06204492226243019 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, slightly low-angled medium shot slowly settles in, revealing a rich fantasy concept art style. The camera focuses on a tall, slender figure seated at a potter's wheel in a dimly lit ancient stone workshop. This entity is wrapped in a living fabric woven from shifting purple shadows and deep blue dusk, with countless tiny, star-like points of light glittering within its undulating folds. In place of a head, a perfectly smooth, featureless sphere of swirling cosmic energy—an enveloping cloud of vibrant hues of mauve, cyan, and gold—serves as the scene's primary dynamic light source. Its slender, ethereal fingers, translucent and trailing a faint purplish energy, gracefully swing over the wheel, crafted from a single block of polished dark obsidian. Moving with steady, deliberate tenderness, the fingers deftly shape a lump of damp, dark clay, flecks of stardust shimmering in the cosmic glow. As the wheel rotates almost silently, the figure's hands delicately grasp and guide the rising mass, leaving faint, shimmering trails of light on its damp surface. The background is softly illuminated, revealing shelves housing previous creations: ceramic containers containing miniature rotating galaxies and perfectly shaped planets. The light from the figures’ heads casts dramatic, shifting shadows on the stone walls, creating an atmosphere of tranquility, mystery, and focused concentration.", + "seed": 28206, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_12/index_8400_seed_28206_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026031104754656552, + "vlm_logprob": -0.026031104754656552, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.026031104754656552, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09211734682321548 + ], + "opposite_logprob_tokens": [ + -2.4671173095703125 + ], + "logprob_reward": -0.09211734682321548, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09211734682321548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.264232337474823, + 0.11526358872652054, + 0.07582632452249527, + 0.06146423518657684, + 0.10968437790870667, + 0.2543911337852478, + 0.033570270985364914, + 0.031110137701034546, + 0.02136482298374176, + 0.03309274837374687 + ], + "credit_map": [ + 0.031229261308908463, + 0.02440229058265686, + 0.040405530482530594, + 0.033295050263404846, + 0.0172189362347126, + 0.03167543560266495, + 0.03200323507189751, + 0.05400259792804718, + 0.012599678710103035, + 0.009192636236548424, + 0.018847841769456863, + 0.011653808876872063, + 0.00790380034595728, + 0.013655184768140316, + 0.017084505409002304, + 0.024326132610440254, + 0.010150611400604248, + 0.007648708298802376, + 0.01349957101047039, + 0.005499154329299927, + 0.005160005763173103, + 0.01241027656942606, + 0.006447473540902138, + 0.015010522678494453, + 0.005276872776448727, + 0.006822394672781229, + 0.010121927596628666, + 0.006899369414895773, + 0.005224531516432762, + 0.007918475195765495, + 0.010324469767510891, + 0.00887619610875845, + 0.009968496859073639, + 0.011124610900878906, + 0.014717405661940575, + 0.012509332038462162, + 0.007846543565392494, + 0.016019871458411217, + 0.0126152653247118, + 0.024882839992642403, + 0.031242668628692627, + 0.016965392976999283, + 0.03996560722589493, + 0.022314440459012985, + 0.013272633776068687, + 0.04150228202342987, + 0.04194142296910286, + 0.047186676412820816, + 0.003621816635131836, + 0.0031438441947102547, + 0.004844846669584513, + 0.00335995783098042, + 0.00267605297267437, + 0.004076819866895676, + 0.003940373659133911, + 0.007906557060778141, + 0.004792870953679085, + 0.0026378233451396227, + 0.004683015868067741, + 0.002798602916300297, + 0.0021079194266349077, + 0.004364623688161373, + 0.0032168980687856674, + 0.006508383899927139, + 0.002076939446851611, + 0.0029526101425290108, + 0.003010708140209317, + 0.0024653535801917315, + 0.0025438060984015465, + 0.0028362651355564594, + 0.0016246294835582376, + 0.0038545094430446625, + 0.0033022565767169, + 0.0036867549642920494, + 0.004923183936625719, + 0.005132459104061127, + 0.0030021723359823227, + 0.0038229492492973804, + 0.004492875188589096, + 0.004730097018182278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.09211734682321548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964457057416439 + ], + "opposite_logprob_tokens": [ + -5.130964279174805 + ], + "logprob_reward": -0.005964457057416439, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964457057416439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.276685893535614, + 0.13352154195308685, + 0.09120511263608932, + 0.06166164577007294, + 0.11234104633331299, + 0.09398183971643448, + 0.036154963076114655, + 0.0703548714518547, + 0.025090165436267853, + 0.09900283068418503 + ], + "credit_map": [ + 0.03287113085389137, + 0.028357496485114098, + 0.03956620395183563, + 0.02735215611755848, + 0.02411729469895363, + 0.037256140261888504, + 0.02834244817495346, + 0.05882304161787033, + 0.01873208023607731, + 0.010128882713615894, + 0.016647322103381157, + 0.012374638579785824, + 0.00930880755186081, + 0.0244760625064373, + 0.017390621826052666, + 0.024463143199682236, + 0.009025937877595425, + 0.009185588918626308, + 0.016568243503570557, + 0.012750997208058834, + 0.006980211939662695, + 0.01113151479512453, + 0.010818869806826115, + 0.014743756502866745, + 0.005966509692370892, + 0.0049200886860489845, + 0.008467164821922779, + 0.006372429430484772, + 0.00433239946141839, + 0.00981215201318264, + 0.008645751513540745, + 0.01314515620470047, + 0.010936617851257324, + 0.011897069402039051, + 0.012607666663825512, + 0.016026636585593224, + 0.008705466985702515, + 0.01689143292605877, + 0.011865966953337193, + 0.023410197347402573, + 0.0167692881077528, + 0.007706856820732355, + 0.008173990063369274, + 0.00687285466119647, + 0.0058213272131979465, + 0.02090095542371273, + 0.013059187680482864, + 0.014677387662231922, + 0.004101946484297514, + 0.005596589297056198, + 0.004160888027399778, + 0.0036634025163948536, + 0.004084632266312838, + 0.0056610796600580215, + 0.002859893487766385, + 0.006026534829288721, + 0.005729814991354942, + 0.009672378189861774, + 0.007918394170701504, + 0.007129007484763861, + 0.007375052664428949, + 0.010628728196024895, + 0.008214562200009823, + 0.013686935417354107, + 0.0025130868889391422, + 0.0027849776670336723, + 0.005028543993830681, + 0.002360736019909382, + 0.0023682131431996822, + 0.003211644012480974, + 0.0027518542483448982, + 0.00407111132517457, + 0.008055225946009159, + 0.009094675071537495, + 0.013856523670256138, + 0.01824692264199257, + 0.00717150466516614, + 0.01269100047647953, + 0.009091699495911598, + 0.020795291289687157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005964457057416439 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004107373300939798 + ], + "opposite_logprob_tokens": [ + -5.504107475280762 + ], + "logprob_reward": -0.004107373300939798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107373300939798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24910679459571838, + 0.17363326251506805, + 0.09849715232849121, + 0.0874016061425209, + 0.051344048231840134, + 0.10541670769453049, + 0.03709174692630768, + 0.046974487602710724, + 0.05069682374596596, + 0.09983738511800766 + ], + "credit_map": [ + 0.03336374834179878, + 0.026324577629566193, + 0.03383561968803406, + 0.02901419810950756, + 0.01951930671930313, + 0.031145399436354637, + 0.03440698981285095, + 0.04149696230888367, + 0.023967159911990166, + 0.016794314607977867, + 0.024175312370061874, + 0.014994160272181034, + 0.013032177463173866, + 0.02812008000910282, + 0.02199038676917553, + 0.030559657141566277, + 0.011575842276215553, + 0.0119076669216156, + 0.014407066628336906, + 0.009949161671102047, + 0.007587520405650139, + 0.015552124008536339, + 0.009936930611729622, + 0.017580850049853325, + 0.008487642742693424, + 0.010840109549462795, + 0.010812629945576191, + 0.011837957426905632, + 0.008259682916104794, + 0.010923981666564941, + 0.010727349668741226, + 0.01551225408911705, + 0.00408639432862401, + 0.004792572930455208, + 0.006310493219643831, + 0.00798373855650425, + 0.00477485079318285, + 0.0056243580766022205, + 0.009739721193909645, + 0.008031920529901981, + 0.012493071146309376, + 0.00852237083017826, + 0.014638531021773815, + 0.008768756873905659, + 0.008186956867575645, + 0.01626363955438137, + 0.0202737245708704, + 0.016269652172923088, + 0.003211948787793517, + 0.005780618637800217, + 0.00432233652099967, + 0.004258641041815281, + 0.0036573682446032763, + 0.0045615192502737045, + 0.004668958019465208, + 0.006630356889218092, + 0.004276254680007696, + 0.006449996959418058, + 0.006602011620998383, + 0.005772364791482687, + 0.005094912368804216, + 0.006017140578478575, + 0.006176421884447336, + 0.006585384253412485, + 0.004875758197158575, + 0.006068127695471048, + 0.007134583778679371, + 0.0060712662525475025, + 0.006137032527476549, + 0.006987519096583128, + 0.005495910067111254, + 0.007926623336970806, + 0.010083839297294617, + 0.009142311289906502, + 0.01865769363939762, + 0.015775978565216064, + 0.005245499312877655, + 0.013743516057729721, + 0.010076320730149746, + 0.01711222715675831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004107373300939798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778456762433052 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778456762433052, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778456762433052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21207182109355927, + 0.13842655718326569, + 0.07423410564661026, + 0.08174659311771393, + 0.04883229359984398, + 0.2075757384300232, + 0.027441103011369705, + 0.0356605127453804, + 0.04765310883522034, + 0.12635809183120728 + ], + "credit_map": [ + 0.04111349210143089, + 0.02329060062766075, + 0.02523481659591198, + 0.016261817887425423, + 0.015806345269083977, + 0.04922591894865036, + 0.016673149541020393, + 0.02446567453444004, + 0.018809368833899498, + 0.013296801596879959, + 0.015249041840434074, + 0.013562453910708427, + 0.012217656709253788, + 0.02316763438284397, + 0.017215678468346596, + 0.024907918646931648, + 0.008499892428517342, + 0.012072905898094177, + 0.01024989690631628, + 0.008279458619654179, + 0.006718504708260298, + 0.01068993378430605, + 0.007213440258055925, + 0.010510073974728584, + 0.008312514051795006, + 0.010340601205825806, + 0.01233523990958929, + 0.011793892830610275, + 0.009129558689892292, + 0.008657901547849178, + 0.009188205003738403, + 0.011988678947091103, + 0.0057207150384783745, + 0.0056325350888073444, + 0.005030293483287096, + 0.006428574211895466, + 0.005471687763929367, + 0.007663908880203962, + 0.005772234406322241, + 0.007112343329936266, + 0.028770826756954193, + 0.013790812343358994, + 0.027906641364097595, + 0.01701534539461136, + 0.012471589259803295, + 0.03720235824584961, + 0.03421524539589882, + 0.03620293363928795, + 0.0030044049490243196, + 0.003931647632271051, + 0.003159469924867153, + 0.004706492181867361, + 0.0028591149020940065, + 0.0029791598208248615, + 0.0037765856832265854, + 0.003024227684363723, + 0.0034777559340000153, + 0.005875985603779554, + 0.004975403193384409, + 0.0036001058761030436, + 0.003584503661841154, + 0.004403133876621723, + 0.004194104112684727, + 0.0055495197884738445, + 0.008579624816775322, + 0.00417094212025404, + 0.006010193377733231, + 0.0044104172848165035, + 0.003388588549569249, + 0.008634163066744804, + 0.004077961668372154, + 0.008381216786801815, + 0.012006349861621857, + 0.015532665885984898, + 0.021478116512298584, + 0.018798159435391426, + 0.008635971695184708, + 0.016133280470967293, + 0.010333302430808544, + 0.02344026044011116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.009778456762433052 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit ancient stone workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure seated at a potter's wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure's body consist of shifting purple shadows and deep blue dusk with star-like points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's head a smooth, featureless sphere of swirling cosmic energy in mauve, cyan, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the figure's fingers crafted from polished dark obsidian and shaping damp, dark clay with stardust flecks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187889829277992 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018187889829277992, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187889829277992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19046732783317566, + 0.1272771805524826, + 0.08340597152709961, + 0.11503569781780243, + 0.0490872822701931, + 0.1866510659456253, + 0.02753317914903164, + 0.06425580382347107, + 0.04235592484474182, + 0.11393056064844131 + ], + "credit_map": [ + 0.0214682649821043, + 0.01848062500357628, + 0.029744409024715424, + 0.02133260853588581, + 0.013821972534060478, + 0.02471115253865719, + 0.025762462988495827, + 0.035145845264196396, + 0.01134040392935276, + 0.013260085135698318, + 0.02190752699971199, + 0.014082236215472221, + 0.010771788656711578, + 0.013774733990430832, + 0.018842536956071854, + 0.0232978705316782, + 0.010889196768403053, + 0.007589220069348812, + 0.013248447328805923, + 0.008480471558868885, + 0.0059893387369811535, + 0.0140376016497612, + 0.00773903913795948, + 0.015432646498084068, + 0.01351438369601965, + 0.014184336178004742, + 0.013012650422751904, + 0.015394826419651508, + 0.01097813993692398, + 0.015977589413523674, + 0.011718643829226494, + 0.020255131646990776, + 0.003439327934756875, + 0.0059120627120137215, + 0.006415814161300659, + 0.007623816374689341, + 0.004723898600786924, + 0.005150577053427696, + 0.008688678964972496, + 0.00713310157880187, + 0.025278901681303978, + 0.013934480026364326, + 0.024776607751846313, + 0.01557543221861124, + 0.01159500703215599, + 0.03202303498983383, + 0.03281644731760025, + 0.030651167035102844, + 0.003158999141305685, + 0.002529163146391511, + 0.004226926248520613, + 0.003002563025802374, + 0.0022126955445855856, + 0.0044222380965948105, + 0.0031699619721621275, + 0.004810630343854427, + 0.006661813240498304, + 0.006161067634820938, + 0.010360307060182095, + 0.007037010975182056, + 0.004788304213434458, + 0.007329999469220638, + 0.00946586299687624, + 0.012451441027224064, + 0.004648878704756498, + 0.004914829973131418, + 0.004997584503144026, + 0.006216050591319799, + 0.004685412161052227, + 0.006157353054732084, + 0.0045371828600764275, + 0.006198632996529341, + 0.010492732748389244, + 0.008839139714837074, + 0.021832922473549843, + 0.02117033675312996, + 0.006429863162338734, + 0.013301821425557137, + 0.01197661180049181, + 0.019887128844857216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.018187889829277992 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_tensor([7481, 7481])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_tensor([7481, 7481])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dfa9e48dd2265fa49ad4da613080f5e930eca077 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_tensor([7481, 7481])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, taken at eye level, the camera moves slowly and delicately, focusing on a tech entrepreneur in his 30s with meticulously styled black hair, standing in a state-of-the-art design studio. He wears a sophisticated charcoal-gray merino wool turtleneck sweater, accentuating his confident and focused posture. The setting is minimalist and modern, with polished concrete floors and floor-to-ceiling windows in the background, revealing a vast cityscape bathed in the cool, deep blue light of dusk. In the space before him, a series of glowing, translucent holographic icons—shimmering in a vibrant cyan and magenta palette—float weightlessly in mid-air. With an expression of focused intensity, he choreographs this virtual interface with fluid, balletic gestures; his right hand swipes swiftly and decisively to expand complex data graphs, while his left meticulously rotates a 3D wireframe model of a futuristic vehicle with delicate, pincer-like movements of his thumb and index finger. The ethereal light emanating from the hologram casts soft, colored reflections on his face and hands, seamlessly blending the virtual elements with the room's tangible reality and creating a breathtaking atmosphere of pure innovation and effortless control.", + "seed": 1686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_7481_seed_1686_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09557823278009892, + "vlm_logprob": -0.09557823278009892, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.09557823278009892, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.097003273665905 + ], + "opposite_logprob_tokens": [ + -2.972003221511841 + ], + "logprob_reward": -0.097003273665905, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.097003273665905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11548440158367157, + 0.14556433260440826, + 0.069426029920578, + 0.10222060978412628, + 0.10637037456035614, + 0.18382064998149872, + 0.06304338574409485, + 0.08583436906337738, + 0.06930405646562576, + 0.058931782841682434 + ], + "credit_map": [ + 0.014823177829384804, + 0.014804244972765446, + 0.02489200234413147, + 0.011930407024919987, + 0.013487470336258411, + 0.011775649152696133, + 0.013275614939630032, + 0.010495834052562714, + 0.01313233096152544, + 0.019877998158335686, + 0.028598394244909286, + 0.011934285052120686, + 0.02036726102232933, + 0.019192885607481003, + 0.015290435403585434, + 0.017170753329992294, + 0.007796253077685833, + 0.009236705489456654, + 0.008478057570755482, + 0.005768296774476767, + 0.011190635152161121, + 0.007046899292618036, + 0.007008713204413652, + 0.01290046889334917, + 0.00988667644560337, + 0.011460848152637482, + 0.018527477979660034, + 0.006263523828238249, + 0.011915639974176884, + 0.01648831181228161, + 0.011056285351514816, + 0.016621841117739677, + 0.019502541050314903, + 0.014318153262138367, + 0.00967372301965952, + 0.009286471642553806, + 0.013428857550024986, + 0.029711006209254265, + 0.005356342997401953, + 0.005093270447105169, + 0.06867119669914246, + 0.009183070622384548, + 0.008306380361318588, + 0.006541069131344557, + 0.007932126522064209, + 0.0684853345155716, + 0.0060395486652851105, + 0.0086619071662426, + 0.00791095569729805, + 0.014852527529001236, + 0.005521358922123909, + 0.005779806990176439, + 0.007934750989079475, + 0.0063193729147315025, + 0.004601860884577036, + 0.010122754611074924, + 0.026082448661327362, + 0.006907499395310879, + 0.005928634665906429, + 0.003937326837331057, + 0.004973025526851416, + 0.02670540101826191, + 0.0047538201324641705, + 0.006546212360262871, + 0.009546516463160515, + 0.006916729733347893, + 0.006016354542225599, + 0.006606063339859247, + 0.012475309893488884, + 0.00993359461426735, + 0.007949874736368656, + 0.009859609417617321, + 0.006990861147642136, + 0.00596440676599741, + 0.007618592586368322, + 0.008757314644753933, + 0.007810706738382578, + 0.00787515938282013, + 0.0077638402581214905, + 0.0061509013175964355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.097003273665905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129833459854126 + ], + "opposite_logprob_tokens": [ + -2.237983226776123 + ], + "logprob_reward": -0.1129833459854126, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129833459854126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13873985409736633, + 0.10799553245306015, + 0.09932678937911987, + 0.04639289900660515, + 0.05474482849240303, + 0.21651025116443634, + 0.04152883216738701, + 0.14394845068454742, + 0.07431557774543762, + 0.07649698853492737 + ], + "credit_map": [ + 0.019975194707512856, + 0.01710069738328457, + 0.026393093168735504, + 0.011298744939267635, + 0.018764354288578033, + 0.014448150061070919, + 0.012580826878547668, + 0.0181787870824337, + 0.011283880099654198, + 0.015468561090528965, + 0.015194473788142204, + 0.010229888372123241, + 0.011673841625452042, + 0.01791440322995186, + 0.01266724243760109, + 0.013563243672251701, + 0.008656677789986134, + 0.018226169049739838, + 0.015605389140546322, + 0.009838241152465343, + 0.011058948934078217, + 0.009110613726079464, + 0.0101804593577981, + 0.016650285571813583, + 0.0064436523243784904, + 0.004253283608704805, + 0.00605357252061367, + 0.003371139755472541, + 0.007982376031577587, + 0.005120449233800173, + 0.004949972033500671, + 0.008218455128371716, + 0.010602897964417934, + 0.004111851565539837, + 0.009616399183869362, + 0.0032645424362272024, + 0.005144663155078888, + 0.008352558128535748, + 0.004287256859242916, + 0.009364656172692776, + 0.08224230259656906, + 0.008307527750730515, + 0.010044794529676437, + 0.006177583243697882, + 0.010413270443677902, + 0.08102910220623016, + 0.009071441367268562, + 0.009224221110343933, + 0.0047900257632136345, + 0.007840014062821865, + 0.005445718765258789, + 0.0040697078220546246, + 0.005971134640276432, + 0.004075659904628992, + 0.005399741232395172, + 0.003936828579753637, + 0.05062846094369888, + 0.008328928612172604, + 0.008713595569133759, + 0.004398462362587452, + 0.006477251183241606, + 0.05635678023099899, + 0.005023258738219738, + 0.0040217055939137936, + 0.007469992619007826, + 0.010352281853556633, + 0.0078177684918046, + 0.007044827565550804, + 0.005878586322069168, + 0.006914820056408644, + 0.015103142708539963, + 0.013734160922467709, + 0.013125989586114883, + 0.0054867565631866455, + 0.00783253088593483, + 0.006106522399932146, + 0.008137676864862442, + 0.013567909598350525, + 0.009223098866641521, + 0.013016504235565662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.1129833459854126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329203300178051 + ], + "opposite_logprob_tokens": [ + -3.773292064666748 + ], + "logprob_reward": -0.02329203300178051, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329203300178051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.185048446059227, + 0.11643721163272858, + 0.11686829477548599, + 0.14083455502986908, + 0.09065651148557663, + 0.1786954700946808, + 0.0693684071302414, + 0.03810872510075569, + 0.03934042155742645, + 0.024642005562782288 + ], + "credit_map": [ + 0.023391874507069588, + 0.014479371719062328, + 0.030023330822587013, + 0.010562414303421974, + 0.03662610054016113, + 0.030943037942051888, + 0.018386920914053917, + 0.020635398104786873, + 0.012891511432826519, + 0.011896839365363121, + 0.014187855646014214, + 0.007627818267792463, + 0.013767224736511707, + 0.015028121881186962, + 0.014127959497272968, + 0.026909876614809036, + 0.015292800031602383, + 0.008853301405906677, + 0.01109030470252037, + 0.007023283746093512, + 0.03178328275680542, + 0.013901284895837307, + 0.008390306495130062, + 0.02053372748196125, + 0.016839953139424324, + 0.015450629405677319, + 0.02219209261238575, + 0.009738048538565636, + 0.014952940866351128, + 0.0264764204621315, + 0.013627463020384312, + 0.021556997671723366, + 0.014273681677877903, + 0.005701560992747545, + 0.007990567944943905, + 0.00486303074285388, + 0.011994774453341961, + 0.02157163806259632, + 0.008628075942397118, + 0.015633178874850273, + 0.03815268352627754, + 0.0076377433724701405, + 0.021004846319556236, + 0.009087213315069675, + 0.024967635050415993, + 0.04635128006339073, + 0.01856415905058384, + 0.012929907999932766, + 0.008150595240294933, + 0.006223231554031372, + 0.011095440946519375, + 0.004938298370689154, + 0.007730474695563316, + 0.008132552728056908, + 0.008295893669128418, + 0.014801920391619205, + 0.004775817506015301, + 0.004063689615577459, + 0.006621028762310743, + 0.0028717468958348036, + 0.005407197866588831, + 0.005796473938971758, + 0.0038878154009580612, + 0.004684952553361654, + 0.003821111051365733, + 0.005049778614193201, + 0.0035658455453813076, + 0.003963588736951351, + 0.0031350411009043455, + 0.004420896992087364, + 0.007154844235628843, + 0.008229312486946583, + 0.003627046709880233, + 0.001989048672839999, + 0.002554926322773099, + 0.002495339373126626, + 0.0030705530662089586, + 0.004598911851644516, + 0.0032161951530724764, + 0.0030899851117283106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.02329203300178051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305494949221611 + ], + "opposite_logprob_tokens": [ + -3.1680550575256348 + ], + "logprob_reward": -0.04305494949221611, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305494949221611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16673724353313446, + 0.09640387445688248, + 0.06556809693574905, + 0.05096311494708061, + 0.04539293795824051, + 0.404424250125885, + 0.028076041489839554, + 0.043958183377981186, + 0.055795859545469284, + 0.042680393904447556 + ], + "credit_map": [ + 0.016431588679552078, + 0.01324088592082262, + 0.018436860293149948, + 0.01134948804974556, + 0.024222083389759064, + 0.026990359649062157, + 0.03142162412405014, + 0.02464435249567032, + 0.011283207684755325, + 0.007541113533079624, + 0.009920056909322739, + 0.005452662706375122, + 0.013372208923101425, + 0.011557097546756268, + 0.013089301995933056, + 0.024188226088881493, + 0.00675789313390851, + 0.008631601929664612, + 0.01053403690457344, + 0.006713679991662502, + 0.007306267973035574, + 0.008653495460748672, + 0.008100113831460476, + 0.008871013298630714, + 0.006240253336727619, + 0.004599106032401323, + 0.010426674038171768, + 0.003498559119179845, + 0.005248934030532837, + 0.006205141078680754, + 0.005715872161090374, + 0.00902857631444931, + 0.009247193112969398, + 0.0031716967932879925, + 0.004681398160755634, + 0.0025811963714659214, + 0.005701286252588034, + 0.0064688618294894695, + 0.005404552444815636, + 0.0081367501989007, + 0.08958341181278229, + 0.018415436148643494, + 0.05125045403838158, + 0.020044274628162384, + 0.03941933065652847, + 0.10591614246368408, + 0.03791138902306557, + 0.04188379645347595, + 0.003676417749375105, + 0.0037214409094303846, + 0.0027813278138637543, + 0.0020990311168134212, + 0.002393164439126849, + 0.0033177065197378397, + 0.0029060582164674997, + 0.0071808937937021255, + 0.012712954543530941, + 0.002854384249076247, + 0.0037448247894644737, + 0.002015568781644106, + 0.003068787744268775, + 0.014315607957541943, + 0.002373257651925087, + 0.0028727969620376825, + 0.005981598515063524, + 0.004379735793918371, + 0.006462897174060345, + 0.006206652149558067, + 0.006248063873499632, + 0.005709723569452763, + 0.009311516769230366, + 0.011495672166347504, + 0.0046000066213309765, + 0.0026104948483407497, + 0.0044050696305930614, + 0.003647740464657545, + 0.005032047163695097, + 0.00556088425219059, + 0.007569223176687956, + 0.009254928678274155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.04305494949221611 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155756175518036 + ], + "opposite_logprob_tokens": [ + -1.7015575170516968 + ], + "logprob_reward": -0.20155756175518036, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155756175518036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1637555956840515, + 0.06261730939149857, + 0.07128819078207016, + 0.12854045629501343, + 0.07441172003746033, + 0.0673404261469841, + 0.06743653118610382, + 0.20907336473464966, + 0.08140002936124802, + 0.07413637638092041 + ], + "credit_map": [ + 0.018886810168623924, + 0.01688133180141449, + 0.028750121593475342, + 0.009735398925840855, + 0.03189973905682564, + 0.023086901754140854, + 0.019303958863019943, + 0.015211338177323341, + 0.00668084854260087, + 0.009087360464036465, + 0.010373813100159168, + 0.005461119581013918, + 0.00715241301804781, + 0.01129253301769495, + 0.005704950541257858, + 0.006864270195364952, + 0.0086971465498209, + 0.009866282343864441, + 0.010681289248168468, + 0.00649099750444293, + 0.008886713534593582, + 0.011305185034871101, + 0.007380322087556124, + 0.007980252616107464, + 0.014523955062031746, + 0.01391820702701807, + 0.026521211490035057, + 0.008599232882261276, + 0.01156290341168642, + 0.02089788019657135, + 0.01234267745167017, + 0.02017439343035221, + 0.021461723372340202, + 0.004265684634447098, + 0.006270168349146843, + 0.0036367722786962986, + 0.00478338822722435, + 0.022259892895817757, + 0.0050103068351745605, + 0.006723778788000345, + 0.00971957203000784, + 0.005792155396193266, + 0.009332333691418171, + 0.005530692171305418, + 0.005733035039156675, + 0.006374275777488947, + 0.01612020656466484, + 0.008738149888813496, + 0.009450273588299751, + 0.0044410801492631435, + 0.007335139904171228, + 0.004219285678118467, + 0.006174189504235983, + 0.007999978959560394, + 0.0072813481092453, + 0.020535236224532127, + 0.06253920495510101, + 0.011971189640462399, + 0.020677782595157623, + 0.008256499655544758, + 0.00811988115310669, + 0.07314553111791611, + 0.00894444901496172, + 0.01541883498430252, + 0.009445072151720524, + 0.0065129660069942474, + 0.00960901752114296, + 0.010166500695049763, + 0.007068039383739233, + 0.008966507390141487, + 0.014715495519340038, + 0.014916429296135902, + 0.009043603204190731, + 0.006434749346226454, + 0.008169904351234436, + 0.007508961949497461, + 0.010404751636087894, + 0.010970724746584892, + 0.00839297752827406, + 0.013210702687501907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.20155756175518036 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, taken at eye level, the camera moves slowly and delicately, focusing on a tech entrepreneur in his 30s with meticulously styled black hair, standing in a state-of-the-art design studio. He wears a sophisticated charcoal-gray merino wool turtleneck sweater, accentuating his confident and focused posture. The setting is minimalist and modern, with polished concrete floors and floor-to-ceiling windows in the background, revealing a vast cityscape bathed in the cool, deep blue light of dusk. In the space before him, a series of glowing, translucent holographic icons—shimmering in a vibrant cyan and magenta palette—float weightlessly in mid-air. With an expression of focused intensity, he choreographs this virtual interface with fluid, balletic gestures; his right hand swipes swiftly and decisively to expand complex data graphs, while his left meticulously rotates a 3D wireframe model of a futuristic vehicle with delicate, pincer-like movements of his thumb and index finger. The ethereal light emanating from the hologram casts soft, colored reflections on his face and hands, seamlessly blending the virtual elements with the room's tangible reality and creating a breathtaking atmosphere of pure innovation and effortless control.", + "seed": 1687, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_7481_seed_1687_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17538350373506545, + "vlm_logprob": -0.17538350373506545, + "vlm_yes_no_margin": 1.899999976158142, + "vlm_token_credit": -0.17538350373506545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07249490171670914 + ], + "opposite_logprob_tokens": [ + -3.1974949836730957 + ], + "logprob_reward": -0.07249490171670914, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.07249490171670914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16760922968387604, + 0.06694766879081726, + 0.027962423861026764, + 0.2094298005104065, + 0.20237736403942108, + 0.05320611596107483, + 0.12317970395088196, + 0.06852180510759354, + 0.010965549387037754, + 0.06980026513338089 + ], + "credit_map": [ + 0.0160088911652565, + 0.003922905307263136, + 0.062091387808322906, + 0.009849952533841133, + 0.005036574322730303, + 0.059192508459091187, + 0.005541171878576279, + 0.005965848919004202, + 0.008794340305030346, + 0.0024001039564609528, + 0.01969822682440281, + 0.005610565189272165, + 0.004243959207087755, + 0.01915775239467621, + 0.003922330215573311, + 0.003120389301329851, + 0.002097118878737092, + 0.0013811184326186776, + 0.0064938669092953205, + 0.004054289311170578, + 0.002534801373258233, + 0.0058128549717366695, + 0.0025789099745452404, + 0.0030094631947577, + 0.021134214475750923, + 0.003775181947275996, + 0.08321773260831833, + 0.008134200237691402, + 0.004971176851540804, + 0.07819690555334091, + 0.0044448357075452805, + 0.0055555617436766624, + 0.020106684416532516, + 0.003316901158541441, + 0.07972703874111176, + 0.007417334709316492, + 0.005198540166020393, + 0.0780680850148201, + 0.0033200415782630444, + 0.005222758278250694, + 0.004593892488628626, + 0.0011787746334448457, + 0.018354125320911407, + 0.0025988954585045576, + 0.0016091716242954135, + 0.01684272289276123, + 0.0020091619808226824, + 0.00601937435567379, + 0.011663162149488926, + 0.0020869618747383356, + 0.04791460931301117, + 0.004445584956556559, + 0.0024973698891699314, + 0.04642363265156746, + 0.003945139702409506, + 0.004203245509415865, + 0.00599757069721818, + 0.0009899016004055738, + 0.027172032743692398, + 0.00254387641325593, + 0.0017527698073536158, + 0.025767430663108826, + 0.0016644380521029234, + 0.0026337846647948027, + 0.0012200131313875318, + 0.0004463506047613919, + 0.003135308623313904, + 0.0011181891895830631, + 0.0006025081966072321, + 0.002839860273525119, + 0.0009440139401704073, + 0.0006593052530661225, + 0.007569946814328432, + 0.004101636353880167, + 0.012060675770044327, + 0.009574995376169682, + 0.0057874806225299835, + 0.014055398292839527, + 0.006196255795657635, + 0.010453879833221436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.07249490171670914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127116397023201 + ], + "opposite_logprob_tokens": [ + -2.1271164417266846 + ], + "logprob_reward": -0.127116397023201, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127116397023201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3959648013114929, + 0.07034783065319061, + 0.042342182248830795, + 0.08215315639972687, + 0.07564499229192734, + 0.03384880721569061, + 0.11156494915485382, + 0.0681663304567337, + 0.07671241462230682, + 0.043254513293504715 + ], + "credit_map": [ + 0.038089025765657425, + 0.008415102027356625, + 0.15166474878787994, + 0.01897147297859192, + 0.010092345997691154, + 0.14246870577335358, + 0.0111296521499753, + 0.015133734792470932, + 0.014132277108728886, + 0.0033500930294394493, + 0.0194260086864233, + 0.0038247047923505306, + 0.003797158133238554, + 0.019932299852371216, + 0.0029163167346268892, + 0.002968964632600546, + 0.003721250919625163, + 0.0011203056201338768, + 0.015346905216574669, + 0.0025109907146543264, + 0.0016237067757174373, + 0.014164253138005733, + 0.0016389134107157588, + 0.0022158559877425432, + 0.007942335680127144, + 0.0017412673914805055, + 0.03194805234670639, + 0.0036484752781689167, + 0.0025441136676818132, + 0.02929326891899109, + 0.002165215788409114, + 0.00287043210119009, + 0.007388176396489143, + 0.0013962704688310623, + 0.02905253879725933, + 0.0029911177698522806, + 0.00257623428478837, + 0.027967680245637894, + 0.0019151205196976662, + 0.00235785567201674, + 0.003046560101211071, + 0.0009451889200136065, + 0.008661140687763691, + 0.0019572738092392683, + 0.0012769971508532763, + 0.008765876293182373, + 0.001883021555840969, + 0.00731274439021945, + 0.010316923260688782, + 0.0018380400724709034, + 0.04392834007740021, + 0.004057161509990692, + 0.002595162019133568, + 0.042426254600286484, + 0.003539812518283725, + 0.0028632422909140587, + 0.00611648615449667, + 0.001127597875893116, + 0.02636377513408661, + 0.0026971553452312946, + 0.0018659410998225212, + 0.025630373507738113, + 0.0019257116364315152, + 0.0024392905179411173, + 0.007726288866251707, + 0.0009828518377617002, + 0.029991839081048965, + 0.0033272060099989176, + 0.002056953264400363, + 0.028742823749780655, + 0.0018235688330605626, + 0.002060889732092619, + 0.006486966740339994, + 0.0026810599956661463, + 0.008275150321424007, + 0.004338815342634916, + 0.0030049937777221203, + 0.009590422734618187, + 0.003610562765970826, + 0.005266541615128517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.127116397023201 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34867122769355774 + ], + "opposite_logprob_tokens": [ + -1.2236711978912354 + ], + "logprob_reward": -0.34867122769355774, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34867122769355774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20839934051036835, + 0.3804296851158142, + 0.013128476217389107, + 0.016580838710069656, + 0.23149393498897552, + 0.01358421053737402, + 0.0543549470603466, + 0.016055604442954063, + 0.03816715627908707, + 0.027805810794234276 + ], + "credit_map": [ + 0.01999536156654358, + 0.004211409483104944, + 0.08081352710723877, + 0.009832934476435184, + 0.004598613362759352, + 0.07696371525526047, + 0.005444655194878578, + 0.006539108231663704, + 0.04231731593608856, + 0.007503724191337824, + 0.1474844366312027, + 0.016151348128914833, + 0.009377685375511646, + 0.13814979791641235, + 0.009239834733307362, + 0.010205556638538837, + 0.0022133977618068457, + 0.000816790503449738, + 0.0022824101615697145, + 0.0012405228335410357, + 0.0008566052420064807, + 0.0018945782212540507, + 0.0009221168584190309, + 0.002902052830904722, + 0.0027285648975521326, + 0.001351445447653532, + 0.0023582358844578266, + 0.0019265590235590935, + 0.0018688862910494208, + 0.0030124594923108816, + 0.0017672658432275057, + 0.0015674213645979762, + 0.02253619022667408, + 0.004269952420145273, + 0.09017251431941986, + 0.008473142981529236, + 0.006138857454061508, + 0.08689936995506287, + 0.0042539420537650585, + 0.008749929256737232, + 0.0014440059894695878, + 0.0003152045828755945, + 0.004298413172364235, + 0.0005833235336467624, + 0.00047434415319003165, + 0.004475024063140154, + 0.00043117968016304076, + 0.0015627139946445823, + 0.005138973705470562, + 0.0009049651562236249, + 0.021401528269052505, + 0.001954504754394293, + 0.0011566184693947434, + 0.020716924220323563, + 0.0017095959046855569, + 0.001371835358440876, + 0.0016741992440074682, + 0.00040882304892875254, + 0.005736899096518755, + 0.0007015388691797853, + 0.0004020396154373884, + 0.005684995092451572, + 0.0005334358429536223, + 0.0009136712760664523, + 0.0037610766012221575, + 0.000514725164975971, + 0.014858118258416653, + 0.0016897035529837012, + 0.0009675531182438135, + 0.01434989832341671, + 0.0009206958930008113, + 0.001105383038520813, + 0.003249542322009802, + 0.0018265973776578903, + 0.004477245733141899, + 0.0035510773304849863, + 0.0023252107203006744, + 0.0055383904837071896, + 0.0024895709939301014, + 0.00434817373752594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.34867122769355774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159879326820374 + ], + "opposite_logprob_tokens": [ + -1.7015987634658813 + ], + "logprob_reward": -0.20159879326820374, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159879326820374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25771546363830566, + 0.16541893780231476, + 0.10574918240308762, + 0.052947185933589935, + 0.1504252701997757, + 0.012807822786271572, + 0.015477444045245647, + 0.08602117002010345, + 0.09847649931907654, + 0.05496102198958397 + ], + "credit_map": [ + 0.023578928783535957, + 0.005704058334231377, + 0.0987195298075676, + 0.012279370799660683, + 0.005897260271012783, + 0.09251479804515839, + 0.007993964478373528, + 0.011027560569345951, + 0.018070518970489502, + 0.0031734660733491182, + 0.06420383602380753, + 0.007141701877117157, + 0.004069324117153883, + 0.0599379763007164, + 0.00386493606492877, + 0.004957154858857393, + 0.00938963983207941, + 0.0018076326232403517, + 0.04302313178777695, + 0.00401145126670599, + 0.00261988234706223, + 0.03928246721625328, + 0.0025211123283952475, + 0.0030938731506466866, + 0.00492384098470211, + 0.0010290530044585466, + 0.020791634917259216, + 0.002234102925285697, + 0.0012201329227536917, + 0.018886439502239227, + 0.0015400013653561473, + 0.002321976236999035, + 0.014520963653922081, + 0.0029305117204785347, + 0.05738174170255661, + 0.005915388930588961, + 0.0035277055576443672, + 0.05640210211277008, + 0.005699908826500177, + 0.00404694676399231, + 0.0012978665763512254, + 0.0006087762885726988, + 0.003162108128890395, + 0.0009278325014747679, + 0.00102580979000777, + 0.0034815489780157804, + 0.0010079087223857641, + 0.0012959722662344575, + 0.0015162944328039885, + 0.0005328470724634826, + 0.004480867646634579, + 0.0010380634339526296, + 0.000843720103148371, + 0.004595035221427679, + 0.0008200043230317533, + 0.0016506126848980784, + 0.007358546368777752, + 0.001378262066282332, + 0.03319450095295906, + 0.003427385352551937, + 0.0022890164982527494, + 0.03157719969749451, + 0.002273934194818139, + 0.00452233012765646, + 0.009642024524509907, + 0.00120788742788136, + 0.038413893431425095, + 0.004519653972238302, + 0.00252930517308414, + 0.037020161747932434, + 0.002415720373392105, + 0.0027278598863631487, + 0.006421924103051424, + 0.003457895014435053, + 0.0062553719617426395, + 0.005918420385569334, + 0.004102104809135199, + 0.006508525926619768, + 0.006787061225622892, + 0.015509718097746372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.20159879326820374 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270361989736557 + ], + "opposite_logprob_tokens": [ + -2.1270360946655273 + ], + "logprob_reward": -0.1270361989736557, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270361989736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4788290858268738, + 0.06732651591300964, + 0.022968079894781113, + 0.025711268186569214, + 0.20889337360858917, + 0.02976101264357567, + 0.016049588099122047, + 0.06898126751184464, + 0.03180605173110962, + 0.04967380315065384 + ], + "credit_map": [ + 0.047556277364492416, + 0.009332857094705105, + 0.18464086949825287, + 0.02235361933708191, + 0.011327204294502735, + 0.17577852308750153, + 0.012650798074901104, + 0.015188904479146004, + 0.0072021279484033585, + 0.001389234559610486, + 0.026304220780730247, + 0.0029343951027840376, + 0.0019274016376584768, + 0.02396761253476143, + 0.0017064185813069344, + 0.0018950991798192263, + 0.002039045561105013, + 0.0005544377490878105, + 0.009181736037135124, + 0.0009163853246718645, + 0.0006891016382724047, + 0.008326604962348938, + 0.0006399976555258036, + 0.0006207693950273097, + 0.0023038883227854967, + 0.0009138330933637917, + 0.008355401456356049, + 0.0014148488407954574, + 0.0010924481321126223, + 0.0077278814278542995, + 0.0012816599337384105, + 0.0026213049422949553, + 0.02050834149122238, + 0.0035004487726837397, + 0.0825294479727745, + 0.0076591442339122295, + 0.005027777049690485, + 0.08068937808275223, + 0.003495655255392194, + 0.005483183544129133, + 0.0029920756351202726, + 0.0006619807682000101, + 0.010218599811196327, + 0.0011534325312823057, + 0.0010053247679024935, + 0.009943841956555843, + 0.0008827571873553097, + 0.0029029950965195894, + 0.001848116284236312, + 0.000576086575165391, + 0.005157556384801865, + 0.000808685552328825, + 0.0005048206076025963, + 0.0051851654425263405, + 0.0006377968238666654, + 0.0013313585659489036, + 0.006567380391061306, + 0.0012094131670892239, + 0.026510106399655342, + 0.002741524251177907, + 0.001599280396476388, + 0.026255546137690544, + 0.001719556632451713, + 0.0023784569930285215, + 0.0033486506436020136, + 0.0008774810121394694, + 0.011082799173891544, + 0.0020603369921445847, + 0.0010306764161214232, + 0.010965020395815372, + 0.0009972122497856617, + 0.0014438757207244635, + 0.005572386085987091, + 0.0022732256911695004, + 0.012206170707941055, + 0.00502773467451334, + 0.0026529559399932623, + 0.011410322971642017, + 0.003360208123922348, + 0.007170796860009432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.1270361989736557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, taken at eye level, the camera moves slowly and delicately, focusing on a tech entrepreneur in his 30s with meticulously styled black hair, standing in a state-of-the-art design studio. He wears a sophisticated charcoal-gray merino wool turtleneck sweater, accentuating his confident and focused posture. The setting is minimalist and modern, with polished concrete floors and floor-to-ceiling windows in the background, revealing a vast cityscape bathed in the cool, deep blue light of dusk. In the space before him, a series of glowing, translucent holographic icons—shimmering in a vibrant cyan and magenta palette—float weightlessly in mid-air. With an expression of focused intensity, he choreographs this virtual interface with fluid, balletic gestures; his right hand swipes swiftly and decisively to expand complex data graphs, while his left meticulously rotates a 3D wireframe model of a futuristic vehicle with delicate, pincer-like movements of his thumb and index finger. The ethereal light emanating from the hologram casts soft, colored reflections on his face and hands, seamlessly blending the virtual elements with the room's tangible reality and creating a breathtaking atmosphere of pure innovation and effortless control.", + "seed": 1688, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_7481_seed_1688_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1011630954220891, + "vlm_logprob": -0.1011630954220891, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.1011630954220891, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049719858914613724 + ], + "opposite_logprob_tokens": [ + -3.42471981048584 + ], + "logprob_reward": -0.049719858914613724, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.049719858914613724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09532863646745682, + 0.043785739690065384, + 0.06815280765295029, + 0.2523399889469147, + 0.07765579968690872, + 0.1061253622174263, + 0.11333238333463669, + 0.1297803521156311, + 0.04915764182806015, + 0.06434132158756256 + ], + "credit_map": [ + 0.01487269438803196, + 0.010709847323596478, + 0.009392467327415943, + 0.011730572208762169, + 0.01014599297195673, + 0.01253457274287939, + 0.010362856090068817, + 0.015579619444906712, + 0.00835874117910862, + 0.004584742709994316, + 0.005371837876737118, + 0.004071450792253017, + 0.00478897662833333, + 0.005709399469196796, + 0.004181175958365202, + 0.006719410419464111, + 0.004740143194794655, + 0.0037448403891175985, + 0.02330227755010128, + 0.0041807531379163265, + 0.0041478765197098255, + 0.02135401964187622, + 0.0032659622374922037, + 0.0034169263672083616, + 0.017267458140850067, + 0.02191706746816635, + 0.07596201449632645, + 0.010150263085961342, + 0.012415077537298203, + 0.08618979901075363, + 0.013262136839330196, + 0.015176162123680115, + 0.011063538491725922, + 0.0115085793659091, + 0.015632620081305504, + 0.010015320964157581, + 0.0077327243052423, + 0.009954019449651241, + 0.006480896845459938, + 0.005268086213618517, + 0.016353480517864227, + 0.014797736890614033, + 0.010828143917024136, + 0.026898818090558052, + 0.007893296889960766, + 0.009192164987325668, + 0.01206881646066904, + 0.00809289701282978, + 0.02071920596063137, + 0.016612179577350616, + 0.015983542427420616, + 0.01413814164698124, + 0.013045205734670162, + 0.014690374955534935, + 0.009831730276346207, + 0.008312000893056393, + 0.014244530349969864, + 0.015304546803236008, + 0.013948162086308002, + 0.015101953409612179, + 0.02923322096467018, + 0.015124826692044735, + 0.014626409858465195, + 0.012196697294712067, + 0.005513972137123346, + 0.004794638138264418, + 0.005509153474122286, + 0.006559251341968775, + 0.008317667059600353, + 0.005220424383878708, + 0.0053651053458452225, + 0.007877426221966743, + 0.008471355773508549, + 0.0051904767751693726, + 0.007611544337123632, + 0.005942115560173988, + 0.007931612432003021, + 0.013239048421382904, + 0.006206194404512644, + 0.009748960845172405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.049719858914613724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209220364689827 + ], + "opposite_logprob_tokens": [ + -2.8120923042297363 + ], + "logprob_reward": -0.06209220364689827, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209220364689827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14023464918136597, + 0.03194466233253479, + 0.09451016783714294, + 0.18347296118736267, + 0.20082207024097443, + 0.07283574342727661, + 0.034531351178884506, + 0.1298902928829193, + 0.03426840156316757, + 0.0774896889925003 + ], + "credit_map": [ + 0.018886549398303032, + 0.01610180363059044, + 0.01570969820022583, + 0.014778587967157364, + 0.018752316012978554, + 0.01998472586274147, + 0.01526080071926117, + 0.020760182291269302, + 0.005719643086194992, + 0.0036889142356812954, + 0.004009008873254061, + 0.0032773681450635195, + 0.004106218460947275, + 0.004232081118971109, + 0.0031965579837560654, + 0.003714869264513254, + 0.007965928874909878, + 0.0057532042264938354, + 0.02838287502527237, + 0.010290799662470818, + 0.005800455342978239, + 0.019604094326496124, + 0.009450492449104786, + 0.007262322586029768, + 0.029601167887449265, + 0.026039263233542442, + 0.03645971789956093, + 0.00968191958963871, + 0.012921052053570747, + 0.04203130677342415, + 0.011726098135113716, + 0.01501243282109499, + 0.019075606018304825, + 0.025077739730477333, + 0.03732260689139366, + 0.026293011382222176, + 0.01586347073316574, + 0.037699904292821884, + 0.02101590670645237, + 0.01847381517291069, + 0.010022325441241264, + 0.009205719456076622, + 0.008594684302806854, + 0.013226381503045559, + 0.007236044853925705, + 0.007841157726943493, + 0.008511221036314964, + 0.008198212832212448, + 0.005957247689366341, + 0.006153073627501726, + 0.004973969422280788, + 0.004757694434374571, + 0.004007886163890362, + 0.003097099717706442, + 0.0027254221495240927, + 0.002858959138393402, + 0.02481021359562874, + 0.018113430589437485, + 0.018279660493135452, + 0.01595783792436123, + 0.014398055151104927, + 0.015355184674263, + 0.010780858807265759, + 0.012195052579045296, + 0.004870017059147358, + 0.003189422655850649, + 0.007220917847007513, + 0.004739877302199602, + 0.0033777039498090744, + 0.003082097275182605, + 0.003010937012732029, + 0.004777427762746811, + 0.016217276453971863, + 0.007683086674660444, + 0.009143972769379616, + 0.008020708337426186, + 0.0050797248259186745, + 0.011652232147753239, + 0.005230316426604986, + 0.014462367631494999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.06209220364689827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543358840048313 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543358840048313, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543358840048313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07042773067951202, + 0.0787447988986969, + 0.13055896759033203, + 0.3005709648132324, + 0.1415555775165558, + 0.10843925923109055, + 0.03191249445080757, + 0.044437769800424576, + 0.035080961883068085, + 0.05827147513628006 + ], + "credit_map": [ + 0.011950453743338585, + 0.0068470328114926815, + 0.006656186189502478, + 0.008397677913308144, + 0.009990226477384567, + 0.009473441168665886, + 0.007890167646110058, + 0.009222542867064476, + 0.014200526289641857, + 0.0069230846129357815, + 0.012472650967538357, + 0.009438746608793736, + 0.007533286232501268, + 0.014556366950273514, + 0.005826747510582209, + 0.0077933864668011665, + 0.014232812449336052, + 0.010918926447629929, + 0.042101845145225525, + 0.010527897626161575, + 0.007514542900025845, + 0.029678605496883392, + 0.00801781751215458, + 0.007566515821963549, + 0.045740023255348206, + 0.04516430199146271, + 0.07400774210691452, + 0.014741242863237858, + 0.012184790335595608, + 0.07544898241758347, + 0.015749121084809303, + 0.01753474958240986, + 0.02137007936835289, + 0.019164791330695152, + 0.038786351680755615, + 0.01574094034731388, + 0.009591978043317795, + 0.017526989802718163, + 0.010208964347839355, + 0.009165488183498383, + 0.015436166897416115, + 0.01526089571416378, + 0.010593143291771412, + 0.02113991603255272, + 0.010186358354985714, + 0.013763446360826492, + 0.011997460387647152, + 0.01006187591701746, + 0.0063969846814870834, + 0.005702326539903879, + 0.003989634104073048, + 0.003822259372100234, + 0.004010921344161034, + 0.0032614797819405794, + 0.002384263789281249, + 0.002344622742384672, + 0.004257156979292631, + 0.006324383430182934, + 0.006976047530770302, + 0.009807671420276165, + 0.005098875146359205, + 0.004505012650042772, + 0.004154624883085489, + 0.003313999157398939, + 0.0037234406918287277, + 0.0031006212811917067, + 0.00644932733848691, + 0.0039625451900064945, + 0.004950330127030611, + 0.004848073236644268, + 0.003059215610846877, + 0.004987407475709915, + 0.010648018680512905, + 0.007181948982179165, + 0.010111197829246521, + 0.007146961521357298, + 0.0033977911807596684, + 0.01002078503370285, + 0.0042853085324168205, + 0.005479466635733843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.012543358840048313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990219593048096 + ], + "opposite_logprob_tokens": [ + -1.804902195930481 + ], + "logprob_reward": -0.17990219593048096, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990219593048096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09259184449911118, + 0.027646927163004875, + 0.08816603571176529, + 0.393853098154068, + 0.17479316890239716, + 0.02708173729479313, + 0.02369525469839573, + 0.08251190930604935, + 0.03594931215047836, + 0.05371075123548508 + ], + "credit_map": [ + 0.012618372216820717, + 0.008355595171451569, + 0.00864080898463726, + 0.013751559890806675, + 0.013559685088694096, + 0.011763074435293674, + 0.011531876400113106, + 0.012370865792036057, + 0.004406028892844915, + 0.0028409557417035103, + 0.0038991328328847885, + 0.002598067745566368, + 0.002589666284620762, + 0.004936535377055407, + 0.002152188215404749, + 0.004224351141601801, + 0.007795798592269421, + 0.005496189929544926, + 0.0313694104552269, + 0.0048109316267073154, + 0.0032692074310034513, + 0.023661943152546883, + 0.005041833966970444, + 0.006720713805407286, + 0.030609382316470146, + 0.03393441438674927, + 0.11814171820878983, + 0.016297373920679092, + 0.012709886766970158, + 0.13996630907058716, + 0.021584346890449524, + 0.02060961164534092, + 0.017232170328497887, + 0.021676119416952133, + 0.03579122573137283, + 0.023742428049445152, + 0.011929045431315899, + 0.02701430208981037, + 0.018295807763934135, + 0.019112052395939827, + 0.004205154720693827, + 0.0034630356822162867, + 0.003200302831828594, + 0.004011981654912233, + 0.0029883189126849174, + 0.002792937681078911, + 0.0033722848165780306, + 0.0030477200634777546, + 0.004997017327696085, + 0.004248014185577631, + 0.0031663402915000916, + 0.0024463851004838943, + 0.0025959096383303404, + 0.002643224783241749, + 0.0015638289041817188, + 0.0020345328375697136, + 0.012061943300068378, + 0.011847183108329773, + 0.011347560212016106, + 0.015636827796697617, + 0.007895298302173615, + 0.009398093447089195, + 0.0073385839350521564, + 0.006986413616687059, + 0.003953433595597744, + 0.003318261355161667, + 0.005504385568201542, + 0.0033002684358507395, + 0.005991755519062281, + 0.0043919021263718605, + 0.004193541593849659, + 0.005295758601278067, + 0.0038712520617991686, + 0.004808313678950071, + 0.0035227895714342594, + 0.007604587357491255, + 0.004724031314253807, + 0.013604330830276012, + 0.0042458148673176765, + 0.01132962666451931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.17990219593048096 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155785977840424 + ], + "opposite_logprob_tokens": [ + -1.7015578746795654 + ], + "logprob_reward": -0.20155785977840424, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155785977840424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1404048055410385, + 0.034326836466789246, + 0.03505318984389305, + 0.1837058961391449, + 0.15552479028701782, + 0.11331907659769058, + 0.07953160256147385, + 0.12159015238285065, + 0.06980859488248825, + 0.06673508137464523 + ], + "credit_map": [ + 0.022303683683276176, + 0.01640043407678604, + 0.013048685155808926, + 0.016145771369338036, + 0.020408783107995987, + 0.01944875530898571, + 0.016440464183688164, + 0.016208218410611153, + 0.00540479552000761, + 0.0033166727516800165, + 0.004725792445242405, + 0.0032022518571466208, + 0.005406097508966923, + 0.004051073919981718, + 0.003075131680816412, + 0.0051450179889798164, + 0.005585016682744026, + 0.004285440314561129, + 0.0045394618064165115, + 0.006738059222698212, + 0.003176445607095957, + 0.0030866090673953295, + 0.0035387533716857433, + 0.004103402607142925, + 0.02063712477684021, + 0.02260413207113743, + 0.036099229007959366, + 0.011597083881497383, + 0.012658828869462013, + 0.04783065244555473, + 0.015501085668802261, + 0.016777750104665756, + 0.01857103407382965, + 0.017308777198195457, + 0.03692450746893883, + 0.019233835861086845, + 0.011239035986363888, + 0.02436738647520542, + 0.015572242438793182, + 0.012307963334023952, + 0.015675444155931473, + 0.013448186218738556, + 0.010827064514160156, + 0.027949005365371704, + 0.008756051771342754, + 0.016604894772171974, + 0.01321767084300518, + 0.006840756628662348, + 0.0253923162817955, + 0.01819468103349209, + 0.008913865312933922, + 0.005907791666686535, + 0.004193384200334549, + 0.007369720842689276, + 0.004135666880756617, + 0.005424175411462784, + 0.015209220349788666, + 0.01915368065237999, + 0.016494451090693474, + 0.021518737077713013, + 0.012994722463190556, + 0.013763269409537315, + 0.009668982587754726, + 0.012787084095180035, + 0.006779471877962351, + 0.0055806622840464115, + 0.010049799457192421, + 0.0066515011712908745, + 0.01240305881947279, + 0.009308192878961563, + 0.008172810077667236, + 0.010863092727959156, + 0.00760893477126956, + 0.004759976174682379, + 0.00996036734431982, + 0.0063254074193537235, + 0.005687718279659748, + 0.013704604469239712, + 0.004129121080040932, + 0.01455894485116005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.20155785977840424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, taken at eye level, the camera moves slowly and delicately, focusing on a tech entrepreneur in his 30s with meticulously styled black hair, standing in a state-of-the-art design studio. He wears a sophisticated charcoal-gray merino wool turtleneck sweater, accentuating his confident and focused posture. The setting is minimalist and modern, with polished concrete floors and floor-to-ceiling windows in the background, revealing a vast cityscape bathed in the cool, deep blue light of dusk. In the space before him, a series of glowing, translucent holographic icons—shimmering in a vibrant cyan and magenta palette—float weightlessly in mid-air. With an expression of focused intensity, he choreographs this virtual interface with fluid, balletic gestures; his right hand swipes swiftly and decisively to expand complex data graphs, while his left meticulously rotates a 3D wireframe model of a futuristic vehicle with delicate, pincer-like movements of his thumb and index finger. The ethereal light emanating from the hologram casts soft, colored reflections on his face and hands, seamlessly blending the virtual elements with the room's tangible reality and creating a breathtaking atmosphere of pure innovation and effortless control.", + "seed": 1689, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_13/index_7481_seed_1689_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11442800611257553, + "vlm_logprob": -0.11442800611257553, + "vlm_yes_no_margin": 2.274999976158142, + "vlm_token_credit": -0.11442800611257553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07142895460128784 + ], + "opposite_logprob_tokens": [ + -3.3214290142059326 + ], + "logprob_reward": -0.07142895460128784, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.07142895460128784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06139608100056648, + 0.035472605377435684, + 0.048916034400463104, + 0.023700620979070663, + 0.0525863915681839, + 0.05518906190991402, + 0.038195110857486725, + 0.041209567338228226, + 0.5740229487419128, + 0.06931158900260925 + ], + "credit_map": [ + 0.009298666380345821, + 0.006104570347815752, + 0.0086308429017663, + 0.012125177308917046, + 0.007391465362161398, + 0.004988992586731911, + 0.006562718655914068, + 0.006293645594269037, + 0.004249018616974354, + 0.002633165568113327, + 0.0049684299156069756, + 0.006327736657112837, + 0.0046858396381139755, + 0.0030732126906514168, + 0.004391795489937067, + 0.005143410060554743, + 0.008044376038014889, + 0.003990148194134235, + 0.009498785249888897, + 0.005777833051979542, + 0.0032893435563892126, + 0.013104686513543129, + 0.002648309338837862, + 0.0025625533889979124, + 0.005023358389735222, + 0.0018876588437706232, + 0.0018926019547507167, + 0.0029750654939562082, + 0.0027529401704669, + 0.005576522089540958, + 0.001965487375855446, + 0.0016269859625026584, + 0.005523438099771738, + 0.0046965074725449085, + 0.009505542926490307, + 0.007383460644632578, + 0.0065519195050001144, + 0.010642576962709427, + 0.0048165591433644295, + 0.003466386580839753, + 0.0074102068319916725, + 0.0051453085616230965, + 0.0070921736769378185, + 0.011716878972947598, + 0.0068997531197965145, + 0.008609279058873653, + 0.003522848943248391, + 0.004792612511664629, + 0.004675119183957577, + 0.0024812063202261925, + 0.0023886444978415966, + 0.011964173056185246, + 0.0036144284531474113, + 0.00671496894210577, + 0.0027637439779937267, + 0.003592824097722769, + 0.007422699127346277, + 0.004617845639586449, + 0.004163648001849651, + 0.009598121978342533, + 0.0035659957211464643, + 0.005803574342280626, + 0.0025953869335353374, + 0.0034422960598021746, + 0.15284988284111023, + 0.02730860561132431, + 0.24666376411914825, + 0.03996836766600609, + 0.015518074855208397, + 0.035424310714006424, + 0.03672728314995766, + 0.01956266537308693, + 0.01318009290844202, + 0.01148692611604929, + 0.005346995312720537, + 0.0071232011541724205, + 0.004593417048454285, + 0.01981080137193203, + 0.004277676343917847, + 0.0034924859646707773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.07142895460128784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298547685146332 + ], + "opposite_logprob_tokens": [ + -2.237985372543335 + ], + "logprob_reward": -0.11298547685146332, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298547685146332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04804461821913719, + 0.015537228435277939, + 0.039465028792619705, + 0.025621609762310982, + 0.11641840636730194, + 0.010872380807995796, + 0.0766332671046257, + 0.012417107820510864, + 0.6316381096839905, + 0.02335219644010067 + ], + "credit_map": [ + 0.006920504383742809, + 0.004664471838623285, + 0.006134499330073595, + 0.007404884323477745, + 0.005508244037628174, + 0.005379323847591877, + 0.006296735256910324, + 0.0057359556667506695, + 0.002059144899249077, + 0.0012373444624245167, + 0.0022001927718520164, + 0.0026801847852766514, + 0.0019225252326577902, + 0.001932806451804936, + 0.002079194411635399, + 0.0014258349547162652, + 0.00987759418785572, + 0.0035611430648714304, + 0.0022255342919379473, + 0.0034269748721271753, + 0.0015151710249483585, + 0.016147425398230553, + 0.0012413149233907461, + 0.001469869865104556, + 0.0037618852220475674, + 0.0013853543205186725, + 0.0031075975857675076, + 0.003816810669377446, + 0.0018665400566533208, + 0.004754305351525545, + 0.004189459141343832, + 0.0027396585792303085, + 0.011985210701823235, + 0.00953043531626463, + 0.030433066189289093, + 0.006682999432086945, + 0.004756060894578695, + 0.04459407553076744, + 0.00374963297508657, + 0.004686923231929541, + 0.0013864845968782902, + 0.0008813041495159268, + 0.001798825222067535, + 0.0018665599636733532, + 0.0014533490175381303, + 0.0011573359370231628, + 0.0008475409704260528, + 0.001480981009081006, + 0.005418047774583101, + 0.0046645826660096645, + 0.004080840386450291, + 0.030193619430065155, + 0.00247773714363575, + 0.02204839698970318, + 0.004898447077721357, + 0.0028515965677797794, + 0.0017242785543203354, + 0.0014688529772683978, + 0.001269637723453343, + 0.0019449532264843583, + 0.0013944990932941437, + 0.002419687109068036, + 0.0011400782968848944, + 0.0010551202576607466, + 0.16859930753707886, + 0.0301617830991745, + 0.2706352472305298, + 0.042343344539403915, + 0.01770954579114914, + 0.03959694877266884, + 0.040114354342222214, + 0.022477589547634125, + 0.002623612293973565, + 0.00445632915943861, + 0.0031126851681619883, + 0.0034323798026889563, + 0.0025582925882190466, + 0.003077511675655842, + 0.002275069011375308, + 0.0018163177883252501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.11298547685146332 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003212109208107 + ], + "opposite_logprob_tokens": [ + -2.3503212928771973 + ], + "logprob_reward": -0.1003212109208107, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003212109208107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0211002379655838, + 0.026498813182115555, + 0.015127373859286308, + 0.011106052435934544, + 0.03180154040455818, + 0.020183781161904335, + 0.027599314227700233, + 0.04935451224446297, + 0.7593942880630493, + 0.037834081798791885 + ], + "credit_map": [ + 0.0028332092333585024, + 0.0022322421427816153, + 0.003599171293899417, + 0.004022468347102404, + 0.002391030080616474, + 0.0019220622489228845, + 0.0018192208372056484, + 0.002280835062265396, + 0.003380995476618409, + 0.00238447985611856, + 0.0036494245287030935, + 0.005701808258891106, + 0.0023049544543027878, + 0.002144074998795986, + 0.004282565321773291, + 0.0026505112182348967, + 0.002508259378373623, + 0.0012935252161696553, + 0.0017275161808356643, + 0.0020321374759078026, + 0.0010655390797182918, + 0.003929526545107365, + 0.0016933042788878083, + 0.0008775661117397249, + 0.0015875183744356036, + 0.000844258232973516, + 0.001349001657217741, + 0.0014042120892554522, + 0.0012824226869270205, + 0.0017727361992001534, + 0.0015499707078561187, + 0.0013159328373149037, + 0.003556240350008011, + 0.0024821870028972626, + 0.008142491802573204, + 0.0017876223428174853, + 0.001572516979649663, + 0.012145346030592918, + 0.0010376115096732974, + 0.0010775215923786163, + 0.00249181198887527, + 0.0021238531917333603, + 0.002665865235030651, + 0.004007815849035978, + 0.002300775144249201, + 0.0038121789693832397, + 0.0012645211536437273, + 0.0015169588150456548, + 0.001895259483717382, + 0.0016655870713293552, + 0.001803558086976409, + 0.010515447705984116, + 0.001777726924046874, + 0.006917646154761314, + 0.0018305195262655616, + 0.0011935693910345435, + 0.009624490514397621, + 0.0045075188390910625, + 0.005222594831138849, + 0.013305608183145523, + 0.003997454419732094, + 0.006261200178414583, + 0.003074364038184285, + 0.0033612786792218685, + 0.20213279128074646, + 0.03678235039114952, + 0.3255068361759186, + 0.054639093577861786, + 0.020380984991788864, + 0.04660831391811371, + 0.048214543610811234, + 0.02512936107814312, + 0.007353599648922682, + 0.007220953702926636, + 0.00242871162481606, + 0.0026988317258656025, + 0.00236982642672956, + 0.011156454682350159, + 0.0031880869064480066, + 0.0014176167314872146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.1003212109208107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270512044429779 + ], + "opposite_logprob_tokens": [ + -2.1270511150360107 + ], + "logprob_reward": -0.1270512044429779, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270512044429779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03137998282909393, + 0.05793483182787895, + 0.057363539934158325, + 0.0439319908618927, + 0.06180647760629654, + 0.04570542648434639, + 0.19319507479667664, + 0.021803580224514008, + 0.35434478521347046, + 0.13253425061702728 + ], + "credit_map": [ + 0.005335856229066849, + 0.002492598257958889, + 0.004845445044338703, + 0.005715741775929928, + 0.0035826358944177628, + 0.0025532131548970938, + 0.003385993419215083, + 0.0034685058053582907, + 0.008539129048585892, + 0.0035635074600577354, + 0.013753354549407959, + 0.010023958049714565, + 0.0052117835730314255, + 0.005491550080478191, + 0.006264835130423307, + 0.005086732096970081, + 0.013625680468976498, + 0.004967899527400732, + 0.003932992462068796, + 0.005125069059431553, + 0.0024053279776126146, + 0.022483063861727715, + 0.002559088869020343, + 0.0022644249256700277, + 0.008801981806755066, + 0.0026692766696214676, + 0.004684268496930599, + 0.006655900739133358, + 0.004074721597135067, + 0.008021828718483448, + 0.005102045368403196, + 0.003921976312994957, + 0.007088363170623779, + 0.00539754331111908, + 0.013532932847738266, + 0.0037920998875051737, + 0.002936308505013585, + 0.023451242595911026, + 0.002314411336556077, + 0.0032935882918536663, + 0.004762206692248583, + 0.004354191478341818, + 0.0053671314381062984, + 0.008905172348022461, + 0.007256129756569862, + 0.007960183545947075, + 0.0038907863199710846, + 0.0032096346840262413, + 0.013330496847629547, + 0.012018839828670025, + 0.009650029242038727, + 0.0667170062661171, + 0.011542809195816517, + 0.06079611927270889, + 0.010816258378326893, + 0.008323552086949348, + 0.0031045624054968357, + 0.002158042509108782, + 0.0018030720530077815, + 0.004590048920363188, + 0.0024911309592425823, + 0.003826407715678215, + 0.0022346414625644684, + 0.0015956781571730971, + 0.09266572445631027, + 0.01793033443391323, + 0.1513034701347351, + 0.026650352403521538, + 0.009765058755874634, + 0.021646620705723763, + 0.022574592381715775, + 0.011808671988546848, + 0.024420371279120445, + 0.011270290240645409, + 0.012341853231191635, + 0.017654573544859886, + 0.011554436758160591, + 0.029653960838913918, + 0.014591912738978863, + 0.011046882718801498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.1270512044429779 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a tech entrepreneur in his 30s with styled black hair wearing a charcoal-gray merino wool turtleneck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing translucent holographic icons in cyan and magenta floating in mid-air before the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the subject perform specific gestures, including swiping with the right hand and rotating a 3D wireframe vehicle with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and delicately throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603531837463379 + ], + "opposite_logprob_tokens": [ + -1.910353183746338 + ], + "logprob_reward": -0.1603531837463379, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603531837463379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03017193265259266, + 0.013508742675185204, + 0.020518645644187927, + 0.016288582235574722, + 0.12093071639537811, + 0.02208145707845688, + 0.04348238185048103, + 0.027422476559877396, + 0.6376885771751404, + 0.06790647655725479 + ], + "credit_map": [ + 0.00431616697460413, + 0.0026555650401860476, + 0.0031376348342746496, + 0.005927450023591518, + 0.003651795443147421, + 0.002098293509334326, + 0.0041466206312179565, + 0.00423840619623661, + 0.0023728751111775637, + 0.0013202200643718243, + 0.0019569399300962687, + 0.0012750222813338041, + 0.0022009003441780806, + 0.0018170239636674523, + 0.0013150873128324747, + 0.0012506741331890225, + 0.0037823799066245556, + 0.0014656624989584088, + 0.003394100349396467, + 0.0026551310438662767, + 0.0017730891704559326, + 0.004338860511779785, + 0.0018096673302352428, + 0.0012997537851333618, + 0.0017454477492719889, + 0.0015491483500227332, + 0.0032036949414759874, + 0.0027779170777648687, + 0.00182481505908072, + 0.0017468915320932865, + 0.0020202319137752056, + 0.001420435612089932, + 0.012746080756187439, + 0.011457426473498344, + 0.02666921354830265, + 0.008945446461439133, + 0.006679616868495941, + 0.043493080884218216, + 0.0053259581327438354, + 0.005613901652395725, + 0.0025657224468886852, + 0.0018210199195891619, + 0.002574429614469409, + 0.004206719342619181, + 0.00399193586781621, + 0.0023705775383859873, + 0.0026060438249260187, + 0.001945009338669479, + 0.0023949819151312113, + 0.002585404785349965, + 0.002631529001519084, + 0.015835553407669067, + 0.0029402582440525293, + 0.011172371916472912, + 0.002673092531040311, + 0.003249190514907241, + 0.0027703449595719576, + 0.003033803077414632, + 0.0022465360816568136, + 0.004902481101453304, + 0.002855842001736164, + 0.005255565047264099, + 0.003664635121822357, + 0.0026932694017887115, + 0.16664628684520721, + 0.0328361876308918, + 0.27274322509765625, + 0.046908993273973465, + 0.017135780304670334, + 0.039369553327560425, + 0.04098326712846756, + 0.021065328270196915, + 0.013899525627493858, + 0.0046118502505123615, + 0.005198569968342781, + 0.0078756557777524, + 0.005060971714556217, + 0.01832563802599907, + 0.007473676931113005, + 0.0054605882614851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.1603531837463379 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_14/index_tensor([1024, 1024])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_14/index_tensor([1024, 1024])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d36f8763e5be352889194218a4e8df0ba52cc15b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_14/index_tensor([1024, 1024])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, taken from a level angle, focuses on an elderly couple sitting side by side on a quaint wooden park bench. Soft afternoon sunlight slants through the golden foliage of the tall trees behind them, casting a dappled, warm light and shadow on them. The elderly man on the left, his face deeply wrinkled by time, his thinning silver hair gleaming softly in the sunlight. He wears a beige knitted cardigan over a light blue shirt. Beside him, his wife, also silver-haired, has a gentle and serene expression. She wears a lavender woolen shawl, adding a soft coolness to the warm scene. The elderly man's gaze drifts off into the distance, as if lost in reminiscence. His chest rises and falls imperceptibly, and then he lets out a slow, almost silent sigh, a blend of nostalgia, contentment, and a touch of melancholy. The main colors of the entire picture are warm gold and brown. Through extremely realistic rendering of details, from the texture of the clothes to the light and shadow of every wrinkle on the skin, they strive to be real, creating an atmosphere of tranquility, affection and full of a sense of time.", + "seed": 1162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_14/index_1024_seed_1162_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018501677364110947, + "vlm_logprob": -0.018501677364110947, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.018501677364110947, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008893521502614021 + ], + "opposite_logprob_tokens": [ + -5.2588934898376465 + ], + "logprob_reward": -0.008893521502614021, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008893521502614021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06382843106985092, + 0.03583264350891113, + 0.04947561398148537, + 0.019944321364164352, + 0.01582190953195095, + 0.021222826093435287, + 0.1281844973564148, + 0.03333055600523949, + 0.03370470926165581, + 0.598654568195343 + ], + "credit_map": [ + 0.01001906767487526, + 0.005813981406390667, + 0.009269959293305874, + 0.005357042420655489, + 0.005774073768407106, + 0.00889784935861826, + 0.008172925561666489, + 0.01052352786064148, + 0.006208557635545731, + 0.003149141324684024, + 0.006303617265075445, + 0.0027407240122556686, + 0.00415600324049592, + 0.005483022890985012, + 0.0034479801543056965, + 0.004343596752732992, + 0.00944506749510765, + 0.005893642548471689, + 0.004815447144210339, + 0.0033791507594287395, + 0.0038123319391161203, + 0.015069520100951195, + 0.0033098063431680202, + 0.0037506488151848316, + 0.002776690758764744, + 0.0035832934081554413, + 0.0017879025544971228, + 0.001855767099186778, + 0.0026183512527495623, + 0.002733890898525715, + 0.0026649530045688152, + 0.0019234726205468178, + 0.002773852553218603, + 0.001594880479387939, + 0.002060806844383478, + 0.0009489632211625576, + 0.002180543728172779, + 0.002530973870307207, + 0.001958308042958379, + 0.0017735807923600078, + 0.0033708082046359777, + 0.002523793838918209, + 0.00205914000980556, + 0.0011968696489930153, + 0.0020970297046005726, + 0.005362770985811949, + 0.0026786334346979856, + 0.001933779800310731, + 0.0233636312186718, + 0.007352705113589764, + 0.014272492378950119, + 0.005097738932818174, + 0.04866107180714607, + 0.01319922972470522, + 0.008418447338044643, + 0.007819177582859993, + 0.00910874642431736, + 0.0022675914224237204, + 0.004809336271136999, + 0.0025676670484244823, + 0.0023417251650243998, + 0.0027997568249702454, + 0.006237404886633158, + 0.003198326099663973, + 0.005268788896501064, + 0.004032705444842577, + 0.004107845481485128, + 0.002720980439335108, + 0.0031731845811009407, + 0.004913514479994774, + 0.004936259239912033, + 0.00455143116414547, + 0.0587160661816597, + 0.02272055856883526, + 0.0424899123609066, + 0.013586549088358879, + 0.2676278352737427, + 0.1158040389418602, + 0.03881276398897171, + 0.038896847516298294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.008893521502614021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014222919009625912 + ], + "opposite_logprob_tokens": [ + -4.264223098754883 + ], + "logprob_reward": -0.014222919009625912, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014222919009625912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0722518116235733, + 0.10845845937728882, + 0.3304888606071472, + 0.0604541189968586, + 0.032735999673604965, + 0.03928620368242264, + 0.04528017342090607, + 0.12932999432086945, + 0.03824131190776825, + 0.14347310364246368 + ], + "credit_map": [ + 0.011548587121069431, + 0.007730976212769747, + 0.012624632567167282, + 0.006332808639854193, + 0.007698072586208582, + 0.009092834778130054, + 0.007534090429544449, + 0.009689810685813427, + 0.012933540157973766, + 0.00805662665516138, + 0.0228972677141428, + 0.004663832485675812, + 0.015889951959252357, + 0.019131461158394814, + 0.013804133981466293, + 0.01108164619654417, + 0.07514727115631104, + 0.02800239622592926, + 0.029787417501211166, + 0.012654062360525131, + 0.02000025101006031, + 0.12435801327228546, + 0.017720969393849373, + 0.022818496450781822, + 0.007958214730024338, + 0.004436492919921875, + 0.005885724443942308, + 0.002772062551230192, + 0.008722074329853058, + 0.01647534966468811, + 0.006489145569503307, + 0.0077150557190179825, + 0.004369329661130905, + 0.003929605707526207, + 0.0028606110718101263, + 0.0029288933146744967, + 0.004180602263659239, + 0.004627796821296215, + 0.0035644415766000748, + 0.006274717394262552, + 0.005961098242551088, + 0.004326669964939356, + 0.006360323168337345, + 0.003705215873196721, + 0.004122899379581213, + 0.005775475408881903, + 0.004702113103121519, + 0.004332407843321562, + 0.008253928273916245, + 0.003764266148209572, + 0.004900873173028231, + 0.0021462934091687202, + 0.011586845852434635, + 0.005351063329726458, + 0.004147653467953205, + 0.005129249766469002, + 0.04652928188443184, + 0.007594821974635124, + 0.024283897131681442, + 0.0062951077707111835, + 0.005996132735162973, + 0.013836905360221863, + 0.015027559362351894, + 0.009766285307705402, + 0.004106046631932259, + 0.0035056660417467356, + 0.0034069810062646866, + 0.0025083578657358885, + 0.010652507655322552, + 0.004685967694967985, + 0.004869413562119007, + 0.004506370518356562, + 0.014747558161616325, + 0.005813907831907272, + 0.015834230929613113, + 0.0037235936615616083, + 0.06198732554912567, + 0.022167993709445, + 0.010822257958352566, + 0.008376243524253368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.014222919009625912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806073755025864 + ], + "opposite_logprob_tokens": [ + -3.529806137084961 + ], + "logprob_reward": -0.029806073755025864, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806073755025864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23803547024726868, + 0.17383602261543274, + 0.08373809605836868, + 0.02833724021911621, + 0.06649626791477203, + 0.023162011057138443, + 0.15601538121700287, + 0.12424732744693756, + 0.031290072947740555, + 0.07484214752912521 + ], + "credit_map": [ + 0.024416916072368622, + 0.015139703638851643, + 0.04349980130791664, + 0.013352920301258564, + 0.036434393376111984, + 0.045115359127521515, + 0.029250606894493103, + 0.030825762078166008, + 0.01777280680835247, + 0.00974203273653984, + 0.03797545284032822, + 0.008799721486866474, + 0.023262396454811096, + 0.031615350395441055, + 0.0282258540391922, + 0.01644241437315941, + 0.017029765993356705, + 0.006038145627826452, + 0.009323428384959698, + 0.004525252152234316, + 0.00921289250254631, + 0.020720699802041054, + 0.009684452787041664, + 0.0072034508921206, + 0.0034744436852633953, + 0.003062056377530098, + 0.0025491947308182716, + 0.0021879090927541256, + 0.004961113445460796, + 0.0038227690383791924, + 0.00460490956902504, + 0.0036748419515788555, + 0.013039147481322289, + 0.004434983711689711, + 0.007498119957745075, + 0.002901137573644519, + 0.006734449416399002, + 0.02042545937001705, + 0.00547029310837388, + 0.005992671009153128, + 0.004327879752963781, + 0.0020859981887042522, + 0.0028592224698513746, + 0.0018563885241746902, + 0.0019633439369499683, + 0.005672072060406208, + 0.0023406713735312223, + 0.0020564314909279346, + 0.02805640548467636, + 0.013714401982724667, + 0.01745162159204483, + 0.00693866889923811, + 0.04891842231154442, + 0.015361043624579906, + 0.011533315293490887, + 0.014041494578123093, + 0.0368281714618206, + 0.00717573007568717, + 0.021776149049401283, + 0.005556953139603138, + 0.007409354671835899, + 0.01265630405396223, + 0.015675613656640053, + 0.0171690434217453, + 0.003511920338496566, + 0.00326487491838634, + 0.0033654477447271347, + 0.0020475792698562145, + 0.009469124488532543, + 0.0037231636233627796, + 0.0032989932224154472, + 0.0026089674793183804, + 0.008864250965416431, + 0.0034084683284163475, + 0.013583033345639706, + 0.00286392355337739, + 0.02789846993982792, + 0.007729628123342991, + 0.00536370137706399, + 0.005130667705088854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.029806073755025864 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787782095372677 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787782095372677, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787782095372677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0724315419793129, + 0.07310891896486282, + 0.10297827422618866, + 0.05762782692909241, + 0.06007394567131996, + 0.07188201695680618, + 0.016798118129372597, + 0.021682988852262497, + 0.09815260767936707, + 0.42526373267173767 + ], + "credit_map": [ + 0.015479085966944695, + 0.007015813607722521, + 0.007777141407132149, + 0.004491610452532768, + 0.010185581631958485, + 0.012364153750240803, + 0.00569248478859663, + 0.009425663389265537, + 0.01853008009493351, + 0.007250302936881781, + 0.012237234972417355, + 0.004031750839203596, + 0.007846790365874767, + 0.00797225534915924, + 0.007504613604396582, + 0.007735888939350843, + 0.017064949497580528, + 0.007743142079561949, + 0.010578552260994911, + 0.003508491674438119, + 0.005541225429624319, + 0.03976709023118019, + 0.011936530470848083, + 0.006838289555162191, + 0.008229110389947891, + 0.010522260330617428, + 0.00751543277874589, + 0.005269430112093687, + 0.00503543158993125, + 0.010463597252964973, + 0.0058986470103263855, + 0.004693913273513317, + 0.008757207542657852, + 0.007462289184331894, + 0.006809061858803034, + 0.004720213357359171, + 0.005894257687032223, + 0.01363934762775898, + 0.006315589416772127, + 0.006475979462265968, + 0.014145685359835625, + 0.007926189340651035, + 0.009825061075389385, + 0.006748962681740522, + 0.007082622963935137, + 0.013112938031554222, + 0.005236852448433638, + 0.007803707383573055, + 0.0017676034476608038, + 0.0021593738347291946, + 0.0017720164032652974, + 0.0017961434787139297, + 0.001720250817015767, + 0.002532605780288577, + 0.0021740603260695934, + 0.0028760633431375027, + 0.0031523192301392555, + 0.002665013074874878, + 0.0020491709001362324, + 0.0014820960350334644, + 0.002291395328938961, + 0.0032329128589481115, + 0.0023979973047971725, + 0.00441208528354764, + 0.014016995206475258, + 0.00979500450193882, + 0.008799762465059757, + 0.00731463311240077, + 0.02411550283432007, + 0.012441578321158886, + 0.011605397798120975, + 0.010063735768198967, + 0.04048487916588783, + 0.015368924476206303, + 0.039049435406923294, + 0.009688597172498703, + 0.1897553950548172, + 0.07653328776359558, + 0.02704693004488945, + 0.02733626216650009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.009787782095372677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979809045791626 + ], + "opposite_logprob_tokens": [ + -3.5297980308532715 + ], + "logprob_reward": -0.02979809045791626, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979809045791626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04142913222312927, + 0.03248129412531853, + 0.045413680374622345, + 0.019117766991257668, + 0.032114461064338684, + 0.025967948138713837, + 0.018887607380747795, + 0.025523681193590164, + 0.022045856341719627, + 0.7370185852050781 + ], + "credit_map": [ + 0.010449363850057125, + 0.0050273193046450615, + 0.008077443577349186, + 0.003464445471763611, + 0.0022646107245236635, + 0.004171804990619421, + 0.0031160602811723948, + 0.004858080297708511, + 0.008665568195283413, + 0.002950998954474926, + 0.0054438430815935135, + 0.00219934806227684, + 0.002434129361063242, + 0.003989421762526035, + 0.0035228021442890167, + 0.0032751827966421843, + 0.008691010065376759, + 0.006630150135606527, + 0.0023893641773611307, + 0.003154197009280324, + 0.0049002389423549175, + 0.004980690311640501, + 0.010626748204231262, + 0.004041280597448349, + 0.0031432337127625942, + 0.0013537677004933357, + 0.0021669960115104914, + 0.0008035225910134614, + 0.002214452950283885, + 0.005621618591248989, + 0.0018995722057297826, + 0.0019146018894389272, + 0.005837708245962858, + 0.0017757655587047338, + 0.00431967806071043, + 0.0013732596999034286, + 0.003232494229450822, + 0.01000563520938158, + 0.002739789430052042, + 0.0028301291167736053, + 0.004814523737877607, + 0.002101251157000661, + 0.003608610015362501, + 0.0014587377663701773, + 0.0031228128354996443, + 0.006029919255524874, + 0.002376637188717723, + 0.0024554564151912928, + 0.0018669062992557883, + 0.002689612563699484, + 0.0012729630107060075, + 0.0012417128309607506, + 0.0037781621795147657, + 0.0028389585204422474, + 0.002070130780339241, + 0.003129160264506936, + 0.002691028406843543, + 0.0028350900392979383, + 0.004815542139112949, + 0.0016413384582847357, + 0.002033141441643238, + 0.0057763694785535336, + 0.002671882975846529, + 0.0030592887196689844, + 0.0027745342813432217, + 0.002172118052840233, + 0.002215711632743478, + 0.001851055771112442, + 0.0021392826456576586, + 0.00299592362716794, + 0.004049800336360931, + 0.003847430692985654, + 0.07416367530822754, + 0.02777949534356594, + 0.05683550238609314, + 0.016732115298509598, + 0.3297271132469177, + 0.13573047518730164, + 0.047796446830034256, + 0.04825374856591225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.02979809045791626 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, taken from a level angle, focuses on an elderly couple sitting side by side on a quaint wooden park bench. Soft afternoon sunlight slants through the golden foliage of the tall trees behind them, casting a dappled, warm light and shadow on them. The elderly man on the left, his face deeply wrinkled by time, his thinning silver hair gleaming softly in the sunlight. He wears a beige knitted cardigan over a light blue shirt. Beside him, his wife, also silver-haired, has a gentle and serene expression. She wears a lavender woolen shawl, adding a soft coolness to the warm scene. The elderly man's gaze drifts off into the distance, as if lost in reminiscence. His chest rises and falls imperceptibly, and then he lets out a slow, almost silent sigh, a blend of nostalgia, contentment, and a touch of melancholy. The main colors of the entire picture are warm gold and brown. Through extremely realistic rendering of details, from the texture of the clothes to the light and shadow of every wrinkle on the skin, they strive to be real, creating an atmosphere of tranquility, affection and full of a sense of time.", + "seed": 1163, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_14/index_1024_seed_1163_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014590925723314285, + "vlm_logprob": -0.014590925723314285, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.014590925723314285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009577162563800812 + ], + "opposite_logprob_tokens": [ + -5.00957727432251 + ], + "logprob_reward": -0.009577162563800812, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009577162563800812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3732526898384094, + 0.048941295593976974, + 0.12988296151161194, + 0.04150492697954178, + 0.09122366458177567, + 0.14068594574928284, + 0.03969036042690277, + 0.08550316840410233, + 0.019044730812311172, + 0.030270220711827278 + ], + "credit_map": [ + 0.03872883319854736, + 0.01790682226419449, + 0.1067555844783783, + 0.025874994695186615, + 0.024357527494430542, + 0.028510794043540955, + 0.1036919355392456, + 0.02742619812488556, + 0.00450701080262661, + 0.0048149083741009235, + 0.009290513582527637, + 0.004464574623852968, + 0.007604783866554499, + 0.007191213313490152, + 0.00747012859210372, + 0.003598165465518832, + 0.013276375830173492, + 0.004952958319336176, + 0.03640512749552727, + 0.006201916374266148, + 0.005298426374793053, + 0.04729872941970825, + 0.006989625282585621, + 0.009459802880883217, + 0.0050450656563043594, + 0.0035191169008612633, + 0.007048603612929583, + 0.0036191665567457676, + 0.005597877316176891, + 0.008504769764840603, + 0.003523795632645488, + 0.004646530374884605, + 0.011303749866783619, + 0.004318317398428917, + 0.03178246319293976, + 0.007737255189567804, + 0.004405813291668892, + 0.015096877701580524, + 0.005802236497402191, + 0.010776950977742672, + 0.017675546929240227, + 0.005442689172923565, + 0.04705866053700447, + 0.013223405927419662, + 0.004887229297310114, + 0.01597069762647152, + 0.007562977261841297, + 0.028864730149507523, + 0.004309024196118116, + 0.005062459036707878, + 0.004026906564831734, + 0.005044231656938791, + 0.005423086229711771, + 0.004698673728853464, + 0.005754841957241297, + 0.005371138919144869, + 0.007983613759279251, + 0.009195281192660332, + 0.011238188482820988, + 0.0111284414306283, + 0.0057910773903131485, + 0.009581997990608215, + 0.014710531570017338, + 0.01587403006851673, + 0.0028411196544766426, + 0.0016783251194283366, + 0.0011579260462895036, + 0.002031650161370635, + 0.0012710865121334791, + 0.001859853626228869, + 0.00258677639067173, + 0.005617992952466011, + 0.004548132419586182, + 0.003663347801193595, + 0.002123375190421939, + 0.002661660313606262, + 0.00277457176707685, + 0.003594901878386736, + 0.004369568545371294, + 0.006534661632031202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.009577162563800812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659431710839272 + ], + "opposite_logprob_tokens": [ + -4.758659362792969 + ], + "logprob_reward": -0.008659431710839272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008659431710839272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21825477480888367, + 0.05572576820850372, + 0.2509155571460724, + 0.06580553203821182, + 0.029869399964809418, + 0.0379449836909771, + 0.033012211322784424, + 0.08785760402679443, + 0.022909680381417274, + 0.19770443439483643 + ], + "credit_map": [ + 0.02098367176949978, + 0.007692620158195496, + 0.06361250579357147, + 0.014595153741538525, + 0.014105729758739471, + 0.018752960488200188, + 0.06219779700040817, + 0.01631433144211769, + 0.005415053106844425, + 0.007419014349579811, + 0.008582616224884987, + 0.005789880640804768, + 0.006106241140514612, + 0.004327497910708189, + 0.010958625935018063, + 0.007126835640519857, + 0.026442483067512512, + 0.008759395219385624, + 0.08052227646112442, + 0.009520738385617733, + 0.009925550781190395, + 0.08791949599981308, + 0.014254411682486534, + 0.013571217656135559, + 0.012367007322609425, + 0.0033506490290164948, + 0.010540190152823925, + 0.005417488515377045, + 0.006480883341282606, + 0.015332814306020737, + 0.006081628613173962, + 0.006234864704310894, + 0.005106073804199696, + 0.0020519394893199205, + 0.005042982287704945, + 0.0026969739701598883, + 0.0026141249109059572, + 0.005422675516456366, + 0.002576600294560194, + 0.004358027596026659, + 0.004631004761904478, + 0.002413678215816617, + 0.012124421074986458, + 0.003707251511514187, + 0.0018644348019734025, + 0.0043394966050982475, + 0.002862431574612856, + 0.006002265959978104, + 0.004642817657440901, + 0.0029919676017016172, + 0.006477771792560816, + 0.003337049623951316, + 0.004723407328128815, + 0.0035430705174803734, + 0.0034217769280076027, + 0.0038743505720049143, + 0.014718127436935902, + 0.004444794729351997, + 0.02874070592224598, + 0.007512000855058432, + 0.003181810723617673, + 0.008066229522228241, + 0.005521875340491533, + 0.01567206159234047, + 0.0030823415145277977, + 0.002869734540581703, + 0.0020701319444924593, + 0.0034412750974297523, + 0.0018720927182585, + 0.004050144460052252, + 0.002507347846403718, + 0.003016610397025943, + 0.04010600224137306, + 0.012210390530526638, + 0.014397778548300266, + 0.017916390672326088, + 0.006782136391848326, + 0.01691913977265358, + 0.022871796041727066, + 0.06650080531835556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.008659431710839272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187304958701134 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187304958701134, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187304958701134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2518412470817566, + 0.021094122901558876, + 0.36409857869148254, + 0.1061115562915802, + 0.015700366348028183, + 0.06481517851352692, + 0.07138319313526154, + 0.031196558848023415, + 0.01999453268945217, + 0.05376468598842621 + ], + "credit_map": [ + 0.026154395192861557, + 0.010442622005939484, + 0.07490108162164688, + 0.01724623516201973, + 0.015048871748149395, + 0.01697690039873123, + 0.07332220673561096, + 0.017748935148119926, + 0.0019037300953641534, + 0.0018649682169780135, + 0.0043289996683597565, + 0.0015478725545108318, + 0.0021570920944213867, + 0.002570987679064274, + 0.004722416400909424, + 0.001998055260628462, + 0.03792678937315941, + 0.011303179897367954, + 0.11204648017883301, + 0.013166019693017006, + 0.013855422846972942, + 0.13261272013187408, + 0.01975533552467823, + 0.023432612419128418, + 0.023060837760567665, + 0.005312390625476837, + 0.019333839416503906, + 0.007526039611548185, + 0.009056002832949162, + 0.026735641062259674, + 0.006638652645051479, + 0.008448147214949131, + 0.0017894526245072484, + 0.0016397422878071666, + 0.002443737117573619, + 0.0019583203829824924, + 0.0021152456756681204, + 0.0019196014618501067, + 0.0019088754197582603, + 0.0019253891659900546, + 0.00824766419827938, + 0.002313733333721757, + 0.0221566092222929, + 0.006120769772678614, + 0.0020344804506748915, + 0.007444810122251511, + 0.0032620837446302176, + 0.013235021382570267, + 0.009126059710979462, + 0.007540127262473106, + 0.012157955206930637, + 0.008724176324903965, + 0.0078815883025527, + 0.0076538873836398125, + 0.007778584491461515, + 0.010520807467401028, + 0.004764122422784567, + 0.0010699440026655793, + 0.009834018535912037, + 0.0027084313333034515, + 0.000908674206584692, + 0.002944965148344636, + 0.0024152356199920177, + 0.006551166530698538, + 0.0025546718388795853, + 0.0018640316557139158, + 0.0015380671247839928, + 0.002184160053730011, + 0.0022523822262883186, + 0.0042639318853616714, + 0.002785589313134551, + 0.002551698824390769, + 0.012538062408566475, + 0.0029250348452478647, + 0.004201042931526899, + 0.0039991396479308605, + 0.0018112780526280403, + 0.00396699970588088, + 0.006332815159112215, + 0.01799030974507332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.018187304958701134 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737846881151199 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737846881151199, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737846881151199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3029947280883789, + 0.14646583795547485, + 0.06782426685094833, + 0.25841909646987915, + 0.043863505125045776, + 0.039708420634269714, + 0.04495428130030632, + 0.017617937177419662, + 0.016356058418750763, + 0.06179579347372055 + ], + "credit_map": [ + 0.025050664320588112, + 0.019091930240392685, + 0.04982991889119148, + 0.02252338081598282, + 0.02376766875386238, + 0.01581420935690403, + 0.10393355786800385, + 0.04298342764377594, + 0.017201172187924385, + 0.010265625081956387, + 0.03906287997961044, + 0.012494342401623726, + 0.01749846152961254, + 0.020893745124340057, + 0.012076099403202534, + 0.016973529011011124, + 0.007072598673403263, + 0.0029009883292019367, + 0.022455332800745964, + 0.0028161522932350636, + 0.0040207174606621265, + 0.021597756072878838, + 0.004284861031919718, + 0.0026758662424981594, + 0.061343032866716385, + 0.014302393421530724, + 0.03921231999993324, + 0.018990347161889076, + 0.02111741155385971, + 0.06723732501268387, + 0.015260203741490841, + 0.02095608226954937, + 0.008627946488559246, + 0.0032497390639036894, + 0.009163253009319305, + 0.004168876446783543, + 0.002937848446890712, + 0.008334220387041569, + 0.0034849003423005342, + 0.0038967267610132694, + 0.010354594327509403, + 0.0030124925542622805, + 0.006000122055411339, + 0.004202605225145817, + 0.0031065719667822123, + 0.005188142415136099, + 0.0038556596264243126, + 0.0039882357232272625, + 0.011379304341971874, + 0.004508558660745621, + 0.007443557493388653, + 0.005434184800833464, + 0.003260268596932292, + 0.004835403524339199, + 0.004457575734704733, + 0.0036354325711727142, + 0.0018837423995137215, + 0.001761849271133542, + 0.0016337697161361575, + 0.002247254131361842, + 0.0014199124416336417, + 0.002399783581495285, + 0.00330058834515512, + 0.0029710382223129272, + 0.0019367404747754335, + 0.0014163247542455792, + 0.0013715799432247877, + 0.0021914581302553415, + 0.0013258948456496, + 0.0030285378452390432, + 0.0021970518864691257, + 0.0028884708881378174, + 0.012838694266974926, + 0.0037659411318600178, + 0.004538391251116991, + 0.004862598143517971, + 0.0030012712813913822, + 0.005333928856998682, + 0.007790353614836931, + 0.019664619117975235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.006737846881151199 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979288250207901 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.02979288250207901, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979288250207901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12810026109218597, + 0.11765474826097488, + 0.4315454065799713, + 0.050956349819898605, + 0.05001579225063324, + 0.036914654076099396, + 0.0391395278275013, + 0.06819825619459152, + 0.02759956568479538, + 0.04987546429038048 + ], + "credit_map": [ + 0.020226791501045227, + 0.006792701780796051, + 0.03395972028374672, + 0.011751637794077396, + 0.010896261781454086, + 0.010287989862263203, + 0.023975038900971413, + 0.010210121981799603, + 0.013668237254023552, + 0.005549202207475901, + 0.037945546209812164, + 0.009187767282128334, + 0.009286750108003616, + 0.02217797189950943, + 0.011787031777203083, + 0.008052249439060688, + 0.04568192735314369, + 0.0131006371229887, + 0.14673593640327454, + 0.015005534514784813, + 0.013602002523839474, + 0.15377458930015564, + 0.024076059460639954, + 0.01956874132156372, + 0.010610771365463734, + 0.004003321751952171, + 0.008413862437009811, + 0.004134533926844597, + 0.0057159229181706905, + 0.010704648680984974, + 0.0038476085755974054, + 0.003525681793689728, + 0.00778603833168745, + 0.00367496395483613, + 0.016107890754938126, + 0.003880675183609128, + 0.002744876779615879, + 0.006949811242520809, + 0.0034966410603374243, + 0.005374896805733442, + 0.005203716456890106, + 0.003270136658102274, + 0.008691602386534214, + 0.0031182661186903715, + 0.0035422970540821552, + 0.0043588546104729176, + 0.004897333215922117, + 0.003832448972389102, + 0.006920922081917524, + 0.002738849725574255, + 0.010166884399950504, + 0.005854228511452675, + 0.0022339923307299614, + 0.0036563328467309475, + 0.0033503363374620676, + 0.004217980429530144, + 0.010549303144216537, + 0.0026340107433497906, + 0.021634669974446297, + 0.006588437594473362, + 0.0023888368159532547, + 0.005577709060162306, + 0.00501999631524086, + 0.013805296272039413, + 0.00290762260556221, + 0.0031601032242178917, + 0.0023498383816331625, + 0.0035127538722008467, + 0.0031764928717166185, + 0.004050679039210081, + 0.004677356220781803, + 0.0037647199351340532, + 0.007403719238936901, + 0.005609074141830206, + 0.005909773986786604, + 0.004315140191465616, + 0.0045325299724936485, + 0.006276289001107216, + 0.00950794667005539, + 0.00632099062204361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02979288250207901 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, taken from a level angle, focuses on an elderly couple sitting side by side on a quaint wooden park bench. Soft afternoon sunlight slants through the golden foliage of the tall trees behind them, casting a dappled, warm light and shadow on them. The elderly man on the left, his face deeply wrinkled by time, his thinning silver hair gleaming softly in the sunlight. He wears a beige knitted cardigan over a light blue shirt. Beside him, his wife, also silver-haired, has a gentle and serene expression. She wears a lavender woolen shawl, adding a soft coolness to the warm scene. The elderly man's gaze drifts off into the distance, as if lost in reminiscence. His chest rises and falls imperceptibly, and then he lets out a slow, almost silent sigh, a blend of nostalgia, contentment, and a touch of melancholy. The main colors of the entire picture are warm gold and brown. Through extremely realistic rendering of details, from the texture of the clothes to the light and shadow of every wrinkle on the skin, they strive to be real, creating an atmosphere of tranquility, affection and full of a sense of time.", + "seed": 1164, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_14/index_1024_seed_1164_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018136934284120797, + "vlm_logprob": -0.018136934284120797, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.018136934284120797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008687440305948257 + ], + "opposite_logprob_tokens": [ + -5.008687496185303 + ], + "logprob_reward": -0.008687440305948257, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008687440305948257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15366800129413605, + 0.07699260115623474, + 0.15848283469676971, + 0.07380335032939911, + 0.08322068303823471, + 0.10330596566200256, + 0.1301959753036499, + 0.18087206780910492, + 0.01938096433877945, + 0.02007749117910862 + ], + "credit_map": [ + 0.020381318405270576, + 0.02139514498412609, + 0.015604383312165737, + 0.016053931787610054, + 0.02376866713166237, + 0.015789728611707687, + 0.020352430641651154, + 0.0203224066644907, + 0.015170327387750149, + 0.008493022061884403, + 0.010551823303103447, + 0.006474403198808432, + 0.011455925181508064, + 0.009797330014407635, + 0.007235271390527487, + 0.007814504206180573, + 0.029575884342193604, + 0.020078901201486588, + 0.015789946541190147, + 0.011038221418857574, + 0.011122841387987137, + 0.03273950144648552, + 0.009300725534558296, + 0.02883681282401085, + 0.010298826731741428, + 0.008599641732871532, + 0.009337127208709717, + 0.003898025956004858, + 0.013032973743975163, + 0.009793058969080448, + 0.008114132098853588, + 0.010729564353823662, + 0.020722132176160812, + 0.007659061811864376, + 0.01575358957052231, + 0.005547770764678717, + 0.006327585782855749, + 0.009754921309649944, + 0.007621660828590393, + 0.009833959862589836, + 0.02409818023443222, + 0.011967923492193222, + 0.013113260269165039, + 0.011645158752799034, + 0.014786767773330212, + 0.0110021298751235, + 0.008400429971516132, + 0.008292116224765778, + 0.024117235094308853, + 0.013765722513198853, + 0.015186838805675507, + 0.01151047833263874, + 0.012152009643614292, + 0.022686006501317024, + 0.020657146349549294, + 0.010120552964508533, + 0.02682877704501152, + 0.021354638040065765, + 0.020764237269759178, + 0.02037794515490532, + 0.014921959489583969, + 0.020918892696499825, + 0.0331827737390995, + 0.022522836923599243, + 0.002588244853541255, + 0.0026533096097409725, + 0.00126165512483567, + 0.0013027245877310634, + 0.004621921107172966, + 0.002349371323361993, + 0.0021212485153228045, + 0.0024824896827340126, + 0.0027058329433202744, + 0.002588056493550539, + 0.002319911727681756, + 0.002034913981333375, + 0.002976394025608897, + 0.0025760699063539505, + 0.001640298985876143, + 0.003236014163121581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.008687440305948257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822852462530136 + ], + "opposite_logprob_tokens": [ + -3.529822826385498 + ], + "logprob_reward": -0.029822852462530136, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822852462530136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13617493212223053, + 0.08319875597953796, + 0.23570133745670319, + 0.07268474251031876, + 0.102518729865551, + 0.13724201917648315, + 0.04432159289717674, + 0.12838783860206604, + 0.026408817619085312, + 0.03336120769381523 + ], + "credit_map": [ + 0.01571178063750267, + 0.018566902726888657, + 0.012706764973700047, + 0.016402535140514374, + 0.023517204448580742, + 0.013077212497591972, + 0.0162067711353302, + 0.019985772669315338, + 0.013916712254285812, + 0.008770192041993141, + 0.010063241235911846, + 0.010904806666076183, + 0.010386419482529163, + 0.01053342130035162, + 0.010581022128462791, + 0.008042935281991959, + 0.03450008109211922, + 0.032795678824186325, + 0.03247032314538956, + 0.034719765186309814, + 0.023747680708765984, + 0.03548971936106682, + 0.017107881605625153, + 0.024870209395885468, + 0.012819923460483551, + 0.009598116390407085, + 0.009902589954435825, + 0.005672377999871969, + 0.008995629847049713, + 0.008315850049257278, + 0.009883840568363667, + 0.007496418431401253, + 0.018894586712121964, + 0.010864587500691414, + 0.01409042626619339, + 0.007485808338969946, + 0.015650242567062378, + 0.011116699315607548, + 0.011374177411198616, + 0.013042192906141281, + 0.04038127884268761, + 0.018315061926841736, + 0.01586885377764702, + 0.011331980116665363, + 0.013004442676901817, + 0.017623310908675194, + 0.008186093531548977, + 0.012531009502708912, + 0.008263028226792812, + 0.005103417672216892, + 0.0033130208030343056, + 0.0034192418679594994, + 0.005041308235377073, + 0.006767806597054005, + 0.007630143780261278, + 0.0047836266458034515, + 0.01876806654036045, + 0.01656075194478035, + 0.011881213635206223, + 0.015045342966914177, + 0.01677039824426174, + 0.014686623588204384, + 0.021143866702914238, + 0.013531571254134178, + 0.0034284391440451145, + 0.0030093861278146505, + 0.0023021791130304337, + 0.001651856815442443, + 0.004204454831779003, + 0.004927709698677063, + 0.0034834493417292833, + 0.003401343245059252, + 0.004935246892273426, + 0.004270895849913359, + 0.004436175804585218, + 0.003107290016487241, + 0.005553885828703642, + 0.0037946857046335936, + 0.0030244796071201563, + 0.004238548688590527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.029822852462530136 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581606775522232 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581606775522232, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581606775522232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14082247018814087, + 0.0598115399479866, + 0.12460775673389435, + 0.10108470171689987, + 0.0575571246445179, + 0.23259826004505157, + 0.16047337651252747, + 0.07608316093683243, + 0.029114900156855583, + 0.01784672960639 + ], + "credit_map": [ + 0.021618662402033806, + 0.01626613736152649, + 0.013194111175835133, + 0.00966904778033495, + 0.02037379890680313, + 0.017570581287145615, + 0.01657227613031864, + 0.025557857006788254, + 0.0057734097354114056, + 0.006469659507274628, + 0.006029506679624319, + 0.007803574204444885, + 0.011539770290255547, + 0.006386198103427887, + 0.008928711526095867, + 0.006880708504468203, + 0.03665956109762192, + 0.011832859367132187, + 0.01960650645196438, + 0.010016455315053463, + 0.00925799086689949, + 0.016853775829076767, + 0.007452401798218489, + 0.012928204610943794, + 0.025607725605368614, + 0.010892665944993496, + 0.016364267095923424, + 0.0058034854009747505, + 0.008349275216460228, + 0.012938364408910275, + 0.00973425805568695, + 0.011394662782549858, + 0.013640769757330418, + 0.007924643345177174, + 0.008262825198471546, + 0.004354218952357769, + 0.006946238689124584, + 0.006315151695162058, + 0.004662119783461094, + 0.0054511562921106815, + 0.047590017318725586, + 0.03453761711716652, + 0.023934844881296158, + 0.03797873109579086, + 0.02057052217423916, + 0.02608324959874153, + 0.02080932818353176, + 0.0210939459502697, + 0.024630986154079437, + 0.014622475951910019, + 0.013342882506549358, + 0.01489910576492548, + 0.009225753135979176, + 0.03628953546285629, + 0.0350092388689518, + 0.012453398667275906, + 0.011735769920051098, + 0.009873590432107449, + 0.009378432296216488, + 0.007771746721118689, + 0.010660766623914242, + 0.009156770072877407, + 0.008031024597585201, + 0.00947506632655859, + 0.005825981497764587, + 0.0034048075322061777, + 0.002505106385797262, + 0.0035678204149007797, + 0.0031453280244022608, + 0.004056804813444614, + 0.003695590188726783, + 0.0029134617652744055, + 0.0018338585505262017, + 0.0023654175456613302, + 0.0021146994549781084, + 0.001981547335162759, + 0.0021399473771452904, + 0.0028530138079077005, + 0.0020245276391506195, + 0.0025337161496281624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.020581606775522232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252846051007509 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252846051007509, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252846051007509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3760550320148468, + 0.1010521799325943, + 0.07718072831630707, + 0.12324405461549759, + 0.03249063715338707, + 0.0844104066491127, + 0.07007257640361786, + 0.06180381029844284, + 0.03339280188083649, + 0.0402977429330349 + ], + "credit_map": [ + 0.041285622864961624, + 0.053539786487817764, + 0.04126525670289993, + 0.05785933509469032, + 0.03160301223397255, + 0.04656355455517769, + 0.036941807717084885, + 0.06699664145708084, + 0.016456469893455505, + 0.014175351709127426, + 0.011251330375671387, + 0.011226736940443516, + 0.012797192670404911, + 0.011490462347865105, + 0.011029317043721676, + 0.012625323608517647, + 0.011808636598289013, + 0.012406822293996811, + 0.010836485773324966, + 0.007753316778689623, + 0.008275505155324936, + 0.011194213293492794, + 0.00626746891066432, + 0.008638276718556881, + 0.03482562303543091, + 0.011440370231866837, + 0.024074336513876915, + 0.008081314153969288, + 0.007695122621953487, + 0.016324929893016815, + 0.007660564035177231, + 0.013141798786818981, + 0.0061781639233231544, + 0.00558784743770957, + 0.0029241417068988085, + 0.003485659137368202, + 0.004644435364753008, + 0.0033384780399501324, + 0.003281373530626297, + 0.0030505384784191847, + 0.015002627857029438, + 0.00886200275272131, + 0.008755368180572987, + 0.0068360744044184685, + 0.011008678935468197, + 0.008952178061008453, + 0.01219768077135086, + 0.012795794755220413, + 0.0077079711481928825, + 0.006143646780401468, + 0.007326124235987663, + 0.0037239284720271826, + 0.007566965185105801, + 0.015625229105353355, + 0.014594554901123047, + 0.007384148892015219, + 0.008551131933927536, + 0.006970892660319805, + 0.005727987736463547, + 0.004598251543939114, + 0.007653177250176668, + 0.008056027814745903, + 0.013608542270958424, + 0.006637799087911844, + 0.0049094101414084435, + 0.004375360440462828, + 0.003980480600148439, + 0.0039556571282446384, + 0.005181517917662859, + 0.003646798897534609, + 0.0038187073078006506, + 0.003524868283420801, + 0.005839350167661905, + 0.004571893718093634, + 0.0044061122462153435, + 0.004334423691034317, + 0.006249862723052502, + 0.004385165870189667, + 0.00476469611749053, + 0.005746238864958286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005252846051007509 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339925825595856 + ], + "opposite_logprob_tokens": [ + -3.6513400077819824 + ], + "logprob_reward": -0.026339925825595856, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339925825595856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25295189023017883, + 0.05285639688372612, + 0.0642002671957016, + 0.07624345272779465, + 0.03520280122756958, + 0.07237695157527924, + 0.33393263816833496, + 0.04723542928695679, + 0.03816414996981621, + 0.026836110278964043 + ], + "credit_map": [ + 0.050510335713624954, + 0.03312769904732704, + 0.02838124893605709, + 0.02507978491485119, + 0.024561740458011627, + 0.033560480922460556, + 0.02283422276377678, + 0.0348963588476181, + 0.012302631512284279, + 0.007058329880237579, + 0.005950169172137976, + 0.0031026345677673817, + 0.005961226765066385, + 0.006814939901232719, + 0.006087197922170162, + 0.005579262040555477, + 0.011946062557399273, + 0.007630604784935713, + 0.008222372271120548, + 0.00428613368421793, + 0.006612492725253105, + 0.012311853468418121, + 0.004262711387127638, + 0.008928032591938972, + 0.01644984260201454, + 0.006954528391361237, + 0.012468640692532063, + 0.00395455677062273, + 0.00814344547688961, + 0.009125017561018467, + 0.006853240076452494, + 0.012294178828597069, + 0.006934183184057474, + 0.004305683076381683, + 0.0046857800334692, + 0.002875374862924218, + 0.003516162047162652, + 0.0045202043838799, + 0.0035097519867122173, + 0.004855656530708075, + 0.011428981088101864, + 0.007667584344744682, + 0.008462093770503998, + 0.0049870251677930355, + 0.013612516224384308, + 0.007163768634200096, + 0.010368280112743378, + 0.00868669617921114, + 0.03232639282941818, + 0.016275646165013313, + 0.024059098213911057, + 0.011431935243308544, + 0.023557040840387344, + 0.100179523229599, + 0.09280673414468765, + 0.03329623118042946, + 0.0066683548502624035, + 0.005398324690759182, + 0.003938976675271988, + 0.0030995577108114958, + 0.010436884127557278, + 0.007434590253978968, + 0.00559445284307003, + 0.004664285574108362, + 0.0045875743962824345, + 0.0058375694788992405, + 0.004337170161306858, + 0.0022648314479738474, + 0.007877661846578121, + 0.005268188659101725, + 0.003492399351671338, + 0.004498751368373632, + 0.005367573816329241, + 0.0031419123988598585, + 0.004043584689497948, + 0.0029125125147402287, + 0.002804016461595893, + 0.0036192936822772026, + 0.0019807214848697186, + 0.0029664929024875164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.026339925825595856 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, taken from a level angle, focuses on an elderly couple sitting side by side on a quaint wooden park bench. Soft afternoon sunlight slants through the golden foliage of the tall trees behind them, casting a dappled, warm light and shadow on them. The elderly man on the left, his face deeply wrinkled by time, his thinning silver hair gleaming softly in the sunlight. He wears a beige knitted cardigan over a light blue shirt. Beside him, his wife, also silver-haired, has a gentle and serene expression. She wears a lavender woolen shawl, adding a soft coolness to the warm scene. The elderly man's gaze drifts off into the distance, as if lost in reminiscence. His chest rises and falls imperceptibly, and then he lets out a slow, almost silent sigh, a blend of nostalgia, contentment, and a touch of melancholy. The main colors of the entire picture are warm gold and brown. Through extremely realistic rendering of details, from the texture of the clothes to the light and shadow of every wrinkle on the skin, they strive to be real, creating an atmosphere of tranquility, affection and full of a sense of time.", + "seed": 1165, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_14/index_1024_seed_1165_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012550043687224389, + "vlm_logprob": -0.012550043687224389, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.012550043687224389, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010466211475431919 + ], + "opposite_logprob_tokens": [ + -4.8854660987854 + ], + "logprob_reward": -0.010466211475431919, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010466211475431919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08875977247953415, + 0.02478862926363945, + 0.0790611207485199, + 0.24026387929916382, + 0.1648603230714798, + 0.0653763860464096, + 0.08162953704595566, + 0.05008183792233467, + 0.17423515021800995, + 0.03094335086643696 + ], + "credit_map": [ + 0.011499758809804916, + 0.015191001817584038, + 0.00998468417674303, + 0.010987415909767151, + 0.010787401348352432, + 0.01151986699551344, + 0.0110376738011837, + 0.007751969620585442, + 0.0038106467109173536, + 0.0031161706428974867, + 0.003096264321357012, + 0.003608888480812311, + 0.0028591169975697994, + 0.0022217980585992336, + 0.00320239900611341, + 0.0028733452782034874, + 0.010631163604557514, + 0.007177416700869799, + 0.006383195985108614, + 0.00856896210461855, + 0.008873457089066505, + 0.023337431252002716, + 0.007120500784367323, + 0.006968991365283728, + 0.01761193573474884, + 0.01490533072501421, + 0.0818728432059288, + 0.013075586408376694, + 0.011041928082704544, + 0.06157269328832626, + 0.018463579937815666, + 0.021719984710216522, + 0.014667642302811146, + 0.031237322837114334, + 0.01882306858897209, + 0.029914861544966698, + 0.01928221434354782, + 0.016424445435404778, + 0.018815329298377037, + 0.01569543592631817, + 0.008923869580030441, + 0.009743556380271912, + 0.013257590122520924, + 0.007032549940049648, + 0.005337253678590059, + 0.00595444580540061, + 0.00730131333693862, + 0.007825808599591255, + 0.010160090401768684, + 0.009823932312428951, + 0.005400866735726595, + 0.008522219955921173, + 0.009644420817494392, + 0.015014315955340862, + 0.01131200511008501, + 0.011751687154173851, + 0.007657037116587162, + 0.015067812986671925, + 0.004081089515239, + 0.0035435159225016832, + 0.007246837485581636, + 0.003849596716463566, + 0.0048616486601531506, + 0.0037742978893220425, + 0.02088899351656437, + 0.07348111271858215, + 0.007488294970244169, + 0.010743488557636738, + 0.027428263798356056, + 0.013499604538083076, + 0.01346432976424694, + 0.007241052575409412, + 0.004892065655440092, + 0.005067884456366301, + 0.0029088668525218964, + 0.003629102138802409, + 0.0041864896193146706, + 0.0035799674224108458, + 0.003908488899469376, + 0.0027704855892807245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.010466211475431919 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006756199989467859 + ], + "opposite_logprob_tokens": [ + -5.00675630569458 + ], + "logprob_reward": -0.006756199989467859, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006756199989467859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06497781723737717, + 0.04608943313360214, + 0.03540833666920662, + 0.6696934700012207, + 0.02877783589065075, + 0.07368278503417969, + 0.038069434463977814, + 0.0143222501501441, + 0.012891164049506187, + 0.016087451949715614 + ], + "credit_map": [ + 0.008808128535747528, + 0.007973753847181797, + 0.006260178051888943, + 0.008364188484847546, + 0.009561071172356606, + 0.006545437034219503, + 0.006803479511290789, + 0.010661584325134754, + 0.004873586818575859, + 0.004766377154737711, + 0.0044211638160049915, + 0.007560838013887405, + 0.00638424139469862, + 0.004670785740017891, + 0.004968080669641495, + 0.008444361388683319, + 0.003453535493463278, + 0.0035064611583948135, + 0.0029947550501674414, + 0.005003673490136862, + 0.004744570702314377, + 0.007155402563512325, + 0.004190292675048113, + 0.004359645303338766, + 0.06758811324834824, + 0.033184658735990524, + 0.28143224120140076, + 0.03058049827814102, + 0.03607486933469772, + 0.1476363241672516, + 0.042817629873752594, + 0.030379120260477066, + 0.004013900179415941, + 0.004210387822240591, + 0.002883272711187601, + 0.004408450797200203, + 0.0028735941741615534, + 0.0031440365128219128, + 0.003499007085338235, + 0.0037451868411153555, + 0.007591179106384516, + 0.009027663618326187, + 0.020103219896554947, + 0.010015168227255344, + 0.007043411955237389, + 0.00667225569486618, + 0.007845831103622913, + 0.005384056828916073, + 0.004996802192181349, + 0.00435788556933403, + 0.00847889669239521, + 0.004547922872006893, + 0.004110578913241625, + 0.0054313503205776215, + 0.0032379652839154005, + 0.002908032387495041, + 0.0026282677426934242, + 0.002625902183353901, + 0.0016141433734446764, + 0.0009639373165555298, + 0.0019430371467024088, + 0.001498786499723792, + 0.0015999386087059975, + 0.0014482372207567096, + 0.0014574421802535653, + 0.0035494703333824873, + 0.001055798027664423, + 0.00109420798253268, + 0.0020231318194419146, + 0.0013099368661642075, + 0.0013364753685891628, + 0.0010647013550624251, + 0.0022859780583530664, + 0.0035056928172707558, + 0.0014855656772851944, + 0.0014840570511296391, + 0.0025389937218278646, + 0.0017302213236689568, + 0.001877169357612729, + 0.001179773360490799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.006756199989467859 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09498516470193863, + 0.03578216955065727, + 0.21887531876564026, + 0.2715236246585846, + 0.08769368380308151, + 0.08414026349782944, + 0.02595100924372673, + 0.06052429974079132, + 0.02509358711540699, + 0.09543092548847198 + ], + "credit_map": [ + 0.010789601132273674, + 0.013624189421534538, + 0.010022995062172413, + 0.01014900952577591, + 0.013972116634249687, + 0.012022816576063633, + 0.014889494515955448, + 0.009514939971268177, + 0.004098972771316767, + 0.00522202393040061, + 0.0029648670461028814, + 0.004803687799721956, + 0.005046301521360874, + 0.003422040492296219, + 0.005449877120554447, + 0.004774399567395449, + 0.025530066341161728, + 0.024636613205075264, + 0.023258909583091736, + 0.02804275043308735, + 0.025627940893173218, + 0.041354116052389145, + 0.023370003327727318, + 0.02705492451786995, + 0.026883142068982124, + 0.015379597432911396, + 0.10739975422620773, + 0.01385441329330206, + 0.014926389791071415, + 0.06361255794763565, + 0.017803192138671875, + 0.011664560064673424, + 0.01560326386243105, + 0.01258755661547184, + 0.008091445080935955, + 0.008489848114550114, + 0.014843218959867954, + 0.009644794277846813, + 0.010690937750041485, + 0.007742622867226601, + 0.008780205622315407, + 0.009011921472847462, + 0.014818483032286167, + 0.009714066982269287, + 0.01062551699578762, + 0.011374370194971561, + 0.008357636630535126, + 0.011458062566816807, + 0.003345299046486616, + 0.0031035947613418102, + 0.004778713453561068, + 0.00277152331545949, + 0.0029449178837239742, + 0.0037886525969952345, + 0.002475208602845669, + 0.0027430991176515818, + 0.011972962878644466, + 0.01691523566842079, + 0.005160646978765726, + 0.004188021644949913, + 0.0074815908446908, + 0.005852150730788708, + 0.005371192004531622, + 0.0035824982915073633, + 0.002385827712714672, + 0.003801025915890932, + 0.0026224451139569283, + 0.0029715236742049456, + 0.0034526558592915535, + 0.002060120226815343, + 0.004156977403908968, + 0.0036430105101317167, + 0.02185823768377304, + 0.011662941426038742, + 0.00830866489559412, + 0.010711648501455784, + 0.011812200769782066, + 0.008967255242168903, + 0.01372175570577383, + 0.008388221263885498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.009783769026398659 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059498813934624195 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.0059498813934624195, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059498813934624195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07699908316135406, + 0.055642616003751755, + 0.11774498224258423, + 0.3261978328227997, + 0.03232648968696594, + 0.050494033843278885, + 0.026599755510687828, + 0.1334065943956375, + 0.15606573224067688, + 0.024522814899683 + ], + "credit_map": [ + 0.014052655547857285, + 0.01010927464812994, + 0.007136109750717878, + 0.006883916910737753, + 0.011677231639623642, + 0.010320494882762432, + 0.010233205743134022, + 0.006586200557649136, + 0.011887955479323864, + 0.006486392579972744, + 0.005600004456937313, + 0.0068042767234146595, + 0.007954230532050133, + 0.005376533139497042, + 0.005788024049252272, + 0.0057452041655778885, + 0.022784776985645294, + 0.009903401136398315, + 0.007982388138771057, + 0.009641611017286777, + 0.007898288778960705, + 0.04355074465274811, + 0.00909348577260971, + 0.006890296004712582, + 0.03379412367939949, + 0.01796519011259079, + 0.13795006275177002, + 0.015343943610787392, + 0.020093608647584915, + 0.056194812059402466, + 0.022039463743567467, + 0.022816648706793785, + 0.005967156030237675, + 0.0045572854578495026, + 0.0034585907123982906, + 0.003540012752637267, + 0.003182657528668642, + 0.004191828891634941, + 0.004626200534403324, + 0.002802762668579817, + 0.005938648711889982, + 0.009071245789527893, + 0.009524489752948284, + 0.006168358027935028, + 0.004842379130423069, + 0.005047385580837727, + 0.005230961367487907, + 0.004670566413551569, + 0.0054184808395802975, + 0.006284222472459078, + 0.002315809251740575, + 0.0027978969737887383, + 0.0021913533564656973, + 0.001699124346487224, + 0.0026605059392750263, + 0.0032323659397661686, + 0.02210545353591442, + 0.03016039915382862, + 0.017732294276356697, + 0.010942845605313778, + 0.01811944507062435, + 0.010938060469925404, + 0.013542082160711288, + 0.009866026230156422, + 0.018922125920653343, + 0.06577570736408234, + 0.008706395514309406, + 0.009828506968915462, + 0.024105612188577652, + 0.01178679708391428, + 0.01109659019857645, + 0.005844002589583397, + 0.005036018323153257, + 0.0027854242362082005, + 0.0024853318464010954, + 0.004718417301774025, + 0.0021231714636087418, + 0.0028120428323745728, + 0.0027095957193523645, + 0.001852815505117178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.0059498813934624195 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people sitting side by side on a wooden park bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man on the left wearing a beige knitted cardigan over a light blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a lavender woolen shawl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft afternoon sunlight creating dappled shadows through golden foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794156551361084 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794156551361084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794156551361084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07795793563127518, + 0.03218628466129303, + 0.05782482400536537, + 0.4282781481742859, + 0.1049165278673172, + 0.049548909068107605, + 0.04029276221990585, + 0.06872080266475677, + 0.09039158374071121, + 0.049882229417562485 + ], + "credit_map": [ + 0.012393246404826641, + 0.012140590697526932, + 0.009793560951948166, + 0.010440940037369728, + 0.008887515403330326, + 0.009123661555349827, + 0.009592399001121521, + 0.005586032755672932, + 0.005325707606971264, + 0.004518389236181974, + 0.0028848329093307257, + 0.004646398592740297, + 0.005720813758671284, + 0.003053941298276186, + 0.0035663614980876446, + 0.0024698409251868725, + 0.010137215256690979, + 0.005797761958092451, + 0.004157884977757931, + 0.004582964815199375, + 0.005110837984830141, + 0.018142422661185265, + 0.005633698310703039, + 0.004262044094502926, + 0.039203811436891556, + 0.02128341794013977, + 0.18002910912036896, + 0.019733978435397148, + 0.024266444146633148, + 0.09278856962919235, + 0.02742629311978817, + 0.023546552285552025, + 0.011520533822476864, + 0.01970578171312809, + 0.01276057492941618, + 0.016618352383375168, + 0.013209600001573563, + 0.011045807041227818, + 0.0113819669932127, + 0.008673912845551968, + 0.006456317845731974, + 0.0064109754748642445, + 0.006214968394488096, + 0.005209855269640684, + 0.007376461289823055, + 0.0060316999442875385, + 0.0055361236445605755, + 0.006312510464340448, + 0.008679281920194626, + 0.006162544246762991, + 0.0032495204359292984, + 0.00341252563521266, + 0.004411883186548948, + 0.0036755141336470842, + 0.004423974547535181, + 0.0062775216065347195, + 0.011543423868715763, + 0.019574876874685287, + 0.006765265483409166, + 0.005727419164031744, + 0.008452823385596275, + 0.005504032131284475, + 0.006519257090985775, + 0.004633711650967598, + 0.011013500392436981, + 0.039089545607566833, + 0.0034267683513462543, + 0.005809676833450794, + 0.014475653879344463, + 0.006119025871157646, + 0.006503937300294638, + 0.0039534796960651875, + 0.009992820210754871, + 0.0055242860689759254, + 0.006843599025160074, + 0.008326173759996891, + 0.005500491708517075, + 0.004164735320955515, + 0.0063084024004638195, + 0.003221724880859256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.029794156551361084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_15/index_tensor([3812, 3812])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_15/index_tensor([3812, 3812])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f9931033d9dbcf66de4f7fa6a6f9f7f6b7137518 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_15/index_tensor([3812, 3812])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up transports the viewer into the exquisitely detailed world of a vibrant goldenrod flower, its forms rendered with the painterly precision of classic Makoto Shinkai's aesthetic. The fiery golden petals fill the frame, their texture appearing paper-thin yet delicately soft, each delicate vein visible as a subtle and intricate network branching across the surface. A fine mist gently falls upon the flower, its moisture slowly gathering into rows of tiny crystal droplets that cling to the petals like miniature jewels. At the flower's core, a dark, velvety center creates a profound, captivating point of contrast. Backlit by the warm, low-angle sun, each droplet captures the light, refracting it into a breathtaking spectacle of shimmering prismatic sparkle. This gives the flower's edges an ethereal, luminous halo, evoking a powerful, poignant atmosphere of tranquility and fleeting beauty.", + "seed": 19118, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_15/index_3812_seed_19118_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0947957654017955, + "vlm_logprob": -0.0947957654017955, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.0947957654017955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031994156539440155 + ], + "opposite_logprob_tokens": [ + -3.531994104385376 + ], + "logprob_reward": -0.031994156539440155, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.031994156539440155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019779261201620102, + 0.9497738480567932, + 0.0017787193646654487, + 0.009194444864988327, + 0.002142788376659155, + 0.0046066949144005775, + 0.0020281809847801924, + 0.0034574270248413086, + 0.001146296039223671, + 0.006092377007007599 + ], + "credit_map": [ + 0.0019147556740790606, + 0.0020568205509334803, + 0.0015030791983008385, + 0.003254897892475128, + 0.0017060472164303064, + 0.004882101435214281, + 0.0013099601492285728, + 0.0031515969894826412, + 0.04916343092918396, + 0.05441894382238388, + 0.30470559000968933, + 0.3902128338813782, + 0.019870635122060776, + 0.08574437350034714, + 0.02577385865151882, + 0.019884169101715088, + 0.00025152857415378094, + 0.00020228070206940174, + 0.00031531116110272706, + 0.00016081541252788156, + 0.00017798312183003873, + 0.00022149425058159977, + 0.00023334569414146245, + 0.00021596014266833663, + 0.0005730583216063678, + 0.0034632037859410048, + 0.0029807821847498417, + 0.00041306932689622045, + 0.00030085264006629586, + 0.0008053387864492834, + 0.00030106527265161276, + 0.0003570730332285166, + 0.0003443466848693788, + 0.00022796672419644892, + 0.0004272663500159979, + 0.00019602151587605476, + 0.00015769731544423848, + 0.0002712514251470566, + 0.0002682055637706071, + 0.00025003260816447437, + 0.0008338732877746224, + 0.0005815993063151836, + 0.000994497095234692, + 0.0005394925829023123, + 0.0003050131199415773, + 0.0005874361377209425, + 0.0004102002130821347, + 0.00035458282218314707, + 0.00028846890199929476, + 0.00020102219423279166, + 0.00033662543864920735, + 0.0003821533755399287, + 0.00015877935220487416, + 0.00017176188703160733, + 0.00028071829001419246, + 0.00020865123951807618, + 0.0002902636188082397, + 0.0008364461245946586, + 0.0009598128963261843, + 0.00024738998035900295, + 0.00023597684048581868, + 0.00036240846384316683, + 0.00021917818230576813, + 0.00030595067073591053, + 0.00015081647143233567, + 0.00020375633903313428, + 0.000198300025658682, + 0.00012886714830528945, + 7.663102587684989e-05, + 0.00019640369282569736, + 8.98524303920567e-05, + 0.0001016687456285581, + 0.0006089446251280606, + 0.0006368821486830711, + 0.0006602713256143034, + 0.0005061750998720527, + 0.0018003347795456648, + 0.0008135964744724333, + 0.0006972792325541377, + 0.00036889268085360527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.031994156539440155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026383349671959877 + ], + "opposite_logprob_tokens": [ + -3.651383399963379 + ], + "logprob_reward": -0.026383349671959877, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026383349671959877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029543746262788773, + 0.8874131441116333, + 0.00916415173560381, + 0.027971934527158737, + 0.007966693490743637, + 0.0073022679425776005, + 0.004137091338634491, + 0.005604906007647514, + 0.006111332681030035, + 0.014784730970859528 + ], + "credit_map": [ + 0.0030491608195006847, + 0.0027181028854101896, + 0.002332794014364481, + 0.005329661536961794, + 0.0034463030751794577, + 0.006306723691523075, + 0.002235712483525276, + 0.004125289618968964, + 0.04593206197023392, + 0.05090693384408951, + 0.2840643525123596, + 0.3647119998931885, + 0.018739791586995125, + 0.08013704419136047, + 0.024005835875868797, + 0.01891515403985977, + 0.0016588367288932204, + 0.0010608587181195617, + 0.0013277262914925814, + 0.0008836131892167032, + 0.0009495226549915969, + 0.0015083025209605694, + 0.0010964326793327928, + 0.0006788591272197664, + 0.0011780090862885118, + 0.010606377385556698, + 0.009151573292911053, + 0.001278448267839849, + 0.0010188870364800096, + 0.002638287376612425, + 0.0010274220257997513, + 0.0010729320347309113, + 0.0012702057138085365, + 0.000815920066088438, + 0.001624235650524497, + 0.0007301500882022083, + 0.0006302121328189969, + 0.0008843161631375551, + 0.0010113544994965196, + 0.0010002998169511557, + 0.0009730732999742031, + 0.0008147957269102335, + 0.001223976374603808, + 0.0008557093678973615, + 0.0006739823729731143, + 0.0009734213235788047, + 0.0009866224136203527, + 0.0008006872376427054, + 0.0005652091349475086, + 0.0007346876082010567, + 0.0006559288594871759, + 0.00040574243757873774, + 0.00026067905128002167, + 0.0004220736736897379, + 0.0006026238552294672, + 0.0004901466309092939, + 0.0005164911854080856, + 0.0007642809650860727, + 0.0008371094008907676, + 0.0006288663716986775, + 0.0009439626010134816, + 0.0005797885823994875, + 0.0006943393382243812, + 0.000640068028587848, + 0.0006501260213553905, + 0.0010574954794719815, + 0.000941287144087255, + 0.0006315284408628941, + 0.0006914669065736234, + 0.0009282043320126832, + 0.0006840492133051157, + 0.0005271753179840744, + 0.0013911352725699544, + 0.00158921186812222, + 0.0014560320414602757, + 0.0012059370055794716, + 0.004437322728335857, + 0.001998909981921315, + 0.0017584782326593995, + 0.0009477051207795739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.026383349671959877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870631158351898 + ], + "opposite_logprob_tokens": [ + -1.1370631456375122 + ], + "logprob_reward": -0.3870631158351898, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870631158351898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04466287046670914, + 0.5304294228553772, + 0.005736308638006449, + 0.054811250418424606, + 0.005329825449734926, + 0.01288034301251173, + 0.010184317827224731, + 0.015354391187429428, + 0.012954269535839558, + 0.3076569437980652 + ], + "credit_map": [ + 0.004125545267015696, + 0.003940857481211424, + 0.0023022464010864496, + 0.007815960794687271, + 0.009727871976792812, + 0.00850538071244955, + 0.003974243067204952, + 0.004270763602107763, + 0.027796784415841103, + 0.03057740069925785, + 0.16931597888469696, + 0.2176646739244461, + 0.011710378341376781, + 0.04764512553811073, + 0.014530016109347343, + 0.011189092881977558, + 0.0008137626573443413, + 0.0009317327057942748, + 0.0008810671279206872, + 0.0005550466594286263, + 0.0005195732228457928, + 0.0007483723456971347, + 0.0006594465812668204, + 0.0006273075705394149, + 0.003890170017257333, + 0.019424017518758774, + 0.016742141917347908, + 0.0027257706969976425, + 0.0020289842505007982, + 0.004611464217305183, + 0.002724225400015712, + 0.0026644717436283827, + 0.0007599358796142042, + 0.0004728866333607584, + 0.0009823134168982506, + 0.0005150386714376509, + 0.0005876870127394795, + 0.0007515400648117065, + 0.0007068223785609007, + 0.0005536009557545185, + 0.002202847972512245, + 0.0014544036239385605, + 0.0026102790143340826, + 0.0015729042934253812, + 0.0009493225370533764, + 0.0018680635839700699, + 0.0012609335826709867, + 0.000961588229984045, + 0.0014038928784430027, + 0.0014606259064748883, + 0.0015440941788256168, + 0.0011772396974265575, + 0.0007941315416246653, + 0.0012280023656785488, + 0.0017038786318153143, + 0.0008724516956135631, + 0.0013210242614150047, + 0.0018746902933344245, + 0.0020889940205961466, + 0.0018494116375222802, + 0.0026091549079865217, + 0.002645639004185796, + 0.0018426324240863323, + 0.0011228451039642096, + 0.0014931332552805543, + 0.0015276159392669797, + 0.001294335350394249, + 0.0011369165731593966, + 0.003208989044651389, + 0.0017813704907894135, + 0.0014935904182493687, + 0.001018318347632885, + 0.0291950311511755, + 0.033550817519426346, + 0.027719151228666306, + 0.022976243868470192, + 0.09538323432207108, + 0.04250532388687134, + 0.03585483133792877, + 0.020472312346100807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.3870631158351898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021972341928631067 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021972341928631067, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021972341928631067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01684885285794735, + 0.887865424156189, + 0.006327847018837929, + 0.017476093024015427, + 0.004095505923032761, + 0.007684051990509033, + 0.007226179353892803, + 0.009459690190851688, + 0.008795008063316345, + 0.034221284091472626 + ], + "credit_map": [ + 0.0019482047064229846, + 0.0018072675447911024, + 0.0015892317751422524, + 0.0024871109053492546, + 0.0018331629689782858, + 0.0030149761587381363, + 0.0014320930931717157, + 0.002736809430643916, + 0.046133510768413544, + 0.05080807954072952, + 0.284568727016449, + 0.36444011330604553, + 0.018783897161483765, + 0.0803145095705986, + 0.024089790880680084, + 0.018726903945207596, + 0.0010866103693842888, + 0.0006359619437716901, + 0.0006561862537637353, + 0.0006724580889567733, + 0.0005311060813255608, + 0.0008314930601045489, + 0.0009926565689966083, + 0.0009213755838572979, + 0.0015051322989165783, + 0.005336579866707325, + 0.004951522219926119, + 0.000983346370048821, + 0.0007716267136856914, + 0.0017222752794623375, + 0.0010923335794359446, + 0.0011132796062156558, + 0.00045109272468835115, + 0.0006006004987284541, + 0.0004880887863691896, + 0.0003863086167257279, + 0.0004778902803082019, + 0.00041837283060885966, + 0.0005614022957161069, + 0.0007117503555491567, + 0.0013169568264856935, + 0.0008042528061196208, + 0.0012891150545328856, + 0.0009688141872175038, + 0.0005692910053767264, + 0.00104498618748039, + 0.0009176710736937821, + 0.0007729658391326666, + 0.000926417123991996, + 0.001133945188485086, + 0.0009691247250884771, + 0.0008031507604755461, + 0.0009336852235719562, + 0.0009848868940025568, + 0.0009066225029528141, + 0.0005683478084392846, + 0.0006783213466405869, + 0.0023574605584144592, + 0.0022709679324179888, + 0.0005664229393005371, + 0.0010666423477232456, + 0.000958475167863071, + 0.0009745827410370111, + 0.0005868179723620415, + 0.0008850473677739501, + 0.0014929393073543906, + 0.0012043818132951856, + 0.0009851129725575447, + 0.0010731901274994016, + 0.0014133565127849579, + 0.0009772792691364884, + 0.0007637015078216791, + 0.0033657446037977934, + 0.0037674475461244583, + 0.0031090709380805492, + 0.002629027236253023, + 0.010589871555566788, + 0.004810965154320002, + 0.0039054211229085922, + 0.002043739892542362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.0021972341928631067 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340970769524574 + ], + "opposite_logprob_tokens": [ + -3.651340961456299 + ], + "logprob_reward": -0.026340970769524574, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340970769524574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0998576283454895, + 0.5506234169006348, + 0.034003328531980515, + 0.04472353309392929, + 0.013878916390240192, + 0.03398573398590088, + 0.02146230824291706, + 0.05559350550174713, + 0.027022920548915863, + 0.11884859204292297 + ], + "credit_map": [ + 0.009989899583160877, + 0.008896686136722565, + 0.006106535438448191, + 0.017917262390255928, + 0.01855999231338501, + 0.019568372517824173, + 0.008604243397712708, + 0.010214656591415405, + 0.029286464676260948, + 0.031660858541727066, + 0.17372579872608185, + 0.22538556158542633, + 0.013119576498866081, + 0.04996519908308983, + 0.01540346723049879, + 0.012076576240360737, + 0.0038037067279219627, + 0.0030725765973329544, + 0.0024720104411244392, + 0.003938229288905859, + 0.005842413287609816, + 0.004987598862498999, + 0.006537023466080427, + 0.003349777776747942, + 0.008636567741632462, + 0.006872858386486769, + 0.004041758831590414, + 0.00436014961451292, + 0.003548900131136179, + 0.0066491710022091866, + 0.005290549248456955, + 0.005323581397533417, + 0.0022519470658153296, + 0.0015835361555218697, + 0.0025761874858289957, + 0.001590151572600007, + 0.0013810427626594901, + 0.0017573628574609756, + 0.0015014596283435822, + 0.0012372315395623446, + 0.004814262501895428, + 0.006193506997078657, + 0.006486272905021906, + 0.003541612531989813, + 0.0025429697707295418, + 0.0038190698251128197, + 0.003690361510962248, + 0.002897681901231408, + 0.0030581478495150805, + 0.00224197912029922, + 0.0032224091701209545, + 0.001956620952114463, + 0.0021516859997063875, + 0.0029964123386889696, + 0.003703558584675193, + 0.002131496090441942, + 0.004567100200802088, + 0.012630036100745201, + 0.014347967691719532, + 0.005396333523094654, + 0.0038612361531704664, + 0.007529550697654486, + 0.004223742987960577, + 0.003037550952285528, + 0.003139705862849951, + 0.0039907521568238735, + 0.003826132509857416, + 0.002603972563520074, + 0.004355739336460829, + 0.004902263637632132, + 0.002514537191018462, + 0.0016898210160434246, + 0.01513940654695034, + 0.01269762683659792, + 0.01332652848213911, + 0.011169841513037682, + 0.02697938308119774, + 0.01423309650272131, + 0.015046810731291771, + 0.010255908593535423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.026340970769524574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up transports the viewer into the exquisitely detailed world of a vibrant goldenrod flower, its forms rendered with the painterly precision of classic Makoto Shinkai's aesthetic. The fiery golden petals fill the frame, their texture appearing paper-thin yet delicately soft, each delicate vein visible as a subtle and intricate network branching across the surface. A fine mist gently falls upon the flower, its moisture slowly gathering into rows of tiny crystal droplets that cling to the petals like miniature jewels. At the flower's core, a dark, velvety center creates a profound, captivating point of contrast. Backlit by the warm, low-angle sun, each droplet captures the light, refracting it into a breathtaking spectacle of shimmering prismatic sparkle. This gives the flower's edges an ethereal, luminous halo, evoking a powerful, poignant atmosphere of tranquility and fleeting beauty.", + "seed": 19119, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_15/index_3812_seed_19119_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19925820007920264, + "vlm_logprob": -0.19925820007920264, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.19925820007920264, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008093536831438541 + ], + "opposite_logprob_tokens": [ + -5.258093357086182 + ], + "logprob_reward": -0.008093536831438541, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008093536831438541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2647443413734436, + 0.14454518258571625, + 0.03732737526297569, + 0.043101269751787186, + 0.05069617182016373, + 0.13607196509838104, + 0.06317412853240967, + 0.13492271304130554, + 0.060841016471385956, + 0.06457588076591492 + ], + "credit_map": [ + 0.030757004395127296, + 0.03560066223144531, + 0.03977728262543678, + 0.026704145595431328, + 0.03598528727889061, + 0.03678571805357933, + 0.026861708611249924, + 0.03227251023054123, + 0.008389411494135857, + 0.029733188450336456, + 0.00966554880142212, + 0.010646837763488293, + 0.028419848531484604, + 0.029485026374459267, + 0.010526036843657494, + 0.017679277807474136, + 0.002888731425628066, + 0.004207680933177471, + 0.0053816065192222595, + 0.004315991885960102, + 0.005888835061341524, + 0.007531941868364811, + 0.0034941639751195908, + 0.003618421498686075, + 0.0034194381441920996, + 0.0048155710101127625, + 0.004498224705457687, + 0.00299818511120975, + 0.006529105361551046, + 0.008428236469626427, + 0.0037730266340076923, + 0.008639480918645859, + 0.004560620523989201, + 0.008199242874979973, + 0.0057809315621852875, + 0.003973348066210747, + 0.005735987331718206, + 0.009888894855976105, + 0.0054586236365139484, + 0.007098518777638674, + 0.01114202756434679, + 0.017671767622232437, + 0.013613425195217133, + 0.014573588036000729, + 0.016727378591895103, + 0.02573678083717823, + 0.017541054636240005, + 0.01906592957675457, + 0.0035584785509854555, + 0.014705885201692581, + 0.005781952291727066, + 0.0038826391100883484, + 0.013940546661615372, + 0.00871842261403799, + 0.004116642288863659, + 0.008469556458294392, + 0.0107673779129982, + 0.016751307994127274, + 0.017822762951254845, + 0.014616553671658039, + 0.020713331177830696, + 0.012739305384457111, + 0.020775683224201202, + 0.02073638141155243, + 0.007903850637376308, + 0.010321758687496185, + 0.00965141598135233, + 0.004461287055164576, + 0.00783609040081501, + 0.00924188457429409, + 0.004332988988608122, + 0.007091735489666462, + 0.005450033117085695, + 0.01117576565593481, + 0.0056893606670200825, + 0.004848617594689131, + 0.011224313639104366, + 0.013409660197794437, + 0.004757499322295189, + 0.008020629175007343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.008093536831438541 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11303818225860596 + ], + "opposite_logprob_tokens": [ + -2.2380380630493164 + ], + "logprob_reward": -0.11303818225860596, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11303818225860596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16597110033035278, + 0.05880098417401314, + 0.03422382101416588, + 0.025980709120631218, + 0.0439646914601326, + 0.24686402082443237, + 0.04750480130314827, + 0.2480517327785492, + 0.07803469896316528, + 0.050603460520505905 + ], + "credit_map": [ + 0.018368437886238098, + 0.025262508541345596, + 0.021339690312743187, + 0.015938997268676758, + 0.02137158066034317, + 0.024578223004937172, + 0.01603398099541664, + 0.023077677935361862, + 0.005463010165840387, + 0.007802198175340891, + 0.004747605416923761, + 0.0044116852805018425, + 0.00904777366667986, + 0.016291070729494095, + 0.004202286712825298, + 0.00683535635471344, + 0.0026937839575111866, + 0.005071860272437334, + 0.004471886903047562, + 0.0033235156442970037, + 0.005757940001785755, + 0.006709229666739702, + 0.0031921211630105972, + 0.0030034836381673813, + 0.002144220285117626, + 0.003109395271167159, + 0.0033747595734894276, + 0.0020067684818059206, + 0.003997840452939272, + 0.002863859059289098, + 0.002381970640271902, + 0.006101893726736307, + 0.0038281502202153206, + 0.007152336649596691, + 0.004881743341684341, + 0.0032612381037324667, + 0.006193791050463915, + 0.008420108817517757, + 0.003838665084913373, + 0.006388655863702297, + 0.019611259922385216, + 0.032179247587919235, + 0.024098578840494156, + 0.026636924594640732, + 0.03067496046423912, + 0.04722694307565689, + 0.03132708743214607, + 0.0351090244948864, + 0.0055845617316663265, + 0.0072316997684538364, + 0.007289744447916746, + 0.003691076999530196, + 0.006736404728144407, + 0.00670010969042778, + 0.004524060059338808, + 0.0057471445761621, + 0.019942132756114006, + 0.03281239792704582, + 0.03069557249546051, + 0.027813151478767395, + 0.03901303559541702, + 0.02278505265712738, + 0.03813771903514862, + 0.03685265779495239, + 0.011890311725437641, + 0.011598438024520874, + 0.01249708142131567, + 0.00682496465742588, + 0.009337755851447582, + 0.009887676686048508, + 0.006977479439228773, + 0.009020990692079067, + 0.004969700239598751, + 0.006685569416731596, + 0.006464448757469654, + 0.0061086323112249374, + 0.0070107802748680115, + 0.006764144171029329, + 0.005243816878646612, + 0.0073563698679208755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.11303818225860596 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261532783508301 + ], + "opposite_logprob_tokens": [ + -0.5761532783508301 + ], + "logprob_reward": -0.8261532783508301, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261532783508301, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2639598548412323, + 0.06439261883497238, + 0.027440186589956284, + 0.0378233939409256, + 0.21771813929080963, + 0.1070946455001831, + 0.0661642998456955, + 0.06037935987114906, + 0.10285666584968567, + 0.052170805633068085 + ], + "credit_map": [ + 0.030872268602252007, + 0.0385315902531147, + 0.0395086444914341, + 0.02749117650091648, + 0.03674154728651047, + 0.03194013610482216, + 0.02661341242492199, + 0.032261110842227936, + 0.006534981075674295, + 0.012870782054960728, + 0.007324211299419403, + 0.005986848380416632, + 0.014515193179249763, + 0.0046303109265863895, + 0.005531701259315014, + 0.006998595781624317, + 0.0020186228211969137, + 0.004102466627955437, + 0.0033067658077925444, + 0.0023154285736382008, + 0.005783153232187033, + 0.004659786354750395, + 0.002467752667143941, + 0.0027862144634127617, + 0.002786525059491396, + 0.00696305837482214, + 0.0043740784749388695, + 0.0038153231143951416, + 0.006727479863911867, + 0.006457625422626734, + 0.0036813528276979923, + 0.0030179531313478947, + 0.01593697816133499, + 0.04681510478258133, + 0.024298440665006638, + 0.014617454260587692, + 0.03765787556767464, + 0.03282402455806732, + 0.01605486497282982, + 0.02951342612504959, + 0.011114131659269333, + 0.015683362260460854, + 0.009977014735341072, + 0.008933968842029572, + 0.015410315245389938, + 0.01992139406502247, + 0.012561740353703499, + 0.013492729514837265, + 0.004479192662984133, + 0.010646183043718338, + 0.007659419905394316, + 0.005844333674758673, + 0.012141793966293335, + 0.007529277354478836, + 0.006568710785359144, + 0.011295394040644169, + 0.005815152078866959, + 0.009626924991607666, + 0.010063794441521168, + 0.005179083440452814, + 0.008624977432191372, + 0.0075717102736234665, + 0.0048522260040044785, + 0.0086454963311553, + 0.007894417271018028, + 0.017200332134962082, + 0.012605688534677029, + 0.006465897895395756, + 0.016954626888036728, + 0.0210472010076046, + 0.006485398393124342, + 0.01420311164110899, + 0.005227150861173868, + 0.008740215562283993, + 0.006250996608287096, + 0.004886695649474859, + 0.0072151911444962025, + 0.00608747499063611, + 0.0045586987398564816, + 0.009204387664794922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.8261532783508301 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950710736215115 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950710736215115, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950710736215115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0667545273900032, + 0.17045994102954865, + 0.07648947089910507, + 0.08121433109045029, + 0.08489512652158737, + 0.1497681587934494, + 0.07645203173160553, + 0.09211574494838715, + 0.08222988992929459, + 0.11962074041366577 + ], + "credit_map": [ + 0.006153266876935959, + 0.008163763210177422, + 0.008013102225959301, + 0.00492435647174716, + 0.008495770394802094, + 0.0076487110927701, + 0.007923060096800327, + 0.015432501211762428, + 0.010667216964066029, + 0.03050919435918331, + 0.01844758726656437, + 0.01750543713569641, + 0.031168490648269653, + 0.026799190789461136, + 0.012916222214698792, + 0.022446604445576668, + 0.005991126410663128, + 0.00796833261847496, + 0.012790820561349392, + 0.007598099764436483, + 0.00981016643345356, + 0.016195038333535194, + 0.007442052476108074, + 0.00869382917881012, + 0.008607695810496807, + 0.007390861865133047, + 0.011057144030928612, + 0.009958488866686821, + 0.011708960868418217, + 0.011624819599092007, + 0.009863382205367088, + 0.01100297924131155, + 0.010251802392303944, + 0.01133592240512371, + 0.012071849778294563, + 0.008229542523622513, + 0.012936963699758053, + 0.008269267156720161, + 0.009872986935079098, + 0.011926795355975628, + 0.014179266057908535, + 0.020397426560521126, + 0.015163756906986237, + 0.0182944443076849, + 0.018285848200321198, + 0.024369578808546066, + 0.018402446061372757, + 0.020675381645560265, + 0.00762916961684823, + 0.011454874649643898, + 0.009989921934902668, + 0.006889774929732084, + 0.009519223123788834, + 0.012411636300384998, + 0.008343013934791088, + 0.010214414447546005, + 0.008435222320258617, + 0.011393943801522255, + 0.012202777899801731, + 0.010855630971491337, + 0.01384363416582346, + 0.008398940786719322, + 0.015130133368074894, + 0.011855456046760082, + 0.006553231738507748, + 0.013742227107286453, + 0.008765332400798798, + 0.007335151545703411, + 0.01281310711055994, + 0.014533109031617641, + 0.0057016536593437195, + 0.012786072678864002, + 0.013909108005464077, + 0.019746486097574234, + 0.013524500653147697, + 0.009206450544297695, + 0.01670083776116371, + 0.012937977910041809, + 0.011549955233931541, + 0.022045426070690155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.005950710736215115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305529221892357 + ], + "opposite_logprob_tokens": [ + -3.168055295944214 + ], + "logprob_reward": -0.04305529221892357, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305529221892357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14669141173362732, + 0.059220898896455765, + 0.030114999040961266, + 0.06357157230377197, + 0.11729898303747177, + 0.18669258058071136, + 0.05903063341975212, + 0.14259584248065948, + 0.10396996140480042, + 0.09081310033798218 + ], + "credit_map": [ + 0.019148074090480804, + 0.020268479362130165, + 0.020974963903427124, + 0.016379602253437042, + 0.017133597284555435, + 0.01613139733672142, + 0.01629067212343216, + 0.020364606752991676, + 0.005264706909656525, + 0.007974185049533844, + 0.00744015583768487, + 0.007494569756090641, + 0.009132889099419117, + 0.0073187160305678844, + 0.007435388397425413, + 0.007160283159464598, + 0.002137841423973441, + 0.003834682982414961, + 0.003406661096960306, + 0.002730834297835827, + 0.005607234314084053, + 0.00481837335973978, + 0.0036529083736240864, + 0.003926458768546581, + 0.006447058171033859, + 0.005866032559424639, + 0.00761144096031785, + 0.008142012171447277, + 0.008592420257627964, + 0.008957182057201862, + 0.008100478909909725, + 0.009854945354163647, + 0.00853890273720026, + 0.0255049429833889, + 0.013099693693220615, + 0.0072800773195922375, + 0.02046789973974228, + 0.017098180949687958, + 0.009220689535140991, + 0.016088584437966347, + 0.01579202711582184, + 0.024346131831407547, + 0.018920663744211197, + 0.02092624083161354, + 0.02281436324119568, + 0.03519117832183838, + 0.0231021661311388, + 0.025599787011742592, + 0.005583084654062986, + 0.008864580653607845, + 0.007619652431458235, + 0.005225726403295994, + 0.007956943474709988, + 0.00924086943268776, + 0.006006715353578329, + 0.008533056825399399, + 0.012530253268778324, + 0.017992567270994186, + 0.01721506379544735, + 0.015015673823654652, + 0.021638743579387665, + 0.0154972979798913, + 0.021255824714899063, + 0.021450400352478027, + 0.009628905914723873, + 0.017569229006767273, + 0.010749809443950653, + 0.007830704562366009, + 0.016306517645716667, + 0.021080700680613518, + 0.006198116112500429, + 0.014605971984565258, + 0.010830679908394814, + 0.014355898834764957, + 0.012337494641542435, + 0.008317156694829464, + 0.013959706760942936, + 0.013574750162661076, + 0.006454865448176861, + 0.010982532985508442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.04305529221892357 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up transports the viewer into the exquisitely detailed world of a vibrant goldenrod flower, its forms rendered with the painterly precision of classic Makoto Shinkai's aesthetic. The fiery golden petals fill the frame, their texture appearing paper-thin yet delicately soft, each delicate vein visible as a subtle and intricate network branching across the surface. A fine mist gently falls upon the flower, its moisture slowly gathering into rows of tiny crystal droplets that cling to the petals like miniature jewels. At the flower's core, a dark, velvety center creates a profound, captivating point of contrast. Backlit by the warm, low-angle sun, each droplet captures the light, refracting it into a breathtaking spectacle of shimmering prismatic sparkle. This gives the flower's edges an ethereal, luminous halo, evoking a powerful, poignant atmosphere of tranquility and fleeting beauty.", + "seed": 19120, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_15/index_3812_seed_19120_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0702358603477478, + "vlm_logprob": -0.0702358603477478, + "vlm_yes_no_margin": 4.424999988079071, + "vlm_token_credit": -0.0702358603477478, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0033830334432423115 + ], + "opposite_logprob_tokens": [ + -6.753383159637451 + ], + "logprob_reward": -0.0033830334432423115, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0033830334432423115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028640391305088997, + 0.022563820704817772, + 0.019242655485868454, + 0.009719553403556347, + 0.5502591729164124, + 0.019064314663410187, + 0.012862621806561947, + 0.31646162271499634, + 0.01484602503478527, + 0.006339803338050842 + ], + "credit_map": [ + 0.0060120546258986, + 0.0026383919175714254, + 0.00405326159670949, + 0.002802568720653653, + 0.0028291691560298204, + 0.0039925049059093, + 0.002203116426244378, + 0.004109323490411043, + 0.004957242868840694, + 0.0014897322980687022, + 0.001985816052183509, + 0.001250498928129673, + 0.0017850252334028482, + 0.004572431091219187, + 0.002295526210218668, + 0.004227547440677881, + 0.004211293067783117, + 0.0015265311812981963, + 0.002639050828292966, + 0.0008126103202812374, + 0.0013003171188756824, + 0.0032397471368312836, + 0.0009454322862438858, + 0.004567673895508051, + 0.0017166773322969675, + 0.001488048816099763, + 0.0012129711685702205, + 0.0009596332092769444, + 0.0009442892041988671, + 0.0011896865908056498, + 0.0009033455862663686, + 0.0013049019034951925, + 0.21593257784843445, + 0.017863834276795387, + 0.02980973944067955, + 0.006138395052403212, + 0.012093108147382736, + 0.05171774700284004, + 0.008106237277388573, + 0.20859751105308533, + 0.004533039405941963, + 0.0014736518496647477, + 0.0029685194604098797, + 0.0007232202915474772, + 0.0014190502697601914, + 0.003963542636483908, + 0.0010608023731037974, + 0.0029224881436675787, + 0.0025182184763252735, + 0.0011945631122216582, + 0.0016595626948401332, + 0.0009446216863580048, + 0.0011254982091486454, + 0.001594287226907909, + 0.0011094743385910988, + 0.002716396702453494, + 0.10676952451467514, + 0.01631939597427845, + 0.009129093028604984, + 0.005036925431340933, + 0.13390563428401947, + 0.016753781586885452, + 0.007305747829377651, + 0.021241525188088417, + 0.0032487947028130293, + 0.0015154654392972589, + 0.001555955270305276, + 0.000655032810755074, + 0.0031914780847728252, + 0.00225298129953444, + 0.0009438390843570232, + 0.0014824795071035624, + 0.0011868445435538888, + 0.0007608701707795262, + 0.0006507757934741676, + 0.00040842487942427397, + 0.0007486757822334766, + 0.0012027781922370195, + 0.0005181722226552665, + 0.0008632618701085448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.0033830334432423115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016096286475658417 + ], + "opposite_logprob_tokens": [ + -4.141096115112305 + ], + "logprob_reward": -0.016096286475658417, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096286475658417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0598892979323864, + 0.028836311772465706, + 0.03350343182682991, + 0.021686794236302376, + 0.39809751510620117, + 0.0495443120598793, + 0.01427142322063446, + 0.3079843819141388, + 0.06858747452497482, + 0.017599117010831833 + ], + "credit_map": [ + 0.014161189086735249, + 0.004496158566325903, + 0.010654020123183727, + 0.0037873319815844297, + 0.004477077163755894, + 0.008703104220330715, + 0.003419519169256091, + 0.01019089762121439, + 0.006888906005769968, + 0.0019057272002100945, + 0.0038327595684677362, + 0.0015208243858069181, + 0.0019209494348615408, + 0.007831954397261143, + 0.0022238141391426325, + 0.0027113775722682476, + 0.00663723424077034, + 0.0037287743762135506, + 0.0032218173146247864, + 0.0017316722078248858, + 0.003224286250770092, + 0.0045241378247737885, + 0.0021074682008475065, + 0.00832804013043642, + 0.003899632953107357, + 0.002525691408663988, + 0.0030930826906114817, + 0.002151814755052328, + 0.0019695917144417763, + 0.0026665576733648777, + 0.0021817097440361977, + 0.003198714694008231, + 0.15308935940265656, + 0.013087019324302673, + 0.021715650334954262, + 0.004774370696395636, + 0.009846131317317486, + 0.03867216035723686, + 0.006289491429924965, + 0.15062332153320312, + 0.010386328212916851, + 0.0037361744325608015, + 0.008846065029501915, + 0.0016699647530913353, + 0.004138156305998564, + 0.011021277867257595, + 0.00223894277587533, + 0.007507400121539831, + 0.003006847808137536, + 0.0008761624922044575, + 0.001607294543646276, + 0.00045771209988743067, + 0.0017765476368367672, + 0.0027279050555080175, + 0.0007314126123674214, + 0.0030875408556312323, + 0.10009302198886871, + 0.016419554129242897, + 0.008997643366456032, + 0.004972103517502546, + 0.1306997835636139, + 0.01633145473897457, + 0.007348574232310057, + 0.02312224544584751, + 0.0185816939920187, + 0.004414187744259834, + 0.00579610001295805, + 0.0017368217231705785, + 0.023034539073705673, + 0.00630645826458931, + 0.0030206507071852684, + 0.005697022657841444, + 0.004217919427901506, + 0.0016528876731172204, + 0.001524955383501947, + 0.0006644955137744546, + 0.003332772757858038, + 0.0021968449000269175, + 0.0008731258567422628, + 0.003136116312816739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.016096286475658417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344038248062134 + ], + "opposite_logprob_tokens": [ + -1.3134403228759766 + ], + "logprob_reward": -0.31344038248062134, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31344038248062134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03189776837825775, + 0.027105078101158142, + 0.05456976965069771, + 0.025377197191119194, + 0.24683263897895813, + 0.0697263553738594, + 0.024170279502868652, + 0.4804169237613678, + 0.02881532907485962, + 0.011088650673627853 + ], + "credit_map": [ + 0.0048311022110283375, + 0.0034121719654649496, + 0.0050082639791071415, + 0.0028812987729907036, + 0.0029105916619300842, + 0.005339484196156263, + 0.002907796995714307, + 0.004607058130204678, + 0.005202137399464846, + 0.0025710132904350758, + 0.0040292213670909405, + 0.0020350127015262842, + 0.002720929216593504, + 0.00427956972271204, + 0.0017096723895519972, + 0.004557523876428604, + 0.01282262522727251, + 0.003961038310080767, + 0.0076744756661355495, + 0.002824687398970127, + 0.003539304481819272, + 0.010475999675691128, + 0.0026004379615187645, + 0.010671200230717659, + 0.006622305605560541, + 0.0027749838773161173, + 0.0032169960904866457, + 0.0022543789818882942, + 0.0022013895213603973, + 0.0033777409698814154, + 0.002038007602095604, + 0.0028913964051753283, + 0.09754674881696701, + 0.007940338924527168, + 0.014184858649969101, + 0.004214798100292683, + 0.005887332372367382, + 0.021279476583003998, + 0.004161703400313854, + 0.09161737561225891, + 0.010893112979829311, + 0.006932666990906, + 0.01228887215256691, + 0.003902107011526823, + 0.00835396721959114, + 0.013659017160534859, + 0.003987657371908426, + 0.009708955883979797, + 0.004693387076258659, + 0.0017758639296516776, + 0.0023739770986139774, + 0.0009969373932108283, + 0.003955406602472067, + 0.004242835566401482, + 0.0012031568912789226, + 0.0049287122674286366, + 0.15799321234226227, + 0.025256581604480743, + 0.01418302021920681, + 0.007715277373790741, + 0.20415419340133667, + 0.02592695690691471, + 0.011182625778019428, + 0.03400503471493721, + 0.006512001156806946, + 0.0024792489130049944, + 0.004187596961855888, + 0.001251877867616713, + 0.006916120182722807, + 0.00335554126650095, + 0.0017574315425008535, + 0.0023555117659270763, + 0.0029904674738645554, + 0.0012331464095041156, + 0.001114016748033464, + 0.0005843662074767053, + 0.0015710598090663552, + 0.0014852517051622272, + 0.0007061510696075857, + 0.0014041911344975233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.31344038248062134 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195806708186865 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195806708186865, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195806708186865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028788287192583084, + 0.03754628822207451, + 0.03432730957865715, + 0.005731410812586546, + 0.6403865218162537, + 0.012722432613372803, + 0.010968157090246677, + 0.20321333408355713, + 0.007195819169282913, + 0.01912045106291771 + ], + "credit_map": [ + 0.004298524931073189, + 0.0033546139020472765, + 0.004470349755138159, + 0.0031353733502328396, + 0.004273159895092249, + 0.0035547958686947823, + 0.00291391066275537, + 0.002787560224533081, + 0.009287022985517979, + 0.002341912593692541, + 0.003806080436334014, + 0.00241368031129241, + 0.0031557457987219095, + 0.006619271822273731, + 0.002552193822339177, + 0.007370381150394678, + 0.007964765653014183, + 0.002866716356948018, + 0.004750846419483423, + 0.002050132257863879, + 0.0021933559328317642, + 0.0060504041612148285, + 0.0018725590780377388, + 0.00657853065058589, + 0.0011800153879448771, + 0.0006267275894060731, + 0.0007815729477442801, + 0.0005729762488044798, + 0.00048223588964901865, + 0.0007755582919344306, + 0.0005263915518298745, + 0.0007859324687160552, + 0.25202786922454834, + 0.020632484927773476, + 0.03511075675487518, + 0.007159325294196606, + 0.014141933061182499, + 0.05978047475218773, + 0.009460007771849632, + 0.2420736849308014, + 0.002783953445032239, + 0.0012763580307364464, + 0.0021277049090713263, + 0.0007323417230509222, + 0.0011264784261584282, + 0.0020847930572927, + 0.0007552591268904507, + 0.0018355438951402903, + 0.0018579578027129173, + 0.001040720148012042, + 0.0009261375525966287, + 0.000677157542668283, + 0.0025489849504083395, + 0.0015969440573826432, + 0.0007591655594296753, + 0.0015610894188284874, + 0.06622696667909622, + 0.010875716805458069, + 0.005819350015372038, + 0.003267691470682621, + 0.08639825135469437, + 0.010598117485642433, + 0.004750690422952175, + 0.015276551246643066, + 0.0013698928523808718, + 0.0008410955197177827, + 0.0010870065307244658, + 0.00046839614515192807, + 0.000957957177888602, + 0.0010841182665899396, + 0.0005548063199967146, + 0.0008325461531057954, + 0.004017460159957409, + 0.002150090876966715, + 0.002319574821740389, + 0.0013549349969252944, + 0.002931166673079133, + 0.0025547172408550978, + 0.0014326893724501133, + 0.002359817735850811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.002195806708186865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063792631030083 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.016063792631030083, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063792631030083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02417564205825329, + 0.027609243988990784, + 0.02728862687945366, + 0.010614761151373386, + 0.377870112657547, + 0.017860906198620796, + 0.011277416720986366, + 0.47889581322669983, + 0.017855605110526085, + 0.006551856640726328 + ], + "credit_map": [ + 0.0037556528113782406, + 0.002514645690098405, + 0.003925689961761236, + 0.0026763584464788437, + 0.002998490585014224, + 0.0029741970356553793, + 0.0025623608380556107, + 0.002768246689811349, + 0.007072708569467068, + 0.0019783328752964735, + 0.003924366552382708, + 0.002579573541879654, + 0.00211940030567348, + 0.005291800480335951, + 0.0017410283908247948, + 0.0029020339716225863, + 0.006470558233559132, + 0.0016277618706226349, + 0.0037915578577667475, + 0.0010574859334155917, + 0.0016365693882107735, + 0.005087609868496656, + 0.0010598186636343598, + 0.006557266227900982, + 0.002175410045310855, + 0.0011602946324273944, + 0.0014172198716551065, + 0.0012947347713634372, + 0.000917085213586688, + 0.0012860685819759965, + 0.0010705937165766954, + 0.0012933540856465697, + 0.14491118490695953, + 0.012165538966655731, + 0.02165166661143303, + 0.004431526642292738, + 0.009112704545259476, + 0.03729122504591942, + 0.005852341186255217, + 0.14245390892028809, + 0.003878579940646887, + 0.001968845259398222, + 0.002816983964294195, + 0.0010592052713036537, + 0.0016607012366876006, + 0.003177824430167675, + 0.0010242204880341887, + 0.0022745467722415924, + 0.0016167615540325642, + 0.0010189000749960542, + 0.0014964130241423845, + 0.0005698113236576319, + 0.002413490554317832, + 0.0018999275052919984, + 0.0008983496227301657, + 0.0013637634692713618, + 0.15481042861938477, + 0.0257773045450449, + 0.014130859635770321, + 0.007686967961490154, + 0.20207171142101288, + 0.02624415047466755, + 0.011630027554929256, + 0.03654434159398079, + 0.0038017318584024906, + 0.0015374200884252787, + 0.0018030359642580152, + 0.0007363077602349222, + 0.004286336246877909, + 0.0025929296389222145, + 0.0011513615027070045, + 0.0019464808283373713, + 0.0012080444721505046, + 0.00045405892888084054, + 0.0006583010545000434, + 0.0002739196643233299, + 0.0011678446317091584, + 0.0008547698380425572, + 0.0005770102143287659, + 0.0013579076621681452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.016063792631030083 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up transports the viewer into the exquisitely detailed world of a vibrant goldenrod flower, its forms rendered with the painterly precision of classic Makoto Shinkai's aesthetic. The fiery golden petals fill the frame, their texture appearing paper-thin yet delicately soft, each delicate vein visible as a subtle and intricate network branching across the surface. A fine mist gently falls upon the flower, its moisture slowly gathering into rows of tiny crystal droplets that cling to the petals like miniature jewels. At the flower's core, a dark, velvety center creates a profound, captivating point of contrast. Backlit by the warm, low-angle sun, each droplet captures the light, refracting it into a breathtaking spectacle of shimmering prismatic sparkle. This gives the flower's edges an ethereal, luminous halo, evoking a powerful, poignant atmosphere of tranquility and fleeting beauty.", + "seed": 19121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_15/index_3812_seed_19121_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1065688518807292, + "vlm_logprob": -0.1065688518807292, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.1065688518807292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0034638189245015383 + ], + "opposite_logprob_tokens": [ + -6.2534637451171875 + ], + "logprob_reward": -0.0034638189245015383, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0034638189245015383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054377105087041855, + 0.027396956458687782, + 0.022390663623809814, + 0.03372761607170105, + 0.09570787847042084, + 0.1117134615778923, + 0.01328201126307249, + 0.011332501657307148, + 0.6000633239746094, + 0.030008438974618912 + ], + "credit_map": [ + 0.00526790414005518, + 0.006488149520009756, + 0.008689404465258121, + 0.00680393585935235, + 0.0054584043100476265, + 0.007362638134509325, + 0.006406071130186319, + 0.007900603115558624, + 0.003582277335226536, + 0.003354024840518832, + 0.00418733898550272, + 0.0029737972654402256, + 0.0027246843092143536, + 0.0035893390886485577, + 0.004437445197254419, + 0.002548050833866, + 0.0022302421275526285, + 0.002987163607031107, + 0.00351240998134017, + 0.002076265634968877, + 0.0033148941583931446, + 0.00377547275274992, + 0.00260330643504858, + 0.0018909084610641003, + 0.0034087745007127523, + 0.0038708678912371397, + 0.004518704488873482, + 0.0037248374428600073, + 0.003657140303403139, + 0.006013641599565744, + 0.004720916505903006, + 0.003812735201790929, + 0.00872866716235876, + 0.011632680892944336, + 0.01763809099793434, + 0.011629091575741768, + 0.008171595633029938, + 0.011185656301677227, + 0.014105087146162987, + 0.012617003172636032, + 0.008934765122830868, + 0.013671413995325565, + 0.020489955320954323, + 0.015768583863973618, + 0.009115382097661495, + 0.013449656777083874, + 0.014100629836320877, + 0.016183072701096535, + 0.0011367001570761204, + 0.0016926303505897522, + 0.0019041831837967038, + 0.0013438075548037887, + 0.002348522422835231, + 0.002140684751793742, + 0.0014160978607833385, + 0.0012993842829018831, + 0.0010557299247011542, + 0.001318582333624363, + 0.0017526313895359635, + 0.0013257378013804555, + 0.0012408153852447867, + 0.0011456573847681284, + 0.0017618286656215787, + 0.0017315183067694306, + 0.020755954086780548, + 0.27074703574180603, + 0.22822508215904236, + 0.018021631985902786, + 0.030129192396998405, + 0.007617854047566652, + 0.007010305766016245, + 0.01755625754594803, + 0.0011676611611619592, + 0.012724731117486954, + 0.010158387012779713, + 0.001201392151415348, + 0.0016986182890832424, + 0.0008020339300855994, + 0.0009129407699219882, + 0.0013426733203232288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.0034638189245015383 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374212980270386 + ], + "opposite_logprob_tokens": [ + -3.4087421894073486 + ], + "logprob_reward": -0.03374212980270386, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374212980270386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13226920366287231, + 0.06388825178146362, + 0.03386615589261055, + 0.017010416835546494, + 0.08075068145990372, + 0.08482315391302109, + 0.016986053436994553, + 0.02221682108938694, + 0.42187434434890747, + 0.12631487846374512 + ], + "credit_map": [ + 0.013523564673960209, + 0.01385718584060669, + 0.019106322899460793, + 0.017996812239289284, + 0.012118673883378506, + 0.01887204870581627, + 0.015647830441594124, + 0.02114677242934704, + 0.005680990871042013, + 0.007044412661343813, + 0.008550981059670448, + 0.007598805241286755, + 0.008710280992090702, + 0.00624160235747695, + 0.008577637374401093, + 0.01148354820907116, + 0.0033999832812696695, + 0.004608370363712311, + 0.004844698589295149, + 0.003758044447749853, + 0.004872960038483143, + 0.0034125109668821096, + 0.005259282421320677, + 0.0037103071808815002, + 0.0017089815810322762, + 0.0019278698600828648, + 0.002534693107008934, + 0.0019732166547328234, + 0.0015508693177253008, + 0.002494797809049487, + 0.0028645277488976717, + 0.0019554607570171356, + 0.005881840363144875, + 0.009533376432955265, + 0.015839241445064545, + 0.009953804314136505, + 0.007482124026864767, + 0.008086205460131168, + 0.01204605307430029, + 0.011928041465580463, + 0.006717395968735218, + 0.01009348127990961, + 0.015116061083972454, + 0.012465277686715126, + 0.006897824350744486, + 0.009778423234820366, + 0.010832360945641994, + 0.012922342866659164, + 0.001468813861720264, + 0.0017808976117521524, + 0.0030493189115077257, + 0.002211633138358593, + 0.0022712720092386007, + 0.0014147100737318397, + 0.0030155403073877096, + 0.0017738672904670238, + 0.0024690001737326384, + 0.0023228421341627836, + 0.0032426631078124046, + 0.0028558033518493176, + 0.0023795049637556076, + 0.0017142889555543661, + 0.004432955291122198, + 0.002799765905365348, + 0.017103280872106552, + 0.18361234664916992, + 0.15667186677455902, + 0.012858670204877853, + 0.022305132821202278, + 0.007332443259656429, + 0.007210313808172941, + 0.01478034071624279, + 0.005418146029114723, + 0.049057092517614365, + 0.039084695279598236, + 0.007463360205292702, + 0.007207460235804319, + 0.004800346679985523, + 0.0061521027237176895, + 0.007131682243198156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.03374212980270386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742666482925415 + ], + "opposite_logprob_tokens": [ + -0.9742666482925415 + ], + "logprob_reward": -0.4742666482925415, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742666482925415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09132024645805359, + 0.020816056057810783, + 0.034605517983436584, + 0.019367752596735954, + 0.025638584047555923, + 0.020169463008642197, + 0.04484047368168831, + 0.10777153074741364, + 0.614594578742981, + 0.020875820890069008 + ], + "credit_map": [ + 0.011596162803471088, + 0.013955596834421158, + 0.016598301008343697, + 0.008132075890898705, + 0.011336795054376125, + 0.01197353471070528, + 0.008501794189214706, + 0.009225980378687382, + 0.0019666252192109823, + 0.00252819643355906, + 0.0023381318897008896, + 0.0021278648637235165, + 0.0028903991915285587, + 0.002157176611945033, + 0.0023045551497489214, + 0.004503107629716396, + 0.0024837476667016745, + 0.004388880915939808, + 0.00562333082780242, + 0.004164411686360836, + 0.0057324995286762714, + 0.0030515557155013084, + 0.00457938015460968, + 0.004581708926707506, + 0.0019465174991637468, + 0.0029425155371427536, + 0.0030301425140351057, + 0.0018260821234434843, + 0.0018223345978185534, + 0.00234555103816092, + 0.0035304820630699396, + 0.0019241278059780598, + 0.002102826489135623, + 0.0028353992383927107, + 0.004352830816060305, + 0.0030972836539149284, + 0.0020697456784546375, + 0.003075685817748308, + 0.0055263168178498745, + 0.0025784955359995365, + 0.002047720132395625, + 0.002587549388408661, + 0.0033176643773913383, + 0.0023536665830761194, + 0.0019403131445869803, + 0.002469740342348814, + 0.0023921779356896877, + 0.0030606298241764307, + 0.0036637159064412117, + 0.0050356704741716385, + 0.007637896109372377, + 0.005968393292278051, + 0.004464302211999893, + 0.0057427408173680305, + 0.006781040225178003, + 0.005546713713556528, + 0.009596498683094978, + 0.011041609570384026, + 0.014913090504705906, + 0.01228207629173994, + 0.01131200697273016, + 0.01328486017882824, + 0.019468089565634727, + 0.015873299911618233, + 0.02804226242005825, + 0.22357018291950226, + 0.19486907124519348, + 0.03784115985035896, + 0.04257446527481079, + 0.02045559138059616, + 0.03959038481116295, + 0.02765142358839512, + 0.0011332316789776087, + 0.006588499061763287, + 0.005374641623347998, + 0.001955993939191103, + 0.0012790376786142588, + 0.0010480002965778112, + 0.0018509574001654983, + 0.0016454588621854782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.4742666482925415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08936089277267456, + 0.03710322827100754, + 0.022105911746621132, + 0.020485050976276398, + 0.013420558534562588, + 0.051999758929014206, + 0.028286781162023544, + 0.04762531816959381, + 0.6490588188171387, + 0.04055369272828102 + ], + "credit_map": [ + 0.009040139615535736, + 0.009113851934671402, + 0.011274319142103195, + 0.012569022364914417, + 0.009201323613524437, + 0.013130735605955124, + 0.012111622840166092, + 0.012919879518449306, + 0.002975677140057087, + 0.0038690681103616953, + 0.005089342128485441, + 0.005238715559244156, + 0.003771434770897031, + 0.0026166443713009357, + 0.00578627735376358, + 0.007756069768220186, + 0.001769660972058773, + 0.002725716447457671, + 0.003219632199034095, + 0.0025898893363773823, + 0.003007042221724987, + 0.0016395376296713948, + 0.004160269629210234, + 0.0029941655229777098, + 0.0014348485274240375, + 0.0026215361431241035, + 0.0039033200591802597, + 0.002254818333312869, + 0.0025222788099199533, + 0.0021160761825740337, + 0.0030168986413627863, + 0.002615273930132389, + 0.0015015574172139168, + 0.00160882156342268, + 0.002382737584412098, + 0.001561728771775961, + 0.001317542977631092, + 0.0011527056340128183, + 0.0017659402219578624, + 0.002129525877535343, + 0.0041803522035479546, + 0.006356677506119013, + 0.0094452528283, + 0.006809354294091463, + 0.004522264003753662, + 0.006371290888637304, + 0.006398946512490511, + 0.007915626280009747, + 0.002329515526071191, + 0.0030114988330751657, + 0.0036157257854938507, + 0.00421185651794076, + 0.0036520983558148146, + 0.0032324797939509153, + 0.004786815494298935, + 0.003446792485192418, + 0.0040837000124156475, + 0.005834170617163181, + 0.008261756971478462, + 0.005432446487247944, + 0.0047911894507706165, + 0.004659039434045553, + 0.007479882799088955, + 0.007083131466060877, + 0.02444126084446907, + 0.28428396582603455, + 0.24153195321559906, + 0.02111896127462387, + 0.03505169600248337, + 0.009222992695868015, + 0.011093231849372387, + 0.02231479249894619, + 0.0025793223176151514, + 0.007887205109000206, + 0.006441496778279543, + 0.006328776478767395, + 0.003189204493537545, + 0.0031432711984962225, + 0.006116956938058138, + 0.004867460113018751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.0031914988067001104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant goldenrod flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the petals appear fiery golden with a paper-thin texture and visible intricate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fine mist falling on the flower, forming tiny crystal droplets on the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower's core dark and velvety, creating a contrast with the petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a warm, low-angle sun causing prismatic sparkle and a luminous halo on the edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180163577198982 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180163577198982, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180163577198982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09254293888807297, + 0.023199327290058136, + 0.023419007658958435, + 0.013211834244430065, + 0.029172485694289207, + 0.014075914397835732, + 0.0347599983215332, + 0.036050960421562195, + 0.6649655103683472, + 0.06860209256410599 + ], + "credit_map": [ + 0.010117348283529282, + 0.014392111450433731, + 0.015841122716665268, + 0.009771446697413921, + 0.011981144547462463, + 0.011369445361196995, + 0.009693130850791931, + 0.009377190843224525, + 0.002160493051633239, + 0.0025994302704930305, + 0.00277267605997622, + 0.002575415186583996, + 0.0029372298158705235, + 0.002636657329276204, + 0.0030159454327076674, + 0.00450148107483983, + 0.0017307590460404754, + 0.002866108901798725, + 0.0035109124146401882, + 0.0030007848981767893, + 0.0034133547451347113, + 0.002222666749730706, + 0.0031992264557629824, + 0.003475195961073041, + 0.0008682251791469753, + 0.0018831791821867228, + 0.002204335294663906, + 0.001421247492544353, + 0.0016959374770522118, + 0.0023057693615555763, + 0.001616592868231237, + 0.0012165468651801348, + 0.0031265649013221264, + 0.003497934201732278, + 0.004989143460988998, + 0.003330746665596962, + 0.0029923978727310896, + 0.004174266010522842, + 0.003700768342241645, + 0.0033606637734919786, + 0.0012812444474548101, + 0.0016745313769206405, + 0.0022481707856059074, + 0.00201476295478642, + 0.0014756801538169384, + 0.0015599699690937996, + 0.0022128194104880095, + 0.0016087358817458153, + 0.002699025673791766, + 0.004331726115196943, + 0.0048779514618217945, + 0.003802259685471654, + 0.005519482772797346, + 0.0031391605734825134, + 0.0065724956803023815, + 0.003817894496023655, + 0.003281485754996538, + 0.003521299222484231, + 0.0047116936184465885, + 0.005432407837361097, + 0.004008684307336807, + 0.0033502038568258286, + 0.007833692245185375, + 0.003911492880433798, + 0.025769084692001343, + 0.28333956003189087, + 0.24327217042446136, + 0.025217700749635696, + 0.03728550300002098, + 0.012390747666358948, + 0.018201587721705437, + 0.01948908530175686, + 0.0042247832752764225, + 0.024004245176911354, + 0.019478587433695793, + 0.004671573173254728, + 0.00434085913002491, + 0.00347410817630589, + 0.0034864170011132956, + 0.004921517800539732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.018180163577198982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_tensor([1141, 1141])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_tensor([1141, 1141])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..41e4cf88e59bdbc7e99b6c3d22347509db3f8a1f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_tensor([1141, 1141])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, macro, dolly shot moves horizontally across the surface of a gleaming, polished black granite kitchen counter, beginning a slow, sweeping journey that transforms the scene into a vast miniature landscape. Warm, golden morning sunlight streams in through an unseen window, casting long, dramatic shadows and illuminating motes of dust that drift like tiny stars. The camera glides over soft drifts and dunes of fine white flour, its powdery texture catching the light and creating the illusion of a snow-capped terrain with delicate peaks and valleys. Stainless steel utensils rise from this floury plain; the tines of forks stand upright like a row of metallic monoliths, while the deep, curved bowls of spoons form a shadowy, reflective basin. The camera continues its steady and deliberate path, revealing a wire whisk lying on its side, its intricate structure resembling a strange, skeletal bush. The entire scene is imbued with a tranquil and magical atmosphere, the warm light interacting on the cool, reflective surfaces of the counter and steel, creating a quiet, awe-inspiring world born from the beautiful chaos of baking.", + "seed": 8955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_1141_seed_8955_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06519515435211361, + "vlm_logprob": -0.06519515435211361, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.06519515435211361, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21318815648555756 + ], + "opposite_logprob_tokens": [ + -1.7131881713867188 + ], + "logprob_reward": -0.21318815648555756, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21318815648555756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09011099487543106, + 0.26813334226608276, + 0.03991531953215599, + 0.08989594876766205, + 0.06495065242052078, + 0.04194888100028038, + 0.1298026293516159, + 0.05202016234397888, + 0.04838087782263756, + 0.17484121024608612 + ], + "credit_map": [ + 0.010234122164547443, + 0.007140837609767914, + 0.012252199463546276, + 0.018338041380047798, + 0.0071868933737277985, + 0.00808845367282629, + 0.012409084476530552, + 0.014461366459727287, + 0.02007311023771763, + 0.011349992826581001, + 0.05678578466176987, + 0.026630792766809464, + 0.01825658418238163, + 0.010187860578298569, + 0.08727807551622391, + 0.03757111728191376, + 0.003455378580838442, + 0.003899315372109413, + 0.0045071193017065525, + 0.003717950312420726, + 0.00576213700696826, + 0.007113409228622913, + 0.00552387023344636, + 0.00593614112585783, + 0.007677872199565172, + 0.009372211992740631, + 0.01051050703972578, + 0.011558135971426964, + 0.011134572327136993, + 0.009952410124242306, + 0.018081603571772575, + 0.011608638800680637, + 0.004206008743494749, + 0.007417592220008373, + 0.0087944520637393, + 0.015119412913918495, + 0.008053760044276714, + 0.0058802226558327675, + 0.008238309063017368, + 0.007240894716233015, + 0.0037981101777404547, + 0.004753557965159416, + 0.005663648247718811, + 0.006853986997157335, + 0.004236482549458742, + 0.003614228218793869, + 0.007428008131682873, + 0.005600858945399523, + 0.016279470175504684, + 0.015177462249994278, + 0.022400524467229843, + 0.02060476876795292, + 0.010096157900989056, + 0.009681092575192451, + 0.017969055101275444, + 0.017594100907444954, + 0.0065801674500107765, + 0.0057122670114040375, + 0.01165533997118473, + 0.010089093819260597, + 0.005343802738934755, + 0.00268959510140121, + 0.0055152419954538345, + 0.004434655420482159, + 0.004473078530281782, + 0.004910269286483526, + 0.007367054931819439, + 0.009003570303320885, + 0.004402632359415293, + 0.0035154768265783787, + 0.006428646855056286, + 0.008280148729681969, + 0.016841838136315346, + 0.01700901798903942, + 0.027380941435694695, + 0.02986188977956772, + 0.024314800277352333, + 0.012023644521832466, + 0.026561066508293152, + 0.020848015323281288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.21318815648555756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870099201798439 + ], + "opposite_logprob_tokens": [ + -3.048701047897339 + ], + "logprob_reward": -0.04870099201798439, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870099201798439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15957516431808472, + 0.1593971997499466, + 0.09100653231143951, + 0.11397607624530792, + 0.06444527208805084, + 0.14157947897911072, + 0.038156937807798386, + 0.06264351308345795, + 0.09356378763914108, + 0.07565593719482422 + ], + "credit_map": [ + 0.014682581648230553, + 0.01176118291914463, + 0.02329159528017044, + 0.03643544018268585, + 0.01183045282959938, + 0.012705892324447632, + 0.020764006301760674, + 0.02810402400791645, + 0.015273881144821644, + 0.009857496246695518, + 0.03467939794063568, + 0.0207836776971817, + 0.01083934586495161, + 0.009701726026833057, + 0.03829420730471611, + 0.019967488944530487, + 0.008411865681409836, + 0.00938763190060854, + 0.012139994651079178, + 0.007456252817064524, + 0.013621560297906399, + 0.015454079955816269, + 0.010598056018352509, + 0.013937104493379593, + 0.011882242746651173, + 0.015171325765550137, + 0.013294561766088009, + 0.010407981462776661, + 0.025547655299305916, + 0.01369235385209322, + 0.011920677497982979, + 0.012059293687343597, + 0.005477686878293753, + 0.006937559228390455, + 0.008588221855461597, + 0.012508001178503036, + 0.007083415985107422, + 0.005998085252940655, + 0.010096639394760132, + 0.007755671162158251, + 0.013055682182312012, + 0.016630589962005615, + 0.020241158083081245, + 0.024081727489829063, + 0.01000417210161686, + 0.012198655866086483, + 0.024812715128064156, + 0.0205547995865345, + 0.006978388875722885, + 0.005320585798472166, + 0.004881923086941242, + 0.004803991876542568, + 0.0032133893109858036, + 0.0037502499762922525, + 0.005976435728371143, + 0.003231974318623543, + 0.007976398803293705, + 0.006784741301089525, + 0.01104830950498581, + 0.01184853632003069, + 0.0043864562176167965, + 0.004407947417348623, + 0.006993952672928572, + 0.009197178296744823, + 0.015205452218651772, + 0.010855719447135925, + 0.016310643404722214, + 0.015332430601119995, + 0.008332338184118271, + 0.005618728697299957, + 0.009205695241689682, + 0.012702787294983864, + 0.007773146033287048, + 0.009029325097799301, + 0.012136406265199184, + 0.01200682483613491, + 0.00927276723086834, + 0.00534201692789793, + 0.010532296262681484, + 0.009563160128891468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.04870099201798439 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028334015514701605 + ], + "opposite_logprob_tokens": [ + -5.877833366394043 + ], + "logprob_reward": -0.0028334015514701605, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028334015514701605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2762589752674103, + 0.06870589405298233, + 0.02747923694550991, + 0.08791682869195938, + 0.05628037452697754, + 0.08539289236068726, + 0.06993965804576874, + 0.0774230808019638, + 0.1344064623117447, + 0.11619659513235092 + ], + "credit_map": [ + 0.03102608025074005, + 0.02410208061337471, + 0.038440316915512085, + 0.067803293466568, + 0.021878348663449287, + 0.022318411618471146, + 0.036588896065950394, + 0.034101568162441254, + 0.005766733083873987, + 0.002310609444975853, + 0.01477342750877142, + 0.006801601033657789, + 0.004337225575000048, + 0.004231089726090431, + 0.02171558327972889, + 0.008769625797867775, + 0.002029749797657132, + 0.00222115870565176, + 0.004098774399608374, + 0.003290056949481368, + 0.0034451244864612818, + 0.00442701019346714, + 0.003916298039257526, + 0.004051066469401121, + 0.006263687741011381, + 0.011480853892862797, + 0.010888226330280304, + 0.00960923358798027, + 0.01737585850059986, + 0.008140519261360168, + 0.017079781740903854, + 0.007078670430928469, + 0.004763112869113684, + 0.0052499920129776, + 0.008494113571941853, + 0.010431363247334957, + 0.006586594507098198, + 0.0036697268951684237, + 0.010399570688605309, + 0.006685899570584297, + 0.008202686905860901, + 0.009497041814029217, + 0.012713605538010597, + 0.012786045670509338, + 0.006147639360278845, + 0.006911688018590212, + 0.01661515422165394, + 0.01251902710646391, + 0.008365790359675884, + 0.005944521632045507, + 0.010750211775302887, + 0.013713483698666096, + 0.004815129563212395, + 0.005903242155909538, + 0.010808657854795456, + 0.009638618677854538, + 0.008366056717932224, + 0.008863781578838825, + 0.010319825261831284, + 0.013979082927107811, + 0.009181134402751923, + 0.004253878723829985, + 0.013238918036222458, + 0.00922040082514286, + 0.016679169610142708, + 0.017609139904379845, + 0.017834389582276344, + 0.02182445488870144, + 0.015085914172232151, + 0.009846013970673084, + 0.014069964177906513, + 0.021457407623529434, + 0.011923234909772873, + 0.01905515417456627, + 0.018021337687969208, + 0.02082056552171707, + 0.017355117946863174, + 0.0067205289378762245, + 0.010713385418057442, + 0.011587277054786682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0028334015514701605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201468512415886 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201468512415886, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201468512415886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1343877911567688, + 0.07584340125322342, + 0.02810083143413067, + 0.12684613466262817, + 0.08021717518568039, + 0.11448159068822861, + 0.1536213755607605, + 0.06152457371354103, + 0.03873833641409874, + 0.18623872101306915 + ], + "credit_map": [ + 0.008752410300076008, + 0.008859753608703613, + 0.016949603334069252, + 0.0309340450912714, + 0.012151074595749378, + 0.010223033837974072, + 0.02010251022875309, + 0.026415368542075157, + 0.007568599656224251, + 0.004004301503300667, + 0.00903232954442501, + 0.01065538078546524, + 0.005635298788547516, + 0.005116682033985853, + 0.019938785582780838, + 0.013892020098865032, + 0.0023245392367243767, + 0.0023415195755660534, + 0.0031370932701975107, + 0.0023714997805655003, + 0.004029665607959032, + 0.005092723295092583, + 0.004761119373142719, + 0.004042672459036112, + 0.010521803982555866, + 0.0159926638007164, + 0.016558440402150154, + 0.015857581049203873, + 0.01310809701681137, + 0.009973359294235706, + 0.029902290552854538, + 0.014931910671293736, + 0.006049405317753553, + 0.005713823717087507, + 0.012967345304787159, + 0.020859306678175926, + 0.00667811231687665, + 0.004886371083557606, + 0.014290621504187584, + 0.008772189728915691, + 0.012162626720964909, + 0.012323764152824879, + 0.017227377742528915, + 0.016672424972057343, + 0.007180301006883383, + 0.009897996671497822, + 0.021651294082403183, + 0.01736580953001976, + 0.018322372809052467, + 0.013459304347634315, + 0.024275686591863632, + 0.0284455306828022, + 0.010144278407096863, + 0.012099718675017357, + 0.022371750324964523, + 0.02450273185968399, + 0.007430416531860828, + 0.007861337624490261, + 0.009950653649866581, + 0.008499766699969769, + 0.004861775320023298, + 0.006983454804867506, + 0.008055653423070908, + 0.00788151752203703, + 0.006296979263424873, + 0.004354196600615978, + 0.007193224038928747, + 0.006166660226881504, + 0.003464742098003626, + 0.0023128753527998924, + 0.004261812195181847, + 0.004687848035246134, + 0.017271412536501884, + 0.02989153377711773, + 0.026946302503347397, + 0.03312120586633682, + 0.0263352207839489, + 0.010475020855665207, + 0.02138124592602253, + 0.02081678993999958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.018201468512415886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305175319314003 + ], + "opposite_logprob_tokens": [ + -3.1680517196655273 + ], + "logprob_reward": -0.04305175319314003, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305175319314003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15205144882202148, + 0.08522988855838776, + 0.03289973735809326, + 0.08933179825544357, + 0.0507345050573349, + 0.045836251229047775, + 0.038569144904613495, + 0.06857824325561523, + 0.32075148820877075, + 0.11601750552654266 + ], + "credit_map": [ + 0.010660411790013313, + 0.0123030636459589, + 0.018045911565423012, + 0.02531566470861435, + 0.020042501389980316, + 0.008780410513281822, + 0.022095937281847, + 0.03480754420161247, + 0.007280171848833561, + 0.003492501564323902, + 0.017702866345643997, + 0.010227964259684086, + 0.007123423274606466, + 0.005980323534458876, + 0.02209550514817238, + 0.011327126063406467, + 0.0029844464734196663, + 0.004090416710823774, + 0.0044549149461090565, + 0.0038291786331683397, + 0.004270204342901707, + 0.004623663146048784, + 0.004387481603771448, + 0.004259430803358555, + 0.006239891517907381, + 0.010628888383507729, + 0.011424924246966839, + 0.007121377158910036, + 0.017468539997935295, + 0.00902049895375967, + 0.015071707777678967, + 0.012355967424809933, + 0.004110296256840229, + 0.006299052853137255, + 0.006989792920649052, + 0.009200118482112885, + 0.006506270729005337, + 0.004039887338876724, + 0.00779059948399663, + 0.005798483733087778, + 0.0044816951267421246, + 0.004197509028017521, + 0.007453079801052809, + 0.009281456470489502, + 0.003675163025036454, + 0.0029674407560378313, + 0.009205407463014126, + 0.004574498161673546, + 0.003581209806725383, + 0.0030163624323904514, + 0.00445889588445425, + 0.006728076841682196, + 0.0032260899897664785, + 0.0033159248996526003, + 0.006231473293155432, + 0.008011114783585072, + 0.007832633331418037, + 0.00888850074261427, + 0.013848516158759594, + 0.011959868483245373, + 0.005694475956261158, + 0.006200452335178852, + 0.006931704003363848, + 0.007222088519483805, + 0.0348762683570385, + 0.035529911518096924, + 0.043606773018836975, + 0.051266029477119446, + 0.03369022533297539, + 0.021513186395168304, + 0.042509157210588455, + 0.057759951800107956, + 0.010810686275362968, + 0.0174836665391922, + 0.01653324067592621, + 0.01959204487502575, + 0.01633889228105545, + 0.006287345662713051, + 0.012372625060379505, + 0.016599006950855255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.04305175319314003 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, macro, dolly shot moves horizontally across the surface of a gleaming, polished black granite kitchen counter, beginning a slow, sweeping journey that transforms the scene into a vast miniature landscape. Warm, golden morning sunlight streams in through an unseen window, casting long, dramatic shadows and illuminating motes of dust that drift like tiny stars. The camera glides over soft drifts and dunes of fine white flour, its powdery texture catching the light and creating the illusion of a snow-capped terrain with delicate peaks and valleys. Stainless steel utensils rise from this floury plain; the tines of forks stand upright like a row of metallic monoliths, while the deep, curved bowls of spoons form a shadowy, reflective basin. The camera continues its steady and deliberate path, revealing a wire whisk lying on its side, its intricate structure resembling a strange, skeletal bush. The entire scene is imbued with a tranquil and magical atmosphere, the warm light interacting on the cool, reflective surfaces of the counter and steel, creating a quiet, awe-inspiring world born from the beautiful chaos of baking.", + "seed": 8956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_1141_seed_8956_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03812635084614158, + "vlm_logprob": -0.03812635084614158, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03812635084614158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09469000995159149 + ], + "opposite_logprob_tokens": [ + -2.5946900844573975 + ], + "logprob_reward": -0.09469000995159149, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09469000995159149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42083802819252014, + 0.07237409055233002, + 0.07188256084918976, + 0.08277655392885208, + 0.0484338216483593, + 0.08929871022701263, + 0.04519229382276535, + 0.08192450553178787, + 0.053737036883831024, + 0.03354237973690033 + ], + "credit_map": [ + 0.025855597108602524, + 0.01727048121392727, + 0.12987832725048065, + 0.03776463121175766, + 0.020051557570695877, + 0.019131634384393692, + 0.14617770910263062, + 0.024708110839128494, + 0.005220521707087755, + 0.005372039973735809, + 0.007425762712955475, + 0.020501889288425446, + 0.006916915997862816, + 0.005236462224274874, + 0.014476308599114418, + 0.007224198896437883, + 0.010007280856370926, + 0.006449913140386343, + 0.009702232666313648, + 0.010400163009762764, + 0.0059491610154509544, + 0.01600141078233719, + 0.007957966066896915, + 0.005414435174316168, + 0.01043478213250637, + 0.006966260261833668, + 0.014752935618162155, + 0.014557979069650173, + 0.008492504246532917, + 0.01255751308053732, + 0.00813184306025505, + 0.006882746238261461, + 0.003790923161432147, + 0.01127080898731947, + 0.0042537967674434185, + 0.010040217079222202, + 0.004500219598412514, + 0.0046306573785841465, + 0.005963681265711784, + 0.003983520437031984, + 0.0067245569080114365, + 0.007703847251832485, + 0.012311003170907497, + 0.010411608032882214, + 0.017468469217419624, + 0.009112415835261345, + 0.015084100887179375, + 0.010482714511454105, + 0.003722443012520671, + 0.00397487822920084, + 0.00876967329531908, + 0.00877115223556757, + 0.003682178445160389, + 0.00404119398444891, + 0.006898434367030859, + 0.00533234653994441, + 0.005608915351331234, + 0.00604631332680583, + 0.009645584970712662, + 0.009422019124031067, + 0.013075594790279865, + 0.01210511103272438, + 0.014441311359405518, + 0.011579661630094051, + 0.003984208218753338, + 0.0051654246635735035, + 0.006504433695226908, + 0.006015821360051632, + 0.0059659602120518684, + 0.007097467314451933, + 0.006306472234427929, + 0.01269725151360035, + 0.0035747240763157606, + 0.003357820212841034, + 0.005487017333507538, + 0.007485282141715288, + 0.002765417331829667, + 0.0021943857427686453, + 0.005583607126027346, + 0.003094129730015993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.09469000995159149 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813941404223442 + ], + "opposite_logprob_tokens": [ + -3.2881393432617188 + ], + "logprob_reward": -0.03813941404223442, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813941404223442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1279628723859787, + 0.2751900553703308, + 0.08026167005300522, + 0.07137922942638397, + 0.049792151898145676, + 0.05726858600974083, + 0.0374835729598999, + 0.09544088691473007, + 0.14722208678722382, + 0.057998836040496826 + ], + "credit_map": [ + 0.008204535581171513, + 0.008711776696145535, + 0.03396937623620033, + 0.015617858618497849, + 0.00967815425246954, + 0.009972449392080307, + 0.03350390866398811, + 0.008304816670715809, + 0.019822269678115845, + 0.014946362003684044, + 0.02035229653120041, + 0.05509806424379349, + 0.023235538974404335, + 0.018151894211769104, + 0.08673293143510818, + 0.0368506982922554, + 0.010507512837648392, + 0.006797501817345619, + 0.008132965303957462, + 0.011296790093183517, + 0.007557600736618042, + 0.0174616277217865, + 0.010463126935064793, + 0.00804454181343317, + 0.007473665289580822, + 0.005277886521071196, + 0.010361043736338615, + 0.0077532515861094, + 0.005589908454567194, + 0.007002602331340313, + 0.013937143608927727, + 0.013983727432787418, + 0.003486127592623234, + 0.006305851507931948, + 0.005758988671004772, + 0.009456055238842964, + 0.0063661327585577965, + 0.007378405891358852, + 0.006303433794528246, + 0.004737160634249449, + 0.003630724735558033, + 0.005832137539982796, + 0.006858187261968851, + 0.007588353473693132, + 0.010996419005095959, + 0.0066265142522752285, + 0.008485449478030205, + 0.007250805385410786, + 0.002861094195395708, + 0.0038415684830397367, + 0.006752543151378632, + 0.004454035311937332, + 0.003983207978308201, + 0.004325747489929199, + 0.00669943168759346, + 0.004565945826470852, + 0.005065558943897486, + 0.009165669791400433, + 0.020240118727087975, + 0.009942114353179932, + 0.011391031555831432, + 0.00829880591481924, + 0.008099977858364582, + 0.02323761023581028, + 0.0118149034678936, + 0.016261516138911247, + 0.01795312948524952, + 0.017690477892756462, + 0.01725439354777336, + 0.01642736606299877, + 0.01970003731548786, + 0.03012027032673359, + 0.004162009805440903, + 0.0040256427600979805, + 0.00937562808394432, + 0.013582047075033188, + 0.006029624026268721, + 0.00474966736510396, + 0.01070130243897438, + 0.005372915416955948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.03813941404223442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003632496576756239 + ], + "opposite_logprob_tokens": [ + -5.628632545471191 + ], + "logprob_reward": -0.003632496576756239, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003632496576756239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2506669759750366, + 0.07329802960157394, + 0.13446208834648132, + 0.1607690304517746, + 0.048413023352622986, + 0.06765276938676834, + 0.0241891797631979, + 0.09003083407878876, + 0.04917396977543831, + 0.1013440489768982 + ], + "credit_map": [ + 0.012213753536343575, + 0.011190825141966343, + 0.08051169663667679, + 0.02333376742899418, + 0.01611543633043766, + 0.009100127965211868, + 0.08137690275907516, + 0.016824498772621155, + 0.004726702813059092, + 0.0030979332514107227, + 0.0070279850624501705, + 0.010256977751851082, + 0.006834533531218767, + 0.005298064090311527, + 0.02367354743182659, + 0.01238228753209114, + 0.013251759111881256, + 0.019225824624300003, + 0.017978686839342117, + 0.013696318492293358, + 0.016538936644792557, + 0.025248754769563675, + 0.013764427974820137, + 0.014757399447262287, + 0.019038191065192223, + 0.014174940064549446, + 0.03555012121796608, + 0.037788208574056625, + 0.010399308055639267, + 0.012628619559109211, + 0.01850697211921215, + 0.012682677246630192, + 0.003972557373344898, + 0.008233503438532352, + 0.004820018075406551, + 0.00916622020304203, + 0.005550024099647999, + 0.006629774812608957, + 0.005941138137131929, + 0.00409978860989213, + 0.0048578521236777306, + 0.006224891170859337, + 0.011510674841701984, + 0.011628464795649052, + 0.009013395756483078, + 0.006374660413712263, + 0.010024653747677803, + 0.008018179796636105, + 0.0018560531316325068, + 0.0022924093063920736, + 0.004633401520550251, + 0.0035503297112882137, + 0.0024039135314524174, + 0.0025790121871978045, + 0.003413865575566888, + 0.0034601972438395023, + 0.006907240021973848, + 0.009071732871234417, + 0.012977154925465584, + 0.011336073279380798, + 0.010373078286647797, + 0.011629878543317318, + 0.018982307985424995, + 0.008753377944231033, + 0.005830733105540276, + 0.0034233159385621548, + 0.006220007315278053, + 0.006424823310226202, + 0.004467995371669531, + 0.0042728399857878685, + 0.007438158616423607, + 0.011096099391579628, + 0.007064096629619598, + 0.007443042006343603, + 0.019085004925727844, + 0.02970222197473049, + 0.00822925753891468, + 0.005447318311780691, + 0.015674931928515434, + 0.00869818963110447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.003632496576756239 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016080450266599655 + ], + "opposite_logprob_tokens": [ + -4.141080379486084 + ], + "logprob_reward": -0.016080450266599655, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016080450266599655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2741726040840149, + 0.07695445418357849, + 0.08792148530483246, + 0.06865079700946808, + 0.06898969411849976, + 0.047346871346235275, + 0.029988151043653488, + 0.12878255546092987, + 0.058707140386104584, + 0.1584862321615219 + ], + "credit_map": [ + 0.015167265199124813, + 0.01512192189693451, + 0.07720595598220825, + 0.02416439913213253, + 0.02466863952577114, + 0.023647140711545944, + 0.0752612054347992, + 0.01893606223165989, + 0.005271941889077425, + 0.006386711727827787, + 0.008035724982619286, + 0.011999613605439663, + 0.0108129121363163, + 0.009042439050972462, + 0.01807992532849312, + 0.00732518220320344, + 0.011826453730463982, + 0.008764581754803658, + 0.012887017801404, + 0.009106653742492199, + 0.009004672057926655, + 0.02116926573216915, + 0.007417826913297176, + 0.007745011709630489, + 0.009795078076422215, + 0.007241087965667248, + 0.009941683150827885, + 0.009794934652745724, + 0.008818869479000568, + 0.009187018498778343, + 0.007976802065968513, + 0.0058953240513801575, + 0.005826030857861042, + 0.014911897480487823, + 0.0073853894136846066, + 0.0163123682141304, + 0.005580617114901543, + 0.005036810413002968, + 0.008450510911643505, + 0.005486073438078165, + 0.004654691554605961, + 0.00403002230450511, + 0.005466945935040712, + 0.007321623619645834, + 0.005613515619188547, + 0.005651690997183323, + 0.009162891656160355, + 0.005445489659905434, + 0.0026785731315612793, + 0.0026517827063798904, + 0.00539606437087059, + 0.004254013765603304, + 0.002511556027457118, + 0.003491949522867799, + 0.0051950812339782715, + 0.0038091286551207304, + 0.00863912794739008, + 0.0114905945956707, + 0.02579287812113762, + 0.01498204655945301, + 0.0154497679322958, + 0.012229696847498417, + 0.022561460733413696, + 0.0176369771361351, + 0.007836000062525272, + 0.004357451573014259, + 0.007665582001209259, + 0.009107869118452072, + 0.005489391274750233, + 0.005371440667659044, + 0.009377130307257175, + 0.009502274915575981, + 0.014374847523868084, + 0.013204499147832394, + 0.02743382193148136, + 0.03845641389489174, + 0.014154348522424698, + 0.009007461369037628, + 0.025547504425048828, + 0.016307344660162926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.016080450266599655 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808938339352608 + ], + "opposite_logprob_tokens": [ + -3.2880892753601074 + ], + "logprob_reward": -0.03808938339352608, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808938339352608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20237307250499725, + 0.07946927845478058, + 0.27723121643066406, + 0.062213439494371414, + 0.14471986889839172, + 0.044710755348205566, + 0.03219114616513252, + 0.044385094195604324, + 0.041486706584692, + 0.07121945917606354 + ], + "credit_map": [ + 0.010345199145376682, + 0.012737545184791088, + 0.05463631823658943, + 0.02335508167743683, + 0.020130159333348274, + 0.015535013750195503, + 0.052621468901634216, + 0.013012275099754333, + 0.004421474877744913, + 0.005493158474564552, + 0.010771222412586212, + 0.023506952449679375, + 0.010053448379039764, + 0.007477729581296444, + 0.009807929396629333, + 0.007937360554933548, + 0.03850417584180832, + 0.02315419912338257, + 0.03997838497161865, + 0.030974838882684708, + 0.02369055710732937, + 0.07499237358570099, + 0.02427443116903305, + 0.021662233397364616, + 0.0063798632472753525, + 0.00519216014072299, + 0.010267202742397785, + 0.009187312796711922, + 0.005840538535267115, + 0.0068060168996453285, + 0.010189513675868511, + 0.00835083145648241, + 0.010963834822177887, + 0.029280249029397964, + 0.014806093648076057, + 0.03265112265944481, + 0.011427112855017185, + 0.01058757770806551, + 0.02294490486383438, + 0.012058970518410206, + 0.0031623877584934235, + 0.00374107388779521, + 0.005004545673727989, + 0.006313446909189224, + 0.00610006507486105, + 0.004511270672082901, + 0.01040131226181984, + 0.005476654041558504, + 0.0029254239052534103, + 0.0037432971876114607, + 0.00565829873085022, + 0.004917861893773079, + 0.002447181148454547, + 0.0034718497190624475, + 0.004509855527430773, + 0.004517378285527229, + 0.0025356379337608814, + 0.0034626934211701155, + 0.005752074997872114, + 0.003975935745984316, + 0.007414426654577255, + 0.007091704290360212, + 0.005987166427075863, + 0.008165456354618073, + 0.0035841616336256266, + 0.0036769607104361057, + 0.004643041640520096, + 0.0043669152073562145, + 0.0050923521630465984, + 0.005041858647018671, + 0.006555664818733931, + 0.008525751531124115, + 0.005390837322920561, + 0.005110730417072773, + 0.01380881667137146, + 0.018674485385417938, + 0.006202805787324905, + 0.0036499444395303726, + 0.011616850271821022, + 0.006764981895685196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.03808938339352608 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, macro, dolly shot moves horizontally across the surface of a gleaming, polished black granite kitchen counter, beginning a slow, sweeping journey that transforms the scene into a vast miniature landscape. Warm, golden morning sunlight streams in through an unseen window, casting long, dramatic shadows and illuminating motes of dust that drift like tiny stars. The camera glides over soft drifts and dunes of fine white flour, its powdery texture catching the light and creating the illusion of a snow-capped terrain with delicate peaks and valleys. Stainless steel utensils rise from this floury plain; the tines of forks stand upright like a row of metallic monoliths, while the deep, curved bowls of spoons form a shadowy, reflective basin. The camera continues its steady and deliberate path, revealing a wire whisk lying on its side, its intricate structure resembling a strange, skeletal bush. The entire scene is imbued with a tranquil and magical atmosphere, the warm light interacting on the cool, reflective surfaces of the counter and steel, creating a quiet, awe-inspiring world born from the beautiful chaos of baking.", + "seed": 8957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_1141_seed_8957_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.036318654823116955, + "vlm_logprob": -0.036318654823116955, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.036318654823116955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0716283842921257 + ], + "opposite_logprob_tokens": [ + -3.0716283321380615 + ], + "logprob_reward": -0.0716283842921257, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0716283842921257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16948550939559937, + 0.06319625675678253, + 0.07080896943807602, + 0.08032316714525223, + 0.048400793224573135, + 0.14610084891319275, + 0.10786906629800797, + 0.12069664150476456, + 0.034569282084703445, + 0.158549502491951 + ], + "credit_map": [ + 0.015251778066158295, + 0.018499787896871567, + 0.017814481630921364, + 0.024957260116934776, + 0.03213013336062431, + 0.01735551655292511, + 0.022140420973300934, + 0.021336114034056664, + 0.006491545587778091, + 0.007499389350414276, + 0.009447488002479076, + 0.0087894257158041, + 0.009068329818546772, + 0.006797628942877054, + 0.007685805670917034, + 0.007416644133627415, + 0.004985596984624863, + 0.006345424801111221, + 0.008701050654053688, + 0.011418039910495281, + 0.006719622761011124, + 0.006585811264812946, + 0.014169291593134403, + 0.011884121224284172, + 0.007812978699803352, + 0.006873585283756256, + 0.009600027464330196, + 0.014978655613958836, + 0.008562812581658363, + 0.008183250203728676, + 0.016259195283055305, + 0.008052661083638668, + 0.003710373304784298, + 0.005142956972122192, + 0.007497078273445368, + 0.009918464347720146, + 0.005569331347942352, + 0.006200676318258047, + 0.0059580327942967415, + 0.004403877072036266, + 0.015336587093770504, + 0.01661253161728382, + 0.01805480383336544, + 0.03012698143720627, + 0.015659868717193604, + 0.017858007922768593, + 0.016027752310037613, + 0.016424305737018585, + 0.00811406783759594, + 0.008667968213558197, + 0.04106630012392998, + 0.006694808602333069, + 0.006887347903102636, + 0.01646411046385765, + 0.007372196763753891, + 0.012602260336279869, + 0.007948089390993118, + 0.011147618293762207, + 0.009296744130551815, + 0.00820854026824236, + 0.012234576977789402, + 0.03326736390590668, + 0.00992958527058363, + 0.028664125129580498, + 0.0037336302921175957, + 0.0059724184684455395, + 0.004250429570674896, + 0.003975775092840195, + 0.0038030357100069523, + 0.003487802343443036, + 0.004269718192517757, + 0.005076470784842968, + 0.01960502751171589, + 0.020225046202540398, + 0.020508917048573494, + 0.021755095571279526, + 0.018239334225654602, + 0.018680935725569725, + 0.018009958788752556, + 0.02152516506612301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0716283842921257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869633540511131 + ], + "opposite_logprob_tokens": [ + -3.048696279525757 + ], + "logprob_reward": -0.04869633540511131, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869633540511131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09274802356958389, + 0.02760675735771656, + 0.028114046901464462, + 0.021715674549341202, + 0.06462586671113968, + 0.0754886344075203, + 0.31268054246902466, + 0.28310704231262207, + 0.027233485132455826, + 0.0666799247264862 + ], + "credit_map": [ + 0.01543615572154522, + 0.014335989952087402, + 0.011661909520626068, + 0.015932196751236916, + 0.007676666136831045, + 0.007051597815006971, + 0.009999468922615051, + 0.010654044337570667, + 0.004700435325503349, + 0.0029800948686897755, + 0.0037581317592412233, + 0.0038239695131778717, + 0.003404232207685709, + 0.0033805908169597387, + 0.0028042371850460768, + 0.0027550666127353907, + 0.002164621837437153, + 0.0023583481088280678, + 0.0030585890635848045, + 0.0038174374494701624, + 0.004904228262603283, + 0.0025356856640428305, + 0.0055146231316030025, + 0.003760511288419366, + 0.0018444382585585117, + 0.0018266815459355712, + 0.0026317229494452477, + 0.004101062659174204, + 0.0022954228334128857, + 0.0023884489201009274, + 0.003787539666518569, + 0.0028403587639331818, + 0.0035984127316623926, + 0.00763574056327343, + 0.010945922695100307, + 0.015081646852195263, + 0.006702189799398184, + 0.010051451623439789, + 0.0061025070026516914, + 0.004507989156991243, + 0.007607671897858381, + 0.0066872877068817616, + 0.01105957105755806, + 0.01634031906723976, + 0.008391850627958775, + 0.006641632877290249, + 0.008736460469663143, + 0.010023836977779865, + 0.01841227151453495, + 0.01318769808858633, + 0.13157087564468384, + 0.019071074202656746, + 0.014643180184066296, + 0.0644247978925705, + 0.018657291308045387, + 0.032713379710912704, + 0.008534148335456848, + 0.007120415568351746, + 0.007921290583908558, + 0.008409913629293442, + 0.008561240509152412, + 0.11730876564979553, + 0.014092930592596531, + 0.11115831136703491, + 0.0028875290881842375, + 0.0035303933545947075, + 0.0035243125166743994, + 0.0036053780931979418, + 0.0032274031545966864, + 0.004400898702442646, + 0.0026026456616818905, + 0.003454923629760742, + 0.0064315274357795715, + 0.007697424851357937, + 0.00844588316977024, + 0.007115380838513374, + 0.007765675894916058, + 0.01202698890119791, + 0.007536099757999182, + 0.00966094620525837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.04869633540511131 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015337619697675109 + ], + "opposite_logprob_tokens": [ + -6.5015339851379395 + ], + "logprob_reward": -0.0015337619697675109, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015337619697675109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22944849729537964, + 0.0694555714726448, + 0.11961925029754639, + 0.24758507311344147, + 0.0682675689458847, + 0.057061851024627686, + 0.04516657441854477, + 0.08241891115903854, + 0.01883753389120102, + 0.06213914602994919 + ], + "credit_map": [ + 0.019670093432068825, + 0.01642017625272274, + 0.02880818583071232, + 0.02645972929894924, + 0.03317611664533615, + 0.028159214183688164, + 0.03241754323244095, + 0.044337447732686996, + 0.009375069290399551, + 0.0072867185808718204, + 0.009686696343123913, + 0.008402597159147263, + 0.010218034498393536, + 0.00869140774011612, + 0.007438150700181723, + 0.008356896229088306, + 0.008444266393780708, + 0.008325863629579544, + 0.012874959036707878, + 0.019121335819363594, + 0.01255269255489111, + 0.010931018739938736, + 0.028213664889335632, + 0.01915544643998146, + 0.020286981016397476, + 0.022116078063845634, + 0.03195401281118393, + 0.05617084354162216, + 0.027593594044446945, + 0.02388475276529789, + 0.04448750615119934, + 0.02109130285680294, + 0.00460802111774683, + 0.006818989757448435, + 0.012641850858926773, + 0.015737144276499748, + 0.006830643862485886, + 0.009024257771670818, + 0.0070638819597661495, + 0.005542776081711054, + 0.004347356967628002, + 0.00401734234765172, + 0.008703472092747688, + 0.008718465454876423, + 0.007312505505979061, + 0.008854147046804428, + 0.008362793363630772, + 0.006745771039277315, + 0.003411952406167984, + 0.003550157183781266, + 0.015398716554045677, + 0.0037488993257284164, + 0.0029855917673557997, + 0.007469115778803825, + 0.003926541656255722, + 0.004675601609051228, + 0.00417738314718008, + 0.004952009301632643, + 0.004680947866290808, + 0.00388615601696074, + 0.006161986384540796, + 0.026831530034542084, + 0.005945214070379734, + 0.025783682242035866, + 0.0020800521597266197, + 0.0019292295910418034, + 0.00213919160887599, + 0.002063541440293193, + 0.002896838588640094, + 0.00249923556111753, + 0.0027907858602702618, + 0.002438659779727459, + 0.007052984554320574, + 0.0076166363433003426, + 0.009391503408551216, + 0.008840110152959824, + 0.00575580820441246, + 0.008238556794822216, + 0.007865694351494312, + 0.00737784756347537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0015337619697675109 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088416911661625 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088416911661625, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088416911661625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0709720328450203, + 0.032387688755989075, + 0.07369078695774078, + 0.03575805574655533, + 0.03248317912220955, + 0.06997226923704147, + 0.24481211602687836, + 0.33401623368263245, + 0.03659224137663841, + 0.06931537389755249 + ], + "credit_map": [ + 0.005891477223485708, + 0.008403974585235119, + 0.008381552994251251, + 0.008954791352152824, + 0.008324914611876011, + 0.008339868858456612, + 0.009917216375470161, + 0.012758240103721619, + 0.0030115321278572083, + 0.004063555039465427, + 0.00550233805552125, + 0.005343951750546694, + 0.0038126581348478794, + 0.0032121092081069946, + 0.003736963029950857, + 0.003704580944031477, + 0.005735114216804504, + 0.005690535996109247, + 0.008268974721431732, + 0.011609094217419624, + 0.008004006929695606, + 0.007340429816395044, + 0.01565416343510151, + 0.011388476006686687, + 0.002980445744469762, + 0.0022479789331555367, + 0.0051706996746361256, + 0.007636789698153734, + 0.004078584257513285, + 0.003911852836608887, + 0.005246850661933422, + 0.004484857432544231, + 0.0021358253434300423, + 0.0032206682953983545, + 0.007001988124102354, + 0.009434426203370094, + 0.0019418535521253943, + 0.002558351494371891, + 0.0036739197093993425, + 0.0025161465164273977, + 0.005197254475206137, + 0.006486834958195686, + 0.009652596898376942, + 0.012745477259159088, + 0.007964514195919037, + 0.008926519192755222, + 0.01054066140204668, + 0.00845841784030199, + 0.0152882169932127, + 0.009090865962207317, + 0.10888887941837311, + 0.014505691826343536, + 0.0074528916738927364, + 0.05785736069083214, + 0.010347471572458744, + 0.021380746737122536, + 0.011282490566372871, + 0.009076321497559547, + 0.0100376782938838, + 0.008960869163274765, + 0.012424841523170471, + 0.1382717788219452, + 0.01640688255429268, + 0.12755538523197174, + 0.0033051669597625732, + 0.004726979415863752, + 0.0035521299578249454, + 0.004687934182584286, + 0.004148546140640974, + 0.005590748507529497, + 0.00558231258764863, + 0.004998428281396627, + 0.007668547332286835, + 0.009169640950858593, + 0.011306306347250938, + 0.012166987173259258, + 0.0073493728414177895, + 0.006477085407823324, + 0.007005726452916861, + 0.008171714842319489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.011088416911661625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864637553691864 + ], + "opposite_logprob_tokens": [ + -3.0486464500427246 + ], + "logprob_reward": -0.04864637553691864, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864637553691864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0889144092798233, + 0.029046395793557167, + 0.045121923089027405, + 0.04861662909388542, + 0.033278267830610275, + 0.0335504375398159, + 0.16664496064186096, + 0.4800240099430084, + 0.03483092784881592, + 0.039972081780433655 + ], + "credit_map": [ + 0.015181398950517178, + 0.008146974258124828, + 0.01419378723949194, + 0.013697783462703228, + 0.00979425385594368, + 0.006881814915686846, + 0.009024355560541153, + 0.011994034051895142, + 0.0031461352482438087, + 0.0025896381121128798, + 0.004615087527781725, + 0.004847405478358269, + 0.003772798925638199, + 0.0029782885685563087, + 0.0034491484984755516, + 0.0036478929687291384, + 0.0029065541457384825, + 0.0031606671400368214, + 0.005155783146619797, + 0.00794302299618721, + 0.004830229561775923, + 0.005747330840677023, + 0.009451979771256447, + 0.00592635665088892, + 0.004616987891495228, + 0.004664243198931217, + 0.005213879514485598, + 0.008822186850011349, + 0.005706509575247765, + 0.004379622638225555, + 0.010422226041555405, + 0.004790972918272018, + 0.003061159746721387, + 0.003627372905611992, + 0.003917780704796314, + 0.005100858397781849, + 0.0035166791640222073, + 0.004446093458682299, + 0.005553630646318197, + 0.004054694902151823, + 0.0026518800295889378, + 0.002770808059722185, + 0.004744155332446098, + 0.004602702334523201, + 0.004213322885334492, + 0.005374378990381956, + 0.004082393366843462, + 0.005110795609652996, + 0.011194022372364998, + 0.007825879380106926, + 0.07449092715978622, + 0.01072263065725565, + 0.005051861982792616, + 0.02526799775660038, + 0.0069413864985108376, + 0.025150258094072342, + 0.014949144795536995, + 0.010719838552176952, + 0.011720628477633, + 0.011767650954425335, + 0.01363764051347971, + 0.20343931019306183, + 0.023389678448438644, + 0.19040007889270782, + 0.003356494242325425, + 0.004805409349501133, + 0.0036741828080266714, + 0.0033088973723351955, + 0.005326368845999241, + 0.004920361563563347, + 0.00483988830819726, + 0.004599326755851507, + 0.004345943219959736, + 0.004484917037189007, + 0.004562453366816044, + 0.006798668298870325, + 0.004650003742426634, + 0.005782107822597027, + 0.004171408712863922, + 0.005176577717065811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.04864637553691864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, macro, dolly shot moves horizontally across the surface of a gleaming, polished black granite kitchen counter, beginning a slow, sweeping journey that transforms the scene into a vast miniature landscape. Warm, golden morning sunlight streams in through an unseen window, casting long, dramatic shadows and illuminating motes of dust that drift like tiny stars. The camera glides over soft drifts and dunes of fine white flour, its powdery texture catching the light and creating the illusion of a snow-capped terrain with delicate peaks and valleys. Stainless steel utensils rise from this floury plain; the tines of forks stand upright like a row of metallic monoliths, while the deep, curved bowls of spoons form a shadowy, reflective basin. The camera continues its steady and deliberate path, revealing a wire whisk lying on its side, its intricate structure resembling a strange, skeletal bush. The entire scene is imbued with a tranquil and magical atmosphere, the warm light interacting on the cool, reflective surfaces of the counter and steel, creating a quiet, awe-inspiring world born from the beautiful chaos of baking.", + "seed": 8958, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_2/index_1141_seed_8958_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04767229398712516, + "vlm_logprob": -0.04767229398712516, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04767229398712516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11248045414686203 + ], + "opposite_logprob_tokens": [ + -2.362480401992798 + ], + "logprob_reward": -0.11248045414686203, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11248045414686203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1833513379096985, + 0.14676138758659363, + 0.04184059798717499, + 0.1432478278875351, + 0.04196900129318237, + 0.12933848798274994, + 0.0665828213095665, + 0.037192992866039276, + 0.12486197054386139, + 0.0848536491394043 + ], + "credit_map": [ + 0.017489779740571976, + 0.02370028756558895, + 0.02113039419054985, + 0.027488112449645996, + 0.02866126224398613, + 0.017921051010489464, + 0.024163782596588135, + 0.02279665879905224, + 0.018446985632181168, + 0.018680082634091377, + 0.019357340410351753, + 0.019889339804649353, + 0.020841382443904877, + 0.01612900383770466, + 0.017177194356918335, + 0.016240056604146957, + 0.006717480253428221, + 0.0064383079297840595, + 0.005347057711333036, + 0.005415101069957018, + 0.004648696631193161, + 0.003808664157986641, + 0.005701125133782625, + 0.003764165099710226, + 0.01810154877603054, + 0.020368289202451706, + 0.012219349853694439, + 0.01510189101099968, + 0.030307523906230927, + 0.013321662321686745, + 0.02232605218887329, + 0.011501510627567768, + 0.005713690537959337, + 0.00545594934374094, + 0.005006034858524799, + 0.004078110679984093, + 0.006189488805830479, + 0.0037332952488213778, + 0.006183085963129997, + 0.005609346553683281, + 0.017008012160658836, + 0.02269560657441616, + 0.01134630385786295, + 0.01391638070344925, + 0.012619001790881157, + 0.019945785403251648, + 0.018533974885940552, + 0.013273423537611961, + 0.006754547357559204, + 0.009755325503647327, + 0.007064000703394413, + 0.01698494888842106, + 0.007909697480499744, + 0.005719935987144709, + 0.006968561559915543, + 0.005425799638032913, + 0.004963630810379982, + 0.004284289665520191, + 0.003996769431978464, + 0.005019219592213631, + 0.0067521738819777966, + 0.003729884047061205, + 0.004097837954759598, + 0.004349185153841972, + 0.012733086943626404, + 0.017417050898075104, + 0.015216012485325336, + 0.012288815341889858, + 0.01677759736776352, + 0.011361193843185902, + 0.019146667793393135, + 0.019921544939279556, + 0.009104076772928238, + 0.010543187148869038, + 0.013503015972673893, + 0.012077814899384975, + 0.010914833284914494, + 0.009520800784230232, + 0.008881503716111183, + 0.010308416560292244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.11248045414686203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209041178226471 + ], + "opposite_logprob_tokens": [ + -2.8120903968811035 + ], + "logprob_reward": -0.06209041178226471, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209041178226471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16435833275318146, + 0.12089303135871887, + 0.06806982308626175, + 0.06370432674884796, + 0.0902002826333046, + 0.040062639862298965, + 0.12088717520236969, + 0.18289054930210114, + 0.06648854911327362, + 0.08244531601667404 + ], + "credit_map": [ + 0.01845385506749153, + 0.020020117983222008, + 0.01758688874542713, + 0.026035139337182045, + 0.018242429941892624, + 0.01534255687147379, + 0.02553732320666313, + 0.02314000390470028, + 0.01664755865931511, + 0.012316920794546604, + 0.014796363189816475, + 0.01469770073890686, + 0.01837831735610962, + 0.011931690387427807, + 0.019040299579501152, + 0.013084175996482372, + 0.009070051833987236, + 0.011618327349424362, + 0.00778945442289114, + 0.009579304605722427, + 0.006537281442433596, + 0.008256851695477962, + 0.008981493301689625, + 0.006237058900296688, + 0.008485965430736542, + 0.00619539013132453, + 0.00762247433885932, + 0.0102424630895257, + 0.007864105515182018, + 0.00700016226619482, + 0.009935135953128338, + 0.006358620245009661, + 0.010359089821577072, + 0.010549544356763363, + 0.012018384411931038, + 0.008601502515375614, + 0.010700138285756111, + 0.008600094355642796, + 0.012139920145273209, + 0.01723160222172737, + 0.004882097244262695, + 0.00762596633285284, + 0.004244429059326649, + 0.004629501607269049, + 0.0035052027087658644, + 0.004168582148849964, + 0.00549025135114789, + 0.00551660405471921, + 0.018439600244164467, + 0.013504588045179844, + 0.011004523374140263, + 0.021033156663179398, + 0.01591956615447998, + 0.011098501272499561, + 0.012408191338181496, + 0.017479032278060913, + 0.025505991652607918, + 0.052541181445121765, + 0.03587425500154495, + 0.010194975882768631, + 0.015734223648905754, + 0.013370322063565254, + 0.015085180290043354, + 0.014584400691092014, + 0.008535834960639477, + 0.0069218287244439125, + 0.009396931156516075, + 0.007401027251034975, + 0.005722638685256243, + 0.004160052165389061, + 0.012577108107507229, + 0.011773125268518925, + 0.00986823532730341, + 0.00863717868924141, + 0.012429590336978436, + 0.013520097360014915, + 0.008939788676798344, + 0.010302134789526463, + 0.010716472752392292, + 0.008031816221773624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.06209041178226471 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004657966550439596 + ], + "opposite_logprob_tokens": [ + -5.379657745361328 + ], + "logprob_reward": -0.004657966550439596, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004657966550439596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21086809039115906, + 0.1655978560447693, + 0.05959330499172211, + 0.07282986491918564, + 0.11194576323032379, + 0.038408055901527405, + 0.03702293708920479, + 0.09657489508390427, + 0.1614619940519333, + 0.04569721221923828 + ], + "credit_map": [ + 0.020438743755221367, + 0.024670375511050224, + 0.02218124456703663, + 0.0363350510597229, + 0.02995278313755989, + 0.020771382376551628, + 0.03126043826341629, + 0.025258062407374382, + 0.02098223753273487, + 0.020453475415706635, + 0.0185147225856781, + 0.035037606954574585, + 0.015852807089686394, + 0.015497036278247833, + 0.01913466304540634, + 0.020125312730669975, + 0.008759817108511925, + 0.010886562056839466, + 0.007182944566011429, + 0.00993846170604229, + 0.00608926173299551, + 0.005102428607642651, + 0.007342760916799307, + 0.004291071090847254, + 0.009168105199933052, + 0.009426587261259556, + 0.006281792651861906, + 0.0076532489620149136, + 0.011526609770953655, + 0.008392899297177792, + 0.013529863208532333, + 0.006850763224065304, + 0.010859298519790173, + 0.012541820295155048, + 0.012109201401472092, + 0.008772337809205055, + 0.012999485246837139, + 0.010227515362203121, + 0.01810295507311821, + 0.026333153247833252, + 0.004158451687544584, + 0.006622411776334047, + 0.005223318934440613, + 0.005664792377501726, + 0.0037402946036309004, + 0.0037014468107372522, + 0.005088407546281815, + 0.004208930768072605, + 0.005222279112786055, + 0.005742479115724564, + 0.0037783284205943346, + 0.006190592423081398, + 0.004884412512183189, + 0.0035999526735395193, + 0.0031603851821273565, + 0.004444507881999016, + 0.014553318731486797, + 0.02562064118683338, + 0.009288502857089043, + 0.0072725554928183556, + 0.015660056844353676, + 0.008523009717464447, + 0.006818027701228857, + 0.008838770911097527, + 0.016839981079101562, + 0.020274857059121132, + 0.019351720809936523, + 0.017230944707989693, + 0.01873551681637764, + 0.01476147398352623, + 0.028066616505384445, + 0.026200875639915466, + 0.006348363123834133, + 0.005486632231622934, + 0.005507636349648237, + 0.007331177592277527, + 0.003481081919744611, + 0.005029718857258558, + 0.00602059718221426, + 0.0064920056611299515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.004657966550439596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079628840088844 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.016079628840088844, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079628840088844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3040691614151001, + 0.1460389643907547, + 0.03612007200717926, + 0.03466595336794853, + 0.11663152277469635, + 0.017704477533698082, + 0.03961224481463432, + 0.0768781304359436, + 0.17934145033359528, + 0.04893803596496582 + ], + "credit_map": [ + 0.02430867776274681, + 0.03813158720731735, + 0.06465353816747665, + 0.03192548453807831, + 0.04069792106747627, + 0.030947944149374962, + 0.040248118340969086, + 0.033155880868434906, + 0.01843542419373989, + 0.020979823544621468, + 0.018330832943320274, + 0.01699645072221756, + 0.02205052226781845, + 0.013362498953938484, + 0.019184952601790428, + 0.016698449850082397, + 0.004776001442223787, + 0.005282283294945955, + 0.005483337212353945, + 0.007688132580369711, + 0.0025792070664465427, + 0.002866020891815424, + 0.005159378983080387, + 0.0022857103031128645, + 0.0038215171080082655, + 0.004091982264071703, + 0.00456573162227869, + 0.004780222196131945, + 0.005545043386518955, + 0.0038837839383631945, + 0.004753238055855036, + 0.0032244331669062376, + 0.012974509038031101, + 0.013295329175889492, + 0.009593361988663673, + 0.009008324705064297, + 0.014818680472671986, + 0.011963829398155212, + 0.019151587039232254, + 0.025825902819633484, + 0.0026599315460771322, + 0.0038030866999179125, + 0.002003836678341031, + 0.0019262712448835373, + 0.0011646710336208344, + 0.0014656023122370243, + 0.0018203746294602752, + 0.002860704669728875, + 0.004401770886033773, + 0.005356186535209417, + 0.006336556281894445, + 0.004692397080361843, + 0.0056135631166398525, + 0.005298214498907328, + 0.004193910863250494, + 0.0037196448538452387, + 0.010232808999717236, + 0.019924717023968697, + 0.014162321574985981, + 0.005345942918211222, + 0.00938583817332983, + 0.006429541390389204, + 0.006518072448670864, + 0.004878883715718985, + 0.02092413604259491, + 0.025696896016597748, + 0.02219393290579319, + 0.02226187102496624, + 0.020987695083022118, + 0.015569068491458893, + 0.027198465541005135, + 0.024509374052286148, + 0.006666619796305895, + 0.0058556366711854935, + 0.007255543489009142, + 0.006814782042056322, + 0.0044281152077019215, + 0.005078798159956932, + 0.007735692895948887, + 0.005102850031107664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.016079628840088844 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle macro view of a kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a horizontal dolly shot moving across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surface covered in fine white flour resembling a miniature landscape with dunes and peaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are stainless steel utensils visible, including upright forks and spoons forming a reflective basin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wire whisk lying on its side visible, resembling a skeletal bush?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305300861597061 + ], + "opposite_logprob_tokens": [ + -3.168052911758423 + ], + "logprob_reward": -0.04305300861597061, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305300861597061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07338287681341171, + 0.04742126911878586, + 0.023750625550746918, + 0.0562872514128685, + 0.032807450741529465, + 0.015791457146406174, + 0.0192223209887743, + 0.07086239010095596, + 0.6068225502967834, + 0.05365182086825371 + ], + "credit_map": [ + 0.005644683726131916, + 0.007800514344125986, + 0.011316610500216484, + 0.011756042949855328, + 0.007857338525354862, + 0.008926530368626118, + 0.011026288382709026, + 0.00905486661940813, + 0.006604387890547514, + 0.005897159222513437, + 0.004658376332372427, + 0.005287935491651297, + 0.010048845782876015, + 0.0043300301767885685, + 0.004854479338973761, + 0.005740056280046701, + 0.003399957437068224, + 0.0025835982523858547, + 0.003646416589617729, + 0.0054354616440832615, + 0.0014602468581870198, + 0.0018357100198045373, + 0.003642807947471738, + 0.0017464287811890244, + 0.006270077079534531, + 0.005245131906121969, + 0.00844302587211132, + 0.008385851979255676, + 0.008058183826506138, + 0.005012371111661196, + 0.00836554728448391, + 0.00650706235319376, + 0.003777250647544861, + 0.00384643767029047, + 0.0038333849515765905, + 0.002451736945658922, + 0.004828970413655043, + 0.003445199690759182, + 0.005737454164773226, + 0.004887018818408251, + 0.002142705023288727, + 0.002775616245344281, + 0.0016263629077002406, + 0.0014623895986005664, + 0.0017774448497220874, + 0.0016401743050664663, + 0.0019471384584903717, + 0.0024196256417781115, + 0.0027486528269946575, + 0.003175639547407627, + 0.0021763669792562723, + 0.002519916743040085, + 0.0021271132864058018, + 0.001733120298013091, + 0.0017278972081840038, + 0.003013613633811474, + 0.00928590353578329, + 0.014195483177900314, + 0.006464865058660507, + 0.006329808384180069, + 0.013765303418040276, + 0.005426419898867607, + 0.006518300622701645, + 0.008876307867467403, + 0.0668976753950119, + 0.08391264826059341, + 0.072349913418293, + 0.0745212733745575, + 0.06678103655576706, + 0.0540161058306694, + 0.10476543009281158, + 0.08357841521501541, + 0.00541304424405098, + 0.007548769004642963, + 0.0076291244477033615, + 0.00754472566768527, + 0.008487476035952568, + 0.004066070541739464, + 0.007888569496572018, + 0.005074039567261934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.04305300861597061 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_tensor([3204, 3204])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_tensor([3204, 3204])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f03e08ddfba8e1c2cc0ac8e9b4f4bde861b97024 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_tensor([3204, 3204])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a scientifically inspired style, the camera executes a steady forward tracking shot, gliding through the deep, inky black of the abyss. The powerful, cool-white floodlight from the camera's light source carves a crisp circle on the seafloor, contrasting starkly against the surrounding impenetrable darkness. High frame rate capture renders the entire scene in ultra-smooth slow motion, highlighting the environment's intricate details with clinical precision. The camera first cruises over a vast abyssal plain, where the light reveals every tiny grain of pale sediment, swirling in soft, lingering clouds as the camera passes. The terrain then transitions to the jagged, oppressive edge of a massive volcanic ridge, its black slate formations casting long, sharp shadows. The water is exceptionally clear, exceptionally crystal clear, and the camera's path leads it to the cliffs of the deep, shadowy trench, offering an unobstructed view; the floodlight illuminates the sharp edge of the abyss before being completely swallowed up, plummeting into unseen depths, creating an atmosphere of breathtaking discovery and profound isolation.", + "seed": 19503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_3204_seed_19503_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19234438426792622, + "vlm_logprob": -0.19234438426792622, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.19234438426792622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2841046154499054 + ], + "opposite_logprob_tokens": [ + -1.409104585647583 + ], + "logprob_reward": -0.2841046154499054, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2841046154499054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1989518105983734, + 0.06958993524312973, + 0.03501720726490021, + 0.22919714450836182, + 0.05010978877544403, + 0.06761685758829117, + 0.03515186905860901, + 0.037906307727098465, + 0.18360303342342377, + 0.09285610169172287 + ], + "credit_map": [ + 0.016263628378510475, + 0.01727350614964962, + 0.04109634459018707, + 0.025326233357191086, + 0.01805264502763748, + 0.030128953978419304, + 0.0232005063444376, + 0.02760997787117958, + 0.007253196556121111, + 0.00478022638708353, + 0.010339291766285896, + 0.012357154861092567, + 0.007032608147710562, + 0.010237820446491241, + 0.008768551051616669, + 0.008821090683341026, + 0.0027094227261841297, + 0.002935412572696805, + 0.007764134090393782, + 0.004267831798642874, + 0.002754813525825739, + 0.004625467583537102, + 0.002743779681622982, + 0.007216344121843576, + 0.016021132469177246, + 0.004879275802522898, + 0.09709758311510086, + 0.009527328424155712, + 0.0047227381728589535, + 0.009924289770424366, + 0.006463528145104647, + 0.08056126534938812, + 0.006097796373069286, + 0.0020475618075579405, + 0.020554767921566963, + 0.0029679113067686558, + 0.0017075754003599286, + 0.0029183365404605865, + 0.0022017417941242456, + 0.011614100076258183, + 0.007052722852677107, + 0.005496894009411335, + 0.014598018489778042, + 0.004978979006409645, + 0.004450528416782618, + 0.01232262421399355, + 0.009418115951120853, + 0.009298973716795444, + 0.005705317948013544, + 0.002263209782540798, + 0.009460492990911007, + 0.0032432882580906153, + 0.002697674324735999, + 0.003628263482823968, + 0.004118611570447683, + 0.004035008605569601, + 0.004056906793266535, + 0.0044860453344881535, + 0.0030657711904495955, + 0.004068156238645315, + 0.0054567959159612656, + 0.006002089940011501, + 0.006265292875468731, + 0.004505247808992863, + 0.013915437273681164, + 0.030276698991656303, + 0.012106003239750862, + 0.04970056936144829, + 0.024359742179512978, + 0.018556635826826096, + 0.019842518493533134, + 0.014845418743789196, + 0.012750834226608276, + 0.011660625226795673, + 0.016255805268883705, + 0.008363439701497555, + 0.008482147008180618, + 0.018531665205955505, + 0.005427556578069925, + 0.011384026147425175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.2841046154499054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048682596534490585 + ], + "opposite_logprob_tokens": [ + -3.048682689666748 + ], + "logprob_reward": -0.048682596534490585, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048682596534490585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060964375734329224, + 0.026331549510359764, + 0.03697787597775459, + 0.7417211532592773, + 0.040517546236515045, + 0.016429468989372253, + 0.03782700374722481, + 0.009355694055557251, + 0.015110474079847336, + 0.01476479321718216 + ], + "credit_map": [ + 0.004859632812440395, + 0.006510959472507238, + 0.007945952005684376, + 0.008477335795760155, + 0.007375261280685663, + 0.013914576731622219, + 0.006397927645593882, + 0.005482735577970743, + 0.002796752844005823, + 0.0021003838628530502, + 0.0045190718956291676, + 0.0030212963465601206, + 0.0019267108291387558, + 0.0025246439035981894, + 0.004029175266623497, + 0.005413517821580172, + 0.0022675618529319763, + 0.001192826428450644, + 0.012685799039900303, + 0.002706731902435422, + 0.0014110777992755175, + 0.0028102893847972155, + 0.002059910912066698, + 0.011843683198094368, + 0.04785699397325516, + 0.014738796278834343, + 0.3160436153411865, + 0.030684208497405052, + 0.013404796831309795, + 0.03088846616446972, + 0.01941837929189205, + 0.2686859965324402, + 0.004370423965156078, + 0.0018645876552909613, + 0.01493879035115242, + 0.002562448848038912, + 0.001541133620776236, + 0.0034239133819937706, + 0.0034666676074266434, + 0.008349587209522724, + 0.0019300642888993025, + 0.0011924270074814558, + 0.004498606082051992, + 0.0012205808889120817, + 0.0010026530362665653, + 0.001762626226991415, + 0.00232672574929893, + 0.002495787339285016, + 0.00602631876245141, + 0.002078182762488723, + 0.009677301160991192, + 0.002047925954684615, + 0.006083478685468435, + 0.004222658928483725, + 0.00419813534244895, + 0.0034930065739899874, + 0.0006389080081135035, + 0.0008367564878426492, + 0.0009898486314341426, + 0.0013139285147190094, + 0.0016213246854022145, + 0.00133613555226475, + 0.001520106801763177, + 0.0010986870620399714, + 0.0014895574422553182, + 0.002158508636057377, + 0.0011405879631638527, + 0.004306437447667122, + 0.0014972611097618937, + 0.0013326043263077736, + 0.0019359895959496498, + 0.0012495310511440039, + 0.0014219541335478425, + 0.0015851703938096762, + 0.0028975990135222673, + 0.002277412684634328, + 0.0017449185252189636, + 0.0018569062231108546, + 0.0011963521828874946, + 0.001784482621587813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.048682596534490585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355136185884476 + ], + "opposite_logprob_tokens": [ + -3.651355028152466 + ], + "logprob_reward": -0.026355136185884476, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355136185884476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18179288506507874, + 0.04740282893180847, + 0.04875145107507706, + 0.33881115913391113, + 0.10462601482868195, + 0.05758003145456314, + 0.057077374309301376, + 0.030199337750673294, + 0.06897463649511337, + 0.06478425115346909 + ], + "credit_map": [ + 0.01150466687977314, + 0.0077861761674284935, + 0.06374779343605042, + 0.021599454805254936, + 0.010968003422021866, + 0.007881496101617813, + 0.016584357246756554, + 0.041720956563949585, + 0.0034557643812149763, + 0.002930059563368559, + 0.007880707271397114, + 0.003853158326819539, + 0.004281819798052311, + 0.008374757133424282, + 0.006596103310585022, + 0.010030456818640232, + 0.003164323978126049, + 0.0013362147146835923, + 0.016848517581820488, + 0.0028466160874813795, + 0.0025864855851978064, + 0.004847025033086538, + 0.002247973345220089, + 0.014874297194182873, + 0.021342238411307335, + 0.0065747895278036594, + 0.1432797759771347, + 0.013903670012950897, + 0.006143713835626841, + 0.014828926883637905, + 0.010186417028307915, + 0.12255165725946426, + 0.012587809935212135, + 0.004063159227371216, + 0.04267977178096771, + 0.005880179814994335, + 0.0036767309065908194, + 0.0055448394268751144, + 0.005033908411860466, + 0.025159623473882675, + 0.005746073555201292, + 0.00289855501614511, + 0.0076949940994381905, + 0.0038008522242307663, + 0.003690634621307254, + 0.008253215812146664, + 0.0125904381275177, + 0.012905267998576164, + 0.00890757143497467, + 0.0034076517913490534, + 0.01374558825045824, + 0.004242691211402416, + 0.004864424932748079, + 0.006338867358863354, + 0.00852800440043211, + 0.007042581681162119, + 0.0027400346007198095, + 0.0032570846378803253, + 0.0022459905594587326, + 0.0024776076897978783, + 0.00591890187934041, + 0.004394341725856066, + 0.006023628171533346, + 0.0031417508143931627, + 0.005602526478469372, + 0.009915712289512157, + 0.005150764714926481, + 0.01565857231616974, + 0.009371847845613956, + 0.007704493124037981, + 0.008189496584236622, + 0.007381230127066374, + 0.006794284097850323, + 0.006306196562945843, + 0.006537826266139746, + 0.0069603282026946545, + 0.007061815820634365, + 0.012741854414343834, + 0.006925438530743122, + 0.01145651564002037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.026355136185884476 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762330293655396 + ], + "opposite_logprob_tokens": [ + -0.8262330293655396 + ], + "logprob_reward": -0.5762330293655396, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762330293655396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13439472019672394, + 0.022776270285248756, + 0.013257142156362534, + 0.6248925924301147, + 0.04824759438633919, + 0.023097073659300804, + 0.02395430952310562, + 0.021893901750445366, + 0.03344067186117172, + 0.054045677185058594 + ], + "credit_map": [ + 0.009642564691603184, + 0.009898500517010689, + 0.032253723591566086, + 0.013480344787240028, + 0.011467894539237022, + 0.02080773375928402, + 0.01541199628263712, + 0.02143196575343609, + 0.0028516482561826706, + 0.0016017466550692916, + 0.003706009592860937, + 0.0022869098465889692, + 0.0017667358042672276, + 0.003301876364275813, + 0.0021082297898828983, + 0.005153115373104811, + 0.0009911871748045087, + 0.0006283724214881659, + 0.004264773800969124, + 0.0010369535302743316, + 0.0006675707991234958, + 0.0012225199025124311, + 0.0006968892412260175, + 0.0037488772068172693, + 0.03928215056657791, + 0.012021003291010857, + 0.26590338349342346, + 0.024981072172522545, + 0.011186007410287857, + 0.02628915384411812, + 0.017039302736520767, + 0.22819063067436218, + 0.004920323379337788, + 0.001944654155522585, + 0.018830114975571632, + 0.002461827825754881, + 0.0017689239466562867, + 0.0037870246451348066, + 0.003231088863685727, + 0.011303644627332687, + 0.002527954988181591, + 0.001717515173368156, + 0.003446602262556553, + 0.0013319721911102533, + 0.0019837399013340473, + 0.005544680170714855, + 0.002750007202848792, + 0.00379460328258574, + 0.0037934461142867804, + 0.0013463987270370126, + 0.007036000024527311, + 0.001806683954782784, + 0.0034641448874026537, + 0.0021518354769796133, + 0.002414178568869829, + 0.0019416240975260735, + 0.002226305892691016, + 0.002961750840768218, + 0.0016806327039375901, + 0.002183508360758424, + 0.0029079467058181763, + 0.002873161341995001, + 0.00504605146124959, + 0.0020145471207797527, + 0.0027035402599722147, + 0.005355400033295155, + 0.0028011540416628122, + 0.007847827859222889, + 0.004190012812614441, + 0.003576004644855857, + 0.0036237204913049936, + 0.0033430138137191534, + 0.005966152995824814, + 0.006216227076947689, + 0.009065313264727592, + 0.007061263546347618, + 0.007197457365691662, + 0.007621778640896082, + 0.00456976005807519, + 0.006347734481096268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.5762330293655396 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346543803811073 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.026346543803811073, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346543803811073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09976374357938766, + 0.022968892008066177, + 0.01896395906805992, + 0.6783087253570557, + 0.02548023872077465, + 0.025288406759500504, + 0.033818647265434265, + 0.03447955474257469, + 0.03201797604560852, + 0.028909901157021523 + ], + "credit_map": [ + 0.00697827385738492, + 0.006240281742066145, + 0.022903986275196075, + 0.012730808928608894, + 0.007116485387086868, + 0.015887759625911713, + 0.01032225601375103, + 0.01758388802409172, + 0.0018261122750118375, + 0.0014942513080313802, + 0.005503967870026827, + 0.003141802968457341, + 0.0019080712227150798, + 0.0028573644813150167, + 0.0023480409290641546, + 0.0038892829325050116, + 0.0014996548416092992, + 0.0008761024218983948, + 0.0055794683285057545, + 0.0013732517836615443, + 0.0010143748950213194, + 0.0018734056502580643, + 0.0010828635422512889, + 0.005664838943630457, + 0.04272529482841492, + 0.013132505118846893, + 0.2889338731765747, + 0.02742132544517517, + 0.012119471095502377, + 0.028118956834077835, + 0.018380766734480858, + 0.24747654795646667, + 0.002811237471178174, + 0.002036504214629531, + 0.003544746432453394, + 0.002241019392386079, + 0.002488973317667842, + 0.00437212735414505, + 0.004238686058670282, + 0.0037469444796442986, + 0.0034822896122932434, + 0.0014394144527614117, + 0.008057351224124432, + 0.0018588007660582662, + 0.0013794403057545424, + 0.0031468675006181, + 0.0022704985458403826, + 0.003653742605820298, + 0.0032837833277881145, + 0.0022808655630797148, + 0.004475489258766174, + 0.004086374770849943, + 0.005237304139882326, + 0.0035894326865673065, + 0.006672998890280724, + 0.004192398861050606, + 0.003040738170966506, + 0.004081695340573788, + 0.003029379528015852, + 0.003637074027210474, + 0.003922513220459223, + 0.0056709665805101395, + 0.00660723727196455, + 0.00448994804173708, + 0.0027318468783050776, + 0.004098489414900541, + 0.0034188535064458847, + 0.006691775284707546, + 0.003826600732281804, + 0.003537906566634774, + 0.0044622644782066345, + 0.0032502408139407635, + 0.003154496429488063, + 0.0026155756786465645, + 0.003571407636627555, + 0.0048638153821229935, + 0.0038429773412644863, + 0.003973488695919514, + 0.003057019552215934, + 0.0038311195094138384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.026346543803811073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scientifically inspired style, the camera executes a steady forward tracking shot, gliding through the deep, inky black of the abyss. The powerful, cool-white floodlight from the camera's light source carves a crisp circle on the seafloor, contrasting starkly against the surrounding impenetrable darkness. High frame rate capture renders the entire scene in ultra-smooth slow motion, highlighting the environment's intricate details with clinical precision. The camera first cruises over a vast abyssal plain, where the light reveals every tiny grain of pale sediment, swirling in soft, lingering clouds as the camera passes. The terrain then transitions to the jagged, oppressive edge of a massive volcanic ridge, its black slate formations casting long, sharp shadows. The water is exceptionally clear, exceptionally crystal clear, and the camera's path leads it to the cliffs of the deep, shadowy trench, offering an unobstructed view; the floodlight illuminates the sharp edge of the abyss before being completely swallowed up, plummeting into unseen depths, creating an atmosphere of breathtaking discovery and profound isolation.", + "seed": 19504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_3204_seed_19504_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.39596310555934905, + "vlm_logprob": -0.39596310555934905, + "vlm_yes_no_margin": 1.7, + "vlm_token_credit": -0.39596310555934905, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07218507677316666 + ], + "opposite_logprob_tokens": [ + -2.6971850395202637 + ], + "logprob_reward": -0.07218507677316666, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07218507677316666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30726954340934753, + 0.10406095534563065, + 0.010341132059693336, + 0.016603481024503708, + 0.07589206099510193, + 0.013603109866380692, + 0.035230640321969986, + 0.34395408630371094, + 0.05746479332447052, + 0.035580143332481384 + ], + "credit_map": [ + 0.02021123468875885, + 0.008297729305922985, + 0.011523024179041386, + 0.01364182773977518, + 0.01125518698245287, + 0.12442708015441895, + 0.008880591951310635, + 0.10903287678956985, + 0.008051819168031216, + 0.0047114090994000435, + 0.006081594154238701, + 0.003774031065404415, + 0.004269647877663374, + 0.035260580480098724, + 0.003342817770317197, + 0.03856906294822693, + 0.001297393231652677, + 0.0011920082615688443, + 0.0015109522501006722, + 0.0007587701547890902, + 0.0010842804331332445, + 0.0019470991101115942, + 0.0008752976427786052, + 0.001675331499427557, + 0.002778019290417433, + 0.0013430762337520719, + 0.001415869453921914, + 0.0010528743732720613, + 0.0015081281308084726, + 0.0033286353573203087, + 0.0010245766025036573, + 0.004152301698923111, + 0.014921436086297035, + 0.013902947306632996, + 0.0052871680818498135, + 0.004854441620409489, + 0.009027853608131409, + 0.008187639527022839, + 0.003644235199317336, + 0.01606633886694908, + 0.0016532543813809752, + 0.002071864902973175, + 0.0012300704838708043, + 0.0022540148347616196, + 0.0022867650259286165, + 0.0011955862864851952, + 0.0018459130078554153, + 0.0010656421072781086, + 0.0027408632449805737, + 0.004088526125997305, + 0.003466481575742364, + 0.007218403276056051, + 0.004855799023061991, + 0.00352981174364686, + 0.00544073898345232, + 0.0038900182116776705, + 0.13764195144176483, + 0.02472451888024807, + 0.12290263921022415, + 0.014198584482073784, + 0.010534169152379036, + 0.011580763384699821, + 0.010076381266117096, + 0.012295091524720192, + 0.006039456930011511, + 0.008447805419564247, + 0.008884386159479618, + 0.008111830800771713, + 0.004724524449557066, + 0.004901331849396229, + 0.008286961354315281, + 0.008068501017987728, + 0.006679586600512266, + 0.005793188698589802, + 0.0054299719631671906, + 0.00269358791410923, + 0.004386825952678919, + 0.002810729667544365, + 0.0032016525510698557, + 0.004584599286317825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.07218507677316666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603946089744568 + ], + "opposite_logprob_tokens": [ + -1.9103946685791016 + ], + "logprob_reward": -0.1603946089744568, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603946089744568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11641982942819595, + 0.0477600134909153, + 0.012503694742918015, + 0.026430653408169746, + 0.12663456797599792, + 0.0842890813946724, + 0.11242437362670898, + 0.3303053379058838, + 0.07784252613782883, + 0.06538987904787064 + ], + "credit_map": [ + 0.017501352354884148, + 0.007103860378265381, + 0.009143305942416191, + 0.008570586331188679, + 0.006813480518758297, + 0.035920217633247375, + 0.006765259895473719, + 0.024601774290204048, + 0.003184797940775752, + 0.0021180741023272276, + 0.0026757470332086086, + 0.0016327626071870327, + 0.0028218606021255255, + 0.015818148851394653, + 0.0021886888425797224, + 0.01731993816792965, + 0.001683044945821166, + 0.0020009607542306185, + 0.001673251623287797, + 0.00130551028996706, + 0.0015118616865947843, + 0.0014018634101375937, + 0.001426082686521113, + 0.0015011202776804566, + 0.0044206902384757996, + 0.0019073180155828595, + 0.002034006640315056, + 0.0018579314928501844, + 0.0024857032112777233, + 0.005398706998676062, + 0.0016310866922140121, + 0.006695212796330452, + 0.020009638741612434, + 0.010303890332579613, + 0.010127225890755653, + 0.007749502081423998, + 0.01319107785820961, + 0.02029469422996044, + 0.008129876106977463, + 0.03682866320014, + 0.008909014984965324, + 0.00996352732181549, + 0.005027819890528917, + 0.011174987070262432, + 0.011838575825095177, + 0.009312337264418602, + 0.01285399030894041, + 0.015208842232823372, + 0.007742736488580704, + 0.0135448407381773, + 0.009944848716259003, + 0.015607407316565514, + 0.026029545813798904, + 0.01060627494007349, + 0.019886836409568787, + 0.009061887860298157, + 0.13463780283927917, + 0.023226657882332802, + 0.1217539981007576, + 0.014009815640747547, + 0.00960349291563034, + 0.007791052106767893, + 0.009025856852531433, + 0.010256679728627205, + 0.0060705747455358505, + 0.008094385266304016, + 0.01190358679741621, + 0.016827397048473358, + 0.00553669361397624, + 0.004637531004846096, + 0.012085706926882267, + 0.012686655856668949, + 0.008026215247809887, + 0.01090505626052618, + 0.008359355852007866, + 0.005676624830812216, + 0.008008521050214767, + 0.006869683973491192, + 0.007383916527032852, + 0.010160508565604687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.1603946089744568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294001340866089 + ], + "opposite_logprob_tokens": [ + -2.2379400730133057 + ], + "logprob_reward": -0.11294001340866089, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294001340866089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05475460737943649, + 0.21293723583221436, + 0.021983355283737183, + 0.01354027446359396, + 0.01546194963157177, + 0.0803765133023262, + 0.05129119008779526, + 0.32416990399360657, + 0.1847558468580246, + 0.040729086846113205 + ], + "credit_map": [ + 0.007328621111810207, + 0.004954116884618998, + 0.005851252004504204, + 0.007356517948210239, + 0.004418364260345697, + 0.011782452464103699, + 0.005320759955793619, + 0.0077425213530659676, + 0.01703214831650257, + 0.010182797908782959, + 0.008882231079041958, + 0.007383356802165508, + 0.008042540401220322, + 0.07422073185443878, + 0.005679956171661615, + 0.08151347190141678, + 0.0027618366293609142, + 0.00203562225215137, + 0.0024399023968726397, + 0.0012629644479602575, + 0.0021605330985039473, + 0.0043821618892252445, + 0.002204341348260641, + 0.004735993687063456, + 0.0013424400240182877, + 0.00093101616948843, + 0.0014012104365974665, + 0.0009618547046557069, + 0.0024573628325015306, + 0.0025097199250012636, + 0.0013834108831360936, + 0.00255325878970325, + 0.0010653802892193198, + 0.0015163017669692636, + 0.0015117176808416843, + 0.0020642548333853483, + 0.0033234995789825916, + 0.002012911019846797, + 0.0020244051702320576, + 0.0019434788264334202, + 0.008553923107683659, + 0.015289782546460629, + 0.005874950904399157, + 0.008124700747430325, + 0.012982279062271118, + 0.007574097253382206, + 0.012064526788890362, + 0.009912255220115185, + 0.0048804208636283875, + 0.005581525154411793, + 0.005496314261108637, + 0.011112889274954796, + 0.007105444557964802, + 0.0054476093500852585, + 0.008230556733906269, + 0.00343643082305789, + 0.1293964385986328, + 0.024603428319096565, + 0.11715293675661087, + 0.01504860445857048, + 0.009219588711857796, + 0.008600622415542603, + 0.009341101162135601, + 0.010807187296450138, + 0.015103875659406185, + 0.019711453467607498, + 0.03210745006799698, + 0.031074928119778633, + 0.015413305722177029, + 0.013562615029513836, + 0.03122204728424549, + 0.02656017243862152, + 0.006480314768850803, + 0.006691938266158104, + 0.005594577640295029, + 0.004308083560317755, + 0.0043210359290242195, + 0.0030884125735610723, + 0.003953167703002691, + 0.006291557569056749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.11294001340866089 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005730628967285 + ], + "opposite_logprob_tokens": [ + -0.22557303309440613 + ], + "logprob_reward": -1.6005730628967285, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005730628967285, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07553233951330185, + 0.04536420479416847, + 0.02436785213649273, + 0.03215232864022255, + 0.15693145990371704, + 0.09826280921697617, + 0.08711455017328262, + 0.3190940320491791, + 0.08020482957363129, + 0.08097562193870544 + ], + "credit_map": [ + 0.0044868867844343185, + 0.004093402996659279, + 0.008065400645136833, + 0.008285504765808582, + 0.003921821713447571, + 0.02020798996090889, + 0.005359438713639975, + 0.021111896261572838, + 0.003444720758125186, + 0.002602323656901717, + 0.005373455584049225, + 0.002460078103467822, + 0.0027934075333178043, + 0.012591731734573841, + 0.0023399130441248417, + 0.013758577406406403, + 0.003134947968646884, + 0.002142456127330661, + 0.0022635480854660273, + 0.0017198032001033425, + 0.002148020314052701, + 0.005303956568241119, + 0.0018078046850860119, + 0.005847316235303879, + 0.005303736310452223, + 0.004149300511926413, + 0.0023099579848349094, + 0.00236512441188097, + 0.0037194264587014914, + 0.005554333329200745, + 0.001998485531657934, + 0.006751967128366232, + 0.027385419234633446, + 0.011262120679020882, + 0.012136557139456272, + 0.006171869579702616, + 0.015792643651366234, + 0.026708116754889488, + 0.008701615035533905, + 0.04877311736345291, + 0.01090454962104559, + 0.01571115292608738, + 0.007302647922188044, + 0.016909116879105568, + 0.014974474906921387, + 0.00813352596014738, + 0.012397639453411102, + 0.011929702013731003, + 0.0076535725966095924, + 0.0098654143512249, + 0.00740777887403965, + 0.01382058672606945, + 0.018589546903967857, + 0.011672478169202805, + 0.01075872965157032, + 0.007346436847001314, + 0.11863553524017334, + 0.02438690885901451, + 0.10767050832509995, + 0.01828203722834587, + 0.011834383010864258, + 0.011690409854054451, + 0.015481364913284779, + 0.011112863197922707, + 0.005653235130012035, + 0.009171758778393269, + 0.010273958556354046, + 0.01550428569316864, + 0.009204551577568054, + 0.008736539632081985, + 0.01338869147002697, + 0.008271806873381138, + 0.014967555180191994, + 0.010269366204738617, + 0.012919571250677109, + 0.007036171853542328, + 0.008196232840418816, + 0.007758206222206354, + 0.008848154917359352, + 0.010980363003909588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -1.6005730628967285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372276574373245 + ], + "opposite_logprob_tokens": [ + -3.4087228775024414 + ], + "logprob_reward": -0.03372276574373245, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372276574373245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05950546637177467, + 0.12788377702236176, + 0.011879351921379566, + 0.010938282124698162, + 0.2223934531211853, + 0.06890139728784561, + 0.07491865754127502, + 0.2341640144586563, + 0.07702888548374176, + 0.11238673329353333 + ], + "credit_map": [ + 0.0030922270379960537, + 0.0051834979094564915, + 0.004265681840479374, + 0.006399643141776323, + 0.003352831117808819, + 0.018024466931819916, + 0.0030554209370166063, + 0.016131700947880745, + 0.010092453099787235, + 0.0064815860241651535, + 0.0059473542496562, + 0.004751691594719887, + 0.004821742884814739, + 0.04388686642050743, + 0.004007928539067507, + 0.04789415001869202, + 0.0014644076582044363, + 0.001351029146462679, + 0.0019082775106653571, + 0.001383354188874364, + 0.0015933223767206073, + 0.0013925728853791952, + 0.0016620797105133533, + 0.0011243082117289305, + 0.0008713376009836793, + 0.0014977366663515568, + 0.0013757632113993168, + 0.0010921887587755919, + 0.0022016125731170177, + 0.0013990476727485657, + 0.0011012121103703976, + 0.0013993834145367146, + 0.03965722396969795, + 0.018776891753077507, + 0.017563633620738983, + 0.009156838990747929, + 0.023160366341471672, + 0.035663455724716187, + 0.012551702558994293, + 0.06586335599422455, + 0.006896959152072668, + 0.011131142266094685, + 0.004965499974787235, + 0.013761547394096851, + 0.009499276988208294, + 0.005151133518666029, + 0.011136477813124657, + 0.006359361112117767, + 0.005748235620558262, + 0.008329560980200768, + 0.0063330731354653835, + 0.011241291649639606, + 0.017662055790424347, + 0.00946697499603033, + 0.008679606951773167, + 0.007457854226231575, + 0.09807422012090683, + 0.015476635657250881, + 0.08618833124637604, + 0.008122441358864307, + 0.006776272784918547, + 0.004997941665351391, + 0.006357564125210047, + 0.0081705953925848, + 0.006360986735671759, + 0.011343744583427906, + 0.012320881709456444, + 0.011609440669417381, + 0.008168096654117107, + 0.005390959791839123, + 0.009865306317806244, + 0.011969463899731636, + 0.016892245039343834, + 0.012654935009777546, + 0.01810179091989994, + 0.009048957377672195, + 0.010526378639042377, + 0.007169215939939022, + 0.017484480515122414, + 0.020508738234639168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.03372276574373245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a scientifically inspired style, the camera executes a steady forward tracking shot, gliding through the deep, inky black of the abyss. The powerful, cool-white floodlight from the camera's light source carves a crisp circle on the seafloor, contrasting starkly against the surrounding impenetrable darkness. High frame rate capture renders the entire scene in ultra-smooth slow motion, highlighting the environment's intricate details with clinical precision. The camera first cruises over a vast abyssal plain, where the light reveals every tiny grain of pale sediment, swirling in soft, lingering clouds as the camera passes. The terrain then transitions to the jagged, oppressive edge of a massive volcanic ridge, its black slate formations casting long, sharp shadows. The water is exceptionally clear, exceptionally crystal clear, and the camera's path leads it to the cliffs of the deep, shadowy trench, offering an unobstructed view; the floodlight illuminates the sharp edge of the abyss before being completely swallowed up, plummeting into unseen depths, creating an atmosphere of breathtaking discovery and profound isolation.", + "seed": 19505, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_3204_seed_19505_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.289299529697746, + "vlm_logprob": -0.289299529697746, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.289299529697746, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004526250530034304 + ], + "opposite_logprob_tokens": [ + -5.754526138305664 + ], + "logprob_reward": -0.004526250530034304, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004526250530034304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1108035072684288, + 0.2623589336872101, + 0.2057689130306244, + 0.03911763057112694, + 0.051222506910562515, + 0.02840602584183216, + 0.03843941539525986, + 0.11081098020076752, + 0.07442537695169449, + 0.07864676415920258 + ], + "credit_map": [ + 0.014311281032860279, + 0.017376216128468513, + 0.015659332275390625, + 0.01906874217092991, + 0.01123621966689825, + 0.012099267914891243, + 0.007170748431235552, + 0.01388170849531889, + 0.018685651943087578, + 0.1025976613163948, + 0.017692970111966133, + 0.007200140971690416, + 0.03735240921378136, + 0.05777063965797424, + 0.006835122127085924, + 0.014224314130842686, + 0.010949341580271721, + 0.0753018707036972, + 0.012887801975011826, + 0.0033551903907209635, + 0.0759618803858757, + 0.018649684265255928, + 0.003782625775784254, + 0.004880516789853573, + 0.002384436782449484, + 0.004210011567920446, + 0.005257314071059227, + 0.004280710127204657, + 0.004810220096260309, + 0.00580464256927371, + 0.004139489494264126, + 0.008230808191001415, + 0.004430430009961128, + 0.007875820621848106, + 0.0106829097494483, + 0.005332893691956997, + 0.007386586628854275, + 0.006551915779709816, + 0.004487294238060713, + 0.0044746557250618935, + 0.003051508916541934, + 0.0044374046847224236, + 0.0043541244231164455, + 0.003740343963727355, + 0.003390066558495164, + 0.0037072645500302315, + 0.0031368141062557697, + 0.0025884981732815504, + 0.0056616258807480335, + 0.003930131439119577, + 0.005457655061036348, + 0.004661886487156153, + 0.003873113077133894, + 0.003911321517080069, + 0.0037212998140603304, + 0.007222378626465797, + 0.011590439826250076, + 0.010350405238568783, + 0.0166404340416193, + 0.015467106364667416, + 0.006871641613543034, + 0.017458273097872734, + 0.012584991753101349, + 0.01984769105911255, + 0.007164707873016596, + 0.011694859713315964, + 0.009687522426247597, + 0.012241947464644909, + 0.008950373157858849, + 0.007255608681589365, + 0.008781109005212784, + 0.008649243041872978, + 0.0077649871818721294, + 0.010445871390402317, + 0.009733223356306553, + 0.01062764972448349, + 0.012289921753108501, + 0.006050590891391039, + 0.009895708411931992, + 0.011838811449706554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.004526250530034304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801357984542847 + ], + "opposite_logprob_tokens": [ + -4.634801387786865 + ], + "logprob_reward": -0.009801357984542847, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801357984542847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18207336962223053, + 0.06012016162276268, + 0.4209219217300415, + 0.0839674174785614, + 0.03611620143055916, + 0.04072272777557373, + 0.04117118567228317, + 0.058608368039131165, + 0.02817635051906109, + 0.048122338950634 + ], + "credit_map": [ + 0.024500222876667976, + 0.03487206995487213, + 0.021495051681995392, + 0.021996738389134407, + 0.02835347130894661, + 0.021072853356599808, + 0.012533441185951233, + 0.017249522730708122, + 0.00961329136043787, + 0.005536009557545185, + 0.007322730962187052, + 0.002754796762019396, + 0.008482852950692177, + 0.018032129853963852, + 0.0031192884780466557, + 0.005259065423160791, + 0.026908623054623604, + 0.1639484316110611, + 0.016390545293688774, + 0.007855379022657871, + 0.16185179352760315, + 0.025097142904996872, + 0.007828562520444393, + 0.011041413992643356, + 0.00592777319252491, + 0.026240339502692223, + 0.004983380436897278, + 0.006873954553157091, + 0.02208022028207779, + 0.0059104301035404205, + 0.005054329056292772, + 0.0068969931453466415, + 0.0029973648488521576, + 0.00599605031311512, + 0.004814258776605129, + 0.0041540395468473434, + 0.006597686558961868, + 0.0048865992575883865, + 0.003257768927142024, + 0.003412432735785842, + 0.004713942762464285, + 0.006544755771756172, + 0.005023622419685125, + 0.006003853399306536, + 0.004315588157624006, + 0.004366148728877306, + 0.004872043617069721, + 0.004882771987468004, + 0.0039347512647509575, + 0.0034843795001506805, + 0.005219186656177044, + 0.005039699375629425, + 0.0037536637391895056, + 0.006027396768331528, + 0.005926464684307575, + 0.007785643916577101, + 0.006496231537312269, + 0.004143721889704466, + 0.009131941944360733, + 0.00878006312996149, + 0.004298892803490162, + 0.011624972335994244, + 0.006491788197308779, + 0.007640758063644171, + 0.0024144896306097507, + 0.00545505341142416, + 0.0038326503708958626, + 0.004254222381860018, + 0.0031380420550704002, + 0.0030831117182970047, + 0.002653676550835371, + 0.0033451037015765905, + 0.004119711462408304, + 0.006283581722527742, + 0.00721708545461297, + 0.007278820499777794, + 0.006904371082782745, + 0.008231114596128464, + 0.0040002865716814995, + 0.004087368957698345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.009801357984542847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781762026250362 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781762026250362, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781762026250362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15050716698169708, + 0.22149817645549774, + 0.1573754847049713, + 0.11123759299516678, + 0.022663868963718414, + 0.03657105937600136, + 0.025132939219474792, + 0.11299395561218262, + 0.07332298904657364, + 0.08869675546884537 + ], + "credit_map": [ + 0.018163060769438744, + 0.028913287445902824, + 0.01794133521616459, + 0.02112608216702938, + 0.022157888859510422, + 0.016345541924238205, + 0.009961557574570179, + 0.01589842140674591, + 0.015667017549276352, + 0.0869140475988388, + 0.017041822895407677, + 0.006575462408363819, + 0.032298725098371506, + 0.04438355937600136, + 0.006149638444185257, + 0.01246790774166584, + 0.01132967509329319, + 0.05693243071436882, + 0.008524171076714993, + 0.0033529354259371758, + 0.061120547354221344, + 0.008259222842752934, + 0.003229892812669277, + 0.004626602400094271, + 0.007878446951508522, + 0.034570835530757904, + 0.006100078579038382, + 0.008316517807543278, + 0.030064592137932777, + 0.008128976449370384, + 0.005917667876929045, + 0.01026048231869936, + 0.0017122384160757065, + 0.0035814212169498205, + 0.003922243602573872, + 0.0023337281309068203, + 0.004483039956539869, + 0.0025943925138562918, + 0.002392879221588373, + 0.0016439248574897647, + 0.00458013080060482, + 0.004138490650802851, + 0.005972243845462799, + 0.005413908511400223, + 0.003981044981628656, + 0.004036350175738335, + 0.0037608924321830273, + 0.0046880003064870834, + 0.0020512109622359276, + 0.003068601479753852, + 0.002860728884115815, + 0.0024986271746456623, + 0.002419518306851387, + 0.003983758855611086, + 0.004272873513400555, + 0.003977620042860508, + 0.011695980094373226, + 0.007186572067439556, + 0.015117255039513111, + 0.012710241600871086, + 0.006687036715447903, + 0.025604741647839546, + 0.010930444113910198, + 0.023061681538820267, + 0.006057641003280878, + 0.01327118743211031, + 0.011430994607508183, + 0.01386676263064146, + 0.007844080217182636, + 0.008381028659641743, + 0.005874760448932648, + 0.006596527528017759, + 0.0051451451145112514, + 0.008051944896578789, + 0.02448228932917118, + 0.02027667686343193, + 0.008724707178771496, + 0.009591353125870228, + 0.005865827668458223, + 0.006558811292052269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.009781762026250362 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.406315803527832 + ], + "opposite_logprob_tokens": [ + -0.2813158631324768 + ], + "logprob_reward": -1.406315803527832, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.406315803527832, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1419525295495987, + 0.17028042674064636, + 0.12389339506626129, + 0.03834209591150284, + 0.03386325761675835, + 0.023388564586639404, + 0.06241775304079056, + 0.13932578265666962, + 0.09198877215385437, + 0.17454741895198822 + ], + "credit_map": [ + 0.019481627270579338, + 0.02380608581006527, + 0.018275102600455284, + 0.019516492262482643, + 0.01885594055056572, + 0.017210403457283974, + 0.009797697886824608, + 0.01500918809324503, + 0.011756118386983871, + 0.06726477295160294, + 0.011997762136161327, + 0.00471982266753912, + 0.02480943314731121, + 0.036575861275196075, + 0.004602164030075073, + 0.008554510772228241, + 0.004419910255819559, + 0.041153013706207275, + 0.01696847938001156, + 0.0026036230847239494, + 0.03770199418067932, + 0.013486245647072792, + 0.002409449080005288, + 0.0051506939344108105, + 0.0030084499157965183, + 0.009886561892926693, + 0.0026938817463815212, + 0.003272023983299732, + 0.009108650498092175, + 0.002978416858240962, + 0.0031089165713638067, + 0.004285197239369154, + 0.0031093352008610964, + 0.004275162238627672, + 0.0036876045633107424, + 0.004873789846897125, + 0.0069809104315936565, + 0.003989670891314745, + 0.004019665531814098, + 0.0029271207749843597, + 0.002652318449690938, + 0.003018950344994664, + 0.0031670485623180866, + 0.0027968306094408035, + 0.002458572154864669, + 0.0026503042317926884, + 0.002792536048218608, + 0.0038520044181495905, + 0.006011344958096743, + 0.0069962735287845135, + 0.006578817032277584, + 0.007533264812082052, + 0.007135152816772461, + 0.008869470097124577, + 0.010074286721646786, + 0.009219148196280003, + 0.012660120613873005, + 0.013488037511706352, + 0.01724962517619133, + 0.02339867874979973, + 0.009774554520845413, + 0.020816607400774956, + 0.014996852725744247, + 0.026941323652863503, + 0.007324857637286186, + 0.016689717769622803, + 0.01402133796364069, + 0.015120179392397404, + 0.012867603451013565, + 0.010767623782157898, + 0.007410314865410328, + 0.007787146605551243, + 0.011980067938566208, + 0.01925290934741497, + 0.0560186468064785, + 0.045508094131946564, + 0.009430729784071445, + 0.010716145858168602, + 0.010745913721621037, + 0.01089491043239832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -1.406315803527832 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072474420070648 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072474420070648, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072474420070648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20468761026859283, + 0.27459514141082764, + 0.0950426384806633, + 0.039128322154283524, + 0.0486639030277729, + 0.01797013357281685, + 0.04031529277563095, + 0.10799498856067657, + 0.05956077203154564, + 0.11204119771718979 + ], + "credit_map": [ + 0.02751966565847397, + 0.03639746457338333, + 0.026343604549765587, + 0.02658277377486229, + 0.02811579965054989, + 0.02385910414159298, + 0.014364750124514103, + 0.021504445001482964, + 0.020114243030548096, + 0.10547855496406555, + 0.019171765074133873, + 0.007363407872617245, + 0.03759418800473213, + 0.06273091584444046, + 0.00666916323825717, + 0.01547290850430727, + 0.007543100509792566, + 0.028673158958554268, + 0.0182923823595047, + 0.0022948274854570627, + 0.021394582465291023, + 0.009886086918413639, + 0.0023348014801740646, + 0.004623704589903355, + 0.0025390968658030033, + 0.010577850043773651, + 0.003273187205195427, + 0.0036462375428527594, + 0.009067410603165627, + 0.0034122723154723644, + 0.0032419513445347548, + 0.003370316233485937, + 0.004289776086807251, + 0.005713941995054483, + 0.009478168562054634, + 0.006332326214760542, + 0.00783101748675108, + 0.0060077086091041565, + 0.005528672598302364, + 0.003482293337583542, + 0.0020138174295425415, + 0.0018908062484115362, + 0.002544064773246646, + 0.0026310232933610678, + 0.0014359484193846583, + 0.0018534226110205054, + 0.001572909182868898, + 0.004028140567243099, + 0.003345620585605502, + 0.00450957752764225, + 0.005164762027561665, + 0.0044108061119914055, + 0.005155081860721111, + 0.005271931644529104, + 0.006846624892205, + 0.005610890220850706, + 0.01075795292854309, + 0.008871073834598064, + 0.012961643747985363, + 0.018335741013288498, + 0.006720413453876972, + 0.0138615183532238, + 0.013679839670658112, + 0.02280680648982525, + 0.0051500736735761166, + 0.011120609939098358, + 0.008670620620250702, + 0.009849043563008308, + 0.0063138697296381, + 0.007258163765072823, + 0.005843065679073334, + 0.005355325061827898, + 0.007664978038519621, + 0.014083904214203358, + 0.032990097999572754, + 0.02607320062816143, + 0.008628001436591148, + 0.005817137658596039, + 0.008764387108385563, + 0.008019505068659782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.016072474420070648 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a scientifically inspired style, the camera executes a steady forward tracking shot, gliding through the deep, inky black of the abyss. The powerful, cool-white floodlight from the camera's light source carves a crisp circle on the seafloor, contrasting starkly against the surrounding impenetrable darkness. High frame rate capture renders the entire scene in ultra-smooth slow motion, highlighting the environment's intricate details with clinical precision. The camera first cruises over a vast abyssal plain, where the light reveals every tiny grain of pale sediment, swirling in soft, lingering clouds as the camera passes. The terrain then transitions to the jagged, oppressive edge of a massive volcanic ridge, its black slate formations casting long, sharp shadows. The water is exceptionally clear, exceptionally crystal clear, and the camera's path leads it to the cliffs of the deep, shadowy trench, offering an unobstructed view; the floodlight illuminates the sharp edge of the abyss before being completely swallowed up, plummeting into unseen depths, creating an atmosphere of breathtaking discovery and profound isolation.", + "seed": 19506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_3/index_3204_seed_19506_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.35876081585884095, + "vlm_logprob": -0.35876081585884095, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.35876081585884095, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03131527826189995 + ], + "opposite_logprob_tokens": [ + -3.531315326690674 + ], + "logprob_reward": -0.03131527826189995, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03131527826189995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021975144743919373, + 0.07934220135211945, + 0.0173720084130764, + 0.12395267933607101, + 0.4759503901004791, + 0.07834263145923615, + 0.026155464351177216, + 0.03675338625907898, + 0.09212910383939743, + 0.04802703112363815 + ], + "credit_map": [ + 0.002639696467667818, + 0.0023000366054475307, + 0.0024227492976933718, + 0.0017700341995805502, + 0.0025494650471955538, + 0.0035032613668590784, + 0.002240526955574751, + 0.004549374338239431, + 0.011738982982933521, + 0.00528987729921937, + 0.004919791594147682, + 0.003027839818969369, + 0.007631721906363964, + 0.02046273462474346, + 0.0037476103752851486, + 0.02252363972365856, + 0.0018374926876276731, + 0.0015316872159019113, + 0.0015583219937980175, + 0.001298033632338047, + 0.0028094889130443335, + 0.0037194674368947744, + 0.0012563540367409587, + 0.0033611610997468233, + 0.008959046564996243, + 0.013203655369579792, + 0.010393630713224411, + 0.008338730782270432, + 0.02093692682683468, + 0.016507413238286972, + 0.01864195428788662, + 0.02697131410241127, + 0.0432395339012146, + 0.027394883334636688, + 0.01885131560266018, + 0.019204869866371155, + 0.09878061711788177, + 0.07349295914173126, + 0.028536925092339516, + 0.16644924879074097, + 0.009366054087877274, + 0.010658475570380688, + 0.006737444084137678, + 0.010360955260694027, + 0.010627565905451775, + 0.011142076924443245, + 0.010471168905496597, + 0.008978882804512978, + 0.0026571163907647133, + 0.003149315482005477, + 0.003995860461145639, + 0.005716066807508469, + 0.0035326112993061543, + 0.002087751869112253, + 0.0033839396201074123, + 0.0016328024212270975, + 0.005456953775137663, + 0.007394437212496996, + 0.004696432501077652, + 0.004396814852952957, + 0.003307238221168518, + 0.002897984581068158, + 0.0034704520367085934, + 0.005133071448653936, + 0.0075395843014121056, + 0.005372495390474796, + 0.035497378557920456, + 0.021920213475823402, + 0.0050097880885005, + 0.00460911076515913, + 0.006425808649510145, + 0.005754712037742138, + 0.004796222783625126, + 0.006981588900089264, + 0.00852962676435709, + 0.008710567839443684, + 0.005971674807369709, + 0.004502676893025637, + 0.005062825512140989, + 0.0034718422684818506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.03131527826189995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062074389308691025 + ], + "opposite_logprob_tokens": [ + -2.8120744228363037 + ], + "logprob_reward": -0.062074389308691025, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062074389308691025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03653942048549652, + 0.2150210589170456, + 0.03866683319211006, + 0.05886060744524002, + 0.31572315096855164, + 0.0626031756401062, + 0.05137708783149719, + 0.1001463532447815, + 0.0934259295463562, + 0.027636338025331497 + ], + "credit_map": [ + 0.005483720451593399, + 0.0033137716818600893, + 0.005125155206769705, + 0.0029569435864686966, + 0.005955337081104517, + 0.004654391203075647, + 0.00445151049643755, + 0.004598590079694986, + 0.024117130786180496, + 0.01159487571567297, + 0.010884910821914673, + 0.007710722275078297, + 0.01746038720011711, + 0.06205873191356659, + 0.010199995711445808, + 0.07099432498216629, + 0.00543373916298151, + 0.0051098172552883625, + 0.003742518834769726, + 0.002554420381784439, + 0.0076505327597260475, + 0.004902415908873081, + 0.003726441413164139, + 0.00554695026949048, + 0.003817284479737282, + 0.00575518561527133, + 0.0032979915849864483, + 0.0031287900637835264, + 0.014463677071034908, + 0.008781574666500092, + 0.007513111922889948, + 0.01210299413651228, + 0.026980319991707802, + 0.017999907955527306, + 0.013129618018865585, + 0.010190707631409168, + 0.06796304136514664, + 0.048398472368717194, + 0.019070511683821678, + 0.11199057102203369, + 0.007138017565011978, + 0.013015815988183022, + 0.006338364910334349, + 0.005759105086326599, + 0.00966603308916092, + 0.006665890570729971, + 0.008877639658749104, + 0.005142308305948973, + 0.0045759417116642, + 0.007422844413667917, + 0.006678623612970114, + 0.011125195771455765, + 0.00493671465665102, + 0.007073030807077885, + 0.0052137961611151695, + 0.004350948613137007, + 0.010616733692586422, + 0.01361225638538599, + 0.016850702464580536, + 0.011417160741984844, + 0.009941832162439823, + 0.010293209925293922, + 0.014006723649799824, + 0.013407738879323006, + 0.0067332093603909016, + 0.009986966848373413, + 0.017971405759453773, + 0.027464643120765686, + 0.007473692763596773, + 0.005281507968902588, + 0.010300248861312866, + 0.008214257657527924, + 0.002457439200952649, + 0.003950750455260277, + 0.004260021727532148, + 0.0051279217004776, + 0.0032604567240923643, + 0.002974172355607152, + 0.0029025846160948277, + 0.002702991710975766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.062074389308691025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002148032188416 + ], + "opposite_logprob_tokens": [ + -2.695021390914917 + ], + "logprob_reward": -0.07002148032188416, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002148032188416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02412538230419159, + 0.09370587766170502, + 0.03633822500705719, + 0.03420655429363251, + 0.31254953145980835, + 0.11411440372467041, + 0.027064986526966095, + 0.08025649189949036, + 0.16679301857948303, + 0.11084555089473724 + ], + "credit_map": [ + 0.003457097103819251, + 0.0033700098283588886, + 0.003973917104303837, + 0.001676643267273903, + 0.0025693338830024004, + 0.002750552259385586, + 0.002726553240790963, + 0.0036012770142406225, + 0.011704840697348118, + 0.005897830706089735, + 0.004981283564120531, + 0.0033528171479701996, + 0.007752617821097374, + 0.026279496029019356, + 0.004326325375586748, + 0.029410669580101967, + 0.004308916162699461, + 0.003770241979509592, + 0.003413212951272726, + 0.0024877581745386124, + 0.006629590410739183, + 0.007419707253575325, + 0.0031283709686249495, + 0.00518042640760541, + 0.0030896365642547607, + 0.003698235610499978, + 0.003572354791685939, + 0.0022004363127052784, + 0.005219951272010803, + 0.003867401508614421, + 0.006437468808144331, + 0.006121067795902491, + 0.026595262810587883, + 0.018458642065525055, + 0.013501687906682491, + 0.011157549917697906, + 0.06616958230733871, + 0.04785009101033211, + 0.01976628415286541, + 0.109050452709198, + 0.01338935736566782, + 0.031442757695913315, + 0.009185276925563812, + 0.009206701070070267, + 0.01245610136538744, + 0.01399532426148653, + 0.009515058249235153, + 0.01492382399737835, + 0.002317995997145772, + 0.0027350089512765408, + 0.003375616855919361, + 0.004284331109374762, + 0.00458019832149148, + 0.003092774422839284, + 0.004719656892120838, + 0.001959402346983552, + 0.007641963195055723, + 0.013669073581695557, + 0.00999768078327179, + 0.011556446552276611, + 0.010843266732990742, + 0.008757377974689007, + 0.010967457666993141, + 0.006823217496275902, + 0.021256715059280396, + 0.01003206055611372, + 0.0690060406923294, + 0.03721983730792999, + 0.005029293708503246, + 0.0060149915516376495, + 0.012460720725357533, + 0.00577335013076663, + 0.007203086744993925, + 0.012169992551207542, + 0.022708550095558167, + 0.032793570309877396, + 0.006571693811565638, + 0.011938236653804779, + 0.011740155518054962, + 0.0057202731259167194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.07002148032188416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005778312683105 + ], + "opposite_logprob_tokens": [ + -0.22557780146598816 + ], + "logprob_reward": -1.6005778312683105, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005778312683105, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.028975242748856544, + 0.01802976429462433, + 0.023286858573555946, + 0.052931997925043106, + 0.5446292757987976, + 0.037080880254507065, + 0.07777082175016403, + 0.02949821949005127, + 0.0835968405008316, + 0.10420007258653641 + ], + "credit_map": [ + 0.0038496884517371655, + 0.0033016016241163015, + 0.004389826208353043, + 0.0023518064990639687, + 0.0033861121628433466, + 0.0038537862710654736, + 0.0028963920194655657, + 0.004946030210703611, + 0.0031828160863369703, + 0.0021246669348329306, + 0.0020169189665466547, + 0.001047423924319446, + 0.0026274239644408226, + 0.002444618381559849, + 0.0015200190246105194, + 0.0030658775940537453, + 0.002841402543708682, + 0.0026476059574633837, + 0.0022060100454837084, + 0.00126190809533, + 0.004798712208867073, + 0.003513964591547847, + 0.0021218927577137947, + 0.0038953640032559633, + 0.0034065365325659513, + 0.004997787531465292, + 0.0033730180002748966, + 0.0024955288972705603, + 0.011501768603920937, + 0.007910341024398804, + 0.007046954240649939, + 0.012200066819787025, + 0.04668392613530159, + 0.0299752876162529, + 0.022861436009407043, + 0.0171587485820055, + 0.11837851256132126, + 0.08400621265172958, + 0.031745146960020065, + 0.1938200443983078, + 0.004060482140630484, + 0.008689247071743011, + 0.0043477993458509445, + 0.003283670637756586, + 0.0059671541675925255, + 0.00343680870719254, + 0.004438001662492752, + 0.0028577176854014397, + 0.007581997197121382, + 0.00975780189037323, + 0.008497420698404312, + 0.014343619346618652, + 0.00914805755019188, + 0.013211837969720364, + 0.008887573145329952, + 0.006342513952404261, + 0.003986067604273558, + 0.0061014085076749325, + 0.004086124245077372, + 0.0034971986897289753, + 0.0034515256993472576, + 0.0028318429831415415, + 0.0030506718903779984, + 0.002493382664397359, + 0.007191505283117294, + 0.005362184718251228, + 0.028830142691731453, + 0.015211129561066628, + 0.006002345122396946, + 0.004207721445709467, + 0.006604519207030535, + 0.010187291540205479, + 0.006183298770338297, + 0.011138458736240864, + 0.021687204018235207, + 0.03130536526441574, + 0.009011993184685707, + 0.01035013236105442, + 0.008998064324259758, + 0.005525560583919287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -1.6005778312683105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep-sea abyssal environment with inky black surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a cool-white floodlight creating a crisp circle of illumination on the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the terrain transition from a vast abyssal plain with pale sediment to a jagged volcanic ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as exceptionally clear and crystal clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady forward tracking shot gliding through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815100133419037 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.029815100133419037, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815100133419037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0429144911468029, + 0.05788528546690941, + 0.02426161617040634, + 0.07809736579656601, + 0.5093218684196472, + 0.05722862109541893, + 0.0356927290558815, + 0.03803398087620735, + 0.09276428073644638, + 0.06379980593919754 + ], + "credit_map": [ + 0.004245116375386715, + 0.00442549167200923, + 0.006076680961996317, + 0.003388793906196952, + 0.00580952363088727, + 0.00626214686781168, + 0.004748717416077852, + 0.007958019152283669, + 0.007228465750813484, + 0.004875710234045982, + 0.004647173453122377, + 0.002510814694687724, + 0.006400225684046745, + 0.01386367529630661, + 0.003355354070663452, + 0.015003863722085953, + 0.002048182999715209, + 0.0019172211177647114, + 0.0015063307946547866, + 0.0011662646429613233, + 0.004402393940836191, + 0.004936521407216787, + 0.0017105462029576302, + 0.006574154831469059, + 0.004937964025884867, + 0.007055983413010836, + 0.004130844958126545, + 0.0032188030891120434, + 0.018419411033391953, + 0.013273440301418304, + 0.009962229058146477, + 0.017098689451813698, + 0.044062089174985886, + 0.027421187609434128, + 0.020910384133458138, + 0.015732189640402794, + 0.11000862717628479, + 0.07918082177639008, + 0.02932126075029373, + 0.18268530070781708, + 0.008066970854997635, + 0.019245130941271782, + 0.005441281478852034, + 0.004608009476214647, + 0.006924168672412634, + 0.004177040886133909, + 0.003881955286487937, + 0.004884068854153156, + 0.003707430325448513, + 0.004133325070142746, + 0.004159900359809399, + 0.007789991330355406, + 0.003791575552895665, + 0.005103767849504948, + 0.0043003372848033905, + 0.002706397557631135, + 0.004391195252537727, + 0.00647205812856555, + 0.006059935316443443, + 0.004941990599036217, + 0.004209768958389759, + 0.0038539457600563765, + 0.004592922516167164, + 0.0035121652763336897, + 0.009166865609586239, + 0.0061590950936079025, + 0.029425689950585365, + 0.02089337445795536, + 0.0059770080260932446, + 0.004555229563266039, + 0.006574008613824844, + 0.010013011284172535, + 0.004857912193983793, + 0.007706765551120043, + 0.01057988777756691, + 0.015052862465381622, + 0.007797630038112402, + 0.006860981695353985, + 0.006016971543431282, + 0.004926794208586216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.029815100133419037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_4/index_tensor([4076, 4076])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_4/index_tensor([4076, 4076])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8522e9314988cb6d466b956b008dad4fc4a27972 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_4/index_tensor([4076, 4076])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot, rendered with surreal detail akin to macro photography, captures a small section of an old painted metal surface under the cool, diffuse light of an overcast day. The primary texture is a thick layer of faded seafoam green paint, its surface a complex map of fine cracks and blisters, revealing the deep orange-brown rust beneath. A gust of wind rusts into the frame, and a loose flake of paint, already curling at the edges, begins to tremble. The camera remains steady as the wind carefully peels the flake away from the rough, oxidized metal layer beneath, slowly and deliberately finally freeing it. Once free, the fragile flake is carried a short distance by the current, tumbling and drifting weightlessly for a brief moment, revealing its painted surface and dark underside. Then, gravity inexorably pulls it down, gently landing on the gritty gray concrete floor below. A tiny, colorful, decaying relic in the desolate industrial setting, it creates a quiet, melancholic atmosphere focused on the passage of time.", + "seed": 11044, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_4/index_4076_seed_11044_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06158193126320839, + "vlm_logprob": -0.06158193126320839, + "vlm_yes_no_margin": 3.424999976158142, + "vlm_token_credit": -0.06158193126320839, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13088853657245636 + ], + "opposite_logprob_tokens": [ + -2.1308884620666504 + ], + "logprob_reward": -0.13088853657245636, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13088853657245636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4800329804420471, + 0.19758322834968567, + 0.06349732726812363, + 0.09194184094667435, + 0.0591944083571434, + 0.03659970313310623, + 0.015508141368627548, + 0.01620359532535076, + 0.02156572788953781, + 0.017873046919703484 + ], + "credit_map": [ + 0.06618122011423111, + 0.08208238333463669, + 0.04173842817544937, + 0.060259874910116196, + 0.07224486023187637, + 0.06548922508955002, + 0.05889017507433891, + 0.033146779984235764, + 0.025542832911014557, + 0.03666668012738228, + 0.015492654405534267, + 0.021821055561304092, + 0.03649837523698807, + 0.023445498198270798, + 0.02443177066743374, + 0.013684369623661041, + 0.006721546873450279, + 0.008069191128015518, + 0.007245954126119614, + 0.00807811040431261, + 0.009764722548425198, + 0.006926758214831352, + 0.009202132932841778, + 0.007488911040127277, + 0.009654517285525799, + 0.01358445268124342, + 0.008314965292811394, + 0.011059996671974659, + 0.01660986617207527, + 0.009167607873678207, + 0.010510021820664406, + 0.013040417805314064, + 0.006464263889938593, + 0.007334317546337843, + 0.006477842107415199, + 0.006743802689015865, + 0.008566534146666527, + 0.005752921104431152, + 0.008779487572610378, + 0.009075237438082695, + 0.007481949869543314, + 0.004894333891570568, + 0.003505873028188944, + 0.003578471951186657, + 0.005271884147077799, + 0.004186610691249371, + 0.003602169919759035, + 0.004078410565853119, + 0.002337111858651042, + 0.002363409148529172, + 0.0011112889042124152, + 0.0017466016579419374, + 0.002469233237206936, + 0.0018238957272842526, + 0.0021453567314893007, + 0.001511245034635067, + 0.002748314756900072, + 0.002734909299761057, + 0.0010854913853108883, + 0.0016456248704344034, + 0.0030234220903366804, + 0.001843145932070911, + 0.001896722475066781, + 0.0012259650975465775, + 0.003095999825745821, + 0.002511544618755579, + 0.0023372634314000607, + 0.0025288721080869436, + 0.003795861965045333, + 0.0026828187983483076, + 0.002319403924047947, + 0.002293963450938463, + 0.0030369258020073175, + 0.0026881094090640545, + 0.0012388306204229593, + 0.001631538849323988, + 0.0033438759855926037, + 0.0022608342114835978, + 0.0022289257030934095, + 0.0014440045924857259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.13088853657245636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036262013018131256 + ], + "opposite_logprob_tokens": [ + -5.628626346588135 + ], + "logprob_reward": -0.0036262013018131256, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036262013018131256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3612172603607178, + 0.1646372228860855, + 0.10342162847518921, + 0.20742562413215637, + 0.038303568959236145, + 0.026280708611011505, + 0.022153766825795174, + 0.026079116389155388, + 0.021014492958784103, + 0.029466653242707253 + ], + "credit_map": [ + 0.05259048566222191, + 0.06649558991193771, + 0.03016718104481697, + 0.04095052182674408, + 0.057242460548877716, + 0.051453396677970886, + 0.0417148619890213, + 0.020602766424417496, + 0.02360502816736698, + 0.031732093542814255, + 0.009780739434063435, + 0.018029307946562767, + 0.03315936401486397, + 0.01838628016412258, + 0.018299801275134087, + 0.011644589714705944, + 0.011067146435379982, + 0.016649004071950912, + 0.00846870057284832, + 0.015189377591013908, + 0.019178742542862892, + 0.010637711733579636, + 0.01289596501737833, + 0.009334963746368885, + 0.020647188648581505, + 0.03197304904460907, + 0.01741959899663925, + 0.029042182490229607, + 0.0360666885972023, + 0.020480958744883537, + 0.025682903826236725, + 0.02611304633319378, + 0.004996927920728922, + 0.005340659525245428, + 0.0040704598650336266, + 0.005541990976780653, + 0.005879812408238649, + 0.004480517469346523, + 0.005110955331474543, + 0.0028822419699281454, + 0.00335898669436574, + 0.0033344163093715906, + 0.0033367641735821962, + 0.0034424548503011465, + 0.00393245043233037, + 0.0028299447149038315, + 0.0037629450671374798, + 0.002282743575051427, + 0.004372652154415846, + 0.0030338051728904247, + 0.002467863028869033, + 0.0021597559098154306, + 0.0038897679187357426, + 0.002538312692195177, + 0.002047667745500803, + 0.0016439396422356367, + 0.0032954178750514984, + 0.004522108938544989, + 0.002077626995742321, + 0.0030800914391875267, + 0.004674428142607212, + 0.0029073236510157585, + 0.003232287708669901, + 0.0022898295428603888, + 0.0039178053848445415, + 0.0030592591501772404, + 0.0019689928740262985, + 0.002222347306087613, + 0.003245298285037279, + 0.0024251306895166636, + 0.0021267877891659737, + 0.0020488696172833443, + 0.004332431126385927, + 0.004086992237716913, + 0.002637013327330351, + 0.0037669208832085133, + 0.004649536218494177, + 0.0032451224979013205, + 0.004066423047333956, + 0.0026822127401828766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0036262013018131256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202654913067818 + ], + "opposite_logprob_tokens": [ + -2.8120265007019043 + ], + "logprob_reward": -0.06202654913067818, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202654913067818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2862091064453125, + 0.11913704872131348, + 0.048297781497240067, + 0.4468946158885956, + 0.01899590902030468, + 0.01504309382289648, + 0.013960160315036774, + 0.021316252648830414, + 0.013281038962304592, + 0.016865041106939316 + ], + "credit_map": [ + 0.03833949193358421, + 0.04827980697154999, + 0.033175546675920486, + 0.03402622416615486, + 0.0432497039437294, + 0.03871605917811394, + 0.0332346148788929, + 0.017187634482979774, + 0.013102415949106216, + 0.02039884217083454, + 0.00963159091770649, + 0.015071679838001728, + 0.023796958848834038, + 0.013119606301188469, + 0.01645444706082344, + 0.0075614978559315205, + 0.005146065726876259, + 0.006277360022068024, + 0.004924047272652388, + 0.007891494780778885, + 0.005954303313046694, + 0.004825240466743708, + 0.007280635181814432, + 0.0059986296109855175, + 0.04042862728238106, + 0.06744145601987839, + 0.03491589054465294, + 0.05633915588259697, + 0.09466177970170975, + 0.04806685075163841, + 0.053030092269182205, + 0.05201071873307228, + 0.0033916987013071775, + 0.003135593608021736, + 0.0019469126127660275, + 0.0019359649159014225, + 0.002801181748509407, + 0.0021008786279708147, + 0.0021656490862369537, + 0.001518028904683888, + 0.0025140915531665087, + 0.0021485353354364634, + 0.0012902902672067285, + 0.0021187334787100554, + 0.0023790127597749233, + 0.0017380643403157592, + 0.0018504585605114698, + 0.0010039068292826414, + 0.002316864440217614, + 0.002296969760209322, + 0.0009073882247321308, + 0.0015079290606081486, + 0.0027727854903787374, + 0.0015117381699383259, + 0.001624900964088738, + 0.0010215839138254523, + 0.0027190709952265024, + 0.0025995858013629913, + 0.0020626778714358807, + 0.0028685787692666054, + 0.0026385225355625153, + 0.002170317806303501, + 0.003312855027616024, + 0.0029446433763951063, + 0.0018743282416835427, + 0.0020292506087571383, + 0.0014085742877796292, + 0.001376566244289279, + 0.0019854374695569277, + 0.0014524651924148202, + 0.0017795971361920238, + 0.0013748196652159095, + 0.002783415373414755, + 0.0025459337048232555, + 0.002160843228921294, + 0.0017155047971755266, + 0.0020194421522319317, + 0.0017970327753573656, + 0.0019455861765891314, + 0.0018972820835188031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.06202654913067818 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067431420087814 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067431420087814, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067431420087814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33492782711982727, + 0.1285146325826645, + 0.11336245387792587, + 0.18458080291748047, + 0.09107782691717148, + 0.05784522742033005, + 0.02089574933052063, + 0.02460782416164875, + 0.021028215065598488, + 0.023159490898251534 + ], + "credit_map": [ + 0.04800024628639221, + 0.056686658412218094, + 0.03549809008836746, + 0.04032498970627785, + 0.048709385097026825, + 0.045135535299777985, + 0.03942454978823662, + 0.02114836312830448, + 0.017041923478245735, + 0.025204524397850037, + 0.009420965798199177, + 0.013434632681310177, + 0.02236981876194477, + 0.01644222065806389, + 0.015698889270424843, + 0.008901670575141907, + 0.01069219782948494, + 0.01804809831082821, + 0.008972728624939919, + 0.017649155110120773, + 0.019586827605962753, + 0.011783960275352001, + 0.014521345496177673, + 0.012108132243156433, + 0.015590193681418896, + 0.027618804946541786, + 0.014580044895410538, + 0.028907056897878647, + 0.03444204851984978, + 0.019799325615167618, + 0.022726314142346382, + 0.0209170151501894, + 0.009353142231702805, + 0.01074483897536993, + 0.010791726410388947, + 0.01563110016286373, + 0.01335839368402958, + 0.009821112267673016, + 0.014495233073830605, + 0.006882274057716131, + 0.006276590749621391, + 0.007304162718355656, + 0.008483910001814365, + 0.007467132527381182, + 0.008673449046909809, + 0.005703353323042393, + 0.0077876439318060875, + 0.006148985121399164, + 0.004180467687547207, + 0.002934213960543275, + 0.002379836980253458, + 0.002048705704510212, + 0.0028395121917128563, + 0.0022747591137886047, + 0.002090035704895854, + 0.0021482184529304504, + 0.002666078507900238, + 0.0027838742826133966, + 0.002422235207632184, + 0.0035178884863853455, + 0.00419209897518158, + 0.002888654125854373, + 0.004202661104500294, + 0.0019343323074281216, + 0.003497063647955656, + 0.0030042175203561783, + 0.002246091840788722, + 0.002889609429985285, + 0.003461324144154787, + 0.002166467485949397, + 0.0022962354123592377, + 0.001467205467633903, + 0.003255423391237855, + 0.0028643477708101273, + 0.002299167914316058, + 0.0035592857748270035, + 0.0029445262625813484, + 0.0022120503708720207, + 0.00382417649962008, + 0.0022005150094628334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.011067431420087814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030093789100647 + ], + "opposite_logprob_tokens": [ + -2.3503010272979736 + ], + "logprob_reward": -0.10030093789100647, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030093789100647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1560405194759369, + 0.16667766869068146, + 0.15542547404766083, + 0.27991729974746704, + 0.07240047305822372, + 0.046483829617500305, + 0.028537744656205177, + 0.020744821056723595, + 0.023971935734152794, + 0.04980025440454483 + ], + "credit_map": [ + 0.019911596551537514, + 0.023657046258449554, + 0.017918281257152557, + 0.015953248366713524, + 0.027356110513210297, + 0.01968258060514927, + 0.01299784705042839, + 0.018563808873295784, + 0.018971271812915802, + 0.03465328365564346, + 0.008956829085946083, + 0.019413195550441742, + 0.03413012996315956, + 0.018997730687260628, + 0.020779862999916077, + 0.010775351896882057, + 0.01627727970480919, + 0.026698347181081772, + 0.013578061014413834, + 0.02021683007478714, + 0.02739410102367401, + 0.015686199069023132, + 0.021872831508517265, + 0.013701830059289932, + 0.02586986869573593, + 0.039386048913002014, + 0.019260838627815247, + 0.03217146545648575, + 0.06183265149593353, + 0.03125051036477089, + 0.03535554185509682, + 0.03479035571217537, + 0.00641126511618495, + 0.009535352699458599, + 0.00738186202943325, + 0.009916795417666435, + 0.011179806664586067, + 0.007389688864350319, + 0.011618862859904766, + 0.008966836147010326, + 0.004484625533223152, + 0.006300848443061113, + 0.004741303157061338, + 0.006061532534658909, + 0.00692688999697566, + 0.004234583117067814, + 0.007502118591219187, + 0.006231926381587982, + 0.003969097975641489, + 0.003986387513577938, + 0.0023631476797163486, + 0.003724136622622609, + 0.005690157879143953, + 0.0031277493108063936, + 0.0035759557504206896, + 0.0021011135540902615, + 0.0019992708694189787, + 0.0028761301655322313, + 0.002027746755629778, + 0.0028287172317504883, + 0.0031196479685604572, + 0.002007190603762865, + 0.003957265056669712, + 0.001928852405399084, + 0.0027242039795964956, + 0.004352465271949768, + 0.0026458012871444225, + 0.0032586827874183655, + 0.0038342666812241077, + 0.0027181117329746485, + 0.0027073940727859735, + 0.0017310091061517596, + 0.004885266534984112, + 0.006976110860705376, + 0.004553862847387791, + 0.00794275663793087, + 0.005293727852404118, + 0.00492697861045599, + 0.009448055177927017, + 0.005773494020104408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.10030093789100647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, rendered with surreal detail akin to macro photography, captures a small section of an old painted metal surface under the cool, diffuse light of an overcast day. The primary texture is a thick layer of faded seafoam green paint, its surface a complex map of fine cracks and blisters, revealing the deep orange-brown rust beneath. A gust of wind rusts into the frame, and a loose flake of paint, already curling at the edges, begins to tremble. The camera remains steady as the wind carefully peels the flake away from the rough, oxidized metal layer beneath, slowly and deliberately finally freeing it. Once free, the fragile flake is carried a short distance by the current, tumbling and drifting weightlessly for a brief moment, revealing its painted surface and dark underside. Then, gravity inexorably pulls it down, gently landing on the gritty gray concrete floor below. A tiny, colorful, decaying relic in the desolate industrial setting, it creates a quiet, melancholic atmosphere focused on the passage of time.", + "seed": 11045, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_4/index_4076_seed_11045_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020055977115407585, + "vlm_logprob": -0.020055977115407585, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.020055977115407585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036045245826244354 + ], + "opposite_logprob_tokens": [ + -3.6610453128814697 + ], + "logprob_reward": -0.036045245826244354, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.036045245826244354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19717994332313538, + 0.4756549596786499, + 0.05676998198032379, + 0.05552702397108078, + 0.04302191361784935, + 0.03361089527606964, + 0.04299937188625336, + 0.042444683611392975, + 0.03746170550584793, + 0.015329549089074135 + ], + "credit_map": [ + 0.01968470960855484, + 0.032899487763643265, + 0.016327280551195145, + 0.023254189640283585, + 0.034157123416662216, + 0.019114414229989052, + 0.03332936763763428, + 0.018413372337818146, + 0.0254104845225811, + 0.18423102796077728, + 0.006113870535045862, + 0.005985196214169264, + 0.031162913888692856, + 0.20024526119232178, + 0.012264116667211056, + 0.010242103599011898, + 0.005412150174379349, + 0.011310799978673458, + 0.004363984800875187, + 0.006919912528246641, + 0.0066469223238527775, + 0.009509037248790264, + 0.006638089194893837, + 0.005969086196273565, + 0.005548462737351656, + 0.007920450530946255, + 0.004400778096169233, + 0.005932142958045006, + 0.009026552550494671, + 0.0061775208450853825, + 0.011854393407702446, + 0.00466672470793128, + 0.004557312931865454, + 0.008525801822543144, + 0.0024782991968095303, + 0.0028116144239902496, + 0.005497978534549475, + 0.010038434527814388, + 0.003925906028598547, + 0.0051865652203559875, + 0.0030432906933128834, + 0.006372010335326195, + 0.0022542630322277546, + 0.003032416570931673, + 0.005442361813038588, + 0.0063647134229540825, + 0.003971300553530455, + 0.0031305395532399416, + 0.005223143845796585, + 0.00719555513933301, + 0.003838223870843649, + 0.004906611517071724, + 0.005940073169767857, + 0.007738709915429354, + 0.003915579058229923, + 0.004241479095071554, + 0.0037290409673005342, + 0.00992487370967865, + 0.0020707750227302313, + 0.003283898811787367, + 0.003736425656825304, + 0.011345408856868744, + 0.003956186585128307, + 0.004398072604089975, + 0.0031622180249542, + 0.01032297033816576, + 0.0014703749911859632, + 0.0017490945756435394, + 0.004434031434357166, + 0.01187625341117382, + 0.0024040641728788614, + 0.002042699372395873, + 0.0018897103145718575, + 0.002942862920463085, + 0.0013153697364032269, + 0.0023679533042013645, + 0.0018663424998521805, + 0.001890785526484251, + 0.0014174095122143626, + 0.0016391152748838067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.036045245826244354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017240673769265413 + ], + "opposite_logprob_tokens": [ + -6.3767242431640625 + ], + "logprob_reward": -0.0017240673769265413, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017240673769265413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0990632101893425, + 0.5339624285697937, + 0.1173209398984909, + 0.08474870771169662, + 0.02311309054493904, + 0.030244741588830948, + 0.03408951312303543, + 0.0449189767241478, + 0.019719362258911133, + 0.012819066643714905 + ], + "credit_map": [ + 0.011285434477031231, + 0.01595694199204445, + 0.007761126849800348, + 0.010903740301728249, + 0.015914198011159897, + 0.011335551738739014, + 0.015854625031352043, + 0.010051579214632511, + 0.028392748907208443, + 0.20782959461212158, + 0.006550778169184923, + 0.006460464559495449, + 0.03585892543196678, + 0.22588101029396057, + 0.012046074494719505, + 0.010942804627120495, + 0.011462302878499031, + 0.01656370609998703, + 0.014164971187710762, + 0.014864710159599781, + 0.016138751059770584, + 0.014045231975615025, + 0.015587040223181248, + 0.014494214206933975, + 0.005222490523010492, + 0.028566183522343636, + 0.00231533613987267, + 0.00317018898203969, + 0.006860863883048296, + 0.030493566766381264, + 0.004822372924536467, + 0.0032976940274238586, + 0.0024065233301371336, + 0.006139623001217842, + 0.001210907124914229, + 0.001758515601977706, + 0.002290428848937154, + 0.0063675022684037685, + 0.0017370147397741675, + 0.0012025730684399605, + 0.0025985988322645426, + 0.006962532643228769, + 0.0015472720842808485, + 0.002138089155778289, + 0.0023231254890561104, + 0.008106853812932968, + 0.0028453751001507044, + 0.00372289284132421, + 0.004842238035053015, + 0.004539693705737591, + 0.003561041783541441, + 0.004805152770131826, + 0.0038271122612059116, + 0.004439861048012972, + 0.0033400722313672304, + 0.004734339192509651, + 0.003748716786503792, + 0.008037485182285309, + 0.003143823705613613, + 0.005010352935642004, + 0.004348262213170528, + 0.010210410691797733, + 0.004035506397485733, + 0.006384414620697498, + 0.002190707251429558, + 0.004194127861410379, + 0.001187954330816865, + 0.0015361834084615111, + 0.0018271345179527998, + 0.004878361709415913, + 0.0017250344390049577, + 0.0021798552479594946, + 0.0019193554762750864, + 0.0021620094776153564, + 0.0010324546601623297, + 0.0012933412799611688, + 0.0022975008469074965, + 0.001429602736607194, + 0.0013811163371428847, + 0.0013036852469667792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0017240673769265413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536648660898209 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536648660898209, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536648660898209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06460380554199219, + 0.666499137878418, + 0.03312106430530548, + 0.1501990109682083, + 0.014157156459987164, + 0.010575379244983196, + 0.029284581542015076, + 0.011800641193985939, + 0.010477695614099503, + 0.009281555190682411 + ], + "credit_map": [ + 0.00588115444406867, + 0.010469427332282066, + 0.005795564968138933, + 0.009347253479063511, + 0.010544095188379288, + 0.006770058535039425, + 0.011013064533472061, + 0.004783186595886946, + 0.03550640866160393, + 0.25888365507125854, + 0.009398219175636768, + 0.009240026585757732, + 0.04454883188009262, + 0.279574990272522, + 0.016162512823939323, + 0.013184517621994019, + 0.0028625440318137407, + 0.006783463526517153, + 0.001891444087959826, + 0.0034605308901518583, + 0.005041148047894239, + 0.0064118788577616215, + 0.0044323233887553215, + 0.002237729961052537, + 0.011432282626628876, + 0.03508565574884415, + 0.008321263827383518, + 0.010335449129343033, + 0.016291584819555283, + 0.04159953072667122, + 0.018448537215590477, + 0.008684714324772358, + 0.001970715122297406, + 0.0023800155613571405, + 0.0011864020489156246, + 0.0015596433077007532, + 0.0020112900529056787, + 0.0016985917463898659, + 0.0017952157650142908, + 0.0015552827389910817, + 0.0012019835412502289, + 0.0018035935936495662, + 0.0009433085797354579, + 0.0008341090870089829, + 0.0017537720268592238, + 0.0014138914411887527, + 0.0011572294170036912, + 0.0014674912672489882, + 0.0026645197067409754, + 0.007929128594696522, + 0.0010027916869148612, + 0.0016790346708148718, + 0.0026053241454064846, + 0.009337351657450199, + 0.0023620985448360443, + 0.0017043324187397957, + 0.0010663974098861217, + 0.002650256035849452, + 0.0007107908604666591, + 0.0010480008786544204, + 0.0012972670374438167, + 0.0028981217183172703, + 0.0011066250735893846, + 0.001023182994686067, + 0.0008573845843784511, + 0.0023988857865333557, + 0.0007775453850626945, + 0.0010842860210686922, + 0.001043097348883748, + 0.002736683702096343, + 0.0007187490700744092, + 0.0008610638324171305, + 0.0009337469236925244, + 0.0019930454436689615, + 0.0006951640243642032, + 0.0009480643784627318, + 0.0008807253325358033, + 0.0020869295112788677, + 0.0006938746082596481, + 0.0010500053176656365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.012536648660898209 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013341344892978668 + ], + "opposite_logprob_tokens": [ + -6.626334190368652 + ], + "logprob_reward": -0.0013341344892978668, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013341344892978668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10472387820482254, + 0.6614407300949097, + 0.034198250621557236, + 0.11520567536354065, + 0.020670952275395393, + 0.009100877679884434, + 0.009539497084915638, + 0.018714508041739464, + 0.015849780291318893, + 0.010555827990174294 + ], + "credit_map": [ + 0.01108596008270979, + 0.01684477925300598, + 0.008950271643698215, + 0.01465713232755661, + 0.014082188718020916, + 0.009631660766899586, + 0.02077820710837841, + 0.00869367178529501, + 0.03517647087574005, + 0.25815320014953613, + 0.008947642520070076, + 0.008638888597488403, + 0.04367360472679138, + 0.2786521315574646, + 0.014964621514081955, + 0.013234199956059456, + 0.0030324722174555063, + 0.006947454996407032, + 0.002479498041793704, + 0.004996511619538069, + 0.003903739620000124, + 0.005620452109724283, + 0.004446438979357481, + 0.0027716830372810364, + 0.010837514884769917, + 0.020070891827344894, + 0.007067464292049408, + 0.009777862578630447, + 0.01603231579065323, + 0.023777859285473824, + 0.01887953281402588, + 0.008762229233980179, + 0.0024462018627673388, + 0.003207326866686344, + 0.0015964765334501863, + 0.0020769531838595867, + 0.0029621755238622427, + 0.002490237122401595, + 0.003165922360494733, + 0.0027256598696112633, + 0.0008633828838355839, + 0.0015647232066839933, + 0.0007523296517319977, + 0.0012264364631846547, + 0.0012885670876130462, + 0.0012692665914073586, + 0.0011199479922652245, + 0.0010162235703319311, + 0.0010263285366818309, + 0.001924628159031272, + 0.0007953886524774134, + 0.0009871742222458124, + 0.0012924394104629755, + 0.0021281603258103132, + 0.0006827032775618136, + 0.000702673860359937, + 0.0016489491099491715, + 0.0043296231888234615, + 0.001033784239552915, + 0.0012934586266055703, + 0.002920952159911394, + 0.004430848173797131, + 0.001832842594012618, + 0.0012240500655025244, + 0.0015370423207059503, + 0.0035377375315874815, + 0.0009741311077959836, + 0.0009390172781422734, + 0.002166885882616043, + 0.004231106489896774, + 0.0015020373975858092, + 0.0009618234471417964, + 0.0012282851384952664, + 0.0019685360603034496, + 0.0010007480159401894, + 0.0012109341332688928, + 0.0012454085517674685, + 0.0019167023710906506, + 0.0008849170408211648, + 0.0011002973187714815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0013341344892978668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863978922367096 + ], + "opposite_logprob_tokens": [ + -3.0486397743225098 + ], + "logprob_reward": -0.04863978922367096, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863978922367096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1635797768831253, + 0.10638222098350525, + 0.05264108255505562, + 0.17985038459300995, + 0.021430155262351036, + 0.04083802178502083, + 0.08392330259084702, + 0.06290219724178314, + 0.19932542741298676, + 0.08912742882966995 + ], + "credit_map": [ + 0.02043195813894272, + 0.030407331883907318, + 0.010844239965081215, + 0.022251026704907417, + 0.026266390457749367, + 0.016957059502601624, + 0.022242629900574684, + 0.014179147779941559, + 0.006524176336824894, + 0.036906830966472626, + 0.0032950551249086857, + 0.002794920466840267, + 0.007723493501543999, + 0.041241828352212906, + 0.004822337068617344, + 0.0030735787004232407, + 0.005357446614652872, + 0.009285814128816128, + 0.005018783267587423, + 0.00825816486030817, + 0.005686759017407894, + 0.005993266589939594, + 0.006546477787196636, + 0.0064943702891469, + 0.011336266063153744, + 0.04994901269674301, + 0.009341386146843433, + 0.010661649517714977, + 0.016979912295937538, + 0.05959976091980934, + 0.015149548649787903, + 0.006832836661487818, + 0.002334169577807188, + 0.004128395114094019, + 0.0021562783513218164, + 0.0023706720676273108, + 0.002897161291912198, + 0.0027173201087862253, + 0.002800836693495512, + 0.002025321824476123, + 0.004213347099721432, + 0.00878874585032463, + 0.0035647235345095396, + 0.004000965505838394, + 0.004959056619554758, + 0.007828309200704098, + 0.004120443016290665, + 0.0033624323550611734, + 0.007587269879877567, + 0.018577616661787033, + 0.004528361838310957, + 0.007118165493011475, + 0.008290228433907032, + 0.022101137787103653, + 0.007127995137125254, + 0.008592532016336918, + 0.004824880510568619, + 0.017770608887076378, + 0.0030001162085682154, + 0.003311936277896166, + 0.005835020914673805, + 0.020685290917754173, + 0.003534926800057292, + 0.003939413465559483, + 0.014811010099947453, + 0.05675400048494339, + 0.007598421070724726, + 0.010679788887500763, + 0.019061530008912086, + 0.06743036210536957, + 0.011472279205918312, + 0.011518042534589767, + 0.007528907153755426, + 0.023547731339931488, + 0.004019927699118853, + 0.005351613741368055, + 0.008392327465116978, + 0.027652889490127563, + 0.005470232106745243, + 0.0071637993678450584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.04863978922367096 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, rendered with surreal detail akin to macro photography, captures a small section of an old painted metal surface under the cool, diffuse light of an overcast day. The primary texture is a thick layer of faded seafoam green paint, its surface a complex map of fine cracks and blisters, revealing the deep orange-brown rust beneath. A gust of wind rusts into the frame, and a loose flake of paint, already curling at the edges, begins to tremble. The camera remains steady as the wind carefully peels the flake away from the rough, oxidized metal layer beneath, slowly and deliberately finally freeing it. Once free, the fragile flake is carried a short distance by the current, tumbling and drifting weightlessly for a brief moment, revealing its painted surface and dark underside. Then, gravity inexorably pulls it down, gently landing on the gritty gray concrete floor below. A tiny, colorful, decaying relic in the desolate industrial setting, it creates a quiet, melancholic atmosphere focused on the passage of time.", + "seed": 11046, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_4/index_4076_seed_11046_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.036635982478037475, + "vlm_logprob": -0.036635982478037475, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.036635982478037475, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0539502389729023 + ], + "opposite_logprob_tokens": [ + -3.178950309753418 + ], + "logprob_reward": -0.0539502389729023, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0539502389729023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.308115690946579, + 0.09971991926431656, + 0.059298135340213776, + 0.09016869217157364, + 0.07730220258235931, + 0.11483902484178543, + 0.08786918967962265, + 0.042818959802389145, + 0.05352991446852684, + 0.06633827090263367 + ], + "credit_map": [ + 0.050074346363544464, + 0.049557045102119446, + 0.02241152711212635, + 0.02846367284655571, + 0.056641336530447006, + 0.040295302867889404, + 0.042838890105485916, + 0.017833594232797623, + 0.017323775216937065, + 0.01943039707839489, + 0.008973310701549053, + 0.011934354901313782, + 0.009815518744289875, + 0.012541745789349079, + 0.011103903874754906, + 0.008596912026405334, + 0.008818596601486206, + 0.007458536885678768, + 0.009083328768610954, + 0.007930101826786995, + 0.004458558280020952, + 0.008038937114179134, + 0.008946927264332771, + 0.004563149064779282, + 0.017773333936929703, + 0.015863249078392982, + 0.009066070429980755, + 0.007902220822870731, + 0.012351320125162601, + 0.010550125502049923, + 0.009047995321452618, + 0.007614373695105314, + 0.013991976156830788, + 0.015142853371798992, + 0.009511158801615238, + 0.0065244645811617374, + 0.007968734949827194, + 0.012197988107800484, + 0.0057385945692658424, + 0.006226432044059038, + 0.020931318402290344, + 0.022220158949494362, + 0.017047226428985596, + 0.009095457382500172, + 0.01358330063521862, + 0.015318499878048897, + 0.0075301071628928185, + 0.009112964384257793, + 0.02012530341744423, + 0.01591417007148266, + 0.008532993495464325, + 0.0076134284026920795, + 0.010728891007602215, + 0.010555123910307884, + 0.007261591963469982, + 0.007137689273804426, + 0.0077609168365597725, + 0.007349269464612007, + 0.004361605737358332, + 0.004548604134470224, + 0.004711151123046875, + 0.006112216040492058, + 0.004111879505217075, + 0.0038633132353425026, + 0.008479791693389416, + 0.007096968591213226, + 0.007292910013347864, + 0.005842498037964106, + 0.005843719933182001, + 0.006739172153174877, + 0.00812930054962635, + 0.004105553030967712, + 0.00791147816926241, + 0.010602213442325592, + 0.009509861469268799, + 0.007882609032094479, + 0.007675059139728546, + 0.009127689525485039, + 0.008451927453279495, + 0.005177434999495745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0539502389729023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019538376946002245 + ], + "opposite_logprob_tokens": [ + -6.251953601837158 + ], + "logprob_reward": -0.0019538376946002245, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019538376946002245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44370582699775696, + 0.07304054498672485, + 0.1316768079996109, + 0.09356069564819336, + 0.04436102509498596, + 0.0494462288916111, + 0.03619784489274025, + 0.036678068339824677, + 0.05207553878426552, + 0.03925751894712448 + ], + "credit_map": [ + 0.07051821798086166, + 0.0813746303319931, + 0.046896494925022125, + 0.04218687117099762, + 0.05963201820850372, + 0.055505331605672836, + 0.056374263018369675, + 0.031217943876981735, + 0.011971525847911835, + 0.014299165457487106, + 0.007464416325092316, + 0.008316624909639359, + 0.006541500799357891, + 0.009177795611321926, + 0.007529370952397585, + 0.007740137167274952, + 0.020061366260051727, + 0.020671270787715912, + 0.023928582668304443, + 0.01142418384552002, + 0.015072137117385864, + 0.018450869247317314, + 0.012198621407151222, + 0.009869762696325779, + 0.018186969682574272, + 0.015776634216308594, + 0.008365618996322155, + 0.008571170270442963, + 0.01191664021462202, + 0.011213032528758049, + 0.00996300671249628, + 0.009567609056830406, + 0.007250356953591108, + 0.008173120208084583, + 0.005218257196247578, + 0.004068475216627121, + 0.004308090079575777, + 0.006789109669625759, + 0.004180143587291241, + 0.00437346613034606, + 0.008201190270483494, + 0.00873304158449173, + 0.006249689497053623, + 0.005018383730202913, + 0.005277623888105154, + 0.008113455958664417, + 0.0036780270747840405, + 0.004174809902906418, + 0.006058582104742527, + 0.005134711042046547, + 0.0036318718921393156, + 0.00373494322411716, + 0.004189793020486832, + 0.005154517479240894, + 0.004553525242954493, + 0.0037398994900286198, + 0.00560890045017004, + 0.005390153732150793, + 0.004811021499335766, + 0.004177280236035585, + 0.003662560135126114, + 0.004450684413313866, + 0.005167636089026928, + 0.003409824799746275, + 0.009917078539729118, + 0.00687562208622694, + 0.007182063534855843, + 0.005396563094109297, + 0.005403445102274418, + 0.006548695731908083, + 0.006359012331813574, + 0.0043930551037192345, + 0.005067263264209032, + 0.005643853917717934, + 0.004927379544824362, + 0.0045213839039206505, + 0.0043780566193163395, + 0.004831345286220312, + 0.006171592511236668, + 0.0037166415713727474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0019538376946002245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057918548583984 + ], + "opposite_logprob_tokens": [ + -3.168057918548584 + ], + "logprob_reward": -0.043057918548583984, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057918548583984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39062678813934326, + 0.07111795991659164, + 0.05809428170323372, + 0.039858609437942505, + 0.046949975192546844, + 0.08801759034395218, + 0.0859169140458107, + 0.14620454609394073, + 0.030428744852542877, + 0.042784541845321655 + ], + "credit_map": [ + 0.06885283440351486, + 0.04706082493066788, + 0.048269227147102356, + 0.03600829839706421, + 0.039515141397714615, + 0.09383951127529144, + 0.03408067673444748, + 0.023000268265604973, + 0.01316832285374403, + 0.01107251737266779, + 0.005799504462629557, + 0.0063222069293260574, + 0.008836297318339348, + 0.009384517557919025, + 0.008732589893043041, + 0.007801997009664774, + 0.00925604347139597, + 0.007031111977994442, + 0.005483873654156923, + 0.004099276382476091, + 0.006532255094498396, + 0.012243084609508514, + 0.009368134662508965, + 0.004080502782016993, + 0.007443602662533522, + 0.006308007054030895, + 0.00382670178078115, + 0.00423705019056797, + 0.004671930335462093, + 0.005837425589561462, + 0.004047285299748182, + 0.003486608387902379, + 0.007051099091768265, + 0.009618159383535385, + 0.0063498071394860744, + 0.004806642420589924, + 0.004935214761644602, + 0.00631922809407115, + 0.003568088635802269, + 0.004301737062633038, + 0.015239718370139599, + 0.015222760848701, + 0.013434230349957943, + 0.0071887243539094925, + 0.01078405138105154, + 0.013197211548686028, + 0.0060302866622805595, + 0.006920613814145327, + 0.012282734736800194, + 0.01718524470925331, + 0.008793983608484268, + 0.00794267375022173, + 0.010537604801356792, + 0.011624421924352646, + 0.00925980880856514, + 0.008290447294712067, + 0.02378448285162449, + 0.024181779474020004, + 0.014343488030135632, + 0.011977825313806534, + 0.022080862894654274, + 0.01752094179391861, + 0.019014157354831696, + 0.013301005586981773, + 0.005435135681182146, + 0.00485275499522686, + 0.0028148444835096598, + 0.0029892041347920895, + 0.002963719656690955, + 0.0037245044950395823, + 0.00451204227283597, + 0.0031365402974188328, + 0.008477558381855488, + 0.006713316310197115, + 0.004570936784148216, + 0.0038827252574265003, + 0.005014806520193815, + 0.0059967441484332085, + 0.004736239556223154, + 0.003392215585336089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.043057918548583984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249407142400742 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249407142400742, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249407142400742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.264191210269928, + 0.0786786675453186, + 0.2867606580257416, + 0.1010633334517479, + 0.04362447187304497, + 0.06581757962703705, + 0.027465295046567917, + 0.04170718044042587, + 0.0534282848238945, + 0.037263352423906326 + ], + "credit_map": [ + 0.0341242291033268, + 0.03271779790520668, + 0.03650544583797455, + 0.02758864127099514, + 0.03208720684051514, + 0.04818594455718994, + 0.03581565245985985, + 0.017166277393698692, + 0.014676275663077831, + 0.013309797272086143, + 0.008733035065233707, + 0.007941997610032558, + 0.00764503562822938, + 0.009357121773064137, + 0.009875460527837276, + 0.007139944937080145, + 0.03712005913257599, + 0.04489612951874733, + 0.04926932603120804, + 0.02666298672556877, + 0.030705852434039116, + 0.038360025733709335, + 0.03452188894152641, + 0.02522440068423748, + 0.01285439170897007, + 0.01653464138507843, + 0.00840072613209486, + 0.01227724738419056, + 0.011238280683755875, + 0.014649788849055767, + 0.015475104562938213, + 0.009633151814341545, + 0.006156398449093103, + 0.007418815512210131, + 0.00530759384855628, + 0.004888984374701977, + 0.003944649826735258, + 0.005248554050922394, + 0.00561470678076148, + 0.005044769495725632, + 0.011038071475923061, + 0.011843165382742882, + 0.010026617906987667, + 0.005165365058928728, + 0.009552162140607834, + 0.009260157123208046, + 0.0038868377450853586, + 0.0050452095456421375, + 0.004711880814284086, + 0.004454255104064941, + 0.0028209846932440996, + 0.0025311452336609364, + 0.0028930557891726494, + 0.003819847945123911, + 0.003449275391176343, + 0.0027848496101796627, + 0.007300723344087601, + 0.004290965851396322, + 0.006954196374863386, + 0.003813805291429162, + 0.00469727860763669, + 0.004968327935785055, + 0.0058321282267570496, + 0.0038497564382851124, + 0.01090759877115488, + 0.006909963674843311, + 0.008111725561320782, + 0.0050073713064193726, + 0.0060575962997972965, + 0.006112988572567701, + 0.005894443951547146, + 0.004426596686244011, + 0.007741281762719154, + 0.005339829251170158, + 0.004808174446225166, + 0.004730545915663242, + 0.0028970465064048767, + 0.005347002763301134, + 0.0029007226694375277, + 0.003498747479170561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.005249407142400742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896851003170013 + ], + "opposite_logprob_tokens": [ + -2.5789685249328613 + ], + "logprob_reward": -0.07896851003170013, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896851003170013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719241738319397, + 0.09516460448503494, + 0.08772943168878555, + 0.11688727140426636, + 0.06318018585443497, + 0.1383814960718155, + 0.15942677855491638, + 0.05626177787780762, + 0.055440351366996765, + 0.055603958666324615 + ], + "credit_map": [ + 0.032144200056791306, + 0.02282659523189068, + 0.013733376748859882, + 0.014712687581777573, + 0.019920984283089638, + 0.030329516157507896, + 0.020774513483047485, + 0.01748230867087841, + 0.017157796770334244, + 0.018199129030108452, + 0.01161237433552742, + 0.009864077903330326, + 0.01134943962097168, + 0.010500790551304817, + 0.009133750572800636, + 0.007347246166318655, + 0.01304666418582201, + 0.013339869678020477, + 0.01416196208447218, + 0.007678174413740635, + 0.00943463109433651, + 0.014197396114468575, + 0.009674739092588425, + 0.006195995956659317, + 0.020265908911824226, + 0.019371958449482918, + 0.015547515824437141, + 0.00915298331528902, + 0.012260423973202705, + 0.015322825871407986, + 0.01114153303205967, + 0.013824112713336945, + 0.011288754642009735, + 0.009819702245295048, + 0.007374025881290436, + 0.007306346204131842, + 0.006207044702023268, + 0.008760386146605015, + 0.006223456002771854, + 0.006200469098985195, + 0.024360625073313713, + 0.02365480363368988, + 0.02134726382791996, + 0.012162016704678535, + 0.016827009618282318, + 0.01999848149716854, + 0.009685736149549484, + 0.010345550253987312, + 0.024214619770646095, + 0.03328607231378555, + 0.021878844127058983, + 0.012439187616109848, + 0.022443002089858055, + 0.018900742754340172, + 0.011297933757305145, + 0.014966374263167381, + 0.008941736072301865, + 0.008460933342576027, + 0.007188057992607355, + 0.005925305187702179, + 0.006330540403723717, + 0.007710826117545366, + 0.006914070807397366, + 0.004790309350937605, + 0.012021362781524658, + 0.008207103237509727, + 0.008607063442468643, + 0.004910726100206375, + 0.005773913115262985, + 0.006289695389568806, + 0.005187173839658499, + 0.004443315789103508, + 0.009703347459435463, + 0.00872571486979723, + 0.007674772292375565, + 0.006710167974233627, + 0.005537756718695164, + 0.007335359696298838, + 0.004970810376107693, + 0.004946031142026186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.07896851003170013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, rendered with surreal detail akin to macro photography, captures a small section of an old painted metal surface under the cool, diffuse light of an overcast day. The primary texture is a thick layer of faded seafoam green paint, its surface a complex map of fine cracks and blisters, revealing the deep orange-brown rust beneath. A gust of wind rusts into the frame, and a loose flake of paint, already curling at the edges, begins to tremble. The camera remains steady as the wind carefully peels the flake away from the rough, oxidized metal layer beneath, slowly and deliberately finally freeing it. Once free, the fragile flake is carried a short distance by the current, tumbling and drifting weightlessly for a brief moment, revealing its painted surface and dark underside. Then, gravity inexorably pulls it down, gently landing on the gritty gray concrete floor below. A tiny, colorful, decaying relic in the desolate industrial setting, it creates a quiet, melancholic atmosphere focused on the passage of time.", + "seed": 11047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_4/index_4076_seed_11047_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027710998663678765, + "vlm_logprob": -0.027710998663678765, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.027710998663678765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02200252003967762 + ], + "opposite_logprob_tokens": [ + -4.147002696990967 + ], + "logprob_reward": -0.02200252003967762, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02200252003967762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3928990364074707, + 0.12837237119674683, + 0.12634560465812683, + 0.0416020043194294, + 0.053685568273067474, + 0.13419580459594727, + 0.03943761810660362, + 0.042168255895376205, + 0.019558703526854515, + 0.0217350572347641 + ], + "credit_map": [ + 0.04455694183707237, + 0.04731220752000809, + 0.037183552980422974, + 0.037938982248306274, + 0.08058369904756546, + 0.07138356566429138, + 0.04548276215791702, + 0.028457336127758026, + 0.014775627292692661, + 0.020250611007213593, + 0.009428798221051693, + 0.010053110308945179, + 0.036656830459833145, + 0.01575080119073391, + 0.01367280911654234, + 0.007783783599734306, + 0.015056195668876171, + 0.014729449525475502, + 0.01163617055863142, + 0.01401432417333126, + 0.01663454994559288, + 0.013488737866282463, + 0.027295874431729317, + 0.013490301556885242, + 0.0051233465783298016, + 0.00594119168817997, + 0.003896344918757677, + 0.004004075191915035, + 0.0052083502523601055, + 0.0063320002518594265, + 0.00619411701336503, + 0.00490257702767849, + 0.007589970249682665, + 0.007433445658534765, + 0.005644978489726782, + 0.00508182542398572, + 0.009906511753797531, + 0.007716177962720394, + 0.00616558687761426, + 0.004147072322666645, + 0.014644237235188484, + 0.020047688856720924, + 0.011016948148608208, + 0.012480929493904114, + 0.023128651082515717, + 0.022609300911426544, + 0.019202208146452904, + 0.011065822094678879, + 0.004192161839455366, + 0.0059052505530416965, + 0.0028512037824839354, + 0.004031138028949499, + 0.006165540777146816, + 0.006144956685602665, + 0.006609817501157522, + 0.0035375526640564203, + 0.004483039956539869, + 0.007293280214071274, + 0.003523287596181035, + 0.0037513505667448044, + 0.007695913780480623, + 0.00556532945483923, + 0.006001030094921589, + 0.003855021670460701, + 0.0023190926294773817, + 0.003038762602955103, + 0.0017330580158159137, + 0.001697878702543676, + 0.0034249427262693644, + 0.0023519936949014664, + 0.0029917368665337563, + 0.0020012359600514174, + 0.0023626191541552544, + 0.0025281275156885386, + 0.0027023630682379007, + 0.001932060346007347, + 0.0038572733756154776, + 0.0037845775950700045, + 0.0027303295210003853, + 0.0018377063097432256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02200252003967762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010530170984566212 + ], + "opposite_logprob_tokens": [ + -6.8760528564453125 + ], + "logprob_reward": -0.0010530170984566212, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010530170984566212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43123868107795715, + 0.1658996045589447, + 0.09082784503698349, + 0.038570553064346313, + 0.05270172655582428, + 0.03964414820075035, + 0.030661839991807938, + 0.0659961849451065, + 0.040415845811367035, + 0.04404361918568611 + ], + "credit_map": [ + 0.03071351908147335, + 0.09359817951917648, + 0.029797345399856567, + 0.03542784973978996, + 0.05595926567912102, + 0.12542293965816498, + 0.03386753052473068, + 0.02645201049745083, + 0.019280817359685898, + 0.027230774983763695, + 0.013394669629633427, + 0.01666918769478798, + 0.03484194725751877, + 0.02271013706922531, + 0.01454442273825407, + 0.01722763292491436, + 0.008835616521537304, + 0.01169884018599987, + 0.0066839479841291904, + 0.007873854599893093, + 0.01448990311473608, + 0.019786955788731575, + 0.01439967006444931, + 0.007059060502797365, + 0.004580806940793991, + 0.004957128316164017, + 0.0033895475789904594, + 0.0035384169314056635, + 0.007151283323764801, + 0.007864344865083694, + 0.004254793282598257, + 0.0028342322912067175, + 0.008106117136776447, + 0.0071172090247273445, + 0.00577503303065896, + 0.004774425178766251, + 0.010041226632893085, + 0.00663762865588069, + 0.005495048128068447, + 0.004755035042762756, + 0.005040045361965895, + 0.006758037488907576, + 0.0036548501811921597, + 0.004196835681796074, + 0.00579786254093051, + 0.004859535489231348, + 0.0053625949658453465, + 0.00397438695654273, + 0.0037784259766340256, + 0.004590840544551611, + 0.0022375949192792177, + 0.0029186983592808247, + 0.005498854443430901, + 0.004743543919175863, + 0.004229305312037468, + 0.002664576517418027, + 0.00559956906363368, + 0.008747830055654049, + 0.005909190513193607, + 0.006076037883758545, + 0.010253572836518288, + 0.014138661324977875, + 0.010323305614292622, + 0.004948014859110117, + 0.004440987017005682, + 0.006960901897400618, + 0.004337318241596222, + 0.0034011490643024445, + 0.004769173916429281, + 0.008685491047799587, + 0.005178363993763924, + 0.0026424594689160585, + 0.004731704946607351, + 0.007179458159953356, + 0.0041551110334694386, + 0.0037067807279527187, + 0.005776720121502876, + 0.009513897821307182, + 0.005931233521550894, + 0.0030487116891890764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0010530170984566212 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282134905457497 + ], + "opposite_logprob_tokens": [ + -3.773282051086426 + ], + "logprob_reward": -0.023282134905457497, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282134905457497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2872852683067322, + 0.11815799027681351, + 0.23572540283203125, + 0.06076027825474739, + 0.06843816488981247, + 0.06223665922880173, + 0.03014872968196869, + 0.040543120354413986, + 0.06077626347541809, + 0.03592810407280922 + ], + "credit_map": [ + 0.0336168073117733, + 0.045264482498168945, + 0.024431800469756126, + 0.02115841768682003, + 0.03640148043632507, + 0.07294522225856781, + 0.033811282366514206, + 0.019655784592032433, + 0.01449360977858305, + 0.020140046253800392, + 0.010305149480700493, + 0.008418235927820206, + 0.014892379753291607, + 0.023552389815449715, + 0.013655099086463451, + 0.012701083905994892, + 0.02561749331653118, + 0.030442478135228157, + 0.01842879131436348, + 0.025211703032255173, + 0.039560310542583466, + 0.028907177969813347, + 0.04870012030005455, + 0.0188573207706213, + 0.004671131260693073, + 0.009150825440883636, + 0.003930248320102692, + 0.0034911867696791887, + 0.011376646347343922, + 0.016216527670621872, + 0.007015936076641083, + 0.004907776135951281, + 0.00752929737791419, + 0.011870103888213634, + 0.005856161471456289, + 0.007087505888193846, + 0.012419566512107849, + 0.008887858130037785, + 0.00797206349670887, + 0.006815608125180006, + 0.00685248663648963, + 0.00883905403316021, + 0.007043382618576288, + 0.006373737473040819, + 0.009430507197976112, + 0.006961555685847998, + 0.010495799593627453, + 0.0062401373870670795, + 0.002303829649463296, + 0.003481136402115226, + 0.0024878326803445816, + 0.0026511356700211763, + 0.004635735414922237, + 0.005833959206938744, + 0.0058560119941830635, + 0.002899089828133583, + 0.004857697989791632, + 0.004991335794329643, + 0.0049694678746163845, + 0.005219560582190752, + 0.005157955456525087, + 0.005304424092173576, + 0.007083762437105179, + 0.0029589177574962378, + 0.006799046881496906, + 0.009194102138280869, + 0.005530652590095997, + 0.005518888123333454, + 0.011304164305329323, + 0.010166672989726067, + 0.0075034815818071365, + 0.00475925300270319, + 0.003913772292435169, + 0.00609120586887002, + 0.0039177569560706615, + 0.0030055746901780367, + 0.007355970796197653, + 0.00541260838508606, + 0.003937198780477047, + 0.0022940137423574924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.023282134905457497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190072951838374 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003190072951838374, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190072951838374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34007614850997925, + 0.19095999002456665, + 0.15827925503253937, + 0.055540699511766434, + 0.039586134254932404, + 0.04219602793455124, + 0.02688712440431118, + 0.05850056931376457, + 0.035203371196985245, + 0.05277062952518463 + ], + "credit_map": [ + 0.0444922000169754, + 0.040892381221055984, + 0.03906969353556633, + 0.03053026832640171, + 0.05385127663612366, + 0.04965787008404732, + 0.05106588825583458, + 0.030516617000102997, + 0.02855043299496174, + 0.027367061004042625, + 0.016091668978333473, + 0.021459253504872322, + 0.03294004499912262, + 0.02242117188870907, + 0.02433806098997593, + 0.017792312428355217, + 0.018555007874965668, + 0.017377836629748344, + 0.013006305322051048, + 0.013915915973484516, + 0.028671948239207268, + 0.029071899130940437, + 0.023663310334086418, + 0.014017055742442608, + 0.005674438551068306, + 0.00687439413741231, + 0.007186915725469589, + 0.005020756274461746, + 0.008253191597759724, + 0.00783634465187788, + 0.008124835789203644, + 0.006569825578480959, + 0.004219070542603731, + 0.005849613342434168, + 0.0032261942978948355, + 0.003687627613544464, + 0.006519812159240246, + 0.0054865130223333836, + 0.007663488853722811, + 0.0029338200110942125, + 0.005217344034463167, + 0.0065738847479224205, + 0.0039628189988434315, + 0.0037125032395124435, + 0.004918170161545277, + 0.006418504286557436, + 0.006466452497988939, + 0.004926353693008423, + 0.002583986846730113, + 0.0041078366339206696, + 0.0019305365858599544, + 0.0014688188675791025, + 0.004917018581181765, + 0.007587709464132786, + 0.0025227279402315617, + 0.0017684911144897342, + 0.004993025679141283, + 0.006814050022512674, + 0.007121330592781305, + 0.005993180442601442, + 0.007588334381580353, + 0.010130359791219234, + 0.010491653345525265, + 0.0053686415776610374, + 0.003736251499503851, + 0.00450343731790781, + 0.004390337970107794, + 0.004255963023751974, + 0.005257729906588793, + 0.0038295139092952013, + 0.006611458491533995, + 0.0026186837349087, + 0.005944333504885435, + 0.009165635332465172, + 0.006405453197658062, + 0.00537008885294199, + 0.008380835875868797, + 0.005733146797865629, + 0.007727526128292084, + 0.004043613560497761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.003190072951838374 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an old painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paint primarily seafoam green with visible cracks, blisters, and deep orange-brown rust underneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a loose flake of paint tremble and peel away from the surface due to a gust of wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the detached paint flake drift weightlessly and then land on a gritty gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static and steady throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902724832296371 + ], + "opposite_logprob_tokens": [ + -2.464027166366577 + ], + "logprob_reward": -0.08902724832296371, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902724832296371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41539323329925537, + 0.10606301575899124, + 0.0890083834528923, + 0.03681548684835434, + 0.07424883544445038, + 0.04402559623122215, + 0.05317035689949989, + 0.07692725956439972, + 0.05258436128497124, + 0.051763471215963364 + ], + "credit_map": [ + 0.04706305265426636, + 0.057473957538604736, + 0.05690619349479675, + 0.049431212246418, + 0.05921221151947975, + 0.04654868692159653, + 0.04981357231736183, + 0.048944324254989624, + 0.01035352610051632, + 0.018625589087605476, + 0.00698376027867198, + 0.009609050117433071, + 0.018648521974682808, + 0.016487522050738335, + 0.012796369381248951, + 0.01255867537111044, + 0.011646362021565437, + 0.016718629747629166, + 0.0074835228733718395, + 0.008547923527657986, + 0.013113641180098057, + 0.013213617727160454, + 0.010522096417844296, + 0.007762589957565069, + 0.004842965863645077, + 0.004788682796061039, + 0.003472735406830907, + 0.0033040002454072237, + 0.005443934351205826, + 0.006892520934343338, + 0.004286572802811861, + 0.003784075379371643, + 0.005949012935161591, + 0.010781842283904552, + 0.005052671767771244, + 0.005504986736923456, + 0.02028278447687626, + 0.012154664844274521, + 0.007975232787430286, + 0.00654763774946332, + 0.005128867458552122, + 0.007352273445576429, + 0.0045068152248859406, + 0.004399942699819803, + 0.007250676397234201, + 0.006949807051569223, + 0.004814481362700462, + 0.0036227309610694647, + 0.004168184939771891, + 0.007444380782544613, + 0.0037296945229172707, + 0.005316257942467928, + 0.008699476718902588, + 0.011395499110221863, + 0.008915536105632782, + 0.0035013232845813036, + 0.009074022062122822, + 0.01028426457196474, + 0.009004258550703526, + 0.008231621235609055, + 0.013602523133158684, + 0.010526064783334732, + 0.009969213046133518, + 0.006235290784388781, + 0.0056168679147958755, + 0.006546138320118189, + 0.0059080165810883045, + 0.006253980565816164, + 0.008684570901095867, + 0.005684667266905308, + 0.00926733948290348, + 0.004622779320925474, + 0.005132048390805721, + 0.007498929742723703, + 0.004920712672173977, + 0.0055465782061219215, + 0.009078663773834705, + 0.005873311311006546, + 0.00938649382442236, + 0.004326733760535717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08902724832296371 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_5/index_tensor([7104, 7104])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_5/index_tensor([7104, 7104])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e77c9337e9e8a820b2aa5d75aa43940016de0e7c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_5/index_tensor([7104, 7104])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static low-angle medium close-up freezes the climactic moment of the action, rendered in a dramatic comic style that mimics a static image sequence. The protagonist is a formidable warrior, his body a coil of primal energy, clad in ornate silver armor that gleams with a metallic sheen. The armor is intricately detailed, with a sharp shield and carved gorget, its polished surface reflecting the chaos of battle. His face, contorted into a mask of pure rage, is the focal point; his eyes are wide, pupils constricted in a wild gaze, while his teeth are clenched so tightly that the muscles in his jaw appear distinct. He is caught in the act, thrusting a powerful spear directly at the viewer. As he drives his weapon forward, his arm blurs in motion, muscles bulging beneath the straps. A dense mass of sharp black and white dynamic speed lines radiating from the harpoon tip and his body visually convey the intense movement—a classic comic technique that conveys immense speed and impact. The scene is drenched in high-contrast, dramatic lighting, creating deep, inky shadows that sculpt his form and emphasize the raw power of his frantic battle against a stark abstract background of explosive energy, creating an intense and violent atmosphere.", + "seed": 21268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_5/index_7104_seed_21268_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028450630232691766, + "vlm_logprob": -0.028450630232691766, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.028450630232691766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015384926460683346 + ], + "opposite_logprob_tokens": [ + -4.390385150909424 + ], + "logprob_reward": -0.015384926460683346, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015384926460683346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15360841155052185, + 0.13068996369838715, + 0.08796744793653488, + 0.07196610420942307, + 0.24120314419269562, + 0.032048024237155914, + 0.07919348031282425, + 0.05077483877539635, + 0.023061582818627357, + 0.12948699295520782 + ], + "credit_map": [ + 0.016168778762221336, + 0.01221503410488367, + 0.03044968843460083, + 0.02073664404451847, + 0.015599123202264309, + 0.020343922078609467, + 0.020706037059426308, + 0.01738918386399746, + 0.015688618645071983, + 0.013797792606055737, + 0.01674446277320385, + 0.01274999137967825, + 0.02269386313855648, + 0.01907612755894661, + 0.017008136957883835, + 0.012930967845022678, + 0.010111140087246895, + 0.00924187432974577, + 0.013125085271894932, + 0.014571454375982285, + 0.012078235857188702, + 0.006495303008705378, + 0.015554618090391159, + 0.006789740175008774, + 0.0075032371096313, + 0.006351700518280268, + 0.008740072138607502, + 0.006177229806780815, + 0.013617385178804398, + 0.009317927062511444, + 0.010637246072292328, + 0.00962131004780531, + 0.08967767655849457, + 0.009655377827584743, + 0.09803322702646255, + 0.010647892951965332, + 0.011182119138538837, + 0.008531076833605766, + 0.007514071185141802, + 0.0059617082588374615, + 0.00506866117939353, + 0.0032193572260439396, + 0.005655129440128803, + 0.00463438592851162, + 0.003645518096163869, + 0.0031501140911132097, + 0.003981126472353935, + 0.002693730406463146, + 0.0155186727643013, + 0.006632164120674133, + 0.01643742248415947, + 0.008944319561123848, + 0.010105031542479992, + 0.006123753730207682, + 0.009601067751646042, + 0.005831039045006037, + 0.008509118109941483, + 0.006055211182683706, + 0.009157151915133, + 0.0060197459533810616, + 0.005074442829936743, + 0.00493997149169445, + 0.005977611988782883, + 0.005041585303843021, + 0.002251060213893652, + 0.002303413348272443, + 0.0031608224380761385, + 0.0035201760474592447, + 0.0026963206473737955, + 0.0029106521978974342, + 0.003759029321372509, + 0.002460111165419221, + 0.016295740380883217, + 0.010943141765892506, + 0.02414594031870365, + 0.016885390505194664, + 0.01766039803624153, + 0.011126688681542873, + 0.020163467153906822, + 0.012266227975487709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.015384926460683346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495136231184006 + ], + "opposite_logprob_tokens": [ + -2.9299514293670654 + ], + "logprob_reward": -0.05495136231184006, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495136231184006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14153826236724854, + 0.10869473963975906, + 0.0830315500497818, + 0.03372085094451904, + 0.03772459551692009, + 0.017017798498272896, + 0.4319505989551544, + 0.06170758232474327, + 0.03359448164701462, + 0.051019515842199326 + ], + "credit_map": [ + 0.015281177125871181, + 0.014104876667261124, + 0.026653306558728218, + 0.01781347021460533, + 0.01732964627444744, + 0.019130736589431763, + 0.018308645114302635, + 0.012916401028633118, + 0.013834211975336075, + 0.010469567961990833, + 0.01768611930310726, + 0.009546088986098766, + 0.014211759902536869, + 0.013179801404476166, + 0.015797695145010948, + 0.01396949589252472, + 0.010273624211549759, + 0.009021295234560966, + 0.013475077226758003, + 0.010445941239595413, + 0.010150427930057049, + 0.0071341609582304955, + 0.0154623594135046, + 0.0070686680264770985, + 0.003819095203652978, + 0.0032360153272747993, + 0.004106557462364435, + 0.0028245793655514717, + 0.005802525673061609, + 0.00371760712005198, + 0.006158773321658373, + 0.004055697005242109, + 0.0051767295226454735, + 0.004267290700227022, + 0.005088366102427244, + 0.0036987238563597202, + 0.006505432538688183, + 0.004575678147375584, + 0.004174809902906418, + 0.004237566143274307, + 0.0017708102241158485, + 0.0016768639907240868, + 0.0019636941142380238, + 0.002932993695139885, + 0.002878520404919982, + 0.0015748190926387906, + 0.0024575339630246162, + 0.0017625640612095594, + 0.17163747549057007, + 0.012431793846189976, + 0.1973002403974533, + 0.015714308246970177, + 0.0077743506990373135, + 0.01356426440179348, + 0.007670106831938028, + 0.005858082324266434, + 0.009505810216069221, + 0.005898412317037582, + 0.01272493228316307, + 0.007948162034153938, + 0.007077387999743223, + 0.004996468313038349, + 0.008858428336679935, + 0.004697982221841812, + 0.004603422246873379, + 0.004116044379770756, + 0.006045293062925339, + 0.004670822061598301, + 0.004070260562002659, + 0.00296539138071239, + 0.004286456853151321, + 0.00283679086714983, + 0.007804470136761665, + 0.004875510465353727, + 0.011874089948832989, + 0.007284396328032017, + 0.005730535369366407, + 0.004390872083604336, + 0.004897594451904297, + 0.0041620503179728985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.05495136231184006 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202811747789383 + ], + "opposite_logprob_tokens": [ + -2.812028169631958 + ], + "logprob_reward": -0.06202811747789383, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202811747789383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05011478066444397, + 0.01461669709533453, + 0.012095102109014988, + 0.007241688668727875, + 0.018337741494178772, + 0.013929586857557297, + 0.8413889408111572, + 0.019870374351739883, + 0.011155704036355019, + 0.01124931126832962 + ], + "credit_map": [ + 0.007193830795586109, + 0.005162887275218964, + 0.012341075576841831, + 0.00669866194948554, + 0.00520880613476038, + 0.004430426750332117, + 0.005619978532195091, + 0.003459118539467454, + 0.0015024368185549974, + 0.0010854232823476195, + 0.001870763604529202, + 0.0014723028289154172, + 0.0019533175509423018, + 0.001550903427414596, + 0.003091533901169896, + 0.002090016845613718, + 0.0015047856140881777, + 0.0011766074458137155, + 0.0016280593117699027, + 0.0013192316982895136, + 0.001781014259904623, + 0.0010926889954134822, + 0.002331286435946822, + 0.001261430443264544, + 0.000990276224911213, + 0.0006226555560715497, + 0.0010736646363511682, + 0.0007186457514762878, + 0.000959582335781306, + 0.000985371763817966, + 0.001099795219488442, + 0.0007916979375295341, + 0.0056978678330779076, + 0.0009309577289968729, + 0.00589753920212388, + 0.0011142479488626122, + 0.0016816495917737484, + 0.0008984961314126849, + 0.0010571929160505533, + 0.0010597920045256615, + 0.0022821971215307713, + 0.0016198634402826428, + 0.0020936569198966026, + 0.0026246882043778896, + 0.0016598176443949342, + 0.001187176676467061, + 0.0014310632832348347, + 0.0010311255464330316, + 0.3357946574687958, + 0.024105597287416458, + 0.38761377334594727, + 0.029700908809900284, + 0.012750163674354553, + 0.027643032371997833, + 0.013646670617163181, + 0.010134260170161724, + 0.0030420476105064154, + 0.001712628174573183, + 0.004058136139065027, + 0.0027646836824715137, + 0.0021850760094821453, + 0.0021856636740267277, + 0.001941127935424447, + 0.0019810127560049295, + 0.0013750599464401603, + 0.0010546103585511446, + 0.0021022965665906668, + 0.001362703973427415, + 0.0015129693783819675, + 0.0011492626508697867, + 0.0015421832213178277, + 0.001056618639267981, + 0.0014470887836068869, + 0.0016864276258274913, + 0.002076150383800268, + 0.0014153766678646207, + 0.0016698697581887245, + 0.0010808539809659123, + 0.0009161466732621193, + 0.0009573989664204419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.06202811747789383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249407142400742 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249407142400742, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249407142400742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12222136557102203, + 0.038380008190870285, + 0.18715828657150269, + 0.02387980744242668, + 0.038529492914676666, + 0.04334654286503792, + 0.34345781803131104, + 0.06245777755975723, + 0.06848542392253876, + 0.07208346575498581 + ], + "credit_map": [ + 0.015525849536061287, + 0.010801832191646099, + 0.028513729572296143, + 0.02376582846045494, + 0.013000532984733582, + 0.013004861772060394, + 0.010452911257743835, + 0.007155818399041891, + 0.003571741282939911, + 0.0031890287064015865, + 0.005250523332506418, + 0.003939157817512751, + 0.0049318671226501465, + 0.004126191604882479, + 0.007546850014477968, + 0.005824650637805462, + 0.025929773226380348, + 0.020403554663062096, + 0.03312750533223152, + 0.02974732778966427, + 0.01690658926963806, + 0.017093390226364136, + 0.02917267195880413, + 0.014777475036680698, + 0.0027847758028656244, + 0.0023039262741804123, + 0.005544931162148714, + 0.0024206137750297785, + 0.002852968405932188, + 0.002496695378795266, + 0.0029741041362285614, + 0.0025017934385687113, + 0.011626466177403927, + 0.002370433881878853, + 0.013066022656857967, + 0.0022617015056312084, + 0.0024790174793452024, + 0.0021868939511477947, + 0.0023312573321163654, + 0.002207699231803417, + 0.005419786088168621, + 0.006093516014516354, + 0.006752231623977423, + 0.008659634739160538, + 0.005151599179953337, + 0.0037229391746222973, + 0.0037622260861098766, + 0.0037846099585294724, + 0.1332377940416336, + 0.010237382724881172, + 0.1521383374929428, + 0.013779236935079098, + 0.007900405675172806, + 0.011047434993088245, + 0.008998892270028591, + 0.006118351593613625, + 0.009752839803695679, + 0.008459335193037987, + 0.009982939809560776, + 0.008354425430297852, + 0.007229953072965145, + 0.006413518451154232, + 0.006344923283904791, + 0.005919842980802059, + 0.012532581575214863, + 0.0066434829495847225, + 0.01687270775437355, + 0.00652653630822897, + 0.0075200581923127174, + 0.005439715925604105, + 0.007534577511250973, + 0.00541576137766242, + 0.010836787521839142, + 0.007238209713250399, + 0.014905850403010845, + 0.011144650168716908, + 0.007948660291731358, + 0.007164939306676388, + 0.006590635050088167, + 0.00625373050570488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.005249407142400742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463933777064085 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463933777064085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463933777064085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015274430625140667, + 0.01222939696162939, + 0.010712138377130032, + 0.005240131169557571, + 0.01334155909717083, + 0.007844979874789715, + 0.8998563885688782, + 0.006943816319108009, + 0.008822817355394363, + 0.019734349101781845 + ], + "credit_map": [ + 0.0020152681972831488, + 0.0018871426582336426, + 0.002922748913988471, + 0.002183542586863041, + 0.0016547326231375337, + 0.0018518882570788264, + 0.0017010887386277318, + 0.0010580196976661682, + 0.0014750304399058223, + 0.0016328393248841166, + 0.0016554314643144608, + 0.0011398590868338943, + 0.0013134714681655169, + 0.0011771608842536807, + 0.0021525537595152855, + 0.0016830514650791883, + 0.001296670874580741, + 0.0013491892023012042, + 0.0014753257855772972, + 0.0018019862473011017, + 0.0013254444347694516, + 0.0010143521940335631, + 0.001508895424194634, + 0.0009402752039022744, + 0.000791999336797744, + 0.0005790961440652609, + 0.0009126645163632929, + 0.0005521484999917448, + 0.0007167243165895343, + 0.0005033899215050042, + 0.0006890601362101734, + 0.0004950488219037652, + 0.004563546273857355, + 0.000752992054913193, + 0.004959296900779009, + 0.0007133902981877327, + 0.0005262860213406384, + 0.000545020098797977, + 0.0007247944013215601, + 0.0005562332225963473, + 0.00168038927949965, + 0.0007426482043229043, + 0.0019214600324630737, + 0.0007894812151789665, + 0.0007443735958077013, + 0.0006233624881133437, + 0.0007046703831292689, + 0.0006385950255207717, + 0.3589201867580414, + 0.025689564645290375, + 0.4139653146266937, + 0.0319938026368618, + 0.013906358741223812, + 0.02920662611722946, + 0.014867027290165424, + 0.011307571083307266, + 0.0010661028791218996, + 0.0008779328200034797, + 0.001039964030496776, + 0.0009369314066134393, + 0.0007856516749598086, + 0.0007454596343450248, + 0.0008715793956071138, + 0.0006201948854140937, + 0.0019868973176926374, + 0.0007021310739219189, + 0.002722569042816758, + 0.0007322427700273693, + 0.0006151524721644819, + 0.0005602440214715898, + 0.0008893072954379022, + 0.0006142739439383149, + 0.0027167578227818012, + 0.0018638675101101398, + 0.003644253360107541, + 0.0034574500750750303, + 0.002641839673742652, + 0.0020059063099324703, + 0.0018579548923298717, + 0.0015463205054402351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.00463933777064085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static low-angle medium close-up freezes the climactic moment of the action, rendered in a dramatic comic style that mimics a static image sequence. The protagonist is a formidable warrior, his body a coil of primal energy, clad in ornate silver armor that gleams with a metallic sheen. The armor is intricately detailed, with a sharp shield and carved gorget, its polished surface reflecting the chaos of battle. His face, contorted into a mask of pure rage, is the focal point; his eyes are wide, pupils constricted in a wild gaze, while his teeth are clenched so tightly that the muscles in his jaw appear distinct. He is caught in the act, thrusting a powerful spear directly at the viewer. As he drives his weapon forward, his arm blurs in motion, muscles bulging beneath the straps. A dense mass of sharp black and white dynamic speed lines radiating from the harpoon tip and his body visually convey the intense movement—a classic comic technique that conveys immense speed and impact. The scene is drenched in high-contrast, dramatic lighting, creating deep, inky shadows that sculpt his form and emphasize the raw power of his frantic battle against a stark abstract background of explosive energy, creating an intense and violent atmosphere.", + "seed": 21269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_5/index_7104_seed_21269_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026598057709634304, + "vlm_logprob": -0.026598057709634304, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.026598057709634304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024215754121541977 + ], + "opposite_logprob_tokens": [ + -3.8992156982421875 + ], + "logprob_reward": -0.024215754121541977, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024215754121541977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2234896570444107, + 0.04949700087308884, + 0.031090889126062393, + 0.030403172597289085, + 0.3499751389026642, + 0.029624704271554947, + 0.028545072302222252, + 0.021789634600281715, + 0.07730485498905182, + 0.1582798808813095 + ], + "credit_map": [ + 0.011238968931138515, + 0.019451119005680084, + 0.009227086789906025, + 0.013639271259307861, + 0.08389006555080414, + 0.01455750036984682, + 0.023169340565800667, + 0.048316292464733124, + 0.004211081191897392, + 0.009096546098589897, + 0.006681568454951048, + 0.0041290647350251675, + 0.005916486494243145, + 0.005026622675359249, + 0.006435371935367584, + 0.008000251837074757, + 0.004921404179185629, + 0.006662487052381039, + 0.0026440750807523727, + 0.0041147442534565926, + 0.0035971319302916527, + 0.0032033375464379787, + 0.002152713481336832, + 0.0037949939724057913, + 0.004789026454091072, + 0.004563507623970509, + 0.0027222775388509035, + 0.0038000973872840405, + 0.004803760442882776, + 0.0031251783948391676, + 0.003459303406998515, + 0.003140021115541458, + 0.017312321811914444, + 0.019051063805818558, + 0.008231712505221367, + 0.034187834709882736, + 0.01927165873348713, + 0.11892934143543243, + 0.013608936220407486, + 0.11938226968050003, + 0.0029030716978013515, + 0.0052123842760920525, + 0.0022001529578119516, + 0.003341061994433403, + 0.004442648496478796, + 0.0027190165128558874, + 0.004741415847092867, + 0.004064949695020914, + 0.004695387091487646, + 0.0032972092740237713, + 0.0023090329486876726, + 0.006824228446930647, + 0.0031114022713154554, + 0.0023068315349519253, + 0.001970689743757248, + 0.004030289594084024, + 0.0018969456432387233, + 0.0034957497846335173, + 0.0032302967738360167, + 0.0024143466725945473, + 0.002031245268881321, + 0.0017259042942896485, + 0.003660573624074459, + 0.003334569977596402, + 0.006999485194683075, + 0.007227678783237934, + 0.00977250188589096, + 0.012928513810038567, + 0.007822321727871895, + 0.010881412774324417, + 0.009621654637157917, + 0.012051278725266457, + 0.01550903171300888, + 0.01398527342826128, + 0.03178023919463158, + 0.01786634884774685, + 0.01906007155776024, + 0.012559155933558941, + 0.030511222779750824, + 0.017008516937494278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.024215754121541977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495418235659599 + ], + "opposite_logprob_tokens": [ + -2.9299542903900146 + ], + "logprob_reward": -0.05495418235659599, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495418235659599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29392924904823303, + 0.09887105971574783, + 0.04791484773159027, + 0.04681571573019028, + 0.16852356493473053, + 0.028672294691205025, + 0.0843302384018898, + 0.029074663296341896, + 0.11829347908496857, + 0.08357486128807068 + ], + "credit_map": [ + 0.018509363755583763, + 0.024001821875572205, + 0.017617151141166687, + 0.018738407641649246, + 0.1106773242354393, + 0.019427843391895294, + 0.02759244479238987, + 0.05736492946743965, + 0.007455980870872736, + 0.015878906473517418, + 0.011294004507362843, + 0.007688378449529409, + 0.02113211899995804, + 0.016239473596215248, + 0.00849098339676857, + 0.010691212490200996, + 0.007716622203588486, + 0.007864922285079956, + 0.004473983775824308, + 0.006517800968140364, + 0.007458690088242292, + 0.0037587720435112715, + 0.004889348987489939, + 0.005234707612544298, + 0.006876248866319656, + 0.006311697419732809, + 0.00415864447131753, + 0.00690863374620676, + 0.007526697125285864, + 0.004714487120509148, + 0.005480641033500433, + 0.00483867060393095, + 0.010483263991773129, + 0.011142401956021786, + 0.007167357485741377, + 0.031204793602228165, + 0.010396256111562252, + 0.03590598329901695, + 0.009129091165959835, + 0.053094420582056046, + 0.003472659969702363, + 0.0033948724158108234, + 0.0022862369660288095, + 0.003042737953364849, + 0.006379502359777689, + 0.003435770282521844, + 0.003808637848123908, + 0.002851880621165037, + 0.0066686104983091354, + 0.029584182426333427, + 0.003583796788007021, + 0.007154169958084822, + 0.006070462055504322, + 0.021096177399158478, + 0.004763380624353886, + 0.00540946377441287, + 0.0027851213235408068, + 0.004197472240775824, + 0.002908211201429367, + 0.0034922161139547825, + 0.002652204828336835, + 0.0032194461673498154, + 0.005354283377528191, + 0.004465712234377861, + 0.014781483449041843, + 0.022369110956788063, + 0.017004942521452904, + 0.012777218595147133, + 0.015739044174551964, + 0.012945366092026234, + 0.010152654722332954, + 0.01252366416156292, + 0.009602824226021767, + 0.009262272156774998, + 0.014696680009365082, + 0.011431094259023666, + 0.012673530727624893, + 0.00943262130022049, + 0.008614636026322842, + 0.00786120630800724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.05495418235659599 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038092367351055145 + ], + "opposite_logprob_tokens": [ + -3.2880923748016357 + ], + "logprob_reward": -0.038092367351055145, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038092367351055145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1514148861169815, + 0.07489142566919327, + 0.05778343975543976, + 0.15002703666687012, + 0.18702003359794617, + 0.05173366144299507, + 0.02948552742600441, + 0.05432751402258873, + 0.15234749019145966, + 0.09096895903348923 + ], + "credit_map": [ + 0.018048731610178947, + 0.018383320420980453, + 0.01737670972943306, + 0.022744184359908104, + 0.030241329222917557, + 0.01319065224379301, + 0.013451087288558483, + 0.017978878691792488, + 0.007947767153382301, + 0.012041030451655388, + 0.0074435328133404255, + 0.005686063785105944, + 0.013086208142340183, + 0.010808270424604416, + 0.0071814707480371, + 0.010697083547711372, + 0.00971517339348793, + 0.011783871799707413, + 0.0058817085810005665, + 0.005268022418022156, + 0.007868411019444466, + 0.004986530635505915, + 0.006312916520982981, + 0.005966807249933481, + 0.020998621359467506, + 0.023500999435782433, + 0.01325950212776661, + 0.013699345290660858, + 0.030153490602970123, + 0.019297737628221512, + 0.01435592770576477, + 0.014761409722268581, + 0.007273405324667692, + 0.00814605038613081, + 0.003881178330630064, + 0.02088119462132454, + 0.009355251677334309, + 0.06041852757334709, + 0.007441927678883076, + 0.06962250918149948, + 0.009808535687625408, + 0.005968614015728235, + 0.004118726123124361, + 0.007136458996683359, + 0.007406572811305523, + 0.005667196121066809, + 0.005198825150728226, + 0.006428736727684736, + 0.004592577926814556, + 0.0042571877129375935, + 0.0037227566353976727, + 0.004421585705131292, + 0.0030421516858041286, + 0.0037094061262905598, + 0.0019315193640068173, + 0.0038083421532064676, + 0.006557408720254898, + 0.006887936964631081, + 0.00493278494104743, + 0.004013188183307648, + 0.007946792058646679, + 0.008725867606699467, + 0.007414031308144331, + 0.00784950703382492, + 0.020510634407401085, + 0.016112299636006355, + 0.01738811284303665, + 0.022183554247021675, + 0.014580439776182175, + 0.02114167995750904, + 0.01611972227692604, + 0.024311048910021782, + 0.005366139579564333, + 0.007059039082378149, + 0.01380489207804203, + 0.007887614890933037, + 0.008437735959887505, + 0.008925299160182476, + 0.024372169747948647, + 0.015116075985133648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.038092367351055145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977774802595377 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.00977774802595377, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977774802595377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4094942808151245, + 0.08154802769422531, + 0.035356055945158005, + 0.044607728719711304, + 0.07646025717258453, + 0.019357474520802498, + 0.03571926802396774, + 0.028774995356798172, + 0.11527764052152634, + 0.15340426564216614 + ], + "credit_map": [ + 0.04143732786178589, + 0.048348892480134964, + 0.03661876916885376, + 0.07506144046783447, + 0.06336945295333862, + 0.04744260758161545, + 0.0354781337082386, + 0.061737656593322754, + 0.005390540696680546, + 0.010324563831090927, + 0.007571472320705652, + 0.006726063787937164, + 0.01288130134344101, + 0.012251805514097214, + 0.012261927127838135, + 0.014140347018837929, + 0.0041043017990887165, + 0.008009728975594044, + 0.0033535948023200035, + 0.003640855895355344, + 0.006341521628201008, + 0.0038190253544598818, + 0.002870496129617095, + 0.0032165315933525562, + 0.007774725090712309, + 0.007161188405007124, + 0.004639527294784784, + 0.005092129111289978, + 0.006092775613069534, + 0.005325927399098873, + 0.004129165783524513, + 0.004392294678837061, + 0.005018206313252449, + 0.007110315840691328, + 0.0031307952012866735, + 0.007456567604094744, + 0.005322553217411041, + 0.023972077295184135, + 0.00450443709269166, + 0.019945304840803146, + 0.0025457735173404217, + 0.0025136980693787336, + 0.0016731438226997852, + 0.002453200053423643, + 0.003478688420727849, + 0.0019803098402917385, + 0.0022890386171638966, + 0.002423621714115143, + 0.004815315827727318, + 0.008453345857560635, + 0.0037663213443011045, + 0.0031650401651859283, + 0.004672556649893522, + 0.004129644017666578, + 0.0031143820378929377, + 0.0036026639863848686, + 0.0036228992976248264, + 0.004628281574696302, + 0.0036627615336328745, + 0.0033088650088757277, + 0.0027303111273795366, + 0.0027429077308624983, + 0.003966314718127251, + 0.004112653899937868, + 0.017699742689728737, + 0.014762802049517632, + 0.015338370576500893, + 0.017417309805750847, + 0.010010865516960621, + 0.015186943113803864, + 0.01011020690202713, + 0.01475139893591404, + 0.02014373056590557, + 0.012879940681159496, + 0.028685957193374634, + 0.015706852078437805, + 0.017046796157956123, + 0.015405581332743168, + 0.020919837057590485, + 0.022615572437644005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.00977774802595377 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950236693024635 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950236693024635, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950236693024635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12258969992399216, + 0.04807073622941971, + 0.027896933257579803, + 0.037841882556676865, + 0.4070235788822174, + 0.020986488088965416, + 0.03356652706861496, + 0.029031772166490555, + 0.14309318363666534, + 0.12989920377731323 + ], + "credit_map": [ + 0.007738942746073008, + 0.010735541582107544, + 0.008428357541561127, + 0.0114736994728446, + 0.03450490161776543, + 0.009016118943691254, + 0.01570897549390793, + 0.024983156472444534, + 0.005822408478707075, + 0.008816036395728588, + 0.004321228712797165, + 0.0038576608058065176, + 0.009537425823509693, + 0.005213833414018154, + 0.005116344895213842, + 0.0053857965394854546, + 0.004369672853499651, + 0.004429147113114595, + 0.0036386402789503336, + 0.0028275365475565195, + 0.0038319730665534735, + 0.0026542902924120426, + 0.0020604818128049374, + 0.004085192922502756, + 0.004357332829385996, + 0.0054268925450742245, + 0.005116230342537165, + 0.0036836385261267424, + 0.006680518388748169, + 0.005168617703020573, + 0.0033897513058036566, + 0.004018897190690041, + 0.016156736761331558, + 0.01752091944217682, + 0.005383641924709082, + 0.037441544234752655, + 0.02101738378405571, + 0.14533261954784393, + 0.014008142985403538, + 0.15016257762908936, + 0.0026612263172864914, + 0.003426471957936883, + 0.0018102991161867976, + 0.0028722386341542006, + 0.003696638159453869, + 0.001870940555818379, + 0.0020416860934346914, + 0.002606986090540886, + 0.0031178852077573538, + 0.006601865403354168, + 0.003131904872134328, + 0.004765346180647612, + 0.0037428250070661306, + 0.006544440519064665, + 0.0021480685099959373, + 0.0035141913685947657, + 0.004153526853770018, + 0.003321582218632102, + 0.003588468302041292, + 0.0027932811062783003, + 0.0035341025795787573, + 0.003702016081660986, + 0.00441409507766366, + 0.0035246978513896465, + 0.017603863030672073, + 0.012148350477218628, + 0.02117195911705494, + 0.0243170578032732, + 0.013465055264532566, + 0.019745169207453728, + 0.014054905623197556, + 0.02058682031929493, + 0.01168904360383749, + 0.005908334162086248, + 0.026862777769565582, + 0.007031076587736607, + 0.009261907078325748, + 0.009275762364268303, + 0.03966093063354492, + 0.02020936645567417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.005950236693024635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static low-angle medium close-up freezes the climactic moment of the action, rendered in a dramatic comic style that mimics a static image sequence. The protagonist is a formidable warrior, his body a coil of primal energy, clad in ornate silver armor that gleams with a metallic sheen. The armor is intricately detailed, with a sharp shield and carved gorget, its polished surface reflecting the chaos of battle. His face, contorted into a mask of pure rage, is the focal point; his eyes are wide, pupils constricted in a wild gaze, while his teeth are clenched so tightly that the muscles in his jaw appear distinct. He is caught in the act, thrusting a powerful spear directly at the viewer. As he drives his weapon forward, his arm blurs in motion, muscles bulging beneath the straps. A dense mass of sharp black and white dynamic speed lines radiating from the harpoon tip and his body visually convey the intense movement—a classic comic technique that conveys immense speed and impact. The scene is drenched in high-contrast, dramatic lighting, creating deep, inky shadows that sculpt his form and emphasize the raw power of his frantic battle against a stark abstract background of explosive energy, creating an intense and violent atmosphere.", + "seed": 21270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_5/index_7104_seed_21270_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01416441025212407, + "vlm_logprob": -0.01416441025212407, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.01416441025212407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009245322085916996 + ], + "opposite_logprob_tokens": [ + -5.0092453956604 + ], + "logprob_reward": -0.009245322085916996, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009245322085916996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30668050050735474, + 0.08551228791475296, + 0.0709858387708664, + 0.06789769232273102, + 0.08591735363006592, + 0.09135279804468155, + 0.026207666844129562, + 0.02905365079641342, + 0.14149253070354462, + 0.0948997437953949 + ], + "credit_map": [ + 0.030892131850123405, + 0.04437687247991562, + 0.03545689582824707, + 0.02451641671359539, + 0.0519837960600853, + 0.06527679413557053, + 0.027067404240369797, + 0.027110189199447632, + 0.009206177666783333, + 0.011946954764425755, + 0.007574834395200014, + 0.007719080895185471, + 0.017017830163240433, + 0.01663554459810257, + 0.00754323648288846, + 0.007868630811572075, + 0.008592757396399975, + 0.01267810259014368, + 0.00965385977178812, + 0.0040968130342662334, + 0.013336723670363426, + 0.010641028173267841, + 0.006623137276619673, + 0.005363416858017445, + 0.004209442529827356, + 0.009305342100560665, + 0.00795792043209076, + 0.006672695279121399, + 0.012487339787185192, + 0.010873989202082157, + 0.008891980163753033, + 0.007498984690755606, + 0.019224591553211212, + 0.008315234445035458, + 0.024612121284008026, + 0.005307051818817854, + 0.008839836344122887, + 0.007189514581114054, + 0.006119346246123314, + 0.006309660617262125, + 0.023504935204982758, + 0.01177180279046297, + 0.020399121567606926, + 0.006444236729294062, + 0.007491998840123415, + 0.006818183232098818, + 0.006770973093807697, + 0.00815154705196619, + 0.0037808469496667385, + 0.0028930744156241417, + 0.004766332916915417, + 0.00218194629997015, + 0.00454501761123538, + 0.0032260434236377478, + 0.002245844341814518, + 0.0025685608852654696, + 0.004359682090580463, + 0.0038314578123390675, + 0.005330447573214769, + 0.0022216110955923796, + 0.004943911451846361, + 0.004028553143143654, + 0.002192145213484764, + 0.0021458424162119627, + 0.01530048530548811, + 0.01662537269294262, + 0.015990065410733223, + 0.018060579895973206, + 0.026341119781136513, + 0.013905747793614864, + 0.010892963036894798, + 0.024376196786761284, + 0.007734360639005899, + 0.012691102921962738, + 0.012083635665476322, + 0.00977469515055418, + 0.018897002562880516, + 0.009828870184719563, + 0.013579780235886574, + 0.01031029224395752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.009245322085916996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305723309516907 + ], + "opposite_logprob_tokens": [ + -3.1680572032928467 + ], + "logprob_reward": -0.04305723309516907, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305723309516907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2666454315185547, + 0.1016114205121994, + 0.06690170615911484, + 0.047088656574487686, + 0.07117148488759995, + 0.06251456588506699, + 0.01841040328145027, + 0.018915753811597824, + 0.2546028792858124, + 0.09213767200708389 + ], + "credit_map": [ + 0.02589654177427292, + 0.04708952084183693, + 0.022195536643266678, + 0.01823926344513893, + 0.03230557590723038, + 0.07399950176477432, + 0.019304534420371056, + 0.02761496789753437, + 0.01681968756020069, + 0.017271162942051888, + 0.014664985239505768, + 0.006768068764358759, + 0.018382059410214424, + 0.010850018821656704, + 0.008845052681863308, + 0.008010388351976871, + 0.006735170725733042, + 0.010190069675445557, + 0.009173627942800522, + 0.0047075338661670685, + 0.0140154380351305, + 0.008875845931470394, + 0.007023151032626629, + 0.0061808680184185505, + 0.0032660597935318947, + 0.006545564625412226, + 0.006032116245478392, + 0.0036730633582919836, + 0.011143429204821587, + 0.006335665937513113, + 0.005378370173275471, + 0.004714387468993664, + 0.0059177326038479805, + 0.0035785846412181854, + 0.026057632640004158, + 0.004129715263843536, + 0.0046458118595182896, + 0.004262183327227831, + 0.004175953101366758, + 0.01840386912226677, + 0.0108712837100029, + 0.011595958843827248, + 0.00905840378254652, + 0.008430189453065395, + 0.005244144704192877, + 0.006609248463064432, + 0.004578674677759409, + 0.006126661319285631, + 0.0033012947533279657, + 0.00410858727991581, + 0.002610601484775543, + 0.0011124308221042156, + 0.002571108750998974, + 0.002195491921156645, + 0.0010928696719929576, + 0.0014180195285007358, + 0.00227403175085783, + 0.002444223966449499, + 0.0022574036847800016, + 0.0013758707791566849, + 0.003935371991246939, + 0.0027675707824528217, + 0.0015172937419265509, + 0.0023439875803887844, + 0.023590324446558952, + 0.032317474484443665, + 0.023045457899570465, + 0.024022774770855904, + 0.05391889438033104, + 0.029317984357476234, + 0.02120860107243061, + 0.04718134552240372, + 0.007824528031051159, + 0.016687465831637383, + 0.013975775800645351, + 0.007428304757922888, + 0.0160976629704237, + 0.015165772289037704, + 0.009258655831217766, + 0.005699511151760817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.04305723309516907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071793735027313 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071793735027313, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071793735027313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17025072872638702, + 0.10122844576835632, + 0.08459661155939102, + 0.04735726863145828, + 0.16815230250358582, + 0.053863897919654846, + 0.01678996905684471, + 0.02203327789902687, + 0.2170761078596115, + 0.11865131556987762 + ], + "credit_map": [ + 0.017588142305612564, + 0.035430632531642914, + 0.011126040481030941, + 0.011840922757983208, + 0.019924771040678024, + 0.04716760665178299, + 0.010432525537908077, + 0.016740091145038605, + 0.012074200436472893, + 0.014639072120189667, + 0.011126870289444923, + 0.010798470117151737, + 0.017990171909332275, + 0.012994927354156971, + 0.012194504961371422, + 0.009410235099494457, + 0.007550876587629318, + 0.010364631190896034, + 0.009373495355248451, + 0.01004798337817192, + 0.016285238787531853, + 0.008116140961647034, + 0.015576127916574478, + 0.007282119709998369, + 0.003607431659474969, + 0.0057609244249761105, + 0.01025843434035778, + 0.006690556183457375, + 0.00608848687261343, + 0.0051743085496127605, + 0.00419797794893384, + 0.005579153075814247, + 0.04037176072597504, + 0.012268111109733582, + 0.01337494421750307, + 0.005655959714204073, + 0.008393936790525913, + 0.017099646851420403, + 0.00970219075679779, + 0.06128576025366783, + 0.014508443884551525, + 0.005158923100680113, + 0.01037054043263197, + 0.0022816667333245277, + 0.005625464953482151, + 0.007018809672445059, + 0.004973011091351509, + 0.0039270417764782906, + 0.002073233015835285, + 0.002877575345337391, + 0.0017485833959653974, + 0.001293072709813714, + 0.0030301292426884174, + 0.00288370531052351, + 0.0013881342019885778, + 0.001495536882430315, + 0.0028721680864691734, + 0.004302455112338066, + 0.003661381546407938, + 0.0013101886725053191, + 0.004163313191384077, + 0.0027758495416492224, + 0.00164276291616261, + 0.0013051597634330392, + 0.02147144451737404, + 0.025416145101189613, + 0.02083687111735344, + 0.024989157915115356, + 0.03494097292423248, + 0.02757878601551056, + 0.017697108909487724, + 0.04414563253521919, + 0.013323486782610416, + 0.016575666144490242, + 0.016472337767481804, + 0.008644096553325653, + 0.02025253139436245, + 0.01912834122776985, + 0.013660170137882233, + 0.010594701394438744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.011071793735027313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195212058722973 + ], + "opposite_logprob_tokens": [ + -6.127195358276367 + ], + "logprob_reward": -0.002195212058722973, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195212058722973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14038319885730743, + 0.05780014768242836, + 0.10102342069149017, + 0.10057888925075531, + 0.25078338384628296, + 0.13076044619083405, + 0.034955207258462906, + 0.04801667109131813, + 0.07915129512548447, + 0.05654725059866905 + ], + "credit_map": [ + 0.011941095814108849, + 0.02508542127907276, + 0.013861610554158688, + 0.011058028787374496, + 0.022955583408474922, + 0.03427020087838173, + 0.01073989924043417, + 0.010471365414559841, + 0.006760118063539267, + 0.007677270565181971, + 0.010318857617676258, + 0.004182278178632259, + 0.010772524401545525, + 0.007348592393100262, + 0.005960455164313316, + 0.004780054558068514, + 0.013256512582302094, + 0.015054834075272083, + 0.01947931945323944, + 0.006233446300029755, + 0.0185931995511055, + 0.011260701343417168, + 0.009769361466169357, + 0.007376052439212799, + 0.009625335223972797, + 0.012910674326121807, + 0.019536374136805534, + 0.0072540524415671825, + 0.019849667325615883, + 0.015557004138827324, + 0.007602113764733076, + 0.008243675343692303, + 0.07734409719705582, + 0.019813619554042816, + 0.03055211715400219, + 0.008950937539339066, + 0.010030470788478851, + 0.018365448340773582, + 0.010854511521756649, + 0.074872225522995, + 0.02401847019791603, + 0.014628884382545948, + 0.031185511499643326, + 0.011192741803824902, + 0.010693947784602642, + 0.015275273472070694, + 0.009184991009533405, + 0.014580639079213142, + 0.005160277243703604, + 0.005478520877659321, + 0.00431499257683754, + 0.0028491835109889507, + 0.004663597326725721, + 0.0061665805988013744, + 0.002402640413492918, + 0.003919420298188925, + 0.005769143346697092, + 0.00557520380243659, + 0.01238231360912323, + 0.003902832977473736, + 0.007772075943648815, + 0.006004608701914549, + 0.003154183505102992, + 0.003456314094364643, + 0.007612802088260651, + 0.011645188555121422, + 0.009622056037187576, + 0.007314841728657484, + 0.02106674388051033, + 0.008801989257335663, + 0.0059289406053721905, + 0.007158743683248758, + 0.005448020529001951, + 0.008406644687056541, + 0.013186580501496792, + 0.004819941706955433, + 0.009469689801335335, + 0.005009093787521124, + 0.005772783886641264, + 0.00443450128659606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.002195212058722973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19422651827335358, + 0.03963373228907585, + 0.07306132465600967, + 0.08846375346183777, + 0.1007402092218399, + 0.1940937489271164, + 0.01584130898118019, + 0.020816657692193985, + 0.22398807108402252, + 0.04913470149040222 + ], + "credit_map": [ + 0.0275589469820261, + 0.04752011224627495, + 0.019351564347743988, + 0.013155280612409115, + 0.016066774725914, + 0.04455959051847458, + 0.009429102763533592, + 0.016585148870944977, + 0.0059875850565731525, + 0.0061311740428209305, + 0.003916082438081503, + 0.0029836641624569893, + 0.006017894018441439, + 0.007786082103848457, + 0.0036020264960825443, + 0.0032092223409563303, + 0.011215774342417717, + 0.01320592500269413, + 0.010794933885335922, + 0.003169137518852949, + 0.011158928275108337, + 0.013374830596148968, + 0.004568113479763269, + 0.005573684815317392, + 0.010947217233479023, + 0.01181681640446186, + 0.01418036874383688, + 0.005870976950973272, + 0.013743558898568153, + 0.016849208623170853, + 0.00782250240445137, + 0.007233099080622196, + 0.010382278822362423, + 0.005916228052228689, + 0.02829659730195999, + 0.0049156323075294495, + 0.007324389647692442, + 0.005259139928966761, + 0.005184960551559925, + 0.033460989594459534, + 0.03655961528420448, + 0.03733297809958458, + 0.024939769878983498, + 0.028199540451169014, + 0.017078090459108353, + 0.016014650464057922, + 0.01281755045056343, + 0.021151544526219368, + 0.0019712119828909636, + 0.002039211569353938, + 0.002727468963712454, + 0.0011543784057721496, + 0.003304304787889123, + 0.002163345692679286, + 0.0014282058691605926, + 0.0010531804291531444, + 0.002874864498153329, + 0.0034598864149302244, + 0.0033301094081252813, + 0.0014150692149996758, + 0.003966770600527525, + 0.0025287303142249584, + 0.0015665763057768345, + 0.001674650819040835, + 0.025989918038249016, + 0.025973249226808548, + 0.01673406921327114, + 0.027659745886921883, + 0.036294374614953995, + 0.024189196527004242, + 0.01790725439786911, + 0.04924025759100914, + 0.006072849966585636, + 0.005355618428438902, + 0.007326617371290922, + 0.0050886995159089565, + 0.008550464175641537, + 0.0060600293800234795, + 0.0061586215160787106, + 0.004521796479821205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.005252490285784006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static low-angle medium close-up freezes the climactic moment of the action, rendered in a dramatic comic style that mimics a static image sequence. The protagonist is a formidable warrior, his body a coil of primal energy, clad in ornate silver armor that gleams with a metallic sheen. The armor is intricately detailed, with a sharp shield and carved gorget, its polished surface reflecting the chaos of battle. His face, contorted into a mask of pure rage, is the focal point; his eyes are wide, pupils constricted in a wild gaze, while his teeth are clenched so tightly that the muscles in his jaw appear distinct. He is caught in the act, thrusting a powerful spear directly at the viewer. As he drives his weapon forward, his arm blurs in motion, muscles bulging beneath the straps. A dense mass of sharp black and white dynamic speed lines radiating from the harpoon tip and his body visually convey the intense movement—a classic comic technique that conveys immense speed and impact. The scene is drenched in high-contrast, dramatic lighting, creating deep, inky shadows that sculpt his form and emphasize the raw power of his frantic battle against a stark abstract background of explosive energy, creating an intense and violent atmosphere.", + "seed": 21271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_5/index_7104_seed_21271_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.048707938054576516, + "vlm_logprob": -0.048707938054576516, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.048707938054576516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026514075696468353 + ], + "opposite_logprob_tokens": [ + -3.7765140533447266 + ], + "logprob_reward": -0.026514075696468353, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026514075696468353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2167610377073288, + 0.062304895371198654, + 0.051916833966970444, + 0.07133356481790543, + 0.22228370606899261, + 0.03249762952327728, + 0.09219720959663391, + 0.12714223563671112, + 0.08093979954719543, + 0.042623113840818405 + ], + "credit_map": [ + 0.020262999460101128, + 0.022213047370314598, + 0.020564619451761246, + 0.043649911880493164, + 0.037317052483558655, + 0.01899108476936817, + 0.03418239578604698, + 0.019579915329813957, + 0.006394482217729092, + 0.00531977042555809, + 0.006353025324642658, + 0.010890981182456017, + 0.009508511982858181, + 0.00750197097659111, + 0.009521610103547573, + 0.006814541295170784, + 0.004949372261762619, + 0.004834596533328295, + 0.0023610431235283613, + 0.013613386079668999, + 0.005259532481431961, + 0.012714879587292671, + 0.00479172682389617, + 0.0033922947477549314, + 0.008005953393876553, + 0.00803728774189949, + 0.004727765452116728, + 0.008606309071183205, + 0.018442317843437195, + 0.008404224179685116, + 0.008648231625556946, + 0.006461475044488907, + 0.06340068578720093, + 0.013189626857638359, + 0.08405289053916931, + 0.009140454232692719, + 0.011666354723274708, + 0.018726229667663574, + 0.010554778389632702, + 0.011552668176591396, + 0.003972421400249004, + 0.0039105829782783985, + 0.0035999317187815905, + 0.004093881696462631, + 0.005962255876511335, + 0.004706074483692646, + 0.004211027640849352, + 0.0020414532627910376, + 0.013770733028650284, + 0.008103767409920692, + 0.011064944788813591, + 0.013531148433685303, + 0.011466036550700665, + 0.013098403811454773, + 0.012143902480602264, + 0.00901826936751604, + 0.015090642496943474, + 0.013135557994246483, + 0.009226426482200623, + 0.01667487993836403, + 0.017111187800765038, + 0.020135557278990746, + 0.018727777525782585, + 0.017040211707353592, + 0.0192182008177042, + 0.0085728345438838, + 0.011279750615358353, + 0.007575348950922489, + 0.006456601433455944, + 0.013039015233516693, + 0.006389522459357977, + 0.00840852316468954, + 0.005279163364320993, + 0.0036926253233104944, + 0.004203288350254297, + 0.006123239174485207, + 0.007738396059721708, + 0.003829296911135316, + 0.007381538860499859, + 0.004375561606138945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.026514075696468353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347124949097633 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347124949097633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347124949097633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06598605215549469, + 0.12316497415304184, + 0.27431535720825195, + 0.04881824925541878, + 0.13718841969966888, + 0.1425241082906723, + 0.0781140998005867, + 0.023850522935390472, + 0.06823140382766724, + 0.03780679777264595 + ], + "credit_map": [ + 0.0062585934065282345, + 0.008483258076012135, + 0.006430341862142086, + 0.010443051345646381, + 0.011430668644607067, + 0.007643067743629217, + 0.009737839922308922, + 0.005559232551604509, + 0.007267356384545565, + 0.010044087655842304, + 0.00970376841723919, + 0.011722635477781296, + 0.009286245331168175, + 0.038313791155815125, + 0.013669471256434917, + 0.023157615214586258, + 0.030828610062599182, + 0.024878861382603645, + 0.009422033093869686, + 0.06695146858692169, + 0.017608582973480225, + 0.09858857840299606, + 0.014410615898668766, + 0.01162660587579012, + 0.007277942728251219, + 0.003732687793672085, + 0.004874992650002241, + 0.007013794034719467, + 0.007559671066701412, + 0.0062250359915196896, + 0.006381290964782238, + 0.005752832628786564, + 0.02175539918243885, + 0.017771316692233086, + 0.018746880814433098, + 0.014468422159552574, + 0.01829606667160988, + 0.020576030015945435, + 0.016221774742007256, + 0.009352518245577812, + 0.011343826539814472, + 0.016916649416089058, + 0.020682478323578835, + 0.01654049940407276, + 0.0236587505787611, + 0.01733076199889183, + 0.024597788229584694, + 0.011453358456492424, + 0.014034238643944263, + 0.006625113543123007, + 0.006971570663154125, + 0.010182575322687626, + 0.012302338145673275, + 0.009437792003154755, + 0.012257273308932781, + 0.006303203757852316, + 0.002507653320208192, + 0.0022102517541497946, + 0.0025429034139961004, + 0.004367182031273842, + 0.0032456594053655863, + 0.0028098386246711016, + 0.003516522003337741, + 0.0026505128480494022, + 0.016781510785222054, + 0.009237486869096756, + 0.009139134548604488, + 0.006425921805202961, + 0.005093609914183617, + 0.009613883681595325, + 0.005053685046732426, + 0.006886170711368322, + 0.0047116270288825035, + 0.0036130461376160383, + 0.003985479939728975, + 0.0053175510838627815, + 0.008489299565553665, + 0.003836877178400755, + 0.004362707491964102, + 0.0034902116749435663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.026347124949097633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798577904701233 + ], + "opposite_logprob_tokens": [ + -1.8048577308654785 + ], + "logprob_reward": -0.1798577904701233, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798577904701233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04821602255105972, + 0.055497437715530396, + 0.3705625534057617, + 0.025941750034689903, + 0.20757679641246796, + 0.07660241425037384, + 0.056355543434619904, + 0.04107998311519623, + 0.0927274227142334, + 0.025440122932195663 + ], + "credit_map": [ + 0.007275450974702835, + 0.0041120704263448715, + 0.00592577550560236, + 0.005887953098863363, + 0.008724970743060112, + 0.0048993052914738655, + 0.007081795018166304, + 0.0043086945079267025, + 0.0031688762828707695, + 0.0034541215281933546, + 0.002858840161934495, + 0.004818370100110769, + 0.005821101367473602, + 0.01638084277510643, + 0.0059699369594454765, + 0.01302533969283104, + 0.0462147556245327, + 0.015013341791927814, + 0.015139507129788399, + 0.11635826528072357, + 0.01616513915359974, + 0.12420891970396042, + 0.019687844440340996, + 0.01777469366788864, + 0.002237616805359721, + 0.0018946630880236626, + 0.0025132615119218826, + 0.005534294527024031, + 0.004120376892387867, + 0.003330221399664879, + 0.0035579472314566374, + 0.002753365784883499, + 0.05073268711566925, + 0.014477440156042576, + 0.07502928376197815, + 0.013342267833650112, + 0.01656867004930973, + 0.01686081849038601, + 0.009943205863237381, + 0.010622388683259487, + 0.007742125075310469, + 0.0107196606695652, + 0.013414482586085796, + 0.008831553161144257, + 0.009514686651527882, + 0.009812786243855953, + 0.010603907518088818, + 0.005963204428553581, + 0.010015691630542278, + 0.006937840487807989, + 0.005851367022842169, + 0.009098274633288383, + 0.0057293567806482315, + 0.006114584859460592, + 0.0056268079206347466, + 0.006981612183153629, + 0.0065955412574112415, + 0.003202970139682293, + 0.0031202377285808325, + 0.006672181189060211, + 0.005056673660874367, + 0.005535649601370096, + 0.0041162725538015366, + 0.006780452094972134, + 0.022183960303664207, + 0.012446167878806591, + 0.013291850686073303, + 0.00592056242749095, + 0.006181255914270878, + 0.015979072079062462, + 0.005528103094547987, + 0.011196441948413849, + 0.0023457971401512623, + 0.002284445334225893, + 0.0024449436459690332, + 0.0035797995515167713, + 0.005129891913384199, + 0.002581549109891057, + 0.00450570322573185, + 0.002567992778494954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.1798577904701233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625991944223642 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625991944223642, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625991944223642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08886063098907471, + 0.09573767334222794, + 0.10336139798164368, + 0.05085153132677078, + 0.2651859223842621, + 0.1511799395084381, + 0.0561160184442997, + 0.04540296271443367, + 0.0912347361445427, + 0.05206925794482231 + ], + "credit_map": [ + 0.01687345653772354, + 0.011209482327103615, + 0.011712536215782166, + 0.011403407901525497, + 0.011800523847341537, + 0.008105331100523472, + 0.00984202791005373, + 0.007913852110505104, + 0.020766759291291237, + 0.012153131887316704, + 0.009940288960933685, + 0.010725784115493298, + 0.01127850916236639, + 0.013417485170066357, + 0.009121590293943882, + 0.008334111422300339, + 0.008692695759236813, + 0.0047233751974999905, + 0.004696833435446024, + 0.03444839268922806, + 0.008094361051917076, + 0.027792511507868767, + 0.008895975537598133, + 0.006017243955284357, + 0.006267522927373648, + 0.005289854947477579, + 0.0045115458779037, + 0.009305932559072971, + 0.008762899786233902, + 0.005352760199457407, + 0.006282650865614414, + 0.005078360438346863, + 0.03439922258257866, + 0.032073166221380234, + 0.0675656795501709, + 0.02989865653216839, + 0.036124855279922485, + 0.03036816231906414, + 0.020388957113027573, + 0.014367207884788513, + 0.014851197600364685, + 0.02044382318854332, + 0.026086626574397087, + 0.019782545045018196, + 0.0169941708445549, + 0.020847352221608162, + 0.02126493491232395, + 0.010909268632531166, + 0.008373924531042576, + 0.006428663618862629, + 0.005715027451515198, + 0.007753525394946337, + 0.007428591139614582, + 0.006070700939744711, + 0.007922300137579441, + 0.006423282902687788, + 0.009110450744628906, + 0.003677997039631009, + 0.004469558130949736, + 0.005290852393954992, + 0.006239891052246094, + 0.0061100805178284645, + 0.00602389732375741, + 0.004480232018977404, + 0.01945924386382103, + 0.010108515620231628, + 0.013819855637848377, + 0.006456773262470961, + 0.006908476818352938, + 0.016505997627973557, + 0.0064237299375236034, + 0.011552137322723866, + 0.008853158913552761, + 0.00341933686286211, + 0.007671886123716831, + 0.006322353146970272, + 0.005001036450266838, + 0.007773684337735176, + 0.006376909092068672, + 0.006650893483310938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.007625991944223642 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the protagonist depicted as a warrior wearing ornate silver armor with a sharp shield and carved gorget?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior's face contorted in rage with wide eyes, constricted pupils, and clenched teeth showing distinct jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior thrusting a powerful spear directly at the viewer with an arm that blurs in motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dense black and white dynamic speed lines radiating from the spear tip and the warrior's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a static comic style with high-contrast dramatic lighting and deep inky shadows against a stark abstract background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031947072129696608 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031947072129696608, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031947072129696608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1450931876897812, + 0.10416295379400253, + 0.14269176125526428, + 0.051409393548965454, + 0.20422568917274475, + 0.18728719651699066, + 0.043705958873033524, + 0.04714200273156166, + 0.03309384733438492, + 0.041188016533851624 + ], + "credit_map": [ + 0.02604866214096546, + 0.015645096078515053, + 0.01959419995546341, + 0.01810525916516781, + 0.014191328547894955, + 0.017099913209676743, + 0.017320213839411736, + 0.017088508233428, + 0.0069361221976578236, + 0.006685856729745865, + 0.011729789897799492, + 0.010256742127239704, + 0.009877596981823444, + 0.03139369934797287, + 0.007117568515241146, + 0.02016557939350605, + 0.022489318624138832, + 0.006735802162438631, + 0.005558365490287542, + 0.0420040637254715, + 0.006563631817698479, + 0.04431512951850891, + 0.007420812733471394, + 0.007604638114571571, + 0.006063883192837238, + 0.006565507501363754, + 0.004595660138875246, + 0.0068371533416211605, + 0.007530618458986282, + 0.007199373096227646, + 0.005862707272171974, + 0.006754486355930567, + 0.040604349225759506, + 0.0184237789362669, + 0.06730746477842331, + 0.017039567232131958, + 0.021399714052677155, + 0.01762365736067295, + 0.011469687335193157, + 0.010357465595006943, + 0.023764023557305336, + 0.022644203156232834, + 0.028842097148299217, + 0.02639572136104107, + 0.018355626612901688, + 0.02935841679573059, + 0.02507897838950157, + 0.012848135083913803, + 0.009121108800172806, + 0.003488409100100398, + 0.004756212700158358, + 0.006560992915183306, + 0.004125671461224556, + 0.005818757228553295, + 0.004378799349069595, + 0.005456005223095417, + 0.007303290069103241, + 0.004822679795324802, + 0.00466860830783844, + 0.007793042343109846, + 0.00536264618858695, + 0.007096210960298777, + 0.006381503771990538, + 0.0037140194326639175, + 0.005703754723072052, + 0.0037348170299082994, + 0.0036309051793068647, + 0.005244482308626175, + 0.004235275089740753, + 0.0031121366191655397, + 0.004493918269872665, + 0.002938557183369994, + 0.005250784568488598, + 0.0045603057369589806, + 0.004096934571862221, + 0.006191590800881386, + 0.005343141965568066, + 0.004221935290843248, + 0.005731034558266401, + 0.005792289040982723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0031947072129696608 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_tensor([7400, 7400])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_tensor([7400, 7400])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0fd466936d0881a16503fce58e0d40140b11103 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_tensor([7400, 7400])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up, at eye level, slowly pans to a man with short, dark hair, his face frozen in shock. He wears a crisp white button-down shirt, its brightness contrasting sharply with the dim, ambiguous interior. The primary light source, a cool digital blue tint from off-camera, casts sharp, dramatic shadows across his features and creates a high-contrast, tense atmosphere of suspense. His eyes are wide, reflecting an invisible blue light, his eyebrows raised in disbelief, and his mouth slightly open, as if caught in a silent gasp. The background is completely out of focus, reduced to deep, soft shadows, isolating the man and heightening the sense of suspense, locking the viewer in his sudden, terrifying realization.", + "seed": 16049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_7400_seed_16049_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1683467805152759, + "vlm_logprob": -0.1683467805152759, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.1683467805152759, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005397153086960316 + ], + "opposite_logprob_tokens": [ + -6.130397319793701 + ], + "logprob_reward": -0.005397153086960316, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005397153086960316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23303130269050598, + 0.1494799703359604, + 0.042706988751888275, + 0.1290988028049469, + 0.03609277680516243, + 0.07003399729728699, + 0.18070177733898163, + 0.03804981708526611, + 0.032306309789419174, + 0.08849820494651794 + ], + "credit_map": [ + 0.02836778201162815, + 0.02504727803170681, + 0.01911083795130253, + 0.021845074370503426, + 0.02818174473941326, + 0.04136985167860985, + 0.026692092418670654, + 0.042416661977767944, + 0.015037352219223976, + 0.014386377297341824, + 0.010402464307844639, + 0.013042422011494637, + 0.02736225724220276, + 0.030924515798687935, + 0.023070190101861954, + 0.015254395082592964, + 0.007123422808945179, + 0.005829834379255772, + 0.004118249285966158, + 0.0046515134163200855, + 0.005010591354221106, + 0.005273184273391962, + 0.004470420069992542, + 0.006229774560779333, + 0.013677208684384823, + 0.018508395180106163, + 0.007613914553076029, + 0.018226468935608864, + 0.01582249253988266, + 0.019062768667936325, + 0.01828809268772602, + 0.01789948157966137, + 0.0041542272083461285, + 0.004086240194737911, + 0.0032796624582260847, + 0.005411240737885237, + 0.0054468619637191296, + 0.005382257979363203, + 0.004102592822164297, + 0.004229696001857519, + 0.0072876568883657455, + 0.007970101200044155, + 0.00551198422908783, + 0.012321995571255684, + 0.005874897353351116, + 0.008525479584932327, + 0.00992209929972887, + 0.012619789689779282, + 0.018602279946208, + 0.012392996810376644, + 0.01202143169939518, + 0.027084829285740852, + 0.0177717674523592, + 0.02370276302099228, + 0.031367894262075424, + 0.037757836282253265, + 0.004651885014027357, + 0.006306440103799105, + 0.00282465061172843, + 0.0071219708770513535, + 0.004464915953576565, + 0.005087269935756922, + 0.004737820941954851, + 0.0028548650443553925, + 0.0060384394600987434, + 0.0028779625426977873, + 0.0022837715223431587, + 0.0025758056435734034, + 0.006605301517993212, + 0.005607459228485823, + 0.003244304796680808, + 0.003073267173022032, + 0.010432031936943531, + 0.00916507001966238, + 0.007509595714509487, + 0.017745226621627808, + 0.011320722289383411, + 0.014034619554877281, + 0.010338768362998962, + 0.00795217975974083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.005397153086960316 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019511011196300387 + ], + "opposite_logprob_tokens": [ + -6.251951217651367 + ], + "logprob_reward": -0.0019511011196300387, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019511011196300387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23352621495723724, + 0.10320181399583817, + 0.0681244507431984, + 0.05776750296354294, + 0.1151338517665863, + 0.10978084057569504, + 0.07167656719684601, + 0.07804443687200546, + 0.06392119079828262, + 0.0988231748342514 + ], + "credit_map": [ + 0.024733571335673332, + 0.03255399316549301, + 0.018221203237771988, + 0.023683937266469002, + 0.03118300251662731, + 0.03786994516849518, + 0.02787574753165245, + 0.03740480914711952, + 0.010233857668936253, + 0.01808650605380535, + 0.008654405362904072, + 0.016600079834461212, + 0.013504170812666416, + 0.013477801345288754, + 0.01111163105815649, + 0.01153335627168417, + 0.0071142916567623615, + 0.011919311247766018, + 0.0053598154336214066, + 0.011167735792696476, + 0.008644831366837025, + 0.009336501359939575, + 0.006663594860583544, + 0.007918375544250011, + 0.005455350503325462, + 0.010912144556641579, + 0.0037826579064130783, + 0.00984989758580923, + 0.006991456728428602, + 0.007936461828649044, + 0.006191810127347708, + 0.006647722329944372, + 0.014210861176252365, + 0.007992889732122421, + 0.0071602798998355865, + 0.013215720653533936, + 0.01208202913403511, + 0.01700327731668949, + 0.01807418465614319, + 0.025394605472683907, + 0.011926600709557533, + 0.011502555571496487, + 0.008581777103245258, + 0.028441455215215683, + 0.012152533046901226, + 0.01279747486114502, + 0.013843908905982971, + 0.01053453516215086, + 0.0077674840576946735, + 0.006402942817658186, + 0.005823871586471796, + 0.007152894511818886, + 0.008718974888324738, + 0.009013723582029343, + 0.01088846568018198, + 0.015908213332295418, + 0.007010813802480698, + 0.008936622180044651, + 0.006444460246711969, + 0.009970951825380325, + 0.008599928580224514, + 0.010794206522405148, + 0.013047988526523113, + 0.01323946937918663, + 0.007145178038626909, + 0.00544785289093852, + 0.005523196421563625, + 0.007227947469800711, + 0.009748416021466255, + 0.011317037045955658, + 0.010455305688083172, + 0.007056260481476784, + 0.01215337123721838, + 0.009006436914205551, + 0.0071964384987950325, + 0.018318604677915573, + 0.015451454557478428, + 0.02107936516404152, + 0.008304174058139324, + 0.00731333252042532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0019511011196300387 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036187181249260902 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036187181249260902, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036187181249260902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18220102787017822, + 0.15513649582862854, + 0.1602824330329895, + 0.1458747684955597, + 0.06736261397600174, + 0.029401512816548347, + 0.13919755816459656, + 0.03496766462922096, + 0.04003952443599701, + 0.045536454766988754 + ], + "credit_map": [ + 0.031392231583595276, + 0.015654990449547768, + 0.01618295907974243, + 0.013150674290955067, + 0.023291567340493202, + 0.030217213556170464, + 0.01995236985385418, + 0.03235900029540062, + 0.026290569454431534, + 0.014660968445241451, + 0.014592614956200123, + 0.019255368039011955, + 0.016679853200912476, + 0.021291036158800125, + 0.01993699185550213, + 0.022429101169109344, + 0.01787184551358223, + 0.021357856690883636, + 0.014097316190600395, + 0.026997139677405357, + 0.019794413819909096, + 0.023903632536530495, + 0.017013905569911003, + 0.019246315583586693, + 0.015960821881890297, + 0.02006343938410282, + 0.010248479433357716, + 0.026256684213876724, + 0.017177682369947433, + 0.02008875459432602, + 0.018244890496134758, + 0.0178340133279562, + 0.007654903456568718, + 0.005090031772851944, + 0.0038901534862816334, + 0.006023874972015619, + 0.0081735048443079, + 0.010430438444018364, + 0.010794773697853088, + 0.0153049286454916, + 0.0032921815291047096, + 0.003501555183902383, + 0.0017588447080925107, + 0.005151408724486828, + 0.0031717661768198013, + 0.003821723395958543, + 0.0031094937585294247, + 0.005594540853053331, + 0.014842349104583263, + 0.009628228843212128, + 0.008709044195711613, + 0.0232413187623024, + 0.013560852967202663, + 0.018642162904143333, + 0.023255815729498863, + 0.027317779138684273, + 0.006439862307161093, + 0.004283958114683628, + 0.0027720178477466106, + 0.00453714607283473, + 0.0036985334008932114, + 0.004859944339841604, + 0.004138300195336342, + 0.00423790467903018, + 0.003999100532382727, + 0.0061042457818984985, + 0.0029888167046010494, + 0.005441328044980764, + 0.006525155622512102, + 0.0064782509580254555, + 0.005194305442273617, + 0.0033083215821534395, + 0.006955594290047884, + 0.004964818246662617, + 0.004640997853130102, + 0.006765022408217192, + 0.0051647936925292015, + 0.005905344616621733, + 0.005296836141496897, + 0.005843046121299267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0036187181249260902 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638507030904293 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638507030904293, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638507030904293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06023135408759117, + 0.04118210822343826, + 0.015220943838357925, + 0.07364008575677872, + 0.07357943803071976, + 0.0507345013320446, + 0.40001150965690613, + 0.07289184629917145, + 0.029378943145275116, + 0.18312928080558777 + ], + "credit_map": [ + 0.008066482841968536, + 0.007145007140934467, + 0.0050027272664010525, + 0.0059524173848330975, + 0.007885457947850227, + 0.008693122304975986, + 0.009036589413881302, + 0.008449545130133629, + 0.0041648284532129765, + 0.003459892002865672, + 0.002486984711140394, + 0.00533434608951211, + 0.006150547415018082, + 0.007609163876622915, + 0.005527001339942217, + 0.006449342239648104, + 0.0019466679077595472, + 0.00177513447124511, + 0.0010238982504233718, + 0.0014750361442565918, + 0.0019465753575786948, + 0.002763135591521859, + 0.0016623986884951591, + 0.0026280966121703386, + 0.0071271369233727455, + 0.009121792390942574, + 0.004296503495424986, + 0.012641455978155136, + 0.00961841456592083, + 0.010536422953009605, + 0.00940258800983429, + 0.010895767249166965, + 0.007228396832942963, + 0.004855798557400703, + 0.00455382838845253, + 0.00964801199734211, + 0.006793324835598469, + 0.010554594919085503, + 0.010217446833848953, + 0.019728031009435654, + 0.006041571963578463, + 0.005089417565613985, + 0.003973746672272682, + 0.015022309496998787, + 0.00525864539667964, + 0.005690791178494692, + 0.003957914188504219, + 0.00570010906085372, + 0.04119657352566719, + 0.02750927209854126, + 0.028902025893330574, + 0.054092276841402054, + 0.035803232342004776, + 0.052764587104320526, + 0.06875012814998627, + 0.09099338948726654, + 0.006850305944681168, + 0.010415671393275261, + 0.004495762288570404, + 0.011342557147145271, + 0.011869950219988823, + 0.01075825560837984, + 0.010306045413017273, + 0.00685329781845212, + 0.004434171598404646, + 0.0028964090161025524, + 0.0021463085431605577, + 0.0052884542383253574, + 0.004198245704174042, + 0.0051194773986935616, + 0.0026515270583331585, + 0.002644347958266735, + 0.017285345122218132, + 0.019285375252366066, + 0.014098909683525562, + 0.03807404264807701, + 0.02744314633309841, + 0.03424407169222832, + 0.019044701009988785, + 0.01365368627011776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.004638507030904293 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261284232139587 + ], + "opposite_logprob_tokens": [ + -0.5761284232139587 + ], + "logprob_reward": -0.8261284232139587, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261284232139587, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09637860208749771, + 0.027100810781121254, + 0.059861887246370316, + 0.13799215853214264, + 0.3111121654510498, + 0.11150914430618286, + 0.0915282666683197, + 0.01969209685921669, + 0.03738638013601303, + 0.10743846744298935 + ], + "credit_map": [ + 0.010123614221811295, + 0.014409070834517479, + 0.008137227036058903, + 0.00924750603735447, + 0.012213451787829399, + 0.013945380225777626, + 0.013441966846585274, + 0.014860386028885841, + 0.0032495995983481407, + 0.004061547573655844, + 0.0025307689793407917, + 0.0037275240756571293, + 0.003125180257484317, + 0.003362185787409544, + 0.003252547699958086, + 0.0037914556451141834, + 0.006460499484091997, + 0.009094015695154667, + 0.005367622710764408, + 0.009403984062373638, + 0.006611938588321209, + 0.008305481635034084, + 0.006548197474330664, + 0.008070147596299648, + 0.014143572188913822, + 0.019304275512695312, + 0.007755936123430729, + 0.022456876933574677, + 0.016519615426659584, + 0.018718736246228218, + 0.018634948879480362, + 0.020458199083805084, + 0.028641901910305023, + 0.021908333525061607, + 0.01940653659403324, + 0.04445941746234894, + 0.029490657150745392, + 0.04396607354283333, + 0.04512863606214523, + 0.07811062037944794, + 0.01179563533514738, + 0.010507861152291298, + 0.008093921467661858, + 0.026512062177062035, + 0.013291917741298676, + 0.013607845641672611, + 0.010641410946846008, + 0.017058493569493294, + 0.01119286846369505, + 0.010461220517754555, + 0.007188514340668917, + 0.015219295397400856, + 0.011225283145904541, + 0.011915276758372784, + 0.012143799103796482, + 0.012182004749774933, + 0.0022857398726046085, + 0.0020595635287463665, + 0.001614851295016706, + 0.00243290769867599, + 0.003305393736809492, + 0.00285897683352232, + 0.002663991181179881, + 0.002470671897754073, + 0.004853474907577038, + 0.0050466228276491165, + 0.0029339243192225695, + 0.004704234190285206, + 0.005172823555767536, + 0.006027036812156439, + 0.004773170221596956, + 0.003875090042129159, + 0.009746638126671314, + 0.011397291906177998, + 0.007803065236657858, + 0.0228122565895319, + 0.016054749488830566, + 0.018297478556632996, + 0.012448069639503956, + 0.008878921158611774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.8261284232139587 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium close-up, at eye level, slowly pans to a man with short, dark hair, his face frozen in shock. He wears a crisp white button-down shirt, its brightness contrasting sharply with the dim, ambiguous interior. The primary light source, a cool digital blue tint from off-camera, casts sharp, dramatic shadows across his features and creates a high-contrast, tense atmosphere of suspense. His eyes are wide, reflecting an invisible blue light, his eyebrows raised in disbelief, and his mouth slightly open, as if caught in a silent gasp. The background is completely out of focus, reduced to deep, soft shadows, isolating the man and heightening the sense of suspense, locking the viewer in his sudden, terrifying realization.", + "seed": 16050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_7400_seed_16050_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1558365186676383, + "vlm_logprob": -0.1558365186676383, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.1558365186676383, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0068216766230762005 + ], + "opposite_logprob_tokens": [ + -5.756821632385254 + ], + "logprob_reward": -0.0068216766230762005, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0068216766230762005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5525319576263428, + 0.21145008504390717, + 0.03835364431142807, + 0.029262349009513855, + 0.025489788502454758, + 0.02149321883916855, + 0.018753930926322937, + 0.0297881867736578, + 0.012839708477258682, + 0.060037098824977875 + ], + "credit_map": [ + 0.06099003180861473, + 0.09919619560241699, + 0.09544160962104797, + 0.05989683419466019, + 0.051509276032447815, + 0.10821419954299927, + 0.030021658167243004, + 0.04726220294833183, + 0.017841391265392303, + 0.026839176192879677, + 0.04096253961324692, + 0.03230131044983864, + 0.025767864659428596, + 0.017603816464543343, + 0.022980445995926857, + 0.027153542265295982, + 0.0036836082581430674, + 0.003953743260353804, + 0.00632076570764184, + 0.006106290966272354, + 0.004415447823703289, + 0.003940916620194912, + 0.0046149967238307, + 0.005317880306392908, + 0.0029419290367513895, + 0.0031976448372006416, + 0.0044271331280469894, + 0.005070067010819912, + 0.003560034791007638, + 0.002177290851250291, + 0.003521640319377184, + 0.004366611130535603, + 0.003360050730407238, + 0.00313315587118268, + 0.004549870733171701, + 0.003683648072183132, + 0.003290493506938219, + 0.0022706002928316593, + 0.0021502389572560787, + 0.0030517305713146925, + 0.002949462039396167, + 0.002561486791819334, + 0.004538514651358128, + 0.0020404283422976732, + 0.0030575115233659744, + 0.0019276448292657733, + 0.002133401809260249, + 0.0022847699001431465, + 0.0022884251084178686, + 0.002300697611644864, + 0.0027872563805431128, + 0.0020931113976985216, + 0.002384796505793929, + 0.0030920859426259995, + 0.0013973887544125319, + 0.002410170156508684, + 0.002998729469254613, + 0.003987445496022701, + 0.004353486001491547, + 0.004377513192594051, + 0.0036532345693558455, + 0.0037883988115936518, + 0.0030783729162067175, + 0.0035510079469531775, + 0.0015165973454713821, + 0.0013837672304362059, + 0.0019012283300980926, + 0.0015216423198580742, + 0.001399167929776013, + 0.0017222805181518197, + 0.0013144111726433039, + 0.002080614445731044, + 0.012846395373344421, + 0.004198863636702299, + 0.022202689200639725, + 0.003597990609705448, + 0.004675912205129862, + 0.0038572032935917377, + 0.0036226685624569654, + 0.005035376641899347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0068216766230762005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005269803572446108 + ], + "opposite_logprob_tokens": [ + -5.255270004272461 + ], + "logprob_reward": -0.005269803572446108, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005269803572446108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.542468249797821, + 0.06291399151086807, + 0.028198812156915665, + 0.04883335530757904, + 0.03605556860566139, + 0.04048764705657959, + 0.020488716661930084, + 0.049065012484788895, + 0.016584498807787895, + 0.15490415692329407 + ], + "credit_map": [ + 0.061889294534921646, + 0.10099354386329651, + 0.08954299241304398, + 0.05295083671808243, + 0.05280296877026558, + 0.11119119077920914, + 0.028082845732569695, + 0.04501459002494812, + 0.006227015983313322, + 0.010763615369796753, + 0.015105098485946655, + 0.007819789461791515, + 0.006496177986264229, + 0.004279556218534708, + 0.007179283071309328, + 0.005043453071266413, + 0.002921221312135458, + 0.0040643224492669106, + 0.0035714644473046064, + 0.0035759611055254936, + 0.004689205437898636, + 0.0031808519270271063, + 0.0031893164850771427, + 0.003006469924002886, + 0.005156317725777626, + 0.0050030904822051525, + 0.006804922129958868, + 0.008892933838069439, + 0.005544492043554783, + 0.004405064973980188, + 0.005390899255871773, + 0.0076356385834515095, + 0.0048846169374883175, + 0.004672926850616932, + 0.005578865762799978, + 0.0048048472963273525, + 0.005390134174376726, + 0.003755072131752968, + 0.0034280098043382168, + 0.003541095880791545, + 0.004214483313262463, + 0.005118616856634617, + 0.006761630065739155, + 0.006540501490235329, + 0.005129639059305191, + 0.0033074221573770046, + 0.004800186958163977, + 0.004615170415490866, + 0.0028512715362012386, + 0.002785170217975974, + 0.0035135122016072273, + 0.0025207598228007555, + 0.0023253438994288445, + 0.002418552991002798, + 0.0019154574256390333, + 0.0021586474031209946, + 0.0062810443341732025, + 0.005527179688215256, + 0.00870948750525713, + 0.006762166507542133, + 0.006477754563093185, + 0.0046449932269752026, + 0.005349075421690941, + 0.005313308909535408, + 0.0018890696810558438, + 0.0018286349950358272, + 0.0028900261968374252, + 0.0020086201839149, + 0.002130074892193079, + 0.001999079482629895, + 0.0016184209380298853, + 0.002220572205260396, + 0.032636091113090515, + 0.010856684297323227, + 0.06257660686969757, + 0.007287519983947277, + 0.007947809994220734, + 0.010164028033614159, + 0.007183551322668791, + 0.016251858323812485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.005269803572446108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464135454967618 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464135454967618, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464135454967618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32214146852493286, + 0.14351974427700043, + 0.15425913035869598, + 0.06250594556331635, + 0.09741267561912537, + 0.057850878685712814, + 0.031596265733242035, + 0.025745924562215805, + 0.023485364392399788, + 0.08148255199193954 + ], + "credit_map": [ + 0.03170160576701164, + 0.0567505918443203, + 0.05467528477311134, + 0.030811740085482597, + 0.03285432234406471, + 0.06171528995037079, + 0.018360301852226257, + 0.035272352397441864, + 0.012092511169612408, + 0.018064789474010468, + 0.03089003637433052, + 0.02249201014637947, + 0.0143608832731843, + 0.01234501414000988, + 0.015235664322972298, + 0.018038859590888023, + 0.014824213460087776, + 0.016713637858629227, + 0.027696410194039345, + 0.022645311430096626, + 0.01907917484641075, + 0.016285164281725883, + 0.016331536695361137, + 0.02068370021879673, + 0.007176415994763374, + 0.007291869726032019, + 0.008856051601469517, + 0.010612974874675274, + 0.007750042248517275, + 0.006313754711300135, + 0.006320094224065542, + 0.008184749633073807, + 0.011602348648011684, + 0.01136036403477192, + 0.01565537601709366, + 0.017183201387524605, + 0.010942717082798481, + 0.008604094386100769, + 0.008239034563302994, + 0.013825546018779278, + 0.00714943977072835, + 0.007224244996905327, + 0.009974629618227482, + 0.008628985844552517, + 0.00645401980727911, + 0.004946468863636255, + 0.005019553937017918, + 0.00845354050397873, + 0.004379549063742161, + 0.003981538582593203, + 0.00581755768507719, + 0.003595709102228284, + 0.0037360694259405136, + 0.004091840703040361, + 0.0026936177164316177, + 0.0033003862481564283, + 0.0036774545442312956, + 0.003449982963502407, + 0.00494187418371439, + 0.0020827387925237417, + 0.0034175871405750513, + 0.003417197149246931, + 0.002053883858025074, + 0.0027052077930420637, + 0.0027085337787866592, + 0.003020497737452388, + 0.0027842314448207617, + 0.003328013001009822, + 0.002957315882667899, + 0.003570114728063345, + 0.002433353802189231, + 0.002683305647224188, + 0.009221933782100677, + 0.006611791905015707, + 0.024257224053144455, + 0.006731738336384296, + 0.006436636671423912, + 0.006593438796699047, + 0.0075279725715518, + 0.014101826585829258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00464135454967618 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046377950347959995 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.0046377950347959995, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046377950347959995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6499593257904053, + 0.07327980548143387, + 0.09554346650838852, + 0.025434421375393867, + 0.020984994247555733, + 0.03404532000422478, + 0.01159234344959259, + 0.008178997784852982, + 0.023013440892100334, + 0.05796797201037407 + ], + "credit_map": [ + 0.07012617588043213, + 0.12103626877069473, + 0.10634998232126236, + 0.06310932338237762, + 0.0643879622220993, + 0.1323736160993576, + 0.036407470703125, + 0.05616842955350876, + 0.005730452947318554, + 0.008908660151064396, + 0.014409786090254784, + 0.012417932040989399, + 0.007861601188778877, + 0.006483528763055801, + 0.008001815527677536, + 0.009466025047004223, + 0.008655055426061153, + 0.01014725025743246, + 0.015475957654416561, + 0.016072651371359825, + 0.00959395244717598, + 0.009960646741092205, + 0.01102738082408905, + 0.014610568061470985, + 0.0034259778913110495, + 0.003079125424847007, + 0.003758952021598816, + 0.0038853329606354237, + 0.0030418639071285725, + 0.0025841468013823032, + 0.002335200784727931, + 0.0033238190226256847, + 0.0022667732555419207, + 0.0023901916574686766, + 0.0034985158126801252, + 0.0025509612169116735, + 0.0032573044300079346, + 0.002154087647795677, + 0.0025910998228937387, + 0.0022760583087801933, + 0.003632989712059498, + 0.004347902722656727, + 0.0055182925425469875, + 0.005295044742524624, + 0.003708182368427515, + 0.0031701172702014446, + 0.0037392894737422466, + 0.004633497912436724, + 0.0016043627401813865, + 0.0017817403422668576, + 0.0017696196446195245, + 0.0013260418782010674, + 0.0013683694414794445, + 0.0015087280189618468, + 0.001101459376513958, + 0.0011320203775539994, + 0.000803719274699688, + 0.0010655492078512907, + 0.001037186710163951, + 0.0011697688605636358, + 0.0010613417252898216, + 0.0008892188197933137, + 0.0009503092733211815, + 0.0012019029818475246, + 0.0025902274064719677, + 0.002973532536998391, + 0.003364695468917489, + 0.0029125420842319727, + 0.0026773589197546244, + 0.002378859557211399, + 0.0030930694192647934, + 0.0030231543350964785, + 0.011384774930775166, + 0.004065499175339937, + 0.022161567583680153, + 0.0034453230910003185, + 0.003604603698477149, + 0.004160391632467508, + 0.003577152034267783, + 0.005568659398704767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0046377950347959995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.757811963558197 + ], + "opposite_logprob_tokens": [ + -0.632811963558197 + ], + "logprob_reward": -0.757811963558197, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.757811963558197, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6328569650650024, + 0.02149883098900318, + 0.07292575389146805, + 0.02410188689827919, + 0.013519699685275555, + 0.02287980355322361, + 0.015023364685475826, + 0.01619495078921318, + 0.015571913681924343, + 0.1654268205165863 + ], + "credit_map": [ + 0.06642631441354752, + 0.12199567258358002, + 0.10324021428823471, + 0.057225171476602554, + 0.06523647904396057, + 0.1277809888124466, + 0.03642493113875389, + 0.054527219384908676, + 0.0017074079951271415, + 0.0036310856230556965, + 0.004321278538554907, + 0.002738406416028738, + 0.0025493118446320295, + 0.0012962009059265256, + 0.002410103799775243, + 0.0028450367972254753, + 0.006956947036087513, + 0.008068929426372051, + 0.012745647691190243, + 0.010731101036071777, + 0.008306846022605896, + 0.008602791465818882, + 0.007513565011322498, + 0.00999993272125721, + 0.003019708674401045, + 0.003295203670859337, + 0.003164854599162936, + 0.0038419440388679504, + 0.002771432977169752, + 0.002749661449342966, + 0.002249615266919136, + 0.0030094669200479984, + 0.0016749755013734102, + 0.001568125793710351, + 0.0015447053592652082, + 0.001990425167605281, + 0.0023814369924366474, + 0.0014386939583346248, + 0.0012237936025485396, + 0.0016975423786789179, + 0.002720389747992158, + 0.00305029540322721, + 0.0036999136209487915, + 0.003568764543160796, + 0.0030111761298030615, + 0.0019052885472774506, + 0.002105506369844079, + 0.002818469889461994, + 0.001857379567809403, + 0.0019111596047878265, + 0.002277553780004382, + 0.0013109524734318256, + 0.0021824734285473824, + 0.0023086220026016235, + 0.0014322607312351465, + 0.0017429626313969493, + 0.001997188199311495, + 0.0019728762563318014, + 0.0030653849244117737, + 0.0018969413358718157, + 0.002123250626027584, + 0.0018388991011306643, + 0.0016413568519055843, + 0.0016590534942224622, + 0.002260905224829912, + 0.0017603153828531504, + 0.0032575507648289204, + 0.0016924942610785365, + 0.0015169803518801928, + 0.002182767726480961, + 0.001499377191066742, + 0.0014015224296599627, + 0.020524127408862114, + 0.014257236383855343, + 0.04892906919121742, + 0.018556300550699234, + 0.01480410061776638, + 0.012106886133551598, + 0.01341555267572403, + 0.022833555936813354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.757811963558197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium close-up, at eye level, slowly pans to a man with short, dark hair, his face frozen in shock. He wears a crisp white button-down shirt, its brightness contrasting sharply with the dim, ambiguous interior. The primary light source, a cool digital blue tint from off-camera, casts sharp, dramatic shadows across his features and creates a high-contrast, tense atmosphere of suspense. His eyes are wide, reflecting an invisible blue light, his eyebrows raised in disbelief, and his mouth slightly open, as if caught in a silent gasp. The background is completely out of focus, reduced to deep, soft shadows, isolating the man and heightening the sense of suspense, locking the viewer in his sudden, terrifying realization.", + "seed": 16051, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_7400_seed_16051_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18779186955653132, + "vlm_logprob": -0.18779186955653132, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.18779186955653132, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00682061119005084 + ], + "opposite_logprob_tokens": [ + -5.7568206787109375 + ], + "logprob_reward": -0.00682061119005084, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00682061119005084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15301787853240967, + 0.11418081820011139, + 0.11919838190078735, + 0.23712320625782013, + 0.060325946658849716, + 0.07312088459730148, + 0.06282845884561539, + 0.08147941529750824, + 0.041376225650310516, + 0.0573488213121891 + ], + "credit_map": [ + 0.014163411222398281, + 0.01242880616337061, + 0.02780623361468315, + 0.019361132755875587, + 0.020747100934386253, + 0.017098363488912582, + 0.01717432402074337, + 0.024238502606749535, + 0.009849563241004944, + 0.01047914195805788, + 0.014164096675813198, + 0.023021753877401352, + 0.014804697595536709, + 0.011308187618851662, + 0.014185414649546146, + 0.0163679588586092, + 0.009651784785091877, + 0.008128426969051361, + 0.02091558277606964, + 0.014018736779689789, + 0.012419580481946468, + 0.02851875126361847, + 0.013922571204602718, + 0.011622948572039604, + 0.024601899087429047, + 0.018342021852731705, + 0.028754377737641335, + 0.03938133642077446, + 0.019948936998844147, + 0.028593003749847412, + 0.028734639286994934, + 0.048767007887363434, + 0.005193020217120647, + 0.004950067028403282, + 0.009549391455948353, + 0.008515200577676296, + 0.005813848692923784, + 0.012667075730860233, + 0.007314146962016821, + 0.0063231950625777245, + 0.008036354556679726, + 0.008061121217906475, + 0.010924309492111206, + 0.00932625588029623, + 0.006749759428203106, + 0.009107578545808792, + 0.010829580947756767, + 0.010085919871926308, + 0.0080558555200696, + 0.00616078358143568, + 0.008431876078248024, + 0.01054556667804718, + 0.005719366949051619, + 0.008401677943766117, + 0.00836960133165121, + 0.007143734022974968, + 0.007063628640025854, + 0.007391209248453379, + 0.012883763760328293, + 0.011891471222043037, + 0.008135100826621056, + 0.015710551291704178, + 0.009524370543658733, + 0.008879316039383411, + 0.005370889324694872, + 0.004252609796822071, + 0.006279738154262304, + 0.0035610287450253963, + 0.0037874667905271053, + 0.008057680912315845, + 0.0043302359990775585, + 0.005736571736633778, + 0.006631155963987112, + 0.005036319140344858, + 0.008588186465203762, + 0.010439706966280937, + 0.004992224276065826, + 0.008769737556576729, + 0.006807946600019932, + 0.006083543412387371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.00682061119005084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369184255599976 + ], + "opposite_logprob_tokens": [ + -3.651369094848633 + ], + "logprob_reward": -0.026369184255599976, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369184255599976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2636229395866394, + 0.24529984593391418, + 0.051335811614990234, + 0.059876374900341034, + 0.04765668511390686, + 0.12585535645484924, + 0.07321183383464813, + 0.08712401241064072, + 0.0211283341050148, + 0.024888796731829643 + ], + "credit_map": [ + 0.02898358181118965, + 0.02078966237604618, + 0.06088804081082344, + 0.02328743040561676, + 0.03526119142770767, + 0.03510558605194092, + 0.021393578499555588, + 0.03791384771466255, + 0.02584904059767723, + 0.019515689462423325, + 0.06047091633081436, + 0.019013267010450363, + 0.02331889420747757, + 0.040492892265319824, + 0.023574570193886757, + 0.03306456282734871, + 0.004635397344827652, + 0.0048767514526844025, + 0.00715031661093235, + 0.006169007159769535, + 0.006664551794528961, + 0.008093282580375671, + 0.006808994337916374, + 0.006937509868294001, + 0.005691232159733772, + 0.005705073941498995, + 0.005602905992418528, + 0.008527582511305809, + 0.005848786793649197, + 0.007123523857444525, + 0.007500069681555033, + 0.013877196237444878, + 0.0054822578094899654, + 0.0038185776211321354, + 0.006580738350749016, + 0.005680748261511326, + 0.00484691746532917, + 0.008664152584969997, + 0.00591557240113616, + 0.006667723413556814, + 0.015447486191987991, + 0.010190465487539768, + 0.018768133595585823, + 0.014687505550682545, + 0.009652490727603436, + 0.018369480967521667, + 0.01905210129916668, + 0.019687702879309654, + 0.007059088908135891, + 0.0069764661602675915, + 0.012869315221905708, + 0.013072749599814415, + 0.005443938542157412, + 0.00938674807548523, + 0.010809440165758133, + 0.007594086695462465, + 0.009105375036597252, + 0.007352697663009167, + 0.01378218736499548, + 0.013404016382992268, + 0.007231314200907946, + 0.012159922160208225, + 0.011821706779301167, + 0.012266796082258224, + 0.0023556877858936787, + 0.0018384406575933099, + 0.004246152937412262, + 0.002736919792369008, + 0.0015956562710925937, + 0.002907539252191782, + 0.0024978260044008493, + 0.0029501093085855246, + 0.002588849514722824, + 0.003236243734136224, + 0.0035734842531383038, + 0.002606517868116498, + 0.002431738656014204, + 0.0030747600831091404, + 0.003844890743494034, + 0.0035323118790984154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.026369184255599976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022003266494721174 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.0022003266494721174, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022003266494721174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18708586692810059, + 0.22453100979328156, + 0.05385145545005798, + 0.062279652804136276, + 0.08595714718103409, + 0.06773193925619125, + 0.17391307651996613, + 0.05554483085870743, + 0.04843056946992874, + 0.04067446291446686 + ], + "credit_map": [ + 0.03285471349954605, + 0.008960646577179432, + 0.06128346174955368, + 0.015258117578923702, + 0.013681520707905293, + 0.03038109838962555, + 0.010435990057885647, + 0.014230323024094105, + 0.02808198891580105, + 0.01234749797731638, + 0.0621294267475605, + 0.031312644481658936, + 0.01613183505833149, + 0.037958819419145584, + 0.017907792702317238, + 0.01866101287305355, + 0.004991550929844379, + 0.0028650767635554075, + 0.012396784499287605, + 0.003387361066415906, + 0.00559248449280858, + 0.01709389127790928, + 0.003421838628128171, + 0.004102466162294149, + 0.0060333930887281895, + 0.004540478345006704, + 0.00828008633106947, + 0.010351627133786678, + 0.006361854262650013, + 0.008846016600728035, + 0.008514071814715862, + 0.009352128021419048, + 0.008584500290453434, + 0.005787396803498268, + 0.01745939813554287, + 0.012008074671030045, + 0.006386868190020323, + 0.01567894034087658, + 0.008504712022840977, + 0.011547255329787731, + 0.007990200072526932, + 0.0053092059679329395, + 0.00964417215436697, + 0.008652911521494389, + 0.005857965908944607, + 0.014690077863633633, + 0.008516917936503887, + 0.007070489693433046, + 0.024145497009158134, + 0.01777908205986023, + 0.025184938684105873, + 0.0256704930216074, + 0.014326357282698154, + 0.026758531108498573, + 0.017207864671945572, + 0.022840308025479317, + 0.006062185391783714, + 0.00474588805809617, + 0.009514504112303257, + 0.007224320899695158, + 0.004959615413099527, + 0.011377566494047642, + 0.005864250939339399, + 0.005796498619019985, + 0.005743780639022589, + 0.003959209658205509, + 0.0097732525318861, + 0.006092563271522522, + 0.0036040653940290213, + 0.007279783021658659, + 0.0059383325278759, + 0.006039578467607498, + 0.004024871625006199, + 0.003240726189687848, + 0.007288984954357147, + 0.006307986564934254, + 0.003360204165801406, + 0.0065538608469069, + 0.005986581090837717, + 0.003911247942596674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0022003266494721174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052535575814545155 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.0052535575814545155, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052535575814545155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21832476556301117, + 0.1316547393798828, + 0.0984136089682579, + 0.13229364156723022, + 0.0916399136185646, + 0.04999249055981636, + 0.058413971215486526, + 0.053905051201581955, + 0.041003305464982986, + 0.12435851991176605 + ], + "credit_map": [ + 0.03204498812556267, + 0.013308092951774597, + 0.05309424176812172, + 0.017745366320014, + 0.025228101760149002, + 0.025261852890253067, + 0.024897268041968346, + 0.026744846254587173, + 0.01690828800201416, + 0.008981440216302872, + 0.03046814724802971, + 0.01781206950545311, + 0.013127586804330349, + 0.016759201884269714, + 0.013313200324773788, + 0.014284802600741386, + 0.008484803140163422, + 0.006721521262079477, + 0.020673831924796104, + 0.00667174207046628, + 0.011759039014577866, + 0.02430140972137451, + 0.009169443510472775, + 0.010631815530359745, + 0.018037350848317146, + 0.008272306062281132, + 0.018633538857102394, + 0.016312913969159126, + 0.010533169843256474, + 0.02605622448027134, + 0.01384375523775816, + 0.020604386925697327, + 0.008648747578263283, + 0.00657234899699688, + 0.0168480034917593, + 0.011639718897640705, + 0.008134597912430763, + 0.020883463323116302, + 0.008981388993561268, + 0.009931639768183231, + 0.00517354067414999, + 0.004169001244008541, + 0.008196932263672352, + 0.005915102083235979, + 0.003796856151893735, + 0.011514008976519108, + 0.005189850926399231, + 0.006037199404090643, + 0.007860115729272366, + 0.005689949728548527, + 0.009589038789272308, + 0.007657970767468214, + 0.00517865177243948, + 0.008850165642797947, + 0.0068666706793010235, + 0.006721407640725374, + 0.005340673960745335, + 0.004942868836224079, + 0.009633345529437065, + 0.003753694472834468, + 0.0041848463006317616, + 0.016032185405492783, + 0.004577008541673422, + 0.0054404293186962605, + 0.004004088696092367, + 0.003463923931121826, + 0.0062182084657251835, + 0.005620610434561968, + 0.0034229662269353867, + 0.007384464610368013, + 0.006029150448739529, + 0.004859895445406437, + 0.014682373963296413, + 0.011563633568584919, + 0.018594207242131233, + 0.0182467270642519, + 0.01110466755926609, + 0.020974252372980118, + 0.016867907717823982, + 0.012324752286076546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0052535575814545155 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983156681060791 + ], + "opposite_logprob_tokens": [ + -0.5233156681060791 + ], + "logprob_reward": -0.8983156681060791, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983156681060791, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.31396734714508057, + 0.11568459868431091, + 0.08969439566135406, + 0.07032683491706848, + 0.09569132328033447, + 0.08076547831296921, + 0.03996461257338524, + 0.02347593382000923, + 0.07276072353124619, + 0.09766878187656403 + ], + "credit_map": [ + 0.04995602369308472, + 0.0155180050060153, + 0.11120446026325226, + 0.021747255697846413, + 0.019996900111436844, + 0.0548575185239315, + 0.018220754340291023, + 0.022466402500867844, + 0.013284249231219292, + 0.008478162810206413, + 0.02566022239625454, + 0.0169189665466547, + 0.008498058654367924, + 0.014033270999789238, + 0.012191361747682095, + 0.016620296984910965, + 0.010130885057151318, + 0.005998167674988508, + 0.016581084579229355, + 0.0059363399632275105, + 0.00863182544708252, + 0.025012820959091187, + 0.0076446752063930035, + 0.009758598171174526, + 0.007215537130832672, + 0.0050997077487409115, + 0.013398785144090652, + 0.00565780233591795, + 0.005975053645670414, + 0.016254765912890434, + 0.006560798268765211, + 0.01016437727957964, + 0.0100888442248106, + 0.007947452366352081, + 0.017028605565428734, + 0.01205749623477459, + 0.008890718221664429, + 0.014405234716832638, + 0.011836540885269642, + 0.013436430133879185, + 0.01185875479131937, + 0.006749570369720459, + 0.013363405130803585, + 0.011381536722183228, + 0.005986272357404232, + 0.010157281532883644, + 0.011149181053042412, + 0.010119468905031681, + 0.0038954121991991997, + 0.0034681069664657116, + 0.006222106050699949, + 0.007012183777987957, + 0.0030871687922626734, + 0.005149343982338905, + 0.0057466947473585606, + 0.005383593030273914, + 0.0027207043021917343, + 0.002042461419478059, + 0.0035870408173650503, + 0.0035200617276132107, + 0.0026019467040896416, + 0.0031067903619259596, + 0.003283443860709667, + 0.002613482531160116, + 0.0076849800534546375, + 0.005543661769479513, + 0.01208561286330223, + 0.009140649810433388, + 0.005827434826642275, + 0.0163497943431139, + 0.008756565861403942, + 0.0073720188811421394, + 0.00995616801083088, + 0.007710064295679331, + 0.010920451022684574, + 0.01617933250963688, + 0.007939712144434452, + 0.011619302444159985, + 0.01427567470818758, + 0.019068075343966484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.8983156681060791 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium close-up, at eye level, slowly pans to a man with short, dark hair, his face frozen in shock. He wears a crisp white button-down shirt, its brightness contrasting sharply with the dim, ambiguous interior. The primary light source, a cool digital blue tint from off-camera, casts sharp, dramatic shadows across his features and creates a high-contrast, tense atmosphere of suspense. His eyes are wide, reflecting an invisible blue light, his eyebrows raised in disbelief, and his mouth slightly open, as if caught in a silent gasp. The background is completely out of focus, reduced to deep, soft shadows, isolating the man and heightening the sense of suspense, locking the viewer in his sudden, terrifying realization.", + "seed": 16052, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_6/index_7400_seed_16052_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19879783163778483, + "vlm_logprob": -0.19879783163778483, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.19879783163778483, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006825583986938 + ], + "opposite_logprob_tokens": [ + -5.7568254470825195 + ], + "logprob_reward": -0.006825583986938, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.006825583986938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3000331521034241, + 0.07274428755044937, + 0.05743204429745674, + 0.05990839749574661, + 0.0872936099767685, + 0.11497177183628082, + 0.07829909771680832, + 0.05461287125945091, + 0.04568672180175781, + 0.12901803851127625 + ], + "credit_map": [ + 0.031733959913253784, + 0.03347700834274292, + 0.039827700704336166, + 0.023413604125380516, + 0.05871035158634186, + 0.042366474866867065, + 0.02925618179142475, + 0.041247881948947906, + 0.008261535316705704, + 0.008147894404828548, + 0.00874828826636076, + 0.004993550479412079, + 0.012760084122419357, + 0.009382303804159164, + 0.01052585057914257, + 0.009924775920808315, + 0.006348667200654745, + 0.007062372751533985, + 0.006574499886482954, + 0.004301599692553282, + 0.009504678659141064, + 0.008405337110161781, + 0.008246136829257011, + 0.006988750770688057, + 0.004965262953191996, + 0.007347918581217527, + 0.008547625504434109, + 0.005690183490514755, + 0.008307014591991901, + 0.008912479504942894, + 0.011337600648403168, + 0.004800308961421251, + 0.009131385944783688, + 0.005426427815109491, + 0.008091411553323269, + 0.0053564137779176235, + 0.009465914219617844, + 0.013480735011398792, + 0.01336173340678215, + 0.022979583591222763, + 0.00902596302330494, + 0.01424719113856554, + 0.013403255492448807, + 0.009336624294519424, + 0.01881519891321659, + 0.017128173261880875, + 0.02243412844836712, + 0.010581237263977528, + 0.00680547533556819, + 0.007612673100084066, + 0.010965989902615547, + 0.0032690351363271475, + 0.010862832888960838, + 0.010750455781817436, + 0.009753580205142498, + 0.0182790569961071, + 0.005366120953112841, + 0.007536545395851135, + 0.006504840683192015, + 0.003675919957458973, + 0.00957068894058466, + 0.007134321611374617, + 0.008667334914207458, + 0.006157096475362778, + 0.006086689420044422, + 0.0047800480388104916, + 0.007383020594716072, + 0.0029125798027962446, + 0.006964995060116053, + 0.0067007048055529594, + 0.007051114924252033, + 0.0038075721822679043, + 0.018416007980704308, + 0.009316119365394115, + 0.03429700434207916, + 0.005962832365185022, + 0.013050793670117855, + 0.02010628953576088, + 0.010462792590260506, + 0.017406204715371132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.006825583986938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108204040676355 + ], + "opposite_logprob_tokens": [ + -5.504108428955078 + ], + "logprob_reward": -0.004108204040676355, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108204040676355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34038251638412476, + 0.11905360966920853, + 0.08783559501171112, + 0.11356412619352341, + 0.07374376803636551, + 0.15902870893478394, + 0.01760098524391651, + 0.035353101789951324, + 0.023311587050557137, + 0.030125953257083893 + ], + "credit_map": [ + 0.028791556134819984, + 0.03957740589976311, + 0.045133087784051895, + 0.029095156118273735, + 0.05056848004460335, + 0.04819565266370773, + 0.04019123315811157, + 0.058829955756664276, + 0.010233576409518719, + 0.011054765433073044, + 0.018978962674736977, + 0.015051133930683136, + 0.014365436509251595, + 0.01503738109022379, + 0.01649436168372631, + 0.017837993800640106, + 0.00742469122633338, + 0.013554119504988194, + 0.01027965359389782, + 0.006482740864157677, + 0.015248304232954979, + 0.012992430478334427, + 0.01266089454293251, + 0.009192760102450848, + 0.011038674041628838, + 0.009659367613494396, + 0.016684945672750473, + 0.006569727789610624, + 0.01269550435245037, + 0.016367390751838684, + 0.015282421372830868, + 0.025266099721193314, + 0.007073138374835253, + 0.0074219065718352795, + 0.00701272115111351, + 0.005993529688566923, + 0.013641934841871262, + 0.011031826958060265, + 0.015201842412352562, + 0.006366870831698179, + 0.013853431679308414, + 0.011852498166263103, + 0.015785368159413338, + 0.00905578676611185, + 0.017595907673239708, + 0.02249186858534813, + 0.02693653106689453, + 0.04145731404423714, + 0.001932640909217298, + 0.0017076795920729637, + 0.0026163216680288315, + 0.0010907360119745135, + 0.003028813749551773, + 0.00236006872728467, + 0.0031106257811188698, + 0.001754099503159523, + 0.0027075805701315403, + 0.004338695667684078, + 0.003813384100794792, + 0.0028429655358195305, + 0.006535928696393967, + 0.004755787085741758, + 0.005841271951794624, + 0.004517488647252321, + 0.0043539199978113174, + 0.0019084032392129302, + 0.0033471137285232544, + 0.0012831451604142785, + 0.0025608919095247984, + 0.0036957443226128817, + 0.0025863805785775185, + 0.003575987182557583, + 0.0036105357576161623, + 0.0029562590643763542, + 0.0051069664768874645, + 0.00225151632912457, + 0.0045368727296590805, + 0.004259159788489342, + 0.0038232882507145405, + 0.0035813553258776665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004108204040676355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028234163764864206 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.0028234163764864206, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028234163764864206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28588420152664185, + 0.15687114000320435, + 0.06525132060050964, + 0.04680213704705238, + 0.0724206194281578, + 0.11259064823389053, + 0.08014674484729767, + 0.07694867253303528, + 0.055226121097803116, + 0.04785830155014992 + ], + "credit_map": [ + 0.03156306594610214, + 0.030709365382790565, + 0.03556983545422554, + 0.02658735029399395, + 0.03585824742913246, + 0.043643854558467865, + 0.03351660817861557, + 0.04843590408563614, + 0.013613895513117313, + 0.019072989001870155, + 0.021802904084324837, + 0.013722731731832027, + 0.01841145195066929, + 0.023125767707824707, + 0.018417801707983017, + 0.028703615069389343, + 0.008855398744344711, + 0.006139106582850218, + 0.013467955403029919, + 0.004686166997998953, + 0.0069630504585802555, + 0.01203650701791048, + 0.007113359402865171, + 0.005989787168800831, + 0.00531034404411912, + 0.005696358159184456, + 0.005485546309500933, + 0.003403197741135955, + 0.00563832139596343, + 0.00662728538736701, + 0.008116114884614944, + 0.006524973548948765, + 0.00707129156216979, + 0.006531054154038429, + 0.005368844605982304, + 0.0036364973057061434, + 0.009991150349378586, + 0.011573572643101215, + 0.010214160196483135, + 0.018034061416983604, + 0.013718548230826855, + 0.010310827754437923, + 0.01336909644305706, + 0.006784704513847828, + 0.01353218499571085, + 0.017040571197867393, + 0.016572866588830948, + 0.021261852234601974, + 0.007676956243813038, + 0.009681103751063347, + 0.007941349409520626, + 0.0045734066516160965, + 0.013165535405278206, + 0.012440647929906845, + 0.012769602239131927, + 0.011898152530193329, + 0.005952903069555759, + 0.009456329979002476, + 0.009484023787081242, + 0.0056142951361835, + 0.012887440621852875, + 0.010462019592523575, + 0.013477619737386703, + 0.009614050388336182, + 0.00525519298389554, + 0.005327555350959301, + 0.006997270043939352, + 0.0046365042217075825, + 0.00848762970417738, + 0.008092171512544155, + 0.010322807356715202, + 0.0061069936491549015, + 0.006664888933300972, + 0.004886428825557232, + 0.009135288186371326, + 0.0027169419918209314, + 0.006186699960380793, + 0.006943976040929556, + 0.00565280532464385, + 0.005671278573572636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0028234163764864206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952132865786552 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952132865786552, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952132865786552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09271473437547684, + 0.05819643661379814, + 0.058440130203962326, + 0.10080170631408691, + 0.20725242793560028, + 0.136479452252388, + 0.03766850382089615, + 0.05684363469481468, + 0.0517311692237854, + 0.19987180829048157 + ], + "credit_map": [ + 0.011106817983090878, + 0.011079376563429832, + 0.01932232826948166, + 0.008514088578522205, + 0.012392228469252586, + 0.014457976445555687, + 0.008284416981041431, + 0.007557498756796122, + 0.004581584595143795, + 0.00560828298330307, + 0.010170510970056057, + 0.005038675386458635, + 0.009054698050022125, + 0.008329770527780056, + 0.007550126872956753, + 0.00786278210580349, + 0.005846336018294096, + 0.005562009755522013, + 0.007116883527487516, + 0.007269073277711868, + 0.00889875739812851, + 0.01026671938598156, + 0.008248236030340195, + 0.005232112016528845, + 0.00847545824944973, + 0.011960002593696117, + 0.013433285057544708, + 0.0071885897777974606, + 0.015674231573939323, + 0.016207758337259293, + 0.018414819613099098, + 0.009447559714317322, + 0.02606760337948799, + 0.018656238913536072, + 0.027161799371242523, + 0.01628219150006771, + 0.027976298704743385, + 0.028617095202207565, + 0.02871764823794365, + 0.03377354145050049, + 0.016724174842238426, + 0.012650122866034508, + 0.019839368760585785, + 0.007794204633682966, + 0.018516244366765022, + 0.01915806531906128, + 0.0164243932813406, + 0.025372881442308426, + 0.006170713342726231, + 0.0035093335900455713, + 0.0054448451846838, + 0.001994034741073847, + 0.005212632939219475, + 0.005436620209366083, + 0.003945206291973591, + 0.005955118220299482, + 0.005131130572408438, + 0.006604520138353109, + 0.007262448314577341, + 0.0043664551340043545, + 0.008669900707900524, + 0.007903067395091057, + 0.010475331917405128, + 0.006430782377719879, + 0.004761400632560253, + 0.006260607857257128, + 0.007142764516174793, + 0.004143976606428623, + 0.007345716468989849, + 0.006964806467294693, + 0.008986284956336021, + 0.006125611253082752, + 0.02860194817185402, + 0.015095697715878487, + 0.048626288771629333, + 0.009066550061106682, + 0.019114278256893158, + 0.02986658550798893, + 0.018221650272607803, + 0.03127879276871681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.005952132865786552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man with short, dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a crisp white button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cool digital blue tint from an off-camera source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely out of focus with deep, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742798209190369 + ], + "opposite_logprob_tokens": [ + -0.47427982091903687 + ], + "logprob_reward": -0.9742798209190369, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742798209190369, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2850937843322754, + 0.11505141854286194, + 0.05275050923228264, + 0.03944429010152817, + 0.057538438588380814, + 0.11540282517671585, + 0.09958019107580185, + 0.035696327686309814, + 0.05224053934216499, + 0.14720171689987183 + ], + "credit_map": [ + 0.030608009546995163, + 0.03360888361930847, + 0.048611756414175034, + 0.027180328965187073, + 0.037685975432395935, + 0.04207756742835045, + 0.02727952040731907, + 0.03804173693060875, + 0.010103166103363037, + 0.009508894756436348, + 0.017830979079008102, + 0.014931460842490196, + 0.015112394466996193, + 0.013030709698796272, + 0.018801555037498474, + 0.01573225110769272, + 0.005014269612729549, + 0.00649737985804677, + 0.005899554118514061, + 0.004457634408026934, + 0.009508783929049969, + 0.006677738390862942, + 0.009821062907576561, + 0.004874077159911394, + 0.006274873856455088, + 0.003996630664914846, + 0.005518019199371338, + 0.0022238027304410934, + 0.0048664831556379795, + 0.006272620055824518, + 0.004917255602777004, + 0.0053746020421385765, + 0.005852089263498783, + 0.0059007154777646065, + 0.006509802304208279, + 0.0033666198141872883, + 0.009382186457514763, + 0.009337705560028553, + 0.007796380668878555, + 0.009392930194735527, + 0.01108618639409542, + 0.010797855444252491, + 0.012178960256278515, + 0.005416610278189182, + 0.01361354161053896, + 0.015788281336426735, + 0.016243310645222664, + 0.03027806431055069, + 0.013794043101370335, + 0.009062455967068672, + 0.014978320337831974, + 0.006649426184594631, + 0.01343611441552639, + 0.014363860711455345, + 0.009585696272552013, + 0.01771026849746704, + 0.0033457656390964985, + 0.0033718447666615248, + 0.004242373164743185, + 0.003795079654082656, + 0.00626251520588994, + 0.005032392684370279, + 0.005251230672001839, + 0.004395123105496168, + 0.006278148852288723, + 0.006522262003272772, + 0.006592868827283382, + 0.0039183394983410835, + 0.007418719585984945, + 0.007429968565702438, + 0.00797083880752325, + 0.006109388079494238, + 0.023279482498764992, + 0.010979774408042431, + 0.03436146304011345, + 0.007829830050468445, + 0.01637120358645916, + 0.021524475887417793, + 0.01337415911257267, + 0.019481319934129715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.9742798209190369 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_tensor([4247, 4247])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_tensor([4247, 4247])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..557fb99819ef1b55583c80c36ca78131c1bc4a70 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_tensor([4247, 4247])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro close-up, captured from a low, ground-level angle, the camera remains stationary, focused on a delicate green sprout, newly emerging from the soil. The sprout, a vibrant, emerald green, its tip slightly bent from its vigorous growth. Tiny, translucent water droplets cling to its translucent stem, shimmering like miniature diamonds in the soft, diffuse morning light. Surrounding it is moist, dark brown, almost black soil, its grains coarse and distinct, its texture glossy from its moisture. With slow yet unmistakable force, the shoot pushes upward, causing tiny cracks to form around its roots, and tiny clumps of soil tumbling from its sides. The entire frame is enveloped by an extremely shallow depth of field, the background completely blurred, leaving the focus squarely on this tiny yet tenacious miracle of life. The delicate light and shadows, and the intense color contrasts create a grand atmosphere of tranquility, hope, and the resilience of life.", + "seed": 11172, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_4247_seed_11172_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05860965959727764, + "vlm_logprob": -0.05860965959727764, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.05860965959727764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0023316359147429466 + ], + "opposite_logprob_tokens": [ + -7.877331733703613 + ], + "logprob_reward": -0.0023316359147429466, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0023316359147429466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.193767711520195, + 0.058139197528362274, + 0.04410477727651596, + 0.14705167710781097, + 0.0678824931383133, + 0.09595496207475662, + 0.030420880764722824, + 0.1722211241722107, + 0.13424630463123322, + 0.05621090158820152 + ], + "credit_map": [ + 0.02653578296303749, + 0.017515230923891068, + 0.023508284240961075, + 0.02278749831020832, + 0.023740889504551888, + 0.03015791065990925, + 0.02741609886288643, + 0.022106017917394638, + 0.00698095140978694, + 0.006735485978424549, + 0.006248502992093563, + 0.005586354527622461, + 0.007337004411965609, + 0.009274771437048912, + 0.010423792526125908, + 0.005552330520004034, + 0.004940036218613386, + 0.0045974962413311005, + 0.0062955571338534355, + 0.0043099611066281796, + 0.005633747670799494, + 0.006921075284481049, + 0.0064649516716599464, + 0.004941953811794519, + 0.016899822279810905, + 0.015802111476659775, + 0.029655596241354942, + 0.016398735344409943, + 0.01883355900645256, + 0.01540463324636221, + 0.018777506425976753, + 0.015279706567525864, + 0.011261200532317162, + 0.00844377651810646, + 0.009253071621060371, + 0.007486244197934866, + 0.005402200389653444, + 0.008149323053658009, + 0.0074596949853003025, + 0.010426982305943966, + 0.011663621291518211, + 0.010768755339086056, + 0.012540210969746113, + 0.01111070066690445, + 0.0117812380194664, + 0.013482537120580673, + 0.012054446153342724, + 0.01255345530807972, + 0.0035924327094107866, + 0.0029464440885931253, + 0.003373041283339262, + 0.0033184965141117573, + 0.004406758584082127, + 0.005102597642689943, + 0.00343055697157979, + 0.0042505525052547455, + 0.017969980835914612, + 0.025690948590636253, + 0.01846059039235115, + 0.013182210735976696, + 0.038735128939151764, + 0.022533126175403595, + 0.019200123846530914, + 0.016449011862277985, + 0.016321295872330666, + 0.020273353904485703, + 0.012528992258012295, + 0.009022701531648636, + 0.03031759150326252, + 0.020605530589818954, + 0.012779971584677696, + 0.01239687018096447, + 0.008147950284183025, + 0.008720595389604568, + 0.005582340992987156, + 0.006083051208406687, + 0.007542280945926905, + 0.008156384341418743, + 0.006728136911988258, + 0.005250157322734594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0023316359147429466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210108667612076 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014210108667612076, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210108667612076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.251107394695282, + 0.0568360798060894, + 0.046977877616882324, + 0.07152215391397476, + 0.08834963291883469, + 0.03680919110774994, + 0.04020491614937782, + 0.0602944940328598, + 0.22952164709568024, + 0.11837656050920486 + ], + "credit_map": [ + 0.02645181678235531, + 0.023470401763916016, + 0.037849605083465576, + 0.025235675275325775, + 0.0267168078571558, + 0.039731331169605255, + 0.05256669595837593, + 0.01908508501946926, + 0.005703656934201717, + 0.006564651150256395, + 0.008268584497272968, + 0.005407072138041258, + 0.008641662076115608, + 0.005037762224674225, + 0.010247205384075642, + 0.006965490989387035, + 0.005092056002467871, + 0.005749398842453957, + 0.004707567393779755, + 0.004103973973542452, + 0.0099265081807971, + 0.004995258525013924, + 0.008026041090488434, + 0.004377075005322695, + 0.008218951523303986, + 0.006581739988178015, + 0.012794680893421173, + 0.008054410107433796, + 0.01050714310258627, + 0.009442984126508236, + 0.00928956363350153, + 0.0066326819360256195, + 0.010773604735732079, + 0.012337975203990936, + 0.011607468128204346, + 0.010177584365010262, + 0.00933397188782692, + 0.00937944557517767, + 0.015246524475514889, + 0.00949306134134531, + 0.006047695875167847, + 0.0038683461025357246, + 0.004709382075816393, + 0.0034170462749898434, + 0.002837774809449911, + 0.007335631642490625, + 0.004406902007758617, + 0.0041864123195409775, + 0.0049043376930058, + 0.0050241355784237385, + 0.005195767153054476, + 0.003091271501034498, + 0.008077423088252544, + 0.005717664957046509, + 0.004804142750799656, + 0.0033901764545589685, + 0.007509979419410229, + 0.008147141896188259, + 0.008889281190931797, + 0.005256429314613342, + 0.00809205137193203, + 0.008315286599099636, + 0.00734111899510026, + 0.006743209902197123, + 0.027002409100532532, + 0.036282461136579514, + 0.018980924040079117, + 0.014032945968210697, + 0.05152928829193115, + 0.03689160570502281, + 0.02601977065205574, + 0.018782256171107292, + 0.012682945467531681, + 0.013363044708967209, + 0.011604612693190575, + 0.011266740970313549, + 0.022879060357809067, + 0.015691706910729408, + 0.012903912924230099, + 0.017984531819820404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.014210108667612076 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295194178819656 + ], + "opposite_logprob_tokens": [ + -2.2379519939422607 + ], + "logprob_reward": -0.11295194178819656, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295194178819656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2774598002433777, + 0.048144131898880005, + 0.06166385859251022, + 0.10066574066877365, + 0.1610509604215622, + 0.0435606874525547, + 0.02828710712492466, + 0.09886229783296585, + 0.1272839605808258, + 0.05302147939801216 + ], + "credit_map": [ + 0.030429938808083534, + 0.03307874873280525, + 0.03462560102343559, + 0.026370538398623466, + 0.037615545094013214, + 0.07010839879512787, + 0.02400072105228901, + 0.021230315789580345, + 0.004911324940621853, + 0.005877761635929346, + 0.0074265641160309315, + 0.004252712242305279, + 0.00662694638594985, + 0.004634397570043802, + 0.007150572258979082, + 0.0072638485580682755, + 0.007920566014945507, + 0.00824758131057024, + 0.006312116980552673, + 0.006206892430782318, + 0.0069177718833088875, + 0.006340173538774252, + 0.010309248231351376, + 0.009409508667886257, + 0.011127007193863392, + 0.010770456865429878, + 0.013867286033928394, + 0.013559190556406975, + 0.01288177166134119, + 0.012536325491964817, + 0.014805321581661701, + 0.011118379421532154, + 0.020240463316440582, + 0.02185741625726223, + 0.021892834454774857, + 0.018593985587358475, + 0.016508741304278374, + 0.015845967456698418, + 0.026028502732515335, + 0.020083054900169373, + 0.00530313840135932, + 0.00559342373162508, + 0.004912029020488262, + 0.003950316458940506, + 0.009266900829970837, + 0.005058217328041792, + 0.005249351728707552, + 0.004227305296808481, + 0.0033266055397689342, + 0.004113435745239258, + 0.0028108437545597553, + 0.002577286446467042, + 0.004468672908842564, + 0.003913470543920994, + 0.003461680840700865, + 0.003615110879763961, + 0.011330824345350266, + 0.012019720859825611, + 0.014447467401623726, + 0.008805307559669018, + 0.01702272705733776, + 0.012574026361107826, + 0.008410933427512646, + 0.014251284301280975, + 0.015484022907912731, + 0.01998119056224823, + 0.010946713387966156, + 0.008239151909947395, + 0.026663852855563164, + 0.02099420316517353, + 0.013474494218826294, + 0.011500326916575432, + 0.006668861955404282, + 0.006764867343008518, + 0.006217489019036293, + 0.005134645849466324, + 0.007423693779855967, + 0.007106016390025616, + 0.005207312759011984, + 0.008498596027493477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.11295194178819656 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189953975379467 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003189953975379467, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189953975379467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25843939185142517, + 0.10215693712234497, + 0.06711038947105408, + 0.09111694991588593, + 0.08913251012563705, + 0.06687802076339722, + 0.05423683673143387, + 0.0914112776517868, + 0.11217000335454941, + 0.0673476904630661 + ], + "credit_map": [ + 0.03496326506137848, + 0.02966279909014702, + 0.03712699934840202, + 0.023047039285302162, + 0.029027121141552925, + 0.05257022753357887, + 0.02354435622692108, + 0.02849760837852955, + 0.011854104697704315, + 0.013887673616409302, + 0.010699410922825336, + 0.009779756888747215, + 0.012322572059929371, + 0.02313668467104435, + 0.008878907188773155, + 0.011597830802202225, + 0.008820130489766598, + 0.010541334748268127, + 0.012428468093276024, + 0.00591931352391839, + 0.004070906434208155, + 0.009030161425471306, + 0.005958701483905315, + 0.010341373272240162, + 0.013494286686182022, + 0.008460633456707, + 0.014114068821072578, + 0.011550634168088436, + 0.005132927559316158, + 0.020032618194818497, + 0.0074333264492452145, + 0.010898462496697903, + 0.011410095728933811, + 0.011168008670210838, + 0.013983718119561672, + 0.011936582624912262, + 0.006729607004672289, + 0.00825449824333191, + 0.012825061567127705, + 0.012824942357838154, + 0.007943300530314445, + 0.006997072137892246, + 0.00914078764617443, + 0.009325173683464527, + 0.007481561042368412, + 0.00807840283960104, + 0.007700398098677397, + 0.01021132804453373, + 0.007050186395645142, + 0.006674140691757202, + 0.005702919326722622, + 0.0050544049590826035, + 0.005790417082607746, + 0.008939177729189396, + 0.007315018214285374, + 0.00771057466045022, + 0.012964312918484211, + 0.013586188666522503, + 0.008735128678381443, + 0.006332689430564642, + 0.020690005272626877, + 0.012459810823202133, + 0.010560339316725731, + 0.006082812324166298, + 0.011922875419259071, + 0.01767417974770069, + 0.009358365088701248, + 0.0075136213563382626, + 0.026106305420398712, + 0.015680493786931038, + 0.013462841510772705, + 0.010451329872012138, + 0.012502418830990791, + 0.00863945297896862, + 0.010700223036110401, + 0.005320892669260502, + 0.007802844513207674, + 0.013831088319420815, + 0.0033864241559058428, + 0.005164349917322397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.003189953975379467 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036465764045715 + ], + "opposite_logprob_tokens": [ + -1.9103646278381348 + ], + "logprob_reward": -0.16036465764045715, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036465764045715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24755552411079407, + 0.03779441863298416, + 0.061135534197092056, + 0.1099076047539711, + 0.056853897869586945, + 0.06697943806648254, + 0.06774942576885223, + 0.03427068144083023, + 0.27218392491340637, + 0.04556949809193611 + ], + "credit_map": [ + 0.0252798181027174, + 0.04243358224630356, + 0.05045167729258537, + 0.027371274307370186, + 0.03547348082065582, + 0.02974100224673748, + 0.01822826638817787, + 0.018576430156826973, + 0.004830227233469486, + 0.0055824448354542255, + 0.004308005794882774, + 0.004331822507083416, + 0.0037216481287032366, + 0.005222458392381668, + 0.004381222650408745, + 0.005416587460786104, + 0.006575209088623524, + 0.006944785825908184, + 0.005530050955712795, + 0.007887630723416805, + 0.011019119061529636, + 0.00594330532476306, + 0.011280636303126812, + 0.0059547992423176765, + 0.01500152051448822, + 0.013026854023337364, + 0.015572762116789818, + 0.012950281612575054, + 0.008908754214644432, + 0.01998170092701912, + 0.01158355176448822, + 0.01288218330591917, + 0.005985893774777651, + 0.007969084195792675, + 0.007398924324661493, + 0.008409772999584675, + 0.007271481677889824, + 0.005320548079907894, + 0.008346177637577057, + 0.006152015645056963, + 0.008852862752974033, + 0.008506253361701965, + 0.006725787650793791, + 0.0068190693855285645, + 0.010319228284060955, + 0.007072040345519781, + 0.009276987053453922, + 0.009407205507159233, + 0.008412196300923824, + 0.009415821172297001, + 0.006080385763198137, + 0.005052649416029453, + 0.012740354053676128, + 0.010434183292090893, + 0.009333308786153793, + 0.006280525587499142, + 0.005033690016716719, + 0.0029517870862036943, + 0.004098235629498959, + 0.002765520941466093, + 0.004455608781427145, + 0.0070503102615475655, + 0.003944722469896078, + 0.003970808815211058, + 0.03075585700571537, + 0.045425646007061005, + 0.02282305806875229, + 0.01654134690761566, + 0.06298685818910599, + 0.0406784750521183, + 0.03212226182222366, + 0.020850440487265587, + 0.005300878081470728, + 0.007923617027699947, + 0.004690042231231928, + 0.0038567776791751385, + 0.008467815816402435, + 0.006692300084978342, + 0.004224409349262714, + 0.004413657356053591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.16036465764045715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a low, ground-level angle, the camera remains stationary, focused on a delicate green sprout, newly emerging from the soil. The sprout, a vibrant, emerald green, its tip slightly bent from its vigorous growth. Tiny, translucent water droplets cling to its translucent stem, shimmering like miniature diamonds in the soft, diffuse morning light. Surrounding it is moist, dark brown, almost black soil, its grains coarse and distinct, its texture glossy from its moisture. With slow yet unmistakable force, the shoot pushes upward, causing tiny cracks to form around its roots, and tiny clumps of soil tumbling from its sides. The entire frame is enveloped by an extremely shallow depth of field, the background completely blurred, leaving the focus squarely on this tiny yet tenacious miracle of life. The delicate light and shadows, and the intense color contrasts create a grand atmosphere of tranquility, hope, and the resilience of life.", + "seed": 11173, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_4247_seed_11173_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04755922053009272, + "vlm_logprob": -0.04755922053009272, + "vlm_yes_no_margin": 4.274999976158142, + "vlm_token_credit": -0.04755922053009272, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002986140549182892 + ], + "opposite_logprob_tokens": [ + -7.627985954284668 + ], + "logprob_reward": -0.002986140549182892, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.002986140549182892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2083979845046997, + 0.24226365983486176, + 0.05542697384953499, + 0.08953453600406647, + 0.05195259675383568, + 0.021937431767582893, + 0.04414176568388939, + 0.12371394783258438, + 0.058681126683950424, + 0.10394994914531708 + ], + "credit_map": [ + 0.03583921492099762, + 0.03855251893401146, + 0.02024783194065094, + 0.02165505290031433, + 0.015028283931314945, + 0.02209325321018696, + 0.03164439648389816, + 0.02333744615316391, + 0.04329676553606987, + 0.04879068210721016, + 0.026130547747015953, + 0.01867125928401947, + 0.028903018683195114, + 0.03342582657933235, + 0.023190148174762726, + 0.019855428487062454, + 0.010079937987029552, + 0.010869903489947319, + 0.004954913165420294, + 0.004749930929392576, + 0.006102108396589756, + 0.009395868517458439, + 0.004983376245945692, + 0.004290939774364233, + 0.00894041359424591, + 0.008376765996217728, + 0.018361758440732956, + 0.013103818520903587, + 0.006032510194927454, + 0.01195929478853941, + 0.006482709664851427, + 0.016277272254228592, + 0.007147539872676134, + 0.005969611927866936, + 0.008784995414316654, + 0.00859659630805254, + 0.004712914116680622, + 0.006836260203272104, + 0.0035262121818959713, + 0.006378465797752142, + 0.002176274312660098, + 0.0012857142137363553, + 0.005091513507068157, + 0.00415851641446352, + 0.002479067537933588, + 0.0023262551985681057, + 0.0017658567521721125, + 0.0026542365085333586, + 0.00491420179605484, + 0.004498254507780075, + 0.004568316973745823, + 0.007429489400237799, + 0.006369719747453928, + 0.006277771200984716, + 0.005075314547866583, + 0.0050086998380720615, + 0.015198448672890663, + 0.018752332776784897, + 0.021528637036681175, + 0.022586049512028694, + 0.011756145395338535, + 0.019071152433753014, + 0.007201545871794224, + 0.007619649171829224, + 0.005193009506911039, + 0.006456824019551277, + 0.011032811366021633, + 0.009489117190241814, + 0.006112862378358841, + 0.008022697642445564, + 0.006215465720742941, + 0.006158346775919199, + 0.014726277440786362, + 0.010820411145687103, + 0.011481226421892643, + 0.011418554931879044, + 0.009928595274686813, + 0.02512865699827671, + 0.009686199948191643, + 0.010760033503174782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.002986140549182892 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088416911661625 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088416911661625, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088416911661625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30536478757858276, + 0.0557403638958931, + 0.2621253430843353, + 0.06244618445634842, + 0.04724791646003723, + 0.042768970131874084, + 0.020516136661171913, + 0.06837689131498337, + 0.05922439694404602, + 0.07618909329175949 + ], + "credit_map": [ + 0.09724640101194382, + 0.09669563919305801, + 0.01584913581609726, + 0.01763620227575302, + 0.017326919361948967, + 0.02114657126367092, + 0.02456083707511425, + 0.014903067611157894, + 0.007622139528393745, + 0.007670494727790356, + 0.007753376848995686, + 0.007338921073824167, + 0.006333858706057072, + 0.006149541586637497, + 0.00751722464337945, + 0.005354810040444136, + 0.04891013726592064, + 0.06276021897792816, + 0.018603190779685974, + 0.039221614599227905, + 0.019669972360134125, + 0.029363485053181648, + 0.020581364631652832, + 0.023015346378087997, + 0.006232502404600382, + 0.005114891566336155, + 0.013906403444707394, + 0.00850338488817215, + 0.005607410334050655, + 0.007317750249058008, + 0.004290723241865635, + 0.01147311832755804, + 0.006200183182954788, + 0.0051797302439808846, + 0.006789561361074448, + 0.004950317088514566, + 0.007529098074883223, + 0.005331822670996189, + 0.005340734962373972, + 0.005926467943936586, + 0.0037006998900324106, + 0.0033355329651385546, + 0.009922854602336884, + 0.01002479624003172, + 0.004851676523685455, + 0.0036521286237984896, + 0.003417176427319646, + 0.003864104161038995, + 0.002298609120771289, + 0.0025056723970919847, + 0.0024424793664366007, + 0.002227857941761613, + 0.003943504765629768, + 0.0031280298717319965, + 0.002239211928099394, + 0.0017307702219113708, + 0.012247852049767971, + 0.010663838125765324, + 0.006456076167523861, + 0.008170217275619507, + 0.007261228747665882, + 0.015735086053609848, + 0.004105070140212774, + 0.003737521357834339, + 0.005715957377105951, + 0.005924704484641552, + 0.011269956827163696, + 0.012740770354866982, + 0.004723099526017904, + 0.007171608507633209, + 0.006043547298759222, + 0.005634755827486515, + 0.011247767135500908, + 0.01015893928706646, + 0.008851499296724796, + 0.007350767496973276, + 0.008047613315284252, + 0.01765529252588749, + 0.0065361130982637405, + 0.006341102533042431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.011088416911661625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890500396490097 + ], + "opposite_logprob_tokens": [ + -2.464050054550171 + ], + "logprob_reward": -0.0890500396490097, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890500396490097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20453956723213196, + 0.44433367252349854, + 0.08597113192081451, + 0.03609700873494148, + 0.07079685479402542, + 0.015564647503197193, + 0.008587528020143509, + 0.03909948468208313, + 0.05930476635694504, + 0.035705339163541794 + ], + "credit_map": [ + 0.06427382677793503, + 0.06309452652931213, + 0.01619032770395279, + 0.011028718203306198, + 0.01223030500113964, + 0.014911985024809837, + 0.013307253830134869, + 0.00950262788683176, + 0.08502103388309479, + 0.097872793674469, + 0.03400031104683876, + 0.03366832807660103, + 0.051023270934820175, + 0.06763094663619995, + 0.034685224294662476, + 0.04043173789978027, + 0.015402435325086117, + 0.01874174363911152, + 0.007556355092674494, + 0.012032270431518555, + 0.007671390660107136, + 0.01198512502014637, + 0.006024866830557585, + 0.0065569463185966015, + 0.003632328473031521, + 0.002776864217594266, + 0.007332677952945232, + 0.00518331304192543, + 0.002811835613101721, + 0.004564194940030575, + 0.003992636222392321, + 0.005803155247122049, + 0.008030815050005913, + 0.007067655213177204, + 0.007731372956186533, + 0.013608415611088276, + 0.009246462024748325, + 0.008972854353487492, + 0.007220544386655092, + 0.008918729610741138, + 0.0017267477232962847, + 0.0014572379877790809, + 0.0020102786365896463, + 0.0028260541148483753, + 0.0023006293922662735, + 0.002631325740367174, + 0.0014831627486273646, + 0.0011292117414996028, + 0.0009262384846806526, + 0.0007459092885255814, + 0.001466349232941866, + 0.0007754357648082078, + 0.0014918933156877756, + 0.0013152353931218386, + 0.0008438309887424111, + 0.0010226357262581587, + 0.006213672459125519, + 0.006113862618803978, + 0.004394297953695059, + 0.0059854197315871716, + 0.0034201168455183506, + 0.008049721829593182, + 0.002178031951189041, + 0.002744359662756324, + 0.006524760276079178, + 0.006781064439564943, + 0.0076186396181583405, + 0.007671808358281851, + 0.006179105024784803, + 0.011438913643360138, + 0.005503893829882145, + 0.007586583495140076, + 0.004719076212495565, + 0.005945182405412197, + 0.003789603943005204, + 0.0032119217794388533, + 0.004653490148484707, + 0.00492793507874012, + 0.00431095203384757, + 0.0041471743024885654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0890500396490097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623862475156784 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623862475156784, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623862475156784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3647233545780182, + 0.08580552041530609, + 0.10184472799301147, + 0.11945236474275589, + 0.058631107211112976, + 0.13928242027759552, + 0.016666408628225327, + 0.02263057231903076, + 0.036460962146520615, + 0.05450253188610077 + ], + "credit_map": [ + 0.11563324183225632, + 0.11702053248882294, + 0.020733213052153587, + 0.017773227766156197, + 0.023443084210157394, + 0.03377562761306763, + 0.019294828176498413, + 0.017049597576260567, + 0.014034776017069817, + 0.016587011516094208, + 0.005487595219165087, + 0.006937934085726738, + 0.010263483971357346, + 0.01663389801979065, + 0.007090083789080381, + 0.008770745247602463, + 0.010569499805569649, + 0.012077533639967442, + 0.015382092446088791, + 0.013769232667982578, + 0.009764136746525764, + 0.015557835809886456, + 0.01040186733007431, + 0.014322532340884209, + 0.010624571703374386, + 0.010475367307662964, + 0.029470298439264297, + 0.02427934668958187, + 0.007620140444487333, + 0.013452671468257904, + 0.008604171685874462, + 0.014925798401236534, + 0.006846162024885416, + 0.00619643134996295, + 0.008132831193506718, + 0.008465997874736786, + 0.006312641780823469, + 0.0075901285745203495, + 0.007108713965862989, + 0.007978195324540138, + 0.014099569991230965, + 0.009692872874438763, + 0.02952536568045616, + 0.029992204159498215, + 0.010516158305108547, + 0.015959767624735832, + 0.010550567880272865, + 0.018945924937725067, + 0.0019160867668688297, + 0.0017821204382926226, + 0.0016053958097472787, + 0.0015035077231004834, + 0.003504301654174924, + 0.0023433584719896317, + 0.0024488407652825117, + 0.001562796300277114, + 0.0027166767977178097, + 0.003715014550834894, + 0.003615529043599963, + 0.0033331529702991247, + 0.001943030278198421, + 0.003922960255295038, + 0.001838411670178175, + 0.0015457958215847611, + 0.004138481803238392, + 0.003271309658885002, + 0.00534244766458869, + 0.005265576299279928, + 0.003158999839797616, + 0.008063833229243755, + 0.00314071262255311, + 0.004079599864780903, + 0.006240001413971186, + 0.006297803483903408, + 0.007907625287771225, + 0.004886295180767775, + 0.0050664665177464485, + 0.007344099693000317, + 0.006670002825558186, + 0.0100902384147048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.007623862475156784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704764306545258 + ], + "opposite_logprob_tokens": [ + -2.127047538757324 + ], + "logprob_reward": -0.12704764306545258, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704764306545258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09967951476573944, + 0.11055164784193039, + 0.13367745280265808, + 0.06061675027012825, + 0.06679892539978027, + 0.0477549210190773, + 0.027829397469758987, + 0.04180638864636421, + 0.3091731071472168, + 0.10211190581321716 + ], + "credit_map": [ + 0.020282650366425514, + 0.018820492550730705, + 0.010737298056483269, + 0.009823250584304333, + 0.006886858958750963, + 0.0091011431068182, + 0.015591884031891823, + 0.008435934782028198, + 0.01624494604766369, + 0.019294502213597298, + 0.007527429144829512, + 0.007578442804515362, + 0.012818248011171818, + 0.020612848922610283, + 0.009577923454344273, + 0.016897311434149742, + 0.015626905485987663, + 0.021022068336606026, + 0.015244624577462673, + 0.025131752714514732, + 0.01340286061167717, + 0.02038809470832348, + 0.013716802932322025, + 0.009144344367086887, + 0.007471132557839155, + 0.00617872504517436, + 0.0117938918992877, + 0.007177635096013546, + 0.005298067815601826, + 0.007537436671555042, + 0.004429822321981192, + 0.010730039328336716, + 0.009447302669286728, + 0.010443526320159435, + 0.008144266903400421, + 0.008620449341833591, + 0.006905056536197662, + 0.008196591399610043, + 0.007510003168135881, + 0.0075317309238016605, + 0.005399573128670454, + 0.0050480058416724205, + 0.005753811914473772, + 0.007073689252138138, + 0.006752368062734604, + 0.009074351750314236, + 0.0042368704453110695, + 0.00441624503582716, + 0.003973197191953659, + 0.0034964687656611204, + 0.0036592313554137945, + 0.002706896048039198, + 0.0025891654659062624, + 0.006060339044779539, + 0.0016797579592093825, + 0.003664341988041997, + 0.005074396263808012, + 0.004371308721601963, + 0.005043941084295511, + 0.004445550497621298, + 0.005639778450131416, + 0.007549318019300699, + 0.005824272520840168, + 0.0038578249514102936, + 0.031244346871972084, + 0.034215208142995834, + 0.05037478730082512, + 0.05874108150601387, + 0.02907714433968067, + 0.045964475721120834, + 0.03068576753139496, + 0.02887030504643917, + 0.01159945223480463, + 0.012530847452580929, + 0.017844492569565773, + 0.009337027557194233, + 0.01005290076136589, + 0.013779519125819206, + 0.012682013213634491, + 0.014285656623542309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.12704764306545258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a low, ground-level angle, the camera remains stationary, focused on a delicate green sprout, newly emerging from the soil. The sprout, a vibrant, emerald green, its tip slightly bent from its vigorous growth. Tiny, translucent water droplets cling to its translucent stem, shimmering like miniature diamonds in the soft, diffuse morning light. Surrounding it is moist, dark brown, almost black soil, its grains coarse and distinct, its texture glossy from its moisture. With slow yet unmistakable force, the shoot pushes upward, causing tiny cracks to form around its roots, and tiny clumps of soil tumbling from its sides. The entire frame is enveloped by an extremely shallow depth of field, the background completely blurred, leaving the focus squarely on this tiny yet tenacious miracle of life. The delicate light and shadows, and the intense color contrasts create a grand atmosphere of tranquility, hope, and the resilience of life.", + "seed": 11174, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_4247_seed_11174_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.079484320897609, + "vlm_logprob": -0.079484320897609, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.079484320897609, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002504785545170307 + ], + "opposite_logprob_tokens": [ + -7.502504825592041 + ], + "logprob_reward": -0.002504785545170307, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.002504785545170307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13181914389133453, + 0.14850455522537231, + 0.11387485265731812, + 0.043675221502780914, + 0.05663860961794853, + 0.024718938395380974, + 0.3352862596511841, + 0.02857874147593975, + 0.026577504351735115, + 0.09032624959945679 + ], + "credit_map": [ + 0.010507442988455296, + 0.03684140369296074, + 0.01024194061756134, + 0.012510271742939949, + 0.009634085930883884, + 0.03298795968294144, + 0.011532488279044628, + 0.007563564460724592, + 0.009817645885050297, + 0.04560449346899986, + 0.008122601546347141, + 0.004806459881365299, + 0.011581710539758205, + 0.05982137471437454, + 0.0043845586478710175, + 0.00436570355668664, + 0.011714767664670944, + 0.02285773493349552, + 0.011177065782248974, + 0.008137987926602364, + 0.009540751576423645, + 0.015314762480556965, + 0.008870803751051426, + 0.026260986924171448, + 0.007457383442670107, + 0.006517382338643074, + 0.0038788034580647945, + 0.004650866612792015, + 0.006350837182253599, + 0.006803218740969896, + 0.0032431313302367926, + 0.004773599561303854, + 0.011863159015774727, + 0.00968929287046194, + 0.005567729473114014, + 0.0037686110008507967, + 0.005904506891965866, + 0.008326519280672073, + 0.005320142023265362, + 0.00619864696636796, + 0.003942517563700676, + 0.002501776674762368, + 0.0035287393257021904, + 0.0021958560682833195, + 0.003216566052287817, + 0.0033482173457741737, + 0.0031823900062590837, + 0.0028028744272887707, + 0.07800830155611038, + 0.011614718474447727, + 0.026160119101405144, + 0.009690669365227222, + 0.035513993352651596, + 0.10886314511299133, + 0.025968974456191063, + 0.03946632146835327, + 0.004858117084950209, + 0.003038763767108321, + 0.003236035816371441, + 0.0019350512884557247, + 0.004557742737233639, + 0.003821559716016054, + 0.0028687843587249517, + 0.0042626867070794106, + 0.0031082446221262217, + 0.003060337156057358, + 0.003123318310827017, + 0.0020699426531791687, + 0.005163468420505524, + 0.003802048973739147, + 0.0037175435572862625, + 0.0025326015893369913, + 0.021651269868016243, + 0.009408865123987198, + 0.01079386193305254, + 0.005831893067806959, + 0.018798980861902237, + 0.011089262552559376, + 0.006100998725742102, + 0.006651117466390133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.002504785545170307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792622178792953 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792622178792953, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792622178792953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09291468560695648, + 0.060249313712120056, + 0.2232954353094101, + 0.04353547468781471, + 0.08419613540172577, + 0.05361562594771385, + 0.3047509789466858, + 0.08929101377725601, + 0.015441970899701118, + 0.032709360122680664 + ], + "credit_map": [ + 0.008505339734256268, + 0.012889627367258072, + 0.007687498349696398, + 0.010306034237146378, + 0.010050226002931595, + 0.021344197914004326, + 0.011784660629928112, + 0.010347099974751472, + 0.004257922992110252, + 0.012154830619692802, + 0.004762916825711727, + 0.004649091046303511, + 0.00664459727704525, + 0.017581675201654434, + 0.0062185246497392654, + 0.003979756496846676, + 0.026929404586553574, + 0.02865639515221119, + 0.01950041390955448, + 0.02855982817709446, + 0.04549511522054672, + 0.031983938068151474, + 0.01837235502898693, + 0.023798001930117607, + 0.006771470420062542, + 0.008593431673943996, + 0.004447626881301403, + 0.003863375401124358, + 0.005688423290848732, + 0.0060299839824438095, + 0.0039200205355882645, + 0.0042211441323161125, + 0.01752859726548195, + 0.01622103713452816, + 0.008750337176024914, + 0.005489820148795843, + 0.009796795435249805, + 0.01587047427892685, + 0.00527794286608696, + 0.005261133890599012, + 0.00848293025046587, + 0.007221704814583063, + 0.0049047926440835, + 0.004674093797802925, + 0.012889504432678223, + 0.00682931300252676, + 0.004496635403484106, + 0.004116649739444256, + 0.08296287059783936, + 0.016517380252480507, + 0.025598449632525444, + 0.011120888404548168, + 0.02784336544573307, + 0.11512664705514908, + 0.01008041761815548, + 0.015500991605222225, + 0.01877768523991108, + 0.010085185058414936, + 0.011067795567214489, + 0.007113478146493435, + 0.014046600088477135, + 0.012365429662168026, + 0.006759082432836294, + 0.009075758047401905, + 0.0029107334557920694, + 0.001631702994927764, + 0.002135926391929388, + 0.0015822877176105976, + 0.0023716366849839687, + 0.0017457309877499938, + 0.0018199944170191884, + 0.0012439570855349302, + 0.004773677792400122, + 0.0037772017531096935, + 0.003687454154714942, + 0.0022926300298422575, + 0.006847033742815256, + 0.005124799441546202, + 0.002430076478049159, + 0.003776487661525607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.009792622178792953 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706926465034485 + ], + "opposite_logprob_tokens": [ + -2.1270692348480225 + ], + "logprob_reward": -0.12706926465034485, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706926465034485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10352347791194916, + 0.07380524277687073, + 0.1418381780385971, + 0.05595412477850914, + 0.08019430190324783, + 0.14074695110321045, + 0.19083204865455627, + 0.0895100012421608, + 0.017122425138950348, + 0.10647329688072205 + ], + "credit_map": [ + 0.011322742328047752, + 0.011577080935239792, + 0.010691164061427116, + 0.018068630248308182, + 0.012325367890298367, + 0.013159860856831074, + 0.013136064633727074, + 0.013242565095424652, + 0.0048862784169614315, + 0.022195514291524887, + 0.005151030141860247, + 0.006731352768838406, + 0.005542906001210213, + 0.01978839561343193, + 0.004948215093463659, + 0.00456155464053154, + 0.019011739641427994, + 0.01997397467494011, + 0.014393274672329426, + 0.015582158230245113, + 0.02027899958193302, + 0.01956190913915634, + 0.012678544037044048, + 0.020357586443424225, + 0.006796176545321941, + 0.006022252142429352, + 0.006071440409868956, + 0.004671649541705847, + 0.008947362191975117, + 0.008530636318027973, + 0.004410859197378159, + 0.010503750294446945, + 0.012652360834181309, + 0.01136078778654337, + 0.006606565322726965, + 0.007103056646883488, + 0.01545700989663601, + 0.01400383934378624, + 0.007530096918344498, + 0.005480589810758829, + 0.02566622570157051, + 0.019228532910346985, + 0.01622060127556324, + 0.013681357726454735, + 0.02926165796816349, + 0.016200758516788483, + 0.01071976963430643, + 0.009768052026629448, + 0.025706978514790535, + 0.012252641841769218, + 0.014391607604920864, + 0.013194841332733631, + 0.034131068736314774, + 0.04251888766884804, + 0.025934891775250435, + 0.022701136767864227, + 0.01366072054952383, + 0.008176736533641815, + 0.009115404449403286, + 0.005574601702392101, + 0.013919766061007977, + 0.011240947060286999, + 0.01265166699886322, + 0.015170161612331867, + 0.0020521634723991156, + 0.0013314400566741824, + 0.0017668032087385654, + 0.0014522391138598323, + 0.0034767368342727423, + 0.0033575266133993864, + 0.001535983057692647, + 0.002149533247575164, + 0.026022635400295258, + 0.012471495196223259, + 0.012378276325762272, + 0.00720395240932703, + 0.01756187714636326, + 0.014570754021406174, + 0.008088121190667152, + 0.008176186122000217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.12706926465034485 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945970769971609 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005945970769971609, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945970769971609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10644974559545517, + 0.18199172616004944, + 0.1864250898361206, + 0.04450635239481926, + 0.06705041229724884, + 0.09917446225881577, + 0.15542709827423096, + 0.06382700055837631, + 0.03322804719209671, + 0.06192007660865784 + ], + "credit_map": [ + 0.010073114186525345, + 0.02166571281850338, + 0.009355234913527966, + 0.009718408808112144, + 0.012297159060835838, + 0.020935777574777603, + 0.010323446244001389, + 0.012080900371074677, + 0.013347310945391655, + 0.05321485921740532, + 0.010820453055202961, + 0.007101738825440407, + 0.013653771951794624, + 0.07104064524173737, + 0.005346574820578098, + 0.007466383744031191, + 0.02244103141129017, + 0.05127078294754028, + 0.014855950139462948, + 0.012135768309235573, + 0.016674824059009552, + 0.03409311920404434, + 0.011843477375805378, + 0.02311014197766781, + 0.006500722374767065, + 0.006427694112062454, + 0.006027277559041977, + 0.0036842089612036943, + 0.005267047788947821, + 0.005345799960196018, + 0.005092221312224865, + 0.006161381956189871, + 0.01113505195826292, + 0.011657795868813992, + 0.007779906969517469, + 0.0052381171844899654, + 0.00586449820548296, + 0.01434099581092596, + 0.005614394787698984, + 0.005419659428298473, + 0.021976826712489128, + 0.012582861818373203, + 0.013165360316634178, + 0.0061299302615225315, + 0.010782814584672451, + 0.012606040574610233, + 0.010621510446071625, + 0.011309116147458553, + 0.035873256623744965, + 0.007823241874575615, + 0.013805191032588482, + 0.005746209062635899, + 0.013313304632902145, + 0.04048381373286247, + 0.014176580123603344, + 0.02420552261173725, + 0.014000744558870792, + 0.007406969554722309, + 0.009243050590157509, + 0.004041607026010752, + 0.007644059136509895, + 0.011325002647936344, + 0.00355695397593081, + 0.006608615629374981, + 0.005632840096950531, + 0.0034243299160152674, + 0.004165761638432741, + 0.0026723151095211506, + 0.005461137276142836, + 0.005657826084643602, + 0.0030550886876881123, + 0.0031587514095008373, + 0.010326201096177101, + 0.004310619551688433, + 0.007261434569954872, + 0.00511466758325696, + 0.008564518764615059, + 0.013348618522286415, + 0.005104967392981052, + 0.007889054715633392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.005945970769971609 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210896134376526 + ], + "opposite_logprob_tokens": [ + -1.5021089315414429 + ], + "logprob_reward": -0.25210896134376526, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210896134376526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07921025156974792, + 0.028258804231882095, + 0.15779440104961395, + 0.04540405422449112, + 0.08371951431035995, + 0.10232292860746384, + 0.4053606390953064, + 0.017293378710746765, + 0.009728915989398956, + 0.07090719789266586 + ], + "credit_map": [ + 0.00500479806214571, + 0.019931552931666374, + 0.00493759335950017, + 0.0042337276972830296, + 0.007337462157011032, + 0.02557692863047123, + 0.00505479471758008, + 0.007133382838219404, + 0.0018474526004865766, + 0.004966110922396183, + 0.0030134841799736023, + 0.0029001787770539522, + 0.0037323187571018934, + 0.006608255673199892, + 0.0028064351063221693, + 0.0023845615796744823, + 0.024366240948438644, + 0.027811937034130096, + 0.015435649082064629, + 0.017587270587682724, + 0.02233126014471054, + 0.028757914900779724, + 0.008967841975390911, + 0.012536249123513699, + 0.007156421430408955, + 0.009289707988500595, + 0.00473338458687067, + 0.003987226169556379, + 0.005527487490326166, + 0.005619921255856752, + 0.004597862251102924, + 0.004492032341659069, + 0.016495008021593094, + 0.017438413575291634, + 0.008540220558643341, + 0.005711516831070185, + 0.011325037106871605, + 0.013700461015105247, + 0.004555137362331152, + 0.0059537035413086414, + 0.019095979630947113, + 0.014748314395546913, + 0.012243585661053658, + 0.008536730892956257, + 0.01715993881225586, + 0.011582322418689728, + 0.009914475493133068, + 0.009041568264365196, + 0.08925771713256836, + 0.011631844565272331, + 0.02848835475742817, + 0.016355525702238083, + 0.063330739736557, + 0.16176344454288483, + 0.01637495681643486, + 0.0181579627096653, + 0.0019892207346856594, + 0.0016606842400506139, + 0.0020973808132112026, + 0.0013009699760004878, + 0.003109375946223736, + 0.0025272155180573463, + 0.002200047019869089, + 0.0024084807373583317, + 0.001433099969290197, + 0.0008581264992244542, + 0.0010077876504510641, + 0.0008229871746152639, + 0.0017796452157199383, + 0.0020301847252994776, + 0.0009113640990108252, + 0.0008857179200276732, + 0.018346697092056274, + 0.00789329782128334, + 0.008969197981059551, + 0.004413370043039322, + 0.012466092593967915, + 0.01003347896039486, + 0.004136315546929836, + 0.004648733418434858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.25210896134376526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a low, ground-level angle, the camera remains stationary, focused on a delicate green sprout, newly emerging from the soil. The sprout, a vibrant, emerald green, its tip slightly bent from its vigorous growth. Tiny, translucent water droplets cling to its translucent stem, shimmering like miniature diamonds in the soft, diffuse morning light. Surrounding it is moist, dark brown, almost black soil, its grains coarse and distinct, its texture glossy from its moisture. With slow yet unmistakable force, the shoot pushes upward, causing tiny cracks to form around its roots, and tiny clumps of soil tumbling from its sides. The entire frame is enveloped by an extremely shallow depth of field, the background completely blurred, leaving the focus squarely on this tiny yet tenacious miracle of life. The delicate light and shadows, and the intense color contrasts create a grand atmosphere of tranquility, hope, and the resilience of life.", + "seed": 11175, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_7/index_4247_seed_11175_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029797309543937445, + "vlm_logprob": -0.029797309543937445, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.029797309543937445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0030578067526221275 + ], + "opposite_logprob_tokens": [ + -8.378057479858398 + ], + "logprob_reward": -0.0030578067526221275, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.0030578067526221275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14630122482776642, + 0.13606765866279602, + 0.09996088594198227, + 0.12246256321668625, + 0.05463395640254021, + 0.0837722048163414, + 0.07314492762088776, + 0.13481995463371277, + 0.11794830858707428, + 0.030888337641954422 + ], + "credit_map": [ + 0.03181247413158417, + 0.01852625608444214, + 0.017535105347633362, + 0.014438342303037643, + 0.01120044942945242, + 0.014133850112557411, + 0.012007875367999077, + 0.02664686180651188, + 0.030299589037895203, + 0.013982328586280346, + 0.012367348186671734, + 0.009595753625035286, + 0.01611214317381382, + 0.010965831577777863, + 0.013691919855773449, + 0.029052739962935448, + 0.012134025804698467, + 0.008539704605937004, + 0.017191890627145767, + 0.020342132076621056, + 0.0109100965783, + 0.009777924045920372, + 0.011606773361563683, + 0.009458332322537899, + 0.016729645431041718, + 0.008925222791731358, + 0.016129590570926666, + 0.01385506335645914, + 0.02405632473528385, + 0.014863836579024792, + 0.01286979392170906, + 0.015033097006380558, + 0.006704449187964201, + 0.006733150687068701, + 0.00787377543747425, + 0.007475333753973246, + 0.008180670440196991, + 0.007966171018779278, + 0.004832914099097252, + 0.004867488984018564, + 0.00936678983271122, + 0.013421894982457161, + 0.012205212377011776, + 0.009160309098660946, + 0.012077665887773037, + 0.009878402575850487, + 0.011220802552998066, + 0.006441120989620686, + 0.00882086344063282, + 0.009021676145493984, + 0.009750057943165302, + 0.007819099351763725, + 0.012321222573518753, + 0.01126893050968647, + 0.008615680038928986, + 0.005527404602617025, + 0.014714563265442848, + 0.014859921298921108, + 0.017491089180111885, + 0.022638026624917984, + 0.018583543598651886, + 0.013933778740465641, + 0.02271825447678566, + 0.009880776517093182, + 0.012471302412450314, + 0.013897756114602089, + 0.017741195857524872, + 0.019419671967625618, + 0.012467030435800552, + 0.013625366613268852, + 0.016984188929200172, + 0.01134179625660181, + 0.0032476880587637424, + 0.0044050198048353195, + 0.0051941778510808945, + 0.0042658476158976555, + 0.0038737724535167217, + 0.003276445670053363, + 0.004338917322456837, + 0.002286471426486969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0030578067526221275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745306774973869 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745306774973869, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745306774973869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27506914734840393, + 0.10094735771417618, + 0.21355785429477692, + 0.07541535049676895, + 0.08178611099720001, + 0.05522463098168373, + 0.030770674347877502, + 0.0391944982111454, + 0.08445856720209122, + 0.04357580095529556 + ], + "credit_map": [ + 0.07195042073726654, + 0.024605339393019676, + 0.020573558285832405, + 0.02055857703089714, + 0.01677565462887287, + 0.019364329054951668, + 0.019125182181596756, + 0.08211609721183777, + 0.021389158442616463, + 0.009701107628643513, + 0.010466628707945347, + 0.008558929897844791, + 0.013256819918751717, + 0.008831631392240524, + 0.008989416062831879, + 0.01975366100668907, + 0.02426045946776867, + 0.017367564141750336, + 0.03125408664345741, + 0.04809222370386124, + 0.03453066945075989, + 0.015055288560688496, + 0.026854904368519783, + 0.016142666339874268, + 0.008710616268217564, + 0.00634470721706748, + 0.009112854488193989, + 0.008315453305840492, + 0.01689962111413479, + 0.008552388288080692, + 0.011137931607663631, + 0.006341779604554176, + 0.009982891380786896, + 0.010458532720804214, + 0.008242763578891754, + 0.010921561159193516, + 0.011539069935679436, + 0.009981496259570122, + 0.0115067632868886, + 0.009153029881417751, + 0.0065644774585962296, + 0.008115692064166069, + 0.006331497337669134, + 0.0096148531883955, + 0.005770508199930191, + 0.006208844482898712, + 0.0081724738702178, + 0.004446282517164946, + 0.003506503766402602, + 0.0027811359614133835, + 0.003497485537081957, + 0.004668188281357288, + 0.0038108204025775194, + 0.0027218249160796404, + 0.006032360717654228, + 0.003752356395125389, + 0.004261523485183716, + 0.005613832268863916, + 0.005282105877995491, + 0.004738123621791601, + 0.004396624397486448, + 0.003900018520653248, + 0.008406861685216427, + 0.0025954057928174734, + 0.008034000173211098, + 0.01445343904197216, + 0.009727908298373222, + 0.00856789667159319, + 0.01263360120356083, + 0.010250624269247055, + 0.013554010540246964, + 0.007237088400870562, + 0.004702297039330006, + 0.005674955900758505, + 0.00714658759534359, + 0.005855295341461897, + 0.0051681287586688995, + 0.005089387763291597, + 0.006066134199500084, + 0.003873012727126479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.006745306774973869 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054962195456027985 + ], + "opposite_logprob_tokens": [ + -2.929962158203125 + ], + "logprob_reward": -0.054962195456027985, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054962195456027985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18636296689510345, + 0.09665863960981369, + 0.1354912519454956, + 0.13072149455547333, + 0.10533054172992706, + 0.09487433731555939, + 0.04541826993227005, + 0.09119920432567596, + 0.050116680562496185, + 0.06382659822702408 + ], + "credit_map": [ + 0.045811668038368225, + 0.0135823218151927, + 0.01374760176986456, + 0.01193627156317234, + 0.013141734525561333, + 0.01358270924538374, + 0.019566889852285385, + 0.05499378219246864, + 0.011130288243293762, + 0.00786193273961544, + 0.011922556906938553, + 0.010397071950137615, + 0.017155766487121582, + 0.01303150411695242, + 0.013896213844418526, + 0.011263316497206688, + 0.02177729643881321, + 0.010758246295154095, + 0.014879212714731693, + 0.01235960889607668, + 0.02212996408343315, + 0.01314837858080864, + 0.013813905417919159, + 0.026624644175171852, + 0.023360738530755043, + 0.011950775049626827, + 0.012164625339210033, + 0.011384001933038235, + 0.023091409355401993, + 0.021053381264209747, + 0.014816963113844395, + 0.012899624183773994, + 0.011904403567314148, + 0.01186448521912098, + 0.013361243531107903, + 0.020310277119278908, + 0.013109310530126095, + 0.011931171640753746, + 0.01568375714123249, + 0.007165906019508839, + 0.01048798393458128, + 0.011647721752524376, + 0.013034227304160595, + 0.014556534588336945, + 0.009432691149413586, + 0.011676779016852379, + 0.016766423359513283, + 0.007271982729434967, + 0.0051504261791706085, + 0.004368855617940426, + 0.0068839420564472675, + 0.0055512734688818455, + 0.005610802676528692, + 0.004147650208324194, + 0.009243817999958992, + 0.004461509175598621, + 0.009522899985313416, + 0.01123203057795763, + 0.011316372081637383, + 0.009269722737371922, + 0.015700243413448334, + 0.00854149367660284, + 0.018588248640298843, + 0.0070282043889164925, + 0.00467958627268672, + 0.007551334798336029, + 0.007012574467808008, + 0.00750895868986845, + 0.006299761589616537, + 0.005101930815726519, + 0.008365773595869541, + 0.0035967668518424034, + 0.006643873639404774, + 0.009181391447782516, + 0.010137484408915043, + 0.00651063397526741, + 0.010014371015131474, + 0.008130254223942757, + 0.008124139159917831, + 0.005084459204226732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.054962195456027985 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249644163995981 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249644163995981, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249644163995981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20576098561286926, + 0.15584509074687958, + 0.12652122974395752, + 0.11632829904556274, + 0.08281039446592331, + 0.12750868499279022, + 0.038772083818912506, + 0.04400520771741867, + 0.03968839347362518, + 0.06275966763496399 + ], + "credit_map": [ + 0.04275534301996231, + 0.02905440889298916, + 0.017960142344236374, + 0.014440571889281273, + 0.008373848162591457, + 0.022107265889644623, + 0.01364041492342949, + 0.057429004460573196, + 0.020365433767437935, + 0.022587304934859276, + 0.02564234845340252, + 0.02003035694360733, + 0.01736011728644371, + 0.017701473087072372, + 0.018027307465672493, + 0.014130746945738792, + 0.02178717963397503, + 0.011499951593577862, + 0.020563378930091858, + 0.021930472925305367, + 0.010321428999304771, + 0.009950314648449421, + 0.009222668595612049, + 0.02124582789838314, + 0.025283312425017357, + 0.00920596532523632, + 0.014921469613909721, + 0.015011696144938469, + 0.007788928225636482, + 0.014847857877612114, + 0.008108100853860378, + 0.021160973235964775, + 0.00937805324792862, + 0.009646881371736526, + 0.01320544071495533, + 0.010603033006191254, + 0.013045145198702812, + 0.009951186366379261, + 0.011015885509550571, + 0.005964765325188637, + 0.015535898506641388, + 0.022493606433272362, + 0.020051339641213417, + 0.015268590301275253, + 0.013569915667176247, + 0.01758153922855854, + 0.014558912254869938, + 0.008448878303170204, + 0.005128650460392237, + 0.0028934411238878965, + 0.007602761499583721, + 0.0048781209625303745, + 0.0034586514811962843, + 0.00425310805439949, + 0.007343598175793886, + 0.0032137492671608925, + 0.004445888102054596, + 0.005036940798163414, + 0.0066492254845798016, + 0.006103313993662596, + 0.005420594941824675, + 0.005124663934111595, + 0.007371887564659119, + 0.0038526933640241623, + 0.005199550651013851, + 0.004755397327244282, + 0.008033418096601963, + 0.003944741562008858, + 0.0035642855800688267, + 0.0061526368372142315, + 0.004493193235248327, + 0.003545172745361924, + 0.006406543310731649, + 0.006886355113238096, + 0.010686305351555347, + 0.010234265588223934, + 0.005246221087872982, + 0.006805851589888334, + 0.010542982257902622, + 0.0059511419385671616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.005249644163995981 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a green sprout emerging from soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprout vibrant emerald green with a tip slightly bent from growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny translucent water droplets visible on the sprout's stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding soil dark brown, coarse, and glossy from moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the sprout push upward causing tiny cracks to form around its roots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897159457206726 + ], + "opposite_logprob_tokens": [ + -2.5789716243743896 + ], + "logprob_reward": -0.07897159457206726, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897159457206726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09701457619667053, + 0.04364284127950668, + 0.077997125685215, + 0.23106299340724945, + 0.31929248571395874, + 0.09167434275150299, + 0.01799657754600048, + 0.042696960270404816, + 0.03862442448735237, + 0.03999771550297737 + ], + "credit_map": [ + 0.023874543607234955, + 0.011053673923015594, + 0.007186900824308395, + 0.007285581436008215, + 0.007322312332689762, + 0.009847813285887241, + 0.00813372153788805, + 0.02231002412736416, + 0.007268136367201805, + 0.004612340126186609, + 0.006363901309669018, + 0.005365826655179262, + 0.005473946686834097, + 0.004210342653095722, + 0.004293636418879032, + 0.0060547091998159885, + 0.011543480679392815, + 0.006045953836292028, + 0.011962193995714188, + 0.013272997923195362, + 0.007594943046569824, + 0.007272835820913315, + 0.008695879951119423, + 0.011608833447098732, + 0.0659594014286995, + 0.0143645741045475, + 0.014807591214776039, + 0.01637430302798748, + 0.01271920558065176, + 0.022265611216425896, + 0.015788737684488297, + 0.06878356635570526, + 0.03286629542708397, + 0.033964015543460846, + 0.03190302848815918, + 0.05154918134212494, + 0.05034941807389259, + 0.03908107429742813, + 0.05652782320976257, + 0.023051636293530464, + 0.011517654173076153, + 0.01129094697535038, + 0.015379693359136581, + 0.0160202719271183, + 0.008819968439638615, + 0.00955023244023323, + 0.01178466435521841, + 0.007310906890779734, + 0.001994466409087181, + 0.0022868735250085592, + 0.0021323561668395996, + 0.0019970443099737167, + 0.0029628421179950237, + 0.002359243342652917, + 0.0025739700067788363, + 0.0016897825989872217, + 0.004066370893269777, + 0.005757310893386602, + 0.005142245441675186, + 0.005122096743434668, + 0.0071831876412034035, + 0.004835396073758602, + 0.007187370676547289, + 0.0034029826056212187, + 0.00396149093285203, + 0.005095608998090029, + 0.004317599348723888, + 0.0044802348129451275, + 0.004260190296918154, + 0.004427685867995024, + 0.008534452877938747, + 0.003547161817550659, + 0.0035040508955717087, + 0.006287858355790377, + 0.0039414395578205585, + 0.0048660594038665295, + 0.006710130255669355, + 0.00454519409686327, + 0.006657399237155914, + 0.0034855843987315893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.07897159457206726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_8/index_tensor([6794, 6794])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_8/index_tensor([6794, 6794])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6c6f03156b7c1803ce0423bbfda45588fdd3234b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_8/index_tensor([6794, 6794])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static shot, taken from directly above, captures a man in clown makeup kneeling alone in the middle of a rain-soaked city street. The dark asphalt, like a shattered mirror, reflects the flickering neon lights of the surrounding skyscrapers and the dappled shadows cast by the dim streetlights. He wears a once vibrant, now soaked, purple suit, the fabric clinging to his body, and his dyed green curls are plastered to his scalp. His signature white clown greaseproof makeup, washed away by the relentless, icy rain, blends with the black eyeliner and blue teardrop patterns, creating dirty, sorrowful furrows across his pale cheeks. His exaggerated red smile becomes blurred, as if melting. His hands weakly support the ground in front of him, his body hunched slightly in uncontrollable grief, his shoulders shaking in a subtle, uncontrollable, desperate tremor that courses through his entire upper body, silently conveying a sense of utter heartbreak and devastation. The entire picture is shrouded in a gloomy and depressing tone, with only artificial light casting a sickly warm color in the puddles, which together create a heartbreaking feeling of being abandoned by the world.", + "seed": 11831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_8/index_6794_seed_11831_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2903358582407236, + "vlm_logprob": -0.2903358582407236, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.2903358582407236, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.242452621459961 + ], + "opposite_logprob_tokens": [ + -0.36745262145996094 + ], + "logprob_reward": -1.242452621459961, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.242452621459961, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24749109148979187, + 0.1247902512550354, + 0.11569926142692566, + 0.1851276159286499, + 0.14161132276058197, + 0.016442973166704178, + 0.03490953519940376, + 0.032089561223983765, + 0.025789985433220863, + 0.07604832202196121 + ], + "credit_map": [ + 0.027829410508275032, + 0.02550281584262848, + 0.01721436344087124, + 0.03633398562669754, + 0.0318254753947258, + 0.028899142518639565, + 0.036773908883333206, + 0.043112024664878845, + 0.009811114519834518, + 0.024661242961883545, + 0.0076585132628679276, + 0.013564303517341614, + 0.013474133796989918, + 0.009458513930439949, + 0.027947798371315002, + 0.018214644864201546, + 0.00977319199591875, + 0.01978847198188305, + 0.008058742620050907, + 0.02324119210243225, + 0.00715307192876935, + 0.008081740699708462, + 0.021224545314908028, + 0.01837831176817417, + 0.016328876838088036, + 0.028279097750782967, + 0.013044559396803379, + 0.04133130982518196, + 0.012022840790450573, + 0.009993762709200382, + 0.03595348447561264, + 0.02817370928823948, + 0.0125883175060153, + 0.01915423385798931, + 0.012241413816809654, + 0.025347864255309105, + 0.018772147595882416, + 0.012790207751095295, + 0.02013942040503025, + 0.020577741786837578, + 0.0011886657448485494, + 0.0014085134025663137, + 0.0016951824072748423, + 0.002256113337352872, + 0.0017896753270179033, + 0.001918311812914908, + 0.003074606182053685, + 0.0031119082123041153, + 0.0035589244216680527, + 0.004269502591341734, + 0.004061681684106588, + 0.005236114375293255, + 0.0038501599337905645, + 0.004092474468052387, + 0.005103472154587507, + 0.004737211391329765, + 0.0017958218231797218, + 0.0018699723295867443, + 0.0013758111745119095, + 0.0018435576930642128, + 0.0017200515139847994, + 0.0030977765563875437, + 0.010125385597348213, + 0.010261190123856068, + 0.003099162131547928, + 0.004066287539899349, + 0.002721417462453246, + 0.0027575704734772444, + 0.002504919655621052, + 0.00356093468144536, + 0.0037075451109558344, + 0.0033721518702805042, + 0.01170209888368845, + 0.014613857492804527, + 0.006708997767418623, + 0.013035762123763561, + 0.004079640377312899, + 0.00647739740088582, + 0.012649890035390854, + 0.006780693773180246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -1.242452621459961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710632383823395 + ], + "opposite_logprob_tokens": [ + -2.1271064281463623 + ], + "logprob_reward": -0.12710632383823395, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12710632383823395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3330342769622803, + 0.09460248053073883, + 0.13123412430286407, + 0.07225789874792099, + 0.08683259785175323, + 0.09783840924501419, + 0.04865257441997528, + 0.06484302133321762, + 0.029206329956650734, + 0.04149828851222992 + ], + "credit_map": [ + 0.04229635372757912, + 0.035314127802848816, + 0.031196672469377518, + 0.03828918933868408, + 0.03077004663646221, + 0.04702582582831383, + 0.04992009699344635, + 0.05822195112705231, + 0.00819617509841919, + 0.01788380742073059, + 0.006117810029536486, + 0.016479220241308212, + 0.010064741596579552, + 0.00664060702547431, + 0.016590286046266556, + 0.012629833072423935, + 0.013500597327947617, + 0.015685126185417175, + 0.008926383219659328, + 0.02789660170674324, + 0.012139962054789066, + 0.009465635754168034, + 0.02404925785958767, + 0.019570564851164818, + 0.009395289234817028, + 0.009691166691482067, + 0.0059612649492919445, + 0.013316498138010502, + 0.006547144614160061, + 0.005755895283073187, + 0.011925559490919113, + 0.009665082208812237, + 0.008710158057510853, + 0.010031231679022312, + 0.008166776970028877, + 0.011989517137408257, + 0.012302928604185581, + 0.00886441022157669, + 0.01283336617052555, + 0.013934209942817688, + 0.010504286736249924, + 0.009780463762581348, + 0.007069610524922609, + 0.011002479121088982, + 0.010107194073498249, + 0.015794768929481506, + 0.016538411378860474, + 0.017041195183992386, + 0.004752734210342169, + 0.005720716435462236, + 0.005310484208166599, + 0.007350696716457605, + 0.005400334484875202, + 0.005482600536197424, + 0.00743631087243557, + 0.007198694162070751, + 0.003997934050858021, + 0.0033653764985501766, + 0.005482308100908995, + 0.0035412085708230734, + 0.0032399031333625317, + 0.006618448533117771, + 0.019853901118040085, + 0.01874394342303276, + 0.0033766061533242464, + 0.003367889439687133, + 0.0026975141372531652, + 0.0033728366252034903, + 0.0030876907985657454, + 0.003979090601205826, + 0.0045621441677212715, + 0.004762560594826937, + 0.0055013541132211685, + 0.0078812837600708, + 0.004054554738104343, + 0.006887663621455431, + 0.0029210131615400314, + 0.0038151054177433252, + 0.0067990305833518505, + 0.0036382812540978193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.12710632383823395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108488067984581 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.01108488067984581, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108488067984581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1325165182352066, + 0.27226772904396057, + 0.13897845149040222, + 0.17061534523963928, + 0.08856525272130966, + 0.03927016630768776, + 0.04617609828710556, + 0.023892661556601524, + 0.049791183322668076, + 0.03792661055922508 + ], + "credit_map": [ + 0.013719465583562851, + 0.018731100484728813, + 0.015781978145241737, + 0.02223929576575756, + 0.01253396738320589, + 0.015891240909695625, + 0.018455514684319496, + 0.015163952484726906, + 0.02095632441341877, + 0.06412466615438461, + 0.015040479600429535, + 0.03598596528172493, + 0.02174201048910618, + 0.015714513137936592, + 0.062479883432388306, + 0.036223892122507095, + 0.01229812204837799, + 0.022398028522729874, + 0.009432085789740086, + 0.031859446316957474, + 0.007103233132511377, + 0.007847771048545837, + 0.026050476357340813, + 0.021989276632666588, + 0.008391501381993294, + 0.0262883473187685, + 0.00839337706565857, + 0.02990887500345707, + 0.013781312853097916, + 0.013028893619775772, + 0.029701853170990944, + 0.041121166199445724, + 0.007760145701467991, + 0.01653485931456089, + 0.008381952531635761, + 0.013221913948655128, + 0.008207025937736034, + 0.008131934329867363, + 0.011781045235693455, + 0.014546371065080166, + 0.0030812385957688093, + 0.007117068860679865, + 0.004510523751378059, + 0.005148379132151604, + 0.0032482717651873827, + 0.004167942795902491, + 0.006766957230865955, + 0.005229780916124582, + 0.0041684238240122795, + 0.004668406210839748, + 0.004532916471362114, + 0.006610528100281954, + 0.006105251144617796, + 0.006710847374051809, + 0.0066665890626609325, + 0.006713134236633778, + 0.0019902789499610662, + 0.0020721880719065666, + 0.0030292849987745285, + 0.0021213560830801725, + 0.0018203799845650792, + 0.002868151990696788, + 0.005031675565987825, + 0.004959342535585165, + 0.005741338711231947, + 0.005727286450564861, + 0.005126324947923422, + 0.0061755976639688015, + 0.004767773672938347, + 0.00571220088750124, + 0.00777796283364296, + 0.008762693963944912, + 0.005195018369704485, + 0.006176794413477182, + 0.002511954167857766, + 0.006162864156067371, + 0.0034001220483332872, + 0.0038640641141682863, + 0.006853583734482527, + 0.003762207692489028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.01108488067984581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608150638639927 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.01608150638639927, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608150638639927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1533392369747162, + 0.3250153064727783, + 0.0836181789636612, + 0.10311426967382431, + 0.06903285533189774, + 0.04738309979438782, + 0.05305680260062218, + 0.04894213005900383, + 0.05221153423190117, + 0.06428660452365875 + ], + "credit_map": [ + 0.018403876572847366, + 0.015205959789454937, + 0.013692107051610947, + 0.0232755858451128, + 0.01778903789818287, + 0.01773684099316597, + 0.023296954110264778, + 0.023938864469528198, + 0.02569289319217205, + 0.07411841303110123, + 0.01804705522954464, + 0.04242350533604622, + 0.02653159573674202, + 0.018960753455758095, + 0.0749899372458458, + 0.044251129031181335, + 0.007603234611451626, + 0.01598169282078743, + 0.006133102811872959, + 0.012961608357727528, + 0.008733457885682583, + 0.005525760818272829, + 0.013321032747626305, + 0.013358285650610924, + 0.008286267518997192, + 0.02510216273367405, + 0.005745733622461557, + 0.017016422003507614, + 0.006605914793908596, + 0.005867458414286375, + 0.01607518643140793, + 0.018415117636322975, + 0.0058838422410190105, + 0.011603704653680325, + 0.005675262771546841, + 0.011762802489101887, + 0.005677016451954842, + 0.005952680949121714, + 0.011247389949858189, + 0.011230153031647205, + 0.005234262906014919, + 0.005318445153534412, + 0.004121959675103426, + 0.005804751999676228, + 0.005069416482001543, + 0.00664215162396431, + 0.008138529025018215, + 0.007053579203784466, + 0.0061202081851661205, + 0.0068749720230698586, + 0.005782510619610548, + 0.008705902844667435, + 0.0052332025952637196, + 0.0062190573662519455, + 0.007145428564399481, + 0.006975519470870495, + 0.0027340769302099943, + 0.002843782538548112, + 0.0037023238837718964, + 0.0027894244994968176, + 0.00266083050519228, + 0.004388189408928156, + 0.014847947284579277, + 0.01497555710375309, + 0.005447808187454939, + 0.007065121550112963, + 0.005814395844936371, + 0.006267619784921408, + 0.0064405035227537155, + 0.0071546523831784725, + 0.006943588610738516, + 0.007077846676111221, + 0.00856674462556839, + 0.01288894098252058, + 0.005477201193571091, + 0.0123665239661932, + 0.0036466738674789667, + 0.0050360895693302155, + 0.011246547102928162, + 0.005057881120592356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01608150638639927 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054953958839178085 + ], + "opposite_logprob_tokens": [ + -2.9299540519714355 + ], + "logprob_reward": -0.054953958839178085, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054953958839178085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1508920043706894, + 0.0854138508439064, + 0.17183348536491394, + 0.15733061730861664, + 0.24593263864517212, + 0.01953921467065811, + 0.06413613259792328, + 0.03972576558589935, + 0.02814917266368866, + 0.037047117948532104 + ], + "credit_map": [ + 0.012702765874564648, + 0.01590787246823311, + 0.014890032820403576, + 0.027614327147603035, + 0.01669968105852604, + 0.017608262598514557, + 0.0207187719643116, + 0.024750296026468277, + 0.005957578308880329, + 0.016572777181863785, + 0.004613705910742283, + 0.007667556870728731, + 0.010172645561397076, + 0.006206055637449026, + 0.020410081371665, + 0.013813449069857597, + 0.013481349684298038, + 0.026650557294487953, + 0.010682717896997929, + 0.04076242446899414, + 0.009908314794301987, + 0.011599177494645119, + 0.03171912208199501, + 0.027029825374484062, + 0.019153978675603867, + 0.01681886613368988, + 0.011539983563125134, + 0.026885610073804855, + 0.010704653337597847, + 0.009952946566045284, + 0.032145723700523376, + 0.03012884594500065, + 0.02168384939432144, + 0.037570178508758545, + 0.02778913639485836, + 0.047458428889513016, + 0.022820524871349335, + 0.016451586037874222, + 0.036436840891838074, + 0.035722095519304276, + 0.0015044893370941281, + 0.0020834875758737326, + 0.0017577250255271792, + 0.002730641281232238, + 0.0024918767157942057, + 0.0026444224640727043, + 0.0031469385139644146, + 0.00317963445559144, + 0.005867606960237026, + 0.007630450185388327, + 0.00681807566434145, + 0.007304453290998936, + 0.008044353686273098, + 0.008434346877038479, + 0.009946313686668873, + 0.010090529918670654, + 0.004062332212924957, + 0.002720900811254978, + 0.003641610499471426, + 0.0027637898456305265, + 0.0030321397352963686, + 0.0035061128437519073, + 0.009817386046051979, + 0.010181490331888199, + 0.0028128898702561855, + 0.002997891278937459, + 0.0034536521416157484, + 0.004187490791082382, + 0.0024168016389012337, + 0.0030319844372570515, + 0.004451559390872717, + 0.004796900786459446, + 0.005588970612734556, + 0.0063743325881659985, + 0.00316424248740077, + 0.005501871462911367, + 0.0027101386804133654, + 0.0036353259347379208, + 0.006195318419486284, + 0.0038769179955124855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.054953958839178085 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, taken from directly above, captures a man in clown makeup kneeling alone in the middle of a rain-soaked city street. The dark asphalt, like a shattered mirror, reflects the flickering neon lights of the surrounding skyscrapers and the dappled shadows cast by the dim streetlights. He wears a once vibrant, now soaked, purple suit, the fabric clinging to his body, and his dyed green curls are plastered to his scalp. His signature white clown greaseproof makeup, washed away by the relentless, icy rain, blends with the black eyeliner and blue teardrop patterns, creating dirty, sorrowful furrows across his pale cheeks. His exaggerated red smile becomes blurred, as if melting. His hands weakly support the ground in front of him, his body hunched slightly in uncontrollable grief, his shoulders shaking in a subtle, uncontrollable, desperate tremor that courses through his entire upper body, silently conveying a sense of utter heartbreak and devastation. The entire picture is shrouded in a gloomy and depressing tone, with only artificial light casting a sickly warm color in the puddles, which together create a heartbreaking feeling of being abandoned by the world.", + "seed": 11832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_8/index_6794_seed_11832_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11381326876580715, + "vlm_logprob": -0.11381326876580715, + "vlm_yes_no_margin": 2.625000011920929, + "vlm_token_credit": -0.11381326876580715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3581945300102234 + ], + "opposite_logprob_tokens": [ + -1.2331945896148682 + ], + "logprob_reward": -0.3581945300102234, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3581945300102234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23682934045791626, + 0.29127874970436096, + 0.103968046605587, + 0.10973620414733887, + 0.056795381009578705, + 0.0387444943189621, + 0.04012797027826309, + 0.039070188999176025, + 0.055938512086868286, + 0.027511129155755043 + ], + "credit_map": [ + 0.02830241620540619, + 0.017633456736803055, + 0.03180266171693802, + 0.041739322245121, + 0.022364987060427666, + 0.036094456911087036, + 0.03344570845365524, + 0.025446347892284393, + 0.022750500589609146, + 0.04480401799082756, + 0.01834632083773613, + 0.02474210411310196, + 0.06729500740766525, + 0.03370474651455879, + 0.034150898456573486, + 0.04548514634370804, + 0.008562661707401276, + 0.02112964540719986, + 0.009374027140438557, + 0.008930707350373268, + 0.02272937260568142, + 0.01233521569520235, + 0.010396022349596024, + 0.010510395281016827, + 0.010756945237517357, + 0.010094920173287392, + 0.00914060603827238, + 0.015560204163193703, + 0.012551271356642246, + 0.01594551093876362, + 0.015295559540390968, + 0.020391181111335754, + 0.004742960445582867, + 0.004838732071220875, + 0.007371608167886734, + 0.0060799093917012215, + 0.008592421188950539, + 0.008856263011693954, + 0.007754698395729065, + 0.008558783680200577, + 0.0035127163864672184, + 0.0036737476475536823, + 0.003821008838713169, + 0.0069030397571623325, + 0.004886622540652752, + 0.00550201116129756, + 0.005890360102057457, + 0.0045549869537353516, + 0.0032622592989355326, + 0.003644980024546385, + 0.005389414262026548, + 0.005818577483296394, + 0.004717594478279352, + 0.0066986181773245335, + 0.005878741387277842, + 0.004717784933745861, + 0.0037980906199663877, + 0.004611880052834749, + 0.002895587356761098, + 0.004859338514506817, + 0.007614716421812773, + 0.005493362434208393, + 0.00604681670665741, + 0.0037503980565816164, + 0.005238143261522055, + 0.005628665909171104, + 0.006770885083824396, + 0.008469026535749435, + 0.005481181666254997, + 0.008730919100344181, + 0.009538820013403893, + 0.006080870982259512, + 0.0020943903364241123, + 0.0018123005283996463, + 0.0019370127702131867, + 0.00323057291097939, + 0.0025044928770512342, + 0.004312437493354082, + 0.006578429602086544, + 0.005041491240262985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.3581945300102234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890793725848198 + ], + "opposite_logprob_tokens": [ + -2.4640793800354004 + ], + "logprob_reward": -0.0890793725848198, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890793725848198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22114554047584534, + 0.08962392807006836, + 0.21004866063594818, + 0.08211003988981247, + 0.08681041747331619, + 0.11166070401668549, + 0.01782139018177986, + 0.024692382663488388, + 0.021879803389310837, + 0.13420714437961578 + ], + "credit_map": [ + 0.038345083594322205, + 0.02263244241476059, + 0.03316551446914673, + 0.027429061010479927, + 0.023318979889154434, + 0.028818869963288307, + 0.023901157081127167, + 0.023534439504146576, + 0.017991285771131516, + 0.007393940817564726, + 0.013187164440751076, + 0.0076605998910963535, + 0.007834505289793015, + 0.011721140705049038, + 0.01128969993442297, + 0.012545591220259666, + 0.019893428310751915, + 0.035994112491607666, + 0.02211579866707325, + 0.015797661617398262, + 0.039678290486335754, + 0.026046935468912125, + 0.022179676219820976, + 0.02834276668727398, + 0.0089950580149889, + 0.006382221356034279, + 0.00808289460837841, + 0.008799605071544647, + 0.009171618148684502, + 0.01212761178612709, + 0.012107684276998043, + 0.016443351283669472, + 0.019807619974017143, + 0.006515036802738905, + 0.01626940444111824, + 0.010032666847109795, + 0.006337556056678295, + 0.010954399593174458, + 0.008505440317094326, + 0.00838829018175602, + 0.023123636841773987, + 0.0073321424424648285, + 0.020350860431790352, + 0.011116649955511093, + 0.0080467090010643, + 0.014952307567000389, + 0.012454940937459469, + 0.01428346149623394, + 0.002804031129926443, + 0.0010141473030671477, + 0.0017584472661837935, + 0.0020312287379056215, + 0.0013907727552577853, + 0.0028152361046522856, + 0.002247244119644165, + 0.0037602835800498724, + 0.0036239305045455694, + 0.002231818391010165, + 0.004601665306836367, + 0.0026568525936454535, + 0.002537808381021023, + 0.003380859736353159, + 0.0026537247467786074, + 0.003005723236128688, + 0.0015746926655992866, + 0.0023049137089401484, + 0.0017091241898015141, + 0.0035184910520911217, + 0.0033103018067777157, + 0.0032521425746381283, + 0.0035482672974467278, + 0.0026618717238307, + 0.023073168471455574, + 0.01273772306740284, + 0.02332601323723793, + 0.014851926825940609, + 0.01217019371688366, + 0.01681450754404068, + 0.016460837796330452, + 0.014772777445614338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0890793725848198 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864705726504326 + ], + "opposite_logprob_tokens": [ + -3.048647165298462 + ], + "logprob_reward": -0.04864705726504326, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864705726504326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24339018762111664, + 0.17434848845005035, + 0.09646634012460709, + 0.15781298279762268, + 0.05631769075989723, + 0.03914616256952286, + 0.02392975240945816, + 0.06225157529115677, + 0.12383528053760529, + 0.022501539438962936 + ], + "credit_map": [ + 0.03007044456899166, + 0.024496112018823624, + 0.022215835750102997, + 0.0412403829395771, + 0.029751548543572426, + 0.03267861157655716, + 0.03483857959508896, + 0.028098680078983307, + 0.013500564731657505, + 0.029231751337647438, + 0.015418894588947296, + 0.016758499667048454, + 0.03534519672393799, + 0.020493511110544205, + 0.022966820746660233, + 0.020633261650800705, + 0.006630246061831713, + 0.01207615714520216, + 0.006605889648199081, + 0.015612049028277397, + 0.014030265621840954, + 0.01379504892975092, + 0.015666471794247627, + 0.012050220742821693, + 0.010143356397747993, + 0.01541429664939642, + 0.013957012444734573, + 0.023381758481264114, + 0.026676716282963753, + 0.02315431460738182, + 0.024602780118584633, + 0.02048276737332344, + 0.005887193605303764, + 0.004231925122439861, + 0.005087115801870823, + 0.009330828674137592, + 0.007477460894733667, + 0.008504272438585758, + 0.008212028071284294, + 0.0075868722051382065, + 0.004226808901876211, + 0.0037933378480374813, + 0.0026810048148036003, + 0.007452136836946011, + 0.004976905416697264, + 0.006287721451371908, + 0.005031510256230831, + 0.004696741700172424, + 0.003152829362079501, + 0.001520892372354865, + 0.0023511224426329136, + 0.0031574575696140528, + 0.0022069173865020275, + 0.0034694639034569263, + 0.003733145771548152, + 0.004337928257882595, + 0.003863357938826084, + 0.005946961697191, + 0.005176554899662733, + 0.007141591049730778, + 0.011053149588406086, + 0.008661131374537945, + 0.009989133104681969, + 0.010419701226055622, + 0.010824778117239475, + 0.012162298895418644, + 0.013530822470784187, + 0.018114494159817696, + 0.017347469925880432, + 0.019862685352563858, + 0.021019140258431435, + 0.010973604395985603, + 0.003269854700192809, + 0.0015334980562329292, + 0.0038043276872485876, + 0.0030160797759890556, + 0.0018135355785489082, + 0.00299014151096344, + 0.0027288098353892565, + 0.0033452934585511684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04864705726504326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192103132605553 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192103132605553, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192103132605553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26654136180877686, + 0.2449445128440857, + 0.19975031912326813, + 0.07894568890333176, + 0.03824996575713158, + 0.03711334988474846, + 0.03267182782292366, + 0.025508766993880272, + 0.05195523798465729, + 0.024318985641002655 + ], + "credit_map": [ + 0.04094091057777405, + 0.022701313719153404, + 0.04973425716161728, + 0.035139258950948715, + 0.02703770436346531, + 0.03482048213481903, + 0.027795612812042236, + 0.028371799737215042, + 0.015660030767321587, + 0.039874035865068436, + 0.01678963378071785, + 0.020692599937319756, + 0.05302468314766884, + 0.03005564771592617, + 0.03245319426059723, + 0.036394670605659485, + 0.02018889971077442, + 0.03600871190428734, + 0.022435318678617477, + 0.015459757298231125, + 0.038150470703840256, + 0.02388363890349865, + 0.02106953226029873, + 0.02255398966372013, + 0.007532790768891573, + 0.006014267448335886, + 0.005872580222785473, + 0.007868233136832714, + 0.009911349043250084, + 0.011475126259028912, + 0.009207227267324924, + 0.021064113825559616, + 0.004983032122254372, + 0.003148294985294342, + 0.0044182464480400085, + 0.005300167948007584, + 0.006112284492701292, + 0.0050796642899513245, + 0.005438396707177162, + 0.0037698803935199976, + 0.004094810225069523, + 0.003664543852210045, + 0.003228733316063881, + 0.005774252582341433, + 0.0056378901936113834, + 0.005844809114933014, + 0.004141085781157017, + 0.004727223888039589, + 0.005627341568470001, + 0.002593348501250148, + 0.0032441923394799232, + 0.0036522301379591227, + 0.00409440603107214, + 0.0044640167616307735, + 0.0035905777476727962, + 0.005405713804066181, + 0.002977494616061449, + 0.002146762562915683, + 0.003713269019499421, + 0.0037552688736468554, + 0.0030215242877602577, + 0.0034835042897611856, + 0.0035137291997671127, + 0.002897211816161871, + 0.005036809016019106, + 0.0096583291888237, + 0.00433720275759697, + 0.0072920252569019794, + 0.008255353197455406, + 0.007713815663009882, + 0.005930118262767792, + 0.0037315860390663147, + 0.0030625909566879272, + 0.0022887494415044785, + 0.003003222169354558, + 0.0036846292205154896, + 0.0024035978130996227, + 0.003521230537444353, + 0.003246795851737261, + 0.0031081687193363905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018192103132605553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054953280836343765 + ], + "opposite_logprob_tokens": [ + -2.9299533367156982 + ], + "logprob_reward": -0.054953280836343765, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054953280836343765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2440100759267807, + 0.1848340779542923, + 0.19828298687934875, + 0.11274070292711258, + 0.046658121049404144, + 0.03293104097247124, + 0.05460461229085922, + 0.028600357472896576, + 0.05104481428861618, + 0.0462932251393795 + ], + "credit_map": [ + 0.021924225613474846, + 0.02076912671327591, + 0.02532968297600746, + 0.04651770740747452, + 0.028929341584444046, + 0.035627011209726334, + 0.03809795528650284, + 0.026815028861165047, + 0.01539220754057169, + 0.028941236436367035, + 0.013044818304479122, + 0.016139565035700798, + 0.04121725633740425, + 0.0216063242405653, + 0.022760095074772835, + 0.02573256753385067, + 0.016850685700774193, + 0.03896761313080788, + 0.014066599309444427, + 0.0159462783485651, + 0.0442628338932991, + 0.02413313463330269, + 0.022843122482299805, + 0.021212728694081306, + 0.010308105498552322, + 0.009995019063353539, + 0.008903445675969124, + 0.01577993482351303, + 0.012681428343057632, + 0.016569742932915688, + 0.014648695476353168, + 0.02385433204472065, + 0.004987550433725119, + 0.0052495901472866535, + 0.004340175073593855, + 0.006572460290044546, + 0.007213585078716278, + 0.00633325194939971, + 0.006671827286481857, + 0.0052896845154464245, + 0.0034629404544830322, + 0.00325202033855021, + 0.0035202798899263144, + 0.004047674126923084, + 0.003908379934728146, + 0.004558640532195568, + 0.00427144393324852, + 0.00590966222807765, + 0.006007438059896231, + 0.00465060118585825, + 0.004231808707118034, + 0.008612108416855335, + 0.006430381443351507, + 0.007476385682821274, + 0.009203503839671612, + 0.00799238495528698, + 0.00337171065621078, + 0.002854263177141547, + 0.0030172253027558327, + 0.003439529100432992, + 0.003966729622334242, + 0.00414064759388566, + 0.0049912575632333755, + 0.0028189930599182844, + 0.0045098597183823586, + 0.005004053469747305, + 0.0056132664903998375, + 0.008267357014119625, + 0.00596705311909318, + 0.00782026443630457, + 0.00880498718470335, + 0.005057974252849817, + 0.004862544592469931, + 0.0035912746097892523, + 0.0040808008052408695, + 0.00741219287738204, + 0.004513989668339491, + 0.00827015470713377, + 0.006075765937566757, + 0.007486501708626747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.054953280836343765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, taken from directly above, captures a man in clown makeup kneeling alone in the middle of a rain-soaked city street. The dark asphalt, like a shattered mirror, reflects the flickering neon lights of the surrounding skyscrapers and the dappled shadows cast by the dim streetlights. He wears a once vibrant, now soaked, purple suit, the fabric clinging to his body, and his dyed green curls are plastered to his scalp. His signature white clown greaseproof makeup, washed away by the relentless, icy rain, blends with the black eyeliner and blue teardrop patterns, creating dirty, sorrowful furrows across his pale cheeks. His exaggerated red smile becomes blurred, as if melting. His hands weakly support the ground in front of him, his body hunched slightly in uncontrollable grief, his shoulders shaking in a subtle, uncontrollable, desperate tremor that courses through his entire upper body, silently conveying a sense of utter heartbreak and devastation. The entire picture is shrouded in a gloomy and depressing tone, with only artificial light casting a sickly warm color in the puddles, which together create a heartbreaking feeling of being abandoned by the world.", + "seed": 11833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_8/index_6794_seed_11833_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13662521056830884, + "vlm_logprob": -0.13662521056830884, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.13662521056830884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5533740520477295 + ], + "opposite_logprob_tokens": [ + -0.9283740520477295 + ], + "logprob_reward": -0.5533740520477295, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5533740520477295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1959773749113083, + 0.06974295526742935, + 0.173949733376503, + 0.06996598094701767, + 0.10503141582012177, + 0.09367550909519196, + 0.09539087116718292, + 0.04898524284362793, + 0.03640078008174896, + 0.11088025569915771 + ], + "credit_map": [ + 0.037567608058452606, + 0.020199717953801155, + 0.019632624462246895, + 0.030554983764886856, + 0.020515289157629013, + 0.02218974195420742, + 0.02246909961104393, + 0.02284829691052437, + 0.00905310083180666, + 0.010943369939923286, + 0.005614012945443392, + 0.007685932796448469, + 0.012696553952991962, + 0.008234928362071514, + 0.007634021807461977, + 0.007881038822233677, + 0.01590837351977825, + 0.021806219592690468, + 0.022818416357040405, + 0.01556300837546587, + 0.0410090908408165, + 0.01801980473101139, + 0.019443631172180176, + 0.019381193444132805, + 0.008958686143159866, + 0.0074875447899103165, + 0.008545099757611752, + 0.0069151753559708595, + 0.005313836969435215, + 0.008777973242104053, + 0.012155311182141304, + 0.011812351644039154, + 0.009965283796191216, + 0.018514694646000862, + 0.00833211001008749, + 0.012536625377833843, + 0.013211426325142384, + 0.006268747616559267, + 0.021152956411242485, + 0.015049569308757782, + 0.009964719414710999, + 0.008825215511023998, + 0.008980984799563885, + 0.017767583951354027, + 0.00733096431940794, + 0.01053512841463089, + 0.016340650618076324, + 0.013930258341133595, + 0.00671367347240448, + 0.010194853879511356, + 0.0063295201398432255, + 0.00873645767569542, + 0.0163746178150177, + 0.008357644081115723, + 0.020877767354249954, + 0.017806341871619225, + 0.00562138669192791, + 0.004726600367575884, + 0.006994318217039108, + 0.005003765225410461, + 0.009520204737782478, + 0.0052294498309493065, + 0.006285754963755608, + 0.005603761412203312, + 0.004690044093877077, + 0.004649903159588575, + 0.0037018703296780586, + 0.004248429089784622, + 0.0036924099549651146, + 0.0040478091686964035, + 0.0049619791097939014, + 0.0064083305187523365, + 0.011799078434705734, + 0.007437034975737333, + 0.0071919159963727, + 0.01932908408343792, + 0.007480822969228029, + 0.007633094675838947, + 0.026171160861849785, + 0.023838061839342117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.5533740520477295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813207149505615 + ], + "opposite_logprob_tokens": [ + -3.2881321907043457 + ], + "logprob_reward": -0.03813207149505615, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813207149505615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08901853114366531, + 0.11830948293209076, + 0.18534517288208008, + 0.14757154881954193, + 0.03376112878322601, + 0.18105603754520416, + 0.01900782622396946, + 0.03636648505926132, + 0.05975800380110741, + 0.1298058032989502 + ], + "credit_map": [ + 0.014069628901779652, + 0.012335441075265408, + 0.011097107082605362, + 0.012296165339648724, + 0.008103751577436924, + 0.009498405270278454, + 0.010896830819547176, + 0.010721188969910145, + 0.01890053227543831, + 0.006762910634279251, + 0.019268827512860298, + 0.010377283208072186, + 0.006338175386190414, + 0.024459343403577805, + 0.013611317612230778, + 0.018591083586215973, + 0.029758457094430923, + 0.021940268576145172, + 0.03425488620996475, + 0.017175637185573578, + 0.02921580709517002, + 0.016672851517796516, + 0.016365177929401398, + 0.019962074235081673, + 0.009296951815485954, + 0.00421253964304924, + 0.007589108310639858, + 0.0077288588508963585, + 0.007590678986161947, + 0.013349058106541634, + 0.05041211098432541, + 0.047392234206199646, + 0.0034273643977940083, + 0.008517120964825153, + 0.004070278722792864, + 0.0037981162313371897, + 0.0028630737215280533, + 0.0036270362325012684, + 0.004827392753213644, + 0.002630744595080614, + 0.018828539177775383, + 0.016787927597761154, + 0.018166635185480118, + 0.0349397249519825, + 0.015109973028302193, + 0.02168956957757473, + 0.030631782487034798, + 0.024901878088712692, + 0.0016129729337990284, + 0.0022236774675548077, + 0.0010071535361930728, + 0.0017722150078043342, + 0.004985358100384474, + 0.0018369014142081141, + 0.003478472353890538, + 0.00209107156842947, + 0.003989355638623238, + 0.002303745597600937, + 0.006486490834504366, + 0.003207152709364891, + 0.003156521590426564, + 0.004914369434118271, + 0.0065083750523626804, + 0.005800471641123295, + 0.007394617889076471, + 0.005562684033066034, + 0.007799067534506321, + 0.007217929698526859, + 0.006412788759917021, + 0.006044401787221432, + 0.01065385527908802, + 0.008672656491398811, + 0.004545112606137991, + 0.003005646402016282, + 0.0019756071269512177, + 0.006347713991999626, + 0.00555420434102416, + 0.008296580985188484, + 0.04925965517759323, + 0.050821274518966675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.03813207149505615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607036218047142 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.01607036218047142, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607036218047142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1334785670042038, + 0.10822177678346634, + 0.20255692303180695, + 0.05613471940159798, + 0.0826735869050026, + 0.08071289956569672, + 0.06706442683935165, + 0.042787522077560425, + 0.02961447834968567, + 0.19675511121749878 + ], + "credit_map": [ + 0.015652401372790337, + 0.012653760612010956, + 0.011173966340720654, + 0.02294699102640152, + 0.015291880816221237, + 0.014135344885289669, + 0.022733034566044807, + 0.01889118365943432, + 0.01869918592274189, + 0.012383844703435898, + 0.014846174977719784, + 0.020350316539406776, + 0.00833844393491745, + 0.009828959591686726, + 0.01241507288068533, + 0.01135978102684021, + 0.033997081220149994, + 0.023397166281938553, + 0.033644549548625946, + 0.024309270083904266, + 0.025745874270796776, + 0.01880873553454876, + 0.02198823168873787, + 0.02066602185368538, + 0.007002890110015869, + 0.0039109354838728905, + 0.009944496676325798, + 0.0035985615104436874, + 0.0035281889140605927, + 0.003425674745813012, + 0.016555311158299446, + 0.0081686582416296, + 0.005213947966694832, + 0.011398626491427422, + 0.005738003179430962, + 0.01307945977896452, + 0.012040117755532265, + 0.007090268190950155, + 0.01536643411964178, + 0.012746729888021946, + 0.007328222505748272, + 0.008271411992609501, + 0.007793055381625891, + 0.014479921199381351, + 0.006224334705621004, + 0.01015846524387598, + 0.013575245626270771, + 0.012882242910563946, + 0.0043892026878893375, + 0.0055969455279409885, + 0.0035064665134996176, + 0.004968099761754274, + 0.015350474044680595, + 0.006227186881005764, + 0.01542032603174448, + 0.01160572562366724, + 0.004569904413074255, + 0.003430691547691822, + 0.005876974668353796, + 0.0065805488266050816, + 0.0031940671615302563, + 0.005285047926008701, + 0.007352359592914581, + 0.006497925613075495, + 0.004289886448532343, + 0.0024949065409600735, + 0.002637348370626569, + 0.0035323137417435646, + 0.002659517340362072, + 0.004866770934313536, + 0.005010821856558323, + 0.0041229138150811195, + 0.0068670534528791904, + 0.0038277623243629932, + 0.0031983850058168173, + 0.00993974320590496, + 0.009033966809511185, + 0.011551052331924438, + 0.07562099397182465, + 0.07671615481376648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.01607036218047142 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059526927769184 + ], + "opposite_logprob_tokens": [ + -3.8955953121185303 + ], + "logprob_reward": -0.02059526927769184, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059526927769184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12694524228572845, + 0.04735991358757019, + 0.5023224949836731, + 0.05772365257143974, + 0.047603219747543335, + 0.02960517816245556, + 0.04193500429391861, + 0.04195913299918175, + 0.040451791137456894, + 0.06409432739019394 + ], + "credit_map": [ + 0.02090137079358101, + 0.012441462837159634, + 0.014624247327446938, + 0.01730160601437092, + 0.013391437008976936, + 0.01364460214972496, + 0.016702933236956596, + 0.01793758012354374, + 0.006345054134726524, + 0.005319288931787014, + 0.007340527139604092, + 0.006356056313961744, + 0.0031044529750943184, + 0.006631629541516304, + 0.005633742548525333, + 0.0066291578114032745, + 0.041093841195106506, + 0.06925870478153229, + 0.05605234205722809, + 0.038593411445617676, + 0.13160373270511627, + 0.05297870934009552, + 0.05332644656300545, + 0.059415318071842194, + 0.0046452064998447895, + 0.0038635109085589647, + 0.005808464717119932, + 0.0045814900659024715, + 0.0031973733566701412, + 0.00495370477437973, + 0.014012633822858334, + 0.0166612658649683, + 0.0074048759415745735, + 0.003700816538184881, + 0.006045367103070021, + 0.004468830302357674, + 0.006976496893912554, + 0.006230511236935854, + 0.0067065623588860035, + 0.0060697575099766254, + 0.003107292577624321, + 0.0030799531377851963, + 0.0029059317894279957, + 0.004934566095471382, + 0.002543453127145767, + 0.004155623726546764, + 0.004880860447883606, + 0.003997496794909239, + 0.0029704400803893805, + 0.004224439617246389, + 0.002264683833345771, + 0.0029855642933398485, + 0.004551485646516085, + 0.003737199353054166, + 0.011072405613958836, + 0.010128787718713284, + 0.004714929964393377, + 0.003796539269387722, + 0.006959311664104462, + 0.005065962206572294, + 0.004495661705732346, + 0.004234149120748043, + 0.007893627509474754, + 0.004798950627446175, + 0.0045226095244288445, + 0.007629429455846548, + 0.003933400847017765, + 0.0052805994637310505, + 0.003471605246886611, + 0.004220930393785238, + 0.006315090227872133, + 0.0050781238824129105, + 0.0035654075909405947, + 0.0028957659378647804, + 0.0026373229920864105, + 0.0037322009447962046, + 0.0029205074533820152, + 0.004724331200122833, + 0.021504877135157585, + 0.02211390994489193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.02059526927769184 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054954297840595245 + ], + "opposite_logprob_tokens": [ + -2.9299542903900146 + ], + "logprob_reward": -0.054954297840595245, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054954297840595245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12641169130802155, + 0.06460056453943253, + 0.13885587453842163, + 0.07505355030298233, + 0.033598266541957855, + 0.031392861157655716, + 0.057065606117248535, + 0.057619716972112656, + 0.07846278697252274, + 0.3369390070438385 + ], + "credit_map": [ + 0.012540382333099842, + 0.012207639403641224, + 0.01158195361495018, + 0.02119566686451435, + 0.01959480158984661, + 0.009838973172008991, + 0.02091565541923046, + 0.01853662170469761, + 0.008187212981283665, + 0.007589783053845167, + 0.005928097292780876, + 0.00798789318650961, + 0.010189431719481945, + 0.009192263707518578, + 0.00788874737918377, + 0.007637142203748226, + 0.01211005263030529, + 0.01758088730275631, + 0.018797889351844788, + 0.01113944873213768, + 0.03318978473544121, + 0.016287880018353462, + 0.01569407805800438, + 0.01405587699264288, + 0.0064428807236254215, + 0.010322301648557186, + 0.007685248274356127, + 0.009018740616738796, + 0.0067220223136246204, + 0.006676360033452511, + 0.015431543812155724, + 0.012754461728036404, + 0.002850349061191082, + 0.004227702505886555, + 0.0022580481600016356, + 0.003887650091201067, + 0.008049705065786839, + 0.003281778423115611, + 0.004932155832648277, + 0.004110881593078375, + 0.0034905641805380583, + 0.003335588611662388, + 0.003096757922321558, + 0.004366312641650438, + 0.0035823334474116564, + 0.0038120814133435488, + 0.00530480919405818, + 0.0044044144451618195, + 0.003381356131285429, + 0.00473475968465209, + 0.0031312890350818634, + 0.0038822884671390057, + 0.010998361743986607, + 0.0050697363913059235, + 0.01410109642893076, + 0.011766723357141018, + 0.006297776009887457, + 0.004122782964259386, + 0.010433541610836983, + 0.004582064691931009, + 0.008539749309420586, + 0.007387880701571703, + 0.009233986958861351, + 0.0070219384506344795, + 0.008023256435990334, + 0.012888295575976372, + 0.006463069934397936, + 0.011524004861712456, + 0.007360382471233606, + 0.009051041677594185, + 0.011410661973059177, + 0.011742074973881245, + 0.011171262711286545, + 0.008440863341093063, + 0.005697662942111492, + 0.018788864836096764, + 0.01426724623888731, + 0.019847935065627098, + 0.12785272300243378, + 0.13087248802185059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.054954297840595245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, taken from directly above, captures a man in clown makeup kneeling alone in the middle of a rain-soaked city street. The dark asphalt, like a shattered mirror, reflects the flickering neon lights of the surrounding skyscrapers and the dappled shadows cast by the dim streetlights. He wears a once vibrant, now soaked, purple suit, the fabric clinging to his body, and his dyed green curls are plastered to his scalp. His signature white clown greaseproof makeup, washed away by the relentless, icy rain, blends with the black eyeliner and blue teardrop patterns, creating dirty, sorrowful furrows across his pale cheeks. His exaggerated red smile becomes blurred, as if melting. His hands weakly support the ground in front of him, his body hunched slightly in uncontrollable grief, his shoulders shaking in a subtle, uncontrollable, desperate tremor that courses through his entire upper body, silently conveying a sense of utter heartbreak and devastation. The entire picture is shrouded in a gloomy and depressing tone, with only artificial light casting a sickly warm color in the puddles, which together create a heartbreaking feeling of being abandoned by the world.", + "seed": 11834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_8/index_6794_seed_11834_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3585310310125351, + "vlm_logprob": -0.3585310310125351, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.3585310310125351, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6211376190185547 + ], + "opposite_logprob_tokens": [ + -0.24613764882087708 + ], + "logprob_reward": -1.6211376190185547, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6211376190185547, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21201300621032715, + 0.062082450836896896, + 0.050204336643218994, + 0.06092977151274681, + 0.236830934882164, + 0.017795519903302193, + 0.18527409434318542, + 0.04335465654730797, + 0.04874145984649658, + 0.08277375996112823 + ], + "credit_map": [ + 0.023920072242617607, + 0.0229162760078907, + 0.01955995336174965, + 0.021182022988796234, + 0.05670478194952011, + 0.016954079270362854, + 0.026783617213368416, + 0.023992210626602173, + 0.008911948651075363, + 0.006845281925052404, + 0.008771614171564579, + 0.008140980266034603, + 0.0077922954224050045, + 0.005593464709818363, + 0.009871984831988811, + 0.006154884118586779, + 0.00411832332611084, + 0.0075318291783332825, + 0.005545296706259251, + 0.007353455759584904, + 0.010321465320885181, + 0.003446172224357724, + 0.006835537496954203, + 0.0050522577948868275, + 0.004626730922609568, + 0.006732329726219177, + 0.006903547327965498, + 0.007072350941598415, + 0.016257312148809433, + 0.0032687715720385313, + 0.008470169268548489, + 0.007598559372127056, + 0.01685120351612568, + 0.013867898844182491, + 0.01605425775051117, + 0.06641045212745667, + 0.02446052059531212, + 0.010023233480751514, + 0.060496266931295395, + 0.028667088598012924, + 0.002356576267629862, + 0.0014791061403229833, + 0.001340752001851797, + 0.0021660346537828445, + 0.002521777292713523, + 0.00184156047180295, + 0.0028294373769313097, + 0.0032602758146822453, + 0.016953187063336372, + 0.023668460547924042, + 0.01597142405807972, + 0.018763761967420578, + 0.046838466078042984, + 0.018330805003643036, + 0.025245167315006256, + 0.019502827897667885, + 0.006705744192004204, + 0.005322426091879606, + 0.00535010127350688, + 0.006608598865568638, + 0.0066568064503371716, + 0.002738280687481165, + 0.005623694974929094, + 0.004349003545939922, + 0.006419480312615633, + 0.007360472809523344, + 0.007410694845020771, + 0.006951996590942144, + 0.004620886407792568, + 0.004996034782379866, + 0.006479810457676649, + 0.004502082709223032, + 0.007868370041251183, + 0.015137155540287495, + 0.008488058112561703, + 0.01730029657483101, + 0.005179001949727535, + 0.005165502429008484, + 0.014474727213382721, + 0.009160645306110382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -1.6211376190185547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501251667737961 + ], + "opposite_logprob_tokens": [ + -2.9300124645233154 + ], + "logprob_reward": -0.05501251667737961, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501251667737961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1013868972659111, + 0.1254255175590515, + 0.01808825321495533, + 0.026294268667697906, + 0.24898456037044525, + 0.011629312299191952, + 0.04562452808022499, + 0.04859447106719017, + 0.12642693519592285, + 0.24754515290260315 + ], + "credit_map": [ + 0.011845514178276062, + 0.009691059589385986, + 0.02067304030060768, + 0.0115101533010602, + 0.013375988230109215, + 0.00992521457374096, + 0.010989568196237087, + 0.013376375660300255, + 0.020016415044665337, + 0.013651524670422077, + 0.016872655600309372, + 0.014726062305271626, + 0.01411354262381792, + 0.016487782821059227, + 0.015075481496751308, + 0.014482083730399609, + 0.0024024436715990305, + 0.0018898616544902325, + 0.0019082333892583847, + 0.0030744203832000494, + 0.0025617596693336964, + 0.0016987172421067953, + 0.0026931290049105883, + 0.001859691459685564, + 0.003359682857990265, + 0.0034602766390889883, + 0.003228304209187627, + 0.003429400036111474, + 0.004384729545563459, + 0.0024580461904406548, + 0.003076969413086772, + 0.0028968642000108957, + 0.021150507032871246, + 0.013792693614959717, + 0.019831053912639618, + 0.06643369048833847, + 0.02442534640431404, + 0.013082240708172321, + 0.06000448390841484, + 0.03026457689702511, + 0.0009688541176728904, + 0.0014074088539928198, + 0.0011700083268806338, + 0.002736434806138277, + 0.0013416087022051215, + 0.0007243959698826075, + 0.001712588593363762, + 0.0015680150827392936, + 0.0035303025506436825, + 0.0036553393583744764, + 0.0029047082643955946, + 0.005424992647022009, + 0.007016441319137812, + 0.002095867181196809, + 0.011199872009456158, + 0.009797015227377415, + 0.007109648548066616, + 0.006198790855705738, + 0.004988780710846186, + 0.009415101259946823, + 0.005660602357238531, + 0.003776567755267024, + 0.0058599007315933704, + 0.005585088394582272, + 0.013725410215556622, + 0.016248181462287903, + 0.01283252239227295, + 0.021896516904234886, + 0.012799564749002457, + 0.014110414311289787, + 0.01883009262382984, + 0.01598426327109337, + 0.037890683859586716, + 0.017229929566383362, + 0.03278229013085365, + 0.03279149532318115, + 0.015703655779361725, + 0.024523410946130753, + 0.04775425046682358, + 0.03886949643492699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.05501251667737961 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420799270272255 + ], + "opposite_logprob_tokens": [ + -4.26420783996582 + ], + "logprob_reward": -0.01420799270272255, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420799270272255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29609113931655884, + 0.058006882667541504, + 0.029564334079623222, + 0.08770523965358734, + 0.3069648742675781, + 0.02637432888150215, + 0.06667745858430862, + 0.03696785494685173, + 0.031195541843771935, + 0.06045232713222504 + ], + "credit_map": [ + 0.025344450026750565, + 0.034444499760866165, + 0.027101654559373856, + 0.0316476933658123, + 0.07774070650339127, + 0.017885232344269753, + 0.04063849896192551, + 0.04128839075565338, + 0.0055562714114785194, + 0.007574787829071283, + 0.0072387102991342545, + 0.010882220230996609, + 0.0067305839620530605, + 0.0055397991091012955, + 0.008414742536842823, + 0.0060697682201862335, + 0.002864374313503504, + 0.003934694919735193, + 0.0028375263791531324, + 0.005684401839971542, + 0.0044571650214493275, + 0.002084498992189765, + 0.004814548883587122, + 0.0028871227987110615, + 0.007564884144812822, + 0.01123996265232563, + 0.008796635083854198, + 0.012772319838404655, + 0.016534307971596718, + 0.0076463595032691956, + 0.010965545661747456, + 0.012185225263237953, + 0.0255647674202919, + 0.017397038638591766, + 0.02446332387626171, + 0.0796520784497261, + 0.03182360157370567, + 0.01599591039121151, + 0.07439309358596802, + 0.037675049155950546, + 0.0021117348223924637, + 0.0038120271638035774, + 0.0034248242154717445, + 0.003344270633533597, + 0.00579493772238493, + 0.0016875595320016146, + 0.0033955073449760675, + 0.0028034683782607317, + 0.0050164684653282166, + 0.006035235710442066, + 0.005452049896121025, + 0.01089979987591505, + 0.013197479769587517, + 0.0064552053809165955, + 0.011000063270330429, + 0.008621154353022575, + 0.004458501003682613, + 0.00555415777489543, + 0.0036618260201066732, + 0.006450730841606855, + 0.004731575958430767, + 0.003005644306540489, + 0.0047718374989926815, + 0.004333578981459141, + 0.004105627071112394, + 0.004375134129077196, + 0.005633566528558731, + 0.005026557482779026, + 0.002454502508044243, + 0.002535817911848426, + 0.0044136024080216885, + 0.002650734968483448, + 0.007921415381133556, + 0.006440484430640936, + 0.008664419874548912, + 0.009625283069908619, + 0.003165682777762413, + 0.005000840872526169, + 0.011738534085452557, + 0.007895667105913162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.01420799270272255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330857142806053 + ], + "opposite_logprob_tokens": [ + -3.773308515548706 + ], + "logprob_reward": -0.02330857142806053, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330857142806053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33630338311195374, + 0.0624094195663929, + 0.017624886706471443, + 0.05610744282603264, + 0.17589883506298065, + 0.02641749009490013, + 0.13583002984523773, + 0.030089380219578743, + 0.045959874987602234, + 0.11335926502943039 + ], + "credit_map": [ + 0.029181798920035362, + 0.034785233438014984, + 0.03578236699104309, + 0.03284770995378494, + 0.08957996219396591, + 0.01899181865155697, + 0.04661635681986809, + 0.04851814731955528, + 0.006891224067658186, + 0.007859201170504093, + 0.007645354140549898, + 0.010365751571953297, + 0.008653836324810982, + 0.0069933924823999405, + 0.007455405779182911, + 0.006545261014252901, + 0.001746929599903524, + 0.0021153101697564125, + 0.0014687295770272613, + 0.0029403928201645613, + 0.0029554241336882114, + 0.0022322742734104395, + 0.002433215733617544, + 0.001732613891363144, + 0.004958460573107004, + 0.005727189593017101, + 0.006392878945916891, + 0.008050834760069847, + 0.014342385344207287, + 0.004480182193219662, + 0.006368664093315601, + 0.005786858033388853, + 0.014880978502333164, + 0.009836164303123951, + 0.01400125864893198, + 0.04574507102370262, + 0.01777106523513794, + 0.009601510129868984, + 0.042364701628685, + 0.02169809862971306, + 0.002792740473523736, + 0.0036803032271564007, + 0.0026689355727285147, + 0.003961358685046434, + 0.004558966960757971, + 0.0021596504375338554, + 0.003280459437519312, + 0.0033150771632790565, + 0.010590777732431889, + 0.013264380395412445, + 0.012512072920799255, + 0.02273908071219921, + 0.028726933524012566, + 0.016438428312540054, + 0.0181086678057909, + 0.013449695892632008, + 0.004263645503669977, + 0.003318891627714038, + 0.0029985967557877302, + 0.004654353484511375, + 0.004814615938812494, + 0.0023879818618297577, + 0.0044081504456698895, + 0.003243148559704423, + 0.00497320294380188, + 0.007901150733232498, + 0.005178408231586218, + 0.006167793180793524, + 0.004964266438037157, + 0.005224641878157854, + 0.006438965909183025, + 0.005111446604132652, + 0.011038042604923248, + 0.01836727187037468, + 0.011600102297961712, + 0.020931782200932503, + 0.007618685252964497, + 0.00862905103713274, + 0.022120961919426918, + 0.01305337529629469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.02330857142806053 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static shot taken directly from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street with dark asphalt reflecting neon lights and streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a soaked purple suit with dyed green curls plastered to his scalp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's white clown makeup washed away by rain, blending with black eyeliner and blue teardrop patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man kneeling with his hands supporting the ground while his shoulders shake in a tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789884552359581 + ], + "opposite_logprob_tokens": [ + -2.578988552093506 + ], + "logprob_reward": -0.0789884552359581, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789884552359581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18092063069343567, + 0.025981416925787926, + 0.02420240081846714, + 0.06050041317939758, + 0.4758203327655792, + 0.02092728577554226, + 0.09451121091842651, + 0.036352284252643585, + 0.034919898957014084, + 0.04586413875222206 + ], + "credit_map": [ + 0.012411531060934067, + 0.023923778906464577, + 0.013670054264366627, + 0.0246686153113842, + 0.045940689742565155, + 0.009659434668719769, + 0.028231466189026833, + 0.022415073588490486, + 0.003182825865224004, + 0.002776263514533639, + 0.002967879408970475, + 0.004440704360604286, + 0.0036263028159737587, + 0.0028349587228149176, + 0.003330968553200364, + 0.0028215115889906883, + 0.002316787140443921, + 0.003250768408179283, + 0.0025055240839719772, + 0.003487412352114916, + 0.004537310916930437, + 0.0017980302218347788, + 0.003959040157496929, + 0.002347525442019105, + 0.004849092103540897, + 0.006575671024620533, + 0.006828881800174713, + 0.006934927776455879, + 0.016662947833538055, + 0.00369876972399652, + 0.008018864318728447, + 0.0069312602281570435, + 0.03652825951576233, + 0.02872859314084053, + 0.03988863155245781, + 0.12014459818601608, + 0.053717322647571564, + 0.022741708904504776, + 0.1167023554444313, + 0.05736885219812393, + 0.0020188128110021353, + 0.002279464853927493, + 0.0015891666989773512, + 0.0034509620163589716, + 0.0029105122666805983, + 0.0019151153974235058, + 0.002924026222899556, + 0.0038392250426113605, + 0.00823754072189331, + 0.008764653466641903, + 0.00748241925612092, + 0.009775836952030659, + 0.01821945421397686, + 0.010657044127583504, + 0.01724855601787567, + 0.014125708490610123, + 0.0038747561629861593, + 0.003973083570599556, + 0.0029922104440629482, + 0.009605087339878082, + 0.0043192352168262005, + 0.002925873501226306, + 0.005259746685624123, + 0.0034022920299321413, + 0.004285392351448536, + 0.005632480606436729, + 0.003370242891833186, + 0.007347595877945423, + 0.0030062056612223387, + 0.002919054590165615, + 0.005510085728019476, + 0.0028488426469266415, + 0.007258519995957613, + 0.0072739701718091965, + 0.006470352411270142, + 0.006574776489287615, + 0.0029170960187911987, + 0.002922598971053958, + 0.006788184866309166, + 0.005658639594912529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.0789884552359581 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_9/index_tensor([2787, 2787])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_9/index_tensor([2787, 2787])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d17d1ba045a84b8e3fda7f7fb0b0dcc5aa7b7088 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_9/index_tensor([2787, 2787])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene unfolds in extreme close-up, focusing on a single, twinkling point of light in the absolute darkness of deep space. A majestic, slow pullback begins, revealing this point of light to be no simple star, but the heart of a cosmic phenomenon. It begins to bloom gracefully, its core intensifying to a brilliant diamond-white before unfurling delicate, petal-like tendrils of sapphire and fiery red-orange clouds of gas, like a time-lapse of a flower blossoming in the sky. As the camera continues its steady, smooth retreat, countless other distant stars in the void begin the same breathtaking process, each a unique cosmic flower, bursting into a brilliant, multi-colored glow. A slow, relentless zoom-out gradually reveals these individual blooming stars as part of a much grander structure, their light merging into the swirling arms of a vast spiral galaxy. The final frame, framed by a breathtaking wide-angle shot, showcases the entire Milky Way in all its splendor—a majestic celestial maelstrom with a dense, fiery golden core, sweeping arms dotted with young blue stars and dark, intricate dust lanes, all set against the silent, star-studded velvet universe, creating an atmosphere of profound awe and tranquil cosmic grandeur.", + "seed": 24247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_9/index_2787_seed_24247_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16682510287500918, + "vlm_logprob": -0.16682510287500918, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.16682510287500918, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16147911548614502 + ], + "opposite_logprob_tokens": [ + -1.911479115486145 + ], + "logprob_reward": -0.16147911548614502, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16147911548614502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17441049218177795, + 0.146065816283226, + 0.28985628485679626, + 0.13116493821144104, + 0.08201245218515396, + 0.02253931760787964, + 0.04057064652442932, + 0.016099104657769203, + 0.07625816017389297, + 0.02102275751531124 + ], + "credit_map": [ + 0.020957404747605324, + 0.019855596125125885, + 0.031175125390291214, + 0.02257119119167328, + 0.024373244494199753, + 0.019357386976480484, + 0.017641013488173485, + 0.01847952790558338, + 0.012933481484651566, + 0.019922679290175438, + 0.024561172351241112, + 0.022362997755408287, + 0.015204117633402348, + 0.015733661130070686, + 0.018353722989559174, + 0.016993990167975426, + 0.10609142482280731, + 0.03146909922361374, + 0.05553452670574188, + 0.011701062321662903, + 0.009856983087956905, + 0.019512996077537537, + 0.014555776491761208, + 0.04113440588116646, + 0.01227665413171053, + 0.015001406893134117, + 0.018273837864398956, + 0.01517294067889452, + 0.01745147444307804, + 0.01886306330561638, + 0.019081953912973404, + 0.015043618157505989, + 0.008219761773943901, + 0.007414135616272688, + 0.012233651243150234, + 0.007587467785924673, + 0.015587040223181248, + 0.009872577153146267, + 0.010511970147490501, + 0.010585848242044449, + 0.0015916448319330812, + 0.003112784819677472, + 0.0027812866028398275, + 0.003066528122872114, + 0.0028654979541897774, + 0.003058946691453457, + 0.003131207777187228, + 0.0029314213898032904, + 0.0028636339120566845, + 0.0037093262653797865, + 0.0041352566331624985, + 0.003234811592847109, + 0.007681495044380426, + 0.006843943148851395, + 0.006612889934331179, + 0.005489290226250887, + 0.0023703884799033403, + 0.0019524741219356656, + 0.002168846782296896, + 0.0017628390341997147, + 0.0012935352278873324, + 0.00210355082526803, + 0.0023205429315567017, + 0.0021269286517053843, + 0.002777926856651902, + 0.001320810872130096, + 0.003201743122190237, + 0.002672309521585703, + 0.004270809702575207, + 0.027643674984574318, + 0.0034016810823231936, + 0.030969208106398582, + 0.002252997597679496, + 0.001917723217047751, + 0.003253048285841942, + 0.0022553023882210255, + 0.002952615963295102, + 0.0028308825567364693, + 0.002877002814784646, + 0.002683185273781419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.16147911548614502 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001348539488390088 + ], + "opposite_logprob_tokens": [ + -6.626348495483398 + ], + "logprob_reward": -0.001348539488390088, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001348539488390088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059251584112644196, + 0.0962669625878334, + 0.13957318663597107, + 0.11172708868980408, + 0.019938940182328224, + 0.029955249279737473, + 0.037162236869335175, + 0.06964359432458878, + 0.3570808470249176, + 0.0794002115726471 + ], + "credit_map": [ + 0.008782432414591312, + 0.008958039805293083, + 0.009145134128630161, + 0.006830033380538225, + 0.0076209744438529015, + 0.005666562356054783, + 0.006178057752549648, + 0.006070357281714678, + 0.009530060924589634, + 0.010939078405499458, + 0.017887301743030548, + 0.013404516503214836, + 0.010247425176203251, + 0.010670684278011322, + 0.012713924050331116, + 0.01087398361414671, + 0.05091505125164986, + 0.015451506711542606, + 0.025799935683608055, + 0.005780305713415146, + 0.004515219479799271, + 0.009445334784686565, + 0.006703849416226149, + 0.02096199244260788, + 0.016758516430854797, + 0.010917668230831623, + 0.01521954033523798, + 0.012251167558133602, + 0.012072599492967129, + 0.0134462034329772, + 0.013432225212454796, + 0.01762918382883072, + 0.002338845981284976, + 0.0020533320493996143, + 0.0024003058206290007, + 0.0018509307410567999, + 0.0036284709349274635, + 0.0026612565852701664, + 0.002423466881737113, + 0.0025823337491601706, + 0.002003156114369631, + 0.003441995708271861, + 0.002960503101348877, + 0.002313792472705245, + 0.006015074905008078, + 0.004124689847230911, + 0.004837492946535349, + 0.004258547443896532, + 0.0030832388438284397, + 0.004144728183746338, + 0.004705430008471012, + 0.003016469068825245, + 0.006317195016890764, + 0.00544354971498251, + 0.005782053805887699, + 0.004669577814638615, + 0.006645791698247194, + 0.002472860738635063, + 0.003939683083444834, + 0.0026302938349545, + 0.004318362567573786, + 0.021395251154899597, + 0.004186810459941626, + 0.024054545909166336, + 0.012344373390078545, + 0.00673677958548069, + 0.02426588162779808, + 0.0124550461769104, + 0.01673208735883236, + 0.1191319152712822, + 0.013572671450674534, + 0.1518421322107315, + 0.010004096664488316, + 0.00654631620272994, + 0.012883090414106846, + 0.007780975196510553, + 0.009430100210011005, + 0.009464542381465435, + 0.01158286165446043, + 0.011708242818713188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.001348539488390088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807578772306442 + ], + "opposite_logprob_tokens": [ + -3.5298075675964355 + ], + "logprob_reward": -0.029807578772306442, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807578772306442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037682145833969116, + 0.04447830095887184, + 0.050332773476839066, + 0.07011373341083527, + 0.047621581703424454, + 0.024277715012431145, + 0.1279974728822708, + 0.13531586527824402, + 0.4102436304092407, + 0.051936835050582886 + ], + "credit_map": [ + 0.005347913131117821, + 0.004463914781808853, + 0.0055401986464858055, + 0.0036869256291538477, + 0.006054689176380634, + 0.004320183303207159, + 0.0043114866130054, + 0.003956829197704792, + 0.003979253116995096, + 0.0054993354715406895, + 0.00669461814686656, + 0.0054499302059412, + 0.005034052766859531, + 0.0067437454126775265, + 0.005034198984503746, + 0.006043164525181055, + 0.016329260542988777, + 0.005253832787275314, + 0.008237053640186787, + 0.00332897761836648, + 0.0026081297546625137, + 0.00397347379475832, + 0.0035636317916214466, + 0.007038408424705267, + 0.016522666439414024, + 0.0044895438477396965, + 0.008382430300116539, + 0.003838911885395646, + 0.0068923188373446465, + 0.008373982273042202, + 0.00705003272742033, + 0.014563842676579952, + 0.005628359504044056, + 0.00680721178650856, + 0.006798784248530865, + 0.004945183172821999, + 0.006389893591403961, + 0.004750680178403854, + 0.006060374900698662, + 0.006241091527044773, + 0.0018311688909307122, + 0.004562928806990385, + 0.00228006555698812, + 0.003146390663459897, + 0.0037764734588563442, + 0.0028451094403862953, + 0.0026640223804861307, + 0.0031715529039502144, + 0.00799576286226511, + 0.014188151806592941, + 0.011740189976990223, + 0.00879458524286747, + 0.02980497106909752, + 0.019665176048874855, + 0.01669747568666935, + 0.01911114528775215, + 0.010319585911929607, + 0.005105376243591309, + 0.007881006225943565, + 0.005867558065801859, + 0.01066703163087368, + 0.04151009023189545, + 0.009793169796466827, + 0.04417204111814499, + 0.013236475177109241, + 0.007527229841798544, + 0.015025997534394264, + 0.014100924134254456, + 0.022260218858718872, + 0.14976821839809418, + 0.01590404286980629, + 0.172420471906662, + 0.007426418364048004, + 0.0040050530806183815, + 0.009666729718446732, + 0.004215340595692396, + 0.006569381803274155, + 0.006796571891754866, + 0.0061625344678759575, + 0.007094803266227245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.029807578772306442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07659972459077835, + 0.0815516859292984, + 0.1880815625190735, + 0.09956551343202591, + 0.05451923981308937, + 0.08359592407941818, + 0.1061253547668457, + 0.10063760727643967, + 0.15748898684978485, + 0.05183437466621399 + ], + "credit_map": [ + 0.009317789226770401, + 0.009390875697135925, + 0.014019805006682873, + 0.008583598770201206, + 0.010872011072933674, + 0.00798496138304472, + 0.008855445310473442, + 0.007575237192213535, + 0.007716215215623379, + 0.01171440351754427, + 0.0135163813829422, + 0.010880079120397568, + 0.010072209872305393, + 0.010300974361598492, + 0.008134864270687103, + 0.009216558188199997, + 0.06781931221485138, + 0.02202150784432888, + 0.03314690291881561, + 0.009673090651631355, + 0.007044217549264431, + 0.012693887576460838, + 0.007660743314772844, + 0.028021911159157753, + 0.017259199172258377, + 0.00781446322798729, + 0.011725729331374168, + 0.008390572853386402, + 0.012228959240019321, + 0.011926845647394657, + 0.012732546776533127, + 0.017487196251749992, + 0.004598104394972324, + 0.005738044157624245, + 0.005972002167254686, + 0.005650961305946112, + 0.009185317903757095, + 0.0075662946328520775, + 0.007291075307875872, + 0.008517439477145672, + 0.006401644088327885, + 0.012777021154761314, + 0.011218009516596794, + 0.009792950004339218, + 0.00856416393071413, + 0.014273939654231071, + 0.009094316512346268, + 0.011473880149424076, + 0.009524945169687271, + 0.00998733565211296, + 0.011466906405985355, + 0.009877068921923637, + 0.01953023299574852, + 0.01797735132277012, + 0.013113514520227909, + 0.014647997915744781, + 0.00829580519348383, + 0.007084283512085676, + 0.006974838674068451, + 0.0043602921068668365, + 0.009385914541780949, + 0.025428321212530136, + 0.011204121634364128, + 0.02790403738617897, + 0.010392464697360992, + 0.007844326086342335, + 0.026152674108743668, + 0.00786555651575327, + 0.01640411652624607, + 0.047295939177274704, + 0.011402097530663013, + 0.03013182431459427, + 0.0046988558024168015, + 0.006017008330672979, + 0.011984791606664658, + 0.004315948113799095, + 0.0050383638590574265, + 0.006452001165598631, + 0.006974005606025457, + 0.006353400647640228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.008639931678771973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328503489494324 + ], + "opposite_logprob_tokens": [ + -0.7578503489494324 + ], + "logprob_reward": -0.6328503489494324, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328503489494324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03350279480218887, + 0.03752506151795387, + 0.20648424327373505, + 0.055872220546007156, + 0.017690038308501244, + 0.0347331166267395, + 0.07349904626607895, + 0.04508382827043533, + 0.34839797019958496, + 0.14721164107322693 + ], + "credit_map": [ + 0.004429640248417854, + 0.00559183768928051, + 0.006296062376350164, + 0.004291918128728867, + 0.0034164946991950274, + 0.002912391209974885, + 0.002832785015925765, + 0.00373166985809803, + 0.003519446821883321, + 0.005640791729092598, + 0.008046162314713001, + 0.005470195319503546, + 0.003625312587246299, + 0.003830434288829565, + 0.003677201457321644, + 0.003715517930686474, + 0.07683509588241577, + 0.023075943812727928, + 0.03930583596229553, + 0.008060303516685963, + 0.007468904834240675, + 0.013132978230714798, + 0.008774531073868275, + 0.029830647632479668, + 0.013044161722064018, + 0.0031292985659092665, + 0.006593051832169294, + 0.0034547769464552402, + 0.004051088355481625, + 0.0068395668640732765, + 0.0057159909047186375, + 0.013044288381934166, + 0.0018912454834207892, + 0.003034950466826558, + 0.0023699188604950905, + 0.002718597650527954, + 0.0019112275913357735, + 0.0015683135716244578, + 0.0018008376937359571, + 0.0023949481546878815, + 0.0029256197158247232, + 0.007376417052000761, + 0.004946430213749409, + 0.00429916987195611, + 0.003329973667860031, + 0.0035310755483806133, + 0.00432334840297699, + 0.0040010870434343815, + 0.0053895702585577965, + 0.005508014466613531, + 0.008801103569567204, + 0.004624067340046167, + 0.017545580863952637, + 0.01100107654929161, + 0.010807336308062077, + 0.0098223015666008, + 0.005196486599743366, + 0.002514334162697196, + 0.0031129566486924887, + 0.0019085888052359223, + 0.003439453663304448, + 0.010810942389070988, + 0.0055179656483232975, + 0.01258309930562973, + 0.01277197990566492, + 0.00566523103043437, + 0.01913510449230671, + 0.012429465539753437, + 0.01954818330705166, + 0.12993164360523224, + 0.014170766808092594, + 0.13474564254283905, + 0.020210329443216324, + 0.009922965429723263, + 0.023527728393673897, + 0.013787832111120224, + 0.018939554691314697, + 0.02003643661737442, + 0.02055390551686287, + 0.020232900977134705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.6328503489494324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in extreme close-up, focusing on a single, twinkling point of light in the absolute darkness of deep space. A majestic, slow pullback begins, revealing this point of light to be no simple star, but the heart of a cosmic phenomenon. It begins to bloom gracefully, its core intensifying to a brilliant diamond-white before unfurling delicate, petal-like tendrils of sapphire and fiery red-orange clouds of gas, like a time-lapse of a flower blossoming in the sky. As the camera continues its steady, smooth retreat, countless other distant stars in the void begin the same breathtaking process, each a unique cosmic flower, bursting into a brilliant, multi-colored glow. A slow, relentless zoom-out gradually reveals these individual blooming stars as part of a much grander structure, their light merging into the swirling arms of a vast spiral galaxy. The final frame, framed by a breathtaking wide-angle shot, showcases the entire Milky Way in all its splendor—a majestic celestial maelstrom with a dense, fiery golden core, sweeping arms dotted with young blue stars and dark, intricate dust lanes, all set against the silent, star-studded velvet universe, creating an atmosphere of profound awe and tranquil cosmic grandeur.", + "seed": 24248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_9/index_2787_seed_24248_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.26817914107814433, + "vlm_logprob": -0.26817914107814433, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.26817914107814433, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5773079991340637 + ], + "opposite_logprob_tokens": [ + -0.8273079991340637 + ], + "logprob_reward": -0.5773079991340637, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5773079991340637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18720735609531403, + 0.36384934186935425, + 0.1317376345396042, + 0.020654940977692604, + 0.015431822277605534, + 0.01605616696178913, + 0.00990488100796938, + 0.020605333149433136, + 0.2198706418275833, + 0.014681907370686531 + ], + "credit_map": [ + 0.02441340498626232, + 0.030444419011473656, + 0.03254823759198189, + 0.033936887979507446, + 0.01800287887454033, + 0.01783113181591034, + 0.012518245726823807, + 0.017512153834104538, + 0.14662061631679535, + 0.016478169709444046, + 0.15276354551315308, + 0.021748118102550507, + 0.007023338694125414, + 0.004141802433878183, + 0.006620076484978199, + 0.008453667163848877, + 0.027116265147924423, + 0.014170533046126366, + 0.04127145931124687, + 0.014503703452646732, + 0.004270507022738457, + 0.005852196365594864, + 0.005670675542205572, + 0.01888228766620159, + 0.0029297040309756994, + 0.0028072756249457598, + 0.003058777190744877, + 0.001988910371437669, + 0.0023984862491488457, + 0.0029493188485503197, + 0.002272115321829915, + 0.002250353340059519, + 0.0010858074529096484, + 0.0010804400080814958, + 0.0010921639623120427, + 0.0011610048823058605, + 0.0015071275411173701, + 0.003970325458794832, + 0.0013921762583777308, + 0.0041427770629525185, + 0.0009165764786303043, + 0.0021872881334275007, + 0.0012418002588674426, + 0.0008435692288912833, + 0.0012715666089206934, + 0.0038336957804858685, + 0.0007925046375021338, + 0.004969165660440922, + 0.001414837664924562, + 0.0007617336814291775, + 0.001315033994615078, + 0.0008054827922023833, + 0.001250626053661108, + 0.0016172403702512383, + 0.0008506488520652056, + 0.001889277482405305, + 0.0034202695824205875, + 0.002254235092550516, + 0.002084786305204034, + 0.002599841682240367, + 0.0013626848813146353, + 0.002627928275614977, + 0.0020143932197242975, + 0.0042411936447024345, + 0.08497877418994904, + 0.007330150809139013, + 0.008454076945781708, + 0.004284375812858343, + 0.008952737785875797, + 0.08140311390161514, + 0.005510496441274881, + 0.018956925719976425, + 0.004186998121440411, + 0.0008839774527586997, + 0.0015187750104814768, + 0.0010785540798678994, + 0.0007183411507867277, + 0.0017012882744893432, + 0.0009113528649322689, + 0.0036826201248914003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.5773079991340637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272767972201109 + ], + "opposite_logprob_tokens": [ + -5.25527286529541 + ], + "logprob_reward": -0.005272767972201109, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272767972201109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1158892884850502, + 0.09317602217197418, + 0.0803927555680275, + 0.0309850312769413, + 0.04276232793927193, + 0.0067118089646101, + 0.011445360258221626, + 0.05314391851425171, + 0.5022396445274353, + 0.06325380504131317 + ], + "credit_map": [ + 0.028189264237880707, + 0.013287861831486225, + 0.03174448385834694, + 0.013965240679681301, + 0.0095358369871974, + 0.005515583790838718, + 0.005906533915549517, + 0.007744492497295141, + 0.03346388414502144, + 0.004709590692073107, + 0.03438416123390198, + 0.0056448038667440414, + 0.003465954214334488, + 0.004304942674934864, + 0.0028021601028740406, + 0.004400528501719236, + 0.020069722086191177, + 0.007298384327441454, + 0.03353498503565788, + 0.005536098964512348, + 0.0026172236539423466, + 0.0019446173682808876, + 0.002856484381482005, + 0.006535243708640337, + 0.004201096016913652, + 0.004212215542793274, + 0.004779231268912554, + 0.0028549612034112215, + 0.0036081611178815365, + 0.0035616587847471237, + 0.0038848337717354298, + 0.003882872872054577, + 0.0015916133997961879, + 0.0011743436334654689, + 0.0019276647362858057, + 0.0014629361685365438, + 0.0020109512843191624, + 0.017244916409254074, + 0.001597532769665122, + 0.015752369537949562, + 0.0007651724736206234, + 0.0011604182654991746, + 0.0007160111563280225, + 0.0007216630037873983, + 0.000677927047945559, + 0.0011994304368272424, + 0.0006011845543980598, + 0.0008700016187503934, + 0.001406585332006216, + 0.0014916181098669767, + 0.0010520749492570758, + 0.0014123722212389112, + 0.001304428675211966, + 0.0013839228777214885, + 0.0012105150381103158, + 0.002183843171223998, + 0.005204491782933474, + 0.0017972681671380997, + 0.0022592737805098295, + 0.0019273096695542336, + 0.0028346339240670204, + 0.01697694882750511, + 0.0030865841545164585, + 0.019057409837841988, + 0.1884157359600067, + 0.01674279384315014, + 0.018706105649471283, + 0.011672652326524258, + 0.020250562578439713, + 0.186266228556633, + 0.014448482543230057, + 0.04573708400130272, + 0.023765770718455315, + 0.0022177835926413536, + 0.0031936534214764833, + 0.002363505307585001, + 0.0020310357213020325, + 0.007818383164703846, + 0.001872298656962812, + 0.019991369917988777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.005272767972201109 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295013129711151 + ], + "opposite_logprob_tokens": [ + -2.237950086593628 + ], + "logprob_reward": -0.11295013129711151, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295013129711151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06328589469194412, + 0.10222560167312622, + 0.0702638328075409, + 0.03306610509753227, + 0.03440912812948227, + 0.009872369468212128, + 0.020946601405739784, + 0.04259244725108147, + 0.5567712187767029, + 0.0665668398141861 + ], + "credit_map": [ + 0.014038889668881893, + 0.005872749723494053, + 0.014267689548432827, + 0.007718907669186592, + 0.0058157057501375675, + 0.0041594430804252625, + 0.004257268738001585, + 0.00715523911640048, + 0.036164745688438416, + 0.006975978147238493, + 0.03688580170273781, + 0.006673483643680811, + 0.004687559325248003, + 0.0039462982676923275, + 0.003045557765290141, + 0.0038461824879050255, + 0.0062678419053554535, + 0.024376999586820602, + 0.019296720623970032, + 0.007442169357091188, + 0.005231755785644054, + 0.0021204480435699224, + 0.002338416175916791, + 0.0031894834246486425, + 0.005267773754894733, + 0.004422005731612444, + 0.004402458667755127, + 0.0027585974894464016, + 0.0031178880017250776, + 0.005310941021889448, + 0.0035277686547487974, + 0.004258672706782818, + 0.0027964916080236435, + 0.002289598574861884, + 0.0028221714310348034, + 0.002290918957442045, + 0.003317652503028512, + 0.00881976168602705, + 0.003365899436175823, + 0.008706633001565933, + 0.0009119813330471516, + 0.000853710516821593, + 0.0011477553052827716, + 0.0009349760366603732, + 0.001173115335404873, + 0.0018486300250515342, + 0.000828165968414396, + 0.0021740347146987915, + 0.0025301300920546055, + 0.002394182374700904, + 0.0029650877695530653, + 0.0030654824804514647, + 0.002393407514318824, + 0.0025466992519795895, + 0.0023471906315535307, + 0.002704419894143939, + 0.0064581818878650665, + 0.0032638160046190023, + 0.004588038194924593, + 0.004176221322268248, + 0.0032473027240484953, + 0.005108141340315342, + 0.005008789245039225, + 0.010741958394646645, + 0.18495634198188782, + 0.016345607116818428, + 0.019493799656629562, + 0.011395333334803581, + 0.021431853994727135, + 0.20988985896110535, + 0.020877471193671227, + 0.07238097488880157, + 0.012340154498815536, + 0.003633539192378521, + 0.012531915679574013, + 0.003827157197520137, + 0.005075915716588497, + 0.012554430402815342, + 0.00358473788946867, + 0.013018989935517311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.11295013129711151 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542770244181156 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542770244181156, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542770244181156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07117722183465958, + 0.2987009882926941, + 0.1206885501742363, + 0.08701635897159576, + 0.08429741114377975, + 0.013020765036344528, + 0.015373220667243004, + 0.12246216833591461, + 0.1303127557039261, + 0.05695055425167084 + ], + "credit_map": [ + 0.016223808750510216, + 0.008063235320150852, + 0.01813497208058834, + 0.008708011358976364, + 0.005934952292591333, + 0.004006440285593271, + 0.00439972011372447, + 0.0057060797698795795, + 0.12017248570919037, + 0.013782755471765995, + 0.12278205901384354, + 0.01750212348997593, + 0.006937726400792599, + 0.004465590696781874, + 0.005935194436460733, + 0.007123053539544344, + 0.02268463931977749, + 0.023801591247320175, + 0.046593599021434784, + 0.008688997477293015, + 0.005457267630845308, + 0.003713182872161269, + 0.0035008410923182964, + 0.006248433608561754, + 0.013282573781907558, + 0.010749829933047295, + 0.013566728681325912, + 0.007825980894267559, + 0.01096141804009676, + 0.00968310609459877, + 0.01114572025835514, + 0.009801000356674194, + 0.004354981705546379, + 0.0031181671656668186, + 0.00491544883698225, + 0.0031409875955432653, + 0.0037319709081202745, + 0.03185632452368736, + 0.003508227411657572, + 0.029671307653188705, + 0.001250329427421093, + 0.0011260277824476361, + 0.0010504608508199453, + 0.0009859356796368957, + 0.0010491402354091406, + 0.0029939517844468355, + 0.001060356036759913, + 0.003504561958834529, + 0.0020683673210442066, + 0.0015130273532122374, + 0.0021340064704418182, + 0.0016861810581758618, + 0.0017283086199313402, + 0.002092378679662943, + 0.0014823719393461943, + 0.002668580273166299, + 0.013986645266413689, + 0.004924105014652014, + 0.006643418688327074, + 0.006177451927214861, + 0.0062263221479952335, + 0.03323693200945854, + 0.008386856876313686, + 0.04288044199347496, + 0.04877649247646332, + 0.005768903996795416, + 0.005709354300051928, + 0.006851929239928722, + 0.005466551519930363, + 0.043186698108911514, + 0.004231676459312439, + 0.01032115239650011, + 0.014337290078401566, + 0.003396189073100686, + 0.007660752162337303, + 0.004316077567636967, + 0.004020560998469591, + 0.010044076479971409, + 0.0029695823322981596, + 0.010206027887761593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.012542770244181156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328220367431641 + ], + "opposite_logprob_tokens": [ + -0.7578220367431641 + ], + "logprob_reward": -0.6328220367431641, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328220367431641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09206827729940414, + 0.09613412618637085, + 0.0860861986875534, + 0.02598244696855545, + 0.026066239923238754, + 0.008622718043625355, + 0.02601972408592701, + 0.5594948530197144, + 0.01891169510781765, + 0.06061375141143799 + ], + "credit_map": [ + 0.02161659672856331, + 0.009130147285759449, + 0.023726940155029297, + 0.01193321030586958, + 0.007171879522502422, + 0.004171420820057392, + 0.005606344435364008, + 0.008711731061339378, + 0.03635648638010025, + 0.00464958930388093, + 0.03771626576781273, + 0.006369651295244694, + 0.002797620603814721, + 0.0021659601479768753, + 0.003091849386692047, + 0.0029866951517760754, + 0.016662871465086937, + 0.01038554310798645, + 0.033058010041713715, + 0.00658971955999732, + 0.004535441752523184, + 0.0030910344794392586, + 0.0033286763355135918, + 0.00843488983809948, + 0.004454200156033039, + 0.0024841129779815674, + 0.0036715047899633646, + 0.0029366984963417053, + 0.0032790047116577625, + 0.003125230548903346, + 0.002581017091870308, + 0.0034506763331592083, + 0.0022457295563071966, + 0.003447282360866666, + 0.003593046683818102, + 0.0018849533516913652, + 0.0022260616533458233, + 0.005245676729828119, + 0.0023703966289758682, + 0.005053090397268534, + 0.0008622817695140839, + 0.0007457114988937974, + 0.0009482155437581241, + 0.0008932192577049136, + 0.0013510112185031176, + 0.0015678262570872903, + 0.0008367090485990047, + 0.001417742227204144, + 0.0038761799223721027, + 0.0024237781763076782, + 0.0026784015353769064, + 0.002390018431469798, + 0.002467552199959755, + 0.00397150032222271, + 0.0026412447914481163, + 0.005571045912802219, + 0.047267165035009384, + 0.023117195814847946, + 0.020792948082089424, + 0.015642328187823296, + 0.0268813855946064, + 0.20699584484100342, + 0.024654854089021683, + 0.19414307177066803, + 0.006579237524420023, + 0.0010740048019215465, + 0.0010570517042651772, + 0.0014318199828267097, + 0.0008851669845171273, + 0.0044967555440962315, + 0.0009316631476394832, + 0.0024559935554862022, + 0.018589114770293236, + 0.002845813985913992, + 0.007327504456043243, + 0.002590510994195938, + 0.0027400334365665913, + 0.006814140826463699, + 0.0028722172137349844, + 0.016834409907460213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.6328220367431641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in extreme close-up, focusing on a single, twinkling point of light in the absolute darkness of deep space. A majestic, slow pullback begins, revealing this point of light to be no simple star, but the heart of a cosmic phenomenon. It begins to bloom gracefully, its core intensifying to a brilliant diamond-white before unfurling delicate, petal-like tendrils of sapphire and fiery red-orange clouds of gas, like a time-lapse of a flower blossoming in the sky. As the camera continues its steady, smooth retreat, countless other distant stars in the void begin the same breathtaking process, each a unique cosmic flower, bursting into a brilliant, multi-colored glow. A slow, relentless zoom-out gradually reveals these individual blooming stars as part of a much grander structure, their light merging into the swirling arms of a vast spiral galaxy. The final frame, framed by a breathtaking wide-angle shot, showcases the entire Milky Way in all its splendor—a majestic celestial maelstrom with a dense, fiery golden core, sweeping arms dotted with young blue stars and dark, intricate dust lanes, all set against the silent, star-studded velvet universe, creating an atmosphere of profound awe and tranquil cosmic grandeur.", + "seed": 24249, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_9/index_2787_seed_24249_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08554524183273315, + "vlm_logprob": -0.08554524183273315, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.08554524183273315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07085529714822769 + ], + "opposite_logprob_tokens": [ + -2.6958553791046143 + ], + "logprob_reward": -0.07085529714822769, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07085529714822769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18661139905452728, + 0.06749577820301056, + 0.14818419516086578, + 0.06667529046535492, + 0.02571721188724041, + 0.10645772516727448, + 0.007423057686537504, + 0.03156551346182823, + 0.025906523689627647, + 0.3339633047580719 + ], + "credit_map": [ + 0.01678755320608616, + 0.02591460756957531, + 0.014669036492705345, + 0.02129349485039711, + 0.030038299039006233, + 0.02690865844488144, + 0.023817239329218864, + 0.02718251384794712, + 0.005363205447793007, + 0.013382462784647942, + 0.006215960253030062, + 0.012295412831008434, + 0.010325168259441853, + 0.00832421611994505, + 0.007360272109508514, + 0.0042290762066841125, + 0.02301601693034172, + 0.014538194052875042, + 0.008033971302211285, + 0.03189920634031296, + 0.008533825166523457, + 0.03688205033540726, + 0.009457794949412346, + 0.015823127701878548, + 0.012147066183388233, + 0.006394774653017521, + 0.007873952388763428, + 0.005010787397623062, + 0.0066267079673707485, + 0.010119586251676083, + 0.00805604737251997, + 0.010446369647979736, + 0.004021334461867809, + 0.0028665887657552958, + 0.0025586767587810755, + 0.005937622394412756, + 0.0018139741150662303, + 0.002493575681000948, + 0.0034592782612890005, + 0.002566160634160042, + 0.005935132037848234, + 0.011170295998454094, + 0.005791490897536278, + 0.03955280780792236, + 0.003973428159952164, + 0.006100167520344257, + 0.021777385845780373, + 0.012157009914517403, + 0.0006562257767654955, + 0.0011420154478400946, + 0.0009441162692382932, + 0.0010586109710857272, + 0.001004982739686966, + 0.0009596660383976996, + 0.000788070319686085, + 0.0008693707641214132, + 0.0027840377297252417, + 0.0031786307226866484, + 0.003345775417983532, + 0.004168869461864233, + 0.00538738165050745, + 0.004056304693222046, + 0.0054037561640143394, + 0.003240758553147316, + 0.0036121008452028036, + 0.003546131309121847, + 0.0013112430460751057, + 0.0017151528736576438, + 0.004215172957628965, + 0.0082266079261899, + 0.0018509520450606942, + 0.0014291624538600445, + 0.04676050320267677, + 0.026384495198726654, + 0.022916972637176514, + 0.028629183769226074, + 0.03938084840774536, + 0.045626070350408554, + 0.10114078968763351, + 0.02312444895505905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07085529714822769 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006753832101821899 + ], + "opposite_logprob_tokens": [ + -5.006753921508789 + ], + "logprob_reward": -0.006753832101821899, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006753832101821899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061061959713697433, + 0.0592891089618206, + 0.3165440261363983, + 0.047030169516801834, + 0.045055828988552094, + 0.017792198807001114, + 0.012081240303814411, + 0.036335289478302, + 0.027748411521315575, + 0.37706172466278076 + ], + "credit_map": [ + 0.008098497055470943, + 0.006646196357905865, + 0.005312866996973753, + 0.0067178718745708466, + 0.008534048683941364, + 0.0077937631867825985, + 0.006864122115075588, + 0.01109459437429905, + 0.00559755926951766, + 0.00964843574911356, + 0.006133413407951593, + 0.009503689594566822, + 0.008585596457123756, + 0.0069259898737072945, + 0.00741712749004364, + 0.005477298051118851, + 0.11343052238225937, + 0.018025562167167664, + 0.01538151130080223, + 0.014018149115145206, + 0.01244368776679039, + 0.09242559224367142, + 0.021358774974942207, + 0.02946021780371666, + 0.005793084856122732, + 0.005349434912204742, + 0.004579020664095879, + 0.006899235770106316, + 0.004849917721003294, + 0.006321340333670378, + 0.006527294404804707, + 0.006710842251777649, + 0.005026643164455891, + 0.005148347467184067, + 0.005261386279016733, + 0.009320671670138836, + 0.004321696236729622, + 0.004899094346910715, + 0.00620860792696476, + 0.004869379103183746, + 0.002010644879192114, + 0.0023030522279441357, + 0.002132955240085721, + 0.0027221543714404106, + 0.0019598458893597126, + 0.0022484285291284323, + 0.0026599669363349676, + 0.0017551507335156202, + 0.0013136828783899546, + 0.0015072927344590425, + 0.001682800822891295, + 0.0015924399485811591, + 0.0015343950362876058, + 0.0017376887844875455, + 0.001394503633491695, + 0.001318436348810792, + 0.005099748261272907, + 0.0041873157024383545, + 0.0034998643677681684, + 0.0033945615869015455, + 0.003241963218897581, + 0.005456964019685984, + 0.00442767608910799, + 0.007027195766568184, + 0.003803798696026206, + 0.004152154084295034, + 0.001716813538223505, + 0.0029742904007434845, + 0.0029347657691687346, + 0.006016199942678213, + 0.0034487973898649216, + 0.002701590768992901, + 0.05237145349383354, + 0.030000580474734306, + 0.02613842487335205, + 0.03283775597810745, + 0.044618282467126846, + 0.0494563914835453, + 0.11549527943134308, + 0.026143554598093033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.006753832101821899 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071652993559837 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016071652993559837, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071652993559837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06299808621406555, + 0.051133058965206146, + 0.06792788952589035, + 0.0435059517621994, + 0.05498482286930084, + 0.15308348834514618, + 0.009707053191959858, + 0.025381222367286682, + 0.02713964134454727, + 0.5041388273239136 + ], + "credit_map": [ + 0.00696803443133831, + 0.005994551349431276, + 0.004793268628418446, + 0.004438536707311869, + 0.0054133920930325985, + 0.010430868715047836, + 0.006700496189296246, + 0.01825892925262451, + 0.0038832162972539663, + 0.010617434978485107, + 0.0038358739111572504, + 0.009014427661895752, + 0.00869210995733738, + 0.004874815233051777, + 0.005196566227823496, + 0.005018610041588545, + 0.00961783342063427, + 0.006935452576726675, + 0.003918815869837999, + 0.014313112944364548, + 0.004493602551519871, + 0.016232891008257866, + 0.00540684862062335, + 0.0070093292742967606, + 0.006517394445836544, + 0.004584937356412411, + 0.0033725877292454243, + 0.005129944067448378, + 0.0035199953708797693, + 0.005845824256539345, + 0.005817735102027655, + 0.00871752854436636, + 0.00799577310681343, + 0.006410409230738878, + 0.005626391619443893, + 0.008426757529377937, + 0.005888775456696749, + 0.008844032883644104, + 0.006572091951966286, + 0.005220587365329266, + 0.008580607362091541, + 0.015892421826720238, + 0.007313116453588009, + 0.05705961212515831, + 0.005858359858393669, + 0.008905568160116673, + 0.03141012042760849, + 0.018063677474856377, + 0.0011208538198843598, + 0.0012482904130592942, + 0.0010630153119564056, + 0.0015118489973247051, + 0.0011068747844547033, + 0.0014516697265207767, + 0.0012856029206886888, + 0.0009188973344862461, + 0.0031250917818397284, + 0.0031684015411883593, + 0.002900246065109968, + 0.0027196023147553205, + 0.0033574437256902456, + 0.0035502545069903135, + 0.0032376085873693228, + 0.0033225712832063437, + 0.0034202688839286566, + 0.0029068083968013525, + 0.0019489438273012638, + 0.003120012581348419, + 0.0036635699216276407, + 0.005016987677663565, + 0.004277255851775408, + 0.0027857909444719553, + 0.07063430547714233, + 0.03970365598797798, + 0.03464653342962265, + 0.04321381822228432, + 0.05908622592687607, + 0.0696607306599617, + 0.15236018598079681, + 0.03483336791396141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.016071652993559837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058674581348896 + ], + "opposite_logprob_tokens": [ + -3.8955867290496826 + ], + "logprob_reward": -0.02058674581348896, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058674581348896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03132794797420502, + 0.052725959569215775, + 0.039650991559028625, + 0.03423233702778816, + 0.027493420988321304, + 0.07458078861236572, + 0.013268506154417992, + 0.07014698535203934, + 0.028502866625785828, + 0.6280702352523804 + ], + "credit_map": [ + 0.004039266612380743, + 0.003546475199982524, + 0.0029898900538682938, + 0.0033025017473846674, + 0.003575032576918602, + 0.004361059051007032, + 0.003215988865122199, + 0.006297728978097439, + 0.005026297643780708, + 0.009273465722799301, + 0.004613467026501894, + 0.007523590233176947, + 0.008075946941971779, + 0.005346565041691065, + 0.005281612277030945, + 0.0075850095599889755, + 0.008976049721240997, + 0.003951016813516617, + 0.003079951275140047, + 0.00713697774335742, + 0.004065170418471098, + 0.0055685886181890965, + 0.004128214903175831, + 0.002745016012340784, + 0.0044987197034060955, + 0.003733279649168253, + 0.003207365283742547, + 0.004913751967251301, + 0.0037731837946921587, + 0.0046449205838143826, + 0.003981188405305147, + 0.005479922518134117, + 0.005541276186704636, + 0.0035562384873628616, + 0.0035650422796607018, + 0.0033775896299630404, + 0.002227283548563719, + 0.002867132890969515, + 0.0031411373056471348, + 0.003217715071514249, + 0.004714184440672398, + 0.007462870795279741, + 0.0038051740266382694, + 0.02711576037108898, + 0.0032137022353708744, + 0.004186565987765789, + 0.015863625332713127, + 0.008218890056014061, + 0.0013209563912823796, + 0.0014125821180641651, + 0.0024788284208625555, + 0.0015306433197110891, + 0.0013840531464666128, + 0.0018188953399658203, + 0.001769963768310845, + 0.0015525815542787313, + 0.00883170310407877, + 0.008090616203844547, + 0.0073095085099339485, + 0.009516208432614803, + 0.009802083484828472, + 0.011355730704963207, + 0.00946903321892023, + 0.005772089120000601, + 0.004158013500273228, + 0.002638711594045162, + 0.002710324712097645, + 0.0034723784774541855, + 0.003137306310236454, + 0.005963327828794718, + 0.004154866095632315, + 0.00226793484762311, + 0.08863773196935654, + 0.049343496561050415, + 0.04365057870745659, + 0.05366690084338188, + 0.07383940368890762, + 0.08570340275764465, + 0.18899989128112793, + 0.044228777289390564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.02058674581348896 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134586811065674 + ], + "opposite_logprob_tokens": [ + -1.3134586811065674 + ], + "logprob_reward": -0.3134586811065674, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134586811065674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018738852813839912, + 0.025000454857945442, + 0.01990078203380108, + 0.015432466752827168, + 0.013953102752566338, + 0.025439705699682236, + 0.012177439406514168, + 0.022000638768076897, + 0.021975791081786156, + 0.8253808617591858 + ], + "credit_map": [ + 0.0016824295744299889, + 0.002171851694583893, + 0.002015460981056094, + 0.0027691612485796213, + 0.0025163921527564526, + 0.002364915329962969, + 0.002670875983312726, + 0.0025477621238678694, + 0.002634455217048526, + 0.00395929766818881, + 0.002324253786355257, + 0.0032607235480099916, + 0.0036682405043393373, + 0.0027271942235529423, + 0.002383623970672488, + 0.004042661748826504, + 0.0032398803159594536, + 0.001718578627333045, + 0.0019222073024138808, + 0.0024628585670143366, + 0.002243026625365019, + 0.003832006361335516, + 0.0024114351253956556, + 0.0020707868970930576, + 0.0016314509557560086, + 0.001564192702062428, + 0.001826775842346251, + 0.0014308952959254384, + 0.0017754948930814862, + 0.002024047775194049, + 0.003008626401424408, + 0.0021709795109927654, + 0.002982382196933031, + 0.0019000188913196325, + 0.0016900255577638745, + 0.0018956613494083285, + 0.0010170477908104658, + 0.0015678595518693328, + 0.0011986111057922244, + 0.0017014933982864022, + 0.0015415506204590201, + 0.0027700376231223345, + 0.0013865085784345865, + 0.008691299706697464, + 0.0011765147792175412, + 0.0019011812983080745, + 0.0046317544765770435, + 0.003340852214023471, + 0.0011911785695701838, + 0.0014243503101170063, + 0.0017332538263872266, + 0.0015894400421530008, + 0.0010837342124432325, + 0.002077873330563307, + 0.0015448201447725296, + 0.0015327862929552794, + 0.0025216536596417427, + 0.0028767320327460766, + 0.0021477348636835814, + 0.0025844851043075323, + 0.003869427600875497, + 0.002518446883186698, + 0.0033140250016003847, + 0.002168128499761224, + 0.001848656334914267, + 0.001780576421879232, + 0.0017134873196482658, + 0.0023507606238126755, + 0.0057518370449543, + 0.003936863504350185, + 0.0027298941276967525, + 0.0018637101165950298, + 0.11574196815490723, + 0.06539363414049149, + 0.05691052973270416, + 0.07106319814920425, + 0.09690959751605988, + 0.11263994127511978, + 0.24960075318813324, + 0.05712111294269562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.3134586811065674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in extreme close-up, focusing on a single, twinkling point of light in the absolute darkness of deep space. A majestic, slow pullback begins, revealing this point of light to be no simple star, but the heart of a cosmic phenomenon. It begins to bloom gracefully, its core intensifying to a brilliant diamond-white before unfurling delicate, petal-like tendrils of sapphire and fiery red-orange clouds of gas, like a time-lapse of a flower blossoming in the sky. As the camera continues its steady, smooth retreat, countless other distant stars in the void begin the same breathtaking process, each a unique cosmic flower, bursting into a brilliant, multi-colored glow. A slow, relentless zoom-out gradually reveals these individual blooming stars as part of a much grander structure, their light merging into the swirling arms of a vast spiral galaxy. The final frame, framed by a breathtaking wide-angle shot, showcases the entire Milky Way in all its splendor—a majestic celestial maelstrom with a dense, fiery golden core, sweeping arms dotted with young blue stars and dark, intricate dust lanes, all set against the silent, star-studded velvet universe, creating an atmosphere of profound awe and tranquil cosmic grandeur.", + "seed": 24250, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000029/group_9/index_2787_seed_24250_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2671284107491374, + "vlm_logprob": -0.2671284107491374, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.2671284107491374, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5769730806350708 + ], + "opposite_logprob_tokens": [ + -0.8269730806350708 + ], + "logprob_reward": -0.5769730806350708, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5769730806350708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14301791787147522, + 0.47369885444641113, + 0.041426628828048706, + 0.17093321681022644, + 0.09132665395736694, + 0.0134591581299901, + 0.020487338304519653, + 0.018745195120573044, + 0.01607436127960682, + 0.010830717161297798 + ], + "credit_map": [ + 0.009109431877732277, + 0.05159608647227287, + 0.010841292329132557, + 0.007309282198548317, + 0.020410113036632538, + 0.0076839132234454155, + 0.024128388613462448, + 0.011939399875700474, + 0.03021732158958912, + 0.13147610425949097, + 0.026626229286193848, + 0.015490817837417126, + 0.028667323291301727, + 0.05382503569126129, + 0.1581907868385315, + 0.029205238446593285, + 0.005458701401948929, + 0.005497507285326719, + 0.006472650915384293, + 0.004811619874089956, + 0.004705379251390696, + 0.004558786284178495, + 0.0038920468650758266, + 0.0060299355536699295, + 0.018462609499692917, + 0.011581100523471832, + 0.02031935378909111, + 0.014679531566798687, + 0.01537313498556614, + 0.01271128561347723, + 0.02827272191643715, + 0.04953347519040108, + 0.00900357123464346, + 0.009586133994162083, + 0.0025808929931372404, + 0.004173837602138519, + 0.005525058135390282, + 0.006314270198345184, + 0.032705482095479965, + 0.02143741212785244, + 0.0011371956206858158, + 0.0017301315674558282, + 0.0009909584186971188, + 0.0017806532559916377, + 0.0022840548772364855, + 0.0012952042743563652, + 0.002131001790985465, + 0.0021099578589200974, + 0.0030607138760387897, + 0.0018111454555764794, + 0.003678280161693692, + 0.0016531043220311403, + 0.001487612258642912, + 0.002389958593994379, + 0.001045961631461978, + 0.0053605614230036736, + 0.001801987993530929, + 0.0014509283937513828, + 0.0031096346210688353, + 0.0013435520231723785, + 0.0016364270122721791, + 0.002789175370708108, + 0.0014330013655126095, + 0.0051804883405566216, + 0.0018616123124957085, + 0.002273996127769351, + 0.0029610544443130493, + 0.0013672130880877376, + 0.001874748500995338, + 0.0019490472041070461, + 0.001578840659931302, + 0.002207848010584712, + 0.0017203395254909992, + 0.0011454006889835, + 0.0015602591447532177, + 0.0012783391866832972, + 0.0011906606378033757, + 0.0012200535275042057, + 0.001307534403167665, + 0.0014081299304962158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.5769730806350708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093958280980587 + ], + "opposite_logprob_tokens": [ + -4.511094093322754 + ], + "logprob_reward": -0.011093958280980587, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093958280980587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17466266453266144, + 0.45657452940940857, + 0.042365219444036484, + 0.10587547719478607, + 0.04874025657773018, + 0.03795767202973366, + 0.025969866663217545, + 0.05298938974738121, + 0.03268949314951897, + 0.022175423800945282 + ], + "credit_map": [ + 0.012482915073633194, + 0.06666819006204605, + 0.007941314950585365, + 0.007689653895795345, + 0.028539836406707764, + 0.00719838310033083, + 0.028852291405200958, + 0.015290088951587677, + 0.0272777508944273, + 0.1689181625843048, + 0.01843218132853508, + 0.012310635298490524, + 0.024741359055042267, + 0.022237593308091164, + 0.15704967081546783, + 0.025607215240597725, + 0.00377261801622808, + 0.00708004180341959, + 0.004897213540971279, + 0.004729472100734711, + 0.005671413615345955, + 0.004657462239265442, + 0.005039401352405548, + 0.00651759747415781, + 0.010976525954902172, + 0.010086622089147568, + 0.011417650617659092, + 0.008307857438921928, + 0.008503743447363377, + 0.007680836599320173, + 0.018956376239657402, + 0.029945868998765945, + 0.005179857835173607, + 0.004735003691166639, + 0.007924236357212067, + 0.0044787065126001835, + 0.004787653684616089, + 0.004367728717625141, + 0.009960041381418705, + 0.007307027466595173, + 0.0036714677698910236, + 0.0062861936166882515, + 0.0053876363672316074, + 0.0036506750620901585, + 0.005179557017982006, + 0.004913918673992157, + 0.002376934979110956, + 0.0064912899397313595, + 0.0043386248871684074, + 0.002903474261984229, + 0.004574814811348915, + 0.0022951809223741293, + 0.001846485654823482, + 0.0029747537337243557, + 0.0016625355929136276, + 0.005373996216803789, + 0.004361926112323999, + 0.006842195987701416, + 0.009438215754926205, + 0.005702654365450144, + 0.006494867615401745, + 0.008426070213317871, + 0.0058197202160954475, + 0.005903739016503096, + 0.0020689729135483503, + 0.0055722384713590145, + 0.004932287614792585, + 0.0041933441534638405, + 0.004142113961279392, + 0.004022631794214249, + 0.003584523918107152, + 0.00417338265106082, + 0.0020687927026301622, + 0.0024019686970859766, + 0.003661961294710636, + 0.0030692184809595346, + 0.0028096388559788465, + 0.002648809226229787, + 0.0024671286810189486, + 0.003047906095162034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.011093958280980587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810613602399826 + ], + "opposite_logprob_tokens": [ + -3.2881062030792236 + ], + "logprob_reward": -0.03810613602399826, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810613602399826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09561942517757416, + 0.23173516988754272, + 0.04777471348643303, + 0.05507120117545128, + 0.36381176114082336, + 0.0904756635427475, + 0.029367271810770035, + 0.016019947826862335, + 0.03199230507016182, + 0.038132574409246445 + ], + "credit_map": [ + 0.007456080988049507, + 0.02962687984108925, + 0.0055039129219949245, + 0.006361585110425949, + 0.0170121006667614, + 0.004727769177407026, + 0.01170669961720705, + 0.013224391266703606, + 0.014089028351008892, + 0.09450801461935043, + 0.007295634597539902, + 0.005507546942681074, + 0.013598151504993439, + 0.0064059155993163586, + 0.07752680033445358, + 0.012804091908037663, + 0.005370329599827528, + 0.0044344337657094, + 0.005766089539974928, + 0.005217116326093674, + 0.006322985980659723, + 0.006879895925521851, + 0.006129227578639984, + 0.007654638960957527, + 0.0058587696403265, + 0.004429034888744354, + 0.005298675503581762, + 0.005971039645373821, + 0.004336748272180557, + 0.0056397272273898125, + 0.00950919184833765, + 0.014028015546500683, + 0.03649967908859253, + 0.03667786717414856, + 0.011913420632481575, + 0.01725805178284645, + 0.020891381427645683, + 0.023466017097234726, + 0.13122619688510895, + 0.0858791396021843, + 0.007663734257221222, + 0.01492156833410263, + 0.010799342766404152, + 0.007166503928601742, + 0.009446815587580204, + 0.016443219035863876, + 0.009958420880138874, + 0.014076066203415394, + 0.0036230599507689476, + 0.004370194394141436, + 0.003907146397978067, + 0.003290932159870863, + 0.004137655254453421, + 0.0040492345578968525, + 0.002404293976724148, + 0.0035847551189363003, + 0.001437100931070745, + 0.0017559491097927094, + 0.002952312119305134, + 0.0015822902787476778, + 0.0019166322890669107, + 0.002474067732691765, + 0.0017252948600798845, + 0.002176299225538969, + 0.004147518426179886, + 0.004861363675445318, + 0.005540930666029453, + 0.0029542900156229734, + 0.0031057046726346016, + 0.0035288387443870306, + 0.0029164114966988564, + 0.004937247838824987, + 0.005275311414152384, + 0.00337799871340394, + 0.005617621820420027, + 0.0038685104809701443, + 0.004709335044026375, + 0.004571691155433655, + 0.0046862089075148106, + 0.0060258968733251095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.03810613602399826 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071535646915436 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071535646915436, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071535646915436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09150994569063187, + 0.11393822729587555, + 0.037491075694561005, + 0.02530740201473236, + 0.3954276740550995, + 0.1120651587843895, + 0.02771885134279728, + 0.06420710682868958, + 0.06821604073047638, + 0.06411856412887573 + ], + "credit_map": [ + 0.0073389033786952496, + 0.030620187520980835, + 0.006324399262666702, + 0.0053592282347381115, + 0.015395103953778744, + 0.004856864456087351, + 0.012294959276914597, + 0.009320286102592945, + 0.007521377876400948, + 0.041046131402254105, + 0.005022282246500254, + 0.003908923361450434, + 0.006621404550969601, + 0.006715067662298679, + 0.03647248074412346, + 0.006630547810345888, + 0.0036904814187437296, + 0.005600372329354286, + 0.006785634905099869, + 0.005240221973508596, + 0.004530714359134436, + 0.004486855585128069, + 0.0031950301490724087, + 0.0039617642760276794, + 0.0031409922521561384, + 0.0035646373871713877, + 0.0024583309423178434, + 0.0029792063869535923, + 0.003840611781924963, + 0.0033366696443408728, + 0.0027628540992736816, + 0.0032240990549325943, + 0.039312899112701416, + 0.03871425613760948, + 0.01002999022603035, + 0.017333555966615677, + 0.022273723036050797, + 0.031464580446481705, + 0.13766181468963623, + 0.09863683581352234, + 0.007980392314493656, + 0.018570050597190857, + 0.014091020449995995, + 0.009341311641037464, + 0.014775346964597702, + 0.01726364716887474, + 0.012169530615210533, + 0.017873845994472504, + 0.004208628088235855, + 0.00372976902872324, + 0.00289520132355392, + 0.0032139879185706377, + 0.0026557703968137503, + 0.0035072709433734417, + 0.001670459983870387, + 0.005837760400027037, + 0.004259750712662935, + 0.003905672812834382, + 0.009120060130953789, + 0.002606504363939166, + 0.002510626567527652, + 0.01455731876194477, + 0.0031364872120320797, + 0.02411067672073841, + 0.004339322913438082, + 0.006218294147402048, + 0.012063453905284405, + 0.009364580735564232, + 0.009622648358345032, + 0.011559469625353813, + 0.006266603246331215, + 0.008781660348176956, + 0.008572599850594997, + 0.0070761204697191715, + 0.009814959950745106, + 0.008302586153149605, + 0.007119758520275354, + 0.006161538418382406, + 0.008276717737317085, + 0.008794277906417847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.016071535646915436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single twinkling point of light in deep space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the central point of light bloom into a diamond-white core surrounded by sapphire and red-orange gas tendrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth pullback and zoom-out to reveal multiple blooming stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition to show the swirling arms of a vast spiral galaxy with a dense golden core and blue stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final frame present a wide-angle shot of the entire Milky Way against a star-studded background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933973431587219 + ], + "opposite_logprob_tokens": [ + -0.6933973431587219 + ], + "logprob_reward": -0.6933973431587219, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933973431587219, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17387406527996063, + 0.1407739520072937, + 0.039883095771074295, + 0.03165551275014877, + 0.14362671971321106, + 0.12821760773658752, + 0.02492746151983738, + 0.10899766534566879, + 0.039959363639354706, + 0.16808463633060455 + ], + "credit_map": [ + 0.010506687685847282, + 0.06681521236896515, + 0.007569748442620039, + 0.008535551838576794, + 0.025757020339369774, + 0.007731067016720772, + 0.03258538991212845, + 0.01437339000403881, + 0.008249797858297825, + 0.053901005536317825, + 0.005885688122361898, + 0.004490652587264776, + 0.007974076084792614, + 0.004490249790251255, + 0.047443050891160965, + 0.00833942648023367, + 0.0028077971655875444, + 0.011201969347894192, + 0.0051499283872544765, + 0.0037887233775109053, + 0.0037024947814643383, + 0.002449547406286001, + 0.00569640938192606, + 0.00508622732013464, + 0.0043868706561625, + 0.0031859225127846003, + 0.004756313282996416, + 0.004447066690772772, + 0.0031123466324061155, + 0.0040870183147490025, + 0.0033548767678439617, + 0.004325101617723703, + 0.01592208445072174, + 0.016338903456926346, + 0.009261716157197952, + 0.009420004673302174, + 0.011759322136640549, + 0.014347166754305363, + 0.041632525622844696, + 0.024944990873336792, + 0.012604901567101479, + 0.02071409858763218, + 0.017728855833411217, + 0.013070894405245781, + 0.021202687174081802, + 0.01584864780306816, + 0.012794343754649162, + 0.014253185130655766, + 0.002917346078902483, + 0.0031545846723020077, + 0.0032000455539673567, + 0.004204849246889353, + 0.0031564137898385525, + 0.0027574794366955757, + 0.002489516045898199, + 0.003047226695343852, + 0.010295867919921875, + 0.012237722985446453, + 0.017685655504465103, + 0.009141690097749233, + 0.011094600893557072, + 0.016287358477711678, + 0.00909566693007946, + 0.023159101605415344, + 0.003982430323958397, + 0.003564335172995925, + 0.00615565525367856, + 0.00491675641387701, + 0.005956267938017845, + 0.007107917219400406, + 0.003916318994015455, + 0.004359681624919176, + 0.021806485950946808, + 0.01577071100473404, + 0.024594875052571297, + 0.019932467490434647, + 0.01857207529246807, + 0.02511991746723652, + 0.018656902015209198, + 0.023631198331713676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.6933973431587219 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_tensor([3397, 3397])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_tensor([3397, 3397])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..55734dba04373439c0cf9ed0beed883febb0a908 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_tensor([3397, 3397])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tightly framed underwater close-up, rendered with the breathtaking detail of high-speed nature photography, tracks sideways in stunning slow motion, capturing the primal moment of predation. A massive, torpedo-shaped longnose gar, its body covered in a cacophony of interwoven, diamond-shaped black scales shimmering with iridescent hues of olive green and bronze, dashes at explosive speed. Sunlight filters through the water, creating shifting caustic patterns across its form. Its long, bony snout flares slightly, revealing rows of terrifyingly sharp, translucent, needle-like teeth. Its target, a much smaller, silver-scaled blacknose fish, gleams in the dim light and desperately lunges sideways, its fins flared and its body contorted in sheer panic. The powerful, whip-like propulsion from the gar's tail violently churns the water, creating visible eddies and trails of tiny, shimmering bubbles and disturbed sediment that swirl around the two fish, all frozen in the surreal clarity of slow motion, creating the tense and dramatic atmosphere of a life-or-death chase.", + "seed": 17916, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_3397_seed_17916_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026686896570026875, + "vlm_logprob": -0.026686896570026875, + "vlm_yes_no_margin": 4.374999952316284, + "vlm_token_credit": -0.026686896570026875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019252223894000053 + ], + "opposite_logprob_tokens": [ + -5.769252300262451 + ], + "logprob_reward": -0.019252223894000053, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.019252223894000053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19724085927009583, + 0.1442405879497528, + 0.12424580007791519, + 0.0817345380783081, + 0.04147867113351822, + 0.06760573387145996, + 0.07184667885303497, + 0.08127369731664658, + 0.08641215413808823, + 0.1039213016629219 + ], + "credit_map": [ + 0.016792792826890945, + 0.0190852303057909, + 0.030329512432217598, + 0.01679626852273941, + 0.011481727473437786, + 0.02522399090230465, + 0.026158520951867104, + 0.051372818648815155, + 0.021095292642712593, + 0.023060977458953857, + 0.02582482621073723, + 0.01287579070776701, + 0.018732501193881035, + 0.014737922698259354, + 0.01362865511327982, + 0.014284629374742508, + 0.02441536635160446, + 0.01856364868581295, + 0.012554577551782131, + 0.012360158376395702, + 0.019075803458690643, + 0.008971021510660648, + 0.015437770634889603, + 0.012867460027337074, + 0.012730217538774014, + 0.010132365860044956, + 0.01362134050577879, + 0.008454700000584126, + 0.009154276922345161, + 0.005677084904164076, + 0.01219248492270708, + 0.009772066958248615, + 0.008368133567273617, + 0.009549019858241081, + 0.003959657158702612, + 0.0033227389212697744, + 0.003917544614523649, + 0.003985730465501547, + 0.003626610152423382, + 0.004749235697090626, + 0.006537881214171648, + 0.010392100550234318, + 0.0070631252601742744, + 0.009433847852051258, + 0.008830663748085499, + 0.009121815674006939, + 0.009445500560104847, + 0.0067807999439537525, + 0.006345493718981743, + 0.011457758024334908, + 0.007092813029885292, + 0.006226390600204468, + 0.017739061266183853, + 0.011133470572531223, + 0.0059605371206998825, + 0.005891157314181328, + 0.0076886010356247425, + 0.00721984077244997, + 0.0050795855931937695, + 0.01797233521938324, + 0.012441360391676426, + 0.006952692288905382, + 0.006394444033503532, + 0.01752484403550625, + 0.005356592126190662, + 0.006947844289243221, + 0.01326791848987341, + 0.020679842680692673, + 0.01154114119708538, + 0.006458369083702564, + 0.012898992747068405, + 0.009261446073651314, + 0.009333259426057339, + 0.007559212855994701, + 0.017072618007659912, + 0.013672257773578167, + 0.010794795118272305, + 0.01218065433204174, + 0.015980804339051247, + 0.01732770726084709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.019252223894000053 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963390693068504 + ], + "opposite_logprob_tokens": [ + -5.130963325500488 + ], + "logprob_reward": -0.005963390693068504, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963390693068504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.576697826385498, + 0.017905764281749725, + 0.03330162540078163, + 0.10038607567548752, + 0.04573756456375122, + 0.05286286026239395, + 0.06190234422683716, + 0.035819146782159805, + 0.03046019934117794, + 0.044926583766937256 + ], + "credit_map": [ + 0.04325004667043686, + 0.02158139832317829, + 0.1410047560930252, + 0.03586680069565773, + 0.022186677902936935, + 0.08641249686479568, + 0.032065317034721375, + 0.1943303346633911, + 0.0024744935799390078, + 0.0030084492173045874, + 0.002935666823759675, + 0.0015130449319258332, + 0.0021824794821441174, + 0.0019036823650822043, + 0.00219188560731709, + 0.001696061808615923, + 0.0070941271260380745, + 0.005493802484124899, + 0.002688755514100194, + 0.0028880268801003695, + 0.005779029801487923, + 0.0025942078791558743, + 0.0036847966257482767, + 0.0030788788571953773, + 0.01572774164378643, + 0.01419705431908369, + 0.011318483389914036, + 0.011421840637922287, + 0.011315848678350449, + 0.011024597100913525, + 0.012842950411140919, + 0.012537555769085884, + 0.0063857766799628735, + 0.008354539051651955, + 0.004741190001368523, + 0.0043372404761612415, + 0.006585036404430866, + 0.0040415553376078606, + 0.005962766706943512, + 0.005329462233930826, + 0.005959338042885065, + 0.012111885473132133, + 0.004499366972595453, + 0.005930387414991856, + 0.007140131667256355, + 0.0056939395144581795, + 0.004909287206828594, + 0.006618525832891464, + 0.0059102317318320274, + 0.00927791278809309, + 0.007591772824525833, + 0.006754768546670675, + 0.011716269887983799, + 0.008249972946941853, + 0.006590140052139759, + 0.005811270326375961, + 0.002495687687769532, + 0.0024071179796010256, + 0.0033400976099073887, + 0.010336305946111679, + 0.003452085191383958, + 0.0028709983453154564, + 0.0035937768407166004, + 0.007323075085878372, + 0.0014306282391771674, + 0.0015216084429994226, + 0.010182933881878853, + 0.0058703492395579815, + 0.005041069351136684, + 0.001736249541863799, + 0.0032647212501615286, + 0.001412637415342033, + 0.004638812504708767, + 0.0029392968863248825, + 0.008988399989902973, + 0.008061978965997696, + 0.003601296339184046, + 0.0069647617638111115, + 0.005141102243214846, + 0.004590933211147785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.005963390693068504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188709393143654 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.018188709393143654, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018188709393143654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.622950553894043, + 0.013409595936536789, + 0.013451249338686466, + 0.06364959478378296, + 0.057510338723659515, + 0.04552055895328522, + 0.016560375690460205, + 0.07347631454467773, + 0.07411535829305649, + 0.01935598812997341 + ], + "credit_map": [ + 0.04425464943051338, + 0.021489599719643593, + 0.1512124091386795, + 0.038064565509557724, + 0.02539934776723385, + 0.09411358833312988, + 0.035296980291604996, + 0.21311941742897034, + 0.0023695838171988726, + 0.001787110697478056, + 0.001793030765838921, + 0.0013522265944629908, + 0.0020834668539464474, + 0.0017759623005986214, + 0.0010465458035469055, + 0.001201669336296618, + 0.003237636759877205, + 0.0021079606376588345, + 0.0012775275390595198, + 0.0012781237019225955, + 0.001961785601451993, + 0.001154736033640802, + 0.0010939172934740782, + 0.0013395611895248294, + 0.009607484564185143, + 0.008744247257709503, + 0.006517563946545124, + 0.005887089762836695, + 0.0076790498569607735, + 0.007160440552979708, + 0.009753052145242691, + 0.008300662972033024, + 0.00785378273576498, + 0.00754793593659997, + 0.008175091817975044, + 0.00668359873816371, + 0.008126257918775082, + 0.00601993128657341, + 0.006119430996477604, + 0.006984309293329716, + 0.004448567517101765, + 0.007430907804518938, + 0.0036646544467657804, + 0.005627140868455172, + 0.0054459404200315475, + 0.0062684533186256886, + 0.006732052192091942, + 0.0059028444811701775, + 0.0017082428093999624, + 0.0020827152766287327, + 0.002110590459778905, + 0.0017947740852832794, + 0.003120851470157504, + 0.002573988400399685, + 0.0013723878655582666, + 0.0017968249740079045, + 0.005125540308654308, + 0.005897753871977329, + 0.007449820172041655, + 0.014639899134635925, + 0.01920250430703163, + 0.007488412782549858, + 0.007259966805577278, + 0.006412409711629152, + 0.0038949402514845133, + 0.004953055176883936, + 0.023524560034275055, + 0.014094831421971321, + 0.008355448953807354, + 0.004149886779487133, + 0.009247394278645515, + 0.005895235110074282, + 0.0019336743280291557, + 0.0014593317173421383, + 0.003133666468784213, + 0.0027214332949370146, + 0.0020055940840393305, + 0.0026129879988729954, + 0.00303600518964231, + 0.002453295048326254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.018188709393143654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895860075950623 + ], + "opposite_logprob_tokens": [ + -2.578958511352539 + ], + "logprob_reward": -0.07895860075950623, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895860075950623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6683293581008911, + 0.020997317507863045, + 0.015031595714390278, + 0.034800656139850616, + 0.027891941368579865, + 0.017277825623750687, + 0.017960254102945328, + 0.028056271374225616, + 0.14981864392757416, + 0.019836105406284332 + ], + "credit_map": [ + 0.05163634568452835, + 0.024640167132019997, + 0.16313882172107697, + 0.04134805127978325, + 0.026441266760230064, + 0.09924894571304321, + 0.03882317245006561, + 0.22305262088775635, + 0.002481838222593069, + 0.0032070926390588284, + 0.002986358478665352, + 0.002216585213318467, + 0.0034825291950255632, + 0.002459093229845166, + 0.002087714383378625, + 0.002076106145977974, + 0.0030895716045051813, + 0.0023175906389951706, + 0.0014023911207914352, + 0.0012360168620944023, + 0.002723956247791648, + 0.0013462821952998638, + 0.0015851653879508376, + 0.0013306207256391644, + 0.006801775190979242, + 0.006346807349473238, + 0.003712719539180398, + 0.003263607155531645, + 0.0034037998411804438, + 0.0024390488397330046, + 0.005132332909852266, + 0.0037005634512752295, + 0.003789712907746434, + 0.004818015266209841, + 0.0028519188053905964, + 0.0031517050229012966, + 0.0033230239059776068, + 0.002572051016613841, + 0.004041478969156742, + 0.0033440347760915756, + 0.0026660196017473936, + 0.003036331385374069, + 0.00153672497253865, + 0.0020300489850342274, + 0.001918564084917307, + 0.0018083372851833701, + 0.0017514313803985715, + 0.0025303680449724197, + 0.001868365565314889, + 0.0028976069297641516, + 0.0021226461976766586, + 0.001941633177921176, + 0.0033480620477348566, + 0.00235934485681355, + 0.0018892131047323346, + 0.0015333821065723896, + 0.002228592289611697, + 0.0027836784720420837, + 0.0023585972376167774, + 0.004249338526278734, + 0.00941932201385498, + 0.0028409049846231937, + 0.0026109141763299704, + 0.0015649209963157773, + 0.006950343493372202, + 0.010428083129227161, + 0.0466449037194252, + 0.025795113295316696, + 0.024982081726193428, + 0.008990483358502388, + 0.016541173681616783, + 0.009486454539000988, + 0.002457418479025364, + 0.001517943455837667, + 0.0037587250117212534, + 0.003779029008001089, + 0.0016083846567198634, + 0.003003779100254178, + 0.0020512857008725405, + 0.0016595389461144805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.07895860075950623 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071558110415936 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071558110415936, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071558110415936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5308932065963745, + 0.08584654331207275, + 0.025756068527698517, + 0.05359925329685211, + 0.046074576675891876, + 0.022635145112872124, + 0.04440326243638992, + 0.04320630803704262, + 0.11655907332897186, + 0.03102652169764042 + ], + "credit_map": [ + 0.04126294329762459, + 0.021331774070858955, + 0.1316075623035431, + 0.03495628759264946, + 0.022014832124114037, + 0.07586562633514404, + 0.033809661865234375, + 0.17004454135894775, + 0.01001260057091713, + 0.012029427103698254, + 0.014804170466959476, + 0.00988634955137968, + 0.009500443935394287, + 0.009950147941708565, + 0.011549719609320164, + 0.008113689720630646, + 0.0035124998539686203, + 0.00343299494124949, + 0.003218042431399226, + 0.0034665248822420835, + 0.003116017207503319, + 0.002144449856132269, + 0.004082970321178436, + 0.0027825688011944294, + 0.007534622214734554, + 0.007383387535810471, + 0.0075064003467559814, + 0.00449161184951663, + 0.010010967962443829, + 0.004182339645922184, + 0.0062529342249035835, + 0.006236989516764879, + 0.005420989356935024, + 0.00468223774805665, + 0.005764014553278685, + 0.005407713353633881, + 0.005848957225680351, + 0.004372512921690941, + 0.007981049828231335, + 0.0065970998257398605, + 0.0028870445676147938, + 0.0055114515125751495, + 0.0019929693080484867, + 0.0021051010116934776, + 0.0029383900109678507, + 0.002446581609547138, + 0.0024807367008179426, + 0.002272870158776641, + 0.004825605545192957, + 0.003964276518672705, + 0.005283322185277939, + 0.005096942186355591, + 0.0076275053434073925, + 0.0056843869388103485, + 0.0056717488914728165, + 0.006249474361538887, + 0.003126394934952259, + 0.004980025812983513, + 0.006108474917709827, + 0.0061798496171832085, + 0.005484543740749359, + 0.004999936558306217, + 0.005259349942207336, + 0.007067730650305748, + 0.005304234102368355, + 0.008762286975979805, + 0.0321650467813015, + 0.018147185444831848, + 0.02661687321960926, + 0.007971001788973808, + 0.011655459180474281, + 0.005936981178820133, + 0.003088275669142604, + 0.002606234746053815, + 0.005829130299389362, + 0.0062928576953709126, + 0.002645112806931138, + 0.004206447396427393, + 0.003364868927747011, + 0.002993592992424965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.011071558110415936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tightly framed underwater close-up, rendered with the breathtaking detail of high-speed nature photography, tracks sideways in stunning slow motion, capturing the primal moment of predation. A massive, torpedo-shaped longnose gar, its body covered in a cacophony of interwoven, diamond-shaped black scales shimmering with iridescent hues of olive green and bronze, dashes at explosive speed. Sunlight filters through the water, creating shifting caustic patterns across its form. Its long, bony snout flares slightly, revealing rows of terrifyingly sharp, translucent, needle-like teeth. Its target, a much smaller, silver-scaled blacknose fish, gleams in the dim light and desperately lunges sideways, its fins flared and its body contorted in sheer panic. The powerful, whip-like propulsion from the gar's tail violently churns the water, creating visible eddies and trails of tiny, shimmering bubbles and disturbed sediment that swirl around the two fish, all frozen in the surreal clarity of slow motion, creating the tense and dramatic atmosphere of a life-or-death chase.", + "seed": 17917, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_3397_seed_17917_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020299911499023438, + "vlm_logprob": -0.020299911499023438, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.020299911499023438, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02611127868294716 + ], + "opposite_logprob_tokens": [ + -5.901111125946045 + ], + "logprob_reward": -0.02611127868294716, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.02611127868294716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10377997159957886, + 0.06860367953777313, + 0.05458161607384682, + 0.15628565847873688, + 0.05556754022836685, + 0.05691364035010338, + 0.17074555158615112, + 0.2194293886423111, + 0.039123356342315674, + 0.07496961951255798 + ], + "credit_map": [ + 0.012148804031312466, + 0.006654021795839071, + 0.007984395138919353, + 0.006578861735761166, + 0.00833935383707285, + 0.017347577959299088, + 0.009011366404592991, + 0.03571559488773346, + 0.006803249008953571, + 0.006335005629807711, + 0.010567491874098778, + 0.00819051917642355, + 0.008255666121840477, + 0.0151780741289258, + 0.005644258577376604, + 0.007629410829395056, + 0.005304344929754734, + 0.008230709470808506, + 0.008957495912909508, + 0.0065374248661100864, + 0.007269205991178751, + 0.005416425410658121, + 0.00696953758597374, + 0.005896473303437233, + 0.03865325078368187, + 0.020555060356855392, + 0.017034010961651802, + 0.013074408285319805, + 0.008433700539171696, + 0.011211828328669071, + 0.014401325024664402, + 0.03292205557227135, + 0.004079035948961973, + 0.005075525492429733, + 0.004113067407160997, + 0.006840093061327934, + 0.005094348452985287, + 0.008489083498716354, + 0.012739930301904678, + 0.00913645327091217, + 0.013008736073970795, + 0.008717925287783146, + 0.006324445828795433, + 0.004916342906653881, + 0.004646250046789646, + 0.005071969237178564, + 0.006118521094322205, + 0.008109452202916145, + 0.010234605520963669, + 0.01691138558089733, + 0.012624475173652172, + 0.027854451909661293, + 0.02896326594054699, + 0.028647977858781815, + 0.022204706445336342, + 0.023304693400859833, + 0.01478820014744997, + 0.03355599567294121, + 0.013242113403975964, + 0.01660790666937828, + 0.04704999178647995, + 0.03438669443130493, + 0.02251804992556572, + 0.03728042170405388, + 0.0034180099610239267, + 0.006353715434670448, + 0.004283369053155184, + 0.004775972571223974, + 0.006563141942024231, + 0.004341342486441135, + 0.0035295640118420124, + 0.0058582439087331295, + 0.008584198541939259, + 0.010162413120269775, + 0.016751138493418694, + 0.01187298633158207, + 0.007551521994173527, + 0.006701246369630098, + 0.0072330753318965435, + 0.006113043520599604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.02611127868294716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041136653162539005 + ], + "opposite_logprob_tokens": [ + -5.504113674163818 + ], + "logprob_reward": -0.0041136653162539005, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041136653162539005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15885019302368164, + 0.02831486612558365, + 0.015376672148704529, + 0.5079318881034851, + 0.016981201246380806, + 0.11243069171905518, + 0.05971807241439819, + 0.0339302197098732, + 0.03704375401139259, + 0.029422439634799957 + ], + "credit_map": [ + 0.01920188218355179, + 0.012575648725032806, + 0.013983757235109806, + 0.00994215626269579, + 0.011200697161257267, + 0.03351927921175957, + 0.010787702165544033, + 0.04763907194137573, + 0.0028084968216717243, + 0.002440120792016387, + 0.004133556038141251, + 0.004543014336377382, + 0.003155553014948964, + 0.004171071574091911, + 0.003780786879360676, + 0.0032822689972817898, + 0.0012787216110154986, + 0.0022377397399395704, + 0.001726269838400185, + 0.001560693490318954, + 0.0028618467040359974, + 0.0020179369021207094, + 0.0018236726755276322, + 0.0018697917694225907, + 0.07490631192922592, + 0.09200787544250488, + 0.07317104190587997, + 0.07276206463575363, + 0.04649542644619942, + 0.05094312131404877, + 0.04268868640065193, + 0.05495737865567207, + 0.001445238129235804, + 0.0026364754885435104, + 0.0013262942666187882, + 0.0020465385168790817, + 0.0016463629435747862, + 0.003455553902313113, + 0.0023428997956216335, + 0.0020818361081182957, + 0.01401438843458891, + 0.01812831498682499, + 0.008217079564929008, + 0.01120168250054121, + 0.010919095017015934, + 0.024128567427396774, + 0.0088652940467, + 0.0169562678784132, + 0.003184310393407941, + 0.006079872604459524, + 0.005797256715595722, + 0.009230101481080055, + 0.006614045239984989, + 0.0116504468023777, + 0.009669346734881401, + 0.007492691744118929, + 0.002613300457596779, + 0.005379013251513243, + 0.002161437878385186, + 0.0024932732339948416, + 0.004737114068120718, + 0.005874198395758867, + 0.004464198835194111, + 0.006207681261003017, + 0.003385004820302129, + 0.002999424235895276, + 0.0043006218038499355, + 0.006645104382187128, + 0.004178603179752827, + 0.005442123394459486, + 0.004333908203989267, + 0.005758963990956545, + 0.0039194198325276375, + 0.004164841957390308, + 0.005099475383758545, + 0.003659129375591874, + 0.0033420687541365623, + 0.0029764766804873943, + 0.0026206718757748604, + 0.0036403576377779245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0041136653162539005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059270069003105 + ], + "opposite_logprob_tokens": [ + -3.89559268951416 + ], + "logprob_reward": -0.02059270069003105, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059270069003105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22413435578346252, + 0.04534824192523956, + 0.04405275732278824, + 0.15364131331443787, + 0.02772671915590763, + 0.12994857132434845, + 0.05337473377585411, + 0.09806716442108154, + 0.05866988003253937, + 0.16503626108169556 + ], + "credit_map": [ + 0.023658351972699165, + 0.015643559396266937, + 0.019190026447176933, + 0.019063929095864296, + 0.017430074512958527, + 0.03422999754548073, + 0.023414358496665955, + 0.07150405645370483, + 0.004091094713658094, + 0.004755931440740824, + 0.007013252004981041, + 0.006107138469815254, + 0.005391378421336412, + 0.009234019555151463, + 0.005121354013681412, + 0.003634072607383132, + 0.00499014463275671, + 0.0052542900666594505, + 0.013590890914201736, + 0.005160095635801554, + 0.004015884827822447, + 0.0029874895699322224, + 0.003920525312423706, + 0.004133439157158136, + 0.031587425619363785, + 0.02462334930896759, + 0.016155319288372993, + 0.01727735437452793, + 0.009603003039956093, + 0.01169261708855629, + 0.013939741998910904, + 0.02876250445842743, + 0.0029178992845118046, + 0.003954085521399975, + 0.0028437005821615458, + 0.004232217092067003, + 0.0029192494694143534, + 0.0036913007497787476, + 0.0037147621624171734, + 0.0034535040613263845, + 0.03346940129995346, + 0.021696504205465317, + 0.016506701707839966, + 0.011120120994746685, + 0.008477938361465931, + 0.013291485607624054, + 0.008662670850753784, + 0.016723738983273506, + 0.0031634815968573093, + 0.005947148893028498, + 0.003423545742407441, + 0.009075651876628399, + 0.00749625451862812, + 0.010391691699624062, + 0.007131483871489763, + 0.006745481863617897, + 0.006877960171550512, + 0.01646396331489086, + 0.0061577195301651955, + 0.005800910759717226, + 0.020574215799570084, + 0.014898317866027355, + 0.010549615137279034, + 0.01674446277320385, + 0.006047769449651241, + 0.00846227165311575, + 0.0057817441411316395, + 0.00973078515380621, + 0.007422320079058409, + 0.007739515975117683, + 0.005351117346435785, + 0.008134355768561363, + 0.02246912568807602, + 0.023481162264943123, + 0.026241665706038475, + 0.02475566416978836, + 0.021045388653874397, + 0.015746086835861206, + 0.013567088171839714, + 0.01773008331656456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02059270069003105 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305095225572586 + ], + "opposite_logprob_tokens": [ + -3.16805100440979 + ], + "logprob_reward": -0.04305095225572586, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305095225572586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12978807091712952, + 0.15729233622550964, + 0.05585138499736786, + 0.10856745392084122, + 0.08506348729133606, + 0.044350214302539825, + 0.11389733105897903, + 0.06021377071738243, + 0.13083888590335846, + 0.11413707584142685 + ], + "credit_map": [ + 0.016965733841061592, + 0.012651508674025536, + 0.012322753667831421, + 0.01027331780642271, + 0.01054725144058466, + 0.018534913659095764, + 0.010495014488697052, + 0.03799756243824959, + 0.015057823620736599, + 0.017193803563714027, + 0.023093195632100105, + 0.0216242503374815, + 0.018702881410717964, + 0.03450546786189079, + 0.015542653389275074, + 0.011572257615625858, + 0.005324845667928457, + 0.0070740398950874805, + 0.01357644610106945, + 0.006561463698744774, + 0.006052175536751747, + 0.004891844466328621, + 0.004737589042633772, + 0.007632979657500982, + 0.021349430084228516, + 0.015927422791719437, + 0.012150580994784832, + 0.010250560939311981, + 0.009367857128381729, + 0.011284839361906052, + 0.010732813738286495, + 0.017503945156931877, + 0.007151513360440731, + 0.012046808376908302, + 0.007947491481900215, + 0.011088227853178978, + 0.006241001188755035, + 0.01982850208878517, + 0.009355834685266018, + 0.011404098942875862, + 0.0044421241618692875, + 0.005878867581486702, + 0.004822201561182737, + 0.005900980904698372, + 0.004398076795041561, + 0.005997123196721077, + 0.0061348117887973785, + 0.006776026450097561, + 0.008601074106991291, + 0.008503568358719349, + 0.010268553160130978, + 0.017309358343482018, + 0.020252708345651627, + 0.0185053963214159, + 0.015456081368029118, + 0.01500059012323618, + 0.004167064093053341, + 0.010370178148150444, + 0.0054547605104744434, + 0.006764038000255823, + 0.009432198479771614, + 0.008957989513874054, + 0.00567855779081583, + 0.009388981387019157, + 0.012601864524185658, + 0.015311525203287601, + 0.016915343701839447, + 0.019149037078022957, + 0.0195505078881979, + 0.015396551229059696, + 0.011626879684627056, + 0.020287171006202698, + 0.01226278580725193, + 0.013626618310809135, + 0.011889183893799782, + 0.013367537409067154, + 0.015453065745532513, + 0.016408223658800125, + 0.01754937693476677, + 0.013580281287431717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.04305095225572586 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1481924057006836, + 0.14100538194179535, + 0.11273819953203201, + 0.07947660237550735, + 0.09959492832422256, + 0.0796278640627861, + 0.06457667052745819, + 0.05577553063631058, + 0.06706932932138443, + 0.1519431322813034 + ], + "credit_map": [ + 0.017282644286751747, + 0.015782032161951065, + 0.020243223756551743, + 0.012964312918484211, + 0.016070952638983727, + 0.018402041867375374, + 0.015590068884193897, + 0.031857118010520935, + 0.014715985395014286, + 0.009961810894310474, + 0.02277989499270916, + 0.019711708649992943, + 0.015880363062024117, + 0.02342655509710312, + 0.02119893953204155, + 0.013330112211406231, + 0.015749245882034302, + 0.01116091012954712, + 0.03570673614740372, + 0.011310670524835587, + 0.010150589048862457, + 0.008580894209444523, + 0.008922068402171135, + 0.011157079599797726, + 0.007647560443729162, + 0.013246318325400352, + 0.012523817829787731, + 0.011769902892410755, + 0.006637102924287319, + 0.007070654537528753, + 0.007565052714198828, + 0.013016194105148315, + 0.00792201329022646, + 0.012020437978208065, + 0.007752516306936741, + 0.011117491871118546, + 0.010188842192292213, + 0.01909647509455681, + 0.0177448857575655, + 0.013752263970673084, + 0.019045686349272728, + 0.012172085233032703, + 0.00901663489639759, + 0.006130793131887913, + 0.0064327046275138855, + 0.0079613933339715, + 0.007633191999047995, + 0.011235373094677925, + 0.006632552482187748, + 0.009830117225646973, + 0.00532270735129714, + 0.008267762139439583, + 0.010721476748585701, + 0.008334319107234478, + 0.008308145217597485, + 0.007159586530178785, + 0.005244192201644182, + 0.01009923405945301, + 0.004599714186042547, + 0.0044191377237439156, + 0.006406471133232117, + 0.009007171727716923, + 0.007175665348768234, + 0.00882394053041935, + 0.011045316234230995, + 0.00879611261188984, + 0.007053655106574297, + 0.0067996433936059475, + 0.012030678801238537, + 0.007586393039673567, + 0.0050149941816926, + 0.008742529898881912, + 0.017164135351777077, + 0.020094240084290504, + 0.01996169239282608, + 0.02347240410745144, + 0.019064342603087425, + 0.0204162560403347, + 0.01466582715511322, + 0.017104217782616615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.007630960550159216 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tightly framed underwater close-up, rendered with the breathtaking detail of high-speed nature photography, tracks sideways in stunning slow motion, capturing the primal moment of predation. A massive, torpedo-shaped longnose gar, its body covered in a cacophony of interwoven, diamond-shaped black scales shimmering with iridescent hues of olive green and bronze, dashes at explosive speed. Sunlight filters through the water, creating shifting caustic patterns across its form. Its long, bony snout flares slightly, revealing rows of terrifyingly sharp, translucent, needle-like teeth. Its target, a much smaller, silver-scaled blacknose fish, gleams in the dim light and desperately lunges sideways, its fins flared and its body contorted in sheer panic. The powerful, whip-like propulsion from the gar's tail violently churns the water, creating visible eddies and trails of tiny, shimmering bubbles and disturbed sediment that swirl around the two fish, all frozen in the surreal clarity of slow motion, creating the tense and dramatic atmosphere of a life-or-death chase.", + "seed": 17918, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_3397_seed_17918_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03360330071300268, + "vlm_logprob": -0.03360330071300268, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03360330071300268, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022534027695655823 + ], + "opposite_logprob_tokens": [ + -6.272533893585205 + ], + "logprob_reward": -0.022534027695655823, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.022534027695655823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025097491219639778, + 0.031250353902578354, + 0.04487362131476402, + 0.13190729916095734, + 0.04407816380262375, + 0.08391726762056351, + 0.0526619553565979, + 0.5281962752342224, + 0.0088017163798213, + 0.04921586811542511 + ], + "credit_map": [ + 0.002308617113158107, + 0.0026539419777691364, + 0.003967749420553446, + 0.0029706249479204416, + 0.0031425864435732365, + 0.0038427498657256365, + 0.0025203165132552385, + 0.003690903540700674, + 0.0036087583284825087, + 0.0025653226766735315, + 0.008011476136744022, + 0.0032140815164893866, + 0.0031081486959010363, + 0.007180471438914537, + 0.001786123146302998, + 0.001775968587026, + 0.014316906221210957, + 0.00319755170494318, + 0.003574318950995803, + 0.003130186116322875, + 0.001868181279860437, + 0.0028862531762570143, + 0.002952032955363393, + 0.012948187068104744, + 0.023375296965241432, + 0.010228719562292099, + 0.006094368640333414, + 0.030506577342748642, + 0.006761275231838226, + 0.007746178191155195, + 0.023264216259121895, + 0.023930659517645836, + 0.005056330934166908, + 0.003257565665990114, + 0.008786626160144806, + 0.003327200887724757, + 0.0037089327815920115, + 0.004053322598338127, + 0.004889178089797497, + 0.010999007150530815, + 0.004069156479090452, + 0.008285130374133587, + 0.006608903408050537, + 0.03300447389483452, + 0.0029607831966131926, + 0.00362664507701993, + 0.02244601584970951, + 0.0029161579441279173, + 0.0026925981510430574, + 0.003592651104554534, + 0.00848691537976265, + 0.008055990561842918, + 0.00962502509355545, + 0.013720570132136345, + 0.003935016691684723, + 0.002553181257098913, + 0.06352566927671432, + 0.023451469838619232, + 0.19188322126865387, + 0.15776954591274261, + 0.017806656658649445, + 0.020755497738718987, + 0.026272358372807503, + 0.026731858029961586, + 0.001294686459004879, + 0.0008409736328758299, + 0.0014719662722200155, + 0.0011184653267264366, + 0.0009385310695506632, + 0.0012735952623188496, + 0.0006883239257149398, + 0.001175174256786704, + 0.004559832625091076, + 0.0036903892178088427, + 0.005670133046805859, + 0.006559362635016441, + 0.009413602761924267, + 0.00924481451511383, + 0.005732739344239235, + 0.0043449923396110535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.022534027695655823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036328528076410294 + ], + "opposite_logprob_tokens": [ + -5.62863302230835 + ], + "logprob_reward": -0.0036328528076410294, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036328528076410294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06211162358522415, + 0.14177243411540985, + 0.09385263919830322, + 0.12256262451410294, + 0.0400485098361969, + 0.13957691192626953, + 0.047538794577121735, + 0.2540782392024994, + 0.031727392226457596, + 0.06673090159893036 + ], + "credit_map": [ + 0.006284913048148155, + 0.0077804261818528175, + 0.009375523775815964, + 0.009413586929440498, + 0.006813185289502144, + 0.008587977848947048, + 0.0056317830458283424, + 0.008224218152463436, + 0.011937393806874752, + 0.017441194504499435, + 0.03481760993599892, + 0.011632162146270275, + 0.012413163669407368, + 0.030999187380075455, + 0.009096402674913406, + 0.013435310684144497, + 0.02004844881594181, + 0.006933485623449087, + 0.00936080515384674, + 0.006165873724967241, + 0.006866551004350185, + 0.009171140380203724, + 0.007730023469775915, + 0.027576297521591187, + 0.013235748745501041, + 0.009908609092235565, + 0.008704273030161858, + 0.0377565436065197, + 0.008471057750284672, + 0.010319575667381287, + 0.022517848759889603, + 0.011648952029645443, + 0.004436913412064314, + 0.003975129686295986, + 0.0053031351417303085, + 0.0034948792308568954, + 0.004582326393574476, + 0.005730500444769859, + 0.006104080006480217, + 0.00642154086381197, + 0.012120168656110764, + 0.009914655238389969, + 0.013588500209152699, + 0.053173039108514786, + 0.006390728987753391, + 0.009531599469482899, + 0.027947526425123215, + 0.0069106705486774445, + 0.0036815162748098373, + 0.0037636447232216597, + 0.006988339591771364, + 0.006772213149815798, + 0.007325101178139448, + 0.00970735214650631, + 0.005207992624491453, + 0.004092629533261061, + 0.03073257952928543, + 0.011029739864170551, + 0.08891010284423828, + 0.07247991114854813, + 0.010336046107113361, + 0.013737122528254986, + 0.013416736386716366, + 0.013435961678624153, + 0.003147050505504012, + 0.0033968279603868723, + 0.004727375227957964, + 0.004376141354441643, + 0.003680017776787281, + 0.004429385997354984, + 0.0043050735257565975, + 0.003665516385808587, + 0.006028782110661268, + 0.005426617339253426, + 0.012562916614115238, + 0.010630680248141289, + 0.010787533596158028, + 0.009238284081220627, + 0.0064772008918225765, + 0.0055788797326385975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.0036328528076410294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003437727689743 + ], + "opposite_logprob_tokens": [ + -2.6950342655181885 + ], + "logprob_reward": -0.07003437727689743, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003437727689743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1148114949464798, + 0.08558198809623718, + 0.06477919220924377, + 0.0793428122997284, + 0.0361458919942379, + 0.06396729499101639, + 0.04291714355349541, + 0.47218409180641174, + 0.01760987564921379, + 0.022660203278064728 + ], + "credit_map": [ + 0.01278702449053526, + 0.016841435804963112, + 0.01702745258808136, + 0.013788155280053616, + 0.013313944451510906, + 0.014408236369490623, + 0.011622235178947449, + 0.015023012645542622, + 0.005643151234835386, + 0.004811808466911316, + 0.029867121949791908, + 0.005402515176683664, + 0.004934977274388075, + 0.027359768748283386, + 0.0034821939188987017, + 0.004080452956259251, + 0.018574004992842674, + 0.0037373483646661043, + 0.004031738732010126, + 0.003415245795622468, + 0.002245049923658371, + 0.003486448433250189, + 0.005002124235033989, + 0.024287233129143715, + 0.0108054643496871, + 0.006518024485558271, + 0.006428379099816084, + 0.0171820055693388, + 0.005442796740680933, + 0.00578128919005394, + 0.015173832885921001, + 0.012011022306978703, + 0.0036403529811650515, + 0.0033713735174387693, + 0.005119268782436848, + 0.0030271271243691444, + 0.003925404977053404, + 0.004637559875845909, + 0.005157973617315292, + 0.007266830187290907, + 0.012131617404520512, + 0.015909966081380844, + 0.0054738204926252365, + 0.0062750098295509815, + 0.005748499650508165, + 0.004536104388535023, + 0.009103180840611458, + 0.004789094906300306, + 0.0020336699672043324, + 0.003128980752080679, + 0.007516713347285986, + 0.008280517533421516, + 0.006468836683779955, + 0.00843766238540411, + 0.004254455212503672, + 0.0027963079046458006, + 0.057158470153808594, + 0.01969822682440281, + 0.16780072450637817, + 0.13810878992080688, + 0.01781114563345909, + 0.021872837096452713, + 0.024370452389121056, + 0.025363456457853317, + 0.0016692329663783312, + 0.002434611087664962, + 0.002331569092348218, + 0.0024665375240147114, + 0.0024862363934516907, + 0.002733121160417795, + 0.002054890850558877, + 0.0014336761087179184, + 0.0017350679263472557, + 0.0017983716679736972, + 0.0035192177165299654, + 0.003826085478067398, + 0.004717451054602861, + 0.002881087129935622, + 0.002182208700105548, + 0.0020007137209177017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.07003437727689743 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062033720314502716 + ], + "opposite_logprob_tokens": [ + -2.8120336532592773 + ], + "logprob_reward": -0.062033720314502716, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062033720314502716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0845891460776329, + 0.1509922593832016, + 0.0739944651722908, + 0.08552850782871246, + 0.12454862147569656, + 0.1643371284008026, + 0.04847194254398346, + 0.10538461804389954, + 0.06960171461105347, + 0.0925515741109848 + ], + "credit_map": [ + 0.007112871389836073, + 0.009779391810297966, + 0.01745280995965004, + 0.007484641391783953, + 0.007049722597002983, + 0.02086101472377777, + 0.006317394319921732, + 0.008531298488378525, + 0.009807647205889225, + 0.006467164494097233, + 0.054437316954135895, + 0.010519693605601788, + 0.00860761571675539, + 0.04954130947589874, + 0.006152595859020948, + 0.005458906292915344, + 0.01252054050564766, + 0.005797816906124353, + 0.008245796896517277, + 0.005435083992779255, + 0.006677665282040834, + 0.008403318002820015, + 0.006986625492572784, + 0.0199276152998209, + 0.009245352819561958, + 0.010657014325261116, + 0.006479751318693161, + 0.018772415816783905, + 0.007964267395436764, + 0.009714148007333279, + 0.007381162606179714, + 0.015314402990043163, + 0.011435236781835556, + 0.009023451246321201, + 0.027439502999186516, + 0.0070843156427145, + 0.009138979017734528, + 0.010748796164989471, + 0.01410722266882658, + 0.03557111322879791, + 0.006645489018410444, + 0.00828317180275917, + 0.01773289032280445, + 0.07124491035938263, + 0.005732453893870115, + 0.007529913913458586, + 0.03968202322721481, + 0.007486290298402309, + 0.0035702933091670275, + 0.005148957017809153, + 0.010741434060037136, + 0.010079546831548214, + 0.004356003366410732, + 0.007058161310851574, + 0.00425188010558486, + 0.0032656688708812, + 0.014370101504027843, + 0.008174510672688484, + 0.022894876077771187, + 0.019259657710790634, + 0.008558608591556549, + 0.011076618917286396, + 0.012025087140500546, + 0.009025152772665024, + 0.007626925595104694, + 0.011817646212875843, + 0.008487206883728504, + 0.008838746696710587, + 0.010986367240548134, + 0.008648449555039406, + 0.008151283487677574, + 0.005045088939368725, + 0.005930705461651087, + 0.006063884124159813, + 0.011897689662873745, + 0.01563984900712967, + 0.02287692204117775, + 0.012379990890622139, + 0.010063405148684978, + 0.007699126843363047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.062033720314502716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978152547031641 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.00978152547031641, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978152547031641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1086215153336525, + 0.08835820108652115, + 0.0550207681953907, + 0.11395055055618286, + 0.037256963551044464, + 0.13505128026008606, + 0.0803849995136261, + 0.2958909273147583, + 0.04024040699005127, + 0.045224349945783615 + ], + "credit_map": [ + 0.008644837886095047, + 0.012204977683722973, + 0.01589246466755867, + 0.01441781222820282, + 0.012175510637462139, + 0.01896454021334648, + 0.01142054796218872, + 0.014900822192430496, + 0.0059004658833146095, + 0.006154745817184448, + 0.0277788694947958, + 0.006450282875448465, + 0.007203469518572092, + 0.026908019557595253, + 0.00383970164693892, + 0.004122648388147354, + 0.011696680448949337, + 0.0043975287117064, + 0.006660728715360165, + 0.004779815208166838, + 0.0032710714731365442, + 0.004851263016462326, + 0.006297048646956682, + 0.013066633604466915, + 0.019362077116966248, + 0.010466894134879112, + 0.0064432453364133835, + 0.030018454417586327, + 0.006657033693045378, + 0.007266566157341003, + 0.01894240826368332, + 0.014793884940445423, + 0.005937665235251188, + 0.0025111273862421513, + 0.006319385953247547, + 0.0034745903685688972, + 0.0031218009535223246, + 0.004643830470740795, + 0.004922232590615749, + 0.006326331757009029, + 0.00889576692134142, + 0.006770761217921972, + 0.011763511225581169, + 0.05543527752161026, + 0.004636166151612997, + 0.006841811817139387, + 0.03437858074903488, + 0.006329420953989029, + 0.0036130284424871206, + 0.006037882994860411, + 0.01296639908105135, + 0.014800219796597958, + 0.013088393025100231, + 0.019092129543423653, + 0.006777762435376644, + 0.004009184427559376, + 0.0347072035074234, + 0.0126225296407938, + 0.1083461120724678, + 0.08892121911048889, + 0.010253974236547947, + 0.013109665364027023, + 0.014476977288722992, + 0.013453265652060509, + 0.0036185102071613073, + 0.005042762495577335, + 0.006208096630871296, + 0.005260382778942585, + 0.0045652007684111595, + 0.006115259137004614, + 0.004827249329537153, + 0.004602947272360325, + 0.00360638159327209, + 0.004424951504915953, + 0.005655871704220772, + 0.0069196270778775215, + 0.006520780734717846, + 0.007703179493546486, + 0.00655350973829627, + 0.003840050660073757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.00978152547031641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tightly framed underwater close-up, rendered with the breathtaking detail of high-speed nature photography, tracks sideways in stunning slow motion, capturing the primal moment of predation. A massive, torpedo-shaped longnose gar, its body covered in a cacophony of interwoven, diamond-shaped black scales shimmering with iridescent hues of olive green and bronze, dashes at explosive speed. Sunlight filters through the water, creating shifting caustic patterns across its form. Its long, bony snout flares slightly, revealing rows of terrifyingly sharp, translucent, needle-like teeth. Its target, a much smaller, silver-scaled blacknose fish, gleams in the dim light and desperately lunges sideways, its fins flared and its body contorted in sheer panic. The powerful, whip-like propulsion from the gar's tail violently churns the water, creating visible eddies and trails of tiny, shimmering bubbles and disturbed sediment that swirl around the two fish, all frozen in the surreal clarity of slow motion, creating the tense and dramatic atmosphere of a life-or-death chase.", + "seed": 17919, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_1/index_3397_seed_17919_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03326196651905775, + "vlm_logprob": -0.03326196651905775, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.03326196651905775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017111873254179955 + ], + "opposite_logprob_tokens": [ + -6.892111778259277 + ], + "logprob_reward": -0.017111873254179955, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.017111873254179955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2933093011379242, + 0.10138483345508575, + 0.05979771912097931, + 0.08188899606466293, + 0.06528819352388382, + 0.052315209060907364, + 0.05438627302646637, + 0.07493055611848831, + 0.09203822910785675, + 0.12466075271368027 + ], + "credit_map": [ + 0.05798238515853882, + 0.048439592123031616, + 0.044924769550561905, + 0.04037095233798027, + 0.024507256224751472, + 0.025787074118852615, + 0.024087851867079735, + 0.02720942348241806, + 0.010948730632662773, + 0.009204213507473469, + 0.015089343301951885, + 0.017119422554969788, + 0.012284008786082268, + 0.007851322181522846, + 0.013333063572645187, + 0.015554724261164665, + 0.005699416622519493, + 0.00918612815439701, + 0.011732218787074089, + 0.010815798304975033, + 0.004347056616097689, + 0.005267533008009195, + 0.007468024268746376, + 0.005281541962176561, + 0.006324057001620531, + 0.005602581892162561, + 0.01320442371070385, + 0.013171875849366188, + 0.011537262238562107, + 0.01298117358237505, + 0.010416872799396515, + 0.008650751784443855, + 0.004267462529242039, + 0.0066801076754927635, + 0.014787725172936916, + 0.011468683369457722, + 0.006769201252609491, + 0.008638604544103146, + 0.008669373579323292, + 0.004007038194686174, + 0.005608357023447752, + 0.003604310564696789, + 0.0069932229816913605, + 0.005938720889389515, + 0.004581638611853123, + 0.006833620369434357, + 0.008255508728325367, + 0.010499829426407814, + 0.008897350169718266, + 0.003988601732999086, + 0.005966637283563614, + 0.006416248623281717, + 0.004111248534172773, + 0.005271977744996548, + 0.007828771136701107, + 0.011905437335371971, + 0.009467773139476776, + 0.005923525430262089, + 0.008408132940530777, + 0.013786211609840393, + 0.006519922520965338, + 0.011017444543540478, + 0.010700157843530178, + 0.009107386693358421, + 0.01390696968883276, + 0.013686060905456543, + 0.014140513725578785, + 0.01064556185156107, + 0.010649117641150951, + 0.011470627970993519, + 0.008524238131940365, + 0.009015137329697609, + 0.01970624551177025, + 0.02330305241048336, + 0.020378967747092247, + 0.015394246205687523, + 0.015704093500971794, + 0.008554127998650074, + 0.01152151357382536, + 0.010098504833877087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.017111873254179955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003207421861588955 + ], + "opposite_logprob_tokens": [ + -5.753207206726074 + ], + "logprob_reward": -0.003207421861588955, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003207421861588955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19776256382465363, + 0.12107302993535995, + 0.030071143060922623, + 0.04752784222364426, + 0.060851238667964935, + 0.055599816143512726, + 0.061835531145334244, + 0.09568746387958527, + 0.20955197513103485, + 0.12003941833972931 + ], + "credit_map": [ + 0.028914393857121468, + 0.028791962191462517, + 0.03377014771103859, + 0.020217789337038994, + 0.01824331283569336, + 0.033182866871356964, + 0.015269512310624123, + 0.019372565671801567, + 0.01122965570539236, + 0.009870386682450771, + 0.019470976665616035, + 0.025959372520446777, + 0.011014353483915329, + 0.010039673186838627, + 0.019268400967121124, + 0.014220209792256355, + 0.0029924402479082346, + 0.004815458320081234, + 0.003488691058009863, + 0.0053468975238502026, + 0.003329542698338628, + 0.0030561794992536306, + 0.0032013587187975645, + 0.0038405759260058403, + 0.0035449443385004997, + 0.00407921988517046, + 0.00682262284681201, + 0.007947312667965889, + 0.005064524710178375, + 0.00843786634504795, + 0.00597199285402894, + 0.00565935717895627, + 0.004683005157858133, + 0.008027341216802597, + 0.013715866953134537, + 0.00829100701957941, + 0.0050873225554823875, + 0.007753270212560892, + 0.00752167496830225, + 0.0057717557065188885, + 0.005590700078755617, + 0.004982863552868366, + 0.006461499258875847, + 0.00957332644611597, + 0.0063701895996928215, + 0.007025548256933689, + 0.008647686801850796, + 0.0069479988887906075, + 0.009296780452132225, + 0.004423520527780056, + 0.006209244951605797, + 0.009629213251173496, + 0.004127967171370983, + 0.007642567157745361, + 0.00851854681968689, + 0.011987688019871712, + 0.012666665017604828, + 0.008492549881339073, + 0.010897965170443058, + 0.010563597083091736, + 0.010366522707045078, + 0.015586220659315586, + 0.013270499184727669, + 0.013843445107340813, + 0.03205886483192444, + 0.03723888844251633, + 0.02584267221391201, + 0.0269938837736845, + 0.028381898999214172, + 0.023504750803112984, + 0.017443282529711723, + 0.01808774657547474, + 0.022955937311053276, + 0.015685072168707848, + 0.016866818070411682, + 0.013727321289479733, + 0.011914554983377457, + 0.014636043459177017, + 0.013906759209930897, + 0.01034691371023655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.003207421861588955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003571093082428 + ], + "opposite_logprob_tokens": [ + -2.695035696029663 + ], + "logprob_reward": -0.07003571093082428, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003571093082428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1905856430530548, + 0.06474227458238602, + 0.049632731825113297, + 0.05593430995941162, + 0.051150355488061905, + 0.07621332257986069, + 0.03448023274540901, + 0.06072433292865753, + 0.20659516751766205, + 0.20994171500205994 + ], + "credit_map": [ + 0.030901553109288216, + 0.034593336284160614, + 0.0317600853741169, + 0.020367488265037537, + 0.018317539244890213, + 0.023949146270751953, + 0.011960997246205807, + 0.0187354888767004, + 0.006252320017665625, + 0.006142660975456238, + 0.010915003716945648, + 0.01069973036646843, + 0.007409664802253246, + 0.005964782554656267, + 0.009037799201905727, + 0.008320299908518791, + 0.0035359812900424004, + 0.005894272588193417, + 0.009756436571478844, + 0.010259302332997322, + 0.0024834272917360067, + 0.004948068410158157, + 0.0073541393503546715, + 0.005401101894676685, + 0.00581643870100379, + 0.004891016986221075, + 0.005422760266810656, + 0.005188941024243832, + 0.0073600332252681255, + 0.009770190343260765, + 0.007238885387778282, + 0.010246037505567074, + 0.0032431459985673428, + 0.004093557130545378, + 0.010245033539831638, + 0.007495866157114506, + 0.005133174359798431, + 0.006291485391557217, + 0.006634037010371685, + 0.008014047518372536, + 0.0115738520398736, + 0.004920029081404209, + 0.009914936497807503, + 0.014423970133066177, + 0.00576797453686595, + 0.008685490116477013, + 0.010951843112707138, + 0.00997521635144949, + 0.005218955688178539, + 0.0028692183550447226, + 0.0035224587190896273, + 0.004524462856352329, + 0.002099081175401807, + 0.004040472209453583, + 0.00594579242169857, + 0.0062597887590527534, + 0.009830703027546406, + 0.005635327193886042, + 0.007966180332005024, + 0.008755049668252468, + 0.004767188336700201, + 0.008368060924112797, + 0.008575932122766972, + 0.006825882475823164, + 0.03202347829937935, + 0.03360534459352493, + 0.023372864350676537, + 0.024183232337236404, + 0.028635701164603233, + 0.0252069178968668, + 0.018575889989733696, + 0.020991718396544456, + 0.03076249733567238, + 0.023805171251296997, + 0.027185553684830666, + 0.027818365022540092, + 0.021688664332032204, + 0.029505642130970955, + 0.029654575511813164, + 0.019521230831742287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.07003571093082428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000458985567093 + ], + "opposite_logprob_tokens": [ + -2.69500470161438 + ], + "logprob_reward": -0.07000458985567093, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000458985567093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1905859112739563, + 0.12153024226427078, + 0.09395278990268707, + 0.05881825461983681, + 0.05992360785603523, + 0.13570617139339447, + 0.051578667014837265, + 0.08651822805404663, + 0.0750725194811821, + 0.12631355226039886 + ], + "credit_map": [ + 0.0191832035779953, + 0.01892222836613655, + 0.027567746117711067, + 0.026935916393995285, + 0.021828433498740196, + 0.028784457594156265, + 0.02196655236184597, + 0.025397393852472305, + 0.010797510854899883, + 0.009101875126361847, + 0.019718503579497337, + 0.02160486951470375, + 0.012419451028108597, + 0.008664948865771294, + 0.02150721475481987, + 0.017715882509946823, + 0.008223721757531166, + 0.008917099796235561, + 0.0188988596200943, + 0.013218008913099766, + 0.01043553464114666, + 0.009059442207217216, + 0.009547390043735504, + 0.015652742236852646, + 0.005130245815962553, + 0.0051979306153953075, + 0.006330611649900675, + 0.006811138708144426, + 0.011284629814326763, + 0.008670295588672161, + 0.006413185503333807, + 0.00898022297769785, + 0.0043676369823515415, + 0.004177755210548639, + 0.008313827216625214, + 0.011441444978117943, + 0.006557750981301069, + 0.009850655682384968, + 0.00965424720197916, + 0.0055602905340492725, + 0.016919998452067375, + 0.008631374686956406, + 0.01992642506957054, + 0.02564193308353424, + 0.009769290685653687, + 0.014185147359967232, + 0.021534284576773643, + 0.019097717478871346, + 0.009317630901932716, + 0.004702528938651085, + 0.004090432543307543, + 0.0055464147590100765, + 0.0038224137388169765, + 0.007115119136869907, + 0.007217024452984333, + 0.009767106734216213, + 0.007768088486045599, + 0.007598549127578735, + 0.013597755692899227, + 0.014824668876826763, + 0.00864595826715231, + 0.011128334328532219, + 0.012019501067698002, + 0.010935378260910511, + 0.010518057271838188, + 0.01720123551785946, + 0.005498803686350584, + 0.008863572031259537, + 0.008713601157069206, + 0.00730513408780098, + 0.0073367212899029255, + 0.009635405614972115, + 0.021293852478265762, + 0.017642412334680557, + 0.016391245648264885, + 0.018842514604330063, + 0.015797367319464684, + 0.015093554742634296, + 0.010938500985503197, + 0.010314112529158592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.07000458985567093 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a tightly framed close-up composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, torpedo-shaped longnose gar with diamond-shaped black scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a smaller, silver-scaled blacknose fish visible as the target of the gar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the gar dashing at explosive speed while the camera tracks sideways in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible eddies, trails of tiny bubbles, and disturbed sediment swirling around the fish due to the tail propulsion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950236693024635 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950236693024635, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950236693024635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27494433522224426, + 0.06045420840382576, + 0.03805489093065262, + 0.06636412441730499, + 0.04274639114737511, + 0.07955857366323471, + 0.048523470759391785, + 0.10821633040904999, + 0.15651975572109222, + 0.12461784482002258 + ], + "credit_map": [ + 0.03253971040248871, + 0.023119620978832245, + 0.05068109929561615, + 0.03155170753598213, + 0.021733270958065987, + 0.03290228918194771, + 0.025339744985103607, + 0.05707689747214317, + 0.006512918509542942, + 0.005375634413212538, + 0.008862669579684734, + 0.009627488441765308, + 0.006540847010910511, + 0.004636391997337341, + 0.009997472167015076, + 0.008900786750018597, + 0.00339921354316175, + 0.0041649192571640015, + 0.006971192080527544, + 0.003866454353556037, + 0.0048773521557450294, + 0.004588959272950888, + 0.0035302387550473213, + 0.0066565657034516335, + 0.004281600471585989, + 0.005138348322361708, + 0.00763735082000494, + 0.007862206548452377, + 0.01161736436188221, + 0.013420706614851952, + 0.008288645185530186, + 0.00811790581792593, + 0.0028983147349208593, + 0.0036627138033509254, + 0.006390805821865797, + 0.007131150458008051, + 0.004735154565423727, + 0.007535397540777922, + 0.005714653059840202, + 0.004678203258663416, + 0.009191333316266537, + 0.0058456226252019405, + 0.009534030221402645, + 0.014888486824929714, + 0.005483245011419058, + 0.009760689921677113, + 0.012712454423308372, + 0.012142711319029331, + 0.009053119458258152, + 0.004181956872344017, + 0.00483241630718112, + 0.004343864042311907, + 0.0037526735104620457, + 0.003965785726904869, + 0.006572658661752939, + 0.011820993386209011, + 0.009838770143687725, + 0.00992243830114603, + 0.018683768808841705, + 0.01936439424753189, + 0.010768156498670578, + 0.01021851971745491, + 0.015789633616805077, + 0.013630648143589497, + 0.021207576617598534, + 0.025391750037670135, + 0.016155321151018143, + 0.016538139432668686, + 0.01821853220462799, + 0.020574111491441727, + 0.018291296437382698, + 0.02014303207397461, + 0.024237902835011482, + 0.017062272876501083, + 0.01769046112895012, + 0.015208967961370945, + 0.014637160114943981, + 0.016224823892116547, + 0.009820649400353432, + 0.009735600091516972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.005950236693024635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_tensor([8273, 8273])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_tensor([8273, 8273])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..65ec127b3675e94d59fcbeccba59034cb74cb729 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_tensor([8273, 8273])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly upward angle depicts the camera in a slow, smooth orbital rotation around a suspended red paper lantern. The lantern, a full circle, is crafted from thin, translucent rice paper. Its vibrant vermilion color stands out against the inky darkness of the night, its slender bamboo framework faintly visible in the light from within. Inside, a warm candle flame burns steadily, emitting a soft orange glow through the paper, transforming the lantern itself into a luminous object and casting dynamic shadows across its surface. Gently rustled by the night breeze, the lantern sways gently and rhythmically in the air, its golden tassels dangling from its base flowing gracefully, their threads shimmering in the dim light. The combination of the camera's relentless rotation and the lantern's gentle trembling creates a hypnotic visual effect, imbuing the image with a tranquil, contemplative, and classically mysterious atmosphere. The out-of-focus black outlines of the traditional building's eaves in the background further enhance this serene feeling.", + "seed": 25619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_8273_seed_25619_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.049086987785995004, + "vlm_logprob": -0.049086987785995004, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.049086987785995004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003252931870520115 + ], + "opposite_logprob_tokens": [ + -7.753252983093262 + ], + "logprob_reward": -0.003252931870520115, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.003252931870520115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10778021067380905, + 0.050181686878204346, + 0.061431583017110825, + 0.07826614379882812, + 0.1317097693681717, + 0.2749938666820526, + 0.08587531000375748, + 0.09588025510311127, + 0.06432714313268661, + 0.04955405741930008 + ], + "credit_map": [ + 0.010947898961603642, + 0.007540386635810137, + 0.010555604472756386, + 0.009509146213531494, + 0.01094137504696846, + 0.01848449744284153, + 0.017014654353260994, + 0.022786637768149376, + 0.005296066403388977, + 0.004529704339802265, + 0.004633444361388683, + 0.007610831409692764, + 0.006663716863840818, + 0.007134940940886736, + 0.005160157568752766, + 0.009152824990451336, + 0.005844346713274717, + 0.007316778413951397, + 0.0057339295744895935, + 0.010163813829421997, + 0.008915267884731293, + 0.009693608619272709, + 0.0051390849985182285, + 0.008624754846096039, + 0.008401746861636639, + 0.009108318015933037, + 0.008204725570976734, + 0.008375005796551704, + 0.010047322139143944, + 0.012512042187154293, + 0.009099862538278103, + 0.012517118826508522, + 0.01721060834825039, + 0.016458263620734215, + 0.017647316679358482, + 0.022500064224004745, + 0.011758402921259403, + 0.016358667984604836, + 0.018310604616999626, + 0.011465842835605145, + 0.0264340378344059, + 0.029696205630898476, + 0.02620411105453968, + 0.03737611323595047, + 0.026783494278788567, + 0.037683866918087006, + 0.06115478277206421, + 0.029661281034350395, + 0.01086997427046299, + 0.011150173842906952, + 0.008323400281369686, + 0.006513099651783705, + 0.009243735112249851, + 0.013944454491138458, + 0.009109445847570896, + 0.016721026971936226, + 0.023493260145187378, + 0.00889662280678749, + 0.008165056817233562, + 0.005319238640367985, + 0.007075640372931957, + 0.015200702473521233, + 0.009018848650157452, + 0.01871088333427906, + 0.007741395849734545, + 0.006144373677670956, + 0.007502482272684574, + 0.007462846580892801, + 0.007288310211151838, + 0.008799513801932335, + 0.008779874071478844, + 0.01060834713280201, + 0.004871191456913948, + 0.006054282654076815, + 0.007195166777819395, + 0.007845102809369564, + 0.0044545978307724, + 0.005313051864504814, + 0.0074125416576862335, + 0.006408120971173048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.003252931870520115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495249107480049 + ], + "opposite_logprob_tokens": [ + -2.929952383041382 + ], + "logprob_reward": -0.05495249107480049, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495249107480049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14241057634353638, + 0.0481395460665226, + 0.1213388442993164, + 0.07529163360595703, + 0.115968719124794, + 0.22028742730617523, + 0.15388959646224976, + 0.07324983924627304, + 0.03319291025400162, + 0.016230899840593338 + ], + "credit_map": [ + 0.008860701695084572, + 0.011764442548155785, + 0.013377831317484379, + 0.01717069372534752, + 0.019919492304325104, + 0.023869667202234268, + 0.021573608741164207, + 0.025874141603708267, + 0.004857430700212717, + 0.0037898768205195665, + 0.005705395247787237, + 0.006353117525577545, + 0.004129508044570684, + 0.008810725063085556, + 0.005034198518842459, + 0.009459299966692924, + 0.02081638015806675, + 0.010694312863051891, + 0.013839298859238625, + 0.0071556237526237965, + 0.016138827428221703, + 0.021037135273218155, + 0.007992198690772057, + 0.02366507425904274, + 0.005822786130011082, + 0.006537608336657286, + 0.009146692231297493, + 0.012194143608212471, + 0.010408109985291958, + 0.011889069341123104, + 0.006056950427591801, + 0.0132362749427557, + 0.01306045800447464, + 0.016943737864494324, + 0.01393711194396019, + 0.019814172759652138, + 0.010309839621186256, + 0.015972048044204712, + 0.014805111102759838, + 0.011126257479190826, + 0.021609816700220108, + 0.019841423258185387, + 0.024075310677289963, + 0.02673257887363434, + 0.018737608566880226, + 0.032819461077451706, + 0.04768335819244385, + 0.028787869960069656, + 0.019529396668076515, + 0.017983736470341682, + 0.01137723308056593, + 0.013269211165606976, + 0.01000728365033865, + 0.028879376128315926, + 0.014131303876638412, + 0.038712065666913986, + 0.018386585637927055, + 0.0077966866083443165, + 0.0066296374425292015, + 0.0033794748596847057, + 0.00663815438747406, + 0.011763921938836575, + 0.0057173827663064, + 0.012938003987073898, + 0.003595625516027212, + 0.0035820933990180492, + 0.0026370801497250795, + 0.003508487017825246, + 0.002791617764160037, + 0.004914580378681421, + 0.007167458534240723, + 0.0049959709867835045, + 0.0017573819495737553, + 0.0020084462594240904, + 0.0021251370199024677, + 0.0014972947537899017, + 0.0011818697676062584, + 0.002145372796803713, + 0.002780045149847865, + 0.0027353542391210794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.05495249107480049 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041186451911926 + ], + "opposite_logprob_tokens": [ + -1.9104118347167969 + ], + "logprob_reward": -0.16041186451911926, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041186451911926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16334660351276398, + 0.19235225021839142, + 0.09098147600889206, + 0.05125649645924568, + 0.10668580234050751, + 0.11659687012434006, + 0.06093651428818703, + 0.057279545813798904, + 0.10613124072551727, + 0.05443317070603371 + ], + "credit_map": [ + 0.01763226091861725, + 0.01897369511425495, + 0.015223072841763496, + 0.021475641056895256, + 0.020560288801789284, + 0.02401280403137207, + 0.021575184538960457, + 0.02389366552233696, + 0.017432790249586105, + 0.027334684506058693, + 0.012170230969786644, + 0.0344722755253315, + 0.022312726825475693, + 0.025702787563204765, + 0.015682706609368324, + 0.03724405914545059, + 0.013266723603010178, + 0.007055132184177637, + 0.008264148607850075, + 0.004610330797731876, + 0.012014929205179214, + 0.01780168153345585, + 0.006716607604175806, + 0.021251937374472618, + 0.0051808119751513, + 0.005213816650211811, + 0.004549793899059296, + 0.005999820772558451, + 0.005119072739034891, + 0.009874100796878338, + 0.00422881543636322, + 0.011090267449617386, + 0.013109217397868633, + 0.01469629630446434, + 0.013581153936684132, + 0.018994269892573357, + 0.010026526637375355, + 0.013481241650879383, + 0.011864660307765007, + 0.01093243807554245, + 0.013664648868143559, + 0.014344229362905025, + 0.01565057970583439, + 0.021599039435386658, + 0.007140968460589647, + 0.015241648070514202, + 0.016286956146359444, + 0.012668806128203869, + 0.006311255507171154, + 0.008476885966956615, + 0.00399687048047781, + 0.004449014086276293, + 0.004215140827000141, + 0.010936394333839417, + 0.007138154003769159, + 0.015412802807986736, + 0.008095461875200272, + 0.004879192914813757, + 0.007115049287676811, + 0.00477336160838604, + 0.004591710399836302, + 0.007819697260856628, + 0.01111290231347084, + 0.008892172947525978, + 0.014669517986476421, + 0.018356909975409508, + 0.010015079751610756, + 0.013172701001167297, + 0.00969649013131857, + 0.014664188027381897, + 0.009136172011494637, + 0.01642019860446453, + 0.005839155055582523, + 0.008058334700763226, + 0.007095057517290115, + 0.0076365708373487, + 0.00557142635807395, + 0.007209884002804756, + 0.005696967244148254, + 0.007325783371925354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.16041186451911926 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17924709618091583, + 0.11780285090208054, + 0.19459015130996704, + 0.08898855745792389, + 0.09521650522947311, + 0.05158577114343643, + 0.10174991190433502, + 0.07053820788860321, + 0.04273664951324463, + 0.05754432454705238 + ], + "credit_map": [ + 0.016383694484829903, + 0.016258040443062782, + 0.017465336248278618, + 0.019244207069277763, + 0.015704652294516563, + 0.02914569340646267, + 0.028745144605636597, + 0.03630032762885094, + 0.011008950881659985, + 0.017547059804201126, + 0.0077263666316866875, + 0.024850016459822655, + 0.01303288247436285, + 0.015150196850299835, + 0.008104396052658558, + 0.02038297988474369, + 0.027447618544101715, + 0.018894298002123833, + 0.021078290417790413, + 0.011840050108730793, + 0.028207551687955856, + 0.03409154713153839, + 0.01463882252573967, + 0.03839198499917984, + 0.009521597065031528, + 0.012737595476210117, + 0.012208306230604649, + 0.01488320343196392, + 0.011334246024489403, + 0.01135141309350729, + 0.00630611227825284, + 0.010646084323525429, + 0.010793913155794144, + 0.013592647388577461, + 0.012147020548582077, + 0.016996221616864204, + 0.009393692016601562, + 0.012749649584293365, + 0.01293235830962658, + 0.006611002143472433, + 0.005035432055592537, + 0.006974717136472464, + 0.0047885216772556305, + 0.005507920868694782, + 0.0033490422647446394, + 0.008443078026175499, + 0.008911225013434887, + 0.008575832471251488, + 0.010508903302252293, + 0.011169754900038242, + 0.007534092757850885, + 0.008607184514403343, + 0.006070691626518965, + 0.019994113594293594, + 0.009681700728833675, + 0.02818346954882145, + 0.010928958654403687, + 0.0056700087152421474, + 0.006293242797255516, + 0.004861867055296898, + 0.0041599576361477375, + 0.011655420064926147, + 0.012538796290755272, + 0.014429953880608082, + 0.005819585174322128, + 0.003985804505646229, + 0.0038883888628333807, + 0.004082677885890007, + 0.005145946983247995, + 0.006672114599496126, + 0.007128010969609022, + 0.006014117505401373, + 0.00689358776435256, + 0.006636310834437609, + 0.01007452979683876, + 0.0065348488278687, + 0.005261710379272699, + 0.007070099003612995, + 0.0076070851646363735, + 0.007466150913387537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00864158570766449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176065757870674 + ], + "opposite_logprob_tokens": [ + -4.018176078796387 + ], + "logprob_reward": -0.018176065757870674, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176065757870674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2399892508983612, + 0.027655353769659996, + 0.028896408155560493, + 0.025491958484053612, + 0.026709547266364098, + 0.0575096569955349, + 0.02734445035457611, + 0.5470083951950073, + 0.011963648721575737, + 0.00743133993819356 + ], + "credit_map": [ + 0.025223493576049805, + 0.02159227803349495, + 0.021692046895623207, + 0.024369211867451668, + 0.027281850576400757, + 0.03754870221018791, + 0.03456475958228111, + 0.047716885805130005, + 0.0021576012950390577, + 0.004910044837743044, + 0.0016802343307062984, + 0.006075055338442326, + 0.003562597557902336, + 0.003273474518209696, + 0.0021407820750027895, + 0.0038555627688765526, + 0.0033381562680006027, + 0.002614021534100175, + 0.002806887263432145, + 0.0036600930616259575, + 0.004007660783827305, + 0.004578207153826952, + 0.0023062899708747864, + 0.005585090257227421, + 0.002623934531584382, + 0.002621882129460573, + 0.003195166354998946, + 0.003874351503327489, + 0.0030409872997552156, + 0.00383383734151721, + 0.0018726506968960166, + 0.0044291457161307335, + 0.001863118028268218, + 0.002960158046334982, + 0.0034444513730704784, + 0.0032272569369524717, + 0.002482885494828224, + 0.004592603538185358, + 0.002996304305270314, + 0.0051427665166556835, + 0.00444863410666585, + 0.006410618312656879, + 0.006200958043336868, + 0.009437392465770245, + 0.004059886559844017, + 0.008036522194743156, + 0.012208627536892891, + 0.006707006599754095, + 0.002380329417064786, + 0.002061162143945694, + 0.0021596939768642187, + 0.002241008449345827, + 0.001877357135526836, + 0.005813728552311659, + 0.003131608013063669, + 0.0076795597560703754, + 0.1512916386127472, + 0.06281742453575134, + 0.04719921946525574, + 0.021848207339644432, + 0.04548725113272667, + 0.08549094200134277, + 0.03706301003694534, + 0.09581062197685242, + 0.0017449397128075361, + 0.0013585066189989448, + 0.0013646320439875126, + 0.00137926428578794, + 0.001021668897010386, + 0.0016384761547669768, + 0.0016947861295193434, + 0.0017613725503906608, + 0.000893854652531445, + 0.0009642005315981805, + 0.001010883366689086, + 0.0007519921637140214, + 0.001086950651369989, + 0.0009999535977840424, + 0.0007141120149753988, + 0.0010093924356624484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.018176065757870674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle depicts the camera in a slow, smooth orbital rotation around a suspended red paper lantern. The lantern, a full circle, is crafted from thin, translucent rice paper. Its vibrant vermilion color stands out against the inky darkness of the night, its slender bamboo framework faintly visible in the light from within. Inside, a warm candle flame burns steadily, emitting a soft orange glow through the paper, transforming the lantern itself into a luminous object and casting dynamic shadows across its surface. Gently rustled by the night breeze, the lantern sways gently and rhythmically in the air, its golden tassels dangling from its base flowing gracefully, their threads shimmering in the dim light. The combination of the camera's relentless rotation and the lantern's gentle trembling creates a hypnotic visual effect, imbuing the image with a tranquil, contemplative, and classically mysterious atmosphere. The out-of-focus black outlines of the traditional building's eaves in the background further enhance this serene feeling.", + "seed": 25620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_8273_seed_25620_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024954154063016175, + "vlm_logprob": -0.024954154063016175, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.024954154063016175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022800182923674583 + ], + "opposite_logprob_tokens": [ + -8.002280235290527 + ], + "logprob_reward": -0.0022800182923674583, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.0022800182923674583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08109009265899658, + 0.10259275883436203, + 0.3583647310733795, + 0.0557878278195858, + 0.0344352200627327, + 0.06338857859373093, + 0.025254154577851295, + 0.1417466700077057, + 0.04009602591395378, + 0.09724389016628265 + ], + "credit_map": [ + 0.013606862165033817, + 0.014443733729422092, + 0.011680524796247482, + 0.012244602665305138, + 0.00732953567057848, + 0.007972044870257378, + 0.00531002227216959, + 0.00850276742130518, + 0.020586054772138596, + 0.0204043909907341, + 0.020521340891718864, + 0.01432418916374445, + 0.005715509876608849, + 0.0064390418119728565, + 0.006865363102406263, + 0.007736878003925085, + 0.033323947340250015, + 0.11302682757377625, + 0.12901948392391205, + 0.03419477865099907, + 0.011566144414246082, + 0.009636721573770046, + 0.012235557660460472, + 0.015361267141997814, + 0.006468447856605053, + 0.005647715646773577, + 0.007806037552654743, + 0.005509998649358749, + 0.007128679193556309, + 0.006888759322464466, + 0.008506406098604202, + 0.007831783965229988, + 0.008299696259200573, + 0.002328268950805068, + 0.006551687605679035, + 0.0053665670566260815, + 0.0030051551293581724, + 0.002754034474492073, + 0.003623235272243619, + 0.0025065764784812927, + 0.01130432728677988, + 0.006248929537832737, + 0.009661351330578327, + 0.005290396511554718, + 0.0052516283467411995, + 0.006025230977684259, + 0.012711924500763416, + 0.006894790101796389, + 0.0024074800312519073, + 0.0028765792958438396, + 0.0030204292852431536, + 0.003196355188265443, + 0.003536656266078353, + 0.0029197956901043653, + 0.004049221985042095, + 0.003247638000175357, + 0.008329551666975021, + 0.030692143365740776, + 0.010646050795912743, + 0.01189612876623869, + 0.022645175457000732, + 0.027998318895697594, + 0.010575622320175171, + 0.01896369457244873, + 0.0030571368988603354, + 0.004050195682793856, + 0.005155689083039761, + 0.0035042890813201666, + 0.0025501088239252567, + 0.008880384266376495, + 0.0033180222380906343, + 0.009580200538039207, + 0.009256521239876747, + 0.01289778295904398, + 0.009685478173196316, + 0.010790709406137466, + 0.010985882952809334, + 0.01059398427605629, + 0.020642131567001343, + 0.012391407042741776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.0022800182923674583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305963218212128 + ], + "opposite_logprob_tokens": [ + -3.1680595874786377 + ], + "logprob_reward": -0.04305963218212128, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305963218212128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0815628245472908, + 0.07588633894920349, + 0.3030552566051483, + 0.09236549586057663, + 0.04361807554960251, + 0.1157940924167633, + 0.06533852964639664, + 0.07522910088300705, + 0.060260288417339325, + 0.08688997477293015 + ], + "credit_map": [ + 0.013850003480911255, + 0.016057642176747322, + 0.015042121522128582, + 0.012613993138074875, + 0.006891165394335985, + 0.005939872935414314, + 0.005895267240703106, + 0.005272751674056053, + 0.0114807840436697, + 0.009805998764932156, + 0.013757421635091305, + 0.011675030924379826, + 0.005250696092844009, + 0.008207310922443867, + 0.006896661594510078, + 0.00881243497133255, + 0.02513970620930195, + 0.10472245514392853, + 0.11131447553634644, + 0.02293234132230282, + 0.01028672605752945, + 0.008337048813700676, + 0.009155534207820892, + 0.01116698794066906, + 0.013255277648568153, + 0.01090584322810173, + 0.011164139956235886, + 0.00857427716255188, + 0.010488053783774376, + 0.011187026277184486, + 0.012625765055418015, + 0.014165107160806656, + 0.006941039580851793, + 0.003756202757358551, + 0.006843539886176586, + 0.005599400494247675, + 0.004719312768429518, + 0.004938934464007616, + 0.005410682410001755, + 0.00540896225720644, + 0.025489626452326775, + 0.01151945348829031, + 0.019555971026420593, + 0.008770979940891266, + 0.007449158933013678, + 0.00792237650603056, + 0.021614739671349525, + 0.013471785932779312, + 0.005955413915216923, + 0.006217560730874538, + 0.008897551335394382, + 0.00728342542424798, + 0.009217506274580956, + 0.006017966195940971, + 0.010654662735760212, + 0.011094445362687111, + 0.004292789846658707, + 0.006637388374656439, + 0.006534748710691929, + 0.006151637993752956, + 0.006918043363839388, + 0.021500518545508385, + 0.00634743832051754, + 0.016846537590026855, + 0.004071730654686689, + 0.004711441230028868, + 0.010746682062745094, + 0.006876594386994839, + 0.004270873963832855, + 0.010908273048698902, + 0.00785887986421585, + 0.010815812274813652, + 0.022245166823267937, + 0.00824431050568819, + 0.01245785690844059, + 0.007051222492009401, + 0.005537859164178371, + 0.011148529127240181, + 0.007906720973551273, + 0.01229830738157034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.04305963218212128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381145142018795 + ], + "opposite_logprob_tokens": [ + -3.288114547729492 + ], + "logprob_reward": -0.0381145142018795, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381145142018795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16306838393211365, + 0.16897106170654297, + 0.2607288956642151, + 0.06289085000753403, + 0.047885164618492126, + 0.04998788982629776, + 0.04439440369606018, + 0.08885189145803452, + 0.06930442154407501, + 0.043916989117860794 + ], + "credit_map": [ + 0.024859337136149406, + 0.0329865887761116, + 0.0236840657889843, + 0.02333947643637657, + 0.014555700123310089, + 0.0145103232935071, + 0.014304506592452526, + 0.014828368090093136, + 0.022346055135130882, + 0.028552686795592308, + 0.03674750775098801, + 0.03223203495144844, + 0.009626109153032303, + 0.01035130862146616, + 0.015139785595238209, + 0.01397557184100151, + 0.019139718264341354, + 0.09958426654338837, + 0.10297266393899918, + 0.01679334230720997, + 0.006027203984558582, + 0.004652538802474737, + 0.004973410163074732, + 0.006585730705410242, + 0.007069569081068039, + 0.011730491183698177, + 0.007303494494408369, + 0.006507823709398508, + 0.009434841573238373, + 0.006763076409697533, + 0.007186626084148884, + 0.00689493166282773, + 0.007826050743460655, + 0.0042790938168764114, + 0.007287370972335339, + 0.005449189804494381, + 0.0055959271267056465, + 0.004723954480141401, + 0.007049453444778919, + 0.005674123764038086, + 0.006029059644788504, + 0.005299323704093695, + 0.007074684835970402, + 0.004489195998758078, + 0.006846716161817312, + 0.0037562295328825712, + 0.011293615214526653, + 0.005199064034968615, + 0.004108367953449488, + 0.005808738991618156, + 0.004968729801476002, + 0.0043275742791593075, + 0.007561841979622841, + 0.003863533027470112, + 0.007196177262812853, + 0.006559445522725582, + 0.00464437622576952, + 0.015503394417464733, + 0.006120269186794758, + 0.0067285033874213696, + 0.01088706310838461, + 0.023391764611005783, + 0.0049532325938344, + 0.016623280942440033, + 0.005600302014499903, + 0.006762074772268534, + 0.009935569949448109, + 0.006403145380318165, + 0.005522123072296381, + 0.013824481517076492, + 0.006535146851092577, + 0.0147215835750103, + 0.005791621282696724, + 0.004796539433300495, + 0.0043626222759485245, + 0.006004271097481251, + 0.0031244948040694, + 0.005565833766013384, + 0.008377404883503914, + 0.005894204135984182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0381145142018795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630723994225264 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630723994225264, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630723994225264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06938754767179489, + 0.07889147102832794, + 0.3773667514324188, + 0.04407532513141632, + 0.05348173528909683, + 0.11377797275781631, + 0.04473808780312538, + 0.06154482811689377, + 0.06666809320449829, + 0.09006813913583755 + ], + "credit_map": [ + 0.008822469972074032, + 0.010432436130940914, + 0.009520050138235092, + 0.010275446809828281, + 0.008124959655106068, + 0.0070515768602490425, + 0.008304905146360397, + 0.006855702493339777, + 0.009975681081414223, + 0.013643785379827023, + 0.018752064555883408, + 0.01655459590256214, + 0.004046476446092129, + 0.0038661037106066942, + 0.006428948137909174, + 0.005623814649879932, + 0.030877461656928062, + 0.1424141228199005, + 0.15900883078575134, + 0.021125100553035736, + 0.006900100037455559, + 0.0044557577930390835, + 0.004592059180140495, + 0.007993332110345364, + 0.004138806369155645, + 0.007752932608127594, + 0.00514307850971818, + 0.0037325099110603333, + 0.0056654317304492, + 0.00500882463529706, + 0.00528128445148468, + 0.007352454587817192, + 0.01008092425763607, + 0.004308181814849377, + 0.007645256817340851, + 0.007677988149225712, + 0.004908732138574123, + 0.004971467424184084, + 0.0078536756336689, + 0.006035507656633854, + 0.0215445663779974, + 0.010202762670814991, + 0.018831295892596245, + 0.010040802881121635, + 0.010392598807811737, + 0.007839522324502468, + 0.021415624767541885, + 0.013510804623365402, + 0.004167592618614435, + 0.004024982452392578, + 0.006965304724872112, + 0.004704490303993225, + 0.006684626452624798, + 0.005692136473953724, + 0.006920019630342722, + 0.0055789346806705, + 0.0032922590617090464, + 0.0109708895906806, + 0.00366394454613328, + 0.004225434735417366, + 0.007873381488025188, + 0.016523677855730057, + 0.003575350157916546, + 0.011419891379773617, + 0.0045340945944190025, + 0.007437217980623245, + 0.009031894616782665, + 0.00804052036255598, + 0.0060269273817539215, + 0.010897061787545681, + 0.009821988642215729, + 0.010878380388021469, + 0.0053462861105799675, + 0.011033209040760994, + 0.01226640585809946, + 0.010333443991839886, + 0.008814028464257717, + 0.01028781570494175, + 0.021244177594780922, + 0.010742769576609135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.007630723994225264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368588164448738 + ], + "opposite_logprob_tokens": [ + -3.4086859226226807 + ], + "logprob_reward": -0.03368588164448738, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368588164448738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0700308158993721, + 0.23748879134655, + 0.5252061486244202, + 0.049937594681978226, + 0.021160516887903214, + 0.05421842262148857, + 0.005179690662771463, + 0.006797783076763153, + 0.009781374596059322, + 0.020198838785290718 + ], + "credit_map": [ + 0.009835614822804928, + 0.011340699158608913, + 0.010753048583865166, + 0.011120722629129887, + 0.006579212378710508, + 0.007249944377690554, + 0.006294258404523134, + 0.00685732439160347, + 0.03708025440573692, + 0.037274524569511414, + 0.053511351346969604, + 0.04257699102163315, + 0.011446892283856869, + 0.014801064506173134, + 0.024774808436632156, + 0.0160229355096817, + 0.0409102663397789, + 0.1845695972442627, + 0.19945977628231049, + 0.04736412316560745, + 0.013196365907788277, + 0.010052700527012348, + 0.011785583570599556, + 0.017867792397737503, + 0.007616711780428886, + 0.005714047234505415, + 0.00679626502096653, + 0.003562144236639142, + 0.004746744874864817, + 0.0056023020297288895, + 0.00758127449080348, + 0.00831811036914587, + 0.004089302849024534, + 0.0013633181806653738, + 0.0035414870362728834, + 0.0033766007982194424, + 0.0020233886316418648, + 0.0015029733767732978, + 0.0022599371150135994, + 0.003003510646522045, + 0.010744478553533554, + 0.004925333429127932, + 0.008585064671933651, + 0.004483561497181654, + 0.0039563472382724285, + 0.003935577347874641, + 0.010940545238554478, + 0.006647520232945681, + 0.0005262552876956761, + 0.0005318968323990703, + 0.0008314484148286283, + 0.0004788265796378255, + 0.0006047413917258382, + 0.0004986737621948123, + 0.0007615549257025123, + 0.0009462941088713706, + 0.00046701700193807483, + 0.000678291660733521, + 0.0006443278980441391, + 0.0006922038737684488, + 0.0007979057845659554, + 0.001461577252484858, + 0.0007089179707691073, + 0.0013475421583279967, + 0.0004310435615479946, + 0.0006269291043281555, + 0.0008902949630282819, + 0.0007466839742846787, + 0.0006509682862088084, + 0.002857295563444495, + 0.0007294761016964912, + 0.0028486838564276695, + 0.0015442498261108994, + 0.0024242436047643423, + 0.0027504705358296633, + 0.0025181507226079702, + 0.0015300217783078551, + 0.0026188353076577187, + 0.004185558762401342, + 0.002627311274409294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03368588164448738 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle depicts the camera in a slow, smooth orbital rotation around a suspended red paper lantern. The lantern, a full circle, is crafted from thin, translucent rice paper. Its vibrant vermilion color stands out against the inky darkness of the night, its slender bamboo framework faintly visible in the light from within. Inside, a warm candle flame burns steadily, emitting a soft orange glow through the paper, transforming the lantern itself into a luminous object and casting dynamic shadows across its surface. Gently rustled by the night breeze, the lantern sways gently and rhythmically in the air, its golden tassels dangling from its base flowing gracefully, their threads shimmering in the dim light. The combination of the camera's relentless rotation and the lantern's gentle trembling creates a hypnotic visual effect, imbuing the image with a tranquil, contemplative, and classically mysterious atmosphere. The out-of-focus black outlines of the traditional building's eaves in the background further enhance this serene feeling.", + "seed": 25621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_8273_seed_25621_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.032666774559766056, + "vlm_logprob": -0.032666774559766056, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.032666774559766056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004529098514467478 + ], + "opposite_logprob_tokens": [ + -7.754528999328613 + ], + "logprob_reward": -0.004529098514467478, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.004529098514467478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05534251034259796, + 0.07742553949356079, + 0.021836042404174805, + 0.028812022879719734, + 0.0662098303437233, + 0.23279528319835663, + 0.08416067063808441, + 0.08833298087120056, + 0.2765364646911621, + 0.06854871660470963 + ], + "credit_map": [ + 0.008825736120343208, + 0.011186303570866585, + 0.007108040153980255, + 0.00462057301774621, + 0.005340289790183306, + 0.004164198879152536, + 0.007659882307052612, + 0.006437483709305525, + 0.012241619639098644, + 0.009612367488443851, + 0.012721149250864983, + 0.0057199737057089806, + 0.007203905377537012, + 0.014704338274896145, + 0.008319038897752762, + 0.006903139874339104, + 0.004814073443412781, + 0.0024342616088688374, + 0.0035285220947116613, + 0.0022736445534974337, + 0.0017287347000092268, + 0.0018238902557641268, + 0.0023919385857880116, + 0.002840974135324359, + 0.004129777196794748, + 0.0033222720958292484, + 0.004435564391314983, + 0.0034401153679937124, + 0.0037156431935727596, + 0.0028922390192747116, + 0.0037555487360805273, + 0.0031208607833832502, + 0.01534656435251236, + 0.005735075566917658, + 0.01136272307485342, + 0.007380385417491198, + 0.006360282190144062, + 0.005807322449982166, + 0.00673970952630043, + 0.007477764040231705, + 0.021419769152998924, + 0.005950187798589468, + 0.06448706984519958, + 0.007488118950277567, + 0.0090789208188653, + 0.09438474476337433, + 0.005443017929792404, + 0.02454342506825924, + 0.007144215051084757, + 0.009522004052996635, + 0.01196928508579731, + 0.006375479977577925, + 0.0076453848741948605, + 0.023158084601163864, + 0.007183236535638571, + 0.011162970215082169, + 0.0050931223668158054, + 0.013524356298148632, + 0.006001232657581568, + 0.011048195883631706, + 0.011605951935052872, + 0.014278269372880459, + 0.014073461294174194, + 0.012708381749689579, + 0.009438544511795044, + 0.11909116804599762, + 0.006962608080357313, + 0.01286591961979866, + 0.08172973990440369, + 0.017997877672314644, + 0.008436479605734348, + 0.020014099776744843, + 0.00546104833483696, + 0.015794213861227036, + 0.006355470046401024, + 0.008617631159722805, + 0.00866924412548542, + 0.0053650387562811375, + 0.010700942948460579, + 0.007585125509649515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.004529098514467478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809638321399689 + ], + "opposite_logprob_tokens": [ + -3.2880964279174805 + ], + "logprob_reward": -0.03809638321399689, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809638321399689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11884775757789612, + 0.07888051867485046, + 0.02535170316696167, + 0.02995625138282776, + 0.05473960191011429, + 0.28625160455703735, + 0.044166725128889084, + 0.12143438309431076, + 0.18896713852882385, + 0.051404327154159546 + ], + "credit_map": [ + 0.02276267111301422, + 0.01842578500509262, + 0.017121709883213043, + 0.013960815966129303, + 0.011156843043863773, + 0.010444249026477337, + 0.011775840073823929, + 0.01319983322173357, + 0.010558758862316608, + 0.011019550263881683, + 0.012048556469380856, + 0.007168839685618877, + 0.009473776444792747, + 0.01031419076025486, + 0.008792519569396973, + 0.009504321031272411, + 0.0034076962620019913, + 0.0033535778056830168, + 0.003281070152297616, + 0.00254301680251956, + 0.002269650576636195, + 0.003011193824931979, + 0.005299543496221304, + 0.002185951918363571, + 0.004099870566278696, + 0.0032896038610488176, + 0.004534257110208273, + 0.004093265626579523, + 0.003201556857675314, + 0.0033524506725370884, + 0.0039026273880153894, + 0.0034826183691620827, + 0.010551458224654198, + 0.005997772794216871, + 0.01531805470585823, + 0.0042230067774653435, + 0.005114881321787834, + 0.0040075769647955894, + 0.0036754400935024023, + 0.005851411260664463, + 0.030421152710914612, + 0.008444760926067829, + 0.07126426696777344, + 0.008856095373630524, + 0.01246747188270092, + 0.11387699842453003, + 0.006721836049109697, + 0.03419902175664902, + 0.004413594026118517, + 0.005959746893495321, + 0.004502370022237301, + 0.006836097687482834, + 0.004970964044332504, + 0.005127045791596174, + 0.00725806038826704, + 0.005098842550069094, + 0.007139027118682861, + 0.02267824299633503, + 0.009335950016975403, + 0.014083403162658215, + 0.019877873361110687, + 0.010147410444915295, + 0.02268729917705059, + 0.01548517681658268, + 0.007548647467046976, + 0.0628892332315445, + 0.007713614497333765, + 0.011168555356562138, + 0.04449149966239929, + 0.029516171663999557, + 0.009916446171700954, + 0.01572294346988201, + 0.0036573121324181557, + 0.010524401441216469, + 0.004695465788245201, + 0.006942853331565857, + 0.00640503503382206, + 0.005525396205484867, + 0.007286394480615854, + 0.006367462687194347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.03809638321399689 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906170725822449 + ], + "opposite_logprob_tokens": [ + -2.464061737060547 + ], + "logprob_reward": -0.08906170725822449, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906170725822449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2702837884426117, + 0.04812319204211235, + 0.03214377164840698, + 0.052802372723817825, + 0.05016934126615524, + 0.07551410049200058, + 0.08754272758960724, + 0.08238613605499268, + 0.19038109481334686, + 0.11065355688333511 + ], + "credit_map": [ + 0.05391567945480347, + 0.03823527321219444, + 0.038309354335069656, + 0.02853205054998398, + 0.02054719068109989, + 0.021850094199180603, + 0.029028773307800293, + 0.03986535593867302, + 0.006226802710443735, + 0.008732235990464687, + 0.00631903437897563, + 0.005555968265980482, + 0.006637938786298037, + 0.0027551911771297455, + 0.007936622947454453, + 0.003959390800446272, + 0.0036531398072838783, + 0.004936360288411379, + 0.004213317297399044, + 0.0029710486996918917, + 0.0030847599264234304, + 0.005572516471147537, + 0.004651295021176338, + 0.0030613297130912542, + 0.00876819621771574, + 0.0054357475601136684, + 0.008291350677609444, + 0.006076553370803595, + 0.0056178574450314045, + 0.007048888597637415, + 0.0065869917161762714, + 0.004976782947778702, + 0.008471137844026089, + 0.003533685812726617, + 0.007980048656463623, + 0.0058987499214708805, + 0.006020484957844019, + 0.005306067410856485, + 0.005507261957973242, + 0.0074519007466733456, + 0.013846021145582199, + 0.005897327791899443, + 0.008658189326524734, + 0.0060295043513178825, + 0.006213225889950991, + 0.01571641117334366, + 0.006223853211849928, + 0.012929563410580158, + 0.007676606997847557, + 0.010184449143707752, + 0.008367682807147503, + 0.010843836702406406, + 0.013456438668072224, + 0.012268781661987305, + 0.013257907703518867, + 0.011487019248306751, + 0.004197447560727596, + 0.013457190245389938, + 0.0038833972066640854, + 0.008773505687713623, + 0.011735977604985237, + 0.01735171675682068, + 0.012553283013403416, + 0.010433611460030079, + 0.006640762090682983, + 0.07814636081457138, + 0.003965931944549084, + 0.008894668892025948, + 0.052387069910764694, + 0.01512106228619814, + 0.005848477128893137, + 0.01937674917280674, + 0.008069288916885853, + 0.021961549296975136, + 0.011811093427240849, + 0.017362723127007484, + 0.012692508287727833, + 0.008697991259396076, + 0.016921279951930046, + 0.013137112371623516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.08906170725822449 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09748177975416183, + 0.07825255393981934, + 0.028741367161273956, + 0.03655318170785904, + 0.047096285969018936, + 0.3528672456741333, + 0.16022920608520508, + 0.105316661298275, + 0.027026750147342682, + 0.06643495708703995 + ], + "credit_map": [ + 0.018359985202550888, + 0.012441533617675304, + 0.01390924584120512, + 0.010197561234235764, + 0.011140838265419006, + 0.007593702059239149, + 0.010792532935738564, + 0.013046386651694775, + 0.009543820284307003, + 0.011008955538272858, + 0.00932371150702238, + 0.0066954898647964, + 0.00939181074500084, + 0.005992902908474207, + 0.015870893374085426, + 0.010424967855215073, + 0.00337545876391232, + 0.003335351590067148, + 0.003408788936212659, + 0.003001399338245392, + 0.0021602159831672907, + 0.0017020321683958173, + 0.0070369732566177845, + 0.00472114747390151, + 0.0066881366074085236, + 0.003511259565129876, + 0.005149850621819496, + 0.004157309886068106, + 0.0047322725877165794, + 0.0037073555868119, + 0.005148605909198523, + 0.0034583930391818285, + 0.0076104095205664635, + 0.004171066451817751, + 0.010739940218627453, + 0.005548486020416021, + 0.004412595648318529, + 0.00415921863168478, + 0.004305882379412651, + 0.006148688495159149, + 0.04088694602251053, + 0.010120963677763939, + 0.07887548208236694, + 0.009754613041877747, + 0.013354409486055374, + 0.14072152972221375, + 0.008834990672767162, + 0.05031830072402954, + 0.01004208717495203, + 0.015745913609862328, + 0.020029207691550255, + 0.016636887565255165, + 0.023442180827260017, + 0.03762391582131386, + 0.017971068620681763, + 0.018737953156232834, + 0.005401731003075838, + 0.01544694509357214, + 0.005247663240879774, + 0.009965511970221996, + 0.01520927157253027, + 0.025375403463840485, + 0.013972149230539799, + 0.014697980135679245, + 0.0017555828671902418, + 0.003625297686085105, + 0.002253493992611766, + 0.0034749305341392756, + 0.0022495025768876076, + 0.006722899619489908, + 0.00304003176279366, + 0.003905011573806405, + 0.0049098338931798935, + 0.012076104059815407, + 0.007355294190347195, + 0.008575621992349625, + 0.007798442151397467, + 0.00543484091758728, + 0.01151235681027174, + 0.008772463537752628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.01107780635356903 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020568877458572388 + ], + "opposite_logprob_tokens": [ + -3.89556884765625 + ], + "logprob_reward": -0.020568877458572388, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020568877458572388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11469040811061859, + 0.02759026363492012, + 0.017280837520956993, + 0.057848669588565826, + 0.03106326423585415, + 0.15409047901630402, + 0.081076480448246, + 0.17689357697963715, + 0.26630982756614685, + 0.0731561928987503 + ], + "credit_map": [ + 0.020149143412709236, + 0.01239427924156189, + 0.017695313319563866, + 0.008229495957493782, + 0.013772047124803066, + 0.008691365830600262, + 0.017149288207292557, + 0.016609471291303635, + 0.004049427807331085, + 0.003948464058339596, + 0.004043286200612783, + 0.0022832064423710108, + 0.002574762562289834, + 0.003625798039138317, + 0.003685358678922057, + 0.003379960311576724, + 0.0018425636226311326, + 0.0027892431244254112, + 0.002167588798329234, + 0.0020895882043987513, + 0.0016748643247410655, + 0.0017708645900711417, + 0.0032365114893764257, + 0.0017096125520765781, + 0.01150739099830389, + 0.00629251915961504, + 0.009648270905017853, + 0.005848389118909836, + 0.006136773154139519, + 0.007877176627516747, + 0.005307053681463003, + 0.005231097340583801, + 0.008574151434004307, + 0.002891162410378456, + 0.004528217017650604, + 0.0036990419030189514, + 0.0032488631550222635, + 0.0025712696369737387, + 0.0025996093172580004, + 0.002950949128717184, + 0.017117369920015335, + 0.005178034771233797, + 0.03391321375966072, + 0.0042168558575212955, + 0.005841135047376156, + 0.06165940314531326, + 0.0035926084965467453, + 0.02257186360657215, + 0.005866548977792263, + 0.008415967226028442, + 0.00914341676980257, + 0.009333543479442596, + 0.010427559725940228, + 0.017863702028989792, + 0.010691153816878796, + 0.009334586560726166, + 0.010219486430287361, + 0.010701502673327923, + 0.011464856564998627, + 0.010149515233933926, + 0.01061878353357315, + 0.07437041401863098, + 0.00975880678743124, + 0.03961022570729256, + 0.010660632513463497, + 0.0939391627907753, + 0.006433771923184395, + 0.014205174520611763, + 0.06453928351402283, + 0.03918557986617088, + 0.012706637382507324, + 0.024639589712023735, + 0.006681730970740318, + 0.014805278740823269, + 0.007578981574624777, + 0.010179898701608181, + 0.008067493326961994, + 0.005943616386502981, + 0.01127687655389309, + 0.008622322231531143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.020568877458572388 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle depicts the camera in a slow, smooth orbital rotation around a suspended red paper lantern. The lantern, a full circle, is crafted from thin, translucent rice paper. Its vibrant vermilion color stands out against the inky darkness of the night, its slender bamboo framework faintly visible in the light from within. Inside, a warm candle flame burns steadily, emitting a soft orange glow through the paper, transforming the lantern itself into a luminous object and casting dynamic shadows across its surface. Gently rustled by the night breeze, the lantern sways gently and rhythmically in the air, its golden tassels dangling from its base flowing gracefully, their threads shimmering in the dim light. The combination of the camera's relentless rotation and the lantern's gentle trembling creates a hypnotic visual effect, imbuing the image with a tranquil, contemplative, and classically mysterious atmosphere. The out-of-focus black outlines of the traditional building's eaves in the background further enhance this serene feeling.", + "seed": 25622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_10/index_8273_seed_25622_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.037529494054615496, + "vlm_logprob": -0.037529494054615496, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.037529494054615496, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00278632715344429 + ], + "opposite_logprob_tokens": [ + -8.127786636352539 + ], + "logprob_reward": -0.00278632715344429, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.00278632715344429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03799283877015114, + 0.01513831689953804, + 0.050431083887815475, + 0.2892417907714844, + 0.016835318878293037, + 0.03310263901948929, + 0.010945555754005909, + 0.02175157330930233, + 0.021845517680048943, + 0.5027153491973877 + ], + "credit_map": [ + 0.008719472214579582, + 0.002889011986553669, + 0.004490554798394442, + 0.0022177977953106165, + 0.0030674885492771864, + 0.0055783819407224655, + 0.0032355599105358124, + 0.007794572506099939, + 0.004115176387131214, + 0.0011598082492128015, + 0.0014563174918293953, + 0.0011029773158952594, + 0.0014560103882104158, + 0.002210717648267746, + 0.001215580152347684, + 0.0024217283353209496, + 0.007257917430251837, + 0.0035227055195719004, + 0.003675463143736124, + 0.0027096157427877188, + 0.002519647590816021, + 0.011580307967960835, + 0.002503348281607032, + 0.016662077978253365, + 0.03889859840273857, + 0.013179493136703968, + 0.023026278242468834, + 0.011726089753210545, + 0.014534013345837593, + 0.0751294493675232, + 0.009428597055375576, + 0.10331928730010986, + 0.002576874103397131, + 0.001202753046527505, + 0.0022472075652331114, + 0.0018871481297537684, + 0.001141305547207594, + 0.0030130266677588224, + 0.0012141573242843151, + 0.0035528468433767557, + 0.005686137825250626, + 0.0029336565639823675, + 0.005247704219073057, + 0.003831474808976054, + 0.0024568052031099796, + 0.0052008372731506824, + 0.0029805495869368315, + 0.004765472374856472, + 0.0009692262392491102, + 0.0010029118275269866, + 0.001912528183311224, + 0.0009614971349947155, + 0.0006818609545007348, + 0.0021215251181274652, + 0.000821813358925283, + 0.0024741922970861197, + 0.0030372769106179476, + 0.001954691717401147, + 0.004397978074848652, + 0.0025630241725593805, + 0.001993719022721052, + 0.003786251414567232, + 0.0012745355488732457, + 0.0027440963312983513, + 0.002930707298219204, + 0.0018795863725245, + 0.002883125329390168, + 0.001486814464442432, + 0.0023768385872244835, + 0.003255206625908613, + 0.0019306152826175094, + 0.005102623254060745, + 0.19708149135112762, + 0.00930998008698225, + 0.21552304923534393, + 0.009053963236510754, + 0.0055322349071502686, + 0.03371259197592735, + 0.008802840486168861, + 0.0236991997808218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.00278632715344429 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494515597820282 + ], + "opposite_logprob_tokens": [ + -2.929945230484009 + ], + "logprob_reward": -0.05494515597820282, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494515597820282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0657922551035881, + 0.028148461133241653, + 0.045432399958372116, + 0.6776058673858643, + 0.027076533064246178, + 0.05087429657578468, + 0.027834221720695496, + 0.03102417290210724, + 0.020168958231806755, + 0.026042908430099487 + ], + "credit_map": [ + 0.01113508827984333, + 0.005142458714544773, + 0.006565490737557411, + 0.0053128814324736595, + 0.004723680671304464, + 0.010658168233931065, + 0.006706492509692907, + 0.01554798148572445, + 0.005736703053116798, + 0.0022078838665038347, + 0.0026567827444523573, + 0.0024445371236652136, + 0.0029600029811263084, + 0.004443922080099583, + 0.002152363071218133, + 0.005546262487769127, + 0.00602125097066164, + 0.006407978944480419, + 0.0047344909980893135, + 0.004123506136238575, + 0.003887018421664834, + 0.007759738247841597, + 0.0024054874666035175, + 0.010092920623719692, + 0.09126600623130798, + 0.03335607796907425, + 0.05440964922308922, + 0.02744803950190544, + 0.035391151905059814, + 0.17316879332065582, + 0.02223629504442215, + 0.2403298169374466, + 0.004208446014672518, + 0.00292975059710443, + 0.0026825368404388428, + 0.003299002768471837, + 0.0024502179585397243, + 0.004525058437138796, + 0.0017542800633236766, + 0.005227237474173307, + 0.004074086435139179, + 0.006536486558616161, + 0.006183550227433443, + 0.00743484403938055, + 0.005175828002393246, + 0.007514437660574913, + 0.00635175546631217, + 0.007603301666676998, + 0.003317382652312517, + 0.0027822046540677547, + 0.004170536063611507, + 0.003216541837900877, + 0.002803757321089506, + 0.0053693149238824844, + 0.002159983152523637, + 0.004014499485492706, + 0.0032466219272464514, + 0.004520720336586237, + 0.006093387026339769, + 0.0032501323148608208, + 0.003604948054999113, + 0.004833021201193333, + 0.0020196049008518457, + 0.0034557320177555084, + 0.002732848981395364, + 0.0012369474861770868, + 0.002577809151262045, + 0.001556099159643054, + 0.0015251495642587543, + 0.00339569547213614, + 0.0013733904343098402, + 0.005771016702055931, + 0.008214344270527363, + 0.0015164449578151107, + 0.00833663810044527, + 0.0010766563937067986, + 0.001952331978827715, + 0.00134899967815727, + 0.001597604830749333, + 0.001999885542318225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.05494515597820282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906040340662003 + ], + "opposite_logprob_tokens": [ + -2.4640603065490723 + ], + "logprob_reward": -0.08906040340662003, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906040340662003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03704546019434929, + 0.035639986395835876, + 0.10787464678287506, + 0.5257524251937866, + 0.12450248003005981, + 0.027409618720412254, + 0.04317623749375343, + 0.034212660044431686, + 0.031913843005895615, + 0.03247267007827759 + ], + "credit_map": [ + 0.006746337283402681, + 0.0031194116454571486, + 0.006187798455357552, + 0.0027263767551630735, + 0.0029465670231729746, + 0.00544806057587266, + 0.003211948787793517, + 0.006658958271145821, + 0.006127974949777126, + 0.004006724338978529, + 0.0030111714731901884, + 0.0033267633989453316, + 0.0036914856173098087, + 0.005176468752324581, + 0.0033912209328264, + 0.006908175069838762, + 0.012782416306436062, + 0.009481999091804028, + 0.008027289994060993, + 0.006666403263807297, + 0.007309970445930958, + 0.024242211133241653, + 0.005479426123201847, + 0.0338849239051342, + 0.0711415633559227, + 0.023968283087015152, + 0.042224377393722534, + 0.021229803562164307, + 0.02702701836824417, + 0.13603036105632782, + 0.017311228439211845, + 0.18681977689266205, + 0.026159372180700302, + 0.009235863573849201, + 0.01118125393986702, + 0.009215823374688625, + 0.00846742745488882, + 0.024071140214800835, + 0.0054440367966890335, + 0.030727555975317955, + 0.004694576840847731, + 0.0017850585281848907, + 0.004183500539511442, + 0.0028881127946078777, + 0.001886670128442347, + 0.004828851670026779, + 0.0018460951978340745, + 0.005296753253787756, + 0.003248587716370821, + 0.004386096727102995, + 0.005590193904936314, + 0.005132726859301329, + 0.003654345404356718, + 0.007942532189190388, + 0.00415943143889308, + 0.009062324650585651, + 0.003029461018741131, + 0.003006116719916463, + 0.005713677033782005, + 0.002872948069125414, + 0.0025823938194662333, + 0.006363782566040754, + 0.0024777755606919527, + 0.00816650316119194, + 0.004229506477713585, + 0.0021341571118682623, + 0.004068177193403244, + 0.002118882490321994, + 0.0024171832483261824, + 0.00641612708568573, + 0.002216037129983306, + 0.008313773199915886, + 0.005029493011534214, + 0.003302263328805566, + 0.006722288206219673, + 0.0035336546134203672, + 0.0034161268267780542, + 0.0025373161770403385, + 0.0042542703449726105, + 0.0036772608291357756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.08906040340662003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074387468397617 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074387468397617, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074387468397617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028018169105052948, + 0.020128926262259483, + 0.055163558572530746, + 0.37516921758651733, + 0.019565653055906296, + 0.02744164690375328, + 0.013518397696316242, + 0.02886696346104145, + 0.02403678186237812, + 0.4080906808376312 + ], + "credit_map": [ + 0.007692242972552776, + 0.0016425381181761622, + 0.0026301387697458267, + 0.001439810497686267, + 0.0021262059453874826, + 0.004698030650615692, + 0.0015514697879552841, + 0.006237734109163284, + 0.002575403079390526, + 0.0018681203946471214, + 0.0015158902388066053, + 0.0016548649873584509, + 0.00178349984344095, + 0.003653197083622217, + 0.0014694244600832462, + 0.005608526989817619, + 0.00805683434009552, + 0.0035807969979941845, + 0.003902596654370427, + 0.0029533179476857185, + 0.002673114649951458, + 0.013028007000684738, + 0.002098220167681575, + 0.018870674073696136, + 0.0501280203461647, + 0.01785203441977501, + 0.030111810192465782, + 0.015785813331604004, + 0.020155729725956917, + 0.09658549726009369, + 0.012474991381168365, + 0.13207535445690155, + 0.002375935669988394, + 0.001645508804358542, + 0.002045318717136979, + 0.0018331189639866352, + 0.001179043552838266, + 0.004399615339934826, + 0.0012260962976142764, + 0.004861018154770136, + 0.004737957380712032, + 0.0020858629141002893, + 0.0023876114282757044, + 0.0022833046969026327, + 0.002490525832399726, + 0.00525039853528142, + 0.0022394463885575533, + 0.005966543685644865, + 0.0013300948776304722, + 0.0017916429787874222, + 0.0017410812433809042, + 0.0017686873907223344, + 0.0011185245821252465, + 0.002287366194650531, + 0.0009789082687348127, + 0.002502093557268381, + 0.0035686418414115906, + 0.0051142810843884945, + 0.003912930842489004, + 0.0031966357491910458, + 0.0026266875211149454, + 0.004575327038764954, + 0.0019933590665459633, + 0.0038791021797806025, + 0.003654173109680414, + 0.0011613526148721576, + 0.0027399437967687845, + 0.0014157792320474982, + 0.0015823275316506624, + 0.004815274849534035, + 0.001595878740772605, + 0.00707205617800355, + 0.15985961258411407, + 0.007573616690933704, + 0.1754651814699173, + 0.007096288725733757, + 0.004614013247191906, + 0.027986308559775352, + 0.007917884737253189, + 0.01757780835032463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.011074387468397617 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suspended red paper lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, smooth orbital rotation around the lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lantern crafted from thin, translucent rice paper with a slender bamboo framework visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lantern sway gently and rhythmically while its golden tassels flow gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of out-of-focus black outlines of traditional building eaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029781196266412735 + ], + "opposite_logprob_tokens": [ + -3.5297811031341553 + ], + "logprob_reward": -0.029781196266412735, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029781196266412735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07328714430332184, + 0.052734993398189545, + 0.12523090839385986, + 0.5524632930755615, + 0.035076942294836044, + 0.013485557399690151, + 0.03862709924578667, + 0.053566932678222656, + 0.04441109299659729, + 0.011116104200482368 + ], + "credit_map": [ + 0.01912068948149681, + 0.005110751837491989, + 0.007453233934938908, + 0.003908630460500717, + 0.0058587705716490746, + 0.01147326361387968, + 0.007275050971657038, + 0.01308675017207861, + 0.01365222129970789, + 0.0040224348194897175, + 0.0037399535067379475, + 0.0031455818098038435, + 0.004919448867440224, + 0.008714022114872932, + 0.0036977645941078663, + 0.010843566618859768, + 0.016710830852389336, + 0.011222104541957378, + 0.008523552678525448, + 0.0069374097511172295, + 0.006828413810580969, + 0.02868272364139557, + 0.005372114013880491, + 0.04095376655459404, + 0.07505064457654953, + 0.026402635499835014, + 0.04450753703713417, + 0.022770952433347702, + 0.02917802706360817, + 0.14193494617938995, + 0.018116887658834457, + 0.19450163841247559, + 0.0049481140449643135, + 0.002080768346786499, + 0.0037683260161429644, + 0.0019224057905375957, + 0.0018426495371386409, + 0.007769144605845213, + 0.0016117938794195652, + 0.011133742518723011, + 0.0030099828727543354, + 0.001099984161555767, + 0.0017100414261221886, + 0.0011838936479762197, + 0.000808103708550334, + 0.002373439259827137, + 0.0006582997157238424, + 0.002641812665387988, + 0.003539741039276123, + 0.0023258759174495935, + 0.0051706223748624325, + 0.002946457825601101, + 0.0023988529574126005, + 0.008401475846767426, + 0.0029121132101863623, + 0.010931961238384247, + 0.005609785672277212, + 0.0055226837284862995, + 0.009906225837767124, + 0.005116513930261135, + 0.005904160439968109, + 0.008915211074054241, + 0.0036039871629327536, + 0.008988365530967712, + 0.006513006053864956, + 0.002291515003889799, + 0.005590174347162247, + 0.002264097100123763, + 0.0027606221847236156, + 0.009114861488342285, + 0.002613671822473407, + 0.013263145461678505, + 0.0036438305396586657, + 0.0005408725701272488, + 0.004023500718176365, + 0.0003599723568186164, + 0.0005403872346505523, + 0.0007362302276305854, + 0.0004031058633700013, + 0.0008682049810886383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.029781196266412735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_11/index_tensor([3058, 3058])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_11/index_tensor([3058, 3058])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9c4f882f5c50e6fe922e3d4f73d2300296c45abf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_11/index_tensor([3058, 3058])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An unsettling overhead shot, framed by a Dutch tilt, captures a tragedy of cosmic proportions in a slow, almost imperceptible drift: a planet, pushed too close to its Roche limit, is being ripped apart by the immense forces of gravity. This former world, its shattered continents and seething ocean core, is violently yet gracefully disintegrating. Vast fragments of its crust snap like bones, while its fiery molten core belches searing streams of material into the inky void. Rather than dispersing, the devastated remains converge into a magnificent ring against the backdrop of the slowly rotating Milky Way, reflecting the cool moonlight of distant stars. Countless fragments, gleaming with metallic luster and icy brilliance, form a cosmic river of devastation. The scene is a scene of stark contrasts: the outer shells of the fragments reflect a cool, silvery glow, while the fractured interiors blaze with a hellish crimson and orange brilliance. The distant Milky Way background is not a simple star point, but a churning nebula filled with dust clouds of ochre, sulfur yellow and deep purple. Ultimately, this entire grand scene of destruction and rebirth is given the kind of sacred, grotesque, and intricate surreal texture in Hieronymus Bosch's paintings, creating an awe-inspiring, fear-filled, and allegorical apocalyptic atmosphere.", + "seed": 8419, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_11/index_3058_seed_8419_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07309735734015703, + "vlm_logprob": -0.07309735734015703, + "vlm_yes_no_margin": 3.349999952316284, + "vlm_token_credit": -0.07309735734015703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18031945824623108 + ], + "opposite_logprob_tokens": [ + -1.9303194284439087 + ], + "logprob_reward": -0.18031945824623108, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.18031945824623108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6103572249412537, + 0.06817840039730072, + 0.059791985899209976, + 0.032718244940042496, + 0.015894554555416107, + 0.1117638349533081, + 0.019291501492261887, + 0.01598648726940155, + 0.018216615542769432, + 0.04780108854174614 + ], + "credit_map": [ + 0.0285564586520195, + 0.010117287747561932, + 0.021876823157072067, + 0.01576491817831993, + 0.020548343658447266, + 0.2538459300994873, + 0.03645588830113411, + 0.22319164872169495, + 0.006381420884281397, + 0.005991523619741201, + 0.0047627827152609825, + 0.008288337849080563, + 0.00844735186547041, + 0.013951221480965614, + 0.008904920890927315, + 0.011450846679508686, + 0.006870724260807037, + 0.007890590466558933, + 0.007367386482656002, + 0.011669117026031017, + 0.004816425498574972, + 0.005741154309362173, + 0.008827961049973965, + 0.006608634255826473, + 0.002946981228888035, + 0.00322322896681726, + 0.003988799639046192, + 0.004273486789315939, + 0.0035230284556746483, + 0.003920549992471933, + 0.005068423226475716, + 0.005773751996457577, + 0.0018861843273043633, + 0.001828096923418343, + 0.0035857725888490677, + 0.0023760285694152117, + 0.0014047020813450217, + 0.0015629589324817061, + 0.0013605012791231275, + 0.0018903105519711971, + 0.010414770804345608, + 0.03239769861102104, + 0.005427111871540546, + 0.01099317241460085, + 0.007497628219425678, + 0.029562057927250862, + 0.0072691114619374275, + 0.00820228923112154, + 0.0026720676105469465, + 0.0020664562471210957, + 0.0027964157052338123, + 0.0019431253895163536, + 0.0018311216263100505, + 0.003881566459313035, + 0.0015813263598829508, + 0.0025194233749061823, + 0.0017546606250107288, + 0.004217813722789288, + 0.002517745830118656, + 0.0016386251663789153, + 0.0016837759176269174, + 0.0010896025924012065, + 0.00079666159581393, + 0.002287602750584483, + 0.002537819556891918, + 0.0017719948664307594, + 0.0015328471781685948, + 0.0018275958718732, + 0.0029744072817265987, + 0.002963834907859564, + 0.0029932826291769743, + 0.0016148366266861558, + 0.008003014139831066, + 0.006057295482605696, + 0.008002267219126225, + 0.005594954825937748, + 0.004503952339291573, + 0.006280532572418451, + 0.004151121713221073, + 0.005207951180636883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.18031945824623108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01823541522026062 + ], + "opposite_logprob_tokens": [ + -4.018235206604004 + ], + "logprob_reward": -0.01823541522026062, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01823541522026062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7453929781913757, + 0.03438790515065193, + 0.03353627398610115, + 0.013981904834508896, + 0.021417079493403435, + 0.09112712740898132, + 0.009359686635434628, + 0.013762492686510086, + 0.020403755828738213, + 0.016630718484520912 + ], + "credit_map": [ + 0.037488605827093124, + 0.011835074052214622, + 0.027617158368229866, + 0.018892107531428337, + 0.02514941431581974, + 0.3133985996246338, + 0.044733766466379166, + 0.26627835631370544, + 0.002676866017282009, + 0.003005880629643798, + 0.0036155786365270615, + 0.005734358914196491, + 0.0034498567692935467, + 0.004226670600473881, + 0.005333288572728634, + 0.006345408968627453, + 0.003686870215460658, + 0.005027574021369219, + 0.0046721878461539745, + 0.005804361775517464, + 0.002373757306486368, + 0.003381423419341445, + 0.005098261404782534, + 0.0034918386954814196, + 0.0011623604223132133, + 0.0015122588956728578, + 0.001741030253469944, + 0.0023529059253633022, + 0.0012604257790371776, + 0.001550533575937152, + 0.0022418913431465626, + 0.002160499570891261, + 0.0025323876179754734, + 0.002227464923635125, + 0.004740857984870672, + 0.0033042621798813343, + 0.002098885364830494, + 0.002155604539439082, + 0.0018472326919436455, + 0.002510383725166321, + 0.008090365678071976, + 0.027920318767428398, + 0.005649827886372805, + 0.006816293578594923, + 0.006447372026741505, + 0.023793144151568413, + 0.006661639548838139, + 0.00574817368760705, + 0.0008512157364748418, + 0.000708777632098645, + 0.0007973915198817849, + 0.0011460748501121998, + 0.000668220454826951, + 0.0029758636374026537, + 0.0007774862460792065, + 0.0014346574898809195, + 0.0009805135196074843, + 0.002628641203045845, + 0.002142823301255703, + 0.0015777692897245288, + 0.0018422440625727177, + 0.001174833858385682, + 0.0017722691409289837, + 0.0016433994751423597, + 0.001887601800262928, + 0.002525589196011424, + 0.002115078968927264, + 0.003339902963489294, + 0.0035202978178858757, + 0.0024671130813658237, + 0.0027573544066399336, + 0.001790819107554853, + 0.0016526342369616032, + 0.0020385412499308586, + 0.0015352481277659535, + 0.0019147899001836777, + 0.002059799386188388, + 0.002455866662785411, + 0.002534214174374938, + 0.0024396267253905535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01823541522026062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200883641839027 + ], + "opposite_logprob_tokens": [ + -4.018200874328613 + ], + "logprob_reward": -0.018200883641839027, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018200883641839027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8579140305519104, + 0.02136547490954399, + 0.013751343823969364, + 0.008647625334560871, + 0.004960136488080025, + 0.0746164619922638, + 0.002775786444544792, + 0.002835863037034869, + 0.0030237212777137756, + 0.010109556838870049 + ], + "credit_map": [ + 0.047049447894096375, + 0.014283477328717709, + 0.03501738980412483, + 0.022679107263684273, + 0.030272267758846283, + 0.3620099127292633, + 0.052689798176288605, + 0.2939126193523407, + 0.0017842019442468882, + 0.0019386418862268329, + 0.002789866179227829, + 0.002696387469768524, + 0.0020581462886184454, + 0.0025174140464514494, + 0.0033833645284175873, + 0.004197451751679182, + 0.0015537285944446921, + 0.0016464455984532833, + 0.0017740066396072507, + 0.002635430311784148, + 0.0012564636999741197, + 0.0014338549226522446, + 0.0017890941817313433, + 0.0016623190604150295, + 0.0008244975470006466, + 0.0008467838633805513, + 0.0012711979215964675, + 0.001287307241000235, + 0.0010623129783198237, + 0.0008964500157162547, + 0.0010670730844140053, + 0.001392001868225634, + 0.0004785810597240925, + 0.0005606889026239514, + 0.0009263634565286338, + 0.0006033561076037586, + 0.000555961043573916, + 0.0005597967538051307, + 0.0004860468616243452, + 0.0007893419242464006, + 0.006617393344640732, + 0.023384442552924156, + 0.00335703045129776, + 0.0056283315643668175, + 0.005673824809491634, + 0.02066654898226261, + 0.0044246758334338665, + 0.004864201880991459, + 0.00034057642915286124, + 0.0002495137741789222, + 0.0002905299770645797, + 0.00029400514904409647, + 0.00021852977806702256, + 0.0006972390692681074, + 0.00023586100724060088, + 0.0004495311004575342, + 0.0003630805585999042, + 0.00031261888216249645, + 0.0006485578487627208, + 0.0002975200768560171, + 0.0002672781120054424, + 0.0002727990213315934, + 0.0001916310575325042, + 0.00048237733426503837, + 0.0003246655687689781, + 0.0003237180062569678, + 0.0003838571719825268, + 0.00044657179387286305, + 0.0004651565395761281, + 0.0003725489950738847, + 0.0004211163613945246, + 0.0002860868407879025, + 0.0014909108867868781, + 0.001239013741724193, + 0.0014718477614223957, + 0.0012919590808451176, + 0.0011294435244053602, + 0.0012281605741009116, + 0.0011711883125826716, + 0.0010870324913412333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.018200883641839027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595568772405386 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.00595568772405386, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595568772405386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5674751400947571, + 0.04154568165540695, + 0.043888166546821594, + 0.035689178854227066, + 0.014478585682809353, + 0.2029648721218109, + 0.008820366114377975, + 0.02465079538524151, + 0.020390590652823448, + 0.04009659215807915 + ], + "credit_map": [ + 0.040113966912031174, + 0.012913241982460022, + 0.03736366704106331, + 0.022225396707654, + 0.023303840309381485, + 0.2319084107875824, + 0.03744184225797653, + 0.162204772233963, + 0.0028511446435004473, + 0.0035477755591273308, + 0.0031008466612547636, + 0.004997065756469965, + 0.004317739512771368, + 0.007551414426416159, + 0.007775282487273216, + 0.00740441121160984, + 0.004799936898052692, + 0.00558056915178895, + 0.005895510781556368, + 0.007895388640463352, + 0.0034635604824870825, + 0.004719324875622988, + 0.0061038341373205185, + 0.005430040415376425, + 0.0032438996713608503, + 0.003379375208169222, + 0.004935829434543848, + 0.005575745832175016, + 0.0035587376914918423, + 0.003873230656608939, + 0.00514264078810811, + 0.005979720503091812, + 0.0016825294587761164, + 0.0012673422461375594, + 0.0026181815192103386, + 0.002101671416312456, + 0.0016120894579216838, + 0.0014797275653108954, + 0.0017154280794784427, + 0.002001615706831217, + 0.021185612305998802, + 0.05655331537127495, + 0.010432529263198376, + 0.01575397327542305, + 0.015576310455799103, + 0.0531279630959034, + 0.01296992413699627, + 0.01736525259912014, + 0.0006842787261120975, + 0.0007046734681352973, + 0.0005708734970539808, + 0.001038147951476276, + 0.0006828696932643652, + 0.002618639264255762, + 0.0007317137788049877, + 0.00178917043376714, + 0.0025894560385495424, + 0.005797822959721088, + 0.0054459357634186745, + 0.0020728670060634613, + 0.002765458542853594, + 0.002148966770619154, + 0.0019721679855138063, + 0.0018581219483166933, + 0.0028187334537506104, + 0.002262961585074663, + 0.0022829666268080473, + 0.002881751162931323, + 0.0030577119905501604, + 0.0023241378366947174, + 0.002507706405594945, + 0.0022546222899109125, + 0.0066587235778570175, + 0.0054061212576925755, + 0.008771137334406376, + 0.0044755213893949986, + 0.0030192583799362183, + 0.005120048765093088, + 0.0028479229658842087, + 0.0037978605832904577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.00595568772405386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277534186840057 + ], + "opposite_logprob_tokens": [ + -2.017775297164917 + ], + "logprob_reward": -0.14277534186840057, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14277534186840057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6269586682319641, + 0.014637600630521774, + 0.005871853791177273, + 0.006913910619914532, + 0.02066788449883461, + 0.2966514229774475, + 0.0050027924589812756, + 0.007489040959626436, + 0.009138965979218483, + 0.006667805835604668 + ], + "credit_map": [ + 0.031244732439517975, + 0.011482948437333107, + 0.023122457787394524, + 0.014611376449465752, + 0.022221539169549942, + 0.26048514246940613, + 0.03724408894777298, + 0.22654642164707184, + 0.0010848845122382045, + 0.0015002103755250573, + 0.002183767268434167, + 0.0013933030422776937, + 0.0011111933272331953, + 0.0025706524029374123, + 0.0024039566051214933, + 0.0023896333295851946, + 0.0006035169935785234, + 0.0006440501310862601, + 0.0007302337908186018, + 0.0007975861662998796, + 0.000520144822075963, + 0.000671251502353698, + 0.0008581645088270307, + 0.0010469064582139254, + 0.0006120337056927383, + 0.0007129345322027802, + 0.000911123410332948, + 0.0009051717352122068, + 0.0008679187158122659, + 0.0007853159913793206, + 0.0009419001289643347, + 0.0011775135062634945, + 0.002655619289726019, + 0.0017295476282015443, + 0.004475717432796955, + 0.003516184166073799, + 0.0018140767933800817, + 0.002153108362108469, + 0.001750230323523283, + 0.00257340120151639, + 0.028871363028883934, + 0.08506334573030472, + 0.014306748285889626, + 0.02361232414841652, + 0.02324836514890194, + 0.08007260411977768, + 0.01939425989985466, + 0.022082414478063583, + 0.0004092188901267946, + 0.00043101265328004956, + 0.0006027839845046401, + 0.0005013629561290145, + 0.0004145060374867171, + 0.0011822390370070934, + 0.0004462221113499254, + 0.0010154475457966328, + 0.001052847015671432, + 0.0014383886009454727, + 0.0014946268638595939, + 0.0009394147782586515, + 0.0006499107694253325, + 0.0006629792042076588, + 0.0004096686898265034, + 0.0008412059978581965, + 0.0009442177833989263, + 0.0010938914492726326, + 0.0008575960528105497, + 0.0011152769438922405, + 0.0014393551973626018, + 0.0013043865328654647, + 0.0014881102833896875, + 0.0008961324929259717, + 0.0009814283112064004, + 0.0007189673488028347, + 0.0011483350535854697, + 0.0012038559652864933, + 0.0005528000765480101, + 0.0006573669379577041, + 0.0006908208597451448, + 0.0007142316899262369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.14277534186840057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsettling overhead shot, framed by a Dutch tilt, captures a tragedy of cosmic proportions in a slow, almost imperceptible drift: a planet, pushed too close to its Roche limit, is being ripped apart by the immense forces of gravity. This former world, its shattered continents and seething ocean core, is violently yet gracefully disintegrating. Vast fragments of its crust snap like bones, while its fiery molten core belches searing streams of material into the inky void. Rather than dispersing, the devastated remains converge into a magnificent ring against the backdrop of the slowly rotating Milky Way, reflecting the cool moonlight of distant stars. Countless fragments, gleaming with metallic luster and icy brilliance, form a cosmic river of devastation. The scene is a scene of stark contrasts: the outer shells of the fragments reflect a cool, silvery glow, while the fractured interiors blaze with a hellish crimson and orange brilliance. The distant Milky Way background is not a simple star point, but a churning nebula filled with dust clouds of ochre, sulfur yellow and deep purple. Ultimately, this entire grand scene of destruction and rebirth is given the kind of sacred, grotesque, and intricate surreal texture in Hieronymus Bosch's paintings, creating an awe-inspiring, fear-filled, and allegorical apocalyptic atmosphere.", + "seed": 8420, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_11/index_3058_seed_8420_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09652110934257507, + "vlm_logprob": -0.09652110934257507, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09652110934257507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29619768261909485 + ], + "opposite_logprob_tokens": [ + -1.4211976528167725 + ], + "logprob_reward": -0.29619768261909485, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29619768261909485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.550999641418457, + 0.21650375425815582, + 0.10890383273363113, + 0.026903415098786354, + 0.011622372083365917, + 0.013832751661539078, + 0.010076072067022324, + 0.01584942266345024, + 0.013124469667673111, + 0.03218429163098335 + ], + "credit_map": [ + 0.06164180487394333, + 0.036030497401952744, + 0.026329414919018745, + 0.03080860525369644, + 0.021879378706216812, + 0.19455432891845703, + 0.020703637972474098, + 0.15905199944972992, + 0.019518591463565826, + 0.019425204023718834, + 0.010201193392276764, + 0.07280644029378891, + 0.015857696533203125, + 0.0425444133579731, + 0.015256822109222412, + 0.020893383771181107, + 0.012171902693808079, + 0.009672961197793484, + 0.009730357676744461, + 0.014859496615827084, + 0.007252943702042103, + 0.023016121238470078, + 0.008164728991687298, + 0.02403533086180687, + 0.0031545180827379227, + 0.0021344523411244154, + 0.0037431137170642614, + 0.002486389596015215, + 0.00421813502907753, + 0.0033021201379597187, + 0.0025998041965067387, + 0.005264881532639265, + 0.0014544270234182477, + 0.0008713313145563006, + 0.0015142052434384823, + 0.0012621378991752863, + 0.0015452163061127067, + 0.0012360922992229462, + 0.0012895397376269102, + 0.002449422376230359, + 0.0012782905250787735, + 0.0010645465226843953, + 0.0014438375364989042, + 0.0012981228064745665, + 0.0011508699972182512, + 0.0013734657550230622, + 0.0011973991058766842, + 0.005026220344007015, + 0.0014304063515737653, + 0.0012838816037401557, + 0.0014756224118173122, + 0.001167171634733677, + 0.0011444083647802472, + 0.0010198896052315831, + 0.001416906132362783, + 0.0011377851478755474, + 0.0013652611523866653, + 0.001319774892181158, + 0.001527454936876893, + 0.0023278940934687853, + 0.0020082597620785236, + 0.003169496776536107, + 0.0023988187313079834, + 0.0017324641812592745, + 0.0016607912257313728, + 0.0011288446839898825, + 0.0012876729015260935, + 0.0011297574965283275, + 0.0013374817790463567, + 0.0021106465719640255, + 0.0018716732738539577, + 0.0025976018514484167, + 0.002643825951963663, + 0.0028611361049115658, + 0.005135789979249239, + 0.0027695270255208015, + 0.00406668009236455, + 0.004967159125953913, + 0.004670902621001005, + 0.005069271195679903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.29619768261909485 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062072791159153 + ], + "opposite_logprob_tokens": [ + -3.895620822906494 + ], + "logprob_reward": -0.02062072791159153, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062072791159153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20201264321804047, + 0.164524644613266, + 0.16079570353031158, + 0.11713311821222305, + 0.0767519474029541, + 0.027675537392497063, + 0.046390384435653687, + 0.0919872522354126, + 0.05819111689925194, + 0.05453767627477646 + ], + "credit_map": [ + 0.017419135197997093, + 0.038704417645931244, + 0.021753817796707153, + 0.031063038855791092, + 0.01760798692703247, + 0.028064755722880363, + 0.021575992926955223, + 0.02582348883152008, + 0.0149356285110116, + 0.0188989769667387, + 0.009509041905403137, + 0.04996485635638237, + 0.013599822297692299, + 0.025560526177287102, + 0.011831803247332573, + 0.02022397518157959, + 0.025565261021256447, + 0.021133890375494957, + 0.019139626994729042, + 0.030151614919304848, + 0.014127472415566444, + 0.016444889828562737, + 0.014706246554851532, + 0.01952669955790043, + 0.021421151235699654, + 0.014299089089035988, + 0.013963216915726662, + 0.01246590819209814, + 0.010961310006678104, + 0.013129624538123608, + 0.009561257436871529, + 0.021331559866666794, + 0.010752789676189423, + 0.007585490122437477, + 0.00822383351624012, + 0.009134181775152683, + 0.010275283828377724, + 0.007420166861265898, + 0.007540092337876558, + 0.015820110216736794, + 0.0025660141836851835, + 0.003294867929071188, + 0.003246464068070054, + 0.0031804845202714205, + 0.0030603986233472824, + 0.0026207577902823687, + 0.0031085435766726732, + 0.006598006002604961, + 0.005980282090604305, + 0.004213540814816952, + 0.0063827624544501305, + 0.007256305310875177, + 0.0055788131430745125, + 0.004447542130947113, + 0.00605732062831521, + 0.006473816931247711, + 0.008718467317521572, + 0.008310072124004364, + 0.010403749532997608, + 0.015449014492332935, + 0.010871331207454205, + 0.009818928316235542, + 0.015732312574982643, + 0.012683377601206303, + 0.007097810041159391, + 0.0042661027982831, + 0.005702320020645857, + 0.005457557737827301, + 0.005851335357874632, + 0.00849800556898117, + 0.011554577387869358, + 0.00976340752094984, + 0.005272254813462496, + 0.005931674037128687, + 0.011188715696334839, + 0.004510563798248768, + 0.005860875826328993, + 0.006828516721725464, + 0.005910232663154602, + 0.00903484970331192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02062072791159153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043090347200632095 + ], + "opposite_logprob_tokens": [ + -3.168090343475342 + ], + "logprob_reward": -0.043090347200632095, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043090347200632095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41543930768966675, + 0.161932572722435, + 0.1129148080945015, + 0.08913974463939667, + 0.03977135196328163, + 0.022974276915192604, + 0.03419283404946327, + 0.03536199778318405, + 0.04179706051945686, + 0.046476054936647415 + ], + "credit_map": [ + 0.05404496565461159, + 0.02634863741695881, + 0.023382732644677162, + 0.02453450672328472, + 0.016655832529067993, + 0.1444934457540512, + 0.017289720475673676, + 0.10868947207927704, + 0.01487133651971817, + 0.01508407574146986, + 0.009370379149913788, + 0.05633585527539253, + 0.01251316349953413, + 0.02655535750091076, + 0.012751265428960323, + 0.014451127499341965, + 0.008675368502736092, + 0.0056096212938427925, + 0.007863896898925304, + 0.008887522853910923, + 0.007038799114525318, + 0.033302921801805496, + 0.00864134170114994, + 0.03289533406496048, + 0.014707867056131363, + 0.008880313485860825, + 0.010925696231424809, + 0.008942239917814732, + 0.010216432623565197, + 0.010701441206037998, + 0.006201289594173431, + 0.01856447197496891, + 0.0032922346144914627, + 0.00253491778858006, + 0.005018926225602627, + 0.003304452635347843, + 0.004988160450011492, + 0.004704964812844992, + 0.004644771572202444, + 0.011282925494015217, + 0.0013223965652287006, + 0.0011305270018056035, + 0.0024593023117631674, + 0.0018205010564997792, + 0.0024077477864921093, + 0.002697044052183628, + 0.0023031895980238914, + 0.00883356761187315, + 0.006249495781958103, + 0.003668467979878187, + 0.003737499238923192, + 0.005877456162124872, + 0.004204881843179464, + 0.002040751278400421, + 0.004968544002622366, + 0.003445737762376666, + 0.003946409095078707, + 0.00313163991086185, + 0.003996227402240038, + 0.0051833027973771095, + 0.004073680844157934, + 0.005350267980247736, + 0.005131823476403952, + 0.004548646509647369, + 0.0030894544906914234, + 0.0027342187240719795, + 0.0032444270327687263, + 0.0038169350009411573, + 0.004732425790280104, + 0.008721519261598587, + 0.008383962325751781, + 0.007074115797877312, + 0.00260467198677361, + 0.0027622354682534933, + 0.0036674828734248877, + 0.0029483302496373653, + 0.005867326632142067, + 0.015181885100901127, + 0.004706467501819134, + 0.008737655356526375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.043090347200632095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097875464707613 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097875464707613, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097875464707613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16135236620903015, + 0.060831937938928604, + 0.33760759234428406, + 0.05651739239692688, + 0.07178570330142975, + 0.022565999999642372, + 0.09788618981838226, + 0.027501067146658897, + 0.05381827801465988, + 0.11013344675302505 + ], + "credit_map": [ + 0.015431737527251244, + 0.00997083354741335, + 0.008597088977694511, + 0.008955803699791431, + 0.006099310237914324, + 0.05599356070160866, + 0.006095748394727707, + 0.050208285450935364, + 0.005901021417230368, + 0.008704524487257004, + 0.005976424552500248, + 0.008116519078612328, + 0.0052154818549752235, + 0.00984051264822483, + 0.005456437822431326, + 0.011621016077697277, + 0.04445990175008774, + 0.03278481587767601, + 0.03216593340039253, + 0.049311380833387375, + 0.030674010515213013, + 0.05868293717503548, + 0.032123830169439316, + 0.057404760271310806, + 0.008797671645879745, + 0.005510764662176371, + 0.007390302140265703, + 0.006748524494469166, + 0.007137005683034658, + 0.0067601571790874004, + 0.006175410468131304, + 0.00799755472689867, + 0.010134333744645119, + 0.005860795266926289, + 0.00991013552993536, + 0.007958552800118923, + 0.00725593464449048, + 0.007254971656948328, + 0.007891888730227947, + 0.015519090928137302, + 0.002233366249129176, + 0.002129610860720277, + 0.0023054461926221848, + 0.002060845261439681, + 0.002535868203267455, + 0.002479935297742486, + 0.001877769478596747, + 0.006943159271031618, + 0.014741561375558376, + 0.011042412370443344, + 0.012465009465813637, + 0.01718362420797348, + 0.008268388919532299, + 0.006955193355679512, + 0.016515925526618958, + 0.010714072734117508, + 0.002857488114386797, + 0.002464606426656246, + 0.0027381249237805605, + 0.005008178763091564, + 0.003605295903980732, + 0.003577386261895299, + 0.004295869264751673, + 0.0029541177209466696, + 0.003222860861569643, + 0.0032776573207229376, + 0.0033640218898653984, + 0.0051771486178040504, + 0.007538228295743465, + 0.011441175825893879, + 0.011920402757823467, + 0.007876783609390259, + 0.005642556119710207, + 0.006798743270337582, + 0.007759192492812872, + 0.008570097386837006, + 0.012023483403027058, + 0.03254609927535057, + 0.018269075080752373, + 0.018524199724197388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0097875464707613 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129092425107956 + ], + "opposite_logprob_tokens": [ + -2.2379093170166016 + ], + "logprob_reward": -0.1129092425107956, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129092425107956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3346914052963257, + 0.17889836430549622, + 0.06140461564064026, + 0.07298235595226288, + 0.22036130726337433, + 0.01857728697359562, + 0.019004009664058685, + 0.012027325108647346, + 0.02214651368558407, + 0.05990682169795036 + ], + "credit_map": [ + 0.03998255357146263, + 0.027374355122447014, + 0.018768716603517532, + 0.02270938642323017, + 0.015741748735308647, + 0.11040528863668442, + 0.01531059667468071, + 0.08439875394105911, + 0.015739474445581436, + 0.01782624050974846, + 0.008023292757570744, + 0.05804595351219177, + 0.013039140962064266, + 0.03709224984049797, + 0.01246531680226326, + 0.016666673123836517, + 0.005545370746403933, + 0.00457919854670763, + 0.003921465948224068, + 0.005148065276443958, + 0.004340480547398329, + 0.01615898311138153, + 0.005259234458208084, + 0.016451815143227577, + 0.0065038856118917465, + 0.005328397266566753, + 0.00798435416072607, + 0.007324208039790392, + 0.016350289806723595, + 0.008754818700253963, + 0.00813655462116003, + 0.012599839828908443, + 0.01643688790500164, + 0.010563990101218224, + 0.032283965498209, + 0.017451956868171692, + 0.0205462034791708, + 0.0243510939180851, + 0.023982126265764236, + 0.07474508136510849, + 0.0019491837592795491, + 0.0013909708941355348, + 0.0022276046220213175, + 0.001945769996382296, + 0.002601775573566556, + 0.0020521332044154406, + 0.0018484770553186536, + 0.004561370704323053, + 0.003988537471741438, + 0.00225175847299397, + 0.0024354937486350536, + 0.002158838091418147, + 0.002077412558719516, + 0.0015744664706289768, + 0.0023037786595523357, + 0.0022137227933853865, + 0.0010621327674016356, + 0.0013519005151465535, + 0.0012989557581022382, + 0.0018107041250914335, + 0.0013073472073301673, + 0.0013194400817155838, + 0.0018454850651323795, + 0.0020313577260822058, + 0.0026221938896924257, + 0.0022775207180529833, + 0.0020870680455118418, + 0.002392987022176385, + 0.002690608147531748, + 0.002972545800730586, + 0.0042823501862585545, + 0.0028212377801537514, + 0.004055369645357132, + 0.0040742685087025166, + 0.004568937234580517, + 0.004085636232048273, + 0.007205641362816095, + 0.015338503755629063, + 0.010690480470657349, + 0.009887981228530407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.1129092425107956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsettling overhead shot, framed by a Dutch tilt, captures a tragedy of cosmic proportions in a slow, almost imperceptible drift: a planet, pushed too close to its Roche limit, is being ripped apart by the immense forces of gravity. This former world, its shattered continents and seething ocean core, is violently yet gracefully disintegrating. Vast fragments of its crust snap like bones, while its fiery molten core belches searing streams of material into the inky void. Rather than dispersing, the devastated remains converge into a magnificent ring against the backdrop of the slowly rotating Milky Way, reflecting the cool moonlight of distant stars. Countless fragments, gleaming with metallic luster and icy brilliance, form a cosmic river of devastation. The scene is a scene of stark contrasts: the outer shells of the fragments reflect a cool, silvery glow, while the fractured interiors blaze with a hellish crimson and orange brilliance. The distant Milky Way background is not a simple star point, but a churning nebula filled with dust clouds of ochre, sulfur yellow and deep purple. Ultimately, this entire grand scene of destruction and rebirth is given the kind of sacred, grotesque, and intricate surreal texture in Hieronymus Bosch's paintings, creating an awe-inspiring, fear-filled, and allegorical apocalyptic atmosphere.", + "seed": 8421, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_11/index_3058_seed_8421_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06097729522734881, + "vlm_logprob": -0.06097729522734881, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.06097729522734881, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11251741647720337 + ], + "opposite_logprob_tokens": [ + -2.3625173568725586 + ], + "logprob_reward": -0.11251741647720337, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11251741647720337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24456702172756195, + 0.1688380390405655, + 0.08520758152008057, + 0.05587877705693245, + 0.0453682616353035, + 0.0759681910276413, + 0.08200584352016449, + 0.11290102452039719, + 0.08757858723402023, + 0.04168665036559105 + ], + "credit_map": [ + 0.02914869412779808, + 0.041621025651693344, + 0.026631463319063187, + 0.02649020403623581, + 0.0342395193874836, + 0.022660721093416214, + 0.034766338765621185, + 0.029009060934185982, + 0.0226362943649292, + 0.012288154102861881, + 0.02268347144126892, + 0.02832367643713951, + 0.01786489225924015, + 0.017314301803708076, + 0.029561392962932587, + 0.01816585473716259, + 0.01462465152144432, + 0.006832331418991089, + 0.010340945795178413, + 0.009399116970598698, + 0.006257995031774044, + 0.015121174976229668, + 0.008624275214970112, + 0.014007091522216797, + 0.0076307631097733974, + 0.0038308838848024607, + 0.006825410760939121, + 0.0057770791463553905, + 0.004932404030114412, + 0.007517306134104729, + 0.0056893168948590755, + 0.01367561612278223, + 0.004237312823534012, + 0.004001758527010679, + 0.008003702387213707, + 0.0055702971294522285, + 0.0038352622650563717, + 0.007816392928361893, + 0.004798604641109705, + 0.007104932330548763, + 0.007757258601486683, + 0.006246980279684067, + 0.00834956206381321, + 0.00945973303169012, + 0.006402524188160896, + 0.008364377543330193, + 0.008769351989030838, + 0.02061840146780014, + 0.011799345724284649, + 0.006894788704812527, + 0.017489012330770493, + 0.006649746093899012, + 0.006303415168076754, + 0.015362934209406376, + 0.00633395416662097, + 0.011172650381922722, + 0.014473017305135727, + 0.010294441133737564, + 0.02093871310353279, + 0.012181700207293034, + 0.009491135366261005, + 0.015135223977267742, + 0.018414229154586792, + 0.011972565203905106, + 0.012762166559696198, + 0.00949569046497345, + 0.01163477823138237, + 0.01091237273067236, + 0.009594415314495564, + 0.01009705476462841, + 0.010766960680484772, + 0.012315146625041962, + 0.006594684440642595, + 0.002798287430778146, + 0.009336243383586407, + 0.004961799364537001, + 0.0029972356278449297, + 0.0045119947753846645, + 0.0047815898433327675, + 0.005704812239855528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.11251741647720337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014227267354726791 + ], + "opposite_logprob_tokens": [ + -4.264227390289307 + ], + "logprob_reward": -0.014227267354726791, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014227267354726791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09685967117547989, + 0.032131534069776535, + 0.06084626540541649, + 0.19456595182418823, + 0.15193098783493042, + 0.11234764754772186, + 0.040257569402456284, + 0.07145421206951141, + 0.11332611739635468, + 0.12627999484539032 + ], + "credit_map": [ + 0.00940961018204689, + 0.015269673429429531, + 0.012630414217710495, + 0.01594448834657669, + 0.01082572340965271, + 0.006670036353170872, + 0.014536131173372269, + 0.011573594994843006, + 0.004417108837515116, + 0.004080126993358135, + 0.004074847791343927, + 0.004062441177666187, + 0.003047443926334381, + 0.0025673385243862867, + 0.0059457444585859776, + 0.003936478402465582, + 0.009565633721649647, + 0.005060945171862841, + 0.006684598047286272, + 0.006663638632744551, + 0.0049148136749863625, + 0.010829687118530273, + 0.006700219586491585, + 0.010426728054881096, + 0.02711152285337448, + 0.012907527387142181, + 0.03809771314263344, + 0.018604958429932594, + 0.015622877515852451, + 0.026954473927617073, + 0.020487023517489433, + 0.034779854118824005, + 0.012761292979121208, + 0.013556228950619698, + 0.02857219986617565, + 0.015339664183557034, + 0.011660168878734112, + 0.02858605608344078, + 0.015189778059720993, + 0.026265591382980347, + 0.009714564308524132, + 0.009339542128145695, + 0.015689373016357422, + 0.0178363099694252, + 0.012217055074870586, + 0.01328838150948286, + 0.016561392694711685, + 0.017701026052236557, + 0.005320622120052576, + 0.004415006842464209, + 0.005674267653375864, + 0.006278506480157375, + 0.0038495634216815233, + 0.004413805436342955, + 0.004782034549862146, + 0.005523761734366417, + 0.009063327684998512, + 0.006348675116896629, + 0.01232876256108284, + 0.007730138022452593, + 0.007365514989942312, + 0.009925900027155876, + 0.010323191992938519, + 0.008368700742721558, + 0.0189800001680851, + 0.007539673708379269, + 0.03262664005160332, + 0.00980591680854559, + 0.009024125523865223, + 0.010744674131274223, + 0.009054692462086678, + 0.015550395473837852, + 0.017492638900876045, + 0.009779419749975204, + 0.020805753767490387, + 0.01722121611237526, + 0.012082724831998348, + 0.016052953898906708, + 0.015565219335258007, + 0.01728007197380066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.014227267354726791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13143640756607056, + 0.12058849632740021, + 0.04310218617320061, + 0.11861000955104828, + 0.16227805614471436, + 0.11785559356212616, + 0.12638355791568756, + 0.08117169141769409, + 0.06104668229818344, + 0.03752730414271355 + ], + "credit_map": [ + 0.012040573172271252, + 0.03187807649374008, + 0.015455957502126694, + 0.015213407576084137, + 0.013582178391516209, + 0.010744375176727772, + 0.01798296719789505, + 0.01453888788819313, + 0.0144095653668046, + 0.010085828602313995, + 0.01724301278591156, + 0.0161935705691576, + 0.014540036208927631, + 0.00970962643623352, + 0.022208349779248238, + 0.016198517754673958, + 0.005329752806574106, + 0.004518932197242975, + 0.0045107570476830006, + 0.004189788363873959, + 0.008977992460131645, + 0.004627699498087168, + 0.00649250065907836, + 0.00445476733148098, + 0.023707853630185127, + 0.010523403994739056, + 0.016225621104240417, + 0.011020975187420845, + 0.009706651791930199, + 0.0158220324665308, + 0.012794174253940582, + 0.018809309229254723, + 0.014322111383080482, + 0.01186016108840704, + 0.041051384061574936, + 0.016430221498012543, + 0.008299553766846657, + 0.020866597071290016, + 0.01419712696224451, + 0.03525090962648392, + 0.010849093087017536, + 0.010294956155121326, + 0.01642710715532303, + 0.01422801986336708, + 0.012976516969501972, + 0.01621905528008938, + 0.01391860656440258, + 0.022942250594496727, + 0.011851447634398937, + 0.008530166000127792, + 0.028379876166582108, + 0.012604907155036926, + 0.00938115268945694, + 0.023036304861307144, + 0.009911201894283295, + 0.022688502445816994, + 0.009553170762956142, + 0.0060159508138895035, + 0.013055196031928062, + 0.011495261453092098, + 0.009447046555578709, + 0.00993659719824791, + 0.008649692870676517, + 0.013018783181905746, + 0.007013808935880661, + 0.007256548386067152, + 0.0071791899390518665, + 0.008042600937187672, + 0.00903991423547268, + 0.0076739974319934845, + 0.007609471213072538, + 0.007231154944747686, + 0.0051910909824073315, + 0.002823351416736841, + 0.006098135840147734, + 0.004305137321352959, + 0.003725252114236355, + 0.006267928518354893, + 0.005041993223130703, + 0.004074420779943466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.01108264084905386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738675758242607 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738675758242607, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738675758242607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1853262037038803, + 0.17173637449741364, + 0.07528650015592575, + 0.11329276114702225, + 0.08711438626050949, + 0.050296757370233536, + 0.051552604883909225, + 0.11725851148366928, + 0.09106633067131042, + 0.05706963688135147 + ], + "credit_map": [ + 0.019587362185120583, + 0.02382800728082657, + 0.024555735290050507, + 0.02189021185040474, + 0.02472250536084175, + 0.01861528865993023, + 0.028192810714244843, + 0.023934295400977135, + 0.02143799141049385, + 0.018153198063373566, + 0.028105225414037704, + 0.02108435519039631, + 0.015143943950533867, + 0.025725632905960083, + 0.024039436131715775, + 0.01804659329354763, + 0.011776858009397984, + 0.006322159431874752, + 0.007501548156142235, + 0.008723357692360878, + 0.005870671942830086, + 0.014629513956606388, + 0.007577758282423019, + 0.012884629890322685, + 0.017250480130314827, + 0.008794184774160385, + 0.013943132013082504, + 0.011298359371721745, + 0.00959897693246603, + 0.016321634873747826, + 0.01218516193330288, + 0.0239008329808712, + 0.009445619769394398, + 0.0076688057743012905, + 0.01758451946079731, + 0.009679533541202545, + 0.007773341611027718, + 0.011885675601661205, + 0.0097114322707057, + 0.01336545217782259, + 0.003606810700148344, + 0.0040216040797531605, + 0.004928318317979574, + 0.0063831377774477005, + 0.005626976024359465, + 0.005878000054508448, + 0.005568195600062609, + 0.01428371760994196, + 0.0067862896248698235, + 0.003781205741688609, + 0.0098677147179842, + 0.004801289178431034, + 0.005417877808213234, + 0.008212472312152386, + 0.005946085322648287, + 0.0067396704107522964, + 0.01243363507091999, + 0.008748874999582767, + 0.021237462759017944, + 0.015009036287665367, + 0.015945708379149437, + 0.01584601029753685, + 0.013620187528431416, + 0.014417596161365509, + 0.014867601916193962, + 0.008127434179186821, + 0.022915523499250412, + 0.00848215539008379, + 0.008842725306749344, + 0.007668045349419117, + 0.007718343287706375, + 0.012444505468010902, + 0.007990925572812557, + 0.005691438913345337, + 0.007693252991884947, + 0.007868514396250248, + 0.005894508678466082, + 0.007005066145211458, + 0.00812236126512289, + 0.006803568452596664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.006738675758242607 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603204756975174 + ], + "opposite_logprob_tokens": [ + -1.910320520401001 + ], + "logprob_reward": -0.1603204756975174, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603204756975174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12561585009098053, + 0.13019846379756927, + 0.03170018643140793, + 0.10814007371664047, + 0.1593993902206421, + 0.11259856820106506, + 0.07300615310668945, + 0.10890424996614456, + 0.08142029494047165, + 0.06901679933071136 + ], + "credit_map": [ + 0.01478579267859459, + 0.013506701216101646, + 0.01547810435295105, + 0.012384487316012383, + 0.01795908436179161, + 0.012601727619767189, + 0.02285730466246605, + 0.016042649745941162, + 0.015085756778717041, + 0.013199592009186745, + 0.018886864185333252, + 0.01924651861190796, + 0.014736957848072052, + 0.013408093713223934, + 0.01936175860464573, + 0.01627291552722454, + 0.002850102260708809, + 0.002508417470380664, + 0.0038279013242572546, + 0.003206016030162573, + 0.0028708954341709614, + 0.004351531621068716, + 0.0034773002844303846, + 0.008608024567365646, + 0.021184084936976433, + 0.0088884886354208, + 0.017104236409068108, + 0.008956138044595718, + 0.008434074930846691, + 0.015285558998584747, + 0.010569358244538307, + 0.017718134447932243, + 0.01646215096116066, + 0.01476243045181036, + 0.03323717042803764, + 0.013689456507563591, + 0.010777998715639114, + 0.028009315952658653, + 0.013797378167510033, + 0.028663482517004013, + 0.01568271964788437, + 0.008714187890291214, + 0.013808580115437508, + 0.013447439298033714, + 0.010544239543378353, + 0.012913493439555168, + 0.012774783186614513, + 0.024713123217225075, + 0.008903680369257927, + 0.006716649979352951, + 0.010821516625583172, + 0.008917934261262417, + 0.007202295586466789, + 0.007592146284878254, + 0.0077387127093970776, + 0.015113217756152153, + 0.01121992152184248, + 0.00837770290672779, + 0.019470475614070892, + 0.012791759334504604, + 0.015764424577355385, + 0.016614263877272606, + 0.012327095493674278, + 0.0123386075720191, + 0.01196103636175394, + 0.0077140070497989655, + 0.010043991729617119, + 0.010548328049480915, + 0.01143084466457367, + 0.007161651272326708, + 0.010349535383284092, + 0.012210904620587826, + 0.009978801012039185, + 0.006383449304848909, + 0.00926325935870409, + 0.007170047145336866, + 0.005324085243046284, + 0.010589071549475193, + 0.010010700672864914, + 0.0102973822504282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.1603204756975174 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsettling overhead shot, framed by a Dutch tilt, captures a tragedy of cosmic proportions in a slow, almost imperceptible drift: a planet, pushed too close to its Roche limit, is being ripped apart by the immense forces of gravity. This former world, its shattered continents and seething ocean core, is violently yet gracefully disintegrating. Vast fragments of its crust snap like bones, while its fiery molten core belches searing streams of material into the inky void. Rather than dispersing, the devastated remains converge into a magnificent ring against the backdrop of the slowly rotating Milky Way, reflecting the cool moonlight of distant stars. Countless fragments, gleaming with metallic luster and icy brilliance, form a cosmic river of devastation. The scene is a scene of stark contrasts: the outer shells of the fragments reflect a cool, silvery glow, while the fractured interiors blaze with a hellish crimson and orange brilliance. The distant Milky Way background is not a simple star point, but a churning nebula filled with dust clouds of ochre, sulfur yellow and deep purple. Ultimately, this entire grand scene of destruction and rebirth is given the kind of sacred, grotesque, and intricate surreal texture in Hieronymus Bosch's paintings, creating an awe-inspiring, fear-filled, and allegorical apocalyptic atmosphere.", + "seed": 8422, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_11/index_3058_seed_8422_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04523648167960346, + "vlm_logprob": -0.04523648167960346, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.04523648167960346, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17253606021404266 + ], + "opposite_logprob_tokens": [ + -1.922536015510559 + ], + "logprob_reward": -0.17253606021404266, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17253606021404266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22031767666339874, + 0.10598532110452652, + 0.1803937405347824, + 0.12477508187294006, + 0.07894965261220932, + 0.023626521229743958, + 0.15935169160366058, + 0.022463809698820114, + 0.035750824958086014, + 0.0483856201171875 + ], + "credit_map": [ + 0.0255996473133564, + 0.024833085015416145, + 0.03084753081202507, + 0.03904159739613533, + 0.01941615343093872, + 0.03814104571938515, + 0.01468432042747736, + 0.027754314243793488, + 0.00956321507692337, + 0.00899618398398161, + 0.015078332275152206, + 0.015648750588297844, + 0.01267934124916792, + 0.017555976286530495, + 0.011792049743235111, + 0.014671480283141136, + 0.0176788829267025, + 0.01786821149289608, + 0.01838645152747631, + 0.029235519468784332, + 0.03142976388335228, + 0.02339763380587101, + 0.024789420887827873, + 0.017607877030968666, + 0.012279353104531765, + 0.011432883329689503, + 0.012885627336800098, + 0.014960365369915962, + 0.011805115267634392, + 0.019352490082383156, + 0.016930250450968742, + 0.02512902207672596, + 0.006704448256641626, + 0.007189487107098103, + 0.00899429526180029, + 0.00848403386771679, + 0.0062111360020935535, + 0.011893605813384056, + 0.016867129132151604, + 0.012605531141161919, + 0.0020421738736331463, + 0.0016238371608778834, + 0.0034670555032789707, + 0.002719075186178088, + 0.004010371398180723, + 0.003090787213295698, + 0.0030610044486820698, + 0.003612220287322998, + 0.009095822460949421, + 0.013631682842969894, + 0.014165048487484455, + 0.013841510750353336, + 0.01696661487221718, + 0.03504372015595436, + 0.04241738095879555, + 0.014189944602549076, + 0.0014107321621850133, + 0.003198648104444146, + 0.0028915381990373135, + 0.004282943904399872, + 0.0028683028649538755, + 0.002669856883585453, + 0.002817310392856598, + 0.0023244814947247505, + 0.00375921162776649, + 0.0032662255689501762, + 0.0037344556767493486, + 0.003954188898205757, + 0.0054578459821641445, + 0.005953254643827677, + 0.005952279083430767, + 0.003673367900773883, + 0.003886748105287552, + 0.004580679815262556, + 0.005894104018807411, + 0.00570039264857769, + 0.006328931078314781, + 0.009496137499809265, + 0.00547004071995616, + 0.007028594147413969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.17253606021404266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005984720308333635 + ], + "opposite_logprob_tokens": [ + -5.130984783172607 + ], + "logprob_reward": -0.005984720308333635, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005984720308333635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2644607722759247, + 0.05230794474482536, + 0.134562149643898, + 0.14487361907958984, + 0.029786165803670883, + 0.06491366773843765, + 0.17601637542247772, + 0.03683982044458389, + 0.04730719327926636, + 0.04893232882022858 + ], + "credit_map": [ + 0.04237903282046318, + 0.03215847909450531, + 0.03553122654557228, + 0.04839158430695534, + 0.023352863267064095, + 0.030361656099557877, + 0.027808530256152153, + 0.024477383121848106, + 0.004566994030028582, + 0.004808300174772739, + 0.0072670006193220615, + 0.00772869074717164, + 0.007572604343295097, + 0.008202637545764446, + 0.0057082511484622955, + 0.006453463342040777, + 0.013445913791656494, + 0.011596551164984703, + 0.015932822600007057, + 0.02280457504093647, + 0.02103753760457039, + 0.017376983538269997, + 0.017746690660715103, + 0.014621065929532051, + 0.011706555262207985, + 0.01598900929093361, + 0.019659342244267464, + 0.025949371978640556, + 0.012689325027167797, + 0.02083493210375309, + 0.01612175814807415, + 0.021923325955867767, + 0.0028577703051269054, + 0.0021479905117303133, + 0.0042417338117957115, + 0.00355699029751122, + 0.0028329493943601847, + 0.004653098527342081, + 0.005525368265807629, + 0.003970262594521046, + 0.008519180119037628, + 0.004938546102494001, + 0.008135545067489147, + 0.0060768225230276585, + 0.009679731912910938, + 0.008176031522452831, + 0.0069412123411893845, + 0.01244659349322319, + 0.010276577435433865, + 0.01581663452088833, + 0.014387944713234901, + 0.014486554078757763, + 0.020613893866539, + 0.0377521812915802, + 0.04739799723029137, + 0.015284593217074871, + 0.00326275615952909, + 0.004884412512183189, + 0.004477532580494881, + 0.005234052427113056, + 0.0052899084985256195, + 0.004911957774311304, + 0.003998556174337864, + 0.0047806380316615105, + 0.0046498472802340984, + 0.005752744618803263, + 0.005013446323573589, + 0.0063232663087546825, + 0.007401745766401291, + 0.006742472760379314, + 0.0065047163516283035, + 0.004918951075524092, + 0.004339883103966713, + 0.0030340226367115974, + 0.003886827966198325, + 0.003840427612885833, + 0.003315483219921589, + 0.013392931781709194, + 0.0037882535252720118, + 0.01333449874073267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.005984720308333635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959243047982454 + ], + "opposite_logprob_tokens": [ + -5.1309590339660645 + ], + "logprob_reward": -0.005959243047982454, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959243047982454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3627748191356659, + 0.1023699939250946, + 0.13269010186195374, + 0.14234015345573425, + 0.04443119838833809, + 0.03350553661584854, + 0.057795386761426926, + 0.043058883398771286, + 0.0268575306981802, + 0.054176364094018936 + ], + "credit_map": [ + 0.024782508611679077, + 0.017820538952946663, + 0.04035034403204918, + 0.02668280154466629, + 0.08272123336791992, + 0.07041572034358978, + 0.028449760749936104, + 0.07155190408229828, + 0.008486490696668625, + 0.008037076331675053, + 0.010719063691794872, + 0.015857186168432236, + 0.0206496249884367, + 0.01445774920284748, + 0.011986210010945797, + 0.012176601216197014, + 0.01498486753553152, + 0.010163404047489166, + 0.01803712546825409, + 0.021664826199412346, + 0.01616840809583664, + 0.016837405040860176, + 0.017814012244343758, + 0.01702006347477436, + 0.015616709366440773, + 0.020116854459047318, + 0.01839613914489746, + 0.016978343948721886, + 0.013727974146604538, + 0.02287444658577442, + 0.013325759209692478, + 0.021303940564393997, + 0.0031817478593438864, + 0.005613310728222132, + 0.0048636384308338165, + 0.006272973958402872, + 0.004654769320040941, + 0.00640758266672492, + 0.009534492157399654, + 0.0039026853628456593, + 0.004585770424455404, + 0.0026368889957666397, + 0.003551103174686432, + 0.004060117527842522, + 0.0032048628199845552, + 0.004338795784860849, + 0.004194143693894148, + 0.006933856289833784, + 0.0035604003351181746, + 0.004467151127755642, + 0.003963383845984936, + 0.0043185921385884285, + 0.004956463351845741, + 0.014902348630130291, + 0.017130712047219276, + 0.0044963425025343895, + 0.0028825183399021626, + 0.0061518969014286995, + 0.0059852213598787785, + 0.005037397146224976, + 0.006777802482247353, + 0.005718333646655083, + 0.004514961503446102, + 0.005990758538246155, + 0.0018626210512593389, + 0.0031526151578873396, + 0.0028791537042707205, + 0.00400427496060729, + 0.0033920612186193466, + 0.00399611284956336, + 0.0037788916379213333, + 0.0037918025627732277, + 0.003610942978411913, + 0.003770167473703623, + 0.003486810252070427, + 0.003919144161045551, + 0.004819828551262617, + 0.015479995869100094, + 0.004758335184305906, + 0.014331144280731678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.005959243047982454 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036193120758980513 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036193120758980513, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036193120758980513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21919625997543335, + 0.15082404017448425, + 0.06637164205312729, + 0.08705615252256393, + 0.09418004006147385, + 0.03709207475185394, + 0.1208009347319603, + 0.06628591567277908, + 0.055192675441503525, + 0.10300015658140182 + ], + "credit_map": [ + 0.021336166188120842, + 0.017386319115757942, + 0.03149055689573288, + 0.04334244504570961, + 0.02178717404603958, + 0.03422445431351662, + 0.017177114263176918, + 0.03245203197002411, + 0.014373313635587692, + 0.008479714393615723, + 0.02462240494787693, + 0.01426944974809885, + 0.029706580564379692, + 0.022923002019524574, + 0.019581545144319534, + 0.016868053004145622, + 0.004773831460624933, + 0.004889967385679483, + 0.008483781479299068, + 0.008573146536946297, + 0.0068338969722390175, + 0.011251554824411869, + 0.0063492776826024055, + 0.015216188505291939, + 0.007961973547935486, + 0.00745469331741333, + 0.016892224550247192, + 0.013948260806500912, + 0.007859150879085064, + 0.013353830203413963, + 0.007768971845507622, + 0.01181705016642809, + 0.010403803549706936, + 0.005314517300575972, + 0.014472649432718754, + 0.011871013790369034, + 0.007358117960393429, + 0.013446761295199394, + 0.020167473703622818, + 0.01114571001380682, + 0.0049567329697310925, + 0.0032167481258511543, + 0.005399979185312986, + 0.0049689169973134995, + 0.0035748539958149195, + 0.004289332311600447, + 0.004930764902383089, + 0.0057547492906451225, + 0.011762984097003937, + 0.013490946963429451, + 0.015137762762606144, + 0.013280327431857586, + 0.02170870639383793, + 0.013339277356863022, + 0.016441144049167633, + 0.01563980057835579, + 0.004931275267153978, + 0.00964550394564867, + 0.008332726545631886, + 0.008700087666511536, + 0.010564008727669716, + 0.00832586083561182, + 0.007531322538852692, + 0.008255133405327797, + 0.005384228657931089, + 0.005871696397662163, + 0.006297244690358639, + 0.008559598587453365, + 0.008004888892173767, + 0.006591236684471369, + 0.008280745707452297, + 0.006203037686645985, + 0.009344245307147503, + 0.006576069165021181, + 0.006500049494206905, + 0.0056602987460792065, + 0.010168424807488918, + 0.030278386548161507, + 0.007108028046786785, + 0.027364656329154968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0036193120758980513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead shot with a Dutch tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a planet being ripped apart by gravity near its Roche limit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do fragments of the planet converge to form a ring against a rotating Milky Way background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments display a stark contrast between cool silvery outer shells and fiery crimson/orange interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a churning nebula with dust clouds of ochre, sulfur yellow, and deep purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808307275176048 + ], + "opposite_logprob_tokens": [ + -3.288083076477051 + ], + "logprob_reward": -0.03808307275176048, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808307275176048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5314199924468994, + 0.09115378558635712, + 0.046442240476608276, + 0.03994618356227875, + 0.01517986785620451, + 0.035734064877033234, + 0.10231421142816544, + 0.04756608232855797, + 0.026708781719207764, + 0.06353481858968735 + ], + "credit_map": [ + 0.037657927721738815, + 0.029532555490732193, + 0.055228084325790405, + 0.0614527128636837, + 0.10420337319374084, + 0.10101433843374252, + 0.03974102810025215, + 0.10258997231721878, + 0.008659757673740387, + 0.006208914332091808, + 0.013372623361647129, + 0.008075187914073467, + 0.013344361446797848, + 0.016969949007034302, + 0.014308403246104717, + 0.010214590467512608, + 0.0034925579093396664, + 0.0033794327173382044, + 0.006591197568923235, + 0.0064143529161810875, + 0.005910966079682112, + 0.007027766667306423, + 0.005003581289201975, + 0.008622388355433941, + 0.003211565315723419, + 0.0036415017675608397, + 0.0040499004535377026, + 0.0035273514222353697, + 0.004043856170028448, + 0.0075246552005410194, + 0.0033983218017965555, + 0.010549028404057026, + 0.0012182705104351044, + 0.0011232178658246994, + 0.0019897199235856533, + 0.0022824450861662626, + 0.001311694155447185, + 0.0019946331158280373, + 0.0033944915048778057, + 0.0018653967417776585, + 0.0048931739293038845, + 0.0027414599899202585, + 0.004656191449612379, + 0.003243881044909358, + 0.0038495834451168776, + 0.004582117777317762, + 0.004006365779787302, + 0.00776129262521863, + 0.008632971905171871, + 0.009174528531730175, + 0.011587461456656456, + 0.009314781986176968, + 0.014815881848335266, + 0.01649302989244461, + 0.02098199725151062, + 0.011313550174236298, + 0.0028060940094292164, + 0.006827684585005045, + 0.006050716619938612, + 0.006239674519747496, + 0.0077624330297112465, + 0.0060952696949243546, + 0.005219627171754837, + 0.006564581301063299, + 0.002467373153194785, + 0.002958807395771146, + 0.002993047470226884, + 0.0034748322796076536, + 0.002838741522282362, + 0.0038729477673768997, + 0.003951352089643478, + 0.004151680041104555, + 0.0044004227966070175, + 0.005088734906166792, + 0.004969724919646978, + 0.004183225799351931, + 0.0058549875393509865, + 0.016758183017373085, + 0.005546931643038988, + 0.016732607036828995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.03808307275176048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_12/index_tensor([5241, 5241])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_12/index_tensor([5241, 5241])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8f9e4ac54f444f000316484f0c215aadd39d65a0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_12/index_tensor([5241, 5241])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, established shot, shot from a first-person perspective, immerses the viewer in a claustrophobic and tense atmosphere. A dazzling, focused, circular flashlight beam suddenly pierces the all-consuming darkness, trembling back and forth across a chaotic pile of old objects in rapid, erratic movements. Each movement of the beam is characterized by anxiety, briefly illuminating dusty wooden boxes, rusted metal tools, stacks of yellowed newspapers, and a tangle of wires. Dust particles dance on the surfaces of these objects in the beam. The cold, white beam jumps nervously among the debris, occasionally pausing abruptly as if examining a particular object. Then, a vague hand cautiously pushes aside the obstruction, revealing something deeper. The stark contrast between light and shadow and the erratic movement of the beam create an atmosphere of intense focus, suspense, and urgency in the search for an object.", + "seed": 25590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_12/index_5241_seed_25590_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04930602163076401, + "vlm_logprob": -0.04930602163076401, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04930602163076401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.108867347240448 + ], + "opposite_logprob_tokens": [ + -2.4838674068450928 + ], + "logprob_reward": -0.108867347240448, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.108867347240448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15654344856739044, + 0.09358760714530945, + 0.17615245282649994, + 0.07295361161231995, + 0.05244603008031845, + 0.17554344236850739, + 0.05458104982972145, + 0.04782387986779213, + 0.08185640722513199, + 0.08851208537817001 + ], + "credit_map": [ + 0.017804542556405067, + 0.02309241145849228, + 0.014025324024260044, + 0.017175860702991486, + 0.026706218719482422, + 0.02169906347990036, + 0.020298926159739494, + 0.015741104260087013, + 0.012658336199820042, + 0.008518785238265991, + 0.0116293765604496, + 0.014477500692009926, + 0.008921479806303978, + 0.008998438715934753, + 0.015011959709227085, + 0.013371728360652924, + 0.02211899496614933, + 0.014080727472901344, + 0.023499654605984688, + 0.012079622596502304, + 0.018360575661063194, + 0.023660151287913322, + 0.043073561042547226, + 0.019279157742857933, + 0.00807199813425541, + 0.00601687328889966, + 0.012545035220682621, + 0.010969577357172966, + 0.007087331265211105, + 0.010543872602283955, + 0.0064214360900223255, + 0.011297491379082203, + 0.005246536806225777, + 0.00695028156042099, + 0.00587758794426918, + 0.0047641173005104065, + 0.00793606135994196, + 0.009991525672376156, + 0.004444468300789595, + 0.007235452998429537, + 0.01338507141917944, + 0.027076955884695053, + 0.013994392938911915, + 0.01605171523988247, + 0.01879224181175232, + 0.030094144865870476, + 0.02124527096748352, + 0.03490365669131279, + 0.00629454618319869, + 0.004604493733495474, + 0.0064398301765322685, + 0.009674186818301678, + 0.005860673729330301, + 0.006919141858816147, + 0.008380943909287453, + 0.006407229695469141, + 0.0058835637755692005, + 0.0032725916244089603, + 0.008932978846132755, + 0.005761814769357443, + 0.004140029661357403, + 0.00778918294236064, + 0.006938240956515074, + 0.005105477757751942, + 0.011293825693428516, + 0.010179171338677406, + 0.010990199632942677, + 0.009927836246788502, + 0.011355627328157425, + 0.011550597846508026, + 0.009293090552091599, + 0.007266055326908827, + 0.010297502391040325, + 0.013036275282502174, + 0.010713962838053703, + 0.011797863058745861, + 0.008993965573608875, + 0.01149389985948801, + 0.011488333344459534, + 0.01069028303027153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.108867347240448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00597393698990345 + ], + "opposite_logprob_tokens": [ + -5.130973815917969 + ], + "logprob_reward": -0.00597393698990345, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00597393698990345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144883394241333, + 0.07886172831058502, + 0.0872369259595871, + 0.15500779449939728, + 0.05820153281092644, + 0.28919482231140137, + 0.04945974051952362, + 0.07204391062259674, + 0.025618283078074455, + 0.0394919216632843 + ], + "credit_map": [ + 0.020022595301270485, + 0.017245419323444366, + 0.012291508726775646, + 0.015237141400575638, + 0.021235134452581406, + 0.0248419176787138, + 0.019136283546686172, + 0.014873386360704899, + 0.011549975723028183, + 0.007875418290495872, + 0.01188821904361248, + 0.012248444370925426, + 0.007503079250454903, + 0.004801570437848568, + 0.011870321817696095, + 0.011124702170491219, + 0.010513275861740112, + 0.009209271520376205, + 0.012665203772485256, + 0.009665078483521938, + 0.007068826351314783, + 0.011988988146185875, + 0.01612187922000885, + 0.010004403069615364, + 0.02758624032139778, + 0.01974496804177761, + 0.020172851160168648, + 0.01415447797626257, + 0.022312406450510025, + 0.01705174334347248, + 0.016442887485027313, + 0.017542215064167976, + 0.005611578933894634, + 0.0063548400066792965, + 0.006193558219820261, + 0.004491154570132494, + 0.00906129740178585, + 0.01150151714682579, + 0.006558462977409363, + 0.008429125882685184, + 0.021546250209212303, + 0.024976056069135666, + 0.01824699155986309, + 0.022199684754014015, + 0.023174943402409554, + 0.07435474544763565, + 0.027360690757632256, + 0.07733545452356339, + 0.004822528921067715, + 0.005318240728229284, + 0.004366962239146233, + 0.007034766487777233, + 0.0058997394517064095, + 0.0067009455524384975, + 0.009109217673540115, + 0.006207343190908432, + 0.010282649658620358, + 0.002666902495548129, + 0.01016966998577118, + 0.01330725010484457, + 0.004139950033277273, + 0.011394545435905457, + 0.011914427392184734, + 0.008168518543243408, + 0.0034710976760834455, + 0.003904789686203003, + 0.002700657118111849, + 0.0031734067015349865, + 0.002780989743769169, + 0.003516169497743249, + 0.0041724336333572865, + 0.0018987389048561454, + 0.0036986360792070627, + 0.005381122697144747, + 0.006149773020297289, + 0.0044969129376113415, + 0.0037505391519516706, + 0.006525670178234577, + 0.00468136090785265, + 0.004807907156646252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.00597393698990345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026365121826529503 + ], + "opposite_logprob_tokens": [ + -3.651365041732788 + ], + "logprob_reward": -0.026365121826529503, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026365121826529503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1416008174419403, + 0.051826149225234985, + 0.07392703741788864, + 0.04276130720973015, + 0.04964909330010414, + 0.2413061410188675, + 0.04663042724132538, + 0.2222830206155777, + 0.04546581953763962, + 0.08455019444227219 + ], + "credit_map": [ + 0.015787066891789436, + 0.017896175384521484, + 0.015405477955937386, + 0.021201565861701965, + 0.019546547904610634, + 0.01701418124139309, + 0.02003009244799614, + 0.014719714410603046, + 0.00760120851919055, + 0.003878978081047535, + 0.006856046617031097, + 0.008740618824958801, + 0.006090046837925911, + 0.0047852047719061375, + 0.006913754623383284, + 0.006960298866033554, + 0.008101947605609894, + 0.008684725500643253, + 0.012122458778321743, + 0.007944266311824322, + 0.006781203206628561, + 0.010010149329900742, + 0.01439940370619297, + 0.005882888101041317, + 0.006201306823641062, + 0.005539225414395332, + 0.0065483953803777695, + 0.0038545315619558096, + 0.005102185998111963, + 0.005716320592910051, + 0.0037257231306284666, + 0.006073623429983854, + 0.0045824358239769936, + 0.010809427127242088, + 0.0034337895922362804, + 0.003248529974371195, + 0.00881985668092966, + 0.007237179670482874, + 0.003946878015995026, + 0.0075710006058216095, + 0.019339559599757195, + 0.04209137335419655, + 0.014228327199816704, + 0.02621767297387123, + 0.037829287350177765, + 0.0340103805065155, + 0.034139715135097504, + 0.033449843525886536, + 0.005168647971004248, + 0.0036925384774804115, + 0.004038795363157988, + 0.006306525785475969, + 0.008747618645429611, + 0.006210651248693466, + 0.006790820509195328, + 0.005674832966178656, + 0.022412052378058434, + 0.013675158843398094, + 0.026834722608327866, + 0.041951071470975876, + 0.013227354735136032, + 0.03220897540450096, + 0.04455487057566643, + 0.027418814599514008, + 0.007215944584459066, + 0.003888092003762722, + 0.0071781170554459095, + 0.005715751089155674, + 0.0036430773325264454, + 0.005900349467992783, + 0.008451170288026333, + 0.0034733223728835583, + 0.009292376227676868, + 0.008302069269120693, + 0.01180294994264841, + 0.013217034749686718, + 0.013640088960528374, + 0.010615007020533085, + 0.009322862140834332, + 0.008357815444469452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.026365121826529503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635188400745392 + ], + "opposite_logprob_tokens": [ + -3.6513519287109375 + ], + "logprob_reward": -0.02635188400745392, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635188400745392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17115473747253418, + 0.08035369962453842, + 0.15284943580627441, + 0.1496141254901886, + 0.07523700594902039, + 0.08894798904657364, + 0.07763238996267319, + 0.06580928713083267, + 0.06668253242969513, + 0.07171886414289474 + ], + "credit_map": [ + 0.023491868749260902, + 0.019728215411305428, + 0.01777760311961174, + 0.014901582151651382, + 0.030435917899012566, + 0.026094796136021614, + 0.01928701251745224, + 0.019437748938798904, + 0.010606872849166393, + 0.006476119160652161, + 0.011035697534680367, + 0.015610596165060997, + 0.007136654574424028, + 0.007113668601959944, + 0.010741235688328743, + 0.011632852256298065, + 0.018181603401899338, + 0.02261343039572239, + 0.020701071247458458, + 0.017179349437355995, + 0.01189490593969822, + 0.020101407542824745, + 0.021404528990387917, + 0.020773138850927353, + 0.01806279830634594, + 0.014090685173869133, + 0.0275784432888031, + 0.018574794754385948, + 0.013993329368531704, + 0.01659293659031391, + 0.016335396096110344, + 0.024385744705796242, + 0.007519937586039305, + 0.008874756284058094, + 0.007420992013067007, + 0.007858281955122948, + 0.013573923148214817, + 0.011856970377266407, + 0.008018284104764462, + 0.010113857686519623, + 0.006648786831647158, + 0.010556393302977085, + 0.007800975814461708, + 0.009065982885658741, + 0.01869177632033825, + 0.012699779123067856, + 0.010485311038792133, + 0.012998979538679123, + 0.009556853212416172, + 0.008413267321884632, + 0.00879279151558876, + 0.008408579975366592, + 0.00926926825195551, + 0.010390372946858406, + 0.012224369682371616, + 0.010576890781521797, + 0.0065878829918801785, + 0.004070674069225788, + 0.007750877644866705, + 0.012039602734148502, + 0.005752589553594589, + 0.009412562474608421, + 0.012831123545765877, + 0.007363973185420036, + 0.010176228359341621, + 0.013756752014160156, + 0.00941541139036417, + 0.006024302449077368, + 0.00525827007368207, + 0.008852805010974407, + 0.006448531523346901, + 0.0067502278834581375, + 0.008883826434612274, + 0.008051336742937565, + 0.01012618001550436, + 0.008903122507035732, + 0.007358945440500975, + 0.010906483978033066, + 0.008811421692371368, + 0.008677546866238117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.02635188400745392 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897181808948517 + ], + "opposite_logprob_tokens": [ + -2.5789718627929688 + ], + "logprob_reward": -0.07897181808948517, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897181808948517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21283769607543945, + 0.16309764981269836, + 0.09832286834716797, + 0.04790716990828514, + 0.037860024720430374, + 0.09142003953456879, + 0.06626616418361664, + 0.07786592096090317, + 0.04898308217525482, + 0.15543946623802185 + ], + "credit_map": [ + 0.01571369357407093, + 0.017971547320485115, + 0.020585857331752777, + 0.018463242799043655, + 0.05244610458612442, + 0.031558696180582047, + 0.03478270396590233, + 0.021315833553671837, + 0.013669431209564209, + 0.010273879393935204, + 0.02290385589003563, + 0.04289723560214043, + 0.01417752355337143, + 0.013601631857454777, + 0.022933783009648323, + 0.022640299052000046, + 0.01087133027613163, + 0.01103908196091652, + 0.01576962321996689, + 0.009869993664324284, + 0.008352530188858509, + 0.01248379610478878, + 0.016380375251173973, + 0.013556129299104214, + 0.006880579516291618, + 0.0050755697302520275, + 0.008799868635833263, + 0.0038180092815309763, + 0.005842340644448996, + 0.005386147648096085, + 0.004583090543746948, + 0.007521561812609434, + 0.003277849406003952, + 0.006768792401999235, + 0.0032438356429338455, + 0.0026863710954785347, + 0.006165625993162394, + 0.006086532026529312, + 0.0038013134617358446, + 0.005829700268805027, + 0.0058674863539636135, + 0.006779157090932131, + 0.0038660005666315556, + 0.00694172503426671, + 0.012764142826199532, + 0.024881193414330482, + 0.007705306634306908, + 0.02261502854526043, + 0.0053605069406330585, + 0.005550856702029705, + 0.00663160253316164, + 0.009537851437926292, + 0.009957579895853996, + 0.009397244080901146, + 0.01306177768856287, + 0.0067687383852899075, + 0.006454747635871172, + 0.003995310515165329, + 0.009610053151845932, + 0.016944078728556633, + 0.0040997592732310295, + 0.012408466078341007, + 0.015705350786447525, + 0.008648158051073551, + 0.006252634339034557, + 0.007187047507613897, + 0.006592228077352047, + 0.006687627639621496, + 0.004922680091112852, + 0.006403821054846048, + 0.006068871822208166, + 0.00486816605553031, + 0.01997258886694908, + 0.013051377609372139, + 0.02398030087351799, + 0.015090099535882473, + 0.014977811835706234, + 0.028037074953317642, + 0.019894542172551155, + 0.020435651764273643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07897181808948517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, established shot, shot from a first-person perspective, immerses the viewer in a claustrophobic and tense atmosphere. A dazzling, focused, circular flashlight beam suddenly pierces the all-consuming darkness, trembling back and forth across a chaotic pile of old objects in rapid, erratic movements. Each movement of the beam is characterized by anxiety, briefly illuminating dusty wooden boxes, rusted metal tools, stacks of yellowed newspapers, and a tangle of wires. Dust particles dance on the surfaces of these objects in the beam. The cold, white beam jumps nervously among the debris, occasionally pausing abruptly as if examining a particular object. Then, a vague hand cautiously pushes aside the obstruction, revealing something deeper. The stark contrast between light and shadow and the erratic movement of the beam create an atmosphere of intense focus, suspense, and urgency in the search for an object.", + "seed": 25591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_12/index_5241_seed_25591_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08703986871987582, + "vlm_logprob": -0.08703986871987582, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.08703986871987582, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17554838955402374 + ], + "opposite_logprob_tokens": [ + -1.9255484342575073 + ], + "logprob_reward": -0.17554838955402374, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17554838955402374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05937489867210388, + 0.03461823984980583, + 0.0513019822537899, + 0.36435481905937195, + 0.04696346074342728, + 0.19949261844158173, + 0.07426851242780685, + 0.06292713433504105, + 0.054285917431116104, + 0.052412424236536026 + ], + "credit_map": [ + 0.006039052736014128, + 0.007270140573382378, + 0.008103091269731522, + 0.004707920365035534, + 0.007402858696877956, + 0.007859901525080204, + 0.007253206800669432, + 0.010738722048699856, + 0.004958241246640682, + 0.010033710859715939, + 0.0038324822671711445, + 0.0038177715614438057, + 0.0021786342840641737, + 0.004380217753350735, + 0.002336106263101101, + 0.0030810770113021135, + 0.007612334098666906, + 0.004739829804748297, + 0.007585367653518915, + 0.005540986079722643, + 0.005256987176835537, + 0.007586478255689144, + 0.006248367018997669, + 0.0067316340282559395, + 0.03566325828433037, + 0.029982004314661026, + 0.03674821928143501, + 0.03112262487411499, + 0.03166346251964569, + 0.060445189476013184, + 0.04882599040865898, + 0.08990409225225449, + 0.005254097282886505, + 0.009195888414978981, + 0.006865938194096088, + 0.004850052762776613, + 0.004230029881000519, + 0.005656671244651079, + 0.004825618118047714, + 0.0060851615853607655, + 0.020700689405202866, + 0.009928250685334206, + 0.008219379000365734, + 0.01039291825145483, + 0.023368747904896736, + 0.03446194529533386, + 0.01929122395813465, + 0.07312945276498795, + 0.0053183045238256454, + 0.008500688709318638, + 0.011329048313200474, + 0.009352047927677631, + 0.0073525384068489075, + 0.00978757068514824, + 0.008739904500544071, + 0.01388840563595295, + 0.006820932496339083, + 0.006453365087509155, + 0.007096054498106241, + 0.010973778553307056, + 0.005495326593518257, + 0.008037024177610874, + 0.010936419479548931, + 0.007114232052117586, + 0.005489258095622063, + 0.004423852078616619, + 0.007375841494649649, + 0.010757618583738804, + 0.005596358794718981, + 0.004509981721639633, + 0.007738003972917795, + 0.00839499942958355, + 0.004238680005073547, + 0.006186432670801878, + 0.005192593205720186, + 0.0038112225010991096, + 0.01400949340313673, + 0.0073206727392971516, + 0.006812341511249542, + 0.004840985406190157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.17554838955402374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008673259057104588 + ], + "opposite_logprob_tokens": [ + -4.758673191070557 + ], + "logprob_reward": -0.008673259057104588, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008673259057104588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14344264566898346, + 0.07743985205888748, + 0.06304572522640228, + 0.15911494195461273, + 0.11060251295566559, + 0.09405910223722458, + 0.10266102105379105, + 0.14128918945789337, + 0.05822186544537544, + 0.05012308806180954 + ], + "credit_map": [ + 0.011722339317202568, + 0.01899034157395363, + 0.01920994557440281, + 0.010602745227515697, + 0.022673938423395157, + 0.01792174205183983, + 0.023428305983543396, + 0.0188932903110981, + 0.009462033398449421, + 0.009422232396900654, + 0.010234358720481396, + 0.01000798400491476, + 0.008611291646957397, + 0.010020720772445202, + 0.011162454262375832, + 0.00851877685636282, + 0.008931824006140232, + 0.007467107381671667, + 0.008866437710821629, + 0.006438111420720816, + 0.004674980882555246, + 0.009260347113013268, + 0.0069770123809576035, + 0.010429903864860535, + 0.013391179963946342, + 0.02277699112892151, + 0.021656330674886703, + 0.018168287351727486, + 0.01578417234122753, + 0.022733092308044434, + 0.01954277604818344, + 0.025062119588255882, + 0.01211891882121563, + 0.026863688603043556, + 0.022149275988340378, + 0.011187748052179813, + 0.010271193459630013, + 0.009945861995220184, + 0.008431226946413517, + 0.009634590707719326, + 0.009620572440326214, + 0.0048921480774879456, + 0.0045736744068562984, + 0.006483164615929127, + 0.00979518797248602, + 0.015440172515809536, + 0.009275364689528942, + 0.03397881239652634, + 0.010504643432796001, + 0.013119472190737724, + 0.013979763723909855, + 0.015466401353478432, + 0.011365708895027637, + 0.014629931189119816, + 0.008489149622619152, + 0.015105956234037876, + 0.016490966081619263, + 0.012264019809663296, + 0.01928146556019783, + 0.027475783601403236, + 0.01646886020898819, + 0.014648500829935074, + 0.021951112896203995, + 0.012708481401205063, + 0.0047287060879170895, + 0.00474017858505249, + 0.008012419566512108, + 0.010369348339736462, + 0.007479283958673477, + 0.005436801817268133, + 0.007927240803837776, + 0.009527886286377907, + 0.004557776730507612, + 0.005513453856110573, + 0.007020370103418827, + 0.005982255097478628, + 0.008526593446731567, + 0.006609505973756313, + 0.005979664623737335, + 0.005933465901762247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.008673259057104588 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812920302152634 + ], + "opposite_logprob_tokens": [ + -3.2881290912628174 + ], + "logprob_reward": -0.03812920302152634, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812920302152634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07758846879005432, + 0.05663593113422394, + 0.16583143174648285, + 0.09230432659387589, + 0.13572737574577332, + 0.14793069660663605, + 0.04117123410105705, + 0.08913417905569077, + 0.10470519214868546, + 0.08897117525339127 + ], + "credit_map": [ + 0.007891367189586163, + 0.008268526755273342, + 0.015453037805855274, + 0.008174371905624866, + 0.008293523453176022, + 0.010979675687849522, + 0.008803492411971092, + 0.009724477306008339, + 0.007917758077383041, + 0.00815151073038578, + 0.004207747522741556, + 0.007234753575176001, + 0.0055501749739050865, + 0.010269849561154842, + 0.006518309935927391, + 0.006785827223211527, + 0.022081846371293068, + 0.019291404634714127, + 0.024206671863794327, + 0.019756294786930084, + 0.016105160117149353, + 0.020036252215504646, + 0.019306642934679985, + 0.025047147646546364, + 0.010819547809660435, + 0.006913698744028807, + 0.007071405183523893, + 0.007034736685454845, + 0.007585387211292982, + 0.016453662887215614, + 0.010902099311351776, + 0.02552379108965397, + 0.015661342069506645, + 0.0177615899592638, + 0.020534582436084747, + 0.010556019842624664, + 0.014153649099171162, + 0.020556295290589333, + 0.01935720257461071, + 0.017146704718470573, + 0.013593115843832493, + 0.0077407448552548885, + 0.005777309648692608, + 0.00896570272743702, + 0.018290484324097633, + 0.024216990917921066, + 0.016200080513954163, + 0.05314626172184944, + 0.003340182127431035, + 0.004624698311090469, + 0.004719571676105261, + 0.007894315756857395, + 0.004524040501564741, + 0.006200454663485289, + 0.004278998821973801, + 0.005588972009718418, + 0.011825405061244965, + 0.011031858623027802, + 0.010199490934610367, + 0.022323468700051308, + 0.006625925190746784, + 0.007225566543638706, + 0.0119745172560215, + 0.007927944883704185, + 0.0075247338972985744, + 0.0075564212165772915, + 0.016801830381155014, + 0.020686812698841095, + 0.013049707747995853, + 0.010680435225367546, + 0.01300236489623785, + 0.01540288981050253, + 0.00818057544529438, + 0.01266510970890522, + 0.009659764356911182, + 0.007885154336690903, + 0.018286358565092087, + 0.01140876766294241, + 0.011297316290438175, + 0.009588134475052357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.03812920302152634 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003170996904373 + ], + "opposite_logprob_tokens": [ + -2.6950316429138184 + ], + "logprob_reward": -0.07003170996904373, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003170996904373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10885077714920044, + 0.05640869960188866, + 0.17836695909500122, + 0.06400072574615479, + 0.0621979758143425, + 0.21704478561878204, + 0.08704628795385361, + 0.10776892304420471, + 0.0710662230849266, + 0.04724861681461334 + ], + "credit_map": [ + 0.012064768001437187, + 0.01171768270432949, + 0.023091666400432587, + 0.009557679295539856, + 0.010972989723086357, + 0.014829425141215324, + 0.010606767609715462, + 0.016009803861379623, + 0.00864366628229618, + 0.009929834865033627, + 0.00498271593824029, + 0.006773650646209717, + 0.003492244053632021, + 0.00956107210367918, + 0.0049183857627213, + 0.008107125759124756, + 0.021864134818315506, + 0.017865661531686783, + 0.021490920335054398, + 0.020085221156477928, + 0.013300320133566856, + 0.03605446219444275, + 0.01956099085509777, + 0.02814524434506893, + 0.005523974075913429, + 0.005900013260543346, + 0.00650362204760313, + 0.005295292474329472, + 0.006428997032344341, + 0.011487595736980438, + 0.006912337150424719, + 0.015948891639709473, + 0.007816697470843792, + 0.009193233214318752, + 0.007065810728818178, + 0.005710904020816088, + 0.006044609472155571, + 0.009090160019695759, + 0.007223340217024088, + 0.010053218342363834, + 0.020430870354175568, + 0.011272400617599487, + 0.008870254270732403, + 0.014170157723128796, + 0.025833113119006157, + 0.036443211138248444, + 0.02071906067430973, + 0.07930571585893631, + 0.011407873593270779, + 0.01053833682090044, + 0.01224382221698761, + 0.00973599124699831, + 0.006753304973244667, + 0.011476052924990654, + 0.008322978392243385, + 0.016567928716540337, + 0.013944949954748154, + 0.0126242870464921, + 0.015308404341340065, + 0.01765649951994419, + 0.008461772464215755, + 0.012272046878933907, + 0.01374438963830471, + 0.01375657320022583, + 0.006288685370236635, + 0.0074264612048864365, + 0.008593466132879257, + 0.012683534994721413, + 0.009752070531249046, + 0.005692330189049244, + 0.010290361940860748, + 0.010339312255382538, + 0.003715573577210307, + 0.008129380643367767, + 0.004283393267542124, + 0.005757633130997419, + 0.006147206760942936, + 0.005621952470391989, + 0.00805511325597763, + 0.0055383602157235146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.07003170996904373 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281678199768066 + ], + "opposite_logprob_tokens": [ + -2.0178167819976807 + ], + "logprob_reward": -0.14281678199768066, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281678199768066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0987120121717453, + 0.07343190163373947, + 0.10969485342502594, + 0.0773586705327034, + 0.05409114807844162, + 0.17068909108638763, + 0.08887632191181183, + 0.143161803483963, + 0.058560729026794434, + 0.12542346119880676 + ], + "credit_map": [ + 0.009493687190115452, + 0.01033035758882761, + 0.011468159034848213, + 0.00769400829449296, + 0.012304854579269886, + 0.011815889738500118, + 0.01591482199728489, + 0.01969023235142231, + 0.011300325393676758, + 0.011321883648633957, + 0.0050288112834095955, + 0.011419959366321564, + 0.005856268107891083, + 0.012586744502186775, + 0.00711255194619298, + 0.008805357851088047, + 0.014420244842767715, + 0.006937236059457064, + 0.01431807316839695, + 0.01130063645541668, + 0.008014112710952759, + 0.022346217185258865, + 0.01281612366437912, + 0.019542213529348373, + 0.008270286954939365, + 0.0092746177688241, + 0.012715799733996391, + 0.007945744320750237, + 0.00784300360828638, + 0.010983348824083805, + 0.00938502512872219, + 0.010940853506326675, + 0.00629378529265523, + 0.007552612107247114, + 0.006234578788280487, + 0.004967213608324528, + 0.0055121248587965965, + 0.010589651763439178, + 0.006716907024383545, + 0.006224276497960091, + 0.016609832644462585, + 0.007098481524735689, + 0.007129098288714886, + 0.010288269259035587, + 0.018306124955415726, + 0.029309658333659172, + 0.016269806772470474, + 0.06567782908678055, + 0.006966402288526297, + 0.010218924842774868, + 0.01193339191377163, + 0.015677589923143387, + 0.008267130702733994, + 0.01164296269416809, + 0.01272687315940857, + 0.011443052440881729, + 0.016032405197620392, + 0.016228513792157173, + 0.016443559899926186, + 0.030416831374168396, + 0.013170547783374786, + 0.013326209038496017, + 0.02096240222454071, + 0.0165813397616148, + 0.004972284659743309, + 0.004461505450308323, + 0.01093936339020729, + 0.008232279680669308, + 0.008686329238116741, + 0.005875891540199518, + 0.006910082884132862, + 0.008482994511723518, + 0.007521548308432102, + 0.017023062333464622, + 0.018624750897288322, + 0.01183854416012764, + 0.024315563961863518, + 0.01704224944114685, + 0.015243826434016228, + 0.013813930563628674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.14281678199768066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, established shot, shot from a first-person perspective, immerses the viewer in a claustrophobic and tense atmosphere. A dazzling, focused, circular flashlight beam suddenly pierces the all-consuming darkness, trembling back and forth across a chaotic pile of old objects in rapid, erratic movements. Each movement of the beam is characterized by anxiety, briefly illuminating dusty wooden boxes, rusted metal tools, stacks of yellowed newspapers, and a tangle of wires. Dust particles dance on the surfaces of these objects in the beam. The cold, white beam jumps nervously among the debris, occasionally pausing abruptly as if examining a particular object. Then, a vague hand cautiously pushes aside the obstruction, revealing something deeper. The stark contrast between light and shadow and the erratic movement of the beam create an atmosphere of intense focus, suspense, and urgency in the search for an object.", + "seed": 25592, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_12/index_5241_seed_25592_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0696302737109363, + "vlm_logprob": -0.0696302737109363, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.0696302737109363, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03881252929568291 + ], + "opposite_logprob_tokens": [ + -4.413812637329102 + ], + "logprob_reward": -0.03881252929568291, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03881252929568291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09000637382268906, + 0.06709930300712585, + 0.07223797589540482, + 0.061959270387887955, + 0.09694790840148926, + 0.028492692857980728, + 0.06623544543981552, + 0.06477901339530945, + 0.15761040151119232, + 0.29463163018226624 + ], + "credit_map": [ + 0.009990646503865719, + 0.01926754228770733, + 0.01436616387218237, + 0.017620114609599113, + 0.006316239479929209, + 0.0053415922448039055, + 0.009155184030532837, + 0.007948883809149265, + 0.010021329857409, + 0.006087664980441332, + 0.01568659394979477, + 0.008887303061783314, + 0.0032902895472943783, + 0.0063344151712954044, + 0.005343003664165735, + 0.011448705568909645, + 0.004893413744866848, + 0.007040505763143301, + 0.011906006373465061, + 0.008716180920600891, + 0.005456281825900078, + 0.01101780030876398, + 0.01000195648521185, + 0.013205824419856071, + 0.005955791566520929, + 0.006382545921951532, + 0.0042539541609585285, + 0.007784986402839422, + 0.0023961057886481285, + 0.008340096101164818, + 0.012951036915183067, + 0.013894752599298954, + 0.010543976910412312, + 0.011816476471722126, + 0.019186777994036674, + 0.017902713268995285, + 0.004856081213802099, + 0.01092483103275299, + 0.009864346124231815, + 0.011852698400616646, + 0.004226772580295801, + 0.004243533127009869, + 0.00625243503600359, + 0.0040269107557833195, + 0.0015229237033054233, + 0.0032198296394199133, + 0.002385638188570738, + 0.0026146499440073967, + 0.008217238821089268, + 0.007017520722001791, + 0.007141072303056717, + 0.008807789534330368, + 0.004218878224492073, + 0.007909630425274372, + 0.008947812020778656, + 0.013975508511066437, + 0.0033090501092374325, + 0.0029041790403425694, + 0.007248447742313147, + 0.0042351000010967255, + 0.003867685329169035, + 0.019669808447360992, + 0.005037911236286163, + 0.018506832420825958, + 0.009333999827504158, + 0.007096295710653067, + 0.05289444699883461, + 0.011819901876151562, + 0.005101276095956564, + 0.03279486671090126, + 0.01706465519964695, + 0.021504949778318405, + 0.011121731251478195, + 0.005396398715674877, + 0.12552949786186218, + 0.014247589744627476, + 0.004737516399472952, + 0.014995772391557693, + 0.00907175149768591, + 0.10953138768672943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.03881252929568291 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005294231232255697 + ], + "opposite_logprob_tokens": [ + -5.255294322967529 + ], + "logprob_reward": -0.005294231232255697, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005294231232255697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19180408120155334, + 0.05410131812095642, + 0.04392963647842407, + 0.0339227169752121, + 0.08825010061264038, + 0.023260075598955154, + 0.052095912396907806, + 0.18255378305912018, + 0.2370038777589798, + 0.09307861328125 + ], + "credit_map": [ + 0.01893288642168045, + 0.024046657606959343, + 0.02794296108186245, + 0.05301819369196892, + 0.013018998317420483, + 0.013564466498792171, + 0.020640049129724503, + 0.020639831200242043, + 0.013258018530905247, + 0.005073084030300379, + 0.010123088955879211, + 0.006725385785102844, + 0.0021444775629788637, + 0.0060321372002363205, + 0.004026468843221664, + 0.0067186481319367886, + 0.003100421978160739, + 0.00394670432433486, + 0.008051753044128418, + 0.007086408790200949, + 0.0040399013087153435, + 0.00536402128636837, + 0.005957433488219976, + 0.006382983177900314, + 0.0034891271498054266, + 0.0037716664373874664, + 0.0031434299889951944, + 0.0049391845241189, + 0.0023910729214549065, + 0.004430771339684725, + 0.007197854109108448, + 0.004559605848044157, + 0.008601660840213299, + 0.013414327055215836, + 0.004417767748236656, + 0.011443138122558594, + 0.007323163561522961, + 0.015297765843570232, + 0.01607104390859604, + 0.011681221425533295, + 0.0023805787786841393, + 0.0030961045995354652, + 0.0030713286250829697, + 0.0028835698030889034, + 0.0016617869259789586, + 0.0032384712249040604, + 0.0034391642548143864, + 0.003489065682515502, + 0.0057976399548351765, + 0.0055167702957987785, + 0.006130936089903116, + 0.007662266027182341, + 0.00347340595908463, + 0.006806084420531988, + 0.009319225326180458, + 0.0073895761743187904, + 0.007980093359947205, + 0.007508994545787573, + 0.006266478914767504, + 0.007384799420833588, + 0.018802547827363014, + 0.06191442534327507, + 0.010525942780077457, + 0.06217045336961746, + 0.011463740840554237, + 0.005626585800200701, + 0.09021822363138199, + 0.011093420907855034, + 0.006005120929330587, + 0.09042435139417648, + 0.00968150980770588, + 0.012490877881646156, + 0.0034048312809318304, + 0.0013710932107642293, + 0.03974289447069168, + 0.004850901197642088, + 0.0016989248106256127, + 0.004571252502501011, + 0.0023675570264458656, + 0.03507114574313164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.005294231232255697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062081336975097656 + ], + "opposite_logprob_tokens": [ + -2.8120813369750977 + ], + "logprob_reward": -0.062081336975097656, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062081336975097656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07252303510904312, + 0.07140535116195679, + 0.09042051434516907, + 0.08195535838603973, + 0.033746104687452316, + 0.07362382113933563, + 0.07739253342151642, + 0.1101040169596672, + 0.30925652384757996, + 0.07957279682159424 + ], + "credit_map": [ + 0.008862475864589214, + 0.009622865356504917, + 0.01247063186019659, + 0.015249288640916348, + 0.003611417021602392, + 0.005721062421798706, + 0.007826469838619232, + 0.009158826433122158, + 0.012495158240199089, + 0.009250888600945473, + 0.011711237020790577, + 0.010500825010240078, + 0.004076975863426924, + 0.0060872542671859264, + 0.005571739748120308, + 0.011711271479725838, + 0.006012048106640577, + 0.009163331240415573, + 0.015125536359846592, + 0.01556225772947073, + 0.005675585474818945, + 0.011128347367048264, + 0.011397119611501694, + 0.01635628379881382, + 0.008505521342158318, + 0.010236725211143494, + 0.007001305464655161, + 0.010025314055383205, + 0.004109938628971577, + 0.011571438051760197, + 0.017674263566732407, + 0.012830857187509537, + 0.003840318415313959, + 0.004770300351083279, + 0.006408765912055969, + 0.005623425357043743, + 0.0015224006492644548, + 0.0038203480653464794, + 0.0038398357573896646, + 0.00392070971429348, + 0.014211087487637997, + 0.011932333931326866, + 0.017391184344887733, + 0.0074617681093513966, + 0.0024274743627756834, + 0.008662629872560501, + 0.005030830390751362, + 0.006506509613245726, + 0.008676602505147457, + 0.006188283674418926, + 0.009841428138315678, + 0.015007384121418, + 0.004684306215494871, + 0.0094253895804286, + 0.014135545119643211, + 0.009433590807020664, + 0.006113553419709206, + 0.005239012185484171, + 0.013226768001914024, + 0.0058600218035280704, + 0.004904299043118954, + 0.03310989961028099, + 0.009443167597055435, + 0.032207295298576355, + 0.016194069758057594, + 0.008510992862284184, + 0.11865120381116867, + 0.018973518162965775, + 0.008993172086775303, + 0.10318879783153534, + 0.014640852808952332, + 0.020103909075260162, + 0.0032224240712821484, + 0.002111986046656966, + 0.03192613273859024, + 0.005445585120469332, + 0.0024061109870672226, + 0.004032472148537636, + 0.002569160657003522, + 0.027858927845954895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.062081336975097656 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991554737091064 + ], + "opposite_logprob_tokens": [ + -1.8049155473709106 + ], + "logprob_reward": -0.17991554737091064, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991554737091064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07738707214593887, + 0.13662882149219513, + 0.08108532428741455, + 0.08010513335466385, + 0.06592704355716705, + 0.12523464858531952, + 0.038022059947252274, + 0.036311209201812744, + 0.29045411944389343, + 0.06884453445672989 + ], + "credit_map": [ + 0.010573484934866428, + 0.01160760223865509, + 0.013649594970047474, + 0.015800850465893745, + 0.005003972910344601, + 0.004690955858677626, + 0.0077323936857283115, + 0.008328219875693321, + 0.034831702709198, + 0.010951911099255085, + 0.026344571262598038, + 0.012610230594873428, + 0.007178846746683121, + 0.018431834876537323, + 0.010046495124697685, + 0.016233226284384727, + 0.0055124168284237385, + 0.009273664094507694, + 0.013074629008769989, + 0.009509262628853321, + 0.005478632636368275, + 0.010176166892051697, + 0.015083931386470795, + 0.012976625934243202, + 0.0079618776217103, + 0.012495143339037895, + 0.005646413192152977, + 0.009225470013916492, + 0.004727701656520367, + 0.011332535184919834, + 0.01462289784103632, + 0.014093099161982536, + 0.006766238249838352, + 0.011481234803795815, + 0.007201419677585363, + 0.008217839524149895, + 0.004086898639798164, + 0.010114790871739388, + 0.010289425030350685, + 0.007769200950860977, + 0.01503458246588707, + 0.013661314733326435, + 0.021377358585596085, + 0.015579842031002045, + 0.008439689874649048, + 0.01842486672103405, + 0.018068261444568634, + 0.0146487420424819, + 0.004828236065804958, + 0.00236684363335371, + 0.004141748882830143, + 0.003163642017170787, + 0.0018660251516848803, + 0.005808211863040924, + 0.00468018464744091, + 0.011167164891958237, + 0.0016872741980478168, + 0.0021072609815746546, + 0.0021203202195465565, + 0.002290382282808423, + 0.003042791038751602, + 0.011584224179387093, + 0.0024008196778595448, + 0.01107813697308302, + 0.013612285256385803, + 0.006601936183869839, + 0.11246120184659958, + 0.013818482868373394, + 0.007209463976323605, + 0.1082831621170044, + 0.012082028202712536, + 0.016385558992624283, + 0.0027542707975953817, + 0.0012322841212153435, + 0.027451196685433388, + 0.004021958447992802, + 0.0018699215725064278, + 0.003831497160717845, + 0.003316813614219427, + 0.0243665911257267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.17991554737091064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062047723680734634 + ], + "opposite_logprob_tokens": [ + -2.8120477199554443 + ], + "logprob_reward": -0.062047723680734634, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062047723680734634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14854606986045837, + 0.035069648176431656, + 0.03483998030424118, + 0.055413421243429184, + 0.22304899990558624, + 0.03275144100189209, + 0.028220098465681076, + 0.03348304331302643, + 0.09160047024488449, + 0.31702685356140137 + ], + "credit_map": [ + 0.01508663035929203, + 0.017990775406360626, + 0.02260410413146019, + 0.04762403294444084, + 0.005928413476794958, + 0.00644559133797884, + 0.016572177410125732, + 0.016294335946440697, + 0.006524394266307354, + 0.0027127228677272797, + 0.006365838460624218, + 0.005288929678499699, + 0.0020461941603571177, + 0.005671392660588026, + 0.0030041737481951714, + 0.003456002101302147, + 0.002554618986323476, + 0.004596581216901541, + 0.006255690939724445, + 0.0042037684470415115, + 0.00233997474424541, + 0.004608893766999245, + 0.006149533670395613, + 0.004130915738642216, + 0.004523861221969128, + 0.006544568110257387, + 0.004596365615725517, + 0.006968197878450155, + 0.0038752127438783646, + 0.008096831850707531, + 0.010014473460614681, + 0.010793905705213547, + 0.021324342116713524, + 0.02704346925020218, + 0.031544726341962814, + 0.03538374602794647, + 0.014058233238756657, + 0.02752326801419258, + 0.03786982223391533, + 0.028301382437348366, + 0.004028008319437504, + 0.004427418112754822, + 0.007699132431298494, + 0.003908675163984299, + 0.0011884634150192142, + 0.004668662324547768, + 0.0034917073789983988, + 0.003339371643960476, + 0.002584510948508978, + 0.002510466380044818, + 0.003553684800863266, + 0.0036718249320983887, + 0.0014945714501664042, + 0.003930917475372553, + 0.0047936951741576195, + 0.005680423695594072, + 0.0030327539425343275, + 0.0029924928676337004, + 0.00415994506329298, + 0.004178864881396294, + 0.003320119809359312, + 0.004375705495476723, + 0.007461744826287031, + 0.003961415030062199, + 0.005740018095821142, + 0.004563626367598772, + 0.029189422726631165, + 0.009922339580953121, + 0.0041100988164544106, + 0.018862003460526466, + 0.006256932392716408, + 0.012956028804183006, + 0.013741695322096348, + 0.008666876703500748, + 0.1279015988111496, + 0.020247643813490868, + 0.005844011437147856, + 0.016006655991077423, + 0.010382483713328838, + 0.11423587799072266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.062047723680734634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, established shot, shot from a first-person perspective, immerses the viewer in a claustrophobic and tense atmosphere. A dazzling, focused, circular flashlight beam suddenly pierces the all-consuming darkness, trembling back and forth across a chaotic pile of old objects in rapid, erratic movements. Each movement of the beam is characterized by anxiety, briefly illuminating dusty wooden boxes, rusted metal tools, stacks of yellowed newspapers, and a tangle of wires. Dust particles dance on the surfaces of these objects in the beam. The cold, white beam jumps nervously among the debris, occasionally pausing abruptly as if examining a particular object. Then, a vague hand cautiously pushes aside the obstruction, revealing something deeper. The stark contrast between light and shadow and the erratic movement of the beam create an atmosphere of intense focus, suspense, and urgency in the search for an object.", + "seed": 25593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_12/index_5241_seed_25593_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05281364377588034, + "vlm_logprob": -0.05281364377588034, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.05281364377588034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07778406143188477 + ], + "opposite_logprob_tokens": [ + -2.8277840614318848 + ], + "logprob_reward": -0.07778406143188477, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07778406143188477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13060660660266876, + 0.06282312422990799, + 0.05614476278424263, + 0.08255311846733093, + 0.24146278202533722, + 0.16179797053337097, + 0.06096940115094185, + 0.11769849807024002, + 0.059737108647823334, + 0.02620656229555607 + ], + "credit_map": [ + 0.013434036634862423, + 0.016468090936541557, + 0.01909605972468853, + 0.010446954518556595, + 0.023036913946270943, + 0.017295539379119873, + 0.008841111324727535, + 0.021987900137901306, + 0.008321871049702168, + 0.006473932880908251, + 0.017354506999254227, + 0.0037323397118598223, + 0.008096727542579174, + 0.008853638544678688, + 0.003429020754992962, + 0.006561084650456905, + 0.008721904829144478, + 0.0091935396194458, + 0.009665904566645622, + 0.00483560049906373, + 0.007804370950907469, + 0.008084550499916077, + 0.0033087381161749363, + 0.004530153702944517, + 0.013088461011648178, + 0.010562627576291561, + 0.013354863971471786, + 0.008726687170565128, + 0.008506248705089092, + 0.013923349790275097, + 0.006672245915979147, + 0.007718631532043219, + 0.03794339671730995, + 0.040758002549409866, + 0.04087568074464798, + 0.03435363247990608, + 0.01813986711204052, + 0.021691562607884407, + 0.022210709750652313, + 0.025489935651421547, + 0.022439543157815933, + 0.02703789621591568, + 0.02185174822807312, + 0.017898615449666977, + 0.019847625866532326, + 0.01550105307251215, + 0.017497936263680458, + 0.01972355507314205, + 0.006638168357312679, + 0.010932945646345615, + 0.006190512795001268, + 0.003401524154469371, + 0.010846435092389584, + 0.007450081408023834, + 0.008428942412137985, + 0.007080788724124432, + 0.020958930253982544, + 0.0186070054769516, + 0.011428206227719784, + 0.011630777269601822, + 0.017900267615914345, + 0.015038537792861462, + 0.008629664778709412, + 0.013505112379789352, + 0.011505445465445518, + 0.013672096654772758, + 0.00643895985558629, + 0.004127041436731815, + 0.00991442147642374, + 0.0056581683456897736, + 0.004118084907531738, + 0.004302890971302986, + 0.00483992462977767, + 0.0037293301429599524, + 0.0034074741415679455, + 0.002429169835522771, + 0.004093480762094259, + 0.002246777294203639, + 0.0023107773158699274, + 0.003149625612422824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.07778406143188477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009809738956391811 + ], + "opposite_logprob_tokens": [ + -4.634809970855713 + ], + "logprob_reward": -0.009809738956391811, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009809738956391811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11174269765615463, + 0.04495425522327423, + 0.08200369030237198, + 0.07109434902667999, + 0.08731009811162949, + 0.08209861814975739, + 0.09699882566928864, + 0.31881073117256165, + 0.06512342393398285, + 0.03986334800720215 + ], + "credit_map": [ + 0.011460190638899803, + 0.015073788352310658, + 0.019991755485534668, + 0.010676647536456585, + 0.01401659194380045, + 0.02057422511279583, + 0.009409221820533276, + 0.010540274903178215, + 0.0059460182674229145, + 0.005714097525924444, + 0.0058341072872281075, + 0.0029437451157718897, + 0.008032510988414288, + 0.006404654588550329, + 0.0027739903889596462, + 0.0073051308281719685, + 0.012203190475702286, + 0.009291108697652817, + 0.016675559803843498, + 0.009017077274620533, + 0.005846149753779173, + 0.011218208819627762, + 0.007581871002912521, + 0.010170520283281803, + 0.011515368707478046, + 0.007529374212026596, + 0.011925329454243183, + 0.006973897106945515, + 0.0066415099427104, + 0.01109314151108265, + 0.007584546692669392, + 0.007831179536879063, + 0.01216065976768732, + 0.014931323006749153, + 0.01509806327521801, + 0.010465269908308983, + 0.008450846187770367, + 0.009315057657659054, + 0.006136969197541475, + 0.010751903057098389, + 0.012007436715066433, + 0.01261382456868887, + 0.010829481296241283, + 0.008981847204267979, + 0.011417816393077374, + 0.009000221267342567, + 0.007984020747244358, + 0.009263974614441395, + 0.008962931111454964, + 0.015519936569035053, + 0.007748664356768131, + 0.005008233245462179, + 0.020479507744312286, + 0.015239251777529716, + 0.010326160117983818, + 0.01371413841843605, + 0.061083607375621796, + 0.04963690787553787, + 0.031247563660144806, + 0.028891194611787796, + 0.0538545586168766, + 0.037622787058353424, + 0.022928329184651375, + 0.03354577720165253, + 0.010008237324655056, + 0.011125841177999973, + 0.008691759780049324, + 0.007082408294081688, + 0.007578219287097454, + 0.009693779982626438, + 0.0050976150669157505, + 0.0058455620892345905, + 0.0073511735536158085, + 0.00342542864382267, + 0.005913801025599241, + 0.004036243073642254, + 0.003431876190006733, + 0.005878856871277094, + 0.0031667749863117933, + 0.006659194827079773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.009809738956391811 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372587636113167 + ], + "opposite_logprob_tokens": [ + -3.4087259769439697 + ], + "logprob_reward": -0.03372587636113167, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372587636113167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16061758995056152, + 0.06869423389434814, + 0.08899860829114914, + 0.03281876817345619, + 0.07357561588287354, + 0.08217291533946991, + 0.1975366175174713, + 0.18123699724674225, + 0.06987833231687546, + 0.04447033256292343 + ], + "credit_map": [ + 0.025072643533349037, + 0.014702338725328445, + 0.030422396957874298, + 0.010487014427781105, + 0.018028361722826958, + 0.03455158323049545, + 0.010374360717833042, + 0.016978885978460312, + 0.010104172863066196, + 0.006966490764170885, + 0.014318166300654411, + 0.006304046604782343, + 0.008576864376664162, + 0.011324248276650906, + 0.004689981695264578, + 0.006410251837223768, + 0.012125195004045963, + 0.010741939768195152, + 0.014649119228124619, + 0.009044001810252666, + 0.007624062243849039, + 0.0157479178160429, + 0.00798116810619831, + 0.011085194535553455, + 0.004636845551431179, + 0.0046348911710083485, + 0.004066771827638149, + 0.004006995819509029, + 0.00320700416341424, + 0.00418766587972641, + 0.004037629812955856, + 0.004040957428514957, + 0.011193119920790195, + 0.014049221761524677, + 0.016429435461759567, + 0.0071533783338963985, + 0.005309008993208408, + 0.007050386629998684, + 0.004669245798140764, + 0.0077218106016516685, + 0.01107396837323904, + 0.015429949387907982, + 0.00981304980814457, + 0.009882576763629913, + 0.012479460798203945, + 0.008234703913331032, + 0.008055287413299084, + 0.00720391096547246, + 0.019940216094255447, + 0.02563842013478279, + 0.027344536036252975, + 0.011323658749461174, + 0.028339272364974022, + 0.026920493692159653, + 0.017564989626407623, + 0.04046500846743584, + 0.034268639981746674, + 0.027616377919912338, + 0.017745938152074814, + 0.01655583456158638, + 0.030264653265476227, + 0.020824693143367767, + 0.013631287962198257, + 0.020329549908638, + 0.009982867166399956, + 0.00974291656166315, + 0.011462140828371048, + 0.007088886108249426, + 0.011075395159423351, + 0.011051816865801811, + 0.003355177817866206, + 0.006119117606431246, + 0.005396099295467138, + 0.00579801993444562, + 0.006651249248534441, + 0.005631508771330118, + 0.00420754449442029, + 0.004325283225625753, + 0.005557927768677473, + 0.0069026947021484375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.03372587636113167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806189239025116 + ], + "opposite_logprob_tokens": [ + -3.529806137084961 + ], + "logprob_reward": -0.029806189239025116, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806189239025116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.069157674908638, + 0.04726595804095268, + 0.5605689883232117, + 0.02004891261458397, + 0.03251709043979645, + 0.0560995489358902, + 0.050083354115486145, + 0.06489875912666321, + 0.07927579432725906, + 0.020084019750356674 + ], + "credit_map": [ + 0.006819272413849831, + 0.007189820520579815, + 0.009382587857544422, + 0.004173927009105682, + 0.011381912045180798, + 0.01255001686513424, + 0.006183970253914595, + 0.011476167477667332, + 0.005989859811961651, + 0.0047459728084504604, + 0.012383217923343182, + 0.0047043319791555405, + 0.004931567702442408, + 0.0062076495960354805, + 0.002327599097043276, + 0.005975762382149696, + 0.09606950730085373, + 0.09013581275939941, + 0.13697220385074615, + 0.044772956520318985, + 0.06584012508392334, + 0.06245996803045273, + 0.02265794388949871, + 0.04166044294834137, + 0.0029786156956106424, + 0.002077060053125024, + 0.0036556008271872997, + 0.002188832499086857, + 0.0012316731736063957, + 0.003747061127796769, + 0.0016823098994791508, + 0.0024877586401998997, + 0.004447822459042072, + 0.0052297161892056465, + 0.005390828475356102, + 0.0031616517808288336, + 0.0036090451758354902, + 0.004087280482053757, + 0.0032426253892481327, + 0.003348120953887701, + 0.00765098724514246, + 0.00962259341031313, + 0.007494388148188591, + 0.007168883457779884, + 0.0070366221480071545, + 0.005917548201978207, + 0.004793399013578892, + 0.006415128242224455, + 0.005259760655462742, + 0.012563337571918964, + 0.004854167811572552, + 0.002492419444024563, + 0.008529718033969402, + 0.007116436026990414, + 0.003364286618307233, + 0.005903230514377356, + 0.012336190789937973, + 0.01298571564257145, + 0.009207643568515778, + 0.005633392836898565, + 0.009135576896369457, + 0.0066474489867687225, + 0.003385973861441016, + 0.005566817242652178, + 0.019393272697925568, + 0.023566244170069695, + 0.010460206307470798, + 0.005908474326133728, + 0.005761560518294573, + 0.006174235604703426, + 0.0031707463786005974, + 0.0048410589806735516, + 0.002907935995608568, + 0.0019106129184365273, + 0.003841094207018614, + 0.0026240141596645117, + 0.002315149176865816, + 0.0020935037173330784, + 0.0017703158082440495, + 0.002621393883600831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.029806189239025116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a circular flashlight beam visible piercing the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dusty wooden boxes, rusted metal tools, yellowed newspapers, and wires visible in the pile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flashlight beam move in rapid, erratic, and trembling motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a vague hand appear to push aside an obstruction to reveal something deeper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294235289096832 + ], + "opposite_logprob_tokens": [ + -2.2379424571990967 + ], + "logprob_reward": -0.11294235289096832, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294235289096832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14146064221858978, + 0.03309763967990875, + 0.14884640276432037, + 0.07471828162670135, + 0.06561354547739029, + 0.10674289613962173, + 0.08852945268154144, + 0.07089819759130478, + 0.2164083868265152, + 0.053684528917074203 + ], + "credit_map": [ + 0.01614958792924881, + 0.017319483682513237, + 0.025660663843154907, + 0.011848115362226963, + 0.02537425421178341, + 0.020100442692637444, + 0.009983825497329235, + 0.015024270862340927, + 0.0037386114709079266, + 0.0043663824908435345, + 0.006268640514463186, + 0.002650406677275896, + 0.005424791015684605, + 0.003774638520553708, + 0.0018573473207652569, + 0.0050168223679065704, + 0.024148985743522644, + 0.02405410073697567, + 0.03446226567029953, + 0.011604339815676212, + 0.016058990731835365, + 0.01916908659040928, + 0.006660593673586845, + 0.01268803235143423, + 0.012267332524061203, + 0.010160579346120358, + 0.008385169319808483, + 0.005679809022694826, + 0.006757709663361311, + 0.01625596545636654, + 0.006395786069333553, + 0.008815932087600231, + 0.010686229914426804, + 0.009984143078327179, + 0.012990452349185944, + 0.008721590973436832, + 0.00660365168005228, + 0.005886247847229242, + 0.005622770171612501, + 0.005118459928780794, + 0.013540571555495262, + 0.017968855798244476, + 0.012808180414140224, + 0.015836799517273903, + 0.010647359304130077, + 0.012246441096067429, + 0.014857834205031395, + 0.008836859837174416, + 0.009495453909039497, + 0.014221923425793648, + 0.01014687679708004, + 0.00428400281816721, + 0.014490457251667976, + 0.012641899287700653, + 0.006790023762732744, + 0.016458820551633835, + 0.012138234451413155, + 0.008663538843393326, + 0.00729743205010891, + 0.006817808840423822, + 0.011955122463405132, + 0.00949218962341547, + 0.006033083889633417, + 0.008500787429511547, + 0.06508097052574158, + 0.07627780735492706, + 0.02378816157579422, + 0.011615248396992683, + 0.013805962167680264, + 0.009896418079733849, + 0.006620253901928663, + 0.009323571808636189, + 0.008945229463279247, + 0.00836370512843132, + 0.008446413092315197, + 0.008211466483771801, + 0.005903156474232674, + 0.004192963242530823, + 0.004766904283314943, + 0.004854691680520773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.11294235289096832 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_15/index_tensor([1012, 1012])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_15/index_tensor([1012, 1012])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a3abed4e79771b2a9e8b94eba1ba10d9081172d4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_15/index_tensor([1012, 1012])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a static, extreme close-up shot from a slightly low angle, the quiet miracle of spring unfolds. A dormant branch, its bark a tapestry of deep, weathered brown and rough, cracked texture, fills the frame against a soft, blurred backdrop of soft greens and earthy tones. Bathed in the warm, golden light of early morning, a swollen bud becomes the center of the action. A new shoot steadily and deliberately pushes forward from its tip, its vibrant, almost translucent lime-green tip piercing the bud's papery outer shell. Two tiny, tightly curled leaves, glistening with tiny dewdrops, begin to unfurl with a slow, relentless grace. As backlight illuminates them from behind, the subtle movement reveals their intricate new vessels, giving their edges a soft, ethereal glow and creating a tranquil and hopeful atmosphere centered on this powerful, silent act of renewal.", + "seed": 7918, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_15/index_1012_seed_7918_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03103164113126695, + "vlm_logprob": -0.03103164113126695, + "vlm_yes_no_margin": 4.599999976158142, + "vlm_token_credit": -0.03103164113126695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010851067490875721 + ], + "opposite_logprob_tokens": [ + -5.01085090637207 + ], + "logprob_reward": -0.010851067490875721, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010851067490875721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08949749916791916, + 0.03069247119128704, + 0.018922677263617516, + 0.20363274216651917, + 0.20197299122810364, + 0.0414997823536396, + 0.30744874477386475, + 0.07291218638420105, + 0.0142357824370265, + 0.019185135141015053 + ], + "credit_map": [ + 0.011415988206863403, + 0.011258510872721672, + 0.011784433387219906, + 0.010728100314736366, + 0.011635344475507736, + 0.006059551611542702, + 0.017066089436411858, + 0.009549479000270367, + 0.0035035433247685432, + 0.00387197220697999, + 0.0032074127811938524, + 0.004076430574059486, + 0.002819087589159608, + 0.0028603977989405394, + 0.006386903114616871, + 0.003966721706092358, + 0.0028936737217009068, + 0.0026625944301486015, + 0.002788232872262597, + 0.0019781277514994144, + 0.0019474539440125227, + 0.0016995106125250459, + 0.002609856193885207, + 0.002343226922675967, + 0.07759978622198105, + 0.07733026146888733, + 0.0067970543168485165, + 0.006704490631818771, + 0.009980127215385437, + 0.00488631846383214, + 0.01187456026673317, + 0.008460153825581074, + 0.07638762146234512, + 0.07665938884019852, + 0.006923843640834093, + 0.006647717207670212, + 0.00852207001298666, + 0.010110135190188885, + 0.006990782916545868, + 0.009731420315802097, + 0.012760353274643421, + 0.014443546533584595, + 0.0023136748932302, + 0.0016258680261671543, + 0.0037605564575642347, + 0.0029971140902489424, + 0.001821674988605082, + 0.0017769940895959735, + 0.11675318330526352, + 0.12755443155765533, + 0.009774720296263695, + 0.008599854074418545, + 0.016230473294854164, + 0.008326156064867973, + 0.009999183006584644, + 0.010210715234279633, + 0.009861634112894535, + 0.014043590053915977, + 0.004859024193137884, + 0.007689913734793663, + 0.007743651047348976, + 0.009749772027134895, + 0.01223294623196125, + 0.006731651723384857, + 0.0020168665796518326, + 0.0023555157240480185, + 0.0010971492156386375, + 0.001708389725536108, + 0.002208818681538105, + 0.0018143699271604419, + 0.001761108753271401, + 0.0012735630152747035, + 0.00251127895899117, + 0.0023717882577329874, + 0.002083687111735344, + 0.003218847792595625, + 0.0023910487070679665, + 0.0019531743600964546, + 0.002658816520124674, + 0.0019964929670095444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.010851067490875721 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707598507404327 + ], + "opposite_logprob_tokens": [ + -2.1270759105682373 + ], + "logprob_reward": -0.12707598507404327, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12707598507404327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07016292214393616, + 0.01940089277923107, + 0.023711124435067177, + 0.05775370076298714, + 0.0480307899415493, + 0.023014932870864868, + 0.659907341003418, + 0.05873478204011917, + 0.017030732706189156, + 0.022252826020121574 + ], + "credit_map": [ + 0.010965042747557163, + 0.009310542605817318, + 0.00682713882997632, + 0.0062433225102722645, + 0.010761908255517483, + 0.00843128003180027, + 0.009647381491959095, + 0.00797629076987505, + 0.003215810749679804, + 0.0025160745717585087, + 0.001896890695206821, + 0.0018581129843369126, + 0.0028567775152623653, + 0.0013657426461577415, + 0.0032675573602318764, + 0.002423923695459962, + 0.006617646664381027, + 0.0073490929789841175, + 0.0016857729060575366, + 0.0012961330357939005, + 0.002285904251039028, + 0.0014216004638001323, + 0.0015885788016021252, + 0.0014663934707641602, + 0.021211694926023483, + 0.020098920911550522, + 0.0032076258212327957, + 0.002222445560619235, + 0.002947188215330243, + 0.0020233436953276396, + 0.003450192976742983, + 0.0025922798085957766, + 0.017837604507803917, + 0.02009592391550541, + 0.0013637980446219444, + 0.0013182449620217085, + 0.0023547611199319363, + 0.0014962013810873032, + 0.0015576878795400262, + 0.0020065607968717813, + 0.006061030086129904, + 0.0070023504085838795, + 0.001349612488411367, + 0.001059681293554604, + 0.003456587204709649, + 0.001331970444880426, + 0.0016194905620068312, + 0.0011342063080519438, + 0.26548951864242554, + 0.28334981203079224, + 0.012094083242118359, + 0.011829887516796589, + 0.033996038138866425, + 0.016939720138907433, + 0.017285631969571114, + 0.018922582268714905, + 0.015164160169661045, + 0.014040157198905945, + 0.0042302547954022884, + 0.003599002491682768, + 0.00561595568433404, + 0.007262743543833494, + 0.004252181388437748, + 0.004570323508232832, + 0.0022197612561285496, + 0.0035128039307892323, + 0.0016725144814699888, + 0.0019782099407166243, + 0.002962375758215785, + 0.0016818078001961112, + 0.0016777669079601765, + 0.0013254891382530332, + 0.0030581944156438112, + 0.003319874871522188, + 0.0033781039528548717, + 0.0025802170857787132, + 0.002349038142710924, + 0.0034306959714740515, + 0.0014263192424550653, + 0.002710379660129547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.12707598507404327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200207905843854 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.002200207905843854, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200207905843854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15554402768611908, + 0.10944900661706924, + 0.07099154591560364, + 0.1577029675245285, + 0.03957488015294075, + 0.031751953065395355, + 0.3541215658187866, + 0.036562856286764145, + 0.01258437242358923, + 0.0317167267203331 + ], + "credit_map": [ + 0.020860997959971428, + 0.01988454908132553, + 0.01721007376909256, + 0.01518555823713541, + 0.0198814794421196, + 0.014955964870750904, + 0.024940311908721924, + 0.02262509986758232, + 0.022083964198827744, + 0.015859749168157578, + 0.012817579321563244, + 0.011330826207995415, + 0.007985014468431473, + 0.01633952185511589, + 0.009734850376844406, + 0.013297510333359241, + 0.011422345414757729, + 0.01474271435290575, + 0.006270149722695351, + 0.00454386742785573, + 0.01095971092581749, + 0.005706949159502983, + 0.011388984508812428, + 0.005956831853836775, + 0.06465422362089157, + 0.06238382309675217, + 0.004407204687595367, + 0.004019932821393013, + 0.006177157163619995, + 0.0030499997083097696, + 0.007400899659842253, + 0.005609756801277399, + 0.012075187638401985, + 0.014461982995271683, + 0.0018804637948051095, + 0.0019849110394716263, + 0.0026463319081813097, + 0.0021243374794721603, + 0.0024632178246974945, + 0.0019384488696232438, + 0.005328434519469738, + 0.0045347330160439014, + 0.0034424299374222755, + 0.0029800196643918753, + 0.004605371970683336, + 0.0036626728251576424, + 0.0032767001539468765, + 0.00392159353941679, + 0.14000457525253296, + 0.15064363181591034, + 0.00756312720477581, + 0.007267060223966837, + 0.018715955317020416, + 0.00939224660396576, + 0.010602407157421112, + 0.009932590648531914, + 0.005650762002915144, + 0.006987208500504494, + 0.0025035187136381865, + 0.004099284298717976, + 0.00407441146671772, + 0.00448622927069664, + 0.005386436358094215, + 0.003375007538124919, + 0.001609462546184659, + 0.002260516630485654, + 0.0014837234048172832, + 0.0011751396814361215, + 0.001919255475513637, + 0.0018508194480091333, + 0.0009831094648689032, + 0.0013023470528423786, + 0.005421302281320095, + 0.005300655495375395, + 0.0027208770625293255, + 0.004976772703230381, + 0.0042177531868219376, + 0.003579346928745508, + 0.002795739099383354, + 0.002704283921048045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.002200207905843854 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252134520560503 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252134520560503, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252134520560503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11782636493444443, + 0.028345979750156403, + 0.11852476000785828, + 0.020858455449342728, + 0.1373332142829895, + 0.0456090047955513, + 0.34003305435180664, + 0.11349762976169586, + 0.020982464775443077, + 0.056989073753356934 + ], + "credit_map": [ + 0.015015706419944763, + 0.015298246406018734, + 0.013869508169591427, + 0.012092637829482555, + 0.01913847029209137, + 0.009906531311571598, + 0.020255843177437782, + 0.012249430641531944, + 0.003393129911273718, + 0.005087409168481827, + 0.002313898876309395, + 0.00313713401556015, + 0.0017808196134865284, + 0.0017233677208423615, + 0.006066407077014446, + 0.004843815695494413, + 0.024259958416223526, + 0.030825022608041763, + 0.00854924600571394, + 0.006778584327548742, + 0.015745384618639946, + 0.008146613836288452, + 0.015295417048037052, + 0.008924541063606739, + 0.004158356226980686, + 0.0034309530165046453, + 0.002025576075538993, + 0.0023831806611269712, + 0.0025938930921256542, + 0.0016154727200046182, + 0.002722491044551134, + 0.0019285340094938874, + 0.05408311262726784, + 0.05604668706655502, + 0.003359269816428423, + 0.004138735588639975, + 0.005732244811952114, + 0.004049867391586304, + 0.004477859940379858, + 0.0054454561322927475, + 0.008273957297205925, + 0.011119889095425606, + 0.004088442772626877, + 0.0034477333538234234, + 0.007442283444106579, + 0.003803324420005083, + 0.003911246079951525, + 0.0035221290308982134, + 0.13600558042526245, + 0.1430804580450058, + 0.00722605362534523, + 0.00645598815754056, + 0.017616458237171173, + 0.009643190540373325, + 0.008966878987848759, + 0.011038442142307758, + 0.021989855915308, + 0.027490627020597458, + 0.00655011460185051, + 0.011331072077155113, + 0.013324594125151634, + 0.01074079517275095, + 0.012374192476272583, + 0.009696386754512787, + 0.0026284113992005587, + 0.002920969622209668, + 0.001279096701182425, + 0.002642381936311722, + 0.0023897220380604267, + 0.0030562663450837135, + 0.003410683013498783, + 0.0026549347676336765, + 0.009655162692070007, + 0.007653445936739445, + 0.005687718745321035, + 0.009110675193369389, + 0.005464581307023764, + 0.007065671496093273, + 0.005829242058098316, + 0.006522581912577152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.005252134520560503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778810665011406 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778810665011406, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778810665011406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1471136510372162, + 0.11864716559648514, + 0.042787402868270874, + 0.18183544278144836, + 0.15601825714111328, + 0.04568176344037056, + 0.15011662244796753, + 0.04585355147719383, + 0.0320383682847023, + 0.07990781217813492 + ], + "credit_map": [ + 0.02364952862262726, + 0.019101466983556747, + 0.016181036829948425, + 0.01440390944480896, + 0.020304406061768532, + 0.014293735846877098, + 0.021786771714687347, + 0.017392784357070923, + 0.028867531567811966, + 0.023258443921804428, + 0.011313661932945251, + 0.01146876160055399, + 0.007290243171155453, + 0.009838140569627285, + 0.015567073598504066, + 0.011043298989534378, + 0.01227988675236702, + 0.014347639866173267, + 0.0033613070845603943, + 0.002599458210170269, + 0.0023535671643912792, + 0.0021815472282469273, + 0.0025986030232161283, + 0.003065392840653658, + 0.07405421137809753, + 0.07271922379732132, + 0.004204129334539175, + 0.004194510634988546, + 0.007491504307836294, + 0.0040886723436415195, + 0.00874058436602354, + 0.006342614069581032, + 0.06093331798911095, + 0.0656294971704483, + 0.0033359581138938665, + 0.004030569456517696, + 0.007019234821200371, + 0.004415339790284634, + 0.0049439179711043835, + 0.005710398778319359, + 0.015290028415620327, + 0.01561540924012661, + 0.0020686176139861345, + 0.002098179655149579, + 0.003141077235341072, + 0.002710675122216344, + 0.0021100598387420177, + 0.0026477170176804066, + 0.055779632180929184, + 0.0612419955432415, + 0.004576260689646006, + 0.0048338002525269985, + 0.009034277871251106, + 0.00428071478381753, + 0.0056054433807730675, + 0.004764502868056297, + 0.008620161563158035, + 0.009766815230250359, + 0.004311049357056618, + 0.005105963908135891, + 0.005167342722415924, + 0.004833482671529055, + 0.0038929800502955914, + 0.0041557601653039455, + 0.0050887418910861015, + 0.005791880190372467, + 0.0033392314799129963, + 0.0026860388461500406, + 0.005056250840425491, + 0.0045946938917040825, + 0.0024454332888126373, + 0.00303609948605299, + 0.011403306387364864, + 0.010392382740974426, + 0.011182431131601334, + 0.010802446864545345, + 0.008633547462522984, + 0.011662513017654419, + 0.006895073223859072, + 0.008936119265854359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.009778810665011406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, extreme close-up shot from a slightly low angle, the quiet miracle of spring unfolds. A dormant branch, its bark a tapestry of deep, weathered brown and rough, cracked texture, fills the frame against a soft, blurred backdrop of soft greens and earthy tones. Bathed in the warm, golden light of early morning, a swollen bud becomes the center of the action. A new shoot steadily and deliberately pushes forward from its tip, its vibrant, almost translucent lime-green tip piercing the bud's papery outer shell. Two tiny, tightly curled leaves, glistening with tiny dewdrops, begin to unfurl with a slow, relentless grace. As backlight illuminates them from behind, the subtle movement reveals their intricate new vessels, giving their edges a soft, ethereal glow and creating a tranquil and hopeful atmosphere centered on this powerful, silent act of renewal.", + "seed": 7919, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_15/index_1012_seed_7919_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022630336321890356, + "vlm_logprob": -0.022630336321890356, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.022630336321890356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016170891001820564 + ], + "opposite_logprob_tokens": [ + -4.391170978546143 + ], + "logprob_reward": -0.016170891001820564, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016170891001820564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17203962802886963, + 0.03565146401524544, + 0.015881501138210297, + 0.017953012138605118, + 0.2324284315109253, + 0.06592478603124619, + 0.051772817969322205, + 0.3600492477416992, + 0.02979477308690548, + 0.01850435696542263 + ], + "credit_map": [ + 0.033257994800806046, + 0.034361179918050766, + 0.01175014954060316, + 0.019433043897151947, + 0.014242896810173988, + 0.015481904149055481, + 0.025755122303962708, + 0.017757318913936615, + 0.006800902541726828, + 0.0050529129803180695, + 0.003066467819735408, + 0.0021284976974129677, + 0.006781009491533041, + 0.0027976459823548794, + 0.0037184604443609715, + 0.005305567756295204, + 0.0019764406606554985, + 0.0014977083774283528, + 0.001531456713564694, + 0.0017921376274898648, + 0.0020619919523596764, + 0.0027195813599973917, + 0.0020862771198153496, + 0.0022159074433147907, + 0.003399851266294718, + 0.002012448152527213, + 0.0024110269732773304, + 0.0022444718051701784, + 0.002040138002485037, + 0.002174269175156951, + 0.0015508384676650167, + 0.002119968645274639, + 0.0323818139731884, + 0.0327235646545887, + 0.018222764134407043, + 0.043940912932157516, + 0.02828257903456688, + 0.03129710257053375, + 0.02290218509733677, + 0.022677486762404442, + 0.010438736528158188, + 0.008145861327648163, + 0.0084322365000844, + 0.008292214944958687, + 0.0073693436570465565, + 0.010474633425474167, + 0.005007739178836346, + 0.007764019537717104, + 0.012826213613152504, + 0.005858834832906723, + 0.0059255678206682205, + 0.004688582848757505, + 0.005448552779853344, + 0.0035762775223702192, + 0.0044297752901911736, + 0.009019016288220882, + 0.11845090985298157, + 0.02538272924721241, + 0.032025787979364395, + 0.025418750941753387, + 0.026750056073069572, + 0.019747646525502205, + 0.03323759511113167, + 0.07903576642274857, + 0.011062661185860634, + 0.002106679603457451, + 0.002043503802269697, + 0.0014571619685739279, + 0.001307113328948617, + 0.00227047735825181, + 0.002049707341939211, + 0.0074974666349589825, + 0.0040385047905147076, + 0.0016430731629952788, + 0.0025805593468248844, + 0.0014090673066675663, + 0.0016053835861384869, + 0.0034069090615957975, + 0.0012207472464069724, + 0.0026001110672950745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.016170891001820564 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789945125579834 + ], + "opposite_logprob_tokens": [ + -2.5789945125579834 + ], + "logprob_reward": -0.0789945125579834, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789945125579834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2068261057138443, + 0.026442160829901695, + 0.029244672507047653, + 0.024146350100636482, + 0.1080174595117569, + 0.03966273367404938, + 0.1442466825246811, + 0.2881307303905487, + 0.10984203964471817, + 0.023441040888428688 + ], + "credit_map": [ + 0.058686450123786926, + 0.0554894395172596, + 0.014446710236370564, + 0.011016834527254105, + 0.009860679507255554, + 0.021373802796006203, + 0.012066447176039219, + 0.02388574555516243, + 0.003927667625248432, + 0.002485929988324642, + 0.0025300097186118364, + 0.004047407768666744, + 0.0028725084848701954, + 0.0031150691211223602, + 0.0036795749329030514, + 0.0037839936558157206, + 0.005359837785363197, + 0.0024248184636235237, + 0.0028125611133873463, + 0.004481195006519556, + 0.003774314885959029, + 0.003114698687568307, + 0.0032483332324773073, + 0.004028913099318743, + 0.00531029375270009, + 0.0018370476318523288, + 0.0031955945305526257, + 0.0021789241582155228, + 0.0027505711186677217, + 0.002795333508402109, + 0.002798910718411207, + 0.0032796720042824745, + 0.013582996092736721, + 0.012304139323532581, + 0.013509167358279228, + 0.015830885618925095, + 0.010101055726408958, + 0.011497113853693008, + 0.017672529444098473, + 0.013519569300115108, + 0.00496431440114975, + 0.0034887029323726892, + 0.007107221987098455, + 0.00504005141556263, + 0.0034818968269973993, + 0.006601149681955576, + 0.005027072038501501, + 0.003952324390411377, + 0.05647234991192818, + 0.010427584871649742, + 0.007245102431625128, + 0.007501260843127966, + 0.0054597025737166405, + 0.0064555201679468155, + 0.009945682249963284, + 0.04073946550488472, + 0.09386436641216278, + 0.023010848090052605, + 0.023597147315740585, + 0.022428074851632118, + 0.022384673357009888, + 0.014733434654772282, + 0.025417136028409004, + 0.06269506365060806, + 0.0403951033949852, + 0.008705352433025837, + 0.010437730699777603, + 0.004859610460698605, + 0.00593407591804862, + 0.006266383919864893, + 0.007003349717706442, + 0.026240436360239983, + 0.004530309233814478, + 0.0018936449196189642, + 0.004912942182272673, + 0.0015506716445088387, + 0.001997079700231552, + 0.004365295171737671, + 0.0016965538961812854, + 0.0024945430923253298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0789945125579834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255692172795534 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255692172795534, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255692172795534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32665860652923584, + 0.04932465776801109, + 0.04368952289223671, + 0.03258313611149788, + 0.07684292644262314, + 0.10862486809492111, + 0.06217026710510254, + 0.11461638659238815, + 0.14578987658023834, + 0.039699725806713104 + ], + "credit_map": [ + 0.11015940457582474, + 0.11500141769647598, + 0.011439884081482887, + 0.012259667739272118, + 0.02400735579431057, + 0.012172166258096695, + 0.017009928822517395, + 0.0246087945997715, + 0.00704736215993762, + 0.006881185341626406, + 0.0042191543616354465, + 0.004139035940170288, + 0.011171435937285423, + 0.004704060964286327, + 0.006820706184953451, + 0.004341721534729004, + 0.006157983094453812, + 0.003912213258445263, + 0.004556693602353334, + 0.00623579416424036, + 0.005285841412842274, + 0.007474434096366167, + 0.005304296035319567, + 0.00476226769387722, + 0.0056020841002464294, + 0.003336656140163541, + 0.004294604063034058, + 0.004381085280328989, + 0.004048634320497513, + 0.003970912657678127, + 0.003575571347028017, + 0.003373589599505067, + 0.009911089204251766, + 0.011106441728770733, + 0.008004914969205856, + 0.011924710124731064, + 0.009438187815248966, + 0.010497166775166988, + 0.008714201860129833, + 0.007246215362101793, + 0.0274363961070776, + 0.009856468997895718, + 0.011596170254051685, + 0.012239404954016209, + 0.0069427527487277985, + 0.008191519416868687, + 0.010694535449147224, + 0.02166762389242649, + 0.015455275774002075, + 0.0034102341160178185, + 0.008277589455246925, + 0.005734047386795282, + 0.00400746101513505, + 0.007407306227833033, + 0.0070844306610524654, + 0.010793921537697315, + 0.0358595997095108, + 0.00825407449156046, + 0.011927824467420578, + 0.008141230791807175, + 0.008575065061450005, + 0.008513059467077255, + 0.00959926750510931, + 0.023746255785226822, + 0.058041490614414215, + 0.0113223260268569, + 0.007996825501322746, + 0.005968017503619194, + 0.0070084575563669205, + 0.007589838933199644, + 0.008641282096505165, + 0.039221636950969696, + 0.006611446850001812, + 0.0025765225291252136, + 0.007026002276688814, + 0.0030620049219578505, + 0.0025406673084944487, + 0.009431110695004463, + 0.002906528767198324, + 0.005545440129935741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.005255692172795534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095145035535097 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095145035535097, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095145035535097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023762734606862068, + 0.05597447231411934, + 0.021551283076405525, + 0.026635175570845604, + 0.06277689337730408, + 0.176926851272583, + 0.17077524960041046, + 0.26825979351997375, + 0.16532015800476074, + 0.028017273172736168 + ], + "credit_map": [ + 0.0032024255488067865, + 0.003265144070610404, + 0.002382984384894371, + 0.003267297288402915, + 0.0040098316967487335, + 0.0021890136413276196, + 0.0029520285315811634, + 0.002494011539965868, + 0.008948034606873989, + 0.006422187201678753, + 0.0030512085650116205, + 0.008028141222894192, + 0.010041773319244385, + 0.00650990242138505, + 0.007380489259958267, + 0.005592742469161749, + 0.0031477664597332478, + 0.002419087802991271, + 0.0017643352039158344, + 0.004016585182398558, + 0.0030596356373280287, + 0.0022785605397075415, + 0.0019372539827600121, + 0.0029280614107847214, + 0.005600765813142061, + 0.0027556554414331913, + 0.002341347513720393, + 0.0027999780140817165, + 0.0026326749939471483, + 0.0029283487237989902, + 0.002595895202830434, + 0.004980515688657761, + 0.00762302428483963, + 0.00767883425578475, + 0.004923573695123196, + 0.013577520847320557, + 0.007069475948810577, + 0.008867780677974224, + 0.006537686102092266, + 0.006499006878584623, + 0.05041448026895523, + 0.014854710549116135, + 0.01713956892490387, + 0.017966000363230705, + 0.010872547514736652, + 0.012745345011353493, + 0.013903273269534111, + 0.03903096541762352, + 0.06488698720932007, + 0.01235781516879797, + 0.010354176163673401, + 0.008957738988101482, + 0.007190709933638573, + 0.00863041914999485, + 0.012521758675575256, + 0.0458756648004055, + 0.09045452624559402, + 0.02032478153705597, + 0.022153180092573166, + 0.019624970853328705, + 0.0203586146235466, + 0.012634575366973877, + 0.02156613953411579, + 0.06114305928349495, + 0.06108463183045387, + 0.013707617297768593, + 0.008979979902505875, + 0.00896479468792677, + 0.010855271480977535, + 0.009022708050906658, + 0.011455840431153774, + 0.04124932736158371, + 0.004304414615035057, + 0.004318032413721085, + 0.0027864628937095404, + 0.0031096164602786303, + 0.004765738267451525, + 0.003313013818114996, + 0.002660991856828332, + 0.002759008901193738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.004095145035535097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635440841317177 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635440841317177, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635440841317177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.149145245552063, + 0.12523844838142395, + 0.029425591230392456, + 0.021758325397968292, + 0.12833990156650543, + 0.043253179639577866, + 0.03913309425115585, + 0.18037493526935577, + 0.24748121201992035, + 0.035850077867507935 + ], + "credit_map": [ + 0.026554903015494347, + 0.029270155355334282, + 0.01040792465209961, + 0.014144561253488064, + 0.02062642015516758, + 0.012402087450027466, + 0.0207295473664999, + 0.015009637922048569, + 0.01774008199572563, + 0.01830488257110119, + 0.008727152831852436, + 0.010469322092831135, + 0.029696667566895485, + 0.012701984494924545, + 0.01610603742301464, + 0.011492306366562843, + 0.0040017832070589066, + 0.002954624593257904, + 0.004279824905097485, + 0.003381694434210658, + 0.0033085087779909372, + 0.0037722166161984205, + 0.003903585020452738, + 0.003823351813480258, + 0.0030756911728531122, + 0.0023142548743635416, + 0.002942553721368313, + 0.0037723714485764503, + 0.003327186917886138, + 0.002442975528538227, + 0.001655889325775206, + 0.002227399032562971, + 0.016412831842899323, + 0.02000110037624836, + 0.010588499717414379, + 0.02352163940668106, + 0.01644558645784855, + 0.016336193308234215, + 0.013361930847167969, + 0.01167211402207613, + 0.006101783365011215, + 0.0063329096883535385, + 0.0051363930106163025, + 0.006116992793977261, + 0.0059567661955952644, + 0.005198901519179344, + 0.004049256443977356, + 0.004360172897577286, + 0.006018765736371279, + 0.0026789712719619274, + 0.006650382187217474, + 0.005114058963954449, + 0.0033159523736685514, + 0.006897611543536186, + 0.004152997862547636, + 0.004304348956793547, + 0.05866112932562828, + 0.01425886433571577, + 0.013615885749459267, + 0.0145408408716321, + 0.014112730510532856, + 0.011605623178184032, + 0.014312783256173134, + 0.03926706314086914, + 0.09690488129854202, + 0.018260879442095757, + 0.013147755526006222, + 0.011212659068405628, + 0.013217979110777378, + 0.015756649896502495, + 0.014903104864060879, + 0.06407728046178818, + 0.004313935991376638, + 0.004966333042830229, + 0.005427642725408077, + 0.004956644028425217, + 0.005029236897826195, + 0.0045007988810539246, + 0.003495048498734832, + 0.00316043384373188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.008635440841317177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, extreme close-up shot from a slightly low angle, the quiet miracle of spring unfolds. A dormant branch, its bark a tapestry of deep, weathered brown and rough, cracked texture, fills the frame against a soft, blurred backdrop of soft greens and earthy tones. Bathed in the warm, golden light of early morning, a swollen bud becomes the center of the action. A new shoot steadily and deliberately pushes forward from its tip, its vibrant, almost translucent lime-green tip piercing the bud's papery outer shell. Two tiny, tightly curled leaves, glistening with tiny dewdrops, begin to unfurl with a slow, relentless grace. As backlight illuminates them from behind, the subtle movement reveals their intricate new vessels, giving their edges a soft, ethereal glow and creating a tranquil and hopeful atmosphere centered on this powerful, silent act of renewal.", + "seed": 7920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_15/index_1012_seed_7920_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03548981044441461, + "vlm_logprob": -0.03548981044441461, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03548981044441461, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01802375726401806 + ], + "opposite_logprob_tokens": [ + -4.14302396774292 + ], + "logprob_reward": -0.01802375726401806, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01802375726401806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3000026047229767, + 0.12345165759325027, + 0.0827735960483551, + 0.02412995882332325, + 0.02267715521156788, + 0.0286724753677845, + 0.2379465103149414, + 0.09211450815200806, + 0.02011796273291111, + 0.0681135505437851 + ], + "credit_map": [ + 0.09570859372615814, + 0.020716393366456032, + 0.01328621432185173, + 0.013496298342943192, + 0.10016585141420364, + 0.026530198752880096, + 0.01313116867095232, + 0.01696789264678955, + 0.03903606906533241, + 0.01705796644091606, + 0.007948007434606552, + 0.014980456791818142, + 0.017169075086712837, + 0.015084889717400074, + 0.006318908650428057, + 0.0058562858030200005, + 0.014001728035509586, + 0.008125916123390198, + 0.00661152508109808, + 0.010162935592234135, + 0.006858856417238712, + 0.015371175482869148, + 0.01179237850010395, + 0.009849081747233868, + 0.004401446785777807, + 0.0027431310154497623, + 0.0025723285507410765, + 0.003523575607687235, + 0.0026363490615040064, + 0.002745726378634572, + 0.0018322585383430123, + 0.0036751430016011, + 0.003363352734595537, + 0.001968286233022809, + 0.00388049497269094, + 0.0028795753605663776, + 0.0020661866292357445, + 0.002627558773383498, + 0.002248425269499421, + 0.0036432736087590456, + 0.005382836796343327, + 0.002368505811318755, + 0.0029002123046666384, + 0.0034618370700627565, + 0.0018848462495952845, + 0.005932929925620556, + 0.0024205411318689585, + 0.004320764448493719, + 0.08859512954950333, + 0.019167257472872734, + 0.005398268811404705, + 0.008171559311449528, + 0.010116408579051495, + 0.09394635260105133, + 0.005833388306200504, + 0.006718131247907877, + 0.012515759095549583, + 0.007069350220263004, + 0.008859256282448769, + 0.00839377474039793, + 0.006016026251018047, + 0.019536377862095833, + 0.009812637232244015, + 0.019911320880055428, + 0.0034854738041758537, + 0.0014873191248625517, + 0.0021046146284788847, + 0.002005799440667033, + 0.0015738310758024454, + 0.0039792670868337154, + 0.0016277633840218186, + 0.0038538947701454163, + 0.009608009830117226, + 0.004511548206210136, + 0.008720490150153637, + 0.007559224031865597, + 0.005131279118359089, + 0.00909407064318657, + 0.00974554754793644, + 0.013743379153311253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.01802375726401806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270873248577118 + ], + "opposite_logprob_tokens": [ + -2.127087354660034 + ], + "logprob_reward": -0.1270873248577118, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270873248577118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15081019699573517, + 0.022383658215403557, + 0.029537610709667206, + 0.08174704760313034, + 0.060637716203927994, + 0.05659543722867966, + 0.5679042339324951, + 0.006900853011757135, + 0.004329763352870941, + 0.01915348507463932 + ], + "credit_map": [ + 0.05599068105220795, + 0.011984867043793201, + 0.005348284263163805, + 0.005719607230275869, + 0.05113128945231438, + 0.010095751844346523, + 0.004215070512145758, + 0.006324649788439274, + 0.00465715816244483, + 0.003147844225168228, + 0.0015823426656425, + 0.0035434411838650703, + 0.00257657328620553, + 0.0025302590802311897, + 0.0018332615727558732, + 0.0025127765256911516, + 0.005496810656040907, + 0.0024443906731903553, + 0.0033064086455851793, + 0.00258048833347857, + 0.0033555945847183466, + 0.005490919109433889, + 0.0030391530599445105, + 0.003823845647275448, + 0.016253512352705002, + 0.009745091199874878, + 0.007241688668727875, + 0.01213728915899992, + 0.006860205437988043, + 0.014094957150518894, + 0.008669830858707428, + 0.006744469981640577, + 0.008746794424951077, + 0.00465446338057518, + 0.008051981218159199, + 0.009069216437637806, + 0.005258768331259489, + 0.008566189557313919, + 0.006128471810370684, + 0.010161831974983215, + 0.010725731961429119, + 0.005129099357873201, + 0.005133475176990032, + 0.009234616532921791, + 0.005220782943069935, + 0.009329874068498611, + 0.006321948952972889, + 0.005499902181327343, + 0.22715304791927338, + 0.04623924195766449, + 0.010482088662683964, + 0.015982860699295998, + 0.019229769706726074, + 0.22920268774032593, + 0.009800678119063377, + 0.009813900105655193, + 0.0009613261790946126, + 0.0007575222407467663, + 0.0006561595946550369, + 0.0010040522320196033, + 0.0008043174748308957, + 0.0012954920530319214, + 0.0008291359408758581, + 0.0005928468890488148, + 0.0006971522816456854, + 0.0004516015178523958, + 0.00044573095510713756, + 0.0005391912418417633, + 0.00043103451025672257, + 0.0008086214656941593, + 0.0003461804299149662, + 0.0006102509796619415, + 0.003699705470353365, + 0.0016045892843976617, + 0.001793829957023263, + 0.0016909241676330566, + 0.001597499824129045, + 0.003277016570791602, + 0.002589570125564933, + 0.0029003487434238195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.1270873248577118 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641940541565418 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641940541565418, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641940541565418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22666867077350616, + 0.012774631381034851, + 0.012765471823513508, + 0.02394457720220089, + 0.02762654423713684, + 0.024868864566087723, + 0.6269396543502808, + 0.030685974285006523, + 0.003169344738125801, + 0.010556268505752087 + ], + "credit_map": [ + 0.0879150778055191, + 0.014197118580341339, + 0.005603299010545015, + 0.006414379458874464, + 0.08457905799150467, + 0.016589023172855377, + 0.005505445413291454, + 0.00586526608094573, + 0.0036154035478830338, + 0.0016484258230775595, + 0.0007747928611934185, + 0.001508719171397388, + 0.0011352559085935354, + 0.0021964507177472115, + 0.0011133780935779214, + 0.0007822048501111567, + 0.002548537217080593, + 0.0012516308343037963, + 0.001238022348843515, + 0.0008541765855625272, + 0.001431534648872912, + 0.003170120995491743, + 0.0011018702061846852, + 0.0011695778230205178, + 0.004076877608895302, + 0.0030765689443796873, + 0.0027644019573926926, + 0.0038999617099761963, + 0.0016473816940560937, + 0.0030682275537401438, + 0.0025511174462735653, + 0.0028600404039025307, + 0.004013035446405411, + 0.00210553128272295, + 0.0034966710954904556, + 0.0038799233734607697, + 0.0018988862866535783, + 0.003770762123167515, + 0.0028446295764297247, + 0.005617103539407253, + 0.0032091981265693903, + 0.0020741603802889585, + 0.0028583998791873455, + 0.0035792714916169643, + 0.002534236293286085, + 0.0034952424466609955, + 0.0035974953789263964, + 0.0035208589397370815, + 0.2500128149986267, + 0.05056268349289894, + 0.011664370074868202, + 0.01726631633937359, + 0.019632792100310326, + 0.2542116343975067, + 0.010064920410513878, + 0.013524092733860016, + 0.00466257706284523, + 0.0033325490076094866, + 0.0040410044603049755, + 0.003379177302122116, + 0.0033642658963799477, + 0.005088831298053265, + 0.003132300451397896, + 0.003685267409309745, + 0.00046151268179528415, + 0.00042019455577246845, + 0.0004607053706422448, + 0.0003978264576289803, + 0.00043749905307777226, + 0.00036961224395781755, + 0.0002366075204918161, + 0.0003853864618577063, + 0.001630463171750307, + 0.0007533945608884096, + 0.001105708535760641, + 0.0010920846834778786, + 0.0007676060195080936, + 0.0012959892628714442, + 0.0013890056870877743, + 0.0025220164097845554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.008641940541565418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629185914993286 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629185914993286, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629185914993286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10717744380235672, + 0.07608631998300552, + 0.036936089396476746, + 0.03565547987818718, + 0.03647157549858093, + 0.02799336425960064, + 0.6230714321136475, + 0.021815959364175797, + 0.008350824005901814, + 0.02644147351384163 + ], + "credit_map": [ + 0.03423493355512619, + 0.008702918887138367, + 0.004893105011433363, + 0.007213106378912926, + 0.03340565413236618, + 0.008936244063079357, + 0.005823053885251284, + 0.00396844232454896, + 0.021263310685753822, + 0.011867638677358627, + 0.004323453642427921, + 0.011275459080934525, + 0.008328771218657494, + 0.010549264959990978, + 0.003956321161240339, + 0.004522114992141724, + 0.005830473266541958, + 0.005323400255292654, + 0.003782109124585986, + 0.0062154619954526424, + 0.003916400019079447, + 0.005032621324062347, + 0.002563765272498131, + 0.0042718625627458096, + 0.006240135058760643, + 0.004287049639970064, + 0.0038776970468461514, + 0.004497164860367775, + 0.002812424208968878, + 0.005159739870578051, + 0.0027236449532210827, + 0.006057628430426121, + 0.004858294036239386, + 0.003636508248746395, + 0.00524304062128067, + 0.005322747863829136, + 0.003042440162971616, + 0.0047398838214576244, + 0.0038341088220477104, + 0.005794558208435774, + 0.004060780629515648, + 0.00233853398822248, + 0.003585290862247348, + 0.004392258822917938, + 0.002253776416182518, + 0.004700664896517992, + 0.0033837840892374516, + 0.0032782782800495625, + 0.2466091811656952, + 0.05112241208553314, + 0.01138679962605238, + 0.017721744254231453, + 0.021532999351620674, + 0.2511625289916992, + 0.011061699129641056, + 0.012474114075303078, + 0.0029092300683259964, + 0.002706756116822362, + 0.0038594333454966545, + 0.003391191828995943, + 0.002388987224549055, + 0.0024094278924167156, + 0.002186842728406191, + 0.001964094815775752, + 0.0011880345409736037, + 0.0006339571555145085, + 0.001249395776540041, + 0.0008156790863722563, + 0.0006258698413148522, + 0.001875642454251647, + 0.0007252702489495277, + 0.0012369764735922217, + 0.00416323309764266, + 0.002903401618823409, + 0.0021104540210217237, + 0.0024051188956946135, + 0.0029424279928207397, + 0.0041924770921468735, + 0.0031316308304667473, + 0.004592733923345804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.007629185914993286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066843643784523 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016066843643784523, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066843643784523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13260188698768616, + 0.031128615140914917, + 0.03915800154209137, + 0.03728054091334343, + 0.03465787693858147, + 0.02872508578002453, + 0.592565655708313, + 0.016298573464155197, + 0.023546287789940834, + 0.06403742730617523 + ], + "credit_map": [ + 0.03318468853831291, + 0.010909933596849442, + 0.0073882704600691795, + 0.011699766851961613, + 0.034924525767564774, + 0.014597688801586628, + 0.011010278016328812, + 0.008886752650141716, + 0.005089316982775927, + 0.005001535173505545, + 0.002707339357584715, + 0.0038449731655418873, + 0.004230656661093235, + 0.0044344570487737656, + 0.0025276364758610725, + 0.0032927049323916435, + 0.008551068603992462, + 0.0039038597606122494, + 0.003397855442017317, + 0.00460338220000267, + 0.0032191367354243994, + 0.007772317621856928, + 0.004310721065849066, + 0.0033996684942394495, + 0.00624841870740056, + 0.004509697202593088, + 0.003652194980531931, + 0.0071226852014660835, + 0.003642858238890767, + 0.004940632730722427, + 0.004409966990351677, + 0.002754094311967492, + 0.0047289906069636345, + 0.0044961231760680676, + 0.00454698083922267, + 0.005748208146542311, + 0.0030332908499985933, + 0.003777560079470277, + 0.004333839751780033, + 0.0039928872138261795, + 0.004614256788045168, + 0.002928365021944046, + 0.0029914516489952803, + 0.004197429399937391, + 0.003167724935337901, + 0.0037462501786649227, + 0.0032949564047157764, + 0.0037846562918275595, + 0.2390109896659851, + 0.0478559173643589, + 0.011041507124900818, + 0.01639874279499054, + 0.018958017230033875, + 0.23821650445461273, + 0.009615777060389519, + 0.01146828755736351, + 0.0018902237061411142, + 0.0014012581668794155, + 0.002883683890104294, + 0.0025069357361644506, + 0.0017991156782954931, + 0.0020156893879175186, + 0.0021222282666713, + 0.0016794431721791625, + 0.004342522006481886, + 0.0017650423105806112, + 0.003238492179661989, + 0.0026710410602390766, + 0.0021560522727668285, + 0.0036509549245238304, + 0.0017489682650193572, + 0.003973217681050301, + 0.010277806781232357, + 0.006090017035603523, + 0.005851801484823227, + 0.006160696502774954, + 0.006080085877329111, + 0.007555670104920864, + 0.00928017683327198, + 0.012741184793412685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.016066843643784523 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, extreme close-up shot from a slightly low angle, the quiet miracle of spring unfolds. A dormant branch, its bark a tapestry of deep, weathered brown and rough, cracked texture, fills the frame against a soft, blurred backdrop of soft greens and earthy tones. Bathed in the warm, golden light of early morning, a swollen bud becomes the center of the action. A new shoot steadily and deliberately pushes forward from its tip, its vibrant, almost translucent lime-green tip piercing the bud's papery outer shell. Two tiny, tightly curled leaves, glistening with tiny dewdrops, begin to unfurl with a slow, relentless grace. As backlight illuminates them from behind, the subtle movement reveals their intricate new vessels, giving their edges a soft, ethereal glow and creating a tranquil and hopeful atmosphere centered on this powerful, silent act of renewal.", + "seed": 7921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_15/index_1012_seed_7921_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.47001599352806805, + "vlm_logprob": -0.47001599352806805, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.47001599352806805, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.128065347671509 + ], + "opposite_logprob_tokens": [ + -0.12806527316570282 + ], + "logprob_reward": -2.128065347671509, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.128065347671509, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14234450459480286, + 0.4253568649291992, + 0.01877324841916561, + 0.04093204811215401, + 0.06148781627416611, + 0.11911406368017197, + 0.052896540611982346, + 0.04348902776837349, + 0.06722643971443176, + 0.02837945707142353 + ], + "credit_map": [ + 0.02109338901937008, + 0.021906733512878418, + 0.019962534308433533, + 0.014693209901452065, + 0.017777683213353157, + 0.008576534688472748, + 0.0225346852093935, + 0.015799738466739655, + 0.12269484996795654, + 0.06008320674300194, + 0.011142479255795479, + 0.008474894799292088, + 0.18237358331680298, + 0.018339065834879875, + 0.010282875038683414, + 0.011965898796916008, + 0.0018775869393721223, + 0.0018257764168083668, + 0.0028149965219199657, + 0.0015836323145776987, + 0.0025823975447565317, + 0.001161025371402502, + 0.004272876773029566, + 0.0026549578178673983, + 0.007158445660024881, + 0.004239171743392944, + 0.002120101125910878, + 0.005880958866328001, + 0.00926905032247305, + 0.0028324162121862173, + 0.0054630013182759285, + 0.0039689005352556705, + 0.007010261062532663, + 0.00888772588223219, + 0.00467393035069108, + 0.0072780996561050415, + 0.00744392117485404, + 0.006226304918527603, + 0.008426827378571033, + 0.011540746316313744, + 0.010133500210940838, + 0.0321010947227478, + 0.012426535598933697, + 0.011271200142800808, + 0.02493276447057724, + 0.007194977719336748, + 0.008527601137757301, + 0.012526394799351692, + 0.008985171094536781, + 0.006781332660466433, + 0.006003112997859716, + 0.004723138641566038, + 0.008633549325168133, + 0.004536070860922337, + 0.006455295719206333, + 0.006778866518288851, + 0.005572840105742216, + 0.004570464137941599, + 0.006436456460505724, + 0.005064436234533787, + 0.006173132453113794, + 0.0039078472182154655, + 0.005394614301621914, + 0.006369240581989288, + 0.007872742600739002, + 0.007036454509943724, + 0.0071705663576722145, + 0.005040809046477079, + 0.009554126299917698, + 0.010916287079453468, + 0.008872679434716702, + 0.010762776248157024, + 0.002918780315667391, + 0.0026984072756022215, + 0.0032188808545470238, + 0.004296391736716032, + 0.003982975147664547, + 0.0036098521668463945, + 0.003944499883800745, + 0.003709672251716256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -2.128065347671509 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1429498791694641 + ], + "opposite_logprob_tokens": [ + -2.0179498195648193 + ], + "logprob_reward": -0.1429498791694641, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1429498791694641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10402879118919373, + 0.636914074420929, + 0.031291231513023376, + 0.018323281779885292, + 0.037721533328294754, + 0.052735164761543274, + 0.03392961248755455, + 0.028072627261281013, + 0.042783841490745544, + 0.014199845492839813 + ], + "credit_map": [ + 0.010359865613281727, + 0.01567022316157818, + 0.01698135957121849, + 0.010547937825322151, + 0.015685521066188812, + 0.006535537540912628, + 0.017544245347380638, + 0.01070410292595625, + 0.18907447159290314, + 0.08688747137784958, + 0.01545601524412632, + 0.010816656053066254, + 0.27505016326904297, + 0.027947505936026573, + 0.015247898176312447, + 0.016433868557214737, + 0.00533378217369318, + 0.003248792141675949, + 0.0025101429782807827, + 0.002605871530249715, + 0.007143497467041016, + 0.0016187960281968117, + 0.00559403607621789, + 0.0032363124191761017, + 0.0021691657602787018, + 0.002615606412291527, + 0.0016162646934390068, + 0.002442115917801857, + 0.003608505707234144, + 0.001265423372387886, + 0.0027920168358832598, + 0.001814182149246335, + 0.0036942949518561363, + 0.0054288627579808235, + 0.003133865538984537, + 0.003898831084370613, + 0.004476079251617193, + 0.003332868916913867, + 0.005267478991299868, + 0.008489254862070084, + 0.005127825774252415, + 0.01270371489226818, + 0.005258241668343544, + 0.004670475143939257, + 0.012368608266115189, + 0.003292276058346033, + 0.003822740400210023, + 0.00549128046259284, + 0.004373437259346247, + 0.005356463138014078, + 0.004049200098961592, + 0.003509021829813719, + 0.0051909382455050945, + 0.002748350379988551, + 0.004822371527552605, + 0.0038798293098807335, + 0.0033730007708072662, + 0.0035883423406630754, + 0.0037129754200577736, + 0.00328823272138834, + 0.00454432750120759, + 0.0019760809373110533, + 0.003616918111220002, + 0.003972747828811407, + 0.005060449242591858, + 0.004298749845474958, + 0.0049421885050833225, + 0.0031319831032305956, + 0.006737794727087021, + 0.007698386907577515, + 0.004352210089564323, + 0.006562083028256893, + 0.0016953524900600314, + 0.001783194369636476, + 0.0015020708087831736, + 0.0012521727476269007, + 0.0027855054941028357, + 0.0015657657058909535, + 0.001728117698803544, + 0.0018876660615205765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.1429498791694641 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794275276362896 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.009794275276362896, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794275276362896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1613525003194809, + 0.37448906898498535, + 0.054661426693201065, + 0.05745429918169975, + 0.07258504629135132, + 0.1053517535328865, + 0.06012849882245064, + 0.04698263481259346, + 0.03347611427307129, + 0.03351860120892525 + ], + "credit_map": [ + 0.017721721902489662, + 0.018651099875569344, + 0.024420732632279396, + 0.01906333677470684, + 0.019566679373383522, + 0.011331988498568535, + 0.031029613688588142, + 0.019567329436540604, + 0.10159661620855331, + 0.05331656336784363, + 0.011990892700850964, + 0.010971968993544579, + 0.1544668972492218, + 0.015713317319750786, + 0.013673773966729641, + 0.012759028933942318, + 0.007872501388192177, + 0.005402799230068922, + 0.005187132395803928, + 0.004615296144038439, + 0.011817301623523235, + 0.002841991139575839, + 0.011228850111365318, + 0.0056955572217702866, + 0.010975787416100502, + 0.005493229255080223, + 0.0038880109786987305, + 0.008184220641851425, + 0.014062400907278061, + 0.0038440292701125145, + 0.006261008325964212, + 0.004745613317936659, + 0.007459667045623064, + 0.009872918017208576, + 0.005263346713036299, + 0.00822985265403986, + 0.010851689614355564, + 0.006569053046405315, + 0.009903091005980968, + 0.01443543191999197, + 0.009283467195928097, + 0.028939686715602875, + 0.011328152380883694, + 0.009699850343167782, + 0.01955515891313553, + 0.006665835622698069, + 0.007884804159402847, + 0.0119948023930192, + 0.009363709948956966, + 0.009954520501196384, + 0.006595509126782417, + 0.005507985129952431, + 0.009693860076367855, + 0.00511183962225914, + 0.006676848046481609, + 0.0072242263704538345, + 0.006213692016899586, + 0.005574433133006096, + 0.005709182005375624, + 0.005458002910017967, + 0.007064610719680786, + 0.0037861389573663473, + 0.006287014577537775, + 0.006889561656862497, + 0.004269937518984079, + 0.0023784053046256304, + 0.006031636614352465, + 0.003028841456398368, + 0.0036124635953456163, + 0.0030619273893535137, + 0.004466583952307701, + 0.006626319605857134, + 0.003676007268950343, + 0.003389186691492796, + 0.0040972852148115635, + 0.004543420393019915, + 0.005116516724228859, + 0.00430141668766737, + 0.004213464912027121, + 0.0041813007555902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.009794275276362896 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602744072675705 + ], + "opposite_logprob_tokens": [ + -3.8956027030944824 + ], + "logprob_reward": -0.020602744072675705, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602744072675705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1853574812412262, + 0.37034764885902405, + 0.027285989373922348, + 0.027141433209180832, + 0.10578326135873795, + 0.06884575635194778, + 0.0700218454003334, + 0.03749799728393555, + 0.07474777102470398, + 0.032970838248729706 + ], + "credit_map": [ + 0.020407486706972122, + 0.023880664259195328, + 0.028904374688863754, + 0.021001648157835007, + 0.025993842631578445, + 0.012269974686205387, + 0.032119616866111755, + 0.020779848098754883, + 0.10288996249437332, + 0.05440915375947952, + 0.010057087987661362, + 0.00804021954536438, + 0.15721002221107483, + 0.015812311321496964, + 0.010358509607613087, + 0.011570352129638195, + 0.0035265302285552025, + 0.002846319228410721, + 0.0038093540351837873, + 0.001937961671501398, + 0.004105494357645512, + 0.001454695826396346, + 0.005653782282024622, + 0.003951848018914461, + 0.0035068546421825886, + 0.0048367539420723915, + 0.0021931410301476717, + 0.002580219879746437, + 0.005068214610219002, + 0.0017321991035714746, + 0.00465856958180666, + 0.0025654782075434923, + 0.009739384986460209, + 0.02177906595170498, + 0.00809592753648758, + 0.011505267582833767, + 0.020522141829133034, + 0.0064725615084171295, + 0.010731902904808521, + 0.016936995089054108, + 0.008258005604147911, + 0.016965802758932114, + 0.007331268861889839, + 0.005631786305457354, + 0.010182540863752365, + 0.005124006886035204, + 0.0060321916826069355, + 0.009320140816271305, + 0.009525572881102562, + 0.012290402315557003, + 0.0074302698485553265, + 0.006600141525268555, + 0.01071250718086958, + 0.006561560090631247, + 0.008785882033407688, + 0.008115501143038273, + 0.005334989633411169, + 0.004459833726286888, + 0.0028309824410825968, + 0.004355607554316521, + 0.0055276937782764435, + 0.004394634626805782, + 0.005062312353402376, + 0.005531937815248966, + 0.00949109811335802, + 0.009211630560457706, + 0.0071953763253986835, + 0.005324516445398331, + 0.012690965086221695, + 0.013004177249968052, + 0.0076240370981395245, + 0.010205968283116817, + 0.003807207802310586, + 0.0034401058219373226, + 0.004919980186969042, + 0.004030290991067886, + 0.005088094621896744, + 0.0037613436579704285, + 0.0039014113135635853, + 0.004022400360554457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.020602744072675705 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dormant branch with deep, weathered brown bark and a rough, cracked texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly low angle looking up at the branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a soft, blurred backdrop of greens and earthy tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a new shoot with a vibrant, translucent lime-green tip shown pushing forward and piercing a papery bud shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, golden backlight that creates an ethereal glow on the unfurling leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866772145032883 + ], + "opposite_logprob_tokens": [ + -3.0486676692962646 + ], + "logprob_reward": -0.04866772145032883, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866772145032883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3821756839752197, + 0.26484885811805725, + 0.036320898681879044, + 0.06755180656909943, + 0.05848899856209755, + 0.04174010828137398, + 0.05845605209469795, + 0.04150776192545891, + 0.02054610848426819, + 0.0283636637032032 + ], + "credit_map": [ + 0.02884593792259693, + 0.10453419387340546, + 0.04037102684378624, + 0.028456412255764008, + 0.09313707798719406, + 0.01836073026061058, + 0.04170658066868782, + 0.02676372416317463, + 0.07439607381820679, + 0.03657068312168121, + 0.008075602352619171, + 0.0065488931722939014, + 0.1104268878698349, + 0.011631962843239307, + 0.007641750853508711, + 0.009557023644447327, + 0.00540856784209609, + 0.003964447416365147, + 0.004316452890634537, + 0.0029942025430500507, + 0.006615102291107178, + 0.002400739584118128, + 0.0038316636346280575, + 0.006789726670831442, + 0.012510540895164013, + 0.01112818531692028, + 0.00405752332881093, + 0.009111188352108002, + 0.011786914430558681, + 0.004909961950033903, + 0.00871976837515831, + 0.005327720195055008, + 0.006589352618902922, + 0.008104108273983002, + 0.005160951986908913, + 0.005130254663527012, + 0.00948480237275362, + 0.00450327480211854, + 0.006229693070054054, + 0.013286566361784935, + 0.0043868739157915115, + 0.01043617445975542, + 0.0035590732004493475, + 0.0043617733754217625, + 0.006636866368353367, + 0.0026243908796459436, + 0.00470354687422514, + 0.0050314124673604965, + 0.007646569982171059, + 0.012157868593931198, + 0.0057172514498233795, + 0.00625390000641346, + 0.009833895601332188, + 0.004813612904399633, + 0.00591493071988225, + 0.0061180260963737965, + 0.004225501324981451, + 0.005423514172434807, + 0.004273431375622749, + 0.004237865563482046, + 0.006742698140442371, + 0.005329264793545008, + 0.006108684465289116, + 0.0051668016240000725, + 0.002018043538555503, + 0.0025203076656907797, + 0.0028879090677946806, + 0.0019147791899740696, + 0.00301114353351295, + 0.0021306073758751154, + 0.003423119895160198, + 0.0026401986833661795, + 0.002939425176009536, + 0.0032823928631842136, + 0.002951772650703788, + 0.003382172202691436, + 0.005090121179819107, + 0.0032304120250046253, + 0.003689353819936514, + 0.0037980147171765566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04866772145032883 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_tensor([6888, 6888])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_tensor([6888, 6888])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c90ff42272207f95c432a98988c8986b939ee1f4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_tensor([6888, 6888])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot horizontally and rendered in slow motion, focuses on a heavy, dusty, dark wooden table. Lying quietly upon the tabletop lies an ancient parchment scroll, its paper varying shades of yellow, its edges frayed and slightly charred, its surface covered in fine wrinkles and the marks of age. Suddenly, a powerful, invisible gust of wind blows violently, instantly lifting a corner of the scroll and sending the paper trembling and rising. This movement stirs the long-settled dust on the tabletop, carrying countless golden particles into the air. Within a dramatic beam of light cast from above, a powerful Tyndall effect, the resulting image instantly forms a swirling, shimmering golden nebula. Each speck of dust shimmers like tiny pieces of gold foil, dancing around the slowly unrolling scroll. The entire scene is shrouded in a mysterious, solemn atmosphere imbued with ancient magic, painted in warm, deep tones.", + "seed": 29256, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_6888_seed_29256_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011212486308068037, + "vlm_logprob": -0.011212486308068037, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.011212486308068037, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03187984973192215 + ], + "opposite_logprob_tokens": [ + -3.7818799018859863 + ], + "logprob_reward": -0.03187984973192215, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03187984973192215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12774808704853058, + 0.06235355883836746, + 0.2654176652431488, + 0.07827116549015045, + 0.08405452966690063, + 0.07452062517404556, + 0.0834154337644577, + 0.04313347861170769, + 0.10605309158563614, + 0.07503242045640945 + ], + "credit_map": [ + 0.022208403795957565, + 0.01471717283129692, + 0.01513743121176958, + 0.010583915747702122, + 0.01811905950307846, + 0.016575634479522705, + 0.016612667590379715, + 0.013793785125017166, + 0.008190741762518883, + 0.007957642897963524, + 0.00759378494694829, + 0.007350832223892212, + 0.007998470216989517, + 0.00819465797394514, + 0.008170709013938904, + 0.006896722596138716, + 0.02236006036400795, + 0.024579128250479698, + 0.019503042101860046, + 0.019807128235697746, + 0.07261966913938522, + 0.03958289697766304, + 0.04740307852625847, + 0.019562629982829094, + 0.00676777632907033, + 0.007762915454804897, + 0.007239466067403555, + 0.0074985879473388195, + 0.0106298066675663, + 0.009202048182487488, + 0.018252750858664513, + 0.010917813517153263, + 0.012528588995337486, + 0.01005844958126545, + 0.009280712343752384, + 0.00692767184227705, + 0.008992008864879608, + 0.011436729691922665, + 0.010186143219470978, + 0.014644230715930462, + 0.008490979671478271, + 0.008880610577762127, + 0.009504668414592743, + 0.011231058277189732, + 0.010343867354094982, + 0.006197250448167324, + 0.008957898244261742, + 0.010914290323853493, + 0.008917600847780704, + 0.011050847359001637, + 0.008265919052064419, + 0.014584191143512726, + 0.015410013496875763, + 0.007330097258090973, + 0.007781537249684334, + 0.010075234808027744, + 0.004798131063580513, + 0.005490942858159542, + 0.004389910958707333, + 0.006419474724680185, + 0.00672953249886632, + 0.006091899238526821, + 0.0031004708725959063, + 0.006113115232437849, + 0.017749253660440445, + 0.016781432554125786, + 0.01103239320218563, + 0.014071028679609299, + 0.010980312712490559, + 0.009928267449140549, + 0.008649550378322601, + 0.016860848292708397, + 0.008659678511321545, + 0.010960525833070278, + 0.014123977161943913, + 0.0064909071661531925, + 0.012477347627282143, + 0.008195541799068451, + 0.00578219722956419, + 0.00834224559366703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.03187984973192215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025022882036864758 + ], + "opposite_logprob_tokens": [ + -6.00250244140625 + ], + "logprob_reward": -0.0025022882036864758, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025022882036864758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055680543184280396, + 0.03393562138080597, + 0.3420664072036743, + 0.05203599855303764, + 0.056327883154153824, + 0.07823377102613449, + 0.15895208716392517, + 0.068479023873806, + 0.028090599924325943, + 0.12619800865650177 + ], + "credit_map": [ + 0.00679917773231864, + 0.005682238377630711, + 0.0041792867705225945, + 0.00676968926563859, + 0.011872472241520882, + 0.007502291351556778, + 0.00648365868255496, + 0.006391730159521103, + 0.0032758486922830343, + 0.004307748284190893, + 0.0022547694388777018, + 0.003590276697650552, + 0.008217427879571915, + 0.005694609135389328, + 0.004069349728524685, + 0.002525590592995286, + 0.023631500080227852, + 0.03156549483537674, + 0.02603215165436268, + 0.023374203592538834, + 0.08843334019184113, + 0.0546284094452858, + 0.06598637998104095, + 0.028414921835064888, + 0.004081448074430227, + 0.00472300685942173, + 0.005034454632550478, + 0.0037461596075445414, + 0.009727521799504757, + 0.0066249780356884, + 0.011546488851308823, + 0.006551941391080618, + 0.004828793462365866, + 0.007780604064464569, + 0.004468974657356739, + 0.004144658800214529, + 0.0079030841588974, + 0.00657882122322917, + 0.007496098522096872, + 0.013126851059496403, + 0.009284431114792824, + 0.007963193580508232, + 0.010710026137530804, + 0.009179896675050259, + 0.012845915742218494, + 0.007063350640237331, + 0.01136963814496994, + 0.009817319922149181, + 0.02261877804994583, + 0.01716422475874424, + 0.020605267956852913, + 0.021109430119395256, + 0.023635128512978554, + 0.012465916574001312, + 0.014529416337609291, + 0.02682393789291382, + 0.005972858984023333, + 0.009286576882004738, + 0.007605815771967173, + 0.009589577093720436, + 0.01578718237578869, + 0.006996117066591978, + 0.005631949752569199, + 0.007608945947140455, + 0.004382268059998751, + 0.003985115792602301, + 0.0030406888108700514, + 0.0035642872098833323, + 0.0033930139616131783, + 0.003034592140465975, + 0.002785872435197234, + 0.0039047629106789827, + 0.013937026262283325, + 0.017779512330889702, + 0.01879269629716873, + 0.016988473013043404, + 0.02005373127758503, + 0.012333361431956291, + 0.01396878995001316, + 0.012344417162239552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.0025022882036864758 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079810559749603 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079810559749603, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079810559749603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0895773246884346, + 0.04665754735469818, + 0.17317639291286469, + 0.03948879614472389, + 0.07165946811437607, + 0.11094480007886887, + 0.11690283566713333, + 0.03881023824214935, + 0.1276337057352066, + 0.1851489096879959 + ], + "credit_map": [ + 0.008926342241466045, + 0.009412615559995174, + 0.00828011054545641, + 0.013101167045533657, + 0.016797997057437897, + 0.012493505142629147, + 0.011592107824981213, + 0.008973478339612484, + 0.004130980931222439, + 0.005878678523004055, + 0.0038458779454231262, + 0.004304897505789995, + 0.0096576027572155, + 0.006672670133411884, + 0.007609308697283268, + 0.004557529930025339, + 0.01557189878076315, + 0.015575318597257137, + 0.016679853200912476, + 0.01610109768807888, + 0.037553466856479645, + 0.02561027556657791, + 0.031683072447776794, + 0.014401406981050968, + 0.0030350929591804743, + 0.004086165223270655, + 0.0038540728855878115, + 0.005137044936418533, + 0.004620624706149101, + 0.005035240203142166, + 0.007310370448976755, + 0.006410185247659683, + 0.006598870735615492, + 0.008545425720512867, + 0.005861874669790268, + 0.005136009305715561, + 0.010900309309363365, + 0.008695908822119236, + 0.007474692538380623, + 0.01844637095928192, + 0.010543362237513065, + 0.012616574764251709, + 0.01200197171419859, + 0.011984807439148426, + 0.02111516147851944, + 0.011181794106960297, + 0.012421808205544949, + 0.019079318270087242, + 0.016438310965895653, + 0.014624358154833317, + 0.016641853377223015, + 0.018440859392285347, + 0.01823394186794758, + 0.008343590423464775, + 0.010973786003887653, + 0.013206138275563717, + 0.005158912856131792, + 0.006058533675968647, + 0.0035563239362090826, + 0.004817361477762461, + 0.008749843575060368, + 0.0038639039266854525, + 0.0028214354533702135, + 0.0037839175201952457, + 0.021940868347883224, + 0.017946120351552963, + 0.014820797368884087, + 0.010720623657107353, + 0.017069976776838303, + 0.014517276547849178, + 0.010961022228002548, + 0.019657008349895477, + 0.022019583731889725, + 0.024233510717749596, + 0.029026992619037628, + 0.024173852056264877, + 0.02777617983520031, + 0.018070852383971214, + 0.021011022850871086, + 0.018836896866559982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.011079810559749603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046443212777376175 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.0046443212777376175, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046443212777376175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07271966338157654, + 0.04494632035493851, + 0.11206153780221939, + 0.06510839611291885, + 0.03646828979253769, + 0.030639348551630974, + 0.06668725609779358, + 0.04743797704577446, + 0.3630179464817047, + 0.16091318428516388 + ], + "credit_map": [ + 0.008036256767809391, + 0.008064109832048416, + 0.011024508625268936, + 0.006053833290934563, + 0.011492273770272732, + 0.010649220086634159, + 0.008543911390006542, + 0.008855560794472694, + 0.0048467619344592094, + 0.005311971064656973, + 0.004182123579084873, + 0.0033723399974405766, + 0.01008179783821106, + 0.007137756794691086, + 0.006096580531448126, + 0.003916991874575615, + 0.009901142679154873, + 0.011160125024616718, + 0.00859715137630701, + 0.00563382962718606, + 0.034255195409059525, + 0.016149045899510384, + 0.01764506660401821, + 0.00871998630464077, + 0.008225517347455025, + 0.006821460090577602, + 0.008125661872327328, + 0.005486751906573772, + 0.008989771828055382, + 0.007068157196044922, + 0.01159659307450056, + 0.008794487453997135, + 0.0034085542429238558, + 0.005218086764216423, + 0.0025340132415294647, + 0.0019702650606632233, + 0.0057370224967598915, + 0.005466692615300417, + 0.004143295343965292, + 0.007990362122654915, + 0.002774507272988558, + 0.0036855442449450493, + 0.0022917240858078003, + 0.0024555809795856476, + 0.007495697122067213, + 0.003507856046780944, + 0.003416928695514798, + 0.005011511035263538, + 0.007243945728987455, + 0.007570105604827404, + 0.011007732711732388, + 0.012518694624304771, + 0.008594730868935585, + 0.004878871608525515, + 0.007262133993208408, + 0.007611047942191362, + 0.004577208776026964, + 0.007219766266644001, + 0.006463393568992615, + 0.006798218935728073, + 0.007939288392663002, + 0.004411360248923302, + 0.003944126423448324, + 0.006084614899009466, + 0.05568571388721466, + 0.054304733872413635, + 0.0372358113527298, + 0.05066891387104988, + 0.044508546590805054, + 0.03763410821557045, + 0.03543355315923691, + 0.047546595335006714, + 0.014071925543248653, + 0.026298614218831062, + 0.02067182958126068, + 0.021005693823099136, + 0.029006831347942352, + 0.017510702833533287, + 0.0173255056142807, + 0.015022099949419498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.0046443212777376175 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956161767244339 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956161767244339, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956161767244339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07935456931591034, + 0.02382626198232174, + 0.20192253589630127, + 0.08684074878692627, + 0.06353144347667694, + 0.13891804218292236, + 0.21346722543239594, + 0.055571820586919785, + 0.03749172016978264, + 0.09907564520835876 + ], + "credit_map": [ + 0.010622669011354446, + 0.00860509555786848, + 0.007385721430182457, + 0.00832802802324295, + 0.01219211146235466, + 0.012778379023075104, + 0.010004145093262196, + 0.009438415057957172, + 0.002086859894916415, + 0.0031196579802781343, + 0.002301509026437998, + 0.0024707159027457237, + 0.004311596509069204, + 0.0036891065537929535, + 0.003259094664826989, + 0.0025877186562865973, + 0.016918495297431946, + 0.01840529963374138, + 0.01453215442597866, + 0.01402617059648037, + 0.05236390605568886, + 0.03079391084611416, + 0.03818368911743164, + 0.01669890061020851, + 0.007353487890213728, + 0.00805631186813116, + 0.008295896463096142, + 0.006373220589011908, + 0.015087113715708256, + 0.010464414954185486, + 0.01970970816910267, + 0.01150058675557375, + 0.005648193880915642, + 0.0095808245241642, + 0.006013878155499697, + 0.0048961634747684, + 0.010829676873981953, + 0.007258749101310968, + 0.008944114670157433, + 0.010359836742281914, + 0.015575536526739597, + 0.016168951988220215, + 0.014335854910314083, + 0.015225518494844437, + 0.028854990378022194, + 0.01176054123789072, + 0.0209345705807209, + 0.01606207899749279, + 0.032176293432712555, + 0.024181341752409935, + 0.02975962869822979, + 0.02678975835442543, + 0.034088220447301865, + 0.016856079921126366, + 0.019156858325004578, + 0.030459025874733925, + 0.005727139301598072, + 0.007823648862540722, + 0.006012939382344484, + 0.006335021927952766, + 0.012478402815759182, + 0.006353168282657862, + 0.004358633421361446, + 0.006482860539108515, + 0.004685777239501476, + 0.00457947188988328, + 0.0055336193181574345, + 0.004008222837001085, + 0.005828380584716797, + 0.004383888561278582, + 0.004254447761923075, + 0.004217908717691898, + 0.008519761264324188, + 0.014381666667759418, + 0.012207106687128544, + 0.013058333657681942, + 0.01619051583111286, + 0.010027105920016766, + 0.011413375847041607, + 0.013277776539325714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.005956161767244339 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot horizontally and rendered in slow motion, focuses on a heavy, dusty, dark wooden table. Lying quietly upon the tabletop lies an ancient parchment scroll, its paper varying shades of yellow, its edges frayed and slightly charred, its surface covered in fine wrinkles and the marks of age. Suddenly, a powerful, invisible gust of wind blows violently, instantly lifting a corner of the scroll and sending the paper trembling and rising. This movement stirs the long-settled dust on the tabletop, carrying countless golden particles into the air. Within a dramatic beam of light cast from above, a powerful Tyndall effect, the resulting image instantly forms a swirling, shimmering golden nebula. Each speck of dust shimmers like tiny pieces of gold foil, dancing around the slowly unrolling scroll. The entire scene is shrouded in a mysterious, solemn atmosphere imbued with ancient magic, painted in warm, deep tones.", + "seed": 29257, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_6888_seed_29257_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008929101540707052, + "vlm_logprob": -0.008929101540707052, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.008929101540707052, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017786525189876556 + ], + "opposite_logprob_tokens": [ + -4.392786502838135 + ], + "logprob_reward": -0.017786525189876556, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017786525189876556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06634577363729477, + 0.13546307384967804, + 0.025335531681776047, + 0.1466677188873291, + 0.4512377381324768, + 0.019706029444932938, + 0.045676637440919876, + 0.043499670922756195, + 0.06015823408961296, + 0.005909576080739498 + ], + "credit_map": [ + 0.008082526735961437, + 0.008168491534888744, + 0.010411389172077179, + 0.009581610560417175, + 0.005396677181124687, + 0.008957630023360252, + 0.008220215328037739, + 0.007527234498411417, + 0.015363890677690506, + 0.01837383583188057, + 0.023986617103219032, + 0.019393954426050186, + 0.019068099558353424, + 0.016763560473918915, + 0.007133512757718563, + 0.015379599295556545, + 0.0021749166771769524, + 0.0028804440516978502, + 0.002075356664136052, + 0.002379294950515032, + 0.004737076349556446, + 0.006669388618320227, + 0.0019866034854203463, + 0.002432453678920865, + 0.010265257209539413, + 0.020821131765842438, + 0.011602749116718769, + 0.017472470179200172, + 0.03845483064651489, + 0.028612826019525528, + 0.007920386269688606, + 0.01151806116104126, + 0.030989129096269608, + 0.062016624957323074, + 0.037250496447086334, + 0.06285233795642853, + 0.09980109333992004, + 0.09188598394393921, + 0.03116423636674881, + 0.0352778397500515, + 0.0016481638886034489, + 0.0029720570892095566, + 0.0022966954857110977, + 0.001829344779253006, + 0.0040915049612522125, + 0.003670027945190668, + 0.0012271713931113482, + 0.001971062971279025, + 0.0033060889691114426, + 0.005240108817815781, + 0.00323567190207541, + 0.002304180758073926, + 0.013979046605527401, + 0.012659951113164425, + 0.001980310073122382, + 0.0029712773393839598, + 0.002365515800192952, + 0.006041507702320814, + 0.002168812323361635, + 0.0013748168712481856, + 0.015548096038401127, + 0.012908349744975567, + 0.0012459417339414358, + 0.00184662826359272, + 0.0034806965850293636, + 0.021560221910476685, + 0.0034459526650607586, + 0.0025961368810385466, + 0.008565089665353298, + 0.015014677308499813, + 0.0019973060116171837, + 0.003498150035738945, + 0.0006453586393035948, + 0.001346759730949998, + 0.000709700514562428, + 0.0004925073590129614, + 0.0009380625560879707, + 0.0008204104378819466, + 0.00036838353844359517, + 0.0005883931298740208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.017786525189876556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015261442167684436 + ], + "opposite_logprob_tokens": [ + -6.501526355743408 + ], + "logprob_reward": -0.0015261442167684436, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015261442167684436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29049620032310486, + 0.08159966766834259, + 0.05196929723024368, + 0.0814349427819252, + 0.34462714195251465, + 0.026762763038277626, + 0.08057724684476852, + 0.023093195632100105, + 0.016232622787356377, + 0.0032069780863821507 + ], + "credit_map": [ + 0.03568486124277115, + 0.033698584884405136, + 0.05603081360459328, + 0.034150902181863785, + 0.029804794117808342, + 0.022849775850772858, + 0.04004041850566864, + 0.03823604807257652, + 0.008655628189444542, + 0.01209677942097187, + 0.011973238550126553, + 0.012223737314343452, + 0.012187936343252659, + 0.01032780110836029, + 0.00494182575494051, + 0.009192719124257565, + 0.0037987756077200174, + 0.009585898369550705, + 0.0037555780727416277, + 0.004289979115128517, + 0.0130697563290596, + 0.009407606907188892, + 0.0029469614382833242, + 0.005114743020385504, + 0.006436878815293312, + 0.010088939219713211, + 0.008028128184378147, + 0.012414059601724148, + 0.01663045771420002, + 0.013736405409872532, + 0.005822659004479647, + 0.008277407847344875, + 0.023491019383072853, + 0.046844661235809326, + 0.027723686769604683, + 0.04749428108334541, + 0.07731017470359802, + 0.07172034680843353, + 0.023484567180275917, + 0.026558389887213707, + 0.0023333763238042593, + 0.004666035994887352, + 0.0032407010439783335, + 0.0020707189105451107, + 0.00485618319362402, + 0.004499693866819143, + 0.0016540358774363995, + 0.00344201922416687, + 0.005846214946359396, + 0.008083686232566833, + 0.0057502370327711105, + 0.0035869802813977003, + 0.025397608056664467, + 0.02331700548529625, + 0.0033672372810542583, + 0.005228282883763313, + 0.0023507375735789537, + 0.005160408094525337, + 0.003544303821399808, + 0.002148048486560583, + 0.0031751052010804415, + 0.0024206500966101885, + 0.0020160949788987637, + 0.002277847845107317, + 0.0012370871845632792, + 0.005067694466561079, + 0.0014316760934889317, + 0.00121136917732656, + 0.00328624970279634, + 0.0018000926356762648, + 0.0009706377750262618, + 0.001227815868332982, + 0.0003155542362947017, + 0.000800398294813931, + 0.0003091380640398711, + 0.00026266867644153535, + 0.00046786791062913835, + 0.0005279472679831088, + 0.0001834832364693284, + 0.00033992042881436646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0015261442167684436 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419694535434246 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.01419694535434246, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419694535434246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18234814703464508, + 0.20779763162136078, + 0.09832867234945297, + 0.09373737871646881, + 0.2901184558868408, + 0.01918397657573223, + 0.026817183941602707, + 0.016809431836009026, + 0.060472242534160614, + 0.0043868692591786385 + ], + "credit_map": [ + 0.022332677617669106, + 0.018145468086004257, + 0.03460203856229782, + 0.023840228095650673, + 0.018820438534021378, + 0.017175354063510895, + 0.025666935369372368, + 0.02176501415669918, + 0.021604496985673904, + 0.03507716581225395, + 0.02996162138879299, + 0.026997016742825508, + 0.03260741010308266, + 0.026828860864043236, + 0.011794915422797203, + 0.02292616479098797, + 0.007228749338537455, + 0.022918248549103737, + 0.006077982950955629, + 0.006359069608151913, + 0.022599441930651665, + 0.017538905143737793, + 0.005247017368674278, + 0.010359256528317928, + 0.006333048455417156, + 0.011782703921198845, + 0.005921494215726852, + 0.01062590628862381, + 0.0251933541148901, + 0.02137334644794464, + 0.005689066369086504, + 0.006818462163209915, + 0.019418252632021904, + 0.040103841572999954, + 0.021911030635237694, + 0.0394129678606987, + 0.06861543655395508, + 0.06133250892162323, + 0.019672896713018417, + 0.019651520997285843, + 0.001512731658294797, + 0.0027327737770974636, + 0.0017926375148817897, + 0.002053031465038657, + 0.003929265774786472, + 0.003917222376912832, + 0.0012668390991166234, + 0.001979476073756814, + 0.0018336031353101134, + 0.0033291764557361603, + 0.0020034469198435545, + 0.001106154639273882, + 0.00846517737954855, + 0.00738088134676218, + 0.0008843590621836483, + 0.0018143843626603484, + 0.0011003012768924236, + 0.0032854704186320305, + 0.0009549089591018856, + 0.0008720551850274205, + 0.005506971850991249, + 0.0035370017867535353, + 0.0005772252916358411, + 0.0009754971833899617, + 0.0043280054815113544, + 0.016253212466835976, + 0.004411839880049229, + 0.0033641329500824213, + 0.006707190070301294, + 0.01862420327961445, + 0.0026321951299905777, + 0.004151463974267244, + 0.0004449646221473813, + 0.0006684288382530212, + 0.00047587245353497565, + 0.0004904078086838126, + 0.0005212351097725332, + 0.0006367071182467043, + 0.0003889028448611498, + 0.0007603504927828908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.01419694535434246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024920618161559105 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.0024920618161559105, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024920618161559105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036135874688625336, + 0.03368326276540756, + 0.049563661217689514, + 0.2654638886451721, + 0.3759649991989136, + 0.04285823553800583, + 0.07749301195144653, + 0.06992176175117493, + 0.03745020925998688, + 0.011465033516287804 + ], + "credit_map": [ + 0.003913864493370056, + 0.0058760372921824455, + 0.006186498794704676, + 0.006262281443923712, + 0.004069142043590546, + 0.003592956578359008, + 0.0036055119708180428, + 0.002629583002999425, + 0.0038211361970752478, + 0.00434836745262146, + 0.007293127477169037, + 0.00521122058853507, + 0.003229981055483222, + 0.0032968767918646336, + 0.002404714236035943, + 0.004077838733792305, + 0.00426965206861496, + 0.010420504957437515, + 0.004477374721318483, + 0.003404193324968219, + 0.010110881179571152, + 0.010063068941235542, + 0.0026994827203452587, + 0.004118505399674177, + 0.018452191725373268, + 0.03520362079143524, + 0.017502909526228905, + 0.036266524344682693, + 0.0647660344839096, + 0.05414767563343048, + 0.018708450719714165, + 0.0204165056347847, + 0.026728413999080658, + 0.0517825148999691, + 0.0337519608438015, + 0.055340517312288284, + 0.07486232370138168, + 0.07372113317251205, + 0.027194470167160034, + 0.03258366137742996, + 0.003180129686370492, + 0.009894096292555332, + 0.0045368787832558155, + 0.0026603457517921925, + 0.008463816717267036, + 0.007964124903082848, + 0.0021726915147155523, + 0.003986150957643986, + 0.005292970687150955, + 0.0082161333411932, + 0.0047204988077282906, + 0.002899909857660532, + 0.026065336540341377, + 0.02358442172408104, + 0.002711972687393427, + 0.004001769702881575, + 0.004153432324528694, + 0.011544991284608841, + 0.004629618488252163, + 0.002466660924255848, + 0.024580728262662888, + 0.01714019663631916, + 0.002360591199249029, + 0.0030455482192337513, + 0.0027009560726583004, + 0.010785598307847977, + 0.0027532593812793493, + 0.0017435320187360048, + 0.0030996554996818304, + 0.01256483606994152, + 0.001588137587532401, + 0.0022142394445836544, + 0.0009475243859924376, + 0.0025030020624399185, + 0.0009924530750140548, + 0.0008602188900113106, + 0.0021369480527937412, + 0.0025285787414759398, + 0.0006596963503398001, + 0.000836611317936331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.0024920618161559105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643831126391888 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643831126391888, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643831126391888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07992250472307205, + 0.027397053316235542, + 0.049509234726428986, + 0.15598434209823608, + 0.36974242329597473, + 0.03668138384819031, + 0.08773264288902283, + 0.0894327387213707, + 0.09978263080120087, + 0.0038150667678564787 + ], + "credit_map": [ + 0.009585320018231869, + 0.009550641290843487, + 0.016195092350244522, + 0.00971287116408348, + 0.008663642220199108, + 0.006460902746766806, + 0.009676970541477203, + 0.010077064856886864, + 0.0025175337214022875, + 0.004889587871730328, + 0.0033098384737968445, + 0.004805528558790684, + 0.003057158086448908, + 0.002981213154271245, + 0.00326356734149158, + 0.0025726251769810915, + 0.004307261668145657, + 0.007240738719701767, + 0.003805313026532531, + 0.004076727200299501, + 0.01069408468902111, + 0.010784666985273361, + 0.0034149514976888895, + 0.005185495130717754, + 0.011252118274569511, + 0.020739471539855003, + 0.01237245462834835, + 0.020981425419449806, + 0.03802124410867691, + 0.029974857345223427, + 0.010396409779787064, + 0.012246355414390564, + 0.025333235040307045, + 0.051187220960855484, + 0.030572915449738503, + 0.052583809942007065, + 0.08159925788640976, + 0.07469335198402405, + 0.026034587994217873, + 0.02773800492286682, + 0.0031115254387259483, + 0.008145367726683617, + 0.003834172384813428, + 0.003058262402191758, + 0.005936897825449705, + 0.005918601993471384, + 0.002190845785662532, + 0.004485709592700005, + 0.005921864416450262, + 0.009677458554506302, + 0.005280695389956236, + 0.0039200191386044025, + 0.028861263766884804, + 0.025633739307522774, + 0.0034988559782505035, + 0.004938747733831406, + 0.004880234599113464, + 0.014092739671468735, + 0.00510714715346694, + 0.002805625321343541, + 0.0324489027261734, + 0.02391258254647255, + 0.0026035765185952187, + 0.003581929486244917, + 0.006082636304199696, + 0.01908048242330551, + 0.0036789679434150457, + 0.0026389630511403084, + 0.019935522228479385, + 0.04178626835346222, + 0.0023956785444170237, + 0.004184114281088114, + 0.0003835463139694184, + 0.0007150442688725889, + 0.00034314816002734005, + 0.0003711394383572042, + 0.0005333541776053607, + 0.0008849417790770531, + 0.00024759609368629754, + 0.0003362962161190808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.008643831126391888 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot horizontally and rendered in slow motion, focuses on a heavy, dusty, dark wooden table. Lying quietly upon the tabletop lies an ancient parchment scroll, its paper varying shades of yellow, its edges frayed and slightly charred, its surface covered in fine wrinkles and the marks of age. Suddenly, a powerful, invisible gust of wind blows violently, instantly lifting a corner of the scroll and sending the paper trembling and rising. This movement stirs the long-settled dust on the tabletop, carrying countless golden particles into the air. Within a dramatic beam of light cast from above, a powerful Tyndall effect, the resulting image instantly forms a swirling, shimmering golden nebula. Each speck of dust shimmers like tiny pieces of gold foil, dancing around the slowly unrolling scroll. The entire scene is shrouded in a mysterious, solemn atmosphere imbued with ancient magic, painted in warm, deep tones.", + "seed": 29258, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_6888_seed_29258_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.005363243422470987, + "vlm_logprob": -0.005363243422470987, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.005363243422470987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011833352968096733 + ], + "opposite_logprob_tokens": [ + -4.761833190917969 + ], + "logprob_reward": -0.011833352968096733, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011833352968096733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13913974165916443, + 0.14647501707077026, + 0.07969444990158081, + 0.06032640114426613, + 0.23696845769882202, + 0.11811421066522598, + 0.06102300435304642, + 0.03665102273225784, + 0.04421747848391533, + 0.0773901492357254 + ], + "credit_map": [ + 0.012680661864578724, + 0.02299201674759388, + 0.009863298386335373, + 0.01880735345184803, + 0.025399651378393173, + 0.02006569691002369, + 0.011118349619209766, + 0.018212709575891495, + 0.017090769484639168, + 0.018993394449353218, + 0.01714092306792736, + 0.018226655200123787, + 0.023809775710105896, + 0.01832539588212967, + 0.013983935117721558, + 0.018904175609350204, + 0.011579042300581932, + 0.00869979802519083, + 0.010570432990789413, + 0.009092654101550579, + 0.011476202867925167, + 0.007930871099233627, + 0.010581589303910732, + 0.009763862937688828, + 0.009851337410509586, + 0.005888415966182947, + 0.009205779060721397, + 0.006163547746837139, + 0.010596022941172123, + 0.007105693221092224, + 0.004854431841522455, + 0.006661179009824991, + 0.023429540917277336, + 0.024246254935860634, + 0.01039513386785984, + 0.014473773539066315, + 0.09061328321695328, + 0.02631443552672863, + 0.02637164294719696, + 0.02112438529729843, + 0.010888470336794853, + 0.016998320817947388, + 0.007304486818611622, + 0.012534884735941887, + 0.03586629778146744, + 0.01484507042914629, + 0.010915628634393215, + 0.008761054836213589, + 0.006981657352298498, + 0.010026670061051846, + 0.005541736260056496, + 0.004600218031555414, + 0.011236514896154404, + 0.006384858395904303, + 0.006790148094296455, + 0.009461202658712864, + 0.003715104889124632, + 0.004871162120252848, + 0.003146142465993762, + 0.0032899859361350536, + 0.010290690697729588, + 0.00390616781078279, + 0.0029208867345005274, + 0.004510882776230574, + 0.004596061073243618, + 0.00890799518674612, + 0.004202511627227068, + 0.0038388788234442472, + 0.010691487230360508, + 0.004012607503682375, + 0.003205762943252921, + 0.004762174561619759, + 0.0069669135846197605, + 0.01923280395567417, + 0.008020708337426186, + 0.007474973797798157, + 0.017492162063717842, + 0.00642717070877552, + 0.005072395317256451, + 0.006703025661408901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.011833352968096733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019477697787806392 + ], + "opposite_logprob_tokens": [ + -6.25194787979126 + ], + "logprob_reward": -0.0019477697787806392, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019477697787806392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12540298700332642, + 0.10033868998289108, + 0.09368923306465149, + 0.03928378224372864, + 0.12580527365207672, + 0.09431702643632889, + 0.19907797873020172, + 0.045591048896312714, + 0.026769788935780525, + 0.14972415566444397 + ], + "credit_map": [ + 0.01487117912620306, + 0.020101916044950485, + 0.010161858052015305, + 0.015350127592682838, + 0.019064778462052345, + 0.014907928183674812, + 0.014488393440842628, + 0.016456810757517815, + 0.011707128956913948, + 0.011529100127518177, + 0.011851108632981777, + 0.014724367298185825, + 0.017272701486945152, + 0.011017275042831898, + 0.009050954133272171, + 0.013186057098209858, + 0.010491609573364258, + 0.01045173965394497, + 0.010423894040286541, + 0.013197129592299461, + 0.014925768598914146, + 0.009216497652232647, + 0.016322113573551178, + 0.008660484105348587, + 0.006280358415096998, + 0.003879149444401264, + 0.004788287449628115, + 0.0029740568716078997, + 0.00708207581192255, + 0.005032598506659269, + 0.0031547562684863806, + 0.006092502735555172, + 0.012430815026164055, + 0.014573651365935802, + 0.006253866944462061, + 0.006966396700590849, + 0.04940351843833923, + 0.014309212565422058, + 0.013377644121646881, + 0.008490155450999737, + 0.00894079077988863, + 0.016646677628159523, + 0.007399887777864933, + 0.008272393606603146, + 0.02447102963924408, + 0.012174339033663273, + 0.008129655383527279, + 0.008282250724732876, + 0.02164377085864544, + 0.03171149268746376, + 0.021990785375237465, + 0.019259246066212654, + 0.041959766298532486, + 0.018075019121170044, + 0.021204747259616852, + 0.023233139887452126, + 0.0039283535443246365, + 0.011843343265354633, + 0.00361348083242774, + 0.003192602191120386, + 0.01104417722672224, + 0.005072182975709438, + 0.0031257665250450373, + 0.003771139308810234, + 0.002362746512517333, + 0.005840890575200319, + 0.0020845334511250257, + 0.002643006620928645, + 0.007410828024148941, + 0.003175921505317092, + 0.0015442060539498925, + 0.0017076560761779547, + 0.013253358192741871, + 0.03983534872531891, + 0.01542509812861681, + 0.01351090520620346, + 0.03287634253501892, + 0.013727591373026371, + 0.008607429452240467, + 0.01248808391392231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.0019477697787806392 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096688237041235 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096688237041235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096688237041235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1974336802959442, + 0.08351726084947586, + 0.09659463167190552, + 0.06519146263599396, + 0.2454010546207428, + 0.06147370859980583, + 0.11318635940551758, + 0.052862267941236496, + 0.03161836415529251, + 0.05272119119763374 + ], + "credit_map": [ + 0.020353158935904503, + 0.0322309210896492, + 0.012410846538841724, + 0.019928351044654846, + 0.03596949577331543, + 0.02708955481648445, + 0.01756385527551174, + 0.0318874828517437, + 0.009899811819195747, + 0.013626067899167538, + 0.010735142044723034, + 0.00851768534630537, + 0.011674475856125355, + 0.009675297886133194, + 0.009170274250209332, + 0.010218506678938866, + 0.012143813073635101, + 0.008504489436745644, + 0.01235144305974245, + 0.01479244977235794, + 0.013078560121357441, + 0.007056763395667076, + 0.01613147184252739, + 0.012535644695162773, + 0.010857593268156052, + 0.007079700473695993, + 0.008752749301493168, + 0.005616231821477413, + 0.012437005527317524, + 0.006303294096142054, + 0.006082030013203621, + 0.008062858134508133, + 0.025954298675060272, + 0.02368287183344364, + 0.015440919436514378, + 0.019487839192152023, + 0.09013429284095764, + 0.027044981718063354, + 0.023460058495402336, + 0.02019578590989113, + 0.0060304803773760796, + 0.01022946834564209, + 0.005007878411561251, + 0.009145594201982021, + 0.012986752204596996, + 0.0067809284664690495, + 0.005444648209959269, + 0.005847955122590065, + 0.009982365183532238, + 0.013578718528151512, + 0.01104958076030016, + 0.012859584763646126, + 0.024688521400094032, + 0.013189819641411304, + 0.012498198077082634, + 0.015339578501880169, + 0.004928346257656813, + 0.005859857890754938, + 0.004799996502697468, + 0.005910728592425585, + 0.013135368935763836, + 0.005982941947877407, + 0.0053149000741541386, + 0.006930127274245024, + 0.0030872488860040903, + 0.004200047347694635, + 0.0020807008258998394, + 0.002537446329370141, + 0.010334753431379795, + 0.004307042341679335, + 0.0024557264987379313, + 0.002615400357171893, + 0.004176505841314793, + 0.009392241947352886, + 0.004977723583579063, + 0.0053854938596487045, + 0.012767428532242775, + 0.00549828726798296, + 0.003919965587556362, + 0.00660354271531105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.004096688237041235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199375070631504 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.002199375070631504, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199375070631504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.167944997549057, + 0.0964096188545227, + 0.054213590919971466, + 0.04035690054297447, + 0.1767868548631668, + 0.11239874362945557, + 0.1267741322517395, + 0.05658522620797157, + 0.08606699854135513, + 0.08246289193630219 + ], + "credit_map": [ + 0.017708951607346535, + 0.020803023129701614, + 0.011068475432693958, + 0.015126630663871765, + 0.025988539680838585, + 0.026927482336759567, + 0.0142300920560956, + 0.03609180450439453, + 0.011848180554807186, + 0.013507579453289509, + 0.010949626564979553, + 0.013442115858197212, + 0.015028049238026142, + 0.010473761707544327, + 0.008321735076606274, + 0.012838577851653099, + 0.005980368237942457, + 0.006208466365933418, + 0.005238981917500496, + 0.008724181912839413, + 0.008231883868575096, + 0.005348936188966036, + 0.008332110941410065, + 0.006148665212094784, + 0.005910816136747599, + 0.004027365706861019, + 0.00516835693269968, + 0.0042784432880580425, + 0.007541025523096323, + 0.0038476702757179737, + 0.004315650090575218, + 0.005267573054879904, + 0.016335677355527878, + 0.018179062753915787, + 0.010147842578589916, + 0.009495090693235397, + 0.07572600990533829, + 0.02066851034760475, + 0.016744503751397133, + 0.009490164928138256, + 0.012269978411495686, + 0.024406572803854942, + 0.00737176975235343, + 0.006711805704981089, + 0.0326894149184227, + 0.015554433688521385, + 0.008141152560710907, + 0.005253626499325037, + 0.014396412298083305, + 0.023264842107892036, + 0.008672263473272324, + 0.010906829498708248, + 0.031532686203718185, + 0.014033956453204155, + 0.008929743431508541, + 0.015037400647997856, + 0.0053369831293821335, + 0.00901959091424942, + 0.005099093541502953, + 0.008758066222071648, + 0.010504120960831642, + 0.005184041801840067, + 0.006130421534180641, + 0.006552911829203367, + 0.008047022856771946, + 0.01253762561827898, + 0.00308039877563715, + 0.007343782112002373, + 0.030062837526202202, + 0.01277014147490263, + 0.007128600031137466, + 0.0050966013222932816, + 0.00850614532828331, + 0.01178792491555214, + 0.003165489761158824, + 0.00646804366260767, + 0.029212798923254013, + 0.01172008365392685, + 0.006501183845102787, + 0.005101222079247236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.002199375070631504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739031057804823 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006739031057804823, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739031057804823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.214483380317688, + 0.10963846743106842, + 0.05779410898685455, + 0.0552937313914299, + 0.1316671371459961, + 0.27854397892951965, + 0.04558924213051796, + 0.03501148521900177, + 0.022376658394932747, + 0.04960182309150696 + ], + "credit_map": [ + 0.02449895441532135, + 0.03596712276339531, + 0.01034067664295435, + 0.019155919551849365, + 0.0374036580324173, + 0.029263677075505257, + 0.022961338981986046, + 0.03489205241203308, + 0.013154714368283749, + 0.01228591613471508, + 0.010886757634580135, + 0.019887631759047508, + 0.014401638880372047, + 0.011581615544855595, + 0.010981163010001183, + 0.0164590273052454, + 0.006782425567507744, + 0.004769925028085709, + 0.007086119148880243, + 0.006601139437407255, + 0.007978852838277817, + 0.006054472178220749, + 0.007644291035830975, + 0.010876882821321487, + 0.009629330597817898, + 0.006630726158618927, + 0.00756239565089345, + 0.0042786928825080395, + 0.008618819527328014, + 0.004755089059472084, + 0.004688854329288006, + 0.009129820391535759, + 0.014467602595686913, + 0.01229851134121418, + 0.009971429593861103, + 0.010374484583735466, + 0.04871845990419388, + 0.014056231826543808, + 0.014499161392450333, + 0.00728125125169754, + 0.028833985328674316, + 0.05804285407066345, + 0.016860373318195343, + 0.022038383409380913, + 0.08141415566205978, + 0.03802388906478882, + 0.018315797671675682, + 0.015014572069048882, + 0.005167539231479168, + 0.004697810858488083, + 0.0047533148899674416, + 0.004908856004476547, + 0.006512359715998173, + 0.005021972116082907, + 0.00585981598123908, + 0.008667576126754284, + 0.003161653643473983, + 0.007147922646254301, + 0.002970711328089237, + 0.003690173616632819, + 0.007843860425055027, + 0.0036902623251080513, + 0.0028323023580014706, + 0.00367459817789495, + 0.002364800777286291, + 0.002952644834294915, + 0.00211187731474638, + 0.0029328051023185253, + 0.004559285473078489, + 0.0026480716187506914, + 0.0023246428463608027, + 0.0024825306609272957, + 0.004162687808275223, + 0.015037079341709614, + 0.0036916874814778566, + 0.0033397777006030083, + 0.013546244241297245, + 0.004432674031704664, + 0.0023796246387064457, + 0.0030120499432086945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.006739031057804823 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot horizontally and rendered in slow motion, focuses on a heavy, dusty, dark wooden table. Lying quietly upon the tabletop lies an ancient parchment scroll, its paper varying shades of yellow, its edges frayed and slightly charred, its surface covered in fine wrinkles and the marks of age. Suddenly, a powerful, invisible gust of wind blows violently, instantly lifting a corner of the scroll and sending the paper trembling and rising. This movement stirs the long-settled dust on the tabletop, carrying countless golden particles into the air. Within a dramatic beam of light cast from above, a powerful Tyndall effect, the resulting image instantly forms a swirling, shimmering golden nebula. Each speck of dust shimmers like tiny pieces of gold foil, dancing around the slowly unrolling scroll. The entire scene is shrouded in a mysterious, solemn atmosphere imbued with ancient magic, painted in warm, deep tones.", + "seed": 29259, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_2/index_6888_seed_29259_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020050454512238504, + "vlm_logprob": -0.020050454512238504, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.020050454512238504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048909660428762436 + ], + "opposite_logprob_tokens": [ + -3.1739096641540527 + ], + "logprob_reward": -0.048909660428762436, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.048909660428762436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22928862273693085, + 0.05309063196182251, + 0.07044609636068344, + 0.06951078772544861, + 0.15239296853542328, + 0.047251079231500626, + 0.041616540402173996, + 0.04954815283417702, + 0.2006985992193222, + 0.08615651726722717 + ], + "credit_map": [ + 0.02128722332417965, + 0.02323313243687153, + 0.02709050476551056, + 0.02234569378197193, + 0.03986034169793129, + 0.02815430983901024, + 0.03586769849061966, + 0.03144971653819084, + 0.005727867130190134, + 0.004675513599067926, + 0.008468138054013252, + 0.0033309150021523237, + 0.00907806120812893, + 0.007744612172245979, + 0.006587616167962551, + 0.007477912120521069, + 0.012967249378561974, + 0.009690389968454838, + 0.005190495867282152, + 0.006538885645568371, + 0.008334321901202202, + 0.00651747127994895, + 0.011449542827904224, + 0.009757742285728455, + 0.012972932308912277, + 0.009878002107143402, + 0.004761368036270142, + 0.006799773313105106, + 0.00840681791305542, + 0.009646005928516388, + 0.007983541116118431, + 0.009062353521585464, + 0.016180871054530144, + 0.049293696880340576, + 0.004087227862328291, + 0.005426180548965931, + 0.044914938509464264, + 0.012159660458564758, + 0.009965752251446247, + 0.010364660061895847, + 0.005839156452566385, + 0.007912804372608662, + 0.0033962754532694817, + 0.005174239631742239, + 0.007341335527598858, + 0.004349014721810818, + 0.005444427020847797, + 0.007793826051056385, + 0.005099239759147167, + 0.005837637465447187, + 0.002986086532473564, + 0.0031630350276827812, + 0.01015052292495966, + 0.006966141052544117, + 0.003412966849282384, + 0.004000912420451641, + 0.004020476248115301, + 0.01099622156471014, + 0.001410729717463255, + 0.001513374736532569, + 0.018404262140393257, + 0.008784059435129166, + 0.002161746146157384, + 0.0022572805173695087, + 0.013273797929286957, + 0.06349258124828339, + 0.0020200808066874743, + 0.0024741729721426964, + 0.08000298589468002, + 0.02993796207010746, + 0.004016272723674774, + 0.005480749532580376, + 0.0047519998624920845, + 0.019905557855963707, + 0.001342911971732974, + 0.0016466276720166206, + 0.03643113374710083, + 0.0178571455180645, + 0.0020063125994056463, + 0.0022148259449750185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.048909660428762436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639951538294554 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.007639951538294554, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639951538294554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27221354842185974, + 0.0720285028219223, + 0.06437772512435913, + 0.11003396660089493, + 0.1430296003818512, + 0.05215132236480713, + 0.045897744596004486, + 0.04585152864456177, + 0.1382935345172882, + 0.05612250044941902 + ], + "credit_map": [ + 0.01879933290183544, + 0.020237181335687637, + 0.03732568770647049, + 0.03142860159277916, + 0.03307972103357315, + 0.03302090987563133, + 0.049876950681209564, + 0.04844518005847931, + 0.008378656581044197, + 0.0074479300528764725, + 0.008937784470617771, + 0.004645220469683409, + 0.01052558422088623, + 0.009735703468322754, + 0.010147957131266594, + 0.01220966037362814, + 0.008244846947491169, + 0.005721887107938528, + 0.0050432877615094185, + 0.004939664155244827, + 0.007965222001075745, + 0.00843617133796215, + 0.012480542995035648, + 0.011546100489795208, + 0.018503384664654732, + 0.015750234946608543, + 0.008389757014811039, + 0.012672063894569874, + 0.013545162044465542, + 0.011889743618667126, + 0.01455497182905674, + 0.01472864393144846, + 0.016260644420981407, + 0.05334211513400078, + 0.002723056823015213, + 0.0024926376063376665, + 0.04734153300523758, + 0.009232250973582268, + 0.005007994826883078, + 0.006629372015595436, + 0.005229804199188948, + 0.008335689082741737, + 0.003258864162489772, + 0.004079868085682392, + 0.008227181620895863, + 0.005731293000280857, + 0.007076167035847902, + 0.010212456807494164, + 0.004555669147521257, + 0.00660107983276248, + 0.003568659769371152, + 0.00432963902130723, + 0.008941484615206718, + 0.005957997404038906, + 0.005341213662177324, + 0.006602001842111349, + 0.003057412803173065, + 0.004780225455760956, + 0.0012788937892764807, + 0.0016534053720533848, + 0.017541931942105293, + 0.011870098300278187, + 0.002178946975618601, + 0.003490615403279662, + 0.012627559714019299, + 0.04205117002129555, + 0.0033859997056424618, + 0.0028313028160482645, + 0.04939725250005722, + 0.01722540333867073, + 0.004364214837551117, + 0.006410631816834211, + 0.0051162950694561005, + 0.01234124694019556, + 0.0023337174206972122, + 0.0033993772231042385, + 0.014203974045813084, + 0.0055358028039336205, + 0.0047791143879294395, + 0.008412973023951054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.007639951538294554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805148020386696 + ], + "opposite_logprob_tokens": [ + -3.5298051834106445 + ], + "logprob_reward": -0.029805148020386696, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805148020386696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15740330517292023, + 0.0925387516617775, + 0.1014622375369072, + 0.05788812413811684, + 0.09573116898536682, + 0.1918446570634842, + 0.03431771695613861, + 0.05855386704206467, + 0.10980092734098434, + 0.10045924037694931 + ], + "credit_map": [ + 0.01607193984091282, + 0.014568203128874302, + 0.021522637456655502, + 0.015835484489798546, + 0.02197297476232052, + 0.017004264518618584, + 0.026525424793362617, + 0.023902365937829018, + 0.0080005694180727, + 0.009281091392040253, + 0.014010475017130375, + 0.005253975745290518, + 0.012051970697939396, + 0.011805436573922634, + 0.014451459050178528, + 0.01768377609550953, + 0.015674810856580734, + 0.010298608802258968, + 0.006561721209436655, + 0.008077790029346943, + 0.011148961260914803, + 0.009532657451927662, + 0.024549970403313637, + 0.015617712400853634, + 0.008928576484322548, + 0.005663160700351, + 0.0034956997260451317, + 0.004792679566890001, + 0.006841979455202818, + 0.011896610260009766, + 0.007757261395454407, + 0.008512157946825027, + 0.01091717928647995, + 0.026743564754724503, + 0.003936511930078268, + 0.00566965714097023, + 0.022462736815214157, + 0.008045532740652561, + 0.00965303648263216, + 0.008302954956889153, + 0.016586527228355408, + 0.02407701127231121, + 0.010869413614273071, + 0.015509537421166897, + 0.03773074597120285, + 0.02179630845785141, + 0.03700247034430504, + 0.028272639960050583, + 0.00494068069383502, + 0.004612862132489681, + 0.002652895636856556, + 0.002718736883252859, + 0.006121833808720112, + 0.004510271362960339, + 0.0045827156864106655, + 0.004177719820290804, + 0.0029095790814608335, + 0.011749902740120888, + 0.0010093169985339046, + 0.0010273506632074714, + 0.024086838588118553, + 0.012625684961676598, + 0.0019090895075351, + 0.0032361040357500315, + 0.010785238817334175, + 0.025968626141548157, + 0.002927024383097887, + 0.0024216047022491693, + 0.03939765691757202, + 0.01838783361017704, + 0.003357494715601206, + 0.006555452942848206, + 0.006933669093996286, + 0.02612970769405365, + 0.0019365422194823623, + 0.0014310578117147088, + 0.04030013829469681, + 0.018147079274058342, + 0.002328283153474331, + 0.003252775641158223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.029805148020386696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255098920315504 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255098920315504, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255098920315504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13902930915355682, + 0.07095290720462799, + 0.05517209321260452, + 0.06513343006372452, + 0.06957748532295227, + 0.12189740687608719, + 0.0667463093996048, + 0.05754191800951958, + 0.12277307361364365, + 0.23117601871490479 + ], + "credit_map": [ + 0.013519287109375, + 0.01598072238266468, + 0.012622842565178871, + 0.01095210574567318, + 0.017643973231315613, + 0.017215127125382423, + 0.026841117069125175, + 0.02425413392484188, + 0.007376123685389757, + 0.006810963153839111, + 0.01228488702327013, + 0.004177872557193041, + 0.008732231333851814, + 0.010524217039346695, + 0.009140902198851109, + 0.011905712075531483, + 0.008346150629222393, + 0.007892347872257233, + 0.003026880556717515, + 0.004889547824859619, + 0.005862266756594181, + 0.007517685182392597, + 0.008262056857347488, + 0.009375163353979588, + 0.01067684218287468, + 0.0134916752576828, + 0.002801349386572838, + 0.004515610169619322, + 0.011068112216889858, + 0.007730709854513407, + 0.007134699262678623, + 0.007714440580457449, + 0.009996173903346062, + 0.016622047871351242, + 0.003407432232052088, + 0.003561008023098111, + 0.014617987908422947, + 0.006551947444677353, + 0.007874283008277416, + 0.006946612149477005, + 0.01073305495083332, + 0.015513723716139793, + 0.006552489474415779, + 0.012137485668063164, + 0.024332629516720772, + 0.015027823857963085, + 0.017360644415020943, + 0.020239563658833504, + 0.007417148910462856, + 0.011671802029013634, + 0.004629030358046293, + 0.004612216725945473, + 0.013807591050863266, + 0.008447641506791115, + 0.006707583554089069, + 0.0094533022493124, + 0.00850823987275362, + 0.00994984619319439, + 0.0031065044458955526, + 0.002665378386154771, + 0.012127552181482315, + 0.013055146671831608, + 0.0030304789543151855, + 0.005098773632198572, + 0.012981810607016087, + 0.03964366763830185, + 0.003881396958604455, + 0.0026903299149125814, + 0.0417555496096611, + 0.012771938927471638, + 0.0042766244150698185, + 0.004771761130541563, + 0.010326603427529335, + 0.0553935244679451, + 0.0016906263772398233, + 0.0025593582540750504, + 0.09446720033884048, + 0.04392269626259804, + 0.007481512147933245, + 0.015334520488977432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.005255098920315504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a heavy, dusty, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient parchment scroll with yellow, frayed, and charred edges visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a gust of wind lift a corner of the scroll and stir golden dust particles into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a swirling, shimmering golden nebula formed by dust within a dramatic beam of light showing a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in slow motion with warm, deep tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642413653433323 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642413653433323, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642413653433323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0852135494351387, + 0.06392271816730499, + 0.05935287103056908, + 0.06533235311508179, + 0.024026691913604736, + 0.18158559501171112, + 0.03087150678038597, + 0.0858498141169548, + 0.11434397101402283, + 0.28950098156929016 + ], + "credit_map": [ + 0.00815814919769764, + 0.01086482871323824, + 0.011129169724881649, + 0.008705692365765572, + 0.011607632972300053, + 0.012498626485466957, + 0.01316085085272789, + 0.009088603779673576, + 0.00580740999430418, + 0.006158425472676754, + 0.008704042062163353, + 0.004588079173117876, + 0.010580982081592083, + 0.008148361928761005, + 0.007736401166766882, + 0.012199013493955135, + 0.01692591980099678, + 0.007897780276834965, + 0.005043625831604004, + 0.004705492407083511, + 0.004787568934261799, + 0.006940513849258423, + 0.006418109405785799, + 0.0066338516771793365, + 0.007161343470215797, + 0.014851201325654984, + 0.003620697883889079, + 0.0035252675879746675, + 0.012083403766155243, + 0.006330265197902918, + 0.009300160221755505, + 0.008460008539259434, + 0.0037124056834727526, + 0.003583748359233141, + 0.001584123820066452, + 0.0013916331809014082, + 0.0028440614696592093, + 0.004928065929561853, + 0.0022146354895085096, + 0.0037680179812014103, + 0.013933742418885231, + 0.02276313491165638, + 0.00860654003918171, + 0.012027979828417301, + 0.03923872485756874, + 0.02268737182021141, + 0.03399622440338135, + 0.028331853449344635, + 0.003244041930884123, + 0.003687653224915266, + 0.002193534979596734, + 0.002277929801493883, + 0.007108961697667837, + 0.0049574850127100945, + 0.004204993136227131, + 0.0031969030387699604, + 0.0039118085987865925, + 0.01689336448907852, + 0.001587501959875226, + 0.00241504842415452, + 0.03523702174425125, + 0.0192521084100008, + 0.002341018058359623, + 0.004211940802633762, + 0.008044470101594925, + 0.03692362830042839, + 0.0014494159258902073, + 0.0012867797631770372, + 0.045349203050136566, + 0.016435643658041954, + 0.002113309456035495, + 0.0027415065560489893, + 0.014692594297230244, + 0.07907356321811676, + 0.0018106993520632386, + 0.002787460805848241, + 0.12677288055419922, + 0.0572819821536541, + 0.003541438840329647, + 0.0035403186921030283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.008642413653433323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_5/index_tensor([5659, 5659])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_5/index_tensor([5659, 5659])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4917e1e4964578496bd1248872e83344bfa6c2dd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_5/index_tensor([5659, 5659])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly dynamic close-up tracking shot, set amidst the hum of machinery in a 1950s Chinese automobile factory, follows the tires of a massive, heavy transport truck as they roll at high speed. The deep treads, etched with dirt and gravel, leave a trail of rapid movement across the rough concrete floor. Suddenly, with a sharp screeching of brakes, the camera screeches to a halt, locking focus precisely on the stationary tire's sidewall. At that moment, a worker's arm, calloused and thick-knuckled, emerges from the rolled-up sleeves of his blue work clothes. Steadily and firmly grasping a stub of white chalk, he decisively scratches a clear registration mark on the rough rubber surface. At the same time, the wood chips and dust that flew up in the air due to the sudden stop of the vehicle seemed to be frozen in time, suspended in the air by the volumetric oblique beams cast from the high windows of the workshop. Every tiny particle was clearly visible, and its irregular shape formed an instant sculpture full of industrial power in the still picture. Under the harsh contrast of light and shadow, the whole scene presents an atmosphere of industrial aesthetics that is focused, powerful and full of the imprint of the times.", + "seed": 21339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_5/index_5659_seed_21339_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1939275985583663, + "vlm_logprob": -0.1939275985583663, + "vlm_yes_no_margin": 2.5249999523162843, + "vlm_token_credit": -0.1939275985583663, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7153248190879822 + ], + "opposite_logprob_tokens": [ + -0.7153248190879822 + ], + "logprob_reward": -0.7153248190879822, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7153248190879822, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16746081411838531, + 0.20007053017616272, + 0.022252844646573067, + 0.11340510845184326, + 0.157615527510643, + 0.05486959591507912, + 0.056953590363264084, + 0.14384685456752777, + 0.04281992465257645, + 0.040705226361751556 + ], + "credit_map": [ + 0.014743461273610592, + 0.014202886261045933, + 0.010684551671147346, + 0.02742822840809822, + 0.02048931084573269, + 0.022472189739346504, + 0.020290914922952652, + 0.03714926168322563, + 0.011713381856679916, + 0.006574979517608881, + 0.011981145478785038, + 0.014216653071343899, + 0.011517182923853397, + 0.05994692072272301, + 0.059578295797109604, + 0.024541957303881645, + 0.0032323335763067007, + 0.0030757870990782976, + 0.0032989562023431063, + 0.0024991000536829233, + 0.002795080654323101, + 0.002351590897887945, + 0.002564478199928999, + 0.00243551773019135, + 0.008682763203978539, + 0.005398612469434738, + 0.012070792727172375, + 0.012602847069501877, + 0.00852679181843996, + 0.019292186945676804, + 0.008901110850274563, + 0.03793000057339668, + 0.010201564989984035, + 0.00978216715157032, + 0.025139739736914635, + 0.030597038567066193, + 0.014278174377977848, + 0.01836235448718071, + 0.01918407715857029, + 0.03007039986550808, + 0.004690084606409073, + 0.006540027447044849, + 0.006742261350154877, + 0.011586969718337059, + 0.005282172001898289, + 0.005392163060605526, + 0.008566691540181637, + 0.006069221533834934, + 0.006417885888367891, + 0.006629110313951969, + 0.005803845357149839, + 0.010434916242957115, + 0.010723053477704525, + 0.006024199537932873, + 0.005810979288071394, + 0.005109588615596294, + 0.012715651653707027, + 0.018701158463954926, + 0.017058445140719414, + 0.013782816007733345, + 0.025300560519099236, + 0.020661480724811554, + 0.015567169524729252, + 0.020059555768966675, + 0.00622759060934186, + 0.006070180796086788, + 0.00540607376024127, + 0.003130455734208226, + 0.004726181272417307, + 0.007412692531943321, + 0.004812593571841717, + 0.005034152884036303, + 0.005599652882665396, + 0.007481392007321119, + 0.005525998771190643, + 0.002323646331205964, + 0.005867854226380587, + 0.006676129065454006, + 0.003782643936574459, + 0.0034479047171771526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.7153248190879822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14288540184497833 + ], + "opposite_logprob_tokens": [ + -2.017885446548462 + ], + "logprob_reward": -0.14288540184497833, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14288540184497833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12023486196994781, + 0.03340841084718704, + 0.046077530831098557, + 0.06671807169914246, + 0.08677886426448822, + 0.0597696416079998, + 0.08323642611503601, + 0.43013057112693787, + 0.023218398913741112, + 0.0504271425306797 + ], + "credit_map": [ + 0.014883254654705524, + 0.010748764500021935, + 0.012267164885997772, + 0.017130734398961067, + 0.019061986356973648, + 0.01480740774422884, + 0.01605536788702011, + 0.015280196443200111, + 0.0036656754091382027, + 0.003027576021850109, + 0.004067291505634785, + 0.0035320143215358257, + 0.007425736170262098, + 0.003856746247038245, + 0.004504526033997536, + 0.0033288472332060337, + 0.007085925899446011, + 0.004812484607100487, + 0.008283347822725773, + 0.006587825249880552, + 0.00466048251837492, + 0.00415400555357337, + 0.005013111978769302, + 0.005480347666889429, + 0.006233385764062405, + 0.005564144812524319, + 0.01076581608504057, + 0.00769036915153265, + 0.006510799750685692, + 0.009094924665987492, + 0.005242705810815096, + 0.015615936368703842, + 0.007207571528851986, + 0.006036980077624321, + 0.007877126336097717, + 0.009331637062132359, + 0.010277947410941124, + 0.012264037504792213, + 0.011744402348995209, + 0.022039169445633888, + 0.004497482441365719, + 0.004550776910036802, + 0.01166126225143671, + 0.012617905624210835, + 0.005200619343668222, + 0.006337685510516167, + 0.00863050576299429, + 0.006273412145674229, + 0.006115019787102938, + 0.004608187358826399, + 0.007467007730156183, + 0.019390247762203217, + 0.00818725861608982, + 0.005522118881344795, + 0.006067287176847458, + 0.025879304856061935, + 0.03583985194563866, + 0.05966303497552872, + 0.049614738672971725, + 0.04289435222744942, + 0.0727248415350914, + 0.059920813888311386, + 0.04943395406007767, + 0.06003899872303009, + 0.002769010839983821, + 0.004903113469481468, + 0.0031709123868495226, + 0.0018567182123661041, + 0.0020959225948899984, + 0.0033584325574338436, + 0.0018688365817070007, + 0.003195454366505146, + 0.007065515033900738, + 0.011294903233647346, + 0.00654192641377449, + 0.002986280247569084, + 0.006603555753827095, + 0.00835281703621149, + 0.004093791823834181, + 0.0034883590415120125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.14288540184497833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214691706001759 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214691706001759, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214691706001759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0659785196185112, + 0.06453031301498413, + 0.04531428590416908, + 0.10093450546264648, + 0.26788729429244995, + 0.0535610094666481, + 0.11608894169330597, + 0.20670858025550842, + 0.029257338494062424, + 0.049739181995391846 + ], + "credit_map": [ + 0.007241507992148399, + 0.005632623564451933, + 0.0066315471194684505, + 0.006345534697175026, + 0.011692184954881668, + 0.00868704542517662, + 0.0118282251060009, + 0.007919860072433949, + 0.004840418230742216, + 0.00786658562719822, + 0.0050123343244194984, + 0.004635583143681288, + 0.015917448326945305, + 0.009516166523098946, + 0.010743048042058945, + 0.0059987385757267475, + 0.0044108484871685505, + 0.006988478358834982, + 0.00527528952807188, + 0.004520706832408905, + 0.0062300353311002254, + 0.007986991666257381, + 0.005190043244510889, + 0.004711897578090429, + 0.006562619004398584, + 0.006259697023779154, + 0.00916803628206253, + 0.009567159228026867, + 0.009565621614456177, + 0.0198644008487463, + 0.007183844689279795, + 0.032763127237558365, + 0.020399905741214752, + 0.01748540624976158, + 0.022005030885338783, + 0.03279036656022072, + 0.030940933153033257, + 0.03780623525381088, + 0.041818998754024506, + 0.06464042514562607, + 0.004082311410456896, + 0.007770748808979988, + 0.010266559198498726, + 0.01557437889277935, + 0.0034236470237374306, + 0.0038252833765000105, + 0.004416077397763729, + 0.004202007781714201, + 0.022541474550962448, + 0.011376740410923958, + 0.02292911522090435, + 0.015176442451775074, + 0.009239389561116695, + 0.009827068075537682, + 0.008290531113743782, + 0.016708187758922577, + 0.01712433435022831, + 0.026703685522079468, + 0.023900218307971954, + 0.020465319976210594, + 0.0361798033118248, + 0.030341004952788353, + 0.022976495325565338, + 0.029017718508839607, + 0.0046600415371358395, + 0.005853206384927034, + 0.003632717067375779, + 0.0022241079714149237, + 0.0030240570195019245, + 0.003701253794133663, + 0.0028435990680009127, + 0.0033183572813868523, + 0.007789053022861481, + 0.012655436061322689, + 0.006860753986984491, + 0.003573240013793111, + 0.0046509793028235435, + 0.00627564825117588, + 0.004217833746224642, + 0.0037162431981414557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.014214691706001759 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901488989591599 + ], + "opposite_logprob_tokens": [ + -2.579014778137207 + ], + "logprob_reward": -0.07901488989591599, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901488989591599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11750032007694244, + 0.09255968034267426, + 0.03199734911322594, + 0.06405569612979889, + 0.13721439242362976, + 0.021388908848166466, + 0.08548630774021149, + 0.08560584485530853, + 0.17805048823356628, + 0.1861409991979599 + ], + "credit_map": [ + 0.010873780585825443, + 0.007525265216827393, + 0.010787833482027054, + 0.03216666728258133, + 0.007582018617540598, + 0.00957658514380455, + 0.010250973515212536, + 0.028737206012010574, + 0.005670466925948858, + 0.00613116379827261, + 0.006028363015502691, + 0.00665398221462965, + 0.011805286630988121, + 0.022326769307255745, + 0.024063603952527046, + 0.009880045428872108, + 0.004775206092745066, + 0.0042949882335960865, + 0.005840035621076822, + 0.004601969383656979, + 0.0029484794940799475, + 0.0032066339626908302, + 0.0031611621379852295, + 0.003168876050040126, + 0.00385844218544662, + 0.005199129693210125, + 0.0053405421786010265, + 0.005912113003432751, + 0.011328337714076042, + 0.010861079208552837, + 0.006376992911100388, + 0.015179061330854893, + 0.0059081148356199265, + 0.005256956908851862, + 0.033183157444000244, + 0.04256484657526016, + 0.008282748982310295, + 0.020180393010377884, + 0.009895868599414825, + 0.011942322365939617, + 0.0019522507209330797, + 0.002469826489686966, + 0.0021501074079424143, + 0.0037393192760646343, + 0.002217170549556613, + 0.003290321212261915, + 0.0032205351162701845, + 0.002349379938095808, + 0.01186530850827694, + 0.010059619322419167, + 0.01031541172415018, + 0.015630638226866722, + 0.011209402233362198, + 0.007632378488779068, + 0.008727332577109337, + 0.010046219453215599, + 0.014937969855964184, + 0.014664236456155777, + 0.012691646814346313, + 0.006962968967854977, + 0.01145041361451149, + 0.009751670993864536, + 0.0069721536710858345, + 0.00817478634417057, + 0.025585200637578964, + 0.050272099673748016, + 0.025023657828569412, + 0.009943564422428608, + 0.017656881362199783, + 0.015511139295995235, + 0.018202947452664375, + 0.01585499756038189, + 0.033724647015333176, + 0.05318902060389519, + 0.03012440726161003, + 0.011830722913146019, + 0.016454122960567474, + 0.016324449330568314, + 0.014409712515771389, + 0.010083942674100399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.07901488989591599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12198701500892639, + 0.31466278433799744, + 0.03605930879712105, + 0.2109227031469345, + 0.07869858294725418, + 0.058943070471286774, + 0.04516957700252533, + 0.05851610749959946, + 0.04481195658445358, + 0.03022889979183674 + ], + "credit_map": [ + 0.013904603198170662, + 0.008834339678287506, + 0.021008597686886787, + 0.026895087212324142, + 0.010433989576995373, + 0.00921582616865635, + 0.01444404199719429, + 0.017250530421733856, + 0.021842053160071373, + 0.009613245725631714, + 0.018408948555588722, + 0.02026478201150894, + 0.016874665394425392, + 0.09340061992406845, + 0.09839664399623871, + 0.03586181625723839, + 0.004632361698895693, + 0.004765940364450216, + 0.004598057828843594, + 0.005227853078395128, + 0.004636525642126799, + 0.0036668472457677126, + 0.0041763074696063995, + 0.004355414770543575, + 0.018988804891705513, + 0.018119731917977333, + 0.018794458359479904, + 0.022465864196419716, + 0.017995044589042664, + 0.035468682646751404, + 0.016214102506637573, + 0.06287601590156555, + 0.005644956603646278, + 0.004663111642003059, + 0.010126493871212006, + 0.013195060193538666, + 0.00787835381925106, + 0.013142404146492481, + 0.010037465021014214, + 0.01401074230670929, + 0.00454636849462986, + 0.00871041975915432, + 0.009522251784801483, + 0.016260260716080666, + 0.003757823957130313, + 0.004228896461427212, + 0.007046394981443882, + 0.0048706550151109695, + 0.007602858357131481, + 0.006434016861021519, + 0.004981982987374067, + 0.006316104903817177, + 0.005046901758760214, + 0.0034678701777011156, + 0.0037685581482946873, + 0.007551281712949276, + 0.007818317972123623, + 0.00998148787766695, + 0.006683497689664364, + 0.005270600318908691, + 0.010417912155389786, + 0.006928244139999151, + 0.004874133504927158, + 0.00654190918430686, + 0.0076677510514855385, + 0.010412562638521194, + 0.0056373015977442265, + 0.0030978736467659473, + 0.0042363121174275875, + 0.005008423235267401, + 0.004247175995260477, + 0.004504555836319923, + 0.00393762718886137, + 0.005443948321044445, + 0.004104372579604387, + 0.0023813035804778337, + 0.004368470050394535, + 0.004224901087582111, + 0.0024767182767391205, + 0.003291558241471648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.01819819025695324 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic close-up tracking shot, set amidst the hum of machinery in a 1950s Chinese automobile factory, follows the tires of a massive, heavy transport truck as they roll at high speed. The deep treads, etched with dirt and gravel, leave a trail of rapid movement across the rough concrete floor. Suddenly, with a sharp screeching of brakes, the camera screeches to a halt, locking focus precisely on the stationary tire's sidewall. At that moment, a worker's arm, calloused and thick-knuckled, emerges from the rolled-up sleeves of his blue work clothes. Steadily and firmly grasping a stub of white chalk, he decisively scratches a clear registration mark on the rough rubber surface. At the same time, the wood chips and dust that flew up in the air due to the sudden stop of the vehicle seemed to be frozen in time, suspended in the air by the volumetric oblique beams cast from the high windows of the workshop. Every tiny particle was clearly visible, and its irregular shape formed an instant sculpture full of industrial power in the still picture. Under the harsh contrast of light and shadow, the whole scene presents an atmosphere of industrial aesthetics that is focused, powerful and full of the imprint of the times.", + "seed": 21340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_5/index_5659_seed_21340_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09540188442915679, + "vlm_logprob": -0.09540188442915679, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.09540188442915679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19502025842666626 + ], + "opposite_logprob_tokens": [ + -1.8200201988220215 + ], + "logprob_reward": -0.19502025842666626, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19502025842666626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09790820628404617, + 0.04367014765739441, + 0.06396426260471344, + 0.21593526005744934, + 0.050671424716711044, + 0.03614222630858421, + 0.21897850930690765, + 0.03530057519674301, + 0.0984385758638382, + 0.13899074494838715 + ], + "credit_map": [ + 0.010962494648993015, + 0.010545448400080204, + 0.0155480420216918, + 0.01752312481403351, + 0.00950595922768116, + 0.00987907312810421, + 0.00976640172302723, + 0.014177672564983368, + 0.00498327799141407, + 0.005352685693651438, + 0.0055359709076583385, + 0.006828304845839739, + 0.004739526659250259, + 0.005083381664007902, + 0.007408542558550835, + 0.0037384622264653444, + 0.007413381244987249, + 0.006295701954513788, + 0.010158177465200424, + 0.013843884691596031, + 0.005726649891585112, + 0.00847590621560812, + 0.00644663767889142, + 0.005603931378573179, + 0.006750278174877167, + 0.009419224224984646, + 0.05952572077512741, + 0.07949390262365341, + 0.009135855361819267, + 0.013858948834240437, + 0.02345222420990467, + 0.014299135655164719, + 0.004121622536331415, + 0.004535323940217495, + 0.007163387257605791, + 0.010520915500819683, + 0.009563145227730274, + 0.0031024033669382334, + 0.008283180184662342, + 0.0033814520575106144, + 0.002837509848177433, + 0.002740158000960946, + 0.006236847955733538, + 0.005137992557138205, + 0.005779081489890814, + 0.003852835623547435, + 0.0062471600249409676, + 0.0033106443006545305, + 0.01214609108865261, + 0.012286962009966373, + 0.0685613751411438, + 0.020098214969038963, + 0.07322702556848526, + 0.014748395420610905, + 0.010817630216479301, + 0.007092831656336784, + 0.0029301191680133343, + 0.0038647009059786797, + 0.007677340880036354, + 0.003927980083972216, + 0.00427615363150835, + 0.004353721626102924, + 0.004020179156213999, + 0.004250382538884878, + 0.011585510335862637, + 0.01531143207103014, + 0.012940023094415665, + 0.016045542433857918, + 0.011130842380225658, + 0.014611338265240192, + 0.0076662455685436726, + 0.009147649630904198, + 0.01094425655901432, + 0.016350876539945602, + 0.01485864445567131, + 0.021667804569005966, + 0.028595441952347755, + 0.021936895325779915, + 0.012822295539081097, + 0.0118145402520895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.19502025842666626 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016642689704895 + ], + "opposite_logprob_tokens": [ + -1.7016642093658447 + ], + "logprob_reward": -0.2016642689704895, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016642689704895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13717994093894958, + 0.15115875005722046, + 0.06815154105424881, + 0.10677848011255264, + 0.08270158618688583, + 0.05873434245586395, + 0.16975867748260498, + 0.057191357016563416, + 0.11054138094186783, + 0.05780389532446861 + ], + "credit_map": [ + 0.01510925404727459, + 0.015035698190331459, + 0.02276398055255413, + 0.017272794619202614, + 0.010392577387392521, + 0.016126833856105804, + 0.016507428139448166, + 0.023971380665898323, + 0.01938624307513237, + 0.028968170285224915, + 0.01987425796687603, + 0.019140468910336494, + 0.015361589379608631, + 0.017322953790426254, + 0.020606445148587227, + 0.01049861777573824, + 0.007106814067810774, + 0.008961126208305359, + 0.010128498077392578, + 0.01354576088488102, + 0.006576055195182562, + 0.00849010981619358, + 0.0076053813099861145, + 0.00573780108243227, + 0.006539194379001856, + 0.0041000694036483765, + 0.0279228575527668, + 0.035004984587430954, + 0.004546248819679022, + 0.006505978759378195, + 0.014289824292063713, + 0.007869336754083633, + 0.0068950834684073925, + 0.013613606803119183, + 0.010986213572323322, + 0.01303660124540329, + 0.013972553424537182, + 0.0077129025012254715, + 0.008187208324670792, + 0.008297428488731384, + 0.003573294961825013, + 0.004409239161759615, + 0.0143556147813797, + 0.00974881649017334, + 0.0054234955459833145, + 0.007127311080694199, + 0.00966082327067852, + 0.004435752984136343, + 0.009118589572608471, + 0.00890334788709879, + 0.055253878235816956, + 0.013882925733923912, + 0.05644356459379196, + 0.009728281758725643, + 0.009520172141492367, + 0.006907941307872534, + 0.003297460963949561, + 0.004916172940284014, + 0.009562262333929539, + 0.006105826236307621, + 0.008039088919758797, + 0.007042632903903723, + 0.009106473065912724, + 0.009121445007622242, + 0.010650837793946266, + 0.014108262024819851, + 0.014151225797832012, + 0.01815485581755638, + 0.009912211447954178, + 0.018918639048933983, + 0.013810569420456886, + 0.010834791697561741, + 0.00787532702088356, + 0.01004758570343256, + 0.007250096648931503, + 0.007747085299342871, + 0.008899519219994545, + 0.006766658741980791, + 0.0047087520360946655, + 0.0045088715851306915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.2016642689704895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005267906002700329 + ], + "opposite_logprob_tokens": [ + -5.255268096923828 + ], + "logprob_reward": -0.005267906002700329, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005267906002700329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1301797330379486, + 0.15369956195354462, + 0.07661796361207962, + 0.07490666955709457, + 0.07313501089811325, + 0.0515286885201931, + 0.1006196141242981, + 0.05846162140369415, + 0.1308594048023224, + 0.1499917358160019 + ], + "credit_map": [ + 0.013307535089552402, + 0.013261989690363407, + 0.02995668724179268, + 0.010708759538829327, + 0.009760799817740917, + 0.01334343384951353, + 0.01039178017526865, + 0.029448751360177994, + 0.018705159425735474, + 0.028197402134537697, + 0.021073590964078903, + 0.019783901050686836, + 0.016648797318339348, + 0.016211392357945442, + 0.021140269935131073, + 0.011939053423702717, + 0.005807765759527683, + 0.008102071471512318, + 0.012922553345561028, + 0.018792249262332916, + 0.006470173131674528, + 0.006837644148617983, + 0.010447126813232899, + 0.007238379679620266, + 0.012581894174218178, + 0.007584072649478912, + 0.012117797508835793, + 0.011216271668672562, + 0.007824922911822796, + 0.006008140277117491, + 0.010272668674588203, + 0.007300897501409054, + 0.006514598149806261, + 0.010293404571712017, + 0.010608081705868244, + 0.015552947297692299, + 0.00882576685398817, + 0.006348665338009596, + 0.007540808524936438, + 0.0074507370591163635, + 0.003510607872158289, + 0.004251526203006506, + 0.013941827230155468, + 0.009373463690280914, + 0.005469750612974167, + 0.005134908016771078, + 0.006517563480883837, + 0.0033290423452854156, + 0.0057313647121191025, + 0.007363028358668089, + 0.026318565011024475, + 0.009662221185863018, + 0.030497590079903603, + 0.00909354630857706, + 0.0071251410990953445, + 0.004828157834708691, + 0.004765699151903391, + 0.004030849784612656, + 0.007306764367967844, + 0.006964624859392643, + 0.009161037392914295, + 0.012786873616278172, + 0.005648464895784855, + 0.007797309197485447, + 0.016114801168441772, + 0.01836324855685234, + 0.017176300287246704, + 0.02004893869161606, + 0.012072019279003143, + 0.026376500725746155, + 0.008011071011424065, + 0.01269651111215353, + 0.016791844740509987, + 0.0192982479929924, + 0.01815812475979328, + 0.022920189425349236, + 0.02625049091875553, + 0.021385181695222855, + 0.012443138286471367, + 0.012744521722197533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.005267906002700329 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868850111961365 + ], + "opposite_logprob_tokens": [ + -3.0486884117126465 + ], + "logprob_reward": -0.04868850111961365, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868850111961365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06594999879598618, + 0.055144768208265305, + 0.06824979186058044, + 0.20765423774719238, + 0.0480135977268219, + 0.07706678658723831, + 0.11466983705759048, + 0.037991274148225784, + 0.0723317340016365, + 0.2529279887676239 + ], + "credit_map": [ + 0.00645909458398819, + 0.007782273460179567, + 0.013019362464547157, + 0.004556970205157995, + 0.005225927568972111, + 0.008224221877753735, + 0.004721248522400856, + 0.015960900112986565, + 0.0061805578880012035, + 0.013017598539590836, + 0.006673079449683428, + 0.006498800590634346, + 0.005843968130648136, + 0.0059311226941645145, + 0.0069139087572693825, + 0.004085732623934746, + 0.005751718766987324, + 0.007162438239902258, + 0.010579526424407959, + 0.006503575015813112, + 0.004733615089207888, + 0.014967173337936401, + 0.006515889894217253, + 0.012035859748721123, + 0.004952719900757074, + 0.006754698697477579, + 0.06239253655076027, + 0.07966616004705429, + 0.0063399821519851685, + 0.012334399856626987, + 0.021980952471494675, + 0.013232781551778316, + 0.003215852892026305, + 0.005197294522076845, + 0.005567981395870447, + 0.008118310943245888, + 0.01027197390794754, + 0.0036097527481615543, + 0.006946577690541744, + 0.005085852462798357, + 0.004758708644658327, + 0.006162320263683796, + 0.018174950033426285, + 0.013789897784590721, + 0.008863929659128189, + 0.006340516731142998, + 0.013407238759100437, + 0.005569226108491421, + 0.006855469662696123, + 0.005826557520776987, + 0.03517219424247742, + 0.010476721450686455, + 0.039061039686203, + 0.00700766546651721, + 0.006554262712597847, + 0.003715936793014407, + 0.003618980059400201, + 0.0031276994850486517, + 0.006286217365413904, + 0.0052448841743171215, + 0.004528258927166462, + 0.004742044024169445, + 0.005133194848895073, + 0.005309993401169777, + 0.007863091304898262, + 0.009719448164105415, + 0.010589232668280602, + 0.010706338100135326, + 0.009485051035881042, + 0.010037877596914768, + 0.0056789037771523, + 0.008251789957284927, + 0.020438171923160553, + 0.025721613317728043, + 0.03366425260901451, + 0.03789620101451874, + 0.048715297132730484, + 0.04811728745698929, + 0.01810302399098873, + 0.020272146910429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.04868850111961365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368487626314163 + ], + "opposite_logprob_tokens": [ + -3.6513683795928955 + ], + "logprob_reward": -0.026368487626314163, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368487626314163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1097036600112915, + 0.10256285220384598, + 0.05164780095219612, + 0.04484904929995537, + 0.057087358087301254, + 0.04633377119898796, + 0.16663673520088196, + 0.05887609347701073, + 0.14324083924293518, + 0.21906182169914246 + ], + "credit_map": [ + 0.010564005933701992, + 0.009788756258785725, + 0.01731009967625141, + 0.008870931342244148, + 0.00690692150965333, + 0.01581772416830063, + 0.009059964679181576, + 0.03138526529073715, + 0.014178997837007046, + 0.01937216706573963, + 0.012901319190859795, + 0.012092849239706993, + 0.009385581128299236, + 0.011007318273186684, + 0.01640915684401989, + 0.007215468678623438, + 0.006167786195874214, + 0.0048064421862363815, + 0.007462901528924704, + 0.010407469235360622, + 0.004404827021062374, + 0.006143543403595686, + 0.006802403833717108, + 0.005452430807054043, + 0.003601578064262867, + 0.004492337349802256, + 0.011835734359920025, + 0.009975084103643894, + 0.004002375062555075, + 0.0022117297630757093, + 0.0063844905234873295, + 0.002345724729821086, + 0.0029535952489823103, + 0.006284045521169901, + 0.007963592186570168, + 0.01270995568484068, + 0.007548775058239698, + 0.004556136671453714, + 0.009581443853676319, + 0.005489816889166832, + 0.003602354321628809, + 0.0033192189875990152, + 0.006067963782697916, + 0.006051560398191214, + 0.006405620835721493, + 0.006863867864012718, + 0.008584804832935333, + 0.005438385531306267, + 0.008039914071559906, + 0.0060507371090352535, + 0.059477318078279495, + 0.014450650662183762, + 0.05889158323407173, + 0.00940430723130703, + 0.006643800530582666, + 0.0036784412804991007, + 0.0038752416148781776, + 0.005286098923534155, + 0.009032954461872578, + 0.008389914408326149, + 0.007157628890126944, + 0.01352014858275652, + 0.005727759096771479, + 0.0058863526210188866, + 0.01602957956492901, + 0.020109346136450768, + 0.02109123021364212, + 0.02394200675189495, + 0.01476610079407692, + 0.02470524050295353, + 0.010366288013756275, + 0.012231050059199333, + 0.026731278747320175, + 0.031205274164676666, + 0.027422897517681122, + 0.031913649290800095, + 0.035568010061979294, + 0.02788039669394493, + 0.020154541358351707, + 0.018185772001743317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.026368487626314163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic close-up tracking shot, set amidst the hum of machinery in a 1950s Chinese automobile factory, follows the tires of a massive, heavy transport truck as they roll at high speed. The deep treads, etched with dirt and gravel, leave a trail of rapid movement across the rough concrete floor. Suddenly, with a sharp screeching of brakes, the camera screeches to a halt, locking focus precisely on the stationary tire's sidewall. At that moment, a worker's arm, calloused and thick-knuckled, emerges from the rolled-up sleeves of his blue work clothes. Steadily and firmly grasping a stub of white chalk, he decisively scratches a clear registration mark on the rough rubber surface. At the same time, the wood chips and dust that flew up in the air due to the sudden stop of the vehicle seemed to be frozen in time, suspended in the air by the volumetric oblique beams cast from the high windows of the workshop. Every tiny particle was clearly visible, and its irregular shape formed an instant sculpture full of industrial power in the still picture. Under the harsh contrast of light and shadow, the whole scene presents an atmosphere of industrial aesthetics that is focused, powerful and full of the imprint of the times.", + "seed": 21341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_5/index_5659_seed_21341_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20544607304036616, + "vlm_logprob": -0.20544607304036616, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.20544607304036616, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2041289508342743 + ], + "opposite_logprob_tokens": [ + -1.8291289806365967 + ], + "logprob_reward": -0.2041289508342743, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.2041289508342743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32144689559936523, + 0.07507480680942535, + 0.0889085978269577, + 0.0768112763762474, + 0.04420478641986847, + 0.07053034752607346, + 0.087908074259758, + 0.06068543717265129, + 0.11665249615907669, + 0.05777725204825401 + ], + "credit_map": [ + 0.04304613918066025, + 0.02218249812722206, + 0.04892825335264206, + 0.03276805207133293, + 0.03873691335320473, + 0.05153157562017441, + 0.05662466958165169, + 0.027628783136606216, + 0.007489569019526243, + 0.0071073174476623535, + 0.011321586556732655, + 0.0122032156214118, + 0.008311362937092781, + 0.006540575064718723, + 0.013326211832463741, + 0.008774968795478344, + 0.009345203638076782, + 0.011106256395578384, + 0.010980951599776745, + 0.014489872381091118, + 0.009364612400531769, + 0.008409938775002956, + 0.017228079959750175, + 0.007983689196407795, + 0.015735629945993423, + 0.009224284440279007, + 0.0060873739421367645, + 0.012380236759781837, + 0.008336660452187061, + 0.0043030851520597935, + 0.01740199327468872, + 0.0033420168329030275, + 0.0031116006430238485, + 0.011830330826342106, + 0.004219523631036282, + 0.007790127769112587, + 0.005777441896498203, + 0.0022830527741461992, + 0.006301660556346178, + 0.00289104669354856, + 0.006461484357714653, + 0.009782892651855946, + 0.009086500853300095, + 0.009624844416975975, + 0.011944189667701721, + 0.00739456620067358, + 0.011859453283250332, + 0.004376411437988281, + 0.011989317834377289, + 0.015170013532042503, + 0.014185081236064434, + 0.007850290276110172, + 0.007028217893093824, + 0.011353444308042526, + 0.007705499418079853, + 0.012626204639673233, + 0.0074158222414553165, + 0.005642844829708338, + 0.012410392984747887, + 0.008279891684651375, + 0.0033969173673540354, + 0.00849913153797388, + 0.007398731075227261, + 0.0076417080126702785, + 0.007658893242478371, + 0.007448600139468908, + 0.03897089511156082, + 0.007919544354081154, + 0.0045704771764576435, + 0.009772054851055145, + 0.007959704846143723, + 0.032352328300476074, + 0.0062347352504730225, + 0.006972678005695343, + 0.012563815340399742, + 0.006357742473483086, + 0.005776860285550356, + 0.005182686727494001, + 0.007901233620941639, + 0.006787507329136133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.2041289508342743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7579030394554138 + ], + "opposite_logprob_tokens": [ + -0.6329030394554138 + ], + "logprob_reward": -0.7579030394554138, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7579030394554138, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3926461935043335, + 0.0615873858332634, + 0.03625134006142616, + 0.046493273228406906, + 0.0774359256029129, + 0.056085631251335144, + 0.05339816212654114, + 0.11145440489053726, + 0.11234978586435318, + 0.05229790881276131 + ], + "credit_map": [ + 0.05301244556903839, + 0.027513742446899414, + 0.054653577506542206, + 0.03911790996789932, + 0.05863958224654198, + 0.057793937623500824, + 0.06907977163791656, + 0.0328352265059948, + 0.005847004242241383, + 0.005553379189223051, + 0.009165755473077297, + 0.00848450232297182, + 0.006581108085811138, + 0.006901919841766357, + 0.010329599492251873, + 0.00872411672025919, + 0.0033352237660437822, + 0.004262735601514578, + 0.0033754673786461353, + 0.004654075484722853, + 0.0037684915587306023, + 0.006175430491566658, + 0.006135139614343643, + 0.0045447759330272675, + 0.005378586705774069, + 0.004106403794139624, + 0.004329727031290531, + 0.007316099479794502, + 0.006707485765218735, + 0.003955812193453312, + 0.011800423264503479, + 0.002898735925555229, + 0.004327164031565189, + 0.017268817871809006, + 0.0090261809527874, + 0.015439766459167004, + 0.011163394898176193, + 0.0054864236153662205, + 0.010912779718637466, + 0.0038113996852189302, + 0.005424260627478361, + 0.008515418507158756, + 0.005936244037002325, + 0.008354661986231804, + 0.010395873337984085, + 0.005748831667006016, + 0.007444801740348339, + 0.004265539813786745, + 0.006217246875166893, + 0.00816018134355545, + 0.011246958747506142, + 0.006814820226281881, + 0.004782096482813358, + 0.005163086112588644, + 0.005471207667142153, + 0.0055425651371479034, + 0.015717819333076477, + 0.017491260543465614, + 0.015175487846136093, + 0.009573680348694324, + 0.006890889722853899, + 0.02080496959388256, + 0.012096372433006763, + 0.013703921809792519, + 0.005885281600058079, + 0.006077930796891451, + 0.038473717868328094, + 0.007972516119480133, + 0.004963994026184082, + 0.009476212784647942, + 0.008356002159416676, + 0.03114412911236286, + 0.003446257207542658, + 0.0049435896798968315, + 0.01345793902873993, + 0.006641796790063381, + 0.004020845051854849, + 0.004018748179078102, + 0.005933067761361599, + 0.009835666976869106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.7579030394554138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061430551111698 + ], + "opposite_logprob_tokens": [ + -3.8956143856048584 + ], + "logprob_reward": -0.02061430551111698, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061430551111698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.265436589717865, + 0.031444087624549866, + 0.04961492121219635, + 0.04348010569810867, + 0.06821289658546448, + 0.0953092873096466, + 0.15594394505023956, + 0.0928557962179184, + 0.1727915108203888, + 0.02491084113717079 + ], + "credit_map": [ + 0.03637222200632095, + 0.019352905452251434, + 0.03662356734275818, + 0.026284581050276756, + 0.04010194167494774, + 0.0389496274292469, + 0.046050943434238434, + 0.021700793877243996, + 0.0037344875745475292, + 0.0025883973576128483, + 0.005080848466604948, + 0.004043182823807001, + 0.0026555638760328293, + 0.004083115141838789, + 0.004584477748721838, + 0.004674013704061508, + 0.004055927507579327, + 0.0049187978729605675, + 0.003553214017301798, + 0.008724682964384556, + 0.0037614426109939814, + 0.01035601831972599, + 0.007235220167785883, + 0.007009617984294891, + 0.007243727799504995, + 0.004788239020854235, + 0.004777830559760332, + 0.005655008368194103, + 0.007566153071820736, + 0.0034455570857971907, + 0.007229658775031567, + 0.0027739310171455145, + 0.0038776835426688194, + 0.019607773050665855, + 0.009774898178875446, + 0.011655822396278381, + 0.006036052480340004, + 0.004698734264820814, + 0.00974227488040924, + 0.0028196610510349274, + 0.009951276704668999, + 0.015078013762831688, + 0.009028474800288677, + 0.017477037385106087, + 0.014312786050140858, + 0.00990450568497181, + 0.01161986868828535, + 0.00793732050806284, + 0.025077976286411285, + 0.019420389086008072, + 0.020631587132811546, + 0.013832374475896358, + 0.012136510573327541, + 0.019847968593239784, + 0.017177822068333626, + 0.027819311246275902, + 0.012323630973696709, + 0.017827223986387253, + 0.013899854384362698, + 0.012247603386640549, + 0.006922557950019836, + 0.01075264997780323, + 0.010050737299025059, + 0.008831537328660488, + 0.008886364288628101, + 0.009078556671738625, + 0.06342258304357529, + 0.010533202439546585, + 0.005305931903421879, + 0.014039859175682068, + 0.008841566741466522, + 0.052683450281620026, + 0.002628420479595661, + 0.003899232717230916, + 0.0037827978376299143, + 0.003302098484709859, + 0.0022875131107866764, + 0.0026633506640791893, + 0.0032590685877949, + 0.003088359022513032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.02061430551111698 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209312111139297 + ], + "opposite_logprob_tokens": [ + -4.018209457397461 + ], + "logprob_reward": -0.018209312111139297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209312111139297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09925071150064468, + 0.0489489883184433, + 0.06465425342321396, + 0.07679474353790283, + 0.07109429687261581, + 0.10676686465740204, + 0.08878589421510696, + 0.11844828724861145, + 0.22240695357322693, + 0.10284900665283203 + ], + "credit_map": [ + 0.012273567728698254, + 0.00726323900744319, + 0.01356884092092514, + 0.013038171455264091, + 0.012064561247825623, + 0.013812566176056862, + 0.018001575022935867, + 0.009228191338479519, + 0.004505355842411518, + 0.0035671074874699116, + 0.0067788842134177685, + 0.00802661757916212, + 0.005529687739908695, + 0.00560368224978447, + 0.006640680134296417, + 0.008296972140669823, + 0.0083278464153409, + 0.009725680574774742, + 0.006677869241684675, + 0.009339220821857452, + 0.005874428898096085, + 0.007701200433075428, + 0.009458993561565876, + 0.007549015339463949, + 0.015611776150763035, + 0.00576971797272563, + 0.009501690976321697, + 0.014393406920135021, + 0.007129868026822805, + 0.004018782638013363, + 0.01723834127187729, + 0.0031311556231230497, + 0.006455984897911549, + 0.02058863453567028, + 0.007880331948399544, + 0.011011062189936638, + 0.007733810693025589, + 0.0037658722139894962, + 0.009901493787765503, + 0.003757105441763997, + 0.008797216229140759, + 0.011428175494074821, + 0.010958326980471611, + 0.013743876479566097, + 0.020023010671138763, + 0.012325526215136051, + 0.02023410238325596, + 0.00925662275403738, + 0.010290293954312801, + 0.01002874132245779, + 0.012112568132579327, + 0.011447766795754433, + 0.009266849607229233, + 0.00991032738238573, + 0.008248834870755672, + 0.017480505630373955, + 0.012415586039423943, + 0.01724494993686676, + 0.018929990008473396, + 0.014704883098602295, + 0.008104310370981693, + 0.014161313883960247, + 0.018080420792102814, + 0.014806837774813175, + 0.02257635071873665, + 0.030270403251051903, + 0.042608268558979034, + 0.022187137976288795, + 0.017072660848498344, + 0.024115681648254395, + 0.025351732969284058, + 0.0382247157394886, + 0.01023037824779749, + 0.01330967340618372, + 0.02208890952169895, + 0.016302209347486496, + 0.0069984085857868195, + 0.007933443412184715, + 0.01348834577947855, + 0.01249763835221529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.018209312111139297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026374757289886475 + ], + "opposite_logprob_tokens": [ + -3.6513748168945312 + ], + "logprob_reward": -0.026374757289886475, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026374757289886475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17853620648384094, + 0.05112176015973091, + 0.03986001014709473, + 0.042159222066402435, + 0.05034293979406357, + 0.05358267202973366, + 0.1671476662158966, + 0.13311852514743805, + 0.18138529360294342, + 0.10274568945169449 + ], + "credit_map": [ + 0.025478458032011986, + 0.015780581161379814, + 0.020156463608145714, + 0.0204282496124506, + 0.02827627770602703, + 0.02575686201453209, + 0.024354666471481323, + 0.018304644152522087, + 0.005478276871144772, + 0.006830879487097263, + 0.006685542408376932, + 0.007752002216875553, + 0.005767105147242546, + 0.003983486909419298, + 0.008580329827964306, + 0.006044136825948954, + 0.003878643736243248, + 0.005510785151273012, + 0.0032885675318539143, + 0.0037188788410276175, + 0.002898546401411295, + 0.008517730981111526, + 0.004981708247214556, + 0.0070651532150805, + 0.0068759857676923275, + 0.0038060087244957685, + 0.005279941018670797, + 0.006534842308610678, + 0.005165840499103069, + 0.004165074788033962, + 0.0074122995138168335, + 0.0029192285146564245, + 0.004362031351774931, + 0.007551745977252722, + 0.006144437938928604, + 0.010620341636240482, + 0.008538613095879555, + 0.0032664472237229347, + 0.0074005224741995335, + 0.0024588010273873806, + 0.007007695268839598, + 0.009993716143071651, + 0.005824033636599779, + 0.008396754041314125, + 0.007238453254103661, + 0.005514620803296566, + 0.005291286390274763, + 0.004316109232604504, + 0.020326806232333183, + 0.034600626677274704, + 0.02599824033677578, + 0.017620671540498734, + 0.01602010987699032, + 0.018085289746522903, + 0.016806529834866524, + 0.01768939197063446, + 0.010378318838775158, + 0.012870007194578648, + 0.019759558141231537, + 0.0183394867926836, + 0.010080331936478615, + 0.017687266692519188, + 0.027242017909884453, + 0.016761543229222298, + 0.015834294259548187, + 0.019446074962615967, + 0.0441056452691555, + 0.02080148085951805, + 0.009350365959107876, + 0.018532343208789825, + 0.017758963629603386, + 0.035556115210056305, + 0.009133955463767052, + 0.009502693079411983, + 0.0227800402790308, + 0.01611640676856041, + 0.009140455164015293, + 0.0077562653459608555, + 0.014368383213877678, + 0.013947482220828533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.026374757289886475 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic close-up tracking shot, set amidst the hum of machinery in a 1950s Chinese automobile factory, follows the tires of a massive, heavy transport truck as they roll at high speed. The deep treads, etched with dirt and gravel, leave a trail of rapid movement across the rough concrete floor. Suddenly, with a sharp screeching of brakes, the camera screeches to a halt, locking focus precisely on the stationary tire's sidewall. At that moment, a worker's arm, calloused and thick-knuckled, emerges from the rolled-up sleeves of his blue work clothes. Steadily and firmly grasping a stub of white chalk, he decisively scratches a clear registration mark on the rough rubber surface. At the same time, the wood chips and dust that flew up in the air due to the sudden stop of the vehicle seemed to be frozen in time, suspended in the air by the volumetric oblique beams cast from the high windows of the workshop. Every tiny particle was clearly visible, and its irregular shape formed an instant sculpture full of industrial power in the still picture. Under the harsh contrast of light and shadow, the whole scene presents an atmosphere of industrial aesthetics that is focused, powerful and full of the imprint of the times.", + "seed": 21342, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_5/index_5659_seed_21342_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.186997514590621, + "vlm_logprob": -0.186997514590621, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.186997514590621, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4509306252002716 + ], + "opposite_logprob_tokens": [ + -1.0759305953979492 + ], + "logprob_reward": -0.4509306252002716, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4509306252002716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1587173193693161, + 0.26887810230255127, + 0.27488481998443604, + 0.0407792404294014, + 0.07486559450626373, + 0.019119461998343468, + 0.045233141630887985, + 0.03360822796821594, + 0.06580081582069397, + 0.01811320334672928 + ], + "credit_map": [ + 0.008199258707463741, + 0.010027131997048855, + 0.05070880800485611, + 0.009488411247730255, + 0.008570383302867413, + 0.013293355703353882, + 0.010474495589733124, + 0.04795549809932709, + 0.014064648188650608, + 0.010099643841385841, + 0.10599271208047867, + 0.025042105466127396, + 0.00894791167229414, + 0.0053724683821201324, + 0.08778668940067291, + 0.011571950279176235, + 0.011751649901270866, + 0.1116073802113533, + 0.07671646028757095, + 0.02520177885890007, + 0.016203999519348145, + 0.015067258849740028, + 0.012913552112877369, + 0.0054227933287620544, + 0.0031586457043886185, + 0.003326546400785446, + 0.004999996162950993, + 0.007171242497861385, + 0.005326302256435156, + 0.005610880441963673, + 0.007150780409574509, + 0.004034852609038353, + 0.007031829096376896, + 0.007397353649139404, + 0.00844577420502901, + 0.010670749470591545, + 0.010820168070495129, + 0.010897846892476082, + 0.009676032699644566, + 0.009925855323672295, + 0.0018881452269852161, + 0.0011067275190725923, + 0.0017617808189243078, + 0.001934162573888898, + 0.0023372184950858355, + 0.0018635467858985066, + 0.005772428587079048, + 0.0024554559495300055, + 0.002835448831319809, + 0.0030039059929549694, + 0.0030085970647633076, + 0.0032775935251265764, + 0.0065115815959870815, + 0.004232359584420919, + 0.01608492247760296, + 0.006278736516833305, + 0.0030537426937371492, + 0.0037504774518311024, + 0.003330020233988762, + 0.005983590614050627, + 0.003522447543218732, + 0.0036646537482738495, + 0.006787767168134451, + 0.003515531774610281, + 0.0029705266933888197, + 0.003394212806597352, + 0.0029196373652666807, + 0.004934518598020077, + 0.022184688597917557, + 0.010498527437448502, + 0.004176164977252483, + 0.014722548425197601, + 0.001427890150807798, + 0.0013567854184657335, + 0.0029577682726085186, + 0.0029496997594833374, + 0.0016948276897892356, + 0.0017775215674191713, + 0.003081592731177807, + 0.0028671210166066885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.4509306252002716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38716593384742737 + ], + "opposite_logprob_tokens": [ + -1.137165904045105 + ], + "logprob_reward": -0.38716593384742737, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38716593384742737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19337977468967438, + 0.22279025614261627, + 0.1948925256729126, + 0.051804132759571075, + 0.0883123129606247, + 0.042159661650657654, + 0.05259302631020546, + 0.04257269948720932, + 0.05760103464126587, + 0.05389450117945671 + ], + "credit_map": [ + 0.016563598066568375, + 0.025006674230098724, + 0.019785074517130852, + 0.024368904531002045, + 0.019614076241850853, + 0.0277011226862669, + 0.03613857179880142, + 0.0242017712444067, + 0.01117311604321003, + 0.009042775258421898, + 0.08739757537841797, + 0.02054545283317566, + 0.00773896649479866, + 0.004967262037098408, + 0.07223597913980484, + 0.009689136408269405, + 0.008411751128733158, + 0.07057175040245056, + 0.0590723380446434, + 0.019052835181355476, + 0.011788103729486465, + 0.008794749155640602, + 0.011119906790554523, + 0.0060811229050159454, + 0.005484787747263908, + 0.0047940704971551895, + 0.00604116590693593, + 0.007270941976457834, + 0.006264207419008017, + 0.008690697140991688, + 0.007254071533679962, + 0.00600419333204627, + 0.009236960671842098, + 0.008908571675419807, + 0.009363200515508652, + 0.010963108390569687, + 0.012999499216675758, + 0.011804741807281971, + 0.012442969717085361, + 0.012593272142112255, + 0.0020141303539276123, + 0.0025740996934473515, + 0.003312227316200733, + 0.009226981550455093, + 0.005536052864044905, + 0.004800018854439259, + 0.0076103126630187035, + 0.007085841614753008, + 0.0035379936452955008, + 0.005098276771605015, + 0.0036923012230545282, + 0.005321004427969456, + 0.006326025817543268, + 0.006775482092052698, + 0.012706990353763103, + 0.009134955704212189, + 0.00341406580992043, + 0.004078071564435959, + 0.0042079780250787735, + 0.009075358510017395, + 0.003918753936886787, + 0.004844730254262686, + 0.00892691407352686, + 0.004106831271201372, + 0.0018709488213062286, + 0.00192624656483531, + 0.00203514052554965, + 0.004208733793348074, + 0.02115710824728012, + 0.009959634393453598, + 0.0037668978329747915, + 0.012676333077251911, + 0.007634949870407581, + 0.006121844053268433, + 0.008184311911463737, + 0.006621283013373613, + 0.006001263856887817, + 0.007401935290545225, + 0.0065469578839838505, + 0.0053819576278328896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.38716593384742737 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020617807283997536 + ], + "opposite_logprob_tokens": [ + -3.895617723464966 + ], + "logprob_reward": -0.020617807283997536, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020617807283997536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15868815779685974, + 0.19665369391441345, + 0.20575004816055298, + 0.02605486661195755, + 0.15224848687648773, + 0.03302392736077309, + 0.07206038385629654, + 0.04211330786347389, + 0.07504712045192719, + 0.03835997357964516 + ], + "credit_map": [ + 0.012917840853333473, + 0.018849845975637436, + 0.022718729451298714, + 0.019220909103751183, + 0.01376368198543787, + 0.024554824456572533, + 0.022590462118387222, + 0.024071861058473587, + 0.009898994117975235, + 0.00797104462981224, + 0.07584797590970993, + 0.015934115275740623, + 0.007574198301881552, + 0.004776417277753353, + 0.06633113324642181, + 0.008319810964167118, + 0.008660412393510342, + 0.08062145113945007, + 0.056469112634658813, + 0.019408630207180977, + 0.012432900257408619, + 0.011421076953411102, + 0.012184444814920425, + 0.0045520225539803505, + 0.0023843736853450537, + 0.0026982545386999846, + 0.0034730802290141582, + 0.0036068938206881285, + 0.004026708658784628, + 0.0039534252136945724, + 0.0039589921943843365, + 0.0019531380385160446, + 0.012735016644001007, + 0.014909494668245316, + 0.01671360246837139, + 0.02126678079366684, + 0.023051802068948746, + 0.021657196804881096, + 0.0196089968085289, + 0.022305602207779884, + 0.0019277754472568631, + 0.0018521343590691686, + 0.002486490411683917, + 0.005290077533572912, + 0.003782605053856969, + 0.003486976260319352, + 0.0098996851593256, + 0.004298183135688305, + 0.004963862244039774, + 0.008609481155872345, + 0.005367886275053024, + 0.007766198832541704, + 0.009104848839342594, + 0.011327531188726425, + 0.010765059851109982, + 0.01415551733225584, + 0.002862590830773115, + 0.003875170135870576, + 0.0036063517909497023, + 0.011151105165481567, + 0.0036537358537316322, + 0.004134886432439089, + 0.006655690260231495, + 0.006173776462674141, + 0.003971039317548275, + 0.0037241936661303043, + 0.003585048718377948, + 0.005146395415067673, + 0.025714129209518433, + 0.012680230662226677, + 0.0038815930020064116, + 0.01634448580443859, + 0.005938325077295303, + 0.005546571221202612, + 0.005993810947984457, + 0.004983175545930862, + 0.0035477608907967806, + 0.0034119561314582825, + 0.004457580391317606, + 0.0044807931408286095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.020617807283997536 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813677653670311 + ], + "opposite_logprob_tokens": [ + -3.2881367206573486 + ], + "logprob_reward": -0.03813677653670311, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813677653670311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14439791440963745, + 0.21595384180545807, + 0.3181065618991852, + 0.029751872643828392, + 0.02820468880236149, + 0.01813478395342827, + 0.0645938366651535, + 0.038159117102622986, + 0.09620428830385208, + 0.0464930385351181 + ], + "credit_map": [ + 0.00726527813822031, + 0.011746135540306568, + 0.03860919550061226, + 0.01273895800113678, + 0.010545270517468452, + 0.013110287487506866, + 0.014813580550253391, + 0.0355691984295845, + 0.011136302724480629, + 0.008554341271519661, + 0.08413796126842499, + 0.018598804250359535, + 0.008031977340579033, + 0.004988125991076231, + 0.07132674753665924, + 0.009179585613310337, + 0.013729540631175041, + 0.1299024522304535, + 0.08939211070537567, + 0.028137078508734703, + 0.018998943269252777, + 0.016992740333080292, + 0.014706184156239033, + 0.006247531156986952, + 0.0025554571766406298, + 0.003152459627017379, + 0.002793149324133992, + 0.005704259965568781, + 0.004478237126022577, + 0.0029499498195946217, + 0.005439672619104385, + 0.0026786867529153824, + 0.0029342856723815203, + 0.0030748392455279827, + 0.002943300409242511, + 0.0038198677357286215, + 0.003869876265525818, + 0.004274760838598013, + 0.0035905083641409874, + 0.003697252133861184, + 0.0017489700112491846, + 0.001085215131752193, + 0.0022669255267828703, + 0.002026910660788417, + 0.0018646613461896777, + 0.002577038947492838, + 0.0026980736292898655, + 0.0038669886998832226, + 0.007691650651395321, + 0.0041325148195028305, + 0.010259021073579788, + 0.005985932890325785, + 0.00845546554774046, + 0.003717507468536496, + 0.015650032088160515, + 0.00870171282440424, + 0.0033083143644034863, + 0.004290597513318062, + 0.004057152196764946, + 0.009905862621963024, + 0.003665459807962179, + 0.0029366272501647472, + 0.0035231802612543106, + 0.006471922155469656, + 0.005159848369657993, + 0.005645022727549076, + 0.005491749849170446, + 0.007325264159590006, + 0.029116127640008926, + 0.01309257186949253, + 0.006894221529364586, + 0.023479485884308815, + 0.007623132783919573, + 0.007690433878451586, + 0.0066884784027934074, + 0.005716398824006319, + 0.004322934430092573, + 0.0036562324967235327, + 0.0053022862412035465, + 0.005493141245096922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03813677653670311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a 1950s Chinese automobile factory with visible machinery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a massive transport truck with deep, dirty treads rolling at high speed on a rough concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track the truck's tires at high speed before suddenly stopping to focus on a stationary tire's sidewall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a worker with calloused arms wearing blue work clothes seen scratching a registration mark on the tire with white chalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and dust particles frozen in mid-air, illuminated by volumetric oblique beams from high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813643008470535 + ], + "opposite_logprob_tokens": [ + -3.2881364822387695 + ], + "logprob_reward": -0.03813643008470535, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813643008470535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08158992975950241, + 0.2347862422466278, + 0.16824299097061157, + 0.03194276615977287, + 0.10076677054166794, + 0.03284763544797897, + 0.035887908190488815, + 0.025390835478901863, + 0.031039992347359657, + 0.2575049102306366 + ], + "credit_map": [ + 0.00648530526086688, + 0.0076475744135677814, + 0.012730985879898071, + 0.009859074838459492, + 0.008799290284514427, + 0.010626976378262043, + 0.013126521371304989, + 0.012314202263951302, + 0.01195688545703888, + 0.00891063455492258, + 0.09167593717575073, + 0.019571032375097275, + 0.00836016982793808, + 0.004876323975622654, + 0.07936529070138931, + 0.010069993324577808, + 0.0075736637227237225, + 0.06478946655988693, + 0.03837062045931816, + 0.019053760915994644, + 0.009988333098590374, + 0.011093629524111748, + 0.012424330227077007, + 0.004949194844812155, + 0.0024460104759782553, + 0.0038010398857295513, + 0.0033428494352847338, + 0.005866485647857189, + 0.003982401452958584, + 0.0041463361121714115, + 0.005591775290668011, + 0.002765871351584792, + 0.009379375725984573, + 0.010066134855151176, + 0.010830443352460861, + 0.013914201408624649, + 0.015092020854353905, + 0.014529979787766933, + 0.0132159274071455, + 0.013738696463406086, + 0.003376025939360261, + 0.0018342797411605716, + 0.004165958147495985, + 0.003994428552687168, + 0.0027071021031588316, + 0.0047348360531032085, + 0.003591592190787196, + 0.00844341516494751, + 0.0038900033105164766, + 0.003861689008772373, + 0.0058316620998084545, + 0.004698926582932472, + 0.0044828299432992935, + 0.0035987726878374815, + 0.005753063131123781, + 0.003770963754504919, + 0.0036068365443497896, + 0.003528895555064082, + 0.003197264624759555, + 0.0030053991358727217, + 0.0029889768920838833, + 0.002919678809121251, + 0.003623739117756486, + 0.0025200475938618183, + 0.0020925358403474092, + 0.0017159609124064445, + 0.0017985565355047584, + 0.002609429182484746, + 0.007937577553093433, + 0.004517093766480684, + 0.0038485045079141855, + 0.006520337890833616, + 0.03295273706316948, + 0.03211861848831177, + 0.034458041191101074, + 0.034416262060403824, + 0.03132052347064018, + 0.027509044855833054, + 0.03250512108206749, + 0.032224588096141815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.03813643008470535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_tensor([4610, 4610])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_tensor([4610, 4610])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d17b76c3284fb67a734fcbe39f0a335678cf5942 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_tensor([4610, 4610])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot follows a female athlete as she desperately stumbles along a reddish-brown all-weather track under the blazing midday sun. Her dark brown hair is pulled back into a tight ponytail, but loose, sweat-soaked strands cling to her temples and forehead. Her face is a mask of pure agony; her jaw clenched, her eyes squeezed shut, her features contorted. She wears a dark blue, sweat-soaked running tank top and matching shorts. Her knees ache and blood, angry red scrapes weeping and caked with the grit of the track's surface. Her forward momentum is uncontrolled, a series of unsteady strides, her legs trembling with fatigue, before finally giving way completely. Her knees buckle, and she collapses hard, her torso hitting the track first in a boneless heap, her arms unable to break the fall. As she plummets, the camera tilts sharply downward, capturing the final moments of impact, creating a raw and visceral atmosphere of utter physical exhaustion and agonizing defeat.", + "seed": 9212, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_4610_seed_9212_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019766456540673973, + "vlm_logprob": -0.019766456540673973, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.019766456540673973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017184410244226456 + ], + "opposite_logprob_tokens": [ + -5.392184257507324 + ], + "logprob_reward": -0.017184410244226456, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.017184410244226456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14387035369873047, + 0.04215196892619133, + 0.029957221820950508, + 0.08684004098176956, + 0.09270022809505463, + 0.13457070291042328, + 0.10142875462770462, + 0.26689210534095764, + 0.05148729681968689, + 0.05010131746530533 + ], + "credit_map": [ + 0.012502885423600674, + 0.009814219549298286, + 0.03150161728262901, + 0.03156686946749687, + 0.01156722754240036, + 0.0239871758967638, + 0.008277017623186111, + 0.014653338119387627, + 0.00467733945697546, + 0.005130159202963114, + 0.0050069247372448444, + 0.004439518321305513, + 0.005182649474591017, + 0.007137737702578306, + 0.005481693893671036, + 0.005095947068184614, + 0.002207704819738865, + 0.003169824369251728, + 0.00316631724126637, + 0.0036025159060955048, + 0.005223636981099844, + 0.003567660925909877, + 0.0040608542039990425, + 0.004958708304911852, + 0.006361918989568949, + 0.008504913188517094, + 0.013103322125971317, + 0.017013080418109894, + 0.00876845046877861, + 0.008105473592877388, + 0.01333671249449253, + 0.011646165512502193, + 0.007138182874768972, + 0.0065505895763635635, + 0.015657713636755943, + 0.012543860822916031, + 0.00989957433193922, + 0.007137722335755825, + 0.01689440943300724, + 0.016878174617886543, + 0.016900023445487022, + 0.014177217148244381, + 0.01640700362622738, + 0.011052915826439857, + 0.010557065717875957, + 0.02763725258409977, + 0.009848077781498432, + 0.02799113653600216, + 0.018205583095550537, + 0.00819074735045433, + 0.008102644234895706, + 0.010527294129133224, + 0.011768276803195477, + 0.013693655841052532, + 0.01469852589070797, + 0.016242021694779396, + 0.03610232472419739, + 0.0178670696914196, + 0.04002196341753006, + 0.04564730077981949, + 0.031331948935985565, + 0.03926726430654526, + 0.024710407480597496, + 0.03194383904337883, + 0.007574391085654497, + 0.0033376351930201054, + 0.006792470812797546, + 0.005704670213162899, + 0.004540251102298498, + 0.0068933116272091866, + 0.008901707828044891, + 0.007742859423160553, + 0.004898815881460905, + 0.004309739917516708, + 0.008751983754336834, + 0.006147138774394989, + 0.005360397044569254, + 0.007360361982136965, + 0.006746077444404364, + 0.0065268054604530334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.017184410244226456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330542728304863 + ], + "opposite_logprob_tokens": [ + -3.7733054161071777 + ], + "logprob_reward": -0.02330542728304863, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330542728304863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08183994889259338, + 0.06841793656349182, + 0.07772234827280045, + 0.06945883482694626, + 0.050159379839897156, + 0.0716390609741211, + 0.10278788208961487, + 0.08057862520217896, + 0.195132315158844, + 0.20226363837718964 + ], + "credit_map": [ + 0.00942338164895773, + 0.010488331317901611, + 0.01021566428244114, + 0.009456844069063663, + 0.007169684860855341, + 0.012631974183022976, + 0.007842120714485645, + 0.014611950144171715, + 0.006305901799350977, + 0.009513354860246181, + 0.0061469110660254955, + 0.006032328587025404, + 0.008057629689574242, + 0.012697276659309864, + 0.009254788979887962, + 0.010409746319055557, + 0.00734034925699234, + 0.006900780834257603, + 0.007794119883328676, + 0.006557880900800228, + 0.009035284630954266, + 0.008758462965488434, + 0.017322571948170662, + 0.014012902975082397, + 0.005820585880428553, + 0.008555918000638485, + 0.008932869881391525, + 0.01087885070592165, + 0.008002368733286858, + 0.0086810989305377, + 0.008747858926653862, + 0.009839281439781189, + 0.003206138266250491, + 0.003535786410793662, + 0.00855710357427597, + 0.007082259748131037, + 0.005703120958060026, + 0.004798575770109892, + 0.008506003767251968, + 0.008770393207669258, + 0.01177111268043518, + 0.009545659646391869, + 0.008284910582005978, + 0.007335331290960312, + 0.0075641777366399765, + 0.01069545652717352, + 0.007382950745522976, + 0.009059462696313858, + 0.020803023129701614, + 0.009350907057523727, + 0.008340885862708092, + 0.011627637781202793, + 0.012078204192221165, + 0.013692478649318218, + 0.010824359022080898, + 0.01607038453221321, + 0.007154273800551891, + 0.00505552114918828, + 0.010283143259584904, + 0.016928184777498245, + 0.01280521135777235, + 0.011892042122781277, + 0.00926805566996336, + 0.00719219958409667, + 0.020127300173044205, + 0.018552497029304504, + 0.04010944440960884, + 0.029720185324549675, + 0.020283168181777, + 0.02856532856822014, + 0.02150651067495346, + 0.016267886385321617, + 0.017947372049093246, + 0.029336025938391685, + 0.028615456074476242, + 0.023946112021803856, + 0.032126862555742264, + 0.027025017887353897, + 0.025418685749173164, + 0.01784811168909073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.02330542728304863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043083954602479935 + ], + "opposite_logprob_tokens": [ + -3.168083906173706 + ], + "logprob_reward": -0.043083954602479935, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043083954602479935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15354853868484497, + 0.035610392689704895, + 0.026281801983714104, + 0.06935032457113266, + 0.06998766958713531, + 0.09925048798322678, + 0.11806151270866394, + 0.25416457653045654, + 0.07951309531927109, + 0.09423164278268814 + ], + "credit_map": [ + 0.014540552161633968, + 0.009865875355899334, + 0.04883161932229996, + 0.05364544689655304, + 0.005588489584624767, + 0.006693252827972174, + 0.005944882985204458, + 0.0084384148940444, + 0.004477748181670904, + 0.00616919482126832, + 0.003935968969017267, + 0.003537494456395507, + 0.0037611995358020067, + 0.004821763373911381, + 0.005151825491338968, + 0.00375519716180861, + 0.002588437171652913, + 0.0032703932374715805, + 0.0023621462751179934, + 0.002907901769503951, + 0.003470333758741617, + 0.003057156689465046, + 0.00447218120098114, + 0.0041532511822879314, + 0.0048289112746715546, + 0.005794999655336142, + 0.011391623876988888, + 0.014830822125077248, + 0.0068937442265450954, + 0.005480394698679447, + 0.00750129297375679, + 0.012628542259335518, + 0.003929151222109795, + 0.005010226275771856, + 0.010345248505473137, + 0.014386432245373726, + 0.005774745251983404, + 0.006180271971970797, + 0.013314732350409031, + 0.011046860367059708, + 0.013504245318472385, + 0.011204472742974758, + 0.015771476551890373, + 0.008481337688863277, + 0.010325996205210686, + 0.014443867839872837, + 0.0097059840336442, + 0.015813108533620834, + 0.022924557328224182, + 0.010082568973302841, + 0.007875831797719002, + 0.00963001511991024, + 0.014635029248893261, + 0.016966000199317932, + 0.013798977248370647, + 0.022148529067635536, + 0.038867782801389694, + 0.018183482810854912, + 0.03239400312304497, + 0.03437238186597824, + 0.03505861014127731, + 0.03755872696638107, + 0.02515762858092785, + 0.032571956515312195, + 0.008146179839968681, + 0.012370726093649864, + 0.009680243209004402, + 0.010223858058452606, + 0.00977201946079731, + 0.01077335700392723, + 0.009292008355259895, + 0.009254704229533672, + 0.009209845215082169, + 0.008203941397368908, + 0.014913873746991158, + 0.010377220809459686, + 0.013435102999210358, + 0.01412158738821745, + 0.011649985797703266, + 0.012320081703364849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.043083954602479935 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628239691257477 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628239691257477, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628239691257477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14819906651973724, + 0.03935915604233742, + 0.015337302349507809, + 0.038065105676651, + 0.07177697122097015, + 0.12366632372140884, + 0.17460522055625916, + 0.12644177675247192, + 0.19759728014469147, + 0.0649518370628357 + ], + "credit_map": [ + 0.009314368478953838, + 0.010083463974297047, + 0.0450478233397007, + 0.05415535718202591, + 0.006399048492312431, + 0.007966914214193821, + 0.006561538204550743, + 0.00867053959518671, + 0.004632624331861734, + 0.005672893021255732, + 0.004160805605351925, + 0.0038229916244745255, + 0.004900189582258463, + 0.0062064542435109615, + 0.005395433399826288, + 0.004567759111523628, + 0.002308413852006197, + 0.0014081632252782583, + 0.0012220351491123438, + 0.0014872974716126919, + 0.001462155836634338, + 0.0023424490354955196, + 0.0024080274160951376, + 0.002698759315535426, + 0.003028852166607976, + 0.003658746834844351, + 0.005805432330816984, + 0.007596044335514307, + 0.003514800686389208, + 0.003208504756912589, + 0.004380512051284313, + 0.006872208323329687, + 0.00467623770236969, + 0.003340959781780839, + 0.01413715910166502, + 0.011177661828696728, + 0.0047166962176561356, + 0.006761079654097557, + 0.016068648546934128, + 0.010898525826632977, + 0.012209348380565643, + 0.008598821237683296, + 0.009719346649944782, + 0.007709878962486982, + 0.006992811802774668, + 0.035369351506233215, + 0.007326957769691944, + 0.035739801824092865, + 0.031883250921964645, + 0.015517966821789742, + 0.01119213830679655, + 0.023204931989312172, + 0.01506246067583561, + 0.02512400783598423, + 0.018688250333070755, + 0.03393221274018288, + 0.016759509220719337, + 0.008983280509710312, + 0.014323820360004902, + 0.01538761705160141, + 0.019843092188239098, + 0.020472316071391106, + 0.01357080228626728, + 0.017101332545280457, + 0.02121439389884472, + 0.017046282067894936, + 0.03890872374176979, + 0.029053015634417534, + 0.021443074569106102, + 0.028361238539218903, + 0.023226963356137276, + 0.018343595787882805, + 0.008860748261213303, + 0.005697482731193304, + 0.01086499448865652, + 0.007739841006696224, + 0.008539318107068539, + 0.0099135497584939, + 0.007352626882493496, + 0.005983276292681694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.007628239691257477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630250882357359 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630250882357359, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630250882357359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18721535801887512, + 0.049739185720682144, + 0.028048034757375717, + 0.08496513217687607, + 0.0911361575126648, + 0.1119825541973114, + 0.09760751575231552, + 0.12248330563306808, + 0.15462172031402588, + 0.07220105081796646 + ], + "credit_map": [ + 0.012481513433158398, + 0.010158829391002655, + 0.06152404099702835, + 0.071625255048275, + 0.005833649076521397, + 0.00853111781179905, + 0.006427265238016844, + 0.010633678175508976, + 0.004525164607912302, + 0.004785772413015366, + 0.005253938492387533, + 0.0040194024331867695, + 0.00612475723028183, + 0.008138301782310009, + 0.008718268945813179, + 0.008173573762178421, + 0.004254767671227455, + 0.004699012264609337, + 0.0026366340462118387, + 0.002209738129749894, + 0.0032592746429145336, + 0.004602289292961359, + 0.003805126529186964, + 0.002581191249191761, + 0.007834656164050102, + 0.010822197422385216, + 0.011089320294559002, + 0.013947911560535431, + 0.010128209367394447, + 0.011844022199511528, + 0.008218131959438324, + 0.011080680415034294, + 0.007619892712682486, + 0.006932244170457125, + 0.013959352858364582, + 0.015637777745723724, + 0.00821230374276638, + 0.00928148441016674, + 0.01374436542391777, + 0.015748733654618263, + 0.016272058710455894, + 0.0075571006163954735, + 0.01812209188938141, + 0.01271449401974678, + 0.006658447906374931, + 0.019116831943392754, + 0.007793391589075327, + 0.0237481277436018, + 0.017510727047920227, + 0.009563949890434742, + 0.008333400823175907, + 0.007931873202323914, + 0.011279286816716194, + 0.013457221910357475, + 0.01468011923134327, + 0.014850936830043793, + 0.01407952606678009, + 0.008375736884772778, + 0.016346225515007973, + 0.013748139142990112, + 0.022691519930958748, + 0.019997721537947655, + 0.011473753489553928, + 0.015770666301250458, + 0.018649019300937653, + 0.012919729575514793, + 0.02374013140797615, + 0.02295832894742489, + 0.017786482349038124, + 0.020742159336805344, + 0.022661816328763962, + 0.015164047479629517, + 0.0075626978650689125, + 0.005857622716575861, + 0.012919029220938683, + 0.008291753008961678, + 0.008973555639386177, + 0.01088604424148798, + 0.01007647905498743, + 0.007633859757333994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.007630250882357359 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a female athlete as she desperately stumbles along a reddish-brown all-weather track under the blazing midday sun. Her dark brown hair is pulled back into a tight ponytail, but loose, sweat-soaked strands cling to her temples and forehead. Her face is a mask of pure agony; her jaw clenched, her eyes squeezed shut, her features contorted. She wears a dark blue, sweat-soaked running tank top and matching shorts. Her knees ache and blood, angry red scrapes weeping and caked with the grit of the track's surface. Her forward momentum is uncontrolled, a series of unsteady strides, her legs trembling with fatigue, before finally giving way completely. Her knees buckle, and she collapses hard, her torso hitting the track first in a boneless heap, her arms unable to break the fall. As she plummets, the camera tilts sharply downward, capturing the final moments of impact, creating a raw and visceral atmosphere of utter physical exhaustion and agonizing defeat.", + "seed": 9213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_4610_seed_9213_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020060194144025444, + "vlm_logprob": -0.020060194144025444, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.020060194144025444, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00958353839814663 + ], + "opposite_logprob_tokens": [ + -6.259583473205566 + ], + "logprob_reward": -0.00958353839814663, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00958353839814663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13333868980407715, + 0.026298681274056435, + 0.04182741418480873, + 0.05893898755311966, + 0.03699527308344841, + 0.2712017893791199, + 0.22769881784915924, + 0.11119436472654343, + 0.043347910046577454, + 0.049158040434122086 + ], + "credit_map": [ + 0.03883165121078491, + 0.036802757531404495, + 0.009530707262456417, + 0.00656094029545784, + 0.008003942668437958, + 0.01430527400225401, + 0.0075332471169531345, + 0.011770167388021946, + 0.0037826625630259514, + 0.002996008610352874, + 0.0026608463376760483, + 0.0023409307468682528, + 0.007435411214828491, + 0.0022328596096485853, + 0.002867675619199872, + 0.0019822875037789345, + 0.007227523718029261, + 0.005537095014005899, + 0.004490650724619627, + 0.00675058551132679, + 0.007112529128789902, + 0.003773850854486227, + 0.003397468477487564, + 0.003537712385877967, + 0.0069624329917132854, + 0.010807376354932785, + 0.005936152301728725, + 0.0064917937852442265, + 0.006831341423094273, + 0.0106046786531806, + 0.006318442989140749, + 0.004986770451068878, + 0.004235061351209879, + 0.0057420916855335236, + 0.003429942764341831, + 0.004689252935349941, + 0.004710232838988304, + 0.004995667841285467, + 0.004890072625130415, + 0.00430294731631875, + 0.01568135991692543, + 0.054067134857177734, + 0.024662703275680542, + 0.029454145580530167, + 0.07204306870698929, + 0.02281256951391697, + 0.02783835493028164, + 0.024642463773489, + 0.04170960932970047, + 0.02588621713221073, + 0.05157684162259102, + 0.013371280394494534, + 0.02703523449599743, + 0.03252864256501198, + 0.012711006216704845, + 0.022879991680383682, + 0.015148406848311424, + 0.010234517976641655, + 0.017823657020926476, + 0.015454244799911976, + 0.010987979359924793, + 0.018653301522135735, + 0.011953867971897125, + 0.010938392952084541, + 0.007145414594560862, + 0.004049769602715969, + 0.00565664516761899, + 0.004431803245097399, + 0.004693950060755014, + 0.005527321249246597, + 0.007337627466768026, + 0.004505375865846872, + 0.008873575367033482, + 0.005478924605995417, + 0.008621432818472385, + 0.003767996560782194, + 0.0057116542011499405, + 0.0077133080922067165, + 0.004509813152253628, + 0.004481334704905748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.00958353839814663 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043088290840387344 + ], + "opposite_logprob_tokens": [ + -3.16808819770813 + ], + "logprob_reward": -0.043088290840387344, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043088290840387344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07260634005069733, + 0.04846259579062462, + 0.04271816089749336, + 0.0411955863237381, + 0.04808724299073219, + 0.15223681926727295, + 0.307635635137558, + 0.1936625987291336, + 0.07589197903871536, + 0.01750301383435726 + ], + "credit_map": [ + 0.010565013624727726, + 0.009140281938016415, + 0.007134747691452503, + 0.006015635561197996, + 0.0097893588244915, + 0.009748220443725586, + 0.013260893523693085, + 0.006952191237360239, + 0.007177348714321852, + 0.006777034606784582, + 0.00412353128194809, + 0.004452012479305267, + 0.011603226885199547, + 0.004702224396169186, + 0.006354080513119698, + 0.003273134818300605, + 0.0044951243326067924, + 0.0030342303216457367, + 0.0051642428152263165, + 0.009155044332146645, + 0.005426572635769844, + 0.0034231548197567463, + 0.005306134931743145, + 0.006713658571243286, + 0.0029659292194992304, + 0.004009744618088007, + 0.00729767233133316, + 0.00691668875515461, + 0.004028886556625366, + 0.005418283399194479, + 0.005646218080073595, + 0.004912163596600294, + 0.004145687445998192, + 0.006498012691736221, + 0.004607596434652805, + 0.005259408615529537, + 0.007049507927149534, + 0.00800058152526617, + 0.004757735878229141, + 0.007768711075186729, + 0.009157557040452957, + 0.03258293867111206, + 0.01686370186507702, + 0.01018625870347023, + 0.04094044491648674, + 0.014695565216243267, + 0.012109879404306412, + 0.01570047065615654, + 0.05684796720743179, + 0.03225388005375862, + 0.08693324029445648, + 0.013417702168226242, + 0.03599460422992706, + 0.041430193930864334, + 0.01475807186216116, + 0.025999974459409714, + 0.02769252099096775, + 0.02369067445397377, + 0.02792913094162941, + 0.020325884222984314, + 0.027525875717401505, + 0.03171775862574577, + 0.0145431412383914, + 0.02023761346936226, + 0.012123432010412216, + 0.005844189319759607, + 0.011304900981485844, + 0.007011080160737038, + 0.00924560334533453, + 0.012412098236382008, + 0.010896039195358753, + 0.007054631598293781, + 0.00312896934337914, + 0.001828148029744625, + 0.002725477796047926, + 0.0016960372449830174, + 0.0013603770639747381, + 0.003524456638842821, + 0.0016833518166095018, + 0.001556195318698883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.043088290840387344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372887521982193 + ], + "opposite_logprob_tokens": [ + -3.408728837966919 + ], + "logprob_reward": -0.03372887521982193, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372887521982193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13099108636379242, + 0.058976009488105774, + 0.10736250132322311, + 0.09574157744646072, + 0.0398525670170784, + 0.19470703601837158, + 0.13517457246780396, + 0.13800416886806488, + 0.06529611349105835, + 0.03389435261487961 + ], + "credit_map": [ + 0.04511462524533272, + 0.044196467846632004, + 0.008574814535677433, + 0.0058358414098620415, + 0.006622593849897385, + 0.008644661866128445, + 0.005560661666095257, + 0.006441424135118723, + 0.008984105661511421, + 0.007641294971108437, + 0.004008741118013859, + 0.005302415229380131, + 0.017029765993356705, + 0.005711375270038843, + 0.006788024678826332, + 0.0035102837719023228, + 0.011411633342504501, + 0.009117555804550648, + 0.014839576557278633, + 0.02608094923198223, + 0.012901253066956997, + 0.006196741014719009, + 0.012735461816191673, + 0.014079337939620018, + 0.010985205881297588, + 0.014106265269219875, + 0.016479380428791046, + 0.012003863230347633, + 0.009168029762804508, + 0.012995381839573383, + 0.008736193180084229, + 0.011267256923019886, + 0.0028908594977110624, + 0.006092279218137264, + 0.003432944416999817, + 0.004721619188785553, + 0.005883294157683849, + 0.005408692639321089, + 0.006788888946175575, + 0.004633986856788397, + 0.010036472231149673, + 0.04331580176949501, + 0.032023049890995026, + 0.017803465947508812, + 0.026673393324017525, + 0.015468594618141651, + 0.017258094623684883, + 0.03212815523147583, + 0.025249049067497253, + 0.013598375953733921, + 0.03414084389805794, + 0.008570797741413116, + 0.014105454087257385, + 0.017994316294789314, + 0.009431552141904831, + 0.012084199115633965, + 0.020276980474591255, + 0.016220126301050186, + 0.021380173042416573, + 0.01572299562394619, + 0.016526740044355392, + 0.016694393008947372, + 0.015018638223409653, + 0.016164137050509453, + 0.010339177213609219, + 0.005614105146378279, + 0.011328255757689476, + 0.005079022608697414, + 0.006738276220858097, + 0.010237595066428185, + 0.010013834573328495, + 0.005945840384811163, + 0.007575559429824352, + 0.003887639380991459, + 0.004749873653054237, + 0.004021935630589724, + 0.0030531480442732573, + 0.004878833424299955, + 0.002789065707474947, + 0.0029382973443716764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.03372887521982193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078514158725739 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078514158725739, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078514158725739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04580444470047951, + 0.027591893449425697, + 0.045768555253744125, + 0.09198488295078278, + 0.0531393364071846, + 0.20708662271499634, + 0.3053887188434601, + 0.12128213047981262, + 0.08391893655061722, + 0.018034519627690315 + ], + "credit_map": [ + 0.009099051356315613, + 0.00804595835506916, + 0.003699931316077709, + 0.003126587485894561, + 0.006116839591413736, + 0.005191458389163017, + 0.005631989799439907, + 0.004892626311630011, + 0.004202783573418856, + 0.00270231906324625, + 0.002490083919838071, + 0.002989373169839382, + 0.005495554301887751, + 0.0026976566296070814, + 0.003919204231351614, + 0.0030949164647608995, + 0.00664557283744216, + 0.0044826180674135685, + 0.004799590911716223, + 0.008159473538398743, + 0.010207113809883595, + 0.0028485688380897045, + 0.00446369918063283, + 0.004161919001489878, + 0.01051097922027111, + 0.008075383491814137, + 0.02313191071152687, + 0.009154348634183407, + 0.0056234621442854404, + 0.014588872902095318, + 0.007766493130475283, + 0.013133424334228039, + 0.004134510178118944, + 0.005957812536507845, + 0.005867504980415106, + 0.006627685856074095, + 0.0061124167405068874, + 0.008855685591697693, + 0.00623196130618453, + 0.009351762011647224, + 0.010176031850278378, + 0.05129355937242508, + 0.023276442661881447, + 0.013373730704188347, + 0.05633361637592316, + 0.013435258530080318, + 0.011031581088900566, + 0.028166389092803, + 0.047145139425992966, + 0.04146721214056015, + 0.06469547748565674, + 0.02045256644487381, + 0.040958620607852936, + 0.03904447704553604, + 0.025669610127806664, + 0.025955624878406525, + 0.023516163229942322, + 0.012853164225816727, + 0.019338473677635193, + 0.010047521442174911, + 0.010520009323954582, + 0.017758270725607872, + 0.010797911323606968, + 0.016450615599751472, + 0.01589193381369114, + 0.00599003117531538, + 0.015873834490776062, + 0.005783981643617153, + 0.00893150083720684, + 0.012426309287548065, + 0.010639684274792671, + 0.008381660096347332, + 0.0028844887856394053, + 0.0014101008418947458, + 0.002435229951515794, + 0.0012291024904698133, + 0.0020467510912567377, + 0.003969138953834772, + 0.0016315706307068467, + 0.002428138628602028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.011078514158725739 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028217521030455828 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.0028217521030455828, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028217521030455828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041306428611278534, + 0.011142361909151077, + 0.026904936879873276, + 0.02504761517047882, + 0.038159824907779694, + 0.64444899559021, + 0.11273197084665298, + 0.04286334663629532, + 0.04669193923473358, + 0.010702516883611679 + ], + "credit_map": [ + 0.010583206079900265, + 0.010744522325694561, + 0.0030138734728097916, + 0.002117594936862588, + 0.0033625029027462006, + 0.005305043887346983, + 0.0032010816503316164, + 0.002978600561618805, + 0.0016023311764001846, + 0.0014536278322339058, + 0.0016653366619721055, + 0.0010267089819535613, + 0.0022014284040778875, + 0.0007614570786245167, + 0.0014092809287831187, + 0.001022190204821527, + 0.0029090875759720802, + 0.002413331065326929, + 0.003228842280805111, + 0.006386441644281149, + 0.003943584859371185, + 0.0019231110345572233, + 0.0029819689225405455, + 0.0031185694970190525, + 0.0016479038167744875, + 0.001734795980155468, + 0.0036044945009052753, + 0.004947314504534006, + 0.0028035782743245363, + 0.0026530048344284296, + 0.004588196519762278, + 0.003068326273933053, + 0.003918690141290426, + 0.0052733649499714375, + 0.003509528236463666, + 0.004809296689927578, + 0.006786742713302374, + 0.005644960328936577, + 0.004047687165439129, + 0.004169556312263012, + 0.03503420203924179, + 0.21233949065208435, + 0.056390028446912766, + 0.05360377952456474, + 0.11229024082422256, + 0.06483378261327744, + 0.04134850576519966, + 0.06860898435115814, + 0.013650891371071339, + 0.010193592868745327, + 0.017436891794204712, + 0.010180921293795109, + 0.014246419072151184, + 0.01707754097878933, + 0.01177891343832016, + 0.018166808411478996, + 0.005753840785473585, + 0.003999648150056601, + 0.007976732216775417, + 0.005132387392222881, + 0.004457986447960138, + 0.00580420158803463, + 0.004306727088987827, + 0.005431822035461664, + 0.008606691844761372, + 0.0032951326575130224, + 0.008845827542245388, + 0.003372951177880168, + 0.005551942158490419, + 0.006986879277974367, + 0.005347470752894878, + 0.004685047082602978, + 0.0017068536253646016, + 0.0009089757222682238, + 0.0013974433531984687, + 0.0011038986267521977, + 0.0012799251126125455, + 0.0020433056633919477, + 0.0010160552337765694, + 0.00124605861492455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.0028217521030455828 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a female athlete as she desperately stumbles along a reddish-brown all-weather track under the blazing midday sun. Her dark brown hair is pulled back into a tight ponytail, but loose, sweat-soaked strands cling to her temples and forehead. Her face is a mask of pure agony; her jaw clenched, her eyes squeezed shut, her features contorted. She wears a dark blue, sweat-soaked running tank top and matching shorts. Her knees ache and blood, angry red scrapes weeping and caked with the grit of the track's surface. Her forward momentum is uncontrolled, a series of unsteady strides, her legs trembling with fatigue, before finally giving way completely. Her knees buckle, and she collapses hard, her torso hitting the track first in a boneless heap, her arms unable to break the fall. As she plummets, the camera tilts sharply downward, capturing the final moments of impact, creating a raw and visceral atmosphere of utter physical exhaustion and agonizing defeat.", + "seed": 9214, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_4610_seed_9214_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028738191816955805, + "vlm_logprob": -0.028738191816955805, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.028738191816955805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01781802624464035 + ], + "opposite_logprob_tokens": [ + -5.267817974090576 + ], + "logprob_reward": -0.01781802624464035, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.01781802624464035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2707776427268982, + 0.1060141772031784, + 0.0691242441534996, + 0.06024216488003731, + 0.03971390426158905, + 0.1261831521987915, + 0.08016092330217361, + 0.09442627429962158, + 0.10528329014778137, + 0.04807424172759056 + ], + "credit_map": [ + 0.02465304732322693, + 0.024413710460066795, + 0.03975120931863785, + 0.056357577443122864, + 0.04404769465327263, + 0.018354829400777817, + 0.032185036689043045, + 0.031014544889330864, + 0.019734563305974007, + 0.013501693494617939, + 0.01248904038220644, + 0.012482898309826851, + 0.013094651512801647, + 0.013194717466831207, + 0.013071508146822453, + 0.00844510830938816, + 0.0073186964727938175, + 0.010603944770991802, + 0.006376197561621666, + 0.009568249806761742, + 0.00904479157179594, + 0.007568292319774628, + 0.008481906726956367, + 0.010162165388464928, + 0.009649538435041904, + 0.007592371199280024, + 0.0063001783564686775, + 0.006323361303657293, + 0.007778752129524946, + 0.00924061518162489, + 0.006179002113640308, + 0.007178348954766989, + 0.005202257074415684, + 0.004474851302802563, + 0.0036782552488148212, + 0.0039915055967867374, + 0.005699190776795149, + 0.006012216676026583, + 0.004812244325876236, + 0.005843381863087416, + 0.010462610982358456, + 0.013861946761608124, + 0.017588850110769272, + 0.02292918972671032, + 0.0189657025039196, + 0.01750010997056961, + 0.01262570172548294, + 0.012249038554728031, + 0.011055760085582733, + 0.011813333258032799, + 0.012817585840821266, + 0.00950664933770895, + 0.00895360205322504, + 0.0069521390832960606, + 0.009118959307670593, + 0.00994289480149746, + 0.015818702057003975, + 0.012183856219053268, + 0.012849176302552223, + 0.008122492581605911, + 0.009872842580080032, + 0.018589362502098083, + 0.009338535368442535, + 0.007651305291801691, + 0.018669236451387405, + 0.010862801223993301, + 0.01783955842256546, + 0.018008820712566376, + 0.011167393065989017, + 0.009260551072657108, + 0.0091015649959445, + 0.010373362340033054, + 0.008479153737425804, + 0.0060605620965361595, + 0.008715650998055935, + 0.004224246367812157, + 0.005376943852752447, + 0.005828517023473978, + 0.004695551469922066, + 0.0046936143189668655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01781802624464035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308212548494339 + ], + "opposite_logprob_tokens": [ + -3.1680822372436523 + ], + "logprob_reward": -0.04308212548494339, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308212548494339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23700031638145447, + 0.1077161654829979, + 0.028183521702885628, + 0.08240470290184021, + 0.06921698153018951, + 0.12247670441865921, + 0.09686210006475449, + 0.1041409969329834, + 0.11548677831888199, + 0.03651166707277298 + ], + "credit_map": [ + 0.022279947996139526, + 0.02262389473617077, + 0.029290927574038506, + 0.03257468715310097, + 0.0569089837372303, + 0.01857617311179638, + 0.029862144961953163, + 0.024883583188056946, + 0.020215129479765892, + 0.0126243457198143, + 0.015400341711938381, + 0.009944227524101734, + 0.012036115862429142, + 0.014808408915996552, + 0.012491636909544468, + 0.010195978917181492, + 0.0024243751540780067, + 0.004203446675091982, + 0.0032880352810025215, + 0.0035875500179827213, + 0.005278761964291334, + 0.003169239265844226, + 0.003732282668352127, + 0.0024998339358717203, + 0.011965564452111721, + 0.00914995651692152, + 0.008697126060724258, + 0.008502621203660965, + 0.010626052506268024, + 0.009755892679095268, + 0.010654743760824203, + 0.013052748516201973, + 0.008419795893132687, + 0.008708263747394085, + 0.006680596154183149, + 0.007182331755757332, + 0.009385377168655396, + 0.010176947340369225, + 0.011583476327359676, + 0.00708020618185401, + 0.008021097630262375, + 0.0119434529915452, + 0.01751904934644699, + 0.020095935091376305, + 0.012052829377353191, + 0.018053628504276276, + 0.018268851563334465, + 0.01652187667787075, + 0.012546910904347897, + 0.01590382680296898, + 0.013176565058529377, + 0.01784813590347767, + 0.010743644088506699, + 0.008036952465772629, + 0.007762928027659655, + 0.01084315124899149, + 0.02568105049431324, + 0.011780722998082638, + 0.0190646480768919, + 0.00956412497907877, + 0.009206180460751057, + 0.013365402817726135, + 0.006722410209476948, + 0.00875647459179163, + 0.017962608486413956, + 0.01157079916447401, + 0.016957895830273628, + 0.0191862341016531, + 0.014652500860393047, + 0.011506292968988419, + 0.010897106491029263, + 0.012753354385495186, + 0.005489886738359928, + 0.0038403430953621864, + 0.0070359548553824425, + 0.003854363225400448, + 0.005743685178458691, + 0.003088848665356636, + 0.0037333532236516476, + 0.003725233254954219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.04308212548494339 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207909435033798 + ], + "opposite_logprob_tokens": [ + -2.8120791912078857 + ], + "logprob_reward": -0.06207909435033798, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207909435033798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12961739301681519, + 0.08775169402360916, + 0.05609256401658058, + 0.028546812012791634, + 0.10986630618572235, + 0.15222805738449097, + 0.0841025784611702, + 0.1394643932580948, + 0.15059511363506317, + 0.06173507496714592 + ], + "credit_map": [ + 0.012629053555428982, + 0.013573932461440563, + 0.018607789650559425, + 0.02425570972263813, + 0.0193722452968359, + 0.008849290199577808, + 0.017399972304701805, + 0.014929397031664848, + 0.016665970906615257, + 0.007404956966638565, + 0.011576109565794468, + 0.008005686104297638, + 0.008413457311689854, + 0.015564165078103542, + 0.009631813503801823, + 0.01048953179270029, + 0.004928980488330126, + 0.006922406610101461, + 0.005075171124190092, + 0.008150982670485973, + 0.008775879628956318, + 0.004851154983043671, + 0.008969539776444435, + 0.008418448269367218, + 0.003912706393748522, + 0.003242677776142955, + 0.0037735733203589916, + 0.0029054528567939997, + 0.002999800955876708, + 0.003347771242260933, + 0.002296574180945754, + 0.0060682534240186214, + 0.009320554323494434, + 0.01549577247351408, + 0.01190323755145073, + 0.013734307140111923, + 0.01795250177383423, + 0.011571547947824001, + 0.012745295651257038, + 0.01714308373630047, + 0.01222244929522276, + 0.022522415965795517, + 0.021608099341392517, + 0.02337646670639515, + 0.01686156913638115, + 0.01986711472272873, + 0.015286804176867008, + 0.020483139902353287, + 0.01360821258276701, + 0.008363453671336174, + 0.013020360842347145, + 0.013784600421786308, + 0.007601460907608271, + 0.0072723994962871075, + 0.009569271467626095, + 0.010882820002734661, + 0.023806285113096237, + 0.014582606963813305, + 0.023057006299495697, + 0.016163840889930725, + 0.014317173510789871, + 0.02078709378838539, + 0.013709762133657932, + 0.01304063480347395, + 0.02574029751121998, + 0.014530864544212818, + 0.024600036442279816, + 0.025065533816814423, + 0.018922889605164528, + 0.015511334873735905, + 0.012847820296883583, + 0.013376334682106972, + 0.008381246589124203, + 0.004946588538587093, + 0.009916938841342926, + 0.005967346951365471, + 0.00977302435785532, + 0.009939763695001602, + 0.006016378756612539, + 0.006793786771595478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.06207909435033798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070714220404625 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.016070714220404625, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070714220404625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09670441597700119, + 0.10649306327104568, + 0.03309343010187149, + 0.03308666869997978, + 0.06526777148246765, + 0.30969953536987305, + 0.11489200592041016, + 0.09537564218044281, + 0.06631742417812347, + 0.07907002419233322 + ], + "credit_map": [ + 0.022067083045840263, + 0.018944203853607178, + 0.012213808484375477, + 0.007551137823611498, + 0.011718797497451305, + 0.007365277502685785, + 0.008842308074235916, + 0.008001797832548618, + 0.020812394097447395, + 0.009083195589482784, + 0.014774675481021404, + 0.010579503141343594, + 0.010449551045894623, + 0.017990535125136375, + 0.012072952464222908, + 0.01073025818914175, + 0.0031117149628698826, + 0.004160833545029163, + 0.003897685557603836, + 0.004578887950628996, + 0.004891911055892706, + 0.004753504879772663, + 0.0036472019273787737, + 0.004051689524203539, + 0.003733644261956215, + 0.004699256736785173, + 0.0037592581938952208, + 0.003564845072105527, + 0.005029041785746813, + 0.004996609874069691, + 0.002938459627330303, + 0.004365552682429552, + 0.0073487418703734875, + 0.00857118796557188, + 0.00518787419423461, + 0.006996920797973871, + 0.011123595759272575, + 0.010467042215168476, + 0.007840803824365139, + 0.0077316053211688995, + 0.02731473185122013, + 0.039047736674547195, + 0.04456687718629837, + 0.05454558506608009, + 0.029160287231206894, + 0.04555037245154381, + 0.03292209282517433, + 0.03659185767173767, + 0.013893707655370235, + 0.016183264553546906, + 0.017104309052228928, + 0.01986275427043438, + 0.0112915663048625, + 0.010752743110060692, + 0.013585085980594158, + 0.012218581512570381, + 0.017565825954079628, + 0.01055844034999609, + 0.015496785752475262, + 0.012604812160134315, + 0.010780558921396732, + 0.00958008412271738, + 0.008968104608356953, + 0.00982103031128645, + 0.01164485327899456, + 0.006088425870984793, + 0.012287190183997154, + 0.008942198939621449, + 0.008885536342859268, + 0.004774325992912054, + 0.006880946923047304, + 0.0068139443174004555, + 0.014919442124664783, + 0.010851221159100533, + 0.015547257848083973, + 0.005834197159856558, + 0.0077972509898245335, + 0.012391199357807636, + 0.005901777651160955, + 0.005827674176543951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.016070714220404625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640998784452677 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640998784452677, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640998784452677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1509704440832138, + 0.04666565731167793, + 0.04017629846930504, + 0.023318594321608543, + 0.06436298787593842, + 0.20956671237945557, + 0.060129787772893906, + 0.23525655269622803, + 0.12700189650058746, + 0.042551081627607346 + ], + "credit_map": [ + 0.03229764848947525, + 0.028085224330425262, + 0.01855490542948246, + 0.01700599677860737, + 0.017958929762244225, + 0.009745228104293346, + 0.01545772235840559, + 0.011864802800118923, + 0.004978240933269262, + 0.009074767120182514, + 0.005608548875898123, + 0.005513259209692478, + 0.005750168580561876, + 0.006415020674467087, + 0.0051993331871926785, + 0.004126317333430052, + 0.0030442913994193077, + 0.005142359063029289, + 0.004233922343701124, + 0.006360764615237713, + 0.006170639302581549, + 0.0038756425492465496, + 0.0063282824121415615, + 0.0050203995779156685, + 0.004381885752081871, + 0.002784203737974167, + 0.0033172140829265118, + 0.0021132382098585367, + 0.00317513570189476, + 0.0026864486280828714, + 0.0021181029733270407, + 0.002742365701124072, + 0.007544760126620531, + 0.010016458109021187, + 0.0058799306862056255, + 0.007361894939094782, + 0.01183544471859932, + 0.008296078070998192, + 0.007545468397438526, + 0.005882950033992529, + 0.019929906353354454, + 0.021051382645964622, + 0.02889312244951725, + 0.03730029985308647, + 0.017029689624905586, + 0.03834107145667076, + 0.025296567007899284, + 0.021724671125411987, + 0.0077786315232515335, + 0.007322130259126425, + 0.008538524620234966, + 0.007818697020411491, + 0.0067150406539440155, + 0.008157719857990742, + 0.005842638202011585, + 0.007956407032907009, + 0.04576728120446205, + 0.02495058998465538, + 0.040685031563043594, + 0.03101116605103016, + 0.026058606803417206, + 0.022950554266572, + 0.02137991227209568, + 0.022453423589468002, + 0.022619035094976425, + 0.014347054995596409, + 0.019226152449846268, + 0.019228331744670868, + 0.015186676755547523, + 0.012375001795589924, + 0.012254327535629272, + 0.011765317060053349, + 0.00965863186866045, + 0.004989935550838709, + 0.009310344234108925, + 0.0025130463764071465, + 0.003882709424942732, + 0.005595449823886156, + 0.0026820432394742966, + 0.003918919246643782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.004640998784452677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a female athlete as she desperately stumbles along a reddish-brown all-weather track under the blazing midday sun. Her dark brown hair is pulled back into a tight ponytail, but loose, sweat-soaked strands cling to her temples and forehead. Her face is a mask of pure agony; her jaw clenched, her eyes squeezed shut, her features contorted. She wears a dark blue, sweat-soaked running tank top and matching shorts. Her knees ache and blood, angry red scrapes weeping and caked with the grit of the track's surface. Her forward momentum is uncontrolled, a series of unsteady strides, her legs trembling with fatigue, before finally giving way completely. Her knees buckle, and she collapses hard, her torso hitting the track first in a boneless heap, her arms unable to break the fall. As she plummets, the camera tilts sharply downward, capturing the final moments of impact, creating a raw and visceral atmosphere of utter physical exhaustion and agonizing defeat.", + "seed": 9215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_6/index_4610_seed_9215_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02135051293298602, + "vlm_logprob": -0.02135051293298602, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02135051293298602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01829558052122593 + ], + "opposite_logprob_tokens": [ + -5.518295764923096 + ], + "logprob_reward": -0.01829558052122593, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.01829558052122593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055478405207395554, + 0.035265274345874786, + 0.029237976297736168, + 0.0991111770272255, + 0.06665614992380142, + 0.06418953090906143, + 0.04612263664603233, + 0.12693694233894348, + 0.04078420251607895, + 0.4362177550792694 + ], + "credit_map": [ + 0.006559968926012516, + 0.0043270522728562355, + 0.012694321572780609, + 0.005484487395733595, + 0.008114299736917019, + 0.006539326161146164, + 0.004630466923117638, + 0.007128478027880192, + 0.0035075941123068333, + 0.005929419305175543, + 0.0063899727538228035, + 0.002562378067523241, + 0.0050883363001048565, + 0.004717664793133736, + 0.003548921085894108, + 0.003520986996591091, + 0.004563441500067711, + 0.004225362092256546, + 0.0036352442111819983, + 0.001905751763843, + 0.004554435610771179, + 0.004063292406499386, + 0.0038985724095255136, + 0.0023918747901916504, + 0.015444966033101082, + 0.005452889949083328, + 0.012653850018978119, + 0.005414444021880627, + 0.005527155473828316, + 0.02129710279405117, + 0.011695804074406624, + 0.02162495069205761, + 0.0037028479855507612, + 0.0037759868428111076, + 0.010290867649018764, + 0.00807997863739729, + 0.006490632891654968, + 0.00456572650000453, + 0.02128281444311142, + 0.008467295207083225, + 0.005598641000688076, + 0.0067350417375564575, + 0.006773257628083229, + 0.00898818764835596, + 0.011126908473670483, + 0.007486159447580576, + 0.009479368105530739, + 0.00800196174532175, + 0.006499611306935549, + 0.003754206234589219, + 0.006598109379410744, + 0.004811101593077183, + 0.00624433346092701, + 0.0049359449185431, + 0.007205761503428221, + 0.00607356196269393, + 0.018255548551678658, + 0.009520038962364197, + 0.012759648263454437, + 0.011209839954972267, + 0.007672312203794718, + 0.039883486926555634, + 0.008578565903007984, + 0.019057491794228554, + 0.005581558216363192, + 0.005333341658115387, + 0.003889653366059065, + 0.005439118482172489, + 0.003637540154159069, + 0.00405119638890028, + 0.006674704607576132, + 0.00617709057405591, + 0.020501529797911644, + 0.01976170390844345, + 0.042629025876522064, + 0.026279769837856293, + 0.031498368829488754, + 0.024159885942935944, + 0.13160808384418488, + 0.13977934420108795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.01829558052122593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982192672789097 + ], + "opposite_logprob_tokens": [ + -3.5298218727111816 + ], + "logprob_reward": -0.02982192672789097, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982192672789097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08489449322223663, + 0.08463069051504135, + 0.03452396020293236, + 0.04800085723400116, + 0.019914068281650543, + 0.023288097232580185, + 0.16073329746723175, + 0.03731992095708847, + 0.026873275637626648, + 0.4798213541507721 + ], + "credit_map": [ + 0.009862525388598442, + 0.007635443937033415, + 0.013739126734435558, + 0.011719224974513054, + 0.008587084710597992, + 0.011976988986134529, + 0.009007500484585762, + 0.012366591952741146, + 0.0068678706884384155, + 0.012303308583796024, + 0.010780188255012035, + 0.0057485345751047134, + 0.010744703933596611, + 0.00889312382787466, + 0.02022509090602398, + 0.009067868813872337, + 0.005080861039459705, + 0.004578633699566126, + 0.00347505952231586, + 0.0032203616574406624, + 0.007273956201970577, + 0.0038944215048104525, + 0.0033746908884495497, + 0.0036259705666452646, + 0.005654375068843365, + 0.004334635101258755, + 0.007614581845700741, + 0.002672314876690507, + 0.002618364989757538, + 0.008056800812482834, + 0.006187664344906807, + 0.01086211483925581, + 0.0014045865973457694, + 0.0018960469169542193, + 0.002864082111045718, + 0.002331087365746498, + 0.002634064294397831, + 0.0023792001884430647, + 0.00360471336171031, + 0.0028002853505313396, + 0.0017118860268965364, + 0.0023061614483594894, + 0.0023971216287463903, + 0.002969477092847228, + 0.004711962305009365, + 0.002785417716950178, + 0.003874744987115264, + 0.002531321719288826, + 0.023232266306877136, + 0.012408184818923473, + 0.029915310442447662, + 0.019042061641812325, + 0.017477640882134438, + 0.015947548672556877, + 0.01686502993106842, + 0.0258452370762825, + 0.0042625595815479755, + 0.004225751850754023, + 0.005409373436123133, + 0.004276908002793789, + 0.0052736373618245125, + 0.0044907210394740105, + 0.004425796214491129, + 0.004955171141773462, + 0.004227032884955406, + 0.0028164461255073547, + 0.0031118381302803755, + 0.0031247632578015327, + 0.0022412254475057125, + 0.0033222052734345198, + 0.003571509150788188, + 0.004458253271877766, + 0.021422874182462692, + 0.024019818753004074, + 0.03729316219687462, + 0.03150205314159393, + 0.030555972829461098, + 0.025899305939674377, + 0.15071986615657806, + 0.1584082692861557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.02982192672789097 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370685875415802 + ], + "opposite_logprob_tokens": [ + -3.4087069034576416 + ], + "logprob_reward": -0.03370685875415802, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370685875415802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06423042714595795, + 0.03762476146221161, + 0.04385826364159584, + 0.07829649746417999, + 0.034041982144117355, + 0.033689651638269424, + 0.08475276082754135, + 0.23904050886631012, + 0.14651785790920258, + 0.2379472851753235 + ], + "credit_map": [ + 0.009041360579431057, + 0.004912865348160267, + 0.01474306732416153, + 0.006712359841912985, + 0.006120441015809774, + 0.008994328789412975, + 0.006521908566355705, + 0.007184096146374941, + 0.005314558744430542, + 0.009714790619909763, + 0.0038256209809333086, + 0.0028832885436713696, + 0.004261721856892109, + 0.00541045842692256, + 0.0027818663511425257, + 0.0034324564039707184, + 0.007221216801553965, + 0.004659098107367754, + 0.002604008186608553, + 0.004445783793926239, + 0.009974959306418896, + 0.005719128996133804, + 0.005393131636083126, + 0.003840937279164791, + 0.012600556015968323, + 0.005061902571469545, + 0.010578223504126072, + 0.003923684358596802, + 0.00453685550019145, + 0.016926012933254242, + 0.008826661854982376, + 0.015842601656913757, + 0.0026928584557026625, + 0.0033893510699272156, + 0.0048508248291909695, + 0.003553603542968631, + 0.004910685587674379, + 0.005789636634290218, + 0.005262416787445545, + 0.0035926057025790215, + 0.0029851985163986683, + 0.002356122713536024, + 0.00405180873349309, + 0.004490980412811041, + 0.005761736538261175, + 0.0035808624234050512, + 0.007059107068926096, + 0.0034038363955914974, + 0.00999963004142046, + 0.007092773914337158, + 0.011563023552298546, + 0.012507673352956772, + 0.010427969507873058, + 0.008701927028596401, + 0.01133730262517929, + 0.013122457079589367, + 0.025656146928668022, + 0.023757582530379295, + 0.019828233867883682, + 0.02451285347342491, + 0.02122780680656433, + 0.05936109647154808, + 0.026766132563352585, + 0.03793065622448921, + 0.01807619072496891, + 0.015455187298357487, + 0.01597665250301361, + 0.022533614188432693, + 0.012339236214756966, + 0.016635224223136902, + 0.02386612631380558, + 0.021635614335536957, + 0.014931797049939632, + 0.019267717376351357, + 0.022575896233320236, + 0.016852565109729767, + 0.02819027565419674, + 0.03371797502040863, + 0.05054749548435211, + 0.05186355859041214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.03370685875415802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189996480941772 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018189996480941772, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189996480941772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08637873083353043, + 0.042561955749988556, + 0.032773468643426895, + 0.05994406342506409, + 0.0307072214782238, + 0.06453953683376312, + 0.23102980852127075, + 0.0793919563293457, + 0.08102834969758987, + 0.2916449010372162 + ], + "credit_map": [ + 0.008964044973254204, + 0.007619077805429697, + 0.018232200294733047, + 0.01051107607781887, + 0.008574169129133224, + 0.010195454582571983, + 0.00879580620676279, + 0.01348690502345562, + 0.00465579004958272, + 0.011130287311971188, + 0.004254891537129879, + 0.0027363623958081007, + 0.00624425895512104, + 0.006729895249009132, + 0.0030978326685726643, + 0.0037126345559954643, + 0.005407092627137899, + 0.003985862247645855, + 0.003280051751062274, + 0.0029143993742763996, + 0.0061439950950443745, + 0.0039176554419100285, + 0.0034920971374958754, + 0.003632314968854189, + 0.010065063834190369, + 0.006127391010522842, + 0.007579082623124123, + 0.0036669368855655193, + 0.003326268633827567, + 0.013172123581171036, + 0.005894051399081945, + 0.010113146156072617, + 0.002155614085495472, + 0.0027849492616951466, + 0.003704553470015526, + 0.0031531071290373802, + 0.004199012648314238, + 0.004075278528034687, + 0.0067875380627810955, + 0.0038471694570034742, + 0.005517779849469662, + 0.008546623401343822, + 0.007494701072573662, + 0.008922976441681385, + 0.007871962152421474, + 0.008432134054601192, + 0.009519177488982677, + 0.008234185166656971, + 0.04319841042160988, + 0.017903927713632584, + 0.04348593205213547, + 0.029132746160030365, + 0.01669425703585148, + 0.018266180530190468, + 0.029709862545132637, + 0.03263849392533302, + 0.009718905203044415, + 0.009661909192800522, + 0.012011870741844177, + 0.01011065486818552, + 0.007325462065637112, + 0.009519397281110287, + 0.009437636472284794, + 0.011606122367084026, + 0.009425705298781395, + 0.011274469085037708, + 0.009559218771755695, + 0.010524644516408443, + 0.00819002278149128, + 0.011081373319029808, + 0.009251149371266365, + 0.011721763759851456, + 0.015811828896403313, + 0.010846053250133991, + 0.0281328484416008, + 0.014108852483332157, + 0.024377357214689255, + 0.020898304879665375, + 0.08587154746055603, + 0.09159810841083527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.018189996480941772 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective following a female athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a dark blue, sweat-soaked running tank top and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are visible angry red scrapes with grit on the athlete's knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the athlete collapse by hitting the track with her torso first in a boneless heap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt sharply downward to capture the final moments of impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738202180713415 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738202180713415, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738202180713415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06324639916419983, + 0.04629329591989517, + 0.04167712479829788, + 0.07978829741477966, + 0.09713989496231079, + 0.11415307968854904, + 0.07567442953586578, + 0.18964225053787231, + 0.046649426221847534, + 0.2457357496023178 + ], + "credit_map": [ + 0.00796067900955677, + 0.008551092818379402, + 0.009700520895421505, + 0.007805515546351671, + 0.006786984391510487, + 0.007300370838493109, + 0.006448935251682997, + 0.008692299015820026, + 0.005033702123910189, + 0.009455561637878418, + 0.006290433928370476, + 0.0040057869628071785, + 0.006509712897241116, + 0.005925529170781374, + 0.004005598835647106, + 0.005066973157227039, + 0.0056755151599645615, + 0.005372398998588324, + 0.0054842634126544, + 0.006514952052384615, + 0.004975006449967623, + 0.004751383792608976, + 0.0035488721914589405, + 0.0053547280840575695, + 0.008800201117992401, + 0.007392790634185076, + 0.011295944452285767, + 0.006065094377845526, + 0.005564830265939236, + 0.009410424157977104, + 0.011310677975416183, + 0.019948337227106094, + 0.00547245005145669, + 0.007090166211128235, + 0.017128588631749153, + 0.006417059805244207, + 0.008975961245596409, + 0.006218967493623495, + 0.03496307507157326, + 0.010873638093471527, + 0.00905671902000904, + 0.009316733106970787, + 0.01289296057075262, + 0.013952585868537426, + 0.02092866040766239, + 0.011088666506111622, + 0.019256047904491425, + 0.017660707235336304, + 0.012620583176612854, + 0.006990675814449787, + 0.015187152661383152, + 0.008282752707600594, + 0.007426249794661999, + 0.007256851997226477, + 0.0067825703881680965, + 0.0111275939270854, + 0.023502536118030548, + 0.02149304747581482, + 0.0173343475908041, + 0.0224351454526186, + 0.015995414927601814, + 0.04573250934481621, + 0.018421217799186707, + 0.024728043004870415, + 0.0055452799424529076, + 0.007134060841053724, + 0.005226421169936657, + 0.005206417292356491, + 0.005170724354684353, + 0.004657462704926729, + 0.007059779483824968, + 0.006649279035627842, + 0.018256716430187225, + 0.02310064621269703, + 0.03932974115014076, + 0.01974782533943653, + 0.026372026652097702, + 0.03356267511844635, + 0.04294423386454582, + 0.04242188110947609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.006738202180713415 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_8/index_tensor([9393, 9393])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_8/index_tensor([9393, 9393])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b659dc67dce1ef71db289a86ecf773daec271dca --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_8/index_tensor([9393, 9393])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion, low-angle tracking shot follows a woman as she makes a perfect dive. The camera first focuses on her bare feet gripping the edge of a rugged, sun-drenched granite cliff, then tilts upward to capture her entire body against a bright, cloudless sky. She is a woman with an athletic, toned figure, her long, dark hair pulled back in a sleek ponytail that scatters momentarily as she thrusts forward. She wears a minimalist black one-piece swimsuit made of matte-finished, form-fitting material. Her body forms a breathtakingly graceful arc in mid-air, her arms extended forward, her legs straight, toes pointed, her facial expression focused yet filled with tranquil bliss. Below her, the water is a deep blue, crystal clear, revealing a sandy seabed and smooth underwater rocks. The bright midday sun casts sharp shadows across the cliff face and creates dazzling, crystalline sparkles on the water's surface. She cuts into the water with powerful precision, creating the smallest splash, her body disappearing seamlessly into the blue depths as the camera captures the expanding concentric ripples as the only evidence of her entry, evoking a profound sense of freedom and natural grace.", + "seed": 25838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_8/index_9393_seed_25838_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.26864915173500775, + "vlm_logprob": -0.26864915173500775, + "vlm_yes_no_margin": 1.875000011920929, + "vlm_token_credit": -0.26864915173500775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4484438896179199 + ], + "opposite_logprob_tokens": [ + -1.07344388961792 + ], + "logprob_reward": -0.4484438896179199, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4484438896179199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04001303017139435, + 0.017715955153107643, + 0.015432901680469513, + 0.015193209983408451, + 0.01561406534165144, + 0.058862946927547455, + 0.042640458792448044, + 0.05790899321436882, + 0.6424055099487305, + 0.09421288222074509 + ], + "credit_map": [ + 0.008088413625955582, + 0.006945291068404913, + 0.00480927899479866, + 0.003938495647162199, + 0.003968596924096346, + 0.0040030889213085175, + 0.003708956064656377, + 0.00455091055482626, + 0.0019731707870960236, + 0.002618079772219062, + 0.0034934086725115776, + 0.0018890624633058906, + 0.002224624389782548, + 0.0019867196679115295, + 0.001565386657603085, + 0.0019655032083392143, + 0.002513190032914281, + 0.0020726744551211596, + 0.0014967729803174734, + 0.0028997750487178564, + 0.0023312983103096485, + 0.0011262651532888412, + 0.0017949732718989253, + 0.0011979526607319713, + 0.0018321077805012465, + 0.0024646224919706583, + 0.0018155050929635763, + 0.0017660108860582113, + 0.0016312886727973819, + 0.0014352293219417334, + 0.002137785078957677, + 0.002110660308972001, + 0.0019170051673427224, + 0.002422553254291415, + 0.0016453302232548594, + 0.0017495912034064531, + 0.0012563865166157484, + 0.001429956522770226, + 0.002788195852190256, + 0.0024050462525337934, + 0.008315431885421276, + 0.009956954047083855, + 0.006328950170427561, + 0.006242097821086645, + 0.00429281173273921, + 0.006534887012094259, + 0.006782475393265486, + 0.010409338399767876, + 0.00591157004237175, + 0.006614589598029852, + 0.0035020997747778893, + 0.00589734734967351, + 0.006314192898571491, + 0.005698713008314371, + 0.004993324633687735, + 0.003708622185513377, + 0.003237830474972725, + 0.0046583921648561954, + 0.0039014420472085476, + 0.004047483205795288, + 0.014340277761220932, + 0.01241370104253292, + 0.010627496987581253, + 0.004682367201894522, + 0.0103589603677392, + 0.006236815359443426, + 0.0076562343165278435, + 0.0051661208271980286, + 0.30427297949790955, + 0.2967014014720917, + 0.007582084741443396, + 0.004430974833667278, + 0.001663355389609933, + 0.0018723204266279936, + 0.002119537675753236, + 0.0013406723737716675, + 0.04244350269436836, + 0.04123036190867424, + 0.001962102949619293, + 0.00158102682325989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.4484438896179199 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289166331291199 + ], + "opposite_logprob_tokens": [ + -1.0539166927337646 + ], + "logprob_reward": -0.4289166331291199, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4289166331291199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039386771619319916, + 0.10079418122768402, + 0.03570025414228439, + 0.01838665083050728, + 0.012845219112932682, + 0.03857672959566116, + 0.05828497186303139, + 0.04961548000574112, + 0.5516494512557983, + 0.0947602391242981 + ], + "credit_map": [ + 0.00400423351675272, + 0.004744780249893665, + 0.006991906091570854, + 0.004610640928149223, + 0.00459787854924798, + 0.004743507597595453, + 0.005473833531141281, + 0.004219996742904186, + 0.02408776246011257, + 0.015351193957030773, + 0.011615014635026455, + 0.0057428497821092606, + 0.01398918591439724, + 0.013700761832296848, + 0.007880422286689281, + 0.008426999673247337, + 0.0044297571294009686, + 0.007068270817399025, + 0.003379885805770755, + 0.004360450431704521, + 0.005459023639559746, + 0.002904757158830762, + 0.004646227229386568, + 0.0034518849570304155, + 0.0027262179646641016, + 0.0030209608376026154, + 0.0029192762449383736, + 0.002314498880878091, + 0.0017401307122781873, + 0.001476594596169889, + 0.0026111272163689137, + 0.001577847171574831, + 0.0025454177521169186, + 0.003167303977534175, + 0.0016767174238339067, + 0.0016150114824995399, + 0.001025424455292523, + 0.0010073879966512322, + 0.0009472763049416244, + 0.0008606810006313026, + 0.005468780640512705, + 0.0073084658943116665, + 0.0045928750187158585, + 0.003538501914590597, + 0.0032649789936840534, + 0.004008582793176174, + 0.005959576927125454, + 0.00443497346714139, + 0.007792001124471426, + 0.0076829432509839535, + 0.0092753442004323, + 0.013262047432363033, + 0.00424471078440547, + 0.00564989959821105, + 0.004684641491621733, + 0.005693389102816582, + 0.0026822483632713556, + 0.004400594159960747, + 0.0025830105878412724, + 0.0034918361343443394, + 0.01459460612386465, + 0.011321109719574451, + 0.007912651635706425, + 0.002629424910992384, + 0.010460898280143738, + 0.008388137444853783, + 0.007105316035449505, + 0.01103798858821392, + 0.25099900364875793, + 0.2459191232919693, + 0.009196081198751926, + 0.008542931638658047, + 0.0032485045958310366, + 0.0030276821926236153, + 0.0048985956236720085, + 0.003397858701646328, + 0.03795577958226204, + 0.036655180156230927, + 0.0028423566836863756, + 0.0027342953253537416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.4289166331291199 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28134799003601074 + ], + "opposite_logprob_tokens": [ + -1.4063479900360107 + ], + "logprob_reward": -0.28134799003601074, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28134799003601074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1853116899728775, + 0.08779497444629669, + 0.039697542786598206, + 0.033300258219242096, + 0.015459400601685047, + 0.03318246081471443, + 0.06199609860777855, + 0.025648655369877815, + 0.44855737686157227, + 0.06905154883861542 + ], + "credit_map": [ + 0.02354210428893566, + 0.020843833684921265, + 0.021268436685204506, + 0.018316136673092842, + 0.02402976155281067, + 0.024479443207383156, + 0.029393693432211876, + 0.023438287898898125, + 0.013817375525832176, + 0.010860194452106953, + 0.013491107150912285, + 0.01031169667840004, + 0.012933499179780483, + 0.008739661425352097, + 0.00965003203600645, + 0.007991413585841656, + 0.0051161847077310085, + 0.00512566976249218, + 0.0057376776821911335, + 0.007056974805891514, + 0.005475436802953482, + 0.003354788525030017, + 0.004411733243614435, + 0.0034190788865089417, + 0.005364357493817806, + 0.008110626600682735, + 0.0035258857533335686, + 0.0033432680647820234, + 0.0034198949579149485, + 0.002605189336463809, + 0.004435969050973654, + 0.0024950688239187002, + 0.0030176315922290087, + 0.003006818238645792, + 0.002222995739430189, + 0.0017227609641849995, + 0.0015880005666986108, + 0.0011541404528543353, + 0.001536430325359106, + 0.0012106226058676839, + 0.0036316870246082544, + 0.005422855727374554, + 0.003315909532830119, + 0.005120947491377592, + 0.002377142198383808, + 0.003742515342310071, + 0.003896441077813506, + 0.005674963817000389, + 0.00897909700870514, + 0.009818105027079582, + 0.006239431910216808, + 0.01041505765169859, + 0.006227927748113871, + 0.007108988706022501, + 0.0058038062416017056, + 0.007403692230582237, + 0.001998908119276166, + 0.0025938579346984625, + 0.0018379020038992167, + 0.002542700618505478, + 0.0041095600463449955, + 0.003622487187385559, + 0.006573841441422701, + 0.002369399182498455, + 0.00782437901943922, + 0.005929355043917894, + 0.005276361480355263, + 0.0029357753228396177, + 0.2106855809688568, + 0.20570524036884308, + 0.0055707888677716255, + 0.004629919305443764, + 0.0016278534894809127, + 0.0014133682707324624, + 0.001231102622114122, + 0.0012146317167207599, + 0.03024314157664776, + 0.029732700437307358, + 0.001740576233714819, + 0.001848182757385075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.28134799003601074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464171078056097 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464171078056097, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464171078056097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08389385789632797, + 0.08931946754455566, + 0.07691309601068497, + 0.04389916732907295, + 0.03440562263131142, + 0.06909718364477158, + 0.0445314459502697, + 0.23504599928855896, + 0.10322222113609314, + 0.21967197954654694 + ], + "credit_map": [ + 0.011078188195824623, + 0.010023572482168674, + 0.01059029158204794, + 0.008435549214482307, + 0.014015288092195988, + 0.00862663984298706, + 0.01086112204939127, + 0.010263205505907536, + 0.011444821022450924, + 0.008491143584251404, + 0.014905148185789585, + 0.010633568279445171, + 0.014244421385228634, + 0.009346510283648968, + 0.010661469772458076, + 0.009592384099960327, + 0.010137101635336876, + 0.011203878559172153, + 0.010019826702773571, + 0.014834222383797169, + 0.0099635636433959, + 0.006647908594459295, + 0.007770002353936434, + 0.006336593069136143, + 0.00685980636626482, + 0.007926116697490215, + 0.004900521598756313, + 0.0057458458468317986, + 0.003671671263873577, + 0.004076340701431036, + 0.0065827155485749245, + 0.004136150237172842, + 0.006175210699439049, + 0.007327789906412363, + 0.005378019064664841, + 0.0053758942522108555, + 0.0024061824660748243, + 0.00266596139408648, + 0.003063604701310396, + 0.0020129606127738953, + 0.01113169640302658, + 0.011789388954639435, + 0.007763806730508804, + 0.008055970072746277, + 0.005642800126224756, + 0.007165396120399237, + 0.009575302712619305, + 0.007972819730639458, + 0.006139306351542473, + 0.00692373514175415, + 0.00446416437625885, + 0.005466754082590342, + 0.0045538535341620445, + 0.006412804592400789, + 0.004327001050114632, + 0.0062438263557851315, + 0.01046078186482191, + 0.014052113518118858, + 0.009689435362815857, + 0.008486748673021793, + 0.09266433119773865, + 0.07140877097845078, + 0.01864699460566044, + 0.009636824950575829, + 0.004622700624167919, + 0.006491788197308779, + 0.005616762675344944, + 0.005155132617801428, + 0.036950040608644485, + 0.034890856593847275, + 0.004510816186666489, + 0.004984112456440926, + 0.0053852577693760395, + 0.00686231255531311, + 0.0070339092053473, + 0.006497525610029697, + 0.09254949539899826, + 0.09023735672235489, + 0.006451526656746864, + 0.004654591903090477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.00464171078056097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798955351114273 + ], + "opposite_logprob_tokens": [ + -1.8048955202102661 + ], + "logprob_reward": -0.1798955351114273, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798955351114273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04782075062394142, + 0.07188183069229126, + 0.023568950593471527, + 0.017536437138915062, + 0.008460117504000664, + 0.0269220732152462, + 0.03427615761756897, + 0.3276752233505249, + 0.41860780119895935, + 0.023250611498951912 + ], + "credit_map": [ + 0.006792519707232714, + 0.005399942398071289, + 0.00653018057346344, + 0.005863206461071968, + 0.004244931973516941, + 0.005794817581772804, + 0.006829950958490372, + 0.006365200970321894, + 0.015255585312843323, + 0.010979893617331982, + 0.0076791075989604, + 0.005098249763250351, + 0.011522816494107246, + 0.00860598124563694, + 0.006457128562033176, + 0.006283071357756853, + 0.0028177679050713778, + 0.0032649533823132515, + 0.0027598964516073465, + 0.0034799606073647738, + 0.003917972091585398, + 0.0022328388877213, + 0.0033127048518508673, + 0.0017828550189733505, + 0.002485646167770028, + 0.004524239804595709, + 0.0017149847699329257, + 0.0021431208588182926, + 0.0017678439617156982, + 0.0015842251013964415, + 0.0021274180617183447, + 0.0011889588786289096, + 0.0015009456546977162, + 0.0011098196264356375, + 0.0014082277193665504, + 0.0009948504157364368, + 0.000793183920904994, + 0.0006620521889999509, + 0.0011734876316040754, + 0.0008175505208782852, + 0.004282951820641756, + 0.0032979759853333235, + 0.0032775939907878637, + 0.0035249076317995787, + 0.0019224395509809256, + 0.0029246413614600897, + 0.0035219965502619743, + 0.00416956702247262, + 0.004986458458006382, + 0.003297431394457817, + 0.005625809542834759, + 0.007252023555338383, + 0.0028070842381566763, + 0.004306897055357695, + 0.0018033317755907774, + 0.0041971225291490555, + 0.014368509873747826, + 0.021014263853430748, + 0.015249429270625114, + 0.01801988109946251, + 0.1132301315665245, + 0.0873045027256012, + 0.043260589241981506, + 0.01522792037576437, + 0.006855324376374483, + 0.00510528264567256, + 0.004945616703480482, + 0.0031546219252049923, + 0.1964174211025238, + 0.19228404760360718, + 0.005160843953490257, + 0.00468464707955718, + 0.0010837046429514885, + 0.0009368474129587412, + 0.0017227274365723133, + 0.0014752575661987066, + 0.007149068638682365, + 0.007515808567404747, + 0.001837273477576673, + 0.0015299239894375205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.1798955351114273 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, low-angle tracking shot follows a woman as she makes a perfect dive. The camera first focuses on her bare feet gripping the edge of a rugged, sun-drenched granite cliff, then tilts upward to capture her entire body against a bright, cloudless sky. She is a woman with an athletic, toned figure, her long, dark hair pulled back in a sleek ponytail that scatters momentarily as she thrusts forward. She wears a minimalist black one-piece swimsuit made of matte-finished, form-fitting material. Her body forms a breathtakingly graceful arc in mid-air, her arms extended forward, her legs straight, toes pointed, her facial expression focused yet filled with tranquil bliss. Below her, the water is a deep blue, crystal clear, revealing a sandy seabed and smooth underwater rocks. The bright midday sun casts sharp shadows across the cliff face and creates dazzling, crystalline sparkles on the water's surface. She cuts into the water with powerful precision, creating the smallest splash, her body disappearing seamlessly into the blue depths as the camera captures the expanding concentric ripples as the only evidence of her entry, evoking a profound sense of freedom and natural grace.", + "seed": 25839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_8/index_9393_seed_25839_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16777541106566787, + "vlm_logprob": -0.16777541106566787, + "vlm_yes_no_margin": 2.500000011920929, + "vlm_token_credit": -0.16777541106566787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1849265992641449 + ], + "opposite_logprob_tokens": [ + -1.9349266290664673 + ], + "logprob_reward": -0.1849265992641449, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1849265992641449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11661992967128754, + 0.07952821254730225, + 0.06603635102510452, + 0.04564599320292473, + 0.09186588227748871, + 0.10845950990915298, + 0.22878216207027435, + 0.053883787244558334, + 0.04403626173734665, + 0.16514192521572113 + ], + "credit_map": [ + 0.016806023195385933, + 0.008365633897483349, + 0.017267882823944092, + 0.021410666406154633, + 0.0065574427135288715, + 0.023368531838059425, + 0.01212368905544281, + 0.010720065794885159, + 0.008874021470546722, + 0.007711374666541815, + 0.012810460291802883, + 0.013027873821556568, + 0.005848628934472799, + 0.013675060123205185, + 0.006992715410888195, + 0.010588075965642929, + 0.008518467657268047, + 0.0066348351538181305, + 0.007478279992938042, + 0.007144103292375803, + 0.008504007942974567, + 0.009822718799114227, + 0.0076313018798828125, + 0.010302639566361904, + 0.004470347426831722, + 0.0071227154694497585, + 0.006045137532055378, + 0.008467349223792553, + 0.0047821225598454475, + 0.0037836411502212286, + 0.00623326376080513, + 0.004741416312754154, + 0.010627832263708115, + 0.010170849040150642, + 0.013294808566570282, + 0.011984300799667835, + 0.012794820591807365, + 0.013292862102389336, + 0.008606558665633202, + 0.011093849316239357, + 0.013122744858264923, + 0.013790199533104897, + 0.017435722053050995, + 0.011596591211855412, + 0.010005517862737179, + 0.0208341833204031, + 0.009980255737900734, + 0.011694290675222874, + 0.037654466927051544, + 0.025026721879839897, + 0.03309990093111992, + 0.03492088243365288, + 0.026856686919927597, + 0.028549088165163994, + 0.022595005109906197, + 0.020079411566257477, + 0.011149346828460693, + 0.00433616666123271, + 0.007382133509963751, + 0.004932679235935211, + 0.008259925991296768, + 0.007817359641194344, + 0.004299018066376448, + 0.005707158707082272, + 0.008291663601994514, + 0.004558367654681206, + 0.005584354046732187, + 0.003633273998275399, + 0.005441965069621801, + 0.007121461443603039, + 0.004477133974432945, + 0.004928039852529764, + 0.028672080487012863, + 0.012027623131871223, + 0.020714322105050087, + 0.01574581116437912, + 0.01661122590303421, + 0.017428142949938774, + 0.035503026098012924, + 0.01843969151377678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.1849265992641449 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289429783821106 + ], + "opposite_logprob_tokens": [ + -1.0539429187774658 + ], + "logprob_reward": -0.4289429783821106, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4289429783821106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10490643978118896, + 0.040492765605449677, + 0.058419644832611084, + 0.05187255144119263, + 0.06406023353338242, + 0.027813512831926346, + 0.08567846566438675, + 0.13592155277729034, + 0.25008293986320496, + 0.1807519644498825 + ], + "credit_map": [ + 0.015252752229571342, + 0.010799918323755264, + 0.012213315814733505, + 0.020624661818146706, + 0.007345104124397039, + 0.01922157034277916, + 0.011634109541773796, + 0.00781499594449997, + 0.005526458378881216, + 0.005116995424032211, + 0.005949929356575012, + 0.005522127263247967, + 0.004231755156069994, + 0.0064753456972539425, + 0.0034882251638919115, + 0.004181926604360342, + 0.008678419515490532, + 0.003764910390600562, + 0.006931847892701626, + 0.008119731210172176, + 0.004999210126698017, + 0.008657040074467659, + 0.00717927236109972, + 0.010089208371937275, + 0.005061623174697161, + 0.005228778813034296, + 0.006638471037149429, + 0.006865646690130234, + 0.006890262011438608, + 0.005369305144995451, + 0.00901784561574459, + 0.006800611037760973, + 0.008230569772422314, + 0.008520320057868958, + 0.008653334341943264, + 0.009807049296796322, + 0.007419131696224213, + 0.009782982058823109, + 0.005844918079674244, + 0.0058019207790493965, + 0.004094463773071766, + 0.003673675237223506, + 0.0041110017336905, + 0.0029535077046602964, + 0.0029709767550230026, + 0.004694024566560984, + 0.0027902473229914904, + 0.0025256152730435133, + 0.00936177745461464, + 0.013782917521893978, + 0.013383449986577034, + 0.020388860255479813, + 0.006507724989205599, + 0.008619376458227634, + 0.00693170353770256, + 0.006702644284814596, + 0.018977327272295952, + 0.023941339924931526, + 0.014254387468099594, + 0.011481855064630508, + 0.02149953506886959, + 0.023048439994454384, + 0.008723347447812557, + 0.013995318673551083, + 0.05998212471604347, + 0.025778725743293762, + 0.03511427342891693, + 0.018972555175423622, + 0.03135526552796364, + 0.03110630251467228, + 0.023015040904283524, + 0.02475861646234989, + 0.0352807492017746, + 0.019699327647686005, + 0.02265748381614685, + 0.02014658972620964, + 0.018023531883955002, + 0.022315237671136856, + 0.018962671980261803, + 0.023666365072131157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.4289429783821106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282742142677307 + ], + "opposite_logprob_tokens": [ + -2.0178275108337402 + ], + "logprob_reward": -0.14282742142677307, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14282742142677307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2543402910232544, + 0.2160346657037735, + 0.24619387090206146, + 0.052252814173698425, + 0.0344662144780159, + 0.018748991191387177, + 0.04762858524918556, + 0.02566915936768055, + 0.04161807522177696, + 0.063047394156456 + ], + "credit_map": [ + 0.0418827086687088, + 0.027191225439310074, + 0.024049749597907066, + 0.03877869248390198, + 0.04031438007950783, + 0.047934770584106445, + 0.019944749772548676, + 0.01424399483948946, + 0.02714071422815323, + 0.02615409344434738, + 0.028907915577292442, + 0.030853968113660812, + 0.02889280952513218, + 0.0305792223662138, + 0.020306961610913277, + 0.02319897897541523, + 0.03673679009079933, + 0.021629031747579575, + 0.033369094133377075, + 0.03616050258278847, + 0.025186916813254356, + 0.03400159254670143, + 0.027470272034406662, + 0.03163965046405792, + 0.006330770906060934, + 0.007263060659170151, + 0.006766910199075937, + 0.007277045864611864, + 0.004070641938596964, + 0.007603278383612633, + 0.006641303654760122, + 0.006299798842519522, + 0.003541169688105583, + 0.004279476124793291, + 0.00452596927061677, + 0.004327916074544191, + 0.004963053856045008, + 0.005165503825992346, + 0.00376327708363533, + 0.003899846924468875, + 0.002183042699471116, + 0.00226716510951519, + 0.0032828757539391518, + 0.0024923409800976515, + 0.0024455625098198652, + 0.0026170199271291494, + 0.0016772457165643573, + 0.0017837374471127987, + 0.0059719267301261425, + 0.007384414318948984, + 0.0058690644800662994, + 0.005624215118587017, + 0.004714123904705048, + 0.00910011027008295, + 0.0036979939322918653, + 0.005266730207949877, + 0.005883056670427322, + 0.002321127336472273, + 0.004228326492011547, + 0.002147431019693613, + 0.0027221052441745996, + 0.003674593288451433, + 0.0017275160644203424, + 0.002965001855045557, + 0.010524094104766846, + 0.004362063482403755, + 0.005911749321967363, + 0.002775351284071803, + 0.004745637532323599, + 0.005752915982156992, + 0.0029708663932979107, + 0.0045753964222967625, + 0.01174116786569357, + 0.00558802206069231, + 0.007500606123358011, + 0.005747429560869932, + 0.007044055033475161, + 0.0075058359652757645, + 0.010488785803318024, + 0.007431487087160349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.14282742142677307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319446949288249 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.00319446949288249, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319446949288249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11444129049777985, + 0.10107968747615814, + 0.06458666920661926, + 0.04137744754552841, + 0.047545723617076874, + 0.16404907405376434, + 0.050305336713790894, + 0.10225656628608704, + 0.19405628740787506, + 0.12030202150344849 + ], + "credit_map": [ + 0.018879536539316177, + 0.012799479998648167, + 0.016270119696855545, + 0.01574617438018322, + 0.014867426827549934, + 0.017847949638962746, + 0.009467445313930511, + 0.008563154377043247, + 0.011905225925147533, + 0.01186063326895237, + 0.017346803098917007, + 0.013625518418848515, + 0.014544651843607426, + 0.014639589935541153, + 0.009232169948518276, + 0.007925083860754967, + 0.008158481679856777, + 0.005978452041745186, + 0.009454270824790001, + 0.0070934537798166275, + 0.006928407587110996, + 0.010936728678643703, + 0.007518371567130089, + 0.008518507704138756, + 0.004463937133550644, + 0.007508257869631052, + 0.00563427247107029, + 0.004869087133556604, + 0.003770965849980712, + 0.006125309504568577, + 0.004907750058919191, + 0.004097865428775549, + 0.005101528950035572, + 0.005331672262400389, + 0.00794288795441389, + 0.007001671940088272, + 0.00377931771799922, + 0.006608691532164812, + 0.0062643964774906635, + 0.005515558645129204, + 0.017078114673495293, + 0.019979003816843033, + 0.02418123558163643, + 0.018224190920591354, + 0.018606504425406456, + 0.029964426532387733, + 0.01840154454112053, + 0.017614055424928665, + 0.005153443664312363, + 0.005806810688227415, + 0.006779421120882034, + 0.007606771308928728, + 0.004716694355010986, + 0.008775156922638416, + 0.00610846932977438, + 0.005358566530048847, + 0.016894754022359848, + 0.01647166535258293, + 0.013392281718552113, + 0.009034302085638046, + 0.012546845711767673, + 0.01473095454275608, + 0.007208893075585365, + 0.011976868845522404, + 0.04550950601696968, + 0.019099673256278038, + 0.026401901617646217, + 0.015510412864387035, + 0.02260398119688034, + 0.027997583150863647, + 0.01746387965977192, + 0.019469337537884712, + 0.030265990644693375, + 0.01378424372524023, + 0.017663532868027687, + 0.00858396477997303, + 0.017065569758415222, + 0.015404252335429192, + 0.009647132828831673, + 0.007887329906225204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.00319446949288249 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898558676242828 + ], + "opposite_logprob_tokens": [ + -2.5789856910705566 + ], + "logprob_reward": -0.07898558676242828, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898558676242828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08794952183961868, + 0.05279995873570442, + 0.0782424658536911, + 0.06122313812375069, + 0.07252012193202972, + 0.05089332163333893, + 0.24222545325756073, + 0.274732381105423, + 0.04188929870724678, + 0.03752434626221657 + ], + "credit_map": [ + 0.014754616655409336, + 0.008152575232088566, + 0.008775594644248486, + 0.010613476857542992, + 0.007980924099683762, + 0.020371293649077415, + 0.010208375751972198, + 0.00709265423938632, + 0.006750669330358505, + 0.006476731039583683, + 0.010399430058896542, + 0.0051253014244139194, + 0.00583796389400959, + 0.007136393804103136, + 0.00449196295812726, + 0.006581502966582775, + 0.00870888028293848, + 0.007239209022372961, + 0.01086441334336996, + 0.011584246531128883, + 0.009975479915738106, + 0.010654263198375702, + 0.011240290477871895, + 0.007975677959620953, + 0.006273890845477581, + 0.007560045458376408, + 0.00759612862020731, + 0.008965186774730682, + 0.005874086171388626, + 0.005118571221828461, + 0.00849930290132761, + 0.011335916817188263, + 0.008622633293271065, + 0.010755396448075771, + 0.01107257790863514, + 0.011714224703609943, + 0.0068152835592627525, + 0.009705135598778725, + 0.007789380848407745, + 0.0060454788617789745, + 0.006112835835665464, + 0.00671769306063652, + 0.006843727082014084, + 0.005158908665180206, + 0.005828114226460457, + 0.008607031777501106, + 0.006493912544101477, + 0.005131090059876442, + 0.022198056802153587, + 0.028011998161673546, + 0.037535618990659714, + 0.06416954100131989, + 0.021397864446043968, + 0.029536258429288864, + 0.017548490315675735, + 0.02182759903371334, + 0.043203361332416534, + 0.04202037304639816, + 0.031532879918813705, + 0.02267727255821228, + 0.04690809175372124, + 0.0427885428071022, + 0.017551708966493607, + 0.028050146996974945, + 0.010629907250404358, + 0.004118835553526878, + 0.006287888623774052, + 0.0026816176250576973, + 0.004853110760450363, + 0.005751892924308777, + 0.0035515448544174433, + 0.004014496225863695, + 0.008355572819709778, + 0.003146182978525758, + 0.004992751404643059, + 0.0034119808115065098, + 0.003728939685970545, + 0.004595599137246609, + 0.005500626750290394, + 0.0037926919758319855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.07898558676242828 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, low-angle tracking shot follows a woman as she makes a perfect dive. The camera first focuses on her bare feet gripping the edge of a rugged, sun-drenched granite cliff, then tilts upward to capture her entire body against a bright, cloudless sky. She is a woman with an athletic, toned figure, her long, dark hair pulled back in a sleek ponytail that scatters momentarily as she thrusts forward. She wears a minimalist black one-piece swimsuit made of matte-finished, form-fitting material. Her body forms a breathtakingly graceful arc in mid-air, her arms extended forward, her legs straight, toes pointed, her facial expression focused yet filled with tranquil bliss. Below her, the water is a deep blue, crystal clear, revealing a sandy seabed and smooth underwater rocks. The bright midday sun casts sharp shadows across the cliff face and creates dazzling, crystalline sparkles on the water's surface. She cuts into the water with powerful precision, creating the smallest splash, her body disappearing seamlessly into the blue depths as the camera captures the expanding concentric ripples as the only evidence of her entry, evoking a profound sense of freedom and natural grace.", + "seed": 25840, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_8/index_9393_seed_25840_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1636479366570711, + "vlm_logprob": -0.1636479366570711, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.1636479366570711, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22042685747146606 + ], + "opposite_logprob_tokens": [ + -1.7204267978668213 + ], + "logprob_reward": -0.22042685747146606, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22042685747146606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07910290360450745, + 0.06791529804468155, + 0.09744276106357574, + 0.2529338300228119, + 0.08063273876905441, + 0.037853777408599854, + 0.1839456558227539, + 0.06675662100315094, + 0.033749405294656754, + 0.09966699033975601 + ], + "credit_map": [ + 0.009456761181354523, + 0.005856677424162626, + 0.015668774023652077, + 0.011008274741470814, + 0.01229312364012003, + 0.009194006212055683, + 0.0077565633691847324, + 0.007868729531764984, + 0.005358715541660786, + 0.011350260116159916, + 0.005171526689082384, + 0.007154328748583794, + 0.01302370149642229, + 0.009809696115553379, + 0.005572110414505005, + 0.01047496311366558, + 0.006168682128190994, + 0.010755524970591068, + 0.0061399079859256744, + 0.01278097927570343, + 0.018055979162454605, + 0.014377005398273468, + 0.019242070615291595, + 0.009922616183757782, + 0.02534395642578602, + 0.023220445960760117, + 0.039503246545791626, + 0.021492650732398033, + 0.06343083828687668, + 0.03754360228776932, + 0.02542315423488617, + 0.016975946724414825, + 0.010004577226936817, + 0.008549779653549194, + 0.015349182300269604, + 0.010744282975792885, + 0.007015800103545189, + 0.009695653803646564, + 0.006333605386316776, + 0.012939867563545704, + 0.00515866931527853, + 0.003672437509521842, + 0.005703180097043514, + 0.004847624804824591, + 0.005230383947491646, + 0.004677025601267815, + 0.0038131920155137777, + 0.004751264583319426, + 0.02821681834757328, + 0.036055151373147964, + 0.02080811932682991, + 0.02055276371538639, + 0.018646782264113426, + 0.023028220981359482, + 0.020121023058891296, + 0.01651678793132305, + 0.007948273792862892, + 0.00823537353426218, + 0.007365870289504528, + 0.006600665859878063, + 0.010044964030385017, + 0.007411566562950611, + 0.011523813009262085, + 0.007626097183674574, + 0.003615780733525753, + 0.003731085453182459, + 0.0038105861749500036, + 0.0053308503702282906, + 0.005561542697250843, + 0.0034495731815695763, + 0.004958157427608967, + 0.0032918292563408613, + 0.007850776426494122, + 0.011377295479178429, + 0.016418594866991043, + 0.01514382753521204, + 0.01834123767912388, + 0.007772805169224739, + 0.012622205540537834, + 0.010140255093574524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.22042685747146606 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035507380962372 + ], + "opposite_logprob_tokens": [ + -2.3503551483154297 + ], + "logprob_reward": -0.10035507380962372, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035507380962372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1517341583967209, + 0.05812181532382965, + 0.23099565505981445, + 0.09475571662187576, + 0.1610923707485199, + 0.03938901796936989, + 0.10634206980466843, + 0.04749276116490364, + 0.05226387456059456, + 0.05781259015202522 + ], + "credit_map": [ + 0.014093583449721336, + 0.015623554587364197, + 0.027445673942565918, + 0.02188117243349552, + 0.028144560754299164, + 0.01966998167335987, + 0.013587389141321182, + 0.011288240551948547, + 0.004974976181983948, + 0.00844657514244318, + 0.004409403074532747, + 0.006217425223439932, + 0.014733363874256611, + 0.009602222591638565, + 0.0040113842114806175, + 0.005726462230086327, + 0.025215838104486465, + 0.030552342534065247, + 0.018073763698339462, + 0.027922941371798515, + 0.045321203768253326, + 0.03050164505839348, + 0.035801712423563004, + 0.01760619692504406, + 0.011101371608674526, + 0.006807791069149971, + 0.020110150799155235, + 0.011197339743375778, + 0.018524501472711563, + 0.01236883644014597, + 0.008776290342211723, + 0.005869434215128422, + 0.025855470448732376, + 0.021044893190264702, + 0.0299588181078434, + 0.028598356992006302, + 0.011657708324491978, + 0.018050141632556915, + 0.010211898013949394, + 0.01571507565677166, + 0.005117480177432299, + 0.005087837111204863, + 0.005809434689581394, + 0.005587691441178322, + 0.005029987543821335, + 0.004231695551425219, + 0.004531234037131071, + 0.003993653692305088, + 0.014589792117476463, + 0.01859450526535511, + 0.012459738180041313, + 0.011820570565760136, + 0.010626672767102718, + 0.011969100683927536, + 0.01515170093625784, + 0.01112997904419899, + 0.006042545195668936, + 0.004671350121498108, + 0.004904452245682478, + 0.004742531105875969, + 0.007363487966358662, + 0.005046885460615158, + 0.008455167524516582, + 0.006266339216381311, + 0.00476350961253047, + 0.006488358601927757, + 0.005316983442753553, + 0.008613918907940388, + 0.008965110406279564, + 0.005310719832777977, + 0.007358300965279341, + 0.00544696906581521, + 0.005428689066320658, + 0.006576462648808956, + 0.011961997486650944, + 0.0055443234741687775, + 0.00813073106110096, + 0.007270058151334524, + 0.006253190338611603, + 0.006647136528044939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.10035507380962372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134585916996002 + ], + "opposite_logprob_tokens": [ + -1.3134585618972778 + ], + "logprob_reward": -0.3134585916996002, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134585916996002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08738268911838531, + 0.3091806173324585, + 0.08662483841180801, + 0.11625053733587265, + 0.05485062673687935, + 0.03810947760939598, + 0.17507757246494293, + 0.0594705305993557, + 0.03835614025592804, + 0.034696899354457855 + ], + "credit_map": [ + 0.008620882406830788, + 0.009180191904306412, + 0.008876608684659004, + 0.00881102867424488, + 0.019615840166807175, + 0.012444115243852139, + 0.01185852475464344, + 0.00797549169510603, + 0.01424971129745245, + 0.058599889278411865, + 0.009757940657436848, + 0.01931021176278591, + 0.1070641279220581, + 0.06654328852891922, + 0.018100637942552567, + 0.01555480994284153, + 0.01016361452639103, + 0.00948265939950943, + 0.009159424342215061, + 0.009213734418153763, + 0.013373155146837234, + 0.010163364931941032, + 0.012458903715014458, + 0.012609985657036304, + 0.014332079328596592, + 0.010542559437453747, + 0.017395013943314552, + 0.008265060372650623, + 0.03161948174238205, + 0.016922147944569588, + 0.010803643614053726, + 0.006370555143803358, + 0.006108527537435293, + 0.0052109709940850735, + 0.008494231849908829, + 0.008383491076529026, + 0.004370399750769138, + 0.007600621785968542, + 0.00405088858678937, + 0.010631495155394077, + 0.006668000482022762, + 0.0051050372421741486, + 0.00780617818236351, + 0.005251970142126083, + 0.0036247707903385162, + 0.003996577113866806, + 0.0020233013201504946, + 0.003633643966168165, + 0.0308461245149374, + 0.024244925007224083, + 0.021647142246365547, + 0.02304467372596264, + 0.013171358034014702, + 0.016570432111620903, + 0.022755127400159836, + 0.02279779128730297, + 0.0067827715538442135, + 0.006663585547357798, + 0.005476613063365221, + 0.0052964286878705025, + 0.01287926360964775, + 0.007995004765689373, + 0.008372661657631397, + 0.006004201248288155, + 0.0033667348325252533, + 0.003982336260378361, + 0.003991920035332441, + 0.010250766761600971, + 0.005904128775000572, + 0.0031664136331528425, + 0.004186924081295729, + 0.0035069147124886513, + 0.0033674822188913822, + 0.0038460760843008757, + 0.004375112242996693, + 0.005824950058013201, + 0.0046572391875088215, + 0.0037375236861407757, + 0.004365803208202124, + 0.004522716626524925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.3134585916996002 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098350182175636 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098350182175636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098350182175636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10234910994768143, + 0.1362094283103943, + 0.06941492855548859, + 0.07589273899793625, + 0.1599142700433731, + 0.15199318528175354, + 0.07283322513103485, + 0.03869624435901642, + 0.024049118161201477, + 0.16864770650863647 + ], + "credit_map": [ + 0.010974314995110035, + 0.012044322676956654, + 0.011121136136353016, + 0.008853121660649776, + 0.026060935109853745, + 0.01629183441400528, + 0.007573507726192474, + 0.009429939091205597, + 0.010537705384194851, + 0.023230571299791336, + 0.00794739555567503, + 0.011326721869409084, + 0.03786816820502281, + 0.024838188663125038, + 0.009063721634447575, + 0.011396951042115688, + 0.007779912557452917, + 0.010762198828160763, + 0.004683909472078085, + 0.00613408163189888, + 0.010360112413764, + 0.008532484993338585, + 0.00911684799939394, + 0.012045379728078842, + 0.008326279930770397, + 0.005721080116927624, + 0.011506333947181702, + 0.00703016109764576, + 0.01599777489900589, + 0.00962577760219574, + 0.008356533944606781, + 0.009328790940344334, + 0.013860083185136318, + 0.012356175109744072, + 0.01761552505195141, + 0.015959663316607475, + 0.012738343328237534, + 0.028829917311668396, + 0.011478478088974953, + 0.04707607999444008, + 0.023662839084863663, + 0.01880502514541149, + 0.03344932571053505, + 0.02236415445804596, + 0.010991722345352173, + 0.016630830243229866, + 0.01093009952455759, + 0.015159185975790024, + 0.012243964709341526, + 0.008988235145807266, + 0.008773105219006538, + 0.00992647185921669, + 0.005690323188900948, + 0.006763158831745386, + 0.009932908229529858, + 0.010515062138438225, + 0.0057662189938127995, + 0.006604636553674936, + 0.004767732694745064, + 0.004036222118884325, + 0.004001623950898647, + 0.004072897136211395, + 0.005281294230371714, + 0.00416561821475625, + 0.0019356992561370134, + 0.0027310315053910017, + 0.0034385661128908396, + 0.003937782719731331, + 0.002752451691776514, + 0.002705218270421028, + 0.003537096781656146, + 0.003011273220181465, + 0.013581743463873863, + 0.019899040460586548, + 0.022811993956565857, + 0.015040243975818157, + 0.035649869590997696, + 0.017686141654849052, + 0.021580101922154427, + 0.0223985742777586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.004098350182175636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990081012248993 + ], + "opposite_logprob_tokens": [ + -1.8049007654190063 + ], + "logprob_reward": -0.17990081012248993, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990081012248993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08445482701063156, + 0.04980878531932831, + 0.19144193828105927, + 0.07209078967571259, + 0.3688620328903198, + 0.049054719507694244, + 0.09851399809122086, + 0.029977142810821533, + 0.013497320003807545, + 0.042298492044210434 + ], + "credit_map": [ + 0.011170335114002228, + 0.0072390614077448845, + 0.015803370624780655, + 0.013947190716862679, + 0.01097750011831522, + 0.008436476811766624, + 0.00855562649667263, + 0.008325262926518917, + 0.00295954174362123, + 0.007050823420286179, + 0.0035795867443084717, + 0.003783626714721322, + 0.013284476473927498, + 0.009009630419313908, + 0.0032769800163805485, + 0.0068641179241240025, + 0.026257134974002838, + 0.031059974804520607, + 0.014464516192674637, + 0.01824062131345272, + 0.02872288227081299, + 0.022848574444651604, + 0.02063666470348835, + 0.029211565852165222, + 0.0062734573148190975, + 0.0061333561316132545, + 0.00923948548734188, + 0.005705990828573704, + 0.019147725775837898, + 0.01107414998114109, + 0.00826549157500267, + 0.006251127924770117, + 0.03286340832710266, + 0.03731260076165199, + 0.025846945121884346, + 0.03481530770659447, + 0.03282880410552025, + 0.06973019242286682, + 0.028703127056360245, + 0.10676164925098419, + 0.0054512303322553635, + 0.004460902418941259, + 0.007252924609929323, + 0.007396993227303028, + 0.007196934428066015, + 0.006191886030137539, + 0.005246160086244345, + 0.005857688840478659, + 0.01365458220243454, + 0.011964940465986729, + 0.012695839628577232, + 0.011467383243143559, + 0.0126179875805974, + 0.010484871454536915, + 0.014094337821006775, + 0.011534054763615131, + 0.0024988059885799885, + 0.0036227507516741753, + 0.004196511581540108, + 0.00309520261362195, + 0.0044555338099598885, + 0.0031536477617919445, + 0.005532995332032442, + 0.003421694738790393, + 0.0014288454549387097, + 0.0012938275467604399, + 0.0013878935715183616, + 0.00249314378015697, + 0.0018169619143009186, + 0.0015076869167387486, + 0.0021338253282010555, + 0.0014351359568536282, + 0.004531556740403175, + 0.005602825433015823, + 0.007083578500896692, + 0.004369097761809826, + 0.00606479262933135, + 0.005604590754956007, + 0.003670367179438472, + 0.005371682345867157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.17990081012248993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, low-angle tracking shot follows a woman as she makes a perfect dive. The camera first focuses on her bare feet gripping the edge of a rugged, sun-drenched granite cliff, then tilts upward to capture her entire body against a bright, cloudless sky. She is a woman with an athletic, toned figure, her long, dark hair pulled back in a sleek ponytail that scatters momentarily as she thrusts forward. She wears a minimalist black one-piece swimsuit made of matte-finished, form-fitting material. Her body forms a breathtakingly graceful arc in mid-air, her arms extended forward, her legs straight, toes pointed, her facial expression focused yet filled with tranquil bliss. Below her, the water is a deep blue, crystal clear, revealing a sandy seabed and smooth underwater rocks. The bright midday sun casts sharp shadows across the cliff face and creates dazzling, crystalline sparkles on the water's surface. She cuts into the water with powerful precision, creating the smallest splash, her body disappearing seamlessly into the blue depths as the camera captures the expanding concentric ripples as the only evidence of her entry, evoking a profound sense of freedom and natural grace.", + "seed": 25841, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000054/group_8/index_9393_seed_25841_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17830553194507956, + "vlm_logprob": -0.17830553194507956, + "vlm_yes_no_margin": 2.400000011920929, + "vlm_token_credit": -0.17830553194507956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24376346170902252 + ], + "opposite_logprob_tokens": [ + -1.6187634468078613 + ], + "logprob_reward": -0.24376346170902252, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24376346170902252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09868092834949493, + 0.02403092570602894, + 0.0945647582411766, + 0.10113062709569931, + 0.06677175313234329, + 0.09911919385194778, + 0.16275493800640106, + 0.05398820713162422, + 0.23293234407901764, + 0.06602635234594345 + ], + "credit_map": [ + 0.012031651102006435, + 0.01957327499985695, + 0.00900101289153099, + 0.011419541202485561, + 0.015447895042598248, + 0.01610863208770752, + 0.00805931817740202, + 0.0070395995862782, + 0.002426193794235587, + 0.0023009495344012976, + 0.002221792470663786, + 0.0025907910894602537, + 0.0025696088559925556, + 0.003606033744290471, + 0.004470208194106817, + 0.0038453475572168827, + 0.027042442932724953, + 0.009358217939734459, + 0.010749312117695808, + 0.006170057225972414, + 0.006541335489600897, + 0.01338696014136076, + 0.007938943803310394, + 0.013377496972680092, + 0.015169783495366573, + 0.011914114467799664, + 0.008031179197132587, + 0.007133665028959513, + 0.009256483986973763, + 0.02745569497346878, + 0.00819903053343296, + 0.013970676809549332, + 0.0159800685942173, + 0.0054679433815181255, + 0.0074692415073513985, + 0.00580518739297986, + 0.005107685923576355, + 0.015832282602787018, + 0.004849638324230909, + 0.006259709596633911, + 0.013430414721369743, + 0.011045876890420914, + 0.010863742791116238, + 0.016143901273608208, + 0.0072097089141607285, + 0.006991070229560137, + 0.011533444747328758, + 0.021901030093431473, + 0.023450247943401337, + 0.019511012360453606, + 0.01821533590555191, + 0.01993941329419613, + 0.01928427442908287, + 0.0133738461881876, + 0.021803325042128563, + 0.027177486568689346, + 0.0070414356887340546, + 0.007443841081112623, + 0.009326201863586903, + 0.00537654384970665, + 0.006186668295413256, + 0.005172299686819315, + 0.006541653536260128, + 0.006899564992636442, + 0.06444211304187775, + 0.03856075182557106, + 0.021166525781154633, + 0.024649405851960182, + 0.026555372402071953, + 0.023365464061498642, + 0.014513044618070126, + 0.01967964507639408, + 0.006845104042440653, + 0.006103390362113714, + 0.0077088093385100365, + 0.010391865856945515, + 0.004720388446003199, + 0.008008088916540146, + 0.013774760067462921, + 0.008473951369524002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.24376346170902252 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710230588912964 + ], + "opposite_logprob_tokens": [ + -1.1371023654937744 + ], + "logprob_reward": -0.38710230588912964, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38710230588912964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2123708575963974, + 0.045551493763923645, + 0.09522166848182678, + 0.04143235832452774, + 0.06336185336112976, + 0.0773840844631195, + 0.2029993236064911, + 0.04822816327214241, + 0.097293421626091, + 0.11615672707557678 + ], + "credit_map": [ + 0.02523837797343731, + 0.04201146960258484, + 0.018456678837537766, + 0.02883816324174404, + 0.030107542872428894, + 0.032213982194662094, + 0.01667325757443905, + 0.018831374123692513, + 0.00320196570828557, + 0.003589747240766883, + 0.005482174921780825, + 0.004721667617559433, + 0.008095696568489075, + 0.010290364734828472, + 0.005074978340417147, + 0.0050949035212397575, + 0.01839805394411087, + 0.012062357738614082, + 0.01051478460431099, + 0.010381481610238552, + 0.011568774469196796, + 0.018004650250077248, + 0.006996698211878538, + 0.007294867653399706, + 0.005405164323747158, + 0.003171604359522462, + 0.003891551401466131, + 0.0035555530339479446, + 0.005300068762153387, + 0.010480066761374474, + 0.003949856385588646, + 0.005678489804267883, + 0.009410264901816845, + 0.007907758466899395, + 0.008758455514907837, + 0.007362270262092352, + 0.006272256374359131, + 0.007600170094519854, + 0.005290402099490166, + 0.01076028123497963, + 0.010383383370935917, + 0.008803279139101505, + 0.00682586245238781, + 0.012352274730801582, + 0.006535266060382128, + 0.005805797874927521, + 0.01155752595514059, + 0.015120702795684338, + 0.019764643162488937, + 0.024899838492274284, + 0.03227371722459793, + 0.03366372734308243, + 0.014249909669160843, + 0.01549617201089859, + 0.03021734394133091, + 0.032433975487947464, + 0.00770212709903717, + 0.007601301651448011, + 0.009281900711357594, + 0.004707204643636942, + 0.004825700540095568, + 0.0044015985913574696, + 0.0053495862521231174, + 0.004358744714409113, + 0.021916115656495094, + 0.016432510688900948, + 0.00966797024011612, + 0.00974776316434145, + 0.015013624913990498, + 0.011057284660637379, + 0.00656864931806922, + 0.006889509968459606, + 0.01441839151084423, + 0.013234959915280342, + 0.013476952910423279, + 0.014388256706297398, + 0.010682795196771622, + 0.01501433365046978, + 0.023495156317949295, + 0.011445875279605389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.38710230588912964 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.201592355966568 + ], + "opposite_logprob_tokens": [ + -1.7015923261642456 + ], + "logprob_reward": -0.201592355966568, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.201592355966568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08630572259426117, + 0.10123130679130554, + 0.27305781841278076, + 0.0934898778796196, + 0.015227832831442356, + 0.05731010437011719, + 0.10579650849103928, + 0.06328520178794861, + 0.07608397305011749, + 0.12821164727210999 + ], + "credit_map": [ + 0.010450063273310661, + 0.01358965877443552, + 0.006651210132986307, + 0.012795371003448963, + 0.013240299187600613, + 0.011936214752495289, + 0.010244764387607574, + 0.007398141082376242, + 0.005124050658196211, + 0.004535018466413021, + 0.016965070739388466, + 0.008666165173053741, + 0.0218809787184, + 0.03088424541056156, + 0.007307432126253843, + 0.005868346430361271, + 0.0748751312494278, + 0.024548383429646492, + 0.02514411136507988, + 0.0176750086247921, + 0.014769777655601501, + 0.02946733869612217, + 0.02848803624510765, + 0.05809001997113228, + 0.014575416222214699, + 0.01190368551760912, + 0.007978416979312897, + 0.006304496433585882, + 0.009512504562735558, + 0.025445129722356796, + 0.006956029217690229, + 0.010814204812049866, + 0.001967797754332423, + 0.0016755019314587116, + 0.0020772984717041254, + 0.001739409170113504, + 0.0016939154593273997, + 0.0012989292154088616, + 0.0015982798067852855, + 0.0031767014879733324, + 0.007668404374271631, + 0.005882542580366135, + 0.005232525523751974, + 0.008642288856208324, + 0.004312516655772924, + 0.003987838979810476, + 0.007563923019915819, + 0.014020062051713467, + 0.011124933138489723, + 0.014359786175191402, + 0.01511856634169817, + 0.016771260648965836, + 0.010692567564547062, + 0.008262158371508121, + 0.015252329409122467, + 0.014214902184903622, + 0.007492630276829004, + 0.008404551073908806, + 0.011191749013960361, + 0.005189289338886738, + 0.006771006155759096, + 0.0063154506497085094, + 0.007172259967774153, + 0.01074826531112194, + 0.019974781200289726, + 0.012618937529623508, + 0.007247054949402809, + 0.008917241357266903, + 0.00743223587051034, + 0.0076765138655900955, + 0.006058958359062672, + 0.00615825317800045, + 0.012877840548753738, + 0.010502224788069725, + 0.015733564272522926, + 0.02667170949280262, + 0.007698576431721449, + 0.015137534588575363, + 0.023839985951781273, + 0.015750207006931305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.201592355966568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040998938493430614 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.0040998938493430614, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040998938493430614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24926449358463287, + 0.07689660787582397, + 0.25725674629211426, + 0.04300883412361145, + 0.03981849551200867, + 0.09414597600698471, + 0.09298606961965561, + 0.043963510543107986, + 0.05505465716123581, + 0.04760462045669556 + ], + "credit_map": [ + 0.03928888961672783, + 0.041713252663612366, + 0.02206335961818695, + 0.02922523207962513, + 0.040680840611457825, + 0.03833417594432831, + 0.02036810666322708, + 0.017590640112757683, + 0.006899597123265266, + 0.004694963805377483, + 0.005512841045856476, + 0.005415475461632013, + 0.017376983538269997, + 0.024065932258963585, + 0.006851501762866974, + 0.006079310551285744, + 0.07077597081661224, + 0.023275410756468773, + 0.024801820516586304, + 0.01739359460771084, + 0.013590197078883648, + 0.02653099037706852, + 0.026973305270075798, + 0.05391545593738556, + 0.005902387667447329, + 0.0030613497365266085, + 0.002974965376779437, + 0.0026983909774571657, + 0.005964298266917467, + 0.01107650063931942, + 0.004434725735336542, + 0.006896217819303274, + 0.005473379977047443, + 0.0034569636918604374, + 0.004610336851328611, + 0.004458867013454437, + 0.003419260261580348, + 0.005142649635672569, + 0.003907397389411926, + 0.009349639527499676, + 0.015319167636334896, + 0.010812326334416866, + 0.010690523311495781, + 0.012817265465855598, + 0.006152670364826918, + 0.00915527157485485, + 0.009939475916326046, + 0.01925928145647049, + 0.010052408091723919, + 0.010615987703204155, + 0.008753204718232155, + 0.013568093068897724, + 0.009039413183927536, + 0.007605373859405518, + 0.012834856286644936, + 0.020516730844974518, + 0.005314374342560768, + 0.006388705689460039, + 0.0056981416419148445, + 0.0039380993694067, + 0.005641316529363394, + 0.00455454783514142, + 0.005803511943668127, + 0.006624809466302395, + 0.012059753760695457, + 0.00999067910015583, + 0.008470101282000542, + 0.007806526031345129, + 0.004626104608178139, + 0.004404779523611069, + 0.003194997785612941, + 0.00450171809643507, + 0.004521442111581564, + 0.0038630468770861626, + 0.005421350244432688, + 0.008565722964704037, + 0.0028869493398815393, + 0.006478060968220234, + 0.008394950069487095, + 0.00747309997677803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0040998938493430614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, looking up at the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a minimalist black one-piece swimsuit made of matte-finished material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a rugged, sun-drenched granite cliff with a deep blue, crystal clear water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward from the woman's feet to capture her entire body against a bright, cloudless sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496964231133461 + ], + "opposite_logprob_tokens": [ + -2.929969549179077 + ], + "logprob_reward": -0.05496964231133461, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496964231133461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08444412052631378, + 0.17533175647258759, + 0.12014249712228775, + 0.1496201604604721, + 0.0809197649359703, + 0.24077266454696655, + 0.027123291045427322, + 0.04665936902165413, + 0.045265860855579376, + 0.029720501974225044 + ], + "credit_map": [ + 0.015198408626019955, + 0.011760295368731022, + 0.00900506041944027, + 0.010933883488178253, + 0.012446265667676926, + 0.010260053910315037, + 0.008382759988307953, + 0.006457394454628229, + 0.014343710616230965, + 0.009160644374787807, + 0.04273521527647972, + 0.02021816372871399, + 0.023259369656443596, + 0.040161509066820145, + 0.014926598407328129, + 0.010526545345783234, + 0.03490627557039261, + 0.013497664593160152, + 0.014617258682847023, + 0.0078001804649829865, + 0.01007502619177103, + 0.021984783932566643, + 0.007257342338562012, + 0.010003968141973019, + 0.02252083644270897, + 0.019998446106910706, + 0.011968432925641537, + 0.010404816828668118, + 0.015460298396646976, + 0.04263991117477417, + 0.010046933777630329, + 0.01658049039542675, + 0.0190582312643528, + 0.0069842408411204815, + 0.008936399593949318, + 0.00700893672183156, + 0.006763189099729061, + 0.01644686609506607, + 0.005594596266746521, + 0.010127302259206772, + 0.041084300726652145, + 0.03110458515584469, + 0.01839475706219673, + 0.025157049298286438, + 0.018412021920084953, + 0.02555418759584427, + 0.03131481260061264, + 0.04975093901157379, + 0.0037112680729478598, + 0.003285597078502178, + 0.0031531620770692825, + 0.005400885362178087, + 0.0028673773631453514, + 0.0020251271780580282, + 0.0039046951569616795, + 0.0027751794550567865, + 0.005426532123237848, + 0.008596072904765606, + 0.00923491083085537, + 0.0039741904474794865, + 0.005799789912998676, + 0.0038218831177800894, + 0.004568034783005714, + 0.005237953271716833, + 0.010400080122053623, + 0.005159289110451937, + 0.005556793417781591, + 0.004037776961922646, + 0.005660620518028736, + 0.005237388890236616, + 0.0034904922358691692, + 0.005723423324525356, + 0.003479649079963565, + 0.003010379383340478, + 0.0029925459530204535, + 0.00532437302172184, + 0.0019516805186867714, + 0.0033994477707892656, + 0.005278768949210644, + 0.004283657763153315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.05496964231133461 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_0/index_tensor([7919, 7919])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_0/index_tensor([7919, 7919])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1e42f4da308a7229c0015225433da94afc002974 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_0/index_tensor([7919, 7919])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a man in his 30s standing confidently at the edge of a luxurious swimming pool on a sunny day. He has short, neatly styled brown hair, and his expression is warm and friendly as he speaks directly to the camera. He wears a light blue, short-sleeved polo shirt made of high-quality chinos, paired with khaki shorts. The bright midday sun casts a soft glow on him, highlighting the texture of his shirt. Behind him, the crystal-clear turquoise waters of the lake ripple gently, creating a dance of light reflections on the surface. The pool is surrounded by a clean, light-colored stone terrace. In the background, several empty lounge chairs and a large white umbrella can be seen, surrounded by lush green tropical foliage. The man occasionally gestures with his right hand in a relaxed and open manner, emphasizing his words. His casual posture and direct eye contact create a warm and engaging atmosphere, as if he is sharing a pleasant story or travel tips.", + "seed": 4851, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_0/index_7919_seed_4851_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01701965434476733, + "vlm_logprob": -0.01701965434476733, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.01701965434476733, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03189763054251671 + ], + "opposite_logprob_tokens": [ + -3.78189754486084 + ], + "logprob_reward": -0.03189763054251671, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03189763054251671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3369719088077545, + 0.08677586168050766, + 0.08147819340229034, + 0.10424972325563431, + 0.053040552884340286, + 0.02618803270161152, + 0.16247786581516266, + 0.028514938428997993, + 0.08614572882652283, + 0.034157175570726395 + ], + "credit_map": [ + 0.06775279343128204, + 0.04940353333950043, + 0.031365811824798584, + 0.021879270672798157, + 0.039180055260658264, + 0.06963125616312027, + 0.034320034086704254, + 0.02343914844095707, + 0.009588305838406086, + 0.007796812802553177, + 0.006331321783363819, + 0.008843029849231243, + 0.017937997356057167, + 0.015401269309222698, + 0.013346654362976551, + 0.0075304750353097916, + 0.01271888893097639, + 0.00717634242027998, + 0.007747327443212271, + 0.008816162124276161, + 0.012714626267552376, + 0.010904362425208092, + 0.010430801659822464, + 0.01096968911588192, + 0.02160564623773098, + 0.01695985719561577, + 0.009362749755382538, + 0.006726379040628672, + 0.01187072042375803, + 0.019256487488746643, + 0.01110466755926609, + 0.007363217417150736, + 0.0060750446282327175, + 0.009386792778968811, + 0.003004681318998337, + 0.00795006100088358, + 0.0077956621535122395, + 0.006439973600208759, + 0.007495387922972441, + 0.0048929559998214245, + 0.005453533958643675, + 0.004045986570417881, + 0.0024082884192466736, + 0.0022224735002964735, + 0.0035330599639564753, + 0.003343933494761586, + 0.0028495953883975744, + 0.002331163501366973, + 0.022584501653909683, + 0.027911333367228508, + 0.013135114684700966, + 0.011076425202190876, + 0.02046012319624424, + 0.0368928462266922, + 0.021840309724211693, + 0.008577227592468262, + 0.004688027780503035, + 0.005958321038633585, + 0.002214584266766906, + 0.002791221719235182, + 0.004038233309984207, + 0.0036106067709624767, + 0.003437866223976016, + 0.0017760804621502757, + 0.011653662659227848, + 0.01611790992319584, + 0.007577990647405386, + 0.007650772575289011, + 0.011271832510828972, + 0.01801115833222866, + 0.008474794216454029, + 0.005387614946812391, + 0.0035708602517843246, + 0.004350506234914064, + 0.002680955221876502, + 0.003739279694855213, + 0.00649691466242075, + 0.005051326006650925, + 0.005640894174575806, + 0.00262644374743104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.03189763054251671 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084645055234432 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084645055234432, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084645055234432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5655917525291443, + 0.11528246104717255, + 0.047727979719638824, + 0.04114070162177086, + 0.026858314871788025, + 0.06120813637971878, + 0.05841143801808357, + 0.03281993046402931, + 0.034762706607580185, + 0.016196662560105324 + ], + "credit_map": [ + 0.09725627303123474, + 0.0644155889749527, + 0.044095251709222794, + 0.03886183351278305, + 0.08473888039588928, + 0.10346092283725739, + 0.08479612320661545, + 0.04796681925654411, + 0.01682659052312374, + 0.011453542858362198, + 0.010861331596970558, + 0.010017598047852516, + 0.017790213227272034, + 0.021868810057640076, + 0.01640579104423523, + 0.010058562271296978, + 0.0064341542311012745, + 0.007870190776884556, + 0.0037410391960293055, + 0.0037494818679988384, + 0.008188447915017605, + 0.005921924952417612, + 0.006847270764410496, + 0.004975463729351759, + 0.006562677212059498, + 0.005356034263968468, + 0.0037262204568833113, + 0.003326383652165532, + 0.006589547265321016, + 0.006522492505609989, + 0.005466101691126823, + 0.0035912380553781986, + 0.0024525297340005636, + 0.0029512171167880297, + 0.0019279150292277336, + 0.0034228486474603415, + 0.006074108649045229, + 0.003244388848543167, + 0.004064034670591354, + 0.002721267519518733, + 0.009019373916089535, + 0.009525262750685215, + 0.005776264239102602, + 0.0038067284040153027, + 0.009183469228446484, + 0.01253732293844223, + 0.006764906458556652, + 0.0045948033221066, + 0.009447887539863586, + 0.012125177308917046, + 0.0044283256866037846, + 0.0037441428285092115, + 0.0075539760291576385, + 0.011685903184115887, + 0.00693846121430397, + 0.00248755537904799, + 0.004912399221211672, + 0.007019041571766138, + 0.00228122528642416, + 0.003555865027010441, + 0.00458672596141696, + 0.00421416200697422, + 0.003915118984878063, + 0.0023353896103799343, + 0.0055351248010993, + 0.005960441194474697, + 0.0034309832844883204, + 0.003302911529317498, + 0.004598457366228104, + 0.006999738048762083, + 0.003062954405322671, + 0.0018720882944762707, + 0.0017129514599218965, + 0.0020888803992420435, + 0.0012427405454218388, + 0.0015449662460014224, + 0.0031694252975285053, + 0.002513347892090678, + 0.002404932165518403, + 0.0015194166917353868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.011084645055234432 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4382922053337097, + 0.08102859556674957, + 0.06959961354732513, + 0.04698075354099274, + 0.03933759778738022, + 0.036445870995521545, + 0.14336252212524414, + 0.06664810329675674, + 0.03756169229745865, + 0.04074309393763542 + ], + "credit_map": [ + 0.07398758828639984, + 0.06925775110721588, + 0.02729210816323757, + 0.03684021160006523, + 0.07112600654363632, + 0.05990254506468773, + 0.06141321733593941, + 0.038472723215818405, + 0.015819737687706947, + 0.009899234399199486, + 0.007460435852408409, + 0.0056417156010866165, + 0.010132493451237679, + 0.016187522560358047, + 0.008337413892149925, + 0.007550024427473545, + 0.008648712188005447, + 0.010823845863342285, + 0.005174737889319658, + 0.00493988860398531, + 0.011844596825540066, + 0.011964518576860428, + 0.009258732199668884, + 0.006944573950022459, + 0.009536324068903923, + 0.008845482021570206, + 0.00420353002846241, + 0.00410907668992877, + 0.004310768097639084, + 0.008891762234270573, + 0.004070000257343054, + 0.003013802692294121, + 0.008022284135222435, + 0.004254468251019716, + 0.003151013981550932, + 0.0027731622103601694, + 0.006676923017948866, + 0.005481420550495386, + 0.005854474846273661, + 0.0031238480005413294, + 0.005295149050652981, + 0.00355486455373466, + 0.002918630838394165, + 0.003849043743684888, + 0.007349212188273668, + 0.005465871188789606, + 0.00501818023622036, + 0.002994913375005126, + 0.022022206336259842, + 0.026291437447071075, + 0.01103911641985178, + 0.008920490741729736, + 0.018618451431393623, + 0.03043954074382782, + 0.018398694694042206, + 0.007632554974406958, + 0.009826011024415493, + 0.010845044627785683, + 0.005658281035721302, + 0.010111777111887932, + 0.00798030011355877, + 0.008557881228625774, + 0.007976188324391842, + 0.005692608188837767, + 0.00402084831148386, + 0.006746023893356323, + 0.0028366814367473125, + 0.004345789086073637, + 0.006418551784008741, + 0.005632179789245129, + 0.004818724934011698, + 0.002742885146290064, + 0.006321423687040806, + 0.00871631596237421, + 0.003305989783257246, + 0.0033017778769135475, + 0.006014512851834297, + 0.006420284043997526, + 0.004006185103207827, + 0.002656597178429365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0024867109023034573 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944785661995411 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944785661995411, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944785661995411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34371063113212585, + 0.10831135511398315, + 0.13489596545696259, + 0.03948784992098808, + 0.05389045551419258, + 0.04213729500770569, + 0.14888407289981842, + 0.032721858471632004, + 0.035684939473867416, + 0.060275495052337646 + ], + "credit_map": [ + 0.06695982813835144, + 0.042100295424461365, + 0.030792171135544777, + 0.024226730689406395, + 0.04523707553744316, + 0.06896962225437164, + 0.03773827850818634, + 0.027686668559908867, + 0.0176854208111763, + 0.011018429882824421, + 0.01084079034626484, + 0.008227109909057617, + 0.013321355916559696, + 0.027778245508670807, + 0.010156693868339062, + 0.009283318184316158, + 0.025569146499037743, + 0.014384269714355469, + 0.014160843566060066, + 0.008272198028862476, + 0.015717579051852226, + 0.03221041336655617, + 0.01286719273775816, + 0.011714331805706024, + 0.00632758904248476, + 0.005837648641318083, + 0.003564259270206094, + 0.00381672871299088, + 0.005681019742041826, + 0.006172723602503538, + 0.005443181376904249, + 0.0026447048876434565, + 0.007298268377780914, + 0.0071578677743673325, + 0.004916691221296787, + 0.006368896458297968, + 0.008143458515405655, + 0.0057495529763400555, + 0.009255561977624893, + 0.00500016612932086, + 0.00631778035312891, + 0.004767629783600569, + 0.003320724470540881, + 0.004692036658525467, + 0.008110810071229935, + 0.007155618630349636, + 0.004437987692654133, + 0.003334709908813238, + 0.02280098758637905, + 0.028446251526474953, + 0.011913707479834557, + 0.009768289513885975, + 0.017330147325992584, + 0.03330935165286064, + 0.018433131277561188, + 0.006882213056087494, + 0.005749574862420559, + 0.005632864776998758, + 0.002335763070732355, + 0.0028465730138123035, + 0.0061246477998793125, + 0.0046661184169352055, + 0.0028663338162004948, + 0.0024999857414513826, + 0.004864387214183807, + 0.00597950117662549, + 0.003637694288045168, + 0.0020565094891935587, + 0.004939395934343338, + 0.007573770359158516, + 0.0032643675804138184, + 0.003369319951161742, + 0.010070376098155975, + 0.0134804742410779, + 0.0051305582746863365, + 0.004394607152789831, + 0.007695804815739393, + 0.010548120364546776, + 0.0055992878042161465, + 0.003356277011334896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.005944785661995411 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368449956178665 + ], + "opposite_logprob_tokens": [ + -3.408684492111206 + ], + "logprob_reward": -0.03368449956178665, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368449956178665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22692923247814178, + 0.2410813719034195, + 0.0555577389895916, + 0.06692036986351013, + 0.08237136155366898, + 0.04774656891822815, + 0.05899951234459877, + 0.09460234642028809, + 0.09958463162183762, + 0.026206865906715393 + ], + "credit_map": [ + 0.04201682657003403, + 0.023662934079766273, + 0.019692089408636093, + 0.018532853573560715, + 0.03135261312127113, + 0.032955266535282135, + 0.03625568374991417, + 0.022460969164967537, + 0.0314520038664341, + 0.030119089409708977, + 0.02055756188929081, + 0.017731845378875732, + 0.03187533840537071, + 0.05996127426624298, + 0.030552342534065247, + 0.018831921741366386, + 0.0076083955354988575, + 0.0077631534077227116, + 0.004989915061742067, + 0.005052349530160427, + 0.008672279305756092, + 0.010658190585672855, + 0.004978142213076353, + 0.005835312884300947, + 0.010659875348210335, + 0.012709886766970158, + 0.00499684177339077, + 0.007120952941477299, + 0.00870397500693798, + 0.009572720155119896, + 0.00691744452342391, + 0.006238670088350773, + 0.012062805704772472, + 0.007416720036417246, + 0.006399966310709715, + 0.00999035406857729, + 0.013814559206366539, + 0.01236580591648817, + 0.013147715479135513, + 0.007173440419137478, + 0.00932353362441063, + 0.004707288928329945, + 0.003610604675486684, + 0.0037837265990674496, + 0.008657214231789112, + 0.007671460974961519, + 0.005739376414567232, + 0.004253365099430084, + 0.009102827869355679, + 0.009990954771637917, + 0.0046707093715667725, + 0.0046294499188661575, + 0.011366555467247963, + 0.008452408947050571, + 0.0074950940907001495, + 0.0032915091142058372, + 0.015400118194520473, + 0.012464744038879871, + 0.008123607374727726, + 0.010213701985776424, + 0.014387140050530434, + 0.013148998841643333, + 0.011510482057929039, + 0.00935355480760336, + 0.011304900981485844, + 0.010048105381429195, + 0.007574131712317467, + 0.01490646693855524, + 0.01633520796895027, + 0.01777365617454052, + 0.013834862038493156, + 0.007807302288711071, + 0.001923613715916872, + 0.0028612418100237846, + 0.0017068602610379457, + 0.003200077684596181, + 0.005375419743359089, + 0.004489797167479992, + 0.004794935695827007, + 0.0018549192463979125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.03368449956178665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a man in his 30s standing confidently at the edge of a luxurious swimming pool on a sunny day. He has short, neatly styled brown hair, and his expression is warm and friendly as he speaks directly to the camera. He wears a light blue, short-sleeved polo shirt made of high-quality chinos, paired with khaki shorts. The bright midday sun casts a soft glow on him, highlighting the texture of his shirt. Behind him, the crystal-clear turquoise waters of the lake ripple gently, creating a dance of light reflections on the surface. The pool is surrounded by a clean, light-colored stone terrace. In the background, several empty lounge chairs and a large white umbrella can be seen, surrounded by lush green tropical foliage. The man occasionally gestures with his right hand in a relaxed and open manner, emphasizing his words. His casual posture and direct eye contact create a warm and engaging atmosphere, as if he is sharing a pleasant story or travel tips.", + "seed": 4852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_0/index_7919_seed_4852_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01976093100383878, + "vlm_logprob": -0.01976093100383878, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.01976093100383878, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042336463928222656 + ], + "opposite_logprob_tokens": [ + -3.6673364639282227 + ], + "logprob_reward": -0.042336463928222656, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.042336463928222656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2871583104133606, + 0.07105876505374908, + 0.08753505349159241, + 0.09862086921930313, + 0.051647648215293884, + 0.046765852719545364, + 0.1612643301486969, + 0.11605735868215561, + 0.0386221669614315, + 0.04126960411667824 + ], + "credit_map": [ + 0.021405089646577835, + 0.025004401803016663, + 0.056577276438474655, + 0.04479772225022316, + 0.04688806086778641, + 0.03145967423915863, + 0.030683843418955803, + 0.030342239886522293, + 0.007100722286850214, + 0.006991209462285042, + 0.010625012218952179, + 0.01197438407689333, + 0.01172694656997919, + 0.009608330205082893, + 0.007789527997374535, + 0.005242628511041403, + 0.0064866505563259125, + 0.006609023083001375, + 0.02124396711587906, + 0.014771108515560627, + 0.015550447627902031, + 0.008590135723352432, + 0.0077926707454025745, + 0.006491050124168396, + 0.008866733871400356, + 0.008059765212237835, + 0.014239990152418613, + 0.018289171159267426, + 0.018986349925398827, + 0.007665485143661499, + 0.013567553833127022, + 0.008945824578404427, + 0.005844777915626764, + 0.005186283495277166, + 0.009464734233915806, + 0.007787277456372976, + 0.008058102801442146, + 0.005303643643856049, + 0.005571715533733368, + 0.004431113135069609, + 0.0037807822227478027, + 0.003965890500694513, + 0.006898202002048492, + 0.008668129332363605, + 0.006965304724872112, + 0.0054607088677585125, + 0.0038384480867534876, + 0.007188383489847183, + 0.019157927483320236, + 0.016620945185422897, + 0.022652972489595413, + 0.025411618873476982, + 0.024863366037607193, + 0.020033854991197586, + 0.019587993621826172, + 0.012935652397572994, + 0.007188582792878151, + 0.006969161331653595, + 0.03177530691027641, + 0.020532747730612755, + 0.014204557053744793, + 0.0097358338534832, + 0.009029930457472801, + 0.016621241346001625, + 0.0036049718037247658, + 0.0026174287777394056, + 0.010328135453164577, + 0.006195596419274807, + 0.004498942755162716, + 0.0033071476500481367, + 0.002745973877608776, + 0.005323972553014755, + 0.005405968986451626, + 0.003422651207074523, + 0.006221074610948563, + 0.004728749860078096, + 0.00909513607621193, + 0.0037899522576481104, + 0.004500472918152809, + 0.004105601459741592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.042336463928222656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980480156838894 + ], + "opposite_logprob_tokens": [ + -3.5298047065734863 + ], + "logprob_reward": -0.02980480156838894, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980480156838894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42220136523246765, + 0.034465909004211426, + 0.02786986529827118, + 0.06871646642684937, + 0.03157215937972069, + 0.020822349935770035, + 0.08264140784740448, + 0.02812713384628296, + 0.05417279154062271, + 0.22941063344478607 + ], + "credit_map": [ + 0.03990766778588295, + 0.03007735125720501, + 0.11297569423913956, + 0.09693696349859238, + 0.04053175449371338, + 0.02991877682507038, + 0.036811068654060364, + 0.035042036324739456, + 0.003351273015141487, + 0.0030201205518096685, + 0.005917895585298538, + 0.006020702421665192, + 0.005169885233044624, + 0.003985726274549961, + 0.004512791056185961, + 0.0024875104427337646, + 0.0028532349970191717, + 0.002302322071045637, + 0.00622033653780818, + 0.004833374638110399, + 0.004662616644054651, + 0.002290305681526661, + 0.002923692110925913, + 0.0017839769134297967, + 0.008390511386096478, + 0.007176164537668228, + 0.011338473297655582, + 0.010468968190252781, + 0.009573240764439106, + 0.005911023821681738, + 0.008976603858172894, + 0.006881476379930973, + 0.0025861391332000494, + 0.0026099318638443947, + 0.004036697559058666, + 0.003740848507732153, + 0.009086430072784424, + 0.0024987594224512577, + 0.00470559811219573, + 0.0023077530786395073, + 0.001846141996793449, + 0.0018656293395906687, + 0.004438076168298721, + 0.002696272684261203, + 0.002413671463727951, + 0.0018841730197891593, + 0.0014752097195014358, + 0.004203171469271183, + 0.010536756366491318, + 0.007990465499460697, + 0.012250763364136219, + 0.01300235278904438, + 0.0123571353033185, + 0.010874294675886631, + 0.009495146572589874, + 0.006134473718702793, + 0.0026718005537986755, + 0.002651632996276021, + 0.005430355668067932, + 0.00374074699357152, + 0.0038181040436029434, + 0.004279842134565115, + 0.002800934249535203, + 0.0027337127830833197, + 0.005080318544059992, + 0.003833746537566185, + 0.017976585775613785, + 0.013212419115006924, + 0.003404117189347744, + 0.0030436241067945957, + 0.002330410061404109, + 0.005291558802127838, + 0.01712653413414955, + 0.01691601425409317, + 0.05255761370062828, + 0.04023829847574234, + 0.0333457812666893, + 0.02469736523926258, + 0.03132224828004837, + 0.013206744566559792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.02980480156838894 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024840945843607187 + ], + "opposite_logprob_tokens": [ + -6.002484321594238 + ], + "logprob_reward": -0.0024840945843607187, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024840945843607187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2755889892578125, + 0.1283104568719864, + 0.07622536271810532, + 0.07839111983776093, + 0.03928344324231148, + 0.06543447822332382, + 0.019832918420433998, + 0.11772767454385757, + 0.12826979160308838, + 0.07093576341867447 + ], + "credit_map": [ + 0.04101518914103508, + 0.026652352884411812, + 0.035870760679244995, + 0.058998528867959976, + 0.025655385106801987, + 0.02773744985461235, + 0.023547587916254997, + 0.036111749708652496, + 0.01350833848118782, + 0.014111081138253212, + 0.024970367550849915, + 0.016237616539001465, + 0.016168024390935898, + 0.013803627341985703, + 0.013194081373512745, + 0.016317343339323997, + 0.00545042660087347, + 0.005402472801506519, + 0.021872738376259804, + 0.018301036208868027, + 0.007825411856174469, + 0.005436907056719065, + 0.005834887269884348, + 0.006101486273109913, + 0.01090757641941309, + 0.008880331180989742, + 0.01678771711885929, + 0.011388173326849937, + 0.008125594817101955, + 0.007077964022755623, + 0.006281365640461445, + 0.008942403830587864, + 0.004189508035778999, + 0.00324059184640646, + 0.01110746432095766, + 0.007772235665470362, + 0.0037966016680002213, + 0.003004850586876273, + 0.002607756294310093, + 0.0035644355230033398, + 0.007266057189553976, + 0.006038853898644447, + 0.019931232556700706, + 0.014002257958054543, + 0.004397605080157518, + 0.006033382378518581, + 0.0027434309013187885, + 0.005021662451326847, + 0.0035651300568133593, + 0.0018824185244739056, + 0.0033802271354943514, + 0.0025268413592129946, + 0.002715006936341524, + 0.0021008916664868593, + 0.002157513052225113, + 0.0015048924833536148, + 0.012218336574733257, + 0.009481151588261127, + 0.038286201655864716, + 0.027270223945379257, + 0.006140853278338909, + 0.007422361057251692, + 0.005581530276685953, + 0.011327026411890984, + 0.016384797170758247, + 0.012357190251350403, + 0.03623058274388313, + 0.02571529522538185, + 0.011542313732206821, + 0.008607963100075722, + 0.00585884740576148, + 0.011572818271815777, + 0.009865911677479744, + 0.005995207466185093, + 0.01698276400566101, + 0.017711762338876724, + 0.006723540835082531, + 0.00543840741738677, + 0.004825369920581579, + 0.0033928074408322573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0024840945843607187 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036112351808696985 + ], + "opposite_logprob_tokens": [ + -5.628611087799072 + ], + "logprob_reward": -0.0036112351808696985, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036112351808696985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18247710168361664, + 0.07802356779575348, + 0.08790450543165207, + 0.13237452507019043, + 0.09701395034790039, + 0.11882176250219345, + 0.05138184130191803, + 0.08008759468793869, + 0.09258390218019485, + 0.07933124899864197 + ], + "credit_map": [ + 0.016453171148896217, + 0.012825331650674343, + 0.041401926428079605, + 0.041069045662879944, + 0.021945832297205925, + 0.021533135324716568, + 0.016270816326141357, + 0.010977850295603275, + 0.010506120510399342, + 0.00921848975121975, + 0.009747189469635487, + 0.010709675960242748, + 0.012964521534740925, + 0.010450558736920357, + 0.008190339431166649, + 0.006236671935766935, + 0.019353479146957397, + 0.00816629733890295, + 0.013612418435513973, + 0.012973063625395298, + 0.010701632127165794, + 0.007454002276062965, + 0.008120804093778133, + 0.007522811647504568, + 0.022246012464165688, + 0.01419519167393446, + 0.02134019136428833, + 0.016063310205936432, + 0.017491940408945084, + 0.01686316542327404, + 0.011943884193897247, + 0.012230821885168552, + 0.012874179519712925, + 0.009853056631982327, + 0.0135916443541646, + 0.014382104389369488, + 0.015348493121564388, + 0.011088731698691845, + 0.012969974428415298, + 0.00690576434135437, + 0.010870412923395634, + 0.009523994289338589, + 0.03256124630570412, + 0.025358151644468307, + 0.011318412609398365, + 0.01600397378206253, + 0.007122769020497799, + 0.006062805186957121, + 0.007175772916525602, + 0.005644980352371931, + 0.008349694311618805, + 0.010728524066507816, + 0.007355471141636372, + 0.003871110500767827, + 0.004589146468788385, + 0.003667145501822233, + 0.007943426258862019, + 0.006035227328538895, + 0.02097437158226967, + 0.014148353599011898, + 0.008126937784254551, + 0.010900386609137058, + 0.005607872735708952, + 0.006351015996187925, + 0.009205477312207222, + 0.007027041167020798, + 0.02852734737098217, + 0.020427705720067024, + 0.007178039755672216, + 0.006017165258526802, + 0.003752701450139284, + 0.010448429733514786, + 0.004728567320853472, + 0.007069672457873821, + 0.016994204372167587, + 0.010096556507050991, + 0.012866195291280746, + 0.00926035363227129, + 0.008629560470581055, + 0.00968613289296627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0036112351808696985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020568059757351875 + ], + "opposite_logprob_tokens": [ + -3.8955681324005127 + ], + "logprob_reward": -0.020568059757351875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020568059757351875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29291650652885437, + 0.10421725362539291, + 0.11369994282722473, + 0.14880385994911194, + 0.0412447452545166, + 0.06463109701871872, + 0.04753316938877106, + 0.022958505898714066, + 0.06748853623867035, + 0.09650640189647675 + ], + "credit_map": [ + 0.01812869869172573, + 0.015464511699974537, + 0.0841655358672142, + 0.06423873454332352, + 0.027413172647356987, + 0.025768890976905823, + 0.02431618981063366, + 0.03342076390981674, + 0.011048456653952599, + 0.007256814744323492, + 0.01675770804286003, + 0.015711931511759758, + 0.019135626032948494, + 0.01243031769990921, + 0.014967304654419422, + 0.0069090877659618855, + 0.01794545352458954, + 0.010283072479069233, + 0.018595004454255104, + 0.014636783860623837, + 0.015891432762145996, + 0.011723359115421772, + 0.01450964156538248, + 0.010115205310285091, + 0.012451251037418842, + 0.011385317891836166, + 0.017582105472683907, + 0.019671514630317688, + 0.039393480867147446, + 0.013574410229921341, + 0.024827061221003532, + 0.009918713942170143, + 0.0051981545984745026, + 0.003442050889134407, + 0.009179124608635902, + 0.006086157634854317, + 0.005028135143220425, + 0.003766099689528346, + 0.004624004475772381, + 0.003921015653759241, + 0.006939441431313753, + 0.006297754589468241, + 0.01787853054702282, + 0.017424674704670906, + 0.0044892774894833565, + 0.005045561119914055, + 0.003480558516457677, + 0.003075302578508854, + 0.00559109216555953, + 0.00334372790530324, + 0.009020453318953514, + 0.005939312279224396, + 0.00708809494972229, + 0.005901754833757877, + 0.006309415213763714, + 0.004339317791163921, + 0.002892465563490987, + 0.0017423488898202777, + 0.0030520737636834383, + 0.0037183791864663363, + 0.003929661586880684, + 0.002938029123470187, + 0.001960793975740671, + 0.002724753227084875, + 0.004998764954507351, + 0.004845554940402508, + 0.01747806929051876, + 0.009575600735843182, + 0.00795904640108347, + 0.006003382615745068, + 0.007367132231593132, + 0.009260979481041431, + 0.004477932583540678, + 0.00804895255714655, + 0.019214943051338196, + 0.011729364283382893, + 0.018879741430282593, + 0.011677319183945656, + 0.013004997745156288, + 0.009473156183958054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.020568059757351875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a man in his 30s standing confidently at the edge of a luxurious swimming pool on a sunny day. He has short, neatly styled brown hair, and his expression is warm and friendly as he speaks directly to the camera. He wears a light blue, short-sleeved polo shirt made of high-quality chinos, paired with khaki shorts. The bright midday sun casts a soft glow on him, highlighting the texture of his shirt. Behind him, the crystal-clear turquoise waters of the lake ripple gently, creating a dance of light reflections on the surface. The pool is surrounded by a clean, light-colored stone terrace. In the background, several empty lounge chairs and a large white umbrella can be seen, surrounded by lush green tropical foliage. The man occasionally gestures with his right hand in a relaxed and open manner, emphasizing his words. His casual posture and direct eye contact create a warm and engaging atmosphere, as if he is sharing a pleasant story or travel tips.", + "seed": 4853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_0/index_7919_seed_4853_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016425574314780533, + "vlm_logprob": -0.016425574314780533, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.016425574314780533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039330825209617615 + ], + "opposite_logprob_tokens": [ + -3.7893307209014893 + ], + "logprob_reward": -0.039330825209617615, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.039330825209617615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24507635831832886, + 0.05247874930500984, + 0.09443751722574234, + 0.17799225449562073, + 0.08691413700580597, + 0.06414203345775604, + 0.1261628121137619, + 0.07513453811407089, + 0.046164821833372116, + 0.03149683028459549 + ], + "credit_map": [ + 0.018799355253577232, + 0.016042722389101982, + 0.019427262246608734, + 0.02700377255678177, + 0.034864358603954315, + 0.025939088314771652, + 0.08755325525999069, + 0.015446529723703861, + 0.005795009434223175, + 0.003842694219201803, + 0.004275172017514706, + 0.005461013410240412, + 0.008394640870392323, + 0.006086931563913822, + 0.013049867004156113, + 0.00557341892272234, + 0.006807765457779169, + 0.006128108128905296, + 0.005722317844629288, + 0.011596409603953362, + 0.018583936616778374, + 0.011209588497877121, + 0.028016474097967148, + 0.006372906267642975, + 0.015571633353829384, + 0.008434637449681759, + 0.012414811179041862, + 0.027053050696849823, + 0.02957868203520775, + 0.01947435922920704, + 0.050492044538259506, + 0.01497301459312439, + 0.013452444225549698, + 0.006750287488102913, + 0.006810636259615421, + 0.010360341519117355, + 0.006827351171523333, + 0.019679628312587738, + 0.018418487161397934, + 0.004614952486008406, + 0.0080072907730937, + 0.004993918351829052, + 0.005696922540664673, + 0.00927175022661686, + 0.008796432055532932, + 0.007857613265514374, + 0.015243244357407093, + 0.00427485816180706, + 0.014474407769739628, + 0.007560630328953266, + 0.008396022021770477, + 0.013585439883172512, + 0.024654490873217583, + 0.01389777660369873, + 0.03323657810688019, + 0.0103574488312006, + 0.016194703057408333, + 0.004659660160541534, + 0.005304527934640646, + 0.004507873672991991, + 0.007378777023404837, + 0.014206431806087494, + 0.016220295801758766, + 0.006662264000624418, + 0.005498017184436321, + 0.002753020264208317, + 0.003240598365664482, + 0.005379361566156149, + 0.008060541935265064, + 0.006497266702353954, + 0.011823389679193497, + 0.002912624506279826, + 0.003907020203769207, + 0.00179178302641958, + 0.002396937692537904, + 0.0036469788756221533, + 0.0046807038597762585, + 0.0061205062083899975, + 0.006795293651521206, + 0.00215760525316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.039330825209617615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079693213105202 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079693213105202, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079693213105202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11233386397361755, + 0.05575753003358841, + 0.2621428072452545, + 0.07673145830631256, + 0.1692148894071579, + 0.03782559186220169, + 0.07562830299139023, + 0.1422245353460312, + 0.02697138860821724, + 0.04116962477564812 + ], + "credit_map": [ + 0.013143016025424004, + 0.010005353949964046, + 0.0113741559907794, + 0.021722815930843353, + 0.017271408811211586, + 0.01305331289768219, + 0.017830556258559227, + 0.007933244109153748, + 0.005155760794878006, + 0.004102155100554228, + 0.004485833924263716, + 0.008113565854728222, + 0.010459770448505878, + 0.005635423585772514, + 0.013013456016778946, + 0.0047915661707520485, + 0.022799156606197357, + 0.017126960679888725, + 0.018788350746035576, + 0.024099871516227722, + 0.035708557814359665, + 0.03436596319079399, + 0.08028463274240494, + 0.02896934188902378, + 0.006263697519898415, + 0.004960689228028059, + 0.007036295719444752, + 0.013421686366200447, + 0.012840273790061474, + 0.008443060331046581, + 0.018450677394866943, + 0.005315074697136879, + 0.024070240557193756, + 0.013524314388632774, + 0.010040509514510632, + 0.0160257238894701, + 0.008546113967895508, + 0.043221648782491684, + 0.04464450851082802, + 0.009141837246716022, + 0.00662901159375906, + 0.0035114935599267483, + 0.0032309191301465034, + 0.0035011242143809795, + 0.004598649218678474, + 0.004699814599007368, + 0.00841617677360773, + 0.0032383990474045277, + 0.010758164338767529, + 0.005548741668462753, + 0.006948606111109257, + 0.0102674700319767, + 0.011441037990152836, + 0.009500871412456036, + 0.011510740034282207, + 0.00965267326682806, + 0.024573156610131264, + 0.008905068971216679, + 0.010682045482099056, + 0.012840689159929752, + 0.016948308795690536, + 0.024547483772039413, + 0.025209423154592514, + 0.018518365919589996, + 0.003304859157651663, + 0.0019174247281625867, + 0.0019570006988942623, + 0.0034237983636558056, + 0.004738259594887495, + 0.0032212038058787584, + 0.00629114406183362, + 0.0021176980808377266, + 0.0061653777956962585, + 0.0030889089684933424, + 0.0028849923983216286, + 0.004980612546205521, + 0.004320802632719278, + 0.009483853355050087, + 0.006830860860645771, + 0.0034142148215323687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.011079693213105202 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017106198938563466 + ], + "opposite_logprob_tokens": [ + -6.376710414886475 + ], + "logprob_reward": -0.0017106198938563466, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017106198938563466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18293051421642303, + 0.07563278824090958, + 0.06710880249738693, + 0.05432269349694252, + 0.21322527527809143, + 0.10323252528905869, + 0.1044761911034584, + 0.0857955813407898, + 0.037059981375932693, + 0.07621565461158752 + ], + "credit_map": [ + 0.02806731127202511, + 0.023601312190294266, + 0.024309491738677025, + 0.02357367053627968, + 0.020181676372885704, + 0.026470109820365906, + 0.020754145458340645, + 0.015972774475812912, + 0.012591028586030006, + 0.009400294162333012, + 0.007557486649602652, + 0.008861111477017403, + 0.008074709214270115, + 0.01200961135327816, + 0.009507420472800732, + 0.007631118409335613, + 0.010847698897123337, + 0.007870770990848541, + 0.0076028136536479, + 0.005490464624017477, + 0.007985735312104225, + 0.008418714627623558, + 0.009040823206305504, + 0.009851779788732529, + 0.007405856158584356, + 0.005597926676273346, + 0.006679900456219912, + 0.00737028568983078, + 0.007514047436416149, + 0.005826096516102552, + 0.009387782774865627, + 0.004540791269391775, + 0.045736685395240784, + 0.015398925170302391, + 0.018981371074914932, + 0.021577367559075356, + 0.012894329614937305, + 0.028741290792822838, + 0.05415109544992447, + 0.01574420928955078, + 0.012841279618442059, + 0.014315551146864891, + 0.014059321023523808, + 0.02117597497999668, + 0.00844950694590807, + 0.012926514260470867, + 0.012507464736700058, + 0.00695691118016839, + 0.013421693816781044, + 0.006494374480098486, + 0.00895022414624691, + 0.013976089656352997, + 0.01644950546324253, + 0.013701246120035648, + 0.02053934708237648, + 0.010943702422082424, + 0.017773952335119247, + 0.010983572341501713, + 0.008438349701464176, + 0.008906363509595394, + 0.007156091276556253, + 0.015970448032021523, + 0.010433255694806576, + 0.006133549381047487, + 0.005544791463762522, + 0.004262670874595642, + 0.0034563355147838593, + 0.004022508393973112, + 0.004051882773637772, + 0.00710042379796505, + 0.006019119638949633, + 0.0026022514794021845, + 0.010948849841952324, + 0.009523672051727772, + 0.012309337966144085, + 0.0073631927371025085, + 0.008299514651298523, + 0.008642803877592087, + 0.010224275290966034, + 0.008904005400836468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.0017106198938563466 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732518319040537 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732518319040537, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732518319040537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06772974878549576, + 0.03448205813765526, + 0.028578193858265877, + 0.03778592124581337, + 0.600365400314331, + 0.010915741324424744, + 0.04248746484518051, + 0.11863492429256439, + 0.016063382849097252, + 0.04295715317130089 + ], + "credit_map": [ + 0.009630364365875721, + 0.005398957524448633, + 0.009031749330461025, + 0.010509832762181759, + 0.007912391796708107, + 0.01065001543611288, + 0.009982218965888023, + 0.004614225123077631, + 0.006424205377697945, + 0.0033287438564002514, + 0.004901291336864233, + 0.00291243614628911, + 0.0038824458606541157, + 0.00437134550884366, + 0.004775791894644499, + 0.003885800950229168, + 0.0076983338221907616, + 0.003366418182849884, + 0.002967446343973279, + 0.002259194618090987, + 0.002654009498655796, + 0.0043601393699646, + 0.0034872458782047033, + 0.0017854090547189116, + 0.007483615539968014, + 0.003943776711821556, + 0.0042947507463395596, + 0.0046910797245800495, + 0.0035474819596856833, + 0.005045528989285231, + 0.005848926492035389, + 0.002930764574557543, + 0.10774087905883789, + 0.04491367191076279, + 0.04161463677883148, + 0.060661476105451584, + 0.03024463728070259, + 0.12219551205635071, + 0.1575188785791397, + 0.035475730895996094, + 0.0016667868476361036, + 0.0012393621727824211, + 0.0013724911259487271, + 0.0018480579601600766, + 0.000914459116756916, + 0.0018787028966471553, + 0.0014379421481862664, + 0.0005579394637607038, + 0.0070112962275743484, + 0.0029294395353645086, + 0.005301477387547493, + 0.008723974227905273, + 0.0034640063531696796, + 0.0061696539632976055, + 0.005964116659015417, + 0.0029235058464109898, + 0.02071194350719452, + 0.007863937877118587, + 0.009194408543407917, + 0.0077132051810622215, + 0.008986554108560085, + 0.01789623312652111, + 0.03753858804702759, + 0.00873007345944643, + 0.0020278242882341146, + 0.0016146465204656124, + 0.001739787170663476, + 0.0022500804625451565, + 0.0021837668027728796, + 0.00208744197152555, + 0.003047114936634898, + 0.001112721860408783, + 0.006303820293396711, + 0.006757992785423994, + 0.006419790908694267, + 0.0049397582188248634, + 0.004172707907855511, + 0.005248616449534893, + 0.005352168343961239, + 0.0037623040843755007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.006732518319040537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023274214938282967 + ], + "opposite_logprob_tokens": [ + -3.7732741832733154 + ], + "logprob_reward": -0.023274214938282967, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023274214938282967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20521511137485504, + 0.05352543294429779, + 0.07113517075777054, + 0.12678034603595734, + 0.06144052371382713, + 0.02524297498166561, + 0.04069817438721657, + 0.2956259846687317, + 0.1079593226313591, + 0.012376920320093632 + ], + "credit_map": [ + 0.017244411632418633, + 0.013777375221252441, + 0.029464568942785263, + 0.04838337004184723, + 0.021233392879366875, + 0.027984529733657837, + 0.03923267126083374, + 0.007894810289144516, + 0.005894209258258343, + 0.0031321365386247635, + 0.003541237208992243, + 0.007501369342207909, + 0.00865299440920353, + 0.005254622083157301, + 0.013698427937924862, + 0.005850441288203001, + 0.00676264613866806, + 0.004804003983736038, + 0.004406333435326815, + 0.007113868370652199, + 0.012968105264008045, + 0.012013761326670647, + 0.017438044771552086, + 0.005628413055092096, + 0.009533797390758991, + 0.00783439539372921, + 0.010487678460776806, + 0.025377770885825157, + 0.02020757459104061, + 0.01253907848149538, + 0.032223932445049286, + 0.008576123043894768, + 0.006535561755299568, + 0.0036822976544499397, + 0.00611779885366559, + 0.011504181660711765, + 0.0062593817710876465, + 0.008482868783175945, + 0.0146356001496315, + 0.0042228396050632, + 0.0034282973501831293, + 0.0018047284102067351, + 0.002767877886071801, + 0.0035538426600396633, + 0.002907667076215148, + 0.002656256314367056, + 0.005464490037411451, + 0.00265981862321496, + 0.005479274317622185, + 0.003122747642919421, + 0.002968445187434554, + 0.0049226959235966206, + 0.007682248018682003, + 0.004328172653913498, + 0.008495034649968147, + 0.003699564142152667, + 0.04716252163052559, + 0.015413686633110046, + 0.022877512499690056, + 0.020528826862573624, + 0.02130323275923729, + 0.054253414273262024, + 0.09178298711776733, + 0.02230381779372692, + 0.011119895614683628, + 0.006473533809185028, + 0.00938424188643694, + 0.020015619695186615, + 0.013236913830041885, + 0.014071661978960037, + 0.024301793426275253, + 0.009355672635138035, + 0.0014510262990370393, + 0.0011401817901059985, + 0.0012621887726709247, + 0.002177549758926034, + 0.002042775508016348, + 0.0016399449668824673, + 0.001750028575770557, + 0.0009132258128374815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.023274214938282967 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a man in his 30s standing confidently at the edge of a luxurious swimming pool on a sunny day. He has short, neatly styled brown hair, and his expression is warm and friendly as he speaks directly to the camera. He wears a light blue, short-sleeved polo shirt made of high-quality chinos, paired with khaki shorts. The bright midday sun casts a soft glow on him, highlighting the texture of his shirt. Behind him, the crystal-clear turquoise waters of the lake ripple gently, creating a dance of light reflections on the surface. The pool is surrounded by a clean, light-colored stone terrace. In the background, several empty lounge chairs and a large white umbrella can be seen, surrounded by lush green tropical foliage. The man occasionally gestures with his right hand in a relaxed and open manner, emphasizing his words. His casual posture and direct eye contact create a warm and engaging atmosphere, as if he is sharing a pleasant story or travel tips.", + "seed": 4854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_0/index_7919_seed_4854_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01621138146147132, + "vlm_logprob": -0.01621138146147132, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.01621138146147132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02714030258357525 + ], + "opposite_logprob_tokens": [ + -4.152140140533447 + ], + "logprob_reward": -0.02714030258357525, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02714030258357525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2511408030986786, + 0.05048999935388565, + 0.03195522725582123, + 0.049474380910396576, + 0.07762360572814941, + 0.05080440640449524, + 0.12194135785102844, + 0.09828189015388489, + 0.22313067317008972, + 0.045157600194215775 + ], + "credit_map": [ + 0.013033755123615265, + 0.02759423479437828, + 0.02700221538543701, + 0.023204071447253227, + 0.06957170367240906, + 0.03687761723995209, + 0.03877648711204529, + 0.015080730430781841, + 0.0051203141920268536, + 0.004837243817746639, + 0.004717400297522545, + 0.005313188303261995, + 0.011582722887396812, + 0.007280865218490362, + 0.008032509125769138, + 0.0036057564429938793, + 0.002993874019011855, + 0.003020449308678508, + 0.004139874596148729, + 0.0033568430226296186, + 0.00685912137851119, + 0.004402216989547014, + 0.004629152826964855, + 0.0025536944158375263, + 0.0048054843209683895, + 0.004954732954502106, + 0.00434885686263442, + 0.004144852980971336, + 0.013689998537302017, + 0.006702605169266462, + 0.0072785490192472935, + 0.0035492994356900454, + 0.007448471151292324, + 0.007625325117260218, + 0.008617717772722244, + 0.013103723526000977, + 0.011741016991436481, + 0.015608367510139942, + 0.008547715842723846, + 0.004931262228637934, + 0.004769999999552965, + 0.006879983004182577, + 0.005036892835050821, + 0.004222747404128313, + 0.012069561518728733, + 0.0063662659376859665, + 0.008051562123000622, + 0.003407391021028161, + 0.009213419631123543, + 0.008375453762710094, + 0.028341952711343765, + 0.012629100121557713, + 0.012619905173778534, + 0.02758452296257019, + 0.011355728842318058, + 0.011821265332400799, + 0.012532914057374, + 0.007325221784412861, + 0.015468306839466095, + 0.014446739107370377, + 0.014087862335145473, + 0.01274307444691658, + 0.012739802710711956, + 0.008937972597777843, + 0.013133399188518524, + 0.015749063342809677, + 0.06566443294286728, + 0.02002865821123123, + 0.020445669069886208, + 0.05443980172276497, + 0.01705797202885151, + 0.016611695289611816, + 0.005189058370888233, + 0.005124570801854134, + 0.005118079949170351, + 0.004034890793263912, + 0.00564960390329361, + 0.006945424247533083, + 0.00516107864677906, + 0.007934893481433392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.02714030258357525 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864654965698719 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.00864654965698719, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864654965698719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1372569352388382, + 0.039993081241846085, + 0.07920216023921967, + 0.03130296245217323, + 0.10234017670154572, + 0.024036670103669167, + 0.41927289962768555, + 0.08466584980487823, + 0.024600083008408546, + 0.05732913687825203 + ], + "credit_map": [ + 0.007704606745392084, + 0.017821552231907845, + 0.020898209884762764, + 0.018896782770752907, + 0.022934703156352043, + 0.01760631799697876, + 0.024575361981987953, + 0.006819418631494045, + 0.004218421410769224, + 0.004338064696639776, + 0.0031409382354468107, + 0.004117122385650873, + 0.008206288330256939, + 0.005957390647381544, + 0.007266192231327295, + 0.0027486691251397133, + 0.007178160827606916, + 0.005723048932850361, + 0.005870670545846224, + 0.006878581829369068, + 0.016371777281165123, + 0.011030355468392372, + 0.013536745682358742, + 0.012612828053534031, + 0.00356872845441103, + 0.003535207826644182, + 0.0038547234144061804, + 0.0030238828621804714, + 0.00570261524990201, + 0.004336609039455652, + 0.004579597618430853, + 0.0027015996165573597, + 0.009111094288527966, + 0.0067771743051707745, + 0.010089876130223274, + 0.022032901644706726, + 0.015410139225423336, + 0.021821869537234306, + 0.010850402526557446, + 0.006246729753911495, + 0.0025074973236769438, + 0.0029187831096351147, + 0.002999665215611458, + 0.0023595409002155066, + 0.0046206284314394, + 0.003754298435524106, + 0.0030663569923490286, + 0.0018099005101248622, + 0.03163259103894234, + 0.028780562803149223, + 0.13656455278396606, + 0.04025761038064957, + 0.038343194872140884, + 0.07667248696088791, + 0.039921704679727554, + 0.027100250124931335, + 0.007392990402877331, + 0.007093498017638922, + 0.012256018817424774, + 0.0244231466203928, + 0.007538858335465193, + 0.010632556863129139, + 0.008267352357506752, + 0.007061433978378773, + 0.00285206432454288, + 0.002511778846383095, + 0.00466403691098094, + 0.0039613195694983006, + 0.0025537859182804823, + 0.002844556700438261, + 0.0026031285524368286, + 0.0026094159111380577, + 0.005007900297641754, + 0.005088335834443569, + 0.007069048006087542, + 0.011104358360171318, + 0.0060230037197470665, + 0.0061173150315880775, + 0.010763867758214474, + 0.006155310198664665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.00864654965698719 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019383705221116543 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019383705221116543, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019383705221116543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1376713067293167, + 0.07335060089826584, + 0.1311211884021759, + 0.06528253108263016, + 0.023615878075361252, + 0.07283671945333481, + 0.06895782053470612, + 0.06739316135644913, + 0.2725965976715088, + 0.08717424422502518 + ], + "credit_map": [ + 0.011109177954494953, + 0.017475999891757965, + 0.021339425817131996, + 0.011159393936395645, + 0.021124955266714096, + 0.02005750499665737, + 0.022083312273025513, + 0.013321518898010254, + 0.013828874565660954, + 0.010240251198410988, + 0.012806587852537632, + 0.006570968311280012, + 0.008489510975778103, + 0.009636271744966507, + 0.004896543920040131, + 0.006881594657897949, + 0.01302210334688425, + 0.009316461160779, + 0.012035523541271687, + 0.008600904606282711, + 0.016057953238487244, + 0.02574707195162773, + 0.012333834543824196, + 0.03400733321905136, + 0.007549542002379894, + 0.008524375036358833, + 0.009472208097577095, + 0.00629426771774888, + 0.007465437985956669, + 0.007965236902236938, + 0.00740364333614707, + 0.010607821866869926, + 0.0023219396825879812, + 0.0026036251801997423, + 0.003811108646914363, + 0.0026630039792507887, + 0.0038608210161328316, + 0.0034053765702992678, + 0.0030371525790542364, + 0.0019128493731841445, + 0.01005265861749649, + 0.008143126964569092, + 0.0103292902931571, + 0.0053453766740858555, + 0.008822953328490257, + 0.018728775903582573, + 0.006062529049813747, + 0.00535200722515583, + 0.006055461708456278, + 0.008573713712394238, + 0.0069747925736010075, + 0.009235481731593609, + 0.012084932997822762, + 0.012248631566762924, + 0.008053647354245186, + 0.005731156561523676, + 0.013293184340000153, + 0.008843403309583664, + 0.006229968275874853, + 0.00557888625189662, + 0.009148012846708298, + 0.00909954123198986, + 0.009449456818401814, + 0.005750703625380993, + 0.01610368676483631, + 0.024326955899596214, + 0.08075810223817825, + 0.021652676165103912, + 0.015107298269867897, + 0.08146803081035614, + 0.016093231737613678, + 0.01708660088479519, + 0.0070360866375267506, + 0.006139848381280899, + 0.015611091628670692, + 0.01397373154759407, + 0.008454974740743637, + 0.008854185231029987, + 0.014249050058424473, + 0.012855282053351402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0019383705221116543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249762907624245 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249762907624245, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249762907624245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09714289009571075, + 0.02879277803003788, + 0.1422518640756607, + 0.07417477667331696, + 0.20963479578495026, + 0.03206157311797142, + 0.11345274746417999, + 0.11816263198852539, + 0.11412110924720764, + 0.07020478695631027 + ], + "credit_map": [ + 0.01955951191484928, + 0.013092500157654285, + 0.010302742943167686, + 0.005518752615898848, + 0.01154593750834465, + 0.013880277052521706, + 0.011109977960586548, + 0.012133195996284485, + 0.003505435772240162, + 0.003045100485906005, + 0.004242392256855965, + 0.0031412497628480196, + 0.005003571975976229, + 0.0037741034757345915, + 0.00357119832187891, + 0.0025097262114286423, + 0.017470816150307655, + 0.01091747172176838, + 0.017006197944283485, + 0.009725803509354591, + 0.015614956617355347, + 0.022269586101174355, + 0.012697904370725155, + 0.036549121141433716, + 0.009238959290087223, + 0.008139953948557377, + 0.012018619105219841, + 0.006982591934502125, + 0.010949854739010334, + 0.012281432747840881, + 0.007031816523522139, + 0.007531547918915749, + 0.027082767337560654, + 0.025667497888207436, + 0.032815102487802505, + 0.021033653989434242, + 0.01659967564046383, + 0.0552905835211277, + 0.01917402632534504, + 0.011971494182944298, + 0.004981745034456253, + 0.003510788781568408, + 0.00501930620521307, + 0.0024722563102841377, + 0.004204896744340658, + 0.0068867867812514305, + 0.0028439012821763754, + 0.0021418945398181677, + 0.009572777897119522, + 0.00943282712250948, + 0.02669435366988182, + 0.012698127888143063, + 0.012409709393978119, + 0.026013273745775223, + 0.008752436377108097, + 0.007879247888922691, + 0.01520369853824377, + 0.016715021803975105, + 0.022651810199022293, + 0.014293402433395386, + 0.011002842336893082, + 0.01650727167725563, + 0.010421518236398697, + 0.011367063038051128, + 0.007023793179541826, + 0.011168005876243114, + 0.030778907239437103, + 0.008541574701666832, + 0.009431696496903896, + 0.03301958739757538, + 0.007372381631284952, + 0.00678516598418355, + 0.006213263142853975, + 0.009274544194340706, + 0.0075174178928136826, + 0.012812010943889618, + 0.007056342903524637, + 0.007021509576588869, + 0.013053679838776588, + 0.007256018463522196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.005249762907624245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a man in his 30s?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light blue short-sleeved polo shirt and khaki shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man standing at the edge of a luxurious swimming pool with crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are empty lounge chairs and a large white umbrella visible in the background surrounded by lush green tropical foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture with his right hand in a relaxed manner while speaking directly to the camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808192163705826 + ], + "opposite_logprob_tokens": [ + -3.2880818843841553 + ], + "logprob_reward": -0.03808192163705826, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808192163705826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09659872949123383, + 0.03403512388467789, + 0.04696178808808327, + 0.03509758040308952, + 0.2587187588214874, + 0.046599239110946655, + 0.19435453414916992, + 0.059861134737730026, + 0.08726499229669571, + 0.14050808548927307 + ], + "credit_map": [ + 0.006187628023326397, + 0.009003125131130219, + 0.013159364461898804, + 0.01108926348388195, + 0.017664194107055664, + 0.014338916167616844, + 0.019305717200040817, + 0.005850518587976694, + 0.003058473812416196, + 0.002860154490917921, + 0.00423408392816782, + 0.0050832005217671394, + 0.006483765318989754, + 0.00399826280772686, + 0.006206473335623741, + 0.0021107119973748922, + 0.00443763192743063, + 0.004531291779130697, + 0.002951305592432618, + 0.0035576317459344864, + 0.012446003966033459, + 0.00634036073461175, + 0.007533503696322441, + 0.005164057482033968, + 0.0030844404827803373, + 0.004792030900716782, + 0.0024953982792794704, + 0.0023411132860928774, + 0.009516789577901363, + 0.005203463137149811, + 0.00555140245705843, + 0.002112944144755602, + 0.017248282209038734, + 0.02051720954477787, + 0.028336966410279274, + 0.07260634750127792, + 0.029182104393839836, + 0.04247936233878136, + 0.02789478749036789, + 0.0204536821693182, + 0.00473364582285285, + 0.003192218951880932, + 0.004535112529993057, + 0.004060848616063595, + 0.010802640579640865, + 0.008798239752650261, + 0.007571504916995764, + 0.002905024914070964, + 0.014843741431832314, + 0.007344112731516361, + 0.06520216166973114, + 0.018680253997445107, + 0.017582949250936508, + 0.04234836995601654, + 0.018048517405986786, + 0.010304431430995464, + 0.010992031544446945, + 0.005194237921386957, + 0.006558756344020367, + 0.009664283134043217, + 0.00519096152856946, + 0.010077199898660183, + 0.006731487810611725, + 0.00545217702165246, + 0.005205780733376741, + 0.0075611830689013, + 0.012506574392318726, + 0.023704493418335915, + 0.007118514738976955, + 0.015161475166678429, + 0.00965078640729189, + 0.006356185767799616, + 0.01565762236714363, + 0.022229500114917755, + 0.0182988028973341, + 0.022832386195659637, + 0.021157683804631233, + 0.014825128950178623, + 0.01650281809270382, + 0.009004148654639721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.03808192163705826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_tensor([4023, 4023])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_tensor([4023, 4023])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d05d6ec2529a6d2e66d4b752fdeafe160013e084 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_tensor([4023, 4023])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot places the viewer in the midst of a primal showdown, depicting a ferocious Siberian tiger and a wild boar locked in a fierce tussle in a deciduous forest. The adult tiger, a massive beast, boasts a resplendent orange-yellow coat streaked with striking black stripes. Its shoulder and back muscles bulge with exertion, its maw gaping, revealing long, sharp canines and a silent roar. Its massive front paws, their nails fully developed, dig like hooks into the boar's thick, dark-brown, bristly flesh. Beneath the tiger's grasp, the equally formidable boar struggles frantically, its massive limbs pawing at the ground in an effort to free itself, twisting its mud-covered head, attempting to strike at the tiger's soft underbelly with its sharp, upturned tusks jutting from the corners of its mouth. Their brutal struggle unleashes a storm on the ground. Countless dry, fragile ochre and russet fallen leaves mingle with the black earth, hurled into the air by their immense force, forming a chaotic and dynamic vortex that flies across the canvas. The soft afternoon sunlight filters through the surrounding bare treetops, casting dappled shadows across the forest floor, perfectly illuminating the airy dust and leaf fragments, infusing the entire scene with an atmosphere of tension, chaos, and a savage, primal vitality.", + "seed": 32936, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_4023_seed_32936_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08393418220803142, + "vlm_logprob": -0.08393418220803142, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.08393418220803142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21171759068965912 + ], + "opposite_logprob_tokens": [ + -1.7117176055908203 + ], + "logprob_reward": -0.21171759068965912, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21171759068965912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05710401013493538, + 0.08156239986419678, + 0.06484277546405792, + 0.049486856907606125, + 0.0411621518433094, + 0.06414724886417389, + 0.5082787275314331, + 0.05218980833888054, + 0.04482356086373329, + 0.03640240803360939 + ], + "credit_map": [ + 0.006653390824794769, + 0.00624087592586875, + 0.01024685986340046, + 0.007887176238000393, + 0.004752293229103088, + 0.007489211857318878, + 0.00924033671617508, + 0.004593860357999802, + 0.009603146463632584, + 0.011116140522062778, + 0.011772681958973408, + 0.011240125633776188, + 0.005823066923767328, + 0.011030597612261772, + 0.008510670624673367, + 0.012465966865420341, + 0.008341272361576557, + 0.012653184123337269, + 0.007902782410383224, + 0.009122700430452824, + 0.007644641678780317, + 0.008402608335018158, + 0.0044990116730332375, + 0.006276573985815048, + 0.009617344476282597, + 0.00901393685489893, + 0.004321816843003035, + 0.004108768422156572, + 0.004420338664203882, + 0.008679530583322048, + 0.004376118537038565, + 0.004949002061039209, + 0.005874585825949907, + 0.008617396466434002, + 0.005965676624327898, + 0.00369144044816494, + 0.004346716683357954, + 0.005041342694312334, + 0.004150034859776497, + 0.003474959172308445, + 0.005361603572964668, + 0.010178906843066216, + 0.00556998560205102, + 0.004801142495125532, + 0.017544493079185486, + 0.010695543140172958, + 0.006490762811154127, + 0.0035048085264861584, + 0.10243469476699829, + 0.06873128563165665, + 0.054576270282268524, + 0.04262226074934006, + 0.08097293227910995, + 0.06335457414388657, + 0.041730642318725586, + 0.05385610833764076, + 0.006930692587047815, + 0.010602740570902824, + 0.0045735533349215984, + 0.0055667441338300705, + 0.0062375376001000404, + 0.004219021648168564, + 0.007627860642969608, + 0.0064316620118916035, + 0.0067840153351426125, + 0.005650938022881746, + 0.003269231878221035, + 0.003067484125494957, + 0.003337326692417264, + 0.005930514074862003, + 0.0036650383844971657, + 0.013119014911353588, + 0.006088616326451302, + 0.006127530708909035, + 0.005528673063963652, + 0.003463596571236849, + 0.003712485544383526, + 0.002930064918473363, + 0.004022587090730667, + 0.00452885078266263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.21171759068965912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004119957331568003 + ], + "opposite_logprob_tokens": [ + -5.504119873046875 + ], + "logprob_reward": -0.004119957331568003, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004119957331568003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042673010379076004, + 0.037156928330659866, + 0.022881919518113136, + 0.029263025149703026, + 0.026754900813102722, + 0.025978969410061836, + 0.7599674463272095, + 0.022699015215039253, + 0.011740698479115963, + 0.02088410034775734 + ], + "credit_map": [ + 0.005734696984291077, + 0.005425577983260155, + 0.007387946825474501, + 0.006311118137091398, + 0.0030504055321216583, + 0.00442280201241374, + 0.006498944945633411, + 0.0038415146991610527, + 0.0042510987259447575, + 0.00703191664069891, + 0.004750263411551714, + 0.0034628715366125107, + 0.0036910302005708218, + 0.00573247205466032, + 0.004135427996516228, + 0.004101844970136881, + 0.003794584423303604, + 0.005532091483473778, + 0.002098398981615901, + 0.002473493805155158, + 0.0022126054391264915, + 0.0035285635385662317, + 0.0015233715530484915, + 0.0017188111087307334, + 0.004290770273655653, + 0.004893459379673004, + 0.0033018263056874275, + 0.0027689996641129255, + 0.003278736025094986, + 0.0061549413949251175, + 0.0024999924935400486, + 0.00207429938018322, + 0.0036310243885964155, + 0.005501348990947008, + 0.004322017543017864, + 0.0026499186642467976, + 0.0029756356962025166, + 0.0031948380637913942, + 0.0022560246288776398, + 0.0022240940015763044, + 0.002920142374932766, + 0.0039504398591816425, + 0.001526535372249782, + 0.0017479914240539074, + 0.007254597265273333, + 0.004439735785126686, + 0.002268054988235235, + 0.0018714737379923463, + 0.16411183774471283, + 0.10602108389139175, + 0.06619155406951904, + 0.05500681325793266, + 0.12435591965913773, + 0.09715002775192261, + 0.06329338997602463, + 0.08383681625127792, + 0.003352533094584942, + 0.006459931842982769, + 0.003374638734385371, + 0.0017637968994677067, + 0.002668013097718358, + 0.0019215036882087588, + 0.0015128428349271417, + 0.0016457547899335623, + 0.001404113369062543, + 0.0014848841819912195, + 0.001481293817050755, + 0.0009500117739662528, + 0.001510682632215321, + 0.0012558657908812165, + 0.0010865411022678018, + 0.002567305928096175, + 0.0029027555137872696, + 0.00518685020506382, + 0.0018158385064452887, + 0.0017601209692656994, + 0.002728731371462345, + 0.00169912981800735, + 0.0032348469831049442, + 0.001555828028358519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.004119957331568003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787782095372677 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787782095372677, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787782095372677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026181941851973534, + 0.022690771147608757, + 0.024265196174383163, + 0.02822667360305786, + 0.03822312504053116, + 0.02197781205177307, + 0.7831794619560242, + 0.013905828818678856, + 0.024944672361016273, + 0.016404522582888603 + ], + "credit_map": [ + 0.0027180591132491827, + 0.0023336075246334076, + 0.0043267663568258286, + 0.004117266274988651, + 0.002210887148976326, + 0.0036459453403949738, + 0.004389649722725153, + 0.0024397613015025854, + 0.002438443247228861, + 0.0034374401438981295, + 0.002641723258420825, + 0.0022042954806238413, + 0.0024922422599047422, + 0.003741710213944316, + 0.0029212322551757097, + 0.002813684055581689, + 0.004468216560781002, + 0.0034285213332623243, + 0.004246468190103769, + 0.0023419398348778486, + 0.0030765405390411615, + 0.002447938546538353, + 0.001872883178293705, + 0.002382685663178563, + 0.004959723446518183, + 0.00506658386439085, + 0.003019205294549465, + 0.0020661416929215193, + 0.0031096050515770912, + 0.005947724916040897, + 0.0017254914855584502, + 0.002332199364900589, + 0.0064337109215557575, + 0.00773858604952693, + 0.006318323314189911, + 0.0037251771427690983, + 0.0035708907525986433, + 0.004189630970358849, + 0.002877843100577593, + 0.003368963021785021, + 0.0029852294828742743, + 0.0031562764197587967, + 0.0014783774968236685, + 0.001479315455071628, + 0.005681121256202459, + 0.0035640387795865536, + 0.001982533372938633, + 0.0016509186243638396, + 0.16869978606700897, + 0.1081015020608902, + 0.07069738954305649, + 0.05809960141777992, + 0.12740465998649597, + 0.09935836493968964, + 0.06495867669582367, + 0.08585943281650543, + 0.0024480847641825676, + 0.002827843651175499, + 0.0030054079834371805, + 0.0010626373114064336, + 0.0011041946709156036, + 0.0010000297334045172, + 0.0013235037913545966, + 0.0011341273784637451, + 0.003051402047276497, + 0.0036390533205121756, + 0.0025507290847599506, + 0.0016804912593215704, + 0.0021024406887590885, + 0.003118357388302684, + 0.0019023597706109285, + 0.006899839732795954, + 0.0021120200399309397, + 0.0035058266948908567, + 0.001932767452672124, + 0.0014530910411849618, + 0.0019043025095015764, + 0.0013572083553299308, + 0.0021644018124789, + 0.001974905375391245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.009787782095372677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336962565779686 + ], + "opposite_logprob_tokens": [ + -3.408696174621582 + ], + "logprob_reward": -0.0336962565779686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336962565779686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03525245934724808, + 0.023477287963032722, + 0.010320826433598995, + 0.03350377455353737, + 0.07707768678665161, + 0.037688154727220535, + 0.6968368291854858, + 0.0323646254837513, + 0.03892280161380768, + 0.014555591158568859 + ], + "credit_map": [ + 0.0030008095782250166, + 0.0034638664219528437, + 0.005040238611400127, + 0.004968259483575821, + 0.0037658335641026497, + 0.00458337040618062, + 0.006772185210138559, + 0.0036578967701643705, + 0.003098483430221677, + 0.003967225551605225, + 0.0026405283715575933, + 0.0026362596545368433, + 0.003143604611977935, + 0.0028335535898804665, + 0.002290190663188696, + 0.0028674437198787928, + 0.0019593320321291685, + 0.001599857467226684, + 0.0008261976181529462, + 0.0014764388324692845, + 0.0013042585924267769, + 0.0010789011139422655, + 0.0010869029210880399, + 0.0009889376815408468, + 0.004217350389808416, + 0.007096945773810148, + 0.0042371973395347595, + 0.0031483240891247988, + 0.003986293915659189, + 0.0058216070756316185, + 0.002670002169907093, + 0.0023260540328919888, + 0.01080030482262373, + 0.013892889022827148, + 0.01579063944518566, + 0.009315981529653072, + 0.0065030320547521114, + 0.008215931244194508, + 0.005636419635266066, + 0.00692248810082674, + 0.003917708061635494, + 0.00593592319637537, + 0.002424727426841855, + 0.002505288692191243, + 0.010423190891742706, + 0.00610555661842227, + 0.0032459727954119444, + 0.003129783784970641, + 0.1512400209903717, + 0.09781183302402496, + 0.060632966458797455, + 0.0503525547683239, + 0.1144866943359375, + 0.0896100327372551, + 0.057483769953250885, + 0.07521887868642807, + 0.005231443792581558, + 0.008343275636434555, + 0.004346427042037249, + 0.002116010058671236, + 0.0033134918194264174, + 0.002842273563146591, + 0.0017880556406453252, + 0.004383646417409182, + 0.004964265041053295, + 0.005103824194520712, + 0.007928621955215931, + 0.004321307875216007, + 0.004889839328825474, + 0.0032736314460635185, + 0.004143656697124243, + 0.004297654144465923, + 0.002332218922674656, + 0.0018875726964324713, + 0.002911246381700039, + 0.001394735649228096, + 0.0015546102076768875, + 0.0013341892044991255, + 0.0015240489738062024, + 0.0016169692389667034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.0336962565779686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034932434558868 + ], + "opposite_logprob_tokens": [ + -1.9103493690490723 + ], + "logprob_reward": -0.16034932434558868, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034932434558868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010916435159742832, + 0.03152739629149437, + 0.009849077090620995, + 0.013977994211018085, + 0.014901846647262573, + 0.013887876644730568, + 0.863875687122345, + 0.009174642153084278, + 0.012159924022853374, + 0.01972910203039646 + ], + "credit_map": [ + 0.0015110305976122618, + 0.001879263436421752, + 0.001298804534599185, + 0.001335163484327495, + 0.0013122232630848885, + 0.0013277378166094422, + 0.0010605276329442859, + 0.001191683579236269, + 0.004695937968790531, + 0.005613056011497974, + 0.004078655503690243, + 0.0030929448548704386, + 0.0046936143189668655, + 0.0032172142527997494, + 0.0031864233314990997, + 0.002949546091258526, + 0.0012671787990257144, + 0.001787654240615666, + 0.0020003714598715305, + 0.001225584070198238, + 0.0008993553929030895, + 0.0011645524064078927, + 0.0006590108969248831, + 0.0008453693590126932, + 0.002326920395717025, + 0.0028279435355216265, + 0.0011794636957347393, + 0.001041100244037807, + 0.0016324457246810198, + 0.0027715137694031, + 0.001147208153270185, + 0.0010513968300074339, + 0.00236310763284564, + 0.0029671762604266405, + 0.0025580173823982477, + 0.0012918759603053331, + 0.0017882181564345956, + 0.0015688129933550954, + 0.0009956301655620337, + 0.0013690079795196652, + 0.0025720407720655203, + 0.002312009921297431, + 0.0012001842260360718, + 0.0011937206145375967, + 0.0020191227085888386, + 0.0016811210662126541, + 0.000996675225906074, + 0.0019130011787638068, + 0.18715398013591766, + 0.1209951862692833, + 0.07415032386779785, + 0.06267008930444717, + 0.14175772666931152, + 0.11068553477525711, + 0.07134846597909927, + 0.09511440247297287, + 0.00178976496681571, + 0.0010532448068261147, + 0.001009077299386263, + 0.0004209452017676085, + 0.0013053294969722629, + 0.0009308054577559233, + 0.0006183813675306737, + 0.0020470928866416216, + 0.0017121090786531568, + 0.0013576700584962964, + 0.0010895829182118177, + 0.0008240938186645508, + 0.0010125057306140661, + 0.0016164189437404275, + 0.0011198269203305244, + 0.0034277159720659256, + 0.0024045431055128574, + 0.002978599863126874, + 0.003234189935028553, + 0.002149058971554041, + 0.002792028011754155, + 0.0017195283435285091, + 0.002777127316221595, + 0.0016740236897021532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.16034932434558868 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot places the viewer in the midst of a primal showdown, depicting a ferocious Siberian tiger and a wild boar locked in a fierce tussle in a deciduous forest. The adult tiger, a massive beast, boasts a resplendent orange-yellow coat streaked with striking black stripes. Its shoulder and back muscles bulge with exertion, its maw gaping, revealing long, sharp canines and a silent roar. Its massive front paws, their nails fully developed, dig like hooks into the boar's thick, dark-brown, bristly flesh. Beneath the tiger's grasp, the equally formidable boar struggles frantically, its massive limbs pawing at the ground in an effort to free itself, twisting its mud-covered head, attempting to strike at the tiger's soft underbelly with its sharp, upturned tusks jutting from the corners of its mouth. Their brutal struggle unleashes a storm on the ground. Countless dry, fragile ochre and russet fallen leaves mingle with the black earth, hurled into the air by their immense force, forming a chaotic and dynamic vortex that flies across the canvas. The soft afternoon sunlight filters through the surrounding bare treetops, casting dappled shadows across the forest floor, perfectly illuminating the airy dust and leaf fragments, infusing the entire scene with an atmosphere of tension, chaos, and a savage, primal vitality.", + "seed": 32937, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_4023_seed_32937_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07158620143309236, + "vlm_logprob": -0.07158620143309236, + "vlm_yes_no_margin": 3.249999976158142, + "vlm_token_credit": -0.07158620143309236, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07985588163137436 + ], + "opposite_logprob_tokens": [ + -2.8298559188842773 + ], + "logprob_reward": -0.07985588163137436, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07985588163137436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08225352317094803, + 0.07369787245988846, + 0.03635832294821739, + 0.030423378571867943, + 0.13998091220855713, + 0.37860679626464844, + 0.08850926160812378, + 0.046132974326610565, + 0.053266558796167374, + 0.07077035307884216 + ], + "credit_map": [ + 0.012853177264332771, + 0.014298438094556332, + 0.01052452065050602, + 0.008162995800375938, + 0.009801384061574936, + 0.010006147436797619, + 0.008086583577096462, + 0.0085202781483531, + 0.014939462766051292, + 0.008052448742091656, + 0.011266365647315979, + 0.009425214491784573, + 0.009540232829749584, + 0.006122197490185499, + 0.00877360813319683, + 0.005578347947448492, + 0.004510211758315563, + 0.004907779395580292, + 0.005647394340485334, + 0.005370486062020063, + 0.0046029649674892426, + 0.003648307640105486, + 0.004161910153925419, + 0.0035092728212475777, + 0.0039058492984622717, + 0.003097772831097245, + 0.005389661993831396, + 0.003843914484605193, + 0.0035353871062397957, + 0.0035275418777018785, + 0.004565358627587557, + 0.002557894913479686, + 0.034029193222522736, + 0.010489100590348244, + 0.0177337396889925, + 0.008652791380882263, + 0.014365315437316895, + 0.02241716906428337, + 0.01565106213092804, + 0.016642559319734573, + 0.04320674389600754, + 0.020263226702809334, + 0.02488170564174652, + 0.01630629412829876, + 0.11844635009765625, + 0.015737658366560936, + 0.10841386765241623, + 0.031351007521152496, + 0.008987421169877052, + 0.013748870231211185, + 0.010102389380335808, + 0.00798191037029028, + 0.010781222954392433, + 0.014594834297895432, + 0.011041199788451195, + 0.011271419003605843, + 0.0047647529281675816, + 0.006753264460712671, + 0.012416486628353596, + 0.003952378872781992, + 0.007509900722652674, + 0.004484076052904129, + 0.0032501909881830215, + 0.0030019255355000496, + 0.005467017646878958, + 0.006108087953180075, + 0.009926977567374706, + 0.004444369580596685, + 0.00618939520791173, + 0.00438976613804698, + 0.012062410824000835, + 0.004678539931774139, + 0.009860605001449585, + 0.007683356758207083, + 0.005629506893455982, + 0.006192450411617756, + 0.013786327093839645, + 0.007649073377251625, + 0.006698310375213623, + 0.013270726427435875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.07985588163137436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041276742704212666 + ], + "opposite_logprob_tokens": [ + -5.504127502441406 + ], + "logprob_reward": -0.0041276742704212666, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041276742704212666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5390897989273071, + 0.07641569525003433, + 0.07145313918590546, + 0.054813042283058167, + 0.08721549063920975, + 0.03181212767958641, + 0.030778607353568077, + 0.0192740298807621, + 0.05768918991088867, + 0.031458910554647446 + ], + "credit_map": [ + 0.0457608625292778, + 0.02534625306725502, + 0.02817671187222004, + 0.20401428639888763, + 0.03372262418270111, + 0.013916399329900742, + 0.03370171785354614, + 0.154450923204422, + 0.013038313016295433, + 0.005853773094713688, + 0.013455391861498356, + 0.010793805122375488, + 0.008154415525496006, + 0.006723626982420683, + 0.011718292720615864, + 0.0066780769266188145, + 0.009421257302165031, + 0.014439448714256287, + 0.009458711370825768, + 0.008594226092100143, + 0.009846414439380169, + 0.005885744001716375, + 0.007648863364011049, + 0.006158473901450634, + 0.0068975347094237804, + 0.00284653645940125, + 0.010211322456598282, + 0.01249412540346384, + 0.0037822260055691004, + 0.00523761473596096, + 0.008406789973378181, + 0.004936889745295048, + 0.019712964072823524, + 0.006587156094610691, + 0.012808498926460743, + 0.005398067645728588, + 0.008954471908509731, + 0.013889949768781662, + 0.009763984009623528, + 0.010100394487380981, + 0.007158050313591957, + 0.004476715810596943, + 0.0026441682130098343, + 0.002843444235622883, + 0.004081019200384617, + 0.004657181445509195, + 0.0030514136888086796, + 0.0029001301154494286, + 0.0028259349055588245, + 0.004675638396292925, + 0.004579097963869572, + 0.003417509375140071, + 0.003769857110455632, + 0.002990060020238161, + 0.005708007141947746, + 0.002812498016282916, + 0.0022311711218208075, + 0.0032885067630559206, + 0.002029634779319167, + 0.002571993740275502, + 0.0030302514787763357, + 0.002385475905612111, + 0.001989394426345825, + 0.0017475993372499943, + 0.007620660588145256, + 0.011915205046534538, + 0.011907665058970451, + 0.005214083474129438, + 0.0040092370472848415, + 0.00540295522660017, + 0.005375918932259083, + 0.006243461277335882, + 0.004410048481076956, + 0.005362948402762413, + 0.004239461850374937, + 0.003503483720123768, + 0.004068635869771242, + 0.0031908471137285233, + 0.003032812150195241, + 0.00365067133679986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0041276742704212666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819830760359764 + ], + "opposite_logprob_tokens": [ + -4.018198490142822 + ], + "logprob_reward": -0.01819830760359764, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819830760359764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4002391993999481, + 0.0741293653845787, + 0.08842374384403229, + 0.02985786646604538, + 0.07174140214920044, + 0.04756862297654152, + 0.02817203477025032, + 0.02270289696753025, + 0.18686844408512115, + 0.050296392291784286 + ], + "credit_map": [ + 0.03373900055885315, + 0.02148108370602131, + 0.020673835650086403, + 0.1467791050672531, + 0.0259366724640131, + 0.01396126113831997, + 0.026688015088438988, + 0.11098024994134903, + 0.009095984511077404, + 0.009029867127537727, + 0.01049846038222313, + 0.012183595448732376, + 0.010295436717569828, + 0.009406800381839275, + 0.008069869130849838, + 0.005549353081732988, + 0.017043525353074074, + 0.01447209995239973, + 0.012657199054956436, + 0.010246866382658482, + 0.009875918738543987, + 0.005677046254277229, + 0.010953797958791256, + 0.007497284095734358, + 0.0036778838839381933, + 0.0030399796087294817, + 0.0034110420383512974, + 0.004905638284981251, + 0.0026336410082876682, + 0.0055733947083354, + 0.003899114904925227, + 0.0027171734254807234, + 0.01601317897439003, + 0.005142529029399157, + 0.009187504649162292, + 0.004640564322471619, + 0.007073911372572184, + 0.010934450663626194, + 0.0071673886850476265, + 0.011581875383853912, + 0.010464845225214958, + 0.006443403661251068, + 0.005367935635149479, + 0.003749910742044449, + 0.007019948214292526, + 0.005313119385391474, + 0.0044722226448357105, + 0.004737239331007004, + 0.0031155815813690424, + 0.0038762625772506, + 0.0038095295894891024, + 0.003312160726636648, + 0.003886470338329673, + 0.0037865641061216593, + 0.0037979574408382177, + 0.002587506314739585, + 0.002322969026863575, + 0.003760583233088255, + 0.002615333767607808, + 0.003314373781904578, + 0.0034511147532612085, + 0.0023106676526367664, + 0.002598890569061041, + 0.002328964415937662, + 0.0191739983856678, + 0.021782271564006805, + 0.05688561499118805, + 0.009689404629170895, + 0.00948469527065754, + 0.00946797151118517, + 0.048160918056964874, + 0.012223565950989723, + 0.006478998810052872, + 0.0062514664605259895, + 0.0037106310483068228, + 0.0033095970284193754, + 0.007157376501709223, + 0.005643841344863176, + 0.006751549895852804, + 0.010992931202054024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.01819830760359764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294683068990707 + ], + "opposite_logprob_tokens": [ + -2.2379467487335205 + ], + "logprob_reward": -0.11294683068990707, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294683068990707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37956899404525757, + 0.07403447479009628, + 0.11706044524908066, + 0.03733714297413826, + 0.058098312467336655, + 0.04961744323372841, + 0.027120893821120262, + 0.019595691934227943, + 0.16553260385990143, + 0.07203400880098343 + ], + "credit_map": [ + 0.03121391497552395, + 0.021824171766638756, + 0.019657550379633904, + 0.1354764997959137, + 0.024691332131624222, + 0.014924958348274231, + 0.028601378202438354, + 0.1031792089343071, + 0.009211583994328976, + 0.007626994978636503, + 0.011406944133341312, + 0.013077843934297562, + 0.01008074451237917, + 0.006144654471427202, + 0.009154661558568478, + 0.007331044878810644, + 0.024395247921347618, + 0.01665104739367962, + 0.017577793449163437, + 0.01377685647457838, + 0.011545788496732712, + 0.007769695017486811, + 0.013963252305984497, + 0.011380759067833424, + 0.006936016492545605, + 0.003360570641234517, + 0.0036293372977524996, + 0.004883874207735062, + 0.003913641907274723, + 0.007260013837367296, + 0.0033978046849370003, + 0.003955882508307695, + 0.01194984745234251, + 0.006407724693417549, + 0.006295426748692989, + 0.0044742850586771965, + 0.00813743844628334, + 0.007403262425214052, + 0.0065372539684176445, + 0.006893075071275234, + 0.010132680647075176, + 0.007618039846420288, + 0.00467812130227685, + 0.0038798595778644085, + 0.00650694128125906, + 0.00653614429756999, + 0.00559889804571867, + 0.00466675590723753, + 0.0037449703086167574, + 0.003192371688783169, + 0.0035872473381459713, + 0.0030069444328546524, + 0.0036306693218648434, + 0.003607864724472165, + 0.0031758060213178396, + 0.003175020217895508, + 0.0017001990927383304, + 0.002918472047895193, + 0.0036987969651818275, + 0.002751137362793088, + 0.002536818152293563, + 0.0015851552598178387, + 0.0023909867741167545, + 0.002014125930145383, + 0.018855644389986992, + 0.03465305641293526, + 0.0365949310362339, + 0.015158846974372864, + 0.013281627558171749, + 0.013478628359735012, + 0.02081526629626751, + 0.01269459817558527, + 0.009799355641007423, + 0.006402104161679745, + 0.005307996645569801, + 0.006488689687103033, + 0.012584209442138672, + 0.008373036980628967, + 0.01083553209900856, + 0.012243087403476238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.11294683068990707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280231297016144 + ], + "opposite_logprob_tokens": [ + -2.0178022384643555 + ], + "logprob_reward": -0.14280231297016144, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280231297016144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22545219957828522, + 0.06505978852510452, + 0.05414233356714249, + 0.04637429490685463, + 0.36154651641845703, + 0.1307215541601181, + 0.0280047245323658, + 0.014635812491178513, + 0.026894675567746162, + 0.04716808721423149 + ], + "credit_map": [ + 0.021820854395627975, + 0.019477184861898422, + 0.016360556706786156, + 0.06601732969284058, + 0.019352607429027557, + 0.015721024945378304, + 0.015777580440044403, + 0.05092505365610123, + 0.011465919204056263, + 0.00794031098484993, + 0.007516516372561455, + 0.00967442151159048, + 0.008604692295193672, + 0.005914858542382717, + 0.008326107636094093, + 0.005616961047053337, + 0.00987162347882986, + 0.005609435960650444, + 0.008514842949807644, + 0.00881893653422594, + 0.004222236108034849, + 0.00503726489841938, + 0.00616107415407896, + 0.0059069180861115456, + 0.005943859461694956, + 0.0034087940584868193, + 0.004819333553314209, + 0.008092949166893959, + 0.0050275600515306, + 0.010085790418088436, + 0.00408081104978919, + 0.004915194120258093, + 0.0694022998213768, + 0.027619697153568268, + 0.04263230040669441, + 0.020517634227871895, + 0.055452991276979446, + 0.05162408947944641, + 0.05011189356446266, + 0.04418560862541199, + 0.014245200902223587, + 0.004197601228952408, + 0.007807260844856501, + 0.004195930901914835, + 0.042547836899757385, + 0.005056178197264671, + 0.04096602648496628, + 0.011705500073730946, + 0.003139602253213525, + 0.003292759647592902, + 0.0038637579418718815, + 0.002583577996119857, + 0.004501000512391329, + 0.003911653533577919, + 0.004282018635421991, + 0.002430350985378027, + 0.0016874620923772454, + 0.0019632489420473576, + 0.0017692908877506852, + 0.0018860058626160026, + 0.0022934600710868835, + 0.0020913102198392153, + 0.0012638046173378825, + 0.0016812296817079186, + 0.004056491889059544, + 0.0030599043238908052, + 0.0024907514452934265, + 0.0022405185736715794, + 0.004714527633041143, + 0.004053301177918911, + 0.0031655875500291586, + 0.003113590646535158, + 0.0065352898091077805, + 0.003717215033248067, + 0.0030620477627962828, + 0.0034309353213757277, + 0.008842524141073227, + 0.004887359216809273, + 0.007746189832687378, + 0.008946522139012814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.14280231297016144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot places the viewer in the midst of a primal showdown, depicting a ferocious Siberian tiger and a wild boar locked in a fierce tussle in a deciduous forest. The adult tiger, a massive beast, boasts a resplendent orange-yellow coat streaked with striking black stripes. Its shoulder and back muscles bulge with exertion, its maw gaping, revealing long, sharp canines and a silent roar. Its massive front paws, their nails fully developed, dig like hooks into the boar's thick, dark-brown, bristly flesh. Beneath the tiger's grasp, the equally formidable boar struggles frantically, its massive limbs pawing at the ground in an effort to free itself, twisting its mud-covered head, attempting to strike at the tiger's soft underbelly with its sharp, upturned tusks jutting from the corners of its mouth. Their brutal struggle unleashes a storm on the ground. Countless dry, fragile ochre and russet fallen leaves mingle with the black earth, hurled into the air by their immense force, forming a chaotic and dynamic vortex that flies across the canvas. The soft afternoon sunlight filters through the surrounding bare treetops, casting dappled shadows across the forest floor, perfectly illuminating the airy dust and leaf fragments, infusing the entire scene with an atmosphere of tension, chaos, and a savage, primal vitality.", + "seed": 32938, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_4023_seed_32938_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05801266524940729, + "vlm_logprob": -0.05801266524940729, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.05801266524940729, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0412876233458519 + ], + "opposite_logprob_tokens": [ + -3.416287660598755 + ], + "logprob_reward": -0.0412876233458519, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0412876233458519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07775965332984924, + 0.11934630572795868, + 0.026003172621130943, + 0.05971372127532959, + 0.07509054988622665, + 0.03465517237782478, + 0.02370033599436283, + 0.3357551097869873, + 0.13657517731189728, + 0.1114007830619812 + ], + "credit_map": [ + 0.006675757933408022, + 0.006787706166505814, + 0.008687395602464676, + 0.008586580865085125, + 0.011787340976297855, + 0.007751908153295517, + 0.005949374753981829, + 0.021533597260713577, + 0.014272844418883324, + 0.019832570105791092, + 0.013729804195463657, + 0.014838841743767262, + 0.020215870812535286, + 0.013445498421788216, + 0.01111905463039875, + 0.011891829781234264, + 0.0030871080234646797, + 0.004019328858703375, + 0.004232865292578936, + 0.00317159341648221, + 0.002374199451878667, + 0.003203385742381215, + 0.003220273181796074, + 0.0026944205164909363, + 0.00887982826679945, + 0.006145089399069548, + 0.006975053809583187, + 0.005993178114295006, + 0.01128369476646185, + 0.00995186809450388, + 0.005643099080771208, + 0.004841914400458336, + 0.010458958335220814, + 0.02338712103664875, + 0.005618572235107422, + 0.005727929063141346, + 0.010810906067490578, + 0.0069333757273852825, + 0.006078195292502642, + 0.006075501907616854, + 0.0034477219451218843, + 0.005187358241528273, + 0.00259661884047091, + 0.0030490197241306305, + 0.007911409251391888, + 0.004737873561680317, + 0.003909419756382704, + 0.003815754083916545, + 0.002779210451990366, + 0.0042572347447276115, + 0.0028963240329176188, + 0.0020540975965559483, + 0.002816788386553526, + 0.0030651355627924204, + 0.0018921175505965948, + 0.003939428366720676, + 0.07374100387096405, + 0.10611682385206223, + 0.014705621637403965, + 0.019024338573217392, + 0.02105766348540783, + 0.04619764909148216, + 0.017407149076461792, + 0.037504892796278, + 0.018742185086011887, + 0.018739277496933937, + 0.016570650041103363, + 0.013775774277746677, + 0.018652144819498062, + 0.017198849469423294, + 0.012635387480258942, + 0.02026090770959854, + 0.0096890265122056, + 0.03888870030641556, + 0.005609484855085611, + 0.0051105208694934845, + 0.004294432699680328, + 0.008706770837306976, + 0.005957042332738638, + 0.03314480558037758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.0412876233458519 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002843862399458885 + ], + "opposite_logprob_tokens": [ + -5.877843856811523 + ], + "logprob_reward": -0.002843862399458885, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002843862399458885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07059184461832047, + 0.08755584806203842, + 0.1277364045381546, + 0.06428870558738708, + 0.09158354252576828, + 0.06769119948148727, + 0.06095500662922859, + 0.14797574281692505, + 0.06288668513298035, + 0.2187349945306778 + ], + "credit_map": [ + 0.010765934363007545, + 0.010605626739561558, + 0.006999135483056307, + 0.004468204919248819, + 0.013337375596165657, + 0.007956421002745628, + 0.007546950131654739, + 0.00891219824552536, + 0.011985437013208866, + 0.013727533631026745, + 0.015107043087482452, + 0.00946652702987194, + 0.010519196279346943, + 0.010667014867067337, + 0.008696812205016613, + 0.007386280223727226, + 0.01787569187581539, + 0.02543436549603939, + 0.02195282280445099, + 0.01860423944890499, + 0.008457047864794731, + 0.01644037291407585, + 0.01080550067126751, + 0.008166365325450897, + 0.011318639852106571, + 0.01024543959647417, + 0.008730432018637657, + 0.00482717901468277, + 0.00582725228741765, + 0.007776862941682339, + 0.004989474546164274, + 0.010573429986834526, + 0.014004616998136044, + 0.023930074647068977, + 0.009801794774830341, + 0.0076956236734986305, + 0.011654823087155819, + 0.00813823938369751, + 0.008246777579188347, + 0.008111590519547462, + 0.00886521302163601, + 0.012894068844616413, + 0.005725718569010496, + 0.005077419336885214, + 0.009095779620110989, + 0.012606586329638958, + 0.00687372824177146, + 0.006552684120833874, + 0.007186013739556074, + 0.009854359552264214, + 0.0042051272466778755, + 0.005289820954203606, + 0.007610253989696503, + 0.011183309368789196, + 0.005937460344284773, + 0.009688661433756351, + 0.0244850255548954, + 0.04925327003002167, + 0.010188025422394276, + 0.010900808498263359, + 0.012163386680185795, + 0.012358591891825199, + 0.010280598886311054, + 0.018346043303608894, + 0.013413512147963047, + 0.009514935314655304, + 0.007249972317367792, + 0.004747138358652592, + 0.009223848581314087, + 0.006503098178654909, + 0.004682397935539484, + 0.00755178602412343, + 0.019819485023617744, + 0.07451357692480087, + 0.010238736867904663, + 0.010468882508575916, + 0.009381129406392574, + 0.017669543623924255, + 0.012108074501156807, + 0.06453558057546616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.002843862399458885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083937250077724 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011083937250077724, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083937250077724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14902035892009735, + 0.10500359535217285, + 0.14416325092315674, + 0.06158534437417984, + 0.07021120935678482, + 0.03705849498510361, + 0.043585482984781265, + 0.07712262868881226, + 0.057630859315395355, + 0.25461870431900024 + ], + "credit_map": [ + 0.01655527390539646, + 0.016802094876766205, + 0.023019801825284958, + 0.02013387903571129, + 0.016439994797110558, + 0.015061668120324612, + 0.011720366775989532, + 0.02928730845451355, + 0.009047704748809338, + 0.01495169848203659, + 0.01190865132957697, + 0.010047021321952343, + 0.013463056646287441, + 0.011785210110247135, + 0.019057219848036766, + 0.01474304124712944, + 0.019321035593748093, + 0.02603474073112011, + 0.025107981637120247, + 0.019889848306775093, + 0.009579174220561981, + 0.017390461638569832, + 0.013947474770247936, + 0.012892549857497215, + 0.010105752386152744, + 0.010452693328261375, + 0.008742352947592735, + 0.004017351660877466, + 0.00650235777720809, + 0.00749666104093194, + 0.004375329241156578, + 0.009892851114273071, + 0.010864198207855225, + 0.015171732753515244, + 0.00683653075248003, + 0.0070920782163739204, + 0.010124750435352325, + 0.00838067103177309, + 0.005439517553895712, + 0.006301734130829573, + 0.00498235272243619, + 0.0060767922550439835, + 0.0028246946167200804, + 0.003251587273553014, + 0.004952937830239534, + 0.006425780244171619, + 0.004837345331907272, + 0.0037070096004754305, + 0.005058953072875738, + 0.006739026866853237, + 0.00460341852158308, + 0.003814012510702014, + 0.0049730208702385426, + 0.007559033110737801, + 0.0038616154342889786, + 0.0069764056243002415, + 0.008150752633810043, + 0.020069951191544533, + 0.007952572777867317, + 0.00690260948613286, + 0.007621609605848789, + 0.008360479027032852, + 0.007227730937302113, + 0.01083693653345108, + 0.00849916785955429, + 0.007796239107847214, + 0.007368441671133041, + 0.00546018872410059, + 0.008985289372503757, + 0.005578841548413038, + 0.008228394202888012, + 0.005714304279536009, + 0.02333194576203823, + 0.08328600972890854, + 0.012103889137506485, + 0.012341362424194813, + 0.010781965218484402, + 0.01953243464231491, + 0.013553558848798275, + 0.07968757301568985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.011083937250077724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495734512805939 + ], + "opposite_logprob_tokens": [ + -2.929957389831543 + ], + "logprob_reward": -0.05495734512805939, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495734512805939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14728786051273346, + 0.14149002730846405, + 0.083262138068676, + 0.0627194344997406, + 0.04842912778258324, + 0.07789729535579681, + 0.030042538419365883, + 0.08986916393041611, + 0.039631716907024384, + 0.2793706953525543 + ], + "credit_map": [ + 0.024690018966794014, + 0.01745680533349514, + 0.022898683324456215, + 0.014894332736730576, + 0.02373535931110382, + 0.017605917528271675, + 0.013639351353049278, + 0.01236737985163927, + 0.01728922501206398, + 0.017715975642204285, + 0.015490779653191566, + 0.017220672219991684, + 0.01793314516544342, + 0.01875447668135166, + 0.014599648304283619, + 0.02248610556125641, + 0.012110267765820026, + 0.012248811312019825, + 0.01591455191373825, + 0.012072645127773285, + 0.006840878631919622, + 0.010155177675187588, + 0.00832303985953331, + 0.005596765782684088, + 0.00832898449152708, + 0.011830613017082214, + 0.006586541887372732, + 0.005910144653171301, + 0.008156903088092804, + 0.0077306306920945644, + 0.004497266374528408, + 0.00967834796756506, + 0.007254601456224918, + 0.006733658257871866, + 0.004891297314316034, + 0.004529866855591536, + 0.006041325628757477, + 0.008105027489364147, + 0.004575068596750498, + 0.006298279855400324, + 0.007064732722938061, + 0.012462396174669266, + 0.004925569519400597, + 0.005350285675376654, + 0.012934381142258644, + 0.01883137971162796, + 0.008769030682742596, + 0.007559523917734623, + 0.003908636048436165, + 0.0040792799554765224, + 0.003961122129112482, + 0.0026121397968381643, + 0.003688549157232046, + 0.004602694418281317, + 0.002936416305601597, + 0.004253700375556946, + 0.018837016075849533, + 0.029081853106617928, + 0.005387147422879934, + 0.0075669181533157825, + 0.006688229274004698, + 0.008881310001015663, + 0.005419801454991102, + 0.008006887510418892, + 0.006723715923726559, + 0.0064610643312335014, + 0.004481216426938772, + 0.004424467217177153, + 0.004603958688676357, + 0.004080301616340876, + 0.004324662033468485, + 0.004532326944172382, + 0.025726135820150375, + 0.09112894535064697, + 0.01359487697482109, + 0.014660860411822796, + 0.013123620301485062, + 0.02098751813173294, + 0.015404745936393738, + 0.08474399894475937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.05495734512805939 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989055812358856 + ], + "opposite_logprob_tokens": [ + -1.804890513420105 + ], + "logprob_reward": -0.17989055812358856, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989055812358856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21253979206085205, + 0.3114761710166931, + 0.05041613429784775, + 0.08500506728887558, + 0.06139029189944267, + 0.026058655232191086, + 0.017893243581056595, + 0.07797110825777054, + 0.04105379804968834, + 0.11619571596384048 + ], + "credit_map": [ + 0.02684820257127285, + 0.021177463233470917, + 0.02930864691734314, + 0.021599944680929184, + 0.03203713521361351, + 0.022467680275440216, + 0.023011917248368263, + 0.036088813096284866, + 0.04314295947551727, + 0.04579293727874756, + 0.05642431974411011, + 0.040825460106134415, + 0.03324427455663681, + 0.035011567175388336, + 0.03336172550916672, + 0.0236729234457016, + 0.007397213019430637, + 0.008527861908078194, + 0.008800904266536236, + 0.0068543218076229095, + 0.0045705013908445835, + 0.0060137431137263775, + 0.004674280993640423, + 0.003577310126274824, + 0.009857465513050556, + 0.014460322447121143, + 0.007666703779250383, + 0.006229339633136988, + 0.018784647807478905, + 0.01238090917468071, + 0.007436659652739763, + 0.008189022541046143, + 0.009314091876149178, + 0.012971153482794762, + 0.007971759885549545, + 0.007195051293820143, + 0.006467884406447411, + 0.006089808885008097, + 0.005233812611550093, + 0.006146733183413744, + 0.0037164848763495684, + 0.004580758512020111, + 0.0020284962374716997, + 0.0020965903531759977, + 0.004278738051652908, + 0.004846648313105106, + 0.0019361292943358421, + 0.0025748100597411394, + 0.002778718713670969, + 0.003332088701426983, + 0.00241972622461617, + 0.0014009118312969804, + 0.001713464967906475, + 0.002327867317944765, + 0.0019010165706276894, + 0.002019451931118965, + 0.016653968021273613, + 0.025674916803836823, + 0.005499147344380617, + 0.005527325440198183, + 0.004250700119882822, + 0.009478160180151463, + 0.00403011916205287, + 0.006856773514300585, + 0.00433468958362937, + 0.0048731425777077675, + 0.004454538691788912, + 0.00428417045623064, + 0.005404499359428883, + 0.004546631593257189, + 0.007192265707999468, + 0.005963862873613834, + 0.010790489614009857, + 0.03846552595496178, + 0.005981795955449343, + 0.006298697087913752, + 0.006458182353526354, + 0.008542378433048725, + 0.006906076800078154, + 0.03275258094072342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.17989055812358856 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot places the viewer in the midst of a primal showdown, depicting a ferocious Siberian tiger and a wild boar locked in a fierce tussle in a deciduous forest. The adult tiger, a massive beast, boasts a resplendent orange-yellow coat streaked with striking black stripes. Its shoulder and back muscles bulge with exertion, its maw gaping, revealing long, sharp canines and a silent roar. Its massive front paws, their nails fully developed, dig like hooks into the boar's thick, dark-brown, bristly flesh. Beneath the tiger's grasp, the equally formidable boar struggles frantically, its massive limbs pawing at the ground in an effort to free itself, twisting its mud-covered head, attempting to strike at the tiger's soft underbelly with its sharp, upturned tusks jutting from the corners of its mouth. Their brutal struggle unleashes a storm on the ground. Countless dry, fragile ochre and russet fallen leaves mingle with the black earth, hurled into the air by their immense force, forming a chaotic and dynamic vortex that flies across the canvas. The soft afternoon sunlight filters through the surrounding bare treetops, casting dappled shadows across the forest floor, perfectly illuminating the airy dust and leaf fragments, infusing the entire scene with an atmosphere of tension, chaos, and a savage, primal vitality.", + "seed": 32939, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_1/index_4023_seed_32939_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05016233865171671, + "vlm_logprob": -0.05016233865171671, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.05016233865171671, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08090727776288986 + ], + "opposite_logprob_tokens": [ + -2.7059073448181152 + ], + "logprob_reward": -0.08090727776288986, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08090727776288986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10949265211820602, + 0.10290661454200745, + 0.15440934896469116, + 0.09625794738531113, + 0.07858114689588547, + 0.052582982927560806, + 0.10704869776964188, + 0.08691243827342987, + 0.1267286241054535, + 0.08507953584194183 + ], + "credit_map": [ + 0.014879727736115456, + 0.017123371362686157, + 0.015191757120192051, + 0.015239356085658073, + 0.010234862565994263, + 0.01674736849963665, + 0.010735317133367062, + 0.009340895339846611, + 0.008405406959354877, + 0.011077331379055977, + 0.021506482735276222, + 0.008679334074258804, + 0.012599504552781582, + 0.01112055778503418, + 0.012838519178330898, + 0.01667947880923748, + 0.02159479446709156, + 0.02416490949690342, + 0.019139690324664116, + 0.014412397518754005, + 0.011157263070344925, + 0.02302667871117592, + 0.027732757851481438, + 0.013180849142372608, + 0.013944828882813454, + 0.01906104013323784, + 0.011250433512032032, + 0.01042261067777872, + 0.008606195449829102, + 0.011199709959328175, + 0.011745625175535679, + 0.010027502663433552, + 0.008440337143838406, + 0.010639559477567673, + 0.010039267130196095, + 0.007866112515330315, + 0.011873411014676094, + 0.009673184715211391, + 0.00796885322779417, + 0.012080422602593899, + 0.0056691477075219154, + 0.006550190970301628, + 0.007727359421551228, + 0.005697465501725674, + 0.006153969094157219, + 0.0080451350659132, + 0.0066984836012125015, + 0.006041231099516153, + 0.016886668279767036, + 0.016811611130833626, + 0.013042569160461426, + 0.012145232409238815, + 0.010883896611630917, + 0.011664394289255142, + 0.014207975007593632, + 0.011406349949538708, + 0.007744889706373215, + 0.01704397238790989, + 0.008230865001678467, + 0.009982693940401077, + 0.014954371377825737, + 0.007913068868219852, + 0.013488439843058586, + 0.007554136216640472, + 0.012985719367861748, + 0.01783282496035099, + 0.01918734796345234, + 0.013834803365170956, + 0.020632408559322357, + 0.010981234721839428, + 0.018139274790883064, + 0.013135010376572609, + 0.020729048177599907, + 0.010991407558321953, + 0.014910598285496235, + 0.009382576681673527, + 0.006652481853961945, + 0.006663328502327204, + 0.0061874124221503735, + 0.009562683291733265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.08090727776288986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004124349914491177 + ], + "opposite_logprob_tokens": [ + -5.504124164581299 + ], + "logprob_reward": -0.004124349914491177, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004124349914491177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1818827986717224, + 0.11580450087785721, + 0.07149770110845566, + 0.0915495902299881, + 0.16684186458587646, + 0.047275397926568985, + 0.06688878685235977, + 0.08247503638267517, + 0.0932520180940628, + 0.0825323760509491 + ], + "credit_map": [ + 0.021683722734451294, + 0.023020122200250626, + 0.025074850767850876, + 0.017919884994626045, + 0.021956011652946472, + 0.020068174228072166, + 0.03392404690384865, + 0.018235981464385986, + 0.012034575454890728, + 0.017123911529779434, + 0.014137350022792816, + 0.019838107749819756, + 0.015094546601176262, + 0.011382891796529293, + 0.013439422473311424, + 0.012753691524267197, + 0.01319104339927435, + 0.013241728767752647, + 0.008368073031306267, + 0.0055668349377810955, + 0.00619964674115181, + 0.011394662782549858, + 0.007568804081529379, + 0.005966911092400551, + 0.012251386418938637, + 0.014202188700437546, + 0.01077689416706562, + 0.011064575053751469, + 0.010216765105724335, + 0.010149373672902584, + 0.013200375251471996, + 0.009688024409115314, + 0.015445315279066563, + 0.03443966433405876, + 0.01918187364935875, + 0.015454483218491077, + 0.022328905761241913, + 0.020749161019921303, + 0.019930347800254822, + 0.01931210793554783, + 0.005580367054790258, + 0.006162144243717194, + 0.006494267378002405, + 0.005358387716114521, + 0.005311275832355022, + 0.005406395997852087, + 0.006886112503707409, + 0.006076447665691376, + 0.0067571913823485374, + 0.017323624342679977, + 0.008570129051804543, + 0.008578728884458542, + 0.0066551268100738525, + 0.007444680202752352, + 0.006533242762088776, + 0.005026055965572596, + 0.010537673719227314, + 0.016235103830695152, + 0.011705677025020123, + 0.009429406374692917, + 0.011064318008720875, + 0.006246919743716717, + 0.008015775121748447, + 0.00924015510827303, + 0.012335912324488163, + 0.0185933206230402, + 0.010410265997052193, + 0.008431787602603436, + 0.01611258089542389, + 0.007990928366780281, + 0.0099612046033144, + 0.009416013024747372, + 0.014773375354707241, + 0.010539033450186253, + 0.018892085179686546, + 0.010328546166419983, + 0.004716385155916214, + 0.007498959545046091, + 0.005295766517519951, + 0.010488219559192657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.004124349914491177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790142998099327 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.009790142998099327, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790142998099327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1673201620578766, + 0.10260618478059769, + 0.07789105176925659, + 0.0855376347899437, + 0.1975998431444168, + 0.0528031624853611, + 0.07465756684541702, + 0.047069791704416275, + 0.07650202512741089, + 0.11801251769065857 + ], + "credit_map": [ + 0.015236071310937405, + 0.027128126472234726, + 0.021945368498563766, + 0.021866295486688614, + 0.017256764695048332, + 0.0233831275254488, + 0.02717834711074829, + 0.01332606840878725, + 0.01223701797425747, + 0.018391694873571396, + 0.013898106291890144, + 0.011554669588804245, + 0.012152384035289288, + 0.00947638414800167, + 0.012565715238451958, + 0.012330212630331516, + 0.013645336031913757, + 0.014263037592172623, + 0.006011982448399067, + 0.007148029748350382, + 0.008033715188503265, + 0.01278179232031107, + 0.008220990188419819, + 0.007786173839122057, + 0.011609125882387161, + 0.016163580119609833, + 0.010323896072804928, + 0.007809199392795563, + 0.010241917334496975, + 0.010141164064407349, + 0.010873732157051563, + 0.008375021629035473, + 0.01861553266644478, + 0.03852704539895058, + 0.021252360194921494, + 0.0172575693577528, + 0.026727603748440742, + 0.02618894912302494, + 0.021781830117106438, + 0.027248961851000786, + 0.0053747957572340965, + 0.007556605152785778, + 0.008881165646016598, + 0.0053624254651367664, + 0.007773655001074076, + 0.0054535334929823875, + 0.006901324260979891, + 0.005499658640474081, + 0.007819239981472492, + 0.016935616731643677, + 0.00815522950142622, + 0.009562517516314983, + 0.008225946687161922, + 0.008270100690424442, + 0.009447758086025715, + 0.006241159979254007, + 0.0059708282351493835, + 0.01035531610250473, + 0.005475831683725119, + 0.0054866718128323555, + 0.005901101045310497, + 0.0036798587534576654, + 0.006148440297693014, + 0.004051744006574154, + 0.009233804419636726, + 0.010403317399322987, + 0.00810370035469532, + 0.008676660247147083, + 0.013091656379401684, + 0.007268922403454781, + 0.010573392733931541, + 0.009150572121143341, + 0.01545282918959856, + 0.014061682857573032, + 0.023236332461237907, + 0.015262274071574211, + 0.009446726180613041, + 0.011835315264761448, + 0.008053058758378029, + 0.02066430076956749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.009790142998099327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306385666131973 + ], + "opposite_logprob_tokens": [ + -3.1680638790130615 + ], + "logprob_reward": -0.04306385666131973, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306385666131973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2310616374015808, + 0.12629377841949463, + 0.17211118340492249, + 0.08392124623060226, + 0.038004614412784576, + 0.052396390587091446, + 0.05219428986310959, + 0.04513278231024742, + 0.06491801887750626, + 0.1339661180973053 + ], + "credit_map": [ + 0.019612250849604607, + 0.03754719719290733, + 0.028267258778214455, + 0.025294233113527298, + 0.02740391343832016, + 0.02549116127192974, + 0.037870459258556366, + 0.02957516349852085, + 0.011623349040746689, + 0.020004048943519592, + 0.013804945163428783, + 0.0157282967120409, + 0.01848604343831539, + 0.016057230532169342, + 0.014248738996684551, + 0.01634112000465393, + 0.02628418803215027, + 0.024453995749354362, + 0.016169961541891098, + 0.016763431951403618, + 0.013971877284348011, + 0.026170887053012848, + 0.02504866197705269, + 0.02324816584587097, + 0.011485553346574306, + 0.011106225661933422, + 0.008386611938476562, + 0.009692355059087276, + 0.007151416502892971, + 0.021582724526524544, + 0.008556901477277279, + 0.005959453992545605, + 0.005056316033005714, + 0.005512365605682135, + 0.0049334377981722355, + 0.004459408577531576, + 0.0038586759474128485, + 0.005970032420009375, + 0.004892313852906227, + 0.003322061151266098, + 0.008967065252363682, + 0.0060095456428825855, + 0.009283254854381084, + 0.005879121366888285, + 0.006537375506013632, + 0.006940704770386219, + 0.0038285870105028152, + 0.0049507287330925465, + 0.0057543860748410225, + 0.010234286077320576, + 0.00650409609079361, + 0.005889894440770149, + 0.00615700613707304, + 0.006027198862284422, + 0.006524439435452223, + 0.0051029762253165245, + 0.003959021531045437, + 0.008549581281840801, + 0.004473901353776455, + 0.004221173468977213, + 0.010032064281404018, + 0.004292501602321863, + 0.005070164799690247, + 0.0045343684032559395, + 0.00804033875465393, + 0.012033095583319664, + 0.007485722191631794, + 0.005990053992718458, + 0.011308743618428707, + 0.005722004920244217, + 0.007403015624731779, + 0.006935037672519684, + 0.032122597098350525, + 0.015979811549186707, + 0.029722323641180992, + 0.01664823107421398, + 0.007818995043635368, + 0.011987020261585712, + 0.006372380070388317, + 0.013314737007021904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.04306385666131973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level medium shot of a Siberian tiger and a wild boar locked in a struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's coat orange-yellow with black stripes, and are its shoulder and back muscles bulging with exertion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar dark-brown and bristly, with mud on its head and sharp upturned tusks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry ochre and russet leaves being hurled into the air by the struggle, forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight filtering through bare treetops, casting dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292606592178345 + ], + "opposite_logprob_tokens": [ + -2.2379260063171387 + ], + "logprob_reward": -0.11292606592178345, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292606592178345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16415603458881378, + 0.08752001076936722, + 0.09431905299425125, + 0.07508401572704315, + 0.2552790939807892, + 0.03548822179436684, + 0.06995172798633575, + 0.06224682182073593, + 0.057868730276823044, + 0.09808630496263504 + ], + "credit_map": [ + 0.019422942772507668, + 0.024206144735217094, + 0.021641340106725693, + 0.015019522048532963, + 0.02304796874523163, + 0.02024339698255062, + 0.02232564613223076, + 0.018249068409204483, + 0.009706590324640274, + 0.013857165351510048, + 0.010756277479231358, + 0.015200154855847359, + 0.01007116213440895, + 0.008349421434104443, + 0.010028275661170483, + 0.009550963528454304, + 0.013726378791034222, + 0.011393990367650986, + 0.010265834629535675, + 0.006863682065159082, + 0.008913400582969189, + 0.016072819009423256, + 0.017965149134397507, + 0.00911780633032322, + 0.008420553058385849, + 0.013170038349926472, + 0.009487430565059185, + 0.009726314805448055, + 0.007101380731910467, + 0.012716964818537235, + 0.007945628836750984, + 0.006515701301395893, + 0.02388462796807289, + 0.04671816900372505, + 0.03437411040067673, + 0.0421900600194931, + 0.027204180136322975, + 0.026541590690612793, + 0.029436448588967323, + 0.02492990344762802, + 0.00585950305685401, + 0.003985462244600058, + 0.005980250425636768, + 0.0035566457081586123, + 0.00361588504165411, + 0.005153903272002935, + 0.0034374420065432787, + 0.0038991309702396393, + 0.00719091622158885, + 0.020146330818533897, + 0.00732377776876092, + 0.008917214348912239, + 0.00663897767663002, + 0.008184188976883888, + 0.006595822051167488, + 0.004954502917826176, + 0.006744416430592537, + 0.01575244963169098, + 0.006824092473834753, + 0.008620048873126507, + 0.006317962892353535, + 0.005822167731821537, + 0.00664452463388443, + 0.00552115635946393, + 0.007466955576092005, + 0.009265710599720478, + 0.007143134716898203, + 0.005605451297014952, + 0.009104598313570023, + 0.004628645721822977, + 0.0059214141219854355, + 0.008732819929718971, + 0.01689193956553936, + 0.011174686253070831, + 0.02233152464032173, + 0.010794732719659805, + 0.005716785788536072, + 0.009333488531410694, + 0.006470851134508848, + 0.01537229958921671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.11292606592178345 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_tensor([5331, 5331])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_tensor([5331, 5331])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dd0b13120dc52c3a9c198cc5d6eed679d2d8f7be --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_tensor([5331, 5331])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, slightly overlooking, slowly pans to the right, following a cuddly giant panda as it clumsily swims deep within a lush, verdant bamboo forest. Soft sunlight filters through the dense foliage, casting dappled, vibrant light across the crystal-clear stream. The panda's signature thick black and white fur, soaked by the stream water, clings to its chubby body, creating a comical effect. Its chubby limbs shift uncoordinatedly in the water, each clumsy stroke a playful struggle against buoyancy rather than efficient propulsion. A look of both bewilderment and self-satisfaction permeates its face. As its body rises and falls, transparent ripples and countless crystalline droplets are stirred up. These droplets shimmer briefly in the sunlight before falling back into the water, creating a melodious sound in the tranquil forest. The stream is so clear that you can see the smooth pebbles at the bottom of the water. The banks are covered with moist moss and ferns. The whole scene is filled with a peaceful, natural and slightly humorous atmosphere in the main tones of vibrant emerald green and clear water, creating a pure and healing sense of picture.", + "seed": 24723, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_5331_seed_24723_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09499952867627144, + "vlm_logprob": -0.09499952867627144, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.09499952867627144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002519886940717697 + ], + "opposite_logprob_tokens": [ + -6.7525200843811035 + ], + "logprob_reward": -0.002519886940717697, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.002519886940717697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3200758993625641, + 0.056588560342788696, + 0.0816584900021553, + 0.04605211317539215, + 0.06386047601699829, + 0.05184192582964897, + 0.055947110056877136, + 0.06180208548903465, + 0.05602750554680824, + 0.20614579319953918 + ], + "credit_map": [ + 0.04178949072957039, + 0.030601371079683304, + 0.029796244576573372, + 0.021005837246775627, + 0.04140298813581467, + 0.02802669256925583, + 0.0416184663772583, + 0.08583485335111618, + 0.01106320507824421, + 0.01122540794312954, + 0.007861891761422157, + 0.0036728910636156797, + 0.005042747128754854, + 0.0051629068329930305, + 0.005080024711787701, + 0.0074794902466237545, + 0.013879011385142803, + 0.01613742671906948, + 0.008061067201197147, + 0.006078388076275587, + 0.005939694121479988, + 0.008018657565116882, + 0.006621242966502905, + 0.016922999173402786, + 0.011806090362370014, + 0.009318937547504902, + 0.004534621257334948, + 0.0033956253901124, + 0.0029583806172013283, + 0.007266071159392595, + 0.0031404010951519012, + 0.0036319857463240623, + 0.012329419143497944, + 0.008840021677315235, + 0.008156171068549156, + 0.005421299021691084, + 0.006993507035076618, + 0.007184973452240229, + 0.008983838371932507, + 0.005951241124421358, + 0.014124873094260693, + 0.008567176759243011, + 0.005090629681944847, + 0.0026660868898034096, + 0.003969830926507711, + 0.008850300684571266, + 0.0038968988228589296, + 0.004676129203289747, + 0.011401425115764141, + 0.006406247615814209, + 0.003549644723534584, + 0.006165866274386644, + 0.0037922237534075975, + 0.009126394987106323, + 0.006384516600519419, + 0.009120790287852287, + 0.01635395549237728, + 0.005587240215390921, + 0.0031730711925774813, + 0.004305812995880842, + 0.0035192403011024, + 0.00561782531440258, + 0.0029925762210041285, + 0.020252365618944168, + 0.008072310127317905, + 0.005479019600898027, + 0.0044798702001571655, + 0.007615056820213795, + 0.009437023662030697, + 0.007135752588510513, + 0.007159729953855276, + 0.006648744456470013, + 0.03147014230489731, + 0.008181996643543243, + 0.006193116307258606, + 0.01078040897846222, + 0.009383906610310078, + 0.06437511742115021, + 0.01864212565124035, + 0.057118989527225494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.002519886940717697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520902454853058 + ], + "opposite_logprob_tokens": [ + -1.5020902156829834 + ], + "logprob_reward": -0.2520902454853058, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520902454853058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02071145921945572, + 0.005212659947574139, + 0.010723874904215336, + 0.007851717993617058, + 0.007793029770255089, + 0.019711215049028397, + 0.007010238245129585, + 0.0034837934654206038, + 0.00317802163772285, + 0.9143239855766296 + ], + "credit_map": [ + 0.0026971837505698204, + 0.0038298512808978558, + 0.0019603136461228132, + 0.0016872627893462777, + 0.00211483146995306, + 0.00200817477889359, + 0.002720039105042815, + 0.003693799488246441, + 0.000745678145904094, + 0.000855370657518506, + 0.0008183245663531125, + 0.000464978045783937, + 0.0004681914288084954, + 0.0004725415783468634, + 0.0005405674455687404, + 0.0008470076718367636, + 0.0014681678730994463, + 0.0015318812802433968, + 0.0016491818241775036, + 0.0007705347961746156, + 0.0009444811730645597, + 0.0012101843021810055, + 0.0018075939733535051, + 0.0013418493326753378, + 0.0010483096120879054, + 0.0010675295488908887, + 0.0007338839350268245, + 0.0009030934306792915, + 0.0015400531701743603, + 0.000971021712757647, + 0.0007838174933567643, + 0.0008040087996050715, + 0.0009710563463158906, + 0.0010433822171762586, + 0.001082362956367433, + 0.000611301395110786, + 0.0006946514477021992, + 0.0008904401329346001, + 0.0016924983356148005, + 0.0008073369390331209, + 0.00593281164765358, + 0.003279826370999217, + 0.001042987685650587, + 0.0008812780142761767, + 0.000880252628121525, + 0.0059324405156075954, + 0.0007127757417038083, + 0.0010488416301086545, + 0.001368250115774572, + 0.0013650540495291352, + 0.00033491020440123975, + 0.0004368811205495149, + 0.00047734417603351176, + 0.001136633101850748, + 0.000691701250616461, + 0.001199463615193963, + 0.0006452419911511242, + 0.0004931664443574846, + 0.0002747905091382563, + 0.0005227613146416843, + 0.00028345242026261985, + 0.0003646745753940195, + 0.00038875435711815953, + 0.0005109517369419336, + 0.0005339265917427838, + 0.0003586831735447049, + 0.00025644086417742074, + 0.0003148430841974914, + 0.0004098723875358701, + 0.0006102218176238239, + 0.00041142525151371956, + 0.0002826082054525614, + 0.14020147919654846, + 0.01756656914949417, + 0.022798264399170876, + 0.03570650517940521, + 0.02013694867491722, + 0.32276827096939087, + 0.07306820899248123, + 0.28207775950431824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.2520902454853058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033710893243551254 + ], + "opposite_logprob_tokens": [ + -3.4087109565734863 + ], + "logprob_reward": -0.033710893243551254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710893243551254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06366145610809326, + 0.019790660589933395, + 0.02373303472995758, + 0.10028107464313507, + 0.02895081415772438, + 0.025600150227546692, + 0.010649902746081352, + 0.02544049732387066, + 0.02320854365825653, + 0.678683876991272 + ], + "credit_map": [ + 0.00897802971303463, + 0.012855136767029762, + 0.005641888827085495, + 0.006335099693387747, + 0.007635747082531452, + 0.005707127042114735, + 0.008732982911169529, + 0.007775448728352785, + 0.003868836909532547, + 0.003919346258044243, + 0.0028319591656327248, + 0.001024563447572291, + 0.0014374436577782035, + 0.0015051643131300807, + 0.0015619592741131783, + 0.0036413874477148056, + 0.002691979520022869, + 0.005145863629877567, + 0.002814176259562373, + 0.0014759490732103586, + 0.002245721872895956, + 0.0022445914801210165, + 0.00401846831664443, + 0.0030962841119617224, + 0.008506287820637226, + 0.0039307433180511, + 0.004260045010596514, + 0.004724695812910795, + 0.005322873126715422, + 0.031731016933918, + 0.005622811149805784, + 0.03618260473012924, + 0.004063315223902464, + 0.003821899415925145, + 0.0026957672089338303, + 0.002136818366125226, + 0.0030011069029569626, + 0.0031633828766644, + 0.005694111343473196, + 0.004374411888420582, + 0.00544685497879982, + 0.00339643401093781, + 0.0017962359124794602, + 0.0013330349465832114, + 0.0020685899071395397, + 0.005789915565401316, + 0.0031605963595211506, + 0.0026084890123456717, + 0.0016655869549140334, + 0.0013831096002832055, + 0.000952729897107929, + 0.00131740712095052, + 0.0009482685127295554, + 0.0014895785134285688, + 0.0014241551980376244, + 0.0014690678799524903, + 0.006679180543869734, + 0.0027989570517092943, + 0.000977319316007197, + 0.001953731756657362, + 0.0011202803580090404, + 0.002257023938000202, + 0.0012898186687380075, + 0.008364184759557247, + 0.004511603619903326, + 0.0017436080379411578, + 0.0015222299844026566, + 0.002079347614198923, + 0.0023723330814391375, + 0.0037606144323945045, + 0.0023318680468946695, + 0.00488694105297327, + 0.10470326989889145, + 0.012603026814758778, + 0.017071278765797615, + 0.02563614770770073, + 0.016575872898101807, + 0.2406211495399475, + 0.05232032388448715, + 0.20915278792381287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.033710893243551254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032961010932922 + ], + "opposite_logprob_tokens": [ + -1.9103295803070068 + ], + "logprob_reward": -0.16032961010932922, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032961010932922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10144232958555222, + 0.013941851444542408, + 0.04146137461066246, + 0.12833404541015625, + 0.06476001441478729, + 0.046440962702035904, + 0.040164053440093994, + 0.14055581390857697, + 0.07089085131883621, + 0.3520087003707886 + ], + "credit_map": [ + 0.012609761208295822, + 0.01595139689743519, + 0.009055141359567642, + 0.007563594728708267, + 0.014013820327818394, + 0.009427465498447418, + 0.016174547374248505, + 0.016646603122353554, + 0.0018232641741633415, + 0.002027172828093171, + 0.0029861966613680124, + 0.001143752015195787, + 0.0011197711573913693, + 0.0012572979321703315, + 0.0012046173214912415, + 0.002379779005423188, + 0.005287090316414833, + 0.009691962040960789, + 0.003845256520435214, + 0.0020706080831587315, + 0.0030236877501010895, + 0.0035183264408260584, + 0.004684671759605408, + 0.009339767508208752, + 0.01194732915610075, + 0.007868521846830845, + 0.005963397677987814, + 0.004388044588267803, + 0.005178161431103945, + 0.039180465042591095, + 0.007934263907372952, + 0.045873869210481644, + 0.01334910374134779, + 0.013743403367698193, + 0.003969208803027868, + 0.004456574097275734, + 0.005573133006691933, + 0.010569406673312187, + 0.006631460040807724, + 0.006467725150287151, + 0.013596187345683575, + 0.009545979090034962, + 0.0032279808074235916, + 0.0020842032972723246, + 0.0035509036388248205, + 0.007660947274416685, + 0.0036648085806518793, + 0.0031099566258490086, + 0.006717328447848558, + 0.007321527693420649, + 0.0029961555264890194, + 0.004911326337605715, + 0.0027450413908809423, + 0.005786639638245106, + 0.00436043506488204, + 0.005325598642230034, + 0.02579789236187935, + 0.009663676843047142, + 0.009455210529267788, + 0.01366645097732544, + 0.018338585272431374, + 0.014977245591580868, + 0.01501226145774126, + 0.03364448621869087, + 0.010585000738501549, + 0.005679637659341097, + 0.003913756925612688, + 0.006336275953799486, + 0.011519070714712143, + 0.011436061933636665, + 0.011431979946792126, + 0.009989067912101746, + 0.05326467007398605, + 0.007859456352889538, + 0.00897421408444643, + 0.0150279076769948, + 0.008063343353569508, + 0.12203941494226456, + 0.02959912270307541, + 0.10718058049678802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.16032961010932922 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347007602453232 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.026347007602453232, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347007602453232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0710626021027565, + 0.013932278379797935, + 0.045762546360492706, + 0.02052338980138302, + 0.06681755185127258, + 0.18182016909122467, + 0.0318194180727005, + 0.034010712057352066, + 0.018478738144040108, + 0.5157726407051086 + ], + "credit_map": [ + 0.010927784256637096, + 0.010611983016133308, + 0.004874628037214279, + 0.00506780156865716, + 0.009921804070472717, + 0.00678638881072402, + 0.00916681531816721, + 0.01370539516210556, + 0.0022736527025699615, + 0.002160795032978058, + 0.0027827660087496042, + 0.001001743134111166, + 0.0011823102831840515, + 0.0012261451920494437, + 0.0010644177673384547, + 0.002240448724478483, + 0.007932888343930244, + 0.009060650132596493, + 0.0059778024442493916, + 0.002995492657646537, + 0.002929724520072341, + 0.004128189757466316, + 0.0038602924905717373, + 0.008877507410943508, + 0.0031648585572838783, + 0.0032787371892482042, + 0.0018695449689403176, + 0.0016524443635717034, + 0.002091530244797468, + 0.0031512808054685593, + 0.0029083837289363146, + 0.002406609943136573, + 0.015236170031130314, + 0.013917043805122375, + 0.004466860555112362, + 0.004660324659198523, + 0.0043098838068544865, + 0.011745112016797066, + 0.004698670003563166, + 0.007783480919897556, + 0.0543416328728199, + 0.0294721070677042, + 0.00837869942188263, + 0.008365733548998833, + 0.007735299412161112, + 0.05971445143222809, + 0.005311975721269846, + 0.008500291034579277, + 0.0077474117279052734, + 0.00537037244066596, + 0.0017810334684327245, + 0.0026004540268331766, + 0.0015319683589041233, + 0.005959704518318176, + 0.0032124158460646868, + 0.00361605710349977, + 0.007095335982739925, + 0.006378458812832832, + 0.0018204896477982402, + 0.003062609815970063, + 0.0035856289323419333, + 0.004586480557918549, + 0.0028966667596250772, + 0.004585040267556906, + 0.001969896722584963, + 0.0016144593246281147, + 0.0020716371946036816, + 0.0035917628556489944, + 0.0023259357549250126, + 0.0018485162872821093, + 0.0023397670593112707, + 0.0027167622465640306, + 0.07923789322376251, + 0.009142044931650162, + 0.01281407754868269, + 0.0191966500133276, + 0.012090158648788929, + 0.1838131994009018, + 0.03934900835156441, + 0.1601296067237854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.026347007602453232 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, slightly overlooking, slowly pans to the right, following a cuddly giant panda as it clumsily swims deep within a lush, verdant bamboo forest. Soft sunlight filters through the dense foliage, casting dappled, vibrant light across the crystal-clear stream. The panda's signature thick black and white fur, soaked by the stream water, clings to its chubby body, creating a comical effect. Its chubby limbs shift uncoordinatedly in the water, each clumsy stroke a playful struggle against buoyancy rather than efficient propulsion. A look of both bewilderment and self-satisfaction permeates its face. As its body rises and falls, transparent ripples and countless crystalline droplets are stirred up. These droplets shimmer briefly in the sunlight before falling back into the water, creating a melodious sound in the tranquil forest. The stream is so clear that you can see the smooth pebbles at the bottom of the water. The banks are covered with moist moss and ferns. The whole scene is filled with a peaceful, natural and slightly humorous atmosphere in the main tones of vibrant emerald green and clear water, creating a pure and healing sense of picture.", + "seed": 24724, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_5331_seed_24724_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.20578711684793233, + "vlm_logprob": -0.20578711684793233, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.20578711684793233, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002155839465558529 + ], + "opposite_logprob_tokens": [ + -6.627155780792236 + ], + "logprob_reward": -0.002155839465558529, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.002155839465558529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1764267534017563, + 0.05652354285120964, + 0.14758837223052979, + 0.1377597153186798, + 0.2276417315006256, + 0.05303700268268585, + 0.04177749902009964, + 0.03204329311847687, + 0.04627428948879242, + 0.08092779666185379 + ], + "credit_map": [ + 0.03580959141254425, + 0.011215124279260635, + 0.01607540249824524, + 0.01316828653216362, + 0.0154805239289999, + 0.03233398124575615, + 0.017836229875683784, + 0.0345076285302639, + 0.010291806422173977, + 0.004950508009642363, + 0.004768251907080412, + 0.006075586657971144, + 0.004560389090329409, + 0.011793163605034351, + 0.008563553914427757, + 0.005520283244550228, + 0.023753006011247635, + 0.013449211604893208, + 0.01574033312499523, + 0.017318274825811386, + 0.014140051789581776, + 0.026391202583909035, + 0.011612332426011562, + 0.025183970108628273, + 0.020761851221323013, + 0.014790165238082409, + 0.03003959357738495, + 0.01288970559835434, + 0.009969839826226234, + 0.02237432263791561, + 0.009964427910745144, + 0.016969824209809303, + 0.03625195100903511, + 0.025149652734398842, + 0.039870280772447586, + 0.021668236702680588, + 0.019723784178495407, + 0.03246130794286728, + 0.019122367724776268, + 0.03339417278766632, + 0.008896080777049065, + 0.005963302683085203, + 0.004585429560393095, + 0.0047059147618710995, + 0.005185492802411318, + 0.010140810161828995, + 0.005300804506987333, + 0.00825916975736618, + 0.009405684657394886, + 0.0026237284764647484, + 0.005433231126517057, + 0.003725468646734953, + 0.002033516764640808, + 0.010982447303831577, + 0.0032792426645755768, + 0.004294178914278746, + 0.0055971574038267136, + 0.003527230815961957, + 0.0026646412443369627, + 0.002945082262158394, + 0.004185042344033718, + 0.0066854809410870075, + 0.002951592206954956, + 0.0034870668314397335, + 0.006574504543095827, + 0.006959653925150633, + 0.003929717931896448, + 0.005212619435042143, + 0.0052819326519966125, + 0.006624066736549139, + 0.007179287262260914, + 0.004512506537139416, + 0.014766085892915726, + 0.01328839547932148, + 0.007694660685956478, + 0.007523605599999428, + 0.007095508277416229, + 0.009870407171547413, + 0.012055067345499992, + 0.008634068071842194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.002155839465558529 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233070850372314 + ], + "opposite_logprob_tokens": [ + -0.8983070850372314 + ], + "logprob_reward": -0.5233070850372314, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233070850372314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7503629922866821, + 0.02650919370353222, + 0.04931255057454109, + 0.02427314966917038, + 0.048875659704208374, + 0.018865317106246948, + 0.01328396238386631, + 0.006565113086253405, + 0.03396810218691826, + 0.02798396162688732 + ], + "credit_map": [ + 0.29358333349227905, + 0.022963644936680794, + 0.017884552478790283, + 0.01818130724132061, + 0.0378153994679451, + 0.32228052616119385, + 0.017190124839544296, + 0.020464075729250908, + 0.0037328239995986223, + 0.00348040577955544, + 0.0021870711352676153, + 0.004132856614887714, + 0.003695100313052535, + 0.0038736460264772177, + 0.0031612436287105083, + 0.0022460443433374166, + 0.01040393766015768, + 0.003927601035684347, + 0.0049316356889903545, + 0.004260099958628416, + 0.0061081405729055405, + 0.008196315728127956, + 0.00610280130058527, + 0.00538201630115509, + 0.003631133818998933, + 0.002128435065969825, + 0.004540270194411278, + 0.0022303867153823376, + 0.0019112202571704984, + 0.004754830617457628, + 0.002265491057187319, + 0.0028113822918385267, + 0.007995545864105225, + 0.005597891751676798, + 0.008455006405711174, + 0.004028320778161287, + 0.00422128988429904, + 0.00817324873059988, + 0.0037692252080887556, + 0.006635131314396858, + 0.0028142742812633514, + 0.0021522887982428074, + 0.001670960569754243, + 0.0017173716332763433, + 0.0015811881748959422, + 0.0033779696095734835, + 0.002260383451357484, + 0.003290879540145397, + 0.002314454410225153, + 0.000956831849180162, + 0.0013843415072187781, + 0.0013829284580424428, + 0.0013153923209756613, + 0.0026730787940323353, + 0.0012009857455268502, + 0.0020559504628181458, + 0.0008458683732897043, + 0.0008401736267842352, + 0.0006296008359640837, + 0.0005361185176298022, + 0.0008384371176362038, + 0.0008394162287004292, + 0.0014909774763509631, + 0.0005445206770673394, + 0.004557314794510603, + 0.004537701141089201, + 0.0035224631428718567, + 0.003881421871483326, + 0.0034182434901595116, + 0.0038204090669751167, + 0.0050313654355704784, + 0.005199180915951729, + 0.004621042869985104, + 0.002243111142888665, + 0.002507186494767666, + 0.002687244676053524, + 0.0036700661294162273, + 0.004106609150767326, + 0.004665210377424955, + 0.0034834900870919228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.5233070850372314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158728957176208 + ], + "opposite_logprob_tokens": [ + -1.7015873193740845 + ], + "logprob_reward": -0.20158728957176208, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158728957176208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5067510008811951, + 0.09308970719575882, + 0.08997014909982681, + 0.026967495679855347, + 0.0682644173502922, + 0.032341018319129944, + 0.029286067932844162, + 0.036982808262109756, + 0.06706787645816803, + 0.049279481172561646 + ], + "credit_map": [ + 0.188340961933136, + 0.01540563628077507, + 0.017139647156000137, + 0.020031491294503212, + 0.02543979324400425, + 0.20198173820972443, + 0.017922155559062958, + 0.020489560440182686, + 0.015354341827332973, + 0.01170769240707159, + 0.006534191779792309, + 0.014721072278916836, + 0.012089135125279427, + 0.013637377880513668, + 0.012203672900795937, + 0.006842218339443207, + 0.018896548077464104, + 0.007006015162914991, + 0.00971249956637621, + 0.009050306864082813, + 0.010436102747917175, + 0.01363440789282322, + 0.011029483750462532, + 0.010204784572124481, + 0.0039109159260988235, + 0.002231351798400283, + 0.0038074548356235027, + 0.002342863939702511, + 0.0024852061178535223, + 0.004103657323867083, + 0.0035330415703356266, + 0.004553002305328846, + 0.011099476367235184, + 0.008165809325873852, + 0.011309881694614887, + 0.005821203347295523, + 0.006026288960129023, + 0.011218775995075703, + 0.00572264613583684, + 0.008900338783860207, + 0.004939883016049862, + 0.004015350714325905, + 0.003172154538333416, + 0.003393215360119939, + 0.0029532448388636112, + 0.004977973643690348, + 0.004215277265757322, + 0.00467391824349761, + 0.004662637133151293, + 0.0018339591333642602, + 0.004232389386743307, + 0.0028031908441334963, + 0.0023622389417141676, + 0.0055098882876336575, + 0.0033607669174671173, + 0.004520999733358622, + 0.009292617440223694, + 0.002852140460163355, + 0.002319170394912362, + 0.0023434460163116455, + 0.0033963844180107117, + 0.0041562472470104694, + 0.003702777437865734, + 0.00892002321779728, + 0.009572039358317852, + 0.008212940767407417, + 0.006475370842963457, + 0.008658775128424168, + 0.007451300974935293, + 0.006427389103919268, + 0.012627305462956429, + 0.007642753887921572, + 0.00920862890779972, + 0.005429313983768225, + 0.003559073433279991, + 0.00444151321426034, + 0.004571808502078056, + 0.010594949126243591, + 0.0050488910637795925, + 0.006425305735319853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20158728957176208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28129804134368896 + ], + "opposite_logprob_tokens": [ + -1.406298041343689 + ], + "logprob_reward": -0.28129804134368896, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28129804134368896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33650702238082886, + 0.0919865220785141, + 0.09118520468473434, + 0.051117800176143646, + 0.06314300000667572, + 0.027201585471630096, + 0.02660449407994747, + 0.0794743001461029, + 0.049868449568748474, + 0.18291161954402924 + ], + "credit_map": [ + 0.13045690953731537, + 0.010316003113985062, + 0.007673996966332197, + 0.006856971420347691, + 0.016979387030005455, + 0.14511622488498688, + 0.007777828723192215, + 0.011329707689583302, + 0.013316147960722446, + 0.009926238097250462, + 0.007743074558675289, + 0.013970455154776573, + 0.010532401502132416, + 0.013261502608656883, + 0.013378765434026718, + 0.009857939556241035, + 0.014544788748025894, + 0.008367029950022697, + 0.008644506335258484, + 0.012475312687456608, + 0.010743538849055767, + 0.015172957442700863, + 0.007858438417315483, + 0.013378636911511421, + 0.008472614921629429, + 0.00617410521954298, + 0.007584226783365011, + 0.006090085953474045, + 0.004732935223728418, + 0.008464259095489979, + 0.00514592370018363, + 0.004453644622117281, + 0.009947878308594227, + 0.0075121656991541386, + 0.009943663142621517, + 0.006069869734346867, + 0.005840370897203684, + 0.00915414560586214, + 0.005775767844170332, + 0.008899137377738953, + 0.004098793957382441, + 0.0029909650329500437, + 0.003937286324799061, + 0.0027385877911001444, + 0.0020698984153568745, + 0.003619717201218009, + 0.0031364804599434137, + 0.004609855357557535, + 0.003590972861275077, + 0.0024722558446228504, + 0.003106321208178997, + 0.002375116338953376, + 0.0027563085313886404, + 0.004835409577935934, + 0.002675502561032772, + 0.004792605526745319, + 0.010482927784323692, + 0.00938404817134142, + 0.006550380028784275, + 0.008068599738180637, + 0.011742496863007545, + 0.011681756936013699, + 0.014833766967058182, + 0.0067303236573934555, + 0.007407207041978836, + 0.006658700294792652, + 0.004746878519654274, + 0.005959718022495508, + 0.0055552818812429905, + 0.006302242632955313, + 0.007838129065930843, + 0.005400289315730333, + 0.024637268856167793, + 0.0175746139138937, + 0.025079168379306793, + 0.019135836511850357, + 0.025457242503762245, + 0.027276158332824707, + 0.024676987901329994, + 0.019074341282248497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.28129804134368896 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058732882142067 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.02058732882142067, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058732882142067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10602565854787827, + 0.11165650188922882, + 0.08091770112514496, + 0.13270074129104614, + 0.25864797830581665, + 0.14529924094676971, + 0.02026611566543579, + 0.053609445691108704, + 0.045112572610378265, + 0.04576407000422478 + ], + "credit_map": [ + 0.023925036191940308, + 0.00611645495519042, + 0.011300616897642612, + 0.00852243509143591, + 0.007717386353760958, + 0.020936282351613045, + 0.011582739651203156, + 0.015924707055091858, + 0.01637398451566696, + 0.014132699929177761, + 0.00863252580165863, + 0.01783888228237629, + 0.011533292941749096, + 0.018318790942430496, + 0.015943467617034912, + 0.008882852271199226, + 0.01736486330628395, + 0.00776574807241559, + 0.007874471135437489, + 0.007753010373562574, + 0.009012741036713123, + 0.013334445655345917, + 0.007986975833773613, + 0.009825449436903, + 0.01973334699869156, + 0.015553310513496399, + 0.024886012077331543, + 0.016141749918460846, + 0.008592710830271244, + 0.02370264194905758, + 0.009839767590165138, + 0.014251185581088066, + 0.04345664754509926, + 0.02249256707727909, + 0.04043536260724068, + 0.024233827367424965, + 0.019886977970600128, + 0.04634851589798927, + 0.02237832359969616, + 0.039415761828422546, + 0.026358388364315033, + 0.011865600012242794, + 0.00975328404456377, + 0.008181620389223099, + 0.01045197807252407, + 0.04585627093911171, + 0.014714744873344898, + 0.018117347732186317, + 0.003821177640929818, + 0.001488840556703508, + 0.0021096537820994854, + 0.002076242584735155, + 0.001479683443903923, + 0.0043855588883161545, + 0.0017173935193568468, + 0.003187565365806222, + 0.013278622180223465, + 0.0054530189372599125, + 0.0048776124604046345, + 0.0038552682381123304, + 0.003973331302404404, + 0.005239388905465603, + 0.0050715855322778225, + 0.01186061929911375, + 0.006265489850193262, + 0.005095973145216703, + 0.0038053011521697044, + 0.005176273640245199, + 0.006276766303926706, + 0.00512863602489233, + 0.008641491644084454, + 0.0047226399183273315, + 0.007916075177490711, + 0.005083632655441761, + 0.00505826435983181, + 0.004854757804423571, + 0.004031550604850054, + 0.009247293695807457, + 0.005139587447047234, + 0.004432912450283766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.02058732882142067 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, slightly overlooking, slowly pans to the right, following a cuddly giant panda as it clumsily swims deep within a lush, verdant bamboo forest. Soft sunlight filters through the dense foliage, casting dappled, vibrant light across the crystal-clear stream. The panda's signature thick black and white fur, soaked by the stream water, clings to its chubby body, creating a comical effect. Its chubby limbs shift uncoordinatedly in the water, each clumsy stroke a playful struggle against buoyancy rather than efficient propulsion. A look of both bewilderment and self-satisfaction permeates its face. As its body rises and falls, transparent ripples and countless crystalline droplets are stirred up. These droplets shimmer briefly in the sunlight before falling back into the water, creating a melodious sound in the tranquil forest. The stream is so clear that you can see the smooth pebbles at the bottom of the water. The banks are covered with moist moss and ferns. The whole scene is filled with a peaceful, natural and slightly humorous atmosphere in the main tones of vibrant emerald green and clear water, creating a pure and healing sense of picture.", + "seed": 24725, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_5331_seed_24725_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08884235564619303, + "vlm_logprob": -0.08884235564619303, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.08884235564619303, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005436279810965061 + ], + "opposite_logprob_tokens": [ + -5.505436420440674 + ], + "logprob_reward": -0.005436279810965061, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005436279810965061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30947640538215637, + 0.07173902541399002, + 0.04938631132245064, + 0.10927257686853409, + 0.03325332701206207, + 0.02211935818195343, + 0.17619512975215912, + 0.0246618390083313, + 0.08312196284532547, + 0.12077407538890839 + ], + "credit_map": [ + 0.04940825328230858, + 0.044202715158462524, + 0.02293962426483631, + 0.02846415527164936, + 0.026566412299871445, + 0.04912266135215759, + 0.04476001113653183, + 0.04401257261633873, + 0.01118494663387537, + 0.009822985157370567, + 0.0055155144073069096, + 0.007431735750287771, + 0.008667150512337685, + 0.01342832762748003, + 0.008765689097344875, + 0.006922679487615824, + 0.008643499575555325, + 0.006979081314057112, + 0.004600189626216888, + 0.0047179944813251495, + 0.006570606958121061, + 0.006796799134463072, + 0.0058466712944209576, + 0.005231468006968498, + 0.01663760282099247, + 0.012066120281815529, + 0.014406295493245125, + 0.01721283607184887, + 0.008820106275379658, + 0.009813617914915085, + 0.017459766939282417, + 0.012856231071054935, + 0.004803574178367853, + 0.005509431939572096, + 0.0037531687412410975, + 0.003672987688332796, + 0.00412300368770957, + 0.004687915090471506, + 0.002598478691652417, + 0.004104766063392162, + 0.003143352223560214, + 0.0023954210337251425, + 0.004984752275049686, + 0.002624764107167721, + 0.002312221098691225, + 0.002972454996779561, + 0.0017290085088461637, + 0.001957384403795004, + 0.060612455010414124, + 0.010970128700137138, + 0.004981782287359238, + 0.0040054889395833015, + 0.008873917162418365, + 0.07429644465446472, + 0.00601970823481679, + 0.0064352136105299, + 0.0038715186528861523, + 0.0060319481417536736, + 0.0030968952924013138, + 0.0021423043217509985, + 0.0023998082615435123, + 0.0022911352571099997, + 0.001978582236915827, + 0.002849648240953684, + 0.012948479503393173, + 0.012520331889390945, + 0.010145691223442554, + 0.009538558311760426, + 0.008721736259758472, + 0.00952253956347704, + 0.0076205190271139145, + 0.012104107066988945, + 0.028686532750725746, + 0.010350039228796959, + 0.008686104789376259, + 0.039892762899398804, + 0.006342512089759111, + 0.011522378772497177, + 0.008416182361543179, + 0.006877559702843428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.005436279810965061 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210803747177124 + ], + "opposite_logprob_tokens": [ + -1.502108097076416 + ], + "logprob_reward": -0.25210803747177124, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210803747177124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2510245740413666, + 0.051685865968465805, + 0.15437088906764984, + 0.04471264034509659, + 0.035446714609861374, + 0.043846648186445236, + 0.2573254704475403, + 0.016895946115255356, + 0.04344988986849785, + 0.1012413501739502 + ], + "credit_map": [ + 0.039362818002700806, + 0.037307146936655045, + 0.01692494936287403, + 0.03444864973425865, + 0.017390627413988113, + 0.04104520380496979, + 0.037110667675733566, + 0.027434512972831726, + 0.008457177318632603, + 0.008944915607571602, + 0.0035942706745117903, + 0.007095364388078451, + 0.0036515400279313326, + 0.010911214165389538, + 0.0044412813149392605, + 0.004590103402733803, + 0.030527351424098015, + 0.028965070843696594, + 0.0070250993594527245, + 0.007905104197561741, + 0.011779550462961197, + 0.04055535048246384, + 0.010958956554532051, + 0.016654398292303085, + 0.007030893582850695, + 0.005469392519444227, + 0.004892149940133095, + 0.008360957726836205, + 0.004140530712902546, + 0.004167506471276283, + 0.006486000493168831, + 0.004165208898484707, + 0.004922741558402777, + 0.005271426402032375, + 0.003484440268948674, + 0.0027892275247722864, + 0.004862721543759108, + 0.005071481689810753, + 0.004510823171585798, + 0.004533855710178614, + 0.007030182052403688, + 0.0032682365272194147, + 0.0066010053269565105, + 0.005465875845402479, + 0.0033415593206882477, + 0.012613845989108086, + 0.0026884570252150297, + 0.0028374867979437113, + 0.08482000231742859, + 0.01675395481288433, + 0.00685134856030345, + 0.006692368071526289, + 0.013642781414091587, + 0.10983845591545105, + 0.00933236163109541, + 0.00939419399946928, + 0.0026637972332537174, + 0.0030780392698943615, + 0.0021140414755791426, + 0.0013950933935120702, + 0.001661776565015316, + 0.0017786407843232155, + 0.001586564932949841, + 0.002617992227897048, + 0.007455949205905199, + 0.010533386841416359, + 0.004035193007439375, + 0.003737077349796891, + 0.006515998858958483, + 0.0037444913759827614, + 0.003162093460559845, + 0.004265700466930866, + 0.023839756846427917, + 0.007035139016807079, + 0.007088304962962866, + 0.03659706935286522, + 0.003872890491038561, + 0.009426495991647243, + 0.007451090030372143, + 0.005930603481829166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.25210803747177124 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330111712217331 + ], + "opposite_logprob_tokens": [ + -3.773301124572754 + ], + "logprob_reward": -0.02330111712217331, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330111712217331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20317797362804413, + 0.06569495797157288, + 0.05843488872051239, + 0.0391671396791935, + 0.029136255383491516, + 0.07275848835706711, + 0.33872514963150024, + 0.01740618795156479, + 0.1183202862739563, + 0.05717867240309715 + ], + "credit_map": [ + 0.03330371901392937, + 0.026595573872327805, + 0.016707520931959152, + 0.03442462161183357, + 0.015548363327980042, + 0.023564817383885384, + 0.03920988366007805, + 0.013823471963405609, + 0.01057933084666729, + 0.016078663989901543, + 0.0048740035854279995, + 0.006652249954640865, + 0.005325059872120619, + 0.008177881129086018, + 0.007823334075510502, + 0.006184441968798637, + 0.010243825614452362, + 0.009209291078150272, + 0.00524059496819973, + 0.00435755355283618, + 0.00503623578697443, + 0.01216148305684328, + 0.005004911683499813, + 0.007180997170507908, + 0.006342953070998192, + 0.006616540253162384, + 0.004847963806241751, + 0.002724749967455864, + 0.003657326102256775, + 0.004588637501001358, + 0.005410983692854643, + 0.0049779885448515415, + 0.003974061459302902, + 0.00433188583701849, + 0.004190259613096714, + 0.0030734592583030462, + 0.0024851833004504442, + 0.004639447201043367, + 0.0038250235375016928, + 0.0026169363409280777, + 0.01358234602957964, + 0.005877047777175903, + 0.00769071327522397, + 0.006353015545755625, + 0.006174391135573387, + 0.02527359314262867, + 0.0030451654456555843, + 0.004762222524732351, + 0.11395604908466339, + 0.021170375868678093, + 0.009038117714226246, + 0.007647314108908176, + 0.017611557617783546, + 0.14490313827991486, + 0.011964973993599415, + 0.01243363507091999, + 0.0025078612379729748, + 0.002150688087567687, + 0.0030868765898048878, + 0.0020314278081059456, + 0.0018892157822847366, + 0.0019528706325218081, + 0.0017167176119983196, + 0.0020705307833850384, + 0.016674231737852097, + 0.023938437923789024, + 0.01346425712108612, + 0.011866319924592972, + 0.01854037493467331, + 0.012897871434688568, + 0.011858678422868252, + 0.009080126881599426, + 0.007796486373990774, + 0.00850820355117321, + 0.006155523005872965, + 0.0060330843552947044, + 0.004822005517780781, + 0.006332436576485634, + 0.0076944949105381966, + 0.009836443699896336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.02330111712217331 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278298616409302 + ], + "opposite_logprob_tokens": [ + -2.0177829265594482 + ], + "logprob_reward": -0.14278298616409302, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278298616409302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15845221281051636, + 0.08213359862565994, + 0.055635593831539154, + 0.10999441891908646, + 0.0475044772028923, + 0.06928583979606628, + 0.3291928470134735, + 0.03369869664311409, + 0.04461774230003357, + 0.06948453187942505 + ], + "credit_map": [ + 0.024290001019835472, + 0.017460724338889122, + 0.01604543998837471, + 0.029772695153951645, + 0.009716879576444626, + 0.021352974697947502, + 0.024437397718429565, + 0.015376096591353416, + 0.011251179501414299, + 0.010836780071258545, + 0.007725869305431843, + 0.012043100781738758, + 0.008711645379662514, + 0.011875818483531475, + 0.01045738160610199, + 0.009231830015778542, + 0.008306523784995079, + 0.00884284544736147, + 0.005584062077105045, + 0.004884987138211727, + 0.006986788474023342, + 0.007354742847383022, + 0.007448770105838776, + 0.006226878147572279, + 0.017438624054193497, + 0.012815727852284908, + 0.01361137256026268, + 0.019975215196609497, + 0.00868795532733202, + 0.01057341042906046, + 0.012405089102685452, + 0.014487035572528839, + 0.007464354857802391, + 0.006203935015946627, + 0.0062384940683841705, + 0.0042887902818620205, + 0.004197907634079456, + 0.0076470389030873775, + 0.004627642221748829, + 0.00683631282299757, + 0.01165318675339222, + 0.010630620643496513, + 0.009036785922944546, + 0.006794264540076256, + 0.00478422362357378, + 0.018194468691945076, + 0.0033337243366986513, + 0.0048585692420601845, + 0.10810431838035583, + 0.021087778732180595, + 0.009080502204596996, + 0.008824965916574001, + 0.01703779585659504, + 0.141015887260437, + 0.012019169516861439, + 0.012022445909678936, + 0.006612427532672882, + 0.008072171360254288, + 0.004211613442748785, + 0.002466255333274603, + 0.0034556512255221605, + 0.0029335080180317163, + 0.0029919794760644436, + 0.0029550918843597174, + 0.00801174994558096, + 0.0072237118147313595, + 0.004806320182979107, + 0.0036266872193664312, + 0.008443640545010567, + 0.004309319891035557, + 0.004717528820037842, + 0.0034787931945174932, + 0.010029220022261143, + 0.008826103061437607, + 0.007403384428471327, + 0.013585022650659084, + 0.006195616442710161, + 0.006130996625870466, + 0.010146783664822578, + 0.007167411502450705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.14278298616409302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10877007246017456, + 0.10891043394804001, + 0.0993032157421112, + 0.04567934200167656, + 0.06390689313411713, + 0.05173991620540619, + 0.36038151383399963, + 0.018904980272054672, + 0.11521389335393906, + 0.027189821004867554 + ], + "credit_map": [ + 0.017833394929766655, + 0.009161424823105335, + 0.009009301662445068, + 0.00966549851000309, + 0.009107674472033978, + 0.023998213931918144, + 0.01570141687989235, + 0.0142931267619133, + 0.016838911920785904, + 0.013819238170981407, + 0.00932613480836153, + 0.014916195534169674, + 0.011598893441259861, + 0.015765488147735596, + 0.015429397113621235, + 0.011216158978641033, + 0.01849946565926075, + 0.02413884364068508, + 0.0065586157143116, + 0.0062515209428966045, + 0.008555000647902489, + 0.017842920497059822, + 0.007387351244688034, + 0.010069483891129494, + 0.007768315728753805, + 0.009467227384448051, + 0.0037708687596023083, + 0.0053239851258695126, + 0.004666774068027735, + 0.005791326053440571, + 0.006096852011978626, + 0.0027939872816205025, + 0.008546868339180946, + 0.00739977927878499, + 0.008874096907675266, + 0.005661746021360159, + 0.005177007522433996, + 0.011602682061493397, + 0.0063892449252307415, + 0.01025545597076416, + 0.010272708721458912, + 0.0065245237201452255, + 0.004607195500284433, + 0.002922122133895755, + 0.004257413558661938, + 0.018428046256303787, + 0.0019816525746136904, + 0.0027462474536150694, + 0.12003641575574875, + 0.023795850574970245, + 0.00936917681246996, + 0.008831636980175972, + 0.018564613536000252, + 0.15422111749649048, + 0.012757089920341969, + 0.012805584818124771, + 0.0035167005844414234, + 0.0059827594086527824, + 0.001887956983409822, + 0.0014244409976527095, + 0.0020903402473777533, + 0.001378857996314764, + 0.0011633139802142978, + 0.001460607280023396, + 0.017598679289221764, + 0.028353113681077957, + 0.0126718208193779, + 0.00973100308328867, + 0.018592247739434242, + 0.008970910683274269, + 0.011378337629139423, + 0.007917760871350765, + 0.004165396559983492, + 0.003801919985562563, + 0.0033378570806235075, + 0.004725163336843252, + 0.003260335186496377, + 0.002285705879330635, + 0.0031423710752278566, + 0.002471068175509572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.02058335766196251 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, slightly overlooking, slowly pans to the right, following a cuddly giant panda as it clumsily swims deep within a lush, verdant bamboo forest. Soft sunlight filters through the dense foliage, casting dappled, vibrant light across the crystal-clear stream. The panda's signature thick black and white fur, soaked by the stream water, clings to its chubby body, creating a comical effect. Its chubby limbs shift uncoordinatedly in the water, each clumsy stroke a playful struggle against buoyancy rather than efficient propulsion. A look of both bewilderment and self-satisfaction permeates its face. As its body rises and falls, transparent ripples and countless crystalline droplets are stirred up. These droplets shimmer briefly in the sunlight before falling back into the water, creating a melodious sound in the tranquil forest. The stream is so clear that you can see the smooth pebbles at the bottom of the water. The banks are covered with moist moss and ferns. The whole scene is filled with a peaceful, natural and slightly humorous atmosphere in the main tones of vibrant emerald green and clear water, creating a pure and healing sense of picture.", + "seed": 24726, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_10/index_5331_seed_24726_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1196237268857658, + "vlm_logprob": -0.1196237268857658, + "vlm_yes_no_margin": 3.125000011920929, + "vlm_token_credit": -0.1196237268857658, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028544417582452297 + ], + "opposite_logprob_tokens": [ + -6.502854347229004 + ], + "logprob_reward": -0.0028544417582452297, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0028544417582452297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08780954778194427, + 0.07720361649990082, + 0.09030556678771973, + 0.06024732440710068, + 0.03465476632118225, + 0.10504026710987091, + 0.3708799481391907, + 0.043892864137887955, + 0.029353734105825424, + 0.10061236470937729 + ], + "credit_map": [ + 0.009313267655670643, + 0.011987044475972652, + 0.012115470133721828, + 0.011594810523092747, + 0.007650011219084263, + 0.01288060937076807, + 0.007592916954308748, + 0.014675409533083439, + 0.009619565680623055, + 0.007419448345899582, + 0.011707049794495106, + 0.011721940711140633, + 0.008945820853114128, + 0.011092843487858772, + 0.007612033747136593, + 0.00908491387963295, + 0.016940828412771225, + 0.014109322801232338, + 0.006029394920915365, + 0.009159177541732788, + 0.008622406981885433, + 0.014357918873429298, + 0.011670362204313278, + 0.009416148997843266, + 0.004623980727046728, + 0.0029610111378133297, + 0.013930853456258774, + 0.004387453198432922, + 0.00501317298039794, + 0.016432542353868484, + 0.006476701237261295, + 0.006421606056392193, + 0.008537479676306248, + 0.002296747174113989, + 0.002606983296573162, + 0.004651914816349745, + 0.003013783833011985, + 0.0054105655290186405, + 0.004789734724909067, + 0.0033475577365607023, + 0.017740918323397636, + 0.012714440003037453, + 0.016272010281682014, + 0.010022993199527264, + 0.00839528813958168, + 0.008524316363036633, + 0.009996539913117886, + 0.02137376181781292, + 0.10917583853006363, + 0.01712050847709179, + 0.1444682478904724, + 0.016574645414948463, + 0.019850505515933037, + 0.023184526711702347, + 0.015823790803551674, + 0.02468191832304001, + 0.007320646662265062, + 0.004090375732630491, + 0.009025431238114834, + 0.004487302154302597, + 0.004587958566844463, + 0.0040956218726933, + 0.004586183000355959, + 0.005699348635971546, + 0.003650936996564269, + 0.0027675640303641558, + 0.0032902585808187723, + 0.004858045373111963, + 0.0030059877317398787, + 0.0034190930891782045, + 0.003865885315462947, + 0.004495962988585234, + 0.027729375287890434, + 0.004509438760578632, + 0.021959815174341202, + 0.006273943930864334, + 0.006114372052252293, + 0.004664241801947355, + 0.009987023659050465, + 0.019374152645468712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.0028544417582452297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870534300804138 + ], + "opposite_logprob_tokens": [ + -1.1370534896850586 + ], + "logprob_reward": -0.3870534300804138, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870534300804138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12493359297513962, + 0.23965595662593842, + 0.12179828435182571, + 0.03876657411456108, + 0.02546134777367115, + 0.04097216948866844, + 0.21291622519493103, + 0.025971809402108192, + 0.03668692708015442, + 0.13283710181713104 + ], + "credit_map": [ + 0.014012238010764122, + 0.013615823350846767, + 0.017595531418919563, + 0.02079441212117672, + 0.010522901080548763, + 0.01539575681090355, + 0.015130962245166302, + 0.017865967005491257, + 0.020803961902856827, + 0.026937013491988182, + 0.04024515673518181, + 0.026017997413873672, + 0.02065184898674488, + 0.036074016243219376, + 0.0230193343013525, + 0.04590661823749542, + 0.02775084786117077, + 0.016414590179920197, + 0.009394239634275436, + 0.01395465712994337, + 0.010573895648121834, + 0.018017737194895744, + 0.013316763564944267, + 0.0123755419626832, + 0.0038166488520801067, + 0.0029926938004791737, + 0.005455615930259228, + 0.004530529025942087, + 0.0034491303376853466, + 0.00761893205344677, + 0.006227319594472647, + 0.004675708245486021, + 0.00441625015810132, + 0.0022310351487249136, + 0.0026617615949362516, + 0.002554657869040966, + 0.002294342964887619, + 0.004190713167190552, + 0.003014438785612583, + 0.004098146688193083, + 0.007926175370812416, + 0.004368165507912636, + 0.005103206727653742, + 0.0038288943469524384, + 0.0024612199049443007, + 0.0038061835803091526, + 0.004122428596019745, + 0.009355892427265644, + 0.05568424612283707, + 0.01047303806990385, + 0.06895039975643158, + 0.01060317549854517, + 0.01163648534566164, + 0.029910055920481682, + 0.010965986177325249, + 0.014692841097712517, + 0.004723304882645607, + 0.00197359686717391, + 0.006115800701081753, + 0.0032886641565710306, + 0.003251788904890418, + 0.002515010302886367, + 0.002033066703006625, + 0.0020705782808363438, + 0.0067026433534920216, + 0.00495369965210557, + 0.008273405954241753, + 0.003557610558345914, + 0.003669187193736434, + 0.002849796088412404, + 0.003550988854840398, + 0.003129597520455718, + 0.030196592211723328, + 0.011639537289738655, + 0.017735743895173073, + 0.01070091687142849, + 0.00988235604017973, + 0.006863679736852646, + 0.014842852018773556, + 0.030975423753261566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.3870534300804138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307698830962181 + ], + "opposite_logprob_tokens": [ + -3.168076992034912 + ], + "logprob_reward": -0.04307698830962181, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307698830962181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07825344800949097, + 0.11171452701091766, + 0.03757806494832039, + 0.11805874109268188, + 0.019079554826021194, + 0.05387182906270027, + 0.3157605826854706, + 0.04015275090932846, + 0.044706959277391434, + 0.18082356452941895 + ], + "credit_map": [ + 0.009542590007185936, + 0.007943259552121162, + 0.013633609749376774, + 0.009754032827913761, + 0.006375494413077831, + 0.010346170514822006, + 0.007838855497539043, + 0.012819433584809303, + 0.012290820479393005, + 0.011420496739447117, + 0.01295741181820631, + 0.012045500800013542, + 0.01260494627058506, + 0.015589951537549496, + 0.01425898540765047, + 0.020546404644846916, + 0.005068590398877859, + 0.00456079188734293, + 0.004555910360068083, + 0.005890795961022377, + 0.0031928967218846083, + 0.005267960485070944, + 0.004333897028118372, + 0.004707217216491699, + 0.008015463128685951, + 0.003923866432160139, + 0.03541488200426102, + 0.008301323279738426, + 0.007902201265096664, + 0.04067298397421837, + 0.006688167806714773, + 0.007139848079532385, + 0.0038166288286447525, + 0.0015737159410491586, + 0.0017511658370494843, + 0.00222327490337193, + 0.0015618392499163747, + 0.0027728115674108267, + 0.002363114384934306, + 0.0030170029494911432, + 0.011021101847290993, + 0.005768276285380125, + 0.011418623849749565, + 0.004957407247275114, + 0.005636268761008978, + 0.004796810448169708, + 0.005064267665147781, + 0.005209066439419985, + 0.07978846877813339, + 0.013735270127654076, + 0.1069699376821518, + 0.016493692994117737, + 0.017702940851449966, + 0.04028436169028282, + 0.014320792630314827, + 0.026465091854333878, + 0.004847360774874687, + 0.003717370331287384, + 0.005033052526414394, + 0.00527194794267416, + 0.007806398905813694, + 0.005668072495609522, + 0.003923327196389437, + 0.0038852153811603785, + 0.0058049401268363, + 0.00232437695376575, + 0.010850997641682625, + 0.005097860936075449, + 0.006014276761561632, + 0.004538699984550476, + 0.005429904442280531, + 0.004645898938179016, + 0.04402082413434982, + 0.0119036128744483, + 0.023348446935415268, + 0.014160145074129105, + 0.012540064752101898, + 0.0094472486525774, + 0.01958942599594593, + 0.04581379517912865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.04307698830962181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702706456184387 + ], + "opposite_logprob_tokens": [ + -2.1270270347595215 + ], + "logprob_reward": -0.12702706456184387, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702706456184387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013509889133274555, + 0.013636800460517406, + 0.02473648078739643, + 0.006467105355113745, + 0.01025470346212387, + 0.008899014443159103, + 0.8364486694335938, + 0.05232064053416252, + 0.028219034895300865, + 0.005507723428308964 + ], + "credit_map": [ + 0.0015415410744026303, + 0.001212020288221538, + 0.0017177367117255926, + 0.0025730272755026817, + 0.0011691056424751878, + 0.0017311528790742159, + 0.0020406979601830244, + 0.001524606253951788, + 0.0015013919910416007, + 0.001540890778414905, + 0.002260980661958456, + 0.0015574670396745205, + 0.0021580455359071493, + 0.0018714523175731301, + 0.001383951399475336, + 0.0013626203872263432, + 0.003182459156960249, + 0.004525414668023586, + 0.0019673043861985207, + 0.0031332543585449457, + 0.002466183388605714, + 0.0037541331257671118, + 0.0028002418112009764, + 0.0029074859339743853, + 0.000441145763034001, + 0.00041331848478876054, + 0.001414758269675076, + 0.0004467647522687912, + 0.0005453781341202557, + 0.0017375694587826729, + 0.0006593430298380554, + 0.000808826822321862, + 0.0015008068876340985, + 0.001089301542378962, + 0.0008268875535577536, + 0.0018073847750201821, + 0.001487293979153037, + 0.001413938240148127, + 0.0013997947098687291, + 0.0007292948430404067, + 0.001215683645568788, + 0.0009691013256087899, + 0.0013107105623930693, + 0.0012995083816349506, + 0.0011245803907513618, + 0.0009693786269053817, + 0.0011478253873065114, + 0.0008622249588370323, + 0.24650853872299194, + 0.03921983763575554, + 0.32893261313438416, + 0.03717202693223953, + 0.04487938433885574, + 0.04789402708411217, + 0.03379689157009125, + 0.05804522708058357, + 0.008323787711560726, + 0.005152079742401838, + 0.011113116517663002, + 0.006279126740992069, + 0.006840667221695185, + 0.005798456259071827, + 0.004511458333581686, + 0.004301944747567177, + 0.005420337431132793, + 0.0026298509910702705, + 0.003852874506264925, + 0.0028612648602575064, + 0.002721859375014901, + 0.0043320078402757645, + 0.0035050956066697836, + 0.002895740559324622, + 0.000823919486720115, + 0.0007254751399159431, + 0.0007047776598483324, + 0.0005476088263094425, + 0.00043483314220793545, + 0.0004773589607793838, + 0.0009898176649585366, + 0.0008039316162467003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.12702706456184387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant panda with thick black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda swimming clumsily in a crystal-clear stream within a lush bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the panda's fur appear soaked and clinging to its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a medium shot that slowly pans to the right following the panda?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the stream banks covered with moist moss and ferns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038106709718704224 + ], + "opposite_logprob_tokens": [ + -3.288106679916382 + ], + "logprob_reward": -0.038106709718704224, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038106709718704224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10454589873552322, + 0.019639931619167328, + 0.11705823242664337, + 0.122189462184906, + 0.06525101512670517, + 0.024881185963749886, + 0.3731939494609833, + 0.02753720059990883, + 0.042822983115911484, + 0.10288014262914658 + ], + "credit_map": [ + 0.015478678978979588, + 0.018489589914679527, + 0.015236643142998219, + 0.009496042504906654, + 0.005782030522823334, + 0.016007887199521065, + 0.009353261440992355, + 0.014701755717396736, + 0.002125480445101857, + 0.0023865317925810814, + 0.0024022439029067755, + 0.0027925779577344656, + 0.0026370494160801172, + 0.002876098733395338, + 0.0024333992041647434, + 0.001986549934372306, + 0.028363509103655815, + 0.013602289371192455, + 0.009756578132510185, + 0.011393498629331589, + 0.0073363035917282104, + 0.02033291570842266, + 0.012435131706297398, + 0.013838003389537334, + 0.012575937435030937, + 0.005353125743567944, + 0.026767654344439507, + 0.00906764529645443, + 0.009184399619698524, + 0.03355421870946884, + 0.011767634190618992, + 0.013918852433562279, + 0.012886466458439827, + 0.00398274278268218, + 0.00860714539885521, + 0.005111026111990213, + 0.0045319534838199615, + 0.013040084391832352, + 0.007339175324887037, + 0.009752422571182251, + 0.004880581982433796, + 0.0032833325676620007, + 0.002578630344942212, + 0.002010104013606906, + 0.003931496292352676, + 0.002207068959251046, + 0.00266777235083282, + 0.0033221999183297157, + 0.11081095039844513, + 0.01761713996529579, + 0.14658482372760773, + 0.0166348647326231, + 0.01999177783727646, + 0.02113186940550804, + 0.015210500918328762, + 0.025212019681930542, + 0.003590960521250963, + 0.002810753881931305, + 0.00515685835853219, + 0.0026942305266857147, + 0.004813516512513161, + 0.003304732497781515, + 0.002063578926026821, + 0.003102568443864584, + 0.009275499731302261, + 0.0054237996228039265, + 0.009719346649944782, + 0.003989238757640123, + 0.003119803499430418, + 0.0026853294111788273, + 0.0040830327197909355, + 0.004526933655142784, + 0.022248271852731705, + 0.006621352396905422, + 0.016275856643915176, + 0.008703431114554405, + 0.009163035079836845, + 0.005734074395149946, + 0.012494746595621109, + 0.021639384329319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.038106709718704224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_11/index_tensor([8388, 8388])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_11/index_tensor([8388, 8388])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0d2e9bfb5c835b07409dabb2f45dc4ad5da66a26 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_11/index_tensor([8388, 8388])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunningly low-angle, static shot, almost touching the damp forest floor strewn with withered yellow leaves and scattered moss, focuses upward on a single emerald-green bitter bamboo. Behind it, the dense bamboo forest forms a deep, dark vertical barrier, shading the surroundings with an extraordinary tranquility. At this moment, only a single beam of warm, golden sunlight, like a stage spotlight, penetrates the gaps in the forest canopy, casting a precise shadow, creating a distinct Tyndall effect in the dim air, with countless tiny particles dancing slowly within the beam of light. The bamboo stalks illuminated by the light take on the translucent texture of jade, their smooth surface tinged with a faint waxy sheen, while the backlit side is cast in soft shadow, outlining its sturdy outline. Under the effect of time-lapse photography, the bamboo pole is performing a phototropic movement in a stable and persistent posture that is imperceptible to the naked eye. Its slender trunk is slowly bending towards the direction of the light column in a tiny but clearly visible arc, and the few tender green bamboo leaves on the top also stretch out, as if silently declaring its desire for light and life. The whole picture is full of indomitable life tension and hope in the ultimate tranquility.", + "seed": 1440, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_11/index_8388_seed_1440_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5213402869179845, + "vlm_logprob": -0.5213402869179845, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.5213402869179845, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.46987974643707275 + ], + "opposite_logprob_tokens": [ + -1.0948797464370728 + ], + "logprob_reward": -0.46987974643707275, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.46987974643707275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22939597070217133, + 0.06187346205115318, + 0.04601268097758293, + 0.051686327904462814, + 0.16519685089588165, + 0.0757293775677681, + 0.09520222991704941, + 0.08836693316698074, + 0.11961622536182404, + 0.06691990047693253 + ], + "credit_map": [ + 0.03205227851867676, + 0.03835897892713547, + 0.04915967211127281, + 0.035928040742874146, + 0.01066782046109438, + 0.025545217096805573, + 0.014425213448703289, + 0.023258766159415245, + 0.006646996363997459, + 0.010427894070744514, + 0.0111008295789361, + 0.00956971850246191, + 0.002662218874320388, + 0.01076324749737978, + 0.004434021655470133, + 0.006268540397286415, + 0.004399668890982866, + 0.007147806230932474, + 0.006690106820315123, + 0.006941802799701691, + 0.004473146516829729, + 0.00512323621660471, + 0.006015024147927761, + 0.005221894010901451, + 0.004308228846639395, + 0.008344156667590141, + 0.016251694411039352, + 0.010345926508307457, + 0.0014846146805211902, + 0.005334905814379454, + 0.002066289773210883, + 0.003550515277311206, + 0.01448531448841095, + 0.010822364129126072, + 0.028851309791207314, + 0.015339565463364124, + 0.005058418493717909, + 0.06030226871371269, + 0.004774355795234442, + 0.02556326799094677, + 0.006740808021277189, + 0.012644341215491295, + 0.012529054656624794, + 0.01149220671504736, + 0.006284613162279129, + 0.010929092764854431, + 0.005422623828053474, + 0.00968664325773716, + 0.006055573932826519, + 0.010193979367613792, + 0.00663629500195384, + 0.00640836451202631, + 0.0037264088168740273, + 0.03125332295894623, + 0.0037550947163254023, + 0.02717318944633007, + 0.014486861415207386, + 0.010838976129889488, + 0.015131230466067791, + 0.010673045180737972, + 0.006800351198762655, + 0.011962300166487694, + 0.0077888257801532745, + 0.010685352608561516, + 0.01559506356716156, + 0.00983845442533493, + 0.021150901913642883, + 0.01645115204155445, + 0.007352166343480349, + 0.0237719863653183, + 0.006790147628635168, + 0.01866636611521244, + 0.008977625519037247, + 0.006846383213996887, + 0.012880443595349789, + 0.008276327513158321, + 0.0031568489503115416, + 0.01000256184488535, + 0.0035310033708810806, + 0.013248705305159092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.46987974643707275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236732244491577 + ], + "opposite_logprob_tokens": [ + -0.3486732244491577 + ], + "logprob_reward": -1.2236732244491577, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236732244491577, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10125470906496048, + 0.05456562340259552, + 0.07778852432966232, + 0.14022105932235718, + 0.08456878364086151, + 0.08815713971853256, + 0.10391411930322647, + 0.1257556676864624, + 0.17511875927448273, + 0.048655662685632706 + ], + "credit_map": [ + 0.012781315483152866, + 0.015986250713467598, + 0.015678107738494873, + 0.019218359142541885, + 0.00418587401509285, + 0.011975034140050411, + 0.007049678359180689, + 0.01438008900731802, + 0.0067906309850513935, + 0.010904654860496521, + 0.008963041938841343, + 0.008050436154007912, + 0.0026624836027622223, + 0.006618613377213478, + 0.00504839513450861, + 0.005527368281036615, + 0.007274986710399389, + 0.014207342639565468, + 0.02289695478975773, + 0.009515183046460152, + 0.0034562728833407164, + 0.008743004873394966, + 0.005611920263618231, + 0.006082861218601465, + 0.010108042508363724, + 0.026908788830041885, + 0.022209133952856064, + 0.034108906984329224, + 0.007257833611220121, + 0.0203764159232378, + 0.007075956556946039, + 0.012175975367426872, + 0.006897220388054848, + 0.014508563093841076, + 0.010955403558909893, + 0.01836421713232994, + 0.008097952231764793, + 0.011705479584634304, + 0.007376396097242832, + 0.006663541775196791, + 0.0077690063044428825, + 0.012112898752093315, + 0.017154797911643982, + 0.014392344281077385, + 0.008936076425015926, + 0.010810990817844868, + 0.009334011003375053, + 0.007647005375474691, + 0.00642880005761981, + 0.009033634327352047, + 0.009318350814282894, + 0.009282490238547325, + 0.007449677679687738, + 0.03198802471160889, + 0.00863612163811922, + 0.021777018904685974, + 0.0158990491181612, + 0.019532445818185806, + 0.02765156887471676, + 0.012421905063092709, + 0.00860515981912613, + 0.014433272182941437, + 0.010707326233386993, + 0.016504937782883644, + 0.013092966750264168, + 0.009850839152932167, + 0.038112714886665344, + 0.02225157618522644, + 0.005242862738668919, + 0.0209171362221241, + 0.0077009438537061214, + 0.057949718087911606, + 0.004706518724560738, + 0.005859263241291046, + 0.0061903721652925014, + 0.012650219723582268, + 0.0037134976591914892, + 0.005604535806924105, + 0.0039716134779155254, + 0.005959641188383102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -1.2236732244491577 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565725483000278 + ], + "opposite_logprob_tokens": [ + -4.387565612792969 + ], + "logprob_reward": -0.012565725483000278, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565725483000278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1735103279352188, + 0.16798657178878784, + 0.0958394780755043, + 0.05664227530360222, + 0.17008228600025177, + 0.09348063170909882, + 0.05265635997056961, + 0.02918974682688713, + 0.051372360438108444, + 0.10923995822668076 + ], + "credit_map": [ + 0.01805763505399227, + 0.026070350781083107, + 0.03149200230836868, + 0.026637472212314606, + 0.011327231302857399, + 0.021836750209331512, + 0.012167483568191528, + 0.02592141181230545, + 0.01714683696627617, + 0.029232069849967957, + 0.02616753615438938, + 0.022902103140950203, + 0.010876401327550411, + 0.023149270564317703, + 0.017080336809158325, + 0.021432016044855118, + 0.006558987777680159, + 0.0198319423943758, + 0.01867843233048916, + 0.011477194726467133, + 0.007141915149986744, + 0.012282243929803371, + 0.009024428203701973, + 0.010844328440725803, + 0.004677752032876015, + 0.008694577030837536, + 0.007291803602129221, + 0.010887941345572472, + 0.004765900783240795, + 0.009359749034047127, + 0.004202411975711584, + 0.0067621381022036076, + 0.01393021922558546, + 0.009340014308691025, + 0.03120841085910797, + 0.016356969252228737, + 0.0062658702954649925, + 0.06246175989508629, + 0.00517069548368454, + 0.025348350405693054, + 0.00924431998282671, + 0.01501998957246542, + 0.013100803829729557, + 0.012386378832161427, + 0.011742082424461842, + 0.012294662185013294, + 0.010096218436956406, + 0.009596171788871288, + 0.0034337223041802645, + 0.007050714921206236, + 0.0048957825638353825, + 0.008630849421024323, + 0.0038483808748424053, + 0.0130930095911026, + 0.003630471881479025, + 0.008073427714407444, + 0.003225248772650957, + 0.005169196054339409, + 0.0041548158042132854, + 0.003701470559462905, + 0.0030665649101138115, + 0.004607273731380701, + 0.001961842877790332, + 0.0033033357467502356, + 0.007980643771588802, + 0.005198436789214611, + 0.007005883846431971, + 0.005357574671506882, + 0.003328902181237936, + 0.011804407462477684, + 0.0026607441250234842, + 0.008035773411393166, + 0.011009416542947292, + 0.010807301849126816, + 0.016362052410840988, + 0.011492091231048107, + 0.007144834380596876, + 0.023164158686995506, + 0.010372078977525234, + 0.01888803392648697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.012565725483000278 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022066310048103333 + ], + "opposite_logprob_tokens": [ + -6.127206802368164 + ], + "logprob_reward": -0.0022066310048103333, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022066310048103333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06980278342962265, + 0.030996903777122498, + 0.04870743677020073, + 0.04433603957295418, + 0.0527074970304966, + 0.047587666660547256, + 0.16518564522266388, + 0.12697190046310425, + 0.36069098114967346, + 0.05301306024193764 + ], + "credit_map": [ + 0.007893324829638004, + 0.012532157823443413, + 0.012985123321413994, + 0.008386796340346336, + 0.005283132661134005, + 0.010625949129462242, + 0.005290924105793238, + 0.0068053812719881535, + 0.0035498072393238544, + 0.004292577039450407, + 0.00579987233504653, + 0.006073594093322754, + 0.00182475452311337, + 0.0038857071194797754, + 0.0028892578557133675, + 0.0026813368313014507, + 0.0036642702762037516, + 0.009292509406805038, + 0.013681107200682163, + 0.0059343017637729645, + 0.0030578100122511387, + 0.004665075335651636, + 0.0031862137839198112, + 0.0052261510863900185, + 0.0034803084563463926, + 0.005172973033040762, + 0.013136075809597969, + 0.009444192051887512, + 0.0016831079265102744, + 0.005341938231140375, + 0.0025179332587867975, + 0.003559513483196497, + 0.0041636135429143906, + 0.005623538512736559, + 0.007456930819898844, + 0.008991245180368423, + 0.0055258674547076225, + 0.012282426469027996, + 0.003141265595331788, + 0.005522612482309341, + 0.002940097823739052, + 0.009267291985452175, + 0.005283776205033064, + 0.005724568385630846, + 0.005842713639140129, + 0.008922092616558075, + 0.0035897819325327873, + 0.006017348729074001, + 0.013534331694245338, + 0.03488834574818611, + 0.01461494155228138, + 0.01593024842441082, + 0.012403580360114574, + 0.033743713051080704, + 0.012942072004079819, + 0.027128424495458603, + 0.016853580251336098, + 0.021737489849328995, + 0.026482265442609787, + 0.015182343311607838, + 0.006893300451338291, + 0.01466680783778429, + 0.005498774815350771, + 0.019657356664538383, + 0.027239879593253136, + 0.01875034160912037, + 0.09436754882335663, + 0.03295397758483887, + 0.00959792174398899, + 0.04631771519780159, + 0.009537125006318092, + 0.12192650884389877, + 0.008286044001579285, + 0.006864864844828844, + 0.005793541669845581, + 0.007365182973444462, + 0.003717671846970916, + 0.010472316294908524, + 0.0037081914488226175, + 0.006805249955505133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.0022066310048103333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983761072158813 + ], + "opposite_logprob_tokens": [ + -0.5233761072158813 + ], + "logprob_reward": -0.8983761072158813, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983761072158813, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10298318415880203, + 0.03889558091759682, + 0.1330500692129135, + 0.06280043721199036, + 0.08944500982761383, + 0.12676605582237244, + 0.07353462278842926, + 0.06531821191310883, + 0.2114773541688919, + 0.09572944045066833 + ], + "credit_map": [ + 0.00954346265643835, + 0.021991949528455734, + 0.012846773490309715, + 0.018620561808347702, + 0.004687475971877575, + 0.015878839418292046, + 0.009280789643526077, + 0.010133328847587109, + 0.004818033892661333, + 0.007725737523287535, + 0.005919246934354305, + 0.005943081807345152, + 0.002335520461201668, + 0.005419284105300903, + 0.0030325413681566715, + 0.003702132496982813, + 0.00941294152289629, + 0.031078815460205078, + 0.04354595020413399, + 0.013558984734117985, + 0.006440913770347834, + 0.008345495909452438, + 0.009607679210603237, + 0.011059295386075974, + 0.005343688651919365, + 0.010487699881196022, + 0.013281388208270073, + 0.015497230924665928, + 0.003152423305436969, + 0.007956979796290398, + 0.0027528060600161552, + 0.0043282173573970795, + 0.007095202337950468, + 0.009906730614602566, + 0.016072241589426994, + 0.009246604517102242, + 0.01033629011362791, + 0.02225578762590885, + 0.0062048668041825294, + 0.008327288553118706, + 0.013180562295019627, + 0.022498473525047302, + 0.023223698139190674, + 0.01845274493098259, + 0.010048853233456612, + 0.020174918696284294, + 0.009012427181005478, + 0.010174384340643883, + 0.0076255397871136665, + 0.011196772567927837, + 0.009093266911804676, + 0.009924441576004028, + 0.008420949801802635, + 0.01114712655544281, + 0.0064834775403141975, + 0.009643049910664558, + 0.010199404321610928, + 0.009077344089746475, + 0.014267373830080032, + 0.009387930855154991, + 0.002909491304308176, + 0.008970323018729687, + 0.0028510468546301126, + 0.007655302062630653, + 0.015290774405002594, + 0.010234222747385502, + 0.05339145287871361, + 0.020272696390748024, + 0.005593856796622276, + 0.02519821561872959, + 0.005029954016208649, + 0.07646619528532028, + 0.008989300578832626, + 0.010060806758701801, + 0.0150649044662714, + 0.013049140572547913, + 0.006007843650877476, + 0.018426692113280296, + 0.00740066496655345, + 0.016730090603232384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.8983761072158813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "stunningly low-angle, static shot, almost touching the damp forest floor strewn with withered yellow leaves and scattered moss, focuses upward on a single emerald-green bitter bamboo. Behind it, the dense bamboo forest forms a deep, dark vertical barrier, shading the surroundings with an extraordinary tranquility. At this moment, only a single beam of warm, golden sunlight, like a stage spotlight, penetrates the gaps in the forest canopy, casting a precise shadow, creating a distinct Tyndall effect in the dim air, with countless tiny particles dancing slowly within the beam of light. The bamboo stalks illuminated by the light take on the translucent texture of jade, their smooth surface tinged with a faint waxy sheen, while the backlit side is cast in soft shadow, outlining its sturdy outline. Under the effect of time-lapse photography, the bamboo pole is performing a phototropic movement in a stable and persistent posture that is imperceptible to the naked eye. Its slender trunk is slowly bending towards the direction of the light column in a tiny but clearly visible arc, and the few tender green bamboo leaves on the top also stretch out, as if silently declaring its desire for light and life. The whole picture is full of indomitable life tension and hope in the ultimate tranquility.", + "seed": 1441, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_11/index_8388_seed_1441_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3524063009768724, + "vlm_logprob": -0.3524063009768724, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.3524063009768724, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09272489696741104 + ], + "opposite_logprob_tokens": [ + -2.7177248001098633 + ], + "logprob_reward": -0.09272489696741104, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09272489696741104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1536238044500351, + 0.05173115059733391, + 0.018679196015000343, + 0.026306791231036186, + 0.048162031918764114, + 0.13827408850193024, + 0.34968623518943787, + 0.08027626574039459, + 0.045347634702920914, + 0.08791282773017883 + ], + "credit_map": [ + 0.018090369179844856, + 0.014644209295511246, + 0.02002009190618992, + 0.013483896851539612, + 0.02921167202293873, + 0.025337964296340942, + 0.015403707511723042, + 0.017431877553462982, + 0.0061485967598855495, + 0.002592368982732296, + 0.008441080339252949, + 0.006099083460867405, + 0.00979282520711422, + 0.006961499806493521, + 0.004820138216018677, + 0.006875551771372557, + 0.0017131997738033533, + 0.001168223563581705, + 0.0027550519444048405, + 0.0022214800119400024, + 0.0032610008493065834, + 0.0026254209224134684, + 0.0023256568238139153, + 0.0026091604959219694, + 0.003066306235268712, + 0.0014556896639987826, + 0.0035393768921494484, + 0.0020807830151170492, + 0.004553001374006271, + 0.004328479059040546, + 0.0030487333424389362, + 0.004234418738633394, + 0.0045422036200761795, + 0.0023014897014945745, + 0.006219706032425165, + 0.0023752443958073854, + 0.011343386024236679, + 0.0075402576476335526, + 0.008619854226708412, + 0.005219884682446718, + 0.01262526586651802, + 0.008705923333764076, + 0.019783005118370056, + 0.011810769326984882, + 0.02735738642513752, + 0.02052089013159275, + 0.024130722507834435, + 0.013340123929083347, + 0.02557496540248394, + 0.012709674425423145, + 0.08310451358556747, + 0.02134447731077671, + 0.050460610538721085, + 0.028206417337059975, + 0.042684633284807205, + 0.08560091257095337, + 0.008874687366187572, + 0.0034259804524481297, + 0.02090153843164444, + 0.004170497879385948, + 0.009821335785090923, + 0.008831572718918324, + 0.009315296076238155, + 0.014935350976884365, + 0.005723495036363602, + 0.0026736410800367594, + 0.013335573486983776, + 0.0024466076865792274, + 0.0032053287141025066, + 0.003966579679399729, + 0.00509230699390173, + 0.00890409853309393, + 0.012997347861528397, + 0.005750042386353016, + 0.01507286075502634, + 0.00492868572473526, + 0.012154837138950825, + 0.0103102196007967, + 0.01280752383172512, + 0.013891298323869705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.09272489696741104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371355056762695 + ], + "opposite_logprob_tokens": [ + -0.38713550567626953 + ], + "logprob_reward": -1.1371355056762695, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371355056762695, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09135118871927261, + 0.07548034191131592, + 0.14937371015548706, + 0.1727255880832672, + 0.09194808453321457, + 0.07171658426523209, + 0.07445535063743591, + 0.1130032166838646, + 0.05209920555353165, + 0.10784672945737839 + ], + "credit_map": [ + 0.010866405442357063, + 0.0072112721391022205, + 0.011065897531807423, + 0.01023410726338625, + 0.020080313086509705, + 0.01584644801914692, + 0.008300120010972023, + 0.007746626157313585, + 0.008406666107475758, + 0.006484788842499256, + 0.00913951825350523, + 0.008130930364131927, + 0.012528616935014725, + 0.015667669475078583, + 0.006622315384447575, + 0.008499834686517715, + 0.01470231730490923, + 0.010251711122691631, + 0.02492624893784523, + 0.02505085989832878, + 0.025231171399354935, + 0.020275909453630447, + 0.014359451830387115, + 0.014576037414371967, + 0.01762588508427143, + 0.012865059077739716, + 0.02007710188627243, + 0.028646212071180344, + 0.021368209272623062, + 0.03030790388584137, + 0.01993647962808609, + 0.02189873345196247, + 0.009886658750474453, + 0.008088412694633007, + 0.011359144002199173, + 0.012489954940974712, + 0.01246035099029541, + 0.014699098654091358, + 0.012791628018021584, + 0.010172830894589424, + 0.008384555578231812, + 0.006558481603860855, + 0.01092985924333334, + 0.0077133383601903915, + 0.011656340211629868, + 0.010304942727088928, + 0.008626108057796955, + 0.0075429584830999374, + 0.00947851873934269, + 0.005890771746635437, + 0.009460262022912502, + 0.009994665160775185, + 0.01454213261604309, + 0.007312288042157888, + 0.009365451522171497, + 0.008411253802478313, + 0.011402552947402, + 0.007215902674943209, + 0.02283616177737713, + 0.009484300389885902, + 0.015652300789952278, + 0.015322478488087654, + 0.013323167338967323, + 0.017766350880265236, + 0.005526662804186344, + 0.003915257286280394, + 0.007625240832567215, + 0.0033056295942515135, + 0.009981337003409863, + 0.008151735179126263, + 0.007962667383253574, + 0.005630672909319401, + 0.012425185181200504, + 0.005045278463512659, + 0.025266094133257866, + 0.004809904377907515, + 0.018020423129200935, + 0.012997523881494999, + 0.016504567116498947, + 0.012777757830917835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -1.1371355056762695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111172165721655 + ], + "opposite_logprob_tokens": [ + -5.504111289978027 + ], + "logprob_reward": -0.004111172165721655, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111172165721655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1976383477449417, + 0.09065144509077072, + 0.06230448931455612, + 0.1047009751200676, + 0.07228361070156097, + 0.14905695617198944, + 0.12638743221759796, + 0.039173658937215805, + 0.06938271224498749, + 0.08842044323682785 + ], + "credit_map": [ + 0.02726905792951584, + 0.021958447992801666, + 0.02592563070356846, + 0.013186660595238209, + 0.03743496537208557, + 0.040464114397764206, + 0.013191798701882362, + 0.018207654356956482, + 0.012040003202855587, + 0.007618242874741554, + 0.015362933278083801, + 0.012266570702195168, + 0.012774589471518993, + 0.008717588149011135, + 0.007432303857058287, + 0.014439202845096588, + 0.005971163045614958, + 0.00497255427762866, + 0.009681208059191704, + 0.010569770820438862, + 0.009146434254944324, + 0.010637073777616024, + 0.005424527917057276, + 0.005901752505451441, + 0.009788045659661293, + 0.007736908737570047, + 0.01336003839969635, + 0.010225007310509682, + 0.018979700282216072, + 0.013790402561426163, + 0.017488131299614906, + 0.013332728296518326, + 0.009847667068243027, + 0.004174394067376852, + 0.009953564964234829, + 0.004845843184739351, + 0.009232567623257637, + 0.013067035935819149, + 0.013277445919811726, + 0.007885088212788105, + 0.018001580610871315, + 0.012431035749614239, + 0.02166217565536499, + 0.015461111441254616, + 0.02513623610138893, + 0.023583265021443367, + 0.016677143052220345, + 0.01610439270734787, + 0.009337148629128933, + 0.006050031632184982, + 0.029461292549967766, + 0.010944617912173271, + 0.015732837840914726, + 0.009472628124058247, + 0.014310483820736408, + 0.03107837773859501, + 0.0041488888673484325, + 0.001430285512469709, + 0.01151102315634489, + 0.0017641374142840505, + 0.0030089905485510826, + 0.0034478555899113417, + 0.0031505355145782232, + 0.010711935348808765, + 0.0069737667217850685, + 0.004568011965602636, + 0.012562334537506104, + 0.005068550817668438, + 0.009773905389010906, + 0.006393086165189743, + 0.013622032478451729, + 0.010421019978821278, + 0.013850042596459389, + 0.007375818211585283, + 0.014361480250954628, + 0.006487424019724131, + 0.012167785316705704, + 0.009009464643895626, + 0.01124369353055954, + 0.013924721628427505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.004111172165721655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647762048989534 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647762048989534, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647762048989534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21991991996765137, + 0.1049848273396492, + 0.08626574277877808, + 0.10481665283441544, + 0.06242530792951584, + 0.09921612590551376, + 0.20115624368190765, + 0.0405762679874897, + 0.04155946522951126, + 0.03907950967550278 + ], + "credit_map": [ + 0.025046445429325104, + 0.027243854478001595, + 0.025270963087677956, + 0.01962757483124733, + 0.03469682112336159, + 0.03428936377167702, + 0.02736479789018631, + 0.026380078867077827, + 0.0112829077988863, + 0.006723407190293074, + 0.017441583797335625, + 0.010167248547077179, + 0.020377913489937782, + 0.016205918043851852, + 0.01002529077231884, + 0.012760546058416367, + 0.007811747491359711, + 0.004623237531632185, + 0.013727644458413124, + 0.00922289676964283, + 0.01981242373585701, + 0.010226299054920673, + 0.010693211108446121, + 0.01014826912432909, + 0.009513339959084988, + 0.009932397864758968, + 0.016288794577121735, + 0.01512950100004673, + 0.011330908164381981, + 0.016577983275055885, + 0.013463756069540977, + 0.01257996540516615, + 0.005916486959904432, + 0.0035870682913810015, + 0.007638304028660059, + 0.006007404997944832, + 0.01020954828709364, + 0.012332427315413952, + 0.00926374364644289, + 0.007470316253602505, + 0.010817677713930607, + 0.006297925021499395, + 0.016042176634073257, + 0.008356930688023567, + 0.013750568963587284, + 0.022694259881973267, + 0.011973688378930092, + 0.009282891638576984, + 0.01671845279633999, + 0.009305592626333237, + 0.05062376707792282, + 0.01379022654145956, + 0.021509479731321335, + 0.01711195521056652, + 0.018883375450968742, + 0.05321338400244713, + 0.004782388918101788, + 0.0023634203244000673, + 0.00719669833779335, + 0.002707196632400155, + 0.0066562495194375515, + 0.0048399632796645164, + 0.005344012752175331, + 0.006686336360871792, + 0.0043798694387078285, + 0.002364445710554719, + 0.01098966971039772, + 0.002640222432091832, + 0.004815725609660149, + 0.004788889084011316, + 0.004079186823219061, + 0.007501454558223486, + 0.005337514448910952, + 0.003142020897939801, + 0.006996410433202982, + 0.0022802117746323347, + 0.0043531968258321285, + 0.003522997722029686, + 0.006583042908459902, + 0.006864110007882118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.004647762048989534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234121680259705 + ], + "opposite_logprob_tokens": [ + -0.8984121680259705 + ], + "logprob_reward": -0.5234121680259705, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234121680259705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08718936145305634, + 0.04199913516640663, + 0.10161767899990082, + 0.1775917112827301, + 0.10289366543292999, + 0.10052421689033508, + 0.10862743109464645, + 0.13087286055088043, + 0.08677183836698532, + 0.06191210821270943 + ], + "credit_map": [ + 0.011201892979443073, + 0.0074353888630867004, + 0.009612197056412697, + 0.007376513909548521, + 0.01809946820139885, + 0.017451200634241104, + 0.006837830413132906, + 0.009174867533147335, + 0.0056258696131408215, + 0.003214577678591013, + 0.007720558438450098, + 0.003090579528361559, + 0.007499421946704388, + 0.004874456208199263, + 0.003527425928041339, + 0.006446249317377806, + 0.01055649109184742, + 0.00494071701541543, + 0.013258418999612331, + 0.01286648865789175, + 0.020217018201947212, + 0.018515106290578842, + 0.008560825139284134, + 0.012702608481049538, + 0.021349573507905006, + 0.01171447429805994, + 0.024117235094308853, + 0.025625314563512802, + 0.028874743729829788, + 0.025386551395058632, + 0.023052094504237175, + 0.01747169904410839, + 0.01041699480265379, + 0.00874574575573206, + 0.013355497270822525, + 0.013947570696473122, + 0.014690093696117401, + 0.01670560985803604, + 0.014768723398447037, + 0.010263429954648018, + 0.011992027051746845, + 0.009225724264979362, + 0.013650280423462391, + 0.012078926898539066, + 0.015985092148184776, + 0.016641903668642044, + 0.01143814716488123, + 0.009512116201221943, + 0.009652562439441681, + 0.006341956090182066, + 0.018472859635949135, + 0.007712673395872116, + 0.019588744267821312, + 0.013156434521079063, + 0.017787093296647072, + 0.01591511070728302, + 0.014159949496388435, + 0.010821996256709099, + 0.019699182361364365, + 0.010347515344619751, + 0.023185554891824722, + 0.015871893614530563, + 0.018604708835482597, + 0.018182067200541496, + 0.01017494685947895, + 0.00579474214464426, + 0.01294065173715353, + 0.004871446639299393, + 0.013224913738667965, + 0.01508057489991188, + 0.014631988480687141, + 0.01005257572978735, + 0.007391239982098341, + 0.003560638288035989, + 0.011770461685955524, + 0.003076768247410655, + 0.009425451047718525, + 0.007762458641082048, + 0.008486581966280937, + 0.010438509285449982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.5234121680259705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunningly low-angle, static shot, almost touching the damp forest floor strewn with withered yellow leaves and scattered moss, focuses upward on a single emerald-green bitter bamboo. Behind it, the dense bamboo forest forms a deep, dark vertical barrier, shading the surroundings with an extraordinary tranquility. At this moment, only a single beam of warm, golden sunlight, like a stage spotlight, penetrates the gaps in the forest canopy, casting a precise shadow, creating a distinct Tyndall effect in the dim air, with countless tiny particles dancing slowly within the beam of light. The bamboo stalks illuminated by the light take on the translucent texture of jade, their smooth surface tinged with a faint waxy sheen, while the backlit side is cast in soft shadow, outlining its sturdy outline. Under the effect of time-lapse photography, the bamboo pole is performing a phototropic movement in a stable and persistent posture that is imperceptible to the naked eye. Its slender trunk is slowly bending towards the direction of the light column in a tiny but clearly visible arc, and the few tender green bamboo leaves on the top also stretch out, as if silently declaring its desire for light and life. The whole picture is full of indomitable life tension and hope in the ultimate tranquility.", + "seed": 1442, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_11/index_8388_seed_1442_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.514023761684075, + "vlm_logprob": -0.514023761684075, + "vlm_yes_no_margin": 2.049999952316284, + "vlm_token_credit": -0.514023761684075, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18230848014354706 + ], + "opposite_logprob_tokens": [ + -1.9323084354400635 + ], + "logprob_reward": -0.18230848014354706, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.18230848014354706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1670326143503189, + 0.5319452285766602, + 0.0293973907828331, + 0.041105639189481735, + 0.03802068904042244, + 0.023618154227733612, + 0.04216747730970383, + 0.032094817608594894, + 0.044081203639507294, + 0.050536807626485825 + ], + "credit_map": [ + 0.02680075727403164, + 0.013161838985979557, + 0.027639079838991165, + 0.01215517707169056, + 0.016879422590136528, + 0.03150101378560066, + 0.014902245253324509, + 0.023993078619241714, + 0.10803678631782532, + 0.028057705610990524, + 0.1125885546207428, + 0.04037066176533699, + 0.1008286327123642, + 0.054634515196084976, + 0.055065616965293884, + 0.03236272558569908, + 0.003744895337149501, + 0.0018951034871861339, + 0.004975913558155298, + 0.0029054959304630756, + 0.003256181487813592, + 0.004665339831262827, + 0.002845004666596651, + 0.005109456833451986, + 0.00802653655409813, + 0.002376982942223549, + 0.00697324750944972, + 0.0028902310878038406, + 0.003154989331960678, + 0.008387944661080837, + 0.004295601975172758, + 0.0050001079216599464, + 0.004205520264804363, + 0.0023167612962424755, + 0.007057446986436844, + 0.001989304320886731, + 0.003120215144008398, + 0.01052749902009964, + 0.003750528208911419, + 0.005053409840911627, + 0.00355538004077971, + 0.0020551150664687157, + 0.003562397323548794, + 0.00134463410358876, + 0.001674209488555789, + 0.003853804199025035, + 0.003950876649469137, + 0.003621738636866212, + 0.005306549370288849, + 0.004750609863549471, + 0.0061586047522723675, + 0.004084895830601454, + 0.005066333804279566, + 0.006681082304567099, + 0.006025638896971941, + 0.004093760158866644, + 0.004588086158037186, + 0.0061332788318395615, + 0.004404837265610695, + 0.0029342849738895893, + 0.003928869962692261, + 0.002857762621715665, + 0.00465506361797452, + 0.002592633245512843, + 0.007690444588661194, + 0.006038900464773178, + 0.008192688226699829, + 0.003888075239956379, + 0.003450187388807535, + 0.007417383138090372, + 0.002651263726875186, + 0.004752262961119413, + 0.0077963052317500114, + 0.005529151763767004, + 0.008655485697090626, + 0.0040351590141654015, + 0.0038285027258098125, + 0.01066936831921339, + 0.0037952561397105455, + 0.006227580830454826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.18230848014354706 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134723901748657 + ], + "opposite_logprob_tokens": [ + -0.31347236037254333 + ], + "logprob_reward": -1.3134723901748657, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134723901748657, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05284082517027855, + 0.5803075432777405, + 0.0267983116209507, + 0.08129514753818512, + 0.061972327530384064, + 0.038962461054325104, + 0.03903189301490784, + 0.07410760223865509, + 0.019124962389469147, + 0.02555900625884533 + ], + "credit_map": [ + 0.008694036863744259, + 0.004559671971946955, + 0.007536441087722778, + 0.006164690013974905, + 0.004471863619983196, + 0.009193673729896545, + 0.006117527838796377, + 0.006102920044213533, + 0.11302513629198074, + 0.034805260598659515, + 0.11606211960315704, + 0.047073569148778915, + 0.10924316942691803, + 0.05745343863964081, + 0.06342732906341553, + 0.039217498153448105, + 0.0038699633441865444, + 0.001751599251292646, + 0.004361017141491175, + 0.0025991962756961584, + 0.0030272388830780983, + 0.004330241121351719, + 0.0031648308504372835, + 0.0036942255683243275, + 0.016286898404359818, + 0.00610438222065568, + 0.014240948483347893, + 0.00500902347266674, + 0.006717487704008818, + 0.015003124251961708, + 0.0067953504621982574, + 0.011137929745018482, + 0.008909091353416443, + 0.004518112167716026, + 0.011627466417849064, + 0.004969087895005941, + 0.005569205153733492, + 0.011671613901853561, + 0.007685755379498005, + 0.007021995261311531, + 0.00600982503965497, + 0.0033629818353801966, + 0.006812126375734806, + 0.0032952905166894197, + 0.002674562158063054, + 0.0049125272780656815, + 0.005306693259626627, + 0.00658845528960228, + 0.004311910830438137, + 0.005294443108141422, + 0.005739766638725996, + 0.0035297933500260115, + 0.0029115970246493816, + 0.008446472696959972, + 0.005430554039776325, + 0.003367353230714798, + 0.009474716149270535, + 0.01002765167504549, + 0.009665552526712418, + 0.007279377430677414, + 0.009904823265969753, + 0.009972754865884781, + 0.008689327165484428, + 0.009093394502997398, + 0.0029870630241930485, + 0.0031524982769042253, + 0.0035028448328375816, + 0.0011597664561122656, + 0.0020657898858189583, + 0.00317934132181108, + 0.0011253957636654377, + 0.0019522630609571934, + 0.002532667014747858, + 0.003180351573973894, + 0.0034118806943297386, + 0.001851743902079761, + 0.004055266734212637, + 0.0033914458472281694, + 0.0024770963937044144, + 0.0046585556119680405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -1.3134723901748657 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018206268548965454 + ], + "opposite_logprob_tokens": [ + -4.0182061195373535 + ], + "logprob_reward": -0.018206268548965454, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018206268548965454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06108676269650459, + 0.6482863426208496, + 0.041884198784828186, + 0.04580909386277199, + 0.03571061044931412, + 0.050661277025938034, + 0.026223663240671158, + 0.024887533858418465, + 0.024669604375958443, + 0.04078083112835884 + ], + "credit_map": [ + 0.009334512986242771, + 0.00565333291888237, + 0.012368416413664818, + 0.006696225143969059, + 0.005064178723841906, + 0.007248383015394211, + 0.0073179807513952255, + 0.007403733674436808, + 0.11246400326490402, + 0.04190795123577118, + 0.11287953704595566, + 0.060397762805223465, + 0.12669377028942108, + 0.06574112176895142, + 0.07868629693984985, + 0.049515943974256516, + 0.005675258114933968, + 0.0036352560855448246, + 0.006654004566371441, + 0.004951763898134232, + 0.0038828954566270113, + 0.007494669407606125, + 0.004338334314525127, + 0.0052520157769322395, + 0.006262451410293579, + 0.0028227237053215504, + 0.008878594264388084, + 0.004245704971253872, + 0.003115713130682707, + 0.007170128170400858, + 0.0064275492914021015, + 0.00688623171299696, + 0.004967500921338797, + 0.0030171386897563934, + 0.00647421320900321, + 0.0035967330913990736, + 0.0038321467582136393, + 0.006431908812373877, + 0.0031566505786031485, + 0.0042343176901340485, + 0.006843247450888157, + 0.004608169663697481, + 0.00844169408082962, + 0.00413156021386385, + 0.004762711934745312, + 0.007467396557331085, + 0.008629269897937775, + 0.005777226760983467, + 0.003404587507247925, + 0.004066291265189648, + 0.003677147440612316, + 0.0020966511219739914, + 0.002381545025855303, + 0.003937226720154285, + 0.003675705287605524, + 0.0029845070093870163, + 0.003182507585734129, + 0.00503443693742156, + 0.0031936352606862783, + 0.002611352363601327, + 0.0025431420654058456, + 0.0032217896077781916, + 0.002818830544129014, + 0.0022818397264927626, + 0.004987221676856279, + 0.003556131152436137, + 0.003675985150039196, + 0.0018304820405319333, + 0.002229767618700862, + 0.003621729789301753, + 0.0011526773450896144, + 0.003615610534325242, + 0.006216902751475573, + 0.006844379939138889, + 0.00544855697080493, + 0.002712148940190673, + 0.0038972520269453526, + 0.006339104380458593, + 0.0023005083203315735, + 0.007021978963166475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.018206268548965454 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022060361225157976 + ], + "opposite_logprob_tokens": [ + -6.127205848693848 + ], + "logprob_reward": -0.0022060361225157976, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022060361225157976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05298864096403122, + 0.5508724451065063, + 0.031019696965813637, + 0.07334989309310913, + 0.07409179955720901, + 0.03125087544322014, + 0.03415152430534363, + 0.07101599127054214, + 0.02998024970293045, + 0.05127888172864914 + ], + "credit_map": [ + 0.009055995382368565, + 0.0038580792024731636, + 0.008121081627905369, + 0.004944506101310253, + 0.00573877664282918, + 0.009107296355068684, + 0.00620168587192893, + 0.005961214192211628, + 0.09640034288167953, + 0.033911559730768204, + 0.09803257882595062, + 0.04765130206942558, + 0.11171063780784607, + 0.05819009244441986, + 0.06271499395370483, + 0.04226086661219597, + 0.0038222225848585367, + 0.0018346562283113599, + 0.0056537967175245285, + 0.0031437312718480825, + 0.0029102438129484653, + 0.004759252071380615, + 0.00410003587603569, + 0.004795752931386232, + 0.011942621320486069, + 0.005577635020017624, + 0.012777215801179409, + 0.005109135061502457, + 0.007886294275522232, + 0.009854263626039028, + 0.012471426278352737, + 0.007731286808848381, + 0.010432732291519642, + 0.007428782992064953, + 0.013232954777777195, + 0.006393167655915022, + 0.005391645710915327, + 0.014558861032128334, + 0.007535917684435844, + 0.009117723442614079, + 0.004436049610376358, + 0.0028910720720887184, + 0.005556462332606316, + 0.0022358123678714037, + 0.00298119755461812, + 0.004210478160530329, + 0.0049323271960020065, + 0.004007473587989807, + 0.005754623096436262, + 0.004579542204737663, + 0.005064495839178562, + 0.00334175699390471, + 0.002366000087931752, + 0.005579526070505381, + 0.0041795349679887295, + 0.003286039922386408, + 0.01045033149421215, + 0.010253719054162502, + 0.010153603740036488, + 0.008122955448925495, + 0.007845887914299965, + 0.006975272204726934, + 0.009969916194677353, + 0.007244304288178682, + 0.003947601653635502, + 0.004137394018471241, + 0.005213622469455004, + 0.0027488006744533777, + 0.002835742896422744, + 0.006110843271017075, + 0.0018043075688183308, + 0.0031819334253668785, + 0.007480570115149021, + 0.008880109526216984, + 0.006977649871259928, + 0.003079703776165843, + 0.003604946192353964, + 0.009036589413881302, + 0.004052676726132631, + 0.008166629821062088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.0022060361225157976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053925633430481 + ], + "opposite_logprob_tokens": [ + -0.42892566323280334 + ], + "logprob_reward": -1.053925633430481, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.053925633430481, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05356723442673683, + 0.6165764927864075, + 0.03295538201928139, + 0.04230114817619324, + 0.07240274548530579, + 0.04125971347093582, + 0.0232550036162138, + 0.028417617082595825, + 0.05108952894806862, + 0.03817519545555115 + ], + "credit_map": [ + 0.008568346500396729, + 0.0038255518302321434, + 0.01003712136298418, + 0.006803491152822971, + 0.004979782272130251, + 0.007960507646203041, + 0.005118899047374725, + 0.006273531820625067, + 0.11079130321741104, + 0.03808821365237236, + 0.1121588721871376, + 0.05342460423707962, + 0.12323784083127975, + 0.062482111155986786, + 0.07182394713163376, + 0.04456959292292595, + 0.0038525669369846582, + 0.0025290772318840027, + 0.004654550459235907, + 0.0033550886437296867, + 0.003149203024804592, + 0.007371458224952221, + 0.003357434645295143, + 0.0046860030852258205, + 0.005811154376715422, + 0.0026307860389351845, + 0.006718747317790985, + 0.0026523598935455084, + 0.003542855614796281, + 0.008992877788841724, + 0.005370920989662409, + 0.006581443827599287, + 0.01067397091537714, + 0.007488805800676346, + 0.012325537391006947, + 0.007410010322928429, + 0.007393678184598684, + 0.011759800836443901, + 0.007797400001436472, + 0.00755354156717658, + 0.005908296909183264, + 0.004305744078010321, + 0.006701951380819082, + 0.004195001907646656, + 0.0028115222230553627, + 0.005658090580254793, + 0.00563888531178236, + 0.00604021642357111, + 0.0028624210972338915, + 0.0023943670094013214, + 0.0035243434831500053, + 0.0016206599539145827, + 0.0021257998887449503, + 0.002904046094045043, + 0.00476201344281435, + 0.0030613515991717577, + 0.0033304935786873102, + 0.0037497570738196373, + 0.004025839269161224, + 0.00257001630961895, + 0.0036055571399629116, + 0.0036394521594047546, + 0.005168186500668526, + 0.002328312024474144, + 0.009370931424200535, + 0.008731656707823277, + 0.00782967172563076, + 0.0047248187474906445, + 0.004347790498286486, + 0.006710214074701071, + 0.003097785171121359, + 0.006276654079556465, + 0.004860532470047474, + 0.004342257510870695, + 0.0053947060368955135, + 0.002261640038341284, + 0.003928616642951965, + 0.005982849281281233, + 0.0025419846642762423, + 0.008862607181072235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -1.053925633430481 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "stunningly low-angle, static shot, almost touching the damp forest floor strewn with withered yellow leaves and scattered moss, focuses upward on a single emerald-green bitter bamboo. Behind it, the dense bamboo forest forms a deep, dark vertical barrier, shading the surroundings with an extraordinary tranquility. At this moment, only a single beam of warm, golden sunlight, like a stage spotlight, penetrates the gaps in the forest canopy, casting a precise shadow, creating a distinct Tyndall effect in the dim air, with countless tiny particles dancing slowly within the beam of light. The bamboo stalks illuminated by the light take on the translucent texture of jade, their smooth surface tinged with a faint waxy sheen, while the backlit side is cast in soft shadow, outlining its sturdy outline. Under the effect of time-lapse photography, the bamboo pole is performing a phototropic movement in a stable and persistent posture that is imperceptible to the naked eye. Its slender trunk is slowly bending towards the direction of the light column in a tiny but clearly visible arc, and the few tender green bamboo leaves on the top also stretch out, as if silently declaring its desire for light and life. The whole picture is full of indomitable life tension and hope in the ultimate tranquility.", + "seed": 1443, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_11/index_8388_seed_1443_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3807025113608688, + "vlm_logprob": -0.3807025113608688, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.3807025113608688, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07476234436035156 + ], + "opposite_logprob_tokens": [ + -3.0747623443603516 + ], + "logprob_reward": -0.07476234436035156, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.07476234436035156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1998370885848999, + 0.051889754831790924, + 0.03598087653517723, + 0.022654922679066658, + 0.054041389375925064, + 0.09354815632104874, + 0.10060974210500717, + 0.08482234925031662, + 0.20066606998443604, + 0.1559496372938156 + ], + "credit_map": [ + 0.02266962267458439, + 0.021620528772473335, + 0.02973340079188347, + 0.02230803109705448, + 0.026206854730844498, + 0.03072720393538475, + 0.014577176421880722, + 0.0319942831993103, + 0.0075776418671011925, + 0.00570313073694706, + 0.006274235434830189, + 0.006382942199707031, + 0.003755809972062707, + 0.007408322300761938, + 0.003460282227024436, + 0.011327392421662807, + 0.003848608350381255, + 0.00421922979876399, + 0.00481682363897562, + 0.0034603034146130085, + 0.0032100831158459187, + 0.004394106101244688, + 0.0027408746536821127, + 0.00929084699600935, + 0.0020019756630063057, + 0.0029620116110891104, + 0.00286851660348475, + 0.0023387852124869823, + 0.0030858018435537815, + 0.002575524849817157, + 0.0021921659354120493, + 0.0046301428228616714, + 0.005292079411447048, + 0.006053900811821222, + 0.005838415585458279, + 0.005917612928897142, + 0.008090920746326447, + 0.008156041614711285, + 0.00646985461935401, + 0.008222565054893494, + 0.010257781483232975, + 0.011033895425498486, + 0.01120968908071518, + 0.007193607743829489, + 0.012319773435592651, + 0.011660700663924217, + 0.018060902133584023, + 0.011811812408268452, + 0.010561495088040829, + 0.0050801620818674564, + 0.020929038524627686, + 0.006448002997785807, + 0.007154169026762247, + 0.03454817458987236, + 0.007664065342396498, + 0.008224631659686565, + 0.012052631005644798, + 0.010182920843362808, + 0.011365910992026329, + 0.008468993939459324, + 0.006770731881260872, + 0.010377345606684685, + 0.008937744423747063, + 0.016666067764163017, + 0.02472076192498207, + 0.03865308687090874, + 0.031300079077482224, + 0.01536306831985712, + 0.01662752963602543, + 0.015876293182373047, + 0.019159793853759766, + 0.03896546736359596, + 0.014774099923670292, + 0.022321205586194992, + 0.030444294214248657, + 0.015838302671909332, + 0.015391401015222073, + 0.02745388261973858, + 0.012490523047745228, + 0.017235925421118736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.07476234436035156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134626150131226 + ], + "opposite_logprob_tokens": [ + -0.31346258521080017 + ], + "logprob_reward": -1.3134626150131226, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134626150131226, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11141891032457352, + 0.03932642564177513, + 0.06201985850930214, + 0.05427024886012077, + 0.17929081618785858, + 0.08365917205810547, + 0.09039180725812912, + 0.0807449072599411, + 0.2242826223373413, + 0.07459519803524017 + ], + "credit_map": [ + 0.014155476354062557, + 0.014427580870687962, + 0.01773744635283947, + 0.008622284047305584, + 0.018706997856497765, + 0.017246361821889877, + 0.007433155085891485, + 0.013089616782963276, + 0.006507935933768749, + 0.003931817598640919, + 0.006017709616571665, + 0.00432155467569828, + 0.004165122751146555, + 0.005495677702128887, + 0.003238448640331626, + 0.005648160818964243, + 0.01022107619792223, + 0.008600258268415928, + 0.009358209557831287, + 0.005713669583201408, + 0.008490298874676228, + 0.008985452353954315, + 0.00538267195224762, + 0.0052682263776659966, + 0.008226756937801838, + 0.006921751890331507, + 0.009682678617537022, + 0.006672449875622988, + 0.006564663257449865, + 0.0061829062178730965, + 0.005290523637086153, + 0.004728517960757017, + 0.028281375765800476, + 0.019760040566325188, + 0.023286934942007065, + 0.017611388117074966, + 0.022257639095187187, + 0.026791708543896675, + 0.019001062959432602, + 0.022300684824585915, + 0.00820835493505001, + 0.006962813902646303, + 0.013232828117907047, + 0.007118941284716129, + 0.009224860928952694, + 0.021602962166070938, + 0.008687938563525677, + 0.00862047541886568, + 0.011356326751410961, + 0.008048021234571934, + 0.017725711688399315, + 0.005610176362097263, + 0.007569614332169294, + 0.02307031862437725, + 0.008539235219359398, + 0.008472401648759842, + 0.010307971388101578, + 0.0075630019418895245, + 0.009571049362421036, + 0.008989301510155201, + 0.008444951847195625, + 0.010499890893697739, + 0.011653540655970573, + 0.013715201988816261, + 0.03231535106897354, + 0.01593843661248684, + 0.04644974693655968, + 0.023979276418685913, + 0.010916807688772678, + 0.020785974338650703, + 0.013356257230043411, + 0.06054076552391052, + 0.008310098201036453, + 0.00745937367901206, + 0.01417483575642109, + 0.009255804121494293, + 0.006720364559441805, + 0.006303761620074511, + 0.008178181946277618, + 0.014192781411111355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -1.3134626150131226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038116004317998886 + ], + "opposite_logprob_tokens": [ + -3.288115978240967 + ], + "logprob_reward": -0.038116004317998886, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038116004317998886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19309179484844208, + 0.1020905077457428, + 0.1029767319560051, + 0.02844490110874176, + 0.04034365713596344, + 0.1419450342655182, + 0.18037191033363342, + 0.06507068127393723, + 0.08568144589662552, + 0.05998336896300316 + ], + "credit_map": [ + 0.02831331640481949, + 0.025919705629348755, + 0.02544424869120121, + 0.015582715161144733, + 0.03340841457247734, + 0.03011205792427063, + 0.012518560513854027, + 0.021792761981487274, + 0.0182980764657259, + 0.012629897333681583, + 0.014554424211382866, + 0.008927643299102783, + 0.013064151629805565, + 0.01355778332799673, + 0.008608018979430199, + 0.012450512498617172, + 0.01491508912295103, + 0.011077501811087132, + 0.015006684698164463, + 0.008332004770636559, + 0.017402438446879387, + 0.014907309785485268, + 0.009143028408288956, + 0.012192669324576855, + 0.0031897348817437887, + 0.003466961905360222, + 0.0032263773027807474, + 0.0028160123620182276, + 0.004357360303401947, + 0.003870758693665266, + 0.002862054156139493, + 0.004655640106648207, + 0.005462379660457373, + 0.004493865184485912, + 0.006892884150147438, + 0.004284475464373827, + 0.004029593896120787, + 0.0056831068359315395, + 0.004505111835896969, + 0.004992240108549595, + 0.010164232924580574, + 0.010540270246565342, + 0.023135153576731682, + 0.009235800243914127, + 0.016113899648189545, + 0.03833014890551567, + 0.019630149006843567, + 0.014795375056564808, + 0.01742623560130596, + 0.008334443904459476, + 0.033277854323387146, + 0.00913198385387659, + 0.010731624439358711, + 0.06170721724629402, + 0.025024453178048134, + 0.014738080091774464, + 0.011991376988589764, + 0.007155356928706169, + 0.008608289062976837, + 0.00906433630734682, + 0.004322273191064596, + 0.008642611093819141, + 0.005924998316913843, + 0.009361435659229755, + 0.01178729347884655, + 0.009203493595123291, + 0.01593729294836521, + 0.00970454141497612, + 0.005742816254496574, + 0.006856830790638924, + 0.007023905869573355, + 0.01942526176571846, + 0.007944297976791859, + 0.006299746688455343, + 0.006358010694384575, + 0.005695966072380543, + 0.005841952282935381, + 0.01099927444010973, + 0.00829747598618269, + 0.008546646684408188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.038116004317998886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028272203635424376 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028272203635424376, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028272203635424376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0800982192158699, + 0.044972870498895645, + 0.04738027974963188, + 0.04228091984987259, + 0.07002584636211395, + 0.1470138281583786, + 0.21602387726306915, + 0.1206386461853981, + 0.16194452345371246, + 0.06962095201015472 + ], + "credit_map": [ + 0.012976676225662231, + 0.011779083870351315, + 0.009972885251045227, + 0.0054025789722800255, + 0.014537087641656399, + 0.011530688963830471, + 0.005947950296103954, + 0.007951266132295132, + 0.006482618860900402, + 0.006781889591366053, + 0.00653492845594883, + 0.003860178869217634, + 0.005427589640021324, + 0.006842855364084244, + 0.0038231443613767624, + 0.005219663958996534, + 0.007962307892739773, + 0.005282269790768623, + 0.007695815525949001, + 0.004617270082235336, + 0.004915403667837381, + 0.007569244131445885, + 0.0039825732819736, + 0.005355394445359707, + 0.005045202095061541, + 0.0044904123060405254, + 0.0077907671220600605, + 0.004199374467134476, + 0.004112031776458025, + 0.006132996641099453, + 0.004472315311431885, + 0.006037821061909199, + 0.00961877778172493, + 0.007313001900911331, + 0.007607426960021257, + 0.006465693935751915, + 0.012658776715397835, + 0.009995084255933762, + 0.008254638873040676, + 0.008112446404993534, + 0.011155505664646626, + 0.010815930552780628, + 0.024348193779587746, + 0.00911768153309822, + 0.017911547794938087, + 0.038580089807510376, + 0.018898453563451767, + 0.016186431050300598, + 0.018824748694896698, + 0.010629425756633282, + 0.03959205374121666, + 0.009819752536714077, + 0.014763394370675087, + 0.07158456742763519, + 0.035376742482185364, + 0.015433189459145069, + 0.011535831727087498, + 0.01190126221626997, + 0.013190473429858685, + 0.011278508231043816, + 0.013915768824517727, + 0.023675279691815376, + 0.014044921845197678, + 0.021096600219607353, + 0.01787119172513485, + 0.011553918942809105, + 0.033889271318912506, + 0.015828445553779602, + 0.009878424927592278, + 0.01483219861984253, + 0.01469248253852129, + 0.04339858889579773, + 0.008739524520933628, + 0.00918128713965416, + 0.009084479883313179, + 0.006993208546191454, + 0.007800669874995947, + 0.012821975164115429, + 0.007768799550831318, + 0.007231013383716345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.0028272203635424376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost touching the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without any panning or moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single emerald-green bamboo stalk the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single beam of warm golden sunlight visible penetrating the canopy and creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bamboo stalk visibly bending slowly towards the light source in a time-lapse style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743443727493286 + ], + "opposite_logprob_tokens": [ + -0.9743443727493286 + ], + "logprob_reward": -0.4743443727493286, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743443727493286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.102840356528759, + 0.05634743720293045, + 0.06404455751180649, + 0.03894516080617905, + 0.21746312081813812, + 0.1396891474723816, + 0.12327732145786285, + 0.0548095665872097, + 0.12527243793010712, + 0.07731087505817413 + ], + "credit_map": [ + 0.011818005703389645, + 0.01593288779258728, + 0.014267098158597946, + 0.009040707722306252, + 0.01645701564848423, + 0.01448633149266243, + 0.008656325750052929, + 0.01218198798596859, + 0.010839326307177544, + 0.006058949511498213, + 0.008976882323622704, + 0.004307278431952, + 0.007369427476078272, + 0.007915276102721691, + 0.005886500235646963, + 0.004993796348571777, + 0.009752468205988407, + 0.007956558838486671, + 0.009022913873195648, + 0.0055345939472317696, + 0.00978846289217472, + 0.009482466615736485, + 0.005313392728567123, + 0.0071937027387320995, + 0.003951731603592634, + 0.004535579588264227, + 0.00512529443949461, + 0.004017000086605549, + 0.004504439886659384, + 0.006128118839114904, + 0.0034974392037838697, + 0.007185555994510651, + 0.03482857346534729, + 0.02397654578089714, + 0.02925492450594902, + 0.02161143533885479, + 0.029626401141285896, + 0.035041652619838715, + 0.020438451319932938, + 0.022685138508677483, + 0.011512620374560356, + 0.01352049969136715, + 0.015817735344171524, + 0.01014329306781292, + 0.028378698974847794, + 0.023793619126081467, + 0.024156814441084862, + 0.012365875765681267, + 0.01236774679273367, + 0.006983923725783825, + 0.013180390000343323, + 0.008814464323222637, + 0.005620066076517105, + 0.03886331617832184, + 0.024139761924743652, + 0.013307657092809677, + 0.0064072092063724995, + 0.0072957612574100494, + 0.00575898215174675, + 0.006396247074007988, + 0.006261082831770182, + 0.00872729904949665, + 0.00657423073425889, + 0.007388753350824118, + 0.018836919218301773, + 0.008352682925760746, + 0.027547188103199005, + 0.009492750279605389, + 0.005745150148868561, + 0.012528551742434502, + 0.006033784709870815, + 0.0367354080080986, + 0.008724549785256386, + 0.008505333214998245, + 0.014519271440804005, + 0.01088848989456892, + 0.006652658339589834, + 0.0059431931003928185, + 0.008484448306262493, + 0.013592934235930443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.4743443727493286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_12/index_tensor([8605, 8605])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_12/index_tensor([8605, 8605])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..930d46544e2d0f8f5ada42d7f75adc748958da7e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_12/index_tensor([8605, 8605])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle drone shot smoothly glides from left to right, clinging to the tops of a vast, golden wheat field. Under the low, evening sun, the field is tinted a warm honey hue. The sky is a deep blue, dotted with a few puffy white clouds whipped by the wind. Each ripe stalk bears a heavy load of plump ears, their delicate awns gleaming in the light. Suddenly, a powerful, invisible gust of wind sweeps through, forcing the stalks to bend uniformly, forming a spectacular, undulating wave. This golden wave of wheat rapidly spreads from one side of the camera toward the distant horizon, one layer chasing another, creating a flowing, rhythmic visual spectacle across the vast field. Light dances and flickers on the swaying ears of wheat, contrasting sharply with the shifting shadows cast by the waves. The entire image is imbued with dynamic vitality and a majestic yet tranquil natural beauty.", + "seed": 7436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_12/index_8605_seed_7436_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.058088432624936105, + "vlm_logprob": -0.058088432624936105, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.058088432624936105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03492220863699913 + ], + "opposite_logprob_tokens": [ + -4.784922122955322 + ], + "logprob_reward": -0.03492220863699913, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.03492220863699913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045555368065834045, + 0.03555682301521301, + 0.33231592178344727, + 0.022937344387173653, + 0.22316399216651917, + 0.2684254050254822, + 0.024182744324207306, + 0.00939824990928173, + 0.01899084448814392, + 0.019473319873213768 + ], + "credit_map": [ + 0.002967642620205879, + 0.003783626714721322, + 0.006246237549930811, + 0.007717432454228401, + 0.00304866349324584, + 0.003117864252999425, + 0.012427941896021366, + 0.006245958618819714, + 0.001632243162021041, + 0.0015247657429426908, + 0.007426929194480181, + 0.008452652022242546, + 0.002167799975723028, + 0.0024362364783883095, + 0.00598223228007555, + 0.005933966021984816, + 0.012625878676772118, + 0.010660837404429913, + 0.09784550219774246, + 0.077337846159935, + 0.01583012379705906, + 0.023735251277685165, + 0.05353063717484474, + 0.04074983671307564, + 0.0015386129962280393, + 0.00395915936678648, + 0.002448070328682661, + 0.0023727372754365206, + 0.002254806226119399, + 0.005639899056404829, + 0.0022903650533407927, + 0.002433694899082184, + 0.019718987867236137, + 0.012054064311087132, + 0.04043542593717575, + 0.042982302606105804, + 0.014945102855563164, + 0.0188260730355978, + 0.05812585726380348, + 0.016076192259788513, + 0.01421220414340496, + 0.05299300700426102, + 0.030563732609152794, + 0.041862133890390396, + 0.026853548362851143, + 0.07299869507551193, + 0.010696128942072392, + 0.018245965242385864, + 0.0011412224266678095, + 0.0022885131184011698, + 0.00431122025474906, + 0.006106190383434296, + 0.0015238557243719697, + 0.0012264238903298974, + 0.004217343870550394, + 0.0033679730258882046, + 0.0009188198018819094, + 0.0016208203742280602, + 0.0010019465116783977, + 0.0012743008555844426, + 0.0011635039700195193, + 0.0009643253870308399, + 0.0011923323618248105, + 0.0012622002977877855, + 0.0009882486192509532, + 0.001000471762381494, + 0.0037329071201384068, + 0.0025365855544805527, + 0.0018056733533740044, + 0.0014102833811193705, + 0.003661130787804723, + 0.003855545539408922, + 0.0009930672822520137, + 0.001549628097563982, + 0.00394553504884243, + 0.003919707611203194, + 0.0025163190439343452, + 0.0013698912225663662, + 0.0033467765897512436, + 0.0018323964904993773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.03492220863699913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293117702007294 + ], + "opposite_logprob_tokens": [ + -2.237931251525879 + ], + "logprob_reward": -0.11293117702007294, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293117702007294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039736583828926086, + 0.06143421679735184, + 0.6130807399749756, + 0.07066450268030167, + 0.12636055052280426, + 0.031433314085006714, + 0.016054803505539894, + 0.009812108241021633, + 0.015958795323967934, + 0.015464379452168941 + ], + "credit_map": [ + 0.003050568513572216, + 0.0037226234562695026, + 0.006406883709132671, + 0.006463335826992989, + 0.0027185287326574326, + 0.0026705064810812473, + 0.009142028167843819, + 0.005562109407037497, + 0.0025738300755620003, + 0.0029354484286159277, + 0.010860281996428967, + 0.018687503412365913, + 0.0035600936971604824, + 0.005199036560952663, + 0.008630705066025257, + 0.00898731593042612, + 0.025198865681886673, + 0.03348185122013092, + 0.18094362318515778, + 0.15344353020191193, + 0.022388296201825142, + 0.024481426924467087, + 0.0974898487329483, + 0.07565327733755112, + 0.0029031289741396904, + 0.009649005718529224, + 0.013821396045386791, + 0.008087404072284698, + 0.006683479994535446, + 0.01528251450508833, + 0.0076055065728724, + 0.006632068660110235, + 0.017440615221858025, + 0.00803324393928051, + 0.03383447229862213, + 0.026108337566256523, + 0.006009511183947325, + 0.008941921405494213, + 0.019623668864369392, + 0.006368785630911589, + 0.0015226179966703057, + 0.003503454150632024, + 0.005476388614624739, + 0.00410654628649354, + 0.003134828293696046, + 0.007433109451085329, + 0.002681286307051778, + 0.0035750865936279297, + 0.0020208386704325676, + 0.0019414308480918407, + 0.004185823258012533, + 0.0025828012730926275, + 0.0013682589633390307, + 0.00095329963369295, + 0.0015262749511748552, + 0.0014760751510038972, + 0.0005243127816356719, + 0.0014148545451462269, + 0.0013688625767827034, + 0.001143465400673449, + 0.0017296129371970892, + 0.0015525053022429347, + 0.0011218782747164369, + 0.0009566160151734948, + 0.0011332256253808737, + 0.0012656666804105043, + 0.002457085531204939, + 0.0018313268665224314, + 0.0016979229403659701, + 0.0030211806297302246, + 0.0021472168155014515, + 0.00240516965277493, + 0.001147580100223422, + 0.0011589049827307463, + 0.0034031367395073175, + 0.002862946130335331, + 0.001136085600592196, + 0.001217008219100535, + 0.0023510281462222338, + 0.00218768953345716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.11293117702007294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581373944878578 + ], + "opposite_logprob_tokens": [ + -3.8955814838409424 + ], + "logprob_reward": -0.020581373944878578, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581373944878578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14487139880657196, + 0.023326445370912552, + 0.11933288723230362, + 0.06951776146888733, + 0.18852558732032776, + 0.09687357395887375, + 0.2857123911380768, + 0.02622545324265957, + 0.008457454852759838, + 0.03715704753994942 + ], + "credit_map": [ + 0.0073143127374351025, + 0.010669868439435959, + 0.021063176915049553, + 0.025981713086366653, + 0.006997767370194197, + 0.008619654923677444, + 0.03629481419920921, + 0.02793009765446186, + 0.0017918472876772285, + 0.0022994615137577057, + 0.004393788520246744, + 0.00636870414018631, + 0.0014180622529238462, + 0.0014106431044638157, + 0.002330741146579385, + 0.0033131956588476896, + 0.005755534861236811, + 0.022411642596125603, + 0.023914486169815063, + 0.0244914498180151, + 0.007408948615193367, + 0.022633766755461693, + 0.006780303083360195, + 0.005936765111982822, + 0.002330723451450467, + 0.006147871259599924, + 0.01965043693780899, + 0.010493888519704342, + 0.005965674761682749, + 0.01386556401848793, + 0.005928151775151491, + 0.005135449580848217, + 0.010937033221125603, + 0.009982752613723278, + 0.0421510674059391, + 0.04158243536949158, + 0.011868034489452839, + 0.008346718735992908, + 0.049003489315509796, + 0.014654064550995827, + 0.0052891443483531475, + 0.018418291583657265, + 0.013339651748538017, + 0.01565912365913391, + 0.009380464442074299, + 0.024216484278440475, + 0.004238158464431763, + 0.006332253105938435, + 0.017139967530965805, + 0.045765865594148636, + 0.04693146049976349, + 0.05389123409986496, + 0.021381864324212074, + 0.012124063447117805, + 0.05318580940365791, + 0.0352921187877655, + 0.0019725339952856302, + 0.0027325740084052086, + 0.008394446223974228, + 0.003434095997363329, + 0.0017472900217399001, + 0.0020859986543655396, + 0.003695049090310931, + 0.002163467463105917, + 0.0007280944264493883, + 0.0008669172530062497, + 0.002483280375599861, + 0.0010989014990627766, + 0.0006800686824135482, + 0.0006579992477782071, + 0.0011251058895140886, + 0.0008170875371433794, + 0.002705559367313981, + 0.0070215738378465176, + 0.008469212800264359, + 0.0055287061259150505, + 0.004732366651296616, + 0.002496063243597746, + 0.0033682631328701973, + 0.002835302846506238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.020581373944878578 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896124571561813 + ], + "opposite_logprob_tokens": [ + -2.578961133956909 + ], + "logprob_reward": -0.07896124571561813, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896124571561813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0454372800886631, + 0.05194602906703949, + 0.4246116578578949, + 0.03966666758060455, + 0.0771930143237114, + 0.2261923849582672, + 0.0637839138507843, + 0.017963608726859093, + 0.04243546724319458, + 0.010769939050078392 + ], + "credit_map": [ + 0.0028689547907561064, + 0.0032970383763313293, + 0.00948338769376278, + 0.007437651976943016, + 0.002538747154176235, + 0.0026616295799613, + 0.011808590032160282, + 0.0053412835113704205, + 0.0020619917195290327, + 0.0027019137050956488, + 0.010630794800817966, + 0.01215103268623352, + 0.003213424701243639, + 0.005194882862269878, + 0.00749394204467535, + 0.008498052135109901, + 0.017154455184936523, + 0.021653681993484497, + 0.1321873962879181, + 0.10436303168535233, + 0.015673182904720306, + 0.01711302064359188, + 0.06641004234552383, + 0.050056926906108856, + 0.001463345717638731, + 0.003732845652848482, + 0.009109452366828918, + 0.0051543028093874454, + 0.003504714462906122, + 0.0073344530537724495, + 0.004634443204849958, + 0.00473311310634017, + 0.006788998376578093, + 0.003966550808399916, + 0.010615121573209763, + 0.012285488657653332, + 0.007429752964526415, + 0.013856595382094383, + 0.016415879130363464, + 0.005834630690515041, + 0.012687914073467255, + 0.04668407887220383, + 0.020062867552042007, + 0.028881685808300972, + 0.024987000972032547, + 0.06727629154920578, + 0.009192220866680145, + 0.016420362517237663, + 0.005738466512411833, + 0.00952224899083376, + 0.008823458105325699, + 0.011898081749677658, + 0.004809397738426924, + 0.003265253035351634, + 0.011350656859576702, + 0.00837635062634945, + 0.0015368959866464138, + 0.001685994560830295, + 0.0031411023810505867, + 0.0028064833022654057, + 0.0014641802990809083, + 0.0015404665609821677, + 0.0028998269699513912, + 0.002888660877943039, + 0.0023421195801347494, + 0.0024846121668815613, + 0.010690352879464626, + 0.006908239331096411, + 0.0022860593162477016, + 0.00420248182490468, + 0.006819960195571184, + 0.006701648700982332, + 0.0009635525057092309, + 0.002283671172335744, + 0.001267233514226973, + 0.0014411351876333356, + 0.0017967505846172571, + 0.000933047675061971, + 0.001036892761476338, + 0.0010476572206243873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.07896124571561813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304615780711174 + ], + "opposite_logprob_tokens": [ + -3.168046236038208 + ], + "logprob_reward": -0.04304615780711174, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304615780711174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07100866734981537, + 0.041655283421278, + 0.36938244104385376, + 0.04239704832434654, + 0.049210935831069946, + 0.23763753473758698, + 0.12482452392578125, + 0.014019111171364784, + 0.035400353372097015, + 0.014464057050645351 + ], + "credit_map": [ + 0.0055025070905685425, + 0.005164622329175472, + 0.013380085118114948, + 0.012012179009616375, + 0.003480643965303898, + 0.0037660920061171055, + 0.016741333529353142, + 0.010961206629872322, + 0.0018462014850229025, + 0.002198132686316967, + 0.006776087451726198, + 0.014212872833013535, + 0.0022886176593601704, + 0.0031191569287329912, + 0.0054483674466609955, + 0.005765845067799091, + 0.014660650864243507, + 0.01968967914581299, + 0.11640288680791855, + 0.09966592490673065, + 0.009625055827200413, + 0.016838936135172844, + 0.05223311111330986, + 0.04026620462536812, + 0.0017105472506955266, + 0.004063003230839968, + 0.010803836397826672, + 0.0064008780755102634, + 0.003404895542189479, + 0.008003108203411102, + 0.0045334878377616405, + 0.0034772909712046385, + 0.002947168191894889, + 0.002818226581439376, + 0.008967387489974499, + 0.009127546101808548, + 0.004832569509744644, + 0.007699259091168642, + 0.009022563695907593, + 0.0037962167989462614, + 0.013224304653704166, + 0.04828183352947235, + 0.024966802448034286, + 0.033069953322410583, + 0.02502277120947838, + 0.06765461713075638, + 0.009280762635171413, + 0.016136491671204567, + 0.009873870760202408, + 0.024752169847488403, + 0.019687646999955177, + 0.0175345316529274, + 0.010069923475384712, + 0.00611903565004468, + 0.021315395832061768, + 0.015471961349248886, + 0.0008437532815150917, + 0.0013235214864835143, + 0.002431815955787897, + 0.0019939099438488483, + 0.0014541250420734286, + 0.0013904499355703592, + 0.002239528112113476, + 0.0023420073557645082, + 0.001641061739064753, + 0.0025077369064092636, + 0.00502014858648181, + 0.0050719063729047775, + 0.003984118811786175, + 0.005419098772108555, + 0.007068141363561153, + 0.004688141867518425, + 0.001343266572803259, + 0.002663056831806898, + 0.0023738869931548834, + 0.0020497054792940617, + 0.0016751840012148023, + 0.0007613538764417171, + 0.0016585624543949962, + 0.001939040725119412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.04304615780711174 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle drone shot smoothly glides from left to right, clinging to the tops of a vast, golden wheat field. Under the low, evening sun, the field is tinted a warm honey hue. The sky is a deep blue, dotted with a few puffy white clouds whipped by the wind. Each ripe stalk bears a heavy load of plump ears, their delicate awns gleaming in the light. Suddenly, a powerful, invisible gust of wind sweeps through, forcing the stalks to bend uniformly, forming a spectacular, undulating wave. This golden wave of wheat rapidly spreads from one side of the camera toward the distant horizon, one layer chasing another, creating a flowing, rhythmic visual spectacle across the vast field. Light dances and flickers on the swaying ears of wheat, contrasting sharply with the shifting shadows cast by the waves. The entire image is imbued with dynamic vitality and a majestic yet tranquil natural beauty.", + "seed": 7437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_12/index_8605_seed_7437_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0817927511408925, + "vlm_logprob": -0.0817927511408925, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.0817927511408925, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051644567400217056 + ], + "opposite_logprob_tokens": [ + -4.051644802093506 + ], + "logprob_reward": -0.051644567400217056, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.051644567400217056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21241562068462372, + 0.10673422366380692, + 0.13875575363636017, + 0.06836947053670883, + 0.131971538066864, + 0.06790148466825485, + 0.05949123576283455, + 0.08906774967908859, + 0.052217673510313034, + 0.07307519763708115 + ], + "credit_map": [ + 0.014018687419593334, + 0.018048230558633804, + 0.053753066807985306, + 0.026561314240098, + 0.019640397280454636, + 0.022739987820386887, + 0.033655840903520584, + 0.023998122662305832, + 0.005098000168800354, + 0.004213194362819195, + 0.02320411056280136, + 0.013159168884158134, + 0.007105272728949785, + 0.006338735111057758, + 0.034456364810466766, + 0.013159391470253468, + 0.008712638169527054, + 0.012608733959496021, + 0.03798828646540642, + 0.02261948585510254, + 0.011499256826937199, + 0.010795379988849163, + 0.01734808087348938, + 0.017183905467391014, + 0.005959200672805309, + 0.013219031505286694, + 0.00836479477584362, + 0.009034795686602592, + 0.007545466534793377, + 0.00673926156014204, + 0.010125069878995419, + 0.007381854113191366, + 0.007275615818798542, + 0.008874899707734585, + 0.04545748978853226, + 0.013257015496492386, + 0.011268120259046555, + 0.011590367183089256, + 0.011457955464720726, + 0.022790085524320602, + 0.003746578935533762, + 0.003612121567130089, + 0.008066750131547451, + 0.014198769815266132, + 0.009069588966667652, + 0.006458954885601997, + 0.013573224656283855, + 0.009175499901175499, + 0.0031220358796417713, + 0.004344166722148657, + 0.007892420515418053, + 0.014208084903657436, + 0.007335135247558355, + 0.004672090522944927, + 0.011223319917917252, + 0.006693986244499683, + 0.0054222578182816505, + 0.006632546428591013, + 0.01841726154088974, + 0.009614954702556133, + 0.0070479935966432095, + 0.009395786561071873, + 0.021553151309490204, + 0.010983807034790516, + 0.005887454375624657, + 0.005446176044642925, + 0.007985143922269344, + 0.00988609530031681, + 0.004107230808585882, + 0.0032688763458281755, + 0.007320387754589319, + 0.008316312916576862, + 0.005690490361303091, + 0.004387153312563896, + 0.011595407500863075, + 0.00946751981973648, + 0.008410352282226086, + 0.009295396506786346, + 0.008414189331233501, + 0.015814704820513725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.051644567400217056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706464529037476 + ], + "opposite_logprob_tokens": [ + -2.1270647048950195 + ], + "logprob_reward": -0.12706464529037476, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706464529037476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21469011902809143, + 0.08968625217676163, + 0.09141615033149719, + 0.08029713481664658, + 0.08902950584888458, + 0.0647357702255249, + 0.07083332538604736, + 0.07879940420389175, + 0.07174426317214966, + 0.1487680822610855 + ], + "credit_map": [ + 0.012425404973328114, + 0.016573866829276085, + 0.046437911689281464, + 0.035274118185043335, + 0.011884885840117931, + 0.020497802644968033, + 0.04878779500722885, + 0.022808333858847618, + 0.004180368967354298, + 0.004040209110826254, + 0.011487043462693691, + 0.01517593115568161, + 0.007562210783362389, + 0.005543833132833242, + 0.017208490520715714, + 0.0244881734251976, + 0.007152614649385214, + 0.009398745372891426, + 0.015098047442734241, + 0.01572331041097641, + 0.008907581679522991, + 0.010588890872895718, + 0.014204919338226318, + 0.010342035442590714, + 0.005522159393876791, + 0.0039249262772500515, + 0.01651206985116005, + 0.011290001682937145, + 0.00771857937797904, + 0.00522617157548666, + 0.020712492987513542, + 0.009390732273459435, + 0.0053433519788086414, + 0.008273826912045479, + 0.029305461794137955, + 0.01181020773947239, + 0.007104095537215471, + 0.00775319105014205, + 0.009200888685882092, + 0.010238484479486942, + 0.004355749115347862, + 0.006164239253848791, + 0.01081795897334814, + 0.012913468293845654, + 0.004640397150069475, + 0.004561775363981724, + 0.011532421223819256, + 0.009749754332005978, + 0.004178127273917198, + 0.0038615632802248, + 0.014139576815068722, + 0.014550602994859219, + 0.006196692120283842, + 0.005552508868277073, + 0.01146952249109745, + 0.010884731076657772, + 0.004930668976157904, + 0.0038647938054054976, + 0.019937966018915176, + 0.011625479906797409, + 0.004613330587744713, + 0.0058069308288395405, + 0.016546864062547684, + 0.011473366990685463, + 0.005171079654246569, + 0.005671469960361719, + 0.011393360793590546, + 0.021833842620253563, + 0.004808655474334955, + 0.004897525999695063, + 0.011306448839604855, + 0.006661879830062389, + 0.006888608448207378, + 0.006418152712285519, + 0.039158012717962265, + 0.02083090879023075, + 0.008393106050789356, + 0.00910684559494257, + 0.03099098429083824, + 0.026981474831700325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.12706464529037476 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201176352798939 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201176352798939, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201176352798939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07528175413608551, + 0.15304763615131378, + 0.3185308277606964, + 0.06328999996185303, + 0.1967918574810028, + 0.03550779074430466, + 0.05674184858798981, + 0.06620901077985764, + 0.01657988503575325, + 0.01801944337785244 + ], + "credit_map": [ + 0.003893704852089286, + 0.005963281728327274, + 0.018003735691308975, + 0.010350566357374191, + 0.007308553904294968, + 0.00736653059720993, + 0.01438989955931902, + 0.008005479350686073, + 0.011858129873871803, + 0.012709449045360088, + 0.024968888610601425, + 0.02035234309732914, + 0.01513804029673338, + 0.0215006023645401, + 0.012734923511743546, + 0.033785246312618256, + 0.024667274206876755, + 0.03570220246911049, + 0.060303423553705215, + 0.046391021460294724, + 0.02944074384868145, + 0.04198889434337616, + 0.04803967475891113, + 0.03199757635593414, + 0.004861898720264435, + 0.004256588872522116, + 0.01082003302872181, + 0.011847673915326595, + 0.0040237922221422195, + 0.0032703483011573553, + 0.015782739967107773, + 0.008426920510828495, + 0.019003959372639656, + 0.02488815039396286, + 0.05575571209192276, + 0.02594691328704357, + 0.015247257426381111, + 0.015382181853055954, + 0.02601630426943302, + 0.01455135177820921, + 0.00178263068664819, + 0.002264959504827857, + 0.004623965825885534, + 0.006473016459494829, + 0.003394803497940302, + 0.0038664527237415314, + 0.007602738682180643, + 0.005499218124896288, + 0.004196631722152233, + 0.004624188877642155, + 0.007740669883787632, + 0.009525073692202568, + 0.00574624864384532, + 0.004321380518376827, + 0.011280626058578491, + 0.009307026863098145, + 0.0047982800751924515, + 0.008165719918906689, + 0.010876387357711792, + 0.007279893383383751, + 0.005308551248162985, + 0.008808924816548824, + 0.010280167683959007, + 0.010691084899008274, + 0.0011813532328233123, + 0.0012576705776154995, + 0.002337049227207899, + 0.004046578425914049, + 0.0010332422098144889, + 0.0012398249236866832, + 0.00277448957785964, + 0.0027096769772469997, + 0.0011083006393164396, + 0.001517935306765139, + 0.002591100288555026, + 0.0031189671717584133, + 0.0019131881417706609, + 0.0015439079143106937, + 0.0029135397635400295, + 0.003312502522021532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.014201176352798939 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902201056480408 + ], + "opposite_logprob_tokens": [ + -2.464021921157837 + ], + "logprob_reward": -0.08902201056480408, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902201056480408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2068125456571579, + 0.050817057490348816, + 0.10918799042701721, + 0.08911393582820892, + 0.1417776644229889, + 0.0818651020526886, + 0.06904885917901993, + 0.11823632568120956, + 0.04140230268239975, + 0.09173819422721863 + ], + "credit_map": [ + 0.010222754441201687, + 0.011988679878413677, + 0.04991259425878525, + 0.03860153257846832, + 0.010356901213526726, + 0.018266156315803528, + 0.04518824815750122, + 0.02227567695081234, + 0.003653498599305749, + 0.006020120345056057, + 0.006286328192800283, + 0.008791008964180946, + 0.004287604708224535, + 0.008435225114226341, + 0.007100897841155529, + 0.0062423706986010075, + 0.007622333709150553, + 0.00995128694921732, + 0.0182308629155159, + 0.017732465639710426, + 0.009909098967909813, + 0.013657232746481895, + 0.01814000867307186, + 0.013944701291620731, + 0.005991199053823948, + 0.012512517161667347, + 0.00978576298803091, + 0.01683632656931877, + 0.007792775519192219, + 0.007678248919546604, + 0.01783222332596779, + 0.010684878565371037, + 0.011305362917482853, + 0.00997807364910841, + 0.023247499018907547, + 0.017272137105464935, + 0.012802588753402233, + 0.015577571466565132, + 0.019012391567230225, + 0.032582033425569534, + 0.003687516786158085, + 0.004022178705781698, + 0.011941535398364067, + 0.019100209698081017, + 0.00771986972540617, + 0.0049339341931045055, + 0.017500905320048332, + 0.012958956882357597, + 0.0039046667516231537, + 0.004147143103182316, + 0.011942556127905846, + 0.01781598851084709, + 0.006409354042261839, + 0.003371855244040489, + 0.011316884309053421, + 0.01014041155576706, + 0.00827698688954115, + 0.006615166086703539, + 0.029688268899917603, + 0.017990034073591232, + 0.0075446804985404015, + 0.008352450095117092, + 0.021409641951322556, + 0.018359094858169556, + 0.002902862848713994, + 0.0026721572503447533, + 0.007649344392120838, + 0.007930796593427658, + 0.0032125941943377256, + 0.003493828931823373, + 0.008740001358091831, + 0.004800716415047646, + 0.003571844659745693, + 0.003058207221329212, + 0.02136414311826229, + 0.012792253866791725, + 0.003873778274282813, + 0.005281167570501566, + 0.020111816003918648, + 0.021684983745217323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.08902201056480408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270313560962677 + ], + "opposite_logprob_tokens": [ + -2.1270313262939453 + ], + "logprob_reward": -0.1270313560962677, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270313560962677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24073779582977295, + 0.07929234206676483, + 0.295560747385025, + 0.04595031216740608, + 0.035466376692056656, + 0.04632958769798279, + 0.04111088067293167, + 0.08676821738481522, + 0.039537228643894196, + 0.0892464742064476 + ], + "credit_map": [ + 0.010201580822467804, + 0.007786449510604143, + 0.09154593199491501, + 0.03468761220574379, + 0.010068956762552261, + 0.011868057772517204, + 0.037815023213624954, + 0.03676421940326691, + 0.0057006836868822575, + 0.006631277967244387, + 0.012052851729094982, + 0.010591058060526848, + 0.008629605174064636, + 0.006869207136332989, + 0.011694889515638351, + 0.017122782766819, + 0.02305956929922104, + 0.03309791162610054, + 0.05641169846057892, + 0.04416817054152489, + 0.026844484731554985, + 0.038523368537425995, + 0.04430919885635376, + 0.029146354645490646, + 0.00359487347304821, + 0.0041673495434224606, + 0.006973092909902334, + 0.008061419241130352, + 0.004040735308080912, + 0.003438257845118642, + 0.009579400531947613, + 0.006095190066844225, + 0.0028155799955129623, + 0.0025936844758689404, + 0.007353332359343767, + 0.004114669747650623, + 0.00366259110160172, + 0.0034448804799467325, + 0.0033793521579355, + 0.008102291263639927, + 0.0028160992078483105, + 0.004554107319563627, + 0.006397504825145006, + 0.007580335717648268, + 0.003776079509407282, + 0.004316467791795731, + 0.008624773472547531, + 0.008264226838946342, + 0.0031231215689331293, + 0.002589860465377569, + 0.004137442447245121, + 0.005875993985682726, + 0.0041708629578351974, + 0.00465912651270628, + 0.007036891300231218, + 0.009517584927380085, + 0.005364201497286558, + 0.006918152328580618, + 0.015933619812130928, + 0.013332249596714973, + 0.0049909851513803005, + 0.00545978918671608, + 0.021857770159840584, + 0.012911457568407059, + 0.003576622111722827, + 0.0047048176638782024, + 0.00528049236163497, + 0.011264721862971783, + 0.002723908517509699, + 0.003613377222791314, + 0.005016796290874481, + 0.003356498433277011, + 0.004104208666831255, + 0.0042489091865718365, + 0.022128446027636528, + 0.014324981719255447, + 0.00749195134267211, + 0.008516300469636917, + 0.01571762189269066, + 0.012714060954749584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.1270313560962677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle drone shot smoothly glides from left to right, clinging to the tops of a vast, golden wheat field. Under the low, evening sun, the field is tinted a warm honey hue. The sky is a deep blue, dotted with a few puffy white clouds whipped by the wind. Each ripe stalk bears a heavy load of plump ears, their delicate awns gleaming in the light. Suddenly, a powerful, invisible gust of wind sweeps through, forcing the stalks to bend uniformly, forming a spectacular, undulating wave. This golden wave of wheat rapidly spreads from one side of the camera toward the distant horizon, one layer chasing another, creating a flowing, rhythmic visual spectacle across the vast field. Light dances and flickers on the swaying ears of wheat, contrasting sharply with the shifting shadows cast by the waves. The entire image is imbued with dynamic vitality and a majestic yet tranquil natural beauty.", + "seed": 7438, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_12/index_8605_seed_7438_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12334448546171188, + "vlm_logprob": -0.12334448546171188, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.12334448546171188, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06745936721563339 + ], + "opposite_logprob_tokens": [ + -3.5674593448638916 + ], + "logprob_reward": -0.06745936721563339, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.06745936721563339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11330302804708481, + 0.04202180728316307, + 0.11831577122211456, + 0.19071641564369202, + 0.043166834861040115, + 0.05106962472200394, + 0.04278520122170448, + 0.06148676946759224, + 0.08787088841199875, + 0.24926364421844482 + ], + "credit_map": [ + 0.00841984897851944, + 0.01832815818488598, + 0.016059216111898422, + 0.01149672083556652, + 0.015094692818820477, + 0.014749356545507908, + 0.015317265875637531, + 0.013837764970958233, + 0.0031365316826850176, + 0.004224380943924189, + 0.006179844960570335, + 0.011062509380280972, + 0.003043889533728361, + 0.0033109167125076056, + 0.006992978043854237, + 0.0040707578882575035, + 0.005061038304120302, + 0.006391963921487331, + 0.018541300669312477, + 0.03173156455159187, + 0.007940891198813915, + 0.00830724649131298, + 0.023763809353113174, + 0.016577959060668945, + 0.007181542459875345, + 0.01212940365076065, + 0.03064674884080887, + 0.03342242166399956, + 0.01965157315135002, + 0.02202891930937767, + 0.04802248999476433, + 0.017633309587836266, + 0.002551440615206957, + 0.004786187317222357, + 0.007427605800330639, + 0.008199152536690235, + 0.0032890024594962597, + 0.003391240257769823, + 0.006563559174537659, + 0.006958644837141037, + 0.002913655247539282, + 0.0045933849178254604, + 0.008516138419508934, + 0.010789395309984684, + 0.003863260382786393, + 0.0039042325224727392, + 0.009846705943346024, + 0.006642854772508144, + 0.002430412918329239, + 0.003677411936223507, + 0.0039975689724087715, + 0.006238633766770363, + 0.004749867133796215, + 0.005432703997939825, + 0.007870561443269253, + 0.008388039655983448, + 0.004613554570823908, + 0.006746230181306601, + 0.00816553644835949, + 0.007282261271029711, + 0.008341747336089611, + 0.00782102718949318, + 0.008696679957211018, + 0.00981973111629486, + 0.005996143911033869, + 0.00786875281482935, + 0.015843000262975693, + 0.015213178470730782, + 0.009960661642253399, + 0.010307183489203453, + 0.011907972395420074, + 0.010773995891213417, + 0.009210219606757164, + 0.01818186603486538, + 0.018466642126441002, + 0.025024911388754845, + 0.05263862386345863, + 0.054883912205696106, + 0.022051243111491203, + 0.04880622774362564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.06745936721563339 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991933226585388 + ], + "opposite_logprob_tokens": [ + -1.8049193620681763 + ], + "logprob_reward": -0.17991933226585388, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991933226585388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06572819501161575, + 0.030883336439728737, + 0.35515424609184265, + 0.06790052354335785, + 0.09435280412435532, + 0.019245849922299385, + 0.046017128974199295, + 0.03337524086236954, + 0.22137586772441864, + 0.06596677005290985 + ], + "credit_map": [ + 0.0041586835868656635, + 0.006999044679105282, + 0.009612922556698322, + 0.012196598574519157, + 0.006852100603282452, + 0.007002468220889568, + 0.011466184630990028, + 0.0074401916936039925, + 0.002389253117144108, + 0.0026037516072392464, + 0.00517856702208519, + 0.007654036860913038, + 0.0024146533105522394, + 0.0025300702545791864, + 0.003806001041084528, + 0.0043070027604699135, + 0.018280787393450737, + 0.012581810355186462, + 0.0471685416996479, + 0.06840311735868454, + 0.02929142862558365, + 0.03409937769174576, + 0.09795565158128738, + 0.04737352207303047, + 0.0032347741071134806, + 0.006630426738411188, + 0.01240028440952301, + 0.010139149613678455, + 0.004466121084988117, + 0.0055272458121180534, + 0.018753746524453163, + 0.006748780142515898, + 0.007268029730767012, + 0.009798667393624783, + 0.017686033621430397, + 0.013713435269892216, + 0.007732014171779156, + 0.008284726180136204, + 0.01881176233291626, + 0.01105814054608345, + 0.0012534069828689098, + 0.0021572988480329514, + 0.003504182444885373, + 0.0025290653575211763, + 0.001882995362393558, + 0.0018419409170746803, + 0.003191468771547079, + 0.0028854908887296915, + 0.0022583010140806437, + 0.0028402863536030054, + 0.0044447025284171104, + 0.0066889552399516106, + 0.00611390545964241, + 0.006545980926603079, + 0.008165811188519001, + 0.008959184400737286, + 0.0017036383505910635, + 0.003179442137479782, + 0.004478226415812969, + 0.005104312673211098, + 0.003108401084318757, + 0.0043128477409482, + 0.006542588118463755, + 0.004945786204189062, + 0.01902106963098049, + 0.03194087743759155, + 0.03257424756884575, + 0.03900334984064102, + 0.023157043382525444, + 0.02569846250116825, + 0.03036154806613922, + 0.019619261845946312, + 0.0036038493271917105, + 0.005225950386375189, + 0.005754123907536268, + 0.006872340105473995, + 0.01715896837413311, + 0.016902804374694824, + 0.004337377846240997, + 0.006111350376158953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.17991933226585388 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865920543670654 + ], + "opposite_logprob_tokens": [ + -3.048659324645996 + ], + "logprob_reward": -0.04865920543670654, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865920543670654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16151943802833557, + 0.11431761085987091, + 0.2414083182811737, + 0.2302445024251938, + 0.04679564759135246, + 0.037219833582639694, + 0.02604341134428978, + 0.021724220365285873, + 0.06140789017081261, + 0.0593191459774971 + ], + "credit_map": [ + 0.015079958364367485, + 0.023542294278740883, + 0.017953546717762947, + 0.018327316269278526, + 0.016982894390821457, + 0.019120683893561363, + 0.020118348300457, + 0.03039439208805561, + 0.009003428742289543, + 0.014154075644910336, + 0.010600650683045387, + 0.02772056870162487, + 0.007877594791352749, + 0.010411201044917107, + 0.02274002879858017, + 0.011810061521828175, + 0.009629826992750168, + 0.009257011115550995, + 0.03329388052225113, + 0.060002222657203674, + 0.017436586320400238, + 0.02106381580233574, + 0.057748060673475266, + 0.03297693654894829, + 0.008813044056296349, + 0.017821870744228363, + 0.02819260209798813, + 0.04168059676885605, + 0.03487744554877281, + 0.0362359844148159, + 0.04423286020755768, + 0.018390078097581863, + 0.002787472680211067, + 0.006861401721835136, + 0.006222206633538008, + 0.007395809516310692, + 0.005517225246876478, + 0.005823525134474039, + 0.006231514271348715, + 0.005956493318080902, + 0.003004029393196106, + 0.006766610778868198, + 0.004966390319168568, + 0.004825014155358076, + 0.005270406138151884, + 0.005699467845261097, + 0.0037273287307471037, + 0.002960587153211236, + 0.00342182326130569, + 0.006163166835904121, + 0.0018150004325434566, + 0.0019068963592872024, + 0.004147429950535297, + 0.00467243418097496, + 0.0018255558097735047, + 0.0020911034662276506, + 0.0019107144325971603, + 0.0040855989791452885, + 0.0022444629576057196, + 0.00318092480301857, + 0.002274115337058902, + 0.0026573818176984787, + 0.00261656753718853, + 0.002754455665126443, + 0.0034166809637099504, + 0.0045652203261852264, + 0.007375003769993782, + 0.01046714000403881, + 0.005802694242447615, + 0.00614940095692873, + 0.0077771274372935295, + 0.015854621306061745, + 0.0026133577339351177, + 0.0050081307999789715, + 0.005364884622395039, + 0.005568698979914188, + 0.01316182129085064, + 0.013818801380693913, + 0.005212330259382725, + 0.008571119979023933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.04865920543670654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034221649169922 + ], + "opposite_logprob_tokens": [ + -1.9103422164916992 + ], + "logprob_reward": -0.16034221649169922, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034221649169922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11396727710962296, + 0.050095804035663605, + 0.27633941173553467, + 0.0767865851521492, + 0.1643269956111908, + 0.061944082379341125, + 0.04212845489382744, + 0.0326695553958416, + 0.10482263565063477, + 0.07691923528909683 + ], + "credit_map": [ + 0.0071657500229775906, + 0.010171118192374706, + 0.019241496920585632, + 0.02618197165429592, + 0.009004858322441578, + 0.010666713118553162, + 0.01468750461935997, + 0.01684786193072796, + 0.004433461930602789, + 0.005532529670745134, + 0.008004900068044662, + 0.005689585581421852, + 0.004811237100511789, + 0.0052796718664467335, + 0.009704462252557278, + 0.006639950443059206, + 0.01699787750840187, + 0.018871905282139778, + 0.03612431511282921, + 0.0353296734392643, + 0.030606746673583984, + 0.033722907304763794, + 0.07145516574382782, + 0.033230800181627274, + 0.004738493822515011, + 0.006718458607792854, + 0.006099799182265997, + 0.00855766050517559, + 0.018542766571044922, + 0.018340203911066055, + 0.0077178641222417355, + 0.006071334704756737, + 0.01180277019739151, + 0.016696114093065262, + 0.029280174523591995, + 0.024648517370224, + 0.013105794787406921, + 0.014521092176437378, + 0.03457963839173317, + 0.01969289593398571, + 0.0025861479807645082, + 0.006143100559711456, + 0.008500374853610992, + 0.012836812064051628, + 0.004439425189048052, + 0.005301198922097683, + 0.014641116373240948, + 0.007495907600969076, + 0.002415999537333846, + 0.0031816011760383844, + 0.005151272285729647, + 0.00603800592944026, + 0.0031546165701001883, + 0.004893655888736248, + 0.008016123436391354, + 0.009277179837226868, + 0.002708975225687027, + 0.0036758757196366787, + 0.005063303746283054, + 0.004192212596535683, + 0.0029958325903862715, + 0.003129473188892007, + 0.004551396239548922, + 0.006352485157549381, + 0.006809608079493046, + 0.011812317185103893, + 0.0160187017172575, + 0.025059029459953308, + 0.007697731722146273, + 0.009407046250998974, + 0.012400245293974876, + 0.015617954544723034, + 0.00334321241825819, + 0.004603946581482887, + 0.006295128725469112, + 0.008771039545536041, + 0.020823368802666664, + 0.02140093967318535, + 0.005256774369627237, + 0.006424823310226202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.16034221649169922 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034230589866638 + ], + "opposite_logprob_tokens": [ + -1.9103423357009888 + ], + "logprob_reward": -0.16034230589866638, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034230589866638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09220051020383835, + 0.04045521095395088, + 0.34670907258987427, + 0.05992233008146286, + 0.09254761785268784, + 0.06205427646636963, + 0.05993085727095604, + 0.017648309469223022, + 0.13867872953414917, + 0.08985308557748795 + ], + "credit_map": [ + 0.007878991775214672, + 0.012671537697315216, + 0.010687493719160557, + 0.013657893985509872, + 0.013185429386794567, + 0.013692939653992653, + 0.01088324561715126, + 0.009542981162667274, + 0.0029912253376096487, + 0.00404025474563241, + 0.006566962227225304, + 0.0059327129274606705, + 0.005294991657137871, + 0.004732060246169567, + 0.005142674781382084, + 0.0057543301954865456, + 0.019436275586485863, + 0.013851424679160118, + 0.046450175344944, + 0.05025050416588783, + 0.03592238202691078, + 0.03768983483314514, + 0.09817730635404587, + 0.044931165874004364, + 0.003788562258705497, + 0.008194888941943645, + 0.007490704767405987, + 0.006786920130252838, + 0.008622630499303341, + 0.00849614106118679, + 0.011694284155964851, + 0.004848201293498278, + 0.005136352032423019, + 0.010239635594189167, + 0.013902815990149975, + 0.014855235815048218, + 0.007806144654750824, + 0.00856240838766098, + 0.01918942481279373, + 0.012855601496994495, + 0.004080192185938358, + 0.006692038383334875, + 0.012277280911803246, + 0.009449386969208717, + 0.005840925965458155, + 0.005423390306532383, + 0.009689923375844955, + 0.008601134642958641, + 0.0035908890422433615, + 0.004370949696749449, + 0.006769819185137749, + 0.008830538019537926, + 0.004796061664819717, + 0.007039931602776051, + 0.011203296482563019, + 0.013329370878636837, + 0.0010310254292562604, + 0.001377907465212047, + 0.002420116448774934, + 0.003960604313760996, + 0.0011925582075491548, + 0.0014827385311946273, + 0.003347632009536028, + 0.002835726598277688, + 0.010338328778743744, + 0.014689045958220959, + 0.02075701206922531, + 0.02556762658059597, + 0.016316648572683334, + 0.017008543014526367, + 0.01918093115091324, + 0.014820593409240246, + 0.006273005157709122, + 0.005806926172226667, + 0.010481176897883415, + 0.010484166443347931, + 0.019833525642752647, + 0.019145645201206207, + 0.007281900849193335, + 0.010546737350523472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.16034230589866638 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle drone shot smoothly glides from left to right, clinging to the tops of a vast, golden wheat field. Under the low, evening sun, the field is tinted a warm honey hue. The sky is a deep blue, dotted with a few puffy white clouds whipped by the wind. Each ripe stalk bears a heavy load of plump ears, their delicate awns gleaming in the light. Suddenly, a powerful, invisible gust of wind sweeps through, forcing the stalks to bend uniformly, forming a spectacular, undulating wave. This golden wave of wheat rapidly spreads from one side of the camera toward the distant horizon, one layer chasing another, creating a flowing, rhythmic visual spectacle across the vast field. Light dances and flickers on the swaying ears of wheat, contrasting sharply with the shifting shadows cast by the waves. The entire image is imbued with dynamic vitality and a majestic yet tranquil natural beauty.", + "seed": 7439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_12/index_8605_seed_7439_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06678729206323623, + "vlm_logprob": -0.06678729206323623, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.06678729206323623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05890877917408943 + ], + "opposite_logprob_tokens": [ + -4.183908939361572 + ], + "logprob_reward": -0.05890877917408943, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.05890877917408943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09213937819004059, + 0.05779145285487175, + 0.4359435737133026, + 0.019414614886045456, + 0.0389944352209568, + 0.03366267308592796, + 0.04763554036617279, + 0.02833804115653038, + 0.0743989422917366, + 0.17168138921260834 + ], + "credit_map": [ + 0.006553334649652243, + 0.004201648291200399, + 0.023007629439234734, + 0.0164048969745636, + 0.005901177879422903, + 0.009129801765084267, + 0.014752990566194057, + 0.012187891639769077, + 0.00376747059635818, + 0.0035741266328841448, + 0.00874447263777256, + 0.010099749080836773, + 0.005267022177577019, + 0.009770204313099384, + 0.008717185817658901, + 0.00785121414810419, + 0.01041780412197113, + 0.009545831941068172, + 0.14415377378463745, + 0.1910010576248169, + 0.006383721716701984, + 0.01780383102595806, + 0.02707909606397152, + 0.02955840900540352, + 0.0009495231206528842, + 0.0009650603751651943, + 0.003427553689107299, + 0.004770229570567608, + 0.0011061903787776828, + 0.001078720553778112, + 0.0038905804976820946, + 0.003226756351068616, + 0.0032060761004686356, + 0.0029497789219021797, + 0.00734576303511858, + 0.005286219995468855, + 0.002970930188894272, + 0.006537546869367361, + 0.0055318172089755535, + 0.005166298244148493, + 0.0024942494928836823, + 0.002433961722999811, + 0.0068161580711603165, + 0.006478582043200731, + 0.002250096295028925, + 0.003006520913913846, + 0.005917901173233986, + 0.0042651984840631485, + 0.0024600084871053696, + 0.0022592684254050255, + 0.013625772669911385, + 0.0069260732270777225, + 0.002332685748115182, + 0.002573031932115555, + 0.006785995792597532, + 0.01067269966006279, + 0.0018250789726153016, + 0.002675022231414914, + 0.0028583852108567953, + 0.0036089341156184673, + 0.0021492464002221823, + 0.005302921403199434, + 0.0040008677169680595, + 0.00591758219525218, + 0.0037701555993407965, + 0.0025767628103494644, + 0.016299638897180557, + 0.011120649054646492, + 0.00413178838789463, + 0.0066336216405034065, + 0.0152397146448493, + 0.01462660264223814, + 0.006204572040587664, + 0.00508497329428792, + 0.06343136727809906, + 0.046784136444330215, + 0.0075317430309951305, + 0.01915091462433338, + 0.011215893551707268, + 0.012277775444090366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.05890877917408943 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032940655946732 + ], + "opposite_logprob_tokens": [ + -2.3503293991088867 + ], + "logprob_reward": -0.10032940655946732, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032940655946732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05746740102767944, + 0.01583886705338955, + 0.8170194625854492, + 0.007161205168813467, + 0.015407351776957512, + 0.020691191777586937, + 0.004856750834733248, + 0.02774827554821968, + 0.024091750383377075, + 0.009717708453536034 + ], + "credit_map": [ + 0.002581012202426791, + 0.0030241806525737047, + 0.012160557322204113, + 0.012538007460534573, + 0.0024114330299198627, + 0.0029219903517514467, + 0.01219655480235815, + 0.009633672423660755, + 0.0013731535291299224, + 0.0010275938548147678, + 0.0033426720183342695, + 0.0034331525675952435, + 0.0011781982611864805, + 0.002816783031448722, + 0.0013911394635215402, + 0.001276175258681178, + 0.018497731536626816, + 0.01774856075644493, + 0.27506425976753235, + 0.3633415102958679, + 0.009258878417313099, + 0.026031244546175003, + 0.051231347024440765, + 0.055846020579338074, + 0.00040857004933059216, + 0.0004999084048904479, + 0.0014781856443732977, + 0.000996846123598516, + 0.0004284651076886803, + 0.0005390686565078795, + 0.0016445688670501113, + 0.0011655930429697037, + 0.0012367651797831059, + 0.0010560019873082638, + 0.0023150485940277576, + 0.002381902188062668, + 0.0012433165684342384, + 0.0023317495360970497, + 0.002406015992164612, + 0.0024365519639104605, + 0.00093045987887308, + 0.0006250220467336476, + 0.006046375259757042, + 0.0026940081734210253, + 0.0006948545342311263, + 0.0008150697685778141, + 0.0022324894089251757, + 0.0066529144532978535, + 0.00036822707625105977, + 0.0003784791042562574, + 0.000668393331579864, + 0.0008116978569887578, + 0.0002800929651129991, + 0.000432845699833706, + 0.0009869963396340609, + 0.0009300188976339996, + 0.0016569335712119937, + 0.0016000635223463178, + 0.0033077045809477568, + 0.0021431331988424063, + 0.0024189050309360027, + 0.006173927336931229, + 0.0025346700567752123, + 0.007912939414381981, + 0.001182857435196638, + 0.0008044276619330049, + 0.0038367609959095716, + 0.004330428317189217, + 0.0014391025761142373, + 0.001206951099447906, + 0.005351483356207609, + 0.005939737427979708, + 0.0003701488603837788, + 0.0003561530029401183, + 0.0036651783157140017, + 0.002579194726422429, + 0.0004736337868962437, + 0.0006883731693960726, + 0.000765712873544544, + 0.0008193149697035551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.10032940655946732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370501473546028 + ], + "opposite_logprob_tokens": [ + -3.408704996109009 + ], + "logprob_reward": -0.03370501473546028, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370501473546028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0651867613196373, + 0.058039341121912, + 0.5441580414772034, + 0.028088664636015892, + 0.0334695503115654, + 0.010792792774736881, + 0.009099303744733334, + 0.2184308022260666, + 0.023623008280992508, + 0.009111828170716763 + ], + "credit_map": [ + 0.005484535824507475, + 0.005217546597123146, + 0.00910200085490942, + 0.0115188704803586, + 0.009492220357060432, + 0.007375001907348633, + 0.009621743112802505, + 0.007374837063252926, + 0.007026210892945528, + 0.003915197215974331, + 0.006052556913346052, + 0.00630336394533515, + 0.007055952679365873, + 0.01706949807703495, + 0.005418804474174976, + 0.005197750870138407, + 0.012234040535986423, + 0.011963996104896069, + 0.18625812232494354, + 0.24175308644771576, + 0.006222720257937908, + 0.016025710850954056, + 0.03340570256114006, + 0.03629456087946892, + 0.0020332790445536375, + 0.0018310091691091657, + 0.004101500380784273, + 0.006768573075532913, + 0.0022625289857387543, + 0.00224619684740901, + 0.004548531956970692, + 0.00429704412817955, + 0.002749055391177535, + 0.002466058824211359, + 0.007024744059890509, + 0.0029767723754048347, + 0.003538736840710044, + 0.008217373862862587, + 0.004326419904828072, + 0.0021703869570046663, + 0.000577349856030196, + 0.0007431631674990058, + 0.002570041222497821, + 0.0014522266574203968, + 0.0005882173427380621, + 0.0010562291136011481, + 0.0014062990667298436, + 0.002399265067651868, + 0.0004458119219634682, + 0.000627198547590524, + 0.0021881370339542627, + 0.0014163041487336159, + 0.00044388885726220906, + 0.0008115234086290002, + 0.0011871801689267159, + 0.0019792583771049976, + 0.01350094098597765, + 0.012713688425719738, + 0.025047967210412025, + 0.015881676226854324, + 0.01933806762099266, + 0.05027582868933678, + 0.020403651520609856, + 0.06126897782087326, + 0.001250475412234664, + 0.0011712644482031465, + 0.0032504629343748093, + 0.004488797392696142, + 0.0014487016014754772, + 0.0016656697262078524, + 0.004854890983551741, + 0.005492744501680136, + 0.0002878842642530799, + 0.00036731737782247365, + 0.003191710449755192, + 0.003488842397928238, + 0.0002699221658986062, + 0.00040640434599481523, + 0.0005189375369809568, + 0.0005808094283565879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.03370501473546028 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896267622709274 + ], + "opposite_logprob_tokens": [ + -2.578962564468384 + ], + "logprob_reward": -0.07896267622709274, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896267622709274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11719975620508194, + 0.06720128655433655, + 0.24695409834384918, + 0.01169588416814804, + 0.04182147607207298, + 0.016325775533914566, + 0.04472442716360092, + 0.2018066644668579, + 0.20770348608493805, + 0.04456711933016777 + ], + "credit_map": [ + 0.006635434925556183, + 0.005994666833430529, + 0.021620016545057297, + 0.021268069744110107, + 0.007474500685930252, + 0.009895846247673035, + 0.02397104725241661, + 0.020340170711278915, + 0.003376439679414034, + 0.002166931051760912, + 0.019376281648874283, + 0.020767858251929283, + 0.0024327996652573347, + 0.0037958319298923016, + 0.008224391378462315, + 0.007060753181576729, + 0.006848734337836504, + 0.0064816405065357685, + 0.07177957892417908, + 0.10215996205806732, + 0.007111506536602974, + 0.018007714301347733, + 0.016481786966323853, + 0.018083183094859123, + 0.000869154348038137, + 0.0009668775019235909, + 0.00190719251986593, + 0.002250872552394867, + 0.0008730216068215668, + 0.000859053514432162, + 0.0019635118078440428, + 0.0020062001422047615, + 0.002668409375473857, + 0.0020284419879317284, + 0.00778917083516717, + 0.006379515398293734, + 0.002371651353314519, + 0.004803029354661703, + 0.00622689351439476, + 0.00955436285585165, + 0.0009087283979170024, + 0.0010656631784513593, + 0.0021999888122081757, + 0.003064373042434454, + 0.0009745659190230072, + 0.0017100977711379528, + 0.003371184691786766, + 0.0030311732552945614, + 0.002458050148561597, + 0.0016637597000226378, + 0.012723327614367008, + 0.005268888548016548, + 0.0018227760447189212, + 0.003265464212745428, + 0.005080991890281439, + 0.012441166676580906, + 0.012683828361332417, + 0.012388871051371098, + 0.02348281629383564, + 0.014309080317616463, + 0.017549319192767143, + 0.046751946210861206, + 0.0183266494423151, + 0.0563141368329525, + 0.010252212174236774, + 0.007170164957642555, + 0.053585488349199295, + 0.03126253932714462, + 0.010740949772298336, + 0.019668234512209892, + 0.04234859719872475, + 0.032675303518772125, + 0.0012232594890519977, + 0.0010086012771353126, + 0.018623068928718567, + 0.015740063041448593, + 0.0006756444345228374, + 0.0009807080496102571, + 0.0031882734037935734, + 0.0031274983193725348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.07896267622709274 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly glide from left to right over a vast golden wheat field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a low evening sun that tints the field a warm honey hue against a deep blue sky with white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend uniformly to form a spectacular undulating wave that spreads across the field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible light dancing and flickering on the swaying ears of wheat contrasting with shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203058362007141 + ], + "opposite_logprob_tokens": [ + -2.812030553817749 + ], + "logprob_reward": -0.06203058362007141, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203058362007141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26973235607147217, + 0.18344537913799286, + 0.1710427701473236, + 0.11070968210697174, + 0.02955503761768341, + 0.027030568569898605, + 0.023473192006349564, + 0.05717206746339798, + 0.059686727821826935, + 0.06815227121114731 + ], + "credit_map": [ + 0.02424161694943905, + 0.014434751123189926, + 0.03743511065840721, + 0.04158994182944298, + 0.03283366188406944, + 0.043633345514535904, + 0.03969639912247658, + 0.03586752712726593, + 0.0210108682513237, + 0.010520011186599731, + 0.027418026700615883, + 0.027420280501246452, + 0.020896518602967262, + 0.0518258735537529, + 0.011373685672879219, + 0.012980098836123943, + 0.008321183733642101, + 0.008664436638355255, + 0.021790821105241776, + 0.052881281822919846, + 0.013103225268423557, + 0.03053171932697296, + 0.018086930736899376, + 0.01766318269073963, + 0.004755878821015358, + 0.004466382320970297, + 0.020192690193653107, + 0.027537597343325615, + 0.006479326635599136, + 0.0059648482128977776, + 0.021303515881299973, + 0.020009446889162064, + 0.0019817573484033346, + 0.0015788296004757285, + 0.004921054933220148, + 0.0028374409303069115, + 0.0024291565641760826, + 0.005315352231264114, + 0.003910799045115709, + 0.006580646149814129, + 0.0016290071653202176, + 0.0015732721658423543, + 0.005153212696313858, + 0.004491385072469711, + 0.001708473777398467, + 0.002119536744430661, + 0.005946955177932978, + 0.0044087255373597145, + 0.0014507505111396313, + 0.001213732291944325, + 0.005133636295795441, + 0.003646365599706769, + 0.0012629131088033319, + 0.0018354534404352307, + 0.003524376545101404, + 0.00540596479550004, + 0.004173322580754757, + 0.00475932564586401, + 0.007496559526771307, + 0.00823940522968769, + 0.004128281958401203, + 0.009359614923596382, + 0.008029834367334843, + 0.010985726490616798, + 0.0035066399723291397, + 0.0028218324296176434, + 0.0181219931691885, + 0.0067983511835336685, + 0.004182400181889534, + 0.01028483733534813, + 0.010244419798254967, + 0.003726255614310503, + 0.0018578707240521908, + 0.0019078474724665284, + 0.02536580339074135, + 0.027345309033989906, + 0.0015520600136369467, + 0.001635782653465867, + 0.004520708229392767, + 0.0039668926037848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06203058362007141 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_tensor([7314, 7314])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_tensor([7314, 7314])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e7c4c67891e3099861f9b7c3c0f920835a7d4108 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_tensor([7314, 7314])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth drone tracking shot, shot from a slightly upward rear-view angle, closely follows a classic white Mitsubishi Lancer Evolution VI as it speeds along a winding coastal road. In the golden hour of evening, the warm sunlight casts a mirror-like glow on the car's paintwork, reflecting the pale orange sunset glow in the sky. The massive carbon fiber wing towering high above the rear, perfectly poised in the airflow, becomes the frame's most striking focal point. As the car enters a sharp S-curve with a nimble and graceful gesture, the camera pans steadily, capturing the car's slight tilt under centrifugal force. The red brake calipers on the bronze multi-spoke racing wheels flash by, and the rear end slides slightly outward in a perfectly controlled arc, its four tires gripping the dark gray asphalt like a virtuoso ballet dancer executing a precise and graceful spin. Then, the car quickly rights itself, unleashing its power without hesitation, sprinting towards the next corner with a deep roar from the exhaust. On one side of the road is a steep rocky cliff, and on the other side is a vast blue ocean with white waves. The whole picture is full of dynamic beauty and driving passion under the warm light and shadow, creating a free, focused and highly cinematic atmosphere.", + "seed": 24800, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_7314_seed_24800_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06421499531716109, + "vlm_logprob": -0.06421499531716109, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.06421499531716109, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1588379591703415 + ], + "opposite_logprob_tokens": [ + -2.1588380336761475 + ], + "logprob_reward": -0.1588379591703415, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1588379591703415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10209960490465164, + 0.07944297045469284, + 0.11128273606300354, + 0.0986383855342865, + 0.054107435047626495, + 0.048015158623456955, + 0.15202216804027557, + 0.20987802743911743, + 0.0622226744890213, + 0.08229091018438339 + ], + "credit_map": [ + 0.008522570133209229, + 0.00872508529573679, + 0.011726245284080505, + 0.017519349232316017, + 0.010609501041471958, + 0.012685148045420647, + 0.02033538743853569, + 0.011976306326687336, + 0.008864440023899078, + 0.007592101115733385, + 0.010021689347922802, + 0.011229317635297775, + 0.008259383030235767, + 0.009089119732379913, + 0.01475038006901741, + 0.009636534377932549, + 0.011021479964256287, + 0.01251447293907404, + 0.015504349023103714, + 0.017912212759256363, + 0.014444601722061634, + 0.010808639228343964, + 0.01016019657254219, + 0.018916770815849304, + 0.009981061331927776, + 0.009693768806755543, + 0.021440837532281876, + 0.008026795461773872, + 0.0048687090165913105, + 0.00784829631447792, + 0.008157947100698948, + 0.02862095646560192, + 0.0051223826594650745, + 0.007754524238407612, + 0.006902230903506279, + 0.006840577349066734, + 0.005873195361346006, + 0.010863540694117546, + 0.0037163023371249437, + 0.00703467708081007, + 0.007763925939798355, + 0.007174426224082708, + 0.007833077572286129, + 0.005648897495120764, + 0.0048645539209246635, + 0.006293460261076689, + 0.00406990572810173, + 0.0043669044971466064, + 0.025365013629198074, + 0.028000952675938606, + 0.02865184284746647, + 0.01638670265674591, + 0.012654198333621025, + 0.022412605583667755, + 0.009347131475806236, + 0.009203696623444557, + 0.03752515837550163, + 0.02982565574347973, + 0.024591391906142235, + 0.02208009734749794, + 0.014097221195697784, + 0.0227822232991457, + 0.018416162580251694, + 0.040560100227594376, + 0.007930319756269455, + 0.008759595453739166, + 0.008666874840855598, + 0.008022630587220192, + 0.00820065289735794, + 0.009212258271872997, + 0.0058461339212954044, + 0.005584208760410547, + 0.011315260082483292, + 0.00577201321721077, + 0.00980462972074747, + 0.008427591063082218, + 0.015241662040352821, + 0.013327774591743946, + 0.00746237114071846, + 0.010939599946141243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.1588379591703415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813379257917404 + ], + "opposite_logprob_tokens": [ + -3.2881338596343994 + ], + "logprob_reward": -0.03813379257917404, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813379257917404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06609296053647995, + 0.0684947520494461, + 0.0434323251247406, + 0.3259415030479431, + 0.05546865984797478, + 0.030947556719183922, + 0.05977049842476845, + 0.1990765631198883, + 0.07264786213636398, + 0.07812732458114624 + ], + "credit_map": [ + 0.005212086718529463, + 0.004096842836588621, + 0.0077045452781021595, + 0.011611453257501125, + 0.0043233479373157024, + 0.013872450217604637, + 0.011285369284451008, + 0.007986864075064659, + 0.007257718127220869, + 0.006466301623731852, + 0.00944199413061142, + 0.011738013476133347, + 0.007157521788030863, + 0.011214891448616982, + 0.006334342993795872, + 0.008883963339030743, + 0.004071193281561136, + 0.0056798723526299, + 0.006449304986745119, + 0.006940288469195366, + 0.005059299990534782, + 0.005230277311056852, + 0.004461037926375866, + 0.00554104708135128, + 0.03137071803212166, + 0.01929893158376217, + 0.059204328805208206, + 0.014673345722258091, + 0.01530049555003643, + 0.04404260590672493, + 0.02917850948870182, + 0.11287254095077515, + 0.005785778164863586, + 0.008534918539226055, + 0.007992144674062729, + 0.006475047208368778, + 0.008431159891188145, + 0.007938397116959095, + 0.004526698496192694, + 0.005784508306533098, + 0.004004038870334625, + 0.004927740897983313, + 0.005046419333666563, + 0.0033738729543983936, + 0.0033126799389719963, + 0.004494091495871544, + 0.0021313924808055162, + 0.003657318651676178, + 0.009318864904344082, + 0.012798094190657139, + 0.009456206113100052, + 0.006410590838640928, + 0.005415031220763922, + 0.007605590391904116, + 0.0052103158086538315, + 0.003555802395567298, + 0.034291964024305344, + 0.026746705174446106, + 0.024066904559731483, + 0.022053347900509834, + 0.013566178269684315, + 0.02339942939579487, + 0.016757071018218994, + 0.03819495812058449, + 0.01793719455599785, + 0.008656618185341358, + 0.008907845243811607, + 0.0068557532504200935, + 0.006452135741710663, + 0.00869386363774538, + 0.006163406651467085, + 0.008981042541563511, + 0.015022494830191135, + 0.0073916492983698845, + 0.010814298875629902, + 0.010067191906273365, + 0.010346690192818642, + 0.007593592628836632, + 0.007725363131612539, + 0.009166030213236809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.03813379257917404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905091136693954 + ], + "opposite_logprob_tokens": [ + -2.4640510082244873 + ], + "logprob_reward": -0.08905091136693954, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905091136693954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11658401787281036, + 0.10960248857736588, + 0.035704612731933594, + 0.026900140568614006, + 0.036758020520210266, + 0.02380991168320179, + 0.15135805308818817, + 0.15322978794574738, + 0.032055772840976715, + 0.31399717926979065 + ], + "credit_map": [ + 0.009470378048717976, + 0.0065226624719798565, + 0.012608947232365608, + 0.03399400785565376, + 0.004537330474704504, + 0.010775706730782986, + 0.01883922889828682, + 0.019835758954286575, + 0.010060775093734264, + 0.011217218823730946, + 0.017708055675029755, + 0.019730979576706886, + 0.010762370191514492, + 0.010281020775437355, + 0.01711280271410942, + 0.012729269452393055, + 0.005000613629817963, + 0.00488899415358901, + 0.0037997490726411343, + 0.0045384555123746395, + 0.004831802099943161, + 0.004169329069554806, + 0.0030455803498625755, + 0.005430091172456741, + 0.0031552298460155725, + 0.005260825157165527, + 0.004711545538157225, + 0.0025446568615734577, + 0.0026902365498244762, + 0.003017221810296178, + 0.0019718033727258444, + 0.0035486225970089436, + 0.0036741618532687426, + 0.003804401494562626, + 0.004998453892767429, + 0.005233849864453077, + 0.003855133196339011, + 0.007650631945580244, + 0.0026821426581591368, + 0.004859250504523516, + 0.0025878252927213907, + 0.004158003255724907, + 0.0036443681456148624, + 0.0032274958211928606, + 0.002141423989087343, + 0.0046634008176624775, + 0.0015310606686398387, + 0.0018563371850177646, + 0.025559771806001663, + 0.029213914647698402, + 0.029217947274446487, + 0.014730311930179596, + 0.012682232074439526, + 0.023469986394047737, + 0.007966064848005772, + 0.008517829701304436, + 0.027629034593701363, + 0.020386066287755966, + 0.01893521472811699, + 0.01796288788318634, + 0.009720176458358765, + 0.016030676662921906, + 0.013686818070709705, + 0.028878914192318916, + 0.004839968401938677, + 0.004787428770214319, + 0.005063682794570923, + 0.003825928783044219, + 0.004452434368431568, + 0.004613199271261692, + 0.0022801838349550962, + 0.0021929496433585882, + 0.04422629252076149, + 0.0238560289144516, + 0.061272528022527695, + 0.03555874899029732, + 0.03752024844288826, + 0.040988460183143616, + 0.030495595186948776, + 0.04007929190993309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.08905091136693954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799941927194595 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799941927194595, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799941927194595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0681421086192131, + 0.07982128113508224, + 0.09768927842378616, + 0.05821308121085167, + 0.08227509260177612, + 0.026179276406764984, + 0.07662458717823029, + 0.19638054072856903, + 0.05836253613233566, + 0.25631219148635864 + ], + "credit_map": [ + 0.00612680334597826, + 0.00585609395056963, + 0.009739844128489494, + 0.012900463305413723, + 0.005921691190451384, + 0.006947807967662811, + 0.012841891497373581, + 0.007807516492903233, + 0.007680022157728672, + 0.0074569787830114365, + 0.013229947537183762, + 0.014767411164939404, + 0.0076095410622656345, + 0.00999513640999794, + 0.009899486787617207, + 0.009182760491967201, + 0.011582434177398682, + 0.013517869636416435, + 0.010086472146213055, + 0.019136451184749603, + 0.011237728409469128, + 0.009986341930925846, + 0.008757881820201874, + 0.013384096324443817, + 0.005936143919825554, + 0.005488707218319178, + 0.008841320872306824, + 0.00408483250066638, + 0.002966322936117649, + 0.008982622995972633, + 0.004591402597725391, + 0.01732172816991806, + 0.009817833080887794, + 0.01215254794806242, + 0.0082548838108778, + 0.006936763878911734, + 0.010927018709480762, + 0.013962479308247566, + 0.006164728198200464, + 0.014058837667107582, + 0.003092970931902528, + 0.003778541926294565, + 0.004210059065371752, + 0.0035645824391394854, + 0.0028983009979128838, + 0.003462479216977954, + 0.002460357267409563, + 0.002711980603635311, + 0.011901105754077435, + 0.013600319623947144, + 0.013258327730000019, + 0.008995347656309605, + 0.007138184737414122, + 0.010638152249157429, + 0.00617970759049058, + 0.004913446959108114, + 0.0363469161093235, + 0.028039343655109406, + 0.022633664309978485, + 0.01863294281065464, + 0.013738074339926243, + 0.022100908681750298, + 0.017262345179915428, + 0.037626348435878754, + 0.008160211145877838, + 0.009275298565626144, + 0.006758063565939665, + 0.007579014170914888, + 0.007617870811372995, + 0.009592832066118717, + 0.004420742858201265, + 0.004958501551300287, + 0.040799736976623535, + 0.0213739275932312, + 0.04125015810132027, + 0.03153334930539131, + 0.03047013096511364, + 0.02907434105873108, + 0.028267156332731247, + 0.03354337066411972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.029799941927194595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252371542155743 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252371542155743, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252371542155743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24591530859470367, + 0.102571502327919, + 0.0971260592341423, + 0.07044053077697754, + 0.07519014179706573, + 0.02725551836192608, + 0.16260305047035217, + 0.062291715294122696, + 0.044666703790426254, + 0.11193942278623581 + ], + "credit_map": [ + 0.014341933652758598, + 0.012223375961184502, + 0.02872128039598465, + 0.08113246411085129, + 0.010535209439694881, + 0.01688975840806961, + 0.04148728772997856, + 0.040584027767181396, + 0.012569338083267212, + 0.011244479566812515, + 0.014997589401900768, + 0.015525948256254196, + 0.00937544833868742, + 0.01440081000328064, + 0.011745321564376354, + 0.0127125708386302, + 0.011566202156245708, + 0.00998984556645155, + 0.011564194224774837, + 0.018339060246944427, + 0.013844850473105907, + 0.007245523855090141, + 0.00988351833075285, + 0.014692872762680054, + 0.007662854623049498, + 0.015200208872556686, + 0.010251199826598167, + 0.006498328410089016, + 0.006810873281210661, + 0.01363440789282322, + 0.003351800376549363, + 0.007030863780528307, + 0.007670250721275806, + 0.009372294880449772, + 0.007548968307673931, + 0.006504174787551165, + 0.010689979419112206, + 0.012326590716838837, + 0.006933807861059904, + 0.014144087210297585, + 0.0033757048659026623, + 0.004577639512717724, + 0.0034628859721124172, + 0.0032403136137872934, + 0.0036545300390571356, + 0.0036309021525084972, + 0.0023619425483047962, + 0.002951601054519415, + 0.02980002760887146, + 0.021747998893260956, + 0.040191009640693665, + 0.01300827320665121, + 0.010932063683867455, + 0.02653512917459011, + 0.008912342600524426, + 0.011476225219666958, + 0.007527396082878113, + 0.006564250681549311, + 0.006773820146918297, + 0.008464422076940536, + 0.005149258766323328, + 0.008475396782159805, + 0.00730014918372035, + 0.012037033215165138, + 0.00725873839110136, + 0.006749947555363178, + 0.005759902764111757, + 0.006151595152914524, + 0.005726133473217487, + 0.006951977498829365, + 0.0026637837290763855, + 0.003404630348086357, + 0.02071443200111389, + 0.011155908927321434, + 0.013639855198562145, + 0.014430439099669456, + 0.0151932118460536, + 0.011465090326964855, + 0.011102651245892048, + 0.014237836934626102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.005252371542155743 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth drone tracking shot, shot from a slightly upward rear-view angle, closely follows a classic white Mitsubishi Lancer Evolution VI as it speeds along a winding coastal road. In the golden hour of evening, the warm sunlight casts a mirror-like glow on the car's paintwork, reflecting the pale orange sunset glow in the sky. The massive carbon fiber wing towering high above the rear, perfectly poised in the airflow, becomes the frame's most striking focal point. As the car enters a sharp S-curve with a nimble and graceful gesture, the camera pans steadily, capturing the car's slight tilt under centrifugal force. The red brake calipers on the bronze multi-spoke racing wheels flash by, and the rear end slides slightly outward in a perfectly controlled arc, its four tires gripping the dark gray asphalt like a virtuoso ballet dancer executing a precise and graceful spin. Then, the car quickly rights itself, unleashing its power without hesitation, sprinting towards the next corner with a deep roar from the exhaust. On one side of the road is a steep rocky cliff, and on the other side is a vast blue ocean with white waves. The whole picture is full of dynamic beauty and driving passion under the warm light and shadow, creating a free, focused and highly cinematic atmosphere.", + "seed": 24801, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_7314_seed_24801_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09031681232154369, + "vlm_logprob": -0.09031681232154369, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.09031681232154369, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21879364550113678 + ], + "opposite_logprob_tokens": [ + -1.8437936305999756 + ], + "logprob_reward": -0.21879364550113678, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21879364550113678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09790768474340439, + 0.05268026143312454, + 0.04577307030558586, + 0.037248462438583374, + 0.06640376150608063, + 0.0926518514752388, + 0.2628527581691742, + 0.1776292771100998, + 0.11168424040079117, + 0.055168669670820236 + ], + "credit_map": [ + 0.008316360414028168, + 0.010982805863022804, + 0.011621452867984772, + 0.01481090858578682, + 0.0094503378495574, + 0.00965140387415886, + 0.016571609303355217, + 0.016502808779478073, + 0.004868223797529936, + 0.00685669993981719, + 0.005748184397816658, + 0.009324944578111172, + 0.0049542090855538845, + 0.004734975751489401, + 0.01079485472291708, + 0.005398173350840807, + 0.005550056230276823, + 0.005933221895247698, + 0.0051097930409014225, + 0.004236643202602863, + 0.00532165402546525, + 0.007126875687390566, + 0.004273304250091314, + 0.0082215191796422, + 0.003886944381520152, + 0.0035605377051979303, + 0.005177814047783613, + 0.003495683427900076, + 0.004384430591017008, + 0.006750129163265228, + 0.0041703712195158005, + 0.0058225542306900024, + 0.006704999599605799, + 0.005344487261027098, + 0.01240751426666975, + 0.007643379271030426, + 0.007323901169002056, + 0.008564280346035957, + 0.008943289518356323, + 0.009471913799643517, + 0.007817559875547886, + 0.00951413344591856, + 0.011514966376125813, + 0.01568971946835518, + 0.010327545925974846, + 0.009069142863154411, + 0.012996042147278786, + 0.015722747892141342, + 0.011420046910643578, + 0.006020123604685068, + 0.10109329223632812, + 0.01367742195725441, + 0.0069075683131814, + 0.016989892348647118, + 0.012442589737474918, + 0.09430183470249176, + 0.018014727160334587, + 0.016885587945580482, + 0.014926932752132416, + 0.02281181700527668, + 0.025643041357398033, + 0.035010699182748795, + 0.01826968789100647, + 0.026066776365041733, + 0.016671977937221527, + 0.00810012873262167, + 0.011429617181420326, + 0.00812608003616333, + 0.012266471982002258, + 0.029140949249267578, + 0.01424191053956747, + 0.011707101948559284, + 0.007345245219767094, + 0.00454008998349309, + 0.007905106991529465, + 0.006904484238475561, + 0.004106390289962292, + 0.005817585624754429, + 0.008727921172976494, + 0.009821843355894089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.21879364550113678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790114775300026 + ], + "opposite_logprob_tokens": [ + -2.5790114402770996 + ], + "logprob_reward": -0.0790114775300026, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790114775300026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2552433907985687, + 0.3221138119697571, + 0.07714539766311646, + 0.034924302250146866, + 0.02986207604408264, + 0.05536564067006111, + 0.0899868831038475, + 0.08466871082782745, + 0.035887062549591064, + 0.014802759513258934 + ], + "credit_map": [ + 0.04195201024413109, + 0.029619615525007248, + 0.033446382731199265, + 0.037109147757291794, + 0.023971065878868103, + 0.024518363177776337, + 0.03952089324593544, + 0.025105901062488556, + 0.03600246086716652, + 0.025484221056103706, + 0.03787534683942795, + 0.03653924912214279, + 0.024009020999073982, + 0.023708300665020943, + 0.05108211189508438, + 0.08741311728954315, + 0.008807491511106491, + 0.012761913239955902, + 0.011201580986380577, + 0.00840854737907648, + 0.010534129105508327, + 0.007245562970638275, + 0.011884755454957485, + 0.006301417946815491, + 0.003436206141486764, + 0.003548927139490843, + 0.005956088658422232, + 0.0032674018293619156, + 0.005120014306157827, + 0.004083471372723579, + 0.004959901794791222, + 0.004552291240543127, + 0.0030171300750225782, + 0.0030463978182524443, + 0.0037006207276135683, + 0.005874159745872021, + 0.0028368989005684853, + 0.0037060179747641087, + 0.003101786831393838, + 0.004579063970595598, + 0.004187887534499168, + 0.006224398035556078, + 0.00598813546821475, + 0.00947712641209364, + 0.006491405889391899, + 0.00563138909637928, + 0.010116951540112495, + 0.007248345762491226, + 0.006370948161929846, + 0.005192859563976526, + 0.02528751641511917, + 0.011238223873078823, + 0.008308861404657364, + 0.00790974497795105, + 0.006300497334450483, + 0.019378231838345528, + 0.008619984611868858, + 0.01023488212376833, + 0.00940629467368126, + 0.010843461379408836, + 0.012850279919803143, + 0.00988752394914627, + 0.008533049374818802, + 0.01429323386400938, + 0.004004932940006256, + 0.003811282804235816, + 0.006518773268908262, + 0.004744121804833412, + 0.0034689251333475113, + 0.003883976023644209, + 0.004253426566720009, + 0.00520162470638752, + 0.0026433563325554132, + 0.0020376730244606733, + 0.0022226611617952585, + 0.0015172837302088737, + 0.0009396225796081126, + 0.0010577364591881633, + 0.0020747878588736057, + 0.0023096378426998854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0790114775300026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704521417617798 + ], + "opposite_logprob_tokens": [ + -2.127045154571533 + ], + "logprob_reward": -0.12704521417617798, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704521417617798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054685045033693314, + 0.014663939364254475, + 0.2792668044567108, + 0.012033781036734581, + 0.009720798581838608, + 0.024677781388163567, + 0.4689271152019501, + 0.06601189821958542, + 0.048636894673109055, + 0.02137591503560543 + ], + "credit_map": [ + 0.007413353305310011, + 0.005480001214891672, + 0.0071723866276443005, + 0.008162708953022957, + 0.005428802687674761, + 0.004700064659118652, + 0.009541853331029415, + 0.006785871926695108, + 0.002095907460898161, + 0.001351076178252697, + 0.001991823548451066, + 0.001992620527744293, + 0.0017966895829886198, + 0.0013832137919962406, + 0.002279144711792469, + 0.0017734644934535027, + 0.027743445709347725, + 0.04429434984922409, + 0.03239904344081879, + 0.03675464168190956, + 0.03619842976331711, + 0.027029816061258316, + 0.04688867926597595, + 0.027958424761891365, + 0.0013243774883449078, + 0.0017263239715248346, + 0.001675022067502141, + 0.0015309243462979794, + 0.0014672522665932775, + 0.0012961311731487513, + 0.0015482186572626233, + 0.0014655316481366754, + 0.0009714698535390198, + 0.0009841385763138533, + 0.0017185371834784746, + 0.0010564159601926804, + 0.0012939379084855318, + 0.0011464288691058755, + 0.001167411683127284, + 0.0013824591878801584, + 0.0026044733822345734, + 0.001970927696675062, + 0.0030574295669794083, + 0.0023133005015552044, + 0.00432372884824872, + 0.004202202428132296, + 0.0029960251413285732, + 0.0032096959184855223, + 0.021168719977140427, + 0.00969119742512703, + 0.18257686495780945, + 0.02495105005800724, + 0.009879294782876968, + 0.028602831065654755, + 0.02308795601129532, + 0.16896921396255493, + 0.005543341860175133, + 0.005251931492239237, + 0.0063608004711568356, + 0.010322767309844494, + 0.005828523077070713, + 0.009327050298452377, + 0.010972395539283752, + 0.012405095621943474, + 0.006510969717055559, + 0.003687788499519229, + 0.004285600036382675, + 0.0038462292868644, + 0.004982023034244776, + 0.012612281367182732, + 0.007288357242941856, + 0.005423649679869413, + 0.0036997683346271515, + 0.004628954455256462, + 0.0029256800189614296, + 0.0022533065639436245, + 0.0012797663221135736, + 0.0014748224057257175, + 0.0026647753547877073, + 0.002448845189064741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.12704521417617798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544183060526848 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.012544183060526848, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544183060526848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26857227087020874, + 0.08031948655843735, + 0.18462109565734863, + 0.02525351755321026, + 0.07994161546230316, + 0.07277311384677887, + 0.07548199594020844, + 0.1444198340177536, + 0.03914973512291908, + 0.02946731261909008 + ], + "credit_map": [ + 0.038507990539073944, + 0.029088007286190987, + 0.034697115421295166, + 0.03966202959418297, + 0.025695044547319412, + 0.02832517959177494, + 0.04292234778404236, + 0.029674559831619263, + 0.013356518000364304, + 0.013245491310954094, + 0.01076852809637785, + 0.009521838277578354, + 0.006853819824755192, + 0.008420360274612904, + 0.007307832129299641, + 0.010845089331269264, + 0.017710283398628235, + 0.042134881019592285, + 0.016861623153090477, + 0.01920176111161709, + 0.028543438762426376, + 0.025874555110931396, + 0.0202503502368927, + 0.014044202864170074, + 0.0028552678413689137, + 0.002892110962420702, + 0.004010877106338739, + 0.0026355949230492115, + 0.003147814190015197, + 0.004144996404647827, + 0.0028073955327272415, + 0.0027594619896262884, + 0.007059869822114706, + 0.007270862814038992, + 0.013529862277209759, + 0.013153750449419022, + 0.009632247500121593, + 0.009342987090349197, + 0.009320512413978577, + 0.010631523095071316, + 0.004501940682530403, + 0.006896923761814833, + 0.008597246371209621, + 0.013224136084318161, + 0.007980396039783955, + 0.007850220426917076, + 0.012333317659795284, + 0.011388931423425674, + 0.0061965505592525005, + 0.0040609706193208694, + 0.02148202806711197, + 0.007778276689350605, + 0.005001092329621315, + 0.0059418934397399426, + 0.004735922906547785, + 0.020285261794924736, + 0.010118488222360611, + 0.009473111480474472, + 0.016629332676529884, + 0.025571612641215324, + 0.013697788119316101, + 0.022291943430900574, + 0.02590790018439293, + 0.02072964794933796, + 0.003993222955614328, + 0.004144026432186365, + 0.006433521863073111, + 0.003986954689025879, + 0.004355778452008963, + 0.007421295624226332, + 0.005320142954587936, + 0.0034947884269058704, + 0.007784467656165361, + 0.00979052297770977, + 0.0026612726505845785, + 0.0015090708620846272, + 0.0014382090885192156, + 0.0017066060099750757, + 0.00195843493565917, + 0.0026187279727309942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.012544183060526848 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189541339874268 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189541339874268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189541339874268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06018293276429176, + 0.04423435404896736, + 0.2677313983440399, + 0.04929346963763237, + 0.09188885241746902, + 0.08077728003263474, + 0.21878504753112793, + 0.05601426959037781, + 0.09247258305549622, + 0.03861977532505989 + ], + "credit_map": [ + 0.006272537633776665, + 0.006293070036917925, + 0.006927207112312317, + 0.009699378162622452, + 0.005483681336045265, + 0.005470651667565107, + 0.009611907415091991, + 0.010424499399960041, + 0.005050469655543566, + 0.003988496493548155, + 0.005465216003358364, + 0.006872930098325014, + 0.003465309739112854, + 0.004653593525290489, + 0.004880093969404697, + 0.009858245961368084, + 0.02690068818628788, + 0.03927110135555267, + 0.0348978117108345, + 0.03467501327395439, + 0.03462963551282883, + 0.025592979043722153, + 0.04586125910282135, + 0.025902917608618736, + 0.005552197340875864, + 0.00513413967564702, + 0.006892661564052105, + 0.005211679730564356, + 0.006026149727404118, + 0.0075602601282298565, + 0.00632873410359025, + 0.006587646435946226, + 0.00757979741320014, + 0.009001652710139751, + 0.014378231018781662, + 0.015368993394076824, + 0.012871358543634415, + 0.010010860860347748, + 0.009679784066975117, + 0.012998173013329506, + 0.005987328942865133, + 0.008665462955832481, + 0.009568782523274422, + 0.015068547800183296, + 0.009542861022055149, + 0.0073946029879152775, + 0.011948788538575172, + 0.012600901536643505, + 0.0124529292806983, + 0.006503150332719088, + 0.0806107372045517, + 0.014772885479032993, + 0.006145127583295107, + 0.012642055749893188, + 0.011297875083982944, + 0.07436030358076096, + 0.006981897167861462, + 0.006385054439306259, + 0.004400258418172598, + 0.006252121180295944, + 0.0060823531821370125, + 0.008035698905587196, + 0.005846130661666393, + 0.01203075423836708, + 0.00917737279087305, + 0.010250160470604897, + 0.018833769485354424, + 0.009569810703396797, + 0.008675926364958286, + 0.011917621828615665, + 0.011596038937568665, + 0.012451882474124432, + 0.0043409159407019615, + 0.0040639168582856655, + 0.007335138041526079, + 0.0043012890964746475, + 0.003495631506666541, + 0.004306880757212639, + 0.005356586072593927, + 0.005419417750090361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.014189541339874268 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth drone tracking shot, shot from a slightly upward rear-view angle, closely follows a classic white Mitsubishi Lancer Evolution VI as it speeds along a winding coastal road. In the golden hour of evening, the warm sunlight casts a mirror-like glow on the car's paintwork, reflecting the pale orange sunset glow in the sky. The massive carbon fiber wing towering high above the rear, perfectly poised in the airflow, becomes the frame's most striking focal point. As the car enters a sharp S-curve with a nimble and graceful gesture, the camera pans steadily, capturing the car's slight tilt under centrifugal force. The red brake calipers on the bronze multi-spoke racing wheels flash by, and the rear end slides slightly outward in a perfectly controlled arc, its four tires gripping the dark gray asphalt like a virtuoso ballet dancer executing a precise and graceful spin. Then, the car quickly rights itself, unleashing its power without hesitation, sprinting towards the next corner with a deep roar from the exhaust. On one side of the road is a steep rocky cliff, and on the other side is a vast blue ocean with white waves. The whole picture is full of dynamic beauty and driving passion under the warm light and shadow, creating a free, focused and highly cinematic atmosphere.", + "seed": 24802, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_7314_seed_24802_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05656981710344553, + "vlm_logprob": -0.05656981710344553, + "vlm_yes_no_margin": 3.424999976158142, + "vlm_token_credit": -0.05656981710344553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16477997601032257 + ], + "opposite_logprob_tokens": [ + -2.0397799015045166 + ], + "logprob_reward": -0.16477997601032257, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.16477997601032257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12564726173877716, + 0.24271348118782043, + 0.044010426849126816, + 0.10833065211772919, + 0.08346052467823029, + 0.032616958022117615, + 0.041600219905376434, + 0.075422003865242, + 0.15454335510730743, + 0.09165508300065994 + ], + "credit_map": [ + 0.01593003235757351, + 0.015852591022849083, + 0.01335203368216753, + 0.02434547431766987, + 0.01280178502202034, + 0.011601299978792667, + 0.016344111412763596, + 0.015419931150972843, + 0.008275170810520649, + 0.0649910643696785, + 0.008604479022324085, + 0.027816953137516975, + 0.09968691319227219, + 0.01329735480248928, + 0.008674449287354946, + 0.011367104016244411, + 0.0038093714974820614, + 0.007133800070732832, + 0.007746540941298008, + 0.013856046833097935, + 0.002430272987112403, + 0.0029553212225437164, + 0.00300940265879035, + 0.0030696713365614414, + 0.010723816230893135, + 0.008960380218923092, + 0.011750034987926483, + 0.022715501487255096, + 0.009033295325934887, + 0.01696452498435974, + 0.009414584375917912, + 0.01876850239932537, + 0.009381295181810856, + 0.018725134432315826, + 0.005507235415279865, + 0.02649565599858761, + 0.004649450071156025, + 0.0065609184093773365, + 0.0056393821723759174, + 0.006501455791294575, + 0.007288340013474226, + 0.002539984881877899, + 0.01027149148285389, + 0.002847061725333333, + 0.0021237714681774378, + 0.0025874366983771324, + 0.002130728680640459, + 0.0028281465638428926, + 0.007640887051820755, + 0.0037050964310765266, + 0.0047233328223228455, + 0.006260832771658897, + 0.0031546619720757008, + 0.003192286007106304, + 0.006685110740363598, + 0.006238010711967945, + 0.010727707296609879, + 0.0034325183369219303, + 0.00724109448492527, + 0.006710246670991182, + 0.004940614569932222, + 0.025821847841143608, + 0.0051340097561478615, + 0.011413965374231339, + 0.023394804447889328, + 0.0152779845520854, + 0.027078591287136078, + 0.013904526829719543, + 0.029313165694475174, + 0.01852445863187313, + 0.012354366481304169, + 0.014695446938276291, + 0.01359888631850481, + 0.017649758607149124, + 0.011488595977425575, + 0.00846795178949833, + 0.01154731959104538, + 0.011261122301220894, + 0.007629403378814459, + 0.010012047365307808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.16477997601032257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023325109854340553 + ], + "opposite_logprob_tokens": [ + -3.773325204849243 + ], + "logprob_reward": -0.023325109854340553, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023325109854340553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14972788095474243, + 0.3879481554031372, + 0.13949088752269745, + 0.022167380899190903, + 0.025779007002711296, + 0.02856656163930893, + 0.023650558665394783, + 0.054010238498449326, + 0.04349943995475769, + 0.12515981495380402 + ], + "credit_map": [ + 0.019256848841905594, + 0.03164242208003998, + 0.01774863339960575, + 0.022965017706155777, + 0.01409983728080988, + 0.011946543119847775, + 0.017601002007722855, + 0.014467564411461353, + 0.021980412304401398, + 0.12181416898965836, + 0.017897311598062515, + 0.032187119126319885, + 0.12606778740882874, + 0.03157777711749077, + 0.014693059958517551, + 0.021730531007051468, + 0.0085604852065444, + 0.03849593922495842, + 0.052850980311632156, + 0.010146264918148518, + 0.0067530227825045586, + 0.0049570538103580475, + 0.005847145803272724, + 0.01188001036643982, + 0.004402026068419218, + 0.0019110721768811345, + 0.003267044899985194, + 0.0032512659672647715, + 0.0017625915352255106, + 0.003250251756981015, + 0.0015408892650157213, + 0.002782240742817521, + 0.0018405786249786615, + 0.005315051414072514, + 0.001930508529767394, + 0.008707795292139053, + 0.0016228100284934044, + 0.0014612361555919051, + 0.0032654593233019114, + 0.0016355660045519471, + 0.0065224142745137215, + 0.004064742475748062, + 0.004213282838463783, + 0.002177867107093334, + 0.003035216359421611, + 0.00223649013787508, + 0.0024146398063749075, + 0.0039019091054797173, + 0.003859811695292592, + 0.002695229137316346, + 0.002933334792032838, + 0.004355911165475845, + 0.0019719472620636225, + 0.0019606102723628283, + 0.0030522379092872143, + 0.002821477362886071, + 0.007794141303747892, + 0.00530731026083231, + 0.006902904249727726, + 0.008903603069484234, + 0.0050217583775520325, + 0.00847293995320797, + 0.00483495369553566, + 0.0067726257257163525, + 0.006688783876597881, + 0.006168605759739876, + 0.004406472202390432, + 0.0029523561242967844, + 0.0075327069498598576, + 0.00656807329505682, + 0.0031971975695341825, + 0.00598524697124958, + 0.02103600464761257, + 0.023367328569293022, + 0.009552613832056522, + 0.00541648780927062, + 0.027541037648916245, + 0.01532839797437191, + 0.012280408293008804, + 0.010637528263032436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.023325109854340553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003537565469742 + ], + "opposite_logprob_tokens": [ + -2.695035457611084 + ], + "logprob_reward": -0.07003537565469742, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003537565469742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05167609453201294, + 0.11977750062942505, + 0.16895222663879395, + 0.03941779583692551, + 0.0351150780916214, + 0.06554583460092545, + 0.07835135608911514, + 0.2795434594154358, + 0.046589139848947525, + 0.11503154039382935 + ], + "credit_map": [ + 0.008442503400146961, + 0.007296087685972452, + 0.004975042771548033, + 0.009167023934423923, + 0.004534659907221794, + 0.005803040228784084, + 0.005150644574314356, + 0.00630709296092391, + 0.0058565097860991955, + 0.044007524847984314, + 0.004048211965709925, + 0.005623938050121069, + 0.04377781227231026, + 0.007731296122074127, + 0.003966664895415306, + 0.004765541758388281, + 0.009889474138617516, + 0.03047795034945011, + 0.03265654668211937, + 0.07209547609090805, + 0.004375899210572243, + 0.004026531241834164, + 0.008932150900363922, + 0.006498189643025398, + 0.005439144093543291, + 0.0031328799668699503, + 0.004423816688358784, + 0.00644312147051096, + 0.0037066724617034197, + 0.007352171000093222, + 0.0035860114730894566, + 0.0053339796140789986, + 0.002782223280519247, + 0.008998672477900982, + 0.002579989843070507, + 0.012195521965622902, + 0.0015858747065067291, + 0.0021082761231809855, + 0.001872834051027894, + 0.002991684712469578, + 0.020566456019878387, + 0.005441255867481232, + 0.00893473718315363, + 0.003803390311077237, + 0.0038664545863866806, + 0.004346118308603764, + 0.004244949668645859, + 0.0143424728885293, + 0.01703318953514099, + 0.007403851952403784, + 0.007882092148065567, + 0.009427204728126526, + 0.004456211347132921, + 0.006548463832587004, + 0.009418635629117489, + 0.01618170738220215, + 0.04278450459241867, + 0.01261937152594328, + 0.025531861931085587, + 0.01534590870141983, + 0.02084042690694332, + 0.09668200463056564, + 0.013826734386384487, + 0.05191265791654587, + 0.007596374023705721, + 0.005589961539953947, + 0.0050070867873728275, + 0.004052367527037859, + 0.010030589066445827, + 0.0063775572925806046, + 0.0034419179428368807, + 0.00449328264221549, + 0.01907791942358017, + 0.01793142966926098, + 0.009192120283842087, + 0.00555357476696372, + 0.024291861802339554, + 0.019428418949246407, + 0.010478469543159008, + 0.009077740833163261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.07003537565469742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072474420070648 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072474420070648, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072474420070648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11627122014760971, + 0.0901491641998291, + 0.2288270890712738, + 0.042884331196546555, + 0.01863176003098488, + 0.027124257758259773, + 0.04192496836185455, + 0.1138685792684555, + 0.04213147610425949, + 0.2781871557235718 + ], + "credit_map": [ + 0.012391151860356331, + 0.01463757548481226, + 0.013776852749288082, + 0.018108252435922623, + 0.01304368395358324, + 0.017457030713558197, + 0.012580200098454952, + 0.014276477508246899, + 0.009708983823657036, + 0.009514198638498783, + 0.006203868426382542, + 0.02336890436708927, + 0.01635592244565487, + 0.009745919145643711, + 0.007808047812432051, + 0.007443324197083712, + 0.014243069104850292, + 0.0630534216761589, + 0.08641679584980011, + 0.012279302813112736, + 0.010231423191726208, + 0.0077927871607244015, + 0.01149493083357811, + 0.023315347731113434, + 0.008080150932073593, + 0.003478894243016839, + 0.0038673649542033672, + 0.006148654501885176, + 0.004671212285757065, + 0.005567402113229036, + 0.0051432219333946705, + 0.005927428137511015, + 0.004061571322381496, + 0.003573971800506115, + 0.001559278927743435, + 0.003103238297626376, + 0.0016321439761668444, + 0.00110190873965621, + 0.0017722909105941653, + 0.0018273575697094202, + 0.0048409393057227135, + 0.0043359058909118176, + 0.0025637231301516294, + 0.0015637906035408378, + 0.0033809191081672907, + 0.0026015935000032187, + 0.0020050068851560354, + 0.00583238061517477, + 0.006312238518148661, + 0.005052125081419945, + 0.005057205446064472, + 0.007252089213579893, + 0.004632107447832823, + 0.00476174708455801, + 0.005510053131729364, + 0.0033474056981503963, + 0.01606873609125614, + 0.006320788525044918, + 0.009316465817391872, + 0.010441062971949577, + 0.009203309193253517, + 0.040020983666181564, + 0.0057204971089959145, + 0.016776742413640022, + 0.006156452931463718, + 0.0033848315943032503, + 0.003859354415908456, + 0.0040438976138830185, + 0.009669264778494835, + 0.0064811608754098415, + 0.0033261235803365707, + 0.0052103884518146515, + 0.0464375875890255, + 0.04142976179718971, + 0.019783107563853264, + 0.010078954510390759, + 0.05920245125889778, + 0.0527627132833004, + 0.02543758414685726, + 0.02305498719215393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.016072474420070648 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863614957779646 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863614957779646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863614957779646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13556836545467377, + 0.11244040727615356, + 0.18944284319877625, + 0.04288187623023987, + 0.029805360361933708, + 0.03469988331198692, + 0.03236927092075348, + 0.12134768068790436, + 0.08362133800983429, + 0.21782293915748596 + ], + "credit_map": [ + 0.017527544870972633, + 0.016083816066384315, + 0.014917688444256783, + 0.030404875054955482, + 0.012070124968886375, + 0.01154195424169302, + 0.015652405098080635, + 0.017369961366057396, + 0.011127681471407413, + 0.0385354645550251, + 0.0051498692482709885, + 0.013463417068123817, + 0.03101532720029354, + 0.0048431698232889175, + 0.0051287878304719925, + 0.0031766972970217466, + 0.0111623490229249, + 0.0548187680542469, + 0.06466004997491837, + 0.015223554335534573, + 0.008923865854740143, + 0.00616580992937088, + 0.00929268915206194, + 0.01919577643275261, + 0.006938072387129068, + 0.0035158784594386816, + 0.0052038091234862804, + 0.006179218180477619, + 0.0034715398214757442, + 0.007706569507718086, + 0.003463776782155037, + 0.006403016392141581, + 0.00227943598292768, + 0.007901076227426529, + 0.002242807764559984, + 0.009875329211354256, + 0.0012286814162507653, + 0.0013833099510520697, + 0.002900628373026848, + 0.0019940929487347603, + 0.007507974747568369, + 0.003746687900274992, + 0.005456932820379734, + 0.0023737007286399603, + 0.0030814888887107372, + 0.004439550917595625, + 0.0032857186160981655, + 0.0048078312538564205, + 0.004450672306120396, + 0.007813346572220325, + 0.003603806719183922, + 0.004392856732010841, + 0.0025899598840624094, + 0.002066398737952113, + 0.004002213012427092, + 0.003450020682066679, + 0.017351141199469566, + 0.005921728443354368, + 0.008932634256780148, + 0.008654284290969372, + 0.0077744703739881516, + 0.04528573155403137, + 0.010580042377114296, + 0.016847655177116394, + 0.013146797195076942, + 0.008561337366700172, + 0.016704615205526352, + 0.009069755673408508, + 0.01152174361050129, + 0.008584382012486458, + 0.007116619031876326, + 0.0089160967618227, + 0.035349518060684204, + 0.02098136767745018, + 0.014086227864027023, + 0.008942365646362305, + 0.035985954105854034, + 0.06351905316114426, + 0.013056344352662563, + 0.025902127847075462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.00863614957779646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth drone tracking shot, shot from a slightly upward rear-view angle, closely follows a classic white Mitsubishi Lancer Evolution VI as it speeds along a winding coastal road. In the golden hour of evening, the warm sunlight casts a mirror-like glow on the car's paintwork, reflecting the pale orange sunset glow in the sky. The massive carbon fiber wing towering high above the rear, perfectly poised in the airflow, becomes the frame's most striking focal point. As the car enters a sharp S-curve with a nimble and graceful gesture, the camera pans steadily, capturing the car's slight tilt under centrifugal force. The red brake calipers on the bronze multi-spoke racing wheels flash by, and the rear end slides slightly outward in a perfectly controlled arc, its four tires gripping the dark gray asphalt like a virtuoso ballet dancer executing a precise and graceful spin. Then, the car quickly rights itself, unleashing its power without hesitation, sprinting towards the next corner with a deep roar from the exhaust. On one side of the road is a steep rocky cliff, and on the other side is a vast blue ocean with white waves. The whole picture is full of dynamic beauty and driving passion under the warm light and shadow, creating a free, focused and highly cinematic atmosphere.", + "seed": 24803, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_13/index_7314_seed_24803_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0699196256697178, + "vlm_logprob": -0.0699196256697178, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.0699196256697178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18306627869606018 + ], + "opposite_logprob_tokens": [ + -2.0580663681030273 + ], + "logprob_reward": -0.18306627869606018, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.18306627869606018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10907023400068283, + 0.057569053024053574, + 0.037014320492744446, + 0.04527764767408371, + 0.05885084345936775, + 0.0897836834192276, + 0.07648032903671265, + 0.14048011600971222, + 0.1052609458565712, + 0.28021278977394104 + ], + "credit_map": [ + 0.01974663883447647, + 0.007236114237457514, + 0.016221964731812477, + 0.017419569194316864, + 0.009189296513795853, + 0.011427648365497589, + 0.012839475646615028, + 0.014989522285759449, + 0.005621977616101503, + 0.005319006275385618, + 0.008689709939062595, + 0.008486025966703892, + 0.0069906022399663925, + 0.004893746692687273, + 0.00917145423591137, + 0.008396528661251068, + 0.0039963433519005775, + 0.003437008708715439, + 0.006279369816184044, + 0.005813680589199066, + 0.0033485086169093847, + 0.004035748541355133, + 0.0054168966598808765, + 0.00468676770105958, + 0.003752513090148568, + 0.005513498559594154, + 0.006237988360226154, + 0.005362893920391798, + 0.005682616960257292, + 0.006458562798798084, + 0.005675217602401972, + 0.006594355218112469, + 0.0037421700544655323, + 0.004702944308519363, + 0.00647327397018671, + 0.008595015853643417, + 0.005098749417811632, + 0.009438605047762394, + 0.012020011432468891, + 0.008780072443187237, + 0.011289193294942379, + 0.010982634499669075, + 0.011847278103232384, + 0.01126103661954403, + 0.005658030044287443, + 0.026381725445389748, + 0.0042579565197229385, + 0.00810582097619772, + 0.007222852203994989, + 0.00836836826056242, + 0.011610141023993492, + 0.01184349786490202, + 0.010086175054311752, + 0.01079987920820713, + 0.0073562925681471825, + 0.00919311959296465, + 0.017759956419467926, + 0.019643809646368027, + 0.01632329821586609, + 0.016713066026568413, + 0.020435694605112076, + 0.014356203377246857, + 0.019307149574160576, + 0.015940936282277107, + 0.031683798879384995, + 0.00546389352530241, + 0.03464736416935921, + 0.005736485123634338, + 0.007376677356660366, + 0.008052128367125988, + 0.0061148423701524734, + 0.006185758858919144, + 0.03102046251296997, + 0.008226142264902592, + 0.11066453158855438, + 0.013746839016675949, + 0.01195657067000866, + 0.08118073642253876, + 0.00774629320949316, + 0.01567123271524906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.18306627869606018 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023320451378822327 + ], + "opposite_logprob_tokens": [ + -3.773320436477661 + ], + "logprob_reward": -0.023320451378822327, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023320451378822327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4461939334869385, + 0.043467093259096146, + 0.06716936826705933, + 0.02429480478167534, + 0.02239111438393593, + 0.028547396883368492, + 0.031227268278598785, + 0.04963211342692375, + 0.034823209047317505, + 0.2522536516189575 + ], + "credit_map": [ + 0.0580485500395298, + 0.02499743551015854, + 0.04666208475828171, + 0.12260231375694275, + 0.03737619146704674, + 0.04759438708424568, + 0.04870568960905075, + 0.06020732223987579, + 0.00798348244279623, + 0.0026626361068338156, + 0.0071598561480641365, + 0.003894696244969964, + 0.0034478334710001945, + 0.00560833141207695, + 0.005161757115274668, + 0.007548503577709198, + 0.007876086980104446, + 0.005039055831730366, + 0.010315279476344585, + 0.006645144894719124, + 0.008631355129182339, + 0.012424600310623646, + 0.007908214814960957, + 0.008329633623361588, + 0.002607915550470352, + 0.0015511439414694905, + 0.0036430242471396923, + 0.002415197202935815, + 0.002391042886301875, + 0.006184483878314495, + 0.002218506531789899, + 0.003283492289483547, + 0.0017328327521681786, + 0.0021346015855669975, + 0.002915649674832821, + 0.002930372953414917, + 0.002777781803160906, + 0.003086881013587117, + 0.0031054182909429073, + 0.0037075751461088657, + 0.00421745702624321, + 0.004298384767025709, + 0.0031123613007366657, + 0.0052453745156526566, + 0.002516167238354683, + 0.004512343555688858, + 0.0023838302586227655, + 0.0022614779882133007, + 0.003475896082818508, + 0.003037901595234871, + 0.004786534234881401, + 0.003810438560321927, + 0.00406474806368351, + 0.004861676134169102, + 0.0027504239697009325, + 0.004439650569111109, + 0.0135914022102952, + 0.004778016824275255, + 0.007847139611840248, + 0.003313901601359248, + 0.004010841250419617, + 0.007774780970066786, + 0.0034675172064453363, + 0.004848513752222061, + 0.00830459501594305, + 0.001819348894059658, + 0.009006437845528126, + 0.0028166831471025944, + 0.0031048236414790154, + 0.002939304569736123, + 0.0031200596131384373, + 0.003711957950145006, + 0.027147816494107246, + 0.005883324425667524, + 0.101717509329319, + 0.01191691868007183, + 0.010334428399801254, + 0.07642852514982224, + 0.006862697657197714, + 0.011962410062551498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.023320451378822327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705382704734802 + ], + "opposite_logprob_tokens": [ + -2.127053737640381 + ], + "logprob_reward": -0.12705382704734802, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12705382704734802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6519626975059509, + 0.018500450998544693, + 0.009754779748618603, + 0.010201306082308292, + 0.0334298349916935, + 0.06512938439846039, + 0.03500715270638466, + 0.031318653374910355, + 0.08785868436098099, + 0.05683710053563118 + ], + "credit_map": [ + 0.08561733365058899, + 0.03742365539073944, + 0.06939652562141418, + 0.1790730357170105, + 0.056270454078912735, + 0.06296759843826294, + 0.07286316156387329, + 0.08835086971521378, + 0.0021558632142841816, + 0.0019107359694316983, + 0.0033057963009923697, + 0.0018049398204311728, + 0.0016258503310382366, + 0.0025103213265538216, + 0.0024491094518452883, + 0.0027378348167985678, + 0.0012019892456009984, + 0.0007323254831135273, + 0.0015090180095285177, + 0.0014440867817029357, + 0.0008249097154475749, + 0.0013801561435684562, + 0.0013793963007628918, + 0.001282897312194109, + 0.0012362048728391528, + 0.0007611644105054438, + 0.0012871056096628308, + 0.0009676925837993622, + 0.0011551112402230501, + 0.0021736882627010345, + 0.0012540173484012485, + 0.001366321579553187, + 0.0017707743681967258, + 0.0018406840972602367, + 0.0034946787636727095, + 0.003022956196218729, + 0.002283361041918397, + 0.003442747751250863, + 0.008904117159545422, + 0.008670510724186897, + 0.008319997228682041, + 0.00792565755546093, + 0.007243385072797537, + 0.007875993847846985, + 0.005332253407686949, + 0.016254212707281113, + 0.004913132172077894, + 0.007264749146997929, + 0.003350654849782586, + 0.004759232979267836, + 0.004188281483948231, + 0.005655841901898384, + 0.00552818737924099, + 0.004265243653208017, + 0.003813308198004961, + 0.0034464015625417233, + 0.005086634773761034, + 0.0029762410558760166, + 0.0037591750733554363, + 0.0030495007522404194, + 0.0048840525560081005, + 0.004097571596503258, + 0.0038518591318279505, + 0.0036136142443865538, + 0.029982801526784897, + 0.0029136177618056536, + 0.036909494549036026, + 0.00376946572214365, + 0.0022227985318750143, + 0.005178707651793957, + 0.0027926245238631964, + 0.004089170601218939, + 0.0060718972235918045, + 0.0016206580912694335, + 0.02139398455619812, + 0.0030501459259539843, + 0.0025898099411278963, + 0.01652907207608223, + 0.0017581656575202942, + 0.003823362523689866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.12705382704734802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542416341602802 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542416341602802, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542416341602802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21213163435459137, + 0.019281776621937752, + 0.005761154927313328, + 0.011202946305274963, + 0.013315470889210701, + 0.011151616461575031, + 0.008876118808984756, + 0.01534108817577362, + 0.10571801662445068, + 0.5972201824188232 + ], + "credit_map": [ + 0.02687653712928295, + 0.01234276033937931, + 0.023023445159196854, + 0.058222696185112, + 0.018598033115267754, + 0.02007787488400936, + 0.024447456002235413, + 0.02854282408952713, + 0.0034689409658312798, + 0.0015631432179361582, + 0.0028158570639789104, + 0.002079428406432271, + 0.0018163721542805433, + 0.0019369327928870916, + 0.0022137919440865517, + 0.00338730844669044, + 0.0006067872163839638, + 0.0005052565829828382, + 0.0008457692456431687, + 0.0006829789490438998, + 0.0007247799076139927, + 0.0009712040773592889, + 0.0007010933477431536, + 0.000723285018466413, + 0.0010089885909110308, + 0.0008129859925247729, + 0.003408049466088414, + 0.0014646409545093775, + 0.0006962834158912301, + 0.0016376512357965112, + 0.0009877760894596577, + 0.0011865691049024463, + 0.0009904876351356506, + 0.0007494972669519484, + 0.0013976921327412128, + 0.0014963651774451137, + 0.0011411075247451663, + 0.0015236036852002144, + 0.002551708370447159, + 0.003465008456259966, + 0.0010945094982162118, + 0.0012148357927799225, + 0.0016399321611970663, + 0.0016763333696871996, + 0.000851024582516402, + 0.002654939191415906, + 0.0007956875488162041, + 0.0012243528617545962, + 0.0008077597594819963, + 0.000753213360439986, + 0.0013674686197191477, + 0.0011376047041267157, + 0.0010522822849452496, + 0.0018263672245666385, + 0.0006201966316439211, + 0.0013112254673615098, + 0.003725537331774831, + 0.0016194774070754647, + 0.0021650257986038923, + 0.0013458691537380219, + 0.0013180415844544768, + 0.002358401659876108, + 0.0010395392309874296, + 0.0017691950779408216, + 0.019824253395199776, + 0.0042089177295565605, + 0.0406409315764904, + 0.005087432451546192, + 0.004561925772577524, + 0.008656097576022148, + 0.003498777514323592, + 0.019239675253629684, + 0.06456682831048965, + 0.013513064943253994, + 0.24319060146808624, + 0.027165662497282028, + 0.0247181486338377, + 0.1800328642129898, + 0.015851425006985664, + 0.028181573376059532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012542416341602802 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic white Mitsubishi Lancer Evolution VI?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car driving on a winding coastal road with a steep rocky cliff on one side and a vast blue ocean with white waves on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car feature a massive carbon fiber wing towering high above the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car perform a sharp S-curve maneuver with a slight tilt under centrifugal force while the camera pans steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a slightly upward rear-view angle with a smooth drone tracking motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036151548847556114 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036151548847556114, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036151548847556114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2921334207057953, + 0.024663299322128296, + 0.06495540589094162, + 0.013613296672701836, + 0.01795029267668724, + 0.025773797184228897, + 0.016472794115543365, + 0.043457239866256714, + 0.14963127672672272, + 0.3513491451740265 + ], + "credit_map": [ + 0.03733747452497482, + 0.016004735603928566, + 0.02889503911137581, + 0.08537480980157852, + 0.024247316643595695, + 0.027328314259648323, + 0.033046986907720566, + 0.03989877924323082, + 0.0034071803092956543, + 0.002028713934123516, + 0.003172596450895071, + 0.00313133979216218, + 0.002834273036569357, + 0.0034053842537105083, + 0.002596022794023156, + 0.004087791778147221, + 0.006973860785365105, + 0.007366081699728966, + 0.012829010374844074, + 0.010915461927652359, + 0.005664083641022444, + 0.006169403437525034, + 0.007808620575815439, + 0.00722888857126236, + 0.0014929716708138585, + 0.0016777783166617155, + 0.0017979759722948074, + 0.0014658697182312608, + 0.001663501956500113, + 0.0020809194538742304, + 0.0016930263955146074, + 0.0017412543529644608, + 0.0012527722865343094, + 0.0014469533925876021, + 0.002194303087890148, + 0.0022210346069186926, + 0.002080779755488038, + 0.003233605995774269, + 0.002291450509801507, + 0.0032293943222612143, + 0.0042567322961986065, + 0.0035996597725898027, + 0.00487288786098361, + 0.0036467313766479492, + 0.0022009843960404396, + 0.00287610269151628, + 0.002147694118320942, + 0.0021730083972215652, + 0.001500593265518546, + 0.0018665334209799767, + 0.002324357395991683, + 0.002841274021193385, + 0.002317913109436631, + 0.002189070452004671, + 0.0017482821131125093, + 0.0016847720835357904, + 0.010015306994318962, + 0.004785818979144096, + 0.005688093136996031, + 0.00388138135895133, + 0.0037089481484144926, + 0.005312963388860226, + 0.004758982453495264, + 0.005305752623826265, + 0.02384890988469124, + 0.013419130817055702, + 0.02271188050508499, + 0.019183162599802017, + 0.02120158262550831, + 0.015719421207904816, + 0.021016668528318405, + 0.012530531734228134, + 0.03832796588540077, + 0.007351252716034651, + 0.14403802156448364, + 0.016053233295679092, + 0.013901757076382637, + 0.10636963695287704, + 0.009354200214147568, + 0.01595311239361763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0036151548847556114 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_14/index_tensor([1881, 1881])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_14/index_tensor([1881, 1881])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..14b17f1285d3114445d047f1cfcaea07cfca7db4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_14/index_tensor([1881, 1881])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping drone shot descends slowly from the inky black sky of Iceland at midnight, capturing a magnificent spectacle of the collision of heaven and earth. Across the frozen earth, a fiery red lava flow tears through the thick ice cap with unstoppable force. Golden-orange cracks spread rapidly across the crystalline surface, instantly sending up a vast cloud of white steam. Meanwhile, a vast, emerald green aurora dances across the sky like a vast, shimmering ribbon, its mysterious pulsing glow creating a stark and stunning visual contrast to the raging fire of the earth below. The camera begins to pan forward, smoothly passing over a deep, blue ice cavern, steam rising continuously from geothermal heat. The ice walls within, illuminated by the combined effects of the lava and the aurora, reflect a hazy interplay of blue, red, and green. Suddenly, the camera pitches downward, plummeting at breakneck speed, towards the roiling lava lake below, seething like the heart of a giant beast. In extreme slow motion, as the perspective immerses, viscous liquid rock is stirred up, and countless golden hot sparks splash upwards. They draw graceful arcs in the air and slowly rise, just meeting, interweaving and merging with the emerald green aurora particles falling in the sky, and finally converge into a surreal and brilliant waterfall composed of pure light and heat. The whole scene climaxes in violent power and extreme beauty, creating a doomsday atmosphere that is both magnificent and poetic.", + "seed": 10775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_14/index_1881_seed_10775_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07361767769325525, + "vlm_logprob": -0.07361767769325525, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.07361767769325525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20183944702148438 + ], + "opposite_logprob_tokens": [ + -1.7018394470214844 + ], + "logprob_reward": -0.20183944702148438, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20183944702148438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10981377959251404, + 0.40097105503082275, + 0.03187735751271248, + 0.12205211818218231, + 0.048235177993774414, + 0.04147831350564957, + 0.06707363575696945, + 0.05494432523846626, + 0.08425242453813553, + 0.03930184990167618 + ], + "credit_map": [ + 0.010206195525825024, + 0.015480410307645798, + 0.014642619527876377, + 0.013727271929383278, + 0.014279745519161224, + 0.016793884336948395, + 0.013813188299536705, + 0.010870449244976044, + 0.03783724829554558, + 0.09321301430463791, + 0.029598413035273552, + 0.04541938751935959, + 0.06876592338085175, + 0.05918943136930466, + 0.03786112740635872, + 0.029086483642458916, + 0.003988638520240784, + 0.00435672327876091, + 0.0035111447796225548, + 0.005240433383733034, + 0.0023985907901078463, + 0.0038435200694948435, + 0.004494606051594019, + 0.004043699707835913, + 0.01237727515399456, + 0.008723385632038116, + 0.023908104747533798, + 0.030226921662688255, + 0.010982546955347061, + 0.00824437104165554, + 0.010676453821361065, + 0.01691305637359619, + 0.005803816951811314, + 0.008806610479950905, + 0.008136660791933537, + 0.006602691486477852, + 0.0038620848208665848, + 0.003933585248887539, + 0.006674583535641432, + 0.004415144212543964, + 0.003406550968065858, + 0.00399793591350317, + 0.0034589036367833614, + 0.003062650328502059, + 0.003394440282136202, + 0.01148518081754446, + 0.002899852814152837, + 0.009772798046469688, + 0.006256352178752422, + 0.012293615378439426, + 0.005090015009045601, + 0.00872298888862133, + 0.0074488515965640545, + 0.011337143369019032, + 0.010725838132202625, + 0.005198824685066938, + 0.007108908612281084, + 0.013330467976629734, + 0.005540046840906143, + 0.0050393142737448215, + 0.006280205212533474, + 0.008166268467903137, + 0.0057074809446930885, + 0.003771629184484482, + 0.008316653780639172, + 0.008573626168072224, + 0.016382664442062378, + 0.00870529841631651, + 0.013660505414009094, + 0.007951680570840836, + 0.007268499117344618, + 0.013393497094511986, + 0.0032337771262973547, + 0.003676587948575616, + 0.007620436139404774, + 0.004121420904994011, + 0.005291840992867947, + 0.00476172985509038, + 0.005696851760149002, + 0.004899200517684221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.20183944702148438 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001719545223750174 + ], + "opposite_logprob_tokens": [ + -6.3767194747924805 + ], + "logprob_reward": -0.001719545223750174, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001719545223750174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0676848292350769, + 0.41991689801216125, + 0.0224445853382349, + 0.05742628872394562, + 0.0470585934817791, + 0.0962190330028534, + 0.06022236868739128, + 0.023202737793326378, + 0.1684192270040512, + 0.03740541636943817 + ], + "credit_map": [ + 0.008716562762856483, + 0.009670054540038109, + 0.009930943138897419, + 0.007528555579483509, + 0.007812448311597109, + 0.009868934750556946, + 0.007668450940400362, + 0.00648887874558568, + 0.03912089392542839, + 0.09417407214641571, + 0.02911609411239624, + 0.047741565853357315, + 0.07009588927030563, + 0.06184544041752815, + 0.04242171719670296, + 0.03540123254060745, + 0.0032497041393071413, + 0.0028404174372553825, + 0.00229069497436285, + 0.0035480314400047064, + 0.002779032802209258, + 0.002823252696543932, + 0.002898690290749073, + 0.0020147613249719143, + 0.006684964522719383, + 0.004832405596971512, + 0.0090714106336236, + 0.013062916696071625, + 0.005431681871414185, + 0.005936501082032919, + 0.005212422925978899, + 0.007193985395133495, + 0.005046330392360687, + 0.006394983734935522, + 0.008305473253130913, + 0.006960839964449406, + 0.005273398011922836, + 0.0042609297670423985, + 0.006435437127947807, + 0.004381200764328241, + 0.007232344709336758, + 0.00807487964630127, + 0.006159066688269377, + 0.004581954795867205, + 0.009200703352689743, + 0.032652564346790314, + 0.004624109715223312, + 0.02369341067969799, + 0.0057774255983531475, + 0.005868021864444017, + 0.007112597115337849, + 0.00678168423473835, + 0.011983679607510567, + 0.01028091087937355, + 0.006981590297073126, + 0.005436461418867111, + 0.004200179595500231, + 0.002793690888211131, + 0.004201515112072229, + 0.0024273900780826807, + 0.0020415065810084343, + 0.002273336285725236, + 0.0026888963766396046, + 0.002576223574578762, + 0.016305267810821533, + 0.01636153645813465, + 0.032390743494033813, + 0.018532976508140564, + 0.029111025854945183, + 0.016176331788301468, + 0.01414561178535223, + 0.02539573796093464, + 0.004307287745177746, + 0.0041419933550059795, + 0.0065828366205096245, + 0.004324722103774548, + 0.004438246134668589, + 0.005352729465812445, + 0.003799963742494583, + 0.004457639064639807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.001719545223750174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900816738605499 + ], + "opposite_logprob_tokens": [ + -2.464008092880249 + ], + "logprob_reward": -0.08900816738605499, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900816738605499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14677149057388306, + 0.26330575346946716, + 0.035166215151548386, + 0.05340017378330231, + 0.04359270632266998, + 0.2299463301897049, + 0.12556643784046173, + 0.03653125464916229, + 0.04668404906988144, + 0.019035620614886284 + ], + "credit_map": [ + 0.01759105920791626, + 0.024025365710258484, + 0.020102649927139282, + 0.01709667779505253, + 0.022612756118178368, + 0.02559600956737995, + 0.011399539187550545, + 0.008347437717020512, + 0.024272888898849487, + 0.0562019981443882, + 0.019343676045536995, + 0.030123626813292503, + 0.044241584837436676, + 0.03788011148571968, + 0.027747174724936485, + 0.023494677618145943, + 0.00594598613679409, + 0.003969261888414621, + 0.00303612626157701, + 0.005498091224581003, + 0.0037721963599324226, + 0.006751460954546928, + 0.003580996533855796, + 0.002612094394862652, + 0.0050704702734947205, + 0.005026839207857847, + 0.008402046747505665, + 0.010483029298484325, + 0.005993668455630541, + 0.00587110873311758, + 0.005498596467077732, + 0.007054415065795183, + 0.004276393447071314, + 0.006485318765044212, + 0.006673850584775209, + 0.0053726364858448505, + 0.005502359475940466, + 0.0058509064838290215, + 0.0056121195666491985, + 0.0038191229104995728, + 0.017931856215000153, + 0.018533295020461082, + 0.012928717769682407, + 0.008052698336541653, + 0.023005768656730652, + 0.08036594837903976, + 0.008656437508761883, + 0.06047159433364868, + 0.009376883506774902, + 0.01079739723354578, + 0.009497974999248981, + 0.014713654294610023, + 0.027671203017234802, + 0.027173830196261406, + 0.016995351761579514, + 0.009340143762528896, + 0.005088614299893379, + 0.007151974830776453, + 0.004083668813109398, + 0.004791149869561195, + 0.002885835012421012, + 0.003970022778958082, + 0.0044253007508814335, + 0.00413468899205327, + 0.004509501159191132, + 0.006496727932244539, + 0.00669782143086195, + 0.00473056873306632, + 0.008891473524272442, + 0.004240351729094982, + 0.004728574771434069, + 0.006389028392732143, + 0.0022513996809720993, + 0.001776273944415152, + 0.0019706531893461943, + 0.0013543062377721071, + 0.004333813674747944, + 0.0019572328310459852, + 0.002569548785686493, + 0.0028223921544849873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.08900816738605499 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057810313999653 + ], + "opposite_logprob_tokens": [ + -3.895578145980835 + ], + "logprob_reward": -0.02057810313999653, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057810313999653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0784497782588005, + 0.08338417112827301, + 0.03518693894147873, + 0.056785862892866135, + 0.02095194347202778, + 0.2874169945716858, + 0.12445824593305588, + 0.07286745309829712, + 0.1872795671224594, + 0.05321894958615303 + ], + "credit_map": [ + 0.007857097312808037, + 0.013249537907540798, + 0.008025960065424442, + 0.009592125192284584, + 0.013273659162223339, + 0.012005165219306946, + 0.008665814064443111, + 0.005780431907624006, + 0.008874176070094109, + 0.018718793988227844, + 0.007981343194842339, + 0.009585310705006123, + 0.014584863558411598, + 0.01263031642884016, + 0.006639844737946987, + 0.004369533620774746, + 0.004488213919103146, + 0.003842991776764393, + 0.004171580076217651, + 0.008336180821061134, + 0.0026001364458352327, + 0.004041225183755159, + 0.004244337789714336, + 0.003462275257334113, + 0.009590161964297295, + 0.005893973167985678, + 0.007021197117865086, + 0.01155136153101921, + 0.005607069469988346, + 0.0051100351847708225, + 0.005941402167081833, + 0.006070668809115887, + 0.0028637235518544912, + 0.004509688355028629, + 0.002158928895369172, + 0.002780972048640251, + 0.0023679411970078945, + 0.0022914810106158257, + 0.001968605211004615, + 0.0020106069277971983, + 0.021515298634767532, + 0.024011746048927307, + 0.017217636108398438, + 0.011183501221239567, + 0.028527885675430298, + 0.10016824305057526, + 0.011659129522740841, + 0.07313360273838043, + 0.008998941630125046, + 0.008104436099529266, + 0.0111178457736969, + 0.017898010089993477, + 0.020829198881983757, + 0.026360057294368744, + 0.02137279510498047, + 0.00977698341012001, + 0.012755105271935463, + 0.007856129668653011, + 0.007883552461862564, + 0.009433833882212639, + 0.007687227800488472, + 0.012459736317396164, + 0.007500361651182175, + 0.00729151489213109, + 0.016703829169273376, + 0.02094634622335434, + 0.037119582295417786, + 0.016512732952833176, + 0.033521752804517746, + 0.01671765185892582, + 0.016251524910330772, + 0.029506174847483635, + 0.004088741727173328, + 0.0046474505215883255, + 0.009049620479345322, + 0.005261329934000969, + 0.006628047674894333, + 0.00580490892753005, + 0.0097853047773242, + 0.007953549735248089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.02057810313999653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494312569499016 + ], + "opposite_logprob_tokens": [ + -2.929943084716797 + ], + "logprob_reward": -0.05494312569499016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494312569499016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10775584727525711, + 0.17604286968708038, + 0.04173378273844719, + 0.03987138718366623, + 0.05841761827468872, + 0.08541436493396759, + 0.06660565733909607, + 0.155370831489563, + 0.16377682983875275, + 0.10501078516244888 + ], + "credit_map": [ + 0.012147380039095879, + 0.015417732298374176, + 0.01798507571220398, + 0.014624078758060932, + 0.01189403422176838, + 0.016322072595357895, + 0.010381214320659637, + 0.008984257467091084, + 0.016105948016047478, + 0.039794329553842545, + 0.013174795545637608, + 0.021540233865380287, + 0.02710956521332264, + 0.02553480677306652, + 0.01693332940340042, + 0.01584986038506031, + 0.007770608179271221, + 0.005144012626260519, + 0.004475873429328203, + 0.004255376756191254, + 0.004497971851378679, + 0.007436458021402359, + 0.0027022412978112698, + 0.005451239179819822, + 0.0031546594109386206, + 0.003697528038173914, + 0.007741783279925585, + 0.006220133975148201, + 0.004424468148499727, + 0.004147248342633247, + 0.004297276493161917, + 0.0061882915906608105, + 0.007219486869871616, + 0.008636044338345528, + 0.009324438869953156, + 0.0070583270862698555, + 0.005719085689634085, + 0.00539470836520195, + 0.008043638430535793, + 0.007021888624876738, + 0.006392247974872589, + 0.006854624021798372, + 0.008635635487735271, + 0.007831933908164501, + 0.009047405794262886, + 0.025297364220023155, + 0.00642979284748435, + 0.014925370924174786, + 0.0065095811150968075, + 0.007505310699343681, + 0.007466944400221109, + 0.009818194434046745, + 0.008792019449174404, + 0.012049469165503979, + 0.00965466070920229, + 0.004809483885765076, + 0.016411511227488518, + 0.0192338228225708, + 0.017002204433083534, + 0.023517971858382225, + 0.019584691151976585, + 0.026072660461068153, + 0.020724689587950706, + 0.012823277153074741, + 0.017256377264857292, + 0.018147483468055725, + 0.01824255846440792, + 0.02893189713358879, + 0.028064865618944168, + 0.01825215481221676, + 0.01706431247293949, + 0.017817191779613495, + 0.013800333254039288, + 0.01092977449297905, + 0.01303127408027649, + 0.011948027648031712, + 0.013983003795146942, + 0.010922189801931381, + 0.0176197849214077, + 0.012776413932442665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.05494312569499016 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot descends slowly from the inky black sky of Iceland at midnight, capturing a magnificent spectacle of the collision of heaven and earth. Across the frozen earth, a fiery red lava flow tears through the thick ice cap with unstoppable force. Golden-orange cracks spread rapidly across the crystalline surface, instantly sending up a vast cloud of white steam. Meanwhile, a vast, emerald green aurora dances across the sky like a vast, shimmering ribbon, its mysterious pulsing glow creating a stark and stunning visual contrast to the raging fire of the earth below. The camera begins to pan forward, smoothly passing over a deep, blue ice cavern, steam rising continuously from geothermal heat. The ice walls within, illuminated by the combined effects of the lava and the aurora, reflect a hazy interplay of blue, red, and green. Suddenly, the camera pitches downward, plummeting at breakneck speed, towards the roiling lava lake below, seething like the heart of a giant beast. In extreme slow motion, as the perspective immerses, viscous liquid rock is stirred up, and countless golden hot sparks splash upwards. They draw graceful arcs in the air and slowly rise, just meeting, interweaving and merging with the emerald green aurora particles falling in the sky, and finally converge into a surreal and brilliant waterfall composed of pure light and heat. The whole scene climaxes in violent power and extreme beauty, creating a doomsday atmosphere that is both magnificent and poetic.", + "seed": 10776, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_14/index_1881_seed_10776_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04218900797422975, + "vlm_logprob": -0.04218900797422975, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.04218900797422975, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03857926279306412 + ], + "opposite_logprob_tokens": [ + -3.288579225540161 + ], + "logprob_reward": -0.03857926279306412, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03857926279306412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0849541574716568, + 0.14185591042041779, + 0.13630728423595428, + 0.06272463500499725, + 0.2613784074783325, + 0.08271102607250214, + 0.02594805136322975, + 0.1112830638885498, + 0.050211101770401, + 0.04262639582157135 + ], + "credit_map": [ + 0.00886594783514738, + 0.01099951658397913, + 0.010568616911768913, + 0.011715499684214592, + 0.0112000722438097, + 0.012294218875467777, + 0.009030080400407314, + 0.010280202142894268, + 0.013894317671656609, + 0.01521376147866249, + 0.009858045727014542, + 0.015544747933745384, + 0.031244082376360893, + 0.03036392293870449, + 0.011174573563039303, + 0.014562458731234074, + 0.009698058478534222, + 0.007439220789819956, + 0.028233099728822708, + 0.007358217611908913, + 0.021273508667945862, + 0.04938014596700668, + 0.0066726552322506905, + 0.006252381484955549, + 0.006402760744094849, + 0.006339203100651503, + 0.01685273088514805, + 0.006402011029422283, + 0.005461130756884813, + 0.006791690830141306, + 0.0075822737999260426, + 0.006892834790050983, + 0.018060026690363884, + 0.020388629287481308, + 0.03257482126355171, + 0.02057083696126938, + 0.06864354014396667, + 0.032114170491695404, + 0.0252570491284132, + 0.043769340962171555, + 0.006110756192356348, + 0.008925083093345165, + 0.014743587002158165, + 0.01065053604543209, + 0.008159054443240166, + 0.011014244519174099, + 0.011568356305360794, + 0.011539407074451447, + 0.0025086740497499704, + 0.0026152103673666716, + 0.003555838018655777, + 0.0026316961739212275, + 0.0018091978272423148, + 0.0033282293006777763, + 0.003876489121466875, + 0.005622717086225748, + 0.02653832733631134, + 0.012850110419094563, + 0.005778692662715912, + 0.004417791031301022, + 0.0124583151191473, + 0.034034859389066696, + 0.006501167546957731, + 0.00870379339903593, + 0.005776401609182358, + 0.004908168688416481, + 0.004063705913722515, + 0.004081667400896549, + 0.008780312724411488, + 0.010481774806976318, + 0.006146151106804609, + 0.0059729148633778095, + 0.0030435435473918915, + 0.0022586865816265345, + 0.003927335608750582, + 0.006819659378379583, + 0.004335538484156132, + 0.005286249797791243, + 0.005773576442152262, + 0.011181802488863468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.03857926279306412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001046467456035316 + ], + "opposite_logprob_tokens": [ + -6.876046657562256 + ], + "logprob_reward": -0.001046467456035316, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001046467456035316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09758095443248749, + 0.21243004500865936, + 0.057909511029720306, + 0.05953541770577431, + 0.09919768571853638, + 0.05873806029558182, + 0.05715249851346016, + 0.27399343252182007, + 0.0372004471719265, + 0.046261921525001526 + ], + "credit_map": [ + 0.007511074189096689, + 0.01004900224506855, + 0.012749243527650833, + 0.017048396170139313, + 0.012579343281686306, + 0.013837402686476707, + 0.010516772978007793, + 0.013289733789861202, + 0.018113138154149055, + 0.015840759500861168, + 0.024103673174977303, + 0.027301106601953506, + 0.03986242413520813, + 0.06290317326784134, + 0.011373517103493214, + 0.0129322474822402, + 0.004177789203822613, + 0.0040217009373009205, + 0.013172167353332043, + 0.0036678083706647158, + 0.00792611762881279, + 0.01636713370680809, + 0.003908025566488504, + 0.004668771289288998, + 0.004519239533692598, + 0.005932488478720188, + 0.012161542661488056, + 0.006504513788968325, + 0.008108214475214481, + 0.00754260690882802, + 0.006526354234665632, + 0.00824046228080988, + 0.006809080019593239, + 0.007620242424309254, + 0.015512518584728241, + 0.009722755290567875, + 0.021815109997987747, + 0.01143803820014, + 0.011442169547080994, + 0.014837777242064476, + 0.004777144640684128, + 0.00907078105956316, + 0.008910281583666801, + 0.005071241874247789, + 0.008573793806135654, + 0.006249861791729927, + 0.007035478949546814, + 0.009049479849636555, + 0.0050886753015220165, + 0.0060497974045574665, + 0.005982099566608667, + 0.004732208792120218, + 0.009752319194376469, + 0.006086598616093397, + 0.005339599680155516, + 0.01412121020257473, + 0.07232945412397385, + 0.03256441652774811, + 0.0120119983330369, + 0.008400620892643929, + 0.026525165885686874, + 0.09221459180116653, + 0.011409700848162174, + 0.018537482246756554, + 0.00402033980935812, + 0.004585704766213894, + 0.004249719902873039, + 0.005891735199838877, + 0.0047443280927836895, + 0.006237081717699766, + 0.003862477373331785, + 0.003609066130593419, + 0.0019103188533335924, + 0.004063751082867384, + 0.0016032155836001039, + 0.0018039809074252844, + 0.015594868920743465, + 0.014760148711502552, + 0.003280156757682562, + 0.0032454838510602713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.001046467456035316 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863252118229866 + ], + "opposite_logprob_tokens": [ + -3.0486326217651367 + ], + "logprob_reward": -0.04863252118229866, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863252118229866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32710403203964233, + 0.13898174464702606, + 0.09970733523368835, + 0.06230127438902855, + 0.059909164905548096, + 0.050834618508815765, + 0.08252139389514923, + 0.05057897791266441, + 0.06602775305509567, + 0.062033746391534805 + ], + "credit_map": [ + 0.062170784920454025, + 0.0682610273361206, + 0.03559388592839241, + 0.03389446437358856, + 0.03645836561918259, + 0.04246869683265686, + 0.022826001048088074, + 0.025430789217352867, + 0.015199673362076283, + 0.01639867015182972, + 0.019261248409748077, + 0.03057130239903927, + 0.014509221538901329, + 0.013891235925257206, + 0.013341155834496021, + 0.015809230506420135, + 0.00703605217859149, + 0.006130296736955643, + 0.022724511101841927, + 0.008586323820054531, + 0.014138304628431797, + 0.024780936539173126, + 0.009179054759442806, + 0.007131851278245449, + 0.005126833450049162, + 0.00515820411965251, + 0.011513007804751396, + 0.008091755211353302, + 0.004925781860947609, + 0.006645056419074535, + 0.010230513289570808, + 0.010610121302306652, + 0.005153351463377476, + 0.006282647140324116, + 0.011882622726261616, + 0.00959106720983982, + 0.006782755721360445, + 0.004785197786986828, + 0.0060732546262443066, + 0.009358267299830914, + 0.004070434719324112, + 0.005382556933909655, + 0.007926544174551964, + 0.008037655614316463, + 0.004585649352520704, + 0.004800115246325731, + 0.00914949644356966, + 0.006882168352603912, + 0.0054477183148264885, + 0.013631443493068218, + 0.005868018139153719, + 0.0054650804959237576, + 0.022629888728260994, + 0.006319310050457716, + 0.007159072905778885, + 0.016000857576727867, + 0.006722743157297373, + 0.008414478041231632, + 0.004237964749336243, + 0.003799856873229146, + 0.0065293507650494576, + 0.005441572517156601, + 0.005494223441928625, + 0.00993878673762083, + 0.007422739174216986, + 0.007642932701855898, + 0.0069850776344537735, + 0.006459676194936037, + 0.009630310349166393, + 0.012329746037721634, + 0.007894224487245083, + 0.0076630511321127415, + 0.002879531355574727, + 0.005380250513553619, + 0.0025521889328956604, + 0.003016676288098097, + 0.020365318283438683, + 0.018041186034679413, + 0.004635551944375038, + 0.005163044668734074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04863252118229866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777866303920746 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.009777866303920746, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777866303920746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16002975404262543, + 0.1322132647037506, + 0.08379033207893372, + 0.13946807384490967, + 0.10372950881719589, + 0.06362350285053253, + 0.06255442649126053, + 0.11117241531610489, + 0.10512407124042511, + 0.038294680416584015 + ], + "credit_map": [ + 0.020994799211621284, + 0.024980144575238228, + 0.020584138110280037, + 0.02526581473648548, + 0.021034792065620422, + 0.019681250676512718, + 0.01641920767724514, + 0.011069589294493198, + 0.012470229528844357, + 0.014445753768086433, + 0.015929682180285454, + 0.021141164004802704, + 0.020934253931045532, + 0.026181669905781746, + 0.010343939997255802, + 0.01076655276119709, + 0.007045956328511238, + 0.017961271107196808, + 0.009565114043653011, + 0.008651796728372574, + 0.011402334086596966, + 0.014631799422204494, + 0.007501492276787758, + 0.00703055877238512, + 0.009970908053219318, + 0.013925442472100258, + 0.0479624941945076, + 0.014734122902154922, + 0.0067551471292972565, + 0.007564262952655554, + 0.010357174091041088, + 0.02819851040840149, + 0.008435784839093685, + 0.010872864164412022, + 0.009976396337151527, + 0.009745561517775059, + 0.024018630385398865, + 0.012780516408383846, + 0.008856810629367828, + 0.019042935222387314, + 0.006021964829415083, + 0.008698157034814358, + 0.009003402665257454, + 0.010012524202466011, + 0.006820594891905785, + 0.005876869894564152, + 0.008916893973946571, + 0.008273091167211533, + 0.005678290035575628, + 0.009319190867245197, + 0.007961024530231953, + 0.006997972726821899, + 0.005655114538967609, + 0.007286424748599529, + 0.008916672319173813, + 0.01073972787708044, + 0.01979936845600605, + 0.01382749155163765, + 0.005975425709038973, + 0.006009693257510662, + 0.012940826825797558, + 0.025721726939082146, + 0.011174678802490234, + 0.015723198652267456, + 0.008307126350700855, + 0.01006404124200344, + 0.007982349023222923, + 0.012816154398024082, + 0.0212869793176651, + 0.020216017961502075, + 0.011698518879711628, + 0.012752881273627281, + 0.0027607057709246874, + 0.0037010852247476578, + 0.0031756185926496983, + 0.0044502620585262775, + 0.005763573106378317, + 0.0070763882249593735, + 0.003910130355507135, + 0.007456914521753788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009777866303920746 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290892213582993 + ], + "opposite_logprob_tokens": [ + -2.2379088401794434 + ], + "logprob_reward": -0.11290892213582993, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290892213582993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06984138488769531, + 0.10355692356824875, + 0.046196553856134415, + 0.05992387607693672, + 0.10529650747776031, + 0.11146873980760574, + 0.07763927429914474, + 0.15464934706687927, + 0.13720734417438507, + 0.13422010838985443 + ], + "credit_map": [ + 0.006383060477674007, + 0.014952915720641613, + 0.007637504022568464, + 0.009158879518508911, + 0.006961557548493147, + 0.012607315555214882, + 0.006722543854266405, + 0.005417599342763424, + 0.012506856583058834, + 0.017106270417571068, + 0.011816752143204212, + 0.010516687296330929, + 0.01540265791118145, + 0.014562978409230709, + 0.00807876605540514, + 0.01356594730168581, + 0.005305349361151457, + 0.0067130704410374165, + 0.0041750711388885975, + 0.004257504362612963, + 0.006471544504165649, + 0.009682935662567616, + 0.00529289897531271, + 0.004298177547752857, + 0.004000545479357243, + 0.004667902365326881, + 0.019228029996156693, + 0.005851011723279953, + 0.005821713712066412, + 0.005352978128939867, + 0.004308546427637339, + 0.010693141259253025, + 0.007104027550667524, + 0.008820191957056522, + 0.015124392695724964, + 0.009667261503636837, + 0.020445074886083603, + 0.012417430058121681, + 0.011780815199017525, + 0.019937291741371155, + 0.007329718675464392, + 0.013788183219730854, + 0.020785106346011162, + 0.01153325941413641, + 0.013305876404047012, + 0.013808679766952991, + 0.012657864019274712, + 0.018260041251778603, + 0.006295900791883469, + 0.014482712373137474, + 0.005476147402077913, + 0.004408623091876507, + 0.02446173131465912, + 0.007970890030264854, + 0.005743317771703005, + 0.00879993848502636, + 0.022447049617767334, + 0.01832260750234127, + 0.011219022795557976, + 0.012270444072782993, + 0.014167610555887222, + 0.0168890543282032, + 0.018317893147468567, + 0.04101564735174179, + 0.01010298915207386, + 0.010452746413648129, + 0.01034022867679596, + 0.015554730780422688, + 0.028520895168185234, + 0.030273977667093277, + 0.017108403146266937, + 0.01485336385667324, + 0.0053968001157045364, + 0.01001117005944252, + 0.0070633660070598125, + 0.00846727192401886, + 0.0407455749809742, + 0.03707199543714523, + 0.011223583482205868, + 0.014240343123674393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.11290892213582993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot descends slowly from the inky black sky of Iceland at midnight, capturing a magnificent spectacle of the collision of heaven and earth. Across the frozen earth, a fiery red lava flow tears through the thick ice cap with unstoppable force. Golden-orange cracks spread rapidly across the crystalline surface, instantly sending up a vast cloud of white steam. Meanwhile, a vast, emerald green aurora dances across the sky like a vast, shimmering ribbon, its mysterious pulsing glow creating a stark and stunning visual contrast to the raging fire of the earth below. The camera begins to pan forward, smoothly passing over a deep, blue ice cavern, steam rising continuously from geothermal heat. The ice walls within, illuminated by the combined effects of the lava and the aurora, reflect a hazy interplay of blue, red, and green. Suddenly, the camera pitches downward, plummeting at breakneck speed, towards the roiling lava lake below, seething like the heart of a giant beast. In extreme slow motion, as the perspective immerses, viscous liquid rock is stirred up, and countless golden hot sparks splash upwards. They draw graceful arcs in the air and slowly rise, just meeting, interweaving and merging with the emerald green aurora particles falling in the sky, and finally converge into a surreal and brilliant waterfall composed of pure light and heat. The whole scene climaxes in violent power and extreme beauty, creating a doomsday atmosphere that is both magnificent and poetic.", + "seed": 10777, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_14/index_1881_seed_10777_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0298831342253834, + "vlm_logprob": -0.0298831342253834, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.0298831342253834, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007001626770943403 + ], + "opposite_logprob_tokens": [ + -5.0070013999938965 + ], + "logprob_reward": -0.007001626770943403, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007001626770943403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11064708977937698, + 0.04218233749270439, + 0.1061074361205101, + 0.10089680552482605, + 0.1767069548368454, + 0.18269085884094238, + 0.08149044215679169, + 0.10927937179803848, + 0.05219902843236923, + 0.0377996489405632 + ], + "credit_map": [ + 0.010233098641037941, + 0.012037184089422226, + 0.008872948586940765, + 0.018330616876482964, + 0.01228534709662199, + 0.011228647083044052, + 0.018444141373038292, + 0.019215114414691925, + 0.003079973626881838, + 0.0032134857028722763, + 0.004162669647485018, + 0.0068400041200220585, + 0.0039245388470590115, + 0.004539707209914923, + 0.008648245595395565, + 0.0077737136743962765, + 0.01123869139701128, + 0.012641891837120056, + 0.010858017951250076, + 0.019907427951693535, + 0.013757802546024323, + 0.00943619292229414, + 0.018894216045737267, + 0.009373199194669724, + 0.009014520794153214, + 0.013995248824357986, + 0.008686870336532593, + 0.014608862809836864, + 0.018541952595114708, + 0.015513194724917412, + 0.011365653946995735, + 0.009170507080852985, + 0.015767890959978104, + 0.015649916604161263, + 0.03607773035764694, + 0.019827770069241524, + 0.016908468678593636, + 0.02102271094918251, + 0.017913946881890297, + 0.033538516610860825, + 0.013238007202744484, + 0.04943107068538666, + 0.01471379492431879, + 0.01814218983054161, + 0.012618516571819782, + 0.05028276890516281, + 0.010725488886237144, + 0.013539030216634274, + 0.01092235092073679, + 0.01453516073524952, + 0.008878680877387524, + 0.006353943608701229, + 0.015009508468210697, + 0.008492724038660526, + 0.007755358703434467, + 0.009542716667056084, + 0.007989311590790749, + 0.009873693808913231, + 0.024926716461777687, + 0.0125508401542902, + 0.017220880836248398, + 0.013086278922855854, + 0.01109542977064848, + 0.012536222115159035, + 0.005563345272094011, + 0.00797788705676794, + 0.007746577262878418, + 0.006198289804160595, + 0.010320146568119526, + 0.0063746715895831585, + 0.004668843932449818, + 0.003349264385178685, + 0.004042309243232012, + 0.004065419547259808, + 0.005612223409116268, + 0.003981530200690031, + 0.0047462377697229385, + 0.006222897674888372, + 0.0045111291110515594, + 0.004617900121957064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.007001626770943403 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003618955845013261 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.003618955845013261, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003618955845013261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06682824343442917, + 0.06955539435148239, + 0.06508742272853851, + 0.08271774649620056, + 0.10627660900354385, + 0.3248080909252167, + 0.14929011464118958, + 0.04079558700323105, + 0.041072696447372437, + 0.0535680428147316 + ], + "credit_map": [ + 0.004980042576789856, + 0.007879083044826984, + 0.007744574453681707, + 0.01130355428904295, + 0.007122333161532879, + 0.008113916963338852, + 0.008138848468661308, + 0.011545899324119091, + 0.004897432401776314, + 0.007178882602602243, + 0.00718391640111804, + 0.009359076619148254, + 0.008410600014030933, + 0.008952206932008266, + 0.009864700958132744, + 0.013708589598536491, + 0.006362082436680794, + 0.0070673441514372826, + 0.005403147544711828, + 0.010538659058511257, + 0.009516519494354725, + 0.010691376402974129, + 0.010074127465486526, + 0.005434173624962568, + 0.0055840895511209965, + 0.009233338758349419, + 0.007609727326780558, + 0.010371146723628044, + 0.017590394243597984, + 0.016193505376577377, + 0.008443661034107208, + 0.007691886741667986, + 0.01062000636011362, + 0.009112131781876087, + 0.023969262838363647, + 0.011423022486269474, + 0.012581068091094494, + 0.010589992627501488, + 0.011106668040156364, + 0.016874469816684723, + 0.03181071579456329, + 0.13252581655979156, + 0.009991415776312351, + 0.01310935989022255, + 0.01918110065162182, + 0.09741976112127304, + 0.007684112526476383, + 0.013085850514471531, + 0.019000204280018806, + 0.02934836409986019, + 0.01731722243130207, + 0.012477497570216656, + 0.023626726120710373, + 0.013624649494886398, + 0.017712419852614403, + 0.01618303544819355, + 0.00405586464330554, + 0.007670223712921143, + 0.005462041590362787, + 0.004362175241112709, + 0.006635628640651703, + 0.00513781001791358, + 0.0030510821379721165, + 0.004420765675604343, + 0.004058496560901403, + 0.006787924561649561, + 0.005349243059754372, + 0.006359544582664967, + 0.007895765826106071, + 0.00411863811314106, + 0.0035176407545804977, + 0.002985447645187378, + 0.004788137041032314, + 0.005037466064095497, + 0.006479197181761265, + 0.004277541767805815, + 0.00825821328908205, + 0.013732337392866611, + 0.005682742223143578, + 0.005312413442879915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.003618955845013261 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304501786828041 + ], + "opposite_logprob_tokens": [ + -3.1680450439453125 + ], + "logprob_reward": -0.04304501786828041, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304501786828041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1070604994893074, + 0.07023800909519196, + 0.07907193154096603, + 0.07713136821985245, + 0.2996918261051178, + 0.0925511047244072, + 0.10251924395561218, + 0.06143760308623314, + 0.06052803248167038, + 0.04977034777402878 + ], + "credit_map": [ + 0.00916915014386177, + 0.012065399438142776, + 0.00951636303216219, + 0.016528232023119926, + 0.012803321704268456, + 0.00973774865269661, + 0.017828645184636116, + 0.01941164955496788, + 0.005725525319576263, + 0.00796506553888321, + 0.008714714087545872, + 0.009996069595217705, + 0.008075115270912647, + 0.010297246277332306, + 0.010089248418807983, + 0.009375025518238544, + 0.005657399073243141, + 0.007186105474829674, + 0.0074988859705626965, + 0.014674789272248745, + 0.01305963285267353, + 0.008678136393427849, + 0.013221953995525837, + 0.009095033630728722, + 0.007206624373793602, + 0.009979226626455784, + 0.0070922463200986385, + 0.00886209961026907, + 0.018661702051758766, + 0.013034643605351448, + 0.005787324160337448, + 0.006507500074803829, + 0.03143231198191643, + 0.029407836496829987, + 0.05100471153855324, + 0.030110448598861694, + 0.041147954761981964, + 0.034769799560308456, + 0.02966788038611412, + 0.052150897681713104, + 0.01012409571558237, + 0.022661982104182243, + 0.008781631477177143, + 0.010002671740949154, + 0.004560266155749559, + 0.018814248964190483, + 0.005864138714969158, + 0.01174207217991352, + 0.010936257429420948, + 0.019330348819494247, + 0.011948766186833382, + 0.008945383131504059, + 0.015392185188829899, + 0.010428620502352715, + 0.014448014087975025, + 0.011089658364653587, + 0.0068666678853333, + 0.006565699819475412, + 0.011074716225266457, + 0.006366283632814884, + 0.008285550400614738, + 0.007515478879213333, + 0.006634896155446768, + 0.008128304034471512, + 0.00709920609369874, + 0.010670074261724949, + 0.006801195442676544, + 0.006771852727979422, + 0.011696086265146732, + 0.00813817698508501, + 0.005022638477385044, + 0.004328798037022352, + 0.0063226474449038506, + 0.004319531377404928, + 0.0049872491508722305, + 0.004024534020572901, + 0.008446171879768372, + 0.013150479644536972, + 0.004235374741256237, + 0.004284359049052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.04304501786828041 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11168569326400757, + 0.03279992938041687, + 0.06343016028404236, + 0.03832897171378136, + 0.17477962374687195, + 0.13155581057071686, + 0.2249375879764557, + 0.08191406726837158, + 0.06088433787226677, + 0.07968375831842422 + ], + "credit_map": [ + 0.007928182370960712, + 0.008603854104876518, + 0.010365513153374195, + 0.01813683845102787, + 0.010843873955309391, + 0.009348821826279163, + 0.024230612441897392, + 0.02222800813615322, + 0.0030031579080969095, + 0.004317176528275013, + 0.003635364817455411, + 0.003976430278271437, + 0.004266940522938967, + 0.0049657076597213745, + 0.004180707968771458, + 0.004454446490854025, + 0.005202004220336676, + 0.006845923140645027, + 0.004555803723633289, + 0.012012421153485775, + 0.010135598480701447, + 0.00615720683708787, + 0.012003958225250244, + 0.006517248693853617, + 0.003201984567567706, + 0.006400615908205509, + 0.0035067007411271334, + 0.003339586313813925, + 0.009098717011511326, + 0.0066988468170166016, + 0.003414989449083805, + 0.0026675332337617874, + 0.013486100360751152, + 0.013921627774834633, + 0.037025000900030136, + 0.018385563045740128, + 0.02547108195722103, + 0.03163076564669609, + 0.017306748777627945, + 0.01755276508629322, + 0.010421969927847385, + 0.042725373059511185, + 0.004056173842400312, + 0.005227049347013235, + 0.008658715523779392, + 0.04358953237533569, + 0.005819067358970642, + 0.011057958006858826, + 0.0299477931112051, + 0.04475924000144005, + 0.025750629603862762, + 0.016252832487225533, + 0.03846493363380432, + 0.025010379031300545, + 0.022325878962874413, + 0.022425932809710503, + 0.008824136108160019, + 0.012166949920356274, + 0.01003813836723566, + 0.009497982449829578, + 0.014489144086837769, + 0.010124401189386845, + 0.006563220173120499, + 0.010210113599896431, + 0.006046399939805269, + 0.006257170811295509, + 0.009052932262420654, + 0.009012272581458092, + 0.010957452468574047, + 0.007773579563945532, + 0.007217597682029009, + 0.0045669409446418285, + 0.007831660099327564, + 0.010508252307772636, + 0.008686770685017109, + 0.009005672298371792, + 0.01340305246412754, + 0.01604553498327732, + 0.0077385748736560345, + 0.0064642587676644325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.006734057795256376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901601284742355 + ], + "opposite_logprob_tokens": [ + -2.4640159606933594 + ], + "logprob_reward": -0.08901601284742355, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901601284742355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06677299737930298, + 0.04571325704455376, + 0.10113449394702911, + 0.05954321473836899, + 0.10882764309644699, + 0.051182202994823456, + 0.09501416981220245, + 0.07415808737277985, + 0.1331169307231903, + 0.2645370662212372 + ], + "credit_map": [ + 0.006219067610800266, + 0.00626351498067379, + 0.007010201923549175, + 0.010266219265758991, + 0.007886216975748539, + 0.006890378892421722, + 0.008899650536477566, + 0.013337749056518078, + 0.0033629683312028646, + 0.005459354259073734, + 0.0044728429056704044, + 0.006718222983181477, + 0.008611717261373997, + 0.006387627683579922, + 0.005796000827103853, + 0.004904520697891712, + 0.008406019769608974, + 0.010716619901359081, + 0.010198723524808884, + 0.020158467814326286, + 0.014810873195528984, + 0.0102444589138031, + 0.0171339213848114, + 0.009465407580137253, + 0.00552706653252244, + 0.00959007628262043, + 0.0060549331828951836, + 0.005502303130924702, + 0.011044926941394806, + 0.010065209120512009, + 0.005752121098339558, + 0.006006577517837286, + 0.008610011078417301, + 0.014854502864181995, + 0.02115016058087349, + 0.010720507241785526, + 0.016910098493099213, + 0.012668064795434475, + 0.009519826620817184, + 0.014394471421837807, + 0.004234085325151682, + 0.013325545005500317, + 0.0052881138399243355, + 0.0064577884040772915, + 0.005176139995455742, + 0.007177096791565418, + 0.004964129067957401, + 0.004559303168207407, + 0.009792515076696873, + 0.01597716100513935, + 0.01259684469550848, + 0.00805962085723877, + 0.019005153328180313, + 0.011897393502295017, + 0.00931844674050808, + 0.00836703460663557, + 0.006770503707230091, + 0.011419364251196384, + 0.0089073171839118, + 0.010957294143736362, + 0.010763213969767094, + 0.009348716586828232, + 0.008669114671647549, + 0.0073225609958171844, + 0.014955396763980389, + 0.017780261114239693, + 0.00895014125853777, + 0.019692452624440193, + 0.03416791558265686, + 0.013762202113866806, + 0.013795418664813042, + 0.010013142600655556, + 0.02609584666788578, + 0.025808868929743767, + 0.021503891795873642, + 0.017957525327801704, + 0.051169171929359436, + 0.0858512669801712, + 0.01726909913122654, + 0.018881386145949364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.08901601284742355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot descends slowly from the inky black sky of Iceland at midnight, capturing a magnificent spectacle of the collision of heaven and earth. Across the frozen earth, a fiery red lava flow tears through the thick ice cap with unstoppable force. Golden-orange cracks spread rapidly across the crystalline surface, instantly sending up a vast cloud of white steam. Meanwhile, a vast, emerald green aurora dances across the sky like a vast, shimmering ribbon, its mysterious pulsing glow creating a stark and stunning visual contrast to the raging fire of the earth below. The camera begins to pan forward, smoothly passing over a deep, blue ice cavern, steam rising continuously from geothermal heat. The ice walls within, illuminated by the combined effects of the lava and the aurora, reflect a hazy interplay of blue, red, and green. Suddenly, the camera pitches downward, plummeting at breakneck speed, towards the roiling lava lake below, seething like the heart of a giant beast. In extreme slow motion, as the perspective immerses, viscous liquid rock is stirred up, and countless golden hot sparks splash upwards. They draw graceful arcs in the air and slowly rise, just meeting, interweaving and merging with the emerald green aurora particles falling in the sky, and finally converge into a surreal and brilliant waterfall composed of pure light and heat. The whole scene climaxes in violent power and extreme beauty, creating a doomsday atmosphere that is both magnificent and poetic.", + "seed": 10778, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_14/index_1881_seed_10778_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04093643454834819, + "vlm_logprob": -0.04093643454834819, + "vlm_yes_no_margin": 3.9749999761581423, + "vlm_token_credit": -0.04093643454834819, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009941233322024345 + ], + "opposite_logprob_tokens": [ + -4.634941101074219 + ], + "logprob_reward": -0.009941233322024345, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009941233322024345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19606168568134308, + 0.06522801518440247, + 0.06524304300546646, + 0.12281946837902069, + 0.2121652215719223, + 0.08255386352539062, + 0.03256302326917648, + 0.08244899660348892, + 0.08978170901536942, + 0.051134973764419556 + ], + "credit_map": [ + 0.040589798241853714, + 0.05448085069656372, + 0.021277369931340218, + 0.02200976386666298, + 0.016499372199177742, + 0.01587698608636856, + 0.013667112216353416, + 0.011660427786409855, + 0.008864171802997589, + 0.006974120158702135, + 0.015301265753805637, + 0.008512263186275959, + 0.008303020149469376, + 0.008159467950463295, + 0.0053315903060138226, + 0.0037821191363036633, + 0.0064608692191541195, + 0.00775332422927022, + 0.008625986985862255, + 0.00888876337558031, + 0.010141562670469284, + 0.008674439042806625, + 0.006462717894464731, + 0.008235378190875053, + 0.009214774705469608, + 0.029187655076384544, + 0.0070691462606191635, + 0.009099334478378296, + 0.013522645458579063, + 0.03750499337911606, + 0.009033500216901302, + 0.008187427185475826, + 0.06961430609226227, + 0.08213219046592712, + 0.007870794273912907, + 0.008454963564872742, + 0.016863547265529633, + 0.010124577209353447, + 0.0092979297041893, + 0.007806916255503893, + 0.01518820971250534, + 0.020071204751729965, + 0.0047182547859847546, + 0.0045881615951657295, + 0.0118760596960783, + 0.014945169910788536, + 0.00546959089115262, + 0.005697212181985378, + 0.007087212987244129, + 0.004763093777000904, + 0.0039130584336817265, + 0.0026341944467276335, + 0.003862718353047967, + 0.003985871560871601, + 0.00370664126239717, + 0.002610231749713421, + 0.016645457595586777, + 0.017645200714468956, + 0.007137201726436615, + 0.007515309378504753, + 0.01220290083438158, + 0.010595321655273438, + 0.004618555307388306, + 0.00608905078843236, + 0.019637050107121468, + 0.036323145031929016, + 0.005028840154409409, + 0.0050888992846012115, + 0.009383654221892357, + 0.007331503089517355, + 0.003915966954082251, + 0.0030726531986147165, + 0.005105114076286554, + 0.007640609983354807, + 0.004596563521772623, + 0.003456422593444586, + 0.00988938007503748, + 0.00900301057845354, + 0.007698925212025642, + 0.003744947724044323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.009941233322024345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410274276509881 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.00410274276509881, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410274276509881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12553134560585022, + 0.2207377552986145, + 0.0786486566066742, + 0.12569604814052582, + 0.22102601826190948, + 0.058236632496118546, + 0.043688204139471054, + 0.0637459084391594, + 0.04327177256345749, + 0.01941765658557415 + ], + "credit_map": [ + 0.02338632568717003, + 0.02613474428653717, + 0.011015868745744228, + 0.01098624523729086, + 0.016603903844952583, + 0.016432834789156914, + 0.010333758778870106, + 0.010637656785547733, + 0.03225576505064964, + 0.028241435065865517, + 0.039400797337293625, + 0.02471640333533287, + 0.03360199183225632, + 0.03301670774817467, + 0.014776069670915604, + 0.014728575013577938, + 0.008333043195307255, + 0.011400881223380566, + 0.012103136628866196, + 0.007984117604792118, + 0.014099137857556343, + 0.011674219742417336, + 0.006406495813280344, + 0.006647615227848291, + 0.015033748932182789, + 0.03482310101389885, + 0.005968465935438871, + 0.0070131621323525906, + 0.012144125998020172, + 0.03185543045401573, + 0.008866594173014164, + 0.00999140553176403, + 0.0722762867808342, + 0.08769491314888, + 0.00724879652261734, + 0.007781284861266613, + 0.019194966182112694, + 0.009872390888631344, + 0.009495927020907402, + 0.007461440749466419, + 0.008897300809621811, + 0.01064268033951521, + 0.0043789236806333065, + 0.005862749181687832, + 0.00876552239060402, + 0.009155121631920338, + 0.006227429956197739, + 0.0043068998493254185, + 0.007186728995293379, + 0.004890057723969221, + 0.005799527745693922, + 0.005053355824202299, + 0.004596095997840166, + 0.004275215324014425, + 0.004801394883543253, + 0.007085820659995079, + 0.00844583846628666, + 0.00980894360691309, + 0.005222269333899021, + 0.00783685315400362, + 0.011225714348256588, + 0.0093279043212533, + 0.005659567192196846, + 0.006218813359737396, + 0.005333509296178818, + 0.006367204245179892, + 0.0038155014626681805, + 0.006069557275623083, + 0.006455831695348024, + 0.006338444538414478, + 0.004323861096054316, + 0.004567862022668123, + 0.001647003460675478, + 0.0026182178407907486, + 0.0026604083832353354, + 0.0017191312508657575, + 0.002956849755719304, + 0.00410348130390048, + 0.00202619144693017, + 0.001686371280811727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.00410274276509881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808307275176048 + ], + "opposite_logprob_tokens": [ + -3.288083076477051 + ], + "logprob_reward": -0.03808307275176048, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808307275176048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21537020802497864, + 0.0798381119966507, + 0.08568917959928513, + 0.14692798256874084, + 0.1585531234741211, + 0.06521352380514145, + 0.07477399706840515, + 0.09627995640039444, + 0.051121268421411514, + 0.026232704520225525 + ], + "credit_map": [ + 0.046951357275247574, + 0.05872369557619095, + 0.020384106785058975, + 0.032864052802324295, + 0.01579931005835533, + 0.010553251951932907, + 0.015580423176288605, + 0.014514019712805748, + 0.007647027261555195, + 0.008444196544587612, + 0.010766112245619297, + 0.012635883875191212, + 0.008281269110739231, + 0.008668078109622002, + 0.012868390418589115, + 0.010527157224714756, + 0.009401282295584679, + 0.009912891313433647, + 0.01098842266947031, + 0.01027198601514101, + 0.011831743642687798, + 0.0140623077750206, + 0.007869281806051731, + 0.01135126780718565, + 0.014135454781353474, + 0.03936898708343506, + 0.007331038825213909, + 0.008960401639342308, + 0.01491127721965313, + 0.045368172228336334, + 0.00870443880558014, + 0.008148206397891045, + 0.0539722703397274, + 0.06396189332008362, + 0.003920362796634436, + 0.0045538609847426414, + 0.014210390858352184, + 0.007127146702259779, + 0.006077475380152464, + 0.004729705862700939, + 0.01009531132876873, + 0.014203767292201519, + 0.0037827061023563147, + 0.006024861708283424, + 0.012919464148581028, + 0.007488211151212454, + 0.00667736679315567, + 0.004021832253783941, + 0.009830639697611332, + 0.010783161036670208, + 0.008269322104752064, + 0.009691915474832058, + 0.008160637691617012, + 0.007836570031940937, + 0.008444597944617271, + 0.011757154949009418, + 0.012524649500846863, + 0.01535402424633503, + 0.011780466884374619, + 0.014934631064534187, + 0.016141673550009727, + 0.009706912562251091, + 0.008060676977038383, + 0.007776923011988401, + 0.007269308436661959, + 0.007167180068790913, + 0.005676199682056904, + 0.005332286935299635, + 0.007912416011095047, + 0.008930366486310959, + 0.00434538908302784, + 0.0044881198555231094, + 0.0017262558685615659, + 0.0020829271525144577, + 0.002665946027263999, + 0.0031034923158586025, + 0.005414403975009918, + 0.0027409011963754892, + 0.006019944790750742, + 0.002478834008798003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.03808307275176048 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779990650713444 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779990650713444, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779990650713444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10275404155254364, + 0.07577832043170929, + 0.05113977566361427, + 0.08993322402238846, + 0.2606058120727539, + 0.12437702715396881, + 0.12011002004146576, + 0.08941849321126938, + 0.05104024335741997, + 0.03484297916293144 + ], + "credit_map": [ + 0.010731196962296963, + 0.015070697292685509, + 0.016729380935430527, + 0.011899341829121113, + 0.013168987818062305, + 0.013458488509058952, + 0.010649182833731174, + 0.011046769097447395, + 0.010544906370341778, + 0.008670879527926445, + 0.0098454300314188, + 0.011117814108729362, + 0.010168813169002533, + 0.00916884932667017, + 0.0059866695664823055, + 0.010274968110024929, + 0.00523789320141077, + 0.00476760184392333, + 0.005983904004096985, + 0.006694153416901827, + 0.005761818494647741, + 0.00866067036986351, + 0.006460705306380987, + 0.007573031820356846, + 0.006451300345361233, + 0.022270211949944496, + 0.003974349237978458, + 0.0049239094369113445, + 0.009178470820188522, + 0.03207354620099068, + 0.005822964012622833, + 0.005238480865955353, + 0.09196196496486664, + 0.1070757582783699, + 0.006925403606146574, + 0.0077279419638216496, + 0.023297075182199478, + 0.009562300518155098, + 0.006535579450428486, + 0.007519824430346489, + 0.02236197330057621, + 0.029525313526391983, + 0.0074305664747953415, + 0.00988183356821537, + 0.02110181376338005, + 0.011830098927021027, + 0.013301187194883823, + 0.008944248780608177, + 0.009538260288536549, + 0.013840071856975555, + 0.013846752233803272, + 0.01573551446199417, + 0.014721674844622612, + 0.01518944650888443, + 0.015431543812155724, + 0.021806765347719193, + 0.016979681327939034, + 0.01821194961667061, + 0.008985644206404686, + 0.01167330238968134, + 0.013022077269852161, + 0.008139432407915592, + 0.007107501849532127, + 0.0052989209070801735, + 0.00740929925814271, + 0.009968381375074387, + 0.005582472309470177, + 0.0041158185340464115, + 0.006361485458910465, + 0.0071567874401807785, + 0.005293598398566246, + 0.005152406636625528, + 0.002909317146986723, + 0.003786786226555705, + 0.003534108167514205, + 0.004321709740906954, + 0.006737268529832363, + 0.005450716242194176, + 0.0056327772326767445, + 0.0024702975060790777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.009779990650713444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a fiery red lava flow visible tearing through a thick ice cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast emerald green aurora visible dancing across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin by panning forward over a deep blue ice cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pitch downward at breakneck speed towards a roiling lava lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden hot sparks shown rising in extreme slow motion to merge with the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277513325214386 + ], + "opposite_logprob_tokens": [ + -2.017775058746338 + ], + "logprob_reward": -0.14277513325214386, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14277513325214386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16273869574069977, + 0.10346318036317825, + 0.05248676985502243, + 0.10152741521596909, + 0.11053378880023956, + 0.11075492948293686, + 0.05809152498841286, + 0.1360294222831726, + 0.11046309024095535, + 0.053911179304122925 + ], + "credit_map": [ + 0.029407091438770294, + 0.037438858300447464, + 0.015249239280819893, + 0.022530604153871536, + 0.014869628474116325, + 0.016509192064404488, + 0.013346790336072445, + 0.013387300074100494, + 0.015166645869612694, + 0.012170640751719475, + 0.01964089460670948, + 0.012519260868430138, + 0.013938982971012592, + 0.011714817956089973, + 0.00821886956691742, + 0.010093068704009056, + 0.005678119603544474, + 0.006926957052201033, + 0.005757926031947136, + 0.005054393783211708, + 0.010997415520250797, + 0.009479900822043419, + 0.004286589566618204, + 0.004305469803512096, + 0.009752023033797741, + 0.03284797444939613, + 0.0034877529833465815, + 0.004075752105563879, + 0.009499659761786461, + 0.033385682851076126, + 0.004480828531086445, + 0.003997742664068937, + 0.03635387867689133, + 0.04005862772464752, + 0.005501901265233755, + 0.0038274279795587063, + 0.010037802159786224, + 0.0053517743945121765, + 0.004908071365207434, + 0.004494301043450832, + 0.01654859445989132, + 0.023088397458195686, + 0.008086775429546833, + 0.009851666167378426, + 0.02173183672130108, + 0.012770420871675014, + 0.012369382195174694, + 0.00630785571411252, + 0.007740580476820469, + 0.009151997976005077, + 0.006501592230051756, + 0.0071616643108427525, + 0.005646435543894768, + 0.004794185049831867, + 0.007365580182522535, + 0.009729491546750069, + 0.02092953957617283, + 0.023505734279751778, + 0.014704199507832527, + 0.012239878997206688, + 0.028855977579951286, + 0.016403861343860626, + 0.010861123912036419, + 0.008529111742973328, + 0.01416164543479681, + 0.01105723436921835, + 0.007590566296130419, + 0.01119064912199974, + 0.02046828903257847, + 0.02190149948000908, + 0.012042369693517685, + 0.012050834484398365, + 0.0051520285196602345, + 0.006052757613360882, + 0.008876233361661434, + 0.006271346937865019, + 0.006669026333838701, + 0.007844468578696251, + 0.006031010299921036, + 0.0070143090561032295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.14277513325214386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_15/index_tensor([1672, 1672])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_15/index_tensor([1672, 1672])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..52103f8666e2ead134b57e334b8b58667a0b621b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_15/index_tensor([1672, 1672])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a horizontal angle, slowly tilts downward, following a suit of ancient, battle-worn steel armor as it descends with a solemn, slow descent into the dark, murky waters. The armor consists of a closed helmet with narrow eye slits, a sturdy breastplate, and tightly connected arm and leg armor. Its dull metal surface is covered with countless scratches and dents of varying depths, silently bearing witness to the fierce battles of the past. Mottled rust seeps through the seams. The only source of light comes from far above the water, casting a faint, flickering, cold beam. The light penetrates the deep blue-black water, revealing countless tiny suspended particles drifting slowly within. As this faint light sweeps across the armor's metal surface, it briefly illuminates the ancient scars. The most eye-catching thing is that a series of fine and continuous silver bubbles are rushing out from the narrow gaps in the helmet's eyelets, forming a clear trajectory and floating towards the unreachable light source above, forming a sharp contrast with the fate of the heavy falling armor. The whole picture is filled with a quiet, lonely and slightly tragic atmosphere, as if witnessing the end of a heroic era.", + "seed": 23930, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_15/index_1672_seed_23930_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03136343229562044, + "vlm_logprob": -0.03136343229562044, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03136343229562044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025800099596381187 + ], + "opposite_logprob_tokens": [ + -3.7757999897003174 + ], + "logprob_reward": -0.025800099596381187, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025800099596381187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1021478995680809, + 0.07045150548219681, + 0.04432293772697449, + 0.1183978021144867, + 0.04981952905654907, + 0.048222485929727554, + 0.05954667553305626, + 0.06994196027517319, + 0.13901233673095703, + 0.29813680052757263 + ], + "credit_map": [ + 0.017859773710370064, + 0.009467532858252525, + 0.016110481694340706, + 0.012198735028505325, + 0.006899269763380289, + 0.011527983471751213, + 0.014382435940206051, + 0.013701685704290867, + 0.011807942762970924, + 0.008445891551673412, + 0.010315095074474812, + 0.007303178776055574, + 0.007604797836393118, + 0.00695849908515811, + 0.009618797339498997, + 0.008397303521633148, + 0.00675328029319644, + 0.005895528011023998, + 0.006369652692228556, + 0.005938442889600992, + 0.004307649563997984, + 0.0061117964796721935, + 0.005329936742782593, + 0.003616650588810444, + 0.01618674397468567, + 0.016284270212054253, + 0.016657939180731773, + 0.01650083065032959, + 0.009432341903448105, + 0.01630125567317009, + 0.01746399700641632, + 0.009570417925715446, + 0.006290316581726074, + 0.005694856401532888, + 0.005069537088274956, + 0.008651494048535824, + 0.005404123570770025, + 0.0056884209625422955, + 0.009355391375720501, + 0.0036653922870755196, + 0.00899624265730381, + 0.005306671839207411, + 0.007190876640379429, + 0.005907405633479357, + 0.004589797928929329, + 0.0071377563290297985, + 0.005980614107102156, + 0.003113121259957552, + 0.006478346884250641, + 0.011919641867280006, + 0.005690321791917086, + 0.008185646496713161, + 0.008671029470860958, + 0.009119393303990364, + 0.005956496577709913, + 0.0035257970448583364, + 0.013213851489126682, + 0.0080966642126441, + 0.00957818329334259, + 0.007129569537937641, + 0.009249801747500896, + 0.012157299555838108, + 0.005490479525178671, + 0.005026109982281923, + 0.020800750702619553, + 0.02412516251206398, + 0.017716005444526672, + 0.016497742384672165, + 0.013308458961546421, + 0.01788054034113884, + 0.013073746114969254, + 0.015609923750162125, + 0.05689495801925659, + 0.03449103608727455, + 0.028622105717658997, + 0.024873146787285805, + 0.051203567534685135, + 0.041156377643346786, + 0.02071024477481842, + 0.04018537700176239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.025800099596381187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108560271561146 + ], + "opposite_logprob_tokens": [ + -5.504108428955078 + ], + "logprob_reward": -0.004108560271561146, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108560271561146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14189335703849792, + 0.10757344961166382, + 0.06280628591775894, + 0.058214250952005386, + 0.05233113467693329, + 0.07006096839904785, + 0.08076531440019608, + 0.06097380444407463, + 0.09755316376686096, + 0.2678282558917999 + ], + "credit_map": [ + 0.02476505935192108, + 0.0166179072111845, + 0.01941530965268612, + 0.016880741342902184, + 0.013349928893148899, + 0.015671217814087868, + 0.017825258895754814, + 0.017367934808135033, + 0.023263294249773026, + 0.011216388083994389, + 0.01569495163857937, + 0.010913686826825142, + 0.009756868705153465, + 0.011731213890016079, + 0.013750556856393814, + 0.011246479116380215, + 0.009087979793548584, + 0.008221469819545746, + 0.008955854922533035, + 0.008345700800418854, + 0.005628903862088919, + 0.01046502310782671, + 0.007394492160528898, + 0.004706863313913345, + 0.007154917810112238, + 0.0037774634547531605, + 0.01354347262531519, + 0.006543588824570179, + 0.0030466592870652676, + 0.01614489033818245, + 0.00393700459972024, + 0.004066250286996365, + 0.006247656885534525, + 0.006057882681488991, + 0.006570815108716488, + 0.008670464158058167, + 0.004936713259667158, + 0.006012763828039169, + 0.01044825091958046, + 0.0033865829464048147, + 0.013791258446872234, + 0.009649383835494518, + 0.010951628908514977, + 0.007670956198126078, + 0.005623308010399342, + 0.00975019671022892, + 0.007832931354641914, + 0.004791307728737593, + 0.007728966418653727, + 0.01291723269969225, + 0.010069978423416615, + 0.013372451066970825, + 0.00984475389122963, + 0.011399896815419197, + 0.009828716516494751, + 0.005603315308690071, + 0.010662062093615532, + 0.006500624120235443, + 0.01036305446177721, + 0.006060308311134577, + 0.0051631745882332325, + 0.009331153705716133, + 0.0060517871752381325, + 0.006841636728495359, + 0.014658723026514053, + 0.01546699833124876, + 0.012271763756871223, + 0.01113863941282034, + 0.010373411700129509, + 0.011371707543730736, + 0.00884152203798294, + 0.013430386781692505, + 0.050019681453704834, + 0.03806526958942413, + 0.020447704941034317, + 0.026704031974077225, + 0.04557555168867111, + 0.04286915436387062, + 0.02122659794986248, + 0.0229202751070261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.004108560271561146 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048661477863788605 + ], + "opposite_logprob_tokens": [ + -3.048661470413208 + ], + "logprob_reward": -0.048661477863788605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048661477863788605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11907840520143509, + 0.07205083966255188, + 0.04872690513730049, + 0.22079181671142578, + 0.08997680246829987, + 0.11699081212282181, + 0.06256910413503647, + 0.03899955749511719, + 0.08844666928052902, + 0.14236903190612793 + ], + "credit_map": [ + 0.01814422756433487, + 0.014391683042049408, + 0.017725465819239616, + 0.013570410199463367, + 0.010011216625571251, + 0.013362796977162361, + 0.01564665138721466, + 0.016225948929786682, + 0.013946066610515118, + 0.00739434314891696, + 0.011468599550426006, + 0.008320274762809277, + 0.006832472514361143, + 0.007182778790593147, + 0.010102950967848301, + 0.0068033491261303425, + 0.006697637028992176, + 0.005451223347336054, + 0.010480329394340515, + 0.005561962258070707, + 0.004830951802432537, + 0.006944181397557259, + 0.005057071335613728, + 0.0037035506684333086, + 0.02903890796005726, + 0.02759302407503128, + 0.032675959169864655, + 0.03066249005496502, + 0.016825418919324875, + 0.03590139374136925, + 0.029662739485502243, + 0.018431901931762695, + 0.009687904268503189, + 0.010081068612635136, + 0.01124162133783102, + 0.01663259230554104, + 0.009516341611742973, + 0.010008947923779488, + 0.016346512362360954, + 0.006461817305535078, + 0.0210895873606205, + 0.013552073389291763, + 0.02063339576125145, + 0.010372322984039783, + 0.009846888482570648, + 0.01816299743950367, + 0.014139093458652496, + 0.009194447658956051, + 0.005704783834517002, + 0.009095221757888794, + 0.006387787405401468, + 0.009180240333080292, + 0.008937004953622818, + 0.009541923180222511, + 0.007562802638858557, + 0.006159346550703049, + 0.004818157758563757, + 0.0036842261906713247, + 0.004250766709446907, + 0.004476276692003012, + 0.00872547086328268, + 0.00519562978297472, + 0.004602798726409674, + 0.0032462312374264, + 0.0113590769469738, + 0.015067842788994312, + 0.011766504496335983, + 0.011560509912669659, + 0.008456406183540821, + 0.011825410649180412, + 0.008611157536506653, + 0.009799765422940254, + 0.027885962277650833, + 0.018934231251478195, + 0.01218586228787899, + 0.013283169828355312, + 0.023599831387400627, + 0.020621158182621002, + 0.011327004991471767, + 0.01453181728720665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.048661477863788605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496174469590187 + ], + "opposite_logprob_tokens": [ + -2.929961681365967 + ], + "logprob_reward": -0.05496174469590187, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496174469590187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1688930094242096, + 0.08267214149236679, + 0.06033199280500412, + 0.0816282108426094, + 0.05690787360072136, + 0.10912758857011795, + 0.17143017053604126, + 0.07765647023916245, + 0.07401951402425766, + 0.11733300983905792 + ], + "credit_map": [ + 0.04065213352441788, + 0.016599932685494423, + 0.02227058820426464, + 0.021322522312402725, + 0.014372874051332474, + 0.021009773015975952, + 0.016835706308484077, + 0.015829481184482574, + 0.029801638796925545, + 0.007357821799814701, + 0.014942599460482597, + 0.006080255378037691, + 0.004153321497142315, + 0.010103045031428337, + 0.004304175730794668, + 0.005929281935095787, + 0.008938445709645748, + 0.01067917700856924, + 0.00823560357093811, + 0.007697542198002338, + 0.006391399074345827, + 0.005842706188559532, + 0.007127108983695507, + 0.005420004483312368, + 0.012894585728645325, + 0.009790254756808281, + 0.013806842267513275, + 0.010600589215755463, + 0.006328869145363569, + 0.011337938718497753, + 0.009687934070825577, + 0.007181193679571152, + 0.008400210179388523, + 0.008996814489364624, + 0.007179315201938152, + 0.006449554115533829, + 0.004071677103638649, + 0.008457310497760773, + 0.009290846064686775, + 0.0040621464140713215, + 0.015100996010005474, + 0.01204494759440422, + 0.022184882313013077, + 0.009585591033101082, + 0.0093615697696805, + 0.017230592668056488, + 0.014008695259690285, + 0.009610313922166824, + 0.018023855984210968, + 0.020074667409062386, + 0.019839581102132797, + 0.016837382689118385, + 0.022546621039509773, + 0.024463677778840065, + 0.025574399158358574, + 0.024069977924227715, + 0.009679547511041164, + 0.012169435620307922, + 0.009869051165878773, + 0.009240164421498775, + 0.01261140312999487, + 0.009477304294705391, + 0.008539856411516666, + 0.006069708149880171, + 0.009822245687246323, + 0.011622072197496891, + 0.010183137841522694, + 0.010988498106598854, + 0.007225118111819029, + 0.009010623209178448, + 0.007635259535163641, + 0.007532557472586632, + 0.021329976618289948, + 0.016927288845181465, + 0.007710549980401993, + 0.014059413224458694, + 0.01789465732872486, + 0.014498610980808735, + 0.009462847374379635, + 0.015449662692844868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.05496174469590187 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232852790504694 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.0232852790504694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232852790504694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.067193403840065, + 0.14275571703910828, + 0.15960979461669922, + 0.11329665780067444, + 0.03263700753450394, + 0.06953462213277817, + 0.1183968260884285, + 0.07564789056777954, + 0.18238866329193115, + 0.038539398461580276 + ], + "credit_map": [ + 0.009757664054632187, + 0.006340072024613619, + 0.009091832675039768, + 0.005662570241838694, + 0.0046960473991930485, + 0.008215107023715973, + 0.010808260180056095, + 0.012621847912669182, + 0.051567260175943375, + 0.013049598783254623, + 0.02861536107957363, + 0.008203054778277874, + 0.006601192522794008, + 0.015970362350344658, + 0.008045117370784283, + 0.010703779757022858, + 0.02485264278948307, + 0.016484610736370087, + 0.030893247574567795, + 0.021109746769070625, + 0.013465630821883678, + 0.02454579621553421, + 0.016574975103139877, + 0.011683141812682152, + 0.01819698140025139, + 0.010209971107542515, + 0.020270373672246933, + 0.01422901451587677, + 0.00591773446649313, + 0.025860507041215897, + 0.010444729588925838, + 0.008167344145476818, + 0.003593727480620146, + 0.004571050871163607, + 0.0034975335001945496, + 0.0047713457606732845, + 0.002744869329035282, + 0.004263266455382109, + 0.006923798006027937, + 0.0022714154329150915, + 0.010134809650480747, + 0.0061265709809958935, + 0.012534336186945438, + 0.00785601232200861, + 0.006646580528467894, + 0.009665783494710922, + 0.009623762220144272, + 0.006946766749024391, + 0.013690484687685966, + 0.01822623237967491, + 0.015493965707719326, + 0.01583094149827957, + 0.01066786702722311, + 0.01656845025718212, + 0.013873905874788761, + 0.014044982381165028, + 0.00860513560473919, + 0.007988184690475464, + 0.010374871082603931, + 0.009155862964689732, + 0.007179060485213995, + 0.00941519346088171, + 0.014365038834512234, + 0.008564537391066551, + 0.024004049599170685, + 0.0246493648737669, + 0.027356648817658424, + 0.02950235642492771, + 0.016410548239946365, + 0.01955004408955574, + 0.022502901032567024, + 0.018412750214338303, + 0.007178395986557007, + 0.008356520906090736, + 0.002998921787366271, + 0.0036482270807027817, + 0.005200446583330631, + 0.005767021328210831, + 0.00259353662841022, + 0.002796328393742442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.0232852790504694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a horizontal angle, slowly tilts downward, following a suit of ancient, battle-worn steel armor as it descends with a solemn, slow descent into the dark, murky waters. The armor consists of a closed helmet with narrow eye slits, a sturdy breastplate, and tightly connected arm and leg armor. Its dull metal surface is covered with countless scratches and dents of varying depths, silently bearing witness to the fierce battles of the past. Mottled rust seeps through the seams. The only source of light comes from far above the water, casting a faint, flickering, cold beam. The light penetrates the deep blue-black water, revealing countless tiny suspended particles drifting slowly within. As this faint light sweeps across the armor's metal surface, it briefly illuminates the ancient scars. The most eye-catching thing is that a series of fine and continuous silver bubbles are rushing out from the narrow gaps in the helmet's eyelets, forming a clear trajectory and floating towards the unreachable light source above, forming a sharp contrast with the fate of the heavy falling armor. The whole picture is filled with a quiet, lonely and slightly tragic atmosphere, as if witnessing the end of a heroic era.", + "seed": 23931, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_15/index_1672_seed_23931_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07068743836134672, + "vlm_logprob": -0.07068743836134672, + "vlm_yes_no_margin": 2.950000023841858, + "vlm_token_credit": -0.07068743836134672, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04555739089846611 + ], + "opposite_logprob_tokens": [ + -3.1705574989318848 + ], + "logprob_reward": -0.04555739089846611, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04555739089846611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043755561113357544, + 0.05391699820756912, + 0.035919345915317535, + 0.01660127006471157, + 0.015137949027121067, + 0.05695075914263725, + 0.06538274884223938, + 0.05590521916747093, + 0.5943183898925781, + 0.062111806124448776 + ], + "credit_map": [ + 0.0036032225470989943, + 0.005475127138197422, + 0.007128837052732706, + 0.0028669461607933044, + 0.003923490643501282, + 0.006557371001690626, + 0.002982614329084754, + 0.011217949911952019, + 0.007245561107993126, + 0.006423200014978647, + 0.00550270127132535, + 0.007085585035383701, + 0.0070861163549125195, + 0.009870839305222034, + 0.004625148605555296, + 0.006077844649553299, + 0.003518566256389022, + 0.0036530205979943275, + 0.0033329578582197428, + 0.006645679008215666, + 0.004967133980244398, + 0.0051824008114635944, + 0.004358166828751564, + 0.0042614201083779335, + 0.0019428875530138612, + 0.0020990371704101562, + 0.0016442202031612396, + 0.00234364066272974, + 0.002134432550519705, + 0.0029437176417559385, + 0.001610150677151978, + 0.001883183722384274, + 0.0019177838694304228, + 0.002035721903666854, + 0.0017488240264356136, + 0.0016104963142424822, + 0.002034829929471016, + 0.0021670800633728504, + 0.0018504386534914374, + 0.0017727733356878161, + 0.0031412665266543627, + 0.005072564817965031, + 0.017269693315029144, + 0.001922635710798204, + 0.003991421777755022, + 0.0201265886425972, + 0.0020659291185438633, + 0.003360654693096876, + 0.011432456783950329, + 0.009278656914830208, + 0.006842844653874636, + 0.004575150087475777, + 0.00716883409768343, + 0.01114623248577118, + 0.006617013365030289, + 0.008321553468704224, + 0.006463473197072744, + 0.00623328099027276, + 0.008611857891082764, + 0.0055048358626663685, + 0.007744383532553911, + 0.009115500375628471, + 0.007718515116721392, + 0.0045133656822144985, + 0.08045963943004608, + 0.09321259707212448, + 0.10891776531934738, + 0.03151558339595795, + 0.02964545227587223, + 0.16543498635292053, + 0.04352180287241936, + 0.041610538959503174, + 0.0030920503195375204, + 0.0037863687612116337, + 0.02201147936284542, + 0.0022128743585199118, + 0.002591207856312394, + 0.02388523705303669, + 0.002857751678675413, + 0.0016748352209106088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.04555739089846611 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798760525882244 + ], + "opposite_logprob_tokens": [ + -4.634798526763916 + ], + "logprob_reward": -0.009798760525882244, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798760525882244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18794460594654083, + 0.0514034777879715, + 0.046794723719358444, + 0.02634485624730587, + 0.026415329426527023, + 0.07602560520172119, + 0.047361135482788086, + 0.09488364309072495, + 0.28077200055122375, + 0.162054643034935 + ], + "credit_map": [ + 0.012607251293957233, + 0.007077591028064489, + 0.04016711935400963, + 0.0044168466702103615, + 0.008525990881025791, + 0.043297912925481796, + 0.004676561336964369, + 0.06717531383037567, + 0.005961890332400799, + 0.007740742526948452, + 0.004872704856097698, + 0.006251363083720207, + 0.008000976406037807, + 0.008953608572483063, + 0.004278065171092749, + 0.00534411845728755, + 0.005415701773017645, + 0.006659814156591892, + 0.004251479171216488, + 0.0074825831688940525, + 0.0076135024428367615, + 0.005062200129032135, + 0.005889449268579483, + 0.004419988952577114, + 0.002522278344258666, + 0.0024471611250191927, + 0.002717234194278717, + 0.003419018117710948, + 0.003276496659964323, + 0.0058038560673594475, + 0.003314490430057049, + 0.0028443194460123777, + 0.0033461449202150106, + 0.003272867761552334, + 0.0028192372992634773, + 0.0029328572563827038, + 0.0037425714544951916, + 0.003738279454410076, + 0.0034629961010068655, + 0.0031003726180642843, + 0.004817361012101173, + 0.005987684708088636, + 0.02278173342347145, + 0.0028681051917374134, + 0.004832299426198006, + 0.025185639038681984, + 0.003861662233248353, + 0.005691113416105509, + 0.006582504604011774, + 0.005983137059956789, + 0.004626707173883915, + 0.00487680546939373, + 0.00588019285351038, + 0.007697362918406725, + 0.006148341577500105, + 0.005566081032156944, + 0.011912548914551735, + 0.012079661712050438, + 0.014034273102879524, + 0.009179782122373581, + 0.010390046052634716, + 0.018325116485357285, + 0.01045781746506691, + 0.008504385128617287, + 0.03802221640944481, + 0.044508058577775955, + 0.050403937697410583, + 0.01497623696923256, + 0.01416732370853424, + 0.07779443264007568, + 0.021113403141498566, + 0.019786395132541656, + 0.007600007578730583, + 0.007935470901429653, + 0.06339064240455627, + 0.004191278479993343, + 0.0038454337045550346, + 0.06658242642879486, + 0.004456676077097654, + 0.004052691161632538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.009798760525882244 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034029930830002 + ], + "opposite_logprob_tokens": [ + -2.3503403663635254 + ], + "logprob_reward": -0.10034029930830002, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034029930830002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05184117332100868, + 0.0795169547200203, + 0.06394871324300766, + 0.04691809043288231, + 0.043942518532276154, + 0.1519211232662201, + 0.06542973220348358, + 0.09303198754787445, + 0.34119218587875366, + 0.062257517129182816 + ], + "credit_map": [ + 0.00442658131942153, + 0.005169281270354986, + 0.009530721232295036, + 0.0038563194684684277, + 0.005154946818947792, + 0.007145570125430822, + 0.0049703847616910934, + 0.011587370187044144, + 0.01024232804775238, + 0.011507484130561352, + 0.008328648284077644, + 0.008038438856601715, + 0.010913534089922905, + 0.014259069226682186, + 0.0066430228762328625, + 0.009584428742527962, + 0.00650645699352026, + 0.0071355076506733894, + 0.005635804962366819, + 0.0113863255828619, + 0.010194528847932816, + 0.008846358396112919, + 0.008689665235579014, + 0.005554066505283117, + 0.004549091216176748, + 0.004931868985295296, + 0.004257457330822945, + 0.006542704999446869, + 0.00605445820838213, + 0.00998770259320736, + 0.005340813659131527, + 0.00525399437174201, + 0.0064084636978805065, + 0.0062241023406386375, + 0.0060026017017662525, + 0.003501848317682743, + 0.004196197260171175, + 0.008149580098688602, + 0.0040122466161847115, + 0.00544747756794095, + 0.007966128177940845, + 0.013043669983744621, + 0.04993179067969322, + 0.004092363175004721, + 0.0058477334678173065, + 0.05981681868433952, + 0.0025418377481400967, + 0.008680799044668674, + 0.01018468290567398, + 0.008128833025693893, + 0.009339484386146069, + 0.004804891999810934, + 0.00676351273432374, + 0.01098594069480896, + 0.00723797595128417, + 0.007984412834048271, + 0.01156612765043974, + 0.012197139672935009, + 0.016261396929621696, + 0.006435025483369827, + 0.007538474630564451, + 0.02209986187517643, + 0.008343072608113289, + 0.008590886369347572, + 0.045938871800899506, + 0.05296727642416954, + 0.06377328932285309, + 0.01822359673678875, + 0.016927413642406464, + 0.09355747699737549, + 0.025537488982081413, + 0.02426677756011486, + 0.005357297603040934, + 0.007117327768355608, + 0.014016706496477127, + 0.00510407192632556, + 0.004312149249017239, + 0.012857209891080856, + 0.009421095252037048, + 0.004071660805493593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.10034029930830002 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054962195456027985 + ], + "opposite_logprob_tokens": [ + -2.929962158203125 + ], + "logprob_reward": -0.054962195456027985, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054962195456027985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04290589690208435, + 0.05836214870214462, + 0.06642358005046844, + 0.021852828562259674, + 0.030026862397789955, + 0.04716699942946434, + 0.08738072216510773, + 0.14841598272323608, + 0.4131024479866028, + 0.08436252176761627 + ], + "credit_map": [ + 0.005414359737187624, + 0.00424547353759408, + 0.007372033782303333, + 0.0029766173101961613, + 0.004402920603752136, + 0.005694927182048559, + 0.004005103372037411, + 0.008794461376965046, + 0.006892351433634758, + 0.008762612007558346, + 0.005538661032915115, + 0.006910238415002823, + 0.008633028715848923, + 0.009936223737895489, + 0.005184460431337357, + 0.006504572927951813, + 0.006871236953884363, + 0.0077962069772183895, + 0.005948233883827925, + 0.009637027978897095, + 0.010112322866916656, + 0.012410764582455158, + 0.006345072295516729, + 0.007302708458155394, + 0.002044247230514884, + 0.0025613962206989527, + 0.0023319090250879526, + 0.0032859682105481625, + 0.003540750127285719, + 0.003068159567192197, + 0.002537771826609969, + 0.0024826261214911938, + 0.004299338907003403, + 0.004643974360078573, + 0.002936862176284194, + 0.003796434262767434, + 0.004774994216859341, + 0.003369830781593919, + 0.0034559648483991623, + 0.002749465638771653, + 0.0035011370200663805, + 0.004050741903483868, + 0.011297966353595257, + 0.0035540671087801456, + 0.004964533261954784, + 0.011891788803040981, + 0.0036120626609772444, + 0.004294699989259243, + 0.012323948554694653, + 0.012261481955647469, + 0.009471846744418144, + 0.009238900616765022, + 0.013965960592031479, + 0.010396912693977356, + 0.010865612886846066, + 0.008856060914695263, + 0.02117595076560974, + 0.021366963163018227, + 0.022641371935606003, + 0.009497207589447498, + 0.008756781928241253, + 0.037142910063266754, + 0.012348233722150326, + 0.015486565418541431, + 0.054276157170534134, + 0.06490680575370789, + 0.07708372920751572, + 0.021224811673164368, + 0.02030116319656372, + 0.1159636378288269, + 0.031909506767988205, + 0.027436628937721252, + 0.004643143154680729, + 0.006038721185177565, + 0.027839725837111473, + 0.0031887413933873177, + 0.005686935968697071, + 0.02893410436809063, + 0.004698042292147875, + 0.0033331098966300488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.054962195456027985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277854561805725 + ], + "opposite_logprob_tokens": [ + -2.0177786350250244 + ], + "logprob_reward": -0.14277854561805725, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14277854561805725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038550231605768204, + 0.007908192463219166, + 0.01730506308376789, + 0.01757178083062172, + 0.007350935600697994, + 0.014131798408925533, + 0.058524735271930695, + 0.041111983358860016, + 0.717419445514679, + 0.08012580871582031 + ], + "credit_map": [ + 0.0031019053421914577, + 0.0012540403986349702, + 0.0077176629565656185, + 0.0007964767282828689, + 0.001831025700084865, + 0.009428867138922215, + 0.0009055489208549261, + 0.013514705002307892, + 0.0011893148766830564, + 0.0009809271432459354, + 0.0009017312549985945, + 0.0007230814662761986, + 0.0010574290063232183, + 0.0012742637190967798, + 0.0006642348016612232, + 0.0011172109516337514, + 0.0017228710930794477, + 0.0020759510807693005, + 0.0014908286975696683, + 0.0028317158576101065, + 0.002647834364324808, + 0.00287733250297606, + 0.001910148886963725, + 0.0017483814153820276, + 0.0016761264996603131, + 0.0019003137713298202, + 0.0014744658255949616, + 0.0021587335504591465, + 0.002070441609248519, + 0.004596914164721966, + 0.0020437880884855986, + 0.0016509955748915672, + 0.0007677106186747551, + 0.0014619590947404504, + 0.0006594055448658764, + 0.0009142335038632154, + 0.0009526397916488349, + 0.000939977471716702, + 0.000842815381474793, + 0.0008121944847516716, + 0.001337248831987381, + 0.0020999314729124308, + 0.0017713764682412148, + 0.0013582127867266536, + 0.0023909560404717922, + 0.0025482620112597942, + 0.0016596291679888964, + 0.0009661819785833359, + 0.010281101800501347, + 0.008181778714060783, + 0.006349945906549692, + 0.0039197299629449844, + 0.005848775617778301, + 0.009984358213841915, + 0.0062895892187952995, + 0.007669458631426096, + 0.00466286251321435, + 0.005575636867433786, + 0.005290478467941284, + 0.0038243685849010944, + 0.005050265695899725, + 0.007990448735654354, + 0.004952012561261654, + 0.0037659108638763428, + 0.09627629816532135, + 0.11254469305276871, + 0.13307151198387146, + 0.03789249807596207, + 0.035590481013059616, + 0.19916483759880066, + 0.053335946053266525, + 0.04954317957162857, + 0.0037323737051337957, + 0.004262331873178482, + 0.03183407336473465, + 0.001735979923978448, + 0.0020304264035075903, + 0.03364454582333565, + 0.0012469657231122255, + 0.00163911166600883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.14277854561805725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a horizontal angle, slowly tilts downward, following a suit of ancient, battle-worn steel armor as it descends with a solemn, slow descent into the dark, murky waters. The armor consists of a closed helmet with narrow eye slits, a sturdy breastplate, and tightly connected arm and leg armor. Its dull metal surface is covered with countless scratches and dents of varying depths, silently bearing witness to the fierce battles of the past. Mottled rust seeps through the seams. The only source of light comes from far above the water, casting a faint, flickering, cold beam. The light penetrates the deep blue-black water, revealing countless tiny suspended particles drifting slowly within. As this faint light sweeps across the armor's metal surface, it briefly illuminates the ancient scars. The most eye-catching thing is that a series of fine and continuous silver bubbles are rushing out from the narrow gaps in the helmet's eyelets, forming a clear trajectory and floating towards the unreachable light source above, forming a sharp contrast with the fate of the heavy falling armor. The whole picture is filled with a quiet, lonely and slightly tragic atmosphere, as if witnessing the end of a heroic era.", + "seed": 23932, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_15/index_1672_seed_23932_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08025625310838222, + "vlm_logprob": -0.08025625310838222, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.08025625310838222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027407221496105194 + ], + "opposite_logprob_tokens": [ + -3.777407169342041 + ], + "logprob_reward": -0.027407221496105194, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027407221496105194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09926346689462662, + 0.2251245379447937, + 0.10235103219747543, + 0.03558134287595749, + 0.04155639186501503, + 0.07944571226835251, + 0.0813525840640068, + 0.11197537183761597, + 0.09613551944494247, + 0.12721405923366547 + ], + "credit_map": [ + 0.014600440859794617, + 0.014905083924531937, + 0.011462358757853508, + 0.014921829104423523, + 0.011985987424850464, + 0.0113761555403471, + 0.008971820585429668, + 0.011039791628718376, + 0.025409305468201637, + 0.028598014265298843, + 0.031016217544674873, + 0.024450339376926422, + 0.03211187198758125, + 0.043562132865190506, + 0.0178510881960392, + 0.02212556079030037, + 0.015211324207484722, + 0.015686068683862686, + 0.013627984561026096, + 0.011405956000089645, + 0.011092845350503922, + 0.01659420318901539, + 0.00953572615981102, + 0.00919691938906908, + 0.004142544697970152, + 0.004646717570722103, + 0.0033746499102562666, + 0.005336269270628691, + 0.004689272493124008, + 0.00419615488499403, + 0.0062914518639445305, + 0.002904279623180628, + 0.005148470867425203, + 0.005757915321737528, + 0.004096632823348045, + 0.005851228255778551, + 0.006253151223063469, + 0.0039726789109408855, + 0.004708793479949236, + 0.0057675219140946865, + 0.006822377908974886, + 0.013017998076975346, + 0.008494661189615726, + 0.00868938583880663, + 0.010805269703269005, + 0.013452696613967419, + 0.010130052454769611, + 0.008033271878957748, + 0.011252057738602161, + 0.013622251339256763, + 0.009975395165383816, + 0.011261900886893272, + 0.006664562039077282, + 0.009482317604124546, + 0.009868673980236053, + 0.009225429967045784, + 0.015124833211302757, + 0.022868916392326355, + 0.011415554210543633, + 0.009152545593678951, + 0.010436991229653358, + 0.020296117290854454, + 0.009562171064317226, + 0.013118254952132702, + 0.00856227520853281, + 0.022102288901805878, + 0.008642444387078285, + 0.008625003509223461, + 0.016052981838583946, + 0.015161875635385513, + 0.0072631752118468285, + 0.009725471958518028, + 0.014789356850087643, + 0.010876939631998539, + 0.02267911098897457, + 0.011508318595588207, + 0.013204674236476421, + 0.027409007772803307, + 0.012402319349348545, + 0.014344333671033382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.027407221496105194 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865730457007885 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.00865730457007885, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865730457007885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14361459016799927, + 0.08807501196861267, + 0.13884498178958893, + 0.058124441653490067, + 0.03088424913585186, + 0.16318921744823456, + 0.05404934287071228, + 0.0529426746070385, + 0.12062765657901764, + 0.14964786171913147 + ], + "credit_map": [ + 0.01986033283174038, + 0.03265789896249771, + 0.013226408511400223, + 0.021656526252627373, + 0.014919695444405079, + 0.015774155035614967, + 0.011362661607563496, + 0.014156904071569443, + 0.011949457228183746, + 0.012377376668155193, + 0.015313663519918919, + 0.008514913730323315, + 0.011236041784286499, + 0.013838564977049828, + 0.005651384126394987, + 0.00919361226260662, + 0.02035125344991684, + 0.020821979269385338, + 0.019574137404561043, + 0.015450040809810162, + 0.017144009470939636, + 0.01849827915430069, + 0.015216097235679626, + 0.01178918220102787, + 0.0065293386578559875, + 0.00659811170771718, + 0.005801645573228598, + 0.007750141434371471, + 0.009336896240711212, + 0.008571146987378597, + 0.008888602256774902, + 0.004648557398468256, + 0.005613365210592747, + 0.0048775398172438145, + 0.003657132387161255, + 0.004036815837025642, + 0.002935882890596986, + 0.0032011037692427635, + 0.003255903022363782, + 0.003306506434455514, + 0.025726577267050743, + 0.019458888098597527, + 0.020540324971079826, + 0.023276254534721375, + 0.016603708267211914, + 0.02685290388762951, + 0.012850591912865639, + 0.017879975959658623, + 0.007195045240223408, + 0.01043738890439272, + 0.005819791462272406, + 0.0062151113525033, + 0.004510624799877405, + 0.00942844245582819, + 0.004432336427271366, + 0.006010602694004774, + 0.004084065556526184, + 0.010706847533583641, + 0.005480356980115175, + 0.005425174720585346, + 0.006807571277022362, + 0.0060517131350934505, + 0.007025551050901413, + 0.007361395284533501, + 0.009337294846773148, + 0.022626124322414398, + 0.013382441364228725, + 0.011644828133285046, + 0.020037811249494553, + 0.024689679965376854, + 0.007271165493875742, + 0.01163832563906908, + 0.018878228962421417, + 0.015882713720202446, + 0.02124299854040146, + 0.01057050097733736, + 0.01675085350871086, + 0.03338588774204254, + 0.010773785412311554, + 0.02216288447380066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.00865730457007885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309936612844467 + ], + "opposite_logprob_tokens": [ + -3.1680994033813477 + ], + "logprob_reward": -0.04309936612844467, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309936612844467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13722741603851318, + 0.06745242327451706, + 0.08260645717382431, + 0.026241935789585114, + 0.03902623802423477, + 0.10429629683494568, + 0.07897082716226578, + 0.0748840793967247, + 0.25144651532173157, + 0.13784781098365784 + ], + "credit_map": [ + 0.017069727182388306, + 0.0214233361184597, + 0.020689131692051888, + 0.013059187680482864, + 0.012771114706993103, + 0.01744234561920166, + 0.008477709256112576, + 0.026294849812984467, + 0.007864181883633137, + 0.009839799255132675, + 0.008943386375904083, + 0.008866850286722183, + 0.008742156438529491, + 0.01042142603546381, + 0.005888923071324825, + 0.006885692477226257, + 0.013689334504306316, + 0.010291671380400658, + 0.012431378476321697, + 0.010784303769469261, + 0.008236448280513287, + 0.013153590261936188, + 0.005580594297498465, + 0.008439128287136555, + 0.005574100650846958, + 0.0036829414311796427, + 0.003189063398167491, + 0.0026223857421427965, + 0.002742051612585783, + 0.002981139114126563, + 0.002366339787840843, + 0.003083914052695036, + 0.0058559998869895935, + 0.006988969631493092, + 0.0039749834686517715, + 0.004717128351330757, + 0.003965636715292931, + 0.004190067294985056, + 0.0039600105956196785, + 0.005373441614210606, + 0.0222042053937912, + 0.017872566357254982, + 0.013010188937187195, + 0.008680056780576706, + 0.007285492494702339, + 0.019402476027607918, + 0.006163270678371191, + 0.009678030386567116, + 0.008542989380657673, + 0.01701551303267479, + 0.007534533739089966, + 0.008719727396965027, + 0.01129333395510912, + 0.01081067230552435, + 0.006396804936230183, + 0.008657249622046947, + 0.011141248978674412, + 0.010847179219126701, + 0.00932462140917778, + 0.006722063757479191, + 0.008701720274984837, + 0.01208097580820322, + 0.0062380945309996605, + 0.009828167036175728, + 0.024237817153334618, + 0.06856895983219147, + 0.0247070100158453, + 0.017162291333079338, + 0.03643655404448509, + 0.04901176691055298, + 0.014560703188180923, + 0.016761409118771553, + 0.012778435833752155, + 0.011864964850246906, + 0.02993234433233738, + 0.007633588742464781, + 0.012086944654583931, + 0.040635257959365845, + 0.008030738681554794, + 0.01488553173840046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.04309936612844467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003693282604218 + ], + "opposite_logprob_tokens": [ + -2.6950368881225586 + ], + "logprob_reward": -0.07003693282604218, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003693282604218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1356220543384552, + 0.22529418766498566, + 0.0410204716026783, + 0.0348721444606781, + 0.04446333274245262, + 0.1285778135061264, + 0.08096639066934586, + 0.061140768229961395, + 0.1541481763124466, + 0.09389469027519226 + ], + "credit_map": [ + 0.0211366917937994, + 0.02339252457022667, + 0.014571798034012318, + 0.013826827518641949, + 0.011974554508924484, + 0.022639200091362, + 0.01001732423901558, + 0.018063127994537354, + 0.04078255966305733, + 0.06541074067354202, + 0.027826178818941116, + 0.013573902659118176, + 0.023393860086798668, + 0.02515299804508686, + 0.007517251186072826, + 0.021636661142110825, + 0.006498614326119423, + 0.005711741745471954, + 0.007120282854884863, + 0.0040852976962924, + 0.0037897236179560423, + 0.006776474416255951, + 0.003445127047598362, + 0.0035932078026235104, + 0.005451424513012171, + 0.006101907230913639, + 0.004036782309412956, + 0.004642348270863295, + 0.002879080595448613, + 0.0050078509375452995, + 0.003248851513490081, + 0.003503895364701748, + 0.004597512539476156, + 0.0058207097463309765, + 0.004806396085768938, + 0.00504350196570158, + 0.005986205767840147, + 0.006424885243177414, + 0.007032417226582766, + 0.004751700907945633, + 0.013865948654711246, + 0.025795305147767067, + 0.013722596690058708, + 0.012014307081699371, + 0.02096063643693924, + 0.01752471923828125, + 0.01151974592357874, + 0.013174544088542461, + 0.009288646280765533, + 0.015897434204816818, + 0.008271194994449615, + 0.010351614095270634, + 0.009773112833499908, + 0.011071177199482918, + 0.006780832540243864, + 0.00953237060457468, + 0.00672911386936903, + 0.011628838256001472, + 0.007585481274873018, + 0.00502384640276432, + 0.004748737905174494, + 0.014629206620156765, + 0.003788965754210949, + 0.007006573025137186, + 0.01446071919053793, + 0.033151134848594666, + 0.017457926645874977, + 0.010476335883140564, + 0.029666120186448097, + 0.027322959154844284, + 0.011215907521545887, + 0.010397061705589294, + 0.012389104813337326, + 0.013378781266510487, + 0.014227520674467087, + 0.006940154358744621, + 0.01418313942849636, + 0.016635961830615997, + 0.0060261995531618595, + 0.010113818570971489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.07003693282604218 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208044052124023 + ], + "opposite_logprob_tokens": [ + -1.5020804405212402 + ], + "logprob_reward": -0.25208044052124023, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208044052124023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08089519292116165, + 0.04254008084535599, + 0.02010110579431057, + 0.01968221738934517, + 0.03260152414441109, + 0.08189154416322708, + 0.04210323467850685, + 0.14043520390987396, + 0.3969619870185852, + 0.14278791844844818 + ], + "credit_map": [ + 0.014413885772228241, + 0.01287370827049017, + 0.011591460555791855, + 0.0073201037012040615, + 0.008222607895731926, + 0.0102603230625391, + 0.005047931335866451, + 0.011165174655616283, + 0.007234614808112383, + 0.0069362204521894455, + 0.004405742511153221, + 0.004354942124336958, + 0.004412735812366009, + 0.008159897290170193, + 0.0030477314721792936, + 0.003988194745033979, + 0.002298664301633835, + 0.003223842941224575, + 0.0020937619265168905, + 0.0038081039674580097, + 0.0021356968209147453, + 0.0023776493035256863, + 0.002114934381097555, + 0.002048452151939273, + 0.001928673475049436, + 0.0026607520412653685, + 0.0020929528400301933, + 0.002783068222925067, + 0.002777191810309887, + 0.00275434460490942, + 0.003009505569934845, + 0.0016757296398282051, + 0.0047697387635707855, + 0.0052142697386443615, + 0.004049236420542002, + 0.003784124506637454, + 0.0045525794848799706, + 0.0043657273054122925, + 0.002770820399746299, + 0.003095026593655348, + 0.010773160494863987, + 0.012296422384679317, + 0.00949947815388441, + 0.00901809986680746, + 0.010971110314130783, + 0.01035728957504034, + 0.008692358620464802, + 0.01028362661600113, + 0.004246448632329702, + 0.008635959587991238, + 0.003794893156737089, + 0.005650934297591448, + 0.0039610606618225574, + 0.006245295517146587, + 0.004610264673829079, + 0.004958377685397863, + 0.018699640408158302, + 0.023619895800948143, + 0.016085952520370483, + 0.01944756880402565, + 0.014873191714286804, + 0.016986165195703506, + 0.015967458486557007, + 0.014755338430404663, + 0.037625234574079514, + 0.09928623586893082, + 0.040840089321136475, + 0.027183081954717636, + 0.06459520012140274, + 0.07642178237438202, + 0.02580815553665161, + 0.02520221285521984, + 0.017638755962252617, + 0.018190249800682068, + 0.019492974504828453, + 0.013792197220027447, + 0.020535042509436607, + 0.020190801471471786, + 0.010748014785349369, + 0.022199884057044983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.25208044052124023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a horizontal angle, slowly tilts downward, following a suit of ancient, battle-worn steel armor as it descends with a solemn, slow descent into the dark, murky waters. The armor consists of a closed helmet with narrow eye slits, a sturdy breastplate, and tightly connected arm and leg armor. Its dull metal surface is covered with countless scratches and dents of varying depths, silently bearing witness to the fierce battles of the past. Mottled rust seeps through the seams. The only source of light comes from far above the water, casting a faint, flickering, cold beam. The light penetrates the deep blue-black water, revealing countless tiny suspended particles drifting slowly within. As this faint light sweeps across the armor's metal surface, it briefly illuminates the ancient scars. The most eye-catching thing is that a series of fine and continuous silver bubbles are rushing out from the narrow gaps in the helmet's eyelets, forming a clear trajectory and floating towards the unreachable light source above, forming a sharp contrast with the fate of the heavy falling armor. The whole picture is filled with a quiet, lonely and slightly tragic atmosphere, as if witnessing the end of a heroic era.", + "seed": 23933, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_15/index_1672_seed_23933_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03148223934695125, + "vlm_logprob": -0.03148223934695125, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03148223934695125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00746721588075161 + ], + "opposite_logprob_tokens": [ + -5.257467269897461 + ], + "logprob_reward": -0.00746721588075161, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00746721588075161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024798085913062096, + 0.12634971737861633, + 0.563125491142273, + 0.05893097445368767, + 0.08290571719408035, + 0.053114939481019974, + 0.0419827476143837, + 0.020204046741127968, + 0.01455281488597393, + 0.014035508036613464 + ], + "credit_map": [ + 0.002135687740519643, + 0.0029193656519055367, + 0.0032576113007962704, + 0.003712647594511509, + 0.003216855926439166, + 0.003153272205963731, + 0.0036828925367444754, + 0.002719749230891466, + 0.0062773507088422775, + 0.003753755008801818, + 0.009536879137158394, + 0.0056784143671393394, + 0.005621662363409996, + 0.042841535061597824, + 0.0076347896829247475, + 0.04500531032681465, + 0.1064382791519165, + 0.019277779385447502, + 0.056823331862688065, + 0.025019271299242973, + 0.0216495580971241, + 0.15212011337280273, + 0.05025693401694298, + 0.13154014945030212, + 0.0071606808342039585, + 0.0027899614069610834, + 0.006827566772699356, + 0.003917342983186245, + 0.0029479151125997305, + 0.01624002680182457, + 0.003970515914261341, + 0.0150769567117095, + 0.0047641051933169365, + 0.0018361590337008238, + 0.02913263812661171, + 0.002865702146664262, + 0.0038491718005388975, + 0.006264959461987019, + 0.003286437364295125, + 0.03090653568506241, + 0.004193627275526524, + 0.004720615688711405, + 0.007655961439013481, + 0.004731345921754837, + 0.011660163290798664, + 0.005303933750838041, + 0.008878186345100403, + 0.00597109692171216, + 0.004452227149158716, + 0.0030173873528838158, + 0.0057577402330935, + 0.0067771850153803825, + 0.005681447219103575, + 0.004027902614325285, + 0.008018494583666325, + 0.004250360187143087, + 0.002227935940027237, + 0.0011638094438239932, + 0.0042048171162605286, + 0.001954621635377407, + 0.0020116327796131372, + 0.0019418998854234815, + 0.0035303866025060415, + 0.0031689428724348545, + 0.0018789329333230853, + 0.0016478636534884572, + 0.001830662484280765, + 0.001354321837425232, + 0.0014768739929422736, + 0.002083124592900276, + 0.0024128640070557594, + 0.0018681699875742197, + 0.001190751325339079, + 0.0012845811434090137, + 0.0017640035366639495, + 0.0016773779643699527, + 0.0020511308684945107, + 0.001422605593688786, + 0.0023691258393228054, + 0.00227593001909554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.00746721588075161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110697191208601 + ], + "opposite_logprob_tokens": [ + -5.504110813140869 + ], + "logprob_reward": -0.004110697191208601, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110697191208601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08890850096940994, + 0.06395535916090012, + 0.20390982925891876, + 0.3835867941379547, + 0.16719064116477966, + 0.026520786806941032, + 0.01928044855594635, + 0.010958774946630001, + 0.010373544879257679, + 0.025315236300230026 + ], + "credit_map": [ + 0.005793479736894369, + 0.0030759137589484453, + 0.006111353635787964, + 0.004209026228636503, + 0.0037606379482895136, + 0.031374093145132065, + 0.005105908028781414, + 0.029478086158633232, + 0.0033225403167307377, + 0.0017002879176288843, + 0.005093990825116634, + 0.0026786657981574535, + 0.002168207196518779, + 0.022444825619459152, + 0.0030647304374724627, + 0.023482108488678932, + 0.03772701695561409, + 0.007484929636120796, + 0.020904485136270523, + 0.01049527432769537, + 0.008250045590102673, + 0.05353575572371483, + 0.019907787442207336, + 0.04560455307364464, + 0.04996713995933533, + 0.010330893099308014, + 0.038538217544555664, + 0.013340041972696781, + 0.011630489490926266, + 0.1240912452340126, + 0.01960967294871807, + 0.11607910692691803, + 0.008440854959189892, + 0.0051392181776463985, + 0.06037634238600731, + 0.006421244703233242, + 0.008547795936465263, + 0.008948339149355888, + 0.006351627875119448, + 0.06296522170305252, + 0.002144619356840849, + 0.0022716340608894825, + 0.0032701469026505947, + 0.00290373875759542, + 0.005418968386948109, + 0.0025692859198898077, + 0.0049405330792069435, + 0.0030018615070730448, + 0.00152136804535985, + 0.001469315029680729, + 0.002646042499691248, + 0.0024750600568950176, + 0.0038108627777546644, + 0.001912099658511579, + 0.0034774700179696083, + 0.0019682322163134813, + 0.0014803579542785883, + 0.0007956804474815726, + 0.0015341416001319885, + 0.0013973390450701118, + 0.001299835741519928, + 0.0011259063612669706, + 0.0018569283420220017, + 0.0014685852220281959, + 0.001131063443608582, + 0.0009980248287320137, + 0.0013769841752946377, + 0.0012070010416209698, + 0.0011378817725926638, + 0.0012461839942261577, + 0.0021744517143815756, + 0.0011019547237083316, + 0.0017746643861755729, + 0.002167320577427745, + 0.003341112984344363, + 0.0025328206829726696, + 0.0032246538903564215, + 0.0032066376879811287, + 0.005034693982452154, + 0.004033333621919155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.004110697191208601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200181424617767 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.018200181424617767, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018200181424617767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014646274037659168, + 0.008327437564730644, + 0.7903306484222412, + 0.08103109896183014, + 0.01961895450949669, + 0.04404798522591591, + 0.01673956960439682, + 0.005831501446664333, + 0.007269411813467741, + 0.012157035991549492 + ], + "credit_map": [ + 0.0011903453851118684, + 0.0010712419170886278, + 0.001774817006662488, + 0.0016142043750733137, + 0.0013734051026403904, + 0.0031284084543585777, + 0.002049692440778017, + 0.0024441590066999197, + 0.0008774251909926534, + 0.0005574721144512296, + 0.0011282797204330564, + 0.000741152442060411, + 0.0005632645916193724, + 0.001377154840156436, + 0.001141283311881125, + 0.0019414055859670043, + 0.15020905435085297, + 0.026876579970121384, + 0.08026769757270813, + 0.03629893437027931, + 0.030531393364071846, + 0.21255336701869965, + 0.07031122595071793, + 0.18328246474266052, + 0.010167508386075497, + 0.0024352010805159807, + 0.008042031899094582, + 0.0029174003284424543, + 0.0027013623621314764, + 0.02617538906633854, + 0.004263083916157484, + 0.024329116567969322, + 0.001712024793960154, + 0.0013700880808755755, + 0.003637909423559904, + 0.0014549605548381805, + 0.0025509728584438562, + 0.0021298895590007305, + 0.002349183429032564, + 0.004413927439600229, + 0.0034248316660523415, + 0.003937243018299341, + 0.005292365327477455, + 0.004459319170564413, + 0.00954808946698904, + 0.003965344280004501, + 0.008661983534693718, + 0.004758808296173811, + 0.0016923759831115603, + 0.0013361983001232147, + 0.00219569425098598, + 0.0017244487535208464, + 0.0023426483385264874, + 0.0018665653187781572, + 0.003866833169013262, + 0.0017148061888292432, + 0.000688293541315943, + 0.00046198730706237257, + 0.0008556105312891304, + 0.0007887986139394343, + 0.0005884708953090012, + 0.0006482165190391243, + 0.0008331202552653849, + 0.000967003870755434, + 0.0008212162647396326, + 0.0006346633308567107, + 0.0009225673857145011, + 0.0007083347300067544, + 0.0009722564136609435, + 0.0011436596978455782, + 0.0013309363275766373, + 0.0007357773138210177, + 0.000855113088618964, + 0.0009348581661470234, + 0.0015753100160509348, + 0.00136562658008188, + 0.0016995953628793359, + 0.0012593473074957728, + 0.002910496899858117, + 0.00155668705701828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.018200181424617767 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898922264575958 + ], + "opposite_logprob_tokens": [ + -2.578989267349243 + ], + "logprob_reward": -0.07898922264575958, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898922264575958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027888992801308632, + 0.018103159964084625, + 0.6264738440513611, + 0.19451025128364563, + 0.059254009276628494, + 0.014749662019312382, + 0.011201156303286552, + 0.005421057343482971, + 0.010087807662785053, + 0.032310061156749725 + ], + "credit_map": [ + 0.0016025216318666935, + 0.0011559722479432821, + 0.0023513787891715765, + 0.002348850015550852, + 0.0011595884570851922, + 0.008655551820993423, + 0.0021630472037941217, + 0.00845208391547203, + 0.0013208092423155904, + 0.0011081387056037784, + 0.0017257874133065343, + 0.0010159832891076803, + 0.0009523425251245499, + 0.004882432986050844, + 0.0015257886843755841, + 0.005571877583861351, + 0.1182524561882019, + 0.021160995587706566, + 0.06337913870811462, + 0.028570881113409996, + 0.024105491116642952, + 0.16915138065814972, + 0.05509903281927109, + 0.14675447344779968, + 0.025181951001286507, + 0.005331852473318577, + 0.01966821402311325, + 0.006846324540674686, + 0.005894085392355919, + 0.06325062364339828, + 0.00980455707758665, + 0.05853264406323433, + 0.003079368732869625, + 0.0010319885332137346, + 0.02244393900036812, + 0.0019149263389408588, + 0.001685348805040121, + 0.0032219255808740854, + 0.0017139110714197159, + 0.024162601679563522, + 0.001170785748399794, + 0.0014180067228153348, + 0.002073186682537198, + 0.0016271823551505804, + 0.0020717610605061054, + 0.0015988288214430213, + 0.0029843002557754517, + 0.0018056095577776432, + 0.0009182605426758528, + 0.0009654152672737837, + 0.0014201720478013158, + 0.0014679133892059326, + 0.0015681067015975714, + 0.0014351651770994067, + 0.002344638342037797, + 0.0010814847191795707, + 0.0005751400603912771, + 0.0004393299750518054, + 0.0008532871142961085, + 0.00059687061002478, + 0.0004909863928332925, + 0.0006080640596337616, + 0.0009193160221911967, + 0.0009380629635415971, + 0.0010287060867995024, + 0.0008039756212383509, + 0.001522672362625599, + 0.0011977292597293854, + 0.0012899250723421574, + 0.0009147311793640256, + 0.0020235779229551554, + 0.00130648969206959, + 0.0022069544065743685, + 0.004200053866952658, + 0.004557259380817413, + 0.004293333739042282, + 0.003271348774433136, + 0.0027270535938441753, + 0.007069310639053583, + 0.003984745126217604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.07898922264575958 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a suit of ancient, battle-worn steel armor descending into water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the armor feature a closed helmet with narrow eye slits and a sturdy breastplate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor's dull metal surface covered with scratches, dents, and mottled rust seeping through the seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine, continuous silver bubbles rushing out from the helmet's eyelets and floating upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot from a horizontal angle and slowly tilt downward to follow the armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864387959241867 + ], + "opposite_logprob_tokens": [ + -3.0486438274383545 + ], + "logprob_reward": -0.04864387959241867, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864387959241867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02083035744726658, + 0.045593518763780594, + 0.3126746416091919, + 0.28328683972358704, + 0.11068335175514221, + 0.08409903943538666, + 0.034577079117298126, + 0.0327083058655262, + 0.01853168196976185, + 0.057015180587768555 + ], + "credit_map": [ + 0.001717493636533618, + 0.002127631101757288, + 0.0030955192632973194, + 0.002523416420444846, + 0.0022174096666276455, + 0.0033056172542274, + 0.0022524234373122454, + 0.0035908485297113657, + 0.0023996613454073668, + 0.002173547400161624, + 0.0038762909825891256, + 0.0024722416419535875, + 0.0021486924961209297, + 0.013768862001597881, + 0.004495819099247456, + 0.014258402399718761, + 0.05750259757041931, + 0.01031620241701603, + 0.031528670340776443, + 0.015033903531730175, + 0.012004124000668526, + 0.08446706831455231, + 0.028220992535352707, + 0.07360105216503143, + 0.03808029368519783, + 0.00837432686239481, + 0.02869282104074955, + 0.01008552499115467, + 0.008669579401612282, + 0.09003347903490067, + 0.014971260912716389, + 0.08437954634428024, + 0.005255061201751232, + 0.0044371807016432285, + 0.03717635199427605, + 0.0046876645646989346, + 0.008059977553784847, + 0.0064965831115841866, + 0.006249746773391962, + 0.038320787250995636, + 0.005700245033949614, + 0.006971163675189018, + 0.009357885457575321, + 0.008481296710669994, + 0.017962459474802017, + 0.007682324852794409, + 0.020083868876099586, + 0.00785980187356472, + 0.004362746607512236, + 0.0028761245775967836, + 0.004845620598644018, + 0.005246238317340612, + 0.0043532708659768105, + 0.0034142346121370792, + 0.005850357469171286, + 0.0036284870002418756, + 0.00464435713365674, + 0.0019614158663898706, + 0.005112780723720789, + 0.003616997506469488, + 0.0032986088190227747, + 0.0036843952257186174, + 0.0044355993159115314, + 0.005954149644821882, + 0.0015665743267163634, + 0.00168883940204978, + 0.0021419443655759096, + 0.002102846512570977, + 0.0026637581177055836, + 0.0015744392294436693, + 0.004453974775969982, + 0.0023393055889755487, + 0.003744588466361165, + 0.004632103256881237, + 0.00827766116708517, + 0.006595120765268803, + 0.005717473104596138, + 0.006287501193583012, + 0.013856432400643826, + 0.007904301397502422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.04864387959241867 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_tensor([4901, 4901])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_tensor([4901, 4901])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2d69d5f46942d4557b0aded4113ac6627d0a42e8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_tensor([4901, 4901])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle shot begins with a tight close-up of a tense, muscular calf in the starting blocks, then quickly pulls back as the starting gun sounds. The athlete, an African man with a chiseled, determined face and short, dark hair, charges forward explosively. He wears a vibrant red sleeveless athletic vest made of lightweight, moisture-wicking fabric and a pair of tight black shorts. As the camera pulls back to a wide shot, it reveals the entire scene: a professional terracotta athletic track with crisp white lane markings, bathed in the brilliant, low afternoon sun. The camera then shifts to a smooth, rapid tracking shot, keeping pace with the leading runner as he enters the first turn of the 400-meter race. His form is impeccable; his arms pump rhythmically, his knees lift, and his body leans into the curve to counteract the force. Sweat gleams on his chiseled shoulders and back, catching the golden light and creating sharp highlights. In the background, the blurred forms of other runners struggle to keep up, while the stadium stands form a soft-focus mosaic of cheering crowds. With decisive power, he steadily passed runners in adjacent lanes, his expression of focus unwavering, creating an intense and exciting atmosphere of peak physical fitness and competitive advantage.", + "seed": 15668, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_4901_seed_15668_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02761864485219121, + "vlm_logprob": -0.02761864485219121, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.02761864485219121, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10775772482156754 + ], + "opposite_logprob_tokens": [ + -2.357757806777954 + ], + "logprob_reward": -0.10775772482156754, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10775772482156754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.711296558380127, + 0.017914671450853348, + 0.018923360854387283, + 0.1463107317686081, + 0.017185622826218605, + 0.012508886866271496, + 0.01996701769530773, + 0.02248489297926426, + 0.014655130915343761, + 0.018753161653876305 + ], + "credit_map": [ + 0.06273217499256134, + 0.04049699380993843, + 0.1864204704761505, + 0.03745657950639725, + 0.022931700572371483, + 0.04251716285943985, + 0.04977408051490784, + 0.2689674198627472, + 0.002955848816782236, + 0.0031613297760486603, + 0.0017034040065482259, + 0.0025441646575927734, + 0.001983755500987172, + 0.0021321841049939394, + 0.0016930943820625544, + 0.0017408893909305334, + 0.0026715912390500307, + 0.0019144867546856403, + 0.003065188182517886, + 0.0029941564425826073, + 0.00188438780605793, + 0.001773267169483006, + 0.0025052865967154503, + 0.0021149960812181234, + 0.014895516447722912, + 0.008853203617036343, + 0.04753908887505531, + 0.00782904215157032, + 0.01625310257077217, + 0.016300944611430168, + 0.009772366844117641, + 0.024867480620741844, + 0.001785490894690156, + 0.0020072893239557743, + 0.002063547493889928, + 0.0020278512965887785, + 0.0026230618823319674, + 0.0018724543042480946, + 0.0028961969073861837, + 0.001909730490297079, + 0.001212099683471024, + 0.001184621243737638, + 0.002612403128296137, + 0.0014166680630296469, + 0.0016159792430698872, + 0.0014096099184826016, + 0.001278506126254797, + 0.001778999576345086, + 0.0032891312148422003, + 0.0021723886020481586, + 0.002446940168738365, + 0.0025169430300593376, + 0.0022522853687405586, + 0.00236288970336318, + 0.0024946255143731833, + 0.002431814093142748, + 0.0020039251539856195, + 0.004114934708923101, + 0.0033131262753158808, + 0.0026990079786628485, + 0.0017161224968731403, + 0.0029853354208171368, + 0.0019096927717328072, + 0.003742747940123081, + 0.0014962187269702554, + 0.0021067264024168253, + 0.0009891404770314693, + 0.0019941197242587805, + 0.0011736475862562656, + 0.0023291087709367275, + 0.0015407215105369687, + 0.003025447018444538, + 0.002431361936032772, + 0.0023215266410261393, + 0.003674377454444766, + 0.0015591677511110902, + 0.0018503374885767698, + 0.0019809154327958822, + 0.001094684936106205, + 0.003840789897367358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.10775772482156754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009818002581596375 + ], + "opposite_logprob_tokens": [ + -4.634818077087402 + ], + "logprob_reward": -0.009818002581596375, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009818002581596375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5950403809547424, + 0.0456051267683506, + 0.0982380211353302, + 0.01564798690378666, + 0.06417488306760788, + 0.01234138011932373, + 0.024330729618668556, + 0.06897509098052979, + 0.017866747453808784, + 0.05777965113520622 + ], + "credit_map": [ + 0.054237473756074905, + 0.03543826565146446, + 0.15854866802692413, + 0.0350547693669796, + 0.020480290055274963, + 0.032284390181303024, + 0.041502825915813446, + 0.2174936980009079, + 0.00538875674828887, + 0.005413915030658245, + 0.00632241927087307, + 0.007609279826283455, + 0.0060729896649718285, + 0.004366503097116947, + 0.005828501656651497, + 0.004602755885571241, + 0.011750900186598301, + 0.009023264050483704, + 0.016410311684012413, + 0.012369035743176937, + 0.0076595149002969265, + 0.011322790756821632, + 0.008569573052227497, + 0.02113262377679348, + 0.0015984593192115426, + 0.0009840348502621055, + 0.003676526714116335, + 0.001956155989319086, + 0.0012528616935014725, + 0.0019197085639461875, + 0.002238297602161765, + 0.0020219413563609123, + 0.004983472637832165, + 0.010170534253120422, + 0.006136790383607149, + 0.008934876881539822, + 0.010466638952493668, + 0.006195561960339546, + 0.011126423254609108, + 0.006160584744066, + 0.0015488050412386656, + 0.0014885346172377467, + 0.0016478594625368714, + 0.0012980010360479355, + 0.0013421685434877872, + 0.0018144960049539804, + 0.0009484721231274307, + 0.002253042533993721, + 0.0025266013108193874, + 0.0024934220127761364, + 0.0036781486123800278, + 0.002761946991086006, + 0.002555122599005699, + 0.0026584418956190348, + 0.0027191757690161467, + 0.004937867633998394, + 0.010882169008255005, + 0.008061105385422707, + 0.007903319783508778, + 0.009074453264474869, + 0.005913597997277975, + 0.009168668650090694, + 0.0067761908285319805, + 0.01119557861238718, + 0.0014320212649181485, + 0.0036994614638388157, + 0.0015760429669171572, + 0.0026257007848471403, + 0.002123314654454589, + 0.0028294913936406374, + 0.0018205627566203475, + 0.001760152168571949, + 0.00579937594011426, + 0.008535160683095455, + 0.007407829165458679, + 0.006872684694826603, + 0.005780365783721209, + 0.008037661202251911, + 0.004038074053823948, + 0.01130849402397871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.009818002581596375 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525664072483778 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.00525664072483778, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525664072483778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43195319175720215, + 0.04431517422199249, + 0.09777514636516571, + 0.03725448250770569, + 0.09144818037748337, + 0.049364399164915085, + 0.041354138404130936, + 0.05912427976727486, + 0.07964105159044266, + 0.06776998937129974 + ], + "credit_map": [ + 0.037841491401195526, + 0.026447897776961327, + 0.10967866331338882, + 0.024651842191815376, + 0.01820092834532261, + 0.030428942292928696, + 0.0331343337893486, + 0.15156905353069305, + 0.009926537983119488, + 0.005778270773589611, + 0.0048568351194262505, + 0.0038572538178414106, + 0.0043620518408715725, + 0.006517067085951567, + 0.0033983492758125067, + 0.0056188046000897884, + 0.0099486680701375, + 0.008842484094202518, + 0.015515008941292763, + 0.017052646726369858, + 0.007513702381402254, + 0.010019540786743164, + 0.009832722134888172, + 0.01905037648975849, + 0.003901327960193157, + 0.0021999874152243137, + 0.01082531362771988, + 0.0032175725791603327, + 0.0037984298542141914, + 0.002973368624225259, + 0.0036898809485137463, + 0.006648599170148373, + 0.01235668733716011, + 0.006786560174077749, + 0.018008287996053696, + 0.010850473307073116, + 0.009781059809029102, + 0.01432069018483162, + 0.007813825272023678, + 0.011530587449669838, + 0.007148948032408953, + 0.004609517753124237, + 0.009560414589941502, + 0.006200532428920269, + 0.003970754332840443, + 0.0057918354868888855, + 0.004835905972868204, + 0.007246487773954868, + 0.008183482103049755, + 0.006564118899405003, + 0.005391059443354607, + 0.0033293133601546288, + 0.0045761847868561745, + 0.004498516209423542, + 0.004623821470886469, + 0.004187639337033033, + 0.006218845956027508, + 0.008160091005265713, + 0.010814312845468521, + 0.007197346072643995, + 0.00434922007843852, + 0.008050941862165928, + 0.00501914881169796, + 0.009314370341598988, + 0.009037346579134464, + 0.009268028661608696, + 0.0073827519081532955, + 0.015287309885025024, + 0.007796136662364006, + 0.011392992921173573, + 0.005830234847962856, + 0.013646245934069157, + 0.010804576799273491, + 0.010096199810504913, + 0.010928257368505001, + 0.00746642192825675, + 0.005628609098494053, + 0.007797256112098694, + 0.0039712111465632915, + 0.011077449657022953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.00525664072483778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631078828126192 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631078828126192, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631078828126192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1458347737789154, + 0.05861189588904381, + 0.042611368000507355, + 0.29717734456062317, + 0.044789087027311325, + 0.06367126107215881, + 0.06731076538562775, + 0.06606914103031158, + 0.09888331592082977, + 0.11504105478525162 + ], + "credit_map": [ + 0.023389561101794243, + 0.014907135628163815, + 0.032532911747694016, + 0.009513895958662033, + 0.011019058525562286, + 0.011468732729554176, + 0.013436440378427505, + 0.029567036777734756, + 0.011561902239918709, + 0.008004546165466309, + 0.0071236751973629, + 0.006980920676141977, + 0.005817655473947525, + 0.006135797128081322, + 0.0054505206644535065, + 0.007536883465945721, + 0.005412759725004435, + 0.004059056285768747, + 0.007342694792896509, + 0.005769795272499323, + 0.005022295750677586, + 0.00436436478048563, + 0.0050711967051029205, + 0.005569207947701216, + 0.03090486116707325, + 0.021199025213718414, + 0.08941817283630371, + 0.01853903941810131, + 0.03657865896821022, + 0.032256342470645905, + 0.02302778698503971, + 0.0452534556388855, + 0.004828469827771187, + 0.004174462053924799, + 0.006907123606652021, + 0.006558072287589312, + 0.004972219001501799, + 0.008891928009688854, + 0.003703383030369878, + 0.004753426648676395, + 0.0063467733561992645, + 0.006330116651952267, + 0.011006969027221203, + 0.007985306903719902, + 0.007954820059239864, + 0.00663497531786561, + 0.008175387978553772, + 0.009236915037035942, + 0.008809581398963928, + 0.007241859100759029, + 0.009255640208721161, + 0.008873868733644485, + 0.009093227796256542, + 0.0072137052193284035, + 0.008715816773474216, + 0.008107064291834831, + 0.008875646628439426, + 0.007542512845247984, + 0.011919035576283932, + 0.011155965737998486, + 0.0042750234715640545, + 0.0075440919026732445, + 0.0051182652823626995, + 0.009638594463467598, + 0.007361066993325949, + 0.023334255442023277, + 0.0057926722802221775, + 0.011472745798528194, + 0.008028050884604454, + 0.01888921856880188, + 0.006401604972779751, + 0.01760370284318924, + 0.01826217770576477, + 0.013456516899168491, + 0.02504294365644455, + 0.009457058273255825, + 0.00962846353650093, + 0.012107285670936108, + 0.007712792605161667, + 0.019373826682567596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.007631078828126192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629777304828167 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629777304828167, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629777304828167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4113297760486603, + 0.0257783904671669, + 0.08453844487667084, + 0.11133107542991638, + 0.04394787549972534, + 0.017365744337439537, + 0.04149812087416649, + 0.08392573893070221, + 0.06575723737478256, + 0.11452760547399521 + ], + "credit_map": [ + 0.03611639142036438, + 0.024942489340901375, + 0.09937849640846252, + 0.024795249104499817, + 0.016064070165157318, + 0.02573264203965664, + 0.03253459557890892, + 0.15176580846309662, + 0.0031433177646249533, + 0.003919057548046112, + 0.003637641202658415, + 0.0035427811089903116, + 0.004062323365360498, + 0.00249202037230134, + 0.002474000211805105, + 0.002507248893380165, + 0.008248345926404, + 0.008363848552107811, + 0.013973630964756012, + 0.014284803532063961, + 0.007787856739014387, + 0.008377720601856709, + 0.007272219285368919, + 0.016230018809437752, + 0.009331726469099522, + 0.004808108322322369, + 0.03893974795937538, + 0.005599816329777241, + 0.01186587568372488, + 0.011962523683905602, + 0.008341822773218155, + 0.020481452345848083, + 0.007379007991403341, + 0.003293081885203719, + 0.010041027329862118, + 0.004154284950345755, + 0.004434905480593443, + 0.0056661078706383705, + 0.0036837642546743155, + 0.005295693874359131, + 0.0023366869427263737, + 0.00256999209523201, + 0.0029641268774867058, + 0.0016685917507857084, + 0.0015448551857843995, + 0.001936732092872262, + 0.0021683720406144857, + 0.0021763883996754885, + 0.003160863183438778, + 0.0048338561318814754, + 0.0037799167912453413, + 0.006267169956117868, + 0.004740569274872541, + 0.005193667486310005, + 0.006744221318513155, + 0.006777853239327669, + 0.009276874363422394, + 0.010253692977130413, + 0.013957035727798939, + 0.013686430640518665, + 0.006286632735282183, + 0.010924041271209717, + 0.008244787342846394, + 0.011296240612864494, + 0.0050613172352313995, + 0.008290043100714684, + 0.0053526870906353, + 0.010554141364991665, + 0.007958460599184036, + 0.008959488943219185, + 0.00947618205100298, + 0.01010491419583559, + 0.01515266951173544, + 0.016858357936143875, + 0.01750876009464264, + 0.016636673361063004, + 0.013406648300588131, + 0.013026311062276363, + 0.009448492899537086, + 0.012489697895944118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.007629777304828167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot begins with a tight close-up of a tense, muscular calf in the starting blocks, then quickly pulls back as the starting gun sounds. The athlete, an African man with a chiseled, determined face and short, dark hair, charges forward explosively. He wears a vibrant red sleeveless athletic vest made of lightweight, moisture-wicking fabric and a pair of tight black shorts. As the camera pulls back to a wide shot, it reveals the entire scene: a professional terracotta athletic track with crisp white lane markings, bathed in the brilliant, low afternoon sun. The camera then shifts to a smooth, rapid tracking shot, keeping pace with the leading runner as he enters the first turn of the 400-meter race. His form is impeccable; his arms pump rhythmically, his knees lift, and his body leans into the curve to counteract the force. Sweat gleams on his chiseled shoulders and back, catching the golden light and creating sharp highlights. In the background, the blurred forms of other runners struggle to keep up, while the stadium stands form a soft-focus mosaic of cheering crowds. With decisive power, he steadily passed runners in adjacent lanes, his expression of focus unwavering, creating an intense and exciting atmosphere of peak physical fitness and competitive advantage.", + "seed": 15669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_4901_seed_15669_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014734081272035838, + "vlm_logprob": -0.014734081272035838, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.014734081272035838, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032313570380210876 + ], + "opposite_logprob_tokens": [ + -3.657313585281372 + ], + "logprob_reward": -0.032313570380210876, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.032313570380210876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06481542438268661, + 0.10495929419994354, + 0.4084164798259735, + 0.024769945070147514, + 0.16409586369991302, + 0.018146954476833344, + 0.0747932493686676, + 0.023082228377461433, + 0.06786516308784485, + 0.04905542731285095 + ], + "credit_map": [ + 0.005556945223361254, + 0.011845107190310955, + 0.0053108977153897285, + 0.009442881681025028, + 0.011758417822420597, + 0.005345280282199383, + 0.006366986781358719, + 0.009188907220959663, + 0.023862523958086967, + 0.024516042321920395, + 0.00900180172175169, + 0.0077600497752428055, + 0.008012506179511547, + 0.00863963644951582, + 0.004846347030252218, + 0.01832037977874279, + 0.030185846611857414, + 0.1538243144750595, + 0.1160048097372055, + 0.027644863352179527, + 0.02001958340406418, + 0.01355504896491766, + 0.012749573215842247, + 0.03443245217204094, + 0.0021582008339464664, + 0.0023846719413995743, + 0.0022354775574058294, + 0.0055118827149271965, + 0.002182952594012022, + 0.0033276821486651897, + 0.0023965630680322647, + 0.004572514910250902, + 0.01516000647097826, + 0.019485244527459145, + 0.016570940613746643, + 0.01946866326034069, + 0.03698284924030304, + 0.017997683957219124, + 0.02342991903424263, + 0.015000548213720322, + 0.0014099200488999486, + 0.002759781200438738, + 0.0020650248043239117, + 0.002345005050301552, + 0.002343312371522188, + 0.003598977578803897, + 0.0012496353592723608, + 0.0023752993438392878, + 0.0033303676173090935, + 0.003776642493903637, + 0.02946670539677143, + 0.009122984483838081, + 0.004122857935726643, + 0.01704307459294796, + 0.004310783930122852, + 0.0036198284942656755, + 0.004375853110104799, + 0.002437459072098136, + 0.0022693986538797617, + 0.0020389731507748365, + 0.002243262715637684, + 0.005445446819067001, + 0.00250609265640378, + 0.0017657423159107566, + 0.004236300475895405, + 0.023301612585783005, + 0.016864189878106117, + 0.00422532157972455, + 0.006527718622237444, + 0.0033218064345419407, + 0.005643482785671949, + 0.003744730493053794, + 0.008783026598393917, + 0.009115159511566162, + 0.0056984880939126015, + 0.004910046700388193, + 0.0033854928333312273, + 0.006762109696865082, + 0.005495764315128326, + 0.004905336536467075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.032313570380210876 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008663804270327091 + ], + "opposite_logprob_tokens": [ + -4.758663654327393 + ], + "logprob_reward": -0.008663804270327091, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008663804270327091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13132940232753754, + 0.04418008774518967, + 0.12436798959970474, + 0.08130046725273132, + 0.2518101632595062, + 0.017368342727422714, + 0.1313379406929016, + 0.06730186939239502, + 0.06830993294715881, + 0.08269380778074265 + ], + "credit_map": [ + 0.012559561058878899, + 0.03023734875023365, + 0.0064191496931016445, + 0.011121723800897598, + 0.029075706377625465, + 0.009145664982497692, + 0.008905950002372265, + 0.023864304646849632, + 0.008271550759673119, + 0.0064192055724561214, + 0.004584658890962601, + 0.004544259048998356, + 0.005241887643933296, + 0.007521979510784149, + 0.0036455385852605104, + 0.003951007965952158, + 0.013363437727093697, + 0.015998762100934982, + 0.02450069785118103, + 0.014641372486948967, + 0.010915292426943779, + 0.018537042662501335, + 0.013401901349425316, + 0.013009476475417614, + 0.005095307715237141, + 0.006559934001415968, + 0.006893729791045189, + 0.011519520543515682, + 0.010385662317276001, + 0.007135807070881128, + 0.020930245518684387, + 0.012780261225998402, + 0.03000820241868496, + 0.04059828072786331, + 0.024217654019594193, + 0.036617908626794815, + 0.05059671774506569, + 0.017376689240336418, + 0.0330205038189888, + 0.019374201074242592, + 0.0014534418005496264, + 0.0025781728327274323, + 0.0021271249279379845, + 0.001725173438899219, + 0.0020959272515028715, + 0.003533437615260482, + 0.001101102912798524, + 0.002753962529823184, + 0.006253615487366915, + 0.007313570939004421, + 0.05082640051841736, + 0.015158585272729397, + 0.0070116519927978516, + 0.0309609305113554, + 0.007339081261307001, + 0.006474108435213566, + 0.005849355366080999, + 0.009666753932833672, + 0.010827482677996159, + 0.008464972488582134, + 0.007722991518676281, + 0.0067769852466881275, + 0.01088283583521843, + 0.007110493257641792, + 0.006697842385619879, + 0.010826568119227886, + 0.006034671328961849, + 0.010349119082093239, + 0.008315901271998882, + 0.00992289837449789, + 0.008687389083206654, + 0.007475547958165407, + 0.006404617335647345, + 0.009916841983795166, + 0.009111450053751469, + 0.013983727432787418, + 0.010957170277833939, + 0.01104755699634552, + 0.012645002454519272, + 0.008627442643046379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.008663804270327091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953436251729727 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953436251729727, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953436251729727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3490619659423828, + 0.10828153789043427, + 0.09866195172071457, + 0.06028692424297333, + 0.025869881734251976, + 0.028062326833605766, + 0.07849095016717911, + 0.10680395364761353, + 0.1066756397485733, + 0.03780481964349747 + ], + "credit_map": [ + 0.023719245567917824, + 0.03909097611904144, + 0.015012199059128761, + 0.027806274592876434, + 0.1031574159860611, + 0.035216301679611206, + 0.03749899938702583, + 0.0675605982542038, + 0.0235472209751606, + 0.02419107034802437, + 0.007683679927140474, + 0.007380190305411816, + 0.009869173169136047, + 0.014091932214796543, + 0.004356165882200003, + 0.017162110656499863, + 0.013642321340739727, + 0.02218504250049591, + 0.02096368558704853, + 0.006898901890963316, + 0.0058861542493104935, + 0.009868571534752846, + 0.008582156151533127, + 0.01063513197004795, + 0.006840019952505827, + 0.004676487762480974, + 0.006721519399434328, + 0.0066245379857718945, + 0.010222171433269978, + 0.008517608046531677, + 0.009771262295544147, + 0.006913319230079651, + 0.003212809329852462, + 0.00530041242018342, + 0.0027409764006733894, + 0.003989305812865496, + 0.003388830693438649, + 0.0018405169248580933, + 0.0026489761658012867, + 0.0027480574790388346, + 0.002345462329685688, + 0.004359328653663397, + 0.0037246299907565117, + 0.0032694891560822725, + 0.0048080310225486755, + 0.004806609358638525, + 0.0016211428446695209, + 0.0031276377849280834, + 0.0053252833895385265, + 0.008003197610378265, + 0.018779059872031212, + 0.008751831948757172, + 0.007960410788655281, + 0.015070547349750996, + 0.00678166514262557, + 0.007818958722054958, + 0.021863851696252823, + 0.006065382622182369, + 0.006100984290242195, + 0.00943100918084383, + 0.00995996780693531, + 0.03377576172351837, + 0.009556349366903305, + 0.01005066279321909, + 0.00896503496915102, + 0.023885425180196762, + 0.012907860800623894, + 0.013228635303676128, + 0.011548337526619434, + 0.017979396507143974, + 0.010456775315105915, + 0.007704184856265783, + 0.005099743604660034, + 0.004495958797633648, + 0.005245651584118605, + 0.005537610966712236, + 0.0040175821632146835, + 0.005748867057263851, + 0.004235657397657633, + 0.0034237534273415804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.005953436251729727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541121803224087 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541121803224087, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541121803224087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15074355900287628, + 0.04670984297990799, + 0.07943802326917648, + 0.06418287754058838, + 0.035381101071834564, + 0.026729432865977287, + 0.3634745180606842, + 0.05134664103388786, + 0.08660759031772614, + 0.09538643062114716 + ], + "credit_map": [ + 0.008733113296329975, + 0.020930375903844833, + 0.006510436534881592, + 0.010484614409506321, + 0.04759421572089195, + 0.016077477484941483, + 0.00982678309082985, + 0.03058653511106968, + 0.010027330368757248, + 0.010528416372835636, + 0.003464046400040388, + 0.003177794162184, + 0.003933861386030912, + 0.006188766565173864, + 0.0023181417491286993, + 0.007071483414620161, + 0.014821164309978485, + 0.010512450709939003, + 0.014121240004897118, + 0.01190205104649067, + 0.006569416727870703, + 0.008777440525591373, + 0.005555529147386551, + 0.00717873265966773, + 0.004368993453681469, + 0.006284791510552168, + 0.006372261326760054, + 0.01343726646155119, + 0.006503556855022907, + 0.0043051578104496, + 0.011735348030924797, + 0.011175509542226791, + 0.0034916207659989595, + 0.008772355504333973, + 0.00336358230561018, + 0.004985070787370205, + 0.003981118556112051, + 0.0021732542663812637, + 0.004246178548783064, + 0.004367918707430363, + 0.002652102615684271, + 0.0047001466155052185, + 0.003839501179754734, + 0.0028386912308633327, + 0.003075782675296068, + 0.00514298677444458, + 0.0016849444946274161, + 0.0027952776290476322, + 0.018157316371798515, + 0.016002187505364418, + 0.14369486272335052, + 0.04047238454222679, + 0.016894197091460228, + 0.09227324277162552, + 0.020154522731900215, + 0.0158257856965065, + 0.007647817023098469, + 0.003933836705982685, + 0.003128723008558154, + 0.0042025428265333176, + 0.005267040338367224, + 0.01639191247522831, + 0.00579229136928916, + 0.004982478450983763, + 0.006705290637910366, + 0.019350288435816765, + 0.014331924729049206, + 0.013138272799551487, + 0.009075537323951721, + 0.00861272681504488, + 0.008755709044635296, + 0.0066378372721374035, + 0.020072419196367264, + 0.020200010389089584, + 0.010928260162472725, + 0.008951561525464058, + 0.007862635888159275, + 0.01090149488300085, + 0.009441806934773922, + 0.007028243970125914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.012541121803224087 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198473654687405 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198473654687405, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198473654687405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11444361507892609, + 0.027380965650081635, + 0.1045687273144722, + 0.040159326046705246, + 0.10680743306875229, + 0.03061620332300663, + 0.4681093096733093, + 0.03427482768893242, + 0.03425655886530876, + 0.03938305377960205 + ], + "credit_map": [ + 0.016010135412216187, + 0.016140727326273918, + 0.0062921615317463875, + 0.007377310190349817, + 0.030097922310233116, + 0.010074542835354805, + 0.008376739919185638, + 0.020074088126420975, + 0.0038376932498067617, + 0.004459210205823183, + 0.002197050955146551, + 0.0033003385178744793, + 0.003614194691181183, + 0.003990193363279104, + 0.0024538314901292324, + 0.0035284548066556454, + 0.008252543397247791, + 0.0361030176281929, + 0.026911811903119087, + 0.006781917996704578, + 0.004631999880075455, + 0.006947510875761509, + 0.005003363825380802, + 0.009936568327248096, + 0.0028796354308724403, + 0.003656040644273162, + 0.004131183493882418, + 0.0060476926155388355, + 0.006505344528704882, + 0.0043236264027655125, + 0.007711471989750862, + 0.004904333967715502, + 0.01042361930012703, + 0.011226599104702473, + 0.010278770700097084, + 0.01283229235559702, + 0.02158442698419094, + 0.012846790254116058, + 0.0169841218739748, + 0.010630830191075802, + 0.002574474783614278, + 0.005676981061697006, + 0.004063127096742392, + 0.004036105237901211, + 0.004721104167401791, + 0.003914012107998133, + 0.002193122636526823, + 0.0034372780937701464, + 0.022559981793165207, + 0.020910557359457016, + 0.18490587174892426, + 0.05151008069515228, + 0.021773435175418854, + 0.11898764967918396, + 0.02657616138458252, + 0.020885590463876724, + 0.005870228633284569, + 0.002622220665216446, + 0.002634322037920356, + 0.0029974228236824274, + 0.00336947338655591, + 0.00946122407913208, + 0.0032767869997769594, + 0.00404315022751689, + 0.0037713905330747366, + 0.003207854460924864, + 0.0037381290458142757, + 0.0028699368704110384, + 0.0048418715596199036, + 0.007728517521172762, + 0.004350572824478149, + 0.0037482872139662504, + 0.006436195224523544, + 0.00606316514313221, + 0.004603052977472544, + 0.005149926990270615, + 0.004181360825896263, + 0.004360889550298452, + 0.004790001083165407, + 0.003798465710133314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.014198473654687405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot begins with a tight close-up of a tense, muscular calf in the starting blocks, then quickly pulls back as the starting gun sounds. The athlete, an African man with a chiseled, determined face and short, dark hair, charges forward explosively. He wears a vibrant red sleeveless athletic vest made of lightweight, moisture-wicking fabric and a pair of tight black shorts. As the camera pulls back to a wide shot, it reveals the entire scene: a professional terracotta athletic track with crisp white lane markings, bathed in the brilliant, low afternoon sun. The camera then shifts to a smooth, rapid tracking shot, keeping pace with the leading runner as he enters the first turn of the 400-meter race. His form is impeccable; his arms pump rhythmically, his knees lift, and his body leans into the curve to counteract the force. Sweat gleams on his chiseled shoulders and back, catching the golden light and creating sharp highlights. In the background, the blurred forms of other runners struggle to keep up, while the stadium stands form a soft-focus mosaic of cheering crowds. With decisive power, he steadily passed runners in adjacent lanes, his expression of focus unwavering, creating an intense and exciting atmosphere of peak physical fitness and competitive advantage.", + "seed": 15670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_4901_seed_15670_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010663648694753647, + "vlm_logprob": -0.010663648694753647, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.010663648694753647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019284265115857124 + ], + "opposite_logprob_tokens": [ + -4.144284248352051 + ], + "logprob_reward": -0.019284265115857124, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019284265115857124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24517948925495148, + 0.31217244267463684, + 0.10389583557844162, + 0.04318421334028244, + 0.08728589862585068, + 0.02377135306596756, + 0.027725525200366974, + 0.01851724460721016, + 0.09713048487901688, + 0.04113752394914627 + ], + "credit_map": [ + 0.034932125359773636, + 0.037127647548913956, + 0.0458979606628418, + 0.020982399582862854, + 0.019215460866689682, + 0.03163117915391922, + 0.028184672817587852, + 0.02720802277326584, + 0.049941789358854294, + 0.029786352068185806, + 0.04834827780723572, + 0.0342927947640419, + 0.04386989772319794, + 0.052916254848241806, + 0.01842348463833332, + 0.0345936119556427, + 0.00645946292206645, + 0.008548128418624401, + 0.015637755393981934, + 0.03196372091770172, + 0.007596888579428196, + 0.0065136137418448925, + 0.008884886279702187, + 0.018291374668478966, + 0.005071441642940044, + 0.006246465258300304, + 0.007524667773395777, + 0.005783140659332275, + 0.00451657036319375, + 0.003741705324500799, + 0.005146098788827658, + 0.005154124926775694, + 0.006179682444781065, + 0.006778240203857422, + 0.0214251596480608, + 0.01807812601327896, + 0.008223416283726692, + 0.004330960102379322, + 0.009809925220906734, + 0.012460384517908096, + 0.0023913863115012646, + 0.0019111527362838387, + 0.003392426762729883, + 0.0040620132349431515, + 0.0024833090137690306, + 0.002542647300288081, + 0.002432214329019189, + 0.0045562018640339375, + 0.004924967885017395, + 0.003843616461381316, + 0.004050089046359062, + 0.0035481154918670654, + 0.0031753506045788527, + 0.001995822647586465, + 0.003370397724211216, + 0.0028171667363494635, + 0.0021437611430883408, + 0.004045238718390465, + 0.0025790741201490164, + 0.0026929073501378298, + 0.0017093594651669264, + 0.00210902513936162, + 0.0013785271439701319, + 0.0018593512941151857, + 0.023133525624871254, + 0.023277362808585167, + 0.01082412339746952, + 0.010917105711996555, + 0.006237255875021219, + 0.009961147792637348, + 0.005978290922939777, + 0.0068016694858670235, + 0.0034161123912781477, + 0.004125217907130718, + 0.011517214588820934, + 0.006115683354437351, + 0.004147000145167112, + 0.003423285437747836, + 0.004453074187040329, + 0.003939938731491566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.019284265115857124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964457057416439 + ], + "opposite_logprob_tokens": [ + -5.130964279174805 + ], + "logprob_reward": -0.005964457057416439, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964457057416439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08262188732624054, + 0.21933868527412415, + 0.22771087288856506, + 0.0734877958893776, + 0.07355403900146484, + 0.03638523817062378, + 0.03878941014409065, + 0.04835033044219017, + 0.1729731261730194, + 0.02678864449262619 + ], + "credit_map": [ + 0.009613743051886559, + 0.010755056515336037, + 0.009472792968153954, + 0.006931323558092117, + 0.007114890497177839, + 0.011061478406190872, + 0.013036934658885002, + 0.014635666273534298, + 0.034753601998090744, + 0.016479279845952988, + 0.030372560024261475, + 0.030506372451782227, + 0.03197071701288223, + 0.026168443262577057, + 0.018990086391568184, + 0.030097628012299538, + 0.016551339998841286, + 0.011094490997493267, + 0.07149088382720947, + 0.07549011707305908, + 0.014753020368516445, + 0.013352363370358944, + 0.012272990308701992, + 0.012705670669674873, + 0.006724483333528042, + 0.005315050482749939, + 0.012455793097615242, + 0.012995917350053787, + 0.006996003910899162, + 0.0054589672945439816, + 0.01404903456568718, + 0.009492545388638973, + 0.00830145925283432, + 0.004890745505690575, + 0.018210288137197495, + 0.007902568206191063, + 0.01184646412730217, + 0.0043273926712572575, + 0.011272798292338848, + 0.006802329793572426, + 0.005483826622366905, + 0.004217659123241901, + 0.004357269033789635, + 0.006202324293553829, + 0.0038532197941094637, + 0.004621412605047226, + 0.003612383035942912, + 0.00403714319691062, + 0.00659263227134943, + 0.006044600624591112, + 0.005526447203010321, + 0.006460610777139664, + 0.003721302840858698, + 0.0027127275243401527, + 0.0036365617997944355, + 0.004094528034329414, + 0.005576950032263994, + 0.011956052854657173, + 0.006902237422764301, + 0.005323873367160559, + 0.003958222456276417, + 0.006397945340722799, + 0.003581711556762457, + 0.004653342068195343, + 0.03985901176929474, + 0.04148242622613907, + 0.018822135403752327, + 0.01999710686504841, + 0.011555791832506657, + 0.019187593832612038, + 0.01058419980108738, + 0.011484860442578793, + 0.0019127376144751906, + 0.0025945995002985, + 0.005992473103106022, + 0.006369317881762981, + 0.0025635792408138514, + 0.0024464812595397234, + 0.002471975050866604, + 0.002437481889501214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.005964457057416439 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099062643945217 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004099062643945217, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099062643945217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11933230608701706, + 0.17662793397903442, + 0.20830537378787994, + 0.08004819601774216, + 0.1316731870174408, + 0.08477341383695602, + 0.03411759063601494, + 0.05528740957379341, + 0.06672222912311554, + 0.04311233386397362 + ], + "credit_map": [ + 0.01836620271205902, + 0.013713068328797817, + 0.011866666376590729, + 0.007941634394228458, + 0.01013686042279005, + 0.01644289121031761, + 0.020419485867023468, + 0.020445503294467926, + 0.028812414035201073, + 0.01728496327996254, + 0.03026132471859455, + 0.02119646593928337, + 0.026494616642594337, + 0.021513182669878006, + 0.010903439484536648, + 0.020161522552371025, + 0.015454822219908237, + 0.015527278184890747, + 0.07316357642412186, + 0.036934416741132736, + 0.009083817712962627, + 0.017818201333284378, + 0.011210735887289047, + 0.029112519696354866, + 0.008822156116366386, + 0.007727958261966705, + 0.017707251012325287, + 0.01654651015996933, + 0.0071562668308615685, + 0.00514500867575407, + 0.009178908541798592, + 0.007764135021716356, + 0.015452577732503414, + 0.00643180450424552, + 0.046941399574279785, + 0.025432039052248, + 0.007368001155555248, + 0.0051848990842700005, + 0.009483348578214645, + 0.015379111282527447, + 0.01045563817024231, + 0.007691841572523117, + 0.012927116826176643, + 0.01310829445719719, + 0.007290394511073828, + 0.013719419948756695, + 0.008923685178160667, + 0.010657018981873989, + 0.00761409942060709, + 0.006598814856261015, + 0.0038850416895002127, + 0.005689017008990049, + 0.003021467011421919, + 0.0025767595507204533, + 0.002352251671254635, + 0.0023801419883966446, + 0.007263041567057371, + 0.013140338473021984, + 0.007131782826036215, + 0.006115271709859371, + 0.004193444270640612, + 0.007952661253511906, + 0.003659025300294161, + 0.0058318451046943665, + 0.015919357538223267, + 0.01701436936855316, + 0.00637064129114151, + 0.007426305674016476, + 0.00395017396658659, + 0.00798451341688633, + 0.003652499057352543, + 0.004404370207339525, + 0.003059044247493148, + 0.004617339000105858, + 0.009604941122233868, + 0.008050475269556046, + 0.004104587249457836, + 0.004833947401493788, + 0.004007083363831043, + 0.004834914114326239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.004099062643945217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193536713719368 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193536713719368, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193536713719368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10442598164081573, + 0.12507547438144684, + 0.08986762166023254, + 0.11375673115253448, + 0.13646121323108673, + 0.05246160551905632, + 0.14902277290821075, + 0.06412359327077866, + 0.07997815310955048, + 0.08482685685157776 + ], + "credit_map": [ + 0.012838245369493961, + 0.011453433893620968, + 0.025317758321762085, + 0.010029408149421215, + 0.007343285717070103, + 0.01406507845968008, + 0.01628725230693817, + 0.007091514300554991, + 0.012234160676598549, + 0.011555950157344341, + 0.01751328445971012, + 0.015275496058166027, + 0.01736786961555481, + 0.015592417679727077, + 0.016999589279294014, + 0.018536709249019623, + 0.009162461385130882, + 0.009712242521345615, + 0.021777605637907982, + 0.01848614402115345, + 0.008884998969733715, + 0.008478539064526558, + 0.0051365504041314125, + 0.008229080587625504, + 0.013607258908450603, + 0.006979521829634905, + 0.020232608541846275, + 0.01718420721590519, + 0.012229803949594498, + 0.009109949693083763, + 0.011879072524607182, + 0.022534312680363655, + 0.01257183775305748, + 0.017268558964133263, + 0.025748470798134804, + 0.01894586719572544, + 0.01688339002430439, + 0.010194405913352966, + 0.019005771726369858, + 0.015842922031879425, + 0.005801258608698845, + 0.0038635157980024815, + 0.007056033238768578, + 0.006893415469676256, + 0.00815083459019661, + 0.0056058866903185844, + 0.008436311036348343, + 0.0066543505527079105, + 0.0332261323928833, + 0.029693158343434334, + 0.016987448558211327, + 0.025738902390003204, + 0.010849072597920895, + 0.010876157321035862, + 0.010531003586947918, + 0.011120899580419064, + 0.007301845122128725, + 0.006190067622810602, + 0.008894546888768673, + 0.011539223603904247, + 0.007755954284220934, + 0.005849403794854879, + 0.0063062384724617004, + 0.010286320000886917, + 0.01978413388133049, + 0.01989998109638691, + 0.007881991565227509, + 0.009503663517534733, + 0.005570431239902973, + 0.006218382623046637, + 0.005625758320093155, + 0.005493814591318369, + 0.008408649824559689, + 0.007861042395234108, + 0.017632920295000076, + 0.01447639986872673, + 0.010581960901618004, + 0.007904539816081524, + 0.008810730651021004, + 0.009150620549917221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.014193536713719368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776921942830086 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009776921942830086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776921942830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13968926668167114, + 0.094000905752182, + 0.16204819083213806, + 0.13683342933654785, + 0.08713177591562271, + 0.14830856025218964, + 0.03826593980193138, + 0.04589417949318886, + 0.10120895504951477, + 0.04661880433559418 + ], + "credit_map": [ + 0.017446497455239296, + 0.017391424626111984, + 0.020139135420322418, + 0.010826583951711655, + 0.01230546459555626, + 0.019199544563889503, + 0.018913131207227707, + 0.023467475548386574, + 0.008833498694002628, + 0.009861018508672714, + 0.013360325247049332, + 0.010279294103384018, + 0.01641956903040409, + 0.013828041031956673, + 0.01040540263056755, + 0.011013763025403023, + 0.01453506387770176, + 0.011819603852927685, + 0.034541793167591095, + 0.021036528050899506, + 0.01232863962650299, + 0.012690515257418156, + 0.018876871094107628, + 0.036219172179698944, + 0.019042616710066795, + 0.013172321952879429, + 0.02595192939043045, + 0.02426147647202015, + 0.012354630045592785, + 0.00676761893555522, + 0.021117445081472397, + 0.014165404252707958, + 0.009718764573335648, + 0.007082771975547075, + 0.024068690836429596, + 0.013692285865545273, + 0.006736847572028637, + 0.005002244841307402, + 0.007943157106637955, + 0.012887006625533104, + 0.01514111552387476, + 0.014384067617356777, + 0.019219854846596718, + 0.020257683470845222, + 0.023562870919704437, + 0.015330547466874123, + 0.02458103932440281, + 0.01583138108253479, + 0.006918824277818203, + 0.005922028794884682, + 0.004850725643336773, + 0.0053971922025084496, + 0.004925311543047428, + 0.0034175454638898373, + 0.003908958751708269, + 0.0029253552202135324, + 0.004041018430143595, + 0.005330216139554977, + 0.009249145165085793, + 0.010920718312263489, + 0.004163871053606272, + 0.0032363354694098234, + 0.00423972075805068, + 0.004713152535259724, + 0.025667238980531693, + 0.024809427559375763, + 0.011379729956388474, + 0.011162207461893559, + 0.006068799179047346, + 0.008233641274273396, + 0.006570123136043549, + 0.007317792624235153, + 0.005903645418584347, + 0.004832187667489052, + 0.010152692906558514, + 0.008256963454186916, + 0.004053232725709677, + 0.00438584852963686, + 0.005324151366949081, + 0.0037100862246006727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.009776921942830086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot begins with a tight close-up of a tense, muscular calf in the starting blocks, then quickly pulls back as the starting gun sounds. The athlete, an African man with a chiseled, determined face and short, dark hair, charges forward explosively. He wears a vibrant red sleeveless athletic vest made of lightweight, moisture-wicking fabric and a pair of tight black shorts. As the camera pulls back to a wide shot, it reveals the entire scene: a professional terracotta athletic track with crisp white lane markings, bathed in the brilliant, low afternoon sun. The camera then shifts to a smooth, rapid tracking shot, keeping pace with the leading runner as he enters the first turn of the 400-meter race. His form is impeccable; his arms pump rhythmically, his knees lift, and his body leans into the curve to counteract the force. Sweat gleams on his chiseled shoulders and back, catching the golden light and creating sharp highlights. In the background, the blurred forms of other runners struggle to keep up, while the stadium stands form a soft-focus mosaic of cheering crowds. With decisive power, he steadily passed runners in adjacent lanes, his expression of focus unwavering, creating an intense and exciting atmosphere of peak physical fitness and competitive advantage.", + "seed": 15671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_2/index_4901_seed_15671_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012457380257546901, + "vlm_logprob": -0.012457380257546901, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.012457380257546901, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027954179793596268 + ], + "opposite_logprob_tokens": [ + -3.7779541015625 + ], + "logprob_reward": -0.027954179793596268, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027954179793596268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32132112979888916, + 0.16167224943637848, + 0.09354564547538757, + 0.06627538800239563, + 0.03808553144335747, + 0.06890084594488144, + 0.08035788685083389, + 0.05764298141002655, + 0.05046433210372925, + 0.06173400953412056 + ], + "credit_map": [ + 0.04582533240318298, + 0.02561131864786148, + 0.033490926027297974, + 0.055767495185136795, + 0.027562085539102554, + 0.032692085951566696, + 0.05191662535071373, + 0.04845527932047844, + 0.01495509222149849, + 0.03645746782422066, + 0.02091696672141552, + 0.014653656631708145, + 0.020993119105696678, + 0.020754721015691757, + 0.01815081760287285, + 0.014790406450629234, + 0.00733284093439579, + 0.011332862079143524, + 0.007440182846039534, + 0.018603945150971413, + 0.010046016424894333, + 0.011733349412679672, + 0.016095008701086044, + 0.01096144039183855, + 0.0061138104647397995, + 0.00678040087223053, + 0.010688290931284428, + 0.005625342950224876, + 0.01144190039485693, + 0.00952756404876709, + 0.010607336647808552, + 0.005490745883435011, + 0.0034468616358935833, + 0.005993060301989317, + 0.003787957364693284, + 0.004092796705663204, + 0.004540644120424986, + 0.005692596081644297, + 0.00390770984813571, + 0.006623905152082443, + 0.005548401735723019, + 0.005559202749282122, + 0.004967199172824621, + 0.006137691903859377, + 0.01576559618115425, + 0.015773026272654533, + 0.009922079741954803, + 0.005227645859122276, + 0.008805342949926853, + 0.008758703246712685, + 0.012246187776327133, + 0.008093048818409443, + 0.00989153329282999, + 0.010558395646512508, + 0.010841473005712032, + 0.011163197457790375, + 0.009431588463485241, + 0.006327240262180567, + 0.006277387961745262, + 0.007861465215682983, + 0.006173329893499613, + 0.0056020296178758144, + 0.006729237735271454, + 0.009240702725946903, + 0.008918661624193192, + 0.007618105039000511, + 0.006541895214468241, + 0.0069963205605745316, + 0.004908094182610512, + 0.003966236021369696, + 0.004763038828969002, + 0.0067519815638661385, + 0.007557573728263378, + 0.006880773231387138, + 0.011361576616764069, + 0.008513900451362133, + 0.003564883954823017, + 0.007198094390332699, + 0.0038383188657462597, + 0.01281888596713543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.027954179793596268 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980360060930252 + ], + "opposite_logprob_tokens": [ + -4.634803771972656 + ], + "logprob_reward": -0.00980360060930252, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980360060930252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31566283106803894, + 0.030808081850409508, + 0.05868291109800339, + 0.10104801505804062, + 0.06555095314979553, + 0.04593595489859581, + 0.061000432819128036, + 0.15244893729686737, + 0.11802498251199722, + 0.0508369542658329 + ], + "credit_map": [ + 0.030834781005978584, + 0.024134153500199318, + 0.027254272252321243, + 0.04350666329264641, + 0.052525244653224945, + 0.03640799969434738, + 0.05480669066309929, + 0.046193040907382965, + 0.00344095123000443, + 0.006427350454032421, + 0.004143860656768084, + 0.0028809960931539536, + 0.003570098662748933, + 0.004187855403870344, + 0.002755270339548588, + 0.003401701571419835, + 0.005317240953445435, + 0.009017380885779858, + 0.005434065591543913, + 0.007895242422819138, + 0.009668446145951748, + 0.004022493958473206, + 0.009951567277312279, + 0.007376471534371376, + 0.009150337427854538, + 0.010104125365614891, + 0.010994555428624153, + 0.011986443772912025, + 0.01746349409222603, + 0.02175246737897396, + 0.011758497916162014, + 0.007838094606995583, + 0.005590747110545635, + 0.014767668209969997, + 0.007263496983796358, + 0.008175619877874851, + 0.009306206367909908, + 0.0062089236453175545, + 0.007103551644831896, + 0.0071347421035170555, + 0.003628868842497468, + 0.00549373310059309, + 0.0041345832869410515, + 0.004832866135984659, + 0.007324674166738987, + 0.01226335484534502, + 0.004970393143594265, + 0.0032874804455786943, + 0.004327676724642515, + 0.006449777167290449, + 0.005875282920897007, + 0.011683355085551739, + 0.009117365814745426, + 0.005327780265361071, + 0.010057540610432625, + 0.008161651901900768, + 0.01803804561495781, + 0.016485163941979408, + 0.015612020157277584, + 0.023388192057609558, + 0.022750038653612137, + 0.01374609675258398, + 0.02049810253083706, + 0.021931272000074387, + 0.022978147491812706, + 0.014560403302311897, + 0.018180452287197113, + 0.01118596363812685, + 0.012495412491261959, + 0.013065149076282978, + 0.011672912165522575, + 0.013886540196835995, + 0.004722631070762873, + 0.004907994065433741, + 0.00709633668884635, + 0.006211288273334503, + 0.007261688821017742, + 0.004722108133137226, + 0.006841736845672131, + 0.009073168970644474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.00980360060930252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255692172795534 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255692172795534, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255692172795534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1490478664636612, + 0.07534284144639969, + 0.0420873649418354, + 0.22727133333683014, + 0.06273864954710007, + 0.1335495263338089, + 0.03568919748067856, + 0.06677940487861633, + 0.06354647129774094, + 0.14394745230674744 + ], + "credit_map": [ + 0.012688941322267056, + 0.013488808646798134, + 0.010479873046278954, + 0.03110123611986637, + 0.01652504876255989, + 0.024226030334830284, + 0.019335893914103508, + 0.02120201662182808, + 0.007829161360859871, + 0.013546924106776714, + 0.009251530282199383, + 0.006154689472168684, + 0.00884178839623928, + 0.012674943543970585, + 0.009652117267251015, + 0.007391670718789101, + 0.003921814262866974, + 0.0031631281599402428, + 0.0035377724561840296, + 0.004237906541675329, + 0.004952185321599245, + 0.0063440375961363316, + 0.009566264227032661, + 0.006364254280924797, + 0.020735882222652435, + 0.012743859551846981, + 0.009312465786933899, + 0.008017662912607193, + 0.06204593554139137, + 0.09410206973552704, + 0.008110474795103073, + 0.012202960439026356, + 0.004756728652864695, + 0.009721147827804089, + 0.010650022886693478, + 0.009513902477920055, + 0.008101848885416985, + 0.00534221064299345, + 0.007531360257416964, + 0.0071214232593774796, + 0.006057949736714363, + 0.009098581038415432, + 0.005271830130368471, + 0.005198849830776453, + 0.025046858936548233, + 0.0530954971909523, + 0.011091591790318489, + 0.018688349053263664, + 0.0033339334186166525, + 0.0036628684028983116, + 0.005386654287576675, + 0.005039576906710863, + 0.0032762084156274796, + 0.0053151254542171955, + 0.004877831786870956, + 0.004796992987394333, + 0.008918666280806065, + 0.005631038453429937, + 0.007998744025826454, + 0.00965570006519556, + 0.008103998377919197, + 0.010253406129777431, + 0.006814445368945599, + 0.009403395466506481, + 0.008321075700223446, + 0.009474048390984535, + 0.008030036464333534, + 0.008636255748569965, + 0.006068344693630934, + 0.0053857299499213696, + 0.006441452540457249, + 0.011189522221684456, + 0.01704525575041771, + 0.023776018992066383, + 0.014544783160090446, + 0.01679917611181736, + 0.007661208510398865, + 0.023274967446923256, + 0.007499330677092075, + 0.03334668651223183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.005255692172795534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736662704497576 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736662704497576, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736662704497576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08609878271818161, + 0.0640513151884079, + 0.0582784004509449, + 0.16466891765594482, + 0.0638161450624466, + 0.15931086242198944, + 0.07161996513605118, + 0.09551719576120377, + 0.171157106757164, + 0.06548137217760086 + ], + "credit_map": [ + 0.00814854446798563, + 0.006701556965708733, + 0.005262718535959721, + 0.014239045791327953, + 0.012433384545147419, + 0.01456255279481411, + 0.011928100138902664, + 0.01282287947833538, + 0.005961093585938215, + 0.013476521708071232, + 0.010562891140580177, + 0.005980817601084709, + 0.00738705275580287, + 0.008586558513343334, + 0.006219412200152874, + 0.005876963026821613, + 0.008802447468042374, + 0.005652349442243576, + 0.0048958780243992805, + 0.006596182473003864, + 0.0066232578828930855, + 0.009334666654467583, + 0.008436227217316628, + 0.007937387563288212, + 0.01239002589136362, + 0.009066648781299591, + 0.009973984211683273, + 0.0067242770455777645, + 0.04276202619075775, + 0.06354796886444092, + 0.01013149879872799, + 0.010072492994368076, + 0.0056981598027050495, + 0.01676957868039608, + 0.006618700455874205, + 0.005071448627859354, + 0.008984130807220936, + 0.0061179883778095245, + 0.00542678264901042, + 0.00912935845553875, + 0.007406315766274929, + 0.01365653146058321, + 0.004894976504147053, + 0.006153088063001633, + 0.034477341920137405, + 0.06468551605939865, + 0.010474536567926407, + 0.017562558874487877, + 0.008376430720090866, + 0.006877039559185505, + 0.010846956633031368, + 0.007763779256492853, + 0.0091888178139925, + 0.009115577675402164, + 0.010783394798636436, + 0.008667965419590473, + 0.012671742588281631, + 0.009249852038919926, + 0.01121334359049797, + 0.015478833578526974, + 0.008301764726638794, + 0.008563581854104996, + 0.015318850986659527, + 0.014719229191541672, + 0.02947983704507351, + 0.021364901214838028, + 0.023905504494905472, + 0.019934896379709244, + 0.014237781055271626, + 0.013305357657372952, + 0.014997804537415504, + 0.033931028097867966, + 0.006974450778216124, + 0.0065431008115410805, + 0.009422382339835167, + 0.010358909144997597, + 0.01023348793387413, + 0.00523360725492239, + 0.007061644922941923, + 0.009653783403337002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.006736662704497576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a tight close-up of a muscular African man in starting blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a vibrant red sleeveless athletic vest and tight black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a professional terracotta athletic track with white lane markings under low afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pull back from a close-up to a wide shot and then shift to a smooth, rapid tracking shot following the runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the runner shown with sweat gleaming on his shoulders and back while pumping his arms and leaning into a curve?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253676600754261 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253676600754261, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253676600754261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10860346257686615, + 0.07518244534730911, + 0.10174910724163055, + 0.07158010452985764, + 0.04097634553909302, + 0.19372524321079254, + 0.09131849557161331, + 0.07945629209280014, + 0.1011524498462677, + 0.13625606894493103 + ], + "credit_map": [ + 0.01154323760420084, + 0.008694762364029884, + 0.009451841004192829, + 0.01602819375693798, + 0.017718948423862457, + 0.009539486840367317, + 0.019411947578191757, + 0.016215041279792786, + 0.007508662063628435, + 0.015909483656287193, + 0.00856283213943243, + 0.006307053379714489, + 0.009474908001720905, + 0.010582489892840385, + 0.009205148555338383, + 0.007631870452314615, + 0.008272427134215832, + 0.013862203806638718, + 0.009345396421849728, + 0.01574542373418808, + 0.01142891775816679, + 0.009648025035858154, + 0.021255193278193474, + 0.012191521003842354, + 0.006198336370289326, + 0.00764624634757638, + 0.007118391804397106, + 0.0050664739683270454, + 0.01579437032341957, + 0.018338507041335106, + 0.005952182691544294, + 0.005465592723339796, + 0.006141965277493, + 0.00491011468693614, + 0.0037133987061679363, + 0.004799359943717718, + 0.005256126634776592, + 0.006155010312795639, + 0.004546082112938166, + 0.005454286467283964, + 0.010920029133558273, + 0.013495816849172115, + 0.009846393950283527, + 0.008136214688420296, + 0.03196016699075699, + 0.07482703775167465, + 0.015255538746714592, + 0.029284046962857246, + 0.007859015837311745, + 0.007584280334413052, + 0.022481057792901993, + 0.010151884518563747, + 0.010751998052001, + 0.00852345023304224, + 0.013146655634045601, + 0.01082015410065651, + 0.008546384051442146, + 0.006492522079497576, + 0.01575779728591442, + 0.008217476308345795, + 0.006676694378256798, + 0.01712876744568348, + 0.007424614857882261, + 0.009212042205035686, + 0.016749275848269463, + 0.011739029549062252, + 0.014866863377392292, + 0.010707320645451546, + 0.009018153883516788, + 0.010901919566094875, + 0.009382813237607479, + 0.017787080258131027, + 0.013891750015318394, + 0.009552673436701298, + 0.014259557239711285, + 0.01396973431110382, + 0.011326014995574951, + 0.030737346038222313, + 0.011182880960404873, + 0.031336113810539246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.01253676600754261 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_tensor([8678, 8678])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_tensor([8678, 8678])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5a3212806c0e8c0f125902893d53b86afa5b64f7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_tensor([8678, 8678])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a slightly overhead angle, the camera remains still, focusing on a weathered male hand, its skin roughened, its knuckles thick with calluses, tightly gripping a traditional car key with a black plastic head. With a steady and decisive push of the wrist, the forearm muscles tense, and the key twists forcefully into the old car's ignition lock, making a crisp, mechanical click. Then, with a deep, powerful roar, the car's engine starts, sending a slight tremor through the entire body. Simultaneously, a layer of dust particles, visible in the slanting sunlight streaming into the car, dance slightly on the previously dim dashboard, and the speedometer and tachometer needles jump sharply before illuminating instantly along with various red and orange indicator lights, imbuing this gritty moment with vitality and creating a focused, powerful, and realistic atmosphere of impending departure.", + "seed": 2845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_8678_seed_2845_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07282469943165779, + "vlm_logprob": -0.07282469943165779, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.07282469943165779, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017283068969845772 + ], + "opposite_logprob_tokens": [ + -4.767282962799072 + ], + "logprob_reward": -0.017283068969845772, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.017283068969845772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1486295610666275, + 0.06659118086099625, + 0.04706713184714317, + 0.05481221526861191, + 0.0622154138982296, + 0.25666099786758423, + 0.04873881861567497, + 0.06261139363050461, + 0.19674459099769592, + 0.05592869594693184 + ], + "credit_map": [ + 0.01240526046603918, + 0.026095565408468246, + 0.023502277210354805, + 0.02510230243206024, + 0.01680873893201351, + 0.010593757964670658, + 0.021436871960759163, + 0.012684785760939121, + 0.010109932161867619, + 0.0103193623945117, + 0.010383575223386288, + 0.012016232125461102, + 0.0061547462828457355, + 0.005950464867055416, + 0.0064611067064106464, + 0.005195763427764177, + 0.00503250677138567, + 0.005078425630927086, + 0.008654134348034859, + 0.008127909153699875, + 0.005016009323298931, + 0.004139715805649757, + 0.00717552425339818, + 0.0038429100532084703, + 0.006092685274779797, + 0.006366769317537546, + 0.010595288127660751, + 0.010677238926291466, + 0.005334971006959677, + 0.0038556356448680162, + 0.00818993803113699, + 0.0036996921990066767, + 0.007588131353259087, + 0.007980494759976864, + 0.012002781964838505, + 0.00919605977833271, + 0.006100946106016636, + 0.006143323145806789, + 0.007866758853197098, + 0.005336925853043795, + 0.026417545974254608, + 0.04436781629920006, + 0.043600894510746, + 0.044428497552871704, + 0.0235606599599123, + 0.02482583373785019, + 0.02691974863409996, + 0.022540008649230003, + 0.0032024416141211987, + 0.005296821240335703, + 0.009969976730644703, + 0.012079861015081406, + 0.0060314941219985485, + 0.0029722528997808695, + 0.005863158497959375, + 0.0033228176180273294, + 0.003917028196156025, + 0.013117322698235512, + 0.011423975229263306, + 0.009172162041068077, + 0.007729473058134317, + 0.00598980812355876, + 0.0061916569247841835, + 0.005069969687610865, + 0.010158319026231766, + 0.07770736515522003, + 0.0481136292219162, + 0.014832491055130959, + 0.01976025104522705, + 0.005025830585509539, + 0.014022222720086575, + 0.007124488707631826, + 0.008276944980025291, + 0.008937530219554901, + 0.010826671496033669, + 0.004755343310534954, + 0.009320035576820374, + 0.00388332805596292, + 0.0057976581156253815, + 0.004131188616156578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.017283068969845772 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299058794975281 + ], + "opposite_logprob_tokens": [ + -2.237990617752075 + ], + "logprob_reward": -0.11299058794975281, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299058794975281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2649220824241638, + 0.06986313313245773, + 0.04746093600988388, + 0.03137267008423805, + 0.08372792601585388, + 0.23663119971752167, + 0.03442094475030899, + 0.06896817684173584, + 0.07114874571561813, + 0.09148415923118591 + ], + "credit_map": [ + 0.028106290847063065, + 0.04938178509473801, + 0.047269850969314575, + 0.029568063095211983, + 0.034010712057352066, + 0.0233813114464283, + 0.03514155000448227, + 0.018062490969896317, + 0.009243408218026161, + 0.013086577877402306, + 0.012433159165084362, + 0.010146473534405231, + 0.0056069218553602695, + 0.005193360615521669, + 0.007877708412706852, + 0.006275526713579893, + 0.00442355964332819, + 0.0050461310893297195, + 0.007447307463735342, + 0.009884854778647423, + 0.005046374164521694, + 0.0027369274757802486, + 0.0054784053936600685, + 0.007397376000881195, + 0.0027121875900775194, + 0.0032221022993326187, + 0.005708345677703619, + 0.006090544629842043, + 0.0033880940172821283, + 0.0018412823555991054, + 0.006040539592504501, + 0.002369570778682828, + 0.008901580236852169, + 0.015146839432418346, + 0.013637806288897991, + 0.010528285056352615, + 0.009297399781644344, + 0.009230340830981731, + 0.009355922229588032, + 0.007629750296473503, + 0.023727066814899445, + 0.04655295982956886, + 0.04056062549352646, + 0.038807112723588943, + 0.026280581951141357, + 0.021347707137465477, + 0.018493270501494408, + 0.020861852914094925, + 0.0027251476421952248, + 0.004287974443286657, + 0.006210006773471832, + 0.007122738752514124, + 0.004140367731451988, + 0.0020045689307153225, + 0.005207565147429705, + 0.0027225723024457693, + 0.005237892735749483, + 0.014931178651750088, + 0.01210692711174488, + 0.007386014796793461, + 0.008286467753350735, + 0.0083568524569273, + 0.006041898857802153, + 0.006620946805924177, + 0.0044481041841208935, + 0.02686888910830021, + 0.017060665413737297, + 0.0037572989240288734, + 0.006718158256262541, + 0.0020811385475099087, + 0.006401837803423405, + 0.003812646260485053, + 0.012393902987241745, + 0.015008071437478065, + 0.011234779842197895, + 0.006742609199136496, + 0.0173573587089777, + 0.011014102026820183, + 0.008390795439481735, + 0.009342538192868233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.11299058794975281 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592467859387398 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592467859387398, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592467859387398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12326237559318542, + 0.06310839205980301, + 0.07490096241235733, + 0.054491110146045685, + 0.0833773985505104, + 0.25740325450897217, + 0.022723430767655373, + 0.10248760133981705, + 0.15766014158725739, + 0.060585279017686844 + ], + "credit_map": [ + 0.011927652172744274, + 0.03274912387132645, + 0.022015506401658058, + 0.010728951543569565, + 0.011430883780121803, + 0.010181448422372341, + 0.012149098329246044, + 0.012079731561243534, + 0.008763477206230164, + 0.010726078413426876, + 0.01075226254761219, + 0.011165090836584568, + 0.005791792646050453, + 0.003452606499195099, + 0.00874848011881113, + 0.0037086093798279762, + 0.007420781999826431, + 0.00849475059658289, + 0.013413196429610252, + 0.013759983703494072, + 0.00729968911036849, + 0.004590669646859169, + 0.011208073236048222, + 0.008713824674487114, + 0.005314715206623077, + 0.006916377227753401, + 0.00971696712076664, + 0.011937215924263, + 0.004931716714054346, + 0.004256893415004015, + 0.006926488131284714, + 0.004490736871957779, + 0.006250482518225908, + 0.010513164103031158, + 0.014761727303266525, + 0.018067460507154465, + 0.00727301137521863, + 0.006832769140601158, + 0.013748560100793839, + 0.005930223502218723, + 0.02460336498916149, + 0.044178832322359085, + 0.05212199315428734, + 0.060109496116638184, + 0.02097022347152233, + 0.01646324433386326, + 0.02380828745663166, + 0.015147806145250797, + 0.001811044872738421, + 0.0036717071197927, + 0.003912795335054398, + 0.004227619152516127, + 0.0034089854452759027, + 0.0016427884111180902, + 0.0022130953148007393, + 0.0018353943014517426, + 0.008654356002807617, + 0.021543314680457115, + 0.017897922545671463, + 0.01779230684041977, + 0.009868926368653774, + 0.007435814943164587, + 0.011955112218856812, + 0.007339854724705219, + 0.007905172184109688, + 0.06554979830980301, + 0.04282993823289871, + 0.007969576865434647, + 0.01555546373128891, + 0.0035368960816413164, + 0.008720310404896736, + 0.0055930051021277905, + 0.006914688274264336, + 0.010704608634114265, + 0.007839495316147804, + 0.004694952163845301, + 0.012676212005317211, + 0.005624681711196899, + 0.006521842908114195, + 0.005608797073364258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.020592467859387398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032013058662415 + ], + "opposite_logprob_tokens": [ + -2.3503201007843018 + ], + "logprob_reward": -0.10032013058662415, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032013058662415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25924989581108093, + 0.07033688575029373, + 0.11356456577777863, + 0.052218083292245865, + 0.11826030164957047, + 0.06954498589038849, + 0.044199857860803604, + 0.11555446684360504, + 0.12565036118030548, + 0.03142060711979866 + ], + "credit_map": [ + 0.022435875609517097, + 0.06510275602340698, + 0.04892561957240105, + 0.027986746281385422, + 0.027525773271918297, + 0.02147645503282547, + 0.02293378859758377, + 0.022862860932946205, + 0.007825460284948349, + 0.012170790694653988, + 0.010894796811044216, + 0.01308494247496128, + 0.00656275637447834, + 0.0061301132664084435, + 0.00748774129897356, + 0.00618027476593852, + 0.014021612703800201, + 0.01470743678510189, + 0.019594082608819008, + 0.0243688952177763, + 0.00895001646131277, + 0.007507200352847576, + 0.014735999517142773, + 0.00967930257320404, + 0.006587629672139883, + 0.005100160371512175, + 0.010787849314510822, + 0.010536721907556057, + 0.004535623826086521, + 0.004146748222410679, + 0.007085886318236589, + 0.0034374562092125416, + 0.009257836267352104, + 0.0164337158203125, + 0.021544357761740685, + 0.02594226971268654, + 0.010754011571407318, + 0.007954472675919533, + 0.017573731020092964, + 0.008799891918897629, + 0.00676813255995512, + 0.007187468931078911, + 0.013598349876701832, + 0.015625832602381706, + 0.005680460017174482, + 0.006950294598937035, + 0.006981301587074995, + 0.006753136869519949, + 0.0043265013955533504, + 0.005459947977215052, + 0.007390682119876146, + 0.007303288206458092, + 0.005557917058467865, + 0.0038850645069032907, + 0.006087445188313723, + 0.004189005121588707, + 0.009813280776143074, + 0.027371544390916824, + 0.021399084478616714, + 0.019332097843289375, + 0.011902456171810627, + 0.005791071802377701, + 0.012820626609027386, + 0.007124290335923433, + 0.006023050751537085, + 0.048041779547929764, + 0.03148938715457916, + 0.009397686459124088, + 0.011926708742976189, + 0.003174351528286934, + 0.010124117136001587, + 0.005473280791193247, + 0.0029552674386650324, + 0.006298203486949205, + 0.006562345661222935, + 0.0023232235107570887, + 0.004428619518876076, + 0.0027336212806403637, + 0.002595437690615654, + 0.0035238827113062143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.10032013058662415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293724179267883 + ], + "opposite_logprob_tokens": [ + -2.2379372119903564 + ], + "logprob_reward": -0.11293724179267883, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293724179267883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12364920973777771, + 0.03775065392255783, + 0.04661723971366882, + 0.06825479865074158, + 0.12777391076087952, + 0.3548106253147125, + 0.06785262376070023, + 0.08197391778230667, + 0.04806330054998398, + 0.04325375705957413 + ], + "credit_map": [ + 0.012736435979604721, + 0.03305094689130783, + 0.02170148678123951, + 0.012933815829455853, + 0.013633656315505505, + 0.00753041822463274, + 0.012499703094363213, + 0.009562735445797443, + 0.006008705124258995, + 0.004323112778365612, + 0.005746225360780954, + 0.006788743659853935, + 0.00404752790927887, + 0.0024698027409613132, + 0.005449965130537748, + 0.0029165681917220354, + 0.00448575196787715, + 0.00465648528188467, + 0.008196369744837284, + 0.008605451323091984, + 0.004111419897526503, + 0.0040315971709787846, + 0.007539993617683649, + 0.0049901725724339485, + 0.006717999000102282, + 0.008807897567749023, + 0.011852812953293324, + 0.01475499477237463, + 0.006712679751217365, + 0.004803495015949011, + 0.009233703836798668, + 0.0053712171502411366, + 0.013300691731274128, + 0.021776532754302025, + 0.02291239984333515, + 0.02315516211092472, + 0.013319514691829681, + 0.008198386989533901, + 0.016260644420981407, + 0.008850578218698502, + 0.034235432744026184, + 0.07340016216039658, + 0.06435056030750275, + 0.07089758664369583, + 0.032049618661403656, + 0.02609967440366745, + 0.03202101215720177, + 0.02175658382475376, + 0.006373444572091103, + 0.009018044918775558, + 0.011083658784627914, + 0.013967353850603104, + 0.010343865491449833, + 0.0038462139200419188, + 0.008350219577550888, + 0.004869818687438965, + 0.006877103820443153, + 0.01079114992171526, + 0.014711408875882626, + 0.017323940992355347, + 0.009501705877482891, + 0.00361368409357965, + 0.01391388475894928, + 0.005241044331341982, + 0.005810825619846582, + 0.006756660062819719, + 0.005673361010849476, + 0.009209727868437767, + 0.006576886400580406, + 0.003714273450896144, + 0.004854739643633366, + 0.005466828588396311, + 0.004832081496715546, + 0.008796325884759426, + 0.007441654335707426, + 0.0030511196237057447, + 0.006288872566074133, + 0.003583837067708373, + 0.003529879031702876, + 0.005729989148676395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.11293724179267883 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly overhead angle, the camera remains still, focusing on a weathered male hand, its skin roughened, its knuckles thick with calluses, tightly gripping a traditional car key with a black plastic head. With a steady and decisive push of the wrist, the forearm muscles tense, and the key twists forcefully into the old car's ignition lock, making a crisp, mechanical click. Then, with a deep, powerful roar, the car's engine starts, sending a slight tremor through the entire body. Simultaneously, a layer of dust particles, visible in the slanting sunlight streaming into the car, dance slightly on the previously dim dashboard, and the speedometer and tachometer needles jump sharply before illuminating instantly along with various red and orange indicator lights, imbuing this gritty moment with vitality and creating a focused, powerful, and realistic atmosphere of impending departure.", + "seed": 2846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_8678_seed_2846_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11499924808740616, + "vlm_logprob": -0.11499924808740616, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.11499924808740616, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05965409427881241 + ], + "opposite_logprob_tokens": [ + -3.0596539974212646 + ], + "logprob_reward": -0.05965409427881241, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05965409427881241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3363981246948242, + 0.05990264564752579, + 0.029882559552788734, + 0.03182355314493179, + 0.01607624813914299, + 0.02778981812298298, + 0.045961931347846985, + 0.02146422117948532, + 0.03395882993936539, + 0.3967421054840088 + ], + "credit_map": [ + 0.03492414206266403, + 0.022247377783060074, + 0.10137461125850677, + 0.02762284316122532, + 0.010528082959353924, + 0.05755285918712616, + 0.0217859148979187, + 0.06036229059100151, + 0.00508119398728013, + 0.0038118911907076836, + 0.00832652859389782, + 0.011281142942607403, + 0.003271780675277114, + 0.007974468171596527, + 0.011192302219569683, + 0.00896333809942007, + 0.0032097860239446163, + 0.0014877674402669072, + 0.005860190372914076, + 0.005284718703478575, + 0.001886536949314177, + 0.002691225614398718, + 0.004985758103430271, + 0.004476578440517187, + 0.002400417812168598, + 0.002186232479289174, + 0.0035204822197556496, + 0.005184720270335674, + 0.0018241918878629804, + 0.004038475453853607, + 0.006819299887865782, + 0.005849732086062431, + 0.0010865749791264534, + 0.0010463632643222809, + 0.0016411540564149618, + 0.0034168469719588757, + 0.0006941036554053426, + 0.0026467577554285526, + 0.002331074560061097, + 0.003213373478502035, + 0.0027971495874226093, + 0.0021187160164117813, + 0.0023577131796628237, + 0.0033914607483893633, + 0.002203514100983739, + 0.005836365278810263, + 0.0049333334900438786, + 0.0041515640914440155, + 0.0039773802272975445, + 0.0033866181038320065, + 0.006832842715084553, + 0.0033408235758543015, + 0.0035921423695981503, + 0.010501841083168983, + 0.0060293469578027725, + 0.008300934918224812, + 0.0016952501609921455, + 0.0020068809390068054, + 0.0027218565810471773, + 0.0026133933570235968, + 0.001686801202595234, + 0.0025596690829843283, + 0.004645534325391054, + 0.0035348362289369106, + 0.004463929682970047, + 0.002899468643590808, + 0.007576742209494114, + 0.0024405675940215588, + 0.001933792605996132, + 0.004232626408338547, + 0.0033717863261699677, + 0.007039919495582581, + 0.06663165986537933, + 0.031315114349126816, + 0.09705772995948792, + 0.12978693842887878, + 0.010783904232084751, + 0.013901473954319954, + 0.03147769346833229, + 0.015787608921527863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.05965409427881241 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10039582848548889 + ], + "opposite_logprob_tokens": [ + -2.350395917892456 + ], + "logprob_reward": -0.10039582848548889, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10039582848548889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09520208090543747, + 0.0526193305850029, + 0.019366580992937088, + 0.0222398079931736, + 0.024166107177734375, + 0.028194451704621315, + 0.05074101686477661, + 0.03968265280127525, + 0.038413334637880325, + 0.629374623298645 + ], + "credit_map": [ + 0.0073231784626841545, + 0.005407677497714758, + 0.01206272840499878, + 0.009958124719560146, + 0.008701993152499199, + 0.02324008010327816, + 0.011805277317762375, + 0.016703028231859207, + 0.004127139691263437, + 0.0036130771040916443, + 0.008906645700335503, + 0.007307449355721474, + 0.002332834992557764, + 0.011588416993618011, + 0.005742440931499004, + 0.009001333266496658, + 0.0015786525327712297, + 0.0015498852590098977, + 0.0021863828878849745, + 0.0020665761549025774, + 0.0008829119033180177, + 0.00466781435534358, + 0.002769737970083952, + 0.003664621850475669, + 0.0015437490073964, + 0.0011417719069868326, + 0.0021925007458776236, + 0.003509911010041833, + 0.0012817970709875226, + 0.004003421403467655, + 0.004437214229255915, + 0.004129445180296898, + 0.0016926429234445095, + 0.0012974933488294482, + 0.0029600909911096096, + 0.003983008675277233, + 0.0016543224919587374, + 0.004287465009838343, + 0.00410535279661417, + 0.004185734316706657, + 0.002781220944598317, + 0.0017651690868660808, + 0.0033733102027326822, + 0.003570407396182418, + 0.00211283378303051, + 0.005040150135755539, + 0.005157493986189365, + 0.004393869079649448, + 0.004402712918817997, + 0.004466652870178223, + 0.004439701326191425, + 0.002925026696175337, + 0.005351899657398462, + 0.014583118259906769, + 0.0059651946648955345, + 0.008606715127825737, + 0.0027268733829259872, + 0.003186961403116584, + 0.007558108307421207, + 0.002435685833916068, + 0.004863743204623461, + 0.007711568847298622, + 0.004154321271926165, + 0.007045396137982607, + 0.0027644489891827106, + 0.003874995745718479, + 0.006179667077958584, + 0.0030293005984276533, + 0.004396369680762291, + 0.006403706967830658, + 0.005895489826798439, + 0.00586936017498374, + 0.10505686700344086, + 0.04403248429298401, + 0.15901288390159607, + 0.20991656184196472, + 0.01634664461016655, + 0.02147349901497364, + 0.04901102930307388, + 0.024524636566638947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.10039582848548889 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060437947511673 + ], + "opposite_logprob_tokens": [ + -3.895604372024536 + ], + "logprob_reward": -0.02060437947511673, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060437947511673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12286471575498581, + 0.08196160197257996, + 0.030632279813289642, + 0.030400466173887253, + 0.02057383768260479, + 0.05371618643403053, + 0.07020414620637894, + 0.041823860257864, + 0.05826639011502266, + 0.48955652117729187 + ], + "credit_map": [ + 0.01433323323726654, + 0.010282511822879314, + 0.021656984463334084, + 0.015112179331481457, + 0.006095965392887592, + 0.02138170599937439, + 0.015129800885915756, + 0.018872344866394997, + 0.005612443666905165, + 0.0058451928198337555, + 0.013180035166442394, + 0.007367881014943123, + 0.004796452820301056, + 0.018159352242946625, + 0.012822775170207024, + 0.014177478849887848, + 0.0029877687338739634, + 0.0016380415763705969, + 0.006409934256225824, + 0.0041425046510994434, + 0.0014312256826087832, + 0.005488981492817402, + 0.004116344265639782, + 0.004417481832206249, + 0.0019491686252877116, + 0.001200013910420239, + 0.00207714457064867, + 0.004371037241071463, + 0.0016189569141715765, + 0.007088146638125181, + 0.005551507230848074, + 0.00654449500143528, + 0.0016945399111136794, + 0.0010635030921548605, + 0.002454704837873578, + 0.004256878513842821, + 0.001195825170725584, + 0.0031471149995923042, + 0.0029186077881604433, + 0.003842664184048772, + 0.0037329220212996006, + 0.003942304756492376, + 0.004255000967532396, + 0.005675649270415306, + 0.004060416016727686, + 0.014116698876023293, + 0.008730231784284115, + 0.009202965535223484, + 0.006838061381131411, + 0.006021508947014809, + 0.012101776897907257, + 0.005001415964215994, + 0.0060365223325788975, + 0.012750687077641487, + 0.009364351630210876, + 0.012089826166629791, + 0.002618294209241867, + 0.003403980517759919, + 0.006149862892925739, + 0.004095925018191338, + 0.003279381198808551, + 0.006314229220151901, + 0.0089294807985425, + 0.007032712921500206, + 0.005568474065512419, + 0.006321191322058439, + 0.014279014430940151, + 0.003858459647744894, + 0.004589166026562452, + 0.008158514276146889, + 0.00534782325848937, + 0.010143755003809929, + 0.08213599771261215, + 0.030095946043729782, + 0.12115846574306488, + 0.16321446001529694, + 0.01579534262418747, + 0.01827848143875599, + 0.03894682973623276, + 0.019931016489863396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.02060437947511673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28138190507888794 + ], + "opposite_logprob_tokens": [ + -1.4063818454742432 + ], + "logprob_reward": -0.28138190507888794, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28138190507888794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20769071578979492, + 0.027245357632637024, + 0.05277799442410469, + 0.023381318897008896, + 0.024307100102305412, + 0.046270664781332016, + 0.03325822949409485, + 0.05686509609222412, + 0.059106141328811646, + 0.46909743547439575 + ], + "credit_map": [ + 0.018950093537569046, + 0.01307586394250393, + 0.059919413179159164, + 0.020629091188311577, + 0.006160689517855644, + 0.0368613600730896, + 0.015003642998635769, + 0.03709058091044426, + 0.0021441562566906214, + 0.0021512380335479975, + 0.005989174358546734, + 0.0036298525519669056, + 0.0010319786379113793, + 0.004566509276628494, + 0.003446955466642976, + 0.004285493399947882, + 0.005333371460437775, + 0.003413262078538537, + 0.009268388152122498, + 0.008142332546412945, + 0.0030255357269197702, + 0.008245861157774925, + 0.007651814725250006, + 0.007697428576648235, + 0.0016153345350176096, + 0.0011169114150106907, + 0.002577462000772357, + 0.004318294581025839, + 0.0012343357084318995, + 0.0032941633835434914, + 0.004691798705607653, + 0.004533017985522747, + 0.0016929416451603174, + 0.001156260841526091, + 0.002588855102658272, + 0.005818905308842659, + 0.0012701785890385509, + 0.003059657523408532, + 0.004434351343661547, + 0.004285948816686869, + 0.003017487470060587, + 0.002806458855047822, + 0.00614927290007472, + 0.00567557942122221, + 0.00349871045909822, + 0.00746543426066637, + 0.009883993305265903, + 0.007773728109896183, + 0.002790565835312009, + 0.0025501796044409275, + 0.004050506744533777, + 0.001967700431123376, + 0.003160416381433606, + 0.009091202169656754, + 0.004023130517452955, + 0.005624528042972088, + 0.004476272035390139, + 0.004655114375054836, + 0.007212517783045769, + 0.00330005562864244, + 0.005948272068053484, + 0.01415200810879469, + 0.00733788451179862, + 0.00978296808898449, + 0.004862112458795309, + 0.005158937070518732, + 0.013218151405453682, + 0.004852995742112398, + 0.004514803644269705, + 0.009225505404174328, + 0.007562798447906971, + 0.009710838086903095, + 0.07699283957481384, + 0.02851763926446438, + 0.11625797301530838, + 0.15356579422950745, + 0.017586776986718178, + 0.0192693043500185, + 0.035521283745765686, + 0.02138582430779934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.28138190507888794 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296003311872482 + ], + "opposite_logprob_tokens": [ + -2.23796010017395 + ], + "logprob_reward": -0.11296003311872482, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296003311872482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1321742683649063, + 0.058110129088163376, + 0.06794504821300507, + 0.02276618219912052, + 0.087611123919487, + 0.06006298214197159, + 0.056508876383304596, + 0.0814262181520462, + 0.04558940231800079, + 0.3878058195114136 + ], + "credit_map": [ + 0.012197115458548069, + 0.007830158807337284, + 0.02036092057824135, + 0.024403465911746025, + 0.006094282492995262, + 0.025898760184645653, + 0.01638299599289894, + 0.01900656521320343, + 0.0062643615528941154, + 0.004083690699189901, + 0.014053146354854107, + 0.008945913054049015, + 0.002174549736082554, + 0.007662765681743622, + 0.006815435364842415, + 0.008110260590910912, + 0.006309684831649065, + 0.004769363906234503, + 0.011466668918728828, + 0.011177968233823776, + 0.0035384399816393852, + 0.009179321117699146, + 0.0116054005920887, + 0.009898186661303043, + 0.0017886522691696882, + 0.0013600020902231336, + 0.0022764240857213736, + 0.003394871950149536, + 0.0012328090379014611, + 0.00490208063274622, + 0.0031833460088819265, + 0.0046279942616820335, + 0.005556437186896801, + 0.004549389239400625, + 0.0075875381007790565, + 0.01571902260184288, + 0.005107399076223373, + 0.01676870696246624, + 0.015349717810750008, + 0.01697291061282158, + 0.005814043339341879, + 0.0042303442023694515, + 0.009692788124084473, + 0.007499708794057369, + 0.005061292089521885, + 0.006545313633978367, + 0.012004854157567024, + 0.009214629419147968, + 0.004755999427288771, + 0.004513541702181101, + 0.008204999379813671, + 0.0028351652435958385, + 0.005541452672332525, + 0.01342339999973774, + 0.007228081580251455, + 0.010006229393184185, + 0.005879547912627459, + 0.0069130160845816135, + 0.015625953674316406, + 0.0066484627313911915, + 0.004253219813108444, + 0.010370621457695961, + 0.01672111451625824, + 0.01501427125185728, + 0.003511691465973854, + 0.003457089653238654, + 0.012790819630026817, + 0.0031115536112338305, + 0.002713405294343829, + 0.0058620092459023, + 0.006121933460235596, + 0.008020889945328236, + 0.06391799449920654, + 0.023155853152275085, + 0.10070554912090302, + 0.13031786680221558, + 0.011840730905532837, + 0.013254212215542793, + 0.028815139085054398, + 0.0157984159886837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.11296003311872482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly overhead angle, the camera remains still, focusing on a weathered male hand, its skin roughened, its knuckles thick with calluses, tightly gripping a traditional car key with a black plastic head. With a steady and decisive push of the wrist, the forearm muscles tense, and the key twists forcefully into the old car's ignition lock, making a crisp, mechanical click. Then, with a deep, powerful roar, the car's engine starts, sending a slight tremor through the entire body. Simultaneously, a layer of dust particles, visible in the slanting sunlight streaming into the car, dance slightly on the previously dim dashboard, and the speedometer and tachometer needles jump sharply before illuminating instantly along with various red and orange indicator lights, imbuing this gritty moment with vitality and creating a focused, powerful, and realistic atmosphere of impending departure.", + "seed": 2847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_8678_seed_2847_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11756695881485939, + "vlm_logprob": -0.11756695881485939, + "vlm_yes_no_margin": 2.174999976158142, + "vlm_token_credit": -0.11756695881485939, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14593538641929626 + ], + "opposite_logprob_tokens": [ + -2.020935297012329 + ], + "logprob_reward": -0.14593538641929626, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14593538641929626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08607938885688782, + 0.0614120252430439, + 0.035920411348342896, + 0.1156889796257019, + 0.024520864710211754, + 0.026473378762602806, + 0.07703503966331482, + 0.08833608776330948, + 0.25856417417526245, + 0.22596970200538635 + ], + "credit_map": [ + 0.007186063099652529, + 0.00841803289949894, + 0.013271245174109936, + 0.008588965982198715, + 0.009880192577838898, + 0.016492638736963272, + 0.009967475198209286, + 0.01227477565407753, + 0.00487173767760396, + 0.00538917351514101, + 0.011493225581943989, + 0.007728309370577335, + 0.007790866773575544, + 0.006092258729040623, + 0.00794965960085392, + 0.01009678840637207, + 0.0026986398734152317, + 0.0026147463358938694, + 0.008294842205941677, + 0.004408790729939938, + 0.0038167370948940516, + 0.004027338232845068, + 0.0029968214221298695, + 0.0070624942891299725, + 0.008960563689470291, + 0.00858690682798624, + 0.02084195241332054, + 0.013263539411127567, + 0.013402423821389675, + 0.012223943136632442, + 0.011812697164714336, + 0.026596952229738235, + 0.0019938971381634474, + 0.0020288582891225815, + 0.0039038502145558596, + 0.0031972869765013456, + 0.003726147348061204, + 0.0020022448152303696, + 0.0028349540662020445, + 0.004833622835576534, + 0.002806133357807994, + 0.0016749214846640825, + 0.0039623393677175045, + 0.0027743366081267595, + 0.0030944852624088526, + 0.0031343884766101837, + 0.0037951467093080282, + 0.00523162679746747, + 0.01114669255912304, + 0.003711109049618244, + 0.01396838016808033, + 0.006416090298444033, + 0.010652170516550541, + 0.007172434590756893, + 0.006589153315871954, + 0.01737901009619236, + 0.005640143062919378, + 0.007146782707422972, + 0.02845839224755764, + 0.0059774890542030334, + 0.012710463255643845, + 0.007115404587239027, + 0.009363704361021519, + 0.011923703365027905, + 0.012830371968448162, + 0.014417449943721294, + 0.01797543838620186, + 0.0726039931178093, + 0.07928769290447235, + 0.022562677040696144, + 0.02732188068330288, + 0.011564645916223526, + 0.021315496414899826, + 0.03223780542612076, + 0.02743574045598507, + 0.031473614275455475, + 0.02550066076219082, + 0.027537085115909576, + 0.02738136239349842, + 0.03308793902397156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.14593538641929626 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08913836628198624 + ], + "opposite_logprob_tokens": [ + -2.4641382694244385 + ], + "logprob_reward": -0.08913836628198624, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08913836628198624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20827026665210724, + 0.127802312374115, + 0.02384413219988346, + 0.05276280269026756, + 0.025951076298952103, + 0.046481840312480927, + 0.055950768291950226, + 0.020944708958268166, + 0.2062036544084549, + 0.23178842663764954 + ], + "credit_map": [ + 0.011992624029517174, + 0.021638521924614906, + 0.03401004150509834, + 0.019511135295033455, + 0.02118370309472084, + 0.02544800564646721, + 0.025525052100419998, + 0.048961181193590164, + 0.008734670467674732, + 0.012576342560350895, + 0.02728024311363697, + 0.01498494390398264, + 0.015228215605020523, + 0.011556489393115044, + 0.014930648729205132, + 0.022510763257741928, + 0.0016938664484769106, + 0.0017404137179255486, + 0.004792449064552784, + 0.003076585242524743, + 0.003379291156306863, + 0.0026226176414638758, + 0.002755577675998211, + 0.0037833310198038816, + 0.003956363536417484, + 0.004349317401647568, + 0.00838348176330328, + 0.007006013300269842, + 0.007303872145712376, + 0.005822944454848766, + 0.005669746082276106, + 0.010271064937114716, + 0.0019385537598282099, + 0.0015069689834490418, + 0.004692505579441786, + 0.002828865544870496, + 0.0024892461951822042, + 0.0026304954662919044, + 0.002799911890178919, + 0.007064528297632933, + 0.004595689941197634, + 0.0029332726262509823, + 0.007870747707784176, + 0.004797138273715973, + 0.006619774736464024, + 0.0041326554492115974, + 0.006146648433059454, + 0.009385915473103523, + 0.004531064536422491, + 0.0032323601190000772, + 0.009393970482051373, + 0.004329265095293522, + 0.007852427661418915, + 0.009614778682589531, + 0.005829968024045229, + 0.011166934855282307, + 0.0018025344470515847, + 0.0023779470939189196, + 0.002752496162429452, + 0.0017458960646763444, + 0.002655707998201251, + 0.0029365329537540674, + 0.004269151482731104, + 0.0024044427555054426, + 0.016283607110381126, + 0.014880739152431488, + 0.021649688482284546, + 0.06309574842453003, + 0.04146257042884827, + 0.018990591168403625, + 0.01636473834514618, + 0.013475964777171612, + 0.019294679164886475, + 0.028547387570142746, + 0.04002295434474945, + 0.02928204834461212, + 0.022278888151049614, + 0.027889784425497055, + 0.02794954553246498, + 0.03652314841747284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.08913836628198624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208984926342964 + ], + "opposite_logprob_tokens": [ + -2.8120899200439453 + ], + "logprob_reward": -0.06208984926342964, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208984926342964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11151697486639023, + 0.16882112622261047, + 0.04116014018654823, + 0.020024729892611504, + 0.03443276509642601, + 0.03194873780012131, + 0.06392399966716766, + 0.04572723060846329, + 0.165638729929924, + 0.31680554151535034 + ], + "credit_map": [ + 0.0070514073595404625, + 0.009782714769244194, + 0.018472865223884583, + 0.01193889882415533, + 0.011227265000343323, + 0.02144734561443329, + 0.012274295091629028, + 0.019322186708450317, + 0.010805053636431694, + 0.010401885956525803, + 0.052626919001340866, + 0.012394863180816174, + 0.014227485284209251, + 0.013146482408046722, + 0.010530728846788406, + 0.04468771442770958, + 0.003039125120267272, + 0.0033793356269598007, + 0.009715541265904903, + 0.005248351488262415, + 0.005091101862490177, + 0.003966929856687784, + 0.003936402965337038, + 0.006783354561775923, + 0.0013092565350234509, + 0.001456387690268457, + 0.0031133312731981277, + 0.00169644586276263, + 0.001984401373192668, + 0.002558543812483549, + 0.0030672710854560137, + 0.004839093424379826, + 0.0026415844913572073, + 0.0028160694055259228, + 0.006092812400311232, + 0.0041495938785374165, + 0.004797755274921656, + 0.00341921579092741, + 0.0033371897879987955, + 0.007178544532507658, + 0.0044721276499331, + 0.002503242576494813, + 0.004864968825131655, + 0.004987139254808426, + 0.0037742597050964832, + 0.003006316488608718, + 0.003964080475270748, + 0.004376602824777365, + 0.006860597990453243, + 0.003790328511968255, + 0.010412953794002533, + 0.005818638484925032, + 0.008662430569529533, + 0.009488716721534729, + 0.006993175484240055, + 0.011897160671651363, + 0.00392650393769145, + 0.0032428214326500893, + 0.015784332528710365, + 0.0030231636483222246, + 0.0046687740832567215, + 0.0037101428024470806, + 0.004893741570413113, + 0.0064777503721416, + 0.008345787413418293, + 0.012309621088206768, + 0.01705900765955448, + 0.04689004644751549, + 0.0302655640989542, + 0.02274382673203945, + 0.019036540761590004, + 0.008988319896161556, + 0.030510446056723595, + 0.03260856866836548, + 0.04327382147312164, + 0.04172424226999283, + 0.027400795370340347, + 0.04910847544670105, + 0.02928243950009346, + 0.0628967434167862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.06208984926342964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016032636165619 + ], + "opposite_logprob_tokens": [ + -1.7016032934188843 + ], + "logprob_reward": -0.2016032636165619, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016032636165619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1327066570520401, + 0.0589950755238533, + 0.04159241542220116, + 0.05046940967440605, + 0.0347408689558506, + 0.03734361380338669, + 0.13454079627990723, + 0.06590543687343597, + 0.05298258364200592, + 0.3907231390476227 + ], + "credit_map": [ + 0.008278720080852509, + 0.011807412840425968, + 0.021308861672878265, + 0.01583757996559143, + 0.016896354034543037, + 0.021415121853351593, + 0.015929879620671272, + 0.021232716739177704, + 0.004694520961493254, + 0.0039012127090245485, + 0.013524293899536133, + 0.007764267735183239, + 0.006675709038972855, + 0.00551633071154356, + 0.0056111253798007965, + 0.01130761206150055, + 0.00286459201015532, + 0.003137819701805711, + 0.009297878481447697, + 0.005885140039026737, + 0.005122595001012087, + 0.004516169894486666, + 0.0037326025776565075, + 0.007035622373223305, + 0.004212045576423407, + 0.004152141977101564, + 0.009294627234339714, + 0.006123121362179518, + 0.006586012430489063, + 0.006010206416249275, + 0.005351727362722158, + 0.008739530108869076, + 0.002858042949810624, + 0.002848271047696471, + 0.005933038424700499, + 0.003839423879981041, + 0.00416989391669631, + 0.003680682508274913, + 0.003514533629640937, + 0.007896983996033669, + 0.005796188488602638, + 0.0023824747186154127, + 0.005165722221136093, + 0.003505697241052985, + 0.005834668874740601, + 0.0038203808944672346, + 0.0046445769257843494, + 0.006193905603140593, + 0.01239174883812666, + 0.007169011980295181, + 0.022731641307473183, + 0.015721065923571587, + 0.01916699856519699, + 0.01620449684560299, + 0.016473673284053802, + 0.024682166054844856, + 0.006306608207523823, + 0.007287442684173584, + 0.013302820734679699, + 0.0053279767744243145, + 0.007269617170095444, + 0.006940213963389397, + 0.013570415787398815, + 0.005900347139686346, + 0.005318371579051018, + 0.006729215383529663, + 0.007350271102041006, + 0.008664692752063274, + 0.004333907272666693, + 0.006945641245692968, + 0.009397827088832855, + 0.004242657218128443, + 0.025651078671216965, + 0.04914224520325661, + 0.060393549501895905, + 0.04044102504849434, + 0.03585226461291313, + 0.0752626359462738, + 0.03292929381132126, + 0.07105106115341187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.2016032636165619 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906792849302292 + ], + "opposite_logprob_tokens": [ + -2.4640679359436035 + ], + "logprob_reward": -0.08906792849302292, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906792849302292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12928394973278046, + 0.1534276008605957, + 0.04018362611532211, + 0.1198105439543724, + 0.04431029036641121, + 0.03429993987083435, + 0.06521014869213104, + 0.10391446202993393, + 0.10960423946380615, + 0.1999550759792328 + ], + "credit_map": [ + 0.006013971753418446, + 0.008544908836483955, + 0.017270881682634354, + 0.010813241824507713, + 0.009358327835798264, + 0.01186535693705082, + 0.0305805504322052, + 0.03483672812581062, + 0.01078474335372448, + 0.011128838174045086, + 0.0392603799700737, + 0.01360215712338686, + 0.014969144016504288, + 0.014140976592898369, + 0.014526333659887314, + 0.0350150428712368, + 0.0026459256187081337, + 0.0021615384612232447, + 0.010743393562734127, + 0.005923760123550892, + 0.004445329774171114, + 0.0023956364020705223, + 0.00402695219963789, + 0.007841094397008419, + 0.008322474546730518, + 0.0082399882376194, + 0.02198716439306736, + 0.012756790965795517, + 0.0138551639392972, + 0.012468460947275162, + 0.012963873334228992, + 0.029216637834906578, + 0.0031048764940351248, + 0.0026123877614736557, + 0.008470570668578148, + 0.0036887216847389936, + 0.005887424573302269, + 0.00466541200876236, + 0.004514213651418686, + 0.011366689577698708, + 0.003581221215426922, + 0.0021693904418498278, + 0.005200811196118593, + 0.00374027737416327, + 0.004510920960456133, + 0.0033798047807067633, + 0.004248918499797583, + 0.007468597963452339, + 0.00912808533757925, + 0.0032231116201728582, + 0.011003479361534119, + 0.008028081618249416, + 0.007643823511898518, + 0.006951252929866314, + 0.009815189987421036, + 0.009417127817869186, + 0.008959370665252209, + 0.010478229261934757, + 0.03461917117238045, + 0.00741236237809062, + 0.009607329964637756, + 0.008792192675173283, + 0.009023998863995075, + 0.01502180751413107, + 0.005848044995218515, + 0.008804867044091225, + 0.008336073718965054, + 0.012915373779833317, + 0.03716069087386131, + 0.012414832599461079, + 0.014448900707066059, + 0.009675468318164349, + 0.014952048659324646, + 0.025308936834335327, + 0.029470833018422127, + 0.030074525624513626, + 0.022088175639510155, + 0.024510232731699944, + 0.02041294053196907, + 0.03313740715384483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.08906792849302292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly overhead angle, the camera remains still, focusing on a weathered male hand, its skin roughened, its knuckles thick with calluses, tightly gripping a traditional car key with a black plastic head. With a steady and decisive push of the wrist, the forearm muscles tense, and the key twists forcefully into the old car's ignition lock, making a crisp, mechanical click. Then, with a deep, powerful roar, the car's engine starts, sending a slight tremor through the entire body. Simultaneously, a layer of dust particles, visible in the slanting sunlight streaming into the car, dance slightly on the previously dim dashboard, and the speedometer and tachometer needles jump sharply before illuminating instantly along with various red and orange indicator lights, imbuing this gritty moment with vitality and creating a focused, powerful, and realistic atmosphere of impending departure.", + "seed": 2848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_3/index_8678_seed_2848_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1100334219634533, + "vlm_logprob": -0.1100334219634533, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.1100334219634533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12948183715343475 + ], + "opposite_logprob_tokens": [ + -2.129481792449951 + ], + "logprob_reward": -0.12948183715343475, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12948183715343475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10889866948127747, + 0.049776624888181686, + 0.030210262164473534, + 0.04846605286002159, + 0.15024405717849731, + 0.16955165565013885, + 0.036140602082014084, + 0.29917293787002563, + 0.07269958406686783, + 0.03483955189585686 + ], + "credit_map": [ + 0.012256034649908543, + 0.006411866284906864, + 0.017170235514640808, + 0.011607986874878407, + 0.007037687115371227, + 0.013827702030539513, + 0.018026141449809074, + 0.022561026737093925, + 0.00517469784244895, + 0.003304556244984269, + 0.006366882938891649, + 0.004385366570204496, + 0.004250312224030495, + 0.00884660892188549, + 0.006333569530397654, + 0.011114634573459625, + 0.0034630075097084045, + 0.001953144557774067, + 0.005228997673839331, + 0.00509283784776926, + 0.0026355173904448748, + 0.002188847865909338, + 0.005003111902624369, + 0.004644799046218395, + 0.004489796236157417, + 0.0021916234400123358, + 0.009482599794864655, + 0.00452291639521718, + 0.006661060731858015, + 0.011948591098189354, + 0.004870781674981117, + 0.0042986879125237465, + 0.017682110890746117, + 0.0169919915497303, + 0.01634783111512661, + 0.016894984990358353, + 0.018961435183882713, + 0.02186119370162487, + 0.024635350331664085, + 0.016869155690073967, + 0.01154671423137188, + 0.014360180124640465, + 0.026022590696811676, + 0.024678854271769524, + 0.013505513779819012, + 0.020818423479795456, + 0.03206302225589752, + 0.026556359604001045, + 0.0021961901802569628, + 0.0025868937373161316, + 0.006169995293021202, + 0.006290528457611799, + 0.0037317012902349234, + 0.0047120507806539536, + 0.005585188046097755, + 0.004868054296821356, + 0.022326674312353134, + 0.018814632669091225, + 0.05531749874353409, + 0.050120193511247635, + 0.04092836380004883, + 0.04967223107814789, + 0.02105867490172386, + 0.04093468561768532, + 0.00478263758122921, + 0.00865242537111044, + 0.008248534984886646, + 0.012001902796328068, + 0.013378199189901352, + 0.007247603498399258, + 0.013305604457855225, + 0.005082671996206045, + 0.003172919386997819, + 0.0049391211941838264, + 0.003522086190059781, + 0.004210760351270437, + 0.004616024903953075, + 0.0047577968798577785, + 0.005156240426003933, + 0.004464603960514069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.12948183715343475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17999112606048584 + ], + "opposite_logprob_tokens": [ + -1.8049911260604858 + ], + "logprob_reward": -0.17999112606048584, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17999112606048584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0514364056289196, + 0.024500606581568718, + 0.024418488144874573, + 0.032343197613954544, + 0.048662230372428894, + 0.19075043499469757, + 0.13096822798252106, + 0.3450072705745697, + 0.09347691386938095, + 0.058436162769794464 + ], + "credit_map": [ + 0.007556685246527195, + 0.00339845335111022, + 0.0067650331184268, + 0.006887530907988548, + 0.0030529708601534367, + 0.006879440508782864, + 0.0075594414956867695, + 0.009336854331195354, + 0.0024671100545674562, + 0.0016987556591629982, + 0.0027975335251539946, + 0.0029213090892881155, + 0.002185317687690258, + 0.0037592926528304815, + 0.003458047052845359, + 0.005213242024183273, + 0.0034496705047786236, + 0.0023864309769123793, + 0.004040859639644623, + 0.0038579946849495173, + 0.0027646911330521107, + 0.001437038998119533, + 0.003830264788120985, + 0.0026515384670346975, + 0.004148964770138264, + 0.002154166344553232, + 0.005216081626713276, + 0.004745701793581247, + 0.0037148308474570513, + 0.00503413425758481, + 0.003977700136601925, + 0.0033516227267682552, + 0.006390668917447329, + 0.0047210464254021645, + 0.006722042802721262, + 0.00501272501423955, + 0.00535346195101738, + 0.007203410845249891, + 0.007725053932517767, + 0.00553382420912385, + 0.013190786354243755, + 0.01678282767534256, + 0.028331812471151352, + 0.028940783813595772, + 0.015573758631944656, + 0.02449873834848404, + 0.03489769622683525, + 0.028534051030874252, + 0.007709565106779337, + 0.009819544851779938, + 0.01899079792201519, + 0.01917620189487934, + 0.017597543075680733, + 0.019209789112210274, + 0.020092299208045006, + 0.018372491002082825, + 0.027776220813393593, + 0.020210495218634605, + 0.0700598880648613, + 0.05176246538758278, + 0.04658244550228119, + 0.05754794180393219, + 0.024419840425252914, + 0.04664803668856621, + 0.006057752296328545, + 0.006832421757280827, + 0.0063418191857635975, + 0.011408274061977863, + 0.01693907007575035, + 0.008809863589704037, + 0.029638169333338737, + 0.0074495431035757065, + 0.006550920661538839, + 0.008475164882838726, + 0.005698957946151495, + 0.006881031673401594, + 0.007803717162460089, + 0.007658144924789667, + 0.008168753236532211, + 0.007199480663985014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.17999112606048584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869258776307106 + ], + "opposite_logprob_tokens": [ + -3.0486927032470703 + ], + "logprob_reward": -0.04869258776307106, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869258776307106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05630282312631607, + 0.03041674941778183, + 0.024227438494563103, + 0.03240959346294403, + 0.045911893248558044, + 0.20444846153259277, + 0.05200677737593651, + 0.32637128233909607, + 0.1739949733018875, + 0.053909942507743835 + ], + "credit_map": [ + 0.007312796078622341, + 0.003507317043840885, + 0.008680780418217182, + 0.007666565012186766, + 0.003904949640855193, + 0.0063342503271996975, + 0.009428703226149082, + 0.00946746114641428, + 0.0033429700415581465, + 0.001844796002842486, + 0.0037241175305098295, + 0.0035794980358332396, + 0.0024582152254879475, + 0.004897556267678738, + 0.003931940998882055, + 0.0066376542672514915, + 0.0014919234672561288, + 0.001799974124878645, + 0.004436146002262831, + 0.004498728085309267, + 0.002774428576231003, + 0.0031372900120913982, + 0.0028329121414572, + 0.0032560345716774464, + 0.0032850990537554026, + 0.0016002659685909748, + 0.006530646234750748, + 0.0034399046562612057, + 0.004314453341066837, + 0.007548601366579533, + 0.0030620864126831293, + 0.0026285366620868444, + 0.005853591021150351, + 0.0047740512527525425, + 0.005979188717901707, + 0.00526062585413456, + 0.006108531728386879, + 0.006315549369901419, + 0.00681100320070982, + 0.004809353034943342, + 0.013911638408899307, + 0.017684021964669228, + 0.030803203582763672, + 0.03043145313858986, + 0.016553930938243866, + 0.026048986241221428, + 0.037854887545108795, + 0.031160330399870872, + 0.005205975845456123, + 0.0031816361006349325, + 0.01449638418853283, + 0.005759812891483307, + 0.005867785774171352, + 0.0038890524301677942, + 0.005246578250080347, + 0.00835955049842596, + 0.024884000420570374, + 0.01909210905432701, + 0.06478116661310196, + 0.05185852572321892, + 0.04369090124964714, + 0.05453024432063103, + 0.022775346413254738, + 0.04475899040699005, + 0.013083347119390965, + 0.019355598837137222, + 0.028000572696328163, + 0.030638322234153748, + 0.02807818166911602, + 0.01404012180864811, + 0.029348501935601234, + 0.011450326070189476, + 0.004226755350828171, + 0.0073705133982002735, + 0.005936162080615759, + 0.0056677572429180145, + 0.0075819846242666245, + 0.006689163856208324, + 0.007638678885996342, + 0.008798927068710327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.04869258776307106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902513444423676 + ], + "opposite_logprob_tokens": [ + -2.5790250301361084 + ], + "logprob_reward": -0.07902513444423676, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902513444423676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04911411181092262, + 0.016109097748994827, + 0.02132382243871689, + 0.19707639515399933, + 0.051742881536483765, + 0.15786178410053253, + 0.12260682880878448, + 0.07832445204257965, + 0.24963007867336273, + 0.05621059983968735 + ], + "credit_map": [ + 0.006076410878449678, + 0.004505169112235308, + 0.006282833404839039, + 0.007457092870026827, + 0.005179259460419416, + 0.004460012074559927, + 0.008724105544388294, + 0.006429221481084824, + 0.001895327353850007, + 0.001472726697102189, + 0.00211891857907176, + 0.0016461623599752784, + 0.0010815117275342345, + 0.002645354950800538, + 0.0022178436629474163, + 0.0030312512535601854, + 0.0023685593623667955, + 0.0014725885121151805, + 0.003993573132902384, + 0.004387079272419214, + 0.002167763886973262, + 0.001436319900676608, + 0.0029859282076358795, + 0.002512008184567094, + 0.014645365998148918, + 0.006273542530834675, + 0.046867627650499344, + 0.016585135832428932, + 0.031711090356111526, + 0.062020376324653625, + 0.007321801967918873, + 0.011651435866951942, + 0.006940183229744434, + 0.0037847997155040503, + 0.007290012203156948, + 0.005942116025835276, + 0.003790762508288026, + 0.004879737738519907, + 0.012931355275213718, + 0.006183908320963383, + 0.010813361033797264, + 0.013547366484999657, + 0.024801025167107582, + 0.0230615995824337, + 0.012708986178040504, + 0.020167745649814606, + 0.028548581525683403, + 0.024213096126914024, + 0.007611753419041634, + 0.008247816003859043, + 0.02175942249596119, + 0.016620222479104996, + 0.012924420647323132, + 0.01688927784562111, + 0.019887110218405724, + 0.018666792660951614, + 0.006292616482824087, + 0.004619229584932327, + 0.01524959784001112, + 0.01333807222545147, + 0.010156050324440002, + 0.012630533427000046, + 0.005401046015322208, + 0.010637297295033932, + 0.01585789956152439, + 0.030303871259093285, + 0.03176024183630943, + 0.049977075308561325, + 0.03927820175886154, + 0.023254519328475, + 0.041817404329776764, + 0.017380837351083755, + 0.009223833680152893, + 0.009353792294859886, + 0.00780699634924531, + 0.005571155808866024, + 0.004659437108784914, + 0.0054368008859455585, + 0.007145185489207506, + 0.007013394962996244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.07902513444423676 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a weathered male hand with rough skin and thick calloused knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand shown tightly gripping a traditional car key with a black plastic head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand perform a decisive wrist push to twist the key into an ignition lock, resulting in a mechanical click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car engine start with a roar, causing dust particles to dance in slanting sunlight while dashboard needles jump and red/orange lights illuminate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead angle and remains still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297642439603806 + ], + "opposite_logprob_tokens": [ + -2.237976312637329 + ], + "logprob_reward": -0.11297642439603806, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297642439603806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0433260053396225, + 0.040224265307188034, + 0.03431476652622223, + 0.2597249150276184, + 0.03301103040575981, + 0.03495099022984505, + 0.2731477916240692, + 0.0994708240032196, + 0.06914713233709335, + 0.11268225312232971 + ], + "credit_map": [ + 0.006918136030435562, + 0.003052724292501807, + 0.005782794672995806, + 0.004979992285370827, + 0.0037136096507310867, + 0.004857101012021303, + 0.007116841152310371, + 0.006904805079102516, + 0.0025130566209554672, + 0.0028385394252836704, + 0.005041474476456642, + 0.005306199658662081, + 0.004892426542937756, + 0.004503101576119661, + 0.00880356878042221, + 0.006325900088995695, + 0.003973107784986496, + 0.003984442912042141, + 0.006533617619425058, + 0.0071080531924963, + 0.0029751616530120373, + 0.003449937328696251, + 0.003157565835863352, + 0.003132882062345743, + 0.01907222904264927, + 0.00848457496613264, + 0.062050286680459976, + 0.022417942062020302, + 0.04147801175713539, + 0.08078133314847946, + 0.009905274957418442, + 0.01553528942167759, + 0.0037654731422662735, + 0.0026455277111381292, + 0.005438313353806734, + 0.0036906821187585592, + 0.0032262762542814016, + 0.003807462751865387, + 0.0059720915742218494, + 0.004465203266590834, + 0.003795943921431899, + 0.0016975848702713847, + 0.005432227160781622, + 0.0038080434314906597, + 0.002211630344390869, + 0.003170195734128356, + 0.00881237257272005, + 0.006022990215569735, + 0.017825067043304443, + 0.01754431426525116, + 0.049382418394088745, + 0.037466831505298615, + 0.029920948669314384, + 0.03781605884432793, + 0.04320603236556053, + 0.03998613730072975, + 0.011032329872250557, + 0.008413386531174183, + 0.020731598138809204, + 0.011297797784209251, + 0.014905756339430809, + 0.014753877185285091, + 0.006642915774136782, + 0.011693155393004417, + 0.0038711486849933863, + 0.00845386367291212, + 0.004209539387375116, + 0.01579592190682888, + 0.011217515915632248, + 0.006518085021525621, + 0.012672598473727703, + 0.0064084529876708984, + 0.017552439123392105, + 0.016909504309296608, + 0.01849948801100254, + 0.012080756016075611, + 0.010815637186169624, + 0.011768624186515808, + 0.011511282995343208, + 0.013544522225856781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.11297642439603806 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_4/index_tensor([7539, 7539])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_4/index_tensor([7539, 7539])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c22ce5e4de8a54cafd80166e0171b1364fda99fc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_4/index_tensor([7539, 7539])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle close-up shot, the camera held still, focuses deep within a masonry hearth. In an instant, a flame erupts from the bottom of the frame, its core a dazzling, incandescent white, rapidly transitioning to vibrant orange and deep red. The flames twist and twirl violently in the air, forming a fluid, translucent body of energy. As this surge of power erupts, countless fiery sparks are swept upward. These tiny, bright golden particles scatter against the near-black background, leaving fleeting, brilliant trails like a miniature meteor shower erupting upward. The flames, the sole source of light, cast a dancing, flickering glow, illuminating the rough, smoke-blackened masonry within and casting rapidly shifting, deep shadows all around. The image is imbued with a sense of raw power and fiery dynamic beauty, creating an atmosphere both warm and awe-inspiring.", + "seed": 16514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_4/index_7539_seed_16514_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21211829585954547, + "vlm_logprob": -0.21211829585954547, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.21211829585954547, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06723692268133163 + ], + "opposite_logprob_tokens": [ + -2.81723690032959 + ], + "logprob_reward": -0.06723692268133163, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06723692268133163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11258494853973389, + 0.053218115121126175, + 0.09322059899568558, + 0.1411413550376892, + 0.18451131880283356, + 0.23392537236213684, + 0.03966141492128372, + 0.05325136333703995, + 0.03899615257978439, + 0.04948939383029938 + ], + "credit_map": [ + 0.02347077988088131, + 0.014973470009863377, + 0.023341398686170578, + 0.006072145886719227, + 0.011619167402386665, + 0.019139016047120094, + 0.007102638948708773, + 0.006866332143545151, + 0.007004397921264172, + 0.005643360782414675, + 0.01043672114610672, + 0.004860555287450552, + 0.005349088925868273, + 0.008154905401170254, + 0.004320948850363493, + 0.0074481419287621975, + 0.01678205095231533, + 0.008761666715145111, + 0.010976341553032398, + 0.019778205081820488, + 0.00487073278054595, + 0.015347628854215145, + 0.007046865299344063, + 0.009657103568315506, + 0.0256646741181612, + 0.01525897067040205, + 0.030832894146442413, + 0.00986718013882637, + 0.010931556113064289, + 0.023476963862776756, + 0.009527319110929966, + 0.015581795945763588, + 0.06739238649606705, + 0.00725627550855279, + 0.010307030752301216, + 0.004857509862631559, + 0.00680370070040226, + 0.07042594254016876, + 0.008816189132630825, + 0.008652282878756523, + 0.03658762574195862, + 0.04297793284058571, + 0.05941934138536453, + 0.02159380540251732, + 0.016778603196144104, + 0.01729828678071499, + 0.01565847173333168, + 0.023611314594745636, + 0.006903734989464283, + 0.004312606528401375, + 0.0066433255560696125, + 0.0033903084695339203, + 0.002534617902711034, + 0.005317411385476589, + 0.0031930021941661835, + 0.00736641138792038, + 0.006610804703086615, + 0.006416303105652332, + 0.007557966746389866, + 0.005612947512418032, + 0.0032416097819805145, + 0.0058615743182599545, + 0.0038090047892183065, + 0.014141149818897247, + 0.006694691255688667, + 0.008711586706340313, + 0.004722998943179846, + 0.005430926568806171, + 0.0037788022309541702, + 0.004096967168152332, + 0.0028399229049682617, + 0.002720255171880126, + 0.008161135949194431, + 0.005241862498223782, + 0.014246649108827114, + 0.005954884923994541, + 0.0031267208978533745, + 0.004810449667274952, + 0.0029626907780766487, + 0.0049850004725158215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.06723692268133163 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743343591690063 + ], + "opposite_logprob_tokens": [ + -0.47433435916900635 + ], + "logprob_reward": -0.9743343591690063, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743343591690063, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09572276473045349, + 0.05969833582639694, + 0.05604579299688339, + 0.053864870220422745, + 0.5320691466331482, + 0.017772605642676353, + 0.029442764818668365, + 0.05334886536002159, + 0.08501097559928894, + 0.017023850232362747 + ], + "credit_map": [ + 0.027257952839136124, + 0.012487568892538548, + 0.010863328352570534, + 0.006074912380427122, + 0.011181665584445, + 0.017985282465815544, + 0.00560913048684597, + 0.004262933973222971, + 0.006191391497850418, + 0.005754698533564806, + 0.007251432165503502, + 0.005040976218879223, + 0.006023072637617588, + 0.008478415198624134, + 0.006054171361029148, + 0.014904183335602283, + 0.008625224232673645, + 0.007401449605822563, + 0.007828773930668831, + 0.005936064291745424, + 0.004262784495949745, + 0.010326804593205452, + 0.0037495375145226717, + 0.007915160618722439, + 0.007840325124561787, + 0.004436299204826355, + 0.014020867645740509, + 0.006829934660345316, + 0.004725843667984009, + 0.007188659626990557, + 0.003951162099838257, + 0.004871780052781105, + 0.18834303319454193, + 0.022868141531944275, + 0.03616938367486, + 0.015688074752688408, + 0.021526925265789032, + 0.19478446245193481, + 0.027084939181804657, + 0.025604205206036568, + 0.002549862489104271, + 0.00309200887568295, + 0.0033471323549747467, + 0.001980474917218089, + 0.0018083560280501842, + 0.0016363715985789895, + 0.0019141013035550714, + 0.0014442994724959135, + 0.006574068684130907, + 0.0035017316695302725, + 0.004649736452847719, + 0.0028576559852808714, + 0.0016709305346012115, + 0.0053728544153273106, + 0.0014809940475970507, + 0.003334795357659459, + 0.0066822064109146595, + 0.0072684381157159805, + 0.008334331214427948, + 0.005912190768867731, + 0.005108670797199011, + 0.005677076056599617, + 0.005814445670694113, + 0.008551510982215405, + 0.015566698275506496, + 0.02602989226579666, + 0.009140437468886375, + 0.008582324720919132, + 0.008141250349581242, + 0.00663988059386611, + 0.005376375745981932, + 0.0055341278202831745, + 0.0026411949656903744, + 0.0022948465775698423, + 0.003098356770351529, + 0.00279816472902894, + 0.0015214913291856647, + 0.0018233705777674913, + 0.0012777283554896712, + 0.0015686992555856705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.9743343591690063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024946779012680054 + ], + "opposite_logprob_tokens": [ + -6.002494812011719 + ], + "logprob_reward": -0.0024946779012680054, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024946779012680054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.207441508769989, + 0.16149607300758362, + 0.07031595706939697, + 0.10079963505268097, + 0.2446882277727127, + 0.028862446546554565, + 0.06830310821533203, + 0.041435301303863525, + 0.04199310392141342, + 0.03466462343931198 + ], + "credit_map": [ + 0.04878391698002815, + 0.021904848515987396, + 0.029289772734045982, + 0.015086078085005283, + 0.03096558339893818, + 0.03542057424783707, + 0.01422087848186493, + 0.011769864708185196, + 0.01824532262980938, + 0.016889946535229683, + 0.02042621374130249, + 0.014485514722764492, + 0.013796063140034676, + 0.023911388590931892, + 0.015612863004207611, + 0.03812875226140022, + 0.01242847926914692, + 0.007996467873454094, + 0.017322013154625893, + 0.005269505083560944, + 0.005659036338329315, + 0.010989904403686523, + 0.003534494200721383, + 0.0071160634979605675, + 0.017453299835324287, + 0.008339857682585716, + 0.027891533449292183, + 0.009671970270574093, + 0.0064527615904808044, + 0.014732727780938148, + 0.005473916418850422, + 0.010783566161990166, + 0.08816447854042053, + 0.008300635032355785, + 0.014288883656263351, + 0.00978387612849474, + 0.007472335826605558, + 0.09384149312973022, + 0.011938273906707764, + 0.010898233391344547, + 0.004324019420892, + 0.006651401985436678, + 0.005855794530361891, + 0.0028366632759571075, + 0.001929048215970397, + 0.002584317233413458, + 0.0021227181423455477, + 0.002558484673500061, + 0.01382511481642723, + 0.006236708257347345, + 0.010705125518143177, + 0.00596671411767602, + 0.0040334551595151424, + 0.011530122719705105, + 0.005553212948143482, + 0.010452660731971264, + 0.006231002509593964, + 0.00389684597030282, + 0.0061446912586688995, + 0.004461480770260096, + 0.0027391240000724792, + 0.005232165567576885, + 0.004293385893106461, + 0.008436605334281921, + 0.0067222523503005505, + 0.009708109311759472, + 0.004607673268765211, + 0.005151451099663973, + 0.004633304663002491, + 0.0035102751571685076, + 0.004836604930460453, + 0.002823431743308902, + 0.005169203504920006, + 0.0030956061091274023, + 0.007017625495791435, + 0.0074397423304617405, + 0.002208481077104807, + 0.002924933098256588, + 0.0023393023293465376, + 0.004469729494303465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0024946779012680054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978483073413372 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.00978483073413372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978483073413372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.608971893787384, + 0.017704075202345848, + 0.029024582356214523, + 0.0348229855298996, + 0.15878956019878387, + 0.01657034642994404, + 0.025390926748514175, + 0.018312042579054832, + 0.04827774688601494, + 0.04213586822152138 + ], + "credit_map": [ + 0.18036271631717682, + 0.055013857781887054, + 0.07547075301408768, + 0.0324387364089489, + 0.0828171968460083, + 0.12114030867815018, + 0.03570684418082237, + 0.026021458208560944, + 0.0023023004177957773, + 0.0019028154201805592, + 0.0030067237094044685, + 0.001237287768162787, + 0.0025166093837469816, + 0.0026377327740192413, + 0.0015678014606237411, + 0.002532804384827614, + 0.00594964437186718, + 0.002763490891084075, + 0.0066400556825101376, + 0.002300326945260167, + 0.001326533849351108, + 0.004334212746471167, + 0.00156966014765203, + 0.004140657838433981, + 0.005804731510579586, + 0.003104365663602948, + 0.008667733520269394, + 0.002766940277069807, + 0.002656548982486129, + 0.005152600817382336, + 0.0023086739238351583, + 0.004361391067504883, + 0.060053445398807526, + 0.004833705723285675, + 0.009162561036646366, + 0.0027088806964457035, + 0.0051939832046628, + 0.062487076967954636, + 0.0072938515804708, + 0.007056056521832943, + 0.0024453727528452873, + 0.002837196458131075, + 0.003310876665636897, + 0.0016927600372582674, + 0.0011091806227341294, + 0.001504028681665659, + 0.0015768903540447354, + 0.0020940410904586315, + 0.005664462223649025, + 0.0021552471444010735, + 0.004816889762878418, + 0.0022860250901430845, + 0.001280844327993691, + 0.0034733491484075785, + 0.0019497621105983853, + 0.003764346009120345, + 0.001961997477337718, + 0.001416924991644919, + 0.002670734655112028, + 0.0020075293723493814, + 0.0012585612712427974, + 0.0018625163938850164, + 0.0016450874973088503, + 0.0054886904545128345, + 0.008667903952300549, + 0.014679307118058205, + 0.004967962391674519, + 0.004839963745325804, + 0.004803527612239122, + 0.0042517357505857944, + 0.0029953971970826387, + 0.0030719484202563763, + 0.0057813613675534725, + 0.00520715769380331, + 0.007009172812104225, + 0.009713692590594292, + 0.0026424063835293055, + 0.0037709344178438187, + 0.00281651527620852, + 0.005194624420255423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.00978483073413372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067406888119876385 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.0067406888119876385, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067406888119876385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30278533697128296, + 0.029065264388918877, + 0.10516949743032455, + 0.14475812017917633, + 0.1132868230342865, + 0.11487561464309692, + 0.021511003375053406, + 0.032535456120967865, + 0.10526946932077408, + 0.030743427574634552 + ], + "credit_map": [ + 0.09119570255279541, + 0.030089542269706726, + 0.03764944523572922, + 0.015611535869538784, + 0.03995492681860924, + 0.05873158574104309, + 0.016656504943966866, + 0.012896099127829075, + 0.004002091009169817, + 0.00327320653013885, + 0.005837674718350172, + 0.0023425414692610502, + 0.003129858523607254, + 0.003929120022803545, + 0.0029135020449757576, + 0.003637268440797925, + 0.016026390716433525, + 0.010793053545057774, + 0.017335591837763786, + 0.018380869179964066, + 0.006789084989577532, + 0.014092610217630863, + 0.008229998871684074, + 0.013521899469196796, + 0.025258347392082214, + 0.014875133521854877, + 0.032355230301618576, + 0.00908845104277134, + 0.01290067471563816, + 0.023975390940904617, + 0.009648822247982025, + 0.016656067222356796, + 0.03975408524274826, + 0.005060909315943718, + 0.007639609277248383, + 0.004629077855497599, + 0.004535838961601257, + 0.04053608700633049, + 0.005834347568452358, + 0.0052968659438192844, + 0.01751725934445858, + 0.023671383038163185, + 0.028332188725471497, + 0.011066284961998463, + 0.007673186715692282, + 0.008097820915281773, + 0.007652910426259041, + 0.01086458284407854, + 0.0030680361669510603, + 0.0031833366956561804, + 0.0030410708859562874, + 0.003012773348018527, + 0.0020657700952142477, + 0.003007537918165326, + 0.0016723291482776403, + 0.002460150746628642, + 0.00409522233530879, + 0.0027930589858442545, + 0.004540270660072565, + 0.003259423654526472, + 0.0019243686692789197, + 0.003412065329030156, + 0.003248911816626787, + 0.009262134321033955, + 0.019341187551617622, + 0.029769500717520714, + 0.01190841943025589, + 0.010556420311331749, + 0.010247799567878246, + 0.0073967594653368, + 0.007358270697295666, + 0.008691110648214817, + 0.004574317019432783, + 0.007815529592335224, + 0.004760654643177986, + 0.0025657149963080883, + 0.0028894050046801567, + 0.0035003190860152245, + 0.0016442312626168132, + 0.002993256552144885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0067406888119876385 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up shot, the camera held still, focuses deep within a masonry hearth. In an instant, a flame erupts from the bottom of the frame, its core a dazzling, incandescent white, rapidly transitioning to vibrant orange and deep red. The flames twist and twirl violently in the air, forming a fluid, translucent body of energy. As this surge of power erupts, countless fiery sparks are swept upward. These tiny, bright golden particles scatter against the near-black background, leaving fleeting, brilliant trails like a miniature meteor shower erupting upward. The flames, the sole source of light, cast a dancing, flickering glow, illuminating the rough, smoke-blackened masonry within and casting rapidly shifting, deep shadows all around. The image is imbued with a sense of raw power and fiery dynamic beauty, creating an atmosphere both warm and awe-inspiring.", + "seed": 16515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_4/index_7539_seed_16515_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.295395917352289, + "vlm_logprob": -0.295395917352289, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.295395917352289, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04779325798153877 + ], + "opposite_logprob_tokens": [ + -3.29779314994812 + ], + "logprob_reward": -0.04779325798153877, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04779325798153877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17508001625537872, + 0.056341640651226044, + 0.0577499195933342, + 0.030592745169997215, + 0.06139832362532616, + 0.08734406530857086, + 0.05164739489555359, + 0.08273066580295563, + 0.28014105558395386, + 0.1169741228222847 + ], + "credit_map": [ + 0.02419639565050602, + 0.02037617191672325, + 0.029832717031240463, + 0.02957959473133087, + 0.019124500453472137, + 0.023868996649980545, + 0.01330128125846386, + 0.014800366945564747, + 0.008486464619636536, + 0.007239961065351963, + 0.01162160187959671, + 0.005799496080726385, + 0.008499358780682087, + 0.006599797867238522, + 0.0033878013491630554, + 0.004707159474492073, + 0.01068721991032362, + 0.009600801393389702, + 0.009250746108591557, + 0.007144410628825426, + 0.006279400084167719, + 0.005458541214466095, + 0.004415273666381836, + 0.004913530312478542, + 0.004816878121346235, + 0.004355687648057938, + 0.004035839810967445, + 0.0034560994245111942, + 0.002877322491258383, + 0.004749976098537445, + 0.0033547761850059032, + 0.0029461688827723265, + 0.006868790835142136, + 0.007834450341761112, + 0.01187153160572052, + 0.011370126157999039, + 0.00625583715736866, + 0.0085486164316535, + 0.003972826059907675, + 0.004676145501434803, + 0.007681629154831171, + 0.00932357832789421, + 0.0151438619941473, + 0.010636826977133751, + 0.012592031620442867, + 0.011030130088329315, + 0.011925406754016876, + 0.009010599926114082, + 0.005892762448638678, + 0.00721541466191411, + 0.008108423091471195, + 0.006184771191328764, + 0.010252108797430992, + 0.006494996137917042, + 0.003865390084683895, + 0.0036335301119834185, + 0.011038877069950104, + 0.01148991659283638, + 0.019377142190933228, + 0.007400414440780878, + 0.008379554376006126, + 0.009989737533032894, + 0.008905262686312199, + 0.006149761378765106, + 0.03786502406001091, + 0.029474016278982162, + 0.046568967401981354, + 0.04702102392911911, + 0.015504073351621628, + 0.05923433601856232, + 0.016147470101714134, + 0.02832615189254284, + 0.018697792664170265, + 0.01601480320096016, + 0.019973764196038246, + 0.01607384718954563, + 0.010273740626871586, + 0.01247934065759182, + 0.010417243465781212, + 0.013043592683970928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.04779325798153877 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063581228256226 + ], + "opposite_logprob_tokens": [ + -0.28135815262794495 + ], + "logprob_reward": -1.4063581228256226, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063581228256226, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1188851073384285, + 0.044356390833854675, + 0.07804492861032486, + 0.08124176412820816, + 0.13468103110790253, + 0.060604359954595566, + 0.0540330670773983, + 0.2399434596300125, + 0.11989080905914307, + 0.06831904500722885 + ], + "credit_map": [ + 0.026575308293104172, + 0.012464997358620167, + 0.015393111854791641, + 0.018966255709528923, + 0.009641163051128387, + 0.012670092284679413, + 0.008914845995604992, + 0.014259342104196548, + 0.006905584130436182, + 0.0077909380197525024, + 0.009448091499507427, + 0.0034548176918178797, + 0.003974461928009987, + 0.0059532239101827145, + 0.002592000411823392, + 0.004237275570631027, + 0.009206825867295265, + 0.014145128428936005, + 0.012231212109327316, + 0.010548352263867855, + 0.010211467742919922, + 0.006185605656355619, + 0.006448257248848677, + 0.009068083018064499, + 0.013163487426936626, + 0.009097008965909481, + 0.012600906193256378, + 0.013541842810809612, + 0.006781487260013819, + 0.009194701910018921, + 0.007399769965559244, + 0.009462567046284676, + 0.018239611759781837, + 0.027456285431981087, + 0.017065394669771194, + 0.01779964379966259, + 0.01552418153733015, + 0.01299122255295515, + 0.012869168072938919, + 0.012735538184642792, + 0.0068644266575574875, + 0.006287657655775547, + 0.01398103591054678, + 0.007423728704452515, + 0.007004313636571169, + 0.009378488175570965, + 0.004367966204881668, + 0.0052967434749007225, + 0.0050483825616538525, + 0.005845554638653994, + 0.014604058116674423, + 0.004032708238810301, + 0.0081632724031806, + 0.0070441472344100475, + 0.004907515365630388, + 0.0043874261900782585, + 0.045418355613946915, + 0.026622645556926727, + 0.06944414973258972, + 0.026866072788834572, + 0.017122993245720863, + 0.016591522842645645, + 0.019757697358727455, + 0.018120039254426956, + 0.014264469966292381, + 0.015767309814691544, + 0.02921634539961815, + 0.017037317156791687, + 0.0073901074938476086, + 0.020084289833903313, + 0.006319317501038313, + 0.009811664000153542, + 0.010941561311483383, + 0.007771333679556847, + 0.014113145880401134, + 0.008488241583108902, + 0.0071241166442632675, + 0.007224924396723509, + 0.004609875846654177, + 0.00804585125297308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -1.4063581228256226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041108159348368645 + ], + "opposite_logprob_tokens": [ + -5.504110813140869 + ], + "logprob_reward": -0.0041108159348368645, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041108159348368645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17033597826957703, + 0.08702734112739563, + 0.07761356234550476, + 0.08910463005304337, + 0.15272314846515656, + 0.054947663098573685, + 0.1195158064365387, + 0.1155591756105423, + 0.08671654760837555, + 0.046456146985292435 + ], + "credit_map": [ + 0.021438494324684143, + 0.01775750145316124, + 0.032433029264211655, + 0.019582118839025497, + 0.020022660493850708, + 0.015104326419532299, + 0.022535908967256546, + 0.021461931988596916, + 0.01328159961849451, + 0.01697446219623089, + 0.012926421128213406, + 0.010404430329799652, + 0.010286666452884674, + 0.010578960180282593, + 0.004683999810367823, + 0.007890801876783371, + 0.009705349802970886, + 0.01787916012108326, + 0.00998266227543354, + 0.010431284084916115, + 0.009524271823465824, + 0.006424851715564728, + 0.006593848578631878, + 0.007072128355503082, + 0.01728411763906479, + 0.009560243226587772, + 0.010580727830529213, + 0.010876928456127644, + 0.006890968419611454, + 0.014782094396650791, + 0.006961722392588854, + 0.012167821638286114, + 0.019507108256220818, + 0.029279960319399834, + 0.021635843440890312, + 0.019105201587080956, + 0.01950882188975811, + 0.014238161966204643, + 0.014056345447897911, + 0.015391700901091099, + 0.005925263278186321, + 0.0059379772283136845, + 0.008633983321487904, + 0.007571028079837561, + 0.009822998195886612, + 0.006911394186317921, + 0.005509508773684502, + 0.004635510500520468, + 0.013699852861464024, + 0.015701964497566223, + 0.026622774079442024, + 0.008761473931372166, + 0.016831643879413605, + 0.013998880051076412, + 0.009367590770125389, + 0.014531629160046577, + 0.015965064987540245, + 0.013304120860993862, + 0.023353876546025276, + 0.016662901267409325, + 0.009145008400082588, + 0.01283480878919363, + 0.012879695743322372, + 0.011413699015974998, + 0.01116094645112753, + 0.010520560666918755, + 0.016264233738183975, + 0.011559653095901012, + 0.005970704834908247, + 0.014675957150757313, + 0.0060455817729234695, + 0.010518908500671387, + 0.006863657850772142, + 0.005447132978588343, + 0.009739753790199757, + 0.006043376866728067, + 0.0043905554339289665, + 0.004472100175917149, + 0.003906920552253723, + 0.005592651199549437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0041108159348368645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081697419285774 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081697419285774, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081697419285774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18733128905296326, + 0.07961256802082062, + 0.144513338804245, + 0.0419803224503994, + 0.18397851288318634, + 0.09055497497320175, + 0.06596114486455917, + 0.06560906022787094, + 0.08927814662456512, + 0.05118072032928467 + ], + "credit_map": [ + 0.0261705182492733, + 0.019550690427422523, + 0.035272449254989624, + 0.03140388801693916, + 0.020566172897815704, + 0.017552338540554047, + 0.017140647396445274, + 0.019674548879265785, + 0.013205980882048607, + 0.013381373137235641, + 0.01228927168995142, + 0.007834021002054214, + 0.009702538140118122, + 0.00836474820971489, + 0.007412535138428211, + 0.007422087248414755, + 0.023218901827931404, + 0.03047822043299675, + 0.024364693090319633, + 0.018271345645189285, + 0.017084196209907532, + 0.012381218373775482, + 0.007629513740539551, + 0.01108523365110159, + 0.008151337504386902, + 0.005426276940852404, + 0.004632121417671442, + 0.003898606402799487, + 0.004625438246876001, + 0.004979570861905813, + 0.003589837346225977, + 0.0066771311685442924, + 0.02486330270767212, + 0.04704153910279274, + 0.025722520425915718, + 0.021703781560063362, + 0.019004181027412415, + 0.016250602900981903, + 0.012675407342612743, + 0.016717158257961273, + 0.009733639657497406, + 0.01038375310599804, + 0.014528430067002773, + 0.011887066066265106, + 0.015739252790808678, + 0.011119874194264412, + 0.008516380563378334, + 0.008646569214761257, + 0.006937349680811167, + 0.008485679514706135, + 0.013104617595672607, + 0.0054115308448672295, + 0.009238899685442448, + 0.007563710678368807, + 0.004774802830070257, + 0.010444548912346363, + 0.010387865826487541, + 0.010019884444773197, + 0.016526691615581512, + 0.007460812572389841, + 0.0049430024810135365, + 0.006036918610334396, + 0.004649299196898937, + 0.005584580823779106, + 0.010051541030406952, + 0.013478703796863556, + 0.017701733857393265, + 0.01075937133282423, + 0.007426867261528969, + 0.013145241886377335, + 0.006049442570656538, + 0.010665236972272396, + 0.008513120003044605, + 0.0050545125268399715, + 0.01047622412443161, + 0.005788789596408606, + 0.0049429829232394695, + 0.007406770251691341, + 0.003242115257307887, + 0.005756198428571224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.011081697419285774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635692600160837 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635692600160837, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635692600160837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17581512033939362, + 0.07890544086694717, + 0.09562361240386963, + 0.042058560997247696, + 0.05659538879990578, + 0.06879803538322449, + 0.08807173371315002, + 0.15286536514759064, + 0.1810920089483261, + 0.06017470732331276 + ], + "credit_map": [ + 0.025460127741098404, + 0.01814461499452591, + 0.03175694867968559, + 0.03700338304042816, + 0.015894578769803047, + 0.01885758712887764, + 0.014291802421212196, + 0.014406081289052963, + 0.011520462110638618, + 0.010200832970440388, + 0.013434354215860367, + 0.010101400315761566, + 0.009669889695942402, + 0.010240605100989342, + 0.005707310512661934, + 0.008030582219362259, + 0.015791786834597588, + 0.019096046686172485, + 0.014524240046739578, + 0.010679180733859539, + 0.011989358812570572, + 0.009504367597401142, + 0.007056764326989651, + 0.006981869228184223, + 0.006398136727511883, + 0.006897779647260904, + 0.005359284579753876, + 0.0072645144537091255, + 0.0032158344984054565, + 0.004845726303756237, + 0.003455323865637183, + 0.004621960688382387, + 0.007669222541153431, + 0.009700167924165726, + 0.0082700839266181, + 0.008462357334792614, + 0.005943487863987684, + 0.0075080301612615585, + 0.003907559905201197, + 0.005134481005370617, + 0.007881129160523415, + 0.007795676589012146, + 0.010419663041830063, + 0.007748440373688936, + 0.010074364021420479, + 0.010332980193197727, + 0.007124721072614193, + 0.007421067915856838, + 0.009420732967555523, + 0.015944473445415497, + 0.019071197137236595, + 0.007824542000889778, + 0.010751353576779366, + 0.010031958110630512, + 0.006799385417252779, + 0.008228088729083538, + 0.037776436656713486, + 0.012842009775340557, + 0.0514976941049099, + 0.010842249728739262, + 0.00950921792536974, + 0.011008917354047298, + 0.010350208729505539, + 0.009038618765771389, + 0.02403283677995205, + 0.02373841404914856, + 0.03246419504284859, + 0.024112867191433907, + 0.013518000021576881, + 0.032004352658987045, + 0.00985358189791441, + 0.021367773413658142, + 0.011060030199587345, + 0.007420566864311695, + 0.007718385197222233, + 0.008118655532598495, + 0.006013525649905205, + 0.008319473825395107, + 0.006240649614483118, + 0.005283422768115997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.007635692600160837 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up shot, the camera held still, focuses deep within a masonry hearth. In an instant, a flame erupts from the bottom of the frame, its core a dazzling, incandescent white, rapidly transitioning to vibrant orange and deep red. The flames twist and twirl violently in the air, forming a fluid, translucent body of energy. As this surge of power erupts, countless fiery sparks are swept upward. These tiny, bright golden particles scatter against the near-black background, leaving fleeting, brilliant trails like a miniature meteor shower erupting upward. The flames, the sole source of light, cast a dancing, flickering glow, illuminating the rough, smoke-blackened masonry within and casting rapidly shifting, deep shadows all around. The image is imbued with a sense of raw power and fiery dynamic beauty, creating an atmosphere both warm and awe-inspiring.", + "seed": 16516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_4/index_7539_seed_16516_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2590155672281981, + "vlm_logprob": -0.2590155672281981, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.2590155672281981, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05161014944314957 + ], + "opposite_logprob_tokens": [ + -3.176610231399536 + ], + "logprob_reward": -0.05161014944314957, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05161014944314957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07414300739765167, + 0.2619059681892395, + 0.036516185849905014, + 0.029968764632940292, + 0.010210635140538216, + 0.03682839125394821, + 0.06540574878454208, + 0.070790596306324, + 0.3933311700820923, + 0.020899469032883644 + ], + "credit_map": [ + 0.010986496694386005, + 0.01240867841988802, + 0.008042207919061184, + 0.007716672960668802, + 0.007732482627034187, + 0.01095939613878727, + 0.00731259398162365, + 0.00898448284715414, + 0.021510275080800056, + 0.0277984868735075, + 0.029305730015039444, + 0.04911034181714058, + 0.030868858098983765, + 0.05094752833247185, + 0.033596646040678024, + 0.018768101930618286, + 0.008102119900286198, + 0.010573320090770721, + 0.0032480836380273104, + 0.0024594140704721212, + 0.0025998586788773537, + 0.0033910798374563456, + 0.002390761859714985, + 0.003751545213162899, + 0.003922094590961933, + 0.0032918171491473913, + 0.007117206696420908, + 0.0022332523949444294, + 0.002995014889165759, + 0.005503255873918533, + 0.00165647198446095, + 0.0032496529165655375, + 0.0018226733664050698, + 0.0011261292966082692, + 0.0026756571605801582, + 0.0007724313181824982, + 0.00095540372421965, + 0.001190170762129128, + 0.000756638590246439, + 0.0009115300490520895, + 0.005079374648630619, + 0.004061756189912558, + 0.004959803074598312, + 0.004055985249578953, + 0.004074391908943653, + 0.005747273564338684, + 0.004761738236993551, + 0.0040880669839680195, + 0.007217087782919407, + 0.007765075657516718, + 0.00971071608364582, + 0.009575242176651955, + 0.006808186415582895, + 0.007260965183377266, + 0.006927103269845247, + 0.010141370818018913, + 0.007619362790137529, + 0.010113944299519062, + 0.010094883851706982, + 0.006596054881811142, + 0.013135829009115696, + 0.008338111452758312, + 0.006953169126063585, + 0.00793924368917942, + 0.041400667279958725, + 0.04727964848279953, + 0.10106641054153442, + 0.03317315876483917, + 0.050702523440122604, + 0.06274985522031784, + 0.021348927170038223, + 0.03560996800661087, + 0.0042732288129627705, + 0.0026194739621132612, + 0.006956234108656645, + 0.0014616327825933695, + 0.001627139514312148, + 0.0021116668358445168, + 0.0011033505434170365, + 0.0007467420073226094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.05161014944314957 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236781120300293 + ], + "opposite_logprob_tokens": [ + -0.3486781120300293 + ], + "logprob_reward": -1.2236781120300293, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236781120300293, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07225021719932556, + 0.2508223354816437, + 0.18717677891254425, + 0.04887942597270012, + 0.11056380718946457, + 0.048762548714876175, + 0.05089191347360611, + 0.03387697786092758, + 0.11274074018001556, + 0.08403526991605759 + ], + "credit_map": [ + 0.010825477540493011, + 0.007484371308237314, + 0.010370166040956974, + 0.008129635825753212, + 0.006008358206599951, + 0.011130879633128643, + 0.010205622762441635, + 0.008095701225101948, + 0.02112274244427681, + 0.02732601761817932, + 0.024601934477686882, + 0.044364478439092636, + 0.032508138567209244, + 0.05415791645646095, + 0.0309221800416708, + 0.015818925574421883, + 0.048848167061805725, + 0.06950340420007706, + 0.01042371615767479, + 0.00755032803863287, + 0.01308062020689249, + 0.01948063261806965, + 0.0063873776234686375, + 0.011902537196874619, + 0.006977184675633907, + 0.0050340695306658745, + 0.010641907341778278, + 0.0036583489272743464, + 0.004575905855745077, + 0.007429890334606171, + 0.003345901845023036, + 0.0072162155993282795, + 0.02120998315513134, + 0.011649674735963345, + 0.03336353600025177, + 0.009190584532916546, + 0.007513825781643391, + 0.01238919235765934, + 0.006717392709106207, + 0.008529601618647575, + 0.00571354990825057, + 0.0062242052517831326, + 0.0073902239091694355, + 0.006899903528392315, + 0.005170789081603289, + 0.006788074970245361, + 0.005258576013147831, + 0.005317220464348793, + 0.007646840065717697, + 0.0060956100933253765, + 0.008671723306179047, + 0.005634665954858065, + 0.004368884954601526, + 0.007470590062439442, + 0.005346111953258514, + 0.005657482426613569, + 0.005403925199061632, + 0.004256684333086014, + 0.003850833745673299, + 0.0031527576502412558, + 0.006557524669915438, + 0.004041819833219051, + 0.00335931614972651, + 0.0032541141845285892, + 0.011538363993167877, + 0.015416332520544529, + 0.027304280549287796, + 0.007569428998976946, + 0.01935632899403572, + 0.018357200548052788, + 0.005349855870008469, + 0.007848948240280151, + 0.016882507130503654, + 0.010016212239861488, + 0.03386141359806061, + 0.004259722772985697, + 0.004955887328833342, + 0.008139593526721, + 0.0029914998449385166, + 0.0029284260235726833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -1.2236781120300293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362489465624094 + ], + "opposite_logprob_tokens": [ + -5.62862491607666 + ], + "logprob_reward": -0.00362489465624094, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362489465624094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04936900734901428, + 0.5695248246192932, + 0.03905981779098511, + 0.03072281740605831, + 0.026919422671198845, + 0.02769671007990837, + 0.014607250690460205, + 0.03369498997926712, + 0.1371307522058487, + 0.07127443701028824 + ], + "credit_map": [ + 0.005257292650640011, + 0.006199271883815527, + 0.007687577977776527, + 0.007015672978013754, + 0.004595569800585508, + 0.005465896334499121, + 0.00646272674202919, + 0.006685000378638506, + 0.04860546067357063, + 0.06404668837785721, + 0.06141674518585205, + 0.10728161782026291, + 0.06769552081823349, + 0.10420934855937958, + 0.07475325465202332, + 0.041516173630952835, + 0.010088924318552017, + 0.014058264903724194, + 0.002631759736686945, + 0.0028454791754484177, + 0.00279794167727232, + 0.0031519222538918257, + 0.001921387272886932, + 0.001564142294228077, + 0.0036735492758452892, + 0.0038718446157872677, + 0.007658400572836399, + 0.0035180626437067986, + 0.002765500918030739, + 0.004157339222729206, + 0.0024449308402836323, + 0.002633189084008336, + 0.0046482388861477375, + 0.0028552256990224123, + 0.006429609376937151, + 0.002431280445307493, + 0.0021104447077959776, + 0.003432059194892645, + 0.002490544691681862, + 0.0025220182724297047, + 0.003771888790652156, + 0.00429147994145751, + 0.003474370576441288, + 0.0035903321113437414, + 0.0033062677830457687, + 0.004479268565773964, + 0.0025153313763439655, + 0.00226777046918869, + 0.0017980458214879036, + 0.0017796828178688884, + 0.002902308711782098, + 0.0017906975699588656, + 0.0013071626890450716, + 0.0017501979600638151, + 0.0014607704943045974, + 0.0018183845095336437, + 0.007183338049799204, + 0.005410185549408197, + 0.00530647998675704, + 0.002489491133019328, + 0.00439466955140233, + 0.004664049483835697, + 0.0021058297716081142, + 0.002140945987775922, + 0.01347353309392929, + 0.017969030886888504, + 0.03465276211500168, + 0.009924967773258686, + 0.021002421155571938, + 0.024106577038764954, + 0.00655055046081543, + 0.009450907818973064, + 0.014293966814875603, + 0.008298938162624836, + 0.02634061500430107, + 0.003687242977321148, + 0.004575653932988644, + 0.0074925203807652, + 0.002700005890801549, + 0.0038854910526424646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.00362489465624094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544417753815651 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544417753815651, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544417753815651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10823950916528702, + 0.17162537574768066, + 0.08901745826005936, + 0.04466952756047249, + 0.0439315102994442, + 0.059894662350416183, + 0.02744322642683983, + 0.08233003318309784, + 0.2096565067768097, + 0.1631922423839569 + ], + "credit_map": [ + 0.014350197277963161, + 0.01587420329451561, + 0.018033258616924286, + 0.011144939810037613, + 0.009908195585012436, + 0.015454920940101147, + 0.011755125597119331, + 0.011718660593032837, + 0.014051693491637707, + 0.020481424406170845, + 0.019551707431674004, + 0.03529024496674538, + 0.01763131655752659, + 0.027779879048466682, + 0.02433527633547783, + 0.012503819540143013, + 0.021975746378302574, + 0.027654001489281654, + 0.00700904056429863, + 0.007266653701663017, + 0.006355038844048977, + 0.00841753464192152, + 0.004823809489607811, + 0.005515626166015863, + 0.005402923095971346, + 0.0045668757520616055, + 0.010255707427859306, + 0.0041664657182991505, + 0.0053438665345311165, + 0.007617524825036526, + 0.0036289566196501255, + 0.003687201766297221, + 0.007672351319342852, + 0.004979001823812723, + 0.01191681157797575, + 0.00403317715972662, + 0.0032801346387714148, + 0.005323342978954315, + 0.0032729373779147863, + 0.0034537536557763815, + 0.007226596120744944, + 0.005254687741398811, + 0.007057172246277332, + 0.012015909887850285, + 0.004826377611607313, + 0.007644273340702057, + 0.009213754907250404, + 0.006655889097601175, + 0.0031093403231352568, + 0.004228036850690842, + 0.004169906955212355, + 0.0034027700312435627, + 0.002652902388945222, + 0.0036150324158370495, + 0.0027579213492572308, + 0.003507315181195736, + 0.019444560632109642, + 0.01124589703977108, + 0.01441215630620718, + 0.004725698847323656, + 0.009233781136572361, + 0.011835919693112373, + 0.006300309672951698, + 0.00513170612975955, + 0.020887115970253944, + 0.027316942811012268, + 0.056603796780109406, + 0.016552936285734177, + 0.028453629463911057, + 0.03371307998895645, + 0.010966253466904163, + 0.015162733383476734, + 0.030707836151123047, + 0.018383702263236046, + 0.06031500920653343, + 0.009646872989833355, + 0.010126059874892235, + 0.017120616510510445, + 0.006574192084372044, + 0.010317949578166008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.012544417753815651 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036202622577548027 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036202622577548027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036202622577548027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04584407061338425, + 0.6174248456954956, + 0.02986150048673153, + 0.02637823298573494, + 0.015826063230633736, + 0.027279717847704887, + 0.04547695815563202, + 0.062069013714790344, + 0.08423298597335815, + 0.04560662433505058 + ], + "credit_map": [ + 0.005818243604153395, + 0.006979664321988821, + 0.008030720986425877, + 0.004269695840775967, + 0.003896957030519843, + 0.006871219724416733, + 0.004537099972367287, + 0.005440474487841129, + 0.05071239918470383, + 0.06582782417535782, + 0.06705069541931152, + 0.11830991506576538, + 0.0728975310921669, + 0.11387494951486588, + 0.08228322118520737, + 0.04646832495927811, + 0.006700955796986818, + 0.008161023259162903, + 0.00232402584515512, + 0.002674258314073086, + 0.0022683467250317335, + 0.002691072877496481, + 0.001708020456135273, + 0.0033338007051497698, + 0.0036824934650212526, + 0.0029402400832623243, + 0.006280133035033941, + 0.0019031974952667952, + 0.0022133986931294203, + 0.005010530352592468, + 0.001984373666346073, + 0.002363868523389101, + 0.0030015758238732815, + 0.0014923922717571259, + 0.0028070686385035515, + 0.001439927495084703, + 0.001713937264867127, + 0.00216507725417614, + 0.0014948861207813025, + 0.0017111990600824356, + 0.004005481954663992, + 0.003160193096846342, + 0.0036557400599122047, + 0.0025197137147188187, + 0.003382078604772687, + 0.004155782517045736, + 0.0031885034404695034, + 0.0032122258562594652, + 0.005027064587920904, + 0.007102807052433491, + 0.0056725009344518185, + 0.006360783241689205, + 0.005391244776546955, + 0.00539796520024538, + 0.005054270848631859, + 0.00547032430768013, + 0.00834429357200861, + 0.008869209326803684, + 0.007869021967053413, + 0.0057340762577950954, + 0.011882675811648369, + 0.007519520819187164, + 0.005344147328287363, + 0.006506071891635656, + 0.00817187037318945, + 0.012000877410173416, + 0.013657890260219574, + 0.005922562908381224, + 0.015379690565168858, + 0.015503952279686928, + 0.0050233532674610615, + 0.008572788909077644, + 0.0077331592328846455, + 0.005159316118806601, + 0.014898158609867096, + 0.003123920178040862, + 0.002956738928332925, + 0.005076506175100803, + 0.0020194414537400007, + 0.004639387596398592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.0036202622577548027 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up shot, the camera held still, focuses deep within a masonry hearth. In an instant, a flame erupts from the bottom of the frame, its core a dazzling, incandescent white, rapidly transitioning to vibrant orange and deep red. The flames twist and twirl violently in the air, forming a fluid, translucent body of energy. As this surge of power erupts, countless fiery sparks are swept upward. These tiny, bright golden particles scatter against the near-black background, leaving fleeting, brilliant trails like a miniature meteor shower erupting upward. The flames, the sole source of light, cast a dancing, flickering glow, illuminating the rough, smoke-blackened masonry within and casting rapidly shifting, deep shadows all around. The image is imbued with a sense of raw power and fiery dynamic beauty, creating an atmosphere both warm and awe-inspiring.", + "seed": 16517, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_4/index_7539_seed_16517_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2583304623141885, + "vlm_logprob": -0.2583304623141885, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.2583304623141885, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04667094349861145 + ], + "opposite_logprob_tokens": [ + -3.171670913696289 + ], + "logprob_reward": -0.04667094349861145, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04667094349861145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28875988721847534, + 0.10569748282432556, + 0.0746021494269371, + 0.03286832571029663, + 0.0873129591345787, + 0.07969736307859421, + 0.07551626861095428, + 0.04623990133404732, + 0.1177118718624115, + 0.09159383922815323 + ], + "credit_map": [ + 0.03794672340154648, + 0.030481556430459023, + 0.041221197694540024, + 0.037154119461774826, + 0.028774842619895935, + 0.022810159251093864, + 0.05313791334629059, + 0.0372333750128746, + 0.010337386280298233, + 0.011800725013017654, + 0.026779964566230774, + 0.01634972169995308, + 0.008107180707156658, + 0.006805539596825838, + 0.010549806989729404, + 0.014967164024710655, + 0.004616600461304188, + 0.010956289246678352, + 0.008478756994009018, + 0.00769826490432024, + 0.01715431548655033, + 0.005585900507867336, + 0.00913033913820982, + 0.01098167896270752, + 0.005033555440604687, + 0.003646988421678543, + 0.005811494309455156, + 0.0032172042410820723, + 0.002250476973131299, + 0.0030502129811793566, + 0.003446271177381277, + 0.006412121932953596, + 0.020177874714136124, + 0.010733816772699356, + 0.01619652658700943, + 0.00785492267459631, + 0.006076305638998747, + 0.0077902995981276035, + 0.005724617280066013, + 0.012758598662912846, + 0.012869450263679028, + 0.009103292599320412, + 0.01120713260024786, + 0.012621880508959293, + 0.010901672765612602, + 0.009371861815452576, + 0.005212412215769291, + 0.008409659378230572, + 0.007831989787518978, + 0.009752385318279266, + 0.013421615585684776, + 0.0067571853287518024, + 0.008129779249429703, + 0.01273398194462061, + 0.006966808810830116, + 0.009922523982822895, + 0.006739899516105652, + 0.0047305370680987835, + 0.008754635229706764, + 0.0053626117296516895, + 0.004105641506612301, + 0.006332943681627512, + 0.005305723287165165, + 0.0049079060554504395, + 0.010251743718981743, + 0.015582007355988026, + 0.03193507343530655, + 0.011397571302950382, + 0.012001763097941875, + 0.013678298331797123, + 0.010614149272441864, + 0.012251258827745914, + 0.009155891835689545, + 0.013103649951517582, + 0.018430938944220543, + 0.010774830356240273, + 0.010600384324789047, + 0.012295974418520927, + 0.005936943925917149, + 0.011295221745967865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.04667094349861145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236560583114624 + ], + "opposite_logprob_tokens": [ + -0.3486560583114624 + ], + "logprob_reward": -1.2236560583114624, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236560583114624, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1344757378101349, + 0.05584867671132088, + 0.1436740756034851, + 0.12121646851301193, + 0.1274261474609375, + 0.09948409348726273, + 0.07312507182359695, + 0.05485530570149422, + 0.057607728987932205, + 0.1322866678237915 + ], + "credit_map": [ + 0.021508891135454178, + 0.012057837098836899, + 0.025969751179218292, + 0.012541762553155422, + 0.011809184215962887, + 0.0078014591708779335, + 0.02421307936310768, + 0.01857377216219902, + 0.0041785165667533875, + 0.008829070255160332, + 0.007880798541009426, + 0.00992603786289692, + 0.00603792630136013, + 0.0049029686488211155, + 0.006244472693651915, + 0.007848884910345078, + 0.013653947971761227, + 0.012423056177794933, + 0.034760117530822754, + 0.011377324350178242, + 0.009375550784170628, + 0.012042933143675327, + 0.009865484200417995, + 0.040175650268793106, + 0.01517720427364111, + 0.01588948629796505, + 0.01798919588327408, + 0.013198202475905418, + 0.00960589200258255, + 0.012277623638510704, + 0.011411876417696476, + 0.025666994974017143, + 0.027239037677645683, + 0.015898240730166435, + 0.025716977193951607, + 0.010117306374013424, + 0.011053705587983131, + 0.01073792390525341, + 0.009431643411517143, + 0.01723131537437439, + 0.012386202812194824, + 0.012702357023954391, + 0.014972136355936527, + 0.014773249626159668, + 0.011556549929082394, + 0.013304561376571655, + 0.0070529212243855, + 0.012736105360090733, + 0.011852338910102844, + 0.007382824085652828, + 0.01975821889936924, + 0.006748615764081478, + 0.005411249585449696, + 0.010233203880488873, + 0.0049372510984539986, + 0.006801370531320572, + 0.00519878463819623, + 0.005143896210938692, + 0.008581542409956455, + 0.005640096962451935, + 0.005915024783462286, + 0.006382640916854143, + 0.012681527994573116, + 0.005311790853738785, + 0.0046613095328211784, + 0.0075671677477657795, + 0.0133106904104352, + 0.004939335398375988, + 0.006449171807616949, + 0.006537650711834431, + 0.005728439893573523, + 0.008413960225880146, + 0.013347414322197437, + 0.025476111099123955, + 0.019878223538398743, + 0.014616318047046661, + 0.019285723567008972, + 0.015106912702322006, + 0.008286976255476475, + 0.016288993880152702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -1.2236560583114624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028267446905374527 + ], + "opposite_logprob_tokens": [ + -5.877826690673828 + ], + "logprob_reward": -0.0028267446905374527, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028267446905374527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4063403010368347, + 0.05933214724063873, + 0.11148663610219955, + 0.08370565623044968, + 0.05743735283613205, + 0.04066000506281853, + 0.06066468730568886, + 0.025804588571190834, + 0.0678231492638588, + 0.0867454931139946 + ], + "credit_map": [ + 0.07218658179044724, + 0.017546629533171654, + 0.029630977660417557, + 0.05491599068045616, + 0.04135642200708389, + 0.0345652811229229, + 0.12631011009216309, + 0.029828334227204323, + 0.006051138509064913, + 0.0054329875856637955, + 0.01355075091123581, + 0.012034212239086628, + 0.004594235215336084, + 0.003647770034149289, + 0.005839758552610874, + 0.008181293494999409, + 0.01244428288191557, + 0.011931988410651684, + 0.019650330767035484, + 0.012594999745488167, + 0.010258153080940247, + 0.011871539056301117, + 0.013213560916483402, + 0.019521787762641907, + 0.008495212532579899, + 0.009625609032809734, + 0.01430155523121357, + 0.009677344001829624, + 0.007067396305501461, + 0.0073900772258639336, + 0.008584667928516865, + 0.018563799560070038, + 0.012606915086507797, + 0.004793714266270399, + 0.012105045840144157, + 0.004889059811830521, + 0.004172318149358034, + 0.004540476948022842, + 0.004424518905580044, + 0.009905301034450531, + 0.005687902215868235, + 0.004926295019686222, + 0.005609328392893076, + 0.005729733966290951, + 0.005882371682673693, + 0.004484021570533514, + 0.0034701661206781864, + 0.004870186559855938, + 0.009293629787862301, + 0.007672567386180162, + 0.01181242149323225, + 0.005061488598585129, + 0.00474271597340703, + 0.009184680879116058, + 0.005866101942956448, + 0.007031083106994629, + 0.0033814460039138794, + 0.0019914675503969193, + 0.004016878083348274, + 0.0032475481275469065, + 0.002376516815274954, + 0.003616798436269164, + 0.003381921211257577, + 0.003792011644691229, + 0.005522891879081726, + 0.007860925979912281, + 0.01694260910153389, + 0.005480033811181784, + 0.007412214297801256, + 0.008255249820649624, + 0.00849071703851223, + 0.007858502678573132, + 0.007507796864956617, + 0.013406157493591309, + 0.014500602148473263, + 0.011418471112847328, + 0.012188060209155083, + 0.009838921017944813, + 0.006980534642934799, + 0.010904951952397823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0028267446905374527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544536031782627 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544536031782627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544536031782627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13936372101306915, + 0.09306871891021729, + 0.1256924420595169, + 0.09152577817440033, + 0.22016017138957977, + 0.1014237105846405, + 0.0472228117287159, + 0.06794346123933792, + 0.037953268736600876, + 0.07564588636159897 + ], + "credit_map": [ + 0.04034681245684624, + 0.012372222729027271, + 0.027033338323235512, + 0.009654556401073933, + 0.007915528491139412, + 0.01149112731218338, + 0.00960052851587534, + 0.02094961702823639, + 0.007038034964352846, + 0.009312805719673634, + 0.015488113276660442, + 0.013466499745845795, + 0.014376822859048843, + 0.006440714001655579, + 0.01331951655447483, + 0.01362620573490858, + 0.01279660128057003, + 0.018753768876194954, + 0.015240076929330826, + 0.010465112514793873, + 0.035475075244903564, + 0.00998474657535553, + 0.014648756943643093, + 0.008328305557370186, + 0.011991262435913086, + 0.01139787957072258, + 0.014630059711635113, + 0.013600685633718967, + 0.006673211231827736, + 0.007324572652578354, + 0.010431855916976929, + 0.015476254746317863, + 0.046545956283807755, + 0.023004276677966118, + 0.040666982531547546, + 0.02148631028831005, + 0.017634134739637375, + 0.018484536558389664, + 0.016438167542219162, + 0.03589979559183121, + 0.01010807603597641, + 0.010339378379285336, + 0.019735831767320633, + 0.011943108402192593, + 0.011186967603862286, + 0.015131313353776932, + 0.0058243293315172195, + 0.017154701054096222, + 0.005292986985296011, + 0.0034466497600078583, + 0.009309726767241955, + 0.004159025847911835, + 0.004436499439179897, + 0.009012729860842228, + 0.004694321658462286, + 0.006870872341096401, + 0.010602299124002457, + 0.012250087223947048, + 0.01298186369240284, + 0.00796959176659584, + 0.006589964032173157, + 0.006610305048525333, + 0.0058548543602228165, + 0.005084499716758728, + 0.0029812685679644346, + 0.005510301794856787, + 0.004667952656745911, + 0.004695226904004812, + 0.004349282011389732, + 0.003857709001749754, + 0.008185772225260735, + 0.0037057565059512854, + 0.007792761083692312, + 0.011385512538254261, + 0.014036077074706554, + 0.008511925116181374, + 0.00989697128534317, + 0.008830083534121513, + 0.004932771902531385, + 0.010259781964123249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.012544536031782627 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focused on a masonry hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the flames erupt from the bottom of the frame with a core that transitions from white to orange and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless tiny golden sparks scattered upward, leaving bright trails against a near-black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rough masonry illuminated by the flames with rapidly shifting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059540290385484695 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.0059540290385484695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059540290385484695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34247905015945435, + 0.07788687944412231, + 0.07432523369789124, + 0.03514091297984123, + 0.07901983708143234, + 0.10830190777778625, + 0.025084514170885086, + 0.039667025208473206, + 0.1273529976606369, + 0.09074166417121887 + ], + "credit_map": [ + 0.05601883679628372, + 0.02618984319269657, + 0.03783784434199333, + 0.044939618557691574, + 0.032630305737257004, + 0.028118405491113663, + 0.08373655378818512, + 0.03300764411687851, + 0.007437060121446848, + 0.00994728971272707, + 0.017626479268074036, + 0.013554103672504425, + 0.006711906287819147, + 0.0047263153828680515, + 0.008769777603447437, + 0.009113945998251438, + 0.006662753410637379, + 0.008253337815403938, + 0.01152661070227623, + 0.007150011602789164, + 0.012902441434562206, + 0.005688839592039585, + 0.00783564057201147, + 0.014305593445897102, + 0.002378959907218814, + 0.003582242876291275, + 0.006167042534798384, + 0.0028922788333147764, + 0.0023976422380656004, + 0.0026102925185114145, + 0.0033746466506272554, + 0.011737807653844357, + 0.01407964900135994, + 0.010485655628144741, + 0.011626902036368847, + 0.008555128239095211, + 0.007601532153785229, + 0.008660205639898777, + 0.005068460013717413, + 0.012942296452820301, + 0.014900412410497665, + 0.014968454837799072, + 0.015648921951651573, + 0.015218197368085384, + 0.011756078340113163, + 0.0145740220323205, + 0.007482364773750305, + 0.013753451406955719, + 0.0040327683091163635, + 0.0028259367682039738, + 0.00603287061676383, + 0.0024214896839112043, + 0.0024194447323679924, + 0.0028486496303230524, + 0.0020247085485607386, + 0.002478645648807287, + 0.006252320017665625, + 0.005424614064395428, + 0.00758573692291975, + 0.004657439887523651, + 0.0031132737640291452, + 0.005129730328917503, + 0.003887886181473732, + 0.003616021014750004, + 0.011792557314038277, + 0.01562059111893177, + 0.03683292493224144, + 0.008630511350929737, + 0.013107601553201675, + 0.015477309003472328, + 0.01114578265696764, + 0.01474571879953146, + 0.009277701377868652, + 0.015661736950278282, + 0.017624935135245323, + 0.00953561719506979, + 0.01064080186188221, + 0.011829005554318428, + 0.0054637049324810505, + 0.010708156041800976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0059540290385484695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_5/index_tensor([887, 887])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_5/index_tensor([887, 887])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..13be1e84850ad513fa9ae2c3553a27ed8d9aacb4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_5/index_tensor([887, 887])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, medium-long shot begins on a mountain terrace shrouded in morning mist. An elderly monk, his smooth head gleaming in the damp air, deep wrinkles etched across his serene face like rings of wisdom. He wears a gray cotton and linen robe, slightly whitish from age, its wide cuffs swaying gently with his movements. His feet are firmly rooted in the moist earth, his body erect and relaxed. His arms rise calmly from his sides, tracing a complete, smooth arc in front of him with a steady, inner-strength rhythm, his fingertips seemingly touching invisible currents of energy. The first soft rays of morning light penetrate the mist, casting a faint golden rim around his figure. The light creates a Tyndall effect within the ethereal haze, creating visible beams of light. His eyes are half-closed, a faint smile lingers on his lips, and his face carries a transcendental peace, as if blending seamlessly with the flowing air and the distant mountains. The painting's cool and subdued tones, combined with slow, continuous movements, create an atmosphere of profound tranquility, solemnity, and a sense of Zen.", + "seed": 19168, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_5/index_887_seed_19168_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02024546395987272, + "vlm_logprob": -0.02024546395987272, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.02024546395987272, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020185641944408417 + ], + "opposite_logprob_tokens": [ + -4.145185470581055 + ], + "logprob_reward": -0.020185641944408417, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020185641944408417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1899765133857727, + 0.07184169441461563, + 0.11426225304603577, + 0.048537369817495346, + 0.13502350449562073, + 0.09414119273424149, + 0.07140142470598221, + 0.08368082344532013, + 0.09029172360897064, + 0.10084344446659088 + ], + "credit_map": [ + 0.022006889805197716, + 0.03437655419111252, + 0.037420570850372314, + 0.017782988026738167, + 0.02223539724946022, + 0.017764490097761154, + 0.016939347609877586, + 0.02145027555525303, + 0.0049043819308280945, + 0.00714534567669034, + 0.011487117037177086, + 0.010422135703265667, + 0.010561339557170868, + 0.007857327349483967, + 0.00907921139150858, + 0.010384838096797466, + 0.008917998522520065, + 0.015330000780522823, + 0.010858950205147266, + 0.015782268717885017, + 0.014756166376173496, + 0.011847196146845818, + 0.022122573107481003, + 0.01464709360152483, + 0.005425138399004936, + 0.004767096601426601, + 0.007440146990120411, + 0.006710992194712162, + 0.007855679839849472, + 0.004489128012210131, + 0.0059411050751805305, + 0.005908086430281401, + 0.012231954373419285, + 0.014558454044163227, + 0.01965496502816677, + 0.01711316406726837, + 0.017788344994187355, + 0.015091245993971825, + 0.016557374969124794, + 0.022028012201189995, + 0.010375435464084148, + 0.01150151714682579, + 0.01641898788511753, + 0.006757800001651049, + 0.015106684528291225, + 0.011259693652391434, + 0.008826862089335918, + 0.013894213363528252, + 0.008333807811141014, + 0.008433066308498383, + 0.0131161417812109, + 0.005263506900519133, + 0.010474815033376217, + 0.00759591581299901, + 0.008025052025914192, + 0.010159115307033062, + 0.015212731435894966, + 0.01102987490594387, + 0.010001554153859615, + 0.00644673639908433, + 0.008879597298800945, + 0.007631656248122454, + 0.01050800271332264, + 0.013970672152936459, + 0.014854409731924534, + 0.014395038597285748, + 0.007335393689572811, + 0.010912512429058552, + 0.009586566127836704, + 0.015891263261437416, + 0.009296072646975517, + 0.008020464330911636, + 0.02038051187992096, + 0.012981537729501724, + 0.01167532429099083, + 0.009817029349505901, + 0.010000252164900303, + 0.011832321994006634, + 0.009706461802124977, + 0.014450005255639553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.020185641944408417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014213516376912594 + ], + "opposite_logprob_tokens": [ + -4.264213562011719 + ], + "logprob_reward": -0.014213516376912594, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014213516376912594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3295154571533203, + 0.12397997826337814, + 0.07184497267007828, + 0.06863744556903839, + 0.0809108093380928, + 0.09361247718334198, + 0.07686502486467361, + 0.05873566493391991, + 0.03838202357292175, + 0.0575161874294281 + ], + "credit_map": [ + 0.02276715449988842, + 0.03008558414876461, + 0.030153220519423485, + 0.03283251076936722, + 0.028684303164482117, + 0.07239028066396713, + 0.03044932149350643, + 0.08215305954217911, + 0.009569171816110611, + 0.01396002247929573, + 0.015503858216106892, + 0.019702667370438576, + 0.018065251410007477, + 0.011866317130625248, + 0.015138676390051842, + 0.020174015313386917, + 0.006287429016083479, + 0.01111272256821394, + 0.006804398261010647, + 0.00777861475944519, + 0.010526799596846104, + 0.006926137022674084, + 0.013171341270208359, + 0.009237526915967464, + 0.006348880007863045, + 0.007192541379481554, + 0.006562358699738979, + 0.00801217369735241, + 0.015680722892284393, + 0.006579977925866842, + 0.010447331704199314, + 0.00781345460563898, + 0.007693673949688673, + 0.007607192266732454, + 0.011299563571810722, + 0.01094573363661766, + 0.008922864682972431, + 0.01077659148722887, + 0.010150957852602005, + 0.013514233753085136, + 0.011824142187833786, + 0.00847075879573822, + 0.008036003448069096, + 0.007200484164059162, + 0.01248938124626875, + 0.022420961409807205, + 0.013511616736650467, + 0.009659131988883018, + 0.011234227567911148, + 0.008950202725827694, + 0.0112594710662961, + 0.00707763247191906, + 0.010238473303616047, + 0.007903145626187325, + 0.01234055869281292, + 0.007861316204071045, + 0.012284664437174797, + 0.00666383933275938, + 0.00761418417096138, + 0.004113159142434597, + 0.005054436158388853, + 0.005860551726073027, + 0.007928384467959404, + 0.009216446429491043, + 0.005447853356599808, + 0.007458005100488663, + 0.0027565769851207733, + 0.0037824553437530994, + 0.005122066475450993, + 0.0049850838258862495, + 0.004645287990570068, + 0.004184693098068237, + 0.0060930694453418255, + 0.008430620655417442, + 0.006307133007794619, + 0.0069688656367361546, + 0.010217382572591305, + 0.00624676700681448, + 0.006528518628329039, + 0.006723830476403236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.014213516376912594 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370685875415802 + ], + "opposite_logprob_tokens": [ + -3.4087069034576416 + ], + "logprob_reward": -0.03370685875415802, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370685875415802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27185550332069397, + 0.10299931466579437, + 0.10915309190750122, + 0.03403248265385628, + 0.07705126702785492, + 0.059189144521951675, + 0.03991629555821419, + 0.07435449212789536, + 0.15236006677150726, + 0.07908832281827927 + ], + "credit_map": [ + 0.02259838953614235, + 0.022062988951802254, + 0.019417915493249893, + 0.018993038684129715, + 0.01874474249780178, + 0.07326009124517441, + 0.027413837611675262, + 0.0693645030260086, + 0.008120831102132797, + 0.012463764287531376, + 0.007758055813610554, + 0.015284016728401184, + 0.015417839400470257, + 0.009716778062283993, + 0.021399477496743202, + 0.012838543392717838, + 0.01030802633613348, + 0.014435188844799995, + 0.01147064846009016, + 0.015554102137684822, + 0.013336950913071632, + 0.009413818828761578, + 0.021583175286650658, + 0.013051172718405724, + 0.0034158164635300636, + 0.002909918548539281, + 0.0023639979772269726, + 0.004011454526335001, + 0.006535911001265049, + 0.003289991756901145, + 0.007099012844264507, + 0.004406379070132971, + 0.006127511616796255, + 0.006203822325915098, + 0.010226077400147915, + 0.010312594473361969, + 0.008736114017665386, + 0.013116410933434963, + 0.012725668027997017, + 0.009603065438568592, + 0.009379883296787739, + 0.005057452246546745, + 0.007379129994660616, + 0.004228734411299229, + 0.005720485467463732, + 0.013414183631539345, + 0.008023797534406185, + 0.005985475145280361, + 0.0056120348162949085, + 0.004593293648213148, + 0.005492907017469406, + 0.005102287046611309, + 0.004883229732513428, + 0.004277117550373077, + 0.005492969416081905, + 0.004462455399334431, + 0.016931677237153053, + 0.007810905575752258, + 0.009333927184343338, + 0.006585759576410055, + 0.00851195678114891, + 0.009745325893163681, + 0.007406370248645544, + 0.008028566837310791, + 0.027988672256469727, + 0.026738198474049568, + 0.01247834786772728, + 0.01815064623951912, + 0.013414928689599037, + 0.023657089099287987, + 0.0160880908370018, + 0.013844099827110767, + 0.010010220110416412, + 0.011568546295166016, + 0.004728338681161404, + 0.010799494571983814, + 0.011539947241544724, + 0.008941154927015305, + 0.01189349964261055, + 0.009607124142348766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.03370685875415802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583242177963257 + ], + "opposite_logprob_tokens": [ + -3.895583152770996 + ], + "logprob_reward": -0.020583242177963257, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583242177963257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11547860503196716, + 0.040860723704099655, + 0.05803907290101051, + 0.050947923213243484, + 0.07058945298194885, + 0.09646598249673843, + 0.04215552285313606, + 0.12180343270301819, + 0.286424458026886, + 0.11723486334085464 + ], + "credit_map": [ + 0.01082570105791092, + 0.008987268432974815, + 0.016197185963392258, + 0.006263527553528547, + 0.008930694311857224, + 0.02678564004600048, + 0.007293484639376402, + 0.030195103958249092, + 0.004698992241173983, + 0.007325780112296343, + 0.006855473853647709, + 0.0036947911139577627, + 0.005042411852627993, + 0.003913697320967913, + 0.004077115561813116, + 0.00525246374309063, + 0.004588668700307608, + 0.006843574345111847, + 0.006032193545252085, + 0.006766123231500387, + 0.00856601633131504, + 0.005915237125009298, + 0.011476239189505577, + 0.007851019501686096, + 0.005590205546468496, + 0.006265164352953434, + 0.007323538884520531, + 0.005079422611743212, + 0.01010471023619175, + 0.005059764254838228, + 0.004450664389878511, + 0.007074452005326748, + 0.006559700705111027, + 0.00586891220882535, + 0.008952858857810497, + 0.009021335281431675, + 0.007952231913805008, + 0.01113999355584383, + 0.010607259348034859, + 0.010487157851457596, + 0.013653076253831387, + 0.005803899839520454, + 0.011467485688626766, + 0.005403085611760616, + 0.009247682057321072, + 0.032658811658620834, + 0.010298281908035278, + 0.0079336604103446, + 0.005721251014620066, + 0.007322723511606455, + 0.00508389575406909, + 0.005168978590518236, + 0.004923233762383461, + 0.0042447540909051895, + 0.005774916615337133, + 0.003915773238986731, + 0.023849844932556152, + 0.01641692779958248, + 0.01335187628865242, + 0.010635972023010254, + 0.012147399596869946, + 0.015987897291779518, + 0.013696646317839622, + 0.01571686379611492, + 0.056758131831884384, + 0.05690232291817665, + 0.02532995492219925, + 0.029332343488931656, + 0.02685502916574478, + 0.037431392818689346, + 0.027068164199590683, + 0.02674711123108864, + 0.031370293349027634, + 0.013620221987366676, + 0.015199990943074226, + 0.009291434660553932, + 0.00826038047671318, + 0.014961127191781998, + 0.00888687465339899, + 0.015644539147615433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.020583242177963257 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538060545921326 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538060545921326, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538060545921326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3202672600746155, + 0.05712776631116867, + 0.07229549437761307, + 0.055502891540527344, + 0.15422897040843964, + 0.16798317432403564, + 0.0421626940369606, + 0.04297132417559624, + 0.029276439920067787, + 0.058184053748846054 + ], + "credit_map": [ + 0.022342754527926445, + 0.028381310403347015, + 0.026166105642914772, + 0.015281367115676403, + 0.02265954576432705, + 0.08637548983097076, + 0.03287152945995331, + 0.0861891359090805, + 0.004583826754242182, + 0.007350343279540539, + 0.007316372357308865, + 0.008021799847483635, + 0.008472665213048458, + 0.005551064852625132, + 0.008520984090864658, + 0.007310699205845594, + 0.005915704648941755, + 0.008986442349851131, + 0.006608766037970781, + 0.007460458669811487, + 0.011940978467464447, + 0.007004343904554844, + 0.014834597706794739, + 0.009544199332594872, + 0.00632197642698884, + 0.005994594190269709, + 0.008980653248727322, + 0.005831995513290167, + 0.010117018595337868, + 0.005331289954483509, + 0.005617506802082062, + 0.007307849824428558, + 0.013264171779155731, + 0.014299549162387848, + 0.01951214298605919, + 0.023051736876368523, + 0.020463673397898674, + 0.0193390604108572, + 0.021781902760267258, + 0.022516721859574318, + 0.024386849254369736, + 0.012428255751729012, + 0.017324630171060562, + 0.010397949256002903, + 0.01555817760527134, + 0.055524252355098724, + 0.018505403771996498, + 0.013857640326023102, + 0.006720514968037605, + 0.004742152523249388, + 0.006274883169680834, + 0.0047316900454461575, + 0.005856648087501526, + 0.004104196093976498, + 0.0051764147356152534, + 0.004556193482130766, + 0.006743938196450472, + 0.004775041714310646, + 0.00547810597345233, + 0.0043220724910497665, + 0.0072814831510186195, + 0.003933726344257593, + 0.005181863903999329, + 0.005255086347460747, + 0.0037332591600716114, + 0.003976847976446152, + 0.0023886810522526503, + 0.004119200631976128, + 0.0035270683001726866, + 0.004381753038614988, + 0.0037351048085838556, + 0.003414520062506199, + 0.007899634540081024, + 0.005876346956938505, + 0.005718314554542303, + 0.00501229427754879, + 0.007688192650675774, + 0.009552506729960442, + 0.00893265102058649, + 0.0075041078962385654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.012538060545921326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium-long shot begins on a mountain terrace shrouded in morning mist. An elderly monk, his smooth head gleaming in the damp air, deep wrinkles etched across his serene face like rings of wisdom. He wears a gray cotton and linen robe, slightly whitish from age, its wide cuffs swaying gently with his movements. His feet are firmly rooted in the moist earth, his body erect and relaxed. His arms rise calmly from his sides, tracing a complete, smooth arc in front of him with a steady, inner-strength rhythm, his fingertips seemingly touching invisible currents of energy. The first soft rays of morning light penetrate the mist, casting a faint golden rim around his figure. The light creates a Tyndall effect within the ethereal haze, creating visible beams of light. His eyes are half-closed, a faint smile lingers on his lips, and his face carries a transcendental peace, as if blending seamlessly with the flowing air and the distant mountains. The painting's cool and subdued tones, combined with slow, continuous movements, create an atmosphere of profound tranquility, solemnity, and a sense of Zen.", + "seed": 19169, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_5/index_887_seed_19169_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013332765363156796, + "vlm_logprob": -0.013332765363156796, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.013332765363156796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026129860430955887 + ], + "opposite_logprob_tokens": [ + -3.776129961013794 + ], + "logprob_reward": -0.026129860430955887, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026129860430955887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1930617094039917, + 0.14245975017547607, + 0.037669871002435684, + 0.11769020557403564, + 0.06369583308696747, + 0.03237905353307724, + 0.1726483404636383, + 0.045440930873155594, + 0.11941570788621902, + 0.07553860545158386 + ], + "credit_map": [ + 0.01945067197084427, + 0.017114464193582535, + 0.021451788023114204, + 0.027919042855501175, + 0.023684095591306686, + 0.02862766943871975, + 0.0230653528124094, + 0.03174864500761032, + 0.013966514728963375, + 0.014842994511127472, + 0.01980150304734707, + 0.014593874104321003, + 0.014678644016385078, + 0.024303289130330086, + 0.013441259041428566, + 0.026831675320863724, + 0.005258296616375446, + 0.0042464639991521835, + 0.004195414017885923, + 0.004123001825064421, + 0.006382312159985304, + 0.004576446488499641, + 0.003362498711794615, + 0.005525439511984587, + 0.027557963505387306, + 0.017885880544781685, + 0.013370351865887642, + 0.01093446183949709, + 0.012553377076983452, + 0.016239898279309273, + 0.007502380292862654, + 0.011645902879536152, + 0.011912615969777107, + 0.009717542678117752, + 0.009030873887240887, + 0.004219010937958956, + 0.0057286471128463745, + 0.012282302603125572, + 0.005192124750465155, + 0.0056127202697098255, + 0.004812649916857481, + 0.004051459953188896, + 0.004877893254160881, + 0.002955226693302393, + 0.003928671590983868, + 0.0047987899743020535, + 0.003920889925211668, + 0.0030334731563925743, + 0.02531631663441658, + 0.01999344304203987, + 0.02086988091468811, + 0.018296057358384132, + 0.020707599818706512, + 0.026846498250961304, + 0.019649101421236992, + 0.02096944861114025, + 0.007491882424801588, + 0.003725648857653141, + 0.012760983780026436, + 0.0034297106321901083, + 0.003200495382770896, + 0.006879359018057585, + 0.002624708227813244, + 0.0053281462751328945, + 0.024644359946250916, + 0.007219088729470968, + 0.04070259630680084, + 0.007758097257465124, + 0.008283952251076698, + 0.008793015033006668, + 0.007722663693130016, + 0.014291933737695217, + 0.011066772975027561, + 0.010788392275571823, + 0.009119877591729164, + 0.008910955861210823, + 0.009884420782327652, + 0.011140335351228714, + 0.006091390270739794, + 0.008536460809409618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.026129860430955887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086177080869675 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.011086177080869675, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086177080869675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19648122787475586, + 0.11122781783342361, + 0.09911958128213882, + 0.05994964763522148, + 0.07849776744842529, + 0.03845832496881485, + 0.12166307121515274, + 0.05890398472547531, + 0.21690796315670013, + 0.01879056543111801 + ], + "credit_map": [ + 0.020957140251994133, + 0.01797446981072426, + 0.02330048754811287, + 0.026551680639386177, + 0.0272836834192276, + 0.02423463761806488, + 0.024616409093141556, + 0.031562723219394684, + 0.01139136590063572, + 0.01085966732352972, + 0.01913314498960972, + 0.015066195279359818, + 0.014183178544044495, + 0.01563614048063755, + 0.011620597913861275, + 0.013337535783648491, + 0.012080837041139603, + 0.010511605069041252, + 0.014640641398727894, + 0.011054793372750282, + 0.010763810947537422, + 0.014544735662639141, + 0.011326429434120655, + 0.014196732081472874, + 0.012854388915002346, + 0.00862863939255476, + 0.007924079895019531, + 0.005120523739606142, + 0.005311785731464624, + 0.010539979673922062, + 0.0045825704000890255, + 0.00498767988756299, + 0.011817692779004574, + 0.010081129148602486, + 0.008159306831657887, + 0.006119771394878626, + 0.010534048080444336, + 0.015288331545889378, + 0.00912967137992382, + 0.007367818150669336, + 0.004074643831700087, + 0.004379487596452236, + 0.007455466315150261, + 0.003315409878268838, + 0.005906753242015839, + 0.0066034612245857716, + 0.003962417598813772, + 0.002760685747489333, + 0.022893540561199188, + 0.017649386078119278, + 0.01596500724554062, + 0.013694864697754383, + 0.01232929341495037, + 0.016084400936961174, + 0.01021979097276926, + 0.01282679196447134, + 0.008546361699700356, + 0.007018771953880787, + 0.011940971948206425, + 0.005065822508186102, + 0.008925243280827999, + 0.006907861214131117, + 0.005640689749270678, + 0.004858267493546009, + 0.044594988226890564, + 0.013635832816362381, + 0.08183659613132477, + 0.013796397484838963, + 0.01036781445145607, + 0.014302977360785007, + 0.009899048134684563, + 0.028474321588873863, + 0.002802714006975293, + 0.002587537979707122, + 0.002486239420250058, + 0.002881965832784772, + 0.0019201886607334018, + 0.0022975378669798374, + 0.0019074081210419536, + 0.0019069749396294355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.011086177080869675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076298960484564304 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.0076298960484564304, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076298960484564304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0694749727845192, + 0.03244728967547417, + 0.034454114735126495, + 0.04729537293314934, + 0.06415548920631409, + 0.033370163291692734, + 0.1809452623128891, + 0.04788843169808388, + 0.32979950308799744, + 0.16016942262649536 + ], + "credit_map": [ + 0.007477758452296257, + 0.0076660774648189545, + 0.009520422667264938, + 0.011506128124892712, + 0.008407678455114365, + 0.007639467716217041, + 0.009183486923575401, + 0.008073952980339527, + 0.004039974883198738, + 0.003676515771076083, + 0.0042123934254050255, + 0.005134777631610632, + 0.0050334567204117775, + 0.0033177302684634924, + 0.003291246248409152, + 0.00374119752086699, + 0.0029840681236237288, + 0.003689986653625965, + 0.0052561271004378796, + 0.004454274661839008, + 0.003751009004190564, + 0.005144227761775255, + 0.0034997311886399984, + 0.00567468861117959, + 0.004218787420541048, + 0.004605610854923725, + 0.006545472424477339, + 0.004988538101315498, + 0.010186384432017803, + 0.006224125158041716, + 0.006436254363507032, + 0.004090199712663889, + 0.010691518895328045, + 0.009304291568696499, + 0.007764449343085289, + 0.005239009391516447, + 0.00843724049627781, + 0.0111751863732934, + 0.0055074794217944145, + 0.0060363151133060455, + 0.004436751827597618, + 0.005132876802235842, + 0.004779289476573467, + 0.0034595977049320936, + 0.002767210593447089, + 0.006404818966984749, + 0.003586528589949012, + 0.002803088165819645, + 0.03483531251549721, + 0.024518651887774467, + 0.024910327047109604, + 0.02329353243112564, + 0.014915237203240395, + 0.025265291333198547, + 0.014308233745396137, + 0.018898669630289078, + 0.005482681095600128, + 0.007034379057586193, + 0.005949424114078283, + 0.0062315682880580425, + 0.005217732861638069, + 0.005610318388789892, + 0.006481251213699579, + 0.005881077144294977, + 0.07134868204593658, + 0.01728592813014984, + 0.1305272877216339, + 0.019187701866030693, + 0.012379801832139492, + 0.02073086053133011, + 0.013973038643598557, + 0.04436619579792023, + 0.020305652171373367, + 0.024099623784422874, + 0.019864320755004883, + 0.021546106785535812, + 0.021062741056084633, + 0.02454369142651558, + 0.016000645235180855, + 0.012746644206345081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.0076298960484564304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076275295577943325 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.0076275295577943325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076275295577943325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23779213428497314, + 0.24619579315185547, + 0.06777486205101013, + 0.026844443753361702, + 0.0254485122859478, + 0.042383160442113876, + 0.04339668154716492, + 0.02922518365085125, + 0.1936570107936859, + 0.08728227764368057 + ], + "credit_map": [ + 0.028540607541799545, + 0.028552958741784096, + 0.034671634435653687, + 0.02961999922990799, + 0.03107445128262043, + 0.03662402555346489, + 0.02056155540049076, + 0.02814687229692936, + 0.01907098852097988, + 0.02401956170797348, + 0.03731589391827583, + 0.0316929891705513, + 0.02831104025244713, + 0.04117732122540474, + 0.021113477647304535, + 0.04349449649453163, + 0.010930567048490047, + 0.007460338529199362, + 0.007553569506853819, + 0.007808017544448376, + 0.008109552785754204, + 0.008099786005914211, + 0.006296616978943348, + 0.011516408063471317, + 0.004491306375712156, + 0.00325584365054965, + 0.002600566018372774, + 0.003125844057649374, + 0.004548349883407354, + 0.0028446754440665245, + 0.003322860924527049, + 0.0026549939066171646, + 0.003014773828908801, + 0.0030622135382145643, + 0.0030032857321202755, + 0.0028662311378866434, + 0.002967176726087928, + 0.003679888555780053, + 0.0032405215315520763, + 0.0036144184414297342, + 0.004614769946783781, + 0.004286403767764568, + 0.007355670444667339, + 0.004260576795786619, + 0.005630227271467447, + 0.007940271869301796, + 0.0049847508780658245, + 0.0033104857429862022, + 0.006634205114096403, + 0.005454619415104389, + 0.005831270478665829, + 0.0059714000672101974, + 0.0040403506718575954, + 0.005842444021254778, + 0.0039583491161465645, + 0.0056640394032001495, + 0.004086439497768879, + 0.003147086128592491, + 0.006127878092229366, + 0.0025657876394689083, + 0.00260877120308578, + 0.003683155169710517, + 0.0027888005133718252, + 0.004217262379825115, + 0.041415195912122726, + 0.010569658130407333, + 0.07546360790729523, + 0.011599299497902393, + 0.00825772900134325, + 0.013537322171032429, + 0.00627947086468339, + 0.026534710079431534, + 0.013578091748058796, + 0.01303754560649395, + 0.011626346036791801, + 0.0098901791498065, + 0.009917916730046272, + 0.01116815023124218, + 0.00781757291406393, + 0.010246466845273972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0076275295577943325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190363697707653 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190363697707653, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190363697707653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15430697798728943, + 0.07058937102556229, + 0.27100786566734314, + 0.05698392912745476, + 0.05612172931432724, + 0.05435720831155777, + 0.07377750426530838, + 0.09937514364719391, + 0.06893274933099747, + 0.09454749524593353 + ], + "credit_map": [ + 0.025149401277303696, + 0.018821818754076958, + 0.016112901270389557, + 0.01608186401426792, + 0.023422397673130035, + 0.01967334933578968, + 0.01984637975692749, + 0.01519886776804924, + 0.006155470386147499, + 0.0068723782896995544, + 0.010167429223656654, + 0.011613740585744381, + 0.008331908844411373, + 0.009081593714654446, + 0.009298122487962246, + 0.009068734012544155, + 0.05358797684311867, + 0.031018424779176712, + 0.030464166775345802, + 0.02969307452440262, + 0.031728629022836685, + 0.03208251670002937, + 0.025026068091392517, + 0.03740699589252472, + 0.006297396030277014, + 0.007091809064149857, + 0.007241143845021725, + 0.0045936815440654755, + 0.007065245416015387, + 0.01376214437186718, + 0.004446991253644228, + 0.006485518999397755, + 0.006803583353757858, + 0.007563639897853136, + 0.006905348505824804, + 0.0056334626860916615, + 0.005066320300102234, + 0.009845254942774773, + 0.006897507235407829, + 0.007406613323837519, + 0.007387926336377859, + 0.008000395260751247, + 0.00829665269702673, + 0.004041035193949938, + 0.004250854719430208, + 0.009416335262358189, + 0.0040655601769685745, + 0.008898451924324036, + 0.008508464321494102, + 0.010023467242717743, + 0.007322575896978378, + 0.00784089881926775, + 0.011029877699911594, + 0.010289713740348816, + 0.00887136161327362, + 0.00989114586263895, + 0.01535328570753336, + 0.00795687548816204, + 0.025414077565073967, + 0.007227805908769369, + 0.009557993151247501, + 0.015605125576257706, + 0.008100278675556183, + 0.010159704834222794, + 0.01344901230186224, + 0.005872416775673628, + 0.017672806978225708, + 0.0075431992299854755, + 0.005460415035486221, + 0.005224184598773718, + 0.006208009086549282, + 0.007502707652747631, + 0.012424280866980553, + 0.012964213266968727, + 0.010458468459546566, + 0.012985730543732643, + 0.013826491311192513, + 0.014245006255805492, + 0.008610869757831097, + 0.009032427333295345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.014190363697707653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium-long shot begins on a mountain terrace shrouded in morning mist. An elderly monk, his smooth head gleaming in the damp air, deep wrinkles etched across his serene face like rings of wisdom. He wears a gray cotton and linen robe, slightly whitish from age, its wide cuffs swaying gently with his movements. His feet are firmly rooted in the moist earth, his body erect and relaxed. His arms rise calmly from his sides, tracing a complete, smooth arc in front of him with a steady, inner-strength rhythm, his fingertips seemingly touching invisible currents of energy. The first soft rays of morning light penetrate the mist, casting a faint golden rim around his figure. The light creates a Tyndall effect within the ethereal haze, creating visible beams of light. His eyes are half-closed, a faint smile lingers on his lips, and his face carries a transcendental peace, as if blending seamlessly with the flowing air and the distant mountains. The painting's cool and subdued tones, combined with slow, continuous movements, create an atmosphere of profound tranquility, solemnity, and a sense of Zen.", + "seed": 19170, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_5/index_887_seed_19170_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019713559933006764, + "vlm_logprob": -0.019713559933006764, + "vlm_yes_no_margin": 4.024999952316284, + "vlm_token_credit": -0.019713559933006764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01390343252569437 + ], + "opposite_logprob_tokens": [ + -4.638903617858887 + ], + "logprob_reward": -0.01390343252569437, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01390343252569437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12198011577129364, + 0.5136438608169556, + 0.04078178107738495, + 0.03274558112025261, + 0.02635882794857025, + 0.03749842941761017, + 0.10989590734243393, + 0.030020713806152344, + 0.04233483597636223, + 0.04473992437124252 + ], + "credit_map": [ + 0.015274896286427975, + 0.01576501689851284, + 0.015055752359330654, + 0.01615048013627529, + 0.013119217939674854, + 0.012717642821371555, + 0.017685512080788612, + 0.016211597248911858, + 0.0754183903336525, + 0.19140957295894623, + 0.042043596506118774, + 0.019409792497754097, + 0.018183838576078415, + 0.12064475566148758, + 0.013678024522960186, + 0.03285593539476395, + 0.007156220730394125, + 0.008510992862284184, + 0.004794144071638584, + 0.002646942622959614, + 0.003075353102758527, + 0.005821645725518465, + 0.0051551759243011475, + 0.0036213055718690157, + 0.003848624648526311, + 0.00511305034160614, + 0.007349194027483463, + 0.005285540595650673, + 0.0023703244514763355, + 0.0030893241055309772, + 0.0024601216427981853, + 0.0032294036354869604, + 0.00251572928391397, + 0.0033813028130680323, + 0.004664071835577488, + 0.0037071097176522017, + 0.0029030051082372665, + 0.003370538353919983, + 0.003021012060344219, + 0.002796061336994171, + 0.0033473516814410686, + 0.006187791470438242, + 0.004132334608584642, + 0.003861265955492854, + 0.002441844204440713, + 0.007343601435422897, + 0.00340783828869462, + 0.006776401773095131, + 0.013217335566878319, + 0.03560899943113327, + 0.010020801797509193, + 0.005582400597631931, + 0.006539113353937864, + 0.026591377332806587, + 0.0072539495304226875, + 0.005081919953227043, + 0.003043816424906254, + 0.0035686821211129427, + 0.0057005081325769424, + 0.0035330953542143106, + 0.003072124905884266, + 0.0030723833478987217, + 0.004144798498600721, + 0.0038853047881275415, + 0.004117611795663834, + 0.004622192587703466, + 0.008102764375507832, + 0.0051300134509801865, + 0.005728861782699823, + 0.005727600306272507, + 0.004064333159476519, + 0.004841456655412912, + 0.009214488789439201, + 0.004912161268293858, + 0.00588638661429286, + 0.004513354040682316, + 0.003849865635856986, + 0.008102679625153542, + 0.0035571514163166285, + 0.004703836049884558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.01390343252569437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820533163845539 + ], + "opposite_logprob_tokens": [ + -4.018205165863037 + ], + "logprob_reward": -0.01820533163845539, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01820533163845539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1684017926454544, + 0.12448493391275406, + 0.07082333415746689, + 0.047074731439352036, + 0.02402445115149021, + 0.023581109941005707, + 0.3722032904624939, + 0.04711015895009041, + 0.04194188863039017, + 0.08035431057214737 + ], + "credit_map": [ + 0.019737709313631058, + 0.022368529811501503, + 0.020329171791672707, + 0.01806965097784996, + 0.020815422758460045, + 0.024934036657214165, + 0.023018313571810722, + 0.019128955900669098, + 0.01646191067993641, + 0.03905577212572098, + 0.012423054315149784, + 0.007653338368982077, + 0.005881958175450563, + 0.02551240287721157, + 0.006534116342663765, + 0.010962380096316338, + 0.008978277444839478, + 0.022162025794386864, + 0.007798281032592058, + 0.0038344981148838997, + 0.004397266078740358, + 0.013849575072526932, + 0.004352786112576723, + 0.0054506235755980015, + 0.0043240501545369625, + 0.010795854963362217, + 0.005964500829577446, + 0.003983256407082081, + 0.004473264794796705, + 0.00993126817047596, + 0.003390389494597912, + 0.004212148487567902, + 0.002397212665528059, + 0.002551126526668668, + 0.0036153451073914766, + 0.00364525499753654, + 0.00290851597674191, + 0.0023070916067808867, + 0.0034631371963769197, + 0.0031367680057883263, + 0.0026999739930033684, + 0.003685880685225129, + 0.003037907648831606, + 0.0022162594832479954, + 0.0019626785069704056, + 0.002861658576875925, + 0.0019192134495824575, + 0.005197538994252682, + 0.05125230923295021, + 0.14393098652362823, + 0.03201571851968765, + 0.012039289809763432, + 0.014800521545112133, + 0.08952786773443222, + 0.01574409194290638, + 0.012892498634755611, + 0.006975973956286907, + 0.004635823890566826, + 0.006548114120960236, + 0.003886888036504388, + 0.006531756371259689, + 0.007764420006424189, + 0.005727254785597324, + 0.0050399284809827805, + 0.004338860511779785, + 0.0040137446485459805, + 0.0053202910348773, + 0.007699677255004644, + 0.005891330540180206, + 0.004577744286507368, + 0.004919575527310371, + 0.005180664360523224, + 0.01167630311101675, + 0.006419521756470203, + 0.010211121290922165, + 0.011168323457241058, + 0.009413492865860462, + 0.009074308909475803, + 0.011555720120668411, + 0.010835522785782814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.01820533163845539 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981000766158104 + ], + "opposite_logprob_tokens": [ + -3.5298099517822266 + ], + "logprob_reward": -0.02981000766158104, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981000766158104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08176165074110031, + 0.12449797242879868, + 0.08942550420761108, + 0.0420202761888504, + 0.12753789126873016, + 0.08721518516540527, + 0.196610227227211, + 0.059043318033218384, + 0.15371668338775635, + 0.03817127272486687 + ], + "credit_map": [ + 0.01088236179202795, + 0.009487868286669254, + 0.010468930006027222, + 0.007265839725732803, + 0.011579396203160286, + 0.00953715667128563, + 0.011257924139499664, + 0.011282169260084629, + 0.018739983439445496, + 0.04386274516582489, + 0.011699930764734745, + 0.005674317479133606, + 0.006242549512535334, + 0.024571841582655907, + 0.005586537066847086, + 0.008120063692331314, + 0.009718082845211029, + 0.01914769597351551, + 0.010479834862053394, + 0.0062043024227023125, + 0.009118677116930485, + 0.013485914096236229, + 0.011256948113441467, + 0.01001405157148838, + 0.003284621983766556, + 0.006006795912981033, + 0.007746885996311903, + 0.00813987199217081, + 0.0025660148821771145, + 0.005716027691960335, + 0.004382372833788395, + 0.0041776844300329685, + 0.012061253190040588, + 0.019660569727420807, + 0.016692059114575386, + 0.0149803776293993, + 0.010188126936554909, + 0.021679827943444252, + 0.01587495021522045, + 0.016400732100009918, + 0.007452848367393017, + 0.009194652549922466, + 0.016640037298202515, + 0.008953476324677467, + 0.008128045126795769, + 0.008739063516259193, + 0.007493116892874241, + 0.02061394415795803, + 0.02431182749569416, + 0.0663633868098259, + 0.014627749100327492, + 0.010536127723753452, + 0.010913895443081856, + 0.04562043398618698, + 0.013234692625701427, + 0.011002106592059135, + 0.005703786388039589, + 0.0047631035558879375, + 0.006306870374828577, + 0.007818229496479034, + 0.008345725014805794, + 0.005662598181515932, + 0.00904533639550209, + 0.011397670954465866, + 0.016849685460329056, + 0.016886146739125252, + 0.024204032495617867, + 0.021342333406209946, + 0.022709155455231667, + 0.01708407700061798, + 0.016471361741423607, + 0.018169887363910675, + 0.004443773999810219, + 0.0034446960780769587, + 0.004800330847501755, + 0.0051851277239620686, + 0.00436920952051878, + 0.005682257004082203, + 0.005709544755518436, + 0.004536331165581942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.02981000766158104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10535407066345215, + 0.5076355338096619, + 0.047613367438316345, + 0.027868662029504776, + 0.031952451914548874, + 0.0505170039832592, + 0.0879058986902237, + 0.09429105371236801, + 0.028763815760612488, + 0.01809811219573021 + ], + "credit_map": [ + 0.011799925938248634, + 0.013068723492324352, + 0.020102672278881073, + 0.006920319516211748, + 0.014177212491631508, + 0.011669577099382877, + 0.012505398131906986, + 0.015110247768461704, + 0.060928478837013245, + 0.1496160924434662, + 0.06158188357949257, + 0.03739950805902481, + 0.025528742000460625, + 0.10301001369953156, + 0.030423089861869812, + 0.03914773836731911, + 0.005579805467277765, + 0.00945316907018423, + 0.00572440167888999, + 0.0041234758682549, + 0.00570098590105772, + 0.005664971191436052, + 0.005466692615300417, + 0.00589986564591527, + 0.0023717896547168493, + 0.0047745052725076675, + 0.00506772194057703, + 0.004519723355770111, + 0.0030394827481359243, + 0.003741688560694456, + 0.0020819089841097593, + 0.0022718405816704035, + 0.0027740513905882835, + 0.004149316810071468, + 0.004066763911396265, + 0.0037735742516815662, + 0.003538336604833603, + 0.004072291310876608, + 0.005123876500874758, + 0.0044542415998876095, + 0.004706012085080147, + 0.0062787821516394615, + 0.010281705297529697, + 0.008137169294059277, + 0.0034079805482178926, + 0.0063584647141397, + 0.004757406190037727, + 0.006589481141418219, + 0.01043469738215208, + 0.029007291421294212, + 0.013322600163519382, + 0.004843113012611866, + 0.004320337902754545, + 0.01619255356490612, + 0.0046473303809762, + 0.005137980915606022, + 0.014031754806637764, + 0.013384675607085228, + 0.013993359170854092, + 0.0066974107176065445, + 0.008579357527196407, + 0.015504264272749424, + 0.011080121621489525, + 0.011020115576684475, + 0.0026664959732443094, + 0.0031907199881970882, + 0.003555259434506297, + 0.005141192581504583, + 0.003494007047265768, + 0.0029542092233896255, + 0.004001891240477562, + 0.003760038875043392, + 0.002928027417510748, + 0.0019072864670306444, + 0.002840177621692419, + 0.001960763242095709, + 0.0022966524120420218, + 0.001992476172745228, + 0.0023092369083315134, + 0.0018634898588061333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.02058335766196251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065670177340508 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065670177340508, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065670177340508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12530091404914856, + 0.2479982078075409, + 0.060551226139068604, + 0.042455483227968216, + 0.051017142832279205, + 0.05974547564983368, + 0.24080421030521393, + 0.04025711491703987, + 0.0714372992515564, + 0.060432951897382736 + ], + "credit_map": [ + 0.012953199446201324, + 0.016612837091088295, + 0.02326587401330471, + 0.010980467312037945, + 0.015892570838332176, + 0.011606703512370586, + 0.015506986528635025, + 0.018482277169823647, + 0.037009332329034805, + 0.09441525489091873, + 0.01760280877351761, + 0.00798741728067398, + 0.008984976448118687, + 0.059470660984516144, + 0.008881543762981892, + 0.013646206818521023, + 0.0063234614208340645, + 0.013317259959876537, + 0.006678570061922073, + 0.004055526573210955, + 0.005773106124252081, + 0.010090476833283901, + 0.007886881940066814, + 0.006425945088267326, + 0.00400479044765234, + 0.008505447767674923, + 0.004778676200658083, + 0.004059627186506987, + 0.004684295505285263, + 0.00839778408408165, + 0.0039606643840670586, + 0.0040641967207193375, + 0.005010093096643686, + 0.007327063474804163, + 0.006236674264073372, + 0.0064194342121481895, + 0.0045563941821455956, + 0.007943048141896725, + 0.006494186818599701, + 0.007030246313661337, + 0.005246354732662439, + 0.007256728131324053, + 0.014559147879481316, + 0.008369396440684795, + 0.0039042748976498842, + 0.006707021966576576, + 0.0054321433417499065, + 0.008270411752164364, + 0.03301830217242241, + 0.08800902962684631, + 0.024917157366871834, + 0.011493326164782047, + 0.009805851615965366, + 0.05177757889032364, + 0.013144310563802719, + 0.008638634346425533, + 0.004992731846868992, + 0.003922894597053528, + 0.005715273320674896, + 0.00435646390542388, + 0.005289676133543253, + 0.005480654072016478, + 0.004851784557104111, + 0.005647636018693447, + 0.007657211739569902, + 0.007603960577398539, + 0.013080587610602379, + 0.007834789343178272, + 0.010058997198939323, + 0.010413335636258125, + 0.007567000109702349, + 0.007221415173262358, + 0.012786460109055042, + 0.005909566767513752, + 0.00685267336666584, + 0.00579115143045783, + 0.005911588203161955, + 0.010403459891676903, + 0.006378510966897011, + 0.006399544421583414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.016065670177340508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium-long shot begins on a mountain terrace shrouded in morning mist. An elderly monk, his smooth head gleaming in the damp air, deep wrinkles etched across his serene face like rings of wisdom. He wears a gray cotton and linen robe, slightly whitish from age, its wide cuffs swaying gently with his movements. His feet are firmly rooted in the moist earth, his body erect and relaxed. His arms rise calmly from his sides, tracing a complete, smooth arc in front of him with a steady, inner-strength rhythm, his fingertips seemingly touching invisible currents of energy. The first soft rays of morning light penetrate the mist, casting a faint golden rim around his figure. The light creates a Tyndall effect within the ethereal haze, creating visible beams of light. His eyes are half-closed, a faint smile lingers on his lips, and his face carries a transcendental peace, as if blending seamlessly with the flowing air and the distant mountains. The painting's cool and subdued tones, combined with slow, continuous movements, create an atmosphere of profound tranquility, solemnity, and a sense of Zen.", + "seed": 19171, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_5/index_887_seed_19171_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023235045187175275, + "vlm_logprob": -0.023235045187175275, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.023235045187175275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03504802659153938 + ], + "opposite_logprob_tokens": [ + -3.535048007965088 + ], + "logprob_reward": -0.03504802659153938, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03504802659153938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2502988874912262, + 0.07954148948192596, + 0.07908318936824799, + 0.09217344224452972, + 0.1373460590839386, + 0.08670130372047424, + 0.10337574779987335, + 0.038617681711912155, + 0.06156665086746216, + 0.07129551470279694 + ], + "credit_map": [ + 0.02258221246302128, + 0.019680209457874298, + 0.05357009917497635, + 0.041091930121183395, + 0.03293750062584877, + 0.03469298034906387, + 0.022278137505054474, + 0.02346581406891346, + 0.007227993104606867, + 0.006283741444349289, + 0.0168756190687418, + 0.01061255857348442, + 0.008655015379190445, + 0.009933496825397015, + 0.010106293484568596, + 0.009846771135926247, + 0.01397096086293459, + 0.005938195623457432, + 0.020731301978230476, + 0.006090013775974512, + 0.005010157357901335, + 0.011577344499528408, + 0.004758032970130444, + 0.011007186956703663, + 0.019746361300349236, + 0.00759198097512126, + 0.0213213749229908, + 0.007334600202739239, + 0.005642889533191919, + 0.01820421777665615, + 0.004281181842088699, + 0.00805083941668272, + 0.030477464199066162, + 0.011588492430746555, + 0.029744695872068405, + 0.011159421876072884, + 0.007241719868034124, + 0.03356614708900452, + 0.005902538541704416, + 0.007665571756660938, + 0.010843831114470959, + 0.006473645102232695, + 0.02373461425304413, + 0.00925484485924244, + 0.00547603378072381, + 0.019380943849682808, + 0.005972621031105518, + 0.005564775317907333, + 0.016767526045441628, + 0.008148682303726673, + 0.017732180655002594, + 0.017200544476509094, + 0.008740851655602455, + 0.00977230817079544, + 0.012070249766111374, + 0.012943407520651817, + 0.00397866265848279, + 0.0036113078240305185, + 0.006798340938985348, + 0.006472050677984953, + 0.005827354267239571, + 0.0035471110604703426, + 0.004257322289049625, + 0.004125531762838364, + 0.007850742898881435, + 0.00648109195753932, + 0.009595475159585476, + 0.009734828025102615, + 0.005714921746402979, + 0.010558201000094414, + 0.006819120142608881, + 0.004812270402908325, + 0.008324086666107178, + 0.010557492263615131, + 0.010273942723870277, + 0.009891333058476448, + 0.009279344230890274, + 0.006778303999453783, + 0.009276632219552994, + 0.0069143809378147125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.03504802659153938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558544985949993 + ], + "opposite_logprob_tokens": [ + -4.387558460235596 + ], + "logprob_reward": -0.012558544985949993, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558544985949993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15768194198608398, + 0.0782482698559761, + 0.08937634527683258, + 0.06024542078375816, + 0.06277507543563843, + 0.09527864307165146, + 0.14067043364048004, + 0.03512746840715408, + 0.08926615118980408, + 0.19133028388023376 + ], + "credit_map": [ + 0.018483441323041916, + 0.014969993382692337, + 0.021131087094545364, + 0.017523832619190216, + 0.025293773040175438, + 0.019513286650180817, + 0.02186610922217369, + 0.018900427967309952, + 0.006831135600805283, + 0.0061777373775839806, + 0.013346539810299873, + 0.009348579682409763, + 0.01356943603605032, + 0.008993539027869701, + 0.012637332081794739, + 0.0073439693078398705, + 0.012000457383692265, + 0.0072920783422887325, + 0.02126489207148552, + 0.007658521644771099, + 0.008652638643980026, + 0.014136944897472858, + 0.006738453172147274, + 0.01163235679268837, + 0.007825962267816067, + 0.004068590234965086, + 0.011670216917991638, + 0.007631405722349882, + 0.005984590854495764, + 0.008253653533756733, + 0.007600004784762859, + 0.007210996467620134, + 0.010462872684001923, + 0.004776800516992807, + 0.009041176177561283, + 0.011213213205337524, + 0.00706138089299202, + 0.006620283238589764, + 0.0077484240755438805, + 0.005850922781974077, + 0.010708995163440704, + 0.00853240117430687, + 0.021503973752260208, + 0.013094274327158928, + 0.0069482289254665375, + 0.01831655390560627, + 0.008910173550248146, + 0.0072640543803572655, + 0.01350207719951868, + 0.011524844914674759, + 0.02057829685509205, + 0.020797599107027054, + 0.019697679206728935, + 0.02278592251241207, + 0.01936938427388668, + 0.012414641678333282, + 0.003483739448711276, + 0.0035393338184803724, + 0.004357724450528622, + 0.005750675685703754, + 0.005446164403110743, + 0.0033838958479464054, + 0.004835921339690685, + 0.004330011084675789, + 0.009570125490427017, + 0.009786978363990784, + 0.010751512832939625, + 0.011882544495165348, + 0.016594067215919495, + 0.01282370463013649, + 0.010023396462202072, + 0.007833827286958694, + 0.014253983274102211, + 0.01999419741332531, + 0.02051202766597271, + 0.03328121453523636, + 0.02972574532032013, + 0.016939911991357803, + 0.034747254103422165, + 0.021875940263271332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.012558544985949993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058861404657364 + ], + "opposite_logprob_tokens": [ + -3.8955886363983154 + ], + "logprob_reward": -0.02058861404657364, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058861404657364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09440217167139053, + 0.09552659839391708, + 0.08669623732566833, + 0.052246708422899246, + 0.08058798313140869, + 0.028157111257314682, + 0.07634500414133072, + 0.06583422422409058, + 0.12334989011287689, + 0.2968541085720062 + ], + "credit_map": [ + 0.008208384737372398, + 0.007225509267300367, + 0.013852657750248909, + 0.01210371870547533, + 0.015683358535170555, + 0.01069032121449709, + 0.013715756125748158, + 0.012922472320497036, + 0.015099147334694862, + 0.008824174292385578, + 0.021853292360901833, + 0.011026659049093723, + 0.010277898982167244, + 0.0098267812281847, + 0.008667423389852047, + 0.009951218031346798, + 0.013643014244735241, + 0.004733249545097351, + 0.01985696516931057, + 0.00825450662523508, + 0.008191857486963272, + 0.00823486503213644, + 0.01031302660703659, + 0.013468749821186066, + 0.007053534034639597, + 0.0031971032731235027, + 0.005452585872262716, + 0.00885134469717741, + 0.0058099268935620785, + 0.004853240214288235, + 0.010106026194989681, + 0.006922949105501175, + 0.016763396561145782, + 0.006375795230269432, + 0.016214583069086075, + 0.006766539067029953, + 0.005057559348642826, + 0.018582729622721672, + 0.00562972342595458, + 0.005197660997509956, + 0.0036677063908427954, + 0.0027975342236459255, + 0.004291238263249397, + 0.0043257190845906734, + 0.002528614364564419, + 0.0035979929380118847, + 0.002985339844599366, + 0.003962965216487646, + 0.008917943574488163, + 0.008645526133477688, + 0.009344570338726044, + 0.010817897506058216, + 0.009112491272389889, + 0.008891339413821697, + 0.007976679131388664, + 0.012638555839657784, + 0.007855596952140331, + 0.006252355873584747, + 0.006600901950150728, + 0.012265053577721119, + 0.007719100918620825, + 0.005633043125271797, + 0.01073553878813982, + 0.008772631175816059, + 0.014107023365795612, + 0.010657678358256817, + 0.016402529552578926, + 0.022174514830112457, + 0.013454497791826725, + 0.02030358836054802, + 0.01613548956811428, + 0.010114574804902077, + 0.02474294789135456, + 0.02120823971927166, + 0.0255440641194582, + 0.0655832290649414, + 0.03936712443828583, + 0.02674158103764057, + 0.06318005919456482, + 0.030486876145005226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.02058861404657364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979542873799801 + ], + "opposite_logprob_tokens": [ + -3.5297954082489014 + ], + "logprob_reward": -0.02979542873799801, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979542873799801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32430288195610046, + 0.07714375108480453, + 0.13904252648353577, + 0.07026907056570053, + 0.10503993928432465, + 0.06375735253095627, + 0.047830335795879364, + 0.022908076643943787, + 0.03606090322136879, + 0.11364517360925674 + ], + "credit_map": [ + 0.029393715783953667, + 0.02763807214796543, + 0.06531310081481934, + 0.061703480780124664, + 0.03815511614084244, + 0.03175988793373108, + 0.038590043783187866, + 0.03174945339560509, + 0.015035577118396759, + 0.010345891118049622, + 0.010328252799808979, + 0.007899942807853222, + 0.007045498117804527, + 0.010320553556084633, + 0.0075254193507134914, + 0.008642621338367462, + 0.026493249461054802, + 0.010936463251709938, + 0.03592411056160927, + 0.010741820558905602, + 0.0065196966752409935, + 0.025303198024630547, + 0.005930989980697632, + 0.01719299703836441, + 0.013200775720179081, + 0.005777515005320311, + 0.01613016240298748, + 0.0061466023325920105, + 0.004423200152814388, + 0.013221366330981255, + 0.004350537899881601, + 0.007018910720944405, + 0.023876864463090897, + 0.008034612983465195, + 0.019783837720751762, + 0.00853984709829092, + 0.005561268888413906, + 0.024213869124650955, + 0.00692018261179328, + 0.00810946524143219, + 0.011350457556545734, + 0.0051388186402618885, + 0.014139793813228607, + 0.007388902362436056, + 0.003675412852317095, + 0.013153809122741222, + 0.004574696533381939, + 0.004335461184382439, + 0.004814340732991695, + 0.0037714713253080845, + 0.0068363710306584835, + 0.009609000757336617, + 0.005074694287031889, + 0.0041758217848837376, + 0.006876930128782988, + 0.0066717080771923065, + 0.00363758928142488, + 0.0031356564722955227, + 0.0028348672203719616, + 0.002458740258589387, + 0.002956738928332925, + 0.00322367693297565, + 0.002635953016579151, + 0.002024852903559804, + 0.0055888015776872635, + 0.0034485196229070425, + 0.006746971048414707, + 0.004674708936363459, + 0.003728780196979642, + 0.005633174441754818, + 0.003248002380132675, + 0.0029919485095888376, + 0.016705017536878586, + 0.012650231830775738, + 0.012051358819007874, + 0.016387436538934708, + 0.013400375843048096, + 0.009655910544097424, + 0.01912008970975876, + 0.013674750924110413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02979542873799801 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a mountain terrace shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly monk with a smooth head, deep wrinkles, and a gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the monk's body remain erect and relaxed while his arms rise in a complete, smooth arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible beams of light present in the scene, creating a Tyndall effect within the haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a slow, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184611573815346 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184611573815346, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184611573815346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3043559789657593, + 0.0864022746682167, + 0.05694042891263962, + 0.054886262863874435, + 0.07624199986457825, + 0.08473218977451324, + 0.14671431481838226, + 0.03931405022740364, + 0.04804283380508423, + 0.10236970335245132 + ], + "credit_map": [ + 0.024995507672429085, + 0.026479586958885193, + 0.03705238550901413, + 0.0662030503153801, + 0.04422697797417641, + 0.028732096776366234, + 0.045076481997966766, + 0.031589899212121964, + 0.009568395093083382, + 0.007189377211034298, + 0.019837548956274986, + 0.011288515292108059, + 0.010457479394972324, + 0.00897073745727539, + 0.009203881025314331, + 0.009886335581541061, + 0.010066056624054909, + 0.0033721323125064373, + 0.012243976816534996, + 0.006040815729647875, + 0.005218308884650469, + 0.006753938738256693, + 0.00560084730386734, + 0.007644350174814463, + 0.007103431969881058, + 0.0040330709889531136, + 0.009987515397369862, + 0.00857547763735056, + 0.004805872682482004, + 0.006004585884511471, + 0.0077529605478048325, + 0.006623344495892525, + 0.009677128866314888, + 0.004346364177763462, + 0.012974661774933338, + 0.010584760457277298, + 0.005830575246363878, + 0.010815251618623734, + 0.011314374394714832, + 0.010698872618377209, + 0.011515891179442406, + 0.00776065094396472, + 0.015146864578127861, + 0.017750659957528114, + 0.0051365382969379425, + 0.013197940774261951, + 0.008395644836127758, + 0.0058279866352677345, + 0.014963500201702118, + 0.01124926470220089, + 0.018817942589521408, + 0.025685176253318787, + 0.01660516858100891, + 0.015223141759634018, + 0.018369991332292557, + 0.025800123810768127, + 0.004577566869556904, + 0.004136952571570873, + 0.0055735791102051735, + 0.005425024777650833, + 0.006478575989603996, + 0.00416148267686367, + 0.005219221115112305, + 0.0037416410632431507, + 0.004760253243148327, + 0.0051746671088039875, + 0.0040344311855733395, + 0.005434983875602484, + 0.012561018578708172, + 0.005051442887634039, + 0.006683950312435627, + 0.004342083819210529, + 0.009234258905053139, + 0.009397336281836033, + 0.012218695133924484, + 0.02176106534898281, + 0.010890948586165905, + 0.008510359562933445, + 0.02182847447693348, + 0.008528551086783409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.018184611573815346 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_tensor([1961, 1961])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_tensor([1961, 1961])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a9b1164c69254bc3d5b67a0f8cda6c99d2faece --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_tensor([1961, 1961])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, moving slowly and steadily forward, almost touching the damp ground, centered on the back of a young Chinese man, kneeling deep in a dense, forgotten forest. His back hunched, his entire body etched with tension and concentration. He wore a frayed, dark gray hooded sweatshirt, the hood nearly obscuring his head and face, and dark, mud-stained trousers. His hands, stained with black earth, carefully and steadily lowered an irregularly shaped object wrapped in crimson silk into a freshly dug pit. The red cloth, illuminated by the single, unsettling, warm yellow light of an old lantern on the ground beside him, took on an overly saturated, almost unnatural brightness, creating a harsh contrast with the surrounding darkness, littered with decaying leaves. The lantern's light cast long, dancing shadows, making the surrounding tangled roots and gnarled trunks appear hideous, while the forest beyond was a deep, impenetrable darkness. Without using any tools, he used his fingers to gently yet decisively push the loose soil back into the pit, attempting to completely conceal the mysterious red bag. The entire scene was steeped in a deep sense of secrecy and unsettling ritual, creating an eerie, dark, and deeply oppressive atmosphere.", + "seed": 8605, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_1961_seed_8605_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05809092205017805, + "vlm_logprob": -0.05809092205017805, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.05809092205017805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16742587089538574 + ], + "opposite_logprob_tokens": [ + -1.9174258708953857 + ], + "logprob_reward": -0.16742587089538574, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16742587089538574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20105908811092377, + 0.13167287409305573, + 0.10766394436359406, + 0.07088646292686462, + 0.21697671711444855, + 0.048502784222364426, + 0.040464941412210464, + 0.044019915163517, + 0.08226097375154495, + 0.05649235472083092 + ], + "credit_map": [ + 0.023211931809782982, + 0.02382325753569603, + 0.022581055760383606, + 0.02243468351662159, + 0.026067359372973442, + 0.036571431905031204, + 0.017713159322738647, + 0.02865620143711567, + 0.01751193404197693, + 0.01442047394812107, + 0.022126536816358566, + 0.019405465573072433, + 0.013725819997489452, + 0.013370666652917862, + 0.01375858299434185, + 0.01735338382422924, + 0.02184727042913437, + 0.006225012708455324, + 0.028993645682930946, + 0.009159674867987633, + 0.008919418789446354, + 0.008102769032120705, + 0.011204572394490242, + 0.013211583718657494, + 0.008946792222559452, + 0.005659468472003937, + 0.013726766221225262, + 0.009098498150706291, + 0.008988451212644577, + 0.007968626916408539, + 0.0074204192496836185, + 0.009077434428036213, + 0.030772404745221138, + 0.019223643466830254, + 0.028341900557279587, + 0.05421169847249985, + 0.020601563155651093, + 0.018132058903574944, + 0.024883344769477844, + 0.020810076966881752, + 0.00478631304576993, + 0.00358021492138505, + 0.007883141748607159, + 0.00932392105460167, + 0.006164240185171366, + 0.005847032647579908, + 0.005647475365549326, + 0.005270442459732294, + 0.003634966444224119, + 0.0038609604816883802, + 0.005687647499144077, + 0.005532111506909132, + 0.005984574090689421, + 0.007099378388375044, + 0.0039906930178403854, + 0.0046746074222028255, + 0.0028801150619983673, + 0.004246213473379612, + 0.005357431247830391, + 0.007330821827054024, + 0.005905233323574066, + 0.00730725796893239, + 0.005718777421861887, + 0.005274059250950813, + 0.01082516834139824, + 0.008538906462490559, + 0.014550833031535149, + 0.015517857857048512, + 0.008931544609367847, + 0.008684600703418255, + 0.005325109697878361, + 0.009886949323117733, + 0.006247245240956545, + 0.005626177415251732, + 0.010090429335832596, + 0.010319123975932598, + 0.005131210666149855, + 0.007387507241219282, + 0.0036117772106081247, + 0.00807887502014637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.16742587089538574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221953883767128 + ], + "opposite_logprob_tokens": [ + -4.018221855163574 + ], + "logprob_reward": -0.018221953883767128, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018221953883767128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1404494047164917, + 0.24902120232582092, + 0.06367242336273193, + 0.03715827688574791, + 0.13775445520877838, + 0.06483330577611923, + 0.03908028453588486, + 0.047189027070999146, + 0.11749768257141113, + 0.10334392637014389 + ], + "credit_map": [ + 0.016300732269883156, + 0.018299955874681473, + 0.016843482851982117, + 0.018141265958547592, + 0.014723197557032108, + 0.019232237711548805, + 0.015457486733794212, + 0.021451031789183617, + 0.029297154396772385, + 0.01939903199672699, + 0.03669796139001846, + 0.025376219302415848, + 0.02408154495060444, + 0.0405978225171566, + 0.03377114608883858, + 0.03980032354593277, + 0.008815411478281021, + 0.005587941501289606, + 0.01120524387806654, + 0.010293304920196533, + 0.005026719532907009, + 0.006157267838716507, + 0.007580762729048729, + 0.009005773812532425, + 0.005198419094085693, + 0.003978526685386896, + 0.006608032155781984, + 0.005818788893520832, + 0.0025818063877522945, + 0.004834695253521204, + 0.002523417817428708, + 0.005614590365439653, + 0.018501684069633484, + 0.011998618021607399, + 0.019123997539281845, + 0.034703027456998825, + 0.01290660910308361, + 0.010019714012742043, + 0.015391471795737743, + 0.01510933879762888, + 0.006972073577344418, + 0.005277253221720457, + 0.012024693191051483, + 0.010840784758329391, + 0.007056661415845156, + 0.007730688434094191, + 0.006758196745067835, + 0.008172952570021152, + 0.00390732754021883, + 0.00300655048340559, + 0.006124987732619047, + 0.005768773145973682, + 0.005800049751996994, + 0.006399704609066248, + 0.003580506192520261, + 0.004492383915930986, + 0.005533118732273579, + 0.004276341758668423, + 0.0070401872508227825, + 0.00861686747521162, + 0.0057854861952364445, + 0.005517994984984398, + 0.004487065132707357, + 0.005931963212788105, + 0.01655152440071106, + 0.01498602144420147, + 0.022012902423739433, + 0.022519899532198906, + 0.009570776484906673, + 0.011616440489888191, + 0.00581092806532979, + 0.014429199509322643, + 0.010308693163096905, + 0.012469756416976452, + 0.017276570200920105, + 0.020875025540590286, + 0.009963003918528557, + 0.011468185111880302, + 0.006068547256290913, + 0.014914143830537796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.018221953883767128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214222319424152 + ], + "opposite_logprob_tokens": [ + -4.264214038848877 + ], + "logprob_reward": -0.014214222319424152, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214222319424152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11654862761497498, + 0.22140420973300934, + 0.11905179917812347, + 0.07800851762294769, + 0.18285232782363892, + 0.059067610651254654, + 0.06093465909361839, + 0.03771846368908882, + 0.04869740083813667, + 0.07571627199649811 + ], + "credit_map": [ + 0.01373439934104681, + 0.009953313507139683, + 0.01645374856889248, + 0.010733420960605145, + 0.01212029717862606, + 0.01875900663435459, + 0.013376761227846146, + 0.021417688578367233, + 0.022095078602433205, + 0.022655200213193893, + 0.023922927677631378, + 0.02835513837635517, + 0.020845841616392136, + 0.028977861627936363, + 0.04071446508169174, + 0.03383771702647209, + 0.026468520984053612, + 0.00557709950953722, + 0.03647177293896675, + 0.011042529717087746, + 0.007000631187111139, + 0.00899276789277792, + 0.009231874719262123, + 0.014266615733504295, + 0.009960404597222805, + 0.006754135247319937, + 0.011414241977036, + 0.01626819372177124, + 0.007246715947985649, + 0.008374933153390884, + 0.008469902910292149, + 0.009519994258880615, + 0.024517929181456566, + 0.016639532521367073, + 0.02333788387477398, + 0.042431533336639404, + 0.020122896879911423, + 0.015150616876780987, + 0.022677065804600716, + 0.01797489821910858, + 0.006891719996929169, + 0.004404095467180014, + 0.008087598718702793, + 0.010318674147129059, + 0.0061153387650847435, + 0.006574724335223436, + 0.010153140872716904, + 0.006522323004901409, + 0.005626948084682226, + 0.004090733826160431, + 0.008178920485079288, + 0.008922124281525612, + 0.008874270133674145, + 0.00913150329142809, + 0.009628040716052055, + 0.006482122931629419, + 0.003417357336729765, + 0.003224026644602418, + 0.006386454217135906, + 0.0063234115950763226, + 0.0034214083570986986, + 0.0046113841235637665, + 0.0037146080285310745, + 0.0066198199056088924, + 0.004401111509650946, + 0.004408461973071098, + 0.006118457764387131, + 0.00983229000121355, + 0.005282488185912371, + 0.006470848806202412, + 0.005018462426960468, + 0.007165282499045134, + 0.0072953286580741405, + 0.007757470477372408, + 0.00883164070546627, + 0.019138136878609657, + 0.005850734189152718, + 0.008234510198235512, + 0.00439379783347249, + 0.014214661903679371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.014214222319424152 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579971373081207 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.020579971373081207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579971373081207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11582774668931961, + 0.18436889350414276, + 0.23008687794208527, + 0.06444111466407776, + 0.06817667186260223, + 0.055319204926490784, + 0.05548018589615822, + 0.14583079516887665, + 0.04059816524386406, + 0.03987039625644684 + ], + "credit_map": [ + 0.013268831185996532, + 0.013810616917908192, + 0.015147446654736996, + 0.01540257129818201, + 0.014274845831096172, + 0.012525989674031734, + 0.01446167565882206, + 0.016935760155320168, + 0.02371232770383358, + 0.014776954427361488, + 0.02629847079515457, + 0.0239318385720253, + 0.017852121964097023, + 0.02518312819302082, + 0.01684209704399109, + 0.03577194735407829, + 0.027805527672171593, + 0.01644578017294407, + 0.030068593099713326, + 0.03127766773104668, + 0.023463090881705284, + 0.022972911596298218, + 0.041216231882572174, + 0.036837078630924225, + 0.007520835846662521, + 0.0045422520488500595, + 0.011214924044907093, + 0.010340816341340542, + 0.005561290308833122, + 0.008785462938249111, + 0.007277790457010269, + 0.009197746403515339, + 0.009163551963865757, + 0.0060705626383423805, + 0.009885145351290703, + 0.010253381915390491, + 0.006768273189663887, + 0.009073407389223576, + 0.008988755755126476, + 0.007973594591021538, + 0.005014732945710421, + 0.004357514902949333, + 0.00951369944959879, + 0.008090276271104813, + 0.005994951352477074, + 0.007300025317817926, + 0.007616553455591202, + 0.007431452628225088, + 0.006561612244695425, + 0.005437193438410759, + 0.006680856924504042, + 0.00840936228632927, + 0.006191523279994726, + 0.005686193238943815, + 0.0087974201887846, + 0.007716026157140732, + 0.012100951746106148, + 0.013814304023981094, + 0.021352900192141533, + 0.023377245292067528, + 0.016233066096901894, + 0.0157525185495615, + 0.02002144232392311, + 0.023178374394774437, + 0.005306519567966461, + 0.005027189385145903, + 0.005721042398363352, + 0.007903539575636387, + 0.0035775601863861084, + 0.004957886878401041, + 0.0032056423369795084, + 0.004898786079138517, + 0.0032313528936356306, + 0.003188349772244692, + 0.006727471947669983, + 0.008480682037770748, + 0.0035588592290878296, + 0.004326752386987209, + 0.003781604813411832, + 0.0065753222443163395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.020579971373081207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001259177923203 + ], + "opposite_logprob_tokens": [ + -2.6950125694274902 + ], + "logprob_reward": -0.07001259177923203, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001259177923203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09583581238985062, + 0.21535678207874298, + 0.11788070946931839, + 0.12435238808393478, + 0.02106350287795067, + 0.04251396283507347, + 0.03551587834954262, + 0.10350054502487183, + 0.07985427975654602, + 0.16412605345249176 + ], + "credit_map": [ + 0.009134095162153244, + 0.007856200449168682, + 0.013441036455333233, + 0.01773759536445141, + 0.011383150704205036, + 0.015277874656021595, + 0.011024829000234604, + 0.009981044568121433, + 0.024476459249854088, + 0.017808517441153526, + 0.02739468775689602, + 0.02822197787463665, + 0.023860737681388855, + 0.02877306379377842, + 0.027779098600149155, + 0.0370422787964344, + 0.019822729751467705, + 0.0063433521427214146, + 0.03671963885426521, + 0.014465381391346455, + 0.006370866671204567, + 0.009676745161414146, + 0.009122263640165329, + 0.015359745360910892, + 0.014210238121449947, + 0.013404744677245617, + 0.017288831993937492, + 0.027233021333813667, + 0.014638643711805344, + 0.012060713022947311, + 0.011847034096717834, + 0.01366917323321104, + 0.0022789654321968555, + 0.0020042124669998884, + 0.0031729983165860176, + 0.003719583386555314, + 0.002864204114302993, + 0.0028707212768495083, + 0.0019267668249085546, + 0.0022260521072894335, + 0.00403548264876008, + 0.004093225579708815, + 0.007007393520325422, + 0.008006670512259007, + 0.005891589913517237, + 0.004778730217367411, + 0.004002244211733341, + 0.00469862949103117, + 0.00318109430372715, + 0.003782447427511215, + 0.004614302422851324, + 0.007881457917392254, + 0.004851588048040867, + 0.0031373195815831423, + 0.0034693554043769836, + 0.004598315805196762, + 0.008808144368231297, + 0.011176702566444874, + 0.015814976766705513, + 0.015572573058307171, + 0.015111930668354034, + 0.014396804384887218, + 0.011048492975533009, + 0.011570927686989307, + 0.008053838275372982, + 0.009555368684232235, + 0.011423743329942226, + 0.015795350074768066, + 0.00864053051918745, + 0.011231871321797371, + 0.006050959229469299, + 0.009102634154260159, + 0.015002497471868992, + 0.016010453924536705, + 0.022728383541107178, + 0.03715907037258148, + 0.014674963429570198, + 0.021343398839235306, + 0.010938454419374466, + 0.026268864050507545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.07001259177923203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, moving slowly and steadily forward, almost touching the damp ground, centered on the back of a young Chinese man, kneeling deep in a dense, forgotten forest. His back hunched, his entire body etched with tension and concentration. He wore a frayed, dark gray hooded sweatshirt, the hood nearly obscuring his head and face, and dark, mud-stained trousers. His hands, stained with black earth, carefully and steadily lowered an irregularly shaped object wrapped in crimson silk into a freshly dug pit. The red cloth, illuminated by the single, unsettling, warm yellow light of an old lantern on the ground beside him, took on an overly saturated, almost unnatural brightness, creating a harsh contrast with the surrounding darkness, littered with decaying leaves. The lantern's light cast long, dancing shadows, making the surrounding tangled roots and gnarled trunks appear hideous, while the forest beyond was a deep, impenetrable darkness. Without using any tools, he used his fingers to gently yet decisively push the loose soil back into the pit, attempting to completely conceal the mysterious red bag. The entire scene was steeped in a deep sense of secrecy and unsettling ritual, creating an eerie, dark, and deeply oppressive atmosphere.", + "seed": 8606, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_1961_seed_8606_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07897335616871715, + "vlm_logprob": -0.07897335616871715, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07897335616871715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28799882531166077 + ], + "opposite_logprob_tokens": [ + -1.4129987955093384 + ], + "logprob_reward": -0.28799882531166077, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28799882531166077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15556997060775757, + 0.03807305544614792, + 0.037768952548503876, + 0.4135247468948364, + 0.04366207867860794, + 0.04898960143327713, + 0.044926468282938004, + 0.04150204360485077, + 0.07898567616939545, + 0.09699746966362 + ], + "credit_map": [ + 0.017586378380656242, + 0.02107253670692444, + 0.013938120566308498, + 0.012540892697870731, + 0.02156713418662548, + 0.017619727179408073, + 0.02538950741291046, + 0.025855669751763344, + 0.005720821674913168, + 0.005004276987165213, + 0.004012996796518564, + 0.0021274862810969353, + 0.004850062541663647, + 0.005685634445399046, + 0.0030220088083297014, + 0.007649766281247139, + 0.004583298694342375, + 0.004626575391739607, + 0.0026778103783726692, + 0.0028698453679680824, + 0.00627055112272501, + 0.0061927358619868755, + 0.004397004377096891, + 0.00615113228559494, + 0.029090868309140205, + 0.025427939370274544, + 0.008425609208643436, + 0.008992732502520084, + 0.017601119354367256, + 0.016772789880633354, + 0.14751386642456055, + 0.15969981253147125, + 0.005573052912950516, + 0.004600604064762592, + 0.003626425052061677, + 0.003775710705667734, + 0.004436739254742861, + 0.010172692127525806, + 0.003723457222804427, + 0.0077533964067697525, + 0.00566839287057519, + 0.0062033627182245255, + 0.005002478137612343, + 0.004939910490065813, + 0.006131879519671202, + 0.007880053482949734, + 0.00728857796639204, + 0.005874946713447571, + 0.004683042410761118, + 0.005061168689280748, + 0.0027678105980157852, + 0.0041626510210335255, + 0.006752051413059235, + 0.008788502775132656, + 0.007844023406505585, + 0.004867215640842915, + 0.003876134753227234, + 0.0037053031846880913, + 0.0029200820717960596, + 0.003628549864515662, + 0.007652992382645607, + 0.0071236793883144855, + 0.007633392233401537, + 0.0049619111232459545, + 0.01008937880396843, + 0.005862419493496418, + 0.005173390731215477, + 0.005784458015114069, + 0.011362828314304352, + 0.020619310438632965, + 0.010131779126822948, + 0.0099621107801795, + 0.007089391816407442, + 0.013981585390865803, + 0.005089754704385996, + 0.013773075304925442, + 0.011579852551221848, + 0.016627898439764977, + 0.015295391902327538, + 0.013560516759753227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.28799882531166077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02986878715455532 + ], + "opposite_logprob_tokens": [ + -3.5298688411712646 + ], + "logprob_reward": -0.02986878715455532, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02986878715455532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22908012568950653, + 0.026096098124980927, + 0.02286750264465809, + 0.3174951374530792, + 0.035080477595329285, + 0.034763310104608536, + 0.05460314452648163, + 0.09841763973236084, + 0.13576462864875793, + 0.04583194479346275 + ], + "credit_map": [ + 0.02584761008620262, + 0.030152395367622375, + 0.024762142449617386, + 0.01780063845217228, + 0.03370141610503197, + 0.025582123547792435, + 0.03346702829003334, + 0.037766776978969574, + 0.0026314714923501015, + 0.0032312320545315742, + 0.004021035972982645, + 0.0027015842497348785, + 0.002707670209929347, + 0.003002526005730033, + 0.002606750465929508, + 0.005193828139454126, + 0.004027297254651785, + 0.0028950199484825134, + 0.0033172734547406435, + 0.0014603595482185483, + 0.003405909985303879, + 0.003159944899380207, + 0.001590013038367033, + 0.003011683002114296, + 0.02235995978116989, + 0.019569922238588333, + 0.006379490252584219, + 0.006834080908447504, + 0.01366118248552084, + 0.012564174830913544, + 0.11331112682819366, + 0.12281519174575806, + 0.004200868774205446, + 0.00502428226172924, + 0.0036426216829568148, + 0.004340555053204298, + 0.004028583876788616, + 0.005627152044326067, + 0.003407045966014266, + 0.004809365142136812, + 0.004246370866894722, + 0.002979511162266135, + 0.004768801387399435, + 0.0028313733637332916, + 0.003237940138205886, + 0.006766744423657656, + 0.004990239627659321, + 0.004942325409501791, + 0.00400584377348423, + 0.004167116247117519, + 0.0019024063367396593, + 0.0032880748622119427, + 0.01569967158138752, + 0.014795675873756409, + 0.005363247357308865, + 0.005381109192967415, + 0.009547718800604343, + 0.012004000134766102, + 0.006489616818726063, + 0.008628621697425842, + 0.016458263620734215, + 0.01492328755557537, + 0.015128917060792446, + 0.015237206593155861, + 0.0059045241214334965, + 0.004365882836282253, + 0.003436862025409937, + 0.004447322804480791, + 0.048731472343206406, + 0.05266046151518822, + 0.007223018445074558, + 0.008995088748633862, + 0.003991438075900078, + 0.007722546346485615, + 0.003900150302797556, + 0.005573614034801722, + 0.005457084160298109, + 0.007401160430163145, + 0.0076056853868067265, + 0.004180267918854952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.02986878715455532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016100158914923668 + ], + "opposite_logprob_tokens": [ + -4.14109992980957 + ], + "logprob_reward": -0.016100158914923668, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016100158914923668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29334694147109985, + 0.08580973744392395, + 0.1863362342119217, + 0.0502450168132782, + 0.05862892419099808, + 0.03618622571229935, + 0.04230808466672897, + 0.04828955978155136, + 0.07071676105260849, + 0.12813255190849304 + ], + "credit_map": [ + 0.04662652686238289, + 0.03431476652622223, + 0.03387182205915451, + 0.023613503202795982, + 0.03925919905304909, + 0.03651537746191025, + 0.03602219372987747, + 0.04312357306480408, + 0.007290374021977186, + 0.009821721352636814, + 0.011801380664110184, + 0.011508124880492687, + 0.010827742516994476, + 0.013248349539935589, + 0.01065268274396658, + 0.010659368708729744, + 0.024631528183817863, + 0.025187848135828972, + 0.013013288378715515, + 0.014088899828493595, + 0.026390984654426575, + 0.030008304864168167, + 0.020861821249127388, + 0.03215355798602104, + 0.0052361926063895226, + 0.00420722458511591, + 0.0029116973746567965, + 0.0029374368023127317, + 0.0032899018842726946, + 0.0025369510985910892, + 0.016423892229795456, + 0.012701720930635929, + 0.0076491753570735455, + 0.006803547032177448, + 0.005754449404776096, + 0.011126356199383736, + 0.005250442773103714, + 0.006051560863852501, + 0.006127247121185064, + 0.009866148233413696, + 0.0037943939678370953, + 0.003705560462549329, + 0.004488736856728792, + 0.0037965234369039536, + 0.00566696934401989, + 0.004913328215479851, + 0.004443615674972534, + 0.0053770942613482475, + 0.002966713160276413, + 0.0040051876567304134, + 0.002390338107943535, + 0.0032708309590816498, + 0.008779698982834816, + 0.012090472504496574, + 0.004651479888707399, + 0.004153364337980747, + 0.004752897657454014, + 0.0050992718897759914, + 0.0029123390559107065, + 0.004152129869908094, + 0.00991017371416092, + 0.007776862941682339, + 0.007446286268532276, + 0.006239595822989941, + 0.0048998454585671425, + 0.0036830902099609375, + 0.003174093784764409, + 0.0036516839172691107, + 0.0182614978402853, + 0.02277422696352005, + 0.0052726068533957005, + 0.008999712765216827, + 0.013468406163156033, + 0.014835990034043789, + 0.01683342270553112, + 0.010712780058383942, + 0.01708870753645897, + 0.01919347234070301, + 0.015234974212944508, + 0.020764801651239395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.016100158914923668 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950355436652899 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950355436652899, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950355436652899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15664111077785492, + 0.058896951377391815, + 0.12145574390888214, + 0.06947048753499985, + 0.056156303733587265, + 0.057637542486190796, + 0.03839825466275215, + 0.0688374787569046, + 0.12289918214082718, + 0.24960699677467346 + ], + "credit_map": [ + 0.026773663237690926, + 0.018930897116661072, + 0.01861627586185932, + 0.015265222638845444, + 0.019633764401078224, + 0.015112093649804592, + 0.018227461725473404, + 0.024081716313958168, + 0.011517690494656563, + 0.007441986817866564, + 0.003194327699020505, + 0.0028708074241876602, + 0.0070067644119262695, + 0.0064681777730584145, + 0.003625597571954131, + 0.01677159033715725, + 0.016372866928577423, + 0.018215563148260117, + 0.01039179041981697, + 0.007813120260834694, + 0.01679188199341297, + 0.018939966335892677, + 0.01392360683530569, + 0.019006935879588127, + 0.009061434306204319, + 0.005076519679278135, + 0.002857684390619397, + 0.0034289993345737457, + 0.004684979561716318, + 0.006785632576793432, + 0.01361047476530075, + 0.023964760825037956, + 0.006965795066207647, + 0.006909634917974472, + 0.005800419952720404, + 0.005343484692275524, + 0.006424409337341785, + 0.011747115291655064, + 0.004839958157390356, + 0.008125481195747852, + 0.006633824203163385, + 0.006583784706890583, + 0.007048638537526131, + 0.005661829374730587, + 0.007044508587568998, + 0.009540772996842861, + 0.006905232556164265, + 0.008218950591981411, + 0.0040156301110982895, + 0.0034766902681440115, + 0.0027847562450915575, + 0.0032594387885183096, + 0.00872967578470707, + 0.005645462311804295, + 0.005715058650821447, + 0.004771541338413954, + 0.0068349288776516914, + 0.006586082279682159, + 0.006135256960988045, + 0.00728896539658308, + 0.01290745660662651, + 0.010137180797755718, + 0.010797597467899323, + 0.008150001987814903, + 0.009226467460393906, + 0.006302144378423691, + 0.0038173203356564045, + 0.006239098031073809, + 0.03417924419045448, + 0.02781994640827179, + 0.014815419912338257, + 0.02049952559173107, + 0.03259950876235962, + 0.02882031910121441, + 0.037284985184669495, + 0.029424209147691727, + 0.0328950360417366, + 0.03864980861544609, + 0.02670232020318508, + 0.023230772465467453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.005950355436652899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054948654025793076 + ], + "opposite_logprob_tokens": [ + -2.929948568344116 + ], + "logprob_reward": -0.054948654025793076, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054948654025793076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09648354351520538, + 0.03039301186800003, + 0.0809536725282669, + 0.04581041634082794, + 0.033172763884067535, + 0.08744865655899048, + 0.20721125602722168, + 0.10723716765642166, + 0.2749757766723633, + 0.03631370887160301 + ], + "credit_map": [ + 0.015049625188112259, + 0.010858041234314442, + 0.012023183517158031, + 0.009298044256865978, + 0.00953675527125597, + 0.00944956298917532, + 0.012964567169547081, + 0.017303766682744026, + 0.0049848598428070545, + 0.0035395422019064426, + 0.0031743301078677177, + 0.003310024505481124, + 0.0033100545406341553, + 0.0036930886562913656, + 0.00296865520067513, + 0.0054124570451676846, + 0.011734681203961372, + 0.012240789830684662, + 0.006627852097153664, + 0.005579653661698103, + 0.011095184832811356, + 0.01158873550593853, + 0.008484160527586937, + 0.013602613471448421, + 0.005233980715274811, + 0.003705359296873212, + 0.0019731896463781595, + 0.0019446945516392589, + 0.003961475100368261, + 0.002957795513793826, + 0.01084426324814558, + 0.015189656987786293, + 0.004747958853840828, + 0.0037968656979501247, + 0.0035237569827586412, + 0.003443642985075712, + 0.002754941815510392, + 0.0054083094000816345, + 0.0029447919223457575, + 0.006552493665367365, + 0.009980627335608006, + 0.009642318822443485, + 0.006322767119854689, + 0.0073270611464977264, + 0.012931582517921925, + 0.016586188226938248, + 0.012413243763148785, + 0.012244859710335732, + 0.01480051502585411, + 0.011686684563755989, + 0.006868540775030851, + 0.011355728842318058, + 0.07422798871994019, + 0.045373085886240005, + 0.027956893667578697, + 0.014941824600100517, + 0.010425621643662453, + 0.014425491914153099, + 0.0077244373969733715, + 0.00960596464574337, + 0.0147669967263937, + 0.01535297092050314, + 0.020690739154815674, + 0.01424494106322527, + 0.010135981254279613, + 0.005378211848437786, + 0.004513636697083712, + 0.005120974499732256, + 0.10722348839044571, + 0.10746228694915771, + 0.014673017896711826, + 0.02046819217503071, + 0.0027448637410998344, + 0.006175504997372627, + 0.0020426735281944275, + 0.0042093172669410706, + 0.005085740238428116, + 0.006224065087735653, + 0.006500599905848503, + 0.0033309427089989185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.054948654025793076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, moving slowly and steadily forward, almost touching the damp ground, centered on the back of a young Chinese man, kneeling deep in a dense, forgotten forest. His back hunched, his entire body etched with tension and concentration. He wore a frayed, dark gray hooded sweatshirt, the hood nearly obscuring his head and face, and dark, mud-stained trousers. His hands, stained with black earth, carefully and steadily lowered an irregularly shaped object wrapped in crimson silk into a freshly dug pit. The red cloth, illuminated by the single, unsettling, warm yellow light of an old lantern on the ground beside him, took on an overly saturated, almost unnatural brightness, creating a harsh contrast with the surrounding darkness, littered with decaying leaves. The lantern's light cast long, dancing shadows, making the surrounding tangled roots and gnarled trunks appear hideous, while the forest beyond was a deep, impenetrable darkness. Without using any tools, he used his fingers to gently yet decisively push the loose soil back into the pit, attempting to completely conceal the mysterious red bag. The entire scene was steeped in a deep sense of secrecy and unsettling ritual, creating an eerie, dark, and deeply oppressive atmosphere.", + "seed": 8607, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_1961_seed_8607_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04476766176521778, + "vlm_logprob": -0.04476766176521778, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.04476766176521778, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11534353345632553 + ], + "opposite_logprob_tokens": [ + -2.2403435707092285 + ], + "logprob_reward": -0.11534353345632553, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11534353345632553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19628559052944183, + 0.10405147820711136, + 0.07720126211643219, + 0.05242869630455971, + 0.0640450119972229, + 0.07939733564853668, + 0.14110371470451355, + 0.16053369641304016, + 0.03799979388713837, + 0.08695340156555176 + ], + "credit_map": [ + 0.022156566381454468, + 0.025224672630429268, + 0.01813373528420925, + 0.024252286180853844, + 0.022372301667928696, + 0.03201206028461456, + 0.02793814428150654, + 0.024195821955800056, + 0.01049705222249031, + 0.017491275444626808, + 0.009964657947421074, + 0.01210572849959135, + 0.01669277995824814, + 0.01916251704096794, + 0.00923059694468975, + 0.00890686921775341, + 0.0085888197645545, + 0.011081844568252563, + 0.00897789467126131, + 0.00901866890490055, + 0.012312795966863632, + 0.011096573434770107, + 0.007462032604962587, + 0.0086626335978508, + 0.00418132683262229, + 0.005323569290339947, + 0.006837617605924606, + 0.005273936316370964, + 0.006311115808784962, + 0.00691998191177845, + 0.0054883575066924095, + 0.012092791497707367, + 0.004129551816731691, + 0.007059423252940178, + 0.009343464858829975, + 0.007720730733126402, + 0.009903568774461746, + 0.009507572278380394, + 0.009511898271739483, + 0.006868799217045307, + 0.006599705666303635, + 0.008916676044464111, + 0.008902284316718578, + 0.012612681835889816, + 0.011211289092898369, + 0.01004891935735941, + 0.012509919703006744, + 0.00859585776925087, + 0.016659541055560112, + 0.028935208916664124, + 0.017396805807948112, + 0.00997373927384615, + 0.021879691630601883, + 0.02126118913292885, + 0.013580290600657463, + 0.011417240835726261, + 0.015545720234513283, + 0.022892285138368607, + 0.01784416288137436, + 0.010602245107293129, + 0.04536154121160507, + 0.026621485128998756, + 0.011072488501667976, + 0.010593770071864128, + 0.0029105348512530327, + 0.005720791406929493, + 0.003976997919380665, + 0.0044621978886425495, + 0.006388166919350624, + 0.0042528752237558365, + 0.006870252545922995, + 0.0034179799258708954, + 0.010972551070153713, + 0.0104923564940691, + 0.009442474693059921, + 0.010376134887337685, + 0.011469332501292229, + 0.01225754152983427, + 0.009167388081550598, + 0.012775616720318794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.11534353345632553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020622828975319862 + ], + "opposite_logprob_tokens": [ + -3.895622730255127 + ], + "logprob_reward": -0.020622828975319862, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020622828975319862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14504842460155487, + 0.14218822121620178, + 0.10275010019540787, + 0.09275595843791962, + 0.04887101799249649, + 0.08755867183208466, + 0.07146266102790833, + 0.0645715594291687, + 0.07036250829696655, + 0.17443083226680756 + ], + "credit_map": [ + 0.02381669543683529, + 0.014443193562328815, + 0.01876973919570446, + 0.018416855484247208, + 0.01736036315560341, + 0.019358260557055473, + 0.015130119398236275, + 0.017753198742866516, + 0.014499345794320107, + 0.025856707245111465, + 0.01464433129876852, + 0.02081490494310856, + 0.020784448832273483, + 0.01962696760892868, + 0.014571222476661205, + 0.01139030046761036, + 0.011481169611215591, + 0.012681527063250542, + 0.014197446405887604, + 0.011856239289045334, + 0.013825556263327599, + 0.01503381785005331, + 0.008038938976824284, + 0.015635408461093903, + 0.009140382520854473, + 0.014156327582895756, + 0.017509056255221367, + 0.011559371836483479, + 0.010976406745612621, + 0.01263967901468277, + 0.008217290975153446, + 0.008557443507015705, + 0.0037550872657448053, + 0.0057787811383605, + 0.0085001764819026, + 0.004682152532041073, + 0.00915258564054966, + 0.007635844871401787, + 0.004363167565315962, + 0.005003220867365599, + 0.007527806796133518, + 0.0111091872677207, + 0.012568647041916847, + 0.009412498213350773, + 0.016468470916152, + 0.011874540708959103, + 0.009044927544891834, + 0.009552596136927605, + 0.006759537849575281, + 0.00957409292459488, + 0.00918193906545639, + 0.006068806629627943, + 0.012266035191714764, + 0.013975756242871284, + 0.007016545161604881, + 0.006619944702833891, + 0.004514461848884821, + 0.0073141856119036674, + 0.011553313583135605, + 0.005975701846182346, + 0.018253905698657036, + 0.006987484637647867, + 0.0054356190375983715, + 0.004536888562142849, + 0.005269662477076054, + 0.008832794614136219, + 0.007903693243861198, + 0.008444364182651043, + 0.011065391823649406, + 0.009400269016623497, + 0.012700001709163189, + 0.006746329367160797, + 0.019319122657179832, + 0.015509204007685184, + 0.025696387514472008, + 0.021199984475970268, + 0.020609261468052864, + 0.02287139929831028, + 0.022459670901298523, + 0.026765797287225723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.020622828975319862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016091711819171906 + ], + "opposite_logprob_tokens": [ + -4.141091823577881 + ], + "logprob_reward": -0.016091711819171906, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016091711819171906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09522194415330887, + 0.13236860930919647, + 0.04767324775457382, + 0.0943610668182373, + 0.024489035829901695, + 0.13973942399024963, + 0.1321859061717987, + 0.09972768276929855, + 0.061791352927684784, + 0.17244182527065277 + ], + "credit_map": [ + 0.007732315454632044, + 0.010526950471103191, + 0.009105348959565163, + 0.01388995535671711, + 0.012291011400520802, + 0.01483103446662426, + 0.016711054369807243, + 0.010134272277355194, + 0.010701521299779415, + 0.021980533376336098, + 0.014352780766785145, + 0.01685977354645729, + 0.01933286339044571, + 0.0242957454174757, + 0.014299659058451653, + 0.010545723140239716, + 0.004314227495342493, + 0.005391649901866913, + 0.004230151418596506, + 0.003600334282964468, + 0.008156190626323223, + 0.006836037151515484, + 0.007829095236957073, + 0.007315561641007662, + 0.009755797684192657, + 0.011503423564136028, + 0.013735262677073479, + 0.009892953559756279, + 0.0095127634704113, + 0.013136301189661026, + 0.012411166913807392, + 0.014413397759199142, + 0.0018567651277408004, + 0.0029811349231749773, + 0.0037969087716192007, + 0.0026430271100252867, + 0.0035331244580447674, + 0.0033301871735602617, + 0.0040007042698562145, + 0.0023471848107874393, + 0.01151316612958908, + 0.0143738379701972, + 0.014830646105110645, + 0.02149607241153717, + 0.02408182993531227, + 0.018035966902971268, + 0.0193070899695158, + 0.016100820153951645, + 0.012996011413633823, + 0.02125714160501957, + 0.017385248094797134, + 0.01709773764014244, + 0.014960149303078651, + 0.014531588181853294, + 0.020354360342025757, + 0.013603665865957737, + 0.008749572560191154, + 0.012777430936694145, + 0.029175397008657455, + 0.008337843231856823, + 0.013264447450637817, + 0.012719335034489632, + 0.007457253057509661, + 0.007246407214552164, + 0.005303489044308662, + 0.008563143201172352, + 0.007460156921297312, + 0.006474081426858902, + 0.008985891006886959, + 0.009195109829306602, + 0.009190509095788002, + 0.006618971936404705, + 0.017701847478747368, + 0.018192315474152565, + 0.016644781455397606, + 0.019803542643785477, + 0.02516579069197178, + 0.02811541222035885, + 0.023873528465628624, + 0.022944601252675056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.016091711819171906 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693719655275345 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.033693719655275345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693719655275345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1151931881904602, + 0.04522790014743805, + 0.08948887139558792, + 0.05598592758178711, + 0.03495314344763756, + 0.11973841488361359, + 0.17180223762989044, + 0.26742058992385864, + 0.0542917437851429, + 0.045897986739873886 + ], + "credit_map": [ + 0.016063455492258072, + 0.010914288461208344, + 0.00886533223092556, + 0.013204365968704224, + 0.012792409397661686, + 0.01796088181436062, + 0.021156104281544685, + 0.014236347749829292, + 0.004109520465135574, + 0.0059737879782915115, + 0.004029839765280485, + 0.004253423307090998, + 0.006520342547446489, + 0.0050698015838861465, + 0.009121442213654518, + 0.006149740424007177, + 0.007694315165281296, + 0.010507315397262573, + 0.008031428791582584, + 0.007959914393723011, + 0.012127025052905083, + 0.013543007895350456, + 0.01690848544239998, + 0.01271736714988947, + 0.005924214608967304, + 0.006396362092345953, + 0.006710754241794348, + 0.005759508349001408, + 0.006662523373961449, + 0.008167540654540062, + 0.00703042047098279, + 0.00933459959924221, + 0.0026349592953920364, + 0.0039655002765357494, + 0.004892528522759676, + 0.003949664533138275, + 0.004362777341157198, + 0.00463039381429553, + 0.0055536674335598946, + 0.004963648039847612, + 0.010868974030017853, + 0.012452035211026669, + 0.012988339178264141, + 0.01742617040872574, + 0.02192273922264576, + 0.014269482344388962, + 0.016023170202970505, + 0.013787494041025639, + 0.020023060962557793, + 0.037151455879211426, + 0.020903578028082848, + 0.011931432411074638, + 0.026414647698402405, + 0.021786218509078026, + 0.017426205798983574, + 0.016165627166628838, + 0.019763408228754997, + 0.028062250465154648, + 0.057485733181238174, + 0.014776677824556828, + 0.07886514067649841, + 0.03204265981912613, + 0.01899297907948494, + 0.01743171736598015, + 0.005327080842107534, + 0.007921356707811356, + 0.0072394805029034615, + 0.004850463941693306, + 0.006374703720211983, + 0.007072703447192907, + 0.006747887469828129, + 0.00875806249678135, + 0.0043248687870800495, + 0.007301487028598785, + 0.0052385092712938786, + 0.004442804493010044, + 0.009082578122615814, + 0.006082690786570311, + 0.004408364649862051, + 0.005016678012907505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.033693719655275345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808651491999626 + ], + "opposite_logprob_tokens": [ + -3.288086414337158 + ], + "logprob_reward": -0.03808651491999626, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808651491999626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07780131697654724, + 0.10887559503316879, + 0.0588492825627327, + 0.0755818784236908, + 0.058341387659311295, + 0.08583580702543259, + 0.15131275355815887, + 0.23882301151752472, + 0.06459712982177734, + 0.07998184114694595 + ], + "credit_map": [ + 0.010888281278312206, + 0.007301428355276585, + 0.008355037309229374, + 0.008856834843754768, + 0.010325578972697258, + 0.011117946356534958, + 0.010834614746272564, + 0.01012159138917923, + 0.008739005774259567, + 0.015669120475649834, + 0.010710231959819794, + 0.01267901062965393, + 0.0202203206717968, + 0.020044894888997078, + 0.012712236493825912, + 0.008100781589746475, + 0.0050169602036476135, + 0.009397079236805439, + 0.008205140940845013, + 0.007299961522221565, + 0.007376221939921379, + 0.00899689644575119, + 0.005684750620275736, + 0.006872272118926048, + 0.007137737702578306, + 0.010840949602425098, + 0.012449262663722038, + 0.008543920703232288, + 0.00824939925223589, + 0.01011116337031126, + 0.01113617792725563, + 0.007113263476639986, + 0.004894840996712446, + 0.006587126292288303, + 0.011131794191896915, + 0.005855086725205183, + 0.009894114919006824, + 0.008496690541505814, + 0.004605202469974756, + 0.0068765319883823395, + 0.012587216682732105, + 0.011927916668355465, + 0.015298801474273205, + 0.009419466368854046, + 0.010578216053545475, + 0.011107168160378933, + 0.008404454216361046, + 0.006512556225061417, + 0.0145783182233572, + 0.033778488636016846, + 0.018844829872250557, + 0.013092434033751488, + 0.02358282543718815, + 0.02073458582162857, + 0.011632307432591915, + 0.015068956650793552, + 0.02323618344962597, + 0.02973896823823452, + 0.03271283954381943, + 0.01395499985665083, + 0.06920429319143295, + 0.03696248307824135, + 0.01676650531589985, + 0.0162467323243618, + 0.007050428073853254, + 0.009743085131049156, + 0.008252749219536781, + 0.005595274735242128, + 0.009581170976161957, + 0.008874247781932354, + 0.006886846385896206, + 0.008613331243395805, + 0.007686561904847622, + 0.008804386481642723, + 0.00952645018696785, + 0.0065468656830489635, + 0.01425964292138815, + 0.009801667183637619, + 0.011538701131939888, + 0.011817564256489277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.03808651491999626 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, moving slowly and steadily forward, almost touching the damp ground, centered on the back of a young Chinese man, kneeling deep in a dense, forgotten forest. His back hunched, his entire body etched with tension and concentration. He wore a frayed, dark gray hooded sweatshirt, the hood nearly obscuring his head and face, and dark, mud-stained trousers. His hands, stained with black earth, carefully and steadily lowered an irregularly shaped object wrapped in crimson silk into a freshly dug pit. The red cloth, illuminated by the single, unsettling, warm yellow light of an old lantern on the ground beside him, took on an overly saturated, almost unnatural brightness, creating a harsh contrast with the surrounding darkness, littered with decaying leaves. The lantern's light cast long, dancing shadows, making the surrounding tangled roots and gnarled trunks appear hideous, while the forest beyond was a deep, impenetrable darkness. Without using any tools, he used his fingers to gently yet decisively push the loose soil back into the pit, attempting to completely conceal the mysterious red bag. The entire scene was steeped in a deep sense of secrecy and unsettling ritual, creating an eerie, dark, and deeply oppressive atmosphere.", + "seed": 8608, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_6/index_1961_seed_8608_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03492152746766806, + "vlm_logprob": -0.03492152746766806, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03492152746766806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09741505980491638 + ], + "opposite_logprob_tokens": [ + -2.472414970397949 + ], + "logprob_reward": -0.09741505980491638, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09741505980491638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3260069489479065, + 0.06918846815824509, + 0.054807331413030624, + 0.040231481194496155, + 0.04555458948016167, + 0.09712811559438705, + 0.16333483159542084, + 0.037474971264600754, + 0.08565372973680496, + 0.08061957359313965 + ], + "credit_map": [ + 0.09806180745363235, + 0.02543456479907036, + 0.033174607902765274, + 0.022831212729215622, + 0.022659100592136383, + 0.0295861829072237, + 0.02246248535811901, + 0.0717969760298729, + 0.008576986379921436, + 0.011431379243731499, + 0.009799810126423836, + 0.00709537835791707, + 0.00747899804264307, + 0.008322513662278652, + 0.009381979703903198, + 0.007101420778781176, + 0.009329404681921005, + 0.004534740466624498, + 0.011237332597374916, + 0.006245594937354326, + 0.0038654180243611336, + 0.004862569272518158, + 0.00765670882537961, + 0.007075566332787275, + 0.004361150320619345, + 0.005277843214571476, + 0.005282572004944086, + 0.003252854337915778, + 0.006115725263953209, + 0.005412383005023003, + 0.004738024435937405, + 0.005790932569652796, + 0.007711207959800959, + 0.006072780583053827, + 0.005164242349565029, + 0.0026659900322556496, + 0.004615478217601776, + 0.0034284181892871857, + 0.008527827449142933, + 0.007368645630776882, + 0.012896985746920109, + 0.013834434561431408, + 0.012602915987372398, + 0.006809968501329422, + 0.00926358811557293, + 0.00927212368696928, + 0.016513537615537643, + 0.01593455672264099, + 0.013221587054431438, + 0.015810392796993256, + 0.010552952066063881, + 0.011734542436897755, + 0.026621676981449127, + 0.022194940596818924, + 0.037065304815769196, + 0.026133431121706963, + 0.004754110239446163, + 0.005297107622027397, + 0.003022446297109127, + 0.0025078216567635536, + 0.004740268923342228, + 0.006067996378988028, + 0.005739203654229641, + 0.005346014630049467, + 0.009541724808514118, + 0.0124247120693326, + 0.008526323363184929, + 0.006491499952971935, + 0.008974512107670307, + 0.016453085467219353, + 0.01288389042019844, + 0.010357982479035854, + 0.01094076782464981, + 0.009057123214006424, + 0.007018746808171272, + 0.0072851525619626045, + 0.00853773858398199, + 0.013614829629659653, + 0.014619127847254276, + 0.00954608153551817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.09741505980491638 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822558231651783 + ], + "opposite_logprob_tokens": [ + -4.01822566986084 + ], + "logprob_reward": -0.01822558231651783, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822558231651783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32326164841651917, + 0.1457638293504715, + 0.08986067026853561, + 0.07974398136138916, + 0.03134024143218994, + 0.04026680439710617, + 0.06163929030299187, + 0.10019863396883011, + 0.04906447231769562, + 0.07886050641536713 + ], + "credit_map": [ + 0.0519942045211792, + 0.03484721481800079, + 0.04659070074558258, + 0.045627325773239136, + 0.03554987534880638, + 0.030315101146697998, + 0.03595010191202164, + 0.04238710552453995, + 0.019720125943422318, + 0.019182253628969193, + 0.024765923619270325, + 0.012917891144752502, + 0.017167501151561737, + 0.015771200880408287, + 0.019525540992617607, + 0.016713373363018036, + 0.011181565001606941, + 0.010929850861430168, + 0.014470619149506092, + 0.008501601405441761, + 0.010983739979565144, + 0.008496514521539211, + 0.012956902384757996, + 0.012339869514107704, + 0.007959810085594654, + 0.006716433446854353, + 0.010489486157894135, + 0.006183957681059837, + 0.008999333716928959, + 0.0097641097381711, + 0.01693466119468212, + 0.012696176767349243, + 0.00589064322412014, + 0.0033584467601031065, + 0.005591312423348427, + 0.003644101321697235, + 0.002928487490862608, + 0.002414627932012081, + 0.003993759863078594, + 0.0035188603214919567, + 0.006358522921800613, + 0.007100643124431372, + 0.005715888924896717, + 0.0032803562935441732, + 0.003406238043680787, + 0.00483559537678957, + 0.004512334242463112, + 0.005057225003838539, + 0.0056644282303750515, + 0.007120898459106684, + 0.004612077493220568, + 0.0032740228343755007, + 0.010557184927165508, + 0.012653288431465626, + 0.008833394385874271, + 0.008923997171223164, + 0.009514428675174713, + 0.016945330426096916, + 0.007984265685081482, + 0.005850623827427626, + 0.01708577200770378, + 0.016713857650756836, + 0.012456088326871395, + 0.013648255728185177, + 0.004443681798875332, + 0.006607071962207556, + 0.004156214185059071, + 0.003258105367422104, + 0.0047123548574745655, + 0.010291743092238903, + 0.008295441046357155, + 0.007299852557480335, + 0.009467707015573978, + 0.011006760410964489, + 0.005863475613296032, + 0.0070800562389194965, + 0.010212847962975502, + 0.014971711672842503, + 0.012749413959681988, + 0.007508531678467989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01822558231651783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093487031757832 + ], + "opposite_logprob_tokens": [ + -4.511093616485596 + ], + "logprob_reward": -0.011093487031757832, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093487031757832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2876030504703522, + 0.13872897624969482, + 0.06948351860046387, + 0.06542016565799713, + 0.03981457278132439, + 0.06139814108610153, + 0.046519648283720016, + 0.12537293136119843, + 0.08601023256778717, + 0.07964879274368286 + ], + "credit_map": [ + 0.0447709746658802, + 0.024330787360668182, + 0.03350803256034851, + 0.03968488425016403, + 0.026042692363262177, + 0.05719174072146416, + 0.0253915973007679, + 0.03668232262134552, + 0.01767006888985634, + 0.014284476637840271, + 0.02517160214483738, + 0.025005586445331573, + 0.011760453693568707, + 0.013069749809801579, + 0.017508339136838913, + 0.014258688315749168, + 0.00899021327495575, + 0.004634257405996323, + 0.011286241933703423, + 0.013056762516498566, + 0.005620511714369059, + 0.006766101811081171, + 0.0120697608217597, + 0.007059669587761164, + 0.011560835875570774, + 0.007030421402305365, + 0.01180087961256504, + 0.008804245851933956, + 0.006650019437074661, + 0.005342492368072271, + 0.008142534643411636, + 0.0060887327417731285, + 0.007268386427313089, + 0.005365136545151472, + 0.005668965633958578, + 0.006162968464195728, + 0.0029792608693242073, + 0.002512735780328512, + 0.0057195271365344524, + 0.0041375914588570595, + 0.007307001389563084, + 0.005404496565461159, + 0.006845586467534304, + 0.006679294630885124, + 0.004084018059074879, + 0.004975813440978527, + 0.01479537133127451, + 0.011306561529636383, + 0.00832811277359724, + 0.004847287200391293, + 0.006228842306882143, + 0.0032015428878366947, + 0.005488941911607981, + 0.0055654640309512615, + 0.006627139635384083, + 0.006232317537069321, + 0.01426149345934391, + 0.021045982837677002, + 0.01432009506970644, + 0.013325615786015987, + 0.011505655013024807, + 0.013786081224679947, + 0.01986604742705822, + 0.01726195216178894, + 0.0101242046803236, + 0.011052777990698814, + 0.008533366955816746, + 0.005478183273226023, + 0.008244536817073822, + 0.01943940296769142, + 0.011932803317904472, + 0.011204958893358707, + 0.009681597352027893, + 0.008540675975382328, + 0.005263960920274258, + 0.006505745928734541, + 0.01494956761598587, + 0.015202097594738007, + 0.011767292395234108, + 0.007737851236015558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011093487031757832 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18570882081985474, + 0.08782748878002167, + 0.23481836915016174, + 0.035757407546043396, + 0.03886858746409416, + 0.09174151718616486, + 0.06318105757236481, + 0.04768800362944603, + 0.12332040816545486, + 0.09108838438987732 + ], + "credit_map": [ + 0.04686599224805832, + 0.014310287311673164, + 0.01513635739684105, + 0.020954124629497528, + 0.009756979532539845, + 0.016589807346463203, + 0.014925635419785976, + 0.04716961830854416, + 0.015016691759228706, + 0.008951733820140362, + 0.012739317491650581, + 0.013078274205327034, + 0.009965506382286549, + 0.006763354875147343, + 0.011999232694506645, + 0.009313372895121574, + 0.022768432274460793, + 0.018961597234010696, + 0.031721800565719604, + 0.03295958787202835, + 0.01596398651599884, + 0.015272460877895355, + 0.06433442234992981, + 0.03283606469631195, + 0.004449914209544659, + 0.003939891699701548, + 0.0037005655467510223, + 0.0036814548075199127, + 0.0029792708810418844, + 0.0031116949394345284, + 0.007983426563441753, + 0.005911185406148434, + 0.0077279978431761265, + 0.005138811655342579, + 0.008214104920625687, + 0.004529543686658144, + 0.0022850679233670235, + 0.0022163044195622206, + 0.005031281150877476, + 0.003725470742210746, + 0.015863735228776932, + 0.015665177255868912, + 0.012933664955198765, + 0.009887141175568104, + 0.006575897801667452, + 0.007332736626267433, + 0.012034304440021515, + 0.011448848061263561, + 0.006185465957969427, + 0.007137668784707785, + 0.006193610839545727, + 0.004952232353389263, + 0.009850301779806614, + 0.01083118561655283, + 0.00964989885687828, + 0.008380685932934284, + 0.006467160303145647, + 0.007513726130127907, + 0.006764630321413279, + 0.005088947247713804, + 0.004523059818893671, + 0.004887006711214781, + 0.006809775717556477, + 0.005633690860122442, + 0.016721412539482117, + 0.012318974360823631, + 0.010856276378035545, + 0.00793171301484108, + 0.01307184249162674, + 0.026047397404909134, + 0.01954670436680317, + 0.016826078295707703, + 0.015665728598833084, + 0.014626452699303627, + 0.008346924558281898, + 0.006629570387303829, + 0.007712706457823515, + 0.012909989804029465, + 0.0170796699821949, + 0.008117333985865116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.00978010892868042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the back of a young Chinese man kneeling in a dense forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a frayed dark gray hooded sweatshirt with the hood obscuring his face and dark mud-stained trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man lowering an irregularly shaped object wrapped in crimson silk into a freshly dug pit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm yellow lantern on the ground casting long dancing shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily forward, almost touching the damp ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809339925646782 + ], + "opposite_logprob_tokens": [ + -3.288093328475952 + ], + "logprob_reward": -0.03809339925646782, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809339925646782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14888730645179749, + 0.05358143895864487, + 0.06886979192495346, + 0.0754862129688263, + 0.038793619722127914, + 0.06313306838274002, + 0.04587584733963013, + 0.1587340086698532, + 0.2210835963487625, + 0.12555508315563202 + ], + "credit_map": [ + 0.017078233882784843, + 0.013766976073384285, + 0.022723183035850525, + 0.017596730962395668, + 0.014392378740012646, + 0.025763852521777153, + 0.013108070008456707, + 0.024457881227135658, + 0.0060743228532373905, + 0.005877825431525707, + 0.008043888956308365, + 0.007934699766337872, + 0.00521738501265645, + 0.0057286350056529045, + 0.0066975136287510395, + 0.008007169701159, + 0.011232297867536545, + 0.007382709067314863, + 0.010175305418670177, + 0.009973815642297268, + 0.005509045906364918, + 0.005250126123428345, + 0.009378970600664616, + 0.009967517107725143, + 0.009044980630278587, + 0.007405241951346397, + 0.008208015002310276, + 0.006077298894524574, + 0.008077673614025116, + 0.006768458057194948, + 0.017244329676032066, + 0.01266021654009819, + 0.0051581235602498055, + 0.0046986606903374195, + 0.004615742713212967, + 0.004269243683665991, + 0.003597068600356579, + 0.004008821211755276, + 0.006153928115963936, + 0.006292030215263367, + 0.008750340901315212, + 0.007400339003652334, + 0.009952900931239128, + 0.008158029988408089, + 0.006284070201218128, + 0.005726509261876345, + 0.00972119439393282, + 0.007139681838452816, + 0.005441815592348576, + 0.004478823393583298, + 0.006496013142168522, + 0.004607858136296272, + 0.00437872763723135, + 0.006918396335095167, + 0.007501227781176567, + 0.006052985787391663, + 0.016261836513876915, + 0.022431375458836555, + 0.01307271234691143, + 0.011316999793052673, + 0.022189687937498093, + 0.02891680598258972, + 0.0215463750064373, + 0.022998223081231117, + 0.021116700023412704, + 0.025627927854657173, + 0.02112426981329918, + 0.01700379140675068, + 0.026151448488235474, + 0.04702774062752724, + 0.03588930517435074, + 0.027142401784658432, + 0.01396964956074953, + 0.013974414207041264, + 0.00979258306324482, + 0.014511985704302788, + 0.013532665558159351, + 0.02392905205488205, + 0.018920881673693657, + 0.016923846676945686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.03809339925646782 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_tensor([3018, 3018])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_tensor([3018, 3018])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e18396a978c2d0eb1448fbd8b57db700b4457226 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_tensor([3018, 3018])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, handheld tracking shot follows two men as they engage in a frantic chase through a narrow, damp alley. Shot entirely in high-contrast black-and-white, the leading man, wearing a thin, rain-soaked windbreaker, his damp black hair plastered to his forehead, glances back in panic, his face a panting, fearful expression. The other, more burly man in a heavy, dark coat, pursues them. His steps are heavy and determined, his face stern, his eyes gleaming with unyielding determination. The cobblestones beneath their feet are stained with filthy water, which splatters violently with each heavy step, creating splashes of silvery water that shimmer in the dim light, like shattered glass. The mottled brick walls on both sides squeeze inward, and the peeling graffiti and damp moss on the walls cast long and distorted shadows under the pale light of a single lamp deep in the alley, intensifying the oppressive sense of space and the urgency of the chase, creating a suffocating and dynamic film noir atmosphere.", + "seed": 25249, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_3018_seed_25249_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03570027882233262, + "vlm_logprob": -0.03570027882233262, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.03570027882233262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04374302551150322 + ], + "opposite_logprob_tokens": [ + -3.918743133544922 + ], + "logprob_reward": -0.04374302551150322, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.04374302551150322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11224792152643204, + 0.08048725128173828, + 0.046084146946668625, + 0.05978832021355629, + 0.10100565105676651, + 0.19885924458503723, + 0.0903332307934761, + 0.06049484759569168, + 0.09944743663072586, + 0.15125198662281036 + ], + "credit_map": [ + 0.016202479600906372, + 0.013411919586360455, + 0.016940435394644737, + 0.025316819548606873, + 0.011000346392393112, + 0.007897799834609032, + 0.011101041920483112, + 0.01037706807255745, + 0.010873797349631786, + 0.007179801817983389, + 0.013154584914445877, + 0.010430301539599895, + 0.00642582168802619, + 0.00921052135527134, + 0.012901601381599903, + 0.010310818441212177, + 0.005148665979504585, + 0.0064726597629487514, + 0.006269564386457205, + 0.0031365645118057728, + 0.004480397794395685, + 0.007452955935150385, + 0.006997697986662388, + 0.006125637795776129, + 0.0077043394558131695, + 0.005217986647039652, + 0.006994686555117369, + 0.007757196202874184, + 0.0047403364442288876, + 0.005691913887858391, + 0.011555925942957401, + 0.010125932283699512, + 0.008672641590237617, + 0.014065207913517952, + 0.01414946187287569, + 0.015230507589876652, + 0.012914130464196205, + 0.013113025575876236, + 0.013254155404865742, + 0.009606519713997841, + 0.022394057363271713, + 0.02537234127521515, + 0.022696809843182564, + 0.019130611792206764, + 0.015516169369220734, + 0.03834341838955879, + 0.022752875462174416, + 0.032652948051691055, + 0.008567063137888908, + 0.006801823154091835, + 0.009115085005760193, + 0.009317384101450443, + 0.009786238893866539, + 0.013701549731194973, + 0.01384346466511488, + 0.019200613722205162, + 0.004616637248545885, + 0.0038510560989379883, + 0.003259951714426279, + 0.006993803195655346, + 0.010632555931806564, + 0.007795270532369614, + 0.012619590386748314, + 0.010725980624556541, + 0.005918582901358604, + 0.00816519744694233, + 0.005798956844955683, + 0.006778039503842592, + 0.021289518103003502, + 0.01431785523891449, + 0.01593443937599659, + 0.021244842559099197, + 0.00995545368641615, + 0.016582312062382698, + 0.012126722373068333, + 0.02387523278594017, + 0.02208784595131874, + 0.015665456652641296, + 0.03029651939868927, + 0.020662423223257065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.04374302551150322 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299825459718704 + ], + "opposite_logprob_tokens": [ + -2.2379982471466064 + ], + "logprob_reward": -0.11299825459718704, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299825459718704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09205909073352814, + 0.11945566534996033, + 0.1070343479514122, + 0.08190908282995224, + 0.08316176384687424, + 0.11545239388942719, + 0.09478870779275894, + 0.0641845166683197, + 0.06482376158237457, + 0.17713062465190887 + ], + "credit_map": [ + 0.01461793202906847, + 0.011579953134059906, + 0.015038324519991875, + 0.015805568546056747, + 0.010475668124854565, + 0.00650959275662899, + 0.00867609865963459, + 0.00935595203191042, + 0.015554850921034813, + 0.013389640487730503, + 0.015494597144424915, + 0.020433178171515465, + 0.00994704756885767, + 0.014279662631452084, + 0.018136901780962944, + 0.01221979409456253, + 0.010036255232989788, + 0.015403720550239086, + 0.019071975722908974, + 0.018532827496528625, + 0.008079081773757935, + 0.012281876057386398, + 0.012183638289570808, + 0.011444978415966034, + 0.011756621301174164, + 0.006644161883741617, + 0.009233903139829636, + 0.008710231631994247, + 0.006518985610455275, + 0.011540405452251434, + 0.01443743146955967, + 0.013067352585494518, + 0.008078393526375294, + 0.011713938787579536, + 0.010882113128900528, + 0.01117493025958538, + 0.009947097860276699, + 0.011097594164311886, + 0.012196624651551247, + 0.008071073330938816, + 0.011266514658927917, + 0.020419199019670486, + 0.011640406213700771, + 0.01368208322674036, + 0.010409612208604813, + 0.013289136812090874, + 0.014361447654664516, + 0.02038399875164032, + 0.01055081095546484, + 0.00548606738448143, + 0.006197378970682621, + 0.008465331979095936, + 0.012897825799882412, + 0.011187153868377209, + 0.019046280533075333, + 0.020957864820957184, + 0.00515548512339592, + 0.004298388957977295, + 0.0039749848656356335, + 0.0101417051628232, + 0.009625373408198357, + 0.00877411849796772, + 0.011602580547332764, + 0.010611881501972675, + 0.008421029895544052, + 0.008191696368157864, + 0.0047676293179392815, + 0.005738700740039349, + 0.006831392180174589, + 0.006164045073091984, + 0.011700000613927841, + 0.013009272515773773, + 0.011352059431374073, + 0.017063841223716736, + 0.012416660785675049, + 0.024289648979902267, + 0.03989697992801666, + 0.019540823996067047, + 0.031056156381964684, + 0.02151445671916008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.11299825459718704 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254053320735693 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.01254053320735693, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254053320735693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13622909784317017, + 0.1379312425851822, + 0.0489533394575119, + 0.04546637460589409, + 0.03377385810017586, + 0.10426194965839386, + 0.10622269660234451, + 0.11736408621072769, + 0.15008629858493805, + 0.11971105635166168 + ], + "credit_map": [ + 0.01713886857032776, + 0.021045057103037834, + 0.018976641818881035, + 0.01465749554336071, + 0.014101574197411537, + 0.014896761626005173, + 0.019025348126888275, + 0.016387348994612694, + 0.015577297657728195, + 0.016635818406939507, + 0.02104673720896244, + 0.018492450937628746, + 0.014558791182935238, + 0.015271247364580631, + 0.021675029769539833, + 0.014673866331577301, + 0.0040704659186303616, + 0.008178722113370895, + 0.006411437876522541, + 0.005675952415913343, + 0.0042721908539533615, + 0.007863613776862621, + 0.005021908786147833, + 0.007459047716110945, + 0.004947517067193985, + 0.00470787612721324, + 0.005680595058947802, + 0.005201911088079214, + 0.003751968964934349, + 0.003962426912039518, + 0.008687845431268215, + 0.008526229299604893, + 0.003372336970642209, + 0.005043522920459509, + 0.004276917316019535, + 0.0047836219891905785, + 0.004311744123697281, + 0.004831463098526001, + 0.003784207161515951, + 0.0033700414933264256, + 0.009187472052872181, + 0.014826355502009392, + 0.010750462301075459, + 0.015104368329048157, + 0.011947840452194214, + 0.01222931407392025, + 0.012834739871323109, + 0.017381390556693077, + 0.013169201090931892, + 0.006712674628943205, + 0.00578754348680377, + 0.00987155083566904, + 0.010587279684841633, + 0.014340925961732864, + 0.02584844082593918, + 0.019905073568224907, + 0.013783503323793411, + 0.008195373229682446, + 0.013487467542290688, + 0.018456118181347847, + 0.02208472415804863, + 0.01142803207039833, + 0.01816542074084282, + 0.011763439513742924, + 0.015337674878537655, + 0.012097496539354324, + 0.012478629127144814, + 0.010904200375080109, + 0.018031282350420952, + 0.025189287960529327, + 0.025483867153525352, + 0.0305638387799263, + 0.007202370557934046, + 0.014152002520859241, + 0.008336277678608894, + 0.012511729262769222, + 0.026629174128174782, + 0.014187069609761238, + 0.02294383943080902, + 0.0137485908344388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.01254053320735693 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002485878299921751 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.002485878299921751, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002485878299921751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08153437077999115, + 0.11233269423246384, + 0.17248624563217163, + 0.06159045919775963, + 0.2567155063152313, + 0.11226151138544083, + 0.0441320464015007, + 0.047558050602674484, + 0.053141236305236816, + 0.058247923851013184 + ], + "credit_map": [ + 0.012613097205758095, + 0.009539713151752949, + 0.01091704424470663, + 0.010233776643872261, + 0.012410469353199005, + 0.007002786733210087, + 0.010867306031286716, + 0.007950170896947384, + 0.01636549085378647, + 0.012597812339663506, + 0.016278382390737534, + 0.0155479172244668, + 0.009482171386480331, + 0.014371310360729694, + 0.015292979776859283, + 0.01239662617444992, + 0.01536262221634388, + 0.02645435743033886, + 0.027506334707140923, + 0.01819748431444168, + 0.01244658138602972, + 0.029804276302456856, + 0.01716153509914875, + 0.025553032755851746, + 0.007660702336579561, + 0.00487107876688242, + 0.0066811381839215755, + 0.00470188120380044, + 0.004182199947535992, + 0.007266803178936243, + 0.014941557310521603, + 0.011285098269581795, + 0.02770295925438404, + 0.026162998750805855, + 0.035320062190294266, + 0.025914423167705536, + 0.04836837574839592, + 0.03531999513506889, + 0.026307454332709312, + 0.031619202345609665, + 0.011122837662696838, + 0.018029997125267982, + 0.01286966260522604, + 0.01422283798456192, + 0.008852467872202396, + 0.018878400325775146, + 0.012217988260090351, + 0.01606730744242668, + 0.004038461949676275, + 0.002892240881919861, + 0.0034462958574295044, + 0.004772099666297436, + 0.004540582653135061, + 0.0059028929099440575, + 0.007558263372629881, + 0.010981203988194466, + 0.006937387399375439, + 0.0037244074046611786, + 0.006054962053894997, + 0.006254798732697964, + 0.007293094415217638, + 0.004570582881569862, + 0.007302514743059874, + 0.005420297849923372, + 0.005770946387201548, + 0.007295107468962669, + 0.004158461932092905, + 0.00613319780677557, + 0.00944107212126255, + 0.0066536483354866505, + 0.0073338039219379425, + 0.006354993674904108, + 0.004468623083084822, + 0.007828322239220142, + 0.004400726407766342, + 0.0065097082406282425, + 0.010963010601699352, + 0.005965229589492083, + 0.01089729368686676, + 0.00721500301733613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.002485878299921751 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067337024956941605 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067337024956941605, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067337024956941605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054073967039585114, + 0.504675567150116, + 0.04164588078856468, + 0.02564210630953312, + 0.0575309656560421, + 0.0718376487493515, + 0.09363177418708801, + 0.048678699880838394, + 0.0655372142791748, + 0.03674619644880295 + ], + "credit_map": [ + 0.008161379024386406, + 0.004879770800471306, + 0.008342061191797256, + 0.008277538232505322, + 0.006369458045810461, + 0.004745620768517256, + 0.007072955835610628, + 0.006225185468792915, + 0.05950620770454407, + 0.053321197628974915, + 0.06669174134731293, + 0.11795118451118469, + 0.02768097259104252, + 0.04710225388407707, + 0.08222204446792603, + 0.050199974328279495, + 0.004694822710007429, + 0.007219755556434393, + 0.005232145078480244, + 0.003390910802409053, + 0.003865801729261875, + 0.007704927586019039, + 0.004734287038445473, + 0.004803232382982969, + 0.0035205110907554626, + 0.002056246856227517, + 0.0034855639096349478, + 0.0022298009134829044, + 0.0021350253373384476, + 0.003302754368633032, + 0.004813595674932003, + 0.00409860722720623, + 0.008610064163804054, + 0.006021310109645128, + 0.009209790267050266, + 0.007941539399325848, + 0.005642930045723915, + 0.007376431953161955, + 0.006781334523111582, + 0.005947566591203213, + 0.00889530498534441, + 0.008843451738357544, + 0.008634262718260288, + 0.007714363280683756, + 0.00520881125703454, + 0.012822757475078106, + 0.00831694807857275, + 0.011401747353374958, + 0.010107017122209072, + 0.004905030131340027, + 0.006758180912584066, + 0.009791573509573936, + 0.00797694455832243, + 0.013473708182573318, + 0.016868658363819122, + 0.023750653490424156, + 0.003842433914542198, + 0.0038635435048490763, + 0.003358169225975871, + 0.006388684269040823, + 0.008580164983868599, + 0.006499538663774729, + 0.00868282001465559, + 0.007463343441486359, + 0.008851672522723675, + 0.00840216875076294, + 0.005922145675867796, + 0.004898215178400278, + 0.012958528473973274, + 0.009306072257459164, + 0.007599348668009043, + 0.007599063217639923, + 0.0027715624310076237, + 0.003454765537753701, + 0.0032464046962559223, + 0.005634665954858065, + 0.0062616015784442425, + 0.003616597969084978, + 0.006948653608560562, + 0.004811943043023348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.0067337024956941605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld tracking shot follows two men as they engage in a frantic chase through a narrow, damp alley. Shot entirely in high-contrast black-and-white, the leading man, wearing a thin, rain-soaked windbreaker, his damp black hair plastered to his forehead, glances back in panic, his face a panting, fearful expression. The other, more burly man in a heavy, dark coat, pursues them. His steps are heavy and determined, his face stern, his eyes gleaming with unyielding determination. The cobblestones beneath their feet are stained with filthy water, which splatters violently with each heavy step, creating splashes of silvery water that shimmer in the dim light, like shattered glass. The mottled brick walls on both sides squeeze inward, and the peeling graffiti and damp moss on the walls cast long and distorted shadows under the pale light of a single lamp deep in the alley, intensifying the oppressive sense of space and the urgency of the chase, creating a suffocating and dynamic film noir atmosphere.", + "seed": 25250, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_3018_seed_25250_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027941651176661253, + "vlm_logprob": -0.027941651176661253, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.027941651176661253, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022298678755760193 + ], + "opposite_logprob_tokens": [ + -4.397298812866211 + ], + "logprob_reward": -0.022298678755760193, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.022298678755760193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19171977043151855, + 0.06339795142412186, + 0.2387065887451172, + 0.041546765714883804, + 0.05232005566358566, + 0.12854357063770294, + 0.059706687927246094, + 0.025227457284927368, + 0.09441295266151428, + 0.10441829264163971 + ], + "credit_map": [ + 0.0372469387948513, + 0.017466427758336067, + 0.028512483462691307, + 0.028239067643880844, + 0.01295015960931778, + 0.02699928544461727, + 0.01838766597211361, + 0.02191772125661373, + 0.009418780915439129, + 0.005310126580297947, + 0.007638332433998585, + 0.009202643297612667, + 0.0068089524284005165, + 0.00877317227423191, + 0.006282029207795858, + 0.009963905438780785, + 0.0651978850364685, + 0.013362755067646503, + 0.07361211627721786, + 0.011005925014615059, + 0.008645533584058285, + 0.041390713304281235, + 0.006420555990189314, + 0.019071079790592194, + 0.007613851688802242, + 0.004574993159621954, + 0.008769729174673557, + 0.0037110683042556047, + 0.004265735857188702, + 0.004059862345457077, + 0.004598709288984537, + 0.00395280821248889, + 0.007714008446782827, + 0.006901755463331938, + 0.007483499590307474, + 0.00509789539501071, + 0.0035910122096538544, + 0.00766785629093647, + 0.005009239073842764, + 0.008854782208800316, + 0.03137664124369621, + 0.008183466270565987, + 0.03580799698829651, + 0.006262618582695723, + 0.00921298936009407, + 0.014405985362827778, + 0.009130151011049747, + 0.01416370365768671, + 0.009593399241566658, + 0.006501100491732359, + 0.0062832883559167385, + 0.004783858545124531, + 0.006045896094292402, + 0.010388377122581005, + 0.006783187389373779, + 0.009327574633061886, + 0.002177488524466753, + 0.0022251587361097336, + 0.002893022960051894, + 0.0033959210850298405, + 0.003462384454905987, + 0.003946506883949041, + 0.0038800055626779795, + 0.0032469655852764845, + 0.010477505624294281, + 0.005008295644074678, + 0.010512040928006172, + 0.007679309695959091, + 0.010396783240139484, + 0.01839553751051426, + 0.0107010118663311, + 0.021242450922727585, + 0.016237063333392143, + 0.008694875985383987, + 0.016014397144317627, + 0.011471959762275219, + 0.013090794906020164, + 0.014063789509236813, + 0.014218778349459171, + 0.010626628994941711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.022298678755760193 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908265084028244 + ], + "opposite_logprob_tokens": [ + -2.464082717895508 + ], + "logprob_reward": -0.08908265084028244, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908265084028244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17889545857906342, + 0.08076527714729309, + 0.31571531295776367, + 0.0843038484454155, + 0.04004674032330513, + 0.019086727872490883, + 0.09687796980142593, + 0.03916547819972038, + 0.048234473913908005, + 0.09690868854522705 + ], + "credit_map": [ + 0.027237044647336006, + 0.024078156799077988, + 0.025427507236599922, + 0.02553296647965908, + 0.01565309427678585, + 0.02076357789337635, + 0.019707955420017242, + 0.020495174452662468, + 0.013293925672769547, + 0.006449041422456503, + 0.010720381513237953, + 0.011439669877290726, + 0.008461738005280495, + 0.009136090986430645, + 0.009008502587676048, + 0.012255929410457611, + 0.09497269243001938, + 0.015216383151710033, + 0.1065942719578743, + 0.013028651475906372, + 0.010024928487837315, + 0.04748808592557907, + 0.0069518196396529675, + 0.021438511088490486, + 0.012901038862764835, + 0.010122655890882015, + 0.01568169891834259, + 0.011463750153779984, + 0.006289360113441944, + 0.008403289131820202, + 0.008747088722884655, + 0.010694969445466995, + 0.005627947393804789, + 0.004340042360126972, + 0.004951911978423595, + 0.0034700168762356043, + 0.0035169527400285006, + 0.005238410551100969, + 0.0041533526964485645, + 0.008748111315071583, + 0.0027754702605307102, + 0.001847191946581006, + 0.002432876266539097, + 0.0012345141731202602, + 0.0019782166928052902, + 0.003625531680881977, + 0.001933125895448029, + 0.003259801771491766, + 0.008678753860294819, + 0.010904017835855484, + 0.010669717565178871, + 0.010805857367813587, + 0.01548436563462019, + 0.011923869140446186, + 0.012662808410823345, + 0.015748582780361176, + 0.004098637029528618, + 0.005043760407716036, + 0.005254367832094431, + 0.004130118526518345, + 0.004833944607526064, + 0.006204833276569843, + 0.005275134928524494, + 0.004324686713516712, + 0.005541509483009577, + 0.002854280872270465, + 0.004332071170210838, + 0.006297885440289974, + 0.003082756418734789, + 0.00555749936029315, + 0.006665850058197975, + 0.013902619481086731, + 0.014973977580666542, + 0.011294218711555004, + 0.015917692333459854, + 0.011171957477927208, + 0.011578358709812164, + 0.010191592387855053, + 0.010657040402293205, + 0.011123857460916042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.08908265084028244 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606590487062931 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.01606590487062931, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606590487062931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12283392995595932, + 0.10905007272958755, + 0.20376138389110565, + 0.0756826400756836, + 0.08867944777011871, + 0.13628725707530975, + 0.07222848385572433, + 0.038428597152233124, + 0.07456724345684052, + 0.07848094403743744 + ], + "credit_map": [ + 0.020047586411237717, + 0.011705943383276463, + 0.018340852111577988, + 0.02170035056769848, + 0.00846243929117918, + 0.0140731455758214, + 0.011748472228646278, + 0.01675514690577984, + 0.01649930141866207, + 0.012208941392600536, + 0.015263975597918034, + 0.013224081136286259, + 0.016084695234894753, + 0.009489061310887337, + 0.014332265593111515, + 0.011947751976549625, + 0.05363792926073074, + 0.013845081441104412, + 0.0544673353433609, + 0.013265419751405716, + 0.017864178866147995, + 0.014351733028888702, + 0.02028759755194187, + 0.016042109578847885, + 0.01677076146006584, + 0.00630548782646656, + 0.017372190952301025, + 0.007952910847961903, + 0.0052289157174527645, + 0.0068416902795434, + 0.006146854721009731, + 0.009063823148608208, + 0.012370767071843147, + 0.00953542347997427, + 0.013104818761348724, + 0.010934889316558838, + 0.0068067223764956, + 0.008909400552511215, + 0.009885640814900398, + 0.017131786793470383, + 0.019965602084994316, + 0.01008943747729063, + 0.017908994108438492, + 0.007868167012929916, + 0.024822400882840157, + 0.01939358003437519, + 0.017051080241799355, + 0.019187988713383675, + 0.008759386837482452, + 0.006329289171844721, + 0.008949452079832554, + 0.005325349047780037, + 0.008856652304530144, + 0.014955124817788601, + 0.008144877851009369, + 0.010908348485827446, + 0.0032190599013119936, + 0.0037697958759963512, + 0.006937697529792786, + 0.004165680147707462, + 0.005299828480929136, + 0.005070617888122797, + 0.00488653127104044, + 0.005079385358840227, + 0.009728857316076756, + 0.0038806749507784843, + 0.005248999688774347, + 0.004378203768283129, + 0.006319843232631683, + 0.017223451286554337, + 0.010929614305496216, + 0.01685759425163269, + 0.014206571504473686, + 0.007446253206580877, + 0.010283922776579857, + 0.008672934956848621, + 0.007650106213986874, + 0.009465532377362251, + 0.012035582214593887, + 0.00872004497796297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.01606590487062931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002484808210283518 + ], + "opposite_logprob_tokens": [ + -6.0024847984313965 + ], + "logprob_reward": -0.002484808210283518, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002484808210283518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2091195285320282, + 0.0581299252808094, + 0.3303333520889282, + 0.028254559263586998, + 0.048353563994169235, + 0.03308887779712677, + 0.08669930696487427, + 0.08527255803346634, + 0.07699621468782425, + 0.04375213012099266 + ], + "credit_map": [ + 0.03936603292822838, + 0.018916642293334007, + 0.03187243640422821, + 0.03162703663110733, + 0.014258435927331448, + 0.029086239635944366, + 0.017397822812199593, + 0.026594888418912888, + 0.008326636627316475, + 0.005973178893327713, + 0.006323433481156826, + 0.006383813451975584, + 0.0078814085572958, + 0.006737262476235628, + 0.006914487574249506, + 0.009589697234332561, + 0.10004434734582901, + 0.016719430685043335, + 0.11084073036909103, + 0.014196879230439663, + 0.007679125294089317, + 0.052545323967933655, + 0.005135104991495609, + 0.023172400891780853, + 0.0038176854141056538, + 0.004051378928124905, + 0.004371473100036383, + 0.0036090665962547064, + 0.0018361210823059082, + 0.0038350874092429876, + 0.001954400911927223, + 0.0047793458215892315, + 0.009930227883160114, + 0.00621237326413393, + 0.006854091305285692, + 0.0047973161563277245, + 0.0028886261861771345, + 0.00492520397529006, + 0.004169216845184565, + 0.008576510474085808, + 0.006587638054043055, + 0.003006678307428956, + 0.005065786652266979, + 0.002081050304695964, + 0.005322802346199751, + 0.004015784244984388, + 0.003054271684959531, + 0.003954869229346514, + 0.01161771547049284, + 0.005549841560423374, + 0.012428185902535915, + 0.004710864275693893, + 0.008971343748271465, + 0.019227396696805954, + 0.010121139697730541, + 0.014072822406888008, + 0.006901036947965622, + 0.007593555375933647, + 0.009302509017288685, + 0.006649836897850037, + 0.012041773647069931, + 0.018684903159737587, + 0.011694717220962048, + 0.012404220178723335, + 0.009875679388642311, + 0.005219197832047939, + 0.007063327357172966, + 0.007226243149489164, + 0.007557545322924852, + 0.020199868828058243, + 0.008300688117742538, + 0.011553664691746235, + 0.009319016709923744, + 0.0036734910681843758, + 0.0056172143667936325, + 0.00646680872887373, + 0.004411680158227682, + 0.005099829286336899, + 0.004586754832416773, + 0.004577333573251963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.002484808210283518 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776213206350803 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776213206350803, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776213206350803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10823776572942734, + 0.038415007293224335, + 0.47105979919433594, + 0.027403729036450386, + 0.03469507396221161, + 0.04761221259832382, + 0.1556348353624344, + 0.04130944609642029, + 0.044218406081199646, + 0.0314137227833271 + ], + "credit_map": [ + 0.022420592606067657, + 0.010049555450677872, + 0.018678536638617516, + 0.015444322489202023, + 0.007241999264806509, + 0.014689318835735321, + 0.007090941537171602, + 0.012622486799955368, + 0.007397232111543417, + 0.003849188331514597, + 0.005864919163286686, + 0.004418934229761362, + 0.005671868100762367, + 0.0038025665562599897, + 0.003218561876565218, + 0.004191739019006491, + 0.14301884174346924, + 0.02141946367919445, + 0.16155248880386353, + 0.022045215591788292, + 0.010590066201984882, + 0.07438168674707413, + 0.006464771926403046, + 0.03158721700310707, + 0.00487509835511446, + 0.0035033205058425665, + 0.004410249646753073, + 0.003836327465251088, + 0.0018432019278407097, + 0.0032296946737915277, + 0.002531246980652213, + 0.0031745904125273228, + 0.005674582906067371, + 0.0034513671416789293, + 0.00484098494052887, + 0.0044944994151592255, + 0.0025960931088775396, + 0.004636284429579973, + 0.002952620619907975, + 0.006048643030226231, + 0.010022406466305256, + 0.004121442325413227, + 0.01184637937694788, + 0.002683563856408, + 0.004251402802765369, + 0.005021785851567984, + 0.003825845429673791, + 0.005839387886226177, + 0.017010850831866264, + 0.01895427145063877, + 0.01678341068327427, + 0.016361474990844727, + 0.023954004049301147, + 0.02007940597832203, + 0.01976456493139267, + 0.02272685058414936, + 0.004401806741952896, + 0.003191678784787655, + 0.005244049709290266, + 0.005219263955950737, + 0.005039536859840155, + 0.007099634036421776, + 0.005413318984210491, + 0.005700157955288887, + 0.004639087710529566, + 0.0029917971696704626, + 0.0041036359034478664, + 0.0036514238454401493, + 0.0038492546882480383, + 0.011950099840760231, + 0.004614592995494604, + 0.00841851532459259, + 0.005175626836717129, + 0.0024178079329431057, + 0.0036531060468405485, + 0.0035768577363342047, + 0.003204462118446827, + 0.004126389976590872, + 0.004682071041315794, + 0.00457739969715476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.009776213206350803 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld tracking shot follows two men as they engage in a frantic chase through a narrow, damp alley. Shot entirely in high-contrast black-and-white, the leading man, wearing a thin, rain-soaked windbreaker, his damp black hair plastered to his forehead, glances back in panic, his face a panting, fearful expression. The other, more burly man in a heavy, dark coat, pursues them. His steps are heavy and determined, his face stern, his eyes gleaming with unyielding determination. The cobblestones beneath their feet are stained with filthy water, which splatters violently with each heavy step, creating splashes of silvery water that shimmer in the dim light, like shattered glass. The mottled brick walls on both sides squeeze inward, and the peeling graffiti and damp moss on the walls cast long and distorted shadows under the pale light of a single lamp deep in the alley, intensifying the oppressive sense of space and the urgency of the chase, creating a suffocating and dynamic film noir atmosphere.", + "seed": 25251, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_3018_seed_25251_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05305800535716117, + "vlm_logprob": -0.05305800535716117, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.05305800535716117, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031870726495981216 + ], + "opposite_logprob_tokens": [ + -4.7818708419799805 + ], + "logprob_reward": -0.031870726495981216, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.031870726495981216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054766375571489334, + 0.05012000352144241, + 0.06799016147851944, + 0.0705108717083931, + 0.07452443987131119, + 0.21673032641410828, + 0.0946272537112236, + 0.20707376301288605, + 0.03170052543282509, + 0.1319563239812851 + ], + "credit_map": [ + 0.008872970938682556, + 0.006675361189991236, + 0.008017283864319324, + 0.00900173932313919, + 0.003762510372325778, + 0.005582510959357023, + 0.007065144367516041, + 0.005788852926343679, + 0.009815167635679245, + 0.0029789551626890898, + 0.00815602857619524, + 0.002843651920557022, + 0.0028830061201006174, + 0.005333041772246361, + 0.004754265770316124, + 0.013355882838368416, + 0.005509314592927694, + 0.0036895950324833393, + 0.009542574174702168, + 0.006362908985465765, + 0.005694813095033169, + 0.00825002696365118, + 0.015023619867861271, + 0.013917310163378716, + 0.007538052275776863, + 0.005429431330412626, + 0.01212079543620348, + 0.005373797379434109, + 0.007825617678463459, + 0.012217693030834198, + 0.01057359017431736, + 0.009431899525225163, + 0.004941376391798258, + 0.006382818799465895, + 0.00824100524187088, + 0.005920979660004377, + 0.017325319349765778, + 0.00967448391020298, + 0.010885464027523994, + 0.01115298829972744, + 0.019777216017246246, + 0.022075286135077477, + 0.02312762290239334, + 0.01635613664984703, + 0.012160148471593857, + 0.02195270173251629, + 0.044625502079725266, + 0.05665571987628937, + 0.005202199332416058, + 0.00401195976883173, + 0.004809592850506306, + 0.0044846311211586, + 0.003910652361810207, + 0.023639312013983727, + 0.021025478839874268, + 0.02754342369735241, + 0.02296837419271469, + 0.008443470112979412, + 0.035083968192338943, + 0.012408672831952572, + 0.0176419485360384, + 0.018896447494626045, + 0.06296069175004959, + 0.028670180588960648, + 0.0026574735529720783, + 0.005170510616153479, + 0.003031226107850671, + 0.003546169027686119, + 0.004133255686610937, + 0.00464029423892498, + 0.004196988884359598, + 0.0043246070854365826, + 0.01063593290746212, + 0.014489918015897274, + 0.009296998381614685, + 0.009918212890625, + 0.027655398473143578, + 0.014333127997815609, + 0.023128366097807884, + 0.022498372942209244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.031870726495981216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016259729862213 + ], + "opposite_logprob_tokens": [ + -1.701625943183899 + ], + "logprob_reward": -0.2016259729862213, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016259729862213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10629214346408844, + 0.07493744045495987, + 0.09135740995407104, + 0.16792698204517365, + 0.048198383301496506, + 0.1233443096280098, + 0.05573953688144684, + 0.13690343499183655, + 0.09122209995985031, + 0.10407833009958267 + ], + "credit_map": [ + 0.015190884470939636, + 0.01221874263137579, + 0.01607903279364109, + 0.010716140270233154, + 0.00873718224465847, + 0.01168420072644949, + 0.012128445319831371, + 0.01953752152621746, + 0.008068852126598358, + 0.009747307747602463, + 0.011037269607186317, + 0.00938208494335413, + 0.007974029518663883, + 0.00809038057923317, + 0.01244727149605751, + 0.008190241642296314, + 0.010644602589309216, + 0.0053238230757415295, + 0.01301625557243824, + 0.012875007465481758, + 0.007274644915014505, + 0.012366550043225288, + 0.01951233670115471, + 0.010344184935092926, + 0.020883435383439064, + 0.009964942932128906, + 0.029913313686847687, + 0.01147438120096922, + 0.01746535487473011, + 0.02901764214038849, + 0.02520258165895939, + 0.024005334824323654, + 0.004767886362969875, + 0.003021595533937216, + 0.008846129290759563, + 0.0034167661797255278, + 0.009505554102361202, + 0.004543861839920282, + 0.006399158388376236, + 0.00769742950797081, + 0.009856907650828362, + 0.007279818877577782, + 0.007829281501471996, + 0.008079886436462402, + 0.010955125093460083, + 0.013757812790572643, + 0.02957252599298954, + 0.03601294755935669, + 0.0031943004578351974, + 0.0020171254873275757, + 0.003721072105690837, + 0.003143617184832692, + 0.003170459996908903, + 0.015253073535859585, + 0.011216816492378712, + 0.014023070223629475, + 0.015486007556319237, + 0.006893593817949295, + 0.02123192697763443, + 0.008510173298418522, + 0.012789858505129814, + 0.014172255992889404, + 0.03987583890557289, + 0.017943790182471275, + 0.006694192532449961, + 0.006477890536189079, + 0.006816441658884287, + 0.004287100397050381, + 0.018243899568915367, + 0.01625184156000614, + 0.017082206904888153, + 0.015368524007499218, + 0.007683413103222847, + 0.008994387462735176, + 0.006763432640582323, + 0.007734368555247784, + 0.019907724112272263, + 0.011303129605948925, + 0.023170581087470055, + 0.018521293997764587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.2016259729862213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540886178612709 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540886178612709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540886178612709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12527096271514893, + 0.11742500215768814, + 0.07033012807369232, + 0.07210703939199448, + 0.1140650063753128, + 0.05746885761618614, + 0.10428904742002487, + 0.05033780634403229, + 0.07989099621772766, + 0.20881518721580505 + ], + "credit_map": [ + 0.01883343607187271, + 0.02214401587843895, + 0.01779385842382908, + 0.013578019104897976, + 0.01105390302836895, + 0.012230459600687027, + 0.00987586285918951, + 0.019761405885219574, + 0.017299851402640343, + 0.013283914886415005, + 0.017789050936698914, + 0.009984120726585388, + 0.013015363365411758, + 0.01053647045046091, + 0.01892871782183647, + 0.016587508842349052, + 0.0077271549962460995, + 0.004647746216505766, + 0.010029485449194908, + 0.00750737776979804, + 0.007257371675223112, + 0.008236067369580269, + 0.015844224020838737, + 0.009080703370273113, + 0.0071562426164746284, + 0.007551014423370361, + 0.01225548330694437, + 0.005630734376609325, + 0.011119639500975609, + 0.008360935375094414, + 0.011248339898884296, + 0.0087846415117383, + 0.007586977910250425, + 0.009624632075428963, + 0.013841895386576653, + 0.008483421988785267, + 0.032178010791540146, + 0.012755066156387329, + 0.016608159989118576, + 0.012986842542886734, + 0.006328405346721411, + 0.0053451755084097385, + 0.007810570765286684, + 0.005263592582195997, + 0.005219253245741129, + 0.009142193011939526, + 0.00867628213018179, + 0.00968338642269373, + 0.005824835039675236, + 0.005720246117562056, + 0.007462291046977043, + 0.004835464991629124, + 0.004907821770757437, + 0.01839403249323368, + 0.02595263347029686, + 0.03119172342121601, + 0.0036573319230228662, + 0.0044169919565320015, + 0.00591433048248291, + 0.004105269443243742, + 0.006023747380822897, + 0.006386987399309874, + 0.008418885059654713, + 0.01141426619142294, + 0.006508577149361372, + 0.008806135505437851, + 0.006929332390427589, + 0.003671926213428378, + 0.015925947576761246, + 0.012202262878417969, + 0.012530017644166946, + 0.013316791504621506, + 0.016864677891135216, + 0.018656516447663307, + 0.012878200970590115, + 0.011995112523436546, + 0.04001227021217346, + 0.029288912191987038, + 0.044116877019405365, + 0.03500262275338173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.012540886178612709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031887658406049013 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.0031887658406049013, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031887658406049013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12560713291168213, + 0.09110675007104874, + 0.06387375295162201, + 0.08561483025550842, + 0.12845325469970703, + 0.10099591314792633, + 0.17291080951690674, + 0.05234993249177933, + 0.03795476630330086, + 0.14113278687000275 + ], + "credit_map": [ + 0.01965435780584812, + 0.017120184376835823, + 0.020436059683561325, + 0.009804910980165005, + 0.010264943353831768, + 0.01655079424381256, + 0.009674427099525928, + 0.02210145629942417, + 0.015658307820558548, + 0.007674361579120159, + 0.01449426356703043, + 0.00758811691775918, + 0.006528164725750685, + 0.011791587807238102, + 0.011022805236279964, + 0.016349157318472862, + 0.007249211892485619, + 0.004097215831279755, + 0.008722848258912563, + 0.007588252890855074, + 0.0057817380875349045, + 0.009830905124545097, + 0.00956131424754858, + 0.011042272672057152, + 0.013734178617596626, + 0.003608137369155884, + 0.021368248388171196, + 0.004815129563212395, + 0.0065468112006783485, + 0.012122167274355888, + 0.011962509714066982, + 0.011457661166787148, + 0.015876738354563713, + 0.009579825215041637, + 0.035972632467746735, + 0.00856064260005951, + 0.01961418241262436, + 0.014195375144481659, + 0.01468938123434782, + 0.00996449775993824, + 0.009490588679909706, + 0.013761525973677635, + 0.01400620024651289, + 0.008447741158306599, + 0.007448819000273943, + 0.01245264895260334, + 0.014392037875950336, + 0.02099636383354664, + 0.008981138467788696, + 0.009313910268247128, + 0.01258276030421257, + 0.008221074938774109, + 0.0071767703630030155, + 0.035981111228466034, + 0.04174886271357536, + 0.04890519753098488, + 0.004286329727619886, + 0.0038276175037026405, + 0.008263764902949333, + 0.005756757687777281, + 0.007551693823188543, + 0.008651180192828178, + 0.007801559288054705, + 0.006211033556610346, + 0.005537558346986771, + 0.0042519220151007175, + 0.0036881945561617613, + 0.002680984092876315, + 0.004045004025101662, + 0.006821067072451115, + 0.005517710465937853, + 0.005412328522652388, + 0.014752421528100967, + 0.01041396427899599, + 0.011535720899701118, + 0.008270147256553173, + 0.023756777867674828, + 0.01745886355638504, + 0.03136616572737694, + 0.023578748106956482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.0031887658406049013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606367528438568 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.01606367528438568, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606367528438568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13555781543254852, + 0.08361369371414185, + 0.05061380937695503, + 0.05900047719478607, + 0.04062504693865776, + 0.16538116335868835, + 0.12349545955657959, + 0.12987111508846283, + 0.03415196016430855, + 0.17768944799900055 + ], + "credit_map": [ + 0.021617768332362175, + 0.01750003919005394, + 0.02323271706700325, + 0.01508848275989294, + 0.012627460062503815, + 0.013044435530900955, + 0.013748686760663986, + 0.018698232248425484, + 0.010027219541370869, + 0.011714696884155273, + 0.012167913839221, + 0.010237355716526508, + 0.006823640316724777, + 0.007985711097717285, + 0.014814571477472782, + 0.009842592291533947, + 0.006221224553883076, + 0.0037910102400928736, + 0.007036082446575165, + 0.0030136138666421175, + 0.006951643619686365, + 0.008479241281747818, + 0.007657940033823252, + 0.007463062182068825, + 0.008246080949902534, + 0.003258949611335993, + 0.014733567833900452, + 0.004340421408414841, + 0.004092097748070955, + 0.007660668808966875, + 0.00800246186554432, + 0.008666233159601688, + 0.004847739823162556, + 0.003229791298508644, + 0.010324249975383282, + 0.0033013327047228813, + 0.004424019251018763, + 0.004258614499121904, + 0.004808832425624132, + 0.005430469289422035, + 0.016632257029414177, + 0.009484263136982918, + 0.01782640442252159, + 0.007210147567093372, + 0.007555746473371983, + 0.015573769807815552, + 0.04054984450340271, + 0.0505487434566021, + 0.007599224336445332, + 0.00854578334838152, + 0.012338975444436073, + 0.009044832549989223, + 0.01080629788339138, + 0.029954712837934494, + 0.021482303738594055, + 0.023723341524600983, + 0.013193689286708832, + 0.006272956263273954, + 0.021731404587626457, + 0.008169026114046574, + 0.01227374654263258, + 0.012668751180171967, + 0.037573423236608505, + 0.01798812486231327, + 0.0038013760931789875, + 0.002353806747123599, + 0.0035540354438126087, + 0.0036117048002779484, + 0.0041799345053732395, + 0.005785738117992878, + 0.004748630803078413, + 0.006116736680269241, + 0.016841193661093712, + 0.01772814802825451, + 0.014815940521657467, + 0.013428602367639542, + 0.026408541947603226, + 0.022816134616732597, + 0.03600836545228958, + 0.029642537236213684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.01606367528438568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld tracking shot follows two men as they engage in a frantic chase through a narrow, damp alley. Shot entirely in high-contrast black-and-white, the leading man, wearing a thin, rain-soaked windbreaker, his damp black hair plastered to his forehead, glances back in panic, his face a panting, fearful expression. The other, more burly man in a heavy, dark coat, pursues them. His steps are heavy and determined, his face stern, his eyes gleaming with unyielding determination. The cobblestones beneath their feet are stained with filthy water, which splatters violently with each heavy step, creating splashes of silvery water that shimmer in the dim light, like shattered glass. The mottled brick walls on both sides squeeze inward, and the peeling graffiti and damp moss on the walls cast long and distorted shadows under the pale light of a single lamp deep in the alley, intensifying the oppressive sense of space and the urgency of the chase, creating a suffocating and dynamic film noir atmosphere.", + "seed": 25252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_7/index_3018_seed_25252_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.032694508973509076, + "vlm_logprob": -0.032694508973509076, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.032694508973509076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018299676477909088 + ], + "opposite_logprob_tokens": [ + -5.518299579620361 + ], + "logprob_reward": -0.018299676477909088, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.018299676477909088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24729126691818237, + 0.211666539311409, + 0.03460231423377991, + 0.0566352941095829, + 0.08152773976325989, + 0.053306933492422104, + 0.07237134128808975, + 0.061892468482255936, + 0.0826530009508133, + 0.09805309027433395 + ], + "credit_map": [ + 0.03011748008430004, + 0.009515912272036076, + 0.07257765531539917, + 0.016251977533102036, + 0.010603527538478374, + 0.02875344641506672, + 0.015999678522348404, + 0.06347158551216125, + 0.013368754647672176, + 0.004845408257097006, + 0.06843839585781097, + 0.00843142345547676, + 0.011309838853776455, + 0.02095090039074421, + 0.009190818294882774, + 0.07513099908828735, + 0.006295529194176197, + 0.00524615403264761, + 0.004297971725463867, + 0.0036466980818659067, + 0.0035513427574187517, + 0.003996050916612148, + 0.003311260137706995, + 0.004257306456565857, + 0.013117426075041294, + 0.005369017831981182, + 0.010659890249371529, + 0.0049363975413143635, + 0.0055838595144450665, + 0.00568897370249033, + 0.004841241519898176, + 0.00643848767504096, + 0.007517694029957056, + 0.00916798785328865, + 0.014322481118142605, + 0.00803410168737173, + 0.005212326068431139, + 0.010517701506614685, + 0.015723681077361107, + 0.011031768284738064, + 0.005310682114213705, + 0.006386701017618179, + 0.0061264121904969215, + 0.007396036759018898, + 0.0058768861927092075, + 0.009815214201807976, + 0.006353634875267744, + 0.00604136660695076, + 0.00832143984735012, + 0.007081145420670509, + 0.01776343211531639, + 0.007416218984872103, + 0.006862853653728962, + 0.00840548425912857, + 0.008718031458556652, + 0.007802737411111593, + 0.007533549331128597, + 0.005587030202150345, + 0.009940293617546558, + 0.005730959586799145, + 0.005737510044127703, + 0.013019997626543045, + 0.00678803026676178, + 0.007555099669843912, + 0.011827954091131687, + 0.004346941132098436, + 0.01804235950112343, + 0.005471204407513142, + 0.006077719386667013, + 0.012930062599480152, + 0.011700466275215149, + 0.01225629635155201, + 0.0155226094648242, + 0.007318999618291855, + 0.014829514548182487, + 0.007719707675278187, + 0.009765435941517353, + 0.011491701006889343, + 0.01588815450668335, + 0.015516969375312328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.018299676477909088 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297696083784103 + ], + "opposite_logprob_tokens": [ + -2.2379770278930664 + ], + "logprob_reward": -0.11297696083784103, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297696083784103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15391993522644043, + 0.16470329463481903, + 0.034923478960990906, + 0.06530751287937164, + 0.12068914622068405, + 0.06306847929954529, + 0.09031611680984497, + 0.07036366313695908, + 0.10231100022792816, + 0.1343974620103836 + ], + "credit_map": [ + 0.017312776297330856, + 0.010490983724594116, + 0.03798506408929825, + 0.014514925889670849, + 0.012069460935890675, + 0.023525161668658257, + 0.01332771871238947, + 0.02469382807612419, + 0.010911263525485992, + 0.0031094886362552643, + 0.05691507086157799, + 0.007350606843829155, + 0.007686033379286528, + 0.01262318342924118, + 0.006638144608587027, + 0.05946949124336243, + 0.003569228108972311, + 0.0040659550577402115, + 0.004901840817183256, + 0.005467122420668602, + 0.004205032251775265, + 0.0046102856285870075, + 0.003434881567955017, + 0.004669129848480225, + 0.009283051826059818, + 0.0063099609687924385, + 0.01259835995733738, + 0.005435391794890165, + 0.007025016006082296, + 0.00974444393068552, + 0.006757304538041353, + 0.008153975009918213, + 0.013502396643161774, + 0.01700817234814167, + 0.02211729995906353, + 0.014397032558918, + 0.007574187126010656, + 0.011167190037667751, + 0.018715042620897293, + 0.016207808628678322, + 0.007081677205860615, + 0.004530315287411213, + 0.008191338740289211, + 0.009054320864379406, + 0.006104237865656614, + 0.012166373431682587, + 0.00607649190351367, + 0.009863721206784248, + 0.009799805469810963, + 0.007954979315400124, + 0.009823046624660492, + 0.012004953809082508, + 0.018272005021572113, + 0.010773208923637867, + 0.011691813357174397, + 0.009996292181313038, + 0.0115765780210495, + 0.0051409900188446045, + 0.0093070762231946, + 0.005036881193518639, + 0.006559173576533794, + 0.013868547976016998, + 0.008993826806545258, + 0.009880583733320236, + 0.01000890415161848, + 0.008090578950941563, + 0.015244550071656704, + 0.005904704798012972, + 0.012843429110944271, + 0.017384016886353493, + 0.01730114407837391, + 0.01553366333246231, + 0.017700985074043274, + 0.014255915768444538, + 0.018131602555513382, + 0.01053337100893259, + 0.013474004343152046, + 0.017217615619301796, + 0.021017396822571754, + 0.022066567093133926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.11297696083784103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160689540207386 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160689540207386, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160689540207386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16280820965766907, + 0.10306762903928757, + 0.07763459533452988, + 0.1033613309264183, + 0.09091515094041824, + 0.09325921535491943, + 0.0608346164226532, + 0.07719648629426956, + 0.11879045516252518, + 0.11213229596614838 + ], + "credit_map": [ + 0.015785256400704384, + 0.013981650583446026, + 0.028483478352427483, + 0.016344161704182625, + 0.015366729348897934, + 0.031231118366122246, + 0.018164748325943947, + 0.023451080545783043, + 0.007935852743685246, + 0.004175882786512375, + 0.02966870367527008, + 0.006313642021268606, + 0.006224776618182659, + 0.009817607700824738, + 0.005433093290776014, + 0.033498071134090424, + 0.00923322606831789, + 0.008292542770504951, + 0.01452714018523693, + 0.009443135932087898, + 0.010223195888102055, + 0.006957654841244221, + 0.009027775377035141, + 0.009929919615387917, + 0.013637548312544823, + 0.014410053379833698, + 0.014410543255507946, + 0.012538391165435314, + 0.015857242047786713, + 0.009832995943725109, + 0.011844917200505733, + 0.01082963403314352, + 0.010608581826090813, + 0.010711760260164738, + 0.015250606462359428, + 0.008472873829305172, + 0.005715060979127884, + 0.01050292607396841, + 0.017945140600204468, + 0.011708205565810204, + 0.009302801452577114, + 0.010538988746702671, + 0.011832617223262787, + 0.012326272204518318, + 0.010717750526964664, + 0.017778417095541954, + 0.010670478455722332, + 0.010091891512274742, + 0.005765644833445549, + 0.0058659459464251995, + 0.00685837073251605, + 0.009428656660020351, + 0.008439157158136368, + 0.0077110277488827705, + 0.0069994693621993065, + 0.009766343981027603, + 0.006724528502672911, + 0.00567519199103117, + 0.010992472060024738, + 0.005209596827626228, + 0.012849840335547924, + 0.017084863036870956, + 0.010153836570680141, + 0.008506160229444504, + 0.011310476809740067, + 0.010445739142596722, + 0.014305755496025085, + 0.006624937057495117, + 0.02084391750395298, + 0.0202176496386528, + 0.01806328073143959, + 0.016978701576590538, + 0.017177430912852287, + 0.008086546324193478, + 0.018526315689086914, + 0.008856059983372688, + 0.01082954928278923, + 0.01280919462442398, + 0.018516194075345993, + 0.01733100786805153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0160689540207386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019369428046047688 + ], + "opposite_logprob_tokens": [ + -6.251936912536621 + ], + "logprob_reward": -0.0019369428046047688, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019369428046047688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.290073424577713, + 0.04624610021710396, + 0.03382418304681778, + 0.060814447700977325, + 0.037868574261665344, + 0.06392191350460052, + 0.03874669596552849, + 0.2229853719472885, + 0.08485376089811325, + 0.12066558748483658 + ], + "credit_map": [ + 0.030433781445026398, + 0.017531566321849823, + 0.08062152564525604, + 0.023123061284422874, + 0.01829449087381363, + 0.06518720835447311, + 0.01446104608476162, + 0.04042075574398041, + 0.0050529008731245995, + 0.004458305891603231, + 0.006618590094149113, + 0.003237666329368949, + 0.005558862816542387, + 0.008468547835946083, + 0.004169801250100136, + 0.008681421168148518, + 0.004971290938556194, + 0.004717440810054541, + 0.004694734234362841, + 0.005558292381465435, + 0.00321325589902699, + 0.003190991235896945, + 0.0035993794444948435, + 0.003878798335790634, + 0.00804345216602087, + 0.005123663228005171, + 0.01009206660091877, + 0.006474407389760017, + 0.006745997816324234, + 0.009330215863883495, + 0.0079222796484828, + 0.0070823607966303825, + 0.004317706450819969, + 0.0034393661189824343, + 0.007606200408190489, + 0.002711239270865917, + 0.0037072172854095697, + 0.005479636136442423, + 0.005236026830971241, + 0.00537118362262845, + 0.006077953148633242, + 0.006771585438400507, + 0.006290251389145851, + 0.009942341595888138, + 0.007430656813085079, + 0.013409889303147793, + 0.005989414174109697, + 0.008009816519916058, + 0.003242973005399108, + 0.0029878304339945316, + 0.003876646049320698, + 0.0032573677599430084, + 0.006039171013981104, + 0.005296370945870876, + 0.0069117737002670765, + 0.007134562358260155, + 0.01879342831671238, + 0.010684696026146412, + 0.025069234892725945, + 0.011323109269142151, + 0.01303831022232771, + 0.08409330993890762, + 0.025818072259426117, + 0.034165192395448685, + 0.009427540935575962, + 0.007213881704956293, + 0.013227376155555248, + 0.008688489906489849, + 0.0072957235388457775, + 0.011394564993679523, + 0.014363299123942852, + 0.013242888264358044, + 0.019582631066441536, + 0.009531689807772636, + 0.018594050779938698, + 0.008560335263609886, + 0.011567577719688416, + 0.017449673265218735, + 0.016952430829405785, + 0.01842719316482544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0019369428046047688 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-contrast black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and is the shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two men visible, one wearing a thin rain-soaked windbreaker and the other a heavy dark coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, damp alley with mottled brick walls and cobblestones stained with filthy water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a single lamp casting long, distorted shadows and silvery water splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190010726451874 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190010726451874, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190010726451874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25987204909324646, + 0.13387055695056915, + 0.043976735323667526, + 0.044820014387369156, + 0.044270697981119156, + 0.04957892373204231, + 0.06048062443733215, + 0.1403377652168274, + 0.09285153448581696, + 0.12994104623794556 + ], + "credit_map": [ + 0.031438883394002914, + 0.008842622861266136, + 0.0890829786658287, + 0.015642600134015083, + 0.009004142135381699, + 0.028936780989170074, + 0.012316307052969933, + 0.06460773199796677, + 0.00856389943510294, + 0.0032757131848484278, + 0.04664764553308487, + 0.005887793842703104, + 0.006777271628379822, + 0.007831929251551628, + 0.006726311054080725, + 0.04816000536084175, + 0.0058081080205738544, + 0.006627511233091354, + 0.005545596592128277, + 0.005292165093123913, + 0.005220613442361355, + 0.006251944694668055, + 0.005251025781035423, + 0.003979772329330444, + 0.00732919666916132, + 0.0035209120251238346, + 0.007816915400326252, + 0.0032544005662202835, + 0.00571803655475378, + 0.007244479842483997, + 0.004493187181651592, + 0.005442887078970671, + 0.003890285035595298, + 0.003011049935594201, + 0.006420660298317671, + 0.004536387976258993, + 0.0033423835411667824, + 0.009104713797569275, + 0.006421975791454315, + 0.007543237879872322, + 0.004303410183638334, + 0.0077766976319253445, + 0.0055525475181639194, + 0.0067487675696611404, + 0.004663973581045866, + 0.00777945714071393, + 0.005773853976279497, + 0.006980213336646557, + 0.008378703147172928, + 0.006015304010361433, + 0.010327046737074852, + 0.007775960490107536, + 0.00548899918794632, + 0.006893011275678873, + 0.007481156848371029, + 0.008120444603264332, + 0.012151177041232586, + 0.006716463714838028, + 0.017181353643536568, + 0.005233196076005697, + 0.007335682865232229, + 0.05300530046224594, + 0.01618865132331848, + 0.022525936365127563, + 0.01237989217042923, + 0.009279693476855755, + 0.015316328033804893, + 0.006980422418564558, + 0.009984221309423447, + 0.012720319442451, + 0.012179925106465816, + 0.014010736718773842, + 0.02082703821361065, + 0.011368791572749615, + 0.02208269201219082, + 0.010492898523807526, + 0.0093257250264287, + 0.012660779058933258, + 0.023196179419755936, + 0.019986944273114204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.014190010726451874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_8/index_tensor([5919, 5919])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_8/index_tensor([5919, 5919])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e791b5866ff74df2d6523a7287d9585387480fc6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_8/index_tensor([5919, 5919])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot shot from an extremely low angle, the camera slowly and steadily tilting upward in sync with the subject's movements, captures a scene imbued with Eastern Zen. In the center of a mirror-like lake, shimmering in a shimmering, inky hue, the surface silently breaks, revealing a massive, divine serpent rising from within. Its scales, deep as obsidian, gleam with a moist, warm sheen under the soft, diffused skylight. As its ancient, majestic head and sturdy, long body ascend with steady composure, countless crystalline droplets, like diamonds, carefully slide along the edges of each smooth, dense scale, finally dripping silently back into the lake, stirring tiny ripples. Distant mountains loom in the mist, the scene awash in ink-like black, white, and gray. Only the serpent's golden, vertical pupils, gleaming with the light of ancient wisdom, imbue this tranquil and majestic scene with a mysterious vitality, creating a sense of ultimate tranquility and power.", + "seed": 25501, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_8/index_5919_seed_25501_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.034513223171234134, + "vlm_logprob": -0.034513223171234134, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.034513223171234134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0727626234292984 + ], + "opposite_logprob_tokens": [ + -2.9477627277374268 + ], + "logprob_reward": -0.0727626234292984, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0727626234292984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10393641889095306, + 0.07103528082370758, + 0.08867651224136353, + 0.3123103678226471, + 0.04831065982580185, + 0.036990199238061905, + 0.2179657220840454, + 0.04151720181107521, + 0.03382270038127899, + 0.045434847474098206 + ], + "credit_map": [ + 0.010832318104803562, + 0.011590462177991867, + 0.012220638804137707, + 0.013194679282605648, + 0.007632972672581673, + 0.008899093605577946, + 0.0218758936971426, + 0.01769036054611206, + 0.005167558789253235, + 0.00670473650097847, + 0.008300173096358776, + 0.01338058803230524, + 0.007488888222724199, + 0.009254863485693932, + 0.011268841102719307, + 0.009469634853303432, + 0.0069302767515182495, + 0.008653240278363228, + 0.007456731516867876, + 0.011758480221033096, + 0.011009219102561474, + 0.012600096873939037, + 0.011373774148523808, + 0.01889469288289547, + 0.01830580271780491, + 0.03465241566300392, + 0.05093088373541832, + 0.047869447618722916, + 0.026597758755087852, + 0.028999583795666695, + 0.05122612789273262, + 0.0537283681333065, + 0.004414650611579418, + 0.0035004224628210068, + 0.006463813129812479, + 0.005731310695409775, + 0.004813112784177065, + 0.008525034412741661, + 0.006797580048441887, + 0.008064732886850834, + 0.004037058912217617, + 0.005714056547731161, + 0.003556150710210204, + 0.005150573328137398, + 0.003044607350602746, + 0.005238695070147514, + 0.004638646263629198, + 0.0056104096584022045, + 0.030661476776003838, + 0.01644965633749962, + 0.04416390135884285, + 0.013509326614439487, + 0.012428470887243748, + 0.02117639034986496, + 0.02010112814605236, + 0.059475380927324295, + 0.007003950420767069, + 0.0026487368158996105, + 0.00582133000716567, + 0.004802360199391842, + 0.00356030510738492, + 0.005996281281113625, + 0.0065470244735479355, + 0.005137215834110975, + 0.005351176019757986, + 0.0026375772431492805, + 0.004915454424917698, + 0.0032445865217596292, + 0.002432514913380146, + 0.0028685173019766808, + 0.003851574845612049, + 0.008521300740540028, + 0.007354434579610825, + 0.002925638807937503, + 0.006583868525922298, + 0.0035961277317255735, + 0.0029238027054816484, + 0.005668812897056341, + 0.006149415858089924, + 0.010232750326395035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0727626234292984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029827827587723732 + ], + "opposite_logprob_tokens": [ + -3.529827833175659 + ], + "logprob_reward": -0.029827827587723732, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029827827587723732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10580041259527206, + 0.08292791992425919, + 0.0922635942697525, + 0.13251681625843048, + 0.05606011301279068, + 0.057508684694767, + 0.15671393275260925, + 0.04932605102658272, + 0.10847298800945282, + 0.1584094911813736 + ], + "credit_map": [ + 0.00778731657192111, + 0.0075309984385967255, + 0.010196262039244175, + 0.013866479508578777, + 0.010497591458261013, + 0.011171559803187847, + 0.02416144497692585, + 0.02058875374495983, + 0.0062842839397490025, + 0.011090794578194618, + 0.008826341480016708, + 0.012013309635221958, + 0.011083675548434258, + 0.010881965979933739, + 0.012222414836287498, + 0.010525126941502094, + 0.0065198116935789585, + 0.00956277921795845, + 0.007579579018056393, + 0.011873728595674038, + 0.0162813700735569, + 0.00947166420519352, + 0.011804904788732529, + 0.0191697608679533, + 0.0075109233148396015, + 0.014368070289492607, + 0.021304607391357422, + 0.02042231149971485, + 0.013176199048757553, + 0.011584151536226273, + 0.02580631710588932, + 0.018344249576330185, + 0.0046613458544015884, + 0.004164091311395168, + 0.005568599794059992, + 0.007488985080271959, + 0.006417962722480297, + 0.007377873174846172, + 0.010021928697824478, + 0.01035932544618845, + 0.005700437352061272, + 0.009991669096052647, + 0.005495666526257992, + 0.007663011085242033, + 0.00548495352268219, + 0.007073951419442892, + 0.006474674679338932, + 0.009624321013689041, + 0.022230258211493492, + 0.010809809900820255, + 0.022956043481826782, + 0.012120984494686127, + 0.010262439958751202, + 0.02127034217119217, + 0.028041962534189224, + 0.02902209386229515, + 0.0070961955934762955, + 0.004320881329476833, + 0.007956331595778465, + 0.007529407739639282, + 0.006347961258143187, + 0.0051818666979670525, + 0.005243203602731228, + 0.0056501999497413635, + 0.01481632050126791, + 0.008151322603225708, + 0.016413070261478424, + 0.020889608189463615, + 0.013471444137394428, + 0.009308801963925362, + 0.01344306766986847, + 0.011979351751506329, + 0.026980644091963768, + 0.01024201512336731, + 0.024832984432578087, + 0.013624918647110462, + 0.011114177294075489, + 0.016011400148272514, + 0.01991545967757702, + 0.0356878787279129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.029827827587723732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525260902941227 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.00525260902941227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525260902941227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21843725442886353, + 0.13481658697128296, + 0.21401424705982208, + 0.09671133756637573, + 0.1234779879450798, + 0.04273652285337448, + 0.10024057328701019, + 0.027301020920276642, + 0.013147341087460518, + 0.029117116704583168 + ], + "credit_map": [ + 0.01720249652862549, + 0.011386347003281116, + 0.022790230810642242, + 0.03135552257299423, + 0.014757405035197735, + 0.013480729423463345, + 0.061174433678388596, + 0.0462900809943676, + 0.009332660585641861, + 0.010859590023756027, + 0.01828886941075325, + 0.028924832120537758, + 0.014648852869868279, + 0.014414804987609386, + 0.023598674684762955, + 0.014748306013643742, + 0.01399731170386076, + 0.020774925127625465, + 0.01663639396429062, + 0.030364928767085075, + 0.03147926554083824, + 0.022795500233769417, + 0.02571626752614975, + 0.052249666303396225, + 0.005366286262869835, + 0.01028336863964796, + 0.014618504792451859, + 0.014556827023625374, + 0.012664524838328362, + 0.009426740929484367, + 0.01585208810865879, + 0.013942996971309185, + 0.00991761963814497, + 0.011310498230159283, + 0.014987598173320293, + 0.014512826688587666, + 0.013871297240257263, + 0.01250939629971981, + 0.023988857865333557, + 0.022379882633686066, + 0.0037261871621012688, + 0.004953787662088871, + 0.004774280823767185, + 0.0065450845286250114, + 0.0029717867728322744, + 0.003414980135858059, + 0.005064031109213829, + 0.011286384426057339, + 0.012309649027884007, + 0.010105046443641186, + 0.019327456131577492, + 0.011231699027121067, + 0.008831790648400784, + 0.010348320938646793, + 0.01143383514136076, + 0.016652781516313553, + 0.0037954181898385286, + 0.0026135179214179516, + 0.0034398322459310293, + 0.0038802712224423885, + 0.002326339017599821, + 0.002421555807814002, + 0.004875356797128916, + 0.003948730416595936, + 0.00198587984777987, + 0.0012029596837237477, + 0.001957924338057637, + 0.0015952474204823375, + 0.001703204121440649, + 0.0012695279438048601, + 0.0017155373934656382, + 0.0017170602222904563, + 0.004872984252870083, + 0.0015374562935903668, + 0.006044832058250904, + 0.001564164413139224, + 0.0015139954630285501, + 0.003259106306359172, + 0.004410699009895325, + 0.005913878325372934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00525260902941227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866102337837219 + ], + "opposite_logprob_tokens": [ + -3.04866099357605 + ], + "logprob_reward": -0.04866102337837219, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866102337837219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29915809631347656, + 0.11029964685440063, + 0.20595237612724304, + 0.05080273374915123, + 0.06718418002128601, + 0.040601134300231934, + 0.0790022611618042, + 0.02407000958919525, + 0.02431161142885685, + 0.09861791878938675 + ], + "credit_map": [ + 0.02536683715879917, + 0.018095530569553375, + 0.031120503321290016, + 0.04421587660908699, + 0.02186998724937439, + 0.0204677302390337, + 0.07733581215143204, + 0.06068584322929382, + 0.00730656273663044, + 0.005956526845693588, + 0.015560219995677471, + 0.024691347032785416, + 0.012815829366445541, + 0.011757557280361652, + 0.019210070371627808, + 0.01300153136253357, + 0.013357815332710743, + 0.019442375749349594, + 0.0155855817720294, + 0.02888023480772972, + 0.030606087297201157, + 0.022376198321580887, + 0.024947235360741615, + 0.05075684189796448, + 0.0030453826766461134, + 0.005542262457311153, + 0.005815103184431791, + 0.007277719210833311, + 0.0050589777529239655, + 0.0055059269070625305, + 0.008526964113116264, + 0.010030399076640606, + 0.005246895365417004, + 0.005880082957446575, + 0.009283028542995453, + 0.008777568116784096, + 0.007864830084145069, + 0.005728996358811855, + 0.011363362893462181, + 0.013039418496191502, + 0.0038375258445739746, + 0.004221557639539242, + 0.0036297705955803394, + 0.003946125507354736, + 0.002989207860082388, + 0.003863382386043668, + 0.004165732301771641, + 0.013947831466794014, + 0.012426338158547878, + 0.0052797128446400166, + 0.00976328644901514, + 0.008424012921750546, + 0.00661744037643075, + 0.007384226191788912, + 0.01532037928700447, + 0.013786868192255497, + 0.004316031467169523, + 0.0017286663642153144, + 0.004500302951782942, + 0.0023461072705686092, + 0.002029381925240159, + 0.0016660001128911972, + 0.002333474811166525, + 0.005150045733898878, + 0.0045781382359564304, + 0.0016708002658560872, + 0.003233406227082014, + 0.00282013276591897, + 0.002330937422811985, + 0.0021377382799983025, + 0.0032933587208390236, + 0.004247099626809359, + 0.016230182722210884, + 0.004554462619125843, + 0.019346721470355988, + 0.005521301180124283, + 0.004163221921771765, + 0.014512091875076294, + 0.012600119225680828, + 0.02168981544673443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04866102337837219 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606203243136406 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.01606203243136406, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606203243136406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19203278422355652, + 0.1335369199514389, + 0.1297132819890976, + 0.08777105808258057, + 0.05971706286072731, + 0.06511364132165909, + 0.04790053516626358, + 0.19899581372737885, + 0.05029255524277687, + 0.034926313906908035 + ], + "credit_map": [ + 0.013975030742585659, + 0.018459469079971313, + 0.03127468004822731, + 0.03658616915345192, + 0.018638014793395996, + 0.015553733333945274, + 0.03212883695960045, + 0.025416860356926918, + 0.009361892007291317, + 0.01577899418771267, + 0.016529133543372154, + 0.02616584487259388, + 0.016283616423606873, + 0.013334061950445175, + 0.02127603441476822, + 0.014807343482971191, + 0.008656897582113743, + 0.012203960679471493, + 0.012053946033120155, + 0.01770802028477192, + 0.019753510132431984, + 0.014618156477808952, + 0.015130924992263317, + 0.029587866738438606, + 0.005420040804892778, + 0.009405591525137424, + 0.01273907907307148, + 0.012541454285383224, + 0.011595642194151878, + 0.008646917529404163, + 0.01390773057937622, + 0.013514603488147259, + 0.004745897836983204, + 0.004629654809832573, + 0.011256483383476734, + 0.0077213807962834835, + 0.007446744479238987, + 0.005820733495056629, + 0.008946715854108334, + 0.00914944987744093, + 0.006350630894303322, + 0.007525355089455843, + 0.006686771754175425, + 0.007030427921563387, + 0.003987933974713087, + 0.0057248263619840145, + 0.006838739849627018, + 0.020968953147530556, + 0.006188970524817705, + 0.004402194172143936, + 0.008346198126673698, + 0.0038962147664278746, + 0.00463218754157424, + 0.004356901161372662, + 0.007027636747807264, + 0.009050227701663971, + 0.03505638986825943, + 0.011972218751907349, + 0.029428306967020035, + 0.012236974202096462, + 0.014331527054309845, + 0.013527961447834969, + 0.03763330355286598, + 0.04480914771556854, + 0.009355463087558746, + 0.0031154905445873737, + 0.007632412947714329, + 0.0036088621709495783, + 0.0024986518546938896, + 0.004733754321932793, + 0.005946907214820385, + 0.013401014730334282, + 0.006631769239902496, + 0.001869753934442997, + 0.005575005430728197, + 0.002811022335663438, + 0.0018541984027251601, + 0.00449464563280344, + 0.0037789293564856052, + 0.007910989224910736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.01606203243136406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly and steadily tilting upward in sync with the subject's movements, captures a scene imbued with Eastern Zen. In the center of a mirror-like lake, shimmering in a shimmering, inky hue, the surface silently breaks, revealing a massive, divine serpent rising from within. Its scales, deep as obsidian, gleam with a moist, warm sheen under the soft, diffused skylight. As its ancient, majestic head and sturdy, long body ascend with steady composure, countless crystalline droplets, like diamonds, carefully slide along the edges of each smooth, dense scale, finally dripping silently back into the lake, stirring tiny ripples. Distant mountains loom in the mist, the scene awash in ink-like black, white, and gray. Only the serpent's golden, vertical pupils, gleaming with the light of ancient wisdom, imbue this tranquil and majestic scene with a mysterious vitality, creating a sense of ultimate tranquility and power.", + "seed": 25502, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_8/index_5919_seed_25502_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13926064185798168, + "vlm_logprob": -0.13926064185798168, + "vlm_yes_no_margin": 2.799999988079071, + "vlm_token_credit": -0.13926064185798168, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.45643478631973267 + ], + "opposite_logprob_tokens": [ + -1.081434726715088 + ], + "logprob_reward": -0.45643478631973267, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.45643478631973267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20705974102020264, + 0.07052560150623322, + 0.055333804339170456, + 0.3513907194137573, + 0.045617979019880295, + 0.0652887150645256, + 0.07334094494581223, + 0.04607076197862625, + 0.04637265205383301, + 0.03899907320737839 + ], + "credit_map": [ + 0.025094585493206978, + 0.022053731605410576, + 0.027046242728829384, + 0.02752511203289032, + 0.024560019373893738, + 0.026170840486884117, + 0.028052490204572678, + 0.026556720957159996, + 0.007497373037040234, + 0.008119920268654823, + 0.007561864331364632, + 0.010449262335896492, + 0.009158920496702194, + 0.010404710657894611, + 0.011682911776006222, + 0.005650639068335295, + 0.006541504058986902, + 0.009034561924636364, + 0.00877238716930151, + 0.005942835472524166, + 0.006476175505667925, + 0.007012896239757538, + 0.0071801673620939255, + 0.004373278468847275, + 0.046404968947172165, + 0.08280761539936066, + 0.03605189174413681, + 0.03355511277914047, + 0.037313371896743774, + 0.05264201760292053, + 0.040516447275877, + 0.022099312394857407, + 0.005345172248780727, + 0.006193045061081648, + 0.004304266069084406, + 0.005150981247425079, + 0.011292412877082825, + 0.005252405535429716, + 0.004205802455544472, + 0.0038738930597901344, + 0.011563538573682308, + 0.005547833628952503, + 0.004964814521372318, + 0.005320135038346052, + 0.008571522310376167, + 0.016564548015594482, + 0.004561030305922031, + 0.008195288479328156, + 0.008840919472277164, + 0.008389136753976345, + 0.007468709256500006, + 0.007705759722739458, + 0.008851929567754269, + 0.005820498801767826, + 0.009161538444459438, + 0.017102450132369995, + 0.0076527101919054985, + 0.005285785533487797, + 0.004021693952381611, + 0.006429797504097223, + 0.004651627968996763, + 0.003925235476344824, + 0.005931600462645292, + 0.008172311820089817, + 0.004792667459696531, + 0.006065013352781534, + 0.0055577922612428665, + 0.006442717742174864, + 0.004642285406589508, + 0.003195705823600292, + 0.00776485912501812, + 0.007911612279713154, + 0.0053209043107926846, + 0.0032369920518249273, + 0.005194989964365959, + 0.005855059251189232, + 0.003965180832892656, + 0.003622801974415779, + 0.007077326532453299, + 0.004725820384919643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.45643478631973267 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062014304101467 + ], + "opposite_logprob_tokens": [ + -3.895620107650757 + ], + "logprob_reward": -0.02062014304101467, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062014304101467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10789035260677338, + 0.05351978912949562, + 0.10159198939800262, + 0.17444296181201935, + 0.07256358861923218, + 0.14777041971683502, + 0.08080004900693893, + 0.11429457366466522, + 0.12507523596286774, + 0.022051021456718445 + ], + "credit_map": [ + 0.01827036216855049, + 0.011054189875721931, + 0.015542552806437016, + 0.015075325965881348, + 0.011317668482661247, + 0.013073069043457508, + 0.014394603669643402, + 0.009162585251033306, + 0.005559817422181368, + 0.005460365209728479, + 0.006179748103022575, + 0.010456928052008152, + 0.005973578430712223, + 0.006330989301204681, + 0.009161032736301422, + 0.00439732801169157, + 0.007945001125335693, + 0.017777228727936745, + 0.01764283888041973, + 0.012474092654883862, + 0.01114840991795063, + 0.011890525929629803, + 0.014345080591738224, + 0.008368819952011108, + 0.026235856115818024, + 0.031280357390642166, + 0.015080869197845459, + 0.02030806802213192, + 0.028914468362927437, + 0.025404920801520348, + 0.012471313588321209, + 0.014747108332812786, + 0.007772192824631929, + 0.007394301239401102, + 0.006950384005904198, + 0.008816099725663662, + 0.01735486276447773, + 0.007633103057742119, + 0.007669537328183651, + 0.008973104879260063, + 0.017735488712787628, + 0.014623009599745274, + 0.014588774181902409, + 0.012603714130818844, + 0.039493128657341, + 0.022071991115808487, + 0.01138217281550169, + 0.01527214702218771, + 0.010001740418374538, + 0.011693903245031834, + 0.005930858664214611, + 0.009945890866219997, + 0.012699701823294163, + 0.007109737489372492, + 0.008987709879875183, + 0.014430512674152851, + 0.023306826129555702, + 0.017312349751591682, + 0.007883979007601738, + 0.013783928006887436, + 0.014782372862100601, + 0.012142638675868511, + 0.009668008424341679, + 0.015414472669363022, + 0.011999960988759995, + 0.014918336644768715, + 0.015118516981601715, + 0.017785102128982544, + 0.01390141248703003, + 0.010066832415759563, + 0.021893292665481567, + 0.019391778856515884, + 0.002406598534435034, + 0.0016214903444051743, + 0.003473974298685789, + 0.0046904198825359344, + 0.0017705101054161787, + 0.0019173361361026764, + 0.0030034068040549755, + 0.003167286515235901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.02062014304101467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286327719688416 + ], + "opposite_logprob_tokens": [ + -3.7732863426208496 + ], + "logprob_reward": -0.023286327719688416, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286327719688416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10182768106460571, + 0.08024461567401886, + 0.33579474687576294, + 0.12880685925483704, + 0.07054273784160614, + 0.0874023586511612, + 0.052681341767311096, + 0.04824270308017731, + 0.03898194432258606, + 0.05547500401735306 + ], + "credit_map": [ + 0.011188234202563763, + 0.008592267520725727, + 0.01259343046694994, + 0.015276246704161167, + 0.013126333244144917, + 0.014287905767560005, + 0.017622342333197594, + 0.009140914306044579, + 0.006208740174770355, + 0.011688325554132462, + 0.006805740296840668, + 0.01376739889383316, + 0.008686752058565617, + 0.013138314709067345, + 0.010409011505544186, + 0.009540331549942493, + 0.03052385151386261, + 0.039840660989284515, + 0.026585903018712997, + 0.031796738505363464, + 0.08798783272504807, + 0.048342473804950714, + 0.04072438180446625, + 0.02999289147555828, + 0.013935328461229801, + 0.025066975504159927, + 0.012259095907211304, + 0.013244039379060268, + 0.015975555405020714, + 0.018278615549206734, + 0.017836885526776314, + 0.012210365384817123, + 0.008227581158280373, + 0.010924307629466057, + 0.005865663755685091, + 0.006635880563408136, + 0.015342309139668941, + 0.008448216132819653, + 0.00709676556289196, + 0.008002013899385929, + 0.01686233840882778, + 0.005838816985487938, + 0.006060025189071894, + 0.00722731975838542, + 0.016058072447776794, + 0.02220238745212555, + 0.005751547869294882, + 0.007401856128126383, + 0.0059296805411577225, + 0.0053521739318966866, + 0.011046390980482101, + 0.006809588056057692, + 0.006350741256028414, + 0.003843597834929824, + 0.00478222593665123, + 0.008566944859921932, + 0.006531050428748131, + 0.007228844799101353, + 0.005500036757439375, + 0.007339907810091972, + 0.006357837934046984, + 0.004624519031494856, + 0.005278284195810556, + 0.005382219795137644, + 0.004684631712734699, + 0.005658057983964682, + 0.0033005166333168745, + 0.0059376489371061325, + 0.003779839491471648, + 0.003664637915790081, + 0.005662491079419851, + 0.006294121500104666, + 0.006804589182138443, + 0.00449919980019331, + 0.005158367101103067, + 0.016002235934138298, + 0.005637506954371929, + 0.0042648003436625, + 0.007641434669494629, + 0.00546687189489603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.023286327719688416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798992156982422 + ], + "opposite_logprob_tokens": [ + -1.8048992156982422 + ], + "logprob_reward": -0.1798992156982422, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798992156982422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18939325213432312, + 0.11103098094463348, + 0.0931834802031517, + 0.1641843467950821, + 0.04844166338443756, + 0.11605671793222427, + 0.07420764863491058, + 0.04928964748978615, + 0.05423640087246895, + 0.09997590631246567 + ], + "credit_map": [ + 0.027635738253593445, + 0.018662994727492332, + 0.01983572170138359, + 0.02207682467997074, + 0.026419784873723984, + 0.025345973670482635, + 0.03249890357255936, + 0.01691729947924614, + 0.009549272246658802, + 0.01231948658823967, + 0.014538835734128952, + 0.02037431113421917, + 0.01125521119683981, + 0.013632489368319511, + 0.015045417472720146, + 0.014315957203507423, + 0.010022515431046486, + 0.014595601707696915, + 0.015354896895587444, + 0.008361225947737694, + 0.01386487577110529, + 0.010567793622612953, + 0.01165135856717825, + 0.008765206672251225, + 0.025857530534267426, + 0.025021078065037727, + 0.012345103546977043, + 0.020632654428482056, + 0.022010819986462593, + 0.018114717677235603, + 0.022080207243561745, + 0.018122225999832153, + 0.004422617144882679, + 0.00667730113491416, + 0.00580780440941453, + 0.005236242897808552, + 0.011304249987006187, + 0.005099096801131964, + 0.004417354241013527, + 0.005476996302604675, + 0.03190989047288895, + 0.006027598865330219, + 0.00803079642355442, + 0.008129342459142208, + 0.010762061923742294, + 0.03676893189549446, + 0.0054331435821950436, + 0.00899494532495737, + 0.010094694793224335, + 0.007575420197099447, + 0.01572192646563053, + 0.007213303819298744, + 0.007838518358767033, + 0.006122145801782608, + 0.0074374862015247345, + 0.01220414973795414, + 0.005906158126890659, + 0.008079434745013714, + 0.004552221391350031, + 0.006014527752995491, + 0.007895194925367832, + 0.004234318155795336, + 0.005284118466079235, + 0.007323674391955137, + 0.006298273801803589, + 0.004791238345205784, + 0.005097797140479088, + 0.008845598436892033, + 0.005406276322901249, + 0.004526295233517885, + 0.008955460973083973, + 0.010315458290278912, + 0.011663440614938736, + 0.00763529259711504, + 0.012564338743686676, + 0.024777770042419434, + 0.009797824546694756, + 0.007801745552569628, + 0.015145022422075272, + 0.010590475983917713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.1798992156982422 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606273651123047 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.01606273651123047, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606273651123047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22155046463012695, + 0.11103541404008865, + 0.2415313571691513, + 0.10197729617357254, + 0.045547083020210266, + 0.13237334787845612, + 0.032113537192344666, + 0.04194394871592522, + 0.027758831158280373, + 0.044168684631586075 + ], + "credit_map": [ + 0.017698558047413826, + 0.02892252616584301, + 0.01783526875078678, + 0.019566329196095467, + 0.041664186865091324, + 0.05801019072532654, + 0.02026536874473095, + 0.017588065937161446, + 0.004951259586960077, + 0.016307039186358452, + 0.008182759396731853, + 0.015397393144667149, + 0.019809015095233917, + 0.026590507477521896, + 0.011659921146929264, + 0.008137526921927929, + 0.03077780269086361, + 0.051167573779821396, + 0.03481432795524597, + 0.024053210392594337, + 0.021085776388645172, + 0.030361352488398552, + 0.03021330200135708, + 0.019058018922805786, + 0.01320409495383501, + 0.017398854717612267, + 0.007586368825286627, + 0.013189355842769146, + 0.01260907668620348, + 0.011398100294172764, + 0.01497140247374773, + 0.011620046570897102, + 0.00908640120178461, + 0.005176669452339411, + 0.004285023547708988, + 0.003627179190516472, + 0.01054250355809927, + 0.0047015235759317875, + 0.003696838393807411, + 0.004430950619280338, + 0.03268042206764221, + 0.007026124279946089, + 0.00752202607691288, + 0.008033697493374348, + 0.02039838582277298, + 0.03963230922818184, + 0.005471282638609409, + 0.01160911563783884, + 0.0045449137687683105, + 0.002975770737975836, + 0.0035707568749785423, + 0.002897298661991954, + 0.004634527955204248, + 0.002807577606290579, + 0.0035837378818541765, + 0.007098956033587456, + 0.004687360022217035, + 0.0055111306719481945, + 0.004459097515791655, + 0.00509762903675437, + 0.004065303131937981, + 0.0048260148614645, + 0.004486192483454943, + 0.0088112261146307, + 0.0032347668893635273, + 0.0030878325924277306, + 0.0025092880241572857, + 0.0036968113854527473, + 0.002682995982468128, + 0.0026585867162793875, + 0.004983949940651655, + 0.004904600791633129, + 0.0055262381210923195, + 0.0035089526791125536, + 0.0033523705787956715, + 0.012030725367367268, + 0.004115129355341196, + 0.0036502040456980467, + 0.006519327871501446, + 0.005465742666274309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.01606273651123047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly and steadily tilting upward in sync with the subject's movements, captures a scene imbued with Eastern Zen. In the center of a mirror-like lake, shimmering in a shimmering, inky hue, the surface silently breaks, revealing a massive, divine serpent rising from within. Its scales, deep as obsidian, gleam with a moist, warm sheen under the soft, diffused skylight. As its ancient, majestic head and sturdy, long body ascend with steady composure, countless crystalline droplets, like diamonds, carefully slide along the edges of each smooth, dense scale, finally dripping silently back into the lake, stirring tiny ripples. Distant mountains loom in the mist, the scene awash in ink-like black, white, and gray. Only the serpent's golden, vertical pupils, gleaming with the light of ancient wisdom, imbue this tranquil and majestic scene with a mysterious vitality, creating a sense of ultimate tranquility and power.", + "seed": 25503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_8/index_5919_seed_25503_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1588964938186109, + "vlm_logprob": -0.1588964938186109, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.1588964938186109, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6534106731414795 + ], + "opposite_logprob_tokens": [ + -0.7784106731414795 + ], + "logprob_reward": -0.6534106731414795, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6534106731414795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1442195326089859, + 0.07332746684551239, + 0.04656495898962021, + 0.38539767265319824, + 0.06904163211584091, + 0.08020683377981186, + 0.09843777865171432, + 0.023046445101499557, + 0.039472538977861404, + 0.04028509184718132 + ], + "credit_map": [ + 0.012344500049948692, + 0.009745802730321884, + 0.041760172694921494, + 0.014565535821020603, + 0.005872458219528198, + 0.010067119263112545, + 0.010674415156245232, + 0.03918954357504845, + 0.00636233389377594, + 0.0069860368967056274, + 0.010252183303236961, + 0.008725643157958984, + 0.0077819726429879665, + 0.008229084312915802, + 0.013618577271699905, + 0.011371641419827938, + 0.010387038812041283, + 0.0038384399376809597, + 0.006742624565958977, + 0.006011867895722389, + 0.0029810748528689146, + 0.003427480813115835, + 0.005697822663933039, + 0.007478613406419754, + 0.09847799688577652, + 0.03236241266131401, + 0.052958253771066666, + 0.0411994569003582, + 0.0392773263156414, + 0.034481361508369446, + 0.04631628096103668, + 0.040324606001377106, + 0.010146948508918285, + 0.005059728864580393, + 0.007308642845600843, + 0.006245211232453585, + 0.009408469311892986, + 0.00868747103959322, + 0.008848090656101704, + 0.013337072916328907, + 0.009979208000004292, + 0.009324769489467144, + 0.011912676505744457, + 0.009794923476874828, + 0.010130606591701508, + 0.009019930846989155, + 0.011874060146510601, + 0.00817067176103592, + 0.01164151169359684, + 0.0120486319065094, + 0.010889099910855293, + 0.009096377529203892, + 0.0110544478520751, + 0.013344001956284046, + 0.018747467547655106, + 0.011616246774792671, + 0.0022551382426172495, + 0.0018584290519356728, + 0.0023348683025687933, + 0.0018684015376493335, + 0.003521440550684929, + 0.0028119387570768595, + 0.005169237963855267, + 0.003226992441341281, + 0.008843486197292805, + 0.0033481537830084562, + 0.0027395307552069426, + 0.003448509145528078, + 0.001805642037652433, + 0.007625390309840441, + 0.0076894015073776245, + 0.003972427453845739, + 0.0038257294800132513, + 0.0035949815064668655, + 0.005181116051971912, + 0.00542260380461812, + 0.004760064650326967, + 0.004869392141699791, + 0.006711592432111502, + 0.005919618997722864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.6534106731414795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909104764461517 + ], + "opposite_logprob_tokens": [ + -2.4640910625457764 + ], + "logprob_reward": -0.08909104764461517, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909104764461517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16265228390693665, + 0.08496095985174179, + 0.06956937909126282, + 0.09210038930177689, + 0.11893944442272186, + 0.11400048434734344, + 0.07483340799808502, + 0.09034710377454758, + 0.06269830465316772, + 0.12989825010299683 + ], + "credit_map": [ + 0.015013928525149822, + 0.013514528051018715, + 0.023425191640853882, + 0.02522786147892475, + 0.022049488499760628, + 0.0170526634901762, + 0.02671218290925026, + 0.01965644396841526, + 0.01033723820000887, + 0.011469532735645771, + 0.01166068110615015, + 0.010401283390820026, + 0.01095821987837553, + 0.011236879974603653, + 0.01036372222006321, + 0.008533407002687454, + 0.010533430613577366, + 0.006648286711424589, + 0.009582308121025562, + 0.008186877705156803, + 0.005149423144757748, + 0.00490017281845212, + 0.01492324285209179, + 0.009645640850067139, + 0.019476093351840973, + 0.005932862404733896, + 0.011510541662573814, + 0.012825636193156242, + 0.006158723495900631, + 0.008042423985898495, + 0.012811719439923763, + 0.015342392958700657, + 0.01861218921840191, + 0.012330460362136364, + 0.017725981771945953, + 0.012333582155406475, + 0.007895505987107754, + 0.02047552727162838, + 0.018798332661390305, + 0.010767876170575619, + 0.012569933198392391, + 0.010477100498974323, + 0.01408354938030243, + 0.013093278743326664, + 0.017057830467820168, + 0.01442800834774971, + 0.020809682086110115, + 0.011481108143925667, + 0.01048322580754757, + 0.008496066555380821, + 0.01092029269784689, + 0.006671997718513012, + 0.00818109791725874, + 0.012137922458350658, + 0.010380353778600693, + 0.007562455255538225, + 0.010812661610543728, + 0.009366504848003387, + 0.009775168262422085, + 0.006356145720928907, + 0.01782461442053318, + 0.011852730996906757, + 0.012101155705749989, + 0.0122581347823143, + 0.011581034399569035, + 0.0038668594788759947, + 0.004449476022273302, + 0.00414752122014761, + 0.004102810751646757, + 0.009340586140751839, + 0.018320683389902115, + 0.006889333948493004, + 0.014678317122161388, + 0.010656828992068768, + 0.020941216498613358, + 0.008255970664322376, + 0.009070299565792084, + 0.019098445773124695, + 0.01514388620853424, + 0.032053280621767044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.08909104764461517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097756464034319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097756464034319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18700015544891357, + 0.11651105433702469, + 0.0777401477098465, + 0.16719476878643036, + 0.05491071939468384, + 0.12941712141036987, + 0.09631257504224777, + 0.024172021076083183, + 0.02920951135456562, + 0.11753188073635101 + ], + "credit_map": [ + 0.016151009127497673, + 0.00901445746421814, + 0.050949834287166595, + 0.018139423802495003, + 0.01046427246183157, + 0.013107453472912312, + 0.01876469887793064, + 0.050409022718667984, + 0.010489881038665771, + 0.010766824707388878, + 0.019694296643137932, + 0.021667294204235077, + 0.010013529099524021, + 0.0125684579834342, + 0.016635263338685036, + 0.014675524085760117, + 0.014903568662703037, + 0.004930674564093351, + 0.01096145436167717, + 0.009270214475691319, + 0.004244028590619564, + 0.005768878385424614, + 0.0168331079185009, + 0.010828232392668724, + 0.034733448177576065, + 0.014174140989780426, + 0.023625308647751808, + 0.020015232264995575, + 0.015076269395649433, + 0.015339554287493229, + 0.025780802592635155, + 0.018450047820806503, + 0.007279980927705765, + 0.003642150666564703, + 0.007410584948956966, + 0.004991380963474512, + 0.0064452593214809895, + 0.009010071866214275, + 0.011039625853300095, + 0.005091671831905842, + 0.015493846498429775, + 0.014226405881345272, + 0.014944243244826794, + 0.01579013466835022, + 0.02172281965613365, + 0.016545085236430168, + 0.017582235857844353, + 0.013112367130815983, + 0.011192538775503635, + 0.01193916518241167, + 0.011355736292898655, + 0.008936013095080853, + 0.009515911340713501, + 0.01289050467312336, + 0.018939752131700516, + 0.011542967520654202, + 0.0032502063550055027, + 0.0023630578070878983, + 0.0022144054528325796, + 0.0018990568351000547, + 0.004090338014066219, + 0.002814013510942459, + 0.004818182438611984, + 0.0027227650862187147, + 0.005337308160960674, + 0.0024936033878475428, + 0.002307260176166892, + 0.0022734508384019136, + 0.002560233697295189, + 0.004193922039121389, + 0.0075350599363446236, + 0.0025086761452257633, + 0.009592474438250065, + 0.01122819073498249, + 0.013023684732615948, + 0.015615309588611126, + 0.020854469388723373, + 0.011477070860564709, + 0.021016208454966545, + 0.01472449116408825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.004097756464034319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810429945588112 + ], + "opposite_logprob_tokens": [ + -3.288104295730591 + ], + "logprob_reward": -0.03810429945588112, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810429945588112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16177679598331451, + 0.24386635422706604, + 0.05951233580708504, + 0.29097747802734375, + 0.053661104291677475, + 0.04938068613409996, + 0.05050697550177574, + 0.02592618763446808, + 0.03810221329331398, + 0.02628980576992035 + ], + "credit_map": [ + 0.01246472168713808, + 0.009762173518538475, + 0.04239260032773018, + 0.017083823680877686, + 0.010603641159832478, + 0.011397619731724262, + 0.016892895102500916, + 0.04117933288216591, + 0.027511179447174072, + 0.02388722635805607, + 0.037217892706394196, + 0.035402677953243256, + 0.021452032029628754, + 0.03280244767665863, + 0.03577500209212303, + 0.029817920178174973, + 0.014292797073721886, + 0.004556911997497082, + 0.008348168805241585, + 0.006793365813791752, + 0.003664659336209297, + 0.005724228452891111, + 0.00923524983227253, + 0.006896961480379105, + 0.07186166197061539, + 0.025392239913344383, + 0.03969157487154007, + 0.030550934374332428, + 0.028222009539604187, + 0.02695712074637413, + 0.03775618225336075, + 0.030545774847269058, + 0.005568450782448053, + 0.004440816584974527, + 0.007260676473379135, + 0.004300860688090324, + 0.004423303063958883, + 0.007656189613044262, + 0.009269516915082932, + 0.010741299018263817, + 0.007356000132858753, + 0.005160072818398476, + 0.006163136102259159, + 0.006383931264281273, + 0.006882709916681051, + 0.00554996682330966, + 0.0062391795217990875, + 0.005645694211125374, + 0.006730614695698023, + 0.0062982793897390366, + 0.005302932113409042, + 0.00435751723125577, + 0.0054761855863034725, + 0.007574856746941805, + 0.008989744819700718, + 0.005776851437985897, + 0.003382169408723712, + 0.0019552172161638737, + 0.002248564036563039, + 0.0018723008688539267, + 0.002780552487820387, + 0.003466786816716194, + 0.00553301302716136, + 0.004687587730586529, + 0.006376899313181639, + 0.002704978920519352, + 0.0033460818231105804, + 0.0040901838801801205, + 0.002955188974738121, + 0.004548061639070511, + 0.010797983035445213, + 0.003282837802544236, + 0.0029811502899974585, + 0.002788675483316183, + 0.0039210692048072815, + 0.0021802803967148066, + 0.0023590344935655594, + 0.0028705098666250706, + 0.004497441928833723, + 0.00469164690002799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.03810429945588112 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977869238704443 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.00977869238704443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977869238704443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33643069863319397, + 0.046421777456998825, + 0.04117919132113457, + 0.157099649310112, + 0.148850679397583, + 0.05112624540925026, + 0.09591250121593475, + 0.01518716849386692, + 0.009434947744011879, + 0.098357193171978 + ], + "credit_map": [ + 0.022578757256269455, + 0.015303987078368664, + 0.1129777655005455, + 0.024104613810777664, + 0.012593072839081287, + 0.022235870361328125, + 0.01877051219344139, + 0.10786612331867218, + 0.006103456486016512, + 0.0044643450528383255, + 0.007220633327960968, + 0.005754172336310148, + 0.0045852987095713615, + 0.004395606461912394, + 0.007964445278048515, + 0.005933815147727728, + 0.008094835095107555, + 0.0038801589980721474, + 0.005295102018862963, + 0.005574618466198444, + 0.003284937934949994, + 0.0031168812420219183, + 0.00683032488450408, + 0.0051023270934820175, + 0.04028088599443436, + 0.012316249310970306, + 0.02423112280666828, + 0.016103526577353477, + 0.015287419781088829, + 0.015573321841657162, + 0.01781510002911091, + 0.015492004342377186, + 0.023267854005098343, + 0.008251135237514973, + 0.017686938866972923, + 0.01145774032920599, + 0.012244261801242828, + 0.026654332876205444, + 0.018160946667194366, + 0.031127460300922394, + 0.007876669056713581, + 0.004738908726722002, + 0.00810034479945898, + 0.005956266075372696, + 0.003739288542419672, + 0.006894497200846672, + 0.007691280450671911, + 0.006128989160060883, + 0.010731399990618229, + 0.012439095415174961, + 0.011103079654276371, + 0.008795670233666897, + 0.00841525848954916, + 0.013812880963087082, + 0.019900962710380554, + 0.010714145377278328, + 0.002514536026865244, + 0.001226981170475483, + 0.001157918362878263, + 0.0009194901213049889, + 0.0017286272486671805, + 0.0036970246583223343, + 0.002246822463348508, + 0.0016957679763436317, + 0.0011870202142745256, + 0.0007550884038209915, + 0.00100555585231632, + 0.0011548487236723304, + 0.0005775652243755758, + 0.0010808368679136038, + 0.0019416924333199859, + 0.0017323397332802415, + 0.010328467935323715, + 0.00826259609311819, + 0.014309866353869438, + 0.007882385514676571, + 0.008768359199166298, + 0.012348407879471779, + 0.011354565620422363, + 0.025102535262703896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.00977869238704443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly and steadily tilting upward in sync with the subject's movements, captures a scene imbued with Eastern Zen. In the center of a mirror-like lake, shimmering in a shimmering, inky hue, the surface silently breaks, revealing a massive, divine serpent rising from within. Its scales, deep as obsidian, gleam with a moist, warm sheen under the soft, diffused skylight. As its ancient, majestic head and sturdy, long body ascend with steady composure, countless crystalline droplets, like diamonds, carefully slide along the edges of each smooth, dense scale, finally dripping silently back into the lake, stirring tiny ripples. Distant mountains loom in the mist, the scene awash in ink-like black, white, and gray. Only the serpent's golden, vertical pupils, gleaming with the light of ancient wisdom, imbue this tranquil and majestic scene with a mysterious vitality, creating a sense of ultimate tranquility and power.", + "seed": 25504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_8/index_5919_seed_25504_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06001094123348594, + "vlm_logprob": -0.06001094123348594, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.06001094123348594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20166504383087158 + ], + "opposite_logprob_tokens": [ + -1.8266650438308716 + ], + "logprob_reward": -0.20166504383087158, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.20166504383087158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2722269892692566, + 0.07702071219682693, + 0.06570224463939667, + 0.0987219288945198, + 0.07626115530729294, + 0.08303677290678024, + 0.07527507841587067, + 0.07532812654972076, + 0.05604124814271927, + 0.12038572877645493 + ], + "credit_map": [ + 0.015599059872329235, + 0.016874346882104874, + 0.016537537798285484, + 0.027897559106349945, + 0.01545486506074667, + 0.025151345878839493, + 0.10074428468942642, + 0.053968001157045364, + 0.010239675641059875, + 0.011506130918860435, + 0.0077782562002539635, + 0.007555610965937376, + 0.00719155790284276, + 0.010441716760396957, + 0.011219999752938747, + 0.011087772436439991, + 0.00921258982270956, + 0.006435034796595573, + 0.010107570327818394, + 0.005314448848366737, + 0.006762874312698841, + 0.01071999128907919, + 0.0073067666962742805, + 0.009842969477176666, + 0.015324316918849945, + 0.00843186117708683, + 0.016695711761713028, + 0.005674257874488831, + 0.008548730053007603, + 0.018097108229994774, + 0.008066526614129543, + 0.017883431166410446, + 0.008767149411141872, + 0.009869157336652279, + 0.0116062480956316, + 0.008774260058999062, + 0.0049699884839355946, + 0.007155758328735828, + 0.011556351557374, + 0.013562249019742012, + 0.010228303261101246, + 0.0076294755563139915, + 0.010227025486528873, + 0.009338091127574444, + 0.007275574840605259, + 0.0175508800894022, + 0.009674926288425922, + 0.011112503707408905, + 0.009741534478962421, + 0.012445002794265747, + 0.007147165946662426, + 0.00913422740995884, + 0.0068419324234128, + 0.009667912498116493, + 0.01194506511092186, + 0.008352247066795826, + 0.008301055990159512, + 0.011651960201561451, + 0.008659680373966694, + 0.0057744174264371395, + 0.008996225893497467, + 0.009274096228182316, + 0.007287957239896059, + 0.015382740646600723, + 0.003931254614144564, + 0.0038822086062282324, + 0.015867315232753754, + 0.0021806503646075726, + 0.00811025220900774, + 0.015399966388940811, + 0.0030271399300545454, + 0.003642462892457843, + 0.02856534719467163, + 0.005128371994942427, + 0.039175428450107574, + 0.005193360149860382, + 0.00850688200443983, + 0.020049363374710083, + 0.004926234483718872, + 0.008840744383633137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.20166504383087158 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005137950181961 + ], + "opposite_logprob_tokens": [ + -2.695051431655884 + ], + "logprob_reward": -0.07005137950181961, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005137950181961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33752161264419556, + 0.025172347202897072, + 0.04981943592429161, + 0.02952762320637703, + 0.032458800822496414, + 0.0839226171374321, + 0.0365263931453228, + 0.028919868171215057, + 0.09520535171031952, + 0.2809258997440338 + ], + "credit_map": [ + 0.0129123879596591, + 0.007452304940670729, + 0.017234042286872864, + 0.032799724489450455, + 0.014528349973261356, + 0.017553308978676796, + 0.1495821177959442, + 0.0854593813419342, + 0.002572764176875353, + 0.002630288014188409, + 0.0027644559741020203, + 0.0024885982275009155, + 0.002672818722203374, + 0.004498299676924944, + 0.003391973441466689, + 0.004153147339820862, + 0.004743614234030247, + 0.004815293475985527, + 0.00769491121172905, + 0.005334271118044853, + 0.004849759861826897, + 0.007278712000697851, + 0.00772857666015625, + 0.007374296896159649, + 0.0037109090480953455, + 0.0031121568754315376, + 0.004421682097017765, + 0.0028621938545256853, + 0.0039869798347353935, + 0.0027078038547188044, + 0.0029127164743840694, + 0.00581318186596036, + 0.0028446451760828495, + 0.0036157886497676373, + 0.0054584648460149765, + 0.005845579318702221, + 0.0030295115429908037, + 0.003262076759710908, + 0.003801262704655528, + 0.004601473454385996, + 0.00717750983312726, + 0.0057856738567352295, + 0.005733238998800516, + 0.01251274161040783, + 0.009721602313220501, + 0.014476713724434376, + 0.013140535913407803, + 0.015374603681266308, + 0.002845081267878413, + 0.00636697094887495, + 0.002762990538030863, + 0.004286894574761391, + 0.004316085949540138, + 0.007229451555758715, + 0.004780394956469536, + 0.003938524518162012, + 0.005982087459415197, + 0.0016404205234721303, + 0.002397858304902911, + 0.0012819691328331828, + 0.002111775102093816, + 0.00991063378751278, + 0.0025438666343688965, + 0.0030512588564306498, + 0.012103858403861523, + 0.002533863065764308, + 0.028394600376486778, + 0.0025171812158077955, + 0.0031758081167936325, + 0.04059910774230957, + 0.001990692922845483, + 0.0038902410306036472, + 0.08614092320203781, + 0.013701953925192356, + 0.12273487448692322, + 0.010283585637807846, + 0.007574350107461214, + 0.01625642366707325, + 0.011540804989635944, + 0.012693007476627827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.07005137950181961 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490753773599863 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.002490753773599863, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490753773599863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3129670023918152, + 0.066317617893219, + 0.015708444640040398, + 0.031655363738536835, + 0.06585043668746948, + 0.03266020864248276, + 0.11260980367660522, + 0.13115531206130981, + 0.09459947049617767, + 0.1364763379096985 + ], + "credit_map": [ + 0.011925220489501953, + 0.009090719744563103, + 0.015903254970908165, + 0.035585399717092514, + 0.013349299319088459, + 0.017727846279740334, + 0.1289605349302292, + 0.08042475581169128, + 0.008162455633282661, + 0.006543722935020924, + 0.007247387431561947, + 0.005728586111217737, + 0.00783323124051094, + 0.012370969168841839, + 0.00871024839580059, + 0.009721013717353344, + 0.0017380517674610019, + 0.001310387160629034, + 0.0022302973084151745, + 0.002064853673800826, + 0.001554880989715457, + 0.0022505184169858694, + 0.0020032930187880993, + 0.002556162653490901, + 0.0032660842407494783, + 0.0026525205466896296, + 0.00544442143291235, + 0.0032123690471053123, + 0.003836692776530981, + 0.0032931016758084297, + 0.00552231352776289, + 0.004427859093993902, + 0.006106387823820114, + 0.011598842218518257, + 0.00923470314592123, + 0.007460223976522684, + 0.006106253247708082, + 0.008552548475563526, + 0.0059270537458360195, + 0.010864418931305408, + 0.00371706485748291, + 0.002865241141989827, + 0.0031802095472812653, + 0.005026056431233883, + 0.004460130352526903, + 0.005556159652769566, + 0.0036054139491170645, + 0.004249934107065201, + 0.010758692398667336, + 0.02125079743564129, + 0.010026529431343079, + 0.014331785961985588, + 0.012266309931874275, + 0.01627270318567753, + 0.01643524132668972, + 0.011267736554145813, + 0.022181322798132896, + 0.008586051873862743, + 0.007785801775753498, + 0.004273174796253443, + 0.00972365029156208, + 0.05109759047627449, + 0.009047423489391804, + 0.018460286781191826, + 0.008492526598274708, + 0.0036384426057338715, + 0.029565652832388878, + 0.002912515075877309, + 0.007468406576663256, + 0.03509550541639328, + 0.0024835446383804083, + 0.004942876286804676, + 0.041790176182985306, + 0.00715279346331954, + 0.05934812128543854, + 0.005683119874447584, + 0.005201918072998524, + 0.0060605150647461414, + 0.005684248637408018, + 0.005555446725338697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.002490753773599863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160688366740942 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160688366740942, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160688366740942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10499163717031479, + 0.03970308229327202, + 0.022536255419254303, + 0.02464381977915764, + 0.03582632914185524, + 0.027297915890812874, + 0.4645557403564453, + 0.023324813693761826, + 0.05607152357697487, + 0.20104894042015076 + ], + "credit_map": [ + 0.008052450604736805, + 0.007762228604406118, + 0.010266751050949097, + 0.010618990287184715, + 0.011581582017242908, + 0.013037213124334812, + 0.027119727805256844, + 0.016552679240703583, + 0.0049174390733242035, + 0.004631733521819115, + 0.0039315493777394295, + 0.004160105716437101, + 0.004588183481246233, + 0.005714780185371637, + 0.005460112821310759, + 0.006299172528088093, + 0.0029197060503065586, + 0.0022309510968625546, + 0.003105913521721959, + 0.002504167379811406, + 0.002624451881274581, + 0.0032961855176836252, + 0.0025431569665670395, + 0.0033117206767201424, + 0.002486514626070857, + 0.002257418120279908, + 0.004095365293323994, + 0.0017502026166766882, + 0.004575325176119804, + 0.0024568596854805946, + 0.0029278709553182125, + 0.004094258416444063, + 0.003138836706057191, + 0.005584846716374159, + 0.004906813148409128, + 0.004530190024524927, + 0.002664990955963731, + 0.0034548596013337374, + 0.005354818422347307, + 0.006190968677401543, + 0.0024788049049675465, + 0.002056267112493515, + 0.0025013084523379803, + 0.003538725897669792, + 0.003494373057037592, + 0.004884927533566952, + 0.0030903955921530724, + 0.005253110080957413, + 0.042721424251794815, + 0.08685775846242905, + 0.040258683264255524, + 0.05844307690858841, + 0.05216052383184433, + 0.07049819082021713, + 0.06564141809940338, + 0.04797462746500969, + 0.002076376462355256, + 0.0020826724357903004, + 0.004368653986603022, + 0.0012054763501510024, + 0.002500317757949233, + 0.00403140764683485, + 0.0016579569783061743, + 0.005401950795203447, + 0.007499425671994686, + 0.0027326380368322134, + 0.011747069656848907, + 0.0019519953057169914, + 0.005439142230898142, + 0.021189967170357704, + 0.0016666946467012167, + 0.003844584571197629, + 0.056236691772937775, + 0.009364061988890171, + 0.08127103745937347, + 0.009022080339491367, + 0.008229055441915989, + 0.020241910591721535, + 0.007993869483470917, + 0.008690210990607738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0160688366740942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward in sync with the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive serpent with obsidian scales rising from a mirror-like lake in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless crystalline droplets sliding along the serpent's scales and dripping back into the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are distant mountains visible in the mist with an ink-like black, white, and gray color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977869238704443 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.00977869238704443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977869238704443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5281041860580444, + 0.013824143446981907, + 0.013184979557991028, + 0.012630786746740341, + 0.029559321701526642, + 0.027870703488588333, + 0.03280363231897354, + 0.21874910593032837, + 0.0355723537504673, + 0.08770078420639038 + ], + "credit_map": [ + 0.020443420857191086, + 0.02020137570798397, + 0.020064737647771835, + 0.051144540309906006, + 0.027026502415537834, + 0.030184805393218994, + 0.21079690754413605, + 0.14824189245700836, + 0.0018100934103131294, + 0.001775321550667286, + 0.0015298972139135003, + 0.00139808829408139, + 0.001312039908953011, + 0.0018528529908508062, + 0.0021058288402855396, + 0.0020400206558406353, + 0.0013193782651796937, + 0.001016726833768189, + 0.0021741273812949657, + 0.0016213710187003016, + 0.0013691583881154656, + 0.002445915946736932, + 0.0017959445249289274, + 0.0014423576649278402, + 0.0012922824826091528, + 0.0012607675744220614, + 0.002319634659215808, + 0.0012338263913989067, + 0.0016364145558327436, + 0.0015042354352772236, + 0.0014999678824096918, + 0.0018836577655747533, + 0.00416834419593215, + 0.0041386280208826065, + 0.0038888645358383656, + 0.003274936694651842, + 0.0025993059389293194, + 0.003388090990483761, + 0.0031712190248072147, + 0.004929934628307819, + 0.003676532069221139, + 0.0027842277195304632, + 0.004063270054757595, + 0.004942620173096657, + 0.00266664638184011, + 0.003309851512312889, + 0.002994030714035034, + 0.0034335246309638023, + 0.003991527482867241, + 0.005412349943071604, + 0.0034874870907515287, + 0.004160990938544273, + 0.0030788432341068983, + 0.0033600525930523872, + 0.005567420274019241, + 0.0037449586670845747, + 0.029789382591843605, + 0.007290371228009462, + 0.016255512833595276, + 0.004779194947332144, + 0.015866823494434357, + 0.08437347412109375, + 0.012537816539406776, + 0.04785653576254845, + 0.005442604422569275, + 0.0011786652030423284, + 0.008854029700160027, + 0.0009534425917081535, + 0.0013871240662410855, + 0.015604771673679352, + 0.0007490974967367947, + 0.0014026195276528597, + 0.027376197278499603, + 0.004382062237709761, + 0.03855694830417633, + 0.003306427737697959, + 0.0025360698346048594, + 0.0041869208216667175, + 0.0036295109894126654, + 0.003726649098098278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.00977869238704443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_9/index_tensor([7057, 7057])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_9/index_tensor([7057, 7057])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..62cd812166885422c510c2c7c5ed6483cf0fcc6a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_9/index_tensor([7057, 7057])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, filmed from a slightly upward angle, captures the erosion and eventual collapse of a muddy trench edge in a calm, documentary style. In the image, icy rain falls densely from a gloomy, leaden sky, lashing the trench's damp, dark brown earthen walls. This wall, a mixture of tiny stones and rotting grass roots, is crisscrossed with tiny channels carved by erosion, through which muddy water flows. The camera focuses on the trench's crumbling upper edge, where rainwater continuously seeps in, carrying away tiny particles of earth. Initially, a few small pieces of soil loosen and fall, before a distinct crack quietly appears in the earthwork and rapidly widens. Finally, driven by gravity and rain, a massive, soaked mass of earth, unsupported, slid down silently and heavily, crashing into the muddy water at the bottom of the trench, splashing a foul spray and leaving a new, jagged gap in the original sidewall. The entire scene is filmed under even but lifeless diffuse light, with a dominant palette of dark brown, ochre, and leaden gray, creating an objective, desolate, and oppressive atmosphere.", + "seed": 19847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_9/index_7057_seed_19847_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18934326607268304, + "vlm_logprob": -0.18934326607268304, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.18934326607268304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08741869777441025 + ], + "opposite_logprob_tokens": [ + -2.587418794631958 + ], + "logprob_reward": -0.08741869777441025, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08741869777441025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09508736431598663, + 0.04849709942936897, + 0.02073388174176216, + 0.07291408628225327, + 0.05698012560606003, + 0.5448163151741028, + 0.06983190774917603, + 0.031105324625968933, + 0.03540157899260521, + 0.02463231049478054 + ], + "credit_map": [ + 0.006035666912794113, + 0.01206404622644186, + 0.013031665235757828, + 0.031096389517188072, + 0.005936179775744677, + 0.00643086526542902, + 0.011918250471353531, + 0.00857430137693882, + 0.004239770583808422, + 0.005926347337663174, + 0.00894220918416977, + 0.008639022707939148, + 0.0052223666571080685, + 0.003873368026688695, + 0.007035046350210905, + 0.0046189697459340096, + 0.00195885868743062, + 0.002666776767000556, + 0.002990501234307885, + 0.00403241440653801, + 0.002062278101220727, + 0.0016782813472673297, + 0.0027924594469368458, + 0.0025523114018142223, + 0.004957069642841816, + 0.014580704271793365, + 0.023720571771264076, + 0.008012098260223866, + 0.004353904165327549, + 0.003192820819094777, + 0.008712084032595158, + 0.00538483215495944, + 0.0024460472632199526, + 0.0055693527683615685, + 0.014829731546342373, + 0.011783131398260593, + 0.004836326465010643, + 0.0024405610747635365, + 0.009464118629693985, + 0.005610855761915445, + 0.014409280382096767, + 0.01377515122294426, + 0.1878701150417328, + 0.23538391292095184, + 0.007488813251256943, + 0.004221820738166571, + 0.04958074912428856, + 0.03208647295832634, + 0.004473160952329636, + 0.005839101504534483, + 0.013880861923098564, + 0.016478920355439186, + 0.006032972596585751, + 0.005773372016847134, + 0.01060302834957838, + 0.00675048865377903, + 0.0021141732577234507, + 0.004429278429597616, + 0.007363070268183947, + 0.0068532987497746944, + 0.0024755268823355436, + 0.0018779552774503827, + 0.0026619373820722103, + 0.0033300858922302723, + 0.0036196119617670774, + 0.006946878042072058, + 0.003643069416284561, + 0.004774465691298246, + 0.0030270549468696117, + 0.0032652602531015873, + 0.00488035473972559, + 0.005244882311671972, + 0.0029053068719804287, + 0.0028125110547989607, + 0.002768290229141712, + 0.0033631883561611176, + 0.0027269944548606873, + 0.00397708173841238, + 0.0025729842018336058, + 0.0035059559158980846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.08741869777441025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060834877192974 + ], + "opposite_logprob_tokens": [ + -3.895608425140381 + ], + "logprob_reward": -0.02060834877192974, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060834877192974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15061698853969574, + 0.04925746098160744, + 0.030781520530581474, + 0.050528477877378464, + 0.0444609597325325, + 0.4664812386035919, + 0.060657452791929245, + 0.04802187532186508, + 0.035052340477705, + 0.064141646027565 + ], + "credit_map": [ + 0.006897319108247757, + 0.012972909025847912, + 0.03513713181018829, + 0.020442524924874306, + 0.008900964632630348, + 0.013106361031532288, + 0.039823248982429504, + 0.013336531817913055, + 0.0038308631628751755, + 0.00693671777844429, + 0.012487431988120079, + 0.0063617718406021595, + 0.004660313483327627, + 0.0032709173392504454, + 0.005808379501104355, + 0.005901068449020386, + 0.0023322233464568853, + 0.003162139793857932, + 0.0061840759590268135, + 0.006160247605293989, + 0.0025132885202765465, + 0.0029521777760237455, + 0.004133876878768206, + 0.003343492280691862, + 0.003426360199227929, + 0.007763533387333155, + 0.007892143912613392, + 0.01234113983809948, + 0.004835283849388361, + 0.0035193595103919506, + 0.006221561226993799, + 0.004529094789177179, + 0.0021563421469181776, + 0.0034221773967146873, + 0.008426687680184841, + 0.014696410857141018, + 0.0031793343368917704, + 0.0028647324070334435, + 0.0057709431275725365, + 0.003944335971027613, + 0.012528330087661743, + 0.016644276678562164, + 0.15393297374248505, + 0.19953052699565887, + 0.008309705182909966, + 0.0042511895298957825, + 0.04374157637357712, + 0.027542700991034508, + 0.003398729022592306, + 0.006826706230640411, + 0.013679266907274723, + 0.014729954302310944, + 0.004414116032421589, + 0.0037783682346343994, + 0.009055639617145061, + 0.004774675238877535, + 0.004749990068376064, + 0.010537337511777878, + 0.007023604121059179, + 0.007450632750988007, + 0.006191427819430828, + 0.0034597422927618027, + 0.004197671078145504, + 0.0044114734046161175, + 0.0045262230560183525, + 0.009306030347943306, + 0.0034454353153705597, + 0.004178899805992842, + 0.00355934863910079, + 0.0023887979332357645, + 0.0029819877818226814, + 0.0046656192280352116, + 0.005727007519453764, + 0.005073815584182739, + 0.00804170686751604, + 0.005940582137554884, + 0.0045695132575929165, + 0.007472019176930189, + 0.008922304026782513, + 0.018394701182842255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.02060834877192974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31352943181991577 + ], + "opposite_logprob_tokens": [ + -1.3135294914245605 + ], + "logprob_reward": -0.31352943181991577, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31352943181991577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06742861866950989, + 0.02261734940111637, + 0.03453736752271652, + 0.08251182734966278, + 0.0510399155318737, + 0.6271778345108032, + 0.02937147580087185, + 0.03407995030283928, + 0.018719147890806198, + 0.03251653164625168 + ], + "credit_map": [ + 0.005842188838869333, + 0.013621736317873001, + 0.005924566183239222, + 0.009147601202130318, + 0.00542967626824975, + 0.0038294296246021986, + 0.015354047529399395, + 0.008279369212687016, + 0.002428804524242878, + 0.0026927173603326082, + 0.0029770517721772194, + 0.0036344125401228666, + 0.001833076006732881, + 0.001948817982338369, + 0.003186310874298215, + 0.00391615554690361, + 0.002824606141075492, + 0.008061831817030907, + 0.010367542505264282, + 0.0032693713437765837, + 0.0024312103632837534, + 0.0015737171052023768, + 0.003572839079424739, + 0.0024362504482269287, + 0.0056561920791864395, + 0.021151237189769745, + 0.02745828963816166, + 0.008853757753968239, + 0.004582520108669996, + 0.0025171723682433367, + 0.007780212443321943, + 0.004512439016252756, + 0.0030141486786305904, + 0.008865661919116974, + 0.009736933745443821, + 0.00787079893052578, + 0.0034468562807887793, + 0.004165963269770145, + 0.008651294745504856, + 0.005288257263600826, + 0.01659769006073475, + 0.01856960728764534, + 0.213849738240242, + 0.2688717544078827, + 0.009773731231689453, + 0.0052871801890432835, + 0.05726197361946106, + 0.036966122686862946, + 0.0014806080143898726, + 0.0022599773947149515, + 0.0069750831462442875, + 0.00808521918952465, + 0.001811658381484449, + 0.0017758164322003722, + 0.0047639599069952965, + 0.0022191517055034637, + 0.003985145129263401, + 0.00861607026308775, + 0.005124358460307121, + 0.004434446804225445, + 0.004396908450871706, + 0.002079318044707179, + 0.0026502786204218864, + 0.0027934208046644926, + 0.002407410414889455, + 0.004349410068243742, + 0.0015152099076658487, + 0.0023994750808924437, + 0.0020201883744448423, + 0.0014420421794056892, + 0.0017788887489587069, + 0.002806522650644183, + 0.0025234543718397617, + 0.0020712947007268667, + 0.003955662250518799, + 0.0035199907142668962, + 0.0021895733661949635, + 0.003970090765506029, + 0.004582661669701338, + 0.0097038010135293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.31352943181991577 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234434008598328 + ], + "opposite_logprob_tokens": [ + -0.8984434008598328 + ], + "logprob_reward": -0.5234434008598328, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234434008598328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1984553188085556, + 0.033731769770383835, + 0.027475228533148766, + 0.1518031358718872, + 0.045768123120069504, + 0.2383851706981659, + 0.04335087910294533, + 0.08122703433036804, + 0.03549094870686531, + 0.14431235194206238 + ], + "credit_map": [ + 0.014671231620013714, + 0.04471903294324875, + 0.06264568120241165, + 0.024251751601696014, + 0.009958159178495407, + 0.009816833771765232, + 0.016200372949242592, + 0.01619226112961769, + 0.0029971033800393343, + 0.004194287583231926, + 0.0044098300859332085, + 0.006816715467721224, + 0.002872155047953129, + 0.0032941987738013268, + 0.004567679483443499, + 0.0045797983184456825, + 0.0021409224718809128, + 0.00529507314786315, + 0.007187539711594582, + 0.004530909936875105, + 0.0014681007014587522, + 0.0016414672136306763, + 0.002840308239683509, + 0.0023709069937467575, + 0.0107513926923275, + 0.03947058692574501, + 0.04903483763337135, + 0.017408953979611397, + 0.008975774981081486, + 0.004794156178832054, + 0.013263262808322906, + 0.008104177191853523, + 0.002636362100020051, + 0.008947660215198994, + 0.007874058559536934, + 0.009789779782295227, + 0.002770618535578251, + 0.0031696325168013573, + 0.006816151551902294, + 0.0037638568319380283, + 0.006243247073143721, + 0.00541188009083271, + 0.08580254763364792, + 0.100849449634552, + 0.003654316533356905, + 0.0020326324738562107, + 0.02030363865196705, + 0.014087452553212643, + 0.0021764994598925114, + 0.0037997234612703323, + 0.009114270098507404, + 0.012275165878236294, + 0.0028193690814077854, + 0.0029210124630481005, + 0.006783136632293463, + 0.003461700165644288, + 0.00745433010160923, + 0.017975036054849625, + 0.013231423683464527, + 0.012148031033575535, + 0.010811466723680496, + 0.005178845953196287, + 0.006863846443593502, + 0.007564053870737553, + 0.0032153124921023846, + 0.003633335465565324, + 0.0038896561600267887, + 0.006200506817549467, + 0.0034864521585404873, + 0.0043616886250674725, + 0.0051694829016923904, + 0.005534512456506491, + 0.010347571223974228, + 0.010429768823087215, + 0.01824553683400154, + 0.01736987940967083, + 0.010050226002931595, + 0.01872035674750805, + 0.020033815875649452, + 0.03911519795656204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.5234434008598328 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017164511373266578 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.0017164511373266578, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017164511373266578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3181650936603546, + 0.04150605574250221, + 0.018206855282187462, + 0.13708363473415375, + 0.041984930634498596, + 0.18966583907604218, + 0.057503681629896164, + 0.026254067197442055, + 0.04349216818809509, + 0.12613770365715027 + ], + "credit_map": [ + 0.02932395227253437, + 0.07488691061735153, + 0.08037194609642029, + 0.03424837067723274, + 0.018548494204878807, + 0.019187457859516144, + 0.03066072426736355, + 0.030937209725379944, + 0.004045277368277311, + 0.006586598698049784, + 0.007178848143666983, + 0.007396703120321035, + 0.003190009854733944, + 0.0031563807278871536, + 0.005103476345539093, + 0.004848763346672058, + 0.0012776260264217854, + 0.0018178358441218734, + 0.0034069481771439314, + 0.0031221171375364065, + 0.0027416900265961885, + 0.00195682467892766, + 0.0021998423617333174, + 0.001683971844613552, + 0.009606538340449333, + 0.034659940749406815, + 0.045492012053728104, + 0.014821185730397701, + 0.007308977656066418, + 0.00390980951488018, + 0.013725134544074535, + 0.007560031954199076, + 0.002162681892514229, + 0.0034110061824321747, + 0.004945837892591953, + 0.013058244250714779, + 0.003340309951454401, + 0.003081541508436203, + 0.008141854777932167, + 0.0038434546440839767, + 0.005654222797602415, + 0.006416242104023695, + 0.06542349606752396, + 0.07912823557853699, + 0.0035851008724421263, + 0.0020357142202556133, + 0.016407346352934837, + 0.011015479452908039, + 0.0048374757170677185, + 0.011498830281198025, + 0.011164438910782337, + 0.00756043428555131, + 0.006660100072622299, + 0.004125696606934071, + 0.006341759581118822, + 0.005314946174621582, + 0.0025547388941049576, + 0.0043062069453299046, + 0.00305429520085454, + 0.003937198780477047, + 0.0029391066636890173, + 0.0023337299935519695, + 0.004012241028249264, + 0.003116549225524068, + 0.004120042081922293, + 0.008625862188637257, + 0.005428527947515249, + 0.006438491400331259, + 0.003961218986660242, + 0.004091694951057434, + 0.00438531581312418, + 0.006441013887524605, + 0.008922520093619823, + 0.0072595831006765366, + 0.015439560636878014, + 0.01711186021566391, + 0.009406921453773975, + 0.016245080158114433, + 0.017774831503629684, + 0.03397735580801964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0017164511373266578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, filmed from a slightly upward angle, captures the erosion and eventual collapse of a muddy trench edge in a calm, documentary style. In the image, icy rain falls densely from a gloomy, leaden sky, lashing the trench's damp, dark brown earthen walls. This wall, a mixture of tiny stones and rotting grass roots, is crisscrossed with tiny channels carved by erosion, through which muddy water flows. The camera focuses on the trench's crumbling upper edge, where rainwater continuously seeps in, carrying away tiny particles of earth. Initially, a few small pieces of soil loosen and fall, before a distinct crack quietly appears in the earthwork and rapidly widens. Finally, driven by gravity and rain, a massive, soaked mass of earth, unsupported, slid down silently and heavily, crashing into the muddy water at the bottom of the trench, splashing a foul spray and leaving a new, jagged gap in the original sidewall. The entire scene is filmed under even but lifeless diffuse light, with a dominant palette of dark brown, ochre, and leaden gray, creating an objective, desolate, and oppressive atmosphere.", + "seed": 19848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_9/index_7057_seed_19848_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2525483140721917, + "vlm_logprob": -0.2525483140721917, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.2525483140721917, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0795498639345169 + ], + "opposite_logprob_tokens": [ + -2.704549789428711 + ], + "logprob_reward": -0.0795498639345169, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0795498639345169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09593767672777176, + 0.042688097804784775, + 0.1085948720574379, + 0.04060177877545357, + 0.1607964187860489, + 0.07594465464353561, + 0.12257349491119385, + 0.21237318217754364, + 0.0487329475581646, + 0.09175682812929153 + ], + "credit_map": [ + 0.008320802822709084, + 0.0130808986723423, + 0.010741344653069973, + 0.014576874673366547, + 0.009523425251245499, + 0.011847402900457382, + 0.015434264205396175, + 0.012412665411829948, + 0.0040147025138139725, + 0.005473633296787739, + 0.005530394613742828, + 0.007283405400812626, + 0.004259126260876656, + 0.00451268395408988, + 0.007400159258395433, + 0.004213992971926928, + 0.010118050500750542, + 0.007312224246561527, + 0.016944676637649536, + 0.01815076358616352, + 0.011818261817097664, + 0.010605414398014545, + 0.02138165570795536, + 0.012263826094567776, + 0.0032777239102870226, + 0.00472303805872798, + 0.004385529085993767, + 0.010377454571425915, + 0.004296962637454271, + 0.002950480440631509, + 0.007395251654088497, + 0.0031953377183526754, + 0.013618622906506062, + 0.011544553562998772, + 0.029206976294517517, + 0.014058439992368221, + 0.01863350346684456, + 0.016565002501010895, + 0.027358798310160637, + 0.029810531064867973, + 0.005544313695281744, + 0.007821665145456791, + 0.018475817516446114, + 0.006137932185083628, + 0.008324231021106243, + 0.0052828118205070496, + 0.00756973959505558, + 0.016788147389888763, + 0.009320388548076153, + 0.009480861946940422, + 0.016892554238438606, + 0.017889903858304024, + 0.016485746949911118, + 0.015696588903665543, + 0.015440825372934341, + 0.021366626024246216, + 0.01654892973601818, + 0.01811634935438633, + 0.05248678848147392, + 0.016481444239616394, + 0.06082585081458092, + 0.014840940944850445, + 0.020394666120409966, + 0.012678205966949463, + 0.004072973504662514, + 0.005839881021529436, + 0.007185489404946566, + 0.006442144513130188, + 0.010418602265417576, + 0.005416210275143385, + 0.004030993673950434, + 0.005326651968061924, + 0.009635379537940025, + 0.00757310725748539, + 0.010118423029780388, + 0.02180841937661171, + 0.010698279365897179, + 0.00976923294365406, + 0.01380404643714428, + 0.0083499476313591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.0795498639345169 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029818454757332802 + ], + "opposite_logprob_tokens": [ + -3.529818534851074 + ], + "logprob_reward": -0.029818454757332802, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029818454757332802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1659039705991745, + 0.048476915806531906, + 0.026164380833506584, + 0.01866072602570057, + 0.0584736131131649, + 0.01928021013736725, + 0.05563370883464813, + 0.5388925671577454, + 0.03850652277469635, + 0.030007412657141685 + ], + "credit_map": [ + 0.016578687354922295, + 0.02021012455224991, + 0.022489555180072784, + 0.021286476403474808, + 0.0192281324416399, + 0.017610887065529823, + 0.02812405675649643, + 0.020376039668917656, + 0.0035760579630732536, + 0.005485885310918093, + 0.0073673599399626255, + 0.009195009246468544, + 0.004759335424751043, + 0.004324419889599085, + 0.008376666344702244, + 0.005392175633460283, + 0.0017028358997777104, + 0.0016095048049464822, + 0.0035712462849915028, + 0.005714838858693838, + 0.0037045939825475216, + 0.002819454064592719, + 0.004247153177857399, + 0.0027947518974542618, + 0.0015822482528164983, + 0.002193031134083867, + 0.002280413405969739, + 0.003735495265573263, + 0.0017689651576802135, + 0.001771023846231401, + 0.0034524064976722, + 0.001877141883596778, + 0.004706109873950481, + 0.0066556730307638645, + 0.0075148954056203365, + 0.006411396898329258, + 0.007067709229886532, + 0.006430594716221094, + 0.012075057253241539, + 0.007612168788909912, + 0.002644643886014819, + 0.0035671554505825043, + 0.002179575152695179, + 0.0019828155636787415, + 0.002982816658914089, + 0.0017379805212840438, + 0.0019220729591324925, + 0.0022631483152508736, + 0.00376765220426023, + 0.0037360216956585646, + 0.009353685192763805, + 0.008195354603230953, + 0.008054225705564022, + 0.006812619045376778, + 0.00613395543769002, + 0.009580189362168312, + 0.03941408172249794, + 0.050108056515455246, + 0.1432836353778839, + 0.04052099213004112, + 0.18001538515090942, + 0.03485942631959915, + 0.027284687384963036, + 0.023406222462654114, + 0.004183391109108925, + 0.003190506249666214, + 0.004810194484889507, + 0.006007086485624313, + 0.005594899877905846, + 0.005145308095961809, + 0.003927316982299089, + 0.00564781716093421, + 0.0022885240614414215, + 0.0016709999181330204, + 0.004068332724273205, + 0.0035983193665742874, + 0.0038869224954396486, + 0.004103777464479208, + 0.003924233373254538, + 0.00646629836410284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.029818454757332802 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983912467956543 + ], + "opposite_logprob_tokens": [ + -0.5233912467956543 + ], + "logprob_reward": -0.8983912467956543, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983912467956543, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2004101425409317, + 0.037972405552864075, + 0.026931511238217354, + 0.040765129029750824, + 0.021081531420350075, + 0.028703048825263977, + 0.020962676033377647, + 0.5762217044830322, + 0.03507300838828087, + 0.011878839693963528 + ], + "credit_map": [ + 0.01934424601495266, + 0.023992016911506653, + 0.02820274606347084, + 0.030405012890696526, + 0.0236381646245718, + 0.020181389525532722, + 0.031130006536841393, + 0.02351657673716545, + 0.002681395737454295, + 0.004483183845877647, + 0.005970287602394819, + 0.007543855346739292, + 0.0036528483033180237, + 0.003639528527855873, + 0.005660387221723795, + 0.0043409173376858234, + 0.0021848760079592466, + 0.0026911695022135973, + 0.003398898523300886, + 0.002856623148545623, + 0.0034676981158554554, + 0.003126302268356085, + 0.00573921762406826, + 0.0034667267464101315, + 0.003713623620569706, + 0.002917129546403885, + 0.00426002312451601, + 0.0046981447376310825, + 0.003519782330840826, + 0.006854257546365261, + 0.008374273777008057, + 0.0064278943464159966, + 0.002008916810154915, + 0.0014838053612038493, + 0.002446829341351986, + 0.003603373421356082, + 0.0034962266217917204, + 0.0023599714040756226, + 0.0037905347999185324, + 0.0018918734276667237, + 0.0036524534225463867, + 0.002301979111507535, + 0.006599687971174717, + 0.001983613008633256, + 0.003194586606696248, + 0.0024503024760633707, + 0.001872173510491848, + 0.0066482508555054665, + 0.0012957103317603469, + 0.0012792947236448526, + 0.004296220373362303, + 0.0025224429555237293, + 0.003303486155346036, + 0.0017523999558761716, + 0.0029636500403285027, + 0.003549472661688924, + 0.04148520529270172, + 0.053673699498176575, + 0.15574730932712555, + 0.04424572363495827, + 0.19200165569782257, + 0.0354299396276474, + 0.028897570446133614, + 0.024740582332015038, + 0.003527119755744934, + 0.003493404481559992, + 0.004370903130620718, + 0.0048527405597269535, + 0.006134600378572941, + 0.004581643268465996, + 0.003860397497192025, + 0.004252199083566666, + 0.0011253667762503028, + 0.0011325253872200847, + 0.0014536627568304539, + 0.001485512824729085, + 0.0015352466143667698, + 0.0018857683753594756, + 0.0016841021133586764, + 0.0015766555443406105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.8983912467956543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25216084718704224 + ], + "opposite_logprob_tokens": [ + -1.5021607875823975 + ], + "logprob_reward": -0.25216084718704224, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25216084718704224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08075380325317383, + 0.03313636779785156, + 0.126168891787529, + 0.03330369293689728, + 0.12974056601524353, + 0.03270625323057175, + 0.2117975950241089, + 0.3130475580692291, + 0.015810638666152954, + 0.023534651845693588 + ], + "credit_map": [ + 0.006129938643425703, + 0.009316207841038704, + 0.013344689272344112, + 0.013638417236506939, + 0.011978230439126492, + 0.009293514303863049, + 0.008763102814555168, + 0.008289700374007225, + 0.0023683637846261263, + 0.003944022580981255, + 0.003782011102885008, + 0.007178116589784622, + 0.004124165046960115, + 0.0032468740828335285, + 0.005109811667352915, + 0.0033830024767667055, + 0.011423860676586628, + 0.010519174858927727, + 0.016897670924663544, + 0.02356553077697754, + 0.013793561607599258, + 0.013742820359766483, + 0.0223468616604805, + 0.013879409059882164, + 0.0032966723665595055, + 0.003300974378362298, + 0.004074652213603258, + 0.007647302467375994, + 0.0030658103059977293, + 0.0031782910227775574, + 0.006102033890783787, + 0.0026379572227597237, + 0.009609179571270943, + 0.01620258204638958, + 0.017576517537236214, + 0.014286411926150322, + 0.01812347024679184, + 0.011561454273760319, + 0.02617638185620308, + 0.016204558312892914, + 0.002368890680372715, + 0.002767596859484911, + 0.0057578240521252155, + 0.003189678071066737, + 0.0033198129385709763, + 0.002463635290041566, + 0.00554376095533371, + 0.007295054383575916, + 0.012236008420586586, + 0.014422048814594746, + 0.03764151409268379, + 0.0281223151832819, + 0.030206751078367233, + 0.025334056466817856, + 0.025124557316303253, + 0.03871035575866699, + 0.022239282727241516, + 0.02963297814130783, + 0.08490965515375137, + 0.024159817025065422, + 0.10626926273107529, + 0.019607314839959145, + 0.013115913607180119, + 0.013113315217196941, + 0.0014902573311701417, + 0.001500505255535245, + 0.001644118339754641, + 0.0026918789371848106, + 0.002494308864697814, + 0.002322996500879526, + 0.0018536095740273595, + 0.0018129640957340598, + 0.0023884985130280256, + 0.0020040106028318405, + 0.0035276596900075674, + 0.0034798355773091316, + 0.003665162483230233, + 0.002732254797592759, + 0.0029603547882288694, + 0.002776875626295805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.25216084718704224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028211576864123344 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028211576864123344, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028211576864123344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12900951504707336, + 0.03245941177010536, + 0.03759116306900978, + 0.04083411023020744, + 0.05144554749131203, + 0.10962525755167007, + 0.15170732140541077, + 0.37726396322250366, + 0.025963034480810165, + 0.04410065710544586 + ], + "credit_map": [ + 0.009197071194648743, + 0.011312351562082767, + 0.01969955302774906, + 0.018379569053649902, + 0.018664920702576637, + 0.014960039407014847, + 0.02317909337580204, + 0.013616921380162239, + 0.004057705868035555, + 0.004087318200618029, + 0.0035947496071457863, + 0.004256583284586668, + 0.004249281715601683, + 0.0029640730936080217, + 0.004632684867829084, + 0.004617016762495041, + 0.0025404556654393673, + 0.005493289791047573, + 0.0062767076306045055, + 0.005131884943693876, + 0.004720428958535194, + 0.0033153798431158066, + 0.0065003749914467335, + 0.00361264287494123, + 0.0033608146477490664, + 0.005230979062616825, + 0.005137279629707336, + 0.007830438204109669, + 0.003509960137307644, + 0.004287686198949814, + 0.007690618745982647, + 0.003786336164921522, + 0.0051293522119522095, + 0.00810754019767046, + 0.005882402416318655, + 0.005898827221244574, + 0.006517705507576466, + 0.004410975147038698, + 0.00898970477283001, + 0.006509042344987392, + 0.006956063210964203, + 0.007031208369880915, + 0.03561359643936157, + 0.008125456981360912, + 0.005710245575755835, + 0.0037541422061622143, + 0.008476224727928638, + 0.0339583158493042, + 0.009487137198448181, + 0.01017045695334673, + 0.025167997926473618, + 0.02036464586853981, + 0.022183913737535477, + 0.018582701683044434, + 0.01887921802699566, + 0.026871249079704285, + 0.027074195444583893, + 0.03504182770848274, + 0.10085756331682205, + 0.028936611488461494, + 0.12393715977668762, + 0.023941131308674812, + 0.020512597635388374, + 0.01696288213133812, + 0.0026991586200892925, + 0.002532648853957653, + 0.0047342670150101185, + 0.00415900070220232, + 0.003887514816597104, + 0.0025960002094507217, + 0.00289523764513433, + 0.0024592073168605566, + 0.003634898690506816, + 0.003600821364670992, + 0.007654632441699505, + 0.006197821348905563, + 0.008691975846886635, + 0.004726271610707045, + 0.00439968379214406, + 0.005194555036723614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0028211576864123344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, filmed from a slightly upward angle, captures the erosion and eventual collapse of a muddy trench edge in a calm, documentary style. In the image, icy rain falls densely from a gloomy, leaden sky, lashing the trench's damp, dark brown earthen walls. This wall, a mixture of tiny stones and rotting grass roots, is crisscrossed with tiny channels carved by erosion, through which muddy water flows. The camera focuses on the trench's crumbling upper edge, where rainwater continuously seeps in, carrying away tiny particles of earth. Initially, a few small pieces of soil loosen and fall, before a distinct crack quietly appears in the earthwork and rapidly widens. Finally, driven by gravity and rain, a massive, soaked mass of earth, unsupported, slid down silently and heavily, crashing into the muddy water at the bottom of the trench, splashing a foul spray and leaving a new, jagged gap in the original sidewall. The entire scene is filmed under even but lifeless diffuse light, with a dominant palette of dark brown, ochre, and leaden gray, creating an objective, desolate, and oppressive atmosphere.", + "seed": 19849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_9/index_7057_seed_19849_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15153446374461055, + "vlm_logprob": -0.15153446374461055, + "vlm_yes_no_margin": 2.8500000119209288, + "vlm_token_credit": -0.15153446374461055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059549421072006226 + ], + "opposite_logprob_tokens": [ + -3.059549331665039 + ], + "logprob_reward": -0.059549421072006226, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.059549421072006226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06022477149963379, + 0.03852420672774315, + 0.10266623646020889, + 0.039732079952955246, + 0.12120387703180313, + 0.25966528058052063, + 0.11190769076347351, + 0.13273277878761292, + 0.04119338467717171, + 0.09214965999126434 + ], + "credit_map": [ + 0.00457006087526679, + 0.00503963278606534, + 0.007635090500116348, + 0.006822416093200445, + 0.007271004840731621, + 0.005482297390699387, + 0.013986553065478802, + 0.009417715482413769, + 0.004365512169897556, + 0.005698099732398987, + 0.0040259407833218575, + 0.004147696774452925, + 0.004680905491113663, + 0.0041468399576842785, + 0.005984236020594835, + 0.005474973004311323, + 0.007191716227680445, + 0.009690562263131142, + 0.00824022013694048, + 0.008988641202449799, + 0.007335871458053589, + 0.007365750148892403, + 0.028882605955004692, + 0.02497086301445961, + 0.0036035485100001097, + 0.004873230122029781, + 0.004020740743726492, + 0.005757857114076614, + 0.0035133035853505135, + 0.0033349930308759212, + 0.006341691128909588, + 0.008286713622510433, + 0.010204926133155823, + 0.004499516915529966, + 0.025328783318400383, + 0.010389064438641071, + 0.004719521850347519, + 0.0033870027400553226, + 0.040994349867105484, + 0.021680722013115883, + 0.023939665406942368, + 0.016840003430843353, + 0.04992348328232765, + 0.02347683720290661, + 0.013456705957651138, + 0.012356964871287346, + 0.07770020514726639, + 0.041971419006586075, + 0.006408607121556997, + 0.0042350259609520435, + 0.024064496159553528, + 0.009579367935657501, + 0.004151752218604088, + 0.002531972248107195, + 0.039575375616550446, + 0.021361093968153, + 0.011051640845835209, + 0.005855257622897625, + 0.03729772940278053, + 0.012202336452901363, + 0.006827265024185181, + 0.008651845157146454, + 0.013000261969864368, + 0.03784644976258278, + 0.0035360469482839108, + 0.002599542960524559, + 0.010087119415402412, + 0.004626758862286806, + 0.00312199373729527, + 0.004252355080097914, + 0.0035678159911185503, + 0.009401751682162285, + 0.008105193264782429, + 0.004017399158328772, + 0.027363238856196404, + 0.009793027304112911, + 0.0034527534153312445, + 0.005451797507703304, + 0.008525307290256023, + 0.025440948083996773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.059549421072006226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026376035064458847 + ], + "opposite_logprob_tokens": [ + -3.6513760089874268 + ], + "logprob_reward": -0.026376035064458847, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026376035064458847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17515483498573303, + 0.10409139096736908, + 0.03535507991909981, + 0.025503722950816154, + 0.059295959770679474, + 0.04276420921087265, + 0.04948895052075386, + 0.35335052013397217, + 0.09346243739128113, + 0.0615328848361969 + ], + "credit_map": [ + 0.014612543396651745, + 0.022647973150014877, + 0.02118191495537758, + 0.020216410979628563, + 0.026481198146939278, + 0.01362165529280901, + 0.029889533296227455, + 0.026503613218665123, + 0.005464359652251005, + 0.004816619213670492, + 0.009790276177227497, + 0.007761805783957243, + 0.0061380285769701, + 0.023888353258371353, + 0.03041696548461914, + 0.015814989805221558, + 0.00252146297134459, + 0.00297991163097322, + 0.006091064773499966, + 0.003434149781242013, + 0.0022552330046892166, + 0.0020980883855372667, + 0.009462024085223675, + 0.0065131476148962975, + 0.0032936204224824905, + 0.002169904764741659, + 0.00453361589461565, + 0.003774939337745309, + 0.0021748444996774197, + 0.0021840643603354692, + 0.0038713656831532717, + 0.0035013698507100344, + 0.005287314765155315, + 0.0027687482070177794, + 0.010969714261591434, + 0.005642754957079887, + 0.003028376027941704, + 0.002989239525049925, + 0.01839013211429119, + 0.010219682939350605, + 0.003905048593878746, + 0.0020628797356039286, + 0.009073355235159397, + 0.0038590682670474052, + 0.002340767066925764, + 0.002206383738666773, + 0.01275987084954977, + 0.006556838285177946, + 0.003815736621618271, + 0.005082352086901665, + 0.007723700720816851, + 0.004487757105380297, + 0.003999215085059404, + 0.004515497479587793, + 0.012048267759382725, + 0.007816425524652004, + 0.03014965169131756, + 0.012524084188044071, + 0.10374049097299576, + 0.029953913763165474, + 0.015568924136459827, + 0.020653149113059044, + 0.03205138444900513, + 0.1087089404463768, + 0.006853786762803793, + 0.0062626018188893795, + 0.021450089290738106, + 0.011502178385853767, + 0.006867558695375919, + 0.009291687980294228, + 0.010103223845362663, + 0.02113131247460842, + 0.0053154341876506805, + 0.004356652032583952, + 0.010952727869153023, + 0.008852846920490265, + 0.004722716752439737, + 0.0068468619138002396, + 0.008250382728874683, + 0.012235268019139767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.026376035064458847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871588110923767 + ], + "opposite_logprob_tokens": [ + -1.1371588706970215 + ], + "logprob_reward": -0.3871588110923767, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3871588110923767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28686386346817017, + 0.0373767726123333, + 0.0251841489225626, + 0.07527554035186768, + 0.032069649547338486, + 0.14173565804958344, + 0.07204531133174896, + 0.16824650764465332, + 0.06920892000198364, + 0.09199365973472595 + ], + "credit_map": [ + 0.0184656772762537, + 0.01670021191239357, + 0.03232410550117493, + 0.024600334465503693, + 0.01791210286319256, + 0.05717337131500244, + 0.0786857008934021, + 0.041002366691827774, + 0.002653103321790695, + 0.0033163316547870636, + 0.004137594718486071, + 0.0035111638717353344, + 0.003047901438549161, + 0.006098009645938873, + 0.008682087063789368, + 0.005930579733103514, + 0.0016433235723525286, + 0.002548422198742628, + 0.002484335098415613, + 0.0029935382772237062, + 0.0019185029668733478, + 0.002955935662612319, + 0.005624189507216215, + 0.005015901755541563, + 0.007571674883365631, + 0.006287482101470232, + 0.015095113776624203, + 0.008741497062146664, + 0.005185113288462162, + 0.005028336774557829, + 0.016963979229331017, + 0.010402346961200237, + 0.004037556238472462, + 0.0028120630886405706, + 0.00462613720446825, + 0.0036108277272433043, + 0.002430954948067665, + 0.0030810232274234295, + 0.004415512550622225, + 0.00705557269975543, + 0.014029267244040966, + 0.010323869995772839, + 0.019652778282761574, + 0.01711203344166279, + 0.009335295297205448, + 0.011447911150753498, + 0.03661156818270683, + 0.023222927004098892, + 0.00403388449922204, + 0.00251180911436677, + 0.01573542132973671, + 0.0060262419283390045, + 0.002507125725969672, + 0.0013438514433801174, + 0.025847764685750008, + 0.014039214700460434, + 0.013878940604627132, + 0.004604607820510864, + 0.048249151557683945, + 0.014799814671278, + 0.006191859021782875, + 0.008408292196691036, + 0.016839811578392982, + 0.055274032056331635, + 0.0054840086959302425, + 0.003965694922953844, + 0.0177321620285511, + 0.009503483772277832, + 0.004182768985629082, + 0.005376575514674187, + 0.006338809616863728, + 0.016625413671135902, + 0.008213155902922153, + 0.0037951820995658636, + 0.02855522371828556, + 0.00784078985452652, + 0.0038653474766761065, + 0.005856966599822044, + 0.008114594966173172, + 0.025752395391464233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.3871588110923767 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28139179944992065 + ], + "opposite_logprob_tokens": [ + -1.4063918590545654 + ], + "logprob_reward": -0.28139179944992065, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28139179944992065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10474175214767456, + 0.03524794802069664, + 0.02638396807014942, + 0.035977356135845184, + 0.026721084490418434, + 0.02891721948981285, + 0.14247485995292664, + 0.452509343624115, + 0.02654694952070713, + 0.12047946453094482 + ], + "credit_map": [ + 0.008996409364044666, + 0.01353264506906271, + 0.014375117607414722, + 0.012170536443591118, + 0.015598439611494541, + 0.014580264687538147, + 0.012123570777475834, + 0.013364764861762524, + 0.0033294979948550463, + 0.004059066530317068, + 0.003937830217182636, + 0.003162900684401393, + 0.003197311656549573, + 0.0047107296995818615, + 0.008024889044463634, + 0.004825725685805082, + 0.003306964412331581, + 0.0033675830345600843, + 0.0027742204256355762, + 0.003040182637050748, + 0.003923333249986172, + 0.0025142524391412735, + 0.0038876112084835768, + 0.003569819265976548, + 0.002919187303632498, + 0.004019892308861017, + 0.004261222667992115, + 0.005605972837656736, + 0.0034255795180797577, + 0.0029940472450107336, + 0.005816765129566193, + 0.006934689357876778, + 0.0027822968550026417, + 0.001999438041821122, + 0.003963099792599678, + 0.002939644269645214, + 0.0019844798371195793, + 0.002721542026847601, + 0.005269661545753479, + 0.00506092282012105, + 0.0024838324170559645, + 0.0024011770728975534, + 0.00470014801248908, + 0.003584659192711115, + 0.0020686262287199497, + 0.001978255808353424, + 0.007585509214550257, + 0.004115012474358082, + 0.008210176602005959, + 0.005841884296387434, + 0.031004514545202255, + 0.01198348868638277, + 0.005710501689463854, + 0.0032557742670178413, + 0.04972437024116516, + 0.026744147762656212, + 0.03804527595639229, + 0.01399869192391634, + 0.13396628201007843, + 0.039014481008052826, + 0.017017658799886703, + 0.025341883301734924, + 0.041497256606817245, + 0.1436278223991394, + 0.002639925805851817, + 0.0016865746583789587, + 0.005156178493052721, + 0.00402580201625824, + 0.0021101890597492456, + 0.002781106159090996, + 0.002857885556295514, + 0.0052892896346747875, + 0.010311262682080269, + 0.005831995978951454, + 0.03478403389453888, + 0.01517097745090723, + 0.005173365585505962, + 0.0071014403365552425, + 0.010649248026311398, + 0.031457144767045975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.28139179944992065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196252044290304 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.003196252044290304, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196252044290304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1646496057510376, + 0.09597407281398773, + 0.026841288432478905, + 0.07334305346012115, + 0.24219581484794617, + 0.0948476567864418, + 0.08298885822296143, + 0.07524556666612625, + 0.05048865079879761, + 0.09342542290687561 + ], + "credit_map": [ + 0.012572218663990498, + 0.012141992338001728, + 0.024235069751739502, + 0.01714177057147026, + 0.016274113208055496, + 0.03127468377351761, + 0.03297131136059761, + 0.018038440495729446, + 0.005614550318568945, + 0.004928983747959137, + 0.009705770760774612, + 0.007679555099457502, + 0.005814698059111834, + 0.021262291818857193, + 0.026529308408498764, + 0.014438917860388756, + 0.0022638505324721336, + 0.0031400839798152447, + 0.00215760525316, + 0.002801154274493456, + 0.003243881743401289, + 0.002458364935591817, + 0.006078989710658789, + 0.004697357304394245, + 0.007516198791563511, + 0.007167062256485224, + 0.012679212726652622, + 0.009353302419185638, + 0.004982577636837959, + 0.005590719636529684, + 0.015827717259526253, + 0.010226259008049965, + 0.019392453134059906, + 0.005683050025254488, + 0.05106557905673981, + 0.01952100731432438, + 0.008868315257132053, + 0.0048116520047187805, + 0.08718906342983246, + 0.045664675533771515, + 0.008545581251382828, + 0.007267430424690247, + 0.015872405841946602, + 0.010522312484681606, + 0.005886615253984928, + 0.005415323190391064, + 0.02606312930583954, + 0.015274862758815289, + 0.004904233850538731, + 0.0035523304250091314, + 0.018475506454706192, + 0.006800109054893255, + 0.003138592466711998, + 0.0026786159723997116, + 0.028219331055879593, + 0.015220141969621181, + 0.005178577266633511, + 0.00502282427623868, + 0.013724318705499172, + 0.01077215000987053, + 0.005760860163718462, + 0.005070360377430916, + 0.015599310398101807, + 0.014117160812020302, + 0.0033224555663764477, + 0.0034113344736397266, + 0.00887995958328247, + 0.008575649000704288, + 0.004176445305347443, + 0.00638123694807291, + 0.006677201483398676, + 0.009064366109669209, + 0.007912171073257923, + 0.004666748456656933, + 0.02587541937828064, + 0.01060903538018465, + 0.0038952925242483616, + 0.006074345204979181, + 0.008423034101724625, + 0.025969380512833595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.003196252044290304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, filmed from a slightly upward angle, captures the erosion and eventual collapse of a muddy trench edge in a calm, documentary style. In the image, icy rain falls densely from a gloomy, leaden sky, lashing the trench's damp, dark brown earthen walls. This wall, a mixture of tiny stones and rotting grass roots, is crisscrossed with tiny channels carved by erosion, through which muddy water flows. The camera focuses on the trench's crumbling upper edge, where rainwater continuously seeps in, carrying away tiny particles of earth. Initially, a few small pieces of soil loosen and fall, before a distinct crack quietly appears in the earthwork and rapidly widens. Finally, driven by gravity and rain, a massive, soaked mass of earth, unsupported, slid down silently and heavily, crashing into the muddy water at the bottom of the trench, splashing a foul spray and leaving a new, jagged gap in the original sidewall. The entire scene is filmed under even but lifeless diffuse light, with a dominant palette of dark brown, ochre, and leaden gray, creating an objective, desolate, and oppressive atmosphere.", + "seed": 19850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000061/group_9/index_7057_seed_19850_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19278097352944315, + "vlm_logprob": -0.19278097352944315, + "vlm_yes_no_margin": 2.699999952316284, + "vlm_token_credit": -0.19278097352944315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11403220146894455 + ], + "opposite_logprob_tokens": [ + -2.36403226852417 + ], + "logprob_reward": -0.11403220146894455, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11403220146894455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16825081408023834, + 0.11491663008928299, + 0.0471925288438797, + 0.13055692613124847, + 0.1016007661819458, + 0.047720178961753845, + 0.05218925327062607, + 0.040131378918886185, + 0.17800791561603546, + 0.11943352967500687 + ], + "credit_map": [ + 0.01805485412478447, + 0.022431377321481705, + 0.026087122038006783, + 0.023443154990673065, + 0.014782645739614964, + 0.017877371981739998, + 0.023702718317508698, + 0.021871568635106087, + 0.01343587040901184, + 0.012203704565763474, + 0.016435902565717697, + 0.013431413099169731, + 0.016852496191859245, + 0.014626174233853817, + 0.013612118549644947, + 0.014318952336907387, + 0.0036717718467116356, + 0.005470958072692156, + 0.004824966657906771, + 0.005086296238005161, + 0.006128582637757063, + 0.005022804718464613, + 0.010059763677418232, + 0.00692738825455308, + 0.010829883627593517, + 0.011698656715452671, + 0.0451698824763298, + 0.014129768125712872, + 0.005875755567103624, + 0.008144881576299667, + 0.011829536408185959, + 0.02287857048213482, + 0.00987923052161932, + 0.01849481277167797, + 0.024406075477600098, + 0.015104020945727825, + 0.007966785691678524, + 0.0070174518041312695, + 0.0106964111328125, + 0.008035974577069283, + 0.0038588198367506266, + 0.004685887135565281, + 0.00841186661273241, + 0.008180336095392704, + 0.005052121356129646, + 0.003781836247071624, + 0.008205433376133442, + 0.005543877370655537, + 0.004514386411756277, + 0.004539990331977606, + 0.006929627154022455, + 0.008366554044187069, + 0.006155144888907671, + 0.004970431327819824, + 0.009538335725665092, + 0.0071747880429029465, + 0.002848077565431595, + 0.0036530185025185347, + 0.007895326241850853, + 0.005225357599556446, + 0.0041212886571884155, + 0.0048689101822674274, + 0.006831536069512367, + 0.004687866661697626, + 0.012096305377781391, + 0.012782150879502296, + 0.024041909724473953, + 0.04030491039156914, + 0.018239198252558708, + 0.01806677132844925, + 0.028791798278689384, + 0.02368488721549511, + 0.007393983192741871, + 0.006372095085680485, + 0.019216954708099365, + 0.034691222012043, + 0.009775117039680481, + 0.0074656205251812935, + 0.01906784437596798, + 0.015450695529580116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.11403220146894455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221601843833923 + ], + "opposite_logprob_tokens": [ + -4.018221378326416 + ], + "logprob_reward": -0.018221601843833923, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018221601843833923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11915304511785507, + 0.18236833810806274, + 0.0910143256187439, + 0.0786534994840622, + 0.12904192507266998, + 0.062021806836128235, + 0.054072581231594086, + 0.03701583296060562, + 0.17152629792690277, + 0.07513239234685898 + ], + "credit_map": [ + 0.007932222448289394, + 0.012228117324411869, + 0.018414201214909554, + 0.014555847272276878, + 0.0165655929595232, + 0.015357136726379395, + 0.019517075270414352, + 0.014582844451069832, + 0.010302205570042133, + 0.014678382314741611, + 0.023268893361091614, + 0.02211819775402546, + 0.031919874250888824, + 0.029514893889427185, + 0.029599644243717194, + 0.02096622996032238, + 0.00577970128506422, + 0.009422590024769306, + 0.011111711151897907, + 0.009189805947244167, + 0.013925984501838684, + 0.011340787634253502, + 0.01717063970863819, + 0.013073087669909, + 0.0066529265604913235, + 0.01548359077423811, + 0.017975669354200363, + 0.009559924714267254, + 0.006412129383534193, + 0.006666706409305334, + 0.009308900684118271, + 0.006593642756342888, + 0.012356455437839031, + 0.024232573807239532, + 0.02621670812368393, + 0.020392853766679764, + 0.010875558480620384, + 0.008962949737906456, + 0.015378265641629696, + 0.010626548901200294, + 0.005755212157964706, + 0.006148523651063442, + 0.007670233957469463, + 0.014026107266545296, + 0.006154521834105253, + 0.005044816993176937, + 0.010207769460976124, + 0.007014612201601267, + 0.005786037538200617, + 0.004197211004793644, + 0.008819048292934895, + 0.007055177818983793, + 0.0065454645082354546, + 0.004512138664722443, + 0.01077202521264553, + 0.0063854726031422615, + 0.0027404711581766605, + 0.0034162523224949837, + 0.007940491661429405, + 0.004627670161426067, + 0.0032391375862061977, + 0.003641595831140876, + 0.00718055060133338, + 0.0042296587489545345, + 0.011142865754663944, + 0.013249360956251621, + 0.02666381374001503, + 0.028104010969400406, + 0.021683424711227417, + 0.020052360370755196, + 0.031150834634900093, + 0.019479621201753616, + 0.005359380505979061, + 0.004161556251347065, + 0.01457497850060463, + 0.021619336679577827, + 0.004557912703603506, + 0.005190521478652954, + 0.010808559134602547, + 0.008860134519636631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.018221601843833923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762757658958435 + ], + "opposite_logprob_tokens": [ + -0.8262757658958435 + ], + "logprob_reward": -0.5762757658958435, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762757658958435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1537095606327057, + 0.21811102330684662, + 0.07167135179042816, + 0.0757753774523735, + 0.14592689275741577, + 0.04224690794944763, + 0.037899043411016464, + 0.022029591724276543, + 0.17083780467510223, + 0.06179247424006462 + ], + "credit_map": [ + 0.016139917075634003, + 0.02027571015059948, + 0.028745882213115692, + 0.020634440705180168, + 0.012202934361994267, + 0.012539967894554138, + 0.023798281326889992, + 0.019372422248125076, + 0.023849334567785263, + 0.02843155339360237, + 0.03685614839196205, + 0.027541784569621086, + 0.022483181208372116, + 0.026496781036257744, + 0.02659420669078827, + 0.025858035311102867, + 0.005311602260917425, + 0.014318928122520447, + 0.006266755051910877, + 0.00723126670345664, + 0.01024753600358963, + 0.007276589050889015, + 0.01022531371563673, + 0.010793356224894524, + 0.005595402326434851, + 0.009402354247868061, + 0.0274369977414608, + 0.009101144038140774, + 0.003616508562117815, + 0.0031772872898727655, + 0.005882491823285818, + 0.011563196778297424, + 0.01492743194103241, + 0.027108732610940933, + 0.04327620938420296, + 0.018675250932574272, + 0.009212758392095566, + 0.007008969318121672, + 0.013372746296226978, + 0.012344797141849995, + 0.003450562944635749, + 0.002890331670641899, + 0.006957328878343105, + 0.007170374039560556, + 0.005008119158446789, + 0.003929658327251673, + 0.008136666379868984, + 0.004703868646174669, + 0.005484686698764563, + 0.003314955858513713, + 0.0049972753040492535, + 0.005923133343458176, + 0.003957363776862621, + 0.0029296884313225746, + 0.005868525709956884, + 0.005423412658274174, + 0.0025503896176815033, + 0.002470513107255101, + 0.0031732467468827963, + 0.0031475836876779795, + 0.002475840039551258, + 0.0019219276728108525, + 0.003930994775146246, + 0.0023590978235006332, + 0.01100225280970335, + 0.013979450799524784, + 0.02261149324476719, + 0.031430214643478394, + 0.021130310371518135, + 0.020277870818972588, + 0.03143155947327614, + 0.018974652513861656, + 0.004415920935571194, + 0.004586304072290659, + 0.010027622804045677, + 0.011997995898127556, + 0.008313882164657116, + 0.006125058513134718, + 0.009340525604784489, + 0.006985166110098362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.5762757658958435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521759569644928 + ], + "opposite_logprob_tokens": [ + -1.5021759271621704 + ], + "logprob_reward": -0.2521759569644928, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521759569644928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06942875683307648, + 0.12450258433818817, + 0.06914679706096649, + 0.10849720984697342, + 0.13049602508544922, + 0.08213499933481216, + 0.05750880017876625, + 0.03299202397465706, + 0.19735758006572723, + 0.12793515622615814 + ], + "credit_map": [ + 0.006114710122346878, + 0.00767176179215312, + 0.010921248234808445, + 0.00990363024175167, + 0.00720598129555583, + 0.007967440411448479, + 0.01057785376906395, + 0.009066136553883553, + 0.006511625833809376, + 0.010970473289489746, + 0.013761885464191437, + 0.015461950562894344, + 0.02059856988489628, + 0.022653141990303993, + 0.020488809794187546, + 0.014056144282221794, + 0.004612820688635111, + 0.0068236044608056545, + 0.007333117071539164, + 0.006926302798092365, + 0.010734292678534985, + 0.008909383788704872, + 0.015169406309723854, + 0.008637873455882072, + 0.0078963041305542, + 0.01771935075521469, + 0.03448512777686119, + 0.0138772614300251, + 0.006033842917531729, + 0.006108116824179888, + 0.009724201634526253, + 0.012653011828660965, + 0.010798170231282711, + 0.021658439189195633, + 0.020510120317339897, + 0.023720167577266693, + 0.012614473700523376, + 0.010025324299931526, + 0.01955956406891346, + 0.011609773151576519, + 0.006593856494873762, + 0.007579704746603966, + 0.010792476125061512, + 0.018390478566288948, + 0.008580089546740055, + 0.006250209640711546, + 0.014828925020992756, + 0.009119265712797642, + 0.005250619258731604, + 0.004425699356943369, + 0.009018242359161377, + 0.009643918834626675, + 0.006791634950786829, + 0.004442437086254358, + 0.011173791252076626, + 0.006762464065104723, + 0.0024580738972872496, + 0.0034677041694521904, + 0.006119806319475174, + 0.006280403584241867, + 0.002723782556131482, + 0.002537504304200411, + 0.0056235999800264835, + 0.0037811510264873505, + 0.014174417592585087, + 0.01526617631316185, + 0.02725898288190365, + 0.035099565982818604, + 0.024198899045586586, + 0.022971786558628082, + 0.03660104051232338, + 0.02178671956062317, + 0.009047330357134342, + 0.008242640644311905, + 0.02116776444017887, + 0.03351330757141113, + 0.010239665396511555, + 0.007939711213111877, + 0.023111334070563316, + 0.014673409052193165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.2521759569644928 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot filmed from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a muddy trench with damp, dark brown earthen walls containing tiny stones and rotting grass roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is icy rain falling densely from a gloomy, leaden sky onto the trench walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a crack appearing in the trench wall and widening until a massive mass of earth slides down and crashes into the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting even and diffuse with a dominant palette of dark brown, ochre, and leaden gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031993414741009474 + ], + "opposite_logprob_tokens": [ + -5.753199577331543 + ], + "logprob_reward": -0.0031993414741009474, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031993414741009474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1416826844215393, + 0.10181750357151031, + 0.10007037222385406, + 0.1315498650074005, + 0.13707399368286133, + 0.0634608268737793, + 0.04175528138875961, + 0.04681922495365143, + 0.10172994434833527, + 0.13404026627540588 + ], + "credit_map": [ + 0.014829561114311218, + 0.016207195818424225, + 0.02279876545071602, + 0.023805182427167892, + 0.011213286779820919, + 0.014932716265320778, + 0.018820589408278465, + 0.019075391814112663, + 0.007186005357652903, + 0.015749486163258553, + 0.009414167143404484, + 0.010681027546525002, + 0.017683865502476692, + 0.011376365087926388, + 0.011718342080712318, + 0.018008237704634666, + 0.007242515683174133, + 0.02325626090168953, + 0.011094191111624241, + 0.008473491296172142, + 0.009253570809960365, + 0.007216830272227526, + 0.010216696187853813, + 0.023316822946071625, + 0.009791930206120014, + 0.018218040466308594, + 0.047546617686748505, + 0.015495603904128075, + 0.006474432535469532, + 0.005361019168049097, + 0.008756035007536411, + 0.01990617997944355, + 0.010541100054979324, + 0.010699375532567501, + 0.04022549092769623, + 0.017398366704583168, + 0.011982133612036705, + 0.009244072251021862, + 0.021260250359773636, + 0.01572321355342865, + 0.005900314077734947, + 0.005643927026540041, + 0.011490239761769772, + 0.013036718592047691, + 0.004178558476269245, + 0.005552161019295454, + 0.008811202831566334, + 0.008847705088555813, + 0.003229392459616065, + 0.002991149667650461, + 0.006728773005306721, + 0.008485279977321625, + 0.004525595810264349, + 0.0035011053550988436, + 0.007200173567980528, + 0.005093812011182308, + 0.003208260051906109, + 0.0034750057384371758, + 0.006830939091742039, + 0.009731736965477467, + 0.005512922070920467, + 0.0048712012358009815, + 0.007539165671914816, + 0.005649997852742672, + 0.008580894209444523, + 0.008878719992935658, + 0.015293382108211517, + 0.01566784270107746, + 0.010037655010819435, + 0.011941187083721161, + 0.01837988756597042, + 0.012950376607477665, + 0.013529375195503235, + 0.008295763283967972, + 0.02082999050617218, + 0.03187224268913269, + 0.010181552730500698, + 0.007005438674241304, + 0.024521982297301292, + 0.017803922295570374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0031993414741009474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_0/index_tensor([8959, 8959])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_0/index_tensor([8959, 8959])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..89ba8497185bf9d4dbfba89625e516757333789e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_0/index_tensor([8959, 8959])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up, the camera slowly advances with a solemn and smooth motion, focusing on a luxurious fountain pen, resting in a Rubens-like shimmer of light and shadow. A warm, intense sidelight carves the sharp outline of the hexagonal white star inlaid on the cap tip and flows along the surface of the precious black resin of the pen barrel, creating a soft, jade-like luster. As the cap is slowly unscrewed by an invisible hand in an almost reverent slow motion, the gilded band connecting the pen barrel and cap reflects a rich, fiery hue, complementing the deep black resin. The camera zooms in, capturing the precise moment as the precision-machined metal threads gradually separate as they rotate. The edge of each thread shimmers in the light, a unique glimmer of mechanical beauty, while the other side sinks into a soft, bottomless shadow. The entire image is a supreme celebration of craftsmanship and quality, creating an atmosphere of focused, luxurious, and anticipatory tranquility.", + "seed": 28930, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_0/index_8959_seed_28930_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07591448277235031, + "vlm_logprob": -0.07591448277235031, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07591448277235031, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1047075018286705 + ], + "opposite_logprob_tokens": [ + -2.4797074794769287 + ], + "logprob_reward": -0.1047075018286705, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1047075018286705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040349505841732025, + 0.008631418459117413, + 0.012730272486805916, + 0.03594476729631424, + 0.10558399558067322, + 0.01920744590461254, + 0.016207894310355186, + 0.6819465756416321, + 0.02198110520839691, + 0.05741705000400543 + ], + "credit_map": [ + 0.006632366683334112, + 0.0053444732911884785, + 0.0070619480684399605, + 0.008118588477373123, + 0.00302185770124197, + 0.0026500781532377005, + 0.004868416581302881, + 0.00265177758410573, + 0.0012599511537700891, + 0.0010203546844422817, + 0.0017947327578440309, + 0.0008849019650369883, + 0.0006763699348084629, + 0.0010019714245572686, + 0.000760643626563251, + 0.00123249227181077, + 0.001986132934689522, + 0.0012548906961455941, + 0.001742958091199398, + 0.0017023785039782524, + 0.0011391082080081105, + 0.0019418834708631039, + 0.0015972853871062398, + 0.001365634729154408, + 0.0036506280303001404, + 0.004608252085745335, + 0.0038965162821114063, + 0.005332998465746641, + 0.007394010666757822, + 0.005970736965537071, + 0.002489466918632388, + 0.0026021606754511595, + 0.011636028997600079, + 0.014080008491873741, + 0.009379645809531212, + 0.01008461881428957, + 0.023703698068857193, + 0.014157826080918312, + 0.010258295573294163, + 0.012283877469599247, + 0.003051330801099539, + 0.005869319662451744, + 0.0012560569448396564, + 0.001989982556551695, + 0.0012521090684458613, + 0.001512365066446364, + 0.001055536326020956, + 0.0032207458280026913, + 0.0018673212034627795, + 0.001365338684991002, + 0.0018518605502322316, + 0.0038761335890740156, + 0.001450583804398775, + 0.0027962157037109137, + 0.0010777340503409505, + 0.0019227075390517712, + 0.08344293385744095, + 0.02844778262078762, + 0.02493751235306263, + 0.19350257515907288, + 0.014743087813258171, + 0.2843415439128876, + 0.031696617603302, + 0.020834481343626976, + 0.0021431122440844774, + 0.002076979260891676, + 0.0014876971254125237, + 0.0028282422572374344, + 0.002047998830676079, + 0.003624711884185672, + 0.004988753702491522, + 0.002783609554171562, + 0.0043975370936095715, + 0.004657988902181387, + 0.005052185617387295, + 0.009665521793067455, + 0.006271721329540014, + 0.0092503996565938, + 0.012630694545805454, + 0.0054910024628043175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.1047075018286705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559292614459991 + ], + "opposite_logprob_tokens": [ + -1.6005929708480835 + ], + "logprob_reward": -0.22559292614459991, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559292614459991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06594908237457275, + 0.02234991267323494, + 0.03697085753083229, + 0.03929165005683899, + 0.04198469594120979, + 0.059149764478206635, + 0.04399891197681427, + 0.6326531171798706, + 0.024266434833407402, + 0.03338564187288284 + ], + "credit_map": [ + 0.013307237066328526, + 0.008034326136112213, + 0.015094205737113953, + 0.01158385630697012, + 0.002573884092271328, + 0.0041975826025009155, + 0.006018172018229961, + 0.005139807239174843, + 0.0034546847455203533, + 0.0032742368057370186, + 0.003960306756198406, + 0.002867595525458455, + 0.0015805498696863651, + 0.002975133480504155, + 0.0017734924331307411, + 0.0024639100302010775, + 0.0076132179237902164, + 0.0038185138255357742, + 0.00397901376709342, + 0.004199577029794455, + 0.003314216621220112, + 0.006190812215209007, + 0.0038421181961894035, + 0.004013383761048317, + 0.006692749448120594, + 0.006060524377971888, + 0.006460227072238922, + 0.004233785904943943, + 0.006005099508911371, + 0.0038778583984822035, + 0.0026457547210156918, + 0.0033156471326947212, + 0.005111172795295715, + 0.004916749428957701, + 0.003289317013695836, + 0.004730683285742998, + 0.00869686622172594, + 0.005910919979214668, + 0.004046940710395575, + 0.005282040685415268, + 0.014180367812514305, + 0.012818015180528164, + 0.005002854857593775, + 0.0066422889940440655, + 0.004592491779476404, + 0.006134629715234041, + 0.003996226936578751, + 0.005782884545624256, + 0.005581630859524012, + 0.004841421265155077, + 0.003821850521489978, + 0.00852343812584877, + 0.004813028499484062, + 0.0055497377179563046, + 0.004859482403844595, + 0.0060083139687776566, + 0.07681971788406372, + 0.0266568586230278, + 0.023029053583741188, + 0.17936496436595917, + 0.013663156889379025, + 0.2634407579898834, + 0.029776107519865036, + 0.01990250125527382, + 0.0019891683477908373, + 0.002259016502648592, + 0.0022344915196299553, + 0.004269243683665991, + 0.0025805996265262365, + 0.0044552236795425415, + 0.0038239536806941032, + 0.0026547338347882032, + 0.002465934259817004, + 0.0030439754482358694, + 0.0026038140058517456, + 0.005074797663837671, + 0.003682762151584029, + 0.005859945900738239, + 0.007077718619257212, + 0.003576690563932061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.22559292614459991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205877669155598 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205877669155598, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205877669155598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022316399961709976, + 0.00972186028957367, + 0.021541191264986992, + 0.05585647374391556, + 0.1689089685678482, + 0.05396898835897446, + 0.08482421934604645, + 0.5021385550498962, + 0.0281039047986269, + 0.05261942744255066 + ], + "credit_map": [ + 0.004186255857348442, + 0.003022237913683057, + 0.005946453660726547, + 0.003191903932020068, + 0.0009713161853142083, + 0.0013880410697311163, + 0.0018310160376131535, + 0.0017791750142350793, + 0.001836688257753849, + 0.001381659065373242, + 0.0023365123197436333, + 0.0011061691911891103, + 0.0005384128890000284, + 0.0010096065234392881, + 0.0005846379790455103, + 0.0009281746461056173, + 0.0034881089814007282, + 0.0019419655436649919, + 0.0029217463452368975, + 0.0030446562450379133, + 0.001700759632512927, + 0.0030318391509354115, + 0.0028064183425158262, + 0.002605697140097618, + 0.0055417753756046295, + 0.00728809367865324, + 0.00533642154186964, + 0.006989657413214445, + 0.010406361892819405, + 0.010802325792610645, + 0.004531147424131632, + 0.004960695747286081, + 0.0322277769446373, + 0.05387584865093231, + 0.008926516398787498, + 0.015100531280040741, + 0.016346506774425507, + 0.014682185836136341, + 0.010892154648900032, + 0.0168574508279562, + 0.012974864803254604, + 0.013655480928719044, + 0.004482790362089872, + 0.006543262861669064, + 0.005209183786064386, + 0.004343758337199688, + 0.0029888199642300606, + 0.003770826617255807, + 0.013590569607913494, + 0.010311571881175041, + 0.010163405910134315, + 0.01062105130404234, + 0.011284777894616127, + 0.00988443661481142, + 0.00704390462487936, + 0.011924496851861477, + 0.06128402799367905, + 0.021977342665195465, + 0.018236327916383743, + 0.14150546491146088, + 0.011093635112047195, + 0.2079598307609558, + 0.02381947450339794, + 0.016262469813227654, + 0.0034014859702438116, + 0.002988552674651146, + 0.0021492713131010532, + 0.004156285431236029, + 0.0028130339924246073, + 0.005854809656739235, + 0.0042799548245966434, + 0.0024605130311101675, + 0.005167398601770401, + 0.004645439796149731, + 0.004486200865358114, + 0.006636212579905987, + 0.006793662440031767, + 0.007868744432926178, + 0.009783494286239147, + 0.007238270249217749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.014205877669155598 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981266938149929 + ], + "opposite_logprob_tokens": [ + -3.5298125743865967 + ], + "logprob_reward": -0.02981266938149929, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981266938149929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03361828997731209, + 0.07618551701307297, + 0.04799751192331314, + 0.06986510008573532, + 0.0897328332066536, + 0.0740499272942543, + 0.05825038626790047, + 0.46711063385009766, + 0.065509133040905, + 0.017680682241916656 + ], + "credit_map": [ + 0.005739300046116114, + 0.0038917516358196735, + 0.005505148321390152, + 0.0068878899328410625, + 0.002308859024196863, + 0.003011210123077035, + 0.0032285370398312807, + 0.0030455931555479765, + 0.012212307192385197, + 0.008828557096421719, + 0.013908104039728642, + 0.009278605692088604, + 0.006671737879514694, + 0.010589216835796833, + 0.007270353846251965, + 0.007426635827869177, + 0.007160582579672337, + 0.005038984585553408, + 0.006863731890916824, + 0.006918241269886494, + 0.004140036180615425, + 0.005908052437007427, + 0.006318769417703152, + 0.0056491149589419365, + 0.01007881574332714, + 0.009284261614084244, + 0.009716437198221684, + 0.0070539722219109535, + 0.012807819060981274, + 0.011353844776749611, + 0.004659043159335852, + 0.0049109081737697124, + 0.012414265424013138, + 0.010032985359430313, + 0.009791168384253979, + 0.01257509645074606, + 0.014387236908078194, + 0.01421133428812027, + 0.00793361198157072, + 0.008387132547795773, + 0.009278317913413048, + 0.012336000800132751, + 0.0053805080242455006, + 0.008486713282763958, + 0.011036188341677189, + 0.014180121943354607, + 0.006108356639742851, + 0.007243718486279249, + 0.008447978645563126, + 0.005151954013854265, + 0.00622192956507206, + 0.010710185393691063, + 0.0054488275200128555, + 0.007622441742569208, + 0.0064934780821204185, + 0.008153590373694897, + 0.05529049038887024, + 0.019774362444877625, + 0.017567269504070282, + 0.13420522212982178, + 0.01118397805839777, + 0.19279567897319794, + 0.021660326048731804, + 0.014633323065936565, + 0.008359480649232864, + 0.007328906562179327, + 0.007792225573211908, + 0.007748481817543507, + 0.006711399648338556, + 0.009027467109262943, + 0.011492975056171417, + 0.007048195227980614, + 0.0022131558507680893, + 0.0015068253269419074, + 0.001556243048980832, + 0.002260729670524597, + 0.0017948222812265158, + 0.0025072635617107153, + 0.0033056605607271194, + 0.0025359834544360638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.02981266938149929 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253438837826252 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253438837826252, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253438837826252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1411716789007187, + 0.07523733377456665, + 0.08144699782133102, + 0.03258778154850006, + 0.108223557472229, + 0.061153993010520935, + 0.05421426519751549, + 0.308430939912796, + 0.03459220379590988, + 0.10294121503829956 + ], + "credit_map": [ + 0.024641839787364006, + 0.0167543925344944, + 0.025837492197752, + 0.026965515688061714, + 0.007051678374409676, + 0.011452313512563705, + 0.0138343321159482, + 0.014634114690124989, + 0.012242679484188557, + 0.007599690463393927, + 0.01961524970829487, + 0.00870133750140667, + 0.005561599507927895, + 0.007589203305542469, + 0.006059080362319946, + 0.007868499495089054, + 0.012592787854373455, + 0.008599835447967052, + 0.010710369795560837, + 0.012432525865733624, + 0.006609699223190546, + 0.0108566889539361, + 0.010565380565822124, + 0.009079708717763424, + 0.005086506251245737, + 0.004427991341799498, + 0.004400963895022869, + 0.006139758042991161, + 0.002734206384047866, + 0.0031806977931410074, + 0.0026631460059434175, + 0.0039545102044939995, + 0.01847958005964756, + 0.021252142265439034, + 0.008058919571340084, + 0.010202198289334774, + 0.017536666244268417, + 0.011210795491933823, + 0.009075206704437733, + 0.012408046051859856, + 0.01051356177777052, + 0.01387861929833889, + 0.00629563070833683, + 0.005755046848207712, + 0.00601524394005537, + 0.007713560946285725, + 0.004689445253461599, + 0.006292884238064289, + 0.007887822575867176, + 0.006872958969324827, + 0.005609831307083368, + 0.008038780651986599, + 0.005076923873275518, + 0.006134352181106806, + 0.005039868876338005, + 0.009553730487823486, + 0.037639979273080826, + 0.013387860730290413, + 0.011281607672572136, + 0.08688787370920181, + 0.008297903463244438, + 0.1264733523130417, + 0.014470468275249004, + 0.009991902858018875, + 0.003737288061529398, + 0.00541189918294549, + 0.002544761635363102, + 0.0042241420596838, + 0.005296614486724138, + 0.0047121308743953705, + 0.006051407661288977, + 0.0026139612309634686, + 0.0058450158685445786, + 0.008459731005132198, + 0.010513200424611568, + 0.019784441217780113, + 0.009786760434508324, + 0.017426103353500366, + 0.023335859179496765, + 0.007790104951709509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.005253438837826252 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera slowly advances with a solemn and smooth motion, focusing on a luxurious fountain pen, resting in a Rubens-like shimmer of light and shadow. A warm, intense sidelight carves the sharp outline of the hexagonal white star inlaid on the cap tip and flows along the surface of the precious black resin of the pen barrel, creating a soft, jade-like luster. As the cap is slowly unscrewed by an invisible hand in an almost reverent slow motion, the gilded band connecting the pen barrel and cap reflects a rich, fiery hue, complementing the deep black resin. The camera zooms in, capturing the precise moment as the precision-machined metal threads gradually separate as they rotate. The edge of each thread shimmers in the light, a unique glimmer of mechanical beauty, while the other side sinks into a soft, bottomless shadow. The entire image is a supreme celebration of craftsmanship and quality, creating an atmosphere of focused, luxurious, and anticipatory tranquility.", + "seed": 28931, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_0/index_8959_seed_28931_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1187132366001606, + "vlm_logprob": -0.1187132366001606, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.1187132366001606, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25997573137283325 + ], + "opposite_logprob_tokens": [ + -1.5099756717681885 + ], + "logprob_reward": -0.25997573137283325, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25997573137283325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10004177689552307, + 0.06000596284866333, + 0.019205624237656593, + 0.07543429732322693, + 0.05509788170456886, + 0.49161496758461, + 0.038680534809827805, + 0.04623202607035637, + 0.03617851808667183, + 0.07750844210386276 + ], + "credit_map": [ + 0.012338380329310894, + 0.01426042802631855, + 0.039502669125795364, + 0.007158743217587471, + 0.0032490112353116274, + 0.00294282753020525, + 0.004658812191337347, + 0.0159309022128582, + 0.009303378872573376, + 0.023185476660728455, + 0.0022839587181806564, + 0.0037670896854251623, + 0.0021322511602193117, + 0.0026876074261963367, + 0.003069610334932804, + 0.013576585799455643, + 0.0024349591694772243, + 0.001956312684342265, + 0.0029676936101168394, + 0.002456970978528261, + 0.001599295879714191, + 0.002537346910685301, + 0.0032428186386823654, + 0.0020102248527109623, + 0.010526313446462154, + 0.008928442373871803, + 0.0075507620349526405, + 0.016867587342858315, + 0.003318802220746875, + 0.003496304852887988, + 0.01246710866689682, + 0.012278974056243896, + 0.006878379732370377, + 0.00638226605951786, + 0.005446399096399546, + 0.008638223633170128, + 0.004478858318179846, + 0.004733248148113489, + 0.012830127961933613, + 0.005710375029593706, + 0.18485461175441742, + 0.035657480359077454, + 0.021363692358136177, + 0.012377523817121983, + 0.009238152764737606, + 0.014340033754706383, + 0.03738962113857269, + 0.17639383673667908, + 0.006545226089656353, + 0.0036067760083824396, + 0.0041959271766245365, + 0.005110009107738733, + 0.0034518917091190815, + 0.0060003530234098434, + 0.005375772714614868, + 0.004394579213112593, + 0.006128213834017515, + 0.004741562530398369, + 0.0073358709923923016, + 0.00565515598282218, + 0.0048495070077478886, + 0.005625810939818621, + 0.005486483685672283, + 0.006409415043890476, + 0.005391917657107115, + 0.005465931724756956, + 0.0045064897276461124, + 0.005335624795407057, + 0.004063062835484743, + 0.003807350993156433, + 0.003697608830407262, + 0.003910533618181944, + 0.011915726587176323, + 0.007479001767933369, + 0.010762923397123814, + 0.014752655290067196, + 0.006602417677640915, + 0.0057713985443115234, + 0.008876421488821507, + 0.011347887106239796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.25997573137283325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814143896102905 + ], + "opposite_logprob_tokens": [ + -1.4064143896102905 + ], + "logprob_reward": -0.2814143896102905, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814143896102905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04925072565674782, + 0.1178293526172638, + 0.01270343642681837, + 0.06014425307512283, + 0.1884494125843048, + 0.2301490157842636, + 0.04779401049017906, + 0.13275137543678284, + 0.0902111604809761, + 0.07071724534034729 + ], + "credit_map": [ + 0.0033588178921490908, + 0.009847590699791908, + 0.016332706436514854, + 0.00461313221603632, + 0.002202612115070224, + 0.0037363518495112658, + 0.0030507445335388184, + 0.006108773872256279, + 0.01197244506329298, + 0.013639152981340885, + 0.013953105546534061, + 0.01689532957971096, + 0.006642317399382591, + 0.019366981461644173, + 0.011814993806183338, + 0.023545024916529655, + 0.0015547937946394086, + 0.0019210465252399445, + 0.0011452703038230538, + 0.0022943844087421894, + 0.0008320192573592067, + 0.0007885076920501888, + 0.002007044618949294, + 0.0021603701170533895, + 0.01010780967772007, + 0.006513345520943403, + 0.009370215237140656, + 0.008869613520801067, + 0.00229440163820982, + 0.004612716380506754, + 0.008256733417510986, + 0.010119418613612652, + 0.022132575511932373, + 0.056005336344242096, + 0.011715997010469437, + 0.02204054780304432, + 0.007844674400985241, + 0.007777650374919176, + 0.01638021133840084, + 0.04455242305994034, + 0.08485359698534012, + 0.017236290499567986, + 0.01013146061450243, + 0.007514054421335459, + 0.004597338382154703, + 0.006678183563053608, + 0.017419476062059402, + 0.08171861618757248, + 0.008035441860556602, + 0.006167241372168064, + 0.005649449769407511, + 0.00952437985688448, + 0.0033623208291828632, + 0.003499399172142148, + 0.006022124085575342, + 0.005533652380108833, + 0.01939723640680313, + 0.009844250977039337, + 0.023442164063453674, + 0.024561874568462372, + 0.01241221558302641, + 0.010613781400024891, + 0.012513278983533382, + 0.019966574385762215, + 0.010841716080904007, + 0.0059670014306902885, + 0.018938515335321426, + 0.018466804176568985, + 0.005828389432281256, + 0.007766874507069588, + 0.009532787837088108, + 0.012869072146713734, + 0.010851320810616016, + 0.007941064424812794, + 0.010724978521466255, + 0.009650367312133312, + 0.005786278285086155, + 0.010166654363274574, + 0.007174755912274122, + 0.008421823382377625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.2814143896102905 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193040043115616 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.018193040043115616, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193040043115616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02275148592889309, + 0.023694641888141632, + 0.016095545142889023, + 0.026191331446170807, + 0.048155635595321655, + 0.5938669443130493, + 0.023035703226923943, + 0.1703958809375763, + 0.03833284601569176, + 0.03748001158237457 + ], + "credit_map": [ + 0.002299344865605235, + 0.0055945864878594875, + 0.00612043309956789, + 0.001882642274722457, + 0.0009775899816304445, + 0.0015124526107683778, + 0.001451836316846311, + 0.0029126014560461044, + 0.004498966969549656, + 0.0031308745965361595, + 0.002788252430036664, + 0.0032851723954081535, + 0.0015281259547919035, + 0.0021029191557317972, + 0.002591540105640888, + 0.003768790978938341, + 0.0022966444958001375, + 0.0021830452606081963, + 0.0019491049461066723, + 0.0033033201470971107, + 0.0008683398482389748, + 0.0017147896578535438, + 0.0021498342975974083, + 0.0016304667806252837, + 0.004320813342928886, + 0.003143309848383069, + 0.00645456463098526, + 0.003412760328501463, + 0.001042305026203394, + 0.0010600811801850796, + 0.003093572799116373, + 0.0036639277823269367, + 0.008138497360050678, + 0.0066193388774991035, + 0.009489231742918491, + 0.007375555112957954, + 0.0021870604250580072, + 0.0020273944828659296, + 0.00575214484706521, + 0.006566410884261131, + 0.21725507080554962, + 0.04386080428957939, + 0.02694791927933693, + 0.020744960755109787, + 0.012027676217257977, + 0.018278665840625763, + 0.044609516859054565, + 0.2101423740386963, + 0.002851005643606186, + 0.001807990251109004, + 0.003601971548050642, + 0.003726726630702615, + 0.00173189549241215, + 0.0029537624213844538, + 0.0031959672924131155, + 0.0031663854606449604, + 0.024402350187301636, + 0.013186431489884853, + 0.028655823320150375, + 0.02755882777273655, + 0.012099414132535458, + 0.013748176395893097, + 0.01930288039147854, + 0.03144198656082153, + 0.004663810599595308, + 0.003534722374752164, + 0.007840130478143692, + 0.008208885788917542, + 0.002393946284428239, + 0.0037694990169256926, + 0.0037825845647603273, + 0.004139270167797804, + 0.004804076626896858, + 0.003961324226111174, + 0.004857035353779793, + 0.006415537558495998, + 0.003764245891943574, + 0.003610173938795924, + 0.00463520921766758, + 0.005432411562651396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.018193040043115616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355018839240074 + ], + "opposite_logprob_tokens": [ + -3.651355028152466 + ], + "logprob_reward": -0.026355018839240074, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355018839240074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10134255141019821, + 0.028645241633057594, + 0.00744412187486887, + 0.016970938071608543, + 0.039994534105062485, + 0.6756265759468079, + 0.021804168820381165, + 0.05050838738679886, + 0.044243890792131424, + 0.013419633731245995 + ], + "credit_map": [ + 0.017766613513231277, + 0.012960586696863174, + 0.043814517557621, + 0.004259180743247271, + 0.0025755390524864197, + 0.002766324207186699, + 0.0021326104179024696, + 0.015067165717482567, + 0.003969516605138779, + 0.0034445433411747217, + 0.0029166832100600004, + 0.007821984589099884, + 0.0016355105908587575, + 0.0024990562815219164, + 0.003174833720549941, + 0.0031831120140850544, + 0.0012899431167170405, + 0.0005330399726517498, + 0.0011026114225387573, + 0.0009152807760983706, + 0.0004726204206235707, + 0.0007164105772972107, + 0.0009992423001676798, + 0.001414972823113203, + 0.0023365868255496025, + 0.001931666978634894, + 0.0021837688982486725, + 0.0026973888743668795, + 0.0011872948380187154, + 0.0015650494024157524, + 0.003305545076727867, + 0.0017636356642469764, + 0.005925256758928299, + 0.01390455849468708, + 0.0032865835819393396, + 0.0026499300729483366, + 0.0015887637855485082, + 0.0012103314511477947, + 0.00275879492983222, + 0.00867031142115593, + 0.25011923909187317, + 0.04897356778383255, + 0.02899150550365448, + 0.022203078493475914, + 0.01333656720817089, + 0.01973973587155342, + 0.051779136061668396, + 0.24048368632793427, + 0.0028542340733110905, + 0.0019552884623408318, + 0.0026828423142433167, + 0.003855084301903844, + 0.0023381721694022417, + 0.002653801580891013, + 0.0030990513041615486, + 0.002365692285820842, + 0.007563265506178141, + 0.005262276623398066, + 0.006965248379856348, + 0.011300316080451012, + 0.0037871028762310743, + 0.003247661981731653, + 0.007554094772785902, + 0.004828415345400572, + 0.0054700979962944984, + 0.0030850707553327084, + 0.009469018317759037, + 0.009381801821291447, + 0.0025750361382961273, + 0.003977478481829166, + 0.004508508834987879, + 0.005776871927082539, + 0.0024565444327890873, + 0.000877059530466795, + 0.0017950517358258367, + 0.002318434650078416, + 0.0010218406096100807, + 0.0011510499753057957, + 0.0019404652994126081, + 0.001859185635112226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.026355018839240074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076280031353235245 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.0076280031353235245, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076280031353235245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18725557625293732, + 0.09034501761198044, + 0.024988681077957153, + 0.017644844949245453, + 0.08989059925079346, + 0.26459628343582153, + 0.031138643622398376, + 0.10987808555364609, + 0.11109078675508499, + 0.0731714516878128 + ], + "credit_map": [ + 0.02778247743844986, + 0.03979652374982834, + 0.08168669044971466, + 0.007797910366207361, + 0.0054822275415062904, + 0.004152974113821983, + 0.005741221830248833, + 0.014815566129982471, + 0.01267184130847454, + 0.014713752083480358, + 0.00928597990423441, + 0.02695990540087223, + 0.004298912361264229, + 0.006554226391017437, + 0.008365951478481293, + 0.007494449615478516, + 0.00392887182533741, + 0.003040231531485915, + 0.002741999225690961, + 0.006727902684360743, + 0.0012410423951223493, + 0.0016144363908097148, + 0.003413416212424636, + 0.0022807845380157232, + 0.0024903500452637672, + 0.002128553343936801, + 0.002318155951797962, + 0.0034958282485604286, + 0.0008744891383685172, + 0.0010167239233851433, + 0.0026398247573524714, + 0.002680920297279954, + 0.010063296183943748, + 0.024439530447125435, + 0.005392611958086491, + 0.01122684683650732, + 0.004088568035513163, + 0.004486578516662121, + 0.00876674149185419, + 0.02142644301056862, + 0.09880216419696808, + 0.01929132267832756, + 0.011965609155595303, + 0.007746844552457333, + 0.004897733684629202, + 0.007970924489200115, + 0.019841203466057777, + 0.09408048540353775, + 0.004334259312599897, + 0.003356000641360879, + 0.005753352772444487, + 0.004911611322313547, + 0.0022450387477874756, + 0.003287231083959341, + 0.0037518013268709183, + 0.0034993491135537624, + 0.016993524506688118, + 0.006914851255714893, + 0.020589595660567284, + 0.016286732628941536, + 0.00696692056953907, + 0.008639219217002392, + 0.013234001584351063, + 0.020253244787454605, + 0.014045356772840023, + 0.007818024605512619, + 0.024988286197185516, + 0.025053054094314575, + 0.005684555508196354, + 0.008177830837666988, + 0.011302651837468147, + 0.014021032489836216, + 0.012415922246873379, + 0.006864666007459164, + 0.011272495612502098, + 0.011443888768553734, + 0.004829131532460451, + 0.007462896406650543, + 0.008237522095441818, + 0.01064493041485548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0076280031353235245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera slowly advances with a solemn and smooth motion, focusing on a luxurious fountain pen, resting in a Rubens-like shimmer of light and shadow. A warm, intense sidelight carves the sharp outline of the hexagonal white star inlaid on the cap tip and flows along the surface of the precious black resin of the pen barrel, creating a soft, jade-like luster. As the cap is slowly unscrewed by an invisible hand in an almost reverent slow motion, the gilded band connecting the pen barrel and cap reflects a rich, fiery hue, complementing the deep black resin. The camera zooms in, capturing the precise moment as the precision-machined metal threads gradually separate as they rotate. The edge of each thread shimmers in the light, a unique glimmer of mechanical beauty, while the other side sinks into a soft, bottomless shadow. The entire image is a supreme celebration of craftsmanship and quality, creating an atmosphere of focused, luxurious, and anticipatory tranquility.", + "seed": 28932, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_0/index_8959_seed_28932_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12147271735593676, + "vlm_logprob": -0.12147271735593676, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.12147271735593676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3020007908344269 + ], + "opposite_logprob_tokens": [ + -1.4270007610321045 + ], + "logprob_reward": -0.3020007908344269, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3020007908344269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23829063773155212, + 0.12349080294370651, + 0.0907847136259079, + 0.05345822870731354, + 0.04502381384372711, + 0.06610867381095886, + 0.190195232629776, + 0.08766144514083862, + 0.06610754877328873, + 0.03887892886996269 + ], + "credit_map": [ + 0.04034635052084923, + 0.02189810946583748, + 0.02868221141397953, + 0.02816009894013405, + 0.028127305209636688, + 0.025259584188461304, + 0.04017867147922516, + 0.02563830092549324, + 0.021133476868271828, + 0.009656073525547981, + 0.022734355181455612, + 0.016771424561738968, + 0.008284779265522957, + 0.012589876540005207, + 0.02039952017366886, + 0.011921301484107971, + 0.01737045869231224, + 0.010304044932126999, + 0.01254927646368742, + 0.011933112516999245, + 0.009532790631055832, + 0.007059511728584766, + 0.01378925982862711, + 0.008246260695159435, + 0.008799214847385883, + 0.0033364249393343925, + 0.006122647784650326, + 0.007335003465414047, + 0.00434275483712554, + 0.00666087307035923, + 0.008131337352097034, + 0.008729971945285797, + 0.007858111523091793, + 0.004042817745357752, + 0.007697648834437132, + 0.007502402178943157, + 0.0027485324535518885, + 0.005452918820083141, + 0.005917058791965246, + 0.0038043258246034384, + 0.011041716672480106, + 0.004852045327425003, + 0.009994903579354286, + 0.01047425251454115, + 0.004601750988513231, + 0.00764182535931468, + 0.01071245688945055, + 0.006789725739508867, + 0.01872151531279087, + 0.015494443476200104, + 0.02060369960963726, + 0.03259630873799324, + 0.021675636991858482, + 0.022572819143533707, + 0.029136372730135918, + 0.02939443662762642, + 0.00655254116281867, + 0.005079719237983227, + 0.011751403100788593, + 0.026214169338345528, + 0.004765772260725498, + 0.010740754194557667, + 0.012799734249711037, + 0.009757356718182564, + 0.006154612638056278, + 0.004187503829598427, + 0.011017152108252048, + 0.008895189501345158, + 0.003868156811222434, + 0.01294226199388504, + 0.010145903564989567, + 0.008896764367818832, + 0.0037237328942865133, + 0.002054864540696144, + 0.004756637383252382, + 0.005348584149032831, + 0.002346751047298312, + 0.0066527109593153, + 0.009451208636164665, + 0.0045444415882229805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.3020007908344269 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215399265289307 + ], + "opposite_logprob_tokens": [ + -1.502153992652893 + ], + "logprob_reward": -0.25215399265289307, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215399265289307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3259029686450958, + 0.06376223266124725, + 0.027592778205871582, + 0.04537628963589668, + 0.058536019176244736, + 0.23263821005821228, + 0.03953803703188896, + 0.06481238454580307, + 0.05100712180137634, + 0.09083397686481476 + ], + "credit_map": [ + 0.05600777640938759, + 0.025590233504772186, + 0.05238201841711998, + 0.02708970010280609, + 0.019760437309741974, + 0.05258984863758087, + 0.03021598979830742, + 0.062266942113637924, + 0.01386002916842699, + 0.004028947092592716, + 0.01715441793203354, + 0.007038440555334091, + 0.002612902782857418, + 0.005553051363676786, + 0.0072474307380616665, + 0.006267006043344736, + 0.005207263398915529, + 0.0024667344987392426, + 0.0038166919257491827, + 0.0028497688472270966, + 0.001984346890822053, + 0.0030141009483486414, + 0.003437054343521595, + 0.004816815722733736, + 0.007347077131271362, + 0.0031209022272378206, + 0.005651223938912153, + 0.004694284405559301, + 0.0033142310567200184, + 0.006026746705174446, + 0.007856049574911594, + 0.007365774363279343, + 0.010869020596146584, + 0.004238502122461796, + 0.011485269293189049, + 0.007228466682136059, + 0.003445700276643038, + 0.006247418001294136, + 0.00774892745539546, + 0.007272710092365742, + 0.06643139570951462, + 0.010838485322892666, + 0.01397751271724701, + 0.01535885687917471, + 0.010838554240763187, + 0.014902600087225437, + 0.08086652308702469, + 0.019424274563789368, + 0.00526256300508976, + 0.0029084980487823486, + 0.0031100758351385593, + 0.0071347737684845924, + 0.005557684227824211, + 0.005292628891766071, + 0.0059978533536195755, + 0.004273954778909683, + 0.005993711296468973, + 0.004811037797480822, + 0.009332959540188313, + 0.011014331132173538, + 0.00563441775739193, + 0.009883904829621315, + 0.008866778574883938, + 0.009275240823626518, + 0.004315149039030075, + 0.0027692886069417, + 0.009273633360862732, + 0.007555131800472736, + 0.0032889011781662703, + 0.009109978564083576, + 0.006663861218839884, + 0.008031174540519714, + 0.011533185839653015, + 0.006289200857281685, + 0.011433863081037998, + 0.013814443722367287, + 0.007068053353577852, + 0.012922353111207485, + 0.015290232375264168, + 0.012482638470828533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.25215399265289307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587796345353127 + ], + "opposite_logprob_tokens": [ + -3.895587682723999 + ], + "logprob_reward": -0.020587796345353127, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587796345353127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2003096342086792, + 0.17978373169898987, + 0.06390095502138138, + 0.05781874060630798, + 0.2006990760564804, + 0.10078278928995132, + 0.044215016067028046, + 0.083627849817276, + 0.03109685517847538, + 0.037765346467494965 + ], + "credit_map": [ + 0.027931593358516693, + 0.014568113721907139, + 0.026218365877866745, + 0.02050725929439068, + 0.016004355624318123, + 0.02280997484922409, + 0.03360142931342125, + 0.03866853564977646, + 0.04508756846189499, + 0.007766617927700281, + 0.055159490555524826, + 0.010512466542422771, + 0.008093818090856075, + 0.014629972167313099, + 0.015626559033989906, + 0.022907231003046036, + 0.014603468589484692, + 0.005557653494179249, + 0.007751998957246542, + 0.005480739753693342, + 0.005414970219135284, + 0.006687290966510773, + 0.01024709828197956, + 0.008157739415764809, + 0.009237997233867645, + 0.004001016728579998, + 0.00667188223451376, + 0.005814950913190842, + 0.005178156308829784, + 0.008794399909675121, + 0.008407068438827991, + 0.00971327256411314, + 0.038545917719602585, + 0.011035855859518051, + 0.03471416234970093, + 0.03029468283057213, + 0.00855950452387333, + 0.026394767686724663, + 0.024432582780718803, + 0.026721617206931114, + 0.02659638598561287, + 0.005090660881251097, + 0.007210944779217243, + 0.011854616925120354, + 0.004885623697191477, + 0.006395429372787476, + 0.02881329134106636, + 0.009935842826962471, + 0.00535613251850009, + 0.0024879612028598785, + 0.0047841984778642654, + 0.008038959465920925, + 0.0044995020143687725, + 0.007021354977041483, + 0.006075639743357897, + 0.0059512704610824585, + 0.007782499771565199, + 0.005617943592369556, + 0.010084055364131927, + 0.014384536072611809, + 0.0063936468213796616, + 0.013812699355185032, + 0.013082552701234818, + 0.012469914741814137, + 0.0035773194395005703, + 0.0019669297616928816, + 0.005280714016407728, + 0.0044412678107619286, + 0.0021754885092377663, + 0.005134712439030409, + 0.0043909489177167416, + 0.004129474516957998, + 0.004294210113584995, + 0.0023876179475337267, + 0.006046425085514784, + 0.004872071091085672, + 0.0030274793971329927, + 0.0058616166934370995, + 0.005771152675151825, + 0.005504772067070007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.020587796345353127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980329655110836 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.02980329655110836, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980329655110836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0760837197303772, + 0.07138688117265701, + 0.07324901968240738, + 0.03482615947723389, + 0.06791748851537704, + 0.26837342977523804, + 0.08183624595403671, + 0.1364688277244568, + 0.12014833092689514, + 0.06970994919538498 + ], + "credit_map": [ + 0.012564619071781635, + 0.005290956702083349, + 0.009307712316513062, + 0.0078002214431762695, + 0.008632097393274307, + 0.010408865287899971, + 0.012111177667975426, + 0.009968061000108719, + 0.011821435764431953, + 0.0043447003699839115, + 0.01913396269083023, + 0.006771753076463938, + 0.0037464818451553583, + 0.007908915169537067, + 0.009387945756316185, + 0.008271681144833565, + 0.015018987469375134, + 0.006372937932610512, + 0.007378544192761183, + 0.00837033148854971, + 0.006644326262176037, + 0.009314130060374737, + 0.011552649550139904, + 0.008597103878855705, + 0.007072610780596733, + 0.0024909896310418844, + 0.0060432651080191135, + 0.0034169238060712814, + 0.002054091077297926, + 0.004394340328872204, + 0.006023809313774109, + 0.0033301273360848427, + 0.011353513225913048, + 0.0034401840530335903, + 0.013438013382256031, + 0.01091885194182396, + 0.0028382318560034037, + 0.006823889445513487, + 0.008907589130103588, + 0.010197212919592857, + 0.07533366978168488, + 0.01214649248868227, + 0.017361896112561226, + 0.022954905405640602, + 0.011997018940746784, + 0.01866406388580799, + 0.08248897641897202, + 0.027426356449723244, + 0.009249797090888023, + 0.00616615591570735, + 0.009220706298947334, + 0.014903794042766094, + 0.008372719399631023, + 0.011530955322086811, + 0.012200621888041496, + 0.010191488079726696, + 0.014147511683404446, + 0.007754593621939421, + 0.019312379881739616, + 0.02590840309858322, + 0.009335590526461601, + 0.017748920246958733, + 0.021445218473672867, + 0.02081618458032608, + 0.011430436745285988, + 0.006017505656927824, + 0.020041193813085556, + 0.017693908885121346, + 0.0074816299602389336, + 0.023839367553591728, + 0.017770173028111458, + 0.015874100849032402, + 0.008195757865905762, + 0.003371255472302437, + 0.0124703673645854, + 0.008469773456454277, + 0.003752731950953603, + 0.011101941578090191, + 0.009827375411987305, + 0.012520745396614075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.02980329655110836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028177103959023952 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028177103959023952, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028177103959023952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17434561252593994, + 0.42916339635849, + 0.05861560255289078, + 0.056108612567186356, + 0.05076388269662857, + 0.04984318092465401, + 0.0642201378941536, + 0.05543220788240433, + 0.029213668778538704, + 0.03229362890124321 + ], + "credit_map": [ + 0.02744174748659134, + 0.014192755334079266, + 0.025164980441331863, + 0.02512657828629017, + 0.014939649030566216, + 0.022311467677354813, + 0.022648852318525314, + 0.022519590333104134, + 0.06654580682516098, + 0.03722350671887398, + 0.06537216901779175, + 0.0707884430885315, + 0.020967427641153336, + 0.051516562700271606, + 0.04562158137559891, + 0.07112796604633331, + 0.010705463588237762, + 0.00352690601721406, + 0.008123080246150494, + 0.012773578986525536, + 0.004489639308303595, + 0.005260889418423176, + 0.008125430904328823, + 0.0056106168776750565, + 0.008938450366258621, + 0.003950049635022879, + 0.008864297531545162, + 0.005257805343717337, + 0.003855512011796236, + 0.008721405640244484, + 0.007803469896316528, + 0.008717627264559269, + 0.008306693285703659, + 0.004480509553104639, + 0.006848101504147053, + 0.005441737361252308, + 0.003090674290433526, + 0.007501035463064909, + 0.00814575981348753, + 0.006949373055249453, + 0.013451830483973026, + 0.0023170514032244682, + 0.0031849034130573273, + 0.004656986333429813, + 0.0023481573443859816, + 0.0031573683954775333, + 0.015723587945103645, + 0.005003298167139292, + 0.007032605819404125, + 0.004514243919402361, + 0.007262786850333214, + 0.011249182745814323, + 0.007193455006927252, + 0.009155040606856346, + 0.010035132989287376, + 0.0077776978723704815, + 0.005355197004973888, + 0.00394989550113678, + 0.0068146162666380405, + 0.008210539817810059, + 0.005087651312351227, + 0.009377413429319859, + 0.008452421985566616, + 0.00818447396159172, + 0.002469634171575308, + 0.0018454000819474459, + 0.005110213533043861, + 0.005358661524951458, + 0.0017059465171769261, + 0.004942603409290314, + 0.004450032487511635, + 0.0033311783336102962, + 0.003800366073846817, + 0.0021650618873536587, + 0.0047814189456403255, + 0.004920803941786289, + 0.002116016112267971, + 0.004938952159136534, + 0.005550173111259937, + 0.004020838998258114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0028177103959023952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera slowly advances with a solemn and smooth motion, focusing on a luxurious fountain pen, resting in a Rubens-like shimmer of light and shadow. A warm, intense sidelight carves the sharp outline of the hexagonal white star inlaid on the cap tip and flows along the surface of the precious black resin of the pen barrel, creating a soft, jade-like luster. As the cap is slowly unscrewed by an invisible hand in an almost reverent slow motion, the gilded band connecting the pen barrel and cap reflects a rich, fiery hue, complementing the deep black resin. The camera zooms in, capturing the precise moment as the precision-machined metal threads gradually separate as they rotate. The edge of each thread shimmers in the light, a unique glimmer of mechanical beauty, while the other side sinks into a soft, bottomless shadow. The entire image is a supreme celebration of craftsmanship and quality, creating an atmosphere of focused, luxurious, and anticipatory tranquility.", + "seed": 28933, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_0/index_8959_seed_28933_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08472027219831943, + "vlm_logprob": -0.08472027219831943, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.08472027219831943, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2335749864578247 + ], + "opposite_logprob_tokens": [ + -1.6085749864578247 + ], + "logprob_reward": -0.2335749864578247, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2335749864578247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29792657494544983, + 0.06098495051264763, + 0.03843545913696289, + 0.1740071326494217, + 0.03147488832473755, + 0.03864189237356186, + 0.11932353675365448, + 0.046851180493831635, + 0.1269380748271942, + 0.06541629135608673 + ], + "credit_map": [ + 0.023422982543706894, + 0.02665974758565426, + 0.017058150842785835, + 0.03221089020371437, + 0.028422774747014046, + 0.03148753568530083, + 0.061594657599925995, + 0.07706984877586365, + 0.005835605785250664, + 0.006588851101696491, + 0.003734619589522481, + 0.004787479527294636, + 0.0060898903757333755, + 0.011269384063780308, + 0.012696624733507633, + 0.009982502087950706, + 0.003633319167420268, + 0.003914460074156523, + 0.0042020296677947044, + 0.005195540841668844, + 0.002939419588074088, + 0.004496157169342041, + 0.009559649042785168, + 0.00449488777667284, + 0.013770357705652714, + 0.018900901079177856, + 0.017511045560240746, + 0.020746896043419838, + 0.014673552475869656, + 0.04316478967666626, + 0.01995345577597618, + 0.02528616413474083, + 0.002806903561577201, + 0.002956024371087551, + 0.003563957056030631, + 0.003952730447053909, + 0.0038242232985794544, + 0.004289217758923769, + 0.006148267071694136, + 0.003933569882065058, + 0.003534559393301606, + 0.0040888357907533646, + 0.004890404641628265, + 0.00577383441850543, + 0.003354176878929138, + 0.007194701582193375, + 0.00553266191855073, + 0.004272721707820892, + 0.011337721720337868, + 0.007997112348675728, + 0.01820209063589573, + 0.014081129804253578, + 0.01060168445110321, + 0.02323739044368267, + 0.021774206310510635, + 0.012092203833162785, + 0.005442937836050987, + 0.004412105772644281, + 0.005497717764228582, + 0.005678629968315363, + 0.00443591084331274, + 0.0065266708843410015, + 0.009438245557248592, + 0.005418962799012661, + 0.010686546564102173, + 0.015762707218527794, + 0.017799144610762596, + 0.01429139543324709, + 0.010167066007852554, + 0.01521832961589098, + 0.023953942582011223, + 0.0190589539706707, + 0.005202427040785551, + 0.009363518096506596, + 0.009175743907690048, + 0.011121324263513088, + 0.005959327798336744, + 0.006188570987433195, + 0.011761542409658432, + 0.006643843837082386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.2335749864578247 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270982325077057 + ], + "opposite_logprob_tokens": [ + -2.127098321914673 + ], + "logprob_reward": -0.1270982325077057, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270982325077057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20206458866596222, + 0.14260312914848328, + 0.0398859828710556, + 0.08762571215629578, + 0.0740882083773613, + 0.13766567409038544, + 0.0599958598613739, + 0.10879401117563248, + 0.0883457213640213, + 0.05893109738826752 + ], + "credit_map": [ + 0.016900837421417236, + 0.016029028221964836, + 0.014064444229006767, + 0.021203014999628067, + 0.026678841561079025, + 0.018957478925585747, + 0.05265786498785019, + 0.035573095083236694, + 0.011012737639248371, + 0.009208054281771183, + 0.007964547723531723, + 0.012466248124837875, + 0.011664936318993568, + 0.020642295479774475, + 0.04406373202800751, + 0.025580575689673424, + 0.004210835322737694, + 0.0038077854551374912, + 0.004057000391185284, + 0.005550615023821592, + 0.003358968999236822, + 0.007073183543980122, + 0.0065593114122748375, + 0.005268283654004335, + 0.009970086626708508, + 0.010126492008566856, + 0.00963136088103056, + 0.012817967683076859, + 0.009015219286084175, + 0.014538055285811424, + 0.014805938117206097, + 0.006720588076859713, + 0.007487018126994371, + 0.00595744838938117, + 0.009360346011817455, + 0.007744475733488798, + 0.010944487527012825, + 0.009613593108952045, + 0.014461614191532135, + 0.008519218303263187, + 0.014504514634609222, + 0.010876798070967197, + 0.0160968154668808, + 0.016539305448532104, + 0.015585380606353283, + 0.019198104739189148, + 0.023214444518089294, + 0.021650314331054688, + 0.005936634261161089, + 0.004981152713298798, + 0.008885428309440613, + 0.0077655259519815445, + 0.006001915317028761, + 0.011014271527528763, + 0.008607841096818447, + 0.006803091615438461, + 0.013059527613222599, + 0.009543146006762981, + 0.012554023414850235, + 0.012032601051032543, + 0.009952528402209282, + 0.013616751879453659, + 0.02127278782427311, + 0.016762644052505493, + 0.009006082080304623, + 0.0066101024858653545, + 0.012565580196678638, + 0.012120399624109268, + 0.006317519582808018, + 0.012428323738276958, + 0.015973810106515884, + 0.013323904946446419, + 0.005251253955066204, + 0.00421125115826726, + 0.009651988744735718, + 0.007204299792647362, + 0.004483737517148256, + 0.008627010509371758, + 0.010071557015180588, + 0.009430000558495522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.1270982325077057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587678998708725 + ], + "opposite_logprob_tokens": [ + -3.895587682723999 + ], + "logprob_reward": -0.020587678998708725, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587678998708725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17574340105056763, + 0.07549004256725311, + 0.03803938627243042, + 0.15234293043613434, + 0.04301874339580536, + 0.09079258143901825, + 0.06471279263496399, + 0.11855810880661011, + 0.10468778014183044, + 0.13661418855190277 + ], + "credit_map": [ + 0.012495909817516804, + 0.0121177202090621, + 0.008775823749601841, + 0.019267689436674118, + 0.015972565859556198, + 0.02238304540514946, + 0.039925605058670044, + 0.04480506107211113, + 0.007911807857453823, + 0.00765931885689497, + 0.009167180396616459, + 0.012621771544218063, + 0.007459310349076986, + 0.009936063550412655, + 0.011455779895186424, + 0.00927880872040987, + 0.0036423709243535995, + 0.003139408538118005, + 0.0030180993489921093, + 0.005823548883199692, + 0.004168081562966108, + 0.008464666083455086, + 0.00551766948774457, + 0.0042655435390770435, + 0.012199165299534798, + 0.013751372694969177, + 0.017243407666683197, + 0.018041174858808517, + 0.02160748839378357, + 0.017227087169885635, + 0.030401352792978287, + 0.021871890872716904, + 0.004032824654132128, + 0.0038646385073661804, + 0.007586223538964987, + 0.004253936465829611, + 0.004537186585366726, + 0.0070291003212332726, + 0.005241993349045515, + 0.006472841836512089, + 0.009786914102733135, + 0.007953115738928318, + 0.011209146119654179, + 0.012099463492631912, + 0.008084072731435299, + 0.012118203565478325, + 0.012273048050701618, + 0.017268624156713486, + 0.006211988162249327, + 0.004478129558265209, + 0.009115951135754585, + 0.009473588317632675, + 0.005135851912200451, + 0.01154891774058342, + 0.011784165166318417, + 0.0069642034359276295, + 0.01021159440279007, + 0.01017723698168993, + 0.01483682170510292, + 0.013897164724767208, + 0.01225616130977869, + 0.015400237403810024, + 0.02365921624004841, + 0.018119679763913155, + 0.012686939910054207, + 0.01210771594196558, + 0.015165749937295914, + 0.011936817318201065, + 0.007472698576748371, + 0.010883052833378315, + 0.013349027372896671, + 0.02108577825129032, + 0.014326250180602074, + 0.012399250641465187, + 0.019737493246793747, + 0.020912352949380875, + 0.010910405777394772, + 0.021417025476694107, + 0.022065328434109688, + 0.014846086502075195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.020587678998708725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033704668283462524 + ], + "opposite_logprob_tokens": [ + -3.4087047576904297 + ], + "logprob_reward": -0.033704668283462524, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033704668283462524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12348870933055878, + 0.07687412947416306, + 0.047451939433813095, + 0.29675012826919556, + 0.07701733708381653, + 0.08544132113456726, + 0.062394894659519196, + 0.075398288667202, + 0.04487483575940132, + 0.110308438539505 + ], + "credit_map": [ + 0.013197067193686962, + 0.014109942130744457, + 0.011328083463013172, + 0.012965847738087177, + 0.020374715328216553, + 0.01348028052598238, + 0.018912749364972115, + 0.019120018929243088, + 0.008449729532003403, + 0.00825130008161068, + 0.007253910880535841, + 0.009530264884233475, + 0.007118803448975086, + 0.013634861446917057, + 0.012878045439720154, + 0.009757203049957752, + 0.004677950870245695, + 0.003648309735581279, + 0.005309548694640398, + 0.005596145987510681, + 0.003727262606844306, + 0.00459189573302865, + 0.009759020991623402, + 0.010141801089048386, + 0.025598712265491486, + 0.023668918758630753, + 0.03140563517808914, + 0.033106233924627304, + 0.02325478196144104, + 0.08052629977464676, + 0.032588519155979156, + 0.04660097137093544, + 0.007702637929469347, + 0.0053249867632985115, + 0.013613959774374962, + 0.010610328055918217, + 0.006642851047217846, + 0.007947600446641445, + 0.012262638658285141, + 0.012912329286336899, + 0.01064217183738947, + 0.009602564387023449, + 0.009981554001569748, + 0.009859541431069374, + 0.009776576422154903, + 0.011908899061381817, + 0.01212729699909687, + 0.01154270302504301, + 0.007214044686406851, + 0.004747319035232067, + 0.008627750910818577, + 0.00792032852768898, + 0.004934700205922127, + 0.01122589036822319, + 0.008362053893506527, + 0.009362800046801567, + 0.006357515696436167, + 0.006827909965068102, + 0.010439510457217693, + 0.010787969455122948, + 0.00967694353312254, + 0.008319826796650887, + 0.014219498261809349, + 0.008769107982516289, + 0.0046069021336734295, + 0.005039148032665253, + 0.006649137940257788, + 0.005939976312220097, + 0.0035775997675955296, + 0.0066758692264556885, + 0.005374938249588013, + 0.007011258043348789, + 0.011377069167792797, + 0.009465436451137066, + 0.016633590683341026, + 0.015382163226604462, + 0.006845285650342703, + 0.016392583027482033, + 0.02101241424679756, + 0.013199876062572002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.033704668283462524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fountain pen with a black resin barrel and a hexagonal white star inlaid on the cap tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show the pen cap being unscrewed by an invisible hand in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a visible gilded band connecting the pen barrel and cap that reflects a rich, fiery hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera advance slowly towards the pen while also zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by a warm, intense sidelight creating sharp outlines and soft, jade-like luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863579474389553 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863579474389553, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863579474389553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1534116119146347, + 0.08620838075876236, + 0.06064249947667122, + 0.05016256868839264, + 0.06283793598413467, + 0.18434493243694305, + 0.06599776446819305, + 0.12158868461847305, + 0.09118775278329849, + 0.12361779063940048 + ], + "credit_map": [ + 0.016013726592063904, + 0.018439121544361115, + 0.009411657229065895, + 0.02162795513868332, + 0.017378363758325577, + 0.020579826086759567, + 0.023946192115545273, + 0.026014765724539757, + 0.00822978001087904, + 0.008701195940375328, + 0.005541934631764889, + 0.009352240711450577, + 0.008529028855264187, + 0.017453061416745186, + 0.014585481956601143, + 0.013815653510391712, + 0.005944557022303343, + 0.004458344075828791, + 0.004907755646854639, + 0.0077435169368982315, + 0.004435229115188122, + 0.008120686747133732, + 0.01758863963186741, + 0.0074437703005969524, + 0.004718303214758635, + 0.004961724393069744, + 0.006185321602970362, + 0.00750195886939764, + 0.006565789692103863, + 0.00461632339283824, + 0.00867560412734747, + 0.0069375415332615376, + 0.006109077017754316, + 0.005912206135690212, + 0.012320799753069878, + 0.00668914383277297, + 0.0044186473824083805, + 0.008584028109908104, + 0.008850714191794395, + 0.009953321889042854, + 0.017778975889086723, + 0.01866580918431282, + 0.01968245394527912, + 0.025440560653805733, + 0.016985349357128143, + 0.03554439917206764, + 0.02597338706254959, + 0.024274000898003578, + 0.00603752676397562, + 0.005661578383296728, + 0.009076307527720928, + 0.008253098465502262, + 0.004937345627695322, + 0.01219745073467493, + 0.012048512697219849, + 0.007785947527736425, + 0.01437685638666153, + 0.010369028896093369, + 0.01472433377057314, + 0.016455145552754402, + 0.009923461824655533, + 0.015094035305082798, + 0.0208243727684021, + 0.019821448251605034, + 0.010727709159255028, + 0.00865309126675129, + 0.013424777425825596, + 0.013353679329156876, + 0.0056189559400081635, + 0.010225361213088036, + 0.012802839279174805, + 0.016381336376070976, + 0.010670977644622326, + 0.014484474435448647, + 0.02023278921842575, + 0.016866279765963554, + 0.007203908171504736, + 0.014470888301730156, + 0.019354145973920822, + 0.020334342494606972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.00863579474389553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_tensor([9638, 9638])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_tensor([9638, 9638])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..63b02990c966839ce0e66bf8f93da20ae5513a08 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_tensor([9638, 9638])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "highly immersive, first-person, subjective, and static shot begins with a slender, white hand slowly reaching forward from the bottom of the frame, as if reaching for something invisible. The background is a nearly pitch-black void, devoid of any discernible objects. As the hand reaches out, countless fragments of memory, shimmering with a soft golden glow, begin to converge from all directions into the space in front of the palm, as if drawn by an invisible gravitational force. These luminous particles, like living stardust, swirl and dance in the air before carefully, steadily, and deliberately approaching and merging. Their light provides the sole source of illumination, casting a warm, dynamic shadow on the outstretched hand. Ultimately, these points of light coalesce and intertwine into a hazy, translucent figure composed of pure light, its facial features difficult to discern. The contrast between the profound darkness and the warm golden glow creates a tranquil, dreamy, and deeply nostalgic atmosphere.", + "seed": 31164, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_9638_seed_31164_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019838265667203814, + "vlm_logprob": -0.019838265667203814, + "vlm_yes_no_margin": 5.65, + "vlm_token_credit": -0.019838265667203814, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09010663628578186 + ], + "opposite_logprob_tokens": [ + -2.715106725692749 + ], + "logprob_reward": -0.09010663628578186, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09010663628578186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12113645672798157, + 0.2538340985774994, + 0.03266362100839615, + 0.019875189289450645, + 0.09085573256015778, + 0.0625508725643158, + 0.14880342781543732, + 0.03133397549390793, + 0.21558056771755219, + 0.023366110399365425 + ], + "credit_map": [ + 0.01824738085269928, + 0.014699866063892841, + 0.018615104258060455, + 0.0215592198073864, + 0.014219926670193672, + 0.014832882210612297, + 0.012251436710357666, + 0.006710636429488659, + 0.013908866792917252, + 0.008003566414117813, + 0.10493108630180359, + 0.005438476800918579, + 0.005688095930963755, + 0.10668440908193588, + 0.004005396272987127, + 0.005174186080694199, + 0.009442298673093319, + 0.002820331836119294, + 0.005173713434487581, + 0.002637361641973257, + 0.002035083482041955, + 0.0047044213861227036, + 0.0025716507807374, + 0.0032787558156996965, + 0.003019785974174738, + 0.002549112541601062, + 0.003718255553394556, + 0.001274179550819099, + 0.002680222038179636, + 0.003558446653187275, + 0.0016045388765633106, + 0.001470646122470498, + 0.008633024990558624, + 0.004295961931347847, + 0.026436004787683487, + 0.004480086266994476, + 0.005326430778950453, + 0.024936703965067863, + 0.006942396983504295, + 0.009805113077163696, + 0.007876992225646973, + 0.00852691289037466, + 0.010004835203289986, + 0.0046984110958874226, + 0.010693777352571487, + 0.010494008660316467, + 0.006121090613305569, + 0.004134838934987783, + 0.017043940722942352, + 0.011842921376228333, + 0.040120452642440796, + 0.007596473675221205, + 0.04422474279999733, + 0.011061001569032669, + 0.009132623672485352, + 0.007781254127621651, + 0.003758983453735709, + 0.0037173063028603792, + 0.005672539118677378, + 0.003607309190556407, + 0.00416050897911191, + 0.004097023978829384, + 0.0031104946974664927, + 0.003209806513041258, + 0.03922472521662712, + 0.012508848682045937, + 0.032155491411685944, + 0.02490352839231491, + 0.017810354009270668, + 0.03517719730734825, + 0.02544357441365719, + 0.028356820344924927, + 0.0026200171560049057, + 0.002036530524492264, + 0.003259453922510147, + 0.0029672745149582624, + 0.0026976889930665493, + 0.0029441548977047205, + 0.003420233028009534, + 0.0034207545686513186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.09010663628578186 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003209085436537862 + ], + "opposite_logprob_tokens": [ + -5.753209114074707 + ], + "logprob_reward": -0.003209085436537862, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003209085436537862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06419240683317184, + 0.855909526348114, + 0.008673766627907753, + 0.009358675219118595, + 0.011158636771142483, + 0.01069329772144556, + 0.007286119274795055, + 0.006356296129524708, + 0.01862476021051407, + 0.007746465038508177 + ], + "credit_map": [ + 0.008315726183354855, + 0.01393373217433691, + 0.010511007159948349, + 0.007163982838392258, + 0.005756325554102659, + 0.010023756884038448, + 0.004663619678467512, + 0.0038242642767727375, + 0.04540581628680229, + 0.023018231615424156, + 0.36270004510879517, + 0.015660950914025307, + 0.012531682848930359, + 0.36880195140838623, + 0.011462552472949028, + 0.01632835529744625, + 0.00160420429892838, + 0.0010148943401873112, + 0.001207677531056106, + 0.0006601557251997292, + 0.0011985868914052844, + 0.0011813773307949305, + 0.0009100805618800223, + 0.000896790181286633, + 0.0022984726820141077, + 0.0009340830729342997, + 0.0021448612678796053, + 0.0005404057446867228, + 0.0008240696042776108, + 0.001247091800905764, + 0.0006550797843374312, + 0.0007146107964217663, + 0.0020240831654518843, + 0.0013103222008794546, + 0.0017499082023277879, + 0.0007821281906217337, + 0.00124663719907403, + 0.0014312478015199304, + 0.0010086789261549711, + 0.0016056317836046219, + 0.0014456524513661861, + 0.0013755070976912975, + 0.0017169738421216607, + 0.0007718041306361556, + 0.0017245972994714975, + 0.0019608307629823685, + 0.0009852187940850854, + 0.0007127127610146999, + 0.0008508316241204739, + 0.0007058626506477594, + 0.0014852642780169845, + 0.0006931241950951517, + 0.001666160998865962, + 0.0006602423964068294, + 0.0007018162868916988, + 0.0005228165537118912, + 0.0009074234985746443, + 0.0004586957220453769, + 0.0007958188070915639, + 0.0007782954489812255, + 0.0009483793401159346, + 0.0011661642929539084, + 0.000522791116964072, + 0.0007787277572788298, + 0.0033888707403093576, + 0.0011304623913019896, + 0.0026208790950477123, + 0.002119756070896983, + 0.001659013913013041, + 0.0029920218512415886, + 0.0022500641644001007, + 0.0024636934977024794, + 0.0011243245098739862, + 0.0005520814447663724, + 0.0010818035807460546, + 0.0007296238327398896, + 0.0009052457171492279, + 0.0009272123570553958, + 0.0009414561791345477, + 0.0014847174752503633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.003209085436537862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013392536202445626 + ], + "opposite_logprob_tokens": [ + -6.626339435577393 + ], + "logprob_reward": -0.0013392536202445626, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013392536202445626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08117885887622833, + 0.2564946115016937, + 0.0775262713432312, + 0.06006412208080292, + 0.20922578871250153, + 0.05292385071516037, + 0.1185612604022026, + 0.04222189635038376, + 0.05727890878915787, + 0.04452439025044441 + ], + "credit_map": [ + 0.008988091722130775, + 0.022983161732554436, + 0.01153549738228321, + 0.009603306651115417, + 0.0066492678597569466, + 0.011816599406301975, + 0.006032071076333523, + 0.003570869565010071, + 0.015296232886612415, + 0.008500557392835617, + 0.10605922341346741, + 0.00511793652549386, + 0.004407137166708708, + 0.1072642058134079, + 0.004764274228364229, + 0.005085044540464878, + 0.021645264700055122, + 0.005960115231573582, + 0.014911283738911152, + 0.004108130000531673, + 0.003915369510650635, + 0.012404341250658035, + 0.004673204384744167, + 0.009908569976687431, + 0.013585726730525494, + 0.0037308114115148783, + 0.01477004587650299, + 0.002782003255560994, + 0.0064531853422522545, + 0.009949865750968456, + 0.003768378170207143, + 0.005024108104407787, + 0.023510167375206947, + 0.008656641468405724, + 0.06068948656320572, + 0.00892007164657116, + 0.013780321925878525, + 0.061487723141908646, + 0.014571829698979855, + 0.017609575763344765, + 0.0075392513535916805, + 0.005952719133347273, + 0.009471369907259941, + 0.00412559462711215, + 0.00845508836209774, + 0.009004373103380203, + 0.005389178637415171, + 0.002986280480399728, + 0.01435309462249279, + 0.010826232843101025, + 0.03220945596694946, + 0.006733226124197245, + 0.03214770182967186, + 0.00879001896828413, + 0.0069367410615086555, + 0.006564800161868334, + 0.008529894053936005, + 0.004386913497000933, + 0.006894352845847607, + 0.003610692685469985, + 0.00442101014778018, + 0.007072352338582277, + 0.0034414930269122124, + 0.003865189151838422, + 0.011313050054013729, + 0.0037424960173666477, + 0.008859788998961449, + 0.006295191589742899, + 0.00481201708316803, + 0.009227320551872253, + 0.005918990354984999, + 0.007110056933015585, + 0.006154073402285576, + 0.0035283078905194998, + 0.006285610143095255, + 0.0044643147848546505, + 0.005086405202746391, + 0.00556547986343503, + 0.005033187568187714, + 0.008407010696828365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.0013392536202445626 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004391897819004953 + ], + "opposite_logprob_tokens": [ + -7.750439167022705 + ], + "logprob_reward": -0.0004391897819004953, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004391897819004953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0695180892944336, + 0.4529915153980255, + 0.018801577389240265, + 0.027803247794508934, + 0.05149533599615097, + 0.017078885808587074, + 0.08201312273740768, + 0.04517601057887077, + 0.17411543428897858, + 0.06100677698850632 + ], + "credit_map": [ + 0.006610432639718056, + 0.012021289207041264, + 0.013759098947048187, + 0.008416028693318367, + 0.006490375380963087, + 0.012006796896457672, + 0.00668650958687067, + 0.003527554450556636, + 0.02218877710402012, + 0.011097023263573647, + 0.1934860348701477, + 0.008160476572811604, + 0.00665549049153924, + 0.19680233299732208, + 0.005910688545554876, + 0.008690676651895046, + 0.003235012758523226, + 0.0021958474535495043, + 0.0032274143304675817, + 0.0020593907684087753, + 0.0011976681416854262, + 0.0033262288197875023, + 0.0016729658236727118, + 0.001887048943899572, + 0.006279221270233393, + 0.0020811189897358418, + 0.007503147702664137, + 0.0010637014638632536, + 0.0034266619477421045, + 0.00421951012685895, + 0.0013514800230041146, + 0.0018784055719152093, + 0.006972853094339371, + 0.0035350306425243616, + 0.010110271163284779, + 0.0030431824270635843, + 0.005723345093429089, + 0.00873279944062233, + 0.00567267881706357, + 0.007705173455178738, + 0.0025513351429253817, + 0.002224925672635436, + 0.0026918030343949795, + 0.001404240378178656, + 0.0025102877989411354, + 0.002181126968935132, + 0.0017574660014361143, + 0.0017576988320797682, + 0.009365739300847054, + 0.006469342391937971, + 0.021536067128181458, + 0.004616554360836744, + 0.024070214480161667, + 0.0061254724860191345, + 0.005302430130541325, + 0.0045272973366081715, + 0.006079963408410549, + 0.0035739371087402105, + 0.004632317461073399, + 0.0055879391729831696, + 0.005553812254220247, + 0.008034711703658104, + 0.0048154680989682674, + 0.006897859275341034, + 0.02741091698408127, + 0.01707402803003788, + 0.018046706914901733, + 0.01932394877076149, + 0.020848479121923447, + 0.024824591353535652, + 0.02212747000157833, + 0.024459272623062134, + 0.00755081232637167, + 0.005154874641448259, + 0.008382239378988743, + 0.007089436519891024, + 0.004763408564031124, + 0.007947401143610477, + 0.010850883089005947, + 0.009267720393836498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.0004391897819004953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097163211554289 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004097163211554289, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097163211554289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3048102855682373, + 0.13671836256980896, + 0.023454587906599045, + 0.03239428251981735, + 0.03913484141230583, + 0.02228930965065956, + 0.06291726231575012, + 0.1084028109908104, + 0.20332060754299164, + 0.0665576383471489 + ], + "credit_map": [ + 0.04178770259022713, + 0.04342346638441086, + 0.0441686287522316, + 0.06162606552243233, + 0.03150269016623497, + 0.038929425179958344, + 0.026536909863352776, + 0.016835425049066544, + 0.006915244273841381, + 0.004920170642435551, + 0.05491594597697258, + 0.003644860815256834, + 0.0033784036058932543, + 0.05620112270116806, + 0.003541533136740327, + 0.003201096784323454, + 0.005515031982213259, + 0.0020821960642933846, + 0.004554219078272581, + 0.0019997626077383757, + 0.001298986142501235, + 0.004515990614891052, + 0.001592968706972897, + 0.0018954349216073751, + 0.008047642186284065, + 0.0034568337723612785, + 0.007124083582311869, + 0.0020727752707898617, + 0.0034241382963955402, + 0.003919567447155714, + 0.0017574175726622343, + 0.00259182951413095, + 0.004068247973918915, + 0.003401659894734621, + 0.005484574940055609, + 0.0029230057261884212, + 0.0047754934057593346, + 0.005003820173442364, + 0.004981279373168945, + 0.008496762253344059, + 0.00361124356277287, + 0.002258667489513755, + 0.004268981982022524, + 0.0021621081978082657, + 0.002238944172859192, + 0.003331520361825824, + 0.0021332171745598316, + 0.0022846299689263105, + 0.0072890897281467915, + 0.0060353828594088554, + 0.016462886705994606, + 0.003650019643828273, + 0.017497321590781212, + 0.004683453589677811, + 0.0038669041823595762, + 0.0034322107676416636, + 0.017085954546928406, + 0.009330889210104942, + 0.027096612378954887, + 0.009123405441641808, + 0.009739646688103676, + 0.019332509487867355, + 0.0062682437710464, + 0.010425561107695103, + 0.046390604227781296, + 0.012778190895915031, + 0.03800591081380844, + 0.018371889367699623, + 0.012485051527619362, + 0.037234432995319366, + 0.015499360859394073, + 0.02255518175661564, + 0.007561719510704279, + 0.005929411854594946, + 0.007604807149618864, + 0.007493611425161362, + 0.007242203690111637, + 0.007837513461709023, + 0.011003079824149609, + 0.011885300278663635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.004097163211554289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly immersive, first-person, subjective, and static shot begins with a slender, white hand slowly reaching forward from the bottom of the frame, as if reaching for something invisible. The background is a nearly pitch-black void, devoid of any discernible objects. As the hand reaches out, countless fragments of memory, shimmering with a soft golden glow, begin to converge from all directions into the space in front of the palm, as if drawn by an invisible gravitational force. These luminous particles, like living stardust, swirl and dance in the air before carefully, steadily, and deliberately approaching and merging. Their light provides the sole source of illumination, casting a warm, dynamic shadow on the outstretched hand. Ultimately, these points of light coalesce and intertwine into a hazy, translucent figure composed of pure light, its facial features difficult to discern. The contrast between the profound darkness and the warm golden glow creates a tranquil, dreamy, and deeply nostalgic atmosphere.", + "seed": 31165, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_9638_seed_31165_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018860644148662686, + "vlm_logprob": -0.018860644148662686, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.018860644148662686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07748778164386749 + ], + "opposite_logprob_tokens": [ + -2.7024877071380615 + ], + "logprob_reward": -0.07748778164386749, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07748778164386749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2036886215209961, + 0.21436439454555511, + 0.168638214468956, + 0.09611731022596359, + 0.10758355259895325, + 0.04213026538491249, + 0.03221893683075905, + 0.04762992635369301, + 0.023986708372831345, + 0.06364203244447708 + ], + "credit_map": [ + 0.01789136976003647, + 0.017405033111572266, + 0.011165907606482506, + 0.01746211014688015, + 0.059587616473436356, + 0.055248457938432693, + 0.012974601238965988, + 0.011953532695770264, + 0.017451733350753784, + 0.012218447402119637, + 0.013538328930735588, + 0.01827062852680683, + 0.068942129611969, + 0.04759937524795532, + 0.01765919104218483, + 0.018684569746255875, + 0.013916920870542526, + 0.017191173508763313, + 0.01849919743835926, + 0.01962502859532833, + 0.02375025674700737, + 0.03466367349028587, + 0.02027994580566883, + 0.020712019875645638, + 0.007354117929935455, + 0.011340122669935226, + 0.00996377319097519, + 0.01100335456430912, + 0.013999172486364841, + 0.020242074504494667, + 0.010096099227666855, + 0.01211860217154026, + 0.017783593386411667, + 0.010126644745469093, + 0.019005220383405685, + 0.012931222096085548, + 0.011795719154179096, + 0.01506154052913189, + 0.00869937613606453, + 0.012180235236883163, + 0.0056505631655454636, + 0.004110374487936497, + 0.006075880490243435, + 0.005508781876415014, + 0.006666114088147879, + 0.005471161566674709, + 0.0038742453325539827, + 0.004773146938532591, + 0.0033620004542171955, + 0.0029719904996454716, + 0.0037020710296928883, + 0.003383486997336149, + 0.004507537931203842, + 0.004502632189542055, + 0.004053619224578142, + 0.005735601764172316, + 0.005674324929714203, + 0.0035013200249522924, + 0.004050520248711109, + 0.008379580453038216, + 0.005366953555494547, + 0.00638446444645524, + 0.006406247615814209, + 0.007866516709327698, + 0.003941952250897884, + 0.0018497430719435215, + 0.002643568441271782, + 0.0017236369894817472, + 0.0025517228059470654, + 0.00415897136554122, + 0.0024369610473513603, + 0.0046801515854895115, + 0.0057048979215323925, + 0.0023637409321963787, + 0.004329540766775608, + 0.002688721055164933, + 0.002632180927321315, + 0.020758505910634995, + 0.0026375569868832827, + 0.02252688631415367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07748778164386749 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028325694147497416 + ], + "opposite_logprob_tokens": [ + -5.877832412719727 + ], + "logprob_reward": -0.0028325694147497416, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028325694147497416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12020403146743774, + 0.13881471753120422, + 0.23533061146736145, + 0.08694588392972946, + 0.07441061735153198, + 0.12809379398822784, + 0.073776975274086, + 0.03671944513916969, + 0.02153831534087658, + 0.0841655358672142 + ], + "credit_map": [ + 0.007052757777273655, + 0.007564907427877188, + 0.007241137325763702, + 0.008717167191207409, + 0.042928192764520645, + 0.031005844473838806, + 0.007947416976094246, + 0.007746610324829817, + 0.011985198594629765, + 0.010318899527192116, + 0.010511129163205624, + 0.015419963747262955, + 0.041798077523708344, + 0.02925005741417408, + 0.01030440628528595, + 0.00922698900103569, + 0.016307024285197258, + 0.019423911347985268, + 0.01870046928524971, + 0.022051796317100525, + 0.03412099927663803, + 0.06442806869745255, + 0.026770560070872307, + 0.03352779150009155, + 0.013285265304148197, + 0.014169779606163502, + 0.009257981553673744, + 0.0093001127243042, + 0.008967753499746323, + 0.01043495163321495, + 0.009157267399132252, + 0.01237277127802372, + 0.00968676432967186, + 0.008186325430870056, + 0.010363446548581123, + 0.008829662576317787, + 0.009930909611284733, + 0.009726723656058311, + 0.008297380991280079, + 0.009389406070113182, + 0.01755985990166664, + 0.01664728857576847, + 0.015371805988252163, + 0.014224831946194172, + 0.024566633626818657, + 0.016132032498717308, + 0.011456866748631, + 0.012134482152760029, + 0.005702564492821693, + 0.011851697228848934, + 0.00818402599543333, + 0.012580798007547855, + 0.008270871825516224, + 0.008381086401641369, + 0.009958197362720966, + 0.008847733959555626, + 0.004542837850749493, + 0.004068579059094191, + 0.003844307968392968, + 0.004816333297640085, + 0.004831396974623203, + 0.005087298806756735, + 0.004147760570049286, + 0.005380931310355663, + 0.0026755465660244226, + 0.002894175471737981, + 0.0019425853388383985, + 0.0026241515297442675, + 0.0022455581929534674, + 0.0026550167240202427, + 0.0033062368165701628, + 0.003195043420419097, + 0.007751135621219873, + 0.0034938149619847536, + 0.007362455129623413, + 0.0032124773133546114, + 0.0046765925362706184, + 0.025667330250144005, + 0.00355418398976326, + 0.02844754233956337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.0028325694147497416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036215689033269882 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036215689033269882, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036215689033269882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08933858573436737, + 0.11486118286848068, + 0.13541419804096222, + 0.0520002618432045, + 0.06755711138248444, + 0.1732284128665924, + 0.06933489441871643, + 0.04724479466676712, + 0.05791157856583595, + 0.19310902059078217 + ], + "credit_map": [ + 0.00753641314804554, + 0.007887129671871662, + 0.00672205351293087, + 0.008462116122245789, + 0.027277249842882156, + 0.018444878980517387, + 0.006064200773835182, + 0.006944537162780762, + 0.009710519574582577, + 0.007786688394844532, + 0.008040019311010838, + 0.011304852552711964, + 0.03482341393828392, + 0.027652841061353683, + 0.007985332980751991, + 0.007557509001344442, + 0.010413753800094128, + 0.011651558801531792, + 0.01055113971233368, + 0.013882971368730068, + 0.019792737439274788, + 0.03712664172053337, + 0.014338577166199684, + 0.017656812444329262, + 0.006370524875819683, + 0.005702860187739134, + 0.006396651268005371, + 0.0062111979350447655, + 0.00509897293522954, + 0.007653329987078905, + 0.007184999063611031, + 0.00738172372803092, + 0.00902114063501358, + 0.006179268471896648, + 0.009647365659475327, + 0.0057579707354307175, + 0.00794240366667509, + 0.010065711103379726, + 0.005942962132394314, + 0.013000283390283585, + 0.02012552320957184, + 0.019336916506290436, + 0.023840539157390594, + 0.01770995371043682, + 0.033949729055166245, + 0.02305462583899498, + 0.013485248200595379, + 0.021725865080952644, + 0.005770458839833736, + 0.011989830993115902, + 0.006088302005082369, + 0.010792921297252178, + 0.008372499607503414, + 0.008886742405593395, + 0.00873070489615202, + 0.008703431114554405, + 0.007158695254474878, + 0.004035558085888624, + 0.0040818266570568085, + 0.00549680320546031, + 0.005559096112847328, + 0.006934723816812038, + 0.005580048076808453, + 0.008398040197789669, + 0.009582084603607655, + 0.005990392062813044, + 0.004076750483363867, + 0.005383186973631382, + 0.00528021901845932, + 0.008747725747525692, + 0.0065653142519295216, + 0.01228590402752161, + 0.012847263365983963, + 0.007316030096262693, + 0.012682696804404259, + 0.011827301234006882, + 0.01066204160451889, + 0.05771895498037338, + 0.011515584774315357, + 0.06853914260864258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.0036215689033269882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742820143699646 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742820143699646, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742820143699646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1062977984547615, + 0.1545466035604477, + 0.051943983882665634, + 0.03181222826242447, + 0.07640071958303452, + 0.10813003033399582, + 0.03541908413171768, + 0.05723834037780762, + 0.143696129322052, + 0.23451508581638336 + ], + "credit_map": [ + 0.011981458403170109, + 0.008608396165072918, + 0.006810799241065979, + 0.009919623844325542, + 0.026127289980649948, + 0.025197653099894524, + 0.008904028683900833, + 0.008748541586101055, + 0.013498472049832344, + 0.011728978715837002, + 0.01202845387160778, + 0.017400510609149933, + 0.04502327740192413, + 0.033797018229961395, + 0.011133990250527859, + 0.009935885667800903, + 0.005881437100470066, + 0.005263359751552343, + 0.005553367082029581, + 0.004771401174366474, + 0.007188674062490463, + 0.010207855142652988, + 0.0065691182389855385, + 0.006508766673505306, + 0.005132748279720545, + 0.0037951671984046698, + 0.0037448296789079905, + 0.0032422421500086784, + 0.003516513155773282, + 0.003698439570143819, + 0.004857722204178572, + 0.003824562532827258, + 0.01153227686882019, + 0.008986838161945343, + 0.010516882874071598, + 0.00610051304101944, + 0.012983813881874084, + 0.009761554189026356, + 0.007177494931966066, + 0.009341336786746979, + 0.010616793297231197, + 0.012770832516252995, + 0.017078956589102745, + 0.013195674866437912, + 0.019125547260046005, + 0.014801302924752235, + 0.010171649977564812, + 0.010369261726737022, + 0.004248150158673525, + 0.003997009247541428, + 0.0031980513595044613, + 0.0033824110869318247, + 0.004455429967492819, + 0.0060217357240617275, + 0.0031300168484449387, + 0.00698627857491374, + 0.007689102087169886, + 0.004532272461801767, + 0.004924944136291742, + 0.005007042083889246, + 0.006870059762150049, + 0.0095034996047616, + 0.0062666116282343864, + 0.012444806285202503, + 0.026085684075951576, + 0.012208997271955013, + 0.010639413259923458, + 0.010951667092740536, + 0.012562080286443233, + 0.024371672421693802, + 0.014928806573152542, + 0.031947799026966095, + 0.016223594546318054, + 0.007750694639980793, + 0.017355185002088547, + 0.010599251836538315, + 0.01024980191141367, + 0.07580097019672394, + 0.010860067792236805, + 0.08567550778388977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.006742820143699646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036184806376695633 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036184806376695633, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036184806376695633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2433011680841446, + 0.09745466709136963, + 0.07804927229881287, + 0.031641919165849686, + 0.047179028391838074, + 0.04320962354540825, + 0.10558779537677765, + 0.12102408707141876, + 0.02294088713824749, + 0.20961156487464905 + ], + "credit_map": [ + 0.017636515200138092, + 0.017974961549043655, + 0.013225823640823364, + 0.020065106451511383, + 0.07358430325984955, + 0.07131122052669525, + 0.01579383760690689, + 0.013709408231079578, + 0.008658770471811295, + 0.006644182838499546, + 0.00691272784024477, + 0.009701503440737724, + 0.03026335872709751, + 0.022098157554864883, + 0.006743367295712233, + 0.006432604510337114, + 0.009667321108281612, + 0.0075326720252633095, + 0.00891896616667509, + 0.007281370460987091, + 0.010134274140000343, + 0.015664048492908478, + 0.008315913379192352, + 0.010534713044762611, + 0.004005418624728918, + 0.003065592609345913, + 0.003404468996450305, + 0.002800020622089505, + 0.0036453709471970797, + 0.005662937648594379, + 0.004295752849429846, + 0.0047623589634895325, + 0.00865384191274643, + 0.004563407506793737, + 0.006693690083920956, + 0.004782281816005707, + 0.006297890562564135, + 0.00673687644302845, + 0.003730901051312685, + 0.005720142275094986, + 0.004398316610604525, + 0.004709090106189251, + 0.007389670237898827, + 0.003540364559739828, + 0.007524058222770691, + 0.006145990453660488, + 0.003653624327853322, + 0.005848514847457409, + 0.017024371773004532, + 0.007847880944609642, + 0.01211816631257534, + 0.008504445664584637, + 0.009169614873826504, + 0.017077714204788208, + 0.00999941024929285, + 0.023846207186579704, + 0.013489826582372189, + 0.007327989675104618, + 0.007793530821800232, + 0.012727887369692326, + 0.011736394837498665, + 0.024489477276802063, + 0.013714926317334175, + 0.029744064435362816, + 0.003098054090514779, + 0.0015675477916374803, + 0.0026725868228822947, + 0.0023480153176933527, + 0.002084301318973303, + 0.003899515373632312, + 0.0024135210551321507, + 0.004857347812503576, + 0.0203350018709898, + 0.00784287229180336, + 0.012819727882742882, + 0.007979468442499638, + 0.009509910829365253, + 0.0705052837729454, + 0.009857989847660065, + 0.0707613155245781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0036184806376695633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly immersive, first-person, subjective, and static shot begins with a slender, white hand slowly reaching forward from the bottom of the frame, as if reaching for something invisible. The background is a nearly pitch-black void, devoid of any discernible objects. As the hand reaches out, countless fragments of memory, shimmering with a soft golden glow, begin to converge from all directions into the space in front of the palm, as if drawn by an invisible gravitational force. These luminous particles, like living stardust, swirl and dance in the air before carefully, steadily, and deliberately approaching and merging. Their light provides the sole source of illumination, casting a warm, dynamic shadow on the outstretched hand. Ultimately, these points of light coalesce and intertwine into a hazy, translucent figure composed of pure light, its facial features difficult to discern. The contrast between the profound darkness and the warm golden glow creates a tranquil, dreamy, and deeply nostalgic atmosphere.", + "seed": 31166, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_9638_seed_31166_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018744522472843528, + "vlm_logprob": -0.018744522472843528, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.018744522472843528, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06602754443883896 + ], + "opposite_logprob_tokens": [ + -3.1910276412963867 + ], + "logprob_reward": -0.06602754443883896, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06602754443883896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18621864914894104, + 0.07950268685817719, + 0.10680517554283142, + 0.040525250136852264, + 0.1011417880654335, + 0.07576760649681091, + 0.07543241232633591, + 0.08835996687412262, + 0.1259007453918457, + 0.12034570425748825 + ], + "credit_map": [ + 0.023958535864949226, + 0.01689707674086094, + 0.020738491788506508, + 0.022709554061293602, + 0.01922297105193138, + 0.031519655138254166, + 0.020836060866713524, + 0.030336303636431694, + 0.011949896812438965, + 0.007053114473819733, + 0.007138971239328384, + 0.0067230514250695705, + 0.007954727858304977, + 0.013804508373141289, + 0.00897343922406435, + 0.015904976055026054, + 0.017354093492031097, + 0.009183711372315884, + 0.010449515655636787, + 0.006790948566049337, + 0.01080167107284069, + 0.017735281959176064, + 0.008101742714643478, + 0.026388203725218773, + 0.004312149714678526, + 0.0035240079741925, + 0.004015927202999592, + 0.0035902338568121195, + 0.004552149213850498, + 0.005081863142549992, + 0.007026745472103357, + 0.008422170765697956, + 0.017543096095323563, + 0.006957240868359804, + 0.014425630681216717, + 0.009399128146469593, + 0.008245601318776608, + 0.017779860645532608, + 0.010231293737888336, + 0.016559932380914688, + 0.007575207389891148, + 0.006270675454288721, + 0.008400125429034233, + 0.009243631735444069, + 0.011787128634750843, + 0.010853542014956474, + 0.011561280116438866, + 0.010076012462377548, + 0.01152018178254366, + 0.005682924296706915, + 0.00737349409610033, + 0.00499641802161932, + 0.006675848737359047, + 0.015159458853304386, + 0.006611399818211794, + 0.017412688583135605, + 0.009039396420121193, + 0.011432920582592487, + 0.016070986166596413, + 0.00796421617269516, + 0.008124265819787979, + 0.01236109808087349, + 0.007681774906814098, + 0.0156853087246418, + 0.0094155203551054, + 0.010409904643893242, + 0.014666209928691387, + 0.026997841894626617, + 0.00589100643992424, + 0.01948358863592148, + 0.01665179431438446, + 0.022384880110621452, + 0.013402695767581463, + 0.006689068395644426, + 0.04347870871424675, + 0.007193959783762693, + 0.008747423067688942, + 0.013392515480518341, + 0.005792329553514719, + 0.02164900116622448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.06602754443883896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091836728155613 + ], + "opposite_logprob_tokens": [ + -4.511091709136963 + ], + "logprob_reward": -0.011091836728155613, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091836728155613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10864931344985962, + 0.11608237773180008, + 0.4054768681526184, + 0.05753077566623688, + 0.06309371441602707, + 0.031824830919504166, + 0.06281351298093796, + 0.022181963548064232, + 0.07290563732385635, + 0.0594409704208374 + ], + "credit_map": [ + 0.011822550557553768, + 0.008292287588119507, + 0.01653948426246643, + 0.011564071290194988, + 0.016519630327820778, + 0.016331246122717857, + 0.01100812666118145, + 0.01657191477715969, + 0.010142337530851364, + 0.015473905019462109, + 0.012812746688723564, + 0.013018041849136353, + 0.012953989207744598, + 0.02375185303390026, + 0.012949720956385136, + 0.014979785308241844, + 0.06432084739208221, + 0.02616705186665058, + 0.044728271663188934, + 0.020148776471614838, + 0.03449486568570137, + 0.0706527978181839, + 0.022366126999258995, + 0.12259814143180847, + 0.006203875876963139, + 0.005582151934504509, + 0.005478636361658573, + 0.004516609478741884, + 0.0069702183827757835, + 0.007996785454452038, + 0.009987854398787022, + 0.010794642381370068, + 0.007176403421908617, + 0.0058499486185610294, + 0.008289821445941925, + 0.006130020599812269, + 0.00941191054880619, + 0.010219194926321507, + 0.007118829060345888, + 0.008897583931684494, + 0.003604327794164419, + 0.0027573578990995884, + 0.005209869239479303, + 0.003236910793930292, + 0.004444754682481289, + 0.00447368249297142, + 0.004042267799377441, + 0.004055659286677837, + 0.009876317344605923, + 0.005864380858838558, + 0.009120967239141464, + 0.005880013573914766, + 0.006479880306869745, + 0.00828325655311346, + 0.007998477667570114, + 0.009310217574238777, + 0.0030071542132645845, + 0.0019560910295695066, + 0.0054939305409789085, + 0.002118254778906703, + 0.0022379597648978233, + 0.002647490706294775, + 0.0017222111346200109, + 0.00299887266010046, + 0.005808171816170216, + 0.0032174792140722275, + 0.02107301726937294, + 0.0037141419015824795, + 0.002512372098863125, + 0.007903826422989368, + 0.003404518123716116, + 0.025272104889154434, + 0.006496797315776348, + 0.0026431430596858263, + 0.02197350189089775, + 0.003619387047365308, + 0.004522588104009628, + 0.0064916349947452545, + 0.002821933012455702, + 0.010871986858546734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.011091836728155613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978719163686037 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.00978719163686037, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978719163686037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06951266527175903, + 0.1401204913854599, + 0.10279428958892822, + 0.042567178606987, + 0.08055742084980011, + 0.0891486182808876, + 0.1328333169221878, + 0.06144827976822853, + 0.2275383472442627, + 0.053479425609111786 + ], + "credit_map": [ + 0.008204563520848751, + 0.006322074681520462, + 0.010027111507952213, + 0.010219226591289043, + 0.007433518301695585, + 0.01084030233323574, + 0.007432304788380861, + 0.009033567272126675, + 0.018338823691010475, + 0.01711086370050907, + 0.01649080589413643, + 0.015045924112200737, + 0.016487013548612595, + 0.02050064504146576, + 0.014981883578002453, + 0.021164532750844955, + 0.015076082199811935, + 0.006838764529675245, + 0.010774273425340652, + 0.005900086835026741, + 0.00859774835407734, + 0.018443012610077858, + 0.006165321450680494, + 0.03099900111556053, + 0.0050512198358774185, + 0.0041956291534006596, + 0.004774260800331831, + 0.0030642191413789988, + 0.005822358652949333, + 0.004904449451714754, + 0.007625034544616938, + 0.007130007725208998, + 0.008181587792932987, + 0.006231782026588917, + 0.011364374309778214, + 0.007309009321033955, + 0.010080032050609589, + 0.013172323815524578, + 0.012986049056053162, + 0.011232268996536732, + 0.013082752004265785, + 0.0068814754486083984, + 0.01837022975087166, + 0.006813250947743654, + 0.007982537150382996, + 0.013548455201089382, + 0.009291188791394234, + 0.013178728520870209, + 0.00944541022181511, + 0.014472518116235733, + 0.011474880389869213, + 0.014936122111976147, + 0.011487381532788277, + 0.024548444896936417, + 0.018797876313328743, + 0.027670670300722122, + 0.008460425771772861, + 0.006304233334958553, + 0.013412125408649445, + 0.006851379759609699, + 0.004811427555978298, + 0.0069860294461250305, + 0.007448371034115553, + 0.007174287922680378, + 0.025346942245960236, + 0.007365388795733452, + 0.08482956141233444, + 0.023082615807652473, + 0.006873331498354673, + 0.028233034536242485, + 0.013144418597221375, + 0.038663070648908615, + 0.005566151812672615, + 0.004305580630898476, + 0.016777517274022102, + 0.0038469540886580944, + 0.003756915219128132, + 0.005557056050747633, + 0.0035923458635807037, + 0.010076905600726604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.00978719163686037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031973214354366064 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.0031973214354366064, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031973214354366064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03256525099277496, + 0.03257371857762337, + 0.17525321245193481, + 0.012075374834239483, + 0.09202592074871063, + 0.056911710649728775, + 0.07790369540452957, + 0.06634090095758438, + 0.37754133343696594, + 0.076808862388134 + ], + "credit_map": [ + 0.0035617086105048656, + 0.002693821908906102, + 0.004160577896982431, + 0.00439663790166378, + 0.004164198879152536, + 0.0044323899783194065, + 0.0027843827847391367, + 0.00637153210118413, + 0.005002351477742195, + 0.002465068595483899, + 0.0038523494731634855, + 0.0023702746257185936, + 0.002671545837074518, + 0.007128695026040077, + 0.002364249899983406, + 0.006719186436384916, + 0.02829160913825035, + 0.010164370760321617, + 0.01926848106086254, + 0.007557953242212534, + 0.012400451116263866, + 0.03261492773890495, + 0.008680257014930248, + 0.05627516657114029, + 0.0017268788069486618, + 0.0011088541941717267, + 0.0012979021994397044, + 0.001132057630456984, + 0.0015195553423836827, + 0.0018219177145510912, + 0.0014826447004452348, + 0.001985563663765788, + 0.00940411351621151, + 0.008574945852160454, + 0.013139392249286175, + 0.009807675145566463, + 0.013399774208664894, + 0.014188289642333984, + 0.012555724009871483, + 0.010956010781228542, + 0.008538789115846157, + 0.004705064930021763, + 0.008213345892727375, + 0.005582342389971018, + 0.007425790652632713, + 0.009183868765830994, + 0.005213899537920952, + 0.008048605173826218, + 0.007933511398732662, + 0.007925650104880333, + 0.00893743522465229, + 0.010075616650283337, + 0.007106484845280647, + 0.012035458348691463, + 0.010837570764124393, + 0.01305196899920702, + 0.006801679264754057, + 0.010935175232589245, + 0.009568948298692703, + 0.009036378934979439, + 0.00668379059061408, + 0.006851637735962868, + 0.008067980408668518, + 0.008395315147936344, + 0.03810815140604973, + 0.01308113057166338, + 0.13675571978092194, + 0.0410897359251976, + 0.01046830415725708, + 0.03959215059876442, + 0.02400325983762741, + 0.07444289326667786, + 0.008714971132576466, + 0.002481746254488826, + 0.02886069193482399, + 0.004192675929516554, + 0.006046869792044163, + 0.008637567050755024, + 0.00346738682128489, + 0.014406953006982803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.0031973214354366064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036187181249260902 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036187181249260902, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036187181249260902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1553908735513687, + 0.025007762014865875, + 0.24772116541862488, + 0.06680174916982651, + 0.03358560428023338, + 0.03584960475564003, + 0.09866072982549667, + 0.06612972915172577, + 0.07641244679689407, + 0.19444027543067932 + ], + "credit_map": [ + 0.032709281891584396, + 0.010938768275082111, + 0.0165939349681139, + 0.014632136560976505, + 0.015835294499993324, + 0.021521858870983124, + 0.014592533931136131, + 0.028567055240273476, + 0.002430930268019438, + 0.002351682400330901, + 0.0036280828062444925, + 0.0024482854641973972, + 0.0021759397350251675, + 0.004527371842414141, + 0.0027519583236426115, + 0.004693512339144945, + 0.03833756968379021, + 0.014512503519654274, + 0.027930043637752533, + 0.01070471666753292, + 0.017960986122488976, + 0.04654616490006447, + 0.012086826376616955, + 0.07964236289262772, + 0.010468895547091961, + 0.005547366105020046, + 0.007152102421969175, + 0.004054222721606493, + 0.007759725674986839, + 0.009298780001699924, + 0.007191265933215618, + 0.015329386107623577, + 0.0068289428018033504, + 0.0025177025236189365, + 0.002805336844176054, + 0.0030438650865107775, + 0.003816475858911872, + 0.005103431176394224, + 0.003181165549904108, + 0.006288685370236635, + 0.003247844986617565, + 0.0035657682456076145, + 0.004035491496324539, + 0.0035413606092333794, + 0.005727837327867746, + 0.00617189472541213, + 0.0041923923417925835, + 0.005367015488445759, + 0.01585405133664608, + 0.006880549248307943, + 0.016428327187895775, + 0.007716177962720394, + 0.00893754605203867, + 0.015929479151964188, + 0.009708309546113014, + 0.017206288874149323, + 0.008917826227843761, + 0.008136515505611897, + 0.011365441605448723, + 0.004603220149874687, + 0.0050120665691792965, + 0.010148115456104279, + 0.005396128166466951, + 0.012550411745905876, + 0.006133971270173788, + 0.0036655415315181017, + 0.021072210744023323, + 0.008296859450638294, + 0.002865152433514595, + 0.006275673862546682, + 0.005386367440223694, + 0.022716669365763664, + 0.01995432749390602, + 0.013045890256762505, + 0.06717479974031448, + 0.014585588127374649, + 0.015522315166890621, + 0.02045154757797718, + 0.01245281845331192, + 0.03125300258398056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0036187181249260902 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly immersive, first-person, subjective, and static shot begins with a slender, white hand slowly reaching forward from the bottom of the frame, as if reaching for something invisible. The background is a nearly pitch-black void, devoid of any discernible objects. As the hand reaches out, countless fragments of memory, shimmering with a soft golden glow, begin to converge from all directions into the space in front of the palm, as if drawn by an invisible gravitational force. These luminous particles, like living stardust, swirl and dance in the air before carefully, steadily, and deliberately approaching and merging. Their light provides the sole source of illumination, casting a warm, dynamic shadow on the outstretched hand. Ultimately, these points of light coalesce and intertwine into a hazy, translucent figure composed of pure light, its facial features difficult to discern. The contrast between the profound darkness and the warm golden glow creates a tranquil, dreamy, and deeply nostalgic atmosphere.", + "seed": 31167, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_1/index_9638_seed_31167_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009060758305713535, + "vlm_logprob": -0.009060758305713535, + "vlm_yes_no_margin": 5.124999952316284, + "vlm_token_credit": -0.009060758305713535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026806116104125977 + ], + "opposite_logprob_tokens": [ + -4.026805877685547 + ], + "logprob_reward": -0.026806116104125977, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.026806116104125977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3172937333583832, + 0.07510045170783997, + 0.07399468123912811, + 0.18297584354877472, + 0.11143425852060318, + 0.09725421667098999, + 0.047758765518665314, + 0.036843009293079376, + 0.0294748917222023, + 0.027870144695043564 + ], + "credit_map": [ + 0.04527692869305611, + 0.03684741258621216, + 0.04423288255929947, + 0.028665408492088318, + 0.03669550269842148, + 0.055358078330755234, + 0.0379299521446228, + 0.03228755667805672, + 0.009095585905015469, + 0.0076753366738557816, + 0.009920314885675907, + 0.008973981253802776, + 0.008520136587321758, + 0.011994552798569202, + 0.009720457717776299, + 0.009200084023177624, + 0.008294826373457909, + 0.010110876522958279, + 0.012329565361142159, + 0.0070059699937701225, + 0.008993998169898987, + 0.011090426705777645, + 0.007781927939504385, + 0.008387087844312191, + 0.015444706194102764, + 0.009098264388740063, + 0.03643615543842316, + 0.043025076389312744, + 0.007669038604944944, + 0.04240752011537552, + 0.014269297942519188, + 0.014625771902501583, + 0.015068798325955868, + 0.00862019695341587, + 0.029639875516295433, + 0.014137948863208294, + 0.006641198880970478, + 0.019597800448536873, + 0.009372220374643803, + 0.008356220088899136, + 0.007822082377970219, + 0.010411749593913555, + 0.014649789780378342, + 0.015292015857994556, + 0.01056738942861557, + 0.013944277539849281, + 0.013681558892130852, + 0.010885360650718212, + 0.006936321966350079, + 0.003588324412703514, + 0.0064122905023396015, + 0.006086576730012894, + 0.0033008651807904243, + 0.007672999985516071, + 0.005122681614011526, + 0.008638706058263779, + 0.0034916659351438284, + 0.0032847383990883827, + 0.005039929412305355, + 0.005617917515337467, + 0.00452993530780077, + 0.004224449396133423, + 0.0059877317398786545, + 0.00466663995757699, + 0.002910217037424445, + 0.002803526818752289, + 0.0036093241069465876, + 0.004720998462289572, + 0.0036051839124411345, + 0.003887985134497285, + 0.004603863228112459, + 0.003333794418722391, + 0.0038068450521677732, + 0.003951017279177904, + 0.0028374111279845238, + 0.0041153281927108765, + 0.003791667055338621, + 0.0028927584644407034, + 0.0031779087148606777, + 0.0032972083427011967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.026806116104125977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954977124929428 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954977124929428, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954977124929428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08941719681024551, + 0.6719003915786743, + 0.08381947875022888, + 0.061204783618450165, + 0.027361780405044556, + 0.024577798321843147, + 0.010048015974462032, + 0.017769213765859604, + 0.005765809677541256, + 0.008135554380714893 + ], + "credit_map": [ + 0.008705741725862026, + 0.012642270885407925, + 0.012498365715146065, + 0.012575719505548477, + 0.011917964555323124, + 0.009949298575520515, + 0.013938535936176777, + 0.007189301773905754, + 0.10387766361236572, + 0.06298185139894485, + 0.0878908634185791, + 0.046632055193185806, + 0.06899961084127426, + 0.1347213238477707, + 0.06742661446332932, + 0.09937040507793427, + 0.011887299828231335, + 0.00684343371540308, + 0.0148955462500453, + 0.012082219123840332, + 0.0059280674904584885, + 0.012560049071907997, + 0.009552069939672947, + 0.01007078681141138, + 0.005619559902697802, + 0.0030697749461978674, + 0.013464509509503841, + 0.012379947118461132, + 0.003085200907662511, + 0.011074282228946686, + 0.0053787012584507465, + 0.0071328068152070045, + 0.0032167951576411724, + 0.0023896275088191032, + 0.006367608439177275, + 0.002827852964401245, + 0.0018724938854575157, + 0.004245986230671406, + 0.0035738046281039715, + 0.002867609728127718, + 0.0023003940004855394, + 0.003037614282220602, + 0.0029938810039311647, + 0.004905666224658489, + 0.0028933994472026825, + 0.002811708487570286, + 0.0033448082394897938, + 0.0022903268691152334, + 0.001208904548548162, + 0.0010502277873456478, + 0.0016793619142845273, + 0.0010110815055668354, + 0.000756479159463197, + 0.0016098064370453358, + 0.0009322374826297164, + 0.0017999173142015934, + 0.001726756920106709, + 0.001881926553323865, + 0.002147380728274584, + 0.0022351176012307405, + 0.0023673151154071093, + 0.002193581312894821, + 0.002698327414691448, + 0.002518807537853718, + 0.0005842396640218794, + 0.0004883208312094212, + 0.0006634659366682172, + 0.0008694051066413522, + 0.0006985579966567457, + 0.0006937049329280853, + 0.0010717681143432856, + 0.0006963469786569476, + 0.000859948166180402, + 0.0010533117456361651, + 0.0007710747304372489, + 0.0010745120234787464, + 0.001274482230655849, + 0.0011079725809395313, + 0.0009819698752835393, + 0.0010122829116880894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.005954977124929428 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097519442439079 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097519442439079, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097519442439079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11407680809497833, + 0.46624982357025146, + 0.09732294827699661, + 0.08192411810159683, + 0.08467301726341248, + 0.025244485586881638, + 0.028705250471830368, + 0.02189696580171585, + 0.031841203570365906, + 0.048065368086099625 + ], + "credit_map": [ + 0.011143154464662075, + 0.012949964962899685, + 0.0155250895768404, + 0.00932014174759388, + 0.02083718404173851, + 0.020024742931127548, + 0.01190259214490652, + 0.012373944744467735, + 0.0698089525103569, + 0.042001873254776, + 0.05977568030357361, + 0.03757229819893837, + 0.05039702355861664, + 0.097771555185318, + 0.05041484534740448, + 0.05850759893655777, + 0.008887246251106262, + 0.01758810132741928, + 0.012783125042915344, + 0.010824414901435375, + 0.012524995021522045, + 0.013715079054236412, + 0.010663536377251148, + 0.01033646147698164, + 0.005942812655121088, + 0.005017779767513275, + 0.019854983314871788, + 0.01527101919054985, + 0.0038718602154403925, + 0.01975337602198124, + 0.006234050262719393, + 0.005978238768875599, + 0.007943144999444485, + 0.00604828167706728, + 0.012111959047615528, + 0.015086174011230469, + 0.004289024509489536, + 0.01973959431052208, + 0.008714868687093258, + 0.010739980265498161, + 0.00284990924410522, + 0.002362340921536088, + 0.003452346194535494, + 0.00313805160112679, + 0.0021528664510697126, + 0.0034643507096916437, + 0.003949523437768221, + 0.003875098191201687, + 0.004347508307546377, + 0.002562409732490778, + 0.004198432434350252, + 0.004044219385832548, + 0.0028510531410574913, + 0.004011340904980898, + 0.0029460578225553036, + 0.0037442313041538, + 0.00248355558142066, + 0.0016623076517134905, + 0.0023427894338965416, + 0.0025593158788979053, + 0.002804878866299987, + 0.0029337117448449135, + 0.0028369694482535124, + 0.004273438826203346, + 0.003149311989545822, + 0.003602167358621955, + 0.0035662874579429626, + 0.004782630130648613, + 0.00418776785954833, + 0.004135136026889086, + 0.005128843244165182, + 0.0032890578731894493, + 0.0044143362902104855, + 0.006436429917812347, + 0.004229706712067127, + 0.007722325623035431, + 0.0064635067246854305, + 0.006397532299160957, + 0.0070509337820112705, + 0.005350602325052023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.004097519442439079 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031935188453644514 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031935188453644514, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031935188453644514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09624654054641724, + 0.3348565101623535, + 0.06911308318376541, + 0.10494593530893326, + 0.07917482405900955, + 0.125777006149292, + 0.053412195295095444, + 0.04316268488764763, + 0.03946053981781006, + 0.05385063961148262 + ], + "credit_map": [ + 0.011225897818803787, + 0.010870648548007011, + 0.013654584996402264, + 0.010057653300464153, + 0.012673319317400455, + 0.015294024720788002, + 0.011476099491119385, + 0.010994313284754753, + 0.04911112040281296, + 0.03191467747092247, + 0.04437534511089325, + 0.02394283376634121, + 0.03630286455154419, + 0.06556764990091324, + 0.0346391387283802, + 0.049002863466739655, + 0.006473775487393141, + 0.006348408292979002, + 0.01123746857047081, + 0.010276325047016144, + 0.0058220261707901955, + 0.013704021461308002, + 0.008256347849965096, + 0.006994708441197872, + 0.008223408833146095, + 0.0053884475491940975, + 0.022015832364559174, + 0.022623823955655098, + 0.004267560318112373, + 0.024729080498218536, + 0.0081432880833745, + 0.00955449789762497, + 0.010296674445271492, + 0.006576899439096451, + 0.017109494656324387, + 0.009483085945248604, + 0.004983930848538876, + 0.012139633297920227, + 0.008351840078830719, + 0.01023327000439167, + 0.028663896024227142, + 0.0067995344288647175, + 0.0324116088449955, + 0.008962354622781277, + 0.005937444511801004, + 0.019653912633657455, + 0.008390883915126324, + 0.014957373961806297, + 0.0054151867516338825, + 0.005105950403958559, + 0.007341154385358095, + 0.006230569910258055, + 0.006716903764754534, + 0.007942263036966324, + 0.00749053992331028, + 0.007169629447162151, + 0.00496375747025013, + 0.004248990677297115, + 0.007109689526259899, + 0.006277566775679588, + 0.004459402058273554, + 0.00632072938606143, + 0.004511852283030748, + 0.005270694848150015, + 0.0038924289401620626, + 0.004091551527380943, + 0.004435102455317974, + 0.005846479907631874, + 0.005263876169919968, + 0.00569763733074069, + 0.006016946397721767, + 0.004216517321765423, + 0.005318668205291033, + 0.0066346353851258755, + 0.0046793087385594845, + 0.00819818489253521, + 0.0072797322645783424, + 0.007091266103088856, + 0.0075100320391356945, + 0.007138810586184263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.0031935188453644514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender white hand visible reaching forward from the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden fragments converge from all directions into the space in front of the palm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the golden particles coalesce and intertwine to form a hazy translucent figure with difficult-to-discern facial features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a nearly pitch-black void devoid of discernible objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3453162908554077, + 0.36135223507881165, + 0.03321389853954315, + 0.026980917900800705, + 0.039181698113679886, + 0.037913933396339417, + 0.1107119470834732, + 0.023022551089525223, + 0.01111591700464487, + 0.011190569959580898 + ], + "credit_map": [ + 0.0596882589161396, + 0.03183169662952423, + 0.048213932663202286, + 0.023893270641565323, + 0.043896324932575226, + 0.07068552076816559, + 0.03638913854956627, + 0.030718132853507996, + 0.054611049592494965, + 0.03468439355492592, + 0.046291884034872055, + 0.026011290028691292, + 0.03934045881032944, + 0.07339038699865341, + 0.035994868725538254, + 0.05102792754769325, + 0.003985353279858828, + 0.0033006074372678995, + 0.005634075962007046, + 0.0037553526926785707, + 0.002423312049359083, + 0.00765720009803772, + 0.0024122276809066534, + 0.0040457709692418575, + 0.0026397884357720613, + 0.0015069266082718968, + 0.00804108940064907, + 0.0035423480439931154, + 0.001141951186582446, + 0.006713989190757275, + 0.0015619415789842606, + 0.0018328864825889468, + 0.005340347532182932, + 0.002566051436588168, + 0.004601922817528248, + 0.009052742272615433, + 0.002151194028556347, + 0.00706745870411396, + 0.004023395478725433, + 0.004378588404506445, + 0.004168040584772825, + 0.00429036607965827, + 0.005595745984464884, + 0.006713277660310268, + 0.003821976250037551, + 0.004903723485767841, + 0.003927592653781176, + 0.004493209533393383, + 0.015490042977035046, + 0.009362194687128067, + 0.014850051142275333, + 0.013696753419935703, + 0.01013384759426117, + 0.015942836180329323, + 0.011331500485539436, + 0.01990473084151745, + 0.002868843963369727, + 0.0026385921519249678, + 0.0032197621185332537, + 0.002248047851026058, + 0.002953282091766596, + 0.0032725115306675434, + 0.0023821976501494646, + 0.003439316526055336, + 0.0012779765529558063, + 0.000957799144089222, + 0.0015778096858412027, + 0.0013679418480023742, + 0.0015056129777804017, + 0.0012858276022598147, + 0.0017076086951419711, + 0.0014353413134813309, + 0.0017393594607710838, + 0.0014153269585222006, + 0.0011691233376041055, + 0.0013017638120800257, + 0.0016696894308552146, + 0.0012931324308738112, + 0.0011685809586197138, + 0.0014335938030853868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.005251660011708736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_tensor([453, 453])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_tensor([453, 453])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..810ea50985386d0bea789dcbc97e71a9af9c5df8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_tensor([453, 453])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "highly tense, eye-level close-up slowly zooms in on a knight standing perfectly still, their form a silent monument of endurance against a desolate, out-of-focus battlefield. The entire scene is bathed in a cold, gloomy sky, creating a somber, desaturated color palette of soft grays and earthy browns. The knight's face, stained with soot and grime, is a mask of stoic determination, their jaw clenched tightly. The steel armor they wear is a tapestry of war, its surface riddled with deep scratches, shallow dents, and a thick layer of dried mud, the metal tarnished in the oppressive darkness. The only movement in the frame is a crystal-clear tear that wells up and begins to slowly, deliberately descend from the corner of one eye. It meticulously carves a clean, shimmering path through the thick mud on their cheek, revealing a line of naked, pale skin beneath. The knights remain motionless, statues of suppressed sorrow, their unwavering stance belying the profound sadness conveyed by this solitary, silent act, creating a powerful and weighty atmosphere of quiet destruction.", + "seed": 27432, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_453_seed_27432_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0359858232550323, + "vlm_logprob": -0.0359858232550323, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.0359858232550323, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08759858459234238 + ], + "opposite_logprob_tokens": [ + -2.5875985622406006 + ], + "logprob_reward": -0.08759858459234238, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08759858459234238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11663541942834854, + 0.21665018796920776, + 0.05271616950631142, + 0.07346924394369125, + 0.08540689945220947, + 0.05565159022808075, + 0.07122994214296341, + 0.06747054308652878, + 0.12216299772262573, + 0.138606995344162 + ], + "credit_map": [ + 0.013194326311349869, + 0.015904275700449944, + 0.017415838316082954, + 0.01101066917181015, + 0.01752106286585331, + 0.011655575595796108, + 0.014286844991147518, + 0.01564682647585869, + 0.02173851989209652, + 0.03883961960673332, + 0.027699699625372887, + 0.016970381140708923, + 0.052957840263843536, + 0.02199081890285015, + 0.0227984432131052, + 0.013654880225658417, + 0.00908959936350584, + 0.004461516626179218, + 0.010220609605312347, + 0.006406284403055906, + 0.00550988782197237, + 0.0061104800552129745, + 0.005115550011396408, + 0.00580223836004734, + 0.007878595031797886, + 0.0070116352289915085, + 0.012210996821522713, + 0.01144928764551878, + 0.008155729621648788, + 0.007287431973963976, + 0.008307999931275845, + 0.011167570017278194, + 0.011336094699800014, + 0.0075752465054392815, + 0.011082321405410767, + 0.013671178370714188, + 0.010742400772869587, + 0.009402584284543991, + 0.010374333709478378, + 0.011222735047340393, + 0.00910892616957426, + 0.005621505435556173, + 0.008052069693803787, + 0.008059807121753693, + 0.004812594503164291, + 0.0085188839584589, + 0.006410970818251371, + 0.005066833924502134, + 0.00611046701669693, + 0.007486185058951378, + 0.009706541895866394, + 0.011591721326112747, + 0.008678260259330273, + 0.005075362045317888, + 0.009364868514239788, + 0.013216536492109299, + 0.004614397417753935, + 0.00918422732502222, + 0.00801080372184515, + 0.011160536669194698, + 0.009565705433487892, + 0.004705877508968115, + 0.01068634632974863, + 0.009542648680508137, + 0.009907104074954987, + 0.016275722533464432, + 0.01664581708610058, + 0.01732601225376129, + 0.015845609828829765, + 0.007664652541279793, + 0.0177987702190876, + 0.020699314773082733, + 0.00790183525532484, + 0.017901327461004257, + 0.025722865015268326, + 0.01694171503186226, + 0.015932634472846985, + 0.007515454664826393, + 0.021431880071759224, + 0.02525927498936653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.08759858459234238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266127642244101 + ], + "opposite_logprob_tokens": [ + -5.255266189575195 + ], + "logprob_reward": -0.005266127642244101, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266127642244101, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05453960597515106, + 0.23508301377296448, + 0.06366516649723053, + 0.22411169111728668, + 0.11413569003343582, + 0.03531708940863609, + 0.061875734478235245, + 0.039624765515327454, + 0.14885656535625458, + 0.022790689021348953 + ], + "credit_map": [ + 0.008151870220899582, + 0.005708044860512018, + 0.010609078221023083, + 0.004628791008144617, + 0.006128582637757063, + 0.006667055655270815, + 0.007246832363307476, + 0.0053993454203009605, + 0.025234539061784744, + 0.03904625400900841, + 0.03315421938896179, + 0.018987245857715607, + 0.05396605655550957, + 0.023500796407461166, + 0.025627879425883293, + 0.015566030517220497, + 0.009367412887513638, + 0.0048033869825303555, + 0.012429834343492985, + 0.008171585388481617, + 0.005437765270471573, + 0.0072171129286289215, + 0.007040579803287983, + 0.009197493083775043, + 0.03673684597015381, + 0.0181701872497797, + 0.03753512725234032, + 0.028928501531481743, + 0.01850797049701214, + 0.025267064571380615, + 0.022481366991996765, + 0.03648461773991585, + 0.01853569597005844, + 0.0066147069446742535, + 0.01687067560851574, + 0.01098458468914032, + 0.008752288296818733, + 0.014338690787553787, + 0.010351469740271568, + 0.027687586843967438, + 0.007621368858963251, + 0.0027967500500380993, + 0.0039474572986364365, + 0.004128139466047287, + 0.0031851560343056917, + 0.00769383367151022, + 0.003331078914925456, + 0.0026133025530725718, + 0.0059134578332304955, + 0.007956872694194317, + 0.00776900677010417, + 0.009947407059371471, + 0.007962370291352272, + 0.004364197142422199, + 0.008001546375453472, + 0.009960874915122986, + 0.0028165732510387897, + 0.004567260853946209, + 0.005988837219774723, + 0.005966953933238983, + 0.0053512705489993095, + 0.0030344692058861256, + 0.0065638478845357895, + 0.005335551220923662, + 0.015816088765859604, + 0.01403122954070568, + 0.027846843004226685, + 0.015350543893873692, + 0.013664929196238518, + 0.010547947138547897, + 0.021114347502589226, + 0.03048463724553585, + 0.0013634666102007031, + 0.0033834900241345167, + 0.002821742556989193, + 0.0037924363277852535, + 0.003507734974846244, + 0.0014347382821142673, + 0.003750577801838517, + 0.0027365037240087986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.005266127642244101 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088653467595577 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088653467595577, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088653467595577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08160869032144547, + 0.2199964076280594, + 0.04697984829545021, + 0.09567306935787201, + 0.05302681028842926, + 0.09965566545724869, + 0.14521905779838562, + 0.1711341142654419, + 0.04203720763325691, + 0.044669099152088165 + ], + "credit_map": [ + 0.008646978065371513, + 0.009965799748897552, + 0.010895933024585247, + 0.013593977317214012, + 0.009823983535170555, + 0.0064786202274262905, + 0.011905803345143795, + 0.010297591798007488, + 0.023363146930933, + 0.03621096909046173, + 0.02675599604845047, + 0.01841658726334572, + 0.0545283779501915, + 0.022124310955405235, + 0.023649215698242188, + 0.014947800897061825, + 0.004974010866135359, + 0.0055604795925319195, + 0.0063766492530703545, + 0.007992486469447613, + 0.006219102069735527, + 0.003922153729945421, + 0.006127774249762297, + 0.005807196255773306, + 0.013228306546807289, + 0.00996222160756588, + 0.010797747410833836, + 0.019496239721775055, + 0.008808782324194908, + 0.009982026182115078, + 0.011129041202366352, + 0.012268704362213612, + 0.0050822473131120205, + 0.007340310141444206, + 0.007119573652744293, + 0.007836463861167431, + 0.00820625014603138, + 0.0041941371746361256, + 0.006857443135231733, + 0.006390385329723358, + 0.018580418080091476, + 0.009104130789637566, + 0.010164035484194756, + 0.009912984445691109, + 0.011838681064546108, + 0.019191501662135124, + 0.011223779059946537, + 0.009640134871006012, + 0.013249076902866364, + 0.021765582263469696, + 0.015211689285933971, + 0.028826475143432617, + 0.020989980548620224, + 0.008746787905693054, + 0.02281809411942959, + 0.01361137256026268, + 0.012486339546740055, + 0.02062651701271534, + 0.017956888303160667, + 0.025472598150372505, + 0.03227182850241661, + 0.015273396857082844, + 0.031139181926846504, + 0.01590736396610737, + 0.004112796857953072, + 0.006295021157711744, + 0.00400207145139575, + 0.0073195467703044415, + 0.004924477078020573, + 0.0024492591619491577, + 0.009010435082018375, + 0.003923601470887661, + 0.0027147873770445585, + 0.0056684729643166065, + 0.006814144551753998, + 0.0048650894314050674, + 0.005992329679429531, + 0.0030076990369707346, + 0.007576933596283197, + 0.008029639720916748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.011088653467595577 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956043489277363 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956043489277363, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956043489277363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0801314115524292, + 0.15736405551433563, + 0.06557610630989075, + 0.05414680019021034, + 0.05006986856460571, + 0.1126856878399849, + 0.12273865193128586, + 0.12277277559041977, + 0.09590684622526169, + 0.13860785961151123 + ], + "credit_map": [ + 0.00744272768497467, + 0.009947729296982288, + 0.013793210498988628, + 0.011160860769450665, + 0.009243684820830822, + 0.007409053388983011, + 0.010699566453695297, + 0.010434579104185104, + 0.013350321911275387, + 0.027184508740901947, + 0.016204454004764557, + 0.016954435035586357, + 0.03805084899067879, + 0.015157531946897507, + 0.018974967300891876, + 0.011486981995403767, + 0.008734405040740967, + 0.008739563636481762, + 0.009677580557763577, + 0.011808042414486408, + 0.007750214543193579, + 0.005530954338610172, + 0.006402734667062759, + 0.0069326055236160755, + 0.009519972838461399, + 0.004331770818680525, + 0.006989759858697653, + 0.0067800492979586124, + 0.004521545954048634, + 0.007446947507560253, + 0.004758723545819521, + 0.009798029437661171, + 0.005621791817247868, + 0.006739378906786442, + 0.005447982344776392, + 0.009435923770070076, + 0.0071078077889978886, + 0.00480355229228735, + 0.006577907595783472, + 0.004335526376962662, + 0.027346184477210045, + 0.008886716328561306, + 0.009043005295097828, + 0.010095876641571522, + 0.010809395462274551, + 0.02791997790336609, + 0.010038956999778748, + 0.008545578457415104, + 0.010416685603559017, + 0.01580628752708435, + 0.019558295607566833, + 0.023167753592133522, + 0.013541040942072868, + 0.007340352050960064, + 0.018169496208429337, + 0.01473874133080244, + 0.009432189166545868, + 0.017144128680229187, + 0.0176689475774765, + 0.015280434861779213, + 0.018442701548337936, + 0.01103776041418314, + 0.016603222116827965, + 0.017163392156362534, + 0.007992375642061234, + 0.010807869955897331, + 0.017837703227996826, + 0.013696430251002312, + 0.010523911565542221, + 0.006546539720147848, + 0.013361001387238503, + 0.015141011215746403, + 0.006769902538508177, + 0.021609090268611908, + 0.02109251171350479, + 0.016616512089967728, + 0.021822113543748856, + 0.010111800394952297, + 0.020143816247582436, + 0.020442107692360878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.005956043489277363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001970708370209 + ], + "opposite_logprob_tokens": [ + -2.6950197219848633 + ], + "logprob_reward": -0.07001970708370209, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001970708370209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09589790552854538, + 0.3743252158164978, + 0.028793765231966972, + 0.061104100197553635, + 0.057450272142887115, + 0.15826986730098724, + 0.08743883669376373, + 0.05879075080156326, + 0.036484941840171814, + 0.04144441708922386 + ], + "credit_map": [ + 0.0100404666736722, + 0.01204879954457283, + 0.016090406104922295, + 0.013844429515302181, + 0.010726560838520527, + 0.009161318652331829, + 0.013153704814612865, + 0.010832207277417183, + 0.033394426107406616, + 0.06301114708185196, + 0.04406246915459633, + 0.033322524279356, + 0.09550315886735916, + 0.03517686203122139, + 0.04331371933221817, + 0.02654087543487549, + 0.003343831514939666, + 0.0028922369237989187, + 0.0053152479231357574, + 0.004606902599334717, + 0.0035384143702685833, + 0.0025776326656341553, + 0.0033092848025262356, + 0.003210210707038641, + 0.008222340606153011, + 0.008009052835404873, + 0.0060586766339838505, + 0.011674794368445873, + 0.005943201947957277, + 0.006572227459400892, + 0.006857356522232294, + 0.0077664488926529884, + 0.005377225577831268, + 0.005537312477827072, + 0.009301123209297657, + 0.010538723319768906, + 0.005700220819562674, + 0.00486390246078372, + 0.007485921029001474, + 0.00864583719521761, + 0.027756158262491226, + 0.011515350081026554, + 0.023969201371073723, + 0.016363751143217087, + 0.014495310373604298, + 0.029884200543165207, + 0.014984825626015663, + 0.019301045686006546, + 0.007371732033789158, + 0.011032350361347198, + 0.011212779209017754, + 0.018198510631918907, + 0.007887794636189938, + 0.005881657358258963, + 0.013420650735497475, + 0.012433345429599285, + 0.004900403320789337, + 0.005704932380467653, + 0.01082628220319748, + 0.0071060312911868095, + 0.007544227410107851, + 0.004195626825094223, + 0.009238751605153084, + 0.009274488314986229, + 0.0029948463197797537, + 0.0051981364376842976, + 0.00761484494432807, + 0.0039001842960715294, + 0.0039009323809295893, + 0.002528879791498184, + 0.00551235256716609, + 0.0048347581177949905, + 0.002405597362667322, + 0.004836440552026033, + 0.006815772969275713, + 0.004628353752195835, + 0.004612766206264496, + 0.003109106095507741, + 0.007029242813587189, + 0.008007131516933441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.07001970708370209 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly tense, eye-level close-up slowly zooms in on a knight standing perfectly still, their form a silent monument of endurance against a desolate, out-of-focus battlefield. The entire scene is bathed in a cold, gloomy sky, creating a somber, desaturated color palette of soft grays and earthy browns. The knight's face, stained with soot and grime, is a mask of stoic determination, their jaw clenched tightly. The steel armor they wear is a tapestry of war, its surface riddled with deep scratches, shallow dents, and a thick layer of dried mud, the metal tarnished in the oppressive darkness. The only movement in the frame is a crystal-clear tear that wells up and begins to slowly, deliberately descend from the corner of one eye. It meticulously carves a clean, shimmering path through the thick mud on their cheek, revealing a line of naked, pale skin beneath. The knights remain motionless, statues of suppressed sorrow, their unwavering stance belying the profound sadness conveyed by this solitary, silent act, creating a powerful and weighty atmosphere of quiet destruction.", + "seed": 27433, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_453_seed_27433_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.050489613134413955, + "vlm_logprob": -0.050489613134413955, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.050489613134413955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14464811980724335 + ], + "opposite_logprob_tokens": [ + -2.1446480751037598 + ], + "logprob_reward": -0.14464811980724335, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14464811980724335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1446395069360733, + 0.05606134980916977, + 0.17599698901176453, + 0.09466372430324554, + 0.14158013463020325, + 0.18166998028755188, + 0.03707542642951012, + 0.039035920053720474, + 0.06625495851039886, + 0.06302197277545929 + ], + "credit_map": [ + 0.014714199118316174, + 0.011607212945818901, + 0.02829965390264988, + 0.02319498360157013, + 0.017858445644378662, + 0.017277516424655914, + 0.018860258162021637, + 0.012827238067984581, + 0.00446162698790431, + 0.0036924120504409075, + 0.010001505725085735, + 0.011027046479284763, + 0.006397546268999577, + 0.005912450607866049, + 0.007605229038745165, + 0.0069635347463190556, + 0.01435437798500061, + 0.012843981385231018, + 0.021610908210277557, + 0.02477118745446205, + 0.021602971479296684, + 0.031763724982738495, + 0.022770829498767853, + 0.026279009878635406, + 0.007446357049047947, + 0.008328915573656559, + 0.015467947348952293, + 0.012480130419135094, + 0.011245956644415855, + 0.01419854536652565, + 0.013063696213066578, + 0.012432173825800419, + 0.00981111265718937, + 0.01040800753980875, + 0.023678041994571686, + 0.024726279079914093, + 0.012257557362318039, + 0.014877074398100376, + 0.021867379546165466, + 0.023954682052135468, + 0.01505156047642231, + 0.01383132953196764, + 0.037367865443229675, + 0.03164367750287056, + 0.01760823279619217, + 0.02124094031751156, + 0.02543700486421585, + 0.01948939636349678, + 0.0027276575565338135, + 0.0038070073351264, + 0.005486181005835533, + 0.006247776560485363, + 0.0031909130048006773, + 0.006291186902672052, + 0.005549374036490917, + 0.00377533002756536, + 0.0029641534201800823, + 0.002946221735328436, + 0.008288674987852573, + 0.004214265383780003, + 0.0036667780950665474, + 0.0070610009133815765, + 0.004343530163168907, + 0.0055512976832687855, + 0.004279706627130508, + 0.006311379373073578, + 0.0079036895185709, + 0.015237623825669289, + 0.010382638312876225, + 0.008732474409043789, + 0.008134305477142334, + 0.005273142363876104, + 0.00469157425686717, + 0.005200892221182585, + 0.010239964351058006, + 0.007904937490820885, + 0.008361862041056156, + 0.008740486577153206, + 0.00806514360010624, + 0.00981711596250534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.14464811980724335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008657658472657204 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.008657658472657204, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008657658472657204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19708354771137238, + 0.3099173903465271, + 0.06569630652666092, + 0.07125287503004074, + 0.1951259821653366, + 0.04439295083284378, + 0.016574418172240257, + 0.021294865757226944, + 0.03796493634581566, + 0.04069676250219345 + ], + "credit_map": [ + 0.02063763700425625, + 0.018400343134999275, + 0.038377545773983, + 0.028350260108709335, + 0.02603553980588913, + 0.021619467064738274, + 0.024515334516763687, + 0.019147425889968872, + 0.022373614832758904, + 0.015995433554053307, + 0.06897816807031631, + 0.08510185778141022, + 0.026633432134985924, + 0.024859637022018433, + 0.039104871451854706, + 0.02687039226293564, + 0.004940578248351812, + 0.003610654268413782, + 0.013719713315367699, + 0.013492686673998833, + 0.0055786496959626675, + 0.007823353633284569, + 0.00840696133673191, + 0.008123699575662613, + 0.0054642739705741405, + 0.00633523752912879, + 0.009617363102734089, + 0.007622324861586094, + 0.009989713318645954, + 0.009761537425220013, + 0.009993813931941986, + 0.012468612752854824, + 0.015779873356223106, + 0.015439862385392189, + 0.03575080633163452, + 0.032258566468954086, + 0.018626857548952103, + 0.03542780131101608, + 0.02237844280898571, + 0.019463777542114258, + 0.002807029290124774, + 0.0034489601384848356, + 0.008179822936654091, + 0.0074239191599190235, + 0.00425048777833581, + 0.004724286030977964, + 0.006969186011701822, + 0.0065892585553228855, + 0.0012302477844059467, + 0.0018015308305621147, + 0.00224682898260653, + 0.00308657786808908, + 0.0012973773991689086, + 0.002675658790394664, + 0.0026319592725485563, + 0.001604236545972526, + 0.00230573583394289, + 0.0019212927436456084, + 0.0034771168138831854, + 0.0020240130834281445, + 0.002614554250612855, + 0.003764765104278922, + 0.00277007557451725, + 0.0024173129349946976, + 0.0027497331611812115, + 0.0028265013825148344, + 0.007844011299312115, + 0.004717883188277483, + 0.004088068380951881, + 0.00632601510733366, + 0.004513482563197613, + 0.00489924056455493, + 0.0044467635452747345, + 0.0032101250253617764, + 0.005166055168956518, + 0.005334207322448492, + 0.0029775926377624273, + 0.005211332347244024, + 0.006606440059840679, + 0.007744248956441879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.008657658472657204 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210578985512257 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210578985512257, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210578985512257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06587491929531097, + 0.04564770311117172, + 0.03370238095521927, + 0.03444460779428482, + 0.592700719833374, + 0.04617293179035187, + 0.023632779717445374, + 0.03647575154900551, + 0.04824433848261833, + 0.07310378551483154 + ], + "credit_map": [ + 0.006135453004390001, + 0.006755848415195942, + 0.012761571444571018, + 0.010088264010846615, + 0.006857125088572502, + 0.006705452222377062, + 0.008628268726170063, + 0.007942930795252323, + 0.003914072178304195, + 0.003478138241916895, + 0.010379817336797714, + 0.007458437699824572, + 0.004582083784043789, + 0.00401242496445775, + 0.006405136547982693, + 0.005417595151811838, + 0.0032207283657044172, + 0.003537939628586173, + 0.004115260671824217, + 0.004094142932444811, + 0.0037901094183325768, + 0.005694684572517872, + 0.00440223328769207, + 0.004847280215471983, + 0.002220044145360589, + 0.0022238853853195906, + 0.005884560756385326, + 0.004730226006358862, + 0.004613159690052271, + 0.0053495015017688274, + 0.004775473847985268, + 0.004647758789360523, + 0.048925820738077164, + 0.04589235782623291, + 0.10616689175367355, + 0.07516320049762726, + 0.06453148275613785, + 0.11420678347349167, + 0.07071960717439651, + 0.06709461659193039, + 0.00290865870192647, + 0.003690119134262204, + 0.004336888901889324, + 0.011039802804589272, + 0.00357401673682034, + 0.007768338546156883, + 0.0074363104067742825, + 0.005418800748884678, + 0.0015764681156724691, + 0.0019774797838181257, + 0.004245557356625795, + 0.0024299793876707554, + 0.0026111926417797804, + 0.004810947459191084, + 0.003249519504606724, + 0.002731634071096778, + 0.0033610546961426735, + 0.0026931718457490206, + 0.006696975324302912, + 0.0031281285919249058, + 0.002993991132825613, + 0.0071405102498829365, + 0.0051917326636612415, + 0.0052701854147017, + 0.004592976998537779, + 0.003714132821187377, + 0.00532714556902647, + 0.009580045007169247, + 0.006646845489740372, + 0.005833100993186235, + 0.007599657867103815, + 0.004950434435158968, + 0.007461998611688614, + 0.006059261970221996, + 0.012450635433197021, + 0.00940233189612627, + 0.006213183514773846, + 0.013012690469622612, + 0.009918561205267906, + 0.008585119619965553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.014210578985512257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954384338110685 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954384338110685, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954384338110685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0800510048866272, + 0.09477388113737106, + 0.08052752912044525, + 0.15227355062961578, + 0.1611950397491455, + 0.05876239761710167, + 0.03605424985289574, + 0.13546977937221527, + 0.041096895933151245, + 0.1597956120967865 + ], + "credit_map": [ + 0.008325603790581226, + 0.006894913502037525, + 0.012423897162079811, + 0.011942902579903603, + 0.011290035210549831, + 0.009351321496069431, + 0.009728393517434597, + 0.010093934834003448, + 0.0065786270424723625, + 0.00453380448743701, + 0.021488986909389496, + 0.01926014944911003, + 0.010087926872074604, + 0.009034301154315472, + 0.013659722171723843, + 0.010130367241799831, + 0.006967373192310333, + 0.007342362776398659, + 0.008198329247534275, + 0.009767420589923859, + 0.011061367578804493, + 0.01551641896367073, + 0.010133367031812668, + 0.011540885083377361, + 0.011285124346613884, + 0.011624056845903397, + 0.02227543666958809, + 0.020169228315353394, + 0.020672915503382683, + 0.02051980048418045, + 0.02171233668923378, + 0.024014662951231003, + 0.01141407247632742, + 0.013175936415791512, + 0.034488752484321594, + 0.021497955545783043, + 0.020840948447585106, + 0.021214839071035385, + 0.020221028476953506, + 0.01834149844944477, + 0.005055909976363182, + 0.003893128829076886, + 0.015571033582091331, + 0.00803840160369873, + 0.004961786326020956, + 0.006493294611573219, + 0.008839062415063381, + 0.005909785628318787, + 0.0024853399954736233, + 0.0028171539306640625, + 0.006832608487457037, + 0.00450429692864418, + 0.00431832717731595, + 0.005145476199686527, + 0.005331492517143488, + 0.004619557410478592, + 0.009049279615283012, + 0.009025242179632187, + 0.04534820467233658, + 0.013756493106484413, + 0.009949155151844025, + 0.015972839668393135, + 0.01735173910856247, + 0.015016821213066578, + 0.004755218513309956, + 0.004378484562039375, + 0.0035481255035847425, + 0.0067639402113854885, + 0.004156521055847406, + 0.0054444195702672005, + 0.008012920618057251, + 0.00403726426884532, + 0.01260522659868002, + 0.010695120319724083, + 0.02734583429992199, + 0.028296519070863724, + 0.012208129279315472, + 0.025159303098917007, + 0.022785097360610962, + 0.02070038765668869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.005954384338110685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789773240685463 + ], + "opposite_logprob_tokens": [ + -2.578977346420288 + ], + "logprob_reward": -0.0789773240685463, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789773240685463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10121375322341919, + 0.06307604908943176, + 0.04904752969741821, + 0.17581363022327423, + 0.15647082030773163, + 0.19356198608875275, + 0.0321449413895607, + 0.10087089240550995, + 0.0708840861916542, + 0.0569162592291832 + ], + "credit_map": [ + 0.01111364085227251, + 0.009070999920368195, + 0.014844652265310287, + 0.011752626858651638, + 0.015464372001588345, + 0.015973161906003952, + 0.012745833955705166, + 0.01024846825748682, + 0.004476102534681559, + 0.0036460186820477247, + 0.019292017444968224, + 0.013594801537692547, + 0.005096780601888895, + 0.003474433673545718, + 0.00881084892898798, + 0.0046850512735545635, + 0.004126209300011396, + 0.004521315451711416, + 0.009877431206405163, + 0.007464023772627115, + 0.004672111943364143, + 0.007565345615148544, + 0.005758328828960657, + 0.005062761716544628, + 0.013905356638133526, + 0.017096318304538727, + 0.026733985170722008, + 0.01876022480428219, + 0.025695163756608963, + 0.023804480209946632, + 0.023888124153017998, + 0.025929968804121017, + 0.012250647880136967, + 0.009075755253434181, + 0.03876595199108124, + 0.018612634390592575, + 0.017270443961024284, + 0.02462807297706604, + 0.0199661236256361, + 0.015901200473308563, + 0.011445763520896435, + 0.015654880553483963, + 0.03445134684443474, + 0.048935920000076294, + 0.017183180898427963, + 0.013186012394726276, + 0.02892206981778145, + 0.02378280647099018, + 0.0026571061462163925, + 0.003201764542609453, + 0.008468126878142357, + 0.004856463056057692, + 0.002869043033570051, + 0.0029205901082605124, + 0.004540386609733105, + 0.002631460316479206, + 0.010429715737700462, + 0.007258860394358635, + 0.012160137295722961, + 0.007817884907126427, + 0.010787122882902622, + 0.022898199036717415, + 0.015054022893309593, + 0.014464949257671833, + 0.00413468899205327, + 0.006864550989121199, + 0.014099285937845707, + 0.011861458420753479, + 0.006373199168592691, + 0.008603074587881565, + 0.009472104720771313, + 0.009475725702941418, + 0.0069845025427639484, + 0.004379344638437033, + 0.00814430508762598, + 0.006511231884360313, + 0.005029238760471344, + 0.00803596805781126, + 0.008743894286453724, + 0.009087774902582169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.0789773240685463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly tense, eye-level close-up slowly zooms in on a knight standing perfectly still, their form a silent monument of endurance against a desolate, out-of-focus battlefield. The entire scene is bathed in a cold, gloomy sky, creating a somber, desaturated color palette of soft grays and earthy browns. The knight's face, stained with soot and grime, is a mask of stoic determination, their jaw clenched tightly. The steel armor they wear is a tapestry of war, its surface riddled with deep scratches, shallow dents, and a thick layer of dried mud, the metal tarnished in the oppressive darkness. The only movement in the frame is a crystal-clear tear that wells up and begins to slowly, deliberately descend from the corner of one eye. It meticulously carves a clean, shimmering path through the thick mud on their cheek, revealing a line of naked, pale skin beneath. The knights remain motionless, statues of suppressed sorrow, their unwavering stance belying the profound sadness conveyed by this solitary, silent act, creating a powerful and weighty atmosphere of quiet destruction.", + "seed": 27434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_453_seed_27434_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.035311943385750055, + "vlm_logprob": -0.035311943385750055, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.035311943385750055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05972597002983093 + ], + "opposite_logprob_tokens": [ + -3.0597259998321533 + ], + "logprob_reward": -0.05972597002983093, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05972597002983093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05846738815307617, + 0.04757453873753548, + 0.024508122354745865, + 0.5054711699485779, + 0.09606333076953888, + 0.10057748854160309, + 0.04722236096858978, + 0.07449366897344589, + 0.021148979663848877, + 0.024472899734973907 + ], + "credit_map": [ + 0.006960527040064335, + 0.008293671533465385, + 0.0089699886739254, + 0.007792472839355469, + 0.006734197027981281, + 0.007402669172734022, + 0.007500345818698406, + 0.004813517909497023, + 0.005613964982330799, + 0.0055634803138673306, + 0.008818395435810089, + 0.004956303164362907, + 0.005440898705273867, + 0.006044214591383934, + 0.007322132587432861, + 0.003815148724243045, + 0.002417151350528002, + 0.003273637732490897, + 0.002919690450653434, + 0.003584189573302865, + 0.0031206952407956123, + 0.0025548215489834547, + 0.0037578404881060123, + 0.002880097134038806, + 0.09319370985031128, + 0.06119820848107338, + 0.04767763242125511, + 0.06094053387641907, + 0.07521521300077438, + 0.0565708763897419, + 0.06669740378856659, + 0.043977607041597366, + 0.012500944547355175, + 0.009143278934061527, + 0.011097114533185959, + 0.010382445529103279, + 0.014362170360982418, + 0.011603240855038166, + 0.010772984474897385, + 0.016201145946979523, + 0.0077236248180270195, + 0.008085664361715317, + 0.009196399711072445, + 0.029717378318309784, + 0.00880933366715908, + 0.007171133998781443, + 0.02399420365691185, + 0.005879752337932587, + 0.0038357977755367756, + 0.007026827894151211, + 0.005281457677483559, + 0.007131360005587339, + 0.007000621408224106, + 0.005173716228455305, + 0.007476526312530041, + 0.004296055529266596, + 0.007506451103836298, + 0.009382501244544983, + 0.008232567459344864, + 0.013714215718209743, + 0.008919411338865757, + 0.0055390894412994385, + 0.013694624416530132, + 0.007504801265895367, + 0.002121166093274951, + 0.001972381491214037, + 0.004134682472795248, + 0.002804545685648918, + 0.0022642966359853745, + 0.0026160578709095716, + 0.0030212081037461758, + 0.0022146396804600954, + 0.001378787448629737, + 0.0037139670457690954, + 0.0032886946573853493, + 0.0037192634772509336, + 0.004520221147686243, + 0.0019126789411529899, + 0.003952838480472565, + 0.001986448187381029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.05972597002983093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764515670016408 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.00764515670016408, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764515670016408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05585206672549248, + 0.022949349135160446, + 0.03438972681760788, + 0.3886939287185669, + 0.24417275190353394, + 0.11157047003507614, + 0.03844611346721649, + 0.05115179717540741, + 0.021503236144781113, + 0.03127063810825348 + ], + "credit_map": [ + 0.007775694597512484, + 0.006630222778767347, + 0.009857689961791039, + 0.007454265374690294, + 0.006351765710860491, + 0.007878079079091549, + 0.005361037794500589, + 0.004543308634310961, + 0.004289837554097176, + 0.0025266148149967194, + 0.0027318070642650127, + 0.00251924991607666, + 0.002271958626806736, + 0.0038181820418685675, + 0.0028021931648254395, + 0.0019895019941031933, + 0.0026505806017667055, + 0.0038141084369271994, + 0.004871943965554237, + 0.005181740503758192, + 0.006003418006002903, + 0.003144466318190098, + 0.005883767269551754, + 0.002839698689058423, + 0.08357445150613785, + 0.042807456105947495, + 0.04245741665363312, + 0.044378530234098434, + 0.04687478020787239, + 0.049868401139974594, + 0.04548574239015579, + 0.03324707970023155, + 0.03022010624408722, + 0.028793761506676674, + 0.03380782529711723, + 0.02585846371948719, + 0.03224010765552521, + 0.030908705666661263, + 0.021012935787439346, + 0.04133082553744316, + 0.009880555793642998, + 0.007403366267681122, + 0.01744447462260723, + 0.024736588820815086, + 0.008981667459011078, + 0.009653748013079166, + 0.022391371428966522, + 0.011078695766627789, + 0.004909181967377663, + 0.004033482167869806, + 0.005244625266641378, + 0.005420343019068241, + 0.003667824435979128, + 0.003755368525162339, + 0.005546040832996368, + 0.005869243294000626, + 0.004897296894341707, + 0.005761739797890186, + 0.0032766549848020077, + 0.010163522325456142, + 0.006044812500476837, + 0.0023899252992123365, + 0.01449985709041357, + 0.004117987118661404, + 0.0017712891567498446, + 0.002517369808629155, + 0.003436075523495674, + 0.0036455485969781876, + 0.0027749398723244667, + 0.0011050289031118155, + 0.003925872966647148, + 0.00232710805721581, + 0.0026851510629057884, + 0.003570407396182418, + 0.005185323301702738, + 0.004482378251850605, + 0.0035694444086402655, + 0.0024264247622340918, + 0.004650583490729332, + 0.004700921941548586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.00764515670016408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420364435762167 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.01420364435762167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420364435762167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26415663957595825, + 0.07689735293388367, + 0.062461357563734055, + 0.0848122239112854, + 0.10409647971391678, + 0.1466742902994156, + 0.06256696581840515, + 0.06593970209360123, + 0.01580808311700821, + 0.11658690124750137 + ], + "credit_map": [ + 0.0340999960899353, + 0.03585074469447136, + 0.037052642554044724, + 0.025943242013454437, + 0.04945319518446922, + 0.03427351266145706, + 0.03222648426890373, + 0.015256818383932114, + 0.00787437241524458, + 0.008283634670078754, + 0.008309298194944859, + 0.01175720151513815, + 0.009805547073483467, + 0.008320036344230175, + 0.016560619696974754, + 0.005986644420772791, + 0.004796069115400314, + 0.007491166237741709, + 0.0068168118596076965, + 0.008716881275177002, + 0.011143690906465054, + 0.005970858968794346, + 0.011153745464980602, + 0.006372131407260895, + 0.014420791529119015, + 0.009205679409205914, + 0.009256252087652683, + 0.011911744251847267, + 0.012513654306530952, + 0.008897710591554642, + 0.011866915971040726, + 0.006739466916769743, + 0.01594025082886219, + 0.011419008485972881, + 0.0134242232888937, + 0.010439368896186352, + 0.014122933149337769, + 0.012572009116411209, + 0.008879072964191437, + 0.01729961857199669, + 0.009019586257636547, + 0.013485664501786232, + 0.01682649925351143, + 0.04118027910590172, + 0.014206444844603539, + 0.008060134947299957, + 0.032931629568338394, + 0.01096404530107975, + 0.004903376568108797, + 0.0056696427054703236, + 0.005168916191905737, + 0.015482092276215553, + 0.006718113087117672, + 0.004478157963603735, + 0.013439670205116272, + 0.006706997752189636, + 0.00528247794136405, + 0.008914128877222538, + 0.0036494287196546793, + 0.010274016298353672, + 0.013393535278737545, + 0.004863368347287178, + 0.014831926673650742, + 0.004730822518467903, + 0.0017370794666931033, + 0.0016557031776756048, + 0.0032663953024894, + 0.0025031138211488724, + 0.001661917893216014, + 0.0009564080974087119, + 0.002389830071479082, + 0.0016376348212361336, + 0.00725203100591898, + 0.017633987590670586, + 0.012110069394111633, + 0.018880566582083702, + 0.0187323410063982, + 0.010440926998853683, + 0.02105727046728134, + 0.01047971285879612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.01420364435762167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595355499535799 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595355499535799, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595355499535799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2078026533126831, + 0.04194675385951996, + 0.038764018565416336, + 0.16837899386882782, + 0.05911673232913017, + 0.14107124507427216, + 0.05339057743549347, + 0.06769949197769165, + 0.04294919595122337, + 0.17888028919696808 + ], + "credit_map": [ + 0.02148270793259144, + 0.03791988268494606, + 0.030949953943490982, + 0.027366025373339653, + 0.03179551661014557, + 0.020152591168880463, + 0.022641068324446678, + 0.015494921244680882, + 0.003760318271815777, + 0.005756845232099295, + 0.004939897917211056, + 0.006372726988047361, + 0.006016967352479696, + 0.004464616067707539, + 0.006785261910408735, + 0.0038501203525811434, + 0.004391059745103121, + 0.00502257514744997, + 0.005216105841100216, + 0.0064332811161875725, + 0.0047026448883116245, + 0.0037148904521018267, + 0.005512186326086521, + 0.0037712783087044954, + 0.0303582064807415, + 0.01907629519701004, + 0.015471446327865124, + 0.022097988054156303, + 0.026086801663041115, + 0.01897738128900528, + 0.023214295506477356, + 0.013096590526401997, + 0.008325745351612568, + 0.0053951027803123, + 0.009657999500632286, + 0.006044362206012011, + 0.007958956062793732, + 0.006546916905790567, + 0.006257949396967888, + 0.008929702453315258, + 0.011079974472522736, + 0.01083484385162592, + 0.013240216299891472, + 0.03990670666098595, + 0.011067240498960018, + 0.012052977457642555, + 0.03363154083490372, + 0.009257761761546135, + 0.003992436453700066, + 0.007508766371756792, + 0.0064039453864097595, + 0.006931055802851915, + 0.008349362760782242, + 0.005121305119246244, + 0.007783945649862289, + 0.007299765013158321, + 0.005646405275911093, + 0.0076113804243505, + 0.006923937704414129, + 0.012319072149693966, + 0.007799754850566387, + 0.0038457224145531654, + 0.015412609092891216, + 0.00814061239361763, + 0.0036910646595060825, + 0.005765515845268965, + 0.004105160944163799, + 0.007316654082387686, + 0.005663739051669836, + 0.0053892480209469795, + 0.006860868539661169, + 0.004156949929893017, + 0.012703990563750267, + 0.02135668322443962, + 0.017627988010644913, + 0.03610903397202492, + 0.023110102862119675, + 0.01631644181907177, + 0.033715665340423584, + 0.01794039085507393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.00595355499535799 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890313908457756 + ], + "opposite_logprob_tokens": [ + -2.464031457901001 + ], + "logprob_reward": -0.0890313908457756, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890313908457756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2267671674489975, + 0.05318991839885712, + 0.08670878410339355, + 0.15601257979869843, + 0.04120882228016853, + 0.22952842712402344, + 0.08087746053934097, + 0.058872103691101074, + 0.019683541730046272, + 0.047151222825050354 + ], + "credit_map": [ + 0.030203012749552727, + 0.028748802840709686, + 0.02909829281270504, + 0.024024508893489838, + 0.04463096708059311, + 0.03049994446337223, + 0.02589261345565319, + 0.013669030740857124, + 0.004470794461667538, + 0.006515853106975555, + 0.00720619922503829, + 0.007345984224230051, + 0.007792328018695116, + 0.005902475211769342, + 0.009701254777610302, + 0.004255026113241911, + 0.005702430848032236, + 0.009553233161568642, + 0.007452091667801142, + 0.013914408162236214, + 0.015893369913101196, + 0.007868519052863121, + 0.01671557128429413, + 0.0096091628074646, + 0.02732604369521141, + 0.017614584416151047, + 0.01571144163608551, + 0.019736943766474724, + 0.02475808747112751, + 0.017184436321258545, + 0.021307216957211494, + 0.01237382274121046, + 0.006101072300225496, + 0.004474801942706108, + 0.006881376262754202, + 0.005899661686271429, + 0.005037540104240179, + 0.003761674277484417, + 0.005214042961597443, + 0.003838651580736041, + 0.01759486459195614, + 0.018564006313681602, + 0.020081467926502228, + 0.06734471023082733, + 0.019742118194699287, + 0.016453757882118225, + 0.05539806932210922, + 0.014349451288580894, + 0.007014954462647438, + 0.007859441451728344, + 0.011751738376915455, + 0.015981420874595642, + 0.009361266158521175, + 0.0075518544763326645, + 0.014824445359408855, + 0.006532339379191399, + 0.0035680015571415424, + 0.006001283880323172, + 0.0037195258773863316, + 0.012530146166682243, + 0.012745518237352371, + 0.003431451739743352, + 0.013434738852083683, + 0.0034414371475577354, + 0.0022546874824911356, + 0.002231796272099018, + 0.003718020161613822, + 0.0030004114378243685, + 0.002035980112850666, + 0.0015850666677579284, + 0.002973731840029359, + 0.001883849035948515, + 0.003080757800489664, + 0.008035596460103989, + 0.0035392525605857372, + 0.007817926816642284, + 0.009540645405650139, + 0.0024590319953858852, + 0.009609471075236797, + 0.0030685400124639273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.0890313908457756 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly tense, eye-level close-up slowly zooms in on a knight standing perfectly still, their form a silent monument of endurance against a desolate, out-of-focus battlefield. The entire scene is bathed in a cold, gloomy sky, creating a somber, desaturated color palette of soft grays and earthy browns. The knight's face, stained with soot and grime, is a mask of stoic determination, their jaw clenched tightly. The steel armor they wear is a tapestry of war, its surface riddled with deep scratches, shallow dents, and a thick layer of dried mud, the metal tarnished in the oppressive darkness. The only movement in the frame is a crystal-clear tear that wells up and begins to slowly, deliberately descend from the corner of one eye. It meticulously carves a clean, shimmering path through the thick mud on their cheek, revealing a line of naked, pale skin beneath. The knights remain motionless, statues of suppressed sorrow, their unwavering stance belying the profound sadness conveyed by this solitary, silent act, creating a powerful and weighty atmosphere of quiet destruction.", + "seed": 27435, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_10/index_453_seed_27435_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04629813944920898, + "vlm_logprob": -0.04629813944920898, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.04629813944920898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12522225081920624 + ], + "opposite_logprob_tokens": [ + -2.2502222061157227 + ], + "logprob_reward": -0.12522225081920624, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12522225081920624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10279680788516998, + 0.03486963361501694, + 0.06392133980989456, + 0.18265502154827118, + 0.1501701921224594, + 0.06353899836540222, + 0.11910228431224823, + 0.07978912442922592, + 0.07791025936603546, + 0.12524627149105072 + ], + "credit_map": [ + 0.01272259745746851, + 0.01354536134749651, + 0.013605853542685509, + 0.01570904813706875, + 0.009998421184718609, + 0.010636442340910435, + 0.013756259344518185, + 0.012822823598980904, + 0.005508025176823139, + 0.004704582970589399, + 0.0044321627356112, + 0.004322676919400692, + 0.00359664810821414, + 0.004801148548722267, + 0.003993751481175423, + 0.003510640235617757, + 0.00823157001286745, + 0.012034080922603607, + 0.008385883644223213, + 0.008951169438660145, + 0.008017481304705143, + 0.0062636882066726685, + 0.007426051422953606, + 0.0046114143915474415, + 0.017845137044787407, + 0.01328892633318901, + 0.03362812474370003, + 0.015586454421281815, + 0.023932838812470436, + 0.01886695809662342, + 0.029918182641267776, + 0.029588384553790092, + 0.013157372362911701, + 0.018903136253356934, + 0.021578878164291382, + 0.021391479298472404, + 0.020046669989824295, + 0.009805372916162014, + 0.025597134605050087, + 0.019690154120326042, + 0.006036966107785702, + 0.004712048452347517, + 0.011009288020431995, + 0.007550222799181938, + 0.00822369009256363, + 0.006688128691166639, + 0.008550918661057949, + 0.01076773926615715, + 0.009241807274520397, + 0.00669432058930397, + 0.013382628560066223, + 0.017340967431664467, + 0.007621141150593758, + 0.012879176996648312, + 0.022292418405413628, + 0.029649829491972923, + 0.00524878641590476, + 0.0037960968911647797, + 0.012575744651257992, + 0.009969671256840229, + 0.00624122517183423, + 0.0115825105458498, + 0.014751815237104893, + 0.015623272396624088, + 0.006039337255060673, + 0.0036877484526485205, + 0.008025139570236206, + 0.008715269155800343, + 0.008009709417819977, + 0.009558431804180145, + 0.018816975876688957, + 0.015057653188705444, + 0.006886408664286137, + 0.007706797681748867, + 0.030355146154761314, + 0.021261194720864296, + 0.009585535153746605, + 0.012902898713946342, + 0.02032649517059326, + 0.016221793368458748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.12522225081920624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005267194472253323 + ], + "opposite_logprob_tokens": [ + -5.255267143249512 + ], + "logprob_reward": -0.005267194472253323, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005267194472253323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26333409547805786, + 0.15707752108573914, + 0.06647185236215591, + 0.10372768342494965, + 0.05839470773935318, + 0.04436227306723595, + 0.05276952311396599, + 0.05345815792679787, + 0.06632758677005768, + 0.13407659530639648 + ], + "credit_map": [ + 0.030792685225605965, + 0.03433617949485779, + 0.03469198942184448, + 0.047976303845644, + 0.027868224307894707, + 0.026022199541330338, + 0.02802467532455921, + 0.033621836453676224, + 0.025098394602537155, + 0.016102705150842667, + 0.02476964145898819, + 0.021495284512639046, + 0.017586786299943924, + 0.021092550829052925, + 0.01380762830376625, + 0.01712452434003353, + 0.009060314856469631, + 0.006018030922859907, + 0.015538685023784637, + 0.006215107627213001, + 0.007555743679404259, + 0.008348504081368446, + 0.007157239597290754, + 0.006578229833394289, + 0.010780050419270992, + 0.006450574845075607, + 0.017737502232193947, + 0.008566729724407196, + 0.018985429778695107, + 0.010278234258294106, + 0.01939290203154087, + 0.011536266654729843, + 0.003671506652608514, + 0.004963453393429518, + 0.008175031282007694, + 0.005238433368504047, + 0.00888927560299635, + 0.003355683758854866, + 0.00894576869904995, + 0.015155556611716747, + 0.002490086480975151, + 0.0035374800208956003, + 0.0070849996991455555, + 0.008845140226185322, + 0.005272517446428537, + 0.0029606574680656195, + 0.0074849920347332954, + 0.006686401087790728, + 0.003841645084321499, + 0.002292549703270197, + 0.007279683835804462, + 0.004873461555689573, + 0.0036863733548671007, + 0.0065705180168151855, + 0.010370609350502491, + 0.013854681514203548, + 0.003407310461625457, + 0.0023161808494478464, + 0.007832302711904049, + 0.004369087051600218, + 0.005265879910439253, + 0.009212598204612732, + 0.011613230220973492, + 0.00944156851619482, + 0.008820326998829842, + 0.003195506287738681, + 0.007295307703316212, + 0.003975464031100273, + 0.004900586325675249, + 0.007883729413151741, + 0.010519377887248993, + 0.019737284630537033, + 0.009080908261239529, + 0.00997685082256794, + 0.025848837569355965, + 0.027301901951432228, + 0.009414355270564556, + 0.01465253159403801, + 0.018037056550383568, + 0.01976415142416954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.005267194472253323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079863533377647 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.016079863533377647, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079863533377647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09932712465524673, + 0.041249074041843414, + 0.050660666078329086, + 0.12390808761119843, + 0.10785035043954849, + 0.049200039356946945, + 0.2315121591091156, + 0.1390189379453659, + 0.08515339344739914, + 0.07212022691965103 + ], + "credit_map": [ + 0.009833871386945248, + 0.01313815824687481, + 0.013654626905918121, + 0.023279178887605667, + 0.009775950573384762, + 0.01093180850148201, + 0.00796276330947876, + 0.010750766843557358, + 0.006628436967730522, + 0.004098483826965094, + 0.006031691562384367, + 0.00435892166569829, + 0.004308484960347414, + 0.005715999752283096, + 0.0047329203225672245, + 0.005374134052544832, + 0.005177025683224201, + 0.004949087742716074, + 0.011544957756996155, + 0.006653126794844866, + 0.005989047233015299, + 0.0053778765723109245, + 0.006290240678936243, + 0.004679302219301462, + 0.01065173838287592, + 0.014092792756855488, + 0.01896250620484352, + 0.018118003383278847, + 0.019342457875609398, + 0.011821015737950802, + 0.017133967950940132, + 0.013785595074295998, + 0.010162396356463432, + 0.00989000964909792, + 0.014896586537361145, + 0.012647331692278385, + 0.013833564706146717, + 0.00898380670696497, + 0.018463745713233948, + 0.01897289790213108, + 0.0037566216196864843, + 0.0039040911942720413, + 0.008169474080204964, + 0.007361916359513998, + 0.0068354359827935696, + 0.003944519441574812, + 0.00709952786564827, + 0.008128450252115726, + 0.013504142872989178, + 0.014045752584934235, + 0.03512074425816536, + 0.05403987318277359, + 0.01933646760880947, + 0.02761821262538433, + 0.03930593281984329, + 0.028541015461087227, + 0.009036310017108917, + 0.0058808112516999245, + 0.011130656115710735, + 0.012606930918991566, + 0.013091038912534714, + 0.033453453332185745, + 0.02781112864613533, + 0.026008600369095802, + 0.005144189111888409, + 0.006146848201751709, + 0.012479543685913086, + 0.013677296228706837, + 0.007901914417743683, + 0.008265433833003044, + 0.018961109220981598, + 0.012577050365507603, + 0.0035145834553986788, + 0.003334581619128585, + 0.015991179272532463, + 0.007544334977865219, + 0.00874538067728281, + 0.01050251629203558, + 0.013474296778440475, + 0.009013350121676922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.016079863533377647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953317973762751 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953317973762751, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953317973762751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14709046483039856, + 0.052116699516773224, + 0.08884432166814804, + 0.0695488229393959, + 0.19404485821723938, + 0.04365827143192291, + 0.1656884402036667, + 0.08743146061897278, + 0.06306208670139313, + 0.0885145515203476 + ], + "credit_map": [ + 0.013164135627448559, + 0.02125173807144165, + 0.020945057272911072, + 0.03252478316426277, + 0.018641090020537376, + 0.011729342862963676, + 0.014245497062802315, + 0.014588813297450542, + 0.0064643872901797295, + 0.006732078734785318, + 0.006412615068256855, + 0.00848527904599905, + 0.006205425597727299, + 0.0046899570152163506, + 0.0063435691408813, + 0.006783391814678907, + 0.007212277967482805, + 0.013002140447497368, + 0.012108277529478073, + 0.011125790886580944, + 0.012522445991635323, + 0.007160387933254242, + 0.010991573333740234, + 0.014721439220011234, + 0.005796192679554224, + 0.0048839738592505455, + 0.012115420773625374, + 0.00788505095988512, + 0.010175956413149834, + 0.005790592636913061, + 0.012690232135355473, + 0.010211406275629997, + 0.014198522083461285, + 0.024213988333940506, + 0.027966635301709175, + 0.03063429705798626, + 0.029521707445383072, + 0.013408256694674492, + 0.030411547049880028, + 0.023689929395914078, + 0.003663708921521902, + 0.003415178507566452, + 0.006374274380505085, + 0.00781520176678896, + 0.00490301288664341, + 0.004104301333427429, + 0.006924204528331757, + 0.006458390969783068, + 0.009781121276319027, + 0.010114258155226707, + 0.015988707542419434, + 0.038747530430555344, + 0.014385106973350048, + 0.020916542038321495, + 0.02831297740340233, + 0.0274422075599432, + 0.005490270908921957, + 0.003661323571577668, + 0.0105297164991498, + 0.00868863146752119, + 0.007928866893053055, + 0.018035607412457466, + 0.017052143812179565, + 0.01604490727186203, + 0.0059259263798594475, + 0.004141615238040686, + 0.008792867884039879, + 0.006296481005847454, + 0.005360845942050219, + 0.007883088663220406, + 0.01127655804157257, + 0.013384709134697914, + 0.006188215222209692, + 0.00539430882781744, + 0.01654140092432499, + 0.012749241665005684, + 0.007660755421966314, + 0.008867288008332253, + 0.01568061113357544, + 0.015432735905051231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.005953317973762751 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a knight standing perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a desolate, out-of-focus battlefield under a cold, gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's face stained with soot and grime, with a jaw clenched tightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the steel armor covered in deep scratches, dents, dried mud, and tarnished by darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a single tear slowly descend from the corner of the eye, carving a path through the mud to reveal pale skin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896807044744492 + ], + "opposite_logprob_tokens": [ + -2.578968048095703 + ], + "logprob_reward": -0.07896807044744492, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896807044744492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15152759850025177, + 0.039916206151247025, + 0.07663179934024811, + 0.1471899002790451, + 0.0710429698228836, + 0.048408590257167816, + 0.09087100625038147, + 0.10402940958738327, + 0.04921363294124603, + 0.22116893529891968 + ], + "credit_map": [ + 0.01614927127957344, + 0.022864514961838722, + 0.02223089523613453, + 0.028946105390787125, + 0.018434913828969002, + 0.014720581471920013, + 0.013427580706775188, + 0.014753730036318302, + 0.00479933014139533, + 0.005303684156388044, + 0.00542243430390954, + 0.004979399032890797, + 0.005883027799427509, + 0.0038809406105428934, + 0.004566941410303116, + 0.005080447532236576, + 0.008446703664958477, + 0.006769038271158934, + 0.019867703318595886, + 0.008133879862725735, + 0.008342078886926174, + 0.007232048083096743, + 0.00979447178542614, + 0.00804587546736002, + 0.01226448081433773, + 0.016639210283756256, + 0.020883316174149513, + 0.019148921594023705, + 0.02965814433991909, + 0.015123989433050156, + 0.0208580419421196, + 0.012613795697689056, + 0.006512523628771305, + 0.006364706438034773, + 0.009519342333078384, + 0.008816273882985115, + 0.010340031236410141, + 0.004791416227817535, + 0.014546460472047329, + 0.010152214206755161, + 0.004215677734464407, + 0.004441749770194292, + 0.006102987099438906, + 0.007491508033126593, + 0.006209030747413635, + 0.004289086442440748, + 0.0077405390329658985, + 0.007918011397123337, + 0.004134512506425381, + 0.00542617030441761, + 0.00913307536393404, + 0.024690276011824608, + 0.007262599654495716, + 0.010809219442307949, + 0.017613735049962997, + 0.011801417917013168, + 0.005343239288777113, + 0.005501735024154186, + 0.01828748732805252, + 0.012508820742368698, + 0.010500944219529629, + 0.01471115741878748, + 0.021020187065005302, + 0.01615583337843418, + 0.0029723176266998053, + 0.0036316215991973877, + 0.006229641381651163, + 0.008732752874493599, + 0.004619869869202375, + 0.004988756030797958, + 0.010337680578231812, + 0.007700994610786438, + 0.012418457306921482, + 0.01629095897078514, + 0.04563303291797638, + 0.044497791677713394, + 0.01420655008405447, + 0.02327674627304077, + 0.033364396542310715, + 0.03148098662495613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.07896807044744492 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_11/index_tensor([6815, 6815])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_11/index_tensor([6815, 6815])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..650afc4ca5c4908cfa5fc59dec7084b5e8e08626 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_11/index_tensor([6815, 6815])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dizzying high-angle tracking shot swoops vertically downward, rapidly following a skeletal biomechanical creature in a terrifyingly steep dive. The creature's form is a Giger-esque fusion of bone and machine; a spindly, ribbed spine visibly merges with hydraulic pistons and dark, pulsating pipes, while its smooth, obsidian skull is eyeless, save for a faint, inner amber glow emanating from its empty eye sockets. Its massive, leathery wings, textured like scarred hide, are supported by metal struts and partially fold back to increase speed, only occasionally adjusting its trajectory with powerful, deliberate flaps. The camera follows its terrifying dive across a narrow, oppressive chasm separating vast floating islands. These islands are no mere rocks, but nightmarish, living landscapes of ribbed, chitinous structures fused with dark, damp stone, their bases dripping with viscous fluids and tangled, tubular organic growths. The walls of the abyss are adorned with twisting metallic vines of alien flora and pulsating bioluminescent sacs, casting a sickly, cool green light. A thick, swirling mist rises from the abyss below, obscuring the final destination. Rendered in a desaturated, monochromatic palette of gunmetal grays and wet blacks, with specular highlights gleaming off the metallic parts of the creatures, they create an atmosphere of profound, cosmic dread and vertigo as both creature and camera plunge inexorably into the unknown.", + "seed": 10105, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_11/index_6815_seed_10105_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.054800821281969544, + "vlm_logprob": -0.054800821281969544, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.054800821281969544, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10876139253377914 + ], + "opposite_logprob_tokens": [ + -2.3587613105773926 + ], + "logprob_reward": -0.10876139253377914, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10876139253377914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3011717200279236, + 0.047485508024692535, + 0.07418259233236313, + 0.13276183605194092, + 0.0291155818849802, + 0.0693807452917099, + 0.032263655215501785, + 0.14771243929862976, + 0.11512518674135208, + 0.050800763070583344 + ], + "credit_map": [ + 0.01626518927514553, + 0.014516708441078663, + 0.032479580491781235, + 0.0283974502235651, + 0.04834062606096268, + 0.057461582124233246, + 0.06153343990445137, + 0.04217712581157684, + 0.0029318074230104685, + 0.0033764485269784927, + 0.004701008088886738, + 0.00919658225029707, + 0.008177796378731728, + 0.0037670363672077656, + 0.010812714695930481, + 0.004522108938544989, + 0.004779383074492216, + 0.005555806215852499, + 0.011071030050516129, + 0.013762342743575573, + 0.009400058537721634, + 0.014068159274756908, + 0.008210293017327785, + 0.007335512433201075, + 0.010477063246071339, + 0.011189600452780724, + 0.028162842616438866, + 0.023745791986584663, + 0.012382571585476398, + 0.012890566140413284, + 0.021166125312447548, + 0.01274727750569582, + 0.0028898734599351883, + 0.003911775071173906, + 0.004488660488277674, + 0.0032938700169324875, + 0.0030791331082582474, + 0.005456781014800072, + 0.003405593102797866, + 0.0025898930616676807, + 0.007889301516115665, + 0.006479150615632534, + 0.008132269605994225, + 0.014473878778517246, + 0.005435441620647907, + 0.010569530539214611, + 0.007455941289663315, + 0.008945220150053501, + 0.0033663997892290354, + 0.0027107764035463333, + 0.004255752544850111, + 0.0044822050258517265, + 0.0032817574683576822, + 0.00370828271843493, + 0.0056270938366651535, + 0.004831384401768446, + 0.007663920987397432, + 0.003965696785598993, + 0.00915435515344143, + 0.006337545346468687, + 0.005300196353346109, + 0.005572723690420389, + 0.05437598004937172, + 0.05534200742840767, + 0.015644187107682228, + 0.010677993297576904, + 0.014508211985230446, + 0.011589150875806808, + 0.014266888611018658, + 0.014023974537849426, + 0.01892252080142498, + 0.015492253936827183, + 0.006281605456024408, + 0.00649886392056942, + 0.007188465911895037, + 0.004102141596376896, + 0.008145060390233994, + 0.005840379279106855, + 0.006138869561254978, + 0.0066053676418960094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.10876139253377914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076431455090641975 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.0076431455090641975, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076431455090641975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05115014687180519, + 0.04843900725245476, + 0.024983329698443413, + 0.040272608399391174, + 0.04301430284976959, + 0.061406027525663376, + 0.03219371289014816, + 0.3980788588523865, + 0.21679769456386566, + 0.08366433531045914 + ], + "credit_map": [ + 0.002916037803515792, + 0.0026625439058989286, + 0.004166215658187866, + 0.004833671264350414, + 0.008131684735417366, + 0.009599843062460423, + 0.011785587295889854, + 0.007054566405713558, + 0.0038752038963139057, + 0.003819440258666873, + 0.005396478809416294, + 0.008403019048273563, + 0.00768831092864275, + 0.004837140440940857, + 0.01108259242027998, + 0.0033368270378559828, + 0.0019063150975853205, + 0.002033645985648036, + 0.003500227816402912, + 0.004026946146041155, + 0.002717785770073533, + 0.004945230670273304, + 0.002638119040057063, + 0.0032150617334991693, + 0.004312521778047085, + 0.004791873972862959, + 0.006106499582529068, + 0.004382925108075142, + 0.0044761039316654205, + 0.003886289894580841, + 0.007480149623006582, + 0.004836244508624077, + 0.002397657837718725, + 0.005549750756472349, + 0.008732199668884277, + 0.004610193427652121, + 0.004519697744399309, + 0.010334400460124016, + 0.004107821267098188, + 0.0027625844813883305, + 0.007300946861505508, + 0.0066211833618581295, + 0.00683410232886672, + 0.01055161003023386, + 0.005587517283856869, + 0.009879929013550282, + 0.00773803424090147, + 0.0068927123211324215, + 0.002657035831362009, + 0.0023219597060233355, + 0.0032862506341189146, + 0.00472421757876873, + 0.004559425171464682, + 0.003365176497027278, + 0.006727529689669609, + 0.004552117548882961, + 0.015557647682726383, + 0.01814914122223854, + 0.016598178073763847, + 0.018113506957888603, + 0.014443358406424522, + 0.024746891111135483, + 0.14171697199344635, + 0.14875316619873047, + 0.028131820261478424, + 0.0216688122600317, + 0.01861465722322464, + 0.02360878884792328, + 0.018527742475271225, + 0.045049916952848434, + 0.0347890667617321, + 0.0264069102704525, + 0.010103274136781693, + 0.01227569580078125, + 0.010200454853475094, + 0.008376743644475937, + 0.012237616814672947, + 0.010043934918940067, + 0.010458935983479023, + 0.00996768195182085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0076431455090641975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430787019431591 + ], + "opposite_logprob_tokens": [ + -3.168078660964966 + ], + "logprob_reward": -0.0430787019431591, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430787019431591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15371458232402802, + 0.08055354654788971, + 0.0729912668466568, + 0.05978570505976677, + 0.17651787400245667, + 0.04479445517063141, + 0.037406425923109055, + 0.2242240458726883, + 0.07169178128242493, + 0.07832030206918716 + ], + "credit_map": [ + 0.009726274758577347, + 0.007121147122234106, + 0.016337957233190536, + 0.015488114207983017, + 0.023943135514855385, + 0.02688336931169033, + 0.03248671442270279, + 0.02172786369919777, + 0.0049024345353245735, + 0.004862469621002674, + 0.006642160005867481, + 0.010751865804195404, + 0.01636766828596592, + 0.008011209778487682, + 0.021756071597337723, + 0.007259670179337263, + 0.005273072514683008, + 0.007558257784694433, + 0.01032724417746067, + 0.010515705682337284, + 0.009407855570316315, + 0.011801596730947495, + 0.011140174232423306, + 0.0069673629477620125, + 0.005421868525445461, + 0.005936685483902693, + 0.012454976327717304, + 0.0073023708537220955, + 0.005960294511169195, + 0.006063040811568499, + 0.008538195863366127, + 0.008108271285891533, + 0.008054499514400959, + 0.006208138540387154, + 0.06376903504133224, + 0.0063763950020074844, + 0.008597485721111298, + 0.0699680894613266, + 0.007870991714298725, + 0.005673250649124384, + 0.006376321893185377, + 0.003629349172115326, + 0.0047650714404881, + 0.007631792686879635, + 0.0026569203473627567, + 0.007380870170891285, + 0.0054664332419633865, + 0.006887699477374554, + 0.003175507765263319, + 0.002706747269257903, + 0.005486528389155865, + 0.0057291691191494465, + 0.004343947861343622, + 0.002825941424816847, + 0.007629677187651396, + 0.005508905742317438, + 0.010533539578318596, + 0.00859504658728838, + 0.009896796196699142, + 0.008768491446971893, + 0.009840608574450016, + 0.011066948063671589, + 0.08075018227100372, + 0.08477243781089783, + 0.009028268977999687, + 0.00786389596760273, + 0.012430963106453419, + 0.005203021224588156, + 0.008872771635651588, + 0.010602892376482487, + 0.010120291262865067, + 0.007569677662104368, + 0.00733319902792573, + 0.01061759889125824, + 0.009439517743885517, + 0.008170568384230137, + 0.011138934642076492, + 0.009564717300236225, + 0.013038377277553082, + 0.009017392992973328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0430787019431591 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419447734951973 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.01419447734951973, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419447734951973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15256229043006897, + 0.045704033225774765, + 0.04665375128388405, + 0.032108042389154434, + 0.08341433852910995, + 0.11563541740179062, + 0.03772376477718353, + 0.3842131495475769, + 0.0569494403898716, + 0.045035794377326965 + ], + "credit_map": [ + 0.007130756508558989, + 0.006457651033997536, + 0.01213084813207388, + 0.012601668015122414, + 0.02639656327664852, + 0.029828447848558426, + 0.03481726720929146, + 0.023199087008833885, + 0.003056913148611784, + 0.003366091987118125, + 0.0055243284441530704, + 0.0057417359203100204, + 0.007829230278730392, + 0.004371811170130968, + 0.012557066045701504, + 0.003256852040067315, + 0.005299380049109459, + 0.005127409007400274, + 0.006370700895786285, + 0.007037427742034197, + 0.005990900099277496, + 0.005428696051239967, + 0.005375174805521965, + 0.00602406170219183, + 0.002820160472765565, + 0.0028958646580576897, + 0.004178463015705347, + 0.004426517058163881, + 0.0041796863079071045, + 0.0037696952931582928, + 0.006044148001819849, + 0.0037935057189315557, + 0.004107877146452665, + 0.0033805393613874912, + 0.02607462927699089, + 0.005388419609516859, + 0.005205160938203335, + 0.02854219824075699, + 0.005298948846757412, + 0.005416571162641048, + 0.015864109620451927, + 0.01279495283961296, + 0.01385619305074215, + 0.018561556935310364, + 0.009005137719213963, + 0.018220052123069763, + 0.01420411467552185, + 0.013129303231835365, + 0.004336883779615164, + 0.003803891595453024, + 0.004927024245262146, + 0.00590858468785882, + 0.00569216487929225, + 0.002683265833184123, + 0.005700506269931793, + 0.004671442788094282, + 0.02004283294081688, + 0.013411330990493298, + 0.0173487588763237, + 0.014646092429757118, + 0.01299883984029293, + 0.015354404225945473, + 0.1421503722667694, + 0.14826053380966187, + 0.00799439288675785, + 0.005108329001814127, + 0.009274129755795002, + 0.00442799087613821, + 0.007022749166935682, + 0.0069337086752057076, + 0.009908068925142288, + 0.006280068773776293, + 0.005306342151015997, + 0.007763371802866459, + 0.004756088834255934, + 0.005168691277503967, + 0.005919175688177347, + 0.004871836397796869, + 0.005500175058841705, + 0.005750110372900963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.01419447734951973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032638907432556 + ], + "opposite_logprob_tokens": [ + -2.3503262996673584 + ], + "logprob_reward": -0.10032638907432556, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032638907432556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29531753063201904, + 0.055639125406742096, + 0.07339887320995331, + 0.060173939913511276, + 0.2501324415206909, + 0.0655696764588356, + 0.08153890073299408, + 0.057412128895521164, + 0.025388291105628014, + 0.03542908653616905 + ], + "credit_map": [ + 0.017057841643691063, + 0.015229295007884502, + 0.02499462477862835, + 0.02498551644384861, + 0.05451090261340141, + 0.055724672973155975, + 0.06458399444818497, + 0.03823067247867584, + 0.003296957118436694, + 0.004814907908439636, + 0.005916560534387827, + 0.008789988234639168, + 0.00865321233868599, + 0.005869810003787279, + 0.009115565568208694, + 0.009182125329971313, + 0.005934469867497683, + 0.007403376046568155, + 0.010642405599355698, + 0.014037925750017166, + 0.008418350480496883, + 0.012175817973911762, + 0.008424442261457443, + 0.006362088955938816, + 0.006909321062266827, + 0.005805795546621084, + 0.007696719840168953, + 0.00733485072851181, + 0.006696474738419056, + 0.006930741481482983, + 0.009003006853163242, + 0.009797032922506332, + 0.008853703737258911, + 0.006219614762812853, + 0.09418535977602005, + 0.00865287333726883, + 0.008108659647405148, + 0.1045643538236618, + 0.011383616365492344, + 0.008164294064044952, + 0.009195096790790558, + 0.007330726366490126, + 0.007778114173561335, + 0.014358232729136944, + 0.004873038735240698, + 0.009089838713407516, + 0.006209117826074362, + 0.006735514383763075, + 0.009959460236132145, + 0.006118936464190483, + 0.01516022253781557, + 0.010549238882958889, + 0.0055382815189659595, + 0.007583473343402147, + 0.014209183864295483, + 0.0124201076105237, + 0.003776669967919588, + 0.007348370738327503, + 0.004734459333121777, + 0.004420979414135218, + 0.004584676586091518, + 0.00431601470336318, + 0.013673213310539722, + 0.01455774623900652, + 0.0035842687357217073, + 0.0023773610591888428, + 0.0027198714669793844, + 0.0022406696807593107, + 0.0023016235791146755, + 0.004354274366050959, + 0.0041557601653039455, + 0.0036544636823236942, + 0.004130283370614052, + 0.003206230467185378, + 0.004736615810543299, + 0.004088544752448797, + 0.003594330744817853, + 0.006282434798777103, + 0.004633273929357529, + 0.004757375922054052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.10032638907432556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dizzying high-angle tracking shot swoops vertically downward, rapidly following a skeletal biomechanical creature in a terrifyingly steep dive. The creature's form is a Giger-esque fusion of bone and machine; a spindly, ribbed spine visibly merges with hydraulic pistons and dark, pulsating pipes, while its smooth, obsidian skull is eyeless, save for a faint, inner amber glow emanating from its empty eye sockets. Its massive, leathery wings, textured like scarred hide, are supported by metal struts and partially fold back to increase speed, only occasionally adjusting its trajectory with powerful, deliberate flaps. The camera follows its terrifying dive across a narrow, oppressive chasm separating vast floating islands. These islands are no mere rocks, but nightmarish, living landscapes of ribbed, chitinous structures fused with dark, damp stone, their bases dripping with viscous fluids and tangled, tubular organic growths. The walls of the abyss are adorned with twisting metallic vines of alien flora and pulsating bioluminescent sacs, casting a sickly, cool green light. A thick, swirling mist rises from the abyss below, obscuring the final destination. Rendered in a desaturated, monochromatic palette of gunmetal grays and wet blacks, with specular highlights gleaming off the metallic parts of the creatures, they create an atmosphere of profound, cosmic dread and vertigo as both creature and camera plunge inexorably into the unknown.", + "seed": 10106, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_11/index_6815_seed_10106_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08676280742511153, + "vlm_logprob": -0.08676280742511153, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.08676280742511153, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32356441020965576 + ], + "opposite_logprob_tokens": [ + -1.3235644102096558 + ], + "logprob_reward": -0.32356441020965576, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32356441020965576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028937049210071564, + 0.025393947958946228, + 0.6305766701698303, + 0.028110696002840996, + 0.04740468040108681, + 0.01509007252752781, + 0.07873348891735077, + 0.09396035969257355, + 0.02952924743294716, + 0.022263703867793083 + ], + "credit_map": [ + 0.0026769235264509916, + 0.0032084013801068068, + 0.003885735059157014, + 0.003194367978721857, + 0.0033130787778645754, + 0.004328172188252211, + 0.004383496008813381, + 0.003946877550333738, + 0.0021767602302134037, + 0.00290876068174839, + 0.003004679223522544, + 0.003553189570084214, + 0.005312016699463129, + 0.003554960247129202, + 0.0030144844204187393, + 0.0018690972356125712, + 0.05355844646692276, + 0.008853599429130554, + 0.2623633146286011, + 0.0156535767018795, + 0.013074781745672226, + 0.03080160729587078, + 0.01373087614774704, + 0.2325405478477478, + 0.0029904916882514954, + 0.0033897841349244118, + 0.004053002223372459, + 0.004615879151970148, + 0.0028610164299607277, + 0.003656307701021433, + 0.0023949299938976765, + 0.004149287473410368, + 0.005831094458699226, + 0.005590768530964851, + 0.006416005082428455, + 0.006890230346471071, + 0.004694532603025436, + 0.006129866931587458, + 0.006393043324351311, + 0.0054591442458331585, + 0.0017824260285124183, + 0.0013647639425471425, + 0.0021368220914155245, + 0.001817454001866281, + 0.0018173079006373882, + 0.0019646144937723875, + 0.0013556159101426601, + 0.0028510696720331907, + 0.005411227699369192, + 0.003714799415320158, + 0.023189444094896317, + 0.006321728229522705, + 0.003794362535700202, + 0.009425743483006954, + 0.0050425417721271515, + 0.021833645179867744, + 0.007339518517255783, + 0.00382151803933084, + 0.03116784244775772, + 0.004731477238237858, + 0.004040096886456013, + 0.008146727457642555, + 0.005343086551874876, + 0.029370103031396866, + 0.003034461522474885, + 0.0031772516667842865, + 0.0029713958501815796, + 0.0034898805897682905, + 0.003779875347390771, + 0.003934449050575495, + 0.004892138764262199, + 0.004249795340001583, + 0.002858856925740838, + 0.0020457562059164047, + 0.0020258170552551746, + 0.0033496576361358166, + 0.0037563254591077566, + 0.00342784496024251, + 0.0022677872329950333, + 0.0025316595565527678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.32356441020965576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002212697174400091 + ], + "opposite_logprob_tokens": [ + -6.1272125244140625 + ], + "logprob_reward": -0.002212697174400091, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002212697174400091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10426072031259537, + 0.06215976923704147, + 0.24831806123256683, + 0.2082078605890274, + 0.03939446061849594, + 0.06186089664697647, + 0.04976062476634979, + 0.025018328800797462, + 0.11829958111047745, + 0.08271973580121994 + ], + "credit_map": [ + 0.008357993327081203, + 0.0061299740336835384, + 0.013011706061661243, + 0.012690217234194279, + 0.015413330867886543, + 0.01645466312766075, + 0.012187272310256958, + 0.020015552639961243, + 0.005124641582369804, + 0.004606836475431919, + 0.006607833784073591, + 0.008430074900388718, + 0.009789576753973961, + 0.007245596963912249, + 0.011220306158065796, + 0.009134897962212563, + 0.01966407708823681, + 0.009572353214025497, + 0.08880600333213806, + 0.010377359576523304, + 0.013471297919750214, + 0.012689691968262196, + 0.015405546873807907, + 0.07833170145750046, + 0.021130437031388283, + 0.029951199889183044, + 0.027211252599954605, + 0.03808486461639404, + 0.020426757633686066, + 0.029988383874297142, + 0.01746280863881111, + 0.02395213581621647, + 0.006951322313398123, + 0.003983229398727417, + 0.005115740466862917, + 0.0054991524666547775, + 0.003882715478539467, + 0.004507829435169697, + 0.003468545153737068, + 0.005985922180116177, + 0.005970603786408901, + 0.007896794937551022, + 0.007794458884745836, + 0.009394248016178608, + 0.0078078159131109715, + 0.00755145912989974, + 0.006542822811752558, + 0.008902688510715961, + 0.005667959805577993, + 0.004552126862108707, + 0.007316139526665211, + 0.008038944564759731, + 0.004465524572879076, + 0.00953088328242302, + 0.004574344027787447, + 0.00561469467356801, + 0.0029395450837910175, + 0.003037786576896906, + 0.0030079279094934464, + 0.003116287523880601, + 0.002443744335323572, + 0.0038649628404527903, + 0.0019192324252799153, + 0.004688841290771961, + 0.018863996490836143, + 0.014598132111132145, + 0.016087869182229042, + 0.013022167608141899, + 0.014133607037365437, + 0.015278765931725502, + 0.012049557641148567, + 0.014265474863350391, + 0.01173466071486473, + 0.008202776312828064, + 0.008095716126263142, + 0.008219120092689991, + 0.011527933180332184, + 0.0141064478084445, + 0.00962354987859726, + 0.011209524236619473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.002212697174400091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807347804307938 + ], + "opposite_logprob_tokens": [ + -3.5298073291778564 + ], + "logprob_reward": -0.029807347804307938, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807347804307938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056475549936294556, + 0.0591350682079792, + 0.3384972810745239, + 0.10548462718725204, + 0.019102780148386955, + 0.049611836671829224, + 0.11094380170106888, + 0.1657211035490036, + 0.05095439776778221, + 0.04407355189323425 + ], + "credit_map": [ + 0.0055109974928200245, + 0.004181071184575558, + 0.007495956029742956, + 0.00855319295078516, + 0.007660462986677885, + 0.00992548931390047, + 0.006169217638671398, + 0.006979166995733976, + 0.0038047118578106165, + 0.007191913202404976, + 0.008649267256259918, + 0.0045593599788844585, + 0.014291582629084587, + 0.007496772799640894, + 0.009666334837675095, + 0.003475127974525094, + 0.027181122452020645, + 0.00818879995495081, + 0.13669630885124207, + 0.009776582941412926, + 0.009142701514065266, + 0.017633480951189995, + 0.010016373358666897, + 0.1198619082570076, + 0.012529860250651836, + 0.014542339369654655, + 0.013806529343128204, + 0.01949380710721016, + 0.009510091505944729, + 0.014684611931443214, + 0.008177859708666801, + 0.012739532627165318, + 0.0033837310038506985, + 0.0022156094200909138, + 0.0025877286680042744, + 0.002551212441176176, + 0.001757679507136345, + 0.0018919390859082341, + 0.00275324541144073, + 0.0019616347271949053, + 0.005762146320194006, + 0.0058262404054403305, + 0.0063726031221449375, + 0.00716830650344491, + 0.004863385576754808, + 0.007707429118454456, + 0.004763806238770485, + 0.00714791938662529, + 0.009886188432574272, + 0.006956022698432207, + 0.026821617037057877, + 0.012335900217294693, + 0.007437102496623993, + 0.013408111408352852, + 0.008516332134604454, + 0.02558252587914467, + 0.012656032107770443, + 0.004784547723829746, + 0.052711039781570435, + 0.008255710825324059, + 0.006594436708837748, + 0.01590454764664173, + 0.008070208132266998, + 0.05674458295106888, + 0.00683807535097003, + 0.008282299153506756, + 0.004816695116460323, + 0.005373121704906225, + 0.004471975844353437, + 0.007070871070027351, + 0.007990820333361626, + 0.006110539194196463, + 0.0031447471119463444, + 0.004337682854384184, + 0.002867685165256262, + 0.006244325079023838, + 0.009000683203339577, + 0.009970156475901604, + 0.0030890395864844322, + 0.0054192328825592995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.029807347804307938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279571905732155 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279571905732155, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279571905732155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15907888114452362, + 0.11994657665491104, + 0.18874308466911316, + 0.03652120381593704, + 0.031056392937898636, + 0.0259991567581892, + 0.06546800583600998, + 0.28876930475234985, + 0.03627920150756836, + 0.04813819006085396 + ], + "credit_map": [ + 0.014354837127029896, + 0.012029262259602547, + 0.02186521142721176, + 0.023181132972240448, + 0.02238973043859005, + 0.03140776604413986, + 0.018159646540880203, + 0.01569128967821598, + 0.01080500241369009, + 0.00801550317555666, + 0.01641417108476162, + 0.010698076337575912, + 0.018042199313640594, + 0.027256447821855545, + 0.01511772908270359, + 0.013597450219094753, + 0.015072164125740528, + 0.004611132200807333, + 0.07546289265155792, + 0.005702508147805929, + 0.0059809437952935696, + 0.00932297483086586, + 0.006268492434173822, + 0.0663219690322876, + 0.004353587981313467, + 0.004710312932729721, + 0.0042739310301840305, + 0.006724422797560692, + 0.003862815210595727, + 0.0036247982643544674, + 0.004482378717511892, + 0.0044889566488564014, + 0.0038923074025660753, + 0.0028285018634051085, + 0.003995376639068127, + 0.00456929299980402, + 0.003699642140418291, + 0.003307692240923643, + 0.004261086694896221, + 0.004502493888139725, + 0.0021682570222765207, + 0.002085337182506919, + 0.0048963879235088825, + 0.003557265968993306, + 0.0027797375805675983, + 0.00410826038569212, + 0.0026925893034785986, + 0.0037113234866410494, + 0.004960998892784119, + 0.003265104489400983, + 0.017977669835090637, + 0.006296251434832811, + 0.00445950124412775, + 0.009109195321798325, + 0.004357621073722839, + 0.015041661448776722, + 0.0222827959805727, + 0.010708565823733807, + 0.096380315721035, + 0.012126185931265354, + 0.014283396303653717, + 0.024825431406497955, + 0.015328255482017994, + 0.09283438324928284, + 0.004033616743981838, + 0.006028468254953623, + 0.0034572870936244726, + 0.004280621185898781, + 0.002972930669784546, + 0.004972151014953852, + 0.0055997963063418865, + 0.004934333730489016, + 0.005731003358960152, + 0.004781236406415701, + 0.004799957852810621, + 0.004262738861143589, + 0.008133475668728352, + 0.008404036052525043, + 0.004942815285176039, + 0.007082928437739611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.023279571905732155 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054950010031461716 + ], + "opposite_logprob_tokens": [ + -2.929949998855591 + ], + "logprob_reward": -0.054950010031461716, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054950010031461716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03329820930957794, + 0.03812435641884804, + 0.6743894219398499, + 0.046409208327531815, + 0.035984642803668976, + 0.0296555757522583, + 0.019249271601438522, + 0.0946170762181282, + 0.016431352123618126, + 0.011840893886983395 + ], + "credit_map": [ + 0.00381786678917706, + 0.002454356988891959, + 0.004653513431549072, + 0.0038180637639015913, + 0.0028349694330245256, + 0.005413992330431938, + 0.0041571687906980515, + 0.0061482759192585945, + 0.0028859612066298723, + 0.0026162376161664724, + 0.005141026806086302, + 0.004336673766374588, + 0.008867071010172367, + 0.004835073370486498, + 0.007223603315651417, + 0.0022187093272805214, + 0.05579575151205063, + 0.010347094386816025, + 0.2790014445781708, + 0.017243357375264168, + 0.01456571463495493, + 0.03192149102687836, + 0.018201656639575958, + 0.24731290340423584, + 0.009278570301830769, + 0.0044512078166007996, + 0.007214323617517948, + 0.004004755523055792, + 0.003776337718591094, + 0.004057909827679396, + 0.0061185769736766815, + 0.007507522590458393, + 0.00574799720197916, + 0.0035903987009078264, + 0.00476954085752368, + 0.00423653656616807, + 0.003794838208705187, + 0.004437481053173542, + 0.0038332263939082623, + 0.005574620794504881, + 0.0035826487001031637, + 0.0030861985869705677, + 0.0036566827911883593, + 0.00451158219948411, + 0.0030015837401151657, + 0.003568218322470784, + 0.0032019687350839376, + 0.005046692211180925, + 0.002959389938041568, + 0.0024072967935353518, + 0.003026073332875967, + 0.002484002383425832, + 0.0018816383089870214, + 0.0021797381341457367, + 0.001959682209417224, + 0.0023514502681791782, + 0.00706203980371356, + 0.003017576178535819, + 0.03378785401582718, + 0.003881205804646015, + 0.002998883370310068, + 0.008694186806678772, + 0.0027318643406033516, + 0.03244346007704735, + 0.00274225533939898, + 0.002443580888211727, + 0.0016259014373645186, + 0.0016670305049046874, + 0.0017392031149938703, + 0.0022126638796180487, + 0.0022696766536682844, + 0.001731039839796722, + 0.0009297093492932618, + 0.0007444773218594491, + 0.0006197773036547005, + 0.0010475234594196081, + 0.003161817789077759, + 0.003582476871088147, + 0.0007497883634641767, + 0.001005323021672666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.054950010031461716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dizzying high-angle tracking shot swoops vertically downward, rapidly following a skeletal biomechanical creature in a terrifyingly steep dive. The creature's form is a Giger-esque fusion of bone and machine; a spindly, ribbed spine visibly merges with hydraulic pistons and dark, pulsating pipes, while its smooth, obsidian skull is eyeless, save for a faint, inner amber glow emanating from its empty eye sockets. Its massive, leathery wings, textured like scarred hide, are supported by metal struts and partially fold back to increase speed, only occasionally adjusting its trajectory with powerful, deliberate flaps. The camera follows its terrifying dive across a narrow, oppressive chasm separating vast floating islands. These islands are no mere rocks, but nightmarish, living landscapes of ribbed, chitinous structures fused with dark, damp stone, their bases dripping with viscous fluids and tangled, tubular organic growths. The walls of the abyss are adorned with twisting metallic vines of alien flora and pulsating bioluminescent sacs, casting a sickly, cool green light. A thick, swirling mist rises from the abyss below, obscuring the final destination. Rendered in a desaturated, monochromatic palette of gunmetal grays and wet blacks, with specular highlights gleaming off the metallic parts of the creatures, they create an atmosphere of profound, cosmic dread and vertigo as both creature and camera plunge inexorably into the unknown.", + "seed": 10107, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_11/index_6815_seed_10107_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.045813147956505415, + "vlm_logprob": -0.045813147956505415, + "vlm_yes_no_margin": 3.950000023841858, + "vlm_token_credit": -0.045813147956505415, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16252292692661285 + ], + "opposite_logprob_tokens": [ + -2.037523031234741 + ], + "logprob_reward": -0.16252292692661285, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.16252292692661285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20509883761405945, + 0.16703589260578156, + 0.05044462904334068, + 0.10454215854406357, + 0.1402374505996704, + 0.07575058937072754, + 0.046097200363874435, + 0.05661482363939285, + 0.11587728559970856, + 0.03830121457576752 + ], + "credit_map": [ + 0.0173927191644907, + 0.02207227796316147, + 0.01890288107097149, + 0.040930166840553284, + 0.017321405932307243, + 0.03488001227378845, + 0.023769792169332504, + 0.029829557985067368, + 0.01623455062508583, + 0.027099108323454857, + 0.011436414904892445, + 0.04230721294879913, + 0.01417680736631155, + 0.023370135575532913, + 0.011897156946361065, + 0.0205144714564085, + 0.0068432725965976715, + 0.005366964265704155, + 0.005465647205710411, + 0.00659336568787694, + 0.008002044633030891, + 0.00638110376894474, + 0.0062826224602758884, + 0.005509603302925825, + 0.005454298574477434, + 0.008763855323195457, + 0.00633595883846283, + 0.006049724295735359, + 0.018281374126672745, + 0.027236301451921463, + 0.005598349496722221, + 0.026822282001376152, + 0.007515604607760906, + 0.0048546870239079, + 0.0071997870691120625, + 0.005400151014328003, + 0.00914415717124939, + 0.04494728893041611, + 0.016672790050506592, + 0.044502973556518555, + 0.004419364966452122, + 0.005414362996816635, + 0.006534949876368046, + 0.008827194571495056, + 0.01000035461038351, + 0.014692586846649647, + 0.011066404171288013, + 0.014795362949371338, + 0.004910193383693695, + 0.003344145370647311, + 0.006445079576224089, + 0.00869639404118061, + 0.0040857079438865185, + 0.005737576633691788, + 0.006664137821644545, + 0.006213962100446224, + 0.008124236948788166, + 0.009247357957065105, + 0.008139869198203087, + 0.007791031617671251, + 0.006091749761253595, + 0.005411040503531694, + 0.005424472037702799, + 0.006385058630257845, + 0.00932223629206419, + 0.011414367705583572, + 0.016167698428034782, + 0.014964791014790535, + 0.011835464276373386, + 0.008591348305344582, + 0.031718891113996506, + 0.011862481012940407, + 0.005392160266637802, + 0.004784501623362303, + 0.00481859315186739, + 0.0042023127898573875, + 0.004514419008046389, + 0.003651553066447377, + 0.0057669756934046745, + 0.005170690827071667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.16252292692661285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001957169035449624 + ], + "opposite_logprob_tokens": [ + -6.251956939697266 + ], + "logprob_reward": -0.001957169035449624, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001957169035449624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09275741875171661, + 0.056666988879442215, + 0.03725707530975342, + 0.04004720598459244, + 0.24143414199352264, + 0.11031762510538101, + 0.11976368725299835, + 0.146214559674263, + 0.11794041097164154, + 0.03760084882378578 + ], + "credit_map": [ + 0.010028200224041939, + 0.009584826417267323, + 0.009650137275457382, + 0.01658938080072403, + 0.009828086942434311, + 0.01181495375931263, + 0.012552183121442795, + 0.01270965114235878, + 0.0063042850233614445, + 0.007135759107768536, + 0.005379293579608202, + 0.010564223863184452, + 0.004680562764406204, + 0.009240084327757359, + 0.005383864976465702, + 0.007978915236890316, + 0.0040527042001485825, + 0.0035939207300543785, + 0.0035406032111495733, + 0.005830714479088783, + 0.0066708954982459545, + 0.005118694622069597, + 0.004309746436774731, + 0.004139797296375036, + 0.0019374329131096601, + 0.002545952796936035, + 0.0017653461545705795, + 0.0036936490796506405, + 0.004922832828015089, + 0.011531013064086437, + 0.0023949157912284136, + 0.011256062425673008, + 0.003955597523599863, + 0.0037914684507995844, + 0.00829285103827715, + 0.004050730261951685, + 0.0066677783615887165, + 0.10455609858036041, + 0.014646269381046295, + 0.09547334909439087, + 0.006517817731946707, + 0.010675407946109772, + 0.006811094935983419, + 0.011965877376496792, + 0.025473568588495255, + 0.019322024658322334, + 0.015603754669427872, + 0.013948075473308563, + 0.011273708194494247, + 0.0072317360900342464, + 0.008762896992266178, + 0.012018369510769844, + 0.010475863702595234, + 0.01235516369342804, + 0.027796266600489616, + 0.02984968014061451, + 0.01783091574907303, + 0.020629629492759705, + 0.02035604789853096, + 0.026577992364764214, + 0.013985778205096722, + 0.015028038062155247, + 0.016676511615514755, + 0.01512964628636837, + 0.0110792750492692, + 0.014602490700781345, + 0.01644248142838478, + 0.015310468152165413, + 0.011918758973479271, + 0.01322098821401596, + 0.022625630721449852, + 0.012740319594740868, + 0.004844517447054386, + 0.006339202169328928, + 0.004829233046621084, + 0.006898683495819569, + 0.0037474236451089382, + 0.0035039482172578573, + 0.0043851337395608425, + 0.0030527086928486824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.001957169035449624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808156192302704 + ], + "opposite_logprob_tokens": [ + -3.5298080444335938 + ], + "logprob_reward": -0.029808156192302704, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808156192302704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10226637870073318, + 0.04634321108460426, + 0.13521936535835266, + 0.07840832322835922, + 0.1399707943201065, + 0.06467167288064957, + 0.2014763057231903, + 0.1057080551981926, + 0.05552152544260025, + 0.07041432708501816 + ], + "credit_map": [ + 0.0130851399153471, + 0.012937180697917938, + 0.008678914979100227, + 0.0182220246642828, + 0.00991033110767603, + 0.016143934801220894, + 0.009275018237531185, + 0.01401383988559246, + 0.00563190458342433, + 0.00570529792457819, + 0.004520188085734844, + 0.009781277738511562, + 0.0032104142010211945, + 0.006940717343240976, + 0.004598569590598345, + 0.005954842548817396, + 0.011912700720131397, + 0.01201885286718607, + 0.01240523625165224, + 0.018216222524642944, + 0.024798346683382988, + 0.024034729227423668, + 0.010246370919048786, + 0.021586917340755463, + 0.00395203847438097, + 0.00749527383595705, + 0.003573421388864517, + 0.0038075426127761602, + 0.023786667734384537, + 0.01596832647919655, + 0.00565933994948864, + 0.014165716245770454, + 0.004988056607544422, + 0.0036655343137681484, + 0.006037732120603323, + 0.0035069570876657963, + 0.006757767405360937, + 0.05433788523077965, + 0.011273994110524654, + 0.049402862787246704, + 0.004480945412069559, + 0.006889534182846546, + 0.005377443041652441, + 0.007880798541009426, + 0.009245096705853939, + 0.010589945130050182, + 0.011757384054362774, + 0.008450527675449848, + 0.02361481636762619, + 0.01215705182403326, + 0.015711072832345963, + 0.021503819152712822, + 0.0185033418238163, + 0.022119006142020226, + 0.04626317322254181, + 0.041604019701480865, + 0.009564418345689774, + 0.012569083832204342, + 0.012665458023548126, + 0.019608527421951294, + 0.012857260182499886, + 0.00968250259757042, + 0.015351506881415844, + 0.013409297913312912, + 0.006289203185588121, + 0.0068815043196082115, + 0.012606650590896606, + 0.007138712797313929, + 0.0056133572943508625, + 0.005075821653008461, + 0.006181941833347082, + 0.005734333768486977, + 0.009721221402287483, + 0.009693610481917858, + 0.009267562068998814, + 0.010075002908706665, + 0.008854408748447895, + 0.008923235349357128, + 0.007450693286955357, + 0.006428596563637257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.029808156192302704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419494766741991 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.01419494766741991, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419494766741991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21824754774570465, + 0.05971047654747963, + 0.1396961212158203, + 0.059155192226171494, + 0.035208810120821, + 0.0618656650185585, + 0.10499362647533417, + 0.06636794656515121, + 0.12130212783813477, + 0.13345257937908173 + ], + "credit_map": [ + 0.01928732544183731, + 0.028067246079444885, + 0.022174878045916557, + 0.0496210977435112, + 0.020231183618307114, + 0.027813300490379333, + 0.02534356713294983, + 0.025708936154842377, + 0.008172061294317245, + 0.005684548988938332, + 0.006319845095276833, + 0.011292117647826672, + 0.0047982181422412395, + 0.008822765201330185, + 0.007186229340732098, + 0.007434689439833164, + 0.01686370186507702, + 0.016378439962863922, + 0.013137426227331161, + 0.014515961520373821, + 0.029759811237454414, + 0.01878463663160801, + 0.01456126756966114, + 0.015694869682192802, + 0.0033859889954328537, + 0.0050299097783863544, + 0.0029029352590441704, + 0.0031449347734451294, + 0.012121535837650299, + 0.014739791862666607, + 0.0035851378925144672, + 0.01424495130777359, + 0.004304484464228153, + 0.002779326168820262, + 0.0043968139216303825, + 0.0030354729387909174, + 0.00565696693956852, + 0.0059601664543151855, + 0.0051734852604568005, + 0.0039020904805511236, + 0.004060444887727499, + 0.005560981575399637, + 0.0055360570549964905, + 0.006409354507923126, + 0.010189940221607685, + 0.010828610509634018, + 0.00927167758345604, + 0.010008596815168858, + 0.004680516663938761, + 0.0031608950812369585, + 0.003963758237659931, + 0.005178652238100767, + 0.008691616356372833, + 0.008194324560463428, + 0.03632891923189163, + 0.034794945269823074, + 0.007141094654798508, + 0.009391603991389275, + 0.010504077188670635, + 0.011581320315599442, + 0.006585163529962301, + 0.006646407302469015, + 0.007005988620221615, + 0.007512285839766264, + 0.0120366420596838, + 0.014777615666389465, + 0.024191632866859436, + 0.012557199224829674, + 0.013863496482372284, + 0.018305057659745216, + 0.01192733645439148, + 0.01364313717931509, + 0.015245543792843819, + 0.016024000942707062, + 0.01895374059677124, + 0.018911996856331825, + 0.01762511022388935, + 0.009693902917206287, + 0.01874273270368576, + 0.018255528062582016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.01419494766741991 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582539960741997 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582539960741997, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582539960741997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17140977084636688, + 0.15002012252807617, + 0.0836191177368164, + 0.050427064299583435, + 0.07241619378328323, + 0.09690692275762558, + 0.09079243242740631, + 0.1673297882080078, + 0.05270400643348694, + 0.064374640583992 + ], + "credit_map": [ + 0.01490797009319067, + 0.023861659690737724, + 0.01737387105822563, + 0.028406186029314995, + 0.0195945855230093, + 0.023703571408987045, + 0.021494124084711075, + 0.02206776663661003, + 0.022779317572712898, + 0.016218559816479683, + 0.01595008373260498, + 0.023526201024651527, + 0.00935785286128521, + 0.026520224288105965, + 0.010801729746162891, + 0.02486613765358925, + 0.005711008328944445, + 0.014732401818037033, + 0.006287757307291031, + 0.01493370346724987, + 0.010709874331951141, + 0.011538108810782433, + 0.00844800379127264, + 0.011258251965045929, + 0.0031988800037652254, + 0.00403711898252368, + 0.0027768267318606377, + 0.002960814628750086, + 0.010016871616244316, + 0.012636612169444561, + 0.0035470817238092422, + 0.011252853088080883, + 0.005351489409804344, + 0.0042551555670797825, + 0.0052908603101968765, + 0.005484443623572588, + 0.005429093725979328, + 0.018497750163078308, + 0.012196909636259079, + 0.01591048203408718, + 0.010660402476787567, + 0.0070557985454797745, + 0.009115951135754585, + 0.010451833717525005, + 0.007923251017928123, + 0.01742074452340603, + 0.017872998490929604, + 0.01640593260526657, + 0.003816177835687995, + 0.003076470224186778, + 0.0048361909575760365, + 0.00856042467057705, + 0.003750087460502982, + 0.005434661637991667, + 0.030677516013383865, + 0.030640896409749985, + 0.016139784827828407, + 0.01727432943880558, + 0.022875651717185974, + 0.02726169116795063, + 0.011351646855473518, + 0.011404621414840221, + 0.03889656811952591, + 0.0221254862844944, + 0.00530144153162837, + 0.007315813563764095, + 0.007797910366207361, + 0.009553959593176842, + 0.006067180540412664, + 0.004530911333858967, + 0.007301768753677607, + 0.004835016094148159, + 0.008867565542459488, + 0.007774990983307362, + 0.0075334832072257996, + 0.01007123850286007, + 0.0077596320770680904, + 0.009962906129658222, + 0.006022788118571043, + 0.006382034160196781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.020582539960741997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dizzying high-angle tracking shot swoops vertically downward, rapidly following a skeletal biomechanical creature in a terrifyingly steep dive. The creature's form is a Giger-esque fusion of bone and machine; a spindly, ribbed spine visibly merges with hydraulic pistons and dark, pulsating pipes, while its smooth, obsidian skull is eyeless, save for a faint, inner amber glow emanating from its empty eye sockets. Its massive, leathery wings, textured like scarred hide, are supported by metal struts and partially fold back to increase speed, only occasionally adjusting its trajectory with powerful, deliberate flaps. The camera follows its terrifying dive across a narrow, oppressive chasm separating vast floating islands. These islands are no mere rocks, but nightmarish, living landscapes of ribbed, chitinous structures fused with dark, damp stone, their bases dripping with viscous fluids and tangled, tubular organic growths. The walls of the abyss are adorned with twisting metallic vines of alien flora and pulsating bioluminescent sacs, casting a sickly, cool green light. A thick, swirling mist rises from the abyss below, obscuring the final destination. Rendered in a desaturated, monochromatic palette of gunmetal grays and wet blacks, with specular highlights gleaming off the metallic parts of the creatures, they create an atmosphere of profound, cosmic dread and vertigo as both creature and camera plunge inexorably into the unknown.", + "seed": 10108, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_11/index_6815_seed_10108_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02130828853696585, + "vlm_logprob": -0.02130828853696585, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.02130828853696585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027959860861301422 + ], + "opposite_logprob_tokens": [ + -3.7779598236083984 + ], + "logprob_reward": -0.027959860861301422, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027959860861301422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06354384869337082, + 0.041210196912288666, + 0.028907356783747673, + 0.03150514140725136, + 0.10701780021190643, + 0.05957481265068054, + 0.562178373336792, + 0.034176018089056015, + 0.03418400511145592, + 0.03770238906145096 + ], + "credit_map": [ + 0.005717165768146515, + 0.003795469645410776, + 0.005293664988130331, + 0.006360327359288931, + 0.008863774128258228, + 0.014542508870363235, + 0.012963618151843548, + 0.006007324904203415, + 0.004033252596855164, + 0.004027982708066702, + 0.006581407506018877, + 0.004340312443673611, + 0.004251757636666298, + 0.006133960094302893, + 0.005866080988198519, + 0.0059754434041678905, + 0.0024382607080042362, + 0.0026788676623255014, + 0.0027460185810923576, + 0.0037260164972394705, + 0.003958221524953842, + 0.004519440233707428, + 0.005378774367272854, + 0.003461760003119707, + 0.0027504649478942156, + 0.0017699021846055984, + 0.00276249460875988, + 0.0023784441873431206, + 0.0019465327495709062, + 0.008668467402458191, + 0.002322135493159294, + 0.008906702511012554, + 0.008678940124809742, + 0.010117515921592712, + 0.009549720212817192, + 0.018773278221488, + 0.019415562972426414, + 0.013675280846655369, + 0.016132138669490814, + 0.010675366036593914, + 0.003917333669960499, + 0.007596130017191172, + 0.007752250414341688, + 0.013723076321184635, + 0.004931477829813957, + 0.006750202737748623, + 0.007904069498181343, + 0.0070002758875489235, + 0.044230084866285324, + 0.013278224505484104, + 0.2282160371541977, + 0.023860834538936615, + 0.03964165598154068, + 0.042050618678331375, + 0.1368902623653412, + 0.03401068598031998, + 0.002902831882238388, + 0.004474100191146135, + 0.004280766472220421, + 0.004185082390904427, + 0.005914734676480293, + 0.0047102319076657295, + 0.004500650800764561, + 0.0032076204661279917, + 0.002958135912194848, + 0.0039647831581532955, + 0.0032664495520293713, + 0.004744538106024265, + 0.00456287432461977, + 0.005840914789587259, + 0.0038602519780397415, + 0.004986061714589596, + 0.0032841507345438004, + 0.005412515252828598, + 0.0037446366623044014, + 0.00853189080953598, + 0.004689121153205633, + 0.0025015175342559814, + 0.004927881993353367, + 0.004610680043697357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.027959860861301422 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647525027394295 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647525027394295, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647525027394295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10271475464105606, + 0.14430253207683563, + 0.03924108296632767, + 0.08367437869310379, + 0.06832150369882584, + 0.02670130878686905, + 0.3469208776950836, + 0.04178507626056671, + 0.06191786006093025, + 0.08442061394453049 + ], + "credit_map": [ + 0.009514179080724716, + 0.013539855368435383, + 0.008541258983314037, + 0.017200740054249763, + 0.011872423812747002, + 0.016914157196879387, + 0.013822535052895546, + 0.011309606954455376, + 0.01600637286901474, + 0.021228622645139694, + 0.017737193033099174, + 0.014209868386387825, + 0.018472157418727875, + 0.012730333022773266, + 0.021610213443636894, + 0.022307779639959335, + 0.004129876382648945, + 0.00474681006744504, + 0.00588116142898798, + 0.0043861824087798595, + 0.0038757233414798975, + 0.003992005251348019, + 0.007055183406919241, + 0.0051741451025009155, + 0.010010466910898685, + 0.007744960021227598, + 0.006964145693928003, + 0.006093661766499281, + 0.007422778755426407, + 0.018022479489445686, + 0.011060066521167755, + 0.01635582186281681, + 0.0026508525479584932, + 0.003541711252182722, + 0.006488428916782141, + 0.014881012961268425, + 0.006697504315525293, + 0.016325276345014572, + 0.013343417085707188, + 0.004393300041556358, + 0.0019996787887066603, + 0.0029100365936756134, + 0.003035135567188263, + 0.004193790722638369, + 0.0026033890899270773, + 0.0046745361760258675, + 0.004184086807072163, + 0.003100653411820531, + 0.028456581756472588, + 0.013086151331663132, + 0.13649098575115204, + 0.014947724528610706, + 0.02499181032180786, + 0.029616456478834152, + 0.0788462832570076, + 0.02048487588763237, + 0.0029088875744491816, + 0.005382861010730267, + 0.007331571076065302, + 0.0034763177391141653, + 0.0062134298495948315, + 0.007389506790786982, + 0.004161990713328123, + 0.004920510575175285, + 0.004405265208333731, + 0.008980361744761467, + 0.006220268085598946, + 0.011291048489511013, + 0.008420895785093307, + 0.007262110710144043, + 0.007893274538218975, + 0.00744463549926877, + 0.006498194765299559, + 0.010814250446856022, + 0.009748694486916065, + 0.012401130981743336, + 0.010612482205033302, + 0.009692294523119926, + 0.014341535978019238, + 0.01031203381717205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.004647525027394295 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294245824217796 + ], + "opposite_logprob_tokens": [ + -3.77329421043396 + ], + "logprob_reward": -0.023294245824217796, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294245824217796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06751736253499985, + 0.047896746546030045, + 0.046119775623083115, + 0.024658063426613808, + 0.07299133390188217, + 0.03101436421275139, + 0.6139330267906189, + 0.023212144151329994, + 0.024530963972210884, + 0.048126135021448135 + ], + "credit_map": [ + 0.005262746009975672, + 0.005610662512481213, + 0.006480370182543993, + 0.008353950455784798, + 0.006819782312959433, + 0.013935364782810211, + 0.013434293679893017, + 0.007620200049132109, + 0.004745242651551962, + 0.0075881448574364185, + 0.005609268322587013, + 0.006054682191461325, + 0.008425314910709858, + 0.004391426686197519, + 0.006312699057161808, + 0.004769971128553152, + 0.005809812806546688, + 0.00446602888405323, + 0.005038393195718527, + 0.0035837385803461075, + 0.005214274860918522, + 0.00552694546058774, + 0.010524448938667774, + 0.005956136621534824, + 0.001966806361451745, + 0.0016424348577857018, + 0.0047335452400147915, + 0.002453886903822422, + 0.0023614682722836733, + 0.0042184870690107346, + 0.003085135715082288, + 0.004196303430944681, + 0.002454801695421338, + 0.002469839993864298, + 0.0025278839748352766, + 0.009119241498410702, + 0.0039496393874287605, + 0.025254571810364723, + 0.023071203380823135, + 0.004144159611314535, + 0.002141009783372283, + 0.0029587633907794952, + 0.004225891083478928, + 0.004455889109522104, + 0.0034406050108373165, + 0.004867075476795435, + 0.005377526395022869, + 0.0035476079210639, + 0.04716945067048073, + 0.014081855304539204, + 0.2501627504825592, + 0.02366514317691326, + 0.04161189869046211, + 0.044196922332048416, + 0.15737976133823395, + 0.03566531091928482, + 0.0017206644406542182, + 0.0022433490958064795, + 0.002689468441531062, + 0.0030559750739485025, + 0.003539073746651411, + 0.005117455497384071, + 0.002800809917971492, + 0.0020453515462577343, + 0.0019147112034261227, + 0.004035608842968941, + 0.0021170491818338633, + 0.003752326825633645, + 0.0036733108572661877, + 0.0026595168747007847, + 0.0033686922397464514, + 0.003009749809280038, + 0.004105316940695047, + 0.006743280682712793, + 0.004341665655374527, + 0.005049283616244793, + 0.010934118181467056, + 0.00429284293204546, + 0.0058065312914550304, + 0.006853101775050163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.023294245824217796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537001632153988 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537001632153988, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537001632153988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11354673653841019, + 0.07844777405261993, + 0.09992694109678268, + 0.10190800577402115, + 0.197114035487175, + 0.03944995254278183, + 0.1317458152770996, + 0.0714377835392952, + 0.08762147277593613, + 0.07880145311355591 + ], + "credit_map": [ + 0.010274654254317284, + 0.01231458131223917, + 0.011792995035648346, + 0.014592280611395836, + 0.023902861401438713, + 0.015063774771988392, + 0.0159074105322361, + 0.009698193520307541, + 0.008859345689415932, + 0.011319607496261597, + 0.008873174898326397, + 0.010686381720006466, + 0.011413869448006153, + 0.0077911741100251675, + 0.01103044580668211, + 0.008473780006170273, + 0.009086553007364273, + 0.009795762598514557, + 0.011567948386073112, + 0.010785620659589767, + 0.013175657950341702, + 0.013527733273804188, + 0.01934126578271389, + 0.012646406888961792, + 0.013286028057336807, + 0.011688983999192715, + 0.017415428534150124, + 0.010361278429627419, + 0.007846657186746597, + 0.014119390398263931, + 0.011875668540596962, + 0.015314578078687191, + 0.01733335293829441, + 0.022660577669739723, + 0.017440231516957283, + 0.029049914330244064, + 0.032315585762262344, + 0.023515168577432632, + 0.02797500044107437, + 0.026824206113815308, + 0.0030892628710716963, + 0.005174367688596249, + 0.0038949009031057358, + 0.006742213852703571, + 0.004686776082962751, + 0.0061218393966555595, + 0.005202637519687414, + 0.004537953995168209, + 0.009711155667901039, + 0.020948810502886772, + 0.01710074581205845, + 0.012529849074780941, + 0.01987466774880886, + 0.01974366046488285, + 0.021534444764256477, + 0.010302489623427391, + 0.004226238466799259, + 0.01135826576501131, + 0.009023829363286495, + 0.005937135778367519, + 0.011968810111284256, + 0.013962970115244389, + 0.00752791203558445, + 0.00743262842297554, + 0.007224514614790678, + 0.013551769778132439, + 0.0077072083950042725, + 0.014713143929839134, + 0.010409587062895298, + 0.00926410872489214, + 0.011475828476250172, + 0.013275321573019028, + 0.006206665188074112, + 0.01069251261651516, + 0.011205397546291351, + 0.0111574437469244, + 0.010755069553852081, + 0.0074028694070875645, + 0.007588151376694441, + 0.013793352991342545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.012537001632153988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a dizzying high-angle tracking shot that swoops vertically downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a skeletal biomechanical creature with a Giger-esque fusion of bone and machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature possess a smooth, obsidian, eyeless skull with a faint inner amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow chasm separating vast floating islands made of ribbed, chitinous structures and dark stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual palette desaturated and monochromatic, dominated by gunmetal grays and wet blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038102809339761734 + ], + "opposite_logprob_tokens": [ + -3.288102865219116 + ], + "logprob_reward": -0.038102809339761734, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038102809339761734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07195581495761871, + 0.03266981616616249, + 0.016427641734480858, + 0.02184866927564144, + 0.06365372985601425, + 0.03181739151477814, + 0.5320291519165039, + 0.053882528096437454, + 0.08331789076328278, + 0.09239735454320908 + ], + "credit_map": [ + 0.007032365072518587, + 0.006320379674434662, + 0.006334929261356592, + 0.00976412184536457, + 0.008898238651454449, + 0.012939992360770702, + 0.014136917889118195, + 0.006528870202600956, + 0.0036883060820400715, + 0.003967875149101019, + 0.0038794903084635735, + 0.0034672515466809273, + 0.0043455976992845535, + 0.00322164548560977, + 0.004774675704538822, + 0.005324976053088903, + 0.001868870691396296, + 0.0015791503246873617, + 0.0018300384981557727, + 0.0015778267988935113, + 0.0013991453452035785, + 0.0027261527720838785, + 0.003459581872448325, + 0.0019868756644427776, + 0.003431741613894701, + 0.002310823881998658, + 0.0023489436134696007, + 0.0030196323059499264, + 0.0018391100456938148, + 0.0026721274480223656, + 0.004066347144544125, + 0.0021599424071609974, + 0.0038136884104460478, + 0.004475747235119343, + 0.0038897597696632147, + 0.008311125449836254, + 0.010134926997125149, + 0.012890034355223179, + 0.014068365097045898, + 0.0060700830072164536, + 0.001988861244171858, + 0.0031751052010804415, + 0.004368965048342943, + 0.005822549574077129, + 0.002433327492326498, + 0.005735018756240606, + 0.004561142064630985, + 0.0037324237637221813, + 0.04208926856517792, + 0.01253572478890419, + 0.21918325126171112, + 0.021578127518296242, + 0.03876311704516411, + 0.041150856763124466, + 0.12499063462018967, + 0.03173819184303284, + 0.004340042360126972, + 0.008087536320090294, + 0.006526059936732054, + 0.0059182592667639256, + 0.008537131361663342, + 0.010385342873632908, + 0.005409883800894022, + 0.004678273573517799, + 0.007887310348451138, + 0.009091684594750404, + 0.007999688386917114, + 0.01409216970205307, + 0.009978762827813625, + 0.013469665311276913, + 0.01226526778191328, + 0.008533341810107231, + 0.008236605674028397, + 0.01395372860133648, + 0.007922351360321045, + 0.01279324758797884, + 0.02314841002225876, + 0.00962755549699068, + 0.007337623275816441, + 0.009377828799188137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.038102809339761734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_12/index_tensor([9072, 9072])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_12/index_tensor([9072, 9072])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..093ab344ccc8c06954542043edd9836670616b14 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_12/index_tensor([9072, 9072])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera pans in a flawless 360-degree pan, beginning from a slightly bird's-eye view, slowly sweeping across a futuristic, intelligent nursing bed. Constructed of a sleek white composite material and translucent blue biogel, the bed's edges shimmer with a soft blue glow. A translucent holographic display hovers above the headboard, flickering a steady pulse of heart rate and vital signs. Lying quietly on the bed, a young woman with cascading silver hair spread across the pillow, reflecting the cool light of the surroundings. She wears a form-fitting white, high-tech nanofiber nightgown with a subtle pearlescent sheen. The camera moves steadily and gracefully, focusing precisely on her delicate features. Her fair skin appears crystal clear, cast against the dreamy purple and cyan light of the nebula outside the window and the bed's own blue glow. At that moment, her long, thick eyelashes begin to flutter, the first subtle sign of awakening from a deep sleep. Immediately, her brow furrowed slightly, her breathing became deep and clear, and her slender fingers unconsciously curled. As the camera panned to her front, her eyelids finally slowly opened, revealing a pair of clear, yet slightly lost, light-blue eyes, trying to focus on the minimalist, futuristic medical pod. The entire scene exuded the quality of a top-tier CG work, with every detail achieving photorealism and every play of light and shadow meticulously nuanced, creating a tranquil, sci-fi, and beautiful atmosphere of awakening.", + "seed": 31473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_12/index_9072_seed_31473_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5419996440410614, + "vlm_logprob": -0.5419996440410614, + "vlm_yes_no_margin": 1.25, + "vlm_token_credit": -0.5419996440410614, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.129659652709961 + ], + "opposite_logprob_tokens": [ + -0.12965965270996094 + ], + "logprob_reward": -2.129659652709961, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.129659652709961, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07010025531053543, + 0.0902462750673294, + 0.03132088854908943, + 0.0503062829375267, + 0.04719104617834091, + 0.12757450342178345, + 0.42337509989738464, + 0.0932542085647583, + 0.03117271140217781, + 0.03545871376991272 + ], + "credit_map": [ + 0.008055614307522774, + 0.008391289040446281, + 0.009747350588440895, + 0.00937686674296856, + 0.009424756281077862, + 0.006954349111765623, + 0.010783940553665161, + 0.007366084959357977, + 0.008425776846706867, + 0.018335439264774323, + 0.009851939976215363, + 0.011558810248970985, + 0.013298806734383106, + 0.009160938672721386, + 0.010935242287814617, + 0.008679321967065334, + 0.002871815348044038, + 0.00361320236697793, + 0.004512831103056669, + 0.003939325455576181, + 0.003794086864218116, + 0.0027293001767247915, + 0.004880690015852451, + 0.004979636054486036, + 0.003162152599543333, + 0.004011515062302351, + 0.009165765717625618, + 0.008244774304330349, + 0.004466522019356489, + 0.002370234113186598, + 0.008964871056377888, + 0.009920451790094376, + 0.005033676978200674, + 0.00486877653747797, + 0.006395693402737379, + 0.006757369730621576, + 0.0031449648085981607, + 0.004736095666885376, + 0.009451383724808693, + 0.0068030874244868755, + 0.008680255152285099, + 0.018174665048718452, + 0.018476400524377823, + 0.024944303557276726, + 0.012702658772468567, + 0.010817802511155605, + 0.017391221597790718, + 0.016387205570936203, + 0.02845795825123787, + 0.04477803409099579, + 0.12026515603065491, + 0.025785524398088455, + 0.08037624508142471, + 0.02136748656630516, + 0.047394897788763046, + 0.05494978278875351, + 0.008435639552772045, + 0.011104496195912361, + 0.011545766144990921, + 0.015176395885646343, + 0.012735815718770027, + 0.010794767178595066, + 0.012542910873889923, + 0.010918418876826763, + 0.0029955757781863213, + 0.0026697597932070494, + 0.004800164606422186, + 0.0043002041056752205, + 0.0033186536747962236, + 0.0024152961559593678, + 0.005741708446294069, + 0.00493134930729866, + 0.003295969683676958, + 0.004583293106406927, + 0.00487595796585083, + 0.004339797887951136, + 0.0064011793583631516, + 0.002463037148118019, + 0.006401688791811466, + 0.0030977874994277954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -2.129659652709961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04314675182104111 + ], + "opposite_logprob_tokens": [ + -3.168146848678589 + ], + "logprob_reward": -0.04314675182104111, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04314675182104111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07238224893808365, + 0.07690637558698654, + 0.04790347069501877, + 0.052131716161966324, + 0.0571724995970726, + 0.11157642304897308, + 0.41091957688331604, + 0.09132512658834457, + 0.027029378339648247, + 0.052653200924396515 + ], + "credit_map": [ + 0.008008605800569057, + 0.007277995347976685, + 0.011183510534465313, + 0.00981337483972311, + 0.011403283104300499, + 0.006138866767287254, + 0.01195891946554184, + 0.006597692146897316, + 0.006916258484125137, + 0.011185179464519024, + 0.010621186345815659, + 0.011735727079212666, + 0.009358961135149002, + 0.008715344592928886, + 0.010039654560387135, + 0.00833407137542963, + 0.0035035014152526855, + 0.0067915720865130424, + 0.00674452306702733, + 0.007268383167684078, + 0.005222196690738201, + 0.004284134600311518, + 0.007915389724075794, + 0.006173769012093544, + 0.0030056473333388567, + 0.003204831387847662, + 0.012497453950345516, + 0.006819568108767271, + 0.004352601710706949, + 0.0029695520643144846, + 0.006518444046378136, + 0.012763618491590023, + 0.004302239045500755, + 0.0030814383644610643, + 0.018335482105612755, + 0.004719021264463663, + 0.001938149449415505, + 0.0036030150949954987, + 0.0053373342379927635, + 0.01585582084953785, + 0.007985366508364677, + 0.014688822440803051, + 0.015479997731745243, + 0.02075422927737236, + 0.012746377848088741, + 0.011063872836530209, + 0.015533207915723324, + 0.01332454476505518, + 0.03230900689959526, + 0.03864911571145058, + 0.12762610614299774, + 0.029239807277917862, + 0.057052064687013626, + 0.021932724863290787, + 0.04755115509033203, + 0.056559596210718155, + 0.007975000888109207, + 0.010968673042953014, + 0.012555321678519249, + 0.013595229014754295, + 0.011996609158813953, + 0.010387513786554337, + 0.01241633016616106, + 0.011430454440414906, + 0.002146233804523945, + 0.002785069402307272, + 0.004169357009232044, + 0.003034447319805622, + 0.0033659639302641153, + 0.002569386502727866, + 0.004921560175716877, + 0.00403736112639308, + 0.0068182554095983505, + 0.004844990558922291, + 0.007374741602689028, + 0.008703199215233326, + 0.006183126010000706, + 0.004183514043688774, + 0.008910141885280609, + 0.005635233595967293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.04314675182104111 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28140789270401 + ], + "opposite_logprob_tokens": [ + -1.4064078330993652 + ], + "logprob_reward": -0.28140789270401, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28140789270401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10337761044502258, + 0.14074642956256866, + 0.08459760993719101, + 0.03818577155470848, + 0.16736207902431488, + 0.09096360206604004, + 0.1437360942363739, + 0.16834275424480438, + 0.033949900418519974, + 0.02873818576335907 + ], + "credit_map": [ + 0.01673337258398533, + 0.010617717169225216, + 0.015807760879397392, + 0.009442629292607307, + 0.02432442642748356, + 0.007693008054047823, + 0.00956244021654129, + 0.00919624324887991, + 0.012183042243123055, + 0.016934752464294434, + 0.018997812643647194, + 0.025185832753777504, + 0.014913706108927727, + 0.013327706605196, + 0.023389320820569992, + 0.015814226120710373, + 0.004915677476674318, + 0.006719875615090132, + 0.01933598890900612, + 0.007686637807637453, + 0.008656967431306839, + 0.00812967587262392, + 0.006979695055633783, + 0.022173084318637848, + 0.0023861275985836983, + 0.005081548821181059, + 0.00609114533290267, + 0.006150899920612574, + 0.00431419862434268, + 0.0024668159894645214, + 0.005664111580699682, + 0.006030918098986149, + 0.011514673009514809, + 0.006424040999263525, + 0.0599699541926384, + 0.011158113367855549, + 0.004457526374608278, + 0.009482254274189472, + 0.019388655200600624, + 0.04496683552861214, + 0.007281168829649687, + 0.007550714537501335, + 0.013445163145661354, + 0.007971699349582195, + 0.008687056601047516, + 0.007092057261615992, + 0.025517676025629044, + 0.013418051414191723, + 0.016627440229058266, + 0.01673157885670662, + 0.03868219256401062, + 0.01198512502014637, + 0.01608787104487419, + 0.010413349606096745, + 0.012481382116675377, + 0.02072713151574135, + 0.01830332539975643, + 0.021979626268148422, + 0.02157467044889927, + 0.028656359761953354, + 0.01807735301554203, + 0.015019497834146023, + 0.028026960790157318, + 0.01670493558049202, + 0.0037102741189301014, + 0.002641879254952073, + 0.0053098429925739765, + 0.0052500455640256405, + 0.004451903980225325, + 0.0028857102151960135, + 0.004825301468372345, + 0.004874938167631626, + 0.0036030765622854233, + 0.0034036985598504543, + 0.004813721403479576, + 0.0032404616940766573, + 0.003188120201230049, + 0.0028590017464011908, + 0.004228775389492512, + 0.0034013260155916214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.28140789270401 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294852942228317 + ], + "opposite_logprob_tokens": [ + -2.237948417663574 + ], + "logprob_reward": -0.11294852942228317, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294852942228317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04867835342884064, + 0.07412790507078171, + 0.03675552085042, + 0.013136706314980984, + 0.06913401931524277, + 0.03567705303430557, + 0.6027001142501831, + 0.04767496511340141, + 0.024361785501241684, + 0.047753557562828064 + ], + "credit_map": [ + 0.008847488090395927, + 0.008019030094146729, + 0.007496845442801714, + 0.004467488266527653, + 0.0063484907150268555, + 0.0031590687576681376, + 0.005869215354323387, + 0.00447072321549058, + 0.005393745843321085, + 0.014952471479773521, + 0.009631040506064892, + 0.0073135243728756905, + 0.012706234119832516, + 0.0078052012249827385, + 0.010787460021674633, + 0.005538229364901781, + 0.003018394811078906, + 0.004686084110289812, + 0.007713399827480316, + 0.0037294444628059864, + 0.003480032319203019, + 0.002113595837727189, + 0.00317985936999321, + 0.008834709413349628, + 0.000940052152145654, + 0.001796829397790134, + 0.0019386322237551212, + 0.0015781944384798408, + 0.001668061362579465, + 0.0012201049830764532, + 0.0016163018299266696, + 0.0023785291705280542, + 0.004926638677716255, + 0.002868582960218191, + 0.024457119405269623, + 0.0052020614966750145, + 0.0017098139505833387, + 0.003879702650010586, + 0.008435712195932865, + 0.017654389142990112, + 0.0028743899893015623, + 0.004644215572625399, + 0.004786263685673475, + 0.006320157088339329, + 0.0034888330847024918, + 0.00409628776833415, + 0.005357449874281883, + 0.004109458066523075, + 0.037359755486249924, + 0.12024009972810745, + 0.08811221271753311, + 0.050464604049921036, + 0.12977993488311768, + 0.04635171219706535, + 0.07902585715055466, + 0.051365926861763, + 0.005043187644332647, + 0.005477705039083958, + 0.00623187143355608, + 0.006198562681674957, + 0.005167852155864239, + 0.007512458600103855, + 0.007024463266134262, + 0.00501886336132884, + 0.0018058109562844038, + 0.0028059426695108414, + 0.003843079786747694, + 0.00293172849342227, + 0.003337748581543565, + 0.00228682579472661, + 0.004000944085419178, + 0.0033497039694339037, + 0.005491209682077169, + 0.005216225516051054, + 0.006491174455732107, + 0.008820608258247375, + 0.005747269839048386, + 0.003471137722954154, + 0.008102549239993095, + 0.0044133844785392284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.11294852942228317 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283539354801178 + ], + "opposite_logprob_tokens": [ + -2.0178353786468506 + ], + "logprob_reward": -0.14283539354801178, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283539354801178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06377582252025604, + 0.16867561638355255, + 0.049184031784534454, + 0.019842732697725296, + 0.03258203715085983, + 0.0511971116065979, + 0.21111300587654114, + 0.11434733867645264, + 0.1932331919670105, + 0.09604915231466293 + ], + "credit_map": [ + 0.007292255759239197, + 0.008769012987613678, + 0.008671187795698643, + 0.008098648861050606, + 0.01197589747607708, + 0.004898884799331427, + 0.007667248602956533, + 0.006402686238288879, + 0.014292720705270767, + 0.03641192615032196, + 0.01958298683166504, + 0.01580847054719925, + 0.028738589957356453, + 0.017536023631691933, + 0.021602455526590347, + 0.014702440239489079, + 0.004524521064013243, + 0.00614179065451026, + 0.007661065552383661, + 0.0074721937999129295, + 0.004836574662476778, + 0.00502116559073329, + 0.006293239537626505, + 0.007233481854200363, + 0.0014256453141570091, + 0.0015568698290735483, + 0.0033644947689026594, + 0.0038120511453598738, + 0.002053122268989682, + 0.001169542782008648, + 0.0029508883599191904, + 0.003510119393467903, + 0.0026272167451679707, + 0.0018303857650607824, + 0.009098375216126442, + 0.0027531038504093885, + 0.0015650481218472123, + 0.002867094473913312, + 0.0033152513206005096, + 0.008525560609996319, + 0.004094117321074009, + 0.007458856329321861, + 0.007389733102172613, + 0.00979885645210743, + 0.0052199820056557655, + 0.003099057823419571, + 0.007472726982086897, + 0.006663782987743616, + 0.018510233610868454, + 0.022701893001794815, + 0.0592719204723835, + 0.016441617161035538, + 0.024409379810094833, + 0.015662571415305138, + 0.025687236338853836, + 0.02842813916504383, + 0.010729953646659851, + 0.0094746183604002, + 0.01728609949350357, + 0.018018152564764023, + 0.015768982470035553, + 0.011289265938103199, + 0.02018195018172264, + 0.011598308570683002, + 0.01501481607556343, + 0.022257300093770027, + 0.026334183290600777, + 0.028781002387404442, + 0.021051449701189995, + 0.014485539868474007, + 0.031616877764463425, + 0.033692024648189545, + 0.0095698032528162, + 0.011383907869458199, + 0.013257070444524288, + 0.015575600787997246, + 0.01361369900405407, + 0.006421314552426338, + 0.017060192301869392, + 0.009167566895484924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.14283539354801178 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera pans in a flawless 360-degree pan, beginning from a slightly bird's-eye view, slowly sweeping across a futuristic, intelligent nursing bed. Constructed of a sleek white composite material and translucent blue biogel, the bed's edges shimmer with a soft blue glow. A translucent holographic display hovers above the headboard, flickering a steady pulse of heart rate and vital signs. Lying quietly on the bed, a young woman with cascading silver hair spread across the pillow, reflecting the cool light of the surroundings. She wears a form-fitting white, high-tech nanofiber nightgown with a subtle pearlescent sheen. The camera moves steadily and gracefully, focusing precisely on her delicate features. Her fair skin appears crystal clear, cast against the dreamy purple and cyan light of the nebula outside the window and the bed's own blue glow. At that moment, her long, thick eyelashes begin to flutter, the first subtle sign of awakening from a deep sleep. Immediately, her brow furrowed slightly, her breathing became deep and clear, and her slender fingers unconsciously curled. As the camera panned to her front, her eyelids finally slowly opened, revealing a pair of clear, yet slightly lost, light-blue eyes, trying to focus on the minimalist, futuristic medical pod. The entire scene exuded the quality of a top-tier CG work, with every detail achieving photorealism and every play of light and shadow meticulously nuanced, creating a tranquil, sci-fi, and beautiful atmosphere of awakening.", + "seed": 31474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_12/index_9072_seed_31474_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5548339627683163, + "vlm_logprob": -0.5548339627683163, + "vlm_yes_no_margin": 1.325, + "vlm_token_credit": -0.5548339627683163, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.241231918334961 + ], + "opposite_logprob_tokens": [ + -0.11623190343379974 + ], + "logprob_reward": -2.241231918334961, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.241231918334961, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.42068618535995483, + 0.12329184263944626, + 0.06498105078935623, + 0.03137139976024628, + 0.07601013034582138, + 0.08490002900362015, + 0.04200286790728569, + 0.031184537336230278, + 0.09676126390695572, + 0.02881070412695408 + ], + "credit_map": [ + 0.03446947783231735, + 0.08393411338329315, + 0.03572689741849899, + 0.07313232868909836, + 0.040132660418748856, + 0.029837748035788536, + 0.06495340913534164, + 0.0584995299577713, + 0.013165892101824284, + 0.019608469679951668, + 0.011401087045669556, + 0.017456745728850365, + 0.015821684151887894, + 0.016458921134471893, + 0.015588830225169659, + 0.013790212571620941, + 0.0036999147851020098, + 0.00848693773150444, + 0.00792710017412901, + 0.008936450816690922, + 0.011487100273370743, + 0.005860618315637112, + 0.010771526023745537, + 0.007811404764652252, + 0.002189275808632374, + 0.0037897906731814146, + 0.00538240559399128, + 0.0042991130612790585, + 0.004227754659950733, + 0.0029928316362202168, + 0.00390793988481164, + 0.004582286346703768, + 0.005905176512897015, + 0.01245899684727192, + 0.01058202888816595, + 0.007490050978958607, + 0.010664008557796478, + 0.010146231390535831, + 0.008316274732351303, + 0.01044735312461853, + 0.011675572022795677, + 0.020670143887400627, + 0.010425735265016556, + 0.008768324740231037, + 0.007097015157341957, + 0.008490566164255142, + 0.008259143680334091, + 0.009513531811535358, + 0.003611975582316518, + 0.004343781154602766, + 0.005570746958255768, + 0.007543828804045916, + 0.0061128949746489525, + 0.0050369794480502605, + 0.004602864384651184, + 0.005179797764867544, + 0.003997798077762127, + 0.0028994812164455652, + 0.006850556470453739, + 0.0049006035551428795, + 0.003436581464484334, + 0.00353768072091043, + 0.002584600355476141, + 0.002977235009893775, + 0.008687135763466358, + 0.010353033430874348, + 0.015605677850544453, + 0.015558790415525436, + 0.01400485634803772, + 0.012730755843222141, + 0.008365641348063946, + 0.011455370113253593, + 0.003025021171197295, + 0.003604045370593667, + 0.005338381044566631, + 0.003367246827110648, + 0.004767226055264473, + 0.003441324457526207, + 0.002754057524725795, + 0.0025134009774774313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -2.241231918334961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038204703480005264 + ], + "opposite_logprob_tokens": [ + -3.2882046699523926 + ], + "logprob_reward": -0.038204703480005264, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038204703480005264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3295455873012543, + 0.08911984413862228, + 0.08928558975458145, + 0.11657067388296127, + 0.09713421761989594, + 0.045059118419885635, + 0.03610707074403763, + 0.07937543094158173, + 0.0897279903292656, + 0.028074517846107483 + ], + "credit_map": [ + 0.027254432439804077, + 0.057296063750982285, + 0.03798067942261696, + 0.06033677980303764, + 0.029388664290308952, + 0.025543345138430595, + 0.04607173055410385, + 0.045673906803131104, + 0.010294354520738125, + 0.011879868805408478, + 0.009029839187860489, + 0.00939530972391367, + 0.01049878355115652, + 0.014672624878585339, + 0.011251553893089294, + 0.012097502127289772, + 0.007416028529405594, + 0.011885106563568115, + 0.008219045586884022, + 0.01356617733836174, + 0.017197921872138977, + 0.008851520717144012, + 0.011305388063192368, + 0.01084439642727375, + 0.00635668495669961, + 0.013616537675261497, + 0.015454701147973537, + 0.016951868310570717, + 0.02398397959768772, + 0.00939227920025587, + 0.014706574380397797, + 0.016108045354485512, + 0.006880894768983126, + 0.01345724519342184, + 0.013695829547941685, + 0.00824524275958538, + 0.016320878639817238, + 0.012583860196173191, + 0.011561956256628036, + 0.014388316310942173, + 0.004638851620256901, + 0.005493846721947193, + 0.005626382771879435, + 0.007361875381320715, + 0.006024225614964962, + 0.0046153259463608265, + 0.006649537477642298, + 0.004649071954190731, + 0.004201757721602917, + 0.005161169916391373, + 0.004416454117745161, + 0.005420733708888292, + 0.004985623527318239, + 0.004458135925233364, + 0.0043237218633294106, + 0.003139474429190159, + 0.010116606950759888, + 0.007436513900756836, + 0.01593133807182312, + 0.01440287847071886, + 0.008580883964896202, + 0.008132318034768105, + 0.007377475034445524, + 0.007397417910397053, + 0.008066822774708271, + 0.010007105767726898, + 0.014076673425734043, + 0.015323512256145477, + 0.0118635855615139, + 0.011945487931370735, + 0.007853856310248375, + 0.010590946301817894, + 0.0028038073796778917, + 0.003264473518356681, + 0.004038285464048386, + 0.004365363623946905, + 0.004450301174074411, + 0.0030773929320275784, + 0.0032666856423020363, + 0.002808206481859088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.038204703480005264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28142058849334717 + ], + "opposite_logprob_tokens": [ + -1.4064205884933472 + ], + "logprob_reward": -0.28142058849334717, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28142058849334717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2904066741466522, + 0.2198457419872284, + 0.06773649901151657, + 0.12990093231201172, + 0.06312602758407593, + 0.07874593883752823, + 0.04634794220328331, + 0.041126709431409836, + 0.030510129407048225, + 0.03225342184305191 + ], + "credit_map": [ + 0.02879323624074459, + 0.03634858876466751, + 0.028344817459583282, + 0.036649562418460846, + 0.033734239637851715, + 0.03939656913280487, + 0.04491877928376198, + 0.04222087189555168, + 0.021274616941809654, + 0.028514014557003975, + 0.029208753257989883, + 0.027578290551900864, + 0.025126295164227486, + 0.02895444631576538, + 0.030459284782409668, + 0.028730032965540886, + 0.004269872792065144, + 0.005990239325910807, + 0.011632466688752174, + 0.010575114749372005, + 0.010996493510901928, + 0.004691899288445711, + 0.011145630851387978, + 0.008434776216745377, + 0.00617056293413043, + 0.01869957521557808, + 0.01734711043536663, + 0.016089068725705147, + 0.02845844440162182, + 0.012541706673800945, + 0.013908497989177704, + 0.016685951501131058, + 0.004783353768289089, + 0.0060849497094750404, + 0.00902621727436781, + 0.006633636076003313, + 0.009746863506734371, + 0.007788400165736675, + 0.008716713637113571, + 0.010345886461436749, + 0.006913116667419672, + 0.008183556608855724, + 0.010013576596975327, + 0.00966996792703867, + 0.00861604418605566, + 0.011970479041337967, + 0.012101525440812111, + 0.011277664452791214, + 0.005383595358580351, + 0.008123107254505157, + 0.005679944064468145, + 0.007074128370732069, + 0.004887969233095646, + 0.0053293295204639435, + 0.004776851274073124, + 0.005093011539429426, + 0.004617663566023111, + 0.004218855872750282, + 0.006629323121160269, + 0.008828752674162388, + 0.006134215276688337, + 0.0038605022709816694, + 0.0032498417422175407, + 0.003587552346289158, + 0.002797864843159914, + 0.00365928141400218, + 0.005132841411978006, + 0.005434330552816391, + 0.004962762352079153, + 0.002575341612100601, + 0.0028511821292340755, + 0.003096523229032755, + 0.002931958297267556, + 0.0042727030813694, + 0.0036175160203129053, + 0.005281028337776661, + 0.006089909467846155, + 0.003139068838208914, + 0.003176963422447443, + 0.0037442720495164394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.28142058849334717 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034256428480148 + ], + "opposite_logprob_tokens": [ + -2.3503425121307373 + ], + "logprob_reward": -0.10034256428480148, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034256428480148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1631067842245102, + 0.04098113626241684, + 0.10855473577976227, + 0.2729122042655945, + 0.08107327669858932, + 0.07077740877866745, + 0.0491926483809948, + 0.10127585381269455, + 0.041900817304849625, + 0.07022514194250107 + ], + "credit_map": [ + 0.017651397734880447, + 0.021886970847845078, + 0.016532042995095253, + 0.02447045035660267, + 0.01685807853937149, + 0.022516325116157532, + 0.017603224143385887, + 0.02558829076588154, + 0.001952385762706399, + 0.008004404604434967, + 0.003912755753844976, + 0.005726248491555452, + 0.0060514830984175205, + 0.00319036771543324, + 0.006859140936285257, + 0.005284343380481005, + 0.007510039955377579, + 0.011952037923038006, + 0.009116919711232185, + 0.016055885702371597, + 0.023812366649508476, + 0.011493561789393425, + 0.016741108149290085, + 0.011872815899550915, + 0.015225255861878395, + 0.028556769713759422, + 0.031877435743808746, + 0.037067919969558716, + 0.06362735480070114, + 0.021614432334899902, + 0.03391832113265991, + 0.04102471098303795, + 0.0048210653476417065, + 0.005646895617246628, + 0.01090855523943901, + 0.012375499121844769, + 0.015131822787225246, + 0.008390022441744804, + 0.013041543774306774, + 0.010757865384221077, + 0.009215479716658592, + 0.01275113970041275, + 0.008712389506399632, + 0.009079345501959324, + 0.006469849031418562, + 0.009408530779182911, + 0.007702296134084463, + 0.00743837421759963, + 0.005190894939005375, + 0.005274404305964708, + 0.006191093008965254, + 0.008559231646358967, + 0.0070673325099051, + 0.0050158123485744, + 0.006031201221048832, + 0.005862671881914139, + 0.009227883070707321, + 0.013734557665884495, + 0.014613443054258823, + 0.015545165166258812, + 0.021755682304501534, + 0.009702702984213829, + 0.007736931089311838, + 0.008959484286606312, + 0.004174449481070042, + 0.005899784155189991, + 0.004243347328156233, + 0.006881904788315296, + 0.0072838072665035725, + 0.005222999025136232, + 0.0037846018094569445, + 0.004409919027239084, + 0.006561581045389175, + 0.009739525616168976, + 0.008921723812818527, + 0.012083679437637329, + 0.012497000396251678, + 0.005391446407884359, + 0.007745111361145973, + 0.007285064551979303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.10034256428480148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297003924846649 + ], + "opposite_logprob_tokens": [ + -2.2379701137542725 + ], + "logprob_reward": -0.11297003924846649, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297003924846649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34792810678482056, + 0.10844335705041885, + 0.06224658340215683, + 0.04058654606342316, + 0.1779537796974182, + 0.04006378725171089, + 0.031512148678302765, + 0.055359791964292526, + 0.07969476282596588, + 0.05621109530329704 + ], + "credit_map": [ + 0.0299619659781456, + 0.06853163242340088, + 0.030154788866639137, + 0.0626496970653534, + 0.027437595650553703, + 0.029577817767858505, + 0.044990066438913345, + 0.05462455749511719, + 0.006995719857513905, + 0.014322703704237938, + 0.00917988270521164, + 0.019980644807219505, + 0.01405988447368145, + 0.01197219267487526, + 0.018148891627788544, + 0.013783439993858337, + 0.0036792412865906954, + 0.008145608007907867, + 0.008092770352959633, + 0.00937079917639494, + 0.011599459685385227, + 0.004890662617981434, + 0.009629452601075172, + 0.006838587112724781, + 0.0025034972932189703, + 0.0035632227081805468, + 0.006633815821260214, + 0.00633614556863904, + 0.006662451662123203, + 0.0035212025977671146, + 0.006010124925523996, + 0.005356083624064922, + 0.010489908047020435, + 0.01272781565785408, + 0.026118885725736618, + 0.022986071184277534, + 0.032941702753305435, + 0.01960396021604538, + 0.025178976356983185, + 0.02790646441280842, + 0.004471592139452696, + 0.006618790328502655, + 0.004659560509026051, + 0.00502987252548337, + 0.005061833653599024, + 0.005248064175248146, + 0.004574485123157501, + 0.004399590194225311, + 0.0033642551861703396, + 0.00442799786105752, + 0.003669183934107423, + 0.004589477553963661, + 0.004662008956074715, + 0.003318034578114748, + 0.004129818174988031, + 0.003351373365148902, + 0.00691266218200326, + 0.0070731146261096, + 0.009682227857410908, + 0.0074992734007537365, + 0.010470905341207981, + 0.005147597752511501, + 0.004021511413156986, + 0.004552497062832117, + 0.006693728268146515, + 0.015071545727550983, + 0.007949071936309338, + 0.013179748319089413, + 0.014258951880037785, + 0.006906834430992603, + 0.007481100037693977, + 0.008153778500854969, + 0.005047737620770931, + 0.007234232034534216, + 0.008356628939509392, + 0.006500557530671358, + 0.007788545917719603, + 0.00788457877933979, + 0.006443138234317303, + 0.006955673918128014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.11297003924846649 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera pans in a flawless 360-degree pan, beginning from a slightly bird's-eye view, slowly sweeping across a futuristic, intelligent nursing bed. Constructed of a sleek white composite material and translucent blue biogel, the bed's edges shimmer with a soft blue glow. A translucent holographic display hovers above the headboard, flickering a steady pulse of heart rate and vital signs. Lying quietly on the bed, a young woman with cascading silver hair spread across the pillow, reflecting the cool light of the surroundings. She wears a form-fitting white, high-tech nanofiber nightgown with a subtle pearlescent sheen. The camera moves steadily and gracefully, focusing precisely on her delicate features. Her fair skin appears crystal clear, cast against the dreamy purple and cyan light of the nebula outside the window and the bed's own blue glow. At that moment, her long, thick eyelashes begin to flutter, the first subtle sign of awakening from a deep sleep. Immediately, her brow furrowed slightly, her breathing became deep and clear, and her slender fingers unconsciously curled. As the camera panned to her front, her eyelids finally slowly opened, revealing a pair of clear, yet slightly lost, light-blue eyes, trying to focus on the minimalist, futuristic medical pod. The entire scene exuded the quality of a top-tier CG work, with every detail achieving photorealism and every play of light and shadow meticulously nuanced, creating a tranquil, sci-fi, and beautiful atmosphere of awakening.", + "seed": 31475, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_12/index_9072_seed_31475_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3881965495646, + "vlm_logprob": -0.3881965495646, + "vlm_yes_no_margin": 1.5750000119209289, + "vlm_token_credit": -0.3881965495646, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2283680438995361 + ], + "opposite_logprob_tokens": [ + -0.3533681035041809 + ], + "logprob_reward": -1.2283680438995361, + "yes_no_margin_reward": -0.8749999403953552, + "token_credit_reward": -1.2283680438995361, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10701195895671844, + 0.029702985659241676, + 0.5516947507858276, + 0.02926931530237198, + 0.03912964463233948, + 0.013059823773801327, + 0.030668538063764572, + 0.01914850063621998, + 0.16822613775730133, + 0.012088402174413204 + ], + "credit_map": [ + 0.009561906568706036, + 0.014924174174666405, + 0.014074771665036678, + 0.014835369773209095, + 0.006254070438444614, + 0.007845801301300526, + 0.01908431202173233, + 0.020431548357009888, + 0.002673172391951084, + 0.003979632630944252, + 0.004727309104055166, + 0.0035415564198046923, + 0.0029503554105758667, + 0.003610241925343871, + 0.003314914647489786, + 0.004905798472464085, + 0.03157391771674156, + 0.0261552594602108, + 0.1344555914402008, + 0.019083986058831215, + 0.04645757004618645, + 0.010846395045518875, + 0.2402621954679489, + 0.04285981506109238, + 0.004362008068710566, + 0.003397509455680847, + 0.004257969092577696, + 0.004055693745613098, + 0.0033889662008732557, + 0.0030042335856705904, + 0.003910106141120195, + 0.002892828080803156, + 0.0032321475446224213, + 0.004912076983600855, + 0.006670307368040085, + 0.005609511397778988, + 0.004905617330223322, + 0.004720177035778761, + 0.005789261311292648, + 0.0032905428670346737, + 0.0013819707091897726, + 0.0013479396002367139, + 0.0013325701002031565, + 0.001801067846827209, + 0.0017626023618504405, + 0.0012240300420671701, + 0.0026242444291710854, + 0.0015853974036872387, + 0.002374112606048584, + 0.005726499482989311, + 0.0032863363157957792, + 0.004665769170969725, + 0.0031926645897328854, + 0.002352621639147401, + 0.00475104246288538, + 0.004319491796195507, + 0.0014167344197630882, + 0.0022359241265803576, + 0.0026073320768773556, + 0.002374306321144104, + 0.0027822472620755434, + 0.00215469510294497, + 0.002633271273225546, + 0.002943988423794508, + 0.00535643519833684, + 0.0034050047397613525, + 0.010053719393908978, + 0.003204574342817068, + 0.013494071550667286, + 0.07151146978139877, + 0.005722925066947937, + 0.05547791346907616, + 0.0012597524328157306, + 0.00184080027975142, + 0.001578761963173747, + 0.0013679413823410869, + 0.0012470632791519165, + 0.0014417613856494427, + 0.0016046848613768816, + 0.001747635891661048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -1.2283680438995361 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038142282515764236 + ], + "opposite_logprob_tokens": [ + -3.288142204284668 + ], + "logprob_reward": -0.038142282515764236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038142282515764236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11440762132406235, + 0.027315929532051086, + 0.30074843764305115, + 0.0381452850997448, + 0.020593563094735146, + 0.032163724303245544, + 0.05248361825942993, + 0.060893990099430084, + 0.3279711604118347, + 0.025276683270931244 + ], + "credit_map": [ + 0.012693255208432674, + 0.012461723759770393, + 0.018829284235835075, + 0.013148196041584015, + 0.006866250187158585, + 0.007368617691099644, + 0.015608939342200756, + 0.02743135206401348, + 0.0023786183446645737, + 0.0033305177930742502, + 0.0051503353752195835, + 0.004024913534522057, + 0.0022592174354940653, + 0.002696635667234659, + 0.0030852698255330324, + 0.004390424117445946, + 0.017103711143136024, + 0.014892048202455044, + 0.06577702611684799, + 0.009953315369784832, + 0.030919639393687248, + 0.0056343087926507, + 0.13219860196113586, + 0.024269796907901764, + 0.004473647102713585, + 0.0050510927103459835, + 0.005237342324107885, + 0.004957069177180529, + 0.004022811073809862, + 0.0032526999711990356, + 0.004813109990209341, + 0.006337514612823725, + 0.00141974410507828, + 0.002343008993193507, + 0.002721653087064624, + 0.0034788872580975294, + 0.0025799348950386047, + 0.002956422045826912, + 0.0028267097659409046, + 0.002267202129587531, + 0.002239946974441409, + 0.003020239295437932, + 0.0033738629426807165, + 0.0027003674767911434, + 0.008881175890564919, + 0.004124094266444445, + 0.0034866640344262123, + 0.0043373736552894115, + 0.003989147022366524, + 0.007306681480258703, + 0.006917942781001329, + 0.008140450343489647, + 0.00558529794216156, + 0.00533660314977169, + 0.007876083254814148, + 0.007331413216888905, + 0.006469170097261667, + 0.005335599649697542, + 0.008614573627710342, + 0.0055645243264734745, + 0.009565495885908604, + 0.007245715241879225, + 0.005536150652915239, + 0.012562762945890427, + 0.010293682105839252, + 0.005831875838339329, + 0.018046947196125984, + 0.005011772736907005, + 0.023809485137462616, + 0.14240144193172455, + 0.011160416528582573, + 0.11141552776098251, + 0.002871018834412098, + 0.0031282855197787285, + 0.0034758865367621183, + 0.0025865333154797554, + 0.002091688569635153, + 0.003916480578482151, + 0.002937911543995142, + 0.004268877673894167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.038142282515764236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233895778656006 + ], + "opposite_logprob_tokens": [ + -0.8983895778656006 + ], + "logprob_reward": -0.5233895778656006, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233895778656006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07963857054710388, + 0.02333429828286171, + 0.4542055130004883, + 0.031162234023213387, + 0.029228080064058304, + 0.10151279717683792, + 0.0422317273914814, + 0.020488034933805466, + 0.17459966242313385, + 0.043599024415016174 + ], + "credit_map": [ + 0.00809420645236969, + 0.014902204275131226, + 0.009293093346059322, + 0.008472097106277943, + 0.0062142955139279366, + 0.0033760303631424904, + 0.015496626496315002, + 0.013790023513138294, + 0.0015848020557314157, + 0.003049183404073119, + 0.003089052392169833, + 0.004721358884125948, + 0.0024637915194034576, + 0.0023989391047507524, + 0.0031219213269650936, + 0.0029052509926259518, + 0.025250405073165894, + 0.022541705518960953, + 0.09546744078397751, + 0.015453913249075413, + 0.05647832527756691, + 0.008439050987362862, + 0.19677434861660004, + 0.03380037099123001, + 0.003223864361643791, + 0.005042240489274263, + 0.0028058600146323442, + 0.003102088114246726, + 0.006106121465563774, + 0.0027811371255666018, + 0.004347604233771563, + 0.0037533193826675415, + 0.0023501794785261154, + 0.004213418811559677, + 0.008042849600315094, + 0.00288231298327446, + 0.002883860608562827, + 0.002581960055977106, + 0.0030112017411738634, + 0.0032622998114675283, + 0.009127273224294186, + 0.007124703843146563, + 0.00833259616047144, + 0.00678345188498497, + 0.034483879804611206, + 0.007896412163972855, + 0.008669215254485607, + 0.019095277413725853, + 0.0038802847266197205, + 0.006032670382410288, + 0.006481778342276812, + 0.006415640469640493, + 0.004572817124426365, + 0.004519667010754347, + 0.004829294513911009, + 0.005499579012393951, + 0.001753604388795793, + 0.003057799767702818, + 0.0031137738842517138, + 0.0023554256185889244, + 0.002201754366979003, + 0.003177969716489315, + 0.0024256217293441296, + 0.002402086276561022, + 0.00619263481348753, + 0.004561955574899912, + 0.010036051273345947, + 0.003981044050306082, + 0.0176906269043684, + 0.07186108827590942, + 0.006718727294355631, + 0.05355755612254143, + 0.004398113116621971, + 0.005771371070295572, + 0.004899291321635246, + 0.005015821661800146, + 0.0038456832990050316, + 0.006854212377220392, + 0.005380102898925543, + 0.007434435188770294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.5233895778656006 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903030306100845 + ], + "opposite_logprob_tokens": [ + -2.4640302658081055 + ], + "logprob_reward": -0.08903030306100845, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903030306100845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06155822053551674, + 0.013213299214839935, + 0.2705173194408417, + 0.04752056300640106, + 0.026600876823067665, + 0.04503370448946953, + 0.0385051891207695, + 0.02396492473781109, + 0.41259393095970154, + 0.0604918971657753 + ], + "credit_map": [ + 0.006700445432215929, + 0.014693386852741241, + 0.0036891563795506954, + 0.00477331830188632, + 0.0028858932200819254, + 0.0026173905935138464, + 0.018473884090781212, + 0.007724744267761707, + 0.0011434454936534166, + 0.0024680846836417913, + 0.0016237891977652907, + 0.001664315233938396, + 0.0009807281894609332, + 0.0015297435456886888, + 0.0015188490506261587, + 0.0022843440528959036, + 0.015814373269677162, + 0.013480044901371002, + 0.07437730580568314, + 0.009801173582673073, + 0.013050414621829987, + 0.006002971902489662, + 0.11663129180669785, + 0.021359747275710106, + 0.004076425451785326, + 0.005512733478099108, + 0.005706743337213993, + 0.005784657318145037, + 0.004913816694170237, + 0.005218186881393194, + 0.0061545539647340775, + 0.010153443552553654, + 0.0018476174445822835, + 0.007100949063897133, + 0.002900804625824094, + 0.003511243499815464, + 0.0030477868858724833, + 0.002558422274887562, + 0.003380476264283061, + 0.0022535768803209066, + 0.0032938504591584206, + 0.006174519192427397, + 0.0055554392747581005, + 0.005593929905444384, + 0.005796877201646566, + 0.005500612780451775, + 0.007601423654705286, + 0.005517053417861462, + 0.0026225207839161158, + 0.006996794603765011, + 0.00534574780613184, + 0.007428435143083334, + 0.0034651681780815125, + 0.0030938657000660896, + 0.005490617826581001, + 0.0040620360523462296, + 0.00173759157769382, + 0.0026181356515735388, + 0.0032150601036846638, + 0.002587947528809309, + 0.0044615017250180244, + 0.002577955601736903, + 0.0033387262374162674, + 0.0034280046820640564, + 0.01239475142210722, + 0.0073978491127491, + 0.023469412699341774, + 0.006426048930734396, + 0.026935627683997154, + 0.18041402101516724, + 0.01316322572529316, + 0.14239300787448883, + 0.005013739690184593, + 0.010575360618531704, + 0.007384892087429762, + 0.009090323932468891, + 0.004093996714800596, + 0.006779298651963472, + 0.0086801303550601, + 0.008874153718352318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.08903030306100845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062052540481090546 + ], + "opposite_logprob_tokens": [ + -2.8120524883270264 + ], + "logprob_reward": -0.062052540481090546, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062052540481090546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07042078673839569, + 0.01437054481357336, + 0.3334091007709503, + 0.05067036673426628, + 0.023378103971481323, + 0.026630984619259834, + 0.07956429570913315, + 0.06632781773805618, + 0.2880162000656128, + 0.04721178859472275 + ], + "credit_map": [ + 0.008203625679016113, + 0.011341921053826809, + 0.009677461348474026, + 0.00838834885507822, + 0.004267435986548662, + 0.0037992955185472965, + 0.010265211574733257, + 0.014477488584816456, + 0.0011475441278889775, + 0.002047070302069187, + 0.0017947505693882704, + 0.001991566503420472, + 0.0014281015610322356, + 0.001998915569856763, + 0.0017214860999956727, + 0.0022411097306758165, + 0.01919255219399929, + 0.016165252774953842, + 0.08469293266534805, + 0.0123721519485116, + 0.02382092922925949, + 0.006276802625507116, + 0.14504721760749817, + 0.025841286405920982, + 0.004342096857726574, + 0.005724175367504358, + 0.005351918749511242, + 0.0061943260952830315, + 0.007582546677440405, + 0.004410711582750082, + 0.007716631982475519, + 0.009347956627607346, + 0.0016522364458069205, + 0.004012040328234434, + 0.0036117194686084986, + 0.0028364588506519794, + 0.0028016590513288975, + 0.0026259361766278744, + 0.002970567438751459, + 0.002867485862225294, + 0.0030292070005089045, + 0.003871095133945346, + 0.003121620276942849, + 0.003746610600501299, + 0.003401806578040123, + 0.002397298812866211, + 0.003685325151309371, + 0.003378018969669938, + 0.0056349895894527435, + 0.01578487455844879, + 0.010005342774093151, + 0.01646420918405056, + 0.0077509088441729546, + 0.004863751586526632, + 0.011612107045948505, + 0.007448114920407534, + 0.004141489043831825, + 0.00937714334577322, + 0.00863936822861433, + 0.01083198469132185, + 0.008430276066064835, + 0.005480395630002022, + 0.011716626584529877, + 0.007710531819611788, + 0.008909255266189575, + 0.005170474760234356, + 0.016181040555238724, + 0.004801924340426922, + 0.020677534863352776, + 0.12513022124767303, + 0.009633908048272133, + 0.09751180559396744, + 0.003784478409215808, + 0.008657767437398434, + 0.005587865132838488, + 0.006397891324013472, + 0.003663107519969344, + 0.005081013776361942, + 0.006349524017423391, + 0.007690142374485731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.062052540481090546 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera pans in a flawless 360-degree pan, beginning from a slightly bird's-eye view, slowly sweeping across a futuristic, intelligent nursing bed. Constructed of a sleek white composite material and translucent blue biogel, the bed's edges shimmer with a soft blue glow. A translucent holographic display hovers above the headboard, flickering a steady pulse of heart rate and vital signs. Lying quietly on the bed, a young woman with cascading silver hair spread across the pillow, reflecting the cool light of the surroundings. She wears a form-fitting white, high-tech nanofiber nightgown with a subtle pearlescent sheen. The camera moves steadily and gracefully, focusing precisely on her delicate features. Her fair skin appears crystal clear, cast against the dreamy purple and cyan light of the nebula outside the window and the bed's own blue glow. At that moment, her long, thick eyelashes begin to flutter, the first subtle sign of awakening from a deep sleep. Immediately, her brow furrowed slightly, her breathing became deep and clear, and her slender fingers unconsciously curled. As the camera panned to her front, her eyelids finally slowly opened, revealing a pair of clear, yet slightly lost, light-blue eyes, trying to focus on the minimalist, futuristic medical pod. The entire scene exuded the quality of a top-tier CG work, with every detail achieving photorealism and every play of light and shadow meticulously nuanced, creating a tranquil, sci-fi, and beautiful atmosphere of awakening.", + "seed": 31476, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_12/index_9072_seed_31476_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.35825724359601735, + "vlm_logprob": -0.35825724359601735, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.35825724359601735, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.607260823249817 + ], + "opposite_logprob_tokens": [ + -0.23226085305213928 + ], + "logprob_reward": -1.607260823249817, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.607260823249817, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05931052938103676, + 0.017372995615005493, + 0.020863989368081093, + 0.022163862362504005, + 0.035994384437799454, + 0.025100942701101303, + 0.0529312826693058, + 0.7341479063034058, + 0.019979538396000862, + 0.012134607881307602 + ], + "credit_map": [ + 0.004436726216226816, + 0.0049708206206560135, + 0.00640821922570467, + 0.008711432106792927, + 0.005987232550978661, + 0.007917962968349457, + 0.011325433850288391, + 0.009552700445055962, + 0.001435428624972701, + 0.0013375673443078995, + 0.002562928246334195, + 0.003783327993005514, + 0.001558581367135048, + 0.0013122238451614976, + 0.003394600236788392, + 0.0019883383065462112, + 0.001512466580606997, + 0.0024525003973394632, + 0.0031401445157825947, + 0.004771578125655651, + 0.0017932422924786806, + 0.002097003860399127, + 0.0024741184897720814, + 0.0026229352224618196, + 0.0009192434954456985, + 0.0023824332747608423, + 0.0028282562270760536, + 0.0027611663099378347, + 0.001501407939940691, + 0.001276394003070891, + 0.005884024780243635, + 0.0046109361574053764, + 0.00126551219727844, + 0.004850946832448244, + 0.006663403939455748, + 0.006933736614882946, + 0.006332713179290295, + 0.0020657379645854235, + 0.004715858958661556, + 0.003166476497426629, + 0.001696330145932734, + 0.003573625348508358, + 0.0032228506170213223, + 0.003988097887486219, + 0.0034260780084878206, + 0.001521834870800376, + 0.003220193088054657, + 0.00445193238556385, + 0.0047319852747023106, + 0.009607043117284775, + 0.0058994959108531475, + 0.010404620319604874, + 0.005204391665756702, + 0.004524005111306906, + 0.0069140843115746975, + 0.0056456527672708035, + 0.04688599333167076, + 0.009458113461732864, + 0.2800246477127075, + 0.030792709439992905, + 0.011743992567062378, + 0.036746103316545486, + 0.028467237949371338, + 0.2900291085243225, + 0.002888208720833063, + 0.002442188560962677, + 0.002116476185619831, + 0.0025754901580512524, + 0.002597844460979104, + 0.002247260417789221, + 0.0028899251483380795, + 0.002222142880782485, + 0.0012724247062578797, + 0.0019015723373740911, + 0.0017644399777054787, + 0.0012788581661880016, + 0.0012374328216537833, + 0.0011714505963027477, + 0.0018038168782368302, + 0.0017046122811734676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -1.607260823249817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026405178010463715 + ], + "opposite_logprob_tokens": [ + -3.651405096054077 + ], + "logprob_reward": -0.026405178010463715, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026405178010463715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04811852052807808, + 0.01689399592578411, + 0.09510424733161926, + 0.01817006804049015, + 0.021437373012304306, + 0.024578064680099487, + 0.0324617475271225, + 0.7065696716308594, + 0.02101714350283146, + 0.015649111941456795 + ], + "credit_map": [ + 0.004164000507444143, + 0.007139339577406645, + 0.006030945107340813, + 0.007752030622214079, + 0.004741721786558628, + 0.005017808172851801, + 0.006686218082904816, + 0.006586454808712006, + 0.0015440480783581734, + 0.0014386452967301011, + 0.0026041637174785137, + 0.0027143803890794516, + 0.00134998036082834, + 0.0013374094851315022, + 0.0032755753491073847, + 0.0026297923177480698, + 0.008935479447245598, + 0.010441705584526062, + 0.011602547019720078, + 0.015415939502418041, + 0.009042326360940933, + 0.008625537157058716, + 0.01389322616159916, + 0.017147492617368698, + 0.0008824749384075403, + 0.002244953066110611, + 0.0036631093826144934, + 0.0029966842848807573, + 0.0016555578913539648, + 0.0013470675330609083, + 0.0028609393630176783, + 0.0025192818138748407, + 0.0009578964090906084, + 0.0035037584602832794, + 0.002976406831294298, + 0.003914277534931898, + 0.0037926482036709785, + 0.0014840421499684453, + 0.002703015459701419, + 0.0021053270902484655, + 0.0015085951890796423, + 0.0033669965341687202, + 0.003246683394536376, + 0.003928898833692074, + 0.003482591826468706, + 0.0015152144478634, + 0.0030727130360901356, + 0.004456372931599617, + 0.003001681761816144, + 0.0060021961107850075, + 0.00377263268455863, + 0.005785287823528051, + 0.0032341305632144213, + 0.0027802949771285057, + 0.004517173394560814, + 0.0033683518413454294, + 0.04486764594912529, + 0.008867550641298294, + 0.2699667811393738, + 0.027955904603004456, + 0.011014378629624844, + 0.03567277640104294, + 0.027398010715842247, + 0.2808266580104828, + 0.0017009677831083536, + 0.0026658903807401657, + 0.001908774022012949, + 0.003371785394847393, + 0.0028428523801267147, + 0.0025223076809197664, + 0.0033843950368463993, + 0.002620171057060361, + 0.0013108601560816169, + 0.002023630077019334, + 0.001736875856295228, + 0.0020418104249984026, + 0.002151135355234146, + 0.0016429821262136102, + 0.0026091153267771006, + 0.0021327028516680002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.026405178010463715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035345703363419 + ], + "opposite_logprob_tokens": [ + -2.350353479385376 + ], + "logprob_reward": -0.10035345703363419, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035345703363419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08540432900190353, + 0.021211611106991768, + 0.12748557329177856, + 0.01770751178264618, + 0.044527724385261536, + 0.037976209074258804, + 0.05441842973232269, + 0.5604857206344604, + 0.03018038533627987, + 0.02060241997241974 + ], + "credit_map": [ + 0.00734225008636713, + 0.014297361485660076, + 0.011297091841697693, + 0.014089908450841904, + 0.008446360938251019, + 0.010201586410403252, + 0.00895016174763441, + 0.010779613628983498, + 0.0016622240655124187, + 0.0016145501285791397, + 0.003885480808094144, + 0.003966453485190868, + 0.0016162273241207004, + 0.0020495238713920116, + 0.0031123554799705744, + 0.0033047962933778763, + 0.012302642688155174, + 0.014629713259637356, + 0.014427761547267437, + 0.022319890558719635, + 0.011822902597486973, + 0.012119417078793049, + 0.01752007193863392, + 0.022343173623085022, + 0.0010807170765474439, + 0.0025818080175668, + 0.00384631659835577, + 0.003322901902720332, + 0.001249570632353425, + 0.0013700913405045867, + 0.002612408949062228, + 0.0016436971491202712, + 0.0021295801270753145, + 0.005230412818491459, + 0.006554520223289728, + 0.0065943896770477295, + 0.009596296586096287, + 0.0034007644280791283, + 0.006219024304300547, + 0.0048027425073087215, + 0.0018642616923898458, + 0.004579140339046717, + 0.004803398158401251, + 0.006846916396170855, + 0.0059465705417096615, + 0.002783598843961954, + 0.004378973506391048, + 0.006773349829018116, + 0.0056063891388475895, + 0.009152625687420368, + 0.007412945851683617, + 0.007150338031351566, + 0.005119405686855316, + 0.004914925899356604, + 0.008335649967193604, + 0.006726154126226902, + 0.034243740141391754, + 0.011544888839125633, + 0.20604807138442993, + 0.026734231039881706, + 0.01238186750560999, + 0.027784176170825958, + 0.02490774728357792, + 0.21684107184410095, + 0.0028803071472793818, + 0.0035122809931635857, + 0.003229432972148061, + 0.004960974212735891, + 0.003802642924711108, + 0.0036812752950936556, + 0.0042828950099647045, + 0.003830577014014125, + 0.0019394289702177048, + 0.0029983902350068092, + 0.0022998556960374117, + 0.002826003124937415, + 0.0025498829782009125, + 0.002280044136568904, + 0.002972061512991786, + 0.002736754482612014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.10035345703363419 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197650365531445 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197650365531445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197650365531445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020483650267124176, + 0.006955359131097794, + 0.04167827218770981, + 0.01967792585492134, + 0.05878782272338867, + 0.022402280941605568, + 0.02013986185193062, + 0.7541385889053345, + 0.0322817862033844, + 0.023454515263438225 + ], + "credit_map": [ + 0.001313253422267735, + 0.0014336392050608993, + 0.00175443512853235, + 0.0032068383879959583, + 0.0015286192065104842, + 0.0024548920337110758, + 0.005000689532607794, + 0.0037912840489298105, + 0.0006378344260156155, + 0.0005994910607114434, + 0.0009389907354488969, + 0.001896763569675386, + 0.0007147977012209594, + 0.0006637669866904616, + 0.0008376618498004973, + 0.0006660527433268726, + 0.004108414053916931, + 0.005354264285415411, + 0.003493757452815771, + 0.007799367420375347, + 0.0027656101156026125, + 0.003908042795956135, + 0.005817345343530178, + 0.008431469090282917, + 0.0008737072930671275, + 0.0022157751955091953, + 0.003173256991431117, + 0.003275316907092929, + 0.0013973157620057464, + 0.0012432814110070467, + 0.003951831255108118, + 0.003547441679984331, + 0.002444396959617734, + 0.009258291684091091, + 0.009984456934034824, + 0.011338175274431705, + 0.009999718517065048, + 0.0030797887593507767, + 0.007574942894279957, + 0.0051080575212836266, + 0.001369175035506487, + 0.003083872376009822, + 0.002637481549754739, + 0.004486788995563984, + 0.00316578708589077, + 0.0016449636314064264, + 0.003103074384853244, + 0.002911137882620096, + 0.0019042551284655929, + 0.003138081869110465, + 0.002466046717017889, + 0.0041597215458750725, + 0.002085308311507106, + 0.0020256363786756992, + 0.0027217885944992304, + 0.0016390230739489198, + 0.04775601252913475, + 0.008975381031632423, + 0.28870493173599243, + 0.029503487050533295, + 0.011503183282911777, + 0.03841337934136391, + 0.02868707664310932, + 0.30059513449668884, + 0.002991495421156287, + 0.0034875122364610434, + 0.0028450533282011747, + 0.006223923992365599, + 0.003557672491297126, + 0.0034706934820860624, + 0.005443500354886055, + 0.004261932801455259, + 0.0016415767604485154, + 0.0023029360454529524, + 0.001437618862837553, + 0.0025615703780204058, + 0.004588881507515907, + 0.0018871101783588529, + 0.006459981668740511, + 0.002574838697910309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.014197650365531445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a 360-degree pan starting from a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the nursing bed constructed of white composite material with translucent blue biogel edges that glow softly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent holographic display hovering above the headboard showing a flickering pulse of vital signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young woman with cascading silver hair wearing a form-fitting white nanofiber nightgown lying on the bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's eyelid slowly open to reveal light-blue eyes as the camera pans to her front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043069109320640564 + ], + "opposite_logprob_tokens": [ + -3.1680691242218018 + ], + "logprob_reward": -0.043069109320640564, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043069109320640564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02738158404827118, + 0.011842953972518444, + 0.027716679498553276, + 0.02557692490518093, + 0.03467513620853424, + 0.032301321625709534, + 0.05832303315401077, + 0.7397110462188721, + 0.02785550057888031, + 0.014615818858146667 + ], + "credit_map": [ + 0.0015401242999359965, + 0.0023637767881155014, + 0.0030926531180739403, + 0.003496995661407709, + 0.00241903611458838, + 0.0025713429786264896, + 0.007261834107339382, + 0.004635821096599102, + 0.0010411243420094252, + 0.0012616553576663136, + 0.00187667366117239, + 0.0020370411220937967, + 0.0010727036278694868, + 0.0011523412540555, + 0.0021464417222887278, + 0.0012549731181934476, + 0.002270587719976902, + 0.003526645712554455, + 0.0031172882299870253, + 0.004678800236433744, + 0.0029702805913984776, + 0.00308477901853621, + 0.0030192555859684944, + 0.0050490424036979675, + 0.001449504168704152, + 0.00283582741394639, + 0.003152168123051524, + 0.0035115473438054323, + 0.0016060270136222243, + 0.0019141010707244277, + 0.005958308931440115, + 0.005149441305547953, + 0.0018603109056130052, + 0.007837880402803421, + 0.005560598336160183, + 0.006110019516199827, + 0.003969849552959204, + 0.0019537762273103, + 0.003551311092451215, + 0.003831392154097557, + 0.001351417857222259, + 0.005194340832531452, + 0.0047973329201340675, + 0.006511569023132324, + 0.003680763766169548, + 0.002376839052885771, + 0.003784037195146084, + 0.004605020862072706, + 0.005959981586784124, + 0.007346646394580603, + 0.008720571175217628, + 0.00908145122230053, + 0.005728249903768301, + 0.009248844347894192, + 0.006638070102781057, + 0.005599213298410177, + 0.047160468995571136, + 0.008958670310676098, + 0.2825537919998169, + 0.02934213913977146, + 0.010982520878314972, + 0.03769221529364586, + 0.027983084321022034, + 0.2950381934642792, + 0.004251033533364534, + 0.0029691641684621572, + 0.0027562077157199383, + 0.00446674507111311, + 0.0034870540257543325, + 0.0032128358725458384, + 0.003402321832254529, + 0.003310137428343296, + 0.0010383536573499441, + 0.0016655675135552883, + 0.001706408686004579, + 0.0015009789494797587, + 0.0024570380337536335, + 0.001231302972882986, + 0.0030394941568374634, + 0.0019766762852668762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.043069109320640564 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_tensor([5792, 5792])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_tensor([5792, 5792])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..347e85aadd59c877e077bb1971c230f21e5e32bd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_tensor([5792, 5792])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up time-lapse shot, the camera held still, focused at eye level on a square glass container placed against a stark black background. A concentrated, cool white overhead light shines vertically, precisely illuminating the center of the container, creating a dramatic, theatrical effect. The container is filled with a clear, yet slightly viscous, unknown liquid. A vibrant, crimson apple, completely submerged, its skin smooth with a natural waxy sheen, its stem still emerald green. As the time-lapse begins, fine cracks begin to appear in the apple's vibrant red skin, and tiny bubbles emerge. Strands of crimson pigment, like ink, cautiously seep through the cracks, slowly diffusing through the still liquid. Then, the dissolution process accelerates rapidly and decisively, and the apple's flesh begins to disintegrate into countless tiny particles and fibers, its solid form gradually becoming blurred and soft. Finally, the entire apple completely disintegrates, and under the penetration of top light, it turns into a pool of brilliant colors that steadily and calmly rotates and mixes in the liquid. Deep red, soft milky white and light yellow are intertwined to form an ever-changing, flowing abstract painting. The whole process has both the precise beauty of scientific experiments and a surreal and charming artistic atmosphere.", + "seed": 32992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_5792_seed_32992_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.031215860578231513, + "vlm_logprob": -0.031215860578231513, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.031215860578231513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06753379851579666 + ], + "opposite_logprob_tokens": [ + -3.3175337314605713 + ], + "logprob_reward": -0.06753379851579666, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06753379851579666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3431851863861084, + 0.06760546565055847, + 0.03070341795682907, + 0.07695677131414413, + 0.03529779613018036, + 0.16193059086799622, + 0.040370017290115356, + 0.02643907256424427, + 0.1379656195640564, + 0.07954610884189606 + ], + "credit_map": [ + 0.03507554158568382, + 0.02181539312005043, + 0.08287370204925537, + 0.030919428914785385, + 0.03458274155855179, + 0.0298578143119812, + 0.03878488019108772, + 0.06927569210529327, + 0.008500814437866211, + 0.00437394343316555, + 0.01294642873108387, + 0.006784592755138874, + 0.007800136227160692, + 0.0068652634508907795, + 0.009946627542376518, + 0.010387659072875977, + 0.0027806616853922606, + 0.002063973108306527, + 0.007674274500459433, + 0.003221635241061449, + 0.0030955511610955, + 0.0017170215724036098, + 0.004692819435149431, + 0.005457479972392321, + 0.006859677843749523, + 0.0064623998478055, + 0.01849517598748207, + 0.0077524068765342236, + 0.007911999709904194, + 0.0053283837623894215, + 0.009468559175729752, + 0.014678169041872025, + 0.002334813354536891, + 0.0021328742150217295, + 0.00998650211840868, + 0.0022995430044829845, + 0.002687129657715559, + 0.002209101105108857, + 0.0024073573295027018, + 0.011240473948419094, + 0.03618714585900307, + 0.012821325100958347, + 0.027207626029849052, + 0.012099018320441246, + 0.015824800357222557, + 0.024277696385979652, + 0.010188378393650055, + 0.023324593901634216, + 0.002622835570946336, + 0.004120666533708572, + 0.004493260756134987, + 0.005893789231777191, + 0.008238092064857483, + 0.0028871686663478613, + 0.005662304349243641, + 0.006451900117099285, + 0.001893803826533258, + 0.003005599370226264, + 0.0030697346664965153, + 0.005101834423840046, + 0.0027420171536505222, + 0.0027806502766907215, + 0.004422392696142197, + 0.003423040732741356, + 0.008346492424607277, + 0.011682962067425251, + 0.007399134803563356, + 0.03438486158847809, + 0.009535407647490501, + 0.010573040693998337, + 0.014643412083387375, + 0.04140030965209007, + 0.008464235812425613, + 0.008594874292612076, + 0.007837055251002312, + 0.011923282407224178, + 0.01106051355600357, + 0.009812156669795513, + 0.01275354903191328, + 0.009100440889596939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.06753379851579666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011102800257503986 + ], + "opposite_logprob_tokens": [ + -4.511102676391602 + ], + "logprob_reward": -0.011102800257503986, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011102800257503986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6352158188819885, + 0.06411980092525482, + 0.019141927361488342, + 0.006926414091140032, + 0.02894390933215618, + 0.1760508269071579, + 0.01983289234340191, + 0.008235997520387173, + 0.012842303141951561, + 0.028690017759799957 + ], + "credit_map": [ + 0.06677974760532379, + 0.043663665652275085, + 0.14409534633159637, + 0.0670047402381897, + 0.07454003393650055, + 0.0604901947081089, + 0.07425893098115921, + 0.1043832004070282, + 0.0072747692465782166, + 0.004128257278352976, + 0.01382734626531601, + 0.005024454090744257, + 0.007588234264403582, + 0.0066862162202596664, + 0.007458863314241171, + 0.012131664901971817, + 0.0024039307609200478, + 0.0014894664054736495, + 0.004761811811476946, + 0.0015388574684038758, + 0.0012809347826987505, + 0.0019461490446701646, + 0.0015024697640910745, + 0.004218308720737696, + 0.000991071225143969, + 0.0005752550787292421, + 0.0008215620764531195, + 0.0009831623174250126, + 0.00089051149552688, + 0.0007203550194390118, + 0.001226286287419498, + 0.0007182110566645861, + 0.0024710907600820065, + 0.0012905338080599904, + 0.009799471125006676, + 0.0015873439842835069, + 0.0017273124540224671, + 0.0017668261425569654, + 0.0014469060115516186, + 0.008854424580931664, + 0.04093684256076813, + 0.013089996762573719, + 0.029724551364779472, + 0.012145458720624447, + 0.013465611264109612, + 0.02817545086145401, + 0.00935108121484518, + 0.02916184812784195, + 0.001951959333382547, + 0.001544717699289322, + 0.003418935230001807, + 0.001793111558072269, + 0.0031521169003099203, + 0.002461347496137023, + 0.0018009409541264176, + 0.003709764452651143, + 0.0005303338984958827, + 0.0009761854307726026, + 0.0006142021156847477, + 0.001646864926442504, + 0.001021097064949572, + 0.0008638398139737546, + 0.0014471483882516623, + 0.0011363262310624123, + 0.0007832027040421963, + 0.001063700532540679, + 0.0009310515015386045, + 0.0022313802037388086, + 0.0013600061647593975, + 0.0009881430305540562, + 0.002627262379974127, + 0.0028575577307492495, + 0.0025529402773827314, + 0.002948845038190484, + 0.002749598352238536, + 0.0046309819445014, + 0.0037564849480986595, + 0.0027583250775933266, + 0.006235332228243351, + 0.0030575133860111237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.011102800257503986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014227267354726791 + ], + "opposite_logprob_tokens": [ + -4.264227390289307 + ], + "logprob_reward": -0.014227267354726791, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014227267354726791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08169008046388626, + 0.061508506536483765, + 0.035160865634679794, + 0.05805748701095581, + 0.10789608210325241, + 0.2905518114566803, + 0.14647088944911957, + 0.0381489023566246, + 0.1241021454334259, + 0.05641328915953636 + ], + "credit_map": [ + 0.007064970210194588, + 0.006099478341639042, + 0.01795288920402527, + 0.010393094271421432, + 0.009745709598064423, + 0.005090133752673864, + 0.01272626407444477, + 0.01261753961443901, + 0.006627946626394987, + 0.003707717638462782, + 0.014588074758648872, + 0.006321929860860109, + 0.0075235567055642605, + 0.005153197795152664, + 0.007076654117554426, + 0.010509423911571503, + 0.004042366053909063, + 0.003299119882285595, + 0.006709978450089693, + 0.004328525625169277, + 0.0042717247270047665, + 0.002571757882833481, + 0.005181116051971912, + 0.004756275564432144, + 0.006351934280246496, + 0.005284156650304794, + 0.014447879046201706, + 0.004835486877709627, + 0.006551698315888643, + 0.004311311524361372, + 0.005057159345597029, + 0.011217854917049408, + 0.00796740036457777, + 0.006510887760668993, + 0.03415019437670708, + 0.008575850166380405, + 0.008061573840677738, + 0.004880047868937254, + 0.007801387459039688, + 0.029948733747005463, + 0.06743209064006805, + 0.02091965265572071, + 0.05019986256957054, + 0.01989489607512951, + 0.021387064829468727, + 0.046550799161195755, + 0.016264328733086586, + 0.047903113067150116, + 0.011674958281219006, + 0.014353975653648376, + 0.012428924441337585, + 0.031153416261076927, + 0.01794741302728653, + 0.011441607028245926, + 0.025251606479287148, + 0.02221897803246975, + 0.0033505288884043694, + 0.004703295882791281, + 0.0034691120963543653, + 0.005256377160549164, + 0.004718773998320103, + 0.00463068438693881, + 0.005345467943698168, + 0.006674660369753838, + 0.008163252845406532, + 0.00942549854516983, + 0.00643780454993248, + 0.03155260160565376, + 0.00874388124793768, + 0.01019282266497612, + 0.01162100862711668, + 0.037965256720781326, + 0.004899836145341396, + 0.0050827753730118275, + 0.00593476602807641, + 0.008114315569400787, + 0.008123017847537994, + 0.008519063703715801, + 0.009452780708670616, + 0.006286728661507368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.014227267354726791 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011813811725005507 + ], + "opposite_logprob_tokens": [ + -6.751181602478027 + ], + "logprob_reward": -0.0011813811725005507, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011813811725005507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06769782304763794, + 0.06877721846103668, + 0.026802200824022293, + 0.025459421798586845, + 0.10435541719198227, + 0.13472440838813782, + 0.15114820003509521, + 0.08556690067052841, + 0.24538810551166534, + 0.09008026868104935 + ], + "credit_map": [ + 0.007845498621463776, + 0.005984381772577763, + 0.011193024925887585, + 0.009999096393585205, + 0.006802703253924847, + 0.006976344157010317, + 0.008797146379947662, + 0.010099629871547222, + 0.0068589565344154835, + 0.004586680792272091, + 0.015417555347084999, + 0.005771418567746878, + 0.0071633923798799515, + 0.006621854845434427, + 0.007397871930152178, + 0.014959489926695824, + 0.003305227728560567, + 0.0024493755772709846, + 0.006042048800736666, + 0.002587349619716406, + 0.003533243201673031, + 0.0017834578175097704, + 0.0030958603601902723, + 0.004005638882517815, + 0.0030742320232093334, + 0.0019158286741003394, + 0.003447736846283078, + 0.003538192017003894, + 0.003979768604040146, + 0.0023369472473859787, + 0.004445991013199091, + 0.002720724558457732, + 0.007125837728381157, + 0.004944344982504845, + 0.03719942644238472, + 0.005081444978713989, + 0.005222170148044825, + 0.004616435617208481, + 0.004784754477441311, + 0.035381019115448, + 0.02688005566596985, + 0.011074249632656574, + 0.02057054452598095, + 0.01067822054028511, + 0.014728053472936153, + 0.01772179827094078, + 0.013469865545630455, + 0.01960163190960884, + 0.010202149860560894, + 0.01406015269458294, + 0.012196071445941925, + 0.02462136745452881, + 0.025707632303237915, + 0.009639859199523926, + 0.023289721459150314, + 0.03143125772476196, + 0.008086645044386387, + 0.00899557489901781, + 0.007754539605230093, + 0.011137035675346851, + 0.008895168080925941, + 0.013085474260151386, + 0.011245576664805412, + 0.01636689156293869, + 0.01526187639683485, + 0.01862335577607155, + 0.013082319870591164, + 0.06531671434640884, + 0.017452184110879898, + 0.019623296335339546, + 0.021244261413812637, + 0.07478412985801697, + 0.008560576476156712, + 0.007759269792586565, + 0.009260939434170723, + 0.011621082201600075, + 0.01307751890271902, + 0.010934894904494286, + 0.01761634461581707, + 0.011249647475779057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.0011813811725005507 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203405559062958 + ], + "opposite_logprob_tokens": [ + -2.8120341300964355 + ], + "logprob_reward": -0.06203405559062958, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203405559062958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12030737102031708, + 0.042288895696401596, + 0.03647250682115555, + 0.08888890594244003, + 0.024252789095044136, + 0.16445280611515045, + 0.08636292070150375, + 0.03352849930524826, + 0.210015207529068, + 0.193430095911026 + ], + "credit_map": [ + 0.013875479809939861, + 0.009037384763360023, + 0.023387974128127098, + 0.01622285321354866, + 0.012963058426976204, + 0.010414219461381435, + 0.016803210601210594, + 0.01760319247841835, + 0.003944751340895891, + 0.0037822064477950335, + 0.008129706606268883, + 0.005698402412235737, + 0.004949022550135851, + 0.0030117793940007687, + 0.005611589644104242, + 0.007161438465118408, + 0.0032230850774794817, + 0.002823636634275317, + 0.00663809385150671, + 0.005190353840589523, + 0.0028686742298305035, + 0.002623859327286482, + 0.006831820122897625, + 0.006272985599935055, + 0.006483504083007574, + 0.006135880947113037, + 0.02576020546257496, + 0.00844657327979803, + 0.008072663098573685, + 0.004630593582987785, + 0.010450112633407116, + 0.01890937052667141, + 0.0024355100467801094, + 0.002784377196803689, + 0.004890106152743101, + 0.003275624942034483, + 0.0027255951426923275, + 0.002044585533440113, + 0.003130026627331972, + 0.0029669636860489845, + 0.038587700575590134, + 0.01232051569968462, + 0.029272107407450676, + 0.011251918971538544, + 0.014198158867657185, + 0.02576892264187336, + 0.009602373465895653, + 0.02345110848546028, + 0.006048836279660463, + 0.006263222545385361, + 0.007117271889001131, + 0.012306426651775837, + 0.015585283748805523, + 0.006576229352504015, + 0.011628919281065464, + 0.02083672396838665, + 0.002969520166516304, + 0.0037442699540406466, + 0.00402575358748436, + 0.006413252558559179, + 0.00423969142138958, + 0.002489790553227067, + 0.005622827913612127, + 0.004023394547402859, + 0.012080901302397251, + 0.015561175532639027, + 0.010548283345997334, + 0.052081115543842316, + 0.01720334216952324, + 0.01609881967306137, + 0.023789269849658012, + 0.06265230476856232, + 0.011459622532129288, + 0.018545281141996384, + 0.014852531254291534, + 0.03230373188853264, + 0.02297384664416313, + 0.012410846538841724, + 0.06013860926032066, + 0.02074563503265381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.06203405559062958 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up time-lapse shot, the camera held still, focused at eye level on a square glass container placed against a stark black background. A concentrated, cool white overhead light shines vertically, precisely illuminating the center of the container, creating a dramatic, theatrical effect. The container is filled with a clear, yet slightly viscous, unknown liquid. A vibrant, crimson apple, completely submerged, its skin smooth with a natural waxy sheen, its stem still emerald green. As the time-lapse begins, fine cracks begin to appear in the apple's vibrant red skin, and tiny bubbles emerge. Strands of crimson pigment, like ink, cautiously seep through the cracks, slowly diffusing through the still liquid. Then, the dissolution process accelerates rapidly and decisively, and the apple's flesh begins to disintegrate into countless tiny particles and fibers, its solid form gradually becoming blurred and soft. Finally, the entire apple completely disintegrates, and under the penetration of top light, it turns into a pool of brilliant colors that steadily and calmly rotates and mixes in the liquid. Deep red, soft milky white and light yellow are intertwined to form an ever-changing, flowing abstract painting. The whole process has both the precise beauty of scientific experiments and a surreal and charming artistic atmosphere.", + "seed": 32993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_5792_seed_32993_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.040043990174308416, + "vlm_logprob": -0.040043990174308416, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.040043990174308416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06890616565942764 + ], + "opposite_logprob_tokens": [ + -3.06890606880188 + ], + "logprob_reward": -0.06890616565942764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06890616565942764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2654689848423004, + 0.08454069495201111, + 0.09902019798755646, + 0.058663610368967056, + 0.028700117021799088, + 0.06693031638860703, + 0.10163235664367676, + 0.08604059368371964, + 0.07154619693756104, + 0.13745690882205963 + ], + "credit_map": [ + 0.032336410135030746, + 0.04987140744924545, + 0.031207004562020302, + 0.032227784395217896, + 0.027359671890735626, + 0.029897205531597137, + 0.03053685463964939, + 0.03203263133764267, + 0.009384398348629475, + 0.007993978448212147, + 0.007965536788105965, + 0.01763703115284443, + 0.010830419138073921, + 0.007919886149466038, + 0.011715610511600971, + 0.01109382789582014, + 0.011679550632834435, + 0.019416073337197304, + 0.013578610494732857, + 0.010386530309915543, + 0.010460450313985348, + 0.013455523177981377, + 0.00992576964199543, + 0.010117681697010994, + 0.005272592883557081, + 0.010722842067480087, + 0.007287558168172836, + 0.005975952371954918, + 0.008664182387292385, + 0.0064684925600886345, + 0.006885015405714512, + 0.007386975921690464, + 0.003701096633449197, + 0.0024783003609627485, + 0.0027819897513836622, + 0.0028938227333128452, + 0.0030505871400237083, + 0.006119873840361834, + 0.0030206479132175446, + 0.004653798416256905, + 0.005094420164823532, + 0.003581615397706628, + 0.0036879878025501966, + 0.012304050847887993, + 0.005341450683772564, + 0.0077402954921126366, + 0.01621698960661888, + 0.012963504530489445, + 0.01316932961344719, + 0.009571553207933903, + 0.007425749208778143, + 0.018340367823839188, + 0.007813508622348309, + 0.01441255398094654, + 0.015195679850876331, + 0.015703614801168442, + 0.014534576795995235, + 0.010812119580805302, + 0.005877369083464146, + 0.009412765502929688, + 0.0076582953333854675, + 0.017357055097818375, + 0.009799961000680923, + 0.010588457807898521, + 0.007215709891170263, + 0.005718852858990431, + 0.006578349974006414, + 0.011514728888869286, + 0.006851512473076582, + 0.008304694667458534, + 0.015115989372134209, + 0.01024636346846819, + 0.01860676519572735, + 0.00787070207297802, + 0.03897185996174812, + 0.027612291276454926, + 0.008659295737743378, + 0.010509815998375416, + 0.014193186536431313, + 0.011032992042601109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.06890616565942764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500597506761551 + ], + "opposite_logprob_tokens": [ + -2.9300060272216797 + ], + "logprob_reward": -0.05500597506761551, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500597506761551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12721216678619385, + 0.18160602450370789, + 0.03416081517934799, + 0.03005245327949524, + 0.14625385403633118, + 0.023349249735474586, + 0.032957155257463455, + 0.11011351644992828, + 0.03860188648104668, + 0.27569282054901123 + ], + "credit_map": [ + 0.01696060784161091, + 0.01750222221016884, + 0.015587781555950642, + 0.016253875568509102, + 0.012676035985350609, + 0.0189159344881773, + 0.013419979251921177, + 0.015895722433924675, + 0.021305691450834274, + 0.017172586172819138, + 0.024700049310922623, + 0.025528524070978165, + 0.013103845529258251, + 0.02247452922165394, + 0.02012551762163639, + 0.03719528019428253, + 0.0033151654060930014, + 0.005265617277473211, + 0.006471444386988878, + 0.0038416548632085323, + 0.002767510013654828, + 0.004815433174371719, + 0.0037606453988701105, + 0.003923347219824791, + 0.002760128816589713, + 0.004673581104725599, + 0.0048376875929534435, + 0.003385921474546194, + 0.00422819284722209, + 0.0033377441577613354, + 0.0031623737886548042, + 0.0036668223328888416, + 0.012943541631102562, + 0.015526611357927322, + 0.012020335532724857, + 0.009937680326402187, + 0.014247605577111244, + 0.03602974861860275, + 0.014982124790549278, + 0.030566219240427017, + 0.002606129040941596, + 0.0014143689768388867, + 0.0012747639557346702, + 0.003702159970998764, + 0.0019823983311653137, + 0.00345645472407341, + 0.005036903545260429, + 0.0038760725874453783, + 0.00429559638723731, + 0.003790795337408781, + 0.0019958470948040485, + 0.006134727504104376, + 0.0028833558317273855, + 0.004401074722409248, + 0.0049317688681185246, + 0.004523989278823137, + 0.02056051604449749, + 0.012568182311952114, + 0.008516128174960613, + 0.014583014883100986, + 0.009262419305741787, + 0.01707608252763748, + 0.01473617646843195, + 0.012810993008315563, + 0.00227607530541718, + 0.002641083672642708, + 0.0032727678772062063, + 0.007055269088596106, + 0.003240484045818448, + 0.008113328367471695, + 0.006961291655898094, + 0.005041586235165596, + 0.03425207734107971, + 0.015406406484544277, + 0.08068059384822845, + 0.05013282224535942, + 0.019290611147880554, + 0.0193470511585474, + 0.025258980691432953, + 0.03132427856326103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.05500597506761551 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108797758817673 + ], + "opposite_logprob_tokens": [ + -5.504108905792236 + ], + "logprob_reward": -0.004108797758817673, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108797758817673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1864897906780243, + 0.27929189801216125, + 0.06051883473992348, + 0.04994349926710129, + 0.03557421639561653, + 0.05963518097996712, + 0.05396348983049393, + 0.07913439720869064, + 0.05396580323576927, + 0.141482874751091 + ], + "credit_map": [ + 0.016591520980000496, + 0.02193467691540718, + 0.020306283608078957, + 0.0356563962996006, + 0.02895660325884819, + 0.017121244221925735, + 0.025252539664506912, + 0.020670542493462563, + 0.029585715383291245, + 0.033691149204969406, + 0.02856171503663063, + 0.05145343765616417, + 0.025896023958921432, + 0.028353741392493248, + 0.036433685570955276, + 0.04531644657254219, + 0.00581503938883543, + 0.006277695298194885, + 0.011889754794538021, + 0.007480471394956112, + 0.005378710106015205, + 0.005603391211479902, + 0.009203994646668434, + 0.008869776502251625, + 0.0049665337428450584, + 0.0032270457595586777, + 0.011878297664225101, + 0.006313066929578781, + 0.0033123381435871124, + 0.003991248551756144, + 0.007754940539598465, + 0.00850002747029066, + 0.003900841111317277, + 0.003233162220567465, + 0.003372992854565382, + 0.005447151605039835, + 0.0035122428089380264, + 0.005683713126927614, + 0.005040342453867197, + 0.005383770912885666, + 0.0046184551902115345, + 0.00643352372571826, + 0.0036651145201176405, + 0.009966964833438396, + 0.005646867211908102, + 0.00792256835848093, + 0.012981422245502472, + 0.008400266058743, + 0.006257759407162666, + 0.005044461227953434, + 0.0038197722751647234, + 0.01072488259524107, + 0.005434996448457241, + 0.004104947671294212, + 0.011621573939919472, + 0.006955096963793039, + 0.010430187918245792, + 0.008904858492314816, + 0.004901462234556675, + 0.009648227132856846, + 0.006950049661099911, + 0.01728430762887001, + 0.010837399400770664, + 0.010177905671298504, + 0.005699002183973789, + 0.00572986202314496, + 0.004893658682703972, + 0.01050519198179245, + 0.0059258281253278255, + 0.005239000543951988, + 0.011695419438183308, + 0.004277838859707117, + 0.01761765219271183, + 0.014046208001673222, + 0.02984764240682125, + 0.018922127783298492, + 0.018465673550963402, + 0.009962163865566254, + 0.013814374804496765, + 0.018807020038366318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.004108797758817673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021984234917908907 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.0021984234917908907, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021984234917908907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13575094938278198, + 0.04607374966144562, + 0.06961575150489807, + 0.05295850336551666, + 0.039670541882514954, + 0.1358097940683365, + 0.07687866687774658, + 0.10012960433959961, + 0.09967869520187378, + 0.24343377351760864 + ], + "credit_map": [ + 0.015979060903191566, + 0.023859208449721336, + 0.0130893150344491, + 0.017908137291669846, + 0.01258843019604683, + 0.019005198031663895, + 0.014484315179288387, + 0.018837295472621918, + 0.00484475726261735, + 0.0036858501844108105, + 0.0045110671781003475, + 0.009104343131184578, + 0.004583586473017931, + 0.005122898146510124, + 0.00700295390561223, + 0.0072182887233793736, + 0.0043149976991117, + 0.00548555376008153, + 0.013468118384480476, + 0.013794999569654465, + 0.0069966623559594154, + 0.004500099923461676, + 0.01293946336954832, + 0.008115861564874649, + 0.004054300487041473, + 0.002849094569683075, + 0.010487324558198452, + 0.0058115278370678425, + 0.0048868171870708466, + 0.002690510591492057, + 0.013708634302020073, + 0.008470294065773487, + 0.0026885687839239836, + 0.002556626684963703, + 0.004428195767104626, + 0.008344716392457485, + 0.0035757573787122965, + 0.004035817924886942, + 0.007720215246081352, + 0.006320645567029715, + 0.01832454837858677, + 0.01134229451417923, + 0.007984135299921036, + 0.024631960317492485, + 0.010882350616157055, + 0.008193940855562687, + 0.036953408271074295, + 0.01749715767800808, + 0.007986274547874928, + 0.005942511837929487, + 0.006454295013099909, + 0.012912151403725147, + 0.007348760962486267, + 0.011029626242816448, + 0.013304958119988441, + 0.011900089681148529, + 0.016900140792131424, + 0.011080588214099407, + 0.00804166216403246, + 0.01414981298148632, + 0.008775237947702408, + 0.01555304229259491, + 0.013854761607944965, + 0.011774357408285141, + 0.007309133652597666, + 0.010686422698199749, + 0.008533413521945477, + 0.018846139311790466, + 0.01229916326701641, + 0.009291623719036579, + 0.024196483194828033, + 0.008516316302120686, + 0.029643626883625984, + 0.012440107762813568, + 0.07083770632743835, + 0.04605906456708908, + 0.014060428366065025, + 0.01609581895172596, + 0.024568775668740273, + 0.029728248715400696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0021984234917908907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000058889389038 + ], + "opposite_logprob_tokens": [ + -2.695000648498535 + ], + "logprob_reward": -0.07000058889389038, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000058889389038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059005044400691986, + 0.053111106157302856, + 0.08322950452566147, + 0.07997935265302658, + 0.03285271301865578, + 0.09654519706964493, + 0.10298367589712143, + 0.026892365887761116, + 0.22959661483764648, + 0.23580442368984222 + ], + "credit_map": [ + 0.00741998665034771, + 0.008043983951210976, + 0.006445831619203091, + 0.0073502324521541595, + 0.008718105964362621, + 0.008614996448159218, + 0.006659383419901133, + 0.005752528086304665, + 0.006607145071029663, + 0.00994227547198534, + 0.005271860398352146, + 0.006383048370480537, + 0.00463251443579793, + 0.007595315109938383, + 0.005691241938620806, + 0.006987704895436764, + 0.007139421068131924, + 0.005638291127979755, + 0.015687929466366768, + 0.013962789438664913, + 0.00727499183267355, + 0.008190745487809181, + 0.014737291261553764, + 0.010598042979836464, + 0.007146752905100584, + 0.005534610245376825, + 0.020129293203353882, + 0.009654096327722073, + 0.005462640430778265, + 0.00580360135063529, + 0.01251105684787035, + 0.013737300410866737, + 0.002700736979022622, + 0.0024297304917126894, + 0.004392119124531746, + 0.006957517936825752, + 0.00278897350654006, + 0.003694141050800681, + 0.004705025348812342, + 0.005184467416256666, + 0.006147927138954401, + 0.006671249400824308, + 0.004715932998806238, + 0.020563675090670586, + 0.007565314881503582, + 0.0071545871905982494, + 0.028391487896442413, + 0.0153350206092, + 0.012885300442576408, + 0.00969664752483368, + 0.006541108246892691, + 0.023538723587989807, + 0.007115747779607773, + 0.008063490502536297, + 0.022232308983802795, + 0.01291035208851099, + 0.0047051385045051575, + 0.0022589974105358124, + 0.0027959258295595646, + 0.00322986813262105, + 0.002336689969524741, + 0.005061933770775795, + 0.0023958056699484587, + 0.004108006600290537, + 0.012890003621578217, + 0.017056405544281006, + 0.015910528600215912, + 0.036037132143974304, + 0.021765856072306633, + 0.041723694652318954, + 0.04102284088730812, + 0.04319014772772789, + 0.029647141695022583, + 0.021163444966077805, + 0.045350030064582825, + 0.04100814834237099, + 0.024670036509633064, + 0.022002970799803734, + 0.023280607536435127, + 0.028682051226496696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.07000058889389038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up time-lapse shot, the camera held still, focused at eye level on a square glass container placed against a stark black background. A concentrated, cool white overhead light shines vertically, precisely illuminating the center of the container, creating a dramatic, theatrical effect. The container is filled with a clear, yet slightly viscous, unknown liquid. A vibrant, crimson apple, completely submerged, its skin smooth with a natural waxy sheen, its stem still emerald green. As the time-lapse begins, fine cracks begin to appear in the apple's vibrant red skin, and tiny bubbles emerge. Strands of crimson pigment, like ink, cautiously seep through the cracks, slowly diffusing through the still liquid. Then, the dissolution process accelerates rapidly and decisively, and the apple's flesh begins to disintegrate into countless tiny particles and fibers, its solid form gradually becoming blurred and soft. Finally, the entire apple completely disintegrates, and under the penetration of top light, it turns into a pool of brilliant colors that steadily and calmly rotates and mixes in the liquid. Deep red, soft milky white and light yellow are intertwined to form an ever-changing, flowing abstract painting. The whole process has both the precise beauty of scientific experiments and a surreal and charming artistic atmosphere.", + "seed": 32994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_5792_seed_32994_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022925118496641517, + "vlm_logprob": -0.022925118496641517, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.022925118496641517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05293635278940201 + ], + "opposite_logprob_tokens": [ + -3.552936315536499 + ], + "logprob_reward": -0.05293635278940201, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.05293635278940201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20623494684696198, + 0.11334040015935898, + 0.028877226635813713, + 0.03213449940085411, + 0.03660158812999725, + 0.029977308586239815, + 0.19407740235328674, + 0.03406462445855141, + 0.03684305399656296, + 0.28784888982772827 + ], + "credit_map": [ + 0.02077001892030239, + 0.01448976993560791, + 0.07629703730344772, + 0.013582693412899971, + 0.010600892826914787, + 0.03896590694785118, + 0.00885860901325941, + 0.022670023143291473, + 0.009872308932244778, + 0.012928750365972519, + 0.010840432718396187, + 0.00809941254556179, + 0.006832476239651442, + 0.04478824883699417, + 0.007525954861193895, + 0.01245281845331192, + 0.0031891141552478075, + 0.0036176908761262894, + 0.004237737040966749, + 0.0042937095277011395, + 0.0029364184010773897, + 0.0019171575549989939, + 0.00544534670189023, + 0.0032400530762970448, + 0.0043489676900208, + 0.0035174021031707525, + 0.0036097022239118814, + 0.004801017697900534, + 0.0026619848795235157, + 0.0032286352943629026, + 0.004408031702041626, + 0.00555875850841403, + 0.003633431624621153, + 0.0032576648518443108, + 0.003108270699158311, + 0.008283555507659912, + 0.003447911236435175, + 0.007250726688653231, + 0.004625803790986538, + 0.002994225360453129, + 0.003113914281129837, + 0.00265593733638525, + 0.0026984948199242353, + 0.006435655988752842, + 0.0037600870709866285, + 0.004398360848426819, + 0.0040399618446826935, + 0.0028748975601047277, + 0.019728338345885277, + 0.008849026635289192, + 0.009785025380551815, + 0.02815520390868187, + 0.02407926134765148, + 0.043509989976882935, + 0.047238271683454514, + 0.012732291594147682, + 0.003027780679985881, + 0.004733190406113863, + 0.003136640414595604, + 0.00499968696385622, + 0.0033671853598207235, + 0.003012845991179347, + 0.0075059509836137295, + 0.004281342960894108, + 0.00385723146609962, + 0.0038047253619879484, + 0.0021812436170876026, + 0.007573690265417099, + 0.007943638600409031, + 0.0038276324048638344, + 0.0038922561798244715, + 0.003762637497857213, + 0.00847584567964077, + 0.012536794878542423, + 0.007593046873807907, + 0.11175072193145752, + 0.06824754923582077, + 0.013148822821676731, + 0.05814579874277115, + 0.00795032363384962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.05293635278940201 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764752272516489 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.00764752272516489, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764752272516489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32907500863075256, + 0.19741640985012054, + 0.06862001866102219, + 0.034353017807006836, + 0.032997503876686096, + 0.08520669490098953, + 0.14426463842391968, + 0.021552639082074165, + 0.018038000911474228, + 0.06847609579563141 + ], + "credit_map": [ + 0.03627723827958107, + 0.022486552596092224, + 0.1190042793750763, + 0.01969400979578495, + 0.01820063404738903, + 0.06793279200792313, + 0.012366718612611294, + 0.033112794160842896, + 0.02136257477104664, + 0.022944550961256027, + 0.013482165522873402, + 0.010943963192403316, + 0.01116210874170065, + 0.07957185804843903, + 0.010437876917421818, + 0.02751130424439907, + 0.006599558517336845, + 0.01002624910324812, + 0.012162607163190842, + 0.005089898128062487, + 0.004350312054157257, + 0.019508812576532364, + 0.0038505331613123417, + 0.007032050751149654, + 0.004341368097811937, + 0.004120286554098129, + 0.006799466907978058, + 0.004158800933510065, + 0.003518151119351387, + 0.004314899910241365, + 0.003349683713167906, + 0.003750361269339919, + 0.0033131770323961973, + 0.0018512511160224676, + 0.0026284712366759777, + 0.007759015075862408, + 0.004083437845110893, + 0.007905291393399239, + 0.0026605422608554363, + 0.0027963207103312016, + 0.010820051655173302, + 0.006648371461778879, + 0.0067345090210437775, + 0.01911778934299946, + 0.0068281530402600765, + 0.0142173171043396, + 0.010119770653545856, + 0.010720734484493732, + 0.012511545792222023, + 0.004811533726751804, + 0.006188860163092613, + 0.01858055405318737, + 0.0234404094517231, + 0.04748069494962692, + 0.020943425595760345, + 0.01030760072171688, + 0.002125304425135255, + 0.002045842818915844, + 0.0024812568444758654, + 0.0026511542964726686, + 0.002111238893121481, + 0.0023575047962367535, + 0.0035871549043804407, + 0.00419318163767457, + 0.0015462280716747046, + 0.0018413362558931112, + 0.0012564481003209949, + 0.004007912706583738, + 0.003909285645931959, + 0.001463213819079101, + 0.002366533037275076, + 0.0016470429254695773, + 0.003049213206395507, + 0.002502687741070986, + 0.0024398891255259514, + 0.028202898800373077, + 0.010031186044216156, + 0.0024324075784534216, + 0.01723625883460045, + 0.0025815542321652174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.00764752272516489 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003205401822924614 + ], + "opposite_logprob_tokens": [ + -5.753205299377441 + ], + "logprob_reward": -0.003205401822924614, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003205401822924614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4739803671836853, + 0.06260085105895996, + 0.027414517477154732, + 0.049420762807130814, + 0.03776541352272034, + 0.01907220296561718, + 0.18548688292503357, + 0.03466290235519409, + 0.02825992740690708, + 0.08133621513843536 + ], + "credit_map": [ + 0.0478394590318203, + 0.03019370138645172, + 0.17253196239471436, + 0.023894064128398895, + 0.023328980430960655, + 0.10940783470869064, + 0.01669127680361271, + 0.05009305104613304, + 0.007273850962519646, + 0.0049752723425626755, + 0.015531501732766628, + 0.005990459118038416, + 0.003772446885704994, + 0.007081025280058384, + 0.0041894251480698586, + 0.013786859810352325, + 0.0027475585229694843, + 0.0032089559826999903, + 0.002938053337857127, + 0.0026017464697360992, + 0.0028616159688681364, + 0.007568778470158577, + 0.0027722036466002464, + 0.0027156041469424963, + 0.0033812846522778273, + 0.004304039757698774, + 0.005249375943094492, + 0.00742481742054224, + 0.0050415457226336, + 0.00967426411807537, + 0.00801389291882515, + 0.006331538315862417, + 0.004504791926592588, + 0.0029714242555201054, + 0.0045708538964390755, + 0.007194747216999531, + 0.00390928890556097, + 0.005377772729843855, + 0.003969070967286825, + 0.005267460364848375, + 0.0016814564587548375, + 0.0014185189502313733, + 0.0013085597893223166, + 0.005114458967000246, + 0.0015379780670627952, + 0.0036781285889446735, + 0.0025654654018580914, + 0.001767635578289628, + 0.014268533326685429, + 0.006973369512706995, + 0.006347555201500654, + 0.021206537261605263, + 0.042067330330610275, + 0.06153470277786255, + 0.020464932546019554, + 0.01262391172349453, + 0.0032008697744458914, + 0.003300076350569725, + 0.0038122939877212048, + 0.004372145049273968, + 0.004315327387303114, + 0.0031304629519581795, + 0.006071196403354406, + 0.006460527423769236, + 0.002350802067667246, + 0.002269785851240158, + 0.001422958099283278, + 0.0074445707723498344, + 0.0070083774626255035, + 0.002304896479472518, + 0.003249543020501733, + 0.0022089905105531216, + 0.0034320137929171324, + 0.0040329694747924805, + 0.0026358517352491617, + 0.029728343710303307, + 0.020809708163142204, + 0.0039007957093417645, + 0.014307745732367039, + 0.002488784259185195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.003205401822924614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021991373505443335 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.0021991373505443335, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021991373505443335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11008412390947342, + 0.04945886880159378, + 0.08294588327407837, + 0.07462869584560394, + 0.08380110561847687, + 0.03818314149975777, + 0.3220180571079254, + 0.07309700548648834, + 0.07305894047021866, + 0.09272420406341553 + ], + "credit_map": [ + 0.008914160542190075, + 0.014126438647508621, + 0.022854547947645187, + 0.010424510575830936, + 0.017634039744734764, + 0.016675734892487526, + 0.009175539948046207, + 0.010279153473675251, + 0.005788004957139492, + 0.006280676927417517, + 0.007498537190258503, + 0.007255217060446739, + 0.007342120166867971, + 0.0036533018574118614, + 0.007288900204002857, + 0.004352109972387552, + 0.008065583184361458, + 0.008271969854831696, + 0.006095126271247864, + 0.004610374104231596, + 0.005568400956690311, + 0.03337106853723526, + 0.006086051929742098, + 0.010877312161028385, + 0.006141332909464836, + 0.009118613786995411, + 0.01094654481858015, + 0.009576969780027866, + 0.006348281167447567, + 0.014154547825455666, + 0.0095596294850111, + 0.008782771416008472, + 0.00590229406952858, + 0.007490301504731178, + 0.005415697116404772, + 0.01705402508378029, + 0.008394075557589531, + 0.021236518397927284, + 0.010229828767478466, + 0.008078360930085182, + 0.003359916852787137, + 0.003660796443000436, + 0.0032644716557115316, + 0.006728234700858593, + 0.004626431968063116, + 0.004449016880244017, + 0.0071901301853358746, + 0.0049041444435715675, + 0.029697826132178307, + 0.011715455912053585, + 0.013676469214260578, + 0.04390082508325577, + 0.04235156998038292, + 0.10468558222055435, + 0.0526767298579216, + 0.0233136098831892, + 0.007556717377156019, + 0.0065418025478720665, + 0.008364424109458923, + 0.010135583579540253, + 0.008404330350458622, + 0.007117380853742361, + 0.011466710828244686, + 0.013510060496628284, + 0.003714436199516058, + 0.005217195488512516, + 0.003197962883859873, + 0.021665431559085846, + 0.019080311059951782, + 0.005793723743408918, + 0.010273060761392117, + 0.004116820637136698, + 0.0057348161935806274, + 0.004806992597877979, + 0.0037813016679137945, + 0.029262283816933632, + 0.025234246626496315, + 0.005986018572002649, + 0.013759944587945938, + 0.00415859604254365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.0021991373505443335 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863717779517174 + ], + "opposite_logprob_tokens": [ + -3.0486371517181396 + ], + "logprob_reward": -0.04863717779517174, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863717779517174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12410515546798706, + 0.06181703135371208, + 0.02225952409207821, + 0.04157542437314987, + 0.13133716583251953, + 0.09019111096858978, + 0.04231293871998787, + 0.08764196932315826, + 0.053098857402801514, + 0.3456607758998871 + ], + "credit_map": [ + 0.008732291869819164, + 0.022006703540682793, + 0.012309140525758266, + 0.01546289213001728, + 0.0194882582873106, + 0.026032978668808937, + 0.011506611481308937, + 0.00856628455221653, + 0.006669819355010986, + 0.005994265433400869, + 0.015369159169495106, + 0.006358429789543152, + 0.004336186684668064, + 0.012066137045621872, + 0.005289631430059671, + 0.005733400583267212, + 0.0026387430261820555, + 0.002269686898216605, + 0.003939143382012844, + 0.002892429241910577, + 0.0020367091055959463, + 0.0024077314883470535, + 0.0038064736872911453, + 0.002268606796860695, + 0.004222836811095476, + 0.004228850360959768, + 0.00926938559859991, + 0.004937610123306513, + 0.0029614411760121584, + 0.007526008412241936, + 0.005042428150773048, + 0.003386865137144923, + 0.009869209490716457, + 0.007074886932969093, + 0.006993885152041912, + 0.03998224809765816, + 0.007083304226398468, + 0.037174955010414124, + 0.013603187166154385, + 0.009555491618812084, + 0.006310964003205299, + 0.005118736997246742, + 0.005449717398732901, + 0.03154889866709709, + 0.004455080721527338, + 0.02054416388273239, + 0.010804838500916958, + 0.005958710331469774, + 0.0035583481658250093, + 0.004601503722369671, + 0.003263648832216859, + 0.006182102020829916, + 0.008811700157821178, + 0.004385908599942923, + 0.006477978080511093, + 0.005031748209148645, + 0.007426489144563675, + 0.013048755936324596, + 0.009588734246790409, + 0.011925378814339638, + 0.011027957312762737, + 0.00483003631234169, + 0.020553074777126312, + 0.009241544641554356, + 0.004424318205565214, + 0.003692213213071227, + 0.002889930037781596, + 0.014922558329999447, + 0.012550006620585918, + 0.004538736771792173, + 0.006358474027365446, + 0.0037226194981485605, + 0.018931519240140915, + 0.014284339733421803, + 0.011987931095063686, + 0.13483306765556335, + 0.06854157149791718, + 0.011679253540933132, + 0.07507473230361938, + 0.010328350588679314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.04863717779517174 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up time-lapse shot, the camera held still, focused at eye level on a square glass container placed against a stark black background. A concentrated, cool white overhead light shines vertically, precisely illuminating the center of the container, creating a dramatic, theatrical effect. The container is filled with a clear, yet slightly viscous, unknown liquid. A vibrant, crimson apple, completely submerged, its skin smooth with a natural waxy sheen, its stem still emerald green. As the time-lapse begins, fine cracks begin to appear in the apple's vibrant red skin, and tiny bubbles emerge. Strands of crimson pigment, like ink, cautiously seep through the cracks, slowly diffusing through the still liquid. Then, the dissolution process accelerates rapidly and decisively, and the apple's flesh begins to disintegrate into countless tiny particles and fibers, its solid form gradually becoming blurred and soft. Finally, the entire apple completely disintegrates, and under the penetration of top light, it turns into a pool of brilliant colors that steadily and calmly rotates and mixes in the liquid. Deep red, soft milky white and light yellow are intertwined to form an ever-changing, flowing abstract painting. The whole process has both the precise beauty of scientific experiments and a surreal and charming artistic atmosphere.", + "seed": 32995, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_13/index_5792_seed_32995_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.044559275778010485, + "vlm_logprob": -0.044559275778010485, + "vlm_yes_no_margin": 4.149999976158142, + "vlm_token_credit": -0.044559275778010485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0597403421998024 + ], + "opposite_logprob_tokens": [ + -3.4347403049468994 + ], + "logprob_reward": -0.0597403421998024, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0597403421998024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038557104766368866, + 0.01842426136136055, + 0.0815984308719635, + 0.059146951884031296, + 0.6023886203765869, + 0.0768134593963623, + 0.025667468085885048, + 0.04290122911334038, + 0.030301878228783607, + 0.024200506508350372 + ], + "credit_map": [ + 0.0035932306200265884, + 0.0037226255517452955, + 0.005093461833894253, + 0.006939850281924009, + 0.0037814409006386995, + 0.006144373212009668, + 0.005227976944297552, + 0.004054146818816662, + 0.001843011355958879, + 0.0014710893156006932, + 0.002687043510377407, + 0.0034132509026676416, + 0.0015131270047277212, + 0.0027568440418690443, + 0.002929446054622531, + 0.001810449524782598, + 0.004144910257309675, + 0.003380931681022048, + 0.004090405069291592, + 0.017466528341174126, + 0.004473731853067875, + 0.006224866956472397, + 0.028698110952973366, + 0.013118951581418514, + 0.0032703066244721413, + 0.0077857221476733685, + 0.0027733284514397383, + 0.005255991127341986, + 0.021561795845627785, + 0.0039013433270156384, + 0.011197158135473728, + 0.003401310183107853, + 0.0249332245439291, + 0.07891446352005005, + 0.01885811612010002, + 0.20681533217430115, + 0.05363725870847702, + 0.03237513452768326, + 0.15828318893909454, + 0.0285719595849514, + 0.006653343793004751, + 0.01222986914217472, + 0.010703128762543201, + 0.011004328727722168, + 0.013346457853913307, + 0.007853705435991287, + 0.00930096860975027, + 0.005721655208617449, + 0.0025321145076304674, + 0.0035409426782280207, + 0.003087362041696906, + 0.0033171062823385, + 0.0035204447340220213, + 0.0017121845157817006, + 0.005773055367171764, + 0.002184260869398713, + 0.0037830458022654057, + 0.004507690668106079, + 0.004974262788891792, + 0.009294139221310616, + 0.005292416084557772, + 0.003838494187220931, + 0.008013401180505753, + 0.003197786631062627, + 0.0032522461842745543, + 0.0034971849527209997, + 0.002909503411501646, + 0.004392541013658047, + 0.004106623120605946, + 0.0030985279008746147, + 0.006528937257826328, + 0.002516314620152116, + 0.0026330247055739164, + 0.0021401531994342804, + 0.002711291192099452, + 0.004713313188403845, + 0.0034554044250398874, + 0.002483380027115345, + 0.003704425413161516, + 0.002359516452997923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.0597403421998024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710149586200714 + ], + "opposite_logprob_tokens": [ + -2.127101421356201 + ], + "logprob_reward": -0.12710149586200714, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12710149586200714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10928498208522797, + 0.13017097115516663, + 0.015436114743351936, + 0.12417338043451309, + 0.434086412191391, + 0.06919790804386139, + 0.019748540595173836, + 0.02209913358092308, + 0.02859458141028881, + 0.04720798134803772 + ], + "credit_map": [ + 0.009028639644384384, + 0.005597778130322695, + 0.011215686798095703, + 0.017225658521056175, + 0.006852168124169111, + 0.01565501280128956, + 0.027645153924822807, + 0.016064884141087532, + 0.008374437689781189, + 0.005577776115387678, + 0.011930610053241253, + 0.022888299077749252, + 0.006621067877858877, + 0.013994822278618813, + 0.039561811834573746, + 0.021222153678536415, + 0.0009521498577669263, + 0.001340196467936039, + 0.0015991894761100411, + 0.0025929557159543037, + 0.001564342062920332, + 0.002164986217394471, + 0.003081388073042035, + 0.002140906872227788, + 0.0077321710996329784, + 0.006342079024761915, + 0.008772441186010838, + 0.026886552572250366, + 0.007277354598045349, + 0.008027683943510056, + 0.03531523793935776, + 0.02381986565887928, + 0.018666740506887436, + 0.061611827462911606, + 0.014297407120466232, + 0.14348338544368744, + 0.041464753448963165, + 0.024262959137558937, + 0.1100030466914177, + 0.020296305418014526, + 0.0062652169726789, + 0.011877394281327724, + 0.006637919694185257, + 0.00956603605300188, + 0.010398700833320618, + 0.010986697860062122, + 0.006831304170191288, + 0.006634638179093599, + 0.0019580647349357605, + 0.002314922632649541, + 0.003259767312556505, + 0.00213264930061996, + 0.0030989162623882294, + 0.0012212253641337156, + 0.003535635070875287, + 0.0022273592185229063, + 0.0023502809926867485, + 0.0017646508058533072, + 0.0029570478945970535, + 0.004001650959253311, + 0.0036011952906847, + 0.0020296694710850716, + 0.0036454421933740377, + 0.0017491960898041725, + 0.002819787710905075, + 0.0033502455335110426, + 0.0033878120593726635, + 0.004398726392537355, + 0.0033823708072304726, + 0.002601191634312272, + 0.005639594979584217, + 0.0030148543883115053, + 0.006447192747145891, + 0.003956606611609459, + 0.008599476888775826, + 0.009130501188337803, + 0.007389018312096596, + 0.00363086792640388, + 0.004988870117813349, + 0.003065448487177491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.12710149586200714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004652745556086302 + ], + "opposite_logprob_tokens": [ + -5.379652976989746 + ], + "logprob_reward": -0.004652745556086302, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004652745556086302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09897004812955856, + 0.05493738874793053, + 0.08301534503698349, + 0.1553489863872528, + 0.29730096459388733, + 0.09670880436897278, + 0.06506561487913132, + 0.08821844309568405, + 0.04030768200755119, + 0.0201267097145319 + ], + "credit_map": [ + 0.006375808734446764, + 0.008226148784160614, + 0.010645945556461811, + 0.01704171486198902, + 0.009714341722428799, + 0.01260701846331358, + 0.015264083631336689, + 0.019094986841082573, + 0.0037115260493010283, + 0.004321471322327852, + 0.005178807303309441, + 0.010982011444866657, + 0.004420243203639984, + 0.005441196728497744, + 0.011016204953193665, + 0.009865925647318363, + 0.0069274213165044785, + 0.005696156527847052, + 0.012176185846328735, + 0.013014699332416058, + 0.0063673462718725204, + 0.014280742034316063, + 0.013189852237701416, + 0.011362941935658455, + 0.007878911681473255, + 0.009712422266602516, + 0.013044427148997784, + 0.029454387724399567, + 0.015015813522040844, + 0.01285422220826149, + 0.03505474701523781, + 0.03233405947685242, + 0.015049651265144348, + 0.05001107230782509, + 0.013696812093257904, + 0.08303411304950714, + 0.03437430039048195, + 0.02042583003640175, + 0.0659005343914032, + 0.014808635227382183, + 0.006137076299637556, + 0.01381103415042162, + 0.01084226742386818, + 0.01793798804283142, + 0.013565381057560444, + 0.007761969231069088, + 0.017883693799376488, + 0.00876939669251442, + 0.004647205583751202, + 0.007656485307961702, + 0.005581552628427744, + 0.008958055637776852, + 0.009965262375772, + 0.004023008048534393, + 0.018027564510703087, + 0.006206480786204338, + 0.006659317761659622, + 0.008931231684982777, + 0.009937116876244545, + 0.018088597804307938, + 0.010454618372023106, + 0.009734687395393848, + 0.01687161810696125, + 0.0075412532314658165, + 0.002282504690811038, + 0.0037411474622786045, + 0.003237019293010235, + 0.005262596067041159, + 0.005138160195201635, + 0.005108356475830078, + 0.011911480687558651, + 0.0036264178343117237, + 0.002422559540718794, + 0.0021176517475396395, + 0.003003377467393875, + 0.0032448535785079002, + 0.0026082652620971203, + 0.0016924689989537, + 0.0035873150918632746, + 0.0014502174453809857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.004652745556086302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015140033792704344 + ], + "opposite_logprob_tokens": [ + -6.501513957977295 + ], + "logprob_reward": -0.0015140033792704344, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015140033792704344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03618362173438072, + 0.06793766468763351, + 0.08843179792165756, + 0.150979146361351, + 0.42082637548446655, + 0.0923147201538086, + 0.04181741178035736, + 0.03762848675251007, + 0.03953275457024574, + 0.024347972124814987 + ], + "credit_map": [ + 0.0031716860830783844, + 0.0022177232895046473, + 0.00438642269000411, + 0.004539639223366976, + 0.002393968403339386, + 0.00599167263135314, + 0.007289073895663023, + 0.0061934334225952625, + 0.004300360102206469, + 0.0032488061115145683, + 0.006727845873683691, + 0.011321079917252064, + 0.0036219169851392508, + 0.007018727716058493, + 0.019355036318302155, + 0.012343894690275192, + 0.0047269053757190704, + 0.003948708530515432, + 0.003727033268660307, + 0.020350195467472076, + 0.005130477249622345, + 0.004514788743108511, + 0.03225195035338402, + 0.013781735673546791, + 0.009485328570008278, + 0.017413675785064697, + 0.005899210460484028, + 0.014673581346869469, + 0.05071096494793892, + 0.007810670882463455, + 0.03645831346511841, + 0.008527413010597229, + 0.014723303727805614, + 0.0475001186132431, + 0.011472073383629322, + 0.1557798981666565, + 0.03318379074335098, + 0.018038228154182434, + 0.12226584553718567, + 0.017863137647509575, + 0.007451660465449095, + 0.017472269013524055, + 0.008626974187791348, + 0.015004212968051434, + 0.01430986262857914, + 0.010503388941287994, + 0.012325040064752102, + 0.006621317006647587, + 0.002940016333013773, + 0.004527892917394638, + 0.003268474480137229, + 0.005237395875155926, + 0.007151505444198847, + 0.0028514761943370104, + 0.012034757994115353, + 0.0038058955688029528, + 0.0030403509736061096, + 0.004239978734403849, + 0.0037080354522913694, + 0.009055184200406075, + 0.004287841264158487, + 0.0031331079080700874, + 0.007074284367263317, + 0.003089709673076868, + 0.0028448032680898905, + 0.004996253177523613, + 0.0023847699631005526, + 0.005769303534179926, + 0.00547029310837388, + 0.004367719870060682, + 0.010959798470139503, + 0.0027398152742534876, + 0.0022478667087852955, + 0.0019025570945814252, + 0.002549490425735712, + 0.0052715796045959, + 0.003978554625064135, + 0.002278645057231188, + 0.004198255017399788, + 0.0019210268510505557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.0015140033792704344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse with a stationary camera held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a square glass container filled with a clear, slightly viscous liquid placed against a stark black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant crimson apple completely submerged in the liquid with a smooth, waxy sheen and an emerald green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple disintegrate into particles and fibers, releasing crimson pigment that diffuses into the liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final result show the liquid turning into a rotating pool of deep red, soft milky white, and light yellow colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787791892886162 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.029787791892886162, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787791892886162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03758466988801956, + 0.05212559178471565, + 0.06230523809790611, + 0.14863181114196777, + 0.20569567382335663, + 0.14992402493953705, + 0.09762582182884216, + 0.053034622222185135, + 0.07464341819286346, + 0.11842907965183258 + ], + "credit_map": [ + 0.00371248135343194, + 0.004011484794318676, + 0.004429066553711891, + 0.007582793012261391, + 0.004593509715050459, + 0.004346038214862347, + 0.0062982914969325066, + 0.0026110054459422827, + 0.004117368720471859, + 0.0034501459449529648, + 0.005837786942720413, + 0.009130217134952545, + 0.0039330534636974335, + 0.006078597158193588, + 0.01275679562240839, + 0.006821624934673309, + 0.0034880773164331913, + 0.0036529244389384985, + 0.0051939901895821095, + 0.015031482093036175, + 0.003519656602293253, + 0.003878450719639659, + 0.02177692949771881, + 0.005763727705925703, + 0.005804027430713177, + 0.01991712488234043, + 0.003970024641603231, + 0.0132847148925066, + 0.0569523386657238, + 0.007886805571615696, + 0.032725948840379715, + 0.008090835064649582, + 0.013673798181116581, + 0.0446489043533802, + 0.010642393492162228, + 0.039380814880132675, + 0.030460640788078308, + 0.01820872165262699, + 0.03467455878853798, + 0.014005822129547596, + 0.01357330847531557, + 0.024905234575271606, + 0.018708467483520508, + 0.020889561623334885, + 0.030197972431778908, + 0.013428635895252228, + 0.017714453861117363, + 0.010506391525268555, + 0.008022880181670189, + 0.012107054702937603, + 0.006369057577103376, + 0.01573799178004265, + 0.013461556285619736, + 0.006690269336104393, + 0.02650696225464344, + 0.00873005110770464, + 0.005363731179386377, + 0.0066491649486124516, + 0.004628995433449745, + 0.008079548366367817, + 0.0076951137743890285, + 0.005879363510757685, + 0.009899121709167957, + 0.004839584231376648, + 0.0038590540643781424, + 0.006596646271646023, + 0.00459753954783082, + 0.009118917398154736, + 0.010586782358586788, + 0.010936160571873188, + 0.021291151642799377, + 0.007657165639102459, + 0.005973042454570532, + 0.016815802082419395, + 0.0072894105687737465, + 0.025197133421897888, + 0.022401422262191772, + 0.010221813805401325, + 0.02155926451086998, + 0.008971195667982101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.029787791892886162 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_14/index_tensor([724, 724])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_14/index_tensor([724, 724])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5c66ce785c11b6eb7351dd4ea92e0d3ab873ce69 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_14/index_tensor([724, 724])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, arcing, panoramic shot begins at a slightly lower angle, sweeping across the grand, cavernous hall of a Gothic stone building, capturing a pivotal moment rendered in classic Dungeons & Dragons art. The scene, defined by heroic poses and dramatic black-and-white lighting, centers on a wise, bearded headmaster and a young wizard. The headmaster, with his flowing white beard falling to his chest and deep, all-seeing eyes, wears ornate dark blue robes embroidered with silver constellations, pointing his finger as he makes his point. Beside him stands the young wizard, recognizable by his tousled black hair, round glasses, and the faint, signature lightning bolt scar on his forehead. He stands there, his expression serious. They are explaining something of great importance to the diverse group of students gathered before them. As the camera continues its smooth, circular path, it reveals the students' resolute faces, illuminated by the warm, flickering glow of hundreds of floating candles, which drift toward the vaulted, shadowed ceiling. On the towering stone walls, portraits of ancient figures subtly move and whisper within their frames. The students, dressed in dark robes of varying colors, listened intently, their bodies in powerful poses of concentration—a girl with tightly braided hair, a boy beside her with a clenched fist, their expressions reflecting a shared solemn purpose in a play of light and shadow, full of both heroism and tension.", + "seed": 29306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_14/index_724_seed_29306_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3607651602476835, + "vlm_logprob": -0.3607651602476835, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.3607651602476835, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024681424722075462 + ], + "opposite_logprob_tokens": [ + -3.899681329727173 + ], + "logprob_reward": -0.024681424722075462, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024681424722075462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12151804566383362, + 0.07529384642839432, + 0.05204705521464348, + 0.046044912189245224, + 0.05377959832549095, + 0.08636795729398727, + 0.37470805644989014, + 0.026132356375455856, + 0.034045878797769547, + 0.13006234169006348 + ], + "credit_map": [ + 0.027465833351016045, + 0.01284182071685791, + 0.01426464319229126, + 0.010816079564392567, + 0.010976958088576794, + 0.019066505134105682, + 0.008336169645190239, + 0.017750030383467674, + 0.013687544502317905, + 0.009971382096409798, + 0.010322684422135353, + 0.009177223779261112, + 0.005132501013576984, + 0.009063592180609703, + 0.0059422594495117664, + 0.01199666503816843, + 0.006197692360728979, + 0.007222451735287905, + 0.006488731596618891, + 0.005035606678575277, + 0.005772764328867197, + 0.009385538287460804, + 0.004292065277695656, + 0.007652205880731344, + 0.004587643314152956, + 0.00495900260284543, + 0.009081580676138401, + 0.009036372415721416, + 0.003928256221115589, + 0.006833554711192846, + 0.003200351493433118, + 0.004418148659169674, + 0.0062451353296637535, + 0.003930820617824793, + 0.00755607383325696, + 0.006882278248667717, + 0.004750965163111687, + 0.006866848561912775, + 0.005479099228978157, + 0.012068375013768673, + 0.013803085312247276, + 0.00883153360337019, + 0.013728131540119648, + 0.00761303398758173, + 0.0065891980193555355, + 0.016025705263018608, + 0.006811581086367369, + 0.01296568475663662, + 0.03889787197113037, + 0.035765066742897034, + 0.05680808424949646, + 0.03288239985704422, + 0.04614230990409851, + 0.07371608912944794, + 0.028743349015712738, + 0.06175287440419197, + 0.0028330294881016016, + 0.004594502504914999, + 0.004767213482409716, + 0.002627403475344181, + 0.0030777950305491686, + 0.0027295839972794056, + 0.0023871425073593855, + 0.003115688916295767, + 0.004301420878618956, + 0.002515332540497184, + 0.005420481786131859, + 0.0028077277820557356, + 0.0037065865471959114, + 0.007127788383513689, + 0.003808793844655156, + 0.004357749596238136, + 0.014814778231084347, + 0.009302916936576366, + 0.01650121621787548, + 0.007938964292407036, + 0.011166936717927456, + 0.03842687979340553, + 0.006271027494221926, + 0.025639608502388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.024681424722075462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035862773656845 + ], + "opposite_logprob_tokens": [ + -2.350358724594116 + ], + "logprob_reward": -0.10035862773656845, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035862773656845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05640162155032158, + 0.04458310082554817, + 0.014629494398832321, + 0.035561684519052505, + 0.0236138217151165, + 0.038422755897045135, + 0.061924923211336136, + 0.013438187539577484, + 0.020926596596837044, + 0.6904978156089783 + ], + "credit_map": [ + 0.00905447918921709, + 0.007089400663971901, + 0.008830578997731209, + 0.006266272161155939, + 0.0065910229459404945, + 0.007220023311674595, + 0.005753738805651665, + 0.005596105474978685, + 0.00827313493937254, + 0.007675126660615206, + 0.005458945874124765, + 0.004947690758854151, + 0.0037816341500729322, + 0.0049995919689536095, + 0.0032637009862810373, + 0.0061832754872739315, + 0.0014698893064633012, + 0.0017417233902961016, + 0.0018588381353765726, + 0.002114564646035433, + 0.0015039151767268777, + 0.0027336913626641035, + 0.00205416907556355, + 0.0011527041206136346, + 0.0027718888595700264, + 0.003543452825397253, + 0.007218711078166962, + 0.009291415102779865, + 0.004270342644304037, + 0.0032012704759836197, + 0.0030170916579663754, + 0.0022475121077150106, + 0.0028488377574831247, + 0.0017235359409824014, + 0.0034446679055690765, + 0.0028447380755096674, + 0.002284912159666419, + 0.003909984603524208, + 0.0023560868576169014, + 0.0042010582983493805, + 0.0034379789140075445, + 0.005533060524612665, + 0.004727999214082956, + 0.0057894280180335045, + 0.0046221064403653145, + 0.0034913807176053524, + 0.005992010701447725, + 0.004828792065382004, + 0.006674177013337612, + 0.006704280618578196, + 0.009625228121876717, + 0.005018473137170076, + 0.005120487418025732, + 0.012739676982164383, + 0.004233939107507467, + 0.011808661743998528, + 0.0017480225069448352, + 0.0014449814334511757, + 0.0018456369871273637, + 0.0013615230564028025, + 0.0019766041077673435, + 0.0015580125618726015, + 0.0016046769451349974, + 0.0018987300572916865, + 0.002139714080840349, + 0.0021203365176916122, + 0.003473505610600114, + 0.0016916218446567655, + 0.0026266055647283792, + 0.004312567878514528, + 0.0021824585273861885, + 0.002379788551479578, + 0.07190748304128647, + 0.06194237992167473, + 0.07630713284015656, + 0.05171269550919533, + 0.09026900678873062, + 0.16466505825519562, + 0.04090365022420883, + 0.13279037177562714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.10035862773656845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371164321899414 + ], + "opposite_logprob_tokens": [ + -0.3871164619922638 + ], + "logprob_reward": -1.1371164321899414, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371164321899414, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2165844589471817, + 0.28535276651382446, + 0.05127238109707832, + 0.04491693526506424, + 0.03082663007080555, + 0.049855317920446396, + 0.057357314974069595, + 0.03337513282895088, + 0.04814549535512924, + 0.18231353163719177 + ], + "credit_map": [ + 0.026298483833670616, + 0.027360297739505768, + 0.02090960182249546, + 0.04000446945428848, + 0.020652098581194878, + 0.029613209888339043, + 0.025835871696472168, + 0.025910429656505585, + 0.06222526729106903, + 0.04918605461716652, + 0.03469569608569145, + 0.031210778281092644, + 0.01798594370484352, + 0.031857289373874664, + 0.013475160114467144, + 0.04471656680107117, + 0.005659718997776508, + 0.006868409924209118, + 0.006051686126738787, + 0.006072016898542643, + 0.008270039223134518, + 0.006335974205285311, + 0.004925272427499294, + 0.007089260034263134, + 0.0044802334159612656, + 0.00401243194937706, + 0.009799200110137463, + 0.012015935964882374, + 0.0037862746976315975, + 0.004209357313811779, + 0.0034589990973472595, + 0.003154502948746085, + 0.00453896215185523, + 0.0018577963346615434, + 0.005861316807568073, + 0.003909086808562279, + 0.002624361775815487, + 0.004089616239070892, + 0.002743616234511137, + 0.005201873369514942, + 0.006223736330866814, + 0.005077793262898922, + 0.007000741548836231, + 0.006348221097141504, + 0.005197905469685793, + 0.0069638644345104694, + 0.007145701441913843, + 0.005897352937608957, + 0.005792638286948204, + 0.0052071441896259785, + 0.006725286599248648, + 0.008204696699976921, + 0.0051113092340528965, + 0.010335619561374187, + 0.004493666812777519, + 0.011486958712339401, + 0.003398336237296462, + 0.0067484620958566666, + 0.004388190805912018, + 0.003944624215364456, + 0.004583710804581642, + 0.00489469850435853, + 0.0022438836749643087, + 0.003173225559294224, + 0.005454555619508028, + 0.003267051186412573, + 0.011514305137097836, + 0.003645507851615548, + 0.0034970154520124197, + 0.01293529849499464, + 0.0033103283494710922, + 0.004521432798355818, + 0.024658065289258957, + 0.01878141239285469, + 0.01850097067654133, + 0.012887551449239254, + 0.013557378202676773, + 0.050463225692510605, + 0.010681522078812122, + 0.03278341889381409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -1.1371164321899414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820743829011917 + ], + "opposite_logprob_tokens": [ + -4.018207550048828 + ], + "logprob_reward": -0.01820743829011917, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820743829011917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1121022179722786, + 0.11494278162717819, + 0.0578853040933609, + 0.12819811701774597, + 0.09531320631504059, + 0.11232359707355499, + 0.06754276901483536, + 0.06660439819097519, + 0.0780993327498436, + 0.16698822379112244 + ], + "credit_map": [ + 0.01789757050573826, + 0.01244516484439373, + 0.01725519262254238, + 0.013998518697917461, + 0.008120378479361534, + 0.015567717142403126, + 0.013027151115238667, + 0.013790538534522057, + 0.023304825648665428, + 0.018461985513567924, + 0.013875878416001797, + 0.012185639701783657, + 0.006194144021719694, + 0.013055014424026012, + 0.008685710839927197, + 0.01917959563434124, + 0.006064070854336023, + 0.007163336034864187, + 0.007761739660054445, + 0.0072666797786951065, + 0.006174370646476746, + 0.01074520219117403, + 0.007771268021315336, + 0.0049386401660740376, + 0.009770308621227741, + 0.01222608145326376, + 0.02470254898071289, + 0.0343184769153595, + 0.012620705179870129, + 0.013709677383303642, + 0.011306645348668098, + 0.009543675929307938, + 0.00947414431720972, + 0.007101274095475674, + 0.01377292163670063, + 0.016549071297049522, + 0.007802953943610191, + 0.012195986695587635, + 0.010153225623071194, + 0.01826363056898117, + 0.021745644509792328, + 0.010935010388493538, + 0.015400593169033527, + 0.009790368378162384, + 0.007784654852002859, + 0.02454305812716484, + 0.00720149977132678, + 0.014922772534191608, + 0.008281839080154896, + 0.007410695310682058, + 0.011060305871069431, + 0.008921165019273758, + 0.007265330292284489, + 0.008122281171381474, + 0.005490221083164215, + 0.010990933515131474, + 0.010147918947041035, + 0.005439344327896833, + 0.013472591526806355, + 0.007255654316395521, + 0.008287165313959122, + 0.008343876339495182, + 0.004814484156668186, + 0.008843372575938702, + 0.008543561212718487, + 0.006665453780442476, + 0.015649450942873955, + 0.0077840108424425125, + 0.007406833115965128, + 0.01864108070731163, + 0.005397927016019821, + 0.008011018857359886, + 0.01594226434826851, + 0.014747192151844501, + 0.01586170122027397, + 0.011798338033258915, + 0.015179910697042942, + 0.04802434518933296, + 0.008322644978761673, + 0.03711182624101639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.01820743829011917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234618782997131 + ], + "opposite_logprob_tokens": [ + -0.8984618782997131 + ], + "logprob_reward": -0.5234618782997131, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234618782997131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09115570783615112, + 0.1856965273618698, + 0.03201402351260185, + 0.08813464641571045, + 0.10239393264055252, + 0.09595111012458801, + 0.1158115342259407, + 0.07360759377479553, + 0.06452285498380661, + 0.15071207284927368 + ], + "credit_map": [ + 0.01735183782875538, + 0.010793869383633137, + 0.008065646514296532, + 0.014133545570075512, + 0.006298900581896305, + 0.011640154756605625, + 0.011655699461698532, + 0.011216051876544952, + 0.0420173816382885, + 0.0335283987224102, + 0.023676207289099693, + 0.022119687870144844, + 0.012249625287950039, + 0.019860047847032547, + 0.008468733169138432, + 0.023776428773999214, + 0.004130250308662653, + 0.005241582170128822, + 0.004017877392470837, + 0.0037506516091525555, + 0.003651859238743782, + 0.004026792477816343, + 0.00416477071121335, + 0.003030239138752222, + 0.008561952039599419, + 0.011484957300126553, + 0.011868850328028202, + 0.021355310454964638, + 0.011110793799161911, + 0.008978071622550488, + 0.008692333474755287, + 0.006082379259169102, + 0.012263894081115723, + 0.006053653545677662, + 0.019119389355182648, + 0.011049753054976463, + 0.007032341789454222, + 0.012877457775175571, + 0.008519640192389488, + 0.025477800518274307, + 0.013041426427662373, + 0.00753812538459897, + 0.014129145070910454, + 0.009010939858853817, + 0.007192086894065142, + 0.015253412537276745, + 0.017111463472247124, + 0.012674511410295963, + 0.011920925229787827, + 0.008485126309096813, + 0.01403927244246006, + 0.011166193522512913, + 0.010270616970956326, + 0.032245852053165436, + 0.009201890788972378, + 0.018481656908988953, + 0.008024077862501144, + 0.015900744125247, + 0.011292052455246449, + 0.007560454308986664, + 0.008396454155445099, + 0.008277303539216518, + 0.006300906650722027, + 0.00785559881478548, + 0.0068514421582221985, + 0.005125993862748146, + 0.013225139118731022, + 0.004615137353539467, + 0.0056389193050563335, + 0.014558217488229275, + 0.008126762695610523, + 0.0063812462612986565, + 0.016539305448532104, + 0.01600467599928379, + 0.019972462207078934, + 0.015466995537281036, + 0.013049682602286339, + 0.026551438495516777, + 0.012853538617491722, + 0.03027397021651268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.5234618782997131 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, arcing, panoramic shot begins at a slightly lower angle, sweeping across the grand, cavernous hall of a Gothic stone building, capturing a pivotal moment rendered in classic Dungeons & Dragons art. The scene, defined by heroic poses and dramatic black-and-white lighting, centers on a wise, bearded headmaster and a young wizard. The headmaster, with his flowing white beard falling to his chest and deep, all-seeing eyes, wears ornate dark blue robes embroidered with silver constellations, pointing his finger as he makes his point. Beside him stands the young wizard, recognizable by his tousled black hair, round glasses, and the faint, signature lightning bolt scar on his forehead. He stands there, his expression serious. They are explaining something of great importance to the diverse group of students gathered before them. As the camera continues its smooth, circular path, it reveals the students' resolute faces, illuminated by the warm, flickering glow of hundreds of floating candles, which drift toward the vaulted, shadowed ceiling. On the towering stone walls, portraits of ancient figures subtly move and whisper within their frames. The students, dressed in dark robes of varying colors, listened intently, their bodies in powerful poses of concentration—a girl with tightly braided hair, a boy beside her with a clenched fist, their expressions reflecting a shared solemn purpose in a play of light and shadow, full of both heroism and tension.", + "seed": 29307, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_14/index_724_seed_29307_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3600019417703152, + "vlm_logprob": -0.3600019417703152, + "vlm_yes_no_margin": 1.850000011920929, + "vlm_token_credit": -0.3600019417703152, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07167166471481323 + ], + "opposite_logprob_tokens": [ + -2.821671724319458 + ], + "logprob_reward": -0.07167166471481323, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07167166471481323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1915801614522934, + 0.01723676361143589, + 0.03658217191696167, + 0.05493210628628731, + 0.014154736883938313, + 0.12825781106948853, + 0.17550867795944214, + 0.1614151895046234, + 0.18165771663188934, + 0.03867465257644653 + ], + "credit_map": [ + 0.02207188308238983, + 0.005484803579747677, + 0.006595471873879433, + 0.010775275528430939, + 0.007361146621406078, + 0.08187900483608246, + 0.04423016309738159, + 0.013182409107685089, + 0.001570680644363165, + 0.002128057647496462, + 0.0024683780502527952, + 0.0022409583907574415, + 0.0017186985351145267, + 0.002365065272897482, + 0.0025798100978136063, + 0.00216511357575655, + 0.006995192263275385, + 0.002739085117354989, + 0.0045912666246294975, + 0.0019687111489474773, + 0.0038788337260484695, + 0.00565866194665432, + 0.003972235135734081, + 0.0067781852558255196, + 0.008062299340963364, + 0.003623242024332285, + 0.004966007545590401, + 0.004742931108921766, + 0.004625388886779547, + 0.00972903985530138, + 0.005652121268212795, + 0.013531074859201908, + 0.0009553268318995833, + 0.001462343498133123, + 0.0017152318032458425, + 0.002424113918095827, + 0.0024516379926353693, + 0.0013750571524724364, + 0.0019578852225095034, + 0.0018131412798538804, + 0.013707752339541912, + 0.007886606268584728, + 0.013172557577490807, + 0.011780318804085255, + 0.016020020470023155, + 0.028115427121520042, + 0.021390080451965332, + 0.01618504337966442, + 0.020269382745027542, + 0.011392544023692608, + 0.025783691555261612, + 0.01040218397974968, + 0.027028709650039673, + 0.03220725432038307, + 0.01900983788073063, + 0.02941507287323475, + 0.0272002425044775, + 0.014714352786540985, + 0.017621085047721863, + 0.017611021175980568, + 0.021148230880498886, + 0.02173527143895626, + 0.0226600281894207, + 0.018724964931607246, + 0.023722480982542038, + 0.01042910385876894, + 0.011784273199737072, + 0.011150668375194073, + 0.028088564053177834, + 0.05490012839436531, + 0.01329895295202732, + 0.02828354202210903, + 0.006912288721650839, + 0.005520110949873924, + 0.0031960324849933386, + 0.003836479038000107, + 0.006239676848053932, + 0.006061311345547438, + 0.002942125778645277, + 0.003966625779867172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.07167166471481323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499942973256111 + ], + "opposite_logprob_tokens": [ + -2.929999351501465 + ], + "logprob_reward": -0.05499942973256111, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499942973256111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05761970952153206, + 0.022399764508008957, + 0.021176595240831375, + 0.024656496942043304, + 0.13264170289039612, + 0.04476424679160118, + 0.0626162588596344, + 0.12303194403648376, + 0.4577368497848511, + 0.05335639789700508 + ], + "credit_map": [ + 0.006026014685630798, + 0.006922841537743807, + 0.0117292245849967, + 0.006420752499252558, + 0.004841919057071209, + 0.00592059688642621, + 0.007370484061539173, + 0.008387877605855465, + 0.0022184057161211967, + 0.0038109265733510256, + 0.003558363998308778, + 0.0034400401636958122, + 0.0019387106876820326, + 0.0018378577660769224, + 0.0030016477685421705, + 0.0025938141625374556, + 0.0029358910396695137, + 0.0021881063003093004, + 0.0027751366142183542, + 0.0021163863129913807, + 0.002438931493088603, + 0.0029412212315946817, + 0.0028563844971358776, + 0.002924541011452675, + 0.0030034766532480717, + 0.0029564653523266315, + 0.0019133709138259292, + 0.0035041035152971745, + 0.002111685462296009, + 0.003945629578083754, + 0.0037278945092111826, + 0.0034938736353069544, + 0.008337503299117088, + 0.00872984528541565, + 0.013920274563133717, + 0.012419008649885654, + 0.038265835493803024, + 0.014794917777180672, + 0.02045614644885063, + 0.01571817509829998, + 0.005177519284188747, + 0.005377137567847967, + 0.004963316954672337, + 0.009749090299010277, + 0.003907900303602219, + 0.004172943066805601, + 0.008060328662395477, + 0.0033560125157237053, + 0.01107851229608059, + 0.008432971313595772, + 0.007545112166553736, + 0.005117040127515793, + 0.004847168922424316, + 0.010175547562539577, + 0.006688009947538376, + 0.008731896057724953, + 0.02065861225128174, + 0.01055251806974411, + 0.01461671944707632, + 0.0131779033690691, + 0.0154731972143054, + 0.01579071395099163, + 0.017375614494085312, + 0.015386675484478474, + 0.054426562041044235, + 0.07622232288122177, + 0.03205171227455139, + 0.05819866061210632, + 0.07838521152734756, + 0.06194399297237396, + 0.04028495028614998, + 0.056223466992378235, + 0.008365063928067684, + 0.008493119850754738, + 0.005507366266101599, + 0.008917441591620445, + 0.006795608438551426, + 0.004662923980504274, + 0.004443491343408823, + 0.006171387620270252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.05499942973256111 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236864566802979 + ], + "opposite_logprob_tokens": [ + -0.3486865162849426 + ], + "logprob_reward": -1.2236864566802979, + "yes_no_margin_reward": -0.8749999403953552, + "token_credit_reward": -1.2236864566802979, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07837343215942383, + 0.11172887682914734, + 0.09816985577344894, + 0.14338073134422302, + 0.053778503090143204, + 0.0613914355635643, + 0.11187653988599777, + 0.14015811681747437, + 0.1325068473815918, + 0.06863570958375931 + ], + "credit_map": [ + 0.007331537082791328, + 0.010789310559630394, + 0.010995177552103996, + 0.01315194834023714, + 0.008997367694973946, + 0.007905521430075169, + 0.010519681498408318, + 0.008682887069880962, + 0.01158836204558611, + 0.01843908056616783, + 0.00962440948933363, + 0.011287407018244267, + 0.020616287365555763, + 0.013489813543856144, + 0.014409448951482773, + 0.012274065054953098, + 0.019781801849603653, + 0.005028293468058109, + 0.018993791192770004, + 0.005979748908430338, + 0.007489415351301432, + 0.016619062051177025, + 0.010934063233435154, + 0.013343672268092632, + 0.021251043304800987, + 0.019902510568499565, + 0.01999342441558838, + 0.01406613364815712, + 0.011748417280614376, + 0.023260224610567093, + 0.01521515566855669, + 0.017943810671567917, + 0.0036203761119395494, + 0.007502469699829817, + 0.006482766941189766, + 0.010066648945212364, + 0.00754306185990572, + 0.0034382897429168224, + 0.008750803768634796, + 0.0063740867190063, + 0.007634194567799568, + 0.008334092795848846, + 0.0057131703943014145, + 0.010634616017341614, + 0.004713629838079214, + 0.007630425039678812, + 0.011171887628734112, + 0.005559420213103294, + 0.015393662266433239, + 0.013180716894567013, + 0.014595823362469673, + 0.007026314735412598, + 0.01373283751308918, + 0.018931671977043152, + 0.0122917415574193, + 0.016723770648241043, + 0.020398877561092377, + 0.01451531145721674, + 0.015076256357133389, + 0.01650501787662506, + 0.02002674527466297, + 0.018870003521442413, + 0.02058236114680767, + 0.014183538034558296, + 0.02020060084760189, + 0.01604984700679779, + 0.00840749405324459, + 0.014920851215720177, + 0.028354935348033905, + 0.01734279841184616, + 0.014153316617012024, + 0.013077009469270706, + 0.011360595934092999, + 0.0105005893856287, + 0.005774805322289467, + 0.006513598840683699, + 0.013985889963805676, + 0.00774050410836935, + 0.0063281976617872715, + 0.00643152417615056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -1.2236864566802979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602159202098846 + ], + "opposite_logprob_tokens": [ + -3.895602226257324 + ], + "logprob_reward": -0.020602159202098846, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602159202098846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2284514307975769, + 0.07313892245292664, + 0.1451297253370285, + 0.05696149915456772, + 0.05794699862599373, + 0.057037319988012314, + 0.08813800662755966, + 0.09433479607105255, + 0.1536819338798523, + 0.04517946392297745 + ], + "credit_map": [ + 0.02392522804439068, + 0.014565341174602509, + 0.01823502965271473, + 0.0196993388235569, + 0.013276946730911732, + 0.07596170902252197, + 0.04549252241849899, + 0.01729527860879898, + 0.006696005817502737, + 0.010799617506563663, + 0.007242628373205662, + 0.008230561390519142, + 0.010300451889634132, + 0.010397098958492279, + 0.009145967662334442, + 0.010326583869755268, + 0.021925512701272964, + 0.011792662553489208, + 0.023280588909983635, + 0.010084063746035099, + 0.02115533873438835, + 0.020618999376893044, + 0.016424622386693954, + 0.019847920164465904, + 0.007160657551139593, + 0.0061922078020870686, + 0.008778641931712627, + 0.007055807858705521, + 0.0044237906113266945, + 0.009217264130711555, + 0.0069784740917384624, + 0.0071546463295817375, + 0.004210338927805424, + 0.005498493555933237, + 0.005950316786766052, + 0.012334635481238365, + 0.00969883892685175, + 0.0037900914903730154, + 0.009570183232426643, + 0.00689409114420414, + 0.00895902793854475, + 0.006758049596101046, + 0.006506267935037613, + 0.005752533674240112, + 0.005020651500672102, + 0.004461595322936773, + 0.014316700398921967, + 0.005262486170977354, + 0.010712851770222187, + 0.011016150005161762, + 0.009239697828888893, + 0.00691246660426259, + 0.012227043509483337, + 0.014125224202871323, + 0.01107114739716053, + 0.01283341646194458, + 0.015689851716160774, + 0.008698908612132072, + 0.008914674632251263, + 0.009609389118850231, + 0.014101315289735794, + 0.013354734517633915, + 0.014071512967348099, + 0.009894398041069508, + 0.02052450366318226, + 0.015137550421059132, + 0.007517698686569929, + 0.011142631061375141, + 0.02818983606994152, + 0.04345228895545006, + 0.011411264538764954, + 0.016306152567267418, + 0.00778771610930562, + 0.0051749576814472675, + 0.005893463268876076, + 0.004008937161415815, + 0.006699282210320234, + 0.0047308458015322685, + 0.0072406199760735035, + 0.0036436377558857203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.020602159202098846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4290499985218048 + ], + "opposite_logprob_tokens": [ + -1.0540499687194824 + ], + "logprob_reward": -0.4290499985218048, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4290499985218048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41119393706321716, + 0.026080461218953133, + 0.04826560243964195, + 0.0927838385105133, + 0.046565942466259, + 0.028925379738211632, + 0.06937337666749954, + 0.15372563898563385, + 0.05756322667002678, + 0.06552263349294662 + ], + "credit_map": [ + 0.04560540243983269, + 0.011133162304759026, + 0.01562550850212574, + 0.02619062177836895, + 0.013865926302969456, + 0.17391681671142578, + 0.0964510589838028, + 0.0284054484218359, + 0.002535661682486534, + 0.0033065988682210445, + 0.00244170892983675, + 0.0037710415199398994, + 0.003934376407414675, + 0.0028788710478693247, + 0.004130202811211348, + 0.0030819985549896955, + 0.008085660636425018, + 0.003674478502944112, + 0.006196130532771349, + 0.003779364051297307, + 0.008364954963326454, + 0.0068082138895988464, + 0.004912084434181452, + 0.006444715429097414, + 0.014856977388262749, + 0.008383465930819511, + 0.011227630078792572, + 0.006868337281048298, + 0.00822417438030243, + 0.015605814754962921, + 0.008340923115611076, + 0.019276514649391174, + 0.0030824788846075535, + 0.005845896899700165, + 0.004445236176252365, + 0.005834415555000305, + 0.010247907601296902, + 0.003404424525797367, + 0.008752305991947651, + 0.004953278228640556, + 0.0031033754348754883, + 0.0032811788842082024, + 0.00358787109144032, + 0.004778798203915358, + 0.0030631592962890863, + 0.0038451403379440308, + 0.004204015713185072, + 0.003061841242015362, + 0.014682845212519169, + 0.006100919563323259, + 0.008415483869612217, + 0.0037355623207986355, + 0.0072034792974591255, + 0.011139989830553532, + 0.007151213474571705, + 0.010943879373371601, + 0.025561204180121422, + 0.013072112575173378, + 0.016722355037927628, + 0.016813868656754494, + 0.021048419177532196, + 0.0202826838940382, + 0.022097570821642876, + 0.018127409741282463, + 0.009570621885359287, + 0.0055710081942379475, + 0.0032616755925118923, + 0.004620542284101248, + 0.011235409416258335, + 0.011950042098760605, + 0.005248083733022213, + 0.00610584020614624, + 0.01091964915394783, + 0.008766981773078442, + 0.0071852183900773525, + 0.005755614023655653, + 0.009514291770756245, + 0.008300213143229485, + 0.008090905845165253, + 0.006989762652665377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.4290499985218048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, arcing, panoramic shot begins at a slightly lower angle, sweeping across the grand, cavernous hall of a Gothic stone building, capturing a pivotal moment rendered in classic Dungeons & Dragons art. The scene, defined by heroic poses and dramatic black-and-white lighting, centers on a wise, bearded headmaster and a young wizard. The headmaster, with his flowing white beard falling to his chest and deep, all-seeing eyes, wears ornate dark blue robes embroidered with silver constellations, pointing his finger as he makes his point. Beside him stands the young wizard, recognizable by his tousled black hair, round glasses, and the faint, signature lightning bolt scar on his forehead. He stands there, his expression serious. They are explaining something of great importance to the diverse group of students gathered before them. As the camera continues its smooth, circular path, it reveals the students' resolute faces, illuminated by the warm, flickering glow of hundreds of floating candles, which drift toward the vaulted, shadowed ceiling. On the towering stone walls, portraits of ancient figures subtly move and whisper within their frames. The students, dressed in dark robes of varying colors, listened intently, their bodies in powerful poses of concentration—a girl with tightly braided hair, a boy beside her with a clenched fist, their expressions reflecting a shared solemn purpose in a play of light and shadow, full of both heroism and tension.", + "seed": 29308, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_14/index_724_seed_29308_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3073061924427748, + "vlm_logprob": -0.3073061924427748, + "vlm_yes_no_margin": 1.7, + "vlm_token_credit": -0.3073061924427748, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09127800911664963 + ], + "opposite_logprob_tokens": [ + -2.591278076171875 + ], + "logprob_reward": -0.09127800911664963, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09127800911664963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1654253751039505, + 0.06330271065235138, + 0.0473581999540329, + 0.05696931108832359, + 0.08759617060422897, + 0.12006088346242905, + 0.031357988715171814, + 0.1361832618713379, + 0.11815806478261948, + 0.17358803749084473 + ], + "credit_map": [ + 0.020832965150475502, + 0.012869875878095627, + 0.025601083412766457, + 0.014812707901000977, + 0.01534252054989338, + 0.03089592605829239, + 0.016089146956801414, + 0.028981143608689308, + 0.007792466785758734, + 0.004938588477671146, + 0.008476587012410164, + 0.008097597397863865, + 0.005321823060512543, + 0.00905489083379507, + 0.007488143630325794, + 0.012132606469094753, + 0.0057851835153996944, + 0.003563057631254196, + 0.007975933142006397, + 0.0039129978977143764, + 0.0036599328741431236, + 0.006583863869309425, + 0.005482219159603119, + 0.010395004414021969, + 0.007494942285120487, + 0.00424253661185503, + 0.014934241771697998, + 0.005424404051154852, + 0.0059556737542152405, + 0.006074253004044294, + 0.004706178791821003, + 0.008137078024446964, + 0.006556912325322628, + 0.005424319766461849, + 0.014638383872807026, + 0.008011837489902973, + 0.010939447209239006, + 0.015242258086800575, + 0.013998155482113361, + 0.012784852646291256, + 0.01662536896765232, + 0.015115907415747643, + 0.014240802265703678, + 0.010576692409813404, + 0.01745535060763359, + 0.015501873567700386, + 0.011278300546109676, + 0.01926657371222973, + 0.0022105034440755844, + 0.002238492714241147, + 0.005300063639879227, + 0.002256289590150118, + 0.002882665256038308, + 0.005863575730472803, + 0.0034054755233228207, + 0.007200919557362795, + 0.021376537159085274, + 0.009041977114975452, + 0.027804717421531677, + 0.013169923797249794, + 0.021976908668875694, + 0.01850748062133789, + 0.010050534270703793, + 0.014255181886255741, + 0.015511365607380867, + 0.0070638456381857395, + 0.01710699498653412, + 0.011422490701079369, + 0.009318388067185879, + 0.020355576649308205, + 0.008883464150130749, + 0.02849593013525009, + 0.027349967509508133, + 0.03630981221795082, + 0.01809966005384922, + 0.02127743698656559, + 0.0154418321326375, + 0.028507050126791, + 0.011998319067060947, + 0.014603953808546066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.09127800911664963 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909071981906891 + ], + "opposite_logprob_tokens": [ + -2.4640908241271973 + ], + "logprob_reward": -0.08909071981906891, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909071981906891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13388417661190033, + 0.15203982591629028, + 0.08404409885406494, + 0.13395918905735016, + 0.09593361616134644, + 0.042038444429636, + 0.04332094267010689, + 0.09975051879882812, + 0.09964974969625473, + 0.11537947505712509 + ], + "credit_map": [ + 0.01264571025967598, + 0.015135431662201881, + 0.01875949464738369, + 0.015329711139202118, + 0.016098210588097572, + 0.019106784835457802, + 0.018666740506887436, + 0.018142087385058403, + 0.029010064899921417, + 0.020883431658148766, + 0.02005419135093689, + 0.016351791098713875, + 0.00982634723186493, + 0.01714089885354042, + 0.01871931180357933, + 0.020053770393133163, + 0.008094679564237595, + 0.016419852152466774, + 0.01042166817933321, + 0.011876864358782768, + 0.01094811037182808, + 0.007209696341305971, + 0.01010868139564991, + 0.00896455254405737, + 0.013805622234940529, + 0.016871796920895576, + 0.038166593760252, + 0.012056345120072365, + 0.00991193950176239, + 0.014271584339439869, + 0.007353164255619049, + 0.021522141993045807, + 0.011316807940602303, + 0.010821775533258915, + 0.021012090146541595, + 0.008379021659493446, + 0.0110867228358984, + 0.012634474784135818, + 0.007251259405165911, + 0.013431462459266186, + 0.003561967285349965, + 0.0058814045041799545, + 0.005102677270770073, + 0.005132326390594244, + 0.004283681977540255, + 0.0056101116351783276, + 0.004310402553528547, + 0.00815587118268013, + 0.0037604854442179203, + 0.007626134902238846, + 0.006508486811071634, + 0.004514004103839397, + 0.006012593396008015, + 0.005255986470729113, + 0.004016604740172625, + 0.005626648664474487, + 0.016929158940911293, + 0.007762479595839977, + 0.012639643624424934, + 0.007702409289777279, + 0.020397530868649483, + 0.015518036670982838, + 0.00887299794703722, + 0.00992825347930193, + 0.011970743536949158, + 0.006182848941534758, + 0.019161825999617577, + 0.010910499840974808, + 0.007359736133366823, + 0.015468002296984196, + 0.00758382910862565, + 0.021012265235185623, + 0.013463357463479042, + 0.026660148054361343, + 0.005403125658631325, + 0.016352199018001556, + 0.021214554086327553, + 0.007567259017378092, + 0.014864785596728325, + 0.009854042902588844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.08909071981906891 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328797340393066 + ], + "opposite_logprob_tokens": [ + -0.7578797340393066 + ], + "logprob_reward": -0.6328797340393066, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328797340393066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13985510170459747, + 0.11082703620195389, + 0.0867055207490921, + 0.17367802560329437, + 0.06838110834360123, + 0.0672028586268425, + 0.046562645584344864, + 0.13532470166683197, + 0.06079289689660072, + 0.11067013442516327 + ], + "credit_map": [ + 0.015971455723047256, + 0.02301083691418171, + 0.026260992512106895, + 0.01646764762699604, + 0.011959783732891083, + 0.014300874434411526, + 0.0165354385972023, + 0.015348055399954319, + 0.010402044281363487, + 0.012819352559745312, + 0.014228836633265018, + 0.014775287359952927, + 0.017044348642230034, + 0.011664222925901413, + 0.017204992473125458, + 0.01268794760107994, + 0.010609841905534267, + 0.005830057431012392, + 0.012991044670343399, + 0.008575968444347382, + 0.008590269833803177, + 0.011172843165695667, + 0.014030933380126953, + 0.014904551208019257, + 0.023785723373293877, + 0.01377856358885765, + 0.022932009771466255, + 0.01505053136497736, + 0.022624563425779343, + 0.030630823224782944, + 0.026602106168866158, + 0.018273673951625824, + 0.0070109120570123196, + 0.007343061733990908, + 0.013747947290539742, + 0.006158041302114725, + 0.007214707788079977, + 0.01200528722256422, + 0.004615643061697483, + 0.010285499505698681, + 0.007452555000782013, + 0.005883558187633753, + 0.009617912583053112, + 0.006176221650093794, + 0.009461632929742336, + 0.011030802503228188, + 0.00865915883332491, + 0.008921004831790924, + 0.004753694403916597, + 0.0038313933182507753, + 0.006967666558921337, + 0.0042288280092179775, + 0.00513280788436532, + 0.01025200542062521, + 0.0041979290544986725, + 0.00719831557944417, + 0.018253738060593605, + 0.01211854163557291, + 0.016623549163341522, + 0.018808513879776, + 0.019701644778251648, + 0.019380463287234306, + 0.015046440064907074, + 0.015391805209219456, + 0.006312211509793997, + 0.0073037100955843925, + 0.01001999992877245, + 0.008962522260844707, + 0.0074020023457705975, + 0.007773255929350853, + 0.004706678446382284, + 0.008312511257827282, + 0.016295192763209343, + 0.012639256194233894, + 0.012320105917751789, + 0.015139187686145306, + 0.01669010892510414, + 0.016675902530550957, + 0.011001897975802422, + 0.009908472187817097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.6328797340393066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820306226611137 + ], + "opposite_logprob_tokens": [ + -3.529820203781128 + ], + "logprob_reward": -0.029820306226611137, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820306226611137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11809109151363373, + 0.05297176167368889, + 0.09128287434577942, + 0.05118029937148094, + 0.04635242000222206, + 0.08761361241340637, + 0.08705087751150131, + 0.08890650421380997, + 0.14194980263710022, + 0.2346007525920868 + ], + "credit_map": [ + 0.01618015393614769, + 0.013943050988018513, + 0.014041892252862453, + 0.016447847709059715, + 0.01361953467130661, + 0.017499200999736786, + 0.013753273524343967, + 0.012606138363480568, + 0.0075146215967834, + 0.0055943639017641544, + 0.007714053615927696, + 0.007269098423421383, + 0.004806132987141609, + 0.008286872878670692, + 0.0047714924439787865, + 0.007015123963356018, + 0.01400283444672823, + 0.009171687066555023, + 0.01413482055068016, + 0.00756814144551754, + 0.007485650945454836, + 0.015841351822018623, + 0.00972307100892067, + 0.013355318456888199, + 0.006195219233632088, + 0.006398648023605347, + 0.010714358650147915, + 0.0038819159381091595, + 0.005087513942271471, + 0.00709752319380641, + 0.004788736812770367, + 0.0070163835771381855, + 0.005630944389849901, + 0.005453633144497871, + 0.0052851298823952675, + 0.0049559432081878185, + 0.010197918862104416, + 0.004982109181582928, + 0.004484279081225395, + 0.005362462252378464, + 0.009224913083016872, + 0.010180527344346046, + 0.011904713697731495, + 0.014625578187406063, + 0.012858640402555466, + 0.008881163783371449, + 0.01092600915580988, + 0.0090120704844594, + 0.0062363604083657265, + 0.011926858685910702, + 0.013062719255685806, + 0.013936996459960938, + 0.008720372803509235, + 0.009377269074320793, + 0.013490832410752773, + 0.010299470275640488, + 0.011175882071256638, + 0.011745652183890343, + 0.008713907562196255, + 0.010010162368416786, + 0.014299441128969193, + 0.012390976771712303, + 0.011261281557381153, + 0.009309201501309872, + 0.01874324306845665, + 0.012935100123286247, + 0.019858309999108315, + 0.013741661794483662, + 0.01815914362668991, + 0.018577365204691887, + 0.011230948381125927, + 0.02870403230190277, + 0.03137655556201935, + 0.0597650408744812, + 0.01712450385093689, + 0.03249254450201988, + 0.029652118682861328, + 0.021896420046687126, + 0.023375429213047028, + 0.018918141722679138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.029820306226611137 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934621930122375 + ], + "opposite_logprob_tokens": [ + -0.6934621930122375 + ], + "logprob_reward": -0.6934621930122375, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934621930122375, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18135596811771393, + 0.1739986687898636, + 0.049419742077589035, + 0.06413212418556213, + 0.08256770670413971, + 0.09388504177331924, + 0.041339173913002014, + 0.08427061140537262, + 0.12806780636310577, + 0.10096311569213867 + ], + "credit_map": [ + 0.01867743954062462, + 0.021743373945355415, + 0.02534358575940132, + 0.01771644316613674, + 0.020662495866417885, + 0.026096390560269356, + 0.024374421685934067, + 0.02674182318150997, + 0.024982579052448273, + 0.010610664263367653, + 0.025594258680939674, + 0.012629535980522633, + 0.013675197027623653, + 0.024009741842746735, + 0.018345173448324203, + 0.044151533395051956, + 0.0062866415828466415, + 0.008262140676379204, + 0.00705714663490653, + 0.004916712641716003, + 0.0047118691727519035, + 0.005893800873309374, + 0.003990683238953352, + 0.008300748653709888, + 0.00806150771677494, + 0.005629740189760923, + 0.01260221004486084, + 0.006083378568291664, + 0.006498977076262236, + 0.0122355492785573, + 0.005314249079674482, + 0.0077065094374120235, + 0.007506870198994875, + 0.0067428844049572945, + 0.015303706750273705, + 0.007752478588372469, + 0.0110032819211483, + 0.010848496109247208, + 0.01003848947584629, + 0.013371502049267292, + 0.011425276286900043, + 0.008457674644887447, + 0.012638355605304241, + 0.01878511905670166, + 0.007956219837069511, + 0.01392973680049181, + 0.008929705247282982, + 0.011762958019971848, + 0.0034061421174556017, + 0.009507018141448498, + 0.005476132035255432, + 0.0058671412989497185, + 0.004844654817134142, + 0.0054741655476391315, + 0.0033316819462925196, + 0.0034322370775043964, + 0.013860184699296951, + 0.007411960978060961, + 0.015031653456389904, + 0.009449048899114132, + 0.009270237758755684, + 0.012168988585472107, + 0.006674107629805803, + 0.010404421016573906, + 0.015217769891023636, + 0.015094922855496407, + 0.013953475281596184, + 0.012310078367590904, + 0.02078775130212307, + 0.026056120172142982, + 0.009449055418372154, + 0.015198624692857265, + 0.013030705973505974, + 0.01925559900701046, + 0.006111671216785908, + 0.012009326368570328, + 0.01325571071356535, + 0.011703799478709698, + 0.015955127775669098, + 0.009641173295676708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.6934621930122375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, arcing, panoramic shot begins at a slightly lower angle, sweeping across the grand, cavernous hall of a Gothic stone building, capturing a pivotal moment rendered in classic Dungeons & Dragons art. The scene, defined by heroic poses and dramatic black-and-white lighting, centers on a wise, bearded headmaster and a young wizard. The headmaster, with his flowing white beard falling to his chest and deep, all-seeing eyes, wears ornate dark blue robes embroidered with silver constellations, pointing his finger as he makes his point. Beside him stands the young wizard, recognizable by his tousled black hair, round glasses, and the faint, signature lightning bolt scar on his forehead. He stands there, his expression serious. They are explaining something of great importance to the diverse group of students gathered before them. As the camera continues its smooth, circular path, it reveals the students' resolute faces, illuminated by the warm, flickering glow of hundreds of floating candles, which drift toward the vaulted, shadowed ceiling. On the towering stone walls, portraits of ancient figures subtly move and whisper within their frames. The students, dressed in dark robes of varying colors, listened intently, their bodies in powerful poses of concentration—a girl with tightly braided hair, a boy beside her with a clenched fist, their expressions reflecting a shared solemn purpose in a play of light and shadow, full of both heroism and tension.", + "seed": 29309, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_14/index_724_seed_29309_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.27841412276029587, + "vlm_logprob": -0.27841412276029587, + "vlm_yes_no_margin": 2.049999952316284, + "vlm_token_credit": -0.27841412276029587, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06609739363193512 + ], + "opposite_logprob_tokens": [ + -2.8160974979400635 + ], + "logprob_reward": -0.06609739363193512, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06609739363193512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1515588015317917, + 0.08316417038440704, + 0.06474796682596207, + 0.0413382314145565, + 0.042276255786418915, + 0.015082291327416897, + 0.07054315507411957, + 0.031684644520282745, + 0.09418163448572159, + 0.4054228663444519 + ], + "credit_map": [ + 0.01700693741440773, + 0.012863568030297756, + 0.020121444016695023, + 0.01594313606619835, + 0.019821910187602043, + 0.025846116244792938, + 0.02111702412366867, + 0.018838653340935707, + 0.007551342714577913, + 0.0073975492268800735, + 0.010086249560117722, + 0.010884590446949005, + 0.01509901974350214, + 0.009772105142474174, + 0.015090427361428738, + 0.00728288060054183, + 0.008357237093150616, + 0.008700945414602757, + 0.005609053187072277, + 0.006907484959810972, + 0.008156429044902325, + 0.008746164850890636, + 0.008367258124053478, + 0.00990338809788227, + 0.00419932184740901, + 0.007260669954121113, + 0.0037387614138424397, + 0.007069198880344629, + 0.005584773141890764, + 0.005334948189556599, + 0.004976742435246706, + 0.0031738101970404387, + 0.0042302850633859634, + 0.005089771002531052, + 0.004645912442356348, + 0.0054609039798378944, + 0.00591065501794219, + 0.0045572747476398945, + 0.00557228596881032, + 0.006809164769947529, + 0.001692617661319673, + 0.0017016958445310593, + 0.0017924398416653275, + 0.0023528789170086384, + 0.002369483234360814, + 0.0017135898815467954, + 0.001354207401163876, + 0.0021053755190223455, + 0.007589506916701794, + 0.020018018782138824, + 0.004493201617151499, + 0.009405734948813915, + 0.008299782872200012, + 0.00904044322669506, + 0.006197763606905937, + 0.005498695187270641, + 0.004562158603221178, + 0.0047598592936992645, + 0.003062691306695342, + 0.004151439294219017, + 0.0019859373569488525, + 0.007100200280547142, + 0.0030435563530772924, + 0.0030187980737537146, + 0.006057739723473787, + 0.012086154893040657, + 0.0056669763289391994, + 0.03508901596069336, + 0.005425150506198406, + 0.004912138916552067, + 0.00904692243784666, + 0.01589752547442913, + 0.032965242862701416, + 0.13924385607242584, + 0.09014873951673508, + 0.027786318212747574, + 0.025747224688529968, + 0.035961512476205826, + 0.033456504344940186, + 0.020113429054617882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.06609739363193512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812748193740845 + ], + "opposite_logprob_tokens": [ + -3.2881274223327637 + ], + "logprob_reward": -0.03812748193740845, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812748193740845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055315352976322174, + 0.22706638276576996, + 0.1212095394730568, + 0.3367615342140198, + 0.056511543691158295, + 0.02833457663655281, + 0.03624691069126129, + 0.04994663596153259, + 0.03264328092336655, + 0.05596417561173439 + ], + "credit_map": [ + 0.006790269166231155, + 0.0073667424730956554, + 0.007653776090592146, + 0.007569077890366316, + 0.005299310199916363, + 0.008398521691560745, + 0.006027677562087774, + 0.006209981627762318, + 0.018802743405103683, + 0.01859167218208313, + 0.02941243350505829, + 0.032428234815597534, + 0.03609567880630493, + 0.027178287506103516, + 0.04429998993873596, + 0.02025735005736351, + 0.010464963503181934, + 0.011476910673081875, + 0.01773683726787567, + 0.013945089653134346, + 0.021812301129102707, + 0.01469532959163189, + 0.016854673624038696, + 0.014223436824977398, + 0.03877728059887886, + 0.07437117397785187, + 0.025835081934928894, + 0.05943720415234566, + 0.03912970423698425, + 0.044202614575624466, + 0.030532008036971092, + 0.024476483464241028, + 0.006703734397888184, + 0.007245225366204977, + 0.00405009463429451, + 0.007086664903908968, + 0.011204122565686703, + 0.007279367186129093, + 0.007476944010704756, + 0.005465393885970116, + 0.003233702853322029, + 0.003999501466751099, + 0.00282496539875865, + 0.0034635053016245365, + 0.0036482454743236303, + 0.004265008959919214, + 0.0032422663643956184, + 0.003657383844256401, + 0.0035374651197344065, + 0.01006554439663887, + 0.002818737179040909, + 0.005117197521030903, + 0.004028107970952988, + 0.004673559684306383, + 0.0031744400039315224, + 0.0028318625409156084, + 0.006701011210680008, + 0.007934533059597015, + 0.004914676770567894, + 0.0058281756937503815, + 0.006768867839127779, + 0.006581516470760107, + 0.008295882493257523, + 0.002921975450590253, + 0.004845813382416964, + 0.00429889839142561, + 0.004209374077618122, + 0.006621190812438726, + 0.0027394520584493876, + 0.0028624064289033413, + 0.003975393250584602, + 0.003090756479650736, + 0.0058161052875220776, + 0.010140202939510345, + 0.010743667371571064, + 0.005875549744814634, + 0.005846531596034765, + 0.005936837289482355, + 0.006297779735177755, + 0.005307507701218128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.03812748193740845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934084296226501 + ], + "opposite_logprob_tokens": [ + -0.6934084296226501 + ], + "logprob_reward": -0.6934084296226501, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934084296226501, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1295183151960373, + 0.1443922221660614, + 0.10187374800443649, + 0.10003439337015152, + 0.05861975625157356, + 0.06102614477276802, + 0.12177801877260208, + 0.04476940259337425, + 0.05552846938371658, + 0.18245962262153625 + ], + "credit_map": [ + 0.017499875277280807, + 0.01208462379872799, + 0.016652924939990044, + 0.013172377832233906, + 0.015643710270524025, + 0.021638287231326103, + 0.01531563512980938, + 0.01751086674630642, + 0.013316476717591286, + 0.013131448067724705, + 0.01817186176776886, + 0.020760945975780487, + 0.020333874970674515, + 0.019877562299370766, + 0.026553615927696228, + 0.012246434576809406, + 0.014134122058749199, + 0.014954069629311562, + 0.010474123060703278, + 0.012890107929706573, + 0.012917044572532177, + 0.012956582009792328, + 0.011995451524853706, + 0.01155224908143282, + 0.012557277455925941, + 0.024188444018363953, + 0.006824136711657047, + 0.015347253531217575, + 0.012366542592644691, + 0.012933710590004921, + 0.007834057323634624, + 0.007982959970831871, + 0.007752535864710808, + 0.008515231311321259, + 0.005885843653231859, + 0.008512533269822598, + 0.005883620586246252, + 0.009059843607246876, + 0.006132006179541349, + 0.006878138054162264, + 0.00755495298653841, + 0.009225351735949516, + 0.005938895978033543, + 0.010118493810296059, + 0.007849710062146187, + 0.009234938770532608, + 0.004933924879878759, + 0.006169871427118778, + 0.010672678239643574, + 0.03129149228334427, + 0.010509774088859558, + 0.017476683482527733, + 0.010442504659295082, + 0.014747068285942078, + 0.014394337311387062, + 0.012243484146893024, + 0.004758480470627546, + 0.006753498688340187, + 0.004894787445664406, + 0.007202181499451399, + 0.003913997206836939, + 0.0045145852491259575, + 0.007725229486823082, + 0.005006643012166023, + 0.0044594681821763515, + 0.008141894824802876, + 0.006719532422721386, + 0.011008432134985924, + 0.005690502934157848, + 0.005040137097239494, + 0.007446337025612593, + 0.007022162899374962, + 0.018128864467144012, + 0.034234799444675446, + 0.03716843202710152, + 0.01206873171031475, + 0.028004240244627, + 0.013416584581136703, + 0.023714514449238777, + 0.01572345942258835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.6934084296226501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197722733020782 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.018197722733020782, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018197722733020782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06920722872018814, + 0.06532513350248337, + 0.060919370502233505, + 0.08796653151512146, + 0.10210400074720383, + 0.08514036238193512, + 0.22643837332725525, + 0.07094895094633102, + 0.15276844799518585, + 0.07918155193328857 + ], + "credit_map": [ + 0.010182691738009453, + 0.009190230630338192, + 0.00888025015592575, + 0.008926275186240673, + 0.006396409124135971, + 0.010332719422876835, + 0.007395156659185886, + 0.007903493009507656, + 0.00696570286527276, + 0.006325699854642153, + 0.00862323772162199, + 0.008451701141893864, + 0.0115891695022583, + 0.006054556462913752, + 0.0108023164793849, + 0.006512743420898914, + 0.006864138413220644, + 0.006372963078320026, + 0.005527616944164038, + 0.005231713410466909, + 0.013044574297964573, + 0.009909571148455143, + 0.008146369829773903, + 0.005822427105158567, + 0.009236471727490425, + 0.011733785271644592, + 0.0067401183769106865, + 0.00917462445795536, + 0.015350143425166607, + 0.01118661742657423, + 0.010547064244747162, + 0.013997701928019524, + 0.010944239795207977, + 0.012208991684019566, + 0.009824247099459171, + 0.014974015764892101, + 0.0184047631919384, + 0.013378563337028027, + 0.012641122564673424, + 0.00972805730998516, + 0.009752362966537476, + 0.009245662949979305, + 0.010878956876695156, + 0.013717465102672577, + 0.008848385885357857, + 0.010666128247976303, + 0.00760563975200057, + 0.014425755478441715, + 0.02260289527475834, + 0.07115622609853745, + 0.018608180806040764, + 0.023818442597985268, + 0.02375471591949463, + 0.03345881775021553, + 0.017916033044457436, + 0.015123060904443264, + 0.009846862405538559, + 0.009699290618300438, + 0.007637230213731527, + 0.009588400833308697, + 0.007156169041991234, + 0.008746616542339325, + 0.013107074424624443, + 0.005167303141206503, + 0.020284384489059448, + 0.015910696238279343, + 0.023943262174725533, + 0.02480766922235489, + 0.017569107934832573, + 0.011630876921117306, + 0.019796663895249367, + 0.01882578618824482, + 0.005537138786166906, + 0.027140580117702484, + 0.016033895313739777, + 0.006011528894305229, + 0.004817628767341375, + 0.007598087657243013, + 0.007523315493017435, + 0.004519375041127205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.018197722733020782 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a grand, cavernous hall of a Gothic stone building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the central characters a wise, bearded headmaster in dark blue robes and a young wizard with a lightning bolt scar visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting style black-and-white with dramatic contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth, circular panoramic shot starting from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are portraits of ancient figures on the walls shown moving and whispering within their frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762395858764648 + ], + "opposite_logprob_tokens": [ + -0.8262395858764648 + ], + "logprob_reward": -0.5762395858764648, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762395858764648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16221539676189423, + 0.1593276858329773, + 0.06717325001955032, + 0.06256648898124695, + 0.06397034227848053, + 0.07662799209356308, + 0.2021375149488449, + 0.08459445089101791, + 0.07730139046907425, + 0.0440855510532856 + ], + "credit_map": [ + 0.01435728371143341, + 0.01986687071621418, + 0.0210625771433115, + 0.02745942771434784, + 0.017654987052083015, + 0.017788270488381386, + 0.021711964160203934, + 0.022314021363854408, + 0.013094788417220116, + 0.013084772042930126, + 0.02113821916282177, + 0.02139781415462494, + 0.026347609236836433, + 0.021174214780330658, + 0.028719279915094376, + 0.014370981603860855, + 0.0069227442145347595, + 0.005815261974930763, + 0.007271988317370415, + 0.00997602753341198, + 0.007699768524616957, + 0.009680736809968948, + 0.012168964371085167, + 0.007637753151357174, + 0.007640593219548464, + 0.00927505362778902, + 0.005445350892841816, + 0.007304979953914881, + 0.009082235395908356, + 0.008060895837843418, + 0.008292033337056637, + 0.0074653481133282185, + 0.0069156126119196415, + 0.007290663663297892, + 0.005940326023846865, + 0.010815722867846489, + 0.009868038818240166, + 0.010182923637330532, + 0.0076364376582205296, + 0.005320614669471979, + 0.00850670412182808, + 0.015590937808156013, + 0.0051552122458815575, + 0.01032683253288269, + 0.009826558642089367, + 0.010237328708171844, + 0.007814673706889153, + 0.00916974525898695, + 0.020342404022812843, + 0.05953110381960869, + 0.014680422842502594, + 0.026644594967365265, + 0.02028064616024494, + 0.02744949981570244, + 0.017015991732478142, + 0.016192840412259102, + 0.012666351161897182, + 0.013040638528764248, + 0.007993828505277634, + 0.009903878904879093, + 0.009702201932668686, + 0.013426415622234344, + 0.012093136087059975, + 0.005768004804849625, + 0.007670651655644178, + 0.008022508583962917, + 0.004925896413624287, + 0.019935324788093567, + 0.00709576765075326, + 0.006034277379512787, + 0.009523970074951649, + 0.014092986471951008, + 0.0057661281898617744, + 0.007323629222810268, + 0.005894181784242392, + 0.005892719607800245, + 0.00473007233813405, + 0.0046609495766460896, + 0.00611699977889657, + 0.0037008682265877724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.5762395858764648 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_15/index_tensor([1842, 1842])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_15/index_tensor([1842, 1842])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..253ae9059a80d161b76999bea9151753cd9fb9f1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_15/index_tensor([1842, 1842])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled medium shot focuses with exquisite clarity on a cluster of vibrant, cup-shaped Caltha palustris flowers, their buttercup-yellow petals taking on a waxy, almost luminescent sheen against soft, diffuse reflections. The shallow depth of field renders the flowers and their central, bright yellow cluster in exquisitely sharp detail. Just beneath the crystalline water's surface, the plant's submerged, heart-shaped green leaves are also clearly visible. From the smooth texture of these leaves, a delicate and steady stream of tiny, shimmering oxygen bubbles meticulously detaches and rises gracefully, their slow, deliberate upward journey creating the only perceptible movement in the frame. As these bubbles rise like miniature pearls, the background completely dissolves into a soft, emerald blur of out-of-focus reeds and damp soil, creating a rich, natural color palette of yellows and greens that establishes an atmosphere of profound tranquility and contemplation, capturing a quiet moment of natural wonder.", + "seed": 9716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_15/index_1842_seed_9716_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017683747736737133, + "vlm_logprob": -0.017683747736737133, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.017683747736737133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003422595327720046 + ], + "opposite_logprob_tokens": [ + -6.003422737121582 + ], + "logprob_reward": -0.003422595327720046, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.003422595327720046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1648264229297638, + 0.0696435421705246, + 0.1456805020570755, + 0.11138343811035156, + 0.16694118082523346, + 0.08775800466537476, + 0.11773442476987839, + 0.05235578864812851, + 0.04129109904170036, + 0.042385637760162354 + ], + "credit_map": [ + 0.016430702060461044, + 0.01999840699136257, + 0.016569672152400017, + 0.01914781518280506, + 0.02648145519196987, + 0.016161242499947548, + 0.02832574024796486, + 0.021711386740207672, + 0.007353808265179396, + 0.007326381281018257, + 0.0068594529293477535, + 0.007452643010765314, + 0.012311426922678947, + 0.009282058104872704, + 0.008532379753887653, + 0.010525385849177837, + 0.014254947192966938, + 0.0207658763974905, + 0.009876118041574955, + 0.016578372567892075, + 0.0324120931327343, + 0.02133171074092388, + 0.012876225635409355, + 0.017585162073373795, + 0.009964332915842533, + 0.014929208904504776, + 0.01050505880266428, + 0.015240666456520557, + 0.01946241781115532, + 0.01330774649977684, + 0.010663599707186222, + 0.017310401424765587, + 0.011233990080654621, + 0.01688203774392605, + 0.008672736585140228, + 0.010030191391706467, + 0.05952489376068115, + 0.024647308513522148, + 0.015295539982616901, + 0.020654475316405296, + 0.009124052710831165, + 0.01606438308954239, + 0.0051225777715444565, + 0.012755685485899448, + 0.009916658513247967, + 0.01157295610755682, + 0.008143893443048, + 0.015057800337672234, + 0.013653459958732128, + 0.02072039246559143, + 0.009232084266841412, + 0.017650572583079338, + 0.013078832067549229, + 0.014252809807658195, + 0.008387532085180283, + 0.0207587368786335, + 0.009627425111830235, + 0.005460611544549465, + 0.004390046466141939, + 0.0068359836004674435, + 0.0058407532051205635, + 0.007214861921966076, + 0.005816934164613485, + 0.007169167045503855, + 0.004142069723457098, + 0.007034302223473787, + 0.005192728713154793, + 0.006971984636038542, + 0.005779573228210211, + 0.00484138447791338, + 0.0039995876140892506, + 0.003329469356685877, + 0.004826453048735857, + 0.004732500296086073, + 0.002947291824966669, + 0.00676548620685935, + 0.0056565734557807446, + 0.006071019452065229, + 0.0036392859183251858, + 0.0077470229007303715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.003422595327720046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590132102370262 + ], + "opposite_logprob_tokens": [ + -3.89559006690979 + ], + "logprob_reward": -0.020590132102370262, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590132102370262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15097805857658386, + 0.035239096730947495, + 0.09407612681388855, + 0.10855292528867722, + 0.14501871168613434, + 0.1309473216533661, + 0.0669889971613884, + 0.10563725978136063, + 0.10435273498296738, + 0.05820869654417038 + ], + "credit_map": [ + 0.014916397631168365, + 0.015012245625257492, + 0.01734054647386074, + 0.01811007596552372, + 0.022283993661403656, + 0.017259305343031883, + 0.016213009133934975, + 0.029842497780919075, + 0.0036555761471390724, + 0.004190576728433371, + 0.002829719800502062, + 0.00535520538687706, + 0.005953205283731222, + 0.004536800552159548, + 0.004290248733013868, + 0.00442776782438159, + 0.007523679174482822, + 0.009501796215772629, + 0.004352483432739973, + 0.0073266588151454926, + 0.03159733861684799, + 0.01363522931933403, + 0.007282020058482885, + 0.01285692397505045, + 0.010387581773102283, + 0.012562371790409088, + 0.007443651556968689, + 0.012761189602315426, + 0.026267709210515022, + 0.013717880472540855, + 0.010973405092954636, + 0.014439145103096962, + 0.010651743970811367, + 0.014726574532687664, + 0.007994151674211025, + 0.008290971629321575, + 0.052707232534885406, + 0.01969333365559578, + 0.011294029653072357, + 0.019660690799355507, + 0.013635613955557346, + 0.016153698787093163, + 0.005965138785541058, + 0.010445362888276577, + 0.042676955461502075, + 0.018649619072675705, + 0.009056171402335167, + 0.014364773407578468, + 0.007974560372531414, + 0.009746800176799297, + 0.004472354892641306, + 0.00679876608774066, + 0.009731738828122616, + 0.008469215594232082, + 0.004341961815953255, + 0.01545360591262579, + 0.01639111153781414, + 0.010872979648411274, + 0.0071595199406147, + 0.009769839234650135, + 0.018888764083385468, + 0.014698288403451443, + 0.015067864209413528, + 0.01278890110552311, + 0.01233131904155016, + 0.009236923418939114, + 0.008346347138285637, + 0.009994402527809143, + 0.029603730887174606, + 0.014358675107359886, + 0.009358689188957214, + 0.011122654192149639, + 0.008281009271740913, + 0.006154813338071108, + 0.0037720140535384417, + 0.005336917005479336, + 0.0074109225533902645, + 0.008594086393713951, + 0.006701032165437937, + 0.011957903392612934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.020590132102370262 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368956968188286 + ], + "opposite_logprob_tokens": [ + -3.408689498901367 + ], + "logprob_reward": -0.03368956968188286, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368956968188286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12621727585792542, + 0.06668858975172043, + 0.3260524272918701, + 0.06015167757868767, + 0.13074073195457458, + 0.05000771954655647, + 0.12627582252025604, + 0.018997201696038246, + 0.07977624982595444, + 0.015092306770384312 + ], + "credit_map": [ + 0.011380571871995926, + 0.014561633579432964, + 0.011605681851506233, + 0.016942504793405533, + 0.017293989658355713, + 0.013729056343436241, + 0.020504025742411613, + 0.020199786871671677, + 0.006454698741436005, + 0.008226693607866764, + 0.005403649993240833, + 0.010604428127408028, + 0.007802451960742474, + 0.008721251972019672, + 0.010043095797300339, + 0.009432320483028889, + 0.033323369920253754, + 0.031036851927638054, + 0.019505474716424942, + 0.0203838087618351, + 0.10510976612567902, + 0.0491136834025383, + 0.023066287860274315, + 0.044513165950775146, + 0.005239482969045639, + 0.00764339929446578, + 0.003724045120179653, + 0.007086887024343014, + 0.012444848194718361, + 0.007502262480556965, + 0.007444046903401613, + 0.009066702798008919, + 0.008559265173971653, + 0.01396553311496973, + 0.006451369728893042, + 0.007481708191335201, + 0.05055086314678192, + 0.019463714212179184, + 0.010211963206529617, + 0.014056305401027203, + 0.004013626370579004, + 0.004572563339024782, + 0.0023697703145444393, + 0.0037805477622896433, + 0.017482604831457138, + 0.007659862749278545, + 0.003676924156025052, + 0.006451817695051432, + 0.012444204650819302, + 0.011274885386228561, + 0.004729236010462046, + 0.0061045982874929905, + 0.04974903166294098, + 0.021686125546693802, + 0.007624544203281403, + 0.012663193978369236, + 0.0025752463843673468, + 0.0019230344332754612, + 0.0013361958554014564, + 0.0018476526020094752, + 0.004559204448014498, + 0.0024266219697892666, + 0.0020851914305239916, + 0.0022440520115196705, + 0.007737109437584877, + 0.007756536360830069, + 0.006314408034086227, + 0.007980640046298504, + 0.024755315855145454, + 0.011036749929189682, + 0.006219030357897282, + 0.007976455613970757, + 0.00220583938062191, + 0.001448040478862822, + 0.0012042203452438116, + 0.0019326085457578301, + 0.0023490386083722115, + 0.002129526576027274, + 0.0015315947821363807, + 0.002291437005624175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.03368956968188286 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253335177898407 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.01253335177898407, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253335177898407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13030382990837097, + 0.05349612608551979, + 0.11863170564174652, + 0.051809512078762054, + 0.30271491408348083, + 0.09169650077819824, + 0.06363924592733383, + 0.08607295900583267, + 0.04058033600449562, + 0.06105484440922737 + ], + "credit_map": [ + 0.01411210186779499, + 0.013438389636576176, + 0.012236284092068672, + 0.014069131575524807, + 0.018027210608124733, + 0.014893569983541965, + 0.02633567899465561, + 0.017191482707858086, + 0.005099400877952576, + 0.008518505841493607, + 0.003993331454694271, + 0.006504948250949383, + 0.009111698716878891, + 0.006798445247113705, + 0.005700437352061272, + 0.007769369054585695, + 0.01934901438653469, + 0.014678568579256535, + 0.009087304584681988, + 0.009578724391758442, + 0.02283024787902832, + 0.016190797090530396, + 0.014147140085697174, + 0.01276993378996849, + 0.003973702900111675, + 0.004271899815648794, + 0.0027744926046580076, + 0.004256472457200289, + 0.015936344861984253, + 0.007805734407156706, + 0.004318857099860907, + 0.008472013287246227, + 0.023456547409296036, + 0.030968226492404938, + 0.0202008169144392, + 0.021303337067365646, + 0.10370761156082153, + 0.043592244386672974, + 0.026618773117661476, + 0.032867394387722015, + 0.00929150078445673, + 0.011392979882657528, + 0.006175550632178783, + 0.009513277560472488, + 0.01778663881123066, + 0.011360002681612968, + 0.01107955351471901, + 0.015097010880708694, + 0.007047581952065229, + 0.005819499958306551, + 0.006167849991470575, + 0.005727893207222223, + 0.01408860832452774, + 0.009341811761260033, + 0.004956104792654514, + 0.010489908047020435, + 0.011067776009440422, + 0.006999005563557148, + 0.008982092142105103, + 0.00904830638319254, + 0.01747172139585018, + 0.0131605239585042, + 0.007789094932377338, + 0.011554449796676636, + 0.005437512882053852, + 0.004117663484066725, + 0.003526493674144149, + 0.0036941594444215298, + 0.008554140105843544, + 0.004971223417669535, + 0.002822765614837408, + 0.007456385530531406, + 0.008074161596596241, + 0.006320440210402012, + 0.0036426219157874584, + 0.004840100649744272, + 0.01220003329217434, + 0.0081085916608572, + 0.006427808664739132, + 0.011441093869507313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.01253335177898407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183089792728424 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183089792728424, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183089792728424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16820089519023895, + 0.07901303470134735, + 0.07505013048648834, + 0.12151089310646057, + 0.11577927321195602, + 0.1398811638355255, + 0.10892732441425323, + 0.06438939273357391, + 0.0882406234741211, + 0.0390072762966156 + ], + "credit_map": [ + 0.016854573041200638, + 0.021525289863348007, + 0.016623681411147118, + 0.019108450040221214, + 0.02821088582277298, + 0.020031584426760674, + 0.018778610974550247, + 0.027067819610238075, + 0.006595870014280081, + 0.008112847805023193, + 0.0057601830922067165, + 0.010926004499197006, + 0.010140628553926945, + 0.009594026021659374, + 0.015665719285607338, + 0.012217762880027294, + 0.008728554472327232, + 0.006000453606247902, + 0.005899769254028797, + 0.005707216914743185, + 0.01332449447363615, + 0.010419233702123165, + 0.008382420055568218, + 0.016587981954216957, + 0.008775013498961926, + 0.0144265191629529, + 0.0069246599450707436, + 0.013758208602666855, + 0.02713286690413952, + 0.01616661250591278, + 0.009649630635976791, + 0.02467738278210163, + 0.01041180919855833, + 0.007633306551724672, + 0.008569272235035896, + 0.009873975068330765, + 0.025457730516791344, + 0.015524245798587799, + 0.012343336828052998, + 0.02596559375524521, + 0.011128129437565804, + 0.011101617477834225, + 0.00674046203494072, + 0.008524967357516289, + 0.04940897598862648, + 0.021410569548606873, + 0.012224428355693817, + 0.019342029467225075, + 0.014002503827214241, + 0.013524303212761879, + 0.0071584065444767475, + 0.013156316243112087, + 0.022678490728139877, + 0.015700217336416245, + 0.007121613249182701, + 0.015585478395223618, + 0.007901826873421669, + 0.006289652548730373, + 0.005445990711450577, + 0.006715340539813042, + 0.014425844885408878, + 0.009469250217080116, + 0.005483039189130068, + 0.008658441714942455, + 0.0099705271422863, + 0.006667214911431074, + 0.004709936212748289, + 0.005378440488129854, + 0.030204322189092636, + 0.01318526454269886, + 0.007686327677220106, + 0.0104385856539011, + 0.0035310753155499697, + 0.005814111325889826, + 0.005566942039877176, + 0.006349305622279644, + 0.005757751874625683, + 0.005127638578414917, + 0.003050429979339242, + 0.003810020163655281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.018183089792728424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot focuses with exquisite clarity on a cluster of vibrant, cup-shaped Caltha palustris flowers, their buttercup-yellow petals taking on a waxy, almost luminescent sheen against soft, diffuse reflections. The shallow depth of field renders the flowers and their central, bright yellow cluster in exquisitely sharp detail. Just beneath the crystalline water's surface, the plant's submerged, heart-shaped green leaves are also clearly visible. From the smooth texture of these leaves, a delicate and steady stream of tiny, shimmering oxygen bubbles meticulously detaches and rises gracefully, their slow, deliberate upward journey creating the only perceptible movement in the frame. As these bubbles rise like miniature pearls, the background completely dissolves into a soft, emerald blur of out-of-focus reeds and damp soil, creating a rich, natural color palette of yellows and greens that establishes an atmosphere of profound tranquility and contemplation, capturing a quiet moment of natural wonder.", + "seed": 9717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_15/index_1842_seed_9717_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019481420842930676, + "vlm_logprob": -0.019481420842930676, + "vlm_yes_no_margin": 4.249999952316284, + "vlm_token_credit": -0.019481420842930676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003547448432072997 + ], + "opposite_logprob_tokens": [ + -6.128547668457031 + ], + "logprob_reward": -0.003547448432072997, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.003547448432072997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.135793998837471, + 0.18255184590816498, + 0.031268466264009476, + 0.16055719554424286, + 0.09453383833169937, + 0.2034061849117279, + 0.05018433928489685, + 0.07560320943593979, + 0.03195362910628319, + 0.03414732590317726 + ], + "credit_map": [ + 0.014152891002595425, + 0.02184775471687317, + 0.01348376739770174, + 0.01772412844002247, + 0.015149109065532684, + 0.016523802652955055, + 0.020672835409641266, + 0.016239702701568604, + 0.01422743871808052, + 0.015493501909077168, + 0.011839203536510468, + 0.04034411907196045, + 0.052103493362665176, + 0.02468498796224594, + 0.01305407751351595, + 0.010805011726915836, + 0.004358678124845028, + 0.004036309197545052, + 0.004595707170665264, + 0.004527443554252386, + 0.0038500011432915926, + 0.0036281482316553593, + 0.0033791621681302786, + 0.0028930159751325846, + 0.020610077306628227, + 0.032172176986932755, + 0.014531608670949936, + 0.01898065395653248, + 0.03134537860751152, + 0.019575703889131546, + 0.012683739885687828, + 0.01065785251557827, + 0.012220152653753757, + 0.014245879836380482, + 0.011450070887804031, + 0.010451426729559898, + 0.01748071052134037, + 0.010987061075866222, + 0.00940663367509842, + 0.008291898295283318, + 0.026560483500361443, + 0.02233194001019001, + 0.021930620074272156, + 0.027190299704670906, + 0.04746029153466225, + 0.025109844282269478, + 0.019433902576565742, + 0.01338877435773611, + 0.006413137540221214, + 0.005253361538052559, + 0.006379575934261084, + 0.007428384386003017, + 0.007651214487850666, + 0.0060892351903021336, + 0.006008218973875046, + 0.004961209371685982, + 0.012615358456969261, + 0.01221553049981594, + 0.009126173332333565, + 0.007973580621182919, + 0.011126781813800335, + 0.00939159095287323, + 0.006297484505921602, + 0.006856698077172041, + 0.003961007576435804, + 0.0030887832399457693, + 0.005983258597552776, + 0.004598911851644516, + 0.0030286225955933332, + 0.004893885459750891, + 0.003384709358215332, + 0.0030144494958221912, + 0.0037961762864142656, + 0.0037919774185866117, + 0.0062597980722785, + 0.0034567546099424362, + 0.007163753267377615, + 0.0041818767786026, + 0.0029934223275631666, + 0.002503566211089492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.003547448432072997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981428988277912 + ], + "opposite_logprob_tokens": [ + -3.5298142433166504 + ], + "logprob_reward": -0.02981428988277912, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981428988277912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23097048699855804, + 0.0822213813662529, + 0.08495708554983139, + 0.020230291411280632, + 0.12341765314340591, + 0.23572203516960144, + 0.053771745413541794, + 0.13035450875759125, + 0.016412189230322838, + 0.02194269374012947 + ], + "credit_map": [ + 0.0268898606300354, + 0.022947313264012337, + 0.03987376019358635, + 0.04069957509636879, + 0.01766802743077278, + 0.028352832421660423, + 0.032330114394426346, + 0.02220897749066353, + 0.00857105664908886, + 0.009340392425656319, + 0.006893275771290064, + 0.019355595111846924, + 0.014437062665820122, + 0.009945367462933064, + 0.00833736639469862, + 0.005341252312064171, + 0.006572195794433355, + 0.008763965219259262, + 0.01447701919823885, + 0.01203475333750248, + 0.007307872641831636, + 0.010185867547988892, + 0.01439611054956913, + 0.011219298467040062, + 0.0031150502618402243, + 0.0021214468870311975, + 0.0024517117999494076, + 0.0035400150809437037, + 0.0027285481337457895, + 0.0026957597583532333, + 0.0022989041171967983, + 0.0012788535095751286, + 0.01735817827284336, + 0.0193635281175375, + 0.01680769957602024, + 0.01442668680101633, + 0.019500579684972763, + 0.015483381226658821, + 0.009343931451439857, + 0.01113365963101387, + 0.028795108199119568, + 0.02575431764125824, + 0.028896301984786987, + 0.0332302451133728, + 0.04919670149683952, + 0.029054706916213036, + 0.02441445179283619, + 0.016380181536078453, + 0.009027133695781231, + 0.005062047857791185, + 0.007541740778833628, + 0.006388860754668713, + 0.009043712168931961, + 0.006510057486593723, + 0.005620290525257587, + 0.004577898886054754, + 0.0229527335613966, + 0.01870003342628479, + 0.019238676875829697, + 0.016773764044046402, + 0.016549261286854744, + 0.014522562734782696, + 0.010714567266404629, + 0.010902896523475647, + 0.002360777696594596, + 0.0024165024515241385, + 0.002435371046885848, + 0.0014324492076411843, + 0.0026827496476471424, + 0.001974883023649454, + 0.0014860493829473853, + 0.0016234060749411583, + 0.0020591015927493572, + 0.002152937464416027, + 0.005778157152235508, + 0.0022614519111812115, + 0.002842529909685254, + 0.003248246619477868, + 0.002283911919221282, + 0.0013163556577637792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.02981428988277912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11952299624681473, + 0.45362037420272827, + 0.02147778496146202, + 0.07319362461566925, + 0.032192811369895935, + 0.1097177192568779, + 0.028454504907131195, + 0.13111528754234314, + 0.017491092905402184, + 0.013213817030191422 + ], + "credit_map": [ + 0.014421680010855198, + 0.013318103738129139, + 0.02241828478872776, + 0.019505051895976067, + 0.01699639856815338, + 0.014471019618213177, + 0.009749114513397217, + 0.00864334125071764, + 0.05072861164808273, + 0.037289250642061234, + 0.03919932618737221, + 0.1117212101817131, + 0.10045256465673447, + 0.0546538345515728, + 0.03572279214859009, + 0.02385278418660164, + 0.0024572068359702826, + 0.0030404147692024708, + 0.002620237646624446, + 0.0028204743284732103, + 0.0019864714704453945, + 0.0026796399615705013, + 0.0035970802418887615, + 0.0022762587759643793, + 0.010526100173592567, + 0.012457892298698425, + 0.008056367747485638, + 0.010216951370239258, + 0.011315073817968369, + 0.009107666090130806, + 0.00661127083003521, + 0.004902305081486702, + 0.005091672297567129, + 0.004358322359621525, + 0.004746666643768549, + 0.003727362025529146, + 0.003958011046051979, + 0.003781002014875412, + 0.0034273904748260975, + 0.003102383343502879, + 0.015424644574522972, + 0.011914490722119808, + 0.011184767819941044, + 0.013437507674098015, + 0.028340110555291176, + 0.013364631682634354, + 0.009437225759029388, + 0.006614341400563717, + 0.005021179094910622, + 0.00319491745904088, + 0.003758990438655019, + 0.004717161878943443, + 0.0023696469143033028, + 0.0036229772958904505, + 0.003421741770580411, + 0.0023478928487747908, + 0.021165577694773674, + 0.015443301759660244, + 0.02496115118265152, + 0.019168302416801453, + 0.014122148044407368, + 0.014779657125473022, + 0.011505566537380219, + 0.009969587437808514, + 0.002809034427627921, + 0.002462172880768776, + 0.002680836245417595, + 0.001882886397652328, + 0.001707175513729453, + 0.0020385989919304848, + 0.0020751834381371737, + 0.0018352060578763485, + 0.0014795869356021285, + 0.0014667174546048045, + 0.0022047904785722494, + 0.0015463406452909112, + 0.002468222053721547, + 0.001618234091438353, + 0.0012436996912583709, + 0.0011862253304570913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794733971357346 + ], + "opposite_logprob_tokens": [ + -3.529794692993164 + ], + "logprob_reward": -0.029794733971357346, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794733971357346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20716330409049988, + 0.0614422932267189, + 0.03818323835730553, + 0.1707552969455719, + 0.21580025553703308, + 0.045997727662324905, + 0.14371702075004578, + 0.0735139548778534, + 0.014082398265600204, + 0.029344497248530388 + ], + "credit_map": [ + 0.025860145688056946, + 0.019483640789985657, + 0.039910223335027695, + 0.036212410777807236, + 0.019144833087921143, + 0.026660330593585968, + 0.025502799078822136, + 0.014388931915163994, + 0.006113150157034397, + 0.005099495407193899, + 0.005109768360853195, + 0.01786164753139019, + 0.010222265496850014, + 0.008906805887818336, + 0.004684928338974714, + 0.0034442292526364326, + 0.003897411050274968, + 0.005901701282709837, + 0.004844998475164175, + 0.0064768111333251, + 0.004688819870352745, + 0.004098706413060427, + 0.004928047768771648, + 0.0033467451576143503, + 0.02199058048427105, + 0.03522734344005585, + 0.015510193072259426, + 0.026697533205151558, + 0.022631626576185226, + 0.020265191793441772, + 0.016209648922085762, + 0.012223176658153534, + 0.035464588552713394, + 0.03645119071006775, + 0.024045037105679512, + 0.021010175347328186, + 0.036629896610975266, + 0.02697088196873665, + 0.01706511713564396, + 0.018163366243243217, + 0.006318643689155579, + 0.006205963436514139, + 0.006232419516891241, + 0.0054536499083042145, + 0.00908045656979084, + 0.0054680053144693375, + 0.004065297544002533, + 0.0031732949428260326, + 0.016327856108546257, + 0.015069084241986275, + 0.01928078942000866, + 0.022313281893730164, + 0.020561067387461662, + 0.016686899587512016, + 0.021077007055282593, + 0.012401032261550426, + 0.013060751371085644, + 0.009221093729138374, + 0.010435869917273521, + 0.011399801820516586, + 0.007750522345304489, + 0.008001463487744331, + 0.007350102532655001, + 0.006294349208474159, + 0.0018875155365094543, + 0.0019184164702892303, + 0.0018367335433140397, + 0.0017406223341822624, + 0.0016913600265979767, + 0.001706465263850987, + 0.0017608378548175097, + 0.0015404478181153536, + 0.004942045547068119, + 0.002789359772577882, + 0.00550937931984663, + 0.003203319851309061, + 0.004100320860743523, + 0.003996431361883879, + 0.00225443160161376, + 0.0025492063723504543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.029794733971357346 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185313791036606 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185313791036606, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018185313791036606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11379898339509964, + 0.11574704945087433, + 0.02047812193632126, + 0.31239110231399536, + 0.03801339864730835, + 0.22551926970481873, + 0.06677120178937912, + 0.06866496056318283, + 0.019026365131139755, + 0.01958954706788063 + ], + "credit_map": [ + 0.014584979973733425, + 0.013837728649377823, + 0.01722930185496807, + 0.013759822584688663, + 0.016868116334080696, + 0.014399799518287182, + 0.013588008470833302, + 0.009531230665743351, + 0.011521100997924805, + 0.01808873564004898, + 0.011435321532189846, + 0.013905529864132404, + 0.03065827675163746, + 0.014913560822606087, + 0.0068747373297810555, + 0.008349789306521416, + 0.0028219977393746376, + 0.0017113283975049853, + 0.0030668743420392275, + 0.00402456521987915, + 0.0016582414973527193, + 0.002571806078776717, + 0.0030125160701572895, + 0.0016107945702970028, + 0.04002033919095993, + 0.05698661506175995, + 0.03391793742775917, + 0.04227689653635025, + 0.0459330677986145, + 0.03959617763757706, + 0.030912576243281364, + 0.022747479379177094, + 0.007384595461189747, + 0.004974474664777517, + 0.0031806151382625103, + 0.004349090624600649, + 0.006011756602674723, + 0.004794743377715349, + 0.0029658537823706865, + 0.004352270159870386, + 0.029492856934666634, + 0.02517819032073021, + 0.021914131939411163, + 0.029164884239435196, + 0.056632015854120255, + 0.02813742496073246, + 0.01989726349711418, + 0.015102503821253777, + 0.00839908979833126, + 0.006701611448079348, + 0.009285126812756062, + 0.008084742352366447, + 0.010550993494689465, + 0.00827210396528244, + 0.009227910079061985, + 0.0062496243044734, + 0.009321890771389008, + 0.009477056562900543, + 0.007603143807500601, + 0.009025996550917625, + 0.011757428757846355, + 0.008324120193719864, + 0.007299285847693682, + 0.005856038071215153, + 0.0031590410508215427, + 0.002692614682018757, + 0.0025272772181779146, + 0.002364258747547865, + 0.0021566913928836584, + 0.002612863900139928, + 0.0017773674335330725, + 0.001736250938847661, + 0.002603106666356325, + 0.0017580423736944795, + 0.004720643628388643, + 0.0024236608296632767, + 0.0016145118279382586, + 0.0025847721844911575, + 0.0022958905901759863, + 0.001588919898495078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.018185313791036606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot focuses with exquisite clarity on a cluster of vibrant, cup-shaped Caltha palustris flowers, their buttercup-yellow petals taking on a waxy, almost luminescent sheen against soft, diffuse reflections. The shallow depth of field renders the flowers and their central, bright yellow cluster in exquisitely sharp detail. Just beneath the crystalline water's surface, the plant's submerged, heart-shaped green leaves are also clearly visible. From the smooth texture of these leaves, a delicate and steady stream of tiny, shimmering oxygen bubbles meticulously detaches and rises gracefully, their slow, deliberate upward journey creating the only perceptible movement in the frame. As these bubbles rise like miniature pearls, the background completely dissolves into a soft, emerald blur of out-of-focus reeds and damp soil, creating a rich, natural color palette of yellows and greens that establishes an atmosphere of profound tranquility and contemplation, capturing a quiet moment of natural wonder.", + "seed": 9718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_15/index_1842_seed_9718_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030237568356096745, + "vlm_logprob": -0.030237568356096745, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.030237568356096745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0056941211223602295 + ], + "opposite_logprob_tokens": [ + -5.3806939125061035 + ], + "logprob_reward": -0.0056941211223602295, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0056941211223602295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1021924689412117, + 0.07112076878547668, + 0.07043818384408951, + 0.20778468251228333, + 0.09369878470897675, + 0.27538275718688965, + 0.00961246807128191, + 0.03678330406546593, + 0.054456863552331924, + 0.0785297155380249 + ], + "credit_map": [ + 0.02966308780014515, + 0.00692841736599803, + 0.008768969215452671, + 0.006325133610516787, + 0.005703977774828672, + 0.02464263141155243, + 0.009956852532923222, + 0.010203398764133453, + 0.010668779723346233, + 0.007821863517165184, + 0.008942198939621449, + 0.013311688788235188, + 0.007588621228933334, + 0.0072838179767131805, + 0.007931742817163467, + 0.0075720506720244884, + 0.008730953559279442, + 0.005538232158869505, + 0.009309458546340466, + 0.006362113170325756, + 0.010991347953677177, + 0.013714805245399475, + 0.008203190751373768, + 0.007588077802211046, + 0.03344855085015297, + 0.009828099049627781, + 0.017643794417381287, + 0.01402181014418602, + 0.04276148974895477, + 0.05435768887400627, + 0.017107980325818062, + 0.018615253269672394, + 0.020805664360523224, + 0.004774472676217556, + 0.006897249724715948, + 0.004190661013126373, + 0.007306907791644335, + 0.023804446682333946, + 0.008039725013077259, + 0.01787964627146721, + 0.06610629707574844, + 0.013143103569746017, + 0.02709304727613926, + 0.01610959880053997, + 0.022681323811411858, + 0.06268100440502167, + 0.025169240310788155, + 0.04239911213517189, + 0.0017555608646944165, + 0.0005715681472793221, + 0.0008991971844807267, + 0.0013202863046899438, + 0.0008421496022492647, + 0.0015970793319866061, + 0.001048748497851193, + 0.0015778762754052877, + 0.009218393824994564, + 0.0024050306528806686, + 0.0020929172169417143, + 0.002100974554196, + 0.002621992025524378, + 0.010436947457492352, + 0.003531254595145583, + 0.004375791642814875, + 0.01472401898354292, + 0.0024861199781298637, + 0.002440826268866658, + 0.0023652201052755117, + 0.0044180345721542835, + 0.02049538865685463, + 0.002883192850276828, + 0.004644060507416725, + 0.014945597387850285, + 0.00759841175749898, + 0.011499962769448757, + 0.008100665174424648, + 0.007602119352668524, + 0.012647234834730625, + 0.008797263726592064, + 0.007338448893278837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.0056941211223602295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207450106739998 + ], + "opposite_logprob_tokens": [ + -2.8120744228363037 + ], + "logprob_reward": -0.06207450106739998, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207450106739998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1578986644744873, + 0.04927414655685425, + 0.03797731176018715, + 0.467948317527771, + 0.12629608809947968, + 0.05091055482625961, + 0.011092282831668854, + 0.019312197342514992, + 0.0451020821928978, + 0.03418824076652527 + ], + "credit_map": [ + 0.04874312877655029, + 0.010391615331172943, + 0.01082111056894064, + 0.008775727823376656, + 0.008511772379279137, + 0.045985233038663864, + 0.01032128557562828, + 0.014348804950714111, + 0.011549694463610649, + 0.0034001022577285767, + 0.005821674130856991, + 0.004123906139284372, + 0.004656323231756687, + 0.0059025525115430355, + 0.005518230609595776, + 0.008301668800413609, + 0.005908086430281401, + 0.003308265469968319, + 0.004087153356522322, + 0.003223385661840439, + 0.00651943776756525, + 0.006855261977761984, + 0.004031955264508724, + 0.004043768160045147, + 0.09001962840557098, + 0.028739379718899727, + 0.034635040909051895, + 0.03859269618988037, + 0.06779628992080688, + 0.1058211699128151, + 0.053232207894325256, + 0.04911194369196892, + 0.01825205236673355, + 0.007209026720374823, + 0.011410090140998363, + 0.007112443912774324, + 0.01538704801350832, + 0.0395258367061615, + 0.012913951650261879, + 0.01448563951998949, + 0.013549243099987507, + 0.002431456232443452, + 0.004735417664051056, + 0.002928102156147361, + 0.003404121147468686, + 0.013806418515741825, + 0.0033003906719386578, + 0.006755411624908447, + 0.0017399017233401537, + 0.0011695901630446315, + 0.0009410128695890307, + 0.0012743924744427204, + 0.0014544788282364607, + 0.0015243468806147575, + 0.0013915997697040439, + 0.0015969619853422046, + 0.0025309992488473654, + 0.0017168214544653893, + 0.002705349586904049, + 0.0023014701437205076, + 0.0019206167198717594, + 0.003299183677881956, + 0.0025227596051990986, + 0.0023149969056248665, + 0.00696393009275198, + 0.0025008758530020714, + 0.005486877169460058, + 0.003910135943442583, + 0.004146542400121689, + 0.010445930063724518, + 0.00537906913086772, + 0.006268729455769062, + 0.006376818288117647, + 0.003340538591146469, + 0.005083141382783651, + 0.0034418876748532057, + 0.0032825269736349583, + 0.005190626252442598, + 0.004156057257205248, + 0.003316648071631789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.06207450106739998 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048649102449417114 + ], + "opposite_logprob_tokens": [ + -3.0486490726470947 + ], + "logprob_reward": -0.048649102449417114, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048649102449417114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07473961263895035, + 0.09824654459953308, + 0.1297096163034439, + 0.28120478987693787, + 0.13193488121032715, + 0.1394006609916687, + 0.015862837433815002, + 0.015533032827079296, + 0.05791319161653519, + 0.05545489862561226 + ], + "credit_map": [ + 0.014634300954639912, + 0.008533144369721413, + 0.008310619741678238, + 0.007061331067234278, + 0.007738277316093445, + 0.008073695003986359, + 0.009187793359160423, + 0.011200442910194397, + 0.021237097680568695, + 0.007483468856662512, + 0.008654529228806496, + 0.015702787786722183, + 0.007834100164473057, + 0.008741241879761219, + 0.010394067503511906, + 0.018199235200881958, + 0.021805712953209877, + 0.009695473127067089, + 0.017058851197361946, + 0.009112519212067127, + 0.01847679913043976, + 0.026393484324216843, + 0.01432815007865429, + 0.012838606722652912, + 0.06368378549814224, + 0.01207091473042965, + 0.017343418672680855, + 0.013891953974962234, + 0.042960140854120255, + 0.09621462970972061, + 0.014865130186080933, + 0.020174764096736908, + 0.018985994160175323, + 0.007937666028738022, + 0.01153191551566124, + 0.006391770206391811, + 0.016273263841867447, + 0.043821219354867935, + 0.011932941153645515, + 0.015060100704431534, + 0.03524867445230484, + 0.006449922453612089, + 0.0136981550604105, + 0.007978830486536026, + 0.011467601172626019, + 0.033950887620449066, + 0.011903626844286919, + 0.018702948465943336, + 0.0024949752259999514, + 0.0009738656226545572, + 0.0014886943390592933, + 0.0014749186811968684, + 0.0027783310506492853, + 0.0018012735527008772, + 0.0017720649484544992, + 0.0030787126161158085, + 0.0022721621207892895, + 0.00172523595392704, + 0.0016523441299796104, + 0.0018940282752737403, + 0.0016253070207312703, + 0.0017418178031221032, + 0.0013699749251827598, + 0.0032521600369364023, + 0.014095309190452099, + 0.003431746270507574, + 0.002563678892329335, + 0.002954561496153474, + 0.005703924223780632, + 0.02007078193128109, + 0.0033432820346206427, + 0.005749898497015238, + 0.010548134334385395, + 0.006106415763497353, + 0.007483316585421562, + 0.006415803916752338, + 0.006015955004841089, + 0.009013877250254154, + 0.005098629742860794, + 0.004772759974002838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.048649102449417114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4317372739315033, + 0.034470733255147934, + 0.09119527786970139, + 0.10888268798589706, + 0.0269341841340065, + 0.09716518968343735, + 0.010027998127043247, + 0.039599016308784485, + 0.0571608692407608, + 0.10282681882381439 + ], + "credit_map": [ + 0.13829568028450012, + 0.019946061074733734, + 0.02153998240828514, + 0.013681801967322826, + 0.02728400006890297, + 0.15914253890514374, + 0.01629427634179592, + 0.035552896559238434, + 0.005333337001502514, + 0.003741955617442727, + 0.004122781101614237, + 0.004328086506575346, + 0.003624269040301442, + 0.0040406836196780205, + 0.0037575066089630127, + 0.005522112362086773, + 0.010852743871510029, + 0.0074715339578688145, + 0.011822482571005821, + 0.008064602501690388, + 0.014708271250128746, + 0.016976991668343544, + 0.010184413753449917, + 0.01111423410475254, + 0.020527029410004616, + 0.006719288881868124, + 0.0094728609547019, + 0.009256999008357525, + 0.018001114949584007, + 0.022857805714011192, + 0.01257631927728653, + 0.009471266530454159, + 0.006158791948109865, + 0.002322235843166709, + 0.002005781978368759, + 0.002424348844215274, + 0.002115584909915924, + 0.0029346125666052103, + 0.0034417430870234966, + 0.0055310847237706184, + 0.022586816921830177, + 0.004945299122482538, + 0.010652665048837662, + 0.0062146843411028385, + 0.009034674614667892, + 0.020390277728438377, + 0.00927239004522562, + 0.014068381860852242, + 0.0015098238363862038, + 0.0008449509623460472, + 0.0007557524368166924, + 0.0011588569032028317, + 0.0012240954674780369, + 0.0012530951062217355, + 0.0014240520540624857, + 0.0018573713023215532, + 0.008234741166234016, + 0.0029652495868504047, + 0.004254213534295559, + 0.0036165169440209866, + 0.0031671614851802588, + 0.009678778238594532, + 0.003939464688301086, + 0.0037428883370012045, + 0.00972240511327982, + 0.003245563944801688, + 0.00629903981462121, + 0.004573811311274767, + 0.005204235669225454, + 0.0144630903378129, + 0.006217917427420616, + 0.007434803061187267, + 0.019052637740969658, + 0.010434535332024097, + 0.015008160844445229, + 0.010985400527715683, + 0.009929162450134754, + 0.016344357281923294, + 0.011282376013696194, + 0.00979018583893776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.014192949049174786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577168092131615 + ], + "opposite_logprob_tokens": [ + -3.8955771923065186 + ], + "logprob_reward": -0.020577168092131615, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577168092131615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07004672288894653, + 0.05560509115457535, + 0.03246913105249405, + 0.4587692618370056, + 0.02149791270494461, + 0.03341473639011383, + 0.01883385330438614, + 0.049612127244472504, + 0.18116606771945953, + 0.07858511060476303 + ], + "credit_map": [ + 0.01610657013952732, + 0.008364247158169746, + 0.0050413524731993675, + 0.005065683741122484, + 0.006881261710077524, + 0.01219614502042532, + 0.00659328605979681, + 0.009798187762498856, + 0.011972276493906975, + 0.005139213055372238, + 0.005320229101926088, + 0.007662290707230568, + 0.004703274462372065, + 0.003969254437834024, + 0.005112289451062679, + 0.01172626856714487, + 0.006163100246340036, + 0.002596564358100295, + 0.0037318093236535788, + 0.002659828867763281, + 0.006142876576632261, + 0.004797152243554592, + 0.003291997592896223, + 0.003085804870352149, + 0.1274661421775818, + 0.020428281277418137, + 0.02027064934372902, + 0.017422065138816833, + 0.0481230765581131, + 0.17789964377880096, + 0.019023291766643524, + 0.028136109933257103, + 0.003728521289303899, + 0.0017253587720915675, + 0.002672274364158511, + 0.0021767730358988047, + 0.0019607762806117535, + 0.0038595052901655436, + 0.0025514590088278055, + 0.002823245245963335, + 0.007735582534223795, + 0.0016969252610579133, + 0.002939032856374979, + 0.0019666904117912054, + 0.003956836182624102, + 0.007395409978926182, + 0.0032397371251136065, + 0.004484524950385094, + 0.003011935856193304, + 0.0014771262649446726, + 0.0015384377911686897, + 0.001617375179193914, + 0.0020812801085412502, + 0.002580110915005207, + 0.003126890165731311, + 0.003400699235498905, + 0.01194769237190485, + 0.0029904525727033615, + 0.002607521368190646, + 0.004221376031637192, + 0.003288992214947939, + 0.01719449833035469, + 0.0033418023958802223, + 0.00401979498565197, + 0.03804761543869972, + 0.009763738140463829, + 0.012926124036312103, + 0.010960026644170284, + 0.019077960401773453, + 0.05476799234747887, + 0.016744161024689674, + 0.018878471106290817, + 0.015287778340280056, + 0.007933896966278553, + 0.011121878400444984, + 0.007961264811456203, + 0.008403586223721504, + 0.012516716495156288, + 0.008299713023006916, + 0.007060282863676548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.020577168092131615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot focuses with exquisite clarity on a cluster of vibrant, cup-shaped Caltha palustris flowers, their buttercup-yellow petals taking on a waxy, almost luminescent sheen against soft, diffuse reflections. The shallow depth of field renders the flowers and their central, bright yellow cluster in exquisitely sharp detail. Just beneath the crystalline water's surface, the plant's submerged, heart-shaped green leaves are also clearly visible. From the smooth texture of these leaves, a delicate and steady stream of tiny, shimmering oxygen bubbles meticulously detaches and rises gracefully, their slow, deliberate upward journey creating the only perceptible movement in the frame. As these bubbles rise like miniature pearls, the background completely dissolves into a soft, emerald blur of out-of-focus reeds and damp soil, creating a rich, natural color palette of yellows and greens that establishes an atmosphere of profound tranquility and contemplation, capturing a quiet moment of natural wonder.", + "seed": 9719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_15/index_1842_seed_9719_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02152318386361003, + "vlm_logprob": -0.02152318386361003, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02152318386361003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003542815800756216 + ], + "opposite_logprob_tokens": [ + -6.003542900085449 + ], + "logprob_reward": -0.003542815800756216, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.003542815800756216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20702773332595825, + 0.0858745276927948, + 0.3130851089954376, + 0.11238712072372437, + 0.0634637102484703, + 0.043771546334028244, + 0.05341918766498566, + 0.03719967231154442, + 0.034188225865364075, + 0.04958324506878853 + ], + "credit_map": [ + 0.02385335974395275, + 0.028120508417487144, + 0.01744743436574936, + 0.0311699490994215, + 0.0358436293900013, + 0.028208864852786064, + 0.018353182822465897, + 0.024030767381191254, + 0.012153086252510548, + 0.006678545381873846, + 0.009847946465015411, + 0.009828692302107811, + 0.011949170380830765, + 0.013337602838873863, + 0.006738969590514898, + 0.015340501442551613, + 0.03201094642281532, + 0.02557355910539627, + 0.0341452956199646, + 0.04198632389307022, + 0.0626639872789383, + 0.04023607075214386, + 0.05409267544746399, + 0.02237623557448387, + 0.012714875862002373, + 0.011428574100136757, + 0.020105630159378052, + 0.012685067020356655, + 0.015936382114887238, + 0.015824811533093452, + 0.014086076058447361, + 0.009605690836906433, + 0.007558528799563646, + 0.005781495478004217, + 0.005978282541036606, + 0.007620751857757568, + 0.007683705072849989, + 0.009432771243155003, + 0.005897859577089548, + 0.013510309159755707, + 0.004772834945470095, + 0.00497451750561595, + 0.004706325940787792, + 0.004632853902876377, + 0.007539020385593176, + 0.006377761252224445, + 0.004124303348362446, + 0.006643926724791527, + 0.008356216363608837, + 0.0044816844165325165, + 0.006465079262852669, + 0.004898847546428442, + 0.005881128367036581, + 0.008994209580123425, + 0.0030131174717098475, + 0.011328896507620811, + 0.0037151742726564407, + 0.0038306433707475662, + 0.004154962953180075, + 0.005610735155642033, + 0.004121452569961548, + 0.00541848735883832, + 0.0037984848022460938, + 0.006549727637320757, + 0.004943313077092171, + 0.002940091770142317, + 0.006623018532991409, + 0.004257604945451021, + 0.00409913482144475, + 0.00498451292514801, + 0.002507658675312996, + 0.003832889022305608, + 0.006237788591533899, + 0.004603908862918615, + 0.005898918490856886, + 0.0048645478673279285, + 0.0054166982881724834, + 0.008382202126085758, + 0.003803094383329153, + 0.010376080870628357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.003542815800756216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298323154449463 + ], + "opposite_logprob_tokens": [ + -3.7732982635498047 + ], + "logprob_reward": -0.023298323154449463, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298323154449463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1046457588672638, + 0.11519335955381393, + 0.2637180685997009, + 0.12443369626998901, + 0.044454727321863174, + 0.045376475900411606, + 0.08716457337141037, + 0.11604586988687515, + 0.05036275088787079, + 0.048604704439640045 + ], + "credit_map": [ + 0.016628682613372803, + 0.010256420820951462, + 0.009088601917028427, + 0.011041510850191116, + 0.008711718022823334, + 0.016291789710521698, + 0.008569144643843174, + 0.024057893082499504, + 0.012396750040352345, + 0.007084917277097702, + 0.012788666412234306, + 0.009554775431752205, + 0.009148973040282726, + 0.019638046622276306, + 0.012213132344186306, + 0.03236809000372887, + 0.0407993346452713, + 0.014798606745898724, + 0.03257810324430466, + 0.025706171989440918, + 0.03109145164489746, + 0.037937652319669724, + 0.03862126171588898, + 0.04218548908829689, + 0.021401075646281242, + 0.009006139822304249, + 0.018898366019129753, + 0.01448738481849432, + 0.008927480317652225, + 0.019884798675775528, + 0.009379340335726738, + 0.022449109703302383, + 0.005910612177103758, + 0.003975541330873966, + 0.005614282563328743, + 0.005041921511292458, + 0.003999435342848301, + 0.0066381217911839485, + 0.004195781424641609, + 0.009079032577574253, + 0.00535565335303545, + 0.004237116780132055, + 0.007044197991490364, + 0.0048363846726715565, + 0.004174558445811272, + 0.006600969471037388, + 0.0045927828177809715, + 0.008534812368452549, + 0.01097724586725235, + 0.007105117663741112, + 0.008537114597856998, + 0.0066861556842923164, + 0.01036875694990158, + 0.014886545948684216, + 0.006568034179508686, + 0.02203560061752796, + 0.009733051061630249, + 0.01344803161919117, + 0.013991506770253181, + 0.019322343170642853, + 0.016009880229830742, + 0.01669936254620552, + 0.011621169745922089, + 0.015220524743199348, + 0.007277959957718849, + 0.0035630117636173964, + 0.008310153149068356, + 0.005524618551135063, + 0.004490339197218418, + 0.00731364032253623, + 0.005742859560996294, + 0.008140170946717262, + 0.007375550456345081, + 0.0034894272685050964, + 0.007633805740624666, + 0.005057850852608681, + 0.005100497975945473, + 0.006856492254883051, + 0.005341947078704834, + 0.007749132812023163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.023298323154449463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864649102091789 + ], + "opposite_logprob_tokens": [ + -3.0486464500427246 + ], + "logprob_reward": -0.04864649102091789, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864649102091789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25607219338417053, + 0.09354278445243835, + 0.05584149435162544, + 0.2688845098018646, + 0.04721834510564804, + 0.039575692266225815, + 0.053001441061496735, + 0.08593442291021347, + 0.04113087058067322, + 0.058798179030418396 + ], + "credit_map": [ + 0.024558063596487045, + 0.045563582330942154, + 0.020515253767371178, + 0.026325112208724022, + 0.022664062678813934, + 0.036996643990278244, + 0.02488364279270172, + 0.05456584692001343, + 0.010399693623185158, + 0.009908458217978477, + 0.008654731325805187, + 0.01338984165340662, + 0.017328577116131783, + 0.013240667060017586, + 0.009362166747450829, + 0.011258644983172417, + 0.010312285274267197, + 0.004823649302124977, + 0.007906037382781506, + 0.00680174957960844, + 0.00676514208316803, + 0.008458561263978481, + 0.004950680769979954, + 0.005823384504765272, + 0.023717079311609268, + 0.026910779997706413, + 0.027636123821139336, + 0.04041881859302521, + 0.038269396871328354, + 0.037938907742500305, + 0.02084904909133911, + 0.05314435809850693, + 0.006177844479680061, + 0.004657477606087923, + 0.004795068874955177, + 0.006535960361361504, + 0.007333668414503336, + 0.005945890210568905, + 0.004121690057218075, + 0.0076507460325956345, + 0.0077414498664438725, + 0.0032150468323379755, + 0.0066263070330023766, + 0.004307706840336323, + 0.0037247613072395325, + 0.00526135228574276, + 0.003250513458624482, + 0.005448553711175919, + 0.006493044551461935, + 0.004197280388325453, + 0.005345351994037628, + 0.003980525303632021, + 0.006093576550483704, + 0.009629611857235432, + 0.0038209091871976852, + 0.01344113890081644, + 0.00827149860560894, + 0.009555725380778313, + 0.010214041918516159, + 0.013156663626432419, + 0.011091101914644241, + 0.012604244984686375, + 0.008621999062597752, + 0.012419137172400951, + 0.0069941747933626175, + 0.0031363442540168762, + 0.006904917303472757, + 0.005742354318499565, + 0.00420227600261569, + 0.00548445014283061, + 0.003260798519477248, + 0.005405556410551071, + 0.008079960942268372, + 0.005553241819143295, + 0.007077365648001432, + 0.006547379773110151, + 0.007927910424768925, + 0.008418633602559566, + 0.004365589004009962, + 0.01082809828221798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.04864649102091789 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606437936425209 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.01606437936425209, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606437936425209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19286929070949554, + 0.05031413584947586, + 0.07816104590892792, + 0.13888944685459137, + 0.04777124896645546, + 0.05236191675066948, + 0.1648339480161667, + 0.20724035799503326, + 0.03882945701479912, + 0.02872920036315918 + ], + "credit_map": [ + 0.020910784602165222, + 0.01841064915060997, + 0.033786095678806305, + 0.019995244219899178, + 0.02056531049311161, + 0.0281367227435112, + 0.017933933064341545, + 0.03313055634498596, + 0.006471294444054365, + 0.004006178583949804, + 0.006467913277447224, + 0.00565097201615572, + 0.006013653241097927, + 0.0075155869126319885, + 0.004687997046858072, + 0.009500542655587196, + 0.009696122258901596, + 0.006418122444301844, + 0.011296845972537994, + 0.00937650352716446, + 0.012077150866389275, + 0.011607407592236996, + 0.006143753882497549, + 0.011545136570930481, + 0.01087441761046648, + 0.015066234394907951, + 0.016777168959379196, + 0.022296322509646416, + 0.021133024245500565, + 0.019372351467609406, + 0.015664050355553627, + 0.017705878242850304, + 0.006540331989526749, + 0.004750201944261789, + 0.005447180941700935, + 0.0050330874510109425, + 0.005100838374346495, + 0.007614966481924057, + 0.004246004857122898, + 0.009038638323545456, + 0.005133486818522215, + 0.0052765426225960255, + 0.006223310716450214, + 0.006908002309501171, + 0.007152509409934282, + 0.006735453847795725, + 0.005685545038431883, + 0.009247063659131527, + 0.020333901047706604, + 0.019610267132520676, + 0.01984039694070816, + 0.01652960106730461, + 0.025011703372001648, + 0.025679752230644226, + 0.00981168169528246, + 0.02801663614809513, + 0.017556628212332726, + 0.0253848098218441, + 0.0248090997338295, + 0.03643173351883888, + 0.030657397583127022, + 0.02980765327811241, + 0.019078461453318596, + 0.023514579981565475, + 0.005265674088150263, + 0.0030825906433165073, + 0.006310952361673117, + 0.004879610147327185, + 0.004866796545684338, + 0.005115804262459278, + 0.004287710413336754, + 0.005020314361900091, + 0.003709529060870409, + 0.00283888285048306, + 0.004093609284609556, + 0.0030404217541217804, + 0.0034133782610297203, + 0.004100386053323746, + 0.0027422087732702494, + 0.0047907838597893715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01606437936425209 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are vibrant, cup-shaped Caltha palustris flowers with buttercup-yellow petals visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are heart-shaped green leaves visible beneath the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady stream of tiny, shimmering oxygen bubbles rising from the submerged leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft, emerald blur of out-of-focus reeds and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063909977674484 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.016063909977674484, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063909977674484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20973066985607147, + 0.14849691092967987, + 0.09613398462533951, + 0.114472895860672, + 0.068438820540905, + 0.07589329034090042, + 0.07439836859703064, + 0.14501430094242096, + 0.033635884523391724, + 0.033784832805395126 + ], + "credit_map": [ + 0.019756358116865158, + 0.025203481316566467, + 0.023957010358572006, + 0.03807394951581955, + 0.024836676195263863, + 0.02924296073615551, + 0.021762419492006302, + 0.02689780481159687, + 0.01018900889903307, + 0.015897415578365326, + 0.015211256220936775, + 0.023510098457336426, + 0.030226977542042732, + 0.019797632470726967, + 0.013596384786069393, + 0.020068123936653137, + 0.005955065134912729, + 0.007000909186899662, + 0.01167517714202404, + 0.013288287445902824, + 0.01742899790406227, + 0.011911274865269661, + 0.02005155384540558, + 0.008822723291814327, + 0.011017929762601852, + 0.009700652211904526, + 0.01353936456143856, + 0.01703076809644699, + 0.019391939043998718, + 0.016178742051124573, + 0.011968870647251606, + 0.015644632279872894, + 0.011574015021324158, + 0.006538220681250095, + 0.007791239768266678, + 0.006600518710911274, + 0.008310270495712757, + 0.009404601529240608, + 0.006246595177799463, + 0.01197336707264185, + 0.013719387352466583, + 0.006175883114337921, + 0.011193173937499523, + 0.006732946261763573, + 0.008490152657032013, + 0.01086964551359415, + 0.004217863082885742, + 0.014494233764708042, + 0.010188187472522259, + 0.006311491597443819, + 0.010508217848837376, + 0.007082383614033461, + 0.008379213511943817, + 0.012362224981188774, + 0.00808270275592804, + 0.011483943089842796, + 0.010953875258564949, + 0.017623495310544968, + 0.016825560480356216, + 0.024594072252511978, + 0.025176744908094406, + 0.020705970004200935, + 0.012548776343464851, + 0.016585811972618103, + 0.0033332917373627424, + 0.0032372153364121914, + 0.004987346939742565, + 0.004198409151285887, + 0.004413413815200329, + 0.005084879230707884, + 0.003162247594445944, + 0.005219077691435814, + 0.0032797332387417555, + 0.0039850641041994095, + 0.004190234001725912, + 0.003942881245166063, + 0.005291055887937546, + 0.0052492935210466385, + 0.003074483945965767, + 0.004772087559103966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016063909977674484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_tensor([2971, 2971])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_tensor([2971, 2971])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..065ca1ec28542a1cc4022fde7b3f5e710557058d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_tensor([2971, 2971])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, long shot captures a breathtaking landscape rendered entirely in Van Gogh's signature swirling, detached style, with each element defined by thick, visible, and emotive brushstrokes. The sky above is a turbulent vortex of deep indigo and fiery yellow, its celestial motion conveyed through powerful swirls of thick paint, and the stars appear as blazing, incandescent orbs. A small, ethereal figure hangs quietly in the absolute center of the composition, her ears delicate and pointed, her white hair flowing so softly and subtly, almost imperceptibly. Her expression is one of profound serenity, her body clad in a simple, translucent white gown that captures the magical light surrounding her. Below her, a dreamy, luminous field of plants sways in a silent, rhythmic dance; each petal and leaf a unique, vibrant color, shimmering with an inner light. A wisp of glowing green energy, rendered as a sinuous ribbon of paint, slowly and gracefully spirals around the floating figure, visually tethering her to the vibrant, vivid landscape and creating an overwhelming atmosphere of mysterious tranquility and cosmic connection.", + "seed": 14970, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_2971_seed_14970_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030202112393453717, + "vlm_logprob": -0.030202112393453717, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.030202112393453717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005138879641890526 + ], + "opposite_logprob_tokens": [ + -5.630138874053955 + ], + "logprob_reward": -0.005138879641890526, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.005138879641890526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1527957022190094, + 0.03008962795138359, + 0.0217573419213295, + 0.38368311524391174, + 0.05209317430853844, + 0.17292681336402893, + 0.09609532356262207, + 0.03822067752480507, + 0.02250278927385807, + 0.029835423454642296 + ], + "credit_map": [ + 0.026186227798461914, + 0.014556476846337318, + 0.013274603523314, + 0.016254182904958725, + 0.010616889223456383, + 0.03670762479305267, + 0.018198927864432335, + 0.017000757157802582, + 0.0041766962967813015, + 0.003587973304092884, + 0.0037773787043988705, + 0.0033629490062594414, + 0.002616463927552104, + 0.004476438742130995, + 0.003625128185376525, + 0.004466599319130182, + 0.0034943250939249992, + 0.003245798172429204, + 0.0017826134571805596, + 0.0021219977643340826, + 0.002602163702249527, + 0.0045504202134907246, + 0.001682325848378241, + 0.002277697902172804, + 0.058446865528821945, + 0.02386455237865448, + 0.09222573786973953, + 0.02250862307846546, + 0.01844216324388981, + 0.06901922821998596, + 0.028408348560333252, + 0.0707676038146019, + 0.013194972649216652, + 0.004499269649386406, + 0.00373693834990263, + 0.0025558453053236008, + 0.004083608742803335, + 0.017325785011053085, + 0.003603988792747259, + 0.0030927604530006647, + 0.026652242988348007, + 0.028162501752376556, + 0.01866903528571129, + 0.015308535657823086, + 0.021902089938521385, + 0.03166908025741577, + 0.015067179687321186, + 0.015496142208576202, + 0.015388472005724907, + 0.011399860493838787, + 0.013524766080081463, + 0.006488669663667679, + 0.010175603441894054, + 0.020680461078882217, + 0.00823467131704092, + 0.010202822275459766, + 0.0075633106753230095, + 0.003659567330032587, + 0.004255668725818396, + 0.0030787421856075525, + 0.004353540483862162, + 0.009074247442185879, + 0.002922944724559784, + 0.0033126554917544127, + 0.0029162077698856592, + 0.002519307192414999, + 0.0034200933296233416, + 0.003077148227021098, + 0.003048650221899152, + 0.0031137270852923393, + 0.0022785968612879515, + 0.0021290588192641735, + 0.005512517411261797, + 0.003945858217775822, + 0.0035752246621996164, + 0.0030879804398864508, + 0.0025504245422780514, + 0.00429252116009593, + 0.002580466214567423, + 0.004290430340915918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.005138879641890526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865727573633194 + ], + "opposite_logprob_tokens": [ + -3.048657178878784 + ], + "logprob_reward": -0.04865727573633194, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865727573633194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08313678950071335, + 0.0412679947912693, + 0.02637130580842495, + 0.29875829815864563, + 0.027365412563085556, + 0.15165045857429504, + 0.06130768731236458, + 0.05807161703705788, + 0.02255038544535637, + 0.22951997816562653 + ], + "credit_map": [ + 0.011618729680776596, + 0.011877263896167278, + 0.008605102077126503, + 0.010203707963228226, + 0.00858579482883215, + 0.009980869479477406, + 0.012109893374145031, + 0.010155432857573032, + 0.0056961397640407085, + 0.006752394139766693, + 0.00420267041772604, + 0.004178712610155344, + 0.004916681442409754, + 0.00589332077652216, + 0.00531659834086895, + 0.004311477765440941, + 0.0034134495072066784, + 0.004066778812557459, + 0.00256459042429924, + 0.0033333497121930122, + 0.0027891136705875397, + 0.004627713933587074, + 0.0027929279021918774, + 0.0027833806816488504, + 0.05208944156765938, + 0.014730983413755894, + 0.06478286534547806, + 0.01593823730945587, + 0.01341832522302866, + 0.0654202327132225, + 0.022107381373643875, + 0.05027083307504654, + 0.003007984021678567, + 0.0019343503518030047, + 0.006803535390645266, + 0.0021691301371902227, + 0.003125402145087719, + 0.003006999148055911, + 0.002990884706377983, + 0.0043271263130009174, + 0.020736008882522583, + 0.028522199019789696, + 0.01696106418967247, + 0.01337345875799656, + 0.023709412664175034, + 0.018243540078401566, + 0.01450925599783659, + 0.015595518052577972, + 0.0113374600186944, + 0.004952770657837391, + 0.006175932474434376, + 0.0076176817528903484, + 0.004252348095178604, + 0.01677798107266426, + 0.00538521446287632, + 0.004808296915143728, + 0.008547733537852764, + 0.008559254929423332, + 0.006682989653199911, + 0.006606452167034149, + 0.0067180730402469635, + 0.010647503659129143, + 0.004937121644616127, + 0.0053724925965070724, + 0.0030236232560127974, + 0.0020053805783391, + 0.0038923085667192936, + 0.0017181377625092864, + 0.0018564335769042373, + 0.0029443339444696903, + 0.001991773722693324, + 0.005118395201861858, + 0.03362163528800011, + 0.02273973822593689, + 0.0479389913380146, + 0.015531331300735474, + 0.015599077567458153, + 0.02547593228518963, + 0.017618756741285324, + 0.050994522869586945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.04865727573633194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016055112704634666 + ], + "opposite_logprob_tokens": [ + -4.141055107116699 + ], + "logprob_reward": -0.016055112704634666, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016055112704634666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08378744125366211, + 0.06324860453605652, + 0.04162599518895149, + 0.23274147510528564, + 0.02595347911119461, + 0.2401122897863388, + 0.26369479298591614, + 0.019126158207654953, + 0.005930048879235983, + 0.023779654875397682 + ], + "credit_map": [ + 0.012250569649040699, + 0.006446619983762503, + 0.00921157281845808, + 0.009984909556806087, + 0.006028071977198124, + 0.018166739493608475, + 0.012881838716566563, + 0.008817129768431187, + 0.008450007997453213, + 0.010353423655033112, + 0.006123468745499849, + 0.0068945931270718575, + 0.007025393657386303, + 0.010269219055771828, + 0.008224179968237877, + 0.005908326245844364, + 0.007715515326708555, + 0.007069738116115332, + 0.002809777157381177, + 0.003212361130863428, + 0.004618400242179632, + 0.010342198424041271, + 0.0031281597912311554, + 0.002729851519688964, + 0.03721744194626808, + 0.01302582211792469, + 0.05473609268665314, + 0.01371876709163189, + 0.010518450289964676, + 0.044710054993629456, + 0.01688103936612606, + 0.0419338196516037, + 0.0036940814461559057, + 0.0028947105165570974, + 0.0034852062817662954, + 0.0033155842684209347, + 0.0025873701088130474, + 0.004629043396562338, + 0.003083499614149332, + 0.0022639844100922346, + 0.038462553173303604, + 0.03791654109954834, + 0.026934955269098282, + 0.023262768983840942, + 0.026713160797953606, + 0.04620280861854553, + 0.02047523856163025, + 0.020144285634160042, + 0.03600199520587921, + 0.04131346195936203, + 0.03541005402803421, + 0.021874506026506424, + 0.03563052788376808, + 0.03929732367396355, + 0.025167297571897507, + 0.028999676927924156, + 0.0026905101258307695, + 0.0028404551558196545, + 0.00234565744176507, + 0.001667026779614389, + 0.0023031658492982388, + 0.0024238526821136475, + 0.0019777417182922363, + 0.0028777490369975567, + 0.0007568431319668889, + 0.0006502386531792581, + 0.0007946944679133594, + 0.0006975178839638829, + 0.0007564045954495668, + 0.0010044518858194351, + 0.0006048801587894559, + 0.0006650190334767103, + 0.0034681716933846474, + 0.0019461140036582947, + 0.0047818319872021675, + 0.0017909582238644361, + 0.0015501080779358745, + 0.0031711626797914505, + 0.0018669192213565111, + 0.005204391200095415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.016055112704634666 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021964015904814005 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021964015904814005, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021964015904814005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09668537229299545, + 0.039381690323352814, + 0.026552418246865273, + 0.24954284727573395, + 0.05512896552681923, + 0.022856837138533592, + 0.40616416931152344, + 0.0651029571890831, + 0.018163135275244713, + 0.020421544089913368 + ], + "credit_map": [ + 0.018964126706123352, + 0.009681973606348038, + 0.006680048070847988, + 0.00911637395620346, + 0.006646618712693453, + 0.02794685773551464, + 0.009955137968063354, + 0.0076942467130720615, + 0.005172529723495245, + 0.003586829174309969, + 0.004268912132829428, + 0.0044190469197928905, + 0.003465153742581606, + 0.006971900351345539, + 0.005990567617118359, + 0.0055067515932023525, + 0.005262649152427912, + 0.0024613654240965843, + 0.0027105414774268866, + 0.0017446817364543676, + 0.0019517714390531182, + 0.007391137070953846, + 0.0017335348529741168, + 0.0032967370934784412, + 0.03969687968492508, + 0.012809548527002335, + 0.0603029802441597, + 0.014050610363483429, + 0.010839205235242844, + 0.046958405524492264, + 0.01848435588181019, + 0.04640087112784386, + 0.013139661401510239, + 0.0050146072171628475, + 0.004418710712343454, + 0.003983353264629841, + 0.003764412133023143, + 0.017247207462787628, + 0.0034989649429917336, + 0.004062056075781584, + 0.0030667625833302736, + 0.0024790982715785503, + 0.004017929546535015, + 0.0020172891672700644, + 0.0025585999246686697, + 0.0034398052375763655, + 0.0021054360549896955, + 0.0031719189137220383, + 0.05670010671019554, + 0.06215989962220192, + 0.05368265509605408, + 0.03314008191227913, + 0.0552133284509182, + 0.06319520622491837, + 0.03827553987503052, + 0.04379737377166748, + 0.010150643065571785, + 0.007427781820297241, + 0.008259736001491547, + 0.005390083882957697, + 0.008879123255610466, + 0.012211236171424389, + 0.006244807969778776, + 0.006539551075547934, + 0.0025291317142546177, + 0.0020561758428812027, + 0.001615103567019105, + 0.0018549941014498472, + 0.002489716513082385, + 0.003942637238651514, + 0.0017590777715668082, + 0.0019163001561537385, + 0.0035962231922894716, + 0.0019479559268802404, + 0.0035068162251263857, + 0.0022641222458332777, + 0.0015593725256621838, + 0.003625358222052455, + 0.0017732165288180113, + 0.00214847968891263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.0021964015904814005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896289229393005 + ], + "opposite_logprob_tokens": [ + -2.578962802886963 + ], + "logprob_reward": -0.07896289229393005, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896289229393005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07111599296331406, + 0.018930353224277496, + 0.01979958824813366, + 0.07780035585165024, + 0.015089217573404312, + 0.12417497485876083, + 0.6209339499473572, + 0.02273387275636196, + 0.014659537933766842, + 0.014762096107006073 + ], + "credit_map": [ + 0.009800969623029232, + 0.007126116659492254, + 0.007849419489502907, + 0.009422150440514088, + 0.006346690934151411, + 0.011964843608438969, + 0.01148530188947916, + 0.0071205017156898975, + 0.002785538323223591, + 0.002273581689223647, + 0.0018285424448549747, + 0.0024683880619704723, + 0.0017270268872380257, + 0.002711793640628457, + 0.003212522715330124, + 0.0019229607423767447, + 0.0037392647936940193, + 0.0015753301559016109, + 0.0023365551605820656, + 0.0014605142641812563, + 0.0010997547069564462, + 0.00527089973911643, + 0.001225662766955793, + 0.0030916105024516582, + 0.008857417851686478, + 0.005244820844382048, + 0.0221131332218647, + 0.005147509276866913, + 0.004728901199996471, + 0.00866973027586937, + 0.006348750554025173, + 0.016690099611878395, + 0.0027655442245304585, + 0.0011721334885805845, + 0.0019075217423960567, + 0.001313558197580278, + 0.0015380928525701165, + 0.0038581406697630882, + 0.0013058420736342669, + 0.0012283867690712214, + 0.02285757288336754, + 0.018434010446071625, + 0.013577076606452465, + 0.010666759684681892, + 0.012847996316850185, + 0.027602778747677803, + 0.009598829783499241, + 0.008589953184127808, + 0.08690125495195389, + 0.09275777637958527, + 0.08319281041622162, + 0.05109183490276337, + 0.08173645287752151, + 0.10005369037389755, + 0.05793367698788643, + 0.0672665387392044, + 0.0032693801913410425, + 0.004297859501093626, + 0.0017224110197275877, + 0.0024849730543792248, + 0.003730181371793151, + 0.0033363208640366793, + 0.0024254368618130684, + 0.0014673135010525584, + 0.002357762074097991, + 0.0015743834665045142, + 0.0026617732364684343, + 0.0013734527165070176, + 0.0012695796322077513, + 0.001538843265734613, + 0.0012245273683220148, + 0.0026592181529849768, + 0.001961410976946354, + 0.0017830897122621536, + 0.002552478574216366, + 0.0017405115067958832, + 0.001426207018084824, + 0.0018717548809945583, + 0.0014814732130616903, + 0.0019451731350272894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.07896289229393005 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot captures a breathtaking landscape rendered entirely in Van Gogh's signature swirling, detached style, with each element defined by thick, visible, and emotive brushstrokes. The sky above is a turbulent vortex of deep indigo and fiery yellow, its celestial motion conveyed through powerful swirls of thick paint, and the stars appear as blazing, incandescent orbs. A small, ethereal figure hangs quietly in the absolute center of the composition, her ears delicate and pointed, her white hair flowing so softly and subtly, almost imperceptibly. Her expression is one of profound serenity, her body clad in a simple, translucent white gown that captures the magical light surrounding her. Below her, a dreamy, luminous field of plants sways in a silent, rhythmic dance; each petal and leaf a unique, vibrant color, shimmering with an inner light. A wisp of glowing green energy, rendered as a sinuous ribbon of paint, slowly and gracefully spirals around the floating figure, visually tethering her to the vibrant, vivid landscape and creating an overwhelming atmosphere of mysterious tranquility and cosmic connection.", + "seed": 14971, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_2971_seed_14971_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015591154620051385, + "vlm_logprob": -0.015591154620051385, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.015591154620051385, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003915977198630571 + ], + "opposite_logprob_tokens": [ + -6.128915786743164 + ], + "logprob_reward": -0.003915977198630571, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.003915977198630571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24772123992443085, + 0.040303707122802734, + 0.016239751130342484, + 0.017050210386514664, + 0.04614788293838501, + 0.06665942817926407, + 0.05810622498393059, + 0.20093561708927155, + 0.2535669505596161, + 0.05326900631189346 + ], + "credit_map": [ + 0.023295853286981583, + 0.029901789501309395, + 0.022562231868505478, + 0.030067676678299904, + 0.05787121504545212, + 0.026434117928147316, + 0.03224441781640053, + 0.025343921035528183, + 0.004181671887636185, + 0.003818552242591977, + 0.006316762883216143, + 0.004913948010653257, + 0.004924834240227938, + 0.004304008558392525, + 0.0067394208163022995, + 0.005104501731693745, + 0.0014588250778615475, + 0.002157628769055009, + 0.0024263844825327396, + 0.0020436448976397514, + 0.0020285984501242638, + 0.002128161257132888, + 0.0022121204528957605, + 0.0017843857640400529, + 0.0021235989406704903, + 0.0018593500135466456, + 0.002449080580845475, + 0.0018491822993382812, + 0.002381028374657035, + 0.001715105609036982, + 0.00289815291762352, + 0.0017747092060744762, + 0.005130060017108917, + 0.0042834291234612465, + 0.00901783537119627, + 0.006631304044276476, + 0.006368787959218025, + 0.004572878126055002, + 0.0054563782177865505, + 0.0046872044913470745, + 0.00496723921969533, + 0.0072706472128629684, + 0.005293992348015308, + 0.004726863000541925, + 0.023135358467698097, + 0.008522296324372292, + 0.007468056865036488, + 0.005274972878396511, + 0.005925789941102266, + 0.003940444439649582, + 0.004746594000607729, + 0.0051109446212649345, + 0.009986091405153275, + 0.015792284160852432, + 0.006000257562845945, + 0.0066038160584867, + 0.022435111925005913, + 0.017592299729585648, + 0.017307667061686516, + 0.01626630686223507, + 0.030391188338398933, + 0.05469117686152458, + 0.02322414331138134, + 0.019027722999453545, + 0.029884112998843193, + 0.021568741649389267, + 0.02260766550898552, + 0.020843805745244026, + 0.03724733740091324, + 0.05304491147398949, + 0.04405592381954193, + 0.024314437061548233, + 0.005324008408933878, + 0.006201513577252626, + 0.003810333553701639, + 0.003787342458963394, + 0.01736939325928688, + 0.007190619595348835, + 0.005872897803783417, + 0.003712889039888978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.003915977198630571 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607329398393631 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607329398393631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607329398393631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29317671060562134, + 0.16542303562164307, + 0.053556960076093674, + 0.010037939064204693, + 0.022827034816145897, + 0.027842426672577858, + 0.10207650810480118, + 0.07002478837966919, + 0.21781976521015167, + 0.037214796990156174 + ], + "credit_map": [ + 0.030513884499669075, + 0.03686155751347542, + 0.047854725271463394, + 0.040776416659355164, + 0.03243224695324898, + 0.02855844981968403, + 0.03894242271780968, + 0.0372370220720768, + 0.02163890190422535, + 0.018604587763547897, + 0.011080368421971798, + 0.01895720139145851, + 0.029481563717126846, + 0.0176871158182621, + 0.03259635716676712, + 0.015376931987702847, + 0.006905788090080023, + 0.006724310573190451, + 0.0034616845659911633, + 0.006952405907213688, + 0.007952548563480377, + 0.006078047212213278, + 0.010615270584821701, + 0.0048669008538126945, + 0.0015026205219328403, + 0.0015601466875523329, + 0.0015151739353314042, + 0.0009996868902817369, + 0.001025561010465026, + 0.0011519715189933777, + 0.001187036861665547, + 0.001095741637982428, + 0.0023070448078215122, + 0.0026868830900639296, + 0.0026387530378997326, + 0.0019512901781126857, + 0.00607499573379755, + 0.0027885742019861937, + 0.0021271395962685347, + 0.002252354519441724, + 0.002817613072693348, + 0.00325957452878356, + 0.00269086635671556, + 0.0027364345733076334, + 0.008020276203751564, + 0.003200474428012967, + 0.002716320799663663, + 0.0024008650798350573, + 0.01312953233718872, + 0.007609694264829159, + 0.006432960741221905, + 0.006615812890231609, + 0.013132565654814243, + 0.031941525638103485, + 0.010365170426666737, + 0.012849248945713043, + 0.00866150576621294, + 0.006112104281783104, + 0.005862952675670385, + 0.005499011836946011, + 0.008256027474999428, + 0.018524721264839172, + 0.009645613841712475, + 0.0074628558941185474, + 0.025799499824643135, + 0.017689863219857216, + 0.020718026906251907, + 0.01855284720659256, + 0.027775222435593605, + 0.04683443903923035, + 0.03890959545969963, + 0.02154027856886387, + 0.00403925497084856, + 0.0029327012598514557, + 0.004220461938530207, + 0.0037573485169559717, + 0.007810446433722973, + 0.00403562793508172, + 0.006807834841310978, + 0.003611120395362377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.01607329398393631 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076191299594938755 + ], + "opposite_logprob_tokens": [ + -4.8826189041137695 + ], + "logprob_reward": -0.0076191299594938755, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076191299594938755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14841753244400024, + 0.07629852741956711, + 0.02711484581232071, + 0.02571394294500351, + 0.044795311987400055, + 0.042480744421482086, + 0.08146465569734573, + 0.23287871479988098, + 0.296347051858902, + 0.02448870986700058 + ], + "credit_map": [ + 0.017841393128037453, + 0.01367864292114973, + 0.017360206693410873, + 0.01618943363428116, + 0.0169548150151968, + 0.01881118305027485, + 0.024908702820539474, + 0.02267315611243248, + 0.006062379106879234, + 0.008696004748344421, + 0.010547337122261524, + 0.007588671986013651, + 0.02029278688132763, + 0.006824680604040623, + 0.007539458107203245, + 0.008747209794819355, + 0.0021819472312927246, + 0.004264420364052057, + 0.0027275767643004656, + 0.0024932597298175097, + 0.00758033012971282, + 0.002870839321985841, + 0.002770203398540616, + 0.002226269105449319, + 0.003093870123848319, + 0.002795641077682376, + 0.0036708081606775522, + 0.0024056287948042154, + 0.005629773251712322, + 0.0028038937598466873, + 0.0028460139874368906, + 0.0024683137889951468, + 0.004868953488767147, + 0.00499448087066412, + 0.0061619095504283905, + 0.00539757963269949, + 0.007500274572521448, + 0.004165406338870525, + 0.007334906607866287, + 0.004371802322566509, + 0.004916421137750149, + 0.006623224820941687, + 0.003909117542207241, + 0.004691051784902811, + 0.009968849830329418, + 0.0042195189744234085, + 0.004768630489706993, + 0.0033839293755590916, + 0.010057294741272926, + 0.005468134302645922, + 0.006522954907268286, + 0.007035992108285427, + 0.0104760667309165, + 0.022565074265003204, + 0.0098422197625041, + 0.009496912360191345, + 0.025733504444360733, + 0.01968107558786869, + 0.02196582593023777, + 0.018351035192608833, + 0.036051370203495026, + 0.06185494735836983, + 0.027155393734574318, + 0.022085562348365784, + 0.034417301416397095, + 0.0249891746789217, + 0.026637308299541473, + 0.02495446614921093, + 0.04024419188499451, + 0.06252344697713852, + 0.053367216140031815, + 0.029213927686214447, + 0.002584798028692603, + 0.003122818423435092, + 0.0019180693197995424, + 0.002980771940201521, + 0.004113599192351103, + 0.002314116572961211, + 0.004816721193492413, + 0.002637815196067095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0076191299594938755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001711214892566204 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.001711214892566204, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001711214892566204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23919403553009033, + 0.1064048632979393, + 0.027121642604470253, + 0.017015116289258003, + 0.029644599184393883, + 0.044089943170547485, + 0.04187439754605293, + 0.19705818593502045, + 0.11551471054553986, + 0.18208250403404236 + ], + "credit_map": [ + 0.024782253429293633, + 0.030085131525993347, + 0.02277248166501522, + 0.018912743777036667, + 0.0678820013999939, + 0.026873191818594933, + 0.022015899419784546, + 0.02587033249437809, + 0.010086136870086193, + 0.013280454091727734, + 0.005258567631244659, + 0.010660036467015743, + 0.030904056504368782, + 0.011257749982178211, + 0.016925634816288948, + 0.00803222507238388, + 0.0027894028462469578, + 0.002990992506965995, + 0.0030354505870491266, + 0.0037464466877281666, + 0.0038188474718481302, + 0.002879495033994317, + 0.005163663066923618, + 0.002697342773899436, + 0.0017671508248895407, + 0.001688153948634863, + 0.002001076005399227, + 0.0024074921384453773, + 0.00240585976280272, + 0.0015455145621672273, + 0.003415473038330674, + 0.0017843967070803046, + 0.0027214244473725557, + 0.003900103969499469, + 0.00471357861533761, + 0.0033623105846345425, + 0.00639089522883296, + 0.002540480811148882, + 0.002933079842478037, + 0.0030827282462269068, + 0.0035046166740357876, + 0.004488522186875343, + 0.004659297876060009, + 0.004894861951470375, + 0.00839659571647644, + 0.005346282850950956, + 0.0077443732880055904, + 0.0050553916953504086, + 0.006217743270099163, + 0.00294459774158895, + 0.0034818847198039293, + 0.0030316731426864862, + 0.00599342305213213, + 0.010323184542357922, + 0.005269950721412897, + 0.004611941520124674, + 0.015302486717700958, + 0.02310284785926342, + 0.01357380859553814, + 0.018977178260684013, + 0.05140139162540436, + 0.025379274040460587, + 0.035137709230184555, + 0.01418347842991352, + 0.009332533925771713, + 0.013850497081875801, + 0.007418929133564234, + 0.008636190555989742, + 0.04073366895318031, + 0.012170431204140186, + 0.01651485450565815, + 0.006857595406472683, + 0.020059289410710335, + 0.02358049340546131, + 0.010936525650322437, + 0.012365137226879597, + 0.06107184663414955, + 0.02270122990012169, + 0.018190406262874603, + 0.013177575543522835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.001711214892566204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863615706562996 + ], + "opposite_logprob_tokens": [ + -3.0486361980438232 + ], + "logprob_reward": -0.04863615706562996, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863615706562996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2718939185142517, + 0.10274714976549149, + 0.0329148955643177, + 0.022575436159968376, + 0.026447778567671776, + 0.023579970002174377, + 0.10667373239994049, + 0.22104980051517487, + 0.11835676431655884, + 0.07376057654619217 + ], + "credit_map": [ + 0.028037946671247482, + 0.036221958696842194, + 0.037525612860918045, + 0.03141546994447708, + 0.033301547169685364, + 0.0345967561006546, + 0.03272278606891632, + 0.03807183355093002, + 0.00968552939593792, + 0.008080330677330494, + 0.018406718969345093, + 0.01303369551897049, + 0.01646915264427662, + 0.010905716568231583, + 0.01103152148425579, + 0.015134483575820923, + 0.005014614202082157, + 0.0026423712261021137, + 0.005488405004143715, + 0.004152288194745779, + 0.004630499519407749, + 0.004408117383718491, + 0.0031400423031300306, + 0.0034385521430522203, + 0.002448605140671134, + 0.002322988584637642, + 0.003653143299743533, + 0.0024872315116226673, + 0.004974505864083767, + 0.00204407493583858, + 0.0021568392403423786, + 0.0024880454875528812, + 0.0031582287047058344, + 0.002875549253076315, + 0.004931177012622356, + 0.003192961448803544, + 0.004303185734897852, + 0.0024978010915219784, + 0.0029264173936098814, + 0.002562456065788865, + 0.0030066855251789093, + 0.0027756269555538893, + 0.0032085240818560123, + 0.0020606687758117914, + 0.0029194275848567486, + 0.0032854301389306784, + 0.0030355509370565414, + 0.003288052976131439, + 0.014177273027598858, + 0.007747184485197067, + 0.00741279311478138, + 0.007400483824312687, + 0.015581322833895683, + 0.03191465511918068, + 0.009283643215894699, + 0.013156367465853691, + 0.02409992180764675, + 0.02007199265062809, + 0.021683257073163986, + 0.01795973628759384, + 0.033516306430101395, + 0.05784599110484123, + 0.024995887652039528, + 0.02087671309709549, + 0.013331546448171139, + 0.009286823682487011, + 0.01102957222610712, + 0.010178841650485992, + 0.0158909410238266, + 0.02388611063361168, + 0.023444566875696182, + 0.011308342218399048, + 0.008466637693345547, + 0.00906018353998661, + 0.004860436078161001, + 0.005667097866535187, + 0.022649791091680527, + 0.009507070295512676, + 0.007324235048145056, + 0.006225125398486853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.04863615706562996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot captures a breathtaking landscape rendered entirely in Van Gogh's signature swirling, detached style, with each element defined by thick, visible, and emotive brushstrokes. The sky above is a turbulent vortex of deep indigo and fiery yellow, its celestial motion conveyed through powerful swirls of thick paint, and the stars appear as blazing, incandescent orbs. A small, ethereal figure hangs quietly in the absolute center of the composition, her ears delicate and pointed, her white hair flowing so softly and subtly, almost imperceptibly. Her expression is one of profound serenity, her body clad in a simple, translucent white gown that captures the magical light surrounding her. Below her, a dreamy, luminous field of plants sways in a silent, rhythmic dance; each petal and leaf a unique, vibrant color, shimmering with an inner light. A wisp of glowing green energy, rendered as a sinuous ribbon of paint, slowly and gracefully spirals around the floating figure, visually tethering her to the vibrant, vivid landscape and creating an overwhelming atmosphere of mysterious tranquility and cosmic connection.", + "seed": 14972, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_2971_seed_14972_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.009585011890158058, + "vlm_logprob": -0.009585011890158058, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.009585011890158058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00205781776458025 + ], + "opposite_logprob_tokens": [ + -6.627058029174805 + ], + "logprob_reward": -0.00205781776458025, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.00205781776458025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36139097809791565, + 0.08032462000846863, + 0.03444191813468933, + 0.061539728194475174, + 0.14640308916568756, + 0.059765979647636414, + 0.0912071019411087, + 0.06619707494974136, + 0.06387194246053696, + 0.03485759347677231 + ], + "credit_map": [ + 0.053584225475788116, + 0.05244434252381325, + 0.03766636922955513, + 0.02782432921230793, + 0.049362920224666595, + 0.05340711772441864, + 0.03540797531604767, + 0.051693644374608994, + 0.015169903635978699, + 0.016868459060788155, + 0.008650083094835281, + 0.009581615217030048, + 0.009131585247814655, + 0.0075320033356547356, + 0.005560982972383499, + 0.007829979062080383, + 0.004160687327384949, + 0.0037677197251468897, + 0.003449907060712576, + 0.004691775888204575, + 0.005167811643332243, + 0.005009229760617018, + 0.0038974841590970755, + 0.004297302104532719, + 0.005391951650381088, + 0.010583294555544853, + 0.006140388082712889, + 0.009400424547493458, + 0.006410311441868544, + 0.005245850421488285, + 0.007047567516565323, + 0.01131993718445301, + 0.009832738898694515, + 0.036444295197725296, + 0.01869835890829563, + 0.010606139898300171, + 0.013798763044178486, + 0.012538858689367771, + 0.008349359035491943, + 0.036134570837020874, + 0.008046184666454792, + 0.006595232989639044, + 0.006242960691452026, + 0.007249068468809128, + 0.00970318354666233, + 0.0092202452942729, + 0.0054838270880281925, + 0.007225273177027702, + 0.008880487643182278, + 0.007565503474324942, + 0.009828579612076283, + 0.013421935960650444, + 0.010295829735696316, + 0.010299631394445896, + 0.015013604424893856, + 0.01590152643620968, + 0.007350055035203695, + 0.009484048932790756, + 0.007589362096041441, + 0.006443768739700317, + 0.008012855425477028, + 0.009550556540489197, + 0.006515302695333958, + 0.011251125484704971, + 0.006674957927316427, + 0.00446588359773159, + 0.0060104308649897575, + 0.00880124419927597, + 0.00699277687817812, + 0.007622722070664167, + 0.012047932483255863, + 0.011255990713834763, + 0.003720788052305579, + 0.0037198010832071304, + 0.0036811332684010267, + 0.007699446752667427, + 0.003683755174279213, + 0.0032100463286042213, + 0.005000702571123838, + 0.004141916520893574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.00205781776458025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639813400804996 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639813400804996, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639813400804996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20330822467803955, + 0.048671893775463104, + 0.03492030128836632, + 0.03584678843617439, + 0.30996471643447876, + 0.18671220541000366, + 0.060496240854263306, + 0.06413472443819046, + 0.0240786150097847, + 0.031866271048784256 + ], + "credit_map": [ + 0.035125214606523514, + 0.030468661338090897, + 0.02523367665708065, + 0.020488886162638664, + 0.02298271656036377, + 0.024926304817199707, + 0.019103551283478737, + 0.024979213252663612, + 0.0068926434032619, + 0.008996409364044666, + 0.004306399263441563, + 0.004974387586116791, + 0.005361150950193405, + 0.005850907880812883, + 0.004684853833168745, + 0.007605140097439289, + 0.005675987806171179, + 0.005045614670962095, + 0.0032138070091605186, + 0.0030934778042137623, + 0.005303021986037493, + 0.0034010845702141523, + 0.003375712549313903, + 0.005811596289277077, + 0.00564933056011796, + 0.003636028850451112, + 0.0031396618578583, + 0.0046279835514724255, + 0.00546957366168499, + 0.004920122679322958, + 0.003976629115641117, + 0.004427461419254541, + 0.021479012444615364, + 0.08070048689842224, + 0.040003880858421326, + 0.018419142812490463, + 0.029094934463500977, + 0.02562999166548252, + 0.015192026272416115, + 0.0794452428817749, + 0.016410740092396736, + 0.04177980497479439, + 0.022608110681176186, + 0.013463601469993591, + 0.025957560166716576, + 0.01672501303255558, + 0.010373416356742382, + 0.03939395770430565, + 0.005231552291661501, + 0.011370059102773666, + 0.008570446632802486, + 0.004344761371612549, + 0.006489155348390341, + 0.0061657908372581005, + 0.0038838479667901993, + 0.01444062776863575, + 0.007693443913012743, + 0.008184117265045643, + 0.005416950676590204, + 0.010619305074214935, + 0.007058627437800169, + 0.007759127300232649, + 0.009917428717017174, + 0.007485724985599518, + 0.0021760386880487204, + 0.002294790931046009, + 0.0032284408807754517, + 0.003694632789120078, + 0.002570054493844509, + 0.004137857351452112, + 0.002930373651906848, + 0.003046429017558694, + 0.005225710570812225, + 0.002683229511603713, + 0.0026300535537302494, + 0.004318880848586559, + 0.00583757646381855, + 0.004319654777646065, + 0.0036676835734397173, + 0.003183483611792326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.008639813400804996 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004630201030522585 + ], + "opposite_logprob_tokens": [ + -5.379630088806152 + ], + "logprob_reward": -0.004630201030522585, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004630201030522585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21121324598789215, + 0.0994909480214119, + 0.08801237493753433, + 0.21226751804351807, + 0.08764895051717758, + 0.028847401961684227, + 0.10789715498685837, + 0.0666445940732956, + 0.030773505568504333, + 0.0672043114900589 + ], + "credit_map": [ + 0.02996107004582882, + 0.04498075693845749, + 0.022312579676508904, + 0.02549607865512371, + 0.021531036123633385, + 0.016621485352516174, + 0.016885804012417793, + 0.03342443332076073, + 0.01713590696454048, + 0.008452390320599079, + 0.01367311179637909, + 0.01391749456524849, + 0.013008126989006996, + 0.011259082704782486, + 0.009454280138015747, + 0.012590554542839527, + 0.009591588750481606, + 0.011243949644267559, + 0.006969347596168518, + 0.018592610955238342, + 0.010432752780616283, + 0.009533394128084183, + 0.012446646578609943, + 0.009202082641422749, + 0.016286419704556465, + 0.037195004522800446, + 0.022330155596137047, + 0.041020091623067856, + 0.01631450653076172, + 0.01562418881803751, + 0.027902109548449516, + 0.03559504821896553, + 0.006749527063220739, + 0.012210402637720108, + 0.009844229556620121, + 0.017153892666101456, + 0.007930632680654526, + 0.00683034211397171, + 0.014118815772235394, + 0.012811116874217987, + 0.005010887049138546, + 0.003549292916432023, + 0.0029644088353961706, + 0.003964064177125692, + 0.003827527165412903, + 0.0031723955180495977, + 0.002619581064209342, + 0.003739244770258665, + 0.010231761261820793, + 0.02136334963142872, + 0.01207701675593853, + 0.011079683899879456, + 0.014093950390815735, + 0.010258384048938751, + 0.013636291027069092, + 0.015156716108322144, + 0.005377945955842733, + 0.007998669520020485, + 0.007478032726794481, + 0.011049916036427021, + 0.00601461436599493, + 0.010487312451004982, + 0.006845134776085615, + 0.01139296218752861, + 0.003664526389911771, + 0.002910781651735306, + 0.004011228680610657, + 0.0037771740462630987, + 0.003141809022054076, + 0.003033251967281103, + 0.0037538509350270033, + 0.006480880081653595, + 0.00953361764550209, + 0.004658068064600229, + 0.007279990240931511, + 0.009479261934757233, + 0.01067317184060812, + 0.005786963738501072, + 0.010236918926239014, + 0.009556317701935768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.004630201030522585 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815451705828309 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.002815451705828309, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815451705828309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14260587096214294, + 0.052509061992168427, + 0.05165453627705574, + 0.09917972981929779, + 0.08356262743473053, + 0.0542990006506443, + 0.19284170866012573, + 0.08654766529798508, + 0.04600519686937332, + 0.19079464673995972 + ], + "credit_map": [ + 0.01341074425727129, + 0.020624862983822823, + 0.014369238168001175, + 0.02316293865442276, + 0.012060943990945816, + 0.023167980834841728, + 0.018077075481414795, + 0.017732081934809685, + 0.00796470232307911, + 0.005046424921602011, + 0.006525993812829256, + 0.007056852802634239, + 0.007282889448106289, + 0.0057065123692154884, + 0.0059576090425252914, + 0.006968075875192881, + 0.0054375045001506805, + 0.00709394970908761, + 0.003916165791451931, + 0.010122567415237427, + 0.005613214336335659, + 0.005061477422714233, + 0.005807389039546251, + 0.008602268062531948, + 0.01132653933018446, + 0.015415488742291927, + 0.008937574923038483, + 0.020388975739479065, + 0.008242393843829632, + 0.007456685416400433, + 0.013460090383887291, + 0.013951978646218777, + 0.0072948685847222805, + 0.017524970695376396, + 0.00949921552091837, + 0.010632609948515892, + 0.00798750389367342, + 0.006192320492118597, + 0.007940206676721573, + 0.016490930691361427, + 0.006059784907847643, + 0.004848815035074949, + 0.005621940828859806, + 0.008634926751255989, + 0.006431966088712215, + 0.008412438444793224, + 0.007578984368592501, + 0.006710140500217676, + 0.017088087275624275, + 0.029183540493249893, + 0.02553238347172737, + 0.01217870693653822, + 0.017042839899659157, + 0.026203196495771408, + 0.012346012517809868, + 0.053266964852809906, + 0.010634359903633595, + 0.011607867665588856, + 0.007959095761179924, + 0.01689676195383072, + 0.008002064190804958, + 0.008684182539582253, + 0.011653654277324677, + 0.011109678074717522, + 0.004878864157944918, + 0.003399504814296961, + 0.0065453327260911465, + 0.00704497005790472, + 0.004805247765034437, + 0.005075927823781967, + 0.005029844120144844, + 0.009225503541529179, + 0.025258775800466537, + 0.015060613863170147, + 0.020463818684220314, + 0.030158769339323044, + 0.033733706921339035, + 0.015506566502153873, + 0.025810252875089645, + 0.024802135303616524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.002815451705828309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029781775549054146 + ], + "opposite_logprob_tokens": [ + -3.5297818183898926 + ], + "logprob_reward": -0.029781775549054146, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029781775549054146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14777617156505585, + 0.10586941987276077, + 0.08916135877370834, + 0.09351540356874466, + 0.15686620771884918, + 0.11119412630796432, + 0.16694758832454681, + 0.06525879353284836, + 0.031444251537323, + 0.03196675702929497 + ], + "credit_map": [ + 0.015769824385643005, + 0.03442863002419472, + 0.008680385537445545, + 0.032912954688072205, + 0.013166353106498718, + 0.013134500943124294, + 0.014138057827949524, + 0.015545452944934368, + 0.01895083114504814, + 0.012794001959264278, + 0.011204523034393787, + 0.010828475467860699, + 0.016088897362351418, + 0.012615323066711426, + 0.006969602312892675, + 0.01641775853931904, + 0.009262466803193092, + 0.021793264895677567, + 0.0075507089495658875, + 0.011885974556207657, + 0.007895812392234802, + 0.007168736774474382, + 0.01109312754124403, + 0.01251125242561102, + 0.007759643252938986, + 0.0070448205806314945, + 0.007867979817092419, + 0.021435169503092766, + 0.008772916160523891, + 0.00874337088316679, + 0.018892083317041397, + 0.01299940887838602, + 0.010045399889349937, + 0.02577870525419712, + 0.017682554200291634, + 0.023437194526195526, + 0.011188793927431107, + 0.011973856948316097, + 0.02526044473052025, + 0.03149924427270889, + 0.009053157642483711, + 0.023327834904193878, + 0.012394195422530174, + 0.019137797877192497, + 0.010720474645495415, + 0.00814804807305336, + 0.010551413521170616, + 0.017861194908618927, + 0.011289583519101143, + 0.009721287526190281, + 0.01939709484577179, + 0.014871075749397278, + 0.013930317014455795, + 0.03111851215362549, + 0.01790800504386425, + 0.04871169105172157, + 0.010646353475749493, + 0.005710497032850981, + 0.006833914201706648, + 0.00894843228161335, + 0.006451928056776524, + 0.008166695944964886, + 0.009609302505850792, + 0.008891663514077663, + 0.0026829359121620655, + 0.0031618825159966946, + 0.0040023839101195335, + 0.004606792703270912, + 0.003387975972145796, + 0.0034193252213299274, + 0.004412544891238213, + 0.00577040808275342, + 0.002918705577030778, + 0.0023567250464111567, + 0.003545827232301235, + 0.005326175130903721, + 0.004937778227031231, + 0.003047618316486478, + 0.003955036401748657, + 0.005878886207938194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.029781775549054146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot captures a breathtaking landscape rendered entirely in Van Gogh's signature swirling, detached style, with each element defined by thick, visible, and emotive brushstrokes. The sky above is a turbulent vortex of deep indigo and fiery yellow, its celestial motion conveyed through powerful swirls of thick paint, and the stars appear as blazing, incandescent orbs. A small, ethereal figure hangs quietly in the absolute center of the composition, her ears delicate and pointed, her white hair flowing so softly and subtly, almost imperceptibly. Her expression is one of profound serenity, her body clad in a simple, translucent white gown that captures the magical light surrounding her. Below her, a dreamy, luminous field of plants sways in a silent, rhythmic dance; each petal and leaf a unique, vibrant color, shimmering with an inner light. A wisp of glowing green energy, rendered as a sinuous ribbon of paint, slowly and gracefully spirals around the floating figure, visually tethering her to the vibrant, vivid landscape and creating an overwhelming atmosphere of mysterious tranquility and cosmic connection.", + "seed": 14973, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_2/index_2971_seed_14973_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030294067272916437, + "vlm_logprob": -0.030294067272916437, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.030294067272916437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0033056873362511396 + ], + "opposite_logprob_tokens": [ + -6.003305912017822 + ], + "logprob_reward": -0.0033056873362511396, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0033056873362511396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35076645016670227, + 0.07712014764547348, + 0.1362980604171753, + 0.08213028311729431, + 0.03415672853589058, + 0.0375896655023098, + 0.09436755627393723, + 0.03040565364062786, + 0.08198728412389755, + 0.0751781165599823 + ], + "credit_map": [ + 0.03546611964702606, + 0.03209904581308365, + 0.05727621167898178, + 0.025258567184209824, + 0.030962225049734116, + 0.09966760128736496, + 0.037739433348178864, + 0.03229726850986481, + 0.009859973564743996, + 0.00784361269325018, + 0.008629662916064262, + 0.010188299231231213, + 0.010528992861509323, + 0.0065734609961509705, + 0.016002964228391647, + 0.007493182085454464, + 0.009594796225428581, + 0.00946127064526081, + 0.005323454737663269, + 0.009809714742004871, + 0.010229629464447498, + 0.03719904273748398, + 0.04503653571009636, + 0.009643631055951118, + 0.005911943037062883, + 0.005811534821987152, + 0.004802499897778034, + 0.005832145921885967, + 0.005731720943003893, + 0.023732541128993034, + 0.02418733760714531, + 0.006120560225099325, + 0.004620105493813753, + 0.002790233353152871, + 0.0043679713271558285, + 0.0030613995622843504, + 0.004366200417280197, + 0.004984622355550528, + 0.007006301078945398, + 0.002959898207336664, + 0.003529755398631096, + 0.003963529597967863, + 0.003313487395644188, + 0.004745072685182095, + 0.006979946978390217, + 0.00670294463634491, + 0.004435868002474308, + 0.003919061738997698, + 0.00862479954957962, + 0.005623965989798307, + 0.013199684210121632, + 0.007519183214753866, + 0.007177284453064203, + 0.014451157301664352, + 0.02667377144098282, + 0.011097704991698265, + 0.0021544687915593386, + 0.001934278174303472, + 0.0038078399375081062, + 0.0022963399533182383, + 0.0027148716617375612, + 0.005840878933668137, + 0.0076944055035710335, + 0.0039625708013772964, + 0.010894560255110264, + 0.006997633259743452, + 0.0198084507137537, + 0.00588685367256403, + 0.008067979477345943, + 0.005044020712375641, + 0.016000740230083466, + 0.009287048131227493, + 0.005218153819441795, + 0.005746262613683939, + 0.003375990316271782, + 0.005983327981084585, + 0.006449425593018532, + 0.021869203075766563, + 0.020506825298070908, + 0.006028925534337759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0033056873362511396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901678025722504 + ], + "opposite_logprob_tokens": [ + -2.4640166759490967 + ], + "logprob_reward": -0.08901678025722504, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901678025722504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5611695647239685, + 0.17757372558116913, + 0.09688528627157211, + 0.02607995830476284, + 0.04182951897382736, + 0.021794039756059647, + 0.009035907685756683, + 0.009041775949299335, + 0.026132434606552124, + 0.030457792803645134 + ], + "credit_map": [ + 0.0393347293138504, + 0.0375601127743721, + 0.031889621168375015, + 0.04146174341440201, + 0.037774987518787384, + 0.1580135077238083, + 0.1721460074186325, + 0.0429888553917408, + 0.01420795638114214, + 0.012201103381812572, + 0.017491597682237625, + 0.017361603677272797, + 0.01007168646901846, + 0.04303419217467308, + 0.047614384442567825, + 0.015591202303767204, + 0.008016394451260567, + 0.006537187844514847, + 0.006107310764491558, + 0.006377100013196468, + 0.0066680945456027985, + 0.029009515419602394, + 0.02652263082563877, + 0.007647043094038963, + 0.0032866308465600014, + 0.002414669608697295, + 0.0022955997847020626, + 0.0027897246181964874, + 0.002199810929596424, + 0.007481320761144161, + 0.0021657589823007584, + 0.0034464444033801556, + 0.0038046701811254025, + 0.003052640939131379, + 0.004363305866718292, + 0.0035782093182206154, + 0.00373185146600008, + 0.007040857337415218, + 0.012667488306760788, + 0.0035904983524233103, + 0.0015590767143294215, + 0.001459540450014174, + 0.0024680893402546644, + 0.0018017531838268042, + 0.001322576659731567, + 0.00581040233373642, + 0.005395557731389999, + 0.0019770434591919184, + 0.0008683577179908752, + 0.000947690277826041, + 0.0016641289694234729, + 0.0013058496406301856, + 0.0011406535049900413, + 0.0007145366398617625, + 0.0013407376827672124, + 0.0010539538925513625, + 0.0013808290241286159, + 0.0010062857763841748, + 0.0009711053571663797, + 0.0014706975780427456, + 0.0010128439171239734, + 0.00130623416043818, + 0.0010117749916389585, + 0.0008820046205073595, + 0.003366289660334587, + 0.0021640784107148647, + 0.004716206807643175, + 0.00175629707518965, + 0.002731068292632699, + 0.00304429791867733, + 0.005981532856822014, + 0.0023726625367999077, + 0.002128480700775981, + 0.0021772596519440413, + 0.0023219140712171793, + 0.002060295781120658, + 0.002267418662086129, + 0.008599136024713516, + 0.008566250093281269, + 0.0023370368871837854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08901678025722504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023267226293683052 + ], + "opposite_logprob_tokens": [ + -3.7732672691345215 + ], + "logprob_reward": -0.023267226293683052, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023267226293683052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30027303099632263, + 0.1791449785232544, + 0.09657394886016846, + 0.1304844468832016, + 0.10711870342493057, + 0.047396544367074966, + 0.01355954259634018, + 0.06011415645480156, + 0.019497213885188103, + 0.0458373948931694 + ], + "credit_map": [ + 0.050768956542015076, + 0.03144754841923714, + 0.071120984852314, + 0.026137758046388626, + 0.03542525693774223, + 0.025521481409668922, + 0.030272062867879868, + 0.029579002410173416, + 0.01212300918996334, + 0.018269645050168037, + 0.01017962396144867, + 0.018058255314826965, + 0.022974083200097084, + 0.03225848823785782, + 0.051461122930049896, + 0.013820752501487732, + 0.010334380902349949, + 0.01090791542083025, + 0.007113263476639986, + 0.011250302195549011, + 0.012085068970918655, + 0.020528709515929222, + 0.012919249944388866, + 0.011435060761868954, + 0.013412666507065296, + 0.01130528561770916, + 0.015803147107362747, + 0.0120681868866086, + 0.014351751655340195, + 0.012593522667884827, + 0.04149189591407776, + 0.009457995183765888, + 0.007004010956734419, + 0.007742277812212706, + 0.004779524635523558, + 0.008730560541152954, + 0.010440048761665821, + 0.02319945953786373, + 0.03717614710330963, + 0.008046677336096764, + 0.0036669231485575438, + 0.004481171257793903, + 0.005086615215986967, + 0.003993924707174301, + 0.008109868504106998, + 0.004800958093255758, + 0.01398905273526907, + 0.0032680348958820105, + 0.00129262066911906, + 0.0011588344350457191, + 0.001178440172225237, + 0.001468401518650353, + 0.0013955568429082632, + 0.0027217944152653217, + 0.002963998820632696, + 0.0013798960717394948, + 0.004105864558368921, + 0.005493122152984142, + 0.003610947635024786, + 0.004341077525168657, + 0.006318580824881792, + 0.014179527759552002, + 0.01718069240450859, + 0.0048843445256352425, + 0.0020048499573022127, + 0.0016550563741475344, + 0.004435338079929352, + 0.0020605905447155237, + 0.00198762328363955, + 0.0028141746297478676, + 0.0020925416611135006, + 0.002447039121761918, + 0.003684374038130045, + 0.002925274893641472, + 0.002280913293361664, + 0.0036706922110170126, + 0.003359588561579585, + 0.012713748030364513, + 0.014254706911742687, + 0.002948097186163068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023267226293683052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195450011640787 + ], + "opposite_logprob_tokens": [ + -6.127195358276367 + ], + "logprob_reward": -0.002195450011640787, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195450011640787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3008653223514557, + 0.1388694941997528, + 0.13277283310890198, + 0.12325302511453629, + 0.06318946182727814, + 0.07022695243358612, + 0.028384866192936897, + 0.031986042857170105, + 0.047441598027944565, + 0.06301034241914749 + ], + "credit_map": [ + 0.03574756532907486, + 0.03540363162755966, + 0.05389571562409401, + 0.04091299697756767, + 0.0312967523932457, + 0.037761446088552475, + 0.03836512938141823, + 0.027482084929943085, + 0.011354115791618824, + 0.010253902524709702, + 0.008659295737743378, + 0.015185725875198841, + 0.015647681429982185, + 0.019871197640895844, + 0.0470617339015007, + 0.010835845023393631, + 0.010422121733427048, + 0.011896062642335892, + 0.009534263983368874, + 0.009391247294843197, + 0.014392335899174213, + 0.035000357776880264, + 0.03137597441673279, + 0.010760481469333172, + 0.010290957987308502, + 0.010283516719937325, + 0.022299176082015038, + 0.008789644576609135, + 0.008157577365636826, + 0.032672397792339325, + 0.018405411392450333, + 0.012354335747659206, + 0.005839155986905098, + 0.006658222991973162, + 0.0037913164123892784, + 0.0057985298335552216, + 0.006045747548341751, + 0.01439683511853218, + 0.0158315971493721, + 0.00482805585488677, + 0.00484274048358202, + 0.006584964692592621, + 0.00733674643561244, + 0.005882723722606897, + 0.012855117209255695, + 0.01215232815593481, + 0.01284601166844368, + 0.007726320531219244, + 0.0034082464408129454, + 0.00275822333060205, + 0.0036988372448831797, + 0.004382008221000433, + 0.002502120565623045, + 0.0025763725861907005, + 0.0059156292118132114, + 0.003143427660688758, + 0.004644934553653002, + 0.005089630372822285, + 0.0027678892947733402, + 0.0035782288759946823, + 0.004343049600720406, + 0.0031341142021119595, + 0.005141779780387878, + 0.00328641664236784, + 0.008363784290850163, + 0.004091421142220497, + 0.01109111774712801, + 0.003959919326007366, + 0.005751121323555708, + 0.003316842019557953, + 0.006376353092491627, + 0.004491036292165518, + 0.010555796325206757, + 0.005958507768809795, + 0.011981895193457603, + 0.004971550777554512, + 0.0060934629291296005, + 0.010462891310453415, + 0.00406109681352973, + 0.008925138972699642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.002195450011640787 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a style featuring thick, visible, and emotive brushstrokes similar to Van Gogh's signature style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky depicted as a turbulent vortex of deep indigo and fiery yellow with powerful swirls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, ethereal figure with pointed ears and flowing white hair positioned in the absolute center of the composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure wearing a simple, translucent white gown and standing above a luminous field of plants with vibrant, shimmering colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a wisp of glowing green energy rendered as a sinuous ribbon of paint slowly spiraling around the floating figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033685192465782166 + ], + "opposite_logprob_tokens": [ + -3.4086852073669434 + ], + "logprob_reward": -0.033685192465782166, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033685192465782166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3952193260192871, + 0.11602722853422165, + 0.1210702508687973, + 0.10893987864255905, + 0.02097475714981556, + 0.06886378675699234, + 0.060835808515548706, + 0.027813002467155457, + 0.0587347038090229, + 0.021521303802728653 + ], + "credit_map": [ + 0.02873416058719158, + 0.026856381446123123, + 0.020598864182829857, + 0.030971143394708633, + 0.036311279982328415, + 0.08512391149997711, + 0.1390395164489746, + 0.027584047988057137, + 0.008580696769058704, + 0.0073735062032938, + 0.013865347020328045, + 0.010628456249833107, + 0.006786446087062359, + 0.028824681416153908, + 0.029740162193775177, + 0.010227932594716549, + 0.008624635636806488, + 0.008530240505933762, + 0.005191244184970856, + 0.009429439902305603, + 0.011423764750361443, + 0.025470098480582237, + 0.04341711476445198, + 0.008983710780739784, + 0.007260680664330721, + 0.007012888789176941, + 0.005815605632960796, + 0.008150516077876091, + 0.009018433280289173, + 0.027920743450522423, + 0.035574160516262054, + 0.008186843246221542, + 0.0015705979894846678, + 0.0020328920800238848, + 0.0015386391896754503, + 0.002764221280813217, + 0.003054416738450527, + 0.0015359197277575731, + 0.006136024370789528, + 0.002342045772820711, + 0.005235896445810795, + 0.005389510188251734, + 0.00460059754550457, + 0.005218222737312317, + 0.007024472113698721, + 0.013966524973511696, + 0.02235853299498558, + 0.005070030223578215, + 0.0071329218335449696, + 0.0047440435737371445, + 0.008358410559594631, + 0.005367807112634182, + 0.005077426787465811, + 0.008448981679975986, + 0.015443066135048866, + 0.006263153161853552, + 0.002428451320156455, + 0.0024637500755488873, + 0.00159727712161839, + 0.0026487659197300673, + 0.00307052256539464, + 0.0038503219839185476, + 0.009070989675819874, + 0.002682923339307308, + 0.004848622251302004, + 0.0037330430932343006, + 0.0066206022165715694, + 0.004640286322683096, + 0.004185613710433245, + 0.012814554385840893, + 0.014365063048899174, + 0.00752691924571991, + 0.0034140150528401136, + 0.002379015553742647, + 0.003008044557645917, + 0.0027369919698685408, + 0.0021851828787475824, + 0.00139284273609519, + 0.00391444331035018, + 0.0024907668121159077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.033685192465782166 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_tensor([3117, 3117])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_tensor([3117, 3117])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9cc795a23d1141a8ae918716cc258afa14737ef7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_tensor([3117, 3117])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle close-up depicts an Asian man, his long black hair fanning out on a dark pillow, his eyes gently closed, his face in a state of calm, almost melancholic repose. He wears a loose, black silk robe, its subtle sheen catching the faint light as it drapes over his still form. He lies on his back on an ornate bed, his head and shoulders just in front of a dark cast-iron headboard, its intricate structure embellished with delicate, shimmering gold filigree accents. The background behind him is a deep, shadowy space, the intricate scarlet and sapphire damask patterns of the walls or tapestries barely discernible, adding a layer of textural depth. His pose is completely still, with only the almost imperceptible, slow rise and fall of his chest indicating the rhythm of his breathing. The scene is lit from the side by a single, dim light source, casting soft highlights on the outline of his face and the gold details of the railing, while plunging the rest of the frame into deep, softly edged shadows. This chiaroscuro lighting, combined with a cool overall color temperature, creates a deeply somber and introspective atmosphere, rendered with a realistic style that captures every subtle texture from cool metals to soft fabrics.", + "seed": 22814, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_3117_seed_22814_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019523890083655716, + "vlm_logprob": -0.019523890083655716, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.019523890083655716, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051503270864486694 + ], + "opposite_logprob_tokens": [ + -3.0515031814575195 + ], + "logprob_reward": -0.051503270864486694, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051503270864486694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3987908959388733, + 0.10974067449569702, + 0.04966239631175995, + 0.05875410512089729, + 0.05780872330069542, + 0.11027075350284576, + 0.1268017590045929, + 0.03419911861419678, + 0.016479993239045143, + 0.037491582334041595 + ], + "credit_map": [ + 0.09024502336978912, + 0.025162864476442337, + 0.07081031799316406, + 0.03805486857891083, + 0.026868781074881554, + 0.05799654871225357, + 0.033160705119371414, + 0.05649174749851227, + 0.01614218018949032, + 0.010141290724277496, + 0.012435718439519405, + 0.013221127912402153, + 0.01458785217255354, + 0.01747194118797779, + 0.009942213073372841, + 0.01579834707081318, + 0.0075696720741689205, + 0.004158745054155588, + 0.004365116823464632, + 0.006966204382479191, + 0.005935050547122955, + 0.008127791807055473, + 0.004635031335055828, + 0.0079047791659832, + 0.010236145928502083, + 0.006899550091475248, + 0.00557636097073555, + 0.007835455238819122, + 0.006320098880678415, + 0.008812015876173973, + 0.004965383093804121, + 0.008109090849757195, + 0.012050820514559746, + 0.006803147029131651, + 0.0057953414507210255, + 0.008766675367951393, + 0.007243306841701269, + 0.005900249350816011, + 0.003960304893553257, + 0.007288873661309481, + 0.01992291398346424, + 0.011654173955321312, + 0.012487943284213543, + 0.01780650205910206, + 0.012659760192036629, + 0.011047229170799255, + 0.01058770902454853, + 0.014104515314102173, + 0.0214635469019413, + 0.01414431445300579, + 0.014362429268658161, + 0.015098493546247482, + 0.018857112154364586, + 0.016656916588544846, + 0.008171798661351204, + 0.01804712601006031, + 0.004292056895792484, + 0.003647483652457595, + 0.00347218569368124, + 0.00812672358006239, + 0.0027198633179068565, + 0.00437199929729104, + 0.0032195805106312037, + 0.004349221009761095, + 0.002401340054348111, + 0.0014950637705624104, + 0.002317858161404729, + 0.002634001662954688, + 0.0019882756751030684, + 0.0020699661690741777, + 0.0013140810187906027, + 0.002259405329823494, + 0.004504118114709854, + 0.0031518421601504087, + 0.005871214438229799, + 0.005122095812112093, + 0.005934194196015596, + 0.004768111277371645, + 0.0030756997875869274, + 0.005064305383712053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.051503270864486694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023319518193602562 + ], + "opposite_logprob_tokens": [ + -3.7733194828033447 + ], + "logprob_reward": -0.023319518193602562, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023319518193602562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14293548464775085, + 0.3291131258010864, + 0.150787815451622, + 0.12562984228134155, + 0.09259077906608582, + 0.02603931911289692, + 0.07704777270555496, + 0.013335642404854298, + 0.016243334859609604, + 0.026276908814907074 + ], + "credit_map": [ + 0.02833246998488903, + 0.009535936638712883, + 0.021428845822811127, + 0.014843674376606941, + 0.012547002173960209, + 0.02304067090153694, + 0.01166963204741478, + 0.021537240594625473, + 0.03577617183327675, + 0.03423905745148659, + 0.04166566953063011, + 0.04370421916246414, + 0.05398331955075264, + 0.05029620975255966, + 0.024442749097943306, + 0.04500570520758629, + 0.01370010431855917, + 0.01180480606853962, + 0.01876293681561947, + 0.016446411609649658, + 0.02796821855008602, + 0.02913564071059227, + 0.010173785500228405, + 0.022795895114541054, + 0.009384164586663246, + 0.00898008793592453, + 0.018637526780366898, + 0.01017170399427414, + 0.024512259289622307, + 0.02596954256296158, + 0.008352979086339474, + 0.01962156407535076, + 0.009473966434597969, + 0.006961334962397814, + 0.012972895056009293, + 0.007101757451891899, + 0.022266196087002754, + 0.015231647528707981, + 0.005851515103131533, + 0.012731464579701424, + 0.0029392403084784746, + 0.0035310883540660143, + 0.0023641278967261314, + 0.0053061326034367085, + 0.002811334328725934, + 0.0033410999458283186, + 0.002315181540325284, + 0.0034311125054955482, + 0.012867070734500885, + 0.006694866809993982, + 0.007981011644005775, + 0.011629227548837662, + 0.010706941597163677, + 0.011159123852849007, + 0.005574278067797422, + 0.010435238480567932, + 0.0015996885485947132, + 0.0012994911521673203, + 0.0017586323665454984, + 0.002379965502768755, + 0.001333150314167142, + 0.0017783730290830135, + 0.0012276929337531328, + 0.0019586465787142515, + 0.0022044305223971605, + 0.0012654858874157071, + 0.0024351608008146286, + 0.0025546757970005274, + 0.0017250727396458387, + 0.0021412596106529236, + 0.0015344048151746392, + 0.0023828439880162477, + 0.003716273931786418, + 0.0022936954628676176, + 0.005147511139512062, + 0.002480007940903306, + 0.0030184618663042784, + 0.003784888656809926, + 0.0017674773698672652, + 0.004068591166287661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.023319518193602562 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072147637605667 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072147637605667, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072147637605667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23141376674175262, + 0.08496562391519547, + 0.0530891939997673, + 0.037671104073524475, + 0.07887288928031921, + 0.15703928470611572, + 0.08456570655107498, + 0.16257783770561218, + 0.048309579491615295, + 0.061495039612054825 + ], + "credit_map": [ + 0.044654976576566696, + 0.017126141116023064, + 0.03648173436522484, + 0.03265504166483879, + 0.01599958725273609, + 0.025368666276335716, + 0.02206890843808651, + 0.03705870360136032, + 0.008601550944149494, + 0.005936167668551207, + 0.012019076384603977, + 0.013468224555253983, + 0.013836861588060856, + 0.010532008484005928, + 0.00968968402594328, + 0.010882049798965454, + 0.005995558574795723, + 0.00470321299508214, + 0.005233963020145893, + 0.005573974922299385, + 0.009475686587393284, + 0.009949901141226292, + 0.004202336072921753, + 0.007954557426273823, + 0.003797847079113126, + 0.0024387426674365997, + 0.00457935081794858, + 0.003907823469489813, + 0.007085838355123997, + 0.007010643370449543, + 0.0031326748430728912, + 0.005718185566365719, + 0.008256715722382069, + 0.005442303605377674, + 0.011300359852612019, + 0.008410217240452766, + 0.017084617167711258, + 0.011965722776949406, + 0.005918692797422409, + 0.010494256392121315, + 0.027652453631162643, + 0.018510842695832253, + 0.015644263476133347, + 0.029198959469795227, + 0.013705707155168056, + 0.0179793369024992, + 0.014513194561004639, + 0.01983453519642353, + 0.015104629099369049, + 0.010899693705141544, + 0.008959710597991943, + 0.011937929317355156, + 0.01049107126891613, + 0.009450897574424744, + 0.005797813646495342, + 0.011923957616090775, + 0.03469735011458397, + 0.011496506631374359, + 0.028796952217817307, + 0.013277364894747734, + 0.016517603769898415, + 0.02001436986029148, + 0.009692251682281494, + 0.02808542363345623, + 0.011238754726946354, + 0.004785383585840464, + 0.008164376020431519, + 0.006625881884247065, + 0.002884071320295334, + 0.004345353692770004, + 0.0033378591760993004, + 0.006927897688001394, + 0.018454940989613533, + 0.004034274257719517, + 0.009535908699035645, + 0.006256215274333954, + 0.004616269841790199, + 0.0065977624617516994, + 0.003731236793100834, + 0.008268430829048157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.011072147637605667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978435855358839 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.00978435855358839, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978435855358839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2643545866012573, + 0.09183179587125778, + 0.07873642444610596, + 0.04967186599969864, + 0.058961525559425354, + 0.09553515166044235, + 0.29047712683677673, + 0.023827243596315384, + 0.023800605908036232, + 0.022803587839007378 + ], + "credit_map": [ + 0.051085975021123886, + 0.01982298120856285, + 0.03112480603158474, + 0.04124584421515465, + 0.01919543370604515, + 0.0349799282848835, + 0.026483722031116486, + 0.040415938943624496, + 0.011835132725536823, + 0.0056172991171479225, + 0.008746333420276642, + 0.01437703613191843, + 0.016442209482192993, + 0.013483952730894089, + 0.008378304541110992, + 0.012951537035405636, + 0.012224447913467884, + 0.004574041347950697, + 0.007095567882061005, + 0.014076792635023594, + 0.007622492033988237, + 0.014119988307356834, + 0.0085847033187747, + 0.01043839380145073, + 0.00832405872642994, + 0.0033246208913624287, + 0.006806564051657915, + 0.005881727673113346, + 0.009150669910013676, + 0.006596560589969158, + 0.003407975658774376, + 0.0061796908266842365, + 0.007502248976379633, + 0.004857326857745647, + 0.005641956347972155, + 0.012097383849322796, + 0.006828492972999811, + 0.008208498358726501, + 0.006160525139421225, + 0.0076651000417768955, + 0.013806518167257309, + 0.014793213456869125, + 0.008605784736573696, + 0.01888905093073845, + 0.009491409175097942, + 0.008777709677815437, + 0.008556409738957882, + 0.012615066021680832, + 0.053299304097890854, + 0.019935069605708122, + 0.036725956946611404, + 0.03437993302941322, + 0.044362545013427734, + 0.04204581677913666, + 0.02013503946363926, + 0.03959348797798157, + 0.004278718493878841, + 0.0019333602394908667, + 0.0035385838709771633, + 0.0025347494520246983, + 0.003164884867146611, + 0.002725851023569703, + 0.0018544066697359085, + 0.0037966903764754534, + 0.004098350182175636, + 0.0015940545126795769, + 0.00417405366897583, + 0.0032160968985408545, + 0.0020755610894411802, + 0.0032383280340582132, + 0.0017572931246832013, + 0.0036468731705099344, + 0.0034737635869532824, + 0.002006629714742303, + 0.0035939549561589956, + 0.00296406471170485, + 0.0020114292856305838, + 0.00305410148575902, + 0.0021680444478988647, + 0.0035316015128046274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.00978435855358839 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019401551689952612 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.0019401551689952612, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019401551689952612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21860475838184357, + 0.08982206135988235, + 0.03709292784333229, + 0.06130153313279152, + 0.09555529803037643, + 0.1450578272342682, + 0.22934268414974213, + 0.06336848437786102, + 0.02572453022003174, + 0.03412994369864464 + ], + "credit_map": [ + 0.04385470971465111, + 0.013715348206460476, + 0.033221613615751266, + 0.03499515727162361, + 0.013614973984658718, + 0.02514505386352539, + 0.022676408290863037, + 0.03138149157166481, + 0.011676616966724396, + 0.011990142054855824, + 0.009953167289495468, + 0.016960851848125458, + 0.012172685004770756, + 0.008288188837468624, + 0.006243252661079168, + 0.012537151575088501, + 0.00496582081541419, + 0.0034869415685534477, + 0.004646365065127611, + 0.004975130315870047, + 0.005891985725611448, + 0.004991516005247831, + 0.003180594416335225, + 0.0049545783549547195, + 0.008836040273308754, + 0.00581671018153429, + 0.006064251065254211, + 0.011052006855607033, + 0.006439086515456438, + 0.009772625751793385, + 0.004899975843727589, + 0.008420832455158234, + 0.010138556361198425, + 0.008776938542723656, + 0.012011312879621983, + 0.0174114890396595, + 0.012046407908201218, + 0.014247600920498371, + 0.008357061073184013, + 0.012565931305289268, + 0.027524534612894058, + 0.014008835889399052, + 0.015780562534928322, + 0.024364544078707695, + 0.015417766757309437, + 0.01486872136592865, + 0.015048235654830933, + 0.018044618889689445, + 0.03524230048060417, + 0.02275014854967594, + 0.02802158147096634, + 0.03284580633044243, + 0.034089021384716034, + 0.028668560087680817, + 0.016808776184916496, + 0.030916493386030197, + 0.00715218111872673, + 0.0038064499385654926, + 0.008068679831922054, + 0.00912732258439064, + 0.009356675669550896, + 0.010862528346478939, + 0.005339369177818298, + 0.009655279107391834, + 0.0040891412645578384, + 0.0014764026273041964, + 0.0034045788925141096, + 0.002978648990392685, + 0.002737842733040452, + 0.005056105554103851, + 0.0018451621290296316, + 0.004136648494750261, + 0.0057148560881614685, + 0.004086146596819162, + 0.0051282113417983055, + 0.0046165441162884235, + 0.0030583487823605537, + 0.003948313184082508, + 0.003090829821303487, + 0.004486694931983948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0019401551689952612 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up depicts an Asian man, his long black hair fanning out on a dark pillow, his eyes gently closed, his face in a state of calm, almost melancholic repose. He wears a loose, black silk robe, its subtle sheen catching the faint light as it drapes over his still form. He lies on his back on an ornate bed, his head and shoulders just in front of a dark cast-iron headboard, its intricate structure embellished with delicate, shimmering gold filigree accents. The background behind him is a deep, shadowy space, the intricate scarlet and sapphire damask patterns of the walls or tapestries barely discernible, adding a layer of textural depth. His pose is completely still, with only the almost imperceptible, slow rise and fall of his chest indicating the rhythm of his breathing. The scene is lit from the side by a single, dim light source, casting soft highlights on the outline of his face and the gold details of the railing, while plunging the rest of the frame into deep, softly edged shadows. This chiaroscuro lighting, combined with a cool overall color temperature, creates a deeply somber and introspective atmosphere, rendered with a realistic style that captures every subtle texture from cool metals to soft fabrics.", + "seed": 22815, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_3117_seed_22815_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02542872540652752, + "vlm_logprob": -0.02542872540652752, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02542872540652752, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043619439005851746 + ], + "opposite_logprob_tokens": [ + -3.168619394302368 + ], + "logprob_reward": -0.043619439005851746, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043619439005851746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4643065333366394, + 0.1666337102651596, + 0.07778257876634598, + 0.048298124223947525, + 0.034082554280757904, + 0.030270300805568695, + 0.04167722538113594, + 0.03799859434366226, + 0.041573382914066315, + 0.05737704038619995 + ], + "credit_map": [ + 0.05221812427043915, + 0.03564443066716194, + 0.05351518094539642, + 0.10903267562389374, + 0.03153096139431, + 0.03697647899389267, + 0.07530175149440765, + 0.07008691877126694, + 0.015675662085413933, + 0.01533561572432518, + 0.018717652186751366, + 0.034534189850091934, + 0.014225286431610584, + 0.013690352439880371, + 0.029882362112402916, + 0.02457260899245739, + 0.008527982048690319, + 0.00806807167828083, + 0.008311008103191853, + 0.01388796977698803, + 0.007353509310632944, + 0.006988278590142727, + 0.013004488311707973, + 0.011641272343695164, + 0.005776264704763889, + 0.004868088290095329, + 0.005452824290841818, + 0.00851335097104311, + 0.004765188787132502, + 0.0038324338383972645, + 0.008247125893831253, + 0.0068428474478423595, + 0.004358855541795492, + 0.0032498096115887165, + 0.0034548444673419, + 0.006313698831945658, + 0.003401099005714059, + 0.0030513948295265436, + 0.0051011755131185055, + 0.005151675548404455, + 0.003984651528298855, + 0.0030996992718428373, + 0.003442239947617054, + 0.0047554257325828075, + 0.0031759089324623346, + 0.0028749010525643826, + 0.004435716662555933, + 0.004501756280660629, + 0.005705372430384159, + 0.004128982778638601, + 0.004260839894413948, + 0.007526817265897989, + 0.004488418344408274, + 0.003892695065587759, + 0.005701992195099592, + 0.005972110200673342, + 0.004493131767958403, + 0.004226857330650091, + 0.003949380479753017, + 0.007664152886718512, + 0.003616733243688941, + 0.0034424965269863605, + 0.005332721397280693, + 0.0052731214091181755, + 0.004266963806003332, + 0.0041360571049153805, + 0.004387033171951771, + 0.011293559335172176, + 0.0032304455526173115, + 0.004006735049188137, + 0.004578501917421818, + 0.005674085579812527, + 0.005941228941082954, + 0.00605465192347765, + 0.006376011297106743, + 0.01457707118242979, + 0.0044185989536345005, + 0.005023511126637459, + 0.00698217423632741, + 0.00800379179418087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.043619439005851746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199479207396507 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.018199479207396507, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199479207396507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30879342555999756, + 0.21319209039211273, + 0.11884423345327377, + 0.050438545644283295, + 0.03894466534256935, + 0.028936656191945076, + 0.05995187908411026, + 0.057996973395347595, + 0.046612419188022614, + 0.0762891098856926 + ], + "credit_map": [ + 0.044047653675079346, + 0.025854498147964478, + 0.03465624153614044, + 0.051855891942977905, + 0.04534652829170227, + 0.03876178339123726, + 0.02577310986816883, + 0.04249775409698486, + 0.032465994358062744, + 0.019444920122623444, + 0.026182768866419792, + 0.025402206927537918, + 0.02912081405520439, + 0.026706432923674583, + 0.022696731612086296, + 0.031172219663858414, + 0.017224980518221855, + 0.012280402705073357, + 0.014542816206812859, + 0.012837221845984459, + 0.01755221001803875, + 0.015066894702613354, + 0.01202110294252634, + 0.01731860637664795, + 0.007635691203176975, + 0.00491351680830121, + 0.006727423053234816, + 0.005398130044341087, + 0.008028881624341011, + 0.0058250934816896915, + 0.0041684843599796295, + 0.007741323672235012, + 0.005664609372615814, + 0.0040069459937512875, + 0.005448014009743929, + 0.004869296681135893, + 0.0048512001521885395, + 0.004349238704890013, + 0.003700579982250929, + 0.006054776255041361, + 0.004969427362084389, + 0.002835243009030819, + 0.002826580312103033, + 0.005627359263598919, + 0.0033791016321629286, + 0.003058329224586487, + 0.002440830459818244, + 0.003799784928560257, + 0.008834107778966427, + 0.005788447335362434, + 0.00826142355799675, + 0.007072745822370052, + 0.008708660490810871, + 0.006716724950820208, + 0.005265289917588234, + 0.009304477833211422, + 0.008139684796333313, + 0.0061831725761294365, + 0.007536397781223059, + 0.008952267467975616, + 0.007128796074539423, + 0.005862940102815628, + 0.005977162625640631, + 0.008216550573706627, + 0.009910071268677711, + 0.005289460998028517, + 0.004103787709027529, + 0.0061600180342793465, + 0.006182747893035412, + 0.004563589580357075, + 0.003489538561552763, + 0.006913207471370697, + 0.011444568634033203, + 0.007278055418282747, + 0.009906072169542313, + 0.012132757343351841, + 0.00950308982282877, + 0.006789518054574728, + 0.00905998982489109, + 0.010175052098929882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018199479207396507 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817360892891884 + ], + "opposite_logprob_tokens": [ + -4.018173694610596 + ], + "logprob_reward": -0.01817360892891884, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817360892891884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3723725974559784, + 0.16011975705623627, + 0.1301872283220291, + 0.05911815166473389, + 0.0331939272582531, + 0.042275384068489075, + 0.03493225574493408, + 0.037318285554647446, + 0.05264366418123245, + 0.07783865928649902 + ], + "credit_map": [ + 0.07273536920547485, + 0.045582134276628494, + 0.03584728017449379, + 0.0383426807820797, + 0.04885327070951462, + 0.034649141132831573, + 0.04615573212504387, + 0.050207022577524185, + 0.032148152589797974, + 0.018283886834979057, + 0.015962490811944008, + 0.01835157908499241, + 0.022640332579612732, + 0.01790054515004158, + 0.014181538484990597, + 0.020651238039135933, + 0.026343658566474915, + 0.01595987379550934, + 0.013667644001543522, + 0.013532686047255993, + 0.017092611640691757, + 0.013607317581772804, + 0.01361997239291668, + 0.01636347360908985, + 0.012218059971928596, + 0.007088479120284319, + 0.0056899054907262325, + 0.006080990191549063, + 0.008061540313065052, + 0.006623196881264448, + 0.005776882637292147, + 0.007579101715236902, + 0.006982820108532906, + 0.0037509347312152386, + 0.0030432085040956736, + 0.0043341596610844135, + 0.0039955624379217625, + 0.004206809215247631, + 0.002784430282190442, + 0.004096005577594042, + 0.008719529956579208, + 0.005090181715786457, + 0.003849305212497711, + 0.004379706922918558, + 0.006096856202930212, + 0.004436296410858631, + 0.003963459748774767, + 0.005740047432482243, + 0.00607728399336338, + 0.0038129386957734823, + 0.003701381618157029, + 0.004808584228157997, + 0.004878836218267679, + 0.0035282557364553213, + 0.0034662799444049597, + 0.004658697638660669, + 0.005129392258822918, + 0.003510270267724991, + 0.004440560936927795, + 0.007000666111707687, + 0.004200755152851343, + 0.0027558624278753996, + 0.005046732723712921, + 0.005234051030129194, + 0.008581005968153477, + 0.005123532377183437, + 0.004682090599089861, + 0.008974057622253895, + 0.007025802042335272, + 0.004766315221786499, + 0.006259039975702763, + 0.007231823168694973, + 0.010908426716923714, + 0.00669363047927618, + 0.008920534513890743, + 0.013054931536316872, + 0.01023919228464365, + 0.007100197020918131, + 0.010593325830996037, + 0.01032843254506588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01817360892891884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043059516698122025 + ], + "opposite_logprob_tokens": [ + -3.1680595874786377 + ], + "logprob_reward": -0.043059516698122025, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043059516698122025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33302080631256104, + 0.1654842346906662, + 0.09583161771297455, + 0.06477347016334534, + 0.04166253283619881, + 0.04642810672521591, + 0.058447763323783875, + 0.0504385270178318, + 0.05972031131386757, + 0.08419269323348999 + ], + "credit_map": [ + 0.04785609245300293, + 0.04033605009317398, + 0.03810102492570877, + 0.051586393266916275, + 0.04281692951917648, + 0.040232472121715546, + 0.02986782044172287, + 0.0422239825129509, + 0.027109995484352112, + 0.019169334322214127, + 0.016729874536395073, + 0.027556397020816803, + 0.01805739849805832, + 0.01725282147526741, + 0.017376158386468887, + 0.022232238203287125, + 0.01599050499498844, + 0.010374466888606548, + 0.010340998880565166, + 0.01313786581158638, + 0.011499390006065369, + 0.011059513315558434, + 0.011114846915006638, + 0.012314017862081528, + 0.012737075798213482, + 0.006546346005052328, + 0.006386658176779747, + 0.008153475821018219, + 0.008919541724026203, + 0.008618365041911602, + 0.00485216872766614, + 0.008559833280742168, + 0.007825636304914951, + 0.0037562598008662462, + 0.005388293415307999, + 0.004589935299009085, + 0.00562910595908761, + 0.00553860841318965, + 0.003849881235510111, + 0.005084811244159937, + 0.009337296709418297, + 0.004956597462296486, + 0.004925638437271118, + 0.006678537931293249, + 0.005010658875107765, + 0.004834424238651991, + 0.004605801776051521, + 0.006079147104173899, + 0.008832598105072975, + 0.005998112726956606, + 0.00684545561671257, + 0.009987867437303066, + 0.007566604297608137, + 0.006285006180405617, + 0.005206110887229443, + 0.007725999224931002, + 0.006986217573285103, + 0.003868442727252841, + 0.0067877694964408875, + 0.011114422231912613, + 0.004912939388304949, + 0.0035945335403084755, + 0.006393649149686098, + 0.006780544761568308, + 0.010021449998021126, + 0.0055615585297346115, + 0.005580848082900047, + 0.009678601287305355, + 0.008304081857204437, + 0.006784666795283556, + 0.005447926465421915, + 0.008341173641383648, + 0.013500173576176167, + 0.0073409732431173325, + 0.008364854380488396, + 0.01175101287662983, + 0.014426781795918941, + 0.007056194823235273, + 0.009822758845984936, + 0.011929932981729507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.043059516698122025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409158319234848 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.00409158319234848, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409158319234848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3355032801628113, + 0.11586278676986694, + 0.06242573261260986, + 0.059907421469688416, + 0.06760496646165848, + 0.07127682119607925, + 0.03586546331644058, + 0.05467507615685463, + 0.06335535645484924, + 0.1335231065750122 + ], + "credit_map": [ + 0.04398384690284729, + 0.033905029296875, + 0.0375034399330616, + 0.06801427900791168, + 0.03129380941390991, + 0.027605941519141197, + 0.046203866600990295, + 0.04699304699897766, + 0.017464691773056984, + 0.012947376817464828, + 0.012723984196782112, + 0.024974830448627472, + 0.012690725736320019, + 0.010265465825796127, + 0.011233643628656864, + 0.013562066480517387, + 0.00836150161921978, + 0.0074053071439266205, + 0.006921147927641869, + 0.014222082681953907, + 0.005936284549534321, + 0.006494006607681513, + 0.006116610951721668, + 0.006968785542994738, + 0.011476435698568821, + 0.007370311766862869, + 0.0058766077272593975, + 0.00814545527100563, + 0.008314457722008228, + 0.006667731329798698, + 0.004955461714416742, + 0.007100959308445454, + 0.008993263356387615, + 0.007437278516590595, + 0.007325088605284691, + 0.014573958702385426, + 0.006954710930585861, + 0.006986420135945082, + 0.006481443997472525, + 0.008852804079651833, + 0.014887606725096703, + 0.008692207746207714, + 0.0061585623770952225, + 0.011886452324688435, + 0.00882719550281763, + 0.006142694037407637, + 0.004906021058559418, + 0.009776076301932335, + 0.0059638191014528275, + 0.004062248859554529, + 0.00357087142765522, + 0.0042685954831540585, + 0.005191686097532511, + 0.0034956710878759623, + 0.004243473522365093, + 0.005069097504019737, + 0.009490038268268108, + 0.006090243346989155, + 0.006646847818046808, + 0.006941367406398058, + 0.007463790476322174, + 0.004731991793960333, + 0.006124546751379967, + 0.007186246570199728, + 0.012634974904358387, + 0.006457771640270948, + 0.005013883113861084, + 0.006778517737984657, + 0.010295212268829346, + 0.007599789649248123, + 0.004890854004770517, + 0.00968435499817133, + 0.026056762784719467, + 0.01410625223070383, + 0.01245012879371643, + 0.011857518926262856, + 0.024998681619763374, + 0.01277341041713953, + 0.01074072066694498, + 0.020539626479148865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.00409158319234848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up depicts an Asian man, his long black hair fanning out on a dark pillow, his eyes gently closed, his face in a state of calm, almost melancholic repose. He wears a loose, black silk robe, its subtle sheen catching the faint light as it drapes over his still form. He lies on his back on an ornate bed, his head and shoulders just in front of a dark cast-iron headboard, its intricate structure embellished with delicate, shimmering gold filigree accents. The background behind him is a deep, shadowy space, the intricate scarlet and sapphire damask patterns of the walls or tapestries barely discernible, adding a layer of textural depth. His pose is completely still, with only the almost imperceptible, slow rise and fall of his chest indicating the rhythm of his breathing. The scene is lit from the side by a single, dim light source, casting soft highlights on the outline of his face and the gold details of the railing, while plunging the rest of the frame into deep, softly edged shadows. This chiaroscuro lighting, combined with a cool overall color temperature, creates a deeply somber and introspective atmosphere, rendered with a realistic style that captures every subtle texture from cool metals to soft fabrics.", + "seed": 22816, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_3117_seed_22816_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02274171942844987, + "vlm_logprob": -0.02274171942844987, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02274171942844987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04501451179385185 + ], + "opposite_logprob_tokens": [ + -3.1700146198272705 + ], + "logprob_reward": -0.04501451179385185, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04501451179385185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4225856065750122, + 0.11845916509628296, + 0.08986510336399078, + 0.040816325694322586, + 0.05017833411693573, + 0.03381652384996414, + 0.06341668218374252, + 0.050726570188999176, + 0.02992910146713257, + 0.10020659118890762 + ], + "credit_map": [ + 0.05130021274089813, + 0.051885250955820084, + 0.053756244480609894, + 0.041632384061813354, + 0.0488809198141098, + 0.06026304513216019, + 0.06502003967761993, + 0.04984751716256142, + 0.017062297090888023, + 0.014932045713067055, + 0.013912470079958439, + 0.008482583798468113, + 0.012355487793684006, + 0.015295412391424179, + 0.017010990530252457, + 0.019407883286476135, + 0.022459054365754128, + 0.00900091789662838, + 0.012678992003202438, + 0.007458078674972057, + 0.00727485865354538, + 0.012407143600285053, + 0.007597249001264572, + 0.010988813824951649, + 0.005901150405406952, + 0.003637113841250539, + 0.0063044787384569645, + 0.0037037392612546682, + 0.0034463333431631327, + 0.005808407440781593, + 0.0048282938078045845, + 0.007186806295067072, + 0.008902233093976974, + 0.005823674611747265, + 0.006368828006088734, + 0.004310957621783018, + 0.005036836490035057, + 0.0067508933134377, + 0.0054777562618255615, + 0.007507154252380133, + 0.005324188619852066, + 0.003722652094438672, + 0.004126259591430426, + 0.003707061754539609, + 0.0031415449921041727, + 0.004895708989351988, + 0.00400501349940896, + 0.004894092213362455, + 0.015193162485957146, + 0.004544765688478947, + 0.009996617212891579, + 0.004649760201573372, + 0.005850460380315781, + 0.009072765707969666, + 0.007091708015650511, + 0.007017442025244236, + 0.008877425454556942, + 0.005096897948533297, + 0.007590710185468197, + 0.005342928227037191, + 0.005296327173709869, + 0.006685557775199413, + 0.005016585346311331, + 0.0068201362155377865, + 0.0052365493029356, + 0.002556748688220978, + 0.0035923088435083628, + 0.003183194901794195, + 0.0021416121162474155, + 0.004952296614646912, + 0.0023197042755782604, + 0.005946686957031488, + 0.02594686858355999, + 0.008951627649366856, + 0.014848741702735424, + 0.009443757124245167, + 0.009091231971979141, + 0.014084205962717533, + 0.006804616656154394, + 0.011035538278520107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.04501451179385185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029837315902113914 + ], + "opposite_logprob_tokens": [ + -3.5298373699188232 + ], + "logprob_reward": -0.029837315902113914, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029837315902113914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3180280029773712, + 0.1372167021036148, + 0.09182171523571014, + 0.05074779689311981, + 0.059067171066999435, + 0.05913379415869713, + 0.05312257260084152, + 0.1246560662984848, + 0.034183796495199203, + 0.07202236354351044 + ], + "credit_map": [ + 0.044475097209215164, + 0.022652054205536842, + 0.05222126841545105, + 0.03223966434597969, + 0.0361536405980587, + 0.047161784023046494, + 0.038442667573690414, + 0.0446818508207798, + 0.01890474557876587, + 0.010772469453513622, + 0.02315191738307476, + 0.015456488355994225, + 0.014255711808800697, + 0.019802803173661232, + 0.017068004235625267, + 0.017804564908146858, + 0.010915560647845268, + 0.00602302048355341, + 0.014561595395207405, + 0.010565483011305332, + 0.009776083752512932, + 0.013704224489629269, + 0.010424258187413216, + 0.015851501375436783, + 0.007634501438587904, + 0.003568411571905017, + 0.009593348950147629, + 0.005168907810002565, + 0.005177083425223827, + 0.007724071852862835, + 0.006284635979682207, + 0.005596838891506195, + 0.007246952038258314, + 0.004112400114536285, + 0.00838898029178381, + 0.005758232902735472, + 0.006838019005954266, + 0.00840709824115038, + 0.005943222437053919, + 0.012372270226478577, + 0.007754109799861908, + 0.0032550443429499865, + 0.008714079856872559, + 0.007337228395044804, + 0.004962404258549213, + 0.008971253409981728, + 0.0055729481391608715, + 0.012566725723445415, + 0.009305140934884548, + 0.004493548534810543, + 0.007677772082388401, + 0.005022440571337938, + 0.005056945141404867, + 0.0076788850128650665, + 0.005621485877782106, + 0.008266358636319637, + 0.026565691456198692, + 0.012111167423427105, + 0.01690666377544403, + 0.012289765290915966, + 0.012588157318532467, + 0.016352593898773193, + 0.012601477093994617, + 0.015240561217069626, + 0.007299286779016256, + 0.0020566394086927176, + 0.004763059318065643, + 0.0032277964055538177, + 0.0032674961257725954, + 0.004789210855960846, + 0.0027729556895792484, + 0.006007354240864515, + 0.012106819078326225, + 0.006822020746767521, + 0.009051109664142132, + 0.00990278646349907, + 0.007482755463570356, + 0.010022429749369621, + 0.005885784514248371, + 0.010748662054538727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029837315902113914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817922666668892 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.01817922666668892, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817922666668892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1704249531030655, + 0.11872131377458572, + 0.10888194292783737, + 0.03867543116211891, + 0.04632109776139259, + 0.0353749580681324, + 0.033096861094236374, + 0.1604565680027008, + 0.05691436305642128, + 0.23113255202770233 + ], + "credit_map": [ + 0.03340015187859535, + 0.019733548164367676, + 0.025086719542741776, + 0.01591125875711441, + 0.013987712562084198, + 0.02331300638616085, + 0.02108989655971527, + 0.017902646213769913, + 0.02663000300526619, + 0.013217194005846977, + 0.019735489040613174, + 0.010951372794806957, + 0.008177699521183968, + 0.016094185411930084, + 0.012415103614330292, + 0.011500263586640358, + 0.03048664517700672, + 0.008859779685735703, + 0.016502317041158676, + 0.008114260621368885, + 0.009354881010949612, + 0.01607671193778515, + 0.009095286019146442, + 0.010392063297331333, + 0.009729625657200813, + 0.0034162786323577166, + 0.0058397091925144196, + 0.003236362710595131, + 0.0027688934933394194, + 0.005739421583712101, + 0.003139953361824155, + 0.004805183503776789, + 0.01335051003843546, + 0.00405995873734355, + 0.006745577324181795, + 0.003124058013781905, + 0.003671107580885291, + 0.006760668475180864, + 0.003111492143943906, + 0.005497725680470467, + 0.01018298789858818, + 0.0030755153857171535, + 0.004838143475353718, + 0.003181563224643469, + 0.0025057008024305105, + 0.005407785065472126, + 0.002445490565150976, + 0.0037377721164375544, + 0.007393601816147566, + 0.0026471090968698263, + 0.0053153568878769875, + 0.0029249140061438084, + 0.0028669051826000214, + 0.004840520676225424, + 0.0036596758291125298, + 0.003448777599260211, + 0.04800416901707649, + 0.013023502193391323, + 0.02567347325384617, + 0.010841241106390953, + 0.011828555725514889, + 0.0238894484937191, + 0.012117967009544373, + 0.015078218653798103, + 0.01600710116326809, + 0.00474353414028883, + 0.00925097893923521, + 0.00548371160402894, + 0.004433043301105499, + 0.00849198643118143, + 0.004308219533413649, + 0.004195793531835079, + 0.052202027291059494, + 0.0243242010474205, + 0.033052265644073486, + 0.02414761297404766, + 0.021131077781319618, + 0.03229327127337456, + 0.01672312058508396, + 0.027258966118097305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.01817922666668892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818976178765297 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.01818976178765297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818976178765297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29691222310066223, + 0.1263362318277359, + 0.09956524521112442, + 0.0324927493929863, + 0.05211125314235687, + 0.05147717520594597, + 0.048413459211587906, + 0.06835683435201645, + 0.04745584353804588, + 0.1768788993358612 + ], + "credit_map": [ + 0.05855654552578926, + 0.03596237674355507, + 0.04279087856411934, + 0.03258804231882095, + 0.029291406273841858, + 0.03727540373802185, + 0.030236447229981422, + 0.030211124569177628, + 0.018536262214183807, + 0.01235560979694128, + 0.0206217672675848, + 0.01590397022664547, + 0.013519836589694023, + 0.01674632541835308, + 0.0113465441390872, + 0.017305925488471985, + 0.01329207606613636, + 0.007835326716303825, + 0.016014855355024338, + 0.012463046237826347, + 0.008968441747128963, + 0.01400707010179758, + 0.008252541534602642, + 0.018731890246272087, + 0.006023567635565996, + 0.003671899437904358, + 0.004473602399230003, + 0.002838652115315199, + 0.0029941669199615717, + 0.004911010153591633, + 0.003516214434057474, + 0.00406363932415843, + 0.009349925443530083, + 0.004567306023091078, + 0.0073035890236496925, + 0.004741998855024576, + 0.004339876119047403, + 0.007229132577776909, + 0.0036893323995172977, + 0.010890097357332706, + 0.00993974506855011, + 0.0035929535515606403, + 0.007308958098292351, + 0.005031909793615341, + 0.003337802365422249, + 0.007694731466472149, + 0.0034143663942813873, + 0.011156708933413029, + 0.009629247710108757, + 0.004357198718935251, + 0.006882785353809595, + 0.005136048886924982, + 0.004802359268069267, + 0.00664003798738122, + 0.005279798526316881, + 0.005685982294380665, + 0.012424122542142868, + 0.007006313186138868, + 0.008761920034885406, + 0.007415322586894035, + 0.005624647717922926, + 0.010418379679322243, + 0.006092982832342386, + 0.010613146238029003, + 0.00977293774485588, + 0.0036242397036403418, + 0.006495035719126463, + 0.0055052670650184155, + 0.004858468659222126, + 0.006318680942058563, + 0.005877462215721607, + 0.0050037517212331295, + 0.03529737889766693, + 0.01523501705378294, + 0.023679085075855255, + 0.025934915989637375, + 0.018361568450927734, + 0.022426720708608627, + 0.014152767136693, + 0.021791450679302216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01818976178765297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21084481477737427, + 0.16115283966064453, + 0.08416105806827545, + 0.04398084431886673, + 0.047471653670072556, + 0.05026539787650108, + 0.08840327709913254, + 0.13820286095142365, + 0.07544795423746109, + 0.10006926208734512 + ], + "credit_map": [ + 0.030375322327017784, + 0.02331969514489174, + 0.025311382487416267, + 0.029863828793168068, + 0.02305573970079422, + 0.0288117416203022, + 0.029052672907710075, + 0.021054446697235107, + 0.032627686858177185, + 0.013906184583902359, + 0.02314133569598198, + 0.01683216169476509, + 0.014020208269357681, + 0.020642267540097237, + 0.01785125397145748, + 0.022131746634840965, + 0.020198939368128777, + 0.0071319956332445145, + 0.011441970244050026, + 0.008213157765567303, + 0.006986148189753294, + 0.011114981956779957, + 0.009981323964893818, + 0.009092539548873901, + 0.009180597960948944, + 0.0038164141587913036, + 0.006545213051140308, + 0.003566999454051256, + 0.0027030492201447487, + 0.006314242724329233, + 0.0044241189025342464, + 0.0074302093125879765, + 0.00718426238745451, + 0.004225750453770161, + 0.006749894004315138, + 0.004982189740985632, + 0.003842167789116502, + 0.007131348364055157, + 0.006196232046931982, + 0.007159804925322533, + 0.011209577322006226, + 0.005206417292356491, + 0.0061613041907548904, + 0.005364884622395039, + 0.004992143251001835, + 0.006595573853701353, + 0.0053290617652237415, + 0.005406435113400221, + 0.014937051571905613, + 0.006706134416162968, + 0.014465395361185074, + 0.008863299153745174, + 0.005903258454054594, + 0.01399245671927929, + 0.005361839197576046, + 0.01817384734749794, + 0.02782934531569481, + 0.01173065509647131, + 0.020824728533625603, + 0.012692990712821484, + 0.014791869558393955, + 0.01902034878730774, + 0.01589086465537548, + 0.01542205922305584, + 0.015286249108612537, + 0.005910523235797882, + 0.009137095883488655, + 0.0072935111820697784, + 0.007418781518936157, + 0.010253792628645897, + 0.008498056791722775, + 0.011649947613477707, + 0.026489129289984703, + 0.007653164677321911, + 0.017211876809597015, + 0.00845770351588726, + 0.009998450987040997, + 0.013418409042060375, + 0.007561971433460712, + 0.009278553538024426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0024877809919416904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up depicts an Asian man, his long black hair fanning out on a dark pillow, his eyes gently closed, his face in a state of calm, almost melancholic repose. He wears a loose, black silk robe, its subtle sheen catching the faint light as it drapes over his still form. He lies on his back on an ornate bed, his head and shoulders just in front of a dark cast-iron headboard, its intricate structure embellished with delicate, shimmering gold filigree accents. The background behind him is a deep, shadowy space, the intricate scarlet and sapphire damask patterns of the walls or tapestries barely discernible, adding a layer of textural depth. His pose is completely still, with only the almost imperceptible, slow rise and fall of his chest indicating the rhythm of his breathing. The scene is lit from the side by a single, dim light source, casting soft highlights on the outline of his face and the gold details of the railing, while plunging the rest of the frame into deep, softly edged shadows. This chiaroscuro lighting, combined with a cool overall color temperature, creates a deeply somber and introspective atmosphere, rendered with a realistic style that captures every subtle texture from cool metals to soft fabrics.", + "seed": 22817, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_3/index_3117_seed_22817_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04917542794719339, + "vlm_logprob": -0.04917542794719339, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04917542794719339, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14490535855293274 + ], + "opposite_logprob_tokens": [ + -2.0199053287506104 + ], + "logprob_reward": -0.14490535855293274, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14490535855293274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3902556598186493, + 0.06597044318914413, + 0.04728629067540169, + 0.053612589836120605, + 0.10195406526327133, + 0.06712077558040619, + 0.041173648089170456, + 0.03267508000135422, + 0.0767243430018425, + 0.12322709709405899 + ], + "credit_map": [ + 0.04508817568421364, + 0.08810257911682129, + 0.03266168758273125, + 0.05706637352705002, + 0.05894272401928902, + 0.04787331819534302, + 0.024997232481837273, + 0.03552358224987984, + 0.007366732228547335, + 0.006839031353592873, + 0.008047853596508503, + 0.009216777049005032, + 0.008795314468443394, + 0.008021516725420952, + 0.006678414531052113, + 0.011004804633557796, + 0.0037439758889377117, + 0.006485702004283667, + 0.004743409808725119, + 0.005962621420621872, + 0.005765710957348347, + 0.005162393674254417, + 0.006574816070497036, + 0.00884766411036253, + 0.006460348144173622, + 0.007410138845443726, + 0.005557909607887268, + 0.005819166079163551, + 0.008289048448204994, + 0.006188752595335245, + 0.005988653749227524, + 0.007898570969700813, + 0.013990462757647038, + 0.02154570259153843, + 0.009213531389832497, + 0.011210333555936813, + 0.014346189796924591, + 0.01269895676523447, + 0.00833938643336296, + 0.010609498247504234, + 0.0068807946518063545, + 0.006975733675062656, + 0.009010694921016693, + 0.0064641814678907394, + 0.012199144810438156, + 0.00817890651524067, + 0.004732553847134113, + 0.012678763829171658, + 0.005291082430630922, + 0.0037158774212002754, + 0.0051972344517707825, + 0.003849645843729377, + 0.0052037169225513935, + 0.0048042284324765205, + 0.005606028251349926, + 0.007505838293582201, + 0.0037898747250437737, + 0.0031885404605418444, + 0.003862344892695546, + 0.004053985234349966, + 0.004413290414959192, + 0.005712838843464851, + 0.003075667191296816, + 0.004578540101647377, + 0.007403113413602114, + 0.006778011564165354, + 0.008787032216787338, + 0.010106200352311134, + 0.013216734863817692, + 0.00944738183170557, + 0.008371936157345772, + 0.012613929808139801, + 0.009050587192177773, + 0.012876255437731743, + 0.011873858980834484, + 0.0127501729875803, + 0.026823675259947777, + 0.020502595230937004, + 0.01180300209671259, + 0.01754694990813732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.14490535855293274 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209825351834297 + ], + "opposite_logprob_tokens": [ + -2.812098264694214 + ], + "logprob_reward": -0.06209825351834297, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209825351834297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2284027338027954, + 0.2420642077922821, + 0.21104487776756287, + 0.0569254569709301, + 0.053189072757959366, + 0.045060835778713226, + 0.051915232092142105, + 0.04854627698659897, + 0.028769638389348984, + 0.03408166393637657 + ], + "credit_map": [ + 0.025234339758753777, + 0.0324474573135376, + 0.02279640920460224, + 0.03656483814120293, + 0.03749444708228111, + 0.029184360057115555, + 0.019297197461128235, + 0.025383682921528816, + 0.021369686350226402, + 0.02299562469124794, + 0.022810431197285652, + 0.028620779514312744, + 0.0370543971657753, + 0.03129256144165993, + 0.03433091565966606, + 0.04358980804681778, + 0.018292563036084175, + 0.01906156912446022, + 0.02046043984591961, + 0.020824884995818138, + 0.02274606004357338, + 0.026000365614891052, + 0.042686838656663895, + 0.04097212851047516, + 0.004255208186805248, + 0.00448595779016614, + 0.006142016965895891, + 0.005370386876165867, + 0.008456841111183167, + 0.006733433809131384, + 0.0071399318985641, + 0.01434167567640543, + 0.004642767831683159, + 0.00826779194176197, + 0.0057073961943387985, + 0.005372545216232538, + 0.010060964152216911, + 0.006688361521810293, + 0.006188666447997093, + 0.0062605757266283035, + 0.004195112269371748, + 0.004472728818655014, + 0.006057735066860914, + 0.0034615888725966215, + 0.010711830109357834, + 0.005497264210134745, + 0.0032063277903944254, + 0.007458245847374201, + 0.005993944592773914, + 0.008894255384802818, + 0.005702147260308266, + 0.005148064345121384, + 0.0066499849781394005, + 0.006389415357261896, + 0.004387956112623215, + 0.008749454282224178, + 0.005622717086225748, + 0.00559999980032444, + 0.0059064831584692, + 0.004878022707998753, + 0.00905986875295639, + 0.006431207060813904, + 0.004058402497321367, + 0.006989572662860155, + 0.0025756945833563805, + 0.0031890859827399254, + 0.0031922124326229095, + 0.0023389006964862347, + 0.005580343771725893, + 0.0037765218876302242, + 0.003242732957005501, + 0.00487414188683033, + 0.00398710323497653, + 0.0034850791562348604, + 0.004127800930291414, + 0.00413488456979394, + 0.004182488191872835, + 0.005452905781567097, + 0.003307001432403922, + 0.0054043931886553764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.06209825351834297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191068708896637 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191068708896637, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191068708896637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6550444960594177, + 0.07945466041564941, + 0.035668328404426575, + 0.02706262841820717, + 0.015317868441343307, + 0.033233750611543655, + 0.031607355922460556, + 0.04865187034010887, + 0.040574412792921066, + 0.033384621143341064 + ], + "credit_map": [ + 0.12730632722377777, + 0.15206597745418549, + 0.06783571094274521, + 0.049130551517009735, + 0.09051468223333359, + 0.0837785080075264, + 0.03450997173786163, + 0.04990268871188164, + 0.013006102293729782, + 0.014323118142783642, + 0.008655371144413948, + 0.009280999191105366, + 0.01084219291806221, + 0.010392560623586178, + 0.0046654557809233665, + 0.008288858458399773, + 0.00705809285864234, + 0.006840871647000313, + 0.00382987130433321, + 0.0038254845421761274, + 0.004491457249969244, + 0.003881871234625578, + 0.001964856870472431, + 0.0037758222315460443, + 0.005830368958413601, + 0.00397865055128932, + 0.0027598789893090725, + 0.0028803686145693064, + 0.003432169323787093, + 0.003240494290366769, + 0.0017092402558773756, + 0.0032314560376107693, + 0.0013809846714138985, + 0.0025325901806354523, + 0.001661773188970983, + 0.0018951751990243793, + 0.002078698482364416, + 0.0020657023414969444, + 0.0020286638755351305, + 0.0016742792213335633, + 0.004151687026023865, + 0.002919579390436411, + 0.0040898253209888935, + 0.002874771598726511, + 0.007754874881356955, + 0.00392505107447505, + 0.002323636319488287, + 0.005194323603063822, + 0.005603426601737738, + 0.0035043610259890556, + 0.004684836603701115, + 0.0036665063817054033, + 0.004329992923885584, + 0.003511083545163274, + 0.002055324614048004, + 0.004251824226230383, + 0.009921534918248653, + 0.00559559091925621, + 0.006684126798063517, + 0.005780373699963093, + 0.005863836035132408, + 0.005187737289816141, + 0.0032393254805356264, + 0.006379341706633568, + 0.0038676836993545294, + 0.004220488481223583, + 0.004257073625922203, + 0.0037294309586286545, + 0.007128594908863306, + 0.0054450323805212975, + 0.004916385747492313, + 0.007009721826761961, + 0.004758333787322044, + 0.0017958657117560506, + 0.004727661143988371, + 0.002596581354737282, + 0.005614928435534239, + 0.008429248817265034, + 0.002205691998824477, + 0.0032563055865466595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.014191068708896637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588263869285583 + ], + "opposite_logprob_tokens": [ + -3.8955881595611572 + ], + "logprob_reward": -0.020588263869285583, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588263869285583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4052122235298157, + 0.08198327571153641, + 0.08631911873817444, + 0.06693907082080841, + 0.04224073514342308, + 0.06940636783838272, + 0.09540725499391556, + 0.05170600116252899, + 0.06654228270053864, + 0.034243643283843994 + ], + "credit_map": [ + 0.06339559704065323, + 0.06830041110515594, + 0.04473212733864784, + 0.03858692944049835, + 0.07260952889919281, + 0.052920788526535034, + 0.031954292207956314, + 0.03271254152059555, + 0.009339832700788975, + 0.009971534833312035, + 0.008848894387483597, + 0.008480307646095753, + 0.011755447834730148, + 0.009318914264440536, + 0.010281604714691639, + 0.0139867402613163, + 0.010892568156123161, + 0.011571630835533142, + 0.008164526894688606, + 0.009677481837570667, + 0.01600891724228859, + 0.01018473319709301, + 0.007701328489929438, + 0.012117937207221985, + 0.004907744005322456, + 0.008154744282364845, + 0.006485996302217245, + 0.007310206536203623, + 0.008594361133873463, + 0.007235951721668243, + 0.010551267303526402, + 0.013698799535632133, + 0.004872363992035389, + 0.005549718625843525, + 0.005392466671764851, + 0.004991812165826559, + 0.005993808154016733, + 0.00508769741281867, + 0.0039975764229893684, + 0.006355288438498974, + 0.008547144010663033, + 0.008593035861849785, + 0.00850585289299488, + 0.006333002354949713, + 0.012960251420736313, + 0.007443453185260296, + 0.004627173766493797, + 0.012396460399031639, + 0.009661057032644749, + 0.025684107095003128, + 0.009594812989234924, + 0.011053631082177162, + 0.011306120082736015, + 0.011533046141266823, + 0.004936381708830595, + 0.011638096533715725, + 0.008621282875537872, + 0.005366368684917688, + 0.008274118416011333, + 0.0049296291545033455, + 0.006283309310674667, + 0.006141458172351122, + 0.004011767916381359, + 0.008078062906861305, + 0.006315367296338081, + 0.0053440178744494915, + 0.00783951673656702, + 0.005145640578120947, + 0.014236939139664173, + 0.010118544101715088, + 0.007719299290329218, + 0.009822954423725605, + 0.004383082967251539, + 0.0027726192492991686, + 0.005190011579543352, + 0.002985866740345955, + 0.00556385051459074, + 0.00449624378234148, + 0.0041391681879758835, + 0.004712801426649094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.020588263869285583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian man with long black hair lying on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose black silk robe with a visible sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background featuring a dark cast-iron headboard with gold filigree accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single dim side light source creating strong chiaroscuro shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly high-angle close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094195086508989 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094195086508989, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094195086508989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33147183060646057, + 0.06037484109401703, + 0.09231965243816376, + 0.10270865261554718, + 0.0770161971449852, + 0.09121928364038467, + 0.0981433093547821, + 0.04216918349266052, + 0.07919563353061676, + 0.02538139373064041 + ], + "credit_map": [ + 0.04382840543985367, + 0.05424303561449051, + 0.031671252101659775, + 0.03004555031657219, + 0.0405363030731678, + 0.04016650468111038, + 0.021855663508176804, + 0.06912513822317123, + 0.006922863889485598, + 0.0060496581718325615, + 0.00791382510215044, + 0.006370186805725098, + 0.007006754633039236, + 0.007765292655676603, + 0.0071744173765182495, + 0.01117184292525053, + 0.01310193445533514, + 0.012466066516935825, + 0.009607652202248573, + 0.007393410429358482, + 0.014579794369637966, + 0.010929971933364868, + 0.010273892432451248, + 0.013966930098831654, + 0.017509693279862404, + 0.014937839470803738, + 0.010667177848517895, + 0.010254550725221634, + 0.01380186714231968, + 0.011478645727038383, + 0.01085764728486538, + 0.013201230205595493, + 0.007141462992876768, + 0.011292314156889915, + 0.008194235153496265, + 0.010786350816488266, + 0.010466036386787891, + 0.011364134028553963, + 0.008496616035699844, + 0.0092750433832407, + 0.009984253905713558, + 0.011645307764410973, + 0.009796444326639175, + 0.011851759627461433, + 0.01778017170727253, + 0.00959661416709423, + 0.007682227063924074, + 0.012882503680884838, + 0.012713441625237465, + 0.01748826913535595, + 0.009906352497637272, + 0.011437298730015755, + 0.010575187392532825, + 0.011563820764422417, + 0.00972604937851429, + 0.014732892625033855, + 0.0038091400638222694, + 0.00343575538136065, + 0.004917604383081198, + 0.004999816417694092, + 0.005036473739892244, + 0.00733470031991601, + 0.004616777878254652, + 0.008018916472792625, + 0.007573199924081564, + 0.006891504395753145, + 0.00954471342265606, + 0.007123467978090048, + 0.017387067899107933, + 0.011228029616177082, + 0.007264066021889448, + 0.012183585204184055, + 0.002223644172772765, + 0.0022633448243141174, + 0.0036240676417946815, + 0.002839575055986643, + 0.003197402460500598, + 0.004296860191971064, + 0.0027618955355137587, + 0.004174603149294853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.004094195086508989 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_4/index_tensor([1408, 1408])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_4/index_tensor([1408, 1408])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7338e345843aa66ed4edaf12a6622c5527eb5984 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_4/index_tensor([1408, 1408])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An eye-level wide-angle shot slowly pans right, immersing the viewer in a classic 2D animated view of a colorful Carnival parade. The scene is defined by exaggerated features and bold black outlines, creating a vibrant, storybook feel. Hundreds of performers joyfully parade through the cobblestone streets, their movements energetic and synchronized. In the foreground, a brass band marches in unison, their stylized faces revealing drum-shaped cheekbones as they play shimmering trumpets and trombones. They wear vibrant scarlet jackets with gold epaulettes and tall, feathered hats. Just behind them, a troupe of dancers gracefully waves a massive silk banner in an arcing kaleidoscope of azure, fuchsia, and sunny yellow. The background is filled with a steady stream of marchers, some on stilts, and whimsical, slightly tilted buildings painted in pastel hues. The sidewalks are filled with simple, smiling people cheering. Confetti and streamers constantly and softly drifted through the air, catching the bright, flat light that saturated the entire frame, perfectly amplifying the extremely lively and joyful atmosphere of the celebration.", + "seed": 6375, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_4/index_1408_seed_6375_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029031199961900712, + "vlm_logprob": -0.029031199961900712, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.029031199961900712, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021762963384389877 + ], + "opposite_logprob_tokens": [ + -3.8967628479003906 + ], + "logprob_reward": -0.021762963384389877, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021762963384389877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18833249807357788, + 0.05119168385863304, + 0.08180961012840271, + 0.047570906579494476, + 0.09919555485248566, + 0.0736311748623848, + 0.05160261690616608, + 0.05341773480176926, + 0.2909482419490814, + 0.062300052493810654 + ], + "credit_map": [ + 0.017683997750282288, + 0.016005825251340866, + 0.030772225931286812, + 0.019510706886649132, + 0.032840240746736526, + 0.024129580706357956, + 0.03283751755952835, + 0.014552410691976547, + 0.00556010939180851, + 0.006027159746736288, + 0.004378725308924913, + 0.005963840987533331, + 0.007169219199568033, + 0.01106762234121561, + 0.005729387514293194, + 0.005295617040246725, + 0.007208236493170261, + 0.010770250111818314, + 0.008458592928946018, + 0.0074906125664711, + 0.009379860945045948, + 0.01828630641102791, + 0.00843154639005661, + 0.011784202419221401, + 0.006879729218780994, + 0.005680318456143141, + 0.00579072767868638, + 0.006077709142118692, + 0.005818340927362442, + 0.006231936160475016, + 0.006462493911385536, + 0.004629647359251976, + 0.014279995113611221, + 0.010851201601326466, + 0.012585727497935295, + 0.012907754629850388, + 0.01549395639449358, + 0.011949514038860798, + 0.010520080104470253, + 0.010607324540615082, + 0.008269506506621838, + 0.019527748227119446, + 0.005170629825443029, + 0.0046797278337180614, + 0.02107234299182892, + 0.0052965059876441956, + 0.004144259728491306, + 0.005470447242259979, + 0.00757812662050128, + 0.007032907567918301, + 0.004777824506163597, + 0.004876979626715183, + 0.007485802285373211, + 0.008026933297514915, + 0.006448597181588411, + 0.005375446751713753, + 0.006558883935213089, + 0.011925444938242435, + 0.007126420270651579, + 0.004482093267142773, + 0.008923560380935669, + 0.006206607446074486, + 0.004293219186365604, + 0.003901503747329116, + 0.02252906560897827, + 0.10101251304149628, + 0.01176968589425087, + 0.012630576267838478, + 0.09504076838493347, + 0.019824765622615814, + 0.017229367047548294, + 0.010911478661000729, + 0.007292024791240692, + 0.008676653727889061, + 0.006638774648308754, + 0.006462504155933857, + 0.008064907044172287, + 0.012409243732690811, + 0.006103424821048975, + 0.006652517709881067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.021762963384389877 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486559122800827 + ], + "opposite_logprob_tokens": [ + -3.0486559867858887 + ], + "logprob_reward": -0.0486559122800827, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486559122800827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04647093638777733, + 0.04498392716050148, + 0.039487507194280624, + 0.019949374720454216, + 0.032884176820516586, + 0.43036481738090515, + 0.04839712381362915, + 0.02954973839223385, + 0.2672049105167389, + 0.040707457810640335 + ], + "credit_map": [ + 0.004703112877905369, + 0.004342070315033197, + 0.008147788234055042, + 0.003721798537299037, + 0.00598194170743227, + 0.007758042775094509, + 0.0072320736944675446, + 0.004584105685353279, + 0.007438856642693281, + 0.005096843000501394, + 0.003586031263694167, + 0.0029876797925680876, + 0.006701883394271135, + 0.011228358373045921, + 0.0039666322991251945, + 0.003977648913860321, + 0.0049116178415715694, + 0.004154290538281202, + 0.006173803936690092, + 0.006949524860829115, + 0.0031894799321889877, + 0.004099508747458458, + 0.0051604085601866245, + 0.004848875571042299, + 0.0029169770423322916, + 0.0026943241246044636, + 0.0027373137418180704, + 0.0019556647166609764, + 0.0027889064513146877, + 0.0024658278562128544, + 0.0021157274022698402, + 0.0022746329195797443, + 0.0038294221740216017, + 0.00393719132989645, + 0.0034745479933917522, + 0.0038719491567462683, + 0.004080342594534159, + 0.003547091269865632, + 0.0065749953500926495, + 0.0035686383489519358, + 0.04641582816839218, + 0.1213098019361496, + 0.02144976146519184, + 0.011975212022662163, + 0.15444107353687286, + 0.027414405718445778, + 0.014435538090765476, + 0.03292318433523178, + 0.0048463125713169575, + 0.005375948268920183, + 0.003299932461231947, + 0.0030865485314279795, + 0.004268407355993986, + 0.010414675809442997, + 0.013676449656486511, + 0.0034288503229618073, + 0.0038932443130761385, + 0.005859973840415478, + 0.003937028348445892, + 0.003304037731140852, + 0.00390704395249486, + 0.003700660541653633, + 0.0028384507168084383, + 0.002109299413859844, + 0.021058499813079834, + 0.09873896092176437, + 0.007883910089731216, + 0.009081042371690273, + 0.09324265271425247, + 0.014651676639914513, + 0.012857754714787006, + 0.009690426290035248, + 0.0040903217159211636, + 0.004880980588495731, + 0.006136016920208931, + 0.0053993030451238155, + 0.004194776061922312, + 0.005606118589639664, + 0.005804830230772495, + 0.004595110658556223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.0486559122800827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24902456998825073, + 0.13879643380641937, + 0.03135611489415169, + 0.080046646296978, + 0.03588005527853966, + 0.23332074284553528, + 0.0935271754860878, + 0.044529881328344345, + 0.06010592356324196, + 0.03341248258948326 + ], + "credit_map": [ + 0.01976906508207321, + 0.01888119801878929, + 0.04152672737836838, + 0.025668244808912277, + 0.026935799047350883, + 0.05110994726419449, + 0.039043109863996506, + 0.02609044872224331, + 0.0152962077409029, + 0.014971060678362846, + 0.014272552914917469, + 0.010167660191655159, + 0.019167648628354073, + 0.03862890228629112, + 0.013456347398459911, + 0.012836053036153316, + 0.00409747613593936, + 0.0036212827544659376, + 0.003306526690721512, + 0.004300439730286598, + 0.003688892349600792, + 0.004710056819021702, + 0.004243699833750725, + 0.0033877387177199125, + 0.010857239365577698, + 0.00846095010638237, + 0.01621145009994507, + 0.011431523598730564, + 0.007811846677213907, + 0.004653457086533308, + 0.012857231311500072, + 0.007762942463159561, + 0.004827938973903656, + 0.004523075185716152, + 0.00386882689781487, + 0.004022280219942331, + 0.004453968722373247, + 0.004333523567765951, + 0.006297742016613483, + 0.003552699461579323, + 0.02579713799059391, + 0.06882305443286896, + 0.010942783206701279, + 0.006947242189198732, + 0.07937714457511902, + 0.015724405646324158, + 0.00809585489332676, + 0.017613109201192856, + 0.009336077608168125, + 0.008385486900806427, + 0.0053452965803444386, + 0.004298240412026644, + 0.009974155575037003, + 0.025996819138526917, + 0.02495749108493328, + 0.005233599804341793, + 0.00644321832805872, + 0.008693725802004337, + 0.0044671352952718735, + 0.0038899939972907305, + 0.008213313296437263, + 0.0046777077950537205, + 0.003955025225877762, + 0.004189755767583847, + 0.004568359814584255, + 0.014102296903729439, + 0.0036983294412493706, + 0.0043190778233110905, + 0.01367128361016512, + 0.010398992337286472, + 0.005618610884994268, + 0.0037289722822606564, + 0.003149976022541523, + 0.004698034375905991, + 0.003252288792282343, + 0.0042342147789895535, + 0.005684127099812031, + 0.007150873076170683, + 0.0023863245733082294, + 0.002856639912351966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004092651419341564 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008628231473267078 + ], + "opposite_logprob_tokens": [ + -4.7586283683776855 + ], + "logprob_reward": -0.008628231473267078, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008628231473267078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16060607135295868, + 0.11986575275659561, + 0.03752955049276352, + 0.06438958644866943, + 0.027984213083982468, + 0.19716137647628784, + 0.1880618929862976, + 0.05542798340320587, + 0.09975515305995941, + 0.049218371510505676 + ], + "credit_map": [ + 0.019538365304470062, + 0.018114790320396423, + 0.021479468792676926, + 0.01302851177752018, + 0.01813085377216339, + 0.03394754230976105, + 0.01928582973778248, + 0.017080720514059067, + 0.01414091233164072, + 0.013098412193357944, + 0.008240382187068462, + 0.007571655325591564, + 0.01676730252802372, + 0.03372647240757942, + 0.011882908642292023, + 0.014437710866332054, + 0.006302004214376211, + 0.003992205951362848, + 0.005700247827917337, + 0.005993390455842018, + 0.003348661120980978, + 0.004150138236582279, + 0.003909996245056391, + 0.004132908768951893, + 0.009854521602392197, + 0.007468207739293575, + 0.011211717501282692, + 0.007026085164397955, + 0.00986487977206707, + 0.007207378279417753, + 0.005704128183424473, + 0.0060526663437485695, + 0.0034066152293235064, + 0.003325043711811304, + 0.0038158860988914967, + 0.0023767659440636635, + 0.004588601645082235, + 0.0029228583443909883, + 0.0034716641530394554, + 0.00407677935436368, + 0.0205407552421093, + 0.04774439334869385, + 0.011676894500851631, + 0.011918424628674984, + 0.054755065590143204, + 0.023857686668634415, + 0.011611147783696651, + 0.015057018026709557, + 0.01732652634382248, + 0.014981216751039028, + 0.009031579829752445, + 0.007117766886949539, + 0.018919935449957848, + 0.055631741881370544, + 0.05406301096081734, + 0.010990102775394917, + 0.006145888939499855, + 0.010854396969079971, + 0.009068777784705162, + 0.007343394216150045, + 0.008017080835998058, + 0.0037999004125595093, + 0.005041796714067459, + 0.005156742874532938, + 0.00786818377673626, + 0.0359736829996109, + 0.0029914334882050753, + 0.003924367018043995, + 0.033570967614650726, + 0.006459241267293692, + 0.004751638043671846, + 0.004215633496642113, + 0.005566496402025223, + 0.006748995278030634, + 0.004927421920001507, + 0.005530519410967827, + 0.006308624986559153, + 0.01160142756998539, + 0.00366764422506094, + 0.004867239855229855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.008628231473267078 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201624125242233 + ], + "opposite_logprob_tokens": [ + -2.812016248703003 + ], + "logprob_reward": -0.06201624125242233, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201624125242233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21169407665729523, + 0.06898653507232666, + 0.100556381046772, + 0.04715020954608917, + 0.04993651062250137, + 0.10824168473482132, + 0.13314585387706757, + 0.057757142931222916, + 0.17943225800991058, + 0.043099336326122284 + ], + "credit_map": [ + 0.0255818422883749, + 0.028241412714123726, + 0.03404538333415985, + 0.02847251296043396, + 0.024695001542568207, + 0.02954094298183918, + 0.024211551994085312, + 0.016905421391129494, + 0.010861092247068882, + 0.008047860115766525, + 0.005640697665512562, + 0.006927465088665485, + 0.0075828079134225845, + 0.01169436052441597, + 0.00964419636875391, + 0.008588049560785294, + 0.01133547443896532, + 0.011810218915343285, + 0.013406996615231037, + 0.013708566315472126, + 0.008488018997013569, + 0.01494651474058628, + 0.012386588379740715, + 0.014473988674581051, + 0.007637777831405401, + 0.004689555149525404, + 0.004282549023628235, + 0.004287507385015488, + 0.00906164851039648, + 0.00817827694118023, + 0.004890076816082001, + 0.004122819751501083, + 0.006654188968241215, + 0.0057139890268445015, + 0.0053121838718652725, + 0.0043746172450482845, + 0.005320210009813309, + 0.005306404083967209, + 0.009550190530717373, + 0.007704723626375198, + 0.013549362309277058, + 0.021777167916297913, + 0.011135933920741081, + 0.008416247554123402, + 0.02443554624915123, + 0.01211623940616846, + 0.007346003782004118, + 0.009465179406106472, + 0.015423203818500042, + 0.013016841374337673, + 0.005997844506055117, + 0.005079357419162989, + 0.016734730452299118, + 0.03751692175865173, + 0.031062116846442223, + 0.008314833045005798, + 0.010116906836628914, + 0.010039865039288998, + 0.004286234732717276, + 0.005907940678298473, + 0.009372353553771973, + 0.008962600491940975, + 0.00443930784240365, + 0.0046319300308823586, + 0.01474698819220066, + 0.06229536980390549, + 0.006772859022021294, + 0.009648196399211884, + 0.0573992095887661, + 0.010125686414539814, + 0.010388500057160854, + 0.00805542804300785, + 0.00454968586564064, + 0.005305593833327293, + 0.004844612907618284, + 0.005020712967962027, + 0.0075639765709638596, + 0.00593430083245039, + 0.005632807034999132, + 0.004247648175805807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.06201624125242233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level wide-angle shot slowly pans right, immersing the viewer in a classic 2D animated view of a colorful Carnival parade. The scene is defined by exaggerated features and bold black outlines, creating a vibrant, storybook feel. Hundreds of performers joyfully parade through the cobblestone streets, their movements energetic and synchronized. In the foreground, a brass band marches in unison, their stylized faces revealing drum-shaped cheekbones as they play shimmering trumpets and trombones. They wear vibrant scarlet jackets with gold epaulettes and tall, feathered hats. Just behind them, a troupe of dancers gracefully waves a massive silk banner in an arcing kaleidoscope of azure, fuchsia, and sunny yellow. The background is filled with a steady stream of marchers, some on stilts, and whimsical, slightly tilted buildings painted in pastel hues. The sidewalks are filled with simple, smiling people cheering. Confetti and streamers constantly and softly drifted through the air, catching the bright, flat light that saturated the entire frame, perfectly amplifying the extremely lively and joyful atmosphere of the celebration.", + "seed": 6376, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_4/index_1408_seed_6376_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0717371828854084, + "vlm_logprob": -0.0717371828854084, + "vlm_yes_no_margin": 3.174999976158142, + "vlm_token_credit": -0.0717371828854084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12802763283252716 + ], + "opposite_logprob_tokens": [ + -2.1280276775360107 + ], + "logprob_reward": -0.12802763283252716, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12802763283252716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2356855869293213, + 0.02898278832435608, + 0.06960826367139816, + 0.06596767157316208, + 0.37194013595581055, + 0.06164955347776413, + 0.019720884039998055, + 0.009525001980364323, + 0.014473828487098217, + 0.12244625389575958 + ], + "credit_map": [ + 0.037384144961833954, + 0.042874500155448914, + 0.02259650081396103, + 0.016297107562422752, + 0.040191616863012314, + 0.03186165541410446, + 0.027746934443712234, + 0.01673312485218048, + 0.004828466102480888, + 0.0053410702385008335, + 0.0023869713768363, + 0.0031547285616397858, + 0.0037665129639208317, + 0.00357792223803699, + 0.0034883609041571617, + 0.0024387561716139317, + 0.012282983399927616, + 0.010520746931433678, + 0.005115070380270481, + 0.007234146352857351, + 0.006674516014754772, + 0.012918004766106606, + 0.0069712246768176556, + 0.007891567423939705, + 0.007289994973689318, + 0.008774351328611374, + 0.006757470779120922, + 0.00495636323466897, + 0.018858563154935837, + 0.010038147680461407, + 0.004581050481647253, + 0.004711735527962446, + 0.034886784851551056, + 0.14011889696121216, + 0.012207542546093464, + 0.010666781105101109, + 0.02415832132101059, + 0.04009009525179863, + 0.08202913403511047, + 0.02778257615864277, + 0.006335285492241383, + 0.01044521201401949, + 0.0032835977617651224, + 0.004429659806191921, + 0.020584022626280785, + 0.010834199376404285, + 0.0029493998736143112, + 0.002788175828754902, + 0.0030573264230042696, + 0.0027268656995147467, + 0.0015156245790421963, + 0.0016731804935261607, + 0.0031474751885980368, + 0.0030587073415517807, + 0.002168661681935191, + 0.002373043680563569, + 0.0014605401083827019, + 0.0012745041167363524, + 0.001037144218571484, + 0.0010180913377553225, + 0.0014089150354266167, + 0.001104825409129262, + 0.0012197927571833134, + 0.0010011887643486261, + 0.0021621347405016422, + 0.002173180691897869, + 0.0019042600179091096, + 0.001660676789470017, + 0.0020237972494214773, + 0.0018271692097187042, + 0.001271016662940383, + 0.00145159219391644, + 0.01393110677599907, + 0.04330265149474144, + 0.006340858060866594, + 0.028764722868800163, + 0.010772348381578922, + 0.00665833568200469, + 0.0074032265692949295, + 0.005273010581731796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.12802763283252716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789848193526268 + ], + "opposite_logprob_tokens": [ + -2.5789847373962402 + ], + "logprob_reward": -0.0789848193526268, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789848193526268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09331674128770828, + 0.09010957926511765, + 0.10239218175411224, + 0.09279440343379974, + 0.2865312397480011, + 0.20219686627388, + 0.030741950497031212, + 0.022359903901815414, + 0.04331258311867714, + 0.036244556307792664 + ], + "credit_map": [ + 0.014534925110638142, + 0.01662628725171089, + 0.00759999779984355, + 0.00666881212964654, + 0.015790367498993874, + 0.014652121812105179, + 0.00902140885591507, + 0.00842281524091959, + 0.02172253094613552, + 0.02793164551258087, + 0.004532331135123968, + 0.005114185158163309, + 0.008330905809998512, + 0.012803187593817711, + 0.004787923768162727, + 0.004886866081506014, + 0.01667325757443905, + 0.013524276204407215, + 0.01157629769295454, + 0.009229017421603203, + 0.0124605568125844, + 0.016808727756142616, + 0.011738233268260956, + 0.010381821542978287, + 0.01402231864631176, + 0.020819542929530144, + 0.0091329887509346, + 0.012494318187236786, + 0.010668058879673481, + 0.011321134865283966, + 0.008253621868789196, + 0.006082420237362385, + 0.030931415036320686, + 0.10124722123146057, + 0.010960561223328114, + 0.008716069161891937, + 0.019934039562940598, + 0.0280056931078434, + 0.06271730363368988, + 0.02401895262300968, + 0.020915526896715164, + 0.035410672426223755, + 0.015052715316414833, + 0.019011229276657104, + 0.046053994446992874, + 0.03623102977871895, + 0.014595948159694672, + 0.014925756491720676, + 0.0038930103182792664, + 0.004036364145576954, + 0.00267798057757318, + 0.0032783132046461105, + 0.00641726516187191, + 0.004512393847107887, + 0.0031328529585152864, + 0.002793770283460617, + 0.003044770797714591, + 0.0031348704360425472, + 0.0015745856799185276, + 0.0015558096347376704, + 0.003561268327757716, + 0.001988691743463278, + 0.003599992487579584, + 0.0038999151438474655, + 0.0065873758867383, + 0.008490696549415588, + 0.005781347397714853, + 0.004983681719750166, + 0.004816580563783646, + 0.004885341040790081, + 0.004019438289105892, + 0.0037481205072253942, + 0.004672295413911343, + 0.007900940254330635, + 0.0037603885866701603, + 0.006996831391006708, + 0.003103414550423622, + 0.003465048037469387, + 0.003234044648706913, + 0.003111592261120677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.0789848193526268 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093838855624199 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.004093838855624199, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093838855624199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23227015137672424, + 0.08062200993299484, + 0.08475720137357712, + 0.08150804042816162, + 0.34662067890167236, + 0.08299632370471954, + 0.022999439388513565, + 0.015267138369381428, + 0.015942147001624107, + 0.03701682388782501 + ], + "credit_map": [ + 0.027009818702936172, + 0.03361598402261734, + 0.041496992111206055, + 0.021942639723420143, + 0.03219279646873474, + 0.03293615207076073, + 0.028807025402784348, + 0.014268730767071247, + 0.01520044356584549, + 0.024204352870583534, + 0.007084586191922426, + 0.005566460080444813, + 0.008364702574908733, + 0.009431041777133942, + 0.006299828179180622, + 0.0044706035405397415, + 0.014833449386060238, + 0.010627023875713348, + 0.006828945130109787, + 0.007671397179365158, + 0.011065704748034477, + 0.01220550388097763, + 0.01070711761713028, + 0.010818058624863625, + 0.011484537273645401, + 0.016286874189972878, + 0.007542455103248358, + 0.010350131429731846, + 0.015130138024687767, + 0.009241863153874874, + 0.005968035664409399, + 0.00550400884822011, + 0.03298566862940788, + 0.13054907321929932, + 0.01190678495913744, + 0.009944534860551357, + 0.021035965532064438, + 0.04400567337870598, + 0.06978762894868851, + 0.02640533819794655, + 0.008164197206497192, + 0.018107827752828598, + 0.004042799584567547, + 0.004245379474014044, + 0.028618114069104195, + 0.011995340697467327, + 0.004924598149955273, + 0.0028980751521885395, + 0.003682715818285942, + 0.0032576287630945444, + 0.0023332389537245035, + 0.002377468394115567, + 0.0033337802160531282, + 0.002772852312773466, + 0.002786584896966815, + 0.002455169567838311, + 0.0024541427846997976, + 0.002267699921503663, + 0.0012478128774091601, + 0.0020368648692965508, + 0.002030241070315242, + 0.001540823606774211, + 0.0020600995048880577, + 0.0016294546658173203, + 0.0023737093433737755, + 0.003695383667945862, + 0.001491332077421248, + 0.0012898368295282125, + 0.0022062233183532953, + 0.001552464789710939, + 0.0019538600463420153, + 0.00137933692894876, + 0.0060240887105464935, + 0.003971436992287636, + 0.003137578023597598, + 0.006235992535948753, + 0.0046969084069132805, + 0.006449512671679258, + 0.0033119029831141233, + 0.0031894014682620764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.004093838855624199 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020565373823046684 + ], + "opposite_logprob_tokens": [ + -3.8955652713775635 + ], + "logprob_reward": -0.020565373823046684, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020565373823046684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12761381268501282, + 0.09067309647798538, + 0.18007996678352356, + 0.07433978468179703, + 0.185883566737175, + 0.15350459516048431, + 0.05756621062755585, + 0.051300134509801865, + 0.04633102938532829, + 0.03270780295133591 + ], + "credit_map": [ + 0.019504372030496597, + 0.02288683131337166, + 0.010709776543080807, + 0.01016459334641695, + 0.023877032101154327, + 0.01820332184433937, + 0.013835308142006397, + 0.008432566188275814, + 0.02602255344390869, + 0.02400779351592064, + 0.004165268503129482, + 0.004453601315617561, + 0.006963577587157488, + 0.012355237267911434, + 0.0056686922907829285, + 0.007036362774670124, + 0.03339681029319763, + 0.024518031626939774, + 0.013194069266319275, + 0.0160113163292408, + 0.018106691539287567, + 0.03084164671599865, + 0.024232160300016403, + 0.019779236987233162, + 0.009835372678935528, + 0.014274612069129944, + 0.005641056224703789, + 0.00559600256383419, + 0.01837046630680561, + 0.0128769027069211, + 0.0038633609656244516, + 0.003882007673382759, + 0.018882950767874718, + 0.06637126952409744, + 0.006899289321154356, + 0.00636697094887495, + 0.014136682264506817, + 0.019155869260430336, + 0.04007207229733467, + 0.013998456299304962, + 0.015052754431962967, + 0.03660731762647629, + 0.009508894756436348, + 0.010149631649255753, + 0.04455417022109032, + 0.022315559908747673, + 0.008826308883726597, + 0.006489950697869062, + 0.00942271389067173, + 0.005400431342422962, + 0.008486458100378513, + 0.0052879746071994305, + 0.010217051953077316, + 0.008217031136155128, + 0.005165151320397854, + 0.005369390361011028, + 0.007889491505920887, + 0.007847877219319344, + 0.0028934443835169077, + 0.003906591329723597, + 0.011487110517919064, + 0.004800072871148586, + 0.006261901929974556, + 0.006213644053786993, + 0.006156417075544596, + 0.006010137964040041, + 0.004177819937467575, + 0.004872687626630068, + 0.0068003516644239426, + 0.008203428238630295, + 0.004639487713575363, + 0.005470696371048689, + 0.004184096585959196, + 0.006581101566553116, + 0.002747277496382594, + 0.007135360036045313, + 0.0028782631270587444, + 0.002536849584430456, + 0.0030838639941066504, + 0.003560988698154688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.020565373823046684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701424956321716 + ], + "opposite_logprob_tokens": [ + -2.12701416015625 + ], + "logprob_reward": -0.12701424956321716, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12701424956321716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19856533408164978, + 0.12761268019676208, + 0.07874686270952225, + 0.03421063721179962, + 0.13544422388076782, + 0.15874148905277252, + 0.05523092672228813, + 0.05927620828151703, + 0.026276342570781708, + 0.12589535117149353 + ], + "credit_map": [ + 0.030927520245313644, + 0.035241175442934036, + 0.020537663251161575, + 0.019811993464827538, + 0.03768785670399666, + 0.02303188666701317, + 0.01828879676759243, + 0.013038440607488155, + 0.045223262161016464, + 0.046251580119132996, + 0.004625312983989716, + 0.0043422868475317955, + 0.00821241270750761, + 0.00828259252011776, + 0.004439718089997768, + 0.00623551569879055, + 0.015249480493366718, + 0.011463906615972519, + 0.004862093832343817, + 0.007639307528734207, + 0.007680325768887997, + 0.014255503192543983, + 0.009714078158140182, + 0.007882167585194111, + 0.005267261993139982, + 0.006680037826299667, + 0.0028611503075808287, + 0.0035529297310858965, + 0.005293970927596092, + 0.004739104304462671, + 0.0028907705564051867, + 0.0029254129622131586, + 0.010336448438465595, + 0.042807966470718384, + 0.007550498936325312, + 0.008467970415949821, + 0.008390571922063828, + 0.04356081411242485, + 0.006273185834288597, + 0.00805674772709608, + 0.015919795259833336, + 0.03145081177353859, + 0.0101036187261343, + 0.011695621535182, + 0.04968571290373802, + 0.023516658693552017, + 0.00788626167923212, + 0.008482999168336391, + 0.007419664412736893, + 0.0064667887054383755, + 0.008536542765796185, + 0.007289523258805275, + 0.007374962791800499, + 0.006772567983716726, + 0.006289463024586439, + 0.005081414710730314, + 0.007651832420378923, + 0.011490778997540474, + 0.006717545911669731, + 0.008937100879848003, + 0.0065499842166900635, + 0.005773397162556648, + 0.006968826055526733, + 0.005186744499951601, + 0.00413903035223484, + 0.005715397652238607, + 0.002406702609732747, + 0.001958646811544895, + 0.004716923460364342, + 0.003148052841424942, + 0.0025165947154164314, + 0.0016749940114095807, + 0.011053829453885555, + 0.029518190771341324, + 0.007117245811969042, + 0.033800020813941956, + 0.009938611648976803, + 0.009387936443090439, + 0.008651664480566978, + 0.01642785221338272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.12701424956321716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level wide-angle shot slowly pans right, immersing the viewer in a classic 2D animated view of a colorful Carnival parade. The scene is defined by exaggerated features and bold black outlines, creating a vibrant, storybook feel. Hundreds of performers joyfully parade through the cobblestone streets, their movements energetic and synchronized. In the foreground, a brass band marches in unison, their stylized faces revealing drum-shaped cheekbones as they play shimmering trumpets and trombones. They wear vibrant scarlet jackets with gold epaulettes and tall, feathered hats. Just behind them, a troupe of dancers gracefully waves a massive silk banner in an arcing kaleidoscope of azure, fuchsia, and sunny yellow. The background is filled with a steady stream of marchers, some on stilts, and whimsical, slightly tilted buildings painted in pastel hues. The sidewalks are filled with simple, smiling people cheering. Confetti and streamers constantly and softly drifted through the air, catching the bright, flat light that saturated the entire frame, perfectly amplifying the extremely lively and joyful atmosphere of the celebration.", + "seed": 6377, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_4/index_1408_seed_6377_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2511355279944837, + "vlm_logprob": -0.2511355279944837, + "vlm_yes_no_margin": 2.474999952316284, + "vlm_token_credit": -0.2511355279944837, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8996686935424805 + ], + "opposite_logprob_tokens": [ + -0.5246686935424805 + ], + "logprob_reward": -0.8996686935424805, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8996686935424805, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22210432589054108, + 0.03945541009306908, + 0.10088734328746796, + 0.05033174902200699, + 0.04625822231173515, + 0.15375301241874695, + 0.09651622176170349, + 0.06733331084251404, + 0.18205682933330536, + 0.0413036085665226 + ], + "credit_map": [ + 0.026309702545404434, + 0.027671489864587784, + 0.021275021135807037, + 0.027441851794719696, + 0.03335863724350929, + 0.036642033606767654, + 0.033800724893808365, + 0.015604868531227112, + 0.003973512910306454, + 0.005391791462898254, + 0.003812622046098113, + 0.00554042961448431, + 0.006047451868653297, + 0.006482941564172506, + 0.0043198601342737675, + 0.003886801889166236, + 0.007809825241565704, + 0.009585228748619556, + 0.01584808900952339, + 0.014653963968157768, + 0.009845064952969551, + 0.012543700635433197, + 0.01718270033597946, + 0.013418763875961304, + 0.0038408609107136726, + 0.005816878285259008, + 0.007919411174952984, + 0.010282442905008793, + 0.004688638728111982, + 0.00514381704851985, + 0.007107635494321585, + 0.005532066337764263, + 0.005644125863909721, + 0.0060061500407755375, + 0.005491459276527166, + 0.004737182520329952, + 0.006332110147923231, + 0.008533700369298458, + 0.004899810999631882, + 0.004613684490323067, + 0.014512361958622932, + 0.016169117763638496, + 0.03912508487701416, + 0.027293510735034943, + 0.018244367092847824, + 0.009104898199439049, + 0.01843479461967945, + 0.010868892073631287, + 0.011020596139132977, + 0.014623318798840046, + 0.022543076425790787, + 0.007786103989928961, + 0.007467338349670172, + 0.011183466762304306, + 0.01644809916615486, + 0.005444232374429703, + 0.008775320835411549, + 0.011716380715370178, + 0.006705016363412142, + 0.009209435433149338, + 0.007039083633571863, + 0.009976033121347427, + 0.005351651459932327, + 0.00856038462370634, + 0.01678401418030262, + 0.013513860292732716, + 0.018153484910726547, + 0.017623847350478172, + 0.009505496360361576, + 0.017437275499105453, + 0.051683150231838226, + 0.03735567629337311, + 0.0039243572391569614, + 0.0061004385352134705, + 0.0050372532568871975, + 0.003305623773485422, + 0.007719296962022781, + 0.007291420362889767, + 0.0042151217348873615, + 0.003710097400471568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.8996686935424805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521187961101532 + ], + "opposite_logprob_tokens": [ + -1.5021188259124756 + ], + "logprob_reward": -0.2521187961101532, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521187961101532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10723079741001129, + 0.04311446100473404, + 0.028743725270032883, + 0.059234488755464554, + 0.1815275400876999, + 0.3056778311729431, + 0.09649526327848434, + 0.053268156945705414, + 0.0824148952960968, + 0.04229281470179558 + ], + "credit_map": [ + 0.011801062151789665, + 0.01732075773179531, + 0.010053399950265884, + 0.01799696497619152, + 0.011849947273731232, + 0.010276745073497295, + 0.020086616277694702, + 0.007845298387110233, + 0.0046995775774121284, + 0.005690675228834152, + 0.0050267428159713745, + 0.006498286500573158, + 0.004434956703335047, + 0.005826735403388739, + 0.006184919737279415, + 0.0047525679692626, + 0.002714913571253419, + 0.005265038460493088, + 0.0034564160741865635, + 0.0038253911770880222, + 0.0031756535172462463, + 0.00398279307410121, + 0.003431897610425949, + 0.0028916210867464542, + 0.004961994010955095, + 0.005277268588542938, + 0.01014717947691679, + 0.01541224680840969, + 0.004235532134771347, + 0.004045670852065086, + 0.008552988059818745, + 0.006601604167371988, + 0.017727075144648552, + 0.06279410421848297, + 0.007906544022262096, + 0.015745680779218674, + 0.012781052850186825, + 0.010319448076188564, + 0.014422886073589325, + 0.03983074799180031, + 0.03510066494345665, + 0.034157589077949524, + 0.09119463711977005, + 0.04514387995004654, + 0.032337214797735214, + 0.015987753868103027, + 0.028032420203089714, + 0.02372368797659874, + 0.018072018399834633, + 0.01374068483710289, + 0.014430353417992592, + 0.008053968660533428, + 0.008318963460624218, + 0.008361898362636566, + 0.01847841590642929, + 0.0070389569737017155, + 0.005419346038252115, + 0.008670959621667862, + 0.006000064313411713, + 0.006652384996414185, + 0.00608664657920599, + 0.005399199202656746, + 0.00545735377818346, + 0.009582203812897205, + 0.00676058791577816, + 0.005430363584309816, + 0.009531321004033089, + 0.009406737051904202, + 0.004431592300534248, + 0.010457533411681652, + 0.022430000826716423, + 0.013966762460768223, + 0.0050763157196342945, + 0.0033731248695403337, + 0.004642200190573931, + 0.003198203630745411, + 0.011548950336873531, + 0.005647094454616308, + 0.004420834593474865, + 0.0043860916048288345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.2521187961101532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742583122104406 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.006742583122104406, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742583122104406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11636018753051758, + 0.06477704644203186, + 0.1123536229133606, + 0.088644839823246, + 0.14394481480121613, + 0.17654243111610413, + 0.06747130304574966, + 0.09508424997329712, + 0.08491366356611252, + 0.04990783706307411 + ], + "credit_map": [ + 0.01079694926738739, + 0.014173312112689018, + 0.011809494346380234, + 0.022079793736338615, + 0.015447353012859821, + 0.016171729192137718, + 0.015142567455768585, + 0.010739000514149666, + 0.007011765148490667, + 0.009713206440210342, + 0.006436832714825869, + 0.009163368493318558, + 0.011322933249175549, + 0.008384985849261284, + 0.006953325122594833, + 0.0057906354777514935, + 0.008466426283121109, + 0.010027576237916946, + 0.015636015683412552, + 0.014864712953567505, + 0.010844787582755089, + 0.017716247588396072, + 0.02006659470498562, + 0.014731267467141151, + 0.00785789918154478, + 0.006939757149666548, + 0.014925148338079453, + 0.019898470491170883, + 0.008013206534087658, + 0.009388772770762444, + 0.0121891675516963, + 0.009432417340576649, + 0.016454510390758514, + 0.04964609816670418, + 0.009662335738539696, + 0.012375828810036182, + 0.011645621620118618, + 0.008642701432108879, + 0.012683714739978313, + 0.02283400669693947, + 0.022023912519216537, + 0.021596359089016914, + 0.04932120814919472, + 0.027970243245363235, + 0.014699887484312057, + 0.009216522797942162, + 0.01814720407128334, + 0.01356709748506546, + 0.011753853410482407, + 0.012564944103360176, + 0.0114142382517457, + 0.005416033323854208, + 0.006238187663257122, + 0.007036072202026844, + 0.007880999706685543, + 0.005166978109627962, + 0.010762174613773823, + 0.009071833454072475, + 0.011421467177569866, + 0.013513543643057346, + 0.006893094629049301, + 0.008966012857854366, + 0.012080001644790173, + 0.022376123815774918, + 0.007674404885619879, + 0.005807443056255579, + 0.005289133638143539, + 0.005474021192640066, + 0.003961734473705292, + 0.006799312308430672, + 0.02788291871547699, + 0.02202470228075981, + 0.005607678554952145, + 0.005079751834273338, + 0.00877187680453062, + 0.004593133460730314, + 0.00671758409589529, + 0.009004691615700722, + 0.005519174505025148, + 0.004613954108208418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.006742583122104406 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176300451159477 + ], + "opposite_logprob_tokens": [ + -4.018176078796387 + ], + "logprob_reward": -0.018176300451159477, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018176300451159477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08310249447822571, + 0.060818105936050415, + 0.06567352265119553, + 0.03532460704445839, + 0.05810161679983139, + 0.06296887248754501, + 0.21457326412200928, + 0.3103298246860504, + 0.04919024929404259, + 0.05991743505001068 + ], + "credit_map": [ + 0.009200097993016243, + 0.009295549243688583, + 0.009813101030886173, + 0.01075222808867693, + 0.011660522781312466, + 0.01444596890360117, + 0.010153084993362427, + 0.007781942840665579, + 0.0069236354902386665, + 0.008513527922332287, + 0.006454687565565109, + 0.007770002353936434, + 0.007652733474969864, + 0.009309021756052971, + 0.0065948073752224445, + 0.007599684409797192, + 0.00517577538266778, + 0.006442984100431204, + 0.010515180416405201, + 0.009513402357697487, + 0.007412535138428211, + 0.007348569110035896, + 0.010584655217826366, + 0.008680423721671104, + 0.003463459201157093, + 0.004422795493155718, + 0.004172493703663349, + 0.00492932740598917, + 0.005167806521058083, + 0.004900339990854263, + 0.004388338886201382, + 0.0038800444453954697, + 0.006049320567399263, + 0.015394932590425014, + 0.005633032415062189, + 0.006150974426418543, + 0.006410283502191305, + 0.005193249322474003, + 0.00618364242836833, + 0.007086181081831455, + 0.00649420078843832, + 0.007405227515846491, + 0.014052493497729301, + 0.009580590762197971, + 0.005916480906307697, + 0.0050361971370875835, + 0.0071389633230865, + 0.0073447139002382755, + 0.028164688497781754, + 0.05392569303512573, + 0.023800130933523178, + 0.012735465541481972, + 0.02875703014433384, + 0.03423765301704407, + 0.02301550842821598, + 0.0099371038377285, + 0.06004450097680092, + 0.05410155653953552, + 0.04337812587618828, + 0.02644694410264492, + 0.03759158402681351, + 0.02933657169342041, + 0.028184479102492332, + 0.03124605305492878, + 0.005429546348750591, + 0.0028507662937045097, + 0.003387591103091836, + 0.00378232728689909, + 0.002605590270832181, + 0.003699455177411437, + 0.015189177356660366, + 0.01224579382687807, + 0.005691117141395807, + 0.006489411927759647, + 0.007924044504761696, + 0.005505220498889685, + 0.011479160748422146, + 0.008290819823741913, + 0.00756689440459013, + 0.006970766466110945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.018176300451159477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.078971266746521 + ], + "opposite_logprob_tokens": [ + -2.5789713859558105 + ], + "logprob_reward": -0.078971266746521, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.078971266746521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2562675476074219, + 0.04933592677116394, + 0.07910545915365219, + 0.03769848123192787, + 0.1435074508190155, + 0.0858478918671608, + 0.1628183275461197, + 0.07705040276050568, + 0.0422302670776844, + 0.06613824516534805 + ], + "credit_map": [ + 0.030656244605779648, + 0.054887838661670685, + 0.01837177760899067, + 0.036450475454330444, + 0.03649260476231575, + 0.03026801347732544, + 0.029869336634874344, + 0.01927129179239273, + 0.004282688722014427, + 0.0056674061343073845, + 0.006217150948941708, + 0.006677036173641682, + 0.00503834942355752, + 0.004995746538043022, + 0.008855028077960014, + 0.007602521684020758, + 0.006835668347775936, + 0.011639775708317757, + 0.011550327762961388, + 0.010477762669324875, + 0.010789275169372559, + 0.009355749003589153, + 0.011227504350244999, + 0.0072293998673558235, + 0.004170962609350681, + 0.004172132350504398, + 0.006033079698681831, + 0.004816974978893995, + 0.0034380448050796986, + 0.0056658401153981686, + 0.005151758901774883, + 0.004249689169228077, + 0.013949144631624222, + 0.04644302651286125, + 0.007036349270492792, + 0.013743475079536438, + 0.009957737289369106, + 0.007508377078920603, + 0.011824925430119038, + 0.03304440528154373, + 0.010059423744678497, + 0.009796333499252796, + 0.023091839626431465, + 0.01145582552999258, + 0.009639238938689232, + 0.006826051510870457, + 0.007783982437103987, + 0.007195192389190197, + 0.018799643963575363, + 0.04107116907835007, + 0.019672486931085587, + 0.0107202660292387, + 0.01626732386648655, + 0.016393527388572693, + 0.03142458572983742, + 0.008469327352941036, + 0.0195733904838562, + 0.00829655583947897, + 0.015674596652388573, + 0.006690258625894785, + 0.005823187064379454, + 0.006777559407055378, + 0.007670465856790543, + 0.006544390227645636, + 0.003417825326323509, + 0.004955930169671774, + 0.00407252786681056, + 0.005931917577981949, + 0.004082208964973688, + 0.004340863786637783, + 0.00812530517578125, + 0.007303687743842602, + 0.005897599272429943, + 0.0072937291115522385, + 0.006261732429265976, + 0.004817905370146036, + 0.01201272290199995, + 0.012756332755088806, + 0.00818758923560381, + 0.00891063641756773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.078971266746521 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level wide-angle shot slowly pans right, immersing the viewer in a classic 2D animated view of a colorful Carnival parade. The scene is defined by exaggerated features and bold black outlines, creating a vibrant, storybook feel. Hundreds of performers joyfully parade through the cobblestone streets, their movements energetic and synchronized. In the foreground, a brass band marches in unison, their stylized faces revealing drum-shaped cheekbones as they play shimmering trumpets and trombones. They wear vibrant scarlet jackets with gold epaulettes and tall, feathered hats. Just behind them, a troupe of dancers gracefully waves a massive silk banner in an arcing kaleidoscope of azure, fuchsia, and sunny yellow. The background is filled with a steady stream of marchers, some on stilts, and whimsical, slightly tilted buildings painted in pastel hues. The sidewalks are filled with simple, smiling people cheering. Confetti and streamers constantly and softly drifted through the air, catching the bright, flat light that saturated the entire frame, perfectly amplifying the extremely lively and joyful atmosphere of the celebration.", + "seed": 6378, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_4/index_1408_seed_6378_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.040544355288147926, + "vlm_logprob": -0.040544355288147926, + "vlm_yes_no_margin": 4.150000023841858, + "vlm_token_credit": -0.040544355288147926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007128993049263954 + ], + "opposite_logprob_tokens": [ + -5.132129192352295 + ], + "logprob_reward": -0.007128993049263954, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007128993049263954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16384972631931305, + 0.19573545455932617, + 0.21492400765419006, + 0.04692406952381134, + 0.0504097156226635, + 0.05225914344191551, + 0.10117533057928085, + 0.05994560569524765, + 0.07698994129896164, + 0.037787050008773804 + ], + "credit_map": [ + 0.024032559245824814, + 0.02128830924630165, + 0.019802935421466827, + 0.014741035178303719, + 0.029970137402415276, + 0.021152973175048828, + 0.018156111240386963, + 0.0147056570276618, + 0.01973389834165573, + 0.019084317609667778, + 0.028504449874162674, + 0.020256374031305313, + 0.02594919502735138, + 0.037460897117853165, + 0.01627902127802372, + 0.028467310592532158, + 0.037390898913145065, + 0.02180631272494793, + 0.022037893533706665, + 0.02819979377090931, + 0.03081497550010681, + 0.030119962990283966, + 0.028200963512063026, + 0.01635321043431759, + 0.004525833297520876, + 0.005008486099541187, + 0.0050721545703709126, + 0.008956553414463997, + 0.004770733881741762, + 0.005862607154995203, + 0.007604265585541725, + 0.005123433656990528, + 0.006391157861799002, + 0.008307123556733131, + 0.0054554701782763, + 0.005845340900123119, + 0.0072019994258880615, + 0.00531429098919034, + 0.007235500495880842, + 0.004658828489482403, + 0.008085325360298157, + 0.006391461007297039, + 0.005764068569988012, + 0.005498227663338184, + 0.01243594754487276, + 0.005178077612072229, + 0.004944569896906614, + 0.003961461596190929, + 0.01558713335543871, + 0.012746364809572697, + 0.009610090404748917, + 0.01482867170125246, + 0.009948397986590862, + 0.011585716158151627, + 0.01471597608178854, + 0.012152977287769318, + 0.005298555362969637, + 0.008017055690288544, + 0.004687534179538488, + 0.00693237641826272, + 0.013547230511903763, + 0.009645109064877033, + 0.006287164054811001, + 0.005530582740902901, + 0.007016167975962162, + 0.011250520125031471, + 0.007250953000038862, + 0.007653353735804558, + 0.010804237797856331, + 0.013701150193810463, + 0.008906903676688671, + 0.01040665153414011, + 0.004712222144007683, + 0.004552335944026709, + 0.004206411074846983, + 0.0033545379992574453, + 0.005606402643024921, + 0.005147031042724848, + 0.005877988412976265, + 0.004330119118094444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.007128993049263954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427871286869049 + ], + "opposite_logprob_tokens": [ + -2.017787218093872 + ], + "logprob_reward": -0.1427871286869049, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1427871286869049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06938523054122925, + 0.107939213514328, + 0.09849786013364792, + 0.06549476832151413, + 0.047412168234586716, + 0.06078631430864334, + 0.15851768851280212, + 0.21752579510211945, + 0.13935276865959167, + 0.0350881926715374 + ], + "credit_map": [ + 0.008209817111492157, + 0.00902139488607645, + 0.008039038628339767, + 0.010920904576778412, + 0.009906970895826817, + 0.008267306722700596, + 0.007127611432224512, + 0.007892189547419548, + 0.00960620492696762, + 0.012796364724636078, + 0.0175265371799469, + 0.007749923970550299, + 0.015258989296853542, + 0.0220020841807127, + 0.011327267624437809, + 0.01167184580117464, + 0.014137299731373787, + 0.012667556293308735, + 0.011235601268708706, + 0.012979308143258095, + 0.01139272004365921, + 0.014100444503128529, + 0.012381288222968578, + 0.00960364006459713, + 0.008243299089372158, + 0.0075858477503061295, + 0.008291126228868961, + 0.01031852513551712, + 0.009401440620422363, + 0.006042295601218939, + 0.008701393380761147, + 0.006910834927111864, + 0.005225107539445162, + 0.005531130824238062, + 0.004375194199383259, + 0.006981879938393831, + 0.007287796586751938, + 0.007740018889307976, + 0.006435339339077473, + 0.003835699986666441, + 0.008747745305299759, + 0.0060347807593643665, + 0.009138333611190319, + 0.006387104280292988, + 0.005217218771576881, + 0.01170398760586977, + 0.006307897623628378, + 0.007249244023114443, + 0.024971654638648033, + 0.019908227026462555, + 0.016379153355956078, + 0.024773091077804565, + 0.023050136864185333, + 0.01721685752272606, + 0.014765347354114056, + 0.01745321787893772, + 0.027288205921649933, + 0.032453231513500214, + 0.01610971987247467, + 0.03487078845500946, + 0.0360313318669796, + 0.036256417632102966, + 0.019266147166490555, + 0.015249941498041153, + 0.010840276256203651, + 0.022879069671034813, + 0.014921776950359344, + 0.01638677343726158, + 0.017637817189097404, + 0.021605804562568665, + 0.017646657302975655, + 0.017434604465961456, + 0.004938633646816015, + 0.0036556401755660772, + 0.00400724308565259, + 0.004102627746760845, + 0.004902841057628393, + 0.005226678680628538, + 0.0038894005119800568, + 0.004365124739706516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.1427871286869049 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021968772634863853 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021968772634863853, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021968772634863853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12296142429113388, + 0.15523818135261536, + 0.36838454008102417, + 0.03942617401480675, + 0.04305322468280792, + 0.04699597507715225, + 0.054117653518915176, + 0.10081019997596741, + 0.03970586881041527, + 0.02930677868425846 + ], + "credit_map": [ + 0.01825050078332424, + 0.013582762330770493, + 0.020554417744278908, + 0.012146511115133762, + 0.014777049422264099, + 0.015293806791305542, + 0.01375376433134079, + 0.014602601528167725, + 0.01878662034869194, + 0.022112494334578514, + 0.027603045105934143, + 0.010673647746443748, + 0.027699660509824753, + 0.02525424398481846, + 0.011421934701502323, + 0.011686529032886028, + 0.06198274716734886, + 0.04049687087535858, + 0.03602205961942673, + 0.0392175130546093, + 0.05463925376534462, + 0.05140047147870064, + 0.04174245148897171, + 0.04288313910365105, + 0.004832510370761156, + 0.004422876518219709, + 0.004351590294390917, + 0.0045920307748019695, + 0.006623406428843737, + 0.005867690779268742, + 0.00462335767224431, + 0.004112706985324621, + 0.005233340431004763, + 0.008237963542342186, + 0.003485936438664794, + 0.004931997042149305, + 0.006347074639052153, + 0.004729547072201967, + 0.005434828344732523, + 0.004652534145861864, + 0.009466017596423626, + 0.0038179310504347086, + 0.005235031247138977, + 0.005092733073979616, + 0.00972110778093338, + 0.004163520876318216, + 0.004629554692655802, + 0.004870078060775995, + 0.007856416516005993, + 0.0064624520018696785, + 0.005385132506489754, + 0.009281580336391926, + 0.006514968816190958, + 0.00632116012275219, + 0.005556293297559023, + 0.006739642005413771, + 0.012269336730241776, + 0.014140944927930832, + 0.008121456950902939, + 0.013623560778796673, + 0.016578832641243935, + 0.017534790560603142, + 0.010498845018446445, + 0.00804242305457592, + 0.0036216923035681248, + 0.006170468404889107, + 0.003359974594786763, + 0.0040252250619232655, + 0.0051923212595283985, + 0.006530010141432285, + 0.005325895734131336, + 0.005480277817696333, + 0.0034102878998965025, + 0.003747638314962387, + 0.002969569992274046, + 0.004226967692375183, + 0.0039608958177268505, + 0.003609269857406616, + 0.003024270525202155, + 0.004357874859124422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0021968772634863853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012525935657322407 + ], + "opposite_logprob_tokens": [ + -4.387526035308838 + ], + "logprob_reward": -0.012525935657322407, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012525935657322407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11156821995973587, + 0.3184428811073303, + 0.2390696406364441, + 0.05656537413597107, + 0.041745997965335846, + 0.03733031079173088, + 0.06275051087141037, + 0.04681870713829994, + 0.06151270866394043, + 0.024195663630962372 + ], + "credit_map": [ + 0.01391464564949274, + 0.012511143460869789, + 0.011149723082780838, + 0.01371617242693901, + 0.014405151829123497, + 0.010552311316132545, + 0.020334424450993538, + 0.014984652400016785, + 0.03469343110918999, + 0.03692216798663139, + 0.05249321833252907, + 0.02447635866701603, + 0.0455065555870533, + 0.057152993977069855, + 0.026566142216324806, + 0.040632009506225586, + 0.03760270029306412, + 0.023082509636878967, + 0.02474214695394039, + 0.03546542674303055, + 0.035767391324043274, + 0.02819107100367546, + 0.033726535737514496, + 0.02049185335636139, + 0.00822609756141901, + 0.006830255035310984, + 0.007126895245164633, + 0.006435195915400982, + 0.007441198453307152, + 0.005972458980977535, + 0.007648387458175421, + 0.006884883623570204, + 0.005782576743513346, + 0.004085279069840908, + 0.00392313115298748, + 0.004804084077477455, + 0.0056755393743515015, + 0.0039062409196048975, + 0.007218376733362675, + 0.006350764073431492, + 0.006717348471283913, + 0.0036637522280216217, + 0.004451176151633263, + 0.0036877081729471684, + 0.00380541174672544, + 0.004568159580230713, + 0.003425959497690201, + 0.007010794244706631, + 0.01123239379376173, + 0.008210225030779839, + 0.008495146408677101, + 0.008498407900333405, + 0.0071829501539468765, + 0.005764744244515896, + 0.006044560577720404, + 0.0073220799677073956, + 0.0050572738982737064, + 0.0068294634111225605, + 0.003662980394437909, + 0.006848780903965235, + 0.009725965559482574, + 0.0064542158506810665, + 0.004754328168928623, + 0.003485695691779256, + 0.005550977773964405, + 0.008472451008856297, + 0.007010948844254017, + 0.008444770239293575, + 0.00715619046241045, + 0.007729272823780775, + 0.008723023347556591, + 0.008425076492130756, + 0.002289845608174801, + 0.0035424993839114904, + 0.0023881159722805023, + 0.002722367411479354, + 0.0037795326206833124, + 0.003267988096922636, + 0.002821494359523058, + 0.003383820876479149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.012525935657322407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a classic 2D animation with bold black outlines and exaggerated features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow pan to the right from an eye-level wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass band in the foreground wearing scarlet jackets with gold epaulettes and tall feathered hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a troupe of dancers visible behind the band waving a massive silk banner with azure, fuchsia, and sunny yellow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are confetti and streamers visible drifting through the air in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808284178376198 + ], + "opposite_logprob_tokens": [ + -3.2880828380584717 + ], + "logprob_reward": -0.03808284178376198, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808284178376198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08218591660261154, + 0.08035872131586075, + 0.05036008730530739, + 0.054753661155700684, + 0.043679334223270416, + 0.042426686733961105, + 0.20207305252552032, + 0.3435298204421997, + 0.06032156944274902, + 0.04031115770339966 + ], + "credit_map": [ + 0.012150587514042854, + 0.009391169995069504, + 0.008886466734111309, + 0.013865623623132706, + 0.011037047952413559, + 0.008916526101529598, + 0.009312407113611698, + 0.008626090362668037, + 0.009029513224959373, + 0.009222586639225483, + 0.012879189103841782, + 0.0059060691855847836, + 0.011417667381465435, + 0.011175494641065598, + 0.010882718488574028, + 0.009845483116805553, + 0.006328741554170847, + 0.005870533175766468, + 0.00509386183694005, + 0.006117009092122316, + 0.007831134833395481, + 0.00789911113679409, + 0.006888351868838072, + 0.004331343341618776, + 0.005670758429914713, + 0.007846188731491566, + 0.006371382623910904, + 0.006248150020837784, + 0.009524431079626083, + 0.00578425545245409, + 0.006187152583152056, + 0.007121344096958637, + 0.004870205651968718, + 0.00535857118666172, + 0.004782655276358128, + 0.006857226602733135, + 0.005473129451274872, + 0.005320029333233833, + 0.006398751866072416, + 0.0046187653206288815, + 0.005406143143773079, + 0.005929380189627409, + 0.004251435864716768, + 0.005640966817736626, + 0.006522501353174448, + 0.003848104504868388, + 0.006410557776689529, + 0.004417595453560352, + 0.03127921745181084, + 0.02948947809636593, + 0.023655755445361137, + 0.03136143088340759, + 0.02210676483809948, + 0.020039817318320274, + 0.019922472536563873, + 0.024218104779720306, + 0.04432285577058792, + 0.04518831521272659, + 0.029285266995429993, + 0.04867644980549812, + 0.051750872284173965, + 0.06166084110736847, + 0.036488957703113556, + 0.026156241074204445, + 0.007149853743612766, + 0.007956400513648987, + 0.006446375977247953, + 0.008243096061050892, + 0.007229550275951624, + 0.006442384794354439, + 0.008783118799328804, + 0.008070786483585835, + 0.0032208196353167295, + 0.00431925430893898, + 0.0040025729686021805, + 0.006905750837177038, + 0.00536202872171998, + 0.005049786064773798, + 0.0049980636686086655, + 0.006452878937125206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.03808284178376198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_5/index_tensor([1949, 1949])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_5/index_tensor([1949, 1949])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..905702a5af7ff65e1b575f0dc07e261e1dcedff1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_5/index_tensor([1949, 1949])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a slightly upward angle, slowly and steadily advances through a tranquil bamboo forest. Bright, silvery moonlight filters through the emerald green bamboo leaves that tower into the night sky, casting dappled shadows on the leaf-covered ground. A thin mist lingers in the forest. At the center of the frame, Hua Mulan, clad in a deep cyan silk Hanfu (Chinese traditional costume), is immersed in her sword dance. Her long, jet-black hair is tied high in a neat bun, strands brushing against her cheeks in the breeze. Her expression is focused and serene, her eyes piercing, and in her hand, she clutches a longsword, gleaming with a cold gleam. Her movements are fluid, her body gracefully twirling, dipping, and leaping in time with the sword's thrusts, her wide sleeves drawing graceful arcs through the air. As she swung her sword, the highly polished blade captured and reflected the cool moonlight, leaving behind trails of light in the air as smooth and shining as liquid mercury, perfectly interpreting the artistic conception of \"the sword shines like water\". The whole scene, under the cool light and shadow, created a cinematic atmosphere that combines power, beauty and loneliness.", + "seed": 11996, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_5/index_1949_seed_11996_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.042122987750917676, + "vlm_logprob": -0.042122987750917676, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.042122987750917676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08407042175531387 + ], + "opposite_logprob_tokens": [ + -2.5840704441070557 + ], + "logprob_reward": -0.08407042175531387, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08407042175531387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5245465040206909, + 0.03531299903988838, + 0.08845540136098862, + 0.07944314926862717, + 0.0768030509352684, + 0.049079425632953644, + 0.03201723471283913, + 0.061901070177555084, + 0.03518051281571388, + 0.017260633409023285 + ], + "credit_map": [ + 0.07216810435056686, + 0.04254373535513878, + 0.08412359654903412, + 0.07384798675775528, + 0.04170385003089905, + 0.03678533062338829, + 0.0710752084851265, + 0.10229872912168503, + 0.005676582455635071, + 0.0045837522484362125, + 0.005177564453333616, + 0.0030978573486208916, + 0.003582063829526305, + 0.004577748477458954, + 0.0026923916302621365, + 0.005925036501139402, + 0.020721793174743652, + 0.013473793864250183, + 0.013345914892852306, + 0.00477081723511219, + 0.00770188681781292, + 0.017172962427139282, + 0.004342116415500641, + 0.006926118861883879, + 0.015159262344241142, + 0.015516352839767933, + 0.009222934953868389, + 0.005048561375588179, + 0.005163997877389193, + 0.005691823549568653, + 0.01058484148234129, + 0.013055375777184963, + 0.014642414636909962, + 0.009626145474612713, + 0.014207805506885052, + 0.004720563068985939, + 0.007216738071292639, + 0.010163907893002033, + 0.004963301587849855, + 0.011262170039117336, + 0.01751776784658432, + 0.0027139338199049234, + 0.00855346117168665, + 0.006122257094830275, + 0.0051373811438679695, + 0.002510374179109931, + 0.004105918109416962, + 0.0024183313362300396, + 0.003813195275142789, + 0.0025614979676902294, + 0.005319670774042606, + 0.004699064418673515, + 0.003304742043837905, + 0.003714531660079956, + 0.003934724722057581, + 0.0046698045916855335, + 0.007589724846184254, + 0.005185449030250311, + 0.010969686321914196, + 0.005096986889839172, + 0.0052021099254488945, + 0.014894425868988037, + 0.005803957115858793, + 0.00715872785076499, + 0.004548920784145594, + 0.003081707516685128, + 0.0033516876865178347, + 0.003109050216153264, + 0.003169118659570813, + 0.010120971128344536, + 0.0041567557491362095, + 0.003642302006483078, + 0.002380677731707692, + 0.0019195934291929007, + 0.0021502282470464706, + 0.0014879773370921612, + 0.0029617694672197104, + 0.003246381413191557, + 0.0017149914056062698, + 0.0013990142615512013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.08407042175531387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004655831027776003 + ], + "opposite_logprob_tokens": [ + -5.379655838012695 + ], + "logprob_reward": -0.004655831027776003, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004655831027776003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11643096804618835, + 0.08211679756641388, + 0.09577701985836029, + 0.38243982195854187, + 0.021439116448163986, + 0.05820465460419655, + 0.08936764299869537, + 0.045255765318870544, + 0.06771980226039886, + 0.04124836623668671 + ], + "credit_map": [ + 0.019463367760181427, + 0.010927380993962288, + 0.016437936574220657, + 0.008976857177913189, + 0.010562026873230934, + 0.016468295827507973, + 0.007843412458896637, + 0.025751691311597824, + 0.009346447885036469, + 0.013581193052232265, + 0.00950393546372652, + 0.010679326951503754, + 0.011978814378380775, + 0.008095413446426392, + 0.009977820329368114, + 0.008953848853707314, + 0.01541910134255886, + 0.013801981694996357, + 0.028190473094582558, + 0.00720655033364892, + 0.00492108752951026, + 0.004870900418609381, + 0.007172028534114361, + 0.014194890856742859, + 0.05796971172094345, + 0.03776725381612778, + 0.05389241501688957, + 0.018051084131002426, + 0.023100728169083595, + 0.02101317048072815, + 0.0633927583694458, + 0.10725268721580505, + 0.004548768047243357, + 0.002748806495219469, + 0.003042852971702814, + 0.0015901344595476985, + 0.0024044548626989126, + 0.0026749419048428535, + 0.0014915931969881058, + 0.0029375641606748104, + 0.018146131187677383, + 0.004785242490470409, + 0.010891212150454521, + 0.005100867711007595, + 0.006153102498501539, + 0.0034324354492127895, + 0.005333538632839918, + 0.004362122621387243, + 0.010585852898657322, + 0.006728341802954674, + 0.01869206130504608, + 0.013034979812800884, + 0.008022375404834747, + 0.008562181144952774, + 0.010901669971644878, + 0.012840178795158863, + 0.004971832502633333, + 0.0046358294785022736, + 0.005183262284845114, + 0.004831318277865648, + 0.005115366540849209, + 0.007332331966608763, + 0.004684319254010916, + 0.00850150641053915, + 0.005345611367374659, + 0.0045878272503614426, + 0.017460815608501434, + 0.007131077814847231, + 0.006080635357648134, + 0.0043680486269295216, + 0.013228983618319035, + 0.009516808204352856, + 0.0046259332448244095, + 0.004667746368795633, + 0.005532387178391218, + 0.005781462881714106, + 0.006301704328507185, + 0.005125235300511122, + 0.005906048696488142, + 0.0033078475389629602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.004655831027776003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192220479249954 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192220479249954, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192220479249954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2494165152311325, + 0.05869892239570618, + 0.156377911567688, + 0.07743361592292786, + 0.03263380378484726, + 0.20413674414157867, + 0.061597537249326706, + 0.058808643370866776, + 0.03786909952759743, + 0.06302712112665176 + ], + "credit_map": [ + 0.029995499178767204, + 0.01813403144478798, + 0.03083622455596924, + 0.03594794496893883, + 0.02014165371656418, + 0.025799619033932686, + 0.033092692494392395, + 0.055468879640102386, + 0.006698431447148323, + 0.009594269096851349, + 0.007504554465413094, + 0.007957946509122849, + 0.0076774610206484795, + 0.008210261352360249, + 0.0054076313972473145, + 0.005648369435220957, + 0.025373801589012146, + 0.022207820788025856, + 0.047533486038446426, + 0.011527993716299534, + 0.006373421289026737, + 0.006993963848799467, + 0.012326478958129883, + 0.024040957912802696, + 0.009172216057777405, + 0.009683754295110703, + 0.009854460135102272, + 0.005323917139321566, + 0.005639044567942619, + 0.005216619931161404, + 0.011639674194157124, + 0.020903928205370903, + 0.0069306837394833565, + 0.0034478900488466024, + 0.004289641045033932, + 0.002753927605226636, + 0.004348789807409048, + 0.005341304000467062, + 0.0028497448656708, + 0.0026718261651694775, + 0.030183615162968636, + 0.01869034953415394, + 0.05389487370848656, + 0.03877251222729683, + 0.015114793553948402, + 0.009808829054236412, + 0.02173824980854988, + 0.015933526679873466, + 0.007664582692086697, + 0.005841535050421953, + 0.012251248583197594, + 0.00851164385676384, + 0.006569640710949898, + 0.005735233426094055, + 0.006965039297938347, + 0.008058618754148483, + 0.00556472223252058, + 0.005401020869612694, + 0.0073587303049862385, + 0.006683841813355684, + 0.005653998348861933, + 0.008776106871664524, + 0.006402210332453251, + 0.012968016788363457, + 0.002803727285936475, + 0.0020417661871761084, + 0.009404980577528477, + 0.00422592693939805, + 0.0024472081568092108, + 0.0029824927914887667, + 0.00872858241200447, + 0.005234417971223593, + 0.00816685613244772, + 0.008011111058294773, + 0.007851370610296726, + 0.00612497515976429, + 0.011264132335782051, + 0.01159560028463602, + 0.00519459368661046, + 0.0048184930346906185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.018192220479249954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370121121406555 + ], + "opposite_logprob_tokens": [ + -3.408701181411743 + ], + "logprob_reward": -0.03370121121406555, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370121121406555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5978299975395203, + 0.036717936396598816, + 0.08408614993095398, + 0.07827132940292358, + 0.021550284698605537, + 0.05794350802898407, + 0.024627437815070152, + 0.03233018517494202, + 0.02198152244091034, + 0.04466173052787781 + ], + "credit_map": [ + 0.08573922514915466, + 0.04813969507813454, + 0.0937078446149826, + 0.08957482874393463, + 0.04890364781022072, + 0.038492657244205475, + 0.08356035500764847, + 0.10971172899007797, + 0.0036673445720225573, + 0.005082526244223118, + 0.003719191299751401, + 0.004955194424837828, + 0.0055032460950315, + 0.00427085068076849, + 0.00539798429235816, + 0.004121600650250912, + 0.013135691173374653, + 0.010058827698230743, + 0.02284937910735607, + 0.005120865534991026, + 0.004433430731296539, + 0.008240530267357826, + 0.006165825296193361, + 0.014081602916121483, + 0.01522187888622284, + 0.007847405038774014, + 0.006398070603609085, + 0.0037473824340850115, + 0.004924374166876078, + 0.004727344494313002, + 0.01253710500895977, + 0.02286776714026928, + 0.003271654713898897, + 0.0031050234101712704, + 0.003169945441186428, + 0.0014846913982182741, + 0.002653616014868021, + 0.0030289909336715937, + 0.0018672988517209888, + 0.0029690635856240988, + 0.0200971607118845, + 0.0034179063513875008, + 0.017046283930540085, + 0.003216474549844861, + 0.005746910348534584, + 0.0027269127313047647, + 0.002642679261043668, + 0.0030491764191538095, + 0.0026717132423073053, + 0.0015235565369948745, + 0.003507533110678196, + 0.0037372882943600416, + 0.0025196908973157406, + 0.0035954350605607033, + 0.0036727574188262224, + 0.003399465000256896, + 0.00350041757337749, + 0.002186516532674432, + 0.0038177981041371822, + 0.0032106032595038414, + 0.00273167435079813, + 0.007914572022855282, + 0.0033477870747447014, + 0.005620815325528383, + 0.0024262685328722, + 0.0018143245251849294, + 0.002606641734018922, + 0.0022070910781621933, + 0.0019281514687463641, + 0.005502906162291765, + 0.0023326361551880836, + 0.003163502784445882, + 0.0065688928589224815, + 0.004915379919111729, + 0.009433169849216938, + 0.004701841622591019, + 0.005928496364504099, + 0.00705826235935092, + 0.0033241899218410254, + 0.002731498098000884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03370121121406555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999525427818298 + ], + "opposite_logprob_tokens": [ + -2.694995164871216 + ], + "logprob_reward": -0.06999525427818298, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999525427818298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08377482742071152, + 0.03264869749546051, + 0.07938094437122345, + 0.4513911306858063, + 0.03247543424367905, + 0.22348886728286743, + 0.012647397816181183, + 0.024218903854489326, + 0.012824569828808308, + 0.04714914411306381 + ], + "credit_map": [ + 0.011706417426466942, + 0.00947565957903862, + 0.012403145432472229, + 0.009262697771191597, + 0.0068306392058730125, + 0.008607195690274239, + 0.006241553463041782, + 0.019247526302933693, + 0.003244316205382347, + 0.004184060730040073, + 0.004051913972944021, + 0.004841065965592861, + 0.0040081157349050045, + 0.003630124032497406, + 0.004566905554383993, + 0.004122197162359953, + 0.013046937063336372, + 0.012611917220056057, + 0.02154986746609211, + 0.006033430341631174, + 0.004897534381598234, + 0.0053585851565003395, + 0.006470101419836283, + 0.009412580169737339, + 0.06479410082101822, + 0.04280836880207062, + 0.06489033252000809, + 0.023835008963942528, + 0.02709057927131653, + 0.023853838443756104, + 0.07626307755708694, + 0.12785589694976807, + 0.006338478531688452, + 0.002925361040979624, + 0.004642616491764784, + 0.0023735244758427143, + 0.0035586596932262182, + 0.004578145686537027, + 0.0031208000145852566, + 0.004937852267175913, + 0.07836062461137772, + 0.013230202719569206, + 0.037732142955064774, + 0.02900911308825016, + 0.022886313498020172, + 0.010904176160693169, + 0.018831178545951843, + 0.01253515388816595, + 0.0016135353362187743, + 0.0010873171268031001, + 0.0014663890469819307, + 0.0017815210158005357, + 0.0013414416462182999, + 0.0021180009935051203, + 0.0019487376557663083, + 0.0012904559262096882, + 0.0025033357087522745, + 0.0015803234418854117, + 0.003381222253665328, + 0.0031661842949688435, + 0.0028146144468337297, + 0.003231843700632453, + 0.00283705978654325, + 0.004704323597252369, + 0.0011296544689685106, + 0.0008387654670514166, + 0.002290163654834032, + 0.001871827756986022, + 0.0013839061139151454, + 0.001831413945183158, + 0.0020085743162781, + 0.0014702652115374804, + 0.007641036994755268, + 0.005656092427670956, + 0.005026108585298061, + 0.003990359604358673, + 0.008495123125612736, + 0.009780500084161758, + 0.003524889936670661, + 0.003035039408132434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.06999525427818298 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, slowly and steadily advances through a tranquil bamboo forest. Bright, silvery moonlight filters through the emerald green bamboo leaves that tower into the night sky, casting dappled shadows on the leaf-covered ground. A thin mist lingers in the forest. At the center of the frame, Hua Mulan, clad in a deep cyan silk Hanfu (Chinese traditional costume), is immersed in her sword dance. Her long, jet-black hair is tied high in a neat bun, strands brushing against her cheeks in the breeze. Her expression is focused and serene, her eyes piercing, and in her hand, she clutches a longsword, gleaming with a cold gleam. Her movements are fluid, her body gracefully twirling, dipping, and leaping in time with the sword's thrusts, her wide sleeves drawing graceful arcs through the air. As she swung her sword, the highly polished blade captured and reflected the cool moonlight, leaving behind trails of light in the air as smooth and shining as liquid mercury, perfectly interpreting the artistic conception of \"the sword shines like water\". The whole scene, under the cool light and shadow, created a cinematic atmosphere that combines power, beauty and loneliness.", + "seed": 11997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_5/index_1949_seed_11997_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016562467627227306, + "vlm_logprob": -0.016562467627227306, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.016562467627227306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006291819736361504 + ], + "opposite_logprob_tokens": [ + -6.13129186630249 + ], + "logprob_reward": -0.006291819736361504, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.006291819736361504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17985230684280396, + 0.07206717133522034, + 0.14430870115756989, + 0.07983257621526718, + 0.058004677295684814, + 0.06250995397567749, + 0.1325811892747879, + 0.18231530487537384, + 0.03253530338406563, + 0.055992789566516876 + ], + "credit_map": [ + 0.0268416628241539, + 0.015232766047120094, + 0.030007213354110718, + 0.013218218460679054, + 0.02073048986494541, + 0.03758314251899719, + 0.014519920572638512, + 0.021718891337513924, + 0.007478427607566118, + 0.009664593264460564, + 0.01203170046210289, + 0.009137069806456566, + 0.007429560646414757, + 0.010102340951561928, + 0.006283998489379883, + 0.009939482435584068, + 0.009403232485055923, + 0.010727071203291416, + 0.014634115621447563, + 0.0447932668030262, + 0.007337386254221201, + 0.040236108005046844, + 0.006104761268943548, + 0.01107275951653719, + 0.010236594825983047, + 0.007730536628514528, + 0.00969002302736044, + 0.011498852632939816, + 0.008929331786930561, + 0.010613884776830673, + 0.010851114988327026, + 0.010282240808010101, + 0.007096798159182072, + 0.011196942068636417, + 0.009316754527390003, + 0.005119237117469311, + 0.006149826571345329, + 0.006836591754108667, + 0.0056547014974057674, + 0.006633821874856949, + 0.007889612577855587, + 0.007989446632564068, + 0.00769165251404047, + 0.006231067236512899, + 0.0104057053104043, + 0.007692474406212568, + 0.008345182985067368, + 0.0062648155726492405, + 0.02405964955687523, + 0.01818503625690937, + 0.0183700080960989, + 0.018920471891760826, + 0.013641875237226486, + 0.017332592979073524, + 0.011217976920306683, + 0.010853582993149757, + 0.023068614304065704, + 0.03283626586198807, + 0.018925487995147705, + 0.023227287456393242, + 0.01596377231180668, + 0.01942186988890171, + 0.02838217467069626, + 0.020489845424890518, + 0.005954866297543049, + 0.004011684563010931, + 0.003926113713532686, + 0.0038099833764135838, + 0.004065196495503187, + 0.0033971865195780993, + 0.004343461245298386, + 0.0030268128030002117, + 0.00794980302453041, + 0.006188172847032547, + 0.0063500297255814075, + 0.007522194180637598, + 0.006596475839614868, + 0.006549053825438023, + 0.00909603014588356, + 0.005741029046475887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.006291819736361504 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748029962182045 + ], + "opposite_logprob_tokens": [ + -5.006748199462891 + ], + "logprob_reward": -0.006748029962182045, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748029962182045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17776073515415192, + 0.09137997776269913, + 0.19929422438144684, + 0.14590631425380707, + 0.07214345037937164, + 0.04774640500545502, + 0.07527657598257065, + 0.0975886806845665, + 0.05991464853286743, + 0.03298893943428993 + ], + "credit_map": [ + 0.020504416897892952, + 0.025425611063838005, + 0.02229607291519642, + 0.017410025000572205, + 0.020407164469361305, + 0.02313665673136711, + 0.017912933602929115, + 0.030667856335639954, + 0.00795015599578619, + 0.008767898194491863, + 0.013007143512368202, + 0.0166657492518425, + 0.009656133130192757, + 0.010865502059459686, + 0.012967748567461967, + 0.011499646119773388, + 0.009820856153964996, + 0.011296521872282028, + 0.014974336139857769, + 0.06491897255182266, + 0.009826702065765858, + 0.06818269938230515, + 0.009365303441882133, + 0.010908837430179119, + 0.01146450825035572, + 0.01313246600329876, + 0.017278878018260002, + 0.027769986540079117, + 0.01946445368230343, + 0.017558041960000992, + 0.021142814308404922, + 0.018095169216394424, + 0.007466306444257498, + 0.010318403132259846, + 0.008407262153923512, + 0.009387683123350143, + 0.010082462802529335, + 0.007263197097927332, + 0.008962463587522507, + 0.010255668312311172, + 0.0043874154798686504, + 0.0036593188997358084, + 0.0064462935552001, + 0.004634595941752195, + 0.007884287275373936, + 0.006515949033200741, + 0.009098466485738754, + 0.005120080430060625, + 0.008450211025774479, + 0.005939289927482605, + 0.009802533313632011, + 0.012196846306324005, + 0.009622504003345966, + 0.01108860969543457, + 0.008287752978503704, + 0.009888825006783009, + 0.014077317900955677, + 0.011554010212421417, + 0.011576378718018532, + 0.012478080578148365, + 0.00889903586357832, + 0.011326663196086884, + 0.01622680202126503, + 0.011450394056737423, + 0.011855426244437695, + 0.005545914638787508, + 0.00904659554362297, + 0.0073720053769648075, + 0.005870994180440903, + 0.005428165663033724, + 0.008133048191666603, + 0.006662498693913221, + 0.0039040904957801104, + 0.005364000331610441, + 0.0035400069318711758, + 0.004539406392723322, + 0.0029262597672641277, + 0.004122912883758545, + 0.005314233712852001, + 0.0032780293840914965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.006748029962182045 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194712042808533 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194712042808533, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194712042808533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15314312279224396, + 0.09960509091615677, + 0.150529682636261, + 0.09803563356399536, + 0.08213230222463608, + 0.12194863706827164, + 0.061626266688108444, + 0.11153141409158707, + 0.06028112769126892, + 0.061166733503341675 + ], + "credit_map": [ + 0.020624982193112373, + 0.018263056874275208, + 0.018201952800154686, + 0.01569373905658722, + 0.024989275261759758, + 0.020329713821411133, + 0.014693397097289562, + 0.02034701220691204, + 0.008470471017062664, + 0.009370771236717701, + 0.012999179773032665, + 0.018121881410479546, + 0.012583030387759209, + 0.013147619552910328, + 0.01237470656633377, + 0.012537424452602863, + 0.006971141789108515, + 0.008558795787394047, + 0.011369374580681324, + 0.04721728339791298, + 0.007756830658763647, + 0.05149352550506592, + 0.007387761492282152, + 0.009774960577487946, + 0.010139509104192257, + 0.008589389733970165, + 0.011381895281374454, + 0.016982978209853172, + 0.011978558264672756, + 0.01402212306857109, + 0.012958914041519165, + 0.0119822658598423, + 0.011749238707125187, + 0.011801376938819885, + 0.010089629329741001, + 0.008151394315063953, + 0.011594587005674839, + 0.008555008098483086, + 0.009919312782585621, + 0.010271756909787655, + 0.01048979815095663, + 0.011104151606559753, + 0.01857583038508892, + 0.013686837628483772, + 0.017796551808714867, + 0.015077347867190838, + 0.021256275475025177, + 0.013961846008896828, + 0.009923643432557583, + 0.008467298932373524, + 0.00650376919656992, + 0.006064502988010645, + 0.010491711087524891, + 0.00726618617773056, + 0.006233776453882456, + 0.00667537609115243, + 0.010502773337066174, + 0.019913669675588608, + 0.011128025129437447, + 0.012797858566045761, + 0.019268156960606575, + 0.01422165334224701, + 0.012825502082705498, + 0.01087377779185772, + 0.012441054917871952, + 0.005562924314290285, + 0.009574669413268566, + 0.006130505818873644, + 0.0061820861883461475, + 0.0056779454462230206, + 0.007616357412189245, + 0.0070955827832221985, + 0.007528447080403566, + 0.005841165781021118, + 0.008996374905109406, + 0.006620645988732576, + 0.007565486244857311, + 0.008888913318514824, + 0.007206370588392019, + 0.00851933192461729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.014194712042808533 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14224740862846375, + 0.10904131829738617, + 0.3572601079940796, + 0.04019550234079361, + 0.09230361878871918, + 0.07159809023141861, + 0.05663545802235603, + 0.03533586487174034, + 0.06478831171989441, + 0.03059428744018078 + ], + "credit_map": [ + 0.01989099755883217, + 0.026920489966869354, + 0.023016521707177162, + 0.012831224128603935, + 0.01436963677406311, + 0.02518216148018837, + 0.008917911909520626, + 0.01111847348511219, + 0.014109892770648003, + 0.014317470602691174, + 0.02566376142203808, + 0.010946650058031082, + 0.008007434196770191, + 0.012651050463318825, + 0.0063838837668299675, + 0.01696118153631687, + 0.01725418120622635, + 0.014923763461411, + 0.029931696131825447, + 0.12864723801612854, + 0.010688104666769505, + 0.12414799630641937, + 0.012250437401235104, + 0.019416693598031998, + 0.004440236836671829, + 0.006799482274800539, + 0.006522579118609428, + 0.00488473754376173, + 0.005344710778445005, + 0.005265220068395138, + 0.0041796364821493626, + 0.0027588973753154278, + 0.01679842732846737, + 0.019398387521505356, + 0.012727008201181889, + 0.007964435033500195, + 0.00857726950198412, + 0.010045137256383896, + 0.008011303842067719, + 0.008781655691564083, + 0.007620770949870348, + 0.01581534743309021, + 0.012106848880648613, + 0.008398877456784248, + 0.008899232372641563, + 0.00823905598372221, + 0.005765341687947512, + 0.004752620123326778, + 0.013019018806517124, + 0.00624253461137414, + 0.006904212757945061, + 0.0056625232100486755, + 0.005650542210787535, + 0.008659160695970058, + 0.004964552354067564, + 0.0055329161696136, + 0.005165994167327881, + 0.005030537024140358, + 0.004571713972836733, + 0.00330149382352829, + 0.006665588356554508, + 0.003941938281059265, + 0.003723767353221774, + 0.0029348337557166815, + 0.01618160307407379, + 0.008320050314068794, + 0.007835017517209053, + 0.007350556552410126, + 0.005982222966849804, + 0.00638550566509366, + 0.006503408774733543, + 0.006229945458471775, + 0.004929074086248875, + 0.0048955874517560005, + 0.0031751864589750767, + 0.003718534717336297, + 0.002697848482057452, + 0.003775406163185835, + 0.0034674208145588636, + 0.003935231361538172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.012536413036286831 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304136335849762 + ], + "opposite_logprob_tokens": [ + -3.168041467666626 + ], + "logprob_reward": -0.04304136335849762, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304136335849762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10325551778078079, + 0.08465298265218735, + 0.26846200227737427, + 0.04882800206542015, + 0.03589124232530594, + 0.1700020283460617, + 0.07393322139978409, + 0.10583513975143433, + 0.04246615245938301, + 0.06667374819517136 + ], + "credit_map": [ + 0.01311162393540144, + 0.009104309603571892, + 0.01800994761288166, + 0.011454680003225803, + 0.012004140764474869, + 0.01383743155747652, + 0.011669750325381756, + 0.01406362745910883, + 0.011406379751861095, + 0.013554418459534645, + 0.013594772666692734, + 0.009983059950172901, + 0.006521244067698717, + 0.01153686735779047, + 0.00719592347741127, + 0.010860309936106205, + 0.01284028310328722, + 0.015050558373332024, + 0.01963471621274948, + 0.09322977811098099, + 0.009136809036135674, + 0.09562017768621445, + 0.009268145076930523, + 0.013681525364518166, + 0.0052488106302917, + 0.005282638128846884, + 0.008142326027154922, + 0.0050039696507155895, + 0.0061270869337022305, + 0.007434883154928684, + 0.00612696772441268, + 0.005461321212351322, + 0.0033797649666666985, + 0.006942784879356623, + 0.005168288014829159, + 0.0040842508897185326, + 0.0034169950522482395, + 0.005110135767608881, + 0.003937887027859688, + 0.0038511347956955433, + 0.022622032091021538, + 0.022684462368488312, + 0.028038877993822098, + 0.01941709592938423, + 0.017784425988793373, + 0.03017541579902172, + 0.018901633098721504, + 0.010378086939454079, + 0.0075523206032812595, + 0.007989472709596157, + 0.009437330067157745, + 0.012551763094961643, + 0.007719991263002157, + 0.012517457827925682, + 0.00653254147619009, + 0.009632347151637077, + 0.013149709440767765, + 0.015741700306534767, + 0.012212310917675495, + 0.013044722378253937, + 0.010516954585909843, + 0.012231139466166496, + 0.01711491495370865, + 0.011823688633739948, + 0.008437841199338436, + 0.0046433876268565655, + 0.004387603607028723, + 0.0057451557368040085, + 0.00568187003955245, + 0.004164363257586956, + 0.004525057505816221, + 0.004880874417722225, + 0.006581953726708889, + 0.007832104340195656, + 0.00679588271304965, + 0.009008560329675674, + 0.006098144222050905, + 0.011218633502721786, + 0.007881798781454563, + 0.011256668716669083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.04304136335849762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, slowly and steadily advances through a tranquil bamboo forest. Bright, silvery moonlight filters through the emerald green bamboo leaves that tower into the night sky, casting dappled shadows on the leaf-covered ground. A thin mist lingers in the forest. At the center of the frame, Hua Mulan, clad in a deep cyan silk Hanfu (Chinese traditional costume), is immersed in her sword dance. Her long, jet-black hair is tied high in a neat bun, strands brushing against her cheeks in the breeze. Her expression is focused and serene, her eyes piercing, and in her hand, she clutches a longsword, gleaming with a cold gleam. Her movements are fluid, her body gracefully twirling, dipping, and leaping in time with the sword's thrusts, her wide sleeves drawing graceful arcs through the air. As she swung her sword, the highly polished blade captured and reflected the cool moonlight, leaving behind trails of light in the air as smooth and shining as liquid mercury, perfectly interpreting the artistic conception of \"the sword shines like water\". The whole scene, under the cool light and shadow, created a cinematic atmosphere that combines power, beauty and loneliness.", + "seed": 11998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_5/index_1949_seed_11998_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011750793643295765, + "vlm_logprob": -0.011750793643295765, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.011750793643295765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00658781174570322 + ], + "opposite_logprob_tokens": [ + -5.506587982177734 + ], + "logprob_reward": -0.00658781174570322, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00658781174570322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16444244980812073, + 0.07188726961612701, + 0.0538431778550148, + 0.0383465513586998, + 0.030938420444726944, + 0.030529404059052467, + 0.026079650968313217, + 0.048852238804101944, + 0.5063546895980835, + 0.028726188465952873 + ], + "credit_map": [ + 0.017540553584694862, + 0.0287728663533926, + 0.020539483055472374, + 0.016392095014452934, + 0.019672762602567673, + 0.02378799393773079, + 0.016720838844776154, + 0.021015845239162445, + 0.011107644997537136, + 0.009041119366884232, + 0.007048080675303936, + 0.005112168379127979, + 0.014466565102338791, + 0.010659699328243732, + 0.008553777821362019, + 0.005898218601942062, + 0.009274303913116455, + 0.00796021893620491, + 0.008093307726085186, + 0.005772901698946953, + 0.006485452875494957, + 0.0073943594470620155, + 0.004724693018943071, + 0.004137942101806402, + 0.006434943992644548, + 0.004521727096289396, + 0.005466746166348457, + 0.00672557670623064, + 0.00372214219532907, + 0.004024431575089693, + 0.004464916419237852, + 0.0029860655777156353, + 0.0051437364891171455, + 0.0031071044504642487, + 0.00599296297878027, + 0.0029036225751042366, + 0.0037821605801582336, + 0.004014262929558754, + 0.002854173770174384, + 0.00314039527438581, + 0.0024222100619226694, + 0.0030922021251171827, + 0.005430806893855333, + 0.004165594931691885, + 0.003753278637304902, + 0.004383428022265434, + 0.0037117027677595615, + 0.0035701808519661427, + 0.002144149737432599, + 0.0023402634542435408, + 0.0036369061563163996, + 0.0034559492487460375, + 0.0031208847649395466, + 0.003400459187105298, + 0.004209096077829599, + 0.003771941876038909, + 0.004368553403764963, + 0.00415344862267375, + 0.00802334863692522, + 0.007861567661166191, + 0.0033893377985805273, + 0.004111055284738541, + 0.005714785773307085, + 0.01123014185577631, + 0.02917654626071453, + 0.05013294890522957, + 0.1921669989824295, + 0.028604891151189804, + 0.0399637334048748, + 0.10945092141628265, + 0.016719425097107887, + 0.04013922065496445, + 0.002486973302438855, + 0.0029270753730088472, + 0.003611674066632986, + 0.0031339041888713837, + 0.004431459121406078, + 0.004248112440109253, + 0.003381757065653801, + 0.0045052324421703815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.00658781174570322 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553129345178604 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553129345178604, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553129345178604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0876760482788086, + 0.1177389919757843, + 0.025439821183681488, + 0.03528539836406708, + 0.03563862666487694, + 0.04557475075125694, + 0.09665676951408386, + 0.12698431313037872, + 0.3956228792667389, + 0.033382393419742584 + ], + "credit_map": [ + 0.0075438241474330425, + 0.010608125478029251, + 0.008848264813423157, + 0.013931086286902428, + 0.01107508223503828, + 0.01078132726252079, + 0.013296009041368961, + 0.011592330411076546, + 0.01190350204706192, + 0.01564146764576435, + 0.013084035366773605, + 0.008868987672030926, + 0.02400803193449974, + 0.016354940831661224, + 0.0185848455876112, + 0.009293187409639359, + 0.0035070532467216253, + 0.0032266625203192234, + 0.0029786156956106424, + 0.0028136789333075285, + 0.0035043773241341114, + 0.003052372485399246, + 0.0041553243063390255, + 0.002201739465817809, + 0.005032659508287907, + 0.0035110614262521267, + 0.0053481548093259335, + 0.005284757353365421, + 0.0033471547067165375, + 0.0037452823016792536, + 0.006149526219815016, + 0.00286680250428617, + 0.00307075004093349, + 0.0036559063009917736, + 0.004691405687481165, + 0.00471875537186861, + 0.005839311517775059, + 0.0052326698787510395, + 0.005254692398011684, + 0.003175140358507633, + 0.0038134329952299595, + 0.00428589666262269, + 0.005515960045158863, + 0.005385930649936199, + 0.008426211774349213, + 0.0062315091490745544, + 0.005771236959844828, + 0.006144579499959946, + 0.006935213692486286, + 0.012959692627191544, + 0.020712045952677727, + 0.00980907678604126, + 0.009630601853132248, + 0.013047521933913231, + 0.008842719718813896, + 0.014719896949827671, + 0.006830740720033646, + 0.008681552484631538, + 0.02691859006881714, + 0.012041918933391571, + 0.01133811753243208, + 0.012548944912850857, + 0.01873677410185337, + 0.02988768368959427, + 0.025421053171157837, + 0.027181897312402725, + 0.06342634558677673, + 0.016736580058932304, + 0.034956641495227814, + 0.12686018645763397, + 0.022590450942516327, + 0.07844974845647812, + 0.002720876829698682, + 0.0033621571492403746, + 0.003911126870661974, + 0.002550232456997037, + 0.004676640033721924, + 0.00495480839163065, + 0.00410707201808691, + 0.007099483627825975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.012553129345178604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642295375466347 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642295375466347, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642295375466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14968296885490417, + 0.04581533744931221, + 0.02852998487651348, + 0.05300407484173775, + 0.07180318236351013, + 0.06150960922241211, + 0.16160021722316742, + 0.06388742476701736, + 0.31031039357185364, + 0.05385680869221687 + ], + "credit_map": [ + 0.012376644648611546, + 0.02021658979356289, + 0.016606437042355537, + 0.019812820479273796, + 0.014780925586819649, + 0.01927453652024269, + 0.022639812901616096, + 0.023975199088454247, + 0.005154449958354235, + 0.007735388353466988, + 0.004622325301170349, + 0.005158029962331057, + 0.0064615653827786446, + 0.005587186198681593, + 0.005502528976649046, + 0.005593862850219011, + 0.0036025166045874357, + 0.003844580613076687, + 0.003294948721304536, + 0.003210595343261957, + 0.002924265805631876, + 0.003479540813714266, + 0.005002469290047884, + 0.003171068150550127, + 0.005025939084589481, + 0.004076383076608181, + 0.00915367528796196, + 0.0059493910521268845, + 0.005055367946624756, + 0.013579330407083035, + 0.006349390372633934, + 0.0038146008737385273, + 0.013051793910562992, + 0.005485204979777336, + 0.012207641266286373, + 0.007655407302081585, + 0.0074951257556676865, + 0.00875391997396946, + 0.0068118334747850895, + 0.010342254303395748, + 0.0038007230032235384, + 0.007468437310308218, + 0.006826456170529127, + 0.009229057468473911, + 0.01537239458411932, + 0.008607800118625164, + 0.0054010022431612015, + 0.0048037366941571236, + 0.013089731335639954, + 0.016020145267248154, + 0.028035003691911697, + 0.019388798624277115, + 0.012072768062353134, + 0.026762433350086212, + 0.01781328022480011, + 0.02841806411743164, + 0.008388696238398552, + 0.007217691745609045, + 0.008904436603188515, + 0.007173795718699694, + 0.005379536654800177, + 0.005371317733079195, + 0.008574028499424458, + 0.012877927161753178, + 0.01808982342481613, + 0.03405139222741127, + 0.08589726686477661, + 0.015902576968073845, + 0.02427798882126808, + 0.08597032725811005, + 0.015844076871871948, + 0.030276929959654808, + 0.005490252282470465, + 0.006552891805768013, + 0.007752290926873684, + 0.004580681677907705, + 0.00847910437732935, + 0.008836831897497177, + 0.005676760338246822, + 0.006487997714430094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.008642295375466347 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636727273464203 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636727273464203, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636727273464203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3917144536972046, + 0.07556811720132828, + 0.0163101926445961, + 0.025982428342103958, + 0.02751503512263298, + 0.04730837047100067, + 0.06828610599040985, + 0.04668601229786873, + 0.2704840898513794, + 0.03014528378844261 + ], + "credit_map": [ + 0.02790493331849575, + 0.04683654010295868, + 0.043988727033138275, + 0.07148232311010361, + 0.028522687032818794, + 0.047742582857608795, + 0.07544075697660446, + 0.04979589954018593, + 0.011375383473932743, + 0.009236741811037064, + 0.0068134102039039135, + 0.005225362256169319, + 0.016262630000710487, + 0.011152802035212517, + 0.008961518295109272, + 0.00654026260599494, + 0.0023167107719928026, + 0.0024642320349812508, + 0.002193344756960869, + 0.0016832840628921986, + 0.0017681680619716644, + 0.00187415792606771, + 0.002340547274798155, + 0.0016697457758709788, + 0.0016197366639971733, + 0.0017013902543112636, + 0.00396569911390543, + 0.004869635216891766, + 0.002476329915225506, + 0.004244742915034294, + 0.0038693242240697145, + 0.00323556549847126, + 0.004439104348421097, + 0.0027517732232809067, + 0.004776886664330959, + 0.0028655906207859516, + 0.0027114995755255222, + 0.004169271793216467, + 0.00236579985357821, + 0.0034351074136793613, + 0.006462541408836842, + 0.005550737958401442, + 0.00831772293895483, + 0.006770642939954996, + 0.004761622287333012, + 0.006286604329943657, + 0.004290119279175997, + 0.004868373274803162, + 0.005358426831662655, + 0.0073666051030159, + 0.014073796570301056, + 0.006036406382918358, + 0.004388386849313974, + 0.010042324662208557, + 0.00902571901679039, + 0.011994435451924801, + 0.008847946301102638, + 0.007861465215682983, + 0.007030976936221123, + 0.006082376465201378, + 0.004286716226488352, + 0.003357430687174201, + 0.004704860504716635, + 0.004514235071837902, + 0.01650097593665123, + 0.023335738107562065, + 0.09934743493795395, + 0.014496062882244587, + 0.022773168981075287, + 0.07229484617710114, + 0.009122241288423538, + 0.012613586150109768, + 0.002855573082342744, + 0.0031696914229542017, + 0.004614363424479961, + 0.0029022882226854563, + 0.005641755647957325, + 0.005008304957300425, + 0.0029705697670578957, + 0.002982734004035592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.004636727273464203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633400447666645 + ], + "opposite_logprob_tokens": [ + -3.651334047317505 + ], + "logprob_reward": -0.02633400447666645, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633400447666645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0897328108549118, + 0.05011551082134247, + 0.04461250454187393, + 0.049611493945121765, + 0.08438057452440262, + 0.09590168297290802, + 0.06292638927698135, + 0.11987543851137161, + 0.36060234904289246, + 0.04224121943116188 + ], + "credit_map": [ + 0.006876808125525713, + 0.015688316896557808, + 0.01112825982272625, + 0.011585780419409275, + 0.010778198949992657, + 0.01101294532418251, + 0.01199549064040184, + 0.010667005553841591, + 0.004881318658590317, + 0.008905704133212566, + 0.005943634081631899, + 0.00499185174703598, + 0.008758621290326118, + 0.006276511121541262, + 0.005475706420838833, + 0.004882162902504206, + 0.004054334945976734, + 0.01058254111558199, + 0.005912410095334053, + 0.005731826648116112, + 0.0057222116738557816, + 0.0058382656425237656, + 0.004085067193955183, + 0.0026858504861593246, + 0.003866097191348672, + 0.005373278632760048, + 0.007711596321314573, + 0.006140836980193853, + 0.005340911913663149, + 0.00966759491711855, + 0.0071549201384186745, + 0.004356252960860729, + 0.016239939257502556, + 0.007233299780637026, + 0.011367762461304665, + 0.010261795483529568, + 0.00871360208839178, + 0.011150013655424118, + 0.00854480266571045, + 0.010869363322854042, + 0.008846528828144073, + 0.013271232135593891, + 0.011840798892080784, + 0.014406601898372173, + 0.018614092841744423, + 0.013049923814833164, + 0.008028602227568626, + 0.007843896746635437, + 0.006399946287274361, + 0.005914055276662111, + 0.01110417116433382, + 0.008447188884019852, + 0.006060814019292593, + 0.00819981750100851, + 0.008452797308564186, + 0.008347601629793644, + 0.017968086525797844, + 0.021759945899248123, + 0.02423333004117012, + 0.011601858772337437, + 0.008746258914470673, + 0.008282857947051525, + 0.008082716725766659, + 0.019200390204787254, + 0.018810417503118515, + 0.05706341937184334, + 0.1359235644340515, + 0.021517008543014526, + 0.025114532560110092, + 0.061128705739974976, + 0.011291039176285267, + 0.029753662645816803, + 0.0036735190078616142, + 0.004467255901545286, + 0.005280036944895983, + 0.003640236333012581, + 0.008389290422201157, + 0.007232239004224539, + 0.0041137379594147205, + 0.005444904789328575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.02633400447666645 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, slowly and steadily advances through a tranquil bamboo forest. Bright, silvery moonlight filters through the emerald green bamboo leaves that tower into the night sky, casting dappled shadows on the leaf-covered ground. A thin mist lingers in the forest. At the center of the frame, Hua Mulan, clad in a deep cyan silk Hanfu (Chinese traditional costume), is immersed in her sword dance. Her long, jet-black hair is tied high in a neat bun, strands brushing against her cheeks in the breeze. Her expression is focused and serene, her eyes piercing, and in her hand, she clutches a longsword, gleaming with a cold gleam. Her movements are fluid, her body gracefully twirling, dipping, and leaping in time with the sword's thrusts, her wide sleeves drawing graceful arcs through the air. As she swung her sword, the highly polished blade captured and reflected the cool moonlight, leaving behind trails of light in the air as smooth and shining as liquid mercury, perfectly interpreting the artistic conception of \"the sword shines like water\". The whole scene, under the cool light and shadow, created a cinematic atmosphere that combines power, beauty and loneliness.", + "seed": 11999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_5/index_1949_seed_11999_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01377557129599154, + "vlm_logprob": -0.01377557129599154, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.01377557129599154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011376039125025272 + ], + "opposite_logprob_tokens": [ + -5.011375904083252 + ], + "logprob_reward": -0.011376039125025272, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011376039125025272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10110349208116531, + 0.04073294624686241, + 0.055072102695703506, + 0.05752255767583847, + 0.01898103952407837, + 0.043160609900951385, + 0.02226576954126358, + 0.04640910029411316, + 0.03019915334880352, + 0.5845533013343811 + ], + "credit_map": [ + 0.010541276074945927, + 0.018787957727909088, + 0.02024095132946968, + 0.007119996473193169, + 0.00976416002959013, + 0.01182076521217823, + 0.007052234839648008, + 0.015776148065924644, + 0.004842462949454784, + 0.0058667901903390884, + 0.004985675681382418, + 0.004874291829764843, + 0.004113589413464069, + 0.007223073858767748, + 0.004213838372379541, + 0.004613220226019621, + 0.005523913539946079, + 0.01892978325486183, + 0.00408697035163641, + 0.0027619763277471066, + 0.0025892502162605524, + 0.003902687691152096, + 0.0060055614449083805, + 0.011271950788795948, + 0.005423614755272865, + 0.005117945838719606, + 0.015027946792542934, + 0.009179801680147648, + 0.0036356865894049406, + 0.007156474515795708, + 0.004588471259921789, + 0.007392611354589462, + 0.0016476999735459685, + 0.005340660456568003, + 0.002929302863776684, + 0.00237744371406734, + 0.001089743571355939, + 0.002026707399636507, + 0.001341666909866035, + 0.0022278130054473877, + 0.003924108576029539, + 0.0028846419882029295, + 0.013091283850371838, + 0.002882074797526002, + 0.002503307070583105, + 0.0020620047580450773, + 0.011920221149921417, + 0.0038929658476263285, + 0.0032181660644710064, + 0.0023426152765750885, + 0.0022231980692595243, + 0.002469937317073345, + 0.0024825939908623695, + 0.0030836095102131367, + 0.002374558011069894, + 0.004071090370416641, + 0.012017753906548023, + 0.002790865022689104, + 0.003581882920116186, + 0.002022812142968178, + 0.0030758394859731197, + 0.003335408866405487, + 0.004413592629134655, + 0.01517094299197197, + 0.002568094292655587, + 0.006929061841219664, + 0.007522513624280691, + 0.0024854475632309914, + 0.002734513720497489, + 0.0022540264762938023, + 0.0029770841356366873, + 0.002728409832343459, + 0.06069166585803032, + 0.15811295807361603, + 0.06653576344251633, + 0.08115839213132858, + 0.03955812007188797, + 0.09948203712701797, + 0.036753419786691666, + 0.04226086661219597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.011376039125025272 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264941602945328 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264941602945328, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264941602945328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14901602268218994, + 0.06708098202943802, + 0.03467746451497078, + 0.07786547392606735, + 0.0531122162938118, + 0.11465036123991013, + 0.057492103427648544, + 0.052649933844804764, + 0.20734800398349762, + 0.18610742688179016 + ], + "credit_map": [ + 0.011301805265247822, + 0.038593463599681854, + 0.016327064484357834, + 0.011160306632518768, + 0.01512299757450819, + 0.014804252423346043, + 0.009923729114234447, + 0.031782422214746475, + 0.008639461360871792, + 0.010233219712972641, + 0.011628703214228153, + 0.006726214196532965, + 0.005078192800283432, + 0.008921335451304913, + 0.00602000392973423, + 0.009833849035203457, + 0.0038787664379924536, + 0.006327616982161999, + 0.0033157391007989645, + 0.002530994825065136, + 0.005832848604768515, + 0.0034551138523966074, + 0.004776936024427414, + 0.004559450317174196, + 0.00554730324074626, + 0.006032048724591732, + 0.018991699442267418, + 0.012166176922619343, + 0.009462782181799412, + 0.005731235723942518, + 0.01138785295188427, + 0.008546375669538975, + 0.007469752337783575, + 0.006047223228961229, + 0.011281461454927921, + 0.008308659307658672, + 0.004915510769933462, + 0.004741391632705927, + 0.005401712842285633, + 0.004946503322571516, + 0.010814458131790161, + 0.005763876251876354, + 0.0427149161696434, + 0.007155309896916151, + 0.0047633228823542595, + 0.004583930131047964, + 0.018024224787950516, + 0.020830316469073296, + 0.006424906197935343, + 0.006790868937969208, + 0.007750018034130335, + 0.011021924205124378, + 0.009650275111198425, + 0.005067205987870693, + 0.006098954938352108, + 0.004687948618084192, + 0.0063727437518537045, + 0.00456934655085206, + 0.007201002445071936, + 0.005663369316607714, + 0.007713760249316692, + 0.006730180699378252, + 0.009433106519281864, + 0.004966426640748978, + 0.01789524033665657, + 0.0773625448346138, + 0.06935058534145355, + 0.01004557590931654, + 0.007757233921438456, + 0.010149200446903706, + 0.0068680401891469955, + 0.007919583469629288, + 0.024206649512052536, + 0.04851320758461952, + 0.021323325112462044, + 0.026273192837834358, + 0.012920360080897808, + 0.02712041325867176, + 0.012755007483065128, + 0.01299527008086443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.005264941602945328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024915861431509256 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024915861431509256, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024915861431509256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11911258846521378, + 0.08405786007642746, + 0.07086881250143051, + 0.1070309728384018, + 0.019677435979247093, + 0.22747379541397095, + 0.015734529122710228, + 0.03149382397532463, + 0.2158609926700592, + 0.10868918895721436 + ], + "credit_map": [ + 0.011927052401006222, + 0.0321437269449234, + 0.014744098298251629, + 0.0066635701805353165, + 0.0075889225117862225, + 0.012403849512338638, + 0.007698871195316315, + 0.025942498818039894, + 0.017456835135817528, + 0.006435387767851353, + 0.010594155639410019, + 0.007944202050566673, + 0.007315740920603275, + 0.013477801345288754, + 0.006985655520111322, + 0.013848084025084972, + 0.005359247326850891, + 0.011875590309500694, + 0.0049003721214830875, + 0.005843783728778362, + 0.009811926633119583, + 0.0110224774107337, + 0.008042099885642529, + 0.014013318344950676, + 0.006237591151148081, + 0.008278872817754745, + 0.018528591841459274, + 0.01528170332312584, + 0.01872081123292446, + 0.009175128303468227, + 0.018677327781915665, + 0.01213094498962164, + 0.001754322205670178, + 0.0033734855242073536, + 0.0031452986877411604, + 0.002711752662435174, + 0.0023244626354426146, + 0.00158259691670537, + 0.00284571829251945, + 0.0019397987052798271, + 0.02072582021355629, + 0.010672079399228096, + 0.07801211625337601, + 0.014743228442966938, + 0.01180688850581646, + 0.008562355302274227, + 0.055117689073085785, + 0.027833621948957443, + 0.001970653422176838, + 0.0015191198326647282, + 0.001754722441546619, + 0.002569513861089945, + 0.002162482589483261, + 0.001452464610338211, + 0.0026129866018891335, + 0.0016925850650295615, + 0.004807842895388603, + 0.002344096777960658, + 0.003119589062407613, + 0.0025635177735239267, + 0.003958710003644228, + 0.0020705019123852253, + 0.004626741632819176, + 0.008002824150025845, + 0.017373284325003624, + 0.07410849630832672, + 0.0671166405081749, + 0.012515067122876644, + 0.011418350972235203, + 0.011876572854816914, + 0.010332177393138409, + 0.011120407842099667, + 0.00997012760490179, + 0.027066921815276146, + 0.01273682713508606, + 0.013553974218666553, + 0.00818018987774849, + 0.021758217364549637, + 0.007703281473368406, + 0.00771965179592371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0024915861431509256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606355793774128 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.01606355793774128, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606355793774128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20480680465698242, + 0.06858835369348526, + 0.07305008918046951, + 0.05595426261425018, + 0.015674084424972534, + 0.023476356640458107, + 0.01021594274789095, + 0.07033058255910873, + 0.18866324424743652, + 0.2892402410507202 + ], + "credit_map": [ + 0.013240644708275795, + 0.07029783725738525, + 0.022350654006004333, + 0.011114591732621193, + 0.014580999501049519, + 0.017452115193009377, + 0.009114641696214676, + 0.04665530472993851, + 0.013265484012663364, + 0.007544752676039934, + 0.008828883990645409, + 0.006366105750203133, + 0.005086553283035755, + 0.014158814214169979, + 0.004638328682631254, + 0.008699426427483559, + 0.006544803734868765, + 0.026179011911153793, + 0.0042010885663330555, + 0.003807265544310212, + 0.004695639945566654, + 0.005198829807341099, + 0.007534190081059933, + 0.014889264479279518, + 0.0037194229662418365, + 0.004986272193491459, + 0.00811147503554821, + 0.007613672874867916, + 0.009600372053682804, + 0.005022732075303793, + 0.010259685106575489, + 0.006640627980232239, + 0.001501495367847383, + 0.002798516768962145, + 0.0030032291542738676, + 0.0028275076765567064, + 0.0014760092599317431, + 0.0015255421167239547, + 0.0013000620529055595, + 0.0012417231919243932, + 0.0025409720838069916, + 0.0022593720350414515, + 0.0033510916400700808, + 0.0018154639983549714, + 0.0036933058872818947, + 0.0024572338443249464, + 0.0045193862169981, + 0.002839529188349843, + 0.0013718329137191176, + 0.00096130435122177, + 0.0010775602422654629, + 0.001321710180491209, + 0.001593763125129044, + 0.0014569992199540138, + 0.001154873170889914, + 0.0012778999516740441, + 0.011877326294779778, + 0.004604381509125233, + 0.006739557720720768, + 0.005910052452236414, + 0.006768842227756977, + 0.004925566725432873, + 0.010870805010199547, + 0.018634052947163582, + 0.015760861337184906, + 0.06603750586509705, + 0.06732859462499619, + 0.009222103282809258, + 0.007226443383842707, + 0.00925184041261673, + 0.0071296729147434235, + 0.006706228479743004, + 0.030816618353128433, + 0.07766307890415192, + 0.033014990389347076, + 0.04013910889625549, + 0.019883349537849426, + 0.048834770917892456, + 0.017982816323637962, + 0.020905522629618645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01606355793774128 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest with emerald green leaves under a night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Hua Mulan wearing a deep cyan silk Hanfu with long sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a sword dance with fluid movements including twirling, dipping, and leaping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the longsword reflect the moonlight to create trails of light resembling liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly advancing through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033681731671094894 + ], + "opposite_logprob_tokens": [ + -3.408681631088257 + ], + "logprob_reward": -0.033681731671094894, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033681731671094894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22415941953659058, + 0.04978016018867493, + 0.07403090596199036, + 0.08162569999694824, + 0.03395263850688934, + 0.07983049005270004, + 0.020276715978980064, + 0.09975593537092209, + 0.18316113948822021, + 0.1534268856048584 + ], + "credit_map": [ + 0.015604679472744465, + 0.09240465611219406, + 0.01997310295701027, + 0.009215728379786015, + 0.009333881549537182, + 0.011431243270635605, + 0.005808698013424873, + 0.06038742512464523, + 0.0059730238281190395, + 0.008187135681509972, + 0.005498782731592655, + 0.008456524461507797, + 0.006003969814628363, + 0.005678728222846985, + 0.00425269640982151, + 0.005729300901293755, + 0.00667142029851675, + 0.019493211060762405, + 0.004009880591183901, + 0.0052783917635679245, + 0.012411931529641151, + 0.0057249972596764565, + 0.00798829086124897, + 0.012452779337763786, + 0.004804492462426424, + 0.006302841939032078, + 0.01364133507013321, + 0.01209377683699131, + 0.013632180169224739, + 0.00599049124866724, + 0.014640096575021744, + 0.010520484298467636, + 0.00393744045868516, + 0.005635354667901993, + 0.007370476145297289, + 0.00435026315972209, + 0.0028774235397577286, + 0.004058006685227156, + 0.0026618177071213722, + 0.00306185451336205, + 0.00942049827426672, + 0.00888090394437313, + 0.015096703544259071, + 0.012913794256746769, + 0.008897452615201473, + 0.0064366660080850124, + 0.008881300687789917, + 0.009303173050284386, + 0.004265216179192066, + 0.0026836746837943792, + 0.0024811076000332832, + 0.001983774360269308, + 0.0021413154900074005, + 0.0021092251408845186, + 0.0019062524661421776, + 0.002706149360165, + 0.019400613382458687, + 0.008522533811628819, + 0.010343638248741627, + 0.006687196437269449, + 0.006654923316091299, + 0.008396602235734463, + 0.012844080105423927, + 0.026906346902251244, + 0.015715651214122772, + 0.06503880023956299, + 0.0641094222664833, + 0.009154866449534893, + 0.006639454048126936, + 0.009637911804020405, + 0.006236717104911804, + 0.0066283103078603745, + 0.016654331237077713, + 0.04084367677569389, + 0.01738208904862404, + 0.020736467093229294, + 0.010617010295391083, + 0.026713822036981583, + 0.009702182374894619, + 0.010777316987514496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.033681731671094894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_tensor([3146, 3146])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_tensor([3146, 3146])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..be4e75106ca74402e007b215682bbe68917fa3ae --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_tensor([3146, 3146])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Amidst the crushing silence and darkness of the ten-thousand-meter-deep ocean, the camera pans at a slow, deliberate downward angle, piercing the dust-like marine snow suspended in the water, ultimately revealing a scene imbued with ancient power and quiet destruction. A massive mosasaur, the size of a mountain, its vast bulk seemingly blending into the boundless abyss. Only as the faint blue light passes over it does its skin, coated in smooth, moist, slate-gray scales, reflect a faint, oily texture. Its streamlined form remains almost perfectly still in the water, its massive tail fin swaying imperceptibly to maintain its hovering position. Before this prehistoric behemoth, an equally massive, ancient dragon egg emanates a faint blue glow. The light seeps from within the eggshell, casting intricate shadows across the surface, riven with ancient cracks and fossilized veins, like the slow beating of a heart. The Mosasaur's long, narrow jaws, imbued with destructive force, opened wide. With a silent yet powerful movement, its jaws, powerful enough to shatter steel, closed with a devastating, steady, and deliberate bite upon the upper half of the dragon egg. Without explosion or sound, in absolute silence, the massive, faintly blue-glowing shell of the dragon egg collapsed inward under the irresistible pressure, shattering. Countless crystal-like fragments silently peeled away, drifting slowly into the bottomless darkness like ghostly continental plates. The faint blue light within the egg flared violently as it was devoured, illuminating the terrifying structure deep within the Mosasaur's throat. Then, it died forever. The entire scene was enveloped in a hushed, profound atmosphere, a witness to creation and destruction.", + "seed": 14648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_3146_seed_14648_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14762979447841645, + "vlm_logprob": -0.14762979447841645, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.14762979447841645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02184426039457321 + ], + "opposite_logprob_tokens": [ + -4.271844387054443 + ], + "logprob_reward": -0.02184426039457321, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02184426039457321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17709700763225555, + 0.08666078746318817, + 0.03843862563371658, + 0.05434906482696533, + 0.09544342756271362, + 0.053648095577955246, + 0.13181550800800323, + 0.14496245980262756, + 0.1622251719236374, + 0.05535982549190521 + ], + "credit_map": [ + 0.009939463809132576, + 0.017403321340680122, + 0.014182684943079948, + 0.012088297866284847, + 0.033828917890787125, + 0.051418278366327286, + 0.010374280624091625, + 0.02786177210509777, + 0.005833705421537161, + 0.006764781195670366, + 0.008103957399725914, + 0.007365624886006117, + 0.009695080108940601, + 0.02278822287917137, + 0.004069814458489418, + 0.02203960157930851, + 0.006242252420634031, + 0.005103216040879488, + 0.004580694250762463, + 0.0040308586321771145, + 0.0036929347552359104, + 0.006456604227423668, + 0.00307442550547421, + 0.00525764049962163, + 0.005313003435730934, + 0.005967575591057539, + 0.0077856858260929585, + 0.006021542474627495, + 0.004410117398947477, + 0.006789051927626133, + 0.008363542146980762, + 0.009698547422885895, + 0.010944332927465439, + 0.008982681669294834, + 0.012986849062144756, + 0.013002386316657066, + 0.01119108498096466, + 0.009796113707125187, + 0.016675498336553574, + 0.011864484287798405, + 0.005127666052430868, + 0.004919861909002066, + 0.007843389175832272, + 0.005800016224384308, + 0.006482963915914297, + 0.0069070616737008095, + 0.00863250158727169, + 0.007934638299047947, + 0.012802427634596825, + 0.012508642859756947, + 0.017170971259474754, + 0.019552825018763542, + 0.022029858082532883, + 0.01613852009177208, + 0.018066097050905228, + 0.013546165078878403, + 0.0067931427620351315, + 0.006736773997545242, + 0.01101578027009964, + 0.006952593103051186, + 0.005097797140479088, + 0.0450422577559948, + 0.016463514417409897, + 0.04686060920357704, + 0.008091223426163197, + 0.01076061837375164, + 0.015700599178671837, + 0.01651306450366974, + 0.0111847510561347, + 0.042945004999637604, + 0.014372461475431919, + 0.04265743866562843, + 0.0031482046470046043, + 0.005244872532784939, + 0.01018108706921339, + 0.009042457677423954, + 0.003148660995066166, + 0.006006457377225161, + 0.009113326668739319, + 0.009474758058786392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.02184426039457321 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309616982936859 + ], + "opposite_logprob_tokens": [ + -3.1680960655212402 + ], + "logprob_reward": -0.04309616982936859, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309616982936859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26127904653549194, + 0.06661800295114517, + 0.04666844382882118, + 0.04958239942789078, + 0.07217742502689362, + 0.0608648806810379, + 0.17936119437217712, + 0.12963983416557312, + 0.08972031623125076, + 0.044088512659072876 + ], + "credit_map": [ + 0.016358541324734688, + 0.018148135393857956, + 0.016292452812194824, + 0.013140442781150341, + 0.042812518775463104, + 0.08729897439479828, + 0.013242175802588463, + 0.05398581176996231, + 0.008655525743961334, + 0.009042912162840366, + 0.007304871920496225, + 0.011808102019131184, + 0.00940246693789959, + 0.008989885449409485, + 0.004280379042029381, + 0.007133854553103447, + 0.006705516949295998, + 0.007696398068219423, + 0.005155991762876511, + 0.005927457008510828, + 0.004222748335450888, + 0.007263198494911194, + 0.004027014598250389, + 0.005670119076967239, + 0.005265218671411276, + 0.005275997798889875, + 0.006714046001434326, + 0.0057230317033827305, + 0.003976522944867611, + 0.0059804231859743595, + 0.00833897478878498, + 0.008308181539177895, + 0.007655545137822628, + 0.006503253243863583, + 0.009812000207602978, + 0.009445294737815857, + 0.008869646117091179, + 0.007178819738328457, + 0.014518269337713718, + 0.008194594644010067, + 0.0066659883596003056, + 0.004714382812380791, + 0.009920239448547363, + 0.007049719803035259, + 0.005742201581597328, + 0.008207007311284542, + 0.008969762362539768, + 0.00959557481110096, + 0.01741090416908264, + 0.016832681372761726, + 0.02566436491906643, + 0.026194093748927116, + 0.029885679483413696, + 0.022229986265301704, + 0.023161446675658226, + 0.017982039600610733, + 0.006771066691726446, + 0.008561920374631882, + 0.015216170810163021, + 0.009279316291213036, + 0.005563472397625446, + 0.031596675515174866, + 0.0187075212597847, + 0.033943694084882736, + 0.004795209039002657, + 0.006516318302601576, + 0.008933138102293015, + 0.008726075291633606, + 0.006136396899819374, + 0.023667018860578537, + 0.007573348935693502, + 0.023372814059257507, + 0.002184513257816434, + 0.004119849298149347, + 0.009206747636198997, + 0.007489136885851622, + 0.0026339110918343067, + 0.004494728986173868, + 0.007344326004385948, + 0.006615299731492996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.04309616982936859 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34877845644950867 + ], + "opposite_logprob_tokens": [ + -1.223778486251831 + ], + "logprob_reward": -0.34877845644950867, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34877845644950867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10584438592195511, + 0.06795070320367813, + 0.08327499777078629, + 0.07778839021921158, + 0.18696996569633484, + 0.06872484087944031, + 0.0666452944278717, + 0.16415783762931824, + 0.12743398547172546, + 0.05120960250496864 + ], + "credit_map": [ + 0.009379740804433823, + 0.017486492171883583, + 0.015969866886734962, + 0.015016513876616955, + 0.012654726393520832, + 0.01240637619048357, + 0.01121586374938488, + 0.011714807711541653, + 0.006442365236580372, + 0.006482438184320927, + 0.007573504000902176, + 0.00876331515610218, + 0.01421920582652092, + 0.014141641557216644, + 0.003787779016420245, + 0.006540453527122736, + 0.013235608115792274, + 0.009505266323685646, + 0.010211996734142303, + 0.009878294542431831, + 0.009141919203102589, + 0.014458676800131798, + 0.006029251962900162, + 0.010813974775373936, + 0.007354266010224819, + 0.007219021674245596, + 0.009431886486709118, + 0.008109968155622482, + 0.008749641478061676, + 0.009337770752608776, + 0.0159470085054636, + 0.011638818308711052, + 0.02060893550515175, + 0.018395455554127693, + 0.02415153756737709, + 0.025777656584978104, + 0.022876230999827385, + 0.01851167157292366, + 0.03642461076378822, + 0.020223859697580338, + 0.007229969371110201, + 0.005905794911086559, + 0.01146632619202137, + 0.006909254938364029, + 0.007201783359050751, + 0.008689476177096367, + 0.010108508169651031, + 0.01121372077614069, + 0.006953482050448656, + 0.005866118706762791, + 0.00946804042905569, + 0.010426189750432968, + 0.008954151533544064, + 0.008318397216498852, + 0.009626087732613087, + 0.0070328181609511375, + 0.011452704668045044, + 0.006657157558947802, + 0.01591910980641842, + 0.010243489407002926, + 0.005792559124529362, + 0.051307134330272675, + 0.009903818368911743, + 0.05288185179233551, + 0.00742330402135849, + 0.009532846510410309, + 0.010223724879324436, + 0.011242512613534927, + 0.009519482031464577, + 0.03464744612574577, + 0.01101046334952116, + 0.03383418917655945, + 0.004272040911018848, + 0.004800721071660519, + 0.00933393370360136, + 0.008054785430431366, + 0.003330979263409972, + 0.0054527828469872475, + 0.007570634596049786, + 0.008393715135753155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.34877845644950867 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310918599367142 + ], + "opposite_logprob_tokens": [ + -3.168109178543091 + ], + "logprob_reward": -0.04310918599367142, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310918599367142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11065206676721573, + 0.11026134341955185, + 0.036847811192274094, + 0.05367676913738251, + 0.028824999928474426, + 0.04699926823377609, + 0.05125976353883743, + 0.26587945222854614, + 0.2047509402036667, + 0.09084750711917877 + ], + "credit_map": [ + 0.007779114879667759, + 0.013251906260848045, + 0.015130303800106049, + 0.012256636284291744, + 0.01943805254995823, + 0.02331317961215973, + 0.008421427570283413, + 0.011061466298997402, + 0.010037091560661793, + 0.012463708408176899, + 0.011159990914165974, + 0.010075354017317295, + 0.02173629030585289, + 0.029320891946554184, + 0.004579851403832436, + 0.010888188146054745, + 0.006755989510565996, + 0.004337838385254145, + 0.004018702544271946, + 0.00427527679130435, + 0.003715785685926676, + 0.005640069022774696, + 0.0026661171577870846, + 0.005438042338937521, + 0.006831794045865536, + 0.005812422372400761, + 0.0056806583888828754, + 0.006323082838207483, + 0.006218661088496447, + 0.006381814368069172, + 0.009263778105378151, + 0.007164569105952978, + 0.0034970594570040703, + 0.0023756837472319603, + 0.0034179668873548508, + 0.0032983713317662477, + 0.0035616559907794, + 0.0033049846533685923, + 0.004967134445905685, + 0.00440214853733778, + 0.00665097963064909, + 0.004521585069596767, + 0.007356919348239899, + 0.005369339603930712, + 0.004839036148041487, + 0.004895152989774942, + 0.006912921089679003, + 0.006453342270106077, + 0.006227829493582249, + 0.0041255340911448, + 0.006636894308030605, + 0.006471829488873482, + 0.00864358153194189, + 0.0053903693333268166, + 0.008595135062932968, + 0.005168598610907793, + 0.014911004342138767, + 0.00969476718455553, + 0.019414221867918968, + 0.012331602163612843, + 0.008230061270296574, + 0.08937991410493851, + 0.017172271385788918, + 0.09474564343690872, + 0.011056879535317421, + 0.012434416450560093, + 0.015330720692873001, + 0.014735774137079716, + 0.012301795184612274, + 0.06058108061552048, + 0.016450682654976845, + 0.061859626322984695, + 0.006809630896896124, + 0.00930454395711422, + 0.014275827445089817, + 0.015180415473878384, + 0.006186706013977528, + 0.009337332099676132, + 0.015541484579443932, + 0.014211589470505714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.04310918599367142 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813208997249603 + ], + "opposite_logprob_tokens": [ + -1.4063209295272827 + ], + "logprob_reward": -0.2813208997249603, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813208997249603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15717987716197968, + 0.21477672457695007, + 0.042110372334718704, + 0.047848623245954514, + 0.0351567342877388, + 0.0483817495405674, + 0.10540106892585754, + 0.12426667660474777, + 0.17294800281524658, + 0.05193008854985237 + ], + "credit_map": [ + 0.011180578731000423, + 0.01760675385594368, + 0.0162979606539011, + 0.014239709824323654, + 0.030760720372200012, + 0.041224293410778046, + 0.00905191246420145, + 0.01681794971227646, + 0.008708768524229527, + 0.0126785384491086, + 0.009373498149216175, + 0.010126437060534954, + 0.03511284291744232, + 0.08006520569324493, + 0.0051603480242192745, + 0.053551092743873596, + 0.006114956457167864, + 0.006679893005639315, + 0.004737043287605047, + 0.004639930557459593, + 0.004511354025453329, + 0.007530464790761471, + 0.0030877795070409775, + 0.004808951169252396, + 0.004743041004985571, + 0.004543105605989695, + 0.004677955526858568, + 0.0053064594976603985, + 0.0060689980164170265, + 0.0056042056530714035, + 0.01166116539388895, + 0.0052436948753893375, + 0.00430286256596446, + 0.00398188317194581, + 0.00443881843239069, + 0.005097041372209787, + 0.004039815627038479, + 0.0034327893517911434, + 0.006406772881746292, + 0.003456749487668276, + 0.0059251743368804455, + 0.004419920966029167, + 0.006745335645973682, + 0.005436464678496122, + 0.006914709694683552, + 0.0061768279410898685, + 0.00477170292288065, + 0.0079916100949049, + 0.012915656901896, + 0.012143542990088463, + 0.016451815143227577, + 0.012750579975545406, + 0.010078812018036842, + 0.011438998393714428, + 0.01497241947799921, + 0.014649241231381893, + 0.013129614293575287, + 0.00785556435585022, + 0.02126597985625267, + 0.013862784020602703, + 0.00764718372374773, + 0.023570498451590538, + 0.012821584939956665, + 0.024113468825817108, + 0.01409054733812809, + 0.014349428005516529, + 0.028060629963874817, + 0.024069786071777344, + 0.018392130732536316, + 0.02451309934258461, + 0.02551744505763054, + 0.02395494095981121, + 0.004333310294896364, + 0.005851867608726025, + 0.008501643314957619, + 0.009578516706824303, + 0.0039190808311104774, + 0.006031669210642576, + 0.006158873438835144, + 0.007555127609521151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.2813208997249603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst the crushing silence and darkness of the ten-thousand-meter-deep ocean, the camera pans at a slow, deliberate downward angle, piercing the dust-like marine snow suspended in the water, ultimately revealing a scene imbued with ancient power and quiet destruction. A massive mosasaur, the size of a mountain, its vast bulk seemingly blending into the boundless abyss. Only as the faint blue light passes over it does its skin, coated in smooth, moist, slate-gray scales, reflect a faint, oily texture. Its streamlined form remains almost perfectly still in the water, its massive tail fin swaying imperceptibly to maintain its hovering position. Before this prehistoric behemoth, an equally massive, ancient dragon egg emanates a faint blue glow. The light seeps from within the eggshell, casting intricate shadows across the surface, riven with ancient cracks and fossilized veins, like the slow beating of a heart. The Mosasaur's long, narrow jaws, imbued with destructive force, opened wide. With a silent yet powerful movement, its jaws, powerful enough to shatter steel, closed with a devastating, steady, and deliberate bite upon the upper half of the dragon egg. Without explosion or sound, in absolute silence, the massive, faintly blue-glowing shell of the dragon egg collapsed inward under the irresistible pressure, shattering. Countless crystal-like fragments silently peeled away, drifting slowly into the bottomless darkness like ghostly continental plates. The faint blue light within the egg flared violently as it was devoured, illuminating the terrifying structure deep within the Mosasaur's throat. Then, it died forever. The entire scene was enveloped in a hushed, profound atmosphere, a witness to creation and destruction.", + "seed": 14649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_3146_seed_14649_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09144682884216308, + "vlm_logprob": -0.09144682884216308, + "vlm_yes_no_margin": 2.925000023841858, + "vlm_token_credit": -0.09144682884216308, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02507782354950905 + ], + "opposite_logprob_tokens": [ + -4.400077819824219 + ], + "logprob_reward": -0.02507782354950905, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02507782354950905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12017049640417099, + 0.23322732746601105, + 0.0344749391078949, + 0.11599954217672348, + 0.03276682645082474, + 0.037412647157907486, + 0.1568034142255783, + 0.10384280979633331, + 0.09372653067111969, + 0.07157544791698456 + ], + "credit_map": [ + 0.009740523062646389, + 0.00696592265740037, + 0.022609570994973183, + 0.007986288517713547, + 0.00776235107332468, + 0.04213098809123039, + 0.007492889184504747, + 0.015481961891055107, + 0.04326624423265457, + 0.010845702141523361, + 0.026051387190818787, + 0.020604202523827553, + 0.012338482774794102, + 0.027973849326372147, + 0.020098386332392693, + 0.07204905152320862, + 0.004360847640782595, + 0.0036377853248268366, + 0.0037799952551722527, + 0.004945777822285891, + 0.002797109307721257, + 0.004383767023682594, + 0.0027847541496157646, + 0.007784898858517408, + 0.010758751071989536, + 0.008206644095480442, + 0.008735542185604572, + 0.008482880890369415, + 0.00547849340364337, + 0.01976020075380802, + 0.012199153192341328, + 0.042377881705760956, + 0.003075994085520506, + 0.0036787984427064657, + 0.003305958118289709, + 0.0035729387309402227, + 0.002743375953286886, + 0.005037098191678524, + 0.0030258081387728453, + 0.00832685362547636, + 0.003837436670437455, + 0.004260428715497255, + 0.0055156610906124115, + 0.006213605869561434, + 0.003480025567114353, + 0.005048676859587431, + 0.003280322765931487, + 0.0057764905504882336, + 0.017685534432530403, + 0.021178308874368668, + 0.020665347576141357, + 0.01911567524075508, + 0.013886435888707638, + 0.020527195185422897, + 0.02194679155945778, + 0.021798137575387955, + 0.012894702143967152, + 0.010891511105000973, + 0.010286732576787472, + 0.010410803370177746, + 0.011702272109687328, + 0.013902658596634865, + 0.014584256336092949, + 0.01916988007724285, + 0.014724069274961948, + 0.009232080541551113, + 0.016753727570176125, + 0.012521776370704174, + 0.005947123281657696, + 0.012567064724862576, + 0.00859778467565775, + 0.013382907956838608, + 0.00785988848656416, + 0.007003023289144039, + 0.012910161167383194, + 0.006804545409977436, + 0.010960976593196392, + 0.009417188353836536, + 0.00841404777020216, + 0.008205612190067768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.02507782354950905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309182986617088 + ], + "opposite_logprob_tokens": [ + -3.1680917739868164 + ], + "logprob_reward": -0.04309182986617088, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309182986617088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07369925826787949, + 0.06364529579877853, + 0.04716608300805092, + 0.03634088486433029, + 0.15049709379673004, + 0.09223508089780807, + 0.19761846959590912, + 0.12831194698810577, + 0.14466841518878937, + 0.0658174455165863 + ], + "credit_map": [ + 0.00733674643561244, + 0.008035070262849331, + 0.008738107979297638, + 0.006669065915048122, + 0.006850915029644966, + 0.017253486439585686, + 0.005135754588991404, + 0.013680112548172474, + 0.00571673596277833, + 0.004265002906322479, + 0.010018933564424515, + 0.00931368675082922, + 0.004428118467330933, + 0.011572793126106262, + 0.005030958913266659, + 0.013299069367349148, + 0.0049204109236598015, + 0.007093471474945545, + 0.004972938448190689, + 0.007161395624279976, + 0.0047826627269387245, + 0.007983042858541012, + 0.004667453467845917, + 0.005584708880633116, + 0.003924712538719177, + 0.0038684450555592775, + 0.004364899825304747, + 0.0036390284076333046, + 0.0035386967938393354, + 0.004727999214082956, + 0.004532407503575087, + 0.007744697388261557, + 0.015261842869222164, + 0.01123726088553667, + 0.012473983690142632, + 0.011810846626758575, + 0.007487645372748375, + 0.029383406043052673, + 0.007487140130251646, + 0.055354975163936615, + 0.008120665326714516, + 0.01159587875008583, + 0.007559794932603836, + 0.008846433833241463, + 0.009399493224918842, + 0.015152251347899437, + 0.006113960407674313, + 0.025446603074669838, + 0.022147823125123978, + 0.02577785961329937, + 0.025350121781229973, + 0.02418649196624756, + 0.01726316101849079, + 0.026015421375632286, + 0.02749156393110752, + 0.029386019334197044, + 0.018960019573569298, + 0.014832617715001106, + 0.01211466919630766, + 0.0117184491828084, + 0.016444530338048935, + 0.016823438927531242, + 0.015448972582817078, + 0.021969253197312355, + 0.027791688218712807, + 0.013944649137556553, + 0.02507675252854824, + 0.01914004608988762, + 0.009064365178346634, + 0.020062776282429695, + 0.009439092129468918, + 0.020149031654000282, + 0.007368527352809906, + 0.006736909970641136, + 0.012285157106816769, + 0.005400859750807285, + 0.011325291357934475, + 0.008873814716935158, + 0.006633382756263018, + 0.007193499710410833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.04309182986617088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14288322627544403 + ], + "opposite_logprob_tokens": [ + -2.01788330078125 + ], + "logprob_reward": -0.14288322627544403, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14288322627544403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1919754594564438, + 0.16889157891273499, + 0.043485354632139206, + 0.13839887082576752, + 0.042615022510290146, + 0.02219599299132824, + 0.13284927606582642, + 0.07561086118221283, + 0.09567926079034805, + 0.08829829841852188 + ], + "credit_map": [ + 0.011854481883347034, + 0.008148145861923695, + 0.019826721400022507, + 0.008743301033973694, + 0.009144111536443233, + 0.07134754955768585, + 0.008304710499942303, + 0.054606448858976364, + 0.03082907758653164, + 0.007431968115270138, + 0.018903499469161034, + 0.013666215352714062, + 0.009428233839571476, + 0.023292547091841698, + 0.014318914152681828, + 0.05102110654115677, + 0.00471313763409853, + 0.007050478830933571, + 0.006041946820914745, + 0.005133085418492556, + 0.004341269843280315, + 0.006569580174982548, + 0.0036273810546845198, + 0.006008473690599203, + 0.013061475940048695, + 0.0104307159781456, + 0.0124060595408082, + 0.010941116139292717, + 0.006565840914845467, + 0.021924465894699097, + 0.013678758405148983, + 0.04939043149352074, + 0.004479825496673584, + 0.0039261397905647755, + 0.003981365356594324, + 0.0038681672886013985, + 0.0023496749345213175, + 0.007600398268550634, + 0.0026062019169330597, + 0.01380324736237526, + 0.0021426002494990826, + 0.002635762095451355, + 0.0020139431580901146, + 0.0018050087383016944, + 0.0021337969228625298, + 0.0038005951792001724, + 0.0017638818826526403, + 0.005900405813008547, + 0.014569094404578209, + 0.017829841002821922, + 0.01812809146940708, + 0.015850495547056198, + 0.011918531730771065, + 0.01747025176882744, + 0.01803753338754177, + 0.01904544048011303, + 0.011434814892709255, + 0.008696694858372211, + 0.007652744650840759, + 0.0069035994820296764, + 0.009317919611930847, + 0.010258051566779613, + 0.008656106889247894, + 0.012690930627286434, + 0.011993801221251488, + 0.009768176823854446, + 0.016921132802963257, + 0.013235236518085003, + 0.006419355049729347, + 0.010328324511647224, + 0.013422597199678421, + 0.013590632937848568, + 0.010721827857196331, + 0.009323516860604286, + 0.015805063769221306, + 0.007667786441743374, + 0.0130479596555233, + 0.008761299774050713, + 0.01086085569113493, + 0.01210999209433794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.14288322627544403 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602509379386902 + ], + "opposite_logprob_tokens": [ + -3.8956024646759033 + ], + "logprob_reward": -0.020602509379386902, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602509379386902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22609837353229523, + 0.03126130253076553, + 0.04733595997095108, + 0.05876152589917183, + 0.09312310814857483, + 0.11201640963554382, + 0.0998409241437912, + 0.07222355157136917, + 0.1367359757423401, + 0.12260283529758453 + ], + "credit_map": [ + 0.013762876391410828, + 0.008713261224329472, + 0.03241850808262825, + 0.012438565492630005, + 0.012502084486186504, + 0.0796363428235054, + 0.010947845876216888, + 0.055678874254226685, + 0.004270199220627546, + 0.003016676986590028, + 0.005053343717008829, + 0.004516837187111378, + 0.0022218646481633186, + 0.0037859154399484396, + 0.003256825730204582, + 0.005139640532433987, + 0.004514196887612343, + 0.00721819419413805, + 0.006123803555965424, + 0.005422715563327074, + 0.005398768465965986, + 0.007523057516664267, + 0.004026931710541248, + 0.0071082934737205505, + 0.005482631269842386, + 0.0043092891573905945, + 0.006579630542546511, + 0.004478874616324902, + 0.004272558726370335, + 0.009391617961227894, + 0.006298340391367674, + 0.017948582768440247, + 0.009213555604219437, + 0.007249138318002224, + 0.008627496659755707, + 0.007975402288138866, + 0.004855126608163118, + 0.017746701836586, + 0.005216339603066444, + 0.03223934397101402, + 0.009657961316406727, + 0.011274654418230057, + 0.008412932977080345, + 0.009757108986377716, + 0.009559370577335358, + 0.020609797909855843, + 0.007044495083391666, + 0.03570008650422096, + 0.011172045022249222, + 0.012714467942714691, + 0.013212247751653194, + 0.013728848658502102, + 0.007573409005999565, + 0.013674530200660229, + 0.012045190669596195, + 0.0157201886177063, + 0.010790056549012661, + 0.008310554549098015, + 0.007443321403115988, + 0.006499149836599827, + 0.00958253163844347, + 0.009693264961242676, + 0.007791312411427498, + 0.0121133578941226, + 0.022015566006302834, + 0.014181082136929035, + 0.017993515357375145, + 0.010136487893760204, + 0.010458488948643208, + 0.01785311847925186, + 0.014941902831196785, + 0.02915581688284874, + 0.013667634688317776, + 0.01169012114405632, + 0.025325186550617218, + 0.00941440649330616, + 0.019926873967051506, + 0.012190905399620533, + 0.01109279878437519, + 0.019294900819659233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.020602509379386902 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557875514030457 + ], + "opposite_logprob_tokens": [ + -1.600578784942627 + ], + "logprob_reward": -0.22557875514030457, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557875514030457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30348581075668335, + 0.025889141485095024, + 0.07990578562021255, + 0.13064557313919067, + 0.1019270122051239, + 0.17200036346912384, + 0.02654179185628891, + 0.05047931522130966, + 0.060416221618652344, + 0.048708993941545486 + ], + "credit_map": [ + 0.01672772318124771, + 0.011024349369108677, + 0.053115490823984146, + 0.016984883695840836, + 0.019287999719381332, + 0.11250679194927216, + 0.011706315912306309, + 0.06213228031992912, + 0.0019864423666149378, + 0.0026700575836002827, + 0.003674783743917942, + 0.00309805478900671, + 0.00248897192068398, + 0.0055068801157176495, + 0.002077931771054864, + 0.00438601803034544, + 0.007347859907895327, + 0.008938031271100044, + 0.009493628516793251, + 0.007398473564535379, + 0.00949519220739603, + 0.015503055416047573, + 0.0065382663160562515, + 0.015191282145678997, + 0.01267702691257, + 0.012244298122823238, + 0.011667742393910885, + 0.010693103075027466, + 0.007112582214176655, + 0.02023487165570259, + 0.01321473065763712, + 0.04280121251940727, + 0.008687552995979786, + 0.014281337149441242, + 0.012394007295370102, + 0.013955092057585716, + 0.010530625469982624, + 0.013281187042593956, + 0.008557876572012901, + 0.02023933082818985, + 0.015130922198295593, + 0.017147967591881752, + 0.013266884721815586, + 0.01073034480214119, + 0.01608939841389656, + 0.03215436637401581, + 0.01247358601540327, + 0.05500689521431923, + 0.0027204016223549843, + 0.003071099752560258, + 0.0027230132836848497, + 0.002318731974810362, + 0.002856329083442688, + 0.004353689961135387, + 0.0025914213620126247, + 0.005907103884965181, + 0.005938305519521236, + 0.005185868125408888, + 0.00548448646441102, + 0.008142150938510895, + 0.005167583003640175, + 0.006422285456210375, + 0.0060212621465325356, + 0.008117372170090675, + 0.008355257101356983, + 0.006041890010237694, + 0.007721101865172386, + 0.004435312002897263, + 0.004995871800929308, + 0.009725352749228477, + 0.0047434065490961075, + 0.014398030936717987, + 0.005781103391200304, + 0.005191585514694452, + 0.00791167002171278, + 0.004834221675992012, + 0.005879618227481842, + 0.0050211697816848755, + 0.006333746016025543, + 0.007755879778414965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.22557875514030457 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst the crushing silence and darkness of the ten-thousand-meter-deep ocean, the camera pans at a slow, deliberate downward angle, piercing the dust-like marine snow suspended in the water, ultimately revealing a scene imbued with ancient power and quiet destruction. A massive mosasaur, the size of a mountain, its vast bulk seemingly blending into the boundless abyss. Only as the faint blue light passes over it does its skin, coated in smooth, moist, slate-gray scales, reflect a faint, oily texture. Its streamlined form remains almost perfectly still in the water, its massive tail fin swaying imperceptibly to maintain its hovering position. Before this prehistoric behemoth, an equally massive, ancient dragon egg emanates a faint blue glow. The light seeps from within the eggshell, casting intricate shadows across the surface, riven with ancient cracks and fossilized veins, like the slow beating of a heart. The Mosasaur's long, narrow jaws, imbued with destructive force, opened wide. With a silent yet powerful movement, its jaws, powerful enough to shatter steel, closed with a devastating, steady, and deliberate bite upon the upper half of the dragon egg. Without explosion or sound, in absolute silence, the massive, faintly blue-glowing shell of the dragon egg collapsed inward under the irresistible pressure, shattering. Countless crystal-like fragments silently peeled away, drifting slowly into the bottomless darkness like ghostly continental plates. The faint blue light within the egg flared violently as it was devoured, illuminating the terrifying structure deep within the Mosasaur's throat. Then, it died forever. The entire scene was enveloped in a hushed, profound atmosphere, a witness to creation and destruction.", + "seed": 14650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_3146_seed_14650_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07861980348825455, + "vlm_logprob": -0.07861980348825455, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.07861980348825455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012000152841210365 + ], + "opposite_logprob_tokens": [ + -5.26200008392334 + ], + "logprob_reward": -0.012000152841210365, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.012000152841210365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08600536733865738, + 0.047994982451200485, + 0.0268995463848114, + 0.0848289206624031, + 0.18871979415416718, + 0.07836394011974335, + 0.2886568605899811, + 0.057092029601335526, + 0.08117451518774033, + 0.06026402488350868 + ], + "credit_map": [ + 0.007160675711929798, + 0.013180984184145927, + 0.014577752910554409, + 0.012067346833646297, + 0.0064398739486932755, + 0.011549604125320911, + 0.010740349069237709, + 0.0102887824177742, + 0.005425949580967426, + 0.006637286860495806, + 0.006342650391161442, + 0.007754848804324865, + 0.004596923943608999, + 0.005987972021102905, + 0.006099651101976633, + 0.005149702075868845, + 0.002638604026287794, + 0.0028336469549685717, + 0.0032923936378210783, + 0.004175201524049044, + 0.0031979482155293226, + 0.003846722189337015, + 0.0027897385880351067, + 0.004125290084630251, + 0.00949481688439846, + 0.009783376008272171, + 0.01209928561002016, + 0.010202210396528244, + 0.007543796207755804, + 0.01309683732688427, + 0.008945237845182419, + 0.01366336178034544, + 0.010599139146506786, + 0.007237821817398071, + 0.012355513870716095, + 0.009599460288882256, + 0.008123197592794895, + 0.07576416432857513, + 0.007275983225554228, + 0.05776451900601387, + 0.010658071376383305, + 0.0040121860802173615, + 0.006172007881104946, + 0.00764284236356616, + 0.0034924927167594433, + 0.01455942913889885, + 0.00303295417688787, + 0.028793949633836746, + 0.029009925201535225, + 0.026517225429415703, + 0.07518413662910461, + 0.037992313504219055, + 0.02624167688190937, + 0.028124239295721054, + 0.01937061734497547, + 0.046216726303100586, + 0.009324868209660053, + 0.005055510904639959, + 0.012246926315128803, + 0.006936054676771164, + 0.004794292151927948, + 0.0061921305023133755, + 0.004171549342572689, + 0.008370696566998959, + 0.007582470308989286, + 0.006626056972891092, + 0.013675335794687271, + 0.009393319487571716, + 0.009234437718987465, + 0.010720227845013142, + 0.011849791742861271, + 0.012092874385416508, + 0.004656864330172539, + 0.006808117963373661, + 0.009233231656253338, + 0.009643420577049255, + 0.003661032300442457, + 0.006721172481775284, + 0.012292801402509212, + 0.007247384171932936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.012000152841210365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598072558641434 + ], + "opposite_logprob_tokens": [ + -3.8955981731414795 + ], + "logprob_reward": -0.020598072558641434, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598072558641434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1550723910331726, + 0.0844595804810524, + 0.1305214911699295, + 0.16434670984745026, + 0.0626809373497963, + 0.13601134717464447, + 0.08692381531000137, + 0.05390899255871773, + 0.08043662458658218, + 0.04563814029097557 + ], + "credit_map": [ + 0.018619824200868607, + 0.025151612237095833, + 0.022903859615325928, + 0.017284614965319633, + 0.016521716490387917, + 0.02094290219247341, + 0.021302834153175354, + 0.012345013208687305, + 0.010268938727676868, + 0.008839897811412811, + 0.011097355745732784, + 0.017567692324519157, + 0.005896724760532379, + 0.010443216189742088, + 0.010166836902499199, + 0.01017890777438879, + 0.01544884778559208, + 0.014728027395904064, + 0.013948588632047176, + 0.020014425739645958, + 0.020330414175987244, + 0.016313565894961357, + 0.01602713204920292, + 0.013710465282201767, + 0.0175109151750803, + 0.018733825534582138, + 0.026644283905625343, + 0.021076368167996407, + 0.015631528571248055, + 0.023040277883410454, + 0.01663477160036564, + 0.02507471665740013, + 0.005044783931225538, + 0.004838371649384499, + 0.00886614341288805, + 0.005035808309912682, + 0.006805527489632368, + 0.013343478552997112, + 0.005838377401232719, + 0.01290843728929758, + 0.01998981647193432, + 0.0135880121961236, + 0.013801705092191696, + 0.01440276950597763, + 0.008304262533783913, + 0.01655547134578228, + 0.008193084970116615, + 0.04117622226476669, + 0.009039875119924545, + 0.00859824288636446, + 0.02178056165575981, + 0.010130544193089008, + 0.007500343024730682, + 0.01002566423267126, + 0.004821896553039551, + 0.015026678331196308, + 0.009282989427447319, + 0.005418450105935335, + 0.011384653858840466, + 0.00823635421693325, + 0.003477133810520172, + 0.005458833184093237, + 0.0035525388084352016, + 0.007098029833287001, + 0.008339068852365017, + 0.007671575527638197, + 0.013659861870110035, + 0.010071945376694202, + 0.007832073606550694, + 0.008522827178239822, + 0.01149433758109808, + 0.012844923883676529, + 0.004224275704473257, + 0.0051566376350820065, + 0.006183486431837082, + 0.0062802634201943874, + 0.0035280657466500998, + 0.0035239774733781815, + 0.0117228077724576, + 0.00501862121745944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.020598072558641434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20162968337535858 + ], + "opposite_logprob_tokens": [ + -1.701629638671875 + ], + "logprob_reward": -0.20162968337535858, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20162968337535858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07941675931215286, + 0.07131968438625336, + 0.14193126559257507, + 0.06510230153799057, + 0.14535604417324066, + 0.06390003114938736, + 0.13910701870918274, + 0.05901461839675903, + 0.14209328591823578, + 0.09275888651609421 + ], + "credit_map": [ + 0.007496942300349474, + 0.009339489042758942, + 0.013536517508327961, + 0.009789659641683102, + 0.007355547975748777, + 0.011011146008968353, + 0.010563347488641739, + 0.010324116796255112, + 0.008352356031537056, + 0.009507780894637108, + 0.008552913554012775, + 0.00998646393418312, + 0.005826165899634361, + 0.010360078886151314, + 0.008681995794177055, + 0.010051939636468887, + 0.018756773322820663, + 0.015834946185350418, + 0.015891186892986298, + 0.01921461522579193, + 0.023474043235182762, + 0.01787186786532402, + 0.01739678531885147, + 0.013491055928170681, + 0.006626551505178213, + 0.007228232454508543, + 0.009875752031803131, + 0.006896201055496931, + 0.008580971509218216, + 0.010370765812695026, + 0.0060305483639240265, + 0.009493288584053516, + 0.008219628594815731, + 0.007121094036847353, + 0.012097058817744255, + 0.009519667364656925, + 0.01061142235994339, + 0.050143446773290634, + 0.010135478340089321, + 0.037508271634578705, + 0.006082865875214338, + 0.007349466439336538, + 0.006913256831467152, + 0.005253274459391832, + 0.00805253442376852, + 0.011584434658288956, + 0.005923015996813774, + 0.01274118758738041, + 0.009663274511694908, + 0.006059314589947462, + 0.04421433433890343, + 0.011685273610055447, + 0.0052444771863520145, + 0.0129585275426507, + 0.0062457784079015255, + 0.04303606227040291, + 0.009897880256175995, + 0.005880554672330618, + 0.009431741200387478, + 0.007806242443621159, + 0.004484507255256176, + 0.007805276196449995, + 0.003706664312630892, + 0.010001754388213158, + 0.016150683164596558, + 0.013237289153039455, + 0.023374512791633606, + 0.01672826148569584, + 0.014528866857290268, + 0.014080749824643135, + 0.01997271738946438, + 0.02402021363377571, + 0.008792595006525517, + 0.012153691612184048, + 0.013053507544100285, + 0.013498038984835148, + 0.006745262071490288, + 0.006143232807517052, + 0.022886035963892937, + 0.00948653370141983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.20162968337535858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016080567613244057 + ], + "opposite_logprob_tokens": [ + -4.141080379486084 + ], + "logprob_reward": -0.016080567613244057, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016080567613244057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04328598082065582, + 0.037445131689310074, + 0.02135366201400757, + 0.04200347512960434, + 0.40870732069015503, + 0.11471358686685562, + 0.18376952409744263, + 0.06333119422197342, + 0.044354505836963654, + 0.04103565961122513 + ], + "credit_map": [ + 0.003850266570225358, + 0.005488654598593712, + 0.007009357213973999, + 0.004295594058930874, + 0.00560925854369998, + 0.0062734209932386875, + 0.004841334652155638, + 0.005918093957006931, + 0.003949563018977642, + 0.004357509780675173, + 0.0048340982757508755, + 0.006090836133807898, + 0.003843912621960044, + 0.0047517274506390095, + 0.004736042581498623, + 0.004881442990154028, + 0.0033159032464027405, + 0.0019736767280846834, + 0.001727868802845478, + 0.0023868430871516466, + 0.002489507431164384, + 0.0031103158835321665, + 0.002709386171773076, + 0.0036401599645614624, + 0.004778795875608921, + 0.00456176046282053, + 0.00601036986336112, + 0.005156178958714008, + 0.0034255441278219223, + 0.00669072987511754, + 0.004904855042695999, + 0.006475238595157862, + 0.021048178896307945, + 0.013962116092443466, + 0.018237603828310966, + 0.019237559288740158, + 0.014226800762116909, + 0.17273737490177155, + 0.010585756972432137, + 0.1386719048023224, + 0.01172865740954876, + 0.006945584435015917, + 0.008233248256146908, + 0.009105298668146133, + 0.010286279954016209, + 0.02337983436882496, + 0.006418647710233927, + 0.03861604258418083, + 0.017374521121382713, + 0.015410576947033405, + 0.0535544827580452, + 0.02113293670117855, + 0.014718788675963879, + 0.016094911843538284, + 0.010170841589570045, + 0.035312455147504807, + 0.008232872001826763, + 0.006321176420897245, + 0.010257214307785034, + 0.006943670567125082, + 0.006040007807314396, + 0.010348915122449398, + 0.004870311357080936, + 0.010317022912204266, + 0.003987979143857956, + 0.003935304936021566, + 0.0076268697157502174, + 0.0050201755948364735, + 0.00474033085629344, + 0.005693657323718071, + 0.006468809675425291, + 0.006881382316350937, + 0.003705376060679555, + 0.005115389358252287, + 0.005209572613239288, + 0.005939347669482231, + 0.0029496117495000362, + 0.003530371468514204, + 0.0094144893810153, + 0.005171502940356731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.016080567613244057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427905410528183 + ], + "opposite_logprob_tokens": [ + -2.0177905559539795 + ], + "logprob_reward": -0.1427905410528183, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427905410528183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04220457002520561, + 0.026630114763975143, + 0.028646433725953102, + 0.02445080131292343, + 0.03717494383454323, + 0.04989851266145706, + 0.6253007650375366, + 0.053005509078502655, + 0.08429773896932602, + 0.028390584513545036 + ], + "credit_map": [ + 0.0051538050174713135, + 0.006770679727196693, + 0.006694368086755276, + 0.003811499336734414, + 0.0051515959203243256, + 0.006333275232464075, + 0.003950356971472502, + 0.004338989965617657, + 0.0032107660081237555, + 0.0036013524513691664, + 0.0028534037992358208, + 0.003050839528441429, + 0.0021576634608209133, + 0.004775626584887505, + 0.00253239250741899, + 0.004448070656508207, + 0.003966004122048616, + 0.003287164494395256, + 0.003019633237272501, + 0.00351690873503685, + 0.005417606793344021, + 0.003509964793920517, + 0.0033333487808704376, + 0.0025958032347261906, + 0.002639530226588249, + 0.0024508971255272627, + 0.0034419819712638855, + 0.0029829854611307383, + 0.0019507423276081681, + 0.0036321424413472414, + 0.002844167873263359, + 0.004508352372795343, + 0.003054661676287651, + 0.00356707233004272, + 0.005960006732493639, + 0.004051961470395327, + 0.0035419596824795008, + 0.005516000557690859, + 0.005183660890907049, + 0.006299622356891632, + 0.005828174762427807, + 0.008168954402208328, + 0.007095653563737869, + 0.00471089081838727, + 0.004601506981998682, + 0.008429156616330147, + 0.004116402007639408, + 0.006947775837033987, + 0.06251271069049835, + 0.05467120185494423, + 0.1641005426645279, + 0.08405574411153793, + 0.05309789627790451, + 0.062169551849365234, + 0.04374629631638527, + 0.10094687342643738, + 0.007705680560320616, + 0.007696291897445917, + 0.0074093821458518505, + 0.00759067153558135, + 0.005869543179869652, + 0.006215839181095362, + 0.0037975118029862642, + 0.006720587145537138, + 0.00977568794041872, + 0.007240773178637028, + 0.013870563358068466, + 0.010405483655631542, + 0.009006423875689507, + 0.009767936542630196, + 0.012300072237849236, + 0.011930791661143303, + 0.0022884749341756105, + 0.0031554249580949545, + 0.005086028017103672, + 0.0055993241257965565, + 0.001797333243303001, + 0.002753338310867548, + 0.004427889361977577, + 0.003282772144302726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.1427905410528183 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst the crushing silence and darkness of the ten-thousand-meter-deep ocean, the camera pans at a slow, deliberate downward angle, piercing the dust-like marine snow suspended in the water, ultimately revealing a scene imbued with ancient power and quiet destruction. A massive mosasaur, the size of a mountain, its vast bulk seemingly blending into the boundless abyss. Only as the faint blue light passes over it does its skin, coated in smooth, moist, slate-gray scales, reflect a faint, oily texture. Its streamlined form remains almost perfectly still in the water, its massive tail fin swaying imperceptibly to maintain its hovering position. Before this prehistoric behemoth, an equally massive, ancient dragon egg emanates a faint blue glow. The light seeps from within the eggshell, casting intricate shadows across the surface, riven with ancient cracks and fossilized veins, like the slow beating of a heart. The Mosasaur's long, narrow jaws, imbued with destructive force, opened wide. With a silent yet powerful movement, its jaws, powerful enough to shatter steel, closed with a devastating, steady, and deliberate bite upon the upper half of the dragon egg. Without explosion or sound, in absolute silence, the massive, faintly blue-glowing shell of the dragon egg collapsed inward under the irresistible pressure, shattering. Countless crystal-like fragments silently peeled away, drifting slowly into the bottomless darkness like ghostly continental plates. The faint blue light within the egg flared violently as it was devoured, illuminating the terrifying structure deep within the Mosasaur's throat. Then, it died forever. The entire scene was enveloped in a hushed, profound atmosphere, a witness to creation and destruction.", + "seed": 14651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_6/index_3146_seed_14651_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08812659904360771, + "vlm_logprob": -0.08812659904360771, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.08812659904360771, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025845173746347427 + ], + "opposite_logprob_tokens": [ + -4.150845050811768 + ], + "logprob_reward": -0.025845173746347427, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.025845173746347427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.265829861164093, + 0.09980899095535278, + 0.11049789935350418, + 0.027363426983356476, + 0.0631832405924797, + 0.21738669276237488, + 0.0754154846072197, + 0.02194283902645111, + 0.05697668343782425, + 0.06159490346908569 + ], + "credit_map": [ + 0.016002073884010315, + 0.022493883967399597, + 0.025216607376933098, + 0.020092345774173737, + 0.07323969900608063, + 0.04075716435909271, + 0.014722737483680248, + 0.05330532416701317, + 0.007768336217850447, + 0.011419212445616722, + 0.011133107356727123, + 0.01607510820031166, + 0.018987419083714485, + 0.010491127148270607, + 0.012830924242734909, + 0.011103753000497818, + 0.012806002050638199, + 0.007199466228485107, + 0.010160778649151325, + 0.01147478073835373, + 0.020607812330126762, + 0.02583976276218891, + 0.005695043131709099, + 0.01671425811946392, + 0.002074265154078603, + 0.0024948487989604473, + 0.003164649708196521, + 0.003695762250572443, + 0.004953240510076284, + 0.004422406200319529, + 0.0033153847325593233, + 0.003242873353883624, + 0.006571067031472921, + 0.007725285366177559, + 0.006699100602418184, + 0.006890520919114351, + 0.006299707107245922, + 0.0072429850697517395, + 0.0148609708994627, + 0.00689360499382019, + 0.019604021683335304, + 0.019606849178671837, + 0.019304383546113968, + 0.021295858547091484, + 0.022759949788451195, + 0.03421027213335037, + 0.034150175750255585, + 0.04645517095923424, + 0.007425261195749044, + 0.006971288006752729, + 0.007558573968708515, + 0.012170744128525257, + 0.00833059474825859, + 0.007453741040080786, + 0.011903381906449795, + 0.013601900078356266, + 0.002030974952504039, + 0.0020200894214212894, + 0.003251579124480486, + 0.003804621286690235, + 0.0018604840151965618, + 0.0024523530155420303, + 0.0017360600177198648, + 0.004786677658557892, + 0.00424930127337575, + 0.0036283701192587614, + 0.008963612839579582, + 0.0084691122174263, + 0.006273450795561075, + 0.008688938803970814, + 0.009467393159866333, + 0.007236506324261427, + 0.006302793975919485, + 0.005142326932400465, + 0.006118010263890028, + 0.009357095696032047, + 0.00897139124572277, + 0.005811730865389109, + 0.01190098561346531, + 0.00799056701362133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.025845173746347427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307640105485916 + ], + "opposite_logprob_tokens": [ + -3.7733075618743896 + ], + "logprob_reward": -0.023307640105485916, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307640105485916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18011397123336792, + 0.14063982665538788, + 0.28228047490119934, + 0.03440253436565399, + 0.07498127967119217, + 0.07447664439678192, + 0.055462226271629333, + 0.04993226379156113, + 0.05407034978270531, + 0.05364035442471504 + ], + "credit_map": [ + 0.019270624965429306, + 0.022595243528485298, + 0.015170027501881123, + 0.010609138756990433, + 0.0493454709649086, + 0.024819830432534218, + 0.00861137080937624, + 0.029692279174923897, + 0.007955815643072128, + 0.020790163427591324, + 0.01593564823269844, + 0.031139489263296127, + 0.01788468100130558, + 0.013321767561137676, + 0.01928873360157013, + 0.01432354561984539, + 0.016835927963256836, + 0.01449088379740715, + 0.018175393342971802, + 0.030947217717766762, + 0.04102084040641785, + 0.08383896946907043, + 0.012612363323569298, + 0.0643589198589325, + 0.0030884176958352327, + 0.004365876317024231, + 0.004207998514175415, + 0.0063020274974405766, + 0.004527555778622627, + 0.00458716182038188, + 0.004051780793815851, + 0.0032717182766646147, + 0.008841494098305702, + 0.008468512445688248, + 0.007812848314642906, + 0.009081702679395676, + 0.00765843503177166, + 0.007886002771556377, + 0.01741156168282032, + 0.007820730097591877, + 0.008622842840850353, + 0.010039284825325012, + 0.007294340059161186, + 0.00633949413895607, + 0.007907711900770664, + 0.009707017801702023, + 0.013070870190858841, + 0.011495091952383518, + 0.006322865374386311, + 0.008243881165981293, + 0.006134187802672386, + 0.00808145571500063, + 0.0056304591707885265, + 0.00699292728677392, + 0.006689718458801508, + 0.007366735953837633, + 0.0034086222294718027, + 0.0027866011951118708, + 0.012480250559747219, + 0.0112201739102602, + 0.0027270186692476273, + 0.00431695906445384, + 0.004511432722210884, + 0.00848121102899313, + 0.003694182727485895, + 0.004021739587187767, + 0.00808680523186922, + 0.008873661048710346, + 0.006451468914747238, + 0.007881337776780128, + 0.008723561652004719, + 0.006337596569210291, + 0.004848816432058811, + 0.004393741022795439, + 0.005281402729451656, + 0.007083530072122812, + 0.008358289487659931, + 0.005792680662125349, + 0.010229266248643398, + 0.0076526314951479435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.023307640105485916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25217828154563904 + ], + "opposite_logprob_tokens": [ + -1.5021783113479614 + ], + "logprob_reward": -0.25217828154563904, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25217828154563904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22893422842025757, + 0.09017680585384369, + 0.18154804408550262, + 0.03680318966507912, + 0.07856027036905289, + 0.07360059022903442, + 0.032112255692481995, + 0.029555771499872208, + 0.03335144370794296, + 0.21535739302635193 + ], + "credit_map": [ + 0.015237999148666859, + 0.02107076160609722, + 0.020053088665008545, + 0.010919995605945587, + 0.07014892250299454, + 0.03552325442433357, + 0.009710622951388359, + 0.04626956582069397, + 0.005217570345848799, + 0.013841859996318817, + 0.009668109007179737, + 0.022203730419278145, + 0.011396978981792927, + 0.007941488176584244, + 0.010691826231777668, + 0.009215244092047215, + 0.00732708303257823, + 0.006890505086630583, + 0.010650165379047394, + 0.011291761882603168, + 0.023631110787391663, + 0.06399412453174591, + 0.0052330587059259415, + 0.05253024399280548, + 0.0034021891187876463, + 0.004092131741344929, + 0.00498793413862586, + 0.006546074524521828, + 0.005125599447637796, + 0.004690289497375488, + 0.004441372584551573, + 0.003517599543556571, + 0.00938809011131525, + 0.00998386088758707, + 0.008509509265422821, + 0.008886104449629784, + 0.007698314264416695, + 0.007813624106347561, + 0.018800921738147736, + 0.007479852065443993, + 0.009027850814163685, + 0.010344398207962513, + 0.006974300369620323, + 0.0057563940063118935, + 0.006636267527937889, + 0.010669397190213203, + 0.010296426713466644, + 0.013895557262003422, + 0.0021494380198419094, + 0.0021308406721800566, + 0.005620116833597422, + 0.003247619839385152, + 0.002428400795906782, + 0.004442843142896891, + 0.0025460454635322094, + 0.00954695139080286, + 0.0020183196756988764, + 0.0014567688340321183, + 0.008063782006502151, + 0.006441280245780945, + 0.0015042179729789495, + 0.002490607788786292, + 0.0023692266549915075, + 0.005211568903177977, + 0.004066580906510353, + 0.002742881653830409, + 0.0044412328861653805, + 0.005540984682738781, + 0.0031383419409394264, + 0.0029239647556096315, + 0.005720468237996101, + 0.00477698864415288, + 0.015293571166694164, + 0.013713461346924305, + 0.02749922126531601, + 0.0262894369661808, + 0.02056177332997322, + 0.022105425596237183, + 0.062362175434827805, + 0.027532311156392097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.25217828154563904 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026371043175458908 + ], + "opposite_logprob_tokens": [ + -3.6513710021972656 + ], + "logprob_reward": -0.026371043175458908, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371043175458908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19795700907707214, + 0.04729548469185829, + 0.2574719488620758, + 0.0472610704600811, + 0.05295968055725098, + 0.09416710585355759, + 0.025434600189328194, + 0.016066301614046097, + 0.07333658635616302, + 0.18805015087127686 + ], + "credit_map": [ + 0.012335450388491154, + 0.019793059676885605, + 0.01624966412782669, + 0.01213165931403637, + 0.05698106065392494, + 0.03276753053069115, + 0.009236298501491547, + 0.038462307304143906, + 0.003044751239940524, + 0.009462359361350536, + 0.004205829463899136, + 0.00813132245093584, + 0.006474492140114307, + 0.006621779408305883, + 0.005233988165855408, + 0.0041209678165614605, + 0.010049235075712204, + 0.007882020436227322, + 0.012975050136446953, + 0.010859228670597076, + 0.035909995436668396, + 0.09671517461538315, + 0.005515565164387226, + 0.07756571471691132, + 0.003810327500104904, + 0.0069641051813960075, + 0.006155937910079956, + 0.00872381217777729, + 0.005520361941307783, + 0.005124772898852825, + 0.0068353996612131596, + 0.004126354120671749, + 0.004833304323256016, + 0.007478902582079172, + 0.004749360028654337, + 0.005195191130042076, + 0.006096709985285997, + 0.005905673839151859, + 0.013283712789416313, + 0.005416829604655504, + 0.009558329358696938, + 0.011878502555191517, + 0.009191653691232204, + 0.008500095456838608, + 0.011454458348453045, + 0.013506818562746048, + 0.014714108780026436, + 0.015363143756985664, + 0.0024564850609749556, + 0.0025647913571447134, + 0.00495482049882412, + 0.0032324206549674273, + 0.0025051687844097614, + 0.0026028165593743324, + 0.003574285889044404, + 0.003543814644217491, + 0.0013408656232059002, + 0.0015210392884910107, + 0.0034674350172281265, + 0.003268999280408025, + 0.001211645663715899, + 0.002085138810798526, + 0.0013941675424575806, + 0.0017770114354789257, + 0.005964323878288269, + 0.005568996071815491, + 0.011522595770657063, + 0.01597796380519867, + 0.007130496669560671, + 0.00724881561473012, + 0.011709277518093586, + 0.008214126341044903, + 0.016995800659060478, + 0.013811323791742325, + 0.025180015712976456, + 0.02570476196706295, + 0.019408592954277992, + 0.01817321591079235, + 0.04682296887040138, + 0.021953482180833817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.026371043175458908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, deep ocean environment with marine snow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive mosasaur with slate-gray scales visible, hovering almost perfectly still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive dragon egg with ancient cracks and a faint blue glow visible before the bite?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mosasaur bite the upper half of the dragon egg, causing the shell to collapse inward without explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan at a slow, deliberate downward angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293085664510727 + ], + "opposite_logprob_tokens": [ + -2.2379307746887207 + ], + "logprob_reward": -0.11293085664510727, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293085664510727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0809069350361824, + 0.042107317596673965, + 0.19764764606952667, + 0.06981499493122101, + 0.04317515715956688, + 0.08101573586463928, + 0.09705294668674469, + 0.260043203830719, + 0.056789495050907135, + 0.07144656777381897 + ], + "credit_map": [ + 0.009701083414256573, + 0.008216316811740398, + 0.006833663675934076, + 0.010682846419513226, + 0.01753186248242855, + 0.013367335312068462, + 0.00565712433308363, + 0.008916705846786499, + 0.0036115162074565887, + 0.007752050645649433, + 0.004787077661603689, + 0.006067456211894751, + 0.004347394686192274, + 0.006265316158533096, + 0.004852441139519215, + 0.00442406116053462, + 0.009762807749211788, + 0.006390859372913837, + 0.012030295096337795, + 0.00830858200788498, + 0.026419350877404213, + 0.07191835343837738, + 0.00482381833717227, + 0.0579935647547245, + 0.004424496553838253, + 0.008534623309969902, + 0.00834200344979763, + 0.01077362336218357, + 0.009602289646863937, + 0.009052990935742855, + 0.008429398760199547, + 0.010655567049980164, + 0.0043000164441764355, + 0.0070730517618358135, + 0.004491605795919895, + 0.003997717052698135, + 0.007884442806243896, + 0.005959860049188137, + 0.005242494400590658, + 0.004225969314575195, + 0.007548712659627199, + 0.008657972328364849, + 0.00847616232931614, + 0.009020701982080936, + 0.01014417503029108, + 0.011533894576132298, + 0.014091565273702145, + 0.011542551219463348, + 0.006623611319810152, + 0.00698899757117033, + 0.011974737979471684, + 0.010948178358376026, + 0.010264909826219082, + 0.01180272176861763, + 0.01095678098499775, + 0.027493009343743324, + 0.01772269420325756, + 0.013532075099647045, + 0.08109548687934875, + 0.06843329221010208, + 0.012295141816139221, + 0.015641607344150543, + 0.019502660259604454, + 0.031820233911275864, + 0.005489106755703688, + 0.004786327946931124, + 0.008431795053184032, + 0.009873001836240292, + 0.005739965941756964, + 0.006597782485187054, + 0.008882947266101837, + 0.006988564971834421, + 0.004852747544646263, + 0.003909587860107422, + 0.008546240627765656, + 0.007448964286595583, + 0.005664434749633074, + 0.0067703030072152615, + 0.024493148550391197, + 0.009761137887835503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.11293085664510727 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_tensor([2001, 2001])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_tensor([2001, 2001])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ce4179cea84e1a4a3d184ecee8b4e9b66338e592 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_tensor([2001, 2001])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shaky, handheld, extreme close-up captures the face of a young woman riding a roller coaster at top speed. Her long, dark brown hair is a violent, whipping blur, flying straight back out of the frame. The immense force of the G-forces pulls the skin of her face taut, her mouth open in a silent scream of both terror and pure exhilaration. Her large, hazel eyes stare forward, tears from the speed and gusts of wind welling up in the corners before immediately curling back to her temples in thin, shimmering trails. The harsh, bright sunlight makes the moisture on her cheeks glisten, illuminating the subtle details of her skin. Behind her, the world is rendered in an indecipherable, violent motion blur—a chaotic vortex composed of streaks of blue in the sky, the green of distant trees, and the metallic gray of the coaster's structure. The entire frame hums and vibrates in sync with the chaotic motion of the ride, creating a visceral, overwhelming, and intensely immersive atmosphere of pure adrenaline and sensory overload.", + "seed": 17926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_2001_seed_17926_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10589697188697755, + "vlm_logprob": -0.10589697188697755, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.10589697188697755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016301434487104416 + ], + "opposite_logprob_tokens": [ + -4.89130163192749 + ], + "logprob_reward": -0.016301434487104416, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.016301434487104416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12875375151634216, + 0.06912275403738022, + 0.06972651928663254, + 0.10877291113138199, + 0.12620647251605988, + 0.05753641948103905, + 0.06355120241641998, + 0.0269670020788908, + 0.13428479433059692, + 0.21507816016674042 + ], + "credit_map": [ + 0.012538340874016285, + 0.010405081324279308, + 0.01220385730266571, + 0.013225837610661983, + 0.020068569108843803, + 0.011763962917029858, + 0.02789149433374405, + 0.02065659500658512, + 0.0063543012365698814, + 0.005761743988841772, + 0.011014354415237904, + 0.0076112402603030205, + 0.009545056149363518, + 0.008764650672674179, + 0.01326888706535101, + 0.006802517920732498, + 0.009824765846133232, + 0.006445270497351885, + 0.007130612153559923, + 0.005677615292370319, + 0.009117099456489086, + 0.00801814254373312, + 0.015098156407475471, + 0.0084148570895195, + 0.02224772237241268, + 0.013171850703656673, + 0.02356439270079136, + 0.006461120210587978, + 0.005542493425309658, + 0.01880592480301857, + 0.007751781493425369, + 0.011227619834244251, + 0.022279413416981697, + 0.015180542133748531, + 0.0158860981464386, + 0.009591360576450825, + 0.008006805554032326, + 0.0324590802192688, + 0.012495571747422218, + 0.010307596065104008, + 0.007064623758196831, + 0.008777153678238392, + 0.007552218157798052, + 0.006071101408451796, + 0.006266939453780651, + 0.008842249400913715, + 0.007890689186751842, + 0.005071442108601332, + 0.005937838461250067, + 0.003542977152392268, + 0.009037207812070847, + 0.007321092300117016, + 0.006742373574525118, + 0.009883062914013863, + 0.006831892300397158, + 0.01425476185977459, + 0.003180621424689889, + 0.0026030330918729305, + 0.004727298393845558, + 0.003959248773753643, + 0.003009391948580742, + 0.003337332746013999, + 0.0032862615771591663, + 0.002863815054297447, + 0.012998245656490326, + 0.009724531322717667, + 0.022602252662181854, + 0.013897574506700039, + 0.020588114857673645, + 0.013295311480760574, + 0.018592681735754013, + 0.022586075589060783, + 0.026796137914061546, + 0.012563396245241165, + 0.011992644518613815, + 0.03001236729323864, + 0.01898238994181156, + 0.07356835901737213, + 0.023046711459755898, + 0.018116168677806854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.016301434487104416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34870922565460205 + ], + "opposite_logprob_tokens": [ + -1.223709225654602 + ], + "logprob_reward": -0.34870922565460205, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34870922565460205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20712201297283173, + 0.05751311779022217, + 0.1896216720342636, + 0.29116764664649963, + 0.02688552625477314, + 0.04635769501328468, + 0.02268451265990734, + 0.03487464785575867, + 0.05060523375868797, + 0.07316792011260986 + ], + "credit_map": [ + 0.02108006365597248, + 0.024839717894792557, + 0.013165589421987534, + 0.019311336800456047, + 0.05933445692062378, + 0.017100870609283447, + 0.031126759946346283, + 0.021163221448659897, + 0.0059007517993450165, + 0.008692055940628052, + 0.01092336606234312, + 0.006204189267009497, + 0.007059643976390362, + 0.00496444758027792, + 0.008209708146750927, + 0.005558952689170837, + 0.029322851449251175, + 0.015481517650187016, + 0.020701024681329727, + 0.01734677143394947, + 0.028299791738390923, + 0.025212280452251434, + 0.03260578215122223, + 0.02065165340900421, + 0.06946124881505966, + 0.027589576318860054, + 0.056196875870227814, + 0.020782634615898132, + 0.011507325805723667, + 0.0632157251238823, + 0.014827638864517212, + 0.02758663520216942, + 0.004313116427510977, + 0.002983868820592761, + 0.00478005688637495, + 0.0016133291646838188, + 0.0016521821962669492, + 0.0065574501641094685, + 0.0017716351430863142, + 0.003213886171579361, + 0.004863904323428869, + 0.003539308672770858, + 0.004936183337122202, + 0.004360031336545944, + 0.005617368500679731, + 0.005229003261774778, + 0.013303332962095737, + 0.004508560989052057, + 0.0021959992591291666, + 0.0020951074548065662, + 0.0033749016001820564, + 0.0025720628909766674, + 0.0031154060270637274, + 0.002185676246881485, + 0.005040607415139675, + 0.002104751765727997, + 0.0032242443412542343, + 0.002795715816318989, + 0.00550590455532074, + 0.00312778539955616, + 0.004393427632749081, + 0.005506440997123718, + 0.005202305503189564, + 0.0051188236102461815, + 0.004331917967647314, + 0.0028104556258767843, + 0.013889445923268795, + 0.0051933652721345425, + 0.011695526540279388, + 0.0038769817911088467, + 0.005244480445981026, + 0.0035630573984235525, + 0.008859733119606972, + 0.0074747358448803425, + 0.012298990972340107, + 0.008437256328761578, + 0.006467544008046389, + 0.01875663548707962, + 0.006224288605153561, + 0.00464874180033803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.34870922565460205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299921572208405 + ], + "opposite_logprob_tokens": [ + -2.237999200820923 + ], + "logprob_reward": -0.11299921572208405, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299921572208405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16816894710063934, + 0.09380282461643219, + 0.290432333946228, + 0.08527616411447525, + 0.06756838411092758, + 0.02678951434791088, + 0.07933269441127777, + 0.0455697663128376, + 0.09636030346155167, + 0.04669908061623573 + ], + "credit_map": [ + 0.01744087226688862, + 0.01579069159924984, + 0.01803065836429596, + 0.013956515118479729, + 0.04787454009056091, + 0.013186108320951462, + 0.025779373943805695, + 0.016110194846987724, + 0.0074133193120360374, + 0.009185392409563065, + 0.01388139184564352, + 0.010524763725697994, + 0.01694430410861969, + 0.005909839179366827, + 0.021870842203497887, + 0.008072971366345882, + 0.03622080758213997, + 0.03183318302035332, + 0.02121378853917122, + 0.017299415543675423, + 0.07939658313989639, + 0.036308590322732925, + 0.04485306143760681, + 0.023306911811232567, + 0.012571807950735092, + 0.010963683016598225, + 0.015158009715378284, + 0.006576733198016882, + 0.006440459284931421, + 0.01315589901059866, + 0.009583882056176662, + 0.01082568522542715, + 0.008079329505562782, + 0.007290619891136885, + 0.010375763289630413, + 0.0043989671394228935, + 0.0034861755557358265, + 0.011533391661942005, + 0.007995102554559708, + 0.014409033581614494, + 0.003081708448007703, + 0.002381185768172145, + 0.003798713907599449, + 0.0020873635075986385, + 0.00337585574015975, + 0.0031932613346725702, + 0.005612203851342201, + 0.0032592201605439186, + 0.00776090519502759, + 0.004866229370236397, + 0.0119899557903409, + 0.007578812539577484, + 0.009522996842861176, + 0.010441794991493225, + 0.011685268022119999, + 0.015486731193959713, + 0.004795957822352648, + 0.004022406414151192, + 0.007292113266885281, + 0.005823327228426933, + 0.005794886965304613, + 0.005537338554859161, + 0.007336196023970842, + 0.0049675400368869305, + 0.008257712237536907, + 0.004747272469103336, + 0.029632676392793655, + 0.010084057226777077, + 0.02317574992775917, + 0.007263827603310347, + 0.008382018655538559, + 0.004816981498152018, + 0.004673983436077833, + 0.005476965103298426, + 0.004846124444156885, + 0.003981884103268385, + 0.008245901204645634, + 0.007130828686058521, + 0.008490625768899918, + 0.0038527671713382006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.11299921572208405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002825199393555522 + ], + "opposite_logprob_tokens": [ + -5.8778252601623535 + ], + "logprob_reward": -0.002825199393555522, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002825199393555522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18476304411888123, + 0.0659789964556694, + 0.13448897004127502, + 0.2426498681306839, + 0.09653127193450928, + 0.07973933964967728, + 0.033577948808670044, + 0.02463439665734768, + 0.06773446500301361, + 0.06990177184343338 + ], + "credit_map": [ + 0.019256500527262688, + 0.01857640966773033, + 0.020125282928347588, + 0.01746521145105362, + 0.041719406843185425, + 0.016494132578372955, + 0.029757020995020866, + 0.021369028836488724, + 0.006334106903523207, + 0.010264687240123749, + 0.013257249258458614, + 0.006940313149243593, + 0.0068909586407244205, + 0.0053687915205955505, + 0.009218839928507805, + 0.007704043760895729, + 0.025883231312036514, + 0.014580823481082916, + 0.011998316273093224, + 0.010855665430426598, + 0.021060386672616005, + 0.015289781615138054, + 0.020662987604737282, + 0.01415775902569294, + 0.03973967581987381, + 0.031220631673932076, + 0.06453686952590942, + 0.013763505034148693, + 0.013374782167375088, + 0.03293913975358009, + 0.014752882532775402, + 0.03232235834002495, + 0.014536631293594837, + 0.01945347711443901, + 0.021795617416501045, + 0.0072930180467665195, + 0.005708662793040276, + 0.016063548624515533, + 0.006524829193949699, + 0.005155467893928289, + 0.009377355687320232, + 0.007442764937877655, + 0.017975881695747375, + 0.005638234782963991, + 0.005853171460330486, + 0.012930801138281822, + 0.011264853179454803, + 0.009256263263523579, + 0.0048849559389054775, + 0.0024037521798163652, + 0.005528062116354704, + 0.004225755110383034, + 0.003417654661461711, + 0.004064879845827818, + 0.006071935873478651, + 0.002980944002047181, + 0.002554962644353509, + 0.0022846574429422617, + 0.0036351464223116636, + 0.004153165500611067, + 0.0028367273043841124, + 0.0028704337310045958, + 0.003551213536411524, + 0.0027480870485305786, + 0.005618862342089415, + 0.006244583521038294, + 0.01379155833274126, + 0.006807217840105295, + 0.014006483368575573, + 0.005390659905970097, + 0.008054912090301514, + 0.007820172235369682, + 0.009717363864183426, + 0.005992508493363857, + 0.005581684876233339, + 0.00954490713775158, + 0.0058613549917936325, + 0.022449420765042305, + 0.005237260367721319, + 0.00551725784316659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.002825199393555522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864978417754173 + ], + "opposite_logprob_tokens": [ + -3.048649787902832 + ], + "logprob_reward": -0.04864978417754173, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864978417754173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18965692818164825, + 0.13160595297813416, + 0.10833431035280228, + 0.19053007662296295, + 0.09459749609231949, + 0.06499041616916656, + 0.054261840879917145, + 0.02389853447675705, + 0.0750957503914833, + 0.06702876091003418 + ], + "credit_map": [ + 0.03052784875035286, + 0.01534468587487936, + 0.027522467076778412, + 0.0182095505297184, + 0.03528846800327301, + 0.020538145676255226, + 0.026572294533252716, + 0.01565343700349331, + 0.012015343643724918, + 0.01535639725625515, + 0.02765483595430851, + 0.012652089819312096, + 0.019991610199213028, + 0.009099585004150867, + 0.023020893335342407, + 0.011815186589956284, + 0.02190619707107544, + 0.010566098615527153, + 0.01837357133626938, + 0.007823352701961994, + 0.00932836253196001, + 0.015175948850810528, + 0.014394430443644524, + 0.010766337625682354, + 0.04258882999420166, + 0.020018339157104492, + 0.026903381571173668, + 0.016718268394470215, + 0.012908251956105232, + 0.0400385819375515, + 0.012683290988206863, + 0.018671123310923576, + 0.012012445367872715, + 0.007510134484618902, + 0.013714741915464401, + 0.008934860117733479, + 0.008077481761574745, + 0.011141769587993622, + 0.013900249265134335, + 0.019305799156427383, + 0.007057193201035261, + 0.007336705923080444, + 0.010583254508674145, + 0.005382678937166929, + 0.006424386985599995, + 0.011050226166844368, + 0.008402099832892418, + 0.008753865957260132, + 0.00645036855712533, + 0.0039873807691037655, + 0.009312351234257221, + 0.006676139310002327, + 0.006137952208518982, + 0.0059582083486020565, + 0.008442445658147335, + 0.007296986412256956, + 0.002766735153272748, + 0.002276706276461482, + 0.003937681671231985, + 0.0027703947853296995, + 0.0030888679903000593, + 0.002623817650601268, + 0.0030976843554526567, + 0.0033366431016474962, + 0.007293875329196453, + 0.005209681577980518, + 0.017219748347997665, + 0.006817949935793877, + 0.013015437871217728, + 0.005724451970309019, + 0.007905050180852413, + 0.01190954726189375, + 0.00831117294728756, + 0.006360921543091536, + 0.004240401554852724, + 0.006871711928397417, + 0.008846627548336983, + 0.013840166851878166, + 0.012386186979711056, + 0.00617156270891428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.04864978417754173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shaky, handheld, extreme close-up captures the face of a young woman riding a roller coaster at top speed. Her long, dark brown hair is a violent, whipping blur, flying straight back out of the frame. The immense force of the G-forces pulls the skin of her face taut, her mouth open in a silent scream of both terror and pure exhilaration. Her large, hazel eyes stare forward, tears from the speed and gusts of wind welling up in the corners before immediately curling back to her temples in thin, shimmering trails. The harsh, bright sunlight makes the moisture on her cheeks glisten, illuminating the subtle details of her skin. Behind her, the world is rendered in an indecipherable, violent motion blur—a chaotic vortex composed of streaks of blue in the sky, the green of distant trees, and the metallic gray of the coaster's structure. The entire frame hums and vibrates in sync with the chaotic motion of the ride, creating a visceral, overwhelming, and intensely immersive atmosphere of pure adrenaline and sensory overload.", + "seed": 17927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_2001_seed_17927_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06712024994194507, + "vlm_logprob": -0.06712024994194507, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.06712024994194507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017840275540947914 + ], + "opposite_logprob_tokens": [ + -5.017840385437012 + ], + "logprob_reward": -0.017840275540947914, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.017840275540947914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04150417074561119, + 0.10215391963720322, + 0.1231815367937088, + 0.06030130013823509, + 0.06031734496355057, + 0.4781278967857361, + 0.09431839734315872, + 0.00990744773298502, + 0.011052189394831657, + 0.019135799258947372 + ], + "credit_map": [ + 0.004907437600195408, + 0.005071789026260376, + 0.007271514739841223, + 0.006535011809319258, + 0.0036883980501443148, + 0.004984682891517878, + 0.006453304085880518, + 0.002592029981315136, + 0.012532742694020271, + 0.012314146384596825, + 0.03463824465870857, + 0.01042168214917183, + 0.014154854230582714, + 0.004456222523003817, + 0.007751910947263241, + 0.005884124897420406, + 0.01354169100522995, + 0.0203893780708313, + 0.019893260672688484, + 0.03323632851243019, + 0.010090046562254429, + 0.009797154925763607, + 0.01087635476142168, + 0.005357317626476288, + 0.004923091270029545, + 0.007041226141154766, + 0.004899901803582907, + 0.011549730785191059, + 0.012529726140201092, + 0.005115096922963858, + 0.008691880851984024, + 0.00555064994841814, + 0.006336885504424572, + 0.007821707054972649, + 0.006476738955825567, + 0.009372168220579624, + 0.0052938698790967464, + 0.008037625811994076, + 0.008498874492943287, + 0.008479475975036621, + 0.09442353993654251, + 0.028582941740751266, + 0.06194675713777542, + 0.02076583355665207, + 0.046528443694114685, + 0.021197598427534103, + 0.15855003893375397, + 0.046132709830999374, + 0.008584290742874146, + 0.007835707627236843, + 0.009933194145560265, + 0.013208117336034775, + 0.009297975338995457, + 0.02135080099105835, + 0.012649276293814182, + 0.011459031142294407, + 0.001141359331086278, + 0.0010383428307250142, + 0.0009870077483356, + 0.001194986398331821, + 0.0012925543123856187, + 0.0010984109248965979, + 0.0017254997510463, + 0.001429286552593112, + 0.0012851633364334702, + 0.0012092016404494643, + 0.0009544779313728213, + 0.001363215851597488, + 0.0008726670639589429, + 0.001484728418290615, + 0.001700482564046979, + 0.002182252239435911, + 0.001975654624402523, + 0.0023009032011032104, + 0.0021831742487847805, + 0.0027572729159146547, + 0.002737530041486025, + 0.001385070034302771, + 0.0041264500468969345, + 0.0016697439132258296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.017840275540947914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038841009140015 + ], + "opposite_logprob_tokens": [ + -1.910388469696045 + ], + "logprob_reward": -0.16038841009140015, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038841009140015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027226725593209267, + 0.1014503613114357, + 0.21266651153564453, + 0.15447981655597687, + 0.03352253884077072, + 0.32686302065849304, + 0.06658145785331726, + 0.0322507806122303, + 0.0121366111561656, + 0.03282216936349869 + ], + "credit_map": [ + 0.0032487495336681604, + 0.00387827237136662, + 0.003336128080263734, + 0.005879392847418785, + 0.0022490620613098145, + 0.00294818845577538, + 0.002487450372427702, + 0.003199480939656496, + 0.011751615442335606, + 0.012031185440719128, + 0.035781364887952805, + 0.010151905938982964, + 0.015752526000142097, + 0.004323551896959543, + 0.008130381815135479, + 0.003527829423546791, + 0.019345803186297417, + 0.033113859593868256, + 0.031043102964758873, + 0.06842344999313354, + 0.016382019966840744, + 0.012198437936604023, + 0.022734034806489944, + 0.009425795637071133, + 0.010364309884607792, + 0.014223767444491386, + 0.014070265926420689, + 0.024264294654130936, + 0.027209462597966194, + 0.023527998477220535, + 0.02924642153084278, + 0.01157329697161913, + 0.0033126643393188715, + 0.004251330159604549, + 0.004431514069437981, + 0.00511547364294529, + 0.0028127001132816076, + 0.004669342655688524, + 0.0044860546477139, + 0.004443458281457424, + 0.06505440175533295, + 0.01934388279914856, + 0.04123850539326668, + 0.01385320071130991, + 0.03162912651896477, + 0.01515512727200985, + 0.1081833764910698, + 0.03240540251135826, + 0.005683050025254488, + 0.005383461248129606, + 0.006683663930743933, + 0.008023920468986034, + 0.007187118753790855, + 0.016972461715340614, + 0.008871927857398987, + 0.00777584919705987, + 0.0043701184913516045, + 0.0027443645521998405, + 0.0043631866574287415, + 0.0030789461452513933, + 0.003029433311894536, + 0.003526116255670786, + 0.008013462647795677, + 0.00312515115365386, + 0.0012723096879199147, + 0.0018688339041545987, + 0.0013056037714704871, + 0.002092357026413083, + 0.0010703044245019555, + 0.0010597376385703683, + 0.0018723534885793924, + 0.0015951116802170873, + 0.0032622474245727062, + 0.003773170057684183, + 0.0036102188751101494, + 0.006397138349711895, + 0.0039235688745975494, + 0.002470012055709958, + 0.006935198791325092, + 0.002450615167617798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.16038841009140015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298313736915588 + ], + "opposite_logprob_tokens": [ + -2.237983226776123 + ], + "logprob_reward": -0.11298313736915588, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298313736915588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11663834750652313, + 0.23552991449832916, + 0.25760525465011597, + 0.061644136905670166, + 0.03449569642543793, + 0.1346808224916458, + 0.04039553552865982, + 0.03137267380952835, + 0.04056790843605995, + 0.04706971347332001 + ], + "credit_map": [ + 0.01456262357532978, + 0.016960877925157547, + 0.013321825303137302, + 0.021636387333273888, + 0.010047251358628273, + 0.010972103103995323, + 0.013678393326699734, + 0.015458879992365837, + 0.024036427959799767, + 0.024430157616734505, + 0.07729977369308472, + 0.020408209413290024, + 0.050164319574832916, + 0.00911941658705473, + 0.023163272067904472, + 0.006908338516950607, + 0.019845405593514442, + 0.04554319009184837, + 0.03625078871846199, + 0.09934341162443161, + 0.008950090035796165, + 0.008635246194899082, + 0.030641430988907814, + 0.008395700715482235, + 0.008721434511244297, + 0.005789533723145723, + 0.008240877650678158, + 0.011332059279084206, + 0.005221056751906872, + 0.00844928715378046, + 0.008885613642632961, + 0.005004274658858776, + 0.0032857602927833796, + 0.005121014080941677, + 0.005159728229045868, + 0.006819605827331543, + 0.0028866399079561234, + 0.003125458722934127, + 0.004409587010741234, + 0.0036879039835184813, + 0.024229589849710464, + 0.008033158257603645, + 0.01944074034690857, + 0.006643855944275856, + 0.012972742319107056, + 0.009523482993245125, + 0.04140962287783623, + 0.01242762990295887, + 0.003624286502599716, + 0.0033266630489379168, + 0.0031014878768473864, + 0.005891223903745413, + 0.007475259713828564, + 0.009972927160561085, + 0.003801496233791113, + 0.0032021936494857073, + 0.0030052494257688522, + 0.0038057505153119564, + 0.004669304005801678, + 0.005098438821732998, + 0.0035928396973758936, + 0.003203749191015959, + 0.003829639870673418, + 0.0041677034460008144, + 0.0055040279403328896, + 0.004525494761765003, + 0.003896046429872513, + 0.004533574916422367, + 0.002983826445415616, + 0.0058090584352612495, + 0.004550741519778967, + 0.008765137754380703, + 0.007149507291615009, + 0.004183634649962187, + 0.004120790399610996, + 0.005783298518508673, + 0.005459398031234741, + 0.005275418981909752, + 0.0116462092846632, + 0.003451453521847725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.11298313736915588 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013397298753261566 + ], + "opposite_logprob_tokens": [ + -6.626339912414551 + ], + "logprob_reward": -0.0013397298753261566, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013397298753261566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04238145053386688, + 0.11881238967180252, + 0.07337073981761932, + 0.14271992444992065, + 0.07042676210403442, + 0.40452274680137634, + 0.09320691227912903, + 0.021753327921032906, + 0.017672043293714523, + 0.015133745037019253 + ], + "credit_map": [ + 0.005369595717638731, + 0.005370515864342451, + 0.005098148714751005, + 0.006772086024284363, + 0.003376545151695609, + 0.004622026812285185, + 0.004626292269676924, + 0.007146234158426523, + 0.014909563586115837, + 0.016105202957987785, + 0.0367252379655838, + 0.013092085719108582, + 0.013576820492744446, + 0.006766674108803272, + 0.008871478028595448, + 0.00876531470566988, + 0.012512454763054848, + 0.008364616893231869, + 0.016195714473724365, + 0.012366642244160175, + 0.006899471860378981, + 0.00669133523479104, + 0.005939190275967121, + 0.004401294980198145, + 0.01012401282787323, + 0.015215695835649967, + 0.009509066119790077, + 0.024407904595136642, + 0.02958662621676922, + 0.01572289876639843, + 0.023681532591581345, + 0.014472168870270252, + 0.007799771148711443, + 0.008520254865288734, + 0.008994201198220253, + 0.01147215161472559, + 0.005915968678891659, + 0.008649874478578568, + 0.009912466630339622, + 0.009162061847746372, + 0.07922251522541046, + 0.02466203272342682, + 0.05092333257198334, + 0.020169008523225784, + 0.039542242884635925, + 0.02000780589878559, + 0.1311650425195694, + 0.0388307049870491, + 0.00821713451296091, + 0.006997090298682451, + 0.010317637585103512, + 0.010280693881213665, + 0.010373103432357311, + 0.025459084659814835, + 0.011400810442864895, + 0.010161342099308968, + 0.0027271611616015434, + 0.0016329160425812006, + 0.0032170480117201805, + 0.0021474245004355907, + 0.0019402410835027695, + 0.002716126386076212, + 0.0036196517758071423, + 0.003752754768356681, + 0.0017543996218591928, + 0.0017575767124071717, + 0.0017014833865687251, + 0.0025361713487654924, + 0.0017145586898550391, + 0.0015017330879345536, + 0.002286410890519619, + 0.004419705830514431, + 0.002571521792560816, + 0.0012779217213392258, + 0.0016829149099066854, + 0.0016449709655717015, + 0.0015959698939695954, + 0.0022484033834189177, + 0.0016913479194045067, + 0.002420690841972828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.0013397298753261566 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304969683289528 + ], + "opposite_logprob_tokens": [ + -3.1680498123168945 + ], + "logprob_reward": -0.04304969683289528, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304969683289528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03870150074362755, + 0.057522937655448914, + 0.46739625930786133, + 0.10799627751111984, + 0.034961145371198654, + 0.20008952915668488, + 0.04539339244365692, + 0.013484359718859196, + 0.02174612134695053, + 0.012708412483334541 + ], + "credit_map": [ + 0.005442862398922443, + 0.00488243205472827, + 0.0039038504473865032, + 0.006544903852045536, + 0.0034863499458879232, + 0.0054065813310444355, + 0.005363689735531807, + 0.003670831909403205, + 0.0060467529110610485, + 0.008636919781565666, + 0.018548892810940742, + 0.00525794830173254, + 0.00691660400480032, + 0.0032660800497978926, + 0.005465532653033733, + 0.003384206909686327, + 0.05218113958835602, + 0.0771036222577095, + 0.06785332411527634, + 0.10286255925893784, + 0.05547858402132988, + 0.04244577884674072, + 0.04189116880297661, + 0.027580086141824722, + 0.007920092903077602, + 0.011809163726866245, + 0.008031435310840607, + 0.01856168545782566, + 0.023590432479977608, + 0.010562852025032043, + 0.01744864694774151, + 0.010071976110339165, + 0.003822552040219307, + 0.003917363937944174, + 0.004357316065579653, + 0.005606713704764843, + 0.0028322830330580473, + 0.005245878826826811, + 0.005226681474596262, + 0.003952356521040201, + 0.03889065608382225, + 0.012179470621049404, + 0.02558775618672371, + 0.009700175374746323, + 0.01996167004108429, + 0.008623656816780567, + 0.06580879539251328, + 0.019337352365255356, + 0.004946304950863123, + 0.003812737064436078, + 0.005640052258968353, + 0.0072747087106108665, + 0.004362206906080246, + 0.007098947186022997, + 0.006552267353981733, + 0.005706171039491892, + 0.0011579339625313878, + 0.0019298545084893703, + 0.0017670694505795836, + 0.0018533653346821666, + 0.001637049368582666, + 0.0014851029263809323, + 0.0015480038709938526, + 0.002105980645865202, + 0.0020958029199391603, + 0.0024680697824805975, + 0.002024872461333871, + 0.003387793665751815, + 0.0023035616613924503, + 0.002033631782978773, + 0.0030035923700779676, + 0.004428796004503965, + 0.0014561937423422933, + 0.0012279225047677755, + 0.0013981469674035907, + 0.0016685775481164455, + 0.001354859210550785, + 0.0014859260991215706, + 0.0022149821743369102, + 0.0019018045859411359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.04304969683289528 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shaky, handheld, extreme close-up captures the face of a young woman riding a roller coaster at top speed. Her long, dark brown hair is a violent, whipping blur, flying straight back out of the frame. The immense force of the G-forces pulls the skin of her face taut, her mouth open in a silent scream of both terror and pure exhilaration. Her large, hazel eyes stare forward, tears from the speed and gusts of wind welling up in the corners before immediately curling back to her temples in thin, shimmering trails. The harsh, bright sunlight makes the moisture on her cheeks glisten, illuminating the subtle details of her skin. Behind her, the world is rendered in an indecipherable, violent motion blur—a chaotic vortex composed of streaks of blue in the sky, the green of distant trees, and the metallic gray of the coaster's structure. The entire frame hums and vibrates in sync with the chaotic motion of the ride, creating a visceral, overwhelming, and intensely immersive atmosphere of pure adrenaline and sensory overload.", + "seed": 17928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_2001_seed_17928_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.239283501310274, + "vlm_logprob": -0.239283501310274, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.239283501310274, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020998427644371986 + ], + "opposite_logprob_tokens": [ + -5.020998477935791 + ], + "logprob_reward": -0.020998427644371986, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.020998427644371986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6083426475524902, + 0.03333771601319313, + 0.05804835632443428, + 0.06758153438568115, + 0.05423080548644066, + 0.026581184938549995, + 0.027667347341775894, + 0.07827255874872208, + 0.015986289829015732, + 0.02995148114860058 + ], + "credit_map": [ + 0.03167557343840599, + 0.016857502982020378, + 0.023544378578662872, + 0.017216859385371208, + 0.1467178612947464, + 0.24178676307201385, + 0.025603605434298515, + 0.10494016110897064, + 0.003938579000532627, + 0.004772710148245096, + 0.004433918744325638, + 0.005360347684472799, + 0.003464221488684416, + 0.003883116180077195, + 0.0038762963376939297, + 0.0036085285246372223, + 0.007332182954996824, + 0.006156248971819878, + 0.004657939076423645, + 0.005456998944282532, + 0.0035802689380943775, + 0.008947808295488358, + 0.006455924827605486, + 0.015460989437997341, + 0.00476227980107069, + 0.017277520149946213, + 0.004710949957370758, + 0.010069859214127064, + 0.0027060098946094513, + 0.007819212973117828, + 0.004938832484185696, + 0.015296880155801773, + 0.004985084291547537, + 0.0029213684611022472, + 0.015882005915045738, + 0.004711502231657505, + 0.0030445116572082043, + 0.011784336529672146, + 0.006697798613458872, + 0.004204204771667719, + 0.003128856187686324, + 0.0026487503200769424, + 0.0031367102637887, + 0.002580570289865136, + 0.0015881048748269677, + 0.005146565847098827, + 0.003259653691202402, + 0.0050919754430651665, + 0.002699196105822921, + 0.004377619829028845, + 0.0037669665180146694, + 0.004623389802873135, + 0.0018187185050919652, + 0.003816031850874424, + 0.003336509922519326, + 0.003228917485103011, + 0.006119492929428816, + 0.011146768927574158, + 0.018845492973923683, + 0.010760065168142319, + 0.00659059127792716, + 0.007601420395076275, + 0.006601375062018633, + 0.010607360862195492, + 0.0016186811262741685, + 0.0021080723963677883, + 0.001723216730169952, + 0.002966467756778002, + 0.0016242606798186898, + 0.002142677316442132, + 0.001953764585778117, + 0.0018491508672013879, + 0.003783678635954857, + 0.0032462654635310173, + 0.00688739912584424, + 0.0029685809276998043, + 0.0020519515965133905, + 0.0042823064140975475, + 0.0031798342242836952, + 0.003551470348611474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.020998427644371986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8984073400497437 + ], + "opposite_logprob_tokens": [ + -0.5234073400497437 + ], + "logprob_reward": -0.8984073400497437, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8984073400497437, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.35521239042282104, + 0.05581556260585785, + 0.04075596109032631, + 0.12823368608951569, + 0.15653133392333984, + 0.05225053057074547, + 0.04448918253183365, + 0.08826705068349838, + 0.043930817395448685, + 0.03451347351074219 + ], + "credit_map": [ + 0.019707370549440384, + 0.014740637503564358, + 0.021270116791129112, + 0.015863021835684776, + 0.13233213126659393, + 0.09935379028320312, + 0.025888964533805847, + 0.026056380942463875, + 0.008248125202953815, + 0.007481906097382307, + 0.005953852087259293, + 0.008086653426289558, + 0.003921326715499163, + 0.008945330046117306, + 0.005134761333465576, + 0.008043605834245682, + 0.004038180224597454, + 0.005545022897422314, + 0.003148849355056882, + 0.006345985922962427, + 0.0030273646116256714, + 0.006664846558123827, + 0.005088622681796551, + 0.006897089071571827, + 0.008310952223837376, + 0.039520327001810074, + 0.005557602271437645, + 0.01370503380894661, + 0.006644472945481539, + 0.014409514144062996, + 0.008026489987969398, + 0.03205929324030876, + 0.014756212942302227, + 0.007863541133701801, + 0.047683294862508774, + 0.007335963658988476, + 0.008173997513949871, + 0.04110674932599068, + 0.01854228787124157, + 0.011069276370108128, + 0.009104294702410698, + 0.004171745385974646, + 0.004364832304418087, + 0.00508540216833353, + 0.0029839188791811466, + 0.005353592801839113, + 0.005114979110658169, + 0.016071762889623642, + 0.004732533358037472, + 0.005840100813657045, + 0.005844101309776306, + 0.008483861573040485, + 0.004011635202914476, + 0.004998821299523115, + 0.005380213260650635, + 0.005197914317250252, + 0.006297854240983725, + 0.008961528539657593, + 0.02564743161201477, + 0.011215395294129848, + 0.005467616952955723, + 0.010004683397710323, + 0.007502473890781403, + 0.013170073740184307, + 0.0027871818747371435, + 0.0052440776489675045, + 0.003809448564425111, + 0.008969173766672611, + 0.0034283760469406843, + 0.004755578935146332, + 0.006670401897281408, + 0.00826658122241497, + 0.0032008627895265818, + 0.0032446549739688635, + 0.004642888437956572, + 0.006559643428772688, + 0.002605883404612541, + 0.005192659795284271, + 0.004269961267709732, + 0.0047969208098948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.8984073400497437 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16044791042804718 + ], + "opposite_logprob_tokens": [ + -1.9104479551315308 + ], + "logprob_reward": -0.16044791042804718, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16044791042804718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2288787066936493, + 0.07003216445446014, + 0.03850635513663292, + 0.10015924274921417, + 0.17744070291519165, + 0.0689925029873848, + 0.10872548073530197, + 0.157695934176445, + 0.034654662013053894, + 0.014914198778569698 + ], + "credit_map": [ + 0.018760675564408302, + 0.011786033399403095, + 0.013096341863274574, + 0.02685089036822319, + 0.07191450893878937, + 0.019442658871412277, + 0.028293607756495476, + 0.03873397782444954, + 0.005136619787663221, + 0.009791220538318157, + 0.008582848124206066, + 0.014423524960875511, + 0.007398733403533697, + 0.0076587083749473095, + 0.00868687778711319, + 0.008353634737432003, + 0.004823378752917051, + 0.004482913296669722, + 0.0028845462948083878, + 0.009080277755856514, + 0.002741281408816576, + 0.005044391844421625, + 0.003781234845519066, + 0.005668334197252989, + 0.007703022100031376, + 0.024293260648846626, + 0.009319272823631763, + 0.014583948068320751, + 0.008712954819202423, + 0.011841366067528725, + 0.007061788812279701, + 0.016643626615405083, + 0.015437424182891846, + 0.006851144600659609, + 0.06203248351812363, + 0.008401291444897652, + 0.00874096155166626, + 0.04378797858953476, + 0.022719521075487137, + 0.009469909593462944, + 0.010381444357335567, + 0.007187685929238796, + 0.005377107299864292, + 0.0059226443991065025, + 0.004669420421123505, + 0.010747887194156647, + 0.008314680308103561, + 0.016391640529036522, + 0.01201891154050827, + 0.01579660177230835, + 0.010996238328516483, + 0.02550322562456131, + 0.006758399773389101, + 0.010417588986456394, + 0.0162473376840353, + 0.010987168177962303, + 0.01351824589073658, + 0.014208154752850533, + 0.051383089274168015, + 0.01396271213889122, + 0.008360907435417175, + 0.023162690922617912, + 0.011524422094225883, + 0.02157571353018284, + 0.0041808211244642735, + 0.004673975054174662, + 0.003441903041675687, + 0.006920401938259602, + 0.002059773774817586, + 0.004127515014261007, + 0.004384417086839676, + 0.004865855444222689, + 0.001610026229172945, + 0.0022507626563310623, + 0.0016800094163045287, + 0.0025742617435753345, + 0.0012602144852280617, + 0.0020447038114070892, + 0.0017760624177753925, + 0.0017181581351906061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.16044791042804718 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036267952527850866 + ], + "opposite_logprob_tokens": [ + -5.628626823425293 + ], + "logprob_reward": -0.0036267952527850866, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036267952527850866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.376413494348526, + 0.05829405412077904, + 0.06690221279859543, + 0.10950423032045364, + 0.13181956112384796, + 0.06107460707426071, + 0.04358942434191704, + 0.09231942147016525, + 0.03939707577228546, + 0.02068590559065342 + ], + "credit_map": [ + 0.01884564943611622, + 0.011960314586758614, + 0.014466460794210434, + 0.016632024198770523, + 0.12138982862234116, + 0.12647652626037598, + 0.02436096780002117, + 0.04228173568844795, + 0.009686586447060108, + 0.006823695730417967, + 0.006686955224722624, + 0.007144033908843994, + 0.003813554532825947, + 0.010809395462274551, + 0.005287984851747751, + 0.008041851222515106, + 0.007408429402858019, + 0.005320187658071518, + 0.0057937027886509895, + 0.005846342071890831, + 0.003969680052250624, + 0.016242219135165215, + 0.00578599376603961, + 0.01653565838932991, + 0.008478754200041294, + 0.032346196472644806, + 0.005377068184316158, + 0.009917015209794044, + 0.005843737628310919, + 0.014435110613703728, + 0.0056467377580702305, + 0.027459608390927315, + 0.011564968153834343, + 0.004719081334769726, + 0.04558226838707924, + 0.0057135336101055145, + 0.006765858270227909, + 0.032381635159254074, + 0.016578586772084236, + 0.008513638749718666, + 0.011844713240861893, + 0.005148714408278465, + 0.004535739775747061, + 0.004197172354906797, + 0.003373282728716731, + 0.007154364604502916, + 0.005377790424972773, + 0.019442830234766006, + 0.0048716748133301735, + 0.005978500470519066, + 0.005980537738651037, + 0.007267898879945278, + 0.003652796149253845, + 0.0045214626006782055, + 0.005490815732628107, + 0.005825739819556475, + 0.008135226555168629, + 0.008733476512134075, + 0.02818892151117325, + 0.010694039054214954, + 0.004874496720731258, + 0.010568485595285892, + 0.00664312532171607, + 0.014481655322015285, + 0.0032470538280904293, + 0.0047154780477285385, + 0.0038393200375139713, + 0.005050652660429478, + 0.0030064762104302645, + 0.009037171490490437, + 0.004894999787211418, + 0.0056059230118989944, + 0.002330722287297249, + 0.001836561132222414, + 0.003321038093417883, + 0.0033202965278178453, + 0.0013726534089073539, + 0.0036076903343200684, + 0.002059764927253127, + 0.00283717829734087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0036267952527850866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293703317642212 + ], + "opposite_logprob_tokens": [ + -2.2379369735717773 + ], + "logprob_reward": -0.11293703317642212, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293703317642212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6054753065109253, + 0.03475509211421013, + 0.0734303891658783, + 0.05111350491642952, + 0.052389487624168396, + 0.04527903348207474, + 0.028021126985549927, + 0.05256921797990799, + 0.036650482565164566, + 0.02031644992530346 + ], + "credit_map": [ + 0.04182375967502594, + 0.026989683508872986, + 0.050817493349313736, + 0.03657569736242294, + 0.12212138622999191, + 0.1850140243768692, + 0.03061412274837494, + 0.11151904612779617, + 0.004689171444624662, + 0.0052061122842133045, + 0.003890829626470804, + 0.005758211016654968, + 0.002469731727614999, + 0.005473937839269638, + 0.0041106087155640125, + 0.00315648689866066, + 0.0071881795302033424, + 0.010433315299451351, + 0.005501854233443737, + 0.012273112311959267, + 0.004394894000142813, + 0.012116099707782269, + 0.009915766306221485, + 0.011607157066464424, + 0.005389377009123564, + 0.008473284542560577, + 0.004915039520710707, + 0.007621145341545343, + 0.0033980468288064003, + 0.007802903186529875, + 0.004861592780798674, + 0.008652105927467346, + 0.0046861059963703156, + 0.0027298247441649437, + 0.016918445006012917, + 0.003099108347669244, + 0.0030433330684900284, + 0.011970814317464828, + 0.006293031387031078, + 0.003648818004876375, + 0.008877655491232872, + 0.003653548425063491, + 0.003992942627519369, + 0.0030621325131505728, + 0.0024490002542734146, + 0.004523055162280798, + 0.003909644670784473, + 0.01481105200946331, + 0.0035704278852790594, + 0.004255326464772224, + 0.0036957035772502422, + 0.004228580743074417, + 0.0018550824606791139, + 0.0036296378821134567, + 0.002937099663540721, + 0.0038492660969495773, + 0.004312832374125719, + 0.00592850474640727, + 0.010620535351336002, + 0.006716378964483738, + 0.00377890607342124, + 0.009993039071559906, + 0.005351821891963482, + 0.0058671943843364716, + 0.002901786006987095, + 0.004887021146714687, + 0.0032858997583389282, + 0.005178849678486586, + 0.0032094474881887436, + 0.004443007986992598, + 0.005450092721730471, + 0.00729437405243516, + 0.0018948545912280679, + 0.0030541943851858377, + 0.0023158034309744835, + 0.0031932767014950514, + 0.0018189051188528538, + 0.003222354454919696, + 0.001804168219678104, + 0.0030128909274935722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.11293703317642212 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shaky, handheld, extreme close-up captures the face of a young woman riding a roller coaster at top speed. Her long, dark brown hair is a violent, whipping blur, flying straight back out of the frame. The immense force of the G-forces pulls the skin of her face taut, her mouth open in a silent scream of both terror and pure exhilaration. Her large, hazel eyes stare forward, tears from the speed and gusts of wind welling up in the corners before immediately curling back to her temples in thin, shimmering trails. The harsh, bright sunlight makes the moisture on her cheeks glisten, illuminating the subtle details of her skin. Behind her, the world is rendered in an indecipherable, violent motion blur—a chaotic vortex composed of streaks of blue in the sky, the green of distant trees, and the metallic gray of the coaster's structure. The entire frame hums and vibrates in sync with the chaotic motion of the ride, creating a visceral, overwhelming, and intensely immersive atmosphere of pure adrenaline and sensory overload.", + "seed": 17929, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_7/index_2001_seed_17929_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17869421802461147, + "vlm_logprob": -0.17869421802461147, + "vlm_yes_no_margin": 3.250000023841858, + "vlm_token_credit": -0.17869421802461147, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009971448220312595 + ], + "opposite_logprob_tokens": [ + -5.759971618652344 + ], + "logprob_reward": -0.009971448220312595, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009971448220312595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06962408125400543, + 0.02298342064023018, + 0.3926257789134979, + 0.10096143186092377, + 0.21740466356277466, + 0.017338652163743973, + 0.020883074030280113, + 0.032926905900239944, + 0.10187066346406937, + 0.02338138036429882 + ], + "credit_map": [ + 0.008328713476657867, + 0.005200647283345461, + 0.009208230301737785, + 0.00791747123003006, + 0.005472519900649786, + 0.006992363836616278, + 0.009055386297404766, + 0.017448734492063522, + 0.002410931047052145, + 0.002510856371372938, + 0.002948296954855323, + 0.0031589381396770477, + 0.0023924573324620724, + 0.003406331641599536, + 0.0021499490831047297, + 0.004005656111985445, + 0.020741214975714684, + 0.11427448689937592, + 0.04110632464289665, + 0.02662499248981476, + 0.022578060626983643, + 0.11325099319219589, + 0.020588185638189316, + 0.033461470156908035, + 0.009918628260493279, + 0.007277622353285551, + 0.014712830074131489, + 0.013563931919634342, + 0.006674528121948242, + 0.016943443566560745, + 0.009574255906045437, + 0.022296175360679626, + 0.027556195855140686, + 0.012514196336269379, + 0.06453488767147064, + 0.02869250811636448, + 0.023158129304647446, + 0.020679060369729996, + 0.027987757697701454, + 0.012281890958547592, + 0.0020837232004851103, + 0.0019417601870372891, + 0.0018431547796353698, + 0.002148974919691682, + 0.0019720210693776608, + 0.003568091429769993, + 0.0015390472253784537, + 0.002241877606138587, + 0.0034868575166910887, + 0.0026396054308861494, + 0.0013536646729335189, + 0.0021592509001493454, + 0.001965811476111412, + 0.0022095455788075924, + 0.0013735853135585785, + 0.00569474883377552, + 0.006214473862200975, + 0.00354554015211761, + 0.0024587842635810375, + 0.0036607058718800545, + 0.007891535758972168, + 0.005183449946343899, + 0.001923738163895905, + 0.0020486728753894567, + 0.005356539972126484, + 0.04022538289427757, + 0.020777709782123566, + 0.016816675662994385, + 0.0076337591744959354, + 0.004422977101057768, + 0.0036716004833579063, + 0.002966000698506832, + 0.003983346279710531, + 0.0026458061765879393, + 0.0032372339628636837, + 0.004276772495359182, + 0.002356728073209524, + 0.0028701978735625744, + 0.002236064989119768, + 0.0017752267885953188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.009971448220312595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934357285499573 + ], + "opposite_logprob_tokens": [ + -0.6934357285499573 + ], + "logprob_reward": -0.6934357285499573, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934357285499573, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.018550392240285873, + 0.028010930866003036, + 0.4030563235282898, + 0.04222945123910904, + 0.27578091621398926, + 0.03127517178654671, + 0.0286425668746233, + 0.017417896538972855, + 0.11511220782995224, + 0.03992418944835663 + ], + "credit_map": [ + 0.001754465512931347, + 0.001919260248541832, + 0.002044794149696827, + 0.0032781516201794147, + 0.002606197725981474, + 0.0015592888230457902, + 0.0023320107720792294, + 0.003056223038583994, + 0.004352665971964598, + 0.0021237218752503395, + 0.003985158633440733, + 0.00288874888792634, + 0.002376782475039363, + 0.006953422911465168, + 0.0024307353887706995, + 0.0028996961191296577, + 0.021787095814943314, + 0.11131144315004349, + 0.042983416467905045, + 0.024555852636694908, + 0.02267611213028431, + 0.12347010523080826, + 0.02039690501987934, + 0.03587540239095688, + 0.0033652347046881914, + 0.0030555534176528454, + 0.006318867672234774, + 0.005836791358888149, + 0.0024353733751922846, + 0.008583728224039078, + 0.003900497453287244, + 0.008733408525586128, + 0.032609403133392334, + 0.015002934262156487, + 0.07910432666540146, + 0.04412822797894478, + 0.028004763647913933, + 0.02529815398156643, + 0.03719402104616165, + 0.01443908829241991, + 0.003777672303840518, + 0.00249480246566236, + 0.003797354642301798, + 0.0031735380180180073, + 0.0027915092650800943, + 0.008547910489141941, + 0.0025410871021449566, + 0.004151300992816687, + 0.004863095004111528, + 0.003937683999538422, + 0.0018282168311998248, + 0.003615459194406867, + 0.0021737487986683846, + 0.005888125393539667, + 0.0019736133981496096, + 0.004362623207271099, + 0.002413151552900672, + 0.002379553159698844, + 0.003072936087846756, + 0.002913776319473982, + 0.001439014682546258, + 0.0024639288894832134, + 0.0013927307445555925, + 0.0013428048696368933, + 0.006583258509635925, + 0.044316116720438004, + 0.03290531411767006, + 0.014907820150256157, + 0.007231751922518015, + 0.004713606089353561, + 0.001830938272178173, + 0.0026234015822410583, + 0.0053136409260332584, + 0.007852666079998016, + 0.005050063598901033, + 0.003944517578929663, + 0.0035716011188924313, + 0.00801275297999382, + 0.0016441222978755832, + 0.004534828010946512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.6934357285499573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14289966225624084 + ], + "opposite_logprob_tokens": [ + -2.017899751663208 + ], + "logprob_reward": -0.14289966225624084, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14289966225624084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04159385338425636, + 0.032931648194789886, + 0.4425860047340393, + 0.06114828586578369, + 0.22315706312656403, + 0.007907111197710037, + 0.045913323760032654, + 0.037091176956892014, + 0.09766639769077301, + 0.010005121119320393 + ], + "credit_map": [ + 0.0056220367550849915, + 0.0029317475855350494, + 0.007197913248091936, + 0.005879401229321957, + 0.0030166886281222105, + 0.004654062911868095, + 0.004819285590201616, + 0.0074727218598127365, + 0.004407898057252169, + 0.003610110841691494, + 0.004596053157001734, + 0.004457779228687286, + 0.004046292044222355, + 0.004166815895587206, + 0.004573415499180555, + 0.0030732799787074327, + 0.02639206312596798, + 0.10748598724603653, + 0.05033792555332184, + 0.02189115434885025, + 0.02442561835050583, + 0.14768266677856445, + 0.019982650876045227, + 0.04438794031739235, + 0.004887217655777931, + 0.0051970528438687325, + 0.010566490702331066, + 0.006910556461662054, + 0.004385566338896751, + 0.01222279667854309, + 0.005319702904671431, + 0.01165890321135521, + 0.023012083023786545, + 0.015243812464177608, + 0.05608319491147995, + 0.05127895995974541, + 0.01842307858169079, + 0.01681809313595295, + 0.03045932576060295, + 0.011838536709547043, + 0.0008405105327256024, + 0.0007000273908488452, + 0.0011155488900840282, + 0.0012502566678449512, + 0.0007195352227427065, + 0.0015608931425958872, + 0.000972588371951133, + 0.0007477509207092226, + 0.01102929376065731, + 0.004481421783566475, + 0.0034827229101210833, + 0.005330522544682026, + 0.0036102712620049715, + 0.011040846817195415, + 0.0025619654916226864, + 0.004376278258860111, + 0.005696004722267389, + 0.0034857208374887705, + 0.006278008688241243, + 0.004429266322404146, + 0.005414975341409445, + 0.006797116249799728, + 0.0024946173653006554, + 0.002495468594133854, + 0.005379986017942429, + 0.03908386826515198, + 0.024881239980459213, + 0.013855138793587685, + 0.006641659885644913, + 0.004179287701845169, + 0.0017044935375452042, + 0.0019407252548262477, + 0.0011607245542109013, + 0.001913824351504445, + 0.0012661372311413288, + 0.0019146489212289453, + 0.0011357415933161974, + 0.0007579347584396601, + 0.0009742616675794125, + 0.000881847576238215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.14289966225624084 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104048945009708 + ], + "opposite_logprob_tokens": [ + -5.504104137420654 + ], + "logprob_reward": -0.004104048945009708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104048945009708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06820991635322571, + 0.03173231706023216, + 0.3741108179092407, + 0.07457032799720764, + 0.2579721510410309, + 0.03362150117754936, + 0.01689181663095951, + 0.06714114546775818, + 0.020640866830945015, + 0.0551091805100441 + ], + "credit_map": [ + 0.005988805089145899, + 0.004940695594996214, + 0.01117030344903469, + 0.007217717822641134, + 0.004470936954021454, + 0.006458844058215618, + 0.008423388935625553, + 0.01953921653330326, + 0.004431697074323893, + 0.0024023097939789295, + 0.005126413889229298, + 0.004032795317471027, + 0.0029614288359880447, + 0.005862402729690075, + 0.0038395486772060394, + 0.003075718181207776, + 0.02141331508755684, + 0.09086981415748596, + 0.042756903916597366, + 0.018355417996644974, + 0.021311702206730843, + 0.12428171932697296, + 0.016992194578051567, + 0.03812970966100693, + 0.007684352342039347, + 0.004796707537025213, + 0.019057070836424828, + 0.006835834588855505, + 0.004417708609253168, + 0.011858186684548855, + 0.005444751121103764, + 0.014475707896053791, + 0.027511905878782272, + 0.016286617144942284, + 0.06601204723119736, + 0.04788229241967201, + 0.0249781534075737, + 0.02784171886742115, + 0.03434716537594795, + 0.013112237676978111, + 0.004498930182307959, + 0.0031321581918746233, + 0.004787455778568983, + 0.0026257350109517574, + 0.0031896615400910378, + 0.009152516722679138, + 0.002536976244300604, + 0.0036980626173317432, + 0.002965734340250492, + 0.002307918621227145, + 0.0013839774765074253, + 0.0023916875943541527, + 0.0012245960533618927, + 0.00203282805159688, + 0.0013538619969040155, + 0.003231212729588151, + 0.008142928592860699, + 0.0072532580234110355, + 0.003665284253656864, + 0.004737094976007938, + 0.02239936962723732, + 0.010036994703114033, + 0.004244944546371698, + 0.00666126050055027, + 0.0016066951211541891, + 0.0028268834576010704, + 0.0033575226552784443, + 0.0025378710124641657, + 0.0017903342377394438, + 0.004097198601812124, + 0.0013833194971084595, + 0.003041040850803256, + 0.0057329474948346615, + 0.013215330429375172, + 0.007254777476191521, + 0.006222563795745373, + 0.005118511617183685, + 0.009311807341873646, + 0.0024439608678221703, + 0.005809275899082422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.004104048945009708 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the face of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's long dark brown hair shown as a violent whipping blur flying straight back out of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the woman's large hazel eyes visible with tears welling up in the corners and curling back to her temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as an indecipherable motion blur containing streaks of blue sky, green trees, and metallic gray structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a shaky, handheld camera effect that hums and vibrates in sync with the chaotic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306020215153694 + ], + "opposite_logprob_tokens": [ + -3.168060302734375 + ], + "logprob_reward": -0.04306020215153694, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306020215153694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07090212404727936, + 0.04100016504526138, + 0.45658618211746216, + 0.09664490818977356, + 0.08094430714845657, + 0.010557488538324833, + 0.01792323775589466, + 0.018988845869898796, + 0.1623118668794632, + 0.04414086043834686 + ], + "credit_map": [ + 0.009443526156246662, + 0.0062173823826014996, + 0.010971014387905598, + 0.010351250879466534, + 0.006501368712633848, + 0.006541077978909016, + 0.008514700457453728, + 0.012361802160739899, + 0.004487144760787487, + 0.003322143107652664, + 0.004822934977710247, + 0.005845874082297087, + 0.004162605851888657, + 0.006786203943192959, + 0.004697110038250685, + 0.006876147352159023, + 0.0260012149810791, + 0.11988388001918793, + 0.051267147064208984, + 0.028955956920981407, + 0.02521486207842827, + 0.14198724925518036, + 0.02236383594572544, + 0.04091205075383186, + 0.007204538676887751, + 0.007495061028748751, + 0.016856854781508446, + 0.008209348656237125, + 0.006648560054600239, + 0.020108511671423912, + 0.008229357190430164, + 0.02189267799258232, + 0.00838551763445139, + 0.005455792881548405, + 0.02137317880988121, + 0.016872946172952652, + 0.006591456010937691, + 0.007963539101183414, + 0.009761554189026356, + 0.004540320951491594, + 0.001281934091821313, + 0.0012255929177626967, + 0.0010444429935887456, + 0.0013581538805738091, + 0.0010785552440211177, + 0.002486062003299594, + 0.0009164067450910807, + 0.0011663412442430854, + 0.001538182608783245, + 0.0023855394683778286, + 0.0011114755179733038, + 0.0029610104393213987, + 0.0013810379896312952, + 0.0024204771034419537, + 0.0014766831882297993, + 0.00464883167296648, + 0.0018383966526016593, + 0.0028866082429885864, + 0.0018103409092873335, + 0.0033700033091008663, + 0.0029427288100123405, + 0.0028642932884395123, + 0.0016397074796259403, + 0.0016367662465199828, + 0.008497078903019428, + 0.06568356603384018, + 0.04086035117506981, + 0.023857491090893745, + 0.010841641575098038, + 0.0057098157703876495, + 0.003138225059956312, + 0.003723695408552885, + 0.004000671207904816, + 0.00663298275321722, + 0.005785269197076559, + 0.007028905209153891, + 0.004532734397798777, + 0.007404047064483166, + 0.003670532489195466, + 0.0050857169553637505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.04306020215153694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_8/index_tensor([9877, 9877])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_8/index_tensor([9877, 9877])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e7798ec1e20e67ff3f9d16957a55c3379d85ca59 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_8/index_tensor([9877, 9877])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot, shot from a slightly overhead angle, draws the viewer's attention to an elderly craftsman. His face, etched with age and deeply lined with wrinkles, his eyes narrowed in intense concentration beneath his thinning silver hair. He wears a washed-out, dark blue cotton double-breasted jacket and sits hunched over a heavy wooden workbench covered with sawdust and various carving tools. His left hand, calloused yet remarkably steady, carefully supports a rudimentary wooden sculpture—perhaps a small bird—while his right grips a sharp carving knife. The knife's tip gleams coldly in the warm light of a low-hanging tungsten bulb above the workbench. With his steady, precise movements, a thin, cicada-wing-thin shavings slowly curl and fall, drifting into the pile of sawdust piled on the workbench. This concentrated light not only illuminates the delicate work in his hands, but also casts deep shadows on his focused profile and rough hands, forming a strong contrast between light and dark and the dim, messy but orderly workshop in the background. The whole picture is immersed in an atmosphere of tranquility, concentration and a sense of time.", + "seed": 31821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_8/index_9877_seed_31821_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009468869306147099, + "vlm_logprob": -0.009468869306147099, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.009468869306147099, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015050427988171577 + ], + "opposite_logprob_tokens": [ + -4.640050411224365 + ], + "logprob_reward": -0.015050427988171577, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015050427988171577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15966443717479706, + 0.057982247322797775, + 0.05969266965985298, + 0.028607800602912903, + 0.06347176432609558, + 0.29657718539237976, + 0.07138782739639282, + 0.09479586035013199, + 0.1080971211194992, + 0.05972309410572052 + ], + "credit_map": [ + 0.024298999458551407, + 0.029617339372634888, + 0.01924019120633602, + 0.012101873755455017, + 0.022247174754738808, + 0.02255086600780487, + 0.01700202375650406, + 0.012605974450707436, + 0.007935428991913795, + 0.01065046526491642, + 0.007254213560372591, + 0.0048249405808746815, + 0.00713387131690979, + 0.006496916059404612, + 0.00548452977091074, + 0.008201884105801582, + 0.008975649252533913, + 0.011514164507389069, + 0.00823071226477623, + 0.005832074675709009, + 0.008932400494813919, + 0.007495929952710867, + 0.0041402848437428474, + 0.004571456927806139, + 0.0038391188718378544, + 0.004652933683246374, + 0.004218623507767916, + 0.0024772542528808117, + 0.003198464633896947, + 0.0038660455029457808, + 0.002077640499919653, + 0.004277719184756279, + 0.011857514269649982, + 0.011692997068166733, + 0.011962786316871643, + 0.006311269477009773, + 0.005677797831594944, + 0.008111382834613323, + 0.003764961613342166, + 0.004093059338629246, + 0.04493652284145355, + 0.05138307809829712, + 0.10558289289474487, + 0.017155498266220093, + 0.014356420375406742, + 0.04334729537367821, + 0.01092829741537571, + 0.008887183852493763, + 0.006484077777713537, + 0.008355478756129742, + 0.020133741199970245, + 0.0041170851327478886, + 0.0029666628688573837, + 0.02173982933163643, + 0.004018956795334816, + 0.003571999492123723, + 0.013278159312903881, + 0.019143953919410706, + 0.014664221554994583, + 0.006653516087681055, + 0.011316507123410702, + 0.018527492880821228, + 0.0055364263243973255, + 0.00567557942122221, + 0.013575860299170017, + 0.009840637445449829, + 0.038608163595199585, + 0.005006249062716961, + 0.0034348247572779655, + 0.03177817538380623, + 0.0026068545412272215, + 0.0032463520765304565, + 0.01119537279009819, + 0.01040302962064743, + 0.009137002751231194, + 0.004155810922384262, + 0.007508444134145975, + 0.010206988081336021, + 0.002974495990201831, + 0.004141945391893387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.015050427988171577 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565136887133121 + ], + "opposite_logprob_tokens": [ + -4.3875651359558105 + ], + "logprob_reward": -0.012565136887133121, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565136887133121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1282007098197937, + 0.040553998202085495, + 0.03029102273285389, + 0.0323808379471302, + 0.08444878458976746, + 0.2153903990983963, + 0.11612722277641296, + 0.10926668345928192, + 0.07387534528970718, + 0.16946494579315186 + ], + "credit_map": [ + 0.01404713373631239, + 0.016243761405348778, + 0.0245420690625906, + 0.012343289330601692, + 0.014212515205144882, + 0.01784542389214039, + 0.01731771230697632, + 0.011648806743323803, + 0.0051407101564109325, + 0.00660104863345623, + 0.004998528398573399, + 0.0040477849543094635, + 0.006991923786699772, + 0.004686468280851841, + 0.0029281347524374723, + 0.005159397609531879, + 0.004608595743775368, + 0.005096187815070152, + 0.004701944068074226, + 0.003441663458943367, + 0.0033359678927809, + 0.0037456073332577944, + 0.0020238147117197514, + 0.003337241942062974, + 0.004312471952289343, + 0.006151015404611826, + 0.004964824300259352, + 0.002857211045920849, + 0.004182471428066492, + 0.004377081524580717, + 0.0021781043615192175, + 0.003357658861204982, + 0.0165188517421484, + 0.015026331879198551, + 0.0164985079318285, + 0.008766336366534233, + 0.005568366963416338, + 0.01180677767843008, + 0.004583853762596846, + 0.005679761990904808, + 0.0343325100839138, + 0.03437505662441254, + 0.07606635242700577, + 0.01182313822209835, + 0.008908848278224468, + 0.035151951014995575, + 0.007388550788164139, + 0.007343986537307501, + 0.012030133977532387, + 0.011294483207166195, + 0.03832905739545822, + 0.005379826296120882, + 0.004289553500711918, + 0.03516748175024986, + 0.005631391890347004, + 0.004005299415439367, + 0.01616380177438259, + 0.015959469601511955, + 0.02370917797088623, + 0.011020041070878506, + 0.005292084999382496, + 0.02571243979036808, + 0.005286059342324734, + 0.0061236098408699036, + 0.013249201700091362, + 0.008282686583697796, + 0.02142966538667679, + 0.006015399936586618, + 0.003457452403381467, + 0.01342998817563057, + 0.003167945658788085, + 0.004843005910515785, + 0.02932385355234146, + 0.033054791390895844, + 0.03134071081876755, + 0.021342556923627853, + 0.013075822964310646, + 0.021503375843167305, + 0.009122798219323158, + 0.01070103608071804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.012565136887133121 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003202312160283327 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.003202312160283327, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003202312160283327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16421107947826385, + 0.05085847154259682, + 0.05112934112548828, + 0.03676861152052879, + 0.05433805286884308, + 0.16157324612140656, + 0.31098175048828125, + 0.07192394137382507, + 0.07525879889726639, + 0.02295670472085476 + ], + "credit_map": [ + 0.018802743405103683, + 0.018675189465284348, + 0.030705755576491356, + 0.014044510200619698, + 0.015424669720232487, + 0.02036348544061184, + 0.030637774616479874, + 0.0155569426715374, + 0.0068710241466760635, + 0.010271891951560974, + 0.006287751253694296, + 0.00617408799007535, + 0.0058575780130922794, + 0.005574724171310663, + 0.0041724503971636295, + 0.005648962687700987, + 0.005578063894063234, + 0.007498610764741898, + 0.007592790760099888, + 0.005561215337365866, + 0.004326506517827511, + 0.005842185113579035, + 0.007323406171053648, + 0.0074065630324184895, + 0.004676112439483404, + 0.006857284344732761, + 0.0070094275288283825, + 0.003750495845451951, + 0.0029185519088059664, + 0.004612871911376715, + 0.0029599633999168873, + 0.003983903210610151, + 0.0049930596724152565, + 0.007441333495080471, + 0.012143956497311592, + 0.0035117340739816427, + 0.004729406908154488, + 0.007858938537538052, + 0.007539455778896809, + 0.00612016674131155, + 0.02567341923713684, + 0.03292065113782883, + 0.057971399277448654, + 0.008541464805603027, + 0.006356613710522652, + 0.019153272733092308, + 0.005972634069621563, + 0.004983785096555948, + 0.04608776047825813, + 0.03225595876574516, + 0.11126678436994553, + 0.014649421907961369, + 0.012233144603669643, + 0.07675792276859283, + 0.00836044829338789, + 0.00937030278146267, + 0.008151402696967125, + 0.012845093384385109, + 0.00968481320887804, + 0.005880996119230986, + 0.007047462742775679, + 0.013910364359617233, + 0.0064589180983603, + 0.007944893091917038, + 0.008673255331814289, + 0.006549667101353407, + 0.026979906484484673, + 0.003190718125551939, + 0.0024822165723890066, + 0.022136317566037178, + 0.0023445396218448877, + 0.0029021792579442263, + 0.0035285630729049444, + 0.0034108103718608618, + 0.0028844615444540977, + 0.0019020827021449804, + 0.0020463441032916307, + 0.0042012473568320274, + 0.0019991688895970583, + 0.002984026214107871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.003202312160283327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742583122104406 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.006742583122104406, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742583122104406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12488164007663727, + 0.04897432029247284, + 0.028238963335752487, + 0.036569032818078995, + 0.04989127814769745, + 0.20648007094860077, + 0.24618175625801086, + 0.05954412370920181, + 0.17809650301933289, + 0.021142277866601944 + ], + "credit_map": [ + 0.012855252251029015, + 0.015697840601205826, + 0.029998961836099625, + 0.011154401116073132, + 0.010728692635893822, + 0.01968107372522354, + 0.012225953862071037, + 0.012539458461105824, + 0.0043958257883787155, + 0.007059118710458279, + 0.0072373924776911736, + 0.00436965050175786, + 0.0052284668199718, + 0.004952641203999519, + 0.00971740297973156, + 0.006013824138790369, + 0.0026911720633506775, + 0.0041747973300516605, + 0.0038179843686521053, + 0.00341864419169724, + 0.0025370060466229916, + 0.003246941836550832, + 0.004822216462343931, + 0.0035302010364830494, + 0.004297807812690735, + 0.005058369133621454, + 0.008520813658833504, + 0.0033887100871652365, + 0.0027732455637305975, + 0.004127574618905783, + 0.003738879458978772, + 0.004663634113967419, + 0.00779107166454196, + 0.0077814022079110146, + 0.011557457968592644, + 0.003944090101867914, + 0.004203454125672579, + 0.005745837464928627, + 0.003574373433366418, + 0.005293591413646936, + 0.034238606691360474, + 0.049187518656253815, + 0.06762152165174484, + 0.012053553946316242, + 0.00865729060024023, + 0.01876145973801613, + 0.008505151607096195, + 0.007454975508153439, + 0.03907691687345505, + 0.028652751818299294, + 0.08732950687408447, + 0.01356343924999237, + 0.009861807338893414, + 0.05397617816925049, + 0.0066168904304504395, + 0.007104262709617615, + 0.004964212421327829, + 0.0134431142359972, + 0.008061232976615429, + 0.004605746362358332, + 0.00717993313446641, + 0.012816124595701694, + 0.0036762231029570103, + 0.004797537345439196, + 0.02314714342355728, + 0.016414152458310127, + 0.06422451138496399, + 0.00814751721918583, + 0.004781628958880901, + 0.05189800634980202, + 0.004158186260610819, + 0.00532537279650569, + 0.00456963200122118, + 0.0037468334194272757, + 0.0026147160679101944, + 0.0019029169343411922, + 0.0020216084085404873, + 0.003095804015174508, + 0.001566614373587072, + 0.0016241520643234253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.006742583122104406 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17830491065979004, + 0.08585881441831589, + 0.07246583700180054, + 0.03360756114125252, + 0.11484161764383316, + 0.20889796316623688, + 0.1058276817202568, + 0.07272893190383911, + 0.06476657092571259, + 0.0627000704407692 + ], + "credit_map": [ + 0.02235930599272251, + 0.024850092828273773, + 0.024698199704289436, + 0.018461640924215317, + 0.01765044592320919, + 0.025085758417844772, + 0.021606042981147766, + 0.023593418300151825, + 0.008176914416253567, + 0.00922972522675991, + 0.011491868644952774, + 0.006887679919600487, + 0.008796888403594494, + 0.009171515703201294, + 0.016879215836524963, + 0.015225008130073547, + 0.007551827933639288, + 0.010908186435699463, + 0.009368122555315495, + 0.006072367541491985, + 0.008357679471373558, + 0.011012067086994648, + 0.008928566239774227, + 0.010267018340528011, + 0.004692190792411566, + 0.00593196414411068, + 0.007983921095728874, + 0.002942981431260705, + 0.003389650257304311, + 0.003962602000683546, + 0.0021973231341689825, + 0.0025069296825677156, + 0.014523759484291077, + 0.014897609129548073, + 0.019802246242761612, + 0.012226622551679611, + 0.01013935822993517, + 0.012400755658745766, + 0.01307115238159895, + 0.017780112102627754, + 0.03469664230942726, + 0.0371834896504879, + 0.07275143265724182, + 0.012248176150023937, + 0.008888913318514824, + 0.027787841856479645, + 0.00783894956111908, + 0.007502519525587559, + 0.01809931918978691, + 0.015765327960252762, + 0.03485758602619171, + 0.00763663649559021, + 0.005032477434724569, + 0.0172652006149292, + 0.003503676038235426, + 0.003667453769594431, + 0.01022128202021122, + 0.00905331689864397, + 0.01325182057917118, + 0.007196059916168451, + 0.007639861200004816, + 0.012118703685700893, + 0.004346485249698162, + 0.008901401422917843, + 0.00803380273282528, + 0.006965873762965202, + 0.021549604833126068, + 0.0032005480024963617, + 0.002376560354605317, + 0.017914429306983948, + 0.0019064487423747778, + 0.002819301327690482, + 0.008922416716814041, + 0.010399417020380497, + 0.011830354109406471, + 0.005130534525960684, + 0.007239603437483311, + 0.010387331247329712, + 0.0034695197828114033, + 0.005320890806615353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.00978388637304306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly overhead angle, draws the viewer's attention to an elderly craftsman. His face, etched with age and deeply lined with wrinkles, his eyes narrowed in intense concentration beneath his thinning silver hair. He wears a washed-out, dark blue cotton double-breasted jacket and sits hunched over a heavy wooden workbench covered with sawdust and various carving tools. His left hand, calloused yet remarkably steady, carefully supports a rudimentary wooden sculpture—perhaps a small bird—while his right grips a sharp carving knife. The knife's tip gleams coldly in the warm light of a low-hanging tungsten bulb above the workbench. With his steady, precise movements, a thin, cicada-wing-thin shavings slowly curl and fall, drifting into the pile of sawdust piled on the workbench. This concentrated light not only illuminates the delicate work in his hands, but also casts deep shadows on his focused profile and rough hands, forming a strong contrast between light and dark and the dim, messy but orderly workshop in the background. The whole picture is immersed in an atmosphere of tranquility, concentration and a sense of time.", + "seed": 31822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_8/index_9877_seed_31822_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.033037299290299416, + "vlm_logprob": -0.033037299290299416, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.033037299290299416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05083785206079483 + ], + "opposite_logprob_tokens": [ + -3.050837755203247 + ], + "logprob_reward": -0.05083785206079483, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05083785206079483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25595229864120483, + 0.12347862124443054, + 0.036487285047769547, + 0.08950243890285492, + 0.19099561870098114, + 0.07721652835607529, + 0.04294472932815552, + 0.04629279673099518, + 0.028416352346539497, + 0.10871338844299316 + ], + "credit_map": [ + 0.03661861643195152, + 0.04830878600478172, + 0.0136941559612751, + 0.017971456050872803, + 0.04047747328877449, + 0.0441448912024498, + 0.019190294668078423, + 0.03554660454392433, + 0.021209225058555603, + 0.013447905890643597, + 0.007429769728332758, + 0.011856493540108204, + 0.01854350045323372, + 0.023188812658190727, + 0.009785231202840805, + 0.018017666414380074, + 0.004313429817557335, + 0.004474696237593889, + 0.001518751378171146, + 0.002779634203761816, + 0.006691248621791601, + 0.005645411554723978, + 0.005482219625264406, + 0.005581892095506191, + 0.02214113622903824, + 0.009832418523728848, + 0.0052635762840509415, + 0.005098062101751566, + 0.013425623998045921, + 0.014402680099010468, + 0.005818365141749382, + 0.013520571403205395, + 0.04233875870704651, + 0.014258520677685738, + 0.008535904809832573, + 0.011244745925068855, + 0.03259783610701561, + 0.03401242941617966, + 0.025606976822018623, + 0.02240043506026268, + 0.010357971303164959, + 0.009294834919273853, + 0.005684472154825926, + 0.005436693783849478, + 0.011445255018770695, + 0.011957507580518723, + 0.005772409029304981, + 0.01726738177239895, + 0.008050361648201942, + 0.00277889147400856, + 0.002093230141326785, + 0.002753456123173237, + 0.007432339712977409, + 0.00579844182357192, + 0.0029112074989825487, + 0.011126795783638954, + 0.007712443824857473, + 0.006124727427959442, + 0.0025418971199542284, + 0.0034123416990041733, + 0.008826330304145813, + 0.005523657891899347, + 0.005636502522975206, + 0.006514894776046276, + 0.003921793308109045, + 0.0027919148560613394, + 0.0016599660739302635, + 0.0029136361554265022, + 0.00409385422244668, + 0.0035886657424271107, + 0.0050491224974393845, + 0.004397400189191103, + 0.011280488222837448, + 0.006474224850535393, + 0.005469975993037224, + 0.00967636052519083, + 0.02788281813263893, + 0.02572857402265072, + 0.005956166423857212, + 0.016244767233729362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.05083785206079483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207158789038658 + ], + "opposite_logprob_tokens": [ + -2.8120715618133545 + ], + "logprob_reward": -0.06207158789038658, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207158789038658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3888862133026123, + 0.13861054182052612, + 0.05022995173931122, + 0.028297925367951393, + 0.1137528270483017, + 0.10737676918506622, + 0.046188194304704666, + 0.018919100984930992, + 0.017766719684004784, + 0.08997181802988052 + ], + "credit_map": [ + 0.05935300886631012, + 0.05671992897987366, + 0.01924177259206772, + 0.024336257949471474, + 0.07794531434774399, + 0.07859137654304504, + 0.027769017964601517, + 0.044929537922143936, + 0.023983722552657127, + 0.014407052658498287, + 0.008052227087318897, + 0.012489426881074905, + 0.022536784410476685, + 0.02681637369096279, + 0.011205945163965225, + 0.01911901868879795, + 0.009377099573612213, + 0.00588179100304842, + 0.002243882277980447, + 0.0031733529176563025, + 0.008458901196718216, + 0.008840382099151611, + 0.005356484092772007, + 0.006898057647049427, + 0.006859237793833017, + 0.003119327127933502, + 0.001953444443643093, + 0.001810551038943231, + 0.004000902641564608, + 0.004446857608854771, + 0.002030064119026065, + 0.004077543038874865, + 0.02262141741812229, + 0.008045694790780544, + 0.005812188610434532, + 0.007689179386943579, + 0.02585470862686634, + 0.024997318163514137, + 0.008705433458089828, + 0.010026886127889156, + 0.015786292031407356, + 0.014186911284923553, + 0.00716541800647974, + 0.007658401969820261, + 0.01509846281260252, + 0.019393954426050186, + 0.0077436831779778, + 0.020343637093901634, + 0.005716114304959774, + 0.004386384040117264, + 0.00273911957629025, + 0.003452456556260586, + 0.007346679922193289, + 0.006421854253858328, + 0.0035035619512200356, + 0.012622023932635784, + 0.0023274917621165514, + 0.0028159681241959333, + 0.0015768930315971375, + 0.0028765383176505566, + 0.0024522668682038784, + 0.002233796985819936, + 0.0016987662529572845, + 0.002937381388619542, + 0.0023982638958841562, + 0.0015944449696689844, + 0.0011255609570071101, + 0.0022314239758998156, + 0.0023753878194838762, + 0.0019201524555683136, + 0.0032798140309751034, + 0.0028416705317795277, + 0.01065005548298359, + 0.006253098137676716, + 0.005605158396065235, + 0.008883610367774963, + 0.0207202211022377, + 0.017985878512263298, + 0.005565110594034195, + 0.014308682642877102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.06207158789038658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020606597885489464 + ], + "opposite_logprob_tokens": [ + -3.895606517791748 + ], + "logprob_reward": -0.020606597885489464, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020606597885489464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1988573521375656, + 0.2584010362625122, + 0.04708760604262352, + 0.04339713975787163, + 0.2433437556028366, + 0.037781111896038055, + 0.046731092035770416, + 0.020404798910021782, + 0.024093279615044594, + 0.07990284264087677 + ], + "credit_map": [ + 0.029269257560372353, + 0.03252578526735306, + 0.01075194962322712, + 0.01473198365420103, + 0.034117333590984344, + 0.036569517105817795, + 0.015380465425550938, + 0.02551104873418808, + 0.04853176698088646, + 0.024762116372585297, + 0.01332620158791542, + 0.02128327079117298, + 0.036375150084495544, + 0.05198836699128151, + 0.021052585914731026, + 0.04108155518770218, + 0.004532217979431152, + 0.0029990256298333406, + 0.004514400847256184, + 0.003738864092156291, + 0.0071463389322161674, + 0.007718366105109453, + 0.006097464822232723, + 0.010340929962694645, + 0.009628134779632092, + 0.004640527069568634, + 0.003863057354465127, + 0.0042004394344985485, + 0.005544499959796667, + 0.006314095109701157, + 0.003254855517297983, + 0.005951534491032362, + 0.05677632987499237, + 0.020565388724207878, + 0.013194781728088856, + 0.01903129182755947, + 0.04493989050388336, + 0.04730070009827614, + 0.018909653648734093, + 0.022625720128417015, + 0.004420480690896511, + 0.00406302185729146, + 0.0027109773363918066, + 0.003807270200923085, + 0.00578591413795948, + 0.00558372912928462, + 0.00283510098233819, + 0.00857461616396904, + 0.006401234306395054, + 0.0045740967616438866, + 0.002724219346418977, + 0.005572063848376274, + 0.006567637901753187, + 0.006472386885434389, + 0.0037844388280063868, + 0.010635017417371273, + 0.0022120620124042034, + 0.0028564995154738426, + 0.0020930778700858355, + 0.0035752614494413137, + 0.003052712883800268, + 0.0021889531053602695, + 0.0018940793816000223, + 0.002532153157517314, + 0.004059855360537767, + 0.001952158403582871, + 0.0014795922907069325, + 0.003339067567139864, + 0.0031837013084441423, + 0.002423743484541774, + 0.004129345994442701, + 0.0035258159041404724, + 0.00811240915209055, + 0.004509603604674339, + 0.002893104450777173, + 0.00263342447578907, + 0.026276536285877228, + 0.025805000215768814, + 0.0024006469175219536, + 0.007272119168192148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.020606597885489464 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108488067984581 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.01108488067984581, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108488067984581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25975796580314636, + 0.18487857282161713, + 0.04079529270529747, + 0.0839640274643898, + 0.07444316148757935, + 0.08156982809305191, + 0.06502176076173782, + 0.06336641311645508, + 0.03151519596576691, + 0.11468768119812012 + ], + "credit_map": [ + 0.043582528829574585, + 0.04422043263912201, + 0.012319626286625862, + 0.016149919480085373, + 0.04391820356249809, + 0.048296164721250534, + 0.019014088436961174, + 0.032257042825222015, + 0.027784116566181183, + 0.021051112562417984, + 0.012010165490210056, + 0.01869894750416279, + 0.031499430537223816, + 0.034284234046936035, + 0.01578354276716709, + 0.02376704104244709, + 0.007112992461770773, + 0.004390490707010031, + 0.002168406033888459, + 0.003640288021415472, + 0.005189100280404091, + 0.005562520120292902, + 0.006528121884912252, + 0.006203378085047007, + 0.012540366500616074, + 0.008294571191072464, + 0.006004019640386105, + 0.008181431330740452, + 0.01584656909108162, + 0.01420576125383377, + 0.0059988368302583694, + 0.01289247628301382, + 0.012917288579046726, + 0.004946351982653141, + 0.004662083461880684, + 0.004709695000201464, + 0.019422190263867378, + 0.016499735414981842, + 0.004584295675158501, + 0.006701529026031494, + 0.014919971115887165, + 0.009445123374462128, + 0.00506105599924922, + 0.005532012786716223, + 0.012771833688020706, + 0.015315691009163857, + 0.004486809950321913, + 0.014037341810762882, + 0.007641219068318605, + 0.007540630176663399, + 0.0038814072031527758, + 0.005208670161664486, + 0.00949314795434475, + 0.009122495539486408, + 0.005599965341389179, + 0.01653422601521015, + 0.014614388346672058, + 0.007144249975681305, + 0.004510452970862389, + 0.004954494535923004, + 0.009594221599400043, + 0.009538254700601101, + 0.003952531144022942, + 0.009057826362550259, + 0.0037224653642624617, + 0.0030672180000692606, + 0.0021496335975825787, + 0.005120669957250357, + 0.004263347014784813, + 0.0035720847081393003, + 0.005168787203729153, + 0.004450996872037649, + 0.009892649948596954, + 0.006121092475950718, + 0.004766896367073059, + 0.004550644662231207, + 0.03768252953886986, + 0.03626694902777672, + 0.004062659572809935, + 0.01134426798671484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.01108488067984581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585577934980392 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.020585577934980392, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585577934980392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35086771845817566, + 0.10574749857187271, + 0.0578419491648674, + 0.061314936727285385, + 0.0622846782207489, + 0.05005420744419098, + 0.04638819769024849, + 0.04236677289009094, + 0.03395945951342583, + 0.18917466700077057 + ], + "credit_map": [ + 0.049283046275377274, + 0.04557573422789574, + 0.01763930916786194, + 0.022044101729989052, + 0.0718037337064743, + 0.07434656471014023, + 0.025967562571167946, + 0.0442076213657856, + 0.017033053562045097, + 0.01705853082239628, + 0.007377923466265202, + 0.009479646570980549, + 0.01849067024886608, + 0.019145073369145393, + 0.0060879867523908615, + 0.011074605397880077, + 0.010038148611783981, + 0.006920630112290382, + 0.0028993727173656225, + 0.004389223176985979, + 0.009648758918046951, + 0.008630222640931606, + 0.008391110226511955, + 0.006924470886588097, + 0.01197126042097807, + 0.006242735311388969, + 0.004636228084564209, + 0.004175438079982996, + 0.009902643039822578, + 0.009471633471548557, + 0.004768861923366785, + 0.01014612428843975, + 0.011667365208268166, + 0.006159419193863869, + 0.00701720779761672, + 0.008069867268204689, + 0.00760101480409503, + 0.007642587646842003, + 0.0057363612577319145, + 0.00839084293693304, + 0.010274452157318592, + 0.00529114156961441, + 0.0029552096966654062, + 0.0027510072104632854, + 0.007922015152871609, + 0.009616529569029808, + 0.0026252486277371645, + 0.008618599735200405, + 0.005273033864796162, + 0.0058917030692100525, + 0.0024378446396440268, + 0.003810883266851306, + 0.007008417043834925, + 0.006269299890846014, + 0.005252706818282604, + 0.010444301180541515, + 0.01149502582848072, + 0.0048988135531544685, + 0.00285961851477623, + 0.002896670252084732, + 0.005899052135646343, + 0.006535103544592857, + 0.002325567649677396, + 0.005456915590912104, + 0.009106307290494442, + 0.0032258108258247375, + 0.0020727270748466253, + 0.002393657574430108, + 0.004598546773195267, + 0.004088982008397579, + 0.003872551489621401, + 0.004600874613970518, + 0.015833407640457153, + 0.010722647421061993, + 0.00959330890327692, + 0.01681903377175331, + 0.0538368746638298, + 0.048684727400541306, + 0.010077467188239098, + 0.023607157170772552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.020585577934980392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly overhead angle, draws the viewer's attention to an elderly craftsman. His face, etched with age and deeply lined with wrinkles, his eyes narrowed in intense concentration beneath his thinning silver hair. He wears a washed-out, dark blue cotton double-breasted jacket and sits hunched over a heavy wooden workbench covered with sawdust and various carving tools. His left hand, calloused yet remarkably steady, carefully supports a rudimentary wooden sculpture—perhaps a small bird—while his right grips a sharp carving knife. The knife's tip gleams coldly in the warm light of a low-hanging tungsten bulb above the workbench. With his steady, precise movements, a thin, cicada-wing-thin shavings slowly curl and fall, drifting into the pile of sawdust piled on the workbench. This concentrated light not only illuminates the delicate work in his hands, but also casts deep shadows on his focused profile and rough hands, forming a strong contrast between light and dark and the dim, messy but orderly workshop in the background. The whole picture is immersed in an atmosphere of tranquility, concentration and a sense of time.", + "seed": 31823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_8/index_9877_seed_31823_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010552443377673627, + "vlm_logprob": -0.010552443377673627, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.010552443377673627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011727320961654186 + ], + "opposite_logprob_tokens": [ + -4.886727333068848 + ], + "logprob_reward": -0.011727320961654186, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011727320961654186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21851250529289246, + 0.09338203817605972, + 0.05370499566197395, + 0.07632479071617126, + 0.06192205473780632, + 0.06740249693393707, + 0.05750911310315132, + 0.034300435334444046, + 0.29537609219551086, + 0.04156549274921417 + ], + "credit_map": [ + 0.03310394287109375, + 0.023484036326408386, + 0.02691740170121193, + 0.027552282437682152, + 0.029162611812353134, + 0.03183264657855034, + 0.01488586701452732, + 0.03157370537519455, + 0.013959714211523533, + 0.019525090232491493, + 0.011588898487389088, + 0.0069111427292227745, + 0.012363594956696033, + 0.013417438603937626, + 0.006589381955564022, + 0.00902677234262228, + 0.008078239858150482, + 0.007371949963271618, + 0.0058109755627810955, + 0.005558195058256388, + 0.00951489619910717, + 0.007280284073203802, + 0.005317593924701214, + 0.004772861488163471, + 0.007243465166538954, + 0.01055708434432745, + 0.007922338321805, + 0.011139103211462498, + 0.009254208765923977, + 0.012065749615430832, + 0.006182965822517872, + 0.011959875002503395, + 0.007025680970400572, + 0.007900973781943321, + 0.006198460701853037, + 0.007143228780478239, + 0.010155911557376385, + 0.00993291474878788, + 0.004619813524186611, + 0.008945070207118988, + 0.010991900227963924, + 0.00788251031190157, + 0.007253366522490978, + 0.005410781130194664, + 0.006296143401414156, + 0.0182736087590456, + 0.006797226145863533, + 0.004496963229030371, + 0.009204740636050701, + 0.009436657652258873, + 0.006661630235612392, + 0.006856437772512436, + 0.007751790806651115, + 0.0088307810947299, + 0.003964154981076717, + 0.004802917130291462, + 0.0035167522728443146, + 0.0040029254741966724, + 0.0045155915431678295, + 0.0036579156294465065, + 0.003929371479898691, + 0.0048868488520383835, + 0.006054810248315334, + 0.003736217739060521, + 0.025765080004930496, + 0.010536261834204197, + 0.1047729030251503, + 0.014295805245637894, + 0.009762485511600971, + 0.02504388988018036, + 0.010210043750703335, + 0.0949895977973938, + 0.005004127975553274, + 0.005952097941190004, + 0.004609611816704273, + 0.00419691251590848, + 0.0049205743707716465, + 0.00772452587261796, + 0.004002836998552084, + 0.005154806189239025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.011727320961654186 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216925017535686 + ], + "opposite_logprob_tokens": [ + -4.264216899871826 + ], + "logprob_reward": -0.014216925017535686, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216925017535686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14352047443389893, + 0.0842130035161972, + 0.04214264079928398, + 0.03130999952554703, + 0.04211628437042236, + 0.036466147750616074, + 0.029955759644508362, + 0.02140122279524803, + 0.5287405252456665, + 0.040133941918611526 + ], + "credit_map": [ + 0.019218802452087402, + 0.018182961270213127, + 0.015361931174993515, + 0.016341557726264, + 0.01674593798816204, + 0.021447675302624702, + 0.010613792575895786, + 0.025607820600271225, + 0.01043605711311102, + 0.017328061163425446, + 0.008563938550651073, + 0.008095595054328442, + 0.011302974075078964, + 0.01250034011900425, + 0.006127585656940937, + 0.009858448058366776, + 0.008818987756967545, + 0.007285160943865776, + 0.005002882797271013, + 0.0033919811248779297, + 0.004476034082472324, + 0.006203110329806805, + 0.003185942769050598, + 0.003778542624786496, + 0.004129314795136452, + 0.005450452212244272, + 0.0034017921425402164, + 0.003460261970758438, + 0.0033341282978653908, + 0.0047241076827049255, + 0.003460200736299157, + 0.0033497398253530264, + 0.005740060005337, + 0.008766420185565948, + 0.004059385042637587, + 0.005605222657322884, + 0.004781687166541815, + 0.006643969099968672, + 0.0025720554403960705, + 0.0039474843069911, + 0.006350558716803789, + 0.0049795713275671005, + 0.007034547161310911, + 0.0027651390992105007, + 0.0050692628137767315, + 0.005592942237854004, + 0.0022182355169206858, + 0.0024558899458497763, + 0.004475122317671776, + 0.005256578326225281, + 0.0032215239480137825, + 0.003706882707774639, + 0.003902238793671131, + 0.005069127306342125, + 0.0021147888619452715, + 0.002209498779848218, + 0.0029951073229312897, + 0.0029273061081767082, + 0.002176120411604643, + 0.002207913901656866, + 0.002390665002167225, + 0.004697372671216726, + 0.002168606733903289, + 0.0018381306435912848, + 0.045308999717235565, + 0.011160587891936302, + 0.1964007019996643, + 0.01574382744729519, + 0.012341600842773914, + 0.044560521841049194, + 0.014024063013494015, + 0.1892002522945404, + 0.005246809218078852, + 0.007713004481047392, + 0.004650567192584276, + 0.00646434398368001, + 0.003405566094443202, + 0.005559094715863466, + 0.0029151502531021833, + 0.0041794064454734325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.014216925017535686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955095402896404 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955095402896404, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955095402896404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24995164573192596, + 0.118921659886837, + 0.1058332771062851, + 0.04818931967020035, + 0.016155391931533813, + 0.02434965781867504, + 0.044906049966812134, + 0.02323923446238041, + 0.3254848122596741, + 0.042969007045030594 + ], + "credit_map": [ + 0.036050185561180115, + 0.029662899672985077, + 0.041455887258052826, + 0.034788504242897034, + 0.020404359325766563, + 0.03745325282216072, + 0.016047758981585503, + 0.03408882021903992, + 0.016613274812698364, + 0.02755129523575306, + 0.01676565408706665, + 0.008156657218933105, + 0.013258718885481358, + 0.016970282420516014, + 0.00738953985273838, + 0.012216242961585522, + 0.014717592857778072, + 0.012845557183027267, + 0.021059995517134666, + 0.011635207571089268, + 0.009640000760555267, + 0.0133367283269763, + 0.011103233322501183, + 0.011494964361190796, + 0.007296856492757797, + 0.006399430334568024, + 0.00791870430111885, + 0.006768921390175819, + 0.0038740208838135004, + 0.006813714746385813, + 0.0041697085835039616, + 0.004947960376739502, + 0.00248840544372797, + 0.002371997805312276, + 0.0018987563671544194, + 0.0015263938112184405, + 0.0018347261939197779, + 0.0023966035805642605, + 0.001432111137546599, + 0.0022063981741666794, + 0.00384097290225327, + 0.0032101499382406473, + 0.003923595417290926, + 0.001950880978256464, + 0.002138127340003848, + 0.00376298138871789, + 0.002483176998794079, + 0.0030397733207792044, + 0.0055986796505749226, + 0.007693033665418625, + 0.005468079354614019, + 0.00484376261010766, + 0.005548636894673109, + 0.008915464393794537, + 0.0029776219744235277, + 0.0038607690948992968, + 0.0024996462743729353, + 0.003082276089116931, + 0.002853409620001912, + 0.002674105344340205, + 0.002870597643777728, + 0.0029215833637863398, + 0.0041222162544727325, + 0.0022153982426971197, + 0.028331443667411804, + 0.008938740938901901, + 0.12034358829259872, + 0.010529384948313236, + 0.008571388199925423, + 0.027673671022057533, + 0.008550824597477913, + 0.11254574358463287, + 0.005026194266974926, + 0.005969261284917593, + 0.007652608677744865, + 0.0047459471970796585, + 0.004514171276241541, + 0.008123982697725296, + 0.0032527539879083633, + 0.0036840871907770634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.005955095402896404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786129929125309 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.009786129929125309, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786129929125309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18221907317638397, + 0.09574390947818756, + 0.08753897249698639, + 0.03748829662799835, + 0.02241378463804722, + 0.044888224452733994, + 0.02832327038049698, + 0.04516655206680298, + 0.40623173117637634, + 0.04998619481921196 + ], + "credit_map": [ + 0.028018666431307793, + 0.02419639192521572, + 0.02618800848722458, + 0.019167110323905945, + 0.014998274855315685, + 0.026884013786911964, + 0.01554068736732006, + 0.0272259209305048, + 0.014857529662549496, + 0.01611359417438507, + 0.0095655657351017, + 0.009248138405382633, + 0.017483152449131012, + 0.012194965034723282, + 0.0072930301539599895, + 0.00898793712258339, + 0.011770103126764297, + 0.013140360824763775, + 0.010305803269147873, + 0.007828261703252792, + 0.01066224928945303, + 0.013522449880838394, + 0.012255589477717876, + 0.008054155856370926, + 0.006309642922133207, + 0.004435272887349129, + 0.00512149091809988, + 0.004035403020679951, + 0.003929251804947853, + 0.0055953124538064, + 0.004405490588396788, + 0.003656430169939995, + 0.0033105378970503807, + 0.003620787290856242, + 0.002189235296100378, + 0.0019155918853357434, + 0.004010830074548721, + 0.003090294310823083, + 0.0014933509519323707, + 0.002783159026876092, + 0.006333693861961365, + 0.0060443757101893425, + 0.007477528881281614, + 0.003749043447896838, + 0.004690093453973532, + 0.007824589498341084, + 0.004418529104441404, + 0.004350374918431044, + 0.0037096920423209667, + 0.005428662057965994, + 0.003162415698170662, + 0.003138067200779915, + 0.0038897586055099964, + 0.004540557041764259, + 0.001788848894648254, + 0.0026652677915990353, + 0.005175192374736071, + 0.005005504470318556, + 0.005367822479456663, + 0.00490671256557107, + 0.006177927367389202, + 0.0060801999643445015, + 0.00818166509270668, + 0.004271526355296373, + 0.038021430373191833, + 0.01365584321320057, + 0.1394491195678711, + 0.023699063807725906, + 0.014021489769220352, + 0.03365921229124069, + 0.04059061035513878, + 0.10313495993614197, + 0.005834993906319141, + 0.007163073867559433, + 0.005621498916298151, + 0.0052687544375658035, + 0.008350885473191738, + 0.007154155056923628, + 0.004636096768081188, + 0.005956733133643866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.009786129929125309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076745577156544 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076745577156544, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076745577156544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2105521559715271, + 0.14625594019889832, + 0.02260986715555191, + 0.02847529947757721, + 0.05339982360601425, + 0.02057746984064579, + 0.027575602754950523, + 0.014293955639004707, + 0.4347783029079437, + 0.04148166999220848 + ], + "credit_map": [ + 0.029612772166728973, + 0.029180951416492462, + 0.022977735847234726, + 0.02697506919503212, + 0.024237317964434624, + 0.03336231783032417, + 0.015374025329947472, + 0.028831971809267998, + 0.027579225599765778, + 0.026892632246017456, + 0.01939307525753975, + 0.010875874198973179, + 0.014097476378083229, + 0.022167516872286797, + 0.009510023519396782, + 0.015740102156996727, + 0.0030911113135516644, + 0.004121038131415844, + 0.0022593082394450903, + 0.001996312290430069, + 0.003112977836281061, + 0.0032780340407043695, + 0.0024728893768042326, + 0.002278195461258292, + 0.0047411806881427765, + 0.004015177022665739, + 0.002736495342105627, + 0.0025345822796225548, + 0.003342613810673356, + 0.004668364766985178, + 0.0018786233849823475, + 0.004558263346552849, + 0.009196841157972813, + 0.01066141203045845, + 0.004884984809905291, + 0.004932447336614132, + 0.005421176087111235, + 0.008195418864488602, + 0.0037319932598620653, + 0.006375551223754883, + 0.002997106174007058, + 0.0026700831949710846, + 0.0028681079857051373, + 0.0014874966582283378, + 0.0027084830217063427, + 0.004223411437124014, + 0.0017006852431222796, + 0.0019220954272896051, + 0.0035456453915685415, + 0.0046415869146585464, + 0.003059308510273695, + 0.0026734601706266403, + 0.003486358793452382, + 0.005245742853730917, + 0.002042115665972233, + 0.002881383989006281, + 0.0018428359180688858, + 0.0019687595777213573, + 0.00165734370239079, + 0.0016374161932617426, + 0.0019614603370428085, + 0.002095068572089076, + 0.0018388625467196107, + 0.0012922089081257582, + 0.038932591676712036, + 0.010493505746126175, + 0.16352029144763947, + 0.015028887428343296, + 0.01042116153985262, + 0.03647780790925026, + 0.013847855851054192, + 0.14605620503425598, + 0.005764426663517952, + 0.0064600384794175625, + 0.006646675523370504, + 0.00511208875104785, + 0.00444716727361083, + 0.006223828997462988, + 0.0027476202230900526, + 0.004079822916537523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.011076745577156544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly overhead angle, draws the viewer's attention to an elderly craftsman. His face, etched with age and deeply lined with wrinkles, his eyes narrowed in intense concentration beneath his thinning silver hair. He wears a washed-out, dark blue cotton double-breasted jacket and sits hunched over a heavy wooden workbench covered with sawdust and various carving tools. His left hand, calloused yet remarkably steady, carefully supports a rudimentary wooden sculpture—perhaps a small bird—while his right grips a sharp carving knife. The knife's tip gleams coldly in the warm light of a low-hanging tungsten bulb above the workbench. With his steady, precise movements, a thin, cicada-wing-thin shavings slowly curl and fall, drifting into the pile of sawdust piled on the workbench. This concentrated light not only illuminates the delicate work in his hands, but also casts deep shadows on his focused profile and rough hands, forming a strong contrast between light and dark and the dim, messy but orderly workshop in the background. The whole picture is immersed in an atmosphere of tranquility, concentration and a sense of time.", + "seed": 31824, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_8/index_9877_seed_31824_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01635918819811195, + "vlm_logprob": -0.01635918819811195, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.01635918819811195, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03690085560083389 + ], + "opposite_logprob_tokens": [ + -3.411900758743286 + ], + "logprob_reward": -0.03690085560083389, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03690085560083389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14595073461532593, + 0.10678671300411224, + 0.0856981873512268, + 0.038575559854507446, + 0.08665397763252258, + 0.0944073498249054, + 0.10839305818080902, + 0.0117656160145998, + 0.04052514582872391, + 0.2812437117099762 + ], + "credit_map": [ + 0.014665849506855011, + 0.017043665051460266, + 0.013783547095954418, + 0.0142503147944808, + 0.026247546076774597, + 0.02018912136554718, + 0.01863507553935051, + 0.02113560400903225, + 0.003925696946680546, + 0.003358815098181367, + 0.004773860331624746, + 0.003967425785958767, + 0.0048166848719120026, + 0.041873782873153687, + 0.004742812365293503, + 0.039327628910541534, + 0.0072932117618620396, + 0.009689579717814922, + 0.004237320739775896, + 0.0057936543598771095, + 0.017816849052906036, + 0.014395559206604958, + 0.014588557183742523, + 0.011883453465998173, + 0.002525225281715393, + 0.003118144813925028, + 0.0017594193341210485, + 0.002321025589480996, + 0.0037428713403642178, + 0.010854436084628105, + 0.0033137574791908264, + 0.010940677486360073, + 0.002690879860892892, + 0.002871046308428049, + 0.002111519454047084, + 0.0015016255201771855, + 0.0065849111415445805, + 0.03417685627937317, + 0.004435049369931221, + 0.032282084226608276, + 0.005276483483612537, + 0.008165253326296806, + 0.004204094409942627, + 0.0047527761198580265, + 0.010834078304469585, + 0.02722487412393093, + 0.006271060090512037, + 0.027678722515702248, + 0.003134789178147912, + 0.0030335683841258287, + 0.0026884404942393303, + 0.0021486571058630943, + 0.004243837203830481, + 0.04600939527153969, + 0.0032471909653395414, + 0.0438871793448925, + 0.0009069820516742766, + 0.0012644382659345865, + 0.0008744667284190655, + 0.0009324673446826637, + 0.0024851274210959673, + 0.0015401528216898441, + 0.0018324415432289243, + 0.0019295387901365757, + 0.0029741236940026283, + 0.005187023431062698, + 0.0015026540495455265, + 0.0020239772275090218, + 0.013060168363153934, + 0.0050787366926670074, + 0.00526136439293623, + 0.005437093321233988, + 0.005876274313777685, + 0.00239127897657454, + 0.005475783720612526, + 0.0020092467311769724, + 0.00813282374292612, + 0.12890511751174927, + 0.004254253581166267, + 0.12419889867305756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.03690085560083389 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020614420995116234 + ], + "opposite_logprob_tokens": [ + -3.8956143856048584 + ], + "logprob_reward": -0.020614420995116234, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020614420995116234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08299050480127335, + 0.1395568698644638, + 0.05058840662240982, + 0.0559379979968071, + 0.14403589069843292, + 0.0778094008564949, + 0.035952840000391006, + 0.015502695925533772, + 0.039542607963085175, + 0.35808271169662476 + ], + "credit_map": [ + 0.007892905734479427, + 0.007667758036404848, + 0.006636577192693949, + 0.00875470694154501, + 0.008015410974621773, + 0.018671181052923203, + 0.008418554440140724, + 0.016933409497141838, + 0.0034973504953086376, + 0.0025657746009528637, + 0.0035041242372244596, + 0.0024699647910892963, + 0.004821120761334896, + 0.06211306154727936, + 0.0031813429668545723, + 0.057404130697250366, + 0.0043340157717466354, + 0.007416358683258295, + 0.003039198461920023, + 0.003303373930975795, + 0.005629881285130978, + 0.010888531804084778, + 0.00487781735137105, + 0.011099233292043209, + 0.0025992647279053926, + 0.002869858406484127, + 0.001649930840358138, + 0.0016527425032109022, + 0.006782974116504192, + 0.018527671694755554, + 0.003502781270071864, + 0.018352773040533066, + 0.00467336643487215, + 0.005718579050153494, + 0.004151333123445511, + 0.004678708966821432, + 0.011492007412016392, + 0.05578159913420677, + 0.005506517831236124, + 0.05203378573060036, + 0.0064536468125879765, + 0.010017244145274162, + 0.003551644505932927, + 0.004425729159265757, + 0.019839538261294365, + 0.01111417543143034, + 0.011969847604632378, + 0.010437576100230217, + 0.0019026362570002675, + 0.0014928028685972095, + 0.0018982910551130772, + 0.0010572097962722182, + 0.001732171163894236, + 0.013407711870968342, + 0.0014629996148869395, + 0.012999014928936958, + 0.0013241032138466835, + 0.0018834867514669895, + 0.0008288488606922328, + 0.0009659001370891929, + 0.0038512444589287043, + 0.0017857888014987111, + 0.002501721028238535, + 0.0023616026155650616, + 0.0033272900618612766, + 0.005403563380241394, + 0.0015733581967651844, + 0.0020244987681508064, + 0.012495326809585094, + 0.005257718730717897, + 0.004594879224896431, + 0.004865975119173527, + 0.007369284518063068, + 0.0030127055943012238, + 0.006935983430594206, + 0.002444096840918064, + 0.010216272436082363, + 0.1644938439130783, + 0.005462806671857834, + 0.15814770758152008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.020614420995116234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019463420612737536 + ], + "opposite_logprob_tokens": [ + -6.251946449279785 + ], + "logprob_reward": -0.0019463420612737536, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019463420612737536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13304157555103302, + 0.20857462286949158, + 0.05651060864329338, + 0.05959148705005646, + 0.03007141500711441, + 0.04430519416928291, + 0.14423415064811707, + 0.010099316015839577, + 0.016119228675961494, + 0.2974523901939392 + ], + "credit_map": [ + 0.01372571475803852, + 0.015320003032684326, + 0.01396356150507927, + 0.013706902042031288, + 0.017712460830807686, + 0.023281440138816833, + 0.015170233324170113, + 0.020161254331469536, + 0.008311934769153595, + 0.006491850130259991, + 0.00947768148034811, + 0.005854596383869648, + 0.01018337719142437, + 0.08349639922380447, + 0.006194202695041895, + 0.07856456935405731, + 0.007838493213057518, + 0.0061198389157652855, + 0.006369537208229303, + 0.005770116113126278, + 0.0069687929935753345, + 0.008678629994392395, + 0.006906348280608654, + 0.007858846336603165, + 0.005558561068028212, + 0.0053721219301223755, + 0.0075723291374742985, + 0.005367950536310673, + 0.01023812871426344, + 0.010301795788109303, + 0.0052306936122477055, + 0.009949900209903717, + 0.0030836621299386024, + 0.003007706021890044, + 0.0028897984884679317, + 0.0023339383769780397, + 0.0069959345273673534, + 0.004145618062466383, + 0.0031402220483869314, + 0.004474533721804619, + 0.0035784104838967323, + 0.0025650118477642536, + 0.002338861580938101, + 0.0016362355090677738, + 0.0027607816737145185, + 0.014553523622453213, + 0.0029726552311331034, + 0.01389971375465393, + 0.0069610984064638615, + 0.008205046877264977, + 0.00923385750502348, + 0.004249244928359985, + 0.01514241099357605, + 0.04722704738378525, + 0.007797084748744965, + 0.04541834071278572, + 0.0010768776992335916, + 0.0010962082305923104, + 0.0011088442988693714, + 0.0007850626134313643, + 0.000943387858569622, + 0.002196284243836999, + 0.0008312360150739551, + 0.00206141360104084, + 0.001507676555775106, + 0.0012277141213417053, + 0.0013993231114000082, + 0.001236122683621943, + 0.002805727533996105, + 0.0031924594659358263, + 0.001457549398764968, + 0.003292654873803258, + 0.005624206271022558, + 0.0024012660142034292, + 0.005718973930925131, + 0.002131369896233082, + 0.009144974872469902, + 0.13641515374183655, + 0.004798301961272955, + 0.1312181055545807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.0019463420612737536 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978908035904169 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.00978908035904169, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978908035904169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1375970095396042, + 0.32373642921447754, + 0.07977726310491562, + 0.028457051143050194, + 0.05379732698202133, + 0.02225438691675663, + 0.10532329231500626, + 0.03381428122520447, + 0.015458202920854092, + 0.19978472590446472 + ], + "credit_map": [ + 0.01860892027616501, + 0.013775208964943886, + 0.02020428329706192, + 0.014146207831799984, + 0.01541139092296362, + 0.020016906782984734, + 0.01513700932264328, + 0.020297087728977203, + 0.010383854620158672, + 0.009849964641034603, + 0.011883602477610111, + 0.008533366024494171, + 0.014535064809024334, + 0.13513065874576569, + 0.007370482198894024, + 0.1260494738817215, + 0.0064047640189528465, + 0.01082143746316433, + 0.005841637030243874, + 0.006499189883470535, + 0.020167164504528046, + 0.009596116840839386, + 0.00855366513133049, + 0.01189329382032156, + 0.001934151747263968, + 0.003067037323489785, + 0.001909542246721685, + 0.0023545522708445787, + 0.006431601010262966, + 0.004233821295201778, + 0.0036677622701972723, + 0.004858584608882666, + 0.0031810523942112923, + 0.0017980360426008701, + 0.002821314614266157, + 0.001302020507864654, + 0.002943709259852767, + 0.02027840167284012, + 0.0024654113221913576, + 0.01900738850235939, + 0.0024722227826714516, + 0.002615067409351468, + 0.0025179239455610514, + 0.00204143556766212, + 0.0033077276311814785, + 0.003611571853980422, + 0.0026958452071994543, + 0.00299259414896369, + 0.00593135692179203, + 0.004547433461993933, + 0.004424622282385826, + 0.004322791006416082, + 0.006727029103785753, + 0.038686808198690414, + 0.0038287239149212837, + 0.03685453534126282, + 0.0026349457912147045, + 0.0038571679033339024, + 0.002402621554210782, + 0.002129222033545375, + 0.00869512278586626, + 0.0040809945203363895, + 0.004918643739074469, + 0.005095565691590309, + 0.0011301604099571705, + 0.001361413043923676, + 0.0015994812129065394, + 0.001102132024243474, + 0.002784186974167824, + 0.0029667632188647985, + 0.001659175381064415, + 0.002854892984032631, + 0.004018926061689854, + 0.002765331417322159, + 0.0041933562606573105, + 0.0017242133617401123, + 0.01000236440449953, + 0.0881243571639061, + 0.004794455133378506, + 0.08416175097227097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.00978908035904169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly craftsman with thinning silver hair and a deeply lined face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a washed-out, dark blue cotton double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman sitting hunched over a heavy wooden workbench covered in sawdust and carving tools?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman using a sharp carving knife to create thin shavings from a wooden sculpture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging tungsten bulb casting deep shadows and strong contrast?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545241974294186 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545241974294186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545241974294186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11393459886312485, + 0.19748537242412567, + 0.06663689017295837, + 0.09729364514350891, + 0.11651036143302917, + 0.07274702191352844, + 0.036054663360118866, + 0.08303789049386978, + 0.024589091539382935, + 0.19171050190925598 + ], + "credit_map": [ + 0.02232564054429531, + 0.010953717865049839, + 0.016616828739643097, + 0.013070736080408096, + 0.011063706129789352, + 0.013712422922253609, + 0.010144568979740143, + 0.016046978533267975, + 0.005331415683031082, + 0.0048445481806993484, + 0.006631823256611824, + 0.003287886967882514, + 0.006403706967830658, + 0.0857660174369812, + 0.005557972472161055, + 0.07966200262308121, + 0.007744458504021168, + 0.007605758961290121, + 0.006484809797257185, + 0.0044482396915555, + 0.01650923490524292, + 0.008332804776728153, + 0.0062562511302530766, + 0.009255331940948963, + 0.005973567720502615, + 0.005486740730702877, + 0.0049035921692848206, + 0.0036153760738670826, + 0.013731975108385086, + 0.029482310637831688, + 0.006229365710169077, + 0.02787071280181408, + 0.0050711603835225105, + 0.003564798505976796, + 0.0047949254512786865, + 0.0028566624969244003, + 0.006057410035282373, + 0.04614756628870964, + 0.0035992690827697515, + 0.044418565928936005, + 0.005283921491354704, + 0.004033421631902456, + 0.005384917836636305, + 0.005296465940773487, + 0.005404917523264885, + 0.021933870390057564, + 0.003900985699146986, + 0.02150852233171463, + 0.0032373243011534214, + 0.0030129672959446907, + 0.003009570064023137, + 0.002435023197904229, + 0.005445427726954222, + 0.008113897405564785, + 0.002747338032349944, + 0.00805311556905508, + 0.010409333743155003, + 0.007302799727767706, + 0.010166150517761707, + 0.007948102429509163, + 0.00867187324911356, + 0.015405592508614063, + 0.008145458996295929, + 0.014988573268055916, + 0.002372348215430975, + 0.002297423779964447, + 0.002505878685042262, + 0.0015734991757199168, + 0.005215836688876152, + 0.004277582745999098, + 0.0026172990910708904, + 0.0037292230408638716, + 0.003851143643260002, + 0.001818358083255589, + 0.0038045707624405622, + 0.0017577147809788585, + 0.005592047702521086, + 0.08733616024255753, + 0.0034908701200038195, + 0.08405962586402893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.012545241974294186 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_9/index_tensor([8776, 8776])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_9/index_tensor([8776, 8776])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e52392a401fd0b13f85f5c13796f8c5ac9cf116c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_9/index_tensor([8776, 8776])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level wide-angle shot captures a scene in a vast, sterile clinical setting, characterized by a seamless, polished white floor that curves into matching white walls, creating an infinite, featureless space. At the center of this space lies a creature resembling a large, bipedal wolf in a prone position. Its powerful, digitally rendered legs are folded beneath a muscular torso covered in a thick, coarse, salt-and-pepper fur, the texture rendered in hyper-realistic detail, displaying individual gray, black, and white lines. Its long, humanoid arms end in black claws, which rest flat on the reflective floor. Its head is lowered, its long snout almost touching the ground, and its pointed ears flattened against its skull. The creature's large, unusually bushy tail, a dense coat of dark gray fur, lies softly on the floor, curling slightly around its flank. The entire scene is illuminated by a bright, even, and completely shadowless light from an unseen overhead source, meticulously emphasizing the complex topography of the creature's unusual anatomy and the dense texture of its fur, reinforcing the cool, impersonal, and information-rich atmosphere of scientific literature.", + "seed": 9920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_9/index_8776_seed_9920_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.43300788355991243, + "vlm_logprob": -0.43300788355991243, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.43300788355991243, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9105465412139893 + ], + "opposite_logprob_tokens": [ + -0.16054651141166687 + ], + "logprob_reward": -1.9105465412139893, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9105465412139893, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.34427475929260254, + 0.09025995433330536, + 0.04570144787430763, + 0.05906882882118225, + 0.09376846253871918, + 0.02643333189189434, + 0.026498982682824135, + 0.02523767575621605, + 0.058001548051834106, + 0.23075494170188904 + ], + "credit_map": [ + 0.08362424373626709, + 0.03694256395101547, + 0.03252829983830452, + 0.020908890292048454, + 0.02623799256980419, + 0.06019054353237152, + 0.024732515215873718, + 0.059109754860401154, + 0.015924984589219093, + 0.009060433134436607, + 0.00989031232893467, + 0.011028294451534748, + 0.009960681200027466, + 0.013979296199977398, + 0.007542220875620842, + 0.012873743660748005, + 0.006233123131096363, + 0.004545003175735474, + 0.007186864968389273, + 0.004251531790941954, + 0.004183607175946236, + 0.006235011853277683, + 0.004847296979278326, + 0.00821901299059391, + 0.00806459877640009, + 0.008647985756397247, + 0.00667089968919754, + 0.006698598153889179, + 0.0069523961283266544, + 0.008564365096390247, + 0.006365299690514803, + 0.007104693446308374, + 0.009904051199555397, + 0.009939649142324924, + 0.019688265398144722, + 0.013370446860790253, + 0.008885571733117104, + 0.015811996534466743, + 0.007129942998290062, + 0.009038551710546017, + 0.002950964728370309, + 0.003190494142472744, + 0.003270781831815839, + 0.003168688854202628, + 0.003486819099634886, + 0.004508644808083773, + 0.0023129598703235388, + 0.003543980186805129, + 0.0030934715177863836, + 0.0038523138500750065, + 0.0030571219976991415, + 0.0024900122079998255, + 0.00329536572098732, + 0.004257790744304657, + 0.0031953249126672745, + 0.0032575849909335375, + 0.004394040443003178, + 0.003114450490102172, + 0.0021266546100378036, + 0.003152616322040558, + 0.00250296201556921, + 0.00440456997603178, + 0.003246425185352564, + 0.0022959590423852205, + 0.008281571790575981, + 0.007818321697413921, + 0.005697511602193117, + 0.013241905719041824, + 0.00531349889934063, + 0.008936247788369656, + 0.005077081732451916, + 0.0036354134790599346, + 0.04666164889931679, + 0.03178730607032776, + 0.03163956478238106, + 0.02032698690891266, + 0.028361428529024124, + 0.03897285461425781, + 0.016776099801063538, + 0.01622908003628254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -1.9105465412139893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709592282772064 + ], + "opposite_logprob_tokens": [ + -2.127095937728882 + ], + "logprob_reward": -0.12709592282772064, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709592282772064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27347442507743835, + 0.10600456595420837, + 0.06596029549837112, + 0.084698386490345, + 0.07521245628595352, + 0.054299890995025635, + 0.0273811724036932, + 0.11323888599872589, + 0.061342597007751465, + 0.1383872777223587 + ], + "credit_map": [ + 0.04815993830561638, + 0.0375140979886055, + 0.027605175971984863, + 0.02297188900411129, + 0.03264186158776283, + 0.0358896479010582, + 0.02654438652098179, + 0.0421474389731884, + 0.013150247745215893, + 0.01495776604861021, + 0.013195187784731388, + 0.018475959077477455, + 0.00924000982195139, + 0.012424487620592117, + 0.012652166187763214, + 0.01190874632447958, + 0.00819859467446804, + 0.01427206676453352, + 0.006167302839457989, + 0.007321555633097887, + 0.00705081969499588, + 0.00804152898490429, + 0.007191612850874662, + 0.007716819178313017, + 0.01055429968982935, + 0.01268664188683033, + 0.010299107059836388, + 0.011232330463826656, + 0.012270152568817139, + 0.012185686267912388, + 0.008573995903134346, + 0.006896178238093853, + 0.006699204910546541, + 0.011586137115955353, + 0.0157484021037817, + 0.008719073608517647, + 0.006490263156592846, + 0.01295800693333149, + 0.005728254094719887, + 0.007283113431185484, + 0.007913484238088131, + 0.006223645526915789, + 0.006856605876237154, + 0.006736647803336382, + 0.00580156035721302, + 0.008786706253886223, + 0.005521934479475021, + 0.006459311116486788, + 0.003339417278766632, + 0.003773629432544112, + 0.002802335424348712, + 0.0028442346956580877, + 0.003082512877881527, + 0.004492930136620998, + 0.0038938543293625116, + 0.003152258228510618, + 0.009465242736041546, + 0.019008655101060867, + 0.011165143921971321, + 0.018201682716608047, + 0.013960734009742737, + 0.020423827692866325, + 0.014256380498409271, + 0.006757220719009638, + 0.010412756353616714, + 0.009345965459942818, + 0.005829791072756052, + 0.0071522751823067665, + 0.006149877328425646, + 0.010130485519766808, + 0.005380774848163128, + 0.0069406707771122456, + 0.034075312316417694, + 0.01866726763546467, + 0.015531686134636402, + 0.01148979365825653, + 0.013642636127769947, + 0.024081114679574966, + 0.00983367394655943, + 0.011065785773098469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.12709592282772064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643027231097221 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.007643027231097221, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643027231097221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29493874311447144, + 0.08486739546060562, + 0.05340046063065529, + 0.06800125539302826, + 0.08833421021699905, + 0.033275846391916275, + 0.07713956385850906, + 0.08433623611927032, + 0.04597385600209236, + 0.16973239183425903 + ], + "credit_map": [ + 0.071623295545578, + 0.03013474866747856, + 0.025118285790085793, + 0.02117851935327053, + 0.020980577915906906, + 0.05245327204465866, + 0.023585407063364983, + 0.04986463487148285, + 0.01367095299065113, + 0.007127034477889538, + 0.009752321988344193, + 0.012224260717630386, + 0.010022017173469067, + 0.012800747528672218, + 0.007920175790786743, + 0.011349888518452644, + 0.006077626254409552, + 0.006396444048732519, + 0.005777047947049141, + 0.0073968651704490185, + 0.006291989237070084, + 0.0063573638908565044, + 0.00730335945263505, + 0.007799762301146984, + 0.009268689900636673, + 0.01008653175085783, + 0.008061458356678486, + 0.007175390142947435, + 0.008372155018150806, + 0.01003494393080473, + 0.007121657021343708, + 0.007880426943302155, + 0.009531107731163502, + 0.011372627690434456, + 0.017326580360531807, + 0.011903765611350536, + 0.008742068894207478, + 0.014460110105574131, + 0.006725402548909187, + 0.00827255379408598, + 0.004520241636782885, + 0.0037753996439278126, + 0.004388113971799612, + 0.004152690526098013, + 0.0034978019539266825, + 0.005757409147918224, + 0.0029364668298512697, + 0.004247721750289202, + 0.006847791839390993, + 0.013005353510379791, + 0.00984868872910738, + 0.008961737155914307, + 0.010586380958557129, + 0.013315071351826191, + 0.0075626736506819725, + 0.007011865731328726, + 0.008438368327915668, + 0.013753777369856834, + 0.009597943164408207, + 0.010789788328111172, + 0.009753033518791199, + 0.014797873795032501, + 0.011569689959287643, + 0.005635763518512249, + 0.007456314750015736, + 0.005936746019870043, + 0.004876383580267429, + 0.009060381911695004, + 0.004460193682461977, + 0.007089745718985796, + 0.0038733305409550667, + 0.003220760729163885, + 0.03680810332298279, + 0.024185502901673317, + 0.01513324212282896, + 0.015936221927404404, + 0.02077108435332775, + 0.0291131604462862, + 0.014121688902378082, + 0.013663378544151783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.007643027231097221 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11301986873149872 + ], + "opposite_logprob_tokens": [ + -2.2380199432373047 + ], + "logprob_reward": -0.11301986873149872, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11301986873149872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17769189178943634, + 0.16864141821861267, + 0.04023009538650513, + 0.060160037130117416, + 0.06238216161727905, + 0.04092439264059067, + 0.07091004401445389, + 0.12715424597263336, + 0.03660650923848152, + 0.21529921889305115 + ], + "credit_map": [ + 0.04219679534435272, + 0.017099928110837936, + 0.017768358811736107, + 0.01418246142566204, + 0.012576132081449032, + 0.03062351420521736, + 0.010606936179101467, + 0.03263776749372482, + 0.02316371351480484, + 0.021486347541213036, + 0.021130602806806564, + 0.01959332264959812, + 0.021646026521921158, + 0.021940449252724648, + 0.018166258931159973, + 0.021514689549803734, + 0.005397176835685968, + 0.004684083163738251, + 0.004344184882938862, + 0.005122532602399588, + 0.0032623119186609983, + 0.005155283957719803, + 0.00502610020339489, + 0.007238420192152262, + 0.008701829239726067, + 0.009075465612113476, + 0.007942219264805317, + 0.006800995208323002, + 0.007494631689041853, + 0.00784057006239891, + 0.006654130294919014, + 0.00565019017085433, + 0.00714949332177639, + 0.0067100245505571365, + 0.012779600918293, + 0.008406007662415504, + 0.006312953773885965, + 0.01044965535402298, + 0.004476457368582487, + 0.006097969599068165, + 0.0036695795133709908, + 0.0060888370499014854, + 0.005994714796543121, + 0.005009053274989128, + 0.004912978038191795, + 0.006964425090700388, + 0.0037882444448769093, + 0.004496559035032988, + 0.008230182342231274, + 0.0114633459597826, + 0.008317592553794384, + 0.00776643818244338, + 0.009054016321897507, + 0.011366419494152069, + 0.0069826277904212475, + 0.0077294171787798405, + 0.013446885161101818, + 0.02110869437456131, + 0.014046685770154, + 0.015774378553032875, + 0.01597321406006813, + 0.02306702733039856, + 0.013296672143042088, + 0.010440679267048836, + 0.003783534048125148, + 0.004675513133406639, + 0.003141014836728573, + 0.009012005291879177, + 0.003355183871462941, + 0.006384164094924927, + 0.0034636533819139004, + 0.002791438950225711, + 0.04265047609806061, + 0.028501000255346298, + 0.030882032588124275, + 0.017221780493855476, + 0.02761419676244259, + 0.035079874098300934, + 0.01920262910425663, + 0.014147208072245121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.11301986873149872 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24908441305160522, + 0.09822415560483932, + 0.04763193801045418, + 0.054353613406419754, + 0.09129641950130463, + 0.045535121113061905, + 0.046781472861766815, + 0.03331206738948822, + 0.0637240931391716, + 0.27005669474601746 + ], + "credit_map": [ + 0.06256575137376785, + 0.022784225642681122, + 0.021374138072133064, + 0.01911425031721592, + 0.016381852328777313, + 0.04586343094706535, + 0.01849343068897724, + 0.04250732436776161, + 0.012625043280422688, + 0.014195387251675129, + 0.008782188408076763, + 0.012252798303961754, + 0.012989222072064877, + 0.012799528427422047, + 0.01193115022033453, + 0.012648840434849262, + 0.0069382027722895145, + 0.005613734479993582, + 0.004649200011044741, + 0.004039255436509848, + 0.007632870227098465, + 0.00626791175454855, + 0.005780928768217564, + 0.006709833163768053, + 0.006511112209409475, + 0.009134193882346153, + 0.005167731549590826, + 0.007043994031846523, + 0.00672883540391922, + 0.007958024740219116, + 0.005978844594210386, + 0.005830875597894192, + 0.009499871172010899, + 0.011485756374895573, + 0.018310237675905228, + 0.012283884920179844, + 0.008609631098806858, + 0.015345527790486813, + 0.007344772107899189, + 0.008416741155087948, + 0.0055820210836827755, + 0.007411369122564793, + 0.005416086874902248, + 0.00458736764267087, + 0.006255424115806818, + 0.007819523103535175, + 0.004394978284835815, + 0.004068346228450537, + 0.004176630638539791, + 0.007797648198902607, + 0.006231104023754597, + 0.005116882734000683, + 0.005751218181103468, + 0.008159724995493889, + 0.004848578944802284, + 0.004699685610830784, + 0.003306035418063402, + 0.005197346676141024, + 0.0032878026831895113, + 0.004649326670914888, + 0.003631686558946967, + 0.005854516290128231, + 0.004851981066167355, + 0.0025333696976304054, + 0.009737427346408367, + 0.008453990332782269, + 0.006304739974439144, + 0.014414288103580475, + 0.005852089263498783, + 0.00982325803488493, + 0.005672149360179901, + 0.00346614932641387, + 0.05555512756109238, + 0.035832107067108154, + 0.03575378283858299, + 0.02302583120763302, + 0.0338781401515007, + 0.04528220742940903, + 0.019292525947093964, + 0.021436970680952072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.006734057795256376 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures a scene in a vast, sterile clinical setting, characterized by a seamless, polished white floor that curves into matching white walls, creating an infinite, featureless space. At the center of this space lies a creature resembling a large, bipedal wolf in a prone position. Its powerful, digitally rendered legs are folded beneath a muscular torso covered in a thick, coarse, salt-and-pepper fur, the texture rendered in hyper-realistic detail, displaying individual gray, black, and white lines. Its long, humanoid arms end in black claws, which rest flat on the reflective floor. Its head is lowered, its long snout almost touching the ground, and its pointed ears flattened against its skull. The creature's large, unusually bushy tail, a dense coat of dark gray fur, lies softly on the floor, curling slightly around its flank. The entire scene is illuminated by a bright, even, and completely shadowless light from an unseen overhead source, meticulously emphasizing the complex topography of the creature's unusual anatomy and the dense texture of its fur, reinforcing the cool, impersonal, and information-rich atmosphere of scientific literature.", + "seed": 9921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_9/index_8776_seed_9921_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1762015056796372, + "vlm_logprob": -0.1762015056796372, + "vlm_yes_no_margin": 3.149999976158142, + "vlm_token_credit": -0.1762015056796372, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6939447522163391 + ], + "opposite_logprob_tokens": [ + -0.6939447522163391 + ], + "logprob_reward": -0.6939447522163391, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6939447522163391, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3590737283229828, + 0.09219060838222504, + 0.022956056520342827, + 0.09396356344223022, + 0.08173494040966034, + 0.029999468475580215, + 0.12878358364105225, + 0.04671700671315193, + 0.03245876356959343, + 0.1121223047375679 + ], + "credit_map": [ + 0.04362507164478302, + 0.026085713878273964, + 0.035043537616729736, + 0.030542155727744102, + 0.029602782800793648, + 0.07461246103048325, + 0.03532418608665466, + 0.08423781394958496, + 0.009785597212612629, + 0.005160490516573191, + 0.005751546006649733, + 0.00484899478033185, + 0.005915310699492693, + 0.02192048914730549, + 0.004291628021746874, + 0.034516558051109314, + 0.0025129972491413355, + 0.0027447794564068317, + 0.0026301732286810875, + 0.003106028540059924, + 0.0019585900008678436, + 0.003111880039796233, + 0.0017346938839182258, + 0.005156914237886667, + 0.010213719680905342, + 0.004345654975622892, + 0.024830231443047523, + 0.00812052097171545, + 0.004063781350851059, + 0.0068254005163908005, + 0.006142093800008297, + 0.029422152787446976, + 0.006190034560859203, + 0.014817865565419197, + 0.008290517143905163, + 0.00785575620830059, + 0.012907244265079498, + 0.009327651932835579, + 0.012284823693335056, + 0.01006105076521635, + 0.002845549723133445, + 0.004347238689661026, + 0.0030019597616046667, + 0.0031423475593328476, + 0.003776974743232131, + 0.0050130607560276985, + 0.0031161969527602196, + 0.004756138660013676, + 0.01076109055429697, + 0.006721072364598513, + 0.02754114381968975, + 0.05072448030114174, + 0.005723673850297928, + 0.007666942197829485, + 0.011603697203099728, + 0.008041474968194962, + 0.005052078980952501, + 0.006508467253297567, + 0.008879739791154861, + 0.006277485750615597, + 0.008615908212959766, + 0.0032608420588076115, + 0.004089012276381254, + 0.004033470526337624, + 0.0028388863429427147, + 0.004289522301405668, + 0.005592138506472111, + 0.004781465977430344, + 0.005235737189650536, + 0.0029711874667555094, + 0.0032030418515205383, + 0.0035467830020934343, + 0.018911341205239296, + 0.013066913932561874, + 0.013836990110576153, + 0.02161361277103424, + 0.007677823305130005, + 0.008529137820005417, + 0.012790839187800884, + 0.01569565385580063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.6939447522163391 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033704210072755814 + ], + "opposite_logprob_tokens": [ + -3.4087042808532715 + ], + "logprob_reward": -0.033704210072755814, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033704210072755814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2732696533203125, + 0.15884068608283997, + 0.05840397626161575, + 0.04979395493865013, + 0.046159375458955765, + 0.02892269939184189, + 0.31886497139930725, + 0.02831924334168434, + 0.012520769611001015, + 0.024904627352952957 + ], + "credit_map": [ + 0.025979934260249138, + 0.03297436237335205, + 0.028598438948392868, + 0.03497369587421417, + 0.025348510593175888, + 0.05167974904179573, + 0.03909797966480255, + 0.03461699187755585, + 0.011498394422233105, + 0.023785840719938278, + 0.010427541099488735, + 0.018673332408070564, + 0.02158665843307972, + 0.02374526858329773, + 0.02403213083744049, + 0.025091510266065598, + 0.007574844639748335, + 0.004645721986889839, + 0.00883532129228115, + 0.005380669608712196, + 0.005259722471237183, + 0.007571840193122625, + 0.0039446535520255566, + 0.01519120391458273, + 0.004184400662779808, + 0.003734591184183955, + 0.010279993526637554, + 0.004989671986550093, + 0.004416040610522032, + 0.004392561037093401, + 0.005022919271141291, + 0.012773779220879078, + 0.00393696129322052, + 0.007214335259050131, + 0.004670296795666218, + 0.006040152627974749, + 0.004928741604089737, + 0.006782648619264364, + 0.006551914382725954, + 0.006034323945641518, + 0.003285411512479186, + 0.003380000591278076, + 0.00575987109914422, + 0.004610735457390547, + 0.002940072910860181, + 0.003041845513507724, + 0.003040963551029563, + 0.0028637985233217478, + 0.02429739013314247, + 0.020668700337409973, + 0.06391756981611252, + 0.11686965823173523, + 0.016783447936177254, + 0.02682490088045597, + 0.025883415713906288, + 0.02361988089978695, + 0.0026383709628134966, + 0.004514370579272509, + 0.004225792828947306, + 0.003493497148156166, + 0.004796287510544062, + 0.0025939312763512135, + 0.0034249802120029926, + 0.002632011193782091, + 0.0012288173893466592, + 0.002108774147927761, + 0.0019041233463212848, + 0.001926634693518281, + 0.0015169252874329686, + 0.001119357766583562, + 0.0012602444039657712, + 0.0014558929251506925, + 0.0037066659424453974, + 0.003132022451609373, + 0.0030699779745191336, + 0.004138332791626453, + 0.003498375415802002, + 0.0028879230376333, + 0.001997958170250058, + 0.0024733711034059525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.033704210072755814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255336407572031 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255336407572031, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255336407572031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31999287009239197, + 0.043487388640642166, + 0.04227206110954285, + 0.11456042528152466, + 0.05111490190029144, + 0.036320578306913376, + 0.2291918843984604, + 0.06552047282457352, + 0.03424733132123947, + 0.06329210102558136 + ], + "credit_map": [ + 0.031642306596040726, + 0.024772286415100098, + 0.02209307812154293, + 0.03017565794289112, + 0.024110985919833183, + 0.07029656320810318, + 0.038727644830942154, + 0.07817435264587402, + 0.004716834519058466, + 0.004290203098207712, + 0.0038451445288956165, + 0.006706377491354942, + 0.0056082019582390785, + 0.0050432090647518635, + 0.007802899461239576, + 0.005474518984556198, + 0.0050879670307040215, + 0.004390999209135771, + 0.004951760172843933, + 0.006530753802508116, + 0.004485101439058781, + 0.006966317538172007, + 0.003986309748142958, + 0.005872852634638548, + 0.00900965090841055, + 0.009837253019213676, + 0.020490054041147232, + 0.016145596280694008, + 0.010850532911717892, + 0.01041661761701107, + 0.012869812548160553, + 0.02494090422987938, + 0.00428109522908926, + 0.0070522017776966095, + 0.005181384272873402, + 0.006636721547693014, + 0.00801808200776577, + 0.00604664022102952, + 0.006397569552063942, + 0.0075012082234025, + 0.004042170476168394, + 0.005175690166652203, + 0.0037108873948454857, + 0.005635262932628393, + 0.005540535319596529, + 0.003268835134804249, + 0.003911577630788088, + 0.005035618785768747, + 0.017457133159041405, + 0.01692037284374237, + 0.04403016343712807, + 0.08046071976423264, + 0.015482536517083645, + 0.01686246320605278, + 0.01939212903380394, + 0.01858634687960148, + 0.007295895367860794, + 0.007963157258927822, + 0.011298454366624355, + 0.010524815879762173, + 0.009090354666113853, + 0.005484337918460369, + 0.0068502104841172695, + 0.007013250607997179, + 0.0031578983180224895, + 0.004971685819327831, + 0.005773712415248156, + 0.0042665577493608, + 0.00654318043962121, + 0.003201712854206562, + 0.0030645423103123903, + 0.0032680367585271597, + 0.012417081743478775, + 0.005968424491584301, + 0.008236528374254704, + 0.012042531743645668, + 0.00581754045560956, + 0.005744623951613903, + 0.005737876053899527, + 0.007327494211494923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.005255336407572031 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428559571504593 + ], + "opposite_logprob_tokens": [ + -2.0178558826446533 + ], + "logprob_reward": -0.1428559571504593, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428559571504593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15586984157562256, + 0.11512051522731781, + 0.04766238480806351, + 0.23421436548233032, + 0.046368543058633804, + 0.062306471168994904, + 0.15928560495376587, + 0.028442978858947754, + 0.06352926790714264, + 0.08720006793737411 + ], + "credit_map": [ + 0.0125477509573102, + 0.023417383432388306, + 0.01369090098887682, + 0.024432603269815445, + 0.017649389803409576, + 0.016435332596302032, + 0.030548274517059326, + 0.017148196697235107, + 0.01320713385939598, + 0.007861537858843803, + 0.00976252555847168, + 0.008357681334018707, + 0.008490914478898048, + 0.0191936194896698, + 0.010398219339549541, + 0.037848882377147675, + 0.005364689044654369, + 0.004996906500309706, + 0.006037408020347357, + 0.006842122878879309, + 0.00369299971498549, + 0.006041552871465683, + 0.0043439846485853195, + 0.01034272089600563, + 0.026952700689435005, + 0.010846013203263283, + 0.06305306404829025, + 0.020216014236211777, + 0.011177416890859604, + 0.016491517424583435, + 0.013937600888311863, + 0.07154002040624619, + 0.004714472219347954, + 0.006916056387126446, + 0.004833388142287731, + 0.007811833173036575, + 0.005631385371088982, + 0.005500406492501497, + 0.0047043864615261555, + 0.006256614811718464, + 0.0062076980248093605, + 0.008846918120980263, + 0.005700601730495691, + 0.009822563268244267, + 0.008304759860038757, + 0.007378541864454746, + 0.006375339347869158, + 0.009670042432844639, + 0.012773645110428333, + 0.011177095584571362, + 0.030318820849061012, + 0.056101806461811066, + 0.009391719475388527, + 0.015447677113115788, + 0.013797465711832047, + 0.010277364403009415, + 0.002886225702241063, + 0.00434162886813283, + 0.004605963360518217, + 0.0042775338515639305, + 0.002945122541859746, + 0.002890828298404813, + 0.0038415163289755583, + 0.00265415757894516, + 0.006172549445182085, + 0.010370816104114056, + 0.011726763099431992, + 0.007973936386406422, + 0.009564713574945927, + 0.006533464882522821, + 0.0054587190970778465, + 0.005728298332542181, + 0.01816842518746853, + 0.00845584087073803, + 0.012800044380128384, + 0.017201783135533333, + 0.005691612139344215, + 0.005857241339981556, + 0.007903852500021458, + 0.011121260933578014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.1428559571504593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247272551059723 + ], + "opposite_logprob_tokens": [ + -5.255247116088867 + ], + "logprob_reward": -0.005247272551059723, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247272551059723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40903240442276, + 0.07991056144237518, + 0.08460859209299088, + 0.07685394585132599, + 0.061725448817014694, + 0.03486122190952301, + 0.07825565338134766, + 0.053906865417957306, + 0.0537567175924778, + 0.06708859652280807 + ], + "credit_map": [ + 0.04149917885661125, + 0.026850540190935135, + 0.03392671048641205, + 0.03171934932470322, + 0.029701367020606995, + 0.10562548786401749, + 0.031575802713632584, + 0.1081339493393898, + 0.0107899010181427, + 0.008276567794382572, + 0.007345336955040693, + 0.009279717691242695, + 0.00934498943388462, + 0.007287923246622086, + 0.012413921765983105, + 0.015172204002737999, + 0.011790095828473568, + 0.0061713154427707195, + 0.010492758825421333, + 0.010320356115698814, + 0.006555161904543638, + 0.009666618891060352, + 0.0076554687693715096, + 0.021956821903586388, + 0.006803230848163366, + 0.006764702498912811, + 0.014545202255249023, + 0.010753532871603966, + 0.006921797059476376, + 0.006763136014342308, + 0.008995703421533108, + 0.015306642279028893, + 0.004658354911953211, + 0.009809658862650394, + 0.005585826467722654, + 0.007701303344219923, + 0.008937851525843143, + 0.00823840033262968, + 0.008005219511687756, + 0.008788828738033772, + 0.004374505486339331, + 0.005065974313765764, + 0.003493998199701309, + 0.004519128706306219, + 0.005261614453047514, + 0.0037520083133131266, + 0.0035838454496115446, + 0.004810146056115627, + 0.008182252757251263, + 0.00787934847176075, + 0.011320521123707294, + 0.022758077830076218, + 0.006147304084151983, + 0.006666879635304213, + 0.006944397930055857, + 0.008356878533959389, + 0.0057642399333417416, + 0.006936367601156235, + 0.007907804101705551, + 0.008166525512933731, + 0.010767115280032158, + 0.004472269210964441, + 0.004466586746275425, + 0.005425958428531885, + 0.005202277563512325, + 0.009029552340507507, + 0.005754395388066769, + 0.006747410167008638, + 0.009402558207511902, + 0.005237116944044828, + 0.006078503560274839, + 0.006304899696260691, + 0.010920967906713486, + 0.0067826430313289165, + 0.008202905766665936, + 0.013049458153545856, + 0.0066745951771736145, + 0.0061338674277067184, + 0.007386238779872656, + 0.00793792400509119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.005247272551059723 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures a scene in a vast, sterile clinical setting, characterized by a seamless, polished white floor that curves into matching white walls, creating an infinite, featureless space. At the center of this space lies a creature resembling a large, bipedal wolf in a prone position. Its powerful, digitally rendered legs are folded beneath a muscular torso covered in a thick, coarse, salt-and-pepper fur, the texture rendered in hyper-realistic detail, displaying individual gray, black, and white lines. Its long, humanoid arms end in black claws, which rest flat on the reflective floor. Its head is lowered, its long snout almost touching the ground, and its pointed ears flattened against its skull. The creature's large, unusually bushy tail, a dense coat of dark gray fur, lies softly on the floor, curling slightly around its flank. The entire scene is illuminated by a bright, even, and completely shadowless light from an unseen overhead source, meticulously emphasizing the complex topography of the creature's unusual anatomy and the dense texture of its fur, reinforcing the cool, impersonal, and information-rich atmosphere of scientific literature.", + "seed": 9922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_9/index_8776_seed_9922_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17271021446213125, + "vlm_logprob": -0.17271021446213125, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.17271021446213125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6332249641418457 + ], + "opposite_logprob_tokens": [ + -0.7582249641418457 + ], + "logprob_reward": -0.6332249641418457, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6332249641418457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11423241347074509, + 0.04508653283119202, + 0.4010232090950012, + 0.16502457857131958, + 0.021471897140145302, + 0.012757135555148125, + 0.07658493518829346, + 0.08860345184803009, + 0.045153431594371796, + 0.030062373727560043 + ], + "credit_map": [ + 0.01323709823191166, + 0.011939659714698792, + 0.014170091599225998, + 0.02264787070453167, + 0.019582675769925117, + 0.007928316481411457, + 0.013086088933050632, + 0.011640617623925209, + 0.008473331108689308, + 0.003803556552156806, + 0.008189511485397816, + 0.005751077085733414, + 0.004971183370798826, + 0.00558976735919714, + 0.004935286473482847, + 0.0033728224225342274, + 0.05763201043009758, + 0.029121246188879013, + 0.05784538760781288, + 0.03571034222841263, + 0.028754889965057373, + 0.05193936452269554, + 0.06816067546606064, + 0.07185930758714676, + 0.021539228036999702, + 0.012125518172979355, + 0.01963154971599579, + 0.016340577974915504, + 0.026011694222688675, + 0.021027635782957077, + 0.023197155445814133, + 0.025151217356324196, + 0.002908581867814064, + 0.0016856009606271982, + 0.0017716389847919345, + 0.0025870304089039564, + 0.002573114586994052, + 0.0027287236880511045, + 0.0036788200959563255, + 0.003538387594744563, + 0.001539471559226513, + 0.0008887493750080466, + 0.0012838405091315508, + 0.0014249080559238791, + 0.0015256949700415134, + 0.0018434261437505484, + 0.001792625873349607, + 0.002458419417962432, + 0.006781723815947771, + 0.0072536347433924675, + 0.0076245274394750595, + 0.012359479442238808, + 0.007914412766695023, + 0.013387216255068779, + 0.015366559848189354, + 0.005897375289350748, + 0.014914166182279587, + 0.007142124697566032, + 0.011960363015532494, + 0.02638419158756733, + 0.007234270218759775, + 0.008941884152591228, + 0.007104345131665468, + 0.004922109190374613, + 0.006816599052399397, + 0.002977479947730899, + 0.0063971239142119884, + 0.007387400139123201, + 0.0027290068101137877, + 0.003542575053870678, + 0.0035740877501666546, + 0.011729156598448753, + 0.0032736421562731266, + 0.002909163711592555, + 0.003978176508098841, + 0.004784361459314823, + 0.0027479580603539944, + 0.004307287745177746, + 0.004043965600430965, + 0.004017816390842199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.6332249641418457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820182055234909 + ], + "opposite_logprob_tokens": [ + -4.01820182800293 + ], + "logprob_reward": -0.01820182055234909, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820182055234909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15443111956119537, + 0.25427722930908203, + 0.062189336866140366, + 0.04296122491359711, + 0.02959354594349861, + 0.023411814123392105, + 0.21673688292503357, + 0.10570323467254639, + 0.06517235189676285, + 0.04552328586578369 + ], + "credit_map": [ + 0.014495969749987125, + 0.01519102044403553, + 0.021054977551102638, + 0.026591213420033455, + 0.023419834673404694, + 0.013544069603085518, + 0.01925954595208168, + 0.02087448164820671, + 0.03816941753029823, + 0.023922458291053772, + 0.03376167640089989, + 0.04224737361073494, + 0.04238953813910484, + 0.031191546469926834, + 0.023679198697209358, + 0.018915990367531776, + 0.0087932413443923, + 0.008959039114415646, + 0.011535678058862686, + 0.006093543954193592, + 0.007734791375696659, + 0.007195345126092434, + 0.0063001057133078575, + 0.005577589850872755, + 0.005497648846358061, + 0.0031285788863897324, + 0.0037417446728795767, + 0.005271075293421745, + 0.005419409833848476, + 0.005450911354273558, + 0.009879721328616142, + 0.004572134464979172, + 0.0030195850413292646, + 0.002810918726027012, + 0.0031838088762015104, + 0.0039150002412498, + 0.007814573124051094, + 0.0031122134532779455, + 0.0033791835885494947, + 0.0023582594003528357, + 0.0022837684955447912, + 0.0026370782870799303, + 0.003001468488946557, + 0.0031732216011732817, + 0.003147238865494728, + 0.0027719950303435326, + 0.0035306375939399004, + 0.0028664041310548782, + 0.0210188627243042, + 0.019999228417873383, + 0.023601466789841652, + 0.03140480816364288, + 0.023454707115888596, + 0.04168611019849777, + 0.03704497218132019, + 0.018526731058955193, + 0.011738183908164501, + 0.008929627947509289, + 0.01568634808063507, + 0.01560145616531372, + 0.009866254404187202, + 0.01979146897792816, + 0.012959595769643784, + 0.01113029196858406, + 0.007209791801869869, + 0.003330813255161047, + 0.00776932667940855, + 0.016886144876480103, + 0.004214299377053976, + 0.004608519375324249, + 0.004700168501585722, + 0.016453281044960022, + 0.006857016589492559, + 0.0041715167462825775, + 0.009209099225699902, + 0.007152851205319166, + 0.005213907919824123, + 0.004835301078855991, + 0.0038563425187021494, + 0.004227250348776579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.01820182055234909 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255573429167271 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255573429167271, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255573429167271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19775685667991638, + 0.029506932944059372, + 0.18951639533042908, + 0.12520337104797363, + 0.025266248732805252, + 0.03020019456744194, + 0.07129078358411789, + 0.220049649477005, + 0.05813710764050484, + 0.053072430193424225 + ], + "credit_map": [ + 0.02741718664765358, + 0.023958852514624596, + 0.02632615901529789, + 0.03415494039654732, + 0.025397472083568573, + 0.014543743804097176, + 0.025695715099573135, + 0.02026277594268322, + 0.004538333974778652, + 0.0029249952640384436, + 0.004638294689357281, + 0.003372849663719535, + 0.0036775977350771427, + 0.005380711518228054, + 0.002673262497410178, + 0.0023008871357887983, + 0.027652809396386147, + 0.013192703016102314, + 0.030030684545636177, + 0.016369657590985298, + 0.013841791078448296, + 0.021595148369669914, + 0.032430898398160934, + 0.03440270200371742, + 0.016677573323249817, + 0.009401029907166958, + 0.01425273809581995, + 0.011748951859772205, + 0.019348295405507088, + 0.015692908316850662, + 0.01923401653766632, + 0.018847860395908356, + 0.0032500484958291054, + 0.002739917952567339, + 0.0022743416484445333, + 0.003816488664597273, + 0.003715661121532321, + 0.004054537508636713, + 0.00294379610568285, + 0.002471458399668336, + 0.0033490904606878757, + 0.0035427326802164316, + 0.004796684719622135, + 0.004164719954133034, + 0.003520459169521928, + 0.004132458940148354, + 0.0036882709246128798, + 0.0030057772528380156, + 0.006528472062200308, + 0.00816027820110321, + 0.008210673928260803, + 0.009193596430122852, + 0.00940544530749321, + 0.010707488283514977, + 0.012119884602725506, + 0.006964951753616333, + 0.03674270212650299, + 0.01991303265094757, + 0.030898820608854294, + 0.06527766585350037, + 0.016045620664954185, + 0.02142876572906971, + 0.017472736537456512, + 0.012270329520106316, + 0.008559029549360275, + 0.005194411147385836, + 0.006901034153997898, + 0.008586740121245384, + 0.005536398384720087, + 0.005210093222558498, + 0.0042998772114515305, + 0.013849521055817604, + 0.0051889666356146336, + 0.004079854115843773, + 0.00801607221364975, + 0.008813237771391869, + 0.004942983388900757, + 0.009241010993719101, + 0.006854953244328499, + 0.005935356020927429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.005255573429167271 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20162120461463928 + ], + "opposite_logprob_tokens": [ + -1.701621174812317 + ], + "logprob_reward": -0.20162120461463928, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20162120461463928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11730407923460007, + 0.07287673652172089, + 0.4206372797489166, + 0.07233531773090363, + 0.043667394667863846, + 0.024197114631533623, + 0.042018789798021317, + 0.029210953041911125, + 0.07020869851112366, + 0.10754365473985672 + ], + "credit_map": [ + 0.01810791715979576, + 0.008282552473247051, + 0.015394083224236965, + 0.0202798992395401, + 0.011492270044982433, + 0.01271732896566391, + 0.018042126670479774, + 0.012987894006073475, + 0.01169260311871767, + 0.006912360433489084, + 0.009450641460716724, + 0.008946098387241364, + 0.01145092397928238, + 0.011371225118637085, + 0.006922242231667042, + 0.006130638998001814, + 0.060630667954683304, + 0.03511866182088852, + 0.0610630176961422, + 0.03826950117945671, + 0.03143982216715813, + 0.06168486177921295, + 0.06107423081994057, + 0.07135649770498276, + 0.008503520861268044, + 0.00640804972499609, + 0.008115576580166817, + 0.0077551971189677715, + 0.015114879235625267, + 0.010960692539811134, + 0.007906177081167698, + 0.007571210619062185, + 0.005712426733225584, + 0.0036972719244658947, + 0.004203681368380785, + 0.005485640838742256, + 0.007574861869215965, + 0.00634325947612524, + 0.005360234994441271, + 0.005290014203637838, + 0.002664041006937623, + 0.002096502110362053, + 0.003029121086001396, + 0.003379115369170904, + 0.004104968626052141, + 0.0038263769820332527, + 0.0027473880909383297, + 0.002349599963054061, + 0.005321336444467306, + 0.006230717524886131, + 0.005351779982447624, + 0.007125085219740868, + 0.005073937587440014, + 0.004345264285802841, + 0.006092751864343882, + 0.0024779117666184902, + 0.0028803336899727583, + 0.0025158121716231108, + 0.004978782031685114, + 0.0041329870000481606, + 0.003190096467733383, + 0.004014758393168449, + 0.0039679789915680885, + 0.0035302010364830494, + 0.013452403247356415, + 0.00455215061083436, + 0.007120094262063503, + 0.008048639632761478, + 0.004180241376161575, + 0.004921603947877884, + 0.006261392496526241, + 0.021672165021300316, + 0.012600007466971874, + 0.008095639757812023, + 0.016075393185019493, + 0.015307284891605377, + 0.007995008490979671, + 0.012261834926903248, + 0.012593092396855354, + 0.022615380585193634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.20162120461463928 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052475095726549625 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052475095726549625, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052475095726549625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2606874108314514, + 0.03469895198941231, + 0.19438305497169495, + 0.17219385504722595, + 0.02842685393989086, + 0.04050358384847641, + 0.07164546847343445, + 0.09286629408597946, + 0.04723505303263664, + 0.057359471917152405 + ], + "credit_map": [ + 0.028783755376935005, + 0.02805345505475998, + 0.03487519174814224, + 0.04492576792836189, + 0.03966861590743065, + 0.01990707963705063, + 0.03308460861444473, + 0.03138893097639084, + 0.0046775368973612785, + 0.0030615644063800573, + 0.0033570006489753723, + 0.003404824296012521, + 0.005171387922018766, + 0.006855226121842861, + 0.004847739357501268, + 0.003323670942336321, + 0.02921583689749241, + 0.011451104655861855, + 0.029133332893252373, + 0.016949517652392387, + 0.012346005067229271, + 0.020728008821606636, + 0.03705277293920517, + 0.03750646486878395, + 0.022969016805291176, + 0.012770973145961761, + 0.020839296281337738, + 0.017052317038178444, + 0.027886250987648964, + 0.022035986185073853, + 0.023405687883496284, + 0.025234321132302284, + 0.00406465120613575, + 0.00226048962213099, + 0.0030510174110531807, + 0.0032168503385037184, + 0.003162387292832136, + 0.003940872382372618, + 0.004413039889186621, + 0.004317545332014561, + 0.004381740931421518, + 0.003702426329255104, + 0.0047446428798139095, + 0.004903709981590509, + 0.009072795510292053, + 0.005155122373253107, + 0.005152767989784479, + 0.0033903736621141434, + 0.007537060882896185, + 0.0063563683070242405, + 0.007949384860694408, + 0.009483488276600838, + 0.008438537828624249, + 0.012761370278894901, + 0.009747766889631748, + 0.009371488355100155, + 0.01592911221086979, + 0.008604609407484531, + 0.012133539654314518, + 0.02612028643488884, + 0.009882615879178047, + 0.00793286133557558, + 0.007838415913283825, + 0.004424849525094032, + 0.005075425375252962, + 0.0047200764529407024, + 0.0070389495231211185, + 0.0063466764986515045, + 0.0059866937808692455, + 0.005577581934630871, + 0.004308396019041538, + 0.008181247860193253, + 0.00572538236156106, + 0.005113712977617979, + 0.008878735825419426, + 0.0094944192096591, + 0.004903214983642101, + 0.009581228718161583, + 0.006424950435757637, + 0.007237824611365795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0052475095726549625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures a scene in a vast, sterile clinical setting, characterized by a seamless, polished white floor that curves into matching white walls, creating an infinite, featureless space. At the center of this space lies a creature resembling a large, bipedal wolf in a prone position. Its powerful, digitally rendered legs are folded beneath a muscular torso covered in a thick, coarse, salt-and-pepper fur, the texture rendered in hyper-realistic detail, displaying individual gray, black, and white lines. Its long, humanoid arms end in black claws, which rest flat on the reflective floor. Its head is lowered, its long snout almost touching the ground, and its pointed ears flattened against its skull. The creature's large, unusually bushy tail, a dense coat of dark gray fur, lies softly on the floor, curling slightly around its flank. The entire scene is illuminated by a bright, even, and completely shadowless light from an unseen overhead source, meticulously emphasizing the complex topography of the creature's unusual anatomy and the dense texture of its fur, reinforcing the cool, impersonal, and information-rich atmosphere of scientific literature.", + "seed": 9923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000066/group_9/index_8776_seed_9923_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026253525726497175, + "vlm_logprob": -0.026253525726497175, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.026253525726497175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038823653012514114 + ], + "opposite_logprob_tokens": [ + -3.2888236045837402 + ], + "logprob_reward": -0.038823653012514114, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038823653012514114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3298414349555969, + 0.13966934382915497, + 0.0905829444527626, + 0.10610707849264145, + 0.07294335216283798, + 0.06252385675907135, + 0.04079339653253555, + 0.06468946486711502, + 0.05149252712726593, + 0.04135662317276001 + ], + "credit_map": [ + 0.04240484908223152, + 0.03680330514907837, + 0.025604968890547752, + 0.01908170059323311, + 0.04181228205561638, + 0.057550787925720215, + 0.024594848975539207, + 0.08198871463537216, + 0.016870927065610886, + 0.020769432187080383, + 0.010909534990787506, + 0.00980711355805397, + 0.023481180891394615, + 0.018703816458582878, + 0.014820960350334644, + 0.024306366220116615, + 0.020022418349981308, + 0.013766197487711906, + 0.00803289096802473, + 0.006247006356716156, + 0.012080290354788303, + 0.012997757643461227, + 0.006832530722022057, + 0.010603859089314938, + 0.020115327090024948, + 0.015096679329872131, + 0.008603100664913654, + 0.007049006409943104, + 0.012151542119681835, + 0.01683463156223297, + 0.008017473854124546, + 0.01823931559920311, + 0.011987142264842987, + 0.008408140391111374, + 0.0060595558024942875, + 0.0051233828999102116, + 0.008296126499772072, + 0.01064907293766737, + 0.006533025298267603, + 0.015886912122368813, + 0.009884588420391083, + 0.008083418942987919, + 0.005235536489635706, + 0.004333158954977989, + 0.006582238245755434, + 0.009771115146577358, + 0.005169244483113289, + 0.013464556075632572, + 0.0066739278845489025, + 0.005294978618621826, + 0.0028238415252417326, + 0.0026452522724866867, + 0.005774612072855234, + 0.006149296648800373, + 0.0036163609474897385, + 0.00781512912362814, + 0.012155537493526936, + 0.008474924601614475, + 0.005957507994025946, + 0.003846479579806328, + 0.006803817581385374, + 0.010511863976716995, + 0.004755477886646986, + 0.01218385249376297, + 0.0055287485010921955, + 0.010291019454598427, + 0.004321493208408356, + 0.0033460708800703287, + 0.00895205419510603, + 0.006741244345903397, + 0.005204683169722557, + 0.007107213139533997, + 0.004810796584933996, + 0.007709518074989319, + 0.003332733642309904, + 0.002246940042823553, + 0.0077758608385920525, + 0.0053526838310062885, + 0.004440997261554003, + 0.00568709010258317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.038823653012514114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099825382232666 + ], + "opposite_logprob_tokens": [ + -3.288099765777588 + ], + "logprob_reward": -0.038099825382232666, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099825382232666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5563685297966003, + 0.18600238859653473, + 0.06459545344114304, + 0.0547344833612442, + 0.02380666509270668, + 0.020796477794647217, + 0.03019215539097786, + 0.02492677979171276, + 0.025365453213453293, + 0.013211608864367008 + ], + "credit_map": [ + 0.05797167867422104, + 0.07307016104459763, + 0.04466992989182472, + 0.043478481471538544, + 0.06410261243581772, + 0.09348265081644058, + 0.038381338119506836, + 0.1412116289138794, + 0.022288164123892784, + 0.03464551270008087, + 0.01621115580201149, + 0.015773801133036613, + 0.026965631172060966, + 0.027694452553987503, + 0.014630425721406937, + 0.027793249115347862, + 0.011027678847312927, + 0.010613436810672283, + 0.004405692685395479, + 0.00416701240465045, + 0.0073446291498839855, + 0.00954438280314207, + 0.007234598509967327, + 0.010258025489747524, + 0.006588163319975138, + 0.012276132591068745, + 0.0040948172099888325, + 0.0035337533336132765, + 0.0068793632090091705, + 0.007684286683797836, + 0.004735929425805807, + 0.00894203595817089, + 0.004177495837211609, + 0.0036663589999079704, + 0.0017810884164646268, + 0.0019829377997666597, + 0.003002868266776204, + 0.003572816727682948, + 0.0015295104822143912, + 0.004093589261174202, + 0.0034423067700117826, + 0.0026831007562577724, + 0.00167366536334157, + 0.001796930213458836, + 0.002636164892464876, + 0.0028543733060359955, + 0.0021356206852942705, + 0.003574314760044217, + 0.0030251957941800356, + 0.005100923124700785, + 0.002455559791997075, + 0.002472829772159457, + 0.00413483614102006, + 0.004595701117068529, + 0.0022762902081012726, + 0.006130821071565151, + 0.003684848779812455, + 0.0032783865462988615, + 0.0020839539356529713, + 0.0020942618139088154, + 0.0031486619263887405, + 0.0038218963891267776, + 0.0015806875890120864, + 0.005234082695096731, + 0.002151552587747574, + 0.005438301712274551, + 0.0022561841178685427, + 0.0014592327643185854, + 0.004177209921181202, + 0.003336435416713357, + 0.002068174770101905, + 0.004478360526263714, + 0.0014408419374376535, + 0.0021060502622276545, + 0.001277570496313274, + 0.001055226195603609, + 0.0017692992696538568, + 0.0018237621989101171, + 0.001119697350077331, + 0.002619160106405616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.038099825382232666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031939942855387926 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.0031939942855387926, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031939942855387926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3641330897808075, + 0.1445591002702713, + 0.0862618237733841, + 0.05849826708436012, + 0.04204923287034035, + 0.05014663189649582, + 0.06648730486631393, + 0.05625620111823082, + 0.09573714435100555, + 0.0358712375164032 + ], + "credit_map": [ + 0.03899315372109413, + 0.04826918989419937, + 0.028121372684836388, + 0.021453723311424255, + 0.03359750285744667, + 0.06645066291093826, + 0.02159014530479908, + 0.10565732419490814, + 0.013689602725207806, + 0.028538430109620094, + 0.012811160646378994, + 0.010858261026442051, + 0.031349845230579376, + 0.017853600904345512, + 0.015214757062494755, + 0.014243429526686668, + 0.013017418794333935, + 0.013880222104489803, + 0.007058251183480024, + 0.005970395635813475, + 0.010734548792243004, + 0.011813431046903133, + 0.0062226890586316586, + 0.017564861103892326, + 0.00800163671374321, + 0.011264082044363022, + 0.005611340049654245, + 0.004613153170794249, + 0.010281851515173912, + 0.0073423804715275764, + 0.005429429467767477, + 0.00595439039170742, + 0.006980078294873238, + 0.00529401283711195, + 0.003179009072482586, + 0.0027219345793128014, + 0.006400039419531822, + 0.005688966251909733, + 0.003784234868362546, + 0.008000954985618591, + 0.009040266275405884, + 0.0068742032162845135, + 0.004802158568054438, + 0.005074352025985718, + 0.005509546026587486, + 0.007372358348220587, + 0.004363860469311476, + 0.0071098883636295795, + 0.008686409331858158, + 0.010195296257734299, + 0.0054204449988901615, + 0.004987758584320545, + 0.007910175248980522, + 0.010513171553611755, + 0.005888992454856634, + 0.012885045260190964, + 0.007373403757810593, + 0.009703236632049084, + 0.005275548901408911, + 0.004361115396022797, + 0.008481210097670555, + 0.007578998804092407, + 0.0041733901016414165, + 0.00930928997695446, + 0.008850841782987118, + 0.020781029015779495, + 0.009671861305832863, + 0.007547278888523579, + 0.016563009470701218, + 0.012698269449174404, + 0.00805935263633728, + 0.011565491557121277, + 0.004453024826943874, + 0.00543531309813261, + 0.003263867925852537, + 0.0022731092758476734, + 0.005269207525998354, + 0.005423868075013161, + 0.003225573105737567, + 0.006527268793433905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0031939942855387926 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048665110021829605 + ], + "opposite_logprob_tokens": [ + -3.0486650466918945 + ], + "logprob_reward": -0.048665110021829605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048665110021829605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2689191997051239, + 0.09537739306688309, + 0.13936540484428406, + 0.13161979615688324, + 0.09091845899820328, + 0.045464977622032166, + 0.0797722190618515, + 0.031095953658223152, + 0.03883061930537224, + 0.07863595336675644 + ], + "credit_map": [ + 0.04067876562476158, + 0.03548572212457657, + 0.018409963697195053, + 0.020282382145524025, + 0.03434886038303375, + 0.041332367807626724, + 0.020790796726942062, + 0.05759035423398018, + 0.01258842647075653, + 0.010885349474847317, + 0.008748849853873253, + 0.00706299114972353, + 0.009482474997639656, + 0.016354966908693314, + 0.007061585318297148, + 0.02319275215268135, + 0.025781983509659767, + 0.021727461367845535, + 0.009955687448382378, + 0.009327149949967861, + 0.01655283384025097, + 0.019671013578772545, + 0.012116880156099796, + 0.024232393130660057, + 0.020945044234395027, + 0.018543891608715057, + 0.009800243191421032, + 0.007115566171705723, + 0.015077401883900166, + 0.021802226081490517, + 0.009147336706519127, + 0.029188083484768867, + 0.015131340362131596, + 0.014322970062494278, + 0.008703331463038921, + 0.00638878857716918, + 0.010243146680295467, + 0.01275758258998394, + 0.006021740846335888, + 0.017349563539028168, + 0.005588265135884285, + 0.007006410509347916, + 0.0045571415685117245, + 0.004083557054400444, + 0.007224850356578827, + 0.006310374941676855, + 0.004104041960090399, + 0.006590334232896566, + 0.006818817928433418, + 0.01671566441655159, + 0.007494938559830189, + 0.007079973351210356, + 0.015004255808889866, + 0.010129733011126518, + 0.00759308272972703, + 0.008935750462114811, + 0.005984498653560877, + 0.005091785453259945, + 0.002279495820403099, + 0.00219647865742445, + 0.004088385961949825, + 0.004502994008362293, + 0.0023331418633461, + 0.004619172308593988, + 0.004948261193931103, + 0.006528906524181366, + 0.003464153967797756, + 0.0029137814417481422, + 0.004331253003329039, + 0.006074805278331041, + 0.0029565864242613316, + 0.007612869143486023, + 0.007530966307967901, + 0.01905365288257599, + 0.00741304736584425, + 0.00588093139231205, + 0.014380023814737797, + 0.009885883890092373, + 0.007295955438166857, + 0.007195495534688234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.048665110021829605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, sterile clinical environment with seamless white floors and walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a large, bipedal wolf-like creature in a prone position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have salt-and-pepper fur with visible gray, black, and white lines and long humanoid arms ending in black claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting bright, even, and completely shadowless, illuminating the scene from an overhead source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the creature's large, bushy dark gray tail lying on the floor and curling slightly around its flank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024850459303706884 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.0024850459303706884, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024850459303706884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26672011613845825, + 0.11504147201776505, + 0.10651428997516632, + 0.05446823313832283, + 0.09909471869468689, + 0.06849273294210434, + 0.05842921882867813, + 0.07956013828516006, + 0.08234936743974686, + 0.06932967156171799 + ], + "credit_map": [ + 0.041682805866003036, + 0.03787342831492424, + 0.022453337907791138, + 0.019537484273314476, + 0.032056774944067, + 0.03929503262042999, + 0.01939575746655464, + 0.05442550405859947, + 0.016660146415233612, + 0.01583779975771904, + 0.008893652819097042, + 0.007584839593619108, + 0.012385881505906582, + 0.018327275291085243, + 0.007272360380738974, + 0.028079530224204063, + 0.013510802760720253, + 0.02009180374443531, + 0.008890066295862198, + 0.007557287346571684, + 0.01759490929543972, + 0.014395731501281261, + 0.008442288264632225, + 0.016031406819820404, + 0.005520745646208525, + 0.008471829816699028, + 0.003951204940676689, + 0.005254743155092001, + 0.006670951843261719, + 0.007994311861693859, + 0.005804365500807762, + 0.010800083167850971, + 0.014306744560599327, + 0.014117098413407803, + 0.007270855829119682, + 0.006853677798062563, + 0.014694283716380596, + 0.014827745035290718, + 0.006575314793735743, + 0.02044900506734848, + 0.008360798470675945, + 0.011908674612641335, + 0.0046694097109138966, + 0.0053922017104923725, + 0.01007053628563881, + 0.010662133805453777, + 0.007758071180433035, + 0.009670914150774479, + 0.007786876056343317, + 0.006663037464022636, + 0.004259937908500433, + 0.003342143027111888, + 0.0061894687823951244, + 0.010563995689153671, + 0.004326009191572666, + 0.015297755599021912, + 0.011752360500395298, + 0.01105593889951706, + 0.0064891851507127285, + 0.007149421144276857, + 0.00975809060037136, + 0.011188548989593983, + 0.007543771993368864, + 0.01462282333523035, + 0.012267012149095535, + 0.013592475093901157, + 0.0067281476221978664, + 0.004487432539463043, + 0.011323033832013607, + 0.012296807020902634, + 0.006668628193438053, + 0.014985843561589718, + 0.010202107951045036, + 0.009097947739064693, + 0.005382172297686338, + 0.005521304905414581, + 0.010446391999721527, + 0.009888093918561935, + 0.006838313769549131, + 0.011953345499932766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0024850459303706884 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_0/index_tensor([4776, 4776])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_0/index_tensor([4776, 4776])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a23020644eafb1cada8ce1d63ab23b5f5c18dc7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_0/index_tensor([4776, 4776])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing underwater shot begins in a deep, tranquil turquoise water. A magnificent Tyndall-effect beam of light pierces vertically from the shimmering surface, illuminating countless suspended particles, forming a sacred golden passageway. As the camera approaches, a massive, ancient white sturgeon majestically emerges. Its body is a pearly, milky white, its skin crisscrossed with subtle wrinkles from time. The bony plates on its back and sides, like ancient armor, reflect a soft luster in the light. With steady and calm demeanor, its massive body gently guides a group of almost completely transparent young fish. These juveniles cluster tightly around its head. As they collectively pass through the sacred golden beam, their nearly invisible bodies refract a faint iridescent glow, like a mass of flowing, shimmering stardust. The entire scene is enveloped in a tranquil, solemn, and sacred atmosphere imbued with the miracle of life. Every gentle movement of its massive tail fin is imbued with the benevolent love of a guardian.", + "seed": 6618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_0/index_4776_seed_6618_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03226929281372577, + "vlm_logprob": -0.03226929281372577, + "vlm_yes_no_margin": 5.3, + "vlm_token_credit": -0.03226929281372577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004526013042777777 + ], + "opposite_logprob_tokens": [ + -7.754526138305664 + ], + "logprob_reward": -0.004526013042777777, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.004526013042777777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0882827639579773, + 0.03861575946211815, + 0.06064758822321892, + 0.03932933509349823, + 0.037318769842386246, + 0.4571303129196167, + 0.06653279811143875, + 0.07163796573877335, + 0.0727260634303093, + 0.06777866184711456 + ], + "credit_map": [ + 0.012908974662423134, + 0.009495366364717484, + 0.014848319813609123, + 0.012827100232243538, + 0.005581172648817301, + 0.009901225566864014, + 0.00947586353868246, + 0.01324473600834608, + 0.004906860180199146, + 0.005162320099771023, + 0.004073715303093195, + 0.0053310394287109375, + 0.0036360416561365128, + 0.007898629643023014, + 0.003458041697740555, + 0.00414911238476634, + 0.008174826391041279, + 0.006650907918810844, + 0.006284311413764954, + 0.006132917478680611, + 0.004302113316953182, + 0.015439292415976524, + 0.00633264658972621, + 0.007330567110329866, + 0.004115227609872818, + 0.002775427419692278, + 0.0038094527553766966, + 0.0029094675555825233, + 0.0039566452614963055, + 0.011670109815895557, + 0.002501361072063446, + 0.007591642905026674, + 0.005165770184248686, + 0.0033996657002717257, + 0.004268539138138294, + 0.004852341488003731, + 0.004676871933043003, + 0.005181344226002693, + 0.004949994385242462, + 0.004824242554605007, + 0.06752826273441315, + 0.11498342454433441, + 0.043975744396448135, + 0.02836819738149643, + 0.042518310248851776, + 0.03515775501728058, + 0.04962998256087303, + 0.07496864348649979, + 0.005543813109397888, + 0.0075835734605789185, + 0.007743488531559706, + 0.0075937239453196526, + 0.00984161626547575, + 0.007535483222454786, + 0.010433623567223549, + 0.010257474146783352, + 0.006604835856705904, + 0.014995811507105827, + 0.005556205287575722, + 0.009368645958602428, + 0.009222275577485561, + 0.009769502095878124, + 0.00861418154090643, + 0.007506506983190775, + 0.007572167553007603, + 0.011367401108145714, + 0.008496467024087906, + 0.008515560068190098, + 0.010436383076012135, + 0.008633117191493511, + 0.009026776999235153, + 0.0086781932041049, + 0.008528083562850952, + 0.007098787929862738, + 0.005819386802613735, + 0.00796852633357048, + 0.012036982923746109, + 0.00960452575236559, + 0.010497151874005795, + 0.0062252129428088665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.004526013042777777 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013434203574433923 + ], + "opposite_logprob_tokens": [ + -6.626343250274658 + ], + "logprob_reward": -0.0013434203574433923, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013434203574433923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2747070789337158, + 0.032434891909360886, + 0.04957878962159157, + 0.14495983719825745, + 0.061610396951436996, + 0.1992356777191162, + 0.08402066677808762, + 0.0494786836206913, + 0.05694359913468361, + 0.04703041538596153 + ], + "credit_map": [ + 0.03172948211431503, + 0.025685451924800873, + 0.029974034056067467, + 0.02293112687766552, + 0.022670183330774307, + 0.09010525047779083, + 0.01852063462138176, + 0.033090896904468536, + 0.004581536632031202, + 0.0033355497289448977, + 0.004697907250374556, + 0.004261440131813288, + 0.002623649314045906, + 0.004777546972036362, + 0.0027196775190532207, + 0.005437580402940512, + 0.005659532733261585, + 0.005689800716936588, + 0.004475140944123268, + 0.00533667765557766, + 0.005493634846061468, + 0.012071339413523674, + 0.005661467555910349, + 0.005191192962229252, + 0.015617878176271915, + 0.01228352077305317, + 0.01690640300512314, + 0.008162015117704868, + 0.013375356793403625, + 0.040247414261102676, + 0.007317216135561466, + 0.031050030142068863, + 0.012268703430891037, + 0.009281422942876816, + 0.004335261415690184, + 0.0070574418641626835, + 0.005424410104751587, + 0.005587647669017315, + 0.005279937293380499, + 0.01237556990236044, + 0.02948869578540325, + 0.04913545027375221, + 0.017157169058918953, + 0.012180991470813751, + 0.020038217306137085, + 0.015543553046882153, + 0.02288760058581829, + 0.03280399367213249, + 0.008121855556964874, + 0.008665284141898155, + 0.01062660664319992, + 0.01008431427180767, + 0.010871045291423798, + 0.00940379872918129, + 0.013726571574807167, + 0.012521190568804741, + 0.004277582745999098, + 0.006336051970720291, + 0.004146867897361517, + 0.005431318189948797, + 0.010960424318909645, + 0.0064223273657262325, + 0.006073321681469679, + 0.005830788519233465, + 0.005435130093246698, + 0.007454389240592718, + 0.006946304813027382, + 0.005546377971768379, + 0.006288125179708004, + 0.011251462623476982, + 0.00547831691801548, + 0.008543487638235092, + 0.006857661530375481, + 0.005677293986082077, + 0.005391114391386509, + 0.005248197354376316, + 0.006936026271432638, + 0.006653934251517057, + 0.005315281916409731, + 0.00495090102776885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0013434203574433923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095738288015127 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095738288015127, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095738288015127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0883888527750969, + 0.06623683124780655, + 0.09645667672157288, + 0.07096979022026062, + 0.09231717139482498, + 0.14035600423812866, + 0.15308167040348053, + 0.13405247032642365, + 0.06943262368440628, + 0.08870787173509598 + ], + "credit_map": [ + 0.013769910670816898, + 0.006706633139401674, + 0.010684586130082607, + 0.014091204851865768, + 0.00647058617323637, + 0.016132768243551254, + 0.00855778157711029, + 0.011975378729403019, + 0.007055384572595358, + 0.006691809743642807, + 0.012336131185293198, + 0.009663811884820461, + 0.003472287440672517, + 0.0062577691860497, + 0.006167322397232056, + 0.014592310413718224, + 0.020841436460614204, + 0.012735625728964806, + 0.010495205409824848, + 0.009494793601334095, + 0.004930775612592697, + 0.013895681127905846, + 0.012803292833268642, + 0.011259874328970909, + 0.00786941684782505, + 0.004719725344330072, + 0.007380616385489702, + 0.008016828447580338, + 0.0068414523266255856, + 0.013966923579573631, + 0.009566538967192173, + 0.012608285993337631, + 0.016457613557577133, + 0.008472232148051262, + 0.006903790403157473, + 0.00792029220610857, + 0.01135136466473341, + 0.015262427739799023, + 0.009526477195322514, + 0.016422975808382034, + 0.020845260471105576, + 0.02801346406340599, + 0.011951801367104053, + 0.008428915403783321, + 0.013594117015600204, + 0.01997644640505314, + 0.014605212025344372, + 0.022940782830119133, + 0.014954286627471447, + 0.021650291979312897, + 0.018619880080223083, + 0.02090899646282196, + 0.018679382279515266, + 0.012896494008600712, + 0.024034524336457253, + 0.02133781649172306, + 0.011972494423389435, + 0.017276903614401817, + 0.023258613422513008, + 0.012678893283009529, + 0.018722880631685257, + 0.01531684398651123, + 0.017368055880069733, + 0.017457779496908188, + 0.006562835071235895, + 0.010606596246361732, + 0.007022701669484377, + 0.005795088596642017, + 0.008113104850053787, + 0.015093273483216763, + 0.005815274082124233, + 0.010423744097352028, + 0.009254403412342072, + 0.009517296217381954, + 0.011535195633769035, + 0.010388524271547794, + 0.011788732372224331, + 0.013281884603202343, + 0.011754240840673447, + 0.011187588796019554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.004095738288015127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008628585375845432 + ], + "opposite_logprob_tokens": [ + -4.7586283683776855 + ], + "logprob_reward": -0.008628585375845432, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008628585375845432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09474047273397446, + 0.07444614171981812, + 0.1871688812971115, + 0.048893511295318604, + 0.06688135862350464, + 0.150083526968956, + 0.12564149498939514, + 0.10138004273176193, + 0.076328344643116, + 0.07443614304065704 + ], + "credit_map": [ + 0.015025325119495392, + 0.007174615282565355, + 0.01092169526964426, + 0.012559215538203716, + 0.0066771856509149075, + 0.02268494851887226, + 0.007652691099792719, + 0.012044806033372879, + 0.016937322914600372, + 0.010791347362101078, + 0.010902582667768002, + 0.006704551633447409, + 0.006210416555404663, + 0.005971217527985573, + 0.005815888289362192, + 0.011112811975181103, + 0.02249148115515709, + 0.017145613208413124, + 0.018604490906000137, + 0.03727661445736885, + 0.011428222991526127, + 0.021242655813694, + 0.03763510286808014, + 0.021344726905226707, + 0.004562815651297569, + 0.004663677420467138, + 0.005518320947885513, + 0.0059295035898685455, + 0.003556922310963273, + 0.009532172232866287, + 0.006723416969180107, + 0.008406683802604675, + 0.010228119790554047, + 0.00992614310234785, + 0.007378321141004562, + 0.009166420437395573, + 0.0073852152563631535, + 0.006607748568058014, + 0.007554288487881422, + 0.008635111153125763, + 0.021538861095905304, + 0.029344551265239716, + 0.009067886509001255, + 0.010935276746749878, + 0.02402394451200962, + 0.01418561302125454, + 0.022245382890105247, + 0.018742017447948456, + 0.012641535140573978, + 0.019038908183574677, + 0.014999940991401672, + 0.01540497038513422, + 0.014950607903301716, + 0.012587686069309711, + 0.017767494544386864, + 0.018250364810228348, + 0.0073129101656377316, + 0.020769914612174034, + 0.009376799687743187, + 0.010953634977340698, + 0.020852230489253998, + 0.010614791885018349, + 0.010080888867378235, + 0.011418872512876987, + 0.008077560923993587, + 0.009248996153473854, + 0.0065519316121935844, + 0.00667599868029356, + 0.008617818355560303, + 0.019000524654984474, + 0.00633027870208025, + 0.01182523649185896, + 0.009408179670572281, + 0.009011167101562023, + 0.008313832804560661, + 0.008802452124655247, + 0.010359915904700756, + 0.010995607823133469, + 0.008078467100858688, + 0.00946652702987194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.008628585375845432 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14275270700454712 + ], + "opposite_logprob_tokens": [ + -2.0177526473999023 + ], + "logprob_reward": -0.14275270700454712, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14275270700454712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08806873112916946, + 0.07619209587574005, + 0.0972776785492897, + 0.043346527963876724, + 0.1050887256860733, + 0.17030294239521027, + 0.23697057366371155, + 0.051207806915044785, + 0.07082410901784897, + 0.0607207790017128 + ], + "credit_map": [ + 0.013197562657296658, + 0.006464753765612841, + 0.011544514447450638, + 0.014282574877142906, + 0.006644810549914837, + 0.014411031268537045, + 0.010266331024467945, + 0.011257152073085308, + 0.015471266582608223, + 0.007848751731216908, + 0.011684372089803219, + 0.007731487974524498, + 0.005201287567615509, + 0.0077285743318498135, + 0.0072190118953585625, + 0.013307345099747181, + 0.015848901122808456, + 0.010181818157434464, + 0.010070581920444965, + 0.008024606853723526, + 0.0061466447077691555, + 0.02668851613998413, + 0.010153399780392647, + 0.01016321312636137, + 0.004880250431597233, + 0.003944456111639738, + 0.005515380762517452, + 0.004265389405190945, + 0.00404965179041028, + 0.009038365446031094, + 0.005221066530793905, + 0.006431972607970238, + 0.020460965111851692, + 0.014750569127500057, + 0.010948658920824528, + 0.013133973814547062, + 0.009292067028582096, + 0.010490216314792633, + 0.01063620112836361, + 0.01537607703357935, + 0.02636251226067543, + 0.031084580346941948, + 0.017301077023148537, + 0.013392312452197075, + 0.013429990038275719, + 0.012832342647016048, + 0.022324683144688606, + 0.03357546404004097, + 0.01869923807680607, + 0.045855309814214706, + 0.02774922549724579, + 0.03471946716308594, + 0.02361886017024517, + 0.02108071558177471, + 0.03161308169364929, + 0.03363468870520592, + 0.0050298552960157394, + 0.005524162668734789, + 0.007231091149151325, + 0.0061325072310864925, + 0.006660167593508959, + 0.007898222655057907, + 0.0061133685521781445, + 0.006618434563279152, + 0.006581928580999374, + 0.00993822980672121, + 0.00582415284588933, + 0.010294151492416859, + 0.007090250961482525, + 0.011124786920845509, + 0.011444113217294216, + 0.008526502177119255, + 0.005109047517180443, + 0.007303724531084299, + 0.005935270804911852, + 0.01066665444523096, + 0.009108252823352814, + 0.0056372592225670815, + 0.007552738767117262, + 0.009407834149897099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.14275270700454712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing underwater shot begins in a deep, tranquil turquoise water. A magnificent Tyndall-effect beam of light pierces vertically from the shimmering surface, illuminating countless suspended particles, forming a sacred golden passageway. As the camera approaches, a massive, ancient white sturgeon majestically emerges. Its body is a pearly, milky white, its skin crisscrossed with subtle wrinkles from time. The bony plates on its back and sides, like ancient armor, reflect a soft luster in the light. With steady and calm demeanor, its massive body gently guides a group of almost completely transparent young fish. These juveniles cluster tightly around its head. As they collectively pass through the sacred golden beam, their nearly invisible bodies refract a faint iridescent glow, like a mass of flowing, shimmering stardust. The entire scene is enveloped in a tranquil, solemn, and sacred atmosphere imbued with the miracle of life. Every gentle movement of its massive tail fin is imbued with the benevolent love of a guardian.", + "seed": 6619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_0/index_4776_seed_6619_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04119526939466596, + "vlm_logprob": -0.04119526939466596, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.04119526939466596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005814424250274897 + ], + "opposite_logprob_tokens": [ + -6.755814552307129 + ], + "logprob_reward": -0.005814424250274897, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.005814424250274897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05950899049639702, + 0.027242328971624374, + 0.02368696965277195, + 0.01846875622868538, + 0.016976453363895416, + 0.011328445747494698, + 0.016939958557486534, + 0.4531222879886627, + 0.08403913676738739, + 0.28868672251701355 + ], + "credit_map": [ + 0.008020048029720783, + 0.009414616972208023, + 0.005213986150920391, + 0.006790153682231903, + 0.00612553209066391, + 0.00758645124733448, + 0.009288126602768898, + 0.007070074789226055, + 0.004359693266451359, + 0.0021941165905445814, + 0.0026957199443131685, + 0.001957145519554615, + 0.0028110085986554623, + 0.006052504293620586, + 0.00253958348184824, + 0.0046325558796525, + 0.003659601788967848, + 0.0038603818975389004, + 0.0022355057299137115, + 0.0021253009326756, + 0.004569614306092262, + 0.001777314580976963, + 0.0023550575133413076, + 0.0031041938345879316, + 0.0031514132861047983, + 0.0017016093479469419, + 0.002163338242098689, + 0.001973765902221203, + 0.001975448802113533, + 0.002204759744927287, + 0.0023192327935248613, + 0.0029791877605021, + 0.002042671199887991, + 0.0021896918769925833, + 0.0018457556143403053, + 0.002331043826416135, + 0.0026814118027687073, + 0.0022850744426250458, + 0.0013395282439887524, + 0.0022612782195210457, + 0.001415648148395121, + 0.0016470910049974918, + 0.0011455444619059563, + 0.0011020404053851962, + 0.0009202759829349816, + 0.001071085105650127, + 0.0021822655107825994, + 0.0018444949528202415, + 0.0014562892029061913, + 0.0018068571807816625, + 0.0022811368107795715, + 0.0028750852216035128, + 0.0015892089577391744, + 0.00271027279086411, + 0.0022290099877864122, + 0.0019920971244573593, + 0.038303762674331665, + 0.09052130579948425, + 0.04418594390153885, + 0.040080826729536057, + 0.020256539806723595, + 0.06234477832913399, + 0.10835102200508118, + 0.04907812178134918, + 0.0036444340366870165, + 0.006077745463699102, + 0.004028123803436756, + 0.0050911675207316875, + 0.002704394981265068, + 0.0029411830473691225, + 0.02974412404000759, + 0.029807956889271736, + 0.014245599508285522, + 0.004011030308902264, + 0.11354636400938034, + 0.019156692549586296, + 0.005079153459519148, + 0.009740384295582771, + 0.016640285030007362, + 0.10626719892024994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.005814424250274897 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607329398393631 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607329398393631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607329398393631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5303646922111511, + 0.10492652654647827, + 0.02155323512852192, + 0.04963931441307068, + 0.026117298752069473, + 0.009819386526942253, + 0.009993341751396656, + 0.06867305189371109, + 0.02506830170750618, + 0.1538449227809906 + ], + "credit_map": [ + 0.06046292185783386, + 0.051943641155958176, + 0.07085476070642471, + 0.06439130008220673, + 0.054751455783843994, + 0.07529368251562119, + 0.06912519037723541, + 0.08354170620441437, + 0.011824822053313255, + 0.012037379667162895, + 0.009266444481909275, + 0.008744594641029835, + 0.01447757612913847, + 0.02278306894004345, + 0.01100060436874628, + 0.014792033471167088, + 0.0031510975677520037, + 0.003327960381284356, + 0.0017444310942664742, + 0.0022834688425064087, + 0.0034002061001956463, + 0.0027037353720515966, + 0.0026513556949794292, + 0.0022909806575626135, + 0.008354583755135536, + 0.00658364687114954, + 0.005580388940870762, + 0.003734535537660122, + 0.007412194274365902, + 0.0046000173315405846, + 0.006080396007746458, + 0.007293549831956625, + 0.0033800762612372637, + 0.003538573859259486, + 0.0020087731536477804, + 0.0020461296662688255, + 0.007540064863860607, + 0.0027240614872425795, + 0.00217518606223166, + 0.002704433398321271, + 0.0013878422323614359, + 0.0024533693213015795, + 0.0011843437096104026, + 0.0009639316122047603, + 0.0009394517401233315, + 0.0008318942273035645, + 0.0010558604262769222, + 0.0010026940144598484, + 0.0013003251515328884, + 0.0010835104621946812, + 0.0008626311901025474, + 0.0019017135491594672, + 0.0008328043622896075, + 0.001634146785363555, + 0.0012722827959805727, + 0.0011059273965656757, + 0.007195655722171068, + 0.016831567510962486, + 0.008452465757727623, + 0.005694265011698008, + 0.003529006615281105, + 0.011748824268579483, + 0.00977241713553667, + 0.005448854062706232, + 0.0020447487477213144, + 0.0034903939813375473, + 0.0024836258962750435, + 0.002266279887408018, + 0.0011924415593966842, + 0.0017750803381204605, + 0.006158863194286823, + 0.005656864959746599, + 0.00785730592906475, + 0.0065092602744698524, + 0.04866228997707367, + 0.012716606259346008, + 0.006160242948681116, + 0.00819696020334959, + 0.015311868861317635, + 0.04843038693070412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01607329398393631 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193756565451622 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193756565451622, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193756565451622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024887481704354286, + 0.020337017253041267, + 0.02968578226864338, + 0.008463485166430473, + 0.01535859890282154, + 0.025792136788368225, + 0.03733157366514206, + 0.05432552099227905, + 0.10199609398841858, + 0.6818222999572754 + ], + "credit_map": [ + 0.0033448084723204374, + 0.002533977385610342, + 0.0034045057836920023, + 0.0025897016748785973, + 0.002621445106342435, + 0.003295508213341236, + 0.004148188047111034, + 0.00294934818521142, + 0.002175069646909833, + 0.0018031407380476594, + 0.002482709940522909, + 0.0025832641404122114, + 0.0018314344342797995, + 0.004032567143440247, + 0.0025209144223481417, + 0.0029079196974635124, + 0.004533035680651665, + 0.005178131628781557, + 0.0026840383652597666, + 0.00282050552777946, + 0.0056914472952485085, + 0.00255564390681684, + 0.0028943028301000595, + 0.0033286826219409704, + 0.0016351666999980807, + 0.0011377353221178055, + 0.0009486963390372694, + 0.0009077432914637029, + 0.001005740836262703, + 0.0007698594708926976, + 0.0008651883108541369, + 0.0011933556525036693, + 0.00214526173658669, + 0.0018641020869836211, + 0.0015153924468904734, + 0.0023312289267778397, + 0.0028900974430143833, + 0.0018645173404365778, + 0.00135968963149935, + 0.0013883111532777548, + 0.0029862686060369015, + 0.0026968144811689854, + 0.0024899858981370926, + 0.0038195347879081964, + 0.002054585376754403, + 0.0026460702065378428, + 0.004582366906106472, + 0.004516512621194124, + 0.003259591991081834, + 0.0034538903273642063, + 0.0037303795106709003, + 0.004877225961536169, + 0.006945646367967129, + 0.006038137245923281, + 0.004096560180187225, + 0.004930148366838694, + 0.0026593238580971956, + 0.004447272978723049, + 0.012490798719227314, + 0.004091683775186539, + 0.0023883883841335773, + 0.0019756790716201067, + 0.020449237897992134, + 0.005823140498250723, + 0.003822202794253826, + 0.004814879037439823, + 0.010840553790330887, + 0.007094397675246, + 0.0028266627341508865, + 0.00392068549990654, + 0.03996293991804123, + 0.028713786974549294, + 0.03274988383054733, + 0.008625399321317673, + 0.26639384031295776, + 0.04569592326879501, + 0.011961672455072403, + 0.0229171309620142, + 0.03963101655244827, + 0.25384747982025146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.003193756565451622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571446046233177 + ], + "opposite_logprob_tokens": [ + -3.89557147026062 + ], + "logprob_reward": -0.020571446046233177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571446046233177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11350686103105545, + 0.0633181557059288, + 0.04084755852818489, + 0.019111815840005875, + 0.04216960817575455, + 0.04246201738715172, + 0.10121874511241913, + 0.23926353454589844, + 0.06639964878559113, + 0.27170199155807495 + ], + "credit_map": [ + 0.01742364652454853, + 0.016133738681674004, + 0.010456807911396027, + 0.012757581658661366, + 0.014337019063532352, + 0.014506866224110126, + 0.016430731862783432, + 0.011460479348897934, + 0.00889662280678749, + 0.012392599135637283, + 0.0051723686046898365, + 0.00626075454056263, + 0.009560305625200272, + 0.006996546871960163, + 0.006401308812201023, + 0.007637660484761, + 0.006400041747838259, + 0.006221469026058912, + 0.004088588058948517, + 0.0031247118022292852, + 0.005686541553586721, + 0.0043953959830105305, + 0.004099920392036438, + 0.006830895785242319, + 0.0027228430844843388, + 0.002698373980820179, + 0.0022446527145802975, + 0.002724071964621544, + 0.002570429816842079, + 0.0016456975135952234, + 0.0020588203333318233, + 0.002446928061544895, + 0.005373168271034956, + 0.004919242113828659, + 0.004328013397753239, + 0.004424357786774635, + 0.00880471058189869, + 0.005137518979609013, + 0.003601245814934373, + 0.0055813598446547985, + 0.005647645331919193, + 0.006942578591406345, + 0.004618029575794935, + 0.005121749825775623, + 0.004253278020769358, + 0.006683307699859142, + 0.004682344850152731, + 0.004513084422796965, + 0.008656318299472332, + 0.01076410710811615, + 0.009060307405889034, + 0.01024695485830307, + 0.02296910248696804, + 0.018993422389030457, + 0.009686646983027458, + 0.010841891169548035, + 0.019436055794358253, + 0.04461235925555229, + 0.021035565063357353, + 0.02159966342151165, + 0.010582040064036846, + 0.03135254234075546, + 0.06075216457247734, + 0.029893169179558754, + 0.004288617987185717, + 0.010888559743762016, + 0.011312647722661495, + 0.005941556300967932, + 0.0035416902974247932, + 0.004443358164280653, + 0.020331351086497307, + 0.005651871673762798, + 0.013886451721191406, + 0.013705506920814514, + 0.04728280380368233, + 0.062429726123809814, + 0.012302042916417122, + 0.015776317566633224, + 0.05069485679268837, + 0.055624306201934814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.020571446046233177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032342612743378 + ], + "opposite_logprob_tokens": [ + -1.9103233814239502 + ], + "logprob_reward": -0.16032342612743378, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032342612743378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062054578214883804, + 0.02833920158445835, + 0.02741481363773346, + 0.07191362977027893, + 0.04676316678524017, + 0.04903285577893257, + 0.1131877601146698, + 0.13203293085098267, + 0.08655893057584763, + 0.38270217180252075 + ], + "credit_map": [ + 0.0076840766705572605, + 0.005983270704746246, + 0.007503448985517025, + 0.006273282691836357, + 0.009054457768797874, + 0.008179351687431335, + 0.010239584371447563, + 0.007137104868888855, + 0.0040049259550869465, + 0.004744181875139475, + 0.0022654877975583076, + 0.00391794741153717, + 0.003235175972804427, + 0.0028306262101978064, + 0.0032632197253406048, + 0.0040776366367936134, + 0.004484973847866058, + 0.003601736156269908, + 0.0024280715733766556, + 0.002271011471748352, + 0.003282529069110751, + 0.0035758703015744686, + 0.003027216996997595, + 0.004743404220789671, + 0.012506880797445774, + 0.011613377369940281, + 0.007510163355618715, + 0.0063887955620884895, + 0.011416416615247726, + 0.004926256835460663, + 0.008593750186264515, + 0.008957984857261181, + 0.005757088307291269, + 0.005033088382333517, + 0.004380149766802788, + 0.004836500622332096, + 0.01094649825245142, + 0.005159303545951843, + 0.004668300971388817, + 0.0059822360053658485, + 0.006693032570183277, + 0.008197862654924393, + 0.004929269198328257, + 0.004633027594536543, + 0.004050721880048513, + 0.004220776259899139, + 0.007230166811496019, + 0.00907799880951643, + 0.017749767750501633, + 0.018237290903925896, + 0.013000046834349632, + 0.01575508527457714, + 0.015631072223186493, + 0.011866935528814793, + 0.01177704706788063, + 0.009170514531433582, + 0.008909909054636955, + 0.013490123674273491, + 0.01883806474506855, + 0.011624935083091259, + 0.004747309722006321, + 0.012257777154445648, + 0.04445357620716095, + 0.017711231485009193, + 0.004121210426092148, + 0.003143856767565012, + 0.011403891257941723, + 0.007802432402968407, + 0.003000543685629964, + 0.004324967972934246, + 0.03286000341176987, + 0.019902030006051064, + 0.018903018906712532, + 0.008995394222438335, + 0.14468547701835632, + 0.0261657927185297, + 0.009650329127907753, + 0.012794882990419865, + 0.027287907898426056, + 0.13421933352947235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.16032342612743378 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing underwater shot begins in a deep, tranquil turquoise water. A magnificent Tyndall-effect beam of light pierces vertically from the shimmering surface, illuminating countless suspended particles, forming a sacred golden passageway. As the camera approaches, a massive, ancient white sturgeon majestically emerges. Its body is a pearly, milky white, its skin crisscrossed with subtle wrinkles from time. The bony plates on its back and sides, like ancient armor, reflect a soft luster in the light. With steady and calm demeanor, its massive body gently guides a group of almost completely transparent young fish. These juveniles cluster tightly around its head. As they collectively pass through the sacred golden beam, their nearly invisible bodies refract a faint iridescent glow, like a mass of flowing, shimmering stardust. The entire scene is enveloped in a tranquil, solemn, and sacred atmosphere imbued with the miracle of life. Every gentle movement of its massive tail fin is imbued with the benevolent love of a guardian.", + "seed": 6620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_0/index_4776_seed_6620_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05289135929197073, + "vlm_logprob": -0.05289135929197073, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.05289135929197073, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005021223332732916 + ], + "opposite_logprob_tokens": [ + -7.880021095275879 + ], + "logprob_reward": -0.005021223332732916, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.005021223332732916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11738380044698715, + 0.23170459270477295, + 0.07011070102453232, + 0.08360452204942703, + 0.13071762025356293, + 0.09268356114625931, + 0.05044034868478775, + 0.03708438575267792, + 0.06811745464801788, + 0.11815297603607178 + ], + "credit_map": [ + 0.011188847944140434, + 0.013328570872545242, + 0.017269957810640335, + 0.01404812466353178, + 0.015828082337975502, + 0.015147476457059383, + 0.018770426511764526, + 0.01180232036858797, + 0.018397079780697823, + 0.005953013896942139, + 0.09190978109836578, + 0.011043291538953781, + 0.004801886156201363, + 0.020283544436097145, + 0.009235928766429424, + 0.07008007913827896, + 0.012302479706704617, + 0.00939647015184164, + 0.011425494216382504, + 0.006719287484884262, + 0.0067876605316996574, + 0.00840945914387703, + 0.00750860758125782, + 0.007561240345239639, + 0.005727212410420179, + 0.009998239576816559, + 0.012434763833880424, + 0.016020242124795914, + 0.005844509694725275, + 0.008660241961479187, + 0.01381197851151228, + 0.011107336729764938, + 0.013598607853055, + 0.014333554543554783, + 0.02008095011115074, + 0.020163027569651604, + 0.012715423479676247, + 0.018941324204206467, + 0.0152426827698946, + 0.015642043203115463, + 0.009780209511518478, + 0.008009019307792187, + 0.017063837498426437, + 0.016761504113674164, + 0.00811468344181776, + 0.0100875124335289, + 0.010816419497132301, + 0.012050369754433632, + 0.006994300987571478, + 0.003973167389631271, + 0.006526418961584568, + 0.007466746959835291, + 0.004891757853329182, + 0.006848874036222696, + 0.007533890660852194, + 0.006205191370099783, + 0.0027619819156825542, + 0.003884892212226987, + 0.005965861026197672, + 0.003730535740032792, + 0.004840637557208538, + 0.006436589173972607, + 0.00400296226143837, + 0.005460925400257111, + 0.009287664666771889, + 0.008586570620536804, + 0.00833576824516058, + 0.005964607931673527, + 0.00957565475255251, + 0.012578895315527916, + 0.004509252496063709, + 0.009279036894440651, + 0.024114124476909637, + 0.012594765052199364, + 0.01397301908582449, + 0.011285305954515934, + 0.011597543023526669, + 0.01666525937616825, + 0.008224382065236568, + 0.019698571413755417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.005021223332732916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102624487131834 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.004102624487131834, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102624487131834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12548518180847168, + 0.5086551904678345, + 0.13605612516403198, + 0.020689385011792183, + 0.05979236587882042, + 0.06014174595475197, + 0.027261972427368164, + 0.02065376378595829, + 0.01277665700763464, + 0.028487592935562134 + ], + "credit_map": [ + 0.009959149174392223, + 0.012307240627706051, + 0.021606074646115303, + 0.01636408641934395, + 0.017788942903280258, + 0.014070689678192139, + 0.01768684573471546, + 0.015702152624726295, + 0.023484593257308006, + 0.013274970464408398, + 0.14858873188495636, + 0.019091010093688965, + 0.015724705532193184, + 0.20238955318927765, + 0.01587766967713833, + 0.07022396475076675, + 0.01581806316971779, + 0.022009089589118958, + 0.017454838380217552, + 0.013464807532727718, + 0.013629437424242496, + 0.01682290807366371, + 0.012467148713767529, + 0.024389823898673058, + 0.0014406159752979875, + 0.002658308017998934, + 0.0032276841811835766, + 0.0022387264762073755, + 0.0017531504854559898, + 0.004093287046998739, + 0.0016085361130535603, + 0.003669076832011342, + 0.004736452829092741, + 0.006438801530748606, + 0.012377171777188778, + 0.009404313750565052, + 0.0055520543828606606, + 0.006481363903731108, + 0.007096095476299524, + 0.00770611222833395, + 0.006323045119643211, + 0.00451429421082139, + 0.01147378608584404, + 0.007457638159394264, + 0.008851980790495872, + 0.0063819484785199165, + 0.006742635741829872, + 0.008396420627832413, + 0.003889522049576044, + 0.0026091530453413725, + 0.004478913266211748, + 0.0030851783230900764, + 0.00252556218765676, + 0.0034552346915006638, + 0.0036847104784101248, + 0.003533696522936225, + 0.0025597771164029837, + 0.001533872913569212, + 0.0019302890868857503, + 0.0029376596212387085, + 0.0023842956870794296, + 0.003512694500386715, + 0.0029020500369369984, + 0.0028931235428899527, + 0.0016176124336197972, + 0.0013275227975100279, + 0.0014444288099184632, + 0.0013965972466394305, + 0.0017667453503236175, + 0.002128665102645755, + 0.0011714004212990403, + 0.001923684380017221, + 0.005574973300099373, + 0.0034522393252700567, + 0.0031456926371902227, + 0.0026421595830470324, + 0.002531454199925065, + 0.004718499258160591, + 0.00227950606495142, + 0.004143068101257086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.004102624487131834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606355793774128 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.01606355793774128, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606355793774128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08399077504873276, + 0.3341294825077057, + 0.043402642011642456, + 0.03685469552874565, + 0.13465921580791473, + 0.21427184343338013, + 0.05530620738863945, + 0.03288442641496658, + 0.0422922782599926, + 0.022208450362086296 + ], + "credit_map": [ + 0.008096289820969105, + 0.01050520222634077, + 0.011968862265348434, + 0.009033882059156895, + 0.01321151852607727, + 0.010904080234467983, + 0.011289537884294987, + 0.008981404826045036, + 0.014528626576066017, + 0.004675987176597118, + 0.09777063131332397, + 0.011778125539422035, + 0.009390072897076607, + 0.13941442966461182, + 0.007459770422428846, + 0.04911183565855026, + 0.004714331589639187, + 0.0045395209454, + 0.005970393307507038, + 0.006507546175271273, + 0.004760880954563618, + 0.0051419856026768684, + 0.006937835365533829, + 0.004830149933695793, + 0.0034765820018947124, + 0.003426465205848217, + 0.004104649648070335, + 0.005074045155197382, + 0.0032408724073320627, + 0.006220328621566296, + 0.0046261996030807495, + 0.006685552652925253, + 0.011838839389383793, + 0.01029545720666647, + 0.03065941296517849, + 0.02117890864610672, + 0.011762174777686596, + 0.01502217911183834, + 0.01703222095966339, + 0.01687001809477806, + 0.02334582805633545, + 0.014631751924753189, + 0.0491662472486496, + 0.03184595704078674, + 0.019351189956068993, + 0.0221649631857872, + 0.024442914873361588, + 0.02932298555970192, + 0.004517598543316126, + 0.004492598120123148, + 0.012492852285504341, + 0.008108320645987988, + 0.004612477496266365, + 0.008375744335353374, + 0.006683301646262407, + 0.006023313384503126, + 0.0028769788332283497, + 0.0026875766925513744, + 0.003508124267682433, + 0.004020450636744499, + 0.004688960034400225, + 0.006420916877686977, + 0.0041333697736263275, + 0.004548048134893179, + 0.0035424265079200268, + 0.004011436365544796, + 0.0052295648492872715, + 0.00590991135686636, + 0.0064397226087749004, + 0.006486883386969566, + 0.004744845908135176, + 0.005927485413849354, + 0.0031532037537544966, + 0.0033395702484995127, + 0.0023693018592894077, + 0.002184490207582712, + 0.002262718044221401, + 0.003926035948097706, + 0.002144017955288291, + 0.002829112811014056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.01606355793774128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895165681838989 + ], + "opposite_logprob_tokens": [ + -2.578951597213745 + ], + "logprob_reward": -0.07895165681838989, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895165681838989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08071967214345932, + 0.3644939661026001, + 0.056794196367263794, + 0.061755601316690445, + 0.1383526772260666, + 0.10809984803199768, + 0.06948521733283997, + 0.018606556579470634, + 0.033781636506319046, + 0.06791067868471146 + ], + "credit_map": [ + 0.008596681989729404, + 0.008884563110768795, + 0.014429951086640358, + 0.008648698218166828, + 0.009128781035542488, + 0.009885387495160103, + 0.007795996032655239, + 0.013349613174796104, + 0.016668247058987617, + 0.004562155343592167, + 0.11614073812961578, + 0.011609348468482494, + 0.008554563857614994, + 0.15175481140613556, + 0.00798861775547266, + 0.04721551015973091, + 0.004790471866726875, + 0.005522267427295446, + 0.007338572293519974, + 0.009110701270401478, + 0.0041402303613722324, + 0.009157192893326283, + 0.009854639880359173, + 0.006880119908601046, + 0.004889385309070349, + 0.0033702899236232042, + 0.010615776292979717, + 0.008196413516998291, + 0.005371322389692068, + 0.010191923938691616, + 0.008060417138040066, + 0.01106007769703865, + 0.01300890278071165, + 0.018581366166472435, + 0.02169034257531166, + 0.020884105935692787, + 0.013160227797925472, + 0.017721321433782578, + 0.016057929024100304, + 0.017248474061489105, + 0.011172430589795113, + 0.008600923232734203, + 0.019367797300219536, + 0.01465833093971014, + 0.015163665637373924, + 0.011728363111615181, + 0.013079428113996983, + 0.014328896068036556, + 0.006906419526785612, + 0.005501257721334696, + 0.014791641384363174, + 0.009261035360395908, + 0.005896431393921375, + 0.008691031485795975, + 0.009259232319891453, + 0.009178166277706623, + 0.0027281809598207474, + 0.0013179319212213159, + 0.0018435240490362048, + 0.002601413754746318, + 0.0020432197488844395, + 0.0032267780043184757, + 0.00240731961093843, + 0.0024381871335208416, + 0.0022920826449990273, + 0.00377340498380363, + 0.0034987227991223335, + 0.0030786930583417416, + 0.004923542495816946, + 0.0076585812494158745, + 0.0034879266750067472, + 0.005068682599812746, + 0.007116294465959072, + 0.01010543666779995, + 0.008299638517200947, + 0.006705647800117731, + 0.006926932372152805, + 0.012182153761386871, + 0.007323056925088167, + 0.009251512587070465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.07895165681838989 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16031773388385773 + ], + "opposite_logprob_tokens": [ + -1.9103177785873413 + ], + "logprob_reward": -0.16031773388385773, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16031773388385773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08129619061946869, + 0.18241797387599945, + 0.06234321370720863, + 0.044365499168634415, + 0.1855567842721939, + 0.09470595419406891, + 0.19108669459819794, + 0.06744005531072617, + 0.07246436923742294, + 0.018323251977562904 + ], + "credit_map": [ + 0.0069526368752121925, + 0.004619174171239138, + 0.019729694351553917, + 0.007103616371750832, + 0.004895008634775877, + 0.011138499714434147, + 0.007375149521976709, + 0.01948241889476776, + 0.013118743896484375, + 0.0036365280393511057, + 0.07733570784330368, + 0.008475907146930695, + 0.004248322919011116, + 0.030093485489487648, + 0.004553312435746193, + 0.040955960750579834, + 0.009181572124361992, + 0.008825418539345264, + 0.008167593739926815, + 0.006911806296557188, + 0.006162680685520172, + 0.0072357202880084515, + 0.008357975631952286, + 0.007500448264181614, + 0.003681323491036892, + 0.0064299744553864, + 0.005337797570973635, + 0.0066423071548342705, + 0.004292896017432213, + 0.00556444376707077, + 0.007200086954981089, + 0.005216672085225582, + 0.017271926626563072, + 0.013608942739665508, + 0.040295641869306564, + 0.03317977115511894, + 0.013998868875205517, + 0.021387824788689613, + 0.022540418431162834, + 0.02327338419854641, + 0.007646908052265644, + 0.009640566073358059, + 0.017354028299450874, + 0.01947682723402977, + 0.009256160818040371, + 0.00964977964758873, + 0.010048095136880875, + 0.011633592657744884, + 0.02280290052294731, + 0.016170023009181023, + 0.03398126736283302, + 0.025464598089456558, + 0.018033944070339203, + 0.021842628717422485, + 0.028988055884838104, + 0.02380327321588993, + 0.008889535441994667, + 0.0053172544576227665, + 0.008729531429708004, + 0.007422499358654022, + 0.007220798172056675, + 0.012476376257836819, + 0.00777020025998354, + 0.00961385853588581, + 0.006270688027143478, + 0.00660304632037878, + 0.010253584012389183, + 0.008106924593448639, + 0.008825864642858505, + 0.013009943068027496, + 0.00853771436959505, + 0.01085660420358181, + 0.0023476132191717625, + 0.0024063820019364357, + 0.0020091182086616755, + 0.0019531724974513054, + 0.0022809933871030807, + 0.002926758723333478, + 0.00223055062815547, + 0.002168663777410984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.16031773388385773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing underwater shot begins in a deep, tranquil turquoise water. A magnificent Tyndall-effect beam of light pierces vertically from the shimmering surface, illuminating countless suspended particles, forming a sacred golden passageway. As the camera approaches, a massive, ancient white sturgeon majestically emerges. Its body is a pearly, milky white, its skin crisscrossed with subtle wrinkles from time. The bony plates on its back and sides, like ancient armor, reflect a soft luster in the light. With steady and calm demeanor, its massive body gently guides a group of almost completely transparent young fish. These juveniles cluster tightly around its head. As they collectively pass through the sacred golden beam, their nearly invisible bodies refract a faint iridescent glow, like a mass of flowing, shimmering stardust. The entire scene is enveloped in a tranquil, solemn, and sacred atmosphere imbued with the miracle of life. Every gentle movement of its massive tail fin is imbued with the benevolent love of a guardian.", + "seed": 6621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_0/index_4776_seed_6621_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017842587316408753, + "vlm_logprob": -0.017842587316408753, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.017842587316408753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032058770302683115 + ], + "opposite_logprob_tokens": [ + -7.2532057762146 + ], + "logprob_reward": -0.0032058770302683115, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0032058770302683115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16578666865825653, + 0.0716584324836731, + 0.07661434262990952, + 0.08068916201591492, + 0.053128935396671295, + 0.08381951600313187, + 0.08187893033027649, + 0.13336873054504395, + 0.17702652513980865, + 0.07602875679731369 + ], + "credit_map": [ + 0.01692100428044796, + 0.019667766988277435, + 0.019291948527097702, + 0.022610535845160484, + 0.019490644335746765, + 0.03135436773300171, + 0.019699374213814735, + 0.01675102673470974, + 0.006560050416737795, + 0.014581440016627312, + 0.008366895839571953, + 0.006847085896879435, + 0.009944956749677658, + 0.009492341428995132, + 0.007154879625886679, + 0.008710785768926144, + 0.008114690892398357, + 0.008096996694803238, + 0.008547595702111721, + 0.011546946130692959, + 0.008844939060509205, + 0.01088100764900446, + 0.01186089776456356, + 0.008721266873180866, + 0.016747061163187027, + 0.016903068870306015, + 0.008033428341150284, + 0.008058881387114525, + 0.006778962444514036, + 0.0058394670486450195, + 0.00760783813893795, + 0.010720454156398773, + 0.006018701940774918, + 0.006980483420193195, + 0.007166924886405468, + 0.006786447949707508, + 0.009262287989258766, + 0.006465707905590534, + 0.005192201118916273, + 0.005256179720163345, + 0.007760141976177692, + 0.01269044540822506, + 0.017763426527380943, + 0.010281847789883614, + 0.01052841916680336, + 0.010437735356390476, + 0.007716885767877102, + 0.0066406154073774815, + 0.009921973571181297, + 0.012586629018187523, + 0.013837171718478203, + 0.008228389546275139, + 0.006965601816773415, + 0.010503830388188362, + 0.010088705457746983, + 0.009746627882122993, + 0.006009156350046396, + 0.027518779039382935, + 0.016011860221624374, + 0.012392379343509674, + 0.01018986850976944, + 0.010212741792201996, + 0.04355870187282562, + 0.007475236430764198, + 0.01361768040806055, + 0.043740980327129364, + 0.017228171229362488, + 0.02678268775343895, + 0.010411331430077553, + 0.03565012291073799, + 0.014146205969154835, + 0.015449351631104946, + 0.008610527031123638, + 0.006194856949150562, + 0.00920658465474844, + 0.011318130418658257, + 0.008914695121347904, + 0.00905534066259861, + 0.009312285110354424, + 0.013416336849331856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0032058770302683115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255336407572031 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255336407572031, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255336407572031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3574134111404419, + 0.17553500831127167, + 0.17134559154510498, + 0.0482499785721302, + 0.024898728355765343, + 0.036368004977703094, + 0.024675197899341583, + 0.0661795362830162, + 0.06950142234563828, + 0.025833163410425186 + ], + "credit_map": [ + 0.03484344854950905, + 0.040793053805828094, + 0.05155052989721298, + 0.03856240585446358, + 0.046015772968530655, + 0.07452033460140228, + 0.031560078263282776, + 0.039567768573760986, + 0.013562200590968132, + 0.044542036950588226, + 0.018771661445498466, + 0.01750912331044674, + 0.02193082496523857, + 0.023256855085492134, + 0.013651316985487938, + 0.022310983389616013, + 0.016056371852755547, + 0.0234520323574543, + 0.016284029930830002, + 0.013278162106871605, + 0.02330140769481659, + 0.0358550064265728, + 0.023260030895471573, + 0.019858548417687416, + 0.008299009874463081, + 0.0074511864222586155, + 0.004252418410032988, + 0.003799355123192072, + 0.00543031794950366, + 0.0043582855723798275, + 0.007904281839728355, + 0.006755119655281305, + 0.003432510420680046, + 0.004420706536620855, + 0.003732928540557623, + 0.002743897493928671, + 0.0022234993521124125, + 0.002694427967071533, + 0.002667100401595235, + 0.0029836578760296106, + 0.0028026988729834557, + 0.004871327895671129, + 0.007822739891707897, + 0.004449302796274424, + 0.005198625847697258, + 0.004620585590600967, + 0.00380519637838006, + 0.0027975281700491905, + 0.0028277270030230284, + 0.003960000351071358, + 0.003630466526374221, + 0.0032365242950618267, + 0.0025416959542781115, + 0.003512253984808922, + 0.0026122599374502897, + 0.0023542726412415504, + 0.0029984107241034508, + 0.013586132787168026, + 0.007128043100237846, + 0.0046227239072322845, + 0.008587637916207314, + 0.009453575126826763, + 0.014304823242127895, + 0.005498189479112625, + 0.005311880726367235, + 0.00567118963226676, + 0.008841034024953842, + 0.02228839322924614, + 0.005403736140578985, + 0.008035162463784218, + 0.010178233496844769, + 0.0037717963568866253, + 0.0027051891665905714, + 0.0020887546706944704, + 0.0026508348528295755, + 0.002896117279306054, + 0.0029342074412852526, + 0.002308941213414073, + 0.00328654982149601, + 0.006962568964809179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005255336407572031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252134520560503 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252134520560503, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252134520560503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.154595747590065, + 0.06563662737607956, + 0.08467432111501694, + 0.07185925543308258, + 0.10467123985290527, + 0.03349265828728676, + 0.038538653403520584, + 0.2784477472305298, + 0.09027956426143646, + 0.07780410349369049 + ], + "credit_map": [ + 0.017540790140628815, + 0.015159972012043, + 0.021689385175704956, + 0.025985227897763252, + 0.014059253968298435, + 0.025440508499741554, + 0.018400266766548157, + 0.016320345923304558, + 0.006536213681101799, + 0.008201307617127895, + 0.007066661026328802, + 0.006628993898630142, + 0.008113964460790157, + 0.015014867298305035, + 0.005799622740596533, + 0.008274995721876621, + 0.013252535834908485, + 0.008991844020783901, + 0.009113381616771221, + 0.010949276387691498, + 0.00844624824821949, + 0.012779396958649158, + 0.012916458770632744, + 0.008225180208683014, + 0.010335725732147694, + 0.011608066968619823, + 0.0068719047121703625, + 0.00788086373358965, + 0.006346346344798803, + 0.008124529384076595, + 0.00999761838465929, + 0.010694197379052639, + 0.012288686819374561, + 0.014818064868450165, + 0.018388111144304276, + 0.01588885858654976, + 0.013239032588899136, + 0.010153365321457386, + 0.010476173833012581, + 0.009418956004083157, + 0.0036092870868742466, + 0.00550634553655982, + 0.004234932363033295, + 0.0037245627027004957, + 0.0035625898744910955, + 0.004282035399228334, + 0.0040710922330617905, + 0.004501811694353819, + 0.00408334331586957, + 0.005810803268104792, + 0.005984888877719641, + 0.003878192976117134, + 0.003681436413899064, + 0.005709734279662371, + 0.0049759941175580025, + 0.004414262715727091, + 0.012825293466448784, + 0.06860055029392242, + 0.025803767144680023, + 0.02502339333295822, + 0.017698369920253754, + 0.013973751105368137, + 0.09890498965978622, + 0.015617634169757366, + 0.00807882472872734, + 0.011131229810416698, + 0.008382967673242092, + 0.007636795751750469, + 0.011490227654576302, + 0.02507895603775978, + 0.010562635958194733, + 0.007917921058833599, + 0.00825557392090559, + 0.006864698138087988, + 0.009603768587112427, + 0.01057060994207859, + 0.007842568680644035, + 0.00946591142565012, + 0.0166219100356102, + 0.008579066954553127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.005252134520560503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020568760111927986 + ], + "opposite_logprob_tokens": [ + -3.89556884765625 + ], + "logprob_reward": -0.020568760111927986, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020568760111927986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15397226810455322, + 0.10423622280359268, + 0.1729036271572113, + 0.040703676640987396, + 0.04144009202718735, + 0.03489784523844719, + 0.03310789912939072, + 0.28137147426605225, + 0.10589573532342911, + 0.03147110715508461 + ], + "credit_map": [ + 0.01771240308880806, + 0.013710668310523033, + 0.022300245240330696, + 0.029728790745139122, + 0.014465338550508022, + 0.020713942125439644, + 0.01893446035683155, + 0.016406424343585968, + 0.013542989268898964, + 0.009860895574092865, + 0.016481516882777214, + 0.015792177990078926, + 0.008313409984111786, + 0.016651978716254234, + 0.010235910303890705, + 0.013357343152165413, + 0.017039194703102112, + 0.013595391064882278, + 0.021153442561626434, + 0.020495323464274406, + 0.01667685993015766, + 0.030259542167186737, + 0.034066952764987946, + 0.01961691491305828, + 0.004716524388641119, + 0.005062845069915056, + 0.0043152677826583385, + 0.0034838963765650988, + 0.007033840287476778, + 0.00485661206766963, + 0.006959607359021902, + 0.004275079816579819, + 0.00476285582408309, + 0.004907434806227684, + 0.009190411306917667, + 0.006362327374517918, + 0.004758693743497133, + 0.003713482990860939, + 0.003610508283600211, + 0.004134379792958498, + 0.0036075180396437645, + 0.004110288340598345, + 0.0051636421121656895, + 0.004259167704731226, + 0.0039055682718753815, + 0.0046046762727200985, + 0.004228993784636259, + 0.005017991177737713, + 0.0028837881982326508, + 0.004472753033041954, + 0.004751442465931177, + 0.005089830607175827, + 0.0035294287372380495, + 0.005511606577783823, + 0.003635979490354657, + 0.0032330683898180723, + 0.010167589411139488, + 0.07612747699022293, + 0.020589066669344902, + 0.026403356343507767, + 0.019395187497138977, + 0.014396104961633682, + 0.09958101809024811, + 0.014711694791913033, + 0.00629659416154027, + 0.012047434225678444, + 0.011158071458339691, + 0.03341028466820717, + 0.007558419834822416, + 0.012995189055800438, + 0.016161980107426643, + 0.006267764139920473, + 0.003878080751746893, + 0.00316714053042233, + 0.003154906677082181, + 0.0031464046332985163, + 0.003951132297515869, + 0.003684602677822113, + 0.005182056687772274, + 0.005306783597916365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.020568760111927986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in deep, tranquil turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vertical beam of light visible piercing from the surface, illuminating suspended particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, ancient white sturgeon with pearly, milky white skin and bony plates visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a group of almost completely transparent young fish clustering around the sturgeon's head visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance towards the sturgeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054930828511714935 + ], + "opposite_logprob_tokens": [ + -2.9299309253692627 + ], + "logprob_reward": -0.054930828511714935, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054930828511714935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1275949776172638, + 0.10207437723875046, + 0.11209319531917572, + 0.09988481551408768, + 0.0978575274348259, + 0.2653577923774719, + 0.043488774448633194, + 0.08223064988851547, + 0.033991314470767975, + 0.03542648255825043 + ], + "credit_map": [ + 0.01349364873021841, + 0.012134179472923279, + 0.017413759604096413, + 0.018362702801823616, + 0.015034671872854233, + 0.01948644034564495, + 0.0165933296084404, + 0.015076258219778538, + 0.012343747541308403, + 0.017690036445856094, + 0.016733011230826378, + 0.010881857946515083, + 0.01013899315148592, + 0.012728679925203323, + 0.00843315850943327, + 0.013124905526638031, + 0.008610780350863934, + 0.012392580509185791, + 0.013935060240328312, + 0.01411518920212984, + 0.012639488093554974, + 0.014697150327265263, + 0.021122531965374947, + 0.01458041276782751, + 0.016330992802977562, + 0.01501049567013979, + 0.008934717625379562, + 0.010094977915287018, + 0.009013350121676922, + 0.009285103529691696, + 0.013749045319855213, + 0.01746612973511219, + 0.011651608161628246, + 0.012060558423399925, + 0.020442649722099304, + 0.012178401462733746, + 0.011131427250802517, + 0.010423244908452034, + 0.010053945705294609, + 0.00991570483893156, + 0.025246424600481987, + 0.04170781373977661, + 0.03541995584964752, + 0.03354209288954735, + 0.03197607398033142, + 0.029225755482912064, + 0.03176617994904518, + 0.03647352755069733, + 0.0035287905484437943, + 0.006759725511074066, + 0.007008526939898729, + 0.0061164479702711105, + 0.00458559999242425, + 0.008358526043593884, + 0.003978645894676447, + 0.0031525155063718557, + 0.004560041707009077, + 0.019289769232273102, + 0.007170639466494322, + 0.008433949202299118, + 0.006049415096640587, + 0.006946002133190632, + 0.0251851174980402, + 0.00459572346881032, + 0.003483541775494814, + 0.007004948798567057, + 0.004111827351152897, + 0.003943130373954773, + 0.003554282709956169, + 0.0032428884878754616, + 0.004583841655403376, + 0.004066855181008577, + 0.0035578347742557526, + 0.0028736540116369724, + 0.003579153213649988, + 0.003920485731214285, + 0.0034145822282880545, + 0.005001985467970371, + 0.00568976067006588, + 0.0073890285566449165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.054930828511714935 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_tensor([2715, 2715])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_tensor([2715, 2715])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..52657316b69dbee2cf03918288584383df75542d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_tensor([2715, 2715])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The shot begins with an extreme close-up of a hand, then tilts vertically upward, ultimately framing a medium shot. The camera focuses on a pair of slender, powerful hands, their fingertips swiping, tapping, and zooming with an elegant yet decisive rhythm across a large, borderless, transparent touchpad suspended in mid-air. As he manipulates, the screen's vibrant UI elements—composed of complex cyan data streams, dynamic geometric icons, and three-dimensional models—respond to his every command in real time. As the camera steadily and slowly pans upward, we see him dressed in a meticulously tailored, dark indigo-blue Neo-Chinese formal suit. The silk and wool blend gleams subtly under the soft overhead lighting. The minimalist stand-up collar and several obsidian buttons embody a modern Eastern aesthetic. He stands tall and composed, sitting calmly in an ergonomic chair. Finally, the camera freezes on his face, a chiseled, calm, oriental figure with immaculately short hair. His eyes are focused and penetrating, and a confident smile evokes a sense of control. As the camera widens, the background behind him becomes fully revealed: a minimalist office atop a skyscraper. Through a wall of floor-to-ceiling glass windows, a magnificent futuristic skyline unfolds. Countless whimsical aircraft dart among the meticulously arranged, monolithic buildings. The distant sky, stained a warm orange-gold by the setting sun, forms a stark contrast to the cool, technological interior. The entire scene exudes a tranquil, focused, and futuristic atmosphere of elite élite.", + "seed": 6818, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_2715_seed_6818_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14666771832853556, + "vlm_logprob": -0.14666771832853556, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.14666771832853556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.582456648349762 + ], + "opposite_logprob_tokens": [ + -0.832456648349762 + ], + "logprob_reward": -0.582456648349762, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.582456648349762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1772957146167755, + 0.07840686291456223, + 0.07542015612125397, + 0.3787366449832916, + 0.08307644724845886, + 0.06649856269359589, + 0.029714388772845268, + 0.059537436813116074, + 0.02832767739892006, + 0.022986071184277534 + ], + "credit_map": [ + 0.012836016714572906, + 0.013664969243109226, + 0.05093424767255783, + 0.017498541623353958, + 0.017924822866916656, + 0.038632832467556, + 0.018584907054901123, + 0.007219373714178801, + 0.009431558661162853, + 0.009400485083460808, + 0.006868015509098768, + 0.010620154440402985, + 0.012590218335390091, + 0.010718399658799171, + 0.009667743928730488, + 0.00911028403788805, + 0.01150138583034277, + 0.013140898197889328, + 0.006647782865911722, + 0.007985209114849567, + 0.014530476182699203, + 0.010278628207743168, + 0.006985655054450035, + 0.004350124858319759, + 0.010915842838585377, + 0.024179328233003616, + 0.0084062609821558, + 0.008060741238296032, + 0.1626463532447815, + 0.05355783924460411, + 0.10034695267677307, + 0.01062332559376955, + 0.011785629205405712, + 0.010432906448841095, + 0.0065871370024979115, + 0.010834199376404285, + 0.010001926682889462, + 0.010136052966117859, + 0.014605669304728508, + 0.008692932315170765, + 0.0121226841583848, + 0.00661380123347044, + 0.011193089187145233, + 0.007610676344484091, + 0.008517174050211906, + 0.005508204456418753, + 0.009053240530192852, + 0.005879691801965237, + 0.002224432770162821, + 0.0038668562192469835, + 0.002931291703134775, + 0.0034491040278226137, + 0.006012928672134876, + 0.0029846681281924248, + 0.004792523570358753, + 0.0034525804221630096, + 0.0050312308594584465, + 0.0035750598181039095, + 0.002748404396697879, + 0.007034292444586754, + 0.007295300718396902, + 0.004242157097905874, + 0.014242535457015038, + 0.015368452295660973, + 0.0024987852666527033, + 0.0025666116271167994, + 0.0025550168938934803, + 0.004047897644340992, + 0.0023814942687749863, + 0.002146035199984908, + 0.00750390999019146, + 0.004627925809472799, + 0.002490864833816886, + 0.0022399025037884712, + 0.0035319807939231396, + 0.0031210666056722403, + 0.0025558967608958483, + 0.0034541834611445665, + 0.0030811973847448826, + 0.0025109786074608564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.582456648349762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009810919873416424 + ], + "opposite_logprob_tokens": [ + -4.634810924530029 + ], + "logprob_reward": -0.009810919873416424, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009810919873416424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16181544959545135, + 0.05234383046627045, + 0.05914264917373657, + 0.15466642379760742, + 0.06280813366174698, + 0.054549459367990494, + 0.13543057441711426, + 0.09742364287376404, + 0.15788815915584564, + 0.06393171101808548 + ], + "credit_map": [ + 0.01252592820674181, + 0.006066369824111462, + 0.051290903240442276, + 0.015930363908410072, + 0.012296599335968494, + 0.037919290363788605, + 0.016667362302541733, + 0.009118632413446903, + 0.005792364478111267, + 0.005630219355225563, + 0.005036471877247095, + 0.005415284540504217, + 0.009620553813874722, + 0.007095003500580788, + 0.007213131058961153, + 0.006540799047797918, + 0.009196075610816479, + 0.006829326506704092, + 0.007156710606068373, + 0.006404826883226633, + 0.008261287584900856, + 0.005943418480455875, + 0.009516804479062557, + 0.005834193900227547, + 0.01037049014121294, + 0.013451701030135155, + 0.009547223336994648, + 0.012741412967443466, + 0.03774473816156387, + 0.030532153323292732, + 0.024950161576271057, + 0.015328541398048401, + 0.006804999895393848, + 0.004574730526655912, + 0.004917147569358349, + 0.008617520332336426, + 0.011271461844444275, + 0.005453450605273247, + 0.01486089639365673, + 0.0063079241663217545, + 0.008755655959248543, + 0.004740752745419741, + 0.009912642650306225, + 0.007484573870897293, + 0.007315128110349178, + 0.003987292759120464, + 0.007597493007779121, + 0.004755920264869928, + 0.016199013218283653, + 0.010527326725423336, + 0.016206031665205956, + 0.017405575141310692, + 0.014648971147835255, + 0.014909191988408566, + 0.027958540245890617, + 0.01757592149078846, + 0.007715476676821709, + 0.008268220350146294, + 0.005952482111752033, + 0.015709415078163147, + 0.012710745446383953, + 0.009925585240125656, + 0.016849840059876442, + 0.020291879773139954, + 0.013008174486458302, + 0.01736471988260746, + 0.016024038195610046, + 0.022330470383167267, + 0.012738130986690521, + 0.013824082911014557, + 0.03926590830087662, + 0.023332633078098297, + 0.007655112538486719, + 0.005999433342367411, + 0.009607799351215363, + 0.006511566694825888, + 0.007990098558366299, + 0.01207782980054617, + 0.006985829211771488, + 0.007104039192199707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.009810919873416424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497584864497185 + ], + "opposite_logprob_tokens": [ + -2.929975748062134 + ], + "logprob_reward": -0.05497584864497185, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497584864497185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1890619993209839, + 0.06620386987924576, + 0.19293645024299622, + 0.18232300877571106, + 0.0676233246922493, + 0.01660063862800598, + 0.07586143165826797, + 0.11996930837631226, + 0.04408344626426697, + 0.04533648118376732 + ], + "credit_map": [ + 0.012384857051074505, + 0.0075156413950026035, + 0.06068139150738716, + 0.020378686487674713, + 0.015878088772296906, + 0.04391173645853996, + 0.020057128742337227, + 0.008254464715719223, + 0.008158639073371887, + 0.006591469049453735, + 0.005818777251988649, + 0.005745332688093185, + 0.012376212514936924, + 0.01106457319110632, + 0.006731871515512466, + 0.00971699133515358, + 0.02371375449001789, + 0.027951739728450775, + 0.01492819283157587, + 0.01647808775305748, + 0.04228886589407921, + 0.029828062281012535, + 0.02348179928958416, + 0.0142659367993474, + 0.0047757928259670734, + 0.013419024646282196, + 0.004577868618071079, + 0.004723764955997467, + 0.07361637055873871, + 0.0227989349514246, + 0.051667045801877975, + 0.006744212005287409, + 0.007354935631155968, + 0.0050298720598220825, + 0.004296653904020786, + 0.00826632883399725, + 0.012271419167518616, + 0.007255638483911753, + 0.014971792697906494, + 0.008176685310900211, + 0.0024534526746720076, + 0.0017462216783314943, + 0.001498233643360436, + 0.0019384782062843442, + 0.0031261707190424204, + 0.0014588352059945464, + 0.002668852685019374, + 0.0017103933496400714, + 0.009444363415241241, + 0.005564295221120119, + 0.010065366514027119, + 0.009842826053500175, + 0.008062023669481277, + 0.007551964372396469, + 0.01599241979420185, + 0.00933817308396101, + 0.007999761961400509, + 0.00519334664568305, + 0.005376698914915323, + 0.01141568087041378, + 0.008784503675997257, + 0.007583052851259708, + 0.03635242581367493, + 0.037263836711645126, + 0.003809203626587987, + 0.0067456369288265705, + 0.00495782308280468, + 0.007114799693226814, + 0.0036510713398456573, + 0.0048006013967096806, + 0.008355642668902874, + 0.004648670554161072, + 0.005278901197016239, + 0.006004849448800087, + 0.0058823237195611, + 0.00346040609292686, + 0.005647741723805666, + 0.011836286634206772, + 0.003990803379565477, + 0.0032351708505302668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.05497584864497185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003548741340637 + ], + "opposite_logprob_tokens": [ + -2.695035457611084 + ], + "logprob_reward": -0.07003548741340637, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003548741340637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11807054281234741, + 0.030539581552147865, + 0.09183796495199203, + 0.08156150579452515, + 0.2555707097053528, + 0.05029991269111633, + 0.03934755548834801, + 0.16034355759620667, + 0.10246149450540543, + 0.0699671059846878 + ], + "credit_map": [ + 0.009356899186968803, + 0.006052433978766203, + 0.03561823070049286, + 0.012460012920200825, + 0.009466692805290222, + 0.02532223053276539, + 0.012533522211015224, + 0.007260523736476898, + 0.003220590529963374, + 0.0030221485067158937, + 0.0026316791772842407, + 0.003407830372452736, + 0.005668162368237972, + 0.005100461188703775, + 0.004731303080916405, + 0.002757406560704112, + 0.012450458481907845, + 0.01630530133843422, + 0.00647404370829463, + 0.010508410632610321, + 0.015571458265185356, + 0.010505707934498787, + 0.011762662790715694, + 0.008259923197329044, + 0.006735088769346476, + 0.010472415946424007, + 0.004494566470384598, + 0.007619520649313927, + 0.021694611757993698, + 0.011039121076464653, + 0.015761129558086395, + 0.0037450608797371387, + 0.02156856097280979, + 0.02098630554974079, + 0.014479149132966995, + 0.034707680344581604, + 0.04617723450064659, + 0.03245670720934868, + 0.06943470239639282, + 0.01576036959886551, + 0.006080033723264933, + 0.0073250047862529755, + 0.00588266970589757, + 0.005777233745902777, + 0.006996207870543003, + 0.006242538336664438, + 0.007648547645658255, + 0.00434767734259367, + 0.004092668183147907, + 0.00446384958922863, + 0.002625273773446679, + 0.003238091478124261, + 0.006036737468093634, + 0.007762639783322811, + 0.006385989487171173, + 0.0047423094511032104, + 0.013051900081336498, + 0.00885122362524271, + 0.008311980403959751, + 0.01768024079501629, + 0.017370715737342834, + 0.011541642248630524, + 0.039706021547317505, + 0.04382983595132828, + 0.011532822623848915, + 0.017271989956498146, + 0.007839340716600418, + 0.013786510564386845, + 0.00869662407785654, + 0.013738296926021576, + 0.018306758254766464, + 0.011289146728813648, + 0.010865112766623497, + 0.006738026160746813, + 0.006128436420112848, + 0.00876030046492815, + 0.010401129722595215, + 0.0116072753444314, + 0.006946465000510216, + 0.008520365692675114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.07003548741340637 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059687361121178 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.016059687361121178, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059687361121178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11836094409227371, + 0.06819569319486618, + 0.037279464304447174, + 0.10725976526737213, + 0.019986838102340698, + 0.023203669115900993, + 0.050977423787117004, + 0.06845811009407043, + 0.26018884778022766, + 0.24608927965164185 + ], + "credit_map": [ + 0.008791868574917316, + 0.008657129481434822, + 0.025685548782348633, + 0.015462753362953663, + 0.01243638340383768, + 0.020717928186058998, + 0.015535709448158741, + 0.011073620989918709, + 0.008470329456031322, + 0.008730357512831688, + 0.0063300468027591705, + 0.00806824117898941, + 0.01165767852216959, + 0.009738877415657043, + 0.007022432051599026, + 0.008177723735570908, + 0.004635845310986042, + 0.004125355742871761, + 0.004126903600990772, + 0.0024760684464126825, + 0.006921245250850916, + 0.004244619980454445, + 0.006315592210739851, + 0.0044338335283100605, + 0.0028377159032970667, + 0.007505614310503006, + 0.0022913459688425064, + 0.003082670969888568, + 0.045973848551511765, + 0.017077213153243065, + 0.025226755067706108, + 0.003264613216742873, + 0.0027367661241441965, + 0.0018017537659034133, + 0.0017241459572687745, + 0.003891277825459838, + 0.0023982597049325705, + 0.0018505584448575974, + 0.002669557463377714, + 0.002914518117904663, + 0.0035711161326617002, + 0.002285162452608347, + 0.0036038244143128395, + 0.0035934452898800373, + 0.003350394545122981, + 0.0016895226435735822, + 0.003245072904974222, + 0.0018651298014447093, + 0.006542797200381756, + 0.004541548900306225, + 0.004076007287949324, + 0.005178763065487146, + 0.006334411446005106, + 0.007102296222001314, + 0.010450097732245922, + 0.0067515065893530846, + 0.006714039947837591, + 0.0048958114348351955, + 0.0060098241083323956, + 0.009089585393667221, + 0.0067670210264623165, + 0.007757631130516529, + 0.011276527307927608, + 0.01594766415655613, + 0.04058590903878212, + 0.04000764340162277, + 0.026738323271274567, + 0.021432625129818916, + 0.02964836172759533, + 0.05224749073386192, + 0.025932248681783676, + 0.023596255108714104, + 0.031614553183317184, + 0.029462769627571106, + 0.038597557693719864, + 0.018994268029928207, + 0.023250389844179153, + 0.06394847482442856, + 0.01833266206085682, + 0.0218886099755764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.016059687361121178 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with an extreme close-up of a hand, then tilts vertically upward, ultimately framing a medium shot. The camera focuses on a pair of slender, powerful hands, their fingertips swiping, tapping, and zooming with an elegant yet decisive rhythm across a large, borderless, transparent touchpad suspended in mid-air. As he manipulates, the screen's vibrant UI elements—composed of complex cyan data streams, dynamic geometric icons, and three-dimensional models—respond to his every command in real time. As the camera steadily and slowly pans upward, we see him dressed in a meticulously tailored, dark indigo-blue Neo-Chinese formal suit. The silk and wool blend gleams subtly under the soft overhead lighting. The minimalist stand-up collar and several obsidian buttons embody a modern Eastern aesthetic. He stands tall and composed, sitting calmly in an ergonomic chair. Finally, the camera freezes on his face, a chiseled, calm, oriental figure with immaculately short hair. His eyes are focused and penetrating, and a confident smile evokes a sense of control. As the camera widens, the background behind him becomes fully revealed: a minimalist office atop a skyscraper. Through a wall of floor-to-ceiling glass windows, a magnificent futuristic skyline unfolds. Countless whimsical aircraft dart among the meticulously arranged, monolithic buildings. The distant sky, stained a warm orange-gold by the setting sun, forms a stark contrast to the cool, technological interior. The entire scene exudes a tranquil, focused, and futuristic atmosphere of elite élite.", + "seed": 6819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_2715_seed_6819_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08934449711814522, + "vlm_logprob": -0.08934449711814522, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.08934449711814522, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2838507294654846 + ], + "opposite_logprob_tokens": [ + -1.4088506698608398 + ], + "logprob_reward": -0.2838507294654846, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2838507294654846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09918317943811417, + 0.04262717813253403, + 0.08282757550477982, + 0.2575075626373291, + 0.04350878298282623, + 0.05938331037759781, + 0.16140396893024445, + 0.14857445657253265, + 0.07384230941534042, + 0.031141724437475204 + ], + "credit_map": [ + 0.016233088448643684, + 0.016372762620449066, + 0.012256555259227753, + 0.012341913767158985, + 0.014269576407968998, + 0.010430010966956615, + 0.008910082280635834, + 0.008369195275008678, + 0.00476921396329999, + 0.006050792057067156, + 0.004409307148307562, + 0.00450935959815979, + 0.005414146464318037, + 0.005934596061706543, + 0.008372937329113483, + 0.0031668278388679028, + 0.00653974199667573, + 0.00969558022916317, + 0.0037166981492191553, + 0.02303999662399292, + 0.014350389130413532, + 0.005471439566463232, + 0.012000055983662605, + 0.008013670332729816, + 0.03664221987128258, + 0.040246449410915375, + 0.020974216982722282, + 0.029865004122257233, + 0.042107436805963516, + 0.03995860740542412, + 0.03247800096869469, + 0.015235617756843567, + 0.004037199076265097, + 0.006398867350071669, + 0.003142728703096509, + 0.005248290486633778, + 0.0072306371293962, + 0.004709518514573574, + 0.006695481948554516, + 0.0060460614040493965, + 0.010380150750279427, + 0.007032390218228102, + 0.008602218702435493, + 0.00677447160705924, + 0.006504276301711798, + 0.007154414895921946, + 0.006102618295699358, + 0.0068327682092785835, + 0.03727833181619644, + 0.007527859415858984, + 0.05668703839182854, + 0.01494772732257843, + 0.006219424773007631, + 0.008725135587155819, + 0.019015232101082802, + 0.011003229767084122, + 0.01749410107731819, + 0.013130293227732182, + 0.03133326396346092, + 0.01978713646531105, + 0.010565404780209064, + 0.018516525626182556, + 0.021368594840168953, + 0.016379116103053093, + 0.01366562768816948, + 0.006002745125442743, + 0.020405475050210953, + 0.006491098552942276, + 0.004888003692030907, + 0.004602909553796053, + 0.008822428062558174, + 0.008964020758867264, + 0.008853791281580925, + 0.0022810513619333506, + 0.005374955479055643, + 0.0030683784279972315, + 0.00208116858266294, + 0.0031772260554134846, + 0.0027629274409264326, + 0.0035422269720584154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.2838507294654846 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219510369002819 + ], + "opposite_logprob_tokens": [ + -4.264219284057617 + ], + "logprob_reward": -0.014219510369002819, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219510369002819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10066523402929306, + 0.03557823598384857, + 0.08724857121706009, + 0.36135777831077576, + 0.10476868599653244, + 0.1445493847131729, + 0.096136175096035, + 0.023938322439789772, + 0.025900783017277718, + 0.01985679380595684 + ], + "credit_map": [ + 0.010759891010820866, + 0.013681349344551563, + 0.012381589971482754, + 0.012147183530032635, + 0.009544288739562035, + 0.01885245554149151, + 0.011339853517711163, + 0.011958613060414791, + 0.004690168891102076, + 0.0049772802740335464, + 0.0038445645477622747, + 0.00466867396607995, + 0.004924294073134661, + 0.004096153192222118, + 0.005470722448080778, + 0.0029063753318041563, + 0.011706450954079628, + 0.01065053977072239, + 0.007393634878098965, + 0.02388635464012623, + 0.015562576241791248, + 0.005364317446947098, + 0.0067617581225931644, + 0.005922939628362656, + 0.05190911144018173, + 0.06307142972946167, + 0.03085106983780861, + 0.0502532422542572, + 0.057446982711553574, + 0.046053726226091385, + 0.043047916144132614, + 0.01872430182993412, + 0.01048275362700224, + 0.014682324603199959, + 0.006644345819950104, + 0.012832775712013245, + 0.019122816622257233, + 0.009560184553265572, + 0.01727360673248768, + 0.014169875532388687, + 0.027139876037836075, + 0.017681023105978966, + 0.022018808871507645, + 0.015380731783807278, + 0.014625495299696922, + 0.017816342413425446, + 0.014081573113799095, + 0.015805525705218315, + 0.013667356222867966, + 0.005303781945258379, + 0.03548466041684151, + 0.011198400519788265, + 0.004147515166550875, + 0.007272981572896242, + 0.012418800964951515, + 0.006642680149525404, + 0.002890735398977995, + 0.0022753500379621983, + 0.004858645610511303, + 0.003052174812182784, + 0.0020586333703249693, + 0.002831097925081849, + 0.002818668494001031, + 0.0031530188862234354, + 0.0032869072165340185, + 0.0023403074592351913, + 0.004468649625778198, + 0.0026814662851393223, + 0.0022472613491117954, + 0.002096792683005333, + 0.002881505759432912, + 0.005897890776395798, + 0.00391878979280591, + 0.0019679400138556957, + 0.0024786340072751045, + 0.002380969701334834, + 0.0018215685850009322, + 0.0018820763798430562, + 0.002661355771124363, + 0.0027454581577330828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.014219510369002819 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062057919800281525 + ], + "opposite_logprob_tokens": [ + -2.8120579719543457 + ], + "logprob_reward": -0.062057919800281525, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062057919800281525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1331947147846222, + 0.08509550988674164, + 0.059111107140779495, + 0.15660586953163147, + 0.05841045081615448, + 0.03215624764561653, + 0.2573533356189728, + 0.1327943205833435, + 0.051441434770822525, + 0.03383703529834747 + ], + "credit_map": [ + 0.011283261701464653, + 0.017102809622883797, + 0.01410034578293562, + 0.020692652091383934, + 0.014539211988449097, + 0.014834927394986153, + 0.023153984919190407, + 0.017487507313489914, + 0.008228331804275513, + 0.010820622555911541, + 0.010479534976184368, + 0.010965879075229168, + 0.011929391883313656, + 0.01513703353703022, + 0.008570199832320213, + 0.008964505977928638, + 0.008129681460559368, + 0.006067997310310602, + 0.005769703537225723, + 0.006642745807766914, + 0.00754163321107626, + 0.008930940181016922, + 0.011243798770010471, + 0.004784602206200361, + 0.022693399339914322, + 0.027374539524316788, + 0.011851371265947819, + 0.02473474107682705, + 0.026195306330919266, + 0.01591556705534458, + 0.016511930152773857, + 0.011329002678394318, + 0.008569911122322083, + 0.011397359892725945, + 0.0060541220009326935, + 0.007453822530806065, + 0.00603299867361784, + 0.0061059570871293545, + 0.008100828155875206, + 0.00469544343650341, + 0.0036752200685441494, + 0.00272569851949811, + 0.002987435320392251, + 0.005523082800209522, + 0.0030956240370869637, + 0.002336960518732667, + 0.008422612212598324, + 0.0033896123059093952, + 0.02379782497882843, + 0.013623670674860477, + 0.09952647238969803, + 0.02568187564611435, + 0.006172036286443472, + 0.038269490003585815, + 0.03497759997844696, + 0.015304358676075935, + 0.014676885679364204, + 0.011326096020638943, + 0.028837937861680984, + 0.018766267225146294, + 0.009767876006662846, + 0.014430501498281956, + 0.019734453409910202, + 0.015254302881658077, + 0.01391641329973936, + 0.003750972216948867, + 0.006900544743984938, + 0.0042639789171516895, + 0.0031397780403494835, + 0.005002227611839771, + 0.0050271423533558846, + 0.009440373629331589, + 0.008326608687639236, + 0.00367366848513484, + 0.004657625220716, + 0.003293344285339117, + 0.0029632921796292067, + 0.004479215946048498, + 0.0025706824380904436, + 0.0038725947961211205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.062057919800281525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897259294986725 + ], + "opposite_logprob_tokens": [ + -2.578972578048706 + ], + "logprob_reward": -0.07897259294986725, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897259294986725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03596577048301697, + 0.007888528518378735, + 0.025775475427508354, + 0.3061474859714508, + 0.05740570276975632, + 0.011805382557213306, + 0.36738714575767517, + 0.09938439726829529, + 0.06728566437959671, + 0.020954405888915062 + ], + "credit_map": [ + 0.003541234415024519, + 0.0048953574150800705, + 0.0034245289862155914, + 0.0052609252743422985, + 0.005138345528393984, + 0.005824975203722715, + 0.004413102753460407, + 0.003467304864898324, + 0.0007022424833849072, + 0.0010660571279004216, + 0.0007049297564662993, + 0.0010523541131988168, + 0.001021481235511601, + 0.0008811588631942868, + 0.0016339072026312351, + 0.0008263980271294713, + 0.003557946067303419, + 0.002422101330012083, + 0.002380386693403125, + 0.003751421347260475, + 0.0033713297452777624, + 0.00345874996855855, + 0.004249620251357555, + 0.0025839193258434534, + 0.05041372776031494, + 0.053378038108348846, + 0.0231661144644022, + 0.03836303949356079, + 0.047902561724185944, + 0.04482860490679741, + 0.033390358090400696, + 0.014705073088407516, + 0.010135912336409092, + 0.011216199025511742, + 0.004940215963870287, + 0.006316340994089842, + 0.007436749059706926, + 0.006305897608399391, + 0.006389454938471317, + 0.004664933774620295, + 0.0017470630118623376, + 0.001470911898650229, + 0.0012782346457242966, + 0.0015751165337860584, + 0.0016270594205707312, + 0.0011546736350283027, + 0.0019077081233263016, + 0.0010446151718497276, + 0.03722608834505081, + 0.020927822217345238, + 0.14262793958187103, + 0.03408833220601082, + 0.01000025775283575, + 0.04837127402424812, + 0.05134284868836403, + 0.022802578285336494, + 0.011745085939764977, + 0.008622479625046253, + 0.021530965343117714, + 0.014950425364077091, + 0.00693473732098937, + 0.010045617818832397, + 0.014639055356383324, + 0.010916028171777725, + 0.019623860716819763, + 0.004524621646851301, + 0.01535083819180727, + 0.005246919579803944, + 0.0030665083322674036, + 0.006090295501053333, + 0.007690923288464546, + 0.005691696424037218, + 0.0036678484175354242, + 0.001804650411941111, + 0.003293110989034176, + 0.0023795729503035545, + 0.0015506515046581626, + 0.002833039965480566, + 0.0018500671721994877, + 0.003575463080778718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.07897259294986725 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621733006089926 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.007621733006089926, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621733006089926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03133556991815567, + 0.013747702352702618, + 0.030208686366677284, + 0.12451787292957306, + 0.04293939098715782, + 0.013811232522130013, + 0.09372543543577194, + 0.3564031720161438, + 0.08699270337820053, + 0.20631825923919678 + ], + "credit_map": [ + 0.0035117557272315025, + 0.004489593673497438, + 0.002919706515967846, + 0.005640240386128426, + 0.0029399492777884007, + 0.004450105130672455, + 0.004034303594380617, + 0.0033499153796583414, + 0.001509681693278253, + 0.0020280429162085056, + 0.0011956535745412111, + 0.001934179337695241, + 0.0018612195271998644, + 0.0015774666098877788, + 0.0024462633300572634, + 0.0011951946653425694, + 0.005122034344822168, + 0.0033188227098435163, + 0.003720494220033288, + 0.0035069831646978855, + 0.004000917077064514, + 0.0030769258737564087, + 0.0051646605134010315, + 0.0022978479973971844, + 0.02067412994801998, + 0.02045784331858158, + 0.00901732873171568, + 0.015144411474466324, + 0.019266681745648384, + 0.020347461104393005, + 0.012897205539047718, + 0.006712818052619696, + 0.00736843328922987, + 0.007469001226127148, + 0.004468501545488834, + 0.00501538859680295, + 0.005283408332616091, + 0.004641752690076828, + 0.005053218919783831, + 0.0036396861542016268, + 0.0019554905593395233, + 0.0023668704088777304, + 0.0010379554005339742, + 0.0014990248018875718, + 0.002117360010743141, + 0.001995169324800372, + 0.001702259760349989, + 0.0011371017899364233, + 0.02104494534432888, + 0.004776418674737215, + 0.03273175284266472, + 0.007022733334451914, + 0.0035538894589990377, + 0.00729383947327733, + 0.010900527238845825, + 0.006401330232620239, + 0.04038352519273758, + 0.03136788308620453, + 0.07609587162733078, + 0.054747361689805984, + 0.025973239913582802, + 0.03364599496126175, + 0.05294862017035484, + 0.04124067351222038, + 0.026619693264365196, + 0.0061800009571015835, + 0.016884062439203262, + 0.00766659714281559, + 0.004547815304249525, + 0.008003391325473785, + 0.009817278012633324, + 0.007273863069713116, + 0.05965296924114227, + 0.014416387304663658, + 0.024793019518256187, + 0.026118915528059006, + 0.012665804475545883, + 0.024710623547434807, + 0.023908706381917, + 0.020051835104823112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.007621733006089926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with an extreme close-up of a hand, then tilts vertically upward, ultimately framing a medium shot. The camera focuses on a pair of slender, powerful hands, their fingertips swiping, tapping, and zooming with an elegant yet decisive rhythm across a large, borderless, transparent touchpad suspended in mid-air. As he manipulates, the screen's vibrant UI elements—composed of complex cyan data streams, dynamic geometric icons, and three-dimensional models—respond to his every command in real time. As the camera steadily and slowly pans upward, we see him dressed in a meticulously tailored, dark indigo-blue Neo-Chinese formal suit. The silk and wool blend gleams subtly under the soft overhead lighting. The minimalist stand-up collar and several obsidian buttons embody a modern Eastern aesthetic. He stands tall and composed, sitting calmly in an ergonomic chair. Finally, the camera freezes on his face, a chiseled, calm, oriental figure with immaculately short hair. His eyes are focused and penetrating, and a confident smile evokes a sense of control. As the camera widens, the background behind him becomes fully revealed: a minimalist office atop a skyscraper. Through a wall of floor-to-ceiling glass windows, a magnificent futuristic skyline unfolds. Countless whimsical aircraft dart among the meticulously arranged, monolithic buildings. The distant sky, stained a warm orange-gold by the setting sun, forms a stark contrast to the cool, technological interior. The entire scene exudes a tranquil, focused, and futuristic atmosphere of elite élite.", + "seed": 6820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_2715_seed_6820_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025558086670935155, + "vlm_logprob": -0.025558086670935155, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.025558086670935155, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011399493552744389 + ], + "opposite_logprob_tokens": [ + -5.011399269104004 + ], + "logprob_reward": -0.011399493552744389, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011399493552744389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23750752210617065, + 0.07643941789865494, + 0.1341930627822876, + 0.1629754602909088, + 0.09098328649997711, + 0.09479229152202606, + 0.07197500765323639, + 0.05991731584072113, + 0.03187207877635956, + 0.03934451937675476 + ], + "credit_map": [ + 0.02016914263367653, + 0.040868841111660004, + 0.026949208229780197, + 0.034365762025117874, + 0.03440507501363754, + 0.029800020158290863, + 0.022262975573539734, + 0.028686480596661568, + 0.011349012143909931, + 0.012263108976185322, + 0.00672109192237258, + 0.010649765841662884, + 0.009045840241014957, + 0.014142269268631935, + 0.006701856851577759, + 0.005566473118960857, + 0.010483845137059689, + 0.023857731372117996, + 0.01573987491428852, + 0.018767688423395157, + 0.019524546340107918, + 0.017984583973884583, + 0.012875998392701149, + 0.01495878305286169, + 0.02145705185830593, + 0.018240911886096, + 0.011511141434311867, + 0.017369018867611885, + 0.03076937049627304, + 0.02432124875485897, + 0.02179981768131256, + 0.017506906762719154, + 0.010792911984026432, + 0.012214874848723412, + 0.009042368270456791, + 0.010934876278042793, + 0.011444208212196827, + 0.012247979640960693, + 0.014418615959584713, + 0.00988745130598545, + 0.012454202398657799, + 0.012193966656923294, + 0.014405513182282448, + 0.014050284400582314, + 0.008932633325457573, + 0.009533319622278214, + 0.01288934051990509, + 0.010333028621971607, + 0.010237258858978748, + 0.012717748060822487, + 0.00968910288065672, + 0.0070952121168375015, + 0.008878852240741253, + 0.008464516140520573, + 0.0075976429507136345, + 0.007294680457562208, + 0.009330854751169682, + 0.008987250737845898, + 0.009812346659600735, + 0.006296135485172272, + 0.005643129348754883, + 0.006259267218410969, + 0.005551781039685011, + 0.00803654920309782, + 0.003924854099750519, + 0.004722642246633768, + 0.0050048138946294785, + 0.0033025983721017838, + 0.003854928072541952, + 0.003075238084420562, + 0.0029409790877252817, + 0.005046024452894926, + 0.0034877390135079622, + 0.003210049122571945, + 0.006129013374447823, + 0.002917308360338211, + 0.004825802519917488, + 0.0035238349810242653, + 0.0037317031528800726, + 0.011519066989421844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.011399493552744389 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018212705850601196 + ], + "opposite_logprob_tokens": [ + -4.018212795257568 + ], + "logprob_reward": -0.018212705850601196, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018212705850601196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16188760101795197, + 0.12726031243801117, + 0.14256775379180908, + 0.15458036959171295, + 0.11907719075679779, + 0.16568979620933533, + 0.05305102467536926, + 0.026293490082025528, + 0.022889044135808945, + 0.026703475043177605 + ], + "credit_map": [ + 0.018016042187809944, + 0.01624426245689392, + 0.020000793039798737, + 0.03791444003582001, + 0.01499176025390625, + 0.016497692093253136, + 0.020548079162836075, + 0.017674520611763, + 0.016197727993130684, + 0.012643064372241497, + 0.012900753878057003, + 0.028538409620523453, + 0.012348857708275318, + 0.014294279739260674, + 0.013338305056095123, + 0.016998903825879097, + 0.012385821901261806, + 0.02327449806034565, + 0.01621447503566742, + 0.018253816291689873, + 0.020984584465622902, + 0.018759379163384438, + 0.014258534647524357, + 0.018436623737215996, + 0.022978143766522408, + 0.015706663951277733, + 0.012860789895057678, + 0.016997741535305977, + 0.027926748618483543, + 0.018714945763349533, + 0.01951853558421135, + 0.019876783713698387, + 0.01480675395578146, + 0.01256114337593317, + 0.011238256469368935, + 0.01308189332485199, + 0.016533643007278442, + 0.019359344616532326, + 0.018062936142086983, + 0.013433217070996761, + 0.016452934592962265, + 0.019616946578025818, + 0.025542227551341057, + 0.02630162611603737, + 0.014658794738352299, + 0.015630114823579788, + 0.026509912684559822, + 0.020977215841412544, + 0.007767613511532545, + 0.005988151300698519, + 0.008186769671738148, + 0.007140197791159153, + 0.006398878525942564, + 0.006379362195730209, + 0.006315921898931265, + 0.004874126985669136, + 0.003926274832338095, + 0.003300009062513709, + 0.003340223105624318, + 0.0034024843480437994, + 0.0032021873630583286, + 0.0029246914200484753, + 0.0026786248199641705, + 0.003518991870805621, + 0.0032583759166300297, + 0.002713375026360154, + 0.0033016190864145756, + 0.002799106063321233, + 0.003994236700236797, + 0.0020630978979170322, + 0.002277208026498556, + 0.002482024021446705, + 0.002545778639614582, + 0.0028165746480226517, + 0.0036727290134876966, + 0.0027243730146437883, + 0.00415834179148078, + 0.002351660281419754, + 0.0024871542118489742, + 0.005946863442659378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.018212705850601196 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358501985669136 + ], + "opposite_logprob_tokens": [ + -3.6513586044311523 + ], + "logprob_reward": -0.026358501985669136, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358501985669136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3240523934364319, + 0.046976979821920395, + 0.137751042842865, + 0.1505078226327896, + 0.0768183171749115, + 0.10658013075590134, + 0.03468337655067444, + 0.07410017400979996, + 0.019384542480111122, + 0.029145224019885063 + ], + "credit_map": [ + 0.02995896339416504, + 0.0399225652217865, + 0.0352759025990963, + 0.07759277522563934, + 0.0377354696393013, + 0.03556448221206665, + 0.03724639117717743, + 0.030755842104554176, + 0.0052522518672049046, + 0.006849940866231918, + 0.005701785441488028, + 0.0073243253864347935, + 0.00573458569124341, + 0.006238386034965515, + 0.0046993098221719265, + 0.005176390055567026, + 0.012503237463533878, + 0.02208230458199978, + 0.011353638023138046, + 0.009777895174920559, + 0.022039664909243584, + 0.02870761789381504, + 0.011111150495707989, + 0.020175525918602943, + 0.022206254303455353, + 0.0200347900390625, + 0.012684107758104801, + 0.014232547022402287, + 0.02107773721218109, + 0.02389841340482235, + 0.01979973539710045, + 0.016574230045080185, + 0.01506903674453497, + 0.007267503067851067, + 0.007150636520236731, + 0.008894280530512333, + 0.010108253918588161, + 0.011324436403810978, + 0.007206717971712351, + 0.009797447361052036, + 0.012142650783061981, + 0.0117280138656497, + 0.014582721516489983, + 0.011957759968936443, + 0.015794340521097183, + 0.015527293086051941, + 0.014773797243833542, + 0.010073546320199966, + 0.005096029490232468, + 0.004705623257905245, + 0.005617198534309864, + 0.0034842980094254017, + 0.004061954561620951, + 0.0034423202741891146, + 0.004236387554556131, + 0.004039564169943333, + 0.012275985442101955, + 0.009507754817605019, + 0.012441382743418217, + 0.007519643288105726, + 0.006349717266857624, + 0.007599830627441406, + 0.0067614223808050156, + 0.011644426733255386, + 0.0020192237570881844, + 0.0026261312887072563, + 0.0026808809489011765, + 0.0022293522488325834, + 0.0033634880091995, + 0.0017900383099913597, + 0.001906012766994536, + 0.002769414335489273, + 0.004426984582096338, + 0.0022498457692563534, + 0.004544237162917852, + 0.0022932991851121187, + 0.0025829786900430918, + 0.0029108435846865177, + 0.002437066752463579, + 0.007699967361986637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.026358501985669136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204469874501228 + ], + "opposite_logprob_tokens": [ + -2.812044620513916 + ], + "logprob_reward": -0.06204469874501228, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204469874501228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11829216033220291, + 0.0579427145421505, + 0.13138079643249512, + 0.11360009014606476, + 0.27554890513420105, + 0.1414320319890976, + 0.059999581426382065, + 0.050935085862874985, + 0.026600316166877747, + 0.024268273264169693 + ], + "credit_map": [ + 0.008591355755925179, + 0.016687210649251938, + 0.01602829620242119, + 0.018625732511281967, + 0.01984916254878044, + 0.015902167186141014, + 0.011616784147918224, + 0.010991459712386131, + 0.009135843254625797, + 0.006902811583131552, + 0.00506257452070713, + 0.007204791530966759, + 0.006868245545774698, + 0.011149480007588863, + 0.006458464544266462, + 0.005160505883395672, + 0.010622330941259861, + 0.021840527653694153, + 0.009491820819675922, + 0.015630727633833885, + 0.026141755282878876, + 0.015968549996614456, + 0.014360661618411541, + 0.017324429005384445, + 0.0196540504693985, + 0.02362222969532013, + 0.008623502217233181, + 0.006835730746388435, + 0.026251086965203285, + 0.014950877986848354, + 0.00657344376668334, + 0.007089169695973396, + 0.03113478422164917, + 0.041443511843681335, + 0.02294991724193096, + 0.03751140460371971, + 0.043145060539245605, + 0.022160187363624573, + 0.049727506935596466, + 0.02747654728591442, + 0.012845268473029137, + 0.016851309686899185, + 0.014540242031216621, + 0.017639901489019394, + 0.02346348762512207, + 0.014962720684707165, + 0.02195521630346775, + 0.019173890352249146, + 0.005652236752212048, + 0.009243510663509369, + 0.0074232532642781734, + 0.008454504422843456, + 0.009013225324451923, + 0.005965390708297491, + 0.007401946000754833, + 0.006845518480986357, + 0.009387675672769547, + 0.00663910573348403, + 0.006650780327618122, + 0.007015058305114508, + 0.00639575719833374, + 0.005476022604852915, + 0.004906814079731703, + 0.004463869612663984, + 0.003332502208650112, + 0.002781183924525976, + 0.004264324437826872, + 0.003652129089459777, + 0.0042999559082090855, + 0.0021395476069301367, + 0.003125034971162677, + 0.0030056394170969725, + 0.0026599320117384195, + 0.002656118478626013, + 0.0034169543068856, + 0.0034402210731059313, + 0.00423238892108202, + 0.002417556941509247, + 0.002549082040786743, + 0.0028960180934518576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.06204469874501228 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775033220648766 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775033220648766, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775033220648766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08086349070072174, + 0.05842761695384979, + 0.15377698838710785, + 0.06564784049987793, + 0.14610575139522552, + 0.11617181450128555, + 0.08491046726703644, + 0.06672442704439163, + 0.13213638961315155, + 0.09523516148328781 + ], + "credit_map": [ + 0.0054052285850048065, + 0.006723384372889996, + 0.014466541819274426, + 0.023586703464388847, + 0.008076661266386509, + 0.008031218312680721, + 0.007390453480184078, + 0.007183300796896219, + 0.006828405428677797, + 0.006669091060757637, + 0.005585978738963604, + 0.010638869367539883, + 0.006827032193541527, + 0.008327152580022812, + 0.005986163392663002, + 0.007564920000731945, + 0.016583425924181938, + 0.026156753301620483, + 0.013231823220849037, + 0.009862027131021023, + 0.03178250417113304, + 0.017245333641767502, + 0.012398148886859417, + 0.026516977697610855, + 0.008364591747522354, + 0.011796767823398113, + 0.004987625405192375, + 0.007005790714174509, + 0.008200005628168583, + 0.010781610384583473, + 0.0088625093922019, + 0.005648937076330185, + 0.022335302084684372, + 0.01633569970726967, + 0.016211673617362976, + 0.019344763830304146, + 0.01759972795844078, + 0.017093919217586517, + 0.022204803302884102, + 0.014979861676692963, + 0.012253455817699432, + 0.013129381462931633, + 0.027957884594798088, + 0.011324173770844936, + 0.010861271061003208, + 0.011143472976982594, + 0.015334106981754303, + 0.01416806410998106, + 0.011416133493185043, + 0.010725125670433044, + 0.012693814933300018, + 0.009872308000922203, + 0.009898184798657894, + 0.009758411906659603, + 0.010027974843978882, + 0.010518516413867474, + 0.008844796568155289, + 0.008522755466401577, + 0.011598614044487476, + 0.008023052476346493, + 0.006534271407872438, + 0.007380163297057152, + 0.005873582791537046, + 0.009947191923856735, + 0.020189739763736725, + 0.010580761358141899, + 0.020759975537657738, + 0.008984730578958988, + 0.01224124152213335, + 0.013506480492651463, + 0.011441720649600029, + 0.03443174436688423, + 0.011195185594260693, + 0.008438467979431152, + 0.013504999689757824, + 0.006121558602899313, + 0.008540317416191101, + 0.010703418403863907, + 0.007919969968497753, + 0.02881123684346676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.009775033220648766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with an extreme close-up of a hand, then tilts vertically upward, ultimately framing a medium shot. The camera focuses on a pair of slender, powerful hands, their fingertips swiping, tapping, and zooming with an elegant yet decisive rhythm across a large, borderless, transparent touchpad suspended in mid-air. As he manipulates, the screen's vibrant UI elements—composed of complex cyan data streams, dynamic geometric icons, and three-dimensional models—respond to his every command in real time. As the camera steadily and slowly pans upward, we see him dressed in a meticulously tailored, dark indigo-blue Neo-Chinese formal suit. The silk and wool blend gleams subtly under the soft overhead lighting. The minimalist stand-up collar and several obsidian buttons embody a modern Eastern aesthetic. He stands tall and composed, sitting calmly in an ergonomic chair. Finally, the camera freezes on his face, a chiseled, calm, oriental figure with immaculately short hair. His eyes are focused and penetrating, and a confident smile evokes a sense of control. As the camera widens, the background behind him becomes fully revealed: a minimalist office atop a skyscraper. Through a wall of floor-to-ceiling glass windows, a magnificent futuristic skyline unfolds. Countless whimsical aircraft dart among the meticulously arranged, monolithic buildings. The distant sky, stained a warm orange-gold by the setting sun, forms a stark contrast to the cool, technological interior. The entire scene exudes a tranquil, focused, and futuristic atmosphere of elite élite.", + "seed": 6821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_1/index_2715_seed_6821_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07323388196527958, + "vlm_logprob": -0.07323388196527958, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.07323388196527958, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25469911098480225 + ], + "opposite_logprob_tokens": [ + -1.5046991109848022 + ], + "logprob_reward": -0.25469911098480225, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25469911098480225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17291775345802307, + 0.14908099174499512, + 0.2006673514842987, + 0.09674543887376785, + 0.08185417950153351, + 0.03424438089132309, + 0.05373455956578255, + 0.06853526085615158, + 0.11593003571033478, + 0.026290029287338257 + ], + "credit_map": [ + 0.01923512853682041, + 0.0281424168497324, + 0.01418788731098175, + 0.015526424162089825, + 0.01589677855372429, + 0.024218708276748657, + 0.030779991298913956, + 0.024930430576205254, + 0.03239162266254425, + 0.01742730289697647, + 0.01491664070636034, + 0.01175543200224638, + 0.024786069989204407, + 0.024296240881085396, + 0.01340917032212019, + 0.01009850762784481, + 0.03357313200831413, + 0.021492088213562965, + 0.02184881642460823, + 0.014390194788575172, + 0.030833737924695015, + 0.030546734109520912, + 0.036663636565208435, + 0.011319011449813843, + 0.007161627057939768, + 0.012178685516119003, + 0.007825275883078575, + 0.008624689653515816, + 0.021384339779615402, + 0.01692209206521511, + 0.014280801638960838, + 0.00836792029440403, + 0.01004901435226202, + 0.006850116420537233, + 0.00939187966287136, + 0.007742155343294144, + 0.01692206785082817, + 0.015972906723618507, + 0.007628331426531076, + 0.007297712843865156, + 0.0036633205600082874, + 0.004188117571175098, + 0.006151970475912094, + 0.0032483316026628017, + 0.0038978278171271086, + 0.0036132631357759237, + 0.00590983172878623, + 0.0035717166028916836, + 0.007555323652923107, + 0.009182976558804512, + 0.00658020144328475, + 0.0038852093275636435, + 0.00819164514541626, + 0.006199207156896591, + 0.007551809307187796, + 0.004588182549923658, + 0.010380842722952366, + 0.011403198353946209, + 0.006751656997948885, + 0.005241051781922579, + 0.011434967629611492, + 0.010249217040836811, + 0.007129020988941193, + 0.005945312324911356, + 0.01532808132469654, + 0.021674225106835365, + 0.0183216854929924, + 0.011537306942045689, + 0.012773958034813404, + 0.017197132110595703, + 0.010437393561005592, + 0.008660262450575829, + 0.005049037281423807, + 0.004955886863172054, + 0.003989014774560928, + 0.0023988361936062574, + 0.0018711090087890625, + 0.004595227539539337, + 0.0018271973822265863, + 0.001603719312697649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.25469911098480225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011105512268841267 + ], + "opposite_logprob_tokens": [ + -4.511105537414551 + ], + "logprob_reward": -0.011105512268841267, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011105512268841267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14391687512397766, + 0.0432242788374424, + 0.08152913302183151, + 0.1821390986442566, + 0.1022467166185379, + 0.08099263161420822, + 0.08933243900537491, + 0.05325502157211304, + 0.1200183629989624, + 0.10334543138742447 + ], + "credit_map": [ + 0.01120186410844326, + 0.02501588501036167, + 0.00949216727167368, + 0.012263172306120396, + 0.013417388312518597, + 0.024242324754595757, + 0.02847660519182682, + 0.0198074858635664, + 0.005425512790679932, + 0.008520388044416904, + 0.0032863093074411154, + 0.004733934532850981, + 0.007085958495736122, + 0.006392539944499731, + 0.003993686754256487, + 0.00378594733774662, + 0.017170917242765427, + 0.007036182098090649, + 0.010105849243700504, + 0.0064829271286726, + 0.009212816134095192, + 0.012762825936079025, + 0.009815200231969357, + 0.008942416869103909, + 0.01796777918934822, + 0.02138080634176731, + 0.016157984733581543, + 0.019249388948082924, + 0.0425560437142849, + 0.03394123539328575, + 0.01679590344429016, + 0.014089947566390038, + 0.00596648408100009, + 0.007969752885401249, + 0.005984615534543991, + 0.01586839370429516, + 0.025313537567853928, + 0.01653839647769928, + 0.016681557521224022, + 0.007923980243504047, + 0.008288788609206676, + 0.008131925947964191, + 0.00893898494541645, + 0.014051727019250393, + 0.009915169328451157, + 0.008754095062613487, + 0.013191568665206432, + 0.009720368310809135, + 0.009559373371303082, + 0.015108741819858551, + 0.009477236308157444, + 0.0077447486110031605, + 0.01895492523908615, + 0.007814360782504082, + 0.01342487521469593, + 0.00724817905575037, + 0.00806758739054203, + 0.011227828450500965, + 0.006192936096340418, + 0.004044249188154936, + 0.009948620572686195, + 0.004303177818655968, + 0.006120813544839621, + 0.0033498117700219154, + 0.023956293240189552, + 0.017415093258023262, + 0.014561466872692108, + 0.011852657422423363, + 0.01687128096818924, + 0.012380999512970448, + 0.013280028477311134, + 0.0097005320712924, + 0.016462741419672966, + 0.025015493854880333, + 0.011462009511888027, + 0.009725062176585197, + 0.007249790243804455, + 0.018547050654888153, + 0.007224254310131073, + 0.0076590231619775295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.011105512268841267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307687282562256 + ], + "opposite_logprob_tokens": [ + -3.168076992034912 + ], + "logprob_reward": -0.04307687282562256, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307687282562256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16481705009937286, + 0.08624911308288574, + 0.34695833921432495, + 0.08078822493553162, + 0.06221194565296173, + 0.043977040797472, + 0.0478043258190155, + 0.03829681873321533, + 0.058788251131772995, + 0.07010892778635025 + ], + "credit_map": [ + 0.01917201280593872, + 0.03125682473182678, + 0.012641758657991886, + 0.01592647098004818, + 0.017636889591813087, + 0.02538236789405346, + 0.020422812551259995, + 0.022377921268343925, + 0.009297827258706093, + 0.010104745626449585, + 0.015427427366375923, + 0.007840573787689209, + 0.011546012945473194, + 0.01970047876238823, + 0.006189138628542423, + 0.006142910104244947, + 0.03092837892472744, + 0.029597917571663857, + 0.02700294367969036, + 0.026258215308189392, + 0.04361160099506378, + 0.05857967957854271, + 0.0990460142493248, + 0.031933579593896866, + 0.008173944428563118, + 0.008137187920510769, + 0.011029156856238842, + 0.007540393620729446, + 0.01362139917910099, + 0.014153581112623215, + 0.011622370220720768, + 0.006510196719318628, + 0.007092637475579977, + 0.006467609200626612, + 0.006008324213325977, + 0.01153850182890892, + 0.011228310875594616, + 0.009108874015510082, + 0.006108645349740982, + 0.0046590412966907024, + 0.004158874042332172, + 0.00458461232483387, + 0.00751424441114068, + 0.007401191629469395, + 0.004448599647730589, + 0.0050277202390134335, + 0.006445110309869051, + 0.004396687727421522, + 0.0051096524111926556, + 0.005981266964226961, + 0.009379438124597073, + 0.004207677207887173, + 0.006136867683380842, + 0.007225519511848688, + 0.0063801719807088375, + 0.0033837314695119858, + 0.0052781119011342525, + 0.008173375390470028, + 0.006180762778967619, + 0.002850822638720274, + 0.005131681449711323, + 0.003975844942033291, + 0.003697056323289871, + 0.003009161911904812, + 0.012782379053533077, + 0.010214854963123798, + 0.009007482789456844, + 0.004925453569740057, + 0.007054999936372042, + 0.005747811868786812, + 0.005858093034476042, + 0.0031971773132681847, + 0.012767213396728039, + 0.016737742349505424, + 0.008927133865654469, + 0.006808701902627945, + 0.0047851502895355225, + 0.007730930577963591, + 0.0049492777325212955, + 0.007402779068797827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.04307687282562256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865625500679016 + ], + "opposite_logprob_tokens": [ + -3.0486562252044678 + ], + "logprob_reward": -0.04865625500679016, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865625500679016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07139712572097778, + 0.09903267025947571, + 0.2617732584476471, + 0.07693599909543991, + 0.14221833646297455, + 0.1288948357105255, + 0.07806888222694397, + 0.03951111063361168, + 0.035152893513441086, + 0.0670149177312851 + ], + "credit_map": [ + 0.010340333916246891, + 0.009933303110301495, + 0.006217595189809799, + 0.007201661355793476, + 0.005709019023925066, + 0.010230072773993015, + 0.011874278075993061, + 0.009890862740576267, + 0.018132725730538368, + 0.009502335451543331, + 0.012034174986183643, + 0.00922731775790453, + 0.013193052262067795, + 0.01737855188548565, + 0.011151510290801525, + 0.008412998169660568, + 0.02837412804365158, + 0.02294890396296978, + 0.017331287264823914, + 0.01999092474579811, + 0.031050126999616623, + 0.04470769688487053, + 0.07447638362646103, + 0.022893797606229782, + 0.008650113828480244, + 0.0053480579517781734, + 0.006647901143878698, + 0.008831492625176907, + 0.01706056296825409, + 0.014682624489068985, + 0.010789026506245136, + 0.0049262153916060925, + 0.008927359245717525, + 0.011447572149336338, + 0.009688255377113819, + 0.013313911855220795, + 0.03765200823545456, + 0.026262342929840088, + 0.020052017644047737, + 0.014874860644340515, + 0.01328333280980587, + 0.012374013662338257, + 0.0136492270976305, + 0.0129139618948102, + 0.020068978890776634, + 0.015425002202391624, + 0.021493982523679733, + 0.019686341285705566, + 0.011983274482190609, + 0.01459509041160345, + 0.011093447916209698, + 0.005694894585758448, + 0.011178803630173206, + 0.006019365042448044, + 0.010802613571286201, + 0.006701385602355003, + 0.0059625194407999516, + 0.006452135741710663, + 0.0042334687896072865, + 0.002546758856624365, + 0.0044150990433990955, + 0.007586727850139141, + 0.004146946128457785, + 0.004167455714195967, + 0.004922683350741863, + 0.004876773804426193, + 0.006057148799300194, + 0.0041981227695941925, + 0.004318187013268471, + 0.003225360531359911, + 0.004729956854134798, + 0.0028246610891073942, + 0.008988487534224987, + 0.007362052332609892, + 0.00783185102045536, + 0.011753850616514683, + 0.00572250597178936, + 0.006749809719622135, + 0.006132279522716999, + 0.01247408241033554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.04865625500679016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shown tilting vertically upward from the close-up to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands shown swiping, tapping, and zooming on a large, borderless, transparent touchpad suspended in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a dark indigo-blue Neo-Chinese formal suit with a minimalist stand-up collar and obsidian buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background reveal a minimalist office atop a skyscraper with a futuristic skyline and warm orange-gold sunset sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631658740341663 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.008631658740341663, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631658740341663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1799893081188202, + 0.07896950840950012, + 0.23522551357746124, + 0.05232436954975128, + 0.06033584102988243, + 0.05045027658343315, + 0.059362541884183884, + 0.08573977649211884, + 0.1025419682264328, + 0.09506089240312576 + ], + "credit_map": [ + 0.023594245314598083, + 0.030781812965869904, + 0.010496189817786217, + 0.01545722596347332, + 0.02044018730521202, + 0.02863181196153164, + 0.03168600797653198, + 0.01890181563794613, + 0.009656399488449097, + 0.012858834117650986, + 0.012140314094722271, + 0.007735594175755978, + 0.010076056234538555, + 0.01474690344184637, + 0.0063596260733902454, + 0.005395777057856321, + 0.026473423466086388, + 0.018139570951461792, + 0.018919222056865692, + 0.018075035884976387, + 0.026467634364962578, + 0.03952610865235329, + 0.06733014434576035, + 0.02029435709118843, + 0.005477504804730415, + 0.005285787396132946, + 0.005859497468918562, + 0.0052527571097016335, + 0.012005215510725975, + 0.009061696007847786, + 0.006072956137359142, + 0.0033089558128267527, + 0.004831031430512667, + 0.0053981295786798, + 0.00434604799374938, + 0.00709323026239872, + 0.012796560302376747, + 0.00934328231960535, + 0.011040476150810719, + 0.005487085320055485, + 0.004653266631066799, + 0.004744687583297491, + 0.008540334179997444, + 0.005414870101958513, + 0.006080242805182934, + 0.005879843141883612, + 0.008504939265549183, + 0.006632091943174601, + 0.005038272123783827, + 0.007109438069164753, + 0.008594347164034843, + 0.005164346192032099, + 0.013748724944889545, + 0.006819659844040871, + 0.0088412556797266, + 0.004046497400850058, + 0.015064732171595097, + 0.0220845527946949, + 0.015604062005877495, + 0.00461776414886117, + 0.009005120024085045, + 0.0075679500587284565, + 0.00733310217037797, + 0.004462492186576128, + 0.016390100121498108, + 0.018870847299695015, + 0.01730913110077381, + 0.007912036031484604, + 0.012007126584649086, + 0.012623789720237255, + 0.010772737674415112, + 0.006656199228018522, + 0.016936279833316803, + 0.016048308461904526, + 0.019709160551428795, + 0.007861019112169743, + 0.008546676486730576, + 0.013774821534752846, + 0.0065938569605350494, + 0.005590763874351978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.008631658740341663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_tensor([9876, 9876])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_tensor([9876, 9876])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d00b0e905b96f07e401819f2a9a2e6ae5df2de71 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_tensor([9876, 9876])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from above focuses on an elderly man, his face etched with the marks of time. His calloused, vein-lined hands firmly grip a wooden chisel, its handle rubbed to a sleek sheen. A warm light, streaming from above and sideways, casts deep, three-dimensional shadows on his ridged face and hands, while also illuminating the sharp, metallic gleam of the blade. His eyes, narrowed slightly, his gaze intense, his gaze fixed on the work in hand. With a steady, deliberate rhythm, the tip of the chisel deftly pushes, whittles, and carves across a block of uncarved, light-colored wood. With each precise movement, tiny, curled, thin chips of wood gracefully peel from the sharp blade, falling like a gentle snowflake onto a workbench cluttered with tools and unfinished work. He was wearing a plain dark blue cotton shirt with the cuffs rolled up to his forearms. The whole scene, under warm lighting, with wood and dark blue as the main colors, presented an atmosphere of extreme tranquility, concentration and craftsmanship.", + "seed": 19537, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_9876_seed_19537_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08284283354878426, + "vlm_logprob": -0.08284283354878426, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.08284283354878426, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10984373837709427 + ], + "opposite_logprob_tokens": [ + -2.3598437309265137 + ], + "logprob_reward": -0.10984373837709427, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10984373837709427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26737597584724426, + 0.05385252460837364, + 0.0436047725379467, + 0.10097114741802216, + 0.03963875770568848, + 0.09603346139192581, + 0.1725578010082245, + 0.14444416761398315, + 0.016475996002554893, + 0.06504538655281067 + ], + "credit_map": [ + 0.023040443658828735, + 0.0173996202647686, + 0.0258257444947958, + 0.0257383082062006, + 0.0759970024228096, + 0.03368382900953293, + 0.016726942732930183, + 0.04896407946944237, + 0.005041705910116434, + 0.0047891270369291306, + 0.005379729438573122, + 0.0037694969214498997, + 0.01235781330615282, + 0.008499911054968834, + 0.006994155701249838, + 0.007020586170256138, + 0.004053089302033186, + 0.004011016339063644, + 0.007264301646500826, + 0.0034768502227962017, + 0.0053078713826835155, + 0.01084744930267334, + 0.0028077554889023304, + 0.005836437921971083, + 0.008263031020760536, + 0.014100160449743271, + 0.00996413268148899, + 0.004948367364704609, + 0.02373029664158821, + 0.02362576499581337, + 0.00878160260617733, + 0.007557795383036137, + 0.004003103356808424, + 0.003506146604195237, + 0.007058779709041119, + 0.004227465018630028, + 0.005864959210157394, + 0.005531732924282551, + 0.005597520619630814, + 0.0038490514270961285, + 0.009393577463924885, + 0.008435669355094433, + 0.009557211771607399, + 0.007333642803132534, + 0.023000134155154228, + 0.0174779761582613, + 0.013077965006232262, + 0.007757284212857485, + 0.016605617478489876, + 0.013953560963273048, + 0.026756485924124718, + 0.018766067922115326, + 0.028100764378905296, + 0.027756342664361, + 0.02455347403883934, + 0.01606549136340618, + 0.016722558066248894, + 0.010591312311589718, + 0.020947208628058434, + 0.016421202570199966, + 0.018589744344353676, + 0.026861092075705528, + 0.01647225208580494, + 0.017838792875409126, + 0.0017020744271576405, + 0.0013677871320396662, + 0.0020539378747344017, + 0.0014970502816140652, + 0.0027248472906649113, + 0.0038054336328059435, + 0.0016307305777445436, + 0.0016941347857937217, + 0.007254956290125847, + 0.005216415971517563, + 0.009119140915572643, + 0.007466311100870371, + 0.007729391101747751, + 0.010221419855952263, + 0.006439297925680876, + 0.011598451063036919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.10984373837709427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909573405981064 + ], + "opposite_logprob_tokens": [ + -2.4640958309173584 + ], + "logprob_reward": -0.08909573405981064, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909573405981064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1861480474472046, + 0.11541581898927689, + 0.039516426622867584, + 0.1287488490343094, + 0.10105409473180771, + 0.10559845715761185, + 0.11263179779052734, + 0.10713902860879898, + 0.028169946745038033, + 0.07557752728462219 + ], + "credit_map": [ + 0.018346788361668587, + 0.015360252000391483, + 0.022416571155190468, + 0.026863835752010345, + 0.037324029952287674, + 0.02176717296242714, + 0.01957010105252266, + 0.02449928969144821, + 0.008792659267783165, + 0.01587182842195034, + 0.011648915708065033, + 0.010445916093885899, + 0.01972874440252781, + 0.01745612546801567, + 0.010386915877461433, + 0.021084710955619812, + 0.003555083414539695, + 0.006148321554064751, + 0.005459876731038094, + 0.0047104270197451115, + 0.006462771445512772, + 0.005308499094098806, + 0.0028539265040308237, + 0.005017517600208521, + 0.010782289318740368, + 0.017619410529732704, + 0.013586947694420815, + 0.006036846898496151, + 0.03273879736661911, + 0.02843133732676506, + 0.010422165505588055, + 0.00913105346262455, + 0.009506742469966412, + 0.009327765554189682, + 0.011667652986943722, + 0.007104032207280397, + 0.011098933406174183, + 0.028617404401302338, + 0.008208969607949257, + 0.015522588044404984, + 0.012162700295448303, + 0.01831553503870964, + 0.012742556631565094, + 0.006452133413404226, + 0.019763080403208733, + 0.021289784461259842, + 0.007070778403431177, + 0.007801887579262257, + 0.008806806057691574, + 0.013163522817194462, + 0.015163667500019073, + 0.011295019648969173, + 0.02754819579422474, + 0.017239557579159737, + 0.008507998660206795, + 0.010907025076448917, + 0.011047075502574444, + 0.0158252976834774, + 0.012834051623940468, + 0.006777254398912191, + 0.027971358969807625, + 0.01686946675181389, + 0.006839797832071781, + 0.008974719792604446, + 0.00276271253824234, + 0.0027331500314176083, + 0.0053365593776106834, + 0.001979572931304574, + 0.006250824313610792, + 0.0051573351956903934, + 0.0018520194571465254, + 0.002097771968692541, + 0.009478208608925343, + 0.005552751012146473, + 0.011737075634300709, + 0.00829303078353405, + 0.009332647547125816, + 0.01591327041387558, + 0.007190981414169073, + 0.008079558610916138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.08909573405981064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295885592699051 + ], + "opposite_logprob_tokens": [ + -2.2379589080810547 + ], + "logprob_reward": -0.11295885592699051, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295885592699051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22122816741466522, + 0.06287974864244461, + 0.06843844801187515, + 0.1497904658317566, + 0.1326477974653244, + 0.04919207841157913, + 0.044480469077825546, + 0.042562466114759445, + 0.03204271197319031, + 0.19673766195774078 + ], + "credit_map": [ + 0.01537785492837429, + 0.011972605250775814, + 0.019318703562021255, + 0.01867799460887909, + 0.054936785250902176, + 0.031971707940101624, + 0.03613859415054321, + 0.03283390775322914, + 0.005617945455014706, + 0.006269098725169897, + 0.00721420394256711, + 0.006482240278273821, + 0.006294518243521452, + 0.009734533727169037, + 0.007017233408987522, + 0.014249974861741066, + 0.0059040929190814495, + 0.004537659231573343, + 0.007599745411425829, + 0.009005594067275524, + 0.005187796428799629, + 0.01428927294909954, + 0.007966043427586555, + 0.013948243111371994, + 0.010562719777226448, + 0.017843667417764664, + 0.019226713106036186, + 0.010770251043140888, + 0.03343712538480759, + 0.02603275328874588, + 0.018357165157794952, + 0.01356007531285286, + 0.010328540578484535, + 0.00776389567181468, + 0.015486713498830795, + 0.011783461086452007, + 0.011805112473666668, + 0.03377620875835419, + 0.019020840525627136, + 0.022683026269078255, + 0.004801941104233265, + 0.004827361088246107, + 0.004835714120417833, + 0.0043370663188397884, + 0.011073273606598377, + 0.009146591648459435, + 0.0058046020567417145, + 0.0043655261397361755, + 0.0034640352241694927, + 0.004672158043831587, + 0.005238868296146393, + 0.0063189547508955, + 0.006761211436241865, + 0.005933692678809166, + 0.0064547001384198666, + 0.005636849440634251, + 0.0048446268774569035, + 0.0037776376120746136, + 0.00545887416228652, + 0.004650198854506016, + 0.006858555134385824, + 0.00669172452762723, + 0.005043118726462126, + 0.005237730219960213, + 0.003481397172436118, + 0.002151288790628314, + 0.0047807516530156136, + 0.00309210317209363, + 0.004405428655445576, + 0.004751865286380053, + 0.005032822489738464, + 0.0043470547534525394, + 0.025694457814097404, + 0.01331738568842411, + 0.028504831716418266, + 0.01893671602010727, + 0.019967790693044662, + 0.03751498833298683, + 0.01601104624569416, + 0.03679043799638748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.11295885592699051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790116935968399 + ], + "opposite_logprob_tokens": [ + -2.5790116786956787 + ], + "logprob_reward": -0.0790116935968399, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790116935968399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09227132052183151, + 0.13459360599517822, + 0.048189807683229446, + 0.09361931681632996, + 0.03271842002868652, + 0.06442388147115707, + 0.11807673424482346, + 0.07534503191709518, + 0.08353883028030396, + 0.25722306966781616 + ], + "credit_map": [ + 0.008529839105904102, + 0.005258030723780394, + 0.009336243383586407, + 0.00678508123382926, + 0.01885577104985714, + 0.010279618203639984, + 0.011268869042396545, + 0.021957872435450554, + 0.01058630645275116, + 0.012389847077429295, + 0.015472895465791225, + 0.014459317550063133, + 0.01805342361330986, + 0.022472886368632317, + 0.019350532442331314, + 0.021808398887515068, + 0.005813938099890947, + 0.00560106011107564, + 0.005142660811543465, + 0.003702440531924367, + 0.009486019611358643, + 0.009753636084496975, + 0.004184330347925425, + 0.004505719989538193, + 0.008563351817429066, + 0.013693970628082752, + 0.009364933706820011, + 0.005064558703452349, + 0.020523682236671448, + 0.021257581189274788, + 0.006512993481010199, + 0.008638239465653896, + 0.003329369006678462, + 0.0029912791214883327, + 0.004286055453121662, + 0.002674223855137825, + 0.005901218857616186, + 0.007171133998781443, + 0.002909749746322632, + 0.0034553890582174063, + 0.006356451660394669, + 0.0056250677444040775, + 0.006732707843184471, + 0.005410708487033844, + 0.015073876827955246, + 0.011265791952610016, + 0.008737146854400635, + 0.005222138483077288, + 0.011454404331743717, + 0.012249493040144444, + 0.014525321312248707, + 0.014349795877933502, + 0.015508840791881084, + 0.016742641106247902, + 0.01795259490609169, + 0.015293640084564686, + 0.0092315673828125, + 0.006996312644332647, + 0.009966818615794182, + 0.00852957833558321, + 0.010684506967663765, + 0.012202270328998566, + 0.008114330470561981, + 0.009619646705687046, + 0.009620926342904568, + 0.007840298116207123, + 0.012031534686684608, + 0.008908580988645554, + 0.012658433988690376, + 0.013238539919257164, + 0.008956024423241615, + 0.010284491814672947, + 0.03403254225850105, + 0.01972954347729683, + 0.043985314667224884, + 0.02749485895037651, + 0.03050096333026886, + 0.03947801515460014, + 0.030180256813764572, + 0.03182158246636391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.0790116935968399 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330414578318596 + ], + "opposite_logprob_tokens": [ + -3.7733042240142822 + ], + "logprob_reward": -0.02330414578318596, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330414578318596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4741179347038269, + 0.0600963793694973, + 0.03618092089891434, + 0.06881771981716156, + 0.07573386281728745, + 0.04734247550368309, + 0.06632764637470245, + 0.06241859868168831, + 0.029346726834774017, + 0.0796176940202713 + ], + "credit_map": [ + 0.023721784353256226, + 0.02465706877410412, + 0.033096540719270706, + 0.02134380303323269, + 0.16549712419509888, + 0.07239008694887161, + 0.02114775776863098, + 0.11226377636194229, + 0.005700042936950922, + 0.006715449038892984, + 0.0052939956076443195, + 0.005637413822114468, + 0.013315043412148952, + 0.01408281922340393, + 0.00286581227555871, + 0.006485801190137863, + 0.0031702604610472918, + 0.002871127799153328, + 0.004366864915937185, + 0.003391926409676671, + 0.0058017005212605, + 0.008825989440083504, + 0.0022913264110684395, + 0.005461724940687418, + 0.005967431701719761, + 0.010181603021919727, + 0.0077234492637217045, + 0.004725123289972544, + 0.016704337671399117, + 0.011684244498610497, + 0.004993036855012178, + 0.00683849211782217, + 0.005445609334856272, + 0.00602908106520772, + 0.005457662511616945, + 0.008478574454784393, + 0.010114278644323349, + 0.016326140612363815, + 0.005885581020265818, + 0.017996933311223984, + 0.004222485702484846, + 0.004562072921544313, + 0.0049151950515806675, + 0.0031100569758564234, + 0.012912620790302753, + 0.009245486930012703, + 0.0045892782509326935, + 0.003785279579460621, + 0.0057417768985033035, + 0.005969429388642311, + 0.008078659884631634, + 0.009879950433969498, + 0.008921264670789242, + 0.009723193012177944, + 0.009817378595471382, + 0.008196000009775162, + 0.007997306063771248, + 0.004883342422544956, + 0.009372061118483543, + 0.00809022318571806, + 0.009787063114345074, + 0.009278925135731697, + 0.0058019207790493965, + 0.007207757327705622, + 0.003058925736695528, + 0.0025039748288691044, + 0.00527340779080987, + 0.003147116396576166, + 0.004075729753822088, + 0.004332580603659153, + 0.0036573451943695545, + 0.003297643968835473, + 0.008205700665712357, + 0.006147804204374552, + 0.01016546506434679, + 0.006244870834052563, + 0.010410327464342117, + 0.01655423454940319, + 0.012217523530125618, + 0.009671764448285103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.02330414578318596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from above focuses on an elderly man, his face etched with the marks of time. His calloused, vein-lined hands firmly grip a wooden chisel, its handle rubbed to a sleek sheen. A warm light, streaming from above and sideways, casts deep, three-dimensional shadows on his ridged face and hands, while also illuminating the sharp, metallic gleam of the blade. His eyes, narrowed slightly, his gaze intense, his gaze fixed on the work in hand. With a steady, deliberate rhythm, the tip of the chisel deftly pushes, whittles, and carves across a block of uncarved, light-colored wood. With each precise movement, tiny, curled, thin chips of wood gracefully peel from the sharp blade, falling like a gentle snowflake onto a workbench cluttered with tools and unfinished work. He was wearing a plain dark blue cotton shirt with the cuffs rolled up to his forearms. The whole scene, under warm lighting, with wood and dark blue as the main colors, presented an atmosphere of extreme tranquility, concentration and craftsmanship.", + "seed": 19538, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_9876_seed_19538_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03934537954628468, + "vlm_logprob": -0.03934537954628468, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.03934537954628468, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0283928532153368 + ], + "opposite_logprob_tokens": [ + -4.278392791748047 + ], + "logprob_reward": -0.0283928532153368, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0283928532153368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15761257708072662, + 0.10796944051980972, + 0.07460344582796097, + 0.09333506971597672, + 0.044418517500162125, + 0.07222310453653336, + 0.02861246094107628, + 0.23117366433143616, + 0.10659505426883698, + 0.08345663547515869 + ], + "credit_map": [ + 0.025644224137067795, + 0.028946658596396446, + 0.010954318568110466, + 0.01596827618777752, + 0.013827885501086712, + 0.017724357545375824, + 0.026150958612561226, + 0.018395913764834404, + 0.01417313702404499, + 0.02164757438004017, + 0.01259356364607811, + 0.014961122535169125, + 0.009579011239111423, + 0.01378408819437027, + 0.008529759012162685, + 0.012701200321316719, + 0.008658025413751602, + 0.012487300671637058, + 0.004302836488932371, + 0.0061085703782737255, + 0.007815065793693066, + 0.01496434211730957, + 0.011233553290367126, + 0.009033753536641598, + 0.01292488444596529, + 0.016530180349946022, + 0.004754970781505108, + 0.008161074481904507, + 0.012252828106284142, + 0.012952962890267372, + 0.012908450327813625, + 0.012849720194935799, + 0.006713975686579943, + 0.008084731176495552, + 0.0037431837990880013, + 0.004530001897364855, + 0.004906110931187868, + 0.005584034603089094, + 0.0051676430739462376, + 0.005688840057700872, + 0.011509374715387821, + 0.012473097071051598, + 0.004950526636093855, + 0.007081950549036264, + 0.005846754647791386, + 0.0112239308655262, + 0.009050016291439533, + 0.010087452828884125, + 0.0034675560891628265, + 0.004113364964723587, + 0.0023475897032767534, + 0.0022254479117691517, + 0.0027476900722831488, + 0.005797585006803274, + 0.0034983260557055473, + 0.004414904396981001, + 0.026074279099702835, + 0.027407005429267883, + 0.01580631174147129, + 0.024873781949281693, + 0.02176462486386299, + 0.06226274371147156, + 0.018648236989974976, + 0.03433670476078987, + 0.009601615369319916, + 0.009167673997581005, + 0.006697629578411579, + 0.010428483597934246, + 0.013217806816101074, + 0.02754177711904049, + 0.012835392728447914, + 0.017104683443903923, + 0.012130111455917358, + 0.013628073036670685, + 0.006690926384180784, + 0.012369326315820217, + 0.006648025009781122, + 0.013270852155983448, + 0.00800265185534954, + 0.01071667019277811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0283928532153368 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901500165462494 + ], + "opposite_logprob_tokens": [ + -2.579015016555786 + ], + "logprob_reward": -0.07901500165462494, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901500165462494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1818777620792389, + 0.10726288706064224, + 0.09387709200382233, + 0.06340804696083069, + 0.09376512467861176, + 0.15297701954841614, + 0.018871871754527092, + 0.13363537192344666, + 0.039643142372369766, + 0.1146816611289978 + ], + "credit_map": [ + 0.030960790812969208, + 0.027483809739351273, + 0.012954083271324635, + 0.01264136377722025, + 0.024541666731238365, + 0.030658302828669548, + 0.018705813214182854, + 0.023931946605443954, + 0.014108553528785706, + 0.020888589322566986, + 0.009532286785542965, + 0.010955285280942917, + 0.010317903943359852, + 0.019804399460554123, + 0.010022702626883984, + 0.011633177287876606, + 0.009363595396280289, + 0.011900604702532291, + 0.005509673152118921, + 0.006368768867105246, + 0.007902611047029495, + 0.027541227638721466, + 0.012604741379618645, + 0.012685876339673996, + 0.007905928418040276, + 0.006962003652006388, + 0.004426291678100824, + 0.005230162758380175, + 0.009212040342390537, + 0.011969191953539848, + 0.009038407355546951, + 0.008664024993777275, + 0.010136240161955357, + 0.008492196910083294, + 0.004978502634912729, + 0.005030062049627304, + 0.008423189632594585, + 0.03224600479006767, + 0.00864164438098669, + 0.015817292034626007, + 0.022002199664711952, + 0.018335694447159767, + 0.00787779875099659, + 0.016676684841513634, + 0.012403811328113079, + 0.03760780021548271, + 0.015810981392860413, + 0.022262055426836014, + 0.002975903917104006, + 0.003311803564429283, + 0.0010325125185772777, + 0.0016063982620835304, + 0.0025431676767766476, + 0.0029459926299750805, + 0.0015764202689751983, + 0.0028796731494367123, + 0.016025464981794357, + 0.013918132521212101, + 0.011549347080290318, + 0.019354570657014847, + 0.012071735225617886, + 0.027310991659760475, + 0.014694949612021446, + 0.01871018297970295, + 0.00402603717520833, + 0.0026066822465509176, + 0.003389085177332163, + 0.006024384405463934, + 0.003399777226150036, + 0.008912386372685432, + 0.005620855838060379, + 0.005663939286023378, + 0.01946748048067093, + 0.012332419864833355, + 0.004622533451765776, + 0.007350632455199957, + 0.012765687890350819, + 0.030375583097338676, + 0.010399154387414455, + 0.017368175089359283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.07901500165462494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307024925947189 + ], + "opposite_logprob_tokens": [ + -3.1680703163146973 + ], + "logprob_reward": -0.04307024925947189, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307024925947189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08351728320121765, + 0.04338977858424187, + 0.09624631702899933, + 0.09857222437858582, + 0.06871656328439713, + 0.04930705949664116, + 0.024492407217621803, + 0.3121916353702545, + 0.06877737492322922, + 0.15478940308094025 + ], + "credit_map": [ + 0.01077378261834383, + 0.0179221760481596, + 0.005678911693394184, + 0.007074484135955572, + 0.007576412055641413, + 0.015062741935253143, + 0.008130506612360477, + 0.011298266239464283, + 0.005635617766529322, + 0.007618884090334177, + 0.004180193413048983, + 0.004924181383103132, + 0.0032433767337352037, + 0.008752251975238323, + 0.003627852536737919, + 0.00540741765871644, + 0.0108583839610219, + 0.020118020474910736, + 0.007781786844134331, + 0.007922293618321419, + 0.00786083284765482, + 0.020517513155937195, + 0.009544512256979942, + 0.011642970144748688, + 0.014461580663919449, + 0.020508598536252975, + 0.004997772164642811, + 0.01009862869977951, + 0.011901960708200932, + 0.012167776934802532, + 0.011804773472249508, + 0.012631128542125225, + 0.007805536035448313, + 0.008969850838184357, + 0.00410416116937995, + 0.008029156364500523, + 0.006374187767505646, + 0.018268153071403503, + 0.005550702568143606, + 0.009614807553589344, + 0.006430292502045631, + 0.006448401138186455, + 0.003995988052338362, + 0.0037557377945631742, + 0.003811257891356945, + 0.012301165610551834, + 0.005671015474945307, + 0.006893199402838945, + 0.003195408498868346, + 0.003356928937137127, + 0.0023444457910954952, + 0.0027695773169398308, + 0.0028326890897005796, + 0.0035759094171226025, + 0.0031061905901879072, + 0.0033112545497715473, + 0.03934971243143082, + 0.03447955474257469, + 0.023699631914496422, + 0.0519794225692749, + 0.02605578862130642, + 0.06107333302497864, + 0.03249974176287651, + 0.043054431676864624, + 0.011027229949831963, + 0.00692096957936883, + 0.004256722051650286, + 0.006200940348207951, + 0.007610601373016834, + 0.014056400395929813, + 0.006816508714109659, + 0.011887998320162296, + 0.025239940732717514, + 0.02258334495127201, + 0.012138882651925087, + 0.011435095220804214, + 0.019498687237501144, + 0.03175028786063194, + 0.012709710747003555, + 0.01943344436585903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.04307024925947189 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370743617415428 + ], + "opposite_logprob_tokens": [ + -3.4087073802948 + ], + "logprob_reward": -0.03370743617415428, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370743617415428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10283946245908737, + 0.0350954532623291, + 0.39343681931495667, + 0.03597181662917137, + 0.03418285772204399, + 0.03290423005819321, + 0.023072360083460808, + 0.1653875857591629, + 0.10596168786287308, + 0.07114770263433456 + ], + "credit_map": [ + 0.019199009984731674, + 0.02038850262761116, + 0.006050340831279755, + 0.006977180950343609, + 0.013765217736363411, + 0.011431502178311348, + 0.009013762697577477, + 0.016013948246836662, + 0.00428472924977541, + 0.005281406454741955, + 0.0037666389252990484, + 0.0038833017461001873, + 0.002892744494602084, + 0.007330383639782667, + 0.003353800857439637, + 0.00430244579911232, + 0.04446776211261749, + 0.08693455904722214, + 0.03226669505238533, + 0.036165714263916016, + 0.03993339091539383, + 0.06600959599018097, + 0.042433276772499084, + 0.0452258363366127, + 0.0045342957600951195, + 0.006532853469252586, + 0.0036013328935950994, + 0.0035473674070090055, + 0.004497735295444727, + 0.004632804077118635, + 0.003912219312041998, + 0.004713208880275488, + 0.0042851814068853855, + 0.002572842873632908, + 0.005495661403983831, + 0.0033211002591997385, + 0.0032652236986905336, + 0.0057299258187413216, + 0.004739378113299608, + 0.004773546010255814, + 0.003383610164746642, + 0.004459780640900135, + 0.0033295603934675455, + 0.0035409608390182257, + 0.0034312212374061346, + 0.006934914272278547, + 0.003948792815208435, + 0.003875390626490116, + 0.002724740421399474, + 0.0032916374038904905, + 0.00250574154779315, + 0.0027652624994516373, + 0.00264555006287992, + 0.0029208131600171328, + 0.0029286800418049097, + 0.0032899349462240934, + 0.019603760913014412, + 0.01976044289767742, + 0.011911499314010143, + 0.029142435640096664, + 0.0122300386428833, + 0.0316624641418457, + 0.018834611400961876, + 0.022242337465286255, + 0.013669689185917377, + 0.011861502192914486, + 0.011028057895600796, + 0.016346465796232224, + 0.011073528788983822, + 0.016448980197310448, + 0.011610042303800583, + 0.013923420570790768, + 0.00988392811268568, + 0.011078068986535072, + 0.006467391271144152, + 0.010871928185224533, + 0.006916827987879515, + 0.010146230459213257, + 0.006102370098233223, + 0.00968096312135458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.03370743617415428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541357427835464 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541357427835464, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541357427835464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16672283411026, + 0.050154559314250946, + 0.30398568511009216, + 0.04699352756142616, + 0.03702632337808609, + 0.044099558144807816, + 0.03178434073925018, + 0.13430951535701752, + 0.10205590724945068, + 0.08286773413419724 + ], + "credit_map": [ + 0.022433403879404068, + 0.026081353425979614, + 0.0136040560901165, + 0.016721157357096672, + 0.01706869527697563, + 0.019647564738988876, + 0.032561395317316055, + 0.01860521361231804, + 0.005971432197839022, + 0.00899912416934967, + 0.0045329974964261055, + 0.005876455456018448, + 0.004419609904289246, + 0.008764366619288921, + 0.006391461472958326, + 0.005199113395065069, + 0.04453955963253975, + 0.059732262045145035, + 0.023183751851320267, + 0.024129578843712807, + 0.03072376735508442, + 0.04719748720526695, + 0.03694510459899902, + 0.037534184753894806, + 0.006337849423289299, + 0.00844445638358593, + 0.002922358689829707, + 0.00513417785987258, + 0.006916467100381851, + 0.005602911580353975, + 0.005711938254535198, + 0.005923369899392128, + 0.0037426864728331566, + 0.002858431776985526, + 0.002458307659253478, + 0.0031772847287356853, + 0.0029174305964261293, + 0.011512722820043564, + 0.004677854478359222, + 0.005681606475263834, + 0.005221459548920393, + 0.00504978746175766, + 0.004932262934744358, + 0.005246543791145086, + 0.003086201846599579, + 0.0091380150988698, + 0.005647990386933088, + 0.005777301732450724, + 0.004318873398005962, + 0.003492050338536501, + 0.003259579185396433, + 0.004609167110174894, + 0.003012024564668536, + 0.004940987564623356, + 0.003918059170246124, + 0.00423359964042902, + 0.01478491723537445, + 0.015266712754964828, + 0.00848653819411993, + 0.01638338901102543, + 0.012703584507107735, + 0.035038743168115616, + 0.012300794944167137, + 0.019344834610819817, + 0.011943738907575607, + 0.014081656001508236, + 0.013452915474772453, + 0.013951133005321026, + 0.00941446889191866, + 0.014455486088991165, + 0.01139369048178196, + 0.013362817466259003, + 0.013419168069958687, + 0.01200570072978735, + 0.005563334096223116, + 0.006678832694888115, + 0.010244282893836498, + 0.017143934965133667, + 0.007080268580466509, + 0.010732214897871017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.012541357427835464 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from above focuses on an elderly man, his face etched with the marks of time. His calloused, vein-lined hands firmly grip a wooden chisel, its handle rubbed to a sleek sheen. A warm light, streaming from above and sideways, casts deep, three-dimensional shadows on his ridged face and hands, while also illuminating the sharp, metallic gleam of the blade. His eyes, narrowed slightly, his gaze intense, his gaze fixed on the work in hand. With a steady, deliberate rhythm, the tip of the chisel deftly pushes, whittles, and carves across a block of uncarved, light-colored wood. With each precise movement, tiny, curled, thin chips of wood gracefully peel from the sharp blade, falling like a gentle snowflake onto a workbench cluttered with tools and unfinished work. He was wearing a plain dark blue cotton shirt with the cuffs rolled up to his forearms. The whole scene, under warm lighting, with wood and dark blue as the main colors, presented an atmosphere of extreme tranquility, concentration and craftsmanship.", + "seed": 19539, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_9876_seed_19539_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0283702589571476, + "vlm_logprob": -0.0283702589571476, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.0283702589571476, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01930086873471737 + ], + "opposite_logprob_tokens": [ + -5.019300937652588 + ], + "logprob_reward": -0.01930086873471737, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01930086873471737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04110186547040939, + 0.01460325624793768, + 0.0339275486767292, + 0.010615472681820393, + 0.704001247882843, + 0.11400886625051498, + 0.02797364816069603, + 0.01853133924305439, + 0.01945900171995163, + 0.01577773503959179 + ], + "credit_map": [ + 0.0075475964695215225, + 0.0037480604369193316, + 0.0030073896050453186, + 0.002675185212865472, + 0.004789320286363363, + 0.009838665835559368, + 0.0049163224175572395, + 0.004579325672239065, + 0.0021365610882639885, + 0.00175786844920367, + 0.0013468062970787287, + 0.001544749247841537, + 0.002153246896341443, + 0.002094076480716467, + 0.0019593574106693268, + 0.001610591309145093, + 0.006726529914885759, + 0.005601816810667515, + 0.003421412082388997, + 0.003755503101274371, + 0.0037635136395692825, + 0.003918930888175964, + 0.0033875852823257446, + 0.003352261381223798, + 0.0013093749294057488, + 0.0015716681955382228, + 0.0011323342332616448, + 0.001267633051611483, + 0.0018615176668390632, + 0.0012982299085706472, + 0.001144330482929945, + 0.0010303861927241087, + 0.029717376455664635, + 0.031057199463248253, + 0.015143090859055519, + 0.05879489704966545, + 0.02083832398056984, + 0.22796250879764557, + 0.013223220594227314, + 0.30726462602615356, + 0.022900402545928955, + 0.014063658192753792, + 0.007061654701828957, + 0.010470309294760227, + 0.009852237068116665, + 0.014430208131670952, + 0.023772938176989555, + 0.011457469314336777, + 0.002881251508370042, + 0.003427274292334914, + 0.002198311500251293, + 0.003673476167023182, + 0.005490404088050127, + 0.002888743532821536, + 0.0029637934640049934, + 0.0044503966346383095, + 0.0036168403457850218, + 0.0017280966276302934, + 0.002280933316797018, + 0.0023107235319912434, + 0.0026328524108976126, + 0.0019350360380485654, + 0.002234152052551508, + 0.0017927074804902077, + 0.002145979320630431, + 0.0036157581489533186, + 0.0018483670428395271, + 0.002838105196133256, + 0.0030726550612598658, + 0.0020233839750289917, + 0.00208756304346025, + 0.0018271940061822534, + 0.0015703266253694892, + 0.0022362989839166403, + 0.002379314973950386, + 0.00196217093616724, + 0.0020364620722830296, + 0.0021685087122023106, + 0.0018474232638254762, + 0.001577231683768332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.01930086873471737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062101613730192184 + ], + "opposite_logprob_tokens": [ + -2.8121016025543213 + ], + "logprob_reward": -0.062101613730192184, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062101613730192184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021989217028021812, + 0.02464098297059536, + 0.022457629442214966, + 0.009398688562214375, + 0.7826192378997803, + 0.044261060655117035, + 0.009261449798941612, + 0.022185131907463074, + 0.0117477523162961, + 0.05143878981471062 + ], + "credit_map": [ + 0.0025444887578487396, + 0.002941790735349059, + 0.0017238811124116182, + 0.002662804676219821, + 0.0036777604836970568, + 0.0031858112197369337, + 0.0031193809118121862, + 0.002133300295099616, + 0.0031645733397454023, + 0.003024109173566103, + 0.0026309494860470295, + 0.0032165003940463066, + 0.0036668607499450445, + 0.0031241702381521463, + 0.003414938924834132, + 0.00239888159558177, + 0.005486370529979467, + 0.0025665988214313984, + 0.0022284123115241528, + 0.0016305102035403252, + 0.001719375140964985, + 0.004055957309901714, + 0.002724446589127183, + 0.00204596109688282, + 0.0010790448868647218, + 0.001651738421060145, + 0.0008766806568019092, + 0.001217146753333509, + 0.0009184716618619859, + 0.001053755753673613, + 0.00143761420622468, + 0.0011642375029623508, + 0.03262023627758026, + 0.034254997968673706, + 0.017015032470226288, + 0.06422045826911926, + 0.02274300344288349, + 0.25435593724250793, + 0.014673409052193165, + 0.34273624420166016, + 0.003999326843768358, + 0.009814545512199402, + 0.003638732247054577, + 0.006281460635364056, + 0.006494748406112194, + 0.005447749979794025, + 0.004480752628296614, + 0.004103749059140682, + 0.0009672104497440159, + 0.0010300204157829285, + 0.000944923609495163, + 0.0012208367697894573, + 0.0017756337765604258, + 0.0009261143277399242, + 0.0010519875213503838, + 0.0013447243254631758, + 0.002647379646077752, + 0.0026451339945197105, + 0.002085540210828185, + 0.0026556653901934624, + 0.0043976460583508015, + 0.003721424611285329, + 0.0022667632438242435, + 0.0017655803821980953, + 0.0016154737677425146, + 0.0020614766981452703, + 0.0013544756220653653, + 0.001249994384124875, + 0.001180935767479241, + 0.0015077624702826142, + 0.0012785050785169005, + 0.0014991298085078597, + 0.005509552545845509, + 0.004636102356016636, + 0.00562679348513484, + 0.004633498843759298, + 0.006949391216039658, + 0.010420520789921284, + 0.008478979580104351, + 0.005183950066566467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.062101613730192184 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811956197023392 + ], + "opposite_logprob_tokens": [ + -3.2881195545196533 + ], + "logprob_reward": -0.03811956197023392, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811956197023392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030349422246217728, + 0.010984688065946102, + 0.06314490735530853, + 0.018927298486232758, + 0.6850187182426453, + 0.13291989266872406, + 0.024267887696623802, + 0.013767017982900143, + 0.005648417863994837, + 0.014971747063100338 + ], + "credit_map": [ + 0.005743914283812046, + 0.0033997809514403343, + 0.0022879375610500574, + 0.0019870211835950613, + 0.0036418701056391, + 0.007230509538203478, + 0.0031265225261449814, + 0.002931862836703658, + 0.0017327870009467006, + 0.0013480742927640676, + 0.0017082863487303257, + 0.0011393445311114192, + 0.0016443736385554075, + 0.0012518307194113731, + 0.0012888897908851504, + 0.0008711019763723016, + 0.01018303632736206, + 0.011264531873166561, + 0.0057685114443302155, + 0.0073324935510754585, + 0.007218518760055304, + 0.008803199976682663, + 0.006744349841028452, + 0.0058302623219788074, + 0.002088123233988881, + 0.003211305709555745, + 0.0017351253191009164, + 0.002025461057201028, + 0.003229705849662423, + 0.0025584872346371412, + 0.0020077135413885117, + 0.002071377355605364, + 0.028991887345910072, + 0.02994159795343876, + 0.014608056284487247, + 0.05568540841341019, + 0.019073782488703728, + 0.22435566782951355, + 0.012350941076874733, + 0.3000113368034363, + 0.027223652228713036, + 0.015332113020122051, + 0.008368462324142456, + 0.011183257214725018, + 0.011964532546699047, + 0.015814058482646942, + 0.02959561161696911, + 0.013438207097351551, + 0.002654062816873193, + 0.002892332151532173, + 0.002284111687913537, + 0.0028504529036581516, + 0.005757804494351149, + 0.0024985922500491142, + 0.0023413009475916624, + 0.0029892295133322477, + 0.0014865094562992454, + 0.0009799535619094968, + 0.0018346039578318596, + 0.0012416671961545944, + 0.0029742035549134016, + 0.0018906123004853725, + 0.0020601660944521427, + 0.0012993008131161332, + 0.0005832358729094267, + 0.0009474093676544726, + 0.0005831359303556383, + 0.0007695742533542216, + 0.0006111354450695217, + 0.0005449902964755893, + 0.0005980748683214188, + 0.0010108617134392262, + 0.001661423477344215, + 0.0025962498039007187, + 0.0018230131827294827, + 0.0016485663363710046, + 0.0018578668823465705, + 0.001886901562102139, + 0.001644885167479515, + 0.0018528400687500834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.03811956197023392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254394743591547 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.01254394743591547, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254394743591547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05781024321913719, + 0.033698566257953644, + 0.17724496126174927, + 0.036098722368478775, + 0.18501417338848114, + 0.4173510670661926, + 0.02349879778921604, + 0.031447891145944595, + 0.010948307812213898, + 0.026887279003858566 + ], + "credit_map": [ + 0.006639127153903246, + 0.007597136311233044, + 0.006759583484381437, + 0.006242584902793169, + 0.008306256495416164, + 0.007787659764289856, + 0.007249678019434214, + 0.007228219881653786, + 0.0036073019728064537, + 0.005324109457433224, + 0.003112305421382189, + 0.003978450782597065, + 0.005830660462379456, + 0.0030712715815752745, + 0.003654485335573554, + 0.005119979381561279, + 0.03447470813989639, + 0.025087960064411163, + 0.019259799271821976, + 0.015451283194124699, + 0.015008935704827309, + 0.02587517909705639, + 0.01829688623547554, + 0.023790201172232628, + 0.0038250917568802834, + 0.0050995382480323315, + 0.005584824364632368, + 0.003755236743018031, + 0.004342345986515284, + 0.004752076230943203, + 0.0044457390904426575, + 0.004293866455554962, + 0.009593377821147442, + 0.008729753084480762, + 0.009704633615911007, + 0.01681150123476982, + 0.007599246688187122, + 0.05428101122379303, + 0.00551666459068656, + 0.07277797162532806, + 0.0843944326043129, + 0.04809824004769325, + 0.026674464344978333, + 0.03856273368000984, + 0.03476596251130104, + 0.05268197879195213, + 0.08791303634643555, + 0.04426025226712227, + 0.0021756121423095465, + 0.0022628300357609987, + 0.0026101379189640284, + 0.0026232111267745495, + 0.0026930482126772404, + 0.0033348952420055866, + 0.0021400523837655783, + 0.005659009795635939, + 0.0048681460320949554, + 0.0028248997405171394, + 0.006509925704449415, + 0.002926682820543647, + 0.005135444924235344, + 0.0039345696568489075, + 0.002653151284903288, + 0.002595070516690612, + 0.0016088727861642838, + 0.001847000326961279, + 0.0013511364813894033, + 0.0010854736901819706, + 0.0009423123556189239, + 0.0012494546826928854, + 0.0013174869818612933, + 0.0015465705655515194, + 0.0031250142492353916, + 0.0035549337044358253, + 0.004068424925208092, + 0.0029673234093934298, + 0.0039462074637413025, + 0.003656707936897874, + 0.0032431255094707012, + 0.0023255404084920883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.01254394743591547 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978530291467905 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.00978530291467905, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978530291467905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02219715341925621, + 0.01202984806150198, + 0.020871873944997787, + 0.008893844671547413, + 0.6442139148712158, + 0.242807537317276, + 0.019484074786305428, + 0.008690892718732357, + 0.00634004408493638, + 0.014470938593149185 + ], + "credit_map": [ + 0.0033505151513963938, + 0.002708704210817814, + 0.002539888257160783, + 0.0028657291550189257, + 0.001946546253748238, + 0.0025688596069812775, + 0.0026970417238771915, + 0.0035198680125176907, + 0.001990489661693573, + 0.0011471339967101812, + 0.001112678786739707, + 0.0012527527287602425, + 0.0019113634480163455, + 0.0020285204518586397, + 0.001641168026253581, + 0.0009457402629777789, + 0.0052282246761024, + 0.0026094422210007906, + 0.0022181360982358456, + 0.0015212652506306767, + 0.0017386465333402157, + 0.003503296757116914, + 0.0020320850890129805, + 0.00202077510766685, + 0.0012724202824756503, + 0.0013979497598484159, + 0.000933463335968554, + 0.000953778566326946, + 0.0008620998123660684, + 0.0012781420955434442, + 0.0011394971515983343, + 0.0010564926778897643, + 0.02729400061070919, + 0.02893454022705555, + 0.012744966894388199, + 0.052357375621795654, + 0.018060684204101562, + 0.2113150656223297, + 0.011701987124979496, + 0.2818051874637604, + 0.051674604415893555, + 0.024480924010276794, + 0.015288125723600388, + 0.02104927971959114, + 0.020408963784575462, + 0.03079370968043804, + 0.05436138063669205, + 0.02475050278007984, + 0.0025809318758547306, + 0.0026739875320345163, + 0.001828294130973518, + 0.002369488589465618, + 0.0032637855038046837, + 0.0021102926693856716, + 0.0021406535524874926, + 0.0025166398845613003, + 0.0014531766064465046, + 0.0008411291055381298, + 0.0017113552894443274, + 0.0008309729746542871, + 0.0009993379935622215, + 0.0009313513874076307, + 0.001214236719533801, + 0.0007093310705386102, + 0.0007713851518929005, + 0.0011932295747101307, + 0.0006939909653738141, + 0.0007649397011846304, + 0.0007342758472077549, + 0.0007613737834617496, + 0.0005198575672693551, + 0.0009009906789287925, + 0.0012201336212456226, + 0.0020745794754475355, + 0.003241780446842313, + 0.0022492888383567333, + 0.0021110642701387405, + 0.0014091611374169588, + 0.0010993819450959563, + 0.001065546995960176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.00978530291467905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from above focuses on an elderly man, his face etched with the marks of time. His calloused, vein-lined hands firmly grip a wooden chisel, its handle rubbed to a sleek sheen. A warm light, streaming from above and sideways, casts deep, three-dimensional shadows on his ridged face and hands, while also illuminating the sharp, metallic gleam of the blade. His eyes, narrowed slightly, his gaze intense, his gaze fixed on the work in hand. With a steady, deliberate rhythm, the tip of the chisel deftly pushes, whittles, and carves across a block of uncarved, light-colored wood. With each precise movement, tiny, curled, thin chips of wood gracefully peel from the sharp blade, falling like a gentle snowflake onto a workbench cluttered with tools and unfinished work. He was wearing a plain dark blue cotton shirt with the cuffs rolled up to his forearms. The whole scene, under warm lighting, with wood and dark blue as the main colors, presented an atmosphere of extreme tranquility, concentration and craftsmanship.", + "seed": 19540, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_10/index_9876_seed_19540_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03566485233604908, + "vlm_logprob": -0.03566485233604908, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03566485233604908, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056229084730148315 + ], + "opposite_logprob_tokens": [ + -3.0562291145324707 + ], + "logprob_reward": -0.056229084730148315, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.056229084730148315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33800625801086426, + 0.02564859390258789, + 0.03725395351648331, + 0.15095451474189758, + 0.039395373314619064, + 0.06385549157857895, + 0.055765602737665176, + 0.06289025396108627, + 0.20655544102191925, + 0.01967449113726616 + ], + "credit_map": [ + 0.057792872190475464, + 0.04741116240620613, + 0.014527086168527603, + 0.022760596126317978, + 0.06465441733598709, + 0.07247214019298553, + 0.027368320152163506, + 0.031019652262330055, + 0.002955303294584155, + 0.004154514987021685, + 0.0018091417150571942, + 0.0027525904588401318, + 0.0038178865797817707, + 0.003173873294144869, + 0.003726264927536249, + 0.0032590199261903763, + 0.0046332115307450294, + 0.005730289500206709, + 0.0025298125110566616, + 0.00460218358784914, + 0.005482790060341358, + 0.004824550822377205, + 0.004878360312432051, + 0.004572761245071888, + 0.048698727041482925, + 0.006106402724981308, + 0.007763964589685202, + 0.00535917840898037, + 0.013259101659059525, + 0.05480353161692619, + 0.007143668830394745, + 0.00781994592398405, + 0.004365553613752127, + 0.003271227702498436, + 0.0025851407554000616, + 0.00391197856515646, + 0.004622347187250853, + 0.0062813484109938145, + 0.005861904006451368, + 0.008495872840285301, + 0.0084431366994977, + 0.004249461926519871, + 0.0036189714446663857, + 0.005373011343181133, + 0.00793711468577385, + 0.007780126761645079, + 0.009912839159369469, + 0.016540832817554474, + 0.005980575457215309, + 0.006251255515962839, + 0.004443290643393993, + 0.005525356158614159, + 0.010595624335110188, + 0.008346094749867916, + 0.008086428977549076, + 0.006536978762596846, + 0.006995548494160175, + 0.005487511400133371, + 0.004007089417427778, + 0.004083887208253145, + 0.0070113311521708965, + 0.00969203095883131, + 0.008410709910094738, + 0.01720215007662773, + 0.024942591786384583, + 0.01711728237569332, + 0.01212376169860363, + 0.022311266511678696, + 0.023271800950169563, + 0.026720859110355377, + 0.0540219321846962, + 0.026045946404337883, + 0.0020940247923135757, + 0.0016219723038375378, + 0.0022261596750468016, + 0.0014030910097062588, + 0.002564050490036607, + 0.0029253740794956684, + 0.002485985169187188, + 0.0043538338504731655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.056229084730148315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500281602144241 + ], + "opposite_logprob_tokens": [ + -2.9300029277801514 + ], + "logprob_reward": -0.05500281602144241, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500281602144241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06762384623289108, + 0.04375151917338371, + 0.07351676374673843, + 0.23882213234901428, + 0.040860746055841446, + 0.04961919039487839, + 0.04534640535712242, + 0.0730874091386795, + 0.26433536410331726, + 0.10303661972284317 + ], + "credit_map": [ + 0.010259388945996761, + 0.0115303173661232, + 0.0035325137432664633, + 0.006003060843795538, + 0.009714721702039242, + 0.013529752381145954, + 0.006228521000593901, + 0.006825574673712254, + 0.005883066449314356, + 0.0030203412752598524, + 0.0019108429551124573, + 0.003115898696705699, + 0.005546147469431162, + 0.006038105580955744, + 0.008733835071325302, + 0.009503284469246864, + 0.008870656602084637, + 0.007566110696643591, + 0.004275605548173189, + 0.006128913257271051, + 0.014094940386712551, + 0.012810845859348774, + 0.010423521511256695, + 0.009346174076199532, + 0.029331248253583908, + 0.027016188949346542, + 0.022823506966233253, + 0.029824312776327133, + 0.0244875717908144, + 0.03320895880460739, + 0.02691345103085041, + 0.04521690681576729, + 0.0052451458759605885, + 0.003179743653163314, + 0.002873068442568183, + 0.005096585024148226, + 0.0046279230155050755, + 0.004894453566521406, + 0.006525930482894182, + 0.008417895995080471, + 0.006123946513980627, + 0.0033839563839137554, + 0.0021899885032325983, + 0.003995049744844437, + 0.006048919633030891, + 0.006387819070369005, + 0.008120998740196228, + 0.013368519023060799, + 0.004833638202399015, + 0.004418808035552502, + 0.004033663775771856, + 0.0032956902869045734, + 0.005811428651213646, + 0.006689735688269138, + 0.009672574698925018, + 0.006590868346393108, + 0.007136670872569084, + 0.007015380542725325, + 0.004839965142309666, + 0.005768681410700083, + 0.013798091560602188, + 0.011363204568624496, + 0.00919342041015625, + 0.013971998356282711, + 0.03210970759391785, + 0.021808216348290443, + 0.014948717318475246, + 0.029260985553264618, + 0.030814461410045624, + 0.03401217609643936, + 0.06700015068054199, + 0.0343809574842453, + 0.015935666859149933, + 0.009544570930302143, + 0.008384029380977154, + 0.007616253104060888, + 0.012978363782167435, + 0.013941442593932152, + 0.018093977123498917, + 0.01654231548309326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.05500281602144241 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866499826312065 + ], + "opposite_logprob_tokens": [ + -3.0486650466918945 + ], + "logprob_reward": -0.04866499826312065, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866499826312065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06222989037632942, + 0.03470536321401596, + 0.2009705752134323, + 0.19417300820350647, + 0.051143910735845566, + 0.04160497337579727, + 0.056586794555187225, + 0.08230531960725784, + 0.19895702600479126, + 0.07732313126325607 + ], + "credit_map": [ + 0.009647411294281483, + 0.009226104244589806, + 0.0034548169933259487, + 0.004551145248115063, + 0.010478799231350422, + 0.009883983992040157, + 0.007517459336668253, + 0.007470166310667992, + 0.005594620946794748, + 0.003815757343545556, + 0.0031421571038663387, + 0.0041126469150185585, + 0.004514527507126331, + 0.004249270539730787, + 0.004960829857736826, + 0.004315551836043596, + 0.022138243541121483, + 0.022549254819750786, + 0.01687721349298954, + 0.018417084589600563, + 0.034383680671453476, + 0.029804302379488945, + 0.03281670808792114, + 0.023984093219041824, + 0.06184510141611099, + 0.012409345246851444, + 0.012981566600501537, + 0.009879217483103275, + 0.01561826653778553, + 0.05452454835176468, + 0.01008735504001379, + 0.016827614977955818, + 0.0059799025766551495, + 0.005844259634613991, + 0.00443625170737505, + 0.006782425567507744, + 0.005572309717535973, + 0.006067823618650436, + 0.00816759467124939, + 0.008293344639241695, + 0.005867588799446821, + 0.0042505753226578236, + 0.003397544613108039, + 0.0047797271981835365, + 0.005829871166497469, + 0.0046496023423969746, + 0.005261418875306845, + 0.0075686415657401085, + 0.007572715170681477, + 0.00458584725856781, + 0.003666029078885913, + 0.005546318832784891, + 0.007936066947877407, + 0.007988316006958485, + 0.009208221919834614, + 0.010083276778459549, + 0.007941232062876225, + 0.009275654330849648, + 0.010653144679963589, + 0.012133323587477207, + 0.009422706440091133, + 0.010037291795015335, + 0.010231577791273594, + 0.012610387057065964, + 0.023736778646707535, + 0.015794407576322556, + 0.012080131098628044, + 0.021612489596009254, + 0.021751122549176216, + 0.026141522452235222, + 0.0535571463406086, + 0.02428343892097473, + 0.010588943026959896, + 0.007071302272379398, + 0.006111355032771826, + 0.0077864849008619785, + 0.012163624167442322, + 0.009767090901732445, + 0.0126081183552742, + 0.011226210743188858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.04866499826312065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642885833978653 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642885833978653, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642885833978653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09719846397638321, + 0.1073971688747406, + 0.0859469473361969, + 0.15758782625198364, + 0.13982588052749634, + 0.04833870381116867, + 0.07548995316028595, + 0.059952810406684875, + 0.050074391067028046, + 0.17818789184093475 + ], + "credit_map": [ + 0.013563711196184158, + 0.011513453908264637, + 0.008310744538903236, + 0.00802167970687151, + 0.016927678138017654, + 0.014085710048675537, + 0.013541948050260544, + 0.01123354397714138, + 0.013693940825760365, + 0.00973551720380783, + 0.006751404143869877, + 0.012680897489190102, + 0.01429651491343975, + 0.013453826308250427, + 0.014798532240092754, + 0.02198653854429722, + 0.012485368177294731, + 0.0065518878400325775, + 0.005678017158061266, + 0.013464356772601604, + 0.0113019784912467, + 0.011125736869871616, + 0.01379699818789959, + 0.011542607098817825, + 0.03272128105163574, + 0.012514499947428703, + 0.025434337556362152, + 0.014756348915398121, + 0.013589724898338318, + 0.030753321945667267, + 0.013683929108083248, + 0.014134380966424942, + 0.01676643081009388, + 0.012999461963772774, + 0.016064541414380074, + 0.01470861118286848, + 0.024206973612308502, + 0.01836656965315342, + 0.015529905445873737, + 0.021183397620916367, + 0.00640697218477726, + 0.005397021304816008, + 0.004969491623342037, + 0.006057175807654858, + 0.006751969922333956, + 0.005894487723708153, + 0.005981077440083027, + 0.006880510598421097, + 0.00788903422653675, + 0.008266723714768887, + 0.00416561821475625, + 0.007567159831523895, + 0.013522611930966377, + 0.010893816128373146, + 0.010845373384654522, + 0.012339618057012558, + 0.005865343846380711, + 0.005296482238918543, + 0.006755224894732237, + 0.009358142502605915, + 0.00851408764719963, + 0.008091380819678307, + 0.006245949771255255, + 0.009826199151575565, + 0.007027484476566315, + 0.004278447479009628, + 0.0053166854195296764, + 0.006223568692803383, + 0.005287672858685255, + 0.005494088865816593, + 0.010737559758126736, + 0.005708884447813034, + 0.02288505621254444, + 0.015977412462234497, + 0.008838302455842495, + 0.018283627927303314, + 0.027223117649555206, + 0.025962812826037407, + 0.027673693373799324, + 0.03134387359023094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.008642885833978653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from above focusing on an elderly man's face and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a plain dark blue cotton shirt with cuffs rolled up to his forearms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man using a wooden chisel with a sleek sheen to carve a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, curled, thin chips of wood seen peeling from the blade and falling onto a cluttered workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm light streaming from above and sideways, casting deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784476831555367 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784476831555367, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784476831555367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24862895905971527, + 0.062483616173267365, + 0.08544213324785233, + 0.24192874133586884, + 0.07890861481428146, + 0.028766335919499397, + 0.07377730309963226, + 0.03955772891640663, + 0.052794210612773895, + 0.08771241456270218 + ], + "credit_map": [ + 0.04101581871509552, + 0.024960573762655258, + 0.01599215343594551, + 0.020901240408420563, + 0.050283025950193405, + 0.04646177217364311, + 0.026413436979055405, + 0.0226009339094162, + 0.00863112136721611, + 0.005560464225709438, + 0.004278101027011871, + 0.0073265316896140575, + 0.010795235633850098, + 0.007336134556680918, + 0.009446750394999981, + 0.009109276346862316, + 0.010840782895684242, + 0.00824654009193182, + 0.010207128711044788, + 0.009316307492554188, + 0.012294139713048935, + 0.011182313784956932, + 0.012600837275385857, + 0.010754086077213287, + 0.08166508376598358, + 0.013568812981247902, + 0.012311041355133057, + 0.01017568539828062, + 0.020218729972839355, + 0.07884328067302704, + 0.01010460127145052, + 0.015041514299809933, + 0.01096982229501009, + 0.007103394716978073, + 0.007378755137324333, + 0.008663356304168701, + 0.009043981321156025, + 0.009171629324555397, + 0.011381986550986767, + 0.015195691958069801, + 0.0035575528163462877, + 0.0021952802781015635, + 0.0020561201963573694, + 0.0019050028640776873, + 0.0036030129995197058, + 0.0035949265584349632, + 0.0038371128030121326, + 0.008017326705157757, + 0.009977474808692932, + 0.008896424435079098, + 0.00726543553173542, + 0.006291472818702459, + 0.008485373109579086, + 0.009607074782252312, + 0.012344579212367535, + 0.010909468866884708, + 0.003726036986336112, + 0.003833113005384803, + 0.0028953079599887133, + 0.004099634941667318, + 0.007929042913019657, + 0.005737472325563431, + 0.005691349972039461, + 0.005645771510899067, + 0.005780115257948637, + 0.004137320443987846, + 0.003728251438587904, + 0.006367025896906853, + 0.005979957990348339, + 0.006691645365208387, + 0.013682438060641289, + 0.006427455693483353, + 0.012044396251440048, + 0.009260917082428932, + 0.00618309061974287, + 0.008765739388763905, + 0.01233893632888794, + 0.010392480529844761, + 0.014771186746656895, + 0.01395566388964653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.009784476831555367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_11/index_tensor([8900, 8900])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_11/index_tensor([8900, 8900])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4b3733d144f3f952112f11b78bf29e2040e3df6d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_11/index_tensor([8900, 8900])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot of an interior, shot from a slightly upward angle, slowly follows the movements of a young goat. The goat's horns are small, pure white, and its body is covered in fluffy, soft, pale golden fur, gleaming softly in the warm glow of a flickering oil lamp. A slightly worn blue fabric collar hangs around its neck. Carefully adjusting a patched, rough-hewn backpack, it stops at the doorway of a simple wooden hut. Slowly turning, its large, clear, yet sorrowful eyes cast a final, fleeting glance at its mother, who lies on a straw bed in the corner, breathing weakly. This glance is filled with reluctance and determination. Then, resolutely, it turns back and, with its front hoof, gently and almost silently pushes the heavy wooden door shut. With a dull click of the latch, the warm orange glow of the room is completely blocked out, leaving only a glimpse of the vast grassland, bathed in cold moonlight, through the crack of the door. The entire scene is imbued with a tranquil, sad, yet courageous atmosphere of departure.", + "seed": 7678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_11/index_8900_seed_7678_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028064068406820297, + "vlm_logprob": -0.028064068406820297, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.028064068406820297, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021412856876850128 + ], + "opposite_logprob_tokens": [ + -4.0214128494262695 + ], + "logprob_reward": -0.021412856876850128, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021412856876850128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1403055489063263, + 0.09397601336240768, + 0.07752002775669098, + 0.06986958533525467, + 0.0660247802734375, + 0.3076864182949066, + 0.024144437164068222, + 0.049895014613866806, + 0.039902590215206146, + 0.1306755393743515 + ], + "credit_map": [ + 0.010090626776218414, + 0.010210148058831692, + 0.017020482569932938, + 0.017820119857788086, + 0.010659020394086838, + 0.028260093182325363, + 0.027921143919229507, + 0.01832392066717148, + 0.007904302328824997, + 0.01913263276219368, + 0.013082273304462433, + 0.013126887381076813, + 0.010594574734568596, + 0.0070059155113995075, + 0.013534421101212502, + 0.009595002979040146, + 0.008481861092150211, + 0.010044769383966923, + 0.008873875252902508, + 0.012595921754837036, + 0.010124261491000652, + 0.00800386629998684, + 0.012680654413998127, + 0.006714819464832544, + 0.004547273740172386, + 0.006235851906239986, + 0.006369064096361399, + 0.004872455727308989, + 0.006262731272727251, + 0.008483693934977055, + 0.021385397762060165, + 0.0117131182923913, + 0.005027144681662321, + 0.005606566555798054, + 0.010886711068451405, + 0.005552735645323992, + 0.017245231196284294, + 0.0072443317621946335, + 0.005895518232136965, + 0.008566543459892273, + 0.01058950275182724, + 0.13127903640270233, + 0.013958272524178028, + 0.09110326319932938, + 0.018938397988677025, + 0.007677013985812664, + 0.009923243895173073, + 0.024217704311013222, + 0.0019540851935744286, + 0.0036723909433931112, + 0.002195480279624462, + 0.003804287640377879, + 0.003418437670916319, + 0.003295786213129759, + 0.0029072195757180452, + 0.00289675104431808, + 0.006250402424484491, + 0.0069711171090602875, + 0.005148685071617365, + 0.006135811097919941, + 0.004256115760654211, + 0.006391469854861498, + 0.007361761759966612, + 0.00737965339794755, + 0.004695101175457239, + 0.00409220065921545, + 0.004590947646647692, + 0.0055945860221982, + 0.004772467538714409, + 0.004457708913832903, + 0.005812760442495346, + 0.0058868201449513435, + 0.025998516008257866, + 0.016234487295150757, + 0.034171994775533676, + 0.020278962329030037, + 0.006439783610403538, + 0.007409211248159409, + 0.008170925080776215, + 0.011971673928201199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.021412856876850128 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496896430850029 + ], + "opposite_logprob_tokens": [ + -2.929969072341919 + ], + "logprob_reward": -0.05496896430850029, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496896430850029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059804800897836685, + 0.06375260651111603, + 0.13318583369255066, + 0.04277217388153076, + 0.03443872928619385, + 0.5440005660057068, + 0.05455778166651726, + 0.016371585428714752, + 0.02615530416369438, + 0.02496069297194481 + ], + "credit_map": [ + 0.004874866455793381, + 0.004965535830706358, + 0.006257241126149893, + 0.009263233281672001, + 0.004455245099961758, + 0.006879535038024187, + 0.01345957163721323, + 0.009649564512073994, + 0.0052834684029221535, + 0.009668049402534962, + 0.010354251600801945, + 0.00807963963598013, + 0.007026765029877424, + 0.007067774888128042, + 0.010427949950098991, + 0.005844703409820795, + 0.015550745651125908, + 0.016881339251995087, + 0.01715235970914364, + 0.022738931700587273, + 0.012912814505398273, + 0.011963741853833199, + 0.02112811990082264, + 0.014857765287160873, + 0.0020432723686099052, + 0.0033374917693436146, + 0.002704768907278776, + 0.0029175749514251947, + 0.003405721392482519, + 0.0040451036766171455, + 0.015158063732087612, + 0.009160173125565052, + 0.005474898964166641, + 0.0031430269591510296, + 0.004385512787848711, + 0.0023995088413357735, + 0.006761134602129459, + 0.0031129231210798025, + 0.002750472165644169, + 0.006411247421056032, + 0.02036341093480587, + 0.23303882777690887, + 0.02044428512454033, + 0.14945392310619354, + 0.03637059032917023, + 0.014700544998049736, + 0.01750176213681698, + 0.05212720111012459, + 0.005002241116017103, + 0.0062333280220627785, + 0.006014716345816851, + 0.006352371070533991, + 0.0065544117242097855, + 0.01127180177718401, + 0.00749505590647459, + 0.005633851513266563, + 0.0017529056640341878, + 0.002447570441290736, + 0.0014820260694250464, + 0.002156549599021673, + 0.0016206169966608286, + 0.0035723962355405092, + 0.0014557641698047519, + 0.0018837543902918696, + 0.003059098497033119, + 0.0026557310484349728, + 0.0025502790231257677, + 0.0028526068199425936, + 0.004481131676584482, + 0.004489905666559935, + 0.0033012726344168186, + 0.0027652764692902565, + 0.0042105261236429214, + 0.003981520421802998, + 0.004656569566577673, + 0.004011232405900955, + 0.0018731941236183047, + 0.00130265683401376, + 0.003098561428487301, + 0.001826427411288023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.05496896430850029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642790675163269 + ], + "opposite_logprob_tokens": [ + -4.88264274597168 + ], + "logprob_reward": -0.007642790675163269, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642790675163269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11560468375682831, + 0.1565742939710617, + 0.04503437504172325, + 0.0475633330643177, + 0.05893167108297348, + 0.2631056010723114, + 0.11019713431596756, + 0.0703992024064064, + 0.0771830603480339, + 0.05540662631392479 + ], + "credit_map": [ + 0.010132208466529846, + 0.00945619959384203, + 0.010256661102175713, + 0.016762128099799156, + 0.009190739132463932, + 0.02723882719874382, + 0.013567239977419376, + 0.01900068111717701, + 0.01290882658213377, + 0.03541111946105957, + 0.016893697902560234, + 0.02450038120150566, + 0.017257820814847946, + 0.012023287825286388, + 0.021407069638371468, + 0.01617208495736122, + 0.004718567710369825, + 0.007451026700437069, + 0.005525024142116308, + 0.006672045215964317, + 0.004717070609331131, + 0.0052658445201814175, + 0.006189428269863129, + 0.0044953664764761925, + 0.0031915996223688126, + 0.004583768546581268, + 0.006509356200695038, + 0.003818494500592351, + 0.004207571968436241, + 0.004636144731193781, + 0.01263925340026617, + 0.007977145723998547, + 0.00574473338201642, + 0.005356878973543644, + 0.011551680974662304, + 0.004230252001434565, + 0.011450828053057194, + 0.007936078123748302, + 0.005222355015575886, + 0.007438863627612591, + 0.009371859952807426, + 0.10751024633646011, + 0.015490682795643806, + 0.07918825000524521, + 0.01592799462378025, + 0.00954765547066927, + 0.010462544858455658, + 0.01560633722692728, + 0.010503947734832764, + 0.01302465982735157, + 0.013805767521262169, + 0.013151759281754494, + 0.011911812238395214, + 0.02083665318787098, + 0.014643781818449497, + 0.01231874618679285, + 0.006781404837965965, + 0.00880611315369606, + 0.00807427428662777, + 0.009413710795342922, + 0.005959772039204836, + 0.013317030854523182, + 0.008376462385058403, + 0.009670436382293701, + 0.009725410491228104, + 0.009577091783285141, + 0.013049590401351452, + 0.011964503675699234, + 0.010697675868868828, + 0.00797855295240879, + 0.007412165403366089, + 0.006778068840503693, + 0.008372181095182896, + 0.009265352040529251, + 0.006232580170035362, + 0.0071324571035802364, + 0.004889382980763912, + 0.005418373271822929, + 0.007952234707772732, + 0.006144065875560045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.007642790675163269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038102924823760986 + ], + "opposite_logprob_tokens": [ + -3.288102865219116 + ], + "logprob_reward": -0.038102924823760986, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038102924823760986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14412134885787964, + 0.07954975962638855, + 0.09936443716287613, + 0.08668028563261032, + 0.02299072965979576, + 0.12596829235553741, + 0.07836977392435074, + 0.1078791543841362, + 0.09562741219997406, + 0.15944887697696686 + ], + "credit_map": [ + 0.011164501309394836, + 0.01368020474910736, + 0.013100125826895237, + 0.024620365351438522, + 0.01094431709498167, + 0.020857611671090126, + 0.026795605197548866, + 0.022958621382713318, + 0.006305961403995752, + 0.014863688498735428, + 0.010439164005219936, + 0.009692578576505184, + 0.008855839259922504, + 0.009999148547649384, + 0.011209644377231598, + 0.008183731697499752, + 0.010422137565910816, + 0.011046997271478176, + 0.010430441237986088, + 0.016794053837656975, + 0.011029220186173916, + 0.013746715150773525, + 0.018010515719652176, + 0.007884352467954159, + 0.005906347185373306, + 0.010700492188334465, + 0.004468497354537249, + 0.005964348558336496, + 0.005763819441199303, + 0.008244304917752743, + 0.02909061498939991, + 0.0165418591350317, + 0.0036547535564750433, + 0.0017116940580308437, + 0.004171339794993401, + 0.0014983417931944132, + 0.0031655547209084034, + 0.0016487514367327094, + 0.002757162554189563, + 0.004383130930364132, + 0.010056019760668278, + 0.014637013897299767, + 0.007598397321999073, + 0.04149438440799713, + 0.012686893343925476, + 0.009475375525653362, + 0.011524157598614693, + 0.01849604956805706, + 0.007621682249009609, + 0.009161964058876038, + 0.00923747755587101, + 0.00894724577665329, + 0.010791350156068802, + 0.015534023754298687, + 0.007805010303854942, + 0.009271019138395786, + 0.01057460904121399, + 0.014157046563923359, + 0.01488433312624693, + 0.014887796714901924, + 0.008868624456226826, + 0.016556214541196823, + 0.01151626743376255, + 0.016434265300631523, + 0.011048564687371254, + 0.010559074580669403, + 0.013225550763309002, + 0.01225708145648241, + 0.01664935238659382, + 0.009884112514555454, + 0.010276859626173973, + 0.01172681339085102, + 0.026890646666288376, + 0.02523740567266941, + 0.023865817114710808, + 0.01978238858282566, + 0.01547968853265047, + 0.013477236032485962, + 0.0163617841899395, + 0.018353918567299843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.038102924823760986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192805349826813 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.018192805349826813, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192805349826813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21342839300632477, + 0.06134892255067825, + 0.1546628773212433, + 0.035096149891614914, + 0.05106087028980255, + 0.12047175318002701, + 0.08134432137012482, + 0.07709334045648575, + 0.08106408268213272, + 0.12442931532859802 + ], + "credit_map": [ + 0.018296685069799423, + 0.016700442880392075, + 0.014194661751389503, + 0.028538787737488747, + 0.013113515451550484, + 0.05532397702336311, + 0.028870845213532448, + 0.03838946670293808, + 0.004725704900920391, + 0.009323147125542164, + 0.010611899197101593, + 0.006846146658062935, + 0.006672155577689409, + 0.006176684983074665, + 0.01057036779820919, + 0.006422818638384342, + 0.016691338270902634, + 0.02306734398007393, + 0.01956779882311821, + 0.029620841145515442, + 0.015602480620145798, + 0.009842411614954472, + 0.026923255994915962, + 0.013347390107810497, + 0.0032950509339571, + 0.0063876560889184475, + 0.0037218413781374693, + 0.004006211180239916, + 0.004274460487067699, + 0.004332449287176132, + 0.005163464229553938, + 0.003915018867701292, + 0.009141676127910614, + 0.0037253606133162975, + 0.0062094805762171745, + 0.0031979926861822605, + 0.007223303895443678, + 0.008310137316584587, + 0.004207993857562542, + 0.00904492475092411, + 0.015544029884040356, + 0.023864131420850754, + 0.004752839915454388, + 0.03957534581422806, + 0.007142442278563976, + 0.01110000442713499, + 0.007028508000075817, + 0.01146444771438837, + 0.009054943919181824, + 0.008671180345118046, + 0.011129231192171574, + 0.010388515889644623, + 0.011772209778428078, + 0.0126749062910676, + 0.009717830456793308, + 0.007935507223010063, + 0.008067150600254536, + 0.009703140705823898, + 0.007568527013063431, + 0.007479027379304171, + 0.007593188900500536, + 0.015393845736980438, + 0.01033373735845089, + 0.010954715311527252, + 0.006911912001669407, + 0.006700671277940273, + 0.016740508377552032, + 0.009538745507597923, + 0.009416867047548294, + 0.016198989003896713, + 0.008368771523237228, + 0.007187612820416689, + 0.024992959573864937, + 0.016376519575715065, + 0.03251387178897858, + 0.014246993698179722, + 0.007079616654664278, + 0.007652274798601866, + 0.010634987615048885, + 0.010932083241641521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.018192805349826813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot of an interior, shot from a slightly upward angle, slowly follows the movements of a young goat. The goat's horns are small, pure white, and its body is covered in fluffy, soft, pale golden fur, gleaming softly in the warm glow of a flickering oil lamp. A slightly worn blue fabric collar hangs around its neck. Carefully adjusting a patched, rough-hewn backpack, it stops at the doorway of a simple wooden hut. Slowly turning, its large, clear, yet sorrowful eyes cast a final, fleeting glance at its mother, who lies on a straw bed in the corner, breathing weakly. This glance is filled with reluctance and determination. Then, resolutely, it turns back and, with its front hoof, gently and almost silently pushes the heavy wooden door shut. With a dull click of the latch, the warm orange glow of the room is completely blocked out, leaving only a glimpse of the vast grassland, bathed in cold moonlight, through the crack of the door. The entire scene is imbued with a tranquil, sad, yet courageous atmosphere of departure.", + "seed": 7679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_11/index_8900_seed_7679_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0296985175460577, + "vlm_logprob": -0.0296985175460577, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.0296985175460577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03301103413105011 + ], + "opposite_logprob_tokens": [ + -3.533010959625244 + ], + "logprob_reward": -0.03301103413105011, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03301103413105011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16344524919986725, + 0.08585536479949951, + 0.14295801520347595, + 0.07118497788906097, + 0.14316044747829437, + 0.11179182678461075, + 0.08729421347379684, + 0.06645628809928894, + 0.07902825623750687, + 0.04882530868053436 + ], + "credit_map": [ + 0.015573645941913128, + 0.02992992289364338, + 0.02989422343671322, + 0.013623150996863842, + 0.023436537012457848, + 0.016919590532779694, + 0.02241748385131359, + 0.01165070291608572, + 0.007987020537257195, + 0.010207067243754864, + 0.008961768820881844, + 0.012027903459966183, + 0.015350110828876495, + 0.011526204645633698, + 0.01055179163813591, + 0.009243492968380451, + 0.0163189098238945, + 0.02221144363284111, + 0.018422624096274376, + 0.01800253428518772, + 0.012499954551458359, + 0.014123179018497467, + 0.016617584973573685, + 0.024761773645877838, + 0.011540597304701805, + 0.005812200251966715, + 0.011072040535509586, + 0.012890836223959923, + 0.0068880836479365826, + 0.005347559228539467, + 0.01260196603834629, + 0.00503169558942318, + 0.024611713364720345, + 0.015375316143035889, + 0.019422197714447975, + 0.012940057553350925, + 0.01894945278763771, + 0.018504450097680092, + 0.013078862801194191, + 0.02027839981019497, + 0.015920398756861687, + 0.01196734607219696, + 0.013533255085349083, + 0.009220751002430916, + 0.013090790249407291, + 0.01846170239150524, + 0.01102153304964304, + 0.01857605390250683, + 0.0134110227227211, + 0.007751910947263241, + 0.020331770181655884, + 0.01114004384726286, + 0.005306646227836609, + 0.00677789282053709, + 0.0065784459002316, + 0.015996485948562622, + 0.007598220370709896, + 0.012452804483473301, + 0.007831241935491562, + 0.010337340645492077, + 0.005806010216474533, + 0.005731217097491026, + 0.007688288111239672, + 0.009011160582304, + 0.009110620245337486, + 0.011261302046477795, + 0.014242589473724365, + 0.009392007254064083, + 0.006063051521778107, + 0.008156788535416126, + 0.00827717687934637, + 0.012524721212685108, + 0.005722541827708483, + 0.009526338428258896, + 0.005174513440579176, + 0.005681575275957584, + 0.007273486349731684, + 0.003982724621891975, + 0.005749483127146959, + 0.005714646074920893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.03301103413105011 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370732069015503 + ], + "opposite_logprob_tokens": [ + -3.4087073802948 + ], + "logprob_reward": -0.03370732069015503, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370732069015503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.140929713845253, + 0.17959141731262207, + 0.13672848045825958, + 0.1263803094625473, + 0.12640753388404846, + 0.08732055127620697, + 0.07002336531877518, + 0.048600539565086365, + 0.032505933195352554, + 0.05151219293475151 + ], + "credit_map": [ + 0.012964610010385513, + 0.023890957236289978, + 0.03350779041647911, + 0.012566906400024891, + 0.01775515079498291, + 0.012869909405708313, + 0.017537744715809822, + 0.009836635552346706, + 0.015308531001210213, + 0.024247724562883377, + 0.022394919767975807, + 0.028077295050024986, + 0.025060515850782394, + 0.020325517281889915, + 0.02166610211133957, + 0.022510815411806107, + 0.015571768395602703, + 0.025446169078350067, + 0.016055414453148842, + 0.012976001016795635, + 0.016695883125066757, + 0.01428253948688507, + 0.015432147309184074, + 0.020268550142645836, + 0.018877902999520302, + 0.013062112033367157, + 0.016763269901275635, + 0.02136453427374363, + 0.011026673018932343, + 0.009375189431011677, + 0.025004614144563675, + 0.010906009003520012, + 0.01647581160068512, + 0.012442028149962425, + 0.01840929500758648, + 0.013936394825577736, + 0.013347527012228966, + 0.010149897076189518, + 0.026398256421089172, + 0.015248322859406471, + 0.013331644237041473, + 0.00867165345698595, + 0.011147976852953434, + 0.0094838198274374, + 0.00868514459580183, + 0.015659265220165253, + 0.00890294462442398, + 0.011438095942139626, + 0.005821696016937494, + 0.007399346213787794, + 0.011032667942345142, + 0.008075756020843983, + 0.005618071649223566, + 0.007026143837720156, + 0.013621893711388111, + 0.01142778992652893, + 0.005529223475605249, + 0.008240788243710995, + 0.007563355844467878, + 0.006418830249458551, + 0.004476405214518309, + 0.0034477210137993097, + 0.006160247139632702, + 0.006763963494449854, + 0.0027671251446008682, + 0.005509246606379747, + 0.004323262255638838, + 0.0035220917779952288, + 0.0034635665360838175, + 0.0028207851573824883, + 0.003952954895794392, + 0.006146902684122324, + 0.006830442231148481, + 0.007091316860169172, + 0.007850231602787971, + 0.0050214738585054874, + 0.004636536817997694, + 0.007857149466872215, + 0.005516930017620325, + 0.0067081148736178875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.03370732069015503 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653759025037289 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.008653759025037289, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653759025037289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380809247493744, + 0.0848788470029831, + 0.2871447503566742, + 0.13700230419635773, + 0.06677477061748505, + 0.06150095537304878, + 0.05862937495112419, + 0.10061420500278473, + 0.034942835569381714, + 0.030431030318140984 + ], + "credit_map": [ + 0.013323551043868065, + 0.04036065191030502, + 0.03539097681641579, + 0.009426688775420189, + 0.009426395408809185, + 0.007374074310064316, + 0.014805088751018047, + 0.007973491214215755, + 0.008309204131364822, + 0.01195038016885519, + 0.011543269269168377, + 0.012599593959748745, + 0.010573373176157475, + 0.011577282100915909, + 0.009528722614049911, + 0.008797035552561283, + 0.028942029923200607, + 0.07017006725072861, + 0.034545253962278366, + 0.035619452595710754, + 0.024359306320548058, + 0.022449808195233345, + 0.03638658672571182, + 0.03467225283384323, + 0.014342881739139557, + 0.01234260480850935, + 0.019785111770033836, + 0.021930694580078125, + 0.015307988971471786, + 0.011802658438682556, + 0.029307696968317032, + 0.012182674370706081, + 0.010726984590291977, + 0.005826505832374096, + 0.008457585237920284, + 0.007750241551548243, + 0.0074799456633627415, + 0.007396724075078964, + 0.010409856215119362, + 0.008726936765015125, + 0.010267160832881927, + 0.0076179527677595615, + 0.007579733617603779, + 0.005061760079115629, + 0.005387701094150543, + 0.012001126073300838, + 0.005433298647403717, + 0.008152224123477936, + 0.007843113504350185, + 0.007570712361484766, + 0.009904810227453709, + 0.0064407167956233025, + 0.004975431598722935, + 0.004935881122946739, + 0.007718787062913179, + 0.0092399250715971, + 0.0071968506090343, + 0.017497416585683823, + 0.008594808168709278, + 0.011960006318986416, + 0.016263708472251892, + 0.0077944472432136536, + 0.018479416146874428, + 0.012827553786337376, + 0.004089663736522198, + 0.006884369067847729, + 0.0037050137761980295, + 0.0036498105619102716, + 0.0034276326186954975, + 0.004933782387524843, + 0.004038952756673098, + 0.00421361206099391, + 0.0027175836730748415, + 0.00637411791831255, + 0.003641993971541524, + 0.0045250896364450455, + 0.0033306493423879147, + 0.0024989915546029806, + 0.003433879232034087, + 0.003908725921064615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.008653759025037289 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204526126384735 + ], + "opposite_logprob_tokens": [ + -2.8120453357696533 + ], + "logprob_reward": -0.06204526126384735, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204526126384735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14538788795471191, + 0.10244506597518921, + 0.0599316842854023, + 0.060470353811979294, + 0.10303782671689987, + 0.0992274060845375, + 0.1442132294178009, + 0.15982593595981598, + 0.05522603914141655, + 0.07023455202579498 + ], + "credit_map": [ + 0.012860693968832493, + 0.04149835556745529, + 0.04024556279182434, + 0.011340051889419556, + 0.011421448551118374, + 0.007673289626836777, + 0.01414123922586441, + 0.0062072621658444405, + 0.013663265854120255, + 0.00874366331845522, + 0.01084181945770979, + 0.01476273499429226, + 0.01767157018184662, + 0.011893208138644695, + 0.012588626705110073, + 0.012280182912945747, + 0.00653320224955678, + 0.01019507460296154, + 0.00757258664816618, + 0.007473894860595465, + 0.006467275787144899, + 0.006300804205238819, + 0.008407714776694775, + 0.0069811297580599785, + 0.006210405379533768, + 0.005038249772042036, + 0.008820099756121635, + 0.006786276586353779, + 0.006863960064947605, + 0.011260868050158024, + 0.00998705718666315, + 0.005503436084836721, + 0.019687211140990257, + 0.009070484898984432, + 0.0164682324975729, + 0.009436853229999542, + 0.01076769270002842, + 0.015884827822446823, + 0.008832190185785294, + 0.01289033330976963, + 0.011528939940035343, + 0.012280777096748352, + 0.011871528811752796, + 0.00903190579265356, + 0.015837091952562332, + 0.015152791514992714, + 0.012010844424366951, + 0.01151352934539318, + 0.022367842495441437, + 0.01860099844634533, + 0.025457000359892845, + 0.012963677756488323, + 0.013677297160029411, + 0.012278332374989986, + 0.01804909110069275, + 0.02081897482275963, + 0.02058323472738266, + 0.029025958850979805, + 0.018318137153983116, + 0.021652506664395332, + 0.014972892589867115, + 0.013513641431927681, + 0.023233620449900627, + 0.018525931984186172, + 0.005925686564296484, + 0.00865781120955944, + 0.006660713814198971, + 0.006990425288677216, + 0.006570748519152403, + 0.005586009006947279, + 0.006995925679802895, + 0.007838715799152851, + 0.005412271246314049, + 0.014142134226858616, + 0.005774538964033127, + 0.010722214356064796, + 0.008954426273703575, + 0.005598932970315218, + 0.008301540277898312, + 0.011328496038913727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.06204526126384735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075212620198727 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075212620198727, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075212620198727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11625608056783676, + 0.08418598026037216, + 0.09460287541151047, + 0.07617674022912979, + 0.07755843549966812, + 0.1852424144744873, + 0.08945722132921219, + 0.14000089466571808, + 0.06391359865665436, + 0.07260576635599136 + ], + "credit_map": [ + 0.012512397952377796, + 0.018354011699557304, + 0.013165404088795185, + 0.011398511938750744, + 0.01341671496629715, + 0.017553457990288734, + 0.016238847747445107, + 0.013616736978292465, + 0.009990015998482704, + 0.020254183560609818, + 0.008335745893418789, + 0.006870172452181578, + 0.011393516324460506, + 0.006751032080501318, + 0.00834926962852478, + 0.012242041528224945, + 0.0112765496596694, + 0.02324303239583969, + 0.01162125263363123, + 0.010326031595468521, + 0.0070257242769002914, + 0.006897230166941881, + 0.013102577067911625, + 0.01111048087477684, + 0.012034795247018337, + 0.007449723314493895, + 0.012518062256276608, + 0.014941612258553505, + 0.0066423360258340836, + 0.004765118006616831, + 0.012509370222687721, + 0.005315722897648811, + 0.012355501763522625, + 0.007683124393224716, + 0.012057577259838581, + 0.006799379363656044, + 0.00956053752452135, + 0.009853635914623737, + 0.010556603781878948, + 0.008692080155014992, + 0.024652380496263504, + 0.02120797336101532, + 0.02189522422850132, + 0.016376245766878128, + 0.02051793597638607, + 0.033974021673202515, + 0.021087173372507095, + 0.02553146332502365, + 0.010928034782409668, + 0.008704095147550106, + 0.012382623739540577, + 0.009744688868522644, + 0.008095178753137589, + 0.008650383912026882, + 0.01553014200180769, + 0.015422074124217033, + 0.014353758655488491, + 0.024101126939058304, + 0.016139082610607147, + 0.020462755113840103, + 0.012065969407558441, + 0.011897644028067589, + 0.020234351977705956, + 0.020746199414134026, + 0.008607675321400166, + 0.011378278955817223, + 0.006807347759604454, + 0.007020444143563509, + 0.0054866657592356205, + 0.007413089741021395, + 0.006987263914197683, + 0.010212833993136883, + 0.008418714627623558, + 0.012296578846871853, + 0.007822899147868156, + 0.012339684180915356, + 0.007907667197287083, + 0.005916779860854149, + 0.007982049137353897, + 0.00992138497531414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.011075212620198727 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot of an interior, shot from a slightly upward angle, slowly follows the movements of a young goat. The goat's horns are small, pure white, and its body is covered in fluffy, soft, pale golden fur, gleaming softly in the warm glow of a flickering oil lamp. A slightly worn blue fabric collar hangs around its neck. Carefully adjusting a patched, rough-hewn backpack, it stops at the doorway of a simple wooden hut. Slowly turning, its large, clear, yet sorrowful eyes cast a final, fleeting glance at its mother, who lies on a straw bed in the corner, breathing weakly. This glance is filled with reluctance and determination. Then, resolutely, it turns back and, with its front hoof, gently and almost silently pushes the heavy wooden door shut. With a dull click of the latch, the warm orange glow of the room is completely blocked out, leaving only a glimpse of the vast grassland, bathed in cold moonlight, through the crack of the door. The entire scene is imbued with a tranquil, sad, yet courageous atmosphere of departure.", + "seed": 7680, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_11/index_8900_seed_7680_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05686563197523355, + "vlm_logprob": -0.05686563197523355, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.05686563197523355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04230903834104538 + ], + "opposite_logprob_tokens": [ + -3.417309045791626 + ], + "logprob_reward": -0.04230903834104538, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04230903834104538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08190880715847015, + 0.06689569354057312, + 0.09187564998865128, + 0.29152873158454895, + 0.18858252465724945, + 0.025015847757458687, + 0.05717336758971214, + 0.04524390026926994, + 0.07330303639173508, + 0.07847248762845993 + ], + "credit_map": [ + 0.007236145902425051, + 0.008703775703907013, + 0.007083050906658173, + 0.008728851564228535, + 0.009466223418712616, + 0.011757181957364082, + 0.0073684221133589745, + 0.021565159782767296, + 0.007385579403489828, + 0.004794329404830933, + 0.005270459223538637, + 0.004505010787397623, + 0.006047374568879604, + 0.012635533697903156, + 0.010094513185322285, + 0.016162892803549767, + 0.008811810985207558, + 0.007159119937568903, + 0.011766818352043629, + 0.007977252826094627, + 0.008878348395228386, + 0.019592653959989548, + 0.008580619469285011, + 0.019109023734927177, + 0.007396206725388765, + 0.006473094690591097, + 0.008676448836922646, + 0.009489831514656544, + 0.01408136822283268, + 0.11373890936374664, + 0.017663655802607536, + 0.11400920897722244, + 0.012909194454550743, + 0.016804099082946777, + 0.01803380250930786, + 0.02124948799610138, + 0.017932895570993423, + 0.026342319324612617, + 0.02704143524169922, + 0.04826928302645683, + 0.003977566491812468, + 0.0025125376414507627, + 0.002172464970499277, + 0.0023825494572520256, + 0.003530723974108696, + 0.004572433885186911, + 0.002620045328512788, + 0.0032475264742970467, + 0.014359197579324245, + 0.006402470171451569, + 0.00825727079063654, + 0.004473663400858641, + 0.003296670038253069, + 0.007874620147049427, + 0.0035011491272598505, + 0.00900832749903202, + 0.00663452735170722, + 0.007890574634075165, + 0.0067011648789048195, + 0.0053230999037623405, + 0.002919195219874382, + 0.002993473783135414, + 0.00466909259557724, + 0.008112776093184948, + 0.014738153666257858, + 0.01194667723029852, + 0.01251335721462965, + 0.007803322747349739, + 0.007091001607477665, + 0.005849483422935009, + 0.006883565336465836, + 0.006477473769336939, + 0.014776377007365227, + 0.00805736891925335, + 0.01162483636289835, + 0.013033818453550339, + 0.00639727246016264, + 0.008716742508113384, + 0.006752807181328535, + 0.009113270789384842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.04230903834104538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033714812248945236 + ], + "opposite_logprob_tokens": [ + -3.408714771270752 + ], + "logprob_reward": -0.033714812248945236, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033714812248945236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12197531759738922, + 0.08502852916717529, + 0.049403999000787735, + 0.16459426283836365, + 0.10475814342498779, + 0.03974022716283798, + 0.07238566130399704, + 0.17089301347732544, + 0.11106311529874802, + 0.08015773445367813 + ], + "credit_map": [ + 0.010498914867639542, + 0.01579274795949459, + 0.020760687068104744, + 0.01646426133811474, + 0.011624759063124657, + 0.01492221187800169, + 0.013879652135074139, + 0.01803208887577057, + 0.006620862986892462, + 0.007786987815052271, + 0.013528321869671345, + 0.008415190503001213, + 0.010434407740831375, + 0.012787846848368645, + 0.0131229218095541, + 0.012331989593803883, + 0.0056053088046610355, + 0.003872738452628255, + 0.006270615383982658, + 0.005644794553518295, + 0.004778893198817968, + 0.008517967537045479, + 0.0030237697064876556, + 0.011689909733831882, + 0.010424057021737099, + 0.011469381861388683, + 0.02259690873324871, + 0.016192801296710968, + 0.019295046105980873, + 0.03610033914446831, + 0.018856002017855644, + 0.02965972013771534, + 0.00621494185179472, + 0.007839057594537735, + 0.00937812402844429, + 0.014790641143918037, + 0.009013442322611809, + 0.015248095616698265, + 0.015699321404099464, + 0.026574520394206047, + 0.005850036628544331, + 0.004217902198433876, + 0.004026900976896286, + 0.004479250404983759, + 0.004454559180885553, + 0.00738986162468791, + 0.0034325553569942713, + 0.0058891610242426395, + 0.012010580860078335, + 0.007743594702333212, + 0.0073207044042646885, + 0.008203666657209396, + 0.007834434509277344, + 0.008877312764525414, + 0.00985592883080244, + 0.010539436712861061, + 0.02229314297437668, + 0.021892108023166656, + 0.023015636950731277, + 0.025649042800068855, + 0.017008865252137184, + 0.013942812569439411, + 0.032536737620830536, + 0.014554668217897415, + 0.029223432764410973, + 0.009012042544782162, + 0.0320257730782032, + 0.012144045904278755, + 0.007301891688257456, + 0.0058206832036376, + 0.007156794425100088, + 0.00837845541536808, + 0.018590770661830902, + 0.003866000333800912, + 0.02296026051044464, + 0.006575420964509249, + 0.007002369500696659, + 0.006266404874622822, + 0.005738614592701197, + 0.009157892316579819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.033714812248945236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014217277988791466 + ], + "opposite_logprob_tokens": [ + -4.264217376708984 + ], + "logprob_reward": -0.014217277988791466, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014217277988791466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12972910702228546, + 0.08054903149604797, + 0.04766366258263588, + 0.1286599338054657, + 0.08883165568113327, + 0.04433731362223625, + 0.0816718116402626, + 0.06932765245437622, + 0.1274992972612381, + 0.20173057913780212 + ], + "credit_map": [ + 0.014310802333056927, + 0.014262919314205647, + 0.015120182186365128, + 0.010918723419308662, + 0.011832485906779766, + 0.01757902465760708, + 0.02150445431470871, + 0.024200506508350372, + 0.0070622023195028305, + 0.00870784092694521, + 0.007440083660185337, + 0.007054300978779793, + 0.009608488529920578, + 0.014570808038115501, + 0.01297754142433405, + 0.013127767480909824, + 0.0038813480641692877, + 0.002420620759949088, + 0.006355231627821922, + 0.005521855782717466, + 0.003147057257592678, + 0.007498489692807198, + 0.004525951575487852, + 0.014313106425106525, + 0.005293943453580141, + 0.005177109036594629, + 0.0073248702101409435, + 0.007485982496291399, + 0.011133644729852676, + 0.04243956878781319, + 0.00927411112934351, + 0.040530718863010406, + 0.004769221413880587, + 0.007165450602769852, + 0.007217851001769304, + 0.010322805494070053, + 0.008613398298621178, + 0.013129151426255703, + 0.011881471611559391, + 0.025732312351465225, + 0.007192028686404228, + 0.004084242507815361, + 0.004086765926331282, + 0.005650275852531195, + 0.006434930022805929, + 0.0058386973105371, + 0.004894285928457975, + 0.0061560883186757565, + 0.014418353326618671, + 0.01463409699499607, + 0.009612668305635452, + 0.006410152185708284, + 0.005505833309143782, + 0.013091499917209148, + 0.006530733313411474, + 0.011468475684523582, + 0.009456606581807137, + 0.007359529845416546, + 0.009348788298666477, + 0.008399306796491146, + 0.010980267077684402, + 0.008021664805710316, + 0.00903803389519453, + 0.006723451893776655, + 0.02715686708688736, + 0.014035187661647797, + 0.031608451157808304, + 0.01603117398917675, + 0.010142059065401554, + 0.007841452956199646, + 0.010275725275278091, + 0.010408390313386917, + 0.0563875213265419, + 0.009884944185614586, + 0.07136519253253937, + 0.013469109311699867, + 0.008520000614225864, + 0.00968223437666893, + 0.008391239680349827, + 0.024030335247516632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.014217277988791466 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988726496696472 + ], + "opposite_logprob_tokens": [ + -1.804887294769287 + ], + "logprob_reward": -0.17988726496696472, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988726496696472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12083309888839722, + 0.06512966006994247, + 0.043294671922922134, + 0.11919860541820526, + 0.04529866576194763, + 0.02557436004281044, + 0.06522174924612045, + 0.041939057409763336, + 0.28745973110198975, + 0.1860504001379013 + ], + "credit_map": [ + 0.0120620746165514, + 0.014116954989731312, + 0.012946694158017635, + 0.013013030402362347, + 0.010261723771691322, + 0.01468215323984623, + 0.01338249258697033, + 0.030367974191904068, + 0.00560388807207346, + 0.0055234553292393684, + 0.005784519948065281, + 0.006281653884798288, + 0.006436868570744991, + 0.011619995348155499, + 0.010416965931653976, + 0.013462314382195473, + 0.0031940496992319822, + 0.003449315670877695, + 0.004356756806373596, + 0.0036644029896706343, + 0.003884754842147231, + 0.008345763199031353, + 0.004023048095405102, + 0.012376579456031322, + 0.003906331490725279, + 0.004976159892976284, + 0.006375197786837816, + 0.006244760472327471, + 0.010066735558211803, + 0.0397549606859684, + 0.008717551827430725, + 0.03915691003203392, + 0.003637648420408368, + 0.003961673006415367, + 0.0047548389993608, + 0.0059952023439109325, + 0.004137914162129164, + 0.00641630869358778, + 0.006240570452064276, + 0.010154509916901588, + 0.004858177155256271, + 0.0023831413127481937, + 0.002112471731379628, + 0.0036937447730451822, + 0.002367881592363119, + 0.0029286399949342012, + 0.002917785895988345, + 0.004312516655772924, + 0.01738358661532402, + 0.008598399348556995, + 0.008833255618810654, + 0.0045662084594368935, + 0.0034541580826044083, + 0.00886374618858099, + 0.004178871866315603, + 0.009343522600829601, + 0.00535651994869113, + 0.005298989824950695, + 0.005785855930298567, + 0.004757629241794348, + 0.005976016167551279, + 0.004213874693959951, + 0.004246470518410206, + 0.006303702015429735, + 0.07684916257858276, + 0.028531517833471298, + 0.08532744646072388, + 0.02568819932639599, + 0.015035593882203102, + 0.012284774333238602, + 0.020732469856739044, + 0.023010561242699623, + 0.04357864707708359, + 0.01669188402593136, + 0.04885011911392212, + 0.02121511474251747, + 0.010257091373205185, + 0.014616933651268482, + 0.00843471847474575, + 0.02240588329732418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.17988726496696472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199766330420971 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199766330420971, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199766330420971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12042679637670517, + 0.08126522600650787, + 0.07350102066993713, + 0.04838791862130165, + 0.08660493791103363, + 0.08234729617834091, + 0.06937296688556671, + 0.07833938300609589, + 0.14869403839111328, + 0.21106041967868805 + ], + "credit_map": [ + 0.015051702968776226, + 0.011421224102377892, + 0.014649195596575737, + 0.01503577921539545, + 0.016093572601675987, + 0.015778467059135437, + 0.01678101345896721, + 0.015615835785865784, + 0.008195809088647366, + 0.007719758432358503, + 0.00677164550870657, + 0.007667151745408773, + 0.006930212490260601, + 0.013377483934164047, + 0.012745040468871593, + 0.017858117818832397, + 0.003925122786313295, + 0.004169227555394173, + 0.008875150233507156, + 0.008669780567288399, + 0.005090469494462013, + 0.012876046821475029, + 0.0049667395651340485, + 0.024928482249379158, + 0.004014302045106888, + 0.0036959678400307894, + 0.004467287566512823, + 0.0044832355342805386, + 0.005992791149765253, + 0.007089574821293354, + 0.008254706859588623, + 0.010390050709247589, + 0.005389604717493057, + 0.006241791881620884, + 0.008293794468045235, + 0.0067191640846431255, + 0.009959263727068901, + 0.013461430557072163, + 0.011383133009076118, + 0.02515675127506256, + 0.013384001329541206, + 0.006755175068974495, + 0.006820909678936005, + 0.01065101008862257, + 0.010909494943916798, + 0.012025115080177784, + 0.008834107778966427, + 0.012967478483915329, + 0.015688134357333183, + 0.007376248948276043, + 0.007809493225067854, + 0.006177276838570833, + 0.0059014419093728065, + 0.008078466169536114, + 0.006152115762233734, + 0.012189785949885845, + 0.013275661505758762, + 0.010187347419559956, + 0.012668406590819359, + 0.009201805107295513, + 0.008681908249855042, + 0.00705490680411458, + 0.007171176839619875, + 0.010098169557750225, + 0.030479371547698975, + 0.016396505758166313, + 0.02523047663271427, + 0.020046928897500038, + 0.016216399148106575, + 0.011328385211527348, + 0.015135841444134712, + 0.013860117644071579, + 0.05062117800116539, + 0.007898421958088875, + 0.06616262346506119, + 0.012982227839529514, + 0.014807088300585747, + 0.01937631331384182, + 0.012289944104850292, + 0.02692262828350067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.014199766330420971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot of an interior, shot from a slightly upward angle, slowly follows the movements of a young goat. The goat's horns are small, pure white, and its body is covered in fluffy, soft, pale golden fur, gleaming softly in the warm glow of a flickering oil lamp. A slightly worn blue fabric collar hangs around its neck. Carefully adjusting a patched, rough-hewn backpack, it stops at the doorway of a simple wooden hut. Slowly turning, its large, clear, yet sorrowful eyes cast a final, fleeting glance at its mother, who lies on a straw bed in the corner, breathing weakly. This glance is filled with reluctance and determination. Then, resolutely, it turns back and, with its front hoof, gently and almost silently pushes the heavy wooden door shut. With a dull click of the latch, the warm orange glow of the room is completely blocked out, leaving only a glimpse of the vast grassland, bathed in cold moonlight, through the crack of the door. The entire scene is imbued with a tranquil, sad, yet courageous atmosphere of departure.", + "seed": 7681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_11/index_8900_seed_7681_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026427875459194183, + "vlm_logprob": -0.026427875459194183, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.026427875459194183, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03835947811603546 + ], + "opposite_logprob_tokens": [ + -3.4133594036102295 + ], + "logprob_reward": -0.03835947811603546, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03835947811603546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06752558052539825, + 0.1865374594926834, + 0.0725349485874176, + 0.13793165981769562, + 0.12060679495334625, + 0.13367359340190887, + 0.0474049337208271, + 0.1492563784122467, + 0.048286307603120804, + 0.036242350935935974 + ], + "credit_map": [ + 0.005725824739784002, + 0.007208218798041344, + 0.011445493437349796, + 0.013589173555374146, + 0.006111512426286936, + 0.00849601998925209, + 0.006167765241116285, + 0.008781570009887218, + 0.014141358435153961, + 0.009082605130970478, + 0.019289443269371986, + 0.045820463448762894, + 0.009651174768805504, + 0.06195143237709999, + 0.009527551010251045, + 0.017073439434170723, + 0.007697472348809242, + 0.0038805704098194838, + 0.018225625157356262, + 0.012156978249549866, + 0.004931292962282896, + 0.007470875512808561, + 0.008397948928177357, + 0.00977418851107359, + 0.009858688339591026, + 0.011175301857292652, + 0.01593012735247612, + 0.015678634867072105, + 0.02920875884592533, + 0.0128870140761137, + 0.018205976113677025, + 0.024987148120999336, + 0.011125900782644749, + 0.012785348109900951, + 0.013912249356508255, + 0.017365863546729088, + 0.011374511756002903, + 0.01796496845781803, + 0.01922682859003544, + 0.01685112528502941, + 0.014870386570692062, + 0.020724831148982048, + 0.015607212670147419, + 0.021301155909895897, + 0.008554914966225624, + 0.00980436522513628, + 0.01840078830718994, + 0.024409940466284752, + 0.005063331685960293, + 0.00495168287307024, + 0.006168263964354992, + 0.006870762445032597, + 0.005274906288832426, + 0.0035098199732601643, + 0.009920516051352024, + 0.00564564811065793, + 0.014559497125446796, + 0.01069680042564869, + 0.03289726376533508, + 0.010736098513007164, + 0.005573991220444441, + 0.0194539837539196, + 0.00894167274236679, + 0.04639706388115883, + 0.005151430610567331, + 0.006178249139338732, + 0.01022125594317913, + 0.006520155351608992, + 0.0041669453494250774, + 0.004186656326055527, + 0.004974849987775087, + 0.006886766292154789, + 0.0033862581476569176, + 0.0025895372964441776, + 0.003275244263932109, + 0.005517832934856415, + 0.0028626611456274986, + 0.007017261348664761, + 0.007241533603519201, + 0.0043520210310816765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.03835947811603546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811061009764671 + ], + "opposite_logprob_tokens": [ + -3.2881104946136475 + ], + "logprob_reward": -0.03811061009764671, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811061009764671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1792469620704651, + 0.19673006236553192, + 0.07643275707960129, + 0.11177077144384384, + 0.12186644226312637, + 0.06819774210453033, + 0.02381671965122223, + 0.08953167498111725, + 0.06264197081327438, + 0.0697648748755455 + ], + "credit_map": [ + 0.02344551309943199, + 0.021048463881015778, + 0.024595575407147408, + 0.028780663385987282, + 0.02130364254117012, + 0.022592950612306595, + 0.017804760485887527, + 0.019675396382808685, + 0.010007998906075954, + 0.01121776457875967, + 0.024108823388814926, + 0.0488276332616806, + 0.009567400440573692, + 0.06381534785032272, + 0.009867202490568161, + 0.019317883998155594, + 0.009434955194592476, + 0.005284473299980164, + 0.014139746315777302, + 0.014613242819905281, + 0.005159957334399223, + 0.008755992166697979, + 0.008907235227525234, + 0.010137151926755905, + 0.010590868070721626, + 0.009881061501801014, + 0.012307720258831978, + 0.011007984168827534, + 0.019011665135622025, + 0.0142494672909379, + 0.01602756977081299, + 0.0186944380402565, + 0.010823523625731468, + 0.011625254526734352, + 0.012983140535652637, + 0.021767744794487953, + 0.012231582775712013, + 0.013023094274103642, + 0.019162673503160477, + 0.02024943009018898, + 0.00568855507299304, + 0.0075075323693454266, + 0.011064044199883938, + 0.012111939489841461, + 0.00480281375348568, + 0.008361062966287136, + 0.009833414107561111, + 0.008828384801745415, + 0.002438584342598915, + 0.002249371027573943, + 0.0027565923519432545, + 0.0041043260134756565, + 0.0020837155170738697, + 0.0024736409541219473, + 0.0038809606339782476, + 0.0038295304402709007, + 0.008721076883375645, + 0.006355307064950466, + 0.01666993275284767, + 0.007893918082118034, + 0.005327619146555662, + 0.01310016494244337, + 0.008001693524420261, + 0.023461971431970596, + 0.0058140940964221954, + 0.005579501390457153, + 0.008170051500201225, + 0.006236381363123655, + 0.005285633262246847, + 0.005925875157117844, + 0.01812710426747799, + 0.007503325119614601, + 0.006241575349122286, + 0.01474457886070013, + 0.007095988839864731, + 0.01426671538501978, + 0.006085904315114021, + 0.00649786414578557, + 0.008922884240746498, + 0.005909367464482784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.03811061009764671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797343984246254 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797343984246254, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797343984246254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06589797884225845, + 0.21764259040355682, + 0.11631762981414795, + 0.03257349878549576, + 0.12441995739936829, + 0.03262033313512802, + 0.16422204673290253, + 0.06903273612260818, + 0.115423783659935, + 0.061849445104599 + ], + "credit_map": [ + 0.008034557104110718, + 0.005291261710226536, + 0.009582030586898327, + 0.009902017191052437, + 0.007350768428295851, + 0.007081512827426195, + 0.007672091946005821, + 0.010983739979565144, + 0.011955788359045982, + 0.009302498772740364, + 0.016597937792539597, + 0.06118262931704521, + 0.008185187354683876, + 0.0796375647187233, + 0.009075157344341278, + 0.021705834195017815, + 0.007706211879849434, + 0.006843515671789646, + 0.02060074731707573, + 0.019874785095453262, + 0.009237064979970455, + 0.019728094339370728, + 0.012389153242111206, + 0.019938062876462936, + 0.002758442424237728, + 0.002500668168067932, + 0.0034652496688067913, + 0.006405852735042572, + 0.0031117871403694153, + 0.004240908659994602, + 0.005540698301047087, + 0.004549894481897354, + 0.012399803847074509, + 0.01927349716424942, + 0.013163254596292973, + 0.019913261756300926, + 0.01113270316272974, + 0.012376292608678341, + 0.018800508230924606, + 0.017360640689730644, + 0.0031600119546055794, + 0.0045714364387094975, + 0.0033827307634055614, + 0.0045110187493264675, + 0.0023041998501867056, + 0.0032104733400046825, + 0.006068974267691374, + 0.005411485675722361, + 0.015628905966877937, + 0.01809161901473999, + 0.02492835372686386, + 0.022027157247066498, + 0.020552335307002068, + 0.015403153374791145, + 0.030901212245225906, + 0.016689304262399673, + 0.005697302520275116, + 0.005670780781656504, + 0.008722689934074879, + 0.009078399278223515, + 0.0054948353208601475, + 0.010512462817132473, + 0.014103555120527744, + 0.009752705693244934, + 0.012170150876045227, + 0.010702278465032578, + 0.03176039084792137, + 0.013672723434865475, + 0.00856065284460783, + 0.0095425546169281, + 0.012462921440601349, + 0.016552114859223366, + 0.006051205564290285, + 0.012584450654685497, + 0.0060116141103208065, + 0.012820001691579819, + 0.00593839306384325, + 0.005429527256637812, + 0.007937582209706306, + 0.005076672416180372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.009797343984246254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800057411193848 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029800057411193848, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800057411193848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11300274729728699, + 0.12094296514987946, + 0.07776786386966705, + 0.0930526852607727, + 0.08796116709709167, + 0.08321423083543777, + 0.0962100476026535, + 0.15489786863327026, + 0.10905014723539352, + 0.06390023976564407 + ], + "credit_map": [ + 0.016121044754981995, + 0.008806080557405949, + 0.01918121799826622, + 0.0150684779509902, + 0.012477460317313671, + 0.011866467073559761, + 0.014267142862081528, + 0.015214861370623112, + 0.012638556770980358, + 0.007551079615950584, + 0.020186353474855423, + 0.031995054334402084, + 0.006571274250745773, + 0.02630871720612049, + 0.006279222667217255, + 0.009412712417542934, + 0.0077215442433953285, + 0.006180040538311005, + 0.010134229436516762, + 0.01236905250698328, + 0.007263118401169777, + 0.015283135697245598, + 0.0072989147156476974, + 0.011517830193042755, + 0.006930612493306398, + 0.008016376756131649, + 0.012498263269662857, + 0.014670668169856071, + 0.012755650095641613, + 0.011088618077337742, + 0.01303398609161377, + 0.014058509841561317, + 0.0077414745464921, + 0.010303092189133167, + 0.008855379186570644, + 0.014065650291740894, + 0.00854957289993763, + 0.00956767424941063, + 0.01518956758081913, + 0.013688765466213226, + 0.006945949513465166, + 0.01021383237093687, + 0.0107857221737504, + 0.017130473628640175, + 0.004844268783926964, + 0.0051031396724283695, + 0.016048062592744827, + 0.012142788618803024, + 0.010911800898611546, + 0.013520117849111557, + 0.011086942628026009, + 0.01174312736839056, + 0.014899231493473053, + 0.008368098177015781, + 0.013907955959439278, + 0.011772776953876019, + 0.014740440994501114, + 0.014412125572562218, + 0.021348053589463234, + 0.017051899805665016, + 0.015485166572034359, + 0.02812199480831623, + 0.022790731862187386, + 0.020947471261024475, + 0.01216722372919321, + 0.011933723464608192, + 0.017524143680930138, + 0.011568042449653149, + 0.008855299092829227, + 0.015819568186998367, + 0.015817930921912193, + 0.015364224091172218, + 0.006504984572529793, + 0.008465672843158245, + 0.006751750595867634, + 0.00797909963876009, + 0.006471763830631971, + 0.007611684035509825, + 0.010029568336904049, + 0.010085714980959892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.029800057411193848 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a simple wooden hut with a straw bed in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young goat with small pure white horns and fluffy pale golden fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the goat wear a slightly worn blue fabric collar and adjust a patched rough-hewn backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the goat push a heavy wooden door shut with its front hoof, blocking the warm orange room light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot from a slightly upward angle that slowly follows the goat's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607188768684864 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607188768684864, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607188768684864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11100693792104721, + 0.1038280799984932, + 0.06415259838104248, + 0.06717655807733536, + 0.10816556960344315, + 0.05231823772192001, + 0.1277392953634262, + 0.09659793227910995, + 0.13572901487350464, + 0.13328580558300018 + ], + "credit_map": [ + 0.009110222570598125, + 0.00840106513351202, + 0.0265209898352623, + 0.025000816211104393, + 0.008116519078612328, + 0.010548314079642296, + 0.01139665488153696, + 0.01191235613077879, + 0.01291675865650177, + 0.0056721824221313, + 0.017563723027706146, + 0.01754368282854557, + 0.005352768115699291, + 0.02626134641468525, + 0.005315598100423813, + 0.01320202462375164, + 0.005616940092295408, + 0.004771190695464611, + 0.011816137470304966, + 0.013446570374071598, + 0.004407780710607767, + 0.007419086992740631, + 0.008264763280749321, + 0.008410128764808178, + 0.005131847225129604, + 0.007242712192237377, + 0.011485066264867783, + 0.011351481080055237, + 0.004488794598728418, + 0.0065504261292517185, + 0.010237044654786587, + 0.010689187794923782, + 0.009391074068844318, + 0.022332701832056046, + 0.010903838090598583, + 0.01593862846493721, + 0.011514979414641857, + 0.010541388764977455, + 0.013423372060060501, + 0.014119590632617474, + 0.005759756080806255, + 0.006389731075614691, + 0.003839312819764018, + 0.008469291031360626, + 0.0033916498068720102, + 0.004291222430765629, + 0.010120291262865067, + 0.01005698461085558, + 0.012425695545971394, + 0.014630459249019623, + 0.017453962936997414, + 0.019809918478131294, + 0.01525270938873291, + 0.011689228005707264, + 0.02361789159476757, + 0.012859438546001911, + 0.008331787772476673, + 0.007977040484547615, + 0.012286643497645855, + 0.00800925213843584, + 0.004782096482813358, + 0.015722135081887245, + 0.01272879634052515, + 0.02676018327474594, + 0.014336780644953251, + 0.01632518693804741, + 0.02749902382493019, + 0.01803329400718212, + 0.009753569960594177, + 0.011336594820022583, + 0.02242492511868477, + 0.01601964607834816, + 0.011096669360995293, + 0.02473120577633381, + 0.014610465615987778, + 0.022644231095910072, + 0.014025034382939339, + 0.013410025276243687, + 0.018838820978999138, + 0.013929353095591068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.01607188768684864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_12/index_tensor([8104, 8104])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_12/index_tensor([8104, 8104])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f71b718e993f2c7b87e15564548f52c215522420 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_12/index_tensor([8104, 8104])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle crane shot begins, slowly zooming in on a classic 8-bit retro game world, unfolding in a bright, shadowless, flat lighting. In the center of the frame, a massive Armand pine tree, composed of dark brown and green pixel blocks, its crown an irregular, jagged umbrella-like outline, is used as a battering ram, violently ramming into the walls of a magnificent castle constructed from gray and dark gray pixel bricks. With each heavy, rhythmic impact, the tree's pixelated trunk trembles violently from the immense recoil, while the castle's stone walls erupt in a cloud of shattered pixels, creating a splattering gray \"debris\" while several black pixel cracks rapidly spread across the wall. In the background, a pure, flat blue-blue sky is framed by a few simple, white, blocky clouds, motionless. As the camera steadily and calmly approaches the chaotic siege scene, the audience can more clearly see the details of the destruction around the impact point. This slow advancement contrasts sharply with the fierce destruction, creating an epic siege atmosphere that is both grand and tense in the limited and vibrant retro tones.", + "seed": 5777, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_12/index_8104_seed_5777_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19515559412539005, + "vlm_logprob": -0.19515559412539005, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.19515559412539005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05186450853943825 + ], + "opposite_logprob_tokens": [ + -3.0518646240234375 + ], + "logprob_reward": -0.05186450853943825, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05186450853943825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1925574094057083, + 0.04517100378870964, + 0.04299405962228775, + 0.15169788897037506, + 0.08572550863027573, + 0.08056815713644028, + 0.14886809885501862, + 0.06230643764138222, + 0.06533345580101013, + 0.12477796524763107 + ], + "credit_map": [ + 0.03240196406841278, + 0.01937905140221119, + 0.01560882292687893, + 0.022161046043038368, + 0.016240671277046204, + 0.030980035662651062, + 0.025665616616606712, + 0.03012019209563732, + 0.004586272872984409, + 0.009012891910970211, + 0.0030118057038635015, + 0.005017692223191261, + 0.006437904667109251, + 0.006278305780142546, + 0.004710752982646227, + 0.006115378811955452, + 0.0038414238952100277, + 0.0054605924524366856, + 0.005177420098334551, + 0.008644242770969868, + 0.006096486933529377, + 0.005017491057515144, + 0.0037674838677048683, + 0.004988916218280792, + 0.016172070056200027, + 0.022214587777853012, + 0.013881517574191093, + 0.014799889177083969, + 0.03115800768136978, + 0.022864051163196564, + 0.013680425472557545, + 0.016927337273955345, + 0.022574763745069504, + 0.009990565478801727, + 0.005936001893132925, + 0.009698520414531231, + 0.01418241299688816, + 0.005649685859680176, + 0.005120175890624523, + 0.01257338747382164, + 0.015592805109918118, + 0.009482776746153831, + 0.007654738612473011, + 0.010817065834999084, + 0.014362354762852192, + 0.0069576045498251915, + 0.009014277718961239, + 0.00668653380125761, + 0.023362010717391968, + 0.012142707593739033, + 0.005715375766158104, + 0.007097215391695499, + 0.05119117349386215, + 0.013480947352945805, + 0.00880091916769743, + 0.027077747508883476, + 0.007688481826335192, + 0.008764169178903103, + 0.005829853005707264, + 0.008714187890291214, + 0.0055814641527831554, + 0.005657704081386328, + 0.012919306755065918, + 0.007151270285248756, + 0.009199011139571667, + 0.011056763119995594, + 0.008903097361326218, + 0.008283267728984356, + 0.006625506095588207, + 0.006827938836067915, + 0.00656781904399395, + 0.007870052941143513, + 0.006074302829802036, + 0.010915673337876797, + 0.0072425189428031445, + 0.008417836390435696, + 0.010434307157993317, + 0.034628648310899734, + 0.011570188216865063, + 0.035494495183229446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.05186450853943825 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814103364944458 + ], + "opposite_logprob_tokens": [ + -1.4064103364944458 + ], + "logprob_reward": -0.2814103364944458, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814103364944458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1630241423845291, + 0.05643899366259575, + 0.07122825086116791, + 0.08326895534992218, + 0.10146097838878632, + 0.06371685117483139, + 0.07562236487865448, + 0.07596739381551743, + 0.20234708487987518, + 0.10692501813173294 + ], + "credit_map": [ + 0.026222731918096542, + 0.013281788676977158, + 0.013914193958044052, + 0.017117951065301895, + 0.013756530359387398, + 0.027972491458058357, + 0.01733696646988392, + 0.03342149406671524, + 0.006051615346223116, + 0.011136217974126339, + 0.004213201347738504, + 0.005202892702072859, + 0.009199090301990509, + 0.007823343388736248, + 0.005298946984112263, + 0.007513686083257198, + 0.0072652907110750675, + 0.008034122176468372, + 0.007095532491803169, + 0.011035668663680553, + 0.01015875767916441, + 0.010721314698457718, + 0.008149197325110435, + 0.008768364787101746, + 0.007532775867730379, + 0.012084960006177425, + 0.0077011464163661, + 0.017506681382656097, + 0.01290859840810299, + 0.008412919007241726, + 0.008733999915421009, + 0.008387867361307144, + 0.015500537119805813, + 0.016805939376354218, + 0.006356075871735811, + 0.012381442822515965, + 0.023378843441605568, + 0.008036088198423386, + 0.009359000250697136, + 0.009643053635954857, + 0.01540047861635685, + 0.006415463984012604, + 0.002117226365953684, + 0.003809596411883831, + 0.02159404195845127, + 0.004386163782328367, + 0.003566548228263855, + 0.006427327636629343, + 0.010427269153296947, + 0.006626377813518047, + 0.00417354516685009, + 0.00526864267885685, + 0.021510837599635124, + 0.00802762433886528, + 0.005925256758928299, + 0.013662807643413544, + 0.007232964038848877, + 0.007025096565485001, + 0.0062167043797671795, + 0.007001700811088085, + 0.012987791560590267, + 0.008308722637593746, + 0.013588765636086464, + 0.013605650514364243, + 0.021268149837851524, + 0.01733638532459736, + 0.00870078057050705, + 0.010729710571467876, + 0.06265109032392502, + 0.024343052878975868, + 0.012211557477712631, + 0.045106347650289536, + 0.008159416727721691, + 0.010360747575759888, + 0.01001990120857954, + 0.011466042138636112, + 0.01752348430454731, + 0.009497019462287426, + 0.02887778915464878, + 0.011020618490874767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.2814103364944458 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.387099951505661 + ], + "opposite_logprob_tokens": [ + -1.1370999813079834 + ], + "logprob_reward": -0.387099951505661, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.387099951505661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061747778207063675, + 0.024321788921952248, + 0.05515279993414879, + 0.25288745760917664, + 0.12884923815727234, + 0.15299774706363678, + 0.014251425862312317, + 0.14154252409934998, + 0.026176659390330315, + 0.1420726329088211 + ], + "credit_map": [ + 0.015571472235023975, + 0.006095673888921738, + 0.005077203270047903, + 0.005343260709196329, + 0.0037921599578112364, + 0.010152601636946201, + 0.006207666359841824, + 0.009507735259830952, + 0.003499576821923256, + 0.003162515116855502, + 0.002261113841086626, + 0.0033055306412279606, + 0.002316539641469717, + 0.0036597074940800667, + 0.003378463676199317, + 0.0027383416891098022, + 0.0047666155733168125, + 0.006995301228016615, + 0.003472178941592574, + 0.011400831863284111, + 0.008097870275378227, + 0.007898681797087193, + 0.004927301779389381, + 0.007594017777591944, + 0.023983901366591454, + 0.054256051778793335, + 0.019254488870501518, + 0.020620934665203094, + 0.05435575917363167, + 0.0320463590323925, + 0.017539875581860542, + 0.030830085277557373, + 0.02272014133632183, + 0.02939729578793049, + 0.003689178731292486, + 0.00778427068144083, + 0.043719880282878876, + 0.006458404939621687, + 0.00515922624617815, + 0.009920839220285416, + 0.032506220042705536, + 0.015873080119490623, + 0.009284891188144684, + 0.014825121499598026, + 0.04338372126221657, + 0.01000719889998436, + 0.01057441160082817, + 0.01654309406876564, + 0.0016379875596612692, + 0.001510255504399538, + 0.0012140471953898668, + 0.0015032527735456824, + 0.0027963502798229456, + 0.0022937722969800234, + 0.0017499028472229838, + 0.0015458576381206512, + 0.014929162338376045, + 0.016021016985177994, + 0.014021206647157669, + 0.018501678481698036, + 0.012581586837768555, + 0.014710051007568836, + 0.035328857600688934, + 0.015448963269591331, + 0.003176306141540408, + 0.002665446139872074, + 0.0017384947277605534, + 0.0018609616672620177, + 0.00719776377081871, + 0.002659541554749012, + 0.0020896231289952993, + 0.004788520280271769, + 0.006411287002265453, + 0.01461760699748993, + 0.00855319481343031, + 0.009910565800964832, + 0.01217830739915371, + 0.03820634260773659, + 0.01255776546895504, + 0.039637546986341476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.387099951505661 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560253739356995 + ], + "opposite_logprob_tokens": [ + -1.6006025075912476 + ], + "logprob_reward": -0.22560253739356995, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560253739356995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07985064387321472, + 0.03395039588212967, + 0.060943637043237686, + 0.23869659006595612, + 0.061015382409095764, + 0.10843405872583389, + 0.03820403292775154, + 0.10820572823286057, + 0.051509611308574677, + 0.21918992698192596 + ], + "credit_map": [ + 0.023186100646853447, + 0.006737012416124344, + 0.005645516328513622, + 0.00655681686475873, + 0.004398023243993521, + 0.014010095037519932, + 0.006952657364308834, + 0.012364417314529419, + 0.0046775019727647305, + 0.004310646094381809, + 0.003654950764030218, + 0.004743280354887247, + 0.003296098904684186, + 0.004538822919130325, + 0.004773683380335569, + 0.003955406602472067, + 0.005587241146713495, + 0.007656036410480738, + 0.0038364313077181578, + 0.011191985569894314, + 0.008828452788293362, + 0.010423921048641205, + 0.005180578213185072, + 0.008238987997174263, + 0.02053924836218357, + 0.058556247502565384, + 0.01485024206340313, + 0.021262191236019135, + 0.05442670360207558, + 0.026505345478653908, + 0.01607840694487095, + 0.026478195562958717, + 0.010944978334009647, + 0.009531413204967976, + 0.004831819795072079, + 0.009095968678593636, + 0.009929035790264606, + 0.00421987334266305, + 0.006353522650897503, + 0.006108766887336969, + 0.020770398899912834, + 0.012632823549211025, + 0.00826488807797432, + 0.010875956155359745, + 0.026920590549707413, + 0.007905518636107445, + 0.009794143959879875, + 0.011269735172390938, + 0.004002783913165331, + 0.0034036969300359488, + 0.003565059741958976, + 0.0034028179943561554, + 0.009724604897201061, + 0.004231747705489397, + 0.0033554495312273502, + 0.006517872679978609, + 0.01062015537172556, + 0.011335592716932297, + 0.010058339685201645, + 0.012869182042777538, + 0.010899987071752548, + 0.012685202993452549, + 0.023195549845695496, + 0.016541706398129463, + 0.004785868339240551, + 0.004640623927116394, + 0.002614592434838414, + 0.0035969289019703865, + 0.014519533142447472, + 0.006543369963765144, + 0.0041611818596720695, + 0.01064750924706459, + 0.010712193325161934, + 0.027070725336670876, + 0.01597323827445507, + 0.017820008099079132, + 0.017563458532094955, + 0.05202895775437355, + 0.020755736157298088, + 0.05726557970046997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.22560253739356995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980063669383526 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980063669383526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980063669383526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0713689923286438, + 0.03294055536389351, + 0.06802695989608765, + 0.18585984408855438, + 0.1466948390007019, + 0.19948482513427734, + 0.06602560728788376, + 0.08739550411701202, + 0.025023316964507103, + 0.11717959493398666 + ], + "credit_map": [ + 0.0180146973580122, + 0.006433432921767235, + 0.007229648996144533, + 0.006612181197851896, + 0.004667453467845917, + 0.01068203430622816, + 0.0067382799461483955, + 0.010991265997290611, + 0.005880337208509445, + 0.002402253681793809, + 0.0033500315621495247, + 0.002378854900598526, + 0.0026379572227597237, + 0.005552584305405617, + 0.004047228954732418, + 0.006691308226436377, + 0.006906006019562483, + 0.009872165508568287, + 0.0073205637745559216, + 0.012044904753565788, + 0.009083694778382778, + 0.008351721800863743, + 0.006879197899252176, + 0.007568704895675182, + 0.019424336031079292, + 0.030729159712791443, + 0.014711249619722366, + 0.020456617698073387, + 0.03424104303121567, + 0.024857230484485626, + 0.01735168695449829, + 0.02408851869404316, + 0.029759664088487625, + 0.028534937649965286, + 0.006981877610087395, + 0.01017431914806366, + 0.03709489479660988, + 0.009601659141480923, + 0.008442862890660763, + 0.016104625537991524, + 0.0410429984331131, + 0.022872190922498703, + 0.010591988451778889, + 0.02558146044611931, + 0.05286448448896408, + 0.011667200364172459, + 0.013302176259458065, + 0.02156231924891472, + 0.009100789204239845, + 0.0064336685463786125, + 0.0028992616571485996, + 0.005363287404179573, + 0.02001829259097576, + 0.0061930459924042225, + 0.004886049777269363, + 0.011131211183965206, + 0.007643597200512886, + 0.010655917227268219, + 0.007408861070871353, + 0.010135361924767494, + 0.009951252490282059, + 0.01035210583359003, + 0.01545753888785839, + 0.015790874138474464, + 0.003564699087291956, + 0.004002640955150127, + 0.0032101860269904137, + 0.0033787437714636326, + 0.0029588707257062197, + 0.0026078689843416214, + 0.002791425446048379, + 0.0025088826660066843, + 0.007699926849454641, + 0.012733074836432934, + 0.009451176039874554, + 0.010067807510495186, + 0.014480482786893845, + 0.021561574190855026, + 0.019843999296426773, + 0.02134155109524727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.02980063669383526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle crane shot begins, slowly zooming in on a classic 8-bit retro game world, unfolding in a bright, shadowless, flat lighting. In the center of the frame, a massive Armand pine tree, composed of dark brown and green pixel blocks, its crown an irregular, jagged umbrella-like outline, is used as a battering ram, violently ramming into the walls of a magnificent castle constructed from gray and dark gray pixel bricks. With each heavy, rhythmic impact, the tree's pixelated trunk trembles violently from the immense recoil, while the castle's stone walls erupt in a cloud of shattered pixels, creating a splattering gray \"debris\" while several black pixel cracks rapidly spread across the wall. In the background, a pure, flat blue-blue sky is framed by a few simple, white, blocky clouds, motionless. As the camera steadily and calmly approaches the chaotic siege scene, the audience can more clearly see the details of the destruction around the impact point. This slow advancement contrasts sharply with the fierce destruction, creating an epic siege atmosphere that is both grand and tense in the limited and vibrant retro tones.", + "seed": 5778, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_12/index_8104_seed_5778_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15744020948186516, + "vlm_logprob": -0.15744020948186516, + "vlm_yes_no_margin": 3.150000011920929, + "vlm_token_credit": -0.15744020948186516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004532777238637209 + ], + "opposite_logprob_tokens": [ + -6.629532814025879 + ], + "logprob_reward": -0.004532777238637209, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.004532777238637209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17118729650974274, + 0.06814907491207123, + 0.07482021301984787, + 0.1289605349302292, + 0.0512622632086277, + 0.05359252169728279, + 0.02574903704226017, + 0.049063149839639664, + 0.1556427925825119, + 0.2215731292963028 + ], + "credit_map": [ + 0.031633079051971436, + 0.01414481457322836, + 0.039469413459300995, + 0.01595514640212059, + 0.013826088048517704, + 0.025386177003383636, + 0.013195622712373734, + 0.01757696084678173, + 0.009350377134978771, + 0.005965006072074175, + 0.013017002493143082, + 0.008635945618152618, + 0.007265132386237383, + 0.009845723398029804, + 0.006254374980926514, + 0.00781551655381918, + 0.01079826895147562, + 0.011641740798950195, + 0.01336270198225975, + 0.006436944007873535, + 0.008312728255987167, + 0.01093147974461317, + 0.006874726619571447, + 0.006461624056100845, + 0.01513578835874796, + 0.029760442674160004, + 0.00885014422237873, + 0.006071026902645826, + 0.043634410947561264, + 0.010356107726693153, + 0.005917530972510576, + 0.009235075674951077, + 0.0035392516292631626, + 0.0073670679703354836, + 0.005523309577256441, + 0.006752585992217064, + 0.009216331876814365, + 0.006000908091664314, + 0.005507074296474457, + 0.0073557328432798386, + 0.00653444929048419, + 0.009503941982984543, + 0.011017999611794949, + 0.004480138421058655, + 0.007491739466786385, + 0.004937150981277227, + 0.0037597983609884977, + 0.005867304280400276, + 0.002714972011744976, + 0.004080762155354023, + 0.004146534949541092, + 0.002353828866034746, + 0.004380037542432547, + 0.0032178531400859356, + 0.0023776856251060963, + 0.0024773625191301107, + 0.004760690499097109, + 0.007306583691388369, + 0.007461707107722759, + 0.004442885052412748, + 0.004569489508867264, + 0.007083123084157705, + 0.005178018938750029, + 0.008260654285550117, + 0.01696331426501274, + 0.013542785309255123, + 0.014926093630492687, + 0.014645388349890709, + 0.02370871603488922, + 0.037998758256435394, + 0.013202541507780552, + 0.02065519243478775, + 0.017523376271128654, + 0.012032654136419296, + 0.010260047391057014, + 0.01783090829849243, + 0.0627003088593483, + 0.013373068533837795, + 0.017974328249692917, + 0.06987844407558441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.004532777238637209 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38712117075920105 + ], + "opposite_logprob_tokens": [ + -1.1371212005615234 + ], + "logprob_reward": -0.38712117075920105, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38712117075920105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13795143365859985, + 0.03376328945159912, + 0.07365141808986664, + 0.035164471715688705, + 0.07112251967191696, + 0.06891094893217087, + 0.019935941323637962, + 0.06369361281394958, + 0.05853421986103058, + 0.4372721314430237 + ], + "credit_map": [ + 0.024236317723989487, + 0.013344066217541695, + 0.024429602548480034, + 0.013102313503623009, + 0.018250538036227226, + 0.018052557483315468, + 0.014942693524062634, + 0.011593356728553772, + 0.005115003790706396, + 0.004020580090582371, + 0.004580371081829071, + 0.003092286642640829, + 0.005060398485511541, + 0.005384732503443956, + 0.003191839437931776, + 0.0033180774189531803, + 0.011040328070521355, + 0.007701788563281298, + 0.015989024192094803, + 0.005415534134954214, + 0.007988665252923965, + 0.013588237576186657, + 0.005624891258776188, + 0.006302951369434595, + 0.0067424471490085125, + 0.004575757309794426, + 0.00415708078071475, + 0.0023348357062786818, + 0.005812130868434906, + 0.0064199562184512615, + 0.00211355066858232, + 0.003008710453286767, + 0.007182215340435505, + 0.010370783507823944, + 0.00957584474235773, + 0.008661339990794659, + 0.012026366777718067, + 0.009075427427887917, + 0.005948768928647041, + 0.008281770162284374, + 0.009538921527564526, + 0.013340962119400501, + 0.009376389905810356, + 0.006168355233967304, + 0.009177927859127522, + 0.008397972211241722, + 0.006857698783278465, + 0.006052715238183737, + 0.0020887828432023525, + 0.00300817028619349, + 0.00265448447316885, + 0.002047573681920767, + 0.0038589010946452618, + 0.0027143284678459167, + 0.0015650111017748713, + 0.0019986897241324186, + 0.008232630789279938, + 0.0108120646327734, + 0.009063851088285446, + 0.005329497158527374, + 0.006527318153530359, + 0.010501922108232975, + 0.00581664964556694, + 0.00740967458114028, + 0.006317140068858862, + 0.005631731357425451, + 0.005725845228880644, + 0.005970303900539875, + 0.008358287625014782, + 0.014348302967846394, + 0.005027593579143286, + 0.0071550109423696995, + 0.025742650032043457, + 0.018139803782105446, + 0.016566399484872818, + 0.03755345568060875, + 0.09590932726860046, + 0.025257376953959465, + 0.03312738239765167, + 0.1849757730960846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.38712117075920105 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583707839250565 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583707839250565, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583707839250565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06352300196886063, + 0.044261228293180466, + 0.16878454387187958, + 0.14875392615795135, + 0.08664047718048096, + 0.09651701152324677, + 0.01800098456442356, + 0.027999503538012505, + 0.05462312325835228, + 0.2908962070941925 + ], + "credit_map": [ + 0.00873618945479393, + 0.006849157623946667, + 0.008199218660593033, + 0.007332026958465576, + 0.011093811132013798, + 0.007749417796730995, + 0.007780846673995256, + 0.005782333668321371, + 0.006685819011181593, + 0.004384829197078943, + 0.007282101083546877, + 0.004908852744847536, + 0.005126964766532183, + 0.005457575898617506, + 0.005016049835830927, + 0.00539903249591589, + 0.03826213255524635, + 0.02125532552599907, + 0.019519660621881485, + 0.014397691935300827, + 0.018195586279034615, + 0.03277558460831642, + 0.011924254707992077, + 0.012454290874302387, + 0.020458269864320755, + 0.03517292067408562, + 0.010850577615201473, + 0.006895518396049738, + 0.05181053653359413, + 0.007840144447982311, + 0.00694763520732522, + 0.008778308518230915, + 0.00733055779710412, + 0.013697423040866852, + 0.011606666259467602, + 0.011346105486154556, + 0.014938882552087307, + 0.010141387581825256, + 0.007824992761015892, + 0.009754454717040062, + 0.014400791376829147, + 0.018789391964673996, + 0.014790426008403301, + 0.0085787083953619, + 0.014496397227048874, + 0.010446830652654171, + 0.007306870073080063, + 0.007707595359534025, + 0.002093723975121975, + 0.002615709789097309, + 0.003574275877326727, + 0.001882081269286573, + 0.002990893553942442, + 0.002344701439142227, + 0.0011265805223956704, + 0.0013730170903727412, + 0.0021317468490451574, + 0.002824000548571348, + 0.006594824139028788, + 0.004224826116114855, + 0.0029232541564852, + 0.003295136848464608, + 0.002847803756594658, + 0.0031579097267240286, + 0.005698332563042641, + 0.003261223202571273, + 0.005375794600695372, + 0.004707181826233864, + 0.008606255054473877, + 0.013151494786143303, + 0.006056768819689751, + 0.0077660647220909595, + 0.01415972039103508, + 0.012897021137177944, + 0.01130446046590805, + 0.024910693988204002, + 0.06953132152557373, + 0.016468459740281105, + 0.02229183353483677, + 0.1193326860666275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.020583707839250565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339460164308548 + ], + "opposite_logprob_tokens": [ + -3.651339530944824 + ], + "logprob_reward": -0.026339460164308548, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339460164308548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04317779093980789, + 0.04945478215813637, + 0.050656747072935104, + 0.056055910885334015, + 0.03388088196516037, + 0.06717123091220856, + 0.016319947317242622, + 0.024317631497979164, + 0.3284018039703369, + 0.3305632174015045 + ], + "credit_map": [ + 0.006573956925421953, + 0.00622600968927145, + 0.006534503772854805, + 0.0037348445039242506, + 0.006957344710826874, + 0.006029145326465368, + 0.0038567923475056887, + 0.0032651980873197317, + 0.00740045215934515, + 0.005083381198346615, + 0.007086112163960934, + 0.005514401942491531, + 0.008648999035358429, + 0.006525611504912376, + 0.004214856307953596, + 0.004980971571058035, + 0.0100789163261652, + 0.004581776447594166, + 0.008869570679962635, + 0.004144231788814068, + 0.004392682109028101, + 0.010110395960509777, + 0.00425468385219574, + 0.004224489443004131, + 0.007188227027654648, + 0.013389197178184986, + 0.003373560030013323, + 0.0015965086640790105, + 0.019636288285255432, + 0.004722177516669035, + 0.0026156490202993155, + 0.003534299088642001, + 0.0034756939858198166, + 0.004890176467597485, + 0.0035488505382090807, + 0.0035939933732151985, + 0.00785907357931137, + 0.003476712154224515, + 0.002737364498898387, + 0.004299018532037735, + 0.00785957369953394, + 0.014163204468786716, + 0.010771340690553188, + 0.006976000964641571, + 0.009711522608995438, + 0.00725594675168395, + 0.005590145941823721, + 0.004843497183173895, + 0.0016021644696593285, + 0.0020078159868717194, + 0.0030125840567052364, + 0.0020886205602437258, + 0.0024289507418870926, + 0.0017773183062672615, + 0.0016203784616664052, + 0.0017821149667724967, + 0.002623599022626877, + 0.003886743215844035, + 0.00494325952604413, + 0.0026470583397895098, + 0.002681986428797245, + 0.003059361595660448, + 0.002236933447420597, + 0.002238689688965678, + 0.036407943814992905, + 0.02943621575832367, + 0.0351867601275444, + 0.038722600787878036, + 0.04163443297147751, + 0.06809795647859573, + 0.03243512287735939, + 0.046480778604745865, + 0.015197807922959328, + 0.01601659320294857, + 0.013480511493980885, + 0.027357211336493492, + 0.08619561046361923, + 0.018695266917347908, + 0.025678172707557678, + 0.1279420703649521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.026339460164308548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862393140792847 + ], + "opposite_logprob_tokens": [ + -1.2236239910125732 + ], + "logprob_reward": -0.34862393140792847, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34862393140792847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09321492910385132, + 0.049476251006126404, + 0.1472352296113968, + 0.05071480572223663, + 0.09580282121896744, + 0.057029444724321365, + 0.07394803315401077, + 0.05500679463148117, + 0.10956212133169174, + 0.2680095136165619 + ], + "credit_map": [ + 0.016594821587204933, + 0.007632145658135414, + 0.018836921080946922, + 0.009452150203287601, + 0.010021490044891834, + 0.011607815511524677, + 0.010224223136901855, + 0.008845360018312931, + 0.00655903248116374, + 0.004099261946976185, + 0.006094605661928654, + 0.00671588396653533, + 0.006048858631402254, + 0.005882323253899813, + 0.006914420519024134, + 0.007161869667470455, + 0.02826380729675293, + 0.013289734721183777, + 0.02793232351541519, + 0.013666282407939434, + 0.015627659857273102, + 0.028400473296642303, + 0.00978066399693489, + 0.010274287313222885, + 0.006518188398331404, + 0.005304849240928888, + 0.008854352869093418, + 0.007747448515146971, + 0.005434536840766668, + 0.007716099265962839, + 0.0035685820039361715, + 0.00557075347751379, + 0.008175162598490715, + 0.015711059793829918, + 0.013878822326660156, + 0.01069960929453373, + 0.013377033174037933, + 0.009104754775762558, + 0.010255575180053711, + 0.014600804075598717, + 0.00684210704639554, + 0.009935845620930195, + 0.008488142862915993, + 0.006066066212952137, + 0.008582438342273235, + 0.00654857512563467, + 0.005143179092556238, + 0.005423093214631081, + 0.006058416794985533, + 0.007339971140027046, + 0.019779138267040253, + 0.008371196687221527, + 0.009674062952399254, + 0.01009424589574337, + 0.0061530196107923985, + 0.006477984134107828, + 0.0063934968784451485, + 0.009249185211956501, + 0.005759290419518948, + 0.005718114320188761, + 0.005445560440421104, + 0.00684016477316618, + 0.006941736675798893, + 0.0086592435836792, + 0.012367711402475834, + 0.009522299282252789, + 0.011517376638948917, + 0.01193898543715477, + 0.01416915562003851, + 0.019088009372353554, + 0.012969574891030788, + 0.017989009618759155, + 0.011871934868395329, + 0.01234995387494564, + 0.010399993509054184, + 0.022216549143195152, + 0.0688512995839119, + 0.015841437503695488, + 0.02146361768245697, + 0.10501473397016525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.34862393140792847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle crane shot begins, slowly zooming in on a classic 8-bit retro game world, unfolding in a bright, shadowless, flat lighting. In the center of the frame, a massive Armand pine tree, composed of dark brown and green pixel blocks, its crown an irregular, jagged umbrella-like outline, is used as a battering ram, violently ramming into the walls of a magnificent castle constructed from gray and dark gray pixel bricks. With each heavy, rhythmic impact, the tree's pixelated trunk trembles violently from the immense recoil, while the castle's stone walls erupt in a cloud of shattered pixels, creating a splattering gray \"debris\" while several black pixel cracks rapidly spread across the wall. In the background, a pure, flat blue-blue sky is framed by a few simple, white, blocky clouds, motionless. As the camera steadily and calmly approaches the chaotic siege scene, the audience can more clearly see the details of the destruction around the impact point. This slow advancement contrasts sharply with the fierce destruction, creating an epic siege atmosphere that is both grand and tense in the limited and vibrant retro tones.", + "seed": 5779, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_12/index_8104_seed_5779_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0430844459682703, + "vlm_logprob": -0.0430844459682703, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.0430844459682703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040113236755132675 + ], + "opposite_logprob_tokens": [ + -7.129011154174805 + ], + "logprob_reward": -0.0040113236755132675, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0040113236755132675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33977940678596497, + 0.07770142704248428, + 0.03945399075746536, + 0.048512205481529236, + 0.049755167216062546, + 0.11074430495500565, + 0.06698009371757507, + 0.05603810027241707, + 0.09537463635206223, + 0.11566060781478882 + ], + "credit_map": [ + 0.040773145854473114, + 0.03793821111321449, + 0.04044599086046219, + 0.03570273891091347, + 0.05097152665257454, + 0.06514576077461243, + 0.03719158098101616, + 0.03161047771573067, + 0.011564904823899269, + 0.011019114404916763, + 0.007236413192003965, + 0.007328335661441088, + 0.007800159510225058, + 0.00971873477101326, + 0.01216641440987587, + 0.01086735725402832, + 0.0033969173673540354, + 0.008751267567276955, + 0.004014237783849239, + 0.005593163426965475, + 0.004860905464738607, + 0.0046419380232691765, + 0.003997428808361292, + 0.004198136739432812, + 0.005775650031864643, + 0.0047532073222100735, + 0.004882023204118013, + 0.005602718330919743, + 0.004056120757013559, + 0.009809939190745354, + 0.003931490238755941, + 0.009701061062514782, + 0.0035136500373482704, + 0.009184369817376137, + 0.008240919560194016, + 0.007148071192204952, + 0.005942716728895903, + 0.0037703928537666798, + 0.005574002396315336, + 0.006381047423928976, + 0.008344234898686409, + 0.013570450246334076, + 0.024589436128735542, + 0.014875595457851887, + 0.013514555990695953, + 0.008897497318685055, + 0.013920928351581097, + 0.013031617738306522, + 0.005329802166670561, + 0.008744727820158005, + 0.009996757842600346, + 0.008884305134415627, + 0.012282710522413254, + 0.0063798557966947556, + 0.0068558319471776485, + 0.008506108075380325, + 0.006946226581931114, + 0.007665889337658882, + 0.007905187085270882, + 0.0060469587333500385, + 0.006930842995643616, + 0.0073119644075632095, + 0.00542989419773221, + 0.007801143452525139, + 0.008257738314568996, + 0.02048877626657486, + 0.016147548332810402, + 0.012297808192670345, + 0.009246664121747017, + 0.00992442574352026, + 0.009436455555260181, + 0.009575232863426208, + 0.010936862789094448, + 0.020876578986644745, + 0.014893402345478535, + 0.010954643599689007, + 0.012800831347703934, + 0.012205183506011963, + 0.01959989219903946, + 0.013393227942287922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0040113236755132675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299698054790497 + ], + "opposite_logprob_tokens": [ + -2.237997055053711 + ], + "logprob_reward": -0.11299698054790497, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299698054790497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4454221725463867, + 0.12865345180034637, + 0.03947777673602104, + 0.05485095456242561, + 0.051928628236055374, + 0.040711887180805206, + 0.05966682359576225, + 0.03518373891711235, + 0.08320821076631546, + 0.060896385461091995 + ], + "credit_map": [ + 0.07140304893255234, + 0.03198232501745224, + 0.043660685420036316, + 0.04498317465186119, + 0.12512141466140747, + 0.07327017933130264, + 0.028651626780629158, + 0.02634965069591999, + 0.025174111127853394, + 0.007282314356416464, + 0.009418861009180546, + 0.006810246501117945, + 0.016966087743639946, + 0.03478146344423294, + 0.011964427307248116, + 0.016255926340818405, + 0.0058800526894629, + 0.005326250568032265, + 0.005758081562817097, + 0.004416019655764103, + 0.0040826392360031605, + 0.003904518438503146, + 0.004226473160088062, + 0.005883736535906792, + 0.005554477218538523, + 0.003707688767462969, + 0.004780367016792297, + 0.006870821118354797, + 0.0036195518914610147, + 0.012150401249527931, + 0.004181812983006239, + 0.013985825702548027, + 0.0045225415378808975, + 0.008919911459088326, + 0.008256746456027031, + 0.00450004730373621, + 0.005523902364075184, + 0.008067921735346317, + 0.005571469198912382, + 0.006566083990037441, + 0.00337586784735322, + 0.005370421335101128, + 0.008496124297380447, + 0.006540808826684952, + 0.004745196551084518, + 0.0032845146488398314, + 0.004411898087710142, + 0.004487049765884876, + 0.009738470427691936, + 0.009037503972649574, + 0.007863451726734638, + 0.006310185883194208, + 0.007654975634068251, + 0.007107407785952091, + 0.0054141622968018055, + 0.006540658883750439, + 0.0036457034293562174, + 0.005371575243771076, + 0.004770680796355009, + 0.005116597283631563, + 0.004599877633154392, + 0.003792300820350647, + 0.003726863767951727, + 0.0041601345874369144, + 0.00663044024258852, + 0.01986040361225605, + 0.015026618726551533, + 0.010659892112016678, + 0.007964011281728745, + 0.009183871559798717, + 0.00818045623600483, + 0.00570250628516078, + 0.005460710730403662, + 0.006979296449571848, + 0.01146294828504324, + 0.006452029105275869, + 0.007776632905006409, + 0.004714738577604294, + 0.009234032593667507, + 0.008815991692245007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.11299698054790497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073528677225113 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073528677225113, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073528677225113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18551050126552582, + 0.07627075165510178, + 0.06803887337446213, + 0.047942016273736954, + 0.08635946363210678, + 0.0635230764746666, + 0.12928356230258942, + 0.0800042450428009, + 0.17683793604373932, + 0.0862295925617218 + ], + "credit_map": [ + 0.02848769724369049, + 0.015217900276184082, + 0.01630643755197525, + 0.0139713604003191, + 0.04812858998775482, + 0.032186124473810196, + 0.01360293198376894, + 0.017609450966119766, + 0.008722209371626377, + 0.01228251587599516, + 0.00795357208698988, + 0.006962657906115055, + 0.012851230800151825, + 0.010215695016086102, + 0.007238433230668306, + 0.010044439695775509, + 0.005765789654105902, + 0.010542580857872963, + 0.008712931536138058, + 0.012097342871129513, + 0.007423941045999527, + 0.007441502995789051, + 0.007536981254816055, + 0.008517797105014324, + 0.005643341690301895, + 0.005010106135159731, + 0.004608724266290665, + 0.009100156836211681, + 0.0034076455049216747, + 0.007294692564755678, + 0.0046997833997011185, + 0.008177567273378372, + 0.007083998993039131, + 0.024176988750696182, + 0.007031439337879419, + 0.00739266537129879, + 0.013758241198956966, + 0.006006261333823204, + 0.006368738133460283, + 0.014541137032210827, + 0.004388168919831514, + 0.00930420309305191, + 0.010470551438629627, + 0.006276189815253019, + 0.01152872946113348, + 0.006632110569626093, + 0.005743452347815037, + 0.009179674088954926, + 0.008738304488360882, + 0.020129617303609848, + 0.02010253444314003, + 0.01925264671444893, + 0.018677519634366035, + 0.017439501360058784, + 0.012812084518373013, + 0.01213135290890932, + 0.006032439414411783, + 0.009282797574996948, + 0.013382086530327797, + 0.009882383979856968, + 0.010787520557641983, + 0.013004926033318043, + 0.01046019233763218, + 0.007171896751970053, + 0.012819677591323853, + 0.04011441767215729, + 0.03384633734822273, + 0.023460689932107925, + 0.014321737922728062, + 0.02274252101778984, + 0.014987964183092117, + 0.014544595032930374, + 0.009013901464641094, + 0.010951448231935501, + 0.013298481702804565, + 0.00773853063583374, + 0.010667077265679836, + 0.00844325590878725, + 0.012606071308255196, + 0.013510826975107193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.016073528677225113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864955693483353 + ], + "opposite_logprob_tokens": [ + -3.048649549484253 + ], + "logprob_reward": -0.04864955693483353, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864955693483353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10841203480958939, + 0.12173222750425339, + 0.027415044605731964, + 0.03350427374243736, + 0.08209791034460068, + 0.0580643005669117, + 0.06935416907072067, + 0.038674574345350266, + 0.19295087456703186, + 0.267794668674469 + ], + "credit_map": [ + 0.015511687844991684, + 0.008936343714594841, + 0.011378726921975613, + 0.008075391873717308, + 0.018199272453784943, + 0.01612664945423603, + 0.014201919548213482, + 0.015982048586010933, + 0.016413584351539612, + 0.009438217617571354, + 0.011436394415795803, + 0.009142468683421612, + 0.014888351783156395, + 0.022703377529978752, + 0.015521549619734287, + 0.022188279777765274, + 0.0029322057962417603, + 0.004691748414188623, + 0.0026844062376767397, + 0.003927925601601601, + 0.00256642559543252, + 0.0037990985438227654, + 0.003411076730117202, + 0.003402157686650753, + 0.0034521252382546663, + 0.005049782805144787, + 0.0038806062657386065, + 0.0058084218762815, + 0.0032052798196673393, + 0.004529337864369154, + 0.003029871266335249, + 0.004548848140984774, + 0.005920269526541233, + 0.020957784727215767, + 0.008745590224862099, + 0.00682926457375288, + 0.011877380311489105, + 0.007246688473969698, + 0.006572313141077757, + 0.013948622159659863, + 0.0041862125508487225, + 0.007375385612249374, + 0.011732468381524086, + 0.008828739635646343, + 0.00894028041511774, + 0.00468948669731617, + 0.006659055128693581, + 0.005652672611176968, + 0.008345244452357292, + 0.01107943244278431, + 0.008279644884169102, + 0.012051090598106384, + 0.009167565032839775, + 0.006147170905023813, + 0.00686578918248415, + 0.007418233435600996, + 0.003429990028962493, + 0.004611165262758732, + 0.008384644985198975, + 0.003963736351579428, + 0.00459982780739665, + 0.0054884715937078, + 0.0035978800151497126, + 0.004598855506628752, + 0.01787525974214077, + 0.04293976351618767, + 0.03135025128722191, + 0.02154848910868168, + 0.023836681619286537, + 0.022376898676156998, + 0.016060257330536842, + 0.016963262110948563, + 0.027299880981445312, + 0.03746035322546959, + 0.03815300017595291, + 0.029501289129257202, + 0.028750209137797356, + 0.02334459125995636, + 0.04183026775717735, + 0.041455067694187164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.04864955693483353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033690840005874634 + ], + "opposite_logprob_tokens": [ + -3.408690929412842 + ], + "logprob_reward": -0.033690840005874634, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033690840005874634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12560920417308807, + 0.06223654747009277, + 0.08197106420993805, + 0.06997755914926529, + 0.06732214242219925, + 0.16220629215240479, + 0.18601801991462708, + 0.10257843136787415, + 0.10816879570484161, + 0.03391198068857193 + ], + "credit_map": [ + 0.016826236620545387, + 0.008819037117064, + 0.010630639269948006, + 0.018102293834090233, + 0.032267920672893524, + 0.021638566628098488, + 0.008867721073329449, + 0.008456791751086712, + 0.005860067903995514, + 0.007985499687492847, + 0.007930101826786995, + 0.0040997848846018314, + 0.010274553671479225, + 0.01152642909437418, + 0.006707138381898403, + 0.007852968759834766, + 0.009802237153053284, + 0.011709317564964294, + 0.006958789192140102, + 0.010249268263578415, + 0.008419872261583805, + 0.01351118553429842, + 0.009718777611851692, + 0.011601620353758335, + 0.005646000616252422, + 0.005358721595257521, + 0.005197963211685419, + 0.006567902863025665, + 0.003912011161446571, + 0.01678144372999668, + 0.0052551827393472195, + 0.021258335560560226, + 0.0039001349359750748, + 0.010188668966293335, + 0.014948293566703796, + 0.009592203423380852, + 0.006737742573022842, + 0.007544018793851137, + 0.007594461552798748, + 0.006816620472818613, + 0.011200902983546257, + 0.023038456216454506, + 0.03255968913435936, + 0.02178877405822277, + 0.01830604113638401, + 0.015871815383434296, + 0.015547207556664944, + 0.02389340288937092, + 0.02064635418355465, + 0.02245355397462845, + 0.03170548379421234, + 0.021519452333450317, + 0.035550471395254135, + 0.018199708312749863, + 0.01660224236547947, + 0.019340762868523598, + 0.00851844809949398, + 0.011718492954969406, + 0.021361805498600006, + 0.014687852934002876, + 0.012437163852155209, + 0.013020996935665607, + 0.010924936272203922, + 0.009908735752105713, + 0.011389065533876419, + 0.020138150081038475, + 0.01659012958407402, + 0.011694182641804218, + 0.014362853951752186, + 0.013652736321091652, + 0.01052303146570921, + 0.00981865357607603, + 0.003061937401071191, + 0.004620551597326994, + 0.0051218741573393345, + 0.0035551991313695908, + 0.003700891975313425, + 0.0034793249797075987, + 0.005141804460436106, + 0.0052303955890238285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.033690840005874634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle crane shot begins, slowly zooming in on a classic 8-bit retro game world, unfolding in a bright, shadowless, flat lighting. In the center of the frame, a massive Armand pine tree, composed of dark brown and green pixel blocks, its crown an irregular, jagged umbrella-like outline, is used as a battering ram, violently ramming into the walls of a magnificent castle constructed from gray and dark gray pixel bricks. With each heavy, rhythmic impact, the tree's pixelated trunk trembles violently from the immense recoil, while the castle's stone walls erupt in a cloud of shattered pixels, creating a splattering gray \"debris\" while several black pixel cracks rapidly spread across the wall. In the background, a pure, flat blue-blue sky is framed by a few simple, white, blocky clouds, motionless. As the camera steadily and calmly approaches the chaotic siege scene, the audience can more clearly see the details of the destruction around the impact point. This slow advancement contrasts sharply with the fierce destruction, creating an epic siege atmosphere that is both grand and tense in the limited and vibrant retro tones.", + "seed": 5780, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_12/index_8104_seed_5780_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21325991936028005, + "vlm_logprob": -0.21325991936028005, + "vlm_yes_no_margin": 2.499999988079071, + "vlm_token_credit": -0.21325991936028005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004351789131760597 + ], + "opposite_logprob_tokens": [ + -6.504351615905762 + ], + "logprob_reward": -0.004351789131760597, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.004351789131760597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09540166705846786, + 0.06662362813949585, + 0.10394962131977081, + 0.15189528465270996, + 0.07638656347990036, + 0.14835883677005768, + 0.0940345972776413, + 0.0905148983001709, + 0.0684627816081047, + 0.10437210649251938 + ], + "credit_map": [ + 0.019390463829040527, + 0.006763801909983158, + 0.012186137959361076, + 0.010378537699580193, + 0.009634657762944698, + 0.016007352620363235, + 0.008278733119368553, + 0.012761988677084446, + 0.011500058695673943, + 0.004777001682668924, + 0.009489323012530804, + 0.008115911856293678, + 0.007253960706293583, + 0.013357643038034439, + 0.00802613701671362, + 0.004103591199964285, + 0.025195175781846046, + 0.005984356626868248, + 0.009531067684292793, + 0.008110973984003067, + 0.006394188851118088, + 0.033444102853536606, + 0.006727532483637333, + 0.008562223985791206, + 0.02385062910616398, + 0.017588777467608452, + 0.01888316683471203, + 0.018628602847456932, + 0.017748447135090828, + 0.021914973855018616, + 0.01832774095237255, + 0.014952946454286575, + 0.009483607485890388, + 0.008267238736152649, + 0.007368060760200024, + 0.01107733603566885, + 0.009840586222708225, + 0.017169257625937462, + 0.007920313626527786, + 0.00526016391813755, + 0.01011983398348093, + 0.008852794766426086, + 0.006622356828302145, + 0.005514318123459816, + 0.018141280859708786, + 0.05370877683162689, + 0.008139977231621742, + 0.03725950047373772, + 0.006615620106458664, + 0.028074705973267555, + 0.006292796693742275, + 0.010989692993462086, + 0.02160496637225151, + 0.007440127898007631, + 0.007485811598598957, + 0.005530873779207468, + 0.024366337805986404, + 0.01242540217936039, + 0.006312264129519463, + 0.007511617615818977, + 0.017680097371339798, + 0.011174282059073448, + 0.004538671113550663, + 0.006506221368908882, + 0.00833750981837511, + 0.011784568428993225, + 0.0049535115249454975, + 0.009568518027663231, + 0.01364342775195837, + 0.00916491448879242, + 0.005878273397684097, + 0.005132060497999191, + 0.010876037180423737, + 0.018443528562784195, + 0.004233237821608782, + 0.02677220292389393, + 0.012601971626281738, + 0.01597408577799797, + 0.00725130271166563, + 0.008219738490879536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.004351789131760597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814243733882904 + ], + "opposite_logprob_tokens": [ + -1.4064244031906128 + ], + "logprob_reward": -0.2814243733882904, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814243733882904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08866624534130096, + 0.07778473943471909, + 0.03574086353182793, + 0.08443918079137802, + 0.04352958872914314, + 0.0506502166390419, + 0.047016311436891556, + 0.18930213153362274, + 0.08220451325178146, + 0.3006662130355835 + ], + "credit_map": [ + 0.011989274993538857, + 0.00988420844078064, + 0.012308673933148384, + 0.012213743291795254, + 0.008684936910867691, + 0.010849574580788612, + 0.01120123639702797, + 0.011534605175256729, + 0.015588830225169659, + 0.005961990915238857, + 0.013274425640702248, + 0.006710968445986509, + 0.008185706101357937, + 0.01644187420606613, + 0.0057556708343327045, + 0.0058652781881392, + 0.007998916320502758, + 0.002653385978192091, + 0.0028652369510382414, + 0.003444124711677432, + 0.002937852870672941, + 0.010541288182139397, + 0.003051101928576827, + 0.0022489563561975956, + 0.012258934788405895, + 0.010470571927726269, + 0.010514508001506329, + 0.012356228195130825, + 0.011199526488780975, + 0.008966186083853245, + 0.010433628223836422, + 0.008239606395363808, + 0.0051793865859508514, + 0.006666931789368391, + 0.005052331835031509, + 0.007178573869168758, + 0.004988616332411766, + 0.005891205742955208, + 0.005433408543467522, + 0.003139139385893941, + 0.0036861777771264315, + 0.00348104489967227, + 0.0018914525862783194, + 0.0022549915593117476, + 0.00676290038973093, + 0.01898934319615364, + 0.002557035768404603, + 0.011027274653315544, + 0.0035606911405920982, + 0.01098916307091713, + 0.0047287847846746445, + 0.007387679070234299, + 0.008052048273384571, + 0.0043021077290177345, + 0.00489437859505415, + 0.003101460402831435, + 0.05760481208562851, + 0.02458108775317669, + 0.007465420290827751, + 0.010172398760914803, + 0.06494131684303284, + 0.01219165325164795, + 0.005870879162102938, + 0.006474574562162161, + 0.01546743419021368, + 0.015942120924592018, + 0.003799783531576395, + 0.007306599989533424, + 0.016655391082167625, + 0.013075117953121662, + 0.004545971751213074, + 0.005412096157670021, + 0.035039592534303665, + 0.050884753465652466, + 0.014791336841881275, + 0.05783203989267349, + 0.044614773243665695, + 0.04442127048969269, + 0.02448137290775776, + 0.02860109694302082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.2814243733882904 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892223596572876 + ], + "opposite_logprob_tokens": [ + -1.053922176361084 + ], + "logprob_reward": -0.42892223596572876, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42892223596572876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18562941253185272, + 0.1998511701822281, + 0.03806036338210106, + 0.05052943900227547, + 0.08768945187330246, + 0.036602769047021866, + 0.05082795396447182, + 0.075588159263134, + 0.09479589015245438, + 0.18042539060115814 + ], + "credit_map": [ + 0.02448214404284954, + 0.016332140192389488, + 0.01551182009279728, + 0.030902551487088203, + 0.027363073080778122, + 0.02283376269042492, + 0.02660425752401352, + 0.021599676460027695, + 0.037510260939598083, + 0.01537749171257019, + 0.028687858954072, + 0.023783646523952484, + 0.019929243251681328, + 0.041807036846876144, + 0.016860533505678177, + 0.015895092859864235, + 0.006409596186131239, + 0.0030501203145831823, + 0.003231300041079521, + 0.004507599864155054, + 0.0040453774854540825, + 0.009342926554381847, + 0.004810084123164415, + 0.0026633592788130045, + 0.006136239971965551, + 0.006544426083564758, + 0.006187634076923132, + 0.008103841915726662, + 0.00727654667571187, + 0.0047831241972744465, + 0.006835802923887968, + 0.004661827348172665, + 0.009355420246720314, + 0.010659296065568924, + 0.008986590430140495, + 0.015788672491908073, + 0.011160769499838352, + 0.014230095781385899, + 0.011367960833013058, + 0.006140649318695068, + 0.005724107846617699, + 0.0040112086571753025, + 0.0034967081155627966, + 0.002458771923556924, + 0.005198623053729534, + 0.003837883472442627, + 0.0026643199380487204, + 0.009211144410073757, + 0.004432199988514185, + 0.01395152136683464, + 0.0033171612303704023, + 0.00805431418120861, + 0.00897749699652195, + 0.004214011132717133, + 0.004389691166579723, + 0.0034915555734187365, + 0.01897434890270233, + 0.008794838562607765, + 0.005454236641526222, + 0.006851299200206995, + 0.014607909135520458, + 0.007998387329280376, + 0.006015140563249588, + 0.006891999393701553, + 0.010636728256940842, + 0.014568059705197811, + 0.007452299352735281, + 0.01388526987284422, + 0.01848718896508217, + 0.011906660161912441, + 0.010685073211789131, + 0.007174608297646046, + 0.021849213168025017, + 0.03478320315480232, + 0.008316567167639732, + 0.026726633310317993, + 0.026610342785716057, + 0.028383230790495872, + 0.010561104863882065, + 0.023195097222924232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.42892223596572876 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135010302066803 + ], + "opposite_logprob_tokens": [ + -1.313501000404358 + ], + "logprob_reward": -0.3135010302066803, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135010302066803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08094194531440735, + 0.05276695266366005, + 0.03450732305645943, + 0.0351259671151638, + 0.06702535599470139, + 0.12428971379995346, + 0.06449344754219055, + 0.231853649020195, + 0.053542643785476685, + 0.2554529905319214 + ], + "credit_map": [ + 0.011197644285857677, + 0.00775130232796073, + 0.010009058751165867, + 0.011188105680048466, + 0.012875020503997803, + 0.007556075695902109, + 0.010926183313131332, + 0.009438551031053066, + 0.008866311982274055, + 0.004314834251999855, + 0.005253057461231947, + 0.008084411732852459, + 0.005846028216183186, + 0.010214661248028278, + 0.0063409036956727505, + 0.003846744541078806, + 0.007265659514814615, + 0.0022105989046394825, + 0.002364183310419321, + 0.004107806831598282, + 0.002378919394686818, + 0.010572276078164577, + 0.0036268015392124653, + 0.0019810821395367384, + 0.0049749319441616535, + 0.003965133335441351, + 0.004522050730884075, + 0.004566705320030451, + 0.004110782872885466, + 0.004657522775232792, + 0.004657695535570383, + 0.0036711429711431265, + 0.008072220720350742, + 0.0100346514955163, + 0.007587085012346506, + 0.010213945992290974, + 0.007814398035407066, + 0.011154147796332836, + 0.0076283738017082214, + 0.004520535469055176, + 0.009575285948812962, + 0.012076838873326778, + 0.008989090099930763, + 0.005209111608564854, + 0.00896932277828455, + 0.03450547531247139, + 0.007231610827147961, + 0.03773297369480133, + 0.005579933989793062, + 0.01853087544441223, + 0.004275637213140726, + 0.010321886278688908, + 0.010351735167205334, + 0.0062405928038060665, + 0.0058282166719436646, + 0.0033645699732005596, + 0.06961117684841156, + 0.03313535824418068, + 0.008180888369679451, + 0.009796330705285072, + 0.08496420830488205, + 0.013290616683661938, + 0.005685377866029739, + 0.007189694792032242, + 0.006726931780576706, + 0.00847786944359541, + 0.0037758895196020603, + 0.007498424965888262, + 0.011040878482162952, + 0.008181552402675152, + 0.004683625418692827, + 0.0031574731692671776, + 0.03051503747701645, + 0.05055444687604904, + 0.01295552309602499, + 0.04658983275294304, + 0.033808398991823196, + 0.04169603809714317, + 0.016378289088606834, + 0.022955432534217834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.3135010302066803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in an 8-bit retro game style with pixelated graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting bright, shadowless, and flat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive Armand pine tree made of dark brown and green pixel blocks ramming into a castle wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the castle wall erupt in a cloud of shattered gray pixels and black cracks upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle crane shot and slowly zoom in steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038100168108940125 + ], + "opposite_logprob_tokens": [ + -3.288100242614746 + ], + "logprob_reward": -0.038100168108940125, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038100168108940125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12427613884210587, + 0.059031058102846146, + 0.06373855471611023, + 0.04893890395760536, + 0.05147797614336014, + 0.13847173750400543, + 0.16152353584766388, + 0.18217740952968597, + 0.04804699495434761, + 0.12231773138046265 + ], + "credit_map": [ + 0.018945736810564995, + 0.010835802182555199, + 0.016936110332608223, + 0.013647577725350857, + 0.014625207521021366, + 0.017088694497942924, + 0.015513550490140915, + 0.01668345369398594, + 0.009822615422308445, + 0.0039144582115113735, + 0.006344641093164682, + 0.007906344719231129, + 0.00526337930932641, + 0.013083985075354576, + 0.008640353567898273, + 0.004055274184793234, + 0.015747612342238426, + 0.004431434907019138, + 0.0037206499837338924, + 0.006223716773092747, + 0.004743389785289764, + 0.019597550854086876, + 0.0055891950614750385, + 0.003684994764626026, + 0.005885006859898567, + 0.005195489153265953, + 0.006934348028153181, + 0.007940029725432396, + 0.007629298605024815, + 0.004534515552222729, + 0.006427292712032795, + 0.004392918664962053, + 0.0059067485854029655, + 0.008615242317318916, + 0.005344504490494728, + 0.00814839731901884, + 0.0063903010450303555, + 0.005734181962907314, + 0.007540008053183556, + 0.003798587014898658, + 0.011706686578691006, + 0.026268387213349342, + 0.0071567147970199585, + 0.010839797556400299, + 0.04423723742365837, + 0.017097804695367813, + 0.012239105999469757, + 0.008925987407565117, + 0.015093350782990456, + 0.03571108728647232, + 0.01872018538415432, + 0.0244502741843462, + 0.02484765276312828, + 0.018358919769525528, + 0.013999313116073608, + 0.010342731140553951, + 0.04585308954119682, + 0.025017941370606422, + 0.011585417203605175, + 0.009869570843875408, + 0.05853261053562164, + 0.010745169594883919, + 0.007065435405820608, + 0.013508159667253494, + 0.005479843821376562, + 0.007902338169515133, + 0.00445616664364934, + 0.007940149866044521, + 0.007146258372813463, + 0.006304279901087284, + 0.0042406716383993626, + 0.004577277693897486, + 0.014232299290597439, + 0.02076486125588417, + 0.009201571345329285, + 0.02259143441915512, + 0.014263110235333443, + 0.02005380019545555, + 0.008971958421170712, + 0.012238695286214352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.038100168108940125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_tensor([2289, 2289])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_tensor([2289, 2289])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bcd18d45456f15247d0657b8bbaa9ea3c469bf12 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_tensor([2289, 2289])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a medium shot, shot from an upward angle, then slowly and steadily pulls back, ultimately framing a sweeping panorama of grandeur. The camera initially focuses on a majestic, middle-aged emperor, his figure as straight as a pine tree, standing alone at the edge of a high platform carved from massive black stone. He clad in a xuan (a deep, solemn black) silk dragon robe, intricately embroidered with dazzling gold thread, a lifelike, five-clawed golden dragon, its fangs bared and claws bared. The dragon's scales reflected a subtle metallic luster in the dim light. His long hair was held high in a pristine jade crown. His features were sharply defined, and his deep eyes shone with a piercing, torch-like light. As the camera zoomed out, he slowly turned his head with the air of a king, his gaze scanning the area below. At this moment, a breathtaking spectacle was fully revealed: beneath the platform, within a vast underground palace, thousands of terracotta warriors and horses stood in meticulous formation. Formations of infantry, archers, and charioteers stretched into the endless darkness as far as the eye could see. Each life-size figure, solemn and silent, exuded the majestic aura of war. The entire scene was illuminated by flickering torches hung high on the walls. These torches cast long, swaying shadows, creating a stark contrast of light and dark on the emperor's face and the warriors' armor, creating an epic atmosphere of solemnity, power, and eternity.", + "seed": 6621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_2289_seed_6621_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025102166971191765, + "vlm_logprob": -0.025102166971191765, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.025102166971191765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055403731763362885 + ], + "opposite_logprob_tokens": [ + -3.055403709411621 + ], + "logprob_reward": -0.055403731763362885, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.055403731763362885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13249996304512024, + 0.21419641375541687, + 0.08895851671695709, + 0.1547752022743225, + 0.060490358620882034, + 0.07682283222675323, + 0.13744686543941498, + 0.03400276228785515, + 0.03583309054374695, + 0.06497403234243393 + ], + "credit_map": [ + 0.019792061299085617, + 0.01134423166513443, + 0.019152475520968437, + 0.02252919040620327, + 0.0099949324503541, + 0.01496628113090992, + 0.014112944714725018, + 0.02060784585773945, + 0.018918476998806, + 0.014274968765676022, + 0.04892979934811592, + 0.05040190741419792, + 0.010252173990011215, + 0.04391513019800186, + 0.019328147172927856, + 0.008175790309906006, + 0.015009202994406223, + 0.008452360518276691, + 0.02292044088244438, + 0.007801389787346125, + 0.006734629161655903, + 0.013981983996927738, + 0.007045819889754057, + 0.00701267970725894, + 0.032526832073926926, + 0.020312106236815453, + 0.042022813111543655, + 0.013624941930174828, + 0.008714222349226475, + 0.023014666512608528, + 0.006930639501661062, + 0.007628967519849539, + 0.010864739306271076, + 0.006300258915871382, + 0.008873211219906807, + 0.0055693769827485085, + 0.009495481848716736, + 0.007261803839355707, + 0.006044686771929264, + 0.006080796476453543, + 0.010707419365644455, + 0.010791441425681114, + 0.01357799768447876, + 0.007577952463179827, + 0.008878018707036972, + 0.010523563250899315, + 0.006517309695482254, + 0.008249117061495781, + 0.020996904000639915, + 0.015041803009808064, + 0.015364213846623898, + 0.016188062727451324, + 0.016373544931411743, + 0.02209559641778469, + 0.014840682037174702, + 0.01654605194926262, + 0.003448616713285446, + 0.007067821454256773, + 0.0033193689305335283, + 0.0031807771883904934, + 0.0032950264867395163, + 0.005218366160988808, + 0.0029407546389847994, + 0.005532030947506428, + 0.0035682516172528267, + 0.006865439470857382, + 0.003035931149497628, + 0.005053081549704075, + 0.0053624482825398445, + 0.005076549481600523, + 0.004256596788764, + 0.002614791039377451, + 0.007519392762333155, + 0.015580416657030582, + 0.012042565271258354, + 0.005299699027091265, + 0.005194280296564102, + 0.007872904650866985, + 0.005498016253113747, + 0.005966748110949993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.055403731763362885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360128074884415 + ], + "opposite_logprob_tokens": [ + -3.651360034942627 + ], + "logprob_reward": -0.026360128074884415, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360128074884415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05267925187945366, + 0.4694421887397766, + 0.02057802490890026, + 0.07161567360162735, + 0.04555871710181236, + 0.05659943073987961, + 0.0898766964673996, + 0.05930536985397339, + 0.06519995629787445, + 0.06914468854665756 + ], + "credit_map": [ + 0.006947261281311512, + 0.003559596138074994, + 0.010022895410656929, + 0.009853378869593143, + 0.003391377395018935, + 0.007202863693237305, + 0.004684291314333677, + 0.0070175910368561745, + 0.04348519816994667, + 0.03684106841683388, + 0.07807231694459915, + 0.14685645699501038, + 0.024999350309371948, + 0.06453465670347214, + 0.05483408644795418, + 0.019819071516394615, + 0.002825828269124031, + 0.00290177995339036, + 0.0035742090549319983, + 0.0028201036620885134, + 0.00191378069575876, + 0.0025286644231528044, + 0.001864453312009573, + 0.0021492047235369682, + 0.011770902201533318, + 0.007062873803079128, + 0.02133060246706009, + 0.007898352108895779, + 0.005079910159111023, + 0.010478598065674305, + 0.005245777778327465, + 0.002748657250776887, + 0.00904475711286068, + 0.005961981602013111, + 0.007784719578921795, + 0.005561004392802715, + 0.005745905917137861, + 0.005001752637326717, + 0.003902634372934699, + 0.002555962884798646, + 0.004802556708455086, + 0.006841725204139948, + 0.007407801225781441, + 0.009681078605353832, + 0.006472749635577202, + 0.00768184382468462, + 0.009192374534904957, + 0.004519299603998661, + 0.012707836925983429, + 0.01155914831906557, + 0.01456857193261385, + 0.010055377148091793, + 0.0069947014562785625, + 0.01587659865617752, + 0.007538598030805588, + 0.01057586632668972, + 0.007822085171937943, + 0.011062359437346458, + 0.005827987566590309, + 0.005734330043196678, + 0.005885067395865917, + 0.009147102944552898, + 0.004655459430068731, + 0.009170977398753166, + 0.009768323972821236, + 0.008602186106145382, + 0.011535347439348698, + 0.008040535263717175, + 0.006513850297778845, + 0.008300858549773693, + 0.005673909559845924, + 0.006764946971088648, + 0.009301308542490005, + 0.017934832721948624, + 0.012360001914203167, + 0.004628530237823725, + 0.005297591909766197, + 0.010008730925619602, + 0.0035667705815285444, + 0.006046921480447054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.026360128074884415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032025498803704977 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.0032025498803704977, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032025498803704977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0517461933195591, + 0.4060259163379669, + 0.029172183945775032, + 0.13842885196208954, + 0.026127416640520096, + 0.02967548370361328, + 0.035235483199357986, + 0.10974006354808807, + 0.07897160947322845, + 0.09487677365541458 + ], + "credit_map": [ + 0.009643751196563244, + 0.004665405955165625, + 0.00832044705748558, + 0.007592794485390186, + 0.003572700312361121, + 0.0065994225442409515, + 0.004203367047011852, + 0.0071483091451227665, + 0.03782731667160988, + 0.0313093401491642, + 0.06650025397539139, + 0.13124799728393555, + 0.021147610619664192, + 0.05240132287144661, + 0.04752342402935028, + 0.018068702891469002, + 0.005863654892891645, + 0.0027510435320436954, + 0.005684365052729845, + 0.0025567365810275078, + 0.002753322711214423, + 0.004392401315271854, + 0.001939214882440865, + 0.0032314483541995287, + 0.04375482350587845, + 0.01596454717218876, + 0.022070033475756645, + 0.011977865360677242, + 0.0091309929266572, + 0.022117938846349716, + 0.007238045800477266, + 0.006174622103571892, + 0.0034638980869203806, + 0.0040570758283138275, + 0.0032663613092154264, + 0.003975530155003071, + 0.0029778017196804285, + 0.003025041427463293, + 0.0036099092103540897, + 0.0017518011154606938, + 0.003165960544720292, + 0.004073453601449728, + 0.0034938380122184753, + 0.004961603321135044, + 0.0030403912533074617, + 0.004115825518965721, + 0.0043274471536278725, + 0.002496967790648341, + 0.002962925471365452, + 0.005943348165601492, + 0.002560934517532587, + 0.004885014146566391, + 0.005083322990685701, + 0.006037137471139431, + 0.004425900988280773, + 0.0033369010780006647, + 0.012988402508199215, + 0.021342789754271507, + 0.01450994610786438, + 0.013586877845227718, + 0.008445185609161854, + 0.017710138112306595, + 0.008179537951946259, + 0.012977194972336292, + 0.012007824145257473, + 0.012686992064118385, + 0.010778614319860935, + 0.007047825492918491, + 0.00886487029492855, + 0.010252678766846657, + 0.008085083216428757, + 0.00924772396683693, + 0.012602288275957108, + 0.013153629377484322, + 0.01838458515703678, + 0.010033861733973026, + 0.005754808429628611, + 0.011334405280649662, + 0.00930853746831417, + 0.014304662123322487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.0032025498803704977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197533018887043 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197533018887043, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197533018887043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0748588964343071, + 0.1106419712305069, + 0.03361779451370239, + 0.10566216707229614, + 0.045071665197610855, + 0.05774324759840965, + 0.17669495940208435, + 0.16658855974674225, + 0.10243531316518784, + 0.12668544054031372 + ], + "credit_map": [ + 0.013493011705577374, + 0.005068305879831314, + 0.013047107495367527, + 0.011263294145464897, + 0.005603041034191847, + 0.009988129138946533, + 0.005130697973072529, + 0.01126530859619379, + 0.010097210295498371, + 0.009220615960657597, + 0.017693765461444855, + 0.036035388708114624, + 0.006669618654996157, + 0.01341030653566122, + 0.012346705421805382, + 0.005168354604393244, + 0.005022639408707619, + 0.003752304706722498, + 0.008984506130218506, + 0.003996262326836586, + 0.0022476844023913145, + 0.005554649978876114, + 0.0019711945205926895, + 0.002088551875203848, + 0.02487458474934101, + 0.010295972228050232, + 0.02641826868057251, + 0.010041228495538235, + 0.008405540138483047, + 0.015012209303677082, + 0.006273060105741024, + 0.0043413047678768635, + 0.008653059601783752, + 0.006847934797406197, + 0.0053315781988203526, + 0.004172405693680048, + 0.006322405766695738, + 0.004768076352775097, + 0.004396502859890461, + 0.004579705651849508, + 0.007602323777973652, + 0.008937118574976921, + 0.006977478042244911, + 0.008464384824037552, + 0.005723627284169197, + 0.008347416296601295, + 0.006691692862659693, + 0.004999204073101282, + 0.03415031358599663, + 0.0196105744689703, + 0.017539167776703835, + 0.023787416517734528, + 0.01579880528151989, + 0.028494251891970634, + 0.016681956127285957, + 0.020632479339838028, + 0.017550742253661156, + 0.03207709267735481, + 0.017154302448034286, + 0.02413933537900448, + 0.016777025535702705, + 0.024289065971970558, + 0.014765040017664433, + 0.019835952669382095, + 0.017511887475848198, + 0.01509767398238182, + 0.015620021149516106, + 0.009666496887803078, + 0.010354750789701939, + 0.01479460671544075, + 0.008371252566576004, + 0.011018624529242516, + 0.015784448012709618, + 0.023135365918278694, + 0.028639206662774086, + 0.010847126133739948, + 0.007452770136296749, + 0.017720161005854607, + 0.01066743116825819, + 0.012438940815627575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.014197533018887043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634689211845398 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.02634689211845398, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634689211845398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24202801287174225, + 0.2931526303291321, + 0.03428879752755165, + 0.09108354896306992, + 0.04653138667345047, + 0.055845554918050766, + 0.038007427006959915, + 0.11019638180732727, + 0.03374942019581795, + 0.055116843432188034 + ], + "credit_map": [ + 0.04541223496198654, + 0.011901000514626503, + 0.03974218666553497, + 0.03378397598862648, + 0.013248253613710403, + 0.042459987103939056, + 0.010984085500240326, + 0.04449630528688431, + 0.027090100571513176, + 0.023501140996813774, + 0.046906132251024246, + 0.09678831696510315, + 0.016703367233276367, + 0.0360955186188221, + 0.03393206745386124, + 0.012136004865169525, + 0.006684114225208759, + 0.003565659746527672, + 0.007107765879482031, + 0.004608308430761099, + 0.002212752355262637, + 0.005028461571782827, + 0.0022549168206751347, + 0.0028268166352063417, + 0.024627599865198135, + 0.010674716904759407, + 0.018166789785027504, + 0.008441473357379436, + 0.005889543332159519, + 0.01348214689642191, + 0.004482274875044823, + 0.005318999756127596, + 0.007943281903862953, + 0.0050307284109294415, + 0.0067779431119561195, + 0.004591469187289476, + 0.008036002516746521, + 0.005487356334924698, + 0.00465976819396019, + 0.004004836082458496, + 0.007403634488582611, + 0.007510968018323183, + 0.00847273413091898, + 0.007356138899922371, + 0.0059401593171060085, + 0.008007231168448925, + 0.005991392768919468, + 0.005163294728845358, + 0.007936390116810799, + 0.004540393594652414, + 0.003704106668010354, + 0.0050259800627827644, + 0.0041833180002868176, + 0.005841212347149849, + 0.0034787855111062527, + 0.0032972402404993773, + 0.010431414470076561, + 0.02870832569897175, + 0.008734673261642456, + 0.009799210354685783, + 0.011000802740454674, + 0.021057985723018646, + 0.007799288257956505, + 0.012664682231843472, + 0.003959802910685539, + 0.004678055644035339, + 0.005976047832518816, + 0.0035666250623762608, + 0.0023759384639561176, + 0.005234257318079472, + 0.0027631353586912155, + 0.0051955594681203365, + 0.006528330966830254, + 0.00841335766017437, + 0.010872671380639076, + 0.004414249211549759, + 0.004906314890831709, + 0.006164420861750841, + 0.005606652703136206, + 0.008210846222937107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.02634689211845398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, shot from an upward angle, then slowly and steadily pulls back, ultimately framing a sweeping panorama of grandeur. The camera initially focuses on a majestic, middle-aged emperor, his figure as straight as a pine tree, standing alone at the edge of a high platform carved from massive black stone. He clad in a xuan (a deep, solemn black) silk dragon robe, intricately embroidered with dazzling gold thread, a lifelike, five-clawed golden dragon, its fangs bared and claws bared. The dragon's scales reflected a subtle metallic luster in the dim light. His long hair was held high in a pristine jade crown. His features were sharply defined, and his deep eyes shone with a piercing, torch-like light. As the camera zoomed out, he slowly turned his head with the air of a king, his gaze scanning the area below. At this moment, a breathtaking spectacle was fully revealed: beneath the platform, within a vast underground palace, thousands of terracotta warriors and horses stood in meticulous formation. Formations of infantry, archers, and charioteers stretched into the endless darkness as far as the eye could see. Each life-size figure, solemn and silent, exuded the majestic aura of war. The entire scene was illuminated by flickering torches hung high on the walls. These torches cast long, swaying shadows, creating a stark contrast of light and dark on the emperor's face and the warriors' armor, creating an epic atmosphere of solemnity, power, and eternity.", + "seed": 6622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_2289_seed_6622_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05217599365860224, + "vlm_logprob": -0.05217599365860224, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05217599365860224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04832043871283531 + ], + "opposite_logprob_tokens": [ + -3.1733205318450928 + ], + "logprob_reward": -0.04832043871283531, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04832043871283531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25865697860717773, + 0.08210217207670212, + 0.029877685010433197, + 0.06296233832836151, + 0.15735307335853577, + 0.06191848963499069, + 0.07297097891569138, + 0.057716306298971176, + 0.10380769520998001, + 0.1126343235373497 + ], + "credit_map": [ + 0.03873181343078613, + 0.03937330096960068, + 0.019409120082855225, + 0.03410932049155235, + 0.027559930458664894, + 0.042327880859375, + 0.019461553543806076, + 0.03768404945731163, + 0.011554235592484474, + 0.008362436667084694, + 0.008017582818865776, + 0.009698848240077496, + 0.013323131017386913, + 0.014103777706623077, + 0.006333953235298395, + 0.010708210989832878, + 0.0048531703650951385, + 0.003116452367976308, + 0.0026497801300138235, + 0.005075093358755112, + 0.003691363614052534, + 0.004583572503179312, + 0.002510609570890665, + 0.0033976417034864426, + 0.008715683594346046, + 0.006496709305793047, + 0.007834179326891899, + 0.007399963680654764, + 0.008712014183402061, + 0.010860432870686054, + 0.005766596645116806, + 0.007176758721470833, + 0.0374901220202446, + 0.013442161493003368, + 0.015387600287795067, + 0.013679265975952148, + 0.039978597313165665, + 0.01240996178239584, + 0.01570415124297142, + 0.009261216036975384, + 0.012122252024710178, + 0.00504313874989748, + 0.01091675739735365, + 0.0062782252207398415, + 0.008826740086078644, + 0.0066463518887758255, + 0.005933594424277544, + 0.006151426117867231, + 0.010834593325853348, + 0.006324386689811945, + 0.011273420415818691, + 0.008020268753170967, + 0.007954778149724007, + 0.01243053562939167, + 0.005554293282330036, + 0.010578694753348827, + 0.011102844029664993, + 0.0071996524930000305, + 0.00740589527413249, + 0.005244392901659012, + 0.005483594723045826, + 0.007571230176836252, + 0.005523598752915859, + 0.008185095153748989, + 0.010735655203461647, + 0.009809372015297413, + 0.022824276238679886, + 0.009858755394816399, + 0.009966486133635044, + 0.01963975839316845, + 0.0075600589625537395, + 0.013413339853286743, + 0.016231374815106392, + 0.011245264671742916, + 0.01642264798283577, + 0.014555803500115871, + 0.012830307707190514, + 0.014524522237479687, + 0.01036101020872593, + 0.016463393345475197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.04832043871283531 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062060385942459106 + ], + "opposite_logprob_tokens": [ + -2.8120603561401367 + ], + "logprob_reward": -0.062060385942459106, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062060385942459106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15980592370033264, + 0.10352522134780884, + 0.06385204941034317, + 0.05441596359014511, + 0.05044155567884445, + 0.09096646308898926, + 0.057778436690568924, + 0.07557246088981628, + 0.25214940309524536, + 0.09149254113435745 + ], + "credit_map": [ + 0.02231292612850666, + 0.018398979678750038, + 0.011455630883574486, + 0.02205771766602993, + 0.017089633271098137, + 0.019175462424755096, + 0.01901336759328842, + 0.03030218556523323, + 0.013555080629885197, + 0.010208322666585445, + 0.013212214224040508, + 0.013945642858743668, + 0.018087105825543404, + 0.015744781121611595, + 0.00847740937024355, + 0.010294662788510323, + 0.008605876006186008, + 0.006231386214494705, + 0.006049595307558775, + 0.011023588478565216, + 0.008556204847991467, + 0.009878560900688171, + 0.00510563375428319, + 0.008401202969253063, + 0.007600353565067053, + 0.0025986700784415007, + 0.0050734891556203365, + 0.004184072837233543, + 0.010639938525855541, + 0.008175359107553959, + 0.0038336727302521467, + 0.012310403399169445, + 0.0077847884967923164, + 0.0038432178553193808, + 0.00661114789545536, + 0.004159897565841675, + 0.006501816678792238, + 0.0068904682993888855, + 0.005786858033388853, + 0.008863357827067375, + 0.021976150572299957, + 0.006166426930576563, + 0.012688507325947285, + 0.004570041783154011, + 0.015822481364011765, + 0.008725366555154324, + 0.007975751534104347, + 0.013041730970144272, + 0.010480080731213093, + 0.0041110580787062645, + 0.008594152517616749, + 0.006099608726799488, + 0.004937118384987116, + 0.014911648817360401, + 0.0036940251011401415, + 0.0049507408402860165, + 0.0124211385846138, + 0.0074019478633999825, + 0.01186809130012989, + 0.005431799218058586, + 0.007345252204686403, + 0.008426797576248646, + 0.007172266952693462, + 0.01550515741109848, + 0.02441573329269886, + 0.013738877139985561, + 0.07094739377498627, + 0.0171852745115757, + 0.01585669443011284, + 0.07808990031480789, + 0.015049448236823082, + 0.016866078600287437, + 0.012472143396735191, + 0.00988771952688694, + 0.009995290078222752, + 0.011708775535225868, + 0.009710217826068401, + 0.011952603235840797, + 0.00831609033048153, + 0.017449693754315376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.062060385942459106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260198377072811 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005260198377072811, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260198377072811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035190608352422714, + 0.03719928860664368, + 0.020402735099196434, + 0.01436326839029789, + 0.015614913776516914, + 0.0146318469196558, + 0.04370018467307091, + 0.06683745235204697, + 0.7129945755004883, + 0.03906507417559624 + ], + "credit_map": [ + 0.0043733431957662106, + 0.003658269066363573, + 0.003046551253646612, + 0.003504747524857521, + 0.0052447193302214146, + 0.007069752085953951, + 0.003453946905210614, + 0.004839277360588312, + 0.0077592418529093266, + 0.002720711287111044, + 0.0030117121059447527, + 0.004334397614002228, + 0.005104276351630688, + 0.005104485433548689, + 0.0031046282965689898, + 0.006059834733605385, + 0.002847373718395829, + 0.0015433793887495995, + 0.0020283693447709084, + 0.0014635409461334348, + 0.003964704927057028, + 0.003473872086033225, + 0.0029750370886176825, + 0.0021064593456685543, + 0.0021252641454339027, + 0.0013610579771921039, + 0.001963522518053651, + 0.0013584686676040292, + 0.0027042434085160494, + 0.0018410927150398493, + 0.0013747357297688723, + 0.0016348822973668575, + 0.0032443711534142494, + 0.0013888472458347678, + 0.0018716556951403618, + 0.0010745212202891707, + 0.0013291450450196862, + 0.0033737916965037584, + 0.0013313903473317623, + 0.002001191023737192, + 0.002254134975373745, + 0.001292069093324244, + 0.0022942263167351484, + 0.001393579412251711, + 0.0020851183217018843, + 0.0019903257489204407, + 0.0013349236687645316, + 0.001987469382584095, + 0.007611080072820187, + 0.003349021077156067, + 0.0052816481329500675, + 0.0034556412138044834, + 0.004245765041559935, + 0.010741316713392735, + 0.0038944759871810675, + 0.005121235270053148, + 0.009327273815870285, + 0.005618577357381582, + 0.00858251191675663, + 0.005385298281908035, + 0.007494854740798473, + 0.008652719669044018, + 0.0056802271865308285, + 0.01609599031507969, + 0.06261507421731949, + 0.03880294784903526, + 0.2408868968486786, + 0.03378649428486824, + 0.04092343524098396, + 0.24337302148342133, + 0.026625411584973335, + 0.025981372222304344, + 0.0064609358087182045, + 0.0038110127206891775, + 0.004921839572489262, + 0.004454178269952536, + 0.0037384473253041506, + 0.0050715371035039425, + 0.00352380215190351, + 0.007083319127559662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.005260198377072811 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185431137681007 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.018185431137681007, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185431137681007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17026643455028534, + 0.1506136953830719, + 0.12833960354328156, + 0.0840526819229126, + 0.05221565067768097, + 0.03978370502591133, + 0.03401567414402962, + 0.04606904834508896, + 0.18654918670654297, + 0.10809432715177536 + ], + "credit_map": [ + 0.02341136336326599, + 0.015667948871850967, + 0.011570489965379238, + 0.015517668798565865, + 0.02733430452644825, + 0.03295333683490753, + 0.014107479713857174, + 0.02970384806394577, + 0.03868699446320534, + 0.010345861315727234, + 0.010554388165473938, + 0.011524598114192486, + 0.02419498935341835, + 0.024242958053946495, + 0.008485527709126472, + 0.02257837913930416, + 0.018167581409215927, + 0.008368825539946556, + 0.012626409530639648, + 0.01129122730344534, + 0.02246406115591526, + 0.024760400876402855, + 0.012970393523573875, + 0.017690706998109818, + 0.009875476360321045, + 0.006238278467208147, + 0.009891975671052933, + 0.005889865569770336, + 0.016043873503804207, + 0.011782187968492508, + 0.0065057664178311825, + 0.017825260758399963, + 0.00976834911853075, + 0.002771904692053795, + 0.004441899247467518, + 0.0026359993498772383, + 0.011471499688923359, + 0.010323689319193363, + 0.0037115931045264006, + 0.007090715691447258, + 0.006828357931226492, + 0.005946398247033358, + 0.003929782193154097, + 0.004901661537587643, + 0.005831572692841291, + 0.003938767593353987, + 0.0036897845566272736, + 0.004717379808425903, + 0.006051232572644949, + 0.0023200623691082, + 0.003873201785609126, + 0.004517153836786747, + 0.0031243604607880116, + 0.0077222115360200405, + 0.0022592470049858093, + 0.004148203879594803, + 0.009937582537531853, + 0.006565178744494915, + 0.004633665084838867, + 0.0038135105278342962, + 0.004829735495150089, + 0.005801263265311718, + 0.004272548481822014, + 0.006215566303580999, + 0.0174690093845129, + 0.010359589010477066, + 0.05844176933169365, + 0.01133294403553009, + 0.011388428509235382, + 0.059643905609846115, + 0.006716140080243349, + 0.0111974086612463, + 0.016683304682374, + 0.011596695519983768, + 0.0136414160951972, + 0.012890039011836052, + 0.01167877484112978, + 0.014002327807247639, + 0.010409076698124409, + 0.017192687839269638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.018185431137681007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705351412296295 + ], + "opposite_logprob_tokens": [ + -2.1270534992218018 + ], + "logprob_reward": -0.12705351412296295, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705351412296295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17323417961597443, + 0.1829863041639328, + 0.08287780731916428, + 0.06922135502099991, + 0.06109913811087608, + 0.05749541148543358, + 0.06039535999298096, + 0.06539884954690933, + 0.1390244960784912, + 0.10826709121465683 + ], + "credit_map": [ + 0.024178339168429375, + 0.016422560438513756, + 0.012785619124770164, + 0.01150760892778635, + 0.026081779971718788, + 0.03701481595635414, + 0.011703059077262878, + 0.03354039788246155, + 0.05568346381187439, + 0.010570738464593887, + 0.014541536569595337, + 0.007151543162763119, + 0.026420675218105316, + 0.029785342514514923, + 0.009108411148190498, + 0.02972460351884365, + 0.012598209083080292, + 0.005971316248178482, + 0.005988106597214937, + 0.004487309139221907, + 0.01685631275177002, + 0.01820576936006546, + 0.009951626881957054, + 0.008819162845611572, + 0.008143218234181404, + 0.0038299981970340014, + 0.008204801008105278, + 0.004384466912597418, + 0.011778337880969048, + 0.013324004597961903, + 0.0060390327125787735, + 0.013517494313418865, + 0.008631735108792782, + 0.003704093163833022, + 0.004471807274967432, + 0.0038989365566521883, + 0.014285455457866192, + 0.014501362107694149, + 0.005103846080601215, + 0.006501898635178804, + 0.014043983072042465, + 0.004830831196159124, + 0.006749528925865889, + 0.0044588265009224415, + 0.010168393142521381, + 0.0053208330646157265, + 0.005218841601163149, + 0.006704173982143402, + 0.011021649464964867, + 0.005349630024284124, + 0.011890733614563942, + 0.005633150227367878, + 0.0046734800562262535, + 0.008821475319564342, + 0.004820766858756542, + 0.00818447582423687, + 0.011152103543281555, + 0.007758082821965218, + 0.011684348806738853, + 0.00557608250528574, + 0.005023069679737091, + 0.009224123321473598, + 0.005710703320801258, + 0.009270338341593742, + 0.014607892371714115, + 0.008966684341430664, + 0.03666442632675171, + 0.009170391596853733, + 0.009376006200909615, + 0.040124211460351944, + 0.007309737149626017, + 0.012805143371224403, + 0.018818214535713196, + 0.012907507829368114, + 0.013370705768465996, + 0.010498040355741978, + 0.012830324470996857, + 0.013510504737496376, + 0.008494272828102112, + 0.017837515100836754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.12705351412296295 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, shot from an upward angle, then slowly and steadily pulls back, ultimately framing a sweeping panorama of grandeur. The camera initially focuses on a majestic, middle-aged emperor, his figure as straight as a pine tree, standing alone at the edge of a high platform carved from massive black stone. He clad in a xuan (a deep, solemn black) silk dragon robe, intricately embroidered with dazzling gold thread, a lifelike, five-clawed golden dragon, its fangs bared and claws bared. The dragon's scales reflected a subtle metallic luster in the dim light. His long hair was held high in a pristine jade crown. His features were sharply defined, and his deep eyes shone with a piercing, torch-like light. As the camera zoomed out, he slowly turned his head with the air of a king, his gaze scanning the area below. At this moment, a breathtaking spectacle was fully revealed: beneath the platform, within a vast underground palace, thousands of terracotta warriors and horses stood in meticulous formation. Formations of infantry, archers, and charioteers stretched into the endless darkness as far as the eye could see. Each life-size figure, solemn and silent, exuded the majestic aura of war. The entire scene was illuminated by flickering torches hung high on the walls. These torches cast long, swaying shadows, creating a stark contrast of light and dark on the emperor's face and the warriors' armor, creating an epic atmosphere of solemnity, power, and eternity.", + "seed": 6623, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_2289_seed_6623_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05755140651017428, + "vlm_logprob": -0.05755140651017428, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.05755140651017428, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08746687322854996 + ], + "opposite_logprob_tokens": [ + -2.5874669551849365 + ], + "logprob_reward": -0.08746687322854996, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08746687322854996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37565577030181885, + 0.05418133735656738, + 0.024380946531891823, + 0.06979119777679443, + 0.02856433391571045, + 0.03556586802005768, + 0.2111612856388092, + 0.08411742001771927, + 0.07130265980958939, + 0.04527917876839638 + ], + "credit_map": [ + 0.026258595287799835, + 0.03722063824534416, + 0.06371594220399857, + 0.060094188898801804, + 0.049349743872880936, + 0.03161584585905075, + 0.049079302698373795, + 0.05832149088382721, + 0.0052136764861643314, + 0.005727545358240604, + 0.008132698945701122, + 0.008462212979793549, + 0.0045189568772912025, + 0.007007593754678965, + 0.00612339423969388, + 0.008995259180665016, + 0.0037246530409902334, + 0.0027510859072208405, + 0.00424659438431263, + 0.0025512578431516886, + 0.0019348686328157783, + 0.0037332153879106045, + 0.0017860598163679242, + 0.0036532115191221237, + 0.007967240177094936, + 0.00556089635938406, + 0.008492928929626942, + 0.01073260698467493, + 0.010159892961382866, + 0.006906907539814711, + 0.01159354392439127, + 0.008377180434763432, + 0.003907177597284317, + 0.002735735848546028, + 0.0034270798787474632, + 0.0031527390237897635, + 0.005476283375173807, + 0.0027870219200849533, + 0.004570895805954933, + 0.0025073986034840345, + 0.0043289572931826115, + 0.0039043936412781477, + 0.005074982065707445, + 0.005704815965145826, + 0.004245172720402479, + 0.004092766437679529, + 0.004866783041507006, + 0.0033479994162917137, + 0.034837137907743454, + 0.03414250165224075, + 0.013793911784887314, + 0.03449010103940964, + 0.005727724172174931, + 0.07455054670572281, + 0.007231290452182293, + 0.006388077046722174, + 0.0085177356377244, + 0.010971338488161564, + 0.007933028973639011, + 0.010862375609576702, + 0.01708589680492878, + 0.008397072553634644, + 0.012322249822318554, + 0.008027723990380764, + 0.0077727376483380795, + 0.014081602916121483, + 0.0060724071227014065, + 0.004486462567001581, + 0.008660819381475449, + 0.01660296879708767, + 0.009798405691981316, + 0.0038272528909146786, + 0.005458037834614515, + 0.0065634422935545444, + 0.004544738680124283, + 0.0040936158038675785, + 0.00857619196176529, + 0.005377624649554491, + 0.00600390462204814, + 0.004661621991544962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.08746687322854996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296162754297256 + ], + "opposite_logprob_tokens": [ + -2.237961530685425 + ], + "logprob_reward": -0.11296162754297256, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296162754297256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1179233267903328, + 0.05928837135434151, + 0.05088509991765022, + 0.03850387781858444, + 0.06054895743727684, + 0.03970407322049141, + 0.4350443482398987, + 0.10687582939863205, + 0.061016786843538284, + 0.03020939975976944 + ], + "credit_map": [ + 0.013919705525040627, + 0.017386889085173607, + 0.015164678916335106, + 0.018767137080430984, + 0.010239674709737301, + 0.013116479851305485, + 0.010654830373823643, + 0.01867392286658287, + 0.007498690392822027, + 0.008443077094852924, + 0.0076106395572423935, + 0.007330223917961121, + 0.006977906450629234, + 0.00678147841244936, + 0.005449294578284025, + 0.009197053499519825, + 0.007598675787448883, + 0.004688193555921316, + 0.005716740153729916, + 0.006478638853877783, + 0.006263270974159241, + 0.008769158273935318, + 0.004312120843678713, + 0.007058293558657169, + 0.007115198764950037, + 0.003562287660315633, + 0.00451726233586669, + 0.00497997784987092, + 0.002744293073192239, + 0.0062928954139351845, + 0.004366390407085419, + 0.0049255662597715855, + 0.015277933329343796, + 0.007522292900830507, + 0.005944410804659128, + 0.005118303932249546, + 0.00636611832305789, + 0.009475435130298138, + 0.005411018151789904, + 0.005433439742773771, + 0.005216027144342661, + 0.0037714990321546793, + 0.00436170632019639, + 0.004159138537943363, + 0.004430541303008795, + 0.00605808338150382, + 0.008452340960502625, + 0.0032547281589359045, + 0.07400916516780853, + 0.07897870987653732, + 0.0317622646689415, + 0.12112818658351898, + 0.012324544601142406, + 0.0761343464255333, + 0.020101185888051987, + 0.020605891942977905, + 0.012009941972792149, + 0.012271715328097343, + 0.00795238371938467, + 0.015389719046652317, + 0.023679358884692192, + 0.013417505659162998, + 0.014806867577135563, + 0.007348324172198772, + 0.007714068982750177, + 0.009204600937664509, + 0.009332677349448204, + 0.005067065823823214, + 0.006072103511542082, + 0.015034148469567299, + 0.0048202453181147575, + 0.003771871328353882, + 0.003853121306747198, + 0.003856630064547062, + 0.0034916263539344072, + 0.0030217375606298447, + 0.0032394793815910816, + 0.003775806399062276, + 0.004688156768679619, + 0.004282839130610228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.11296162754297256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084998026490211 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011084998026490211, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084998026490211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1390567272901535, + 0.06049618870019913, + 0.054589565843343735, + 0.05217772722244263, + 0.11578921973705292, + 0.036441948264837265, + 0.36109599471092224, + 0.08934324234724045, + 0.061407580971717834, + 0.02960188314318657 + ], + "credit_map": [ + 0.01771613582968712, + 0.017356975004076958, + 0.019267050549387932, + 0.02549590729176998, + 0.017917167395353317, + 0.011077304370701313, + 0.015033821575343609, + 0.015192350372672081, + 0.009756957180798054, + 0.006910537835210562, + 0.009009601548314095, + 0.006521385628730059, + 0.004664080683141947, + 0.007631703279912472, + 0.004847098141908646, + 0.011154822073876858, + 0.010168926790356636, + 0.005378568544983864, + 0.009014167822897434, + 0.006589837372303009, + 0.004128145519644022, + 0.007351656910032034, + 0.003813577350229025, + 0.0081446822732687, + 0.011973671615123749, + 0.004796395543962717, + 0.006079790182411671, + 0.0052079795859754086, + 0.004198937211185694, + 0.009852276183664799, + 0.003516196273267269, + 0.006552478298544884, + 0.012854456901550293, + 0.004824358969926834, + 0.03705039620399475, + 0.00901828333735466, + 0.005690063815563917, + 0.03297547996044159, + 0.005723300855606794, + 0.007652867119759321, + 0.004603527951985598, + 0.00478768814355135, + 0.003934308886528015, + 0.003948325756937265, + 0.004309284035116434, + 0.004806755110621452, + 0.0059057497419416904, + 0.004146306775510311, + 0.03419386222958565, + 0.0666627436876297, + 0.026276974007487297, + 0.08457258343696594, + 0.010846501216292381, + 0.10953609645366669, + 0.015655124559998512, + 0.013352089561522007, + 0.012863020412623882, + 0.010962316766381264, + 0.009076590649783611, + 0.008055420592427254, + 0.018220752477645874, + 0.008390049450099468, + 0.015167312696576118, + 0.006607771385461092, + 0.0075585003942251205, + 0.009652938693761826, + 0.005855639930814505, + 0.005678515415638685, + 0.005640572868287563, + 0.01772848702967167, + 0.005528038367629051, + 0.003764877561479807, + 0.0040800608694553375, + 0.0034116243477910757, + 0.0024465566966682673, + 0.002331665251404047, + 0.004643217194825411, + 0.004759056959301233, + 0.005402155686169863, + 0.0025275421794503927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.011084998026490211 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196592383086681 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196592383086681, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196592383086681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06241847947239876, + 0.032480377703905106, + 0.02692009136080742, + 0.05398299917578697, + 0.14017930626869202, + 0.01727433316409588, + 0.36595892906188965, + 0.05948136746883392, + 0.18351905047893524, + 0.05778510123491287 + ], + "credit_map": [ + 0.008843081071972847, + 0.00973886251449585, + 0.008674245327711105, + 0.009504138492047787, + 0.006190339103341103, + 0.005987122189253569, + 0.006478342693299055, + 0.0070023504085838795, + 0.004048062488436699, + 0.004004125948995352, + 0.004309211857616901, + 0.004673038609325886, + 0.002573327161371708, + 0.004622641950845718, + 0.0042504980228841305, + 0.003999473061412573, + 0.004154716618359089, + 0.0033745167311280966, + 0.0032366798259317875, + 0.003287808271124959, + 0.002946974942460656, + 0.004049886949360371, + 0.002164017641916871, + 0.0037054915446788073, + 0.00931288581341505, + 0.005464979913085699, + 0.010274362750351429, + 0.005507498048245907, + 0.004345965571701527, + 0.008870094083249569, + 0.0036749725695699453, + 0.006532241590321064, + 0.019007354974746704, + 0.0057177855633199215, + 0.04342656582593918, + 0.008235061541199684, + 0.004741741344332695, + 0.04646238684654236, + 0.004781673662364483, + 0.007806743029505014, + 0.0022854222916066647, + 0.0018188877729699016, + 0.0019321056315675378, + 0.002596698235720396, + 0.0016403190093114972, + 0.0027842565905302763, + 0.002301163272932172, + 0.0019154828041791916, + 0.03656899183988571, + 0.07014168053865433, + 0.027246804907917976, + 0.09525881707668304, + 0.009874281473457813, + 0.09511860460042953, + 0.01622418314218521, + 0.01552559994161129, + 0.007919235154986382, + 0.007873903028666973, + 0.006349788047373295, + 0.004653573036193848, + 0.010643364861607552, + 0.008262420073151588, + 0.007918743416666985, + 0.005860345903784037, + 0.024598417803645134, + 0.03534471616148949, + 0.019670207053422928, + 0.01068788394331932, + 0.011943885125219822, + 0.06376054883003235, + 0.008748904801905155, + 0.00876450352370739, + 0.00801054947078228, + 0.007551555521786213, + 0.006736872252076864, + 0.005649986211210489, + 0.005726547446101904, + 0.010067392140626907, + 0.007034762762486935, + 0.007007439620792866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.014196592383086681 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204694136977196 + ], + "opposite_logprob_tokens": [ + -2.812047004699707 + ], + "logprob_reward": -0.06204694136977196, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204694136977196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24327751994132996, + 0.0974133089184761, + 0.035316549241542816, + 0.10736102610826492, + 0.1075764149427414, + 0.025771358981728554, + 0.20761971175670624, + 0.06539778411388397, + 0.043533965945243835, + 0.06673246622085571 + ], + "credit_map": [ + 0.024811819195747375, + 0.02138346992433071, + 0.041240543127059937, + 0.04888094589114189, + 0.023924527689814568, + 0.028670603409409523, + 0.020547114312648773, + 0.033818505704402924, + 0.012545391917228699, + 0.011215337552130222, + 0.013589302077889442, + 0.010616675950586796, + 0.007033531088382006, + 0.019320713356137276, + 0.01069696620106697, + 0.012395387515425682, + 0.006721768993884325, + 0.004131456837058067, + 0.003684816649183631, + 0.004503164440393448, + 0.0030789049342274666, + 0.00496198795735836, + 0.003234405070543289, + 0.005000046454370022, + 0.019029922783374786, + 0.012045473791658878, + 0.01569698564708233, + 0.013581349514424801, + 0.00851417426019907, + 0.017418120056390762, + 0.009086467325687408, + 0.01198853924870491, + 0.013275278732180595, + 0.007755215745419264, + 0.034177739173173904, + 0.008476329036056995, + 0.004565093200653791, + 0.027319129556417465, + 0.004311583004891872, + 0.007696051616221666, + 0.004354856442660093, + 0.0032554869540035725, + 0.0024035486858338118, + 0.0027347607538104057, + 0.002817116677761078, + 0.0034351148642599583, + 0.004634704440832138, + 0.0021357694640755653, + 0.022165164351463318, + 0.03968748822808266, + 0.0156863946467638, + 0.05487826094031334, + 0.005511246155947447, + 0.051305919885635376, + 0.009657547809183598, + 0.008727689273655415, + 0.009961530566215515, + 0.007485885638743639, + 0.007314459886401892, + 0.007643916178494692, + 0.011148489080369473, + 0.006215182598680258, + 0.008504021912813187, + 0.0071242996491491795, + 0.005289097782224417, + 0.006855044048279524, + 0.003996947314590216, + 0.003003595629706979, + 0.003949803300201893, + 0.01104273647069931, + 0.006174673326313496, + 0.0032220694702118635, + 0.01063946820795536, + 0.01258846279233694, + 0.008409143425524235, + 0.006326451897621155, + 0.008089837618172169, + 0.008507667109370232, + 0.005995843559503555, + 0.006175593473017216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.06204694136977196 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, shot from an upward angle, then slowly and steadily pulls back, ultimately framing a sweeping panorama of grandeur. The camera initially focuses on a majestic, middle-aged emperor, his figure as straight as a pine tree, standing alone at the edge of a high platform carved from massive black stone. He clad in a xuan (a deep, solemn black) silk dragon robe, intricately embroidered with dazzling gold thread, a lifelike, five-clawed golden dragon, its fangs bared and claws bared. The dragon's scales reflected a subtle metallic luster in the dim light. His long hair was held high in a pristine jade crown. His features were sharply defined, and his deep eyes shone with a piercing, torch-like light. As the camera zoomed out, he slowly turned his head with the air of a king, his gaze scanning the area below. At this moment, a breathtaking spectacle was fully revealed: beneath the platform, within a vast underground palace, thousands of terracotta warriors and horses stood in meticulous formation. Formations of infantry, archers, and charioteers stretched into the endless darkness as far as the eye could see. Each life-size figure, solemn and silent, exuded the majestic aura of war. The entire scene was illuminated by flickering torches hung high on the walls. These torches cast long, swaying shadows, creating a stark contrast of light and dark on the emperor's face and the warriors' armor, creating an epic atmosphere of solemnity, power, and eternity.", + "seed": 6624, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_13/index_2289_seed_6624_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04194468669593334, + "vlm_logprob": -0.04194468669593334, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04194468669593334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08557865768671036 + ], + "opposite_logprob_tokens": [ + -2.585578680038452 + ], + "logprob_reward": -0.08557865768671036, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08557865768671036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18915285170078278, + 0.09448602050542831, + 0.0985008105635643, + 0.10198301821947098, + 0.061500806361436844, + 0.07503252476453781, + 0.09190905094146729, + 0.041870906949043274, + 0.20090508460998535, + 0.04465893283486366 + ], + "credit_map": [ + 0.029526222497224808, + 0.022561339661478996, + 0.022397199645638466, + 0.03149093687534332, + 0.022084854543209076, + 0.024035444483160973, + 0.020193452015519142, + 0.016863396391272545, + 0.0116145433858037, + 0.013559994287788868, + 0.01301447581499815, + 0.010960311628878117, + 0.008735337294638157, + 0.016605202108621597, + 0.007340023759752512, + 0.012656123377382755, + 0.010336495004594326, + 0.012737629003822803, + 0.015364053659141064, + 0.00918084941804409, + 0.013928028754889965, + 0.011754784733057022, + 0.013331671245396137, + 0.01186730433255434, + 0.008733470924198627, + 0.01318779680877924, + 0.0136957336217165, + 0.0131983095780015, + 0.011059177108108997, + 0.02127351053059101, + 0.00943552516400814, + 0.011399500072002411, + 0.006212614011019468, + 0.014544435776770115, + 0.010989019647240639, + 0.003675075713545084, + 0.005758057348430157, + 0.0033735684119164944, + 0.008804487995803356, + 0.008143549785017967, + 0.0070388829335570335, + 0.0077891843393445015, + 0.008230632171034813, + 0.015240425243973732, + 0.011705988086760044, + 0.006976818200200796, + 0.009390090592205524, + 0.008660503663122654, + 0.006892477627843618, + 0.01161642000079155, + 0.009965574368834496, + 0.01171916350722313, + 0.008654577657580376, + 0.0072609493508934975, + 0.02002674899995327, + 0.015773136168718338, + 0.004424390848726034, + 0.007400241214782, + 0.007040782365947962, + 0.004509055055677891, + 0.003894967492669821, + 0.0047933184541761875, + 0.0034231706522405148, + 0.0063849822618067265, + 0.01641598530113697, + 0.023903314024209976, + 0.049138639122247696, + 0.03628043085336685, + 0.008933250792324543, + 0.008949555456638336, + 0.019543612375855446, + 0.03774028643965721, + 0.0044729020446538925, + 0.009452794678509235, + 0.0046263188123703, + 0.0036227183882147074, + 0.003548262408003211, + 0.005905108992010355, + 0.006576017010957003, + 0.006454810965806246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.08557865768671036 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062056463211774826 + ], + "opposite_logprob_tokens": [ + -2.812056541442871 + ], + "logprob_reward": -0.062056463211774826, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062056463211774826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1663973182439804, + 0.15962865948677063, + 0.11485308408737183, + 0.09207309782505035, + 0.057290706783533096, + 0.12441687285900116, + 0.056899089366197586, + 0.04292379319667816, + 0.1322282999753952, + 0.053289126604795456 + ], + "credit_map": [ + 0.02754841186106205, + 0.0174667127430439, + 0.023382430896162987, + 0.018230797722935677, + 0.020380087196826935, + 0.028774797916412354, + 0.010980886407196522, + 0.019633186981081963, + 0.02829277701675892, + 0.0239508505910635, + 0.021851057186722755, + 0.013936794362962246, + 0.015156238339841366, + 0.025482546538114548, + 0.011432054452598095, + 0.01952633075416088, + 0.012335853651165962, + 0.01391019020229578, + 0.012126177549362183, + 0.01613389141857624, + 0.01469066645950079, + 0.01951187290251255, + 0.011542879976332188, + 0.014601551927626133, + 0.00848621316254139, + 0.012244395911693573, + 0.015071303583681583, + 0.010341214016079903, + 0.012081236578524113, + 0.013101103715598583, + 0.008002317510545254, + 0.012745313346385956, + 0.005837711505591869, + 0.010763656347990036, + 0.010092535056173801, + 0.005728756543248892, + 0.0069366758689284325, + 0.003938894718885422, + 0.006175793241709471, + 0.007816684432327747, + 0.011852087453007698, + 0.011812999844551086, + 0.014987906441092491, + 0.019047435373067856, + 0.022518189623951912, + 0.01541577372699976, + 0.015043789520859718, + 0.013738683424890041, + 0.004853724502027035, + 0.011176018975675106, + 0.006730676628649235, + 0.007937637157738209, + 0.0035803441423922777, + 0.004597637802362442, + 0.006730738561600447, + 0.011292312294244766, + 0.00350417522713542, + 0.00422705290839076, + 0.007144187577068806, + 0.0054566022008657455, + 0.004046432673931122, + 0.0070259966887533665, + 0.005417123436927795, + 0.00610222015529871, + 0.010819293558597565, + 0.014755572192370892, + 0.03302560746669769, + 0.023030465468764305, + 0.005297405179589987, + 0.007800877559930086, + 0.012857107445597649, + 0.02464197389781475, + 0.0069523765705525875, + 0.009811229072511196, + 0.009720511734485626, + 0.006197165697813034, + 0.003170883748680353, + 0.0046686334535479546, + 0.005328394938260317, + 0.007439932785928249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.062056463211774826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024946779012680054 + ], + "opposite_logprob_tokens": [ + -6.002494812011719 + ], + "logprob_reward": -0.0024946779012680054, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024946779012680054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12447017431259155, + 0.056581757962703705, + 0.0883929505944252, + 0.03402145579457283, + 0.07771429419517517, + 0.08382267504930496, + 0.1972685009241104, + 0.10943201929330826, + 0.2003125101327896, + 0.027983661741018295 + ], + "credit_map": [ + 0.025142734870314598, + 0.009323665872216225, + 0.020550858229398727, + 0.006261324044317007, + 0.008049202151596546, + 0.028876399621367455, + 0.00719366455450654, + 0.019072329625487328, + 0.006741313263773918, + 0.011468358337879181, + 0.006864331196993589, + 0.005819614976644516, + 0.006270247511565685, + 0.007633119355887175, + 0.005094366613775492, + 0.006690406706184149, + 0.00911317951977253, + 0.010912871919572353, + 0.01379892136901617, + 0.016473766416311264, + 0.008267016150057316, + 0.01423751749098301, + 0.006450125481933355, + 0.00913955271244049, + 0.0030275091994553804, + 0.004695108626037836, + 0.006187154445797205, + 0.0038764721248298883, + 0.0037796390242874622, + 0.004325516987591982, + 0.00348975439555943, + 0.004640304017812014, + 0.006785031873732805, + 0.01961030811071396, + 0.013010495342314243, + 0.004995173308998346, + 0.006903572473675013, + 0.005397751461714506, + 0.009114007465541363, + 0.011897956021130085, + 0.008275915868580341, + 0.008853658102452755, + 0.008665923029184341, + 0.009089422412216663, + 0.011933285742998123, + 0.014181449078023434, + 0.009194782935082912, + 0.013628246262669563, + 0.018450533971190453, + 0.03725481778383255, + 0.02143823355436325, + 0.028449971228837967, + 0.01621258072555065, + 0.016135884448885918, + 0.03373616933822632, + 0.025590311735868454, + 0.013165508396923542, + 0.01364802848547697, + 0.02241661585867405, + 0.014364481903612614, + 0.011857234872877598, + 0.010441410355269909, + 0.008586994372308254, + 0.014951743185520172, + 0.015319058671593666, + 0.018855435773730278, + 0.05202072113752365, + 0.03555532172322273, + 0.008996610529720783, + 0.008804449811577797, + 0.019461052492260933, + 0.041299864649772644, + 0.002603936940431595, + 0.003112263511866331, + 0.004452667199075222, + 0.003649759339168668, + 0.002035240177065134, + 0.0032326276414096355, + 0.005066323559731245, + 0.003830842673778534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.0024946779012680054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796700924634933 + ], + "opposite_logprob_tokens": [ + -3.529796600341797 + ], + "logprob_reward": -0.029796700924634933, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796700924634933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19962668418884277, + 0.06547082960605621, + 0.113221175968647, + 0.0645049661397934, + 0.057556696236133575, + 0.06112372875213623, + 0.13890761137008667, + 0.14869417250156403, + 0.08846699446439743, + 0.062427155673503876 + ], + "credit_map": [ + 0.04042661190032959, + 0.021072303876280785, + 0.02594093047082424, + 0.01691131293773651, + 0.018984174355864525, + 0.03535471111536026, + 0.013317981734871864, + 0.027618659660220146, + 0.010201772674918175, + 0.009084810502827168, + 0.008791287429630756, + 0.004850420169532299, + 0.005226300563663244, + 0.012465865351259708, + 0.0038596459198743105, + 0.010990729555487633, + 0.013674763031303883, + 0.017045937478542328, + 0.013385227881371975, + 0.012928643263876438, + 0.014260385185480118, + 0.017205219715833664, + 0.013522094115614891, + 0.011198905296623707, + 0.006225725170224905, + 0.007870283909142017, + 0.010829463601112366, + 0.005992757622152567, + 0.008833933621644974, + 0.01032478827983141, + 0.006335467100143433, + 0.008092546835541725, + 0.005583044607192278, + 0.012577762827277184, + 0.009909282438457012, + 0.0034792947117239237, + 0.004953440744429827, + 0.00606194045394659, + 0.006306848954409361, + 0.008685080334544182, + 0.006051075179129839, + 0.006831309758126736, + 0.006889318581670523, + 0.011243336834013462, + 0.008828696794807911, + 0.0062599689699709415, + 0.0069638267159461975, + 0.008056195452809334, + 0.011637107469141483, + 0.023374954238533974, + 0.01537393219769001, + 0.02670075185596943, + 0.012515285983681679, + 0.012278061360120773, + 0.013146606273949146, + 0.023880911991000175, + 0.014859622344374657, + 0.025564823299646378, + 0.02392761968076229, + 0.01667345128953457, + 0.01309798564761877, + 0.017803410068154335, + 0.015324011445045471, + 0.021443244069814682, + 0.007416599430143833, + 0.01353151723742485, + 0.020133329555392265, + 0.01147074718028307, + 0.0034807506017386913, + 0.006072686519473791, + 0.009365029633045197, + 0.01699632965028286, + 0.005396020598709583, + 0.008376391604542732, + 0.010510096326470375, + 0.006463538855314255, + 0.0050805010832846165, + 0.009196293540298939, + 0.008887400850653648, + 0.008516914211213589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.029796700924634933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged emperor wearing a black silk dragon robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emperor standing alone on a high platform carved from massive black stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a medium shot from an upward angle and slowly pull back to reveal a wide panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of terracotta warriors and horses visible in a vast underground palace beneath the platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by flickering torches that cast long, swaying shadows on the figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796933755278587 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796933755278587, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796933755278587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10206976532936096, + 0.09101814776659012, + 0.061988312751054764, + 0.07745441049337387, + 0.16412584483623505, + 0.07594853639602661, + 0.09586644172668457, + 0.099573515355587, + 0.1866721361875534, + 0.04528288543224335 + ], + "credit_map": [ + 0.021334735676646233, + 0.010936573147773743, + 0.014999816194176674, + 0.008580481633543968, + 0.008767562918365002, + 0.013608471490442753, + 0.00953687820583582, + 0.01430524606257677, + 0.010742670856416225, + 0.005560835357755423, + 0.014148625545203686, + 0.005736163351684809, + 0.010219485498964787, + 0.02263539284467697, + 0.005597117822617292, + 0.01637786254286766, + 0.007202541455626488, + 0.007406963035464287, + 0.007737007457762957, + 0.007450880482792854, + 0.007024246267974377, + 0.010034773498773575, + 0.008276405744254589, + 0.006855495739728212, + 0.006921006832271814, + 0.01057130005210638, + 0.012806864455342293, + 0.00881727784872055, + 0.009001373313367367, + 0.012564709410071373, + 0.006646738853305578, + 0.01012514065951109, + 0.018217815086245537, + 0.02884506992995739, + 0.014608119614422321, + 0.013784050941467285, + 0.018823539838194847, + 0.026758961379528046, + 0.02294236421585083, + 0.020145917311310768, + 0.007520465180277824, + 0.010380295105278492, + 0.007164521608501673, + 0.018360989168286324, + 0.0068517159670591354, + 0.008666593581438065, + 0.008700840175151825, + 0.008303113281726837, + 0.008055525831878185, + 0.017823385074734688, + 0.013752101920545101, + 0.012077482417225838, + 0.0076658823527395725, + 0.009974835440516472, + 0.011465033516287804, + 0.015052197501063347, + 0.010802878998219967, + 0.012904828414320946, + 0.016258636489510536, + 0.008803396485745907, + 0.011744811199605465, + 0.01664360985159874, + 0.007933816872537136, + 0.014481544494628906, + 0.017329005524516106, + 0.025578314438462257, + 0.028189362958073616, + 0.02536134235560894, + 0.013076210394501686, + 0.03332379832863808, + 0.016424110159277916, + 0.02738998644053936, + 0.005454115569591522, + 0.0125422403216362, + 0.00469682551920414, + 0.003971720580011606, + 0.0029255077242851257, + 0.005128847900778055, + 0.004432880785316229, + 0.00613074516877532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.029796933755278587 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_14/index_tensor([581, 581])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_14/index_tensor([581, 581])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4b76d15e590013a7abc75d49e9b628030be1cf8d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_14/index_tensor([581, 581])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, extreme close-up focuses on a piece of weathered wood, its surface covered in cracked and flaking sky-blue paint, laid against the dark, grainy texture of asphalt pavement. The scene, baked in the harsh, direct midday sun, causes a visible heat haze to flicker above the ground. The camera remains perfectly still, observing as subtle fluctuations begin to disturb the flat, matte surface of the paint. Driven by the hot, expanding gases trapped beneath, the paint begins to expand upward from a central point. The blue layer stretches relentlessly, its color thinning slightly to form a smooth, tense dome. The process culminates in a stable, convex bubble that now sits prominently on the wood's surface. Its perfectly hemispherical shape captures the bright, mirror-like highlights of the sunlight and casts a tiny, sharp shadow on the chipped paint nearby, creating a quietly mesmerizing atmosphere of intense heat and slow transformation.", + "seed": 9737, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_14/index_581_seed_9737_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08689128700643778, + "vlm_logprob": -0.08689128700643778, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.08689128700643778, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06252847611904144 + ], + "opposite_logprob_tokens": [ + -2.937528371810913 + ], + "logprob_reward": -0.06252847611904144, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06252847611904144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06562690436840057, + 0.011659393087029457, + 0.012046810239553452, + 0.016873611137270927, + 0.03062715008854866, + 0.02834472618997097, + 0.020914994180202484, + 0.016905466094613075, + 0.7852420806884766, + 0.011758871376514435 + ], + "credit_map": [ + 0.01179119385778904, + 0.018184516578912735, + 0.004957215860486031, + 0.004758927971124649, + 0.008530628867447376, + 0.010075759142637253, + 0.003155831480398774, + 0.004172833636403084, + 0.001980177126824856, + 0.0020631016232073307, + 0.0010465990053489804, + 0.0011385666439309716, + 0.0016593694454059005, + 0.0018681138753890991, + 0.0006904230685904622, + 0.001213041483424604, + 0.001745324581861496, + 0.0022407034412026405, + 0.0009599409531801939, + 0.0007527289562858641, + 0.002431304892525077, + 0.002112944843247533, + 0.0008483424317091703, + 0.0009555198485031724, + 0.0021955466363579035, + 0.0024700877256691456, + 0.0017252707621082664, + 0.002171894535422325, + 0.003134744009003043, + 0.002070849062874913, + 0.0016205244464799762, + 0.0014846924459561706, + 0.003395688720047474, + 0.005767452530562878, + 0.0037863089237362146, + 0.004677842371165752, + 0.003028427017852664, + 0.0034046154469251633, + 0.0034961134660989046, + 0.003070702077820897, + 0.009985135868191719, + 0.002742836019024253, + 0.0022159218788146973, + 0.001826772466301918, + 0.005997150205075741, + 0.002603975124657154, + 0.001292537315748632, + 0.0016803962644189596, + 0.00537547841668129, + 0.00222064764238894, + 0.0029809707775712013, + 0.002435571514070034, + 0.0015168049139901996, + 0.0031982960645109415, + 0.0013768476201221347, + 0.001810377580113709, + 0.002322146203368902, + 0.001905427547171712, + 0.002229224191978574, + 0.0019403304904699326, + 0.002168251434341073, + 0.0021689848508685827, + 0.0019924913067370653, + 0.0021786096040159464, + 0.0813155323266983, + 0.049899592995643616, + 0.03445923328399658, + 0.10651874542236328, + 0.17708958685398102, + 0.2559618651866913, + 0.047181516885757446, + 0.032815977931022644, + 0.0010939511703327298, + 0.001355344313196838, + 0.0015210171695798635, + 0.002643979387357831, + 0.0010041693458333611, + 0.0016410216921940446, + 0.0017289543757215142, + 0.0007704335730522871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.06252847611904144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562508285045624 + ], + "opposite_logprob_tokens": [ + -1.6006250381469727 + ], + "logprob_reward": -0.22562508285045624, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562508285045624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08273676782846451, + 0.05567878857254982, + 0.03129144012928009, + 0.01790666952729225, + 0.048406001180410385, + 0.17930790781974792, + 0.12643729150295258, + 0.2266748696565628, + 0.16614200174808502, + 0.06541828066110611 + ], + "credit_map": [ + 0.014143652282655239, + 0.01580902375280857, + 0.009321560151875019, + 0.006637211889028549, + 0.010479310527443886, + 0.014008567668497562, + 0.0047820801846683025, + 0.007555366959422827, + 0.0076045203022658825, + 0.011578606441617012, + 0.00487675704061985, + 0.0042128474451601505, + 0.01020548865199089, + 0.008278409019112587, + 0.0030264060478657484, + 0.005895757582038641, + 0.005208307411521673, + 0.005049639847129583, + 0.0023183245211839676, + 0.002362960483878851, + 0.0038742637261748314, + 0.00754594849422574, + 0.002253834856674075, + 0.002678163116797805, + 0.0024572783149778843, + 0.002985160332173109, + 0.0017928631277754903, + 0.0022441414184868336, + 0.0019010743126273155, + 0.0029268902726471424, + 0.001752827549353242, + 0.001846436643972993, + 0.006752472370862961, + 0.010467403568327427, + 0.005058594048023224, + 0.005966839846223593, + 0.005823200102895498, + 0.005409532692283392, + 0.004747915547341108, + 0.004180044401437044, + 0.05217220261693001, + 0.017319533973932266, + 0.014827563427388668, + 0.008720719255506992, + 0.04809734597802162, + 0.01980951800942421, + 0.006797655951231718, + 0.011563392356038094, + 0.027857717126607895, + 0.014020085334777832, + 0.016143403947353363, + 0.013429773040115833, + 0.011263648979365826, + 0.017033066600561142, + 0.010059276595711708, + 0.016630318015813828, + 0.05227028578519821, + 0.022792968899011612, + 0.033447157591581345, + 0.02613069675862789, + 0.017842508852481842, + 0.031489502638578415, + 0.01856468804180622, + 0.02413705550134182, + 0.02810842916369438, + 0.011981537565588951, + 0.007288709748536348, + 0.02268403209745884, + 0.044963955879211426, + 0.03442190960049629, + 0.009693368338048458, + 0.007000056095421314, + 0.009841345250606537, + 0.007156653795391321, + 0.004670736845582724, + 0.013905253261327744, + 0.004526571836322546, + 0.013910979963839054, + 0.008023771457374096, + 0.0033829757012426853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.22562508285045624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006418551784008741 + ], + "opposite_logprob_tokens": [ + -7.375641822814941 + ], + "logprob_reward": -0.0006418551784008741, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006418551784008741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08155717700719833, + 0.056198589503765106, + 0.03974879905581474, + 0.050844382494688034, + 0.06925547122955322, + 0.12023705989122391, + 0.07898151129484177, + 0.1534656435251236, + 0.3039955794811249, + 0.0457158125936985 + ], + "credit_map": [ + 0.012562641873955727, + 0.013233385048806667, + 0.009173184633255005, + 0.006853344384580851, + 0.008836370892822742, + 0.01616654545068741, + 0.0052201105281710625, + 0.009511593729257584, + 0.0074006798677146435, + 0.010736324824392796, + 0.005201329942792654, + 0.005268245469778776, + 0.009757548570632935, + 0.007645526435226202, + 0.003844460239633918, + 0.006344472989439964, + 0.004799901507794857, + 0.0072656311094760895, + 0.003709041979163885, + 0.0036953489761799574, + 0.00780126079916954, + 0.004904212895780802, + 0.0034730741754174232, + 0.004100325983017683, + 0.006036140490323305, + 0.011412329971790314, + 0.005387566983699799, + 0.0067514693364501, + 0.0071068257093429565, + 0.005774349439889193, + 0.004019956104457378, + 0.004355747718364, + 0.009707662276923656, + 0.011189146898686886, + 0.010102490894496441, + 0.007766809314489365, + 0.007001176942139864, + 0.008404065854847431, + 0.007472326047718525, + 0.007611797656863928, + 0.019721215590834618, + 0.016247114166617393, + 0.017358746379613876, + 0.012385567650198936, + 0.01360308937728405, + 0.017618674784898758, + 0.008605088107287884, + 0.01469756942242384, + 0.013473907485604286, + 0.01017797365784645, + 0.010734453797340393, + 0.00874590314924717, + 0.008708411827683449, + 0.012496124021708965, + 0.006098364945501089, + 0.008546372875571251, + 0.0231613889336586, + 0.018910549581050873, + 0.01907041296362877, + 0.021462636068463326, + 0.015175567008554935, + 0.02431461401283741, + 0.014527545310556889, + 0.016842924058437347, + 0.04218876361846924, + 0.018354613333940506, + 0.016138827428221703, + 0.052101340144872665, + 0.04170149192214012, + 0.09758405387401581, + 0.02186596766114235, + 0.014060506597161293, + 0.009741946123540401, + 0.005756075028330088, + 0.0035454074386507273, + 0.012357092462480068, + 0.002865710062906146, + 0.0034620696678757668, + 0.005745202302932739, + 0.0022423097398132086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.0006418551784008741 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821395406499505 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821395406499505, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821395406499505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03358244150876999, + 0.02497309260070324, + 0.012828716076910496, + 0.018868790939450264, + 0.021321970969438553, + 0.18678100407123566, + 0.037330012768507004, + 0.14818312227725983, + 0.47050604224205017, + 0.04562477394938469 + ], + "credit_map": [ + 0.004545166622847319, + 0.0074701206758618355, + 0.002372436923906207, + 0.002693385584279895, + 0.006861258298158646, + 0.0045115225948393345, + 0.002098473021760583, + 0.003030079184100032, + 0.003533007809892297, + 0.00522005558013916, + 0.0020161878783255816, + 0.0016698817489668727, + 0.005646467674523592, + 0.003658569185063243, + 0.001313459244556725, + 0.0019154640613123775, + 0.002047229092568159, + 0.002217134926468134, + 0.0013930657878518105, + 0.000865043664816767, + 0.002274317666888237, + 0.0022387844510376453, + 0.000717307673767209, + 0.0010758332209661603, + 0.003012305125594139, + 0.0031720770057290792, + 0.002233922015875578, + 0.001763999229297042, + 0.0022561897058039904, + 0.002935799304395914, + 0.0015838878462091088, + 0.0019106109393760562, + 0.0022995539475232363, + 0.0043356106616556644, + 0.0023103475105017424, + 0.0029854511376470327, + 0.0027528624050319195, + 0.0029514527413994074, + 0.001970424549654126, + 0.0017162678996101022, + 0.06454102694988251, + 0.01645844057202339, + 0.010481497272849083, + 0.009445409290492535, + 0.05549824237823486, + 0.01796679012477398, + 0.0058764806017279625, + 0.006513130385428667, + 0.007419545669108629, + 0.004115493502467871, + 0.005049943923950195, + 0.004017733037471771, + 0.0046227360144257545, + 0.005618556868284941, + 0.002714170143008232, + 0.0037718333769589663, + 0.022479791194200516, + 0.015644067898392677, + 0.02613208442926407, + 0.018282052129507065, + 0.01218294445425272, + 0.01951681263744831, + 0.013727896846830845, + 0.02021748572587967, + 0.049385398626327515, + 0.030040619894862175, + 0.020218877121806145, + 0.059562213718891144, + 0.10978980362415314, + 0.15515735745429993, + 0.027321044355630875, + 0.019030708819627762, + 0.00785741489380598, + 0.008885606192052364, + 0.004446965642273426, + 0.00911973137408495, + 0.0036577784921973944, + 0.0037054638378322124, + 0.005276159383356571, + 0.002675653202459216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.002821395406499505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283962547779083 + ], + "opposite_logprob_tokens": [ + -2.0178396701812744 + ], + "logprob_reward": -0.14283962547779083, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283962547779083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11315460503101349, + 0.06793387234210968, + 0.041531000286340714, + 0.06908416002988815, + 0.15546800196170807, + 0.1680389791727066, + 0.06149696186184883, + 0.05582033097743988, + 0.20219595730304718, + 0.06527604162693024 + ], + "credit_map": [ + 0.019914159551262856, + 0.022097917273640633, + 0.009606968611478806, + 0.01129901222884655, + 0.012455602176487446, + 0.021070687100291252, + 0.006684581749141216, + 0.010025693103671074, + 0.01108513679355383, + 0.011105455458164215, + 0.006388638634234667, + 0.005568575579673052, + 0.012039118446409702, + 0.010541630908846855, + 0.004120127763599157, + 0.007085200399160385, + 0.005412771832197905, + 0.005996865686029196, + 0.0034303502179682255, + 0.0037294568028301, + 0.007192648947238922, + 0.008521023206412792, + 0.0034116192255169153, + 0.003836272284388542, + 0.008089101873338223, + 0.017638452351093292, + 0.00831846334040165, + 0.007292306981980801, + 0.007797339465469122, + 0.00856252945959568, + 0.00563293881714344, + 0.005753037054091692, + 0.017714951187372208, + 0.03421622887253761, + 0.021577782928943634, + 0.02008122391998768, + 0.013480382040143013, + 0.015707967802882195, + 0.017662236467003822, + 0.015027240850031376, + 0.0503147654235363, + 0.016165001317858696, + 0.017142893746495247, + 0.015540055930614471, + 0.020396122708916664, + 0.020688971504569054, + 0.012976858764886856, + 0.014814325608313084, + 0.0165010467171669, + 0.00428564939647913, + 0.005965136922895908, + 0.004207275342196226, + 0.006637177895754576, + 0.013002573512494564, + 0.003222459927201271, + 0.00767565006390214, + 0.0162319578230381, + 0.006421525031328201, + 0.005429674871265888, + 0.006908942013978958, + 0.004212039988487959, + 0.008387974463403225, + 0.0048203724436461926, + 0.003407846437767148, + 0.03408480063080788, + 0.01250541117042303, + 0.008828855119645596, + 0.013615532778203487, + 0.046681735664606094, + 0.06879357993602753, + 0.009063022211194038, + 0.008623038418591022, + 0.012897899374365807, + 0.009795956313610077, + 0.009036862291395664, + 0.009774361737072468, + 0.004862678702920675, + 0.007282778155058622, + 0.007121985778212547, + 0.004503525327891111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.14283962547779083 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme close-up focuses on a piece of weathered wood, its surface covered in cracked and flaking sky-blue paint, laid against the dark, grainy texture of asphalt pavement. The scene, baked in the harsh, direct midday sun, causes a visible heat haze to flicker above the ground. The camera remains perfectly still, observing as subtle fluctuations begin to disturb the flat, matte surface of the paint. Driven by the hot, expanding gases trapped beneath, the paint begins to expand upward from a central point. The blue layer stretches relentlessly, its color thinning slightly to form a smooth, tense dome. The process culminates in a stable, convex bubble that now sits prominently on the wood's surface. Its perfectly hemispherical shape captures the bright, mirror-like highlights of the sunlight and casts a tiny, sharp shadow on the chipped paint nearby, creating a quietly mesmerizing atmosphere of intense heat and slow transformation.", + "seed": 9738, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_14/index_581_seed_9738_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13375197877176107, + "vlm_logprob": -0.13375197877176107, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.13375197877176107, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03307054936885834 + ], + "opposite_logprob_tokens": [ + -4.4080705642700195 + ], + "logprob_reward": -0.03307054936885834, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03307054936885834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09810645133256912, + 0.21162456274032593, + 0.04145315662026405, + 0.15025320649147034, + 0.18445463478565216, + 0.06587143987417221, + 0.04986899718642235, + 0.06631164997816086, + 0.04757174849510193, + 0.08448413759469986 + ], + "credit_map": [ + 0.014817614108324051, + 0.011022687889635563, + 0.008378427475690842, + 0.013704660348594189, + 0.010338131338357925, + 0.007773350924253464, + 0.015542253851890564, + 0.016529327258467674, + 0.042355749756097794, + 0.03222329169511795, + 0.013279182836413383, + 0.018300244584679604, + 0.05347869545221329, + 0.012259343639016151, + 0.019094178453087807, + 0.020633865147829056, + 0.005553963128477335, + 0.004333485849201679, + 0.0050431773997843266, + 0.006985348649322987, + 0.0044560180976986885, + 0.005083226133137941, + 0.00393318198621273, + 0.006064752582460642, + 0.01517157256603241, + 0.028484171256422997, + 0.01736736297607422, + 0.024830028414726257, + 0.013043278828263283, + 0.016105400398373604, + 0.016755754128098488, + 0.018495634198188782, + 0.02346789836883545, + 0.016482064500451088, + 0.025481997057795525, + 0.022671997547149658, + 0.01930648274719715, + 0.021819671615958214, + 0.02522602118551731, + 0.02999848499894142, + 0.008239789865911007, + 0.005946957040578127, + 0.010922936722636223, + 0.009777897968888283, + 0.006756501272320747, + 0.007405797950923443, + 0.008910136297345161, + 0.00791141763329506, + 0.007994200102984905, + 0.011120758019387722, + 0.004431920126080513, + 0.005849877838045359, + 0.006055554375052452, + 0.003713893471285701, + 0.004393678158521652, + 0.006309109274297953, + 0.009893170557916164, + 0.005587245803326368, + 0.008640835992991924, + 0.007459891028702259, + 0.006114637013524771, + 0.00841798447072506, + 0.01281319186091423, + 0.007384685333818197, + 0.007307851687073708, + 0.003217251505702734, + 0.004829741083085537, + 0.005319411400705576, + 0.004687400069087744, + 0.004653502255678177, + 0.011801977641880512, + 0.00575460959225893, + 0.011823460459709167, + 0.007593108806759119, + 0.007802161853760481, + 0.015355669893324375, + 0.007969609461724758, + 0.010794240050017834, + 0.013583713211119175, + 0.009562172926962376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.03307054936885834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761656165122986 + ], + "opposite_logprob_tokens": [ + -0.8261656165122986 + ], + "logprob_reward": -0.5761656165122986, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761656165122986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22259175777435303, + 0.07595159858465195, + 0.13734154403209686, + 0.05658698454499245, + 0.07319971919059753, + 0.1670922487974167, + 0.09707223623991013, + 0.05015867203474045, + 0.05040089413523674, + 0.06960436701774597 + ], + "credit_map": [ + 0.03344527631998062, + 0.03306916356086731, + 0.023217666894197464, + 0.03040400892496109, + 0.025368863716721535, + 0.0220490749925375, + 0.022443359717726707, + 0.032594360411167145, + 0.01366062369197607, + 0.008868982084095478, + 0.006209959276020527, + 0.012622899375855923, + 0.010901469737291336, + 0.006658794358372688, + 0.008158823475241661, + 0.008870047517120838, + 0.018772071227431297, + 0.013000532984733582, + 0.01658189482986927, + 0.024869108572602272, + 0.01818705163896084, + 0.014551397413015366, + 0.016222966834902763, + 0.015156512148678303, + 0.006536959670484066, + 0.005893743131309748, + 0.00811838824301958, + 0.011063524521887302, + 0.004246946424245834, + 0.005813938565552235, + 0.00613798014819622, + 0.008775503374636173, + 0.011609788984060287, + 0.00766008859500289, + 0.010685560293495655, + 0.010455405339598656, + 0.008224154822528362, + 0.008603475987911224, + 0.007313623558729887, + 0.008647622540593147, + 0.021192241460084915, + 0.017345402389764786, + 0.024155735969543457, + 0.02141587808728218, + 0.013131540268659592, + 0.01999303698539734, + 0.017843930050730705, + 0.03201448172330856, + 0.012141217477619648, + 0.022451236844062805, + 0.01420776266604662, + 0.008836028166115284, + 0.009050811640918255, + 0.006668564397841692, + 0.008119715377688408, + 0.015596896409988403, + 0.008486900478601456, + 0.005816622171550989, + 0.0051615070551633835, + 0.004443398676812649, + 0.006233804393559694, + 0.00428800331428647, + 0.010143851861357689, + 0.005584585014730692, + 0.010054162703454494, + 0.004080315586179495, + 0.0049567073583602905, + 0.0046180253848433495, + 0.003319328650832176, + 0.004038553684949875, + 0.012251616455614567, + 0.0070821852423250675, + 0.008190137334167957, + 0.0050734905526041985, + 0.007363433018326759, + 0.014758793637156487, + 0.0064588687382638454, + 0.005475042387843132, + 0.013602515682578087, + 0.008682086132466793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.5761656165122986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656476624310017 + ], + "opposite_logprob_tokens": [ + -4.7586565017700195 + ], + "logprob_reward": -0.008656476624310017, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656476624310017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09239768236875534, + 0.10280326753854752, + 0.1593713015317917, + 0.03635134920477867, + 0.09669896215200424, + 0.07516665756702423, + 0.04901915788650513, + 0.24916669726371765, + 0.06613937765359879, + 0.07288548350334167 + ], + "credit_map": [ + 0.015993302688002586, + 0.011834397912025452, + 0.0078438026830554, + 0.009975649416446686, + 0.012249824590981007, + 0.009238261729478836, + 0.01565885730087757, + 0.009603599086403847, + 0.012913437560200691, + 0.010820173658430576, + 0.010827944613993168, + 0.01669718325138092, + 0.01867094077169895, + 0.009772239252924919, + 0.013787529431283474, + 0.009313835762441158, + 0.02049386128783226, + 0.016177522018551826, + 0.020916320383548737, + 0.02491186559200287, + 0.019976438954472542, + 0.017510078847408295, + 0.021538665518164635, + 0.017846571281552315, + 0.0039607444778084755, + 0.003511432558298111, + 0.006351579446345568, + 0.0049944110214710236, + 0.003686192212626338, + 0.006222161930054426, + 0.004152612295001745, + 0.003472219919785857, + 0.0125504145398736, + 0.007620350457727909, + 0.01392357051372528, + 0.01259704027324915, + 0.007441691588610411, + 0.01621847413480282, + 0.010752742178738117, + 0.015594694763422012, + 0.0077206725254654884, + 0.008325349539518356, + 0.0099549675360322, + 0.01127021573483944, + 0.007630590815097094, + 0.010128305293619633, + 0.009654002264142036, + 0.010482565499842167, + 0.007679486647248268, + 0.01064483355730772, + 0.004984022583812475, + 0.005192372016608715, + 0.005633245687931776, + 0.0037348177284002304, + 0.0050329104997217655, + 0.0061174724251031876, + 0.04145677387714386, + 0.030934898182749748, + 0.01671820692718029, + 0.026006413623690605, + 0.030828921124339104, + 0.02104059047996998, + 0.05647222697734833, + 0.025708673521876335, + 0.01456238143146038, + 0.006164167076349258, + 0.005592739675194025, + 0.007050022482872009, + 0.004761161282658577, + 0.005468850489705801, + 0.0149231543764472, + 0.0076169115491211414, + 0.010975224897265434, + 0.0053304219618439674, + 0.006253205705434084, + 0.01350583229213953, + 0.007153893820941448, + 0.006261700298637152, + 0.015143802389502525, + 0.008261406794190407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.008656476624310017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200207905843854 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.002200207905843854, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200207905843854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11200045794248581, + 0.26715022325515747, + 0.03478478267788887, + 0.062031518667936325, + 0.1459880918264389, + 0.06952665746212006, + 0.06408026069402695, + 0.16007326543331146, + 0.03145907446742058, + 0.0529056191444397 + ], + "credit_map": [ + 0.02258463017642498, + 0.013614354655146599, + 0.011463354341685772, + 0.009466739371418953, + 0.014774946495890617, + 0.011156048625707626, + 0.01155561488121748, + 0.01738477125763893, + 0.056139688938856125, + 0.039593569934368134, + 0.013554071076214314, + 0.024840878322720528, + 0.06317702680826187, + 0.016090191900730133, + 0.027941003441810608, + 0.025813786312937737, + 0.004820553585886955, + 0.004071102477610111, + 0.002641774248331785, + 0.005901883821934462, + 0.005400196649134159, + 0.0035511155147105455, + 0.003970086574554443, + 0.004428069107234478, + 0.00566204683855176, + 0.005930871702730656, + 0.009228430688381195, + 0.013911750167608261, + 0.005406474694609642, + 0.007650291547179222, + 0.007570987567305565, + 0.006670662201941013, + 0.022883616387844086, + 0.012176429852843285, + 0.021150438115000725, + 0.015987318009138107, + 0.012713911943137646, + 0.02364584244787693, + 0.016042251139879227, + 0.02138829231262207, + 0.007289493922144175, + 0.012168897315859795, + 0.007880302146077156, + 0.009945288300514221, + 0.00659251119941473, + 0.010755711235105991, + 0.0060104671865701675, + 0.008883990347385406, + 0.007446156349033117, + 0.005964803975075483, + 0.005775436293333769, + 0.009189562872052193, + 0.006604159716516733, + 0.010704588145017624, + 0.0063253529369831085, + 0.012070209719240665, + 0.026242995634675026, + 0.01834600418806076, + 0.009790941141545773, + 0.015288167633116245, + 0.02052438072860241, + 0.0123396385461092, + 0.04038999229669571, + 0.01715114712715149, + 0.004701104946434498, + 0.0034162476658821106, + 0.003431155579164624, + 0.0052158464677631855, + 0.003123115049675107, + 0.003770769340917468, + 0.004024071618914604, + 0.0037767658941447735, + 0.005978076718747616, + 0.005555594805628061, + 0.005978735629469156, + 0.009302175603806973, + 0.006010814104229212, + 0.005496130790561438, + 0.008921526372432709, + 0.005662564188241959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.002200207905843854 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866704344749451 + ], + "opposite_logprob_tokens": [ + -3.0486669540405273 + ], + "logprob_reward": -0.04866704344749451, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866704344749451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10525373369455338, + 0.27814623713493347, + 0.18130502104759216, + 0.0985201820731163, + 0.11711142957210541, + 0.03045119158923626, + 0.037216976284980774, + 0.06581079214811325, + 0.05285192281007767, + 0.03333253413438797 + ], + "credit_map": [ + 0.015168760903179646, + 0.012072904035449028, + 0.01308966614305973, + 0.012051278725266457, + 0.011039696633815765, + 0.010622375644743443, + 0.013578698970377445, + 0.017630353569984436, + 0.05722057446837425, + 0.04324981942772865, + 0.013398364186286926, + 0.024883458390831947, + 0.06665048748254776, + 0.01693631522357464, + 0.02947227843105793, + 0.026334941387176514, + 0.022533245384693146, + 0.019222354516386986, + 0.020462900400161743, + 0.03011273220181465, + 0.02551659755408764, + 0.020281318575143814, + 0.02411106415092945, + 0.019064804539084435, + 0.011483386158943176, + 0.019253959879279137, + 0.012317161075770855, + 0.011839041486382484, + 0.011411369778215885, + 0.011205830611288548, + 0.00983674917370081, + 0.011172686703503132, + 0.011150228790938854, + 0.011770574375987053, + 0.021421579644083977, + 0.011762102134525776, + 0.011386029422283173, + 0.0167861245572567, + 0.012746681459248066, + 0.02008812129497528, + 0.004200194962322712, + 0.0027780032251030207, + 0.003818655386567116, + 0.0031603830866515636, + 0.003428520867601037, + 0.00601195078343153, + 0.003073709085583687, + 0.003979776985943317, + 0.005093834828585386, + 0.004853691440075636, + 0.0040740640833973885, + 0.004211840685456991, + 0.0035558114759624004, + 0.0063477326184511185, + 0.003010139334946871, + 0.006069862749427557, + 0.010143667459487915, + 0.007588774431496859, + 0.0072682746686041355, + 0.007910165935754776, + 0.0076011125929653645, + 0.006776982918381691, + 0.010313370265066624, + 0.008208447135984898, + 0.010051739402115345, + 0.005043379031121731, + 0.00534922955557704, + 0.006167879793792963, + 0.004457955248653889, + 0.006057548802345991, + 0.00992483738809824, + 0.005799349397420883, + 0.0039121308363974094, + 0.003285961225628853, + 0.00397855369374156, + 0.004940500948578119, + 0.0045268721878528595, + 0.004594969097524881, + 0.0037122529465705156, + 0.004381296690553427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.04866704344749451 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme close-up focuses on a piece of weathered wood, its surface covered in cracked and flaking sky-blue paint, laid against the dark, grainy texture of asphalt pavement. The scene, baked in the harsh, direct midday sun, causes a visible heat haze to flicker above the ground. The camera remains perfectly still, observing as subtle fluctuations begin to disturb the flat, matte surface of the paint. Driven by the hot, expanding gases trapped beneath, the paint begins to expand upward from a central point. The blue layer stretches relentlessly, its color thinning slightly to form a smooth, tense dome. The process culminates in a stable, convex bubble that now sits prominently on the wood's surface. Its perfectly hemispherical shape captures the bright, mirror-like highlights of the sunlight and casts a tiny, sharp shadow on the chipped paint nearby, creating a quietly mesmerizing atmosphere of intense heat and slow transformation.", + "seed": 9739, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_14/index_581_seed_9739_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05945914809126407, + "vlm_logprob": -0.05945914809126407, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.05945914809126407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0302268173545599 + ], + "opposite_logprob_tokens": [ + -4.280226707458496 + ], + "logprob_reward": -0.0302268173545599, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0302268173545599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27652037143707275, + 0.0923653393983841, + 0.021470604464411736, + 0.05346858873963356, + 0.02343139424920082, + 0.046612825244665146, + 0.09126009792089462, + 0.33964335918426514, + 0.03239833191037178, + 0.022829120978713036 + ], + "credit_map": [ + 0.07106266915798187, + 0.10204426199197769, + 0.013091696426272392, + 0.01985539123415947, + 0.03307848423719406, + 0.018386587500572205, + 0.00711592473089695, + 0.011885332874953747, + 0.013230212964117527, + 0.012276658788323402, + 0.01333237811923027, + 0.0052703809924423695, + 0.008003508672118187, + 0.020334487780928612, + 0.006188674829900265, + 0.013729036785662174, + 0.004455271176993847, + 0.003398844739422202, + 0.002921727718785405, + 0.0020392087753862143, + 0.0016670676413923502, + 0.0035671480000019073, + 0.0013230947079136968, + 0.002098241588100791, + 0.006247723940759897, + 0.006471176631748676, + 0.004345929715782404, + 0.0052530900575220585, + 0.007956343702971935, + 0.009917840361595154, + 0.007461200933903456, + 0.0058152806013822556, + 0.003634362481534481, + 0.0026391600258648396, + 0.0031948473770171404, + 0.002086387248709798, + 0.002766924910247326, + 0.003804712323471904, + 0.0027474185917526484, + 0.0025575808249413967, + 0.006289217155426741, + 0.004886372480541468, + 0.0038169838953763247, + 0.00278224959038198, + 0.00459660030901432, + 0.014419392682611942, + 0.005147865507751703, + 0.004674144089221954, + 0.01524081826210022, + 0.010481229983270168, + 0.005568172316998243, + 0.0060300384648144245, + 0.008499953895807266, + 0.03279958292841911, + 0.006159723270684481, + 0.006480575539171696, + 0.08830951154232025, + 0.12775398790836334, + 0.010275891982018948, + 0.00980722438544035, + 0.0542534701526165, + 0.028577156364917755, + 0.011937981471419334, + 0.008728149347007275, + 0.004883996676653624, + 0.0033518834970891476, + 0.002438717056065798, + 0.0026720641180872917, + 0.005995609797537327, + 0.005582372657954693, + 0.003481165738776326, + 0.003992518875747919, + 0.002937235403805971, + 0.0018162510823458433, + 0.003918735310435295, + 0.0015047913184389472, + 0.003865377278998494, + 0.0037069208920001984, + 0.0024842158891260624, + 0.0025955948513001204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0302268173545599 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256295531988144 + ], + "opposite_logprob_tokens": [ + -1.6006295680999756 + ], + "logprob_reward": -0.2256295531988144, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256295531988144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17839650809764862, + 0.0645410418510437, + 0.0867924615740776, + 0.1426311880350113, + 0.12598443031311035, + 0.16010864078998566, + 0.04216283559799194, + 0.10043709725141525, + 0.06813281029462814, + 0.030812975019216537 + ], + "credit_map": [ + 0.03626738861203194, + 0.04118739813566208, + 0.01683861017227173, + 0.020290184766054153, + 0.014260167255997658, + 0.016794566065073013, + 0.013615400530397892, + 0.019142799079418182, + 0.011292747221887112, + 0.006972923409193754, + 0.00874208565801382, + 0.004965322092175484, + 0.007154144812375307, + 0.011668113060295582, + 0.0061942716129124165, + 0.007551438175141811, + 0.0196990929543972, + 0.00822365190833807, + 0.013327624648809433, + 0.006076639983803034, + 0.006842771079391241, + 0.02016715332865715, + 0.004891875199973583, + 0.007563650608062744, + 0.02673807181417942, + 0.015896150842308998, + 0.017543181777000427, + 0.011930464766919613, + 0.013059474527835846, + 0.02859555184841156, + 0.013519438914954662, + 0.015348863787949085, + 0.021102819591760635, + 0.01235765591263771, + 0.01668047532439232, + 0.012725530192255974, + 0.013846109621226788, + 0.022450799122452736, + 0.014060349203646183, + 0.012760692276060581, + 0.024109644815325737, + 0.03260132670402527, + 0.014255160465836525, + 0.011555670760571957, + 0.021343547850847244, + 0.01876484602689743, + 0.021510470658540726, + 0.015967968851327896, + 0.006843386683613062, + 0.005610693711787462, + 0.0026998822577297688, + 0.002882765606045723, + 0.004409224726259708, + 0.013151764869689941, + 0.0031536926981061697, + 0.0034114266745746136, + 0.024972375482320786, + 0.035567719489336014, + 0.0035750323440879583, + 0.002660543192178011, + 0.015491549856960773, + 0.009694884531199932, + 0.003437646431848407, + 0.005037342663854361, + 0.011694475077092648, + 0.005816314835101366, + 0.014109163545072079, + 0.004633844364434481, + 0.009017525240778923, + 0.010787403210997581, + 0.005497749429196119, + 0.006576329469680786, + 0.004591334145516157, + 0.003099072026088834, + 0.0032896283082664013, + 0.002240808680653572, + 0.003924751188606024, + 0.005704635754227638, + 0.004649328533560038, + 0.003313417313620448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.2256295531988144 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008237544680014253 + ], + "opposite_logprob_tokens": [ + -7.125823974609375 + ], + "logprob_reward": -0.0008237544680014253, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008237544680014253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34395843744277954, + 0.10459213703870773, + 0.0249499399214983, + 0.08329591155052185, + 0.029037848114967346, + 0.03502172976732254, + 0.06507286429405212, + 0.1857246607542038, + 0.08144649118185043, + 0.046900056302547455 + ], + "credit_map": [ + 0.09608062356710434, + 0.12835830450057983, + 0.013977889902889729, + 0.02641225978732109, + 0.03649972379207611, + 0.01791883446276188, + 0.010095903649926186, + 0.014614858664572239, + 0.014553922228515148, + 0.016820430755615234, + 0.01971222274005413, + 0.007828159257769585, + 0.008148826658725739, + 0.018358157947659492, + 0.008182770572602749, + 0.010987634770572186, + 0.0036856881342828274, + 0.002830956596881151, + 0.003587479004636407, + 0.0024930343497544527, + 0.0027974932454526424, + 0.0035255427937954664, + 0.0033486736938357353, + 0.0026810697745531797, + 0.009774976409971714, + 0.009347796440124512, + 0.008337512612342834, + 0.009242560714483261, + 0.011446538381278515, + 0.014841494150459766, + 0.011487252078950405, + 0.00881777610629797, + 0.004855033475905657, + 0.0032496617641299963, + 0.005284793209284544, + 0.002328702248632908, + 0.003473462536931038, + 0.004797530360519886, + 0.002394335810095072, + 0.002654326381161809, + 0.00600920245051384, + 0.006501853000372648, + 0.004118714947253466, + 0.0022919701877981424, + 0.0035497487988322973, + 0.00505058653652668, + 0.0034109887201339006, + 0.004088662564754486, + 0.009690592996776104, + 0.008310453966259956, + 0.008327442221343517, + 0.005445417482405901, + 0.008751274086534977, + 0.01380566693842411, + 0.005633969325572252, + 0.005108033772557974, + 0.04434899613261223, + 0.06030300632119179, + 0.0075875623151659966, + 0.006354553159326315, + 0.02780774235725403, + 0.022879676893353462, + 0.008752567693591118, + 0.007690547965466976, + 0.013648980297148228, + 0.007984672673046589, + 0.010204972699284554, + 0.00606312183663249, + 0.013339503668248653, + 0.013050923123955727, + 0.0075559476390480995, + 0.009598360396921635, + 0.007208989001810551, + 0.004886014387011528, + 0.006987981032580137, + 0.004492451436817646, + 0.006515901070088148, + 0.0071475692093372345, + 0.005319522228091955, + 0.004341623280197382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0008237544680014253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491348423063755 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.002491348423063755, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491348423063755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31987428665161133, + 0.08148452639579773, + 0.07862255722284317, + 0.11906856298446655, + 0.04757881164550781, + 0.06991341710090637, + 0.05585043132305145, + 0.06470529735088348, + 0.07429485768079758, + 0.0886072888970375 + ], + "credit_map": [ + 0.0841243639588356, + 0.11676386743783951, + 0.01326906681060791, + 0.024970518425107002, + 0.03908957540988922, + 0.020638296380639076, + 0.009511886164546013, + 0.01150669064372778, + 0.010605156421661377, + 0.0104755824431777, + 0.010604416951537132, + 0.006852590013295412, + 0.009237096644937992, + 0.010397341102361679, + 0.007961893454194069, + 0.01535044889897108, + 0.009639162570238113, + 0.01044569630175829, + 0.0068597435019910336, + 0.008146465755999088, + 0.009511439129710197, + 0.013057636097073555, + 0.009019548073410988, + 0.011942866258323193, + 0.014387753792107105, + 0.01368451863527298, + 0.007508605718612671, + 0.011652583256363869, + 0.016045819967985153, + 0.024503909051418304, + 0.016881946474313736, + 0.014403427951037884, + 0.006108446046710014, + 0.004335302393883467, + 0.005904864985495806, + 0.003244833555072546, + 0.006447634659707546, + 0.006698169279843569, + 0.00504516763612628, + 0.009794393554329872, + 0.012082891538739204, + 0.014799851924180984, + 0.006558142136782408, + 0.005360982846468687, + 0.008770432323217392, + 0.008325866423547268, + 0.008654692210257053, + 0.0053605553694069386, + 0.007511960342526436, + 0.009490982629358768, + 0.005635226611047983, + 0.004045318346470594, + 0.008296415209770203, + 0.01142219640314579, + 0.004970026668161154, + 0.0044783055782318115, + 0.008367221802473068, + 0.007783658802509308, + 0.006378963589668274, + 0.004187239799648523, + 0.011985006742179394, + 0.009269828908145428, + 0.006590310484170914, + 0.010143065825104713, + 0.011346607469022274, + 0.008759329095482826, + 0.008262940682470798, + 0.007119353394955397, + 0.011454861611127853, + 0.011648726649582386, + 0.007831607945263386, + 0.007871435955166817, + 0.01234526839107275, + 0.011103428900241852, + 0.009336821734905243, + 0.008277073502540588, + 0.015882808715105057, + 0.010064318776130676, + 0.011468801647424698, + 0.010128763504326344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.002491348423063755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038124267011880875 + ], + "opposite_logprob_tokens": [ + -3.2881243228912354 + ], + "logprob_reward": -0.038124267011880875, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038124267011880875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5196521878242493, + 0.061857983469963074, + 0.05327191576361656, + 0.20020543038845062, + 0.028380023315548897, + 0.03868855535984039, + 0.033823393285274506, + 0.01822931505739689, + 0.025449609383940697, + 0.020441625267267227 + ], + "credit_map": [ + 0.14455220103263855, + 0.19512978196144104, + 0.023005185648798943, + 0.03920940309762955, + 0.054505761712789536, + 0.02748183347284794, + 0.011925621889531612, + 0.023842327296733856, + 0.008380822837352753, + 0.007601730991154909, + 0.004793804604560137, + 0.003987809643149376, + 0.007978749461472034, + 0.013170238584280014, + 0.005460036918520927, + 0.010484781116247177, + 0.005908560939133167, + 0.00730479508638382, + 0.00391771737486124, + 0.005438623484224081, + 0.006817464716732502, + 0.011139949783682823, + 0.004917852580547333, + 0.007826948538422585, + 0.026035521179437637, + 0.023145781829953194, + 0.020604737102985382, + 0.02227761037647724, + 0.027351252734661102, + 0.03431076556444168, + 0.02720026858150959, + 0.019279463216662407, + 0.00392672186717391, + 0.0027034603990614414, + 0.0033334342297166586, + 0.002172002801671624, + 0.002871482167392969, + 0.005582962650805712, + 0.002833455102518201, + 0.004956502467393875, + 0.007732695434242487, + 0.00839812308549881, + 0.002673571463674307, + 0.0026671660598367453, + 0.003991098143160343, + 0.007371812127530575, + 0.003682434093207121, + 0.002171653788536787, + 0.004345239140093327, + 0.004759521223604679, + 0.004449751228094101, + 0.0030631946865469217, + 0.005034095607697964, + 0.004868585616350174, + 0.0036001114640384912, + 0.0037028903607279062, + 0.002051515271887183, + 0.00266490806825459, + 0.003006639890372753, + 0.0013258806429803371, + 0.00241417670622468, + 0.0039025009609758854, + 0.0013123616809025407, + 0.0015513302059844136, + 0.00428532250225544, + 0.0029142177663743496, + 0.002003457397222519, + 0.001785216387361288, + 0.004291191231459379, + 0.005108700133860111, + 0.0021376365330070257, + 0.002923864871263504, + 0.0031737028621137142, + 0.001552349072881043, + 0.002512129256501794, + 0.0017309561371803284, + 0.002647518878802657, + 0.0034687255974859, + 0.0020190614741295576, + 0.0033371797762811184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.038124267011880875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme close-up focuses on a piece of weathered wood, its surface covered in cracked and flaking sky-blue paint, laid against the dark, grainy texture of asphalt pavement. The scene, baked in the harsh, direct midday sun, causes a visible heat haze to flicker above the ground. The camera remains perfectly still, observing as subtle fluctuations begin to disturb the flat, matte surface of the paint. Driven by the hot, expanding gases trapped beneath, the paint begins to expand upward from a central point. The blue layer stretches relentlessly, its color thinning slightly to form a smooth, tense dome. The process culminates in a stable, convex bubble that now sits prominently on the wood's surface. Its perfectly hemispherical shape captures the bright, mirror-like highlights of the sunlight and casts a tiny, sharp shadow on the chipped paint nearby, creating a quietly mesmerizing atmosphere of intense heat and slow transformation.", + "seed": 9740, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_14/index_581_seed_9740_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1343701643869281, + "vlm_logprob": -0.1343701643869281, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.1343701643869281, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10281050205230713 + ], + "opposite_logprob_tokens": [ + -2.4778103828430176 + ], + "logprob_reward": -0.10281050205230713, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10281050205230713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10044516623020172, + 0.11804535239934921, + 0.0721684992313385, + 0.037550125271081924, + 0.41603660583496094, + 0.09878817200660706, + 0.055100712925195694, + 0.02540726400911808, + 0.014522924087941647, + 0.06193515285849571 + ], + "credit_map": [ + 0.011614890769124031, + 0.02698393166065216, + 0.008532078936696053, + 0.006573358550667763, + 0.008882337249815464, + 0.02242024429142475, + 0.005005477461963892, + 0.010432844050228596, + 0.024198150262236595, + 0.015654750168323517, + 0.02559397742152214, + 0.011115290224552155, + 0.009036170318722725, + 0.011459133587777615, + 0.009875152260065079, + 0.011112731881439686, + 0.011359619908034801, + 0.013610505498945713, + 0.009708059951663017, + 0.007106227800250053, + 0.006431201938539743, + 0.0074418578296899796, + 0.008111132308840752, + 0.008399897255003452, + 0.007021637633442879, + 0.003829267108812928, + 0.006025651469826698, + 0.002767072292044759, + 0.005105937365442514, + 0.0043174191378057, + 0.004041632637381554, + 0.004441507160663605, + 0.07112981379032135, + 0.06191950663924217, + 0.05620252713561058, + 0.03216497227549553, + 0.038803402334451675, + 0.04366690292954445, + 0.06616596132516861, + 0.04598354920744896, + 0.01720188371837139, + 0.015077893622219563, + 0.010278301313519478, + 0.00927516259253025, + 0.014108614064753056, + 0.010913596488535404, + 0.008902556262910366, + 0.01303016021847725, + 0.009780045598745346, + 0.00821650866419077, + 0.005637410096824169, + 0.00501763541251421, + 0.006992332171648741, + 0.007334768772125244, + 0.004513838794082403, + 0.007608172483742237, + 0.0033570341765880585, + 0.003983348608016968, + 0.0023986513260751963, + 0.0024133888073265553, + 0.004054133780300617, + 0.003052736632525921, + 0.0018445155583322048, + 0.004303454887121916, + 0.002642784034833312, + 0.0020588010083884, + 0.0011789773125201464, + 0.0016097635962069035, + 0.001817893935367465, + 0.0015758099034428596, + 0.001218051416799426, + 0.002420843346044421, + 0.00960896909236908, + 0.010866070166230202, + 0.004000446759164333, + 0.0033225531224161386, + 0.008460183627903461, + 0.009628145024180412, + 0.00619922112673521, + 0.00984956230968237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.10281050205230713 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487025201320648 + ], + "opposite_logprob_tokens": [ + -1.2237025499343872 + ], + "logprob_reward": -0.3487025201320648, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3487025201320648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5867348909378052, + 0.039282266050577164, + 0.023043116554617882, + 0.01663786917924881, + 0.16491514444351196, + 0.010995051823556423, + 0.027130095288157463, + 0.020379114896059036, + 0.10019362717866898, + 0.010688859038054943 + ], + "credit_map": [ + 0.04134565591812134, + 0.23770339787006378, + 0.013849650509655476, + 0.02424430660903454, + 0.047644294798374176, + 0.18686626851558685, + 0.009905096143484116, + 0.02517624758183956, + 0.006933082360774279, + 0.004067704081535339, + 0.005957091227173805, + 0.004329290706664324, + 0.004860996268689632, + 0.0038890759460628033, + 0.004140305332839489, + 0.005104723386466503, + 0.003453334793448448, + 0.002673765644431114, + 0.004288821015506983, + 0.0025802759919315577, + 0.0023518064990639687, + 0.00199378514662385, + 0.0036067997571080923, + 0.002094528404995799, + 0.0018528607906773686, + 0.0025244264397770166, + 0.0013430010294541717, + 0.0013457178138196468, + 0.003612631233409047, + 0.0020819250494241714, + 0.0026270963717252016, + 0.0012502102181315422, + 0.02930738590657711, + 0.019133729860186577, + 0.023883379995822906, + 0.012880382128059864, + 0.015148458071053028, + 0.017029156908392906, + 0.02756591886281967, + 0.01996673084795475, + 0.0016519530909135938, + 0.0010820659808814526, + 0.0019695605151355267, + 0.0008795196190476418, + 0.001542996265925467, + 0.0014799064956605434, + 0.0008497221861034632, + 0.001539328950457275, + 0.003981376998126507, + 0.004415910691022873, + 0.0018503193277865648, + 0.0029283470939844847, + 0.0031209627632051706, + 0.00410345196723938, + 0.002689291024580598, + 0.004040437284857035, + 0.003891019383445382, + 0.002216465538367629, + 0.0017036786302924156, + 0.0017206199700012803, + 0.002920488128438592, + 0.0025788038037717342, + 0.0019111904548481107, + 0.0034368515480309725, + 0.012611327692866325, + 0.012424941174685955, + 0.008094140328466892, + 0.009537598118185997, + 0.01859975792467594, + 0.01492034550756216, + 0.007743757218122482, + 0.016261762008070946, + 0.0018745751585811377, + 0.0016975777689367533, + 0.000771646446082741, + 0.00047733093379065394, + 0.00128272850997746, + 0.0019059821497648954, + 0.0007563662366010249, + 0.0019226525910198689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.3487025201320648 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093958280980587 + ], + "opposite_logprob_tokens": [ + -4.511094093322754 + ], + "logprob_reward": -0.011093958280980587, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093958280980587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09399113059043884, + 0.04989456757903099, + 0.06148878484964371, + 0.03730723261833191, + 0.36255720257759094, + 0.04573049768805504, + 0.08148058503866196, + 0.11210432648658752, + 0.08971702307462692, + 0.06572866439819336 + ], + "credit_map": [ + 0.013247072696685791, + 0.01805594563484192, + 0.009134330786764622, + 0.008356171660125256, + 0.011490714736282825, + 0.016659028828144073, + 0.008587747812271118, + 0.008460119366645813, + 0.006137680262327194, + 0.005895708221942186, + 0.006106711458414793, + 0.007327633909881115, + 0.00715449545532465, + 0.005286604631692171, + 0.0073003885336220264, + 0.0046853478997945786, + 0.007934837602078915, + 0.012250596657395363, + 0.005283620674163103, + 0.007839934900403023, + 0.0066092354245483875, + 0.0056949183344841, + 0.007427004165947437, + 0.008448634296655655, + 0.0070732939057052135, + 0.00398766016587615, + 0.005546732805669308, + 0.00267318869009614, + 0.007018362171947956, + 0.004256157670170069, + 0.003820994636043906, + 0.002930839778855443, + 0.04177569970488548, + 0.07915804535150528, + 0.03242098167538643, + 0.03018566034734249, + 0.04766988009214401, + 0.03758934885263443, + 0.05827823281288147, + 0.0354793407022953, + 0.007456157356500626, + 0.007754932157695293, + 0.003972149919718504, + 0.004806495737284422, + 0.006707430817186832, + 0.004743230063468218, + 0.005345701705664396, + 0.004944400861859322, + 0.011002155020833015, + 0.014565086923539639, + 0.006573045160621405, + 0.0075998990796506405, + 0.006589676719158888, + 0.013862461782991886, + 0.008939210325479507, + 0.012349048629403114, + 0.016066793352365494, + 0.01882123202085495, + 0.011729828082025051, + 0.010665763169527054, + 0.022380193695425987, + 0.012506227940320969, + 0.006608334835618734, + 0.013325957581400871, + 0.00872139260172844, + 0.017808031290769577, + 0.004563031252473593, + 0.006804264150559902, + 0.023692425340414047, + 0.010430957190692425, + 0.00670405151322484, + 0.010992869734764099, + 0.008279284462332726, + 0.014622680842876434, + 0.004150169901549816, + 0.004855634178966284, + 0.00868706963956356, + 0.011803878471255302, + 0.004641994833946228, + 0.008687952533364296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.011093958280980587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637940347194672 + ], + "opposite_logprob_tokens": [ + -4.882637977600098 + ], + "logprob_reward": -0.007637940347194672, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637940347194672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31926360726356506, + 0.07037841528654099, + 0.07264935970306396, + 0.04221979156136513, + 0.23939095437526703, + 0.028558216989040375, + 0.03706212341785431, + 0.03459220007061958, + 0.10306096076965332, + 0.05282445251941681 + ], + "credit_map": [ + 0.02347678691148758, + 0.12483317404985428, + 0.009497102349996567, + 0.013357499614357948, + 0.02649584226310253, + 0.09890774637460709, + 0.00702753895893693, + 0.015667911618947983, + 0.0073133413679897785, + 0.007660074159502983, + 0.009655880741775036, + 0.009177894331514835, + 0.009843126870691776, + 0.007322714664041996, + 0.010334398597478867, + 0.009070977568626404, + 0.009649734012782574, + 0.013101286254823208, + 0.011136953718960285, + 0.008317840285599232, + 0.00764300674200058, + 0.0069207120686769485, + 0.009068948216736317, + 0.006810876540839672, + 0.00948207639157772, + 0.004904427099972963, + 0.008001490496098995, + 0.004378600977361202, + 0.004137134645134211, + 0.00400200393050909, + 0.0033276767935603857, + 0.003986380062997341, + 0.04504460096359253, + 0.0326039083302021, + 0.03755396604537964, + 0.020612874999642372, + 0.026155686005949974, + 0.02496284991502762, + 0.026730835437774658, + 0.02572622522711754, + 0.005640479736030102, + 0.003118738532066345, + 0.0036528967320919037, + 0.002529868157580495, + 0.0031531236600130796, + 0.0033942093141376972, + 0.0029166247695684433, + 0.00415227422490716, + 0.0043638404458761215, + 0.005077798385173082, + 0.0033549380023032427, + 0.002618451602756977, + 0.005880871322005987, + 0.005040775053203106, + 0.005805476568639278, + 0.004919969476759434, + 0.005615388974547386, + 0.00475696986541152, + 0.005603875499218702, + 0.0029711073730140924, + 0.004631459712982178, + 0.0041031548753380775, + 0.0022637215442955494, + 0.00464652432128787, + 0.010248619131743908, + 0.019665833562612534, + 0.005794083699584007, + 0.008432970382273197, + 0.026066742837429047, + 0.012371270917356014, + 0.007648841477930546, + 0.012832599692046642, + 0.00655602291226387, + 0.00856196228414774, + 0.0037555203307420015, + 0.0036050535272806883, + 0.006025372538715601, + 0.010512054897844791, + 0.0037827161140739918, + 0.010025747120380402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.007637940347194672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, high-angle, extreme close-up of weathered wood with cracked sky-blue paint on asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by harsh, direct midday sun causing a visible heat haze to flicker above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paint expand upward from a central point to form a smooth, tense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final result show a stable, convex hemispherical bubble with mirror-like highlights and a sharp shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016059011220932 + ], + "opposite_logprob_tokens": [ + -1.7016059160232544 + ], + "logprob_reward": -0.2016059011220932, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016059011220932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4013856053352356, + 0.07881199568510056, + 0.04661080986261368, + 0.04581611230969429, + 0.3150807023048401, + 0.02351391315460205, + 0.011669706553220749, + 0.030371330678462982, + 0.01485314592719078, + 0.03188673034310341 + ], + "credit_map": [ + 0.03382270410656929, + 0.15665946900844574, + 0.016492851078510284, + 0.01690490171313286, + 0.030802756547927856, + 0.12180357426404953, + 0.007253037299960852, + 0.0176463034003973, + 0.009607892483472824, + 0.012041226960718632, + 0.009692471474409103, + 0.007656979374587536, + 0.009487715549767017, + 0.00878400169312954, + 0.00853902567178011, + 0.013002678751945496, + 0.006947105750441551, + 0.010412361472845078, + 0.004183144774287939, + 0.0038075868505984545, + 0.0052720652893185616, + 0.004957135301083326, + 0.0033152932301163673, + 0.0077161178924143314, + 0.01072749774903059, + 0.00411341805011034, + 0.009655618108808994, + 0.0039381845854222775, + 0.0039802053943276405, + 0.004671329632401466, + 0.00290400767698884, + 0.005825853906571865, + 0.056894004344940186, + 0.04088997095823288, + 0.04481719806790352, + 0.022632479667663574, + 0.030081938952207565, + 0.033857185393571854, + 0.048409003764390945, + 0.03749890625476837, + 0.004176383372396231, + 0.003533397102728486, + 0.0028152610175311565, + 0.001777283614501357, + 0.0030293241143226624, + 0.002572352299466729, + 0.002140806522220373, + 0.0034691058099269867, + 0.0014846815029159188, + 0.0016600692179054022, + 0.0012258944334462285, + 0.0013676631497219205, + 0.0017693958943709731, + 0.001476383302360773, + 0.001000395743176341, + 0.0016852236585691571, + 0.004535062704235315, + 0.004670914728194475, + 0.00389591371640563, + 0.0026652393862605095, + 0.005411720369011164, + 0.0033935236278921366, + 0.0018991973483934999, + 0.003899757284671068, + 0.003037472255527973, + 0.002045895904302597, + 0.0012850835919380188, + 0.0014318759785965085, + 0.0014439186779782176, + 0.0018327549332752824, + 0.0012352545745670795, + 0.002540889196097851, + 0.004661917220801115, + 0.005827196873724461, + 0.0027975160628557205, + 0.002016803715378046, + 0.0032848052214831114, + 0.004664482083171606, + 0.002904862631112337, + 0.0057291435077786446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.2016059011220932 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_15/index_tensor([2224, 2224])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_15/index_tensor([2224, 2224])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2d6f1f288cf8db2b67a99b15b71f33ff6059b485 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_15/index_tensor([2224, 2224])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in high-speed slow motion, a low-angle, medium shot pans steadily to the right, tracking the front wheel of a mountain bike as it speeds down a wet city alley. The bike's thick, knobby tires, caked with a thin layer of mud, deliberately cut a large, dark puddle into the dark, reflective asphalt. As the wheel tumbles through the water, it sprays outward and upward in spectacular fan-shaped splashes. Individual droplets, perfectly formed in the soft, diffuse light of an overcast sky, hang suspended in the air in a crystalline arc. The scene, dominated by a cool, desaturated palette of grays and browns against a backdrop of wet sidewalks and brick walls, makes the violent, dynamic action of the splashing water both raw and artistically captured, conveying a powerful sense of kinetic energy.", + "seed": 3709, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_15/index_2224_seed_3709_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040456851944327354, + "vlm_logprob": -0.040456851944327354, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.040456851944327354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10939044505357742 + ], + "opposite_logprob_tokens": [ + -2.8593904972076416 + ], + "logprob_reward": -0.10939044505357742, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.10939044505357742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.219927117228508, + 0.07290180027484894, + 0.038609132170677185, + 0.0926780104637146, + 0.05867768079042435, + 0.06879525631666183, + 0.19559608399868011, + 0.05397341027855873, + 0.1563575714826584, + 0.04248396307229996 + ], + "credit_map": [ + 0.01741756871342659, + 0.03005814179778099, + 0.026216084137558937, + 0.048520006239414215, + 0.02752814069390297, + 0.020702408626675606, + 0.02187621034681797, + 0.027608560398221016, + 0.008667622692883015, + 0.00818687304854393, + 0.007441014517098665, + 0.008846921846270561, + 0.008168363012373447, + 0.00894882995635271, + 0.015331270173192024, + 0.00731090409681201, + 0.0032662132289260626, + 0.002962229773402214, + 0.004935775883495808, + 0.0032868601847440004, + 0.003848912427201867, + 0.006355896592140198, + 0.009134690277278423, + 0.004818551242351532, + 0.016481375321745872, + 0.006515112239867449, + 0.013646912761032581, + 0.011466827243566513, + 0.011973111890256405, + 0.010082037188112736, + 0.009877231903374195, + 0.012635401450097561, + 0.006337333004921675, + 0.004161297343671322, + 0.015789858996868134, + 0.007010165136307478, + 0.0044283391907811165, + 0.005403570365160704, + 0.006923868786543608, + 0.008623247966170311, + 0.007455679588019848, + 0.004220867529511452, + 0.005569095723330975, + 0.0055507696233689785, + 0.004040739499032497, + 0.00493247527629137, + 0.02629188634455204, + 0.010733738541603088, + 0.011539828032255173, + 0.015869442373514175, + 0.032967668026685715, + 0.014403024688363075, + 0.022333752363920212, + 0.045190099626779556, + 0.02827194146811962, + 0.025020327419042587, + 0.004609975963830948, + 0.007515691686421633, + 0.010217340663075447, + 0.005359310191124678, + 0.005868272855877876, + 0.007029337342828512, + 0.004881769418716431, + 0.00849170982837677, + 0.01035680528730154, + 0.011448517441749573, + 0.04270375892519951, + 0.014339547604322433, + 0.008913073688745499, + 0.011015261523425579, + 0.02815311774611473, + 0.029427479952573776, + 0.005542901810258627, + 0.004961022641509771, + 0.004332030192017555, + 0.004646341782063246, + 0.0031970965210348368, + 0.004424629732966423, + 0.009020489640533924, + 0.0063594551756978035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.10939044505357742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054996833205223083 + ], + "opposite_logprob_tokens": [ + -2.9299967288970947 + ], + "logprob_reward": -0.054996833205223083, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054996833205223083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16504919528961182, + 0.047243453562259674, + 0.07126303762197495, + 0.15147991478443146, + 0.07142669707536697, + 0.060076743364334106, + 0.05273331701755524, + 0.07377272844314575, + 0.23057807981967926, + 0.07637684792280197 + ], + "credit_map": [ + 0.012384332716464996, + 0.035761795938014984, + 0.01987239345908165, + 0.02097633294761181, + 0.022130705416202545, + 0.012923354282975197, + 0.018261779099702835, + 0.022738507017493248, + 0.004118591081351042, + 0.004973154980689287, + 0.004586180206388235, + 0.004793517291545868, + 0.0036991529632359743, + 0.010474997572600842, + 0.008210715837776661, + 0.006387143395841122, + 0.009194125421345234, + 0.0059377350844442844, + 0.010982844047248363, + 0.007998602464795113, + 0.006773687433451414, + 0.013178423047065735, + 0.010085467249155045, + 0.007112151011824608, + 0.02241862565279007, + 0.009734680876135826, + 0.03519429266452789, + 0.018340496346354485, + 0.01248371321707964, + 0.017987987026572227, + 0.0214348454028368, + 0.013885270804166794, + 0.005036390386521816, + 0.004221352748572826, + 0.007604958489537239, + 0.013247660361230373, + 0.00654602749273181, + 0.0139976991340518, + 0.008228147402405739, + 0.012544463388621807, + 0.006564588285982609, + 0.0038774320855736732, + 0.0057044015266001225, + 0.007164000999182463, + 0.0034244698472321033, + 0.006683278828859329, + 0.01755746826529503, + 0.009101102128624916, + 0.0040390146896243095, + 0.007805828005075455, + 0.005583603400737047, + 0.007937581278383732, + 0.005700245965272188, + 0.008395323529839516, + 0.006618009880185127, + 0.006653711665421724, + 0.007375803310424089, + 0.008219380863010883, + 0.01056724414229393, + 0.008414926938712597, + 0.0075174542143940926, + 0.014185206964612007, + 0.008753449656069279, + 0.008739259093999863, + 0.009938728995621204, + 0.010023100301623344, + 0.033653661608695984, + 0.01499075535684824, + 0.010672015137970448, + 0.012690779753029346, + 0.07218657433986664, + 0.06642244756221771, + 0.00873135682195425, + 0.009869612753391266, + 0.009049315936863422, + 0.01270768977701664, + 0.007112090941518545, + 0.006615367252379656, + 0.010624285787343979, + 0.011667122133076191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.054996833205223083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075288876891136 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075288876891136, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075288876891136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18814504146575928, + 0.07363806664943695, + 0.09064334630966187, + 0.1768084317445755, + 0.021256309002637863, + 0.039405014365911484, + 0.06362250447273254, + 0.07283440232276917, + 0.24405023455619812, + 0.02959655597805977 + ], + "credit_map": [ + 0.013206027448177338, + 0.02064981497824192, + 0.04074845090508461, + 0.0258589256554842, + 0.01158632431179285, + 0.018898779526352882, + 0.021632671356201172, + 0.035564079880714417, + 0.004918145481497049, + 0.007959295995533466, + 0.0111527219414711, + 0.01104639284312725, + 0.0050282906740903854, + 0.010473090223968029, + 0.012886924669146538, + 0.010173219256103039, + 0.010262368246912956, + 0.010699030011892319, + 0.01245993934571743, + 0.014974702149629593, + 0.007624141871929169, + 0.009117020294070244, + 0.010193993337452412, + 0.015312163159251213, + 0.011346717365086079, + 0.005594427231699228, + 0.008146102540194988, + 0.00792289525270462, + 0.011449756100773811, + 0.017472654581069946, + 0.06393042951822281, + 0.0509454645216465, + 0.001851275679655373, + 0.0014340563211590052, + 0.005200986284762621, + 0.0018539458978921175, + 0.0018049184000119567, + 0.0036495565436780453, + 0.0025658023077994585, + 0.002895771525800228, + 0.0035316769499331713, + 0.0026256595738232136, + 0.005218228325247765, + 0.006252307910472155, + 0.003211060306057334, + 0.005149332340806723, + 0.006174321752041578, + 0.00724242627620697, + 0.004868923686444759, + 0.008742292411625385, + 0.006628471426665783, + 0.0070192888379096985, + 0.008150609210133553, + 0.01070520281791687, + 0.00927253719419241, + 0.008235185407102108, + 0.005563992075622082, + 0.00984061136841774, + 0.005357066635042429, + 0.008474956266582012, + 0.008088397793471813, + 0.012429218739271164, + 0.013832202181220055, + 0.009247957728803158, + 0.010702031664550304, + 0.009348887950181961, + 0.034615837037563324, + 0.018087876960635185, + 0.010201751254498959, + 0.012785917147994041, + 0.07745348662137985, + 0.07085447013378143, + 0.0026332815177738667, + 0.0024092928506433964, + 0.0030764900147914886, + 0.002471438143402338, + 0.0033768764697015285, + 0.004876189399510622, + 0.004842860158532858, + 0.00591012928634882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.016075288876891136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076333265751600266 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.0076333265751600266, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076333265751600266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18098041415214539, + 0.14543849229812622, + 0.08153633028268814, + 0.10513927787542343, + 0.040404535830020905, + 0.025673143565654755, + 0.0800727978348732, + 0.06398390233516693, + 0.24395327270030975, + 0.03281776234507561 + ], + "credit_map": [ + 0.019312571734189987, + 0.04480092227458954, + 0.015181876718997955, + 0.028447994962334633, + 0.022465087473392487, + 0.017306242138147354, + 0.015499123372137547, + 0.017966601997613907, + 0.01203632727265358, + 0.017492573708295822, + 0.019709503278136253, + 0.016447456553578377, + 0.01313659269362688, + 0.022633880376815796, + 0.026693537831306458, + 0.017288630828261375, + 0.011871054768562317, + 0.005934146232903004, + 0.00968074705451727, + 0.009651903994381428, + 0.005971578415483236, + 0.010304629802703857, + 0.010308124125003815, + 0.017814157530665398, + 0.011172164231538773, + 0.003675619838759303, + 0.00813157856464386, + 0.005923818331211805, + 0.007800361141562462, + 0.016270093619823456, + 0.0281590037047863, + 0.02400664985179901, + 0.0031505513470619917, + 0.002393913920968771, + 0.008549446240067482, + 0.008401772938668728, + 0.002807506825774908, + 0.0034972545690834522, + 0.004185445141047239, + 0.007418646011501551, + 0.0033160944003611803, + 0.0019250623881816864, + 0.003471011994406581, + 0.0031167224515229464, + 0.0015459386631846428, + 0.0018596115987747908, + 0.006032311823219061, + 0.004406391642987728, + 0.00595046067610383, + 0.011185442097485065, + 0.015470992773771286, + 0.008635620586574078, + 0.005290111992508173, + 0.010468716733157635, + 0.01025310903787613, + 0.012818344868719578, + 0.0053407480008900166, + 0.009000841528177261, + 0.005060409661382437, + 0.006815621163696051, + 0.008044385351240635, + 0.013290849514305592, + 0.008551943115890026, + 0.007879113778471947, + 0.01223075296729803, + 0.012381600216031075, + 0.04880611598491669, + 0.01808810420334339, + 0.01123824529349804, + 0.014256397262215614, + 0.0656801387667656, + 0.061271943151950836, + 0.004889973439276218, + 0.0034442052710801363, + 0.004975906573235989, + 0.003938703797757626, + 0.0027506391052156687, + 0.0036456005182117224, + 0.004083921201527119, + 0.005088816396892071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0076333265751600266 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188366010785103 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188366010785103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188366010785103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11712267249822617, + 0.03425677865743637, + 0.04387325048446655, + 0.2570782005786896, + 0.03520230948925018, + 0.07174865156412125, + 0.15656442940235138, + 0.09872764348983765, + 0.10489846765995026, + 0.08052759617567062 + ], + "credit_map": [ + 0.01074935868382454, + 0.013539296574890614, + 0.017845947295427322, + 0.01588200405240059, + 0.01384726632386446, + 0.014472359791398048, + 0.01298369187861681, + 0.01780274137854576, + 0.003823024919256568, + 0.0037420550361275673, + 0.004316635429859161, + 0.004224627744406462, + 0.0030463149305433035, + 0.004358597565442324, + 0.005350227002054453, + 0.005395295564085245, + 0.003684268333017826, + 0.005554975010454655, + 0.004761233925819397, + 0.007399819325655699, + 0.004120395518839359, + 0.005564356688410044, + 0.006606665905565023, + 0.006181538570672274, + 0.014664184302091599, + 0.00785867404192686, + 0.014476967975497246, + 0.014032858423888683, + 0.011264408007264137, + 0.017862536013126373, + 0.09447216987609863, + 0.08244638890028, + 0.0030217492021620274, + 0.0025258122477680445, + 0.006725040730088949, + 0.0038342352490872145, + 0.0035174828954041004, + 0.00489303981885314, + 0.0046515995636582375, + 0.006033345591276884, + 0.01005861908197403, + 0.0043331328779459, + 0.0052715265192091465, + 0.006943362299352884, + 0.004040671046823263, + 0.00683613633736968, + 0.025636740028858185, + 0.008628466166555882, + 0.01114030834287405, + 0.020218370482325554, + 0.030080324038863182, + 0.01405235193669796, + 0.014239205978810787, + 0.028287168592214584, + 0.020465265959501266, + 0.018081441521644592, + 0.007638022769242525, + 0.0101939020678401, + 0.009335442446172237, + 0.010355941019952297, + 0.011197305284440517, + 0.015235739760100842, + 0.022959472611546516, + 0.011811813339591026, + 0.00427605863660574, + 0.005113649182021618, + 0.010890335775911808, + 0.006650397554039955, + 0.0054078069515526295, + 0.006795173976570368, + 0.03405427187681198, + 0.03171077370643616, + 0.00816638208925724, + 0.0060781389474868774, + 0.007304274942725897, + 0.008116494864225388, + 0.007263750769197941, + 0.00936832744628191, + 0.01825152523815632, + 0.015978705137968063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.014188366010785103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in high-speed slow motion, a low-angle, medium shot pans steadily to the right, tracking the front wheel of a mountain bike as it speeds down a wet city alley. The bike's thick, knobby tires, caked with a thin layer of mud, deliberately cut a large, dark puddle into the dark, reflective asphalt. As the wheel tumbles through the water, it sprays outward and upward in spectacular fan-shaped splashes. Individual droplets, perfectly formed in the soft, diffuse light of an overcast sky, hang suspended in the air in a crystalline arc. The scene, dominated by a cool, desaturated palette of grays and browns against a backdrop of wet sidewalks and brick walls, makes the violent, dynamic action of the splashing water both raw and artistically captured, conveying a powerful sense of kinetic energy.", + "seed": 3710, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_15/index_2224_seed_3710_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.031027917098253966, + "vlm_logprob": -0.031027917098253966, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.031027917098253966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051761165261268616 + ], + "opposite_logprob_tokens": [ + -3.4267611503601074 + ], + "logprob_reward": -0.051761165261268616, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.051761165261268616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17158736288547516, + 0.06027062237262726, + 0.06354682147502899, + 0.047138337045907974, + 0.03870769962668419, + 0.04528827592730522, + 0.07551088184118271, + 0.1742142140865326, + 0.20773836970329285, + 0.11599735170602798 + ], + "credit_map": [ + 0.016630681231617928, + 0.021090572699904442, + 0.021894000470638275, + 0.028249366208910942, + 0.015530175529420376, + 0.03501666709780693, + 0.019267765805125237, + 0.013908148743212223, + 0.005060916766524315, + 0.007104574702680111, + 0.006679370068013668, + 0.006195137742906809, + 0.006862794514745474, + 0.006882404908537865, + 0.005875114817172289, + 0.015610313042998314, + 0.004949662834405899, + 0.007321939337998629, + 0.00675152987241745, + 0.009675832465291023, + 0.011577364057302475, + 0.009926188737154007, + 0.007321468088775873, + 0.006022829562425613, + 0.003776454832404852, + 0.007434106431901455, + 0.005081790499389172, + 0.007775940466672182, + 0.006276459898799658, + 0.006801768206059933, + 0.0052544064819812775, + 0.00473741116002202, + 0.0034861876629292965, + 0.007453893776983023, + 0.005168095231056213, + 0.004818211775273085, + 0.00341425696387887, + 0.004959327634423971, + 0.005207698326557875, + 0.004200030118227005, + 0.0054060970433056355, + 0.005774984136223793, + 0.005551512818783522, + 0.004513770341873169, + 0.006875535007566214, + 0.0042395577766001225, + 0.00859468337148428, + 0.004332136828452349, + 0.014314710162580013, + 0.009633439593017101, + 0.007946555502712727, + 0.010826142504811287, + 0.006220703013241291, + 0.00702268723398447, + 0.012050776742398739, + 0.0074958717450499535, + 0.03236653655767441, + 0.020145364105701447, + 0.01966651901602745, + 0.028840942308306694, + 0.028452260419726372, + 0.017023861408233643, + 0.012424290180206299, + 0.015294445678591728, + 0.014032738283276558, + 0.01268160529434681, + 0.06780485808849335, + 0.05790826305747032, + 0.012599068693816662, + 0.012074324302375317, + 0.017593733966350555, + 0.013043764978647232, + 0.012077929452061653, + 0.009929055348038673, + 0.024569375440478325, + 0.023531874641776085, + 0.007493881508708, + 0.010609515942633152, + 0.013824577443301678, + 0.013961142860352993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.051761165261268616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900387048721313 + ], + "opposite_logprob_tokens": [ + -2.5790038108825684 + ], + "logprob_reward": -0.07900387048721313, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900387048721313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09652519226074219, + 0.10043071210384369, + 0.04354836419224739, + 0.041342031210660934, + 0.053555168211460114, + 0.14900049567222595, + 0.37467101216316223, + 0.04882111772894859, + 0.04191853851079941, + 0.050187382847070694 + ], + "credit_map": [ + 0.010649063624441624, + 0.011063228361308575, + 0.011661317199468613, + 0.01996283233165741, + 0.012423979118466377, + 0.011285467073321342, + 0.01185267511755228, + 0.007626634556800127, + 0.009585089981555939, + 0.011498440988361835, + 0.013836403377354145, + 0.0136532848700881, + 0.008487415499985218, + 0.010465043596923351, + 0.008967160247266293, + 0.02393786795437336, + 0.0055802082642912865, + 0.005137179512530565, + 0.004145126789808273, + 0.007161539513617754, + 0.005334447603672743, + 0.004282320849597454, + 0.005372762214392424, + 0.006534777581691742, + 0.004433479160070419, + 0.005841503385454416, + 0.0040395911782979965, + 0.0064815497025847435, + 0.005338459741324186, + 0.005562151782214642, + 0.004710758104920387, + 0.004934541415423155, + 0.006061837077140808, + 0.006693124771118164, + 0.010540815070271492, + 0.006942737381905317, + 0.005861710291355848, + 0.005808784160763025, + 0.005595298018306494, + 0.006050862371921539, + 0.020482860505580902, + 0.02574019879102707, + 0.014518375508487225, + 0.015295695513486862, + 0.019437409937381744, + 0.016643762588500977, + 0.021808147430419922, + 0.015074040740728378, + 0.06842876970767975, + 0.052490320056676865, + 0.07269269973039627, + 0.047652535140514374, + 0.037225913256406784, + 0.018746444955468178, + 0.050536639988422394, + 0.026897676289081573, + 0.009488049894571304, + 0.004732130561023951, + 0.0048564113676548, + 0.009212331846356392, + 0.005742230918258429, + 0.005358015652745962, + 0.004718229174613953, + 0.004713717848062515, + 0.004398841876536608, + 0.0033405267167836428, + 0.005416784901171923, + 0.008747127838432789, + 0.003744031535461545, + 0.004609938710927963, + 0.006346146110445261, + 0.005315141752362251, + 0.0071039278991520405, + 0.004175112582743168, + 0.010402399115264416, + 0.009982275776565075, + 0.0034711314365267754, + 0.005153074394911528, + 0.005105942953377962, + 0.004793520551174879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.07900387048721313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595545070245862 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.00595545070245862, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595545070245862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15391027927398682, + 0.19816836714744568, + 0.03705150634050369, + 0.02057202160358429, + 0.028958892449736595, + 0.08753185719251633, + 0.06718531996011734, + 0.077804334461689, + 0.14329621195793152, + 0.1855212301015854 + ], + "credit_map": [ + 0.015582477673888206, + 0.015139580704271793, + 0.026328489184379578, + 0.020293891429901123, + 0.026062168180942535, + 0.015547636896371841, + 0.021385909989476204, + 0.013570145703852177, + 0.021983161568641663, + 0.022874290123581886, + 0.04612763971090317, + 0.01616648957133293, + 0.010600270703434944, + 0.01867794618010521, + 0.013408978469669819, + 0.04832960665225983, + 0.0032213192898780107, + 0.0044002654030919075, + 0.004271287005394697, + 0.0030878453981131315, + 0.006362070795148611, + 0.004926887806504965, + 0.004011240322142839, + 0.006770592648535967, + 0.0021357545629143715, + 0.0027810775209218264, + 0.0022920267656445503, + 0.0021497078705579042, + 0.003332635620608926, + 0.0024094795808196068, + 0.003169019939377904, + 0.002302320208400488, + 0.0024691750295460224, + 0.003801202867180109, + 0.004318984225392342, + 0.0033430939074605703, + 0.003538592951372266, + 0.00472648860886693, + 0.003737538820132613, + 0.0030238176696002483, + 0.009968378581106663, + 0.01383877545595169, + 0.01215776801109314, + 0.011013280600309372, + 0.010744377970695496, + 0.007445578463375568, + 0.013933580368757248, + 0.008430128917098045, + 0.01405592542141676, + 0.007224848959594965, + 0.009196416474878788, + 0.009283985942602158, + 0.005524803418666124, + 0.006538030691444874, + 0.0064308117143809795, + 0.00893049780279398, + 0.011890919879078865, + 0.006580723915249109, + 0.008129829540848732, + 0.012296867556869984, + 0.009703628718852997, + 0.01113506406545639, + 0.007026430685073137, + 0.011040878482162952, + 0.00826937798410654, + 0.008489054627716541, + 0.04478543996810913, + 0.04035014286637306, + 0.011659136973321438, + 0.006338711827993393, + 0.01179945282638073, + 0.011604893952608109, + 0.02283359132707119, + 0.011519377119839191, + 0.03685762733221054, + 0.027628976851701736, + 0.011615773662924767, + 0.019093496724963188, + 0.03138035908341408, + 0.024592041969299316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.00595545070245862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642295375466347 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642295375466347, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642295375466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15820834040641785, + 0.12286801636219025, + 0.10917992889881134, + 0.03823715075850487, + 0.07180093973875046, + 0.06765343993902206, + 0.05924630165100098, + 0.06366147845983505, + 0.2623134255409241, + 0.0468309223651886 + ], + "credit_map": [ + 0.01623314432799816, + 0.020916907116770744, + 0.03248836472630501, + 0.019081708043813705, + 0.028043070808053017, + 0.014132097363471985, + 0.016620131209492683, + 0.010692915879189968, + 0.010138490237295628, + 0.014707772992551327, + 0.02280315011739731, + 0.013877077959477901, + 0.010080233216285706, + 0.011874901130795479, + 0.009995727799832821, + 0.02939065545797348, + 0.008931616321206093, + 0.013178574852645397, + 0.009436107240617275, + 0.01322133932262659, + 0.015033751726150513, + 0.0211827103048563, + 0.012653391808271408, + 0.015542430803179741, + 0.004186231177300215, + 0.004975832067430019, + 0.005296370480209589, + 0.006402045022696257, + 0.005136843770742416, + 0.004821381531655788, + 0.0037299604155123234, + 0.003688486758619547, + 0.005943243857473135, + 0.014096150174736977, + 0.006835785694420338, + 0.008478068746626377, + 0.0077656893990933895, + 0.008364830166101456, + 0.01327361911535263, + 0.0070435525849461555, + 0.00784964207559824, + 0.007693041115999222, + 0.010890171863138676, + 0.0062981536611914635, + 0.010813239961862564, + 0.007214365992695093, + 0.01142293494194746, + 0.005471891723573208, + 0.006893988698720932, + 0.006203056778758764, + 0.007525173015892506, + 0.006964797154068947, + 0.0067038568668067455, + 0.00649060495197773, + 0.011651893146336079, + 0.0068129319697618484, + 0.01142887957394123, + 0.00740602845326066, + 0.006085431203246117, + 0.00873076356947422, + 0.008844065479934216, + 0.007059953175485134, + 0.008662811480462551, + 0.0054435427300632, + 0.016019387170672417, + 0.016769737005233765, + 0.08368460088968277, + 0.07485326379537582, + 0.013282439671456814, + 0.016834737733006477, + 0.02617633529007435, + 0.014692909084260464, + 0.0037779442500323057, + 0.0049593159928917885, + 0.0086017195135355, + 0.009974810294806957, + 0.003378453431650996, + 0.003970019053667784, + 0.006853920873254538, + 0.0053147380240261555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.008642295375466347 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977680366486311 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.00977680366486311, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977680366486311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1759931594133377, + 0.0670260637998581, + 0.043899066746234894, + 0.04459388926625252, + 0.06036016345024109, + 0.07186413556337357, + 0.10196641087532043, + 0.08763701468706131, + 0.24053972959518433, + 0.10612031817436218 + ], + "credit_map": [ + 0.019091395661234856, + 0.019299963489174843, + 0.017813637852668762, + 0.03886761888861656, + 0.019530754536390305, + 0.01965906098484993, + 0.02252446487545967, + 0.019206257537007332, + 0.010525171644985676, + 0.0112006152048707, + 0.014513148926198483, + 0.007861483842134476, + 0.006983509287238121, + 0.0037758739199489355, + 0.005700274370610714, + 0.0064659882336854935, + 0.0038276866544038057, + 0.004639940336346626, + 0.0038211955688893795, + 0.00595741206780076, + 0.0047123306430876255, + 0.009700600989162922, + 0.006111430004239082, + 0.00512846652418375, + 0.003586269449442625, + 0.0054993475787341595, + 0.003998335916548967, + 0.0049438499845564365, + 0.006879614200443029, + 0.006799467373639345, + 0.0076193492859601974, + 0.005267653148621321, + 0.007083325181156397, + 0.009658937342464924, + 0.01076808013021946, + 0.008003650233149529, + 0.0055787935853004456, + 0.007284266408532858, + 0.005443582311272621, + 0.006539522670209408, + 0.0074121421203017235, + 0.008683915250003338, + 0.01471560075879097, + 0.009470412507653236, + 0.006930715870112181, + 0.005462384317070246, + 0.0124385179951787, + 0.006750441621989012, + 0.01318016555160284, + 0.014494351111352444, + 0.012308350764214993, + 0.014962275512516499, + 0.011125859804451466, + 0.009165849536657333, + 0.01628320850431919, + 0.010446347296237946, + 0.01107062678784132, + 0.008992783725261688, + 0.009044287726283073, + 0.020147090777754784, + 0.007957961410284042, + 0.009710351005196571, + 0.011327209882438183, + 0.009386701509356499, + 0.017800217494368553, + 0.015933211892843246, + 0.07278575003147125, + 0.06936734914779663, + 0.014695009216666222, + 0.011742744594812393, + 0.024242235347628593, + 0.013973221182823181, + 0.013583742082118988, + 0.007820211350917816, + 0.019533352926373482, + 0.01952481083571911, + 0.007689510937780142, + 0.012497502379119396, + 0.013345926068723202, + 0.012125266715884209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.00977680366486311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in high-speed slow motion, a low-angle, medium shot pans steadily to the right, tracking the front wheel of a mountain bike as it speeds down a wet city alley. The bike's thick, knobby tires, caked with a thin layer of mud, deliberately cut a large, dark puddle into the dark, reflective asphalt. As the wheel tumbles through the water, it sprays outward and upward in spectacular fan-shaped splashes. Individual droplets, perfectly formed in the soft, diffuse light of an overcast sky, hang suspended in the air in a crystalline arc. The scene, dominated by a cool, desaturated palette of grays and browns against a backdrop of wet sidewalks and brick walls, makes the violent, dynamic action of the splashing water both raw and artistically captured, conveying a powerful sense of kinetic energy.", + "seed": 3711, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_15/index_2224_seed_3711_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05226500141434372, + "vlm_logprob": -0.05226500141434372, + "vlm_yes_no_margin": 4.175000023841858, + "vlm_token_credit": -0.05226500141434372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10393701493740082 + ], + "opposite_logprob_tokens": [ + -2.8539369106292725 + ], + "logprob_reward": -0.10393701493740082, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.10393701493740082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10892252624034882, + 0.05957605317234993, + 0.17960618436336517, + 0.09509846568107605, + 0.18928787112236023, + 0.032038379460573196, + 0.07180187106132507, + 0.05714073032140732, + 0.12551327049732208, + 0.08101465553045273 + ], + "credit_map": [ + 0.01005509402602911, + 0.014610893093049526, + 0.015590322203934193, + 0.016413405537605286, + 0.017792651429772377, + 0.010806308127939701, + 0.015424330718815327, + 0.008229521103203297, + 0.007086365949362516, + 0.005062850657850504, + 0.008696351200342178, + 0.009326695464551449, + 0.005001906305551529, + 0.005344443488866091, + 0.013102834112942219, + 0.005954605061560869, + 0.008578780107200146, + 0.008654702454805374, + 0.020290395244956017, + 0.019087867811322212, + 0.007984613999724388, + 0.015593517571687698, + 0.04721605405211449, + 0.052200254052877426, + 0.005950744729489088, + 0.009398987516760826, + 0.009966103360056877, + 0.012185582891106606, + 0.00555210979655385, + 0.008989010006189346, + 0.01984310708940029, + 0.023212827742099762, + 0.013736337423324585, + 0.018089618533849716, + 0.02586456388235092, + 0.04446452483534813, + 0.023891311138868332, + 0.018776236101984978, + 0.01916266232728958, + 0.025302613154053688, + 0.0029998691752552986, + 0.0030607888475060463, + 0.0047152405604720116, + 0.005134701728820801, + 0.006077858153730631, + 0.00329168071039021, + 0.003114022547379136, + 0.003644219832494855, + 0.008763223886489868, + 0.0069955759681761265, + 0.006985625252127647, + 0.008850432932376862, + 0.008676237426698208, + 0.012670946307480335, + 0.008695690892636776, + 0.010164142586290836, + 0.004157761577516794, + 0.0051183272153139114, + 0.004778197966516018, + 0.009100276976823807, + 0.008953196927905083, + 0.008118395693600178, + 0.008668787777423859, + 0.00824578758329153, + 0.00914428848773241, + 0.014528161846101284, + 0.010385328903794289, + 0.017826875671744347, + 0.01812168024480343, + 0.010403125546872616, + 0.027456680312752724, + 0.017647134140133858, + 0.004783452022820711, + 0.010485348291695118, + 0.0059489961713552475, + 0.011263151653110981, + 0.008162806741893291, + 0.007560988888144493, + 0.013180587440729141, + 0.01962932012975216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.10393701493740082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283455908298492 + ], + "opposite_logprob_tokens": [ + -2.0178346633911133 + ], + "logprob_reward": -0.14283455908298492, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14283455908298492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09867387264966965, + 0.09770675003528595, + 0.14628301560878754, + 0.057059597223997116, + 0.04769374802708626, + 0.031145036220550537, + 0.10322482138872147, + 0.0971437469124794, + 0.2541666626930237, + 0.06690266728401184 + ], + "credit_map": [ + 0.009539796970784664, + 0.012122626416385174, + 0.014388597570359707, + 0.01680889166891575, + 0.015714697539806366, + 0.008336823433637619, + 0.014087537303566933, + 0.007674905937165022, + 0.012984857894480228, + 0.010912248864769936, + 0.012378658168017864, + 0.019054649397730827, + 0.00936162006109953, + 0.007177015300840139, + 0.016677243635058403, + 0.009160465560853481, + 0.007380676921457052, + 0.010877817869186401, + 0.015044741332530975, + 0.012898223474621773, + 0.005919997114688158, + 0.009031212888658047, + 0.04115470126271248, + 0.04397565498948097, + 0.004684344865381718, + 0.005225895904004574, + 0.0055363718420267105, + 0.00903072115033865, + 0.0040623703971505165, + 0.006599388085305691, + 0.010711570270359516, + 0.011208942160010338, + 0.0037369297351688147, + 0.00528316292911768, + 0.006674226373434067, + 0.01300284918397665, + 0.0035903193056583405, + 0.0035832782741636038, + 0.004383579827845097, + 0.007439401000738144, + 0.0036491781938821077, + 0.0024881791323423386, + 0.004425917752087116, + 0.0050885328091681, + 0.003726118477061391, + 0.004337017424404621, + 0.003386510768905282, + 0.004043583292514086, + 0.012452664785087109, + 0.009176718071103096, + 0.009937900118529797, + 0.01375996507704258, + 0.009112276136875153, + 0.012307093478739262, + 0.015748262405395508, + 0.02072995901107788, + 0.007106688339263201, + 0.01037666667252779, + 0.008143873885273933, + 0.013128640130162239, + 0.00798159372061491, + 0.012166263535618782, + 0.018260357901453972, + 0.019979678094387054, + 0.02031080611050129, + 0.04147530719637871, + 0.020146818831562996, + 0.03112545609474182, + 0.03465577960014343, + 0.033753324300050735, + 0.03714577108621597, + 0.03555342182517052, + 0.003911158535629511, + 0.006455676164478064, + 0.005373944062739611, + 0.007176515180617571, + 0.006553153041750193, + 0.009864142164587975, + 0.009343059733510017, + 0.01822502911090851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.14283455908298492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002826506970450282 + ], + "opposite_logprob_tokens": [ + -5.877826690673828 + ], + "logprob_reward": -0.002826506970450282, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002826506970450282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19638144969940186, + 0.14741206169128418, + 0.14788447320461273, + 0.09688494354486465, + 0.1441040337085724, + 0.03341921418905258, + 0.06642569601535797, + 0.05268869176506996, + 0.05482819303870201, + 0.05997123941779137 + ], + "credit_map": [ + 0.02450162172317505, + 0.0185486301779747, + 0.02877994254231453, + 0.030281325802206993, + 0.018510660156607628, + 0.027496563270688057, + 0.035727400332689285, + 0.012535315938293934, + 0.022731488570570946, + 0.014825739897787571, + 0.020726870745420456, + 0.01679493486881256, + 0.01946602761745453, + 0.010919753462076187, + 0.026936620473861694, + 0.015010621398687363, + 0.008127923123538494, + 0.0066541205160319805, + 0.011463484726846218, + 0.009322100318968296, + 0.0067860097624361515, + 0.010750286281108856, + 0.0446278378367424, + 0.050152719020843506, + 0.009400902315974236, + 0.013669424690306187, + 0.01540305931121111, + 0.011494473554193974, + 0.009131346829235554, + 0.016808122396469116, + 0.009571324102580547, + 0.011406287550926208, + 0.017732497304677963, + 0.0170731358230114, + 0.017325036227703094, + 0.023815875872969627, + 0.01565566286444664, + 0.020914867520332336, + 0.012499047443270683, + 0.019087916240096092, + 0.003167624119669199, + 0.003868022933602333, + 0.004248760174959898, + 0.004419553559273481, + 0.0046513257548213005, + 0.004360579419881105, + 0.0046496219001710415, + 0.004053724464029074, + 0.0065895067527890205, + 0.0065465886145830154, + 0.005804433487355709, + 0.006980369333177805, + 0.008070006035268307, + 0.006356676574796438, + 0.010528738610446453, + 0.015549372881650925, + 0.003714190796017647, + 0.0048211244866251945, + 0.00470674317330122, + 0.00899994932115078, + 0.005439917556941509, + 0.009820223785936832, + 0.00742292357608676, + 0.00776361720636487, + 0.005693222861737013, + 0.010451736859977245, + 0.0054792556911706924, + 0.0094445226714015, + 0.004944704473018646, + 0.004564378876239061, + 0.007531508803367615, + 0.006718861870467663, + 0.005696058738976717, + 0.007337629329413176, + 0.008059360086917877, + 0.009528005495667458, + 0.005472204647958279, + 0.003124105976894498, + 0.013151949271559715, + 0.00760192284360528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.002826506970450282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633799687027931 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633799687027931, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633799687027931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1172475516796112, + 0.08643237501382828, + 0.23217430710792542, + 0.13151846826076508, + 0.13059648871421814, + 0.049912936985492706, + 0.06654246896505356, + 0.05762099474668503, + 0.08672043681144714, + 0.04123399779200554 + ], + "credit_map": [ + 0.014647270552814007, + 0.011003902181982994, + 0.019501598551869392, + 0.015706872567534447, + 0.016508223488926888, + 0.013331540860235691, + 0.01914742961525917, + 0.0074007161892950535, + 0.01093489769846201, + 0.0072938078083097935, + 0.012532233260571957, + 0.014013243839144707, + 0.011685633100569248, + 0.0069737122394144535, + 0.014860772527754307, + 0.008138072676956654, + 0.012067579664289951, + 0.008267402648925781, + 0.020584547892212868, + 0.012754132971167564, + 0.009104362688958645, + 0.0171281099319458, + 0.07267171889543533, + 0.07959645986557007, + 0.009215163998305798, + 0.010210833512246609, + 0.008895356208086014, + 0.011758212931454182, + 0.010646476410329342, + 0.012100688181817532, + 0.03176458552479744, + 0.03692713752388954, + 0.013014561496675014, + 0.011197658255696297, + 0.018805595114827156, + 0.02858915738761425, + 0.010368111543357372, + 0.01557113416492939, + 0.014458630234003067, + 0.018591636791825294, + 0.00468814792111516, + 0.006117623299360275, + 0.005500418599694967, + 0.006069260183721781, + 0.008239122107625008, + 0.006566376890987158, + 0.007108293939381838, + 0.0056236921809613705, + 0.009052579291164875, + 0.007610497064888477, + 0.0061013903468847275, + 0.008122558705508709, + 0.010801905766129494, + 0.007994436658918858, + 0.008561959490180016, + 0.008297138847410679, + 0.004524440038949251, + 0.006529051810503006, + 0.004793514963239431, + 0.008553707972168922, + 0.005298275966197252, + 0.013826715759932995, + 0.0063446215353906155, + 0.007750669028609991, + 0.00807258952409029, + 0.012216337025165558, + 0.00825583841651678, + 0.015391605906188488, + 0.0074479603208601475, + 0.0067743523977696896, + 0.01715434342622757, + 0.011407413519918919, + 0.0023175105452537537, + 0.004282624460756779, + 0.0043349373154342175, + 0.008356564678251743, + 0.005333921872079372, + 0.00390796409919858, + 0.007318255491554737, + 0.005382218863815069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.007633799687027931 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093126393854618 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093126393854618, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093126393854618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1895788311958313, + 0.09802765399217606, + 0.3011208176612854, + 0.052278563380241394, + 0.027305843308568, + 0.02342909947037697, + 0.08635345101356506, + 0.06939023733139038, + 0.050537899136543274, + 0.10197767615318298 + ], + "credit_map": [ + 0.018515998497605324, + 0.023989658802747726, + 0.030352812260389328, + 0.04809661954641342, + 0.018739737570285797, + 0.00880128238350153, + 0.026240233331918716, + 0.014842495322227478, + 0.01195148192346096, + 0.006522070616483688, + 0.01576370932161808, + 0.011957954615354538, + 0.006130869034677744, + 0.006606515031307936, + 0.0265145655721426, + 0.012580479495227337, + 0.016264287754893303, + 0.008508363738656044, + 0.02157629281282425, + 0.015266379341483116, + 0.013210296630859375, + 0.023692218586802483, + 0.09525533765554428, + 0.10734760761260986, + 0.005271902773529291, + 0.005794762633740902, + 0.010296092368662357, + 0.006194300949573517, + 0.004924424458295107, + 0.00790087878704071, + 0.005115279462188482, + 0.006780917290598154, + 0.0027749803848564625, + 0.0025607498828321695, + 0.003350689308717847, + 0.003920911345630884, + 0.002311876742169261, + 0.0055997371673583984, + 0.002945943968370557, + 0.003840953577309847, + 0.0020235455594956875, + 0.0025150415021926165, + 0.0031724933069199324, + 0.0038641809951514006, + 0.002851989818736911, + 0.003060566261410713, + 0.0030848351307213306, + 0.0028564464300870895, + 0.010846756398677826, + 0.0071976520121097565, + 0.007287812419235706, + 0.010693584568798542, + 0.008996134623885155, + 0.012258202768862247, + 0.013519705273211002, + 0.01555359736084938, + 0.004379198420792818, + 0.006661282852292061, + 0.005673882085829973, + 0.009805357083678246, + 0.008461629040539265, + 0.013329016976058483, + 0.0101391626521945, + 0.010940702632069588, + 0.005312205292284489, + 0.01305299811065197, + 0.0052472795359790325, + 0.006119458936154842, + 0.0059951175935566425, + 0.0038843273650854826, + 0.006160071585327387, + 0.004766438622027636, + 0.009398028254508972, + 0.013493495993316174, + 0.007525226101279259, + 0.010647687129676342, + 0.011187457479536533, + 0.01595480740070343, + 0.013560592196881771, + 0.020210379734635353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.004093126393854618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in high-speed slow motion, a low-angle, medium shot pans steadily to the right, tracking the front wheel of a mountain bike as it speeds down a wet city alley. The bike's thick, knobby tires, caked with a thin layer of mud, deliberately cut a large, dark puddle into the dark, reflective asphalt. As the wheel tumbles through the water, it sprays outward and upward in spectacular fan-shaped splashes. Individual droplets, perfectly formed in the soft, diffuse light of an overcast sky, hang suspended in the air in a crystalline arc. The scene, dominated by a cool, desaturated palette of grays and browns against a backdrop of wet sidewalks and brick walls, makes the violent, dynamic action of the splashing water both raw and artistically captured, conveying a powerful sense of kinetic energy.", + "seed": 3712, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_15/index_2224_seed_3712_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.051416529808193445, + "vlm_logprob": -0.051416529808193445, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.051416529808193445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1744726300239563 + ], + "opposite_logprob_tokens": [ + -2.0494725704193115 + ], + "logprob_reward": -0.1744726300239563, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1744726300239563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05465935543179512, + 0.010767536237835884, + 0.012718094512820244, + 0.05332951620221138, + 0.020125681534409523, + 0.040234845131635666, + 0.031651128083467484, + 0.7415509819984436, + 0.019050464034080505, + 0.01591244339942932 + ], + "credit_map": [ + 0.00669119181111455, + 0.005206271540373564, + 0.007447437848895788, + 0.008625233545899391, + 0.004823042079806328, + 0.0073134866543114185, + 0.007671522442251444, + 0.006881173700094223, + 0.00164528819732368, + 0.0010032212594524026, + 0.0017096578376367688, + 0.0010573521722108126, + 0.0009217237820848823, + 0.001522521604783833, + 0.0014413975877687335, + 0.0014663743786513805, + 0.0016873196000233293, + 0.0012231565779075027, + 0.0019023704808205366, + 0.002021321328356862, + 0.001441525761038065, + 0.0016797741409391165, + 0.001531445886939764, + 0.0012311803875491023, + 0.005843923892825842, + 0.004577011801302433, + 0.009736542589962482, + 0.006708482280373573, + 0.007601180113852024, + 0.006391842383891344, + 0.008268972858786583, + 0.00420156167820096, + 0.002190342405810952, + 0.0012587850214913487, + 0.004126461688429117, + 0.0025263403076678514, + 0.0018611770356073976, + 0.00119896256364882, + 0.0024071664083749056, + 0.004556446801871061, + 0.00511764595285058, + 0.003253671806305647, + 0.005384403746575117, + 0.010451633483171463, + 0.002773240441456437, + 0.002524528419598937, + 0.006076266523450613, + 0.004653453826904297, + 0.0037447637878358364, + 0.0028863458428531885, + 0.005390911363065243, + 0.003174998564645648, + 0.0021999948658049107, + 0.0040827044285833836, + 0.0038807124365121126, + 0.006290694233030081, + 0.0935162678360939, + 0.03585751727223396, + 0.19038760662078857, + 0.042483121156692505, + 0.027171006426215172, + 0.037967827171087265, + 0.07544493675231934, + 0.23872268199920654, + 0.002024347661063075, + 0.002363045932725072, + 0.0027516414411365986, + 0.0021995268762111664, + 0.0028079282492399216, + 0.0016124813118949533, + 0.003035382367670536, + 0.002256110543385148, + 0.0021582774352282286, + 0.0010571405291557312, + 0.002341696061193943, + 0.0017378428019583225, + 0.001409434713423252, + 0.0026644719764590263, + 0.002301356988027692, + 0.0022422242909669876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.1744726300239563 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208615377545357 + ], + "opposite_logprob_tokens": [ + -2.8120861053466797 + ], + "logprob_reward": -0.06208615377545357, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208615377545357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041653744876384735, + 0.012253225781023502, + 0.015505628660321236, + 0.11638215184211731, + 0.016964469105005264, + 0.028322391211986542, + 0.049092572182416916, + 0.6704255938529968, + 0.03002319298684597, + 0.01937701925635338 + ], + "credit_map": [ + 0.008836592547595501, + 0.003599221585318446, + 0.007323927711695433, + 0.004953489173203707, + 0.00206345459446311, + 0.002703016856685281, + 0.004391595255583525, + 0.0077824462205171585, + 0.001973278820514679, + 0.0010726258624345064, + 0.0023340717889368534, + 0.0010609753662720323, + 0.0010018893517553806, + 0.001116088591516018, + 0.001198416342958808, + 0.0024958790745586157, + 0.0029970016330480576, + 0.001262459671124816, + 0.0017814364982768893, + 0.0020814621821045876, + 0.0024356867652386427, + 0.0014271592954173684, + 0.0022380619775503874, + 0.0012823602883145213, + 0.01244671456515789, + 0.011578078381717205, + 0.01624138467013836, + 0.012502938508987427, + 0.018408263102173805, + 0.021243833005428314, + 0.012232216075062752, + 0.011728727258741856, + 0.001595214707776904, + 0.0012163894716650248, + 0.0026859655044972897, + 0.002923672553151846, + 0.002181335585191846, + 0.0014440675731748343, + 0.0027666829992085695, + 0.0021511418744921684, + 0.0033648137468844652, + 0.003277549520134926, + 0.0037291902117431164, + 0.008438551798462868, + 0.001644282485358417, + 0.002185127232223749, + 0.004124028608202934, + 0.0015588473761454225, + 0.006250025238841772, + 0.002899907762184739, + 0.013285351917147636, + 0.005361143033951521, + 0.0024265176616609097, + 0.0032458214554935694, + 0.0035973310004919767, + 0.012026476673781872, + 0.08311878144741058, + 0.03275381401181221, + 0.1673436313867569, + 0.03882389888167381, + 0.022988148033618927, + 0.03544817119836807, + 0.07227484881877899, + 0.21767430007457733, + 0.004760904237627983, + 0.0027500190772116184, + 0.00408411817625165, + 0.0033845228608697653, + 0.003954897169023752, + 0.0024005542509257793, + 0.0050425054505467415, + 0.003645670600235462, + 0.0026834227610379457, + 0.002171851461753249, + 0.003443663939833641, + 0.002349490998312831, + 0.0016784146428108215, + 0.002152650849893689, + 0.0026888290885835886, + 0.0022086973767727613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.06208615377545357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639123592525721 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007639123592525721, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639123592525721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09857477247714996, + 0.03448247164487839, + 0.026005301624536514, + 0.10175572335720062, + 0.041271958500146866, + 0.08935713022947311, + 0.05408748239278793, + 0.511739194393158, + 0.027282878756523132, + 0.015443103387951851 + ], + "credit_map": [ + 0.013332360424101353, + 0.0075650992803275585, + 0.02000909112393856, + 0.008865420706570148, + 0.00390281667932868, + 0.010210475884377956, + 0.01129500288516283, + 0.02339450642466545, + 0.005092495121061802, + 0.004590606316924095, + 0.006353266071528196, + 0.004069111775606871, + 0.0026234223041683435, + 0.0029108289163559675, + 0.00470691965892911, + 0.004135824274271727, + 0.0034522749483585358, + 0.0019126575207337737, + 0.004421704448759556, + 0.0035140186082571745, + 0.0031307018361985683, + 0.0031475783325731754, + 0.0030055027455091476, + 0.003420864464715123, + 0.011876307427883148, + 0.009354967623949051, + 0.014755838550627232, + 0.009899137541651726, + 0.013320667669177055, + 0.02126455120742321, + 0.01088941004127264, + 0.010394846089184284, + 0.003392135025933385, + 0.002924475120380521, + 0.007891768589615822, + 0.006826081778854132, + 0.004797132220119238, + 0.0030584956984966993, + 0.0058190119452774525, + 0.006562854163348675, + 0.010741934180259705, + 0.008155597373843193, + 0.011984223499894142, + 0.024464990943670273, + 0.006310116034001112, + 0.005375566426664591, + 0.014803155325353146, + 0.007521543651819229, + 0.00658545084297657, + 0.0026992945931851864, + 0.014559851959347725, + 0.005713227204978466, + 0.0030083688907325268, + 0.004312328528612852, + 0.0040069278329610825, + 0.013202032074332237, + 0.05955757200717926, + 0.027188269421458244, + 0.11454252898693085, + 0.030285604298114777, + 0.019765863195061684, + 0.027673110365867615, + 0.08311521261930466, + 0.14961101114749908, + 0.004385027568787336, + 0.001807154156267643, + 0.004480008501559496, + 0.0036637792363762856, + 0.0019516340689733624, + 0.002384552266448736, + 0.004960069432854652, + 0.0036506536416709423, + 0.0021668539848178625, + 0.0010579454246908426, + 0.002219236921519041, + 0.0017395250033587217, + 0.00120463443454355, + 0.002319695195183158, + 0.0028851008974015713, + 0.0018501123413443565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.007639123592525721 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763675756752491 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.00763675756752491, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763675756752491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11591852456331253, + 0.01439281739294529, + 0.03204932063817978, + 0.05469910427927971, + 0.06932711601257324, + 0.016954563558101654, + 0.031518395990133286, + 0.6086820363998413, + 0.017961258068680763, + 0.03849683701992035 + ], + "credit_map": [ + 0.01702432706952095, + 0.009285331703722477, + 0.034932561218738556, + 0.011944198980927467, + 0.004926235415041447, + 0.007801002822816372, + 0.00988065917044878, + 0.020124200731515884, + 0.0020318219903856516, + 0.0014417454367503524, + 0.0021137914154678583, + 0.001744303503073752, + 0.0008937722886912525, + 0.0021441797725856304, + 0.0020224638283252716, + 0.00200073909945786, + 0.0037934735883027315, + 0.0029324982315301895, + 0.004963123705238104, + 0.006386012304574251, + 0.0029454135801643133, + 0.0041202702559530735, + 0.0037716790102422237, + 0.0031368492636829615, + 0.0056192087940871716, + 0.004123703110963106, + 0.009291731752455235, + 0.006367031950503588, + 0.006925601977854967, + 0.006961415521800518, + 0.011069941334426403, + 0.004340470768511295, + 0.007251403760164976, + 0.004528807010501623, + 0.011507144197821617, + 0.007851649075746536, + 0.00913319829851389, + 0.007187604438513517, + 0.009599335491657257, + 0.012267977930605412, + 0.0020279402378946543, + 0.0016200138488784432, + 0.0020840635988861322, + 0.0048570274375379086, + 0.0011189572978764772, + 0.0014712928095832467, + 0.002347905421629548, + 0.0014273623237386346, + 0.003958769608289003, + 0.004453455563634634, + 0.004892098251730204, + 0.005269352812319994, + 0.0019487967947497964, + 0.0038205711171031, + 0.003528427332639694, + 0.003646924626082182, + 0.07687686383724213, + 0.02974119409918785, + 0.1559661477804184, + 0.03479594737291336, + 0.022287186235189438, + 0.031790971755981445, + 0.05972861871123314, + 0.19749508798122406, + 0.002402466954663396, + 0.0021563072223216295, + 0.0025818850845098495, + 0.001949432771652937, + 0.002308283466845751, + 0.0013570711016654968, + 0.0031571355648338795, + 0.0020486756693571806, + 0.00449134549126029, + 0.0029898344073444605, + 0.005012996029108763, + 0.003243820508942008, + 0.004252539481967688, + 0.008824563585221767, + 0.004687190055847168, + 0.00499454652890563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.00763675756752491 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera panning steadily to the right while tracking a mountain bike's front wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike's front wheel cut through a large, dark puddle on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the splashes form a fan-shaped arc with individual droplets hanging suspended in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, desaturated palette of grays and browns with a backdrop of wet sidewalks and brick walls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247984081506729 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005247984081506729, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247984081506729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23902533948421478, + 0.06791450083255768, + 0.03475925698876381, + 0.1944349706172943, + 0.029414264485239983, + 0.053188107907772064, + 0.06606735289096832, + 0.14642329514026642, + 0.04607709124684334, + 0.12269581854343414 + ], + "credit_map": [ + 0.03451640158891678, + 0.03620629385113716, + 0.056214235723018646, + 0.028462400659918785, + 0.01150173507630825, + 0.019147507846355438, + 0.023987578228116035, + 0.02898917905986309, + 0.007170048542320728, + 0.009938741102814674, + 0.007560651283711195, + 0.007469512056559324, + 0.008247742429375648, + 0.006595177110284567, + 0.009261170402169228, + 0.01167146023362875, + 0.006246481090784073, + 0.004170458763837814, + 0.005570142529904842, + 0.004940790589898825, + 0.003987041302025318, + 0.003188049653545022, + 0.0037671171594411135, + 0.0028891724068671465, + 0.019927198067307472, + 0.020450064912438393, + 0.026703428477048874, + 0.02439079061150551, + 0.03131215646862984, + 0.030922498553991318, + 0.020857660099864006, + 0.01987117901444435, + 0.003174207638949156, + 0.0025470980908721685, + 0.004741129465401173, + 0.005066057201474905, + 0.0030047204345464706, + 0.0034745780285447836, + 0.004384823143482208, + 0.0030216495506465435, + 0.007376586087048054, + 0.004915261175483465, + 0.006087955553084612, + 0.00998789630830288, + 0.004299083724617958, + 0.004481105133891106, + 0.00787857361137867, + 0.008161644451320171, + 0.009328649379312992, + 0.007670145481824875, + 0.011020569130778313, + 0.007065696641802788, + 0.007331488188356161, + 0.006422009319067001, + 0.00966666080057621, + 0.007562133017927408, + 0.01846027374267578, + 0.007129073608666658, + 0.033522482961416245, + 0.009823692962527275, + 0.008030814118683338, + 0.007371419109404087, + 0.018571557477116585, + 0.04351397976279259, + 0.012411735951900482, + 0.003583022626116872, + 0.008761431090533733, + 0.004231960512697697, + 0.0019375365227460861, + 0.003998687025159597, + 0.005547614302486181, + 0.0056051043793559074, + 0.012270270846784115, + 0.008543726988136768, + 0.017872076481580734, + 0.00807082187384367, + 0.015261618420481682, + 0.03296786919236183, + 0.011913610622286797, + 0.01579582877457142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.005247984081506729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_tensor([859, 859])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_tensor([859, 859])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..10f51f12bfa332d758a48b42b701281dac4dbd80 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_tensor([859, 859])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunningly static close-up, fixed from a level angle against a completely black, blurred background, reveals a smooth, porcelain-like white mask. A single, cold spotlight shines diagonally from above, carving a stark contrast of light and dark into its surface, highlighting the hollow, black eye sockets and a fixed, slightly eerie smile. Suddenly, tiny, spider-web-like cracks, cautiously starting at the corners of the smiling mouth, spread rapidly and silently, instantly covering the entire smooth surface. Then, without warning, the mask violently and outwardly explodes into countless white fragments of varying sizes. As if time has slowed down, these sharp ceramic shards fly through the air, each reflecting a sharp light before dissipating into the darkness, creating a powerful visual impact and tense atmosphere, a transition from oppressive calm to instantaneous collapse.", + "seed": 26490, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_859_seed_26490_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.003124979929998517, + "vlm_logprob": -0.003124979929998517, + "vlm_yes_no_margin": 6.625, + "vlm_token_credit": -0.003124979929998517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003759936196729541 + ], + "opposite_logprob_tokens": [ + -7.503759860992432 + ], + "logprob_reward": -0.003759936196729541, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.003759936196729541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07073182612657547, + 0.025749072432518005, + 0.04155033826828003, + 0.032234348356723785, + 0.06042107194662094, + 0.05546651780605316, + 0.1288052499294281, + 0.11563089489936829, + 0.14351704716682434, + 0.32589367032051086 + ], + "credit_map": [ + 0.008900891989469528, + 0.007619032170623541, + 0.00800870917737484, + 0.005529365502297878, + 0.011714771389961243, + 0.006400789599865675, + 0.010073529556393623, + 0.012484734877943993, + 0.005477577913552523, + 0.0016803309554234147, + 0.0033838534727692604, + 0.0016314577078446746, + 0.002900846069678664, + 0.0025017813313752413, + 0.0029557310044765472, + 0.005217493511736393, + 0.00839293748140335, + 0.00408823462203145, + 0.004942414350807667, + 0.0029116179794073105, + 0.00649648904800415, + 0.003530194517225027, + 0.007133447099477053, + 0.004055002238601446, + 0.004200894385576248, + 0.0031361361034214497, + 0.00325609277933836, + 0.0037362854927778244, + 0.00574637483805418, + 0.003021138720214367, + 0.00587814673781395, + 0.003259277669712901, + 0.0053771138191223145, + 0.004209122154861689, + 0.005284537095576525, + 0.007045015227049589, + 0.010088280774652958, + 0.013215872459113598, + 0.007927918806672096, + 0.00727321021258831, + 0.0032264916226267815, + 0.010145326144993305, + 0.004028109833598137, + 0.00805756263434887, + 0.006140675861388445, + 0.007105722092092037, + 0.011142384260892868, + 0.00562024163082242, + 0.009018857963383198, + 0.01903623715043068, + 0.011367443017661572, + 0.026594221591949463, + 0.01713521219789982, + 0.01700926199555397, + 0.016119571402668953, + 0.012524439021945, + 0.008816073648631573, + 0.00996642466634512, + 0.011972881853580475, + 0.018312783911824226, + 0.018575625494122505, + 0.012441771104931831, + 0.01747078448534012, + 0.01807454787194729, + 0.015408373437821865, + 0.025616686791181564, + 0.012430047616362572, + 0.01576884090900421, + 0.020374583080410957, + 0.019870219752192497, + 0.013854188844561577, + 0.020194102078676224, + 0.027889814227819443, + 0.035570178180933, + 0.026102859526872635, + 0.05804337561130524, + 0.07238331437110901, + 0.03874770551919937, + 0.035541050136089325, + 0.03161538392305374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.003759936196729541 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865529477596283 + ], + "opposite_logprob_tokens": [ + -4.758655071258545 + ], + "logprob_reward": -0.00865529477596283, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865529477596283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03919830545783043, + 0.06293810158967972, + 0.11194253712892532, + 0.026353687047958374, + 0.05751023441553116, + 0.01863420382142067, + 0.08612210303544998, + 0.05281738191843033, + 0.07115380465984344, + 0.47332969307899475 + ], + "credit_map": [ + 0.006537550128996372, + 0.003553486429154873, + 0.005205502733588219, + 0.002932308241724968, + 0.006125652696937323, + 0.003831698326393962, + 0.005825462751090527, + 0.0051866392605006695, + 0.012324272654950619, + 0.006501750089228153, + 0.007906651124358177, + 0.0045570735819637775, + 0.011622036807239056, + 0.004677338991314173, + 0.008933640085160732, + 0.006415337324142456, + 0.020687323063611984, + 0.01251222938299179, + 0.013861705549061298, + 0.005803761072456837, + 0.019726261496543884, + 0.009719324298202991, + 0.01605968363583088, + 0.013572240248322487, + 0.00259046396240592, + 0.002641771687194705, + 0.0020113815553486347, + 0.0028396816924214363, + 0.004857743624597788, + 0.003055113134905696, + 0.004926689434796572, + 0.0034308393951505423, + 0.008904931135475636, + 0.004486553370952606, + 0.007001043297350407, + 0.004464066121727228, + 0.007521210238337517, + 0.006644731853157282, + 0.008885113522410393, + 0.009602581150829792, + 0.0017262400360777974, + 0.003572860499843955, + 0.0016750283539295197, + 0.002454178174957633, + 0.0019301206339150667, + 0.0018099808366969228, + 0.0032119788229465485, + 0.0022538146004080772, + 0.006217478774487972, + 0.016311511397361755, + 0.010690093040466309, + 0.015499444678425789, + 0.011017915792763233, + 0.006403939798474312, + 0.013716443441808224, + 0.0062652695924043655, + 0.005308335181325674, + 0.010462300851941109, + 0.007084934506565332, + 0.009343912824988365, + 0.004923003725707531, + 0.0047373645938932896, + 0.007025526370853186, + 0.003931997809559107, + 0.006564302369952202, + 0.01196556817740202, + 0.0062641482800245285, + 0.008968518115580082, + 0.013143657706677914, + 0.009430023841559887, + 0.006824912969022989, + 0.007992662489414215, + 0.04416694492101669, + 0.050635360181331635, + 0.039017632603645325, + 0.07213345170021057, + 0.09877172857522964, + 0.0591890849173069, + 0.05189286917448044, + 0.057522598654031754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 6 + ], + "selected_reward": -0.00865529477596283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001720973290503025 + ], + "opposite_logprob_tokens": [ + -6.376720905303955 + ], + "logprob_reward": -0.001720973290503025, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001720973290503025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05418842285871506, + 0.06427066028118134, + 0.051082246005535126, + 0.03286158666014671, + 0.1608019918203354, + 0.0283354502171278, + 0.13754360377788544, + 0.08870977908372879, + 0.10849256068468094, + 0.2737136781215668 + ], + "credit_map": [ + 0.008550203405320644, + 0.004188168793916702, + 0.006378726102411747, + 0.0048935324884951115, + 0.00919950008392334, + 0.0050466773100197315, + 0.00749594671651721, + 0.008435670286417007, + 0.009741161949932575, + 0.005372266285121441, + 0.00899576023221016, + 0.004722947254776955, + 0.009481510147452354, + 0.007116822991520166, + 0.011496818624436855, + 0.007343372795730829, + 0.007435033563524485, + 0.004966442473232746, + 0.005468960385769606, + 0.003583374433219433, + 0.009857717901468277, + 0.005047079641371965, + 0.008189070969820023, + 0.0065345666371285915, + 0.004372797440737486, + 0.0032871905714273453, + 0.002528346376493573, + 0.00261297426186502, + 0.004920610226690769, + 0.0031375547405332327, + 0.005144608207046986, + 0.006857504136860371, + 0.015403437428176403, + 0.014109413139522076, + 0.011351870372891426, + 0.015654779970645905, + 0.03516795113682747, + 0.01693776063621044, + 0.02584673836827278, + 0.026330040767788887, + 0.003887590952217579, + 0.004543300252407789, + 0.0029539521783590317, + 0.002970343455672264, + 0.003591270186007023, + 0.003236692864447832, + 0.004253818653523922, + 0.0028984793461859226, + 0.009919059462845325, + 0.019507018849253654, + 0.014299917034804821, + 0.022496318444609642, + 0.020847614854574203, + 0.015371348708868027, + 0.021760517731308937, + 0.013341812416911125, + 0.007512770127505064, + 0.015667077153921127, + 0.0109782163053751, + 0.014572354964911938, + 0.011051663197577, + 0.008498596958816051, + 0.010253879241645336, + 0.010175221599638462, + 0.009799819439649582, + 0.0150359021499753, + 0.010031580924987793, + 0.014628511853516102, + 0.01962803676724434, + 0.012741116806864738, + 0.011493653059005737, + 0.01513394434005022, + 0.026631496846675873, + 0.029386043548583984, + 0.02384861372411251, + 0.04199296981096268, + 0.05452199652791023, + 0.03405538573861122, + 0.027699489146471024, + 0.0355776883661747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.001720973290503025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009234455064870417 + ], + "opposite_logprob_tokens": [ + -7.0009236335754395 + ], + "logprob_reward": -0.0009234455064870417, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009234455064870417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047719232738018036, + 0.023927219212055206, + 0.011356078088283539, + 0.01582370512187481, + 0.12774181365966797, + 0.03426080569624901, + 0.25484272837638855, + 0.08034229278564453, + 0.12992003560066223, + 0.2740660607814789 + ], + "credit_map": [ + 0.006720549892634153, + 0.005262245889753103, + 0.005243903957307339, + 0.0025295044761151075, + 0.00963867548853159, + 0.004892746452242136, + 0.005638224072754383, + 0.007793381344527006, + 0.0036750773433595896, + 0.0020677256397902966, + 0.0027576247230172157, + 0.0019590426236391068, + 0.004651044029742479, + 0.0024003596045076847, + 0.0034542360808700323, + 0.0029621077701449394, + 0.0020422632806003094, + 0.0010424769716337323, + 0.0012110642855986953, + 0.000913417199626565, + 0.001678017433732748, + 0.001161650987342, + 0.0017366259125992656, + 0.0015705629484727979, + 0.0022943008225411177, + 0.0013847296359017491, + 0.0015014776727184653, + 0.0012923048343509436, + 0.0028681231196969748, + 0.0018882850417867303, + 0.0024088886566460133, + 0.0021855952218174934, + 0.010637864470481873, + 0.01222244929522276, + 0.0072437166236341, + 0.010730748996138573, + 0.031199777498841286, + 0.01930244080722332, + 0.017944447696208954, + 0.018460357561707497, + 0.004635029938071966, + 0.003911599051207304, + 0.0034188462886959314, + 0.003118614898994565, + 0.005366487894207239, + 0.004835553001612425, + 0.005010297521948814, + 0.003964378032833338, + 0.018977569416165352, + 0.05043729394674301, + 0.03211279958486557, + 0.04803640767931938, + 0.023304613307118416, + 0.018642202019691467, + 0.045612070709466934, + 0.01771978661417961, + 0.008781369775533676, + 0.007581419311463833, + 0.011682825163006783, + 0.01309975516051054, + 0.00777267012745142, + 0.00931527465581894, + 0.011241150088608265, + 0.010867823846638203, + 0.011209368705749512, + 0.015405891463160515, + 0.013488017953932285, + 0.020782044157385826, + 0.02012661285698414, + 0.014778298325836658, + 0.01684194803237915, + 0.017287859693169594, + 0.030858080834150314, + 0.035120707005262375, + 0.019950605928897858, + 0.040993962436914444, + 0.060606248676776886, + 0.03378744795918465, + 0.025573832914233208, + 0.027175169438123703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.0009234455064870417 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000565249880310148 + ], + "opposite_logprob_tokens": [ + -7.500565052032471 + ], + "logprob_reward": -0.000565249880310148, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.000565249880310148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0669223815202713, + 0.017405951395630836, + 0.014419162645936012, + 0.01518942229449749, + 0.12848521769046783, + 0.034649964421987534, + 0.16022251546382904, + 0.11759966611862183, + 0.2219679355621338, + 0.2231377214193344 + ], + "credit_map": [ + 0.0098670469596982, + 0.007227315567433834, + 0.008030310273170471, + 0.004038298036903143, + 0.013288738206028938, + 0.007189525756984949, + 0.008380132727324963, + 0.008901010267436504, + 0.0023285753559321165, + 0.0017054579220712185, + 0.0019018296152353287, + 0.001183417160063982, + 0.003838670440018177, + 0.002080436795949936, + 0.002385598374530673, + 0.0019819659646600485, + 0.0018210868583992124, + 0.0016551013104617596, + 0.0015169756952673197, + 0.001045581535436213, + 0.0028072227723896503, + 0.0016199074452742934, + 0.0020774046424776316, + 0.0018758829683065414, + 0.0020147948525846004, + 0.0011668801307678223, + 0.0011444923002272844, + 0.0015469808131456375, + 0.002578070620074868, + 0.0021954032126814127, + 0.002381059341132641, + 0.002161741256713867, + 0.011229981668293476, + 0.013197105377912521, + 0.006945843808352947, + 0.010982370935380459, + 0.034483227878808975, + 0.014549151062965393, + 0.018797511234879494, + 0.018300026655197144, + 0.003284317208454013, + 0.00395182054489851, + 0.003216835670173168, + 0.003464947920292616, + 0.004929160233587027, + 0.005711067467927933, + 0.00471995584666729, + 0.005371862556785345, + 0.010875278152525425, + 0.022169429808855057, + 0.022020168602466583, + 0.024509795010089874, + 0.020957067608833313, + 0.015620000660419464, + 0.02581881545484066, + 0.018251977860927582, + 0.015453438274562359, + 0.018489519134163857, + 0.013353881426155567, + 0.014561926014721394, + 0.010947629809379578, + 0.012509108521044254, + 0.01564842090010643, + 0.016635751351714134, + 0.027148166671395302, + 0.038575634360313416, + 0.01371016539633274, + 0.01817236840724945, + 0.04220634326338768, + 0.03030833788216114, + 0.01933089643716812, + 0.03251604363322258, + 0.03153586760163307, + 0.032555874437093735, + 0.011802945286035538, + 0.0458175353705883, + 0.03498848155140877, + 0.0318152979016304, + 0.013684665784239769, + 0.02093707025051117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.000565249880310148 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunningly static close-up, fixed from a level angle against a completely black, blurred background, reveals a smooth, porcelain-like white mask. A single, cold spotlight shines diagonally from above, carving a stark contrast of light and dark into its surface, highlighting the hollow, black eye sockets and a fixed, slightly eerie smile. Suddenly, tiny, spider-web-like cracks, cautiously starting at the corners of the smiling mouth, spread rapidly and silently, instantly covering the entire smooth surface. Then, without warning, the mask violently and outwardly explodes into countless white fragments of varying sizes. As if time has slowed down, these sharp ceramic shards fly through the air, each reflecting a sharp light before dissipating into the darkness, creating a powerful visual impact and tense atmosphere, a transition from oppressive calm to instantaneous collapse.", + "seed": 26491, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_859_seed_26491_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.005487132933922112, + "vlm_logprob": -0.005487132933922112, + "vlm_yes_no_margin": 5.725, + "vlm_token_credit": -0.005487132933922112, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004546305630356073 + ], + "opposite_logprob_tokens": [ + -7.004546165466309 + ], + "logprob_reward": -0.004546305630356073, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.004546305630356073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15705648064613342, + 0.043353158980607986, + 0.04155026748776436, + 0.06908188760280609, + 0.03386718034744263, + 0.04044105485081673, + 0.058604512363672256, + 0.10015413165092468, + 0.3084774315357208, + 0.14741398394107819 + ], + "credit_map": [ + 0.029719389975070953, + 0.019169345498085022, + 0.02221914939582348, + 0.024985235184431076, + 0.014087066985666752, + 0.012220613658428192, + 0.01666097342967987, + 0.017994696274399757, + 0.00814905483275652, + 0.005458450876176357, + 0.00492451898753643, + 0.00631168857216835, + 0.004681115038692951, + 0.004271716345101595, + 0.004199500661343336, + 0.005357113666832447, + 0.00925042200833559, + 0.003757167374715209, + 0.006676455494016409, + 0.005225138273090124, + 0.003582865698263049, + 0.004004841670393944, + 0.0037337075918912888, + 0.005319666583091021, + 0.0100637748837471, + 0.005620225798338652, + 0.009055710397660732, + 0.013875160366296768, + 0.007131965830922127, + 0.0059909760020673275, + 0.006635186728090048, + 0.010708878748118877, + 0.004195026587694883, + 0.0032741583418101072, + 0.005205031484365463, + 0.006941542960703373, + 0.0028063433710485697, + 0.0025216168724000454, + 0.0038232319056987762, + 0.005100225564092398, + 0.007185132708400488, + 0.002745092147961259, + 0.007078645750880241, + 0.00395354675129056, + 0.002342422027140856, + 0.004433991853147745, + 0.0035121191758662462, + 0.009190102107822895, + 0.0142646050080657, + 0.005069872830063105, + 0.01028395351022482, + 0.005901049822568893, + 0.004300004802644253, + 0.004821170587092638, + 0.005449902731925249, + 0.008513948880136013, + 0.013637077063322067, + 0.011566909961402416, + 0.017621418461203575, + 0.0145700853317976, + 0.011093557812273502, + 0.011195888742804527, + 0.010190069675445557, + 0.010279103182256222, + 0.046571191400289536, + 0.05113663896918297, + 0.04205428063869476, + 0.03704351186752319, + 0.02157820388674736, + 0.031199509277939796, + 0.03603772819042206, + 0.042856331914663315, + 0.024128606542944908, + 0.019556721672415733, + 0.017727995291352272, + 0.014078505337238312, + 0.014772511087357998, + 0.013071548193693161, + 0.020171383395791054, + 0.023906700313091278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.004546305630356073 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009806433692574501 + ], + "opposite_logprob_tokens": [ + -4.6348066329956055 + ], + "logprob_reward": -0.009806433692574501, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009806433692574501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.097431480884552, + 0.09786038845777512, + 0.08081414550542831, + 0.05595872551202774, + 0.056358106434345245, + 0.07360445708036423, + 0.2551085352897644, + 0.08524011075496674, + 0.12406229972839355, + 0.07356172800064087 + ], + "credit_map": [ + 0.016802890226244926, + 0.01205269992351532, + 0.013766034506261349, + 0.016289906576275826, + 0.007398038636893034, + 0.009854567237198353, + 0.0098013486713171, + 0.011465992778539658, + 0.019138585776090622, + 0.009252911433577538, + 0.012877176515758038, + 0.017205767333507538, + 0.0062715462408959866, + 0.010753396898508072, + 0.009971554391086102, + 0.012389452196657658, + 0.014048969373106956, + 0.007585206534713507, + 0.011830929666757584, + 0.015525968745350838, + 0.005423652473837137, + 0.0073503656312823296, + 0.008938654325902462, + 0.010110399685800076, + 0.016086220741271973, + 0.0036094719544053078, + 0.007693827617913485, + 0.0074792238883674145, + 0.0029378444887697697, + 0.005323841702193022, + 0.0035892087034881115, + 0.009239085018634796, + 0.010899998247623444, + 0.0025052670389413834, + 0.012178298085927963, + 0.007255133241415024, + 0.0022601275704801083, + 0.004537981003522873, + 0.0033976221457123756, + 0.013323681429028511, + 0.01220322959125042, + 0.0036197274457663298, + 0.01644451916217804, + 0.006290128454566002, + 0.003667666809633374, + 0.009318341501057148, + 0.004491235129535198, + 0.017569608986377716, + 0.08103092014789581, + 0.02097075618803501, + 0.05090712010860443, + 0.021275658160448074, + 0.014022747054696083, + 0.021070165559649467, + 0.014013377018272877, + 0.031817786395549774, + 0.012723032385110855, + 0.006828929763287306, + 0.01772359199821949, + 0.011686158366501331, + 0.006189960055053234, + 0.01620171219110489, + 0.008010929450392723, + 0.005875797476619482, + 0.018424663692712784, + 0.019136764109134674, + 0.016899079084396362, + 0.014945898205041885, + 0.010266449302434921, + 0.01387537270784378, + 0.015206199139356613, + 0.01530787255614996, + 0.012713956646621227, + 0.012613452039659023, + 0.009731980971992016, + 0.009518059901893139, + 0.007432945538312197, + 0.006133709102869034, + 0.008573404513299465, + 0.0068442197516560555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.009806433692574501 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013448490062728524 + ], + "opposite_logprob_tokens": [ + -6.626344680786133 + ], + "logprob_reward": -0.0013448490062728524, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013448490062728524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12364600598812103, + 0.07920065522193909, + 0.057319629937410355, + 0.048440102487802505, + 0.0175666231662035, + 0.044978342950344086, + 0.025688253343105316, + 0.40509033203125, + 0.11870146542787552, + 0.07936858385801315 + ], + "credit_map": [ + 0.022391268983483315, + 0.01483928319066763, + 0.017564354464411736, + 0.026356618851423264, + 0.008920825086534023, + 0.0096884211525321, + 0.012445278465747833, + 0.011439958587288857, + 0.011114101856946945, + 0.007685587741434574, + 0.01138096023350954, + 0.016889937222003937, + 0.006465916987508535, + 0.0070042856968939304, + 0.009216629900038242, + 0.009443235583603382, + 0.008883276954293251, + 0.007514725439250469, + 0.007218471262603998, + 0.011074728332459927, + 0.004806267563253641, + 0.004152882378548384, + 0.006478185765445232, + 0.007191094104200602, + 0.007045594044029713, + 0.004513788968324661, + 0.007030345033854246, + 0.009550012648105621, + 0.004606966394931078, + 0.003944833762943745, + 0.004944948013871908, + 0.006803616415709257, + 0.002596914768218994, + 0.0013975127367302775, + 0.0024780307430773973, + 0.002990986220538616, + 0.0016225444851443172, + 0.0019888735841959715, + 0.0015617316821590066, + 0.002930029295384884, + 0.008630487136542797, + 0.0029831165447831154, + 0.007941464893519878, + 0.003441021777689457, + 0.0031643626280128956, + 0.004753496032208204, + 0.0031158688943833113, + 0.010948524810373783, + 0.003538528922945261, + 0.002964692423120141, + 0.004358951933681965, + 0.003407699055969715, + 0.0020349491387605667, + 0.002593985293060541, + 0.0026559040416032076, + 0.004133543930947781, + 0.03672712668776512, + 0.030093461275100708, + 0.11557448655366898, + 0.033245235681533813, + 0.041820019483566284, + 0.03800778463482857, + 0.025971287861466408, + 0.08365096896886826, + 0.014102513901889324, + 0.013542004860937595, + 0.016711203381419182, + 0.011933583766222, + 0.010619583539664745, + 0.016850128769874573, + 0.016617141664028168, + 0.018325308337807655, + 0.010284258984029293, + 0.011949591338634491, + 0.009169758297502995, + 0.011558577418327332, + 0.008484569378197193, + 0.007664768025279045, + 0.01159024890512228, + 0.008666816167533398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0013448490062728524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101555794477463 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101555794477463, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101555794477463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07520981132984161, + 0.03795605152845383, + 0.037611912935972214, + 0.043326836079359055, + 0.0274229533970356, + 0.05893469974398613, + 0.06826024502515793, + 0.28099292516708374, + 0.2448190599679947, + 0.1254654824733734 + ], + "credit_map": [ + 0.011592634953558445, + 0.010999354533851147, + 0.010484210215508938, + 0.014634175226092339, + 0.005469949450343847, + 0.006693300791084766, + 0.007568693254143, + 0.007767497096210718, + 0.006805762182921171, + 0.004455387592315674, + 0.0047709085047245026, + 0.00693938322365284, + 0.003093443578109145, + 0.003535938449203968, + 0.0034966517705470324, + 0.004858576226979494, + 0.0049410066567361355, + 0.004681825637817383, + 0.0055120256729424, + 0.008109534159302711, + 0.0031369675416499376, + 0.0028476619627326727, + 0.004049736075103283, + 0.004333154298365116, + 0.006659798789769411, + 0.004670280031859875, + 0.006901274900883436, + 0.006875417660921812, + 0.004246656782925129, + 0.00401642918586731, + 0.003892807988449931, + 0.006064168643206358, + 0.004076871555298567, + 0.002899926621466875, + 0.0029384642839431763, + 0.005288726184517145, + 0.00319170905277133, + 0.0025435597635805607, + 0.0028134749736636877, + 0.00367022049613297, + 0.010589885525405407, + 0.003712292993441224, + 0.01075694803148508, + 0.0044296481646597385, + 0.0036569482181221247, + 0.005911810789257288, + 0.004383293446153402, + 0.015493875369429588, + 0.015637468546628952, + 0.007034298498183489, + 0.012453975155949593, + 0.007580708712339401, + 0.0045552002266049385, + 0.006343421991914511, + 0.005627655889838934, + 0.009027515538036823, + 0.02624092996120453, + 0.01913011446595192, + 0.06378869712352753, + 0.028742453083395958, + 0.031133603304624557, + 0.028101321309804916, + 0.02978229708969593, + 0.05407349765300751, + 0.03292220085859299, + 0.028674151748418808, + 0.03501676395535469, + 0.02346615493297577, + 0.021664060652256012, + 0.03345045819878578, + 0.034007132053375244, + 0.035618141293525696, + 0.020080557093024254, + 0.0231646578758955, + 0.01458748709410429, + 0.013683431781828403, + 0.010651844553649426, + 0.011468521319329739, + 0.01665647327899933, + 0.015172515995800495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.004101555794477463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763652054592967 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.00763652054592967, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763652054592967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03877565637230873, + 0.019573086872696877, + 0.012960023246705532, + 0.02078692801296711, + 0.042682524770498276, + 0.027957754209637642, + 0.024245761334896088, + 0.29937437176704407, + 0.3854011595249176, + 0.12824273109436035 + ], + "credit_map": [ + 0.007501556538045406, + 0.005091821309179068, + 0.00582171231508255, + 0.006453217007219791, + 0.0026041846722364426, + 0.0034629597794264555, + 0.0035819292534142733, + 0.004258275032043457, + 0.004797970876097679, + 0.0014586937613785267, + 0.003199800616130233, + 0.0024699403438717127, + 0.0014951031189411879, + 0.001930416445247829, + 0.001604930846951902, + 0.0026162315625697374, + 0.002693518530577421, + 0.0011947470484301448, + 0.0019547638949006796, + 0.0017715550493448973, + 0.001116087194532156, + 0.001336667570285499, + 0.0013238051906228065, + 0.0015688783023506403, + 0.003299692878499627, + 0.00187730323523283, + 0.003342747688293457, + 0.0027270475402474403, + 0.0022020472679287195, + 0.002221164293587208, + 0.002050012582913041, + 0.0030669141560792923, + 0.004259099252521992, + 0.004106509033590555, + 0.007684935815632343, + 0.006167433690279722, + 0.004785436671227217, + 0.003971518017351627, + 0.003931034356355667, + 0.007776557933539152, + 0.005372559651732445, + 0.001543485326692462, + 0.00520339934155345, + 0.001973663456737995, + 0.0014531684573739767, + 0.0029891717713326216, + 0.0019562027882784605, + 0.007466104347258806, + 0.005338314455002546, + 0.0024518028367310762, + 0.004264961928129196, + 0.0027893914375454187, + 0.0016540992073714733, + 0.002415102906525135, + 0.00238333223387599, + 0.00294875493273139, + 0.026476599276065826, + 0.02433614805340767, + 0.0846167653799057, + 0.02526790089905262, + 0.030935967341065407, + 0.03700167313218117, + 0.020488275215029716, + 0.050251029431819916, + 0.04222273826599121, + 0.05337017402052879, + 0.0532531812787056, + 0.03954476863145828, + 0.03408413007855415, + 0.05185597762465477, + 0.053006354719400406, + 0.058063846081495285, + 0.015298417769372463, + 0.02281099557876587, + 0.01296284794807434, + 0.01867145672440529, + 0.01903211884200573, + 0.012690546922385693, + 0.015279055573046207, + 0.011497287079691887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.00763652054592967 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunningly static close-up, fixed from a level angle against a completely black, blurred background, reveals a smooth, porcelain-like white mask. A single, cold spotlight shines diagonally from above, carving a stark contrast of light and dark into its surface, highlighting the hollow, black eye sockets and a fixed, slightly eerie smile. Suddenly, tiny, spider-web-like cracks, cautiously starting at the corners of the smiling mouth, spread rapidly and silently, instantly covering the entire smooth surface. Then, without warning, the mask violently and outwardly explodes into countless white fragments of varying sizes. As if time has slowed down, these sharp ceramic shards fly through the air, each reflecting a sharp light before dissipating into the darkness, creating a powerful visual impact and tense atmosphere, a transition from oppressive calm to instantaneous collapse.", + "seed": 26492, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_859_seed_26492_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.007091739773750305, + "vlm_logprob": -0.007091739773750305, + "vlm_yes_no_margin": 5.625, + "vlm_token_credit": -0.007091739773750305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005464260466396809 + ], + "opposite_logprob_tokens": [ + -7.13046407699585 + ], + "logprob_reward": -0.005464260466396809, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.005464260466396809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.161410391330719, + 0.08820940554141998, + 0.03943600878119469, + 0.0834679827094078, + 0.08258840441703796, + 0.09840546548366547, + 0.09998326003551483, + 0.07517674565315247, + 0.13981053233146667, + 0.1315118372440338 + ], + "credit_map": [ + 0.018620381131768227, + 0.01820492185652256, + 0.01396912056952715, + 0.010046374052762985, + 0.025285134091973305, + 0.02814648486673832, + 0.018060801550745964, + 0.02907717414200306, + 0.011140981689095497, + 0.010143092833459377, + 0.011193986050784588, + 0.005559710320085287, + 0.015304667875170708, + 0.012289867736399174, + 0.009703407995402813, + 0.012873680330812931, + 0.005615010857582092, + 0.00478591863065958, + 0.0048021492548286915, + 0.0027376513462513685, + 0.006052174139767885, + 0.0048539359122514725, + 0.004447628278285265, + 0.006141538731753826, + 0.013119382783770561, + 0.007268805056810379, + 0.013737070374190807, + 0.004790081642568111, + 0.015169800259172916, + 0.01081118918955326, + 0.007457349449396133, + 0.011114292778074741, + 0.011244340799748898, + 0.007113893516361713, + 0.011677004396915436, + 0.004382146988064051, + 0.010529992170631886, + 0.013986657373607159, + 0.006540392525494099, + 0.017113972455263138, + 0.011445785872638226, + 0.012625905685126781, + 0.00928965862840414, + 0.006842292845249176, + 0.016469476744532585, + 0.014643617905676365, + 0.010350247845053673, + 0.01673847809433937, + 0.015268695540726185, + 0.007814855314791203, + 0.018721165135502815, + 0.00683183316141367, + 0.009527255780994892, + 0.014747481793165207, + 0.0115848109126091, + 0.015487147495150566, + 0.010140755213797092, + 0.004827415104955435, + 0.00946222897619009, + 0.004607043229043484, + 0.007472727447748184, + 0.01525680348277092, + 0.008087247610092163, + 0.01532252412289381, + 0.015828581526875496, + 0.014282620511949062, + 0.016084346920251846, + 0.0157567597925663, + 0.018191952258348465, + 0.020038779824972153, + 0.022570529952645302, + 0.01705694943666458, + 0.011236924678087234, + 0.01299658790230751, + 0.014691382646560669, + 0.01862596906721592, + 0.014635542407631874, + 0.022175107151269913, + 0.019351113587617874, + 0.017799202352762222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.005464260466396809 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018216218799352646 + ], + "opposite_logprob_tokens": [ + -4.018216133117676 + ], + "logprob_reward": -0.018216218799352646, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018216218799352646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14608493447303772, + 0.1603807508945465, + 0.11485108733177185, + 0.04255059361457825, + 0.1328779011964798, + 0.05050188675522804, + 0.041515644639730453, + 0.030609631910920143, + 0.09526839107275009, + 0.1853591799736023 + ], + "credit_map": [ + 0.02113613486289978, + 0.015289882197976112, + 0.0169084370136261, + 0.00881864782422781, + 0.02251618728041649, + 0.022464988753199577, + 0.013153782114386559, + 0.025796862319111824, + 0.022851450368762016, + 0.009354582987725735, + 0.030491411685943604, + 0.007768849842250347, + 0.020265601575374603, + 0.025073716416954994, + 0.010649784468114376, + 0.03392534330487251, + 0.024284224957227707, + 0.007620302960276604, + 0.023100558668375015, + 0.009290803223848343, + 0.010469251312315464, + 0.014041503891348839, + 0.009602827951312065, + 0.016441622748970985, + 0.006576987449079752, + 0.00220013246871531, + 0.005802398547530174, + 0.0023331984411925077, + 0.00398296769708395, + 0.008408777415752411, + 0.003623298369348049, + 0.009622832760214806, + 0.013353770598769188, + 0.0056766243651509285, + 0.016751036047935486, + 0.006067759357392788, + 0.010394658893346786, + 0.0337742380797863, + 0.00837631430476904, + 0.03848349675536156, + 0.004735363647341728, + 0.004970976151525974, + 0.0043143630027771, + 0.002764585195109248, + 0.005111871287226677, + 0.011468852870166302, + 0.004643318708986044, + 0.012492557056248188, + 0.004573560319840908, + 0.002929938957095146, + 0.005578835494816303, + 0.0022330365609377623, + 0.0035343323834240437, + 0.009598206728696823, + 0.0031729654874652624, + 0.009894770570099354, + 0.004111797548830509, + 0.0020863825920969248, + 0.003637883346527815, + 0.0026216229889541864, + 0.0037138257175683975, + 0.004499645903706551, + 0.005822476465255022, + 0.004115997813642025, + 0.009852270595729351, + 0.011538158170878887, + 0.009920366108417511, + 0.013473967090249062, + 0.011946974322199821, + 0.010109921917319298, + 0.017309918999671936, + 0.011116813868284225, + 0.01789248175919056, + 0.015526587143540382, + 0.03732709586620331, + 0.02698749303817749, + 0.0177130326628685, + 0.030331676825881004, + 0.016774127259850502, + 0.022806696593761444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.018216218799352646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022082962095737457 + ], + "opposite_logprob_tokens": [ + -6.127208232879639 + ], + "logprob_reward": -0.0022082962095737457, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022082962095737457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13760140538215637, + 0.050271447747945786, + 0.03508971258997917, + 0.04073134809732437, + 0.08684535324573517, + 0.05473539978265762, + 0.11040806770324707, + 0.046472907066345215, + 0.24151059985160828, + 0.19633378088474274 + ], + "credit_map": [ + 0.015457729808986187, + 0.012718431651592255, + 0.01515822485089302, + 0.00936410203576088, + 0.021010423079133034, + 0.01969478279352188, + 0.019406141713261604, + 0.02479156106710434, + 0.009398117661476135, + 0.0038482248783111572, + 0.008429289795458317, + 0.003774933749809861, + 0.005754620768129826, + 0.00664770882576704, + 0.005825882777571678, + 0.006592671852558851, + 0.006277739070355892, + 0.0027890473138540983, + 0.005137448199093342, + 0.002516168402507901, + 0.004496408626437187, + 0.005600208882242441, + 0.003349620383232832, + 0.0049230726435780525, + 0.004703887272626162, + 0.0026656033005565405, + 0.006230723578482866, + 0.0030130730010569096, + 0.005678258370608091, + 0.007151931524276733, + 0.005403746385127306, + 0.005884123966097832, + 0.015160543844103813, + 0.007363630924373865, + 0.011610765010118484, + 0.004672909155488014, + 0.009096304886043072, + 0.01319613866508007, + 0.008849935606122017, + 0.016895122826099396, + 0.006412323098629713, + 0.005244705826044083, + 0.00410941056907177, + 0.003028174629434943, + 0.006236087065190077, + 0.012737712822854519, + 0.003991938196122646, + 0.012975044548511505, + 0.00853080302476883, + 0.010985629633069038, + 0.00796311441808939, + 0.007957159541547298, + 0.00938174594193697, + 0.027263471856713295, + 0.010787587612867355, + 0.027538558468222618, + 0.005047470796853304, + 0.004271024372428656, + 0.006468158680945635, + 0.003512510797008872, + 0.005227616522461176, + 0.006808949168771505, + 0.007821140810847282, + 0.0073160333558917046, + 0.022327158600091934, + 0.027889154851436615, + 0.028277790173888206, + 0.03429032489657402, + 0.028858279809355736, + 0.024022985249757767, + 0.04672785848379135, + 0.02911706268787384, + 0.017355836927890778, + 0.017982086166739464, + 0.023360712453722954, + 0.024378003552556038, + 0.02538527175784111, + 0.040693316608667374, + 0.021468540653586388, + 0.025710009038448334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0022082962095737457 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742346566170454 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.006742346566170454, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742346566170454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06056744605302811, + 0.03625502437353134, + 0.032952044159173965, + 0.02809825725853443, + 0.021386919543147087, + 0.025283485651016235, + 0.06555920839309692, + 0.14250218868255615, + 0.18554019927978516, + 0.4018552899360657 + ], + "credit_map": [ + 0.008385292254388332, + 0.004961919505149126, + 0.013297872617840767, + 0.004450108855962753, + 0.008168273605406284, + 0.006084949243813753, + 0.005686459597200155, + 0.009532561525702477, + 0.00553991599008441, + 0.003976538777351379, + 0.00603501545265317, + 0.002550079021602869, + 0.005637967027723789, + 0.004465240053832531, + 0.0029111127369105816, + 0.005139152053743601, + 0.006949798669666052, + 0.0026240020524710417, + 0.0059997886419296265, + 0.0023478891234844923, + 0.004725029692053795, + 0.0033442918211221695, + 0.002754073590040207, + 0.004207169637084007, + 0.0062482417561113834, + 0.0019920344930142164, + 0.005100988317281008, + 0.0020006736740469933, + 0.002951287431642413, + 0.003447270253673196, + 0.003069712780416012, + 0.0032880434300750494, + 0.0028778754640370607, + 0.001795486081391573, + 0.0031714874785393476, + 0.0012604902731254697, + 0.002590357093140483, + 0.00360571569763124, + 0.002050278941169381, + 0.004035226535052061, + 0.0047347936779260635, + 0.0024147762451320887, + 0.003384295152500272, + 0.001724951551295817, + 0.0024553609546273947, + 0.0035900436341762543, + 0.0024710397701710463, + 0.004508221987634897, + 0.00786253809928894, + 0.00432948861271143, + 0.010092141106724739, + 0.004330597352236509, + 0.00740437489002943, + 0.012748990207910538, + 0.006302727852016687, + 0.012488347478210926, + 0.019598163664340973, + 0.012463982217013836, + 0.011739321984350681, + 0.012865732423961163, + 0.01190169993788004, + 0.028289102017879486, + 0.01328760851174593, + 0.03235657140612602, + 0.01566070131957531, + 0.025542786344885826, + 0.02345116063952446, + 0.02488522045314312, + 0.019828645512461662, + 0.01764714904129505, + 0.036638591438531876, + 0.02188592590391636, + 0.0406467467546463, + 0.042605187743902206, + 0.050082843750715256, + 0.051177892833948135, + 0.039135050028562546, + 0.0797952264547348, + 0.04771745204925537, + 0.05069485306739807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.006742346566170454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028275768272578716 + ], + "opposite_logprob_tokens": [ + -5.8778276443481445 + ], + "logprob_reward": -0.0028275768272578716, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028275768272578716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045859195291996, + 0.024702157825231552, + 0.018252218142151833, + 0.01634194701910019, + 0.014730751514434814, + 0.03246919438242912, + 0.038220569491386414, + 0.18169015645980835, + 0.1933756023645401, + 0.4343582093715668 + ], + "credit_map": [ + 0.004981901030987501, + 0.004599031060934067, + 0.005397978704422712, + 0.0034459643065929413, + 0.0065324269235134125, + 0.00749728549271822, + 0.005031171720474958, + 0.008373437449336052, + 0.004377256613224745, + 0.002105977851897478, + 0.0038958885706961155, + 0.0013946526451036334, + 0.003341322997584939, + 0.0034050645772367716, + 0.00213589402846992, + 0.004046101588755846, + 0.0032241621520370245, + 0.0012234047753736377, + 0.0030182546470314264, + 0.0013705637538805604, + 0.0019345874898135662, + 0.0028239742387086153, + 0.0021184412762522697, + 0.002538830740377307, + 0.0025057257153093815, + 0.00146670441608876, + 0.0014146794565021992, + 0.0009840470738708973, + 0.002154045272618532, + 0.0032187728211283684, + 0.0014510893961414695, + 0.003146884497255087, + 0.0026148930191993713, + 0.0010109596187248826, + 0.00197801785543561, + 0.0009883551392704248, + 0.001670881174504757, + 0.002406054874882102, + 0.0013912089634686708, + 0.0026703807525336742, + 0.005763851571828127, + 0.0015282864915207028, + 0.004038461018353701, + 0.001579783158376813, + 0.002852223115041852, + 0.007102930452674627, + 0.002057313919067383, + 0.007546343374997377, + 0.0030952307861298323, + 0.0025528951082378626, + 0.0029143933206796646, + 0.0021287654526531696, + 0.002881810301914811, + 0.010665666311979294, + 0.0033340505324304104, + 0.010647762566804886, + 0.02598349191248417, + 0.01435716450214386, + 0.016843117773532867, + 0.01675988733768463, + 0.016282588243484497, + 0.03297760710120201, + 0.02257526107132435, + 0.03591103106737137, + 0.014379095286130905, + 0.022623011842370033, + 0.016800852492451668, + 0.018434299156069756, + 0.0382867231965065, + 0.029223544523119926, + 0.03126908466219902, + 0.022358981892466545, + 0.04274357855319977, + 0.04424627125263214, + 0.04517386853694916, + 0.05076877027750015, + 0.051615774631500244, + 0.0982872024178505, + 0.04986456781625748, + 0.05165819078683853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0028275768272578716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunningly static close-up, fixed from a level angle against a completely black, blurred background, reveals a smooth, porcelain-like white mask. A single, cold spotlight shines diagonally from above, carving a stark contrast of light and dark into its surface, highlighting the hollow, black eye sockets and a fixed, slightly eerie smile. Suddenly, tiny, spider-web-like cracks, cautiously starting at the corners of the smiling mouth, spread rapidly and silently, instantly covering the entire smooth surface. Then, without warning, the mask violently and outwardly explodes into countless white fragments of varying sizes. As if time has slowed down, these sharp ceramic shards fly through the air, each reflecting a sharp light before dissipating into the darkness, creating a powerful visual impact and tense atmosphere, a transition from oppressive calm to instantaneous collapse.", + "seed": 26493, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_2/index_859_seed_26493_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.004766056197695434, + "vlm_logprob": -0.004766056197695434, + "vlm_yes_no_margin": 6.125, + "vlm_token_credit": -0.004766056197695434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002444495679810643 + ], + "opposite_logprob_tokens": [ + -7.002444267272949 + ], + "logprob_reward": -0.002444495679810643, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.002444495679810643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14506272971630096, + 0.04365762323141098, + 0.06722861528396606, + 0.19105584919452667, + 0.1435903012752533, + 0.06049659475684166, + 0.06748069822788239, + 0.03737719729542732, + 0.10843497514724731, + 0.13561545312404633 + ], + "credit_map": [ + 0.00972482468932867, + 0.005215120501816273, + 0.03375811129808426, + 0.008633351884782314, + 0.010440144687891006, + 0.008636893704533577, + 0.04253120720386505, + 0.02612307108938694, + 0.004152136389166117, + 0.00260186567902565, + 0.011705192737281322, + 0.003379685804247856, + 0.0035614671651273966, + 0.0028758561238646507, + 0.007058444898575544, + 0.008322971872985363, + 0.009330227971076965, + 0.0034456131979823112, + 0.01985095627605915, + 0.002999169286340475, + 0.003253006609156728, + 0.005075193475931883, + 0.007249081507325172, + 0.016025366261601448, + 0.025253133848309517, + 0.008954377844929695, + 0.0741119459271431, + 0.008746548555791378, + 0.007815956138074398, + 0.010488581843674183, + 0.02670019306242466, + 0.02898510918021202, + 0.02035204879939556, + 0.006834196858108044, + 0.05332601070404053, + 0.006287231110036373, + 0.006175596732646227, + 0.009014863520860672, + 0.014414769597351551, + 0.027185579761862755, + 0.007196495775133371, + 0.0031673512421548367, + 0.017870405688881874, + 0.0031685028225183487, + 0.003813584800809622, + 0.00523707689717412, + 0.0077611845917999744, + 0.012281994335353374, + 0.006234475877135992, + 0.005028005223721266, + 0.010508696548640728, + 0.006901329383254051, + 0.008263804018497467, + 0.004691943991929293, + 0.014973621815443039, + 0.010878819972276688, + 0.002866156166419387, + 0.003845737548545003, + 0.004853877238929272, + 0.004175318870693445, + 0.004572027362883091, + 0.003700322937220335, + 0.006539825350046158, + 0.006823930889368057, + 0.00954421330243349, + 0.006964850705116987, + 0.009560395032167435, + 0.008907055482268333, + 0.03578054904937744, + 0.010350607335567474, + 0.018046816810965538, + 0.009280484169721603, + 0.013385745696723461, + 0.007862396538257599, + 0.02410592883825302, + 0.010493171401321888, + 0.0219732578843832, + 0.008372618816792965, + 0.01628446951508522, + 0.03313785791397095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.002444495679810643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160913597792387 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.0160913597792387, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160913597792387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11586125940084457, + 0.05842221900820732, + 0.18567338585853577, + 0.0623466894030571, + 0.164420023560524, + 0.030077828094363213, + 0.04692191258072853, + 0.05295052006840706, + 0.08394979685544968, + 0.1993764489889145 + ], + "credit_map": [ + 0.014402501285076141, + 0.006795323453843594, + 0.04309665784239769, + 0.007066736929118633, + 0.006375209894031286, + 0.007010323461145163, + 0.012320443987846375, + 0.018794063478708267, + 0.006804106757044792, + 0.0033603236079216003, + 0.020020078867673874, + 0.004435856826603413, + 0.003395219799131155, + 0.0028763851150870323, + 0.013379544019699097, + 0.004150697495788336, + 0.029977723956108093, + 0.010604443028569221, + 0.02917674370110035, + 0.010904799215495586, + 0.01420488953590393, + 0.016000239178538322, + 0.028795715421438217, + 0.046008821576833725, + 0.008679443970322609, + 0.0030457840766757727, + 0.02054459974169731, + 0.002983632730320096, + 0.0034138576593250036, + 0.004383757710456848, + 0.007559779100120068, + 0.011735827662050724, + 0.02176808752119541, + 0.0072943982668221, + 0.06470510363578796, + 0.01100139133632183, + 0.008792047388851643, + 0.009039005264639854, + 0.025611136108636856, + 0.01620885357260704, + 0.003083043033257127, + 0.002897540805861354, + 0.00460211094468832, + 0.0028022138867527246, + 0.0026937120128422976, + 0.003739304142072797, + 0.004592347890138626, + 0.005667554680258036, + 0.0036017729435116053, + 0.004394978284835815, + 0.006138854194432497, + 0.006191697437316179, + 0.006846778094768524, + 0.004585675895214081, + 0.0069672041572630405, + 0.008194944821298122, + 0.003330858424305916, + 0.0038394455332309008, + 0.00924636609852314, + 0.0061330790631473064, + 0.00409687077626586, + 0.003914912696927786, + 0.011434831656515598, + 0.010954149067401886, + 0.008255922235548496, + 0.003921885509043932, + 0.005298946518450975, + 0.005441241431981325, + 0.03164501488208771, + 0.007485794369131327, + 0.01634436473250389, + 0.005556621588766575, + 0.016289426013827324, + 0.012288561090826988, + 0.0364655964076519, + 0.016140146180987358, + 0.03148723393678665, + 0.012079136446118355, + 0.025279449298977852, + 0.04934687912464142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.0160913597792387 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010489681735634804 + ], + "opposite_logprob_tokens": [ + -6.876049041748047 + ], + "logprob_reward": -0.0010489681735634804, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010489681735634804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10997214913368225, + 0.06803403794765472, + 0.09325297176837921, + 0.07274375110864639, + 0.03664539381861687, + 0.14693675935268402, + 0.07592692971229553, + 0.2734376788139343, + 0.058598246425390244, + 0.06445206701755524 + ], + "credit_map": [ + 0.012798930518329144, + 0.006861488800495863, + 0.011584981344640255, + 0.007347471546381712, + 0.008310839533805847, + 0.00920598953962326, + 0.02467440627515316, + 0.029188035055994987, + 0.004539805930107832, + 0.0033541759476065636, + 0.01234935037791729, + 0.0047153872437775135, + 0.004477596841752529, + 0.004666313529014587, + 0.019915176555514336, + 0.014016225934028625, + 0.013514915481209755, + 0.005037625320255756, + 0.015006473287940025, + 0.004560308530926704, + 0.005788578651845455, + 0.008233916945755482, + 0.015521363355219364, + 0.02558978460729122, + 0.009246334433555603, + 0.003197798039764166, + 0.030066728591918945, + 0.003928868565708399, + 0.0033358149230480194, + 0.003243550891056657, + 0.012145964428782463, + 0.007578686811029911, + 0.004626046400517225, + 0.0019007058581337333, + 0.006592949852347374, + 0.0018864248413592577, + 0.0025876814033836126, + 0.003156048245728016, + 0.00595029117539525, + 0.009945244528353214, + 0.018359940499067307, + 0.0093781603500247, + 0.021448424085974693, + 0.007629544008523226, + 0.011850923299789429, + 0.014432969503104687, + 0.02343597076833248, + 0.040400803089141846, + 0.00615786062553525, + 0.006185682490468025, + 0.010067473165690899, + 0.010413141921162605, + 0.010759547352790833, + 0.007123175077140331, + 0.013233351521193981, + 0.011986689642071724, + 0.022952962666749954, + 0.019334938377141953, + 0.024232499301433563, + 0.015437847003340721, + 0.024136485531926155, + 0.034370459616184235, + 0.04195411503314972, + 0.09101834148168564, + 0.005925500765442848, + 0.006110835820436478, + 0.007471511140465736, + 0.009276344440877438, + 0.01019263919442892, + 0.0067632184363901615, + 0.0077706617303192616, + 0.0050875283777713776, + 0.005588707514107227, + 0.006939427927136421, + 0.008567272685468197, + 0.010504673235118389, + 0.01027288194745779, + 0.006756243761628866, + 0.007880288176238537, + 0.007942567579448223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.0010489681735634804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031967272516340017 + ], + "opposite_logprob_tokens": [ + -5.753196716308594 + ], + "logprob_reward": -0.0031967272516340017, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031967272516340017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03114738129079342, + 0.05446084588766098, + 0.07651713490486145, + 0.045289214700460434, + 0.0762496367096901, + 0.028828315436840057, + 0.022596552968025208, + 0.33234062790870667, + 0.18246841430664062, + 0.1501019299030304 + ], + "credit_map": [ + 0.003942748066037893, + 0.001925118500366807, + 0.004851495381444693, + 0.001998628955334425, + 0.002386578358709812, + 0.002420655218884349, + 0.00689441105350852, + 0.006727747153490782, + 0.006792915519326925, + 0.0027004536241292953, + 0.007891098968684673, + 0.003363773925229907, + 0.003521895734593272, + 0.003898846684023738, + 0.013221141882240772, + 0.013070717453956604, + 0.012570125982165337, + 0.004499209113419056, + 0.014719441533088684, + 0.003481355495750904, + 0.004236361477524042, + 0.006603689398616552, + 0.01003244984894991, + 0.02037450298666954, + 0.006863337475806475, + 0.002198893344029784, + 0.015020268969237804, + 0.0026471298187971115, + 0.0020396269392222166, + 0.002984070684760809, + 0.00610073609277606, + 0.0074351513758301735, + 0.010849427431821823, + 0.0037497084122151136, + 0.026390312239527702, + 0.003155370941385627, + 0.003252710448578, + 0.005060276482254267, + 0.006940273102372885, + 0.016851551830768585, + 0.003610512474551797, + 0.0016167304711416364, + 0.008828694932162762, + 0.0016907253302633762, + 0.0020562356803566217, + 0.0024937514681369066, + 0.004485490266233683, + 0.004046175163239241, + 0.00205729017034173, + 0.0013985548866912723, + 0.003568765940144658, + 0.00260818749666214, + 0.002439166884869337, + 0.0023654275573790073, + 0.0032837872859090567, + 0.004875371232628822, + 0.0227299015969038, + 0.021632913500070572, + 0.035451725125312805, + 0.01875501498579979, + 0.0389549694955349, + 0.04168585687875748, + 0.05273948609828949, + 0.1003907322883606, + 0.01589840091764927, + 0.011649588122963905, + 0.01058568712323904, + 0.011041280813515186, + 0.06521756947040558, + 0.016539795324206352, + 0.03501071780920029, + 0.016525372862815857, + 0.012681875377893448, + 0.009371197782456875, + 0.029157226905226707, + 0.013831219635903835, + 0.024018211290240288, + 0.009630288928747177, + 0.020101848989725113, + 0.031310055404901505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0031967272516340017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of a smooth, porcelain-like white mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background completely black and blurred with a single cold spotlight shining diagonally from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mask feature hollow black eye sockets and a fixed, slightly eerie smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny spider-web-like cracks start at the corners of the mouth and rapidly spread to cover the entire surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mask violently explode into countless white ceramic shards that fly through the air before dissipating?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010487301042303443 + ], + "opposite_logprob_tokens": [ + -6.876048564910889 + ], + "logprob_reward": -0.0010487301042303443, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010487301042303443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05524838715791702, + 0.050614453852176666, + 0.0803438127040863, + 0.058953844010829926, + 0.08156485110521317, + 0.03139983490109444, + 0.04283563420176506, + 0.2022980898618698, + 0.08294965326786041, + 0.313791424036026 + ], + "credit_map": [ + 0.0052327546291053295, + 0.0029980987310409546, + 0.009066218510270119, + 0.003097429173067212, + 0.0037958326283842325, + 0.004613677505403757, + 0.012397630140185356, + 0.014046750031411648, + 0.005331601481884718, + 0.002305156085640192, + 0.017829975113272667, + 0.0031675659120082855, + 0.0027460551355034113, + 0.0021703068632632494, + 0.01153462752699852, + 0.005529163405299187, + 0.013072306290268898, + 0.004690452478826046, + 0.016364002600312233, + 0.0037618237547576427, + 0.004212082363665104, + 0.006792942062020302, + 0.010264841839671135, + 0.021185362711548805, + 0.008046872913837433, + 0.002636483171954751, + 0.023300230503082275, + 0.0032213495578616858, + 0.0025463667698204517, + 0.003064691787585616, + 0.00913581158965826, + 0.007002036552876234, + 0.011035839095711708, + 0.0037303322460502386, + 0.03187054023146629, + 0.0035829534754157066, + 0.0033410917967557907, + 0.00485612265765667, + 0.009447540156543255, + 0.013700432144105434, + 0.003931152168661356, + 0.001624340657144785, + 0.008098476566374302, + 0.001723601366393268, + 0.0022775521501898766, + 0.0026185454335063696, + 0.003692171536386013, + 0.007433992810547352, + 0.0037937997840344906, + 0.004151697736233473, + 0.004493277054280043, + 0.006933942437171936, + 0.006446141749620438, + 0.004543081391602755, + 0.006028209812939167, + 0.006445483770221472, + 0.0132920416072011, + 0.013750219717621803, + 0.020613381639122963, + 0.014548043720424175, + 0.022344978526234627, + 0.02454645000398159, + 0.031700607389211655, + 0.061502378433942795, + 0.007051371969282627, + 0.004506060853600502, + 0.004285104107111692, + 0.004506515804678202, + 0.032896388322114944, + 0.00588965555652976, + 0.01872917264699936, + 0.005085383076220751, + 0.02457026019692421, + 0.015248731710016727, + 0.05751924216747284, + 0.017827630043029785, + 0.05102817341685295, + 0.020068539306521416, + 0.043289002031087875, + 0.08423985540866852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.0010487301042303443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_tensor([9647, 9647])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_tensor([9647, 9647])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ee39caa9b5d3fbc736d1a116769769da65b96b8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_tensor([9647, 9647])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful, low-angle, low-angle tracking shot captures the peak moment of a massive monster truck's leap over a towering sand dune. The truck, clad in a gleaming electric blue, adorned with a striking white lightning bolt, its massive chrome suspension and massive shock absorbers fully exposed in the air, a display of pure mechanical power. Under the glaring afternoon sun, the fine golden sand contrasts starkly against the azure sky, the sunlight reflecting off the truck's sleek metal hull in dazzling flares. As the truck briefly hovers at the apex of its leap, its four massive, sand-encrusted tires, deep in tread, break free from the ground and spin madly through the air, creating a visual blur and flinging sand backward in a subtle trail. The truck's massive bulk carves a majestic arc through the air, its deep shadow cast across the dune slope below. The entire scene is imbued with a sense of raw power and adrenaline-pumping dynamic aesthetic, creating a thrilling atmosphere of transcending limits and defying gravity.", + "seed": 14008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_9647_seed_14008_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009154827357269824, + "vlm_logprob": -0.009154827357269824, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.009154827357269824, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014693381264805794 + ], + "opposite_logprob_tokens": [ + -4.514693260192871 + ], + "logprob_reward": -0.014693381264805794, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014693381264805794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13503380119800568, + 0.0930296778678894, + 0.2269389033317566, + 0.08058138191699982, + 0.06002861261367798, + 0.0340191014111042, + 0.049916189163923264, + 0.038646962493658066, + 0.06730140745639801, + 0.2145039290189743 + ], + "credit_map": [ + 0.010432510636746883, + 0.014858439564704895, + 0.019276315346360207, + 0.03067012131214142, + 0.00861553754657507, + 0.014939101412892342, + 0.01411511655896902, + 0.022126659750938416, + 0.0075968485325574875, + 0.006292878650128841, + 0.016262320801615715, + 0.017007751390337944, + 0.006668010260909796, + 0.009809301234781742, + 0.015406210906803608, + 0.01398635283112526, + 0.0122244693338871, + 0.014330307021737099, + 0.03449253737926483, + 0.045071084052324295, + 0.012548133730888367, + 0.04502784088253975, + 0.033813778311014175, + 0.029430758208036423, + 0.005919937044382095, + 0.00796483550220728, + 0.010462530888617039, + 0.008770652115345001, + 0.014180036261677742, + 0.010465706698596478, + 0.0108634689822793, + 0.011954206973314285, + 0.006583654787391424, + 0.00635046511888504, + 0.007888366468250751, + 0.005151478108018637, + 0.007332474924623966, + 0.009072193875908852, + 0.007751864846795797, + 0.009898114018142223, + 0.004811753518879414, + 0.00492793507874012, + 0.004126731771975756, + 0.003456976730376482, + 0.00466860830783844, + 0.004869299940764904, + 0.003999348729848862, + 0.003158448962494731, + 0.009014442563056946, + 0.005325856152921915, + 0.010035717859864235, + 0.004362666513770819, + 0.004943785723298788, + 0.005586578045040369, + 0.004356107674539089, + 0.006291036494076252, + 0.0075039430521428585, + 0.005018333904445171, + 0.0054370625875890255, + 0.00588208856061101, + 0.003696717321872711, + 0.0033445374574512243, + 0.0033763099927455187, + 0.004387971945106983, + 0.007689466699957848, + 0.010534112341701984, + 0.009008703753352165, + 0.007127550896257162, + 0.009416607208549976, + 0.005082389339804649, + 0.008669543080031872, + 0.009773032739758492, + 0.033716849982738495, + 0.036928609013557434, + 0.03190169855952263, + 0.029987702146172523, + 0.022684359923005104, + 0.013968287967145443, + 0.024586590006947517, + 0.020729821175336838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.014693381264805794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205172657966614 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014205172657966614, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205172657966614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1947375237941742, + 0.10194229334592819, + 0.13251858949661255, + 0.30968502163887024, + 0.09106897562742233, + 0.027768515050411224, + 0.041623372584581375, + 0.02269500121474266, + 0.025489574298262596, + 0.052471134811639786 + ], + "credit_map": [ + 0.014224301092326641, + 0.023489437997341156, + 0.02570449374616146, + 0.036262962967157364, + 0.016827018931508064, + 0.017376137897372246, + 0.023540683090686798, + 0.037312477827072144, + 0.005441895220428705, + 0.006577633786946535, + 0.02176838368177414, + 0.024174034595489502, + 0.006612530443817377, + 0.008562332019209862, + 0.018661024048924446, + 0.010144454427063465, + 0.00846732035279274, + 0.00914928037673235, + 0.01889795996248722, + 0.019028393551707268, + 0.009301966056227684, + 0.0304409172385931, + 0.014933711849153042, + 0.022299040108919144, + 0.022228987887501717, + 0.033321481198072433, + 0.02601122297346592, + 0.03421797603368759, + 0.05256430804729462, + 0.04520440474152565, + 0.04049002006649971, + 0.05564660206437111, + 0.012262689881026745, + 0.010251265950500965, + 0.008512549102306366, + 0.006173667497932911, + 0.01887894980609417, + 0.012886002659797668, + 0.009679440408945084, + 0.012424406595528126, + 0.005337451584637165, + 0.003652045503258705, + 0.00344213773496449, + 0.0027930738870054483, + 0.0037092622369527817, + 0.0045646363869309425, + 0.0020730048418045044, + 0.0021969019435346127, + 0.007017322350293398, + 0.0038765522185713053, + 0.009782693348824978, + 0.0034189855214208364, + 0.0037630321457982063, + 0.004721506033092737, + 0.0028214093763381243, + 0.0062218657694756985, + 0.004537679720669985, + 0.0027799573726952076, + 0.0031885458156466484, + 0.0029755099676549435, + 0.002498196205124259, + 0.0022207750007510185, + 0.0023239878937602043, + 0.0021703490056097507, + 0.003494866658002138, + 0.0033702042419463396, + 0.0038349530659615993, + 0.0033296162728220224, + 0.0033362223766744137, + 0.0017447441350668669, + 0.004051741678267717, + 0.002327222377061844, + 0.007927777245640755, + 0.009108291007578373, + 0.007219091057777405, + 0.006835949141532183, + 0.008447869680821896, + 0.004201180767267942, + 0.003062402131035924, + 0.005668570753186941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.014205172657966614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096688237041235 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096688237041235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096688237041235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1341509371995926, + 0.1819804310798645, + 0.04175516590476036, + 0.05144353583455086, + 0.2634827792644501, + 0.11315974593162537, + 0.04689031094312668, + 0.06685766577720642, + 0.06204891949892044, + 0.03823056444525719 + ], + "credit_map": [ + 0.013134768232703209, + 0.015939507633447647, + 0.017210183665156364, + 0.02400449849665165, + 0.009700982831418514, + 0.012131117284297943, + 0.017517749220132828, + 0.024512112140655518, + 0.014515669085085392, + 0.014320182614028454, + 0.029426565393805504, + 0.036461856216192245, + 0.01463024877011776, + 0.02097596973180771, + 0.0255732461810112, + 0.026076676324009895, + 0.0027158057782799006, + 0.0034436811693012714, + 0.005760392174124718, + 0.00697380630299449, + 0.003495783545076847, + 0.008248042315244675, + 0.0061843679286539555, + 0.004933283198624849, + 0.0027719102799892426, + 0.0044603352434933186, + 0.007768571842461824, + 0.005486875772476196, + 0.004084540996700525, + 0.011556826531887054, + 0.007870051078498363, + 0.007444420363754034, + 0.052897676825523376, + 0.013059590943157673, + 0.02181149460375309, + 0.011804593726992607, + 0.010447962209582329, + 0.0251180250197649, + 0.0402674525976181, + 0.08807595819234848, + 0.015030561946332455, + 0.011139387264847755, + 0.023140033707022667, + 0.009009639732539654, + 0.011502123437821865, + 0.014249696396291256, + 0.014390245079994202, + 0.01469803974032402, + 0.007160490844398737, + 0.004011768382042646, + 0.01079643052071333, + 0.0036379925440996885, + 0.004000997636467218, + 0.0052942452020943165, + 0.004069543909281492, + 0.007918840274214745, + 0.011581509374082088, + 0.008365779183804989, + 0.009135319851338863, + 0.009288918226957321, + 0.006786946672946215, + 0.00526427710428834, + 0.0070792739279568195, + 0.009355638176202774, + 0.00961692538112402, + 0.009445969015359879, + 0.007769607938826084, + 0.007314022164791822, + 0.006958280224353075, + 0.006103878375142813, + 0.006556537933647633, + 0.008283696137368679, + 0.005899686366319656, + 0.007275853771716356, + 0.005629296880215406, + 0.005116511136293411, + 0.00430965656414628, + 0.002942216582596302, + 0.003219006583094597, + 0.0038383312057703733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.004096688237041235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001712523982860148 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.001712523982860148, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001712523982860148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11983464658260345, + 0.14725780487060547, + 0.12603048980236053, + 0.08506761491298676, + 0.0972655788064003, + 0.06739699095487595, + 0.11797881871461868, + 0.043987877666950226, + 0.07396860420703888, + 0.12121160328388214 + ], + "credit_map": [ + 0.011393804103136063, + 0.01318306289613247, + 0.0154257295653224, + 0.019473180174827576, + 0.0058116368018090725, + 0.009468048810958862, + 0.017344161868095398, + 0.027735015377402306, + 0.014186698943376541, + 0.010832095518708229, + 0.02293211966753006, + 0.023758210241794586, + 0.010766169987618923, + 0.01889432780444622, + 0.02134125865995884, + 0.024546919390559196, + 0.00502387760207057, + 0.005432125646620989, + 0.023358218371868134, + 0.04263186454772949, + 0.006471455097198486, + 0.008439753204584122, + 0.026347743347287178, + 0.008325442671775818, + 0.005544853396713734, + 0.009357775561511517, + 0.007125054020434618, + 0.011813976801931858, + 0.012601200491189957, + 0.011245918460190296, + 0.011815986596047878, + 0.015562846325337887, + 0.016709931194782257, + 0.007402818650007248, + 0.011838516220450401, + 0.005729301832616329, + 0.007359481882303953, + 0.01124766655266285, + 0.012978305108845234, + 0.02399955503642559, + 0.01032453216612339, + 0.008036226965487003, + 0.008819513022899628, + 0.006421786732971668, + 0.008926104754209518, + 0.00906718336045742, + 0.008669070899486542, + 0.007132567930966616, + 0.018857229501008987, + 0.0143311507999897, + 0.021023308858275414, + 0.01131828036159277, + 0.014100136235356331, + 0.012073791585862637, + 0.010386995039880276, + 0.015887919813394547, + 0.007675507105886936, + 0.004447929561138153, + 0.006251875776797533, + 0.00672165909782052, + 0.00494619645178318, + 0.0037354675587266684, + 0.004261078778654337, + 0.005948160775005817, + 0.009687047451734543, + 0.009694650769233704, + 0.009985013864934444, + 0.007397495675832033, + 0.012281621806323528, + 0.008225755766034126, + 0.00772826187312603, + 0.00896875187754631, + 0.0162162184715271, + 0.025519825518131256, + 0.01484366599470377, + 0.012812933884561062, + 0.014465628191828728, + 0.011027371510863304, + 0.01051262766122818, + 0.015813319012522697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.001712523982860148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066370643675327 + ], + "opposite_logprob_tokens": [ + -4.511066436767578 + ], + "logprob_reward": -0.011066370643675327, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066370643675327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12435663491487503, + 0.1985517144203186, + 0.05608123168349266, + 0.2335381954908371, + 0.07683702558279037, + 0.05302773043513298, + 0.04095612093806267, + 0.061661120504140854, + 0.12156733870506287, + 0.03342296555638313 + ], + "credit_map": [ + 0.012231026776134968, + 0.012518567964434624, + 0.014939508400857449, + 0.020021721720695496, + 0.008583847433328629, + 0.013858133926987648, + 0.016490835696458817, + 0.02571297250688076, + 0.014149420894682407, + 0.012152465991675854, + 0.03739064186811447, + 0.038444120436906815, + 0.012771042995154858, + 0.021661968901753426, + 0.033490896224975586, + 0.028491150587797165, + 0.005813149735331535, + 0.0037296738009899855, + 0.008311019279062748, + 0.011097010225057602, + 0.0038650818169116974, + 0.00749451806768775, + 0.008417508564889431, + 0.007353262510150671, + 0.013608734123408794, + 0.014685812406241894, + 0.03487684950232506, + 0.035832010209560394, + 0.017432596534490585, + 0.030503619462251663, + 0.041196323931217194, + 0.045402225106954575, + 0.012944702990353107, + 0.0035332441329956055, + 0.009061329066753387, + 0.005006278399378061, + 0.003671706886962056, + 0.00755414878949523, + 0.014199145138263702, + 0.02086646296083927, + 0.007855253294110298, + 0.005732450634241104, + 0.007254889700561762, + 0.004106476437300444, + 0.0063111684285104275, + 0.009378597140312195, + 0.006345527246594429, + 0.006043360102921724, + 0.006321250926703215, + 0.0042478712275624275, + 0.008517825044691563, + 0.003459645202383399, + 0.004498991649597883, + 0.0049886638298630714, + 0.0036778878420591354, + 0.005243981722742319, + 0.011022533290088177, + 0.007173921912908554, + 0.00858637597411871, + 0.009353987872600555, + 0.007156258914619684, + 0.004560742527246475, + 0.006929956842213869, + 0.006877340376377106, + 0.018051695078611374, + 0.014055296778678894, + 0.0169072188436985, + 0.01582668349146843, + 0.012631991878151894, + 0.009337859228253365, + 0.016002172604203224, + 0.018754417076706886, + 0.004939097911119461, + 0.007207408081740141, + 0.005354668013751507, + 0.0033326696138828993, + 0.003713141893967986, + 0.00290268799290061, + 0.0022476688027381897, + 0.003725616028532386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.011066370643675327 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, low-angle, low-angle tracking shot captures the peak moment of a massive monster truck's leap over a towering sand dune. The truck, clad in a gleaming electric blue, adorned with a striking white lightning bolt, its massive chrome suspension and massive shock absorbers fully exposed in the air, a display of pure mechanical power. Under the glaring afternoon sun, the fine golden sand contrasts starkly against the azure sky, the sunlight reflecting off the truck's sleek metal hull in dazzling flares. As the truck briefly hovers at the apex of its leap, its four massive, sand-encrusted tires, deep in tread, break free from the ground and spin madly through the air, creating a visual blur and flinging sand backward in a subtle trail. The truck's massive bulk carves a majestic arc through the air, its deep shadow cast across the dune slope below. The entire scene is imbued with a sense of raw power and adrenaline-pumping dynamic aesthetic, creating a thrilling atmosphere of transcending limits and defying gravity.", + "seed": 14009, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_9647_seed_14009_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.005385286966338753, + "vlm_logprob": -0.005385286966338753, + "vlm_yes_no_margin": 5.949999904632568, + "vlm_token_credit": -0.005385286966338753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00558802904561162 + ], + "opposite_logprob_tokens": [ + -8.005587577819824 + ], + "logprob_reward": -0.00558802904561162, + "yes_no_margin_reward": 7.999999523162842, + "token_credit_reward": -0.00558802904561162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10565018653869629, + 0.05128546804189682, + 0.07897287607192993, + 0.13692645728588104, + 0.10838208347558975, + 0.10905840992927551, + 0.06296443939208984, + 0.11936663091182709, + 0.17397595942020416, + 0.05341744422912598 + ], + "credit_map": [ + 0.013462630100548267, + 0.016842301934957504, + 0.013974499888718128, + 0.013155123218894005, + 0.01123829185962677, + 0.013091986067593098, + 0.011797702871263027, + 0.01208765059709549, + 0.0060903578996658325, + 0.006246773526072502, + 0.0061418344266712666, + 0.00593685545027256, + 0.007882881909608841, + 0.005684948991984129, + 0.008008540607988834, + 0.00529327429831028, + 0.018118448555469513, + 0.004638425074517727, + 0.007290209643542767, + 0.005910434760153294, + 0.007840138860046864, + 0.010791699402034283, + 0.011925522238016129, + 0.012457998469471931, + 0.020202163606882095, + 0.02283109910786152, + 0.015164860524237156, + 0.012587112374603748, + 0.021106190979480743, + 0.018014011904597282, + 0.011211351491510868, + 0.015809668228030205, + 0.010170813649892807, + 0.01506131887435913, + 0.019609080627560616, + 0.015240289270877838, + 0.01369697879999876, + 0.0113001624122262, + 0.012375294230878353, + 0.010928149335086346, + 0.02666347287595272, + 0.012307953089475632, + 0.024866847321391106, + 0.009691068902611732, + 0.009052157402038574, + 0.012107297778129578, + 0.005543916020542383, + 0.008825693279504776, + 0.008427605032920837, + 0.008630125783383846, + 0.005756292957812548, + 0.00934251956641674, + 0.009348418563604355, + 0.009332321584224701, + 0.007011450361460447, + 0.005115708336234093, + 0.016887445002794266, + 0.021585926413536072, + 0.007676779292523861, + 0.017269643023610115, + 0.012670978903770447, + 0.021456941962242126, + 0.009593178518116474, + 0.012225741520524025, + 0.026874741539359093, + 0.028118586167693138, + 0.03872628137469292, + 0.018105965107679367, + 0.01618293486535549, + 0.012100225314497948, + 0.016658075153827667, + 0.017209142446517944, + 0.00791709404438734, + 0.009462251327931881, + 0.005645320750772953, + 0.007336675189435482, + 0.006392485927790403, + 0.0051408871077001095, + 0.0064037214033305645, + 0.005119008012115955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.00558802904561162 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956872832030058 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956872832030058, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956872832030058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05917057767510414, + 0.04375029355287552, + 0.09356804937124252, + 0.04630552977323532, + 0.06563110649585724, + 0.14722543954849243, + 0.10163719207048416, + 0.17706841230392456, + 0.23712511360645294, + 0.028518272563815117 + ], + "credit_map": [ + 0.007575486786663532, + 0.010831590741872787, + 0.006503027398139238, + 0.005461033899337053, + 0.007496962323784828, + 0.007126680109649897, + 0.006771969608962536, + 0.007403826806694269, + 0.00546456640586257, + 0.006690466310828924, + 0.004118806682527065, + 0.004838237538933754, + 0.007388452999293804, + 0.0057874820195138454, + 0.005606240127235651, + 0.003856041468679905, + 0.010075724683701992, + 0.0078023625537753105, + 0.012440936639904976, + 0.010900849476456642, + 0.010206788778305054, + 0.010211985558271408, + 0.021651189774274826, + 0.010278214700520039, + 0.0070128110237419605, + 0.004422035999596119, + 0.006773578468710184, + 0.004767258185893297, + 0.004343569278717041, + 0.004900082480162382, + 0.005248188506811857, + 0.008838005363941193, + 0.006721359211951494, + 0.006337577942758799, + 0.008019047789275646, + 0.004879734944552183, + 0.0178238358348608, + 0.010346075519919395, + 0.005794399883598089, + 0.005709078162908554, + 0.027948621660470963, + 0.012141410261392593, + 0.024840522557497025, + 0.011409997940063477, + 0.015241227112710476, + 0.03207647427916527, + 0.009722917340695858, + 0.013844283297657967, + 0.016011754050850868, + 0.01568259671330452, + 0.00945450272411108, + 0.010092013515532017, + 0.0189307052642107, + 0.015829740092158318, + 0.008409513160586357, + 0.007226369809359312, + 0.02650638110935688, + 0.027849024161696434, + 0.00971970520913601, + 0.026009073480963707, + 0.024529090151190758, + 0.02756061777472496, + 0.015426196157932281, + 0.019468313083052635, + 0.035702552646398544, + 0.03696221858263016, + 0.04033247008919716, + 0.024878203868865967, + 0.03216070309281349, + 0.02459479309618473, + 0.022003352642059326, + 0.02049081027507782, + 0.003442778717726469, + 0.003438354004174471, + 0.003924908582121134, + 0.00406854972243309, + 0.002550909761339426, + 0.0029079518280923367, + 0.005314457230269909, + 0.002870361553505063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.005956872832030058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03549974411725998, + 0.05230286717414856, + 0.06343632191419601, + 0.07232268899679184, + 0.04178359732031822, + 0.0811922624707222, + 0.10289076715707779, + 0.14069347083568573, + 0.37896621227264404, + 0.030912047252058983 + ], + "credit_map": [ + 0.006194377318024635, + 0.004951910115778446, + 0.0037315357476472855, + 0.0031157841440290213, + 0.004170978907495737, + 0.005016581155359745, + 0.0032176878303289413, + 0.005100891459733248, + 0.00796726904809475, + 0.005432393867522478, + 0.005418798886239529, + 0.005422479938715696, + 0.0067413002252578735, + 0.0065756612457334995, + 0.007714210078120232, + 0.00703075248748064, + 0.007069002836942673, + 0.006266778334975243, + 0.007150337100028992, + 0.006423614919185638, + 0.008543387986719608, + 0.007559637073427439, + 0.01433068048208952, + 0.006092884577810764, + 0.012317140586674213, + 0.007895372807979584, + 0.010109064169228077, + 0.006545451004058123, + 0.006160583347082138, + 0.009295286610722542, + 0.007398467510938644, + 0.012601328082382679, + 0.004014825914055109, + 0.005975080654025078, + 0.005811466835439205, + 0.005271261557936668, + 0.008149745874106884, + 0.004628413822501898, + 0.004022960551083088, + 0.003909839317202568, + 0.01798763871192932, + 0.007336688693612814, + 0.01809890754520893, + 0.006338050123304129, + 0.006510716397315264, + 0.0135728120803833, + 0.004733960144221783, + 0.00661347946152091, + 0.016160978004336357, + 0.0242674108594656, + 0.007493073120713234, + 0.018143121153116226, + 0.00853461492806673, + 0.010146028362214565, + 0.009321440942585468, + 0.008824099786579609, + 0.018112050369381905, + 0.025304434821009636, + 0.010021328926086426, + 0.02238032966852188, + 0.015164640732109547, + 0.020189378410577774, + 0.012414281256496906, + 0.017107022926211357, + 0.056363534182310104, + 0.06000296398997307, + 0.07306421548128128, + 0.04052411764860153, + 0.04432287439703941, + 0.03294568508863449, + 0.03730766475200653, + 0.03443516418337822, + 0.003473324468359351, + 0.0034142276272177696, + 0.005193457007408142, + 0.006545640993863344, + 0.003545824671164155, + 0.0026442648377269506, + 0.003718820633366704, + 0.0023764860816299915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.00978010892868042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015112657565623522 + ], + "opposite_logprob_tokens": [ + -6.501511096954346 + ], + "logprob_reward": -0.0015112657565623522, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015112657565623522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08517083525657654, + 0.054233480244874954, + 0.08230603486299515, + 0.0884338766336441, + 0.15018165111541748, + 0.15397359430789948, + 0.11047641932964325, + 0.10077857971191406, + 0.1324603110551834, + 0.04198522865772247 + ], + "credit_map": [ + 0.01233676541596651, + 0.017171749845147133, + 0.008987873792648315, + 0.0070265308022499084, + 0.010336430743336678, + 0.012961138039827347, + 0.00803699903190136, + 0.008313341997563839, + 0.006961441598832607, + 0.006385178305208683, + 0.004478195682168007, + 0.004318087827414274, + 0.006006140727549791, + 0.007190112955868244, + 0.0071756974793970585, + 0.011718623340129852, + 0.01199472788721323, + 0.00629866449162364, + 0.009748503565788269, + 0.008717799559235573, + 0.008799320086836815, + 0.01018968503922224, + 0.017366787418723106, + 0.00919053889811039, + 0.015266481786966324, + 0.01681140623986721, + 0.008565401658415794, + 0.006678079254925251, + 0.012853394262492657, + 0.013181909918785095, + 0.0071294791996479034, + 0.00794772058725357, + 0.01373059581965208, + 0.01664232276380062, + 0.017478682100772858, + 0.019761817529797554, + 0.035953547805547714, + 0.019334888085722923, + 0.015191812068223953, + 0.012087990529835224, + 0.021261155605316162, + 0.025442158803343773, + 0.019859375432133675, + 0.020037349313497543, + 0.015397053211927414, + 0.021380534395575523, + 0.011437948793172836, + 0.0191580168902874, + 0.01649206504225731, + 0.020082497969269753, + 0.008656791411340237, + 0.01618501916527748, + 0.00948986317962408, + 0.014826055616140366, + 0.01528939325362444, + 0.009454725310206413, + 0.015174737200140953, + 0.014543435536324978, + 0.009198159910738468, + 0.013228506781160831, + 0.012827545404434204, + 0.013920024037361145, + 0.011261158622801304, + 0.010625011287629604, + 0.01986616477370262, + 0.01884162239730358, + 0.021243656054139137, + 0.014124983921647072, + 0.017866846174001694, + 0.013582819141447544, + 0.01475600153207779, + 0.012178205884993076, + 0.00562812527641654, + 0.0048813470639288425, + 0.006537469569593668, + 0.007422045338898897, + 0.0038061593659222126, + 0.003364863805472851, + 0.005931704770773649, + 0.004413512535393238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.0015112657565623522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040901582688093185 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.0040901582688093185, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040901582688093185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05753404274582863, + 0.043148357421159744, + 0.04389546811580658, + 0.04866745322942734, + 0.1611497551202774, + 0.2873191237449646, + 0.0658685564994812, + 0.050980765372514725, + 0.21370826661586761, + 0.027728283777832985 + ], + "credit_map": [ + 0.007131743244826794, + 0.012030170299112797, + 0.005482570268213749, + 0.003823703620582819, + 0.007062842603772879, + 0.008721604943275452, + 0.006263484712690115, + 0.007017916068434715, + 0.005986507516354322, + 0.0038379381876438856, + 0.0037951883859932423, + 0.003886812599375844, + 0.002948224311694503, + 0.006358189508318901, + 0.005532938055694103, + 0.010802554897964, + 0.007049115840345621, + 0.00397950503975153, + 0.004876827821135521, + 0.004063144791871309, + 0.006014705169945955, + 0.006450405344367027, + 0.006343493238091469, + 0.005118264351040125, + 0.007882371544837952, + 0.008449447341263294, + 0.004520233720541, + 0.004869543015956879, + 0.00595077546313405, + 0.00682642450556159, + 0.004574492108076811, + 0.005594164598733187, + 0.01600770652294159, + 0.018890105187892914, + 0.018597939983010292, + 0.019588347524404526, + 0.03826798498630524, + 0.02045155130326748, + 0.016239024698734283, + 0.01310708001255989, + 0.0557827465236187, + 0.044655099511146545, + 0.055034089833498, + 0.034332480281591415, + 0.024550244212150574, + 0.02530757524073124, + 0.018411362543702126, + 0.029245518147945404, + 0.011080116964876652, + 0.014682312496006489, + 0.004763421602547169, + 0.010073303245007992, + 0.00704508600756526, + 0.005040204152464867, + 0.007699804380536079, + 0.0054843067191541195, + 0.005797101650387049, + 0.005764473229646683, + 0.00891611073166132, + 0.006689881440252066, + 0.004372583236545324, + 0.006071425974369049, + 0.007875682786107063, + 0.005493502598255873, + 0.03486757352948189, + 0.03588368371129036, + 0.03488876670598984, + 0.0213698111474514, + 0.02058618888258934, + 0.016901513561606407, + 0.026131603866815567, + 0.02307910844683647, + 0.003948762081563473, + 0.003914088010787964, + 0.004573405254632235, + 0.0050559500232338905, + 0.0021165821235626936, + 0.0017449313309043646, + 0.0036182983312755823, + 0.002756264293566346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.0040901582688093185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, low-angle, low-angle tracking shot captures the peak moment of a massive monster truck's leap over a towering sand dune. The truck, clad in a gleaming electric blue, adorned with a striking white lightning bolt, its massive chrome suspension and massive shock absorbers fully exposed in the air, a display of pure mechanical power. Under the glaring afternoon sun, the fine golden sand contrasts starkly against the azure sky, the sunlight reflecting off the truck's sleek metal hull in dazzling flares. As the truck briefly hovers at the apex of its leap, its four massive, sand-encrusted tires, deep in tread, break free from the ground and spin madly through the air, creating a visual blur and flinging sand backward in a subtle trail. The truck's massive bulk carves a majestic arc through the air, its deep shadow cast across the dune slope below. The entire scene is imbued with a sense of raw power and adrenaline-pumping dynamic aesthetic, creating a thrilling atmosphere of transcending limits and defying gravity.", + "seed": 14010, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_9647_seed_14010_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0056683451402932405, + "vlm_logprob": -0.0056683451402932405, + "vlm_yes_no_margin": 5.7, + "vlm_token_credit": -0.0056683451402932405, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007910826243460178 + ], + "opposite_logprob_tokens": [ + -6.75791072845459 + ], + "logprob_reward": -0.007910826243460178, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.007910826243460178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20170152187347412, + 0.1146736890077591, + 0.08846908807754517, + 0.05345567688345909, + 0.058375630527734756, + 0.20535431802272797, + 0.07517282664775848, + 0.06431011110544205, + 0.07248934358358383, + 0.06599780172109604 + ], + "credit_map": [ + 0.038504306226968765, + 0.015086907893419266, + 0.03393423929810524, + 0.017840774729847908, + 0.015487614087760448, + 0.04621162638068199, + 0.01864207722246647, + 0.015993990004062653, + 0.01578822173178196, + 0.018851542845368385, + 0.013339655473828316, + 0.015303708612918854, + 0.015029660426080227, + 0.014425224624574184, + 0.011725690215826035, + 0.010209983214735985, + 0.010057683102786541, + 0.012797643430531025, + 0.011670747771859169, + 0.009736839681863785, + 0.013979500159621239, + 0.009000496938824654, + 0.011379150673747063, + 0.009847017005085945, + 0.009068992920219898, + 0.00467572221532464, + 0.012454502284526825, + 0.0046528661623597145, + 0.00631284061819315, + 0.006532413884997368, + 0.005138360895216465, + 0.004619975574314594, + 0.009043730795383453, + 0.00801021233201027, + 0.007466632407158613, + 0.009403318166732788, + 0.009416974149644375, + 0.006516468245536089, + 0.003622603602707386, + 0.004895688034594059, + 0.03057604841887951, + 0.023877792060375214, + 0.03163920342922211, + 0.02425268292427063, + 0.026524191722273827, + 0.026415444910526276, + 0.01589992269873619, + 0.026169026270508766, + 0.00898982398211956, + 0.009340939112007618, + 0.009959772229194641, + 0.00961212906986475, + 0.010478829964995384, + 0.00971419457346201, + 0.00669996440410614, + 0.01037717703729868, + 0.01059636753052473, + 0.005660302471369505, + 0.008963785134255886, + 0.0078555503860116, + 0.004909532144665718, + 0.009559258818626404, + 0.005786426831036806, + 0.010978884063661098, + 0.009068736806511879, + 0.012022518552839756, + 0.007384103257209063, + 0.0070022582076489925, + 0.006789408158510923, + 0.011983384378254414, + 0.006108841393142939, + 0.012130090035498142, + 0.007330676540732384, + 0.008911279030144215, + 0.007693880703300238, + 0.010574987158179283, + 0.008955066092312336, + 0.007518071215599775, + 0.0074121905490756035, + 0.007601654622703791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.007910826243460178 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790142998099327 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.009790142998099327, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790142998099327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10026127099990845, + 0.1442662626504898, + 0.2706233561038971, + 0.12484735250473022, + 0.04738707095384598, + 0.16505782306194305, + 0.08034858852624893, + 0.031950488686561584, + 0.019776606932282448, + 0.015481185168027878 + ], + "credit_map": [ + 0.020754752680659294, + 0.01359737291932106, + 0.012051080353558064, + 0.007831753231585026, + 0.009868849068880081, + 0.0221208818256855, + 0.007669483311474323, + 0.006367097608745098, + 0.016812222078442574, + 0.020433513447642326, + 0.013656988739967346, + 0.019376931712031364, + 0.024200959131121635, + 0.01605960913002491, + 0.017967589199543, + 0.0157584547996521, + 0.03365596756339073, + 0.028322089463472366, + 0.0394144207239151, + 0.021167347207665443, + 0.04077804833650589, + 0.026020284742116928, + 0.04470794275403023, + 0.036557238548994064, + 0.02342013642191887, + 0.009047139436006546, + 0.03709876164793968, + 0.006436939351260662, + 0.01613221876323223, + 0.018625764176249504, + 0.005769575946033001, + 0.008316813968122005, + 0.006871264427900314, + 0.006498424801975489, + 0.006131377071142197, + 0.006916239857673645, + 0.008289999328553677, + 0.005370596889406443, + 0.0037267461884766817, + 0.003582426579669118, + 0.029520638287067413, + 0.018849270418286324, + 0.027400871738791466, + 0.01737799309194088, + 0.020908527076244354, + 0.02090402878820896, + 0.011166579090058804, + 0.018929919227957726, + 0.00813323725014925, + 0.011369682848453522, + 0.011908743530511856, + 0.00993963610380888, + 0.010792402550578117, + 0.009895379655063152, + 0.00582606066018343, + 0.0124834468588233, + 0.0038783687632530928, + 0.004452188033610582, + 0.004405319225043058, + 0.004528058227151632, + 0.002697138348594308, + 0.005318887997418642, + 0.002641078317537904, + 0.004029450006783009, + 0.0025272706989198923, + 0.003472898155450821, + 0.00225774641148746, + 0.00220540352165699, + 0.001744597451761365, + 0.0037752201315015554, + 0.0015217317268252373, + 0.0022717383690178394, + 0.0015177293680608273, + 0.0023711449466645718, + 0.0019080586498603225, + 0.0020124379079788923, + 0.0013846164802089334, + 0.0020596750546246767, + 0.002362513681873679, + 0.0018650086130946875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.009790142998099327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196252044290304 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.003196252044290304, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196252044290304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07771559059619904, + 0.12585175037384033, + 0.20093418657779694, + 0.04134810343384743, + 0.10789817571640015, + 0.17464390397071838, + 0.10972032696008682, + 0.07242884486913681, + 0.04545820131897926, + 0.04400090500712395 + ], + "credit_map": [ + 0.013906258158385754, + 0.006422535516321659, + 0.012285197153687477, + 0.008815156295895576, + 0.007206034380942583, + 0.011728832498192787, + 0.009373621083796024, + 0.007977962493896484, + 0.021377990022301674, + 0.020943373441696167, + 0.013038241304457188, + 0.011165927164256573, + 0.012308751232922077, + 0.020834963768720627, + 0.012176694348454475, + 0.014005823992192745, + 0.02087787538766861, + 0.025284336879849434, + 0.02212173119187355, + 0.02214420773088932, + 0.030828695744276047, + 0.01934080198407173, + 0.038135264068841934, + 0.02220127359032631, + 0.0058721862733364105, + 0.003958021756261587, + 0.006552479229867458, + 0.004005039110779762, + 0.0063091786578297615, + 0.005333033390343189, + 0.004978894256055355, + 0.004339275881648064, + 0.015139451250433922, + 0.013520771637558937, + 0.01379530131816864, + 0.01187183242291212, + 0.018892956897616386, + 0.01174038928002119, + 0.008295170031487942, + 0.014642306603491306, + 0.024945970624685287, + 0.01891336590051651, + 0.023653877899050713, + 0.01706146076321602, + 0.02923581562936306, + 0.019718032330274582, + 0.013840045779943466, + 0.02727534808218479, + 0.013682516291737556, + 0.01067941915243864, + 0.02037750743329525, + 0.010152467526495457, + 0.008808411657810211, + 0.015502509661018848, + 0.0071900528855621815, + 0.023327456787228584, + 0.009095598943531513, + 0.009987921454012394, + 0.010970786213874817, + 0.010131588205695152, + 0.006052692420780659, + 0.008351904340088367, + 0.005303326994180679, + 0.012535031884908676, + 0.007062762975692749, + 0.006117277313023806, + 0.005997654050588608, + 0.004986828658729792, + 0.0045038554817438126, + 0.004197698086500168, + 0.0044388012029230595, + 0.008153324015438557, + 0.004331639967858791, + 0.008393166586756706, + 0.004711302928626537, + 0.007353936787694693, + 0.00417289836332202, + 0.005057099740952253, + 0.005984281189739704, + 0.003996585495769978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.003196252044290304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021964015904814005 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021964015904814005, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021964015904814005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06662067025899887, + 0.07214438170194626, + 0.10031124204397202, + 0.04319367557764053, + 0.10880497843027115, + 0.14882490038871765, + 0.08923953771591187, + 0.1914680153131485, + 0.1111295074224472, + 0.06826309114694595 + ], + "credit_map": [ + 0.012166706845164299, + 0.007235421799123287, + 0.01025571208447218, + 0.005097412038594484, + 0.0069223493337631226, + 0.014662333764135838, + 0.004664038773626089, + 0.00561669422313571, + 0.008169574663043022, + 0.008880740962922573, + 0.010533315129578114, + 0.008436801843345165, + 0.008375588804483414, + 0.012632033787667751, + 0.005367103032767773, + 0.009749225340783596, + 0.012333153747022152, + 0.013489559292793274, + 0.012008066289126873, + 0.012766794301569462, + 0.011029938235878944, + 0.011086219921708107, + 0.016649767756462097, + 0.01094774715602398, + 0.006020044907927513, + 0.004873529076576233, + 0.008530240505933762, + 0.004442318808287382, + 0.005968913901597261, + 0.004763770382851362, + 0.004515800625085831, + 0.004079058300703764, + 0.015003662556409836, + 0.01790691912174225, + 0.014731628820300102, + 0.013780221343040466, + 0.01870620623230934, + 0.012036853469908237, + 0.009423763491213322, + 0.007215721998363733, + 0.024619270116090775, + 0.016240829601883888, + 0.02585122548043728, + 0.01341301016509533, + 0.019574958831071854, + 0.018307840451598167, + 0.011076829396188259, + 0.01974094659090042, + 0.009467508643865585, + 0.009403541684150696, + 0.017718946561217308, + 0.009051830507814884, + 0.009649756364524364, + 0.013441186398267746, + 0.005972901359200478, + 0.014533864334225655, + 0.03555578365921974, + 0.014126735739409924, + 0.024248993024230003, + 0.021139871329069138, + 0.012105973437428474, + 0.039675552397966385, + 0.012116760015487671, + 0.032498352229595184, + 0.014625412411987782, + 0.016190901398658752, + 0.014142029918730259, + 0.014535943046212196, + 0.01239259447902441, + 0.015166851691901684, + 0.010906357318162918, + 0.0131694246083498, + 0.0060273269191384315, + 0.013457059860229492, + 0.005876497365534306, + 0.010286066681146622, + 0.008526456542313099, + 0.010756946168839931, + 0.007504839450120926, + 0.0058279018849134445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.0021964015904814005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248102825134993 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005248102825134993, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248102825134993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1492476612329483, + 0.10293322056531906, + 0.0704088881611824, + 0.03644203767180443, + 0.10140108317136765, + 0.14370225369930267, + 0.07146941125392914, + 0.10378534346818924, + 0.13713999092578888, + 0.08347013592720032 + ], + "credit_map": [ + 0.018237685784697533, + 0.014981506392359734, + 0.029942134395241737, + 0.013827706687152386, + 0.013847919180989265, + 0.02759137563407421, + 0.014919143170118332, + 0.01590019278228283, + 0.013057447038590908, + 0.013173236511647701, + 0.012612762860953808, + 0.009834042750298977, + 0.01358774397522211, + 0.01855478622019291, + 0.009078300558030605, + 0.013034893199801445, + 0.008891229517757893, + 0.01007764507085085, + 0.009071395732462406, + 0.006430418696254492, + 0.0081002963706851, + 0.008603149093687534, + 0.011469414457678795, + 0.007765337359160185, + 0.005569886416196823, + 0.0049139210022985935, + 0.004161925055086613, + 0.0027628683019429445, + 0.0073208133690059185, + 0.005202434491366148, + 0.003522478276863694, + 0.0029877093620598316, + 0.017175253480672836, + 0.02057463489472866, + 0.010484804399311543, + 0.011657368391752243, + 0.016504570841789246, + 0.008224178105592728, + 0.006359356455504894, + 0.010420913808047771, + 0.0230479147285223, + 0.017583565786480904, + 0.02531001716852188, + 0.013576170429587364, + 0.016161782667040825, + 0.01984008029103279, + 0.011952540837228298, + 0.016230180859565735, + 0.009799341671168804, + 0.007851115427911282, + 0.01104525662958622, + 0.00811210460960865, + 0.009653235785663128, + 0.009113473817706108, + 0.0042385803535580635, + 0.011656307615339756, + 0.014007379300892353, + 0.00810941495001316, + 0.014438091777265072, + 0.013943986035883427, + 0.007568735629320145, + 0.01577117294073105, + 0.007665799930691719, + 0.02228076383471489, + 0.016682511195540428, + 0.014022705145180225, + 0.02705267444252968, + 0.007784003391861916, + 0.007788076065480709, + 0.029542524367570877, + 0.005996712949126959, + 0.02827078476548195, + 0.007913785055279732, + 0.013452359475195408, + 0.005393934436142445, + 0.013874010182917118, + 0.012949087657034397, + 0.0077440328896045685, + 0.014344471506774426, + 0.007798453792929649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.005248102825134993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, low-angle, low-angle tracking shot captures the peak moment of a massive monster truck's leap over a towering sand dune. The truck, clad in a gleaming electric blue, adorned with a striking white lightning bolt, its massive chrome suspension and massive shock absorbers fully exposed in the air, a display of pure mechanical power. Under the glaring afternoon sun, the fine golden sand contrasts starkly against the azure sky, the sunlight reflecting off the truck's sleek metal hull in dazzling flares. As the truck briefly hovers at the apex of its leap, its four massive, sand-encrusted tires, deep in tread, break free from the ground and spin madly through the air, creating a visual blur and flinging sand backward in a subtle trail. The truck's massive bulk carves a majestic arc through the air, its deep shadow cast across the dune slope below. The entire scene is imbued with a sense of raw power and adrenaline-pumping dynamic aesthetic, creating a thrilling atmosphere of transcending limits and defying gravity.", + "seed": 14011, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_3/index_9647_seed_14011_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.003743776911869645, + "vlm_logprob": -0.003743776911869645, + "vlm_yes_no_margin": 6.075, + "vlm_token_credit": -0.003743776911869645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0048996382392942905 + ], + "opposite_logprob_tokens": [ + -6.379899501800537 + ], + "logprob_reward": -0.0048996382392942905, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0048996382392942905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07723519206047058, + 0.034861452877521515, + 0.46765220165252686, + 0.04116996005177498, + 0.05275951325893402, + 0.04007744789123535, + 0.025093676522374153, + 0.043891895562410355, + 0.1319500207901001, + 0.08530859649181366 + ], + "credit_map": [ + 0.007777008693665266, + 0.010578008368611336, + 0.011642586439847946, + 0.006669818423688412, + 0.014556773006916046, + 0.007643233519047499, + 0.011751540936529636, + 0.006616226397454739, + 0.006157052703201771, + 0.005124928895384073, + 0.003715548664331436, + 0.0034180020447820425, + 0.004695114213973284, + 0.005777973216027021, + 0.0035748619120568037, + 0.0023979677353054285, + 0.16265839338302612, + 0.03276814892888069, + 0.031758688390254974, + 0.03772502765059471, + 0.03165024518966675, + 0.12748700380325317, + 0.02183537557721138, + 0.021769322454929352, + 0.006212430074810982, + 0.008238201960921288, + 0.003950698301196098, + 0.005581027362495661, + 0.004688859451562166, + 0.0037788546178489923, + 0.0041410187259316444, + 0.0045788707211613655, + 0.006952341180294752, + 0.010138406418263912, + 0.009755279868841171, + 0.00588552700355649, + 0.006834498606622219, + 0.00404374347999692, + 0.004677451215684414, + 0.004472265485674143, + 0.004718673415482044, + 0.004320479929447174, + 0.0060172718949615955, + 0.005685945972800255, + 0.004799810703843832, + 0.004439179785549641, + 0.004292787052690983, + 0.005803300067782402, + 0.002984457416459918, + 0.003032801439985633, + 0.003383661387488246, + 0.003007906721904874, + 0.002948838984593749, + 0.0029446142725646496, + 0.004523604176938534, + 0.0022677911911159754, + 0.004548555705696344, + 0.009743084199726582, + 0.004565145820379257, + 0.004692245740443468, + 0.004034122917801142, + 0.007729672361165285, + 0.002914749551564455, + 0.00566431600600481, + 0.01423825602978468, + 0.020929094403982162, + 0.014773337170481682, + 0.009934302419424057, + 0.026556607335805893, + 0.013097971677780151, + 0.01384097058326006, + 0.01857948489487171, + 0.010468073189258575, + 0.008845972828567028, + 0.006674140226095915, + 0.011790310963988304, + 0.008025244809687138, + 0.024657798931002617, + 0.008972126059234142, + 0.005874933674931526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.0048996382392942905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642531931400299 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642531931400299, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642531931400299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11658334732055664, + 0.055684152990579605, + 0.22309260070323944, + 0.05547376722097397, + 0.0636899322271347, + 0.03483344987034798, + 0.2251606583595276, + 0.11525758355855942, + 0.06965462863445282, + 0.040569968521595 + ], + "credit_map": [ + 0.013516915030777454, + 0.013581017032265663, + 0.03268156200647354, + 0.012220018543303013, + 0.014161600731313229, + 0.014418900944292545, + 0.006818200461566448, + 0.009185124188661575, + 0.00758133502677083, + 0.004461157601326704, + 0.012118050828576088, + 0.0076757571659982204, + 0.005319446325302124, + 0.006799224764108658, + 0.005035476293414831, + 0.006693699862807989, + 0.0761510580778122, + 0.015308479778468609, + 0.01573519967496395, + 0.017147649079561234, + 0.014628439210355282, + 0.06064179539680481, + 0.011553778313100338, + 0.011926171369850636, + 0.00598000967875123, + 0.008339558728039265, + 0.009073520079255104, + 0.00554010272026062, + 0.008707181550562382, + 0.005039876792579889, + 0.00591009808704257, + 0.006883414927870035, + 0.008051262237131596, + 0.009508606977760792, + 0.00813413318246603, + 0.006361601408571005, + 0.009814979508519173, + 0.006596942897886038, + 0.00871534924954176, + 0.006507052108645439, + 0.003346337703987956, + 0.004484864883124828, + 0.004633004777133465, + 0.005487963557243347, + 0.0038969472516328096, + 0.003991709090769291, + 0.004950225353240967, + 0.004042394459247589, + 0.02445005439221859, + 0.02273382991552353, + 0.028488831594586372, + 0.01960570365190506, + 0.032796461135149, + 0.029717406257987022, + 0.045471761375665665, + 0.021896585822105408, + 0.012212921865284443, + 0.01460697315633297, + 0.02520466223359108, + 0.01009245403110981, + 0.009107640013098717, + 0.01846935600042343, + 0.011390356346964836, + 0.01417321152985096, + 0.006324035581201315, + 0.011547460220754147, + 0.00928656104952097, + 0.005539105273783207, + 0.013417583890259266, + 0.008741145953536034, + 0.007279696874320507, + 0.007519031409174204, + 0.005164738744497299, + 0.006676057819277048, + 0.005123559385538101, + 0.005637160502374172, + 0.006874490063637495, + 0.003910657949745655, + 0.004476861096918583, + 0.0027064390014857054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.008642531931400299 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019407500512897968 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019407500512897968, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019407500512897968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11911405622959137, + 0.26066485047340393, + 0.23388312757015228, + 0.06815245747566223, + 0.07273918390274048, + 0.04909889027476311, + 0.08470934629440308, + 0.03617692366242409, + 0.048204757273197174, + 0.02725645899772644 + ], + "credit_map": [ + 0.009228741750121117, + 0.023238034918904305, + 0.021526167169213295, + 0.008103469386696815, + 0.017656998708844185, + 0.019534654915332794, + 0.012601821683347225, + 0.007224158849567175, + 0.03320678696036339, + 0.019029688090085983, + 0.05917156860232353, + 0.038622185587882996, + 0.024685807526111603, + 0.030489783734083176, + 0.02283078245818615, + 0.03262820839881897, + 0.07786262780427933, + 0.017880475148558617, + 0.01626514457166195, + 0.02208666130900383, + 0.016613103449344635, + 0.06202400103211403, + 0.011311577633023262, + 0.009839506819844246, + 0.00821470096707344, + 0.012316381558775902, + 0.005933801177889109, + 0.00809694454073906, + 0.008507631719112396, + 0.006575389299541712, + 0.00846617016941309, + 0.010041426867246628, + 0.009054885245859623, + 0.013687364757061005, + 0.011444564908742905, + 0.008283892646431923, + 0.011949216015636921, + 0.0069559495896101, + 0.005742127075791359, + 0.005621175281703472, + 0.006645164918154478, + 0.004912139382213354, + 0.004671418108046055, + 0.005525212734937668, + 0.004814359825104475, + 0.005565428175032139, + 0.006542283575981855, + 0.010422876104712486, + 0.010993736796081066, + 0.016901249065995216, + 0.0059935362078249454, + 0.006379900965839624, + 0.015618030913174152, + 0.013457991182804108, + 0.008064622059464455, + 0.007300274446606636, + 0.0032872147858142853, + 0.010083661414682865, + 0.003776668570935726, + 0.00415315106511116, + 0.004317695740610361, + 0.004710494074970484, + 0.0024996816646307707, + 0.0033483505249023438, + 0.0036102747544646263, + 0.005659167189151049, + 0.00924193486571312, + 0.006056936923414469, + 0.008489719592034817, + 0.004752840846776962, + 0.005350291728973389, + 0.00504358671605587, + 0.0032353296410292387, + 0.0037301762495189905, + 0.0042406655848026276, + 0.0045197284780442715, + 0.003698056098073721, + 0.0027428925968706608, + 0.0027765228878706694, + 0.0023130816407501698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0019407500512897968 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010414659045636654 + ], + "opposite_logprob_tokens": [ + -6.876041412353516 + ], + "logprob_reward": -0.0010414659045636654, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010414659045636654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11426685750484467, + 0.044836707413196564, + 0.07002250850200653, + 0.05493247136473656, + 0.10532039403915405, + 0.03856581822037697, + 0.044500384479761124, + 0.09699226915836334, + 0.3551434874534607, + 0.075419120490551 + ], + "credit_map": [ + 0.010946868918836117, + 0.01668364182114601, + 0.03328274190425873, + 0.008650037460029125, + 0.010066875256597996, + 0.01696684956550598, + 0.009356698952615261, + 0.008313138969242573, + 0.005764494650065899, + 0.005645311903208494, + 0.007135241758078337, + 0.004910606425255537, + 0.007108794990926981, + 0.0062420363537967205, + 0.004205234814435244, + 0.00382498512044549, + 0.014968576841056347, + 0.009582854807376862, + 0.005659959279000759, + 0.009199033491313457, + 0.010234200395643711, + 0.01123074721544981, + 0.004691743291914463, + 0.00445539690554142, + 0.00744448509067297, + 0.006495910696685314, + 0.00921968650072813, + 0.004601845517754555, + 0.008146868087351322, + 0.005504808854311705, + 0.006357904989272356, + 0.007160957437008619, + 0.011752291582524776, + 0.014538002200424671, + 0.015336810611188412, + 0.011701161041855812, + 0.01973157562315464, + 0.010009260848164558, + 0.013177283108234406, + 0.009074007160961628, + 0.004007943905889988, + 0.005716160871088505, + 0.006902684923261404, + 0.005295806564390659, + 0.0034188551362603903, + 0.00395397562533617, + 0.004638975486159325, + 0.004631416406482458, + 0.00525305699557066, + 0.005842957179993391, + 0.004019869491457939, + 0.005472088232636452, + 0.0047668335027992725, + 0.0052107274532318115, + 0.010090396739542484, + 0.0038444569800049067, + 0.008758602663874626, + 0.026893239468336105, + 0.018424302339553833, + 0.00871073454618454, + 0.008250715211033821, + 0.007467483635991812, + 0.006627392955124378, + 0.01185979600995779, + 0.030847016721963882, + 0.05593845993280411, + 0.036985401064157486, + 0.03417717292904854, + 0.08195460587739944, + 0.04445547237992287, + 0.03806265816092491, + 0.03272269666194916, + 0.00921470858156681, + 0.009166838601231575, + 0.006822453811764717, + 0.008658391423523426, + 0.00918212067335844, + 0.020768631249666214, + 0.006022640969604254, + 0.005583330523222685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0010414659045636654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monster truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck clad in electric blue with a white lightning bolt design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the truck captured in the peak moment of leaping over a towering sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the truck's massive chrome suspension and shock absorbers fully exposed in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot taken from a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021944984328001738 + ], + "opposite_logprob_tokens": [ + -6.127194404602051 + ], + "logprob_reward": -0.0021944984328001738, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021944984328001738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12643903493881226, + 0.050105735659599304, + 0.2028384506702423, + 0.09897321462631226, + 0.2842191755771637, + 0.029451385140419006, + 0.05703799054026604, + 0.02040995843708515, + 0.09079691022634506, + 0.03972815349698067 + ], + "credit_map": [ + 0.01044145692139864, + 0.025915196165442467, + 0.02642720751464367, + 0.010394472628831863, + 0.012872899882495403, + 0.02293187752366066, + 0.009391553699970245, + 0.008064358495175838, + 0.010169130750000477, + 0.0038586496375501156, + 0.008045237511396408, + 0.006968936882913113, + 0.0037812276277691126, + 0.009209667332470417, + 0.003545367857441306, + 0.0045275152660906315, + 0.06851977854967117, + 0.01440820749849081, + 0.015237870626151562, + 0.013861091807484627, + 0.013922760263085365, + 0.05157666653394699, + 0.011770629324018955, + 0.01354142464697361, + 0.0162756759673357, + 0.014602947048842907, + 0.015487061813473701, + 0.009492162615060806, + 0.013843923807144165, + 0.008981368504464626, + 0.009513171389698982, + 0.010776899755001068, + 0.04248461499810219, + 0.04910888522863388, + 0.06392466276884079, + 0.020601749420166016, + 0.03139718249440193, + 0.0234286580234766, + 0.023754294961690903, + 0.029519109055399895, + 0.004323613829910755, + 0.004250121768563986, + 0.003567526815459132, + 0.00432360777631402, + 0.0032168386969715357, + 0.004410257562994957, + 0.0034422914031893015, + 0.0019171254243701696, + 0.006305749528110027, + 0.006562994327396154, + 0.008215573616325855, + 0.005381444469094276, + 0.006523112766444683, + 0.006200036499649286, + 0.011972269974648952, + 0.005876803305000067, + 0.0019163574324920774, + 0.004208117723464966, + 0.0034726678859442472, + 0.0018137138104066253, + 0.0021886182948946953, + 0.0023094001226127148, + 0.0017499056411907077, + 0.00275117508135736, + 0.007306646555662155, + 0.013052623718976974, + 0.012512177228927612, + 0.008682731539011002, + 0.01870584860444069, + 0.011052698828279972, + 0.010646509937942028, + 0.008837665431201458, + 0.004843913950026035, + 0.0034671532921493053, + 0.0038631013594567776, + 0.00481784250587225, + 0.004171587992459536, + 0.01145907212048769, + 0.0038051737938076258, + 0.003300305688753724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0021944984328001738 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_4/index_tensor([4109, 4109])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_4/index_tensor([4109, 4109])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..65555229596e324b6119b2683b47e582ba8068ed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_4/index_tensor([4109, 4109])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled, medium close-up captures an elderly scholar, hunched over a dark lacquered wooden altar. His long, thinning white hair is meticulously tied into a traditional headdress, and his face, a map of deep wrinkles, reveals a profound concentration. He wears a deep indigo silk robe, its cuffs and collar embroidered with silver thread depicting clouds and cranes, the fabric softened by age. The scene is bathed in warm, low-key lighting, cast by an invisible oil lamp, which creates deep shadows and distinct textures. On the table, next to a half-unrolled scroll of yellowed rice paper and a jade brush holder, lies a heavy, dark slate inkstone. His rough, age-stained right hand rests reverently above it, a pool of freshly ground ink gleaming with a thick, dark sheen. A subtle, persistent tremor passes through his fingers, testifying to his advanced age and the profound weight of the moment, creating an atmosphere of quiet, tense concentration, an ancient ritual.", + "seed": 10560, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_4/index_4109_seed_10560_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12924837730824948, + "vlm_logprob": -0.12924837730824948, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.12924837730824948, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20226731896400452 + ], + "opposite_logprob_tokens": [ + -1.7022672891616821 + ], + "logprob_reward": -0.20226731896400452, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20226731896400452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46348533034324646, + 0.13045842945575714, + 0.07706204056739807, + 0.05704142153263092, + 0.04932654649019241, + 0.03811170533299446, + 0.04142535477876663, + 0.050458043813705444, + 0.04018060117959976, + 0.05245061591267586 + ], + "credit_map": [ + 0.07941824942827225, + 0.057835258543491364, + 0.04725031927227974, + 0.054297443479299545, + 0.06249889358878136, + 0.040792301297187805, + 0.05763079971075058, + 0.06376203149557114, + 0.02919185347855091, + 0.015714777633547783, + 0.01159806177020073, + 0.015584729611873627, + 0.01796557568013668, + 0.010805044323205948, + 0.011377941817045212, + 0.018220439553260803, + 0.01621258072555065, + 0.008470327593386173, + 0.008961397223174572, + 0.00784621573984623, + 0.01088888943195343, + 0.006037450395524502, + 0.008656143210828304, + 0.009989039972424507, + 0.010207174345850945, + 0.006263891700655222, + 0.0064694080501794815, + 0.0063611045479774475, + 0.008321043103933334, + 0.004979475401341915, + 0.007398687768727541, + 0.007040632423013449, + 0.009464958682656288, + 0.004639358259737492, + 0.006509362254291773, + 0.0052328710444271564, + 0.007471853867173195, + 0.004434563685208559, + 0.0054053934291005135, + 0.006168184336274862, + 0.006979383993893862, + 0.003836490912362933, + 0.004945625551044941, + 0.003496163757517934, + 0.006259912624955177, + 0.0035349091049283743, + 0.004590027965605259, + 0.004469189792871475, + 0.010807516053318977, + 0.003975133411586285, + 0.0030080312862992287, + 0.00370581797324121, + 0.005654475651681423, + 0.003432313445955515, + 0.0042554340325295925, + 0.00658663222566247, + 0.011148759163916111, + 0.003994135186076164, + 0.005840267986059189, + 0.006005693227052689, + 0.006191950757056475, + 0.004052037838846445, + 0.006929067429155111, + 0.006296132225543261, + 0.008500080555677414, + 0.003281776327639818, + 0.002779443748295307, + 0.004029508214443922, + 0.004545190837234259, + 0.0035906468983739614, + 0.007046225015074015, + 0.006407730747014284, + 0.01119176298379898, + 0.0034051649272441864, + 0.006103574763983488, + 0.006144283339381218, + 0.005774489603936672, + 0.003144452814012766, + 0.008245115168392658, + 0.008441770449280739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20226731896400452 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134857714176178 + ], + "opposite_logprob_tokens": [ + -1.3134857416152954 + ], + "logprob_reward": -0.3134857714176178, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134857714176178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4104495048522949, + 0.09920307993888855, + 0.06255432963371277, + 0.04200273007154465, + 0.04567983001470566, + 0.03659422695636749, + 0.0394611582159996, + 0.13777253031730652, + 0.05357243865728378, + 0.07271017879247665 + ], + "credit_map": [ + 0.06085092946887016, + 0.05538083240389824, + 0.0394330769777298, + 0.05441350117325783, + 0.05724427476525307, + 0.03611348941922188, + 0.051129452884197235, + 0.05588394030928612, + 0.01790602132678032, + 0.012544563040137291, + 0.00807924848049879, + 0.014243953861296177, + 0.013942576013505459, + 0.008604754693806171, + 0.010596491396427155, + 0.013285469263792038, + 0.013324877247214317, + 0.00671866862103343, + 0.006349284667521715, + 0.007738466840237379, + 0.008917665109038353, + 0.004405600018799305, + 0.006508811376988888, + 0.00859095435589552, + 0.005539898760616779, + 0.004706911742687225, + 0.003945401404052973, + 0.005954530090093613, + 0.007176520302891731, + 0.003976109437644482, + 0.005458733066916466, + 0.0052446238696575165, + 0.006685848347842693, + 0.004921869840472937, + 0.005374039988964796, + 0.00605216808617115, + 0.00779562396928668, + 0.004412155132740736, + 0.005257771350443363, + 0.00518035190179944, + 0.0065488615073263645, + 0.0036174599081277847, + 0.002867046045139432, + 0.004246865399181843, + 0.006356147583574057, + 0.0037066135555505753, + 0.00507243862375617, + 0.004178794100880623, + 0.009380421601235867, + 0.0032807898242026567, + 0.003400702029466629, + 0.004182284697890282, + 0.005909651517868042, + 0.0030283916275948286, + 0.0045446003787219524, + 0.005734317936003208, + 0.02919011376798153, + 0.013324916362762451, + 0.012246927246451378, + 0.021655000746250153, + 0.016357533633708954, + 0.00953021738678217, + 0.018956467509269714, + 0.0165113415569067, + 0.008449042215943336, + 0.005869265180081129, + 0.002946150489151478, + 0.005281022749841213, + 0.007154909428209066, + 0.006196856033056974, + 0.010081342421472073, + 0.007593848276883364, + 0.012868478894233704, + 0.006311837118119001, + 0.006317554507404566, + 0.010098004713654518, + 0.012569168582558632, + 0.007346306461840868, + 0.006997569929808378, + 0.010201256722211838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.3134857714176178 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3496946096420288, + 0.14908677339553833, + 0.07807571440935135, + 0.07653134316205978, + 0.09423758089542389, + 0.06622304022312164, + 0.049064815044403076, + 0.0583539754152298, + 0.034402478486299515, + 0.044329673051834106 + ], + "credit_map": [ + 0.055980026721954346, + 0.04695368558168411, + 0.03693776577711105, + 0.038783978670835495, + 0.050161588937044144, + 0.031201189383864403, + 0.042106743901968, + 0.047569628804922104, + 0.03815825656056404, + 0.01611587591469288, + 0.015865769237279892, + 0.01463908702135086, + 0.018872909247875214, + 0.010541442781686783, + 0.011623458936810493, + 0.02326996624469757, + 0.01749003492295742, + 0.008302894420921803, + 0.010720907710492611, + 0.007392445113509893, + 0.010286361910402775, + 0.006140010431408882, + 0.006995632778853178, + 0.010747427120804787, + 0.013628729619085789, + 0.008691254071891308, + 0.011221359483897686, + 0.007885812781751156, + 0.010149453766644001, + 0.007171786390244961, + 0.008021771907806396, + 0.009761176072061062, + 0.018298881128430367, + 0.010497627779841423, + 0.014142199419438839, + 0.009419223293662071, + 0.011310010217130184, + 0.00795145146548748, + 0.010591617785394192, + 0.012026570737361908, + 0.012811923399567604, + 0.007810212671756744, + 0.009564513340592384, + 0.006199462804943323, + 0.008368517272174358, + 0.005680676084011793, + 0.0073014372028410435, + 0.008486302569508553, + 0.009886267594993114, + 0.004854270257055759, + 0.003920187707990408, + 0.007076900918036699, + 0.006417059805244207, + 0.004387956112623215, + 0.005894490517675877, + 0.006627683062106371, + 0.013900702819228172, + 0.005097994580864906, + 0.008604534901678562, + 0.0055589573457837105, + 0.006101070903241634, + 0.005411032121628523, + 0.0061301058158278465, + 0.007549574598670006, + 0.008680181577801704, + 0.0027170313987880945, + 0.003222312778234482, + 0.0025288828182965517, + 0.0036815726198256016, + 0.0028319547418504953, + 0.0050631058402359486, + 0.005677434615790844, + 0.009091324172914028, + 0.0029172305949032307, + 0.00720919668674469, + 0.0036886956077069044, + 0.005387663375586271, + 0.003275687340646982, + 0.005731726530939341, + 0.007028150837868452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02058347500860691 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495136231184006 + ], + "opposite_logprob_tokens": [ + -2.9299514293670654 + ], + "logprob_reward": -0.05495136231184006, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495136231184006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3970828056335449, + 0.1315229833126068, + 0.06578397750854492, + 0.06233292445540428, + 0.07008087635040283, + 0.053700368851423264, + 0.05697745457291603, + 0.05568176135420799, + 0.05219350755214691, + 0.05464329943060875 + ], + "credit_map": [ + 0.06326659023761749, + 0.04771091043949127, + 0.039439260959625244, + 0.049243271350860596, + 0.05560779571533203, + 0.0418342649936676, + 0.04473932087421417, + 0.055241361260414124, + 0.03225855901837349, + 0.015784040093421936, + 0.011438467539846897, + 0.01474074274301529, + 0.017361218109726906, + 0.010881970636546612, + 0.009583801031112671, + 0.01947418786585331, + 0.014828707091510296, + 0.007120654918253422, + 0.00638175243511796, + 0.006322784814983606, + 0.010283762589097023, + 0.0065804338082671165, + 0.005000812001526356, + 0.009265070781111717, + 0.010942967608571053, + 0.0071979002095758915, + 0.00764909852296114, + 0.006773625500500202, + 0.008886117488145828, + 0.006926624570041895, + 0.006124215666204691, + 0.007832375355064869, + 0.013089342974126339, + 0.0076186987571418285, + 0.009658114053308964, + 0.007954301312565804, + 0.010794011875987053, + 0.006163827143609524, + 0.006432263646274805, + 0.00837031938135624, + 0.01040646992623806, + 0.006555358879268169, + 0.006601306609809399, + 0.005503456108272076, + 0.008266118355095387, + 0.00462373485788703, + 0.005057791247963905, + 0.006686131004244089, + 0.015055178664624691, + 0.004259404726326466, + 0.004375080112367868, + 0.005192815791815519, + 0.007825418375432491, + 0.005638611502945423, + 0.004567384719848633, + 0.010063561610877514, + 0.011362385004758835, + 0.004404309671372175, + 0.00613874988630414, + 0.00416029617190361, + 0.009319058619439602, + 0.006155654322355986, + 0.006874373182654381, + 0.007266937289386988, + 0.018439048901200294, + 0.0036843325942754745, + 0.003038703231140971, + 0.00424443231895566, + 0.004884508904069662, + 0.004167134407907724, + 0.003620762610808015, + 0.010114585049450397, + 0.015848353505134583, + 0.003386411117389798, + 0.0029121534898877144, + 0.0036692859139293432, + 0.005487761460244656, + 0.004927451256662607, + 0.007559732533991337, + 0.010852149687707424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.05495136231184006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054953958839178085 + ], + "opposite_logprob_tokens": [ + -2.9299540519714355 + ], + "logprob_reward": -0.054953958839178085, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054953958839178085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1926109939813614, + 0.0652916356921196, + 0.05518566071987152, + 0.06994328647851944, + 0.06879102438688278, + 0.0627823919057846, + 0.19055646657943726, + 0.22117696702480316, + 0.037669867277145386, + 0.035991717129945755 + ], + "credit_map": [ + 0.02784881740808487, + 0.028544077649712563, + 0.018392911180853844, + 0.021372638642787933, + 0.029019832611083984, + 0.01896769367158413, + 0.022182445973157883, + 0.02628258988261223, + 0.01321520283818245, + 0.00805513747036457, + 0.005742613226175308, + 0.0059887380339205265, + 0.008686683140695095, + 0.005374842789024115, + 0.0093971136957407, + 0.008831308223307133, + 0.008572332561016083, + 0.005103156901896, + 0.004162385128438473, + 0.007261915598064661, + 0.008011994883418083, + 0.006322820670902729, + 0.008513452485203743, + 0.007237602956593037, + 0.012495127506554127, + 0.006800204981118441, + 0.006091063376516104, + 0.005451846867799759, + 0.011915505863726139, + 0.006184423342347145, + 0.01120782271027565, + 0.009797291830182076, + 0.011799474246799946, + 0.0063532134518027306, + 0.0050066933035850525, + 0.00647996598854661, + 0.008825314231216908, + 0.008590794168412685, + 0.012796130031347275, + 0.008939439430832863, + 0.01134555134922266, + 0.006029876414686441, + 0.010202867910265923, + 0.00607616500928998, + 0.007647453807294369, + 0.005616921465843916, + 0.0065549397841095924, + 0.009308612905442715, + 0.032984409481287, + 0.014288237318396568, + 0.01574910804629326, + 0.024034742265939713, + 0.031419724225997925, + 0.023571914061903954, + 0.02316616289317608, + 0.02534218318760395, + 0.04310385510325432, + 0.024038733914494514, + 0.01589462161064148, + 0.024652713909745216, + 0.03375261649489403, + 0.023592012003064156, + 0.02794174663722515, + 0.0282006673514843, + 0.011098269373178482, + 0.0028656604699790478, + 0.002497548470273614, + 0.004146640654653311, + 0.003878578543663025, + 0.0026586614549160004, + 0.004129704087972641, + 0.006394803058356047, + 0.008569524623453617, + 0.0028136330656707287, + 0.00246823881752789, + 0.003423406509682536, + 0.004559085704386234, + 0.0032497078645974398, + 0.004343016073107719, + 0.0065651023760437965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.054953958839178085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, medium close-up captures an elderly scholar, hunched over a dark lacquered wooden altar. His long, thinning white hair is meticulously tied into a traditional headdress, and his face, a map of deep wrinkles, reveals a profound concentration. He wears a deep indigo silk robe, its cuffs and collar embroidered with silver thread depicting clouds and cranes, the fabric softened by age. The scene is bathed in warm, low-key lighting, cast by an invisible oil lamp, which creates deep shadows and distinct textures. On the table, next to a half-unrolled scroll of yellowed rice paper and a jade brush holder, lies a heavy, dark slate inkstone. His rough, age-stained right hand rests reverently above it, a pool of freshly ground ink gleaming with a thick, dark sheen. A subtle, persistent tremor passes through his fingers, testifying to his advanced age and the profound weight of the moment, creating an atmosphere of quiet, tense concentration, an ancient ritual.", + "seed": 10561, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_4/index_4109_seed_10561_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04906806275248528, + "vlm_logprob": -0.04906806275248528, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.04906806275248528, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03389137238264084 + ], + "opposite_logprob_tokens": [ + -3.533891439437866 + ], + "logprob_reward": -0.03389137238264084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03389137238264084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5142333507537842, + 0.13941265642642975, + 0.06554839015007019, + 0.04291776567697525, + 0.0441272146999836, + 0.036055464297533035, + 0.03965345770120621, + 0.033825065940618515, + 0.03612758219242096, + 0.04809906333684921 + ], + "credit_map": [ + 0.06713307648897171, + 0.04444122314453125, + 0.05084947124123573, + 0.0348101370036602, + 0.12976576387882233, + 0.06695575267076492, + 0.05803187936544418, + 0.06224607303738594, + 0.014900380745530128, + 0.01035976130515337, + 0.016999391838908195, + 0.010317785665392876, + 0.03412751853466034, + 0.0191520843654871, + 0.015591190196573734, + 0.01796453632414341, + 0.0066694226115942, + 0.007892646826803684, + 0.0069929989986121655, + 0.004800141789019108, + 0.014430341310799122, + 0.009724204428493977, + 0.0076719531789422035, + 0.007366673089563847, + 0.004446834325790405, + 0.003933701664209366, + 0.00466806348413229, + 0.003209786955267191, + 0.009087781421840191, + 0.006023243069648743, + 0.0052164276130497456, + 0.006331929471343756, + 0.004668273497372866, + 0.004670834168791771, + 0.0037247061263769865, + 0.0037826045881956816, + 0.009204556234180927, + 0.0056307194754481316, + 0.0060151610523462296, + 0.0064303576946258545, + 0.004221104085445404, + 0.0034572025761008263, + 0.003293849527835846, + 0.003244466846808791, + 0.005533291958272457, + 0.006066190078854561, + 0.0039944835007190704, + 0.006244880612939596, + 0.005862624850124121, + 0.004397187381982803, + 0.005398273468017578, + 0.003145614406093955, + 0.0058848438784480095, + 0.005769135896116495, + 0.003987346310168505, + 0.0052084303461015224, + 0.004723785910755396, + 0.004269252996891737, + 0.0038084283005446196, + 0.0032993422355502844, + 0.005591288674622774, + 0.003950244747102261, + 0.003994729369878769, + 0.004187993239611387, + 0.0033212441485375166, + 0.0038641232531517744, + 0.003837412456050515, + 0.0031669267918914557, + 0.005011789966374636, + 0.004725910257548094, + 0.006367928814142942, + 0.0058322446420788765, + 0.006726482417434454, + 0.00520725455135107, + 0.004271847661584616, + 0.004054235760122538, + 0.00714951241388917, + 0.0068833609111607075, + 0.006435538176447153, + 0.007370833307504654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03389137238264084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300091445446014 + ], + "opposite_logprob_tokens": [ + -2.2380008697509766 + ], + "logprob_reward": -0.11300091445446014, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300091445446014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42042577266693115, + 0.1498517096042633, + 0.08499056845903397, + 0.048704613000154495, + 0.053012341260910034, + 0.028048263862729073, + 0.04073735326528549, + 0.050979454070329666, + 0.0464295968413353, + 0.07682032138109207 + ], + "credit_map": [ + 0.04802682623267174, + 0.046367667615413666, + 0.03434119373559952, + 0.03451156988739967, + 0.10587863624095917, + 0.057254523038864136, + 0.045412588864564896, + 0.04863277077674866, + 0.020667750388383865, + 0.014941569417715073, + 0.015320513397455215, + 0.01138768158853054, + 0.026767682284116745, + 0.02018033154308796, + 0.018053315579891205, + 0.022532869130373, + 0.010898330248892307, + 0.007059301715344191, + 0.0069227986969053745, + 0.005536943674087524, + 0.012244746088981628, + 0.01252571027725935, + 0.012624157592654228, + 0.01717858389019966, + 0.0053148106671869755, + 0.0046837893314659595, + 0.003831620095297694, + 0.0028369552455842495, + 0.010551372542977333, + 0.006906567141413689, + 0.006749021355062723, + 0.007830479182302952, + 0.007153282407671213, + 0.005235528573393822, + 0.005439997185021639, + 0.0051450179889798164, + 0.008551777340471745, + 0.006497575901448727, + 0.007548321504145861, + 0.007440841756761074, + 0.002858218038454652, + 0.0025267296005040407, + 0.002702520927414298, + 0.003673524595797062, + 0.004348418675363064, + 0.003719120752066374, + 0.0040002004243433475, + 0.0042195310816168785, + 0.004718491807579994, + 0.003510391805320978, + 0.006175221875309944, + 0.004452756606042385, + 0.0059576453641057014, + 0.0054550012573599815, + 0.004303857684135437, + 0.0061639826744794846, + 0.0047510177828371525, + 0.0033986028283834457, + 0.0060721952468156815, + 0.0042414311319589615, + 0.007966701872646809, + 0.007483564782887697, + 0.006941584404557943, + 0.01012435369193554, + 0.0038926939014345407, + 0.0033117092680186033, + 0.004712861962616444, + 0.00403126934543252, + 0.008570609614253044, + 0.00647896621376276, + 0.006543637253344059, + 0.008887846022844315, + 0.006142147351056337, + 0.006815606262534857, + 0.005721313413232565, + 0.00467000063508749, + 0.014214396476745605, + 0.011100475676357746, + 0.013448270969092846, + 0.014708111062645912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11300091445446014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786956012248993 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.009786956012248993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786956012248993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2929385006427765, + 0.11508303880691528, + 0.07803209871053696, + 0.05532606691122055, + 0.07480709254741669, + 0.1052577868103981, + 0.0758528932929039, + 0.0650184154510498, + 0.04770124703645706, + 0.08998285233974457 + ], + "credit_map": [ + 0.04784395545721054, + 0.028492793440818787, + 0.043337639421224594, + 0.021854519844055176, + 0.04406130686402321, + 0.036428213119506836, + 0.03567729517817497, + 0.035242799669504166, + 0.018577413633465767, + 0.009943132288753986, + 0.014996829442679882, + 0.009426712989807129, + 0.01909821480512619, + 0.01394747942686081, + 0.014384706504642963, + 0.0147085627540946, + 0.013508345000445843, + 0.006837531924247742, + 0.008612624369561672, + 0.00604578061029315, + 0.015789760276675224, + 0.009416146203875542, + 0.009249973110854626, + 0.008571935817599297, + 0.010586404241621494, + 0.004936753306537867, + 0.0052635883912444115, + 0.003982452675700188, + 0.009956805035471916, + 0.006573654245585203, + 0.006344016175717115, + 0.007682392373681068, + 0.013643849641084671, + 0.007550724316388369, + 0.010073022916913033, + 0.006098640151321888, + 0.011933661997318268, + 0.009863130748271942, + 0.006740002892911434, + 0.008904057554900646, + 0.01429043523967266, + 0.02005135640501976, + 0.011109811253845692, + 0.007526549976319075, + 0.016544152051210403, + 0.013816207647323608, + 0.010688980109989643, + 0.01123029738664627, + 0.007878893055021763, + 0.008348933421075344, + 0.006903188303112984, + 0.005344771780073643, + 0.013508455827832222, + 0.009389654733240604, + 0.011417943052947521, + 0.013061054050922394, + 0.007697984576225281, + 0.00887461006641388, + 0.01040941383689642, + 0.006166188046336174, + 0.008859213441610336, + 0.008604336529970169, + 0.006646215450018644, + 0.007760455831885338, + 0.006418296601623297, + 0.010115017183125019, + 0.004516055341809988, + 0.0034902391489595175, + 0.007023184094578028, + 0.00576540594920516, + 0.004790806211531162, + 0.005582240875810385, + 0.011111381463706493, + 0.007999119348824024, + 0.010243100114166737, + 0.007028303109109402, + 0.01850421540439129, + 0.011880198493599892, + 0.009364623576402664, + 0.013851906172931194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.009786956012248993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033702824264764786 + ], + "opposite_logprob_tokens": [ + -3.408702850341797 + ], + "logprob_reward": -0.033702824264764786, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033702824264764786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2822539508342743, + 0.10111051052808762, + 0.05877502262592316, + 0.07525799423456192, + 0.1093326210975647, + 0.10186481475830078, + 0.06584293395280838, + 0.053520265966653824, + 0.06617514044046402, + 0.08586673438549042 + ], + "credit_map": [ + 0.032422810792922974, + 0.02140216715633869, + 0.04487715661525726, + 0.028831813484430313, + 0.04972434416413307, + 0.036415427923202515, + 0.035280536860227585, + 0.03329967334866524, + 0.012877020984888077, + 0.008736254647374153, + 0.01679033786058426, + 0.010396708734333515, + 0.014309659600257874, + 0.012373608537018299, + 0.011636481620371342, + 0.013990435749292374, + 0.006345772184431553, + 0.00555816013365984, + 0.010592855513095856, + 0.005686848424375057, + 0.00948780681937933, + 0.0062753125093877316, + 0.00877462886273861, + 0.006053636781871319, + 0.013182494789361954, + 0.006459227297455072, + 0.013049059547483921, + 0.007879968732595444, + 0.007361728698015213, + 0.00936570018529892, + 0.008253591135144234, + 0.009706228040158749, + 0.018997302278876305, + 0.012364579364657402, + 0.013493158854544163, + 0.00828755646944046, + 0.0172171201556921, + 0.014463226310908794, + 0.011865874752402306, + 0.012643792666494846, + 0.013112249784171581, + 0.016005931422114372, + 0.010731893591582775, + 0.008139829151332378, + 0.012781471014022827, + 0.013287942856550217, + 0.01488556619733572, + 0.012919934466481209, + 0.005335794296115637, + 0.004168962128460407, + 0.007822923362255096, + 0.004477239679545164, + 0.013303685002028942, + 0.008297497406601906, + 0.011915933340787888, + 0.010520903393626213, + 0.004286392591893673, + 0.005537279415875673, + 0.008137136697769165, + 0.004771188832819462, + 0.008187484927475452, + 0.007618934381753206, + 0.006479592528194189, + 0.008502257987856865, + 0.006948207505047321, + 0.009549603797495365, + 0.007373251020908356, + 0.005637688562273979, + 0.00893375463783741, + 0.009191427379846573, + 0.008841568604111671, + 0.009699641726911068, + 0.009124217554926872, + 0.008206316269934177, + 0.008092690259218216, + 0.006074007134884596, + 0.01865035854279995, + 0.011831448413431644, + 0.00976106058806181, + 0.014126632362604141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.033702824264764786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054958246648311615 + ], + "opposite_logprob_tokens": [ + -2.9299583435058594 + ], + "logprob_reward": -0.054958246648311615, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054958246648311615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2507994472980499, + 0.08358805626630783, + 0.08719395101070404, + 0.050250034779310226, + 0.05026964098215103, + 0.049040500074625015, + 0.0932331532239914, + 0.14808125793933868, + 0.14832474291324615, + 0.03921918943524361 + ], + "credit_map": [ + 0.027634890750050545, + 0.02805195562541485, + 0.024530084803700447, + 0.018336784094572067, + 0.04752378910779953, + 0.037159375846385956, + 0.02756962925195694, + 0.03999295085668564, + 0.011797375045716763, + 0.008868135511875153, + 0.010229205712676048, + 0.006938327103853226, + 0.01282484084367752, + 0.011441266164183617, + 0.0090365344658494, + 0.012452378869056702, + 0.014583676122128963, + 0.009319563396275043, + 0.006449721287935972, + 0.005929809994995594, + 0.01602768525481224, + 0.010402441024780273, + 0.013071171008050442, + 0.011409898288547993, + 0.007638353854417801, + 0.0044333795085549355, + 0.005172836594283581, + 0.004465029574930668, + 0.007508951239287853, + 0.00700299721211195, + 0.0068868049420416355, + 0.00714168231934309, + 0.004393426235765219, + 0.00594136118888855, + 0.006629330106079578, + 0.0059912120923399925, + 0.0082187969237566, + 0.0064673288725316525, + 0.006005904637277126, + 0.0066222865134477615, + 0.005745979491621256, + 0.00610304856672883, + 0.004703049547970295, + 0.006135045550763607, + 0.009084579534828663, + 0.005838674493134022, + 0.005770002957433462, + 0.005660124588757753, + 0.01095670647919178, + 0.01671765372157097, + 0.0097334124147892, + 0.007934189401566982, + 0.018335072323679924, + 0.011681385338306427, + 0.008777616545557976, + 0.009097126312553883, + 0.013118084520101547, + 0.008771446533501148, + 0.012962699867784977, + 0.012445244938135147, + 0.029380910098552704, + 0.021012455224990845, + 0.024096159264445305, + 0.02629426307976246, + 0.011240205727517605, + 0.008242922835052013, + 0.01519702933728695, + 0.01605050079524517, + 0.02935628592967987, + 0.02081696316599846, + 0.022561365738511086, + 0.024859489873051643, + 0.003513709641993046, + 0.0034027481451630592, + 0.004041665233671665, + 0.0029058479703962803, + 0.006021222099661827, + 0.006357535719871521, + 0.004941457416862249, + 0.008035006001591682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.054958246648311615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, medium close-up captures an elderly scholar, hunched over a dark lacquered wooden altar. His long, thinning white hair is meticulously tied into a traditional headdress, and his face, a map of deep wrinkles, reveals a profound concentration. He wears a deep indigo silk robe, its cuffs and collar embroidered with silver thread depicting clouds and cranes, the fabric softened by age. The scene is bathed in warm, low-key lighting, cast by an invisible oil lamp, which creates deep shadows and distinct textures. On the table, next to a half-unrolled scroll of yellowed rice paper and a jade brush holder, lies a heavy, dark slate inkstone. His rough, age-stained right hand rests reverently above it, a pool of freshly ground ink gleaming with a thick, dark sheen. A subtle, persistent tremor passes through his fingers, testifying to his advanced age and the profound weight of the moment, creating an atmosphere of quiet, tense concentration, an ancient ritual.", + "seed": 10562, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_4/index_4109_seed_10562_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0898686183616519, + "vlm_logprob": -0.0898686183616519, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.0898686183616519, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20283225178718567 + ], + "opposite_logprob_tokens": [ + -1.7028322219848633 + ], + "logprob_reward": -0.20283225178718567, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20283225178718567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3341687023639679, + 0.05743969976902008, + 0.07867205888032913, + 0.03621064871549606, + 0.06670921295881271, + 0.1274784356355667, + 0.024270402267575264, + 0.026424357667565346, + 0.019324226304888725, + 0.22930221259593964 + ], + "credit_map": [ + 0.023364488035440445, + 0.023545125499367714, + 0.058733511716127396, + 0.037170350551605225, + 0.06089053303003311, + 0.04326092079281807, + 0.0556630976498127, + 0.03154071420431137, + 0.006306712981313467, + 0.005284385289996862, + 0.0070611401461064816, + 0.006760651245713234, + 0.006502419710159302, + 0.01024339348077774, + 0.010845029726624489, + 0.004435972776263952, + 0.007212704047560692, + 0.006964865606278181, + 0.013020633719861507, + 0.00966212060302496, + 0.011154166422784328, + 0.010133842006325722, + 0.012403452768921852, + 0.008120279759168625, + 0.0038224903400987387, + 0.003258123528212309, + 0.003879852592945099, + 0.003959628287702799, + 0.004595467820763588, + 0.004125689156353474, + 0.007803595159202814, + 0.004765805788338184, + 0.005750128533691168, + 0.007709840778261423, + 0.0064504919573664665, + 0.007231449242681265, + 0.011327707208693027, + 0.00792607106268406, + 0.013434461317956448, + 0.006879073102027178, + 0.016533954069018364, + 0.03737994283437729, + 0.007238514721393585, + 0.011850487440824509, + 0.012620343826711178, + 0.013393089175224304, + 0.01958479732275009, + 0.00887730810791254, + 0.0027011174242943525, + 0.002939750673249364, + 0.0025014476850628853, + 0.002094206167384982, + 0.003820199752226472, + 0.0036793206818401814, + 0.004537321627140045, + 0.00199704198166728, + 0.003243168815970421, + 0.003275016089901328, + 0.0030455535743385553, + 0.003062792122364044, + 0.0037002686876803637, + 0.004721478093415499, + 0.003657514462247491, + 0.0017185690812766552, + 0.002126800362020731, + 0.002211703220382333, + 0.0019995165057480335, + 0.0016536788316443563, + 0.002944710897281766, + 0.0034279816318303347, + 0.0025158454664051533, + 0.0024439916014671326, + 0.01781228370964527, + 0.019521810114383698, + 0.03535700961947441, + 0.018236910924315453, + 0.02429839037358761, + 0.03029017522931099, + 0.012588313780725002, + 0.07119734585285187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.20283225178718567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604897528886795 + ], + "opposite_logprob_tokens": [ + -1.910489797592163 + ], + "logprob_reward": -0.1604897528886795, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604897528886795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13727277517318726, + 0.03494725003838539, + 0.049073584377765656, + 0.025856753811240196, + 0.05299708619713783, + 0.05494704842567444, + 0.01120622269809246, + 0.023446550592780113, + 0.01481503527611494, + 0.5954376459121704 + ], + "credit_map": [ + 0.009125163778662682, + 0.010227840393781662, + 0.023299846798181534, + 0.014130267314612865, + 0.0252398569136858, + 0.018326006829738617, + 0.023242473602294922, + 0.01368133258074522, + 0.0037479896564036608, + 0.0035506433341652155, + 0.004515594802796841, + 0.0044382717460393906, + 0.0034396115224808455, + 0.005709422752261162, + 0.006485543679445982, + 0.003060174873098731, + 0.003918749280273914, + 0.003980230074375868, + 0.008707129396498203, + 0.005272326059639454, + 0.007750402204692364, + 0.006349431816488504, + 0.007580210920423269, + 0.005515105556696653, + 0.0025794401299208403, + 0.00199846550822258, + 0.0035542072728276253, + 0.003850717796012759, + 0.002522327471524477, + 0.0031282352283596992, + 0.004073029849678278, + 0.004150330554693937, + 0.004875741433352232, + 0.0056051635183393955, + 0.005021016113460064, + 0.007615541573613882, + 0.006196083500981331, + 0.009519355371594429, + 0.010111748240888119, + 0.004052433650940657, + 0.0059276060201227665, + 0.008720843121409416, + 0.0051428452134132385, + 0.0053513855673372746, + 0.007659852504730225, + 0.007408778183162212, + 0.009715046733617783, + 0.0050206915475428104, + 0.0015284792752936482, + 0.0014392236480489373, + 0.00111471361014992, + 0.001522775273770094, + 0.0012442524312064052, + 0.0014962211716920137, + 0.0017425890546292067, + 0.001117969281040132, + 0.0025550411082804203, + 0.0029233123641461134, + 0.0021422640420496464, + 0.003127405885607004, + 0.0034987793769687414, + 0.003975425381213427, + 0.0030443572904914618, + 0.002179966773837805, + 0.002092311391606927, + 0.0017899005906656384, + 0.0012206585379317403, + 0.0014399125939235091, + 0.0022743188310414553, + 0.002675172407180071, + 0.0019964491948485374, + 0.0013263116125017405, + 0.04933159798383713, + 0.03559571132063866, + 0.13780537247657776, + 0.0297541506588459, + 0.060970667749643326, + 0.04151338338851929, + 0.027304554358124733, + 0.21316221356391907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.1604897528886795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206229709088802 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206229709088802, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206229709088802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3013906478881836, + 0.030468443408608437, + 0.07257824391126633, + 0.08331351727247238, + 0.07475722581148148, + 0.08294033259153366, + 0.02074379287660122, + 0.020112857222557068, + 0.013877524062991142, + 0.29981735348701477 + ], + "credit_map": [ + 0.02129950188100338, + 0.024350790306925774, + 0.05130452290177345, + 0.03450492024421692, + 0.056591782718896866, + 0.037751030176877975, + 0.044864144176244736, + 0.030723977833986282, + 0.0025853822007775307, + 0.0024275316391140223, + 0.003957958891987801, + 0.003736217273399234, + 0.003184711327776313, + 0.005383959971368313, + 0.00670810928568244, + 0.002484573982656002, + 0.008429670706391335, + 0.007913222536444664, + 0.010589507408440113, + 0.01480947807431221, + 0.0074325748719275, + 0.009429549798369408, + 0.009478638879954815, + 0.004495612345635891, + 0.007942935451865196, + 0.008461365476250648, + 0.01060298178344965, + 0.014095919206738472, + 0.00881008617579937, + 0.008718278259038925, + 0.013929350301623344, + 0.010752602480351925, + 0.008143961429595947, + 0.0095875458791852, + 0.012947090901434422, + 0.011050683446228504, + 0.00866703875362873, + 0.006952999159693718, + 0.012782322242856026, + 0.0046255881898105145, + 0.008828038349747658, + 0.013072020374238491, + 0.00834293570369482, + 0.009328179061412811, + 0.010506152175366879, + 0.00993107445538044, + 0.01553237996995449, + 0.007399559486657381, + 0.0030823745764791965, + 0.002634866163134575, + 0.0028043382335454226, + 0.0020499643869698048, + 0.003541430691257119, + 0.0025257407687604427, + 0.002727895276620984, + 0.0013771834783256054, + 0.0018705471884459257, + 0.002712824149057269, + 0.002297203056514263, + 0.0025850243400782347, + 0.0029912979807704687, + 0.0036346290726214647, + 0.0025362966116517782, + 0.001485036569647491, + 0.0015944242477416992, + 0.0016104085370898247, + 0.0014629911165684462, + 0.0015011935029178858, + 0.0017895003547891974, + 0.0021193281281739473, + 0.0017482370603829622, + 0.0020514437928795815, + 0.024505218490958214, + 0.019147837534546852, + 0.06914511322975159, + 0.014521190896630287, + 0.0306644719094038, + 0.0186894740909338, + 0.01353509072214365, + 0.1096089631319046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.014206229709088802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371262177824974 + ], + "opposite_logprob_tokens": [ + -3.40871262550354 + ], + "logprob_reward": -0.03371262177824974, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371262177824974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29325154423713684, + 0.05822465196251869, + 0.1619597226381302, + 0.10434044152498245, + 0.08317362517118454, + 0.08320160955190659, + 0.030383864417672157, + 0.02167684957385063, + 0.02111111395061016, + 0.14267653226852417 + ], + "credit_map": [ + 0.017636870965361595, + 0.021160177886486053, + 0.04688907414674759, + 0.03224354237318039, + 0.05884763225913048, + 0.03953176364302635, + 0.047046490013599396, + 0.029896005988121033, + 0.011731848120689392, + 0.00742894783616066, + 0.007456125225871801, + 0.006658627185970545, + 0.004237052984535694, + 0.009260609745979309, + 0.00657598115503788, + 0.004875458776950836, + 0.02033107914030552, + 0.01500158291310072, + 0.024938369169831276, + 0.02116018906235695, + 0.010874946601688862, + 0.02286459691822529, + 0.021519741043448448, + 0.02526921033859253, + 0.009252491407096386, + 0.015468908473849297, + 0.022575903683900833, + 0.011961744166910648, + 0.010019089095294476, + 0.010016854852437973, + 0.014130993746221066, + 0.010914462618529797, + 0.008917944505810738, + 0.011138839647173882, + 0.013422716408967972, + 0.010464638471603394, + 0.00967296864837408, + 0.00799473375082016, + 0.014363233000040054, + 0.007198554463684559, + 0.018038857728242874, + 0.028324678540229797, + 0.005243720952421427, + 0.00471388129517436, + 0.0073345680721104145, + 0.0073963673785328865, + 0.00851704552769661, + 0.0036324828397482634, + 0.002648675814270973, + 0.0040755015797913074, + 0.002293455181643367, + 0.0026976505760103464, + 0.004776724614202976, + 0.003946660552173853, + 0.007441650610417128, + 0.002503545954823494, + 0.0042412858456373215, + 0.0030445854645222425, + 0.0023457848001271486, + 0.0020889630541205406, + 0.002747986698523164, + 0.002094700001180172, + 0.002462449250742793, + 0.002651095623150468, + 0.0028561537619680166, + 0.0025598264764994383, + 0.002143053337931633, + 0.0029347632080316544, + 0.00261453608982265, + 0.002283995971083641, + 0.0026240453589707613, + 0.0030947402119636536, + 0.012716356664896011, + 0.009748130105435848, + 0.031451307237148285, + 0.011360248550772667, + 0.014874221757054329, + 0.015664326027035713, + 0.006275963503867388, + 0.040585972368717194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.03371262177824974 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810223564505577 + ], + "opposite_logprob_tokens": [ + -3.288102149963379 + ], + "logprob_reward": -0.03810223564505577, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810223564505577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1990252137184143, + 0.022455770522356033, + 0.026907410472631454, + 0.04763180390000343, + 0.05938737094402313, + 0.1044343113899231, + 0.025559088215231895, + 0.023072784766554832, + 0.021727710962295532, + 0.4697985351085663 + ], + "credit_map": [ + 0.017259158194065094, + 0.012663627974689007, + 0.043019939213991165, + 0.021575018763542175, + 0.025221114978194237, + 0.031108452007174492, + 0.02998475171625614, + 0.01819315180182457, + 0.0028332220390439034, + 0.002700903220102191, + 0.0037930707912892103, + 0.0024339824449270964, + 0.002304484834894538, + 0.00375189958140254, + 0.0032309533562511206, + 0.0014072535559535027, + 0.0033972819801419973, + 0.003186147892847657, + 0.004610328935086727, + 0.004121201112866402, + 0.002943007042631507, + 0.002573419827967882, + 0.002695859409868717, + 0.003380164969712496, + 0.004757799208164215, + 0.004621078260242939, + 0.007892067544162273, + 0.009879374876618385, + 0.005099148489534855, + 0.003986510448157787, + 0.006922207772731781, + 0.00447361683472991, + 0.005844171158969402, + 0.005807064939290285, + 0.007707925047725439, + 0.013213823549449444, + 0.004418362397700548, + 0.011204765178263187, + 0.007153143174946308, + 0.004038112238049507, + 0.018741339445114136, + 0.03511673957109451, + 0.005574772134423256, + 0.007349603343755007, + 0.008676668629050255, + 0.010205532424151897, + 0.012879961170256138, + 0.005889691412448883, + 0.004457215778529644, + 0.0032591265626251698, + 0.0032636397518217564, + 0.0030078899580985308, + 0.003278491087257862, + 0.003110746620222926, + 0.0028061969205737114, + 0.0023757796734571457, + 0.004930265713483095, + 0.002638923702761531, + 0.0031196745112538338, + 0.0022249834146350622, + 0.002689110115170479, + 0.0021532895043492317, + 0.0028132167644798756, + 0.0025033208075910807, + 0.003494046162813902, + 0.002292305463925004, + 0.0036124177277088165, + 0.002396806376054883, + 0.0028498168103396893, + 0.00276160822249949, + 0.0023704252671450377, + 0.0019502859795466065, + 0.038110483437776566, + 0.02831825613975525, + 0.10732641071081161, + 0.02419503964483738, + 0.05012323334813118, + 0.02925798110663891, + 0.020077330991625786, + 0.17238979041576385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.03810223564505577 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, medium close-up captures an elderly scholar, hunched over a dark lacquered wooden altar. His long, thinning white hair is meticulously tied into a traditional headdress, and his face, a map of deep wrinkles, reveals a profound concentration. He wears a deep indigo silk robe, its cuffs and collar embroidered with silver thread depicting clouds and cranes, the fabric softened by age. The scene is bathed in warm, low-key lighting, cast by an invisible oil lamp, which creates deep shadows and distinct textures. On the table, next to a half-unrolled scroll of yellowed rice paper and a jade brush holder, lies a heavy, dark slate inkstone. His rough, age-stained right hand rests reverently above it, a pool of freshly ground ink gleaming with a thick, dark sheen. A subtle, persistent tremor passes through his fingers, testifying to his advanced age and the profound weight of the moment, creating an atmosphere of quiet, tense concentration, an ancient ritual.", + "seed": 10563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_4/index_4109_seed_10563_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.047021484375, + "vlm_logprob": -0.047021484375, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.047021484375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045244116336107254 + ], + "opposite_logprob_tokens": [ + -3.1702442169189453 + ], + "logprob_reward": -0.045244116336107254, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045244116336107254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4154987633228302, + 0.1292770355939865, + 0.04865065589547157, + 0.06626763194799423, + 0.07796072214841843, + 0.04611906781792641, + 0.03910215571522713, + 0.07795233279466629, + 0.0437345914542675, + 0.055437058210372925 + ], + "credit_map": [ + 0.05997904762625694, + 0.043190352618694305, + 0.08752377331256866, + 0.07556983083486557, + 0.024423005059361458, + 0.04869359731674194, + 0.053680915385484695, + 0.02243821881711483, + 0.017702843993902206, + 0.01672946847975254, + 0.017784785479307175, + 0.016484349966049194, + 0.016236916184425354, + 0.014305687509477139, + 0.020620660856366158, + 0.009412317536771297, + 0.00677137216553092, + 0.00579310255125165, + 0.00661231717094779, + 0.008013763464987278, + 0.004724398721009493, + 0.00537088280543685, + 0.007658211048692465, + 0.003706609131768346, + 0.009417859837412834, + 0.011099044233560562, + 0.007328683044761419, + 0.00882483460009098, + 0.010476626455783844, + 0.005584103520959616, + 0.007215206976979971, + 0.006321280263364315, + 0.010381015948951244, + 0.010181065648794174, + 0.009153527207672596, + 0.017178496345877647, + 0.006798923015594482, + 0.007521288003772497, + 0.01160258986055851, + 0.00514382217079401, + 0.006174663547426462, + 0.005471406038850546, + 0.005523144267499447, + 0.011608936823904514, + 0.004419517703354359, + 0.0036816727370023727, + 0.005466167815029621, + 0.003773559583351016, + 0.0052672806195914745, + 0.00684231985360384, + 0.0049619534984230995, + 0.00632860790938139, + 0.004823450930416584, + 0.0041033425368368626, + 0.004264111164957285, + 0.002511086408048868, + 0.012051538564264774, + 0.00969241838902235, + 0.014894118532538414, + 0.009930755011737347, + 0.006564750336110592, + 0.007393084466457367, + 0.011061304248869419, + 0.006364369764924049, + 0.008305993862450123, + 0.005630758125334978, + 0.005826227832585573, + 0.007472137454897165, + 0.005236485973000526, + 0.0038528540171682835, + 0.004468471277505159, + 0.0029416645411401987, + 0.00848324317485094, + 0.007642982993274927, + 0.007893557660281658, + 0.007217905018478632, + 0.008048773743212223, + 0.00481528602540493, + 0.007450201082974672, + 0.0038851092103868723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.045244116336107254 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054974041879177094 + ], + "opposite_logprob_tokens": [ + -2.92997407913208 + ], + "logprob_reward": -0.054974041879177094, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054974041879177094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32975006103515625, + 0.12934958934783936, + 0.07694772630929947, + 0.10904552042484283, + 0.06391233950853348, + 0.05824531987309456, + 0.0607353039085865, + 0.10082642734050751, + 0.03346923738718033, + 0.03771846741437912 + ], + "credit_map": [ + 0.0546262152493, + 0.03218001499772072, + 0.08219626545906067, + 0.04228190332651138, + 0.02638748288154602, + 0.04077571630477905, + 0.03204706311225891, + 0.019255399703979492, + 0.021703554317355156, + 0.01640373095870018, + 0.026837747544050217, + 0.012153539806604385, + 0.01372762955725193, + 0.015069343149662018, + 0.013191121630370617, + 0.010262925177812576, + 0.011749917641282082, + 0.01095443032681942, + 0.013502808287739754, + 0.008374915458261967, + 0.0069455043412745, + 0.009053788147866726, + 0.009303496219217777, + 0.007062864024192095, + 0.015635285526514053, + 0.0194070003926754, + 0.013738141395151615, + 0.010739926248788834, + 0.018770771101117134, + 0.008924653753638268, + 0.01193059142678976, + 0.009899148717522621, + 0.008079111576080322, + 0.008765537291765213, + 0.007703547365963459, + 0.01402437686920166, + 0.0062391916289925575, + 0.006206074263900518, + 0.008736450225114822, + 0.004158051684498787, + 0.007610091473907232, + 0.010454936884343624, + 0.00838283821940422, + 0.010093004442751408, + 0.007048045750707388, + 0.004494414199143648, + 0.005661304108798504, + 0.004500680137425661, + 0.009777846746146679, + 0.010619310662150383, + 0.009012567810714245, + 0.005464356858283281, + 0.00989481434226036, + 0.0047308881767094135, + 0.007179706357419491, + 0.004055812489241362, + 0.01460598036646843, + 0.01935119926929474, + 0.013611724600195885, + 0.009108568541705608, + 0.018641743808984756, + 0.00761360302567482, + 0.01066435594111681, + 0.007229257375001907, + 0.005832298193126917, + 0.003961156588047743, + 0.00528574176132679, + 0.004270369652658701, + 0.004147641826421022, + 0.0029623566661030054, + 0.0037401902955025434, + 0.003269480774179101, + 0.007371825631707907, + 0.0049159936606884, + 0.007042482495307922, + 0.003687209915369749, + 0.004217103589326143, + 0.0037535410374403, + 0.003799735801294446, + 0.0029305769130587578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.054974041879177094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606954075396061 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.01606954075396061, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606954075396061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.394524484872818, + 0.194094717502594, + 0.08625923097133636, + 0.07657168060541153, + 0.04018670693039894, + 0.040659770369529724, + 0.027376454323530197, + 0.04374014586210251, + 0.0593416765332222, + 0.03724510595202446 + ], + "credit_map": [ + 0.06487064808607101, + 0.037434887140989304, + 0.07904932647943497, + 0.07093264907598495, + 0.028608638793230057, + 0.04871256649494171, + 0.04231932386755943, + 0.022596463561058044, + 0.031201085075736046, + 0.020450007170438766, + 0.02851918712258339, + 0.04779493063688278, + 0.016154993325471878, + 0.02060645818710327, + 0.01815052330493927, + 0.011217526160180569, + 0.013304187916219234, + 0.00945335067808628, + 0.011849483475089073, + 0.014122006483376026, + 0.011577002704143524, + 0.008976606652140617, + 0.011365390382707119, + 0.00561120780184865, + 0.010375723242759705, + 0.009151986800134182, + 0.009106162004172802, + 0.01955641806125641, + 0.008820596151053905, + 0.006935797166079283, + 0.0075130220502614975, + 0.005111966747790575, + 0.007055336143821478, + 0.004675934556871653, + 0.008549937047064304, + 0.0057471091859042645, + 0.003519563004374504, + 0.003909612540155649, + 0.004741940647363663, + 0.001987273572012782, + 0.005969610996544361, + 0.003928462974727154, + 0.005955710541456938, + 0.012440294958651066, + 0.0028287693858146667, + 0.003896962618455291, + 0.0037390808574855328, + 0.0019008792005479336, + 0.004596714861690998, + 0.002936475444585085, + 0.004342155065387487, + 0.004301793873310089, + 0.0033467828761786222, + 0.0027174456045031548, + 0.0031466286163777113, + 0.0019884565845131874, + 0.005725243128836155, + 0.006392900366336107, + 0.007194309961050749, + 0.007331922184675932, + 0.005809822119772434, + 0.0037648524157702923, + 0.00468766363337636, + 0.002833431353792548, + 0.009939869865775108, + 0.006988657638430595, + 0.0083610899746418, + 0.011641192249953747, + 0.0065731480717659, + 0.005371002946048975, + 0.006490809842944145, + 0.0039759064093232155, + 0.007231248542666435, + 0.003459895495325327, + 0.006027895957231522, + 0.006212197244167328, + 0.004058778751641512, + 0.003395454492419958, + 0.004308021627366543, + 0.0025516110472381115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01606954075396061 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795313253998756 + ], + "opposite_logprob_tokens": [ + -3.5297954082489014 + ], + "logprob_reward": -0.029795313253998756, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795313253998756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3808574676513672, + 0.14473626017570496, + 0.09363406896591187, + 0.053870636969804764, + 0.07366561144590378, + 0.04551161080598831, + 0.05278433486819267, + 0.06035241112112999, + 0.055208757519721985, + 0.03937888517975807 + ], + "credit_map": [ + 0.05060715600848198, + 0.05850394442677498, + 0.07075820118188858, + 0.07196668535470963, + 0.031236175447702408, + 0.03405313193798065, + 0.04299851506948471, + 0.020733648911118507, + 0.018808642402291298, + 0.017873944714665413, + 0.01873692311346531, + 0.018652047961950302, + 0.022320225834846497, + 0.014816961251199245, + 0.022334342822432518, + 0.011193159967660904, + 0.012507865205407143, + 0.009628874249756336, + 0.013299895450472832, + 0.02490154467523098, + 0.008173895068466663, + 0.008892960846424103, + 0.01096285879611969, + 0.005266167223453522, + 0.007196713238954544, + 0.007802568841725588, + 0.006233859341591597, + 0.006746539380401373, + 0.008425458334386349, + 0.005342606920748949, + 0.007891319692134857, + 0.004231567028909922, + 0.010163369588553905, + 0.011763062328100204, + 0.010223784483969212, + 0.008509567938745022, + 0.011191234923899174, + 0.007383167278021574, + 0.009346172213554382, + 0.005085252225399017, + 0.005714050028473139, + 0.005376795306801796, + 0.0064564356580376625, + 0.011073803529143333, + 0.0038116180803626776, + 0.0040289792232215405, + 0.005130064208060503, + 0.003919858951121569, + 0.00852929800748825, + 0.007982286624610424, + 0.009215222671627998, + 0.0046523623168468475, + 0.006030993536114693, + 0.005727952811866999, + 0.005551984533667564, + 0.005094227381050587, + 0.008289555087685585, + 0.006361703854054213, + 0.009305666200816631, + 0.009172514081001282, + 0.006252795457839966, + 0.0058363256976008415, + 0.01025212649255991, + 0.004881717264652252, + 0.011828967370092869, + 0.005426359362900257, + 0.013401759788393974, + 0.007071902044117451, + 0.004987488500773907, + 0.005421312525868416, + 0.004017213359475136, + 0.0030537510756403208, + 0.0071855951100587845, + 0.0039199828170239925, + 0.00715358043089509, + 0.003842522855848074, + 0.0036839130334556103, + 0.0047377017326653, + 0.0056715188547968864, + 0.0031840659212321043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029795313253998756 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly scholar with long, thinning white hair tied in a traditional headdress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a deep indigo silk robe with silver thread embroidery of clouds and cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at a dark lacquered wooden altar containing a half-unrolled yellowed rice paper scroll, a jade brush holder, and a dark slate inkstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scholar's right hand resting above the inkstone with a pool of freshly ground ink showing a thick, dark sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a subtle, persistent tremor in the scholar's fingers while being bathed in warm, low-key lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902440965175629 + ], + "opposite_logprob_tokens": [ + -2.464024305343628 + ], + "logprob_reward": -0.08902440965175629, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902440965175629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4063536822795868, + 0.13905611634254456, + 0.055932629853487015, + 0.04282262921333313, + 0.045438218861818314, + 0.054262034595012665, + 0.036431293934583664, + 0.07813236117362976, + 0.10387638956308365, + 0.037694647908210754 + ], + "credit_map": [ + 0.05682460963726044, + 0.06455356627702713, + 0.0475449301302433, + 0.053072426468133926, + 0.06860294938087463, + 0.03889245167374611, + 0.05616793408989906, + 0.020694823935627937, + 0.01977413520216942, + 0.02282528579235077, + 0.017305582761764526, + 0.015383952297270298, + 0.02415071614086628, + 0.01232834067195654, + 0.01924476958811283, + 0.008043334819376469, + 0.008311470970511436, + 0.006977591663599014, + 0.009679710492491722, + 0.012189988046884537, + 0.004025960806757212, + 0.0048249997198581696, + 0.00640000170096755, + 0.0035229010973125696, + 0.006463845260441303, + 0.006243207026273012, + 0.008572663180530071, + 0.0045845117419958115, + 0.004546585958451033, + 0.004986155312508345, + 0.004037456586956978, + 0.0033882013522088528, + 0.010210215114057064, + 0.005312602035701275, + 0.009178760461509228, + 0.0042174383997917175, + 0.004932657815515995, + 0.004171281587332487, + 0.005224252585321665, + 0.002191005041822791, + 0.008066374808549881, + 0.009140166454017162, + 0.007852905429899693, + 0.00627440121024847, + 0.009971351362764835, + 0.004710161592811346, + 0.005006504710763693, + 0.003240167163312435, + 0.005337720271199942, + 0.0048356312327086926, + 0.005852343514561653, + 0.007692132610827684, + 0.0037851580418646336, + 0.0038151873741298914, + 0.002872304990887642, + 0.0022408145014196634, + 0.011012457311153412, + 0.014928912743926048, + 0.010180133394896984, + 0.010069910436868668, + 0.012726533226668835, + 0.006604356691241264, + 0.00875061471015215, + 0.003859441028907895, + 0.02080744318664074, + 0.020350256934762, + 0.019399959594011307, + 0.009352819994091988, + 0.011668060906231403, + 0.009129464626312256, + 0.007898388430476189, + 0.005269990302622318, + 0.006573437247425318, + 0.005812756717205048, + 0.007367420010268688, + 0.0031509066466242075, + 0.0039830501191318035, + 0.004586354363709688, + 0.0027643931098282337, + 0.003456331556662917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.08902440965175629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_5/index_tensor([720, 720])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_5/index_tensor([720, 720])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cfd0f98a50a575ca840e501b5def969942ce61b5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_5/index_tensor([720, 720])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This extreme close-up, rendered in mesmerizing slow motion, begins by tracking a gleaming, polished chrome sphere as it makes contact with a dark, richly textured wooden surface. The camera, perfectly synchronized with the action, precisely tracks the sphere's movement. The sphere itself is impeccably smooth, its mirror-like surface reflecting a distorted, elongated image of a single, bright studio softbox light against a dark, out-of-focus background. It bounces off a dark, matte-finished wooden board, its deep, pronounced texture creating a rich, organic contrast to the sphere's industrial perfection. As the sphere silently and deliberately lifts off the surface, it begins a slow, majestic rotation, the camera sliding laterally to track its graceful arc through the air. The scene is bathed in high-contrast light, with the single light source creating brilliant white highlights along the sphere's curve and casting deep shadows that accentuate the wood's grain, creating a cool, minimalist color palette of silver and near-black. The overall atmosphere is one of serene focus and hypnotic elegance, a pure visual study of form, texture, and movement.", + "seed": 21081, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_5/index_720_seed_21081_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02009346355916932, + "vlm_logprob": -0.02009346355916932, + "vlm_yes_no_margin": 4.874999952316284, + "vlm_token_credit": -0.02009346355916932, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005273479502648115 + ], + "opposite_logprob_tokens": [ + -6.755273342132568 + ], + "logprob_reward": -0.005273479502648115, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.005273479502648115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18467102944850922, + 0.07836727797985077, + 0.07941442728042603, + 0.10814955085515976, + 0.09024260938167572, + 0.06698615103960037, + 0.06043606996536255, + 0.07732504606246948, + 0.09457407891750336, + 0.15983377397060394 + ], + "credit_map": [ + 0.029275882989168167, + 0.015829356387257576, + 0.036038000136613846, + 0.04093930125236511, + 0.00964334886521101, + 0.013824481517076492, + 0.016500301659107208, + 0.022620348259806633, + 0.010030914098024368, + 0.008716092444956303, + 0.018224947154521942, + 0.01143775973469019, + 0.006608175113797188, + 0.005856157746165991, + 0.006008364725857973, + 0.011484864167869091, + 0.011811891570687294, + 0.011363827623426914, + 0.0149725042283535, + 0.008270984515547752, + 0.005808345507830381, + 0.00640117097645998, + 0.010569870471954346, + 0.010215830989181995, + 0.019163738936185837, + 0.017239153385162354, + 0.021538758650422096, + 0.013476998545229435, + 0.004872285760939121, + 0.006123831029981375, + 0.0071393162943422794, + 0.01859547197818756, + 0.008787429891526699, + 0.01249935757368803, + 0.015237579122185707, + 0.01920147053897381, + 0.006444225553423166, + 0.008067246526479721, + 0.007035137619823217, + 0.012970159761607647, + 0.009458079002797604, + 0.009966069832444191, + 0.008464155718684196, + 0.0075589753687381744, + 0.005346423480659723, + 0.00972718931734562, + 0.004878452979028225, + 0.011586803011596203, + 0.008169207721948624, + 0.005926546175032854, + 0.009136997163295746, + 0.007675993721932173, + 0.005200021434575319, + 0.004982999991625547, + 0.005743363872170448, + 0.013600945472717285, + 0.01358171459287405, + 0.016824893653392792, + 0.011458894237875938, + 0.00810131523758173, + 0.005899073556065559, + 0.004688789136707783, + 0.007654910441488028, + 0.009115458466112614, + 0.01547033991664648, + 0.007434506434947252, + 0.018153872340917587, + 0.011799131520092487, + 0.01119320560246706, + 0.008398596197366714, + 0.011032497510313988, + 0.011091923341155052, + 0.02177049033343792, + 0.014101681299507618, + 0.02304236590862274, + 0.015334655530750751, + 0.01593896932899952, + 0.01850874163210392, + 0.02378653734922409, + 0.027350332587957382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.005273479502648115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810567781329155 + ], + "opposite_logprob_tokens": [ + -3.2881057262420654 + ], + "logprob_reward": -0.03810567781329155, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810567781329155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052380308508872986, + 0.03462299332022667, + 0.056711770594120026, + 0.14360429346561432, + 0.34499555826187134, + 0.03320728987455368, + 0.17768211662769318, + 0.0463167168200016, + 0.06493169814348221, + 0.0455472432076931 + ], + "credit_map": [ + 0.007799849845468998, + 0.004357038997113705, + 0.009762998670339584, + 0.00807631853967905, + 0.003432739060372114, + 0.007054654415696859, + 0.0053397538140416145, + 0.0065569630824029446, + 0.004836462438106537, + 0.0036589677911251783, + 0.010067573748528957, + 0.0042246002703905106, + 0.0017803385853767395, + 0.002506893826648593, + 0.0022522425279021263, + 0.005295915994793177, + 0.00804983265697956, + 0.007461997214704752, + 0.011645595543086529, + 0.009736713953316212, + 0.0033067038748413324, + 0.002953519579023123, + 0.005116216838359833, + 0.008441193960607052, + 0.024720780551433563, + 0.026454927399754524, + 0.03218773007392883, + 0.02212269976735115, + 0.006888572592288256, + 0.006844338960945606, + 0.008401565253734589, + 0.015983691439032555, + 0.12846925854682922, + 0.13301366567611694, + 0.01522382628172636, + 0.015176135115325451, + 0.0090919379144907, + 0.020973483100533485, + 0.00919062178581953, + 0.013856666162610054, + 0.006911491043865681, + 0.006810986902564764, + 0.00437763100489974, + 0.004485586192458868, + 0.0021876979153603315, + 0.0036152389366179705, + 0.0017898209625855088, + 0.003028837963938713, + 0.02862657979130745, + 0.02602565847337246, + 0.03915970027446747, + 0.020508650690317154, + 0.011786816641688347, + 0.016918987035751343, + 0.015455066226422787, + 0.01920066587626934, + 0.006645568646490574, + 0.005742174107581377, + 0.009952177293598652, + 0.006744772661477327, + 0.003060580464079976, + 0.003757565515115857, + 0.004675473552197218, + 0.005738406907767057, + 0.009891538880765438, + 0.006392777897417545, + 0.012811636552214622, + 0.007081644609570503, + 0.006735920440405607, + 0.00598475057631731, + 0.008757435716688633, + 0.007275999989360571, + 0.007002216763794422, + 0.0033009101171046495, + 0.007958550937473774, + 0.003027377650141716, + 0.005165659822523594, + 0.006227799225598574, + 0.005822339095175266, + 0.007042392622679472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.03810567781329155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.265505850315094, + 0.056822992861270905, + 0.0530724935233593, + 0.13102281093597412, + 0.15061987936496735, + 0.03221764788031578, + 0.07544352859258652, + 0.03823151811957359, + 0.06040405109524727, + 0.13665920495986938 + ], + "credit_map": [ + 0.03906060382723808, + 0.022710248827934265, + 0.03475424647331238, + 0.03210875391960144, + 0.026629863306879997, + 0.04736841470003128, + 0.020461905747652054, + 0.04241182655096054, + 0.008173567242920399, + 0.006006283685564995, + 0.01456732489168644, + 0.008767141960561275, + 0.005313650704920292, + 0.003924896474927664, + 0.003725373651832342, + 0.006344759836792946, + 0.006442917976528406, + 0.006931135430932045, + 0.009249315597116947, + 0.007343309931457043, + 0.00483537744730711, + 0.0035247341729700565, + 0.007118513807654381, + 0.007627193816006184, + 0.040784791111946106, + 0.03880171850323677, + 0.012841416522860527, + 0.00968049094080925, + 0.004481293261051178, + 0.0049384236335754395, + 0.005953073967248201, + 0.013541619293391705, + 0.05587110295891762, + 0.056449342519044876, + 0.00655921408906579, + 0.006904589012265205, + 0.00430951127782464, + 0.00971018336713314, + 0.0043419417925179005, + 0.006474006455391645, + 0.00499894842505455, + 0.00453483359888196, + 0.003161374945193529, + 0.003028566250577569, + 0.002589674200862646, + 0.006108636502176523, + 0.002694634022191167, + 0.005100980866700411, + 0.010763290338218212, + 0.011869845911860466, + 0.010398836806416512, + 0.007908972911536694, + 0.006043546367436647, + 0.011492252349853516, + 0.004676086362451315, + 0.012290705926716328, + 0.005336368922144175, + 0.0037423274479806423, + 0.006113815121352673, + 0.004638694226741791, + 0.003000768832862377, + 0.00481343874707818, + 0.004641816020011902, + 0.005944292992353439, + 0.010441120713949203, + 0.0041091591119766235, + 0.013670212589204311, + 0.007052281405776739, + 0.0054096318781375885, + 0.005538837518543005, + 0.006555836647748947, + 0.007626982871443033, + 0.01986234076321125, + 0.017227858304977417, + 0.027905277907848358, + 0.012835401110351086, + 0.013653472997248173, + 0.013235737569630146, + 0.017712324857711792, + 0.014226813800632954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.01819819025695324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038075268268585205 + ], + "opposite_logprob_tokens": [ + -3.2880752086639404 + ], + "logprob_reward": -0.038075268268585205, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038075268268585205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09035838395357132, + 0.03355410322546959, + 0.03632373362779617, + 0.10941219329833984, + 0.29363590478897095, + 0.02190922200679779, + 0.1291571706533432, + 0.07565110176801682, + 0.06934679299592972, + 0.14065131545066833 + ], + "credit_map": [ + 0.012960747815668583, + 0.006367717403918505, + 0.017392776906490326, + 0.01723996177315712, + 0.007390529848635197, + 0.011899757198989391, + 0.0061434474773705006, + 0.010963452979922295, + 0.005086636170744896, + 0.003771305549889803, + 0.00849114079028368, + 0.0040657673962414265, + 0.0023641830775886774, + 0.0036577386781573296, + 0.0025086309760808945, + 0.0036087024491280317, + 0.004539770539849997, + 0.005154729355126619, + 0.005676732864230871, + 0.005146828480064869, + 0.0027826689183712006, + 0.002931928727775812, + 0.004206390120089054, + 0.005884688813239336, + 0.03449539095163345, + 0.033651746809482574, + 0.010063709691166878, + 0.008185501210391521, + 0.0035741885658353567, + 0.004043309483677149, + 0.004210474900901318, + 0.011187873780727386, + 0.10811050236225128, + 0.10849140584468842, + 0.013698996976017952, + 0.01542726531624794, + 0.007765553425997496, + 0.017950940877199173, + 0.007919403724372387, + 0.014271853491663933, + 0.003571568289771676, + 0.0036863808054476976, + 0.002364460611715913, + 0.003641139715909958, + 0.0016531662549823523, + 0.001996138831600547, + 0.0019040198531001806, + 0.0030923481099307537, + 0.02005431056022644, + 0.012381626293063164, + 0.02444838546216488, + 0.019683638587594032, + 0.009287536144256592, + 0.010872721672058105, + 0.008247236721217632, + 0.02418171986937523, + 0.012831084430217743, + 0.011447868309915066, + 0.016492502763867378, + 0.010145053267478943, + 0.005431333556771278, + 0.004759784787893295, + 0.006714036222547293, + 0.00782945565879345, + 0.010566896758973598, + 0.005370300263166428, + 0.01635049469769001, + 0.007794477045536041, + 0.005947323050349951, + 0.005291631445288658, + 0.007907104678452015, + 0.010118572041392326, + 0.01993270218372345, + 0.01567870005965233, + 0.02164381928741932, + 0.013799622654914856, + 0.016331331804394722, + 0.01830112561583519, + 0.017988724634051323, + 0.016975294798612595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.038075268268585205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008147019543685019 + ], + "opposite_logprob_tokens": [ + -7.125814914703369 + ], + "logprob_reward": -0.0008147019543685019, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008147019543685019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08640759438276291, + 0.0278747770935297, + 0.019572358578443527, + 0.10675785690546036, + 0.4048922061920166, + 0.013227108865976334, + 0.09506524354219437, + 0.05009769648313522, + 0.04892615228891373, + 0.1471790075302124 + ], + "credit_map": [ + 0.014224240556359291, + 0.005581723526120186, + 0.018031837418675423, + 0.013411675579845905, + 0.0052173868753015995, + 0.011079922318458557, + 0.006559697445482016, + 0.012301111593842506, + 0.003167912596836686, + 0.00434663612395525, + 0.005528760142624378, + 0.00391006376594305, + 0.0017774550942704082, + 0.002547221491113305, + 0.0023458260111510754, + 0.004250902682542801, + 0.002541320165619254, + 0.002398789394646883, + 0.002485391916707158, + 0.0024137510918080807, + 0.002174892695620656, + 0.0020863546524196863, + 0.0028178361244499683, + 0.002654022304341197, + 0.028117598965764046, + 0.026635870337486267, + 0.014538242481648922, + 0.010038587264716625, + 0.0045423349365592, + 0.004413161892443895, + 0.005602393765002489, + 0.012869661673903465, + 0.15119673311710358, + 0.15464308857917786, + 0.017528193071484566, + 0.02106231264770031, + 0.010838323272764683, + 0.024483516812324524, + 0.00905216857790947, + 0.016087869182229042, + 0.0015356263611465693, + 0.002384684979915619, + 0.002162705408409238, + 0.002016497543081641, + 0.0011977963149547577, + 0.0012082034954801202, + 0.001043067080900073, + 0.0016785277985036373, + 0.014874506741762161, + 0.009748931974172592, + 0.023102110251784325, + 0.0138546172529459, + 0.007196987979114056, + 0.0076540736481547356, + 0.0061792414635419846, + 0.012454773299396038, + 0.008074585348367691, + 0.009001151658594608, + 0.008866150863468647, + 0.0048094564117491245, + 0.004295783583074808, + 0.004730068612843752, + 0.004178847186267376, + 0.006141650024801493, + 0.0072699179872870445, + 0.0061499630101025105, + 0.011051630601286888, + 0.00509169464930892, + 0.003952842205762863, + 0.004283389542251825, + 0.004828388802707195, + 0.006298328284174204, + 0.021898634731769562, + 0.01801924966275692, + 0.022652925923466682, + 0.013127658516168594, + 0.013184855692088604, + 0.015129528939723969, + 0.01472631935030222, + 0.028439827263355255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.0008147019543685019 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extreme close-up, rendered in mesmerizing slow motion, begins by tracking a gleaming, polished chrome sphere as it makes contact with a dark, richly textured wooden surface. The camera, perfectly synchronized with the action, precisely tracks the sphere's movement. The sphere itself is impeccably smooth, its mirror-like surface reflecting a distorted, elongated image of a single, bright studio softbox light against a dark, out-of-focus background. It bounces off a dark, matte-finished wooden board, its deep, pronounced texture creating a rich, organic contrast to the sphere's industrial perfection. As the sphere silently and deliberately lifts off the surface, it begins a slow, majestic rotation, the camera sliding laterally to track its graceful arc through the air. The scene is bathed in high-contrast light, with the single light source creating brilliant white highlights along the sphere's curve and casting deep shadows that accentuate the wood's grain, creating a cool, minimalist color palette of silver and near-black. The overall atmosphere is one of serene focus and hypnotic elegance, a pure visual study of form, texture, and movement.", + "seed": 21082, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_5/index_720_seed_21082_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06756467715604231, + "vlm_logprob": -0.06756467715604231, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.06756467715604231, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008059362880885601 + ], + "opposite_logprob_tokens": [ + -7.758059501647949 + ], + "logprob_reward": -0.008059362880885601, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.008059362880885601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29642558097839355, + 0.04698587954044342, + 0.05200207233428955, + 0.051398105919361115, + 0.05166614428162575, + 0.06069160997867584, + 0.1337766945362091, + 0.11231739073991776, + 0.15516825020313263, + 0.0395682230591774 + ], + "credit_map": [ + 0.0310431569814682, + 0.0598791278898716, + 0.10182168334722519, + 0.029622169211506844, + 0.01935393363237381, + 0.021126337349414825, + 0.01895485632121563, + 0.014624345116317272, + 0.0033712792210280895, + 0.005334421992301941, + 0.005786146502941847, + 0.005524765700101852, + 0.008458144962787628, + 0.0052414764650166035, + 0.007392760366201401, + 0.0058768922463059425, + 0.006110659334808588, + 0.004358774516731501, + 0.009022718295454979, + 0.00475424574688077, + 0.0076140412129461765, + 0.007868295535445213, + 0.005556140560656786, + 0.006717204581946135, + 0.006041836459189653, + 0.006411678157746792, + 0.006732419598847628, + 0.0052847228944301605, + 0.006774865090847015, + 0.005498811602592468, + 0.007361347787082195, + 0.007292431779205799, + 0.004070440772920847, + 0.006317386869341135, + 0.008653702214360237, + 0.0056365784257650375, + 0.007935700938105583, + 0.004382738843560219, + 0.010104816406965256, + 0.004564789589494467, + 0.0037223666440695524, + 0.0064558833837509155, + 0.0069371555000543594, + 0.006146513856947422, + 0.011152518913149834, + 0.008449300192296505, + 0.009558693505823612, + 0.00826918613165617, + 0.016807999461889267, + 0.022477537393569946, + 0.023811066523194313, + 0.015436102636158466, + 0.014662190340459347, + 0.015973970293998718, + 0.010717731900513172, + 0.013890108093619347, + 0.017274919897317886, + 0.005576182156801224, + 0.009739469736814499, + 0.00627678819000721, + 0.002870104741305113, + 0.009002587758004665, + 0.017863048240542412, + 0.04371429979801178, + 0.012727302499115467, + 0.011059927754104137, + 0.011832574382424355, + 0.015747003257274628, + 0.007271284237504005, + 0.013850534334778786, + 0.05577753484249115, + 0.02690211869776249, + 0.005708407144993544, + 0.006816818378865719, + 0.006093424744904041, + 0.005367076024413109, + 0.0027599597815424204, + 0.0030459873378276825, + 0.004292525351047516, + 0.005484029650688171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.008059362880885601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556239366531372 + ], + "opposite_logprob_tokens": [ + -1.600562334060669 + ], + "logprob_reward": -0.22556239366531372, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556239366531372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2740633189678192, + 0.08079105615615845, + 0.039440739899873734, + 0.09763406217098236, + 0.019953636452555656, + 0.04978720843791962, + 0.10610409826040268, + 0.06467468291521072, + 0.20826613903045654, + 0.05928505212068558 + ], + "credit_map": [ + 0.025580788031220436, + 0.06904173642396927, + 0.07960336655378342, + 0.03112826682627201, + 0.017033575102686882, + 0.02199927717447281, + 0.017311669886112213, + 0.012364629656076431, + 0.008043773472309113, + 0.013883412815630436, + 0.00945671834051609, + 0.01271913107484579, + 0.011544019915163517, + 0.006465521175414324, + 0.008331529796123505, + 0.010346941649913788, + 0.005376679357141256, + 0.003306312719359994, + 0.006654114928096533, + 0.003087724093347788, + 0.004984935279935598, + 0.006380682345479727, + 0.004413846880197525, + 0.005236441735178232, + 0.01587950624525547, + 0.0092629324644804, + 0.01515298243612051, + 0.009984749369323254, + 0.008553526364266872, + 0.008552836254239082, + 0.010164023377001286, + 0.02008349634706974, + 0.001602534088306129, + 0.0025744454469531775, + 0.0030736965127289295, + 0.002025855937972665, + 0.003476283513009548, + 0.0014792593428865075, + 0.004245065618306398, + 0.0014764926163479686, + 0.0038004848174750805, + 0.004213288892060518, + 0.005144395399838686, + 0.004343083593994379, + 0.009626081213355064, + 0.007502512075006962, + 0.008758991956710815, + 0.006398367695510387, + 0.021446188911795616, + 0.00821506604552269, + 0.01616761088371277, + 0.008634239435195923, + 0.010365120135247707, + 0.020506909117102623, + 0.008088882081210613, + 0.012680074200034142, + 0.010737852193415165, + 0.004139141645282507, + 0.009110786020755768, + 0.004472400061786175, + 0.0022025490179657936, + 0.005086051300168037, + 0.00472105061635375, + 0.024204842746257782, + 0.014795741066336632, + 0.013687980361282825, + 0.010227176360785961, + 0.02133122645318508, + 0.008338609710335732, + 0.011243862099945545, + 0.07569701969623566, + 0.05294450744986534, + 0.006806907244026661, + 0.007583119906485081, + 0.007188218180090189, + 0.007845073007047176, + 0.003710357239469886, + 0.0035206982865929604, + 0.012793746776878834, + 0.009836922399699688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.22556239366531372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486411526799202 + ], + "opposite_logprob_tokens": [ + -3.0486412048339844 + ], + "logprob_reward": -0.0486411526799202, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486411526799202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4743342101573944, + 0.022385504096746445, + 0.045668065547943115, + 0.021106744185090065, + 0.016373930498957634, + 0.05361917242407799, + 0.08582784980535507, + 0.048611052334308624, + 0.16172237694263458, + 0.07035107165575027 + ], + "credit_map": [ + 0.07080922275781631, + 0.1589733213186264, + 0.14022381603717804, + 0.03413469344377518, + 0.021636946126818657, + 0.022495729848742485, + 0.013787384144961834, + 0.012273109517991543, + 0.002414653543382883, + 0.002840887289494276, + 0.0026476539205759764, + 0.0033475838135927916, + 0.002092338865622878, + 0.0023938179947435856, + 0.002114734845235944, + 0.004533832892775536, + 0.005965180695056915, + 0.007648534141480923, + 0.006796372588723898, + 0.0036902171559631824, + 0.0054919118992984295, + 0.006504354532808065, + 0.0050096409395337105, + 0.004561851732432842, + 0.0027733149472624063, + 0.0019721437711268663, + 0.003281376324594021, + 0.0021520706359297037, + 0.0018967910436913371, + 0.0019546858966350555, + 0.002287123817950487, + 0.004789238795638084, + 0.000992681598290801, + 0.0026287254877388477, + 0.002417138312011957, + 0.0025954819284379482, + 0.0031736211385577917, + 0.0008360340143553913, + 0.0025902094785124063, + 0.0011400390649214387, + 0.005684466101229191, + 0.0083145871758461, + 0.004262267611920834, + 0.005015382543206215, + 0.011920836754143238, + 0.007364869117736816, + 0.006201635580509901, + 0.004855127073824406, + 0.011242911219596863, + 0.011252148076891899, + 0.010605134069919586, + 0.013350237160921097, + 0.009591395035386086, + 0.011008352041244507, + 0.008889379911124706, + 0.009888297878205776, + 0.006859117187559605, + 0.006040790118277073, + 0.007058834657073021, + 0.007484226021915674, + 0.003292820416390896, + 0.006646779831498861, + 0.005050999578088522, + 0.00617748498916626, + 0.011665979400277138, + 0.00627670343965292, + 0.007596527226269245, + 0.014590284787118435, + 0.007212318480014801, + 0.009600176475942135, + 0.0663025826215744, + 0.03847780451178551, + 0.008467020466923714, + 0.005819118116050959, + 0.010934894904494286, + 0.0070574041455984116, + 0.007237073965370655, + 0.010779031552374363, + 0.011912145651876926, + 0.008144385181367397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0486411526799202 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492541193962097 + ], + "opposite_logprob_tokens": [ + -2.9299254417419434 + ], + "logprob_reward": -0.05492541193962097, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492541193962097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1692505180835724, + 0.03962160274386406, + 0.046968527138233185, + 0.07426547259092331, + 0.08165941387414932, + 0.13347499072551727, + 0.09457322955131531, + 0.2340158075094223, + 0.0773332417011261, + 0.048837170004844666 + ], + "credit_map": [ + 0.02699838951230049, + 0.04803289845585823, + 0.039481569081544876, + 0.013107754290103912, + 0.009646023623645306, + 0.01101065892726183, + 0.011571635492146015, + 0.009401594288647175, + 0.00406322767958045, + 0.005671291146427393, + 0.004480568692088127, + 0.0066049289889633656, + 0.0043602026998996735, + 0.0030401276890188456, + 0.004260044079273939, + 0.007141213398426771, + 0.005162855610251427, + 0.003551868489012122, + 0.010339257307350636, + 0.003257783828303218, + 0.005594004411250353, + 0.007007736247032881, + 0.004932821728289127, + 0.0071222009137272835, + 0.010637461207807064, + 0.006588075775653124, + 0.01727827824652195, + 0.005028663668781519, + 0.00941445492208004, + 0.004265730734914541, + 0.008238393813371658, + 0.01281441655009985, + 0.004609691444784403, + 0.011638662777841091, + 0.00959666259586811, + 0.0186286810785532, + 0.009537767618894577, + 0.0050737266428768635, + 0.014812210574746132, + 0.007762014400213957, + 0.011822346597909927, + 0.017239727079868317, + 0.012542624957859516, + 0.015114491805434227, + 0.02459237352013588, + 0.016484547406435013, + 0.019911764189600945, + 0.01576712168753147, + 0.012088317424058914, + 0.01362532377243042, + 0.010044205002486706, + 0.01508140005171299, + 0.008920760825276375, + 0.013293635100126266, + 0.01294914074242115, + 0.008570458739995956, + 0.03153872862458229, + 0.018652163445949554, + 0.02118406444787979, + 0.027511049062013626, + 0.012097010388970375, + 0.023132845759391785, + 0.03365381807088852, + 0.0662461370229721, + 0.010027886368334293, + 0.02260817401111126, + 0.008253025822341442, + 0.011748072691261768, + 0.0032731404062360525, + 0.00516257481649518, + 0.006421674974262714, + 0.009838690981268883, + 0.00658380426466465, + 0.007294435985386372, + 0.005114917643368244, + 0.006165180820971727, + 0.0028924865182489157, + 0.0027196763549000025, + 0.011372403241693974, + 0.0066942693665623665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.05492541193962097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006350646144710481 + ], + "opposite_logprob_tokens": [ + -7.375635147094727 + ], + "logprob_reward": -0.0006350646144710481, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006350646144710481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6420702338218689, + 0.037309933453798294, + 0.03422291949391365, + 0.026286263018846512, + 0.032611873000860214, + 0.0519007183611393, + 0.020169837400317192, + 0.03133794665336609, + 0.05663753300905228, + 0.06745274364948273 + ], + "credit_map": [ + 0.0896126851439476, + 0.2196316123008728, + 0.20841792225837708, + 0.05212985351681709, + 0.024309484288096428, + 0.021165894344449043, + 0.012524428777396679, + 0.01427834015339613, + 0.0026477316860109568, + 0.006451501976698637, + 0.004425958730280399, + 0.005953346844762564, + 0.0052274721674621105, + 0.0033231396228075027, + 0.004056949634104967, + 0.005223835818469524, + 0.004431909415870905, + 0.0031290987972170115, + 0.007145564071834087, + 0.0029400987550616264, + 0.0037826108746230602, + 0.004002301953732967, + 0.0035659181885421276, + 0.0052254158072173595, + 0.004756482318043709, + 0.0030270223505795, + 0.0037820565048605204, + 0.0021693215239793062, + 0.002300594002008438, + 0.0024718863423913717, + 0.002557561732828617, + 0.005221338011324406, + 0.0019170520827174187, + 0.004267525393515825, + 0.005230430979281664, + 0.005336622707545757, + 0.005162977613508701, + 0.002041560597717762, + 0.006413538474589586, + 0.002242164220660925, + 0.005860012955963612, + 0.008807321079075336, + 0.004541737958788872, + 0.004374484531581402, + 0.010867300443351269, + 0.006521865725517273, + 0.005289264023303986, + 0.005638734437525272, + 0.0031394697725772858, + 0.002056537428870797, + 0.0025490832049399614, + 0.0029282488394528627, + 0.0015409644693136215, + 0.0023053414188325405, + 0.001931779901497066, + 0.0037184113170951605, + 0.005720632616430521, + 0.003662297734990716, + 0.004748326726257801, + 0.0035087857395410538, + 0.002293527591973543, + 0.004254956729710102, + 0.0026725896168500185, + 0.004476829431951046, + 0.00643295980989933, + 0.008540432900190353, + 0.004703586455434561, + 0.010036081075668335, + 0.0018655774183571339, + 0.0023596310056746006, + 0.013590754941105843, + 0.009108510799705982, + 0.00843732338398695, + 0.00784274935722351, + 0.00882965512573719, + 0.007789429277181625, + 0.004405890125781298, + 0.005715448409318924, + 0.0167418010532856, + 0.007690449710935354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0006350646144710481 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extreme close-up, rendered in mesmerizing slow motion, begins by tracking a gleaming, polished chrome sphere as it makes contact with a dark, richly textured wooden surface. The camera, perfectly synchronized with the action, precisely tracks the sphere's movement. The sphere itself is impeccably smooth, its mirror-like surface reflecting a distorted, elongated image of a single, bright studio softbox light against a dark, out-of-focus background. It bounces off a dark, matte-finished wooden board, its deep, pronounced texture creating a rich, organic contrast to the sphere's industrial perfection. As the sphere silently and deliberately lifts off the surface, it begins a slow, majestic rotation, the camera sliding laterally to track its graceful arc through the air. The scene is bathed in high-contrast light, with the single light source creating brilliant white highlights along the sphere's curve and casting deep shadows that accentuate the wood's grain, creating a cool, minimalist color palette of silver and near-black. The overall atmosphere is one of serene focus and hypnotic elegance, a pure visual study of form, texture, and movement.", + "seed": 21083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_5/index_720_seed_21083_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02654375075362623, + "vlm_logprob": -0.02654375075362623, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.02654375075362623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010497826151549816 + ], + "opposite_logprob_tokens": [ + -7.260498046875 + ], + "logprob_reward": -0.010497826151549816, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.010497826151549816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21261632442474365, + 0.08390502631664276, + 0.12008257955312729, + 0.10936643183231354, + 0.08093312382698059, + 0.07128267735242844, + 0.09148083627223969, + 0.07268732041120529, + 0.08418405055999756, + 0.07346168160438538 + ], + "credit_map": [ + 0.02088923007249832, + 0.020171944051980972, + 0.02949841134250164, + 0.03705158457159996, + 0.016383308917284012, + 0.013481282629072666, + 0.035883452743291855, + 0.03925709053874016, + 0.004540134686976671, + 0.010553285479545593, + 0.0057296971790492535, + 0.023442832753062248, + 0.005253840237855911, + 0.004419621545821428, + 0.015504280105233192, + 0.014461333863437176, + 0.017079604789614677, + 0.0188424214720726, + 0.017665762454271317, + 0.027529815211892128, + 0.00533558102324605, + 0.005490309093147516, + 0.013829391449689865, + 0.014309684745967388, + 0.006959647871553898, + 0.02338302880525589, + 0.010087736882269382, + 0.03774220496416092, + 0.004859426524490118, + 0.0056186942383646965, + 0.008979004807770252, + 0.011736680753529072, + 0.011176591739058495, + 0.01408686675131321, + 0.013936868868768215, + 0.016184868291020393, + 0.0046457564458251, + 0.005628223065286875, + 0.007264665327966213, + 0.008009272627532482, + 0.012604177929461002, + 0.013753498904407024, + 0.013548354618251324, + 0.011582793667912483, + 0.0026999791152775288, + 0.003184877336025238, + 0.007442759815603495, + 0.006466228049248457, + 0.016174301505088806, + 0.015949975699186325, + 0.010766780935227871, + 0.01239131297916174, + 0.00712190568447113, + 0.01083850022405386, + 0.011581999249756336, + 0.0066560544073581696, + 0.012252075597643852, + 0.008803015574812889, + 0.017554955556988716, + 0.012070215307176113, + 0.004376292694360018, + 0.004273165483027697, + 0.008017463609576225, + 0.00534012820571661, + 0.011929460801184177, + 0.007442612200975418, + 0.01886063441634178, + 0.01304674707353115, + 0.004517287481576204, + 0.00635706027969718, + 0.0140288220718503, + 0.008001419715583324, + 0.011942162178456783, + 0.007961160503327847, + 0.015860624611377716, + 0.005791315343230963, + 0.005741864442825317, + 0.011560625396668911, + 0.006391406990587711, + 0.008212515152990818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.010497826151549816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368720456957817 + ], + "opposite_logprob_tokens": [ + -3.6513686180114746 + ], + "logprob_reward": -0.026368720456957817, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368720456957817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13192783296108246, + 0.04163890331983566, + 0.08354145288467407, + 0.047226469963788986, + 0.21014533936977386, + 0.08165328949689865, + 0.13578923046588898, + 0.0929965153336525, + 0.09822206944227219, + 0.07685881853103638 + ], + "credit_map": [ + 0.010529771447181702, + 0.012457815930247307, + 0.012706044130027294, + 0.01833931729197502, + 0.010826138779520988, + 0.008895590901374817, + 0.027352990582585335, + 0.030820172280073166, + 0.0032834087032824755, + 0.005414821673184633, + 0.003617631271481514, + 0.008296617306768894, + 0.0034852384123951197, + 0.0032434852328151464, + 0.006194123066961765, + 0.00810358114540577, + 0.0073076109401881695, + 0.014517586678266525, + 0.008900756016373634, + 0.025618325918912888, + 0.004255260806530714, + 0.004742973018437624, + 0.009512729942798615, + 0.008686208166182041, + 0.003073583822697401, + 0.007858527824282646, + 0.006861095782369375, + 0.009695187211036682, + 0.0022586986888200045, + 0.0035916040651500225, + 0.0083171920850873, + 0.0055705830454826355, + 0.023483136668801308, + 0.03869936242699623, + 0.043753575533628464, + 0.03645485267043114, + 0.009367287158966064, + 0.014849456958472729, + 0.02273509092628956, + 0.020802590996026993, + 0.009300795383751392, + 0.013538826256990433, + 0.017514754086732864, + 0.016822010278701782, + 0.003445267677307129, + 0.006166602484881878, + 0.008674909360706806, + 0.0061901286244392395, + 0.02404453046619892, + 0.020084530115127563, + 0.022871635854244232, + 0.019614942371845245, + 0.007379428017884493, + 0.010567313991487026, + 0.02014104463160038, + 0.011085817590355873, + 0.01897295005619526, + 0.010479526594281197, + 0.02196696400642395, + 0.009525583125650883, + 0.007864161394536495, + 0.00782241765409708, + 0.008376678451895714, + 0.007988235913217068, + 0.018203875049948692, + 0.007959396578371525, + 0.03092237561941147, + 0.00903406273573637, + 0.00623244047164917, + 0.008146904408931732, + 0.00777641823515296, + 0.009946595877408981, + 0.010586148127913475, + 0.007269659079611301, + 0.011898952536284924, + 0.0072919283993542194, + 0.010677366517484188, + 0.010354140773415565, + 0.010580991394817829, + 0.008199638687074184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.026368720456957817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381014309823513 + ], + "opposite_logprob_tokens": [ + -3.2881014347076416 + ], + "logprob_reward": -0.0381014309823513, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381014309823513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20570921897888184, + 0.0813559740781784, + 0.11739923059940338, + 0.05610405281186104, + 0.10820570588111877, + 0.05486176535487175, + 0.07634817063808441, + 0.05664965882897377, + 0.07828378677368164, + 0.16508245468139648 + ], + "credit_map": [ + 0.015225418843328953, + 0.0232617799192667, + 0.02183057740330696, + 0.04126856476068497, + 0.015050327405333519, + 0.01340175699442625, + 0.03754344955086708, + 0.038127351552248, + 0.0036856099031865597, + 0.009985014796257019, + 0.006671413779258728, + 0.018701765686273575, + 0.005814183037728071, + 0.004870110657066107, + 0.0206600334495306, + 0.010967842303216457, + 0.008687361143529415, + 0.02063157968223095, + 0.0077357422560453415, + 0.030220327898859978, + 0.006066019181162119, + 0.006327561568468809, + 0.019236570224165916, + 0.01849406398832798, + 0.0045500872656702995, + 0.008974381722509861, + 0.0064275916665792465, + 0.01393065508455038, + 0.0029101541731506586, + 0.003310814267024398, + 0.006694435607641935, + 0.009305935353040695, + 0.011533972807228565, + 0.0189644955098629, + 0.01749567687511444, + 0.02522529475390911, + 0.006330746691673994, + 0.008481700904667377, + 0.011935747228562832, + 0.008238071575760841, + 0.006228860002011061, + 0.00735276285558939, + 0.010729659348726273, + 0.010035837069153786, + 0.0037941792979836464, + 0.0037816010881215334, + 0.008836367167532444, + 0.004102497361600399, + 0.013405350968241692, + 0.01074149552732706, + 0.012312373146414757, + 0.009553446434438229, + 0.006789826322346926, + 0.007441668771207333, + 0.00886633899062872, + 0.007237665820866823, + 0.009498059749603271, + 0.006693450268357992, + 0.012658314779400826, + 0.009460978209972382, + 0.0029928849544376135, + 0.003162817796692252, + 0.005986730568110943, + 0.006196426693350077, + 0.016085904091596603, + 0.007777475751936436, + 0.02201969176530838, + 0.0068665058352053165, + 0.006009350996464491, + 0.007120213937014341, + 0.007184592541307211, + 0.00522005558013916, + 0.026424935087561607, + 0.012846517376601696, + 0.034507568925619125, + 0.01348587404936552, + 0.015458475798368454, + 0.021016800776124, + 0.020388394594192505, + 0.020953884348273277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0381014309823513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549311637878418 + ], + "opposite_logprob_tokens": [ + -2.929931163787842 + ], + "logprob_reward": -0.0549311637878418, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549311637878418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1511000692844391, + 0.04137120023369789, + 0.08098021149635315, + 0.08986571431159973, + 0.2243480682373047, + 0.06997185200452805, + 0.09372828900814056, + 0.05638692155480385, + 0.11315479129552841, + 0.07909289002418518 + ], + "credit_map": [ + 0.012259937822818756, + 0.012552052736282349, + 0.014247922226786613, + 0.02872191183269024, + 0.010954754427075386, + 0.009397393092513084, + 0.029921365901827812, + 0.0330447293817997, + 0.0027066508773714304, + 0.00656849006190896, + 0.003327412297949195, + 0.009172536432743073, + 0.0028064469806849957, + 0.0023637493140995502, + 0.007665389217436314, + 0.0067605250515043736, + 0.00798586755990982, + 0.012366021983325481, + 0.0071441782638430595, + 0.02045384608209133, + 0.004060723818838596, + 0.004399176687002182, + 0.01341167464852333, + 0.011158722452819347, + 0.007025706581771374, + 0.01310896035283804, + 0.01732514426112175, + 0.017637578770518303, + 0.005438222084194422, + 0.008967412635684013, + 0.01117828767746687, + 0.009184406138956547, + 0.022367972880601883, + 0.03925369679927826, + 0.04365847632288933, + 0.04061194509267807, + 0.012793060392141342, + 0.021858971565961838, + 0.02285842038691044, + 0.02094552479684353, + 0.007529573980718851, + 0.010074225254356861, + 0.012427899986505508, + 0.013550064526498318, + 0.006105845328420401, + 0.004569569602608681, + 0.010081466287374496, + 0.005633207503706217, + 0.012809254229068756, + 0.011692477390170097, + 0.01577615924179554, + 0.014311281032860279, + 0.007749737706035376, + 0.006117986980825663, + 0.017564183101058006, + 0.007707206066697836, + 0.008545045740902424, + 0.005087374709546566, + 0.012191488407552242, + 0.0063592661172151566, + 0.00641571544110775, + 0.005708710290491581, + 0.00752755394205451, + 0.004551767371594906, + 0.017854899168014526, + 0.009205974638462067, + 0.02742319367825985, + 0.011120442301034927, + 0.011606111191213131, + 0.013445301912724972, + 0.013617160730063915, + 0.008881720714271069, + 0.012383566237986088, + 0.005731528624892235, + 0.017570123076438904, + 0.007296077907085419, + 0.008621633984148502, + 0.007656126283109188, + 0.008683419786393642, + 0.011150416918098927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0549311637878418 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028196123894304037 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.0028196123894304037, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028196123894304037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24734798073768616, + 0.07836975902318954, + 0.11956918239593506, + 0.07208018749952316, + 0.05933825299143791, + 0.07206007093191147, + 0.10833366960287094, + 0.10162980109453201, + 0.07507980614900589, + 0.06619121134281158 + ], + "credit_map": [ + 0.02609594538807869, + 0.03258476033806801, + 0.033562708646059036, + 0.047354284673929214, + 0.016700435429811478, + 0.01353435032069683, + 0.03750570863485336, + 0.040009792894124985, + 0.005100478418171406, + 0.013545749709010124, + 0.007171994540840387, + 0.019434966146945953, + 0.005313209258019924, + 0.005507944151759148, + 0.014059088192880154, + 0.008236335590481758, + 0.0070423525758087635, + 0.02233489789068699, + 0.008480099961161613, + 0.034724488854408264, + 0.005551349371671677, + 0.005080406554043293, + 0.017656801268458366, + 0.018698787316679955, + 0.005837165750563145, + 0.014068106189370155, + 0.00872684083878994, + 0.023132234811782837, + 0.003134559839963913, + 0.0032920814119279385, + 0.008178863674402237, + 0.0057103377766907215, + 0.007042530458420515, + 0.010972624644637108, + 0.009246326051652431, + 0.013592522591352463, + 0.0025531647261232138, + 0.005743773654103279, + 0.006358284503221512, + 0.003829029155895114, + 0.009135475382208824, + 0.013002332299947739, + 0.012046225368976593, + 0.009838552214205265, + 0.0041200751438736916, + 0.007882351987063885, + 0.010399102233350277, + 0.00563596049323678, + 0.019765911623835564, + 0.014514310285449028, + 0.03493816778063774, + 0.012919699773192406, + 0.005041711498051882, + 0.0058314744383096695, + 0.008590511977672577, + 0.006731887347996235, + 0.02237425558269024, + 0.01744144968688488, + 0.023863939568400383, + 0.011152110993862152, + 0.005006674211472273, + 0.0064164926297962666, + 0.007925869897007942, + 0.007449006196111441, + 0.014589134603738785, + 0.006248628254979849, + 0.021725505590438843, + 0.006317149382084608, + 0.005418570712208748, + 0.006777408067137003, + 0.007696847897022963, + 0.006306559778749943, + 0.009653739631175995, + 0.006371912080794573, + 0.011987864039838314, + 0.006627530790865421, + 0.007752937730401754, + 0.009034231305122375, + 0.008096742443740368, + 0.006666252855211496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0028196123894304037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This extreme close-up, rendered in mesmerizing slow motion, begins by tracking a gleaming, polished chrome sphere as it makes contact with a dark, richly textured wooden surface. The camera, perfectly synchronized with the action, precisely tracks the sphere's movement. The sphere itself is impeccably smooth, its mirror-like surface reflecting a distorted, elongated image of a single, bright studio softbox light against a dark, out-of-focus background. It bounces off a dark, matte-finished wooden board, its deep, pronounced texture creating a rich, organic contrast to the sphere's industrial perfection. As the sphere silently and deliberately lifts off the surface, it begins a slow, majestic rotation, the camera sliding laterally to track its graceful arc through the air. The scene is bathed in high-contrast light, with the single light source creating brilliant white highlights along the sphere's curve and casting deep shadows that accentuate the wood's grain, creating a cool, minimalist color palette of silver and near-black. The overall atmosphere is one of serene focus and hypnotic elegance, a pure visual study of form, texture, and movement.", + "seed": 21084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_5/index_720_seed_21084_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.038936767750419675, + "vlm_logprob": -0.038936767750419675, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.038936767750419675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012896348722279072 + ], + "opposite_logprob_tokens": [ + -5.262896537780762 + ], + "logprob_reward": -0.012896348722279072, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.012896348722279072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3501674234867096, + 0.06102379783987999, + 0.09681075811386108, + 0.10623714327812195, + 0.13371045887470245, + 0.06691932678222656, + 0.05235883593559265, + 0.05327506363391876, + 0.03376057744026184, + 0.04573654755949974 + ], + "credit_map": [ + 0.02739621140062809, + 0.015051078051328659, + 0.045899491757154465, + 0.019379720091819763, + 0.014092405326664448, + 0.011616849340498447, + 0.11135022342205048, + 0.10538142919540405, + 0.0063216788694262505, + 0.009637119248509407, + 0.011501152068376541, + 0.006558704189956188, + 0.0044351308606565, + 0.008263131603598595, + 0.005299716256558895, + 0.009007168002426624, + 0.012714509852230549, + 0.013824586756527424, + 0.02111688256263733, + 0.009998091496527195, + 0.009776839055120945, + 0.010061830282211304, + 0.008256792090833187, + 0.011061226949095726, + 0.012474678456783295, + 0.012906035408377647, + 0.015196433290839195, + 0.013759251683950424, + 0.017154080793261528, + 0.01150540355592966, + 0.009986937046051025, + 0.01325433049350977, + 0.01810028776526451, + 0.01990480348467827, + 0.017746740952134132, + 0.024398118257522583, + 0.01603500358760357, + 0.011463596485555172, + 0.014179632999002934, + 0.01188227441161871, + 0.012900163419544697, + 0.005596920847892761, + 0.013362755998969078, + 0.016498418524861336, + 0.0041471682488918304, + 0.004948233254253864, + 0.004119606222957373, + 0.00534606259316206, + 0.008906696923077106, + 0.0043451013043522835, + 0.008662787266075611, + 0.01226808875799179, + 0.004270809702575207, + 0.0049850428476929665, + 0.0037740832194685936, + 0.005146225448697805, + 0.009440486319363117, + 0.007268197368830442, + 0.005215119570493698, + 0.007036185823380947, + 0.005988063290715218, + 0.007017130963504314, + 0.0040991720743477345, + 0.007210709620267153, + 0.004113639239221811, + 0.0031284596771001816, + 0.003953218460083008, + 0.007802736014127731, + 0.00515971053391695, + 0.002583627589046955, + 0.00241539697162807, + 0.0046037910506129265, + 0.00645560072734952, + 0.006010384298861027, + 0.00810222327709198, + 0.0056162946857512, + 0.004355519078671932, + 0.004383736290037632, + 0.006114739924669266, + 0.0046980478800833225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.012896348722279072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905113488435745 + ], + "opposite_logprob_tokens": [ + -2.4640512466430664 + ], + "logprob_reward": -0.08905113488435745, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905113488435745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21320851147174835, + 0.20571333169937134, + 0.031791891902685165, + 0.047270406037569046, + 0.26163041591644287, + 0.06452099978923798, + 0.03850672394037247, + 0.03487143665552139, + 0.039071522653102875, + 0.06341477483510971 + ], + "credit_map": [ + 0.029369352385401726, + 0.01871586963534355, + 0.06208430603146553, + 0.02591799944639206, + 0.015468711033463478, + 0.019564928486943245, + 0.01662568934261799, + 0.025461647659540176, + 0.02446061000227928, + 0.03876858949661255, + 0.04358292371034622, + 0.02279985137283802, + 0.019723504781723022, + 0.013201367110013962, + 0.017572836950421333, + 0.025603627786040306, + 0.004189422819763422, + 0.0030715768225491047, + 0.009597476571798325, + 0.003579753451049328, + 0.0025028816889971495, + 0.002538848901167512, + 0.0032394560985267162, + 0.00307247182354331, + 0.00521146459504962, + 0.005777993239462376, + 0.007971880957484245, + 0.005798202008008957, + 0.008066695183515549, + 0.004080853424966335, + 0.005569073837250471, + 0.004794238600879908, + 0.042227957397699356, + 0.02839663252234459, + 0.03528027608990669, + 0.07292209565639496, + 0.021127738058567047, + 0.013654915615916252, + 0.01842690072953701, + 0.029593899846076965, + 0.011935481801629066, + 0.005905542988330126, + 0.011172444559633732, + 0.015201002359390259, + 0.004669221118092537, + 0.005287792533636093, + 0.004738243296742439, + 0.005611262284219265, + 0.006306395400315523, + 0.005924316588789225, + 0.004978686571121216, + 0.005806558299809694, + 0.004420096054673195, + 0.005531237926334143, + 0.002496884437277913, + 0.0030425472650676966, + 0.006532005034387112, + 0.0026301464531570673, + 0.004045591223984957, + 0.008781892247498035, + 0.002159333322197199, + 0.004306163173168898, + 0.002467459999024868, + 0.00394884217530489, + 0.00522261019796133, + 0.005272087175399065, + 0.005732789169996977, + 0.007312783505767584, + 0.003695915685966611, + 0.0026738515589386225, + 0.0030251697171479464, + 0.006136308889836073, + 0.008822102099657059, + 0.0041520376689732075, + 0.011294438503682613, + 0.012173379771411419, + 0.005195521749556065, + 0.006982152815908194, + 0.007420816924422979, + 0.007374322973191738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.08905113488435745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063629418611526 + ], + "opposite_logprob_tokens": [ + -3.1680636405944824 + ], + "logprob_reward": -0.043063629418611526, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063629418611526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1284627467393875, + 0.19184115529060364, + 0.04564530402421951, + 0.061078086495399475, + 0.08535873144865036, + 0.04879344627261162, + 0.2865836024284363, + 0.03116811439394951, + 0.07804061472415924, + 0.04302818700671196 + ], + "credit_map": [ + 0.019124597311019897, + 0.012366055510938168, + 0.03275832161307335, + 0.01949225924909115, + 0.010441038757562637, + 0.011054840870201588, + 0.011757578700780869, + 0.01146805752068758, + 0.023540010675787926, + 0.036755867302417755, + 0.037249818444252014, + 0.019954262301325798, + 0.019140304997563362, + 0.012876813299953938, + 0.016519363969564438, + 0.025804724544286728, + 0.005453723482787609, + 0.006426660809665918, + 0.009681565687060356, + 0.004667090252041817, + 0.004057649057358503, + 0.005007653031498194, + 0.00494973361492157, + 0.005401231814175844, + 0.007746023591607809, + 0.008814745582640171, + 0.007175824139267206, + 0.0076407091692090034, + 0.010607597418129444, + 0.005403412040323019, + 0.005962650291621685, + 0.007727127522230148, + 0.011970479972660542, + 0.010667743161320686, + 0.008907238021492958, + 0.012303562834858894, + 0.013912023976445198, + 0.010185345076024532, + 0.009559099562466145, + 0.007853247225284576, + 0.008295328356325626, + 0.004147413186728954, + 0.00770283117890358, + 0.012918169610202312, + 0.004360696766525507, + 0.003491251729428768, + 0.0036564706824719906, + 0.004221288021653891, + 0.04659182205796242, + 0.03147440403699875, + 0.048047132790088654, + 0.06416422128677368, + 0.025246160104870796, + 0.03640181943774223, + 0.01602509617805481, + 0.01863294467329979, + 0.005163626745343208, + 0.0036652078852057457, + 0.0032712300308048725, + 0.004062881227582693, + 0.0030601557809859514, + 0.005415380001068115, + 0.003131774254143238, + 0.003397861262783408, + 0.008615237660706043, + 0.009303565137088299, + 0.008308998309075832, + 0.013035457581281662, + 0.015249660238623619, + 0.006757568102329969, + 0.006805690936744213, + 0.009964440949261189, + 0.006090547889471054, + 0.00590897724032402, + 0.007485297042876482, + 0.0062995837070047855, + 0.004277538042515516, + 0.004054156597703695, + 0.0036854890640825033, + 0.005226601846516132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.043063629418611526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048629116266965866 + ], + "opposite_logprob_tokens": [ + -3.04862904548645 + ], + "logprob_reward": -0.048629116266965866, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048629116266965866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19392640888690948, + 0.07689019292593002, + 0.04531880095601082, + 0.15191805362701416, + 0.12081646174192429, + 0.04049070551991463, + 0.05253220722079277, + 0.14369602501392365, + 0.08731639385223389, + 0.0870947614312172 + ], + "credit_map": [ + 0.01894492283463478, + 0.012435948476195335, + 0.028927519917488098, + 0.015241073444485664, + 0.01035465206950903, + 0.012530440464615822, + 0.04660509154200554, + 0.04888675734400749, + 0.011199034750461578, + 0.01272808201611042, + 0.01259175967425108, + 0.006662245839834213, + 0.007051659282296896, + 0.00595570495352149, + 0.008896110579371452, + 0.011805590242147446, + 0.005200336221605539, + 0.00527587067335844, + 0.00950810220092535, + 0.0049708252772688866, + 0.00330512598156929, + 0.005994816310703754, + 0.0048820860683918, + 0.006181634031236172, + 0.01764565147459507, + 0.021013347432017326, + 0.020970366895198822, + 0.01979878544807434, + 0.025217559188604355, + 0.014119588769972324, + 0.015834243968129158, + 0.017318502068519592, + 0.01739872433245182, + 0.013752244412899017, + 0.011295941658318043, + 0.017877008765935898, + 0.018613316118717194, + 0.016021480783820152, + 0.01443394087255001, + 0.011423787102103233, + 0.005692680366337299, + 0.00428707106038928, + 0.0064459689892828465, + 0.00609176279976964, + 0.004652842879295349, + 0.004986650310456753, + 0.004242343828082085, + 0.004091382492333651, + 0.009839601814746857, + 0.005143909715116024, + 0.007244921755045652, + 0.008768419735133648, + 0.005311732180416584, + 0.005337308626621962, + 0.004417724441736937, + 0.006468586158007383, + 0.02825869992375374, + 0.013890959322452545, + 0.023932455107569695, + 0.0294529739767313, + 0.00862402655184269, + 0.012842689640820026, + 0.010756355710327625, + 0.01593785360455513, + 0.012113725766539574, + 0.013318185694515705, + 0.013102098368108273, + 0.010064374655485153, + 0.010644710622727871, + 0.005500056780874729, + 0.007773504592478275, + 0.014799731783568859, + 0.012211957015097141, + 0.011895356699824333, + 0.02102329023182392, + 0.010961011983454227, + 0.0059608956798911095, + 0.008230101317167282, + 0.007485536392778158, + 0.009326613508164883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.048629116266965866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming, polished chrome sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere bounce off a dark, matte-finished wooden board with visible grain texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sphere's mirror-like surface reflect a distorted, elongated image of a single bright studio softbox light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track the sphere's movement laterally as it lifts off and rotates through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out-of-focus, creating a high-contrast scene with a cool, minimalist color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010436094598844647 + ], + "opposite_logprob_tokens": [ + -6.876043796539307 + ], + "logprob_reward": -0.0010436094598844647, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010436094598844647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21476790308952332, + 0.11527788639068604, + 0.03580455854535103, + 0.03854174539446831, + 0.08419448882341385, + 0.11230190843343735, + 0.09572374820709229, + 0.2044152021408081, + 0.020214306190609932, + 0.07875830680131912 + ], + "credit_map": [ + 0.025040796026587486, + 0.013046777807176113, + 0.03590800613164902, + 0.016578227281570435, + 0.009590125642716885, + 0.016486236825585365, + 0.04936403036117554, + 0.048753682523965836, + 0.01590576581656933, + 0.0228765569627285, + 0.019557800143957138, + 0.011955645866692066, + 0.011401193216443062, + 0.007475362624973059, + 0.009298033080995083, + 0.016807520762085915, + 0.004947207402437925, + 0.0034896263387054205, + 0.009280399419367313, + 0.004409965593367815, + 0.0022929953411221504, + 0.004681837745010853, + 0.0028803327586501837, + 0.0038221878930926323, + 0.004556915257126093, + 0.005989708472043276, + 0.008225994184613228, + 0.006537335459142923, + 0.0039275470189750195, + 0.003351310035213828, + 0.0030739353969693184, + 0.0028789942152798176, + 0.013108464889228344, + 0.009987994097173214, + 0.012171744368970394, + 0.01860685646533966, + 0.009321262128651142, + 0.0065386537462472916, + 0.0061085219494998455, + 0.008350993506610394, + 0.020692914724349976, + 0.011983765289187431, + 0.02148929424583912, + 0.024305805563926697, + 0.007966061122715473, + 0.008592227473855019, + 0.007981077767908573, + 0.009290753863751888, + 0.015976540744304657, + 0.013411028310656548, + 0.015478426590561867, + 0.018460961058735847, + 0.010973332449793816, + 0.009677729569375515, + 0.006067581474781036, + 0.0056781466118991375, + 0.03979343920946121, + 0.028027521446347237, + 0.013784551993012428, + 0.03633178398013115, + 0.016018133610486984, + 0.021816983819007874, + 0.016954325139522552, + 0.03168843314051628, + 0.0025950211565941572, + 0.0028476878069341183, + 0.0025995182804763317, + 0.003104276955127716, + 0.003239037236198783, + 0.00133599154651165, + 0.0021331210155040026, + 0.002359650796279311, + 0.011805596761405468, + 0.0139104463160038, + 0.013442190364003181, + 0.007514898665249348, + 0.006904904730618, + 0.007526909466832876, + 0.005554745439440012, + 0.012098605744540691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0010436094598844647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_tensor([8579, 8579])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_tensor([8579, 8579])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8cb67acbb98d3bcb5191aada96d62cd96a3f8ccc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_tensor([8579, 8579])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a slightly elevated angle, slowly zooms in, focusing on the sharply defined, solemn face of the divine general Yang Jian. He wears an intricately carved three-mountain flying phoenix hat, its tassel woven with gold and silver thread swaying gently in the wind. He wears a suit of golden chain mail armor engraved with the patterns of ancient mythical beasts, its plates reflecting a cool metallic luster in the dim light. Against the backdrop, dark storm clouds swirl, occasionally punctuated by silent flashes of lightning, creating an oppressive, impending sense of impending storm. As the camera draws closer, the tightly closed vertical divine pattern at the center of his forehead begins to tremble slightly, then calmly and majestically splits to the sides, revealing a pupilless divine eye composed entirely of radiant golden energy. In an instant, a dazzling golden divine light composed of pure energy poured down like molten gold. Countless tiny shimmering rune particles were mixed in the beam, illuminating the lower half of his face and the armor on his chest, forming a strong contrast between light and shadow. The whole picture was filled with a sacred, majestic and awe-inspiring epic atmosphere.", + "seed": 8901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_8579_seed_8901_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03058717679232359, + "vlm_logprob": -0.03058717679232359, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03058717679232359, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013936114497482777 + ], + "opposite_logprob_tokens": [ + -4.5139360427856445 + ], + "logprob_reward": -0.013936114497482777, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013936114497482777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32815584540367126, + 0.05822928622364998, + 0.04041922837495804, + 0.027049927040934563, + 0.0426352322101593, + 0.05765046179294586, + 0.044539373368024826, + 0.11178038269281387, + 0.14348900318145752, + 0.146051287651062 + ], + "credit_map": [ + 0.060687776654958725, + 0.06944676488637924, + 0.043874043971300125, + 0.017086250707507133, + 0.03607112914323807, + 0.051392536610364914, + 0.027991561219096184, + 0.02160579152405262, + 0.007842818275094032, + 0.01193300262093544, + 0.006300392560660839, + 0.0038955037016421556, + 0.009062755852937698, + 0.008412104099988937, + 0.006353398784995079, + 0.004429311957210302, + 0.004449145402759314, + 0.0053162965923547745, + 0.004293191712349653, + 0.003519552992656827, + 0.006532417144626379, + 0.006369922775775194, + 0.006237523630261421, + 0.003701176028698683, + 0.00601537711918354, + 0.004304281901568174, + 0.002878772094845772, + 0.001981563400477171, + 0.0025201975367963314, + 0.0040858895517885685, + 0.0028947386890649796, + 0.0023691074457019567, + 0.004756529815495014, + 0.003847541054710746, + 0.003762199077755213, + 0.004835131578147411, + 0.006218901835381985, + 0.006175104528665543, + 0.007429816294461489, + 0.005610007327049971, + 0.008318626321852207, + 0.005725234746932983, + 0.004903040360659361, + 0.006631493102759123, + 0.008034203201532364, + 0.00797676295042038, + 0.009077565744519234, + 0.006983535829931498, + 0.005154616665095091, + 0.004983867518603802, + 0.005546438042074442, + 0.00421159015968442, + 0.008907767944037914, + 0.005410533398389816, + 0.005731359589844942, + 0.004593199584633112, + 0.013702789321541786, + 0.01970144361257553, + 0.008459224365651608, + 0.0104255685582757, + 0.019787205383181572, + 0.012479919008910656, + 0.021406516432762146, + 0.005817718803882599, + 0.015834948047995567, + 0.012294196523725986, + 0.015099987387657166, + 0.013377033174037933, + 0.03384312614798546, + 0.013290793634951115, + 0.02879696898162365, + 0.010951952077448368, + 0.019412396475672722, + 0.030074123293161392, + 0.02560589648783207, + 0.014174551703035831, + 0.012974759563803673, + 0.01721639186143875, + 0.014102606102824211, + 0.012490561231970787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.013936114497482777 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903989404439926 + ], + "opposite_logprob_tokens": [ + -2.5790398120880127 + ], + "logprob_reward": -0.07903989404439926, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903989404439926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26272571086883545, + 0.25821056962013245, + 0.1084480881690979, + 0.04864426329731941, + 0.03543541580438614, + 0.029116658493876457, + 0.03169191628694534, + 0.07197852432727814, + 0.04494307562708855, + 0.10880585014820099 + ], + "credit_map": [ + 0.047522421926259995, + 0.046228472143411636, + 0.0356028713285923, + 0.015340184792876244, + 0.032510653138160706, + 0.04340901970863342, + 0.02653949335217476, + 0.015572564676404, + 0.040853146463632584, + 0.022418593987822533, + 0.0425737202167511, + 0.020824292674660683, + 0.04088485240936279, + 0.04056644067168236, + 0.030445128679275513, + 0.0196443572640419, + 0.012552514672279358, + 0.010342083871364594, + 0.011223684065043926, + 0.01052696444094181, + 0.01652652584016323, + 0.015401075594127178, + 0.017507603392004967, + 0.014367621392011642, + 0.006586690898984671, + 0.00433752965182066, + 0.005295604467391968, + 0.005325501784682274, + 0.007600471843034029, + 0.006680133752524853, + 0.007439037784934044, + 0.0053792898543179035, + 0.0033720831852406263, + 0.003937950357794762, + 0.005353097338229418, + 0.0036800962407141924, + 0.0048341332003474236, + 0.005316159222275019, + 0.0036692216526716948, + 0.0052726673893630505, + 0.003495842218399048, + 0.0025124484673142433, + 0.003396959975361824, + 0.0036267582327127457, + 0.004598597530275583, + 0.00410982221364975, + 0.00340070272795856, + 0.0039755236357450485, + 0.004563373979181051, + 0.008407406508922577, + 0.004071086645126343, + 0.0021763888653367758, + 0.003570804838091135, + 0.003565054852515459, + 0.002811916172504425, + 0.002525879768654704, + 0.0077190580777823925, + 0.009839016944169998, + 0.00707687484100461, + 0.010327211581170559, + 0.007144805509597063, + 0.007468667812645435, + 0.01052854023873806, + 0.011874335817992687, + 0.005021261051297188, + 0.004185146186500788, + 0.0054766274988651276, + 0.004101826809346676, + 0.010006152093410492, + 0.0044490983709692955, + 0.00670575350522995, + 0.004997204057872295, + 0.013595357537269592, + 0.017835838720202446, + 0.017063606530427933, + 0.01091906800866127, + 0.016039017587900162, + 0.013102472759783268, + 0.01324829738587141, + 0.007002180442214012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07903989404439926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290986195206642 + ], + "opposite_logprob_tokens": [ + -3.7732908725738525 + ], + "logprob_reward": -0.023290986195206642, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290986195206642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.445321261882782, + 0.11772546917200089, + 0.07630634307861328, + 0.08056751638650894, + 0.019511740654706955, + 0.016185909509658813, + 0.016492566093802452, + 0.031268760561943054, + 0.09140419960021973, + 0.10521620512008667 + ], + "credit_map": [ + 0.09122157841920853, + 0.10425271093845367, + 0.07661621272563934, + 0.020337577909231186, + 0.032617729157209396, + 0.07476628571748734, + 0.027758164331316948, + 0.017750997096300125, + 0.02369719371199608, + 0.026566600427031517, + 0.02398460917174816, + 0.005124546121805906, + 0.006461955141276121, + 0.020454399287700653, + 0.0060325078666210175, + 0.0054036532528698444, + 0.0199835617095232, + 0.01618725061416626, + 0.010395347140729427, + 0.0032149136532098055, + 0.005124191753566265, + 0.014189314097166061, + 0.004400003235787153, + 0.002811760874465108, + 0.00936231017112732, + 0.008767456747591496, + 0.007351900916546583, + 0.007634636480361223, + 0.014340397901833057, + 0.012476760894060135, + 0.012088733725249767, + 0.00854532141238451, + 0.004286475013941526, + 0.002744516124948859, + 0.0023384864907711744, + 0.0017614415846765041, + 0.0019201146205887198, + 0.002768144942820072, + 0.002156919799745083, + 0.0015356414951384068, + 0.0024060169234871864, + 0.0014032181352376938, + 0.0013424924109131098, + 0.001448386232368648, + 0.002192708430811763, + 0.002281400840729475, + 0.00298769841901958, + 0.0021239877678453922, + 0.0021113401744514704, + 0.003521922044456005, + 0.001438513514585793, + 0.001464950735680759, + 0.002499834867194295, + 0.0019155159825459123, + 0.0020442879758775234, + 0.0014961998676881194, + 0.005073924548923969, + 0.005188257433474064, + 0.0030180378817021847, + 0.0038067300338298082, + 0.0030647728126496077, + 0.003761477302759886, + 0.0033692168071866035, + 0.0039863428100943565, + 0.0100641418248415, + 0.009331196546554565, + 0.0142580047249794, + 0.011606145650148392, + 0.011481636203825474, + 0.009552367962896824, + 0.01383073441684246, + 0.011279969476163387, + 0.015336398966610432, + 0.020587176084518433, + 0.014785015024244785, + 0.008075542747974396, + 0.009916386567056179, + 0.01854611560702324, + 0.008027329109609127, + 0.009942242875695229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.023290986195206642 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060122601687908 + ], + "opposite_logprob_tokens": [ + -3.895601272583008 + ], + "logprob_reward": -0.02060122601687908, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060122601687908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19539397954940796, + 0.06519344449043274, + 0.057456374168395996, + 0.062373917549848557, + 0.05548414960503578, + 0.037537384778261185, + 0.049704939126968384, + 0.20037822425365448, + 0.10208341479301453, + 0.17439408600330353 + ], + "credit_map": [ + 0.03859422355890274, + 0.03866329789161682, + 0.02170838974416256, + 0.011731699109077454, + 0.02200891077518463, + 0.029026983305811882, + 0.02012474276125431, + 0.013535736128687859, + 0.0066850692965090275, + 0.008408546447753906, + 0.011312957853078842, + 0.004373413044959307, + 0.010760808363556862, + 0.008951169438660145, + 0.00901755690574646, + 0.005683918483555317, + 0.0052841962315142155, + 0.005277838557958603, + 0.005103191360831261, + 0.005164030939340591, + 0.01157332956790924, + 0.009337483905255795, + 0.010078483261168003, + 0.00563782500103116, + 0.007130806799978018, + 0.005657603032886982, + 0.006662905216217041, + 0.006230648141354322, + 0.010400529950857162, + 0.00860332977026701, + 0.011841067112982273, + 0.005847025662660599, + 0.007741679437458515, + 0.004919394850730896, + 0.005667292978614569, + 0.005595702677965164, + 0.009152489714324474, + 0.008523988537490368, + 0.009402249939739704, + 0.004481349140405655, + 0.005595611408352852, + 0.0036945093888789415, + 0.0033433260396122932, + 0.0032061173114925623, + 0.006309018470346928, + 0.005910933483392, + 0.005377631168812513, + 0.00410023657605052, + 0.004873418249189854, + 0.005690353456884623, + 0.0044355690479278564, + 0.007349222898483276, + 0.007706578820943832, + 0.006210370920598507, + 0.00479104183614254, + 0.008648382499814034, + 0.02603018283843994, + 0.03053768165409565, + 0.018040012568235397, + 0.026449045166373253, + 0.024694429710507393, + 0.02208235114812851, + 0.0268999096006155, + 0.025644615292549133, + 0.021631982177495956, + 0.011817234568297863, + 0.01300005242228508, + 0.010556534864008427, + 0.01106312032788992, + 0.010475436225533485, + 0.010553733445703983, + 0.012985319830477238, + 0.023882487788796425, + 0.039941538125276566, + 0.024310065433382988, + 0.0136414784938097, + 0.014823108911514282, + 0.02707727812230587, + 0.015750493854284286, + 0.014967632479965687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.02060122601687908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067663207650185 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067663207650185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067663207650185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14049798250198364, + 0.07368611544370651, + 0.06435874104499817, + 0.07287099212408066, + 0.03564855083823204, + 0.044166047126054764, + 0.035207245498895645, + 0.16481468081474304, + 0.19424985349178314, + 0.1744997799396515 + ], + "credit_map": [ + 0.021868668496608734, + 0.03279489278793335, + 0.017246991395950317, + 0.00862925499677658, + 0.01406753808259964, + 0.020742293447256088, + 0.015254702419042587, + 0.009893646463751793, + 0.01551559567451477, + 0.014959452673792839, + 0.009393089450895786, + 0.003762464039027691, + 0.006625197362154722, + 0.012854434549808502, + 0.006775380112230778, + 0.0038005083333700895, + 0.009298007935285568, + 0.007785229943692684, + 0.006765883415937424, + 0.004429098684340715, + 0.012981386855244637, + 0.009763741865754128, + 0.0092474315315485, + 0.004087971057742834, + 0.007567146327346563, + 0.008395124226808548, + 0.005606526974588633, + 0.007300980854779482, + 0.01444418728351593, + 0.010294830426573753, + 0.012014943175017834, + 0.007247250992804766, + 0.004915621597319841, + 0.0041304645128548145, + 0.0033616831060498953, + 0.003450426273047924, + 0.005650964565575123, + 0.004950040951371193, + 0.0057703908532857895, + 0.003418960142880678, + 0.006076427176594734, + 0.004442148841917515, + 0.0045692771673202515, + 0.0034919476602226496, + 0.007971107959747314, + 0.006454537156969309, + 0.0070395623333752155, + 0.004121046978980303, + 0.0057550715282559395, + 0.0069260490126907825, + 0.0034258279483765364, + 0.0031896925065666437, + 0.0037677863147109747, + 0.00415030075237155, + 0.003793190699070692, + 0.004199329297989607, + 0.015757376328110695, + 0.012436030432581902, + 0.015876026824116707, + 0.022426309064030647, + 0.023148542270064354, + 0.018574725836515427, + 0.032797034829854965, + 0.023798637092113495, + 0.03272980824112892, + 0.020796842873096466, + 0.030818911269307137, + 0.022642306983470917, + 0.0220437403768301, + 0.0194693673402071, + 0.01749931462109089, + 0.028249574825167656, + 0.02797733247280121, + 0.023613879457116127, + 0.026356324553489685, + 0.019548773765563965, + 0.017667969688773155, + 0.02960870787501335, + 0.015869656577706337, + 0.013857138343155384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.016067663207650185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly elevated angle, slowly zooms in, focusing on the sharply defined, solemn face of the divine general Yang Jian. He wears an intricately carved three-mountain flying phoenix hat, its tassel woven with gold and silver thread swaying gently in the wind. He wears a suit of golden chain mail armor engraved with the patterns of ancient mythical beasts, its plates reflecting a cool metallic luster in the dim light. Against the backdrop, dark storm clouds swirl, occasionally punctuated by silent flashes of lightning, creating an oppressive, impending sense of impending storm. As the camera draws closer, the tightly closed vertical divine pattern at the center of his forehead begins to tremble slightly, then calmly and majestically splits to the sides, revealing a pupilless divine eye composed entirely of radiant golden energy. In an instant, a dazzling golden divine light composed of pure energy poured down like molten gold. Countless tiny shimmering rune particles were mixed in the beam, illuminating the lower half of his face and the armor on his chest, forming a strong contrast between light and shadow. The whole picture was filled with a sacred, majestic and awe-inspiring epic atmosphere.", + "seed": 8902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_8579_seed_8902_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024564376659691335, + "vlm_logprob": -0.024564376659691335, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.024564376659691335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012025241740047932 + ], + "opposite_logprob_tokens": [ + -5.012025356292725 + ], + "logprob_reward": -0.012025241740047932, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012025241740047932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16656482219696045, + 0.09027183800935745, + 0.044958338141441345, + 0.044846922159194946, + 0.018604163080453873, + 0.05150012671947479, + 0.18863226473331451, + 0.0906587690114975, + 0.2426041066646576, + 0.06135868281126022 + ], + "credit_map": [ + 0.03465510532259941, + 0.019800452515482903, + 0.021184243261814117, + 0.0161809790879488, + 0.02580058015882969, + 0.013952450826764107, + 0.021823585033416748, + 0.013167411088943481, + 0.013343733735382557, + 0.013050056993961334, + 0.009050260297954082, + 0.007848657667636871, + 0.013343635946512222, + 0.011245052330195904, + 0.014085428789258003, + 0.008305004797875881, + 0.0074796234257519245, + 0.008589448407292366, + 0.0031294808723032475, + 0.004390199203044176, + 0.00971166044473648, + 0.004261794500052929, + 0.004611689131706953, + 0.0027844393625855446, + 0.0070993052795529366, + 0.004030210897326469, + 0.005287460051476955, + 0.005307128187268972, + 0.003936443943530321, + 0.0071677216328680515, + 0.005339916329830885, + 0.006678731180727482, + 0.003195319790393114, + 0.0032547207083553076, + 0.0016354541294276714, + 0.0023260910529643297, + 0.0018127762014046311, + 0.0033072191290557384, + 0.001745332032442093, + 0.0013272492215037346, + 0.007011751178652048, + 0.008262480609118938, + 0.007543797604739666, + 0.00518962275236845, + 0.006141563877463341, + 0.0056337956339120865, + 0.005486735608428717, + 0.00623037526383996, + 0.06763215363025665, + 0.012037746608257294, + 0.007945728488266468, + 0.007006530649960041, + 0.00825583003461361, + 0.06622879207134247, + 0.007630120497196913, + 0.011895358562469482, + 0.00640336936339736, + 0.014683565124869347, + 0.007113367784768343, + 0.01988857425749302, + 0.010843929834663868, + 0.007662286050617695, + 0.013314494863152504, + 0.01074917335063219, + 0.012465517036616802, + 0.03613944724202156, + 0.019997091963887215, + 0.04726914316415787, + 0.051584962755441666, + 0.02050742693245411, + 0.03697255253791809, + 0.017667952924966812, + 0.00807979516685009, + 0.006832062266767025, + 0.0062697757966816425, + 0.010513205081224442, + 0.008844247087836266, + 0.008023466914892197, + 0.006623788271099329, + 0.006172338034957647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.012025241740047932 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700753852725029 + ], + "opposite_logprob_tokens": [ + -2.695075273513794 + ], + "logprob_reward": -0.0700753852725029, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700753852725029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1881738305091858, + 0.15754881501197815, + 0.06486188620328903, + 0.0462106391787529, + 0.02998257614672184, + 0.04405895620584488, + 0.056657642126083374, + 0.0669645443558693, + 0.2704037129878998, + 0.07513740658760071 + ], + "credit_map": [ + 0.04838986322283745, + 0.01332593522965908, + 0.055679742246866226, + 0.013994479551911354, + 0.018364937976002693, + 0.013395940884947777, + 0.015439220704138279, + 0.009583734907209873, + 0.02663416974246502, + 0.016108786687254906, + 0.025496501475572586, + 0.01997981034219265, + 0.019467800855636597, + 0.014428619295358658, + 0.01889742724597454, + 0.016535701230168343, + 0.009058445692062378, + 0.007657444570213556, + 0.007945151999592781, + 0.007092846091836691, + 0.010347569361329079, + 0.005700858775526285, + 0.010030657984316349, + 0.007028911728411913, + 0.007787097245454788, + 0.0067604221403598785, + 0.005495368503034115, + 0.004773409105837345, + 0.005702574737370014, + 0.007949264720082283, + 0.004266029689460993, + 0.0034764776937663555, + 0.004945026710629463, + 0.005343306809663773, + 0.0027636883314698935, + 0.0032553132623434067, + 0.0032875181641429663, + 0.005079600960016251, + 0.002898510778322816, + 0.0024096115957945585, + 0.007360120303928852, + 0.007052903529256582, + 0.0046227662824094296, + 0.004487915895879269, + 0.004058797378093004, + 0.006050372961908579, + 0.005441955290734768, + 0.004984129685908556, + 0.006324294954538345, + 0.0071725500747561455, + 0.005315187852829695, + 0.007871517911553383, + 0.008995544165372849, + 0.005175502505153418, + 0.011091138236224651, + 0.004711908288300037, + 0.006742183119058609, + 0.0053683118894696236, + 0.009592656046152115, + 0.010884950868785381, + 0.006668698973953724, + 0.007664052303880453, + 0.011004040017724037, + 0.00903965625911951, + 0.016078876331448555, + 0.039410967379808426, + 0.017882291227579117, + 0.04762687161564827, + 0.05235988274216652, + 0.03130650520324707, + 0.040328290313482285, + 0.025410043075680733, + 0.006797023583203554, + 0.01213393360376358, + 0.008913105353713036, + 0.013683782890439034, + 0.00995788536965847, + 0.009191703051328659, + 0.009667719714343548, + 0.0047922637313604355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0700753852725029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634747326374054 + ], + "opposite_logprob_tokens": [ + -3.6513473987579346 + ], + "logprob_reward": -0.02634747326374054, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634747326374054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12297342717647552, + 0.12090350687503815, + 0.07927563786506653, + 0.05440138280391693, + 0.044257983565330505, + 0.12332523614168167, + 0.06417351961135864, + 0.07134256511926651, + 0.2141687422990799, + 0.10517798364162445 + ], + "credit_map": [ + 0.017549589276313782, + 0.010066046379506588, + 0.018743528053164482, + 0.016322914510965347, + 0.01411160547286272, + 0.015808535739779472, + 0.01820726878941059, + 0.01216394267976284, + 0.019429968670010567, + 0.016940400004386902, + 0.008054489269852638, + 0.012725361622869968, + 0.017870016396045685, + 0.022926373407244682, + 0.015338100492954254, + 0.007618790026754141, + 0.013197650201618671, + 0.008653094060719013, + 0.00904824584722519, + 0.00772350188344717, + 0.009082316420972347, + 0.01687425747513771, + 0.009089772589504719, + 0.005606800317764282, + 0.009873478673398495, + 0.005541995167732239, + 0.007047695107758045, + 0.006103219464421272, + 0.006469902582466602, + 0.007381036877632141, + 0.004552081692963839, + 0.007431969977915287, + 0.007721461355686188, + 0.0067793941125273705, + 0.004723902326077223, + 0.004387012682855129, + 0.00452579976990819, + 0.008310939185321331, + 0.0039816005155444145, + 0.0038278757128864527, + 0.02054770290851593, + 0.015443803742527962, + 0.020697100088000298, + 0.010959912091493607, + 0.010398360900580883, + 0.015545635484158993, + 0.01135330181568861, + 0.018379420042037964, + 0.017208175733685493, + 0.005718132480978966, + 0.003409796627238393, + 0.009498062543570995, + 0.004783313721418381, + 0.013324097730219364, + 0.006918216589838266, + 0.0033137176651507616, + 0.005513645242899656, + 0.012160557322204113, + 0.008023562841117382, + 0.01233658567070961, + 0.008786384016275406, + 0.006002530921250582, + 0.012331832200288773, + 0.006187468767166138, + 0.015487249940633774, + 0.034442733973264694, + 0.016283109784126282, + 0.029541445896029472, + 0.0564483217895031, + 0.017580829560756683, + 0.03158218413591385, + 0.012802870944142342, + 0.010832478292286396, + 0.013973616994917393, + 0.010408082976937294, + 0.01922062784433365, + 0.015371985733509064, + 0.011824405752122402, + 0.014020539820194244, + 0.009526251815259457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.02634747326374054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067454250529408455 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.0067454250529408455, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067454250529408455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10065249353647232, + 0.025694716721773148, + 0.04287711903452873, + 0.031215570867061615, + 0.03692561760544777, + 0.03539491072297096, + 0.05086654797196388, + 0.10136288404464722, + 0.489108681678772, + 0.08590144664049149 + ], + "credit_map": [ + 0.02747535891830921, + 0.007518914062529802, + 0.02432529255747795, + 0.008384309709072113, + 0.007941237650811672, + 0.00840082112699747, + 0.010631508193910122, + 0.005975039675831795, + 0.0036330455914139748, + 0.0031892494298517704, + 0.00272596743889153, + 0.002836589701473713, + 0.003530002199113369, + 0.004383579362183809, + 0.003223579144105315, + 0.0021727043204009533, + 0.0070355124771595, + 0.004853411577641964, + 0.003910175058990717, + 0.005531315226107836, + 0.006914332043379545, + 0.0054082926362752914, + 0.006287071388214827, + 0.002937006764113903, + 0.005001004319638014, + 0.005531614646315575, + 0.003228825284168124, + 0.003037127899006009, + 0.003965750336647034, + 0.0051709990948438644, + 0.0028491655830293894, + 0.0024310804437845945, + 0.00700958538800478, + 0.0050883824005723, + 0.003282548626884818, + 0.004040892235934734, + 0.0029027170967310667, + 0.007918569259345531, + 0.0035965880379080772, + 0.003086334327235818, + 0.00495755672454834, + 0.005365600343793631, + 0.0045645637437701225, + 0.00419427827000618, + 0.003354791086167097, + 0.004246377851814032, + 0.004414536990225315, + 0.004297202918678522, + 0.005295193288475275, + 0.007695634849369526, + 0.006443076301366091, + 0.009577232412993908, + 0.00522705540060997, + 0.006710176356136799, + 0.00602329196408391, + 0.0038948862347751856, + 0.009057121351361275, + 0.016251258552074432, + 0.010929392650723457, + 0.013925104402005672, + 0.013746144250035286, + 0.013052980415523052, + 0.014597035944461823, + 0.009803839959204197, + 0.030192172154784203, + 0.0686192512512207, + 0.03571147844195366, + 0.09414493292570114, + 0.09289532154798508, + 0.04984437674283981, + 0.06355427950620651, + 0.054146867245435715, + 0.0058027408085763454, + 0.011944531463086605, + 0.009454133920371532, + 0.019200822338461876, + 0.011043424718081951, + 0.005600918550044298, + 0.01526709459722042, + 0.0075877755880355835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0067454250529408455 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628357969224453 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628357969224453, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628357969224453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047038495540618896, + 0.02021607756614685, + 0.025512071326375008, + 0.058739274740219116, + 0.029874520376324654, + 0.02494748868048191, + 0.12548606097698212, + 0.10286298394203186, + 0.5123931169509888, + 0.05292992293834686 + ], + "credit_map": [ + 0.012203821912407875, + 0.003021039767190814, + 0.013635672628879547, + 0.0049669137224555016, + 0.0035570745822042227, + 0.004297907464206219, + 0.002829380566254258, + 0.0025266886223107576, + 0.0033633136190474033, + 0.0022290570195764303, + 0.002030111849308014, + 0.0020590319763869047, + 0.0019963844679296017, + 0.004770405124872923, + 0.0019393868278712034, + 0.0018283870304003358, + 0.004232970532029867, + 0.0024157045409083366, + 0.0029922553803771734, + 0.0031168018467724323, + 0.0032802652567625046, + 0.0030081842560321093, + 0.0040204389952123165, + 0.0024454519152641296, + 0.010048234835267067, + 0.007227090187370777, + 0.007496797479689121, + 0.0060733589343726635, + 0.008448680862784386, + 0.006571261677891016, + 0.0055368016473948956, + 0.007337047718465328, + 0.005965234246104956, + 0.004105953965336084, + 0.0027940012514591217, + 0.003223068080842495, + 0.002923790831118822, + 0.0047104028053581715, + 0.0034360969439148903, + 0.0027159724850207567, + 0.0036528699565678835, + 0.0034502968192100525, + 0.0032370882108807564, + 0.002917014993727207, + 0.003552245907485485, + 0.0028743513394147158, + 0.003120772074908018, + 0.0021428486797958612, + 0.04004211723804474, + 0.008092392235994339, + 0.010812371037900448, + 0.006829814054071903, + 0.005908409599214792, + 0.039207737892866135, + 0.005542777478694916, + 0.009050440043210983, + 0.008418739773333073, + 0.016792824491858482, + 0.01200631819665432, + 0.015312273986637592, + 0.011885338462889194, + 0.01517779566347599, + 0.011140255257487297, + 0.01212944183498621, + 0.026875607669353485, + 0.07760051637887955, + 0.02810991182923317, + 0.08921549469232559, + 0.11575359851121902, + 0.05415543541312218, + 0.0816132053732872, + 0.03906935453414917, + 0.0065924376249313354, + 0.006133439019322395, + 0.004299928899854422, + 0.008805878460407257, + 0.008527317084372044, + 0.006155659910291433, + 0.007251035422086716, + 0.005164223723113537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.007628357969224453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly elevated angle, slowly zooms in, focusing on the sharply defined, solemn face of the divine general Yang Jian. He wears an intricately carved three-mountain flying phoenix hat, its tassel woven with gold and silver thread swaying gently in the wind. He wears a suit of golden chain mail armor engraved with the patterns of ancient mythical beasts, its plates reflecting a cool metallic luster in the dim light. Against the backdrop, dark storm clouds swirl, occasionally punctuated by silent flashes of lightning, creating an oppressive, impending sense of impending storm. As the camera draws closer, the tightly closed vertical divine pattern at the center of his forehead begins to tremble slightly, then calmly and majestically splits to the sides, revealing a pupilless divine eye composed entirely of radiant golden energy. In an instant, a dazzling golden divine light composed of pure energy poured down like molten gold. Countless tiny shimmering rune particles were mixed in the beam, illuminating the lower half of his face and the armor on his chest, forming a strong contrast between light and shadow. The whole picture was filled with a sacred, majestic and awe-inspiring epic atmosphere.", + "seed": 8903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_8579_seed_8903_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023272641561925413, + "vlm_logprob": -0.023272641561925413, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.023272641561925413, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008906636387109756 + ], + "opposite_logprob_tokens": [ + -5.258906841278076 + ], + "logprob_reward": -0.008906636387109756, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008906636387109756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12105001509189606, + 0.08753210306167603, + 0.1543242633342743, + 0.13472743332386017, + 0.10543223470449448, + 0.06253370642662048, + 0.049253035336732864, + 0.05548594519495964, + 0.047976333647966385, + 0.1816849559545517 + ], + "credit_map": [ + 0.018794827163219452, + 0.026885630562901497, + 0.013433204963803291, + 0.009661193937063217, + 0.01685609482228756, + 0.016903014853596687, + 0.009863176383078098, + 0.008652863092720509, + 0.011120669543743134, + 0.01723083108663559, + 0.007054653484374285, + 0.00900261104106903, + 0.015845179557800293, + 0.011424054391682148, + 0.009707004763185978, + 0.006147099658846855, + 0.018181823194026947, + 0.028007591143250465, + 0.008837312459945679, + 0.014594851061701775, + 0.042972151190042496, + 0.01991378329694271, + 0.013481847941875458, + 0.00833488255739212, + 0.009722334332764149, + 0.020490385591983795, + 0.01515919342637062, + 0.017635364085435867, + 0.02496129274368286, + 0.017162853851914406, + 0.01750069111585617, + 0.012095305137336254, + 0.012918205000460148, + 0.011648940853774548, + 0.007974699139595032, + 0.013338940218091011, + 0.01369516272097826, + 0.014785041101276875, + 0.022185068577528, + 0.008886178024113178, + 0.010085150599479675, + 0.010633676312863827, + 0.005777735263109207, + 0.007516989018768072, + 0.007901758886873722, + 0.008792263455688953, + 0.005132264923304319, + 0.006693865638226271, + 0.008749954402446747, + 0.006586542818695307, + 0.007003904785960913, + 0.005739614367485046, + 0.003982635214924812, + 0.004938120953738689, + 0.005708054173737764, + 0.006544203031808138, + 0.005983253009617329, + 0.007252377457916737, + 0.003782553132623434, + 0.007109323050826788, + 0.006764776073396206, + 0.0056022466160357, + 0.014755377545952797, + 0.00423602806404233, + 0.005619899369776249, + 0.0059674265794456005, + 0.004863052163273096, + 0.006176075432449579, + 0.008661486208438873, + 0.004581664223223925, + 0.006434083450585604, + 0.005672641564160585, + 0.019436173141002655, + 0.028567135334014893, + 0.021134771406650543, + 0.024881480261683464, + 0.026493968442082405, + 0.022975707426667213, + 0.02110825479030609, + 0.017087452113628387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.008906636387109756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870961979031563 + ], + "opposite_logprob_tokens": [ + -3.0487096309661865 + ], + "logprob_reward": -0.04870961979031563, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870961979031563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1406274437904358, + 0.07245856523513794, + 0.1382482647895813, + 0.29434987902641296, + 0.12883228063583374, + 0.04421775043010712, + 0.040239837020635605, + 0.022388331592082977, + 0.045558929443359375, + 0.07307872921228409 + ], + "credit_map": [ + 0.016819043084979057, + 0.031532444059848785, + 0.010905874893069267, + 0.011393376626074314, + 0.026095228269696236, + 0.018973050639033318, + 0.019383752718567848, + 0.005524663720279932, + 0.008216753602027893, + 0.011082670651376247, + 0.006308766081929207, + 0.00966318417340517, + 0.012635496445000172, + 0.00913875550031662, + 0.011223314329981804, + 0.0041896202601492405, + 0.012553730979561806, + 0.025358496233820915, + 0.009009365923702717, + 0.01306163240224123, + 0.038854967802762985, + 0.018286801874637604, + 0.013637847267091274, + 0.007485421374440193, + 0.02948138676583767, + 0.03078746795654297, + 0.05880673974752426, + 0.03361186385154724, + 0.036290377378463745, + 0.03643377497792244, + 0.028152737766504288, + 0.0407855398952961, + 0.011175381951034069, + 0.015222717076539993, + 0.01104174368083477, + 0.015000003390014172, + 0.03123515099287033, + 0.016610976308584213, + 0.019304409623146057, + 0.009241893887519836, + 0.006671092938631773, + 0.0049265529960393906, + 0.004024770576506853, + 0.006875544786453247, + 0.005055002402514219, + 0.005699268076568842, + 0.006916257087141275, + 0.0040492648258805275, + 0.006388691253960133, + 0.003975476138293743, + 0.005410367157310247, + 0.004288056865334511, + 0.0036150694359093904, + 0.0042816828936338425, + 0.0038958603981882334, + 0.00838463008403778, + 0.0017713188426569104, + 0.003036228474229574, + 0.0018239081837236881, + 0.003074766369536519, + 0.0026593334041535854, + 0.002765201497823, + 0.005395903717726469, + 0.0018616702873259783, + 0.004065460059791803, + 0.005147745832800865, + 0.006806500721722841, + 0.006120292004197836, + 0.005725912284106016, + 0.004439354408532381, + 0.006547986995428801, + 0.006705680396407843, + 0.011769999749958515, + 0.010643996298313141, + 0.007162058725953102, + 0.010559109970927238, + 0.007572112139314413, + 0.006978537421673536, + 0.012397704645991325, + 0.0059952083975076675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.04870961979031563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190816044807434 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018190816044807434, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190816044807434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0754537284374237, + 0.09830498695373535, + 0.031924448907375336, + 0.07154329121112823, + 0.08779102563858032, + 0.03701091185212135, + 0.06345423310995102, + 0.07113563269376755, + 0.1384093165397644, + 0.3249724507331848 + ], + "credit_map": [ + 0.013614121824502945, + 0.013144068419933319, + 0.006611584220081568, + 0.007697415538132191, + 0.00965285673737526, + 0.011267674155533314, + 0.007466583047062159, + 0.005999423563480377, + 0.024691740050911903, + 0.01946972869336605, + 0.008769371546804905, + 0.009949848987162113, + 0.008816546760499477, + 0.015360471792519093, + 0.006474919617176056, + 0.004772369284182787, + 0.004774775821715593, + 0.004522023722529411, + 0.0034498516470193863, + 0.004583763889968395, + 0.003999789711087942, + 0.004662723746150732, + 0.0031200512312352657, + 0.002811469603329897, + 0.01203039288520813, + 0.008249407634139061, + 0.007076928857713938, + 0.0068605560809373856, + 0.012276304885745049, + 0.008688460104167461, + 0.010052118450403214, + 0.006309119984507561, + 0.018307410180568695, + 0.01186347659677267, + 0.009409858845174313, + 0.008821090683341026, + 0.007870268076658249, + 0.012633731588721275, + 0.006137026473879814, + 0.012748164124786854, + 0.007554500829428434, + 0.004362092353403568, + 0.003485166933387518, + 0.004463379271328449, + 0.005083857104182243, + 0.004594213794916868, + 0.005141331348568201, + 0.002326371381059289, + 0.006973611190915108, + 0.01368406880646944, + 0.0075409626588225365, + 0.0072373258881270885, + 0.007402075920253992, + 0.006537961773574352, + 0.0075853243470191956, + 0.00649290019646287, + 0.004788191523402929, + 0.011539708822965622, + 0.005395815707743168, + 0.008006877265870571, + 0.01170217152684927, + 0.010124873369932175, + 0.014795856550335884, + 0.0047821360640227795, + 0.04131115972995758, + 0.013991933315992355, + 0.0145716005936265, + 0.012806374579668045, + 0.011603484861552715, + 0.02167004533112049, + 0.009120391681790352, + 0.013334322720766068, + 0.08808309584856033, + 0.03000515140593052, + 0.02772701531648636, + 0.03871731460094452, + 0.03125045448541641, + 0.043519143015146255, + 0.03975697234272957, + 0.0259133018553257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.018190816044807434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360012590885162 + ], + "opposite_logprob_tokens": [ + -3.651360034942627 + ], + "logprob_reward": -0.026360012590885162, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360012590885162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08441320806741714, + 0.03830628842115402, + 0.025600995868444443, + 0.18269577622413635, + 0.04994328320026398, + 0.047094885259866714, + 0.05267888680100441, + 0.049439892172813416, + 0.058675896376371384, + 0.41115090250968933 + ], + "credit_map": [ + 0.011101186275482178, + 0.011442596092820168, + 0.0060136341489851475, + 0.009311346337199211, + 0.01536827813833952, + 0.011123796924948692, + 0.015492557547986507, + 0.004559812135994434, + 0.006816804874688387, + 0.005780974868685007, + 0.003755593206733465, + 0.004259687382727861, + 0.005377321504056454, + 0.005205484107136726, + 0.004154770169407129, + 0.0029556492809206247, + 0.003941898699849844, + 0.0030739102512598038, + 0.002717776456847787, + 0.003375378903001547, + 0.0037050386890769005, + 0.0035430651623755693, + 0.0027108772192150354, + 0.0025330507196485996, + 0.0161516685038805, + 0.02506798692047596, + 0.019171850755810738, + 0.020202746614813805, + 0.03990329056978226, + 0.02072545699775219, + 0.024559076875448227, + 0.01691369153559208, + 0.0064543867483735085, + 0.007485078647732735, + 0.004744089674204588, + 0.00620706332847476, + 0.009152431041002274, + 0.007030881475657225, + 0.005378265865147114, + 0.003491086419671774, + 0.007761548273265362, + 0.005407585296779871, + 0.004509196616709232, + 0.009300677105784416, + 0.004789877682924271, + 0.007148525211960077, + 0.004315563943237066, + 0.00386191438883543, + 0.005595585331320763, + 0.006103092338889837, + 0.008642415516078472, + 0.006326495204120874, + 0.004577824380248785, + 0.005689564626663923, + 0.0056069642305374146, + 0.010136947967112064, + 0.009127408266067505, + 0.007430719211697578, + 0.005311505403369665, + 0.0063120704144239426, + 0.005865942221134901, + 0.0045942580327391624, + 0.006150770466774702, + 0.004647219553589821, + 0.006336354650557041, + 0.006530786398798227, + 0.006168042309582233, + 0.007013068068772554, + 0.008444459177553654, + 0.005995968356728554, + 0.011912344954907894, + 0.006274874322116375, + 0.02998211607336998, + 0.06990465521812439, + 0.04097602143883705, + 0.06516870856285095, + 0.04993117228150368, + 0.0400983951985836, + 0.07159555703401566, + 0.043494269251823425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.026360012590885162 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196122996509075 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196122996509075, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196122996509075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07579407095909119, + 0.041754499077796936, + 0.03300941735506058, + 0.0794406607747078, + 0.09544498473405838, + 0.06297522783279419, + 0.07390043884515762, + 0.06909343600273132, + 0.22096700966358185, + 0.24762025475502014 + ], + "credit_map": [ + 0.00962913129478693, + 0.014711882919073105, + 0.00580732012167573, + 0.006377730052918196, + 0.01249905489385128, + 0.010288582183420658, + 0.010289603844285011, + 0.006190769840031862, + 0.007353855296969414, + 0.006228676065802574, + 0.0039039473049342632, + 0.005356006324291229, + 0.005004481412470341, + 0.005346114747226238, + 0.0052438764832913876, + 0.0033175433054566383, + 0.004531260579824448, + 0.005853001028299332, + 0.0031682620756328106, + 0.004118243232369423, + 0.004649683367460966, + 0.004741760436445475, + 0.0033045359887182713, + 0.0026426692493259907, + 0.008687694557011127, + 0.011881659738719463, + 0.009743582457304, + 0.01153294276446104, + 0.00960023608058691, + 0.01124559622257948, + 0.010340305045247078, + 0.006408641580492258, + 0.014317767694592476, + 0.013144029304385185, + 0.008392809890210629, + 0.012443029321730137, + 0.01003258302807808, + 0.014262358658015728, + 0.012461327016353607, + 0.01039107609540224, + 0.009995860978960991, + 0.0065901740454137325, + 0.00557083310559392, + 0.009780745021998882, + 0.005794182885438204, + 0.008614963851869106, + 0.012071521952748299, + 0.004556938540190458, + 0.007318910211324692, + 0.012122818268835545, + 0.007942387834191322, + 0.009795663878321648, + 0.010626173578202724, + 0.008426381275057793, + 0.009793311357498169, + 0.007874792441725731, + 0.01462907437235117, + 0.008070509880781174, + 0.008332843892276287, + 0.009799736551940441, + 0.005467930808663368, + 0.0075584701262414455, + 0.009264307096600533, + 0.005970565136522055, + 0.02526104263961315, + 0.022043751552700996, + 0.02150994911789894, + 0.02929808758199215, + 0.026456940919160843, + 0.022368529811501503, + 0.0424848273396492, + 0.03154385834932327, + 0.017853137105703354, + 0.04742228239774704, + 0.02452235482633114, + 0.04330313205718994, + 0.03097463771700859, + 0.02116928994655609, + 0.040835488587617874, + 0.02153993397951126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.014196122996509075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly elevated angle, slowly zooms in, focusing on the sharply defined, solemn face of the divine general Yang Jian. He wears an intricately carved three-mountain flying phoenix hat, its tassel woven with gold and silver thread swaying gently in the wind. He wears a suit of golden chain mail armor engraved with the patterns of ancient mythical beasts, its plates reflecting a cool metallic luster in the dim light. Against the backdrop, dark storm clouds swirl, occasionally punctuated by silent flashes of lightning, creating an oppressive, impending sense of impending storm. As the camera draws closer, the tightly closed vertical divine pattern at the center of his forehead begins to tremble slightly, then calmly and majestically splits to the sides, revealing a pupilless divine eye composed entirely of radiant golden energy. In an instant, a dazzling golden divine light composed of pure energy poured down like molten gold. Countless tiny shimmering rune particles were mixed in the beam, illuminating the lower half of his face and the armor on his chest, forming a strong contrast between light and shadow. The whole picture was filled with a sacred, majestic and awe-inspiring epic atmosphere.", + "seed": 8904, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_6/index_8579_seed_8904_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025908711925148965, + "vlm_logprob": -0.025908711925148965, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.025908711925148965, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007310071028769016 + ], + "opposite_logprob_tokens": [ + -5.757309913635254 + ], + "logprob_reward": -0.007310071028769016, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007310071028769016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19875842332839966, + 0.06517240405082703, + 0.03979518637061119, + 0.06116262078285217, + 0.04746205359697342, + 0.06107989698648453, + 0.07954917103052139, + 0.1150948703289032, + 0.10767869651317596, + 0.22424668073654175 + ], + "credit_map": [ + 0.03547351434826851, + 0.04742328077554703, + 0.016076641157269478, + 0.018924541771411896, + 0.024307750165462494, + 0.028937818482518196, + 0.014758259057998657, + 0.012856625020503998, + 0.008834991604089737, + 0.0071703302673995495, + 0.006597275845706463, + 0.007669086568057537, + 0.010739385150372982, + 0.007664557080715895, + 0.01088636927306652, + 0.00561041384935379, + 0.00575596746057272, + 0.00600720290094614, + 0.0030520912259817123, + 0.0050060017965734005, + 0.006036348175257444, + 0.00507501931861043, + 0.005072049330919981, + 0.0037905119825154543, + 0.007059507071971893, + 0.011673586443066597, + 0.006374903954565525, + 0.005082168150693178, + 0.009347648359835148, + 0.00891195423901081, + 0.006846386473625898, + 0.0058664679527282715, + 0.006029206793755293, + 0.005680848844349384, + 0.00487487530335784, + 0.005873463582247496, + 0.007966725155711174, + 0.005564545281231403, + 0.007776620332151651, + 0.003695770399644971, + 0.007779616862535477, + 0.010264994576573372, + 0.007114951964467764, + 0.008113491348922253, + 0.006180807948112488, + 0.00814128015190363, + 0.007380400784313679, + 0.006104357540607452, + 0.0068740397691726685, + 0.011510767973959446, + 0.010083467699587345, + 0.008736361749470234, + 0.013515856117010117, + 0.0074803526513278484, + 0.011522727087140083, + 0.009825603105127811, + 0.01139780692756176, + 0.016540413722395897, + 0.01387014426290989, + 0.01449559349566698, + 0.019127806648612022, + 0.010018529370427132, + 0.01617567054927349, + 0.013468914665281773, + 0.008841762319207191, + 0.013893910683691502, + 0.00855286605656147, + 0.014103743247687817, + 0.02468043752014637, + 0.011443487368524075, + 0.014566720463335514, + 0.011595768854022026, + 0.018124854192137718, + 0.06872870028018951, + 0.016845375299453735, + 0.023921595886349678, + 0.02613973058760166, + 0.013203301467001438, + 0.01858809031546116, + 0.03869505971670151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.007310071028769016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700620487332344 + ], + "opposite_logprob_tokens": [ + -2.6950621604919434 + ], + "logprob_reward": -0.0700620487332344, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700620487332344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1488451212644577, + 0.11589738726615906, + 0.0937323048710823, + 0.043977636843919754, + 0.07488875091075897, + 0.15708325803279877, + 0.06484481692314148, + 0.07500550895929337, + 0.1419098973274231, + 0.08381535857915878 + ], + "credit_map": [ + 0.020021595060825348, + 0.019067911431193352, + 0.01480790413916111, + 0.013816329650580883, + 0.020909547805786133, + 0.020901544019579887, + 0.02029263786971569, + 0.019027646631002426, + 0.01157496776431799, + 0.011376574635505676, + 0.010596005246043205, + 0.016698479652404785, + 0.01630554348230362, + 0.013989825733006, + 0.022130832076072693, + 0.013225152157247066, + 0.00909570511430502, + 0.014010819606482983, + 0.010333972051739693, + 0.011350288987159729, + 0.017042091116309166, + 0.012772091664373875, + 0.01177174597978592, + 0.007355596870183945, + 0.0061520664021372795, + 0.006854952313005924, + 0.003233253490179777, + 0.0049697887152433395, + 0.007057345937937498, + 0.005716417450457811, + 0.006772100459784269, + 0.003221712075173855, + 0.008032971061766148, + 0.008844912052154541, + 0.007866757921874523, + 0.010254867374897003, + 0.01205344870686531, + 0.009385045617818832, + 0.010934526100754738, + 0.007516223005950451, + 0.01599007658660412, + 0.02242319844663143, + 0.019412675872445107, + 0.02295292355120182, + 0.021149232983589172, + 0.017696168273687363, + 0.02249469980597496, + 0.01496428158134222, + 0.009402177296578884, + 0.008420243859291077, + 0.01246240921318531, + 0.0054842205718159676, + 0.005981506314128637, + 0.008222786709666252, + 0.007572849281132221, + 0.00729862367734313, + 0.006366922985762358, + 0.011898038908839226, + 0.0066050696186721325, + 0.00987978558987379, + 0.014718816615641117, + 0.007231011986732483, + 0.013665836304426193, + 0.0046400269493460655, + 0.008292155340313911, + 0.030796680599451065, + 0.010307101532816887, + 0.017391730099916458, + 0.027653172612190247, + 0.012560123577713966, + 0.023481208831071854, + 0.011427728459239006, + 0.014358262531459332, + 0.01248716190457344, + 0.010109457187354565, + 0.009778222069144249, + 0.009977839887142181, + 0.008384523913264275, + 0.008755046874284744, + 0.009964841417968273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0700620487332344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349330320954323 + ], + "opposite_logprob_tokens": [ + -3.6513493061065674 + ], + "logprob_reward": -0.026349330320954323, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349330320954323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4354010224342346, + 0.05238828808069229, + 0.057616617530584335, + 0.037965819239616394, + 0.03395766019821167, + 0.07871385663747787, + 0.05122999846935272, + 0.04402574896812439, + 0.08489460498094559, + 0.1238064095377922 + ], + "credit_map": [ + 0.058545004576444626, + 0.12833766639232635, + 0.044867221266031265, + 0.03599061816930771, + 0.05542568489909172, + 0.06469886004924774, + 0.030071258544921875, + 0.01746470481157303, + 0.010835308581590652, + 0.008235791698098183, + 0.004978881683200598, + 0.004259121138602495, + 0.006927642971277237, + 0.00837410893291235, + 0.005369208753108978, + 0.0034082261845469475, + 0.01000453345477581, + 0.011819972656667233, + 0.004969592206180096, + 0.004308380652219057, + 0.0067155929282307625, + 0.008756052702665329, + 0.0052241249941289425, + 0.005818368401378393, + 0.0058752079494297504, + 0.004478295799344778, + 0.0042821490205824375, + 0.003737182356417179, + 0.003758470993489027, + 0.006705082021653652, + 0.005194462835788727, + 0.0039349705912172794, + 0.0036359746009111404, + 0.006893071345984936, + 0.0032532925251871347, + 0.0035871120635420084, + 0.004991425201296806, + 0.00448501855134964, + 0.004265087191015482, + 0.0028466808144003153, + 0.00689605763182044, + 0.009656841866672039, + 0.010564691387116909, + 0.00983874499797821, + 0.011617164127528667, + 0.009226923808455467, + 0.011584704741835594, + 0.009328730404376984, + 0.006061170250177383, + 0.006523787509649992, + 0.00838960986584425, + 0.005630603991448879, + 0.0071455067954957485, + 0.00532557163387537, + 0.006920493207871914, + 0.005233257543295622, + 0.004196312744170427, + 0.006713368929922581, + 0.003968542907387018, + 0.004770386032760143, + 0.007766904775053263, + 0.005022737663239241, + 0.0073548476211726665, + 0.004232645500451326, + 0.008158170618116856, + 0.013308354653418064, + 0.006714501418173313, + 0.010474249720573425, + 0.01668657548725605, + 0.007254098076373339, + 0.012017716653645039, + 0.010280941613018513, + 0.00832218024879694, + 0.03052414581179619, + 0.013348868116736412, + 0.01063669752329588, + 0.022744856774806976, + 0.012292375788092613, + 0.011679166927933693, + 0.014258116483688354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.026349330320954323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192805349826813 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.018192805349826813, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192805349826813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16289320588111877, + 0.0399043932557106, + 0.05495652183890343, + 0.04401000216603279, + 0.0574178546667099, + 0.09087882936000824, + 0.12231489270925522, + 0.1658915877342224, + 0.17119136452674866, + 0.09054134786128998 + ], + "credit_map": [ + 0.02693437971174717, + 0.030147921293973923, + 0.0128641901537776, + 0.015915710479021072, + 0.02275129035115242, + 0.021919170394539833, + 0.020651141181588173, + 0.01170941349118948, + 0.004817876499146223, + 0.007934315130114555, + 0.003271472407504916, + 0.0034962522331625223, + 0.005589294712990522, + 0.005128062795847654, + 0.00667779054492712, + 0.002989332191646099, + 0.005471214186400175, + 0.005995678249746561, + 0.0059415968134999275, + 0.005519296042621136, + 0.011255919933319092, + 0.007582888938486576, + 0.008612656965851784, + 0.004577271174639463, + 0.005727224051952362, + 0.006403819657862186, + 0.004558882676064968, + 0.003554298309609294, + 0.007941058836877346, + 0.005977974738925695, + 0.0059875138103961945, + 0.0038592286873608828, + 0.007330256048589945, + 0.00824642088264227, + 0.0050988285802304745, + 0.005355881992727518, + 0.009519629180431366, + 0.007822290994226933, + 0.007162613328546286, + 0.00688193691894412, + 0.008831488899886608, + 0.011604108847677708, + 0.011102218180894852, + 0.012044616974890232, + 0.013034844771027565, + 0.010803467594087124, + 0.01566259376704693, + 0.007795488927513361, + 0.01403100322932005, + 0.014195061288774014, + 0.02031468227505684, + 0.011830969713628292, + 0.014773380942642689, + 0.015378267504274845, + 0.017167603597044945, + 0.014623934403061867, + 0.01592942699790001, + 0.025813035666942596, + 0.017023013904690742, + 0.022187286987900734, + 0.03027919866144657, + 0.017197873443365097, + 0.023242229595780373, + 0.014219527132809162, + 0.009518889710307121, + 0.03409545496106148, + 0.012018335051834583, + 0.0217177364975214, + 0.03464757651090622, + 0.016357384622097015, + 0.02732286974787712, + 0.015513136051595211, + 0.009532758966088295, + 0.014846901409327984, + 0.010370531119406223, + 0.007693793624639511, + 0.01612643338739872, + 0.013528481125831604, + 0.010120555758476257, + 0.00832189992070198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.018192805349826813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a man's face wearing a three-mountain flying phoenix hat with gold and silver tassels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing golden chain mail armor engraved with ancient mythical beast patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show dark storm clouds with silent flashes of lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the vertical divine pattern on the man's forehead split to reveal a pupilless golden energy eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629304192960262 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629304192960262, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629304192960262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14551624655723572, + 0.04730122908949852, + 0.036585044115781784, + 0.030723977833986282, + 0.06339957565069199, + 0.0654255747795105, + 0.0557042732834816, + 0.08724668622016907, + 0.13455447554588318, + 0.33354291319847107 + ], + "credit_map": [ + 0.026092329993844032, + 0.02359071560204029, + 0.01155217457562685, + 0.014515125192701817, + 0.02018568105995655, + 0.023851661011576653, + 0.01648085191845894, + 0.009247715584933758, + 0.006493251770734787, + 0.009693781845271587, + 0.004216501023620367, + 0.005197904538363218, + 0.00586990499868989, + 0.0064776320941746235, + 0.006570650264620781, + 0.0027816002257168293, + 0.004160286393016577, + 0.005378594622015953, + 0.0035878182388842106, + 0.003895772388204932, + 0.005853111855685711, + 0.004981555975973606, + 0.005295301787555218, + 0.003432601923123002, + 0.003447223687544465, + 0.004819027613848448, + 0.0037340745329856873, + 0.0034619003999978304, + 0.0037729954347014427, + 0.003527942346408963, + 0.004377169068902731, + 0.0035836452152580023, + 0.008619487285614014, + 0.009096810594201088, + 0.0059174057096242905, + 0.007983813993632793, + 0.00844008382409811, + 0.008548867888748646, + 0.007783096749335527, + 0.0070100100710988045, + 0.006801625248044729, + 0.00985721591860056, + 0.008607165887951851, + 0.009772122837603092, + 0.007093628402799368, + 0.00803298782557249, + 0.007351695094257593, + 0.007909135892987251, + 0.005470466334372759, + 0.009897355921566486, + 0.006510557606816292, + 0.005866439081728458, + 0.00781673938035965, + 0.005933539941906929, + 0.006758588366210461, + 0.007450588047504425, + 0.012390879914164543, + 0.007786109112203121, + 0.01280587911605835, + 0.0077058542519807816, + 0.00949571281671524, + 0.011680479161441326, + 0.012984365224838257, + 0.01239740289747715, + 0.01119180303066969, + 0.018795890733599663, + 0.010774610564112663, + 0.018155351281166077, + 0.02207450568675995, + 0.018263572826981544, + 0.020024465397000313, + 0.015274283476173878, + 0.04872031882405281, + 0.08228175342082977, + 0.03694967180490494, + 0.029550805687904358, + 0.034819189459085464, + 0.027107877656817436, + 0.02656131610274315, + 0.04755198955535889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.007629304192960262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_tensor([940, 940])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_tensor([940, 940])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..896dca99773e7fd8bd8d1629952b5159496a07e9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_tensor([940, 940])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "camera, imbued with the quality of a humanistic documentary, pans in a slow, steady rhythm, arcing around the central figure. On an ancient battlefield, dusty and torn, the reformer Shang Yang, clad in the deep black uniform of a Qin dynasty official, the simple embroidery on his collar and cuffs exudes a sense of authority amidst the chaos. His robes are stained with yellow dust, his cheeks stained with sweat and dust. His angular face, lips pursed, his meticulously tied bun slightly dishevelled, yet his deep, eagle-like eyes, piercing the dense smoke and dust, survey the battlefield with unwavering determination. The background is a purgatory of swords, swordplay, and the tumbling of men and horses. Soldiers clash fiercely, leaving broken weapons and fallen bodies scattered everywhere. The glaring sunlight, filtered through the thick smoke and dust, casts a scorching, warm yellow hue, imbuing the entire scene with a harsh, textured quality. As the camera moves steadily, Shang Yang holds the hilt of the ancient bronze sword tightly in one hand, while the other hand swings forward with a steady and decisive gesture, as if silently commanding thousands of troops. His unmoved command posture in the extreme chaos forms a sharp contrast with the bloody fighting around him, creating a tense, epic and solemn atmosphere.", + "seed": 6155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_940_seed_6155_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07278641760349273, + "vlm_logprob": -0.07278641760349273, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.07278641760349273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027943747118115425 + ], + "opposite_logprob_tokens": [ + -3.7779438495635986 + ], + "logprob_reward": -0.027943747118115425, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027943747118115425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1508091241121292, + 0.07834692299365997, + 0.09493956714868546, + 0.05715582147240639, + 0.0488755963742733, + 0.08070649951696396, + 0.1260806769132614, + 0.25074532628059387, + 0.062304601073265076, + 0.05003579333424568 + ], + "credit_map": [ + 0.01313693542033434, + 0.016364390030503273, + 0.019619353115558624, + 0.022268079221248627, + 0.030303552746772766, + 0.010709243826568127, + 0.01570250652730465, + 0.022705087438225746, + 0.0060896012000739574, + 0.009291915223002434, + 0.00898608285933733, + 0.006037702318280935, + 0.019787514582276344, + 0.008423668332397938, + 0.011106409132480621, + 0.008624039590358734, + 0.01391263585537672, + 0.012702950276434422, + 0.013078958727419376, + 0.008199963718652725, + 0.007370844017714262, + 0.026319360360503197, + 0.005779157392680645, + 0.007575706113129854, + 0.005441735498607159, + 0.00658941688016057, + 0.008815336972475052, + 0.008885785937309265, + 0.009109550155699253, + 0.005728352349251509, + 0.007346789352595806, + 0.005238865502178669, + 0.004791267681866884, + 0.005370770581066608, + 0.0065284231677651405, + 0.00505034951493144, + 0.007259701378643513, + 0.008423003368079662, + 0.007033985108137131, + 0.004418102093040943, + 0.007596299517899752, + 0.01807704009115696, + 0.008066224865615368, + 0.007269649766385555, + 0.012494617141783237, + 0.010525858961045742, + 0.008584565483033657, + 0.008092252537608147, + 0.009293726645410061, + 0.014597252011299133, + 0.012450918555259705, + 0.015364941209554672, + 0.025854825973510742, + 0.013421393930912018, + 0.019906969740986824, + 0.015190660953521729, + 0.02410942316055298, + 0.067189060151577, + 0.013591073453426361, + 0.010561722330749035, + 0.0173487588763237, + 0.09156966209411621, + 0.011480042710900307, + 0.014895591884851456, + 0.007366132456809282, + 0.007074305322021246, + 0.009242461994290352, + 0.007234632503241301, + 0.013276557438075542, + 0.006085013505071402, + 0.005613804794847965, + 0.006411699112504721, + 0.005154646933078766, + 0.00747491093352437, + 0.005822054110467434, + 0.009685199707746506, + 0.00819601584225893, + 0.002977787284180522, + 0.005591790657490492, + 0.005133394617587328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.027943747118115425 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079863533377647 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.016079863533377647, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079863533377647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21154175698757172, + 0.047647032886743546, + 0.17433205246925354, + 0.07776249200105667, + 0.03188112750649452, + 0.02512686513364315, + 0.02921491302549839, + 0.28578320145606995, + 0.04281645268201828, + 0.07389410585165024 + ], + "credit_map": [ + 0.016482971608638763, + 0.026382960379123688, + 0.02053653635084629, + 0.04385402426123619, + 0.037830717861652374, + 0.012394741177558899, + 0.02742932364344597, + 0.02663048915565014, + 0.0033725888933986425, + 0.007509725634008646, + 0.005705949850380421, + 0.0046214452013373375, + 0.007819464430212975, + 0.005500854458659887, + 0.006375863216817379, + 0.006741141900420189, + 0.019122246652841568, + 0.06572648882865906, + 0.01922411285340786, + 0.009307827800512314, + 0.010358343832194805, + 0.03476419299840927, + 0.0063019744120538235, + 0.009526868350803852, + 0.008186990395188332, + 0.011522318236529827, + 0.012319343164563179, + 0.010331402532756329, + 0.011856974102556705, + 0.008229495957493782, + 0.008991198614239693, + 0.006324765272438526, + 0.002665482461452484, + 0.0037587673868983984, + 0.004436276387423277, + 0.003534706076607108, + 0.0051941657438874245, + 0.005628366954624653, + 0.003735446371138096, + 0.0029279158916324377, + 0.002520199166610837, + 0.006060452666133642, + 0.003468393813818693, + 0.002530227182433009, + 0.003581508295610547, + 0.003462845692411065, + 0.0019227993907406926, + 0.0015804398572072387, + 0.002698255702853203, + 0.004534425213932991, + 0.0028868564404547215, + 0.0037738692481070757, + 0.005751744844019413, + 0.0036638802848756313, + 0.003091808408498764, + 0.002814072882756591, + 0.029384277760982513, + 0.08303496986627579, + 0.011896882206201553, + 0.012996385805308819, + 0.015895547345280647, + 0.10495971143245697, + 0.012039665132761002, + 0.015575745142996311, + 0.005705239251255989, + 0.003793037962168455, + 0.00750293955206871, + 0.0061950162053108215, + 0.005332770757377148, + 0.003899749368429184, + 0.005515531171113253, + 0.004872171673923731, + 0.006882558576762676, + 0.012416674755513668, + 0.008073256351053715, + 0.012570802122354507, + 0.011791889555752277, + 0.004371210932731628, + 0.010767979547381401, + 0.0070197321474552155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.016079863533377647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001403719186783 + ], + "opposite_logprob_tokens": [ + -2.695013999938965 + ], + "logprob_reward": -0.07001403719186783, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001403719186783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1921520084142685, + 0.04410145431756973, + 0.12730513513088226, + 0.038142669945955276, + 0.05347621813416481, + 0.033884238451719284, + 0.04261128976941109, + 0.11403302848339081, + 0.231894388794899, + 0.12239956110715866 + ], + "credit_map": [ + 0.018468165770173073, + 0.02469080500304699, + 0.02244950272142887, + 0.034456152468919754, + 0.021898575127124786, + 0.009655329398810863, + 0.022770123556256294, + 0.037763360887765884, + 0.004116412717849016, + 0.005400562658905983, + 0.00696187186986208, + 0.00622527627274394, + 0.0057829804718494415, + 0.004586823284626007, + 0.0055815777741372585, + 0.005445947404950857, + 0.006704546045511961, + 0.04211501404643059, + 0.008505473844707012, + 0.004851768724620342, + 0.00782189704477787, + 0.04811016842722893, + 0.005061683245003223, + 0.004134586546570063, + 0.0038044096436351538, + 0.004997374024242163, + 0.006386965978890657, + 0.006087993737310171, + 0.00465967645868659, + 0.004453722387552261, + 0.00452168844640255, + 0.0032308355439454317, + 0.0043389503844082355, + 0.0062814997509121895, + 0.00758729875087738, + 0.005179113708436489, + 0.009944628924131393, + 0.006837774068117142, + 0.007202069740742445, + 0.006104880012571812, + 0.0027982795145362616, + 0.005143440794199705, + 0.0035052222665399313, + 0.0038146066945046186, + 0.006959747523069382, + 0.003890623804181814, + 0.004313673358410597, + 0.0034586451947689056, + 0.0050902157090604305, + 0.007144357077777386, + 0.0042520444840192795, + 0.004930825438350439, + 0.008173331618309021, + 0.004590021446347237, + 0.004822900518774986, + 0.0036075927782803774, + 0.02444605715572834, + 0.012928378768265247, + 0.01017480343580246, + 0.009275509044528008, + 0.013888562098145485, + 0.023348450660705566, + 0.008937186561524868, + 0.011034074239432812, + 0.0243845134973526, + 0.018243689090013504, + 0.045946523547172546, + 0.03763417899608612, + 0.02191133238375187, + 0.022420158609747887, + 0.02794053964316845, + 0.03341344743967056, + 0.010581456124782562, + 0.02016501873731613, + 0.014152780175209045, + 0.016889538615942, + 0.02347879856824875, + 0.007553033530712128, + 0.015505755320191383, + 0.014073170721530914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.07001403719186783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989841103553772 + ], + "opposite_logprob_tokens": [ + -1.8048983812332153 + ], + "logprob_reward": -0.17989841103553772, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989841103553772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3002370297908783, + 0.19214411079883575, + 0.19159714877605438, + 0.0674397423863411, + 0.05036959797143936, + 0.039012543857097626, + 0.018716201186180115, + 0.03929654881358147, + 0.05221362039446831, + 0.048973504453897476 + ], + "credit_map": [ + 0.03402179479598999, + 0.03386909142136574, + 0.05138127878308296, + 0.05154038220643997, + 0.03973212093114853, + 0.021909009665250778, + 0.02966703288257122, + 0.038116298615932465, + 0.020841138437390327, + 0.02173558808863163, + 0.03762480989098549, + 0.029276955872774124, + 0.024048831313848495, + 0.018447495996952057, + 0.019033357501029968, + 0.02113594114780426, + 0.014320109970867634, + 0.08038462698459625, + 0.013652524910867214, + 0.007074139080941677, + 0.008215623907744884, + 0.058809906244277954, + 0.004240219946950674, + 0.00489999633282423, + 0.006739456672221422, + 0.008237971924245358, + 0.012695970013737679, + 0.011408694088459015, + 0.007923766039311886, + 0.007206904236227274, + 0.007604989688843489, + 0.00562199205160141, + 0.005149364471435547, + 0.00631533982232213, + 0.012211077846586704, + 0.005337991286069155, + 0.006421498488634825, + 0.00703780073672533, + 0.00451855082064867, + 0.003377971239387989, + 0.0037470893003046513, + 0.006641247309744358, + 0.0058705187402665615, + 0.003775953548029065, + 0.007267532870173454, + 0.0037743861321359873, + 0.003906426951289177, + 0.004029387608170509, + 0.0023450220469385386, + 0.0038413754664361477, + 0.002874121069908142, + 0.0021099362056702375, + 0.002201157622039318, + 0.0024495162069797516, + 0.0013882950879633427, + 0.0015067770145833492, + 0.006552492268383503, + 0.007132021710276604, + 0.005872535519301891, + 0.0030893501825630665, + 0.005706582218408585, + 0.0039542303420603275, + 0.003408192889764905, + 0.0035811462439596653, + 0.004408255685120821, + 0.004465283360332251, + 0.011451668106019497, + 0.008345581591129303, + 0.005298470612615347, + 0.005267823114991188, + 0.0066339485347270966, + 0.006342589855194092, + 0.004968415945768356, + 0.005364350508898497, + 0.007861793041229248, + 0.006737988907843828, + 0.005602020770311356, + 0.0037813824601471424, + 0.00531920837238431, + 0.009338345378637314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.17989841103553772 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999602913856506 + ], + "opposite_logprob_tokens": [ + -2.6949961185455322 + ], + "logprob_reward": -0.06999602913856506, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999602913856506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11978162825107574, + 0.1148100271821022, + 0.37531426548957825, + 0.08918079733848572, + 0.04883057251572609, + 0.023259645327925682, + 0.023457828909158707, + 0.06445743888616562, + 0.1267309933900833, + 0.01417683344334364 + ], + "credit_map": [ + 0.010141930542886257, + 0.01458155456930399, + 0.009513456374406815, + 0.01869954913854599, + 0.031653132289648056, + 0.008454442024230957, + 0.01242746226489544, + 0.014310095459222794, + 0.008318903855979443, + 0.01187280472368002, + 0.010836240835487843, + 0.010002559050917625, + 0.03099161386489868, + 0.010993940755724907, + 0.019464408978819847, + 0.012329542078077793, + 0.018603861331939697, + 0.15264831483364105, + 0.018494611606001854, + 0.010482731275260448, + 0.01628248766064644, + 0.13815028965473175, + 0.011872130446135998, + 0.008779809810221195, + 0.00784996710717678, + 0.013693530112504959, + 0.013134513050317764, + 0.011879888363182545, + 0.01643870212137699, + 0.008095436729490757, + 0.010561146773397923, + 0.007527610752731562, + 0.004342136438935995, + 0.005330448970198631, + 0.008146378211677074, + 0.005563812330365181, + 0.009938680566847324, + 0.005361931398510933, + 0.005918757524341345, + 0.00422842288389802, + 0.0018847802421078086, + 0.003698627697303891, + 0.002811272395774722, + 0.00288855261169374, + 0.0048365662805736065, + 0.00265571940690279, + 0.0021561384201049805, + 0.002327987225726247, + 0.001864809775725007, + 0.002784394659101963, + 0.0023969546891748905, + 0.002592138946056366, + 0.005841248203068972, + 0.0021912637166678905, + 0.0027378001250326633, + 0.0030492180958390236, + 0.012643867172300816, + 0.01212265994399786, + 0.0078054144978523254, + 0.006783736869692802, + 0.0061599924229085445, + 0.005524827633053064, + 0.005497060716152191, + 0.007919876836240292, + 0.013231678865849972, + 0.009993699379265308, + 0.024444954469799995, + 0.020242366939783096, + 0.012728167697787285, + 0.012712388299405575, + 0.015821602195501328, + 0.017556127160787582, + 0.001685126917436719, + 0.0018518591532483697, + 0.002256372245028615, + 0.002088281325995922, + 0.002115045441314578, + 0.0008347828988917172, + 0.0014231683453544974, + 0.0019221949623897672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.06999602913856506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera, imbued with the quality of a humanistic documentary, pans in a slow, steady rhythm, arcing around the central figure. On an ancient battlefield, dusty and torn, the reformer Shang Yang, clad in the deep black uniform of a Qin dynasty official, the simple embroidery on his collar and cuffs exudes a sense of authority amidst the chaos. His robes are stained with yellow dust, his cheeks stained with sweat and dust. His angular face, lips pursed, his meticulously tied bun slightly dishevelled, yet his deep, eagle-like eyes, piercing the dense smoke and dust, survey the battlefield with unwavering determination. The background is a purgatory of swords, swordplay, and the tumbling of men and horses. Soldiers clash fiercely, leaving broken weapons and fallen bodies scattered everywhere. The glaring sunlight, filtered through the thick smoke and dust, casts a scorching, warm yellow hue, imbuing the entire scene with a harsh, textured quality. As the camera moves steadily, Shang Yang holds the hilt of the ancient bronze sword tightly in one hand, while the other hand swings forward with a steady and decisive gesture, as if silently commanding thousands of troops. His unmoved command posture in the extreme chaos forms a sharp contrast with the bloody fighting around him, creating a tense, epic and solemn atmosphere.", + "seed": 6156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_940_seed_6156_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06444162428379059, + "vlm_logprob": -0.06444162428379059, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.06444162428379059, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02917097881436348 + ], + "opposite_logprob_tokens": [ + -3.6541709899902344 + ], + "logprob_reward": -0.02917097881436348, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02917097881436348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14855389297008514, + 0.07376430928707123, + 0.09970560669898987, + 0.18739815056324005, + 0.0751471221446991, + 0.08914880454540253, + 0.21874292194843292, + 0.057239316403865814, + 0.026013638824224472, + 0.024286290630698204 + ], + "credit_map": [ + 0.011191361583769321, + 0.016650501638650894, + 0.015940194949507713, + 0.023007964715361595, + 0.02411828003823757, + 0.015959354117512703, + 0.016773462295532227, + 0.024912770837545395, + 0.009362085722386837, + 0.006919465959072113, + 0.011877662502229214, + 0.007932822220027447, + 0.012855101376771927, + 0.005763617809861898, + 0.007562802638858557, + 0.011490748263895512, + 0.008368747308850288, + 0.009667432866990566, + 0.010002308525145054, + 0.013691572472453117, + 0.024196594953536987, + 0.010567283257842064, + 0.0138562535867095, + 0.009355415590107441, + 0.02074439637362957, + 0.022376343607902527, + 0.014725307002663612, + 0.015629766508936882, + 0.05506334453821182, + 0.018221525475382805, + 0.022946612909436226, + 0.017690837383270264, + 0.009529819712042809, + 0.008098622784018517, + 0.008356017991900444, + 0.009331827983260155, + 0.010149016045033932, + 0.008053772151470184, + 0.013955442234873772, + 0.007672598585486412, + 0.009330900385975838, + 0.014003651216626167, + 0.007786199916154146, + 0.014013326726853848, + 0.013976783491671085, + 0.00832776166498661, + 0.01308926660567522, + 0.0086209112778306, + 0.025481240823864937, + 0.03385841101408005, + 0.025287959724664688, + 0.030976109206676483, + 0.028355974704027176, + 0.03644739091396332, + 0.021870123222470284, + 0.016465695574879646, + 0.005136948544532061, + 0.006835072301328182, + 0.01487039215862751, + 0.0053595867939293385, + 0.00906539335846901, + 0.00738433888182044, + 0.0038245825562626123, + 0.0047630006447434425, + 0.002993750385940075, + 0.003693140810355544, + 0.003871392458677292, + 0.0029475728515535593, + 0.0038943402469158173, + 0.003405940253287554, + 0.002094913274049759, + 0.0031125852838158607, + 0.0027904314920306206, + 0.002957836491987109, + 0.0024157606530934572, + 0.0025822739116847515, + 0.0021625286899507046, + 0.0027538463473320007, + 0.0028347126208245754, + 0.005788898561149836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.02917097881436348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033705245703458786 + ], + "opposite_logprob_tokens": [ + -3.408705234527588 + ], + "logprob_reward": -0.033705245703458786, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033705245703458786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14999331533908844, + 0.07577648758888245, + 0.1065187156200409, + 0.10186046361923218, + 0.065757617354393, + 0.28902819752693176, + 0.09245308488607407, + 0.05148685351014137, + 0.04323306679725647, + 0.023892199620604515 + ], + "credit_map": [ + 0.015841858461499214, + 0.01380483154207468, + 0.0207720585167408, + 0.017861604690551758, + 0.01863933354616165, + 0.019979327917099, + 0.016862232238054276, + 0.026232067495584488, + 0.01078175287693739, + 0.00669008307158947, + 0.011767699383199215, + 0.00968104787170887, + 0.013377310708165169, + 0.005570299923419952, + 0.009349891915917397, + 0.008558405563235283, + 0.015794746577739716, + 0.013757460750639439, + 0.018213847652077675, + 0.01099985372275114, + 0.014326789416372776, + 0.011009490117430687, + 0.012074708938598633, + 0.010341831482946873, + 0.01811031624674797, + 0.010348315350711346, + 0.011809156276285648, + 0.006353873293846846, + 0.011622331105172634, + 0.010854021646082401, + 0.012958170846104622, + 0.01980428956449032, + 0.008815115317702293, + 0.00803475920110941, + 0.0069226971827447414, + 0.008203940466046333, + 0.009701204486191273, + 0.007371526211500168, + 0.011350654065608978, + 0.005357733462005854, + 0.038885217159986496, + 0.04233956336975098, + 0.05315698683261871, + 0.02990913763642311, + 0.031073318794369698, + 0.03748715668916702, + 0.024432918056845665, + 0.03174389898777008, + 0.010126103647053242, + 0.009549694135785103, + 0.012500789016485214, + 0.008509072475135326, + 0.011907504871487617, + 0.013027183711528778, + 0.006731903180480003, + 0.02010084129869938, + 0.006197588052600622, + 0.006177915260195732, + 0.012481684796512127, + 0.004854151513427496, + 0.007270409259945154, + 0.005828286521136761, + 0.004004784394055605, + 0.0046720365062355995, + 0.004259787034243345, + 0.0040155258029699326, + 0.005532596260309219, + 0.007515996694564819, + 0.00547823728993535, + 0.005550621077418327, + 0.006091014482080936, + 0.004789293278008699, + 0.0022054410073906183, + 0.002795381238684058, + 0.00238686241209507, + 0.0036788389552384615, + 0.003173865145072341, + 0.002816110849380493, + 0.0036198701709508896, + 0.0032158291433006525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.033705245703458786 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030137002468109 + ], + "opposite_logprob_tokens": [ + -2.3503012657165527 + ], + "logprob_reward": -0.10030137002468109, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030137002468109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3166787326335907, + 0.05551200732588768, + 0.04835345968604088, + 0.17257854342460632, + 0.06754652410745621, + 0.05901863053441048, + 0.05589624494314194, + 0.09924467653036118, + 0.07947912067174911, + 0.045692022889852524 + ], + "credit_map": [ + 0.025301378220319748, + 0.030951132997870445, + 0.04206078127026558, + 0.05049710348248482, + 0.03372518718242645, + 0.03516780957579613, + 0.031656477600336075, + 0.06731890141963959, + 0.007580198813229799, + 0.004204075783491135, + 0.01169822458177805, + 0.006087329238653183, + 0.005702055059373379, + 0.0041857208125293255, + 0.005302655044943094, + 0.010751748457551003, + 0.006449483335018158, + 0.005083187948912382, + 0.0076711419969797134, + 0.005523123312741518, + 0.006161411292850971, + 0.004988677799701691, + 0.005173925776034594, + 0.007302513811737299, + 0.026966538280248642, + 0.015321576967835426, + 0.021085457876324654, + 0.013902082107961178, + 0.014689669944345951, + 0.019665436819195747, + 0.02328464761376381, + 0.037663139402866364, + 0.005672474391758442, + 0.007842406630516052, + 0.00867284182459116, + 0.007669379469007254, + 0.013015258125960827, + 0.007824231870472431, + 0.01018657349050045, + 0.006663368549197912, + 0.008115910924971104, + 0.009597411379218102, + 0.01082142349332571, + 0.005313106346875429, + 0.00859935861080885, + 0.006489578168839216, + 0.005191728938370943, + 0.00489011500030756, + 0.007387210614979267, + 0.005954543594270945, + 0.008138767443597317, + 0.0051305051892995834, + 0.008137439377605915, + 0.007814979180693626, + 0.00407642824575305, + 0.009256371296942234, + 0.018898053094744682, + 0.013161818496882915, + 0.01949932798743248, + 0.010716509073972702, + 0.009109613485634327, + 0.012215997092425823, + 0.0061479113064706326, + 0.009495454840362072, + 0.008743719197809696, + 0.01198944728821516, + 0.010249743238091469, + 0.009976631961762905, + 0.01108846627175808, + 0.010729331523180008, + 0.008680464699864388, + 0.008021321147680283, + 0.004102634731680155, + 0.005795170087367296, + 0.006653347983956337, + 0.0065925512462854385, + 0.0044702147133648396, + 0.005491826683282852, + 0.005460055079311132, + 0.007126224227249622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.10030137002468109 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903738856315613 + ], + "opposite_logprob_tokens": [ + -2.4640374183654785 + ], + "logprob_reward": -0.08903738856315613, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903738856315613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0962093323469162, + 0.09318952262401581, + 0.27234122157096863, + 0.06851562857627869, + 0.06957053393125534, + 0.1487315148115158, + 0.06305443495512009, + 0.1108030304312706, + 0.04608515277504921, + 0.03149964660406113 + ], + "credit_map": [ + 0.01116634439677, + 0.01016800757497549, + 0.018599379807710648, + 0.01062026247382164, + 0.012549048289656639, + 0.010491220280528069, + 0.009754667989909649, + 0.012860394082963467, + 0.015675760805606842, + 0.009525715373456478, + 0.02203471213579178, + 0.010786137543618679, + 0.0066437870264053345, + 0.00681914621964097, + 0.011310570873320103, + 0.01039369311183691, + 0.04914219677448273, + 0.03811878710985184, + 0.0674496591091156, + 0.021579302847385406, + 0.013043614104390144, + 0.0320088230073452, + 0.022155441343784332, + 0.02884339913725853, + 0.015260901302099228, + 0.007668950594961643, + 0.009457028470933437, + 0.005078569520264864, + 0.00786280445754528, + 0.006823936011642218, + 0.007072535343468189, + 0.009290900081396103, + 0.007918202318251133, + 0.008033089339733124, + 0.005806528497487307, + 0.008981484919786453, + 0.013369704596698284, + 0.008186726830899715, + 0.010126743465662003, + 0.007148056291043758, + 0.017301587387919426, + 0.023541297763586044, + 0.027986867353320122, + 0.014475276693701744, + 0.02101466991007328, + 0.016520559787750244, + 0.014671383425593376, + 0.013219875283539295, + 0.012841730378568172, + 0.005684203933924437, + 0.009100914001464844, + 0.005645869765430689, + 0.007059886585921049, + 0.010748567059636116, + 0.00354330544359982, + 0.008429956622421741, + 0.01546485722064972, + 0.008602100424468517, + 0.03608524426817894, + 0.006422178354114294, + 0.008643814362585545, + 0.018384145572781563, + 0.005761557258665562, + 0.011439134366810322, + 0.004536907188594341, + 0.004837546963244677, + 0.008473003283143044, + 0.006172117777168751, + 0.0059998854994773865, + 0.006084388587623835, + 0.004812165629118681, + 0.005169139243662357, + 0.003363177878782153, + 0.00369273335672915, + 0.004099006298929453, + 0.0036547647323459387, + 0.002845646347850561, + 0.003392639569938183, + 0.0032621235586702824, + 0.007189555559307337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.08903738856315613 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999313831329346 + ], + "opposite_logprob_tokens": [ + -2.694993019104004 + ], + "logprob_reward": -0.06999313831329346, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999313831329346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12453176826238632, + 0.1003732979297638, + 0.08543403446674347, + 0.1211993470788002, + 0.14509698748588562, + 0.22097758948802948, + 0.0842399150133133, + 0.057750631123781204, + 0.04728875309228897, + 0.013107694685459137 + ], + "credit_map": [ + 0.01376406755298376, + 0.012019893154501915, + 0.01965297758579254, + 0.015561040490865707, + 0.021887319162487984, + 0.012926475144922733, + 0.012543248012661934, + 0.01617673970758915, + 0.010597006417810917, + 0.009105240926146507, + 0.010028338059782982, + 0.012038452550768852, + 0.023487932980060577, + 0.007950281724333763, + 0.014997382648289204, + 0.012168659828603268, + 0.010888492688536644, + 0.0093603003770113, + 0.01752740703523159, + 0.008726155385375023, + 0.013770273886620998, + 0.009102672338485718, + 0.008827118203043938, + 0.007231604307889938, + 0.017311768606305122, + 0.012738910503685474, + 0.011388897895812988, + 0.010262012481689453, + 0.025228582322597504, + 0.014101201668381691, + 0.012704315595328808, + 0.017463646829128265, + 0.017720386385917664, + 0.014697100035846233, + 0.01281584519892931, + 0.020963145419955254, + 0.026383085176348686, + 0.01661810465157032, + 0.021742768585681915, + 0.014156540855765343, + 0.024426866322755814, + 0.03286724537611008, + 0.042054977267980576, + 0.030348582193255424, + 0.024504326283931732, + 0.02167404815554619, + 0.02369467169046402, + 0.021406862884759903, + 0.010769378393888474, + 0.013254638761281967, + 0.011281255632638931, + 0.010591481812298298, + 0.00945670623332262, + 0.012088505551218987, + 0.00728227011859417, + 0.009515678510069847, + 0.011034531518816948, + 0.006240445654839277, + 0.011349642649292946, + 0.005556334741413593, + 0.006600507535040379, + 0.007447581738233566, + 0.004292954225093126, + 0.005228627007454634, + 0.004887559916824102, + 0.0080762580037117, + 0.005618496797978878, + 0.0060469708405435085, + 0.006510860752314329, + 0.006171614397317171, + 0.0043581342324614525, + 0.005618852097541094, + 0.0012447313638404012, + 0.0015424087177962065, + 0.0020405310206115246, + 0.001493025221861899, + 0.001006560167297721, + 0.0014054592465981841, + 0.0013930402928963304, + 0.0029819391202181578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.06999313831329346 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera, imbued with the quality of a humanistic documentary, pans in a slow, steady rhythm, arcing around the central figure. On an ancient battlefield, dusty and torn, the reformer Shang Yang, clad in the deep black uniform of a Qin dynasty official, the simple embroidery on his collar and cuffs exudes a sense of authority amidst the chaos. His robes are stained with yellow dust, his cheeks stained with sweat and dust. His angular face, lips pursed, his meticulously tied bun slightly dishevelled, yet his deep, eagle-like eyes, piercing the dense smoke and dust, survey the battlefield with unwavering determination. The background is a purgatory of swords, swordplay, and the tumbling of men and horses. Soldiers clash fiercely, leaving broken weapons and fallen bodies scattered everywhere. The glaring sunlight, filtered through the thick smoke and dust, casts a scorching, warm yellow hue, imbuing the entire scene with a harsh, textured quality. As the camera moves steadily, Shang Yang holds the hilt of the ancient bronze sword tightly in one hand, while the other hand swings forward with a steady and decisive gesture, as if silently commanding thousands of troops. His unmoved command posture in the extreme chaos forms a sharp contrast with the bloody fighting around him, creating a tense, epic and solemn atmosphere.", + "seed": 6157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_940_seed_6157_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04543364606797695, + "vlm_logprob": -0.04543364606797695, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.04543364606797695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01702890172600746 + ], + "opposite_logprob_tokens": [ + -4.26702880859375 + ], + "logprob_reward": -0.01702890172600746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01702890172600746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10675616562366486, + 0.06448936462402344, + 0.08346017450094223, + 0.035519957542419434, + 0.0455157533288002, + 0.04444175213575363, + 0.14943769574165344, + 0.097665935754776, + 0.2295875996351242, + 0.14312569797039032 + ], + "credit_map": [ + 0.011098050512373447, + 0.012472704984247684, + 0.016896242275834084, + 0.01420108787715435, + 0.0203564390540123, + 0.012462337501347065, + 0.00967639684677124, + 0.00959290936589241, + 0.009971344843506813, + 0.006332604680210352, + 0.009435990825295448, + 0.006068658083677292, + 0.010318400338292122, + 0.01068543829023838, + 0.006102562416344881, + 0.005574362818151712, + 0.008450375869870186, + 0.010972273536026478, + 0.009553711861371994, + 0.007190839387476444, + 0.018243854865431786, + 0.007959147915244102, + 0.007345077581703663, + 0.013744884170591831, + 0.005175694357603788, + 0.004022845067083836, + 0.007446332834661007, + 0.0047551048919558525, + 0.0033051075879484415, + 0.002559410175308585, + 0.0039958590641617775, + 0.004259602632373571, + 0.005630733445286751, + 0.005802590865641832, + 0.004820944741368294, + 0.004479099065065384, + 0.008358675986528397, + 0.005197544116526842, + 0.005839732941240072, + 0.00538643216714263, + 0.005706962663680315, + 0.0041870237328112125, + 0.005219090264290571, + 0.0032642551232129335, + 0.008728853426873684, + 0.005604410544037819, + 0.00652309600263834, + 0.005208059214055538, + 0.027456197887659073, + 0.012872635386884212, + 0.017173489555716515, + 0.017150260508060455, + 0.011816763319075108, + 0.016584984958171844, + 0.019886542111635208, + 0.02649681642651558, + 0.010169225744903088, + 0.008720248006284237, + 0.014536548405885696, + 0.013167097233235836, + 0.013955394737422466, + 0.009446029551327229, + 0.016722138971090317, + 0.010949243791401386, + 0.02051551453769207, + 0.03039940819144249, + 0.04362110793590546, + 0.0280743595212698, + 0.02919643558561802, + 0.020947687327861786, + 0.02313760109245777, + 0.03369545936584473, + 0.011573194526135921, + 0.024163205176591873, + 0.014353959821164608, + 0.013287548907101154, + 0.02114550583064556, + 0.0268417876213789, + 0.015554115176200867, + 0.016206378117203712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.01702890172600746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607552357017994 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.01607552357017994, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607552357017994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18907274305820465, + 0.09079093486070633, + 0.13145780563354492, + 0.0596066452562809, + 0.055938683450222015, + 0.05088483542203903, + 0.07069291919469833, + 0.08689019829034805, + 0.14524534344673157, + 0.1194198876619339 + ], + "credit_map": [ + 0.017905322834849358, + 0.025431590154767036, + 0.03171741962432861, + 0.024043694138526917, + 0.0339488685131073, + 0.02439623884856701, + 0.015606003813445568, + 0.016023576259613037, + 0.014206180348992348, + 0.009228467009961605, + 0.010049630887806416, + 0.007880942896008492, + 0.01806863583624363, + 0.012301050126552582, + 0.009280789643526077, + 0.009775221347808838, + 0.013627780601382256, + 0.014330732636153698, + 0.016859732568264008, + 0.014392750337719917, + 0.02651277370750904, + 0.011126037687063217, + 0.016570817679166794, + 0.018037183210253716, + 0.007100271992385387, + 0.00795183889567852, + 0.012194234877824783, + 0.006453976966440678, + 0.009277451783418655, + 0.004281897097826004, + 0.006484312936663628, + 0.00586265604943037, + 0.005575033836066723, + 0.005224914755672216, + 0.006471788976341486, + 0.006106053479015827, + 0.0089718047529459, + 0.005116695072501898, + 0.010782799683511257, + 0.007689585909247398, + 0.007613555993884802, + 0.004241896793246269, + 0.006122741382569075, + 0.005965452641248703, + 0.006398899480700493, + 0.0070837377570569515, + 0.00814880058169365, + 0.005309747066348791, + 0.01082267239689827, + 0.005764869507402182, + 0.011846963316202164, + 0.00644514849409461, + 0.009192421101033688, + 0.008185827173292637, + 0.011855185963213444, + 0.006579829379916191, + 0.010130578652024269, + 0.007050753571093082, + 0.013445629738271236, + 0.011106767691671848, + 0.010361249558627605, + 0.009504846297204494, + 0.014222930185496807, + 0.011067435145378113, + 0.013888105750083923, + 0.019190974533557892, + 0.02697562985122204, + 0.015509845688939095, + 0.016853317618370056, + 0.02505643479526043, + 0.015523229725658894, + 0.012247802689671516, + 0.008153825998306274, + 0.024074843153357506, + 0.011112034320831299, + 0.014484070241451263, + 0.013786083087325096, + 0.022519515827298164, + 0.015123617835342884, + 0.010165899991989136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.01607552357017994 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203416734933853 + ], + "opposite_logprob_tokens": [ + -2.8120341300964355 + ], + "logprob_reward": -0.06203416734933853, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203416734933853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20766101777553558, + 0.05628085136413574, + 0.09469716995954514, + 0.05252705514431, + 0.04634492099285126, + 0.06258619576692581, + 0.04726842790842056, + 0.12379840761423111, + 0.24845604598522186, + 0.06037990376353264 + ], + "credit_map": [ + 0.01597372256219387, + 0.017101220786571503, + 0.02556973695755005, + 0.028732122853398323, + 0.039711061865091324, + 0.024340864270925522, + 0.025854436680674553, + 0.03037785179913044, + 0.007062557153403759, + 0.006625165231525898, + 0.009757785126566887, + 0.005129609256982803, + 0.009146342054009438, + 0.007556835189461708, + 0.005022338125854731, + 0.005980215035378933, + 0.009233656339347363, + 0.009894558228552341, + 0.010751345194876194, + 0.010338066145777702, + 0.01372349914163351, + 0.008860042318701744, + 0.01192612573504448, + 0.0199698768556118, + 0.005782978143543005, + 0.006746135652065277, + 0.007740775588899851, + 0.006691675633192062, + 0.008075748570263386, + 0.0037018083967268467, + 0.008075878024101257, + 0.0057120537385344505, + 0.004720928147435188, + 0.005316402297466993, + 0.007447369396686554, + 0.0036939082201570272, + 0.009482413530349731, + 0.005631482228636742, + 0.0044624824076890945, + 0.005589933600276709, + 0.006699037738144398, + 0.005298574920743704, + 0.008647874929010868, + 0.006501419004052877, + 0.006381340324878693, + 0.009223892353475094, + 0.009716550819575787, + 0.010117506608366966, + 0.006608457770198584, + 0.003899500472471118, + 0.005133026745170355, + 0.005818214733153582, + 0.00512098241597414, + 0.005658713635057211, + 0.007906126789748669, + 0.0071234023198485374, + 0.013817842118442059, + 0.012301875278353691, + 0.016446519643068314, + 0.015667390078306198, + 0.018212072551250458, + 0.009290959686040878, + 0.02455584704875946, + 0.013505900278687477, + 0.01573333702981472, + 0.030553990975022316, + 0.04064120724797249, + 0.024921119213104248, + 0.026141276583075523, + 0.0660269558429718, + 0.018796000629663467, + 0.025642145425081253, + 0.004351743031293154, + 0.012494558468461037, + 0.006541806738823652, + 0.006796636153012514, + 0.007348362356424332, + 0.012843534350395203, + 0.006256221793591976, + 0.003747043665498495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.06203416734933853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001003623008728 + ], + "opposite_logprob_tokens": [ + -2.69500994682312 + ], + "logprob_reward": -0.07001003623008728, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001003623008728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14669829607009888, + 0.2091692090034485, + 0.16969655454158783, + 0.05175440385937691, + 0.18318302929401398, + 0.027821749448776245, + 0.038529522716999054, + 0.027624132111668587, + 0.08892066031694412, + 0.05660240352153778 + ], + "credit_map": [ + 0.012623063288629055, + 0.018861988559365273, + 0.019268332049250603, + 0.015844043344259262, + 0.02748998999595642, + 0.02170359157025814, + 0.015897925943136215, + 0.015009379014372826, + 0.030127035453915596, + 0.023457974195480347, + 0.042039111256599426, + 0.019403111189603806, + 0.025367503985762596, + 0.02112269029021263, + 0.018246853724122047, + 0.029404953122138977, + 0.02515755407512188, + 0.019839249551296234, + 0.03814804553985596, + 0.01652996800839901, + 0.020443802699446678, + 0.016724202781915665, + 0.013240153901278973, + 0.019613606855273247, + 0.007208627182990313, + 0.006557583808898926, + 0.007824108004570007, + 0.006683876737952232, + 0.007300572004169226, + 0.0036073161754757166, + 0.006214485038071871, + 0.006357837934046984, + 0.03123558685183525, + 0.019854143261909485, + 0.03135797381401062, + 0.018132764846086502, + 0.016678443178534508, + 0.020860671997070312, + 0.0223332978785038, + 0.02273017354309559, + 0.004121751524507999, + 0.0025413306429982185, + 0.0031458085868507624, + 0.0022541286889463663, + 0.003474996890872717, + 0.0039828745648264885, + 0.0031981177162379026, + 0.005102744325995445, + 0.007081157527863979, + 0.002940124599263072, + 0.007105955854058266, + 0.003928626887500286, + 0.0029172166250646114, + 0.004345645662397146, + 0.005462470464408398, + 0.0047483243979513645, + 0.0027573818806558847, + 0.0026638470590114594, + 0.0030930782668292522, + 0.0031878771260380745, + 0.0034097693860530853, + 0.0024966036435216665, + 0.005280464421957731, + 0.004735111258924007, + 0.009894571267068386, + 0.013478743843734264, + 0.014462579973042011, + 0.016197098419070244, + 0.007567781489342451, + 0.008126459084451199, + 0.008319145068526268, + 0.01087428629398346, + 0.0052288793958723545, + 0.010206750594079494, + 0.00879701878875494, + 0.01308723259717226, + 0.004089111927896738, + 0.006472408305853605, + 0.004588200710713863, + 0.004132802132517099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.07001003623008728 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062019601464271545 + ], + "opposite_logprob_tokens": [ + -2.8120195865631104 + ], + "logprob_reward": -0.062019601464271545, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062019601464271545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1422749012708664, + 0.09415853023529053, + 0.05409834533929825, + 0.042329661548137665, + 0.039000481367111206, + 0.17351554334163666, + 0.22380879521369934, + 0.028585853055119514, + 0.15882036089897156, + 0.04340758174657822 + ], + "credit_map": [ + 0.014062766917049885, + 0.016853876411914825, + 0.019037356600165367, + 0.015001079067587852, + 0.030915990471839905, + 0.021449759602546692, + 0.014195973053574562, + 0.010758095420897007, + 0.009954064153134823, + 0.009542366489768028, + 0.010288292542099953, + 0.011079872958362103, + 0.022496124729514122, + 0.009388319216668606, + 0.012732487171888351, + 0.008676993660628796, + 0.006921600084751844, + 0.005006247200071812, + 0.006598186679184437, + 0.006584468763321638, + 0.008900951594114304, + 0.004044693429023027, + 0.007771485950797796, + 0.008270707912743092, + 0.004381868522614241, + 0.004822412505745888, + 0.005271047819405794, + 0.004862845875322819, + 0.007541859056800604, + 0.0030943190213292837, + 0.006330853793770075, + 0.006024452392011881, + 0.00538673298433423, + 0.003197501413524151, + 0.005363482981920242, + 0.003819533158093691, + 0.005245858337730169, + 0.004338334314525127, + 0.004717872943729162, + 0.006931162904947996, + 0.02986939437687397, + 0.012504421174526215, + 0.023558322340250015, + 0.017710329964756966, + 0.016748439520597458, + 0.028081992641091347, + 0.019602304324507713, + 0.025440337136387825, + 0.042891692370176315, + 0.018942251801490784, + 0.03320292755961418, + 0.023556385189294815, + 0.018406657502055168, + 0.02370794117450714, + 0.02983793430030346, + 0.03326299414038658, + 0.0032239400316029787, + 0.002419056836515665, + 0.0033376046922057867, + 0.003911815583705902, + 0.0032212042715400457, + 0.0026903185062110424, + 0.006164934486150742, + 0.003616977483034134, + 0.009005285799503326, + 0.01831871084868908, + 0.026092126965522766, + 0.017495568841695786, + 0.014557880349457264, + 0.037789613008499146, + 0.017057543620467186, + 0.018503613770008087, + 0.0031427787616848946, + 0.011017631739377975, + 0.004781358875334263, + 0.0049850489012897015, + 0.004560449160635471, + 0.008337480016052723, + 0.004026406444609165, + 0.002556425519287586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.062019601464271545 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera, imbued with the quality of a humanistic documentary, pans in a slow, steady rhythm, arcing around the central figure. On an ancient battlefield, dusty and torn, the reformer Shang Yang, clad in the deep black uniform of a Qin dynasty official, the simple embroidery on his collar and cuffs exudes a sense of authority amidst the chaos. His robes are stained with yellow dust, his cheeks stained with sweat and dust. His angular face, lips pursed, his meticulously tied bun slightly dishevelled, yet his deep, eagle-like eyes, piercing the dense smoke and dust, survey the battlefield with unwavering determination. The background is a purgatory of swords, swordplay, and the tumbling of men and horses. Soldiers clash fiercely, leaving broken weapons and fallen bodies scattered everywhere. The glaring sunlight, filtered through the thick smoke and dust, casts a scorching, warm yellow hue, imbuing the entire scene with a harsh, textured quality. As the camera moves steadily, Shang Yang holds the hilt of the ancient bronze sword tightly in one hand, while the other hand swings forward with a steady and decisive gesture, as if silently commanding thousands of troops. His unmoved command posture in the extreme chaos forms a sharp contrast with the bloody fighting around him, creating a tense, epic and solemn atmosphere.", + "seed": 6158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_7/index_940_seed_6158_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04750343970954418, + "vlm_logprob": -0.04750343970954418, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.04750343970954418, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06482340395450592 + ], + "opposite_logprob_tokens": [ + -2.8148233890533447 + ], + "logprob_reward": -0.06482340395450592, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06482340395450592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20281730592250824, + 0.05844082310795784, + 0.03255678340792656, + 0.03619460389018059, + 0.1505313217639923, + 0.03595685213804245, + 0.035149913281202316, + 0.11262986809015274, + 0.15346810221672058, + 0.18225443363189697 + ], + "credit_map": [ + 0.02442830801010132, + 0.028925299644470215, + 0.03290397301316261, + 0.022480176761746407, + 0.022845523431897163, + 0.022675830870866776, + 0.026250218972563744, + 0.022307969629764557, + 0.006024140398949385, + 0.006715261843055487, + 0.010580570437014103, + 0.00670791557058692, + 0.006274460814893246, + 0.00613901624456048, + 0.006777363363653421, + 0.009222099557518959, + 0.003026483114808798, + 0.005120369605720043, + 0.0050242142751812935, + 0.0037961036432534456, + 0.0034199394285678864, + 0.003414218546822667, + 0.005938732530921698, + 0.002816722495481372, + 0.005346701946109533, + 0.005904421675950289, + 0.003498474135994911, + 0.003648436861112714, + 0.0037950044497847557, + 0.0052067493088543415, + 0.004827098920941353, + 0.003967718221247196, + 0.044108957052230835, + 0.04715175926685333, + 0.012232177890837193, + 0.0055504171177744865, + 0.013202405534684658, + 0.010526676662266254, + 0.009183543734252453, + 0.008575394749641418, + 0.005285519175231457, + 0.0038574247155338526, + 0.00898786075413227, + 0.0038796206936240196, + 0.002942033112049103, + 0.0034326480235904455, + 0.0029711229726672173, + 0.0046006240881979465, + 0.00411200150847435, + 0.003681252710521221, + 0.003954884596168995, + 0.004105615429580212, + 0.004824977368116379, + 0.00566798634827137, + 0.004530330188572407, + 0.004272863734513521, + 0.017495673149824142, + 0.00789673626422882, + 0.012958988547325134, + 0.009266238659620285, + 0.012573356740176678, + 0.029657788574695587, + 0.010225286707282066, + 0.0125558041036129, + 0.0176344383507967, + 0.015354787930846214, + 0.017721619457006454, + 0.017914213240146637, + 0.026630518957972527, + 0.015902500599622726, + 0.02811427041888237, + 0.014195757918059826, + 0.01751534268260002, + 0.02390195243060589, + 0.019315140321850777, + 0.0204208604991436, + 0.03162836283445358, + 0.023501403629779816, + 0.027877651154994965, + 0.01809372939169407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.06482340395450592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200181424617767 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.018200181424617767, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018200181424617767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14372186362743378, + 0.04503537714481354, + 0.04400055110454559, + 0.04193924367427826, + 0.07109580934047699, + 0.060721833258867264, + 0.034159645438194275, + 0.16225884854793549, + 0.12047705054283142, + 0.2765897512435913 + ], + "credit_map": [ + 0.018073316663503647, + 0.015337150543928146, + 0.030049428343772888, + 0.016998497769236565, + 0.014016331173479557, + 0.015065544284880161, + 0.01986159384250641, + 0.014319990761578083, + 0.004786294419318438, + 0.006524636410176754, + 0.007124688010662794, + 0.006608252879232168, + 0.004159603267908096, + 0.00441202986985445, + 0.0059006717056035995, + 0.005519202910363674, + 0.004689369350671768, + 0.010649396106600761, + 0.007979687303304672, + 0.003904947079718113, + 0.0029190245550125837, + 0.0059511177241802216, + 0.005268387030810118, + 0.002638621721416712, + 0.004150650463998318, + 0.006039219442754984, + 0.004481143783777952, + 0.005082071293145418, + 0.003870931686833501, + 0.005278970114886761, + 0.006922143045812845, + 0.006114116869866848, + 0.01550464890897274, + 0.01667197048664093, + 0.008323346264660358, + 0.003478200174868107, + 0.005516787990927696, + 0.010473889298737049, + 0.005199403967708349, + 0.005927564110606909, + 0.01557568646967411, + 0.006140902638435364, + 0.008535606786608696, + 0.008915074169635773, + 0.0059223901480436325, + 0.004314019810408354, + 0.006313009187579155, + 0.005005148705095053, + 0.004086862783879042, + 0.003307134611532092, + 0.0035111973993480206, + 0.0046894727274775505, + 0.005702400580048561, + 0.004601834341883659, + 0.005307168699800968, + 0.0029535761568695307, + 0.031780801713466644, + 0.009824669919908047, + 0.014034358784556389, + 0.01176074892282486, + 0.008205884136259556, + 0.05469789728522301, + 0.011562278494238853, + 0.020392203703522682, + 0.012592105194926262, + 0.012417446821928024, + 0.02433004416525364, + 0.014940704219043255, + 0.012478180229663849, + 0.013146735727787018, + 0.013936132192611694, + 0.016635702922940254, + 0.025454629212617874, + 0.04186931625008583, + 0.03259306028485298, + 0.028888793662190437, + 0.04731824994087219, + 0.041149888187646866, + 0.03711199015378952, + 0.022203806787729263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.018200181424617767 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897589355707169 + ], + "opposite_logprob_tokens": [ + -2.5789759159088135 + ], + "logprob_reward": -0.07897589355707169, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897589355707169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21584098041057587, + 0.08845482766628265, + 0.04701521620154381, + 0.06178363040089607, + 0.0650797188282013, + 0.08524563163518906, + 0.03200523927807808, + 0.12608902156352997, + 0.1658545583486557, + 0.1126311719417572 + ], + "credit_map": [ + 0.02143196389079094, + 0.03138205409049988, + 0.03281959891319275, + 0.025573400780558586, + 0.026217125356197357, + 0.02806594967842102, + 0.025408251211047173, + 0.02494261972606182, + 0.008599336259067059, + 0.014306717552244663, + 0.01361229456961155, + 0.011484230868518353, + 0.011257713660597801, + 0.008560548536479473, + 0.012760680168867111, + 0.007873310707509518, + 0.004854144062846899, + 0.006876232568174601, + 0.006963019259274006, + 0.005338224582374096, + 0.004783506039530039, + 0.006334654986858368, + 0.007649629842489958, + 0.004215804859995842, + 0.006482899188995361, + 0.009099442511796951, + 0.006386716850101948, + 0.007617778144776821, + 0.006993121933192015, + 0.008058382198214531, + 0.010708359070122242, + 0.006436923984438181, + 0.012305399402976036, + 0.015062060207128525, + 0.009088118560612202, + 0.005293521098792553, + 0.006733802612870932, + 0.007773990277200937, + 0.004321346990764141, + 0.0045014796778559685, + 0.01133520994335413, + 0.007330539636313915, + 0.013169510290026665, + 0.014025215059518814, + 0.009631776250898838, + 0.012950380332767963, + 0.008385411463677883, + 0.008417589589953423, + 0.0044859787449240685, + 0.0034176374319940805, + 0.0037799726705998182, + 0.0043511721305549145, + 0.005252343602478504, + 0.003362490562722087, + 0.0033867366146296263, + 0.0039689065888524055, + 0.013309594243764877, + 0.01350349560379982, + 0.019733291119337082, + 0.017141040414571762, + 0.017784420400857925, + 0.011549694463610649, + 0.016678452491760254, + 0.016389038413763046, + 0.01617939956486225, + 0.01599043421447277, + 0.03185151889920235, + 0.02733493037521839, + 0.01574363186955452, + 0.01841539517045021, + 0.020146967843174934, + 0.02019229158759117, + 0.01046814676374197, + 0.014135017059743404, + 0.019583970308303833, + 0.011519404128193855, + 0.010544446296989918, + 0.02445806935429573, + 0.01242094486951828, + 0.009501166641712189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.07897589355707169 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577752962708473 + ], + "opposite_logprob_tokens": [ + -3.8955776691436768 + ], + "logprob_reward": -0.020577752962708473, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577752962708473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5690277814865112, + 0.06550576537847519, + 0.09149639308452606, + 0.04111932963132858, + 0.034144170582294464, + 0.07145056873559952, + 0.021544262766838074, + 0.03764360770583153, + 0.041936539113521576, + 0.026131581515073776 + ], + "credit_map": [ + 0.07077472656965256, + 0.10090070217847824, + 0.128875270485878, + 0.05479259788990021, + 0.03007584810256958, + 0.06735070794820786, + 0.03517335653305054, + 0.08108455687761307, + 0.007786262780427933, + 0.008724204264581203, + 0.014064357616007328, + 0.007537062745541334, + 0.0046446602791547775, + 0.007434796541929245, + 0.007678123190999031, + 0.007636300288140774, + 0.012049143202602863, + 0.021285662427544594, + 0.016424309462308884, + 0.00898800790309906, + 0.0071122050285339355, + 0.011216159909963608, + 0.0090824319049716, + 0.005338467191904783, + 0.006149188615381718, + 0.007428818382322788, + 0.004766823258250952, + 0.0034642883110791445, + 0.003198023419827223, + 0.007778774946928024, + 0.005291517823934555, + 0.0030418988317251205, + 0.006213193293660879, + 0.0058380793780088425, + 0.006201396696269512, + 0.0017168201738968492, + 0.002835517283529043, + 0.004941351246088743, + 0.002249406185001135, + 0.004148404113948345, + 0.009476425126194954, + 0.009466247633099556, + 0.015045576728880405, + 0.00947282463312149, + 0.004383584950119257, + 0.008371300995349884, + 0.006094960495829582, + 0.00913963932543993, + 0.002948778448626399, + 0.0021734642796218395, + 0.0021517740096896887, + 0.002328878501430154, + 0.0026289676316082478, + 0.0030271525029093027, + 0.00360257551074028, + 0.0026826721150428057, + 0.004577144980430603, + 0.0034135421738028526, + 0.006636005360633135, + 0.0038717694114893675, + 0.005513278301805258, + 0.005322408862411976, + 0.004333812743425369, + 0.0039756447076797485, + 0.004319263156503439, + 0.0045283786021173, + 0.005935845896601677, + 0.005422394257038832, + 0.006875212304294109, + 0.004867457319051027, + 0.005705379415303469, + 0.004282607231289148, + 0.0022643981501460075, + 0.0035647291224449873, + 0.003357304260134697, + 0.002292163670063019, + 0.0036161262542009354, + 0.004204110708087683, + 0.003191706258803606, + 0.003641044022515416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.020577752962708473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a man wearing a deep black uniform with simple embroidery on the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's face stained with yellow dust and sweat, with an angular face and pursed lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding the hilt of an ancient bronze sword in one hand while swinging the other hand forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background filled with soldiers clashing, broken weapons, and fallen bodies in a dusty battlefield setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan in a slow, steady rhythm, arcing around the central figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493996664881706 + ], + "opposite_logprob_tokens": [ + -2.9299399852752686 + ], + "logprob_reward": -0.05493996664881706, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493996664881706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12147621810436249, + 0.045149315148591995, + 0.03213850036263466, + 0.045663584023714066, + 0.034350357949733734, + 0.04795539006590843, + 0.08383656293153763, + 0.4276452362537384, + 0.0957227498292923, + 0.06606212258338928 + ], + "credit_map": [ + 0.016549749299883842, + 0.017074571922421455, + 0.023799682036042213, + 0.012289296835660934, + 0.011720573529601097, + 0.013250065967440605, + 0.010080091655254364, + 0.01671217940747738, + 0.005053014028817415, + 0.007713432423770428, + 0.007617097347974777, + 0.005964966025203466, + 0.0027961188461631536, + 0.004546022973954678, + 0.005174525082111359, + 0.006284133531153202, + 0.003637793706730008, + 0.004586631432175636, + 0.0033086356706917286, + 0.003922865726053715, + 0.003737145569175482, + 0.003966016229242086, + 0.006370167247951031, + 0.0026092426851391792, + 0.00550280325114727, + 0.00740239629521966, + 0.004905404523015022, + 0.005171731114387512, + 0.003538374323397875, + 0.006748775485903025, + 0.0068739247508347034, + 0.005520170088857412, + 0.004269697237759829, + 0.004343945533037186, + 0.008783149532973766, + 0.002438515657559037, + 0.003189191920682788, + 0.004676612094044685, + 0.0023086683358997107, + 0.004340573213994503, + 0.0070036365650594234, + 0.005367606412619352, + 0.008994562551379204, + 0.004456992261111736, + 0.005260830279439688, + 0.007667864207178354, + 0.003720221109688282, + 0.005483668763190508, + 0.013355939649045467, + 0.0062392824329435825, + 0.012555907480418682, + 0.009765089489519596, + 0.010614492930471897, + 0.011062861420214176, + 0.010230151005089283, + 0.01001283060759306, + 0.07891049236059189, + 0.02932392805814743, + 0.040086597204208374, + 0.03595878928899765, + 0.03663727268576622, + 0.1245715469121933, + 0.032659295946359634, + 0.049497246742248535, + 0.010931411758065224, + 0.010318593122065067, + 0.011766387149691582, + 0.012396343052387238, + 0.01579234190285206, + 0.00943672750145197, + 0.014519166201353073, + 0.010561775416135788, + 0.006607200484722853, + 0.010234967805445194, + 0.008114907890558243, + 0.007011214271187782, + 0.009867196902632713, + 0.010740232653915882, + 0.008380922488868237, + 0.005105473101139069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.05493996664881706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_8/index_tensor([9119, 9119])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_8/index_tensor([9119, 9119])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa79b7fc84a5059638d1bf1a2159d4cf8b89eec7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_8/index_tensor([9119, 9119])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot taken from a level angle captures a chubby orange tabby cat curled up in a perfect circle, peacefully sleeping on a wide wooden windowsill. Warm afternoon sunlight streams through the old-fashioned window panes, casting a brilliant golden halo around the edges of the cat's soft, fluffy fur, making its orange and white stripes appear particularly vibrant. Its eyes are closed, its small pink nose twitches slightly, its whiskers still in the air, and its entire body rises and falls in a barely perceptible, rhythmic rhythm with its deep, steady breathing. The sunlight casts soft, long shadows on the slightly mottled white lacquer of the windowsill, and tiny dust particles floating in the air glint in the beams of light. The entire scene is immersed in a tranquil, nostalgic atmosphere composed of golden light and warm colors, exuding a warm, vintage style.", + "seed": 13053, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_8/index_9119_seed_13053_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0257331270724535, + "vlm_logprob": -0.0257331270724535, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.0257331270724535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014394726604223251 + ], + "opposite_logprob_tokens": [ + -4.889394760131836 + ], + "logprob_reward": -0.014394726604223251, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.014394726604223251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2621767818927765, + 0.0905284732580185, + 0.12267803400754929, + 0.04639437794685364, + 0.058921072632074356, + 0.05583657696843147, + 0.05825424939393997, + 0.10011059045791626, + 0.12043493986129761, + 0.08466491103172302 + ], + "credit_map": [ + 0.028866684064269066, + 0.06777837127447128, + 0.03161994740366936, + 0.0317864865064621, + 0.030273910611867905, + 0.017528191208839417, + 0.025724638253450394, + 0.02859858050942421, + 0.011992696672677994, + 0.01665467582643032, + 0.010615856386721134, + 0.011510148644447327, + 0.011001769453287125, + 0.00878827553242445, + 0.010020555928349495, + 0.009944487363100052, + 0.015583637170493603, + 0.026667166501283646, + 0.015607415698468685, + 0.012172543443739414, + 0.012987328693270683, + 0.011252840049564838, + 0.016115982085466385, + 0.012291112914681435, + 0.004124004859477282, + 0.006704062223434448, + 0.00491888215765357, + 0.007223440334200859, + 0.00445296335965395, + 0.003751420183107257, + 0.008408124558627605, + 0.006811478640884161, + 0.008522804826498032, + 0.009070279076695442, + 0.0062042404897511005, + 0.006530283018946648, + 0.0066106584854424, + 0.005448841955512762, + 0.009955279529094696, + 0.006578685715794563, + 0.005598314572125673, + 0.009663082659244537, + 0.005991790443658829, + 0.008685537613928318, + 0.008142570033669472, + 0.004848119802772999, + 0.006747219245880842, + 0.006159939803183079, + 0.007781321182847023, + 0.01174213271588087, + 0.00497856829315424, + 0.007100927177816629, + 0.007043301127851009, + 0.005049302242696285, + 0.006673313211649656, + 0.007885381579399109, + 0.011194092221558094, + 0.02916634827852249, + 0.010688986629247665, + 0.00955977849662304, + 0.010871694423258305, + 0.00807526521384716, + 0.010098608210682869, + 0.010455816984176636, + 0.014039412140846252, + 0.037277016788721085, + 0.011143393814563751, + 0.010856865905225277, + 0.012881070375442505, + 0.011027362197637558, + 0.008371804840862751, + 0.014838016591966152, + 0.012344861403107643, + 0.014317825436592102, + 0.007061048410832882, + 0.009440993890166283, + 0.011426545679569244, + 0.008173122070729733, + 0.011974351480603218, + 0.009926159866154194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.014394726604223251 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108629535883665 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.01108629535883665, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108629535883665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2176128476858139, + 0.14899955689907074, + 0.11711585521697998, + 0.046660009771585464, + 0.08418919891119003, + 0.08527164906263351, + 0.06388325989246368, + 0.07713846117258072, + 0.06818755716085434, + 0.09094163030385971 + ], + "credit_map": [ + 0.019161373376846313, + 0.03296990692615509, + 0.029891863465309143, + 0.038404207676649094, + 0.024236520752310753, + 0.015145850367844105, + 0.02780458889901638, + 0.029998522251844406, + 0.01392973493784666, + 0.01988787017762661, + 0.024581842124462128, + 0.020466942340135574, + 0.016333943232893944, + 0.011834218166768551, + 0.02005409263074398, + 0.021910903975367546, + 0.010500880889594555, + 0.021145274862647057, + 0.01416891347616911, + 0.019434567540884018, + 0.012408223003149033, + 0.009194944053888321, + 0.014120123349130154, + 0.016142921522259712, + 0.0050824955105781555, + 0.009301615878939629, + 0.004213353618979454, + 0.005437731742858887, + 0.00644833967089653, + 0.003944852389395237, + 0.006354546640068293, + 0.005877070594578981, + 0.0073602525517344475, + 0.021885650232434273, + 0.008313172496855259, + 0.009688865393400192, + 0.009432747960090637, + 0.007459378335624933, + 0.011436485685408115, + 0.008612648583948612, + 0.01183678675442934, + 0.023805713281035423, + 0.008280867710709572, + 0.006975496653467417, + 0.00946212001144886, + 0.008269041776657104, + 0.007411419879645109, + 0.009230198338627815, + 0.0075271595269441605, + 0.015015299431979656, + 0.005431778728961945, + 0.006668072659522295, + 0.007717224769294262, + 0.005420779809355736, + 0.008993800729513168, + 0.007109140977263451, + 0.00712753226980567, + 0.013691358268260956, + 0.010533983819186687, + 0.010895775631070137, + 0.006722891237586737, + 0.008749482221901417, + 0.010575985535979271, + 0.008841444738209248, + 0.009720160625874996, + 0.01195561233907938, + 0.006238649133592844, + 0.008511742576956749, + 0.007725874427706003, + 0.00754909124225378, + 0.008570635691285133, + 0.00791578646749258, + 0.014387570321559906, + 0.015551887452602386, + 0.007390689570456743, + 0.009209760464727879, + 0.011537333019077778, + 0.010195792652666569, + 0.012221917510032654, + 0.010446670465171337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01108629535883665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036202622577548027 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036202622577548027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036202622577548027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25776830315589905, + 0.12958474457263947, + 0.15941067039966583, + 0.05750105157494545, + 0.12617464363574982, + 0.08544474840164185, + 0.0388018935918808, + 0.055474571883678436, + 0.062210407108068466, + 0.02762889303267002 + ], + "credit_map": [ + 0.03992396965622902, + 0.04534502699971199, + 0.0317426472902298, + 0.020615991204977036, + 0.027913717553019524, + 0.029774414375424385, + 0.03520480915904045, + 0.027247760444879532, + 0.020129866898059845, + 0.018920963630080223, + 0.016944753006100655, + 0.011309551075100899, + 0.01497931033372879, + 0.015723688527941704, + 0.017731236293911934, + 0.013845391571521759, + 0.020890135318040848, + 0.022686028853058815, + 0.021458089351654053, + 0.016522563993930817, + 0.016890540719032288, + 0.014922073110938072, + 0.026865068823099136, + 0.01917618326842785, + 0.006387099623680115, + 0.008883722126483917, + 0.006937946658581495, + 0.006761075928807259, + 0.006866422016173601, + 0.006061158142983913, + 0.009606495499610901, + 0.005997137166559696, + 0.01888684183359146, + 0.019180821254849434, + 0.01471585500985384, + 0.01073808129876852, + 0.014544468373060226, + 0.014674520120024681, + 0.019395632669329643, + 0.01403842680156231, + 0.012857439927756786, + 0.01695309765636921, + 0.010483012534677982, + 0.00795686710625887, + 0.00898818951100111, + 0.009889615699648857, + 0.008188994601368904, + 0.010127535089850426, + 0.004709111992269754, + 0.008619669824838638, + 0.003857729723677039, + 0.005071853753179312, + 0.004038538783788681, + 0.004013108089566231, + 0.0028370930813252926, + 0.0056547909043729305, + 0.006779339164495468, + 0.01035140361636877, + 0.007423629052937031, + 0.005576514173299074, + 0.005797538440674543, + 0.004785525146871805, + 0.008725653402507305, + 0.006034976802766323, + 0.008888723328709602, + 0.012475231662392616, + 0.007046850863844156, + 0.0058520277962088585, + 0.006584674119949341, + 0.007131957449018955, + 0.007279393263161182, + 0.006951551418751478, + 0.004228975158184767, + 0.0035863309167325497, + 0.0030703311786055565, + 0.0024221453350037336, + 0.0033146855421364307, + 0.0031233374029397964, + 0.0040330407209694386, + 0.0038500467780977488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0036202622577548027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899010181427002 + ], + "opposite_logprob_tokens": [ + -2.5789899826049805 + ], + "logprob_reward": -0.07899010181427002, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899010181427002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18482138216495514, + 0.06660735607147217, + 0.11534283310174942, + 0.07747028023004532, + 0.0801873728632927, + 0.06047266349196434, + 0.13219241797924042, + 0.12231412529945374, + 0.10653968155384064, + 0.05405184626579285 + ], + "credit_map": [ + 0.02044629491865635, + 0.042655326426029205, + 0.020147649571299553, + 0.019694996997714043, + 0.0210576169192791, + 0.01448935829102993, + 0.024168526753783226, + 0.022161614149808884, + 0.007052138913422823, + 0.014722493477165699, + 0.008092975243926048, + 0.00934239849448204, + 0.008158441632986069, + 0.005142982117831707, + 0.0076631768606603146, + 0.006432750262320042, + 0.016346214339137077, + 0.018549956381320953, + 0.013244877569377422, + 0.01843762770295143, + 0.011358913965523243, + 0.0111204469576478, + 0.014100427739322186, + 0.012184367515146732, + 0.008190863765776157, + 0.023107249289751053, + 0.006262928247451782, + 0.00896601751446724, + 0.008695426397025585, + 0.005618027877062559, + 0.008663465268909931, + 0.007966297678649426, + 0.00997123122215271, + 0.015504442155361176, + 0.009676053188741207, + 0.010489650070667267, + 0.00768137676641345, + 0.007000309880822897, + 0.0105355279520154, + 0.009328777901828289, + 0.006489886902272701, + 0.0081957271322608, + 0.008544334210455418, + 0.01085527054965496, + 0.006287294905632734, + 0.0058997273445129395, + 0.007367502897977829, + 0.0068329195491969585, + 0.018304720520973206, + 0.0165634136646986, + 0.01636444218456745, + 0.01737588830292225, + 0.013530472293496132, + 0.016326889395713806, + 0.015468759462237358, + 0.01825782097876072, + 0.016896499320864677, + 0.01853632740676403, + 0.01418541744351387, + 0.012068558484315872, + 0.013579078018665314, + 0.01354251429438591, + 0.019398357719182968, + 0.01410737819969654, + 0.01421641930937767, + 0.024644488468766212, + 0.011465209536254406, + 0.011718546971678734, + 0.011336521245539188, + 0.011321062222123146, + 0.010158668272197247, + 0.011678770184516907, + 0.003658832050859928, + 0.015639731660485268, + 0.00453551858663559, + 0.007068340200930834, + 0.005735892802476883, + 0.0032218005508184433, + 0.006117464974522591, + 0.008074267767369747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.07899010181427002 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057424932718277 + ], + "opposite_logprob_tokens": [ + -3.8955743312835693 + ], + "logprob_reward": -0.02057424932718277, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057424932718277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21123413741588593, + 0.04728725180029869, + 0.07859884947538376, + 0.04313070327043533, + 0.10056035965681076, + 0.11379338800907135, + 0.10927771031856537, + 0.1531430035829544, + 0.09941236674785614, + 0.04356221854686737 + ], + "credit_map": [ + 0.020159253850579262, + 0.069203682243824, + 0.017823098227381706, + 0.022120971232652664, + 0.024225100874900818, + 0.01824299991130829, + 0.019263410940766335, + 0.020195623859763145, + 0.005295541137456894, + 0.008652863092720509, + 0.00693920161575079, + 0.006514811888337135, + 0.004639455582946539, + 0.004862524569034576, + 0.00557224964722991, + 0.004810607060790062, + 0.011101695708930492, + 0.011874756775796413, + 0.008919652551412582, + 0.006668295245617628, + 0.009082207456231117, + 0.00848312210291624, + 0.012905082665383816, + 0.009564036503434181, + 0.004988967441022396, + 0.008218345232307911, + 0.003896190784871578, + 0.0054533579386770725, + 0.004640104714781046, + 0.004783793818205595, + 0.004059209488332272, + 0.007090733852237463, + 0.014439590275287628, + 0.020251421257853508, + 0.010544932447373867, + 0.010242640972137451, + 0.01103382371366024, + 0.011443235911428928, + 0.01047554612159729, + 0.012129170820116997, + 0.016823729500174522, + 0.02705814689397812, + 0.010695305652916431, + 0.010579508729279041, + 0.012021821923553944, + 0.012623409740626812, + 0.009490192867815495, + 0.014501269906759262, + 0.015133234672248363, + 0.02725474163889885, + 0.012935804203152657, + 0.008374522440135479, + 0.011368263512849808, + 0.009927473962306976, + 0.012724812142550945, + 0.011558853089809418, + 0.02155889943242073, + 0.026606649160385132, + 0.017964789643883705, + 0.013228807598352432, + 0.01670590415596962, + 0.01647508330643177, + 0.02296620048582554, + 0.017636671662330627, + 0.007764544803649187, + 0.03570449352264404, + 0.008224854245781898, + 0.009957090951502323, + 0.011848589405417442, + 0.005547392647713423, + 0.009341680444777012, + 0.01102372631430626, + 0.004445970524102449, + 0.013128376565873623, + 0.004046451300382614, + 0.004994802176952362, + 0.004853945691138506, + 0.0035069521982222795, + 0.003663662588223815, + 0.0049220542423427105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.02057424932718277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot taken from a level angle captures a chubby orange tabby cat curled up in a perfect circle, peacefully sleeping on a wide wooden windowsill. Warm afternoon sunlight streams through the old-fashioned window panes, casting a brilliant golden halo around the edges of the cat's soft, fluffy fur, making its orange and white stripes appear particularly vibrant. Its eyes are closed, its small pink nose twitches slightly, its whiskers still in the air, and its entire body rises and falls in a barely perceptible, rhythmic rhythm with its deep, steady breathing. The sunlight casts soft, long shadows on the slightly mottled white lacquer of the windowsill, and tiny dust particles floating in the air glint in the beams of light. The entire scene is immersed in a tranquil, nostalgic atmosphere composed of golden light and warm colors, exuding a warm, vintage style.", + "seed": 13054, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_8/index_9119_seed_13054_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022855877596884965, + "vlm_logprob": -0.022855877596884965, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.022855877596884965, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025834834203124046 + ], + "opposite_logprob_tokens": [ + -4.025835037231445 + ], + "logprob_reward": -0.025834834203124046, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.025834834203124046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5858176946640015, + 0.05403997376561165, + 0.023253563791513443, + 0.0326479971408844, + 0.01335231214761734, + 0.06393589079380035, + 0.02927863784134388, + 0.01484596822410822, + 0.01553319487720728, + 0.16729475557804108 + ], + "credit_map": [ + 0.046575479209423065, + 0.025099143385887146, + 0.17197024822235107, + 0.22951608896255493, + 0.019023694097995758, + 0.029650051146745682, + 0.050767090171575546, + 0.013215873390436172, + 0.005562013015151024, + 0.004221975803375244, + 0.008480671793222427, + 0.01840396225452423, + 0.00336457509547472, + 0.004268350545316935, + 0.007268018554896116, + 0.0024704034440219402, + 0.001762183615937829, + 0.001831395784392953, + 0.0050864736549556255, + 0.004978707060217857, + 0.0023827403783798218, + 0.0024116700515151024, + 0.0029709329828619957, + 0.0018294613109901547, + 0.002754140179604292, + 0.001866613165475428, + 0.0070309764705598354, + 0.011107835918664932, + 0.0015010085189715028, + 0.0031515476293861866, + 0.003522403771057725, + 0.0017134720692411065, + 0.0013266138266772032, + 0.001215929165482521, + 0.002035170793533325, + 0.0029087576549500227, + 0.001135370577685535, + 0.0015054774703457952, + 0.002144566038623452, + 0.001080426387488842, + 0.004687033127993345, + 0.0026137710083276033, + 0.021229304373264313, + 0.024049539119005203, + 0.0016902002971619368, + 0.0025040709879249334, + 0.005437400657683611, + 0.0017245725030079484, + 0.0036326085682958364, + 0.0018965295748785138, + 0.004855920560657978, + 0.010303737595677376, + 0.0015184420626610518, + 0.0021761285606771708, + 0.0033695665188133717, + 0.0015257025370374322, + 0.0017309781396761537, + 0.0014430618612095714, + 0.0018567886436358094, + 0.002571320626884699, + 0.001528618624433875, + 0.002373988041654229, + 0.0021595756988972425, + 0.001181635889224708, + 0.0013396932044997811, + 0.0011095801601186395, + 0.0031337772961705923, + 0.0041532618924975395, + 0.0011213164543733, + 0.001226869528181851, + 0.0022397153079509735, + 0.0012089824303984642, + 0.013613338582217693, + 0.006938579957932234, + 0.05006781592965126, + 0.06714959442615509, + 0.003914335276931524, + 0.008095961064100266, + 0.014152977615594864, + 0.003362160176038742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.025834834203124046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552540749311447 + ], + "opposite_logprob_tokens": [ + -4.387552738189697 + ], + "logprob_reward": -0.012552540749311447, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552540749311447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4878377616405487, + 0.047619786113500595, + 0.08165735751390457, + 0.03571632131934166, + 0.06234182044863701, + 0.10072930157184601, + 0.03971870243549347, + 0.024608226493000984, + 0.025218278169631958, + 0.09455247968435287 + ], + "credit_map": [ + 0.041578661650419235, + 0.022251445800065994, + 0.12333321571350098, + 0.19100849330425262, + 0.014926795847713947, + 0.03274061530828476, + 0.048060547560453415, + 0.013937986455857754, + 0.004991031717509031, + 0.004391486756503582, + 0.006877786945551634, + 0.015234713442623615, + 0.0035022671800106764, + 0.004309107083827257, + 0.005424532573670149, + 0.0028888594824820757, + 0.0071217031218111515, + 0.003618527902290225, + 0.018262343481183052, + 0.025176385417580605, + 0.004584173671901226, + 0.010300376452505589, + 0.007390767335891724, + 0.005203081760555506, + 0.0048921070992946625, + 0.0018744218396022916, + 0.005797569639980793, + 0.011501249857246876, + 0.0018339632079005241, + 0.0042506796307861805, + 0.0034124429803341627, + 0.002153889276087284, + 0.007527639623731375, + 0.0029517188668251038, + 0.011965121142566204, + 0.02524859458208084, + 0.0018823114223778248, + 0.004898282699286938, + 0.006015247665345669, + 0.0018529121298342943, + 0.008596810512244701, + 0.004225348588079214, + 0.0296010784804821, + 0.040756795555353165, + 0.002262290334329009, + 0.004530464764684439, + 0.008616644889116287, + 0.0021398733370006084, + 0.005241505801677704, + 0.002266010968014598, + 0.0074811712838709354, + 0.014703413471579552, + 0.0014557897811755538, + 0.0023365961387753487, + 0.004041815642267466, + 0.0021923999302089214, + 0.003449610434472561, + 0.0015121838077902794, + 0.0021802058909088373, + 0.00724145770072937, + 0.0015495768748223782, + 0.004072217736393213, + 0.0029720189049839973, + 0.0016309557249769568, + 0.002851547673344612, + 0.0014936593361198902, + 0.002670199377462268, + 0.007195355836302042, + 0.002056117169559002, + 0.004322155378758907, + 0.0026937753427773714, + 0.0019354693358764052, + 0.007212153170257807, + 0.0037988093681633472, + 0.028816424310207367, + 0.037609223276376724, + 0.0023477524518966675, + 0.005060435272753239, + 0.00786709226667881, + 0.0018405938753858209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.012552540749311447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102386999875307 + ], + "opposite_logprob_tokens": [ + -5.5041022300720215 + ], + "logprob_reward": -0.004102386999875307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102386999875307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6566657423973083, + 0.039554741233587265, + 0.023217692971229553, + 0.023567458614706993, + 0.06082529574632645, + 0.08384966850280762, + 0.026558123528957367, + 0.02732405997812748, + 0.013347527012228966, + 0.045089587569236755 + ], + "credit_map": [ + 0.05480628088116646, + 0.03052743338048458, + 0.16483932733535767, + 0.25983917713165283, + 0.02187909372150898, + 0.05200626701116562, + 0.056914158165454865, + 0.01585404761135578, + 0.005961088929325342, + 0.0035219599958509207, + 0.005770959425717592, + 0.006050459109246731, + 0.0037120291963219643, + 0.0047964188270270824, + 0.005968417041003704, + 0.0037734101060777903, + 0.0031773087102919817, + 0.001980704255402088, + 0.0031144802924245596, + 0.003430494572967291, + 0.002364435000345111, + 0.003992829006165266, + 0.0029756727162748575, + 0.002181770047172904, + 0.0023123505525290966, + 0.0013570289593189955, + 0.004880341235548258, + 0.006488296203315258, + 0.001524949329905212, + 0.002868733601644635, + 0.0025317268446087837, + 0.0016040299087762833, + 0.006643231026828289, + 0.002683382946997881, + 0.01216907612979412, + 0.02306411601603031, + 0.0023031621240079403, + 0.0056178756058216095, + 0.005613614805042744, + 0.002730830805376172, + 0.006338692735880613, + 0.0033691104035824537, + 0.02735724486410618, + 0.033005181699991226, + 0.0019173651235178113, + 0.0029921166133135557, + 0.007059083320200443, + 0.0018108676886186004, + 0.003972246311604977, + 0.001852623769082129, + 0.00302499788813293, + 0.008065362460911274, + 0.0014452298637479544, + 0.0027024024166166782, + 0.003405422205105424, + 0.0020898382645100355, + 0.0033098983112722635, + 0.0015639736084267497, + 0.004102753940969706, + 0.008899621665477753, + 0.0016623043920844793, + 0.0031121461652219296, + 0.0032765218056738377, + 0.0013968400890007615, + 0.001613487838767469, + 0.0008288478129543364, + 0.0019619613885879517, + 0.0038989956956356764, + 0.0009079115698114038, + 0.001597136608324945, + 0.0017032577889040112, + 0.0008359284838661551, + 0.0038314983248710632, + 0.0019338304409757257, + 0.012160825543105602, + 0.017816966399550438, + 0.0013018082827329636, + 0.0029255361296236515, + 0.004009203985333443, + 0.001109920791350305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.004102386999875307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099825382232666 + ], + "opposite_logprob_tokens": [ + -3.288099765777588 + ], + "logprob_reward": -0.038099825382232666, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099825382232666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6930014491081238, + 0.04475535452365875, + 0.03253173828125, + 0.009242295287549496, + 0.051960889250040054, + 0.011445661075413227, + 0.006661094259470701, + 0.02656077966094017, + 0.006509918253868818, + 0.11733084172010422 + ], + "credit_map": [ + 0.053717635571956635, + 0.02901325188577175, + 0.20021390914916992, + 0.27585989236831665, + 0.019900048151612282, + 0.03925801068544388, + 0.05845329910516739, + 0.016585344448685646, + 0.0050919619388878345, + 0.002863615518435836, + 0.008344207890331745, + 0.014915556646883488, + 0.002726614009588957, + 0.0030895937234163284, + 0.005390401463955641, + 0.0023334012366831303, + 0.0030263939406722784, + 0.001454186625778675, + 0.00830149371176958, + 0.01234421506524086, + 0.0012645231327041984, + 0.0018520912854000926, + 0.003150181844830513, + 0.0011386530240997672, + 0.0008989981142804027, + 0.0008269859827123582, + 0.0013853682903572917, + 0.002362258965149522, + 0.0008799427305348217, + 0.0007055328460410237, + 0.0014394903555512428, + 0.0007437172462232411, + 0.005572139751166105, + 0.002381418365985155, + 0.010896909981966019, + 0.019935131072998047, + 0.0018245070241391659, + 0.004177944269031286, + 0.004849962890148163, + 0.002322869375348091, + 0.0012939635198563337, + 0.0007550940499641001, + 0.0025451998226344585, + 0.0019766236655414104, + 0.0008465145365335047, + 0.0014539456460624933, + 0.0015611599665135145, + 0.0010131580056622624, + 0.0010011361446231604, + 0.0006031482480466366, + 0.0007584794075228274, + 0.0012422287836670876, + 0.0006577518652193248, + 0.0005158866988494992, + 0.0012911587255075574, + 0.0005913036875426769, + 0.003103220835328102, + 0.001459050108678639, + 0.004496586509048939, + 0.00943907443434, + 0.0013205184368416667, + 0.0025244811549782753, + 0.0028204733971506357, + 0.0013973708264529705, + 0.0008062411216087639, + 0.0004876114253420383, + 0.0009288263390772045, + 0.0015675552422180772, + 0.000502052076626569, + 0.0007420628680847585, + 0.0009537213481962681, + 0.0005218475707806647, + 0.008887596428394318, + 0.004637842997908592, + 0.03775995224714279, + 0.046686168760061264, + 0.0026116068474948406, + 0.0048545231111347675, + 0.009671620093286037, + 0.00222152192145586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.038099825382232666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368980064988136 + ], + "opposite_logprob_tokens": [ + -3.4086897373199463 + ], + "logprob_reward": -0.03368980064988136, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368980064988136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7557045817375183, + 0.03308148309588432, + 0.03517014533281326, + 0.007516131736338139, + 0.022278999909758568, + 0.03257887437939644, + 0.00835575070232153, + 0.019170217216014862, + 0.008937489241361618, + 0.07720626145601273 + ], + "credit_map": [ + 0.06055547669529915, + 0.02982720360159874, + 0.23375768959522247, + 0.304589182138443, + 0.018854936584830284, + 0.0320727676153183, + 0.06101308390498161, + 0.015034297481179237, + 0.003963920753449202, + 0.0014846579870209098, + 0.007370466366410255, + 0.011008189991116524, + 0.0013515123864635825, + 0.002168444450944662, + 0.0035820237826555967, + 0.002152268774807453, + 0.0031874729320406914, + 0.0017198511632159352, + 0.008253617212176323, + 0.013028951361775398, + 0.00170806550886482, + 0.0027441633865237236, + 0.0032755553256720304, + 0.0012524662306532264, + 0.0009541510371491313, + 0.0007869557593949139, + 0.0010794255649670959, + 0.001032080501317978, + 0.0008084365981630981, + 0.0009928898653015494, + 0.001253595226444304, + 0.0006085975328460336, + 0.0026405586395412683, + 0.0015300720697268844, + 0.0033515961840748787, + 0.0076062423177063465, + 0.001443767687305808, + 0.0016451934352517128, + 0.0027867802418768406, + 0.0012747884029522538, + 0.0023600845597684383, + 0.0012474647955968976, + 0.01146969199180603, + 0.01103707030415535, + 0.0010246180463582277, + 0.0014753317227587104, + 0.0029179181437939405, + 0.0010466909734532237, + 0.0014029431622475386, + 0.0007084888638928533, + 0.001002342440187931, + 0.0019811466336250305, + 0.000712239125277847, + 0.0006382712163031101, + 0.0012065087212249637, + 0.000703810655977577, + 0.0023078983649611473, + 0.001671628444455564, + 0.0026378000620752573, + 0.004930545575916767, + 0.0016279141418635845, + 0.0017474074847996235, + 0.0027430166956037283, + 0.0015040059806779027, + 0.0010315331164747477, + 0.0009819333208724856, + 0.0012566958321258426, + 0.0016043157083913684, + 0.0007892428548075259, + 0.0008967297035269439, + 0.001567272236570716, + 0.0008097660029307008, + 0.005770685151219368, + 0.0031579674687236547, + 0.024340180680155754, + 0.030538706108927727, + 0.002035303972661495, + 0.0030077449046075344, + 0.006664317101240158, + 0.0016913600265979767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.03368980064988136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot taken from a level angle captures a chubby orange tabby cat curled up in a perfect circle, peacefully sleeping on a wide wooden windowsill. Warm afternoon sunlight streams through the old-fashioned window panes, casting a brilliant golden halo around the edges of the cat's soft, fluffy fur, making its orange and white stripes appear particularly vibrant. Its eyes are closed, its small pink nose twitches slightly, its whiskers still in the air, and its entire body rises and falls in a barely perceptible, rhythmic rhythm with its deep, steady breathing. The sunlight casts soft, long shadows on the slightly mottled white lacquer of the windowsill, and tiny dust particles floating in the air glint in the beams of light. The entire scene is immersed in a tranquil, nostalgic atmosphere composed of golden light and warm colors, exuding a warm, vintage style.", + "seed": 13055, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_8/index_9119_seed_13055_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028413727693259717, + "vlm_logprob": -0.028413727693259717, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.028413727693259717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019744873046875 + ], + "opposite_logprob_tokens": [ + -4.519744873046875 + ], + "logprob_reward": -0.019744873046875, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.019744873046875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3258785009384155, + 0.20864859223365784, + 0.10201478004455566, + 0.0664169043302536, + 0.05357809364795685, + 0.057740382850170135, + 0.06632456928491592, + 0.036993686109781265, + 0.05103335902094841, + 0.0313710942864418 + ], + "credit_map": [ + 0.03528463467955589, + 0.05582712963223457, + 0.029178159311413765, + 0.03459782525897026, + 0.04261980205774307, + 0.04541066661477089, + 0.04193371534347534, + 0.04102658852934837, + 0.026451347395777702, + 0.0298335999250412, + 0.03026224486529827, + 0.034594643861055374, + 0.01936185546219349, + 0.028203792870044708, + 0.01876992918550968, + 0.02117115817964077, + 0.01322524156421423, + 0.016741855069994926, + 0.013924609869718552, + 0.01144791767001152, + 0.009895090013742447, + 0.017947396263480186, + 0.009635595604777336, + 0.00919707678258419, + 0.007548804860562086, + 0.009139711037278175, + 0.008716650307178497, + 0.008918169885873795, + 0.006632609758526087, + 0.011841284111142159, + 0.006619110703468323, + 0.007000562269240618, + 0.006608879193663597, + 0.009260767139494419, + 0.0051662311889231205, + 0.0062745362520217896, + 0.006385794375091791, + 0.007393769454210997, + 0.0057571507059037685, + 0.006730964872986078, + 0.006660348270088434, + 0.01116880401968956, + 0.006365352310240269, + 0.007700799964368343, + 0.004994255490601063, + 0.008369436487555504, + 0.006040140520781279, + 0.006441246252506971, + 0.009898512624204159, + 0.009107392281293869, + 0.008053366094827652, + 0.0065955487079918385, + 0.007457311265170574, + 0.009961486794054508, + 0.007279601413756609, + 0.007971351034939289, + 0.005328896455466747, + 0.0057221767492592335, + 0.0034332366194576025, + 0.004040506202727556, + 0.0037285478319972754, + 0.0063430932350456715, + 0.0033956619445234537, + 0.005001564975827932, + 0.006791411433368921, + 0.008662655018270016, + 0.004129449836909771, + 0.004121687263250351, + 0.008188562467694283, + 0.007279496639966965, + 0.005975252483040094, + 0.005884848535060883, + 0.004458131734281778, + 0.004470120649784803, + 0.002859425498172641, + 0.0029048877768218517, + 0.003856744384393096, + 0.0042677344754338264, + 0.00360292661935091, + 0.0049511222168803215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.019744873046875 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067501612938940525 + ], + "opposite_logprob_tokens": [ + -5.006750106811523 + ], + "logprob_reward": -0.0067501612938940525, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067501612938940525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2211955040693283, + 0.11351138353347778, + 0.10195033252239227, + 0.08186708390712738, + 0.10171797126531601, + 0.1620292365550995, + 0.04057422652840614, + 0.05202966555953026, + 0.066068634390831, + 0.05905589088797569 + ], + "credit_map": [ + 0.030990073457360268, + 0.023540468886494637, + 0.025049060583114624, + 0.022381657734513283, + 0.028226692229509354, + 0.03891000524163246, + 0.02986229956150055, + 0.02223527990281582, + 0.016456706449389458, + 0.01982380822300911, + 0.012917784973978996, + 0.011952673085033894, + 0.014547346159815788, + 0.013657735660672188, + 0.013983335345983505, + 0.010172005742788315, + 0.013665162958204746, + 0.014194462448358536, + 0.011875252239406109, + 0.010307963006198406, + 0.011556752026081085, + 0.01691770926117897, + 0.01322421059012413, + 0.010208836756646633, + 0.009118907153606415, + 0.015273000113666058, + 0.007178718224167824, + 0.007435535080730915, + 0.012460485100746155, + 0.011709906160831451, + 0.010931737720966339, + 0.007758797146379948, + 0.01267552562057972, + 0.01763458549976349, + 0.010681363753974438, + 0.010088047944009304, + 0.01235968992114067, + 0.01671282947063446, + 0.011597029864788055, + 0.009968908503651619, + 0.020109981298446655, + 0.02959403395652771, + 0.01449575461447239, + 0.016046399250626564, + 0.018635224550962448, + 0.024223921820521355, + 0.024340543895959854, + 0.014583386480808258, + 0.006268776953220367, + 0.004452700726687908, + 0.0049680001102387905, + 0.004493896383792162, + 0.004060050472617149, + 0.007436871528625488, + 0.004074468277394772, + 0.004819468129426241, + 0.006351415999233723, + 0.005493893753737211, + 0.005746076349169016, + 0.005886375904083252, + 0.006107613909989595, + 0.01131860539317131, + 0.005589473061263561, + 0.005536218173801899, + 0.008042020723223686, + 0.010631841607391834, + 0.006942111067473888, + 0.00510178180411458, + 0.009824722073972225, + 0.01015304122120142, + 0.006518772337585688, + 0.008854355663061142, + 0.009337188675999641, + 0.01007034257054329, + 0.0058874767273664474, + 0.005992446094751358, + 0.007298219483345747, + 0.008317939005792141, + 0.007191271521151066, + 0.004961020313203335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0067501612938940525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956398788839579 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956398788839579, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956398788839579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3191174864768982, + 0.1266707479953766, + 0.09933529049158096, + 0.07503268122673035, + 0.09955089539289474, + 0.08584598451852798, + 0.05696014314889908, + 0.04072529822587967, + 0.04311557486653328, + 0.05364583060145378 + ], + "credit_map": [ + 0.044365596026182175, + 0.04333893209695816, + 0.032112590968608856, + 0.03023112192749977, + 0.03767121955752373, + 0.06378138065338135, + 0.030905939638614655, + 0.036710746586322784, + 0.019755562767386436, + 0.014016857370734215, + 0.01348091196268797, + 0.012831354513764381, + 0.012305492535233498, + 0.028665248304605484, + 0.012597219087183475, + 0.013018114492297173, + 0.014826500788331032, + 0.011540228500962257, + 0.009732939302921295, + 0.010068750940263271, + 0.010498101823031902, + 0.0215049646794796, + 0.010304300114512444, + 0.01085950993001461, + 0.008071193471550941, + 0.010028384625911713, + 0.008031100034713745, + 0.007660175208002329, + 0.009227641858160496, + 0.01482238620519638, + 0.008549977093935013, + 0.008641829714179039, + 0.014527051709592342, + 0.014638413675129414, + 0.009202741086483002, + 0.009331285953521729, + 0.010659261606633663, + 0.01785336248576641, + 0.011529364623129368, + 0.011809427291154861, + 0.011279664002358913, + 0.01721142791211605, + 0.007575510069727898, + 0.009313992224633694, + 0.00933132041245699, + 0.012080814689397812, + 0.009752357378602028, + 0.009300907142460346, + 0.008396309800446033, + 0.009254560805857182, + 0.004640606697648764, + 0.004838534630835056, + 0.006160730961710215, + 0.009982557967305183, + 0.0062520732171833515, + 0.0074347746558487415, + 0.006085025612264872, + 0.005469511263072491, + 0.004467213060706854, + 0.003897490445524454, + 0.0037681369576603174, + 0.007554216310381889, + 0.0040918271988630295, + 0.005391882266849279, + 0.005051743239164352, + 0.00634631235152483, + 0.004816124681383371, + 0.005875858012586832, + 0.005101339891552925, + 0.006257067434489727, + 0.004835927858948708, + 0.004831206053495407, + 0.007172739133238792, + 0.008091961033642292, + 0.004681298974901438, + 0.004683288279920816, + 0.00734785245731473, + 0.010008449666202068, + 0.005531041417270899, + 0.006129204295575619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005956398788839579 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904622495174408 + ], + "opposite_logprob_tokens": [ + -2.4640462398529053 + ], + "logprob_reward": -0.08904622495174408, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904622495174408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22442638874053955, + 0.08260851353406906, + 0.041213516145944595, + 0.037500765174627304, + 0.12509660422801971, + 0.23532912135124207, + 0.073366679251194, + 0.06901489943265915, + 0.06897387653589249, + 0.04246954247355461 + ], + "credit_map": [ + 0.03664775565266609, + 0.03780270740389824, + 0.023014117032289505, + 0.02053908444941044, + 0.02387762814760208, + 0.029310617595911026, + 0.029953239485621452, + 0.023281250149011612, + 0.012072578072547913, + 0.013765707612037659, + 0.007019680924713612, + 0.009805198758840561, + 0.008723235689103603, + 0.011876721866428852, + 0.01203986257314682, + 0.007305528502911329, + 0.0059862625785171986, + 0.005728116724640131, + 0.003332369029521942, + 0.003267369233071804, + 0.004090775270015001, + 0.007707098498940468, + 0.004491765517741442, + 0.006609759759157896, + 0.005118002649396658, + 0.005867402534931898, + 0.004015113692730665, + 0.004257178399711847, + 0.004032438155263662, + 0.005302316043525934, + 0.0054913475178182125, + 0.003416964551433921, + 0.01719691790640354, + 0.023047877475619316, + 0.011596682481467724, + 0.013508669100701809, + 0.014425414614379406, + 0.017511850222945213, + 0.016443634405732155, + 0.011365555226802826, + 0.031730618327856064, + 0.03713783994317055, + 0.022212142124772072, + 0.023401733487844467, + 0.02384120412170887, + 0.03919095918536186, + 0.028396328911185265, + 0.029418298974633217, + 0.009157783351838589, + 0.011742787435650826, + 0.006530799437314272, + 0.006707941647619009, + 0.008192358538508415, + 0.012585614807903767, + 0.00908522866666317, + 0.009364165365695953, + 0.008910704404115677, + 0.011784601956605911, + 0.006169960368424654, + 0.006956189870834351, + 0.0064738159999251366, + 0.011169230565428734, + 0.009099438786506653, + 0.008450958877801895, + 0.008844472467899323, + 0.010826120153069496, + 0.006558703258633614, + 0.007036923430860043, + 0.006495741195976734, + 0.01192645262926817, + 0.008164208382368088, + 0.00912125688046217, + 0.005575512070208788, + 0.008817951194941998, + 0.003818029770627618, + 0.004054022021591663, + 0.005423772614449263, + 0.005369807593524456, + 0.004243265837430954, + 0.0051671783439815044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.08904622495174408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057098038494587 + ], + "opposite_logprob_tokens": [ + -3.895570993423462 + ], + "logprob_reward": -0.02057098038494587, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057098038494587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2955072522163391, + 0.055652935057878494, + 0.054488394409418106, + 0.060790736228227615, + 0.09717974066734314, + 0.09104406833648682, + 0.05053042620420456, + 0.07825897634029388, + 0.14778856933116913, + 0.06875894963741302 + ], + "credit_map": [ + 0.03561428189277649, + 0.04362793266773224, + 0.025473229587078094, + 0.022173574194312096, + 0.042783789336681366, + 0.04903467744588852, + 0.03312041237950325, + 0.04367933049798012, + 0.006371143274009228, + 0.010844253934919834, + 0.004686097148805857, + 0.005811038892716169, + 0.006852029822766781, + 0.008108342997729778, + 0.006759208161383867, + 0.006220814306288958, + 0.007287842687219381, + 0.006692992523312569, + 0.004936159588396549, + 0.004022909328341484, + 0.006860523018985987, + 0.011745890602469444, + 0.006566488184034824, + 0.006375583820044994, + 0.006557939108461142, + 0.010348249226808548, + 0.005090681370347738, + 0.007271356415003538, + 0.008660350926220417, + 0.008391737006604671, + 0.009078926406800747, + 0.005391492508351803, + 0.011452622711658478, + 0.015521521680057049, + 0.011342369951307774, + 0.010116292163729668, + 0.010847775265574455, + 0.014908269979059696, + 0.009795336052775383, + 0.013195549137890339, + 0.01193963922560215, + 0.014297179877758026, + 0.010731888003647327, + 0.012889116071164608, + 0.00967616681009531, + 0.011486231349408627, + 0.011722518131136894, + 0.008301324211061, + 0.0070185791701078415, + 0.007591521367430687, + 0.005190662108361721, + 0.005613976623862982, + 0.004839227069169283, + 0.008626431226730347, + 0.006394135765731335, + 0.005255891475826502, + 0.00861548725515604, + 0.012103920802474022, + 0.006595198065042496, + 0.00648903613910079, + 0.0110976817086339, + 0.014039666391909122, + 0.009300307370722294, + 0.010017668828368187, + 0.021683530882000923, + 0.02106567844748497, + 0.015814345329999924, + 0.016045480966567993, + 0.013798203319311142, + 0.025899427011609077, + 0.01666959747672081, + 0.01681230030953884, + 0.008238254114985466, + 0.010972656309604645, + 0.0059669953770935535, + 0.006511797662824392, + 0.010857989080250263, + 0.011477132327854633, + 0.006223074626177549, + 0.008511044085025787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.02057098038494587 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot taken from a level angle captures a chubby orange tabby cat curled up in a perfect circle, peacefully sleeping on a wide wooden windowsill. Warm afternoon sunlight streams through the old-fashioned window panes, casting a brilliant golden halo around the edges of the cat's soft, fluffy fur, making its orange and white stripes appear particularly vibrant. Its eyes are closed, its small pink nose twitches slightly, its whiskers still in the air, and its entire body rises and falls in a barely perceptible, rhythmic rhythm with its deep, steady breathing. The sunlight casts soft, long shadows on the slightly mottled white lacquer of the windowsill, and tiny dust particles floating in the air glint in the beams of light. The entire scene is immersed in a tranquil, nostalgic atmosphere composed of golden light and warm colors, exuding a warm, vintage style.", + "seed": 13056, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_8/index_9119_seed_13056_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04736008830368519, + "vlm_logprob": -0.04736008830368519, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.04736008830368519, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020881570875644684 + ], + "opposite_logprob_tokens": [ + -4.520881652832031 + ], + "logprob_reward": -0.020881570875644684, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.020881570875644684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3485074043273926, + 0.0753951296210289, + 0.03290187567472458, + 0.03000892698764801, + 0.06506489962339401, + 0.025851180776953697, + 0.1326880306005478, + 0.06825387477874756, + 0.0681084394454956, + 0.15322023630142212 + ], + "credit_map": [ + 0.027117528021335602, + 0.03788266330957413, + 0.06019388884305954, + 0.05016094073653221, + 0.049348004162311554, + 0.026208585128188133, + 0.048973310738801956, + 0.0486224964261055, + 0.00663649570196867, + 0.008267052471637726, + 0.016102498397231102, + 0.010829782113432884, + 0.00884181261062622, + 0.004530960228294134, + 0.009748021140694618, + 0.010438504628837109, + 0.002589584095403552, + 0.00395160960033536, + 0.005669393576681614, + 0.004577350802719593, + 0.004551225807517767, + 0.0024752833414822817, + 0.004618152976036072, + 0.004469272214919329, + 0.002433336805552244, + 0.0036754996981471777, + 0.005865996237844229, + 0.002912599593400955, + 0.003812426468357444, + 0.002860080450773239, + 0.0033653213176876307, + 0.005083662923425436, + 0.0052562435157597065, + 0.007392799481749535, + 0.014499183744192123, + 0.0075940522365272045, + 0.006350345443934202, + 0.0037289655301719904, + 0.01205114834010601, + 0.008192156441509724, + 0.0017491290345788002, + 0.004034634679555893, + 0.0019878114107996225, + 0.005597020965069532, + 0.0025558634661138058, + 0.00239733443595469, + 0.0038061507511883974, + 0.0037232344038784504, + 0.009185679256916046, + 0.01645735278725624, + 0.026617126539349556, + 0.02130196988582611, + 0.01520402729511261, + 0.0083844605833292, + 0.01647041365504265, + 0.01906699314713478, + 0.005166336428374052, + 0.009768512099981308, + 0.009427935816347599, + 0.016688181087374687, + 0.005870289634913206, + 0.004043681547045708, + 0.00960344448685646, + 0.007685493677854538, + 0.0058364663273096085, + 0.008880500681698322, + 0.013313780538737774, + 0.01185314729809761, + 0.005916605703532696, + 0.0040717157535254955, + 0.009659040719270706, + 0.008577175438404083, + 0.013733318075537682, + 0.01893303170800209, + 0.021809358149766922, + 0.026858795434236526, + 0.01893150620162487, + 0.011428210884332657, + 0.021133797243237495, + 0.02039221301674843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.020881570875644684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029828520491719246 + ], + "opposite_logprob_tokens": [ + -3.5298285484313965 + ], + "logprob_reward": -0.029828520491719246, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029828520491719246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22941797971725464, + 0.11082232743501663, + 0.05112951993942261, + 0.038321346044540405, + 0.08702661842107773, + 0.046185772866010666, + 0.12827198207378387, + 0.14175163209438324, + 0.11511168628931046, + 0.051961153745651245 + ], + "credit_map": [ + 0.01952429860830307, + 0.02843649871647358, + 0.03992868959903717, + 0.035637080669403076, + 0.026959313079714775, + 0.01896679773926735, + 0.0318099670112133, + 0.028155343607068062, + 0.011062046512961388, + 0.01416363287717104, + 0.016774537041783333, + 0.02049672231078148, + 0.012519627809524536, + 0.008604004047811031, + 0.015043990686535835, + 0.012157764285802841, + 0.0038869306445121765, + 0.006899076513946056, + 0.006341342814266682, + 0.008367461152374744, + 0.004680682439357042, + 0.005348690785467625, + 0.007231648080050945, + 0.0083736851811409, + 0.0032482799142599106, + 0.006499960087239742, + 0.005412077531218529, + 0.005323145072907209, + 0.0038234840612858534, + 0.0034450863022357225, + 0.005818979814648628, + 0.004750332795083523, + 0.007129731588065624, + 0.009885472245514393, + 0.01784036122262478, + 0.012743283063173294, + 0.006220408249646425, + 0.006294514983892441, + 0.013790728524327278, + 0.013122121803462505, + 0.004407200030982494, + 0.005691110622137785, + 0.0067320154048502445, + 0.008064951747655869, + 0.004611798096448183, + 0.002955483039841056, + 0.0062390645034611225, + 0.00748414546251297, + 0.00960003025829792, + 0.018279772251844406, + 0.019314590841531754, + 0.02062087692320347, + 0.011269581504166126, + 0.009111830964684486, + 0.01893923431634903, + 0.021136069670319557, + 0.011361298151314259, + 0.016818974167108536, + 0.029653292149305344, + 0.014107838273048401, + 0.014891710132360458, + 0.009943487122654915, + 0.021956363692879677, + 0.023018663749098778, + 0.009162128902971745, + 0.013663035817444324, + 0.02720792219042778, + 0.012470231391489506, + 0.011936330236494541, + 0.007728058844804764, + 0.019231155514717102, + 0.01371282059699297, + 0.004369598813354969, + 0.006945065688341856, + 0.009357975795865059, + 0.007189162541180849, + 0.005756649188697338, + 0.004252652171999216, + 0.0071507906541228294, + 0.006939256563782692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.029828520491719246 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202170729637146 + ], + "opposite_logprob_tokens": [ + -4.018202304840088 + ], + "logprob_reward": -0.018202170729637146, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018202170729637146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.348666787147522, + 0.12384870648384094, + 0.080121248960495, + 0.044680871069431305, + 0.03714417666196823, + 0.025103703141212463, + 0.16264201700687408, + 0.07501398026943207, + 0.04645553603768349, + 0.05632294341921806 + ], + "credit_map": [ + 0.03193335980176926, + 0.04434628039598465, + 0.0586235411465168, + 0.04967575520277023, + 0.046237342059612274, + 0.02547253668308258, + 0.04509915038943291, + 0.04727881774306297, + 0.009154495783150196, + 0.01337090041488409, + 0.02488173171877861, + 0.019226906821131706, + 0.011408220045268536, + 0.008718143217265606, + 0.01467302069067955, + 0.022415289655327797, + 0.007270583882927895, + 0.008954903110861778, + 0.010904544033110142, + 0.011160221882164478, + 0.009554189629852772, + 0.00876692496240139, + 0.00921406876295805, + 0.014295817352831364, + 0.0033574551343917847, + 0.004945426248013973, + 0.0060432967729866505, + 0.004965754225850105, + 0.004946900065988302, + 0.005021854303777218, + 0.005969882942736149, + 0.009430301375687122, + 0.0030384028796106577, + 0.004009193740785122, + 0.005706353113055229, + 0.004926386754959822, + 0.004610718227922916, + 0.003662083763629198, + 0.00692943437024951, + 0.004261604510247707, + 0.0028576035983860493, + 0.003197086276486516, + 0.0018454859964549541, + 0.004396745003759861, + 0.002321249106898904, + 0.0029654158279299736, + 0.0027560964226722717, + 0.004764019511640072, + 0.011646239086985588, + 0.0225460696965456, + 0.030807776376605034, + 0.02232104353606701, + 0.01766630820930004, + 0.010589118115603924, + 0.02160503715276718, + 0.025460433214902878, + 0.00664322916418314, + 0.009777376428246498, + 0.011837439611554146, + 0.008986431173980236, + 0.009186375886201859, + 0.005399039015173912, + 0.010589373297989368, + 0.012594715692102909, + 0.004538825713098049, + 0.005822107661515474, + 0.009353475645184517, + 0.006112454459071159, + 0.003829661523923278, + 0.0030151247046887875, + 0.007946178317070007, + 0.005837709177285433, + 0.0036928155459463596, + 0.005575708579272032, + 0.011309143155813217, + 0.008272157981991768, + 0.006494338624179363, + 0.004661834333091974, + 0.008519806899130344, + 0.007797140162438154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.018202170729637146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295001953840256 + ], + "opposite_logprob_tokens": [ + -2.237950086593628 + ], + "logprob_reward": -0.11295001953840256, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295001953840256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3129440248012543, + 0.07085207849740982, + 0.030751660466194153, + 0.0381053201854229, + 0.07705259323120117, + 0.055436067283153534, + 0.19169916212558746, + 0.08151620626449585, + 0.07011907547712326, + 0.07152380794286728 + ], + "credit_map": [ + 0.029086053371429443, + 0.03746388107538223, + 0.04201957955956459, + 0.034058284014463425, + 0.03989275172352791, + 0.027295904234051704, + 0.034159865230321884, + 0.06896771490573883, + 0.008958227932453156, + 0.010432218201458454, + 0.008991323411464691, + 0.01183874811977148, + 0.007438248489052057, + 0.007773383520543575, + 0.00657295435667038, + 0.00884697213768959, + 0.004058916121721268, + 0.004825259558856487, + 0.00277989124879241, + 0.005833735223859549, + 0.0028376616537570953, + 0.003321913769468665, + 0.003684315364807844, + 0.003409968689084053, + 0.0045689805410802364, + 0.004615544341504574, + 0.002709814114496112, + 0.005954936146736145, + 0.0038279201835393906, + 0.0038874512538313866, + 0.00625961646437645, + 0.006281055044382811, + 0.005129662808030844, + 0.008932637982070446, + 0.01864353008568287, + 0.009527277201414108, + 0.007463831454515457, + 0.006079849787056446, + 0.012785039842128754, + 0.008490762673318386, + 0.004682454746216536, + 0.00676339864730835, + 0.00971687026321888, + 0.009841266088187695, + 0.004548380617052317, + 0.002774106105789542, + 0.009329807944595814, + 0.007779780309647322, + 0.013956185430288315, + 0.02481057308614254, + 0.0392167866230011, + 0.026236139237880707, + 0.02230258658528328, + 0.01256269309669733, + 0.025911759585142136, + 0.026702437549829483, + 0.005768636707216501, + 0.011553055606782436, + 0.01429982203990221, + 0.014798402786254883, + 0.006819312926381826, + 0.006383421365171671, + 0.01221535261720419, + 0.00967820081859827, + 0.005612136796116829, + 0.00826064683496952, + 0.016438717022538185, + 0.008115729317069054, + 0.006567619740962982, + 0.005216829478740692, + 0.012106134556233883, + 0.007801264524459839, + 0.004705977626144886, + 0.008785496465861797, + 0.014199373312294483, + 0.009968996979296207, + 0.0070323399268090725, + 0.006179881747812033, + 0.007899054326117039, + 0.012752688489854336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.11295001953840256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a chubby orange tabby cat curled in a circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat sleeping on a wide wooden windowsill with mottled white lacquer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight stream through old-fashioned window panes creating a golden halo on the cat's fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny dust particles visible floating in the air and glinting in the light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493815988302231 + ], + "opposite_logprob_tokens": [ + -2.9299380779266357 + ], + "logprob_reward": -0.05493815988302231, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493815988302231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14828325808048248, + 0.05237595736980438, + 0.039026401937007904, + 0.03597686067223549, + 0.08140364289283752, + 0.04216082766652107, + 0.28788191080093384, + 0.05358659848570824, + 0.12690752744674683, + 0.13239696621894836 + ], + "credit_map": [ + 0.016766108572483063, + 0.01991971768438816, + 0.013264276087284088, + 0.02737985923886299, + 0.01796099916100502, + 0.014663648791611195, + 0.01971633918583393, + 0.01861230842769146, + 0.004212218802422285, + 0.005772184580564499, + 0.006372678559273481, + 0.005787230562418699, + 0.0061089880764484406, + 0.005492162425071001, + 0.008471200242638588, + 0.010159295983612537, + 0.0038229769561439753, + 0.005396956577897072, + 0.002916715806350112, + 0.0043165599927306175, + 0.003996087238192558, + 0.0044728233478963375, + 0.005149135831743479, + 0.008955149911344051, + 0.003819830249994993, + 0.004953274503350258, + 0.004549016244709492, + 0.005232271272689104, + 0.003932720050215721, + 0.0028742582071572542, + 0.005811252165585756, + 0.00480424240231514, + 0.005800250917673111, + 0.010177899152040482, + 0.016376957297325134, + 0.010697646997869015, + 0.006921672262251377, + 0.005733958911150694, + 0.011058101430535316, + 0.014637162908911705, + 0.003876646049320698, + 0.00492460373789072, + 0.0063069844618439674, + 0.0051236264407634735, + 0.0036184547934681177, + 0.003662389237433672, + 0.0069315615110099316, + 0.00771656259894371, + 0.021406041458249092, + 0.03509349003434181, + 0.05893498286604881, + 0.042387910187244415, + 0.03397035226225853, + 0.017761319875717163, + 0.04324960708618164, + 0.0350782610476017, + 0.005027902778238058, + 0.00710381381213665, + 0.004017707891762257, + 0.01315811462700367, + 0.005080614238977432, + 0.0033345443662256002, + 0.008370712399482727, + 0.007493191864341497, + 0.01093450840562582, + 0.015701016411185265, + 0.026798874139785767, + 0.013980413787066936, + 0.013656644150614738, + 0.008582877926528454, + 0.021982921287417412, + 0.015270278789103031, + 0.0109854219481349, + 0.01743793673813343, + 0.02816205844283104, + 0.014961841516196728, + 0.013396629132330418, + 0.011152501218020916, + 0.016526462510228157, + 0.01977412961423397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.05493815988302231 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_9/index_tensor([3890, 3890])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_9/index_tensor([3890, 3890])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5a4faa63a46b9b88ca5edc09dfeb4b7f0bd004e5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_9/index_tensor([3890, 3890])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a cozy, humanistic interior, a fluid single shot begins at table level, then slowly and gracefully pans rightward across a warm dining scene. A long, dark wooden table, covered with a beige linen tablecloth, is laden with half-eaten home-cooked dishes and scattered crumbs. Several friends, dressed in cozy sweaters and casual shirts, gather around the table. Their faces are blurred by a soft bokeh, yet their genuine smiles and lively gestures, engaged in conversation, are clearly visible. As the camera pans, the focus gradually and seamlessly shifts from the joyful gathering in the background to the central subject in the foreground: a woman's hand, delicately lifting a glass of white wine from the table. Steady and composed, her nails immaculately manicured, the hand elegantly grasps the slender stem of a classic crystal goblet. The glass is about one-third full of pale golden white wine, and under the soft light cast by a low, warm-toned chandelier overhead, the wine's clarity is translucent, shimmering with sparkling shimmer. As the wine glasses are slowly raised to the center of the picture, the smiling faces and figures of the friends in the background are completely blurred into warm spots of light (Bokeh) and soft blocks of color. The whole picture is wrapped in an intimate, peaceful and joyful atmosphere. The main colors are warm amber and cream, and the contrast of light and shadow is soft, perfectly capturing this peaceful and sincere moment of celebration.", + "seed": 29048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_9/index_3890_seed_29048_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024692443758249284, + "vlm_logprob": -0.024692443758249284, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.024692443758249284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023985689505934715 + ], + "opposite_logprob_tokens": [ + -4.648985862731934 + ], + "logprob_reward": -0.023985689505934715, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.023985689505934715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23056253790855408, + 0.06564466655254364, + 0.03646525740623474, + 0.3053536117076874, + 0.04305034503340721, + 0.03967694193124771, + 0.04953499138355255, + 0.10567454248666763, + 0.08827686309814453, + 0.035760216414928436 + ], + "credit_map": [ + 0.021356984972953796, + 0.031542208045721054, + 0.027997052296996117, + 0.030955657362937927, + 0.0244105476886034, + 0.038359902799129486, + 0.03227344900369644, + 0.02366672269999981, + 0.009506521746516228, + 0.013798344880342484, + 0.007511007599532604, + 0.005483722779899836, + 0.008276009000837803, + 0.00933806598186493, + 0.005577277857810259, + 0.006153716705739498, + 0.004372600466012955, + 0.007027207873761654, + 0.00329038267955184, + 0.005522295366972685, + 0.00451253866776824, + 0.003589772852137685, + 0.003666920820251107, + 0.004483538214117289, + 0.1126793697476387, + 0.017812350764870644, + 0.013434022665023804, + 0.018952850252389908, + 0.010478327050805092, + 0.10007476061582565, + 0.010955564677715302, + 0.020966339856386185, + 0.010234703309834003, + 0.004147429950535297, + 0.00463801808655262, + 0.0038357467856258154, + 0.004021150525659323, + 0.004411181900650263, + 0.00407393928617239, + 0.00768817076459527, + 0.004483524709939957, + 0.0036184084601700306, + 0.008329713717103004, + 0.0036442119162529707, + 0.004130421672016382, + 0.005012851674109697, + 0.0040594288147985935, + 0.006398383062332869, + 0.010093004442751408, + 0.00514827948063612, + 0.005978121887892485, + 0.0036358903162181377, + 0.006715910509228706, + 0.0057428446598351, + 0.0038031828589737415, + 0.008417759090662003, + 0.012218469753861427, + 0.008584337309002876, + 0.008760357275605202, + 0.008000897243618965, + 0.009036191739141941, + 0.02553503029048443, + 0.011066236533224583, + 0.022473018616437912, + 0.019078239798545837, + 0.006015472579747438, + 0.007107493933290243, + 0.006543186493217945, + 0.008232075721025467, + 0.01297775935381651, + 0.014842632226645947, + 0.013480008579790592, + 0.006236246321350336, + 0.003319827141240239, + 0.003951488062739372, + 0.0025097711477428675, + 0.004602378699928522, + 0.003411069745197892, + 0.008112083189189434, + 0.003617352107539773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.023985689505934715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032106302678585052 + ], + "opposite_logprob_tokens": [ + -5.753210544586182 + ], + "logprob_reward": -0.0032106302678585052, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032106302678585052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15783822536468506, + 0.15243256092071533, + 0.043694499880075455, + 0.11791452765464783, + 0.07952885329723358, + 0.06524457037448883, + 0.09522021561861038, + 0.07417246699333191, + 0.11542992293834686, + 0.09852417558431625 + ], + "credit_map": [ + 0.018517954275012016, + 0.019183820113539696, + 0.016023864969611168, + 0.015588030219078064, + 0.015624498948454857, + 0.0255079735070467, + 0.016151173040270805, + 0.031240902841091156, + 0.021094338968396187, + 0.020714590325951576, + 0.025371603667736053, + 0.01692787930369377, + 0.015929672867059708, + 0.016260676085948944, + 0.017573390156030655, + 0.01856040395796299, + 0.00602180790156126, + 0.006130878813564777, + 0.005026333965361118, + 0.006157285999506712, + 0.0048164743930101395, + 0.005050237290561199, + 0.005120437126606703, + 0.005371044389903545, + 0.03519979119300842, + 0.010154057294130325, + 0.006351320538669825, + 0.008960992097854614, + 0.007082940544933081, + 0.02942105568945408, + 0.0066091581247746944, + 0.014135207049548626, + 0.02301696501672268, + 0.010403929278254509, + 0.007825098000466824, + 0.005893797613680363, + 0.007219614926725626, + 0.0067711700685322285, + 0.00459930719807744, + 0.013798972591757774, + 0.007537711877375841, + 0.006646926514804363, + 0.009249432012438774, + 0.008497528731822968, + 0.006112429779022932, + 0.009982912801206112, + 0.007486037909984589, + 0.009731585159897804, + 0.009137625806033611, + 0.023745130747556686, + 0.012742227874696255, + 0.00734814815223217, + 0.008337162435054779, + 0.011387256905436516, + 0.006390866823494434, + 0.016131795942783356, + 0.009510645642876625, + 0.007702907547354698, + 0.009894792921841145, + 0.006956963334232569, + 0.01008420530706644, + 0.013447624631226063, + 0.006756442133337259, + 0.00981888361275196, + 0.01780278980731964, + 0.006475160364061594, + 0.009546359069645405, + 0.011314893141388893, + 0.013508272357285023, + 0.016659123823046684, + 0.02701278030872345, + 0.013110537081956863, + 0.012221661396324635, + 0.007172172889113426, + 0.008532536216080189, + 0.006173452362418175, + 0.020166506990790367, + 0.014251377433538437, + 0.013551183044910431, + 0.016455283388495445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0032106302678585052 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494357645511627 + ], + "opposite_logprob_tokens": [ + -2.929943561553955 + ], + "logprob_reward": -0.05494357645511627, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494357645511627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24418826401233673, + 0.07245594263076782, + 0.04988117516040802, + 0.24858279526233673, + 0.06027057021856308, + 0.10092021524906158, + 0.08625016361474991, + 0.05239854007959366, + 0.054245080798864365, + 0.030807236209511757 + ], + "credit_map": [ + 0.021471133455634117, + 0.030751464888453484, + 0.032822128385305405, + 0.032995209097862244, + 0.022633885964751244, + 0.04298868030309677, + 0.030588988214731216, + 0.029936788603663445, + 0.013165228068828583, + 0.01429693028330803, + 0.010017970576882362, + 0.005880516022443771, + 0.006162902805954218, + 0.00912332534790039, + 0.005305232014507055, + 0.008503840304911137, + 0.005027409177273512, + 0.008406038396060467, + 0.004325117915868759, + 0.0103257205337286, + 0.004140191711485386, + 0.005332125350832939, + 0.006405794993042946, + 0.005918778013437986, + 0.07664545625448227, + 0.015524697490036488, + 0.012663649395108223, + 0.016576087102293968, + 0.010886549018323421, + 0.08032519370317459, + 0.014857769012451172, + 0.021103395149111748, + 0.008595641702413559, + 0.006461774464696646, + 0.009632773697376251, + 0.006732319016009569, + 0.006866975221782923, + 0.006792095024138689, + 0.006892655044794083, + 0.00829633790999651, + 0.010256155394017696, + 0.0062935794703662395, + 0.033126551657915115, + 0.011400625109672546, + 0.00646792259067297, + 0.01536029577255249, + 0.006820269860327244, + 0.01119481772184372, + 0.01313704065978527, + 0.02162834443151951, + 0.0061199236661195755, + 0.003973487298935652, + 0.006030725780874491, + 0.009338230825960636, + 0.007350992411375046, + 0.018671423196792603, + 0.006728006526827812, + 0.004004607908427715, + 0.006382104940712452, + 0.003024701029062271, + 0.0038304647896438837, + 0.014041847549378872, + 0.005371913779526949, + 0.009014888666570187, + 0.009916498325765133, + 0.0035524186678230762, + 0.005359980743378401, + 0.005628643557429314, + 0.006248584948480129, + 0.007078817114233971, + 0.00929782260209322, + 0.007162314839661121, + 0.003814432304352522, + 0.003211340168491006, + 0.003593054134398699, + 0.0026020873337984085, + 0.00454825209453702, + 0.003246533451601863, + 0.004961276426911354, + 0.004830258898437023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.05494357645511627 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369337320327759 + ], + "opposite_logprob_tokens": [ + -3.408693313598633 + ], + "logprob_reward": -0.03369337320327759, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369337320327759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17159296572208405, + 0.22166816890239716, + 0.06855979561805725, + 0.08929459750652313, + 0.05351032316684723, + 0.07276066392660141, + 0.056778501719236374, + 0.1845591515302658, + 0.048823729157447815, + 0.032452039420604706 + ], + "credit_map": [ + 0.015806643292307854, + 0.022795574739575386, + 0.0173263531178236, + 0.023919740691781044, + 0.012669091112911701, + 0.03172780200839043, + 0.023222988471388817, + 0.024124788120388985, + 0.04074956104159355, + 0.060892265290021896, + 0.025402989238500595, + 0.01800515502691269, + 0.01662723906338215, + 0.021523263305425644, + 0.01948857493698597, + 0.0189791452139616, + 0.008505554869771004, + 0.012190471403300762, + 0.005367910955101252, + 0.01094024907797575, + 0.007973246276378632, + 0.006845807656645775, + 0.00780839379876852, + 0.008928172290325165, + 0.02384111098945141, + 0.009607690386474133, + 0.006844665389508009, + 0.005543362814933062, + 0.009800789877772331, + 0.017957502976059914, + 0.005715391598641872, + 0.009984097443521023, + 0.009546754881739616, + 0.004787123296409845, + 0.0049062082543969154, + 0.00415817741304636, + 0.004192885477095842, + 0.005541959311813116, + 0.0040822518058121204, + 0.016294967383146286, + 0.011708363890647888, + 0.00589334312826395, + 0.01451800111681223, + 0.013464325107634068, + 0.0042340406216681, + 0.0056013138964772224, + 0.010348044335842133, + 0.006993238348513842, + 0.007485144305974245, + 0.005129641387611628, + 0.01000153087079525, + 0.004462400916963816, + 0.004577803425490856, + 0.005484044086188078, + 0.008464314043521881, + 0.01117362268269062, + 0.024778615683317184, + 0.011713491752743721, + 0.01906610280275345, + 0.013744402676820755, + 0.02288701757788658, + 0.0319560170173645, + 0.01857527531683445, + 0.04183823987841606, + 0.006477105896919966, + 0.0031697615049779415, + 0.005350419785827398, + 0.006512284744530916, + 0.004920244682580233, + 0.007408433128148317, + 0.008050634525716305, + 0.006934850011020899, + 0.0043859961442649364, + 0.001989847980439663, + 0.004072710406035185, + 0.002569776028394699, + 0.006373787298798561, + 0.0035501273814588785, + 0.0037875326815992594, + 0.005722262896597385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.03369337320327759 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628949359059334 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.007628949359059334, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628949359059334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05978682264685631, + 0.08413642644882202, + 0.07443810254335403, + 0.3846229314804077, + 0.028890516608953476, + 0.21187713742256165, + 0.03337611258029938, + 0.05924610421061516, + 0.04030900448560715, + 0.023316804319620132 + ], + "credit_map": [ + 0.006575427018105984, + 0.009637919254601002, + 0.008163770660758018, + 0.006796372588723898, + 0.006551808677613735, + 0.009579649195075035, + 0.005491490475833416, + 0.006990383844822645, + 0.010808444581925869, + 0.016552399843931198, + 0.009723060764372349, + 0.009734943509101868, + 0.008519481867551804, + 0.009789522737264633, + 0.0089624784886837, + 0.010046101175248623, + 0.007050335872918367, + 0.010294496081769466, + 0.006417304277420044, + 0.01551605574786663, + 0.008408309891819954, + 0.006445861421525478, + 0.00923154130578041, + 0.011074204929172993, + 0.09355925023555756, + 0.039901088923215866, + 0.0229530930519104, + 0.024766955524683, + 0.04333234578371048, + 0.10709623992443085, + 0.023569481447339058, + 0.02944449707865715, + 0.0036256248131394386, + 0.0038780246395617723, + 0.0029943983536213636, + 0.0029075206257402897, + 0.004574490711092949, + 0.004222193732857704, + 0.0027340396773070097, + 0.003954226151108742, + 0.024024613201618195, + 0.014445839449763298, + 0.049678508192300797, + 0.019680961966514587, + 0.024944329634308815, + 0.03622842952609062, + 0.018179360777139664, + 0.02469511330127716, + 0.004249102436006069, + 0.008933753706514835, + 0.004232003353536129, + 0.0028272001072764397, + 0.0026589208282530308, + 0.0035616701934486628, + 0.0022216164506971836, + 0.0046918499283492565, + 0.00727065047249198, + 0.004535190295428038, + 0.007969656027853489, + 0.003972380422055721, + 0.00734350923448801, + 0.006935115437954664, + 0.0056180162355303764, + 0.01560158934444189, + 0.004901173058897257, + 0.0028388192877173424, + 0.0035728083457797766, + 0.004210414364933968, + 0.00475272024050355, + 0.006841375958174467, + 0.009371290914714336, + 0.0038204053416848183, + 0.0032630469650030136, + 0.002235105959698558, + 0.0020794961601495743, + 0.0016703816363587976, + 0.0034400091972202063, + 0.0028451334219425917, + 0.0023485852871090174, + 0.005435049533843994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.007628949359059334 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a cozy, humanistic interior, a fluid single shot begins at table level, then slowly and gracefully pans rightward across a warm dining scene. A long, dark wooden table, covered with a beige linen tablecloth, is laden with half-eaten home-cooked dishes and scattered crumbs. Several friends, dressed in cozy sweaters and casual shirts, gather around the table. Their faces are blurred by a soft bokeh, yet their genuine smiles and lively gestures, engaged in conversation, are clearly visible. As the camera pans, the focus gradually and seamlessly shifts from the joyful gathering in the background to the central subject in the foreground: a woman's hand, delicately lifting a glass of white wine from the table. Steady and composed, her nails immaculately manicured, the hand elegantly grasps the slender stem of a classic crystal goblet. The glass is about one-third full of pale golden white wine, and under the soft light cast by a low, warm-toned chandelier overhead, the wine's clarity is translucent, shimmering with sparkling shimmer. As the wine glasses are slowly raised to the center of the picture, the smiling faces and figures of the friends in the background are completely blurred into warm spots of light (Bokeh) and soft blocks of color. The whole picture is wrapped in an intimate, peaceful and joyful atmosphere. The main colors are warm amber and cream, and the contrast of light and shadow is soft, perfectly capturing this peaceful and sincere moment of celebration.", + "seed": 29049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_9/index_3890_seed_29049_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021255077281966805, + "vlm_logprob": -0.021255077281966805, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.021255077281966805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015264946967363358 + ], + "opposite_logprob_tokens": [ + -4.890264987945557 + ], + "logprob_reward": -0.015264946967363358, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.015264946967363358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1759108603000641, + 0.055609773844480515, + 0.0505792498588562, + 0.09254157543182373, + 0.049837447702884674, + 0.04367814213037491, + 0.09209280461072922, + 0.22101600468158722, + 0.18546025454998016, + 0.03327387198805809 + ], + "credit_map": [ + 0.03333448991179466, + 0.023203125223517418, + 0.014062897302210331, + 0.009706196375191212, + 0.06021780148148537, + 0.012656591832637787, + 0.012625310570001602, + 0.010104442946612835, + 0.009491683915257454, + 0.005103972274810076, + 0.016163045540452003, + 0.004415359813719988, + 0.0031031472608447075, + 0.005635922309011221, + 0.004117803182452917, + 0.007578839082270861, + 0.006721194367855787, + 0.007895231246948242, + 0.0053329430520534515, + 0.00478749955072999, + 0.004028498660773039, + 0.0061869132332503796, + 0.004815616644918919, + 0.010811352171003819, + 0.028709782287478447, + 0.005523200612515211, + 0.029149802401661873, + 0.004950493108481169, + 0.00641873013228178, + 0.004397196229547262, + 0.006377921439707279, + 0.007014450617134571, + 0.006393229588866234, + 0.005957823246717453, + 0.0059907399117946625, + 0.011248872615396976, + 0.003576155286282301, + 0.004757111892104149, + 0.005812672432512045, + 0.00610083993524313, + 0.005473319441080093, + 0.007230924442410469, + 0.005976004991680384, + 0.002575729275122285, + 0.00247585098259151, + 0.013546234928071499, + 0.0032832783181220293, + 0.0031167962588369846, + 0.035950180143117905, + 0.004539141431450844, + 0.005671794526278973, + 0.0020061174873262644, + 0.0033845845609903336, + 0.005517090205103159, + 0.004623030312359333, + 0.03040086291730404, + 0.0680190697312355, + 0.008132320828735828, + 0.011212551034986973, + 0.004082437604665756, + 0.004779806360602379, + 0.027080314233899117, + 0.00773421349003911, + 0.0899752825498581, + 0.03621610254049301, + 0.012308632023632526, + 0.017798395827412605, + 0.011416361667215824, + 0.03467007353901863, + 0.038278158754110336, + 0.01564258337020874, + 0.019129939377307892, + 0.006965288892388344, + 0.0021968018263578415, + 0.0029979399405419827, + 0.002667198423296213, + 0.00490596704185009, + 0.00775485672056675, + 0.002402324927970767, + 0.0033834956120699644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.015264946967363358 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002500861417502165 + ], + "opposite_logprob_tokens": [ + -6.002501010894775 + ], + "logprob_reward": -0.002500861417502165, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002500861417502165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21143938601016998, + 0.09421349316835403, + 0.06561720371246338, + 0.08609133213758469, + 0.05865120142698288, + 0.10016581416130066, + 0.13621024787425995, + 0.059613391757011414, + 0.1537138819694519, + 0.03428398072719574 + ], + "credit_map": [ + 0.040759630501270294, + 0.029956169426441193, + 0.016605127602815628, + 0.013747290708124638, + 0.06914900988340378, + 0.015613975003361702, + 0.014694476500153542, + 0.010913732461631298, + 0.016189750283956528, + 0.011180673725903034, + 0.009929260239005089, + 0.00988808460533619, + 0.009201791137456894, + 0.02158140204846859, + 0.008977057412266731, + 0.007265486754477024, + 0.011679756455123425, + 0.012089964002370834, + 0.006859530229121447, + 0.004570577759295702, + 0.007397782057523727, + 0.009059231728315353, + 0.0065841879695653915, + 0.0073761846870183945, + 0.02505137398838997, + 0.0077072493731975555, + 0.020758524537086487, + 0.005565921310335398, + 0.007920184172689915, + 0.005158830434083939, + 0.007294694427400827, + 0.006634569261223078, + 0.010532785207033157, + 0.00833724346011877, + 0.004792983643710613, + 0.00715623889118433, + 0.0068488470278680325, + 0.006129151675850153, + 0.006408065557479858, + 0.008445889689028263, + 0.007861761376261711, + 0.020564163103699684, + 0.008605288341641426, + 0.005275229923427105, + 0.007187467534095049, + 0.0353989414870739, + 0.008463112637400627, + 0.006809860467910767, + 0.0524672195315361, + 0.009107143618166447, + 0.009189710021018982, + 0.004990801680833101, + 0.004993824288249016, + 0.007832170464098454, + 0.008554604835808277, + 0.039074793457984924, + 0.01982293464243412, + 0.003470050171017647, + 0.0053789145313203335, + 0.002084946958348155, + 0.003799899248406291, + 0.0077290283516049385, + 0.002796962857246399, + 0.014530659653246403, + 0.03791077062487602, + 0.00736213568598032, + 0.01141643337905407, + 0.00770002044737339, + 0.022172288969159126, + 0.043132372200489044, + 0.009808889590203762, + 0.014210988767445087, + 0.008168310858309269, + 0.0021713022142648697, + 0.004137059673666954, + 0.0026974540669471025, + 0.003396021667867899, + 0.007119419053196907, + 0.002574165817350149, + 0.004020252265036106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.002500861417502165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062033720314502716 + ], + "opposite_logprob_tokens": [ + -2.8120336532592773 + ], + "logprob_reward": -0.062033720314502716, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062033720314502716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06593824177980423, + 0.07133232057094574, + 0.03797302395105362, + 0.12231782078742981, + 0.07165907323360443, + 0.049810945987701416, + 0.09943363815546036, + 0.10408959537744522, + 0.32380345463752747, + 0.053641900420188904 + ], + "credit_map": [ + 0.008260791189968586, + 0.00867545511573553, + 0.011938259936869144, + 0.0062935492023825645, + 0.012065526098012924, + 0.006640991661697626, + 0.005280031356960535, + 0.006783633027225733, + 0.01123827788978815, + 0.00842515379190445, + 0.01922032982110977, + 0.006453734822571278, + 0.006193989422172308, + 0.008338387124240398, + 0.006329171825200319, + 0.005133278667926788, + 0.007319004274904728, + 0.006718096323311329, + 0.002940499922260642, + 0.00244510336779058, + 0.003140887012705207, + 0.004612388554960489, + 0.0026460110675543547, + 0.008151032961905003, + 0.03656644746661186, + 0.011334986425936222, + 0.0318082757294178, + 0.007743009366095066, + 0.00965885166078806, + 0.008824426680803299, + 0.008452557027339935, + 0.00792927760630846, + 0.017276061698794365, + 0.019622229039669037, + 0.005547917447984219, + 0.0055619231425225735, + 0.004759351257234812, + 0.0040548620745539665, + 0.007132029160857201, + 0.007704696152359247, + 0.006948841270059347, + 0.008443158119916916, + 0.006238702218979597, + 0.0034591485746204853, + 0.003947563003748655, + 0.01068753469735384, + 0.004410926252603531, + 0.005675074644386768, + 0.03138503059744835, + 0.005405062343925238, + 0.005830670706927776, + 0.0048161218874156475, + 0.005213914439082146, + 0.006709770765155554, + 0.005060914438217878, + 0.035012152045965195, + 0.030904289335012436, + 0.004705984611064196, + 0.005197076592594385, + 0.0018725654808804393, + 0.0031549176201224327, + 0.011645415797829628, + 0.003958454821258783, + 0.04265088960528374, + 0.07358364015817642, + 0.016762079671025276, + 0.027783218771219254, + 0.017738087102770805, + 0.05013742297887802, + 0.08275628089904785, + 0.021450335159897804, + 0.03359237685799599, + 0.012062194757163525, + 0.0024329067673534155, + 0.003440429922193289, + 0.003008293453603983, + 0.006818077061325312, + 0.014223894104361534, + 0.004062344320118427, + 0.007593761198222637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.062033720314502716 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283764719963074 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283764719963074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283764719963074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07229553163051605, + 0.06046849861741066, + 0.05997925251722336, + 0.05455101653933525, + 0.0554032064974308, + 0.24344517290592194, + 0.2734174430370331, + 0.05798713490366936, + 0.08568663895130157, + 0.036766160279512405 + ], + "credit_map": [ + 0.01294909231364727, + 0.00923353061079979, + 0.005621074698865414, + 0.007009162101894617, + 0.01945270225405693, + 0.00670153833925724, + 0.005635772366076708, + 0.005692654754966497, + 0.010520359501242638, + 0.00758120184764266, + 0.005382075440138578, + 0.006259429268538952, + 0.004524353425949812, + 0.016036247834563255, + 0.005377030931413174, + 0.004787798039615154, + 0.01059751957654953, + 0.00811858382076025, + 0.006923398934304714, + 0.004796857945621014, + 0.004557755310088396, + 0.00450358958914876, + 0.005870846565812826, + 0.01461070217192173, + 0.016664277762174606, + 0.0040394896641373634, + 0.016987795010209084, + 0.003263874677941203, + 0.003111619735136628, + 0.002477994654327631, + 0.003994459751993418, + 0.004011507146060467, + 0.012918893247842789, + 0.014152338728308678, + 0.00549268489703536, + 0.004545996896922588, + 0.0027745149563997984, + 0.0045045181177556515, + 0.0037114289589226246, + 0.007302828133106232, + 0.02051023580133915, + 0.05372699350118637, + 0.019196514040231705, + 0.00790192000567913, + 0.016114989295601845, + 0.09273950010538101, + 0.01738998107612133, + 0.015865037217736244, + 0.09206964820623398, + 0.01142381876707077, + 0.014379864558577538, + 0.013402577489614487, + 0.013447032310068607, + 0.01649639941751957, + 0.017441602423787117, + 0.09475649893283844, + 0.019455429166555405, + 0.002581567270681262, + 0.004185584839433432, + 0.0016146920388564467, + 0.002597968792542815, + 0.004458134528249502, + 0.003233061172068119, + 0.01986069791018963, + 0.012624315917491913, + 0.006479652598500252, + 0.008391578681766987, + 0.00689679617062211, + 0.01911064051091671, + 0.011103448458015919, + 0.012848122045397758, + 0.008232085034251213, + 0.006067546084523201, + 0.0026439267676323652, + 0.004678007680922747, + 0.004666932392865419, + 0.005508963484317064, + 0.0038458858616650105, + 0.004729363601654768, + 0.0046255337074398994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.023283764719963074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003192092990502715 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003192092990502715, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192092990502715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09470315277576447, + 0.04953047260642052, + 0.2924802601337433, + 0.07615784555673599, + 0.05408395454287529, + 0.03551355376839638, + 0.1837211698293686, + 0.08856327831745148, + 0.06282203644514084, + 0.06242423877120018 + ], + "credit_map": [ + 0.012404427863657475, + 0.020744910463690758, + 0.009905221872031689, + 0.010091242380440235, + 0.016455117613077164, + 0.0064922538585960865, + 0.009235557168722153, + 0.009374426677823067, + 0.008827652782201767, + 0.008016467094421387, + 0.006185173988342285, + 0.004162679426372051, + 0.0043525975197553635, + 0.00849211122840643, + 0.004494397900998592, + 0.004999388940632343, + 0.05340411886572838, + 0.03439861908555031, + 0.03953137248754501, + 0.022819748148322105, + 0.01745443418622017, + 0.031150855123996735, + 0.028012273833155632, + 0.06570884585380554, + 0.019693756476044655, + 0.006318230181932449, + 0.020068272948265076, + 0.005369239021092653, + 0.006042362190783024, + 0.007043068762868643, + 0.005471863318234682, + 0.006151058245450258, + 0.01225308421999216, + 0.012441430240869522, + 0.00487868906930089, + 0.004733182489871979, + 0.0034152751322835684, + 0.004413384944200516, + 0.0046017225831747055, + 0.007347190286964178, + 0.005847318563610315, + 0.002046625828370452, + 0.006160373333841562, + 0.003444086527451873, + 0.0029099390376359224, + 0.005447817035019398, + 0.004878959618508816, + 0.004778432194143534, + 0.06668858230113983, + 0.007094910368323326, + 0.009543433785438538, + 0.005404288414865732, + 0.007187332957983017, + 0.010869819670915604, + 0.010588867589831352, + 0.06634394079446793, + 0.023722926154732704, + 0.003781608073040843, + 0.006103749852627516, + 0.002167216269299388, + 0.003846559440717101, + 0.011324211955070496, + 0.004309313837438822, + 0.033307697623968124, + 0.01204993762075901, + 0.0045233373530209064, + 0.009089146740734577, + 0.006401647347956896, + 0.007223604246973991, + 0.012420258484780788, + 0.004667078610509634, + 0.006447024177759886, + 0.013882945291697979, + 0.0035238286945968866, + 0.005316064227372408, + 0.005957744549959898, + 0.007677245419472456, + 0.01565365120768547, + 0.005328257102519274, + 0.00508450111374259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.003192092990502715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a cozy, humanistic interior, a fluid single shot begins at table level, then slowly and gracefully pans rightward across a warm dining scene. A long, dark wooden table, covered with a beige linen tablecloth, is laden with half-eaten home-cooked dishes and scattered crumbs. Several friends, dressed in cozy sweaters and casual shirts, gather around the table. Their faces are blurred by a soft bokeh, yet their genuine smiles and lively gestures, engaged in conversation, are clearly visible. As the camera pans, the focus gradually and seamlessly shifts from the joyful gathering in the background to the central subject in the foreground: a woman's hand, delicately lifting a glass of white wine from the table. Steady and composed, her nails immaculately manicured, the hand elegantly grasps the slender stem of a classic crystal goblet. The glass is about one-third full of pale golden white wine, and under the soft light cast by a low, warm-toned chandelier overhead, the wine's clarity is translucent, shimmering with sparkling shimmer. As the wine glasses are slowly raised to the center of the picture, the smiling faces and figures of the friends in the background are completely blurred into warm spots of light (Bokeh) and soft blocks of color. The whole picture is wrapped in an intimate, peaceful and joyful atmosphere. The main colors are warm amber and cream, and the contrast of light and shadow is soft, perfectly capturing this peaceful and sincere moment of celebration.", + "seed": 29050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_9/index_3890_seed_29050_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025647666351869703, + "vlm_logprob": -0.025647666351869703, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.025647666351869703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03035958670079708 + ], + "opposite_logprob_tokens": [ + -3.9053595066070557 + ], + "logprob_reward": -0.03035958670079708, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03035958670079708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19532379508018494, + 0.10336258262395859, + 0.027096712961792946, + 0.06286746263504028, + 0.19022324681282043, + 0.05564015731215477, + 0.031837210059165955, + 0.04929755628108978, + 0.19385480880737305, + 0.0904965028166771 + ], + "credit_map": [ + 0.038530319929122925, + 0.014475107192993164, + 0.018878567963838577, + 0.019491566345095634, + 0.018380092456936836, + 0.014173907227814198, + 0.028344206511974335, + 0.043050020933151245, + 0.010502602905035019, + 0.019272275269031525, + 0.015692204236984253, + 0.009886994026601315, + 0.012536599300801754, + 0.0120366420596838, + 0.011729602701961994, + 0.011705647222697735, + 0.004095497541129589, + 0.0028884417843073606, + 0.0024867006577551365, + 0.004465981852263212, + 0.0022907592356204987, + 0.0036782242823392153, + 0.003134904894977808, + 0.004056200850754976, + 0.004803546704351902, + 0.006602735724300146, + 0.007098474074155092, + 0.009589419700205326, + 0.007921514101326466, + 0.006934501696377993, + 0.005789186339825392, + 0.014128080569207668, + 0.018158674240112305, + 0.03919642046093941, + 0.01814623549580574, + 0.02082805335521698, + 0.0299872737377882, + 0.01657501608133316, + 0.02528763934969902, + 0.022043928503990173, + 0.003174737561494112, + 0.017199406400322914, + 0.0030291415750980377, + 0.006377970799803734, + 0.0060633146204054356, + 0.012084112502634525, + 0.0044305226765573025, + 0.003280949080362916, + 0.0033422019332647324, + 0.0061660087667405605, + 0.0032385464292019606, + 0.00420876732096076, + 0.004331499338150024, + 0.002917329780757427, + 0.004639077931642532, + 0.002993775764480233, + 0.004991388414055109, + 0.007009935099631548, + 0.004458962939679623, + 0.007312257774174213, + 0.0043739150278270245, + 0.006103513762354851, + 0.009375030174851418, + 0.005672546569257975, + 0.007727827411144972, + 0.06686427444219589, + 0.010358210653066635, + 0.011180207133293152, + 0.0746433213353157, + 0.007935370318591595, + 0.00977027602493763, + 0.005375311244279146, + 0.00320060015656054, + 0.03447601571679115, + 0.0029243624303489923, + 0.006566040217876434, + 0.03403211385011673, + 0.004622738342732191, + 0.002824041061103344, + 0.0018505892949178815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.03035958670079708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108085297048092 + ], + "opposite_logprob_tokens": [ + -5.50410795211792 + ], + "logprob_reward": -0.004108085297048092, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108085297048092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14045360684394836, + 0.0445987693965435, + 0.04317239299416542, + 0.08955156803131104, + 0.13546979427337646, + 0.09043118357658386, + 0.05379365012049675, + 0.0863606184720993, + 0.2244330793619156, + 0.09173532575368881 + ], + "credit_map": [ + 0.026092002168297768, + 0.01241215132176876, + 0.01080638449639082, + 0.010322469286620617, + 0.0261685848236084, + 0.015455036424100399, + 0.01342743169516325, + 0.0257695484906435, + 0.005177336744964123, + 0.0064787790179252625, + 0.0052312216721475124, + 0.004965224768966436, + 0.00445560971274972, + 0.005451761186122894, + 0.004553916864097118, + 0.00828491896390915, + 0.0038279048167169094, + 0.0048668235540390015, + 0.005465445574373007, + 0.006144002079963684, + 0.005108317825943232, + 0.007203933782875538, + 0.004512614104896784, + 0.006043352652341127, + 0.012926638126373291, + 0.01133146695792675, + 0.008614775724709034, + 0.008433271199464798, + 0.01598179154098034, + 0.008168739266693592, + 0.0095823397859931, + 0.014512543566524982, + 0.014645352959632874, + 0.023294489830732346, + 0.013067878782749176, + 0.01639600098133087, + 0.017966771498322487, + 0.01344433706253767, + 0.017746487632393837, + 0.01890847086906433, + 0.0054466393776237965, + 0.027339918538928032, + 0.005333962384611368, + 0.009848506189882755, + 0.010308239609003067, + 0.01967967115342617, + 0.007543292362242937, + 0.004930954892188311, + 0.005277691874653101, + 0.013274415396153927, + 0.004556390456855297, + 0.0057134306989610195, + 0.0073948572389781475, + 0.005637764930725098, + 0.006641671061515808, + 0.00529742706567049, + 0.0057487646117806435, + 0.014401642605662346, + 0.007970611564815044, + 0.011128520593047142, + 0.007823181338608265, + 0.009450757876038551, + 0.020159920677542686, + 0.00967722199857235, + 0.008283162489533424, + 0.0828118547797203, + 0.010685007087886333, + 0.011560085229575634, + 0.09049458801746368, + 0.009104954078793526, + 0.007128875702619553, + 0.0043645622208714485, + 0.003897476475685835, + 0.033362437039613724, + 0.003541405312716961, + 0.006585293915122747, + 0.0329299159348011, + 0.005590641405433416, + 0.0031859460286796093, + 0.0026422098744660616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.004108085297048092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000081241130829 + ], + "opposite_logprob_tokens": [ + -2.6950008869171143 + ], + "logprob_reward": -0.07000081241130829, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000081241130829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39720794558525085, + 0.062264591455459595, + 0.19919143617153168, + 0.02386380173265934, + 0.1053176000714302, + 0.05109996721148491, + 0.036594439297914505, + 0.03697095811367035, + 0.07047317177057266, + 0.017016058787703514 + ], + "credit_map": [ + 0.096523217856884, + 0.03414377197623253, + 0.03126079961657524, + 0.029553651809692383, + 0.05173691362142563, + 0.03341562673449516, + 0.04205307736992836, + 0.07852085679769516, + 0.008384727872908115, + 0.006455590482801199, + 0.0070946551859378815, + 0.005268811248242855, + 0.007148395292460918, + 0.009289190173149109, + 0.009061027318239212, + 0.00956218782812357, + 0.018182829022407532, + 0.01438414677977562, + 0.017948944121599197, + 0.02961006574332714, + 0.020906999707221985, + 0.03556489199399948, + 0.017753630876541138, + 0.04483991861343384, + 0.0027810768224298954, + 0.002605822402983904, + 0.0018841495038941503, + 0.00268882280215621, + 0.001984033966436982, + 0.0028933389112353325, + 0.0031948320101946592, + 0.0058317226357758045, + 0.01004934124648571, + 0.0192574430257082, + 0.01219579204916954, + 0.010707391425967216, + 0.014133227989077568, + 0.01105367112904787, + 0.013207978568971157, + 0.014712749049067497, + 0.003450660267844796, + 0.018176648765802383, + 0.002423623576760292, + 0.0034394769463688135, + 0.006079999264329672, + 0.011374127119779587, + 0.0040007708594202995, + 0.00215465622022748, + 0.004410956986248493, + 0.0073935664258897305, + 0.002907324582338333, + 0.004230998922139406, + 0.005030437838286161, + 0.003351154737174511, + 0.00422277394682169, + 0.005047221202403307, + 0.0036806922871619463, + 0.006816097069531679, + 0.0026842730585485697, + 0.004499696660786867, + 0.004318395629525185, + 0.0048535591922700405, + 0.004783012438565493, + 0.005335227120667696, + 0.0035786281805485487, + 0.02138473279774189, + 0.0036896448582410812, + 0.0051084840670228004, + 0.02551063522696495, + 0.004243992734700441, + 0.0035656008403748274, + 0.003391451667994261, + 0.0015845318557694554, + 0.0031405515037477016, + 0.0011352448491379619, + 0.0022846697829663754, + 0.0025976658798754215, + 0.002272373763844371, + 0.0025101997889578342, + 0.0014908213634043932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.07000081241130829 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057763561606407 + ], + "opposite_logprob_tokens": [ + -3.8955776691436768 + ], + "logprob_reward": -0.02057763561606407, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057763561606407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18551242351531982, + 0.019118182361125946, + 0.028618045151233673, + 0.08049993216991425, + 0.029678838327527046, + 0.03286221995949745, + 0.020791420713067055, + 0.06689020991325378, + 0.38209015130996704, + 0.15393862128257751 + ], + "credit_map": [ + 0.040951937437057495, + 0.01475746463984251, + 0.015042603015899658, + 0.014427085407078266, + 0.02304704487323761, + 0.015808526426553726, + 0.019271252676844597, + 0.04220651462674141, + 0.0018970412202179432, + 0.0026514444034546614, + 0.002243337919935584, + 0.0020472765900194645, + 0.0019773656968027353, + 0.00366804376244545, + 0.0023421712685376406, + 0.002291498938575387, + 0.003331365529447794, + 0.004097724333405495, + 0.003597398055717349, + 0.00416205869987607, + 0.0035426525864750147, + 0.0037551033310592175, + 0.0026149768382310867, + 0.0035167650785297155, + 0.007691585458815098, + 0.009496715851128101, + 0.007739164400845766, + 0.008806826546788216, + 0.018280979245901108, + 0.010639733634889126, + 0.009061678312718868, + 0.00878324918448925, + 0.0028459173627197742, + 0.0041583385318517685, + 0.003348505822941661, + 0.0035639747511595488, + 0.00458779139444232, + 0.003222384722903371, + 0.004445468075573444, + 0.0035064576659351587, + 0.003646208206191659, + 0.0060214754194021225, + 0.004196330439299345, + 0.0038383137434720993, + 0.003527430584654212, + 0.004393859766423702, + 0.004122392740100622, + 0.003116207430139184, + 0.0023866158444434404, + 0.0036844604182988405, + 0.001997930696234107, + 0.0024386311415582895, + 0.0030130925588309765, + 0.0027145484928041697, + 0.002290082396939397, + 0.002266059396788478, + 0.0040023233741521835, + 0.006812272593379021, + 0.005700838286429644, + 0.00826197024434805, + 0.006825900636613369, + 0.009592507965862751, + 0.016844402998685837, + 0.00884999055415392, + 0.014030512422323227, + 0.13977454602718353, + 0.018201887607574463, + 0.01953386701643467, + 0.15467515587806702, + 0.015472391620278358, + 0.012859161011874676, + 0.007542649284005165, + 0.005316894967108965, + 0.056890539824962616, + 0.005042712669819593, + 0.010061334818601608, + 0.058885443955659866, + 0.008993458934128284, + 0.005229523405432701, + 0.0035187143366783857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.02057763561606407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031922117341309786 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031922117341309786, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031922117341309786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20768867433071136, + 0.07270899415016174, + 0.029673581942915916, + 0.11443422734737396, + 0.03687373176217079, + 0.04055259749293327, + 0.04350754991173744, + 0.05580900236964226, + 0.3325548470020294, + 0.06619679927825928 + ], + "credit_map": [ + 0.050257280468940735, + 0.019045494496822357, + 0.014490429311990738, + 0.012934865429997444, + 0.03590643033385277, + 0.018819091841578484, + 0.020497722551226616, + 0.03573734685778618, + 0.014137445017695427, + 0.008547087199985981, + 0.0073013147339224815, + 0.006172338500618935, + 0.010964850895106792, + 0.00854754913598299, + 0.007752878125756979, + 0.00928553007543087, + 0.003702180925756693, + 0.0036823530681431293, + 0.0037717483937740326, + 0.004458978772163391, + 0.003598026465624571, + 0.0032340020406991243, + 0.00311502767726779, + 0.004111264832317829, + 0.013504112139344215, + 0.010472474619746208, + 0.011480696499347687, + 0.013589716516435146, + 0.018429722636938095, + 0.011528363451361656, + 0.012592584826052189, + 0.022836556658148766, + 0.0027104015462100506, + 0.013680699281394482, + 0.001961984671652317, + 0.0024261660873889923, + 0.007007668260484934, + 0.0037669807206839323, + 0.002668156521394849, + 0.0026516763027757406, + 0.002610717201605439, + 0.014364507980644703, + 0.0018711758311837912, + 0.003292773151770234, + 0.005570672452449799, + 0.00760588189586997, + 0.0031915882136672735, + 0.002045280998572707, + 0.004354931879788637, + 0.01159355603158474, + 0.002802088391035795, + 0.0045162588357925415, + 0.006277638953179121, + 0.004819229245185852, + 0.005327895283699036, + 0.003815951757133007, + 0.0039201839827001095, + 0.007205597590655088, + 0.0051121641881763935, + 0.008415031246840954, + 0.004846467170864344, + 0.006970238871872425, + 0.013117080554366112, + 0.00622223736718297, + 0.012164285406470299, + 0.12239927798509598, + 0.015839800238609314, + 0.01698233000934124, + 0.1344665139913559, + 0.013440746814012527, + 0.010862383991479874, + 0.006399497855454683, + 0.0022412652615457773, + 0.023737255483865738, + 0.002079572295770049, + 0.0046552084386348724, + 0.022894522175192833, + 0.004809295292943716, + 0.003183715743944049, + 0.002595961792394519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.0031922117341309786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a cozy, humanistic interior, a fluid single shot begins at table level, then slowly and gracefully pans rightward across a warm dining scene. A long, dark wooden table, covered with a beige linen tablecloth, is laden with half-eaten home-cooked dishes and scattered crumbs. Several friends, dressed in cozy sweaters and casual shirts, gather around the table. Their faces are blurred by a soft bokeh, yet their genuine smiles and lively gestures, engaged in conversation, are clearly visible. As the camera pans, the focus gradually and seamlessly shifts from the joyful gathering in the background to the central subject in the foreground: a woman's hand, delicately lifting a glass of white wine from the table. Steady and composed, her nails immaculately manicured, the hand elegantly grasps the slender stem of a classic crystal goblet. The glass is about one-third full of pale golden white wine, and under the soft light cast by a low, warm-toned chandelier overhead, the wine's clarity is translucent, shimmering with sparkling shimmer. As the wine glasses are slowly raised to the center of the picture, the smiling faces and figures of the friends in the background are completely blurred into warm spots of light (Bokeh) and soft blocks of color. The whole picture is wrapped in an intimate, peaceful and joyful atmosphere. The main colors are warm amber and cream, and the contrast of light and shadow is soft, perfectly capturing this peaceful and sincere moment of celebration.", + "seed": 29051, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000082/group_9/index_3890_seed_29051_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016848475206643343, + "vlm_logprob": -0.016848475206643343, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.016848475206643343, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017409835010766983 + ], + "opposite_logprob_tokens": [ + -4.892409801483154 + ], + "logprob_reward": -0.017409835010766983, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017409835010766983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11311447620391846, + 0.064588762819767, + 0.06516168266534805, + 0.26792746782302856, + 0.05715011805295944, + 0.030096502974629402, + 0.11954519897699356, + 0.1731119304895401, + 0.07640575617551804, + 0.03289804607629776 + ], + "credit_map": [ + 0.016805145889520645, + 0.015108572319149971, + 0.017759237438440323, + 0.01206202432513237, + 0.010952604934573174, + 0.013448876328766346, + 0.011475288309156895, + 0.015502729453146458, + 0.009190346114337444, + 0.009128491394221783, + 0.009375467896461487, + 0.007875730283558369, + 0.0070927804335951805, + 0.007406943012028933, + 0.005697414744645357, + 0.008821582421660423, + 0.005975398235023022, + 0.008597166277468204, + 0.0048850784078240395, + 0.007087737321853638, + 0.007453584112226963, + 0.008342810906469822, + 0.009004049003124237, + 0.013815858401358128, + 0.021996183320879936, + 0.031191721558570862, + 0.025414271280169487, + 0.023639455437660217, + 0.03696012124419212, + 0.07182592898607254, + 0.013624534010887146, + 0.04327522963285446, + 0.00452381419017911, + 0.006850924808532, + 0.00828525796532631, + 0.0064161098562181, + 0.00603407621383667, + 0.007266134489327669, + 0.009394695982336998, + 0.008379106409847736, + 0.0027191918343305588, + 0.004173625260591507, + 0.003768694121390581, + 0.002742271637544036, + 0.004504470620304346, + 0.004092723596841097, + 0.004859852138906717, + 0.0032356728333979845, + 0.013245838694274426, + 0.01427997276186943, + 0.012676496990025043, + 0.01758803427219391, + 0.011660195887088776, + 0.020149391144514084, + 0.016877150163054466, + 0.013068125583231449, + 0.010905642062425613, + 0.020597539842128754, + 0.013016937300562859, + 0.03993989899754524, + 0.02029375731945038, + 0.022051561623811722, + 0.03218662738800049, + 0.014119964092969894, + 0.006421880330890417, + 0.010596358217298985, + 0.01043368224054575, + 0.00750507228076458, + 0.00650797551497817, + 0.01943936198949814, + 0.007266013417392969, + 0.008235406130552292, + 0.005043565761297941, + 0.0037875392008572817, + 0.00403194036334753, + 0.003954450599849224, + 0.0028791790828108788, + 0.004642972722649574, + 0.004278848879039288, + 0.004279551561921835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.017409835010766983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036295270547270775 + ], + "opposite_logprob_tokens": [ + -5.628629684448242 + ], + "logprob_reward": -0.0036295270547270775, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036295270547270775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10831224173307419, + 0.059802547097206116, + 0.2168537676334381, + 0.17558430135250092, + 0.063482366502285, + 0.06258776783943176, + 0.07218117266893387, + 0.06402135640382767, + 0.09602683782577515, + 0.08114766329526901 + ], + "credit_map": [ + 0.017782039940357208, + 0.012065248563885689, + 0.011941798962652683, + 0.011530578136444092, + 0.010193631052970886, + 0.012194630689918995, + 0.009902114048600197, + 0.02270219661295414, + 0.0075818076729774475, + 0.0081928800791502, + 0.0075665684416890144, + 0.010035241022706032, + 0.006338585168123245, + 0.006951526738703251, + 0.0058303107507526875, + 0.007305629085749388, + 0.02477584220468998, + 0.04041663929820061, + 0.02344159223139286, + 0.02008642815053463, + 0.022558990865945816, + 0.0317348912358284, + 0.011250386945903301, + 0.04258899390697479, + 0.01392381265759468, + 0.020151080563664436, + 0.014727435074746609, + 0.011683662422001362, + 0.023055480793118477, + 0.05305531993508339, + 0.009689195081591606, + 0.029298318549990654, + 0.007814357988536358, + 0.006574976723641157, + 0.00679940078407526, + 0.007254819851368666, + 0.010150052607059479, + 0.009793638251721859, + 0.007848691195249557, + 0.007246426306664944, + 0.005250738468021154, + 0.010866993106901646, + 0.006308846641331911, + 0.005362810101360083, + 0.007781299762427807, + 0.00986005924642086, + 0.009768842719495296, + 0.007388178259134293, + 0.008165641687810421, + 0.008605540730059147, + 0.011482789181172848, + 0.006372341886162758, + 0.01108289323747158, + 0.01276299450546503, + 0.008141291327774525, + 0.0055676838383078575, + 0.007781259249895811, + 0.009318196214735508, + 0.00755891902372241, + 0.008217789232730865, + 0.00782580953091383, + 0.007954581640660763, + 0.006975443102419376, + 0.008389363065361977, + 0.00936993770301342, + 0.023884564638137817, + 0.009507028385996819, + 0.009957965463399887, + 0.010531648062169552, + 0.017255539074540138, + 0.008023036643862724, + 0.007497116923332214, + 0.008377891965210438, + 0.008199130184948444, + 0.010742297396063805, + 0.010115043260157108, + 0.007713780738413334, + 0.009947513230144978, + 0.012745670042932034, + 0.013306335546076298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0036295270547270775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792189598083496 + ], + "opposite_logprob_tokens": [ + -3.529792308807373 + ], + "logprob_reward": -0.029792189598083496, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792189598083496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0531272366642952, + 0.05874985456466675, + 0.059823863208293915, + 0.12337258458137512, + 0.10878632962703705, + 0.186255544424057, + 0.1323339343070984, + 0.16118094325065613, + 0.07723701000213623, + 0.039132680743932724 + ], + "credit_map": [ + 0.007391959894448519, + 0.010679333470761776, + 0.009098716080188751, + 0.004056149162352085, + 0.006455655675381422, + 0.007032334804534912, + 0.004260741174221039, + 0.004152352921664715, + 0.012249991297721863, + 0.008540602400898933, + 0.007285391446202993, + 0.007131339982151985, + 0.006371492054313421, + 0.005758308805525303, + 0.004997748415917158, + 0.006414985284209251, + 0.008908078074455261, + 0.008982598781585693, + 0.008138121105730534, + 0.006615119986236095, + 0.00577847333624959, + 0.006577090825885534, + 0.0066600157879292965, + 0.008164368569850922, + 0.011163674294948578, + 0.01203103456646204, + 0.014575565233826637, + 0.009431849233806133, + 0.019397342577576637, + 0.03355461731553078, + 0.006732412148267031, + 0.01648610271513462, + 0.01184894423931837, + 0.009823659434914589, + 0.014163549058139324, + 0.011708812788128853, + 0.01614093780517578, + 0.01700306497514248, + 0.01565869152545929, + 0.012438672594726086, + 0.017048843204975128, + 0.029847435653209686, + 0.019876493141055107, + 0.01549839973449707, + 0.025976598262786865, + 0.026798706501722336, + 0.03227253258228302, + 0.018936537206172943, + 0.015675310045480728, + 0.014775469899177551, + 0.01844317466020584, + 0.012772899121046066, + 0.017822321504354477, + 0.017660025507211685, + 0.019971661269664764, + 0.015213081613183022, + 0.013992929831147194, + 0.015170518308877945, + 0.02734900824725628, + 0.030521534383296967, + 0.012292408384382725, + 0.029436448588967323, + 0.02014361508190632, + 0.012274487875401974, + 0.00822401512414217, + 0.012539971619844437, + 0.013349920511245728, + 0.007658844348043203, + 0.006775388028472662, + 0.013061773031949997, + 0.008069776929914951, + 0.007557323668152094, + 0.006554026622325182, + 0.003925425000488758, + 0.005401306319981813, + 0.005879178177565336, + 0.003542113583534956, + 0.004084345884621143, + 0.005625664256513119, + 0.004120624624192715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.029792189598083496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795313253998756 + ], + "opposite_logprob_tokens": [ + -3.5297954082489014 + ], + "logprob_reward": -0.029795313253998756, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795313253998756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10786443948745728, + 0.08634717017412186, + 0.0855199545621872, + 0.09861550480127335, + 0.052369143813848495, + 0.16147646307945251, + 0.09392370283603668, + 0.09510453045368195, + 0.10685852915048599, + 0.1119205504655838 + ], + "credit_map": [ + 0.016709383577108383, + 0.01094707753509283, + 0.011210755445063114, + 0.008987462148070335, + 0.010694067925214767, + 0.012146897614002228, + 0.01337759755551815, + 0.023791199550032616, + 0.010470695793628693, + 0.017254158854484558, + 0.011246451176702976, + 0.010458854958415031, + 0.0101138511672616, + 0.00879888515919447, + 0.007627125363796949, + 0.010377142578363419, + 0.008870000950992107, + 0.010362989269196987, + 0.011911791749298573, + 0.012945973314344883, + 0.008989139460027218, + 0.009281880222260952, + 0.011858801357448101, + 0.011299378238618374, + 0.006780365016311407, + 0.012945957481861115, + 0.006307719741016626, + 0.006191516760736704, + 0.009717007167637348, + 0.032761264592409134, + 0.00627082958817482, + 0.017640845850110054, + 0.004853959660977125, + 0.005894599948078394, + 0.007203612942248583, + 0.008023141883313656, + 0.00655526714399457, + 0.006747008766978979, + 0.005511250346899033, + 0.0075803049840033054, + 0.01716916449368, + 0.02072472870349884, + 0.018198614940047264, + 0.012933812104165554, + 0.023949654772877693, + 0.01517723873257637, + 0.04107644781470299, + 0.012246797792613506, + 0.009221295826137066, + 0.010840843431651592, + 0.008679471909999847, + 0.012931163422763348, + 0.012082635425031185, + 0.016699353232979774, + 0.00965763907879591, + 0.013811299577355385, + 0.006763567682355642, + 0.010425502434372902, + 0.015334662981331348, + 0.013050601817667484, + 0.012589719146490097, + 0.018349094316363335, + 0.010943218134343624, + 0.007648165337741375, + 0.010740615427494049, + 0.013626279309391975, + 0.011219682171940804, + 0.008219900541007519, + 0.01096763089299202, + 0.03083481267094612, + 0.008486920036375523, + 0.012762698344886303, + 0.019406279549002647, + 0.010521642863750458, + 0.012971876189112663, + 0.014183987863361835, + 0.013110287487506866, + 0.015260593965649605, + 0.015145028941333294, + 0.011320851743221283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.029795313253998756 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dining room with a long dark wooden table covered by a beige linen tablecloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several friends wearing cozy sweaters and casual shirts gathered around the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow pan from left to right starting at table level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman's hand in the foreground lifting a glass of white wine that is about one-third full?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the faces of the friends in the background blurred into warm spots of light (bokeh) while the hand remains in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10089506208896637, + 0.03996381163597107, + 0.038647234439849854, + 0.09570664167404175, + 0.02965349145233631, + 0.10989321023225784, + 0.2730904817581177, + 0.13195250928401947, + 0.07959894090890884, + 0.10059858858585358 + ], + "credit_map": [ + 0.01486575324088335, + 0.01160678081214428, + 0.01629304327070713, + 0.009756105951964855, + 0.01009324286133051, + 0.010410048998892307, + 0.009540162049233913, + 0.0183299221098423, + 0.005289655178785324, + 0.005561964586377144, + 0.004486003424972296, + 0.006381381768733263, + 0.0048372093588113785, + 0.00437166029587388, + 0.0036718123592436314, + 0.005364126525819302, + 0.005421333946287632, + 0.005053404718637466, + 0.004091889597475529, + 0.004342579282820225, + 0.004394129384309053, + 0.0065222536213696, + 0.003720944980159402, + 0.005100699607282877, + 0.010114110074937344, + 0.015864308923482895, + 0.01366485096514225, + 0.0112287737429142, + 0.014048750512301922, + 0.014587725512683392, + 0.006311729084700346, + 0.009886393323540688, + 0.0030226651579141617, + 0.0031684713903814554, + 0.002481853822246194, + 0.002578380052000284, + 0.003285840153694153, + 0.005770024377852678, + 0.00473398994654417, + 0.004612267948687077, + 0.012082510627806187, + 0.015441885218024254, + 0.009947750717401505, + 0.009863653220236301, + 0.015513470396399498, + 0.014525016769766808, + 0.020824147388339043, + 0.0116947703063488, + 0.06008777394890785, + 0.03698292747139931, + 0.03502347692847252, + 0.029113145545125008, + 0.02521417662501335, + 0.03494224697351456, + 0.022902756929397583, + 0.0288239736109972, + 0.020735688507556915, + 0.019298763945698738, + 0.021000411361455917, + 0.018489757552742958, + 0.013817316852509975, + 0.01622120290994644, + 0.01386197842657566, + 0.008527386002242565, + 0.007130709011107683, + 0.02178768813610077, + 0.009080898948013783, + 0.007460068445652723, + 0.006886755581945181, + 0.018201224505901337, + 0.0038094688206911087, + 0.005242126528173685, + 0.01790233887732029, + 0.009213191457092762, + 0.01105054747313261, + 0.016325771808624268, + 0.009759632870554924, + 0.01246274821460247, + 0.011038996279239655, + 0.012845364399254322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.003615511115640402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_0/index_tensor([7637, 7637])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_0/index_tensor([7637, 7637])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d1582750a5d91883b4932e0f48f079dd7e9650 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_0/index_tensor([7637, 7637])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, executed in high-fidelity 3D rendering, slowly zooms in on a dense cluster of hawthorn berries hanging from a delicate, dark branch. Each spherical berry is a vibrant crimson, perfectly encased in a thick, translucent, crystalline crust of ice, the result of recent freezing rain. The camera's movement is nearly imperceptible, allowing the viewer to absorb every minute detail. A soft, diffuse light from a pale, overcast sky illuminates the scene, capturing the perfectly smooth surface of the ice and creating subtle, luminous highlights. This light refracts through the frozen glaze, imbuing the red berries beneath with an intense inherent warmth. On each berry's curved surface, the photorealistic rendering captures a distorted fish-eye reflection of the surrounding bare, skeletal winter branches against the soft white sky, with the distorted miniature model adding a layer of intricate beauty. The ice's crystalline texture is meticulously rendered with glassy perfection, creating a powerful visual and atmospheric contrast between the cold, static nature of the ice and the vibrant life it retains, establishing an atmosphere of tranquility and breathtaking beauty.", + "seed": 15724, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_0/index_7637_seed_15724_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09426504774019122, + "vlm_logprob": -0.09426504774019122, + "vlm_yes_no_margin": 4.074999988079071, + "vlm_token_credit": -0.09426504774019122, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017027145251631737 + ], + "opposite_logprob_tokens": [ + -5.142027378082275 + ], + "logprob_reward": -0.017027145251631737, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.017027145251631737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06501315534114838, + 0.11890970915555954, + 0.08418930321931839, + 0.055417489260435104, + 0.05905718356370926, + 0.06939134746789932, + 0.0806741788983345, + 0.0634894147515297, + 0.10851392894983292, + 0.2953442931175232 + ], + "credit_map": [ + 0.011583705432713032, + 0.006653936579823494, + 0.01033860445022583, + 0.00940819550305605, + 0.005861321464180946, + 0.006175910588353872, + 0.007261295337229967, + 0.007730187848210335, + 0.02797740139067173, + 0.014776348136365414, + 0.02410556748509407, + 0.01606902852654457, + 0.008415714837610722, + 0.008634616620838642, + 0.009129051119089127, + 0.009801974520087242, + 0.014367060735821724, + 0.011999906040728092, + 0.011049260385334492, + 0.010436361655592918, + 0.007689270190894604, + 0.015269223600625992, + 0.005675453692674637, + 0.007702765520662069, + 0.014149374328553677, + 0.005711062345653772, + 0.007086720317602158, + 0.014747404493391514, + 0.0030610011890530586, + 0.0035472100134938955, + 0.003511151997372508, + 0.0036035606171935797, + 0.009498167783021927, + 0.01043064333498478, + 0.008349885232746601, + 0.0074312323704361916, + 0.005383044946938753, + 0.006298585794866085, + 0.005321941338479519, + 0.006343677174299955, + 0.01190295908600092, + 0.008045176975429058, + 0.009144749492406845, + 0.015508051961660385, + 0.00570998340845108, + 0.007942724041640759, + 0.006315664853900671, + 0.004822035785764456, + 0.009316084906458855, + 0.0074828676879405975, + 0.021925831213593483, + 0.020669208839535713, + 0.0033731162548065186, + 0.006728603504598141, + 0.006887176539748907, + 0.004291287157684565, + 0.011202453635632992, + 0.013677852228283882, + 0.008635944686830044, + 0.007258806377649307, + 0.005573693662881851, + 0.003901949618011713, + 0.008016828447580338, + 0.005221880506724119, + 0.01486929226666689, + 0.01007189229130745, + 0.014034860767424107, + 0.013660227879881859, + 0.015222049318253994, + 0.010627452284097672, + 0.021077927201986313, + 0.008950239047408104, + 0.04539767652750015, + 0.032830581068992615, + 0.03783634677529335, + 0.0442538782954216, + 0.037696197628974915, + 0.027623876929283142, + 0.0415828600525856, + 0.02812289632856846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.017027145251631737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005275614093989134 + ], + "opposite_logprob_tokens": [ + -5.255275726318359 + ], + "logprob_reward": -0.005275614093989134, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005275614093989134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22294054925441742, + 0.14539100229740143, + 0.1727134734392166, + 0.050367847084999084, + 0.029489872977137566, + 0.048354726284742355, + 0.06084514409303665, + 0.11644548177719116, + 0.10158201307058334, + 0.05186983197927475 + ], + "credit_map": [ + 0.042596910148859024, + 0.028487080708146095, + 0.037755243480205536, + 0.02524951472878456, + 0.017934665083885193, + 0.0230184905230999, + 0.022152496501803398, + 0.025746140629053116, + 0.03238857164978981, + 0.018377920612692833, + 0.03617382422089577, + 0.0246539618819952, + 0.008761539123952389, + 0.008704852312803268, + 0.007210530806332827, + 0.009119792841374874, + 0.03856148570775986, + 0.023695072159171104, + 0.027363691478967667, + 0.025046536698937416, + 0.01426591444760561, + 0.01674136519432068, + 0.014366612769663334, + 0.012672800570726395, + 0.010869741439819336, + 0.0057166810147464275, + 0.007649946957826614, + 0.01186061929911375, + 0.002791802631691098, + 0.0034203159157186747, + 0.0043396614491939545, + 0.0037190793082118034, + 0.005293375346809626, + 0.0038442108780145645, + 0.005546910688281059, + 0.0058344691060483456, + 0.0024425152223557234, + 0.0018593226559460163, + 0.0025021478068083525, + 0.002166922204196453, + 0.007304198574274778, + 0.003545221872627735, + 0.004964966792613268, + 0.010573029518127441, + 0.0039579058066010475, + 0.005544291343539953, + 0.00762335816398263, + 0.004841753747314215, + 0.006371283903717995, + 0.006346598267555237, + 0.01265973411500454, + 0.01170828565955162, + 0.0053681256249547005, + 0.006762980949133635, + 0.007406448479741812, + 0.004221686627715826, + 0.022465212270617485, + 0.03232785686850548, + 0.013363559730350971, + 0.01190468855202198, + 0.013198548927903175, + 0.00780391413718462, + 0.008819745853543282, + 0.00656194519251585, + 0.018068622797727585, + 0.008760031312704086, + 0.013818378560245037, + 0.01274107862263918, + 0.013881216756999493, + 0.009300507605075836, + 0.01606251485645771, + 0.008949656039476395, + 0.006130712106823921, + 0.008224350400269032, + 0.006301907356828451, + 0.007076887879520655, + 0.005588497966527939, + 0.0059045772068202496, + 0.00655241496860981, + 0.006090485956519842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.005275614093989134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288693070411682 + ], + "opposite_logprob_tokens": [ + -1.0538692474365234 + ], + "logprob_reward": -0.4288693070411682, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4288693070411682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11865050345659256, + 0.06491609662771225, + 0.05967077985405922, + 0.0456971749663353, + 0.03621470183134079, + 0.059811782091856, + 0.08702797442674637, + 0.07420463860034943, + 0.2948625981807709, + 0.15894368290901184 + ], + "credit_map": [ + 0.020177112892270088, + 0.01898961141705513, + 0.017386900261044502, + 0.02546907216310501, + 0.007308135274797678, + 0.005373185966163874, + 0.00927041471004486, + 0.014676081947982311, + 0.012501014396548271, + 0.007245054934173822, + 0.00788150168955326, + 0.009340769611299038, + 0.005022502038627863, + 0.008935336023569107, + 0.005681740120053291, + 0.008308181539177895, + 0.01361050270497799, + 0.006813735701143742, + 0.007729563396424055, + 0.011971201747655869, + 0.003978657070547342, + 0.006652861833572388, + 0.0038428164552897215, + 0.005071447230875492, + 0.010237341746687889, + 0.005724255461245775, + 0.004841426853090525, + 0.011469538323581219, + 0.0025915910955518484, + 0.004157261922955513, + 0.0037475144490599632, + 0.0029282462783157825, + 0.007626554928719997, + 0.0035656311083585024, + 0.0035535835195332766, + 0.007525024004280567, + 0.0029017182532697916, + 0.00332164135761559, + 0.0043925940990448, + 0.003327957820147276, + 0.008930862881243229, + 0.005520313046872616, + 0.009546477347612381, + 0.008391257375478745, + 0.005646017380058765, + 0.0060614896938204765, + 0.008665469475090504, + 0.007049903739243746, + 0.009726153686642647, + 0.00847734697163105, + 0.022965842857956886, + 0.02420208789408207, + 0.0036608059890568256, + 0.007734927348792553, + 0.006091166287660599, + 0.0041696480475366116, + 0.011969263665378094, + 0.02056393027305603, + 0.010549457743763924, + 0.007896596565842628, + 0.007722212467342615, + 0.005532368551939726, + 0.00508843595162034, + 0.004882375709712505, + 0.06158168241381645, + 0.029017651453614235, + 0.03606606647372246, + 0.03268478438258171, + 0.039157234132289886, + 0.025736983865499496, + 0.04348757490515709, + 0.02713065594434738, + 0.02978256158530712, + 0.015151728875935078, + 0.019054554402828217, + 0.019508324563503265, + 0.02118319272994995, + 0.016864530742168427, + 0.020280852913856506, + 0.017117954790592194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.4288693070411682 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949762649834156 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949762649834156, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949762649834156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11398039758205414, + 0.10688551515340805, + 0.1058691143989563, + 0.031615808606147766, + 0.026688849553465843, + 0.043807532638311386, + 0.06650595366954803, + 0.1839343011379242, + 0.15278474986553192, + 0.1679278165102005 + ], + "credit_map": [ + 0.017148951068520546, + 0.023311953991651535, + 0.019569218158721924, + 0.01749834604561329, + 0.011418606154620647, + 0.00869598425924778, + 0.007447967305779457, + 0.00888937246054411, + 0.031237775459885597, + 0.012548893690109253, + 0.014285842888057232, + 0.01945933699607849, + 0.007871374487876892, + 0.006714563816785812, + 0.006133915390819311, + 0.008633814752101898, + 0.02720344252884388, + 0.012116038240492344, + 0.01320002693682909, + 0.020053809508681297, + 0.007272444199770689, + 0.009952821768820286, + 0.006969018839299679, + 0.00910151470452547, + 0.0062886495143175125, + 0.00468764454126358, + 0.004667406901717186, + 0.0044343010522425175, + 0.0022491118870675564, + 0.0035718304570764303, + 0.003328828839585185, + 0.002388033317402005, + 0.0046013216488063335, + 0.0039191218093037605, + 0.0036497812252491713, + 0.005492123309522867, + 0.00178657250944525, + 0.00246420712210238, + 0.003188092028722167, + 0.0015876288525760174, + 0.008047395385801792, + 0.006463019177317619, + 0.005734275560826063, + 0.007619048003107309, + 0.003695926396176219, + 0.0041557843796908855, + 0.004250799771398306, + 0.003841281635686755, + 0.011492074467241764, + 0.006072660908102989, + 0.014468750916421413, + 0.008748960681259632, + 0.005119652021676302, + 0.008189101703464985, + 0.006612899247556925, + 0.005801853723824024, + 0.03509221225976944, + 0.03925327584147453, + 0.025772448629140854, + 0.019531184807419777, + 0.014987392351031303, + 0.014576095156371593, + 0.013903901912271976, + 0.020817795768380165, + 0.03706013783812523, + 0.014643726870417595, + 0.01968642883002758, + 0.016576332971453667, + 0.019696934148669243, + 0.015143640339374542, + 0.017329832538962364, + 0.012647714465856552, + 0.027517391368746758, + 0.016902614384889603, + 0.021001141518354416, + 0.024069905281066895, + 0.02231764793395996, + 0.017178582027554512, + 0.023295164108276367, + 0.015645358711481094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.005949762649834156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203409664332867 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203409664332867, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203409664332867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1170426607131958, + 0.13751602172851562, + 0.07493162900209427, + 0.05411095172166824, + 0.036514002829790115, + 0.039545170962810516, + 0.07194249331951141, + 0.1723693311214447, + 0.07621770352125168, + 0.21980996429920197 + ], + "credit_map": [ + 0.014545870013535023, + 0.025019383057951927, + 0.018411075696349144, + 0.013343879021704197, + 0.012134191580116749, + 0.01215267926454544, + 0.009448814205825329, + 0.011986792087554932, + 0.029790375381708145, + 0.015667995437979698, + 0.03373260796070099, + 0.02417723461985588, + 0.007702662609517574, + 0.010978199541568756, + 0.005759319290518761, + 0.009707664139568806, + 0.015417308546602726, + 0.009149035438895226, + 0.014574742875993252, + 0.012636397033929825, + 0.005145201459527016, + 0.006749821361154318, + 0.005138975102454424, + 0.0061201597563922405, + 0.013200143352150917, + 0.005331902764737606, + 0.00611939886584878, + 0.015867602080106735, + 0.002343676518648863, + 0.0031134355813264847, + 0.0036820308305323124, + 0.004452776629477739, + 0.007813706994056702, + 0.005300380755215883, + 0.003372647799551487, + 0.009271538816392422, + 0.002093868562951684, + 0.002823006361722946, + 0.0032562518026679754, + 0.0025826108176261187, + 0.005881223827600479, + 0.0032190424390137196, + 0.004799853079020977, + 0.006266002077609301, + 0.0042379372753202915, + 0.004984071012586355, + 0.006190522573888302, + 0.003966526128351688, + 0.007885041646659374, + 0.008398990146815777, + 0.01258959248661995, + 0.009145410731434822, + 0.005358045920729637, + 0.01217013131827116, + 0.008876972831785679, + 0.007518318947404623, + 0.03819126635789871, + 0.02426416240632534, + 0.02389342710375786, + 0.02079954743385315, + 0.011199237778782845, + 0.013521541841328144, + 0.019235437735915184, + 0.02126474678516388, + 0.013010431081056595, + 0.006732551380991936, + 0.009618462063372135, + 0.00919343251734972, + 0.012858389876782894, + 0.00923721119761467, + 0.009089848026633263, + 0.006477389018982649, + 0.04024122655391693, + 0.0235743410885334, + 0.030525915324687958, + 0.032368697226047516, + 0.024968639016151428, + 0.01971920020878315, + 0.02985769882798195, + 0.018554292619228363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.014203409664332867 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, executed in high-fidelity 3D rendering, slowly zooms in on a dense cluster of hawthorn berries hanging from a delicate, dark branch. Each spherical berry is a vibrant crimson, perfectly encased in a thick, translucent, crystalline crust of ice, the result of recent freezing rain. The camera's movement is nearly imperceptible, allowing the viewer to absorb every minute detail. A soft, diffuse light from a pale, overcast sky illuminates the scene, capturing the perfectly smooth surface of the ice and creating subtle, luminous highlights. This light refracts through the frozen glaze, imbuing the red berries beneath with an intense inherent warmth. On each berry's curved surface, the photorealistic rendering captures a distorted fish-eye reflection of the surrounding bare, skeletal winter branches against the soft white sky, with the distorted miniature model adding a layer of intricate beauty. The ice's crystalline texture is meticulously rendered with glassy perfection, creating a powerful visual and atmospheric contrast between the cold, static nature of the ice and the vibrant life it retains, establishing an atmosphere of tranquility and breathtaking beauty.", + "seed": 15725, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_0/index_7637_seed_15725_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008962435508146882, + "vlm_logprob": -0.008962435508146882, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.008962435508146882, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011260882019996643 + ], + "opposite_logprob_tokens": [ + -5.636260986328125 + ], + "logprob_reward": -0.011260882019996643, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.011260882019996643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10280748456716537, + 0.029279081150889397, + 0.11954822391271591, + 0.02580145001411438, + 0.1202167198061943, + 0.0513351671397686, + 0.13697218894958496, + 0.22647814452648163, + 0.05064474791288376, + 0.13691683113574982 + ], + "credit_map": [ + 0.015239512547850609, + 0.010725653730332851, + 0.01119229942560196, + 0.009112467058002949, + 0.014901477843523026, + 0.017699623480439186, + 0.01686006225645542, + 0.00707638543099165, + 0.005226542241871357, + 0.002461822237819433, + 0.007357703056186438, + 0.003279496915638447, + 0.003036312758922577, + 0.0031071475241333246, + 0.0029181665740907192, + 0.001891891472041607, + 0.04257415980100632, + 0.005114205181598663, + 0.007246309891343117, + 0.004572963807731867, + 0.0060063512064516544, + 0.04567575827240944, + 0.003868652507662773, + 0.004489826038479805, + 0.003941373433917761, + 0.002391671296209097, + 0.0045440602116286755, + 0.002296745777130127, + 0.0023775706067681313, + 0.006050038151443005, + 0.0019164232071489096, + 0.0022835670970380306, + 0.01575065590441227, + 0.013435374945402145, + 0.018058553338050842, + 0.0187147855758667, + 0.01059933751821518, + 0.014467756263911724, + 0.012217607349157333, + 0.016972646117210388, + 0.0070711965672671795, + 0.006246133707463741, + 0.007557720877230167, + 0.008794017136096954, + 0.0070412373170256615, + 0.0038322831969708204, + 0.005189306102693081, + 0.005603270139545202, + 0.015279043465852737, + 0.01536382082849741, + 0.01671580784022808, + 0.021109025925397873, + 0.016882730647921562, + 0.015500789508223534, + 0.02052539773285389, + 0.01559557393193245, + 0.056398142129182816, + 0.020158233121037483, + 0.03505624085664749, + 0.017957517877221107, + 0.0417865552008152, + 0.015531078912317753, + 0.01629839837551117, + 0.02329198271036148, + 0.005261621437966824, + 0.007941700518131256, + 0.006891132798045874, + 0.010889659635722637, + 0.006570448633283377, + 0.002699056174606085, + 0.00559377484023571, + 0.004797353409230709, + 0.014955731108784676, + 0.018039505928754807, + 0.017222678288817406, + 0.017655551433563232, + 0.02726767398416996, + 0.016361625865101814, + 0.014554438181221485, + 0.010859627276659012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.011260882019996643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001524715917184949 + ], + "opposite_logprob_tokens": [ + -6.501524925231934 + ], + "logprob_reward": -0.001524715917184949, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001524715917184949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08060619980096817, + 0.14498616755008698, + 0.3530372977256775, + 0.019157512113451958, + 0.03888637199997902, + 0.08808446675539017, + 0.09359284490346909, + 0.05504393205046654, + 0.0727582648396492, + 0.05384695157408714 + ], + "credit_map": [ + 0.013293364085257053, + 0.007711965125054121, + 0.009875875897705555, + 0.006528319325298071, + 0.012242496944963932, + 0.014047177508473396, + 0.010066612623631954, + 0.006840389221906662, + 0.026981769129633904, + 0.0076347035355865955, + 0.0527525469660759, + 0.012077130377292633, + 0.00909467227756977, + 0.01584005355834961, + 0.00894020963460207, + 0.011665069498121738, + 0.12728038430213928, + 0.015109783038496971, + 0.02348197065293789, + 0.011018921621143818, + 0.016804562881588936, + 0.13231059908866882, + 0.012561741285026073, + 0.014469319023191929, + 0.0028478242456912994, + 0.0018379329703748226, + 0.003170857671648264, + 0.0019976748153567314, + 0.0025296674575656652, + 0.003096819156780839, + 0.0019272917415946722, + 0.0017494440544396639, + 0.005225732922554016, + 0.0038446325343102217, + 0.0041769216768443584, + 0.008051699958741665, + 0.0038763564079999924, + 0.004343373700976372, + 0.005035644397139549, + 0.0043320124968886375, + 0.01050229649990797, + 0.009947541169822216, + 0.017710940912365913, + 0.014799023047089577, + 0.012779194861650467, + 0.0070889233611524105, + 0.007914579473435879, + 0.007341965101659298, + 0.01062264759093523, + 0.010235507041215897, + 0.012864768505096436, + 0.012373054400086403, + 0.0177326537668705, + 0.007598843891173601, + 0.011038502678275108, + 0.011126860976219177, + 0.0069214110262691975, + 0.0072788698598742485, + 0.013879398815333843, + 0.0062392461113631725, + 0.00753049599006772, + 0.003969722427427769, + 0.0055536311119794846, + 0.003671156009659171, + 0.010134277865290642, + 0.00742451474070549, + 0.013097688555717468, + 0.007951552048325539, + 0.015893563628196716, + 0.007520841434597969, + 0.00604952871799469, + 0.00468629552051425, + 0.00456236582249403, + 0.006717711221426725, + 0.006880836095660925, + 0.008551979437470436, + 0.012574601918458939, + 0.005268868990242481, + 0.005631461273878813, + 0.003659125417470932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.001524715917184949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288773372769356 + ], + "opposite_logprob_tokens": [ + -3.7732887268066406 + ], + "logprob_reward": -0.023288773372769356, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288773372769356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09666944295167923, + 0.037384506314992905, + 0.13684679567813873, + 0.03615366294980049, + 0.06604897975921631, + 0.0720200166106224, + 0.14984451234340668, + 0.12781091034412384, + 0.23777656257152557, + 0.03944461792707443 + ], + "credit_map": [ + 0.01485628355294466, + 0.011727896519005299, + 0.01143131498247385, + 0.010601136833429337, + 0.012002424336969852, + 0.017361454665660858, + 0.012317675165832043, + 0.0063712578266859055, + 0.007682876195758581, + 0.003255448304116726, + 0.0075658210553228855, + 0.0036240711342543364, + 0.003082999726757407, + 0.004907585214823484, + 0.004163156263530254, + 0.003102549584582448, + 0.049614787101745605, + 0.005668484605848789, + 0.006948291789740324, + 0.004472910892218351, + 0.006163985002785921, + 0.05227689445018768, + 0.005044241901487112, + 0.006657194346189499, + 0.004883733112365007, + 0.003564309561625123, + 0.006377887446433306, + 0.004495991859585047, + 0.0034849338699132204, + 0.004612521734088659, + 0.0048097530379891396, + 0.003924533259123564, + 0.005255760625004768, + 0.008263048715889454, + 0.009205147624015808, + 0.010149967856705189, + 0.007188423536717892, + 0.005842779763042927, + 0.00724171893671155, + 0.012902123853564262, + 0.012573018670082092, + 0.008550374768674374, + 0.010999227873980999, + 0.01170271821320057, + 0.0067849475890398026, + 0.009022709913551807, + 0.006494565866887569, + 0.005892449524253607, + 0.012361864559352398, + 0.014095939695835114, + 0.044508516788482666, + 0.017254384234547615, + 0.012570656836032867, + 0.012539731338620186, + 0.024557417258620262, + 0.01195599790662527, + 0.03229112923145294, + 0.011921808123588562, + 0.019629565998911858, + 0.011270066723227501, + 0.022309161722660065, + 0.00796467438340187, + 0.009486636146903038, + 0.01293786708265543, + 0.023366933688521385, + 0.03202779218554497, + 0.03789956495165825, + 0.04125380888581276, + 0.03672192990779877, + 0.02496838942170143, + 0.02220146358013153, + 0.01933668740093708, + 0.003433041740208864, + 0.005479131359606981, + 0.0050031812861561775, + 0.005398678593337536, + 0.008428167551755905, + 0.005171855911612511, + 0.003925158176571131, + 0.002605401910841465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.023288773372769356 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098112694919109 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004098112694919109, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098112694919109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11788816750049591, + 0.05545857548713684, + 0.09990346431732178, + 0.04204539582133293, + 0.07509978860616684, + 0.07081615924835205, + 0.1829667091369629, + 0.21947097778320312, + 0.06017465144395828, + 0.07617617398500443 + ], + "credit_map": [ + 0.020824851468205452, + 0.011550397612154484, + 0.018821800127625465, + 0.00972170103341341, + 0.014992864802479744, + 0.019359420984983444, + 0.01415721420198679, + 0.008459909819066525, + 0.009430653415620327, + 0.00424536457285285, + 0.013154148124158382, + 0.0039688656106591225, + 0.0038672590162605047, + 0.009878098964691162, + 0.0038370355032384396, + 0.0070771509781479836, + 0.03169562295079231, + 0.006271666381508112, + 0.010552309453487396, + 0.005483192391693592, + 0.004507432691752911, + 0.032737284898757935, + 0.003927768208086491, + 0.0047281826846301556, + 0.006252771709114313, + 0.0030563008040189743, + 0.00988365150988102, + 0.004413660615682602, + 0.003892375621944666, + 0.006794491782784462, + 0.004936590790748596, + 0.0028155518230050802, + 0.008293396793305874, + 0.008625865913927555, + 0.009075924754142761, + 0.01684005744755268, + 0.008561952970921993, + 0.007425629533827305, + 0.009239432401955128, + 0.007037527859210968, + 0.009937560185790062, + 0.009405262768268585, + 0.010440233163535595, + 0.010955453850328922, + 0.00892361719161272, + 0.00522192008793354, + 0.008482503704726696, + 0.007449607830494642, + 0.02305060252547264, + 0.02532169595360756, + 0.023347200825810432, + 0.02760852687060833, + 0.021950535476207733, + 0.012596365995705128, + 0.025236474350094795, + 0.0238553024828434, + 0.050977855920791626, + 0.0196059662848711, + 0.039466872811317444, + 0.02239239774644375, + 0.03556351363658905, + 0.015453843399882317, + 0.01488404255360365, + 0.02112646773457527, + 0.01295972429215908, + 0.006903069093823433, + 0.01326515432447195, + 0.006134792696684599, + 0.0048557124100625515, + 0.007386293262243271, + 0.004694780800491571, + 0.003975127823650837, + 0.008135898038744926, + 0.010704448446631432, + 0.009194008074700832, + 0.01034496258944273, + 0.016864776611328125, + 0.008442725986242294, + 0.007283840794116259, + 0.005205515772104263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.004098112694919109 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639693535864353 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639693535864353, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639693535864353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10735767334699631, + 0.028252501040697098, + 0.20821663737297058, + 0.020458193495869637, + 0.09664121270179749, + 0.09952431172132492, + 0.17106324434280396, + 0.13116970658302307, + 0.09532135725021362, + 0.04199514538049698 + ], + "credit_map": [ + 0.02593565359711647, + 0.009001719765365124, + 0.011902913451194763, + 0.006961980368942022, + 0.009319962002336979, + 0.026649050414562225, + 0.009243736974895, + 0.008342647925019264, + 0.005183998495340347, + 0.0025337820407003164, + 0.007556806784123182, + 0.002347770379856229, + 0.00171894661616534, + 0.003079184563830495, + 0.0023031013552099466, + 0.003528905799612403, + 0.07830820232629776, + 0.007668789010494947, + 0.014515032060444355, + 0.006695125252008438, + 0.006568489596247673, + 0.08128084987401962, + 0.00611070217564702, + 0.0070694442838430405, + 0.002896505407989025, + 0.0021271025761961937, + 0.0032375750597566366, + 0.0019422860350459814, + 0.0025082624051719904, + 0.0032217656262218952, + 0.002332426607608795, + 0.0021922679152339697, + 0.020024791359901428, + 0.009971357882022858, + 0.018261469900608063, + 0.010041291825473309, + 0.008036158978939056, + 0.011591069400310516, + 0.007070748135447502, + 0.01164432242512703, + 0.012912306003272533, + 0.014618617482483387, + 0.013513835147023201, + 0.01786992698907852, + 0.011679667048156261, + 0.011159314773976803, + 0.009522163309156895, + 0.00824847724288702, + 0.013231159187853336, + 0.02331673912703991, + 0.0471905842423439, + 0.023742742836475372, + 0.020899983122944832, + 0.013599777594208717, + 0.01946599781513214, + 0.009616256691515446, + 0.033292584121227264, + 0.011305469088256359, + 0.02111581712961197, + 0.011263467371463776, + 0.023065241053700447, + 0.008840424008667469, + 0.008898922242224216, + 0.013387775979936123, + 0.020511876791715622, + 0.009883974678814411, + 0.020146768540143967, + 0.009258308447897434, + 0.011182073503732681, + 0.011064445599913597, + 0.0065491111017763615, + 0.006724794860929251, + 0.00394830759614706, + 0.006539578549563885, + 0.004824282601475716, + 0.0044197216629981995, + 0.010980288498103619, + 0.004092498682439327, + 0.004075682256370783, + 0.0031147815752774477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.004639693535864353 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, executed in high-fidelity 3D rendering, slowly zooms in on a dense cluster of hawthorn berries hanging from a delicate, dark branch. Each spherical berry is a vibrant crimson, perfectly encased in a thick, translucent, crystalline crust of ice, the result of recent freezing rain. The camera's movement is nearly imperceptible, allowing the viewer to absorb every minute detail. A soft, diffuse light from a pale, overcast sky illuminates the scene, capturing the perfectly smooth surface of the ice and creating subtle, luminous highlights. This light refracts through the frozen glaze, imbuing the red berries beneath with an intense inherent warmth. On each berry's curved surface, the photorealistic rendering captures a distorted fish-eye reflection of the surrounding bare, skeletal winter branches against the soft white sky, with the distorted miniature model adding a layer of intricate beauty. The ice's crystalline texture is meticulously rendered with glassy perfection, creating a powerful visual and atmospheric contrast between the cold, static nature of the ice and the vibrant life it retains, establishing an atmosphere of tranquility and breathtaking beauty.", + "seed": 15726, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_0/index_7637_seed_15726_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02828178619965911, + "vlm_logprob": -0.02828178619965911, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.02828178619965911, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020866861566901207 + ], + "opposite_logprob_tokens": [ + -4.520866870880127 + ], + "logprob_reward": -0.020866861566901207, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.020866861566901207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19963036477565765, + 0.1332743614912033, + 0.05582338199019432, + 0.07513558864593506, + 0.06890223175287247, + 0.02680038847029209, + 0.14553160965442657, + 0.15052913129329681, + 0.0865330845117569, + 0.05783984065055847 + ], + "credit_map": [ + 0.03691423311829567, + 0.022441204637289047, + 0.025196142494678497, + 0.05339411273598671, + 0.012783837504684925, + 0.012543155811727047, + 0.020299118012189865, + 0.016058551147580147, + 0.0202435702085495, + 0.026230212301015854, + 0.02408020570874214, + 0.01940772496163845, + 0.008331190794706345, + 0.01276099868118763, + 0.01298112515360117, + 0.00923934206366539, + 0.008461800403892994, + 0.005656488239765167, + 0.013331903144717216, + 0.006334454752504826, + 0.004885844420641661, + 0.0060837529599666595, + 0.0057640839368104935, + 0.005305057857185602, + 0.008900955319404602, + 0.008801561780273914, + 0.012499172240495682, + 0.007846377789974213, + 0.010115810669958591, + 0.007031296845525503, + 0.008057545870542526, + 0.011882863938808441, + 0.009464593604207039, + 0.006028829608112574, + 0.014353813603520393, + 0.007335928734391928, + 0.00814854633063078, + 0.01138751208782196, + 0.0066531626507639885, + 0.005529852584004402, + 0.003573774127289653, + 0.003505453234538436, + 0.005374187603592873, + 0.002939940197393298, + 0.00389980711042881, + 0.0024680315982550383, + 0.0026450532022863626, + 0.002394142560660839, + 0.0215207077562809, + 0.011546103283762932, + 0.020766600966453552, + 0.025637520477175713, + 0.013871386647224426, + 0.027093281969428062, + 0.010401849634945393, + 0.014694159850478172, + 0.041779693216085434, + 0.00636009918525815, + 0.05191786587238312, + 0.014673721045255661, + 0.0070210667327046394, + 0.008414954878389835, + 0.012135461904108524, + 0.008226260542869568, + 0.00904507003724575, + 0.009552482515573502, + 0.008121632970869541, + 0.02397223748266697, + 0.00926713366061449, + 0.006098652724176645, + 0.010440048761665821, + 0.01003582589328289, + 0.01165404636412859, + 0.003402847098186612, + 0.01405419036746025, + 0.005351375788450241, + 0.007270953617990017, + 0.004030267708003521, + 0.004735248163342476, + 0.0073409113101661205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.020866861566901207 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003211461938917637 + ], + "opposite_logprob_tokens": [ + -5.753211498260498 + ], + "logprob_reward": -0.003211461938917637, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003211461938917637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18482810258865356, + 0.13280563056468964, + 0.053649839013814926, + 0.054743874818086624, + 0.07781008630990982, + 0.0480925478041172, + 0.0756395161151886, + 0.07818097621202469, + 0.17811766266822815, + 0.11613168567419052 + ], + "credit_map": [ + 0.03616912662982941, + 0.02022659219801426, + 0.01931980438530445, + 0.041165947914123535, + 0.012059839442372322, + 0.019942285493016243, + 0.018572965636849403, + 0.017371565103530884, + 0.02177184820175171, + 0.027341019362211227, + 0.02106291614472866, + 0.018171260133385658, + 0.01034726481884718, + 0.010691807605326176, + 0.011672616004943848, + 0.011746914125978947, + 0.008588801138103008, + 0.010967908427119255, + 0.009517288766801357, + 0.0056186639703810215, + 0.004333114251494408, + 0.005417913664132357, + 0.003951123915612698, + 0.005255026277154684, + 0.00814913958311081, + 0.006656200625002384, + 0.009798982180655003, + 0.005038641858845949, + 0.005526253487914801, + 0.00923977792263031, + 0.004647274035960436, + 0.00568760558962822, + 0.011363247409462929, + 0.008001329377293587, + 0.011996138840913773, + 0.00855677668005228, + 0.007291871123015881, + 0.011981320567429066, + 0.010200039483606815, + 0.008419368416070938, + 0.0055642761290073395, + 0.003597147297114134, + 0.011225949041545391, + 0.004306186456233263, + 0.006251540966331959, + 0.005768900737166405, + 0.004004247020930052, + 0.007374303415417671, + 0.01259076502174139, + 0.008129803463816643, + 0.013993562199175358, + 0.007520611863583326, + 0.0065235658548772335, + 0.011872226372361183, + 0.007758281193673611, + 0.007250704802572727, + 0.012698638252913952, + 0.007932145148515701, + 0.014570535160601139, + 0.010414115153253078, + 0.00834276806563139, + 0.007909771054983139, + 0.009177741594612598, + 0.007135263178497553, + 0.05149709805846214, + 0.010237183421850204, + 0.04856547340750694, + 0.02409473806619644, + 0.008899562060832977, + 0.009996871463954449, + 0.012387153692543507, + 0.012439591810107231, + 0.033336423337459564, + 0.006670984905213118, + 0.036619432270526886, + 0.007339983247220516, + 0.009506577625870705, + 0.006272381171584129, + 0.006223663222044706, + 0.01016224455088377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.003211461938917637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.100298672914505 + ], + "opposite_logprob_tokens": [ + -2.3502986431121826 + ], + "logprob_reward": -0.100298672914505, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.100298672914505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18501296639442444, + 0.07167553901672363, + 0.04745592549443245, + 0.1068878024816513, + 0.09365686029195786, + 0.047102347016334534, + 0.1303364634513855, + 0.22660201787948608, + 0.028223058208823204, + 0.06304702162742615 + ], + "credit_map": [ + 0.033557288348674774, + 0.017728177830576897, + 0.015838099643588066, + 0.043048154562711716, + 0.015202943235635757, + 0.016835171729326248, + 0.022509777918457985, + 0.020293353125452995, + 0.011374169960618019, + 0.013926628977060318, + 0.009544575586915016, + 0.011259770952165127, + 0.004648040980100632, + 0.008707373403012753, + 0.006510639563202858, + 0.005704346112906933, + 0.0055191474966704845, + 0.005303315352648497, + 0.007737413980066776, + 0.005504784174263477, + 0.006127643398940563, + 0.005274416413158178, + 0.0063521950505673885, + 0.005637009162455797, + 0.01138338353484869, + 0.01352442055940628, + 0.0146176777780056, + 0.009097031317651272, + 0.017320595681667328, + 0.009544906206429005, + 0.013707717880606651, + 0.017692072317004204, + 0.01600455865263939, + 0.009121712297201157, + 0.020001301541924477, + 0.011461465619504452, + 0.006851763464510441, + 0.013431909494102001, + 0.007397797424346209, + 0.009386351332068443, + 0.004942652769386768, + 0.0038308470975607634, + 0.007918758317828178, + 0.005749668926000595, + 0.007631784770637751, + 0.005982063710689545, + 0.005266355350613594, + 0.005780217703431845, + 0.01593608967959881, + 0.013202288188040257, + 0.021634763106703758, + 0.020496467128396034, + 0.014814787544310093, + 0.016004813835024834, + 0.014854727312922478, + 0.013392536900937557, + 0.06356564164161682, + 0.00942561961710453, + 0.07822900265455246, + 0.022313101217150688, + 0.009203379973769188, + 0.012448218651115894, + 0.019274290651082993, + 0.012142769061028957, + 0.003839032491669059, + 0.002395825693383813, + 0.00634737266227603, + 0.0065294299274683, + 0.0025175693444907665, + 0.001924666459672153, + 0.0021836969535797834, + 0.0024854636285454035, + 0.01596633717417717, + 0.0030918975826352835, + 0.01833346113562584, + 0.005445701070129871, + 0.005783847067505121, + 0.006011136341840029, + 0.0033173735719174147, + 0.005097269080579281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.100298672914505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950355436652899 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950355436652899, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950355436652899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23684324324131012, + 0.07382460683584213, + 0.03092869371175766, + 0.04279943183064461, + 0.02656661719083786, + 0.0395062118768692, + 0.1618189960718155, + 0.10766930878162384, + 0.2299654483795166, + 0.050077516585588455 + ], + "credit_map": [ + 0.03989080339670181, + 0.05419659987092018, + 0.022707395255565643, + 0.05070630460977554, + 0.019100939854979515, + 0.016154497861862183, + 0.020551137626171112, + 0.013535557314753532, + 0.012547762133181095, + 0.014309495687484741, + 0.013591861352324486, + 0.00868131686002016, + 0.0061481669545173645, + 0.005394740495830774, + 0.007068165112286806, + 0.0060830977745354176, + 0.0038173941429704428, + 0.004055345896631479, + 0.005874975118786097, + 0.003691562917083502, + 0.0034705286379903555, + 0.0032615417148917913, + 0.0030892558861523867, + 0.0036680882330983877, + 0.007371657062321901, + 0.0037176383193582296, + 0.011022795923054218, + 0.0035286438651382923, + 0.0030970294028520584, + 0.006231680512428284, + 0.0027969121001660824, + 0.005033074412494898, + 0.0035240796860307455, + 0.0032328281085938215, + 0.004362843465059996, + 0.0037801170255988836, + 0.002458774484694004, + 0.0038108795415610075, + 0.0027203639037907124, + 0.00267673097550869, + 0.004278337117284536, + 0.004568996839225292, + 0.005554467439651489, + 0.004234944004565477, + 0.007376787252724171, + 0.004109680652618408, + 0.003268677741289139, + 0.0061143189668655396, + 0.019939087331295013, + 0.01612839661538601, + 0.02061701938509941, + 0.03253055736422539, + 0.017294833436608315, + 0.0255582258105278, + 0.012710673734545708, + 0.01704019494354725, + 0.022803861647844315, + 0.007938768714666367, + 0.028162667527794838, + 0.013875577598810196, + 0.007537527941167355, + 0.009443050250411034, + 0.010165663436055183, + 0.00774219399318099, + 0.07862941175699234, + 0.007940693758428097, + 0.08497407287359238, + 0.01885150745511055, + 0.007678954396396875, + 0.007502526976168156, + 0.014728165231645107, + 0.009660128504037857, + 0.013974436558783054, + 0.003464123234152794, + 0.01478329487144947, + 0.0036918974947184324, + 0.004406943451613188, + 0.002651000628247857, + 0.0030053979717195034, + 0.004100424703210592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.005950355436652899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081579141318798 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081579141318798, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081579141318798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1498912274837494, + 0.061525359749794006, + 0.030116425827145576, + 0.11725557595491409, + 0.07795459032058716, + 0.06944520026445389, + 0.1297217607498169, + 0.11034301668405533, + 0.23359259963035583, + 0.02015424519777298 + ], + "credit_map": [ + 0.02962447889149189, + 0.0135577954351902, + 0.013529997318983078, + 0.0345771498978138, + 0.010231759399175644, + 0.020876172930002213, + 0.015583029016852379, + 0.011910837143659592, + 0.008671601302921772, + 0.010399779304862022, + 0.01047442015260458, + 0.006592382676899433, + 0.006051169242709875, + 0.006872145924717188, + 0.0062273903749883175, + 0.006236474961042404, + 0.003974905237555504, + 0.0051259687170386314, + 0.005942372605204582, + 0.002713814377784729, + 0.0026663220487535, + 0.0033697239123284817, + 0.0032637433614581823, + 0.0030595753341913223, + 0.0122526865452528, + 0.015953563153743744, + 0.014567958191037178, + 0.011214314959943295, + 0.016195688396692276, + 0.011628060601651669, + 0.012355957180261612, + 0.02308734506368637, + 0.012404251843690872, + 0.008848311379551888, + 0.02040504664182663, + 0.007758872117847204, + 0.005445366259664297, + 0.010438903234899044, + 0.004413723945617676, + 0.008240110240876675, + 0.008267931640148163, + 0.006013781297951937, + 0.01589646190404892, + 0.0066850152797997, + 0.00869238842278719, + 0.00767937395721674, + 0.00633232481777668, + 0.009877924807369709, + 0.015666617080569267, + 0.013403377495706081, + 0.022352352738380432, + 0.0249843280762434, + 0.01585507020354271, + 0.014027770608663559, + 0.009693530388176441, + 0.013738708570599556, + 0.01651780866086483, + 0.011109748855233192, + 0.02107321098446846, + 0.016602251678705215, + 0.00882668700069189, + 0.01109477411955595, + 0.016680259257555008, + 0.008438276126980782, + 0.07437074929475784, + 0.011660783551633358, + 0.0712287500500679, + 0.029339099302887917, + 0.009666753932833672, + 0.008753244765102863, + 0.015105996280908585, + 0.013467208482325077, + 0.0030485978350043297, + 0.002147714840248227, + 0.002188517479225993, + 0.002523405011743307, + 0.0027989214286208153, + 0.0021499658469110727, + 0.0025111574213951826, + 0.0027859658002853394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.011081579141318798 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, executed in high-fidelity 3D rendering, slowly zooms in on a dense cluster of hawthorn berries hanging from a delicate, dark branch. Each spherical berry is a vibrant crimson, perfectly encased in a thick, translucent, crystalline crust of ice, the result of recent freezing rain. The camera's movement is nearly imperceptible, allowing the viewer to absorb every minute detail. A soft, diffuse light from a pale, overcast sky illuminates the scene, capturing the perfectly smooth surface of the ice and creating subtle, luminous highlights. This light refracts through the frozen glaze, imbuing the red berries beneath with an intense inherent warmth. On each berry's curved surface, the photorealistic rendering captures a distorted fish-eye reflection of the surrounding bare, skeletal winter branches against the soft white sky, with the distorted miniature model adding a layer of intricate beauty. The ice's crystalline texture is meticulously rendered with glassy perfection, creating a powerful visual and atmospheric contrast between the cold, static nature of the ice and the vibrant life it retains, establishing an atmosphere of tranquility and breathtaking beauty.", + "seed": 15727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_0/index_7637_seed_15727_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01794171454384923, + "vlm_logprob": -0.01794171454384923, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.01794171454384923, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015727292746305466 + ], + "opposite_logprob_tokens": [ + -5.140727519989014 + ], + "logprob_reward": -0.015727292746305466, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.015727292746305466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11809290945529938, + 0.045575905591249466, + 0.05630185827612877, + 0.07832540571689606, + 0.05664444714784622, + 0.04479691758751869, + 0.0937628298997879, + 0.1851234883069992, + 0.2810775637626648, + 0.04029866307973862 + ], + "credit_map": [ + 0.017429238185286522, + 0.02236596681177616, + 0.01622854359447956, + 0.01144277211278677, + 0.01597166433930397, + 0.009116456843912601, + 0.013996719382703304, + 0.011541545391082764, + 0.006126512307673693, + 0.004878680221736431, + 0.005642244592308998, + 0.004915979225188494, + 0.004896193742752075, + 0.004480287432670593, + 0.006609585601836443, + 0.008026418276131153, + 0.008026869036257267, + 0.005745731294155121, + 0.008748915046453476, + 0.005361391697078943, + 0.004533757921308279, + 0.0054832035675644875, + 0.005974766332656145, + 0.012427224777638912, + 0.016000106930732727, + 0.005957937799394131, + 0.025476614013314247, + 0.005815993063151836, + 0.009174271486699581, + 0.004961489699780941, + 0.005650448612868786, + 0.0052885450422763824, + 0.008677328005433083, + 0.006361039821058512, + 0.014847435988485813, + 0.005067141726613045, + 0.006471436470746994, + 0.005542390514165163, + 0.005434494931250811, + 0.004243180155754089, + 0.006469294428825378, + 0.005673026666045189, + 0.00896379817277193, + 0.0061989896930754185, + 0.0026633176021277905, + 0.004492366686463356, + 0.0063547431491315365, + 0.003981378395110369, + 0.010496518574655056, + 0.01190670020878315, + 0.015167677775025368, + 0.009904799982905388, + 0.008334197103977203, + 0.013996333815157413, + 0.011869573965668678, + 0.012087036855518818, + 0.028059078380465508, + 0.03720659017562866, + 0.026163358241319656, + 0.024306315928697586, + 0.013768371194601059, + 0.02051178552210331, + 0.01722932606935501, + 0.017878664657473564, + 0.04097932577133179, + 0.06283532083034515, + 0.030246425420045853, + 0.038559459149837494, + 0.026234449818730354, + 0.01863666996359825, + 0.02732183039188385, + 0.0362640805542469, + 0.0025199756491929293, + 0.01060155313462019, + 0.004672920797020197, + 0.006629048381000757, + 0.004043308552354574, + 0.0035904801916331053, + 0.0034599383361637592, + 0.004781440366059542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.015727292746305466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003204688895493746 + ], + "opposite_logprob_tokens": [ + -5.753204822540283 + ], + "logprob_reward": -0.003204688895493746, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003204688895493746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1891781985759735, + 0.07607875019311905, + 0.06277494877576828, + 0.211781308054924, + 0.06707090139389038, + 0.06263818591833115, + 0.023240139707922935, + 0.11719078570604324, + 0.046651046723127365, + 0.14339570701122284 + ], + "credit_map": [ + 0.02268853411078453, + 0.024684330448508263, + 0.01630268804728985, + 0.011939488351345062, + 0.05030754581093788, + 0.01707017980515957, + 0.021635014563798904, + 0.024550428614020348, + 0.0095139155164361, + 0.0126234395429492, + 0.00714913522824645, + 0.00761755183339119, + 0.007351698819547892, + 0.006701943464577198, + 0.00897178053855896, + 0.016149284318089485, + 0.011113246902823448, + 0.00546553498134017, + 0.01236153207719326, + 0.005141341593116522, + 0.005594548769295216, + 0.004671627655625343, + 0.00823180004954338, + 0.010195317678153515, + 0.0427611842751503, + 0.01845017820596695, + 0.07767488807439804, + 0.014559556730091572, + 0.01418271567672491, + 0.017181091010570526, + 0.01265238132327795, + 0.01431932020932436, + 0.008295638486742973, + 0.00999508798122406, + 0.008545592427253723, + 0.00919939111918211, + 0.006311282515525818, + 0.010084195993840694, + 0.007816762663424015, + 0.006822956260293722, + 0.008219017647206783, + 0.00960079487413168, + 0.011832796968519688, + 0.00787400733679533, + 0.005367088597267866, + 0.006708771456032991, + 0.0072730365209281445, + 0.0057626753114163876, + 0.002932284725829959, + 0.0025795914698392153, + 0.0048250299878418446, + 0.0031060753390192986, + 0.0019050425617024302, + 0.002659113146364689, + 0.0022119644563645124, + 0.003021039068698883, + 0.017284244298934937, + 0.020871393382549286, + 0.017016403377056122, + 0.015246379189193249, + 0.008238622918725014, + 0.01332250889390707, + 0.011935273185372353, + 0.013275965116918087, + 0.006534102372825146, + 0.009977118112146854, + 0.006430669687688351, + 0.005410333164036274, + 0.004770227242261171, + 0.004322702996432781, + 0.004218692891299725, + 0.0049872007220983505, + 0.015257151797413826, + 0.02889864332973957, + 0.020922590047121048, + 0.017713667824864388, + 0.012065951712429523, + 0.012975252233445644, + 0.014102092012763023, + 0.021460358053445816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.003204688895493746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204032897949219 + ], + "opposite_logprob_tokens": [ + -2.812040328979492 + ], + "logprob_reward": -0.06204032897949219, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204032897949219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041505567729473114, + 0.06383833289146423, + 0.039663542062044144, + 0.06469380855560303, + 0.05742993950843811, + 0.05683528631925583, + 0.09889111667871475, + 0.43871206045150757, + 0.07531677931547165, + 0.06311353296041489 + ], + "credit_map": [ + 0.006665100809186697, + 0.004509828984737396, + 0.010216246359050274, + 0.004013481084257364, + 0.0038575187791138887, + 0.0027219541370868683, + 0.004472989123314619, + 0.005048452876508236, + 0.007898668758571148, + 0.008983059786260128, + 0.00897576380521059, + 0.0048674060963094234, + 0.00501538161188364, + 0.005299326963722706, + 0.006427142769098282, + 0.016371585428714752, + 0.005612269043922424, + 0.0030694121960550547, + 0.00887630507349968, + 0.0029806557577103376, + 0.003451921511441469, + 0.0032343862112611532, + 0.004753268789499998, + 0.0076853265054523945, + 0.010794834233820438, + 0.005234348122030497, + 0.02091255597770214, + 0.004741917364299297, + 0.005994067993015051, + 0.0057373857125639915, + 0.004412402398884296, + 0.006866304203867912, + 0.010845961980521679, + 0.004382766783237457, + 0.020683353766798973, + 0.004291445016860962, + 0.003685307689011097, + 0.004738515242934227, + 0.004614581819623709, + 0.0041880072094500065, + 0.007287532091140747, + 0.01182529516518116, + 0.011004816740751266, + 0.00696409959346056, + 0.006346479058265686, + 0.0036749232094734907, + 0.005324556026607752, + 0.0044075907208025455, + 0.009556818753480911, + 0.021567530930042267, + 0.011791246011853218, + 0.010518272407352924, + 0.014034937135875225, + 0.013233866542577744, + 0.010378695093095303, + 0.007809753064066172, + 0.10424959659576416, + 0.05990922823548317, + 0.1108483374118805, + 0.044861800968647, + 0.02240927703678608, + 0.035527553409338, + 0.028130073100328445, + 0.03277624770998955, + 0.01232095155864954, + 0.014823895879089832, + 0.01050630770623684, + 0.01026333961635828, + 0.0057792458683252335, + 0.006014948710799217, + 0.0055401381105184555, + 0.010067958384752274, + 0.005589414853602648, + 0.0158278476446867, + 0.007433673366904259, + 0.007912395521998405, + 0.006293587852269411, + 0.005609063897281885, + 0.007341817021369934, + 0.007105735596269369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.06204032897949219 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095382057130337 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095382057130337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095382057130337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13487236201763153, + 0.046499479562044144, + 0.14563006162643433, + 0.07891745865345001, + 0.1352422535419464, + 0.060191161930561066, + 0.030035350471735, + 0.25657153129577637, + 0.062072087079286575, + 0.04996826872229576 + ], + "credit_map": [ + 0.018616026267409325, + 0.028844524174928665, + 0.025902189314365387, + 0.011500788852572441, + 0.010523376986384392, + 0.0112144248560071, + 0.011203103698790073, + 0.017067918553948402, + 0.006757848896086216, + 0.005958962719887495, + 0.011172396130859852, + 0.004076118115335703, + 0.0032573265489190817, + 0.0033746061380952597, + 0.003622752381488681, + 0.008279469795525074, + 0.021339787170290947, + 0.010185305029153824, + 0.04440010339021683, + 0.013121829368174076, + 0.012715534307062626, + 0.014062358066439629, + 0.016550080850720406, + 0.013255068100988865, + 0.009381008334457874, + 0.009352813474833965, + 0.01140880212187767, + 0.007867543026804924, + 0.012411801144480705, + 0.008947986178100109, + 0.007851113565266132, + 0.011696389876306057, + 0.015480116941034794, + 0.01920926757156849, + 0.025616342201828957, + 0.015218837186694145, + 0.01646595261991024, + 0.015550127252936363, + 0.01484993752092123, + 0.01285167969763279, + 0.009101053699851036, + 0.009722741320729256, + 0.014654085040092468, + 0.0064043281599879265, + 0.004487042780965567, + 0.005936926230788231, + 0.005032781511545181, + 0.00485220504924655, + 0.0034715007059276104, + 0.004282903857529163, + 0.004428358282893896, + 0.0025755148380994797, + 0.0026360005140304565, + 0.0035639393609017134, + 0.003364437259733677, + 0.005712694022804499, + 0.07401055842638016, + 0.02129961922764778, + 0.0817907452583313, + 0.022219914942979813, + 0.010756190866231918, + 0.015811681747436523, + 0.013916990719735622, + 0.016765840351581573, + 0.009715480729937553, + 0.011859067715704441, + 0.007845841348171234, + 0.007237490266561508, + 0.00612470880150795, + 0.0046075559221208096, + 0.005478824954479933, + 0.009203115478157997, + 0.004691031761467457, + 0.009323961101472378, + 0.007499872241169214, + 0.00667176116257906, + 0.004543579649180174, + 0.004992511123418808, + 0.005081558600068092, + 0.007163994014263153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.004095382057130337 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a dense cluster of hawthorn berries on a dark branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the berries spherical, vibrant crimson, and encased in a thick translucent crystalline crust of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly zoom in on the berries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft diffuse light from a pale overcast sky, creating luminous highlights on the ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the berries show a distorted fish-eye reflection of bare skeletal winter branches against a soft white sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640880040824413 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640880040824413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640880040824413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18153414130210876, + 0.03586987406015396, + 0.02152734436094761, + 0.07318443804979324, + 0.18149778246879578, + 0.03450202941894531, + 0.06907723844051361, + 0.23702320456504822, + 0.07445432245731354, + 0.09132955968379974 + ], + "credit_map": [ + 0.0320456400513649, + 0.01820693351328373, + 0.04959665238857269, + 0.012783469632267952, + 0.019373513758182526, + 0.01154432911425829, + 0.016770506277680397, + 0.02121308632194996, + 0.0055999308824539185, + 0.003491354873403907, + 0.008574726060032845, + 0.00274701532907784, + 0.0033869503531605005, + 0.002794150495901704, + 0.003283095546066761, + 0.0059926495887339115, + 0.0032474768813699484, + 0.0017889720620587468, + 0.006130221299827099, + 0.002068996662274003, + 0.002142206998541951, + 0.0017932966584339738, + 0.0021201998461037874, + 0.0022359718568623066, + 0.008199675939977169, + 0.00661424407735467, + 0.012320758774876595, + 0.008038777858018875, + 0.008219682611525059, + 0.010816982015967369, + 0.007900459691882133, + 0.011073855683207512, + 0.021828746423125267, + 0.028338005766272545, + 0.03232938423752785, + 0.01955697126686573, + 0.01817796751856804, + 0.026414338499307632, + 0.018824761733412743, + 0.01602761819958687, + 0.004891914781183004, + 0.004687980283051729, + 0.005799547769129276, + 0.004059338942170143, + 0.003842695849016309, + 0.0035162675194442272, + 0.004583514295518398, + 0.0031207692809402943, + 0.007354429457336664, + 0.011481872759759426, + 0.008139634504914284, + 0.007825786247849464, + 0.00923837162554264, + 0.009906072169542313, + 0.007617135066539049, + 0.007513932418078184, + 0.07382357865571976, + 0.019641146063804626, + 0.0803273394703865, + 0.014202258549630642, + 0.008594789542257786, + 0.012693839147686958, + 0.012065699324011803, + 0.015674574300646782, + 0.013249145820736885, + 0.010333101265132427, + 0.010849368758499622, + 0.009984963573515415, + 0.006706384941935539, + 0.005587860941886902, + 0.007445187773555517, + 0.010298311710357666, + 0.007493085693567991, + 0.02615690790116787, + 0.009785357862710953, + 0.011873343028128147, + 0.008036081679165363, + 0.006461402866989374, + 0.0078382957726717, + 0.013685077428817749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.004640880040824413 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_tensor([6754, 6754])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_tensor([6754, 6754])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..01e67bec9383c8c998e4139e47424a5fc44b1150 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_tensor([6754, 6754])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle camera slowly tilts downward and zooms in, revealing a vast and desolate alien world. The entire scene accurately captures the distinctive, jagged, rough, polygonal textures of the PS1 era. A massive, ring-shaped spaceship, constructed from simple, low-polygons, rests quietly atop a sepia-toned volcanic landscape composed of sharp, angular polygons. Its metal hull appears a dark gray texture with pixelated noise and a simple lightmap. Surrounding the ship is a flowing river of incandescent lava. Its surface, far from a smooth liquid, consists of bright orange-red and yellow blocks of pixels, slowly creeping in a choppy, skipping pattern, providing the scene's sole primary source of light. This fiery glow casts flickering, pulsating orange shadows across the dark polygonal surface of the spaceship and the surrounding rocks. The lighting calculations are simple and direct, with no soft transitions. The deep purple sky, a simple, cloudless color gradient, further enhances this retro aesthetic of early 3D rendering, creating a powerful, desolate atmosphere that is both mysterious and imbued with a sense of digital nostalgia.", + "seed": 12821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_6754_seed_12821_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03423204803839326, + "vlm_logprob": -0.03423204803839326, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03423204803839326, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04405072331428528 + ], + "opposite_logprob_tokens": [ + -3.294050693511963 + ], + "logprob_reward": -0.04405072331428528, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04405072331428528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05165321007370949, + 0.03208204358816147, + 0.028506923466920853, + 0.05052358657121658, + 0.09903831779956818, + 0.11504267156124115, + 0.06653378903865814, + 0.12315192818641663, + 0.04636331647634506, + 0.38710424304008484 + ], + "credit_map": [ + 0.005536774639040232, + 0.006726142484694719, + 0.0053041367791593075, + 0.0071120308712124825, + 0.0066108908504247665, + 0.005802231840789318, + 0.007266635540872812, + 0.007294368464499712, + 0.0035839248448610306, + 0.003647363046184182, + 0.0031469680834561586, + 0.00322296773083508, + 0.006564532872289419, + 0.0026622081641107798, + 0.004491094499826431, + 0.004762981552630663, + 0.002354481490328908, + 0.0030955348629504442, + 0.00359801365993917, + 0.004164823330938816, + 0.00523896049708128, + 0.002493673237040639, + 0.003634430468082428, + 0.003927006386220455, + 0.003994687460362911, + 0.006304106675088406, + 0.006850940641015768, + 0.00579633517190814, + 0.011545932851731777, + 0.005024721845984459, + 0.0057292720302939415, + 0.005277590826153755, + 0.006852574646472931, + 0.010853869840502739, + 0.019186465069651604, + 0.008660753257572651, + 0.024722183123230934, + 0.010567509569227695, + 0.0088810371235013, + 0.009313921444118023, + 0.010360384359955788, + 0.016766877844929695, + 0.02563057467341423, + 0.006193755194544792, + 0.02099168859422207, + 0.01486910693347454, + 0.006456831935793161, + 0.013773449696600437, + 0.005289500579237938, + 0.007134552113711834, + 0.009553786367177963, + 0.0033505174797028303, + 0.01707492209970951, + 0.006842557340860367, + 0.00366632966324687, + 0.013621626421809196, + 0.006189460400491953, + 0.007347347214818001, + 0.032899413257837296, + 0.005481100641191006, + 0.021203147247433662, + 0.007811821531504393, + 0.027565525844693184, + 0.014654116705060005, + 0.0027705307584255934, + 0.0031510426197201014, + 0.002783467061817646, + 0.0022507766261696815, + 0.004340454936027527, + 0.0029274660628288984, + 0.011869334615767002, + 0.016270244494080544, + 0.013239949941635132, + 0.013208216987550259, + 0.051016248762607574, + 0.02408900484442711, + 0.01806630752980709, + 0.024481533095240593, + 0.09642282128334045, + 0.14658018946647644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.04405072331428528 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004113190341740847 + ], + "opposite_logprob_tokens": [ + -5.50411319732666 + ], + "logprob_reward": -0.004113190341740847, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004113190341740847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0869985967874527, + 0.09234126657247543, + 0.21180954575538635, + 0.059316541999578476, + 0.07315769046545029, + 0.06122583895921707, + 0.10049707442522049, + 0.09648143500089645, + 0.055937908589839935, + 0.16223403811454773 + ], + "credit_map": [ + 0.009423762559890747, + 0.008065630681812763, + 0.014115510508418083, + 0.01467973180115223, + 0.01006247941404581, + 0.006640053819864988, + 0.01187187060713768, + 0.012139559723436832, + 0.008083987049758434, + 0.009022963233292103, + 0.008823777548968792, + 0.01813366636633873, + 0.013188212178647518, + 0.0079758046194911, + 0.013325081206858158, + 0.013787777163088322, + 0.020398452877998352, + 0.027036137878894806, + 0.028832323849201202, + 0.022549424320459366, + 0.03513580188155174, + 0.02153792418539524, + 0.024849919602274895, + 0.03146955370903015, + 0.004259590990841389, + 0.007524682208895683, + 0.009374712593853474, + 0.007797962985932827, + 0.007312735542654991, + 0.005639444105327129, + 0.00755546148866415, + 0.009851952083408833, + 0.004699381534010172, + 0.007832588627934456, + 0.014531828463077545, + 0.005991537123918533, + 0.01787104457616806, + 0.00836874172091484, + 0.006080615799874067, + 0.007781952153891325, + 0.004605254624038935, + 0.006271854974329472, + 0.010778970085084438, + 0.0063881040550768375, + 0.013195018284022808, + 0.007399954367429018, + 0.005397109314799309, + 0.00718957232311368, + 0.0068180495873093605, + 0.009661225602030754, + 0.01026987936347723, + 0.0036769649013876915, + 0.030888834968209267, + 0.011561878025531769, + 0.005359663162380457, + 0.022260580211877823, + 0.005906728561967611, + 0.004475114401429892, + 0.03670520707964897, + 0.00476904446259141, + 0.006811987608671188, + 0.00355304847471416, + 0.01976279355585575, + 0.014497515745460987, + 0.0030965562909841537, + 0.004985209554433823, + 0.002953240415081382, + 0.0021639864426106215, + 0.008850881829857826, + 0.005040113814175129, + 0.010805810801684856, + 0.01804211176931858, + 0.010463692247867584, + 0.018468545749783516, + 0.01082083210349083, + 0.012653783895075321, + 0.013929987326264381, + 0.02179257571697235, + 0.04447547718882561, + 0.029629157856106758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.004113190341740847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349563151597977 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349563151597977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349563151597977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08934717625379562, + 0.22494231164455414, + 0.10003072023391724, + 0.030495338141918182, + 0.10362482070922852, + 0.03524298593401909, + 0.08222216367721558, + 0.08878310769796371, + 0.13055320084095, + 0.11475817859172821 + ], + "credit_map": [ + 0.008505580946803093, + 0.00871868897229433, + 0.01319941971451044, + 0.014445624314248562, + 0.008053455501794815, + 0.009902981109917164, + 0.013024681247770786, + 0.013496747240424156, + 0.02206587791442871, + 0.0195881649851799, + 0.030573764815926552, + 0.04831375181674957, + 0.03317221626639366, + 0.012547757476568222, + 0.03327756002545357, + 0.025403212755918503, + 0.007665128447115421, + 0.008578410372138023, + 0.016463980078697205, + 0.024217238649725914, + 0.006135453935712576, + 0.007233734708279371, + 0.017835361883044243, + 0.011901404708623886, + 0.0026918870862573385, + 0.0027443254366517067, + 0.005025951657444239, + 0.004131884779781103, + 0.0032964020501822233, + 0.002050114097073674, + 0.0043854727409780025, + 0.006169300060719252, + 0.008766630664467812, + 0.011898377910256386, + 0.021930573508143425, + 0.01262421254068613, + 0.013774841092526913, + 0.008831665851175785, + 0.009484866634011269, + 0.016313649713993073, + 0.003032740205526352, + 0.0038907101843506098, + 0.005638834089040756, + 0.006489336956292391, + 0.0034321199636906385, + 0.0028716870583593845, + 0.0046533276326954365, + 0.00523422984406352, + 0.005280612502247095, + 0.005772240459918976, + 0.022855622693896294, + 0.004685215186327696, + 0.014271597377955914, + 0.007407049182802439, + 0.004973807837814093, + 0.016976015642285347, + 0.005502996034920216, + 0.0066709090024232864, + 0.01640835963189602, + 0.004344037268310785, + 0.008485684171319008, + 0.004953743889927864, + 0.010843558236956596, + 0.0315738208591938, + 0.009921830147504807, + 0.008757749572396278, + 0.0068259513936936855, + 0.0046598538756370544, + 0.011973033659160137, + 0.00851041916757822, + 0.02783684991300106, + 0.05206751078367233, + 0.007204217836260796, + 0.006765717174857855, + 0.009605041705071926, + 0.004804066848009825, + 0.005691394209861755, + 0.009567941538989544, + 0.038989461958408356, + 0.03213034197688103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.026349563151597977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628239691257477 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628239691257477, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628239691257477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08244287222623825, + 0.06188926100730896, + 0.1258612871170044, + 0.18064697086811066, + 0.11221536993980408, + 0.07550442963838577, + 0.04838929697871208, + 0.131114661693573, + 0.08604899793863297, + 0.09588683396577835 + ], + "credit_map": [ + 0.0074957916513085365, + 0.0073923347517848015, + 0.011533219367265701, + 0.011225983500480652, + 0.012665211223065853, + 0.007438902277499437, + 0.015174690634012222, + 0.009516743011772633, + 0.004770671483129263, + 0.00529318954795599, + 0.008912136778235435, + 0.013506518676877022, + 0.0062040407210588455, + 0.003892275970429182, + 0.012708866968750954, + 0.006601567380130291, + 0.009891999885439873, + 0.008331647142767906, + 0.018160898238420486, + 0.034661825746297836, + 0.007712842896580696, + 0.005614535883069038, + 0.016539214178919792, + 0.024948343634605408, + 0.013330419547855854, + 0.017255326732993126, + 0.030941912904381752, + 0.02713623456656933, + 0.020769409835338593, + 0.013168400153517723, + 0.031728681176900864, + 0.026316607370972633, + 0.007819687947630882, + 0.013109329156577587, + 0.020603753626346588, + 0.009323574602603912, + 0.027291886508464813, + 0.010802301578223705, + 0.009820420295000076, + 0.013444419950246811, + 0.006550118792802095, + 0.0077688549645245075, + 0.017747802659869194, + 0.0066072093322873116, + 0.013608581386506557, + 0.007829217240214348, + 0.0055670770816504955, + 0.009825572371482849, + 0.00441738311201334, + 0.005038258153945208, + 0.010045199654996395, + 0.004755072295665741, + 0.006331603042781353, + 0.005655220244079828, + 0.004108884837478399, + 0.00803767517209053, + 0.011250961571931839, + 0.010123390704393387, + 0.040621068328619, + 0.011509918607771397, + 0.014129436574876308, + 0.005675580818206072, + 0.022365713492035866, + 0.015438598580658436, + 0.006399752572178841, + 0.004742066841572523, + 0.007790710311383009, + 0.005170181393623352, + 0.007812247611582279, + 0.004744320642203093, + 0.021801749244332314, + 0.027587980031967163, + 0.005117629189044237, + 0.007193176541477442, + 0.009081616066396236, + 0.005614602006971836, + 0.00681066932156682, + 0.014439551159739494, + 0.028677158057689667, + 0.018952440470457077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.007628239691257477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901852369308472 + ], + "opposite_logprob_tokens": [ + -2.4640185832977295 + ], + "logprob_reward": -0.08901852369308472, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901852369308472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14683759212493896, + 0.07052268087863922, + 0.08366677165031433, + 0.09424848109483719, + 0.12734302878379822, + 0.05593235790729523, + 0.026465516537427902, + 0.2002718299627304, + 0.04225710779428482, + 0.15245462954044342 + ], + "credit_map": [ + 0.016464538872241974, + 0.01765863597393036, + 0.020753612741827965, + 0.01821531169116497, + 0.01680465042591095, + 0.01523854024708271, + 0.021237073466181755, + 0.020465228706598282, + 0.006783992052078247, + 0.006323512643575668, + 0.00711699016392231, + 0.013542386703193188, + 0.009842293336987495, + 0.007391978520900011, + 0.010577265173196793, + 0.008944260887801647, + 0.006695503368973732, + 0.008041624911129475, + 0.013414228335022926, + 0.017353208735585213, + 0.009921891614794731, + 0.006087123416364193, + 0.011591674759984016, + 0.010561518371105194, + 0.008457661606371403, + 0.012010592967271805, + 0.013276973739266396, + 0.011184526607394218, + 0.01357466820627451, + 0.009609696455299854, + 0.012538242153823376, + 0.013596120290458202, + 0.017220789566636086, + 0.00899972952902317, + 0.02707561105489731, + 0.015587333589792252, + 0.018423806875944138, + 0.008207433857023716, + 0.01509919948875904, + 0.01672913134098053, + 0.005271946080029011, + 0.005950367543846369, + 0.012281186878681183, + 0.008011527359485626, + 0.0059488750994205475, + 0.0037119644694030285, + 0.0059824115596711636, + 0.008774078451097012, + 0.0028103140648454428, + 0.0032481891103088856, + 0.004157175309956074, + 0.0014905219431966543, + 0.005077669862657785, + 0.003101855982095003, + 0.0016222493723034859, + 0.004957540892064571, + 0.01237388327717781, + 0.008355102501809597, + 0.07642979919910431, + 0.011628329753875732, + 0.011702137999236584, + 0.007326404098421335, + 0.0330423042178154, + 0.03941386193037033, + 0.003529035020619631, + 0.0024744640104472637, + 0.004179074894636869, + 0.0020730975084006786, + 0.004515145905315876, + 0.00286451680585742, + 0.00822557881474495, + 0.014396192505955696, + 0.00648539699614048, + 0.005272880662232637, + 0.016629347577691078, + 0.009526356123387814, + 0.007016124669462442, + 0.009360658936202526, + 0.05505439266562462, + 0.0431094616651535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.08901852369308472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle camera slowly tilts downward and zooms in, revealing a vast and desolate alien world. The entire scene accurately captures the distinctive, jagged, rough, polygonal textures of the PS1 era. A massive, ring-shaped spaceship, constructed from simple, low-polygons, rests quietly atop a sepia-toned volcanic landscape composed of sharp, angular polygons. Its metal hull appears a dark gray texture with pixelated noise and a simple lightmap. Surrounding the ship is a flowing river of incandescent lava. Its surface, far from a smooth liquid, consists of bright orange-red and yellow blocks of pixels, slowly creeping in a choppy, skipping pattern, providing the scene's sole primary source of light. This fiery glow casts flickering, pulsating orange shadows across the dark polygonal surface of the spaceship and the surrounding rocks. The lighting calculations are simple and direct, with no soft transitions. The deep purple sky, a simple, cloudless color gradient, further enhances this retro aesthetic of early 3D rendering, creating a powerful, desolate atmosphere that is both mysterious and imbued with a sense of digital nostalgia.", + "seed": 12822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_6754_seed_12822_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21301420275121927, + "vlm_logprob": -0.21301420275121927, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.21301420275121927, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4391086995601654 + ], + "opposite_logprob_tokens": [ + -1.0641087293624878 + ], + "logprob_reward": -0.4391086995601654, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4391086995601654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14879095554351807, + 0.19514083862304688, + 0.056836169213056564, + 0.08996864408254623, + 0.062400564551353455, + 0.10411151498556137, + 0.05390544235706329, + 0.018586350604891777, + 0.11224537342786789, + 0.15801411867141724 + ], + "credit_map": [ + 0.007777715567499399, + 0.011522100307047367, + 0.024390822276473045, + 0.015049947425723076, + 0.011822043918073177, + 0.015155544504523277, + 0.01769336126744747, + 0.0453794039785862, + 0.013336928561329842, + 0.024993961676955223, + 0.060206979513168335, + 0.01622249372303486, + 0.01226585078984499, + 0.01392519660294056, + 0.025606147944927216, + 0.028583262115716934, + 0.0031836123671382666, + 0.003454569960013032, + 0.005183206405490637, + 0.004870245698839426, + 0.007948661223053932, + 0.010590937919914722, + 0.007681312970817089, + 0.013923614285886288, + 0.0034239422529935837, + 0.0033012444619089365, + 0.01758028194308281, + 0.006610550452023745, + 0.004276746418327093, + 0.004761362448334694, + 0.03199746832251549, + 0.018017038702964783, + 0.006062197033315897, + 0.004289377946406603, + 0.006187487859278917, + 0.007334050722420216, + 0.010595581494271755, + 0.013824795372784138, + 0.006469076965004206, + 0.007637995760887861, + 0.006181683391332626, + 0.00657282629981637, + 0.009891586378216743, + 0.024787461385130882, + 0.008220613934099674, + 0.012968840077519417, + 0.00999792106449604, + 0.025490576401352882, + 0.007155723869800568, + 0.004789769649505615, + 0.011080637574195862, + 0.006281923037022352, + 0.007517789024859667, + 0.007002183701843023, + 0.004826460964977741, + 0.005250950343906879, + 0.0027685377281159163, + 0.003926762845367193, + 0.0021823637653142214, + 0.0016051921993494034, + 0.002622739179059863, + 0.0026205074973404408, + 0.0012447467306628823, + 0.0016154976328834891, + 0.007983478717505932, + 0.007672554347664118, + 0.019654318690299988, + 0.037975892424583435, + 0.007436244282871485, + 0.013466864824295044, + 0.007376152090728283, + 0.010679863393306732, + 0.011552523821592331, + 0.010884629562497139, + 0.015889085829257965, + 0.010811255313456059, + 0.018594874069094658, + 0.03869965299963951, + 0.023678366094827652, + 0.027903709560632706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.4391086995601654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600641146302223 + ], + "opposite_logprob_tokens": [ + -3.8956005573272705 + ], + "logprob_reward": -0.020600641146302223, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600641146302223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08080184459686279, + 0.03961664438247681, + 0.08009066432714462, + 0.025402842089533806, + 0.024635672569274902, + 0.07416477054357529, + 0.01274751964956522, + 0.038878053426742554, + 0.021227413788437843, + 0.602434515953064 + ], + "credit_map": [ + 0.007070386316627264, + 0.008040581829845905, + 0.010024403221905231, + 0.004504683427512646, + 0.00955469161272049, + 0.014269656501710415, + 0.009695596992969513, + 0.017641844227910042, + 0.003252298105508089, + 0.003105424577370286, + 0.011307250708341599, + 0.003084595547989011, + 0.003140179207548499, + 0.0037477423902601004, + 0.005220015533268452, + 0.006759141106158495, + 0.004479247611016035, + 0.004851558245718479, + 0.0070577929727733135, + 0.009042599238455296, + 0.013265457935631275, + 0.01581956073641777, + 0.008901726454496384, + 0.016672730445861816, + 0.0013613480841740966, + 0.0011385653633624315, + 0.008230381645262241, + 0.0017554436344653368, + 0.0013834508135914803, + 0.0020641591399908066, + 0.006131667178124189, + 0.0033378282096236944, + 0.001890636864118278, + 0.0017396120820194483, + 0.003912689629942179, + 0.00304972380399704, + 0.003767975838854909, + 0.003271603723987937, + 0.0037866830825805664, + 0.003216750454157591, + 0.0044220369309186935, + 0.0044000097550451756, + 0.009545484557747841, + 0.015449346043169498, + 0.00538940355181694, + 0.009869926609098911, + 0.008458706550300121, + 0.016629863530397415, + 0.001104768249206245, + 0.0013276938116177917, + 0.0022704799193888903, + 0.0012603029608726501, + 0.002406500745564699, + 0.0015410836786031723, + 0.0011467650765553117, + 0.0016899271868169308, + 0.006267115473747253, + 0.006477884482592344, + 0.005912101827561855, + 0.004870814271271229, + 0.004434344358742237, + 0.0037884481716901064, + 0.0038841653149574995, + 0.003243184182792902, + 0.0020774006843566895, + 0.0020305507350713015, + 0.0025325550232082605, + 0.003486941335722804, + 0.0023004808463156223, + 0.00394427077844739, + 0.0016440716572105885, + 0.0032111448235809803, + 0.04154739901423454, + 0.06055757403373718, + 0.05550745129585266, + 0.042556747794151306, + 0.07060061395168304, + 0.1390153020620346, + 0.0901668444275856, + 0.10248265415430069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.020600641146302223 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584292709827423 + ], + "opposite_logprob_tokens": [ + -3.8955843448638916 + ], + "logprob_reward": -0.020584292709827423, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584292709827423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11284943670034409, + 0.04951182380318642, + 0.13771605491638184, + 0.17535856366157532, + 0.1514226794242859, + 0.16841617226600647, + 0.030699919909238815, + 0.039630696177482605, + 0.025660648941993713, + 0.10873407125473022 + ], + "credit_map": [ + 0.008185131475329399, + 0.010774376802146435, + 0.014698057435452938, + 0.009062733501195908, + 0.013171383179724216, + 0.01968439854681492, + 0.01192323025316, + 0.02535010688006878, + 0.003300350159406662, + 0.004123213700950146, + 0.009075002744793892, + 0.006607815623283386, + 0.003874538466334343, + 0.004281030967831612, + 0.008339907974004745, + 0.009909958578646183, + 0.010311821475625038, + 0.0068794796243309975, + 0.010628823190927505, + 0.006704067811369896, + 0.008203948847949505, + 0.009498781524598598, + 0.04116315394639969, + 0.044325970113277435, + 0.005980171728879213, + 0.006699050310999155, + 0.04744650796055794, + 0.013561195693910122, + 0.01689775101840496, + 0.011023803614079952, + 0.051999229937791824, + 0.021750831976532936, + 0.010769380256533623, + 0.009560979902744293, + 0.021391931921243668, + 0.019742103293538094, + 0.022516801953315735, + 0.02447143942117691, + 0.021192770451307297, + 0.021777262911200523, + 0.012663889676332474, + 0.01179429516196251, + 0.025139082223176956, + 0.02760970965027809, + 0.013921528123319149, + 0.02677479200065136, + 0.022968316450715065, + 0.027544546872377396, + 0.0023101516999304295, + 0.0027450129855424166, + 0.005012053065001965, + 0.0034304503351449966, + 0.004930190276354551, + 0.0063822609372437, + 0.003100849222391844, + 0.002788947429507971, + 0.0040259710513055325, + 0.0037088587414473295, + 0.005603073164820671, + 0.005052496679127216, + 0.004222141578793526, + 0.005233008414506912, + 0.007600792683660984, + 0.004184351302683353, + 0.0021878122352063656, + 0.0022483188658952713, + 0.0030556160490959883, + 0.0029946339782327414, + 0.002829583827406168, + 0.005701387766748667, + 0.0028689326718449593, + 0.0037743612192571163, + 0.011781868524849415, + 0.010183027945458889, + 0.012330571189522743, + 0.008521600626409054, + 0.013429625891149044, + 0.02503756619989872, + 0.011844131164252758, + 0.01560566108673811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.020584292709827423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636976592242718 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636976592242718, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636976592242718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15442639589309692, + 0.0210158322006464, + 0.23901323974132538, + 0.16046969592571259, + 0.1323326826095581, + 0.08424346148967743, + 0.03804956376552582, + 0.022360647097229958, + 0.0625070184469223, + 0.08558144420385361 + ], + "credit_map": [ + 0.00899801217019558, + 0.011710499413311481, + 0.03333737328648567, + 0.010509644635021687, + 0.0081973597407341, + 0.011009090580046177, + 0.01692577451467514, + 0.053738635033369064, + 0.0018498137360438704, + 0.002778749680146575, + 0.0036228483077138662, + 0.0029331648256629705, + 0.001935418345965445, + 0.0021886646281927824, + 0.0023465664125978947, + 0.003360604867339134, + 0.014015729539096355, + 0.01157582737505436, + 0.017813635990023613, + 0.01607975922524929, + 0.022051967680454254, + 0.028653902933001518, + 0.06944626569747925, + 0.059376128017902374, + 0.004981345497071743, + 0.005381380207836628, + 0.05086382478475571, + 0.012686492875218391, + 0.010758841410279274, + 0.00962308794260025, + 0.05110654979944229, + 0.015068160369992256, + 0.010007276199758053, + 0.006623686756938696, + 0.02347807213664055, + 0.019719121977686882, + 0.010833565145730972, + 0.00933250691741705, + 0.029874905943870544, + 0.022463545203208923, + 0.003529986133798957, + 0.0028404558543115854, + 0.0091661112383008, + 0.02564784698188305, + 0.003466545371338725, + 0.006104636006057262, + 0.008883398957550526, + 0.024604471400380135, + 0.0037134001031517982, + 0.005964889191091061, + 0.004608896095305681, + 0.00552135007455945, + 0.0049165100790560246, + 0.005414519924670458, + 0.004146918188780546, + 0.003763077547773719, + 0.002103047212585807, + 0.0014787568943575025, + 0.004588442854583263, + 0.0034500262700021267, + 0.0020562768913805485, + 0.0024822638370096684, + 0.0027152167167514563, + 0.0034866142086684704, + 0.004634518176317215, + 0.004224675241857767, + 0.011688708327710629, + 0.011122864671051502, + 0.005690847523510456, + 0.010270635597407818, + 0.006751110311597586, + 0.00812364649027586, + 0.007691413629800081, + 0.014083274640142918, + 0.006218871101737022, + 0.008925508707761765, + 0.011304887942969799, + 0.01470143161714077, + 0.01256168819963932, + 0.010094360448420048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.008636976592242718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761404037475586 + ], + "opposite_logprob_tokens": [ + -0.8261404037475586 + ], + "logprob_reward": -0.5761404037475586, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761404037475586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05674579367041588, + 0.12070636451244354, + 0.3658154308795929, + 0.16747958958148956, + 0.135544553399086, + 0.034750018268823624, + 0.02042704075574875, + 0.009943523444235325, + 0.012721587903797626, + 0.07586609572172165 + ], + "credit_map": [ + 0.005226951092481613, + 0.00422573322430253, + 0.008263195864856243, + 0.008073992095887661, + 0.0059059117920696735, + 0.006076449528336525, + 0.007260293699800968, + 0.011713261716067791, + 0.010876054875552654, + 0.01537021528929472, + 0.03867211192846298, + 0.007758821826428175, + 0.007085410412400961, + 0.00891805812716484, + 0.014374642632901669, + 0.01765104942023754, + 0.016384784132242203, + 0.013253104873001575, + 0.02375105582177639, + 0.017406288534402847, + 0.03336593136191368, + 0.04745142534375191, + 0.10049095004796982, + 0.1137118935585022, + 0.011155891232192516, + 0.009312864392995834, + 0.017855964601039886, + 0.012295542284846306, + 0.02352803945541382, + 0.03069569543004036, + 0.035589054226875305, + 0.02704654447734356, + 0.014672267250716686, + 0.016009604558348656, + 0.011464971117675304, + 0.01913323998451233, + 0.020176751539111137, + 0.018723374232649803, + 0.013189815916121006, + 0.022174526005983353, + 0.001414053258486092, + 0.0013995168264955282, + 0.003265790641307831, + 0.009992948733270168, + 0.0016009897226467729, + 0.0034337330143898726, + 0.0032474894542247057, + 0.010395498014986515, + 0.0016706710448488593, + 0.0020974494982510805, + 0.0025104242376983166, + 0.0034279723186045885, + 0.002453756984323263, + 0.002884628251194954, + 0.0027129275258630514, + 0.002669212408363819, + 0.0014106155140325427, + 0.0012138059828430414, + 0.00130746909417212, + 0.0012460927246138453, + 0.0012553684646263719, + 0.0011220801388844848, + 0.0013484444934874773, + 0.0010396471479907632, + 0.0009649064159020782, + 0.000815815117675811, + 0.00219446187838912, + 0.004522599279880524, + 0.0005676702712662518, + 0.0010543003445491195, + 0.0006926400237716734, + 0.0019091946305707097, + 0.005103344097733498, + 0.00643674423918128, + 0.007307654712349176, + 0.00538621935993433, + 0.008875206112861633, + 0.017643677070736885, + 0.011820883490145206, + 0.013292372226715088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.5761404037475586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle camera slowly tilts downward and zooms in, revealing a vast and desolate alien world. The entire scene accurately captures the distinctive, jagged, rough, polygonal textures of the PS1 era. A massive, ring-shaped spaceship, constructed from simple, low-polygons, rests quietly atop a sepia-toned volcanic landscape composed of sharp, angular polygons. Its metal hull appears a dark gray texture with pixelated noise and a simple lightmap. Surrounding the ship is a flowing river of incandescent lava. Its surface, far from a smooth liquid, consists of bright orange-red and yellow blocks of pixels, slowly creeping in a choppy, skipping pattern, providing the scene's sole primary source of light. This fiery glow casts flickering, pulsating orange shadows across the dark polygonal surface of the spaceship and the surrounding rocks. The lighting calculations are simple and direct, with no soft transitions. The deep purple sky, a simple, cloudless color gradient, further enhances this retro aesthetic of early 3D rendering, creating a powerful, desolate atmosphere that is both mysterious and imbued with a sense of digital nostalgia.", + "seed": 12823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_6754_seed_12823_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10642593316733837, + "vlm_logprob": -0.10642593316733837, + "vlm_yes_no_margin": 2.8499999761581423, + "vlm_token_credit": -0.10642593316733837, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29533475637435913 + ], + "opposite_logprob_tokens": [ + -1.420334815979004 + ], + "logprob_reward": -0.29533475637435913, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29533475637435913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1698186695575714, + 0.22461213171482086, + 0.12260286509990692, + 0.11446397006511688, + 0.098385289311409, + 0.0865757167339325, + 0.029996655881404877, + 0.03021443448960781, + 0.06637898832559586, + 0.056951235979795456 + ], + "credit_map": [ + 0.01971156895160675, + 0.020931093022227287, + 0.024356236681342125, + 0.026354975998401642, + 0.014785077422857285, + 0.015807747840881348, + 0.02126878686249256, + 0.02660318650305271, + 0.02761424332857132, + 0.025380274280905724, + 0.03528997674584389, + 0.03399835154414177, + 0.016371533274650574, + 0.01748192310333252, + 0.026045260950922966, + 0.042430561035871506, + 0.008232115767896175, + 0.014212558977305889, + 0.01897307112812996, + 0.020375804975628853, + 0.009540721774101257, + 0.011631689965724945, + 0.013656141236424446, + 0.025980764999985695, + 0.011371859349310398, + 0.009852809831500053, + 0.018181629478931427, + 0.013791583478450775, + 0.015555176883935928, + 0.014824596233665943, + 0.013471730053424835, + 0.017414582893252373, + 0.011851329356431961, + 0.008964004926383495, + 0.014725791290402412, + 0.012384074740111828, + 0.010026494972407818, + 0.009940155781805515, + 0.013415361754596233, + 0.017078077420592308, + 0.0071552824229002, + 0.006866605952382088, + 0.013960608281195164, + 0.013956017792224884, + 0.011166571639478207, + 0.008555050939321518, + 0.010499042458832264, + 0.014416533522307873, + 0.0032788936514407396, + 0.002814633073285222, + 0.004851882811635733, + 0.005661285948008299, + 0.0028811264783143997, + 0.002696392824873328, + 0.0033159530721604824, + 0.004496488720178604, + 0.0038441421929746866, + 0.0039514158852398396, + 0.0036875938531011343, + 0.005096320062875748, + 0.0038695328403264284, + 0.0027126939967274666, + 0.003401024267077446, + 0.003651711158454418, + 0.003546173684298992, + 0.0038298750296235085, + 0.008802525699138641, + 0.01807762309908867, + 0.005166683346033096, + 0.005331027787178755, + 0.009069215506315231, + 0.012555868364870548, + 0.005260032135993242, + 0.005890366621315479, + 0.009032664820551872, + 0.008969765156507492, + 0.0069148060865700245, + 0.006269332021474838, + 0.006355862133204937, + 0.008258405141532421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.29533475637435913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331206575036049 + ], + "opposite_logprob_tokens": [ + -3.7733120918273926 + ], + "logprob_reward": -0.02331206575036049, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331206575036049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15757067501544952, + 0.20506389439105988, + 0.0681314691901207, + 0.054071806371212006, + 0.09239929169416428, + 0.031872790306806564, + 0.027338417246937752, + 0.2266165018081665, + 0.08039145171642303, + 0.056543655693531036 + ], + "credit_map": [ + 0.016690775752067566, + 0.019563166424632072, + 0.02055235393345356, + 0.02729376219213009, + 0.013700843788683414, + 0.020084885880351067, + 0.017700249329209328, + 0.021984634920954704, + 0.04456855356693268, + 0.07282492518424988, + 0.01774553954601288, + 0.014016506262123585, + 0.01369053591042757, + 0.013307934626936913, + 0.01326528750360012, + 0.015644611790776253, + 0.004882538691163063, + 0.005126808304339647, + 0.015154424123466015, + 0.01159945409744978, + 0.004641528706997633, + 0.00718333525583148, + 0.007771903648972511, + 0.011771476827561855, + 0.005688289646059275, + 0.0037586388643831015, + 0.00972743146121502, + 0.006902213208377361, + 0.004467901773750782, + 0.007431145757436752, + 0.005987582262605429, + 0.010108600370585918, + 0.006764217279851437, + 0.009463507682085037, + 0.015085753053426743, + 0.015790509060025215, + 0.010696101002395153, + 0.009017273783683777, + 0.011219173669815063, + 0.014362758956849575, + 0.002840764354914427, + 0.0032610506750643253, + 0.005360680166631937, + 0.002840725937858224, + 0.0032960926182568073, + 0.0037935669533908367, + 0.003084881231188774, + 0.007395025808364153, + 0.0028029654640704393, + 0.0040674153715372086, + 0.004121648147702217, + 0.0031980390194803476, + 0.002783677075058222, + 0.003450358984991908, + 0.002510395832359791, + 0.004403918981552124, + 0.025209352374076843, + 0.03853569179773331, + 0.03548213094472885, + 0.015759995207190514, + 0.03261547163128853, + 0.0359497107565403, + 0.015582891181111336, + 0.027481267228722572, + 0.005465282592922449, + 0.007240136619657278, + 0.010982953011989594, + 0.014086376875638962, + 0.008135437965393066, + 0.008515164256095886, + 0.00979722011834383, + 0.016168879345059395, + 0.003658886766061187, + 0.005908756982535124, + 0.0073854210786521435, + 0.00810975581407547, + 0.007988570258021355, + 0.0059495787136256695, + 0.005418295506387949, + 0.012124385684728622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.02331206575036049 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204727664589882 + ], + "opposite_logprob_tokens": [ + -2.812047243118286 + ], + "logprob_reward": -0.06204727664589882, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204727664589882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09678477048873901, + 0.3558485805988312, + 0.12210587412118912, + 0.09674058109521866, + 0.07409755885601044, + 0.03448716923594475, + 0.0314987376332283, + 0.08445209264755249, + 0.052228283137083054, + 0.0517563559114933 + ], + "credit_map": [ + 0.010369830764830112, + 0.011879472993314266, + 0.01574459858238697, + 0.014417741447687149, + 0.007084211800247431, + 0.01055001188069582, + 0.014888247475028038, + 0.011850651353597641, + 0.057812564074993134, + 0.07784376293420792, + 0.04891550913453102, + 0.043125126510858536, + 0.02142998017370701, + 0.022277969866991043, + 0.037710852921009064, + 0.0467328242957592, + 0.009130380116403103, + 0.011407301761209965, + 0.01918870210647583, + 0.021690351888537407, + 0.013719703070819378, + 0.015245204791426659, + 0.014941261149942875, + 0.01678297109901905, + 0.010617632418870926, + 0.007374701090157032, + 0.015086480416357517, + 0.0084104398265481, + 0.014565516263246536, + 0.01709705963730812, + 0.010763944126665592, + 0.01282480824738741, + 0.006648301612585783, + 0.006611017510294914, + 0.012415112927556038, + 0.011567818000912666, + 0.007526210974901915, + 0.007836699485778809, + 0.008271757513284683, + 0.013220642693340778, + 0.003320130053907633, + 0.00227134651504457, + 0.005223460961133242, + 0.005582713987678289, + 0.0029808906838297844, + 0.0037455782294273376, + 0.004714925307780504, + 0.006648126523941755, + 0.002657025819644332, + 0.002542020520195365, + 0.004424923565238714, + 0.007126559969037771, + 0.004836666863411665, + 0.0028379589784890413, + 0.0032905994448810816, + 0.0037829831708222628, + 0.008495190180838108, + 0.011779679916799068, + 0.013751273974776268, + 0.008988398127257824, + 0.010553374886512756, + 0.012070666998624802, + 0.006472111213952303, + 0.012341396883130074, + 0.004364479333162308, + 0.004035771358758211, + 0.007233913987874985, + 0.010517366230487823, + 0.004045707173645496, + 0.005659772548824549, + 0.006809835322201252, + 0.009561438113451004, + 0.003283888567239046, + 0.003605637699365616, + 0.00794314593076706, + 0.011509832926094532, + 0.0037143859080970287, + 0.0039934078231453896, + 0.0066445367410779, + 0.01106152031570673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.06204727664589882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640168234705925 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640168234705925, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640168234705925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08767569810152054, + 0.27386483550071716, + 0.11228540539741516, + 0.05457981303334236, + 0.09791779518127441, + 0.03705733269453049, + 0.035208892077207565, + 0.12578584253787994, + 0.0649864673614502, + 0.11063791066408157 + ], + "credit_map": [ + 0.014981881715357304, + 0.01123015210032463, + 0.012231186963617802, + 0.010217376053333282, + 0.007704110816121101, + 0.00961692538112402, + 0.010297848843038082, + 0.011396230198442936, + 0.038776036351919174, + 0.05303372070193291, + 0.035434380173683167, + 0.042187515646219254, + 0.019090596586465836, + 0.024084370583295822, + 0.03418803587555885, + 0.027070187032222748, + 0.007625304628163576, + 0.009905061684548855, + 0.022486673668026924, + 0.020968185737729073, + 0.008700442500412464, + 0.012907176278531551, + 0.012661991640925407, + 0.01703057624399662, + 0.006827279459685087, + 0.004054291173815727, + 0.010169479064643383, + 0.007599188480526209, + 0.005707756616175175, + 0.0072930846363306046, + 0.005593846086412668, + 0.007334891241043806, + 0.008202407509088516, + 0.009891027584671974, + 0.015136178582906723, + 0.016807282343506813, + 0.01197939831763506, + 0.01063485536724329, + 0.012386777438223362, + 0.01287987269461155, + 0.003221676219254732, + 0.0030534686520695686, + 0.005920132156461477, + 0.005310356616973877, + 0.004233731888234615, + 0.004915633238852024, + 0.004038351587951183, + 0.006363984197378159, + 0.003306959057226777, + 0.00410702358931303, + 0.0038533895276486874, + 0.0055612181313335896, + 0.00561736011877656, + 0.004120140802115202, + 0.0035922976676374674, + 0.005050505045801401, + 0.02777712419629097, + 0.00977808702737093, + 0.015666630119085312, + 0.016696421429514885, + 0.011914346367120743, + 0.012331416830420494, + 0.013658932410180569, + 0.01796288788318634, + 0.006876075174659491, + 0.007996708154678345, + 0.008658435195684433, + 0.00962383858859539, + 0.007446887902915478, + 0.007327863946557045, + 0.008755653165280819, + 0.008301005698740482, + 0.011703506112098694, + 0.009860297664999962, + 0.012963310815393925, + 0.01408388651907444, + 0.015433547087013721, + 0.01070372387766838, + 0.012814667075872421, + 0.023074980825185776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.008640168234705925 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427953988313675 + ], + "opposite_logprob_tokens": [ + -2.0177953243255615 + ], + "logprob_reward": -0.1427953988313675, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1427953988313675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.199018657207489, + 0.22888706624507904, + 0.15984348952770233, + 0.17777389287948608, + 0.06898993998765945, + 0.04121948406100273, + 0.028313258662819862, + 0.03151221200823784, + 0.04803234711289406, + 0.01640963740646839 + ], + "credit_map": [ + 0.03823864832520485, + 0.02718503400683403, + 0.02446354366838932, + 0.02124754525721073, + 0.012708340771496296, + 0.018477993085980415, + 0.023125452920794487, + 0.033572107553482056, + 0.034103576093912125, + 0.05492236465215683, + 0.03153945133090019, + 0.025976281613111496, + 0.015084458515048027, + 0.01740618795156479, + 0.02017066441476345, + 0.029684066772460938, + 0.0164687167853117, + 0.022874297574162483, + 0.02993893437087536, + 0.021412350237369537, + 0.01156333927065134, + 0.015206165611743927, + 0.01897875778377056, + 0.023400940001010895, + 0.018998736515641212, + 0.02086026407778263, + 0.0300630871206522, + 0.024954792112112045, + 0.019325848668813705, + 0.017967605963349342, + 0.022351661697030067, + 0.023251891136169434, + 0.006356113590300083, + 0.006853053811937571, + 0.010687438771128654, + 0.01448044553399086, + 0.006235760636627674, + 0.00551249785348773, + 0.008381707593798637, + 0.010482924059033394, + 0.0034733840730041265, + 0.0037935299333184958, + 0.005815966986119747, + 0.004487987607717514, + 0.005873711314052343, + 0.007189829833805561, + 0.004003976471722126, + 0.00658109737560153, + 0.00275583378970623, + 0.002436235314235091, + 0.004047642461955547, + 0.0063444627448916435, + 0.0026149398181587458, + 0.002914303680881858, + 0.0028167772106826305, + 0.004383062943816185, + 0.0025999941863119602, + 0.003132243873551488, + 0.00462730135768652, + 0.005481643136590719, + 0.004594387952238321, + 0.002822698326781392, + 0.004033202305436134, + 0.00422074506059289, + 0.0036134927067905664, + 0.004724194761365652, + 0.006460640579462051, + 0.009830128401517868, + 0.004796285647898912, + 0.004708676598966122, + 0.0058944993652403355, + 0.008004429750144482, + 0.0018704967806115746, + 0.001740584964863956, + 0.0022860874887555838, + 0.0018268732819706202, + 0.002372871385887265, + 0.001761272782459855, + 0.0018335183849558234, + 0.0027179336175322533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.1427953988313675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle camera slowly tilts downward and zooms in, revealing a vast and desolate alien world. The entire scene accurately captures the distinctive, jagged, rough, polygonal textures of the PS1 era. A massive, ring-shaped spaceship, constructed from simple, low-polygons, rests quietly atop a sepia-toned volcanic landscape composed of sharp, angular polygons. Its metal hull appears a dark gray texture with pixelated noise and a simple lightmap. Surrounding the ship is a flowing river of incandescent lava. Its surface, far from a smooth liquid, consists of bright orange-red and yellow blocks of pixels, slowly creeping in a choppy, skipping pattern, providing the scene's sole primary source of light. This fiery glow casts flickering, pulsating orange shadows across the dark polygonal surface of the spaceship and the surrounding rocks. The lighting calculations are simple and direct, with no soft transitions. The deep purple sky, a simple, cloudless color gradient, further enhances this retro aesthetic of early 3D rendering, creating a powerful, desolate atmosphere that is both mysterious and imbued with a sense of digital nostalgia.", + "seed": 12824, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_1/index_6754_seed_12824_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03265453027561307, + "vlm_logprob": -0.03265453027561307, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03265453027561307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029232457280158997 + ], + "opposite_logprob_tokens": [ + -3.7792325019836426 + ], + "logprob_reward": -0.029232457280158997, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.029232457280158997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04070620611310005, + 0.03533404320478439, + 0.034987449645996094, + 0.07759833335876465, + 0.5365955233573914, + 0.17560753226280212, + 0.03569072484970093, + 0.0240146704018116, + 0.01960361748933792, + 0.019861958920955658 + ], + "credit_map": [ + 0.0036246604286134243, + 0.00558424973860383, + 0.005720452405512333, + 0.004542202223092318, + 0.0029614430386573076, + 0.0058644236996769905, + 0.00551227480173111, + 0.0068965028040111065, + 0.006401520222425461, + 0.00415606377646327, + 0.003850503358989954, + 0.0036694896407425404, + 0.00454016262665391, + 0.006294370628893375, + 0.003188671078532934, + 0.0032332614064216614, + 0.0031098767649382353, + 0.0035400891210883856, + 0.004559502936899662, + 0.0056373667903244495, + 0.005174164194613695, + 0.0032077617943286896, + 0.0035562238190323114, + 0.006202463526278734, + 0.0065802619792521, + 0.00580274872481823, + 0.009348786436021328, + 0.011317191645503044, + 0.008450688794255257, + 0.005251269321888685, + 0.020792359486222267, + 0.010055026970803738, + 0.03863949328660965, + 0.07534816861152649, + 0.02704286389052868, + 0.09065977483987808, + 0.08006644248962402, + 0.04382668435573578, + 0.12152861803770065, + 0.05948348715901375, + 0.019170979037880898, + 0.02277575619518757, + 0.010670162737369537, + 0.02201276645064354, + 0.02154509909451008, + 0.02717474289238453, + 0.02062985487282276, + 0.031628161668777466, + 0.0045133065432310104, + 0.004804541822522879, + 0.003160957945510745, + 0.0026655897963792086, + 0.004444332327693701, + 0.006674675736576319, + 0.004173453897237778, + 0.0052538649179041386, + 0.002419928787276149, + 0.0056401449255645275, + 0.0008851421298459172, + 0.0013883570209145546, + 0.0036716272588819265, + 0.004727461375296116, + 0.0019353338284417987, + 0.0033466739114373922, + 0.003670665668323636, + 0.0030375344213098288, + 0.0012608973775058985, + 0.0014078716048970819, + 0.0027655092999339104, + 0.0032871027942746878, + 0.0013083552476018667, + 0.0028656814247369766, + 0.002419259399175644, + 0.0025407064240425825, + 0.0017763297073543072, + 0.00125320500228554, + 0.004152079578489065, + 0.003982509020715952, + 0.0013588382862508297, + 0.002379031851887703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.029232457280158997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651040494441986 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651040494441986, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651040494441986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1147330105304718, + 0.1315062791109085, + 0.0833822712302208, + 0.09360989183187485, + 0.15386582911014557, + 0.07587285339832306, + 0.08263187110424042, + 0.13264289498329163, + 0.04987947270274162, + 0.08187562972307205 + ], + "credit_map": [ + 0.011504781432449818, + 0.01667756773531437, + 0.012936019338667393, + 0.017709584906697273, + 0.013927971012890339, + 0.014505643397569656, + 0.015555090270936489, + 0.01191635336726904, + 0.023295322433114052, + 0.011026821099221706, + 0.015922807157039642, + 0.014598029665648937, + 0.011460920795798302, + 0.029383782297372818, + 0.012008371762931347, + 0.013810219243168831, + 0.00913609191775322, + 0.014200529083609581, + 0.008664645254611969, + 0.010378344915807247, + 0.012923895381391048, + 0.007283834740519524, + 0.01066360343247652, + 0.010131331160664558, + 0.007187221199274063, + 0.01108302641659975, + 0.008800274692475796, + 0.018391909077763557, + 0.013157282024621964, + 0.008154908195137978, + 0.017025833949446678, + 0.009809439070522785, + 0.011579927988350391, + 0.015430284664034843, + 0.012722638435661793, + 0.029417747631669044, + 0.011413009837269783, + 0.011088288389146328, + 0.02168157324194908, + 0.04053236544132233, + 0.007135061081498861, + 0.008618265390396118, + 0.005074311047792435, + 0.01014223787933588, + 0.007346350234001875, + 0.01001909002661705, + 0.010606569238007069, + 0.016930967569351196, + 0.007129746023565531, + 0.009776563383638859, + 0.0037901911418884993, + 0.004704260267317295, + 0.008459380827844143, + 0.017491914331912994, + 0.011352150700986385, + 0.01992766000330448, + 0.013568207621574402, + 0.025282245129346848, + 0.006117642857134342, + 0.007350322790443897, + 0.01836979389190674, + 0.0317506305873394, + 0.010065835900604725, + 0.020138215273618698, + 0.006695096381008625, + 0.008273274637758732, + 0.002287548501044512, + 0.0021411986090242863, + 0.0077273109927773476, + 0.011930587701499462, + 0.003407080424949527, + 0.007417372893542051, + 0.012295665219426155, + 0.01314715575426817, + 0.005134535487741232, + 0.0032597023528069258, + 0.010801546275615692, + 0.013727753423154354, + 0.004973208997398615, + 0.018536055460572243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.008651040494441986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865250736474991 + ], + "opposite_logprob_tokens": [ + -3.048652410507202 + ], + "logprob_reward": -0.04865250736474991, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865250736474991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1099228709936142, + 0.0769055113196373, + 0.1288307011127472, + 0.14729659259319305, + 0.3237139880657196, + 0.06993299722671509, + 0.06302410364151001, + 0.03465281054377556, + 0.011319611221551895, + 0.03440083563327789 + ], + "credit_map": [ + 0.01144535280764103, + 0.01320718415081501, + 0.017491647973656654, + 0.013600916601717472, + 0.011446202173829079, + 0.01248074509203434, + 0.011619207449257374, + 0.01863161101937294, + 0.02200459875166416, + 0.00663141580298543, + 0.006443760823458433, + 0.003277862211689353, + 0.003748370334506035, + 0.026183873414993286, + 0.003521314822137356, + 0.005094317253679037, + 0.012893927283585072, + 0.014692064374685287, + 0.014240173622965813, + 0.02128850482404232, + 0.01711404323577881, + 0.014348345808684826, + 0.01796620897948742, + 0.016287418082356453, + 0.0071307276375591755, + 0.008263152092695236, + 0.013544238172471523, + 0.031067319214344025, + 0.007864927873015404, + 0.0070901173166930676, + 0.040383271872997284, + 0.03195281699299812, + 0.027149736881256104, + 0.05621959641575813, + 0.018929455429315567, + 0.05221203342080116, + 0.04788754880428314, + 0.02472127228975296, + 0.05701841413974762, + 0.03957592323422432, + 0.009801780804991722, + 0.0088462820276618, + 0.0061473967507481575, + 0.008202890865504742, + 0.007047904189676046, + 0.007672647479921579, + 0.00950581580400467, + 0.012708282098174095, + 0.006486645899713039, + 0.01043215673416853, + 0.004284185357391834, + 0.005309535190463066, + 0.008452055044472218, + 0.011786457151174545, + 0.00958634726703167, + 0.006686713546514511, + 0.0029740449972450733, + 0.005805011838674545, + 0.0015933888498693705, + 0.00202655466273427, + 0.005363760981708765, + 0.008574428968131542, + 0.0029694759286940098, + 0.005346142221242189, + 0.0009696561610326171, + 0.00195560185238719, + 0.0006971765542402864, + 0.0007881885394454002, + 0.0021921470761299133, + 0.0024199443869292736, + 0.0009832850191742182, + 0.0013136105844751, + 0.005132712889462709, + 0.005440202541649342, + 0.0018510398222133517, + 0.0013204161077737808, + 0.004722302779555321, + 0.0061863865703344345, + 0.0020020490046590567, + 0.007745726965367794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.04865250736474991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08359306305646896, + 0.11147282272577286, + 0.04945039749145508, + 0.04151473566889763, + 0.45088252425193787, + 0.08853454142808914, + 0.06049596890807152, + 0.0330105759203434, + 0.04206397756934166, + 0.03898130729794502 + ], + "credit_map": [ + 0.009913728572428226, + 0.009617812931537628, + 0.013294684700667858, + 0.010405100882053375, + 0.00972446333616972, + 0.011142468079924583, + 0.008380020968616009, + 0.01111479103565216, + 0.03507884964346886, + 0.008175202645361423, + 0.005294227506965399, + 0.004871752578765154, + 0.005153511185199022, + 0.04343065619468689, + 0.004579700529575348, + 0.0048889284953475, + 0.005390445236116648, + 0.006854564417153597, + 0.007483262103050947, + 0.0054978844709694386, + 0.007676903624087572, + 0.0046471175737679005, + 0.005134221166372299, + 0.0067660026252269745, + 0.004298655316233635, + 0.003636154578998685, + 0.00581378024071455, + 0.006198961287736893, + 0.004902371671050787, + 0.0034966710954904556, + 0.00555810471996665, + 0.00761004164814949, + 0.0352625697851181, + 0.07738295942544937, + 0.02149219810962677, + 0.06857860833406448, + 0.07996664196252823, + 0.03568565100431442, + 0.09445741772651672, + 0.03805652633309364, + 0.008654325269162655, + 0.010573308914899826, + 0.006138076540082693, + 0.011509783565998077, + 0.009204504080116749, + 0.011570819653570652, + 0.01609061285853386, + 0.014793121255934238, + 0.005896304734051228, + 0.006967584602534771, + 0.003342111362144351, + 0.0042929514311254025, + 0.006249658763408661, + 0.011666159145534039, + 0.008500861935317516, + 0.013580341823399067, + 0.0034066459629684687, + 0.0057807620614767075, + 0.0019765745382755995, + 0.0022857082076370716, + 0.00461463863030076, + 0.005223474930971861, + 0.002545667812228203, + 0.00717710517346859, + 0.005832912866026163, + 0.006623025983572006, + 0.002475400920957327, + 0.0027072725351899862, + 0.006645787972956896, + 0.008349328301846981, + 0.0033699257764965296, + 0.00606032507494092, + 0.005815689451992512, + 0.005037314724177122, + 0.002341760555282235, + 0.001464378321543336, + 0.006446491926908493, + 0.0064034368842840195, + 0.0024624706711620092, + 0.009009769186377525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.006735833827406168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a high-angle shot that slowly tilts downward and zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive ring-shaped spaceship constructed from simple low-polygons visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spaceship resting on a sepia-toned volcanic landscape composed of sharp, angular polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flowing river of incandescent lava made of bright orange-red and yellow pixel blocks surrounding the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a deep purple colorless gradient with no clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000081241130829 + ], + "opposite_logprob_tokens": [ + -2.6950008869171143 + ], + "logprob_reward": -0.07000081241130829, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000081241130829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17070826888084412, + 0.2353358268737793, + 0.13182257115840912, + 0.13264411687850952, + 0.18562950193881989, + 0.04728778824210167, + 0.035638466477394104, + 0.026980731636285782, + 0.017472561448812485, + 0.01648017019033432 + ], + "credit_map": [ + 0.032722778618335724, + 0.02405523881316185, + 0.019330020993947983, + 0.018477363511919975, + 0.012629855424165726, + 0.029360059648752213, + 0.012513047084212303, + 0.021619899198412895, + 0.07470017671585083, + 0.015700003132224083, + 0.012670855969190598, + 0.011680208146572113, + 0.010029775090515614, + 0.09082238376140594, + 0.008702460676431656, + 0.011029952205717564, + 0.017910176888108253, + 0.011478288099169731, + 0.013937195762991905, + 0.015179253183305264, + 0.010175184346735477, + 0.022279320284724236, + 0.018568091094493866, + 0.02229505404829979, + 0.014764884486794472, + 0.00894812773913145, + 0.015683261677622795, + 0.024293307214975357, + 0.012415194883942604, + 0.01112260203808546, + 0.02261975221335888, + 0.022796988487243652, + 0.011841543018817902, + 0.026429202407598495, + 0.009373370558023453, + 0.029085829854011536, + 0.03336627781391144, + 0.014552620239555836, + 0.04316400736570358, + 0.01781664788722992, + 0.004655822645872831, + 0.004202422220259905, + 0.004067559726536274, + 0.0058572543784976006, + 0.003851710818707943, + 0.007675322238355875, + 0.006339679472148418, + 0.010638009756803513, + 0.004156726412475109, + 0.005013371352106333, + 0.002359365578740835, + 0.0026293140836060047, + 0.004765033256262541, + 0.006100199650973082, + 0.004969665315002203, + 0.005644785240292549, + 0.0023121677804738283, + 0.004637111909687519, + 0.0014437889913097024, + 0.0014089101459831, + 0.004354089964181185, + 0.007103105075657368, + 0.0020820177160203457, + 0.003639538772404194, + 0.001655539614148438, + 0.0026382138021290302, + 0.0010949061252176762, + 0.0009966216748580337, + 0.004316119942814112, + 0.003170556388795376, + 0.0012045885669067502, + 0.00239601475186646, + 0.002777167595922947, + 0.0023935833014547825, + 0.0013624057173728943, + 0.0009369160397909582, + 0.002421660115942359, + 0.002587202936410904, + 0.0010348166106268764, + 0.0029664160683751106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.07000081241130829 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_tensor([1714, 1714])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_tensor([1714, 1714])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0437a8f68b64a78e20847590ab1d17920f02cc4c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_tensor([1714, 1714])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot smoothly follows a cheerful, bright yellow snowplow, rendered in the charming style of Pixar animation, as it steadily advances. The plow's rounded edges and smooth finish gleam in the soft, unfocused daylight. Its two large, round headlights, designed like smiling eyes, emit a warm, friendly glow, adding to its anthropomorphic character. Its sturdy black caterpillar trails rhythmically churn, leaving deep marks as they propel the vehicle through a thick, untouched blanket of fresh, powdery snow. The massive front blades meticulously carve a clean path, causing the snow to curl and roll, forming two perfect, fluffy banks on either side. A very shallow depth of field is employed, beautifully blurring the background of tall, snow-covered pine trees into a soft, ethereal winter mist composed of muted greens and whites. The overall atmosphere is one of whimsical tranquility and joy, dominated by the vibrant yellow of the plow and the bright white of the landscape.", + "seed": 240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_1714_seed_240_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07346575483679771, + "vlm_logprob": -0.07346575483679771, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.07346575483679771, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0064351484179496765 + ], + "opposite_logprob_tokens": [ + -6.756434917449951 + ], + "logprob_reward": -0.0064351484179496765, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0064351484179496765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13799884915351868, + 0.06514647603034973, + 0.10014820098876953, + 0.0786377415060997, + 0.02875005453824997, + 0.31270623207092285, + 0.13705341517925262, + 0.0387420691549778, + 0.04937518388032913, + 0.05144186317920685 + ], + "credit_map": [ + 0.02393866330385208, + 0.014900828711688519, + 0.023672327399253845, + 0.019906068220734596, + 0.018167611211538315, + 0.012320268899202347, + 0.013433048501610756, + 0.011660023592412472, + 0.01198398694396019, + 0.006723000202327967, + 0.010238446295261383, + 0.007973209954798222, + 0.009688425809144974, + 0.005834287032485008, + 0.006784957367926836, + 0.005920156836509705, + 0.013165139593183994, + 0.010675533674657345, + 0.01940029487013817, + 0.017682228237390518, + 0.013484020717442036, + 0.008899707347154617, + 0.00990980863571167, + 0.0069314585998654366, + 0.006594502367079258, + 0.011462878435850143, + 0.01108953170478344, + 0.009961443021893501, + 0.016136446967720985, + 0.00793779268860817, + 0.008796168491244316, + 0.006658972706645727, + 0.003696081228554249, + 0.0029139581602066755, + 0.004295654594898224, + 0.005719499196857214, + 0.004574130289256573, + 0.0022660037502646446, + 0.003129165153950453, + 0.0021555579733103514, + 0.023614637553691864, + 0.01282848883420229, + 0.014025546610355377, + 0.13629227876663208, + 0.009075836278498173, + 0.008897247724235058, + 0.10006094723939896, + 0.00791118759661913, + 0.01618894934654236, + 0.00964664388448, + 0.025983508676290512, + 0.019240403547883034, + 0.01726687327027321, + 0.01462110411375761, + 0.021542930975556374, + 0.012562975287437439, + 0.004678222816437483, + 0.003262074664235115, + 0.0046731578186154366, + 0.00622783787548542, + 0.007236094679683447, + 0.004083212930709124, + 0.005488473456352949, + 0.003092993749305606, + 0.005988720338791609, + 0.006191014777868986, + 0.008568068034946918, + 0.004902724176645279, + 0.005697045475244522, + 0.004811008460819721, + 0.005503461696207523, + 0.007713135797530413, + 0.007236884906888008, + 0.006126588210463524, + 0.010229292325675488, + 0.0056834653951227665, + 0.006875072605907917, + 0.006261811591684818, + 0.005533577874302864, + 0.0034951623529195786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0064351484179496765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086550429463387 + ], + "opposite_logprob_tokens": [ + -4.141086578369141 + ], + "logprob_reward": -0.016086550429463387, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086550429463387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17831610143184662, + 0.058964554220438004, + 0.16553495824337006, + 0.14429469406604767, + 0.02918405830860138, + 0.06874676048755646, + 0.16899457573890686, + 0.0622723214328289, + 0.02675659768283367, + 0.09693539887666702 + ], + "credit_map": [ + 0.02356242947280407, + 0.011929177679121494, + 0.026247229427099228, + 0.057785894721746445, + 0.02021341770887375, + 0.01151698175817728, + 0.018363211303949356, + 0.008697754703462124, + 0.009763196110725403, + 0.005006982944905758, + 0.01479932852089405, + 0.008792048320174217, + 0.006527074612677097, + 0.004216605797410011, + 0.005320184398442507, + 0.004539132118225098, + 0.019809793680906296, + 0.015161646530032158, + 0.04249688982963562, + 0.033995818346738815, + 0.01623939536511898, + 0.01271623931825161, + 0.012808701023459435, + 0.012306476943194866, + 0.021385828033089638, + 0.016287818551063538, + 0.019759800285100937, + 0.029028194025158882, + 0.018265122547745705, + 0.013055131770670414, + 0.016363341361284256, + 0.010149456560611725, + 0.004016675520688295, + 0.0029147686436772346, + 0.004992211703211069, + 0.004296252969652414, + 0.003782199462875724, + 0.002992415800690651, + 0.0035732020623981953, + 0.002616333309561014, + 0.00590251199901104, + 0.0031164430547505617, + 0.005136409774422646, + 0.02710319124162197, + 0.0030732599552720785, + 0.0023130858317017555, + 0.020022081211209297, + 0.0020797783508896828, + 0.017239641398191452, + 0.017932960763573647, + 0.017339883372187614, + 0.021049266681075096, + 0.04099017754197121, + 0.016853956505656242, + 0.019766639918088913, + 0.017822057008743286, + 0.01028719637542963, + 0.004556258209049702, + 0.009853632189333439, + 0.012640178203582764, + 0.005010347813367844, + 0.006353361066430807, + 0.008041766472160816, + 0.005529582500457764, + 0.004110024310648441, + 0.002589661628007889, + 0.004067886155098677, + 0.002901414642110467, + 0.0031533404253423214, + 0.0028583863750100136, + 0.003795416560024023, + 0.003280465956777334, + 0.019298624247312546, + 0.006811262108385563, + 0.009676850400865078, + 0.0299950260668993, + 0.007170228753238916, + 0.007546134293079376, + 0.010895301587879658, + 0.005541970022022724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.016086550429463387 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813184857368469 + ], + "opposite_logprob_tokens": [ + -1.4063184261322021 + ], + "logprob_reward": -0.2813184857368469, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813184857368469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48031532764434814, + 0.07490717619657516, + 0.06332380324602127, + 0.05305827036499977, + 0.030216291546821594, + 0.027015311643481255, + 0.15043741464614868, + 0.05129852890968323, + 0.01955566555261612, + 0.04987221956253052 + ], + "credit_map": [ + 0.07659618556499481, + 0.04067220166325569, + 0.07458430528640747, + 0.12149255722761154, + 0.04308859631419182, + 0.03473721817135811, + 0.05499722436070442, + 0.034146975725889206, + 0.008667580783367157, + 0.006166818551719189, + 0.014558661729097366, + 0.010997351258993149, + 0.008329424075782299, + 0.007513906806707382, + 0.010813954286277294, + 0.007859484292566776, + 0.015371096320450306, + 0.006331398151814938, + 0.009515583515167236, + 0.0066680144518613815, + 0.008797656744718552, + 0.006315660197287798, + 0.005207519978284836, + 0.005116869695484638, + 0.008739136159420013, + 0.005234821233898401, + 0.006634252145886421, + 0.01470782421529293, + 0.004624665714800358, + 0.0028067731764167547, + 0.007000392768532038, + 0.0033104035537689924, + 0.004355365876108408, + 0.0029170187190175056, + 0.0050318860448896885, + 0.00546790286898613, + 0.0024885418824851513, + 0.0031002117320895195, + 0.003569535445421934, + 0.003285827348008752, + 0.003534000599756837, + 0.001815434661693871, + 0.0027801217511296272, + 0.006952950730919838, + 0.0023696899879723787, + 0.0018512408714741468, + 0.006180694326758385, + 0.0015311799943447113, + 0.01659836433827877, + 0.00941191054880619, + 0.01490640640258789, + 0.03612501919269562, + 0.011047914624214172, + 0.00808220449835062, + 0.047787752002477646, + 0.006477836053818464, + 0.006764817517250776, + 0.004155609756708145, + 0.008934464305639267, + 0.00916467234492302, + 0.006516247056424618, + 0.0054646264761686325, + 0.005671457853168249, + 0.004626632668077946, + 0.0022354209795594215, + 0.0025017382577061653, + 0.0023093149065971375, + 0.002117812167853117, + 0.003246128559112549, + 0.002406977117061615, + 0.002301275497302413, + 0.002436998300254345, + 0.008737338706851006, + 0.0037456813734024763, + 0.006845534313470125, + 0.013737925328314304, + 0.00395444268360734, + 0.0037166534457355738, + 0.006686170119792223, + 0.002448469866067171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.2813184857368469 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799478128552437 + ], + "opposite_logprob_tokens": [ + -3.529799461364746 + ], + "logprob_reward": -0.029799478128552437, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799478128552437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1435631960630417, + 0.05524394288659096, + 0.05857807397842407, + 0.04734377935528755, + 0.019622180610895157, + 0.24870993196964264, + 0.09971761703491211, + 0.055754404515028, + 0.014279638417065144, + 0.2571871876716614 + ], + "credit_map": [ + 0.02011849731206894, + 0.01515984907746315, + 0.030195657163858414, + 0.019935915246605873, + 0.014750377275049686, + 0.013555869460105896, + 0.01566898636519909, + 0.014178045094013214, + 0.008461788296699524, + 0.00445632403716445, + 0.01320397574454546, + 0.008332907222211361, + 0.00503354799002409, + 0.004959092475473881, + 0.0057859173975884914, + 0.005010388791561127, + 0.011091777123510838, + 0.00525510311126709, + 0.013559076935052872, + 0.010519785806536674, + 0.0036861293483525515, + 0.005313255824148655, + 0.005257583688944578, + 0.0038953647017478943, + 0.007565416861325502, + 0.005023466423153877, + 0.0053392923437058926, + 0.012893731705844402, + 0.0058376784436404705, + 0.0032443185336887836, + 0.005094548687338829, + 0.002345328452065587, + 0.0029906502459198236, + 0.0017842247616499662, + 0.0023113954812288284, + 0.004010898061096668, + 0.0021926548797637224, + 0.0021992595866322517, + 0.0021881042048335075, + 0.0019449940882623196, + 0.019476156681776047, + 0.01019764132797718, + 0.011758618988096714, + 0.10635191947221756, + 0.007643629796802998, + 0.008107718080282211, + 0.07799599319696426, + 0.0071782311424613, + 0.008198550902307034, + 0.005748806055635214, + 0.013418560847640038, + 0.022562317550182343, + 0.010365746915340424, + 0.004239323548972607, + 0.030303901061415672, + 0.004880407825112343, + 0.009694927372038364, + 0.0046617416664958, + 0.009190929122269154, + 0.010516704991459846, + 0.005357105750590563, + 0.005304073449224234, + 0.006633018609136343, + 0.004395907279103994, + 0.0022855012211948633, + 0.001779251266270876, + 0.002124055288732052, + 0.0016073433216661215, + 0.002229636535048485, + 0.0017838417552411556, + 0.0013587878784164786, + 0.0011112220818176866, + 0.05231551453471184, + 0.020429197698831558, + 0.026016561314463615, + 0.0786455050110817, + 0.018196675926446915, + 0.01872928999364376, + 0.02966470643877983, + 0.013189738616347313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.029799478128552437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368911147117615 + ], + "opposite_logprob_tokens": [ + -3.408689022064209 + ], + "logprob_reward": -0.03368911147117615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368911147117615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1239931583404541, + 0.035312220454216, + 0.2858566641807556, + 0.052174802869558334, + 0.02668839320540428, + 0.31096652150154114, + 0.0998707190155983, + 0.021546626463532448, + 0.016915423795580864, + 0.02667541801929474 + ], + "credit_map": [ + 0.024014966562390327, + 0.013163787312805653, + 0.015703029930591583, + 0.021012473851442337, + 0.01639694906771183, + 0.012314533814787865, + 0.011614006944000721, + 0.009773429483175278, + 0.005588049069046974, + 0.002844782080501318, + 0.006298711523413658, + 0.003940509632229805, + 0.005323129240423441, + 0.003529908834025264, + 0.004246327560395002, + 0.003540807170793414, + 0.04828203469514847, + 0.024187354370951653, + 0.08103292435407639, + 0.05662580952048302, + 0.01600131019949913, + 0.02316666580736637, + 0.018796700984239578, + 0.0177638977766037, + 0.007866519503295422, + 0.004820230882614851, + 0.0070541840977966785, + 0.012355051003396511, + 0.005876649636775255, + 0.004174013156443834, + 0.006306166294962168, + 0.003721992950886488, + 0.004978206008672714, + 0.0022434801794588566, + 0.0040403190068900585, + 0.006499676965177059, + 0.0019930959679186344, + 0.0015214387094601989, + 0.00351422606036067, + 0.0018979563610628247, + 0.024225713685154915, + 0.01256657112389803, + 0.013937658630311489, + 0.13617099821567535, + 0.008237110450863838, + 0.00838936772197485, + 0.09988034516572952, + 0.007558786775916815, + 0.013110952451825142, + 0.00966215692460537, + 0.013182601891458035, + 0.015127538703382015, + 0.010505164042115211, + 0.010933613404631615, + 0.014731856063008308, + 0.012616848573088646, + 0.004435080103576183, + 0.0022483114153146744, + 0.0033938142005354166, + 0.0030199773609638214, + 0.002996639348566532, + 0.0018109694356098771, + 0.0020237991120666265, + 0.0016180373495444655, + 0.002666611224412918, + 0.0017286061774939299, + 0.0023281730245798826, + 0.0018627488752827048, + 0.0024775860365480185, + 0.0024014022201299667, + 0.001641715643927455, + 0.001808583503589034, + 0.005447497125715017, + 0.0021861896384507418, + 0.004268405959010124, + 0.0059870448894798756, + 0.0018300903029739857, + 0.002453739522024989, + 0.0029357916209846735, + 0.0015666636172682047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.03368911147117615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot smoothly follows a cheerful, bright yellow snowplow, rendered in the charming style of Pixar animation, as it steadily advances. The plow's rounded edges and smooth finish gleam in the soft, unfocused daylight. Its two large, round headlights, designed like smiling eyes, emit a warm, friendly glow, adding to its anthropomorphic character. Its sturdy black caterpillar trails rhythmically churn, leaving deep marks as they propel the vehicle through a thick, untouched blanket of fresh, powdery snow. The massive front blades meticulously carve a clean path, causing the snow to curl and roll, forming two perfect, fluffy banks on either side. A very shallow depth of field is employed, beautifully blurring the background of tall, snow-covered pine trees into a soft, ethereal winter mist composed of muted greens and whites. The overall atmosphere is one of whimsical tranquility and joy, dominated by the vibrant yellow of the plow and the bright white of the landscape.", + "seed": 241, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_1714_seed_241_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08072309903800487, + "vlm_logprob": -0.08072309903800487, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.08072309903800487, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011710118502378464 + ], + "opposite_logprob_tokens": [ + -7.386710166931152 + ], + "logprob_reward": -0.011710118502378464, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.011710118502378464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060551922768354416, + 0.04762649908661842, + 0.04510583356022835, + 0.018231119960546494, + 0.04195871204137802, + 0.015276818536221981, + 0.13600310683250427, + 0.20478995144367218, + 0.04454568773508072, + 0.38591039180755615 + ], + "credit_map": [ + 0.008979929611086845, + 0.004750972148030996, + 0.011217562481760979, + 0.006593368481844664, + 0.005688333883881569, + 0.007378698792308569, + 0.007105804048478603, + 0.008837255649268627, + 0.004327967297285795, + 0.0030518099665641785, + 0.009995120577514172, + 0.00611068494617939, + 0.005657171364873648, + 0.0036492743529379368, + 0.011250561103224754, + 0.003583911107853055, + 0.005338282324373722, + 0.003387914039194584, + 0.013833757489919662, + 0.005069729872047901, + 0.002949009882286191, + 0.003741207765415311, + 0.007130630314350128, + 0.0036553051322698593, + 0.0022214502096176147, + 0.0028315347153693438, + 0.0021292183082550764, + 0.0018947514472529292, + 0.0024253642186522484, + 0.0018267394043505192, + 0.0034696971997618675, + 0.0014323638752102852, + 0.0036872962955385447, + 0.0038440327625721693, + 0.0033061003778129816, + 0.006226512137800455, + 0.007745259907096624, + 0.005504040978848934, + 0.007971678860485554, + 0.0036737904883921146, + 0.0017516951775178313, + 0.0015401646960526705, + 0.001913818414323032, + 0.0025277030654251575, + 0.0022843291517347097, + 0.0018754263874143362, + 0.0018907448975369334, + 0.0014929374447092414, + 0.019432349130511284, + 0.008573184721171856, + 0.005305313039571047, + 0.012342069298028946, + 0.010102548636496067, + 0.03742373362183571, + 0.03527728468179703, + 0.007546615786850452, + 0.0170943271368742, + 0.011455768719315529, + 0.011148200370371342, + 0.025419531390070915, + 0.014518260955810547, + 0.04488912969827652, + 0.06816723197698593, + 0.012097508646547794, + 0.00960548222064972, + 0.003454944584518671, + 0.0031978036276996136, + 0.004615738056600094, + 0.0029527884908020496, + 0.011669255793094635, + 0.005764104891568422, + 0.0032855705358088017, + 0.055058687925338745, + 0.021008554846048355, + 0.15746764838695526, + 0.07506110519170761, + 0.017727885395288467, + 0.016361849382519722, + 0.03332412615418434, + 0.009900541976094246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.011710118502378464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553717941045761 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553717941045761, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553717941045761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15598396956920624, + 0.026956304907798767, + 0.05301954597234726, + 0.02365286648273468, + 0.04373262822628021, + 0.013436500914394855, + 0.13815832138061523, + 0.278659462928772, + 0.11056961864233017, + 0.1558307707309723 + ], + "credit_map": [ + 0.009915437549352646, + 0.010486614890396595, + 0.020339660346508026, + 0.014182928949594498, + 0.04498504102230072, + 0.015706269070506096, + 0.02535117045044899, + 0.015016845427453518, + 0.0030276398174464703, + 0.002118421485647559, + 0.005045308731496334, + 0.0029307568911463022, + 0.0036040383856743574, + 0.0027722495142370462, + 0.005182979628443718, + 0.0022749111521989107, + 0.008305288851261139, + 0.0026621161960065365, + 0.018500639125704765, + 0.0032237654086202383, + 0.0029764268547296524, + 0.004063364118337631, + 0.009296407923102379, + 0.003991535399109125, + 0.002545723458752036, + 0.004281010944396257, + 0.002767202677205205, + 0.002337421989068389, + 0.0027277334593236446, + 0.003132440382614732, + 0.0033475172240287066, + 0.0025138151831924915, + 0.00815869402140379, + 0.005078330636024475, + 0.0036961783189326525, + 0.004037315957248211, + 0.0035858124028891325, + 0.004435753915458918, + 0.010884265415370464, + 0.0038562784902751446, + 0.0029357688035815954, + 0.001894723274745047, + 0.0012687561102211475, + 0.0015109179075807333, + 0.0014175177784636617, + 0.001713483245112002, + 0.0014455264899879694, + 0.00124980672262609, + 0.018988877534866333, + 0.008450654335319996, + 0.004997628275305033, + 0.011758513748645782, + 0.010055460967123508, + 0.04117431864142418, + 0.03521028533577919, + 0.007522576954215765, + 0.02857399359345436, + 0.0171968974173069, + 0.01574670895934105, + 0.03240542858839035, + 0.03449564799666405, + 0.04801245778799057, + 0.07919090986251831, + 0.023037416860461235, + 0.025472033768892288, + 0.007131458260118961, + 0.008591135032474995, + 0.009768995456397533, + 0.0070695457980036736, + 0.02770877070724964, + 0.016539985314011574, + 0.008287686854600906, + 0.022985387593507767, + 0.008511144667863846, + 0.06250952184200287, + 0.028585562482476234, + 0.007941053248941898, + 0.007398421410471201, + 0.013439473696053028, + 0.004460195079445839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.012553717941045761 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486328423023224 + ], + "opposite_logprob_tokens": [ + -1.2236328125 + ], + "logprob_reward": -0.3486328423023224, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486328423023224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09967128187417984, + 0.025573277845978737, + 0.09155050665140152, + 0.028994422405958176, + 0.11286461353302002, + 0.01272578164935112, + 0.07585815340280533, + 0.06929709017276764, + 0.13838538527488708, + 0.3450794219970703 + ], + "credit_map": [ + 0.007063102908432484, + 0.006593347527086735, + 0.011667872779071331, + 0.009705856442451477, + 0.028202300891280174, + 0.010576216503977776, + 0.01553976722061634, + 0.010322821326553822, + 0.0022701567504554987, + 0.0016526191029697657, + 0.004718737211078405, + 0.0024207925889641047, + 0.0033484601881355047, + 0.00207600393332541, + 0.007210992742329836, + 0.0018755177734419703, + 0.012527608312666416, + 0.006283191032707691, + 0.026663638651371002, + 0.007478339597582817, + 0.008061949163675308, + 0.008385684341192245, + 0.014968061819672585, + 0.007182043045759201, + 0.0026606563478708267, + 0.002129422267898917, + 0.0026326864026486874, + 0.003095461055636406, + 0.0029427327681332827, + 0.005510163493454456, + 0.0078017134219408035, + 0.002221588045358658, + 0.022666437551379204, + 0.012486572377383709, + 0.009104018099606037, + 0.010235538706183434, + 0.010013783350586891, + 0.010213353671133518, + 0.029003750532865524, + 0.009141163900494576, + 0.00280635803937912, + 0.001533751841634512, + 0.0017364168306812644, + 0.0012343453709036112, + 0.00133962940890342, + 0.0013255917001515627, + 0.0015685813268646598, + 0.0011811080621555448, + 0.00981448870152235, + 0.0050692385993897915, + 0.0030846891459077597, + 0.00787652563303709, + 0.007187803741544485, + 0.01915828324854374, + 0.019690347835421562, + 0.003976786509156227, + 0.006544092670083046, + 0.0035098427906632423, + 0.0033625441137701273, + 0.009081832133233547, + 0.006745962891727686, + 0.01449376530945301, + 0.02236562967300415, + 0.0031934252474457026, + 0.02698778361082077, + 0.00581926666200161, + 0.03086422197520733, + 0.008287680335342884, + 0.007308480329811573, + 0.022825855761766434, + 0.02916686236858368, + 0.007125236093997955, + 0.04849805310368538, + 0.01894613727927208, + 0.14013144373893738, + 0.0661030188202858, + 0.016524506732821465, + 0.015276632271707058, + 0.029879603534936905, + 0.009720053523778915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.3486328423023224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539120391011238 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012539120391011238, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539120391011238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22372248768806458, + 0.0350143238902092, + 0.07439300417900085, + 0.0427965447306633, + 0.11323132365942001, + 0.009308967739343643, + 0.06608409434556961, + 0.0316944494843483, + 0.0396210215985775, + 0.3641338646411896 + ], + "credit_map": [ + 0.014395666308701038, + 0.01437679398804903, + 0.03174935653805733, + 0.020491046831011772, + 0.06110329553484917, + 0.020619479939341545, + 0.04001077637076378, + 0.020976049825549126, + 0.0031646394636482, + 0.0028469215612858534, + 0.006263887509703636, + 0.004421745426952839, + 0.004064152482897043, + 0.003471736330538988, + 0.007587176747620106, + 0.003194062504917383, + 0.0106401601806283, + 0.004982230719178915, + 0.021963603794574738, + 0.0058999997563660145, + 0.005948776379227638, + 0.006465184967964888, + 0.012585392221808434, + 0.005907650105655193, + 0.0042543839663267136, + 0.010219845920801163, + 0.0038859378546476364, + 0.003958357032388449, + 0.002893112599849701, + 0.004800637252628803, + 0.009369853883981705, + 0.003414412494748831, + 0.022634711116552353, + 0.010089879855513573, + 0.00907664094120264, + 0.009815293364226818, + 0.010732794180512428, + 0.009574380703270435, + 0.03342829644680023, + 0.007879318669438362, + 0.001099692308343947, + 0.001055665547028184, + 0.001051645609550178, + 0.0009464708273299038, + 0.001515056355856359, + 0.0014252775581553578, + 0.0010956990299746394, + 0.0011194598628208041, + 0.009570781141519547, + 0.004246382508426905, + 0.002188953338190913, + 0.005780806299299002, + 0.005013528745621443, + 0.019241340458393097, + 0.016532912850379944, + 0.0035093906335532665, + 0.0038231986109167337, + 0.0021044714376330376, + 0.002192017389461398, + 0.0038619632832705975, + 0.002916521392762661, + 0.0058902110904455185, + 0.008070498704910278, + 0.0028355654794722795, + 0.0074063921347260475, + 0.0027845357544720173, + 0.0061085461638867855, + 0.004375625401735306, + 0.003291807137429714, + 0.00730371056124568, + 0.005248675122857094, + 0.0031017267610877752, + 0.05237673223018646, + 0.019323041662573814, + 0.14884544909000397, + 0.06976727396249771, + 0.017155535519123077, + 0.01566794328391552, + 0.030938152223825455, + 0.010059687308967113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.012539120391011238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179696053266525 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.018179696053266525, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018179696053266525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15478456020355225, + 0.02018064819276333, + 0.015917232260107994, + 0.04308786615729332, + 0.05599301680922508, + 0.027361618354916573, + 0.09281150251626968, + 0.2831200659275055, + 0.16033269464969635, + 0.14641082286834717 + ], + "credit_map": [ + 0.014684204012155533, + 0.012908849865198135, + 0.01891362853348255, + 0.013915237970650196, + 0.039194636046886444, + 0.01757066883146763, + 0.023668086156249046, + 0.013929245993494987, + 0.0023485207930207253, + 0.0031027370132505894, + 0.0022784823086112738, + 0.0020174370147287846, + 0.0036424652207642794, + 0.0027136856224387884, + 0.0023459147196263075, + 0.0017314048018306494, + 0.001753921969793737, + 0.0023459249641746283, + 0.0018617782043293118, + 0.002430892316624522, + 0.001610576524399221, + 0.0027814810164272785, + 0.0014136610552668571, + 0.0017189972568303347, + 0.004845758434385061, + 0.004579240456223488, + 0.005170363932847977, + 0.004688008688390255, + 0.004548069555312395, + 0.005679852329194546, + 0.009968006052076817, + 0.003608568338677287, + 0.009319351986050606, + 0.005432842765003443, + 0.005145116243511438, + 0.005841465201228857, + 0.006708994973450899, + 0.005347791127860546, + 0.013772102072834969, + 0.0044253552332520485, + 0.0035795627627521753, + 0.00349023868329823, + 0.0026779158506542444, + 0.002974354662001133, + 0.004728575702756643, + 0.0038779296446591616, + 0.003301252145320177, + 0.002731789369136095, + 0.012336577288806438, + 0.0057137650437653065, + 0.0037103404756635427, + 0.009330607950687408, + 0.006955210119485855, + 0.024893026798963547, + 0.024817364290356636, + 0.005054614506661892, + 0.030997635796666145, + 0.016013801097869873, + 0.015896186232566833, + 0.035402797162532806, + 0.026367120444774628, + 0.0531807579100132, + 0.08637409657239914, + 0.018887639045715332, + 0.0384456142783165, + 0.009604782797396183, + 0.02106326073408127, + 0.012325849384069443, + 0.00974132027477026, + 0.03323019668459892, + 0.024392060935497284, + 0.011529620736837387, + 0.017223922535777092, + 0.00878010131418705, + 0.058489229530096054, + 0.030480999499559402, + 0.0078042903915047646, + 0.006502307951450348, + 0.012639937922358513, + 0.0044900341890752316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.018179696053266525 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot smoothly follows a cheerful, bright yellow snowplow, rendered in the charming style of Pixar animation, as it steadily advances. The plow's rounded edges and smooth finish gleam in the soft, unfocused daylight. Its two large, round headlights, designed like smiling eyes, emit a warm, friendly glow, adding to its anthropomorphic character. Its sturdy black caterpillar trails rhythmically churn, leaving deep marks as they propel the vehicle through a thick, untouched blanket of fresh, powdery snow. The massive front blades meticulously carve a clean path, causing the snow to curl and roll, forming two perfect, fluffy banks on either side. A very shallow depth of field is employed, beautifully blurring the background of tall, snow-covered pine trees into a soft, ethereal winter mist composed of muted greens and whites. The overall atmosphere is one of whimsical tranquility and joy, dominated by the vibrant yellow of the plow and the bright white of the landscape.", + "seed": 242, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_1714_seed_242_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0867654975503683, + "vlm_logprob": -0.0867654975503683, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0867654975503683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012799020856618881 + ], + "opposite_logprob_tokens": [ + -6.38779878616333 + ], + "logprob_reward": -0.012799020856618881, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.012799020856618881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16736747324466705, + 0.13745719194412231, + 0.1332986205816269, + 0.03834933042526245, + 0.08109936863183975, + 0.21904049813747406, + 0.0973016768693924, + 0.0187528133392334, + 0.030018458142876625, + 0.07731460779905319 + ], + "credit_map": [ + 0.012153339572250843, + 0.010766132734715939, + 0.025333266705274582, + 0.015861794352531433, + 0.026175973936915398, + 0.05429649353027344, + 0.011039834469556808, + 0.011740623041987419, + 0.010210433974862099, + 0.009861558675765991, + 0.017392778769135475, + 0.026767054572701454, + 0.018441539257764816, + 0.022770503535866737, + 0.0189217422157526, + 0.013091577216982841, + 0.011873828247189522, + 0.005011602770537138, + 0.019486622884869576, + 0.009415457025170326, + 0.018138060346245766, + 0.050877731293439865, + 0.013120877556502819, + 0.005374427884817123, + 0.0031156353652477264, + 0.003343060379847884, + 0.0029129181057214737, + 0.008952150121331215, + 0.007031157612800598, + 0.00444638030603528, + 0.005539780016988516, + 0.0030082473531365395, + 0.009294153191149235, + 0.007537933066487312, + 0.009539704769849777, + 0.012471655383706093, + 0.00955892913043499, + 0.014132395386695862, + 0.009548314847052097, + 0.00901627168059349, + 0.022387869656085968, + 0.012085947208106518, + 0.04246906563639641, + 0.013764909468591213, + 0.026842355728149414, + 0.07629024982452393, + 0.01783652789890766, + 0.007363553158938885, + 0.010236273519694805, + 0.004892235156148672, + 0.019295288249850273, + 0.0068646324798464775, + 0.017049776390194893, + 0.026942379772663116, + 0.006978507619351149, + 0.005042571574449539, + 0.0029494650661945343, + 0.0020924052223563194, + 0.0021393350325524807, + 0.0036024434957653284, + 0.0021530643571168184, + 0.0023412450682371855, + 0.001902728108689189, + 0.00157212617341429, + 0.004183847922831774, + 0.003181661479175091, + 0.005507103633135557, + 0.003960710018873215, + 0.0030680112540721893, + 0.004721538163721561, + 0.002144407480955124, + 0.0032511770259588957, + 0.00763622485101223, + 0.004056754522025585, + 0.015737492591142654, + 0.006971381139010191, + 0.010094179771840572, + 0.01973474770784378, + 0.0077714864164590836, + 0.005312332883477211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.012799020856618881 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210364505648613 + ], + "opposite_logprob_tokens": [ + -4.018210411071777 + ], + "logprob_reward": -0.018210364505648613, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018210364505648613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.163185715675354, + 0.22014260292053223, + 0.11868367344141006, + 0.021129680797457695, + 0.05741400271654129, + 0.06085021421313286, + 0.25719964504241943, + 0.03317561373114586, + 0.042981550097465515, + 0.025237325578927994 + ], + "credit_map": [ + 0.013304503634572029, + 0.011676410213112831, + 0.03202976658940315, + 0.016255652531981468, + 0.023988105356693268, + 0.04890844225883484, + 0.009169593453407288, + 0.0078532500192523, + 0.01640116050839424, + 0.014411349780857563, + 0.029775049537420273, + 0.03528807312250137, + 0.030727507546544075, + 0.04861845076084137, + 0.02676106058061123, + 0.01815994828939438, + 0.012373024597764015, + 0.00601767236366868, + 0.015542863868176937, + 0.0058660730719566345, + 0.017015479505062103, + 0.04633910208940506, + 0.011064202524721622, + 0.0044652423821389675, + 0.0021068574860692024, + 0.001415316597558558, + 0.003421083791181445, + 0.0027620515320450068, + 0.002573384903371334, + 0.005141609348356724, + 0.0022860486060380936, + 0.001423328067176044, + 0.005550433415919542, + 0.005354014690965414, + 0.004860414192080498, + 0.00899210199713707, + 0.008411037735641003, + 0.009534798562526703, + 0.007842899300158024, + 0.006868300028145313, + 0.007626303471624851, + 0.004763391800224781, + 0.011106979101896286, + 0.00395174091681838, + 0.007475779391825199, + 0.017885541543364525, + 0.004153072368353605, + 0.003887398401275277, + 0.02042170614004135, + 0.009267413057386875, + 0.061123624444007874, + 0.014048712328076363, + 0.036411505192518234, + 0.08981845527887344, + 0.01769505813717842, + 0.008413144387304783, + 0.006300237961113453, + 0.0037168553099036217, + 0.0041335872374475, + 0.004529254045337439, + 0.003458211198449135, + 0.005216378252953291, + 0.0026583049912005663, + 0.0031627847347408533, + 0.004454640671610832, + 0.0037968545220792294, + 0.004776981193572283, + 0.007088357117027044, + 0.0064376783557236195, + 0.006666037254035473, + 0.004185772966593504, + 0.005575225222855806, + 0.002376462798565626, + 0.001091084093786776, + 0.004908340517431498, + 0.001985896145924926, + 0.0034084455110132694, + 0.007660122122615576, + 0.0021134556736797094, + 0.0016935169696807861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.018210364505648613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486495018005371 + ], + "opposite_logprob_tokens": [ + -1.223649501800537 + ], + "logprob_reward": -0.3486495018005371, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486495018005371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21806462109088898, + 0.07722928375005722, + 0.30900681018829346, + 0.05557786300778389, + 0.04749027639627457, + 0.09511344879865646, + 0.11366169154644012, + 0.04109901562333107, + 0.01558386068791151, + 0.027173062786459923 + ], + "credit_map": [ + 0.01725168339908123, + 0.015189976431429386, + 0.04227299615740776, + 0.016646431758999825, + 0.03209136798977852, + 0.07403974235057831, + 0.010797528550028801, + 0.009774896316230297, + 0.0063368212431669235, + 0.005615179892629385, + 0.010801891796290874, + 0.013068004511296749, + 0.00853855349123478, + 0.015538379549980164, + 0.009440184570848942, + 0.007890263572335243, + 0.031724490225315094, + 0.014007048681378365, + 0.04441245645284653, + 0.010727491229772568, + 0.043177224695682526, + 0.12485896795988083, + 0.029246153309941292, + 0.010852974839508533, + 0.005533394869416952, + 0.002777889370918274, + 0.00742442486807704, + 0.005029706750065088, + 0.00694769574329257, + 0.019111402332782745, + 0.006196076050400734, + 0.0025572706945240498, + 0.005604091100394726, + 0.005347440950572491, + 0.004512656945735216, + 0.007169865071773529, + 0.005926432553678751, + 0.007300588767975569, + 0.006197118200361729, + 0.0054320828057825565, + 0.011641154997050762, + 0.007992947474122047, + 0.01601484976708889, + 0.0056768497452139854, + 0.01144021563231945, + 0.028307735919952393, + 0.007457010913640261, + 0.006582687143236399, + 0.010121637023985386, + 0.0046051982790231705, + 0.026567619293928146, + 0.006465233396738768, + 0.015213638544082642, + 0.038446106016635895, + 0.007602689787745476, + 0.004639570135623217, + 0.009251989424228668, + 0.004841597750782967, + 0.005205539520829916, + 0.005926568526774645, + 0.004209019243717194, + 0.00453526433557272, + 0.003338718554005027, + 0.0037903194315731525, + 0.002176719717681408, + 0.0015421815915033221, + 0.0018295476911589503, + 0.0023444066755473614, + 0.0018069588113576174, + 0.0028199953958392143, + 0.0013097439659759402, + 0.0017543082358315587, + 0.002508231671527028, + 0.001943632960319519, + 0.004816064611077309, + 0.0032136717345565557, + 0.003541720798239112, + 0.006295213010162115, + 0.0029274693224579096, + 0.0019270578632131219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.3486495018005371 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810257837176323 + ], + "opposite_logprob_tokens": [ + -3.288102626800537 + ], + "logprob_reward": -0.03810257837176323, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810257837176323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17423269152641296, + 0.09485691785812378, + 0.18299198150634766, + 0.022413887083530426, + 0.027381515130400658, + 0.21342733502388, + 0.1382286101579666, + 0.03427060693502426, + 0.01671074889600277, + 0.09548579156398773 + ], + "credit_map": [ + 0.01303467433899641, + 0.015104134567081928, + 0.032213542610406876, + 0.0173532422631979, + 0.025862669572234154, + 0.0494169220328331, + 0.011544039472937584, + 0.009703461080789566, + 0.006870302837342024, + 0.005641141440719366, + 0.01420981902629137, + 0.014881843701004982, + 0.011031217873096466, + 0.02217419072985649, + 0.011167393997311592, + 0.008881006389856339, + 0.017900817096233368, + 0.007298297714442015, + 0.026675330474972725, + 0.006197370123118162, + 0.02601640112698078, + 0.07583294063806534, + 0.017826426774263382, + 0.005244384054094553, + 0.0019713800866156816, + 0.0011942877899855375, + 0.0024945256300270557, + 0.0019499680493026972, + 0.0033289319835603237, + 0.0074720242992043495, + 0.002652697963640094, + 0.0013500695349648595, + 0.003554969560354948, + 0.003167100716382265, + 0.002380315912887454, + 0.004057265818119049, + 0.004550138488411903, + 0.0035015472676604986, + 0.0030156609136611223, + 0.003154514590278268, + 0.022577950730919838, + 0.012310675345361233, + 0.04260639846324921, + 0.011024368926882744, + 0.025763684883713722, + 0.07483917474746704, + 0.015880819410085678, + 0.008424249477684498, + 0.012216751463711262, + 0.005627371370792389, + 0.03247491270303726, + 0.008300143294036388, + 0.019234683364629745, + 0.04491555690765381, + 0.009409724734723568, + 0.006049441173672676, + 0.006256308872252703, + 0.0034001870080828667, + 0.004600099753588438, + 0.004495383705943823, + 0.0036582937464118004, + 0.005002015735954046, + 0.0029838639311492443, + 0.00387445162050426, + 0.0021229134872555733, + 0.0015917911659926176, + 0.0017892763717100024, + 0.002319860039278865, + 0.0020468130242079496, + 0.0030050671193748713, + 0.0013648847816511989, + 0.0024701410438865423, + 0.008083241060376167, + 0.0050544883124530315, + 0.01799086295068264, + 0.011254251934587955, + 0.012558219954371452, + 0.0210858266800642, + 0.009655282832682133, + 0.009803618304431438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.03810257837176323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066022217273712 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066022217273712, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066022217273712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4462835192680359, + 0.11347364634275436, + 0.04186476767063141, + 0.06907551735639572, + 0.03545409068465233, + 0.1192391887307167, + 0.08538789302110672, + 0.05720065161585808, + 0.01975221559405327, + 0.01226856466382742 + ], + "credit_map": [ + 0.03245364502072334, + 0.023973800241947174, + 0.08537952601909637, + 0.03010678105056286, + 0.06728343665599823, + 0.16264604032039642, + 0.02353765442967415, + 0.02090262994170189, + 0.010516906157135963, + 0.006246811710298061, + 0.02171442098915577, + 0.007939470000565052, + 0.014771101996302605, + 0.03844247758388519, + 0.00787640456110239, + 0.005966059863567352, + 0.0060107815079391, + 0.00389310996979475, + 0.004029474686831236, + 0.006877850275486708, + 0.005874346941709518, + 0.005513057578355074, + 0.005288491025567055, + 0.0043776556849479675, + 0.006906668655574322, + 0.003916108515113592, + 0.009633351117372513, + 0.005753750912845135, + 0.008154699578881264, + 0.024012956768274307, + 0.007420751731842756, + 0.00327722798101604, + 0.005551314912736416, + 0.004031646531075239, + 0.004520032089203596, + 0.0030232674907892942, + 0.004395862575620413, + 0.008325155824422836, + 0.003399637993425131, + 0.002207172801718116, + 0.013297307305037975, + 0.008817668072879314, + 0.022621292620897293, + 0.006384425330907106, + 0.014206252060830593, + 0.040372397750616074, + 0.009294053539633751, + 0.0042457920499145985, + 0.006888901349157095, + 0.0033310698345303535, + 0.020088160410523415, + 0.005146573763340712, + 0.011736583895981312, + 0.029563328251242638, + 0.005821811500936747, + 0.0028114670421928167, + 0.008313087746500969, + 0.005665282253175974, + 0.006809186656028032, + 0.009494791738688946, + 0.0061798095703125, + 0.0073153129778802395, + 0.0046132029965519905, + 0.008809980005025864, + 0.0027297339402139187, + 0.0021756156347692013, + 0.0021973305847495794, + 0.002599746687337756, + 0.002775984350591898, + 0.003407767042517662, + 0.0014579762937501073, + 0.0024080604780465364, + 0.0018365903524681926, + 0.0011349612614139915, + 0.001666978932917118, + 0.0014168238267302513, + 0.001435997779481113, + 0.0015515139093622565, + 0.0015881378203630447, + 0.0016375614795833826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.016066022217273712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot smoothly follows a cheerful, bright yellow snowplow, rendered in the charming style of Pixar animation, as it steadily advances. The plow's rounded edges and smooth finish gleam in the soft, unfocused daylight. Its two large, round headlights, designed like smiling eyes, emit a warm, friendly glow, adding to its anthropomorphic character. Its sturdy black caterpillar trails rhythmically churn, leaving deep marks as they propel the vehicle through a thick, untouched blanket of fresh, powdery snow. The massive front blades meticulously carve a clean path, causing the snow to curl and roll, forming two perfect, fluffy banks on either side. A very shallow depth of field is employed, beautifully blurring the background of tall, snow-covered pine trees into a soft, ethereal winter mist composed of muted greens and whites. The overall atmosphere is one of whimsical tranquility and joy, dominated by the vibrant yellow of the plow and the bright white of the landscape.", + "seed": 243, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_10/index_1714_seed_243_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11482153907418251, + "vlm_logprob": -0.11482153907418251, + "vlm_yes_no_margin": 3.400000011920929, + "vlm_token_credit": -0.11482153907418251, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0101179089397192 + ], + "opposite_logprob_tokens": [ + -6.010118007659912 + ], + "logprob_reward": -0.0101179089397192, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0101179089397192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22150558233261108, + 0.13091015815734863, + 0.15961092710494995, + 0.07628146559000015, + 0.07935136556625366, + 0.03676113858819008, + 0.03502396494150162, + 0.0679657831788063, + 0.11824191361665726, + 0.07434767484664917 + ], + "credit_map": [ + 0.027694854885339737, + 0.027972962707281113, + 0.020398244261741638, + 0.062062282115221024, + 0.01972201280295849, + 0.008545875549316406, + 0.04510372504591942, + 0.01000563707202673, + 0.02211514301598072, + 0.023473141714930534, + 0.01759808138012886, + 0.014574963599443436, + 0.01753607764840126, + 0.01084189210087061, + 0.013111667707562447, + 0.011659215204417706, + 0.021207673475146294, + 0.011105620302259922, + 0.02615216188132763, + 0.037797961384058, + 0.016297047957777977, + 0.009375481866300106, + 0.02821965329349041, + 0.009455339051783085, + 0.011282799765467644, + 0.008202469907701015, + 0.009448700584471226, + 0.007203558925539255, + 0.010768620297312737, + 0.008608153089880943, + 0.008225974626839161, + 0.012541195377707481, + 0.009113546460866928, + 0.012791695073246956, + 0.011300510726869106, + 0.00948142260313034, + 0.01511668972671032, + 0.005224824417382479, + 0.008306206203997135, + 0.00801647175103426, + 0.004644498229026794, + 0.005765104200690985, + 0.003789182985201478, + 0.005243259482085705, + 0.008223206736147404, + 0.002483212389051914, + 0.004276135470718145, + 0.0023365404922515154, + 0.003476535901427269, + 0.006560744252055883, + 0.005183226428925991, + 0.003744119079783559, + 0.005469774827361107, + 0.002443996723741293, + 0.005129245109856129, + 0.0030163226183503866, + 0.007555823307484388, + 0.009741749614477158, + 0.007000190205872059, + 0.010051879100501537, + 0.013569114729762077, + 0.006329566705971956, + 0.008733049035072327, + 0.0049844118766486645, + 0.016676463186740875, + 0.018882721662521362, + 0.017910566180944443, + 0.013494934886693954, + 0.014246068894863129, + 0.01122551690787077, + 0.012734300456941128, + 0.013071353547275066, + 0.012834521941840649, + 0.007219568360596895, + 0.014280908741056919, + 0.010127191431820393, + 0.006545383483171463, + 0.00899623241275549, + 0.007570789661258459, + 0.006773076020181179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0101179089397192 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212694019079208 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212694019079208, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212694019079208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23331549763679504, + 0.2136455923318863, + 0.17572025954723358, + 0.0705006793141365, + 0.051388613879680634, + 0.024518167600035667, + 0.02610384114086628, + 0.032979849725961685, + 0.11624787747859955, + 0.055579569190740585 + ], + "credit_map": [ + 0.02787052094936371, + 0.03872671723365784, + 0.022094491869211197, + 0.05180719867348671, + 0.023848867043852806, + 0.014882770366966724, + 0.0365467332303524, + 0.017538243904709816, + 0.033182866871356964, + 0.04476366564631462, + 0.02260306477546692, + 0.021041734144091606, + 0.037861473858356476, + 0.019126178696751595, + 0.02094779908657074, + 0.014118836261332035, + 0.025389401242136955, + 0.011120634153485298, + 0.023528460413217545, + 0.050507690757513046, + 0.014948943629860878, + 0.007244812790304422, + 0.034737590700387955, + 0.008242757990956306, + 0.008404972031712532, + 0.007110299076884985, + 0.008133416064083576, + 0.006745384074747562, + 0.018541483208537102, + 0.004521016031503677, + 0.009308201260864735, + 0.007735915016382933, + 0.006651786155998707, + 0.00839728768914938, + 0.006887158844619989, + 0.006951701361685991, + 0.009203857742249966, + 0.0036790177691727877, + 0.004945235326886177, + 0.004672577604651451, + 0.0030152080580592155, + 0.0032710344530642033, + 0.004077640827745199, + 0.0028489104006439447, + 0.004614307079464197, + 0.0020951852202415466, + 0.0028415857814252377, + 0.001754298689775169, + 0.003006913233548403, + 0.0034913504496216774, + 0.0057005309499800205, + 0.0032921957317739725, + 0.0032992830965667963, + 0.002182008232921362, + 0.0031968767289072275, + 0.0019346870249137282, + 0.0024739226792007685, + 0.005577658303081989, + 0.004055849742144346, + 0.003895991947501898, + 0.007027268875390291, + 0.002386381383985281, + 0.004918098449707031, + 0.002644680440425873, + 0.012340781278908253, + 0.015748871490359306, + 0.019000105559825897, + 0.014727959409356117, + 0.02203395776450634, + 0.00981207937002182, + 0.013978933915495872, + 0.00860520452260971, + 0.007143303286284208, + 0.009657332673668861, + 0.009553232230246067, + 0.007013358175754547, + 0.0066536301746964455, + 0.0061014737002551556, + 0.0055703637190163136, + 0.003886881750077009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014212694019079208 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890065908432007 + ], + "opposite_logprob_tokens": [ + -1.0539007186889648 + ], + "logprob_reward": -0.42890065908432007, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42890065908432007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10786446928977966, + 0.10429942607879639, + 0.13888810575008392, + 0.07107977569103241, + 0.04634260758757591, + 0.044305652379989624, + 0.05861067771911621, + 0.036962974816560745, + 0.3019145131111145, + 0.08973178267478943 + ], + "credit_map": [ + 0.014668299816548824, + 0.012676388025283813, + 0.01213163323700428, + 0.01883992739021778, + 0.01794154942035675, + 0.008096779696643353, + 0.014057020656764507, + 0.009452875703573227, + 0.011985820718109608, + 0.012779513373970985, + 0.01408399362117052, + 0.009365670382976532, + 0.02309361658990383, + 0.014465657994151115, + 0.0079639982432127, + 0.010561151430010796, + 0.01800597831606865, + 0.012748903594911098, + 0.023778563365340233, + 0.02545011416077614, + 0.02188940718770027, + 0.0051920656114816666, + 0.020724879577755928, + 0.011098203249275684, + 0.0071012601256370544, + 0.010667427442967892, + 0.00925507303327322, + 0.0075886608101427555, + 0.01329864002764225, + 0.004681003279983997, + 0.006896608043462038, + 0.011591099202632904, + 0.006096969824284315, + 0.0055794925428926945, + 0.006510117091238499, + 0.005664556287229061, + 0.007633163593709469, + 0.005012307316064835, + 0.004501760937273502, + 0.00534423952922225, + 0.006122719030827284, + 0.007849505171179771, + 0.0043482473120093346, + 0.0060677239671349525, + 0.009194372221827507, + 0.0027535269036889076, + 0.005860462319105864, + 0.0021090982481837273, + 0.007089997176080942, + 0.008520210161805153, + 0.005605481564998627, + 0.007393504958599806, + 0.011563421227037907, + 0.00716704735532403, + 0.006375931669026613, + 0.0048950836062431335, + 0.00444129528477788, + 0.008035476319491863, + 0.003238917561247945, + 0.0034821408335119486, + 0.006783731281757355, + 0.00355553743429482, + 0.004629420582205057, + 0.0027964545879513025, + 0.036671269685029984, + 0.03591474890708923, + 0.06511717289686203, + 0.03894725441932678, + 0.04600038379430771, + 0.022104883566498756, + 0.034480441361665726, + 0.022678358480334282, + 0.009874788112938404, + 0.013681285083293915, + 0.010918638668954372, + 0.009991563856601715, + 0.01135129202157259, + 0.015128786675632, + 0.009627951309084892, + 0.009157477878034115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.42890065908432007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15679024159908295, + 0.07409952580928802, + 0.17825128138065338, + 0.09568223357200623, + 0.05504444241523743, + 0.107066310942173, + 0.0731368139386177, + 0.06958230584859848, + 0.0902147889137268, + 0.100132055580616 + ], + "credit_map": [ + 0.022822562605142593, + 0.01865890435874462, + 0.01288112998008728, + 0.037663232535123825, + 0.01830301247537136, + 0.008197297342121601, + 0.027138471603393555, + 0.011125626973807812, + 0.010249428451061249, + 0.010310872457921505, + 0.012194606475532055, + 0.005604641046375036, + 0.012115024030208588, + 0.009655974805355072, + 0.0058542401529848576, + 0.008114742115139961, + 0.02428562380373478, + 0.009713082574307919, + 0.022250907495617867, + 0.052701711654663086, + 0.010274012573063374, + 0.014042546041309834, + 0.035926125943660736, + 0.009057262912392616, + 0.014645647257566452, + 0.011346343904733658, + 0.013411046005785465, + 0.010023905895650387, + 0.014605282805860043, + 0.007851089350879192, + 0.011069678701460361, + 0.012729239650070667, + 0.00771574629470706, + 0.006058738566935062, + 0.007841010577976704, + 0.006646708585321903, + 0.007604700978845358, + 0.006519373506307602, + 0.007123004179447889, + 0.005535159260034561, + 0.014786254614591599, + 0.009843015111982822, + 0.01635521650314331, + 0.02147749625146389, + 0.011912870220839977, + 0.009640245698392391, + 0.014760669320821762, + 0.008290547877550125, + 0.009981420822441578, + 0.009385772049427032, + 0.009284746833145618, + 0.010161764919757843, + 0.012101029977202415, + 0.008259419351816177, + 0.008897033520042896, + 0.005065623205155134, + 0.007160089444369078, + 0.01033554133027792, + 0.00715651037171483, + 0.008489695377647877, + 0.013790919445455074, + 0.006590802222490311, + 0.009411021135747433, + 0.006647726055234671, + 0.009128742851316929, + 0.012052520178258419, + 0.012159355916082859, + 0.013227514922618866, + 0.01539112813770771, + 0.008821510709822178, + 0.01128609012812376, + 0.008147918619215488, + 0.014323935844004154, + 0.009208377450704575, + 0.015767453238368034, + 0.021364305168390274, + 0.009259955957531929, + 0.008475902490317822, + 0.014165768399834633, + 0.007566363085061312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.02058347500860691 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow snowplow with rounded edges and a smooth finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowplow have two large, round headlights that glow warmly like smiling eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the snowplow's black caterpillar trails visible churning rhythmically and leaving deep marks in the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front blade carve a clean path creating two perfect, fluffy banks of snow on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a very shallow depth of field blurring the background pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029295831918716 + ], + "opposite_logprob_tokens": [ + -2.350292921066284 + ], + "logprob_reward": -0.10029295831918716, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029295831918716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14600518345832825, + 0.058787330985069275, + 0.0732775628566742, + 0.1637924313545227, + 0.12443254888057709, + 0.07223685830831528, + 0.04321340471506119, + 0.08310199528932571, + 0.1453673243522644, + 0.08978529274463654 + ], + "credit_map": [ + 0.023020809516310692, + 0.027049530297517776, + 0.014640610665082932, + 0.021873820573091507, + 0.02206435613334179, + 0.009905027225613594, + 0.01864789053797722, + 0.008803148753941059, + 0.006064241286367178, + 0.006886241491883993, + 0.009203964844346046, + 0.005938598420470953, + 0.00962323509156704, + 0.010397341102361679, + 0.00497837271541357, + 0.0056953439489007, + 0.011807107366621494, + 0.006281806156039238, + 0.006847756914794445, + 0.018529998138546944, + 0.010171560570597649, + 0.002949250629171729, + 0.012379852123558521, + 0.004310240037739277, + 0.022491177543997765, + 0.02613467536866665, + 0.023196503520011902, + 0.0185152105987072, + 0.029316555708646774, + 0.010657557286322117, + 0.017525063827633858, + 0.015955710783600807, + 0.018082227557897568, + 0.016580313444137573, + 0.019815409556031227, + 0.019562896341085434, + 0.015480771660804749, + 0.009685390628874302, + 0.014429224655032158, + 0.010796338319778442, + 0.0067146578803658485, + 0.009553602896630764, + 0.008958695456385612, + 0.019470883533358574, + 0.009276630356907845, + 0.003828412853181362, + 0.01068570651113987, + 0.0037482723128050566, + 0.005335596855729818, + 0.005056302063167095, + 0.00864574033766985, + 0.005252970848232508, + 0.006421111989766359, + 0.0038060303777456284, + 0.004892356693744659, + 0.003803296247497201, + 0.011837118305265903, + 0.010655433870851994, + 0.012448002584278584, + 0.008960756473243237, + 0.014193583279848099, + 0.007481739390641451, + 0.010658386163413525, + 0.006866981741040945, + 0.020883353427052498, + 0.015122463926672935, + 0.030796140432357788, + 0.018923461437225342, + 0.01731245405972004, + 0.011487359181046486, + 0.018315574154257774, + 0.012526536360383034, + 0.01332119945436716, + 0.00972055085003376, + 0.011431094259023666, + 0.014844617806375027, + 0.008299724198877811, + 0.01294485293328762, + 0.01091830711811781, + 0.008304959163069725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.10029295831918716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_11/index_tensor([3526, 3526])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_11/index_tensor([3526, 3526])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b5a6d50441d4c07627c617f77e30e31a17759a74 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_11/index_tensor([3526, 3526])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle, uninterrupted tracking shot follows a teenager on an extreme skateboard at breakneck speed, close to the ground. The teenager has shaggy black hair with a hint of silver-gray highlights, and his eyes are focused and piercing. He wears a loose black hoodie with abstract, non-figurative white graffiti printed on the back, dark khaki work pants, and a pair of well-worn professional skateboard shoes. The skateboard's deck is covered in colorful, wild street graffiti, and the colorful wheels are almost invisible at the high speed. He moves almost to the ground, deftly lowering his body as he speeds along the city's main thoroughfare at dusk. Every tilt and shift of his weight is filled with rhythm and explosive power. As he precisely controlled the edge of his skateboard, it slashed across the asphalt. The custom-made wheels rubbed against the ground, instantly releasing a dazzling trail of golden and orange sparks. These particles trailed a dazzling trail of light in front of the camera, illuminating the world beneath his feet. The background was a city scene at a magical hour. The distant skyline burned with a deep orange-purple gradient, and the silhouettes of the tall buildings were gilded with a golden edge. The newly lit neon signs and streetlights on both sides of the street cast highly saturated, slightly haloed beams of light, creating visible volumetric light in the dust-filled air and a gorgeous anamorphic lens flare as the camera moved. The entire scene was imbued with Ufotable's signature visual style: an incredibly smooth, high-frame-rate hand-drawn Sakuga-style animation, crisp lines, and gorgeous light and shadow seamlessly integrated with CG effects. It perfectly combined the youthful energy of a young man with the psychedelic atmosphere of an urban dusk, creating an adrenaline-pumping visual feast.", + "seed": 29619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_11/index_3526_seed_29619_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02721091969870031, + "vlm_logprob": -0.02721091969870031, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02721091969870031, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026904527097940445 + ], + "opposite_logprob_tokens": [ + -3.651904582977295 + ], + "logprob_reward": -0.026904527097940445, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026904527097940445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24444490671157837, + 0.1998714804649353, + 0.03956691548228264, + 0.04633275792002678, + 0.14110136032104492, + 0.03051389753818512, + 0.0371139831840992, + 0.08620113134384155, + 0.085837222635746, + 0.0890163853764534 + ], + "credit_map": [ + 0.03163306042551994, + 0.03474186360836029, + 0.02865440957248211, + 0.03124113567173481, + 0.033979278057813644, + 0.03375856205821037, + 0.02483273483812809, + 0.025603875517845154, + 0.0346834696829319, + 0.05707563832402229, + 0.01520552672445774, + 0.01606551744043827, + 0.019752180203795433, + 0.030079498887062073, + 0.010663463734090328, + 0.016346175223588943, + 0.004556447267532349, + 0.005415968131273985, + 0.006414512172341347, + 0.006675063632428646, + 0.005253797397017479, + 0.0037824336905032396, + 0.003114491468295455, + 0.004354199860244989, + 0.005396381951868534, + 0.003630181308835745, + 0.00543089397251606, + 0.004798990674316883, + 0.00556759350001812, + 0.01003916747868061, + 0.004470461048185825, + 0.006999087519943714, + 0.02330508455634117, + 0.02877962216734886, + 0.010358911007642746, + 0.013859392143785954, + 0.017330897971987724, + 0.011282176710665226, + 0.013710161671042442, + 0.022475119680166245, + 0.0035003952216356993, + 0.0023120171390473843, + 0.0043859020806849, + 0.002317162463441491, + 0.003051749197766185, + 0.007108489517122507, + 0.003033664310351014, + 0.004804516676813364, + 0.004052413161844015, + 0.004339365754276514, + 0.005543735809624195, + 0.004253629595041275, + 0.003323542419821024, + 0.004357350990176201, + 0.0042654359713196754, + 0.006978508085012436, + 0.007181450724601746, + 0.019113389775156975, + 0.010519463568925858, + 0.0094337472692132, + 0.008135751821100712, + 0.012072669342160225, + 0.011509174481034279, + 0.00823548436164856, + 0.008668843656778336, + 0.02037062682211399, + 0.01078363973647356, + 0.008107523433864117, + 0.01144481636583805, + 0.00975708942860365, + 0.009496879763901234, + 0.007207801099866629, + 0.010286744683980942, + 0.01768246851861477, + 0.008685244247317314, + 0.008009890094399452, + 0.01596103422343731, + 0.009407089091837406, + 0.010051188059151173, + 0.00893272366374731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.026904527097940445 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784476831555367 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784476831555367, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784476831555367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12474935501813889, + 0.1757279485464096, + 0.05713050067424774, + 0.0768633484840393, + 0.0682283416390419, + 0.028625966981053352, + 0.024845631793141365, + 0.05866619944572449, + 0.09236101806163788, + 0.29280170798301697 + ], + "credit_map": [ + 0.016367200762033463, + 0.016942068934440613, + 0.016246046870946884, + 0.01612698659300804, + 0.017484933137893677, + 0.01711602322757244, + 0.011792543344199657, + 0.012673542834818363, + 0.03434291481971741, + 0.05776083469390869, + 0.01310146413743496, + 0.013396259397268295, + 0.01882277801632881, + 0.019489917904138565, + 0.009381734766066074, + 0.009432055056095123, + 0.007056029047816992, + 0.009744743816554546, + 0.0078066191636025906, + 0.007324981037527323, + 0.006821040995419025, + 0.006935490295290947, + 0.004874780774116516, + 0.006566816009581089, + 0.009159496054053307, + 0.008930541574954987, + 0.009072664193809032, + 0.007099607959389687, + 0.017199333757162094, + 0.011112404987215996, + 0.008890784345567226, + 0.00539851188659668, + 0.006530368234962225, + 0.010032794438302517, + 0.006352590397000313, + 0.007747172843664885, + 0.011698794551193714, + 0.007474182639271021, + 0.008889683522284031, + 0.009502752684056759, + 0.0034289127215743065, + 0.0039558690041303635, + 0.004310368560254574, + 0.002988791326060891, + 0.0024396225344389677, + 0.004437515512108803, + 0.004062923137098551, + 0.0030019653495401144, + 0.002344046253710985, + 0.0027785536367446184, + 0.0037117262836545706, + 0.0032036970369517803, + 0.002811187645420432, + 0.0021738684736192226, + 0.004020262975245714, + 0.0038022883236408234, + 0.005243311636149883, + 0.008409908972680569, + 0.009616303257644176, + 0.007507688831537962, + 0.006124603562057018, + 0.007560963276773691, + 0.007784154266119003, + 0.00641926284879446, + 0.01021368894726038, + 0.014102425426244736, + 0.00877958070486784, + 0.009211750701069832, + 0.010975138284265995, + 0.01281001977622509, + 0.014074168168008327, + 0.01219424232840538, + 0.03319631144404411, + 0.051485490053892136, + 0.030655885115265846, + 0.027980579063296318, + 0.049650635570287704, + 0.03383973240852356, + 0.03196615353226662, + 0.034026920795440674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.009784476831555367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736899726092815 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736899726092815, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736899726092815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08186694979667664, + 0.3251892030239105, + 0.08499133586883545, + 0.06604824960231781, + 0.16925108432769775, + 0.030743403360247612, + 0.01991765946149826, + 0.04819178953766823, + 0.03150676563382149, + 0.14229349792003632 + ], + "credit_map": [ + 0.008791565895080566, + 0.009200015105307102, + 0.011630116030573845, + 0.014299577102065086, + 0.00770952133461833, + 0.011196417734026909, + 0.008665263652801514, + 0.010374474339187145, + 0.06124535948038101, + 0.10395936667919159, + 0.024708341807127, + 0.028675949200987816, + 0.033784668892621994, + 0.04196755960583687, + 0.01682741567492485, + 0.014020537957549095, + 0.010378123261034489, + 0.012204446829855442, + 0.012212599627673626, + 0.012887494638562202, + 0.009486008435487747, + 0.008708331733942032, + 0.008198373019695282, + 0.010915952734649181, + 0.010275576263666153, + 0.007709946483373642, + 0.005945414770394564, + 0.0066098058596253395, + 0.013229995034635067, + 0.009249572642147541, + 0.007710867095738649, + 0.005317076109349728, + 0.024230780079960823, + 0.03279729187488556, + 0.013483277522027493, + 0.019092317670583725, + 0.0233725905418396, + 0.01498612854629755, + 0.018893476575613022, + 0.022395219653844833, + 0.0030333411414176226, + 0.0025005475617945194, + 0.004808858968317509, + 0.002760502742603421, + 0.0032323908526450396, + 0.006675558164715767, + 0.0031361831352114677, + 0.004596021492034197, + 0.002021525986492634, + 0.0025085152592509985, + 0.004387032240629196, + 0.00239625945687294, + 0.002010972471907735, + 0.0022002756595611572, + 0.0019499610643833876, + 0.0024431163910776377, + 0.004473318345844746, + 0.008636457845568657, + 0.005915414076298475, + 0.0046591125428676605, + 0.005830898880958557, + 0.007123136892914772, + 0.006979529745876789, + 0.00457392493262887, + 0.0032708693761378527, + 0.003134081605821848, + 0.004453421104699373, + 0.0030573629774153233, + 0.00417328579351306, + 0.004688179586082697, + 0.0055843619629740715, + 0.003145203460007906, + 0.01941259391605854, + 0.020038828253746033, + 0.013657021336257458, + 0.022504407912492752, + 0.01891980692744255, + 0.01526670902967453, + 0.008302241563796997, + 0.024191902950406075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.006736899726092815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036128980573266745 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.0036128980573266745, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036128980573266745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08979229629039764, + 0.07072559744119644, + 0.025176728144288063, + 0.039500270038843155, + 0.5793160200119019, + 0.012708674184978008, + 0.010155058465898037, + 0.026313763111829758, + 0.010465339757502079, + 0.13584625720977783 + ], + "credit_map": [ + 0.009599844925105572, + 0.006102955434471369, + 0.012715183198451996, + 0.016681792214512825, + 0.00557000283151865, + 0.018807008862495422, + 0.007848972454667091, + 0.012466534040868282, + 0.011628489010035992, + 0.01936042122542858, + 0.005722648464143276, + 0.00617068400606513, + 0.00889628380537033, + 0.00811407994478941, + 0.005118276458233595, + 0.00571470707654953, + 0.0027349446900188923, + 0.002971911570057273, + 0.004676403012126684, + 0.0034249324817210436, + 0.0032887959387153387, + 0.0028335023671388626, + 0.002955589909106493, + 0.002290649339556694, + 0.004294442478567362, + 0.005590516608208418, + 0.006948721595108509, + 0.003407512092962861, + 0.004344709683209658, + 0.0055927406065166, + 0.004195602145045996, + 0.005126027390360832, + 0.08534550666809082, + 0.10810325294733047, + 0.04093515872955322, + 0.06242745369672775, + 0.07535308599472046, + 0.052381012588739395, + 0.06678630411624908, + 0.08798426389694214, + 0.0013462393544614315, + 0.0013160643866285682, + 0.001991109922528267, + 0.0011706436052918434, + 0.0017851208103820682, + 0.002117581432685256, + 0.001238497206941247, + 0.0017434179317206144, + 0.0010954082245007157, + 0.0014255521818995476, + 0.0019087308319285512, + 0.0014051110483705997, + 0.0010680380510166287, + 0.0009890060173347592, + 0.0012577298330143094, + 0.001005482510663569, + 0.0020199106074869633, + 0.003332817228510976, + 0.003704689908772707, + 0.002897699363529682, + 0.003393984166905284, + 0.0034206598065793514, + 0.004875167738646269, + 0.002668834989890456, + 0.0012178230099380016, + 0.0015083535108715296, + 0.0023582722060382366, + 0.0011075080838054419, + 0.0011121747083961964, + 0.0010364478221163154, + 0.0011432073079049587, + 0.000981552293524146, + 0.017351530492305756, + 0.01957111991941929, + 0.012288125231862068, + 0.020200494676828384, + 0.018543556332588196, + 0.014171715825796127, + 0.010432244278490543, + 0.023287467658519745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.0036128980573266745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901579678058624 + ], + "opposite_logprob_tokens": [ + -2.4640157222747803 + ], + "logprob_reward": -0.08901579678058624, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901579678058624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18984287977218628, + 0.14799878001213074, + 0.12060348689556122, + 0.04747747629880905, + 0.23598262667655945, + 0.03681332617998123, + 0.032091666013002396, + 0.074092335999012, + 0.04416274279356003, + 0.07093464583158493 + ], + "credit_map": [ + 0.021660011261701584, + 0.012207468040287495, + 0.02764762192964554, + 0.03293028101325035, + 0.015163321979343891, + 0.03747444599866867, + 0.020423267036676407, + 0.022336479276418686, + 0.024898042902350426, + 0.041998159140348434, + 0.012256390415132046, + 0.015122361481189728, + 0.017107894644141197, + 0.016492491587996483, + 0.011152212508022785, + 0.00897122360765934, + 0.014851651154458523, + 0.02282632887363434, + 0.024856168776750565, + 0.014070791192352772, + 0.012722880579531193, + 0.013407919555902481, + 0.009307361207902431, + 0.008560380898416042, + 0.005653459578752518, + 0.006642007734626532, + 0.004286638926714659, + 0.0031953337602317333, + 0.008404118940234184, + 0.010862541384994984, + 0.004150775261223316, + 0.004282600246369839, + 0.03605659306049347, + 0.0494467057287693, + 0.01738334260880947, + 0.024303171783685684, + 0.027754196897149086, + 0.01957053132355213, + 0.026608876883983612, + 0.03485920652747154, + 0.003924628719687462, + 0.006287756841629744, + 0.005841037258505821, + 0.003266626037657261, + 0.003229184541851282, + 0.005806800909340382, + 0.004277578555047512, + 0.004179713316261768, + 0.0033332037273794413, + 0.003149866359308362, + 0.006289883051067591, + 0.003667422803118825, + 0.0035202400758862495, + 0.0034263269044458866, + 0.004403674043715, + 0.004301050212234259, + 0.007246844470500946, + 0.014386690221726894, + 0.011585522443056107, + 0.01014308724552393, + 0.006765875034034252, + 0.010788336396217346, + 0.006285769399255514, + 0.006890220567584038, + 0.004756981041282415, + 0.009134366177022457, + 0.004693819209933281, + 0.004500160459429026, + 0.006010937504470348, + 0.004614784847944975, + 0.006127731874585152, + 0.004323962144553661, + 0.013512593694031239, + 0.007843497209250927, + 0.009732309728860855, + 0.00981173850595951, + 0.006891022436320782, + 0.008804225362837315, + 0.006582613103091717, + 0.007756647188216448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.08901579678058624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, uninterrupted tracking shot follows a teenager on an extreme skateboard at breakneck speed, close to the ground. The teenager has shaggy black hair with a hint of silver-gray highlights, and his eyes are focused and piercing. He wears a loose black hoodie with abstract, non-figurative white graffiti printed on the back, dark khaki work pants, and a pair of well-worn professional skateboard shoes. The skateboard's deck is covered in colorful, wild street graffiti, and the colorful wheels are almost invisible at the high speed. He moves almost to the ground, deftly lowering his body as he speeds along the city's main thoroughfare at dusk. Every tilt and shift of his weight is filled with rhythm and explosive power. As he precisely controlled the edge of his skateboard, it slashed across the asphalt. The custom-made wheels rubbed against the ground, instantly releasing a dazzling trail of golden and orange sparks. These particles trailed a dazzling trail of light in front of the camera, illuminating the world beneath his feet. The background was a city scene at a magical hour. The distant skyline burned with a deep orange-purple gradient, and the silhouettes of the tall buildings were gilded with a golden edge. The newly lit neon signs and streetlights on both sides of the street cast highly saturated, slightly haloed beams of light, creating visible volumetric light in the dust-filled air and a gorgeous anamorphic lens flare as the camera moved. The entire scene was imbued with Ufotable's signature visual style: an incredibly smooth, high-frame-rate hand-drawn Sakuga-style animation, crisp lines, and gorgeous light and shadow seamlessly integrated with CG effects. It perfectly combined the youthful energy of a young man with the psychedelic atmosphere of an urban dusk, creating an adrenaline-pumping visual feast.", + "seed": 29620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_11/index_3526_seed_29620_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.037633609399199484, + "vlm_logprob": -0.037633609399199484, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.037633609399199484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04340463876724243 + ], + "opposite_logprob_tokens": [ + -3.1684045791625977 + ], + "logprob_reward": -0.04340463876724243, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04340463876724243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16910675168037415, + 0.12530842423439026, + 0.16038118302822113, + 0.032068587839603424, + 0.10648012906312943, + 0.07349059730768204, + 0.08338262885808945, + 0.02315281517803669, + 0.1507323682308197, + 0.07589658349752426 + ], + "credit_map": [ + 0.01391702238470316, + 0.020454125478863716, + 0.014811631292104721, + 0.01779644377529621, + 0.03643886744976044, + 0.020113937556743622, + 0.02852887474000454, + 0.01704585552215576, + 0.013626858592033386, + 0.018202360719442368, + 0.012191501446068287, + 0.012948889285326004, + 0.012963145039975643, + 0.014552977867424488, + 0.013450119644403458, + 0.0273725688457489, + 0.025137368589639664, + 0.010922632180154324, + 0.013376107439398766, + 0.059321507811546326, + 0.008670742623507977, + 0.0117806326597929, + 0.01868821121752262, + 0.012483981437981129, + 0.0025711827911436558, + 0.004047098103910685, + 0.0057175359688699245, + 0.002646945882588625, + 0.004097776021808386, + 0.003145538503304124, + 0.006289943121373653, + 0.003552566282451153, + 0.022612346336245537, + 0.00701378658413887, + 0.0043654232285916805, + 0.00521498778834939, + 0.03923608362674713, + 0.008004753850400448, + 0.013452315703034401, + 0.006580428685992956, + 0.01173372007906437, + 0.015118414536118507, + 0.0044686091132462025, + 0.007282890845090151, + 0.010046575218439102, + 0.003157124388962984, + 0.014256779104471207, + 0.0074264882132411, + 0.008588369935750961, + 0.014805360697209835, + 0.010996450670063496, + 0.006840842310339212, + 0.01412635575979948, + 0.006770614068955183, + 0.012459720484912395, + 0.008794913999736309, + 0.002514324150979519, + 0.0039040185511112213, + 0.0017506028525531292, + 0.0019447766244411469, + 0.0050105866976082325, + 0.0028418556321412325, + 0.002641105093061924, + 0.0025455448776483536, + 0.03449917212128639, + 0.005584558472037315, + 0.06218082085251808, + 0.014963284134864807, + 0.014630881138145924, + 0.007226133719086647, + 0.005449032410979271, + 0.006198483519256115, + 0.006995369680225849, + 0.004949753172695637, + 0.015598603524267673, + 0.007173105608671904, + 0.009750787168741226, + 0.01575053296983242, + 0.00836975034326315, + 0.007308680098503828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.04340463876724243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539356015622616 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539356015622616, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539356015622616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13689754903316498, + 0.07242907583713531, + 0.28065964579582214, + 0.022847212851047516, + 0.12738189101219177, + 0.011734936386346817, + 0.0192401185631752, + 0.05966419354081154, + 0.21411928534507751, + 0.05502607673406601 + ], + "credit_map": [ + 0.011528643779456615, + 0.021442808210849762, + 0.014765034429728985, + 0.014522567391395569, + 0.028374558314681053, + 0.020177289843559265, + 0.01551789976656437, + 0.010568746365606785, + 0.008347698487341404, + 0.011010460555553436, + 0.006553544197231531, + 0.011504278518259525, + 0.007223948370665312, + 0.00606391578912735, + 0.00823240727186203, + 0.013492819853127003, + 0.0401734784245491, + 0.020053010433912277, + 0.021842483431100845, + 0.10443460941314697, + 0.015049717389047146, + 0.01870126836001873, + 0.03262830525636673, + 0.027776775881648064, + 0.002856989623978734, + 0.003178018145263195, + 0.0037795768585056067, + 0.0020372685976326466, + 0.0033695341553539038, + 0.0024120225571095943, + 0.0033018493559211493, + 0.0019119534408673644, + 0.02569272555410862, + 0.009519902057945728, + 0.006413127761334181, + 0.0075964671559631824, + 0.045254990458488464, + 0.009487295523285866, + 0.015836531296372414, + 0.007580841891467571, + 0.0012821444543078542, + 0.0018442281289026141, + 0.0013583448017016053, + 0.0013930759159848094, + 0.0019222864648327231, + 0.0006263889954425395, + 0.001901215291582048, + 0.0014072527410462499, + 0.002895324258133769, + 0.002568948082625866, + 0.001792617724277079, + 0.0012680803192779422, + 0.0038298405706882477, + 0.0026320426259189844, + 0.002386626088991761, + 0.0018666384276002645, + 0.008742238394916058, + 0.007382688112556934, + 0.004821190144866705, + 0.005027386359870434, + 0.008192094974219799, + 0.012104885652661324, + 0.008142000995576382, + 0.005251709837466478, + 0.051688019186258316, + 0.009706499055027962, + 0.08367319405078888, + 0.018317338079214096, + 0.019024765118956566, + 0.012063256464898586, + 0.009544536471366882, + 0.0101016815751791, + 0.005265990737825632, + 0.005379044450819492, + 0.010543200187385082, + 0.00513284420594573, + 0.0064421482384204865, + 0.007225783541798592, + 0.008166243322193623, + 0.006870816461741924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.012539356015622616 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293431848287582 + ], + "opposite_logprob_tokens": [ + -3.7732934951782227 + ], + "logprob_reward": -0.023293431848287582, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293431848287582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10269233584403992, + 0.05011759698390961, + 0.25511640310287476, + 0.023401064798235893, + 0.25038695335388184, + 0.049089860171079636, + 0.026285655796527863, + 0.023634711280465126, + 0.15407109260559082, + 0.06520434468984604 + ], + "credit_map": [ + 0.0072197262197732925, + 0.012741787359118462, + 0.009641657583415508, + 0.010741268284618855, + 0.019147515296936035, + 0.014081567525863647, + 0.01849004253745079, + 0.01062876358628273, + 0.00479816785082221, + 0.005391278769820929, + 0.011914322152733803, + 0.005002728663384914, + 0.004071192350238562, + 0.006974923890084028, + 0.005635546520352364, + 0.00632943632081151, + 0.029490763321518898, + 0.017587490379810333, + 0.024431368336081505, + 0.09944332391023636, + 0.012258048169314861, + 0.017323048785328865, + 0.02855471335351467, + 0.02602764032781124, + 0.0032377305906265974, + 0.0036954774986952543, + 0.003766946727409959, + 0.0023623602464795113, + 0.0027524069882929325, + 0.002157412702217698, + 0.002918536076322198, + 0.00251019187271595, + 0.03875872492790222, + 0.024527205154299736, + 0.013816055841743946, + 0.020693132653832436, + 0.07592059671878815, + 0.018753357231616974, + 0.04153788089752197, + 0.016380002722144127, + 0.007401658222079277, + 0.009780011139810085, + 0.0031128423288464546, + 0.004678717348724604, + 0.006597507279366255, + 0.0020358855836093426, + 0.010896864347159863, + 0.004586376715451479, + 0.0044703492894768715, + 0.005012547597289085, + 0.0024904164019972086, + 0.0019640387035906315, + 0.005011647008359432, + 0.0024029721971601248, + 0.0033733185846358538, + 0.001560365897603333, + 0.003070492297410965, + 0.00248510017991066, + 0.0030562265310436487, + 0.0027996492572128773, + 0.0024501318112015724, + 0.003312141401693225, + 0.0024657605681568384, + 0.003995209001004696, + 0.034915611147880554, + 0.006212461739778519, + 0.06245107203722, + 0.016163378953933716, + 0.0150021743029356, + 0.007700814865529537, + 0.005774619523435831, + 0.005850964225828648, + 0.006996528711169958, + 0.003998137079179287, + 0.01472885999828577, + 0.005225321743637323, + 0.007122511509805918, + 0.01403674017637968, + 0.0072633386589586735, + 0.005832903552800417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.023293431848287582 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863614957779646 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863614957779646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863614957779646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09740756452083588, + 0.053199589252471924, + 0.22761037945747375, + 0.03426665440201759, + 0.2301957756280899, + 0.05544775724411011, + 0.08575909584760666, + 0.027849361300468445, + 0.08537697046995163, + 0.1028868779540062 + ], + "credit_map": [ + 0.011905323714017868, + 0.017347829416394234, + 0.01256381906569004, + 0.008430720306932926, + 0.015149296261370182, + 0.013794712722301483, + 0.009837761521339417, + 0.008378102444112301, + 0.004188735503703356, + 0.004821819718927145, + 0.00730133568868041, + 0.006431934889405966, + 0.0028522696811705828, + 0.009439167566597462, + 0.004509038757532835, + 0.013655286282300949, + 0.037595588713884354, + 0.015368793159723282, + 0.01924568973481655, + 0.08630254864692688, + 0.01110224612057209, + 0.01647225394845009, + 0.025426022708415985, + 0.016097236424684525, + 0.004633985925465822, + 0.006068490445613861, + 0.005023378878831863, + 0.003288764040917158, + 0.004678578115999699, + 0.0027378227096050978, + 0.00494755245745182, + 0.0028880825266242027, + 0.02641155757009983, + 0.026560116559267044, + 0.01807786524295807, + 0.027896782383322716, + 0.05511913821101189, + 0.018444746732711792, + 0.04219401255249977, + 0.015491537749767303, + 0.007255997974425554, + 0.010568485595285892, + 0.003222045488655567, + 0.005485664587467909, + 0.00798836164176464, + 0.0026552348863333464, + 0.014447052963078022, + 0.00382491247728467, + 0.01133978646248579, + 0.011578376404941082, + 0.007122997660189867, + 0.006544118747115135, + 0.01986459270119667, + 0.010820815339684486, + 0.011303797364234924, + 0.007184612099081278, + 0.0027246458921581507, + 0.004187228158116341, + 0.0025581892114132643, + 0.002378246281296015, + 0.0063430932350456715, + 0.003724338486790657, + 0.003441974287852645, + 0.0024916469119489193, + 0.015253767371177673, + 0.007103396113961935, + 0.029354924336075783, + 0.008830301463603973, + 0.00933659914880991, + 0.005275203846395016, + 0.005831257440149784, + 0.004391514230519533, + 0.010885042138397694, + 0.009478111751377583, + 0.01870415173470974, + 0.010829634964466095, + 0.017394723370671272, + 0.013072398491203785, + 0.010777362622320652, + 0.011745455674827099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.00863614957779646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029447078704834 + ], + "opposite_logprob_tokens": [ + -2.350294589996338 + ], + "logprob_reward": -0.10029447078704834, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029447078704834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11859719455242157, + 0.10949204117059708, + 0.06666899472475052, + 0.0619732104241848, + 0.1573227494955063, + 0.07842767238616943, + 0.05821545794606209, + 0.2028166651725769, + 0.07734723389148712, + 0.0691387876868248 + ], + "credit_map": [ + 0.015433664433658123, + 0.01073357556015253, + 0.01952487602829933, + 0.010824469849467278, + 0.019377952441573143, + 0.013936127535998821, + 0.016433894634246826, + 0.012332622893154621, + 0.009024779312312603, + 0.012346725910902023, + 0.03133418411016464, + 0.009140269830822945, + 0.008603481575846672, + 0.012537865899503231, + 0.00852988101541996, + 0.017974842339754105, + 0.010679789818823338, + 0.0049196709878742695, + 0.009523711167275906, + 0.005928189493715763, + 0.005167870316654444, + 0.008374601602554321, + 0.008198444731533527, + 0.01387671660631895, + 0.007678984198719263, + 0.013089428655803204, + 0.014614420011639595, + 0.005625794176012278, + 0.004925966262817383, + 0.004213400650769472, + 0.007619208190590143, + 0.0042060078121721745, + 0.026684975251555443, + 0.015697196125984192, + 0.011990944854915142, + 0.01755298674106598, + 0.023383723571896553, + 0.013696841895580292, + 0.03411431610584259, + 0.014201757498085499, + 0.01850084215402603, + 0.00956976879388094, + 0.005200832616537809, + 0.006801064591854811, + 0.00918857753276825, + 0.003607085905969143, + 0.010615699924528599, + 0.014943801797926426, + 0.007293717935681343, + 0.009137707762420177, + 0.00814728531986475, + 0.005105022806674242, + 0.008155270479619503, + 0.0069207982160151005, + 0.007043482735753059, + 0.006412175949662924, + 0.02531297691166401, + 0.02576218545436859, + 0.014047586359083652, + 0.015203441493213177, + 0.03397697955369949, + 0.04485628381371498, + 0.027993910014629364, + 0.015663303434848785, + 0.013483003713190556, + 0.004426295403391123, + 0.02543342299759388, + 0.010315594263374805, + 0.008581912145018578, + 0.00547992717474699, + 0.004253019578754902, + 0.0053740497678518295, + 0.00965468306094408, + 0.0066014439798891544, + 0.010403584688901901, + 0.006960084196180105, + 0.00860381405800581, + 0.011869133450090885, + 0.006115368567407131, + 0.00893067754805088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.10029447078704834 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, uninterrupted tracking shot follows a teenager on an extreme skateboard at breakneck speed, close to the ground. The teenager has shaggy black hair with a hint of silver-gray highlights, and his eyes are focused and piercing. He wears a loose black hoodie with abstract, non-figurative white graffiti printed on the back, dark khaki work pants, and a pair of well-worn professional skateboard shoes. The skateboard's deck is covered in colorful, wild street graffiti, and the colorful wheels are almost invisible at the high speed. He moves almost to the ground, deftly lowering his body as he speeds along the city's main thoroughfare at dusk. Every tilt and shift of his weight is filled with rhythm and explosive power. As he precisely controlled the edge of his skateboard, it slashed across the asphalt. The custom-made wheels rubbed against the ground, instantly releasing a dazzling trail of golden and orange sparks. These particles trailed a dazzling trail of light in front of the camera, illuminating the world beneath his feet. The background was a city scene at a magical hour. The distant skyline burned with a deep orange-purple gradient, and the silhouettes of the tall buildings were gilded with a golden edge. The newly lit neon signs and streetlights on both sides of the street cast highly saturated, slightly haloed beams of light, creating visible volumetric light in the dust-filled air and a gorgeous anamorphic lens flare as the camera moved. The entire scene was imbued with Ufotable's signature visual style: an incredibly smooth, high-frame-rate hand-drawn Sakuga-style animation, crisp lines, and gorgeous light and shadow seamlessly integrated with CG effects. It perfectly combined the youthful energy of a young man with the psychedelic atmosphere of an urban dusk, creating an adrenaline-pumping visual feast.", + "seed": 29621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_11/index_3526_seed_29621_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026408389024436475, + "vlm_logprob": -0.026408389024436475, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.026408389024436475, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027065236121416092 + ], + "opposite_logprob_tokens": [ + -3.6520652770996094 + ], + "logprob_reward": -0.027065236121416092, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027065236121416092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.105326347053051, + 0.04868531972169876, + 0.05245693773031235, + 0.01889237016439438, + 0.017789289355278015, + 0.286734402179718, + 0.03679903969168663, + 0.014881954528391361, + 0.38675251603126526, + 0.03168182447552681 + ], + "credit_map": [ + 0.013852165080606937, + 0.013784895651042461, + 0.010710012167692184, + 0.010784938000142574, + 0.01506122201681137, + 0.020424962043762207, + 0.008629565127193928, + 0.012078584171831608, + 0.0037204434629529715, + 0.0036811684258282185, + 0.006635670550167561, + 0.002416248433291912, + 0.00447018863633275, + 0.01297585666179657, + 0.001850046799518168, + 0.012935695238411427, + 0.007449793163686991, + 0.006228489801287651, + 0.007339464966207743, + 0.0037605613470077515, + 0.009194178506731987, + 0.006274004001170397, + 0.004836308304220438, + 0.007374133914709091, + 0.002523343078792095, + 0.0023858624044805765, + 0.002199941547587514, + 0.0018236865289509296, + 0.0021163667552173138, + 0.00264570745639503, + 0.002584828296676278, + 0.0026126333978027105, + 0.0015279158251360059, + 0.0023908005096018314, + 0.0015495490515604615, + 0.0016644919523969293, + 0.002602572087198496, + 0.002135535702109337, + 0.0016131451120600104, + 0.0043052788823843, + 0.042729295790195465, + 0.015144114382565022, + 0.017685482278466225, + 0.011665942147374153, + 0.01903708651661873, + 0.09928731620311737, + 0.011225826106965542, + 0.06995932757854462, + 0.0023932468611747026, + 0.0036949666682630777, + 0.007399782072752714, + 0.0055949571542441845, + 0.008114432916045189, + 0.00346097513101995, + 0.00314265675842762, + 0.0029980193357914686, + 0.0013942576479166746, + 0.0021483541931957006, + 0.0018104423070326447, + 0.0014400917571038008, + 0.0021315058693289757, + 0.0016338523710146546, + 0.0025359594728797674, + 0.0017874898621812463, + 0.022740021347999573, + 0.024352053180336952, + 0.05091174691915512, + 0.1432904750108719, + 0.009716554544866085, + 0.08933255076408386, + 0.022168291732668877, + 0.02424081228673458, + 0.0017034767661243677, + 0.002822635229676962, + 0.003025650978088379, + 0.004633639473468065, + 0.004057363141328096, + 0.002938950201496482, + 0.007682096213102341, + 0.004818010609596968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.027065236121416092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739504635334015 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739504635334015, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739504635334015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07783245295286179, + 0.062140386551618576, + 0.04612910747528076, + 0.026584453880786896, + 0.024585001170635223, + 0.13727296888828278, + 0.13418804109096527, + 0.025971807539463043, + 0.42120125889778137, + 0.044094543904066086 + ], + "credit_map": [ + 0.007769668009132147, + 0.006997639779001474, + 0.013446914032101631, + 0.009608399122953415, + 0.007780511863529682, + 0.01261474285274744, + 0.005690202582627535, + 0.0139243695884943, + 0.004385388456285, + 0.00419021537527442, + 0.009223545901477337, + 0.0030354682821780443, + 0.004933623131364584, + 0.01637144759297371, + 0.0024546796921640635, + 0.017546016722917557, + 0.006255130283534527, + 0.006816977635025978, + 0.008262199349701405, + 0.0037120901979506016, + 0.005321089643985033, + 0.005808534100651741, + 0.004129542037844658, + 0.005823545623570681, + 0.0040062121115624905, + 0.004359512124210596, + 0.0026693888939917088, + 0.0022085881792008877, + 0.004119350574910641, + 0.0038484418764710426, + 0.002900109626352787, + 0.002472850028425455, + 0.0024321204982697964, + 0.003858724841848016, + 0.002855550032109022, + 0.0023332382552325726, + 0.0034287371672689915, + 0.0025952288415282965, + 0.0033920358400791883, + 0.0036893654614686966, + 0.01375134289264679, + 0.006114038173109293, + 0.011003541760146618, + 0.006521480157971382, + 0.008762168698012829, + 0.040297213941812515, + 0.009146230295300484, + 0.04167696088552475, + 0.007431005593389273, + 0.007604660466313362, + 0.05267081782221794, + 0.008689581416547298, + 0.012549865059554577, + 0.006093485280871391, + 0.007354737725108862, + 0.03179388865828514, + 0.0026786255184561014, + 0.004223266616463661, + 0.00340454513207078, + 0.0033039997797459364, + 0.0026268847286701202, + 0.003600089577957988, + 0.00339401769451797, + 0.0027403784915804863, + 0.022605253383517265, + 0.024586524814367294, + 0.05975779518485069, + 0.1640869677066803, + 0.010777261108160019, + 0.0804334282875061, + 0.02867993526160717, + 0.03027411922812462, + 0.0027167610824108124, + 0.003262702375650406, + 0.0026979525573551655, + 0.00544700538739562, + 0.00651753693819046, + 0.004391727037727833, + 0.010953306220471859, + 0.00810755044221878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.006739504635334015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540768831968307 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540768831968307, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540768831968307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07116241008043289, + 0.026439018547534943, + 0.01632756181061268, + 0.021130360662937164, + 0.053628109395504, + 0.05557588115334511, + 0.13112494349479675, + 0.02808435633778572, + 0.5520215034484863, + 0.04450589045882225 + ], + "credit_map": [ + 0.009245045483112335, + 0.010688290931284428, + 0.008809036575257778, + 0.009025552310049534, + 0.010406643152236938, + 0.00937858410179615, + 0.004800376482307911, + 0.008808889426290989, + 0.0028149853460490704, + 0.0024339675437659025, + 0.0036410363391041756, + 0.002288440940901637, + 0.0037113570142537355, + 0.004380641970783472, + 0.001859135227277875, + 0.005309456493705511, + 0.0041835736483335495, + 0.0024605928920209408, + 0.0016701671993359923, + 0.0012572589330375195, + 0.0017223756294697523, + 0.00250639533624053, + 0.001209357869811356, + 0.0013178415829315782, + 0.0036497858818620443, + 0.0032938148360699415, + 0.002433012705296278, + 0.0016705618472769856, + 0.002847556723281741, + 0.0028359810821712017, + 0.0024010587949305773, + 0.0019985910039395094, + 0.004746390040963888, + 0.008074440062046051, + 0.005214930046349764, + 0.005214854143559933, + 0.006303541827946901, + 0.005862997844815254, + 0.005997359752655029, + 0.012213606387376785, + 0.007622642908245325, + 0.003486096393316984, + 0.0063291285187006, + 0.004958783742040396, + 0.00558489141985774, + 0.01449659839272499, + 0.0035506831482052803, + 0.009547058492898941, + 0.007587758358567953, + 0.0040648621506989, + 0.05284900963306427, + 0.006323463283479214, + 0.014750305563211441, + 0.007485155016183853, + 0.0062501076608896255, + 0.03181428834795952, + 0.0032764861825853586, + 0.004381359089165926, + 0.0025954830925911665, + 0.0026400212664157152, + 0.006737387273460627, + 0.0027234621811658144, + 0.002961728721857071, + 0.00276843155734241, + 0.031138628721237183, + 0.029562683776021004, + 0.09057282656431198, + 0.2230347841978073, + 0.012475556693971157, + 0.08567680418491364, + 0.03827221691608429, + 0.04128800705075264, + 0.002391861053183675, + 0.003465876914560795, + 0.0034048648085445166, + 0.005874297581613064, + 0.00792922917753458, + 0.005103078205138445, + 0.006898446008563042, + 0.009438239969313145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.012540768831968307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732991896569729 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732991896569729, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732991896569729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05933352932333946, + 0.0213658157736063, + 0.02948683314025402, + 0.0170188769698143, + 0.04382077232003212, + 0.11463562399148941, + 0.06127183139324188, + 0.017628291621804237, + 0.6143125891685486, + 0.021125858649611473 + ], + "credit_map": [ + 0.008415257558226585, + 0.007475785445421934, + 0.011202437803149223, + 0.007394476793706417, + 0.0055328733287751675, + 0.008152357302606106, + 0.004252331331372261, + 0.006908009294420481, + 0.0017332920106127858, + 0.0013149695005267859, + 0.002785817952826619, + 0.0009053298272192478, + 0.001509444904513657, + 0.006403484381735325, + 0.0008277615415863693, + 0.0058857137337327, + 0.004220864735543728, + 0.0044822352938354015, + 0.004139139782637358, + 0.0027854840736836195, + 0.0035910666920244694, + 0.00366516993381083, + 0.0025102603249251842, + 0.004092610441148281, + 0.004891295451670885, + 0.0021960833109915257, + 0.0018644537776708603, + 0.0010812283726409078, + 0.001605179626494646, + 0.0021341692190617323, + 0.0008525117882527411, + 0.0023939553648233414, + 0.004492517095059156, + 0.005610279273241758, + 0.006385135464370251, + 0.004287254065275192, + 0.004899540916085243, + 0.004613729193806648, + 0.0061919186264276505, + 0.007340394426137209, + 0.01450236514210701, + 0.00454756710678339, + 0.006612446624785662, + 0.006097347941249609, + 0.006552183069288731, + 0.041298314929008484, + 0.0043702987022697926, + 0.03065509907901287, + 0.003399905748665333, + 0.0027575339190661907, + 0.02456800267100334, + 0.0034151719883084297, + 0.0025164007674902678, + 0.0028141771908849478, + 0.0031717445235699415, + 0.018628893420100212, + 0.001602983451448381, + 0.001721897628158331, + 0.002539538312703371, + 0.0015451457584276795, + 0.0013029641704633832, + 0.004490931984037161, + 0.001674424041993916, + 0.002750406041741371, + 0.031257353723049164, + 0.032659780234098434, + 0.10193264484405518, + 0.24619713425636292, + 0.013573416508734226, + 0.10541155189275742, + 0.04047199711203575, + 0.04280872642993927, + 0.0012816410744562745, + 0.001964815426617861, + 0.001390615594573319, + 0.0027853320352733135, + 0.0028886664658784866, + 0.0029474180191755295, + 0.0027792411856353283, + 0.0050881290808320045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.006732991896569729 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896344363689423 + ], + "opposite_logprob_tokens": [ + -2.5789635181427 + ], + "logprob_reward": -0.07896344363689423, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896344363689423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17951467633247375, + 0.04628530517220497, + 0.04830477386713028, + 0.03759464994072914, + 0.038822464644908905, + 0.21776364743709564, + 0.19142372906208038, + 0.06596539914608002, + 0.09226875007152557, + 0.08205657452344894 + ], + "credit_map": [ + 0.018570903688669205, + 0.03123173676431179, + 0.019046558067202568, + 0.026867900043725967, + 0.021333908662199974, + 0.0241056177765131, + 0.016018623486161232, + 0.02233944460749626, + 0.0041066305711865425, + 0.00335977366194129, + 0.006604257505387068, + 0.002969788620248437, + 0.0038109691813588142, + 0.011436086148023605, + 0.0027847622986882925, + 0.01121304091066122, + 0.007409939542412758, + 0.007060104515403509, + 0.005081662442535162, + 0.005197558086365461, + 0.008768558502197266, + 0.006922587752342224, + 0.004270066041499376, + 0.0035943035036325455, + 0.010979505255818367, + 0.004887481685727835, + 0.004103288520127535, + 0.0017967801541090012, + 0.0031361207365989685, + 0.006523648742586374, + 0.0026285741478204727, + 0.00353925465606153, + 0.00467892549932003, + 0.0051682619377970695, + 0.004412476904690266, + 0.0029733171686530113, + 0.005672934930771589, + 0.003991311881691217, + 0.002936432370916009, + 0.008988809771835804, + 0.02887486293911934, + 0.011393017135560513, + 0.013512908481061459, + 0.012860593385994434, + 0.015133202075958252, + 0.07588627934455872, + 0.011811194941401482, + 0.048291608691215515, + 0.0170988030731678, + 0.02553238347172737, + 0.043501902371644974, + 0.018111996352672577, + 0.03647904098033905, + 0.02046840265393257, + 0.012390056625008583, + 0.01784118264913559, + 0.006011482328176498, + 0.0074448357336223125, + 0.009376001544296741, + 0.008007923141121864, + 0.007528883405029774, + 0.012244857847690582, + 0.005173135083168745, + 0.010178281925618649, + 0.008702820166945457, + 0.005497776437550783, + 0.012005497701466084, + 0.02690238133072853, + 0.0059673055075109005, + 0.0182466059923172, + 0.007205217145383358, + 0.007741148583590984, + 0.007990509271621704, + 0.013901016674935818, + 0.00584439467638731, + 0.00866291020065546, + 0.014366112649440765, + 0.011197682470083237, + 0.01079800445586443, + 0.009295952506363392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.07896344363689423 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, uninterrupted tracking shot follows a teenager on an extreme skateboard at breakneck speed, close to the ground. The teenager has shaggy black hair with a hint of silver-gray highlights, and his eyes are focused and piercing. He wears a loose black hoodie with abstract, non-figurative white graffiti printed on the back, dark khaki work pants, and a pair of well-worn professional skateboard shoes. The skateboard's deck is covered in colorful, wild street graffiti, and the colorful wheels are almost invisible at the high speed. He moves almost to the ground, deftly lowering his body as he speeds along the city's main thoroughfare at dusk. Every tilt and shift of his weight is filled with rhythm and explosive power. As he precisely controlled the edge of his skateboard, it slashed across the asphalt. The custom-made wheels rubbed against the ground, instantly releasing a dazzling trail of golden and orange sparks. These particles trailed a dazzling trail of light in front of the camera, illuminating the world beneath his feet. The background was a city scene at a magical hour. The distant skyline burned with a deep orange-purple gradient, and the silhouettes of the tall buildings were gilded with a golden edge. The newly lit neon signs and streetlights on both sides of the street cast highly saturated, slightly haloed beams of light, creating visible volumetric light in the dust-filled air and a gorgeous anamorphic lens flare as the camera moved. The entire scene was imbued with Ufotable's signature visual style: an incredibly smooth, high-frame-rate hand-drawn Sakuga-style animation, crisp lines, and gorgeous light and shadow seamlessly integrated with CG effects. It perfectly combined the youthful energy of a young man with the psychedelic atmosphere of an urban dusk, creating an adrenaline-pumping visual feast.", + "seed": 29622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_11/index_3526_seed_29622_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0301766955293715, + "vlm_logprob": -0.0301766955293715, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.0301766955293715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04344858601689339 + ], + "opposite_logprob_tokens": [ + -3.1684486865997314 + ], + "logprob_reward": -0.04344858601689339, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04344858601689339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1708158552646637, + 0.07095346599817276, + 0.3080594539642334, + 0.06421475112438202, + 0.023998204618692398, + 0.014310475438833237, + 0.08203034102916718, + 0.03611571341753006, + 0.13391463458538055, + 0.09558706730604172 + ], + "credit_map": [ + 0.012088458053767681, + 0.01580946333706379, + 0.020613493397831917, + 0.03165367990732193, + 0.029290052130818367, + 0.027439873665571213, + 0.019618839025497437, + 0.014302002266049385, + 0.0069517469964921474, + 0.006823893170803785, + 0.01151394471526146, + 0.01203560084104538, + 0.0045936391688883305, + 0.010868427343666553, + 0.011609167791903019, + 0.006557040382176638, + 0.011218409053981304, + 0.020938124507665634, + 0.040375836193561554, + 0.1206885576248169, + 0.010482336394488811, + 0.06785339117050171, + 0.02065027877688408, + 0.015852509066462517, + 0.005062614567577839, + 0.00434867525473237, + 0.011182653717696667, + 0.010546838864684105, + 0.009812168776988983, + 0.005429101642221212, + 0.008629435673356056, + 0.009203261695802212, + 0.002879152074456215, + 0.0024985065683722496, + 0.003213637275621295, + 0.005961184855550528, + 0.0013719698181375861, + 0.0019679302349686623, + 0.001420970307663083, + 0.00468485476449132, + 0.0011553809745237231, + 0.0011768469121307135, + 0.0015241100918501616, + 0.002158619463443756, + 0.0022194066550582647, + 0.0014378216583281755, + 0.0028298816177994013, + 0.001808407949283719, + 0.005355928558856249, + 0.0037655914202332497, + 0.03241347149014473, + 0.01432661060243845, + 0.003381256712600589, + 0.015282356180250645, + 0.0031402192544192076, + 0.004364908207207918, + 0.00468824990093708, + 0.004531136713922024, + 0.00267785950563848, + 0.006707014515995979, + 0.0036474827211350203, + 0.0035228438209742308, + 0.0049082254990935326, + 0.0054329014383256435, + 0.010914183221757412, + 0.014120465144515038, + 0.01639447547495365, + 0.01973831281065941, + 0.018605554476380348, + 0.021816842257976532, + 0.014683675952255726, + 0.017641130834817886, + 0.012735258787870407, + 0.009756578132510185, + 0.012157067656517029, + 0.011133444495499134, + 0.016875943168997765, + 0.015026405453681946, + 0.009321071207523346, + 0.008581296540796757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.04344858601689339 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076745577156544 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076745577156544, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076745577156544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11835747212171555, + 0.1787959337234497, + 0.24207624793052673, + 0.08841215074062347, + 0.05972779169678688, + 0.032542817294597626, + 0.08071006089448929, + 0.037958648055791855, + 0.08769367635250092, + 0.07372523099184036 + ], + "credit_map": [ + 0.012307267636060715, + 0.010608508251607418, + 0.00871227402240038, + 0.02172316424548626, + 0.013630297034978867, + 0.024822808802127838, + 0.011723559349775314, + 0.014829578809440136, + 0.016183605417609215, + 0.022567646577954292, + 0.01951296254992485, + 0.030320726335048676, + 0.016993168741464615, + 0.02751057781279087, + 0.029861049726605415, + 0.015846189111471176, + 0.009225892834365368, + 0.01726006530225277, + 0.030944693833589554, + 0.09104392677545547, + 0.010391202755272388, + 0.052977170795202255, + 0.01723674312233925, + 0.01299655344337225, + 0.00794731080532074, + 0.005677837412804365, + 0.01457262970507145, + 0.012124315835535526, + 0.014667726121842861, + 0.005044927820563316, + 0.017845353111624718, + 0.010532042011618614, + 0.003357290057465434, + 0.00292448652908206, + 0.010695518925786018, + 0.024356914684176445, + 0.0016940768109634519, + 0.004289080388844013, + 0.002357442630454898, + 0.010052978992462158, + 0.00282831653021276, + 0.001867678016424179, + 0.0036713972222059965, + 0.004316573962569237, + 0.0065299212001264095, + 0.003212417708709836, + 0.00627270108088851, + 0.0038438080810010433, + 0.007222265005111694, + 0.007955039851367474, + 0.02429294027388096, + 0.009409378282725811, + 0.007026145234704018, + 0.015116618946194649, + 0.005030571948736906, + 0.00465710274875164, + 0.0048462385311722755, + 0.005184818059206009, + 0.002830654848366976, + 0.007560045924037695, + 0.005893548484891653, + 0.0030537215061485767, + 0.0039128330536186695, + 0.004676784388720989, + 0.0097388019785285, + 0.011627987958490849, + 0.010577704757452011, + 0.013841264881193638, + 0.008990037254989147, + 0.013518290594220161, + 0.00758398650214076, + 0.011815591715276241, + 0.006266573444008827, + 0.007142677903175354, + 0.012013902887701988, + 0.014953408390283585, + 0.009488043375313282, + 0.004487800411880016, + 0.013784180395305157, + 0.005588637664914131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.011076745577156544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780935011804104 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780935011804104, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780935011804104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18218760192394257, + 0.05366582050919533, + 0.2604551613330841, + 0.05586167424917221, + 0.07106396555900574, + 0.04104004055261612, + 0.16109004616737366, + 0.027855798602104187, + 0.062428660690784454, + 0.08435128629207611 + ], + "credit_map": [ + 0.0181831456720829, + 0.01573375053703785, + 0.017129354178905487, + 0.03779695928096771, + 0.013740072026848793, + 0.038699448108673096, + 0.01846281811594963, + 0.022442055866122246, + 0.004682203754782677, + 0.0068618860095739365, + 0.0047387718223035336, + 0.00958187971264124, + 0.005485700909048319, + 0.007966519333422184, + 0.009028741158545017, + 0.005320118740200996, + 0.010615338571369648, + 0.02135559916496277, + 0.03501268848776817, + 0.09977122396230698, + 0.009756200015544891, + 0.052973803132772446, + 0.016566185280680656, + 0.014404111541807652, + 0.0053742206655442715, + 0.006829523481428623, + 0.0067472378723323345, + 0.006668575573712587, + 0.0061235311441123486, + 0.006311454810202122, + 0.00906636007130146, + 0.008740766905248165, + 0.0038562598638236523, + 0.0030329450964927673, + 0.009848657995462418, + 0.02832375280559063, + 0.002389154862612486, + 0.0061728754080832005, + 0.003339892253279686, + 0.014100424014031887, + 0.0031762083526700735, + 0.0020678513683378696, + 0.004966843407601118, + 0.007038798648864031, + 0.005971342790871859, + 0.005385947413742542, + 0.006081697065383196, + 0.0063513522036373615, + 0.008436503820121288, + 0.0041310726664960384, + 0.06626094877719879, + 0.025916341692209244, + 0.0061687203124165535, + 0.034651752561330795, + 0.00567563297227025, + 0.009849083609879017, + 0.0045619080774486065, + 0.0049703000113368034, + 0.0028195136692374945, + 0.003788461210206151, + 0.0031597439665347338, + 0.0025395292323082685, + 0.0024098227731883526, + 0.0036065198946744204, + 0.004856710322201252, + 0.005941097158938646, + 0.00852323416620493, + 0.008955222554504871, + 0.006050250492990017, + 0.01296655461192131, + 0.006712368689477444, + 0.008423218503594398, + 0.011357947252690792, + 0.009700778871774673, + 0.009963377378880978, + 0.012180005200207233, + 0.011542950756847858, + 0.006491689942777157, + 0.016384562477469444, + 0.006729970686137676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.009780935011804104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622797507792711 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622797507792711, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622797507792711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21726259589195251, + 0.04004398733377457, + 0.14536237716674805, + 0.03125211223959923, + 0.023525049909949303, + 0.02396438457071781, + 0.16326338052749634, + 0.01849455013871193, + 0.2668398320674896, + 0.06999180465936661 + ], + "credit_map": [ + 0.019141266122460365, + 0.0133919483050704, + 0.014308609999716282, + 0.02922116592526436, + 0.04194799065589905, + 0.05162467435002327, + 0.02630029059946537, + 0.021326623857021332, + 0.00586363859474659, + 0.0038372264243662357, + 0.007642013486474752, + 0.005468148738145828, + 0.0024548773653805256, + 0.006060990039259195, + 0.0048279971815645695, + 0.003889093641191721, + 0.00571470009163022, + 0.008365892805159092, + 0.01760939136147499, + 0.05953313410282135, + 0.002981491619721055, + 0.03369815647602081, + 0.009813525713980198, + 0.007646074518561363, + 0.004950945731252432, + 0.0020536992233246565, + 0.004342405125498772, + 0.0051242755725979805, + 0.0026228937786072493, + 0.0026390203274786472, + 0.005032994318753481, + 0.00448587303981185, + 0.0024916790425777435, + 0.0015315011842176318, + 0.0037836944684386253, + 0.005654881242662668, + 0.0026771309785544872, + 0.0032182077411562204, + 0.0018615761073306203, + 0.002306377049535513, + 0.0024486184120178223, + 0.0020299383904784918, + 0.0026173775549978018, + 0.0037527272943407297, + 0.003055713139474392, + 0.0022237496450543404, + 0.004078385885804892, + 0.0037578728515654802, + 0.009123953990638256, + 0.005020223557949066, + 0.06784351915121078, + 0.028027597814798355, + 0.0052198427729308605, + 0.033723797649145126, + 0.005876951385289431, + 0.008427482098340988, + 0.002089433604851365, + 0.0015728678554296494, + 0.0014815570320934057, + 0.004118185490369797, + 0.001504358951933682, + 0.001266118953935802, + 0.0015033555682748556, + 0.004958673380315304, + 0.024293774738907814, + 0.029717545956373215, + 0.033593788743019104, + 0.04823840782046318, + 0.031268395483493805, + 0.03982006385922432, + 0.02348196506500244, + 0.03642585873603821, + 0.008270593360066414, + 0.006466211285442114, + 0.009761792607605457, + 0.009727745316922665, + 0.010765221901237965, + 0.011461523361504078, + 0.00694273691624403, + 0.00659597385674715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.007622797507792711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenager with shaggy black hair featuring silver-gray highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teenager wearing a loose black hoodie with abstract white graffiti on the back and dark khaki work pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the skateboard have a deck covered in colorful wild street graffiti and wheels that are nearly invisible due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skateboard's wheels rub against the asphalt to create a trail of golden and orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a Ufotable-style animation with crisp lines, high frame rate, and anamorphic lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895441353321075 + ], + "opposite_logprob_tokens": [ + -2.5789544582366943 + ], + "logprob_reward": -0.07895441353321075, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895441353321075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18445782363414764, + 0.07086960971355438, + 0.05364585295319557, + 0.03410593047738075, + 0.0352349616587162, + 0.05138202756643295, + 0.13009324669837952, + 0.03074951469898224, + 0.3703800439834595, + 0.03908100351691246 + ], + "credit_map": [ + 0.019226012751460075, + 0.02241923101246357, + 0.01267941389232874, + 0.021095747128129005, + 0.03286679461598396, + 0.03927105665206909, + 0.02037472277879715, + 0.016524845734238625, + 0.00802885927259922, + 0.0072403461672365665, + 0.014444626867771149, + 0.010148761793971062, + 0.004421819467097521, + 0.010373832657933235, + 0.010262581519782543, + 0.005948776379227638, + 0.0038066988345235586, + 0.004904760979115963, + 0.00720282644033432, + 0.01796351745724678, + 0.003085925942286849, + 0.008978557772934437, + 0.004528691526502371, + 0.0031748705077916384, + 0.005551124922931194, + 0.0035485895350575447, + 0.0042460025288164616, + 0.004005918744951487, + 0.0033811181783676147, + 0.0026384678203612566, + 0.005195118021219969, + 0.0055395918898284435, + 0.0032165837474167347, + 0.0026447030249983072, + 0.003457340644672513, + 0.009276416152715683, + 0.0034252020996063948, + 0.004162635188549757, + 0.0030571373645216227, + 0.00599494157359004, + 0.0050620706751942635, + 0.004313729703426361, + 0.006323878187686205, + 0.007971690036356449, + 0.00807279534637928, + 0.005529351066797972, + 0.008399027399718761, + 0.005709486547857523, + 0.008236626163125038, + 0.004050479736179113, + 0.051568035036325455, + 0.021464210003614426, + 0.006097303703427315, + 0.026551969349384308, + 0.004679148551076651, + 0.0074454667046666145, + 0.004135738592594862, + 0.0037733197677880526, + 0.003613302018493414, + 0.005262089893221855, + 0.003938488662242889, + 0.0023596591781824827, + 0.00296175223775208, + 0.004705164581537247, + 0.035585980862379074, + 0.04460323601961136, + 0.042326439172029495, + 0.050186313688755035, + 0.05525200068950653, + 0.055352695286273956, + 0.042234260588884354, + 0.04483910650014877, + 0.006556440144777298, + 0.004143116995692253, + 0.005128286778926849, + 0.005542002618312836, + 0.006429446395486593, + 0.0024849767796695232, + 0.00602256553247571, + 0.002774166176095605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.07895441353321075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_12/index_tensor([4911, 4911])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_12/index_tensor([4911, 4911])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..17c71d8b6d31173c55873e308f41828cd81bc645 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_12/index_tensor([4911, 4911])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a burly man, his short, dark hair drenched in sweat, standing in the center of a brightly lit, modern gym. He wears a charcoal-gray, sweat-wicking tank top and black gym shorts. Initially, he rhythmically sways his body from side to side, gesturing vigorously with his hands, sometimes with open palms, sometimes with pointed fingers to emphasize his words, speaking with energy and passion. The background is filled with rows of dumbbell racks and weight machines, slightly blurred to keep the focus on him. The atmosphere is vibrant and the conversation intense. Then, his movements pause, his entire demeanor shifts; he takes a deep breath, his attention turned inward. Then, steadily and deliberately, he bends forward, placing his hands on his knees for support, and slowly and reverently lowers his head toward the loaded dumbbells placed on the squat rack directly in front of him. Cool fluorescent lights overhead gleam on the metal equipment, creating an atmosphere of austere, focused attention, reflecting both the physical challenge and the mental preparation to come.", + "seed": 29857, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_12/index_4911_seed_29857_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03429847452789545, + "vlm_logprob": -0.03429847452789545, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03429847452789545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019299698993563652 + ], + "opposite_logprob_tokens": [ + -5.019299507141113 + ], + "logprob_reward": -0.019299698993563652, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.019299698993563652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0890069231390953, + 0.03784472867846489, + 0.026194186881184578, + 0.1468125432729721, + 0.04965381696820259, + 0.11934290081262589, + 0.043163299560546875, + 0.0774516612291336, + 0.06742987781763077, + 0.34310004115104675 + ], + "credit_map": [ + 0.00818962138146162, + 0.010005229152739048, + 0.011847198940813541, + 0.015377736650407314, + 0.012472613714635372, + 0.011420031078159809, + 0.011107145808637142, + 0.008587336167693138, + 0.0038898037746548653, + 0.003794895950704813, + 0.005032561253756285, + 0.007445828057825565, + 0.006061897147446871, + 0.003051650244742632, + 0.003965961746871471, + 0.004602127708494663, + 0.0035366532392799854, + 0.002495870226994157, + 0.0033512276131659746, + 0.0035135610960423946, + 0.004292936995625496, + 0.0028771753422915936, + 0.0032768617384135723, + 0.0028498994652181864, + 0.015997055917978287, + 0.009154241532087326, + 0.023064512759447098, + 0.019455010071396828, + 0.018840858712792397, + 0.013214413076639175, + 0.03572835400700569, + 0.011358099989593029, + 0.006503334268927574, + 0.008497238159179688, + 0.008067338727414608, + 0.006488960701972246, + 0.005091472528874874, + 0.005242058541625738, + 0.004968325607478619, + 0.004795089364051819, + 0.0088085001334548, + 0.008825467899441719, + 0.010282623581588268, + 0.006197014357894659, + 0.04127998650074005, + 0.028066514059901237, + 0.008398330770432949, + 0.007484460715204477, + 0.0038489336147904396, + 0.0036653028801083565, + 0.004381829407066107, + 0.006218534428626299, + 0.00863545574247837, + 0.004752475768327713, + 0.007916558533906937, + 0.0037442052271217108, + 0.008442078717052937, + 0.0054995352402329445, + 0.010297369211912155, + 0.006705875508487225, + 0.01068867091089487, + 0.007089488673955202, + 0.020458431914448738, + 0.008270206861197948, + 0.012609771452844143, + 0.006033946294337511, + 0.006424476858228445, + 0.004271283280104399, + 0.009622803889214993, + 0.01099435705691576, + 0.00917566567659378, + 0.008297573775053024, + 0.07254364341497421, + 0.0315997451543808, + 0.06656447052955627, + 0.01887742429971695, + 0.04262157529592514, + 0.04332086816430092, + 0.03578246384859085, + 0.03178985416889191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.019299698993563652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012555013410747051 + ], + "opposite_logprob_tokens": [ + -4.387555122375488 + ], + "logprob_reward": -0.012555013410747051, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012555013410747051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39505282044410706, + 0.022952066734433174, + 0.0242872703820467, + 0.053484220057725906, + 0.026731258258223534, + 0.1057736873626709, + 0.049697622656822205, + 0.0879335030913353, + 0.1271509975194931, + 0.10693658143281937 + ], + "credit_map": [ + 0.028618469834327698, + 0.047360822558403015, + 0.056140560656785965, + 0.0506291538476944, + 0.042751289904117584, + 0.08071278035640717, + 0.0407198965549469, + 0.04811982065439224, + 0.0023026359267532825, + 0.0024549763184040785, + 0.002722906181588769, + 0.0046266526915133, + 0.0031244519632309675, + 0.0022651255130767822, + 0.0025802659802138805, + 0.002875052159652114, + 0.00282334559597075, + 0.002076784847304225, + 0.0044474294409155846, + 0.003298054216429591, + 0.002521710703149438, + 0.002269031945616007, + 0.004431603010743856, + 0.0024193101562559605, + 0.004885250236839056, + 0.0029620015993714333, + 0.00859961286187172, + 0.00904887542128563, + 0.007032592780888081, + 0.0032195295207202435, + 0.012462741695344448, + 0.0052736178040504456, + 0.0022797256242483854, + 0.0023155445232987404, + 0.005752844270318747, + 0.0024249120615422726, + 0.0025282008573412895, + 0.004267851822078228, + 0.0031523487996309996, + 0.004009830299764872, + 0.00973859615623951, + 0.009828523732721806, + 0.021489128470420837, + 0.010612478479743004, + 0.026339799165725708, + 0.00788899790495634, + 0.009351793676614761, + 0.010524366982281208, + 0.004128220956772566, + 0.005160951055586338, + 0.008039242587983608, + 0.005650158040225506, + 0.007445382419973612, + 0.00864986889064312, + 0.0054105245508253574, + 0.005213271360844374, + 0.00982747320085764, + 0.01163190696388483, + 0.01305666845291853, + 0.0075837853364646435, + 0.008722458966076374, + 0.015003613196313381, + 0.014192376285791397, + 0.007915221154689789, + 0.019106443971395493, + 0.00672573409974575, + 0.036831945180892944, + 0.007194698788225651, + 0.010293246246874332, + 0.010910293087363243, + 0.02719888649880886, + 0.008889744058251381, + 0.022329168394207954, + 0.009847022593021393, + 0.0225236713886261, + 0.006022180430591106, + 0.012374159879982471, + 0.012711421586573124, + 0.011067814193665981, + 0.010061153210699558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.012555013410747051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028211576864123344 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028211576864123344, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028211576864123344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4654083251953125, + 0.03771909326314926, + 0.04053338244557381, + 0.036786530166864395, + 0.02979779988527298, + 0.09647475928068161, + 0.05559392273426056, + 0.04699545353651047, + 0.09294407814741135, + 0.09774661064147949 + ], + "credit_map": [ + 0.040268946439027786, + 0.05542908236384392, + 0.069316565990448, + 0.0595284067094326, + 0.043451834470033646, + 0.0927545353770256, + 0.04870555177330971, + 0.055953435599803925, + 0.004146122839301825, + 0.0034501070622354746, + 0.0045042880810797215, + 0.008932383731007576, + 0.005346611142158508, + 0.002551325364038348, + 0.004466735292226076, + 0.004321520682424307, + 0.0033383553382009268, + 0.003917645663022995, + 0.008358202874660492, + 0.0036107676569372416, + 0.004631437361240387, + 0.004117249045521021, + 0.00897835474461317, + 0.0035813713911920786, + 0.0035934385377913713, + 0.0033216639421880245, + 0.004657215904444456, + 0.0058989375829696655, + 0.005320739466696978, + 0.004055732861161232, + 0.006025693379342556, + 0.003913112450391054, + 0.00360822188667953, + 0.005162084475159645, + 0.005530386231839657, + 0.004120315425097942, + 0.003591979155316949, + 0.0022388407960534096, + 0.002852047560736537, + 0.0026939252857118845, + 0.007955687120556831, + 0.007935302332043648, + 0.005417889449745417, + 0.0056032645516097546, + 0.03584767132997513, + 0.01987546868622303, + 0.0066091520711779594, + 0.007230323739349842, + 0.0045327916741371155, + 0.0052535561844706535, + 0.008604047819972038, + 0.0078103626146912575, + 0.007099106442183256, + 0.008923402056097984, + 0.00810831319540739, + 0.005262347869575024, + 0.009662199765443802, + 0.004257093649357557, + 0.005612116772681475, + 0.003806802211329341, + 0.006802460644394159, + 0.0037945157382637262, + 0.008347601629793644, + 0.004712667316198349, + 0.011661341413855553, + 0.00495185237377882, + 0.026776179671287537, + 0.004463111981749535, + 0.007807278074324131, + 0.0068700434640049934, + 0.024643022567033768, + 0.005771251395344734, + 0.02097744680941105, + 0.009983144700527191, + 0.01980912685394287, + 0.006606520619243383, + 0.011971939355134964, + 0.009613982401788235, + 0.009779038839042187, + 0.009005425497889519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0028211576864123344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779990650713444 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779990650713444, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779990650713444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11462913453578949, + 0.030604980885982513, + 0.04003511369228363, + 0.051780037581920624, + 0.04098972678184509, + 0.12446077167987823, + 0.05291128531098366, + 0.1433020532131195, + 0.33617842197418213, + 0.06510847061872482 + ], + "credit_map": [ + 0.009074214845895767, + 0.012361284345388412, + 0.01758612133562565, + 0.01329322624951601, + 0.011969106271862984, + 0.024580560624599457, + 0.012852326966822147, + 0.01291228923946619, + 0.0030876356177031994, + 0.0026768671814352274, + 0.004150751046836376, + 0.005064941477030516, + 0.0053382436744868755, + 0.002728509483858943, + 0.003969488199800253, + 0.003588541178032756, + 0.003857027506455779, + 0.004263574257493019, + 0.008409203961491585, + 0.004423285368829966, + 0.0051371511071920395, + 0.004664701409637928, + 0.004958289209753275, + 0.004321881104260683, + 0.004758766386657953, + 0.003818679368123412, + 0.006388009525835514, + 0.007997458800673485, + 0.00626276945695281, + 0.0034390240907669067, + 0.014945872128009796, + 0.004169457592070103, + 0.004532286431640387, + 0.005202415864914656, + 0.0076830144971609116, + 0.004352536518126726, + 0.0039595020934939384, + 0.00587889039888978, + 0.0036367662250995636, + 0.005744316149502993, + 0.01011610496789217, + 0.011386679485440254, + 0.008902102708816528, + 0.008658125065267086, + 0.04504428058862686, + 0.021336546167731285, + 0.009393750689923763, + 0.009623173624277115, + 0.0032942513935267925, + 0.005505773238837719, + 0.006664727348834276, + 0.010145355015993118, + 0.006589926313608885, + 0.005407056305557489, + 0.010223855264484882, + 0.005080340895801783, + 0.01420816220343113, + 0.013585972599685192, + 0.014675059355795383, + 0.013045080937445164, + 0.019323766231536865, + 0.01602078415453434, + 0.03536892309784889, + 0.01707429811358452, + 0.04316103830933571, + 0.014312869869172573, + 0.1035575270652771, + 0.018343115225434303, + 0.023489443585276604, + 0.014018125832080841, + 0.1004633679986, + 0.018832961097359657, + 0.011122670024633408, + 0.004807596560567617, + 0.013039317913353443, + 0.003695264458656311, + 0.00625584926456213, + 0.008799194358289242, + 0.007686862256377935, + 0.009701722301542759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.009779990650713444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703651189804077 + ], + "opposite_logprob_tokens": [ + -2.1270365715026855 + ], + "logprob_reward": -0.12703651189804077, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703651189804077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16068805754184723, + 0.02458294667303562, + 0.02326645702123642, + 0.04627928137779236, + 0.0509934239089489, + 0.06716194748878479, + 0.07083019614219666, + 0.09426648169755936, + 0.1819000244140625, + 0.2800312340259552 + ], + "credit_map": [ + 0.011151290498673916, + 0.019421780481934547, + 0.018608223646879196, + 0.026297640055418015, + 0.019811632111668587, + 0.028636248782277107, + 0.01625816896557808, + 0.02050308883190155, + 0.002594665391370654, + 0.0022292586509138346, + 0.0035713843535631895, + 0.00435457332059741, + 0.003927563317120075, + 0.002032645046710968, + 0.0028670162428170443, + 0.0030058398842811584, + 0.002533844206482172, + 0.0028425271157175303, + 0.003912222571671009, + 0.0038071204908192158, + 0.0025435681454837322, + 0.0028692588675767183, + 0.0022230707108974457, + 0.0025348470080643892, + 0.004153495654463768, + 0.003479504957795143, + 0.006804234813898802, + 0.007638262584805489, + 0.004662681836634874, + 0.00454377569258213, + 0.011306059546768665, + 0.0036912700161337852, + 0.004671494942158461, + 0.006179773714393377, + 0.006012295372784138, + 0.007435893174260855, + 0.005606982856988907, + 0.006335249170660973, + 0.009093496017158031, + 0.005658242851495743, + 0.006668944843113422, + 0.0078021325170993805, + 0.014689932577311993, + 0.007568371016532183, + 0.003543829545378685, + 0.010301847942173481, + 0.008073834702372551, + 0.008513056673109531, + 0.004447244107723236, + 0.006770622450858355, + 0.009135005995631218, + 0.010243384167551994, + 0.009550369344651699, + 0.010373279452323914, + 0.013558106496930122, + 0.006752174347639084, + 0.020977500826120377, + 0.013097919523715973, + 0.015529275871813297, + 0.005844215862452984, + 0.010450690053403378, + 0.009565310552716255, + 0.012509903870522976, + 0.00629165954887867, + 0.02525094524025917, + 0.008778693154454231, + 0.04676900431513786, + 0.010481154546141624, + 0.01500395406037569, + 0.013454465195536613, + 0.05188467726111412, + 0.010277128778398037, + 0.06540574133396149, + 0.02710404060781002, + 0.05933684855699539, + 0.016353491693735123, + 0.029840441420674324, + 0.031177738681435585, + 0.0273332167416811, + 0.023479701951146126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.12703651189804077 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a burly man, his short, dark hair drenched in sweat, standing in the center of a brightly lit, modern gym. He wears a charcoal-gray, sweat-wicking tank top and black gym shorts. Initially, he rhythmically sways his body from side to side, gesturing vigorously with his hands, sometimes with open palms, sometimes with pointed fingers to emphasize his words, speaking with energy and passion. The background is filled with rows of dumbbell racks and weight machines, slightly blurred to keep the focus on him. The atmosphere is vibrant and the conversation intense. Then, his movements pause, his entire demeanor shifts; he takes a deep breath, his attention turned inward. Then, steadily and deliberately, he bends forward, placing his hands on his knees for support, and slowly and reverently lowers his head toward the loaded dumbbells placed on the squat rack directly in front of him. Cool fluorescent lights overhead gleam on the metal equipment, creating an atmosphere of austere, focused attention, reflecting both the physical challenge and the mental preparation to come.", + "seed": 29858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_12/index_4911_seed_29858_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03315718909725547, + "vlm_logprob": -0.03315718909725547, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.03315718909725547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01652694121003151 + ], + "opposite_logprob_tokens": [ + -5.01652717590332 + ], + "logprob_reward": -0.01652694121003151, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01652694121003151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05816246569156647, + 0.01944713294506073, + 0.1491687297821045, + 0.07608810812234879, + 0.02347872033715248, + 0.016473544761538506, + 0.05419153347611427, + 0.10592789202928543, + 0.02673175558447838, + 0.4703300893306732 + ], + "credit_map": [ + 0.005286240018904209, + 0.004317165818065405, + 0.0069802007637917995, + 0.009430228732526302, + 0.009287383407354355, + 0.007393352221697569, + 0.007591687608510256, + 0.007876202464103699, + 0.0015414010267704725, + 0.0018772301264107227, + 0.0020062876865267754, + 0.0021807742305099964, + 0.0037890577223151922, + 0.002369059482589364, + 0.0040819160640239716, + 0.0016014042776077986, + 0.005730421282351017, + 0.005246289540082216, + 0.004418116994202137, + 0.05899466946721077, + 0.007600696291774511, + 0.005555263254791498, + 0.05137845501303673, + 0.010244820266962051, + 0.00571692269295454, + 0.006690105423331261, + 0.01217871904373169, + 0.015730850398540497, + 0.010321707464754581, + 0.008039961569011211, + 0.009281857870519161, + 0.008127976208925247, + 0.0015281523810699582, + 0.0022111190482974052, + 0.0036205018404871225, + 0.0033665699884295464, + 0.003832611721009016, + 0.0024563532788306475, + 0.004453705623745918, + 0.002009707735851407, + 0.0011327523970976472, + 0.0017002903623506427, + 0.0017645570915192366, + 0.00335810799151659, + 0.0023602242581546307, + 0.00129670649766922, + 0.003249208442866802, + 0.001611696439795196, + 0.0046379962004721165, + 0.0058484626933932304, + 0.007583898957818747, + 0.008090943098068237, + 0.006437257397919893, + 0.0055151935666799545, + 0.00962507352232933, + 0.0064527071081101894, + 0.008849899284541607, + 0.007228045258671045, + 0.031133409589529037, + 0.007284937426447868, + 0.0070709409192204475, + 0.005731835030019283, + 0.008608420379459858, + 0.03002040460705757, + 0.0028459238819777966, + 0.0036316122859716415, + 0.003545326879248023, + 0.004161279648542404, + 0.0034067046362906694, + 0.003373574698343873, + 0.0030808867886662483, + 0.0026864472310990095, + 0.12577597796916962, + 0.05127737298607826, + 0.09466847032308578, + 0.03261398896574974, + 0.040532588958740234, + 0.03309115767478943, + 0.0483754463493824, + 0.04399508610367775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.01652694121003151 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656240999698639 + ], + "opposite_logprob_tokens": [ + -4.758656024932861 + ], + "logprob_reward": -0.008656240999698639, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656240999698639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030880024656653404, + 0.018961593508720398, + 0.6028702259063721, + 0.03196071460843086, + 0.009423512034118176, + 0.015771474689245224, + 0.04686540737748146, + 0.03264491632580757, + 0.03386157006025314, + 0.17676059901714325 + ], + "credit_map": [ + 0.0029410244897007942, + 0.003335678717121482, + 0.0031919078901410103, + 0.0050253476947546005, + 0.004203952383249998, + 0.004093943163752556, + 0.004777118097990751, + 0.003311051754280925, + 0.0021476061083376408, + 0.0017593511147424579, + 0.003914759028702974, + 0.00293532176874578, + 0.001930366619490087, + 0.002001559128984809, + 0.0023467394057661295, + 0.0019258891697973013, + 0.02320520021021366, + 0.019963791593909264, + 0.017415061593055725, + 0.2387576848268509, + 0.02834971994161606, + 0.024682436138391495, + 0.20975685119628906, + 0.04073953628540039, + 0.0022143826354295015, + 0.00291829789057374, + 0.004591826349496841, + 0.008955529890954494, + 0.0035394870210438967, + 0.0029031995218247175, + 0.004166360013186932, + 0.002671630820259452, + 0.0008240483002737164, + 0.0007225899607874453, + 0.0009429294150322676, + 0.0018008205806836486, + 0.0013523747911676764, + 0.0008590874494984746, + 0.002007466508075595, + 0.0009141957270912826, + 0.0013745834585279226, + 0.0019255621591582894, + 0.002048320835456252, + 0.0017546125454828143, + 0.0018772648181766272, + 0.0015723679680377245, + 0.0038104031700640917, + 0.0014083593850955367, + 0.003934298176318407, + 0.005201889201998711, + 0.00581246055662632, + 0.007922300137579441, + 0.005759652704000473, + 0.004650921560823917, + 0.007973004132509232, + 0.0056108771823346615, + 0.0029387411195784807, + 0.002343610394746065, + 0.008689885959029198, + 0.0031409701332449913, + 0.002464534714818001, + 0.0019830313976854086, + 0.003011516062542796, + 0.0080726258456707, + 0.003393813967704773, + 0.006609064061194658, + 0.005209433846175671, + 0.0035239376593381166, + 0.00614690687507391, + 0.002404723549261689, + 0.0032084693666547537, + 0.0033652205020189285, + 0.047098852694034576, + 0.017230110242962837, + 0.03186149150133133, + 0.011470898054540157, + 0.015009336173534393, + 0.013543237000703812, + 0.02230178378522396, + 0.018244881182909012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.008656240999698639 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024911104701459408 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.0024911104701459408, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024911104701459408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035806506872177124, + 0.03016888163983822, + 0.08870776742696762, + 0.0215263944119215, + 0.02793850190937519, + 0.0200412105768919, + 0.05944596230983734, + 0.2093961387872696, + 0.026120468974113464, + 0.48084816336631775 + ], + "credit_map": [ + 0.0032548604067415, + 0.003023674478754401, + 0.006362854037433863, + 0.004465766716748476, + 0.0037167000118643045, + 0.0038759938906878233, + 0.006657585501670837, + 0.00444907508790493, + 0.0025082645006477833, + 0.002737317932769656, + 0.004987755324691534, + 0.005470045376569033, + 0.002726570935919881, + 0.0037986021488904953, + 0.004745346959680319, + 0.003194981487467885, + 0.003442948218435049, + 0.0028229630552232265, + 0.0029399439226835966, + 0.0345308817923069, + 0.004563448019325733, + 0.0039062127470970154, + 0.030695371329784393, + 0.00580600555986166, + 0.0016990917501971126, + 0.0020470304880291224, + 0.002537098480388522, + 0.006409021094441414, + 0.0023177610710263252, + 0.0016694811638444662, + 0.0031218340154737234, + 0.0017250757664442062, + 0.0025183872785419226, + 0.0025046642404049635, + 0.0039465040899813175, + 0.0043214610777795315, + 0.004125718958675861, + 0.0028977415058761835, + 0.005171599332243204, + 0.0024524270556867123, + 0.001857686205767095, + 0.0026201053988188505, + 0.0027455631643533707, + 0.004378249868750572, + 0.002365285763517022, + 0.0021901752334088087, + 0.002336219884455204, + 0.001547928201034665, + 0.005348577164113522, + 0.007358446251600981, + 0.006689223926514387, + 0.009500772692263126, + 0.00612484198063612, + 0.006639996077865362, + 0.00918773002922535, + 0.00859637837857008, + 0.018169382587075233, + 0.010187572799623013, + 0.06699736416339874, + 0.011833541095256805, + 0.011093185283243656, + 0.010631265118718147, + 0.0114168431609869, + 0.06906700134277344, + 0.003059311071410775, + 0.0030221748165786266, + 0.005725755821913481, + 0.0020627095364034176, + 0.003829791909083724, + 0.002612929092720151, + 0.0024071873631328344, + 0.0034006121568381786, + 0.1660768687725067, + 0.0490645132958889, + 0.10744577646255493, + 0.027035431936383247, + 0.03234070912003517, + 0.024877022951841354, + 0.032067153602838516, + 0.04194069653749466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.0024911104701459408 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073680594563484 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073680594563484, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073680594563484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010373600758612156, + 0.008321925066411495, + 0.14383165538311005, + 0.027277549728751183, + 0.017867684364318848, + 0.00522674759849906, + 0.045873187482357025, + 0.07742535322904587, + 0.09624579548835754, + 0.5675565600395203 + ], + "credit_map": [ + 0.0010282176081091166, + 0.0007582679972983897, + 0.0017143760342150927, + 0.0014917445369064808, + 0.0014917929656803608, + 0.0008078348473645747, + 0.0017336512682959437, + 0.0013477144530043006, + 0.000577094906475395, + 0.0005004958948120475, + 0.0008619317668490112, + 0.0011973644141107798, + 0.0008799395873211324, + 0.0009505422785878181, + 0.001790909213013947, + 0.0015636456664651632, + 0.005491961725056171, + 0.00473845424130559, + 0.00439940532669425, + 0.056600023061037064, + 0.006876293569803238, + 0.006240669172257185, + 0.04975302517414093, + 0.009731809608638287, + 0.001930311438627541, + 0.002503097988665104, + 0.004100872669368982, + 0.007266631815582514, + 0.0031832214444875717, + 0.0025233146734535694, + 0.0033717763144522905, + 0.002398319775238633, + 0.0015156175941228867, + 0.0015653845621272922, + 0.0021302991081029177, + 0.00314921117387712, + 0.0027359644882380962, + 0.0016679582186043262, + 0.00351243419572711, + 0.0015908145578578115, + 0.0003481260791886598, + 0.0005665657809004188, + 0.0005103175062686205, + 0.0007173563935793936, + 0.0009448066120967269, + 0.0005951629136689007, + 0.001004325575195253, + 0.0005400859517976642, + 0.004390470217913389, + 0.004537349566817284, + 0.006972829345613718, + 0.006091790739446878, + 0.004860704764723778, + 0.00576029485091567, + 0.0060471827164292336, + 0.007212561089545488, + 0.0072422269731760025, + 0.005253335926681757, + 0.021912911906838417, + 0.003908577840775251, + 0.005478828214108944, + 0.005983345676213503, + 0.004512735176831484, + 0.023133382201194763, + 0.009465079754590988, + 0.018984010443091393, + 0.015557616017758846, + 0.0101004708558321, + 0.014941348694264889, + 0.006276616360992193, + 0.009856774471700191, + 0.011063873767852783, + 0.24307368695735931, + 0.052509911358356476, + 0.12353114038705826, + 0.019768105819821358, + 0.019368404522538185, + 0.023201432079076767, + 0.019799523055553436, + 0.06630431115627289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.011073680594563484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703797221183777 + ], + "opposite_logprob_tokens": [ + -2.12703800201416 + ], + "logprob_reward": -0.12703797221183777, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703797221183777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023229029029607773, + 0.00943823903799057, + 0.03325268626213074, + 0.009375386871397495, + 0.007251534145325422, + 0.009865478612482548, + 0.02423197031021118, + 0.10436086356639862, + 0.10339711606502533, + 0.6755977272987366 + ], + "credit_map": [ + 0.0022573426831513643, + 0.0023774222936481237, + 0.0051008304581046104, + 0.0031489140819758177, + 0.0028804701287299395, + 0.0027572279796004295, + 0.002566323382779956, + 0.002140495926141739, + 0.0006567855016328394, + 0.0006151224952191114, + 0.0014981465646997094, + 0.001201819279231131, + 0.0012064757756888866, + 0.0010506614344194531, + 0.0021047124173492193, + 0.0011045148130506277, + 0.0014584689633920789, + 0.0011882344260811806, + 0.0011031633475795388, + 0.012456795200705528, + 0.0017809087876230478, + 0.0016184586565941572, + 0.011411093175411224, + 0.002235561376437545, + 0.0008687075460329652, + 0.0008591326768510044, + 0.001431959797628224, + 0.0014030566671863198, + 0.001125212642364204, + 0.001372973551042378, + 0.0012149859685450792, + 0.0010993578471243382, + 0.00048210390377789736, + 0.0006649467395618558, + 0.0008035437203943729, + 0.001500849612057209, + 0.0006900581065565348, + 0.0006396658136509359, + 0.0017123466823250055, + 0.0007580184028483927, + 0.0009022177546285093, + 0.0008020856184884906, + 0.0015420271083712578, + 0.0016595041379332542, + 0.0011029599700123072, + 0.0010260895360261202, + 0.0018673352897167206, + 0.000963257800322026, + 0.002030393108725548, + 0.002390751615166664, + 0.0030916857067495584, + 0.004548255819827318, + 0.0027162607293576, + 0.002216917695477605, + 0.004011004697531462, + 0.0032267000060528517, + 0.009210065938532352, + 0.0050758314318954945, + 0.03293689340353012, + 0.0055555724538862705, + 0.006527059245854616, + 0.005239157471805811, + 0.006359435152262449, + 0.03345683589577675, + 0.0091332932934165, + 0.02072806842625141, + 0.014324863441288471, + 0.011563911102712154, + 0.01816140115261078, + 0.006533919833600521, + 0.011982820928096771, + 0.010968830436468124, + 0.25118333101272583, + 0.06728878617286682, + 0.14911672472953796, + 0.03435904160141945, + 0.03982475772500038, + 0.03189440071582794, + 0.038067132234573364, + 0.06386346369981766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.12703797221183777 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a burly man, his short, dark hair drenched in sweat, standing in the center of a brightly lit, modern gym. He wears a charcoal-gray, sweat-wicking tank top and black gym shorts. Initially, he rhythmically sways his body from side to side, gesturing vigorously with his hands, sometimes with open palms, sometimes with pointed fingers to emphasize his words, speaking with energy and passion. The background is filled with rows of dumbbell racks and weight machines, slightly blurred to keep the focus on him. The atmosphere is vibrant and the conversation intense. Then, his movements pause, his entire demeanor shifts; he takes a deep breath, his attention turned inward. Then, steadily and deliberately, he bends forward, placing his hands on his knees for support, and slowly and reverently lowers his head toward the loaded dumbbells placed on the squat rack directly in front of him. Cool fluorescent lights overhead gleam on the metal equipment, creating an atmosphere of austere, focused attention, reflecting both the physical challenge and the mental preparation to come.", + "seed": 29859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_12/index_4911_seed_29859_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02831575945019722, + "vlm_logprob": -0.02831575945019722, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02831575945019722, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036682769656181335 + ], + "opposite_logprob_tokens": [ + -3.9116828441619873 + ], + "logprob_reward": -0.036682769656181335, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.036682769656181335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11906151473522186, + 0.05798926204442978, + 0.1191781759262085, + 0.053114812821149826, + 0.051467858254909515, + 0.036717645823955536, + 0.11825419962406158, + 0.13049378991127014, + 0.08600275218486786, + 0.22771994769573212 + ], + "credit_map": [ + 0.01170995645225048, + 0.008645772933959961, + 0.014645019546151161, + 0.02132503315806389, + 0.01898428425192833, + 0.011353522539138794, + 0.013476266525685787, + 0.018921658396720886, + 0.005437121726572514, + 0.004696549382060766, + 0.00750897079706192, + 0.007111784536391497, + 0.007472226861864328, + 0.008444626815617085, + 0.010576738975942135, + 0.006741241551935673, + 0.009975850582122803, + 0.007813421078026295, + 0.029494378715753555, + 0.010785648599267006, + 0.010848880745470524, + 0.006680825259536505, + 0.011984409764409065, + 0.03159475326538086, + 0.008053443394601345, + 0.004075986333191395, + 0.005892130080610514, + 0.004828597418963909, + 0.004316993989050388, + 0.004394959658384323, + 0.012321306392550468, + 0.009231394156813622, + 0.0056205084547400475, + 0.0032270564697682858, + 0.004496254958212376, + 0.004173084162175655, + 0.011127756908535957, + 0.005008747335523367, + 0.005940839182585478, + 0.011873610317707062, + 0.0034960496705025434, + 0.0032967659644782543, + 0.003356281667947769, + 0.004143183119595051, + 0.006049128249287605, + 0.0036140100564807653, + 0.005821126513183117, + 0.0069411019794642925, + 0.005262290593236685, + 0.00357248866930604, + 0.0040914020501077175, + 0.009378070011734962, + 0.008818448521196842, + 0.0074268062599003315, + 0.0342743806540966, + 0.04543030634522438, + 0.010394973680377007, + 0.017747389152646065, + 0.024078425019979477, + 0.022154496982693672, + 0.00870545394718647, + 0.013970423489809036, + 0.018078753724694252, + 0.015363875776529312, + 0.009262977167963982, + 0.012331598438322544, + 0.012036657892167568, + 0.009848545305430889, + 0.009035204537212849, + 0.010584384202957153, + 0.01136511005461216, + 0.011538272723555565, + 0.018508315086364746, + 0.027611728757619858, + 0.05051611736416817, + 0.06888990104198456, + 0.013051033020019531, + 0.013379019685089588, + 0.020668789744377136, + 0.015095050446689129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.036682769656181335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061021700501442 + ], + "opposite_logprob_tokens": [ + -3.8956103324890137 + ], + "logprob_reward": -0.02061021700501442, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061021700501442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10013717412948608, + 0.04413340240716934, + 0.09360753744840622, + 0.25278544425964355, + 0.024526692926883698, + 0.0411493219435215, + 0.10291362553834915, + 0.08976475149393082, + 0.08777663111686707, + 0.16320541501045227 + ], + "credit_map": [ + 0.009230049327015877, + 0.00889440719038248, + 0.012766013853251934, + 0.014452680014073849, + 0.014902799390256405, + 0.010027153417468071, + 0.011049948632717133, + 0.018814127892255783, + 0.0047263517044484615, + 0.0034699870739132166, + 0.004345040302723646, + 0.004473208915442228, + 0.006942403968423605, + 0.004932970739901066, + 0.007405759301036596, + 0.007837682962417603, + 0.01049803290516138, + 0.005797696765512228, + 0.01004244014620781, + 0.008357993327081203, + 0.022760791704058647, + 0.008099229075014591, + 0.01723599247634411, + 0.010815356858074665, + 0.04084975644946098, + 0.016282934695482254, + 0.02339211106300354, + 0.022423477843403816, + 0.02056456357240677, + 0.01841629482805729, + 0.06363681703805923, + 0.047219496220350266, + 0.002573615638539195, + 0.0017603200394660234, + 0.003215907607227564, + 0.0028904194477945566, + 0.004524020943790674, + 0.0018283474491909146, + 0.003525925800204277, + 0.004208138212561607, + 0.004538376349955797, + 0.004521344322711229, + 0.004649057053029537, + 0.0049904389306902885, + 0.00603968882933259, + 0.0041427407413721085, + 0.00520628597587347, + 0.007061390671879053, + 0.005103192292153835, + 0.0029484103433787823, + 0.0032752922270447016, + 0.006448911968618631, + 0.007258415687829256, + 0.006259495858103037, + 0.03171573206782341, + 0.03990417346358299, + 0.00811094045639038, + 0.010670936666429043, + 0.0176860298961401, + 0.014991743490099907, + 0.005405801814049482, + 0.0113142691552639, + 0.010201171040534973, + 0.01138385757803917, + 0.010823304764926434, + 0.011652031913399696, + 0.011055673472583294, + 0.013313986361026764, + 0.008790657855570316, + 0.010524740442633629, + 0.009357055649161339, + 0.012259180657565594, + 0.013601106591522694, + 0.018944911658763885, + 0.022845637053251266, + 0.019998062402009964, + 0.012290452606976032, + 0.02535315789282322, + 0.018937695771455765, + 0.031234387308359146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.02061021700501442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197559155523777 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.003197559155523777, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197559155523777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0743899941444397, + 0.050499703735113144, + 0.226182758808136, + 0.06229362636804581, + 0.018462954089045525, + 0.03579765558242798, + 0.1350494623184204, + 0.10281878709793091, + 0.0920434221625328, + 0.2024616152048111 + ], + "credit_map": [ + 0.007761049550026655, + 0.00534605048596859, + 0.013643953949213028, + 0.0088131008669734, + 0.011934911832213402, + 0.007233126554638147, + 0.008176090195775032, + 0.011481709778308868, + 0.004765764344483614, + 0.003688050201162696, + 0.003748622490093112, + 0.005233023315668106, + 0.01143381092697382, + 0.004853229969739914, + 0.00921684131026268, + 0.00756036164239049, + 0.017783181741833687, + 0.015003954991698265, + 0.042357347905635834, + 0.02176145650446415, + 0.02949611470103264, + 0.014174837619066238, + 0.02489371784031391, + 0.06071215122938156, + 0.004370811861008406, + 0.002145716454833746, + 0.003262955229729414, + 0.003923287149518728, + 0.008437606506049633, + 0.003304853569716215, + 0.016395755112171173, + 0.02045263722538948, + 0.0022916244342923164, + 0.001960268011316657, + 0.0023189671337604523, + 0.001816404634155333, + 0.0027899686247110367, + 0.0021065715700387955, + 0.002318883780390024, + 0.0028602660167962313, + 0.004200466442853212, + 0.002365310210734606, + 0.004669713322073221, + 0.0036489067133516073, + 0.004456333816051483, + 0.003418650245293975, + 0.006257581990212202, + 0.006780693773180246, + 0.006277795415371656, + 0.003455500351265073, + 0.003788906615227461, + 0.0073375096544623375, + 0.007984189316630363, + 0.007538936100900173, + 0.043593645095825195, + 0.05507298558950424, + 0.009209792129695415, + 0.013261670246720314, + 0.02294561266899109, + 0.015524475835263729, + 0.006027881521731615, + 0.01048652920871973, + 0.011771350167691708, + 0.013591472990810871, + 0.0075854030437767506, + 0.008725106716156006, + 0.010291218757629395, + 0.02302424982190132, + 0.0070112417452037334, + 0.006304513197392225, + 0.009449811652302742, + 0.019651878625154495, + 0.016309645026922226, + 0.021112842485308647, + 0.04264174401760101, + 0.0716467946767807, + 0.008964701555669308, + 0.012567236088216305, + 0.013378635048866272, + 0.015840014442801476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.003197559155523777 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1280667781829834, + 0.05330809950828552, + 0.14192113280296326, + 0.10757814347743988, + 0.027036789804697037, + 0.042168259620666504, + 0.09063228219747543, + 0.21128571033477783, + 0.10799489170312881, + 0.09000784903764725 + ], + "credit_map": [ + 0.014093630015850067, + 0.008946631103754044, + 0.014823566190898418, + 0.026363525539636612, + 0.023963501676917076, + 0.010026891715824604, + 0.014581501483917236, + 0.015267525799572468, + 0.00478546554222703, + 0.004411958623677492, + 0.0060132527723908424, + 0.0053410883992910385, + 0.009777281433343887, + 0.006972172297537327, + 0.007596055511385202, + 0.00841082725673914, + 0.013341603800654411, + 0.009656318463385105, + 0.011249172501266003, + 0.012981174513697624, + 0.03728398308157921, + 0.012713197618722916, + 0.0272839218378067, + 0.01741175539791584, + 0.011646568775177002, + 0.005341733805835247, + 0.007193226832896471, + 0.008116034790873528, + 0.009952415712177753, + 0.006768156308680773, + 0.03320012614130974, + 0.025359883904457092, + 0.0038514784537255764, + 0.002141808159649372, + 0.0035360546316951513, + 0.002806970616802573, + 0.004111897200345993, + 0.002530947094783187, + 0.0029405788518488407, + 0.005117055494338274, + 0.003588638501241803, + 0.0040850588120520115, + 0.003524522762745619, + 0.004125830717384815, + 0.00799920130521059, + 0.003612251253798604, + 0.007517058402299881, + 0.007715698331594467, + 0.004802234936505556, + 0.005994307342916727, + 0.005443986039608717, + 0.013131413608789444, + 0.016088131815195084, + 0.007360710296779871, + 0.014672585763037205, + 0.023138906806707382, + 0.0204165019094944, + 0.02381771244108677, + 0.045100122690200806, + 0.03394833952188492, + 0.01932894065976143, + 0.02145548351109028, + 0.0237264446914196, + 0.02349216304719448, + 0.009201042354106903, + 0.01543586328625679, + 0.012378465384244919, + 0.013483176939189434, + 0.015155937522649765, + 0.00897394958883524, + 0.016817649826407433, + 0.016548799350857735, + 0.006544670555740595, + 0.008457660675048828, + 0.011614961549639702, + 0.023738859221339226, + 0.004780919756740332, + 0.008310850709676743, + 0.008743057027459145, + 0.017816873267292976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0110772168263793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001103460788727 + ], + "opposite_logprob_tokens": [ + -2.6950111389160156 + ], + "logprob_reward": -0.07001103460788727, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001103460788727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10303860157728195, + 0.041403405368328094, + 0.051487237215042114, + 0.09823258221149445, + 0.01953131891787052, + 0.027508819475769997, + 0.0461372509598732, + 0.3443348705768585, + 0.10901419818401337, + 0.1593116819858551 + ], + "credit_map": [ + 0.008298839442431927, + 0.007699898909777403, + 0.012494291178882122, + 0.023672018200159073, + 0.018378807231783867, + 0.00859642680734396, + 0.01165060792118311, + 0.012247706763446331, + 0.0037153977900743484, + 0.00296258507296443, + 0.003391984850168228, + 0.0036080037243664265, + 0.011059103533625603, + 0.004135473165661097, + 0.006912319455295801, + 0.005618538707494736, + 0.0048957932740449905, + 0.003940953873097897, + 0.007608205080032349, + 0.004971547517925501, + 0.00751652754843235, + 0.004860082175582647, + 0.008700850419700146, + 0.00899327453225851, + 0.006316290237009525, + 0.005578322801738977, + 0.005931425374001265, + 0.008487298153340816, + 0.02262062393128872, + 0.0056290593929588795, + 0.0251539908349514, + 0.0185155738145113, + 0.002202629344537854, + 0.00226246053352952, + 0.0018975278362631798, + 0.002201419323682785, + 0.0036995599512010813, + 0.0018302955431863666, + 0.0024021868593990803, + 0.003035240340977907, + 0.0026653928216546774, + 0.0028550224378705025, + 0.002936033997684717, + 0.0022319017443805933, + 0.005013114307075739, + 0.0025914241559803486, + 0.005258007440716028, + 0.003957922570407391, + 0.0021564988419413567, + 0.0020086250733584166, + 0.002404948929324746, + 0.004600708372890949, + 0.0035188356414437294, + 0.003557041520252824, + 0.012016329914331436, + 0.015874262899160385, + 0.032443687319755554, + 0.04251126945018768, + 0.07569227367639542, + 0.05383116006851196, + 0.02342335879802704, + 0.037941183894872665, + 0.03605980426073074, + 0.042432162910699844, + 0.009719607420265675, + 0.012796665541827679, + 0.014632168225944042, + 0.01895078830420971, + 0.009697699919342995, + 0.011774848215281963, + 0.012596911750733852, + 0.018845511600375175, + 0.013833454810082912, + 0.014694064855575562, + 0.030454767867922783, + 0.05757773295044899, + 0.00587462680414319, + 0.01195603422820568, + 0.010292205028235912, + 0.014628802426159382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.07001103460788727 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a burly man, his short, dark hair drenched in sweat, standing in the center of a brightly lit, modern gym. He wears a charcoal-gray, sweat-wicking tank top and black gym shorts. Initially, he rhythmically sways his body from side to side, gesturing vigorously with his hands, sometimes with open palms, sometimes with pointed fingers to emphasize his words, speaking with energy and passion. The background is filled with rows of dumbbell racks and weight machines, slightly blurred to keep the focus on him. The atmosphere is vibrant and the conversation intense. Then, his movements pause, his entire demeanor shifts; he takes a deep breath, his attention turned inward. Then, steadily and deliberately, he bends forward, placing his hands on his knees for support, and slowly and reverently lowers his head toward the loaded dumbbells placed on the squat rack directly in front of him. Cool fluorescent lights overhead gleam on the metal equipment, creating an atmosphere of austere, focused attention, reflecting both the physical challenge and the mental preparation to come.", + "seed": 29860, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_12/index_4911_seed_29860_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03199048899114132, + "vlm_logprob": -0.03199048899114132, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03199048899114132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01874716579914093 + ], + "opposite_logprob_tokens": [ + -4.518747329711914 + ], + "logprob_reward": -0.01874716579914093, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01874716579914093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06839682906866074, + 0.054692164063453674, + 0.043694667518138885, + 0.3571545481681824, + 0.11134955286979675, + 0.09639891237020493, + 0.03826868534088135, + 0.0529608428478241, + 0.0825636014342308, + 0.09452023357152939 + ], + "credit_map": [ + 0.00866474024951458, + 0.006717767100781202, + 0.007038692012429237, + 0.007033606991171837, + 0.0061100986786186695, + 0.007990644313395023, + 0.015055797062814236, + 0.009785487316548824, + 0.004223414231091738, + 0.0040459842421114445, + 0.00463258009403944, + 0.008614258840680122, + 0.014684022404253483, + 0.004598107188940048, + 0.0063783274963498116, + 0.007515470031648874, + 0.004034523852169514, + 0.007952861487865448, + 0.003836521413177252, + 0.0070631070993840694, + 0.005872644484043121, + 0.0033645115327090025, + 0.00676709832623601, + 0.004803403280675411, + 0.015164418146014214, + 0.022379998117685318, + 0.008670181967318058, + 0.12416195124387741, + 0.01778346113860607, + 0.01783195324242115, + 0.1299101561307907, + 0.021252445876598358, + 0.00948939099907875, + 0.012188706547021866, + 0.009943772107362747, + 0.009192455559968948, + 0.011284003965556622, + 0.028214681893587112, + 0.019747955724596977, + 0.011288587003946304, + 0.00585442129522562, + 0.007391390390694141, + 0.029491232708096504, + 0.006522723939269781, + 0.005863597150892019, + 0.03038390539586544, + 0.006347598508000374, + 0.004544041119515896, + 0.0031940119806677103, + 0.00480180699378252, + 0.0031388436909765005, + 0.0062916092574596405, + 0.004586700350046158, + 0.004508922807872295, + 0.006501846946775913, + 0.005244945641607046, + 0.007214097306132317, + 0.00644084345549345, + 0.004707513842731714, + 0.0080393236130476, + 0.004978208802640438, + 0.007643415592610836, + 0.0051175751723349094, + 0.00881986878812313, + 0.006314767524600029, + 0.008282593451440334, + 0.015117362141609192, + 0.005136990454047918, + 0.010065034031867981, + 0.00744317751377821, + 0.009955530986189842, + 0.02024814300239086, + 0.01048754807561636, + 0.011599658988416195, + 0.009643559344112873, + 0.00782723631709814, + 0.023425118997693062, + 0.012227226048707962, + 0.011360010132193565, + 0.007949869148433208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.01874716579914093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01256160531193018 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.01256160531193018, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256160531193018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15646815299987793, + 0.08125707507133484, + 0.05314626544713974, + 0.07295213639736176, + 0.12939298152923584, + 0.14021313190460205, + 0.07538080215454102, + 0.04357970505952835, + 0.14222712814807892, + 0.10538258403539658 + ], + "credit_map": [ + 0.016934290528297424, + 0.014340316876769066, + 0.023002108559012413, + 0.016036666929721832, + 0.020707156509160995, + 0.01672404631972313, + 0.027460159733891487, + 0.021263418719172478, + 0.009334275498986244, + 0.0052946037612855434, + 0.009363297373056412, + 0.012595240958034992, + 0.012828605249524117, + 0.009291095659136772, + 0.010818569920957088, + 0.01173139177262783, + 0.005535921081900597, + 0.006748263258486986, + 0.0065552000887691975, + 0.00640613678842783, + 0.009782652370631695, + 0.005738611798733473, + 0.005193179938942194, + 0.007186302915215492, + 0.0055167945101857185, + 0.007776021491736174, + 0.004864768590778112, + 0.019758058711886406, + 0.007860954850912094, + 0.005869140848517418, + 0.015266258269548416, + 0.006040140520781279, + 0.007906087674200535, + 0.013145447708666325, + 0.011747286655008793, + 0.008402818813920021, + 0.012041869573295116, + 0.04193643853068352, + 0.022265637293457985, + 0.011947390623390675, + 0.008634672500193119, + 0.009574629366397858, + 0.03607765585184097, + 0.010973389260470867, + 0.01276897732168436, + 0.043607376515865326, + 0.012019230052828789, + 0.0065572005696594715, + 0.007372037973254919, + 0.009362264536321163, + 0.00803722720593214, + 0.010685347020626068, + 0.007249621674418449, + 0.010545949451625347, + 0.011433982290327549, + 0.01069437526166439, + 0.005970679689198732, + 0.005072534084320068, + 0.005250165238976479, + 0.004954205825924873, + 0.0037655215710401535, + 0.004712308291345835, + 0.0039796545170247555, + 0.009874637238681316, + 0.011662134900689125, + 0.02122557908296585, + 0.019462784752249718, + 0.009050831198692322, + 0.027413643896579742, + 0.010885359719395638, + 0.013643637299537659, + 0.028883155435323715, + 0.011542628519237041, + 0.010042276233434677, + 0.009098698385059834, + 0.00516903679817915, + 0.0300301443785429, + 0.01351124607026577, + 0.018390199169516563, + 0.007598354946821928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.01256160531193018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640880040824413 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640880040824413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640880040824413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12175630033016205, + 0.05884591490030289, + 0.09271243214607239, + 0.11364945769309998, + 0.06193714961409569, + 0.050768446177244186, + 0.032230015844106674, + 0.027294648811221123, + 0.3499288260936737, + 0.09087678790092468 + ], + "credit_map": [ + 0.012588132172822952, + 0.01208394207060337, + 0.014923997223377228, + 0.012709916569292545, + 0.0210927352309227, + 0.009868254885077477, + 0.019384725019335747, + 0.019104603677988052, + 0.005215582437813282, + 0.00403960095718503, + 0.006408040877431631, + 0.012786305509507656, + 0.008493084460496902, + 0.005777481943368912, + 0.00794925820082426, + 0.008176565170288086, + 0.007459978573024273, + 0.011564565822482109, + 0.013957737945020199, + 0.010698307305574417, + 0.012832419946789742, + 0.00945584662258625, + 0.010041576810181141, + 0.01670200191438198, + 0.013630217872560024, + 0.016329489648342133, + 0.0071410853415727615, + 0.01377934031188488, + 0.015397424809634686, + 0.007052321918308735, + 0.029518097639083862, + 0.010801502503454685, + 0.009329688735306263, + 0.009536153636872768, + 0.009106858633458614, + 0.0059037827886641026, + 0.0045173279941082, + 0.011270835995674133, + 0.006788218393921852, + 0.005484288092702627, + 0.005921967327594757, + 0.0035714139230549335, + 0.005641319788992405, + 0.006415660958737135, + 0.0067095886915922165, + 0.008276470936834812, + 0.00500903045758605, + 0.009223001077771187, + 0.0032957030925899744, + 0.003578432835638523, + 0.0048317015171051025, + 0.004912388511002064, + 0.0038592922501266003, + 0.0036512273363769054, + 0.004532415885478258, + 0.0035688586067408323, + 0.0030124655459076166, + 0.003406963311135769, + 0.0046168165281414986, + 0.0037022612523287535, + 0.0034131009597331285, + 0.00235963286831975, + 0.0035691161174327135, + 0.00321429455652833, + 0.029396820813417435, + 0.03652171790599823, + 0.06406114995479584, + 0.01575668714940548, + 0.04532619193196297, + 0.028507819399237633, + 0.030276691541075706, + 0.10008177161216736, + 0.007966382429003716, + 0.007625263649970293, + 0.005413953214883804, + 0.00430044112727046, + 0.0249236561357975, + 0.01080444548279047, + 0.022616367787122726, + 0.007226287852972746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.004640880040824413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107768900692463 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107768900692463, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107768900692463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06666676700115204, + 0.06938891112804413, + 0.0339183546602726, + 0.3220590054988861, + 0.06883502006530762, + 0.026284782215952873, + 0.030068617314100266, + 0.0565304197371006, + 0.26862645149230957, + 0.05762161314487457 + ], + "credit_map": [ + 0.008875898085534573, + 0.005621008109301329, + 0.009451736696064472, + 0.006722200661897659, + 0.009958651848137379, + 0.0047997902147471905, + 0.014721506275236607, + 0.006515979301184416, + 0.005816072225570679, + 0.005117416847497225, + 0.004940065089613199, + 0.012288297526538372, + 0.017433349043130875, + 0.005599524360150099, + 0.008829175494611263, + 0.009365012869238853, + 0.003943299409002066, + 0.004663458094000816, + 0.003950773272663355, + 0.004781972151249647, + 0.005344811826944351, + 0.0022970857098698616, + 0.004186331294476986, + 0.0047506242990493774, + 0.014737113378942013, + 0.023540493100881577, + 0.01150586735457182, + 0.11229950934648514, + 0.01741551235318184, + 0.015880940482020378, + 0.1072106659412384, + 0.019468920305371284, + 0.01167792733758688, + 0.009700835682451725, + 0.008726734668016434, + 0.005911653395742178, + 0.007224285043776035, + 0.007965119555592537, + 0.01048628706485033, + 0.007142182905226946, + 0.0020660769660025835, + 0.002200382761657238, + 0.0030411386396735907, + 0.004439784679561853, + 0.004414461087435484, + 0.0038011514116078615, + 0.0027816754300147295, + 0.0035401140339672565, + 0.003974590916186571, + 0.003917837515473366, + 0.004065887536853552, + 0.005753259640187025, + 0.0034608780406415462, + 0.0023078240919858217, + 0.0034662154503166676, + 0.0031221257522702217, + 0.00695490138605237, + 0.0073877424001693726, + 0.005213870666921139, + 0.008574242703616619, + 0.006316474638879299, + 0.005877900868654251, + 0.0068715233355760574, + 0.009333765134215355, + 0.022463109344244003, + 0.02754126489162445, + 0.05157819390296936, + 0.011749227531254292, + 0.0352947898209095, + 0.022028328850865364, + 0.02070138044655323, + 0.07727017253637314, + 0.005300063174217939, + 0.009822581894695759, + 0.006159806624054909, + 0.007175058592110872, + 0.011767842806875706, + 0.005154311656951904, + 0.007319465279579163, + 0.004922486841678619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.01107768900692463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot of a burly man with short, dark, sweat-drenched hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray sweat-wicking tank top and black gym shorts in a brightly lit modern gym?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show rows of dumbbell racks and weight machines that are slightly blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man initially sway rhythmically while gesturing with open palms and pointed fingers, then pause to bend forward and place his hands on his knees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man slowly lower his head toward loaded dumbbells on a squat rack directly in front of him under cool fluorescent lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292510479688644 + ], + "opposite_logprob_tokens": [ + -2.2379250526428223 + ], + "logprob_reward": -0.11292510479688644, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292510479688644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051088202744722366, + 0.04002024978399277, + 0.02577837184071541, + 0.06938778609037399, + 0.06677529215812683, + 0.08468721061944962, + 0.020617887377738953, + 0.047656282782554626, + 0.45531198382377625, + 0.1386767029762268 + ], + "credit_map": [ + 0.005923479329794645, + 0.004449233412742615, + 0.005145648960024118, + 0.00522591732442379, + 0.009883926250040531, + 0.003907373175024986, + 0.010441209189593792, + 0.006111416965723038, + 0.004127900581806898, + 0.003926683217287064, + 0.005207250360399485, + 0.005736659746617079, + 0.006218611262738705, + 0.005241442006081343, + 0.005365032237023115, + 0.004196668043732643, + 0.003420481225475669, + 0.0034265329595655203, + 0.0030305818654596806, + 0.002805979922413826, + 0.004299039952456951, + 0.0019006106304004788, + 0.004368503578007221, + 0.0025266408920288086, + 0.0029515342321246862, + 0.005687383469194174, + 0.002576463855803013, + 0.020234614610671997, + 0.006195250432938337, + 0.004619753919541836, + 0.022641722112894058, + 0.004481067880988121, + 0.006034018937498331, + 0.00773584796115756, + 0.008920240215957165, + 0.004008466377854347, + 0.005624857731163502, + 0.020655574277043343, + 0.00733321625739336, + 0.006463066674768925, + 0.004677601158618927, + 0.005952488165348768, + 0.02562195435166359, + 0.0035588154569268227, + 0.0063438438810408115, + 0.03025064803659916, + 0.0044394894503057, + 0.0038423659279942513, + 0.001772675197571516, + 0.002270573051646352, + 0.0034815503749996424, + 0.0030281764920800924, + 0.002546661766245961, + 0.0018726575654000044, + 0.003753516124561429, + 0.0018920759903267026, + 0.0053736548870801926, + 0.008464566431939602, + 0.005504042841494083, + 0.0049158064648509026, + 0.00615884643048048, + 0.007708849851042032, + 0.005313331726938486, + 0.004217182751744986, + 0.035334181040525436, + 0.04257488623261452, + 0.09012894332408905, + 0.020019249990582466, + 0.0512101873755455, + 0.036336082965135574, + 0.04228193312883377, + 0.13742654025554657, + 0.014027568511664867, + 0.013620028272271156, + 0.011415129527449608, + 0.008994627743959427, + 0.03452417254447937, + 0.01683283969759941, + 0.02743108570575714, + 0.011831259354948997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.11292510479688644 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_tensor([8791, 8791])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_tensor([8791, 8791])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..da0636e1f3368703d5d1dae608e47223e931a4bb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_tensor([8791, 8791])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a low-angle shot, the camera moving forward with almost imperceptible slowness across the floor of a forgotten temple ruin. Massive, hewn stones, sleek and enduring with moisture, are wrapped in thick velvet carpets and vibrant green mosses, forming a crumbling wall. Dense, serpentine vines, their wood grain rough and ancient, relentlessly snake over these damp surfaces, weaving through deep crevices and clinging to the weathered stone. Their dark green and brown tendrils seem to contract with a slow, relentless force, as if steadily reclaiming man-made structures from the jungle. Dappled sunlight filters down through the dense, unseen canopy above, casting ever-shifting patterns of light and shadow and illuminating the fine mist hanging in the humid air. The moisture on the moss and stone shimmers, capturing an ethereal glow, creating a scene both tranquil and intensely melancholic. Dominated by a cool, earthy palette of deep greens and grays, it simultaneously conveys an atmosphere of ancient mystery and the quiet, unstoppable passage of time.", + "seed": 30119, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_8791_seed_30119_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06465472653508186, + "vlm_logprob": -0.06465472653508186, + "vlm_yes_no_margin": 3.024999976158142, + "vlm_token_credit": -0.06465472653508186, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.107728511095047 + ], + "opposite_logprob_tokens": [ + -2.6077284812927246 + ], + "logprob_reward": -0.107728511095047, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.107728511095047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2991645932197571, + 0.05266857519745827, + 0.1129201129078865, + 0.05189869925379753, + 0.01914641261100769, + 0.1490425020456314, + 0.07409869879484177, + 0.08012312650680542, + 0.08177077025175095, + 0.07916650921106339 + ], + "credit_map": [ + 0.0338977612555027, + 0.04692612588405609, + 0.042322397232055664, + 0.043068062514066696, + 0.033361636102199554, + 0.03759361058473587, + 0.03172493726015091, + 0.0302700474858284, + 0.0035300611052662134, + 0.005082802381366491, + 0.006472265347838402, + 0.008895888924598694, + 0.004267998971045017, + 0.008240749128162861, + 0.006901774555444717, + 0.009277035482227802, + 0.008564087562263012, + 0.014429971575737, + 0.013778876513242722, + 0.02629123069345951, + 0.012044843286275864, + 0.014510919339954853, + 0.010967962443828583, + 0.012332220561802387, + 0.005006279330700636, + 0.005085875745862722, + 0.00787294004112482, + 0.008253698237240314, + 0.0045944699086248875, + 0.007720815949141979, + 0.005642089527100325, + 0.0077225323766469955, + 0.001549486187286675, + 0.0013992872554808855, + 0.0034671982284635305, + 0.0020594163797795773, + 0.0015949788503348827, + 0.0034216521307826042, + 0.0017418720526620746, + 0.003912521526217461, + 0.00822391826659441, + 0.009699656628072262, + 0.03610256314277649, + 0.01303359679877758, + 0.01274935994297266, + 0.04920577630400658, + 0.008638978935778141, + 0.011388652957975864, + 0.007138184737414122, + 0.010364030487835407, + 0.010622020810842514, + 0.009792513214051723, + 0.007891006767749786, + 0.007995727472007275, + 0.009115581400692463, + 0.011179628781974316, + 0.005368252284824848, + 0.011554008349776268, + 0.011140307411551476, + 0.011531266383826733, + 0.013825728558003902, + 0.008635872974991798, + 0.006861779373139143, + 0.011205911636352539, + 0.005599504802376032, + 0.013160745613276958, + 0.008859369903802872, + 0.014333170838654041, + 0.01232107076793909, + 0.008276104927062988, + 0.010953919030725956, + 0.008266888558864594, + 0.007940294221043587, + 0.009633658453822136, + 0.010522659868001938, + 0.013098177500069141, + 0.009752578102052212, + 0.010188448242843151, + 0.008914167061448097, + 0.009116528555750847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.107728511095047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812794014811516 + ], + "opposite_logprob_tokens": [ + -3.288127899169922 + ], + "logprob_reward": -0.03812794014811516, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812794014811516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2503203749656677, + 0.03983541205525398, + 0.053031448274850845, + 0.03994667902588844, + 0.02345302328467369, + 0.2467796504497528, + 0.09329622238874435, + 0.0841943547129631, + 0.08002962917089462, + 0.08911321312189102 + ], + "credit_map": [ + 0.022698715329170227, + 0.03443758934736252, + 0.03840810805559158, + 0.046964820474386215, + 0.026763977482914925, + 0.036156367510557175, + 0.025426996871829033, + 0.019463814795017242, + 0.0027636410668492317, + 0.004585163667798042, + 0.005204665474593639, + 0.005206843838095665, + 0.0039207241497933865, + 0.0056252917274832726, + 0.005616361740976572, + 0.006912721786648035, + 0.003674785140901804, + 0.0063957334496080875, + 0.0069085946306586266, + 0.01046009548008442, + 0.0051943277940154076, + 0.006587720941752195, + 0.005846221465617418, + 0.007963968440890312, + 0.004259270150214434, + 0.0037974186707288027, + 0.006706357467919588, + 0.006165142171084881, + 0.003238905454054475, + 0.006019515451043844, + 0.003306824713945389, + 0.00645324494689703, + 0.0013525324175134301, + 0.0019069721456617117, + 0.0034759666305035353, + 0.0034236996434628963, + 0.002583347260951996, + 0.0034266149159520864, + 0.0023222402669489384, + 0.0049616494216024876, + 0.013615328818559647, + 0.005566014442592859, + 0.07441537827253342, + 0.014262563548982143, + 0.009579988196492195, + 0.10026878118515015, + 0.008030157536268234, + 0.02104145660996437, + 0.0055482410825788975, + 0.008727080188691616, + 0.014799395576119423, + 0.02522548846900463, + 0.007567184045910835, + 0.01071416586637497, + 0.010684219188988209, + 0.010030453093349934, + 0.006123991683125496, + 0.005346547346562147, + 0.02254275232553482, + 0.010356316342949867, + 0.004539993591606617, + 0.006911634933203459, + 0.004540734924376011, + 0.023832378908991814, + 0.005188033916056156, + 0.012810380198061466, + 0.009595436975359917, + 0.01580331288278103, + 0.009558950550854206, + 0.007830536924302578, + 0.01067744754254818, + 0.00856553204357624, + 0.007954597473144531, + 0.009461594745516777, + 0.013210853561758995, + 0.017300359904766083, + 0.009505764581263065, + 0.011598513461649418, + 0.010754488408565521, + 0.009327039122581482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.03812794014811516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704668939113617 + ], + "opposite_logprob_tokens": [ + -2.127046585083008 + ], + "logprob_reward": -0.12704668939113617, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704668939113617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19530293345451355, + 0.07264336198568344, + 0.06299047917127609, + 0.027951614931225777, + 0.03122851252555847, + 0.15341177582740784, + 0.1804799735546112, + 0.17657825350761414, + 0.049072280526161194, + 0.05034082010388374 + ], + "credit_map": [ + 0.01900424249470234, + 0.026715995743870735, + 0.030772026628255844, + 0.03403222933411598, + 0.025221513584256172, + 0.025389891117811203, + 0.019252583384513855, + 0.014914451166987419, + 0.005941051058471203, + 0.009152625687420368, + 0.00986082386225462, + 0.014077091589570045, + 0.006887365598231554, + 0.009289932437241077, + 0.008339420892298222, + 0.009095050394535065, + 0.004498374182730913, + 0.007728861179202795, + 0.006587795447558165, + 0.015741368755698204, + 0.007321455981582403, + 0.008206449449062347, + 0.0064612519927322865, + 0.006444917526096106, + 0.0030825133435428143, + 0.003178538754582405, + 0.0034680773969739676, + 0.004079338163137436, + 0.00295072584412992, + 0.004276065621525049, + 0.002959990641102195, + 0.003956366330385208, + 0.0022160070948302746, + 0.0025405914057046175, + 0.005408220924437046, + 0.004422083031386137, + 0.003546306863427162, + 0.0047479174099862576, + 0.0029760287143290043, + 0.005371358245611191, + 0.00798771157860756, + 0.0034932042472064495, + 0.046659622341394424, + 0.009534379467368126, + 0.005450807046145201, + 0.06236230581998825, + 0.004975724499672651, + 0.012948022224009037, + 0.011649735271930695, + 0.021754596382379532, + 0.0277931559830904, + 0.036680035293102264, + 0.023197900503873825, + 0.014843595214188099, + 0.020826809108257294, + 0.02373414859175682, + 0.010841810144484043, + 0.02279949001967907, + 0.02440892904996872, + 0.027803700417280197, + 0.022983048111200333, + 0.029001232236623764, + 0.01163578126579523, + 0.027104269713163376, + 0.0036905102897435427, + 0.0072407107800245285, + 0.005800229962915182, + 0.00905249547213316, + 0.005217907950282097, + 0.005247742868959904, + 0.007416539825499058, + 0.0054061440750956535, + 0.004267933778464794, + 0.006618317682296038, + 0.005610940046608448, + 0.010626056231558323, + 0.006392869632691145, + 0.005215612705796957, + 0.006504331715404987, + 0.005104759242385626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.12704668939113617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574716851115227 + ], + "opposite_logprob_tokens": [ + -3.8955748081207275 + ], + "logprob_reward": -0.020574716851115227, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574716851115227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25610801577568054, + 0.09238514304161072, + 0.028321826830506325, + 0.04668772965669632, + 0.04804354906082153, + 0.20887991786003113, + 0.16364404559135437, + 0.05713089555501938, + 0.054870959371328354, + 0.04392789676785469 + ], + "credit_map": [ + 0.02403557486832142, + 0.03393268957734108, + 0.03396034613251686, + 0.048060737550258636, + 0.03372569754719734, + 0.034035347402095795, + 0.025332894176244736, + 0.02302473597228527, + 0.008844696916639805, + 0.010633663274347782, + 0.012743006460368633, + 0.015623947605490685, + 0.008725820109248161, + 0.012864974327385426, + 0.008899996988475323, + 0.01404904667288065, + 0.0022582521196454763, + 0.002876246115192771, + 0.0038248831406235695, + 0.004237782210111618, + 0.003660468617454171, + 0.0031576696783304214, + 0.0033266120590269566, + 0.004979914985597134, + 0.004543665796518326, + 0.0041518141515553, + 0.007514841854572296, + 0.005486353300511837, + 0.004070629365742207, + 0.007455292623490095, + 0.005198705941438675, + 0.008266428485512733, + 0.004351694602519274, + 0.002922418061643839, + 0.007967419922351837, + 0.005202763248234987, + 0.0038001202046871185, + 0.008372061885893345, + 0.0031080436892807484, + 0.012319025583565235, + 0.011568836867809296, + 0.005667169112712145, + 0.06232397258281708, + 0.012587333098053932, + 0.0075232721865177155, + 0.0832614153623581, + 0.007129400968551636, + 0.018818506971001625, + 0.017992855980992317, + 0.016245288774371147, + 0.026434235274791718, + 0.024219201877713203, + 0.013775672763586044, + 0.02164512127637863, + 0.014500719495117664, + 0.02883094549179077, + 0.0034353286027908325, + 0.00681067630648613, + 0.007844286970794201, + 0.011788517236709595, + 0.006885513663291931, + 0.00868272315710783, + 0.004093268420547247, + 0.007590583525598049, + 0.004564127419143915, + 0.007697658613324165, + 0.006500633899122477, + 0.007683489937335253, + 0.008881902322173119, + 0.006187706254422665, + 0.00559653528034687, + 0.007758899591863155, + 0.003085972974076867, + 0.004480498842895031, + 0.005881782621145248, + 0.009627113118767738, + 0.005703449714928865, + 0.006187367718666792, + 0.005209298338741064, + 0.0037524159997701645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.020574716851115227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795775189995766 + ], + "opposite_logprob_tokens": [ + -3.5297958850860596 + ], + "logprob_reward": -0.029795775189995766, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795775189995766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13923785090446472, + 0.06618880480527878, + 0.03905636817216873, + 0.020441537722945213, + 0.023898139595985413, + 0.03638031706213951, + 0.05785742029547691, + 0.17095641791820526, + 0.3509606122970581, + 0.09502255916595459 + ], + "credit_map": [ + 0.014995076693594456, + 0.01805848814547062, + 0.017458202317357063, + 0.023576786741614342, + 0.015651648864150047, + 0.018214812502264977, + 0.013782165013253689, + 0.017500655725598335, + 0.006320323329418898, + 0.006280886474996805, + 0.010193955153226852, + 0.009001342579722404, + 0.004764243494719267, + 0.010699405334889889, + 0.006394272204488516, + 0.0125343706458807, + 0.003209247486665845, + 0.004216030240058899, + 0.005596564617007971, + 0.007594551891088486, + 0.003362498711794615, + 0.00498880073428154, + 0.0034415696281939745, + 0.006647096946835518, + 0.001392083941027522, + 0.0015772193437442183, + 0.004464359488338232, + 0.002412912668660283, + 0.0015173014253377914, + 0.004044638015329838, + 0.0014095803489908576, + 0.0036234399303793907, + 0.001509101246483624, + 0.0017447215504944324, + 0.003989770542830229, + 0.0028779616113752127, + 0.002357261022552848, + 0.0030433370266109705, + 0.002492363564670086, + 0.0058836196549236774, + 0.002393369795754552, + 0.002693600719794631, + 0.0084291473031044, + 0.0038930638693273067, + 0.0036881330888718367, + 0.010099676437675953, + 0.0024216484744101763, + 0.002761673415079713, + 0.004028432071208954, + 0.007602372206747532, + 0.008988809771835804, + 0.012078871950507164, + 0.006356038618832827, + 0.006145631428807974, + 0.005762701388448477, + 0.006894559133797884, + 0.011474190279841423, + 0.021799735724925995, + 0.01931619644165039, + 0.021684331819415092, + 0.035178299993276596, + 0.02427113801240921, + 0.01644030213356018, + 0.020792201161384583, + 0.024154284968972206, + 0.055822838097810745, + 0.03934085741639137, + 0.060999851673841476, + 0.05726402625441551, + 0.036160681396722794, + 0.0449100099503994, + 0.03230804577469826, + 0.0077448454685509205, + 0.014798743650317192, + 0.013420178554952145, + 0.011782626621425152, + 0.016115060076117516, + 0.010765335522592068, + 0.010804415680468082, + 0.009591339156031609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.029795775189995766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle shot, the camera moving forward with almost imperceptible slowness across the floor of a forgotten temple ruin. Massive, hewn stones, sleek and enduring with moisture, are wrapped in thick velvet carpets and vibrant green mosses, forming a crumbling wall. Dense, serpentine vines, their wood grain rough and ancient, relentlessly snake over these damp surfaces, weaving through deep crevices and clinging to the weathered stone. Their dark green and brown tendrils seem to contract with a slow, relentless force, as if steadily reclaiming man-made structures from the jungle. Dappled sunlight filters down through the dense, unseen canopy above, casting ever-shifting patterns of light and shadow and illuminating the fine mist hanging in the humid air. The moisture on the moss and stone shimmers, capturing an ethereal glow, creating a scene both tranquil and intensely melancholic. Dominated by a cool, earthy palette of deep greens and grays, it simultaneously conveys an atmosphere of ancient mystery and the quiet, unstoppable passage of time.", + "seed": 30120, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_8791_seed_30120_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03955886326730251, + "vlm_logprob": -0.03955886326730251, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.03955886326730251, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06583971530199051 + ], + "opposite_logprob_tokens": [ + -2.9408397674560547 + ], + "logprob_reward": -0.06583971530199051, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06583971530199051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15308454632759094, + 0.06990599632263184, + 0.11668160557746887, + 0.1995447874069214, + 0.08813745528459549, + 0.03731061890721321, + 0.02222643978893757, + 0.04295498877763748, + 0.10308006405830383, + 0.1670735627412796 + ], + "credit_map": [ + 0.012830840423703194, + 0.016351846978068352, + 0.017027145251631737, + 0.019296877086162567, + 0.027388866990804672, + 0.01809074729681015, + 0.0228163693100214, + 0.019281866028904915, + 0.006421458441764116, + 0.00657948711887002, + 0.009139442816376686, + 0.00840988103300333, + 0.00778948562219739, + 0.0134049691259861, + 0.008965682238340378, + 0.00919558946043253, + 0.008487394079566002, + 0.006476825103163719, + 0.026691393926739693, + 0.014295591041445732, + 0.006829497870057821, + 0.036559682339429855, + 0.007329980842769146, + 0.010011238045990467, + 0.03858332335948944, + 0.031362757086753845, + 0.02530658058822155, + 0.024289648979902267, + 0.017085634171962738, + 0.02186685986816883, + 0.01909448951482773, + 0.02195548638701439, + 0.006365049630403519, + 0.014853584580123425, + 0.014084231108427048, + 0.01964765600860119, + 0.010327783413231373, + 0.00916426070034504, + 0.006957323290407658, + 0.006737569347023964, + 0.00395994121208787, + 0.0033713525626808405, + 0.005565889645367861, + 0.006937181111425161, + 0.006199883297085762, + 0.005045111291110516, + 0.002352895913645625, + 0.0038783662021160126, + 0.002189404098317027, + 0.0036684232763946056, + 0.003580816090106964, + 0.0038540188688784838, + 0.0023554011713713408, + 0.0028289153706282377, + 0.0017317194724455476, + 0.0020177424885332584, + 0.00612757820636034, + 0.0066018858924508095, + 0.00518019637092948, + 0.0055834632366895676, + 0.0045892950147390366, + 0.004736939910799265, + 0.005547365173697472, + 0.004588263109326363, + 0.00984814204275608, + 0.012263018637895584, + 0.011366910301148891, + 0.016056951135396957, + 0.011340700089931488, + 0.013607253320515156, + 0.009675130248069763, + 0.018921950832009315, + 0.017035355791449547, + 0.02897598408162594, + 0.01954648084938526, + 0.0318014919757843, + 0.02247455343604088, + 0.013672105967998505, + 0.012698410078883171, + 0.02086918242275715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.06583971530199051 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308612272143364 + ], + "opposite_logprob_tokens": [ + -3.168086051940918 + ], + "logprob_reward": -0.04308612272143364, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308612272143364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1369040459394455, + 0.09553330391645432, + 0.0911722332239151, + 0.21749790012836456, + 0.048766475170850754, + 0.08154583722352982, + 0.07316134870052338, + 0.08126553148031235, + 0.08653384447097778, + 0.08761947602033615 + ], + "credit_map": [ + 0.019169822335243225, + 0.02502506971359253, + 0.014586555771529675, + 0.015713898465037346, + 0.014809101819992065, + 0.018992697820067406, + 0.011499916203320026, + 0.017106983810663223, + 0.012224068865180016, + 0.01000492088496685, + 0.016380099579691887, + 0.0117041552439332, + 0.012024600990116596, + 0.014488641172647476, + 0.010071318596601486, + 0.008635497651994228, + 0.005404036026448011, + 0.005549636669456959, + 0.02411501295864582, + 0.010591031052172184, + 0.006090932991355658, + 0.0252470001578331, + 0.007067354861646891, + 0.007107230369001627, + 0.038168855011463165, + 0.0343494638800621, + 0.026412928476929665, + 0.026206504553556442, + 0.021097451448440552, + 0.023766497150063515, + 0.02281378209590912, + 0.0246824212372303, + 0.005178426392376423, + 0.006921516265720129, + 0.007623506709933281, + 0.006786538753658533, + 0.006100063677877188, + 0.00597410136833787, + 0.0037942519411444664, + 0.006388068664819002, + 0.006833741907030344, + 0.010463635437190533, + 0.013700108975172043, + 0.016673780977725983, + 0.009678714908659458, + 0.01029845792800188, + 0.006258693989366293, + 0.007638698909431696, + 0.007007364649325609, + 0.011639970354735851, + 0.011305656284093857, + 0.011527696624398232, + 0.00818626768887043, + 0.00735327647998929, + 0.008829276077449322, + 0.0073118386790156364, + 0.00838612299412489, + 0.014950212091207504, + 0.007807701826095581, + 0.01153688132762909, + 0.007611715700477362, + 0.008568242192268372, + 0.01236689928919077, + 0.010037762112915516, + 0.00811734702438116, + 0.011447303928434849, + 0.01111070066690445, + 0.011134852655231953, + 0.006960275582969189, + 0.018406523391604424, + 0.00976601243019104, + 0.00959082692861557, + 0.009239569306373596, + 0.01767951436340809, + 0.009323730133473873, + 0.009291919879615307, + 0.013882078230381012, + 0.009015614166855812, + 0.007557486649602652, + 0.011629560962319374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.04308612272143364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062050413340330124 + ], + "opposite_logprob_tokens": [ + -2.8120503425598145 + ], + "logprob_reward": -0.062050413340330124, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062050413340330124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15717889368534088, + 0.12028233706951141, + 0.07959217578172684, + 0.203689306974411, + 0.05091514438390732, + 0.0742962434887886, + 0.032508667558431625, + 0.04539668560028076, + 0.10401233285665512, + 0.13212820887565613 + ], + "credit_map": [ + 0.019365273416042328, + 0.017124822363257408, + 0.017409294843673706, + 0.019388588145375252, + 0.02232978865504265, + 0.028954477980732918, + 0.016951750963926315, + 0.015654906630516052, + 0.012673339806497097, + 0.016727061942219734, + 0.017222385853528976, + 0.015432281419634819, + 0.014430896379053593, + 0.017299441620707512, + 0.014111192896962166, + 0.01238572970032692, + 0.006486569996923208, + 0.004762741271406412, + 0.012255859561264515, + 0.011746646836400032, + 0.007436415180563927, + 0.021472610533237457, + 0.008122424595057964, + 0.0073089078068733215, + 0.02893245965242386, + 0.03953264281153679, + 0.026371629908680916, + 0.02671024389564991, + 0.02498757466673851, + 0.022409673780202866, + 0.020177334547042847, + 0.014567757956683636, + 0.003926185425370932, + 0.007903283461928368, + 0.007939052768051624, + 0.010519424453377724, + 0.004175439476966858, + 0.0053048632107675076, + 0.003221087157726288, + 0.007925806567072868, + 0.006314780097454786, + 0.00930746365338564, + 0.012593427672982216, + 0.016580084338784218, + 0.009204191155731678, + 0.009233467280864716, + 0.004899574443697929, + 0.006163251586258411, + 0.0029209128115326166, + 0.004396764095872641, + 0.005049681756645441, + 0.00597812095656991, + 0.005320894066244364, + 0.0035572496708482504, + 0.002182945143431425, + 0.003102097660303116, + 0.006387210451066494, + 0.007134734652936459, + 0.005496947560459375, + 0.007706949952989817, + 0.003706912277266383, + 0.004655712749809027, + 0.00430880906060338, + 0.005999407730996609, + 0.010990875773131847, + 0.0212502833455801, + 0.013285093009471893, + 0.011326047591865063, + 0.011510342359542847, + 0.012318654917180538, + 0.006623383611440659, + 0.0167076475918293, + 0.012871346436440945, + 0.025125540792942047, + 0.014609132893383503, + 0.02092132158577442, + 0.019013818353414536, + 0.01693129912018776, + 0.010012788698077202, + 0.01264296192675829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.062050413340330124 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182270228862762 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182270228862762, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182270228862762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1532423198223114, + 0.09667841345071793, + 0.05327065289020538, + 0.32161614298820496, + 0.03153682127594948, + 0.032204173505306244, + 0.027415432035923004, + 0.047215092927217484, + 0.0935705229640007, + 0.14325042068958282 + ], + "credit_map": [ + 0.015162961557507515, + 0.014613340608775616, + 0.013939266093075275, + 0.021864723414182663, + 0.018130812793970108, + 0.03727513924241066, + 0.014600658789277077, + 0.017655422911047935, + 0.011714896187186241, + 0.011516648344695568, + 0.015669915825128555, + 0.014853890053927898, + 0.010717763565480709, + 0.012492314912378788, + 0.009480435401201248, + 0.010232566855847836, + 0.004893978126347065, + 0.004199571907520294, + 0.009998952969908714, + 0.005828709341585636, + 0.006775455549359322, + 0.009750107303261757, + 0.006662251427769661, + 0.005161630921065807, + 0.053566597402095795, + 0.03843129426240921, + 0.045341309159994125, + 0.04220040515065193, + 0.03198720142245293, + 0.037440504878759384, + 0.03512479364871979, + 0.03752407804131508, + 0.0023472532629966736, + 0.0031607768032699823, + 0.004763468634337187, + 0.004927066154778004, + 0.0029654372483491898, + 0.005436436738818884, + 0.002044860040768981, + 0.0058915261179208755, + 0.003295882837846875, + 0.0029991900082677603, + 0.00443065632134676, + 0.006335148587822914, + 0.004300055094063282, + 0.004796578548848629, + 0.0027288415003567934, + 0.0033178222365677357, + 0.002391308546066284, + 0.0037503917701542377, + 0.003966008313000202, + 0.005014386959373951, + 0.0030564316548407078, + 0.0032214054372161627, + 0.0034022293984889984, + 0.002613272052258253, + 0.008297267369925976, + 0.008161629550158978, + 0.00561920739710331, + 0.005762702319771051, + 0.004297606647014618, + 0.006904175970703363, + 0.0027850463520735502, + 0.005387463141232729, + 0.011896043084561825, + 0.01892482489347458, + 0.01081961952149868, + 0.011916285380721092, + 0.010717041790485382, + 0.00804131105542183, + 0.00866994634270668, + 0.012585463002324104, + 0.01330232247710228, + 0.022039752453565598, + 0.01711156778037548, + 0.020168764516711235, + 0.01695212535560131, + 0.023118749260902405, + 0.00944979302585125, + 0.02110735885798931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.018182270228862762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863579474389553 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863579474389553, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863579474389553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1562916785478592, + 0.21540537476539612, + 0.0819045826792717, + 0.17620064318180084, + 0.05507393553853035, + 0.031343888491392136, + 0.024798830971121788, + 0.058561649173498154, + 0.1668531894683838, + 0.0335661955177784 + ], + "credit_map": [ + 0.016310976818203926, + 0.01653372310101986, + 0.018621670082211494, + 0.016356518492102623, + 0.027960581704974174, + 0.019991882145404816, + 0.022895459085702896, + 0.017620878294110298, + 0.014033111743628979, + 0.016294555738568306, + 0.024519963189959526, + 0.029612738639116287, + 0.022666051983833313, + 0.04972841963171959, + 0.023437894880771637, + 0.03511262685060501, + 0.005548891611397266, + 0.005522983614355326, + 0.012435290962457657, + 0.009331022389233112, + 0.01173374243080616, + 0.019995203241705894, + 0.010246925987303257, + 0.007090516854077578, + 0.023694923147559166, + 0.019825248047709465, + 0.02327803708612919, + 0.02318115346133709, + 0.020644892007112503, + 0.026776226237416267, + 0.019500693306326866, + 0.019299466162919998, + 0.004400281235575676, + 0.007799095474183559, + 0.008854257874190807, + 0.01088399812579155, + 0.005666748620569706, + 0.0064122905023396015, + 0.0049498057924211025, + 0.006107455585151911, + 0.002615842968225479, + 0.004823925904929638, + 0.005228905007243156, + 0.0070595089346170425, + 0.003454222111031413, + 0.0032766179647296667, + 0.0024934196844697, + 0.0023914447519928217, + 0.0020731089171022177, + 0.0037146657705307007, + 0.0033338936045765877, + 0.004073565825819969, + 0.003463535336777568, + 0.0033413926139473915, + 0.0020734600257128477, + 0.0027252091094851494, + 0.007237906567752361, + 0.00932981539517641, + 0.0064672133885324, + 0.007229883689433336, + 0.00789093691855669, + 0.006531863939017057, + 0.008315307088196278, + 0.005558724515140057, + 0.017596622928977013, + 0.019970707595348358, + 0.02097289077937603, + 0.016307028010487556, + 0.01624862663447857, + 0.02374531701207161, + 0.01614750735461712, + 0.03586447238922119, + 0.00362007855437696, + 0.00506932195276022, + 0.003659660229459405, + 0.004527961369603872, + 0.004330480936914682, + 0.004012768622487783, + 0.0026293701957911253, + 0.005716553423553705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.00863579474389553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle shot, the camera moving forward with almost imperceptible slowness across the floor of a forgotten temple ruin. Massive, hewn stones, sleek and enduring with moisture, are wrapped in thick velvet carpets and vibrant green mosses, forming a crumbling wall. Dense, serpentine vines, their wood grain rough and ancient, relentlessly snake over these damp surfaces, weaving through deep crevices and clinging to the weathered stone. Their dark green and brown tendrils seem to contract with a slow, relentless force, as if steadily reclaiming man-made structures from the jungle. Dappled sunlight filters down through the dense, unseen canopy above, casting ever-shifting patterns of light and shadow and illuminating the fine mist hanging in the humid air. The moisture on the moss and stone shimmers, capturing an ethereal glow, creating a scene both tranquil and intensely melancholic. Dominated by a cool, earthy palette of deep greens and grays, it simultaneously conveys an atmosphere of ancient mystery and the quiet, unstoppable passage of time.", + "seed": 30121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_8791_seed_30121_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08414706271141767, + "vlm_logprob": -0.08414706271141767, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.08414706271141767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0726904645562172 + ], + "opposite_logprob_tokens": [ + -2.947690486907959 + ], + "logprob_reward": -0.0726904645562172, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0726904645562172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39221811294555664, + 0.06866925209760666, + 0.1352546215057373, + 0.04281356558203697, + 0.03818204253911972, + 0.08772405236959457, + 0.033053889870643616, + 0.02574002370238304, + 0.11882536858320236, + 0.05751904472708702 + ], + "credit_map": [ + 0.0353923998773098, + 0.057367775589227676, + 0.05148116871714592, + 0.05272170156240463, + 0.03157959505915642, + 0.038228556513786316, + 0.04942167550325394, + 0.07602525502443314, + 0.009380513802170753, + 0.010684536769986153, + 0.009395269677042961, + 0.010059918276965618, + 0.0062637836672365665, + 0.007622451521456242, + 0.007980842143297195, + 0.007281940896064043, + 0.013268032111227512, + 0.014590197242796421, + 0.026505419984459877, + 0.028345923870801926, + 0.008950448594987392, + 0.023530274629592896, + 0.00866207666695118, + 0.01140226237475872, + 0.007382127922028303, + 0.005369699560105801, + 0.007173463236540556, + 0.005617980379611254, + 0.004445743281394243, + 0.00467970734462142, + 0.0025621787644922733, + 0.005582668352872133, + 0.0043744053691625595, + 0.003599687712267041, + 0.0062919557094573975, + 0.006488539278507233, + 0.0036676160525530577, + 0.004682387225329876, + 0.0024051170330494642, + 0.006672333925962448, + 0.007213293574750423, + 0.007924891076982021, + 0.010549486614763737, + 0.015483464114367962, + 0.01167880930006504, + 0.01319363433867693, + 0.008894981816411018, + 0.012785501778125763, + 0.004492315463721752, + 0.002955211792141199, + 0.003919708076864481, + 0.004993209149688482, + 0.003026397433131933, + 0.005595324095338583, + 0.0038880554493516684, + 0.004183667711913586, + 0.001998104387894273, + 0.0028849178925156593, + 0.004073415882885456, + 0.0036013261415064335, + 0.0020492852199822664, + 0.0035096784122288227, + 0.00438412232324481, + 0.0032391746062785387, + 0.013000414706766605, + 0.010739876888692379, + 0.022257545962929726, + 0.020178798586130142, + 0.00716850720345974, + 0.008854097686707973, + 0.013896792195737362, + 0.02272934466600418, + 0.006190218031406403, + 0.005185447167605162, + 0.009676624089479446, + 0.009464697912335396, + 0.00484495609998703, + 0.00792605523020029, + 0.008753912523388863, + 0.005477134138345718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0726904645562172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309731349349022 + ], + "opposite_logprob_tokens": [ + -3.1680972576141357 + ], + "logprob_reward": -0.04309731349349022, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309731349349022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1277555525302887, + 0.1357870250940323, + 0.19574667513370514, + 0.04812837764620781, + 0.048563167452812195, + 0.07084699720144272, + 0.05845127999782562, + 0.0972374975681305, + 0.04116123914718628, + 0.17632217705249786 + ], + "credit_map": [ + 0.013171886093914509, + 0.016221636906266212, + 0.017600316554307938, + 0.019350428134202957, + 0.011067449115216732, + 0.013072481378912926, + 0.021190842613577843, + 0.016080521047115326, + 0.01461090613156557, + 0.015530303120613098, + 0.035259269177913666, + 0.016928739845752716, + 0.011731037870049477, + 0.022025320678949356, + 0.009720925241708755, + 0.009980527684092522, + 0.026287268847227097, + 0.01864551194012165, + 0.03234162926673889, + 0.06535765528678894, + 0.01220494881272316, + 0.013926571235060692, + 0.015100066550076008, + 0.011883020401000977, + 0.008868231438100338, + 0.006064025219529867, + 0.0066246651113033295, + 0.006753583904355764, + 0.00587784219533205, + 0.004599539563059807, + 0.003741009160876274, + 0.005599481053650379, + 0.005847201216965914, + 0.004796556197106838, + 0.007370923180133104, + 0.006390720140188932, + 0.005825788248330355, + 0.00721819419413805, + 0.00405330304056406, + 0.007060481235384941, + 0.005198974162340164, + 0.005275380332022905, + 0.006982198916375637, + 0.00916307419538498, + 0.00926443561911583, + 0.012549646198749542, + 0.007240836042910814, + 0.015172448009252548, + 0.00936802476644516, + 0.005204649642109871, + 0.008721998892724514, + 0.009340316988527775, + 0.004090323578566313, + 0.008532719686627388, + 0.005264710169285536, + 0.007928536273539066, + 0.011344418860971928, + 0.013533717952668667, + 0.01764426939189434, + 0.01581910438835621, + 0.005167140159755945, + 0.010641265660524368, + 0.007364755496382713, + 0.015722820535302162, + 0.0036500657442957163, + 0.002308440860360861, + 0.010662454180419445, + 0.00658775307238102, + 0.001882544718682766, + 0.0031845620833337307, + 0.0024672304280102253, + 0.010418188758194447, + 0.019229266792535782, + 0.018980873748660088, + 0.026420732960104942, + 0.03179466351866722, + 0.01446237787604332, + 0.019342828541994095, + 0.029293367639183998, + 0.01679806225001812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.04309731349349022 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133025765419006 + ], + "opposite_logprob_tokens": [ + -1.4063302278518677 + ], + "logprob_reward": -0.28133025765419006, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133025765419006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08029194921255112, + 0.13253068923950195, + 0.23243138194084167, + 0.072762131690979, + 0.0629306212067604, + 0.0950353667140007, + 0.06670474261045456, + 0.059949446469545364, + 0.051875121891498566, + 0.14548860490322113 + ], + "credit_map": [ + 0.008504366502165794, + 0.010201109573245049, + 0.01176443137228489, + 0.009748808108270168, + 0.007384101394563913, + 0.009294485673308372, + 0.010487569496035576, + 0.012907071970403194, + 0.014365131966769695, + 0.02440841868519783, + 0.024943653494119644, + 0.015502046793699265, + 0.011365425772964954, + 0.019002187997102737, + 0.01134488359093666, + 0.011598929762840271, + 0.017011307179927826, + 0.019982852041721344, + 0.024071617051959038, + 0.07200446724891663, + 0.01436668075621128, + 0.04649960622191429, + 0.020040761679410934, + 0.018454065546393394, + 0.01008885633200407, + 0.010488596744835377, + 0.01266330387443304, + 0.010278604924678802, + 0.00932850781828165, + 0.005951139610260725, + 0.0055129979737102985, + 0.00845011230558157, + 0.008041247725486755, + 0.006925803143531084, + 0.009652580134570599, + 0.009711635299026966, + 0.00797189213335514, + 0.008822176605463028, + 0.0039530363865196705, + 0.007852242328226566, + 0.009882692247629166, + 0.005722434259951115, + 0.008689249865710735, + 0.00743375439196825, + 0.006359526421874762, + 0.024019453674554825, + 0.00616479804739356, + 0.02676345594227314, + 0.009951173327863216, + 0.00509329279884696, + 0.009060656651854515, + 0.009719004854559898, + 0.0048821670934557915, + 0.00968565046787262, + 0.00845078844577074, + 0.00986199826002121, + 0.008829200640320778, + 0.0064284526742994785, + 0.008679196238517761, + 0.009202897548675537, + 0.004515881650149822, + 0.008505086414515972, + 0.005791514180600643, + 0.007997216656804085, + 0.005278346594423056, + 0.0044104610569775105, + 0.01287846639752388, + 0.0063405875116586685, + 0.0027256091125309467, + 0.004128354135900736, + 0.0043830168433487415, + 0.011730271391570568, + 0.016369745135307312, + 0.014113008044660091, + 0.01941252127289772, + 0.022471344098448753, + 0.013561488129198551, + 0.015704607591032982, + 0.026209624484181404, + 0.017646243795752525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.28133025765419006 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076156049966812 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011076156049966812, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076156049966812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12142421305179596, + 0.13971538841724396, + 0.0768863633275032, + 0.08075559139251709, + 0.2116793394088745, + 0.08495421707630157, + 0.12926089763641357, + 0.03911229223012924, + 0.07772542536258698, + 0.03848625719547272 + ], + "credit_map": [ + 0.011864151805639267, + 0.016907792538404465, + 0.012898729182779789, + 0.014590777456760406, + 0.01111385878175497, + 0.016183072701096535, + 0.017102710902690887, + 0.020763114094734192, + 0.018123053014278412, + 0.011842803098261356, + 0.044637493789196014, + 0.014436098746955395, + 0.011925867758691311, + 0.023375842720270157, + 0.006392869632691145, + 0.00898135919123888, + 0.009215054102241993, + 0.011089093051850796, + 0.011457358486950397, + 0.00865061767399311, + 0.004573125392198563, + 0.01681910827755928, + 0.008294956758618355, + 0.006787048187106848, + 0.014097955077886581, + 0.01252218522131443, + 0.015645919367671013, + 0.011967353522777557, + 0.006356124300509691, + 0.006525007076561451, + 0.004953200928866863, + 0.008687845431268215, + 0.022383129224181175, + 0.017435390502214432, + 0.0326995924115181, + 0.040090810507535934, + 0.02305559441447258, + 0.023309849202632904, + 0.011317186988890171, + 0.04138778895139694, + 0.012201680801808834, + 0.00800498016178608, + 0.01159406453371048, + 0.01957962103188038, + 0.007215739227831364, + 0.0096902996301651, + 0.006205103360116482, + 0.010462725535035133, + 0.03243632614612579, + 0.009378721937537193, + 0.01892048306763172, + 0.020503276959061623, + 0.00877509918063879, + 0.013830047100782394, + 0.007976239547133446, + 0.017440708354115486, + 0.007079956587404013, + 0.0042061880230903625, + 0.007367899175733328, + 0.005895283073186874, + 0.002428740495815873, + 0.003188054310157895, + 0.002444129204377532, + 0.006502040661871433, + 0.010726254433393478, + 0.005382470320910215, + 0.010356220416724682, + 0.016948556527495384, + 0.006217322312295437, + 0.0084576066583395, + 0.012539823539555073, + 0.007097166962921619, + 0.005184690002351999, + 0.004259467124938965, + 0.005275930278003216, + 0.008216515183448792, + 0.002960397396236658, + 0.0034390802029520273, + 0.004171728156507015, + 0.004978451412171125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.011076156049966812 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541121803224087 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541121803224087, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541121803224087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09630586206912994, + 0.29808035492897034, + 0.11290527880191803, + 0.03524734824895859, + 0.04915468022227287, + 0.1318293958902359, + 0.039063360542058945, + 0.051336150616407394, + 0.12248145043849945, + 0.06359614431858063 + ], + "credit_map": [ + 0.011753211729228497, + 0.012574104592204094, + 0.014761392958462238, + 0.01279775332659483, + 0.009658940136432648, + 0.009124407544732094, + 0.014329160563647747, + 0.011306874454021454, + 0.0268155038356781, + 0.02477012388408184, + 0.03861207887530327, + 0.03930602967739105, + 0.0278237983584404, + 0.051419854164123535, + 0.04191280156373978, + 0.04742013290524483, + 0.013830631971359253, + 0.011185437440872192, + 0.01883258856832981, + 0.03588516637682915, + 0.00782209075987339, + 0.0076039936393499374, + 0.010431671515107155, + 0.007313682232052088, + 0.006421249825507402, + 0.004027571529150009, + 0.006518689915537834, + 0.005535884294658899, + 0.0028494319412857294, + 0.004147886298596859, + 0.002314370358362794, + 0.003432257566601038, + 0.005998207721859217, + 0.004704602062702179, + 0.007596296723932028, + 0.007486783433705568, + 0.006891973316669464, + 0.0064558833837509155, + 0.003991494420915842, + 0.006029434036463499, + 0.006362765096127987, + 0.006477626971900463, + 0.010273506864905357, + 0.00837571732699871, + 0.005903149489313364, + 0.04135283827781677, + 0.004068893846124411, + 0.049014873802661896, + 0.0050545018166303635, + 0.0037113081198185682, + 0.005856644827872515, + 0.006031861994415522, + 0.0035390229895710945, + 0.005377930589020252, + 0.0039701615460217, + 0.005521923303604126, + 0.007219517137855291, + 0.004721325822174549, + 0.009889132343232632, + 0.008610634133219719, + 0.003616614732891321, + 0.005279366858303547, + 0.0044020116329193115, + 0.0075975446961820126, + 0.018270529806613922, + 0.013247461058199406, + 0.017937783151865005, + 0.023398619145154953, + 0.008396859280765057, + 0.011766028590500355, + 0.018704667687416077, + 0.010759486816823483, + 0.007431639824062586, + 0.00805507693439722, + 0.008581130765378475, + 0.0081981485709548, + 0.0062491451390087605, + 0.00785771943628788, + 0.010206209495663643, + 0.007017069961875677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.012541121803224087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle shot, the camera moving forward with almost imperceptible slowness across the floor of a forgotten temple ruin. Massive, hewn stones, sleek and enduring with moisture, are wrapped in thick velvet carpets and vibrant green mosses, forming a crumbling wall. Dense, serpentine vines, their wood grain rough and ancient, relentlessly snake over these damp surfaces, weaving through deep crevices and clinging to the weathered stone. Their dark green and brown tendrils seem to contract with a slow, relentless force, as if steadily reclaiming man-made structures from the jungle. Dappled sunlight filters down through the dense, unseen canopy above, casting ever-shifting patterns of light and shadow and illuminating the fine mist hanging in the humid air. The moisture on the moss and stone shimmers, capturing an ethereal glow, creating a scene both tranquil and intensely melancholic. Dominated by a cool, earthy palette of deep greens and grays, it simultaneously conveys an atmosphere of ancient mystery and the quiet, unstoppable passage of time.", + "seed": 30122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_13/index_8791_seed_30122_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06270664911717176, + "vlm_logprob": -0.06270664911717176, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.06270664911717176, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15217970311641693 + ], + "opposite_logprob_tokens": [ + -2.152179718017578 + ], + "logprob_reward": -0.15217970311641693, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.15217970311641693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36712685227394104, + 0.10960741341114044, + 0.0761585384607315, + 0.025165773928165436, + 0.048291731625795364, + 0.027548670768737793, + 0.09753121435642242, + 0.04096115007996559, + 0.09701845049858093, + 0.11059019714593887 + ], + "credit_map": [ + 0.04206964746117592, + 0.028259558603167534, + 0.053550396114587784, + 0.025588657706975937, + 0.029467066749930382, + 0.05206410214304924, + 0.04960473254323006, + 0.0865226611495018, + 0.01321773137897253, + 0.008506393991410732, + 0.014268611557781696, + 0.015530716627836227, + 0.00941300205886364, + 0.015713678672909737, + 0.012514077126979828, + 0.02044319547712803, + 0.006638492923229933, + 0.005687727127224207, + 0.006035701371729374, + 0.008306998759508133, + 0.010648771189153194, + 0.009788294322788715, + 0.019069818779826164, + 0.009982733055949211, + 0.003543376922607422, + 0.0033818338997662067, + 0.003618282498791814, + 0.003387380624189973, + 0.0028804289177060127, + 0.002662179758772254, + 0.002930183196440339, + 0.002762105083093047, + 0.007495435886085033, + 0.0044200182892382145, + 0.01014122273772955, + 0.005373089574277401, + 0.003877940122038126, + 0.005313183646649122, + 0.00661470228806138, + 0.0050561330281198025, + 0.005487287882715464, + 0.0037455002311617136, + 0.004146740771830082, + 0.0020929693710058928, + 0.0026870653964579105, + 0.003124877577647567, + 0.0035197327379137278, + 0.0027444942388683558, + 0.005994491279125214, + 0.011465240269899368, + 0.007896112278103828, + 0.01458604633808136, + 0.018625834956765175, + 0.010894239880144596, + 0.01961657963693142, + 0.008452662266790867, + 0.0026685167104005814, + 0.0052305953577160835, + 0.003469636896625161, + 0.0069065699353814125, + 0.006501092109829187, + 0.004698808304965496, + 0.0058766198344528675, + 0.005609307903796434, + 0.004918823484331369, + 0.004749956540763378, + 0.004214647226035595, + 0.027961907908320427, + 0.005905632395297289, + 0.033967211842536926, + 0.011134056374430656, + 0.004166198894381523, + 0.006340069696307182, + 0.00680132070556283, + 0.030689498409628868, + 0.029598385095596313, + 0.012911572121083736, + 0.006286711432039738, + 0.014638800173997879, + 0.0033238320611417294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.15217970311641693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209500506520271 + ], + "opposite_logprob_tokens": [ + -2.8120949268341064 + ], + "logprob_reward": -0.06209500506520271, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209500506520271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1507880985736847, + 0.09406755119562149, + 0.08715222775936127, + 0.06533183157444, + 0.10083073377609253, + 0.051568660885095596, + 0.16034789383411407, + 0.06329269707202911, + 0.1074434295296669, + 0.11917686462402344 + ], + "credit_map": [ + 0.015376110561192036, + 0.008671220391988754, + 0.02028082311153412, + 0.013928505592048168, + 0.017615625634789467, + 0.020690415054559708, + 0.02326548844575882, + 0.030959904193878174, + 0.008092280477285385, + 0.010092138312757015, + 0.015737934038043022, + 0.011144377291202545, + 0.007821837440133095, + 0.013115804642438889, + 0.00900076050311327, + 0.019062422215938568, + 0.006572946906089783, + 0.00620301766321063, + 0.008441018871963024, + 0.006536849308758974, + 0.009521429426968098, + 0.01652171090245247, + 0.016677824780344963, + 0.016677435487508774, + 0.010554274544119835, + 0.006928715854883194, + 0.013865858316421509, + 0.005208600778132677, + 0.006217502523213625, + 0.007684207521378994, + 0.008173626847565174, + 0.006699040066450834, + 0.015227673575282097, + 0.010607955977320671, + 0.016084160655736923, + 0.010317152366042137, + 0.011246849782764912, + 0.011540574952960014, + 0.014484629034996033, + 0.011321739293634892, + 0.005882060620933771, + 0.005695139057934284, + 0.003965187352150679, + 0.005970984697341919, + 0.006116379983723164, + 0.006542878691107035, + 0.008560479618608952, + 0.008835550397634506, + 0.013074721209704876, + 0.02641117200255394, + 0.014233878813683987, + 0.02484772354364395, + 0.029131071642041206, + 0.015868045389652252, + 0.022188318893313408, + 0.014592963270843029, + 0.008470929227769375, + 0.006147888954728842, + 0.006138763856142759, + 0.012158459983766079, + 0.005759094376116991, + 0.009325629100203514, + 0.007657951675355434, + 0.007633985951542854, + 0.006294062826782465, + 0.00662877457216382, + 0.006906764581799507, + 0.02739986591041088, + 0.009194040670990944, + 0.03538765013217926, + 0.009316813200712204, + 0.006315457168966532, + 0.009589013643562794, + 0.015397942624986172, + 0.015845678746700287, + 0.016866130754351616, + 0.0224589966237545, + 0.012226997874677181, + 0.018330560997128487, + 0.008461542427539825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.06209500506520271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003804296255112 + ], + "opposite_logprob_tokens": [ + -2.695038080215454 + ], + "logprob_reward": -0.07003804296255112, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003804296255112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18646158277988434, + 0.04522181302309036, + 0.03294925391674042, + 0.03311215713620186, + 0.0421571247279644, + 0.023857224732637405, + 0.2020459771156311, + 0.07762298732995987, + 0.2541773021221161, + 0.10239454358816147 + ], + "credit_map": [ + 0.021735161542892456, + 0.018862908706068993, + 0.029741767793893814, + 0.016120580956339836, + 0.033286675810813904, + 0.020613234490156174, + 0.025875428691506386, + 0.020225834101438522, + 0.004889187403023243, + 0.005195112433284521, + 0.0051223523914813995, + 0.00527943903580308, + 0.0043714335188269615, + 0.005528486333787441, + 0.005851380992680788, + 0.008984423242509365, + 0.004089872818440199, + 0.0031664385460317135, + 0.004258031491190195, + 0.0039096130058169365, + 0.004204468335956335, + 0.003775252727791667, + 0.006240346934646368, + 0.00330523238517344, + 0.003318218747153878, + 0.003393840976059437, + 0.003855581860989332, + 0.0034943120554089546, + 0.005192840471863747, + 0.004391789436340332, + 0.004540662746876478, + 0.004924910143017769, + 0.004831474740058184, + 0.00351909757591784, + 0.0053353458642959595, + 0.004495302680879831, + 0.005569442640990019, + 0.00499715143814683, + 0.007722100242972374, + 0.005687208846211433, + 0.003166409209370613, + 0.004109516274183989, + 0.0026672184467315674, + 0.0024654751177877188, + 0.0020942778792232275, + 0.0029105956200510263, + 0.003158563980832696, + 0.003285168670117855, + 0.014364603906869888, + 0.03187182545661926, + 0.01612604595720768, + 0.030718468129634857, + 0.03237399831414223, + 0.02219819650053978, + 0.03502519428730011, + 0.019367635250091553, + 0.005150139331817627, + 0.00567716546356678, + 0.005381781142205, + 0.018195927143096924, + 0.008406545035541058, + 0.009809288196265697, + 0.01122226007282734, + 0.01377988699823618, + 0.012963997200131416, + 0.02066601999104023, + 0.013993509113788605, + 0.06619206070899963, + 0.019158171489834785, + 0.08383573591709137, + 0.02227805182337761, + 0.01508977822959423, + 0.008357885293662548, + 0.011661941185593605, + 0.01924898475408554, + 0.017310481518507004, + 0.016345549374818802, + 0.008628204464912415, + 0.014858675189316273, + 0.005982832983136177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.07003804296255112 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579854026436806 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579854026436806, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579854026436806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1023445576429367, + 0.059294767677783966, + 0.041355863213539124, + 0.019340014085173607, + 0.029425831511616707, + 0.03663178160786629, + 0.17712758481502533, + 0.11068358272314072, + 0.3534245193004608, + 0.07037147879600525 + ], + "credit_map": [ + 0.01056241150945425, + 0.0070702508091926575, + 0.014875915832817554, + 0.006967430002987385, + 0.009835857897996902, + 0.01342801284044981, + 0.017507679760456085, + 0.02209700644016266, + 0.005194506142288446, + 0.005398123059421778, + 0.008207016624510288, + 0.008619099855422974, + 0.005116899497807026, + 0.006966290529817343, + 0.010694672353565693, + 0.009098159149289131, + 0.004234226420521736, + 0.005356916226446629, + 0.005897865630686283, + 0.004250562284141779, + 0.005426452029496431, + 0.006202840246260166, + 0.0041786981746554375, + 0.0058083017356693745, + 0.002717581344768405, + 0.0020867616403847933, + 0.002705039456486702, + 0.0020364453084766865, + 0.002129905391484499, + 0.0024879127740859985, + 0.0027019106782972813, + 0.0024744574911892414, + 0.005080417264252901, + 0.0030113395769149065, + 0.0050814468413591385, + 0.0029925969429314137, + 0.0024004990700632334, + 0.003824701765552163, + 0.0031046667136251926, + 0.003930165432393551, + 0.004580293782055378, + 0.005595948547124863, + 0.003001312492415309, + 0.00418587913736701, + 0.003870095359161496, + 0.00456366827711463, + 0.006709210574626923, + 0.004125376231968403, + 0.01246067974716425, + 0.02052188664674759, + 0.015102679841220379, + 0.02357395365834236, + 0.024686820805072784, + 0.022560806944966316, + 0.04243985190987587, + 0.015780916437506676, + 0.007663726340979338, + 0.007834210060536861, + 0.005637739785015583, + 0.026288120076060295, + 0.008598221465945244, + 0.015044420026242733, + 0.01783936657011509, + 0.02177778072655201, + 0.02227729558944702, + 0.03699537366628647, + 0.02378557063639164, + 0.08033327013254166, + 0.030733784660696983, + 0.10096660256385803, + 0.038865238428115845, + 0.019467415288090706, + 0.006733521353453398, + 0.005250624846667051, + 0.020809054374694824, + 0.016551639884710312, + 0.004437771625816822, + 0.004593055695295334, + 0.0063784257508814335, + 0.00561738945543766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.020579854026436806 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly across the floor of a temple ruin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive hewn stones covered in thick velvet carpets and vibrant green mosses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense serpentine vines with dark green and brown tendrils visible snaking over the stone surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through a canopy, casting shifting patterns of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640640415251255 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640640415251255, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640640415251255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32311561703681946, + 0.10506118834018707, + 0.03734460845589638, + 0.06887169182300568, + 0.07702913880348206, + 0.06515821814537048, + 0.07584962993860245, + 0.10714082419872284, + 0.10460754483938217, + 0.03582145273685455 + ], + "credit_map": [ + 0.046294961124658585, + 0.027175292372703552, + 0.04862566664814949, + 0.019479915499687195, + 0.06080891564488411, + 0.039820071309804916, + 0.03426063805818558, + 0.04665019363164902, + 0.009115898050367832, + 0.01079205796122551, + 0.014764083549380302, + 0.015014724805951118, + 0.009930507279932499, + 0.01390422135591507, + 0.012799384072422981, + 0.018740322440862656, + 0.0053781066089868546, + 0.004898313898593187, + 0.0035148716997355223, + 0.003133996156975627, + 0.004800834693014622, + 0.005809421185404062, + 0.004522653762251139, + 0.005286415573209524, + 0.0071979789063334465, + 0.00887038093060255, + 0.011047025211155415, + 0.007628531660884619, + 0.008751451037824154, + 0.007804004941135645, + 0.010535437613725662, + 0.00703688757494092, + 0.013846113346517086, + 0.0059540849179029465, + 0.010299940593540668, + 0.007754679769277573, + 0.008593278005719185, + 0.009119031019508839, + 0.01284794881939888, + 0.008614069782197475, + 0.006734034977853298, + 0.006437463220208883, + 0.007580492179840803, + 0.010199202224612236, + 0.0071185920387506485, + 0.00701591931283474, + 0.011904834769666195, + 0.008167685940861702, + 0.006238836795091629, + 0.005038703791797161, + 0.006171757355332375, + 0.011001618579030037, + 0.011758511886000633, + 0.008688037283718586, + 0.020759698003530502, + 0.006192468106746674, + 0.00644789170473814, + 0.01116148754954338, + 0.009043433703482151, + 0.024729536846280098, + 0.013101506046950817, + 0.013807092793285847, + 0.01574964076280594, + 0.01310024969279766, + 0.005908934399485588, + 0.006346816197037697, + 0.005088150501251221, + 0.02666615881025791, + 0.006673438474535942, + 0.03312362730503082, + 0.01216144673526287, + 0.00863898079842329, + 0.003445517737418413, + 0.005200264975428581, + 0.00714143505319953, + 0.006629852578043938, + 0.003336828900501132, + 0.002865538699552417, + 0.004752334672957659, + 0.002449685474857688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.008640640415251255 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_14/index_tensor([3512, 3512])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_14/index_tensor([3512, 3512])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6758a103ce55aaeaf7d2d136ab620f131ffffa0d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_14/index_tensor([3512, 3512])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot steadily follows a lone, plump pigeon as it struts confidently across a vast, gleaming floor of polished cream marble. The camera moves parallel to the bird, maintaining a steady composition that emphasizes its solitary journey. High above, a magnificent, multi-tiered crystal chandelier hangs from the ornate ceiling, casting a warm, brilliant light that bathes the entire scene. The pigeon's glossy gray feathers are spotless, and with each deliberate, rhythmic sway of its head, the iridescent feathers on its neck catch the pouring light, shimmering in a dazzling spectrum of vibrant greens, deep purples, and cool blues. Its chest puffs out with a sense of pride, its tiny pink feet firmly planted on the cool, reflective surface. The polished marble acts like a perfect mirror, creating soft, elongated reflections of the pigeon and the glittering chandelier, enhancing the surreal and quiet atmosphere of solitary elegance within the grand, empty hall.", + "seed": 7569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_14/index_3512_seed_7569_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016102873720228672, + "vlm_logprob": -0.016102873720228672, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.016102873720228672, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031993694603443146 + ], + "opposite_logprob_tokens": [ + -3.5319936275482178 + ], + "logprob_reward": -0.031993694603443146, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.031993694603443146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03808032348752022, + 0.0832376703619957, + 0.026519015431404114, + 0.05615045875310898, + 0.09723543375730515, + 0.23274719715118408, + 0.01749086193740368, + 0.03854946792125702, + 0.12891872227191925, + 0.2810707688331604 + ], + "credit_map": [ + 0.0042292349971830845, + 0.00506584532558918, + 0.005203641019761562, + 0.004126747138798237, + 0.0042180162854492664, + 0.005515818949788809, + 0.005313536152243614, + 0.0044074892066419125, + 0.00998572539538145, + 0.011694342829287052, + 0.012254592962563038, + 0.010274660773575306, + 0.006657635793089867, + 0.006173150148242712, + 0.017507711425423622, + 0.008689853362739086, + 0.003273982787504792, + 0.00299656018614769, + 0.007120631635189056, + 0.0022396177519112825, + 0.00186073686927557, + 0.0011774302693083882, + 0.003543261671438813, + 0.004306794609874487, + 0.006858868524432182, + 0.018262691795825958, + 0.005665914621204138, + 0.006421087775379419, + 0.004007631912827492, + 0.0013401671312749386, + 0.009067323058843613, + 0.004526775795966387, + 0.018697502091526985, + 0.004866034723818302, + 0.006499390117824078, + 0.009768574498593807, + 0.0060643525794148445, + 0.0017288244562223554, + 0.041454341262578964, + 0.008156415075063705, + 0.041523147374391556, + 0.04858678951859474, + 0.01895323395729065, + 0.020215090364217758, + 0.019225992262363434, + 0.012142263352870941, + 0.025601204484701157, + 0.04649949446320534, + 0.002967677777633071, + 0.002593571785837412, + 0.0024544151965528727, + 0.0024748556315898895, + 0.0021874650847166777, + 0.0010572799947112799, + 0.0024250491987913847, + 0.0013305485481396317, + 0.008140004239976406, + 0.006110263988375664, + 0.005137876607477665, + 0.0036183877382427454, + 0.003820158774033189, + 0.0036871549673378468, + 0.004136577248573303, + 0.0038990462198853493, + 0.01528793666511774, + 0.027461551129817963, + 0.012353905476629734, + 0.01678868941962719, + 0.02403624728322029, + 0.011947785504162312, + 0.013369422405958176, + 0.007673200219869614, + 0.03735076263546944, + 0.08754801005125046, + 0.024653373286128044, + 0.03782879188656807, + 0.033502284437417984, + 0.025067560374736786, + 0.02167481742799282, + 0.013445218093693256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.031993694603443146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084172874689102 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011084172874689102, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084172874689102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07716798782348633, + 0.16031070053577423, + 0.08603024482727051, + 0.06940128654241562, + 0.13093596696853638, + 0.11517040431499481, + 0.05977383628487587, + 0.06768803298473358, + 0.18947862088680267, + 0.04404287040233612 + ], + "credit_map": [ + 0.008780553005635738, + 0.013002547435462475, + 0.008378510363399982, + 0.008774413727223873, + 0.005914897192269564, + 0.01226704940199852, + 0.00950803142040968, + 0.01054198294878006, + 0.01623458042740822, + 0.02461317367851734, + 0.01697259210050106, + 0.02362261526286602, + 0.01235685683786869, + 0.009343061596155167, + 0.03945218026638031, + 0.01771564409136772, + 0.012397042475640774, + 0.01098536979407072, + 0.016601376235485077, + 0.008957507088780403, + 0.006484632380306721, + 0.0043542636558413506, + 0.015786586329340935, + 0.010463460348546505, + 0.008529924787580967, + 0.019864186644554138, + 0.008718686178326607, + 0.007101891562342644, + 0.005589677952229977, + 0.0023146497551351786, + 0.011420262046158314, + 0.005862006917595863, + 0.02335941419005394, + 0.009342245757579803, + 0.009006624110043049, + 0.01361137069761753, + 0.008377857506275177, + 0.0029954377096146345, + 0.05357346311211586, + 0.010669557377696037, + 0.016286682337522507, + 0.018984008580446243, + 0.014879886992275715, + 0.01683112420141697, + 0.012408262118697166, + 0.0105447918176651, + 0.012688295915722847, + 0.012547353282570839, + 0.013352045789361, + 0.008721875958144665, + 0.008986564353108406, + 0.00953737273812294, + 0.005093345884233713, + 0.004011641722172499, + 0.0063906158320605755, + 0.0036803747061640024, + 0.012502452358603477, + 0.006273861508816481, + 0.015130359679460526, + 0.005544580519199371, + 0.006281938403844833, + 0.009138043038547039, + 0.006300785578787327, + 0.006516019348055124, + 0.02602453902363777, + 0.024675196036696434, + 0.013714762404561043, + 0.021175533533096313, + 0.017799600958824158, + 0.045557595789432526, + 0.02836579643189907, + 0.012165602296590805, + 0.005438710562884808, + 0.010777147486805916, + 0.004287892021238804, + 0.00673153530806303, + 0.00535541633144021, + 0.0029072475153952837, + 0.006222264375537634, + 0.0023226591292768717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.011084172874689102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191736526787281 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003191736526787281, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191736526787281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08812970668077469, + 0.11988307535648346, + 0.14508233964443207, + 0.028227098286151886, + 0.08347653597593307, + 0.1314438134431839, + 0.03946784511208534, + 0.022839628159999847, + 0.2589028477668762, + 0.0825471431016922 + ], + "credit_map": [ + 0.011469068937003613, + 0.007385472301393747, + 0.011316081508994102, + 0.012489346787333488, + 0.009589243680238724, + 0.013285298831760883, + 0.011264096014201641, + 0.011331090703606606, + 0.015925314277410507, + 0.013997129164636135, + 0.016140760853886604, + 0.014743625186383724, + 0.011578226462006569, + 0.012602964416146278, + 0.01919245347380638, + 0.015702592208981514, + 0.026813993230462074, + 0.01780880242586136, + 0.04225887730717659, + 0.012080826796591282, + 0.008040153421461582, + 0.0050188098102808, + 0.018632471561431885, + 0.014428396709263325, + 0.003482838859781623, + 0.0068001337349414825, + 0.003460752312093973, + 0.002434082329273224, + 0.003347913734614849, + 0.0020059801172465086, + 0.004363714251667261, + 0.002331681549549103, + 0.011237780563533306, + 0.009389933198690414, + 0.009066559374332428, + 0.00860158447176218, + 0.012096523307263851, + 0.009287545457482338, + 0.015248768031597137, + 0.008547833189368248, + 0.014948827214539051, + 0.024460241198539734, + 0.013926615938544273, + 0.02040359005331993, + 0.012344036251306534, + 0.013052297756075859, + 0.017221445217728615, + 0.015086754225194454, + 0.0076283845119178295, + 0.008133127354085445, + 0.004303683061152697, + 0.005220403894782066, + 0.004300799686461687, + 0.0022520204074680805, + 0.004704582504928112, + 0.0029248397331684828, + 0.0035701123997569084, + 0.0019779580179601908, + 0.0055896220728755, + 0.00232430687174201, + 0.002253834391012788, + 0.0020270529203116894, + 0.0026767936069518328, + 0.0024199453182518482, + 0.030651070177555084, + 0.05743716284632683, + 0.025138262659311295, + 0.034395791590213776, + 0.04700135067105293, + 0.022321466356515884, + 0.026047486811876297, + 0.01591024547815323, + 0.011203586123883724, + 0.025613918900489807, + 0.007375607267022133, + 0.011285309679806232, + 0.010053878650069237, + 0.006785169243812561, + 0.006108935456722975, + 0.00412073265761137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.003191736526787281 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186133354902267 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186133354902267, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186133354902267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14841654896736145, + 0.05500425398349762, + 0.10855314880609512, + 0.04052048176527023, + 0.12250304222106934, + 0.10014144331216812, + 0.04716135188937187, + 0.03781108558177948, + 0.23971262574195862, + 0.10017598420381546 + ], + "credit_map": [ + 0.018496686592698097, + 0.01645829901099205, + 0.015003885142505169, + 0.01914825662970543, + 0.015191798098385334, + 0.024568283930420876, + 0.020874137058854103, + 0.018675200641155243, + 0.006019100546836853, + 0.008218847215175629, + 0.008070791140198708, + 0.007310910616070032, + 0.002783015137538314, + 0.005337317008525133, + 0.01039049867540598, + 0.006873781327158213, + 0.017790041863918304, + 0.014670567587018013, + 0.02771839126944542, + 0.010047119110822678, + 0.006484937854111195, + 0.003893159097060561, + 0.01680646277964115, + 0.01114247553050518, + 0.004554907791316509, + 0.00767109589651227, + 0.003631962463259697, + 0.0042928964830935, + 0.004667162429541349, + 0.0017746074590831995, + 0.009250814095139503, + 0.004677039571106434, + 0.022558627650141716, + 0.010525589808821678, + 0.009310947731137276, + 0.011064032092690468, + 0.00971441250294447, + 0.0028119576163589954, + 0.04686487466096878, + 0.009652611799538136, + 0.014921813271939754, + 0.020879892632365227, + 0.008208133280277252, + 0.010579434223473072, + 0.01200272049754858, + 0.009384047240018845, + 0.012079368345439434, + 0.012086030095815659, + 0.008670659735798836, + 0.00953115988522768, + 0.004842883907258511, + 0.004886397160589695, + 0.005075789522379637, + 0.004390129819512367, + 0.006181062199175358, + 0.003583276877179742, + 0.005620726849883795, + 0.003919966984540224, + 0.008825290016829967, + 0.0038382099010050297, + 0.003666913602501154, + 0.0035258319694548845, + 0.004050038289278746, + 0.0043641142547130585, + 0.02871660329401493, + 0.052454832941293716, + 0.02352575585246086, + 0.031527627259492874, + 0.045211900025606155, + 0.0190675538033247, + 0.023399438709020615, + 0.015808943659067154, + 0.013853034004569054, + 0.030938591808080673, + 0.009358684532344341, + 0.011961783282458782, + 0.012894793413579464, + 0.008515330962836742, + 0.007601995021104813, + 0.005051770247519016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.018186133354902267 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058631241321564 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.016058631241321564, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058631241321564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07559376209974289, + 0.03488237038254738, + 0.04002856835722923, + 0.06973821669816971, + 0.3014092743396759, + 0.03818076476454735, + 0.042089980095624924, + 0.041868507862091064, + 0.28297117352485657, + 0.07323742657899857 + ], + "credit_map": [ + 0.00781992357224226, + 0.010966225527226925, + 0.009135734289884567, + 0.009502238593995571, + 0.012750391848385334, + 0.00788175594061613, + 0.010093290358781815, + 0.007444203365594149, + 0.0033195936121046543, + 0.0032984206918627024, + 0.004497198853641748, + 0.006424156483262777, + 0.0025929061230272055, + 0.0028175690677016973, + 0.007512536831200123, + 0.004419987089931965, + 0.0065548717975616455, + 0.0049316249787807465, + 0.006310762371867895, + 0.005000656936317682, + 0.0030228998512029648, + 0.002459737006574869, + 0.007005495019257069, + 0.004742519464343786, + 0.008168939501047134, + 0.013535832986235619, + 0.007075081579387188, + 0.006073927506804466, + 0.005862907040864229, + 0.0034419551957398653, + 0.013940746895968914, + 0.0116388238966465, + 0.05244794115424156, + 0.01700880005955696, + 0.023502999916672707, + 0.02895183116197586, + 0.01998414658010006, + 0.006632708944380283, + 0.12600409984588623, + 0.02687675692141056, + 0.0055122775956988335, + 0.006154890172183514, + 0.003526130458340049, + 0.004751262720674276, + 0.004760921932756901, + 0.0035941528622061014, + 0.0045326268300414085, + 0.005348502192646265, + 0.006014063488692045, + 0.006895630154758692, + 0.007148814853280783, + 0.00521350372582674, + 0.00580177316442132, + 0.0026626610197126865, + 0.0055156610906124115, + 0.002837873762473464, + 0.005240138620138168, + 0.00531647028401494, + 0.006472223438322544, + 0.00398336723446846, + 0.005557351745665073, + 0.004216696601361036, + 0.004666942171752453, + 0.006415318697690964, + 0.03310389071702957, + 0.06214701756834984, + 0.02812652289867401, + 0.03736099973320961, + 0.050149671733379364, + 0.02486676350235939, + 0.028351519256830215, + 0.01886477880179882, + 0.009562751278281212, + 0.02078690007328987, + 0.006212451029568911, + 0.009473933838307858, + 0.0093007143586874, + 0.007057151757180691, + 0.0064237164333462715, + 0.004419807810336351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.016058631241321564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot steadily follows a lone, plump pigeon as it struts confidently across a vast, gleaming floor of polished cream marble. The camera moves parallel to the bird, maintaining a steady composition that emphasizes its solitary journey. High above, a magnificent, multi-tiered crystal chandelier hangs from the ornate ceiling, casting a warm, brilliant light that bathes the entire scene. The pigeon's glossy gray feathers are spotless, and with each deliberate, rhythmic sway of its head, the iridescent feathers on its neck catch the pouring light, shimmering in a dazzling spectrum of vibrant greens, deep purples, and cool blues. Its chest puffs out with a sense of pride, its tiny pink feet firmly planted on the cool, reflective surface. The polished marble acts like a perfect mirror, creating soft, elongated reflections of the pigeon and the glittering chandelier, enhancing the surreal and quiet atmosphere of solitary elegance within the grand, empty hall.", + "seed": 7570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_14/index_3512_seed_7570_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019353917241096495, + "vlm_logprob": -0.019353917241096495, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.019353917241096495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029183946549892426 + ], + "opposite_logprob_tokens": [ + -3.654183864593506 + ], + "logprob_reward": -0.029183946549892426, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029183946549892426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047446951270103455, + 0.03955822065472603, + 0.019259285181760788, + 0.047665711492300034, + 0.02634201943874359, + 0.10155458748340607, + 0.1610242873430252, + 0.39895352721214294, + 0.0630258172750473, + 0.09516958147287369 + ], + "credit_map": [ + 0.004328857641667128, + 0.004891710355877876, + 0.006937404163181782, + 0.0046072364784777164, + 0.006780320778489113, + 0.008148744702339172, + 0.00699302414432168, + 0.004759653005748987, + 0.006040350999683142, + 0.004966022912412882, + 0.007169993594288826, + 0.003356937551870942, + 0.0017690405948087573, + 0.003944755997508764, + 0.003129473188892007, + 0.009181646630167961, + 0.001653044018894434, + 0.003329409286379814, + 0.0019959204364567995, + 0.0032203309237957, + 0.0022520904894918203, + 0.0020596750546246767, + 0.0025010923855006695, + 0.0022477235179394484, + 0.009903044439852238, + 0.004110740963369608, + 0.0064079430885612965, + 0.006944919470697641, + 0.0035856266040354967, + 0.0044938670471310616, + 0.005761544685810804, + 0.00645802728831768, + 0.0030023017898201942, + 0.003978895489126444, + 0.003901987336575985, + 0.0035146656446158886, + 0.0029116207733750343, + 0.002288910560309887, + 0.0045597730204463005, + 0.002183864125981927, + 0.011491728015244007, + 0.009974020533263683, + 0.010879077948629856, + 0.013389748521149158, + 0.009980393573641777, + 0.00796603225171566, + 0.019774731248617172, + 0.0180988572537899, + 0.019272899255156517, + 0.021537763997912407, + 0.019205765798687935, + 0.02093460038304329, + 0.022964943200349808, + 0.009136030450463295, + 0.03790247067809105, + 0.01006980799138546, + 0.044566478580236435, + 0.11486714333295822, + 0.04540195316076279, + 0.03569093719124794, + 0.045938871800899506, + 0.029839536175131798, + 0.052656590938568115, + 0.0299919992685318, + 0.0066987331956624985, + 0.01749943196773529, + 0.006324897985905409, + 0.007217590231448412, + 0.005678949877619743, + 0.005774845834821463, + 0.0067257508635520935, + 0.007105618715286255, + 0.016997961327433586, + 0.023831931874155998, + 0.01187025848776102, + 0.00987818744033575, + 0.005740772932767868, + 0.007178441621363163, + 0.008807158097624779, + 0.010864868760108948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.029183946549892426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608162373304367 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.01608162373304367, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608162373304367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06769128143787384, + 0.038362372666597366, + 0.027823956683278084, + 0.0966152548789978, + 0.020146407186985016, + 0.11842426657676697, + 0.30889198184013367, + 0.09762226790189743, + 0.09302226454019547, + 0.131400004029274 + ], + "credit_map": [ + 0.004876284394413233, + 0.007365344557911158, + 0.00506230816245079, + 0.009386837482452393, + 0.01137480791658163, + 0.008855181746184826, + 0.013229692354798317, + 0.007540818303823471, + 0.00802829209715128, + 0.004373303148895502, + 0.0048398785293102264, + 0.003812826471403241, + 0.003365874756127596, + 0.0048540630377829075, + 0.003376665757969022, + 0.005711468867957592, + 0.002765408717095852, + 0.0032281423918902874, + 0.0025811458472162485, + 0.0031092078424990177, + 0.0036204056814312935, + 0.0031075195875018835, + 0.003753037890419364, + 0.00565909082069993, + 0.018413245677947998, + 0.007896143943071365, + 0.013741252943873405, + 0.00895425770431757, + 0.012735115364193916, + 0.011727371253073215, + 0.01314615923911333, + 0.010001708753407001, + 0.0026081877294927835, + 0.002393194241449237, + 0.0030623541679233313, + 0.0041042715311050415, + 0.0021036032121628523, + 0.0015249449061229825, + 0.0030397309456020594, + 0.0013101196382194757, + 0.014365037903189659, + 0.01983526349067688, + 0.01294887438416481, + 0.014912189915776253, + 0.012717511504888535, + 0.00781800877302885, + 0.02208556793630123, + 0.01374181266874075, + 0.0991353988647461, + 0.009950687177479267, + 0.1297513097524643, + 0.019617460668087006, + 0.013416789472103119, + 0.007437761407345533, + 0.017134590074419975, + 0.012447967194020748, + 0.009915122762322426, + 0.013484761118888855, + 0.018953582271933556, + 0.008841986767947674, + 0.012551737949252129, + 0.007397753652185202, + 0.014876527711749077, + 0.011600797064602375, + 0.004909535404294729, + 0.0068817962892353535, + 0.0057647801004350185, + 0.005803235340863466, + 0.005957797635346651, + 0.03099088743329048, + 0.005411725025624037, + 0.02730250358581543, + 0.01600547321140766, + 0.03157570958137512, + 0.01445505116134882, + 0.01180290337651968, + 0.009617040865123272, + 0.01016623992472887, + 0.016860641539096832, + 0.020916951820254326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.01608162373304367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067411623895168304 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.0067411623895168304, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067411623895168304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12144023180007935, + 0.09916278719902039, + 0.05953352898359299, + 0.058659955859184265, + 0.14912451803684235, + 0.14751797914505005, + 0.16118226945400238, + 0.13170474767684937, + 0.027384303510189056, + 0.04428970441222191 + ], + "credit_map": [ + 0.007913932204246521, + 0.013132755644619465, + 0.007652726955711842, + 0.025882689282298088, + 0.01759106107056141, + 0.023342400789260864, + 0.016562292352318764, + 0.00936238281428814, + 0.014184370636940002, + 0.012717542238533497, + 0.010652024298906326, + 0.012673430144786835, + 0.01054403930902481, + 0.011844711378216743, + 0.012751332484185696, + 0.013795336708426476, + 0.005626843776553869, + 0.008154438808560371, + 0.005414282437413931, + 0.008045933209359646, + 0.007578627206385136, + 0.005669096950441599, + 0.0082984808832407, + 0.010745828039944172, + 0.0055709402076900005, + 0.005242970306426287, + 0.006498374510556459, + 0.010531436651945114, + 0.007237010635435581, + 0.007127398159354925, + 0.007983732037246227, + 0.008468091487884521, + 0.01900705322623253, + 0.02432970330119133, + 0.024237288162112236, + 0.021960502490401268, + 0.019656987860798836, + 0.011486413888633251, + 0.01676061376929283, + 0.011685969308018684, + 0.018050285056233406, + 0.01617860421538353, + 0.016507091000676155, + 0.019038496538996696, + 0.025723999366164207, + 0.016357557848095894, + 0.019472312182188034, + 0.01618962548673153, + 0.05060315504670143, + 0.006953299045562744, + 0.06153425574302673, + 0.009345403872430325, + 0.010117636062204838, + 0.007254104129970074, + 0.00900221150368452, + 0.006372205913066864, + 0.016835913062095642, + 0.019449785351753235, + 0.017453506588935852, + 0.010560620576143265, + 0.018755357712507248, + 0.018876727670431137, + 0.01527247205376625, + 0.01450036745518446, + 0.002225670497864485, + 0.004696924705058336, + 0.0033688158728182316, + 0.0036718184128403664, + 0.0022887717932462692, + 0.004195152781903744, + 0.003655662527307868, + 0.0032814862206578255, + 0.004742041230201721, + 0.009763251058757305, + 0.004645101726055145, + 0.005200650542974472, + 0.004447883926331997, + 0.006034014280885458, + 0.005347476806491613, + 0.004109285771846771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.0067411623895168304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1782107651233673, + 0.026028402149677277, + 0.04593927785754204, + 0.08336606621742249, + 0.054068971425294876, + 0.08074191957712173, + 0.09559950232505798, + 0.16873206198215485, + 0.1384427845478058, + 0.12887020409107208 + ], + "credit_map": [ + 0.010869123041629791, + 0.0165280532091856, + 0.010455631650984287, + 0.03451438248157501, + 0.018689941614866257, + 0.05041103810071945, + 0.018314728513360023, + 0.018427859991788864, + 0.003756066085770726, + 0.0032631142530590296, + 0.00206409371457994, + 0.0025343159213662148, + 0.00318955909460783, + 0.003581216325983405, + 0.004371961113065481, + 0.00326807564124465, + 0.004490335937589407, + 0.011224654503166676, + 0.0045233070850372314, + 0.004758836701512337, + 0.006399103440344334, + 0.004566394258290529, + 0.005251897964626551, + 0.004724746569991112, + 0.02388608641922474, + 0.006776801310479641, + 0.013316532596945763, + 0.005830833688378334, + 0.005915435962378979, + 0.0069839307107031345, + 0.007782379165291786, + 0.012874070554971695, + 0.006829258985817432, + 0.007767924573272467, + 0.005554327741265297, + 0.006592442747205496, + 0.008976963348686695, + 0.005579331889748573, + 0.008363688364624977, + 0.00440503703430295, + 0.009560367092490196, + 0.012557846494019032, + 0.00718180276453495, + 0.0095912991091609, + 0.01325513981282711, + 0.0055243042297661304, + 0.014430355280637741, + 0.008640804328024387, + 0.011072158813476562, + 0.01116949412971735, + 0.015607846900820732, + 0.011997148394584656, + 0.015756404027342796, + 0.005746488459408283, + 0.01898590661585331, + 0.0052640545181930065, + 0.019011061638593674, + 0.03244401142001152, + 0.023436972871422768, + 0.01812327466905117, + 0.019058160483837128, + 0.013603424653410912, + 0.02976566180586815, + 0.0132894953712821, + 0.01049885991960764, + 0.020537784323096275, + 0.007903202436864376, + 0.008747907355427742, + 0.009062136523425579, + 0.03862215206027031, + 0.008170169778168201, + 0.034900568425655365, + 0.021354159340262413, + 0.02334577403962612, + 0.014135480858385563, + 0.011929865926504135, + 0.006280587054789066, + 0.012625272385776043, + 0.01494623627513647, + 0.02425282448530197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.011076509952545166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368634358048439 + ], + "opposite_logprob_tokens": [ + -3.408686399459839 + ], + "logprob_reward": -0.03368634358048439, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368634358048439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08523723483085632, + 0.061608389019966125, + 0.030429702252149582, + 0.04853881523013115, + 0.07695811986923218, + 0.15589183568954468, + 0.31581994891166687, + 0.05985332652926445, + 0.028256341814994812, + 0.13740628957748413 + ], + "credit_map": [ + 0.0059744566679000854, + 0.005944821052253246, + 0.012441884726285934, + 0.026595590636134148, + 0.0068318406119942665, + 0.01439182460308075, + 0.0064271921291947365, + 0.006629633717238903, + 0.010013263672590256, + 0.005131210666149855, + 0.018863297998905182, + 0.005327480845153332, + 0.0033007818274199963, + 0.004965885076671839, + 0.004470892250537872, + 0.009535574354231358, + 0.0027954925317317247, + 0.0026293389964848757, + 0.004846012685447931, + 0.0056795887649059296, + 0.002743540331721306, + 0.0027707687113434076, + 0.003526100190356374, + 0.005438858177512884, + 0.005434297490864992, + 0.003529961919412017, + 0.006951593328267336, + 0.006090646144002676, + 0.005625170189887285, + 0.004950218368321657, + 0.005672637838870287, + 0.01028428878635168, + 0.009258966892957687, + 0.014837084338068962, + 0.008813640102744102, + 0.012104149907827377, + 0.008580394089221954, + 0.0049675763584673405, + 0.012613710947334766, + 0.005782595835626125, + 0.018070930615067482, + 0.016274843364953995, + 0.012074287049472332, + 0.017247678712010384, + 0.025374628603458405, + 0.013800371438264847, + 0.029742036014795303, + 0.0233070719987154, + 0.09194380789995193, + 0.0120396688580513, + 0.12508121132850647, + 0.020515337586402893, + 0.021840928122401237, + 0.012299862690269947, + 0.017190944403409958, + 0.014908193610608578, + 0.008492589928209782, + 0.01144357118755579, + 0.007273446302860975, + 0.005863633006811142, + 0.0063633061945438385, + 0.008848116733133793, + 0.00608452782034874, + 0.005484134890139103, + 0.0031430867966264486, + 0.004365194588899612, + 0.0031772858928889036, + 0.0037300640251487494, + 0.002328541362658143, + 0.003994697704911232, + 0.004643146879971027, + 0.0028743231669068336, + 0.020640507340431213, + 0.021831192076206207, + 0.015023577027022839, + 0.013226933777332306, + 0.009426014497876167, + 0.016614440828561783, + 0.01799919269979, + 0.02264443226158619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.03368634358048439 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot steadily follows a lone, plump pigeon as it struts confidently across a vast, gleaming floor of polished cream marble. The camera moves parallel to the bird, maintaining a steady composition that emphasizes its solitary journey. High above, a magnificent, multi-tiered crystal chandelier hangs from the ornate ceiling, casting a warm, brilliant light that bathes the entire scene. The pigeon's glossy gray feathers are spotless, and with each deliberate, rhythmic sway of its head, the iridescent feathers on its neck catch the pouring light, shimmering in a dazzling spectrum of vibrant greens, deep purples, and cool blues. Its chest puffs out with a sense of pride, its tiny pink feet firmly planted on the cool, reflective surface. The polished marble acts like a perfect mirror, creating soft, elongated reflections of the pigeon and the glittering chandelier, enhancing the surreal and quiet atmosphere of solitary elegance within the grand, empty hall.", + "seed": 7571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_14/index_3512_seed_7571_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014899095334112645, + "vlm_logprob": -0.014899095334112645, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.014899095334112645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034464623779058456 + ], + "opposite_logprob_tokens": [ + -3.5344645977020264 + ], + "logprob_reward": -0.034464623779058456, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.034464623779058456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1829565167427063, + 0.1599617451429367, + 0.23721300065517426, + 0.08378828316926956, + 0.0651622861623764, + 0.04899987950921059, + 0.10636476427316666, + 0.016803724691271782, + 0.022367777302861214, + 0.07638201117515564 + ], + "credit_map": [ + 0.016235915943980217, + 0.01920783333480358, + 0.014400068670511246, + 0.02469765767455101, + 0.03290995582938194, + 0.009429318830370903, + 0.037753935903310776, + 0.02832183428108692, + 0.01516555342823267, + 0.021850988268852234, + 0.015779105946421623, + 0.021815303713083267, + 0.024811824783682823, + 0.019617628306150436, + 0.025847669690847397, + 0.015073686838150024, + 0.04487761855125427, + 0.02121502347290516, + 0.0606922022998333, + 0.017847515642642975, + 0.028593406081199646, + 0.022281045094132423, + 0.028434373438358307, + 0.013271825388073921, + 0.009603637270629406, + 0.014812593348324299, + 0.022756580263376236, + 0.005410976242274046, + 0.008255732245743275, + 0.010041940025985241, + 0.007293289992958307, + 0.005613536108285189, + 0.008781224489212036, + 0.006097409874200821, + 0.007436381187289953, + 0.004582117311656475, + 0.010605715215206146, + 0.01267825998365879, + 0.0077399020083248615, + 0.007241274230182171, + 0.0154251204803586, + 0.004970951937139034, + 0.004132775589823723, + 0.00259247119538486, + 0.003727839095517993, + 0.006596188526600599, + 0.003190048737451434, + 0.008364482782781124, + 0.015802796930074692, + 0.011851008050143719, + 0.019783882424235344, + 0.009092969819903374, + 0.009346270002424717, + 0.01337374746799469, + 0.015741847455501556, + 0.011372238397598267, + 0.002052696654573083, + 0.001852197339758277, + 0.0012723709223791957, + 0.0012870726641267538, + 0.0033473039511591196, + 0.003035171190276742, + 0.002071322174742818, + 0.0018855889793485403, + 0.002782962052151561, + 0.0029387385584414005, + 0.0019126356346532702, + 0.0020688436925411224, + 0.00411696033552289, + 0.0027115410193800926, + 0.0034311958588659763, + 0.002404902596026659, + 0.008383942767977715, + 0.01127983070909977, + 0.006406497210264206, + 0.0089258486405015, + 0.014382680878043175, + 0.009742082096636295, + 0.011149466969072819, + 0.006111661903560162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.034464623779058456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658249862492085 + ], + "opposite_logprob_tokens": [ + -4.758658409118652 + ], + "logprob_reward": -0.008658249862492085, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658249862492085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1108168438076973, + 0.20398591458797455, + 0.20606555044651031, + 0.07519703358411789, + 0.09304329752922058, + 0.1254832148551941, + 0.04785972833633423, + 0.023029139265418053, + 0.05253466218709946, + 0.061984628438949585 + ], + "credit_map": [ + 0.011225077323615551, + 0.011232498101890087, + 0.009690689854323864, + 0.011912132613360882, + 0.02366326004266739, + 0.005165113136172295, + 0.019918950274586678, + 0.018009120598435402, + 0.019767025485634804, + 0.023065967485308647, + 0.017516106367111206, + 0.02434597909450531, + 0.027832742780447006, + 0.026903335005044937, + 0.0404614694416523, + 0.024093298241496086, + 0.037310466170310974, + 0.018313001841306686, + 0.05098004639148712, + 0.01582123339176178, + 0.025558259338140488, + 0.019698653370141983, + 0.026347214356064796, + 0.012036649510264397, + 0.008625272661447525, + 0.011155050247907639, + 0.010353420861065388, + 0.006274325307458639, + 0.010963384993374348, + 0.011666867882013321, + 0.010791005566716194, + 0.005367704201489687, + 0.010330102406442165, + 0.010157441720366478, + 0.01159739401191473, + 0.007632005959749222, + 0.01773594506084919, + 0.012092381715774536, + 0.015711799263954163, + 0.007786229252815247, + 0.044374894350767136, + 0.009711850434541702, + 0.009401962161064148, + 0.004857093095779419, + 0.007287684828042984, + 0.020417487248778343, + 0.007839879021048546, + 0.02159236930310726, + 0.006896145176142454, + 0.005401110276579857, + 0.006084095221012831, + 0.0035719899460673332, + 0.006635277532041073, + 0.008995244279503822, + 0.004947586916387081, + 0.0053282808512449265, + 0.004128171131014824, + 0.0031661088578402996, + 0.0015831916825845838, + 0.001472370931878686, + 0.0029525943100452423, + 0.004647504538297653, + 0.0022796127013862133, + 0.0027995833661407232, + 0.005535933189094067, + 0.007032405119389296, + 0.0034835641272366047, + 0.005013918504118919, + 0.011450657621026039, + 0.006825129967182875, + 0.009248956106603146, + 0.0039440966211259365, + 0.007771838456392288, + 0.00897669605910778, + 0.0046431198716163635, + 0.007104826159775257, + 0.010456540621817112, + 0.007536584977060556, + 0.009494632482528687, + 0.0060003940016031265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.008658249862492085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640880040824413 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640880040824413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640880040824413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18957898020744324, + 0.2515769302845001, + 0.0570896714925766, + 0.034148164093494415, + 0.09448088705539703, + 0.06260160356760025, + 0.14319100975990295, + 0.07112840563058853, + 0.051658425480127335, + 0.044545941054821014 + ], + "credit_map": [ + 0.02683527022600174, + 0.027088185772299767, + 0.018901493400335312, + 0.02266196720302105, + 0.021585851907730103, + 0.014381781220436096, + 0.032834477722644806, + 0.025289950892329216, + 0.027889110147953033, + 0.03563489392399788, + 0.031026778742671013, + 0.028694165870547295, + 0.033058878034353256, + 0.03150470182299614, + 0.04395468533039093, + 0.019813720136880875, + 0.009812084957957268, + 0.00562530430033803, + 0.011894172988831997, + 0.005358108319342136, + 0.006921015214174986, + 0.005327727645635605, + 0.006975438445806503, + 0.005175823345780373, + 0.005980846006423235, + 0.006890673656016588, + 0.004075170960277319, + 0.0024932073429226875, + 0.004097344353795052, + 0.004121046047657728, + 0.00334688788279891, + 0.003142986912280321, + 0.013357769697904587, + 0.013851240277290344, + 0.01923852413892746, + 0.004383496008813381, + 0.009610987268388271, + 0.009645485319197178, + 0.008279724046587944, + 0.016113657504320145, + 0.019865145906805992, + 0.005490955896675587, + 0.005488045047968626, + 0.003070390783250332, + 0.004397915676236153, + 0.009041896089911461, + 0.0038631781935691833, + 0.011384069919586182, + 0.02587790973484516, + 0.016161533072590828, + 0.020660176873207092, + 0.008175419643521309, + 0.01175764948129654, + 0.032694026827812195, + 0.010407756082713604, + 0.0174565352499485, + 0.010950368829071522, + 0.010222015902400017, + 0.005568780470639467, + 0.005135339684784412, + 0.010143334977328777, + 0.015040542930364609, + 0.005958306137472391, + 0.008109718561172485, + 0.008501129224896431, + 0.007011465262621641, + 0.005762104410678148, + 0.003789336886256933, + 0.007448974996805191, + 0.006780385039746761, + 0.0051283021457493305, + 0.007236729376018047, + 0.003858455689623952, + 0.005113424267619848, + 0.007016039919108152, + 0.003530398942530155, + 0.004727140534669161, + 0.006233060732483864, + 0.007968894205987453, + 0.006098524201661348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.004640880040824413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542181648314 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542181648314, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542181648314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2418915182352066, + 0.09426015615463257, + 0.18850038945674896, + 0.06131790578365326, + 0.07250398397445679, + 0.05693735554814339, + 0.07055561244487762, + 0.030152585357427597, + 0.05035225674510002, + 0.13352827727794647 + ], + "credit_map": [ + 0.026186881586909294, + 0.022344596683979034, + 0.01815163344144821, + 0.03330839052796364, + 0.054434698075056076, + 0.018147645518183708, + 0.04325683414936066, + 0.026060789823532104, + 0.010626798495650291, + 0.014785713516175747, + 0.009484819136559963, + 0.01039316039532423, + 0.019810441881418228, + 0.007479023188352585, + 0.013376642018556595, + 0.008303539827466011, + 0.03758411481976509, + 0.017409225925803185, + 0.052938997745513916, + 0.014340389519929886, + 0.019799064844846725, + 0.015351136215031147, + 0.019958719611167908, + 0.011118731461465359, + 0.006273775361478329, + 0.007650906685739756, + 0.009230835363268852, + 0.005636417772620916, + 0.0084042027592659, + 0.009308811277151108, + 0.007578378077596426, + 0.007234570104628801, + 0.009261631406843662, + 0.006317367311567068, + 0.017445597797632217, + 0.003730656346306205, + 0.007083968259394169, + 0.009536287747323513, + 0.005504723638296127, + 0.01362374797463417, + 0.013378503732383251, + 0.006507371086627245, + 0.005806303583085537, + 0.004640217870473862, + 0.005177767015993595, + 0.008037783205509186, + 0.005109088961035013, + 0.008280315436422825, + 0.009217489510774612, + 0.007000609766691923, + 0.009027005173265934, + 0.004581792280077934, + 0.012238086201250553, + 0.014044919051229954, + 0.006968711037188768, + 0.007476998958736658, + 0.005797667894512415, + 0.0028807397466152906, + 0.0022431162651628256, + 0.001992849400267005, + 0.002912679919973016, + 0.00675397040322423, + 0.003086024895310402, + 0.0044855340383946896, + 0.006401962600648403, + 0.0067329686135053635, + 0.004125614650547504, + 0.005497629754245281, + 0.008001703768968582, + 0.006957741919904947, + 0.008384810760617256, + 0.004249819088727236, + 0.012479403987526894, + 0.012557399459183216, + 0.013230632059276104, + 0.010645180009305477, + 0.032661519944667816, + 0.016747938469052315, + 0.02286207489669323, + 0.01234412007033825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.012542181648314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189541339874268 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189541339874268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189541339874268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10795344412326813, + 0.2005457580089569, + 0.05206742137670517, + 0.07216226309537888, + 0.11509538441896439, + 0.08348652720451355, + 0.1556561291217804, + 0.03275710344314575, + 0.11090879142284393, + 0.06936716288328171 + ], + "credit_map": [ + 0.010138150304555893, + 0.013103747740387917, + 0.010240146890282631, + 0.014600303024053574, + 0.014753401279449463, + 0.00689878361299634, + 0.020676633343100548, + 0.01754228211939335, + 0.020891062915325165, + 0.02038015052676201, + 0.022032832726836205, + 0.023125849664211273, + 0.024788163602352142, + 0.029033426195383072, + 0.034658562391996384, + 0.02563570626080036, + 0.008507659658789635, + 0.007248178124427795, + 0.012063983827829361, + 0.004469969775527716, + 0.00537584163248539, + 0.004365326836705208, + 0.006103652063757181, + 0.00393280852586031, + 0.009970220737159252, + 0.012000621296465397, + 0.017603164538741112, + 0.004800367169082165, + 0.006339854095131159, + 0.006286421325057745, + 0.006080988794565201, + 0.009080625139176846, + 0.012103954330086708, + 0.00956164300441742, + 0.01619059219956398, + 0.010701309889554977, + 0.02074776031076908, + 0.01343410275876522, + 0.019490057602524757, + 0.012865966185927391, + 0.02091738022863865, + 0.009310378693044186, + 0.007086961530148983, + 0.0066449823789298534, + 0.007505714427679777, + 0.012246194295585155, + 0.007726216223090887, + 0.012048694305121899, + 0.025958647951483727, + 0.025588948279619217, + 0.015608800575137138, + 0.011579693295061588, + 0.015511990524828434, + 0.03518739715218544, + 0.01080482080578804, + 0.01541583240032196, + 0.005616024136543274, + 0.0030576891731470823, + 0.0022190362215042114, + 0.002204743679612875, + 0.007499893195927143, + 0.005118001252412796, + 0.0032480189111083746, + 0.00379369524307549, + 0.013986518606543541, + 0.016523655503988266, + 0.009938642382621765, + 0.009136886335909367, + 0.022285645827651024, + 0.014069728553295135, + 0.014372086152434349, + 0.010595635510981083, + 0.009424560703337193, + 0.009290625341236591, + 0.007894832640886307, + 0.006558326072990894, + 0.009393933229148388, + 0.008533854968845844, + 0.009835085831582546, + 0.008435949683189392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.014189541339874268 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot steadily follows a lone, plump pigeon as it struts confidently across a vast, gleaming floor of polished cream marble. The camera moves parallel to the bird, maintaining a steady composition that emphasizes its solitary journey. High above, a magnificent, multi-tiered crystal chandelier hangs from the ornate ceiling, casting a warm, brilliant light that bathes the entire scene. The pigeon's glossy gray feathers are spotless, and with each deliberate, rhythmic sway of its head, the iridescent feathers on its neck catch the pouring light, shimmering in a dazzling spectrum of vibrant greens, deep purples, and cool blues. Its chest puffs out with a sense of pride, its tiny pink feet firmly planted on the cool, reflective surface. The polished marble acts like a perfect mirror, creating soft, elongated reflections of the pigeon and the glittering chandelier, enhancing the surreal and quiet atmosphere of solitary elegance within the grand, empty hall.", + "seed": 7572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_14/index_3512_seed_7572_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012955154199153185, + "vlm_logprob": -0.012955154199153185, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.012955154199153185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010462789796292782 + ], + "opposite_logprob_tokens": [ + -4.885462760925293 + ], + "logprob_reward": -0.010462789796292782, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010462789796292782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1300184726715088, + 0.15092328190803528, + 0.31229734420776367, + 0.09952298551797867, + 0.07152753323316574, + 0.050073422491550446, + 0.07696906477212906, + 0.03215945139527321, + 0.022092366591095924, + 0.054416123777627945 + ], + "credit_map": [ + 0.013571619056165218, + 0.03716927394270897, + 0.02246045507490635, + 0.015055102296173573, + 0.013278492726385593, + 0.005312660709023476, + 0.012251640670001507, + 0.010919217951595783, + 0.014372039586305618, + 0.018965207040309906, + 0.030404238030314445, + 0.027710167691111565, + 0.01789429411292076, + 0.011120887473225594, + 0.021231843158602715, + 0.00922459363937378, + 0.02428589016199112, + 0.05417345464229584, + 0.10498201847076416, + 0.023922082036733627, + 0.02920638769865036, + 0.02276894636452198, + 0.03488980233669281, + 0.018068764358758926, + 0.012955526821315289, + 0.013307212851941586, + 0.010646602138876915, + 0.023022741079330444, + 0.008440141566097736, + 0.007286223117262125, + 0.017241096124053, + 0.006623433902859688, + 0.010754261165857315, + 0.008578240871429443, + 0.0088815251365304, + 0.016523921862244606, + 0.004735326860100031, + 0.007741236127912998, + 0.0059114545583724976, + 0.008401568047702312, + 0.006359927821904421, + 0.005347547121345997, + 0.004652813542634249, + 0.01121599692851305, + 0.005302020348608494, + 0.007015688810497522, + 0.0034754883963614702, + 0.006703934632241726, + 0.006999161094427109, + 0.007238752208650112, + 0.014057298190891743, + 0.014391439966857433, + 0.010216115973889828, + 0.009859764948487282, + 0.007743918802589178, + 0.006462608464062214, + 0.0026895070914179087, + 0.0034920736216008663, + 0.0055544557981193066, + 0.004410012625157833, + 0.00420428765937686, + 0.004664269741624594, + 0.003418469103053212, + 0.003726373193785548, + 0.0019053883152082562, + 0.001491179340519011, + 0.0021778519731014967, + 0.0036736465990543365, + 0.002514452440664172, + 0.0036356004420667887, + 0.00386661640368402, + 0.0028276280499994755, + 0.0033339231740683317, + 0.004307902418076992, + 0.006717139855027199, + 0.01197474543005228, + 0.007018292788416147, + 0.009099727496504784, + 0.007343968376517296, + 0.00462042260915041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.010462789796292782 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744240876287222 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744240876287222, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744240876287222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1085117980837822, + 0.07158563286066055, + 0.2957717478275299, + 0.17072054743766785, + 0.08773835003376007, + 0.06041713431477547, + 0.05526558682322502, + 0.07945362478494644, + 0.032443251460790634, + 0.03809233382344246 + ], + "credit_map": [ + 0.01624867506325245, + 0.026346061378717422, + 0.015701238065958023, + 0.009413736872375011, + 0.013693041168153286, + 0.0085685970261693, + 0.012117517180740833, + 0.006422939710319042, + 0.00924368854612112, + 0.008133204653859138, + 0.009310823865234852, + 0.008744006045162678, + 0.011993492022156715, + 0.008574160747230053, + 0.010180394165217876, + 0.0054058656096458435, + 0.023111287504434586, + 0.06824421137571335, + 0.10985101759433746, + 0.022078143432736397, + 0.020679665729403496, + 0.0123075470328331, + 0.02399364858865738, + 0.015506220050156116, + 0.016138415783643723, + 0.015084837563335896, + 0.028471248224377632, + 0.027443932369351387, + 0.012623888440430164, + 0.01377486065030098, + 0.04370693489909172, + 0.013476422987878323, + 0.018557000905275345, + 0.007638993207365274, + 0.021823545917868614, + 0.009221145883202553, + 0.00566058699041605, + 0.008137041702866554, + 0.0054002078250050545, + 0.011299831792712212, + 0.006748686544597149, + 0.005181086249649525, + 0.0063988519832491875, + 0.011859315447509289, + 0.0067328340373933315, + 0.008500608615577221, + 0.004761058371514082, + 0.010234689339995384, + 0.006381182465702295, + 0.005737401079386473, + 0.006980708800256252, + 0.009127620607614517, + 0.008731003850698471, + 0.007816550321877003, + 0.004809185396879911, + 0.005681931506842375, + 0.008592786267399788, + 0.008313159458339214, + 0.012603618204593658, + 0.0118462098762393, + 0.00824386440217495, + 0.01016971841454506, + 0.009323159232735634, + 0.010361114516854286, + 0.002075038617476821, + 0.002048072637990117, + 0.003511275863274932, + 0.004962348844856024, + 0.003480432787910104, + 0.006173275411128998, + 0.006063728127628565, + 0.004129079636186361, + 0.0034104231745004654, + 0.00507639953866601, + 0.0046743350103497505, + 0.007932034321129322, + 0.003694578306749463, + 0.005514588672667742, + 0.004761347081512213, + 0.0030286284163594246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.006744240876287222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095619544386864 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095619544386864, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095619544386864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2990528643131256, + 0.13632510602474213, + 0.2623012065887451, + 0.07980360090732574, + 0.04791303724050522, + 0.04826035350561142, + 0.024648720398545265, + 0.045658014714717865, + 0.02507644146680832, + 0.03096061572432518 + ], + "credit_map": [ + 0.04226122424006462, + 0.07289742678403854, + 0.03401761129498482, + 0.02913796715438366, + 0.03997684642672539, + 0.02305474691092968, + 0.03710467368364334, + 0.02060236781835556, + 0.024768367409706116, + 0.015323339030146599, + 0.028824904933571815, + 0.014405637979507446, + 0.013080930337309837, + 0.016500601544976234, + 0.014990000985562801, + 0.008431321009993553, + 0.02474001981317997, + 0.053243059664964676, + 0.09246837347745895, + 0.018881233409047127, + 0.015197808854281902, + 0.02314789406955242, + 0.019725320860743523, + 0.014897523447871208, + 0.01187414675951004, + 0.008163546212017536, + 0.013535949401557446, + 0.012572799809277058, + 0.004472149070352316, + 0.005366825964301825, + 0.017539476975798607, + 0.006278710905462503, + 0.006375416647642851, + 0.00584723986685276, + 0.004759531002491713, + 0.008013058453798294, + 0.006401460617780685, + 0.0075266542844474316, + 0.004311750642955303, + 0.004677933175116777, + 0.0072863055393099785, + 0.005460826680064201, + 0.005154037848114967, + 0.008079229854047298, + 0.006490782834589481, + 0.007450720760971308, + 0.003817932913079858, + 0.004520517773926258, + 0.0023952978663146496, + 0.0025005226489156485, + 0.0031218386720865965, + 0.003200247883796692, + 0.004962966311722994, + 0.0035751904360949993, + 0.0023474888876080513, + 0.002545170718804002, + 0.004278581589460373, + 0.006709069479256868, + 0.007121219765394926, + 0.005679733119904995, + 0.005477484315633774, + 0.006054361816495657, + 0.0048076496459543705, + 0.005529917776584625, + 0.002117918338626623, + 0.0017502300906926394, + 0.003161017317324877, + 0.004263115581125021, + 0.002273437799885869, + 0.0036277612671256065, + 0.004001370165497065, + 0.003881592070683837, + 0.0029334640130400658, + 0.0029668696224689484, + 0.003841569647192955, + 0.005506626330316067, + 0.003965452313423157, + 0.0046135252341628075, + 0.004136877600103617, + 0.0029962328262627125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004095619544386864 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336986780166626 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.0336986780166626, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336986780166626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18385277688503265, + 0.27696460485458374, + 0.169797882437706, + 0.1186400055885315, + 0.0364367738366127, + 0.07308412343263626, + 0.031416721642017365, + 0.03698451444506645, + 0.03609561175107956, + 0.036727018654346466 + ], + "credit_map": [ + 0.022869128733873367, + 0.030144844204187393, + 0.025029370561242104, + 0.018953774124383926, + 0.026968013495206833, + 0.013982641510665417, + 0.031283050775527954, + 0.014621945098042488, + 0.061693813651800156, + 0.03693770617246628, + 0.06481746584177017, + 0.03976214677095413, + 0.014667866751551628, + 0.01506028138101101, + 0.029854536056518555, + 0.014170760288834572, + 0.0202217735350132, + 0.039681192487478256, + 0.03498025983572006, + 0.012868979945778847, + 0.022452950477600098, + 0.011304724961519241, + 0.01966218091547489, + 0.00862580444663763, + 0.01852761022746563, + 0.014534599147737026, + 0.016244810074567795, + 0.020927121862769127, + 0.009174922481179237, + 0.007393422070890665, + 0.02392846904695034, + 0.007909025996923447, + 0.006150979083031416, + 0.005053739529103041, + 0.003358036745339632, + 0.0056257895193994045, + 0.004240910056978464, + 0.005901608616113663, + 0.003269052132964134, + 0.002836655592545867, + 0.00870403740555048, + 0.008738583885133266, + 0.00787814799696207, + 0.009738325141370296, + 0.012675981037318707, + 0.00852159783244133, + 0.007755696773529053, + 0.00907174777239561, + 0.002844578120857477, + 0.002711561042815447, + 0.004722814075648785, + 0.003842582693323493, + 0.006626009475439787, + 0.004375683609396219, + 0.00339689152315259, + 0.00289659621194005, + 0.0036211793776601553, + 0.0038231939543038607, + 0.0056590368039906025, + 0.005218622274696827, + 0.004261537920683622, + 0.005408591125160456, + 0.003943825606256723, + 0.005048523657023907, + 0.003085741074755788, + 0.0034000559244304895, + 0.004270602483302355, + 0.007690575905144215, + 0.003148767165839672, + 0.005245457869023085, + 0.0048528085462749004, + 0.004401597194373608, + 0.002172360196709633, + 0.003667652141302824, + 0.004651427268981934, + 0.006840048357844353, + 0.004852366168051958, + 0.005602759309113026, + 0.00566092599183321, + 0.003279475262388587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0336986780166626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, plump pigeon with glossy gray feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pigeon strutting confidently with a rhythmic head sway and puffed chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast hall with a polished cream marble floor and a multi-tiered crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pigeon's neck feathers shimmer with iridescent greens, purples, and blues due to the lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle tracking shot moving parallel to the pigeon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977444276213646 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.00977444276213646, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977444276213646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.155364528298378, + 0.28614750504493713, + 0.19283238053321838, + 0.1221262589097023, + 0.03963980823755264, + 0.05986909940838814, + 0.0483786016702652, + 0.042027100920677185, + 0.026171507313847542, + 0.02744324691593647 + ], + "credit_map": [ + 0.019362730905413628, + 0.01904553733766079, + 0.026435235515236855, + 0.01673872023820877, + 0.023397522047162056, + 0.014544188976287842, + 0.020023634657263756, + 0.015816960483789444, + 0.07073898613452911, + 0.03857344016432762, + 0.0662892535328865, + 0.04393448680639267, + 0.01489140372723341, + 0.009168969467282295, + 0.029726073145866394, + 0.012824904173612595, + 0.016543574631214142, + 0.03804625943303108, + 0.07072994858026505, + 0.014506994746625423, + 0.013206364586949348, + 0.014868096448481083, + 0.013391668908298016, + 0.01153948251157999, + 0.01552579365670681, + 0.011814317665994167, + 0.017797719687223434, + 0.0172040443867445, + 0.012659033760428429, + 0.008426465094089508, + 0.02843628264963627, + 0.010262595489621162, + 0.011223172768950462, + 0.004094438161700964, + 0.009942874312400818, + 0.0033951636869460344, + 0.002832908881828189, + 0.003333300817757845, + 0.001772611984051764, + 0.003045337274670601, + 0.008126531727612019, + 0.006530605256557465, + 0.0074607147835195065, + 0.007085456047207117, + 0.005744097754359245, + 0.007880191318690777, + 0.005256555508822203, + 0.011784950271248817, + 0.006303377449512482, + 0.005398194305598736, + 0.008149917237460613, + 0.008576780557632446, + 0.004576866049319506, + 0.00613567465916276, + 0.004887207876890898, + 0.004350584466010332, + 0.0033574544358998537, + 0.006501269992440939, + 0.0053968592546880245, + 0.005614010151475668, + 0.005587729159742594, + 0.005712100304663181, + 0.004546194337308407, + 0.005311483982950449, + 0.0026381262578070164, + 0.0021011463832110167, + 0.003994388971477747, + 0.004563425667583942, + 0.0028508887626230717, + 0.003786335466429591, + 0.003294950583949685, + 0.002942245453596115, + 0.0018221597420051694, + 0.0023900591768324375, + 0.0033846015576273203, + 0.00699868006631732, + 0.003408536547794938, + 0.0034401139710098505, + 0.003585204714909196, + 0.002413890790194273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.00977444276213646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_15/index_tensor([6777, 6777])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_15/index_tensor([6777, 6777])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b63b8fe48aa761f5bb44947d79b1977c02fc7e9d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_15/index_tensor([6777, 6777])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, almost imperceptible medium close-up shot from a slightly high angle captures a middle-aged chef, his entire person radiating an aura of intense concentration. He wears a plain, starched white, double-breasted chef's jacket with contrasting black buttons, and a traditional top hat perfectly perched atop his neatly combed black hair. His brow is slightly furrowed, his eyes narrowed in concentration as he leans over a large, gleaming stainless steel counter, which reflects the bright, even light of a professional kitchen like a perfect mirror. In the softly blurred background, a rack of polished copper pots and pans hangs in an orderly fashion. He holds a perfectly folded, clean white cloth in his hand, carefully wiping the already flawless surface. The movement isn't hurried, but a series of steady, deliberate motions. His arm glides in long, smooth, overlapping strokes, applying just enough pressure to polish the metal, leaving behind a fleeting, streak-free trace of moisture that disappears almost immediately. The scene, dominated by cool whites and metallic silver tones, creates a quiet, almost meditative atmosphere, one of absolute precision and unwavering devotion to craft.", + "seed": 31464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_15/index_6777_seed_31464_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04921934213489294, + "vlm_logprob": -0.04921934213489294, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.04921934213489294, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03310537710785866 + ], + "opposite_logprob_tokens": [ + -4.408105373382568 + ], + "logprob_reward": -0.03310537710785866, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03310537710785866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3099428415298462, + 0.08925675600767136, + 0.051107510924339294, + 0.03784908726811409, + 0.045445144176483154, + 0.06896395981311798, + 0.03710562735795975, + 0.14353495836257935, + 0.054957497864961624, + 0.16183656454086304 + ], + "credit_map": [ + 0.07203223556280136, + 0.06916848570108414, + 0.03072754479944706, + 0.02913517691195011, + 0.02151145599782467, + 0.03352315351366997, + 0.01936381682753563, + 0.03448097035288811, + 0.007237166166305542, + 0.005032489541918039, + 0.01067373901605606, + 0.0058599296025931835, + 0.007189126219600439, + 0.010382759384810925, + 0.02164515294134617, + 0.021236391738057137, + 0.011353007517755032, + 0.005608896724879742, + 0.004984935745596886, + 0.005452306009829044, + 0.004097860772162676, + 0.004045228939503431, + 0.00543462997302413, + 0.010130644775927067, + 0.003792102914303541, + 0.0034892349503934383, + 0.004290958866477013, + 0.005414777901023626, + 0.0037685984279960394, + 0.005143410060554743, + 0.00535119604319334, + 0.0065988097339868546, + 0.0038868170231580734, + 0.0029782068450003862, + 0.006044093053787947, + 0.00542479008436203, + 0.0038203394506126642, + 0.007136837113648653, + 0.00906409416347742, + 0.007089965511113405, + 0.007091010455042124, + 0.009418762288987637, + 0.011547066271305084, + 0.006034495308995247, + 0.006846393458545208, + 0.008550387807190418, + 0.01151636615395546, + 0.007959485054016113, + 0.003199148690328002, + 0.004739839117974043, + 0.004740413744002581, + 0.003667398588731885, + 0.004391223192214966, + 0.005660567432641983, + 0.005194795783609152, + 0.0055122398771345615, + 0.008640793152153492, + 0.0050439476035535336, + 0.024464178830385208, + 0.015828242525458336, + 0.01015709713101387, + 0.022807911038398743, + 0.024560531601309776, + 0.032032258808612823, + 0.005702935624867678, + 0.008299548178911209, + 0.00700061209499836, + 0.009332598187029362, + 0.005157002247869968, + 0.0058228750713169575, + 0.005988984368741512, + 0.0076529462821781635, + 0.013845982030034065, + 0.00846065767109394, + 0.03695620968937874, + 0.021114910021424294, + 0.013910988345742226, + 0.019340533763170242, + 0.02217758819460869, + 0.026029691100120544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.03310537710785866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603740006685257 + ], + "opposite_logprob_tokens": [ + -1.9103740453720093 + ], + "logprob_reward": -0.1603740006685257, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603740006685257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18044932186603546, + 0.19576705992221832, + 0.10002902895212173, + 0.03079814463853836, + 0.13353250920772552, + 0.09235475212335587, + 0.08481846749782562, + 0.06529634445905685, + 0.06660350412130356, + 0.05035080388188362 + ], + "credit_map": [ + 0.018470657989382744, + 0.020947180688381195, + 0.01679375395178795, + 0.019087180495262146, + 0.022923870012164116, + 0.029111243784427643, + 0.02767832577228546, + 0.025437116622924805, + 0.013702642172574997, + 0.015025065280497074, + 0.03555301949381828, + 0.020520295947790146, + 0.026772934943437576, + 0.024086641147732735, + 0.029369303956627846, + 0.03073718585073948, + 0.01956007443368435, + 0.011759103275835514, + 0.011958212591707706, + 0.01100887544453144, + 0.007207750808447599, + 0.011622034944593906, + 0.01003203820437193, + 0.0168809462338686, + 0.0036596874706447124, + 0.002234435174614191, + 0.004470586311072111, + 0.003437664592638612, + 0.003223722567781806, + 0.004400282166898251, + 0.005113180261105299, + 0.004258590284734964, + 0.013738631270825863, + 0.009964408352971077, + 0.01908254250884056, + 0.012690018862485886, + 0.01665264368057251, + 0.024792419746518135, + 0.014890735037624836, + 0.021721126511693, + 0.008938882499933243, + 0.0056075481697916985, + 0.010531088337302208, + 0.007161997724324465, + 0.01536689791828394, + 0.014851010404527187, + 0.012216837145388126, + 0.01768050342798233, + 0.008379791863262653, + 0.00553143210709095, + 0.011363773606717587, + 0.0074630496092140675, + 0.005978984292596579, + 0.011682125739753246, + 0.01673024706542492, + 0.017689071595668793, + 0.005052946507930756, + 0.005199883133172989, + 0.009535613469779491, + 0.007517162710428238, + 0.006752055138349533, + 0.009692653082311153, + 0.00849153008311987, + 0.013054507784545422, + 0.007680248003453016, + 0.004501079209148884, + 0.008105933666229248, + 0.013568305410444736, + 0.006511901970952749, + 0.007113791070878506, + 0.008378768339753151, + 0.010743485763669014, + 0.0041713835671544075, + 0.0031120101921260357, + 0.007251334842294455, + 0.009368984028697014, + 0.0071154022589325905, + 0.004904347006231546, + 0.0060540842823684216, + 0.008373260498046875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.1603740006685257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194829389452934 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014194829389452934, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194829389452934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16851717233657837, + 0.09506399184465408, + 0.10317941755056381, + 0.020856013521552086, + 0.02943492867052555, + 0.06893156468868256, + 0.03564143180847168, + 0.22077377140522003, + 0.041726578027009964, + 0.21587516367435455 + ], + "credit_map": [ + 0.016866590827703476, + 0.013141726143658161, + 0.02552654594182968, + 0.022776860743761063, + 0.010303754359483719, + 0.029187170788645744, + 0.03096499852836132, + 0.019749514758586884, + 0.0073964414186775684, + 0.003871180349960923, + 0.010688630864024162, + 0.006913249380886555, + 0.006723648402839899, + 0.0126796318218112, + 0.023423049598932266, + 0.02336815744638443, + 0.013035009615123272, + 0.010514592751860619, + 0.012531078420579433, + 0.008538547903299332, + 0.007878685370087624, + 0.01444887276738882, + 0.01407762710005045, + 0.022154999896883965, + 0.002131338231265545, + 0.00193133600987494, + 0.0026060836389660835, + 0.0016626520082354546, + 0.00144674233160913, + 0.0034578877966850996, + 0.00386401591822505, + 0.0037559582851827145, + 0.0023498674854636192, + 0.0019392677349969745, + 0.003418315900489688, + 0.0038098497316241264, + 0.0027181811165064573, + 0.00306647177785635, + 0.0066958279348909855, + 0.005437148734927177, + 0.005755301099270582, + 0.006094492971897125, + 0.013814999721944332, + 0.005301774013787508, + 0.007053017150610685, + 0.007625238969922066, + 0.014582575298845768, + 0.008704166859388351, + 0.003047727746888995, + 0.0038601767737418413, + 0.003067371202632785, + 0.007612769026309252, + 0.0025390975642949343, + 0.004747974220663309, + 0.005717541556805372, + 0.005048772320151329, + 0.013756626285612583, + 0.011550107970833778, + 0.03137983754277229, + 0.01728643663227558, + 0.01723516918718815, + 0.035900238901376724, + 0.030127456411719322, + 0.06353789567947388, + 0.003843689104542136, + 0.006119333207607269, + 0.005727851763367653, + 0.006887746509164572, + 0.0028044283390045166, + 0.005520679522305727, + 0.005080129951238632, + 0.005742719396948814, + 0.028216727077960968, + 0.01422322541475296, + 0.04141656309366226, + 0.03874175250530243, + 0.016416281461715698, + 0.023947305977344513, + 0.027114229276776314, + 0.02579907886683941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.014194829389452934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633785881102085 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633785881102085, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633785881102085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10884252935647964, + 0.11460228264331818, + 0.3995280861854553, + 0.08464693278074265, + 0.016304997727274895, + 0.08158475905656815, + 0.024048281833529472, + 0.0864749550819397, + 0.031209267675876617, + 0.052757926285266876 + ], + "credit_map": [ + 0.018095025792717934, + 0.01894763857126236, + 0.010028369724750519, + 0.00886049959808588, + 0.007011731155216694, + 0.013365371152758598, + 0.017858067527413368, + 0.014675828628242016, + 0.012097539380192757, + 0.012725168839097023, + 0.010356541723012924, + 0.008484046906232834, + 0.007122832350432873, + 0.01700066216289997, + 0.018340861424803734, + 0.028474627062678337, + 0.06985725462436676, + 0.029800403863191605, + 0.047217611223459244, + 0.03451237082481384, + 0.032290585339069366, + 0.03736737743020058, + 0.05958133935928345, + 0.08890113234519958, + 0.01836518384516239, + 0.01128318440169096, + 0.008395436219871044, + 0.008283961564302444, + 0.005544391460716724, + 0.012237121351063251, + 0.010776979848742485, + 0.009760678745806217, + 0.0015041378792375326, + 0.0017355754971504211, + 0.0017605285393074155, + 0.0021802897099405527, + 0.0016528121195733547, + 0.0020890887826681137, + 0.0029276367276906967, + 0.0024549278896301985, + 0.007289745844900608, + 0.006085012573748827, + 0.02232070453464985, + 0.007122209295630455, + 0.0056587764993309975, + 0.006044425535947084, + 0.019294602796435356, + 0.007769286632537842, + 0.0030886868480592966, + 0.00256207212805748, + 0.0030286756809800863, + 0.003159663639962673, + 0.0017269772943109274, + 0.0028568312991410494, + 0.003603236051276326, + 0.004022140055894852, + 0.00473714480176568, + 0.003233478171750903, + 0.00888758059591055, + 0.009698765352368355, + 0.00515838572755456, + 0.009793124161660671, + 0.014039719477295876, + 0.030926749110221863, + 0.0027302035596221685, + 0.005171454511582851, + 0.004041531588882208, + 0.003667599055916071, + 0.0026919622905552387, + 0.004843065049499273, + 0.0033959613647311926, + 0.004667491652071476, + 0.004181230440735817, + 0.00327632506377995, + 0.011203496716916561, + 0.008151315152645111, + 0.0033006304875016212, + 0.006458945572376251, + 0.010691486299037933, + 0.005494493059813976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.008633785881102085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978871762752533 + ], + "opposite_logprob_tokens": [ + -3.5297887325286865 + ], + "logprob_reward": -0.02978871762752533, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978871762752533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15022164583206177, + 0.08889851719141006, + 0.0879470631480217, + 0.028392549604177475, + 0.022010397166013718, + 0.11681806296110153, + 0.03943128511309624, + 0.15430879592895508, + 0.01960749179124832, + 0.2923641800880432 + ], + "credit_map": [ + 0.013275952078402042, + 0.016218218952417374, + 0.01692437194287777, + 0.020409373566508293, + 0.008216570131480694, + 0.02648700214922428, + 0.03142312169075012, + 0.017267029732465744, + 0.010661857202649117, + 0.006015392020344734, + 0.00849115289747715, + 0.0061714849434792995, + 0.007365449331700802, + 0.009294066578149796, + 0.020280491560697556, + 0.020618608221411705, + 0.007549822796136141, + 0.00820919219404459, + 0.010681969113647938, + 0.008910111151635647, + 0.006760078482329845, + 0.013109228573739529, + 0.014047015458345413, + 0.018679631873965263, + 0.003152900841087103, + 0.0025216315407305956, + 0.003992442972958088, + 0.002889759372919798, + 0.002694065449759364, + 0.003033893648535013, + 0.006368719041347504, + 0.003739133710041642, + 0.00208368175663054, + 0.001791463466361165, + 0.0035829409025609493, + 0.002024369314312935, + 0.0018651585560292006, + 0.004708955995738506, + 0.003012190805748105, + 0.0029416342731565237, + 0.011136896908283234, + 0.005776440259069204, + 0.03305724263191223, + 0.007520813960582018, + 0.006951810792088509, + 0.01010256540030241, + 0.02746008150279522, + 0.014812201261520386, + 0.0043178037740290165, + 0.0026562046259641647, + 0.006621442269533873, + 0.003987846430391073, + 0.002528657903894782, + 0.004869027994573116, + 0.006726884748786688, + 0.007723414339125156, + 0.00856426265090704, + 0.007223015651106834, + 0.01854148507118225, + 0.013028871268033981, + 0.012996458448469639, + 0.01934140734374523, + 0.022350668907165527, + 0.05226261541247368, + 0.002064284635707736, + 0.001886170357465744, + 0.002245806623250246, + 0.003734576515853405, + 0.0020637272391468287, + 0.0025385047774761915, + 0.002088029868900776, + 0.002986390143632889, + 0.03654957562685013, + 0.018157554790377617, + 0.05671592801809311, + 0.0488247349858284, + 0.01987478695809841, + 0.0341719426214695, + 0.0423099547624588, + 0.03575967624783516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.02978871762752533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, almost imperceptible medium close-up shot from a slightly high angle captures a middle-aged chef, his entire person radiating an aura of intense concentration. He wears a plain, starched white, double-breasted chef's jacket with contrasting black buttons, and a traditional top hat perfectly perched atop his neatly combed black hair. His brow is slightly furrowed, his eyes narrowed in concentration as he leans over a large, gleaming stainless steel counter, which reflects the bright, even light of a professional kitchen like a perfect mirror. In the softly blurred background, a rack of polished copper pots and pans hangs in an orderly fashion. He holds a perfectly folded, clean white cloth in his hand, carefully wiping the already flawless surface. The movement isn't hurried, but a series of steady, deliberate motions. His arm glides in long, smooth, overlapping strokes, applying just enough pressure to polish the metal, leaving behind a fleeting, streak-free trace of moisture that disappears almost immediately. The scene, dominated by cool whites and metallic silver tones, creates a quiet, almost meditative atmosphere, one of absolute precision and unwavering devotion to craft.", + "seed": 31465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_15/index_6777_seed_31465_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07240543030202388, + "vlm_logprob": -0.07240543030202388, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07240543030202388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04392170161008835 + ], + "opposite_logprob_tokens": [ + -4.29392147064209 + ], + "logprob_reward": -0.04392170161008835, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.04392170161008835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05471120774745941, + 0.0180809423327446, + 0.48556458950042725, + 0.010118226520717144, + 0.05571259185671806, + 0.03042065165936947, + 0.02644692361354828, + 0.02563343569636345, + 0.2134188413619995, + 0.07989250123500824 + ], + "credit_map": [ + 0.003443538909777999, + 0.0032204086892306805, + 0.0050928834825754166, + 0.007198636420071125, + 0.007453677710145712, + 0.008346461690962315, + 0.011416980065405369, + 0.008538627065718174, + 0.0014689426170662045, + 0.0013972348533570766, + 0.00147530110552907, + 0.0037507168017327785, + 0.0020895444322377443, + 0.0023171300999820232, + 0.004048821981996298, + 0.0015332517214119434, + 0.018864501267671585, + 0.012897144071757793, + 0.014990399591624737, + 0.19755133986473083, + 0.17113037407398224, + 0.017697665840387344, + 0.038112297654151917, + 0.014320895075798035, + 0.000904285057913512, + 0.001028313534334302, + 0.0016277504619210958, + 0.002008363138884306, + 0.0010497046168893576, + 0.0010640815598890185, + 0.0012759952805936337, + 0.0011597351403906941, + 0.005249583162367344, + 0.005035250913351774, + 0.005556231830269098, + 0.007501171436160803, + 0.009350710548460484, + 0.009271197021007538, + 0.0069959647953510284, + 0.006752489600330591, + 0.002090847585350275, + 0.0028797700069844723, + 0.00318437023088336, + 0.004499515052884817, + 0.0036478801630437374, + 0.004587371833622456, + 0.00641610287129879, + 0.0031147985719144344, + 0.0021460263524204493, + 0.0030557194259017706, + 0.0026417591143399477, + 0.0031859909649938345, + 0.003647960489615798, + 0.0039551337249577045, + 0.004484539385885, + 0.0033297960180789232, + 0.0016056883614510298, + 0.0022379339206963778, + 0.0027391896583139896, + 0.003346771001815796, + 0.005178984720259905, + 0.0028519551269710064, + 0.005468776915222406, + 0.0022041392512619495, + 0.004426798317581415, + 0.002981262980028987, + 0.0036034919321537018, + 0.008455823175609112, + 0.08359894156455994, + 0.010425358079373837, + 0.005700572393834591, + 0.09422662109136581, + 0.005470071919262409, + 0.008927259594202042, + 0.007453771773725748, + 0.007092345040291548, + 0.014139346778392792, + 0.017302287742495537, + 0.008212707936763763, + 0.011294720694422722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.04392170161008835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799127608537674 + ], + "opposite_logprob_tokens": [ + -1.804912805557251 + ], + "logprob_reward": -0.1799127608537674, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799127608537674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05688157305121422, + 0.017026394605636597, + 0.7137711644172668, + 0.0228851530700922, + 0.01883796975016594, + 0.03099161945283413, + 0.058316875249147415, + 0.010745340026915073, + 0.04952941834926605, + 0.021014492958784103 + ], + "credit_map": [ + 0.004099290817975998, + 0.0047118631191551685, + 0.0043990230187773705, + 0.00660778908059001, + 0.011448395438492298, + 0.008460519835352898, + 0.008181319572031498, + 0.008973369374871254, + 0.0010858429595828056, + 0.0013909009285271168, + 0.001413049059920013, + 0.004466440994292498, + 0.002333712065592408, + 0.0020167247857898474, + 0.0025489055551588535, + 0.001770819304510951, + 0.02745654806494713, + 0.018872184678912163, + 0.02109631523489952, + 0.29103949666023254, + 0.25551357865333557, + 0.024411719292402267, + 0.05653427541255951, + 0.01884710043668747, + 0.0009225871763192117, + 0.0025921862106770277, + 0.0019163404358550906, + 0.003109908662736416, + 0.0028397992718964815, + 0.002369406633079052, + 0.003093923907727003, + 0.006041000131517649, + 0.001384112169034779, + 0.0018068533390760422, + 0.0016813535476103425, + 0.0034210234880447388, + 0.003499831072986126, + 0.002986186183989048, + 0.0022706850431859493, + 0.0017879250226542354, + 0.0014958991669118404, + 0.0033400633838027716, + 0.002177163027226925, + 0.003655475564301014, + 0.005985052790492773, + 0.003822196973487735, + 0.00676753930747509, + 0.003748230403289199, + 0.002001327695325017, + 0.002614129101857543, + 0.0035753410775214434, + 0.02284885011613369, + 0.0026538039091974497, + 0.0024536065757274628, + 0.0039358786307275295, + 0.01823394000530243, + 0.0005568726337514818, + 0.00128757965285331, + 0.0008022901020012796, + 0.0013572752941399813, + 0.0018610777333378792, + 0.001995344879105687, + 0.0017665608320385218, + 0.0011183388996869326, + 0.001298516639508307, + 0.0010954263852909207, + 0.0014658353757113218, + 0.0032701052259653807, + 0.017600901424884796, + 0.002954930067062378, + 0.0027839418035000563, + 0.019059760496020317, + 0.0008862126269377768, + 0.001606691861525178, + 0.0017996933311223984, + 0.0028190817683935165, + 0.0037601108197122812, + 0.002943284809589386, + 0.0034679011441767216, + 0.003731516655534506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.1799127608537674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080282740294933 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080282740294933, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080282740294933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037249065935611725, + 0.031203554943203926, + 0.26329731941223145, + 0.023152168840169907, + 0.11542904376983643, + 0.22326694428920746, + 0.0463554821908474, + 0.014824500307440758, + 0.19449584186077118, + 0.050726018846035004 + ], + "credit_map": [ + 0.0033062193542718887, + 0.0025855328422039747, + 0.0034513866994529963, + 0.004975398536771536, + 0.009165126830339432, + 0.003690956626087427, + 0.005721189547330141, + 0.004353255964815617, + 0.001740241888910532, + 0.0020391568541526794, + 0.00222613662481308, + 0.006875049788504839, + 0.003533114679157734, + 0.003968081437051296, + 0.007214196491986513, + 0.0036075778771191835, + 0.011154664680361748, + 0.007773054763674736, + 0.01110301073640585, + 0.10432538390159607, + 0.08989082276821136, + 0.009456769563257694, + 0.020688176155090332, + 0.00890545267611742, + 0.0014558667317032814, + 0.002963598119094968, + 0.0028375391848385334, + 0.0030700231436640024, + 0.002788329031318426, + 0.0034317204263061285, + 0.002584227593615651, + 0.004020866006612778, + 0.006930325645953417, + 0.009130164980888367, + 0.009236457757651806, + 0.022529063746333122, + 0.020481055602431297, + 0.01991073600947857, + 0.015416287817060947, + 0.011794944293797016, + 0.010748052969574928, + 0.026758477091789246, + 0.01697617582976818, + 0.027504581958055496, + 0.03735821321606636, + 0.03141079097986221, + 0.04591790959239006, + 0.026592746376991272, + 0.003953114617615938, + 0.0077758305706083775, + 0.0025095962919294834, + 0.0058810897171497345, + 0.005911102518439293, + 0.004986524116247892, + 0.004565541632473469, + 0.01077268272638321, + 0.0010212979977950454, + 0.0012831899803131819, + 0.0013783889589831233, + 0.0015186052769422531, + 0.0027074222452938557, + 0.002601746702566743, + 0.002573709934949875, + 0.0017401390941813588, + 0.004119107499718666, + 0.0028234387282282114, + 0.0031009565573185682, + 0.006739424075931311, + 0.07671498507261276, + 0.00995649117976427, + 0.005704688839614391, + 0.08533676713705063, + 0.0024334071204066277, + 0.004430601838976145, + 0.0042418548837304115, + 0.006526689510792494, + 0.008389712311327457, + 0.008573749102652073, + 0.008562656119465828, + 0.007567343767732382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.011080282740294933 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194007031619549 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194007031619549, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194007031619549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0490783154964447, + 0.0518210232257843, + 0.16014528274536133, + 0.030537348240613937, + 0.05136450380086899, + 0.05017007142305374, + 0.08750468492507935, + 0.014804708771407604, + 0.4171536862850189, + 0.08742041140794754 + ], + "credit_map": [ + 0.003531003138050437, + 0.0037980922497808933, + 0.006061408203095198, + 0.007296963594853878, + 0.010956508107483387, + 0.0053087701089680195, + 0.00730956019833684, + 0.004816008731722832, + 0.0038596102967858315, + 0.004914138000458479, + 0.004770307801663876, + 0.011878765188157558, + 0.005030965898185968, + 0.005179658532142639, + 0.009518146514892578, + 0.006669425405561924, + 0.009874342940747738, + 0.006397665478289127, + 0.005973907187581062, + 0.060832519084215164, + 0.05233117565512657, + 0.007265914231538773, + 0.012064813636243343, + 0.00540493568405509, + 0.0017247655196115375, + 0.004339803010225296, + 0.003628582926467061, + 0.0072425175458192825, + 0.00290009961463511, + 0.004170237574726343, + 0.003686297917738557, + 0.0028450421523302794, + 0.0041341292671859264, + 0.005501220002770424, + 0.005410065874457359, + 0.010446622967720032, + 0.007151314523071051, + 0.006433326285332441, + 0.00649938965216279, + 0.005788430571556091, + 0.005484491121023893, + 0.0023463198449462652, + 0.005147784948348999, + 0.004470545798540115, + 0.010150552727282047, + 0.005695374216884375, + 0.010544970631599426, + 0.006330026779323816, + 0.009932160377502441, + 0.01876729726791382, + 0.007931985892355442, + 0.013596553355455399, + 0.008860811591148376, + 0.007086700294166803, + 0.009435310959815979, + 0.011893868446350098, + 0.00108940654899925, + 0.0013801301829516888, + 0.0018675813917070627, + 0.002046669367700815, + 0.002611719071865082, + 0.0021454517263919115, + 0.002115005860105157, + 0.0015487432247027755, + 0.0085062924772501, + 0.005900724790990353, + 0.006321599241346121, + 0.015072597190737724, + 0.16481326520442963, + 0.020474908873438835, + 0.01126676145941019, + 0.1847975105047226, + 0.006267928052693605, + 0.009890087880194187, + 0.01001884788274765, + 0.010602040216326714, + 0.010191455483436584, + 0.007789533119648695, + 0.014061540365219116, + 0.018598970025777817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.014194007031619549 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291839927434921 + ], + "opposite_logprob_tokens": [ + -2.2379183769226074 + ], + "logprob_reward": -0.11291839927434921, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291839927434921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07365532964468002, + 0.020828671753406525, + 0.2935112714767456, + 0.015147543512284756, + 0.095222108066082, + 0.03434217348694801, + 0.178047776222229, + 0.01562214270234108, + 0.16599515080451965, + 0.10762784630060196 + ], + "credit_map": [ + 0.005130365956574678, + 0.0036520427092909813, + 0.006574280094355345, + 0.008416963741183281, + 0.012229361571371555, + 0.016432277858257294, + 0.013333340175449848, + 0.007886702194809914, + 0.001696513849310577, + 0.001924610696732998, + 0.0017672920366749167, + 0.0031135468743741512, + 0.0029075536876916885, + 0.0024056294932961464, + 0.0033315301407128572, + 0.003681996138766408, + 0.011349616572260857, + 0.008360244333744049, + 0.008677997626364231, + 0.11618807911872864, + 0.10531019419431686, + 0.01092543825507164, + 0.024673892185091972, + 0.008025798946619034, + 0.0011408569989725947, + 0.002777655143290758, + 0.0014454019255936146, + 0.0025983655359596014, + 0.0018509309738874435, + 0.0013952215667814016, + 0.0023660557344555855, + 0.0015730555169284344, + 0.006933323107659817, + 0.008551547303795815, + 0.00840346235781908, + 0.015221679583191872, + 0.018132973462343216, + 0.01556576881557703, + 0.012411895208060741, + 0.01000146009027958, + 0.0021639917977154255, + 0.002724746009334922, + 0.002851535566151142, + 0.0035638075787574053, + 0.005969586316496134, + 0.003858269890770316, + 0.0077133323065936565, + 0.005496903322637081, + 0.005623535718768835, + 0.0063195424154400826, + 0.010855474509298801, + 0.06875625252723694, + 0.008923972956836224, + 0.008082746528089046, + 0.012075243517756462, + 0.05741100385785103, + 0.0008208721992559731, + 0.001317680929787457, + 0.001512109418399632, + 0.0025701699778437614, + 0.0031954811420291662, + 0.0022968794219195843, + 0.0023567616008222103, + 0.0015521874884143472, + 0.0036401201505213976, + 0.002485689939931035, + 0.0028241921681910753, + 0.006594780366867781, + 0.06505575776100159, + 0.008479799143970013, + 0.005033496767282486, + 0.07188130915164948, + 0.005041241180151701, + 0.00875784456729889, + 0.010379249230027199, + 0.01414093654602766, + 0.017740514129400253, + 0.0202189888805151, + 0.017622489482164383, + 0.013726579956710339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.11291839927434921 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, almost imperceptible medium close-up shot from a slightly high angle captures a middle-aged chef, his entire person radiating an aura of intense concentration. He wears a plain, starched white, double-breasted chef's jacket with contrasting black buttons, and a traditional top hat perfectly perched atop his neatly combed black hair. His brow is slightly furrowed, his eyes narrowed in concentration as he leans over a large, gleaming stainless steel counter, which reflects the bright, even light of a professional kitchen like a perfect mirror. In the softly blurred background, a rack of polished copper pots and pans hangs in an orderly fashion. He holds a perfectly folded, clean white cloth in his hand, carefully wiping the already flawless surface. The movement isn't hurried, but a series of steady, deliberate motions. His arm glides in long, smooth, overlapping strokes, applying just enough pressure to polish the metal, leaving behind a fleeting, streak-free trace of moisture that disappears almost immediately. The scene, dominated by cool whites and metallic silver tones, creates a quiet, almost meditative atmosphere, one of absolute precision and unwavering devotion to craft.", + "seed": 31466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_15/index_6777_seed_31466_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0883150378242135, + "vlm_logprob": -0.0883150378242135, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.0883150378242135, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04445411637425423 + ], + "opposite_logprob_tokens": [ + -4.044454097747803 + ], + "logprob_reward": -0.04445411637425423, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.04445411637425423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2762274146080017, + 0.10304755717515945, + 0.10144217312335968, + 0.06500299274921417, + 0.03925285488367081, + 0.05050353333353996, + 0.08156195282936096, + 0.134051114320755, + 0.05226663500070572, + 0.09664378315210342 + ], + "credit_map": [ + 0.027695026248693466, + 0.038811005651950836, + 0.020558878779411316, + 0.026775069534778595, + 0.0254282895475626, + 0.04558392986655235, + 0.0487070232629776, + 0.04266819730401039, + 0.010086447931826115, + 0.015487689524888992, + 0.00751105509698391, + 0.01597507670521736, + 0.01298835501074791, + 0.01037075649946928, + 0.01886364445090294, + 0.011764531955122948, + 0.010137461125850677, + 0.014834719710052013, + 0.009203512221574783, + 0.01050441525876522, + 0.013364166021347046, + 0.013506353832781315, + 0.013923102989792824, + 0.0159684419631958, + 0.007442013826221228, + 0.0060484688729047775, + 0.005838868208229542, + 0.009194190613925457, + 0.010201015509665012, + 0.00756380520761013, + 0.011780156753957272, + 0.006934473291039467, + 0.003956245724111795, + 0.007118963170796633, + 0.003020991338416934, + 0.0040481737814843655, + 0.004490773659199476, + 0.005331825464963913, + 0.007217074278742075, + 0.0040688058361411095, + 0.004732171539217234, + 0.005584804341197014, + 0.009984488599002361, + 0.007109752390533686, + 0.009904121980071068, + 0.004438157659024, + 0.0037052861880511045, + 0.005044752266258001, + 0.010007504373788834, + 0.014304148964583874, + 0.0072629800997674465, + 0.008890305645763874, + 0.010864946991205215, + 0.008959156461060047, + 0.012703911401331425, + 0.008568999357521534, + 0.012928206473588943, + 0.011573253199458122, + 0.012764154933393002, + 0.016036489978432655, + 0.02323845587670803, + 0.017638348042964935, + 0.016843173652887344, + 0.023029033094644547, + 0.006693507079035044, + 0.005515394266694784, + 0.009027314372360706, + 0.006334064994007349, + 0.005905529484152794, + 0.006131425965577364, + 0.007609458640217781, + 0.005049939267337322, + 0.013275151140987873, + 0.01151779294013977, + 0.017370879650115967, + 0.011393057182431221, + 0.015612995252013206, + 0.00934088509529829, + 0.007496973499655724, + 0.010636045597493649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.04445411637425423 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281355619430542 + ], + "opposite_logprob_tokens": [ + -1.406355619430542 + ], + "logprob_reward": -0.281355619430542, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281355619430542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15896771848201752, + 0.07005446404218674, + 0.08723098039627075, + 0.10876550525426865, + 0.028880584985017776, + 0.16494786739349365, + 0.01752731390297413, + 0.2697698473930359, + 0.023787036538124084, + 0.07006866484880447 + ], + "credit_map": [ + 0.015871666371822357, + 0.018978316336870193, + 0.013206914067268372, + 0.018442729488015175, + 0.01985352858901024, + 0.018720710650086403, + 0.02828413061797619, + 0.02560972049832344, + 0.006521501578390598, + 0.013070694170892239, + 0.00741252675652504, + 0.007580709643661976, + 0.009354780428111553, + 0.0071656848303973675, + 0.011390356346964836, + 0.007558204233646393, + 0.012396188452839851, + 0.018297674134373665, + 0.010418583638966084, + 0.008358216844499111, + 0.010907410643994808, + 0.007749782409518957, + 0.00822076853364706, + 0.0108823636546731, + 0.011526569724082947, + 0.006956991273909807, + 0.016040543094277382, + 0.0219852477312088, + 0.012488394975662231, + 0.014942092821002007, + 0.012440485879778862, + 0.012385183945298195, + 0.0034000847954303026, + 0.0027641060296446085, + 0.0028023812919855118, + 0.0029885100666433573, + 0.0036656383890658617, + 0.00412646122276783, + 0.005590301007032394, + 0.0035431040450930595, + 0.01768849417567253, + 0.019723903387784958, + 0.014502476900815964, + 0.02316376566886902, + 0.034333210438489914, + 0.01704544574022293, + 0.01918814890086651, + 0.019302425906062126, + 0.0026198525447398424, + 0.002868747105821967, + 0.0015030705835670233, + 0.0018946119816973805, + 0.002454303903505206, + 0.0020989973563700914, + 0.001973486738279462, + 0.0021142433397471905, + 0.022863775491714478, + 0.04186872020363808, + 0.015324905514717102, + 0.03683611378073692, + 0.05924449488520622, + 0.04348188266158104, + 0.02673613280057907, + 0.023413822054862976, + 0.002331734402105212, + 0.002058119047433138, + 0.0031669442541897297, + 0.0038713784888386726, + 0.0036891582421958447, + 0.0030956733971834183, + 0.003043962176889181, + 0.002530067227780819, + 0.0057458095252513885, + 0.008265241980552673, + 0.007141835056245327, + 0.012468960136175156, + 0.007815546356141567, + 0.00992863904684782, + 0.009808140806853771, + 0.00889449380338192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.281355619430542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198591001331806 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.014198591001331806, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198591001331806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2485700100660324, + 0.128843754529953, + 0.17432184517383575, + 0.03698437288403511, + 0.04470133036375046, + 0.11180977523326874, + 0.04287264123558998, + 0.12174227833747864, + 0.028957974165678024, + 0.06119609624147415 + ], + "credit_map": [ + 0.023383285850286484, + 0.0345463864505291, + 0.018477262929081917, + 0.016517065465450287, + 0.036480020731687546, + 0.053262125700712204, + 0.027031593024730682, + 0.03887227550148964, + 0.011994357220828533, + 0.026705823838710785, + 0.00978025607764721, + 0.017068373039364815, + 0.018043888732790947, + 0.011113905347883701, + 0.021283982321619987, + 0.012853174470365047, + 0.013106174767017365, + 0.02257559262216091, + 0.013760924339294434, + 0.017475293949246407, + 0.0409386083483696, + 0.024434473365545273, + 0.020450446754693985, + 0.021580329164862633, + 0.004382007289677858, + 0.005244992207735777, + 0.0037728806491941214, + 0.00664625596255064, + 0.0036852078046649694, + 0.003817813703790307, + 0.006258451845496893, + 0.003176762955263257, + 0.003736494341865182, + 0.005638020113110542, + 0.0033329948782920837, + 0.007301876787096262, + 0.005597558803856373, + 0.005807877518236637, + 0.007805101107805967, + 0.005481405183672905, + 0.020564017817378044, + 0.00971168838441372, + 0.01828598603606224, + 0.014309648424386978, + 0.011099342256784439, + 0.009842419065535069, + 0.010988930240273476, + 0.01700774021446705, + 0.005573050584644079, + 0.0033645988442003727, + 0.0066816541366279125, + 0.0068571013398468494, + 0.007110172882676125, + 0.0034666990395635366, + 0.006075743585824966, + 0.0037436208222061396, + 0.011230584233999252, + 0.01572437398135662, + 0.010900840163230896, + 0.015234109945595264, + 0.022424599155783653, + 0.02583368867635727, + 0.010788846760988235, + 0.00960522796958685, + 0.0025873531121760607, + 0.002991889836266637, + 0.0028589058201760054, + 0.004473437555134296, + 0.003413533791899681, + 0.004146575927734375, + 0.003641833085566759, + 0.0048444452695548534, + 0.006066356785595417, + 0.005754084791988134, + 0.00941812340170145, + 0.013229778036475182, + 0.007216892205178738, + 0.006079763639718294, + 0.0068693566136062145, + 0.006561737507581711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014198591001331806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253841444849968 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.01253841444849968, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253841444849968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.233401820063591, + 0.19193844497203827, + 0.09534810483455658, + 0.03732377663254738, + 0.07862495630979538, + 0.04412052035331726, + 0.040504518896341324, + 0.15466433763504028, + 0.07309751212596893, + 0.05097605660557747 + ], + "credit_map": [ + 0.025987515226006508, + 0.04376092553138733, + 0.019785618409514427, + 0.016772890463471413, + 0.02503088116645813, + 0.03547320514917374, + 0.027348358184099197, + 0.039242424070835114, + 0.02646907977759838, + 0.06260110437870026, + 0.0161791630089283, + 0.018696367740631104, + 0.01590496487915516, + 0.014017565175890923, + 0.020301993936300278, + 0.01776820793747902, + 0.012920727953314781, + 0.020694639533758163, + 0.011887475848197937, + 0.009324507787823677, + 0.010102149099111557, + 0.009401246905326843, + 0.01279458124190569, + 0.008222774602472782, + 0.00526392599567771, + 0.004391627851873636, + 0.005093012470752001, + 0.004775958601385355, + 0.004596571438014507, + 0.0038909546565264463, + 0.005654576234519482, + 0.0036571475211530924, + 0.007253900170326233, + 0.016179321333765984, + 0.006222513038665056, + 0.0074480758048594, + 0.012161101214587688, + 0.008034469559788704, + 0.012420701794326305, + 0.008904878050088882, + 0.005621031392365694, + 0.005639225710183382, + 0.004951510112732649, + 0.006240309681743383, + 0.006500966381281614, + 0.004476378206163645, + 0.00456278258934617, + 0.006128313485532999, + 0.0038659016136080027, + 0.004181249067187309, + 0.005039357114583254, + 0.004502949304878712, + 0.0064697605557739735, + 0.004122579004615545, + 0.008310751058161259, + 0.004011973273009062, + 0.013071391731500626, + 0.00989607349038124, + 0.008403907530009747, + 0.014071764424443245, + 0.049106795340776443, + 0.04093492776155472, + 0.010805191472172737, + 0.008374284952878952, + 0.009276128374040127, + 0.007064727135002613, + 0.013941174373030663, + 0.011493313126266003, + 0.008163743652403355, + 0.007231167517602444, + 0.00957668386399746, + 0.006350575480610132, + 0.0063297632150352, + 0.004733530338853598, + 0.005893580615520477, + 0.008821147494018078, + 0.009360441006720066, + 0.00555845070630312, + 0.005343900062143803, + 0.004935246426612139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01253841444849968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902844786643982 + ], + "opposite_logprob_tokens": [ + -2.4640283584594727 + ], + "logprob_reward": -0.08902844786643982, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902844786643982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17320866882801056, + 0.09447294473648071, + 0.133307084441185, + 0.05784992501139641, + 0.06102811172604561, + 0.09471094608306885, + 0.036929916590452194, + 0.12334223091602325, + 0.031066808849573135, + 0.1940833479166031 + ], + "credit_map": [ + 0.01645418256521225, + 0.018328562378883362, + 0.016213301569223404, + 0.01398671418428421, + 0.02784327231347561, + 0.03459241986274719, + 0.02346900850534439, + 0.02232120744884014, + 0.009905891492962837, + 0.016115505248308182, + 0.007806251756846905, + 0.008415321819484234, + 0.010280877351760864, + 0.007096016313880682, + 0.026362672448158264, + 0.008490406908094883, + 0.009343404322862625, + 0.012013876810669899, + 0.010102389380335808, + 0.016545889899134636, + 0.027052847668528557, + 0.02204005792737007, + 0.018079690635204315, + 0.018128929659724236, + 0.006281472742557526, + 0.007488796953111887, + 0.006500187795609236, + 0.010234503075480461, + 0.005615613888949156, + 0.004584457725286484, + 0.011410925537347794, + 0.005733971018344164, + 0.004359186161309481, + 0.005912224296480417, + 0.00438696937635541, + 0.009328946471214294, + 0.007851075381040573, + 0.008515791036188602, + 0.01266554556787014, + 0.008008372038602829, + 0.01672811061143875, + 0.008171087130904198, + 0.013773844577372074, + 0.010127791203558445, + 0.009043357335031033, + 0.008691539987921715, + 0.010357829742133617, + 0.017817387357354164, + 0.0037498329766094685, + 0.003136036451905966, + 0.0027210356201976538, + 0.005174683872610331, + 0.005972252693027258, + 0.004792965482920408, + 0.007631882093846798, + 0.0037512239068746567, + 0.017592117190361023, + 0.008375799283385277, + 0.02537691406905651, + 0.013292787596583366, + 0.019459031522274017, + 0.015759091824293137, + 0.01109030470252037, + 0.012396181933581829, + 0.0039046674501150846, + 0.005448875017464161, + 0.004955630749464035, + 0.002689960179850459, + 0.003734058700501919, + 0.0029332225676625967, + 0.003786601824685931, + 0.0036137921269983053, + 0.019598081707954407, + 0.020363878458738327, + 0.026601847261190414, + 0.04114875942468643, + 0.02953930012881756, + 0.01812603883445263, + 0.01895330287516117, + 0.019752137362957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.08902844786643982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, almost imperceptible medium close-up shot from a slightly high angle captures a middle-aged chef, his entire person radiating an aura of intense concentration. He wears a plain, starched white, double-breasted chef's jacket with contrasting black buttons, and a traditional top hat perfectly perched atop his neatly combed black hair. His brow is slightly furrowed, his eyes narrowed in concentration as he leans over a large, gleaming stainless steel counter, which reflects the bright, even light of a professional kitchen like a perfect mirror. In the softly blurred background, a rack of polished copper pots and pans hangs in an orderly fashion. He holds a perfectly folded, clean white cloth in his hand, carefully wiping the already flawless surface. The movement isn't hurried, but a series of steady, deliberate motions. His arm glides in long, smooth, overlapping strokes, applying just enough pressure to polish the metal, leaving behind a fleeting, streak-free trace of moisture that disappears almost immediately. The scene, dominated by cool whites and metallic silver tones, creates a quiet, almost meditative atmosphere, one of absolute precision and unwavering devotion to craft.", + "seed": 31467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_15/index_6777_seed_31467_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06792592434212566, + "vlm_logprob": -0.06792592434212566, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.06792592434212566, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03071443736553192 + ], + "opposite_logprob_tokens": [ + -4.405714511871338 + ], + "logprob_reward": -0.03071443736553192, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03071443736553192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14748802781105042, + 0.06343279778957367, + 0.059534914791584015, + 0.06352951377630234, + 0.03727119788527489, + 0.0448034442961216, + 0.05603731796145439, + 0.35864850878715515, + 0.07550487667322159, + 0.09374939650297165 + ], + "credit_map": [ + 0.008304986171424389, + 0.010049670934677124, + 0.02079070545732975, + 0.013688065111637115, + 0.02416689693927765, + 0.020702414214611053, + 0.02026476338505745, + 0.029520515352487564, + 0.005974264349788427, + 0.004844826180487871, + 0.005749108735471964, + 0.009115103632211685, + 0.009768101386725903, + 0.008450585417449474, + 0.010547751560807228, + 0.008983053267002106, + 0.005221717990934849, + 0.005437787622213364, + 0.0050910827703773975, + 0.006479448173195124, + 0.01000847015529871, + 0.013584315776824951, + 0.006903792731463909, + 0.0068082972429692745, + 0.003040383569896221, + 0.006368257105350494, + 0.007370585110038519, + 0.008679492399096489, + 0.010286493226885796, + 0.01020644698292017, + 0.009058780036866665, + 0.008519063703715801, + 0.0027622669003903866, + 0.00251568085514009, + 0.004591803066432476, + 0.004559917841106653, + 0.0055124456994235516, + 0.006327458657324314, + 0.006396264303475618, + 0.004605362191796303, + 0.002011945005506277, + 0.001498867990449071, + 0.013221317902207375, + 0.0040855552069842815, + 0.004782167263329029, + 0.004183518700301647, + 0.006173083558678627, + 0.008846987038850784, + 0.003586944192647934, + 0.004992957226932049, + 0.0052170525304973125, + 0.005885803606361151, + 0.0061447275802493095, + 0.00845420453697443, + 0.011446438729763031, + 0.010309185832738876, + 0.013359276577830315, + 0.007730225101113319, + 0.12433889508247375, + 0.016342245042324066, + 0.02022969350218773, + 0.029917307198047638, + 0.017246918752789497, + 0.12948395311832428, + 0.0034672552719712257, + 0.00494877016171813, + 0.016633290797472, + 0.016606943681836128, + 0.009489561431109905, + 0.010466265492141247, + 0.008811299689114094, + 0.005081491079181433, + 0.004476340487599373, + 0.007104671560227871, + 0.008231398649513721, + 0.02124796435236931, + 0.013926124200224876, + 0.008856234140694141, + 0.011691802181303501, + 0.01821485161781311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.03071443736553192 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521187961101532 + ], + "opposite_logprob_tokens": [ + -1.5021188259124756 + ], + "logprob_reward": -0.2521187961101532, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521187961101532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28233250975608826, + 0.14309334754943848, + 0.13690532743930817, + 0.04796773940324783, + 0.0524522066116333, + 0.046409886330366135, + 0.030133748427033424, + 0.1344834715127945, + 0.08265317231416702, + 0.04356854408979416 + ], + "credit_map": [ + 0.011880224570631981, + 0.01746712438762188, + 0.02193850465118885, + 0.024035973474383354, + 0.04045451804995537, + 0.039209067821502686, + 0.05071559548377991, + 0.07663150876760483, + 0.010564927943050861, + 0.011077288538217545, + 0.0160379596054554, + 0.014849254861474037, + 0.0203932523727417, + 0.02650274895131588, + 0.021932177245616913, + 0.02173573523759842, + 0.0070585343055427074, + 0.0100826695561409, + 0.018403921276330948, + 0.019467558711767197, + 0.02266504243016243, + 0.021850747987627983, + 0.021399997174739838, + 0.01597684994339943, + 0.0022393232211470604, + 0.002428555628284812, + 0.0113141518086195, + 0.0038729747757315636, + 0.0073456657119095325, + 0.008197399787604809, + 0.0042165727354586124, + 0.008353098295629025, + 0.0030871776398271322, + 0.0039845965802669525, + 0.005371139384806156, + 0.008502502925693989, + 0.008939487859606743, + 0.008391963317990303, + 0.007359635084867477, + 0.006815701723098755, + 0.0022268653847277164, + 0.002817140193656087, + 0.008927677758038044, + 0.004592484328895807, + 0.00911171454936266, + 0.004451787564903498, + 0.008802970871329308, + 0.005479250568896532, + 0.002113716211169958, + 0.0026877352502197027, + 0.006427557207643986, + 0.0025343592278659344, + 0.0038958771619945765, + 0.0029560974799096584, + 0.004731310531497002, + 0.004787096753716469, + 0.005337137263268232, + 0.0025426745414733887, + 0.04628397151827812, + 0.006430949084460735, + 0.008198348805308342, + 0.010707740671932697, + 0.0064201029017567635, + 0.04856255277991295, + 0.002996123628690839, + 0.004795996006578207, + 0.011701077222824097, + 0.016215363517403603, + 0.011888221837580204, + 0.008045111782848835, + 0.019152721390128136, + 0.007858560420572758, + 0.0022836546413600445, + 0.003322448581457138, + 0.0029710226226598024, + 0.011784819886088371, + 0.004828569479286671, + 0.0043904222548007965, + 0.0037145602982491255, + 0.010273047722876072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2521187961101532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076745577156544 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076745577156544, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076745577156544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050795551389455795, + 0.03339123725891113, + 0.034578461199998856, + 0.02429181896150112, + 0.04166384041309357, + 0.06272822618484497, + 0.059956252574920654, + 0.6057654023170471, + 0.0762588158249855, + 0.010570361278951168 + ], + "credit_map": [ + 0.003262175712734461, + 0.003986568655818701, + 0.004903730470687151, + 0.005336755886673927, + 0.008137065917253494, + 0.005955414846539497, + 0.008294617757201195, + 0.010919223539531231, + 0.0023141568526625633, + 0.002751972759142518, + 0.004836251959204674, + 0.004488753620535135, + 0.0032148766331374645, + 0.005560863763093948, + 0.005405088420957327, + 0.004819274414330721, + 0.0029956025537103415, + 0.00432065362110734, + 0.00378617737442255, + 0.0049013108946383, + 0.004340964835137129, + 0.006194637157022953, + 0.0041168988682329655, + 0.003922217059880495, + 0.0017024604603648186, + 0.0011503231944516301, + 0.003887815633788705, + 0.003876908915117383, + 0.0025914516299962997, + 0.0036476280074566603, + 0.002145074075087905, + 0.0052901566959917545, + 0.0033764054533094168, + 0.003069190541282296, + 0.008196601644158363, + 0.00752659235149622, + 0.004485331941395998, + 0.005438758991658688, + 0.004654136952012777, + 0.004916821140795946, + 0.0026270693633705378, + 0.0032215372193604708, + 0.015355917625129223, + 0.005266523454338312, + 0.010411894880235195, + 0.006277837790548801, + 0.010848538018763065, + 0.00871890876442194, + 0.002624817891046405, + 0.004109365399926901, + 0.0059656500816345215, + 0.0056992447935044765, + 0.006587438751012087, + 0.008968333713710308, + 0.015875203534960747, + 0.010126196779310703, + 0.021714670583605766, + 0.0075545222498476505, + 0.2260536402463913, + 0.021655714139342308, + 0.022631142288446426, + 0.052010972052812576, + 0.027739185839891434, + 0.2264055460691452, + 0.0029992375057190657, + 0.004789391532540321, + 0.01939038187265396, + 0.014287537895143032, + 0.010747237130999565, + 0.006844733376055956, + 0.012018661014735699, + 0.005181633867323399, + 0.0006389266927726567, + 0.000765384582336992, + 0.0017809576820582151, + 0.0022768264170736074, + 0.0011278947349637747, + 0.00128213653806597, + 0.0009854172822088003, + 0.0017128176987171173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.011076745577156544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626701612025499 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626701612025499, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626701612025499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03628558665513992, + 0.05573269724845886, + 0.03574562072753906, + 0.03917435184121132, + 0.03075948916375637, + 0.0650549903512001, + 0.04972539842128754, + 0.6364361643791199, + 0.03660883381962776, + 0.014476871117949486 + ], + "credit_map": [ + 0.0025629070587456226, + 0.0031550221610814333, + 0.004442925564944744, + 0.0057225837372243404, + 0.0067853061482310295, + 0.005374202970415354, + 0.003564054612070322, + 0.00467858649790287, + 0.006544673349708319, + 0.00528630381450057, + 0.010034164413809776, + 0.005986414849758148, + 0.007769791409373283, + 0.005995685234665871, + 0.007735017221421003, + 0.006380645092576742, + 0.004305531736463308, + 0.006128878798335791, + 0.004245273768901825, + 0.006198648363351822, + 0.0032340241596102715, + 0.003936859779059887, + 0.003947109915316105, + 0.003749291878193617, + 0.001876113936305046, + 0.0036712398286908865, + 0.006340180989354849, + 0.004899777937680483, + 0.005963008850812912, + 0.0047764163464307785, + 0.005367398262023926, + 0.006280211731791496, + 0.002788043348118663, + 0.0033793007023632526, + 0.004845139570534229, + 0.0040821172297000885, + 0.0033504643943160772, + 0.003674502018839121, + 0.00422319071367383, + 0.00441672932356596, + 0.0025873633567243814, + 0.002852523699402809, + 0.01733168587088585, + 0.005634131375700235, + 0.010551023297011852, + 0.006010131910443306, + 0.01059006992727518, + 0.009498057886958122, + 0.0049270023591816425, + 0.005504150874912739, + 0.004468308761715889, + 0.005012577399611473, + 0.0060109361074864864, + 0.006160271354019642, + 0.009675371460616589, + 0.0079667828977108, + 0.023040922358632088, + 0.008078317157924175, + 0.2370147705078125, + 0.022106124088168144, + 0.024225082248449326, + 0.054613545536994934, + 0.028745032846927643, + 0.23861238360404968, + 0.0017477701185271144, + 0.0025608947034925222, + 0.00978103093802929, + 0.007807310204952955, + 0.004430786240845919, + 0.003173881210386753, + 0.004877874627709389, + 0.0022292819339782, + 0.0008822858799248934, + 0.0011006497079506516, + 0.0012426208704710007, + 0.0028741112910211086, + 0.0015398322138935328, + 0.0017325043445453048, + 0.0016432960983365774, + 0.003461569780483842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.007626701612025499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged chef wearing a white double-breasted jacket with black buttons and a top hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the chef leaning over a large stainless steel counter while holding a white cloth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred showing a rack of polished copper pots and pans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the chef perform slow, deliberate arm strokes to polish the counter surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809294104576111 + ], + "opposite_logprob_tokens": [ + -3.288092851638794 + ], + "logprob_reward": -0.03809294104576111, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809294104576111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05869292467832565, + 0.04170028492808342, + 0.028021009638905525, + 0.022729018703103065, + 0.044535718858242035, + 0.06167434900999069, + 0.09450533986091614, + 0.5282562375068665, + 0.0950295552611351, + 0.0248555988073349 + ], + "credit_map": [ + 0.0032892131712287664, + 0.004917266312986612, + 0.005467946641147137, + 0.008253969252109528, + 0.006815151311457157, + 0.008692112751305103, + 0.00802688766270876, + 0.013230378739535809, + 0.0030614573042839766, + 0.003873400390148163, + 0.004626843146979809, + 0.005567347630858421, + 0.004581955261528492, + 0.007509869523346424, + 0.006822406314313412, + 0.00565700326114893, + 0.001882882439531386, + 0.0017415548209100962, + 0.003945908974856138, + 0.005373534746468067, + 0.0031500468030571938, + 0.004694548435509205, + 0.0035386632662266493, + 0.0036938695702701807, + 0.0020922808907926083, + 0.002038802718743682, + 0.0025791404768824577, + 0.003131633624434471, + 0.0028784407768398523, + 0.003871355904266238, + 0.0030746732372790575, + 0.0030626910738646984, + 0.004141607787460089, + 0.0033959499560296535, + 0.005485099274665117, + 0.007351926527917385, + 0.0065710339695215225, + 0.005969436839222908, + 0.006601634435355663, + 0.005019030999392271, + 0.003837409894913435, + 0.00613143527880311, + 0.007548892870545387, + 0.006883356254547834, + 0.013214579783380032, + 0.0086030513048172, + 0.008603347465395927, + 0.006852280348539352, + 0.004511706531047821, + 0.007523483131080866, + 0.011384401470422745, + 0.012549193575978279, + 0.013958663679659367, + 0.01615217886865139, + 0.014303465373814106, + 0.014122245833277702, + 0.019017742946743965, + 0.006838182453066111, + 0.19619618356227875, + 0.0193190798163414, + 0.020309245213866234, + 0.04548383504152298, + 0.023580530658364296, + 0.1975114494562149, + 0.003900129348039627, + 0.006239031441509724, + 0.024920372292399406, + 0.017360948026180267, + 0.012506160885095596, + 0.009472861886024475, + 0.014527936466038227, + 0.0061021121218800545, + 0.0014424258843064308, + 0.0026622358709573746, + 0.003588245017454028, + 0.0038177717942744493, + 0.0028468479868024588, + 0.0040066721849143505, + 0.0026398426853120327, + 0.0038515578489750624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.03809294104576111 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_tensor([8916, 8916])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_tensor([8916, 8916])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f291aa79d3d090c955aa35e706ec3250b8c26cd4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_tensor([8916, 8916])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, eye-level tracking shot slides slowly and smoothly from left to right, parallel to a massive, towering wall of hydroponic plants within a futuristic agricultural skyscraper. Thousands of vibrant green and crimson heads of lettuce and herbs sprout from a grid of pristine white modules, their leaves lush and perfectly shaped. The entire scene is bathed in the bright, even glow of full-spectrum LED grow lights, eliminating harsh shadows and giving the footage the crisp, observational quality of a high-end documentary. As the camera moves, a fine misting system meticulously activates; nearly invisible nozzles embedded in the structure release a gentle, atomized spray that drifts downward and blankets the foliage. The water vapor catches the light, creating a shimmering veil across the leaves, leaving behind shimmering microdroplets that capture in serene detail the perfect fusion of advanced technology and pristine nature.", + "seed": 12205, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_8916_seed_12205_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06316667394712568, + "vlm_logprob": -0.06316667394712568, + "vlm_yes_no_margin": 3.250000023841858, + "vlm_token_credit": -0.06316667394712568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14784853160381317 + ], + "opposite_logprob_tokens": [ + -2.022848606109619 + ], + "logprob_reward": -0.14784853160381317, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14784853160381317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24880987405776978, + 0.05245937034487724, + 0.02139976993203163, + 0.2666282653808594, + 0.06666463613510132, + 0.015662264078855515, + 0.04037555307149887, + 0.09627629071474075, + 0.10321886092424393, + 0.08850502222776413 + ], + "credit_map": [ + 0.020436223596334457, + 0.028343060985207558, + 0.03217890486121178, + 0.037990324199199677, + 0.036373622715473175, + 0.029600754380226135, + 0.026977650821208954, + 0.03690936416387558, + 0.005041054449975491, + 0.0045755538158118725, + 0.007269516587257385, + 0.0051796347834169865, + 0.006436440162360668, + 0.00981704518198967, + 0.004794599954038858, + 0.00934553612023592, + 0.002165701240301132, + 0.0019768388010561466, + 0.0021008190233260393, + 0.0030434392392635345, + 0.002630563685670495, + 0.003080286318436265, + 0.0037944128271192312, + 0.0026077113579958677, + 0.016256531700491905, + 0.010772455483675003, + 0.023577513173222542, + 0.01917349547147751, + 0.01845412515103817, + 0.06983727216720581, + 0.08880618959665298, + 0.019750716164708138, + 0.005782683379948139, + 0.00859695766121149, + 0.007283139042556286, + 0.007930085062980652, + 0.010296051390469074, + 0.006192994304001331, + 0.013710892759263515, + 0.0068718381226062775, + 0.0019213884370401502, + 0.0016013502608984709, + 0.001940095447935164, + 0.0020563078578561544, + 0.0014618878485634923, + 0.0015297754434868693, + 0.002689496148377657, + 0.002461964963003993, + 0.005021304357796907, + 0.004472020082175732, + 0.005631950218230486, + 0.004933984950184822, + 0.004581150598824024, + 0.005305037833750248, + 0.005922713782638311, + 0.004507396370172501, + 0.007641598582267761, + 0.007078053895384073, + 0.020066335797309875, + 0.0156710222363472, + 0.009467251598834991, + 0.011541428975760937, + 0.0116631630808115, + 0.013147449120879173, + 0.006723083555698395, + 0.00621944572776556, + 0.020188387483358383, + 0.011566001921892166, + 0.024669524282217026, + 0.007570624817162752, + 0.015499078668653965, + 0.010782727040350437, + 0.010562165640294552, + 0.007415864150971174, + 0.011453439481556416, + 0.014400694519281387, + 0.014370100572705269, + 0.0070126173086464405, + 0.015892785042524338, + 0.007397362031042576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.14784853160381317 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046614077873528 + ], + "opposite_logprob_tokens": [ + -5.379661560058594 + ], + "logprob_reward": -0.0046614077873528, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046614077873528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20075862109661102, + 0.08873426169157028, + 0.024096276611089706, + 0.08680282533168793, + 0.06385842710733414, + 0.04979323968291283, + 0.03700593486428261, + 0.0471632182598114, + 0.33292093873023987, + 0.06886628270149231 + ], + "credit_map": [ + 0.017088940367102623, + 0.02542394958436489, + 0.03568435460329056, + 0.046596918255090714, + 0.023375194519758224, + 0.016554221510887146, + 0.01626327820122242, + 0.019771767780184746, + 0.012613818980753422, + 0.011015992611646652, + 0.010775872506201267, + 0.010621676221489906, + 0.009291845373809338, + 0.014524034224450588, + 0.007876604795455933, + 0.012014421634376049, + 0.0019485370721668005, + 0.0023782141506671906, + 0.0024808808229863644, + 0.003121933899819851, + 0.003489293623715639, + 0.004020653199404478, + 0.0029492308385670185, + 0.0037075323052704334, + 0.005122381262481213, + 0.003752500284463167, + 0.00836773682385683, + 0.006635347381234169, + 0.004538536537438631, + 0.019785545766353607, + 0.029850171878933907, + 0.008750607259571552, + 0.005407368298619986, + 0.009662387892603874, + 0.007628139574080706, + 0.010081292130053043, + 0.009367961436510086, + 0.006677315570414066, + 0.009746086783707142, + 0.005287873558700085, + 0.0050344811752438545, + 0.004367892164736986, + 0.006962403655052185, + 0.005647820420563221, + 0.006693762261420488, + 0.006304214708507061, + 0.005270866211503744, + 0.009511798620223999, + 0.0043407706543803215, + 0.0050680264830589294, + 0.004005934577435255, + 0.00480800261721015, + 0.0051870448514819145, + 0.004495367873460054, + 0.00480444822460413, + 0.004296337254345417, + 0.00377725251019001, + 0.004497098736464977, + 0.006024633068591356, + 0.006874943617731333, + 0.004636179190129042, + 0.004209860693663359, + 0.008954648859798908, + 0.00818860623985529, + 0.019600743427872658, + 0.020058700814843178, + 0.08856178820133209, + 0.040936585515737534, + 0.07330705225467682, + 0.022942084819078445, + 0.025367774069309235, + 0.04214619845151901, + 0.006401496008038521, + 0.004997764248400927, + 0.005672320257872343, + 0.010794956237077713, + 0.009590451605618, + 0.004619859624654055, + 0.022412460297346115, + 0.004376970697194338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0046614077873528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308804258704185 + ], + "opposite_logprob_tokens": [ + -3.773308753967285 + ], + "logprob_reward": -0.023308804258704185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308804258704185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18937689065933228, + 0.15450456738471985, + 0.025903888046741486, + 0.10403358936309814, + 0.025963321328163147, + 0.03632863238453865, + 0.01986047439277172, + 0.06472993642091751, + 0.2753714323043823, + 0.10392728447914124 + ], + "credit_map": [ + 0.01598476804792881, + 0.02053764835000038, + 0.025362998247146606, + 0.020517773926258087, + 0.02864772081375122, + 0.028215019032359123, + 0.02131783403456211, + 0.02879314124584198, + 0.017582332715392113, + 0.011448959819972515, + 0.020007921382784843, + 0.014520436525344849, + 0.035369809716939926, + 0.032603852450847626, + 0.010014569386839867, + 0.012956681661307812, + 0.0027488016057759523, + 0.0027023525908589363, + 0.0036054467782378197, + 0.003420720575377345, + 0.003626895369961858, + 0.0040589445270597935, + 0.0028119978960603476, + 0.0029287284705787897, + 0.005998039152473211, + 0.004149672109633684, + 0.009191480465233326, + 0.007294151466339827, + 0.006431120913475752, + 0.026412785053253174, + 0.035614609718322754, + 0.008941725827753544, + 0.0024552950635552406, + 0.003450616728514433, + 0.0027946210466325283, + 0.003566042287275195, + 0.0044660791754722595, + 0.0026954966597259045, + 0.0033863584976643324, + 0.003148810938000679, + 0.00481750350445509, + 0.0034256342332810163, + 0.005081250797957182, + 0.004182383418083191, + 0.005254221148788929, + 0.004069921560585499, + 0.003507930552586913, + 0.005989788565784693, + 0.0025702903512865305, + 0.0021428600884974003, + 0.0029205952305346727, + 0.0029426023829728365, + 0.00202987645752728, + 0.003064714139327407, + 0.002190002705901861, + 0.001999532338231802, + 0.0049186768010258675, + 0.004480022005736828, + 0.013024089857935905, + 0.009437655098736286, + 0.004241118207573891, + 0.006386732216924429, + 0.00844403263181448, + 0.013797612860798836, + 0.0164150632917881, + 0.016585662961006165, + 0.07179995626211166, + 0.033985696732997894, + 0.0610722191631794, + 0.018799100071191788, + 0.02195327915251255, + 0.034760482609272, + 0.007372955325990915, + 0.005013209767639637, + 0.012096298858523369, + 0.03696376085281372, + 0.009429055266082287, + 0.005551064386963844, + 0.023609252646565437, + 0.0038916850462555885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.023308804258704185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002314925193787 + ], + "opposite_logprob_tokens": [ + -2.6950230598449707 + ], + "logprob_reward": -0.07002314925193787, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002314925193787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04676678031682968, + 0.018675843253731728, + 0.03259233757853508, + 0.10805851966142654, + 0.04388941824436188, + 0.010867095552384853, + 0.011441645212471485, + 0.15479211509227753, + 0.5026066899299622, + 0.07030950486660004 + ], + "credit_map": [ + 0.004733867011964321, + 0.010299086570739746, + 0.0053094783797860146, + 0.003226757515221834, + 0.007418708875775337, + 0.007649528328329325, + 0.0034082545898854733, + 0.0047210995107889175, + 0.003557999152690172, + 0.001408034935593605, + 0.0017115662340074778, + 0.0018757289508357644, + 0.0030742131639271975, + 0.0032434493768960238, + 0.0020645102486014366, + 0.001740340841934085, + 0.004231767728924751, + 0.002447030507028103, + 0.0036855998914688826, + 0.00427868589758873, + 0.006834113504737616, + 0.004094907082617283, + 0.004099186044186354, + 0.0029210494831204414, + 0.006373238749802113, + 0.004255708307027817, + 0.009586033411324024, + 0.007607004139572382, + 0.006544200703501701, + 0.028211040422320366, + 0.03674202412366867, + 0.00873926654458046, + 0.0032656495459377766, + 0.004395241849124432, + 0.004204773344099522, + 0.005195037927478552, + 0.007048220373690128, + 0.004720794502645731, + 0.009316161274909973, + 0.005743540357798338, + 0.0013206872390583158, + 0.0011425850680097938, + 0.0013170299353078008, + 0.0013042920036241412, + 0.001348823425360024, + 0.001470971037633717, + 0.001059451955370605, + 0.0019032545387744904, + 0.0012694156030192971, + 0.0013337060809135437, + 0.0013973035383969545, + 0.001619779272004962, + 0.0016619092784821987, + 0.0016353398095816374, + 0.001457884325645864, + 0.0010663069551810622, + 0.008439255878329277, + 0.00910115521401167, + 0.027285538613796234, + 0.02298997901380062, + 0.009231498464941978, + 0.01577456295490265, + 0.02643624134361744, + 0.035533878952264786, + 0.029087845236063004, + 0.029544424265623093, + 0.14309462904930115, + 0.0609150230884552, + 0.10740530490875244, + 0.03450068086385727, + 0.03443678095936775, + 0.06362202763557434, + 0.004677533637732267, + 0.004035400226712227, + 0.00763957342132926, + 0.023336419835686684, + 0.007249393034726381, + 0.003952929750084877, + 0.01540351565927267, + 0.004014741163700819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.07002314925193787 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999147683382034 + ], + "opposite_logprob_tokens": [ + -2.6949915885925293 + ], + "logprob_reward": -0.06999147683382034, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999147683382034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05955057218670845, + 0.1372314691543579, + 0.08090394735336304, + 0.23280268907546997, + 0.023781217634677887, + 0.037423618137836456, + 0.01672416925430298, + 0.030906250700354576, + 0.31850144267082214, + 0.06217466667294502 + ], + "credit_map": [ + 0.01149973925203085, + 0.009239054284989834, + 0.007493352517485619, + 0.006511784624308348, + 0.005792354233562946, + 0.0063179959543049335, + 0.005356858484447002, + 0.00733943423256278, + 0.01185925304889679, + 0.011323938146233559, + 0.013582950457930565, + 0.015104211866855621, + 0.044065553694963455, + 0.02335280366241932, + 0.010264419950544834, + 0.007678341120481491, + 0.00795185286551714, + 0.006067809648811817, + 0.012526441365480423, + 0.009904349222779274, + 0.00873335637152195, + 0.011697406880557537, + 0.005974464118480682, + 0.018048271536827087, + 0.013627956621348858, + 0.008872504346072674, + 0.020884305238723755, + 0.017297761514782906, + 0.012184688821434975, + 0.05748662352561951, + 0.0820322260260582, + 0.020416615530848503, + 0.0021567184012383223, + 0.0023554086219519377, + 0.0032217015977948904, + 0.003166967537254095, + 0.0030278097838163376, + 0.0024962688330560923, + 0.004286283627152443, + 0.003070058301091194, + 0.004854688886553049, + 0.0034441570751369, + 0.005484799388796091, + 0.004163730423897505, + 0.006100930273532867, + 0.004225585609674454, + 0.003984030801802874, + 0.005165694281458855, + 0.0015512341633439064, + 0.0022172550670802593, + 0.002973855473101139, + 0.001938130590133369, + 0.0016799996374174953, + 0.0018835634691640735, + 0.0024192812852561474, + 0.0020608496852219105, + 0.0021696852054446936, + 0.0022386854980140924, + 0.003944222815334797, + 0.004479097668081522, + 0.0026508180890232325, + 0.002679317956790328, + 0.00529404217377305, + 0.007450381759554148, + 0.018617667257785797, + 0.017132868990302086, + 0.09136560559272766, + 0.038338564336299896, + 0.07187157869338989, + 0.020568329840898514, + 0.022108489647507668, + 0.03849835321307182, + 0.006217655260115862, + 0.0030150495003908873, + 0.007741667330265045, + 0.020270241424441338, + 0.008333655074238777, + 0.003191760741174221, + 0.010581578128039837, + 0.0028230594471096992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.06999147683382034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, eye-level tracking shot slides slowly and smoothly from left to right, parallel to a massive, towering wall of hydroponic plants within a futuristic agricultural skyscraper. Thousands of vibrant green and crimson heads of lettuce and herbs sprout from a grid of pristine white modules, their leaves lush and perfectly shaped. The entire scene is bathed in the bright, even glow of full-spectrum LED grow lights, eliminating harsh shadows and giving the footage the crisp, observational quality of a high-end documentary. As the camera moves, a fine misting system meticulously activates; nearly invisible nozzles embedded in the structure release a gentle, atomized spray that drifts downward and blankets the foliage. The water vapor catches the light, creating a shimmering veil across the leaves, leaving behind shimmering microdroplets that capture in serene detail the perfect fusion of advanced technology and pristine nature.", + "seed": 12206, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_8916_seed_12206_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.061402378417551515, + "vlm_logprob": -0.061402378417551515, + "vlm_yes_no_margin": 3.149999976158142, + "vlm_token_credit": -0.061402378417551515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07282746583223343 + ], + "opposite_logprob_tokens": [ + -2.6978275775909424 + ], + "logprob_reward": -0.07282746583223343, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07282746583223343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19012139737606049, + 0.08777238428592682, + 0.05898754298686981, + 0.024967260658740997, + 0.02524641901254654, + 0.044952139258384705, + 0.017550012096762657, + 0.48617908358573914, + 0.045105595141649246, + 0.019118154421448708 + ], + "credit_map": [ + 0.010343266651034355, + 0.010556560009717941, + 0.017480188980698586, + 0.06397929042577744, + 0.009367424063384533, + 0.055381663143634796, + 0.013973793014883995, + 0.009039212949573994, + 0.00515966909006238, + 0.005785011220723391, + 0.0047678034752607346, + 0.012007063254714012, + 0.007218458689749241, + 0.03292703256011009, + 0.004913387820124626, + 0.01499395165592432, + 0.003948558121919632, + 0.0033947378396987915, + 0.005280274897813797, + 0.00380111881531775, + 0.016305232420563698, + 0.004003698006272316, + 0.004662228748202324, + 0.017591698095202446, + 0.00507648941129446, + 0.002240290166810155, + 0.0023738103918731213, + 0.0020117037929594517, + 0.0021647403482347727, + 0.003946038894355297, + 0.0023055600468069315, + 0.004848627373576164, + 0.0023410881403833628, + 0.004222717601805925, + 0.0024825911968946457, + 0.003175608580932021, + 0.0029179202392697334, + 0.0044234259985387325, + 0.0024115764535963535, + 0.003271488007158041, + 0.0033954214304685593, + 0.0039029764011502266, + 0.006162995472550392, + 0.0071007185615599155, + 0.003875000635161996, + 0.0034649628214538097, + 0.012349026277661324, + 0.004701036959886551, + 0.002401285571977496, + 0.0020740388426929712, + 0.00206733332015574, + 0.0012711333110928535, + 0.0013742567971348763, + 0.0021123411133885384, + 0.004092811606824398, + 0.0021568117663264275, + 0.03221847862005234, + 0.11215997487306595, + 0.04330231621861458, + 0.028797341510653496, + 0.0354815274477005, + 0.056621234863996506, + 0.12053627520799637, + 0.05706193298101425, + 0.0040375348180532455, + 0.004309921991080046, + 0.004491407889872789, + 0.011459610424935818, + 0.003278423333540559, + 0.004701129626482725, + 0.009149910882115364, + 0.003677656874060631, + 0.0018876990070566535, + 0.002129036234691739, + 0.003650071332231164, + 0.0026605543680489063, + 0.0017871552845463157, + 0.0020642441231757402, + 0.0030921311117708683, + 0.0018472621450200677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.07282746583223343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011101149953901768 + ], + "opposite_logprob_tokens": [ + -4.511101245880127 + ], + "logprob_reward": -0.011101149953901768, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011101149953901768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2763151228427887, + 0.05537602677941322, + 0.023131107911467552, + 0.05126034840941429, + 0.02433081902563572, + 0.018725208938121796, + 0.026028111577033997, + 0.2944057583808899, + 0.0650755986571312, + 0.16535189747810364 + ], + "credit_map": [ + 0.018106132745742798, + 0.014750364236533642, + 0.02274898812174797, + 0.0858117938041687, + 0.01575450785458088, + 0.08323224633932114, + 0.021445874124765396, + 0.014465205371379852, + 0.0043607973493635654, + 0.003629380604252219, + 0.005246857646852732, + 0.007740833330899477, + 0.005144152324646711, + 0.016290422528982162, + 0.004261931870132685, + 0.00870165228843689, + 0.0023546183947473764, + 0.002218511188402772, + 0.004174760542809963, + 0.002830397803336382, + 0.0026101693511009216, + 0.002698221243917942, + 0.004338054917752743, + 0.0019063748186454177, + 0.011316856369376183, + 0.004170105792582035, + 0.005519803147763014, + 0.006212200503796339, + 0.006140619516372681, + 0.008274051360785961, + 0.004756132606416941, + 0.004870578646659851, + 0.0019412846304476261, + 0.002752755070105195, + 0.0019163495162501931, + 0.002715387614443898, + 0.0038872717414051294, + 0.004369691014289856, + 0.0035995214711874723, + 0.0031485564541071653, + 0.002464685123413801, + 0.0016621701652184129, + 0.0023068394511938095, + 0.0023273194674402475, + 0.0019180889939889312, + 0.0017546969465911388, + 0.004286636132746935, + 0.0020047721918672323, + 0.0025205938145518303, + 0.00448969891294837, + 0.0024019868578761816, + 0.002852225210517645, + 0.0023116962984204292, + 0.003575925249606371, + 0.004511483944952488, + 0.0033645012881606817, + 0.021960482001304626, + 0.048227835446596146, + 0.0303722582757473, + 0.026203889399766922, + 0.027962855994701385, + 0.0532226637005806, + 0.05199635028839111, + 0.03445940092206001, + 0.007919303141534328, + 0.004102621227502823, + 0.0077447956427931786, + 0.016772154718637466, + 0.004960944876074791, + 0.009268279187381268, + 0.010437693446874619, + 0.003869809443131089, + 0.005614973604679108, + 0.006548725068569183, + 0.05658937245607376, + 0.01546669751405716, + 0.007661889772862196, + 0.009348577819764614, + 0.057045139372348785, + 0.00707651861011982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.011101149953901768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822569966316223 + ], + "opposite_logprob_tokens": [ + -4.01822566986084 + ], + "logprob_reward": -0.01822569966316223, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822569966316223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12171679735183716, + 0.13365378975868225, + 0.143400639295578, + 0.022089282050728798, + 0.014269493520259857, + 0.014892120845615864, + 0.02654232271015644, + 0.39800456166267395, + 0.07128600031137466, + 0.05414501577615738 + ], + "credit_map": [ + 0.007832682691514492, + 0.006693373899906874, + 0.014106296934187412, + 0.03947902470827103, + 0.00740748131647706, + 0.030335646122694016, + 0.009280452504754066, + 0.006581833586096764, + 0.0084163723513484, + 0.009905087761580944, + 0.005643912591040134, + 0.040271759033203125, + 0.007320757489651442, + 0.04521757736802101, + 0.005243766587227583, + 0.011634551919996738, + 0.010467356070876122, + 0.009714190848171711, + 0.01763080433011055, + 0.009304906241595745, + 0.0398295596241951, + 0.017194095999002457, + 0.010748661123216152, + 0.028511060401797295, + 0.002554125152528286, + 0.0016758048441261053, + 0.0020875022746622562, + 0.004053968470543623, + 0.001659252680838108, + 0.0018218461191281676, + 0.0028104959055781364, + 0.005426282528787851, + 0.0011572487419471145, + 0.0020649791695177555, + 0.0014000111259520054, + 0.0016448464011773467, + 0.002086505526676774, + 0.002506326651200652, + 0.001851544831879437, + 0.0015580293256789446, + 0.002171512460336089, + 0.0011937993112951517, + 0.00235016830265522, + 0.0018660572823137045, + 0.0016287952894344926, + 0.0017152231885120273, + 0.0024522608146071434, + 0.0015143034979701042, + 0.0026433896273374557, + 0.00641381461173296, + 0.003713042940944433, + 0.002103441162034869, + 0.0019598391372710466, + 0.0026347367092967033, + 0.002835312392562628, + 0.004238744266331196, + 0.028486868366599083, + 0.0909472107887268, + 0.03640484809875488, + 0.023078877478837967, + 0.029509345069527626, + 0.042033471167087555, + 0.09682254493236542, + 0.05072135105729103, + 0.007805963978171349, + 0.005924913100898266, + 0.008706721477210522, + 0.011209573596715927, + 0.007473595440387726, + 0.01642192341387272, + 0.007202598731964827, + 0.006540709640830755, + 0.0032647186890244484, + 0.007787784095853567, + 0.008917300961911678, + 0.010009640827775002, + 0.004383321851491928, + 0.0059854681603610516, + 0.008325343020260334, + 0.005471436772495508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.01822569966316223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283382892608643 + ], + "opposite_logprob_tokens": [ + -2.017833709716797 + ], + "logprob_reward": -0.14283382892608643, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283382892608643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.174396350979805, + 0.08966076374053955, + 0.05544936656951904, + 0.026880210265517235, + 0.025362348183989525, + 0.034065406769514084, + 0.1368899643421173, + 0.27136093378067017, + 0.08033304661512375, + 0.10560169070959091 + ], + "credit_map": [ + 0.011434853076934814, + 0.010989670641720295, + 0.01751594804227352, + 0.053167279809713364, + 0.011158229783177376, + 0.0388772189617157, + 0.02035539597272873, + 0.010897751897573471, + 0.006844368763267994, + 0.007356811314821243, + 0.00619040010496974, + 0.009808618575334549, + 0.005630724597722292, + 0.02826971374452114, + 0.005761135369539261, + 0.01979898288846016, + 0.0067129661329090595, + 0.004824502393603325, + 0.009514814242720604, + 0.005058567970991135, + 0.006460090633481741, + 0.00805472582578659, + 0.006882073357701302, + 0.007941617630422115, + 0.0039695692248642445, + 0.0019307414768263698, + 0.0031004101037979126, + 0.003950277343392372, + 0.0022387357894331217, + 0.0031745824962854385, + 0.0036270387936383486, + 0.004888851195573807, + 0.0017748530954122543, + 0.0038501545786857605, + 0.002651349874213338, + 0.005073458421975374, + 0.004319635685533285, + 0.002959667704999447, + 0.0025127148255705833, + 0.0022205107379704714, + 0.004302016459405422, + 0.0036612378899008036, + 0.004701718688011169, + 0.006734825205057859, + 0.002809964818879962, + 0.003586812410503626, + 0.004609586670994759, + 0.0036592420656234026, + 0.016641413792967796, + 0.030663609504699707, + 0.012338048778474331, + 0.010304448194801807, + 0.010226731188595295, + 0.013375460170209408, + 0.02186870574951172, + 0.021471532061696053, + 0.0147619117051363, + 0.08334405720233917, + 0.01613670401275158, + 0.01803421787917614, + 0.017205676063895226, + 0.016010968014597893, + 0.0874217227101326, + 0.018445629626512527, + 0.009207162074744701, + 0.005611265078186989, + 0.012481966987252235, + 0.011503051035106182, + 0.010188966989517212, + 0.018901053816080093, + 0.005881867837160826, + 0.006557708140462637, + 0.003661198541522026, + 0.004821105394512415, + 0.03515060991048813, + 0.01148621179163456, + 0.0047414167784154415, + 0.007092291954904795, + 0.03371709957718849, + 0.004931747447699308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.14283382892608643 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062023747712373734 + ], + "opposite_logprob_tokens": [ + -2.812023639678955 + ], + "logprob_reward": -0.062023747712373734, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062023747712373734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17494401335716248, + 0.08308328688144684, + 0.06987275183200836, + 0.1486102193593979, + 0.026601463556289673, + 0.026552319526672363, + 0.0341431200504303, + 0.25112953782081604, + 0.12182588875293732, + 0.06323739886283875 + ], + "credit_map": [ + 0.01886785961687565, + 0.019070740789175034, + 0.02501109428703785, + 0.027074355632066727, + 0.028684573248028755, + 0.021406596526503563, + 0.018292952328920364, + 0.01653582602739334, + 0.013678819872438908, + 0.004940328188240528, + 0.00669849431142211, + 0.011842925101518631, + 0.005698580294847488, + 0.024344542995095253, + 0.005284366197884083, + 0.010595228523015976, + 0.0068078311160206795, + 0.004181746859103441, + 0.006710510700941086, + 0.007232866249978542, + 0.021018998697400093, + 0.010601022280752659, + 0.006745800841599703, + 0.006573970429599285, + 0.0287631843239069, + 0.0103379525244236, + 0.015209908597171307, + 0.017184296622872353, + 0.013046175241470337, + 0.01731942966580391, + 0.016496920958161354, + 0.030252352356910706, + 0.0026357152964919806, + 0.0033040379639714956, + 0.0027495541144162416, + 0.004496660549193621, + 0.0030097977723926306, + 0.0033371872268617153, + 0.00269299210049212, + 0.00437551736831665, + 0.0030498134437948465, + 0.002288691932335496, + 0.0031841970048844814, + 0.004241695627570152, + 0.0025196082424372435, + 0.0025546427350491285, + 0.005674127954989672, + 0.0030395425856113434, + 0.0035964695271104574, + 0.008606821298599243, + 0.004283595830202103, + 0.003055645851418376, + 0.0024838594254106283, + 0.0038736602291464806, + 0.00320631661452353, + 0.005036754533648491, + 0.015434226021170616, + 0.06254616379737854, + 0.021184956654906273, + 0.014879904687404633, + 0.01752830110490322, + 0.022386176511645317, + 0.07126930356025696, + 0.025900501757860184, + 0.018364479765295982, + 0.009068552404642105, + 0.01923607848584652, + 0.018484575673937798, + 0.012574908323585987, + 0.02292715758085251, + 0.012727338820695877, + 0.008442792110145092, + 0.002466809470206499, + 0.0027194046415388584, + 0.022600509226322174, + 0.006729204207658768, + 0.003517865901812911, + 0.0027720031794160604, + 0.019684016704559326, + 0.002747576916590333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.062023747712373734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, eye-level tracking shot slides slowly and smoothly from left to right, parallel to a massive, towering wall of hydroponic plants within a futuristic agricultural skyscraper. Thousands of vibrant green and crimson heads of lettuce and herbs sprout from a grid of pristine white modules, their leaves lush and perfectly shaped. The entire scene is bathed in the bright, even glow of full-spectrum LED grow lights, eliminating harsh shadows and giving the footage the crisp, observational quality of a high-end documentary. As the camera moves, a fine misting system meticulously activates; nearly invisible nozzles embedded in the structure release a gentle, atomized spray that drifts downward and blankets the foliage. The water vapor catches the light, creating a shimmering veil across the leaves, leaving behind shimmering microdroplets that capture in serene detail the perfect fusion of advanced technology and pristine nature.", + "seed": 12207, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_8916_seed_12207_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03146708104759455, + "vlm_logprob": -0.03146708104759455, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03146708104759455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00955956894904375 + ], + "opposite_logprob_tokens": [ + -5.009559631347656 + ], + "logprob_reward": -0.00955956894904375, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00955956894904375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08690867573022842, + 0.09633786976337433, + 0.11247618496417999, + 0.33168238401412964, + 0.06415587663650513, + 0.10017572343349457, + 0.022166645154356956, + 0.024669254198670387, + 0.0933523178100586, + 0.0680750384926796 + ], + "credit_map": [ + 0.013042247854173183, + 0.006451321765780449, + 0.012057334184646606, + 0.010758322663605213, + 0.008801712654531002, + 0.010262762196362019, + 0.012595316395163536, + 0.012939654290676117, + 0.007098691537976265, + 0.00695975823327899, + 0.006294117774814367, + 0.020255781710147858, + 0.008719277568161488, + 0.013016587123274803, + 0.024084463715553284, + 0.0099091911688447, + 0.007387333549559116, + 0.02567474916577339, + 0.015327390283346176, + 0.034917835146188736, + 0.009076881222426891, + 0.004382758401334286, + 0.009415295906364918, + 0.006293948274105787, + 0.011944690719246864, + 0.021076368167996407, + 0.056584570556879044, + 0.1020822525024414, + 0.058679189532995224, + 0.023497577756643295, + 0.03553425893187523, + 0.022283511236310005, + 0.005445564165711403, + 0.006024875678122044, + 0.009873979724943638, + 0.009092160500586033, + 0.004250630270689726, + 0.006343832239508629, + 0.016486531123518944, + 0.006638302933424711, + 0.009238150902092457, + 0.028181077912449837, + 0.030758032575249672, + 0.0076201437041163445, + 0.006029111333191395, + 0.004391717724502087, + 0.009818265214562416, + 0.004139227326959372, + 0.002450929954648018, + 0.0023027595598250628, + 0.0025486513040959835, + 0.0032091971952468157, + 0.0028151816222816706, + 0.003093369072303176, + 0.0037513594143092632, + 0.0019951967988163233, + 0.0032635866664350033, + 0.002830513985827565, + 0.003979543689638376, + 0.002145809354260564, + 0.003243580926209688, + 0.003040800802409649, + 0.0023254589177668095, + 0.0038399607874453068, + 0.014523537829518318, + 0.010364513844251633, + 0.025679387152194977, + 0.01018130499869585, + 0.0074436538852751255, + 0.006742355413734913, + 0.009049992077052593, + 0.009367577731609344, + 0.005529568064957857, + 0.00626167980954051, + 0.01699099689722061, + 0.006841899361461401, + 0.006409462075680494, + 0.015559817664325237, + 0.004539887886494398, + 0.005941726267337799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.00955956894904375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649660550057888 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.004649660550057888, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649660550057888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06587205827236176, + 0.035358890891075134, + 0.05992168188095093, + 0.053619056940078735, + 0.20235717296600342, + 0.2327156513929367, + 0.01806948333978653, + 0.04617888852953911, + 0.04177059605717659, + 0.24413646757602692 + ], + "credit_map": [ + 0.006213497370481491, + 0.007723163813352585, + 0.010756497271358967, + 0.009337110444903374, + 0.007725061848759651, + 0.00693177804350853, + 0.007241975050419569, + 0.00994297955185175, + 0.002017177874222398, + 0.0025058616884052753, + 0.002394257578998804, + 0.008643083274364471, + 0.00347489258274436, + 0.004517916589975357, + 0.007747625466436148, + 0.004058081656694412, + 0.0027850191108882427, + 0.014615418389439583, + 0.0045157019048929214, + 0.013885541819036007, + 0.0075984918512403965, + 0.004488728009164333, + 0.005478086415678263, + 0.006554704625159502, + 0.0034838097635656595, + 0.0058862194418907166, + 0.009532555937767029, + 0.009557983838021755, + 0.0054139643907547, + 0.007117400411516428, + 0.005895571783185005, + 0.006731560919433832, + 0.01785832643508911, + 0.018144972622394562, + 0.030471140518784523, + 0.03599705919623375, + 0.015066291205585003, + 0.0207065399736166, + 0.04431850463151932, + 0.01979435607790947, + 0.020585544407367706, + 0.06635920703411102, + 0.07157407701015472, + 0.016219381242990494, + 0.015191820450127125, + 0.010999897494912148, + 0.02277798391878605, + 0.009007752873003483, + 0.00312143936753273, + 0.0024459687992930412, + 0.002176933456212282, + 0.0019706282764673233, + 0.0016952011501416564, + 0.0030166374053806067, + 0.0021994358394294977, + 0.0014432412572205067, + 0.00573792913928628, + 0.007271182257682085, + 0.006035655736923218, + 0.004654475022107363, + 0.004788958001881838, + 0.005421920213848352, + 0.008056732825934887, + 0.004212042782455683, + 0.004604578483849764, + 0.007910272106528282, + 0.0037294405046850443, + 0.004268672317266464, + 0.00419414509087801, + 0.0069887274876236916, + 0.005595559254288673, + 0.004479201976209879, + 0.016849089413881302, + 0.014997915364801884, + 0.06725908815860748, + 0.02238069847226143, + 0.019067445769906044, + 0.07257601618766785, + 0.014133689925074577, + 0.016872551292181015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.004649660550057888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088770814239979 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088770814239979, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088770814239979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04597876965999603, + 0.0716363936662674, + 0.08953700214624405, + 0.16070802509784698, + 0.23514673113822937, + 0.07048797607421875, + 0.014294973574578762, + 0.05210047960281372, + 0.05318839102983475, + 0.2069213092327118 + ], + "credit_map": [ + 0.005915971472859383, + 0.0038213070947676897, + 0.005186744965612888, + 0.004545167554169893, + 0.005939007271081209, + 0.007583487778902054, + 0.006832658313214779, + 0.006154422648251057, + 0.0049239457584917545, + 0.005770454183220863, + 0.005007370375096798, + 0.009157727472484112, + 0.010739769786596298, + 0.00849348958581686, + 0.01869591325521469, + 0.008847721852362156, + 0.004907428286969662, + 0.013098797760903835, + 0.016314826905727386, + 0.036364614963531494, + 0.004136768635362387, + 0.0024967652279883623, + 0.007813761942088604, + 0.004404028411954641, + 0.007874966599047184, + 0.013992293737828732, + 0.03306279703974724, + 0.04501095041632652, + 0.0179771538823843, + 0.012833542190492153, + 0.017372500151395798, + 0.012583816424012184, + 0.020289694890379906, + 0.020059674978256226, + 0.034496646374464035, + 0.04115893319249153, + 0.018860088661313057, + 0.0257596243172884, + 0.05074867978692055, + 0.023773381486535072, + 0.005444711539894342, + 0.01978898048400879, + 0.02173253707587719, + 0.0054473793134093285, + 0.00377134932205081, + 0.0035897253546863794, + 0.006883861031383276, + 0.003829427296295762, + 0.0024554345291107893, + 0.002072884701192379, + 0.001651486149057746, + 0.001560123753733933, + 0.001606808160431683, + 0.0024105955380946398, + 0.001351035083644092, + 0.0011866053100675344, + 0.005808145273476839, + 0.007176033221185207, + 0.01141224056482315, + 0.006399619858711958, + 0.007185067981481552, + 0.005284036975353956, + 0.004058996681123972, + 0.004776338580995798, + 0.007552895694971085, + 0.004608928691595793, + 0.017254941165447235, + 0.005859760567545891, + 0.0038250291254371405, + 0.005195522215217352, + 0.004077472724020481, + 0.004813842009752989, + 0.016529954969882965, + 0.013577057980000973, + 0.07005348056554794, + 0.023254910483956337, + 0.012270557694137096, + 0.04637613520026207, + 0.011540863662958145, + 0.013318341225385666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.011088770814239979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001937180757523 + ], + "opposite_logprob_tokens": [ + -2.695019483566284 + ], + "logprob_reward": -0.07001937180757523, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001937180757523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03736048936843872, + 0.027868352830410004, + 0.11105555295944214, + 0.08247624337673187, + 0.06533307582139969, + 0.2685225009918213, + 0.010430562309920788, + 0.019444016739726067, + 0.07111372798681259, + 0.3063954710960388 + ], + "credit_map": [ + 0.004364222288131714, + 0.004138633143156767, + 0.003912897780537605, + 0.00527196004986763, + 0.005389902740716934, + 0.004467325285077095, + 0.004710140638053417, + 0.005105412565171719, + 0.002803713083267212, + 0.002385140396654606, + 0.003583369078114629, + 0.0037469530943781137, + 0.004063939675688744, + 0.0033150978852063417, + 0.0026520329993218184, + 0.005318108014762402, + 0.005696635693311691, + 0.020857717841863632, + 0.018439145758748055, + 0.043798547238111496, + 0.004027539864182472, + 0.00248871766962111, + 0.009401359595358372, + 0.006345890928059816, + 0.0040147206746041775, + 0.006340694148093462, + 0.007873116992413998, + 0.02464507333934307, + 0.024553207680583, + 0.005051637999713421, + 0.005390578415244818, + 0.004607221577316523, + 0.0045026387088000774, + 0.005108713638037443, + 0.014292972162365913, + 0.012012475170195103, + 0.0039243027567863464, + 0.004532692488282919, + 0.015757065266370773, + 0.005202223546802998, + 0.020652420818805695, + 0.07877247035503387, + 0.0868203192949295, + 0.022323094308376312, + 0.012704209424555302, + 0.014159798622131348, + 0.020907865837216377, + 0.012182318605482578, + 0.0013526752591133118, + 0.0011900325771421194, + 0.0017586398171260953, + 0.0018003612058237195, + 0.0007330563967116177, + 0.0012927482603117824, + 0.0015014823293313384, + 0.0008015672792680562, + 0.0014271041145548224, + 0.0019757598638534546, + 0.004352488089352846, + 0.0019238963723182678, + 0.005280895624309778, + 0.0016505770618095994, + 0.0018362075788900256, + 0.0009970898972824216, + 0.010201117023825645, + 0.0050951638258993626, + 0.021586939692497253, + 0.007698101922869682, + 0.004721279256045818, + 0.006834058091044426, + 0.008828084915876389, + 0.006148981861770153, + 0.018397601321339607, + 0.016850490123033524, + 0.09240203350782394, + 0.027767688035964966, + 0.021673956885933876, + 0.09589323401451111, + 0.014620687812566757, + 0.01878979429602623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.07001937180757523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201803311705589 + ], + "opposite_logprob_tokens": [ + -2.8120179176330566 + ], + "logprob_reward": -0.06201803311705589, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201803311705589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05081070214509964, + 0.0575212761759758, + 0.28503042459487915, + 0.1382489651441574, + 0.10506634414196014, + 0.060191962867975235, + 0.053230032324790955, + 0.049910832196474075, + 0.15578335523605347, + 0.04420606046915054 + ], + "credit_map": [ + 0.006329332012683153, + 0.003938484005630016, + 0.006692130118608475, + 0.0062844217754900455, + 0.0066681900061666965, + 0.006282830145210028, + 0.006158188451081514, + 0.008457131683826447, + 0.0058295829221606255, + 0.005437430460005999, + 0.004849775228649378, + 0.007835953496396542, + 0.00672944076359272, + 0.008986727334558964, + 0.009147019125521183, + 0.008705352433025837, + 0.02192790061235428, + 0.04619813710451126, + 0.05489201843738556, + 0.10005155950784683, + 0.01240263506770134, + 0.009335732087492943, + 0.025143137201666832, + 0.015079336240887642, + 0.005314668174833059, + 0.011600240133702755, + 0.013734757900238037, + 0.043546274304389954, + 0.04387635737657547, + 0.007268768735229969, + 0.007451324257999659, + 0.005456581246107817, + 0.0075554451905190945, + 0.010575325228273869, + 0.021587591618299484, + 0.028136897832155228, + 0.007054452784359455, + 0.008166058920323849, + 0.013637249357998371, + 0.008353332057595253, + 0.004750916734337807, + 0.018322521820664406, + 0.020094189792871475, + 0.0038473899476230145, + 0.0038621684070676565, + 0.003123416332527995, + 0.004056711681187153, + 0.0021346493158489466, + 0.006538121495395899, + 0.007180863991379738, + 0.007580731064081192, + 0.007691034581512213, + 0.005816491320729256, + 0.006010900251567364, + 0.008894267491996288, + 0.003517625154927373, + 0.005691962782293558, + 0.007632807362824678, + 0.010272854939103127, + 0.004221275448799133, + 0.008906113915145397, + 0.0051222736947238445, + 0.003777118632569909, + 0.004286426119506359, + 0.022637201473116875, + 0.009757035411894321, + 0.052817851305007935, + 0.016307765617966652, + 0.010351154021918774, + 0.019033964723348618, + 0.013007013127207756, + 0.01187138631939888, + 0.005165998358279467, + 0.0030967756174504757, + 0.014379631727933884, + 0.004855978302657604, + 0.0028445078060030937, + 0.008122354745864868, + 0.00263834185898304, + 0.003102480201050639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.06201803311705589 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, eye-level tracking shot slides slowly and smoothly from left to right, parallel to a massive, towering wall of hydroponic plants within a futuristic agricultural skyscraper. Thousands of vibrant green and crimson heads of lettuce and herbs sprout from a grid of pristine white modules, their leaves lush and perfectly shaped. The entire scene is bathed in the bright, even glow of full-spectrum LED grow lights, eliminating harsh shadows and giving the footage the crisp, observational quality of a high-end documentary. As the camera moves, a fine misting system meticulously activates; nearly invisible nozzles embedded in the structure release a gentle, atomized spray that drifts downward and blankets the foliage. The water vapor catches the light, creating a shimmering veil across the leaves, leaving behind shimmering microdroplets that capture in serene detail the perfect fusion of advanced technology and pristine nature.", + "seed": 12208, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_2/index_8916_seed_12208_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.053405587002635, + "vlm_logprob": -0.053405587002635, + "vlm_yes_no_margin": 3.2749999523162843, + "vlm_token_credit": -0.053405587002635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028571976348757744 + ], + "opposite_logprob_tokens": [ + -3.6535720825195312 + ], + "logprob_reward": -0.028571976348757744, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028571976348757744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05699874460697174, + 0.0806644931435585, + 0.04702390730381012, + 0.03956430032849312, + 0.0819726511836052, + 0.05256768688559532, + 0.022853974252939224, + 0.02580234408378601, + 0.5330696702003479, + 0.05948229506611824 + ], + "credit_map": [ + 0.007485162001103163, + 0.005555215757340193, + 0.010676667094230652, + 0.007182049565017223, + 0.004836095962673426, + 0.007647750899195671, + 0.0075942398980259895, + 0.006021552719175816, + 0.006642310414463282, + 0.017872357740998268, + 0.01031911838799715, + 0.014432732947170734, + 0.008542348630726337, + 0.010158589109778404, + 0.00638923142105341, + 0.006307794246822596, + 0.004614587873220444, + 0.009163602255284786, + 0.004301811568439007, + 0.006402986589819193, + 0.004275816958397627, + 0.0036463409196585417, + 0.009194955229759216, + 0.00542379729449749, + 0.005560999736189842, + 0.003654185216873884, + 0.007367196958512068, + 0.006534263491630554, + 0.0029885913245379925, + 0.005183952394872904, + 0.003737937193363905, + 0.004537165630608797, + 0.016648482531309128, + 0.00513416389003396, + 0.011109406128525734, + 0.007096823304891586, + 0.004868510644882917, + 0.02347390726208687, + 0.004179882351309061, + 0.009461455047130585, + 0.010752120055258274, + 0.004062462132424116, + 0.013681905344128609, + 0.005468576680868864, + 0.004333809949457645, + 0.005515179131180048, + 0.004016675520688295, + 0.004736950621008873, + 0.0035237129777669907, + 0.0020869900472462177, + 0.005413922481238842, + 0.0021984726190567017, + 0.0024838249664753675, + 0.0023992196656763554, + 0.0028117999900132418, + 0.001936025801114738, + 0.004808478523045778, + 0.0040501197800040245, + 0.00544537790119648, + 0.0022049755789339542, + 0.0032384046353399754, + 0.0020823983941227198, + 0.0017206210177391768, + 0.0022519638296216726, + 0.18494729697704315, + 0.03799249231815338, + 0.1893244981765747, + 0.033565301448106766, + 0.022622542455792427, + 0.02417314425110817, + 0.02019610069692135, + 0.020248234272003174, + 0.017038831487298012, + 0.0041281720623373985, + 0.017679715529084206, + 0.007338352967053652, + 0.003505316562950611, + 0.0038838060572743416, + 0.003115482162684202, + 0.0027926089242100716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.028571976348757744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215044677257538 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.014215044677257538, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215044677257538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08530337363481522, + 0.08426527678966522, + 0.04910790175199509, + 0.0798458606004715, + 0.06620349735021591, + 0.06253522634506226, + 0.08186209946870804, + 0.042893800884485245, + 0.37407612800598145, + 0.07390681654214859 + ], + "credit_map": [ + 0.009067628532648087, + 0.006487907841801643, + 0.011122852563858032, + 0.012705997563898563, + 0.008971154689788818, + 0.010959601029753685, + 0.009484975598752499, + 0.01650325022637844, + 0.008582968264818192, + 0.0112454229965806, + 0.014092326164245605, + 0.012291776947677135, + 0.007718658074736595, + 0.010752731002867222, + 0.006258044857531786, + 0.013323348015546799, + 0.006369166541844606, + 0.008718499913811684, + 0.005055880174040794, + 0.00615026755258441, + 0.0045128571800887585, + 0.004627152346074581, + 0.008141249418258667, + 0.005532826762646437, + 0.013142119161784649, + 0.00556940445676446, + 0.012467565946280956, + 0.01338464580476284, + 0.005606032442301512, + 0.01030578650534153, + 0.008248616941273212, + 0.011121691204607487, + 0.011831492185592651, + 0.00438357749953866, + 0.009928626008331776, + 0.007964946329593658, + 0.004147331230342388, + 0.015285843051970005, + 0.004701792728155851, + 0.00795989204198122, + 0.012514672242105007, + 0.004502372816205025, + 0.014264099299907684, + 0.006639047991484404, + 0.005668774247169495, + 0.006102297455072403, + 0.00537146907299757, + 0.007472492754459381, + 0.011139600537717342, + 0.010433654300868511, + 0.013528544455766678, + 0.011458106338977814, + 0.009236360900104046, + 0.007833598181605339, + 0.010145207867026329, + 0.008087019436061382, + 0.006379220634698868, + 0.011223101057112217, + 0.004739317111670971, + 0.005595202092081308, + 0.006806971505284309, + 0.002791533013805747, + 0.002611897885799408, + 0.0027465580496937037, + 0.15468721091747284, + 0.017854411154985428, + 0.15750841796398163, + 0.02082429826259613, + 0.005242385435849428, + 0.004335655830800533, + 0.006213000509887934, + 0.00741076935082674, + 0.021174421533942223, + 0.004060801118612289, + 0.02241736836731434, + 0.0071836113929748535, + 0.005403438117355108, + 0.004810565151274204, + 0.00533266132697463, + 0.003523948136717081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.014215044677257538 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018208257853984833 + ], + "opposite_logprob_tokens": [ + -4.018208026885986 + ], + "logprob_reward": -0.018208257853984833, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018208257853984833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04603281617164612, + 0.03496525436639786, + 0.018786663189530373, + 0.05694645270705223, + 0.026461124420166016, + 0.02964397333562374, + 0.032406702637672424, + 0.05092143267393112, + 0.4823947846889496, + 0.2214408665895462 + ], + "credit_map": [ + 0.005805406719446182, + 0.006884729955345392, + 0.007159873377531767, + 0.005223044194281101, + 0.003889041254296899, + 0.004654119722545147, + 0.005243970081210136, + 0.007172629237174988, + 0.0036724002566188574, + 0.007599198259413242, + 0.004186198115348816, + 0.004442150704562664, + 0.004354396369308233, + 0.003910842351615429, + 0.002985073486343026, + 0.0038149922620505095, + 0.002528772223740816, + 0.0020473382901400328, + 0.0019229252357035875, + 0.004073781426995993, + 0.001400362583808601, + 0.0026653294917196035, + 0.0017281859181821346, + 0.0024199681356549263, + 0.009627515450119972, + 0.004932604264467955, + 0.009329346939921379, + 0.01118194404989481, + 0.003869521664455533, + 0.006708945147693157, + 0.005163280759006739, + 0.006133291870355606, + 0.004857094492763281, + 0.0023719107266515493, + 0.004461224656552076, + 0.002786701312288642, + 0.002620002254843712, + 0.004586704540997744, + 0.001648416044190526, + 0.0031290685292333364, + 0.0062858061864972115, + 0.0023356187157332897, + 0.005900487769395113, + 0.0033217757008969784, + 0.002994627458974719, + 0.0025828636717051268, + 0.002510601421818137, + 0.003712188685312867, + 0.004156653303653002, + 0.003706154879182577, + 0.005158514715731144, + 0.004726462997496128, + 0.00425936933606863, + 0.003074103966355324, + 0.0044476804323494434, + 0.0028777604456990957, + 0.007593603804707527, + 0.012268112041056156, + 0.006400912534445524, + 0.006907125003635883, + 0.007747167721390724, + 0.0034109156113117933, + 0.0031244989950209856, + 0.003469096263870597, + 0.19231827557086945, + 0.023594284430146217, + 0.19512709975242615, + 0.029348265379667282, + 0.00804069172590971, + 0.01154233142733574, + 0.010194755159318447, + 0.012229063548147678, + 0.08757928013801575, + 0.011948509141802788, + 0.09146265685558319, + 0.010759865865111351, + 0.004494517110288143, + 0.003804504871368408, + 0.004995322786271572, + 0.006396173499524593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.018208257853984833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270725131034851 + ], + "opposite_logprob_tokens": [ + -2.12707257270813 + ], + "logprob_reward": -0.1270725131034851, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270725131034851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09364719688892365, + 0.08577242493629456, + 0.08784779161214828, + 0.1319284588098526, + 0.05278797447681427, + 0.06998114287853241, + 0.029923992231488228, + 0.0682656541466713, + 0.125065416097641, + 0.25477996468544006 + ], + "credit_map": [ + 0.009281405247747898, + 0.009390142746269703, + 0.01290199626237154, + 0.023317556828260422, + 0.007027267012745142, + 0.008192072622478008, + 0.011187039315700531, + 0.012349717319011688, + 0.00941446516662836, + 0.01411654893308878, + 0.011830142699182034, + 0.01282822247594595, + 0.008467392064630985, + 0.008647607639431953, + 0.008259646594524384, + 0.012208394706249237, + 0.008372818119823933, + 0.017190096899867058, + 0.008903415873646736, + 0.014826086349785328, + 0.006573100108653307, + 0.005775643978267908, + 0.017907926812767982, + 0.008298708125948906, + 0.015433505177497864, + 0.012908754870295525, + 0.022644592449069023, + 0.032665569335222244, + 0.010935459285974503, + 0.00929541140794754, + 0.012866677716374397, + 0.015178486704826355, + 0.009028496220707893, + 0.0054625095799565315, + 0.010701002553105354, + 0.004917769227176905, + 0.005958970636129379, + 0.006757188588380814, + 0.004822242073714733, + 0.0051397900097072124, + 0.0133152324706316, + 0.0046157389879226685, + 0.01643168181180954, + 0.006804629694670439, + 0.006134402472525835, + 0.007632583845406771, + 0.00500378618016839, + 0.010043083690106869, + 0.004263579845428467, + 0.003766047302633524, + 0.004206595476716757, + 0.0034823338501155376, + 0.003160597523674369, + 0.0038667814806103706, + 0.0032472724560648203, + 0.003930784296244383, + 0.009468386881053448, + 0.02050183340907097, + 0.006559763569384813, + 0.00889578927308321, + 0.00946940016001463, + 0.004640377126634121, + 0.0049655986949801445, + 0.003764505498111248, + 0.046137597411870956, + 0.006699038669466972, + 0.04719853028655052, + 0.009162969887256622, + 0.0031785648316144943, + 0.0045095039531588554, + 0.004165606573224068, + 0.004013600293546915, + 0.10486332327127457, + 0.011644893325865269, + 0.10917527973651886, + 0.01119064912199974, + 0.004206342622637749, + 0.00359016889706254, + 0.0043460652232170105, + 0.005763225257396698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.1270725131034851 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a futuristic agricultural skyscraper with a massive wall of hydroponic plants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of vibrant green and crimson heads of lettuce and herbs visible growing from a grid of pristine white modules?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by bright, even full-spectrum LED grow lights that eliminate harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a fine misting system activate, releasing a gentle spray that creates a shimmering veil of water vapor across the foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level tracking shot that slides slowly and smoothly from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896014302968979 + ], + "opposite_logprob_tokens": [ + -2.5789601802825928 + ], + "logprob_reward": -0.07896014302968979, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896014302968979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08729678392410278, + 0.05135754123330116, + 0.17732541263103485, + 0.06909049302339554, + 0.039397723972797394, + 0.04094356298446655, + 0.030085870996117592, + 0.07331608235836029, + 0.21984928846359253, + 0.21133728325366974 + ], + "credit_map": [ + 0.008011860772967339, + 0.0070960624143481255, + 0.013142944313585758, + 0.010575910098850727, + 0.008579808287322521, + 0.018054453656077385, + 0.010762806981801987, + 0.011072930879890919, + 0.00496387155726552, + 0.010802899487316608, + 0.006871463730931282, + 0.008998342789709568, + 0.006299760192632675, + 0.005050542298704386, + 0.003879653988406062, + 0.004491004627197981, + 0.01868467777967453, + 0.035712871700525284, + 0.01788286119699478, + 0.02168622612953186, + 0.016286741942167282, + 0.012561887502670288, + 0.03737469017505646, + 0.01713545434176922, + 0.010275532491505146, + 0.005877989809960127, + 0.011676717549562454, + 0.017315935343503952, + 0.005065439268946648, + 0.0056495144963264465, + 0.007389902602881193, + 0.00583946006372571, + 0.006079497747123241, + 0.0042851772159338, + 0.007678051013499498, + 0.00427598599344492, + 0.005734188482165337, + 0.0034514819271862507, + 0.002665103180333972, + 0.00522823678329587, + 0.008920540101826191, + 0.004013031721115112, + 0.008693224750459194, + 0.003971479833126068, + 0.004203573800623417, + 0.0036466368474066257, + 0.004128038417547941, + 0.0033670393750071526, + 0.004544416908174753, + 0.003033625427633524, + 0.005091944709420204, + 0.003864469937980175, + 0.0022134208120405674, + 0.00571770966053009, + 0.002774248830974102, + 0.0028460349421948195, + 0.01042719092220068, + 0.01730979047715664, + 0.007238047197461128, + 0.008374422788619995, + 0.010397844947874546, + 0.006055500358343124, + 0.008999844081699848, + 0.004513447172939777, + 0.08384151756763458, + 0.014482883736491203, + 0.0844997763633728, + 0.013772060163319111, + 0.004327643662691116, + 0.006555517669767141, + 0.0064375740475952625, + 0.005932309199124575, + 0.07614690065383911, + 0.009840982034802437, + 0.07994801551103592, + 0.009510282427072525, + 0.009318209253251553, + 0.009606710635125637, + 0.008559314534068108, + 0.00840687844902277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.07896014302968979 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_tensor([3775, 3775])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_tensor([3775, 3775])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c813bc98e4b3d3578e46828274f3018ab34655d4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_tensor([3775, 3775])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot follows a figure as they walk through a dense, ancient forest at a steady, cautious pace in the depths of night. The camera moves smoothly with them, capturing their journey along a narrow, winding dirt path covered with damp, fallen leaves. The figure is completely enveloped in a heavy, dark, floor-length, hooded cloak made of a rough wool material, its folds swaying gently as they move, their face hidden in deep shadow. In their right hand, they hold a rustic black iron lantern, swinging rhythmically, a soft flame visible through the glass window. This lone light source casts a dynamic, warm, golden-orange glow, illuminating the nearby ground and capturing the textures of gnarled tree roots and damp leaves, while the rest of the forest is plunged into near-total darkness. A thick, ethereal mist covers the forest floor, swirling around the figure's ankles, catching the lantern's light and creating a soft, hazy atmosphere. The swaying motion of the lanterns causes light, long dancing shadows to wander across the mossy trunks of the surrounding trees, creating a mysterious, serene, and atmospheric scene of a solitary passage.", + "seed": 15606, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_3775_seed_15606_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11959670670330524, + "vlm_logprob": -0.11959670670330524, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.11959670670330524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28976133465766907 + ], + "opposite_logprob_tokens": [ + -2.0397613048553467 + ], + "logprob_reward": -0.28976133465766907, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.28976133465766907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13757507503032684, + 0.02781248837709427, + 0.020240573212504387, + 0.0586928129196167, + 0.22972629964351654, + 0.03223438560962677, + 0.21817044913768768, + 0.20084348320960999, + 0.04628393054008484, + 0.02842053584754467 + ], + "credit_map": [ + 0.012910180725157261, + 0.017260950058698654, + 0.013693638145923615, + 0.01760830357670784, + 0.01747213490307331, + 0.04129493236541748, + 0.006007816176861525, + 0.011327100917696953, + 0.0024167343508452177, + 0.009029826149344444, + 0.0023246004711836576, + 0.0044953059405088425, + 0.0026698391884565353, + 0.0042694187723100185, + 0.0010413728887215257, + 0.0015653870068490505, + 0.0015043828170746565, + 0.0014057555235922337, + 0.0020490791648626328, + 0.003709196113049984, + 0.005933499429374933, + 0.002997002098709345, + 0.0015020905993878841, + 0.0011395657202228904, + 0.004807669203728437, + 0.004711184650659561, + 0.00574732618406415, + 0.011162985116243362, + 0.012859716080129147, + 0.005732744466513395, + 0.010570559650659561, + 0.0031006194185465574, + 0.016516657546162605, + 0.043009866029024124, + 0.020509138703346252, + 0.08321678638458252, + 0.026384513825178146, + 0.012639901600778103, + 0.014880289323627949, + 0.012569120153784752, + 0.004960232879966497, + 0.004559104796499014, + 0.005427131894975901, + 0.004153063055127859, + 0.0036029594484716654, + 0.002599199302494526, + 0.0027368967421352863, + 0.004195798188447952, + 0.01574833132326603, + 0.03187236934900284, + 0.042010847479104996, + 0.017385631799697876, + 0.08387375622987747, + 0.009987953118979931, + 0.010493655689060688, + 0.00679788226261735, + 0.013811133801937103, + 0.027978762984275818, + 0.024133997038006783, + 0.024175306782126427, + 0.03448868542909622, + 0.03689359501004219, + 0.019969098269939423, + 0.019392888993024826, + 0.003271261928603053, + 0.009081542491912842, + 0.008061898872256279, + 0.010936594568192959, + 0.0029334763530641794, + 0.00480831740424037, + 0.0038435643073171377, + 0.0033472718205302954, + 0.0019693931099027395, + 0.0037161807995289564, + 0.0043638925999403, + 0.0059776343405246735, + 0.002900018822401762, + 0.003976055886596441, + 0.0023475850466638803, + 0.0031697736121714115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.28976133465766907 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14308689534664154 + ], + "opposite_logprob_tokens": [ + -2.0180869102478027 + ], + "logprob_reward": -0.14308689534664154, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14308689534664154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45364615321159363, + 0.04962949827313423, + 0.01382957398891449, + 0.03147263824939728, + 0.01859910972416401, + 0.034483540803194046, + 0.12751741707324982, + 0.17078115046024323, + 0.0772416815161705, + 0.022799290716648102 + ], + "credit_map": [ + 0.04573472961783409, + 0.0815357118844986, + 0.04458637535572052, + 0.07553745061159134, + 0.09803318977355957, + 0.06546842306852341, + 0.019122308120131493, + 0.023627961054444313, + 0.002469700062647462, + 0.017043597996234894, + 0.0033290523570030928, + 0.01593485288321972, + 0.0038122243713587523, + 0.0038123424164950848, + 0.001474185031838715, + 0.0017535460647195578, + 0.0020885150879621506, + 0.0015189703553915024, + 0.0017388563137501478, + 0.0022223866544663906, + 0.0020559863187372684, + 0.0020665889605879784, + 0.0010886554373428226, + 0.001049614860676229, + 0.003197544952854514, + 0.002327161142602563, + 0.003840056248009205, + 0.005053058732300997, + 0.006366994231939316, + 0.0035380686167627573, + 0.005446412600576878, + 0.0017033397452905774, + 0.002180099254474044, + 0.0018138923915103078, + 0.002100948244333267, + 0.003958864603191614, + 0.002840693574398756, + 0.0014060954563319683, + 0.0027949109207838774, + 0.001503605511970818, + 0.0015026851324364543, + 0.0029179020784795284, + 0.002302426379173994, + 0.010200339369475842, + 0.013051281683146954, + 0.0015270524891093373, + 0.001734321122057736, + 0.0012475305702537298, + 0.008593447506427765, + 0.017327480018138885, + 0.025920424610376358, + 0.008975577540695667, + 0.05195794627070427, + 0.005203313659876585, + 0.006384757813066244, + 0.0031544652301818132, + 0.015394165180623531, + 0.035364508628845215, + 0.020687779411673546, + 0.028029851615428925, + 0.021225648000836372, + 0.024082276970148087, + 0.014970655553042889, + 0.01102626882493496, + 0.005177648738026619, + 0.016065381467342377, + 0.014941244386136532, + 0.019855841994285583, + 0.0029806853272020817, + 0.008186371996998787, + 0.0062342481687664986, + 0.0038002554792910814, + 0.0022848527878522873, + 0.0033638225868344307, + 0.0037629008293151855, + 0.003571566194295883, + 0.002607013564556837, + 0.003226348664611578, + 0.002090832684189081, + 0.0018919549183920026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.14308689534664154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10037490725517273 + ], + "opposite_logprob_tokens": [ + -2.350374937057495 + ], + "logprob_reward": -0.10037490725517273, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10037490725517273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07985266298055649, + 0.12140761315822601, + 0.01530089695006609, + 0.07703201472759247, + 0.1322893351316452, + 0.13095127046108246, + 0.11979997903108597, + 0.20130017399787903, + 0.08488626033067703, + 0.03717983141541481 + ], + "credit_map": [ + 0.005835202056914568, + 0.008997964672744274, + 0.008354466408491135, + 0.009350975044071674, + 0.008689677342772484, + 0.025285717099905014, + 0.004666697699576616, + 0.008671955205500126, + 0.0029766694642603397, + 0.050970952957868576, + 0.004476670641452074, + 0.04478060454130173, + 0.007537883706390858, + 0.005612018518149853, + 0.0027433522045612335, + 0.0023094541393220425, + 0.0017051311442628503, + 0.0010196242947131395, + 0.001980549655854702, + 0.0020394816529005766, + 0.002659248188138008, + 0.002901792759075761, + 0.0014207650674507022, + 0.0015743037220090628, + 0.004432599525898695, + 0.0060838451609015465, + 0.006589548662304878, + 0.010377361439168453, + 0.013707481324672699, + 0.015493752434849739, + 0.015011530369520187, + 0.005335887894034386, + 0.01284826174378395, + 0.01718847267329693, + 0.011466342955827713, + 0.040680892765522, + 0.015490132384002209, + 0.012121588923037052, + 0.009370991960167885, + 0.01312263309955597, + 0.006765540689229965, + 0.01058126986026764, + 0.008227607235312462, + 0.04084809124469757, + 0.05005795136094093, + 0.00473394151777029, + 0.006825980264693499, + 0.0029108829330652952, + 0.008215595968067646, + 0.01535565685480833, + 0.025525392964482307, + 0.010299804620444775, + 0.0450994074344635, + 0.006228748243302107, + 0.005882096942514181, + 0.0031932578422129154, + 0.01881544105708599, + 0.04237823188304901, + 0.025912733748555183, + 0.03497582674026489, + 0.02910035103559494, + 0.026304831728339195, + 0.014054372906684875, + 0.009758375585079193, + 0.006660090293735266, + 0.017459524795413017, + 0.02057548426091671, + 0.015184789896011353, + 0.005096303764730692, + 0.008828580379486084, + 0.005261189769953489, + 0.005820294842123985, + 0.0036959368735551834, + 0.004612661898136139, + 0.00612676190212369, + 0.007435676176100969, + 0.0029405297245830297, + 0.005961147136986256, + 0.003104477422311902, + 0.0033026353921741247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.10037490725517273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016096873208880424 + ], + "opposite_logprob_tokens": [ + -4.141097068786621 + ], + "logprob_reward": -0.016096873208880424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096873208880424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14458931982517242, + 0.03877894952893257, + 0.014158390462398529, + 0.05687684565782547, + 0.06683631241321564, + 0.02285131998360157, + 0.10159613192081451, + 0.46457770466804504, + 0.056630294770002365, + 0.0331047922372818 + ], + "credit_map": [ + 0.01394401490688324, + 0.022708414122462273, + 0.025133440271019936, + 0.017674677073955536, + 0.021053437143564224, + 0.028826462104916573, + 0.005526568274945021, + 0.009722287766635418, + 0.0032483055256307125, + 0.010476811788976192, + 0.0036562359891831875, + 0.010511272586882114, + 0.0036641163751482964, + 0.004091989248991013, + 0.001248399494215846, + 0.0018818109529092908, + 0.001450805109925568, + 0.0015961186727508903, + 0.0017988120671361685, + 0.00263724522665143, + 0.0025031687691807747, + 0.0022192003671079874, + 0.0009218586492352188, + 0.001031179679557681, + 0.006955916061997414, + 0.005202117376029491, + 0.006982750725001097, + 0.010054850950837135, + 0.010161745361983776, + 0.007157699670642614, + 0.007378211710602045, + 0.0029835428576916456, + 0.0032453215681016445, + 0.017574066296219826, + 0.004705456551164389, + 0.024772558361291885, + 0.004339166916906834, + 0.0039159576408565044, + 0.004277572501450777, + 0.004006205592304468, + 0.0033395136706531048, + 0.0029697429854422808, + 0.00427332054823637, + 0.0033460347913205624, + 0.002608419628813863, + 0.00183648313395679, + 0.0018501156009733677, + 0.002627686131745577, + 0.0075073991902172565, + 0.015997160226106644, + 0.018124626949429512, + 0.009055287577211857, + 0.037966288626194, + 0.0046909041702747345, + 0.005126320756971836, + 0.0031281292904168367, + 0.03847979009151459, + 0.08073458820581436, + 0.05708937719464302, + 0.06500551104545593, + 0.06838057935237885, + 0.0750119611620903, + 0.03803296387195587, + 0.04184287413954735, + 0.003994353581219912, + 0.011810263618826866, + 0.012658941559493542, + 0.012471955269575119, + 0.0031740167178213596, + 0.00600812304764986, + 0.0036529824137687683, + 0.002859652740880847, + 0.004238040652126074, + 0.004008245188742876, + 0.004860239569097757, + 0.004691219888627529, + 0.0027186505030840635, + 0.005778140388429165, + 0.0035413186997175217, + 0.003268931061029434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.016096873208880424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866352304816246 + ], + "opposite_logprob_tokens": [ + -3.04866361618042 + ], + "logprob_reward": -0.04866352304816246, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866352304816246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15457239747047424, + 0.07841204106807709, + 0.062356892973184586, + 0.03491279482841492, + 0.2165602147579193, + 0.04376424476504326, + 0.16202569007873535, + 0.08170539140701294, + 0.14012554287910461, + 0.025564758107066154 + ], + "credit_map": [ + 0.014840729534626007, + 0.024904821068048477, + 0.017269734293222427, + 0.024073859676718712, + 0.02230115607380867, + 0.03346101567149162, + 0.007638473063707352, + 0.010082626715302467, + 0.003893933491781354, + 0.016551580280065536, + 0.005832877475768328, + 0.0255267471075058, + 0.008284871466457844, + 0.010995934717357159, + 0.0030231710989028215, + 0.0043029277585446835, + 0.0041836644522845745, + 0.004584942013025284, + 0.006091432645916939, + 0.01379307173192501, + 0.01840241253376007, + 0.008494672365486622, + 0.003955523483455181, + 0.002851178403943777, + 0.0032580127008259296, + 0.002803537528961897, + 0.003772157710045576, + 0.005713007878512144, + 0.008197806775569916, + 0.00416571693494916, + 0.00457626162096858, + 0.002426298102363944, + 0.01773696206510067, + 0.02974742464721203, + 0.020667804405093193, + 0.0718778669834137, + 0.0231152530759573, + 0.018112068995833397, + 0.01271869894117117, + 0.02258416637778282, + 0.0038157403469085693, + 0.004389723297208548, + 0.004639118444174528, + 0.010066498070955276, + 0.011033650487661362, + 0.0035988260060548782, + 0.003275725757703185, + 0.002944967010989785, + 0.010686769150197506, + 0.02345876581966877, + 0.03622853010892868, + 0.020301533862948418, + 0.049108024686574936, + 0.008046825416386127, + 0.007371044252067804, + 0.006824200972914696, + 0.006112434435635805, + 0.012876462191343307, + 0.013444946147501469, + 0.012064099311828613, + 0.009977001696825027, + 0.010502781718969345, + 0.00751840602606535, + 0.00920927431434393, + 0.008191435597836971, + 0.035365261137485504, + 0.03965503349900246, + 0.021830042824149132, + 0.0064619602635502815, + 0.012217888608574867, + 0.008012581616640091, + 0.00839135143905878, + 0.0029805172234773636, + 0.0045881113037467, + 0.00374016584828496, + 0.003779761493206024, + 0.002406585728749633, + 0.0036134880501776934, + 0.002391698071733117, + 0.0020644322503358126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.04866352304816246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a figure as they walk through a dense, ancient forest at a steady, cautious pace in the depths of night. The camera moves smoothly with them, capturing their journey along a narrow, winding dirt path covered with damp, fallen leaves. The figure is completely enveloped in a heavy, dark, floor-length, hooded cloak made of a rough wool material, its folds swaying gently as they move, their face hidden in deep shadow. In their right hand, they hold a rustic black iron lantern, swinging rhythmically, a soft flame visible through the glass window. This lone light source casts a dynamic, warm, golden-orange glow, illuminating the nearby ground and capturing the textures of gnarled tree roots and damp leaves, while the rest of the forest is plunged into near-total darkness. A thick, ethereal mist covers the forest floor, swirling around the figure's ankles, catching the lantern's light and creating a soft, hazy atmosphere. The swaying motion of the lanterns causes light, long dancing shadows to wander across the mossy trunks of the surrounding trees, creating a mysterious, serene, and atmospheric scene of a solitary passage.", + "seed": 15607, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_3775_seed_15607_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0694687376730144, + "vlm_logprob": -0.0694687376730144, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.0694687376730144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15899822115898132 + ], + "opposite_logprob_tokens": [ + -2.4089982509613037 + ], + "logprob_reward": -0.15899822115898132, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.15899822115898132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.167952299118042, + 0.17302349209785461, + 0.040009066462516785, + 0.05251850187778473, + 0.1457364559173584, + 0.09595037251710892, + 0.12419579923152924, + 0.03933633863925934, + 0.09871529787778854, + 0.06256245821714401 + ], + "credit_map": [ + 0.013509934768080711, + 0.016064023599028587, + 0.02721208520233631, + 0.018561948090791702, + 0.027395939454436302, + 0.018978910520672798, + 0.03164895251393318, + 0.014580490998923779, + 0.007743261754512787, + 0.020186640322208405, + 0.008635272271931171, + 0.013693120330572128, + 0.04974748194217682, + 0.015886714681982994, + 0.03408576548099518, + 0.02304523065686226, + 0.0026246937923133373, + 0.005861630663275719, + 0.005518283694982529, + 0.005997239612042904, + 0.007347226608544588, + 0.0036891980562359095, + 0.005427236668765545, + 0.0035435508470982313, + 0.002701808465644717, + 0.011906680651009083, + 0.0038797289598733187, + 0.01245612371712923, + 0.007295144256204367, + 0.0036220699548721313, + 0.005772235803306103, + 0.0048847016878426075, + 0.005649962462484837, + 0.043887410312891006, + 0.015569055452942848, + 0.04205401986837387, + 0.007666600868105888, + 0.01076097134500742, + 0.007300261873751879, + 0.012848170474171638, + 0.0061834524385631084, + 0.027673693373799324, + 0.008779548108577728, + 0.02341487631201744, + 0.009119002148509026, + 0.00598681066185236, + 0.006086310837417841, + 0.00870666652917862, + 0.009363632649183273, + 0.016899265348911285, + 0.019300615414977074, + 0.04339956119656563, + 0.005755327176302671, + 0.013154158368706703, + 0.007318786345422268, + 0.00900444108992815, + 0.005834481213241816, + 0.005052234511822462, + 0.00544630317017436, + 0.0039028972387313843, + 0.0030973961111158133, + 0.005631876178085804, + 0.004064582288265228, + 0.006306565832346678, + 0.010599171742796898, + 0.00933772511780262, + 0.013729848898947239, + 0.013413509353995323, + 0.013350077904760838, + 0.007828875444829464, + 0.016686903312802315, + 0.0137691805139184, + 0.005142678506672382, + 0.004560356494039297, + 0.009018771350383759, + 0.008277189917862415, + 0.009669309481978416, + 0.008119122125208378, + 0.0128273144364357, + 0.004947701469063759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.15899822115898132 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06216670200228691 + ], + "opposite_logprob_tokens": [ + -2.812166690826416 + ], + "logprob_reward": -0.06216670200228691, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06216670200228691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09250456839799881, + 0.10275563597679138, + 0.05614129453897476, + 0.09753554314374924, + 0.07217679172754288, + 0.13551202416419983, + 0.062162965536117554, + 0.048771701753139496, + 0.270817369222641, + 0.061622120440006256 + ], + "credit_map": [ + 0.008824633434414864, + 0.008238117210566998, + 0.020200354978442192, + 0.010635028593242168, + 0.014428069815039635, + 0.009362954646348953, + 0.012390398420393467, + 0.008425014093518257, + 0.009093997068703175, + 0.010598217137157917, + 0.013130784966051579, + 0.012470814399421215, + 0.017599904909729958, + 0.008437320590019226, + 0.015205930918455124, + 0.01621866598725319, + 0.004381758626550436, + 0.009612847119569778, + 0.006396172568202019, + 0.009698661975562572, + 0.006424714345484972, + 0.006972045637667179, + 0.007875203154981136, + 0.004779892973601818, + 0.00562761165201664, + 0.02058517187833786, + 0.011530623771250248, + 0.023095954209566116, + 0.010430071502923965, + 0.006283152382820845, + 0.00942155346274376, + 0.010561401955783367, + 0.0026270670350641012, + 0.024828124791383743, + 0.006586418021470308, + 0.024716677144169807, + 0.003449550596997142, + 0.0031035568099468946, + 0.0030248756520450115, + 0.0038405186496675014, + 0.008822198025882244, + 0.028456497937440872, + 0.01287307683378458, + 0.02661435678601265, + 0.013238740153610706, + 0.007059130351990461, + 0.019776150584220886, + 0.01867186650633812, + 0.006028672680258751, + 0.010112248361110687, + 0.009301908314228058, + 0.01231446210294962, + 0.0042739673517644405, + 0.006070130970329046, + 0.004959100857377052, + 0.009102477692067623, + 0.006459547206759453, + 0.003377835964784026, + 0.007821314036846161, + 0.004106269218027592, + 0.006626185029745102, + 0.006832313258200884, + 0.007768274284899235, + 0.0057799615897238255, + 0.030961452051997185, + 0.030754731968045235, + 0.041747283190488815, + 0.040220748633146286, + 0.029393969103693962, + 0.02171696536242962, + 0.0445181205868721, + 0.03150411322712898, + 0.0069265346974134445, + 0.006347263231873512, + 0.009078264236450195, + 0.009199190884828568, + 0.008915492333471775, + 0.006562915164977312, + 0.007886846549808979, + 0.006705612409859896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.06216670200228691 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033070296049118 + ], + "opposite_logprob_tokens": [ + -2.3503305912017822 + ], + "logprob_reward": -0.10033070296049118, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033070296049118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24635115265846252, + 0.1393774300813675, + 0.08511991798877716, + 0.07857472449541092, + 0.15890640020370483, + 0.07240262627601624, + 0.024674711748957634, + 0.057726386934518814, + 0.09576389938592911, + 0.041102707386016846 + ], + "credit_map": [ + 0.017088159918785095, + 0.02606385201215744, + 0.03140707686543465, + 0.041216012090444565, + 0.046059444546699524, + 0.023273397237062454, + 0.039196427911520004, + 0.022046785801649094, + 0.0068906499072909355, + 0.013147556222975254, + 0.00910235196352005, + 0.010083544999361038, + 0.04778093099594116, + 0.01119207963347435, + 0.0328967347741127, + 0.008283584378659725, + 0.004074481315910816, + 0.01627032645046711, + 0.00486359978094697, + 0.013687439262866974, + 0.01670321822166443, + 0.009894057177007198, + 0.011663135141134262, + 0.007963661104440689, + 0.005379880778491497, + 0.01311829686164856, + 0.009303468279540539, + 0.014506587758660316, + 0.008489488624036312, + 0.010746534913778305, + 0.006409880239516497, + 0.010620585642755032, + 0.0076448265463113785, + 0.04146004840731621, + 0.0113434549421072, + 0.03890876844525337, + 0.01685917377471924, + 0.014004156924784184, + 0.016257090494036674, + 0.012428874149918556, + 0.004972434602677822, + 0.016595425084233284, + 0.007889476604759693, + 0.015169460326433182, + 0.006784403696656227, + 0.005128440912812948, + 0.007485331501811743, + 0.00837765634059906, + 0.0029667068738490343, + 0.0033188993111252785, + 0.003653861116617918, + 0.004304979927837849, + 0.002538853557780385, + 0.0031649935990571976, + 0.0025667608715593815, + 0.0021596571896225214, + 0.006481275428086519, + 0.006252235267311335, + 0.009116940200328827, + 0.004904141649603844, + 0.005492064636200666, + 0.0062241628766059875, + 0.0081013273447752, + 0.011154239997267723, + 0.010597694665193558, + 0.011304018087685108, + 0.01165709737688303, + 0.015566997230052948, + 0.012950628995895386, + 0.010013114660978317, + 0.013514641672372818, + 0.010159705765545368, + 0.0053079393692314625, + 0.007515041623264551, + 0.005576878320425749, + 0.005581595469266176, + 0.0035918676294386387, + 0.004025318659842014, + 0.004676709417253733, + 0.004827356897294521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.10033070296049118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264704581350088 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264704581350088, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264704581350088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06225144863128662, + 0.1029890701174736, + 0.07117785513401031, + 0.09590000659227371, + 0.09236189723014832, + 0.09681244939565659, + 0.2743968963623047, + 0.0747581496834755, + 0.08648604154586792, + 0.04286620393395424 + ], + "credit_map": [ + 0.005190098658204079, + 0.005624261684715748, + 0.010595534928143024, + 0.006478363182395697, + 0.009752730838954449, + 0.008621876128017902, + 0.006167510058730841, + 0.009821070358157158, + 0.008117753081023693, + 0.010568044148385525, + 0.018663212656974792, + 0.010889089666306973, + 0.023181475698947906, + 0.009455151855945587, + 0.013526558876037598, + 0.00858776830136776, + 0.004331103526055813, + 0.011374475434422493, + 0.008701595477759838, + 0.011098871938884258, + 0.01380516029894352, + 0.007143505848944187, + 0.010036599822342396, + 0.004686531610786915, + 0.004865223076194525, + 0.01910059154033661, + 0.011049497872591019, + 0.022352319210767746, + 0.009408386424183846, + 0.009786251932382584, + 0.009717177599668503, + 0.009620553813874722, + 0.002933506155386567, + 0.03174462914466858, + 0.006428624503314495, + 0.0307761300355196, + 0.006150561384856701, + 0.004762051627039909, + 0.004574298858642578, + 0.004992091096937656, + 0.007673988118767738, + 0.020320452749729156, + 0.01472723949700594, + 0.018546350300312042, + 0.007771933451294899, + 0.00702333590015769, + 0.009440257214009762, + 0.011308876797556877, + 0.01414563786238432, + 0.036343276500701904, + 0.04362499341368675, + 0.09248653054237366, + 0.017278240993618965, + 0.026863813400268555, + 0.020295657217502594, + 0.02335873246192932, + 0.007223324850201607, + 0.004741854965686798, + 0.01536473911255598, + 0.006497801747173071, + 0.008362696506083012, + 0.009156354703009129, + 0.014236828312277794, + 0.009174542501568794, + 0.015842901542782784, + 0.007288837805390358, + 0.01509034913033247, + 0.007381462026387453, + 0.009535902179777622, + 0.011548858135938644, + 0.01039840467274189, + 0.009399322792887688, + 0.006931244861334562, + 0.0045877136290073395, + 0.006481726188212633, + 0.0036894965451210737, + 0.0037113879807293415, + 0.005424600560218096, + 0.0063506909646093845, + 0.005689338780939579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.005264704581350088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1283048391342163, + 0.22760339081287384, + 0.10938002169132233, + 0.04470241442322731, + 0.05807318910956383, + 0.0624384880065918, + 0.07091379910707474, + 0.06603769958019257, + 0.14189156889915466, + 0.09065456688404083 + ], + "credit_map": [ + 0.013350007124245167, + 0.013928358443081379, + 0.021031849086284637, + 0.01307735126465559, + 0.0216366034001112, + 0.01449130941182375, + 0.01656940206885338, + 0.014219970442354679, + 0.01770259067416191, + 0.02247968688607216, + 0.04134348779916763, + 0.02277207560837269, + 0.05720701441168785, + 0.017780516296625137, + 0.03409138694405556, + 0.014226636849343777, + 0.008477509953081608, + 0.01777433604001999, + 0.018025193363428116, + 0.01531932782381773, + 0.01778429187834263, + 0.00873654056340456, + 0.014955606311559677, + 0.008307216688990593, + 0.002737311413511634, + 0.007307490333914757, + 0.0059674945659935474, + 0.010428348556160927, + 0.004296515602618456, + 0.005564365070313215, + 0.003348291851580143, + 0.005052594467997551, + 0.00384480319917202, + 0.011308329179883003, + 0.007699590642005205, + 0.009910614229738712, + 0.007296787574887276, + 0.006018023006618023, + 0.005008218809962273, + 0.006986822001636028, + 0.0048895347863435745, + 0.014639480970799923, + 0.007216814439743757, + 0.012275496497750282, + 0.006220535840839148, + 0.003842507489025593, + 0.006648306269198656, + 0.006705806590616703, + 0.006615467835217714, + 0.014309807680547237, + 0.008133020251989365, + 0.012775002047419548, + 0.007328194100409746, + 0.0062867566011846066, + 0.005941519979387522, + 0.009524032473564148, + 0.006442013196647167, + 0.0062471614219248295, + 0.010251813568174839, + 0.007518719881772995, + 0.008423782885074615, + 0.007119919639080763, + 0.010320370085537434, + 0.009713920764625072, + 0.020841281861066818, + 0.01819959469139576, + 0.01873645931482315, + 0.025970740243792534, + 0.01612153649330139, + 0.0141018470749259, + 0.017541343346238136, + 0.010378770530223846, + 0.01065242663025856, + 0.00942834373563528, + 0.01084991917014122, + 0.01209280826151371, + 0.01358183566480875, + 0.009780877269804478, + 0.014979013241827488, + 0.009289341978728771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.02058335766196251 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a figure as they walk through a dense, ancient forest at a steady, cautious pace in the depths of night. The camera moves smoothly with them, capturing their journey along a narrow, winding dirt path covered with damp, fallen leaves. The figure is completely enveloped in a heavy, dark, floor-length, hooded cloak made of a rough wool material, its folds swaying gently as they move, their face hidden in deep shadow. In their right hand, they hold a rustic black iron lantern, swinging rhythmically, a soft flame visible through the glass window. This lone light source casts a dynamic, warm, golden-orange glow, illuminating the nearby ground and capturing the textures of gnarled tree roots and damp leaves, while the rest of the forest is plunged into near-total darkness. A thick, ethereal mist covers the forest floor, swirling around the figure's ankles, catching the lantern's light and creating a soft, hazy atmosphere. The swaying motion of the lanterns causes light, long dancing shadows to wander across the mossy trunks of the surrounding trees, creating a mysterious, serene, and atmospheric scene of a solitary passage.", + "seed": 15608, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_3775_seed_15608_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09105788869783282, + "vlm_logprob": -0.09105788869783282, + "vlm_yes_no_margin": 2.924999976158142, + "vlm_token_credit": -0.09105788869783282, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1645391285419464 + ], + "opposite_logprob_tokens": [ + -2.289539098739624 + ], + "logprob_reward": -0.1645391285419464, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1645391285419464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14562475681304932, + 0.11085812747478485, + 0.14393970370292664, + 0.05957811325788498, + 0.05846822261810303, + 0.09369850903749466, + 0.15025711059570312, + 0.0648684874176979, + 0.12248009443283081, + 0.05022687092423439 + ], + "credit_map": [ + 0.014360720291733742, + 0.014348779805004597, + 0.019477112218737602, + 0.02446058765053749, + 0.014150184579193592, + 0.016185935586690903, + 0.020833930000662804, + 0.021807517856359482, + 0.011560821905732155, + 0.015938978642225266, + 0.015651579946279526, + 0.019692936912178993, + 0.00848004687577486, + 0.013325300998985767, + 0.009285206906497478, + 0.01692325249314308, + 0.008999736979603767, + 0.037888962775468826, + 0.01972358115017414, + 0.01816953904926777, + 0.01716487482190132, + 0.010209331288933754, + 0.01675105281174183, + 0.015032628551125526, + 0.005321819800883532, + 0.007646229583770037, + 0.010066970251500607, + 0.00666345190256834, + 0.006109106354415417, + 0.011314183473587036, + 0.006466905120760202, + 0.005989445839077234, + 0.006214202847331762, + 0.008854600600898266, + 0.010188888758420944, + 0.00689249113202095, + 0.007419218774884939, + 0.006276951637119055, + 0.005951870698481798, + 0.006669999100267887, + 0.007377562578767538, + 0.014568828977644444, + 0.013354567810893059, + 0.007943635806441307, + 0.010558520443737507, + 0.019681772217154503, + 0.008827876299619675, + 0.011385748162865639, + 0.01680811122059822, + 0.02222871594130993, + 0.014410948380827904, + 0.015663808211684227, + 0.030634844675660133, + 0.02539634145796299, + 0.014972002245485783, + 0.010142340324819088, + 0.0055845859460532665, + 0.010143054649233818, + 0.005535557400435209, + 0.017587842419743538, + 0.008432508446276188, + 0.0069152251817286015, + 0.005950990132987499, + 0.0047187223099172115, + 0.011089270934462547, + 0.017377112060785294, + 0.01984362304210663, + 0.01618254743516445, + 0.017055587843060493, + 0.015170745551586151, + 0.012782502919435501, + 0.01297870371490717, + 0.004391743801534176, + 0.007920533418655396, + 0.006047545932233334, + 0.0065871430560946465, + 0.006289622280746698, + 0.008640644140541553, + 0.00498033594340086, + 0.005369303748011589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.1645391285419464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1271401196718216 + ], + "opposite_logprob_tokens": [ + -2.1271400451660156 + ], + "logprob_reward": -0.1271401196718216, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1271401196718216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11311052739620209, + 0.14369279146194458, + 0.09498231112957001, + 0.04999151825904846, + 0.03484996408224106, + 0.23023805022239685, + 0.05171492323279381, + 0.05898548662662506, + 0.04208843782544136, + 0.18034595251083374 + ], + "credit_map": [ + 0.008839583955705166, + 0.011639023199677467, + 0.01839708723127842, + 0.021104106679558754, + 0.015892211347818375, + 0.01342210453003645, + 0.014083435758948326, + 0.009732969105243683, + 0.019673699513077736, + 0.01982002891600132, + 0.023352809250354767, + 0.018865574151277542, + 0.011306935921311378, + 0.021309496834874153, + 0.012081599794328213, + 0.017282651737332344, + 0.005987608339637518, + 0.020524879917502403, + 0.015251442790031433, + 0.013650576584041119, + 0.010483762249350548, + 0.011004310101270676, + 0.009886465035378933, + 0.008193264715373516, + 0.0037274330388754606, + 0.013518884778022766, + 0.008268984965980053, + 0.008541857823729515, + 0.003274044720456004, + 0.004378886893391609, + 0.004571057856082916, + 0.003710370510816574, + 0.0023728159721940756, + 0.004892683122307062, + 0.00545887928456068, + 0.0033127327915281057, + 0.005364916753023863, + 0.0037011082749813795, + 0.0049124532379209995, + 0.004834373015910387, + 0.016166359186172485, + 0.034887053072452545, + 0.028650516644120216, + 0.018805895000696182, + 0.02593747153878212, + 0.055792294442653656, + 0.020637577399611473, + 0.029360882937908173, + 0.005098208785057068, + 0.008216594345867634, + 0.00657507311552763, + 0.0041015720926225185, + 0.009830324910581112, + 0.007960350252687931, + 0.004837857559323311, + 0.0050949398428201675, + 0.0037679884117096663, + 0.008990252390503883, + 0.007149075623601675, + 0.009116606786847115, + 0.010496334172785282, + 0.007081932388246059, + 0.0053748637437820435, + 0.0070084333419799805, + 0.00329232239164412, + 0.006300169508904219, + 0.00525720976293087, + 0.006125628482550383, + 0.006761570461094379, + 0.006411340087652206, + 0.0036763364914804697, + 0.004263860173523426, + 0.01792910508811474, + 0.033545490354299545, + 0.01908830925822258, + 0.021809594705700874, + 0.02288387157022953, + 0.028061144053936005, + 0.021263904869556427, + 0.01576453633606434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.1271401196718216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706254422664642 + ], + "opposite_logprob_tokens": [ + -2.1270625591278076 + ], + "logprob_reward": -0.12706254422664642, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706254422664642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13764624297618866, + 0.05248502641916275, + 0.15468208491802216, + 0.08064024150371552, + 0.11653502285480499, + 0.1997351348400116, + 0.044933516532182693, + 0.131609708070755, + 0.047070711851119995, + 0.03466225415468216 + ], + "credit_map": [ + 0.013015645556151867, + 0.019708018749952316, + 0.0241672582924366, + 0.021928898990154266, + 0.012533912435173988, + 0.008921677246689796, + 0.015353580936789513, + 0.02201727218925953, + 0.003582115052267909, + 0.006940877065062523, + 0.008444154635071754, + 0.0069093299098312855, + 0.004777488764375448, + 0.0066199833527207375, + 0.006275416351854801, + 0.008935662917792797, + 0.009396787732839584, + 0.038290444761514664, + 0.023798782378435135, + 0.029899142682552338, + 0.02246343530714512, + 0.006812770385295153, + 0.014639222994446754, + 0.00938151590526104, + 0.006286573596298695, + 0.010333619080483913, + 0.013590618968009949, + 0.008861059322953224, + 0.012497209012508392, + 0.014694462530314922, + 0.007358370814472437, + 0.007018335163593292, + 0.009615142829716206, + 0.02189268358051777, + 0.019919805228710175, + 0.013017826713621616, + 0.009632918052375317, + 0.009626497514545918, + 0.017763929441571236, + 0.015066219493746758, + 0.014578469097614288, + 0.029021674767136574, + 0.0252978652715683, + 0.013888947665691376, + 0.02712203748524189, + 0.044069379568099976, + 0.019163284450769424, + 0.026593493297696114, + 0.004070209339261055, + 0.006843313574790955, + 0.004746887367218733, + 0.003845799481496215, + 0.008580685593187809, + 0.006069155875593424, + 0.006658919155597687, + 0.004118549171835184, + 0.009257773868739605, + 0.01956578902900219, + 0.008219359442591667, + 0.034490957856178284, + 0.020387904718518257, + 0.016936661675572395, + 0.012111932039260864, + 0.010639329440891743, + 0.0036043827421963215, + 0.00642029196023941, + 0.005334773100912571, + 0.009073453024029732, + 0.007161219138652086, + 0.005525465123355389, + 0.00575646385550499, + 0.00419466570019722, + 0.0027529119979590178, + 0.0068368082866072655, + 0.0037804460152983665, + 0.004976687021553516, + 0.004196454770863056, + 0.005643852986395359, + 0.0026205014437437057, + 0.00385459465906024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.12706254422664642 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067492141388356686 + ], + "opposite_logprob_tokens": [ + -5.006749153137207 + ], + "logprob_reward": -0.0067492141388356686, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067492141388356686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08024655282497406, + 0.08334574848413467, + 0.12909770011901855, + 0.09518375247716904, + 0.07339929789304733, + 0.0654987171292305, + 0.33217811584472656, + 0.05649632588028908, + 0.026755856350064278, + 0.05779797211289406 + ], + "credit_map": [ + 0.006204575765877962, + 0.008284823037683964, + 0.010780679993331432, + 0.01394588965922594, + 0.008957216516137123, + 0.00525195337831974, + 0.01628817431628704, + 0.010533231310546398, + 0.008708184584975243, + 0.010880882851779461, + 0.01546456478536129, + 0.008797249756753445, + 0.006580433342605829, + 0.011626984924077988, + 0.007047238294035196, + 0.014240197837352753, + 0.008035462349653244, + 0.017986146733164787, + 0.015395293943583965, + 0.01914418302476406, + 0.01654849573969841, + 0.018257783725857735, + 0.018441343680024147, + 0.015288980677723885, + 0.005404809024184942, + 0.02606818825006485, + 0.029043452814221382, + 0.00818239152431488, + 0.005482745356857777, + 0.008577787317335606, + 0.00637235539034009, + 0.006052018143236637, + 0.006569317076355219, + 0.010044834576547146, + 0.011046206578612328, + 0.008579272776842117, + 0.007653022184967995, + 0.005720018874853849, + 0.012810684740543365, + 0.01097593829035759, + 0.0060327560640871525, + 0.010817251168191433, + 0.009527210146188736, + 0.005736405495554209, + 0.008641799911856651, + 0.010797755792737007, + 0.006380284205079079, + 0.00756524782627821, + 0.030378034338355064, + 0.04535192623734474, + 0.03433908149600029, + 0.03581719473004341, + 0.04525583237409592, + 0.044335756450891495, + 0.04977422207593918, + 0.046926070004701614, + 0.004344250541180372, + 0.00873816292732954, + 0.004126086365431547, + 0.018374726176261902, + 0.006807988043874502, + 0.0035070173908025026, + 0.006535228807479143, + 0.004062858875840902, + 0.0024148975498974323, + 0.0039563910104334354, + 0.0031166670378297567, + 0.0045336768962442875, + 0.00313877547159791, + 0.00388183631002903, + 0.002519112778827548, + 0.003194497199729085, + 0.005807225592434406, + 0.008824766613543034, + 0.006452467292547226, + 0.008062511682510376, + 0.008394679985940456, + 0.008196823298931122, + 0.0072861649096012115, + 0.00477333040907979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.0067492141388356686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798436909914017 + ], + "opposite_logprob_tokens": [ + -3.5297985076904297 + ], + "logprob_reward": -0.029798436909914017, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798436909914017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1596449911594391, + 0.08657243847846985, + 0.0850735530257225, + 0.06504864245653152, + 0.11731308698654175, + 0.1366311013698578, + 0.05043594911694527, + 0.08082462847232819, + 0.039076127111911774, + 0.17937949299812317 + ], + "credit_map": [ + 0.009749800898134708, + 0.01701446808874607, + 0.02348466031253338, + 0.034244805574417114, + 0.01676034741103649, + 0.013088684529066086, + 0.019743207842111588, + 0.025559017434716225, + 0.012511400505900383, + 0.010962581261992455, + 0.0160871259868145, + 0.009409432299435139, + 0.006624914705753326, + 0.01152170728892088, + 0.007641545031219721, + 0.011813735589385033, + 0.006550298072397709, + 0.02400180697441101, + 0.01238935999572277, + 0.009462746791541576, + 0.010789268650114536, + 0.00712303351610899, + 0.007967530749738216, + 0.006789508741348982, + 0.003944584634155035, + 0.016440413892269135, + 0.01834280602633953, + 0.005028105806559324, + 0.006028630770742893, + 0.004881823901087046, + 0.0044085620902478695, + 0.005973713472485542, + 0.0091535784304142, + 0.02098623663187027, + 0.018520496785640717, + 0.011669447645545006, + 0.009906573221087456, + 0.0064189620316028595, + 0.019901162013411522, + 0.020756635814905167, + 0.014863656833767891, + 0.020570408552885056, + 0.01635708473622799, + 0.01290391106158495, + 0.025550711899995804, + 0.018555600196123123, + 0.011681921780109406, + 0.016147809103131294, + 0.004472550470381975, + 0.0074225920252501965, + 0.007755320519208908, + 0.007723466958850622, + 0.0041806260123848915, + 0.0067998869344592094, + 0.006677308585494757, + 0.005404196679592133, + 0.006684294436126947, + 0.013097062706947327, + 0.007912784814834595, + 0.018101688474416733, + 0.012598603963851929, + 0.008018027991056442, + 0.006500508636236191, + 0.007911665365099907, + 0.0030018838588148355, + 0.00558112096041441, + 0.004932544194161892, + 0.008513150736689568, + 0.005001785699278116, + 0.005089208483695984, + 0.003216987242922187, + 0.0037394447717815638, + 0.015857752412557602, + 0.02781718038022518, + 0.024774648249149323, + 0.028175609186291695, + 0.018811672925949097, + 0.03182515501976013, + 0.017362745478749275, + 0.014754725620150566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.029798436909914017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a figure as they walk through a dense, ancient forest at a steady, cautious pace in the depths of night. The camera moves smoothly with them, capturing their journey along a narrow, winding dirt path covered with damp, fallen leaves. The figure is completely enveloped in a heavy, dark, floor-length, hooded cloak made of a rough wool material, its folds swaying gently as they move, their face hidden in deep shadow. In their right hand, they hold a rustic black iron lantern, swinging rhythmically, a soft flame visible through the glass window. This lone light source casts a dynamic, warm, golden-orange glow, illuminating the nearby ground and capturing the textures of gnarled tree roots and damp leaves, while the rest of the forest is plunged into near-total darkness. A thick, ethereal mist covers the forest floor, swirling around the figure's ankles, catching the lantern's light and creating a soft, hazy atmosphere. The swaying motion of the lanterns causes light, long dancing shadows to wander across the mossy trunks of the surrounding trees, creating a mysterious, serene, and atmospheric scene of a solitary passage.", + "seed": 15609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_3/index_3775_seed_15609_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14398094564676284, + "vlm_logprob": -0.14398094564676284, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.14398094564676284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43361103534698486 + ], + "opposite_logprob_tokens": [ + -1.5586110353469849 + ], + "logprob_reward": -0.43361103534698486, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.43361103534698486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11750964820384979, + 0.07357867062091827, + 0.14657166600227356, + 0.09452158212661743, + 0.10743337124586105, + 0.1350349336862564, + 0.03943903371691704, + 0.055038128048181534, + 0.06105240434408188, + 0.16982053220272064 + ], + "credit_map": [ + 0.019316213205456734, + 0.015322857536375523, + 0.017176970839500427, + 0.016809282824397087, + 0.012974992394447327, + 0.008409460075199604, + 0.01581185683608055, + 0.01168802473694086, + 0.0054774293676018715, + 0.012968061491847038, + 0.011207394301891327, + 0.011296525597572327, + 0.008872265927493572, + 0.006309422664344311, + 0.005710802972316742, + 0.01173677109181881, + 0.014724401757121086, + 0.01246648095548153, + 0.03657380864024162, + 0.060089752078056335, + 0.011023996397852898, + 0.0029741832986474037, + 0.005066401325166225, + 0.003652637591585517, + 0.006343740038573742, + 0.015559077262878418, + 0.014426934532821178, + 0.011387932114303112, + 0.02207353338599205, + 0.005913243163377047, + 0.014494230970740318, + 0.0043228804133832455, + 0.015842627733945847, + 0.011538054794073105, + 0.03453056886792183, + 0.006619690451771021, + 0.00897564459592104, + 0.01130058616399765, + 0.011850968934595585, + 0.006775228772312403, + 0.015190308913588524, + 0.023039797320961952, + 0.04546090215444565, + 0.012008517049252987, + 0.007499374449253082, + 0.011110136285424232, + 0.007140984293073416, + 0.01358492486178875, + 0.0035287896171212196, + 0.00498617859557271, + 0.005194803234189749, + 0.006723841652274132, + 0.004007744137197733, + 0.0043325708247721195, + 0.004503085743635893, + 0.006162021309137344, + 0.00611672829836607, + 0.008269275538623333, + 0.0070938351564109325, + 0.014906026422977448, + 0.003759658895432949, + 0.00622615497559309, + 0.004006338771432638, + 0.004660110920667648, + 0.005071026738733053, + 0.007901215925812721, + 0.007092527113854885, + 0.007830619812011719, + 0.010324777103960514, + 0.007541080471128225, + 0.009790596552193165, + 0.005500557832419872, + 0.020524969324469566, + 0.017443079501390457, + 0.03628508374094963, + 0.014727829955518246, + 0.008554658852517605, + 0.019606906920671463, + 0.01086512953042984, + 0.041812870651483536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.43361103534698486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06220691651105881 + ], + "opposite_logprob_tokens": [ + -2.812206983566284 + ], + "logprob_reward": -0.06220691651105881, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06220691651105881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07965156435966492, + 0.06016556918621063, + 0.15805384516716003, + 0.03275885432958603, + 0.21065305173397064, + 0.04053867235779762, + 0.09297822415828705, + 0.10653413087129593, + 0.035481661558151245, + 0.1831844002008438 + ], + "credit_map": [ + 0.007738118059933186, + 0.012822816148400307, + 0.014652318321168423, + 0.010511929169297218, + 0.006050115916877985, + 0.006873476784676313, + 0.00887185987085104, + 0.012130928225815296, + 0.004799998365342617, + 0.008088148199021816, + 0.013538357801735401, + 0.009601835161447525, + 0.003876811359077692, + 0.007861846126616001, + 0.0030574677512049675, + 0.00934110302478075, + 0.014012914150953293, + 0.018387682735919952, + 0.04178965464234352, + 0.057566411793231964, + 0.010490518063306808, + 0.003849749220535159, + 0.006859018001705408, + 0.005097903311252594, + 0.002109689638018608, + 0.004076497163623571, + 0.003917634952813387, + 0.0038267746567726135, + 0.008537773974239826, + 0.00234467932023108, + 0.006098921410739422, + 0.0018468840280547738, + 0.03375644609332085, + 0.01971730776131153, + 0.07632017135620117, + 0.015554027631878853, + 0.012558693997561932, + 0.021548070013523102, + 0.019017469137907028, + 0.012180873192846775, + 0.004553381819278002, + 0.00476152915507555, + 0.008341377601027489, + 0.0048560090363025665, + 0.0037612314336001873, + 0.004686514846980572, + 0.003794662654399872, + 0.005783964414149523, + 0.00957188569009304, + 0.012362203560769558, + 0.018262969329953194, + 0.01226455345749855, + 0.0115855373442173, + 0.008826443925499916, + 0.010564465075731277, + 0.009540162980556488, + 0.01128376740962267, + 0.015589808113873005, + 0.01288503035902977, + 0.029618028551340103, + 0.008953429758548737, + 0.010332245379686356, + 0.009575813077390194, + 0.008296007290482521, + 0.003714914433658123, + 0.0028109392151236534, + 0.007997921667993069, + 0.0034018447622656822, + 0.003521977225318551, + 0.004084567539393902, + 0.0035474500618875027, + 0.006402048282325268, + 0.017733439803123474, + 0.026013439521193504, + 0.030643610283732414, + 0.023096974939107895, + 0.019446659833192825, + 0.02793119102716446, + 0.016439946368336678, + 0.02187914028763771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.06220691651105881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604040563106537 + ], + "opposite_logprob_tokens": [ + -1.910404086112976 + ], + "logprob_reward": -0.1604040563106537, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604040563106537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10781586915254593, + 0.06973952054977417, + 0.10363856703042984, + 0.1466204822063446, + 0.25876161456108093, + 0.04887673631310463, + 0.0338602215051651, + 0.06447239220142365, + 0.09493973851203918, + 0.07127486914396286 + ], + "credit_map": [ + 0.008254778571426868, + 0.02231394685804844, + 0.011931379325687885, + 0.02756238356232643, + 0.010278858244419098, + 0.007298795040696859, + 0.009625889360904694, + 0.010549824684858322, + 0.005223462823778391, + 0.008124030195176601, + 0.009842772036790848, + 0.008296040818095207, + 0.012496826238930225, + 0.00596584053710103, + 0.009783638641238213, + 0.010006902739405632, + 0.009039480239152908, + 0.010411376133561134, + 0.02859584614634514, + 0.03534381091594696, + 0.007579623255878687, + 0.003261076519265771, + 0.0053223236463963985, + 0.004085023421794176, + 0.010119413956999779, + 0.013277566060423851, + 0.012554795481264591, + 0.01312022190541029, + 0.03554787486791611, + 0.013799510896205902, + 0.03237390145659447, + 0.015827173367142677, + 0.04069951921701431, + 0.024154813960194588, + 0.09445446729660034, + 0.01728030852973461, + 0.016162728890776634, + 0.025190168991684914, + 0.023166418075561523, + 0.017653176560997963, + 0.0050145904533565044, + 0.005416942294687033, + 0.01011939812451601, + 0.005457964260131121, + 0.004557985346764326, + 0.005748222582042217, + 0.007138153072446585, + 0.005423473659902811, + 0.0030629101675003767, + 0.00473439646884799, + 0.00421475525945425, + 0.006478263530880213, + 0.0036777162458747625, + 0.003610228653997183, + 0.003905937308445573, + 0.004176008049398661, + 0.007809676229953766, + 0.00797684770077467, + 0.007262740284204483, + 0.014593459665775299, + 0.006818122696131468, + 0.00727327074855566, + 0.007446940056979656, + 0.005291323643177748, + 0.009850705042481422, + 0.006311333738267422, + 0.03760784491896629, + 0.004761714953929186, + 0.0033206946682184935, + 0.006408998277038336, + 0.0043112244457006454, + 0.022367214784026146, + 0.007762393914163113, + 0.011417139321565628, + 0.01782285049557686, + 0.005918470211327076, + 0.00475299684330821, + 0.008709156885743141, + 0.008832688443362713, + 0.006059168837964535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.1604040563106537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020613837987184525 + ], + "opposite_logprob_tokens": [ + -3.8956139087677 + ], + "logprob_reward": -0.020613837987184525, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020613837987184525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09227360039949417, + 0.09173958003520966, + 0.27103284001350403, + 0.043050963431596756, + 0.20730048418045044, + 0.06647878140211105, + 0.060061559081077576, + 0.03246280550956726, + 0.07770456373691559, + 0.05789469927549362 + ], + "credit_map": [ + 0.0071872989647090435, + 0.013751896098256111, + 0.01251186616718769, + 0.019215472042560577, + 0.008308776654303074, + 0.00912855751812458, + 0.008843093179166317, + 0.013326641172170639, + 0.007184581831097603, + 0.013417175970971584, + 0.017594296485185623, + 0.015279733575880527, + 0.00899328850209713, + 0.00929492898285389, + 0.006208889186382294, + 0.0137666966766119, + 0.027071893215179443, + 0.02813512459397316, + 0.06611069291830063, + 0.1044052392244339, + 0.021096158772706985, + 0.0065397960133850574, + 0.0098800677806139, + 0.007793889380991459, + 0.0033123025204986334, + 0.004959093406796455, + 0.005832736846059561, + 0.004950145725160837, + 0.00938497856259346, + 0.003085124772042036, + 0.008410834707319736, + 0.003115749219432473, + 0.03372059389948845, + 0.019963204860687256, + 0.07854077965021133, + 0.014724060893058777, + 0.010033557191491127, + 0.019249150529503822, + 0.021129444241523743, + 0.009939709678292274, + 0.006921690423041582, + 0.008095826022326946, + 0.015995614230632782, + 0.007267788052558899, + 0.005119382869452238, + 0.006875995080918074, + 0.005189656745642424, + 0.011012839153409004, + 0.006119939032942057, + 0.008057840168476105, + 0.009808039292693138, + 0.007850371301174164, + 0.006473226472735405, + 0.007402752060443163, + 0.005668411031365395, + 0.008680983446538448, + 0.003297715215012431, + 0.004572581499814987, + 0.004451137501746416, + 0.0035899574868381023, + 0.0029016840271651745, + 0.0065339189022779465, + 0.0029136145021766424, + 0.00420219823718071, + 0.00874047540128231, + 0.006217960733920336, + 0.02404901571571827, + 0.004639357794076204, + 0.006101162638515234, + 0.0070253112353384495, + 0.007599745411425829, + 0.013331543654203415, + 0.005529286339879036, + 0.00738673098385334, + 0.015031707473099232, + 0.006827929522842169, + 0.004660047125071287, + 0.004841351881623268, + 0.005252777133136988, + 0.008364874869585037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.020613837987184525 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a heavy, dark, floor-length, hooded cloak made of rough wool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a rustic black iron lantern in their right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a dense, ancient forest at night with a narrow, winding dirt path covered in damp, fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly with the figure as they walk at a steady, cautious pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306888207793236 + ], + "opposite_logprob_tokens": [ + -3.1680688858032227 + ], + "logprob_reward": -0.04306888207793236, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306888207793236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05813382565975189, + 0.03608304634690285, + 0.16986724734306335, + 0.038278527557849884, + 0.1615539789199829, + 0.238369882106781, + 0.026978930458426476, + 0.016202745959162712, + 0.05374125391244888, + 0.20079058408737183 + ], + "credit_map": [ + 0.0061986977234482765, + 0.011242658831179142, + 0.008136370219290257, + 0.009776725433766842, + 0.005379911977797747, + 0.003636726411059499, + 0.0065528713166713715, + 0.007209853734821081, + 0.002709450898692012, + 0.0046332525089383125, + 0.00531061040237546, + 0.005094451829791069, + 0.004509939346462488, + 0.0042050667107105255, + 0.0030683432705700397, + 0.0065519255585968494, + 0.016359223052859306, + 0.014940681867301464, + 0.044532377272844315, + 0.0694170892238617, + 0.011394449509680271, + 0.0033344386611133814, + 0.00556945288553834, + 0.00431952066719532, + 0.0030110273510217667, + 0.006039274390786886, + 0.00491672707721591, + 0.0038460392970591784, + 0.006672129034996033, + 0.004455848131328821, + 0.004738454706966877, + 0.004599018953740597, + 0.02525956556200981, + 0.015852224081754684, + 0.05745156854391098, + 0.01084137987345457, + 0.0100613534450531, + 0.01511374395340681, + 0.015660645440220833, + 0.011313479393720627, + 0.024170516058802605, + 0.028656508773565292, + 0.062126968055963516, + 0.028799191117286682, + 0.019777467474341393, + 0.027714930474758148, + 0.021978961303830147, + 0.025145312771201134, + 0.002730695065110922, + 0.005003390833735466, + 0.004937680903822184, + 0.0033257994800806046, + 0.0029033946339040995, + 0.0028221493121236563, + 0.002408302156254649, + 0.002847513882443309, + 0.002205678028985858, + 0.0027191853150725365, + 0.0022781784646213055, + 0.0030584807973355055, + 0.001411684905178845, + 0.0016358791617676616, + 0.001389367156662047, + 0.0015042895684018731, + 0.005831621121615171, + 0.0034929662942886353, + 0.015379373915493488, + 0.0059652179479599, + 0.004843657370656729, + 0.0059609487652778625, + 0.004602921660989523, + 0.0076645417138934135, + 0.021141253411769867, + 0.01781846024096012, + 0.05949743837118149, + 0.01346833910793066, + 0.005839618854224682, + 0.01986418105661869, + 0.006731096189469099, + 0.0564301572740078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.04306888207793236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_4/index_tensor([1595, 1595])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_4/index_tensor([1595, 1595])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7281afd174121e78c6c7cbb6675ccf7e27669fd5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_4/index_tensor([1595, 1595])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle wide shot, as if captured from a smoothly tracking drone, follows a sleek, modern white yacht as it glides smoothly across vast, calm waters. The yacht's polished fiberglass hull gleams in the dusk light, effortlessly cutting through the deep blue surface, churning out crisp, white, V-shaped foam that gently dissipates in its wake. The sky above is a breathtaking sunset canvas, soft, feathery clouds painted in vibrant strokes of fiery orange and deep burgundy, blending into the twilight-blue horizon. The warm golden light of the setting sun bathes the entire scene, casting a warm glow on the yacht and gleaming on its metallic fixtures. In the distance, the tiny silhouettes of several other vessels dot the water, while the hazy, deep green coastline and dense tree line are faintly visible in the fading light, creating a profound sense of scale and a serene, majestic tranquility.", + "seed": 7731, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_4/index_1595_seed_7731_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012965545803308488, + "vlm_logprob": -0.012965545803308488, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.012965545803308488, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007736599538475275 + ], + "opposite_logprob_tokens": [ + -6.007736682891846 + ], + "logprob_reward": -0.007736599538475275, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007736599538475275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1534562110900879, + 0.2593314051628113, + 0.10793190449476242, + 0.05344094708561897, + 0.05479345843195915, + 0.07125967741012573, + 0.13048821687698364, + 0.0392230860888958, + 0.06746731698513031, + 0.06260783225297928 + ], + "credit_map": [ + 0.02245916798710823, + 0.01983778364956379, + 0.011713001877069473, + 0.01647716760635376, + 0.020283978432416916, + 0.027984479442238808, + 0.01395539939403534, + 0.02074521966278553, + 0.04146208614110947, + 0.03285372629761696, + 0.020451728254556656, + 0.03835054859519005, + 0.029234660789370537, + 0.04461981728672981, + 0.024334905669093132, + 0.02802390418946743, + 0.014665260910987854, + 0.01630663312971592, + 0.008848275057971478, + 0.02007214166224003, + 0.013485584408044815, + 0.009742281399667263, + 0.010525476187467575, + 0.014286239631474018, + 0.009579794481396675, + 0.007807103451341391, + 0.005319949705153704, + 0.008138252422213554, + 0.0056657022796571255, + 0.0063238851726055145, + 0.004805928096175194, + 0.005800324957817793, + 0.008783318102359772, + 0.011614536866545677, + 0.006225327029824257, + 0.008867618627846241, + 0.006425171159207821, + 0.004782223608344793, + 0.0026750750839710236, + 0.005420186556875706, + 0.012200942263007164, + 0.012074539437890053, + 0.006996322888880968, + 0.013105146586894989, + 0.008257306180894375, + 0.008279099129140377, + 0.004557924345135689, + 0.005788388662040234, + 0.026539020240306854, + 0.024753503501415253, + 0.01237886305898428, + 0.02862434647977352, + 0.012030608020722866, + 0.01184298750013113, + 0.008254741318523884, + 0.006064132787287235, + 0.006719558499753475, + 0.008833905681967735, + 0.004968303721398115, + 0.006045474670827389, + 0.004455546382814646, + 0.0032786086667329073, + 0.002101295627653599, + 0.002820390509441495, + 0.014634317718446255, + 0.01820339821279049, + 0.006809563841670752, + 0.008931568823754787, + 0.006414435803890228, + 0.004547592718154192, + 0.003820062382146716, + 0.004106367006897926, + 0.014047647826373577, + 0.0156667772680521, + 0.009478540159761906, + 0.008335741236805916, + 0.004657397512346506, + 0.0030523946043103933, + 0.003779203863814473, + 0.0035901227965950966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.007736599538475275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747438106685877 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747438106685877, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747438106685877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1971929371356964, + 0.1200401708483696, + 0.05634867027401924, + 0.1489216536283493, + 0.1906660944223404, + 0.0684085264801979, + 0.05969022959470749, + 0.06537263840436935, + 0.026644237339496613, + 0.06671489030122757 + ], + "credit_map": [ + 0.02592703141272068, + 0.02618050016462803, + 0.015754003077745438, + 0.025597210973501205, + 0.023953983560204506, + 0.030207069590687752, + 0.016499552875757217, + 0.03307357802987099, + 0.023732636123895645, + 0.020631054416298866, + 0.009598161093890667, + 0.012795117683708668, + 0.01475321315228939, + 0.019833309575915337, + 0.008084369823336601, + 0.010612310841679573, + 0.008780272677540779, + 0.00975273922085762, + 0.004505934659391642, + 0.008786761201918125, + 0.006507298909127712, + 0.005978506989777088, + 0.004194401670247316, + 0.007842754945158958, + 0.023595716804265976, + 0.03232111409306526, + 0.018992053344845772, + 0.025651514530181885, + 0.017625859007239342, + 0.00926284957677126, + 0.010453671216964722, + 0.01101887971162796, + 0.036659590899944305, + 0.036039799451828, + 0.012182516977190971, + 0.025280239060521126, + 0.021807502955198288, + 0.029805725440382957, + 0.01294737495481968, + 0.015943346545100212, + 0.011106816120445728, + 0.012955674901604652, + 0.011413044296205044, + 0.01097473967820406, + 0.0067836809903383255, + 0.008130631409585476, + 0.0031780521385371685, + 0.0038658876437693834, + 0.011212986893951893, + 0.009032288566231728, + 0.0058641014620661736, + 0.011548468843102455, + 0.00666913902387023, + 0.0063368105329573154, + 0.00458192965015769, + 0.004444505088031292, + 0.01097020786255598, + 0.009323323145508766, + 0.006672023329883814, + 0.012830233201384544, + 0.006664070766419172, + 0.009069005027413368, + 0.004654000047594309, + 0.005189776886254549, + 0.005372458137571812, + 0.006092166993767023, + 0.002700482727959752, + 0.0035003081429749727, + 0.0027424683794379234, + 0.0030032231006771326, + 0.0015320252859964967, + 0.0017011049203574657, + 0.014345357194542885, + 0.016051163896918297, + 0.008234330452978611, + 0.008439477533102036, + 0.0063490536995232105, + 0.0050365133211016655, + 0.0036233761347830296, + 0.004635616671293974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.006747438106685877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809947893023491 + ], + "opposite_logprob_tokens": [ + -3.288099527359009 + ], + "logprob_reward": -0.03809947893023491, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809947893023491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2004450261592865, + 0.1558796912431717, + 0.0628272220492363, + 0.09123263508081436, + 0.21278023719787598, + 0.07350782305002213, + 0.06947705149650574, + 0.05729151889681816, + 0.03254270553588867, + 0.04401613771915436 + ], + "credit_map": [ + 0.03381020948290825, + 0.03893674537539482, + 0.015017826110124588, + 0.022318216040730476, + 0.029250187799334526, + 0.02405540831387043, + 0.014616711065173149, + 0.022439690306782722, + 0.03530062362551689, + 0.02735060639679432, + 0.012681851163506508, + 0.013400368392467499, + 0.018704388290643692, + 0.024055803194642067, + 0.01076565869152546, + 0.013620375655591488, + 0.008260896429419518, + 0.008708558045327663, + 0.004907516296952963, + 0.008585771545767784, + 0.011182318441569805, + 0.00763020059093833, + 0.00702692661434412, + 0.00652503315359354, + 0.018335647881031036, + 0.023215562105178833, + 0.015256768092513084, + 0.00952692236751318, + 0.010017042979598045, + 0.006302969064563513, + 0.003911671694368124, + 0.004666040651500225, + 0.03876600041985512, + 0.034837041050195694, + 0.01725345104932785, + 0.029535751789808273, + 0.024739621207118034, + 0.03408779948949814, + 0.015251485630869865, + 0.01830907352268696, + 0.012618898414075375, + 0.012903135269880295, + 0.011888429522514343, + 0.013604020699858665, + 0.007469181902706623, + 0.004765279125422239, + 0.0037963977083563805, + 0.006462469231337309, + 0.012573635205626488, + 0.010722078382968903, + 0.007146009244024754, + 0.014584868215024471, + 0.006913098506629467, + 0.006869529839605093, + 0.0053478446789085865, + 0.005319981370121241, + 0.00921591091901064, + 0.009128755889832973, + 0.008120243437588215, + 0.00914845336228609, + 0.008036657236516476, + 0.006181406322866678, + 0.0037756990641355515, + 0.003684387542307377, + 0.007659685332328081, + 0.009424380958080292, + 0.002784739714115858, + 0.003293560352176428, + 0.0031161243095993996, + 0.0017997360555455089, + 0.0025315440725535154, + 0.0019329333445057273, + 0.008765810169279575, + 0.009730310179293156, + 0.005159918684512377, + 0.005672975443303585, + 0.005606637801975012, + 0.0033802699763327837, + 0.0028678623493760824, + 0.002832346595823765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.03809947893023491 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036149173974990845 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036149173974990845, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036149173974990845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1960482895374298, + 0.2208501398563385, + 0.057883370667696, + 0.032409146428108215, + 0.17016321420669556, + 0.0999072715640068, + 0.11170019954442978, + 0.03798066824674606, + 0.02123684249818325, + 0.05182087421417236 + ], + "credit_map": [ + 0.04067036136984825, + 0.03075326234102249, + 0.01942957192659378, + 0.0172254741191864, + 0.02282886579632759, + 0.03296254575252533, + 0.011375664733350277, + 0.02080254629254341, + 0.03293384984135628, + 0.02938862517476082, + 0.017585627734661102, + 0.0360008142888546, + 0.024583542719483376, + 0.03606746345758438, + 0.021462325006723404, + 0.02282789535820484, + 0.010390721261501312, + 0.01030349638313055, + 0.004861139692366123, + 0.008086982183158398, + 0.00586657552048564, + 0.008119250647723675, + 0.003936369903385639, + 0.006318835541605949, + 0.00572578702121973, + 0.00579372001811862, + 0.0034999812487512827, + 0.006111532915383577, + 0.0032029214780777693, + 0.0025877582374960184, + 0.0021404630970209837, + 0.003346978919580579, + 0.029324840754270554, + 0.03356160223484039, + 0.012004212476313114, + 0.021852416917681694, + 0.01991584710776806, + 0.02830781787633896, + 0.011724247597157955, + 0.013472223654389381, + 0.02097499556839466, + 0.025325288996100426, + 0.006450911983847618, + 0.011804476380348206, + 0.011364723555743694, + 0.0130731500685215, + 0.005247755907475948, + 0.005665963049978018, + 0.02240198478102684, + 0.019472291693091393, + 0.010875712148845196, + 0.021024592220783234, + 0.011665609665215015, + 0.011535665020346642, + 0.007642688229680061, + 0.007081658113747835, + 0.00969663355499506, + 0.01116428803652525, + 0.0025365501642227173, + 0.003824889659881592, + 0.0041193291544914246, + 0.003474606666713953, + 0.0014994912780821323, + 0.001664880313910544, + 0.0033786362037062645, + 0.003346499288454652, + 0.0032554965000599623, + 0.003983425907790661, + 0.0018946971977129579, + 0.002755593042820692, + 0.0013275352539494634, + 0.0012949591036885977, + 0.009462128393352032, + 0.007885180413722992, + 0.006134242285043001, + 0.008003830909729004, + 0.0074723511934280396, + 0.004382656421512365, + 0.004557414911687374, + 0.003923070151358843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0036149173974990845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00862929504364729 + ], + "opposite_logprob_tokens": [ + -4.758629322052002 + ], + "logprob_reward": -0.00862929504364729, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00862929504364729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09915914386510849, + 0.12090446054935455, + 0.07495027035474777, + 0.16813945770263672, + 0.042896341532468796, + 0.09612152725458145, + 0.07169295847415924, + 0.1072784960269928, + 0.06475777179002762, + 0.15409959852695465 + ], + "credit_map": [ + 0.018512792885303497, + 0.013499269261956215, + 0.011556979268789291, + 0.01328520942479372, + 0.01102065946906805, + 0.0119908032938838, + 0.00425806874409318, + 0.015035363845527172, + 0.020587293431162834, + 0.015907658264040947, + 0.009474311023950577, + 0.01753382384777069, + 0.01423045713454485, + 0.02041742019355297, + 0.011198648251593113, + 0.011554842814803123, + 0.012827632948756218, + 0.01130970660597086, + 0.005887159146368504, + 0.011033272370696068, + 0.009026920422911644, + 0.00964964646846056, + 0.006750309839844704, + 0.008465621620416641, + 0.02715480886399746, + 0.030918940901756287, + 0.016070794314146042, + 0.029092196375131607, + 0.01694939285516739, + 0.020282549783587456, + 0.01283364649862051, + 0.014837120659649372, + 0.007364662364125252, + 0.008254125714302063, + 0.005351749714463949, + 0.006570427678525448, + 0.0050248983316123486, + 0.004229811020195484, + 0.00270490231923759, + 0.0033957629930227995, + 0.01980775222182274, + 0.01923150010406971, + 0.007503303699195385, + 0.016681427136063576, + 0.008830221369862556, + 0.012343068607151508, + 0.005332895554602146, + 0.006391366012394428, + 0.012440002523362637, + 0.00975770503282547, + 0.007462203036993742, + 0.015102395787835121, + 0.007165052462369204, + 0.0076310220174491405, + 0.005864440929144621, + 0.006270132958889008, + 0.015867624431848526, + 0.013770549558103085, + 0.01365088950842619, + 0.025346489623188972, + 0.011326028034090996, + 0.011123377829790115, + 0.00874421838670969, + 0.007449322380125523, + 0.009695097804069519, + 0.009606879204511642, + 0.010515944100916386, + 0.012561982497572899, + 0.006911117117851973, + 0.006617327220737934, + 0.005141358822584152, + 0.003708071308210492, + 0.026957107707858086, + 0.021930575370788574, + 0.022609161213040352, + 0.02176269330084324, + 0.020984970033168793, + 0.013487580232322216, + 0.01387687586247921, + 0.012490646913647652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.00862929504364729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle wide shot, as if captured from a smoothly tracking drone, follows a sleek, modern white yacht as it glides smoothly across vast, calm waters. The yacht's polished fiberglass hull gleams in the dusk light, effortlessly cutting through the deep blue surface, churning out crisp, white, V-shaped foam that gently dissipates in its wake. The sky above is a breathtaking sunset canvas, soft, feathery clouds painted in vibrant strokes of fiery orange and deep burgundy, blending into the twilight-blue horizon. The warm golden light of the setting sun bathes the entire scene, casting a warm glow on the yacht and gleaming on its metallic fixtures. In the distance, the tiny silhouettes of several other vessels dot the water, while the hazy, deep green coastline and dense tree line are faintly visible in the fading light, creating a profound sense of scale and a serene, majestic tranquility.", + "seed": 7732, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_4/index_1595_seed_7732_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.29008213672786953, + "vlm_logprob": -0.29008213672786953, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.29008213672786953, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.054538607597351 + ], + "opposite_logprob_tokens": [ + -0.4295385777950287 + ], + "logprob_reward": -1.054538607597351, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.054538607597351, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.592592716217041, + 0.08040380477905273, + 0.008757106028497219, + 0.014449178241193295, + 0.0633290559053421, + 0.03926742821931839, + 0.03225315734744072, + 0.08259778469800949, + 0.03298138827085495, + 0.05336839705705643 + ], + "credit_map": [ + 0.21532857418060303, + 0.04447057843208313, + 0.1840807944536209, + 0.050889644771814346, + 0.022923434153199196, + 0.04035140573978424, + 0.018498744815587997, + 0.016049526631832123, + 0.009588044136762619, + 0.02974032238125801, + 0.014329501427710056, + 0.00651994813233614, + 0.008725040592253208, + 0.00559453247115016, + 0.002832951722666621, + 0.003073463449254632, + 0.0014368980191648006, + 0.0011578573612496257, + 0.0010868983808904886, + 0.0008875323110260069, + 0.0012375329388305545, + 0.0007310174405574799, + 0.001185133820399642, + 0.0010342361638322473, + 0.002177407266572118, + 0.002621616469696164, + 0.0017250091768801212, + 0.0019320142455399036, + 0.001550982124172151, + 0.0014606824843212962, + 0.0013407953083515167, + 0.001640670350752771, + 0.01274875272065401, + 0.016372209414839745, + 0.004782103933393955, + 0.005154639482498169, + 0.009740889072418213, + 0.006160542368888855, + 0.0036164557095617056, + 0.004753462504595518, + 0.006300865672528744, + 0.008731937035918236, + 0.004706942941993475, + 0.004110884387046099, + 0.0061693075112998486, + 0.003230494447052479, + 0.0026093630585819483, + 0.0034076343290507793, + 0.004425502382218838, + 0.007424850482493639, + 0.004197402391582727, + 0.004833603277802467, + 0.0032413231674581766, + 0.003163054585456848, + 0.0024836312513798475, + 0.002483788412064314, + 0.010402926243841648, + 0.030244166031479836, + 0.013870139606297016, + 0.006920266896486282, + 0.009450939483940601, + 0.0050957584753632545, + 0.0041857268661260605, + 0.002427853876724839, + 0.007154052145779133, + 0.007877437397837639, + 0.0038404774386435747, + 0.004391791764646769, + 0.0026215326506644487, + 0.0018688157433643937, + 0.002892012009397149, + 0.0023352692369371653, + 0.006076958030462265, + 0.010005161166191101, + 0.007422158960253, + 0.008666773326694965, + 0.005308627150952816, + 0.007207693066447973, + 0.004353396128863096, + 0.004327626433223486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -1.054538607597351 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039003431797028 + ], + "opposite_logprob_tokens": [ + -1.910390019416809 + ], + "logprob_reward": -0.16039003431797028, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039003431797028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7389371991157532, + 0.03301910683512688, + 0.011844298802316189, + 0.02677948959171772, + 0.07392046600580215, + 0.006971871014684439, + 0.010272719897329807, + 0.05963198095560074, + 0.02751990407705307, + 0.011102922260761261 + ], + "credit_map": [ + 0.2723359167575836, + 0.047888386994600296, + 0.23343369364738464, + 0.0666605681180954, + 0.030063241720199585, + 0.048673246055841446, + 0.020628882572054863, + 0.01925325207412243, + 0.00431875791400671, + 0.012025980278849602, + 0.0046511245891451836, + 0.00238266889937222, + 0.004467952996492386, + 0.0025410489179193974, + 0.0012768981978297234, + 0.0013546752743422985, + 0.0018617857713252306, + 0.0013525328831747174, + 0.0010166483698412776, + 0.0011729473480954766, + 0.0024007700849324465, + 0.0012638517655432224, + 0.001659877598285675, + 0.0011158845154568553, + 0.004805939272046089, + 0.006333285477012396, + 0.0021418666001409292, + 0.0024664176162332296, + 0.002822340000420809, + 0.0028203818947076797, + 0.0025596360210329294, + 0.002829623408615589, + 0.014281238429248333, + 0.019763220101594925, + 0.006599415093660355, + 0.0066000353544950485, + 0.009951455518603325, + 0.0058051468804478645, + 0.004416834097355604, + 0.006503117270767689, + 0.0011587670305743814, + 0.001491616596467793, + 0.0008020292734727263, + 0.0011137601686641574, + 0.0007045309175737202, + 0.0006469676736742258, + 0.0005711273988708854, + 0.0004830716352444142, + 0.001473639509640634, + 0.001585533143952489, + 0.0021744014229625463, + 0.0011738251196220517, + 0.0010681189596652985, + 0.001332920859567821, + 0.0006429080385714769, + 0.0008213728433474898, + 0.007035938557237387, + 0.02101389691233635, + 0.011101160198450089, + 0.004914029035717249, + 0.006916154641658068, + 0.004026198294013739, + 0.003020581789314747, + 0.0016040183836594224, + 0.009128655306994915, + 0.006954661104828119, + 0.003021673997864127, + 0.0021546375937759876, + 0.0019057096214964986, + 0.0016350210644304752, + 0.0015358930686488748, + 0.0011836523190140724, + 0.0015716501511633396, + 0.002036877442151308, + 0.0013715929817408323, + 0.001543185324408114, + 0.0012112045660614967, + 0.0013710075290873647, + 0.0009762674453668296, + 0.0010211371118202806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.16039003431797028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799004077911377 + ], + "opposite_logprob_tokens": [ + -1.8049004077911377 + ], + "logprob_reward": -0.1799004077911377, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799004077911377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6992690563201904, + 0.0890352725982666, + 0.0184390340000391, + 0.02607116289436817, + 0.01688813976943493, + 0.010022802278399467, + 0.018950827419757843, + 0.06019260734319687, + 0.0486517958343029, + 0.012479276396334171 + ], + "credit_map": [ + 0.25822097063064575, + 0.0484750010073185, + 0.21876750886440277, + 0.060035496950149536, + 0.027170665562152863, + 0.05049394816160202, + 0.02010491117835045, + 0.016000637784600258, + 0.010418824851512909, + 0.03421485424041748, + 0.016688989475369453, + 0.006974073126912117, + 0.00925138033926487, + 0.005509440787136555, + 0.003011116059496999, + 0.002966600703075528, + 0.00249898387119174, + 0.004181671421974897, + 0.0018291918095201254, + 0.0015769379679113626, + 0.002753807231783867, + 0.002065579639747739, + 0.0015638039913028479, + 0.00196905923075974, + 0.0038037423510104418, + 0.005814827512949705, + 0.002652138937264681, + 0.0027737414930015802, + 0.0030419810209423304, + 0.002960153855383396, + 0.0028379173018038273, + 0.002186661818996072, + 0.0021148573141545057, + 0.0021277260966598988, + 0.0023408017586916685, + 0.002213910920545459, + 0.0034514276776462793, + 0.002191069768741727, + 0.0013161912793293595, + 0.0011321563506498933, + 0.0015148795209825039, + 0.0014924411661922932, + 0.0012980675091966987, + 0.00152896239887923, + 0.001555196475237608, + 0.0009642327786423266, + 0.0008305141818709671, + 0.0008385093533433974, + 0.003129439428448677, + 0.003392207669094205, + 0.0023003474343568087, + 0.0024735871702432632, + 0.002746010897681117, + 0.0020193769596517086, + 0.0016261966666206717, + 0.0012636625906452537, + 0.007004101760685444, + 0.021648632362484932, + 0.011539014987647533, + 0.00550459697842598, + 0.006543938536196947, + 0.0034017725847661495, + 0.0028650611639022827, + 0.001685494091361761, + 0.01629413105547428, + 0.008261851035058498, + 0.009066437371075153, + 0.004800259135663509, + 0.0019230685429647565, + 0.003021948505192995, + 0.0027423405554145575, + 0.0025417644064873457, + 0.0014140779385343194, + 0.0029775435104966164, + 0.0015573229175060987, + 0.0020902168471366167, + 0.0008716986631043255, + 0.0014810897409915924, + 0.0011226512724533677, + 0.0009646760881878436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.1799004077911377 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539473362267017 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539473362267017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539473362267017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5332002639770508, + 0.07376163452863693, + 0.037385907024145126, + 0.02694067172706127, + 0.08865907788276672, + 0.029841769486665726, + 0.039363354444503784, + 0.06805219501256943, + 0.07365601509809494, + 0.029139075428247452 + ], + "credit_map": [ + 0.19614338874816895, + 0.03716756030917168, + 0.14006857573986053, + 0.0401315912604332, + 0.019385917112231255, + 0.06864694505929947, + 0.016396326944231987, + 0.015260015614330769, + 0.01132807694375515, + 0.028093567118048668, + 0.008837603963911533, + 0.006153811700642109, + 0.006939286831766367, + 0.0064336745999753475, + 0.0033678635954856873, + 0.002607746282592416, + 0.005936911329627037, + 0.004686604719609022, + 0.0038099130615592003, + 0.004269022960215807, + 0.003952604252845049, + 0.004166408441960812, + 0.005975819192826748, + 0.004588624462485313, + 0.004064430482685566, + 0.004690238274633884, + 0.0034067160449922085, + 0.002693876391276717, + 0.0029707499779760838, + 0.0026813356671482325, + 0.0024058783892542124, + 0.004027446731925011, + 0.015723368152976036, + 0.02361132949590683, + 0.003980023320764303, + 0.004876140970736742, + 0.018957968801259995, + 0.01032919343560934, + 0.00512557290494442, + 0.006055484525859356, + 0.00414899829775095, + 0.004495052620768547, + 0.005470312666147947, + 0.004084710497409105, + 0.00285345152951777, + 0.0034673945046961308, + 0.0020802065264433622, + 0.0032416428439319134, + 0.005551821086555719, + 0.009267724119126797, + 0.0055216336622834206, + 0.005338043440133333, + 0.003945164382457733, + 0.003966310061514378, + 0.0027845429722219706, + 0.0029881135560572147, + 0.011515886522829533, + 0.025503620505332947, + 0.006925051100552082, + 0.004357081837952137, + 0.008288037963211536, + 0.005099654663354158, + 0.0033792625181376934, + 0.002983597805723548, + 0.024420153349637985, + 0.01959994062781334, + 0.00796029157936573, + 0.006652445532381535, + 0.003604498691856861, + 0.003467416390776634, + 0.004295037593692541, + 0.003656234359368682, + 0.002958339173346758, + 0.0045382119715213776, + 0.003552286885678768, + 0.004666284658014774, + 0.003173677483573556, + 0.005151169374585152, + 0.0024172761477530003, + 0.0026818292681127787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.012539473362267017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304216057062149 + ], + "opposite_logprob_tokens": [ + -3.1680421829223633 + ], + "logprob_reward": -0.04304216057062149, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304216057062149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7702836990356445, + 0.07211261987686157, + 0.025617452338337898, + 0.018635593354701996, + 0.01528351753950119, + 0.014357509091496468, + 0.01760963164269924, + 0.035062436014413834, + 0.009109118953347206, + 0.02192842960357666 + ], + "credit_map": [ + 0.2869745194911957, + 0.0523194819688797, + 0.23313814401626587, + 0.06311219930648804, + 0.028823189437389374, + 0.06575251370668411, + 0.02260197326540947, + 0.01756160892546177, + 0.0097105223685503, + 0.02887844108045101, + 0.01010288204997778, + 0.005636600311845541, + 0.008101597428321838, + 0.0046165939420461655, + 0.00265164440497756, + 0.002414340153336525, + 0.0032710679806768894, + 0.0036775420885533094, + 0.0023580805864185095, + 0.0021568164229393005, + 0.004165207035839558, + 0.0033869342878460884, + 0.0026349853724241257, + 0.003966816235333681, + 0.0033873117063194513, + 0.004013173747807741, + 0.0017951574409380555, + 0.0019595741759985685, + 0.0019350199727341533, + 0.0019351361552253366, + 0.0014140033163130283, + 0.0021962150931358337, + 0.0029109190218150616, + 0.0026004575192928314, + 0.00174886139575392, + 0.0016144652618095279, + 0.0019272500649094582, + 0.0020472463220357895, + 0.0011054574279114604, + 0.001328859361819923, + 0.002328094094991684, + 0.0027991344686597586, + 0.0015062387101352215, + 0.001609312603250146, + 0.002142710145562887, + 0.0013569352449849248, + 0.0009687506244517863, + 0.0016463318606838584, + 0.00355203403159976, + 0.004236766602844, + 0.0018731328891590238, + 0.0023520931135863066, + 0.0018212798750028014, + 0.0014581304276362062, + 0.0012546837097033858, + 0.001061508897691965, + 0.006029813084751368, + 0.008554038591682911, + 0.0022151560988277197, + 0.002579456428065896, + 0.006267354357987642, + 0.004120879806578159, + 0.0027823348063975573, + 0.0025133981835097075, + 0.001570008578710258, + 0.0010821253526955843, + 0.0013131878804415464, + 0.0012640557251870632, + 0.0013050141278654337, + 0.0009072478278540075, + 0.0008823437383398414, + 0.0007851347327232361, + 0.0021827169694006443, + 0.003111734753474593, + 0.003181509207934141, + 0.004032133147120476, + 0.002412610687315464, + 0.003546658670529723, + 0.001863766578026116, + 0.0015972960973158479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.04304216057062149 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle wide shot, as if captured from a smoothly tracking drone, follows a sleek, modern white yacht as it glides smoothly across vast, calm waters. The yacht's polished fiberglass hull gleams in the dusk light, effortlessly cutting through the deep blue surface, churning out crisp, white, V-shaped foam that gently dissipates in its wake. The sky above is a breathtaking sunset canvas, soft, feathery clouds painted in vibrant strokes of fiery orange and deep burgundy, blending into the twilight-blue horizon. The warm golden light of the setting sun bathes the entire scene, casting a warm glow on the yacht and gleaming on its metallic fixtures. In the distance, the tiny silhouettes of several other vessels dot the water, while the hazy, deep green coastline and dense tree line are faintly visible in the fading light, creating a profound sense of scale and a serene, majestic tranquility.", + "seed": 7733, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_4/index_1595_seed_7733_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010003439756110311, + "vlm_logprob": -0.010003439756110311, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.010003439756110311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008070123381912708 + ], + "opposite_logprob_tokens": [ + -5.88306999206543 + ], + "logprob_reward": -0.008070123381912708, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.008070123381912708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0553203709423542, + 0.00769853824749589, + 0.010716154240071774, + 0.008737856522202492, + 0.007526374887675047, + 0.007548236753791571, + 0.03210294619202614, + 0.05508315563201904, + 0.8067181706428528, + 0.008548136800527573 + ], + "credit_map": [ + 0.004144493490457535, + 0.005254562944173813, + 0.0016714050434529781, + 0.002145984908565879, + 0.02089080400764942, + 0.016051622107625008, + 0.0020928301382809877, + 0.003068671328946948, + 0.000817976426333189, + 0.0009342298726551235, + 0.0007690981146879494, + 0.0008742414065636694, + 0.0010354744736105204, + 0.0012193670263513923, + 0.0010530863655731082, + 0.0009950643870979548, + 0.0008575898827984929, + 0.000894941040314734, + 0.0007247577304951847, + 0.0007834361749701202, + 0.0031978855840861797, + 0.0027444546576589346, + 0.0007231358322314918, + 0.0007899542688392103, + 0.0007194752688519657, + 0.0009313439950346947, + 0.0008394468459300697, + 0.0010489479172974825, + 0.0014554322697222233, + 0.0010693735675886273, + 0.0013256140518933535, + 0.0013482224894687533, + 0.0010558194480836391, + 0.0012332522310316563, + 0.0006215419853106141, + 0.0007339727599173784, + 0.0011426490964367986, + 0.0009545869543217123, + 0.0008337705512531102, + 0.0009507814538665116, + 0.0008826869307085872, + 0.0008823293610475957, + 0.0010421781335026026, + 0.0012843807926401496, + 0.0007236480014398694, + 0.0010446106316521764, + 0.0009233399177901447, + 0.0007650633342564106, + 0.005720183253288269, + 0.0037101926282048225, + 0.0017107449239119887, + 0.0017436096677556634, + 0.006957691162824631, + 0.0060988799668848515, + 0.0024738179054111242, + 0.00368782551959157, + 0.010781385935842991, + 0.00847917702049017, + 0.0037077392917126417, + 0.003465592861175537, + 0.0101359523832798, + 0.008040416054427624, + 0.005378591828048229, + 0.0050942981615662575, + 0.024929428473114967, + 0.03947979956865311, + 0.008720898069441319, + 0.0063133640214800835, + 0.36069032549858093, + 0.33635666966438293, + 0.009105872362852097, + 0.021121790632605553, + 0.001961416332051158, + 0.0012883987510576844, + 0.0006679651560261846, + 0.0006458337884396315, + 0.0014872110914438963, + 0.0009532113326713443, + 0.0007610868779011071, + 0.0007830133545212448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.008070123381912708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003625251119956374 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.003625251119956374, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003625251119956374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06479489058256149, + 0.019345276057720184, + 0.010613890364766121, + 0.008923815563321114, + 0.011189677752554417, + 0.013752286322414875, + 0.05094963312149048, + 0.038498226553201675, + 0.7377678751945496, + 0.04416448250412941 + ], + "credit_map": [ + 0.006073944270610809, + 0.0070018828846514225, + 0.002912208903580904, + 0.003989802207797766, + 0.021347248926758766, + 0.015646053478121758, + 0.0040247272700071335, + 0.0037990224082022905, + 0.0021954774856567383, + 0.0024401391856372356, + 0.0012571358820423484, + 0.0018895819084718823, + 0.0032893021125346422, + 0.0036548322532325983, + 0.0025929883122444153, + 0.002025818219408393, + 0.0021006655879318714, + 0.001755021046847105, + 0.0005470879841595888, + 0.0008330635610036552, + 0.0016398592852056026, + 0.0016070452984422445, + 0.0010152276372537017, + 0.0011159201385453343, + 0.0007056263275444508, + 0.0007570204907096922, + 0.0006661750376224518, + 0.0009604902006685734, + 0.001261471537873149, + 0.0012405840680003166, + 0.0017074663192033768, + 0.001624980359338224, + 0.0013128898572176695, + 0.0015960076125338674, + 0.0007429872639477253, + 0.0010988458525389433, + 0.0015824303263798356, + 0.0014624687610194087, + 0.0017036051722243428, + 0.0016904417425394058, + 0.001290317508392036, + 0.0015834892401471734, + 0.0020400385838001966, + 0.0024289460852742195, + 0.0012290452141314745, + 0.0015590529656037688, + 0.001945588388480246, + 0.0016758074052631855, + 0.00918192882090807, + 0.005885260179638863, + 0.0023392869625240564, + 0.0028845572378486395, + 0.012019243091344833, + 0.009215299040079117, + 0.00396494148299098, + 0.005459107458591461, + 0.007326055318117142, + 0.005509850103408098, + 0.0026861883234232664, + 0.0024458502884954214, + 0.006832379847764969, + 0.00568816065788269, + 0.004191677086055279, + 0.0038180642295628786, + 0.02359691448509693, + 0.0366150438785553, + 0.008114784955978394, + 0.005983038805425167, + 0.3302343189716339, + 0.304988831281662, + 0.009078414179384708, + 0.019156470894813538, + 0.008449832908809185, + 0.004847005009651184, + 0.003766374196857214, + 0.0035987708251923323, + 0.006983124651014805, + 0.005374399479478598, + 0.005381878465414047, + 0.005763096734881401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.003625251119956374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740096490830183 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006740096490830183, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740096490830183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02112393267452717, + 0.00877463724464178, + 0.004777366295456886, + 0.0076606073416769505, + 0.008328397758305073, + 0.009560584090650082, + 0.02845524065196514, + 0.007148559670895338, + 0.8747256398200989, + 0.029444986954331398 + ], + "credit_map": [ + 0.002244609873741865, + 0.0028878115117549896, + 0.0013181261019781232, + 0.0020632275845855474, + 0.005363221280276775, + 0.003789332462474704, + 0.001929930760525167, + 0.0015276705380529165, + 0.0010589193552732468, + 0.0009395444067195058, + 0.000695104303304106, + 0.0010897888569161296, + 0.0012871501967310905, + 0.0014041331596672535, + 0.0013640869874507189, + 0.0009359083487652242, + 0.0006403849110938609, + 0.000569106952752918, + 0.0003859674616251141, + 0.0005943812429904938, + 0.000682211306411773, + 0.0007126462878659368, + 0.0006487846258096397, + 0.0005438830121420324, + 0.0010862124618142843, + 0.001149593386799097, + 0.0005931208725087345, + 0.0009621697827242315, + 0.0010578654473647475, + 0.0009938763687387109, + 0.0010080389911308885, + 0.0008097292156890035, + 0.0008345979731529951, + 0.0010880554327741265, + 0.0009967307560145855, + 0.0010291205253452063, + 0.0010680884588509798, + 0.0012562532210722566, + 0.0010863334173336625, + 0.000969216984231025, + 0.0015551713295280933, + 0.0009946409845724702, + 0.0008788236300460994, + 0.0008959159604273736, + 0.0010386112844571471, + 0.0020820582285523415, + 0.0009115674765780568, + 0.0012037933338433504, + 0.0049935984425246716, + 0.0032388686668127775, + 0.001407101284712553, + 0.0015129250241443515, + 0.006618407089263201, + 0.005333575885742903, + 0.002167060039937496, + 0.003183701541274786, + 0.001212405040860176, + 0.0008855746709741652, + 0.0007052377914078534, + 0.0005585262551903725, + 0.0010962324449792504, + 0.0010757425334304571, + 0.0009073074907064438, + 0.0007075325702317059, + 0.02689553052186966, + 0.04299364238977432, + 0.009560960344970226, + 0.0068579427897930145, + 0.39136189222335815, + 0.3641749620437622, + 0.010124529711902142, + 0.02275610715150833, + 0.006929623428732157, + 0.00357071147300303, + 0.0019941336940973997, + 0.0019517148612067103, + 0.005448690615594387, + 0.003482671920210123, + 0.002931385999545455, + 0.0031360522843897343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.006740096490830183 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052522532641887665 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.0052522532641887665, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052522532641887665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2887904644012451, + 0.0896463543176651, + 0.08960716426372528, + 0.014554274268448353, + 0.01599317602813244, + 0.03810696676373482, + 0.056791193783283234, + 0.038772232830524445, + 0.34003448486328125, + 0.027703719213604927 + ], + "credit_map": [ + 0.021235644817352295, + 0.023823924362659454, + 0.007619272917509079, + 0.008868965320289135, + 0.11375774443149567, + 0.09057049453258514, + 0.010567931458353996, + 0.012346424162387848, + 0.01115428563207388, + 0.009642313234508038, + 0.00515304459258914, + 0.007345343008637428, + 0.024496302008628845, + 0.018643666058778763, + 0.0065167853608727455, + 0.006694610696285963, + 0.006600718945264816, + 0.005596877541393042, + 0.004350912757217884, + 0.004532103892415762, + 0.033130086958408356, + 0.02805023267865181, + 0.003145336639136076, + 0.004200885072350502, + 0.0024208782706409693, + 0.0016622318653389812, + 0.0009082118631340563, + 0.0012878401903435588, + 0.002994248177856207, + 0.002044599736109376, + 0.0016274265944957733, + 0.0016088348347693682, + 0.0031949200201779604, + 0.0018027711194008589, + 0.0012655387399718165, + 0.0014377660118043423, + 0.0023958974052220583, + 0.0033222706988453865, + 0.0010707584442570806, + 0.0015032511437311769, + 0.006573929451406002, + 0.0046855914406478405, + 0.002273391466587782, + 0.0029396002646535635, + 0.0048750098794698715, + 0.009666474536061287, + 0.0027504051104187965, + 0.004342563450336456, + 0.00973349530249834, + 0.006657831370830536, + 0.002352082170546055, + 0.003090860554948449, + 0.015391668304800987, + 0.010774311609566212, + 0.0037067984230816364, + 0.005084141623228788, + 0.007947307080030441, + 0.005913321860134602, + 0.0025652856566011906, + 0.0024870315100997686, + 0.006669586058706045, + 0.005809912923723459, + 0.004064679145812988, + 0.0033151039388030767, + 0.010662646032869816, + 0.016332099214196205, + 0.004143199417740107, + 0.0032485022675246, + 0.15105752646923065, + 0.14098656177520752, + 0.00460601132363081, + 0.008997919037938118, + 0.005753284320235252, + 0.002704615006223321, + 0.0016495658783242106, + 0.0016642851987853646, + 0.006350700277835131, + 0.003867450403049588, + 0.002826923504471779, + 0.002886892529204488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0052522532641887665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02632947452366352 + ], + "opposite_logprob_tokens": [ + -3.651329517364502 + ], + "logprob_reward": -0.02632947452366352, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02632947452366352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03229557350277901, + 0.012154364958405495, + 0.03510769084095955, + 0.007321050390601158, + 0.004762162920087576, + 0.007358631119132042, + 0.06119375675916672, + 0.02336495742201805, + 0.8055440783500671, + 0.010897696018218994 + ], + "credit_map": [ + 0.002843959257006645, + 0.003737247083336115, + 0.0018162191845476627, + 0.0024396118242293596, + 0.008798298425972462, + 0.007059062831103802, + 0.0032564657740294933, + 0.0023447114508599043, + 0.0012949861120432615, + 0.001643984578549862, + 0.0012218097690492868, + 0.0017024815315380692, + 0.0019451862899586558, + 0.0019074470037594438, + 0.0013247303431853652, + 0.0011137403780594468, + 0.0027520302683115005, + 0.002295076148584485, + 0.002352201845496893, + 0.002278604544699192, + 0.01150140818208456, + 0.009942812845110893, + 0.0016750871436670423, + 0.00231047160923481, + 0.000954630842898041, + 0.0008959746919572353, + 0.0006742309778928757, + 0.0007889042026363313, + 0.0012236986076459289, + 0.0012438867706805468, + 0.0007945961551740766, + 0.0007451286073774099, + 0.0005733386497013271, + 0.0005117809632793069, + 0.0003718094958458096, + 0.00047923109377734363, + 0.0007264452870003879, + 0.0006656191544607282, + 0.000735622423235327, + 0.0006983159692026675, + 0.0012424925807863474, + 0.0011556586250662804, + 0.0004781346651725471, + 0.0008310067350976169, + 0.0008322158246301115, + 0.0016657205997034907, + 0.0004914528108201921, + 0.0006619493360631168, + 0.010450154542922974, + 0.00689634820446372, + 0.002978047588840127, + 0.0031127987895160913, + 0.013983876444399357, + 0.011650421656668186, + 0.00504517275840044, + 0.0070769404992461205, + 0.004153889138251543, + 0.0036163171753287315, + 0.0016553635941818357, + 0.001712059834972024, + 0.0045097642578184605, + 0.003495041513815522, + 0.0018833506619557738, + 0.0023391726426780224, + 0.024617688730359077, + 0.03914676979184151, + 0.008588421158492565, + 0.006318239960819483, + 0.3603026866912842, + 0.33646857738494873, + 0.008981091901659966, + 0.021120641380548477, + 0.00121662812307477, + 0.0009766262955963612, + 0.0011984575539827347, + 0.0012004373129457235, + 0.001516935764811933, + 0.0017079656245186925, + 0.001435807324014604, + 0.0016448389505967498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.02632947452366352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle wide shot, as if captured from a smoothly tracking drone, follows a sleek, modern white yacht as it glides smoothly across vast, calm waters. The yacht's polished fiberglass hull gleams in the dusk light, effortlessly cutting through the deep blue surface, churning out crisp, white, V-shaped foam that gently dissipates in its wake. The sky above is a breathtaking sunset canvas, soft, feathery clouds painted in vibrant strokes of fiery orange and deep burgundy, blending into the twilight-blue horizon. The warm golden light of the setting sun bathes the entire scene, casting a warm glow on the yacht and gleaming on its metallic fixtures. In the distance, the tiny silhouettes of several other vessels dot the water, while the hazy, deep green coastline and dense tree line are faintly visible in the fading light, creating a profound sense of scale and a serene, majestic tranquility.", + "seed": 7734, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_4/index_1595_seed_7734_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009395640157163144, + "vlm_logprob": -0.009395640157163144, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.009395640157163144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011718484573066235 + ], + "opposite_logprob_tokens": [ + -6.261718273162842 + ], + "logprob_reward": -0.011718484573066235, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.011718484573066235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19934546947479248, + 0.0645279586315155, + 0.04204549267888069, + 0.06262796372175217, + 0.12747035920619965, + 0.1567700058221817, + 0.08277817815542221, + 0.04330429062247276, + 0.12798887491226196, + 0.09314136207103729 + ], + "credit_map": [ + 0.023865602910518646, + 0.03222170099616051, + 0.019094711169600487, + 0.025589849799871445, + 0.020799828693270683, + 0.01797698438167572, + 0.035533517599105835, + 0.024263281375169754, + 0.007481374777853489, + 0.010880360379815102, + 0.007352384738624096, + 0.008386337198317051, + 0.00808610487729311, + 0.00809764489531517, + 0.005941985175013542, + 0.008301772177219391, + 0.006873737554997206, + 0.006428317632526159, + 0.004648098722100258, + 0.005584173370152712, + 0.004495385102927685, + 0.005853651557117701, + 0.0045957318507134914, + 0.003566398983821273, + 0.009086390025913715, + 0.012788914144039154, + 0.00836727675050497, + 0.009022856131196022, + 0.005109032150357962, + 0.00636636046692729, + 0.006821314338594675, + 0.005065820179879665, + 0.019388454034924507, + 0.02531316503882408, + 0.007441152818500996, + 0.009500580839812756, + 0.01884596236050129, + 0.025492820888757706, + 0.008201022632420063, + 0.013287205249071121, + 0.02813185192644596, + 0.02724209614098072, + 0.010613847523927689, + 0.0133836530148983, + 0.02663794346153736, + 0.016385957598686218, + 0.019769348204135895, + 0.014605317264795303, + 0.01803862676024437, + 0.012185312807559967, + 0.004905507434159517, + 0.004640788305550814, + 0.019450660794973373, + 0.01112032774835825, + 0.00616695499047637, + 0.006269998382776976, + 0.004901655949652195, + 0.0074692671187222, + 0.003931660205125809, + 0.0031895367428660393, + 0.008706903085112572, + 0.006544339936226606, + 0.004620078485459089, + 0.003940850030630827, + 0.01374040450900793, + 0.014654003083705902, + 0.01487535797059536, + 0.018199386075139046, + 0.011595210991799831, + 0.024538669735193253, + 0.018028829246759415, + 0.012357010506093502, + 0.015327838249504566, + 0.015161357820034027, + 0.006766863167285919, + 0.0073784468695521355, + 0.016780104488134384, + 0.0156219732016325, + 0.007721202913671732, + 0.008383568376302719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.011718484573066235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041066608391702175 + ], + "opposite_logprob_tokens": [ + -5.504106521606445 + ], + "logprob_reward": -0.0041066608391702175, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041066608391702175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1197403222322464, + 0.062027331441640854, + 0.03368724137544632, + 0.015492105856537819, + 0.07539582997560501, + 0.29785388708114624, + 0.11503395438194275, + 0.06723353266716003, + 0.17916378378868103, + 0.03437206149101257 + ], + "credit_map": [ + 0.015721088275313377, + 0.01885882206261158, + 0.010258720256388187, + 0.01246409211307764, + 0.015240163542330265, + 0.018627824261784554, + 0.014149192720651627, + 0.0144204031676054, + 0.008640604093670845, + 0.009985137730836868, + 0.005285294726490974, + 0.00655766949057579, + 0.007682705298066139, + 0.009359639137983322, + 0.009036928415298462, + 0.005479345563799143, + 0.006384062580764294, + 0.004908823873847723, + 0.0033934898674488068, + 0.0047962358221411705, + 0.003488744143396616, + 0.0034993754234164953, + 0.004099023528397083, + 0.0031174803152680397, + 0.0023824451491236687, + 0.0024339109659194946, + 0.0010574450716376305, + 0.0016121232183650136, + 0.0025295009836554527, + 0.0023494528140872717, + 0.0019633397459983826, + 0.0011638860451057553, + 0.00822349265217781, + 0.01507563702762127, + 0.003727411152794957, + 0.00541375856846571, + 0.014861074276268482, + 0.015421616844832897, + 0.005160780623555183, + 0.00751204788684845, + 0.07071425020694733, + 0.06350584328174591, + 0.014492283575236797, + 0.016941135749220848, + 0.05785379186272621, + 0.03389620780944824, + 0.021207375451922417, + 0.01924295909702778, + 0.025467563420534134, + 0.018633736297488213, + 0.005697587039321661, + 0.008092059753835201, + 0.020646246150135994, + 0.015159151516854763, + 0.011795745231211185, + 0.009541846811771393, + 0.008096062578260899, + 0.010375208221375942, + 0.007339178118854761, + 0.007007851265370846, + 0.006237925961613655, + 0.012846793048083782, + 0.008444590494036674, + 0.006885911338031292, + 0.021185405552387238, + 0.023209847509860992, + 0.02369661070406437, + 0.024912633001804352, + 0.01784016750752926, + 0.028521180152893066, + 0.024419259279966354, + 0.015378653071820736, + 0.0028842168394476175, + 0.0025910455733537674, + 0.0014393345918506384, + 0.0026342379860579967, + 0.010619373060762882, + 0.009521806612610817, + 0.0024227085523307323, + 0.002259332686662674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0041066608391702175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950236693024635 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950236693024635, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950236693024635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16791272163391113, + 0.09337938576936722, + 0.0507584847509861, + 0.056446053087711334, + 0.35891416668891907, + 0.043013256043195724, + 0.05231281742453575, + 0.0490710623562336, + 0.10690857470035553, + 0.021283382549881935 + ], + "credit_map": [ + 0.02335406094789505, + 0.028772814199328423, + 0.010803796350955963, + 0.01677890680730343, + 0.02416212484240532, + 0.030042387545108795, + 0.018111346289515495, + 0.015887286514043808, + 0.010043426416814327, + 0.01669706217944622, + 0.00940603855997324, + 0.012069790624082088, + 0.010102590546011925, + 0.011455527506768703, + 0.016296137124300003, + 0.007308820728212595, + 0.007095377892255783, + 0.007543162442743778, + 0.006787531543523073, + 0.006022804882377386, + 0.00586117384955287, + 0.004532001446932554, + 0.00676163611933589, + 0.006154803559184074, + 0.006478765979409218, + 0.010287811979651451, + 0.00875031016767025, + 0.007796856574714184, + 0.006563824601471424, + 0.006547289434820414, + 0.0056609478779137135, + 0.004360253922641277, + 0.03840693086385727, + 0.06725567579269409, + 0.024419240653514862, + 0.03569632023572922, + 0.06368184834718704, + 0.0678320899605751, + 0.023654082790017128, + 0.037968024611473083, + 0.005437666550278664, + 0.006162036210298538, + 0.003247199347242713, + 0.0050700451247394085, + 0.005956330802291632, + 0.005680887494236231, + 0.00586376478895545, + 0.005595330614596605, + 0.0075224582105875015, + 0.008484226651489735, + 0.0035191229544579983, + 0.005647926591336727, + 0.009298634715378284, + 0.006416832562536001, + 0.006487223319709301, + 0.0049363975413143635, + 0.0064004515297710896, + 0.00753292441368103, + 0.005181435961276293, + 0.005418001674115658, + 0.005064454860985279, + 0.007782237604260445, + 0.006027352064847946, + 0.005664207506924868, + 0.011213085614144802, + 0.01079832948744297, + 0.014126462861895561, + 0.014292804524302483, + 0.009992793202400208, + 0.01765286736190319, + 0.016731731593608856, + 0.012100506573915482, + 0.003181816078722477, + 0.0028335563838481903, + 0.001971737714484334, + 0.0020843062084168196, + 0.003939170856028795, + 0.003150998381897807, + 0.0018880711868405342, + 0.0022337292321026325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.005950236693024635 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046377950347959995 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.0046377950347959995, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046377950347959995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16845634579658508, + 0.14165471494197845, + 0.053600385785102844, + 0.07381299883127213, + 0.15038129687309265, + 0.062294453382492065, + 0.1099962666630745, + 0.10035828500986099, + 0.07018609344959259, + 0.0692591667175293 + ], + "credit_map": [ + 0.021845314651727676, + 0.0266278013586998, + 0.01198731642216444, + 0.01584262028336525, + 0.027455821633338928, + 0.03527349606156349, + 0.012438097037374973, + 0.016985885798931122, + 0.01844698190689087, + 0.024458447471261024, + 0.0085179153829813, + 0.008412051014602184, + 0.028467217460274696, + 0.035364970564842224, + 0.008304072543978691, + 0.009683062322437763, + 0.007544822990894318, + 0.010641591623425484, + 0.005204958841204643, + 0.0057182009331882, + 0.00662523927167058, + 0.008420995436608791, + 0.004542713053524494, + 0.0049018654972314835, + 0.011439797468483448, + 0.011755102314054966, + 0.00677569629624486, + 0.006459204945713282, + 0.012777434661984444, + 0.013051475398242474, + 0.005480767693370581, + 0.006073526572436094, + 0.023007920011878014, + 0.02833147533237934, + 0.007360802497714758, + 0.013207857497036457, + 0.02935189940035343, + 0.024417942389845848, + 0.009262220934033394, + 0.015441170893609524, + 0.011175472289323807, + 0.012576033361256123, + 0.004453066270798445, + 0.005790538154542446, + 0.008554313331842422, + 0.006372587755322456, + 0.007537348195910454, + 0.005835092626512051, + 0.020146820694208145, + 0.01774468645453453, + 0.00568672688677907, + 0.008819527924060822, + 0.023729031905531883, + 0.013506135903298855, + 0.01170764584094286, + 0.008655698038637638, + 0.011324305087327957, + 0.017781490460038185, + 0.006273970473557711, + 0.00663745729252696, + 0.025980381295084953, + 0.01736392080783844, + 0.008280567824840546, + 0.006716193165630102, + 0.008908366784453392, + 0.015538611449301243, + 0.004054071847349405, + 0.006144328508526087, + 0.01277918554842472, + 0.01227775402367115, + 0.0058367070741951466, + 0.004647067282348871, + 0.012266157194972038, + 0.014204364269971848, + 0.003516812575981021, + 0.005423614755272865, + 0.012649107724428177, + 0.011974569410085678, + 0.003908158279955387, + 0.00531638041138649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0046377950347959995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle wide shot following a sleek, modern white yacht?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the yacht moving smoothly across vast, calm waters with a polished fiberglass hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yacht create crisp, white, V-shaped foam in its wake that dissipates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a sunset scene with soft, feathery clouds in fiery orange and deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny silhouettes of other vessels and a hazy, deep green coastline visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056502364575863 + ], + "opposite_logprob_tokens": [ + -3.8955650329589844 + ], + "logprob_reward": -0.02056502364575863, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056502364575863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1342490166425705, + 0.3285973072052002, + 0.03636384382843971, + 0.06852971762418747, + 0.08826553821563721, + 0.0670318678021431, + 0.07576830685138702, + 0.06096278131008148, + 0.07802468538284302, + 0.062206871807575226 + ], + "credit_map": [ + 0.01734745129942894, + 0.023440653458237648, + 0.009907886385917664, + 0.015180965885519981, + 0.017440570518374443, + 0.017284104600548744, + 0.0220755897462368, + 0.01157181803137064, + 0.0680442675948143, + 0.0647498220205307, + 0.019846629351377487, + 0.03543207421898842, + 0.03585829585790634, + 0.047513823956251144, + 0.03837990388274193, + 0.018772536888718605, + 0.004151147790253162, + 0.005655846558511257, + 0.004025734961032867, + 0.005099826958030462, + 0.004758983384817839, + 0.004136682953685522, + 0.005179092288017273, + 0.0033565331250429153, + 0.009875328280031681, + 0.012316341511905193, + 0.009793035686016083, + 0.010088585317134857, + 0.007093172986060381, + 0.005837153177708387, + 0.007722028996795416, + 0.0058040739968419075, + 0.013057087548077106, + 0.014980151318013668, + 0.00550637673586607, + 0.006922678090631962, + 0.0143023282289505, + 0.01796708069741726, + 0.005710051394999027, + 0.009819789789617062, + 0.01277463324368, + 0.013410456478595734, + 0.006169653031975031, + 0.007315680384635925, + 0.006443752441555262, + 0.006963460240513086, + 0.007831662893295288, + 0.006122570484876633, + 0.009140729904174805, + 0.010642250999808311, + 0.0045761605724692345, + 0.006441845558583736, + 0.01804434135556221, + 0.011197516694664955, + 0.008183382451534271, + 0.007542081642895937, + 0.006544520612806082, + 0.010403582826256752, + 0.004887809045612812, + 0.006799967493861914, + 0.005338451359421015, + 0.013619624078273773, + 0.006773750297725201, + 0.006595078390091658, + 0.010788273066282272, + 0.01713881455361843, + 0.004823054652661085, + 0.006224420852959156, + 0.011489835567772388, + 0.015220540575683117, + 0.005685470998287201, + 0.006654278375208378, + 0.008520949631929398, + 0.009772020392119884, + 0.0052978359162807465, + 0.004384422209113836, + 0.012275486253201962, + 0.01173553615808487, + 0.0038930876180529594, + 0.006327538285404444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.02056502364575863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_5/index_tensor([7207, 7207])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_5/index_tensor([7207, 7207])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1e39d7e824f08f3968de9954d991f703b5f174cb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_5/index_tensor([7207, 7207])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a low-angle perspective, the camera begins a slow, deliberate, winding tracking shot, gliding weightlessly through a narrow canyon carved by architectural giants. These massive skyscrapers rise like an ancient, silent forest, their peaks lost in a clouded sky and soaring, invisible ceilings. Soft, indirect light filters down from above, bathing their massive forms in a diffuse glow and capturing the various textures of their exteriors—the cool, smooth sheen of deep blue reflective glass, the matte, porous surface of brutalist concrete, and the intricate metal grids and steel exoskeletons. The camera drifts smoothly forward, slowly tilting across the canyon, revealing new perspectives on this dense urban jungle: skybridges winding like fallen logs between the towers, recessed windows forming intricate, shadowed patterns, and a subtle play of diffused light creating gentle gradients across the vast, steep surfaces. The overall color palette is a cool, desaturated mix of slate gray, deep blue, and soft silver, creating a profound atmosphere of quiet, contemplative awe and tranquil solitude in the heart of this monumental, sleepy metropolis.", + "seed": 19067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_5/index_7207_seed_19067_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028567178267985584, + "vlm_logprob": -0.028567178267985584, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.028567178267985584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011372385546565056 + ], + "opposite_logprob_tokens": [ + -8.136372566223145 + ], + "logprob_reward": -0.011372385546565056, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.011372385546565056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030663974583148956, + 0.026090167462825775, + 0.036645080894231796, + 0.01674245484173298, + 0.07738756388425827, + 0.04390082508325577, + 0.6564116477966309, + 0.03774872049689293, + 0.029979107901453972, + 0.04443046450614929 + ], + "credit_map": [ + 0.005817094352096319, + 0.0042173294350504875, + 0.004697746131569147, + 0.0032019400969147682, + 0.0020404434762895107, + 0.0036702349316328764, + 0.003586356993764639, + 0.0034328284673392773, + 0.002700177254155278, + 0.0029545777942985296, + 0.004530796781182289, + 0.003841208526864648, + 0.0020934036001563072, + 0.0033630384132266045, + 0.0037545054219663143, + 0.002852460602298379, + 0.004202136769890785, + 0.003107273019850254, + 0.004205228295177221, + 0.004889790900051594, + 0.0026164951268583536, + 0.004847621079534292, + 0.004851729609072208, + 0.007924805395305157, + 0.0020512142218649387, + 0.0010670361807569861, + 0.0026635010726749897, + 0.002807961544021964, + 0.0017015727935358882, + 0.0025873242411762476, + 0.0021792035549879074, + 0.0016846400685608387, + 0.004394340328872204, + 0.004415581934154034, + 0.027537371963262558, + 0.009695209562778473, + 0.006672958377748728, + 0.005390964914113283, + 0.00586112542077899, + 0.013420013710856438, + 0.011507051065564156, + 0.00551662128418684, + 0.005348546430468559, + 0.005361215211451054, + 0.004417032469063997, + 0.003706294344738126, + 0.0038022552616894245, + 0.004241806920617819, + 0.05032731220126152, + 0.03386741131544113, + 0.27037444710731506, + 0.03039039671421051, + 0.015972739085555077, + 0.042313914746046066, + 0.024785256013274193, + 0.1883801519870758, + 0.0031102686189115047, + 0.0038682292215526104, + 0.0041432068683207035, + 0.003210066817700863, + 0.004726050421595573, + 0.009815790690481663, + 0.0046916198916733265, + 0.0041834888979792595, + 0.004100515507161617, + 0.0034565026871860027, + 0.006178758572787046, + 0.0039006106089800596, + 0.0022430226672440767, + 0.0036137199494987726, + 0.0031086443923413754, + 0.0033773346804082394, + 0.006266816519200802, + 0.00544204143807292, + 0.006728457286953926, + 0.005325719714164734, + 0.0037265103310346603, + 0.002625011606141925, + 0.006898093968629837, + 0.007417818531394005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.011372385546565056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369971364736557 + ], + "opposite_logprob_tokens": [ + -3.4086997509002686 + ], + "logprob_reward": -0.03369971364736557, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369971364736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043047867715358734, + 0.0440574549138546, + 0.049750883132219315, + 0.021055376157164574, + 0.24889646470546722, + 0.031338538974523544, + 0.28811609745025635, + 0.11875852197408676, + 0.08605863153934479, + 0.06892016530036926 + ], + "credit_map": [ + 0.005729262251406908, + 0.004802975337952375, + 0.006578121799975634, + 0.005297189112752676, + 0.003648084821179509, + 0.0056275660172104836, + 0.005494697485119104, + 0.0058699725195765495, + 0.0069353156723082066, + 0.006817810237407684, + 0.005170186050236225, + 0.006372321397066116, + 0.004047259222716093, + 0.004715004935860634, + 0.005490305833518505, + 0.004509249236434698, + 0.00425708619877696, + 0.0035355433356016874, + 0.005847474560141563, + 0.004319850821048021, + 0.003987498115748167, + 0.007594740018248558, + 0.007623552810400724, + 0.012585134245455265, + 0.002543065929785371, + 0.002346801571547985, + 0.003951886668801308, + 0.0038713752292096615, + 0.0019212380284443498, + 0.002551056444644928, + 0.002135440008714795, + 0.0017345123924314976, + 0.013898944482207298, + 0.007912092842161655, + 0.10555528104305267, + 0.06198696047067642, + 0.017430182546377182, + 0.01059531606733799, + 0.012729348614811897, + 0.01878831721842289, + 0.004402361810207367, + 0.002687284955754876, + 0.004688490182161331, + 0.0043049887754023075, + 0.0025006760843098164, + 0.0032772596459835768, + 0.00467752106487751, + 0.004799957387149334, + 0.023241881281137466, + 0.014235816895961761, + 0.11797032505273819, + 0.013467314653098583, + 0.008519058115780354, + 0.017986563965678215, + 0.012060308828949928, + 0.08063481003046036, + 0.008292961865663528, + 0.0106063112616539, + 0.01377512514591217, + 0.011880820617079735, + 0.01199769414961338, + 0.036280300468206406, + 0.00908223632723093, + 0.016843067482113838, + 0.010987544432282448, + 0.011158296838402748, + 0.01904989965260029, + 0.014063738286495209, + 0.004589777905493975, + 0.007342135068029165, + 0.010644582100212574, + 0.008222660981118679, + 0.007995697669684887, + 0.005861224140971899, + 0.01351804006844759, + 0.012234491296112537, + 0.0049444278702139854, + 0.0038600151892751455, + 0.0075750467367470264, + 0.012931221164762974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.03369971364736557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795892536640167 + ], + "opposite_logprob_tokens": [ + -3.5297958850860596 + ], + "logprob_reward": -0.029795892536640167, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795892536640167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01741931214928627, + 0.024173835292458534, + 0.020359763875603676, + 0.00872872769832611, + 0.02592400461435318, + 0.06200776249170303, + 0.7127116918563843, + 0.07852882146835327, + 0.028874538838863373, + 0.02127157896757126 + ], + "credit_map": [ + 0.002807173179462552, + 0.0024301984813064337, + 0.002335545839741826, + 0.0016065349336713552, + 0.001394130871631205, + 0.0024916050024330616, + 0.0014142143772915006, + 0.0029399103950709105, + 0.002755236579105258, + 0.0027201748453080654, + 0.003325065830722451, + 0.0027716807089746, + 0.002070613205432892, + 0.004147286061197519, + 0.0026202728040516376, + 0.003763507353141904, + 0.0015211568679660559, + 0.00164198805578053, + 0.0020619540009647608, + 0.002081868704408407, + 0.0016714995726943016, + 0.0033695772290229797, + 0.002546944422647357, + 0.005464775953441858, + 0.0008876555948518217, + 0.0008828071295283735, + 0.0011437443317845464, + 0.0014254482230171561, + 0.0007988979341462255, + 0.0013779113069176674, + 0.0010483663063496351, + 0.0011638967553153634, + 0.0015683714300394058, + 0.0012866106117144227, + 0.009595438838005066, + 0.0034850204829126596, + 0.0023435542825609446, + 0.002000359585508704, + 0.0016499209450557828, + 0.003994727972894907, + 0.016407586634159088, + 0.009237242862582207, + 0.008552885614335537, + 0.007133872248232365, + 0.005529888439923525, + 0.004766614176332951, + 0.004774510394781828, + 0.0056051649153232574, + 0.05715916305780411, + 0.034731827676296234, + 0.2940702438354492, + 0.033069320023059845, + 0.018772682175040245, + 0.04536930471658707, + 0.028713881969451904, + 0.20082524418830872, + 0.006228817161172628, + 0.009600761346518993, + 0.007808194495737553, + 0.0069635906256735325, + 0.0062657808884978294, + 0.023257676512002945, + 0.004999442026019096, + 0.01340455561876297, + 0.0036416598595678806, + 0.003000974655151367, + 0.00439211493358016, + 0.005422633606940508, + 0.0022372398525476456, + 0.003892058739438653, + 0.003383434610441327, + 0.0029044211842119694, + 0.0030320112127810717, + 0.0021723047830164433, + 0.003718082793056965, + 0.003282311838120222, + 0.0019702573772519827, + 0.0009250343427993357, + 0.0020943228155374527, + 0.0040772538632154465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.029795892536640167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202240288257599 + ], + "opposite_logprob_tokens": [ + -2.8120224475860596 + ], + "logprob_reward": -0.06202240288257599, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202240288257599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009056073613464832, + 0.014685438945889473, + 0.011370099149644375, + 0.009323987178504467, + 0.03146130591630936, + 0.02026553265750408, + 0.6203960180282593, + 0.20721761882305145, + 0.06633800268173218, + 0.009885972365736961 + ], + "credit_map": [ + 0.001129789394326508, + 0.0009779363172128797, + 0.001255278941243887, + 0.0008732856367714703, + 0.0009000925347208977, + 0.0014192847302183509, + 0.0008996529504656792, + 0.001600753515958786, + 0.0023748998064547777, + 0.002146360231563449, + 0.0016674279468134046, + 0.0018287698039785028, + 0.0013026196975260973, + 0.0021101508755236864, + 0.0017747568199411035, + 0.0014804541133344173, + 0.0011357521871104836, + 0.0008907165029086173, + 0.0018556524300947785, + 0.0010193632915616035, + 0.0007212061318568885, + 0.0018038979033008218, + 0.001514418050646782, + 0.0024290920700877905, + 0.0011154661187902093, + 0.0009039451251737773, + 0.0015550339594483376, + 0.0018744554836302996, + 0.0008520684204995632, + 0.0010458031902089715, + 0.0011135709937661886, + 0.0008636437123641372, + 0.0028084779623895884, + 0.002426479710265994, + 0.011893955990672112, + 0.005286060739308596, + 0.0028377578128129244, + 0.0020979195833206177, + 0.002188674872741103, + 0.0019219788955524564, + 0.004550415091216564, + 0.0025055604055523872, + 0.003457765793427825, + 0.0023841869551688433, + 0.0015177674358710647, + 0.0018608425743877888, + 0.0016886221710592508, + 0.0023003725800663233, + 0.047690555453300476, + 0.030670326203107834, + 0.2557717263698578, + 0.029341239482164383, + 0.015706254169344902, + 0.040255993604660034, + 0.023470038548111916, + 0.17748987674713135, + 0.017330747097730637, + 0.02222026325762272, + 0.03728234022855759, + 0.03132107853889465, + 0.017509520053863525, + 0.04211962968111038, + 0.01595664955675602, + 0.023477384820580482, + 0.008142351172864437, + 0.009893376380205154, + 0.00982461217790842, + 0.009630327112972736, + 0.005123047158122063, + 0.007265815045684576, + 0.008149800822138786, + 0.008308677934110165, + 0.0017471452010795474, + 0.0014156430261209607, + 0.0016066690441220999, + 0.0011608386412262917, + 0.0007007138920016587, + 0.00039391592144966125, + 0.000988677842542529, + 0.00187236862257123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.06202240288257599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594549672678113 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.00594549672678113, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594549672678113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041126549243927, + 0.03902196139097214, + 0.036264222115278244, + 0.02207998000085354, + 0.044272467494010925, + 0.05226525291800499, + 0.5441126823425293, + 0.13127733767032623, + 0.0434051975607872, + 0.04617432504892349 + ], + "credit_map": [ + 0.008108570240437984, + 0.006236131303012371, + 0.00593489408493042, + 0.004350761882960796, + 0.0031371568329632282, + 0.004634473938494921, + 0.004422257654368877, + 0.0043023074977099895, + 0.0044379886239767075, + 0.00433880602940917, + 0.0048100766725838184, + 0.00394277973100543, + 0.002929334994405508, + 0.007551194634288549, + 0.00490728160366416, + 0.006104502826929092, + 0.003422322915866971, + 0.003657733090221882, + 0.004472671542316675, + 0.005691400729119778, + 0.00275179510936141, + 0.004544256255030632, + 0.0043976143933832645, + 0.007326431106775999, + 0.0019767882768064737, + 0.001419917563907802, + 0.0024491269141435623, + 0.003898096038028598, + 0.0022631629835814238, + 0.00396078173071146, + 0.0040730200707912445, + 0.0020390902645885944, + 0.0023657765705138445, + 0.002905984641984105, + 0.013993622735142708, + 0.012730885297060013, + 0.0029337378218770027, + 0.0027624184731394053, + 0.0033131579402834177, + 0.003266882849857211, + 0.010361241176724434, + 0.008045697584748268, + 0.008175557479262352, + 0.006284141913056374, + 0.005099245347082615, + 0.004924534820020199, + 0.004629568662494421, + 0.004745271522551775, + 0.041611287742853165, + 0.027884671464562416, + 0.2238244116306305, + 0.025891359895467758, + 0.013790843077003956, + 0.0347827710211277, + 0.020395375788211823, + 0.1559319645166397, + 0.01272627618163824, + 0.01578601822257042, + 0.021680612117052078, + 0.026288239285349846, + 0.008928065188229084, + 0.017808200791478157, + 0.0104255061596632, + 0.017634432762861252, + 0.006047607399523258, + 0.003960004076361656, + 0.011090438812971115, + 0.006388992536813021, + 0.002951942617073655, + 0.0038087554275989532, + 0.0047102393582463264, + 0.004447219893336296, + 0.00782306119799614, + 0.003952385857701302, + 0.008288949728012085, + 0.007881136611104012, + 0.003912514541298151, + 0.0025345126632601023, + 0.005028425715863705, + 0.006753345485776663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.00594549672678113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a low-angle perspective, the camera begins a slow, deliberate, winding tracking shot, gliding weightlessly through a narrow canyon carved by architectural giants. These massive skyscrapers rise like an ancient, silent forest, their peaks lost in a clouded sky and soaring, invisible ceilings. Soft, indirect light filters down from above, bathing their massive forms in a diffuse glow and capturing the various textures of their exteriors—the cool, smooth sheen of deep blue reflective glass, the matte, porous surface of brutalist concrete, and the intricate metal grids and steel exoskeletons. The camera drifts smoothly forward, slowly tilting across the canyon, revealing new perspectives on this dense urban jungle: skybridges winding like fallen logs between the towers, recessed windows forming intricate, shadowed patterns, and a subtle play of diffused light creating gentle gradients across the vast, steep surfaces. The overall color palette is a cool, desaturated mix of slate gray, deep blue, and soft silver, creating a profound atmosphere of quiet, contemplative awe and tranquil solitude in the heart of this monumental, sleepy metropolis.", + "seed": 19068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_5/index_7207_seed_19068_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04569156104698777, + "vlm_logprob": -0.04569156104698777, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.04569156104698777, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009901338256895542 + ], + "opposite_logprob_tokens": [ + -9.00990104675293 + ], + "logprob_reward": -0.009901338256895542, + "yes_no_margin_reward": 9.0, + "token_credit_reward": -0.009901338256895542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057339057326316833, + 0.07809656858444214, + 0.04412229731678963, + 0.2851675748825073, + 0.29494336247444153, + 0.024465162307024002, + 0.11275482922792435, + 0.03288816660642624, + 0.04357583820819855, + 0.02664710022509098 + ], + "credit_map": [ + 0.005986494943499565, + 0.005088826175779104, + 0.007418602705001831, + 0.010582393035292625, + 0.004262859933078289, + 0.007382933981716633, + 0.008946220390498638, + 0.007670729421079159, + 0.00968764629215002, + 0.009478791616857052, + 0.011187704280018806, + 0.01014731079339981, + 0.00513219740241766, + 0.012872708030045033, + 0.009161770343780518, + 0.010428435169160366, + 0.005807258188724518, + 0.004700059536844492, + 0.006516904570162296, + 0.004901673179119825, + 0.006937142927199602, + 0.005639785900712013, + 0.005528240464627743, + 0.004091231152415276, + 0.016335798427462578, + 0.014512769877910614, + 0.04926152899861336, + 0.056467846035957336, + 0.010806730948388577, + 0.012122899293899536, + 0.07495790719985962, + 0.05070209875702858, + 0.01616496406495571, + 0.00778228510171175, + 0.11243810504674911, + 0.12024430930614471, + 0.006868177093565464, + 0.005379484035074711, + 0.012412198819220066, + 0.01365381758660078, + 0.002640875754877925, + 0.0018531088717281818, + 0.005055968649685383, + 0.004913086071610451, + 0.0041833799332380295, + 0.0015083919279277325, + 0.0026658237911760807, + 0.0016445259097963572, + 0.032259151339530945, + 0.011322733014822006, + 0.024314861744642258, + 0.01517893560230732, + 0.007585061714053154, + 0.005780814681202173, + 0.0068177953362464905, + 0.009495483711361885, + 0.0068452199921011925, + 0.0032512471079826355, + 0.005117298103868961, + 0.004499264992773533, + 0.004147843457758427, + 0.0019982291851192713, + 0.004078615456819534, + 0.00295044993981719, + 0.0063875955529510975, + 0.006214486435055733, + 0.005207980051636696, + 0.006157950032502413, + 0.006640486419200897, + 0.0023647365160286427, + 0.005665870849043131, + 0.004936727695167065, + 0.005682901944965124, + 0.003454940626397729, + 0.003532146103680134, + 0.0028138277120888233, + 0.0025779514107853174, + 0.002438680501654744, + 0.002644941443577409, + 0.0035017102491110563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.009901338256895542 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306294396519661 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306294396519661, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306294396519661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053608767688274384, + 0.02502281591296196, + 0.08715862780809402, + 0.135281503200531, + 0.5152757167816162, + 0.023238707333803177, + 0.08648774772882462, + 0.024197274819016457, + 0.014078333042562008, + 0.035650450736284256 + ], + "credit_map": [ + 0.00544968293979764, + 0.004695042036473751, + 0.007400472182780504, + 0.012308097444474697, + 0.003938659094274044, + 0.005864654202014208, + 0.0072919330559670925, + 0.0066602276638150215, + 0.002053799107670784, + 0.002235143445432186, + 0.0032688924111425877, + 0.0030393090564757586, + 0.002276354469358921, + 0.0033005084842443466, + 0.004519851412624121, + 0.004328959155827761, + 0.008065275847911835, + 0.009487410075962543, + 0.014392956160008907, + 0.01576729491353035, + 0.010055778548121452, + 0.0091240918263793, + 0.011236198246479034, + 0.009029624983668327, + 0.008191152475774288, + 0.006004075985401869, + 0.026489755138754845, + 0.021719682961702347, + 0.005542732309550047, + 0.0057389577850699425, + 0.04297187179327011, + 0.01862327568233013, + 0.030099963769316673, + 0.016461776569485664, + 0.1957416832447052, + 0.2053494155406952, + 0.012907253578305244, + 0.010327881202101707, + 0.022379279136657715, + 0.022008536383509636, + 0.0051772440783679485, + 0.002494593383744359, + 0.003537380136549473, + 0.0031678429804742336, + 0.002252937527373433, + 0.001451245741918683, + 0.0033208446111530066, + 0.001836619689129293, + 0.02601364254951477, + 0.009706531651318073, + 0.019202785566449165, + 0.012411740608513355, + 0.005194050259888172, + 0.0035566063597798347, + 0.00471983989700675, + 0.005682553164660931, + 0.004734660964459181, + 0.002250576624646783, + 0.004248281009495258, + 0.0035337423905730247, + 0.0036259223707020283, + 0.0015608599642291665, + 0.0027091638185083866, + 0.001534068607725203, + 0.0022214422933757305, + 0.002096747048199177, + 0.0016985514666885138, + 0.0018136607250198722, + 0.002110673114657402, + 0.0010488942498341203, + 0.0016973359743133187, + 0.0013910293346270919, + 0.007392270490527153, + 0.004855846054852009, + 0.004740563221275806, + 0.0030584060586988926, + 0.003014169866219163, + 0.0038471545558422804, + 0.0042430926114320755, + 0.004498948343098164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.04306294396519661 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054952941834926605 + ], + "opposite_logprob_tokens": [ + -2.92995285987854 + ], + "logprob_reward": -0.054952941834926605, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054952941834926605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028853334486484528, + 0.02471267245709896, + 0.039514198899269104, + 0.31120485067367554, + 0.3087652921676636, + 0.03102654218673706, + 0.12872043251991272, + 0.06656908988952637, + 0.038641780614852905, + 0.021991780027747154 + ], + "credit_map": [ + 0.002050637034699321, + 0.0026362675707787275, + 0.0029265719931572676, + 0.003950329963117838, + 0.002990925684571266, + 0.0028283162973821163, + 0.0060020615346729755, + 0.005468226037919521, + 0.0023960003163665533, + 0.0014077241066843271, + 0.003793277544900775, + 0.0038988892920315266, + 0.0016601895913481712, + 0.0037810124922543764, + 0.0039948052726686, + 0.003780773840844631, + 0.0048299143090844154, + 0.00357992434874177, + 0.006124093662947416, + 0.005723613779991865, + 0.004690621048212051, + 0.004836831707507372, + 0.005048587452620268, + 0.004680613987147808, + 0.01931840367615223, + 0.01078783255070448, + 0.04815177991986275, + 0.059240762144327164, + 0.010704859159886837, + 0.012297493405640125, + 0.12243050336837769, + 0.028273209929466248, + 0.0182622279971838, + 0.008912285789847374, + 0.11615849286317825, + 0.12549474835395813, + 0.0067318035289645195, + 0.00524800643324852, + 0.01375018060207367, + 0.01420755684375763, + 0.006082698702812195, + 0.003945303615182638, + 0.004845687188208103, + 0.0047417618334293365, + 0.0031844116747379303, + 0.0016826310893520713, + 0.003173812758177519, + 0.0033702340442687273, + 0.042291734367609024, + 0.013155782595276833, + 0.03030019998550415, + 0.015999810770154, + 0.006234499625861645, + 0.005222192965447903, + 0.0064725750125944614, + 0.009043633937835693, + 0.01725904643535614, + 0.005648073274642229, + 0.01249750517308712, + 0.009967890568077564, + 0.005562294274568558, + 0.0034804700408130884, + 0.006658308673650026, + 0.005495496094226837, + 0.006127937696874142, + 0.0037784334272146225, + 0.004374722484499216, + 0.006534164305776358, + 0.004970072768628597, + 0.0025029403623193502, + 0.005554921459406614, + 0.0047985888086259365, + 0.0051564970053732395, + 0.0024848703760653734, + 0.003207889851182699, + 0.0024443853180855513, + 0.002192022977396846, + 0.0017456322675570846, + 0.002331394702196121, + 0.0024290871806442738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.054952941834926605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291328817605972 + ], + "opposite_logprob_tokens": [ + -2.2379133701324463 + ], + "logprob_reward": -0.11291328817605972, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291328817605972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06311845034360886, + 0.05952182412147522, + 0.21531741321086884, + 0.20464125275611877, + 0.11205800622701645, + 0.05021248757839203, + 0.14601267874240875, + 0.09478060901165009, + 0.02076028846204281, + 0.033577028661966324 + ], + "credit_map": [ + 0.006301965098828077, + 0.0075447228737175465, + 0.008050303906202316, + 0.00868168007582426, + 0.005819124169647694, + 0.007358862087130547, + 0.00999842956662178, + 0.00936336349695921, + 0.00682915048673749, + 0.005067360121756792, + 0.00631236145272851, + 0.006121957208961248, + 0.004331745207309723, + 0.008664225228130817, + 0.011694562621414661, + 0.01050045806914568, + 0.036561451852321625, + 0.024046380072832108, + 0.031385038048028946, + 0.021629396826028824, + 0.02893972583115101, + 0.02808263711631298, + 0.025001898407936096, + 0.019670909270644188, + 0.012592068873345852, + 0.014785815961658955, + 0.022497359663248062, + 0.06630423665046692, + 0.013295518234372139, + 0.012968496419489384, + 0.033096954226493835, + 0.029100779443979263, + 0.007259913254529238, + 0.0047289119102060795, + 0.03909981623291969, + 0.04058413580060005, + 0.005400739144533873, + 0.0033113807439804077, + 0.005936027970165014, + 0.0057370844297111034, + 0.011076288297772408, + 0.005290871951729059, + 0.008991587907075882, + 0.008405612781643867, + 0.004541030619293451, + 0.002780642593279481, + 0.004877968691289425, + 0.004248486366122961, + 0.04825907200574875, + 0.012892676517367363, + 0.03592994064092636, + 0.017445122823119164, + 0.007126398850232363, + 0.0053421445190906525, + 0.008194408379495144, + 0.01082291267812252, + 0.02481669746339321, + 0.00937301479279995, + 0.015361344441771507, + 0.01484993752092123, + 0.007528588641434908, + 0.004979625344276428, + 0.01089825201779604, + 0.006973145063966513, + 0.003585532307624817, + 0.0031481231562793255, + 0.002583820605650544, + 0.0025475199799984694, + 0.0018493164097890258, + 0.0016473809955641627, + 0.0014871492749080062, + 0.0039114453829824924, + 0.005633264780044556, + 0.004931482952088118, + 0.004320783540606499, + 0.0033953646197915077, + 0.004228351172059774, + 0.0026673756074160337, + 0.0037827324122190475, + 0.004617674741894007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.11291328817605972 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762729300186038 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.00762729300186038, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762729300186038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048321839421987534, + 0.08451413363218307, + 0.02356339804828167, + 0.24190683662891388, + 0.306377649307251, + 0.03926979377865791, + 0.14327487349510193, + 0.05895630270242691, + 0.028725450858473778, + 0.025089723989367485 + ], + "credit_map": [ + 0.005140489432960749, + 0.004404986277222633, + 0.005606784485280514, + 0.007641361095011234, + 0.003741804976016283, + 0.005595480091869831, + 0.007815783843398094, + 0.008375151082873344, + 0.00822035875171423, + 0.009624119848012924, + 0.011695384979248047, + 0.010751264169812202, + 0.007024903316050768, + 0.010563762858510017, + 0.012660741806030273, + 0.013973596505820751, + 0.0019395083654671907, + 0.002702123485505581, + 0.003099563531577587, + 0.0031649316661059856, + 0.003493515308946371, + 0.0031285572331398726, + 0.0034171538427472115, + 0.0026180448476225138, + 0.01492119301110506, + 0.012265539728105068, + 0.03812377527356148, + 0.03388969972729683, + 0.00972777884453535, + 0.01073931809514761, + 0.0938146710395813, + 0.028424851596355438, + 0.016720520332455635, + 0.00868741050362587, + 0.1148700937628746, + 0.12561146914958954, + 0.006968211382627487, + 0.005296394694596529, + 0.013584750704467297, + 0.014638800173997879, + 0.008738548494875431, + 0.004982746206223965, + 0.00735812820494175, + 0.005727631971240044, + 0.004788408987224102, + 0.0019430656684562564, + 0.003217036370187998, + 0.0025142249651253223, + 0.04371006041765213, + 0.014634404331445694, + 0.0314163938164711, + 0.018939126282930374, + 0.009339448064565659, + 0.007203387562185526, + 0.007753755431622267, + 0.010278288275003433, + 0.013008966110646725, + 0.005136467516422272, + 0.009803512133657932, + 0.008410505950450897, + 0.006372785661369562, + 0.003811922622844577, + 0.00718972273170948, + 0.005222419276833534, + 0.003636854002252221, + 0.002943899482488632, + 0.0029741397593170404, + 0.0051399460062384605, + 0.004323581233620644, + 0.0018731855088844895, + 0.004390542861074209, + 0.003443303983658552, + 0.0051017762161791325, + 0.00431737769395113, + 0.0032724649645388126, + 0.003261399222537875, + 0.003118970664218068, + 0.0015767617151141167, + 0.002278324216604233, + 0.0021626497618854046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.00762729300186038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a low-angle perspective, the camera begins a slow, deliberate, winding tracking shot, gliding weightlessly through a narrow canyon carved by architectural giants. These massive skyscrapers rise like an ancient, silent forest, their peaks lost in a clouded sky and soaring, invisible ceilings. Soft, indirect light filters down from above, bathing their massive forms in a diffuse glow and capturing the various textures of their exteriors—the cool, smooth sheen of deep blue reflective glass, the matte, porous surface of brutalist concrete, and the intricate metal grids and steel exoskeletons. The camera drifts smoothly forward, slowly tilting across the canyon, revealing new perspectives on this dense urban jungle: skybridges winding like fallen logs between the towers, recessed windows forming intricate, shadowed patterns, and a subtle play of diffused light creating gentle gradients across the vast, steep surfaces. The overall color palette is a cool, desaturated mix of slate gray, deep blue, and soft silver, creating a profound atmosphere of quiet, contemplative awe and tranquil solitude in the heart of this monumental, sleepy metropolis.", + "seed": 19069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_5/index_7207_seed_19069_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04299867749214172, + "vlm_logprob": -0.04299867749214172, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.04299867749214172, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012980958446860313 + ], + "opposite_logprob_tokens": [ + -7.762980937957764 + ], + "logprob_reward": -0.012980958446860313, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.012980958446860313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09213024377822876, + 0.06413821876049042, + 0.0555654875934124, + 0.06388768553733826, + 0.031213179230690002, + 0.5000991821289062, + 0.08071143925189972, + 0.039415761828422546, + 0.03212199732661247, + 0.04071684181690216 + ], + "credit_map": [ + 0.010087096132338047, + 0.00973264966160059, + 0.013620623387396336, + 0.021925313398241997, + 0.006536242086440325, + 0.008943134918808937, + 0.012994221411645412, + 0.008290962316095829, + 0.016456810757517815, + 0.011303579434752464, + 0.008920692838728428, + 0.006945507135242224, + 0.004227298777550459, + 0.005033678840845823, + 0.007480868604034185, + 0.003769782604649663, + 0.010794563218951225, + 0.005413603503257036, + 0.00973490346223116, + 0.004496251232922077, + 0.0028936825692653656, + 0.006089042406529188, + 0.0034716431982815266, + 0.012671797536313534, + 0.006515616085380316, + 0.003650394966825843, + 0.0037237955257296562, + 0.0028875544667243958, + 0.004255098290741444, + 0.017259659245610237, + 0.00379541190341115, + 0.021800149232149124, + 0.004070491995662451, + 0.0030333923641592264, + 0.003333853790536523, + 0.0030688075348734856, + 0.0022137865889817476, + 0.0035791078116744757, + 0.004863782785832882, + 0.007049955893307924, + 0.0274062342941761, + 0.08337020874023438, + 0.029168715700507164, + 0.019617605954408646, + 0.02254641056060791, + 0.0712258517742157, + 0.05609806254506111, + 0.190666064620018, + 0.007432934828102589, + 0.007582101505249739, + 0.013310856185853481, + 0.01089203730225563, + 0.006760154850780964, + 0.010946876369416714, + 0.011324217543005943, + 0.012462264858186245, + 0.004525706637650728, + 0.0025496995076537132, + 0.0043531982228159904, + 0.004515217617154121, + 0.0040202694945037365, + 0.008456099778413773, + 0.005724606104195118, + 0.005270966794341803, + 0.0033790606539696455, + 0.0023476581554859877, + 0.005732348654419184, + 0.0038326370995491743, + 0.002872046083211899, + 0.003913000226020813, + 0.005656903143972158, + 0.0043883430771529675, + 0.003509263740852475, + 0.0028041722252964973, + 0.005563443060964346, + 0.0046644434332847595, + 0.004881652072072029, + 0.006006598938256502, + 0.006986480206251144, + 0.0063007851131260395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.012980958446860313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306488484144211 + ], + "opposite_logprob_tokens": [ + -3.168064832687378 + ], + "logprob_reward": -0.04306488484144211, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306488484144211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05809469893574715, + 0.026630161330103874, + 0.035106029361486435, + 0.07193435728549957, + 0.016920536756515503, + 0.6989802122116089, + 0.037608224898576736, + 0.024386728182435036, + 0.013540223240852356, + 0.016798807308077812 + ], + "credit_map": [ + 0.011002439074218273, + 0.005857669282704592, + 0.011305744759738445, + 0.006765828002244234, + 0.004833662882447243, + 0.004554811865091324, + 0.008588080294430256, + 0.005186459980905056, + 0.004407826345413923, + 0.003300093114376068, + 0.0037036752328276634, + 0.002961499849334359, + 0.002989973872900009, + 0.0026073548942804337, + 0.004281744360923767, + 0.0023779955226927996, + 0.006537824869155884, + 0.0024642038624733686, + 0.005509285256266594, + 0.0022188681177794933, + 0.001836310955695808, + 0.005002239719033241, + 0.002145623555406928, + 0.009391671046614647, + 0.007948588579893112, + 0.005097936373203993, + 0.00331336772069335, + 0.003207715228199959, + 0.00442692544311285, + 0.02029598504304886, + 0.0037528418470174074, + 0.02389099821448326, + 0.0023797659669071436, + 0.0013160734670236707, + 0.0027081414591521025, + 0.002543496433645487, + 0.0013844642089679837, + 0.002163092838600278, + 0.002014755504205823, + 0.002410746645182371, + 0.03945987671613693, + 0.11687635630369186, + 0.041544754058122635, + 0.02714812196791172, + 0.03075731173157692, + 0.09912845492362976, + 0.07725834846496582, + 0.26680704951286316, + 0.003118955297395587, + 0.002110164612531662, + 0.0043743327260017395, + 0.0034996620379388332, + 0.005655884277075529, + 0.007300505880266428, + 0.005491950549185276, + 0.006056768819689751, + 0.003978143911808729, + 0.001954750157892704, + 0.0026785784866660833, + 0.0022857722360640764, + 0.002298993058502674, + 0.004316362086683512, + 0.0035720772575587034, + 0.0033020528499037027, + 0.0013006560038775206, + 0.000986849656328559, + 0.0016779417637735605, + 0.0015875652898102999, + 0.001531466725282371, + 0.0019120753277093172, + 0.0024352797772735357, + 0.0021083892788738012, + 0.001703432179056108, + 0.0010678445687517524, + 0.002098919590935111, + 0.0014734093565493822, + 0.0019805736374109983, + 0.0021734582260251045, + 0.0035912361927330494, + 0.0027099340222775936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.04306488484144211 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289820179343224 + ], + "opposite_logprob_tokens": [ + -3.773289918899536 + ], + "logprob_reward": -0.023289820179343224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289820179343224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.256253182888031, + 0.058144956827163696, + 0.09077690541744232, + 0.07732189446687698, + 0.037233415991067886, + 0.23955364525318146, + 0.12377120554447174, + 0.07342855632305145, + 0.021363841369748116, + 0.022152401506900787 + ], + "credit_map": [ + 0.043381545692682266, + 0.036922283470630646, + 0.05476388707756996, + 0.04740799218416214, + 0.019264783710241318, + 0.016294511035084724, + 0.021491779014468193, + 0.01672641560435295, + 0.013137751258909702, + 0.00729641318321228, + 0.007691793143749237, + 0.005578134674578905, + 0.004916892386972904, + 0.00595596618950367, + 0.007976514287292957, + 0.00559149170294404, + 0.015101996250450611, + 0.005749780219048262, + 0.013418941758573055, + 0.008001557551324368, + 0.004840217065066099, + 0.013022633269429207, + 0.006682182662189007, + 0.023959601297974586, + 0.006882856134325266, + 0.006886465009301901, + 0.004879799205809832, + 0.0049170381389558315, + 0.004239423666149378, + 0.017113177105784416, + 0.004451243672519922, + 0.02795189432799816, + 0.004119419492781162, + 0.0032802377827465534, + 0.003840359393507242, + 0.003627792466431856, + 0.0037160804495215416, + 0.007469202391803265, + 0.005759533029049635, + 0.00542079284787178, + 0.013015304692089558, + 0.03983307257294655, + 0.014438516460359097, + 0.009358507581055164, + 0.01106253545731306, + 0.03389737755060196, + 0.026468563824892044, + 0.09147976338863373, + 0.009037582203745842, + 0.008759163320064545, + 0.012350511737167835, + 0.014075509272515774, + 0.015285233035683632, + 0.024868877604603767, + 0.017862170934677124, + 0.02153216302394867, + 0.015620175749063492, + 0.0053956164047122, + 0.00899900496006012, + 0.007705796509981155, + 0.0060836500488221645, + 0.011575249023735523, + 0.008704211562871933, + 0.009344852529466152, + 0.0022101644426584244, + 0.0010978167410939932, + 0.002432401990517974, + 0.0022623285185545683, + 0.002945039886981249, + 0.0032113578636199236, + 0.0038561003748327494, + 0.0033486306201666594, + 0.0023834509775042534, + 0.00096484663663432, + 0.002714843489229679, + 0.0033989667426794767, + 0.0027453145012259483, + 0.004079652018845081, + 0.003256376599892974, + 0.002608950948342681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.023289820179343224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702086567878723 + ], + "opposite_logprob_tokens": [ + -2.127020835876465 + ], + "logprob_reward": -0.12702086567878723, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702086567878723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31900984048843384, + 0.05109249800443649, + 0.06695342063903809, + 0.09523041546344757, + 0.028227481991052628, + 0.16614529490470886, + 0.0644925907254219, + 0.0624714158475399, + 0.09669337421655655, + 0.04968371242284775 + ], + "credit_map": [ + 0.057071443647146225, + 0.03742099180817604, + 0.07087613642215729, + 0.05900914967060089, + 0.02496137097477913, + 0.020560983568429947, + 0.02745761163532734, + 0.02165214717388153, + 0.012291230261325836, + 0.008594847284257412, + 0.007968130521476269, + 0.00558240944519639, + 0.00421936996281147, + 0.003232203656807542, + 0.0052934680134058, + 0.003910834901034832, + 0.011106344871222973, + 0.005663455929607153, + 0.010617021471261978, + 0.00792980007827282, + 0.003732449607923627, + 0.009257566183805466, + 0.004510552156716585, + 0.014136227779090405, + 0.008943994529545307, + 0.005115226376801729, + 0.01605408638715744, + 0.006383123807609081, + 0.0034496004227548838, + 0.017194025218486786, + 0.004938265308737755, + 0.03315208852291107, + 0.002628873335197568, + 0.0033245247323065996, + 0.003988068085163832, + 0.002653345698490739, + 0.0022300437558442354, + 0.003325114259496331, + 0.002984230173751712, + 0.007093279622495174, + 0.010189652442932129, + 0.026949003338813782, + 0.010632000863552094, + 0.00681030610576272, + 0.00794081762433052, + 0.023509610444307327, + 0.018281931057572365, + 0.06183198094367981, + 0.00688614696264267, + 0.0045644622296094894, + 0.010430442169308662, + 0.009779602289199829, + 0.005294125527143478, + 0.00882586371153593, + 0.00806550495326519, + 0.010646437294781208, + 0.011001135222613811, + 0.005231850314885378, + 0.007586600724607706, + 0.006008148193359375, + 0.007268935441970825, + 0.009359720163047314, + 0.007464428432285786, + 0.008550596423447132, + 0.007317176554352045, + 0.0058126794174313545, + 0.008544600568711758, + 0.013353025540709496, + 0.012186427600681782, + 0.018202343955636024, + 0.01724770851433277, + 0.01402941532433033, + 0.0050448160618543625, + 0.002720666117966175, + 0.007293356582522392, + 0.006287525873631239, + 0.004318087827414274, + 0.006707523483783007, + 0.009040988981723785, + 0.008270742371678352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.12702086567878723 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636858314275742 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636858314275742, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636858314275742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2393300086259842, + 0.0791868045926094, + 0.0434424914419651, + 0.13190819323062897, + 0.04044158384203911, + 0.18237565457820892, + 0.0750466138124466, + 0.08391324430704117, + 0.03158262372016907, + 0.09277273714542389 + ], + "credit_map": [ + 0.05000881850719452, + 0.02693210355937481, + 0.04665903374552727, + 0.03297141194343567, + 0.01981145329773426, + 0.01752975955605507, + 0.02781655080616474, + 0.01760089583694935, + 0.019955893978476524, + 0.010148250497877598, + 0.012134808115661144, + 0.007484175264835358, + 0.006358720827847719, + 0.0062696016393601894, + 0.009698478505015373, + 0.007136881817132235, + 0.00922022107988596, + 0.004470421466976404, + 0.0061231995932757854, + 0.009994473308324814, + 0.0026775901205837727, + 0.003084406489506364, + 0.004342993255704641, + 0.003529187059029937, + 0.016737036406993866, + 0.012941934168338776, + 0.00694865919649601, + 0.00601173983886838, + 0.009438064880669117, + 0.035339221358299255, + 0.006795823108404875, + 0.03769572451710701, + 0.0052774446085095406, + 0.0037376768887043, + 0.003855001414194703, + 0.0037805165629833937, + 0.0038718280848115683, + 0.0073603615164756775, + 0.004986596293747425, + 0.007572162430733442, + 0.010584967210888863, + 0.02854234166443348, + 0.010590877383947372, + 0.007274772971868515, + 0.010176856070756912, + 0.026500452309846878, + 0.01983564905822277, + 0.06886974722146988, + 0.005178503226488829, + 0.005167120136320591, + 0.008584748953580856, + 0.009316711686551571, + 0.00910258200019598, + 0.014927593991160393, + 0.010583818890154362, + 0.012185544706881046, + 0.009456806816160679, + 0.006207005120813847, + 0.012114807963371277, + 0.01208323985338211, + 0.008669593371450901, + 0.013970067724585533, + 0.011149335652589798, + 0.01026239711791277, + 0.003085122210904956, + 0.002405676757916808, + 0.0057134442031383514, + 0.0036207549273967743, + 0.0027463671285659075, + 0.0035269379150122404, + 0.004823474213480949, + 0.005660848692059517, + 0.008265245705842972, + 0.005544029176235199, + 0.01250371988862753, + 0.013738161884248257, + 0.011454914696514606, + 0.01380050741136074, + 0.014809750951826572, + 0.012656415812671185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.008636858314275742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a low-angle perspective, the camera begins a slow, deliberate, winding tracking shot, gliding weightlessly through a narrow canyon carved by architectural giants. These massive skyscrapers rise like an ancient, silent forest, their peaks lost in a clouded sky and soaring, invisible ceilings. Soft, indirect light filters down from above, bathing their massive forms in a diffuse glow and capturing the various textures of their exteriors—the cool, smooth sheen of deep blue reflective glass, the matte, porous surface of brutalist concrete, and the intricate metal grids and steel exoskeletons. The camera drifts smoothly forward, slowly tilting across the canyon, revealing new perspectives on this dense urban jungle: skybridges winding like fallen logs between the towers, recessed windows forming intricate, shadowed patterns, and a subtle play of diffused light creating gentle gradients across the vast, steep surfaces. The overall color palette is a cool, desaturated mix of slate gray, deep blue, and soft silver, creating a profound atmosphere of quiet, contemplative awe and tranquil solitude in the heart of this monumental, sleepy metropolis.", + "seed": 19070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_5/index_7207_seed_19070_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.041595544386655095, + "vlm_logprob": -0.041595544386655095, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.041595544386655095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023551134392619133 + ], + "opposite_logprob_tokens": [ + -8.273550987243652 + ], + "logprob_reward": -0.023551134392619133, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.023551134392619133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3238409459590912, + 0.10964518785476685, + 0.10563718527555466, + 0.07114224880933762, + 0.041495323181152344, + 0.038862839341163635, + 0.09886208921670914, + 0.0744270384311676, + 0.05850186198949814, + 0.07758522778749466 + ], + "credit_map": [ + 0.02267995849251747, + 0.032439980655908585, + 0.03416486084461212, + 0.07205241173505783, + 0.0287758931517601, + 0.04300304502248764, + 0.049762457609176636, + 0.04096236452460289, + 0.016675367951393127, + 0.011195734143257141, + 0.014219880104064941, + 0.012983085587620735, + 0.012463403865695, + 0.012854870408773422, + 0.015927381813526154, + 0.013325465843081474, + 0.015627682209014893, + 0.01894988678395748, + 0.014074600301682949, + 0.012151814997196198, + 0.00804242305457592, + 0.011859772726893425, + 0.01250153873115778, + 0.01242948416620493, + 0.00916247721761465, + 0.008810115046799183, + 0.009974656626582146, + 0.0066817523911595345, + 0.007615308742970228, + 0.010045281611382961, + 0.008697429671883583, + 0.01015523448586464, + 0.004659249912947416, + 0.0054914080537855625, + 0.004206861834973097, + 0.005514383316040039, + 0.004133734852075577, + 0.006138133816421032, + 0.004582488443702459, + 0.006769063416868448, + 0.005606502760201693, + 0.004268600605428219, + 0.004520692862570286, + 0.0058318511582911015, + 0.0029797793831676245, + 0.004742563236504793, + 0.006712343078106642, + 0.004200506489723921, + 0.01537706982344389, + 0.011837647296488285, + 0.012894384562969208, + 0.011518939398229122, + 0.009036634117364883, + 0.01356729120016098, + 0.014470438472926617, + 0.0101596899330616, + 0.009621254168450832, + 0.010086644440889359, + 0.011074773035943508, + 0.007987483404576778, + 0.010700283572077751, + 0.008125001564621925, + 0.008763667196035385, + 0.008067935705184937, + 0.006381259299814701, + 0.005867353640496731, + 0.006199462804943323, + 0.007520200219005346, + 0.006743947509676218, + 0.00917208380997181, + 0.007249790709465742, + 0.009367765858769417, + 0.010201903991401196, + 0.00810637604445219, + 0.012075220234692097, + 0.007737255189567804, + 0.007967141456902027, + 0.010795051231980324, + 0.009349598549306393, + 0.011352689005434513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023551134392619133 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038104183971881866 + ], + "opposite_logprob_tokens": [ + -3.288104295730591 + ], + "logprob_reward": -0.038104183971881866, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038104183971881866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1549721658229828, + 0.05808154493570328, + 0.1851058006286621, + 0.052083782851696014, + 0.07184520363807678, + 0.11255311965942383, + 0.099020816385746, + 0.08566228300333023, + 0.06143980473279953, + 0.11923547089099884 + ], + "credit_map": [ + 0.012866687029600143, + 0.015075021423399448, + 0.02011658437550068, + 0.023000504821538925, + 0.0174957737326622, + 0.020648732781410217, + 0.025503266602754593, + 0.02026558667421341, + 0.007672698702663183, + 0.006321984808892012, + 0.007270900532603264, + 0.008003444410860538, + 0.006229387130588293, + 0.006309626158326864, + 0.00790460966527462, + 0.008368891663849354, + 0.02655121497809887, + 0.029978405684232712, + 0.022864380851387978, + 0.019785163924098015, + 0.01253348495811224, + 0.02447085827589035, + 0.02707289159297943, + 0.021849405020475388, + 0.004033551085740328, + 0.005160550121217966, + 0.005545476451516151, + 0.006973940413445234, + 0.002942252205684781, + 0.012251962907612324, + 0.008735422976315022, + 0.006440627388656139, + 0.007409808691591024, + 0.009490307420492172, + 0.006491038482636213, + 0.008732056245207787, + 0.009479673579335213, + 0.010179388336837292, + 0.01049481425434351, + 0.00956811849027872, + 0.011528315022587776, + 0.009131651371717453, + 0.00894180778414011, + 0.01665126159787178, + 0.010195596143603325, + 0.01623852550983429, + 0.02606988325715065, + 0.013796073384582996, + 0.018824772909283638, + 0.013052009046077728, + 0.010754019021987915, + 0.010633829981088638, + 0.011654768139123917, + 0.012515297159552574, + 0.012891336344182491, + 0.008694789372384548, + 0.007655274588614702, + 0.010922138579189777, + 0.010829294100403786, + 0.007696999702602625, + 0.016327092424035072, + 0.010848096571862698, + 0.011690350249409676, + 0.009693038649857044, + 0.011718008667230606, + 0.004629412200301886, + 0.011572710238397121, + 0.005698861554265022, + 0.004221706185489893, + 0.008629672229290009, + 0.005124336574226618, + 0.009845099411904812, + 0.011492731980979443, + 0.010032548569142818, + 0.015677310526371002, + 0.01511713769286871, + 0.013688014820218086, + 0.019616955891251564, + 0.02049327827990055, + 0.01311749592423439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.038104183971881866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979959361255169 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979959361255169, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979959361255169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20431654155254364, + 0.1691274642944336, + 0.06787722557783127, + 0.036956120282411575, + 0.053187090903520584, + 0.047293320298194885, + 0.08717644214630127, + 0.09507150202989578, + 0.17318235337734222, + 0.06581193208694458 + ], + "credit_map": [ + 0.0175598356872797, + 0.020964832976460457, + 0.023611415177583694, + 0.027892593294382095, + 0.03230341896414757, + 0.02965494617819786, + 0.025399981066584587, + 0.02692953124642372, + 0.02446424774825573, + 0.020653029903769493, + 0.027208928018808365, + 0.014604903757572174, + 0.01859285496175289, + 0.01865747570991516, + 0.024586539715528488, + 0.020359482616186142, + 0.008975286036729813, + 0.011825429275631905, + 0.00762898800894618, + 0.00785322766751051, + 0.005009793676435947, + 0.008199003525078297, + 0.010084631852805614, + 0.00830086413770914, + 0.003984643612056971, + 0.003511890536174178, + 0.004766511265188456, + 0.004840194247663021, + 0.0025634397752583027, + 0.005306577309966087, + 0.007086700294166803, + 0.004896164406090975, + 0.005907343700528145, + 0.00610431469976902, + 0.005779753439128399, + 0.009158389642834663, + 0.004973114468157291, + 0.008122986182570457, + 0.006565043702721596, + 0.00657614367082715, + 0.007111642975360155, + 0.005293059628456831, + 0.005616269540041685, + 0.006347906310111284, + 0.003102015471085906, + 0.00635623000562191, + 0.007636873982846737, + 0.005829320754855871, + 0.008651002310216427, + 0.00772962998598814, + 0.009753669612109661, + 0.021510781720280647, + 0.0058149294927716255, + 0.011117016896605492, + 0.013686450198292732, + 0.008912958204746246, + 0.01924867182970047, + 0.008170065470039845, + 0.01662047766149044, + 0.00616587745025754, + 0.015013017691671848, + 0.011202333495020866, + 0.008368092589080334, + 0.010282971896231174, + 0.016733216121792793, + 0.010117541067302227, + 0.02099798619747162, + 0.0119045814499259, + 0.007974101230502129, + 0.03485780954360962, + 0.02147056721150875, + 0.04912655055522919, + 0.00895272009074688, + 0.005936429835855961, + 0.011871306225657463, + 0.006113036070019007, + 0.005260081496089697, + 0.009778053499758244, + 0.008300784043967724, + 0.009599520824849606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.02979959361255169 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129075363278389 + ], + "opposite_logprob_tokens": [ + -2.237907648086548 + ], + "logprob_reward": -0.1129075363278389, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129075363278389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21333077549934387, + 0.06171112880110741, + 0.13960039615631104, + 0.04460372403264046, + 0.0654105693101883, + 0.10451267659664154, + 0.08201222866773605, + 0.05590105056762695, + 0.06523531675338745, + 0.16768212616443634 + ], + "credit_map": [ + 0.015505283139646053, + 0.026436615735292435, + 0.019117312505841255, + 0.03509595990180969, + 0.037868957966566086, + 0.023647133260965347, + 0.02928844466805458, + 0.02637106366455555, + 0.008333040401339531, + 0.00812000222504139, + 0.007811898831278086, + 0.007262545637786388, + 0.006201687268912792, + 0.006625622045248747, + 0.009809235110878944, + 0.007547098211944103, + 0.01801615208387375, + 0.02021063305437565, + 0.016734853386878967, + 0.016511796042323112, + 0.009521519765257835, + 0.020405452698469162, + 0.020192990079522133, + 0.018007004633545876, + 0.0074397544376552105, + 0.004613583907485008, + 0.005616893991827965, + 0.004835871513932943, + 0.003939713351428509, + 0.006533693056553602, + 0.006128123961389065, + 0.0054960898123681545, + 0.007813300006091595, + 0.007278895005583763, + 0.008307586424052715, + 0.0087063517421484, + 0.006697445642203093, + 0.010093758814036846, + 0.007494979538023472, + 0.009018254466354847, + 0.009540976956486702, + 0.009274841286242008, + 0.00643781665712595, + 0.016907066106796265, + 0.007639674469828606, + 0.011737409979104996, + 0.02893918938934803, + 0.014035697095096111, + 0.008769187144935131, + 0.01130987610667944, + 0.007440878078341484, + 0.014806807041168213, + 0.010744750499725342, + 0.010013985447585583, + 0.011431161314249039, + 0.00749558862298727, + 0.006570128723978996, + 0.006399465259164572, + 0.005735420156270266, + 0.004749976098537445, + 0.013394271954894066, + 0.006646137218922377, + 0.0064681535586714745, + 0.005937498528510332, + 0.005849665030837059, + 0.0055121565237641335, + 0.007930214516818523, + 0.011455927044153214, + 0.006981993559747934, + 0.00855540856719017, + 0.009570533409714699, + 0.009379422292113304, + 0.01658974215388298, + 0.015497968532145023, + 0.024539295583963394, + 0.020565567538142204, + 0.01850668154656887, + 0.023304715752601624, + 0.024978647008538246, + 0.02369951829314232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.1129075363278389 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective looking up at massive skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, winding tracking shot that glides forward through the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the skyscrapers depicted with deep blue reflective glass and matte brutalist concrete surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are skybridges visible winding between the towers like fallen logs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a cool, desaturated mix of slate gray, deep blue, and soft silver?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615273628383875 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615273628383875, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615273628383875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5548034906387329, + 0.056070201098918915, + 0.04322327300906181, + 0.02022521384060383, + 0.021374939009547234, + 0.020553529262542725, + 0.038859397172927856, + 0.1237248107790947, + 0.06419705599546432, + 0.05696800351142883 + ], + "credit_map": [ + 0.038200743496418, + 0.07523249089717865, + 0.04485390707850456, + 0.08073101937770844, + 0.07528818398714066, + 0.06496724486351013, + 0.10104687511920929, + 0.07448302954435349, + 0.006190245971083641, + 0.006153763737529516, + 0.005202723201364279, + 0.007069198880344629, + 0.0053159259259700775, + 0.011154650710523129, + 0.007176719140261412, + 0.0078069730661809444, + 0.004462620243430138, + 0.005361414980143309, + 0.0035165054723620415, + 0.00500417361035943, + 0.005115077830851078, + 0.007447832264006138, + 0.005388115532696247, + 0.00692753354087472, + 0.0018996535800397396, + 0.0019801787566393614, + 0.0024167769588530064, + 0.003215507138520479, + 0.0015089497901499271, + 0.0033493314404040575, + 0.0029751555994153023, + 0.0028796608094125986, + 0.0032876196783035994, + 0.0025058246683329344, + 0.0022413909900933504, + 0.002190381521359086, + 0.002729735802859068, + 0.002721525961533189, + 0.0030611094553023577, + 0.00263735163025558, + 0.002266180468723178, + 0.002323302673175931, + 0.0021865388844162226, + 0.0030789999291300774, + 0.0015333377523347735, + 0.002526307012885809, + 0.004062824882566929, + 0.0025760389398783445, + 0.004152567591518164, + 0.0037323706783354282, + 0.004522612784057856, + 0.004901225678622723, + 0.004100142978131771, + 0.0055200946517288685, + 0.007492425851523876, + 0.004437959752976894, + 0.02705785632133484, + 0.010313224978744984, + 0.02012295089662075, + 0.013882443308830261, + 0.010919805616140366, + 0.015110787004232407, + 0.010760894045233727, + 0.015556850470602512, + 0.006744835525751114, + 0.004564593080431223, + 0.006609525065869093, + 0.0037381472066044807, + 0.006298891734331846, + 0.011227036826312542, + 0.007769610732793808, + 0.01724441722035408, + 0.010278302244842052, + 0.00478737335652113, + 0.010332740843296051, + 0.004151513800024986, + 0.004357367288321257, + 0.009299208410084248, + 0.005969804711639881, + 0.007791691459715366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.003615273628383875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_tensor([4210, 4210])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_tensor([4210, 4210])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0924b72405325474f6452bbfd423515c8e820929 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_tensor([4210, 4210])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up immerses the viewer in the dark, humid incubation chamber, focusing with exquisite clarity on a single, perfect oyster sprouting from a shimmering, dark bottom pocket. The entire scene is visualized with the meticulous, meticulous precision of fine brushwork and vibrant mineral pigments, each detail rendered with a surreal yet stylized elegance. The mushroom's pearlescent fan-shaped cap and delicate, paper-thin gills are defined by sharp, calligraphic lines. A quiet, intense fire begins to spread from the base of the mushroom's stem, its small orange flames casting a warm, flickering glow, creating deep, dancing shadows within the gills. As the relentless heat causes the delicate edges of the gills to first shrivel, then slowly curl inward and begin to char, the transformation unfolds like a writhing ink painting, from pristine white to a fragile, deep black. These colors, from the fiery crimson of the flames to the eroded charcoal, possess the saturated vibrancy of mineral pigments, creating a stark contrast that underscores the silent, destructive process. The atmosphere is one of concentrated, tragic beauty, capturing the elegant consumption of a perfect natural form as thin wisps of smoke rise in delicate, calligraphic curls against a dark, damp background.", + "seed": 11236, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_4210_seed_11236_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0350826732814312, + "vlm_logprob": -0.0350826732814312, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.0350826732814312, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017442872747778893 + ], + "opposite_logprob_tokens": [ + -4.89244270324707 + ], + "logprob_reward": -0.017442872747778893, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017442872747778893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10637805610895157, + 0.07037650793790817, + 0.08396708220243454, + 0.1746225208044052, + 0.03625432401895523, + 0.061279501765966415, + 0.10786280781030655, + 0.19833755493164062, + 0.1304130107164383, + 0.030508587136864662 + ], + "credit_map": [ + 0.0124041847884655, + 0.010289355181157589, + 0.021674050018191338, + 0.009553693234920502, + 0.010757696814835072, + 0.01158651988953352, + 0.012099935673177242, + 0.018012620508670807, + 0.006128211040049791, + 0.009074044413864613, + 0.01122357975691557, + 0.0035719822626560926, + 0.015779685229063034, + 0.010789619758725166, + 0.006388810928910971, + 0.007420576177537441, + 0.004504804033786058, + 0.008151414804160595, + 0.011852205730974674, + 0.00529645849019289, + 0.017462190240621567, + 0.0074491617269814014, + 0.00880820956081152, + 0.020442640408873558, + 0.01640535518527031, + 0.01039282139390707, + 0.04119432717561722, + 0.01604566164314747, + 0.015887169167399406, + 0.026256002485752106, + 0.01400966290384531, + 0.03443153575062752, + 0.00340857170522213, + 0.0030078773852437735, + 0.005480938591063023, + 0.005712233483791351, + 0.003582464763894677, + 0.0031790120992809534, + 0.004767753649502993, + 0.007115476299077272, + 0.006185838021337986, + 0.006681432481855154, + 0.010062966495752335, + 0.009471969678997993, + 0.0061762984842062, + 0.005787814501672983, + 0.007490098010748625, + 0.009423092007637024, + 0.008124451152980328, + 0.007897041738033295, + 0.021261025220155716, + 0.025639452040195465, + 0.006756866816431284, + 0.010415665805339813, + 0.010930048301815987, + 0.016838259994983673, + 0.021143697202205658, + 0.013656040653586388, + 0.0318051353096962, + 0.05638865754008293, + 0.01807173527777195, + 0.016412323340773582, + 0.016238946467638016, + 0.024621037766337395, + 0.019171610474586487, + 0.009284557774662971, + 0.02285267785191536, + 0.013844441622495651, + 0.01098889485001564, + 0.0213141031563282, + 0.012517686001956463, + 0.020439036190509796, + 0.003736177459359169, + 0.0031517238821834326, + 0.005059573333710432, + 0.003474755212664604, + 0.004005861468613148, + 0.0022818725556135178, + 0.0044877915643155575, + 0.004310834687203169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.017442872747778893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08911655843257904 + ], + "opposite_logprob_tokens": [ + -2.4641165733337402 + ], + "logprob_reward": -0.08911655843257904, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08911655843257904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1093926951289177, + 0.10999055206775665, + 0.2901224195957184, + 0.10856585949659348, + 0.047252267599105835, + 0.06569043546915054, + 0.05705146864056587, + 0.0977276936173439, + 0.039993077516555786, + 0.07421351969242096 + ], + "credit_map": [ + 0.012464050203561783, + 0.009996244683861732, + 0.023237217217683792, + 0.013500578701496124, + 0.010350004769861698, + 0.012958569452166557, + 0.014969919808208942, + 0.01191613171249628, + 0.0050462051294744015, + 0.013324713334441185, + 0.015030273236334324, + 0.006543079856783152, + 0.024928543716669083, + 0.024951396510004997, + 0.011821966618299484, + 0.008344383910298347, + 0.024354076012969017, + 0.013869134709239006, + 0.09086627513170242, + 0.017389511689543724, + 0.020536581054329872, + 0.032363053411245346, + 0.018536120653152466, + 0.07220767438411713, + 0.011151008307933807, + 0.006244910880923271, + 0.027299391105771065, + 0.00942781288176775, + 0.00873907282948494, + 0.016039878129959106, + 0.008001238107681274, + 0.021662553772330284, + 0.005097737070173025, + 0.004206101875752211, + 0.008914653211832047, + 0.00720073189586401, + 0.004257965832948685, + 0.005065592937171459, + 0.004649203736335039, + 0.007860285229980946, + 0.005586760584264994, + 0.00564740551635623, + 0.011574114672839642, + 0.017401771619915962, + 0.0048439945094287395, + 0.004716471303254366, + 0.005605448503047228, + 0.010314473882317543, + 0.005542367231100798, + 0.0038963123224675655, + 0.01117880642414093, + 0.01472177729010582, + 0.003686422947794199, + 0.004374871030449867, + 0.004549873527139425, + 0.009101039730012417, + 0.011679683811962605, + 0.008922872133553028, + 0.014725488610565662, + 0.02057657390832901, + 0.011144670657813549, + 0.008723757229745388, + 0.009069112129509449, + 0.012885541655123234, + 0.00576769607141614, + 0.0026747111696749926, + 0.0075287348590791225, + 0.003255660180002451, + 0.00501521211117506, + 0.0040612914599478245, + 0.0039960420690476894, + 0.007693731691688299, + 0.009983657859265804, + 0.009221856482326984, + 0.012277694419026375, + 0.005726238712668419, + 0.008944857865571976, + 0.009779178537428379, + 0.006540353409945965, + 0.011739687994122505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.08911655843257904 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204526126384735 + ], + "opposite_logprob_tokens": [ + -2.8120453357696533 + ], + "logprob_reward": -0.06204526126384735, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204526126384735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12977872788906097, + 0.0610201470553875, + 0.06868643313646317, + 0.15384499728679657, + 0.11942840367555618, + 0.12246375530958176, + 0.053848400712013245, + 0.167870432138443, + 0.05767495185136795, + 0.06538379192352295 + ], + "credit_map": [ + 0.01804380863904953, + 0.013615037314593792, + 0.020429274067282677, + 0.01401236280798912, + 0.014392506331205368, + 0.01254034973680973, + 0.017442336305975914, + 0.01930304802954197, + 0.004108596593141556, + 0.004649692215025425, + 0.011081459932029247, + 0.006055500358343124, + 0.005433367099612951, + 0.012392123229801655, + 0.009324243292212486, + 0.007975160144269466, + 0.005389893427491188, + 0.0068608918227255344, + 0.01892920210957527, + 0.005054563283920288, + 0.013508603908121586, + 0.008131003938615322, + 0.005466411355882883, + 0.005345854442566633, + 0.01704859361052513, + 0.0071220360696315765, + 0.039776671677827835, + 0.017137710005044937, + 0.00936310738325119, + 0.017502978444099426, + 0.017543962225317955, + 0.02834993228316307, + 0.011151266284286976, + 0.009882433339953423, + 0.019421657547354698, + 0.024828840047121048, + 0.007063542027026415, + 0.008964819833636284, + 0.023201733827590942, + 0.014914102852344513, + 0.00961630791425705, + 0.009265999309718609, + 0.020131170749664307, + 0.04099886864423752, + 0.008242778480052948, + 0.008901305496692657, + 0.01028327364474535, + 0.015024052001535892, + 0.003928355872631073, + 0.004193664062768221, + 0.009522074833512306, + 0.010369733907282352, + 0.004514205735176802, + 0.006370641756802797, + 0.005837942007929087, + 0.009111781604588032, + 0.017558902502059937, + 0.013023189269006252, + 0.02531576342880726, + 0.040567703545093536, + 0.02089824713766575, + 0.014387906529009342, + 0.014847378246486187, + 0.021271321922540665, + 0.007781713269650936, + 0.004167607519775629, + 0.010727539658546448, + 0.005828185006976128, + 0.004939918406307697, + 0.009048647247254848, + 0.005669221747666597, + 0.009512113407254219, + 0.00807871576398611, + 0.008480321615934372, + 0.010476759634912014, + 0.005907884798943996, + 0.007701197639107704, + 0.009041673503816128, + 0.005649985745549202, + 0.010047252289950848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.06204526126384735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193756565451622 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193756565451622, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193756565451622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08284343034029007, + 0.10483714193105698, + 0.08319937437772751, + 0.1043243333697319, + 0.06388495862483978, + 0.1063082367181778, + 0.06654857099056244, + 0.24207811057567596, + 0.04881812632083893, + 0.09715767949819565 + ], + "credit_map": [ + 0.011341939680278301, + 0.007999692112207413, + 0.01576181687414646, + 0.008262710645794868, + 0.009716524742543697, + 0.009039531461894512, + 0.010223188437521458, + 0.010498030111193657, + 0.012191642075777054, + 0.007480820640921593, + 0.02927534654736519, + 0.008614091202616692, + 0.008748218417167664, + 0.014414769597351551, + 0.014343667775392532, + 0.00976858101785183, + 0.008521695621311665, + 0.005665099713951349, + 0.02384462207555771, + 0.008010792545974255, + 0.005729825235903263, + 0.009878085926175117, + 0.006817383226007223, + 0.014731873758137226, + 0.009267207235097885, + 0.007696706335991621, + 0.025004303082823753, + 0.009837714955210686, + 0.010792414657771587, + 0.014643479138612747, + 0.00916994921863079, + 0.01791256107389927, + 0.005386544391512871, + 0.0034842975437641144, + 0.012986001558601856, + 0.008277187123894691, + 0.005834345705807209, + 0.005219612270593643, + 0.010127793997526169, + 0.012569175101816654, + 0.009291674941778183, + 0.005374799482524395, + 0.017740000039339066, + 0.03811754658818245, + 0.005687518510967493, + 0.0068786730989813805, + 0.008729159832000732, + 0.014488857239484787, + 0.006239897105842829, + 0.0045617069117724895, + 0.012761360965669155, + 0.013287262059748173, + 0.005149406846612692, + 0.006532771047204733, + 0.007553485222160816, + 0.010462675243616104, + 0.030317306518554688, + 0.01711318828165531, + 0.042038287967443466, + 0.041962917894124985, + 0.017627183347940445, + 0.024879680946469307, + 0.01623467728495598, + 0.05190487205982208, + 0.006666963454335928, + 0.0038135643117129803, + 0.008561437949538231, + 0.005339504685252905, + 0.004885803442448378, + 0.0056721544824540615, + 0.004895621445029974, + 0.008983079344034195, + 0.0099095543846488, + 0.011369938962161541, + 0.01728430576622486, + 0.010280580259859562, + 0.011189456097781658, + 0.010098139755427837, + 0.011091315187513828, + 0.01593438722193241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.003193756565451622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036149173974990845 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036149173974990845, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036149173974990845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1405719667673111, + 0.18503662943840027, + 0.0805310383439064, + 0.07728271931409836, + 0.0366196408867836, + 0.13475188612937927, + 0.07592899352312088, + 0.09773906320333481, + 0.09696898609399796, + 0.07456912845373154 + ], + "credit_map": [ + 0.019386062398552895, + 0.014237894676625729, + 0.022483637556433678, + 0.017334796488285065, + 0.014969722367823124, + 0.01380822155624628, + 0.02351119928061962, + 0.014840427786111832, + 0.010194230824708939, + 0.014547765254974365, + 0.03405258059501648, + 0.01255389116704464, + 0.020509837195277214, + 0.051707979291677475, + 0.020142193883657455, + 0.02132813259959221, + 0.008143951185047626, + 0.004675337113440037, + 0.027226289734244347, + 0.005102729424834251, + 0.004981412086635828, + 0.01046271063387394, + 0.004633869044482708, + 0.01530473679304123, + 0.00924696959555149, + 0.004096159245818853, + 0.015496762469410896, + 0.006545636337250471, + 0.005681958049535751, + 0.008634353056550026, + 0.007732970640063286, + 0.01984790712594986, + 0.003460125532001257, + 0.0031580461654812098, + 0.00683983089402318, + 0.004974563140422106, + 0.0039854831993579865, + 0.0030952407978475094, + 0.004515137989073992, + 0.006591211538761854, + 0.011665210127830505, + 0.006918466649949551, + 0.02557114139199257, + 0.045209046453237534, + 0.007663239724934101, + 0.007453486323356628, + 0.010386775247752666, + 0.019884519279003143, + 0.00558047741651535, + 0.00558856874704361, + 0.014688914641737938, + 0.018647031858563423, + 0.004953300580382347, + 0.007254412397742271, + 0.008422515355050564, + 0.010793771594762802, + 0.011930642649531364, + 0.007220798637717962, + 0.015751391649246216, + 0.02374780923128128, + 0.008398146368563175, + 0.007425706833600998, + 0.007548386696726084, + 0.015716183930635452, + 0.011545340530574322, + 0.007503679487854242, + 0.016131699085235596, + 0.0114845409989357, + 0.008252300322055817, + 0.01637672819197178, + 0.01077846996486187, + 0.01489623636007309, + 0.008257152512669563, + 0.005940825212746859, + 0.012653712183237076, + 0.008916349150240421, + 0.009362773969769478, + 0.011858480982482433, + 0.0063831862062215805, + 0.011196651495993137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0036149173974990845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up immerses the viewer in the dark, humid incubation chamber, focusing with exquisite clarity on a single, perfect oyster sprouting from a shimmering, dark bottom pocket. The entire scene is visualized with the meticulous, meticulous precision of fine brushwork and vibrant mineral pigments, each detail rendered with a surreal yet stylized elegance. The mushroom's pearlescent fan-shaped cap and delicate, paper-thin gills are defined by sharp, calligraphic lines. A quiet, intense fire begins to spread from the base of the mushroom's stem, its small orange flames casting a warm, flickering glow, creating deep, dancing shadows within the gills. As the relentless heat causes the delicate edges of the gills to first shrivel, then slowly curl inward and begin to char, the transformation unfolds like a writhing ink painting, from pristine white to a fragile, deep black. These colors, from the fiery crimson of the flames to the eroded charcoal, possess the saturated vibrancy of mineral pigments, creating a stark contrast that underscores the silent, destructive process. The atmosphere is one of concentrated, tragic beauty, capturing the elegant consumption of a perfect natural form as thin wisps of smoke rise in delicate, calligraphic curls against a dark, damp background.", + "seed": 11237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_4210_seed_11237_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04250885080546141, + "vlm_logprob": -0.04250885080546141, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.04250885080546141, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01743970811367035 + ], + "opposite_logprob_tokens": [ + -4.892439842224121 + ], + "logprob_reward": -0.01743970811367035, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01743970811367035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1650882214307785, + 0.06627191603183746, + 0.18781404197216034, + 0.03993993625044823, + 0.14811627566814423, + 0.04629996791481972, + 0.07615448534488678, + 0.13872437179088593, + 0.069637730717659, + 0.061952996999025345 + ], + "credit_map": [ + 0.026391619816422462, + 0.013515002094209194, + 0.02783222682774067, + 0.022206265479326248, + 0.014930847100913525, + 0.015770576894283295, + 0.021358270198106766, + 0.023083418607711792, + 0.00909577775746584, + 0.00899222120642662, + 0.011185222305357456, + 0.007571527734398842, + 0.006467729341238737, + 0.005919757764786482, + 0.006961306557059288, + 0.010078376159071922, + 0.021850578486919403, + 0.022543542087078094, + 0.029903678223490715, + 0.017820090055465698, + 0.017607431858778, + 0.014851248823106289, + 0.03916788846254349, + 0.024069607257843018, + 0.004116464871913195, + 0.0031020448077470064, + 0.007706866599619389, + 0.004028897266834974, + 0.0039016513619571924, + 0.0045539638958871365, + 0.008234284818172455, + 0.004295766353607178, + 0.018171779811382294, + 0.011937922798097134, + 0.028806235641241074, + 0.017448220402002335, + 0.011365021578967571, + 0.014145948924124241, + 0.022457631304860115, + 0.02378353849053383, + 0.005090324208140373, + 0.006077878177165985, + 0.0065644290298223495, + 0.004683755803853273, + 0.0068648820742964745, + 0.005440196022391319, + 0.00559248449280858, + 0.005986018106341362, + 0.0063805487006902695, + 0.010732513852417469, + 0.017718274146318436, + 0.0063429782167077065, + 0.008600711822509766, + 0.005278220400214195, + 0.006706817541271448, + 0.014394432306289673, + 0.015027628280222416, + 0.011701153591275215, + 0.03512170538306236, + 0.015573466196656227, + 0.012955749407410622, + 0.014758494682610035, + 0.009774474427103996, + 0.023811714723706245, + 0.008487359620630741, + 0.006952793803066015, + 0.015902450308203697, + 0.006766028702259064, + 0.008447687141597271, + 0.004876685794442892, + 0.006146847270429134, + 0.01205788366496563, + 0.006400920916348696, + 0.008364176377654076, + 0.009301753714680672, + 0.008810963481664658, + 0.009907516650855541, + 0.006498263217508793, + 0.006501807831227779, + 0.006167597137391567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.01743970811367035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1800057590007782 + ], + "opposite_logprob_tokens": [ + -1.8050057888031006 + ], + "logprob_reward": -0.1800057590007782, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1800057590007782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32109540700912476, + 0.156339630484581, + 0.11144950240850449, + 0.04279359057545662, + 0.10795332491397858, + 0.06430148333311081, + 0.04546886309981346, + 0.052991293370723724, + 0.04986746609210968, + 0.047739505767822266 + ], + "credit_map": [ + 0.047101229429244995, + 0.02627110853791237, + 0.06392966210842133, + 0.02741435170173645, + 0.032425832003355026, + 0.02597590908408165, + 0.048880282789468765, + 0.04909702017903328, + 0.02024763822555542, + 0.01423563901335001, + 0.030923712998628616, + 0.013032706454396248, + 0.0193663127720356, + 0.015399658121168613, + 0.01858890801668167, + 0.02454504184424877, + 0.013762271031737328, + 0.011798669584095478, + 0.020285755395889282, + 0.012032353319227695, + 0.013316615484654903, + 0.012089760974049568, + 0.011915578506886959, + 0.016248485073447227, + 0.004416562616825104, + 0.004354842938482761, + 0.0065619987435638905, + 0.004209415055811405, + 0.005123973358422518, + 0.007287328597158194, + 0.005798777099698782, + 0.005040684249252081, + 0.01089383289217949, + 0.009356485679745674, + 0.021503834053874016, + 0.014132222160696983, + 0.014620264060795307, + 0.010826749727129936, + 0.009940681979060173, + 0.016679249703884125, + 0.006450931075960398, + 0.008755835704505444, + 0.010305410251021385, + 0.006283131893724203, + 0.007355121895670891, + 0.009346379898488522, + 0.008384418673813343, + 0.007420254405587912, + 0.007384550292044878, + 0.005284459330141544, + 0.006799890659749508, + 0.004379708785563707, + 0.005622140597552061, + 0.0045723202638328075, + 0.0051658558659255505, + 0.00625993125140667, + 0.007766647730022669, + 0.004241881426423788, + 0.011959390714764595, + 0.004723915830254555, + 0.005601240787655115, + 0.005569811444729567, + 0.004591052886098623, + 0.008537347428500652, + 0.005323277786374092, + 0.004153098911046982, + 0.010228583589196205, + 0.005140823777765036, + 0.006845136638730764, + 0.005070140585303307, + 0.004762921016663313, + 0.008343478664755821, + 0.006517394911497831, + 0.005556529853492975, + 0.007530456408858299, + 0.006397262215614319, + 0.005990275647491217, + 0.005941851530224085, + 0.0043546720407903194, + 0.005451057571917772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1800057590007782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639005314558744 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007639005314558744, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639005314558744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09997131675481796, + 0.11986110359430313, + 0.3027886748313904, + 0.037265460938215256, + 0.16949741542339325, + 0.06443564593791962, + 0.04287631809711456, + 0.04636094346642494, + 0.03865525498986244, + 0.07828785479068756 + ], + "credit_map": [ + 0.012586936354637146, + 0.011284287087619305, + 0.01844591088593006, + 0.009402269497513771, + 0.011437845416367054, + 0.010946553200483322, + 0.01456969790160656, + 0.01129781175404787, + 0.012382027693092823, + 0.013553071767091751, + 0.01765882782638073, + 0.011726773343980312, + 0.018978167325258255, + 0.014591634273529053, + 0.01836107298731804, + 0.012609533965587616, + 0.032398100942373276, + 0.03400459885597229, + 0.045828670263290405, + 0.029592309147119522, + 0.029248574748635292, + 0.02489076368510723, + 0.07011328637599945, + 0.036712370812892914, + 0.005137118510901928, + 0.003040672978386283, + 0.007966727949678898, + 0.0033973415847867727, + 0.002970271511003375, + 0.003718961263075471, + 0.006471047643572092, + 0.004563317634165287, + 0.018036680296063423, + 0.015978075563907623, + 0.028021400794386864, + 0.022946368902921677, + 0.021137496456503868, + 0.018443742766976357, + 0.022570980712771416, + 0.02236267738044262, + 0.0070725237019360065, + 0.00813904870301485, + 0.010521620512008667, + 0.006689935922622681, + 0.008551212958991528, + 0.009119450114667416, + 0.006046741269528866, + 0.008295107632875443, + 0.007543491665273905, + 0.005309919361025095, + 0.00702426303178072, + 0.0036983778700232506, + 0.005379684269428253, + 0.004635239485651255, + 0.0037903934717178345, + 0.0054949508048594, + 0.004838996566832066, + 0.0039768582209944725, + 0.009461816400289536, + 0.005681151058524847, + 0.004999331198632717, + 0.004760943353176117, + 0.00589307677000761, + 0.006748769897967577, + 0.004818282555788755, + 0.0035835138987749815, + 0.006641759071499109, + 0.003956052009016275, + 0.004111821763217449, + 0.005914806388318539, + 0.003454222809523344, + 0.0061747971922159195, + 0.011052262969315052, + 0.007094452157616615, + 0.015040489844977856, + 0.008724494837224483, + 0.011929811909794807, + 0.0065909563563764095, + 0.009486420080065727, + 0.008368962444365025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.007639005314558744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820206806063652 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820206806063652, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820206806063652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059007033705711365, + 0.043285612016916275, + 0.2029595524072647, + 0.03769947960972786, + 0.098658986389637, + 0.09219072759151459, + 0.06854724138975143, + 0.06251165270805359, + 0.08946720510721207, + 0.24567250907421112 + ], + "credit_map": [ + 0.00832206942141056, + 0.007417559158056974, + 0.011033914051949978, + 0.0052863662131130695, + 0.007158010266721249, + 0.005728260148316622, + 0.0067503429017961025, + 0.0073105148039758205, + 0.006047563161700964, + 0.005706749390810728, + 0.006483407225459814, + 0.003975544590502977, + 0.005693493876606226, + 0.004717167466878891, + 0.00579014653339982, + 0.004871542565524578, + 0.027265755459666252, + 0.023635558784008026, + 0.03697965666651726, + 0.01719469577074051, + 0.016207315027713776, + 0.017429599538445473, + 0.03824112191796303, + 0.026005852967500687, + 0.0046244594268500805, + 0.00342003651894629, + 0.006822468712925911, + 0.004387343302369118, + 0.0038362755440175533, + 0.00526822917163372, + 0.004107970744371414, + 0.005232696421444416, + 0.010358486324548721, + 0.008060408756136894, + 0.01981952227652073, + 0.01092554908245802, + 0.011326303705573082, + 0.010972210206091404, + 0.01414785347878933, + 0.013048652559518814, + 0.010093394666910172, + 0.007636786438524723, + 0.012538726441562176, + 0.008946285583078861, + 0.014804561622440815, + 0.010404196567833424, + 0.016258113086223602, + 0.011508665047585964, + 0.010772379115223885, + 0.00859627965837717, + 0.010629297234117985, + 0.00623526144772768, + 0.00912204198539257, + 0.00789991021156311, + 0.006752764340490103, + 0.008539309725165367, + 0.005040090065449476, + 0.004602039232850075, + 0.01149081438779831, + 0.006586545612663031, + 0.007101061753928661, + 0.008328155614435673, + 0.008494743146002293, + 0.01086820475757122, + 0.01107370387762785, + 0.0071092951111495495, + 0.0176133681088686, + 0.010224900208413601, + 0.008511506021022797, + 0.012334542348980904, + 0.007282667327672243, + 0.015317227691411972, + 0.036282651126384735, + 0.024926992133259773, + 0.06188524514436722, + 0.01664879359304905, + 0.03055962547659874, + 0.021317705512046814, + 0.023260492831468582, + 0.030790982767939568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.002820206806063652 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463957479223609 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.00463957479223609, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463957479223609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12737417221069336, + 0.1055411696434021, + 0.22454515099525452, + 0.050942592322826385, + 0.07095521688461304, + 0.09753744304180145, + 0.06096924468874931, + 0.05641390010714531, + 0.0899338573217392, + 0.11578726023435593 + ], + "credit_map": [ + 0.01783672347664833, + 0.012693451717495918, + 0.02313109301030636, + 0.01352450717240572, + 0.013473761267960072, + 0.013410218060016632, + 0.017601029947400093, + 0.015703387558460236, + 0.010559563525021076, + 0.012030571699142456, + 0.015826214104890823, + 0.011934617534279823, + 0.01480778492987156, + 0.010898935608565807, + 0.016988536342978477, + 0.012494947761297226, + 0.034519340842962265, + 0.022459479048848152, + 0.051773473620414734, + 0.02102695219218731, + 0.016751307994127274, + 0.017405793070793152, + 0.03476662188768387, + 0.02584218420088291, + 0.0053747957572340965, + 0.00497708423063159, + 0.008144800551235676, + 0.005637074820697308, + 0.005524202715605497, + 0.007558044977486134, + 0.007344391196966171, + 0.00638220040127635, + 0.006850813049823046, + 0.006649382412433624, + 0.012929342687129974, + 0.00927401427179575, + 0.007346692029386759, + 0.006898555438965559, + 0.00978199765086174, + 0.011224412359297276, + 0.01291587669402361, + 0.008943379856646061, + 0.014999198727309704, + 0.009746845811605453, + 0.014656738378107548, + 0.010241908021271229, + 0.014802348800003529, + 0.01123114861547947, + 0.01144736260175705, + 0.006831157486885786, + 0.008855784311890602, + 0.005003965459764004, + 0.007883396930992603, + 0.006186159793287516, + 0.005761185195297003, + 0.009000237099826336, + 0.007555647753179073, + 0.003987244330346584, + 0.013232138007879257, + 0.005431200843304396, + 0.004879416432231665, + 0.005156456958502531, + 0.007307390682399273, + 0.008864403702318668, + 0.009580940939486027, + 0.008928608149290085, + 0.015414189547300339, + 0.008901233784854412, + 0.012047470547258854, + 0.010591674596071243, + 0.009806763380765915, + 0.014662979170680046, + 0.014667131006717682, + 0.01460678968578577, + 0.021367764100432396, + 0.016406767070293427, + 0.011296943761408329, + 0.01439575757831335, + 0.011595213785767555, + 0.011450890451669693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.00463957479223609 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up immerses the viewer in the dark, humid incubation chamber, focusing with exquisite clarity on a single, perfect oyster sprouting from a shimmering, dark bottom pocket. The entire scene is visualized with the meticulous, meticulous precision of fine brushwork and vibrant mineral pigments, each detail rendered with a surreal yet stylized elegance. The mushroom's pearlescent fan-shaped cap and delicate, paper-thin gills are defined by sharp, calligraphic lines. A quiet, intense fire begins to spread from the base of the mushroom's stem, its small orange flames casting a warm, flickering glow, creating deep, dancing shadows within the gills. As the relentless heat causes the delicate edges of the gills to first shrivel, then slowly curl inward and begin to char, the transformation unfolds like a writhing ink painting, from pristine white to a fragile, deep black. These colors, from the fiery crimson of the flames to the eroded charcoal, possess the saturated vibrancy of mineral pigments, creating a stark contrast that underscores the silent, destructive process. The atmosphere is one of concentrated, tragic beauty, capturing the elegant consumption of a perfect natural form as thin wisps of smoke rise in delicate, calligraphic curls against a dark, damp background.", + "seed": 11238, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_4210_seed_11238_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05201000953093171, + "vlm_logprob": -0.05201000953093171, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.05201000953093171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023720765486359596 + ], + "opposite_logprob_tokens": [ + -4.148720741271973 + ], + "logprob_reward": -0.023720765486359596, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.023720765486359596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15370860695838928, + 0.07746917754411697, + 0.2551431953907013, + 0.1581532210111618, + 0.0805022194981575, + 0.09055186808109283, + 0.022274620831012726, + 0.04030440002679825, + 0.08156465739011765, + 0.04032811149954796 + ], + "credit_map": [ + 0.01705232448875904, + 0.01879723183810711, + 0.015392369590699673, + 0.016086233779788017, + 0.02111058682203293, + 0.021220693364739418, + 0.023674489930272102, + 0.020374685525894165, + 0.005752605851739645, + 0.011007731780409813, + 0.005912773311138153, + 0.009383128024637699, + 0.009945848025381565, + 0.013388399966061115, + 0.01510593295097351, + 0.006972754839807749, + 0.03673459589481354, + 0.022812461480498314, + 0.06838417053222656, + 0.02049616351723671, + 0.043802179396152496, + 0.016464123502373695, + 0.02295696921646595, + 0.02349252998828888, + 0.011098156683146954, + 0.02073507569730282, + 0.023456117138266563, + 0.020787972956895828, + 0.01866769604384899, + 0.013632146641612053, + 0.026232456788420677, + 0.023543596267700195, + 0.006140145938843489, + 0.0061969817616045475, + 0.014040088281035423, + 0.007934012450277805, + 0.009379505179822445, + 0.005012231878936291, + 0.02078358829021454, + 0.011015670374035835, + 0.006391127128154039, + 0.007066359743475914, + 0.012918536551296711, + 0.012303312309086323, + 0.011819646693766117, + 0.005257925949990749, + 0.02560397796332836, + 0.009190982207655907, + 0.0022537217009812593, + 0.002170874970033765, + 0.0030284603126347065, + 0.002539257053285837, + 0.0035966800060123205, + 0.001998417777940631, + 0.00302587216719985, + 0.003661336610093713, + 0.0031097473111003637, + 0.004315864760428667, + 0.007596094626933336, + 0.0032991005573421717, + 0.0062694973312318325, + 0.002518574707210064, + 0.004645640030503273, + 0.008549878373742104, + 0.005618589464575052, + 0.013720919378101826, + 0.014120321720838547, + 0.008683007210493088, + 0.01293200347572565, + 0.0038150434847921133, + 0.008699527941644192, + 0.013975243084132671, + 0.0033820457756519318, + 0.003467451548203826, + 0.009041978977620602, + 0.004473500419408083, + 0.00281685683876276, + 0.002982339821755886, + 0.004813727457076311, + 0.009350212290883064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.023720765486359596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20165890455245972 + ], + "opposite_logprob_tokens": [ + -1.7016589641571045 + ], + "logprob_reward": -0.20165890455245972, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20165890455245972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14109587669372559, + 0.032045867294073105, + 0.0954078882932663, + 0.05168488621711731, + 0.09198368340730667, + 0.05356226861476898, + 0.022315897047519684, + 0.07370810210704803, + 0.34486961364746094, + 0.09332598000764847 + ], + "credit_map": [ + 0.014286925084888935, + 0.01524338498711586, + 0.013780707493424416, + 0.014789569191634655, + 0.019767997786402702, + 0.021423829719424248, + 0.023343142122030258, + 0.01846032403409481, + 0.002594469580799341, + 0.0059778341092169285, + 0.0044372109696269035, + 0.0034694853238761425, + 0.004133121110498905, + 0.004097427241504192, + 0.003960661590099335, + 0.003375658765435219, + 0.009313314221799374, + 0.010423705913126469, + 0.018243802711367607, + 0.011266353540122509, + 0.016311578452587128, + 0.008587243035435677, + 0.013161749579012394, + 0.008100141771137714, + 0.00413499167189002, + 0.005962186958640814, + 0.007381503935903311, + 0.006135885138064623, + 0.005721810273826122, + 0.004631293471902609, + 0.01133902370929718, + 0.006378191988915205, + 0.00866356398910284, + 0.007765129208564758, + 0.012483926489949226, + 0.008051241748034954, + 0.010427840054035187, + 0.006722175050526857, + 0.022139932960271835, + 0.0157298780977726, + 0.006165896542370319, + 0.007019617594778538, + 0.007978085428476334, + 0.00527263805270195, + 0.006200889591127634, + 0.004986903630197048, + 0.008010034449398518, + 0.007928203791379929, + 0.0019838858861476183, + 0.002196198096498847, + 0.0032237833365797997, + 0.0038713885005563498, + 0.002359433099627495, + 0.0020288859959691763, + 0.0034941767808049917, + 0.003158144187182188, + 0.0064493161626160145, + 0.0075017958879470825, + 0.013359062373638153, + 0.00883074663579464, + 0.008683468215167522, + 0.006163728889077902, + 0.010768376290798187, + 0.011951606720685959, + 0.033830467611551285, + 0.05336301028728485, + 0.0513448603451252, + 0.03874911740422249, + 0.05019718036055565, + 0.02818840555846691, + 0.03765740618109703, + 0.05153915286064148, + 0.007178960368037224, + 0.005902980454266071, + 0.01597963273525238, + 0.012523354962468147, + 0.00879200641065836, + 0.006968827452510595, + 0.020820563659071922, + 0.015159658156335354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.20165890455245972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293664678931236 + ], + "opposite_logprob_tokens": [ + -3.7732937335968018 + ], + "logprob_reward": -0.023293664678931236, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293664678931236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09560909867286682, + 0.08002510666847229, + 0.09154058992862701, + 0.0636146292090416, + 0.09591849148273468, + 0.04779263213276863, + 0.03491075336933136, + 0.18607445061206818, + 0.25838229060173035, + 0.04613187909126282 + ], + "credit_map": [ + 0.011498132720589638, + 0.011646358296275139, + 0.009038995951414108, + 0.00917503610253334, + 0.01254064217209816, + 0.015022947452962399, + 0.01542699709534645, + 0.011260000057518482, + 0.007795535493642092, + 0.01515956874936819, + 0.009106656536459923, + 0.009173846803605556, + 0.010635549202561378, + 0.009288628585636616, + 0.00880061648786068, + 0.010064715519547462, + 0.008908355608582497, + 0.009497425518929958, + 0.018245430663228035, + 0.008408789522945881, + 0.01747979037463665, + 0.007808331400156021, + 0.011355645954608917, + 0.009836830198764801, + 0.004499942064285278, + 0.008165603503584862, + 0.008589591830968857, + 0.007620632648468018, + 0.007443911395967007, + 0.005839333403855562, + 0.015359430573880672, + 0.006096192169934511, + 0.010164918377995491, + 0.010052354075014591, + 0.012752764858305454, + 0.009955840185284615, + 0.012424953281879425, + 0.007612451910972595, + 0.017853420227766037, + 0.015101796016097069, + 0.006317087449133396, + 0.0060418155044317245, + 0.006283164955675602, + 0.0052996063604950905, + 0.00660987664014101, + 0.0039094737730920315, + 0.005735504440963268, + 0.007596106734126806, + 0.00350492843426764, + 0.0049896277487277985, + 0.004910054616630077, + 0.005959655158221722, + 0.004592211451381445, + 0.0027944189496338367, + 0.004093638155609369, + 0.0040662228129804134, + 0.014214079827070236, + 0.023063957691192627, + 0.029454898089170456, + 0.018105866387486458, + 0.0272125992923975, + 0.013868295587599277, + 0.028800953179597855, + 0.03135381266474724, + 0.02163495309650898, + 0.045884814113378525, + 0.036650191992521286, + 0.028613753616809845, + 0.046648602932691574, + 0.01951068826019764, + 0.02645779773592949, + 0.0329815149307251, + 0.0026494376361370087, + 0.0026246702764183283, + 0.014210882596671581, + 0.0038353868294507265, + 0.0029190294444561005, + 0.0027561774477362633, + 0.003625372424721718, + 0.013510925695300102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.023293664678931236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640405531972647 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640405531972647, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640405531972647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09208311140537262, + 0.0636197179555893, + 0.11978461593389511, + 0.042101845145225525, + 0.06689085811376572, + 0.06157094985246658, + 0.036677226424217224, + 0.08623934537172318, + 0.06346018612384796, + 0.36757218837738037 + ], + "credit_map": [ + 0.01201708521693945, + 0.0079704774543643, + 0.0126938596367836, + 0.009929222986102104, + 0.011211402714252472, + 0.013848035596311092, + 0.012987353838980198, + 0.011425672098994255, + 0.004663861822336912, + 0.010085063055157661, + 0.0063285911455750465, + 0.007140244357287884, + 0.01038005854934454, + 0.007403539959341288, + 0.011780232191085815, + 0.005838126875460148, + 0.015958767384290695, + 0.009823272936046124, + 0.03195464238524437, + 0.009866210632026196, + 0.024307498708367348, + 0.007626072969287634, + 0.01042358297854662, + 0.009824576787650585, + 0.0034110229462385178, + 0.005829595495015383, + 0.005904691759496927, + 0.004932005424052477, + 0.006647932808846235, + 0.004182131960988045, + 0.007935190573334694, + 0.00325927441008389, + 0.0053026918321847916, + 0.009984851814806461, + 0.011481552384793758, + 0.00786587968468666, + 0.007820830680429935, + 0.007790874689817429, + 0.009430211037397385, + 0.007213965989649296, + 0.007382099516689777, + 0.0062025804072618484, + 0.00988224521279335, + 0.007082873955368996, + 0.0076052164658904076, + 0.004961871076375246, + 0.009424480609595776, + 0.00902958121150732, + 0.003955895081162453, + 0.005256536416709423, + 0.004991529043763876, + 0.005482776090502739, + 0.004925708286464214, + 0.0025454952847212553, + 0.005574177019298077, + 0.003945107106119394, + 0.005865190643817186, + 0.010130632668733597, + 0.017575761303305626, + 0.011348302476108074, + 0.011212979443371296, + 0.0050143348053097725, + 0.008439122699201107, + 0.016653025522828102, + 0.005100508686155081, + 0.007102934177964926, + 0.012049461714923382, + 0.007974418811500072, + 0.0072425734251737595, + 0.004944482818245888, + 0.008159277960658073, + 0.010886533185839653, + 0.023818714544177055, + 0.02351546473801136, + 0.10373373329639435, + 0.026636961847543716, + 0.024028681218624115, + 0.019297201186418533, + 0.039530638605356216, + 0.10701078921556473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.004640405531972647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673630740493536 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.00673630740493536, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673630740493536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09947419166564941, + 0.04121655598282814, + 0.06519447267055511, + 0.0935245156288147, + 0.06100333109498024, + 0.03689569979906082, + 0.030983278527855873, + 0.07491346448659897, + 0.30077382922172546, + 0.19602063298225403 + ], + "credit_map": [ + 0.011467400938272476, + 0.014414732344448566, + 0.009608194231987, + 0.01036852691322565, + 0.013734308071434498, + 0.01668519154191017, + 0.01273881085216999, + 0.010457021184265614, + 0.003617245936766267, + 0.00817172322422266, + 0.005342948250472546, + 0.004719430580735207, + 0.005457815248519182, + 0.005475458689033985, + 0.004843288101255894, + 0.003588646650314331, + 0.00707973213866353, + 0.006015151273459196, + 0.014707203023135662, + 0.006511639803647995, + 0.013874453492462635, + 0.00511792441830039, + 0.007066020276397467, + 0.004822347778826952, + 0.006591609679162502, + 0.01089442428201437, + 0.013633649796247482, + 0.012783325277268887, + 0.009926499798893929, + 0.00871575903147459, + 0.01949048414826393, + 0.011488762684166431, + 0.005766903515905142, + 0.005255996715277433, + 0.010613104328513145, + 0.005196780432015657, + 0.007971681654453278, + 0.005723598878830671, + 0.007529745809733868, + 0.012945524416863918, + 0.0037488264497369528, + 0.0044220490381121635, + 0.004929800983518362, + 0.0038363924250006676, + 0.0052400194108486176, + 0.0038877269253134727, + 0.005424979608505964, + 0.005405906122177839, + 0.0027180942706763744, + 0.0035227788612246513, + 0.005233832634985447, + 0.003733453806489706, + 0.0037972452118992805, + 0.0021767241414636374, + 0.005059561226516962, + 0.004741588141769171, + 0.006909260526299477, + 0.010460346937179565, + 0.010033680126070976, + 0.0071008652448654175, + 0.012504027225077152, + 0.005995618645101786, + 0.010976282879710197, + 0.010933379642665386, + 0.028751321136951447, + 0.056460894644260406, + 0.03861111402511597, + 0.02951716259121895, + 0.05377459153532982, + 0.025804435834288597, + 0.031377583742141724, + 0.036476727575063705, + 0.013385872356593609, + 0.00853111781179905, + 0.053145699203014374, + 0.012666392140090466, + 0.010801222175359726, + 0.011624529026448727, + 0.03325487673282623, + 0.05261093005537987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.00673630740493536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up immerses the viewer in the dark, humid incubation chamber, focusing with exquisite clarity on a single, perfect oyster sprouting from a shimmering, dark bottom pocket. The entire scene is visualized with the meticulous, meticulous precision of fine brushwork and vibrant mineral pigments, each detail rendered with a surreal yet stylized elegance. The mushroom's pearlescent fan-shaped cap and delicate, paper-thin gills are defined by sharp, calligraphic lines. A quiet, intense fire begins to spread from the base of the mushroom's stem, its small orange flames casting a warm, flickering glow, creating deep, dancing shadows within the gills. As the relentless heat causes the delicate edges of the gills to first shrivel, then slowly curl inward and begin to char, the transformation unfolds like a writhing ink painting, from pristine white to a fragile, deep black. These colors, from the fiery crimson of the flames to the eroded charcoal, possess the saturated vibrancy of mineral pigments, creating a stark contrast that underscores the silent, destructive process. The atmosphere is one of concentrated, tragic beauty, capturing the elegant consumption of a perfect natural form as thin wisps of smoke rise in delicate, calligraphic curls against a dark, damp background.", + "seed": 11239, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_6/index_4210_seed_11239_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03335385276004672, + "vlm_logprob": -0.03335385276004672, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03335385276004672, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017447207123041153 + ], + "opposite_logprob_tokens": [ + -4.892446994781494 + ], + "logprob_reward": -0.017447207123041153, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017447207123041153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1911238431930542, + 0.11272576451301575, + 0.06807274371385574, + 0.05852673947811127, + 0.2187279313802719, + 0.1671987920999527, + 0.02811850607395172, + 0.08932408690452576, + 0.028123781085014343, + 0.03805778920650482 + ], + "credit_map": [ + 0.026018254458904266, + 0.015382097102701664, + 0.033751703798770905, + 0.026871321722865105, + 0.01891307532787323, + 0.014276089146733284, + 0.027958523482084274, + 0.027952784672379494, + 0.009484008885920048, + 0.011995785869657993, + 0.017752571031451225, + 0.01395768765360117, + 0.015180077403783798, + 0.009489643387496471, + 0.019710833206772804, + 0.015155162662267685, + 0.005739178042858839, + 0.006710559129714966, + 0.009947092272341251, + 0.00989439245313406, + 0.008796081878244877, + 0.004674567840993404, + 0.013426326215267181, + 0.008884540759027004, + 0.004564216360449791, + 0.00674912566319108, + 0.009974954649806023, + 0.007329272571951151, + 0.00602959468960762, + 0.005464461632072926, + 0.010811462067067623, + 0.007603651378303766, + 0.012972316704690456, + 0.0227926354855299, + 0.03445711359381676, + 0.024026956409215927, + 0.044553179293870926, + 0.012109434232115746, + 0.04223589971661568, + 0.0255803894251585, + 0.009540090337395668, + 0.01597171276807785, + 0.024262651801109314, + 0.01616450771689415, + 0.03765413910150528, + 0.0184800922870636, + 0.0163167305290699, + 0.028808869421482086, + 0.00341232493519783, + 0.0036660991609096527, + 0.002536339219659567, + 0.0035949188750237226, + 0.0033219989854842424, + 0.004421928431838751, + 0.003703906200826168, + 0.0034609916619956493, + 0.007371711544692516, + 0.008738911710679531, + 0.019018234685063362, + 0.01563352346420288, + 0.00799246970564127, + 0.005111126694828272, + 0.016950268298387527, + 0.00850784219801426, + 0.0026845261454582214, + 0.005385276861488819, + 0.004764793906360865, + 0.0035763890482485294, + 0.0026728773955255747, + 0.0026456317864358425, + 0.004011651501059532, + 0.0023826356045901775, + 0.0031896820291876793, + 0.005844524130225182, + 0.0054953559301793575, + 0.004358974751085043, + 0.0055391318164765835, + 0.004311595112085342, + 0.005043847020715475, + 0.004274678882211447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.017447207123041153 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11302008479833603 + ], + "opposite_logprob_tokens": [ + -2.238020181655884 + ], + "logprob_reward": -0.11302008479833603, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11302008479833603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11345835775136948, + 0.05164334550499916, + 0.0425625778734684, + 0.040480054914951324, + 0.43720316886901855, + 0.12031504511833191, + 0.0465894415974617, + 0.07659119367599487, + 0.0301959365606308, + 0.0409608855843544 + ], + "credit_map": [ + 0.013032271526753902, + 0.010877746157348156, + 0.01976141892373562, + 0.013125642202794552, + 0.009784631431102753, + 0.007790395524352789, + 0.02285725064575672, + 0.01622900553047657, + 0.005319337360560894, + 0.004481288138777018, + 0.010279000736773014, + 0.005925032310187817, + 0.004155748523771763, + 0.004749011248350143, + 0.00853074248880148, + 0.008203186094760895, + 0.0035081494133919477, + 0.004002089612185955, + 0.005491004325449467, + 0.006776227615773678, + 0.004427989479154348, + 0.0041069937869906425, + 0.009040052071213722, + 0.005210077855736017, + 0.0027443855069577694, + 0.00355474092066288, + 0.00967427808791399, + 0.004480825737118721, + 0.004872458521276712, + 0.002574050100520253, + 0.006649483926594257, + 0.005929834675043821, + 0.023692656308412552, + 0.048455722630023956, + 0.06615617871284485, + 0.038184940814971924, + 0.10673677176237106, + 0.023702720180153847, + 0.07480928301811218, + 0.05546490103006363, + 0.0067893704399466515, + 0.013243707828223705, + 0.015206806361675262, + 0.009964967146515846, + 0.031223345547914505, + 0.01260845921933651, + 0.011231173761188984, + 0.020047223195433617, + 0.002752791391685605, + 0.0033782143145799637, + 0.004040849395096302, + 0.003597166622057557, + 0.0037291005719453096, + 0.0033966635819524527, + 0.016503695398569107, + 0.009190963581204414, + 0.005382074508816004, + 0.004700902849435806, + 0.01693784072995186, + 0.01299698743969202, + 0.004997711628675461, + 0.0076927486807107925, + 0.011903360486030579, + 0.011979569680988789, + 0.002534200670197606, + 0.00357536762021482, + 0.004565357230603695, + 0.0036586171481758356, + 0.002681024605408311, + 0.0024354164488613605, + 0.006984965410083532, + 0.0037609899882227182, + 0.004438034724444151, + 0.006490099243819714, + 0.00613004993647337, + 0.006084698252379894, + 0.0046135662123560905, + 0.0035634145606309175, + 0.0043597654439508915, + 0.005281256046146154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.11302008479833603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058861404657364 + ], + "opposite_logprob_tokens": [ + -3.8955886363983154 + ], + "logprob_reward": -0.02058861404657364, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058861404657364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04711252078413963, + 0.027543850243091583, + 0.021850772202014923, + 0.01654096320271492, + 0.7528849244117737, + 0.021946905180811882, + 0.021576473489403725, + 0.04091501235961914, + 0.010228258557617664, + 0.03940039128065109 + ], + "credit_map": [ + 0.005066760815680027, + 0.004723167046904564, + 0.007883354090154171, + 0.0071100350469350815, + 0.004654147662222385, + 0.00325794517993927, + 0.008064956404268742, + 0.006352157797664404, + 0.0024598122108727694, + 0.002930196700617671, + 0.005738336592912674, + 0.003086867742240429, + 0.0029720754828304052, + 0.0021819754038006067, + 0.004660726524889469, + 0.003513859584927559, + 0.001368009834550321, + 0.002589601557701826, + 0.004066920839250088, + 0.0029463835526257753, + 0.0035077054053545, + 0.0014886780409142375, + 0.003612320637330413, + 0.002271153498440981, + 0.0013498484622687101, + 0.0016059083864092827, + 0.0028634616173803806, + 0.0021563968621194363, + 0.0018494923133403063, + 0.0010350269731134176, + 0.003291015513241291, + 0.0023898121435195208, + 0.03924522548913956, + 0.08389613032341003, + 0.10818439722061157, + 0.06069202721118927, + 0.19105671346187592, + 0.04223015531897545, + 0.12573498487472534, + 0.10184523463249207, + 0.0014270299579948187, + 0.00187047163490206, + 0.0044677541591227055, + 0.002730172360315919, + 0.0027891481295228004, + 0.0023245997726917267, + 0.002801334485411644, + 0.0035363943316042423, + 0.001420312561094761, + 0.0011945540318265557, + 0.001507140346802771, + 0.0015896856784820557, + 0.0013410962419584394, + 0.0010564702097326517, + 0.008233155123889446, + 0.005234059877693653, + 0.003736974438652396, + 0.004097446799278259, + 0.004743627272546291, + 0.007389609701931477, + 0.003108516102656722, + 0.004558120854198933, + 0.008006079122424126, + 0.005274641327559948, + 0.0008156922413036227, + 0.001413262216374278, + 0.0013439520262181759, + 0.0013597350334748626, + 0.0010412109550088644, + 0.0013289081398397684, + 0.0017550765769556165, + 0.0011704213684424758, + 0.002982114674523473, + 0.004008318763226271, + 0.006353577133268118, + 0.007721357513219118, + 0.004179987590759993, + 0.0034517135936766863, + 0.004936675541102886, + 0.005766645539551973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.02058861404657364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076156049966812 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011076156049966812, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076156049966812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08044309169054031, + 0.03990587219595909, + 0.06900696456432343, + 0.019914411008358, + 0.23007389903068542, + 0.3471473157405853, + 0.11592024564743042, + 0.05395085737109184, + 0.011549606919288635, + 0.032087694853544235 + ], + "credit_map": [ + 0.017794128507375717, + 0.006321580149233341, + 0.013854825869202614, + 0.007524972781538963, + 0.0044436147436499596, + 0.004854047205299139, + 0.01058816909790039, + 0.015061751939356327, + 0.003537176875397563, + 0.005000120960175991, + 0.004378264304250479, + 0.004151375498622656, + 0.007531981915235519, + 0.0031360783614218235, + 0.007511759176850319, + 0.004659116268157959, + 0.005645853467285633, + 0.00871427170932293, + 0.00685854209586978, + 0.005837602075189352, + 0.014458778314292431, + 0.005317003466188908, + 0.014073369093239307, + 0.008101546205580235, + 0.0010739583522081375, + 0.00257208407856524, + 0.004122900310903788, + 0.0023837171029299498, + 0.002329273149371147, + 0.0016145439585670829, + 0.0032446435652673244, + 0.002573291538283229, + 0.013194035738706589, + 0.0233506690710783, + 0.03830323368310928, + 0.023051723837852478, + 0.04919412359595299, + 0.011994680389761925, + 0.044660620391368866, + 0.026324832811951637, + 0.0227072611451149, + 0.03579212725162506, + 0.05596671998500824, + 0.04098572954535484, + 0.062142856419086456, + 0.033210720866918564, + 0.043798018246889114, + 0.05254386365413666, + 0.010242746211588383, + 0.007745427545160055, + 0.009166126139461994, + 0.009610720910131931, + 0.011117829941213131, + 0.00966283306479454, + 0.0344603955745697, + 0.023914162069559097, + 0.0050705187022686005, + 0.004786378238350153, + 0.010792383924126625, + 0.0093886349350214, + 0.004127261228859425, + 0.006029109470546246, + 0.009080342948436737, + 0.004676227457821369, + 0.0010257858084514737, + 0.0016722625587135553, + 0.001804451341740787, + 0.0015510726952925324, + 0.0012509454973042011, + 0.000915083393920213, + 0.0018404090078547597, + 0.0014895969070494175, + 0.0030464138835668564, + 0.00415080226957798, + 0.0046578142791986465, + 0.00435010576620698, + 0.003939520101994276, + 0.0028997838962823153, + 0.004971887916326523, + 0.004071368370205164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.011076156049966812 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a single oyster mushroom sprouting from a dark bottom pocket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style characterized by fine brushwork, vibrant mineral pigments, and sharp calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an orange fire start at the base of the mushroom stem and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mushroom's gills transform from pristine white to deep black as they shrivel, curl, and char?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are thin wisps of smoke rising in delicate, calligraphic curls against a dark, damp background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046372017823159695 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.0046372017823159695, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046372017823159695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10867605358362198, + 0.028801830485463142, + 0.02398780547082424, + 0.014892819337546825, + 0.5618994832038879, + 0.06669089198112488, + 0.034810058772563934, + 0.07310418784618378, + 0.014358029700815678, + 0.0727788433432579 + ], + "credit_map": [ + 0.016118478029966354, + 0.010352476499974728, + 0.01818203367292881, + 0.013563715852797031, + 0.009405463002622128, + 0.00738434586673975, + 0.016921624541282654, + 0.016747919842600822, + 0.002493681386113167, + 0.002749877516180277, + 0.0053060841746628284, + 0.003718131687492132, + 0.0023755007423460484, + 0.003027945989742875, + 0.005215734709054232, + 0.003914874512702227, + 0.002265141112729907, + 0.0024844061117619276, + 0.005328113678842783, + 0.002777775516733527, + 0.0019480366026982665, + 0.0014552581124007702, + 0.004104629158973694, + 0.00362444669008255, + 0.0013558331411331892, + 0.0014665370108559728, + 0.0026616051327437162, + 0.0017299252795055509, + 0.0014418577775359154, + 0.0011502584675326943, + 0.0029567612800747156, + 0.00213004183024168, + 0.03171136602759361, + 0.06228352710604668, + 0.08865468204021454, + 0.04761066660284996, + 0.13479813933372498, + 0.030600206926465034, + 0.09535551071166992, + 0.07088537514209747, + 0.0039019326213747263, + 0.007059619762003422, + 0.008030347526073456, + 0.00720645347610116, + 0.016989726573228836, + 0.005432258825749159, + 0.006109595764428377, + 0.0119609534740448, + 0.00245514209382236, + 0.0028968018013983965, + 0.0029678812716156244, + 0.004211162682622671, + 0.0031269777100533247, + 0.0032383997458964586, + 0.009794596582651138, + 0.006119097582995892, + 0.004294972866773605, + 0.003983401693403721, + 0.022490808740258217, + 0.012504701502621174, + 0.00355547689832747, + 0.003739507170394063, + 0.010282360017299652, + 0.012252956628799438, + 0.0011947512393817306, + 0.002033576602116227, + 0.0020039048977196217, + 0.0021283880341798067, + 0.0014955465449020267, + 0.0009363534627482295, + 0.0030249811243265867, + 0.0015405280282720923, + 0.0056087723933160305, + 0.00567964231595397, + 0.012396139092743397, + 0.014080667868256569, + 0.00712395366281271, + 0.0066887810826301575, + 0.009209087118506432, + 0.01199179794639349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.0046372017823159695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_tensor([1316, 1316])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_tensor([1316, 1316])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..15bac524fd8d62dc38f59c13402bc678b01dcf5a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_tensor([1316, 1316])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot, shot from a slightly upward angle, reveals a rich, warm golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame. It casts a brilliant shaft of light, precisely enveloping a figure huddled in a corner. The figure is a man smeared with crude white makeup, a huge, ear-to-ear grin painted in bright red, while black paint surrounds his eyes in sad, drooping triangles. He wears a faded gray hoodie and dusty dark overalls. He hugs his knees to his chest, his head resting limply against the mottled, cracked wall. Within this sacred beam, countless tiny dust particles slowly float, swirl, and dance. The light softly outlines his tousled hair, only to make his comical makeup appear even more melancholic. The whole picture is filled with an extreme sense of tranquility and loneliness against the backdrop of warm colors. The warm light and the sad posture of the characters form a strong visual and emotional contrast, creating a heartbreaking yet beautiful atmosphere.", + "seed": 390, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_1316_seed_390_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01184812029823661, + "vlm_logprob": -0.01184812029823661, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.01184812029823661, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030735937878489494 + ], + "opposite_logprob_tokens": [ + -4.030735969543457 + ], + "logprob_reward": -0.030735937878489494, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.030735937878489494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2679869532585144, + 0.08924421668052673, + 0.10638528317213058, + 0.049303218722343445, + 0.06622091680765152, + 0.05800233036279678, + 0.06553780287504196, + 0.11134174466133118, + 0.12798641622066498, + 0.057991109788417816 + ], + "credit_map": [ + 0.02432294934988022, + 0.030203277245163918, + 0.04443740099668503, + 0.035655677318573, + 0.030641188845038414, + 0.024567771703004837, + 0.04450609162449837, + 0.03365258127450943, + 0.007204598281532526, + 0.011915531940758228, + 0.014252795837819576, + 0.014719984494149685, + 0.008048757910728455, + 0.0065566604025661945, + 0.0141725679859519, + 0.01237332634627819, + 0.008367168717086315, + 0.013631213456392288, + 0.013237332925200462, + 0.022809535264968872, + 0.009693632833659649, + 0.008820859715342522, + 0.016784459352493286, + 0.013041079975664616, + 0.004537642467767, + 0.006463071797043085, + 0.007411045953631401, + 0.010110298171639442, + 0.00427665188908577, + 0.001918475958518684, + 0.007800051476806402, + 0.006785982754081488, + 0.00612920057028532, + 0.005845305509865284, + 0.012526841834187508, + 0.012812250293791294, + 0.0041624098084867, + 0.003959574270993471, + 0.012112491764128208, + 0.008672844618558884, + 0.004811322782188654, + 0.005115507170557976, + 0.006802267860621214, + 0.007378383539617062, + 0.006181048229336739, + 0.008855100721120834, + 0.011258979327976704, + 0.007599718403071165, + 0.008589024655520916, + 0.00617116317152977, + 0.01948838122189045, + 0.0080331452190876, + 0.003926506731659174, + 0.002953881863504648, + 0.00842819269746542, + 0.007947510108351707, + 0.010026968084275723, + 0.007492685224860907, + 0.03221285343170166, + 0.015564247034490108, + 0.006886532064527273, + 0.014641806483268738, + 0.012511956505477428, + 0.012004701420664787, + 0.008971850387752056, + 0.008987905457615852, + 0.03541819751262665, + 0.01892138458788395, + 0.007317306008189917, + 0.027624929323792458, + 0.011148858815431595, + 0.009595978073775768, + 0.0042839134112000465, + 0.00625090766698122, + 0.007779543753713369, + 0.010288342833518982, + 0.0038631500210613012, + 0.003267269115895033, + 0.014021558687090874, + 0.008236425928771496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.030735937878489494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082093119621277 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.016082093119621277, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082093119621277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36903318762779236, + 0.15140007436275482, + 0.148969367146492, + 0.04258076101541519, + 0.049690693616867065, + 0.04020535200834274, + 0.04765123128890991, + 0.06204240396618843, + 0.03594796359539032, + 0.05247896909713745 + ], + "credit_map": [ + 0.025766875594854355, + 0.030648844316601753, + 0.05328090488910675, + 0.04379953816533089, + 0.028291981667280197, + 0.026192395016551018, + 0.04416010528802872, + 0.11689253151416779, + 0.010868885554373264, + 0.020661719143390656, + 0.020233727991580963, + 0.023881249129772186, + 0.01049874909222126, + 0.012938611209392548, + 0.023083986714482307, + 0.029233146458864212, + 0.017024550586938858, + 0.01608564332127571, + 0.021881135180592537, + 0.031947728246450424, + 0.008759268559515476, + 0.006702732294797897, + 0.02571256458759308, + 0.020855743438005447, + 0.003828696208074689, + 0.0065011074766516685, + 0.005576263181865215, + 0.008052054792642593, + 0.003536210162565112, + 0.0027548307552933693, + 0.005760734900832176, + 0.006570861674845219, + 0.0054925293661653996, + 0.005308380350470543, + 0.006472894921898842, + 0.008361470885574818, + 0.0035281600430607796, + 0.0034932882990688086, + 0.011167202144861221, + 0.005866766441613436, + 0.003057673806324601, + 0.0049570705741643906, + 0.0083017498254776, + 0.007652672939002514, + 0.003007975174114108, + 0.00198668846860528, + 0.00684646749868989, + 0.0043950555846095085, + 0.00494932709261775, + 0.00366212148219347, + 0.01358767319470644, + 0.0084172822535038, + 0.002460809890180826, + 0.0031377682462334633, + 0.006721052806824446, + 0.004715196788311005, + 0.00444240914657712, + 0.006227091886103153, + 0.010785415768623352, + 0.00846888404339552, + 0.0043158321641385555, + 0.010113590396940708, + 0.009393386542797089, + 0.00829579308629036, + 0.003576756687834859, + 0.004575102590024471, + 0.006562443915754557, + 0.006024437956511974, + 0.002072281436994672, + 0.002401305129751563, + 0.006082410458475351, + 0.004653224255889654, + 0.00435725087299943, + 0.005852140486240387, + 0.0096395593136549, + 0.008308783173561096, + 0.003790138754993677, + 0.003783304011449218, + 0.010246096178889275, + 0.0065016974695026875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016082093119621277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744122598320246 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.006744122598320246, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744122598320246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4164453446865082, + 0.11029576510190964, + 0.06837116181850433, + 0.03757943958044052, + 0.04474266245961189, + 0.039342377334833145, + 0.040306176990270615, + 0.15431922674179077, + 0.055687788873910904, + 0.03290996700525284 + ], + "credit_map": [ + 0.0356573723256588, + 0.03144893795251846, + 0.05860476568341255, + 0.04707838594913483, + 0.027034413069486618, + 0.026489783078432083, + 0.05237499997019768, + 0.13775669038295746, + 0.007844364270567894, + 0.015317732468247414, + 0.015126918442547321, + 0.019718458876013756, + 0.009323577396571636, + 0.007203539367765188, + 0.02006177231669426, + 0.01569940522313118, + 0.00610964372754097, + 0.007824164815247059, + 0.009319091215729713, + 0.01472720131278038, + 0.005249666515737772, + 0.004302316810935736, + 0.013651612214744091, + 0.007187467534095049, + 0.004825614858418703, + 0.0040433937683701515, + 0.005203599110245705, + 0.005683481227606535, + 0.0033092047087848186, + 0.0029407418332993984, + 0.00682253809645772, + 0.0047508650459349155, + 0.003492855466902256, + 0.005822403822094202, + 0.007828889414668083, + 0.008000047877430916, + 0.003268505446612835, + 0.0033749660942703485, + 0.008625012822449207, + 0.004329977557063103, + 0.004137932788580656, + 0.00480538047850132, + 0.007428829092532396, + 0.006371452007442713, + 0.003084923606365919, + 0.0022517528850585222, + 0.007598733529448509, + 0.003663374576717615, + 0.0045114541426301, + 0.004286664538085461, + 0.00852798018604517, + 0.006956950761377811, + 0.0024338457733392715, + 0.003023766679689288, + 0.0053360192105174065, + 0.005229494534432888, + 0.011592934839427471, + 0.008389484137296677, + 0.04728565737605095, + 0.023485783487558365, + 0.00956309400498867, + 0.024351945146918297, + 0.015857823193073273, + 0.013792499899864197, + 0.0063138678669929504, + 0.007283225655555725, + 0.009059078060090542, + 0.009830661118030548, + 0.0036368672735989094, + 0.0040690163150429726, + 0.008824833668768406, + 0.006670239847153425, + 0.0031032320111989975, + 0.003618723712861538, + 0.00420767767354846, + 0.005673516076058149, + 0.0017514497740194201, + 0.002608679933473468, + 0.007005969528108835, + 0.004940717946738005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.006744122598320246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031907858792692423 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031907858792692423, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031907858792692423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23814362287521362, + 0.06792737543582916, + 0.23096027970314026, + 0.05153847485780716, + 0.05165064334869385, + 0.04250131919980049, + 0.09297484159469604, + 0.09960104525089264, + 0.10275764018297195, + 0.02194473333656788 + ], + "credit_map": [ + 0.01484275609254837, + 0.02334425039589405, + 0.03707928955554962, + 0.02815624140202999, + 0.016796529293060303, + 0.01988431066274643, + 0.022555848583579063, + 0.07548441737890244, + 0.00535696092993021, + 0.010202491655945778, + 0.008555456064641476, + 0.007252159528434277, + 0.007205577101558447, + 0.01088741049170494, + 0.009023216553032398, + 0.009444111958146095, + 0.013249308802187443, + 0.029965702444314957, + 0.0337754487991333, + 0.051905520260334015, + 0.017613016068935394, + 0.011528891511261463, + 0.05016137287020683, + 0.022761043161153793, + 0.0048292335122823715, + 0.00452478788793087, + 0.009217887185513973, + 0.010135979391634464, + 0.005273142829537392, + 0.0038254535757005215, + 0.0068692779168486595, + 0.006862720474600792, + 0.003622452262789011, + 0.005591471679508686, + 0.008649863302707672, + 0.009104923345148563, + 0.005185261368751526, + 0.0025806657504290342, + 0.010647659189999104, + 0.006268350873142481, + 0.003164646215736866, + 0.005834404844790697, + 0.005427350755780935, + 0.0072133904322981834, + 0.004780092742294073, + 0.0022437938023358583, + 0.008884526789188385, + 0.004953117109835148, + 0.008707793429493904, + 0.007691570091992617, + 0.020863791927695274, + 0.01637500897049904, + 0.006699338089674711, + 0.005809243768453598, + 0.014935215935111046, + 0.011892893351614475, + 0.00990199949592352, + 0.005942437332123518, + 0.029557179659605026, + 0.014498387463390827, + 0.007108562160283327, + 0.007957899942994118, + 0.011808871291577816, + 0.012825717218220234, + 0.008154834620654583, + 0.006341015920042992, + 0.03315267339348793, + 0.01653727889060974, + 0.005256963893771172, + 0.01689203456044197, + 0.009183026850223541, + 0.00723982322961092, + 0.0017036814242601395, + 0.002468477003276348, + 0.003837847849354148, + 0.0035214279778301716, + 0.0015949727967381477, + 0.001658475841395557, + 0.004030000884085894, + 0.0031298522371798754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0031907858792692423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024876620154827833 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024876620154827833, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024876620154827833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26332002878189087, + 0.15955863893032074, + 0.09166048467159271, + 0.05364159122109413, + 0.06325914710760117, + 0.05342525616288185, + 0.07914824783802032, + 0.11332213133573532, + 0.07430975884199142, + 0.048354730010032654 + ], + "credit_map": [ + 0.019328083842992783, + 0.026792366057634354, + 0.038942210376262665, + 0.029940588399767876, + 0.016984203830361366, + 0.024040641263127327, + 0.02611960843205452, + 0.08117233216762543, + 0.012301054783165455, + 0.022969158366322517, + 0.021290600299835205, + 0.023482125252485275, + 0.021341316401958466, + 0.009407497942447662, + 0.03214413300156593, + 0.016622750088572502, + 0.007732086349278688, + 0.0066294060088694096, + 0.01464098785072565, + 0.0194744560867548, + 0.006288670934736729, + 0.006115408148616552, + 0.015151378698647022, + 0.015628090128302574, + 0.005867731291800737, + 0.003512430004775524, + 0.012282567098736763, + 0.010601429268717766, + 0.003294465597718954, + 0.0032353452406823635, + 0.008260109461843967, + 0.006587516516447067, + 0.006295442581176758, + 0.0035952182952314615, + 0.019344238564372063, + 0.009598233737051487, + 0.003442322602495551, + 0.006263955030590296, + 0.0075195422396063805, + 0.007200188003480434, + 0.006185300648212433, + 0.002806494478136301, + 0.016142504289746284, + 0.009572147391736507, + 0.002601610729470849, + 0.0035516798961907625, + 0.005894886329770088, + 0.006670634727925062, + 0.00818433053791523, + 0.005572830326855183, + 0.020683616399765015, + 0.014507167041301727, + 0.005276801995933056, + 0.0051629552617669106, + 0.010685795918107033, + 0.009074747562408447, + 0.00841457862406969, + 0.007111679296940565, + 0.032124850898981094, + 0.016774926334619522, + 0.007889497093856335, + 0.016648804768919945, + 0.013033435679972172, + 0.011324355378746986, + 0.004819006659090519, + 0.004166875034570694, + 0.02316046692430973, + 0.013194740749895573, + 0.0035935703199356794, + 0.012248724699020386, + 0.0071645681746304035, + 0.005961807444691658, + 0.005466995760798454, + 0.0031850705854594707, + 0.013550498522818089, + 0.008258200250566006, + 0.0022772455122321844, + 0.00396867236122489, + 0.00638089282438159, + 0.0052671548910439014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0024876620154827833 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly upward angle, reveals a rich, warm golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame. It casts a brilliant shaft of light, precisely enveloping a figure huddled in a corner. The figure is a man smeared with crude white makeup, a huge, ear-to-ear grin painted in bright red, while black paint surrounds his eyes in sad, drooping triangles. He wears a faded gray hoodie and dusty dark overalls. He hugs his knees to his chest, his head resting limply against the mottled, cracked wall. Within this sacred beam, countless tiny dust particles slowly float, swirl, and dance. The light softly outlines his tousled hair, only to make his comical makeup appear even more melancholic. The whole picture is filled with an extreme sense of tranquility and loneliness against the backdrop of warm colors. The warm light and the sad posture of the characters form a strong visual and emotional contrast, creating a heartbreaking yet beautiful atmosphere.", + "seed": 391, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_1316_seed_391_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020800952706485985, + "vlm_logprob": -0.020800952706485985, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.020800952706485985, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043775659054517746 + ], + "opposite_logprob_tokens": [ + -3.7937755584716797 + ], + "logprob_reward": -0.043775659054517746, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.043775659054517746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14395050704479218, + 0.07254333049058914, + 0.17733778059482574, + 0.09868162125349045, + 0.05685238912701607, + 0.152825728058815, + 0.02901064231991768, + 0.13894543051719666, + 0.04371577128767967, + 0.08613674342632294 + ], + "credit_map": [ + 0.015073771588504314, + 0.018286190927028656, + 0.02049909718334675, + 0.020055389031767845, + 0.01813345029950142, + 0.016661377623677254, + 0.01456355582922697, + 0.02067767269909382, + 0.0062653254717588425, + 0.0036197255831211805, + 0.010675220750272274, + 0.006778396666049957, + 0.0055102817714214325, + 0.013841467909514904, + 0.006210421212017536, + 0.019642498344182968, + 0.004643464460968971, + 0.003324168035760522, + 0.01005898229777813, + 0.006160137243568897, + 0.005340232979506254, + 0.06124361604452133, + 0.014946660958230495, + 0.07162052392959595, + 0.0033129160292446613, + 0.004306137096136808, + 0.006833407562226057, + 0.0036117341369390488, + 0.003298196941614151, + 0.0319036990404129, + 0.00783782172948122, + 0.037577711045742035, + 0.002019295934587717, + 0.001733657787553966, + 0.005621884483844042, + 0.003393197664991021, + 0.0031659649685025215, + 0.016979997977614403, + 0.004421152174472809, + 0.019517241045832634, + 0.00383742549456656, + 0.0030604922212660313, + 0.00751857552677393, + 0.005558433476835489, + 0.003564529586583376, + 0.053785040974617004, + 0.011772341094911098, + 0.06372888386249542, + 0.0020571649074554443, + 0.002677577082067728, + 0.004051979165524244, + 0.004892871715128422, + 0.003420501481741667, + 0.003477460704743862, + 0.004404447041451931, + 0.004028639756143093, + 0.005857779644429684, + 0.005021828226745129, + 0.01359168067574501, + 0.008144432678818703, + 0.008451194502413273, + 0.036901142448186874, + 0.011658714152872562, + 0.049318671226501465, + 0.0018917527049779892, + 0.0013981988886371255, + 0.006203165277838707, + 0.002595131052657962, + 0.002173899905756116, + 0.01164272055029869, + 0.0036264746449887753, + 0.014184431172907352, + 0.00457410654053092, + 0.004714844282716513, + 0.00549792405217886, + 0.00388454576022923, + 0.003472001990303397, + 0.024888064712285995, + 0.006642783991992474, + 0.0324624739587307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.043775659054517746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381275936961174 + ], + "opposite_logprob_tokens": [ + -3.2881276607513428 + ], + "logprob_reward": -0.0381275936961174, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381275936961174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10969652980566025, + 0.1274644285440445, + 0.2571428120136261, + 0.19949783384799957, + 0.021304812282323837, + 0.029079079627990723, + 0.04570826515555382, + 0.13522674143314362, + 0.04029902070760727, + 0.03458043560385704 + ], + "credit_map": [ + 0.009324409998953342, + 0.008717697113752365, + 0.024914652109146118, + 0.014502950944006443, + 0.014386004768311977, + 0.01101374626159668, + 0.013177297078073025, + 0.013659766875207424, + 0.005803815554827452, + 0.007626374252140522, + 0.008944256231188774, + 0.008282804861664772, + 0.007045527454465628, + 0.03566453233361244, + 0.010572157800197601, + 0.043524958193302155, + 0.006512698717415333, + 0.005395914893597364, + 0.013498188927769661, + 0.007984231226146221, + 0.00671398313716054, + 0.08942891657352448, + 0.02081434056162834, + 0.10679455101490021, + 0.004510364029556513, + 0.004431725945323706, + 0.010019290260970592, + 0.006316957995295525, + 0.005357109941542149, + 0.07007434219121933, + 0.016824951395392418, + 0.08196310698986053, + 0.0016122291563078761, + 0.001989270793274045, + 0.00248709786683321, + 0.00383653212338686, + 0.0028329479973763227, + 0.0029774627182632685, + 0.0023894361220300198, + 0.003179834457114339, + 0.0023093547206372023, + 0.001641681301407516, + 0.004847012460231781, + 0.004612112417817116, + 0.0029698656871914864, + 0.004531276412308216, + 0.003109878860414028, + 0.0050578974187374115, + 0.003679105546325445, + 0.002343269530683756, + 0.004933349788188934, + 0.004147517494857311, + 0.005327705293893814, + 0.008925524540245533, + 0.005323636345565319, + 0.011028154753148556, + 0.006515625398606062, + 0.004211533349007368, + 0.011330381035804749, + 0.006235900335013866, + 0.0075191352516412735, + 0.03865969553589821, + 0.010678999125957489, + 0.0500754751265049, + 0.0019285213202238083, + 0.0010306554613634944, + 0.00668751448392868, + 0.0025269126053899527, + 0.0016654051141813397, + 0.010462483391165733, + 0.0032656854018568993, + 0.012731839902698994, + 0.0021369103342294693, + 0.0014564032899215817, + 0.0048698121681809425, + 0.0027359589003026485, + 0.0016425708308815956, + 0.009179091081023216, + 0.0028025582432746887, + 0.009757130406796932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.0381275936961174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216454699635506 + ], + "opposite_logprob_tokens": [ + -4.264216423034668 + ], + "logprob_reward": -0.014216454699635506, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216454699635506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11932725459337234, + 0.0500340461730957, + 0.19136273860931396, + 0.20860467851161957, + 0.07727037370204926, + 0.15499797463417053, + 0.048024047166109085, + 0.08606032282114029, + 0.027925366535782814, + 0.036393169313669205 + ], + "credit_map": [ + 0.012388271279633045, + 0.011907304637134075, + 0.023932572454214096, + 0.01545849721878767, + 0.010714140720665455, + 0.015668215230107307, + 0.012926607392728329, + 0.016331644728779793, + 0.004855746868997812, + 0.0059729949571192265, + 0.005137468222528696, + 0.0070065720938146114, + 0.0051436880603432655, + 0.006868591997772455, + 0.0052882912568748, + 0.009760690852999687, + 0.005944141652435064, + 0.004138259217143059, + 0.011918684467673302, + 0.007374489214271307, + 0.006672201212495565, + 0.06344571709632874, + 0.01556279044598341, + 0.07630644738674164, + 0.005235678981989622, + 0.005672745406627655, + 0.010452975519001484, + 0.0061413682997226715, + 0.0044175731018185616, + 0.07336645573377609, + 0.017364749684929848, + 0.08595314621925354, + 0.0031063812784850597, + 0.003030401188880205, + 0.0060259681195020676, + 0.002776402747258544, + 0.0035910599399358034, + 0.023586342111229897, + 0.005835608579218388, + 0.029318207874894142, + 0.004179216921329498, + 0.0033091148361563683, + 0.007691013161092997, + 0.007548807188868523, + 0.0041262139566242695, + 0.0535532683134079, + 0.012091648764908314, + 0.062498707324266434, + 0.0052960519678890705, + 0.003221380291506648, + 0.007922003977000713, + 0.00609807763248682, + 0.006507371086627245, + 0.005659028422087431, + 0.006185227073729038, + 0.007134905084967613, + 0.003988492768257856, + 0.0024615824222564697, + 0.01075008511543274, + 0.004550503566861153, + 0.004986747168004513, + 0.023063424974679947, + 0.007216035854071379, + 0.02904345653951168, + 0.0014912512851879, + 0.00106123648583889, + 0.00478973425924778, + 0.003340788185596466, + 0.002446327591314912, + 0.00540763046592474, + 0.002885308815166354, + 0.006503089796751738, + 0.0030960168223828077, + 0.0019704385194927454, + 0.005358686670660973, + 0.0032605419401079416, + 0.0021462521981447935, + 0.008125713095068932, + 0.003467654576525092, + 0.008967868052423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.014216454699635506 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019393223337829113 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019393223337829113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019393223337829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12167910486459732, + 0.07581577450037003, + 0.09314797073602676, + 0.09245830774307251, + 0.15173037350177765, + 0.12453856319189072, + 0.046405356377363205, + 0.20146772265434265, + 0.016052350401878357, + 0.07670444995164871 + ], + "credit_map": [ + 0.009578291326761246, + 0.005926149897277355, + 0.01865110918879509, + 0.011532053351402283, + 0.011478351429104805, + 0.022298235446214676, + 0.014584009535610676, + 0.027630912140011787, + 0.0031666099093854427, + 0.0027482749428600073, + 0.005859000142663717, + 0.004165126010775566, + 0.0029452848248183727, + 0.02287101559340954, + 0.006794665940105915, + 0.027265796437859535, + 0.0035519760567694902, + 0.002696538344025612, + 0.007796634454280138, + 0.0034632610622793436, + 0.003024168312549591, + 0.02970260940492153, + 0.0075531103648245335, + 0.03535967692732811, + 0.0034730888437479734, + 0.0046367645263671875, + 0.004778439644724131, + 0.003249482484534383, + 0.0023398851044476032, + 0.03096681274473667, + 0.0074221296235919, + 0.03559170290827751, + 0.003732152748852968, + 0.0028734963852912188, + 0.00796983391046524, + 0.003623352153226733, + 0.003120590467005968, + 0.05394202843308449, + 0.011992246843874454, + 0.06447666883468628, + 0.003005512524396181, + 0.002272561425343156, + 0.005931970197707415, + 0.004105335101485252, + 0.0026666908524930477, + 0.044617947190999985, + 0.009765935130417347, + 0.05217261612415314, + 0.004905316978693008, + 0.0026432983577251434, + 0.008547397330403328, + 0.007122649811208248, + 0.0063410066068172455, + 0.004669752437621355, + 0.005790287628769875, + 0.006385649088770151, + 0.008547677658498287, + 0.0068345945328474045, + 0.017102397978305817, + 0.010270827449858189, + 0.01134931668639183, + 0.055347636342048645, + 0.016590654850006104, + 0.07542462646961212, + 0.0015957673313096166, + 0.0008352956501767039, + 0.004111941903829575, + 0.001885060453787446, + 0.0012060604058206081, + 0.0017292362172156572, + 0.0017783557996153831, + 0.0029106324072927237, + 0.002532300539314747, + 0.00221015652641654, + 0.005910148378461599, + 0.0027813224587589502, + 0.0021742985118180513, + 0.024513114243745804, + 0.005996130872517824, + 0.030586982145905495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.0019393223337829113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059457337483763695 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.0059457337483763695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059457337483763695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16188494861125946, + 0.20169556140899658, + 0.3475886285305023, + 0.06624048203229904, + 0.033126868307590485, + 0.06318966299295425, + 0.0372043214738369, + 0.031396493315696716, + 0.0346972793340683, + 0.022975843399763107 + ], + "credit_map": [ + 0.014671792276203632, + 0.018909577280282974, + 0.017917776480317116, + 0.015481943264603615, + 0.012768703512847424, + 0.024111833423376083, + 0.017223255708813667, + 0.0408000648021698, + 0.009985355660319328, + 0.011782487854361534, + 0.01174700167030096, + 0.008983652107417583, + 0.007825598120689392, + 0.060912322252988815, + 0.016881374642252922, + 0.07357776165008545, + 0.008803718723356724, + 0.0073097883723676205, + 0.016620492562651634, + 0.00987248308956623, + 0.007906747050583363, + 0.12128368020057678, + 0.02950979582965374, + 0.14628194272518158, + 0.002301489468663931, + 0.003437910694628954, + 0.004621877800673246, + 0.003136045066639781, + 0.0023747587110847235, + 0.020995261147618294, + 0.005304630845785141, + 0.02406850829720497, + 0.0028931486885994673, + 0.003232762450352311, + 0.004392070695757866, + 0.003368833102285862, + 0.0027293553575873375, + 0.005810908041894436, + 0.0037620875518769026, + 0.006937701720744371, + 0.00236703478731215, + 0.0020767648238688707, + 0.00496062682941556, + 0.0036559898871928453, + 0.0035663284361362457, + 0.019591297954320908, + 0.004849150311201811, + 0.022122463211417198, + 0.001582205994054675, + 0.001539400196634233, + 0.003926961682736874, + 0.0023975931107997894, + 0.001999241067096591, + 0.009954371489584446, + 0.0031755545642226934, + 0.012628993019461632, + 0.0031547299586236477, + 0.002330948133021593, + 0.005719675216823816, + 0.003820665879175067, + 0.0032164743170142174, + 0.00464606424793601, + 0.003418141510337591, + 0.00508979381993413, + 0.0013724376913160086, + 0.0008593779639340937, + 0.0021919601131230593, + 0.0014930928591638803, + 0.001325261895544827, + 0.01133890450000763, + 0.002540787449106574, + 0.013575457967817783, + 0.0024796477518975735, + 0.0022754380479454994, + 0.004372210241854191, + 0.0022833975963294506, + 0.0022020800970494747, + 0.0033576958812773228, + 0.0022314030211418867, + 0.0037739721592515707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0059457337483763695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly upward angle, reveals a rich, warm golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame. It casts a brilliant shaft of light, precisely enveloping a figure huddled in a corner. The figure is a man smeared with crude white makeup, a huge, ear-to-ear grin painted in bright red, while black paint surrounds his eyes in sad, drooping triangles. He wears a faded gray hoodie and dusty dark overalls. He hugs his knees to his chest, his head resting limply against the mottled, cracked wall. Within this sacred beam, countless tiny dust particles slowly float, swirl, and dance. The light softly outlines his tousled hair, only to make his comical makeup appear even more melancholic. The whole picture is filled with an extreme sense of tranquility and loneliness against the backdrop of warm colors. The warm light and the sad posture of the characters form a strong visual and emotional contrast, creating a heartbreaking yet beautiful atmosphere.", + "seed": 392, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_1316_seed_392_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01287759724073112, + "vlm_logprob": -0.01287759724073112, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.01287759724073112, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034239672124385834 + ], + "opposite_logprob_tokens": [ + -4.159239768981934 + ], + "logprob_reward": -0.034239672124385834, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.034239672124385834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1678873747587204, + 0.20750005543231964, + 0.04434523731470108, + 0.052626706659793854, + 0.029697012156248093, + 0.0666498914361, + 0.09914154559373856, + 0.14953787624835968, + 0.11545640230178833, + 0.06715787947177887 + ], + "credit_map": [ + 0.010793293826282024, + 0.007045028731226921, + 0.028462231159210205, + 0.029835321009159088, + 0.01894564926624298, + 0.01403950434178114, + 0.03890318423509598, + 0.01986316591501236, + 0.02481570839881897, + 0.012556253932416439, + 0.06459012627601624, + 0.02029239386320114, + 0.01341394241899252, + 0.015745578333735466, + 0.01806393824517727, + 0.03802211210131645, + 0.0023950233589857817, + 0.0016788376960903406, + 0.007315882947295904, + 0.006881742272526026, + 0.004405522719025612, + 0.006616026163101196, + 0.007784642744809389, + 0.0072675589472055435, + 0.0035737529397010803, + 0.0024887353647500277, + 0.007768279407173395, + 0.007217023056000471, + 0.004768901504576206, + 0.007782998960465193, + 0.01078590378165245, + 0.008241110481321812, + 0.004265955649316311, + 0.0018018386326730251, + 0.004494914319366217, + 0.0035938199143856764, + 0.0020570866763591766, + 0.002853976795449853, + 0.004906964488327503, + 0.005722455680370331, + 0.007663378491997719, + 0.0030337078496813774, + 0.01816040463745594, + 0.009737922810018063, + 0.0037062063347548246, + 0.006735270842909813, + 0.00767504470422864, + 0.009937955997884274, + 0.009270253591239452, + 0.004799235612154007, + 0.019035615026950836, + 0.019616272300481796, + 0.007815293036401272, + 0.010962833650410175, + 0.013695107772946358, + 0.013946927152574062, + 0.007849614135921001, + 0.0055149020627141, + 0.028836533427238464, + 0.03762092813849449, + 0.00891853030771017, + 0.01183408498764038, + 0.02377244643867016, + 0.02519083023071289, + 0.005614988040179014, + 0.003370084799826145, + 0.036915700882673264, + 0.00686268974095583, + 0.004947475157678127, + 0.009394632652401924, + 0.012726545333862305, + 0.03562429919838905, + 0.005796312820166349, + 0.004342404194176197, + 0.015294697135686874, + 0.008943038992583752, + 0.0061371116898953915, + 0.007154461462050676, + 0.009477997198700905, + 0.010011854581534863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.034239672124385834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014218100346624851 + ], + "opposite_logprob_tokens": [ + -4.264218330383301 + ], + "logprob_reward": -0.014218100346624851, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014218100346624851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14985507726669312, + 0.19730141758918762, + 0.14274391531944275, + 0.051894403994083405, + 0.03500313311815262, + 0.04321473091840744, + 0.06043132767081261, + 0.10234762728214264, + 0.15104356408119202, + 0.06616471707820892 + ], + "credit_map": [ + 0.012110351584851742, + 0.008984765037894249, + 0.025712639093399048, + 0.026696577668190002, + 0.022579997777938843, + 0.012611356563866138, + 0.022954562678933144, + 0.018204843625426292, + 0.019270680844783783, + 0.010323169641196728, + 0.054988812655210495, + 0.026685930788517, + 0.01523792278021574, + 0.018430441617965698, + 0.018850095570087433, + 0.03351437672972679, + 0.011147228069603443, + 0.005324363708496094, + 0.026546092703938484, + 0.02325899340212345, + 0.013362785801291466, + 0.02281731739640236, + 0.02147701196372509, + 0.018810132518410683, + 0.0037652947939932346, + 0.0022430343087762594, + 0.009888000786304474, + 0.005263130646198988, + 0.00443659070879221, + 0.008653676137328148, + 0.009035633876919746, + 0.008609048090875149, + 0.0028464009519666433, + 0.001166920643299818, + 0.011971802450716496, + 0.003500182880088687, + 0.0015253808815032244, + 0.0026314633432775736, + 0.0034293814096599817, + 0.007931606844067574, + 0.002806502627208829, + 0.0017018067883327603, + 0.007466695737093687, + 0.00892557855695486, + 0.0032634211238473654, + 0.005128820426762104, + 0.00785625260323286, + 0.0060656568966805935, + 0.008548744022846222, + 0.0028567134868353605, + 0.018119879066944122, + 0.006689436733722687, + 0.0025648893788456917, + 0.00474204495549202, + 0.006805825978517532, + 0.010103796608746052, + 0.004860449582338333, + 0.0031526454258710146, + 0.02444107085466385, + 0.018475834280252457, + 0.004048108588904142, + 0.008852237835526466, + 0.011998905800282955, + 0.026518389582633972, + 0.006289896089583635, + 0.0034363246522843838, + 0.05153634399175644, + 0.01050030067563057, + 0.009632604196667671, + 0.010221283882856369, + 0.013635014183819294, + 0.04579181596636772, + 0.0027285278774797916, + 0.002614154014736414, + 0.012290336191654205, + 0.006865768227726221, + 0.009197100065648556, + 0.009693551808595657, + 0.011347807012498379, + 0.011427481658756733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.014218100346624851 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648795075714588 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648795075714588, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648795075714588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19781742990016937, + 0.11823470145463943, + 0.13141605257987976, + 0.0632382482290268, + 0.0627448558807373, + 0.09251455217599869, + 0.12192694842815399, + 0.07296709716320038, + 0.06060300022363663, + 0.07853716611862183 + ], + "credit_map": [ + 0.0122141158208251, + 0.00837990827858448, + 0.03159584105014801, + 0.024936094880104065, + 0.031553544104099274, + 0.019075796008110046, + 0.045070093125104904, + 0.024992015212774277, + 0.00730676157400012, + 0.0049093314446508884, + 0.021746736019849777, + 0.023686859756708145, + 0.012033551931381226, + 0.01338957343250513, + 0.019255872815847397, + 0.01590600050985813, + 0.01068150531500578, + 0.005063317250460386, + 0.026645377278327942, + 0.016739774495363235, + 0.00674959784373641, + 0.014698414131999016, + 0.017251351848244667, + 0.03358669951558113, + 0.004963025916367769, + 0.004060209728777409, + 0.009516996331512928, + 0.007116721943020821, + 0.005538078490644693, + 0.010995796881616116, + 0.011656675487756729, + 0.009390740655362606, + 0.003994090482592583, + 0.002442775759845972, + 0.017756761983036995, + 0.006977671291679144, + 0.004275668878108263, + 0.0044379280880093575, + 0.007624913938343525, + 0.015235033817589283, + 0.011202462948858738, + 0.004902942571789026, + 0.02079436369240284, + 0.01395175140351057, + 0.006605083588510752, + 0.007432024925947189, + 0.014763543382287025, + 0.012862366624176502, + 0.011332026682794094, + 0.005399998743087053, + 0.03356613963842392, + 0.01743767224252224, + 0.009162024594843388, + 0.009183828718960285, + 0.013718088157474995, + 0.022127149626612663, + 0.004493356216698885, + 0.0025333331432193518, + 0.018830537796020508, + 0.010115083307027817, + 0.003791138296946883, + 0.0059453691355884075, + 0.007333060726523399, + 0.019925205036997795, + 0.0044919573701918125, + 0.0023891404271125793, + 0.009030178189277649, + 0.005906587932258844, + 0.008512267842888832, + 0.009368428960442543, + 0.0130736343562603, + 0.00783079955726862, + 0.004969541449099779, + 0.0025744657032191753, + 0.01789063774049282, + 0.006871462799608707, + 0.006385664455592632, + 0.006848207209259272, + 0.015449865721166134, + 0.017547309398651123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.008648795075714588 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190310439094901 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190310439094901, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190310439094901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1736375242471695, + 0.20974047482013702, + 0.09831520915031433, + 0.03547070920467377, + 0.1748345047235489, + 0.07799725979566574, + 0.055833883583545685, + 0.05095077306032181, + 0.05743599683046341, + 0.06578361243009567 + ], + "credit_map": [ + 0.009913294576108456, + 0.009094922803342342, + 0.03227321803569794, + 0.023437080904841423, + 0.023849505931138992, + 0.017083240672945976, + 0.037534311413764954, + 0.02045195735991001, + 0.019024895504117012, + 0.010471378453075886, + 0.04229545593261719, + 0.03282131999731064, + 0.021034203469753265, + 0.023027686402201653, + 0.027902141213417053, + 0.03316342085599899, + 0.009184486232697964, + 0.004506010562181473, + 0.018916189670562744, + 0.011578655801713467, + 0.005176646169275045, + 0.011530288495123386, + 0.012839768081903458, + 0.02458316646516323, + 0.002465607365593314, + 0.002217002445831895, + 0.007314664777368307, + 0.005529928021132946, + 0.003629395505413413, + 0.004185767844319344, + 0.005575084593147039, + 0.004553262609988451, + 0.0108593525364995, + 0.0038546896539628506, + 0.05667397752404213, + 0.014167462475597858, + 0.0075615025125443935, + 0.010873747989535332, + 0.017579134553670883, + 0.05326465144753456, + 0.008146055042743683, + 0.0036399313248693943, + 0.015258080326020718, + 0.012743416242301464, + 0.00813538022339344, + 0.007736427243798971, + 0.012333802878856659, + 0.010004173964262009, + 0.00827062502503395, + 0.002994319424033165, + 0.010579629801213741, + 0.007666849531233311, + 0.004957264754921198, + 0.005941503215581179, + 0.007908374071121216, + 0.007515322882682085, + 0.0024579253513365984, + 0.0014726396184414625, + 0.01687062904238701, + 0.004632025957107544, + 0.00207069655880332, + 0.0028330376371741295, + 0.004889071919023991, + 0.015724748373031616, + 0.004513617139309645, + 0.0023516209330409765, + 0.00881448294967413, + 0.006107931956648827, + 0.006810311693698168, + 0.007414196152240038, + 0.013791798613965511, + 0.0076320418156683445, + 0.004299215506762266, + 0.0027253313455730677, + 0.010002327151596546, + 0.006796167232096195, + 0.006732824724167585, + 0.006455899216234684, + 0.016312744468450546, + 0.012459107674658298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.003190310439094901 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091108217835426 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091108217835426, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091108217835426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2472372055053711, + 0.12106537073850632, + 0.09270733594894409, + 0.06070006266236305, + 0.03672490641474724, + 0.08177017420530319, + 0.10034043341875076, + 0.04349849373102188, + 0.10927430540323257, + 0.10668167471885681 + ], + "credit_map": [ + 0.02694076858460903, + 0.027089184150099754, + 0.046829331666231155, + 0.031194807961583138, + 0.02108743228018284, + 0.027184879407286644, + 0.030263207852840424, + 0.036647625267505646, + 0.019901486113667488, + 0.008245679549872875, + 0.030958082526922226, + 0.013827246613800526, + 0.007046201266348362, + 0.008897664956748486, + 0.012340270914137363, + 0.019848745316267014, + 0.009307236410677433, + 0.01115648727864027, + 0.015354449860751629, + 0.013143885880708694, + 0.008264527656137943, + 0.01235536951571703, + 0.011352277360856533, + 0.011773115955293179, + 0.004478131420910358, + 0.0036507900804281235, + 0.01015963964164257, + 0.006452018395066261, + 0.005246713757514954, + 0.01014922745525837, + 0.010704108513891697, + 0.009859434328973293, + 0.0039853001944720745, + 0.0018723919056355953, + 0.005526446271687746, + 0.005520973354578018, + 0.002592000411823392, + 0.0034481261391192675, + 0.005907359533011913, + 0.007872315123677254, + 0.00955105945467949, + 0.0042847441509366035, + 0.02704988420009613, + 0.008323192596435547, + 0.0031356869731098413, + 0.004376463126391172, + 0.00811217725276947, + 0.01693696528673172, + 0.01125416625291109, + 0.004437964875251055, + 0.027255969122052193, + 0.01577571965754032, + 0.006211754400283098, + 0.00807039812207222, + 0.010962549597024918, + 0.016371918842196465, + 0.0035967540461570024, + 0.00170760543551296, + 0.013802265748381615, + 0.004807643126696348, + 0.0022684705909341574, + 0.0034899122547358274, + 0.0036358258221298456, + 0.010190018452703953, + 0.008696288801729679, + 0.003371816361322999, + 0.036653634160757065, + 0.007752304896712303, + 0.006106548011302948, + 0.008703909814357758, + 0.011135322973132133, + 0.026854485273361206, + 0.009837916120886803, + 0.004997249227017164, + 0.023899225518107414, + 0.010082772001624107, + 0.011378075927495956, + 0.011437117122113705, + 0.015146453864872456, + 0.019902868196368217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.004091108217835426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a slightly upward angle, reveals a rich, warm golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame. It casts a brilliant shaft of light, precisely enveloping a figure huddled in a corner. The figure is a man smeared with crude white makeup, a huge, ear-to-ear grin painted in bright red, while black paint surrounds his eyes in sad, drooping triangles. He wears a faded gray hoodie and dusty dark overalls. He hugs his knees to his chest, his head resting limply against the mottled, cracked wall. Within this sacred beam, countless tiny dust particles slowly float, swirl, and dance. The light softly outlines his tousled hair, only to make his comical makeup appear even more melancholic. The whole picture is filled with an extreme sense of tranquility and loneliness against the backdrop of warm colors. The warm light and the sad posture of the characters form a strong visual and emotional contrast, creating a heartbreaking yet beautiful atmosphere.", + "seed": 393, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_7/index_1316_seed_393_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04857321567833424, + "vlm_logprob": -0.04857321567833424, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04857321567833424, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14050975441932678 + ], + "opposite_logprob_tokens": [ + -2.140509843826294 + ], + "logprob_reward": -0.14050975441932678, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14050975441932678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30568239092826843, + 0.1032477393746376, + 0.1909795105457306, + 0.07526882737874985, + 0.04859308525919914, + 0.10005176067352295, + 0.053027357906103134, + 0.0526600144803524, + 0.023892393335700035, + 0.04659688472747803 + ], + "credit_map": [ + 0.014130333438515663, + 0.018201211467385292, + 0.05001431703567505, + 0.06268947571516037, + 0.02607685513794422, + 0.028440523892641068, + 0.052358996123075485, + 0.05377069115638733, + 0.005660883150994778, + 0.005847907159477472, + 0.020054636523127556, + 0.018735000863671303, + 0.009616203606128693, + 0.008576088584959507, + 0.01493605226278305, + 0.01982097327709198, + 0.006581111345440149, + 0.00448965048417449, + 0.07547108829021454, + 0.010796093381941319, + 0.002888335380703211, + 0.007366120349615812, + 0.009522504173219204, + 0.07386462390422821, + 0.005305315833538771, + 0.004696425516158342, + 0.013411504216492176, + 0.012303502298891544, + 0.006924773100763559, + 0.006430709268897772, + 0.01058974489569664, + 0.015606851316988468, + 0.0025083194486796856, + 0.00326277082785964, + 0.00914976466447115, + 0.005605300888419151, + 0.0042991130612790585, + 0.004735766444355249, + 0.008713987655937672, + 0.010318062268197536, + 0.00802062638103962, + 0.008452272042632103, + 0.015104657970368862, + 0.014176293276250362, + 0.009860381484031677, + 0.008344589732587337, + 0.021929599344730377, + 0.014163345098495483, + 0.003113517304882407, + 0.003387739649042487, + 0.00490980688482523, + 0.006346139591187239, + 0.0056329588405787945, + 0.005018268711864948, + 0.0165561530739069, + 0.008062775246798992, + 0.002588359173387289, + 0.0032601915299892426, + 0.0065446775406599045, + 0.007885723374783993, + 0.004140425939112902, + 0.004381152801215649, + 0.015410693362355232, + 0.008448793552815914, + 0.0016541298246011138, + 0.0022409744560718536, + 0.0034307322930544615, + 0.0031536193564534187, + 0.0021181541960686445, + 0.0031701698899269104, + 0.004113787319511175, + 0.004010827746242285, + 0.002223326126113534, + 0.002683630445972085, + 0.004793747793883085, + 0.007920834235846996, + 0.004370429553091526, + 0.005608369130641222, + 0.01115846261382103, + 0.007838086225092411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.14050975441932678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900354266166687 + ], + "opposite_logprob_tokens": [ + -2.5790035724639893 + ], + "logprob_reward": -0.07900354266166687, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900354266166687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22377468645572662, + 0.13323456048965454, + 0.10953554511070251, + 0.09653904289007187, + 0.05289286747574806, + 0.10898838937282562, + 0.037704505026340485, + 0.06247812509536743, + 0.05414638668298721, + 0.12070591002702713 + ], + "credit_map": [ + 0.012454165145754814, + 0.01488584280014038, + 0.027043135836720467, + 0.05113378167152405, + 0.021645722910761833, + 0.022428641095757484, + 0.043945979326963425, + 0.03023741953074932, + 0.0073961252346634865, + 0.00652781268581748, + 0.028054263442754745, + 0.023909766227006912, + 0.013198710046708584, + 0.0105248698964715, + 0.018970699980854988, + 0.024652305990457535, + 0.0052368598990142345, + 0.0070153833366930485, + 0.026131540536880493, + 0.01119036041200161, + 0.006984235718846321, + 0.006419935263693333, + 0.015365010127425194, + 0.031192224472761154, + 0.006917855702340603, + 0.006652079056948423, + 0.014184586703777313, + 0.014748957008123398, + 0.010357941500842571, + 0.009438099339604378, + 0.019358690828084946, + 0.014880834147334099, + 0.00338093563914299, + 0.003998337779194117, + 0.011771924793720245, + 0.00590670108795166, + 0.00317169027402997, + 0.0032783106435090303, + 0.006751749664545059, + 0.01463321503251791, + 0.010165706276893616, + 0.009883088059723377, + 0.017293885350227356, + 0.013619320467114449, + 0.008601375855505466, + 0.008562767878174782, + 0.02522202394902706, + 0.015640215948224068, + 0.002046718029305339, + 0.002659633057191968, + 0.0038611339405179024, + 0.006269233301281929, + 0.0039061561692506075, + 0.004054415971040726, + 0.009174021892249584, + 0.005733192898333073, + 0.0034172346349805593, + 0.004163407254964113, + 0.008550143800675869, + 0.011768871918320656, + 0.005489055532962084, + 0.00615929439663887, + 0.013752268627285957, + 0.00917784869670868, + 0.002983140992000699, + 0.003441380336880684, + 0.005361147690564394, + 0.008981742896139622, + 0.0052612763829529285, + 0.005968471057713032, + 0.013333156704902649, + 0.008816070854663849, + 0.005240194499492645, + 0.0069280099123716354, + 0.011518830433487892, + 0.020814377814531326, + 0.011884134262800217, + 0.014287806116044521, + 0.029330391436815262, + 0.020702172070741653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07900354266166687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089949868619442 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089949868619442, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089949868619442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20063446462154388, + 0.04907941445708275, + 0.3701174259185791, + 0.027565432712435722, + 0.018813343718647957, + 0.07464439421892166, + 0.17343971133232117, + 0.04489617422223091, + 0.016155708581209183, + 0.024653907865285873 + ], + "credit_map": [ + 0.012731255032122135, + 0.012442483566701412, + 0.05445607379078865, + 0.025772886350750923, + 0.01224349532276392, + 0.016690481454133987, + 0.024187488481402397, + 0.04211030900478363, + 0.0027415668591856956, + 0.0029738801531493664, + 0.007231794763356447, + 0.00965193286538124, + 0.005786161404103041, + 0.0047568329609930515, + 0.008892967365682125, + 0.00704427994787693, + 0.013356405310332775, + 0.008618975058197975, + 0.1476011872291565, + 0.019205771386623383, + 0.004545752424746752, + 0.01294941920787096, + 0.01935475505888462, + 0.14448514580726624, + 0.0025599158834666014, + 0.0028662276454269886, + 0.0029924563132226467, + 0.004561509471386671, + 0.0030833135824650526, + 0.002689435612410307, + 0.004760864190757275, + 0.004051710944622755, + 0.0011981477728113532, + 0.0013498534681275487, + 0.002644893014803529, + 0.002898489125072956, + 0.0019799591973423958, + 0.0018840155098587275, + 0.003283972619101405, + 0.003574012080207467, + 0.006533326581120491, + 0.00632950384169817, + 0.01143847405910492, + 0.009391427971422672, + 0.005982750095427036, + 0.005518485326319933, + 0.01862657070159912, + 0.010823852382600307, + 0.005751697346568108, + 0.008761484175920486, + 0.017278485000133514, + 0.020949700847268105, + 0.018037155270576477, + 0.015873514115810394, + 0.060960911214351654, + 0.025826767086982727, + 0.0016740019200369716, + 0.0024190363474190235, + 0.0046485117636621, + 0.006858876906335354, + 0.0034024666529148817, + 0.003141068620607257, + 0.015362223610281944, + 0.007389988284558058, + 0.0015859071863815188, + 0.0012398399412631989, + 0.0017401240766048431, + 0.0024754286278039217, + 0.0012876834953203797, + 0.0020718013402074575, + 0.0033249161206185818, + 0.002430006628856063, + 0.0016303963493555784, + 0.0012983593624085188, + 0.0036918274126946926, + 0.0032992607448250055, + 0.0018946198979392648, + 0.0028328094631433487, + 0.005923919845372438, + 0.004082715138792992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.011089949868619442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628121413290501 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007628121413290501, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628121413290501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23220883309841156, + 0.05628038942813873, + 0.39739057421684265, + 0.047348521649837494, + 0.028805559501051903, + 0.02493264153599739, + 0.10479342937469482, + 0.048287369310855865, + 0.024865517392754555, + 0.03508712351322174 + ], + "credit_map": [ + 0.014875638298690319, + 0.011221719905734062, + 0.07279122620820999, + 0.02691388688981533, + 0.010400868020951748, + 0.0173695869743824, + 0.02265881560742855, + 0.05597708746790886, + 0.004260455258190632, + 0.004993127658963203, + 0.007673685438930988, + 0.009554545395076275, + 0.0069795167073607445, + 0.00586314732208848, + 0.010350004769861698, + 0.006605904549360275, + 0.013997613452374935, + 0.009325715713202953, + 0.1602279245853424, + 0.021306894719600677, + 0.0033791589085012674, + 0.014000278897583485, + 0.018510615453124046, + 0.1566423922777176, + 0.00277542881667614, + 0.002577973296865821, + 0.0123673677444458, + 0.004489674232900143, + 0.0023035486228764057, + 0.0026545897126197815, + 0.004198900889605284, + 0.0159810408949852, + 0.003229399910196662, + 0.0020375261083245277, + 0.0038406511303037405, + 0.004195427522063255, + 0.001821301644667983, + 0.002779993461444974, + 0.0067577846348285675, + 0.00414347555488348, + 0.0031567756086587906, + 0.0019093219889327884, + 0.0038429780397564173, + 0.003928024787455797, + 0.0017672105459496379, + 0.0019256728701293468, + 0.005105244927108288, + 0.0032974141649901867, + 0.0047655655071139336, + 0.006214577704668045, + 0.011485875584185123, + 0.012255091220140457, + 0.01027611643075943, + 0.008773496374487877, + 0.03542473912239075, + 0.01559796929359436, + 0.0016775298863649368, + 0.002262665657326579, + 0.004063398577272892, + 0.005662146490067244, + 0.004420414101332426, + 0.003833527909591794, + 0.0180650744587183, + 0.00830260943621397, + 0.001941666821949184, + 0.0015036341501399875, + 0.00238735880702734, + 0.003721301443874836, + 0.0020347312092781067, + 0.002580930944532156, + 0.006467381492257118, + 0.0042285132221877575, + 0.002466748468577862, + 0.002418715739622712, + 0.004093323368579149, + 0.004642056301236153, + 0.0025623536203056574, + 0.003465804737061262, + 0.008573058992624283, + 0.006865060422569513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.007628121413290501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man visible wearing a faded gray hoodie and dusty dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man have crude white makeup with a huge red ear-to-ear grin and black paint in sad, drooping triangles around his eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man huddled in a corner with his knees hugged to his chest and head resting against a mottled, cracked wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset beam slanting through a dusty, old-fashioned window with a peeling wood frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634710028767586 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634710028767586, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634710028767586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45324450731277466, + 0.1575755923986435, + 0.05319809168577194, + 0.05307702347636223, + 0.059246644377708435, + 0.04658331349492073, + 0.052750326693058014, + 0.051088761538267136, + 0.03908252716064453, + 0.03415321931242943 + ], + "credit_map": [ + 0.02377336286008358, + 0.027453836053609848, + 0.0892840102314949, + 0.07366753369569778, + 0.03940489515662193, + 0.042719755321741104, + 0.0779639184474945, + 0.07897721230983734, + 0.008918719366192818, + 0.009456614963710308, + 0.026519978418946266, + 0.02853974513709545, + 0.01777568645775318, + 0.014836826361715794, + 0.027144722640514374, + 0.024383310228586197, + 0.0040322281420230865, + 0.005553028080612421, + 0.007429197896271944, + 0.009903675876557827, + 0.005607160273939371, + 0.004001640249043703, + 0.007277659606188536, + 0.00939350575208664, + 0.004970400128513575, + 0.0035123543348163366, + 0.010973741300404072, + 0.007098651025444269, + 0.0047401986084878445, + 0.003762178821489215, + 0.006803040858358145, + 0.011216461658477783, + 0.004351173527538776, + 0.003273814218118787, + 0.008766411803662777, + 0.008920116350054741, + 0.005568811669945717, + 0.0050657219253480434, + 0.01355341449379921, + 0.009747187606990337, + 0.0033210916444659233, + 0.003748428076505661, + 0.005588127765804529, + 0.007364214863628149, + 0.004986475687474012, + 0.004919635597616434, + 0.009870709851384163, + 0.00678463326767087, + 0.004084547981619835, + 0.004662994295358658, + 0.009739474393427372, + 0.00952577218413353, + 0.003430851735174656, + 0.0033592169638723135, + 0.00996093638241291, + 0.007986539974808693, + 0.003638880793005228, + 0.00470893457531929, + 0.007177561055868864, + 0.00672350637614727, + 0.0040008011274039745, + 0.003970494493842125, + 0.01274664793163538, + 0.008121942169964314, + 0.004288624972105026, + 0.002661849604919553, + 0.0046491315588355064, + 0.005796870682388544, + 0.0033004095312207937, + 0.003279783995822072, + 0.009148777462542057, + 0.005957084242254496, + 0.0031935633160173893, + 0.002690761350095272, + 0.003516824683174491, + 0.0060209473595023155, + 0.002958297962322831, + 0.0032207041513174772, + 0.007220987230539322, + 0.0053311362862586975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.004634710028767586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_8/index_tensor([939, 939])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_8/index_tensor([939, 939])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..be6d1732cb48f1f19dddb720fb04864b167f8872 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_8/index_tensor([939, 939])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle drone smoothly orbits and tracks a mountain bike as it descends a steep, powder-covered slope. The riders, wearing vibrant red weatherproof jackets and black pants with visible kneepads, maintain an aggressive, low-slung stance. Their full-face helmets are a contrasting matte black, with mirrored visors that reflect the brilliant blue sky. This full-suspension mountain bike, with its chunky, knobby tires and dark carbon fiber frame, plunges decisively into the deep, untouched snow. As the rider expertly executes a sharp, flowing turn, the rear wheel pops out, releasing a massive, arcing plume of fine powder that explodes into the air. A bright, crisp winter sun, low in the sky, backlights the spray, making individual ice formations shimmer against the vast, pristine white slopes. The entire scene is imbued with a vibrant, thrilling atmosphere, capturing a moment of pure adrenaline and controlled chaos amidst the serene, epic beauty of the alpine wilderness.", + "seed": 8184, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_8/index_939_seed_8184_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06499572359025478, + "vlm_logprob": -0.06499572359025478, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.06499572359025478, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13021206855773926 + ], + "opposite_logprob_tokens": [ + -2.7552120685577393 + ], + "logprob_reward": -0.13021206855773926, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.13021206855773926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2212870866060257, + 0.10767693817615509, + 0.04078252613544464, + 0.17665505409240723, + 0.050328467041254044, + 0.07254685461521149, + 0.0720156654715538, + 0.02795250155031681, + 0.09163715690374374, + 0.13911771774291992 + ], + "credit_map": [ + 0.023530079051852226, + 0.018926629796624184, + 0.04009217396378517, + 0.01995239593088627, + 0.02155289053916931, + 0.04179355129599571, + 0.02300037443637848, + 0.03243899717926979, + 0.006518673617392778, + 0.01027696393430233, + 0.012364714406430721, + 0.008338325656950474, + 0.014865700155496597, + 0.028732161968946457, + 0.012798243202269077, + 0.01378215104341507, + 0.0041588069871068, + 0.006472534965723753, + 0.0021475711837410927, + 0.0026895664632320404, + 0.009133294224739075, + 0.0077290902845561504, + 0.0034365691244602203, + 0.005015090107917786, + 0.00972533319145441, + 0.011402714066207409, + 0.013391852378845215, + 0.010489190928637981, + 0.05030601844191551, + 0.054903749376535416, + 0.011360282078385353, + 0.015075907111167908, + 0.005276785232126713, + 0.004564822185784578, + 0.004436742980033159, + 0.0037783137522637844, + 0.006926658563315868, + 0.01169811561703682, + 0.003704130882397294, + 0.009942902252078056, + 0.004098084755241871, + 0.005789976101368666, + 0.003082282142713666, + 0.0020557919051498175, + 0.02240176498889923, + 0.013233702629804611, + 0.003660043003037572, + 0.018225213512778282, + 0.003666348522529006, + 0.006201126612722874, + 0.00476450240239501, + 0.0049814945086836815, + 0.010403193533420563, + 0.023469051346182823, + 0.007658763322979212, + 0.010871184058487415, + 0.0028486372902989388, + 0.002855929546058178, + 0.003623907221481204, + 0.0033317063935101032, + 0.00316412840038538, + 0.005564760882407427, + 0.0034532237332314253, + 0.003110206685960293, + 0.009275155141949654, + 0.012745747342705727, + 0.007677828427404165, + 0.008947256952524185, + 0.014075466431677341, + 0.014920874498784542, + 0.012655150145292282, + 0.01133967749774456, + 0.015250887721776962, + 0.012253345921635628, + 0.015923360362648964, + 0.012979472056031227, + 0.014917261898517609, + 0.024179687723517418, + 0.026713304221630096, + 0.016900410875678062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.13021206855773926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033731523901224136 + ], + "opposite_logprob_tokens": [ + -3.408731460571289 + ], + "logprob_reward": -0.033731523901224136, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033731523901224136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17483751475811005, + 0.06021634116768837, + 0.07427287101745605, + 0.11933530867099762, + 0.08907024562358856, + 0.13370433449745178, + 0.08659110963344574, + 0.11498011648654938, + 0.057619884610176086, + 0.08937229961156845 + ], + "credit_map": [ + 0.01570827327668667, + 0.015590092167258263, + 0.04387304559350014, + 0.011653954163193703, + 0.01973133347928524, + 0.03167220950126648, + 0.01154148019850254, + 0.025067130103707314, + 0.004697865806519985, + 0.008297118358314037, + 0.008542317897081375, + 0.006495296489447355, + 0.006925934925675392, + 0.006676778197288513, + 0.010767582803964615, + 0.007813449017703533, + 0.0059226639568805695, + 0.010723968967795372, + 0.005177277140319347, + 0.006738049443811178, + 0.015245755203068256, + 0.013003586791455746, + 0.00922273937612772, + 0.008238828741014004, + 0.010288194753229618, + 0.014118099585175514, + 0.010303648188710213, + 0.0062379976734519005, + 0.02599102444946766, + 0.02959473617374897, + 0.008951347321271896, + 0.013850261457264423, + 0.009394219145178795, + 0.009750640019774437, + 0.007600501645356417, + 0.007204880937933922, + 0.013216539286077023, + 0.021579839289188385, + 0.008437057957053185, + 0.011886571533977985, + 0.00522179901599884, + 0.008966516703367233, + 0.004639506805688143, + 0.004600294400006533, + 0.04331611841917038, + 0.014911685138940811, + 0.006273207254707813, + 0.045775216072797775, + 0.00809127651154995, + 0.006338376086205244, + 0.008071952499449253, + 0.006026138551533222, + 0.011072062887251377, + 0.026517149060964584, + 0.009359939955174923, + 0.011114205233752728, + 0.01947268284857273, + 0.012085228227078915, + 0.01852269098162651, + 0.011257763020694256, + 0.01085888221859932, + 0.01838800124824047, + 0.012559116818010807, + 0.011835753917694092, + 0.004074875731021166, + 0.006013302598148584, + 0.004936089739203453, + 0.004977549891918898, + 0.009172128513455391, + 0.011775929480791092, + 0.008957265876233578, + 0.007712748367339373, + 0.006812770385295153, + 0.0067418161779642105, + 0.010141098871827126, + 0.007520781364291906, + 0.011149921454489231, + 0.019390085712075233, + 0.015910442918539047, + 0.011705382727086544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.033731523901224136 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706999480724335 + ], + "opposite_logprob_tokens": [ + -2.1270699501037598 + ], + "logprob_reward": -0.12706999480724335, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706999480724335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10471092909574509, + 0.12224724143743515, + 0.053971998393535614, + 0.14041128754615784, + 0.05035458505153656, + 0.09311149269342422, + 0.14727206528186798, + 0.15534424781799316, + 0.059002283960580826, + 0.07357383519411087 + ], + "credit_map": [ + 0.008937080390751362, + 0.0073745292611420155, + 0.02829013019800186, + 0.006690483074635267, + 0.010862216353416443, + 0.020690811797976494, + 0.006009494885802269, + 0.015856193378567696, + 0.012368791736662388, + 0.009804400615394115, + 0.016949305310845375, + 0.013673193752765656, + 0.012954894453287125, + 0.02485000155866146, + 0.015338563360273838, + 0.016308091580867767, + 0.00467165419831872, + 0.007362397387623787, + 0.005319723393768072, + 0.005926288664340973, + 0.00844863336533308, + 0.007161737885326147, + 0.010612620040774345, + 0.004468944855034351, + 0.011703573167324066, + 0.014671510085463524, + 0.014797002077102661, + 0.012082912027835846, + 0.025384392589330673, + 0.030359618365764618, + 0.01382417418062687, + 0.017588112503290176, + 0.006474365945905447, + 0.0034050983376801014, + 0.00639055110514164, + 0.004279310815036297, + 0.006350476760417223, + 0.009392899461090565, + 0.006008551921695471, + 0.00805333349853754, + 0.00809431355446577, + 0.008921463042497635, + 0.006732958368957043, + 0.004129158332943916, + 0.018720170482993126, + 0.015872566029429436, + 0.00865829549729824, + 0.021982567384839058, + 0.014404804445803165, + 0.008844759315252304, + 0.013780441135168076, + 0.00952139776200056, + 0.017098406329751015, + 0.04872306063771248, + 0.014413597993552685, + 0.020485613495111465, + 0.027665071189403534, + 0.016181284561753273, + 0.02634047344326973, + 0.017522547394037247, + 0.01514288317412138, + 0.021183546632528305, + 0.01559761818498373, + 0.01571083441376686, + 0.003990956582129002, + 0.004728360567241907, + 0.005371939856559038, + 0.006506582722067833, + 0.007933414541184902, + 0.012029970064759254, + 0.010005516931414604, + 0.008435552008450031, + 0.004702287260442972, + 0.003924942575395107, + 0.005087230820208788, + 0.005937415640801191, + 0.011172842234373093, + 0.014492790214717388, + 0.01727668009698391, + 0.01097965333610773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.12706999480724335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628239691257477 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628239691257477, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628239691257477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17746217548847198, + 0.06266508251428604, + 0.061554037034511566, + 0.13778743147850037, + 0.14911912381649017, + 0.07104789465665817, + 0.05052073299884796, + 0.030968377366662025, + 0.13957886397838593, + 0.11929627507925034 + ], + "credit_map": [ + 0.01681179367005825, + 0.01562962681055069, + 0.0320914164185524, + 0.014695591293275356, + 0.018448874354362488, + 0.03392847254872322, + 0.01913485676050186, + 0.026721559464931488, + 0.0037788893096148968, + 0.006205128505825996, + 0.00967995636165142, + 0.005695090629160404, + 0.007998905144631863, + 0.014484898187220097, + 0.006710539106279612, + 0.008111677132546902, + 0.006673973985016346, + 0.009211165830492973, + 0.005293406546115875, + 0.004558266140520573, + 0.010593228973448277, + 0.008584470488131046, + 0.006561914924532175, + 0.01007760502398014, + 0.011247237212955952, + 0.008632581681013107, + 0.015047729015350342, + 0.01632736809551716, + 0.029812119901180267, + 0.031340111047029495, + 0.012402815744280815, + 0.012977471575140953, + 0.01639426127076149, + 0.018274834379553795, + 0.013957519084215164, + 0.009851676411926746, + 0.021035904064774513, + 0.03227338194847107, + 0.012306937016546726, + 0.02502460964024067, + 0.0038010780699551105, + 0.0069179777055978775, + 0.00307668698951602, + 0.0032562552951276302, + 0.017589852213859558, + 0.014812016859650612, + 0.005658027250319719, + 0.015936002135276794, + 0.002557812724262476, + 0.003265540348365903, + 0.00371585157699883, + 0.0036194515414536, + 0.006816658191382885, + 0.015815235674381256, + 0.007700961083173752, + 0.007029220927506685, + 0.0030662643257528543, + 0.003269738517701626, + 0.004004796035587788, + 0.0034250598400831223, + 0.0031936862505972385, + 0.006085813976824284, + 0.004005047958344221, + 0.003917966969311237, + 0.006911050528287888, + 0.011344453319907188, + 0.008362903259694576, + 0.011331277899444103, + 0.025225594639778137, + 0.03420184180140495, + 0.02253054827451706, + 0.01967119798064232, + 0.010700038634240627, + 0.009312247857451439, + 0.010789691470563412, + 0.00983948353677988, + 0.015447149984538555, + 0.02170482836663723, + 0.024542195722460747, + 0.016960639506578445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.007628239691257477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263367909938097 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.0263367909938097, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263367909938097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13078747689723969, + 0.0593595951795578, + 0.08686723560094833, + 0.1940091848373413, + 0.11114205420017242, + 0.11055253446102142, + 0.0881752073764801, + 0.02533390372991562, + 0.09891864657402039, + 0.09485413134098053 + ], + "credit_map": [ + 0.013362471014261246, + 0.009963472373783588, + 0.0273794773966074, + 0.008024689741432667, + 0.015020348131656647, + 0.030853131785988808, + 0.008050068281590939, + 0.01813380979001522, + 0.004621696192771196, + 0.008945099078118801, + 0.00946654099971056, + 0.004835892468690872, + 0.007431522011756897, + 0.010870235972106457, + 0.00538402795791626, + 0.007804577238857746, + 0.005372100044041872, + 0.007755720522254705, + 0.004397324752062559, + 0.003927797079086304, + 0.02099066972732544, + 0.02080538496375084, + 0.007758894003927708, + 0.015859344974160194, + 0.015575295314192772, + 0.01467946171760559, + 0.013405845500528812, + 0.014704471454024315, + 0.04985492303967476, + 0.046174485236406326, + 0.017642125487327576, + 0.021972591057419777, + 0.012685014866292477, + 0.015862954780459404, + 0.005744431167840958, + 0.007250557187944651, + 0.015393096953630447, + 0.024982357397675514, + 0.007827382534742355, + 0.021396253257989883, + 0.00858108140528202, + 0.011469035409390926, + 0.006679480895400047, + 0.007885721512138844, + 0.016315821558237076, + 0.02768382988870144, + 0.013202977366745472, + 0.018734578043222427, + 0.006234045140445232, + 0.005493621341884136, + 0.008159380406141281, + 0.007477123755961657, + 0.01198050007224083, + 0.024066492915153503, + 0.013445625081658363, + 0.011318420991301537, + 0.001543288817629218, + 0.0017700070748105645, + 0.0024266273248940706, + 0.0023205257020890713, + 0.004108686465770006, + 0.006508576683700085, + 0.003208101261407137, + 0.003448090748861432, + 0.005102337803691626, + 0.008626865223050117, + 0.005623237229883671, + 0.0064150867983698845, + 0.015540426596999168, + 0.024873096495866776, + 0.016427231952548027, + 0.016310371458530426, + 0.005039821844547987, + 0.004254459403455257, + 0.005692456848919392, + 0.008206026628613472, + 0.014046620577573776, + 0.016256405040621758, + 0.025964032858610153, + 0.015394304879009724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0263367909938097 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle drone smoothly orbits and tracks a mountain bike as it descends a steep, powder-covered slope. The riders, wearing vibrant red weatherproof jackets and black pants with visible kneepads, maintain an aggressive, low-slung stance. Their full-face helmets are a contrasting matte black, with mirrored visors that reflect the brilliant blue sky. This full-suspension mountain bike, with its chunky, knobby tires and dark carbon fiber frame, plunges decisively into the deep, untouched snow. As the rider expertly executes a sharp, flowing turn, the rear wheel pops out, releasing a massive, arcing plume of fine powder that explodes into the air. A bright, crisp winter sun, low in the sky, backlights the spray, making individual ice formations shimmer against the vast, pristine white slopes. The entire scene is imbued with a vibrant, thrilling atmosphere, capturing a moment of pure adrenaline and controlled chaos amidst the serene, epic beauty of the alpine wilderness.", + "seed": 8185, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_8/index_939_seed_8185_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07944507002830506, + "vlm_logprob": -0.07944507002830506, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.07944507002830506, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09244822710752487 + ], + "opposite_logprob_tokens": [ + -2.9674482345581055 + ], + "logprob_reward": -0.09244822710752487, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.09244822710752487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19819313287734985, + 0.07234439253807068, + 0.040723610669374466, + 0.04197343438863754, + 0.0844557136297226, + 0.19064974784851074, + 0.16912704706192017, + 0.05326037108898163, + 0.08567903935909271, + 0.0635935515165329 + ], + "credit_map": [ + 0.013637463562190533, + 0.014470057561993599, + 0.03446059674024582, + 0.0539824552834034, + 0.011107027530670166, + 0.011884193867444992, + 0.030401382595300674, + 0.02824992686510086, + 0.008060944266617298, + 0.005799286533147097, + 0.014506709761917591, + 0.013706986792385578, + 0.005408598575741053, + 0.004983899183571339, + 0.01128972414880991, + 0.008588239550590515, + 0.0040867323987185955, + 0.00623464398086071, + 0.007760254666209221, + 0.0062556485645473, + 0.004311949014663696, + 0.0026857545599341393, + 0.0040274993516504765, + 0.005361125338822603, + 0.004248716868460178, + 0.004932555835694075, + 0.006202035117894411, + 0.0070486655458807945, + 0.004229070618748665, + 0.00478505901992321, + 0.005032343789935112, + 0.005494986195117235, + 0.011495896615087986, + 0.02184627391397953, + 0.009469239972531796, + 0.006662625353783369, + 0.009889605455100536, + 0.006576126907020807, + 0.009205994196236134, + 0.00930994376540184, + 0.023627610877156258, + 0.01824190840125084, + 0.02782789245247841, + 0.019431158900260925, + 0.015885520726442337, + 0.04005982354283333, + 0.022396322339773178, + 0.023179488256573677, + 0.036670517176389694, + 0.04952483996748924, + 0.03390013426542282, + 0.013442259281873703, + 0.011352202855050564, + 0.008052995428442955, + 0.007146312855184078, + 0.009037773124873638, + 0.0072235423140227795, + 0.006335339974611998, + 0.009186971932649612, + 0.00474263122305274, + 0.004982287995517254, + 0.009859084151685238, + 0.004260311834514141, + 0.006670198868960142, + 0.011483912356197834, + 0.006964234169572592, + 0.019510673359036446, + 0.009187508374452591, + 0.007953875698149204, + 0.014170372858643532, + 0.006757248193025589, + 0.009651212953031063, + 0.00907234475016594, + 0.0057367864064872265, + 0.011318081989884377, + 0.00654182443395257, + 0.00821809470653534, + 0.009347465820610523, + 0.0073479702696204185, + 0.006010984070599079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.09244822710752487 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371873125433922 + ], + "opposite_logprob_tokens": [ + -3.4087188243865967 + ], + "logprob_reward": -0.03371873125433922, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371873125433922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1665606051683426, + 0.07608196884393692, + 0.08878830075263977, + 0.097327321767807, + 0.05839725211262703, + 0.09317217022180557, + 0.10684288293123245, + 0.09502213448286057, + 0.11474177986383438, + 0.10306552797555923 + ], + "credit_map": [ + 0.02072434313595295, + 0.01685328595340252, + 0.028611671179533005, + 0.02392461523413658, + 0.012281685136258602, + 0.02082548476755619, + 0.01860756240785122, + 0.024731958284974098, + 0.0048697651363909245, + 0.005644415505230427, + 0.012122489511966705, + 0.01591607928276062, + 0.006873403675854206, + 0.0063025811687111855, + 0.01380043476819992, + 0.010552803054451942, + 0.011354872956871986, + 0.01151270791888237, + 0.014256906695663929, + 0.015562986023724079, + 0.009200053289532661, + 0.0085631487891078, + 0.007473587524145842, + 0.010864031501114368, + 0.009660070762038231, + 0.01187802478671074, + 0.015415600501000881, + 0.01619074121117592, + 0.011461952701210976, + 0.007382736075669527, + 0.009854566305875778, + 0.015483630821108818, + 0.007305252831429243, + 0.012904196046292782, + 0.007174389436841011, + 0.0050499760545790195, + 0.006961511448025703, + 0.004296963568776846, + 0.007666897028684616, + 0.007038063835352659, + 0.01294025033712387, + 0.010399861261248589, + 0.011872339062392712, + 0.00946593377739191, + 0.011772672645747662, + 0.015689503401517868, + 0.010985154658555984, + 0.010046455077826977, + 0.013875243254005909, + 0.012981807813048363, + 0.015784241259098053, + 0.014674944803118706, + 0.01331287994980812, + 0.011355205439031124, + 0.010632951743900776, + 0.014225608669221401, + 0.011834139935672283, + 0.012989629060029984, + 0.014115544967353344, + 0.009145921096205711, + 0.020617561414837837, + 0.007287407293915749, + 0.0095352279022336, + 0.009496706537902355, + 0.014318881556391716, + 0.013631178066134453, + 0.023253098130226135, + 0.011863994412124157, + 0.016051102429628372, + 0.012030905112624168, + 0.010244891978800297, + 0.013347725383937359, + 0.014764217659831047, + 0.008907574228942394, + 0.01791677065193653, + 0.006581800524145365, + 0.01427981536835432, + 0.018863743171095848, + 0.012731716968119144, + 0.009019894525408745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.03371873125433922 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038617491722107 + ], + "opposite_logprob_tokens": [ + -1.9103862047195435 + ], + "logprob_reward": -0.16038617491722107, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038617491722107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14556056261062622, + 0.06293176114559174, + 0.050002917647361755, + 0.06401520222425461, + 0.04751570522785187, + 0.17211613059043884, + 0.2553974986076355, + 0.09252665936946869, + 0.0295023862272501, + 0.08043115586042404 + ], + "credit_map": [ + 0.012004924938082695, + 0.012010782025754452, + 0.02481633611023426, + 0.03349979594349861, + 0.00960739515721798, + 0.012809633277356625, + 0.017923010513186455, + 0.022888697683811188, + 0.005577719304710627, + 0.006674655247479677, + 0.009766793809831142, + 0.007053988054394722, + 0.006123623810708523, + 0.007153478916734457, + 0.00801409874111414, + 0.012567410245537758, + 0.006220844108611345, + 0.006500469520688057, + 0.00832982175052166, + 0.008000615984201431, + 0.006568996701389551, + 0.00431912811473012, + 0.0042365300469100475, + 0.005826512351632118, + 0.006948141381144524, + 0.007124691735953093, + 0.009402688592672348, + 0.009860379621386528, + 0.007745447568595409, + 0.005002250429242849, + 0.00679808109998703, + 0.011133521795272827, + 0.005063354969024658, + 0.008984694257378578, + 0.0052454168908298016, + 0.004376143682748079, + 0.006445931736379862, + 0.005122967530041933, + 0.005997087340801954, + 0.006280111614614725, + 0.023647382855415344, + 0.013972151093184948, + 0.024562010541558266, + 0.017834624275565147, + 0.014440199360251427, + 0.03767218813300133, + 0.019338490441441536, + 0.020649097859859467, + 0.06518030166625977, + 0.08802954107522964, + 0.05164702609181404, + 0.018590066581964493, + 0.009396467357873917, + 0.0063133854418993, + 0.006119521800428629, + 0.010121205821633339, + 0.011080453172326088, + 0.013482937589287758, + 0.01520940475165844, + 0.008563379757106304, + 0.01957264356315136, + 0.008593738079071045, + 0.008336883969604969, + 0.0076872240751981735, + 0.00447991443797946, + 0.0039049703627824783, + 0.005278463941067457, + 0.0028746805619448423, + 0.004060171078890562, + 0.0029491891618818045, + 0.002620976883918047, + 0.00333402375690639, + 0.01172372605651617, + 0.006218917667865753, + 0.01262463815510273, + 0.006558799184858799, + 0.011143101379275322, + 0.014417732134461403, + 0.01027613040059805, + 0.007468118332326412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.16038617491722107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864353686571121 + ], + "opposite_logprob_tokens": [ + -3.0486435890197754 + ], + "logprob_reward": -0.04864353686571121, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864353686571121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14349687099456787, + 0.1085154339671135, + 0.1698317676782608, + 0.041348107159137726, + 0.042443834245204926, + 0.09670597314834595, + 0.09130347520112991, + 0.0946502536535263, + 0.09756962209939957, + 0.11413460969924927 + ], + "credit_map": [ + 0.014881384558975697, + 0.012184740044176579, + 0.019840097054839134, + 0.025231007486581802, + 0.008851482532918453, + 0.012794403359293938, + 0.023054907098412514, + 0.026658862829208374, + 0.006732181645929813, + 0.009052165783941746, + 0.018250880762934685, + 0.023548828437924385, + 0.009312517009675503, + 0.007415471598505974, + 0.02178632654249668, + 0.012417064048349857, + 0.028488818556070328, + 0.025563625618815422, + 0.028314510360360146, + 0.027910178527235985, + 0.014336058869957924, + 0.014274639077484608, + 0.01213932316750288, + 0.018804604187607765, + 0.004201325587928295, + 0.00480377022176981, + 0.0064614214934408665, + 0.007674779277294874, + 0.0041191005147993565, + 0.0035670469515025616, + 0.004838271532207727, + 0.005682394374161959, + 0.004980110563337803, + 0.008155087940394878, + 0.00483290571719408, + 0.0035305593628436327, + 0.0049692727625370026, + 0.003831456881016493, + 0.006641827058047056, + 0.005502612330019474, + 0.016079826280474663, + 0.010327959433197975, + 0.013113204389810562, + 0.010041496716439724, + 0.01089926715940237, + 0.011202332563698292, + 0.01329314149916172, + 0.011748741380870342, + 0.013754529878497124, + 0.01299776416271925, + 0.017176436260342598, + 0.010394219309091568, + 0.011019466444849968, + 0.008080068975687027, + 0.006768787745386362, + 0.011112203821539879, + 0.014043083414435387, + 0.010334421880543232, + 0.015412291511893272, + 0.00872018001973629, + 0.016850121319293976, + 0.010316316969692707, + 0.008726882748305798, + 0.010246961377561092, + 0.013648061081767082, + 0.007813045755028725, + 0.02477344125509262, + 0.009501627646386623, + 0.012019293382763863, + 0.011731662787497044, + 0.006603165064007044, + 0.011479327455163002, + 0.016143623739480972, + 0.01051903236657381, + 0.01979292370378971, + 0.012832061387598515, + 0.00952302198857069, + 0.01602860353887081, + 0.018088947981595993, + 0.011206399649381638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.04864353686571121 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620286799967289 + ], + "opposite_logprob_tokens": [ + -2.812028646469116 + ], + "logprob_reward": -0.0620286799967289, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620286799967289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18507710099220276, + 0.09368136525154114, + 0.143677219748497, + 0.06323782354593277, + 0.03150094300508499, + 0.0690024197101593, + 0.10653530061244965, + 0.14564202725887299, + 0.04227948188781738, + 0.11936628818511963 + ], + "credit_map": [ + 0.013284585438668728, + 0.012918708845973015, + 0.03020198456943035, + 0.04106194153428078, + 0.011424791067838669, + 0.014567708596587181, + 0.029143333435058594, + 0.03247404471039772, + 0.009239993058145046, + 0.008592737838625908, + 0.022665195167064667, + 0.019229909405112267, + 0.0060182553716003895, + 0.0049455491825938225, + 0.013641349971294403, + 0.009348375722765923, + 0.018953366205096245, + 0.02120889350771904, + 0.02737978659570217, + 0.02388308383524418, + 0.011363529600203037, + 0.013415791094303131, + 0.011125260032713413, + 0.016347505152225494, + 0.005407477263361216, + 0.006891155615448952, + 0.011710746213793755, + 0.014336909167468548, + 0.004622177220880985, + 0.004874439910054207, + 0.0051216185092926025, + 0.010273302905261517, + 0.004307569470256567, + 0.005827378015965223, + 0.0041524977423250675, + 0.003163907676935196, + 0.0034739102702587843, + 0.0029917338397353888, + 0.004610390402376652, + 0.002973558148369193, + 0.007044657599180937, + 0.008548218756914139, + 0.010553088039159775, + 0.005361148621886969, + 0.006412212736904621, + 0.014220882207155228, + 0.009003742597997189, + 0.00785847194492817, + 0.028327083215117455, + 0.03143724799156189, + 0.016865354031324387, + 0.007926619611680508, + 0.005727469455450773, + 0.004971652757376432, + 0.005099480971693993, + 0.0061803958378732204, + 0.019231164827942848, + 0.01485272217541933, + 0.01768084429204464, + 0.016873996704816818, + 0.01802765391767025, + 0.01715063489973545, + 0.017756620422005653, + 0.024068396538496017, + 0.006500546354800463, + 0.0028652315959334373, + 0.010499880649149418, + 0.0034730518236756325, + 0.004038281738758087, + 0.007214860524982214, + 0.002531622536480427, + 0.005156006198376417, + 0.01657959260046482, + 0.009590975940227509, + 0.02066398411989212, + 0.008377094753086567, + 0.015975581482052803, + 0.023089734837412834, + 0.01487897802144289, + 0.010210346430540085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0620286799967289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle drone smoothly orbits and tracks a mountain bike as it descends a steep, powder-covered slope. The riders, wearing vibrant red weatherproof jackets and black pants with visible kneepads, maintain an aggressive, low-slung stance. Their full-face helmets are a contrasting matte black, with mirrored visors that reflect the brilliant blue sky. This full-suspension mountain bike, with its chunky, knobby tires and dark carbon fiber frame, plunges decisively into the deep, untouched snow. As the rider expertly executes a sharp, flowing turn, the rear wheel pops out, releasing a massive, arcing plume of fine powder that explodes into the air. A bright, crisp winter sun, low in the sky, backlights the spray, making individual ice formations shimmer against the vast, pristine white slopes. The entire scene is imbued with a vibrant, thrilling atmosphere, capturing a moment of pure adrenaline and controlled chaos amidst the serene, epic beauty of the alpine wilderness.", + "seed": 8186, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_8/index_939_seed_8186_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.061275864765048024, + "vlm_logprob": -0.061275864765048024, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.061275864765048024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13582691550254822 + ], + "opposite_logprob_tokens": [ + -3.135826826095581 + ], + "logprob_reward": -0.13582691550254822, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.13582691550254822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11381397396326065, + 0.03645818307995796, + 0.03148578107357025, + 0.05780754238367081, + 0.4631862938404083, + 0.0601595975458622, + 0.08506790548563004, + 0.056079376488924026, + 0.04615377262234688, + 0.04978759214282036 + ], + "credit_map": [ + 0.010143553838133812, + 0.014358851127326488, + 0.01569168083369732, + 0.012171825394034386, + 0.01718512363731861, + 0.01671055518090725, + 0.013361241668462753, + 0.014191142283380032, + 0.00412031402811408, + 0.005062259268015623, + 0.004747255239635706, + 0.004881974775344133, + 0.004810551647096872, + 0.004583724774420261, + 0.004704596009105444, + 0.0035475066397339106, + 0.0033761279191821814, + 0.0044072652235627174, + 0.0036828627344220877, + 0.003545014886185527, + 0.0035514170303940773, + 0.0057739028707146645, + 0.004165158607065678, + 0.002984030172228813, + 0.007294530048966408, + 0.006369386799633503, + 0.007034022826701403, + 0.006028687581419945, + 0.008151628077030182, + 0.0081074433401227, + 0.008435125462710857, + 0.006386716850101948, + 0.024225052446126938, + 0.031182510778307915, + 0.009446143172681332, + 0.011084900237619877, + 0.12192655354738235, + 0.04070276767015457, + 0.08183036744594574, + 0.14278799295425415, + 0.003805378917604685, + 0.004589870572090149, + 0.004423218779265881, + 0.0058434270322322845, + 0.009089154191315174, + 0.012136905454099178, + 0.010645575821399689, + 0.009626067243516445, + 0.005671216640621424, + 0.006892419885843992, + 0.004863109439611435, + 0.005375376436859369, + 0.016182269901037216, + 0.021560437977313995, + 0.012163082137703896, + 0.012359987944364548, + 0.011955788359045982, + 0.006613017991185188, + 0.004145768936723471, + 0.002554374048486352, + 0.013302757404744625, + 0.012797790579497814, + 0.002332663396373391, + 0.0023772178683429956, + 0.00742062134668231, + 0.004592602141201496, + 0.00448937714099884, + 0.003456499194726348, + 0.009788772091269493, + 0.00740647129714489, + 0.0050196112133562565, + 0.00397981982678175, + 0.004848639480769634, + 0.004932406824082136, + 0.008817272260785103, + 0.008404591120779514, + 0.004662472754716873, + 0.005420773755759001, + 0.005918918643146753, + 0.006782514974474907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.13582691550254822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330787293612957 + ], + "opposite_logprob_tokens": [ + -3.7733078002929688 + ], + "logprob_reward": -0.02330787293612957, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330787293612957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12880823016166687, + 0.05549319088459015, + 0.04884133115410805, + 0.0925590842962265, + 0.4793727397918701, + 0.03631112352013588, + 0.03812788426876068, + 0.048652634024620056, + 0.0295573603361845, + 0.04227635636925697 + ], + "credit_map": [ + 0.0108787314966321, + 0.018747767433524132, + 0.022163912653923035, + 0.014853672124445438, + 0.014732927083969116, + 0.020077787339687347, + 0.013020453043282032, + 0.014332996681332588, + 0.006090175826102495, + 0.006278845481574535, + 0.01111353188753128, + 0.007883919402956963, + 0.0052247969433665276, + 0.006614712532609701, + 0.005766190122812986, + 0.0065210177563130856, + 0.005983226466923952, + 0.0073045967146754265, + 0.006719289813190699, + 0.003948282450437546, + 0.008281276561319828, + 0.008121930994093418, + 0.0042472765780985355, + 0.00423545204102993, + 0.014696257188916206, + 0.007222354877740145, + 0.013419057242572308, + 0.009832766838371754, + 0.0137174716219306, + 0.011681335046887398, + 0.012636707164347172, + 0.009353133849799633, + 0.025472432374954224, + 0.03080989420413971, + 0.009839196689426899, + 0.012343963608145714, + 0.12454614788293839, + 0.04152951017022133, + 0.08558071404695511, + 0.14925089478492737, + 0.003776025725528598, + 0.004718209616839886, + 0.003940535243600607, + 0.003749543335288763, + 0.00450443010777235, + 0.005670848768204451, + 0.0058204494416713715, + 0.004131079651415348, + 0.00359632377512753, + 0.0033686005044728518, + 0.002539740176871419, + 0.0029248171485960484, + 0.0057965111918747425, + 0.006358003243803978, + 0.00608703913167119, + 0.007456852123141289, + 0.011146296747028828, + 0.004551910795271397, + 0.006043539848178625, + 0.0036226955708116293, + 0.006658982019871473, + 0.007391073741018772, + 0.004808938130736351, + 0.0044292001985013485, + 0.0033011154737323523, + 0.004523353185504675, + 0.003577592084184289, + 0.0030843771528452635, + 0.004283791407942772, + 0.004318494349718094, + 0.003592228749766946, + 0.0028764079324901104, + 0.004204911645501852, + 0.005647694692015648, + 0.00593964708968997, + 0.009335828013718128, + 0.004855544306337833, + 0.0038296023849397898, + 0.004328873474150896, + 0.004134257324039936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.02330787293612957 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033383220434189 + ], + "opposite_logprob_tokens": [ + -2.3503339290618896 + ], + "logprob_reward": -0.10033383220434189, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033383220434189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21790780127048492, + 0.07756589353084564, + 0.09262600541114807, + 0.09540406614542007, + 0.08895184099674225, + 0.07016219943761826, + 0.04180707409977913, + 0.05773571878671646, + 0.07921241968870163, + 0.17862698435783386 + ], + "credit_map": [ + 0.018922368064522743, + 0.03195205330848694, + 0.040753670036792755, + 0.02328048087656498, + 0.02897137962281704, + 0.03039558418095112, + 0.021641818806529045, + 0.021990453824400902, + 0.007011655252426863, + 0.007918440736830235, + 0.014266322366893291, + 0.011321615427732468, + 0.009217323735356331, + 0.009393205866217613, + 0.008967188186943531, + 0.009470142424106598, + 0.011593211442232132, + 0.011270551942288876, + 0.011099754832684994, + 0.0057135410606861115, + 0.014244813472032547, + 0.020826444029808044, + 0.009390812367200851, + 0.008486873470246792, + 0.013105369172990322, + 0.010531904175877571, + 0.016592631116509438, + 0.012071000412106514, + 0.010366326197981834, + 0.010651827789843082, + 0.01217461097985506, + 0.00991040002554655, + 0.005693511571735144, + 0.007792256306856871, + 0.0030882004648447037, + 0.004650259390473366, + 0.02016718126833439, + 0.008078242652118206, + 0.014559296891093254, + 0.024922896176576614, + 0.007970861159265041, + 0.009009761735796928, + 0.007047259248793125, + 0.0073909093625843525, + 0.007741983979940414, + 0.01064987014979124, + 0.0114028574898839, + 0.008948690257966518, + 0.0045524947345256805, + 0.00553607614710927, + 0.0032960548996925354, + 0.00342487171292305, + 0.007109722588211298, + 0.005678954068571329, + 0.0057000648230314255, + 0.006508829537779093, + 0.010353052988648415, + 0.008307071402668953, + 0.00562382722273469, + 0.00411567185074091, + 0.01019583735615015, + 0.012449750676751137, + 0.003168487688526511, + 0.0035220179706811905, + 0.005125120282173157, + 0.011219803243875504, + 0.012552407570183277, + 0.00945120770484209, + 0.013345461338758469, + 0.011653425171971321, + 0.007978691719472408, + 0.007886306382715702, + 0.01646038331091404, + 0.018966589123010635, + 0.03492562845349312, + 0.030849358066916466, + 0.019256794825196266, + 0.01547956746071577, + 0.019673282280564308, + 0.02301538549363613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.10033383220434189 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057623490691185 + ], + "opposite_logprob_tokens": [ + -3.895576238632202 + ], + "logprob_reward": -0.02057623490691185, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057623490691185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2057749629020691, + 0.10976764559745789, + 0.05256273224949837, + 0.09520886093378067, + 0.27318817377090454, + 0.062426235526800156, + 0.035645682364702225, + 0.02768418751657009, + 0.06819941103458405, + 0.06954208761453629 + ], + "credit_map": [ + 0.02002047933638096, + 0.028958655893802643, + 0.039248451590538025, + 0.025273283943533897, + 0.025226060301065445, + 0.032184846699237823, + 0.016420049592852592, + 0.01844312995672226, + 0.008721289224922657, + 0.013818074949085712, + 0.012815942987799644, + 0.016558578237891197, + 0.01509205624461174, + 0.013792229816317558, + 0.013353285379707813, + 0.01561618223786354, + 0.006653814576566219, + 0.00707643199712038, + 0.005199812352657318, + 0.005285661201924086, + 0.010630273260176182, + 0.007314260583370924, + 0.006309382151812315, + 0.004093094263225794, + 0.013919643126428127, + 0.015149051323533058, + 0.010427181608974934, + 0.008582991547882557, + 0.014336018823087215, + 0.012207605876028538, + 0.011152006685733795, + 0.009434363804757595, + 0.01391549687832594, + 0.017763976007699966, + 0.006966925226151943, + 0.007538501638919115, + 0.07007504999637604, + 0.023939982056617737, + 0.04833703488111496, + 0.08465120941400528, + 0.0062276809476315975, + 0.005475498270243406, + 0.004648779518902302, + 0.005900393705815077, + 0.008852627128362656, + 0.013109655119478703, + 0.010149055160582066, + 0.008062541484832764, + 0.0037361786235123873, + 0.005453384481370449, + 0.0024588920641690493, + 0.0028718062676489353, + 0.0054467469453811646, + 0.007056031376123428, + 0.004937442019581795, + 0.0036852010525763035, + 0.0027287709526717663, + 0.0026464471593499184, + 0.002844453090801835, + 0.0017832174198701978, + 0.006723838858306408, + 0.006600819993764162, + 0.002342158230021596, + 0.0020144812297075987, + 0.008556991815567017, + 0.009220661595463753, + 0.011119433678686619, + 0.006347982678562403, + 0.009715077467262745, + 0.009263472631573677, + 0.007316577713936567, + 0.006659216247498989, + 0.005293488968163729, + 0.008185659535229206, + 0.010318298824131489, + 0.01866733282804489, + 0.007055474910885096, + 0.005613506305962801, + 0.0071121761575341225, + 0.0072961486876010895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.02057623490691185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633446827530861 + ], + "opposite_logprob_tokens": [ + -3.651334524154663 + ], + "logprob_reward": -0.02633446827530861, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633446827530861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25206100940704346, + 0.05593898519873619, + 0.06007847189903259, + 0.07505255937576294, + 0.3641106188297272, + 0.035314399749040604, + 0.03431091830134392, + 0.02452663891017437, + 0.051281899213790894, + 0.04732455685734749 + ], + "credit_map": [ + 0.03083134815096855, + 0.04062190279364586, + 0.0421338714659214, + 0.030507728457450867, + 0.022106794640421867, + 0.038104936480522156, + 0.020910097286105156, + 0.026844333857297897, + 0.005955996457487345, + 0.007144195958971977, + 0.007459132466465235, + 0.007126292679458857, + 0.004032404161989689, + 0.007886512205004692, + 0.005426081363111734, + 0.010908368974924088, + 0.008087941445410252, + 0.004093911498785019, + 0.008081301115453243, + 0.0038902282249182463, + 0.0056617772206664085, + 0.017717711627483368, + 0.005929435137659311, + 0.006616166327148676, + 0.01435895636677742, + 0.006004834547638893, + 0.006139611359685659, + 0.0067215426824986935, + 0.01352253183722496, + 0.010520825162529945, + 0.00962893757969141, + 0.008155319839715958, + 0.018893444910645485, + 0.02380567416548729, + 0.008431862108409405, + 0.00905698724091053, + 0.09545301645994186, + 0.03203080967068672, + 0.06412956118583679, + 0.11230926215648651, + 0.002983290236443281, + 0.003194025717675686, + 0.002820219611749053, + 0.0022951788268983364, + 0.00479742418974638, + 0.0069497013464570045, + 0.007198469713330269, + 0.005076094530522823, + 0.0028375224210321903, + 0.0024508934002369642, + 0.0022018523886799812, + 0.002110631437972188, + 0.005119195207953453, + 0.007472505792975426, + 0.006260383408516645, + 0.0058579337783157825, + 0.00274887983687222, + 0.0027396853547543287, + 0.002263635164126754, + 0.0016563715180382133, + 0.003355668857693672, + 0.004544463939964771, + 0.0037054906133562326, + 0.003512444207444787, + 0.0067359767854213715, + 0.0034938559401780367, + 0.0065939887426793575, + 0.0033117097336798906, + 0.007824020460247993, + 0.010064828209578991, + 0.0075881266966462135, + 0.00566939078271389, + 0.003885549958795309, + 0.00494229793548584, + 0.007708702702075243, + 0.007039802148938179, + 0.004937100224196911, + 0.004806723911315203, + 0.0073416149243712425, + 0.006662761326879263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.02633446827530861 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle drone smoothly orbits and tracks a mountain bike as it descends a steep, powder-covered slope. The riders, wearing vibrant red weatherproof jackets and black pants with visible kneepads, maintain an aggressive, low-slung stance. Their full-face helmets are a contrasting matte black, with mirrored visors that reflect the brilliant blue sky. This full-suspension mountain bike, with its chunky, knobby tires and dark carbon fiber frame, plunges decisively into the deep, untouched snow. As the rider expertly executes a sharp, flowing turn, the rear wheel pops out, releasing a massive, arcing plume of fine powder that explodes into the air. A bright, crisp winter sun, low in the sky, backlights the spray, making individual ice formations shimmer against the vast, pristine white slopes. The entire scene is imbued with a vibrant, thrilling atmosphere, capturing a moment of pure adrenaline and controlled chaos amidst the serene, epic beauty of the alpine wilderness.", + "seed": 8187, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_8/index_939_seed_8187_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07695148400962352, + "vlm_logprob": -0.07695148400962352, + "vlm_yes_no_margin": 2.875000023841858, + "vlm_token_credit": -0.07695148400962352, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12569543719291687 + ], + "opposite_logprob_tokens": [ + -2.5006954669952393 + ], + "logprob_reward": -0.12569543719291687, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.12569543719291687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1367475390434265, + 0.026264015585184097, + 0.012883009389042854, + 0.0265052393078804, + 0.02124571055173874, + 0.03781863674521446, + 0.036776766180992126, + 0.026802197098731995, + 0.5820249319076538, + 0.09293202310800552 + ], + "credit_map": [ + 0.013147723861038685, + 0.012973302975296974, + 0.017574511468410492, + 0.015132351778447628, + 0.013921978883445263, + 0.021613871678709984, + 0.015274119563400745, + 0.0271096620708704, + 0.00461399694904685, + 0.0032317270524799824, + 0.0034205366391688585, + 0.0030019874684512615, + 0.0024796302895992994, + 0.003212447976693511, + 0.0036923412699252367, + 0.0026113453786820173, + 0.0009242034866474569, + 0.001726000220514834, + 0.0018404001602903008, + 0.001620702212676406, + 0.0016720034182071686, + 0.0020313721615821123, + 0.0018113059923052788, + 0.0012570208637043834, + 0.003429250093176961, + 0.003246513893827796, + 0.004408052656799555, + 0.002869686344638467, + 0.002524436917155981, + 0.003858695039525628, + 0.0033126624766737223, + 0.002855938160791993, + 0.0019141681259498, + 0.0026606780011206865, + 0.0040079024620354176, + 0.003209551563486457, + 0.0015624886145815253, + 0.00203719618730247, + 0.002065066248178482, + 0.0037886560894548893, + 0.004024991765618324, + 0.004695232026278973, + 0.006658454891294241, + 0.0046708849258720875, + 0.003423823742195964, + 0.0037407036870718002, + 0.006081400439143181, + 0.004523139446973801, + 0.004389356356114149, + 0.004376481287181377, + 0.004246066324412823, + 0.005182045977562666, + 0.0035871192812919617, + 0.003400687826797366, + 0.005631804466247559, + 0.005963202565908432, + 0.0035579337272793055, + 0.004207673016935587, + 0.0026993288192898035, + 0.0035767504014074802, + 0.0018315664492547512, + 0.006079424172639847, + 0.0018583547789603472, + 0.002991161774843931, + 0.22481460869312286, + 0.2410992532968521, + 0.04320891201496124, + 0.02103421464562416, + 0.018287859857082367, + 0.01084723137319088, + 0.008279232308268547, + 0.014453588984906673, + 0.009798944927752018, + 0.017237190157175064, + 0.031850073486566544, + 0.010524840094149113, + 0.003391590900719166, + 0.009565051645040512, + 0.005474654491990805, + 0.005089670419692993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.12569543719291687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813149780035019 + ], + "opposite_logprob_tokens": [ + -3.2881314754486084 + ], + "logprob_reward": -0.03813149780035019, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813149780035019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1847607046365738, + 0.11235897243022919, + 0.03659028932452202, + 0.05427508056163788, + 0.08902239054441452, + 0.04504365473985672, + 0.07309412956237793, + 0.05406518653035164, + 0.2826840281486511, + 0.06810548901557922 + ], + "credit_map": [ + 0.020397445186972618, + 0.01880674995481968, + 0.02154069021344185, + 0.020474828779697418, + 0.021034561097621918, + 0.03962076082825661, + 0.02628789283335209, + 0.016597779467701912, + 0.01779913157224655, + 0.012948209419846535, + 0.015065225772559643, + 0.012037969194352627, + 0.008593508042395115, + 0.015127269551157951, + 0.011425245553255081, + 0.019362425431609154, + 0.003559218952432275, + 0.003557609859853983, + 0.003988081589341164, + 0.004320736974477768, + 0.0046409619972109795, + 0.00655583618208766, + 0.005562847945839167, + 0.004404999781399965, + 0.0050486549735069275, + 0.005414406303316355, + 0.00827596802264452, + 0.009047274477779865, + 0.004324663430452347, + 0.008113977499306202, + 0.007688654121011496, + 0.006361483130604029, + 0.007773924618959427, + 0.010216785594820976, + 0.01458851620554924, + 0.0105711305513978, + 0.0119435153901577, + 0.010571347549557686, + 0.00978137832134962, + 0.013575810939073563, + 0.003724624402821064, + 0.004652906209230423, + 0.007641117088496685, + 0.00628754124045372, + 0.0048862844705581665, + 0.004601460415869951, + 0.0063969846814870834, + 0.006852738559246063, + 0.007826014421880245, + 0.007005197927355766, + 0.008626913651823997, + 0.00936744175851345, + 0.010834337212145329, + 0.00666196970269084, + 0.012188028544187546, + 0.010584233328700066, + 0.004747752100229263, + 0.007388906553387642, + 0.005829771514981985, + 0.007259811740368605, + 0.005721050314605236, + 0.008266838267445564, + 0.007575000636279583, + 0.007276063784956932, + 0.10444152355194092, + 0.11428670585155487, + 0.025700600817799568, + 0.010908820666372776, + 0.00873936153948307, + 0.006204930134117603, + 0.004478218499571085, + 0.007923886179924011, + 0.0075203292071819305, + 0.011880610138177872, + 0.020266419276595116, + 0.007346468977630138, + 0.005517062731087208, + 0.005509678274393082, + 0.0051276786252856255, + 0.004937250632792711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.03813149780035019 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283455908298492 + ], + "opposite_logprob_tokens": [ + -2.0178346633911133 + ], + "logprob_reward": -0.14283455908298492, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14283455908298492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15134897828102112, + 0.09290354698896408, + 0.024700703099370003, + 0.04801536723971367, + 0.043634817004203796, + 0.04998621344566345, + 0.023819733411073685, + 0.0408901609480381, + 0.4901234209537506, + 0.03457707166671753 + ], + "credit_map": [ + 0.01628154329955578, + 0.016368363052606583, + 0.016469215974211693, + 0.016285132616758347, + 0.020959779620170593, + 0.02793101593852043, + 0.019737133756279945, + 0.0173167921602726, + 0.018554748967289925, + 0.009689646773040295, + 0.011881194077432156, + 0.01075925026088953, + 0.0077867042273283005, + 0.012073084712028503, + 0.00819335225969553, + 0.013965566642582417, + 0.001877669245004654, + 0.003175542689859867, + 0.002601010724902153, + 0.0025749453343451023, + 0.0020829697605222464, + 0.004304822534322739, + 0.0036950341891497374, + 0.004388708155602217, + 0.004898813087493181, + 0.004944829270243645, + 0.007943076081573963, + 0.007149129640311003, + 0.0036574946716427803, + 0.007884283550083637, + 0.005844223778694868, + 0.00569352088496089, + 0.0035537858493626118, + 0.004127552267163992, + 0.007738413754850626, + 0.006698914337903261, + 0.005152054131031036, + 0.004258500877767801, + 0.005429994780570269, + 0.006675600539892912, + 0.003974704537540674, + 0.004565396811813116, + 0.009263131767511368, + 0.00766712473705411, + 0.006100654602050781, + 0.005178533960133791, + 0.00653198454529047, + 0.006704678758978844, + 0.0021636157762259245, + 0.0030023541767150164, + 0.004597102757543325, + 0.0035606215242296457, + 0.0020446497946977615, + 0.0017372046131640673, + 0.0021796838846057653, + 0.004534501116722822, + 0.005658906418830156, + 0.006279670633375645, + 0.0041452920995652676, + 0.005028449464589357, + 0.0033187121152877808, + 0.008293033577501774, + 0.003076651133596897, + 0.005089442245662212, + 0.18632134795188904, + 0.201534241437912, + 0.03994682431221008, + 0.018078502267599106, + 0.015060617588460445, + 0.009553186595439911, + 0.006893448997288942, + 0.012735254131257534, + 0.00425328454002738, + 0.005141978617757559, + 0.009314807131886482, + 0.0038410124834626913, + 0.0029921159148216248, + 0.004827738739550114, + 0.0020713212434202433, + 0.002134813228622079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.14283455908298492 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068367287516594 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068367287516594, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068367287516594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10064058750867844, + 0.06665460020303726, + 0.07580819725990295, + 0.07193193584680557, + 0.19685021042823792, + 0.04507705196738243, + 0.04998038709163666, + 0.019512802362442017, + 0.3566865026950836, + 0.01685776188969612 + ], + "credit_map": [ + 0.011814328841865063, + 0.011548387818038464, + 0.011810040101408958, + 0.013370773755013943, + 0.011512603610754013, + 0.01773699000477791, + 0.012644227594137192, + 0.010203222744166851, + 0.008406288921833038, + 0.01425354927778244, + 0.008058966137468815, + 0.007491511292755604, + 0.004240821581333876, + 0.011869711801409721, + 0.005422424525022507, + 0.006911326199769974, + 0.00517238536849618, + 0.011455805040895939, + 0.010520590469241142, + 0.008466489613056183, + 0.004671996459364891, + 0.013442636467516422, + 0.007413309998810291, + 0.014664975926280022, + 0.007952326908707619, + 0.007793112192302942, + 0.012274076230823994, + 0.015630042180418968, + 0.004452166613191366, + 0.008933125995099545, + 0.006098482292145491, + 0.008798597380518913, + 0.02149282954633236, + 0.025268495082855225, + 0.040552712976932526, + 0.03144432231783867, + 0.008570767007768154, + 0.022183842957019806, + 0.011131824925541878, + 0.03620539978146553, + 0.004608435090631247, + 0.0038721507880836725, + 0.007818623445928097, + 0.006640343926846981, + 0.0036612744443118572, + 0.005569020286202431, + 0.006553540006279945, + 0.006353665143251419, + 0.004180930554866791, + 0.0050547365099191666, + 0.006240904796868563, + 0.006395792122930288, + 0.007837016135454178, + 0.005989185068756342, + 0.008009446784853935, + 0.006272372789680958, + 0.0022677481174468994, + 0.003078630892559886, + 0.002402563812211156, + 0.0024921915028244257, + 0.0016976846382021904, + 0.00342430523596704, + 0.001701758592389524, + 0.0024479178246110678, + 0.13868452608585358, + 0.14765548706054688, + 0.02477695606648922, + 0.012783917598426342, + 0.011032603681087494, + 0.006890004500746727, + 0.005467502400279045, + 0.009395470842719078, + 0.0018511234084144235, + 0.0027469214983284473, + 0.005252624861896038, + 0.0018432112410664558, + 0.0013532775919884443, + 0.0015935789560899138, + 0.0011227066861465573, + 0.0010943157831206918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.016068367287516594 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and does it orbit and track a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the riders wearing vibrant red weatherproof jackets and black pants with visible kneepads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the riders wear full-face helmets that are matte black with mirrored visors reflecting a blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bike have chunky knobby tires and a dark carbon fiber frame while releasing a massive plume of powder during a sharp turn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a low bright sun causing individual ice formations to shimmer against white slopes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202755868434906 + ], + "opposite_logprob_tokens": [ + -2.8120274543762207 + ], + "logprob_reward": -0.06202755868434906, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202755868434906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19939422607421875, + 0.0819053202867508, + 0.1030532568693161, + 0.05442069470882416, + 0.06330234557390213, + 0.072663813829422, + 0.03400149941444397, + 0.04219023883342743, + 0.29487335681915283, + 0.05419524013996124 + ], + "credit_map": [ + 0.024767262861132622, + 0.02302790991961956, + 0.024267489090561867, + 0.01870189607143402, + 0.014920462854206562, + 0.03605780377984047, + 0.019948024302721024, + 0.03770337998867035, + 0.014781290665268898, + 0.013759726658463478, + 0.00971999205648899, + 0.009091800078749657, + 0.0056501696817576885, + 0.010964871384203434, + 0.008150455541908741, + 0.009787010960280895, + 0.008215043693780899, + 0.014166286215186119, + 0.013615827076137066, + 0.010007504373788834, + 0.007316907402127981, + 0.017974328249692917, + 0.010492510162293911, + 0.02126484364271164, + 0.007332318462431431, + 0.006131219211965799, + 0.008866468444466591, + 0.007014043629169464, + 0.0043496559374034405, + 0.0070477738045156, + 0.00619279220700264, + 0.0074864220805466175, + 0.007220710162073374, + 0.010254500433802605, + 0.009144009090960026, + 0.00784046296030283, + 0.004831449128687382, + 0.005999717395752668, + 0.006857309490442276, + 0.01115418691188097, + 0.006772336084395647, + 0.009863810613751411, + 0.014181367121636868, + 0.008520126342773438, + 0.005928040016442537, + 0.008560516871511936, + 0.008676661178469658, + 0.010160958394408226, + 0.0028956776950508356, + 0.003823278471827507, + 0.005899013485759497, + 0.004490261897444725, + 0.0027669353876262903, + 0.004940148442983627, + 0.003708503907546401, + 0.005477679427713156, + 0.005376865155994892, + 0.007062960881739855, + 0.005268166307359934, + 0.0055874986574053764, + 0.003303329925984144, + 0.007611650042235851, + 0.0029622805304825306, + 0.005017489660531282, + 0.10528223216533661, + 0.11730989068746567, + 0.03094690479338169, + 0.011743576265871525, + 0.0086982948705554, + 0.008270611055195332, + 0.004605153109878302, + 0.008016715757548809, + 0.006068894173949957, + 0.010085765272378922, + 0.020238664001226425, + 0.006284172646701336, + 0.00232088309712708, + 0.004926529712975025, + 0.0020656546112149954, + 0.0022046782542020082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.06202755868434906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_9/index_tensor([5345, 5345])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_9/index_tensor([5345, 5345])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c3a0c445e5ec43b2aece64e5ece61bb67a2f6b40 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_9/index_tensor([5345, 5345])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot, taken at eye level, captures a moment of intense natural drama. A vibrant, furry caterpillar, its body patterned in distinct lime green and canary yellow, streaked with velvety black, rises from a dense, emerald green plant stem. Its entire form is covered in fine, translucent white hairs that catch the bright, warm sunlight, creating a glowing, halo-like effect. The caterpillar's body is visibly tense, its back arched, poised to strike with palpable energy, while its tiny, jet-black jaws are parted just enough to reveal a gleaming sharpness. The shallow depth of field renders the background into a soft, painterly dappled light and foliage bokeh, ensuring that all focus remains on the subject's dramatic pose. The entire scene is imbued with a focused, suspended tension, a microsecond of confrontation frozen in time.", + "seed": 12033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_9/index_5345_seed_12033_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05265625168103725, + "vlm_logprob": -0.05265625168103725, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.05265625168103725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03233792632818222 + ], + "opposite_logprob_tokens": [ + -4.0323381423950195 + ], + "logprob_reward": -0.03233792632818222, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03233792632818222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12814834713935852, + 0.07078664749860764, + 0.09438635408878326, + 0.1568136364221573, + 0.1420372575521469, + 0.04480184614658356, + 0.06082528457045555, + 0.23686382174491882, + 0.039769548922777176, + 0.025567326694726944 + ], + "credit_map": [ + 0.009626787155866623, + 0.013691752217710018, + 0.014013310894370079, + 0.016621561720967293, + 0.017225973308086395, + 0.020992156118154526, + 0.014368411153554916, + 0.021608388051390648, + 0.005409202538430691, + 0.008609340526163578, + 0.009303679689764977, + 0.007940842770040035, + 0.010022267699241638, + 0.014204653911292553, + 0.007047063205391169, + 0.008249587379395962, + 0.007747058756649494, + 0.011805616319179535, + 0.01536566298455, + 0.011482550762593746, + 0.014981681481003761, + 0.011365240439772606, + 0.012085583060979843, + 0.009552955627441406, + 0.01844770833849907, + 0.02110951952636242, + 0.02242923341691494, + 0.019629381597042084, + 0.02481302246451378, + 0.019466295838356018, + 0.014193805865943432, + 0.016724657267332077, + 0.01383458822965622, + 0.018562138080596924, + 0.023800544440746307, + 0.020701859146356583, + 0.0179146695882082, + 0.01935824565589428, + 0.010601038113236427, + 0.01726415939629078, + 0.004084295127540827, + 0.003782320534810424, + 0.004442954435944557, + 0.003895057598128915, + 0.00410326337441802, + 0.008776737377047539, + 0.011385886929929256, + 0.004331325646489859, + 0.006718785036355257, + 0.005166256334632635, + 0.00751999719068408, + 0.008694548159837723, + 0.006392905954271555, + 0.012012509629130363, + 0.004333136603236198, + 0.009987137280404568, + 0.020238885655999184, + 0.023439545184373856, + 0.0242113396525383, + 0.03435460105538368, + 0.029132409021258354, + 0.04778514429926872, + 0.02959979884326458, + 0.02810206077992916, + 0.00359569420106709, + 0.002954182680696249, + 0.003322177566587925, + 0.002599896863102913, + 0.0034069418907165527, + 0.009366000071167946, + 0.008907806128263474, + 0.005616841372102499, + 0.0033838716335594654, + 0.0028731473721563816, + 0.0031129319686442614, + 0.0033429176546633244, + 0.0027113426476716995, + 0.0033767824061214924, + 0.0026975597720593214, + 0.0040687695145606995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.03233792632818222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048673514276742935 + ], + "opposite_logprob_tokens": [ + -3.048673629760742 + ], + "logprob_reward": -0.048673514276742935, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048673514276742935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17507977783679962, + 0.2539476752281189, + 0.10759448260068893, + 0.08147263526916504, + 0.05512415990233421, + 0.03987300395965576, + 0.022409776225686073, + 0.18658843636512756, + 0.024892855435609818, + 0.053017161786556244 + ], + "credit_map": [ + 0.018895061686635017, + 0.029338452965021133, + 0.026928730309009552, + 0.019696056842803955, + 0.026463478803634644, + 0.020369544625282288, + 0.016200313344597816, + 0.01718813367187977, + 0.025528743863105774, + 0.032657891511917114, + 0.04903982952237129, + 0.03001227229833603, + 0.040844447910785675, + 0.0326002798974514, + 0.02044929750263691, + 0.02281494438648224, + 0.0107562355697155, + 0.013016852550208569, + 0.017684046179056168, + 0.01307295635342598, + 0.019081896170973778, + 0.01277128141373396, + 0.012387709692120552, + 0.008823501877486706, + 0.0069497437216341496, + 0.013596158474683762, + 0.010797863826155663, + 0.009360693395137787, + 0.008109556511044502, + 0.016139471903443336, + 0.0075806076638400555, + 0.008938534185290337, + 0.004699513781815767, + 0.008651704527437687, + 0.006953627336770296, + 0.006775124929845333, + 0.007296133320778608, + 0.010798514820635319, + 0.004465320613235235, + 0.005484216380864382, + 0.0038949097506701946, + 0.0063399383798241615, + 0.004908919334411621, + 0.005385500844568014, + 0.004752875305712223, + 0.0049353670328855515, + 0.0051017156802117825, + 0.0045537748374044895, + 0.002872427925467491, + 0.003483003471046686, + 0.0030321914236992598, + 0.00266321306116879, + 0.0024954380933195353, + 0.003232295159250498, + 0.0018279770156368613, + 0.002803230192512274, + 0.016111070290207863, + 0.019200626760721207, + 0.0197598896920681, + 0.027544111013412476, + 0.025111310184001923, + 0.03729356452822685, + 0.020417995750904083, + 0.021149881184101105, + 0.0018078588182106614, + 0.0021331259049475193, + 0.0029586879536509514, + 0.0020775669254362583, + 0.002911401679739356, + 0.004968262277543545, + 0.004440920893102884, + 0.003595032263547182, + 0.005772777833044529, + 0.01066224928945303, + 0.00835486501455307, + 0.005781013984233141, + 0.00568510452285409, + 0.0071471878327429295, + 0.004233411513268948, + 0.00538055831566453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.048673514276742935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585928112268448 + ], + "opposite_logprob_tokens": [ + -3.8955860137939453 + ], + "logprob_reward": -0.020585928112268448, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585928112268448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05008444935083389, + 0.05603143572807312, + 0.05296238511800766, + 0.05870407447218895, + 0.01660754159092903, + 0.02145049162209034, + 0.017297232523560524, + 0.6910477876663208, + 0.014912152662873268, + 0.020902402698993683 + ], + "credit_map": [ + 0.005568120628595352, + 0.007693165447562933, + 0.007068520411849022, + 0.005627154838293791, + 0.007943411357700825, + 0.0059784650802612305, + 0.004311627242714167, + 0.005893984343856573, + 0.006100588012486696, + 0.005353723652660847, + 0.013730155304074287, + 0.007097275927662849, + 0.007814347743988037, + 0.007008183281868696, + 0.004381521604955196, + 0.004545640200376511, + 0.005459466017782688, + 0.009323353879153728, + 0.008883357048034668, + 0.005732410121709108, + 0.005778479855507612, + 0.007392397616058588, + 0.0035732209216803312, + 0.006819704547524452, + 0.004957898054271936, + 0.007945275865495205, + 0.009028431959450245, + 0.008095531724393368, + 0.010129187256097794, + 0.007662937510758638, + 0.00503918481990695, + 0.0058456226252019405, + 0.0018292017048224807, + 0.002107570180669427, + 0.0022650714963674545, + 0.00218420522287488, + 0.0019988459534943104, + 0.002392581896856427, + 0.0015024368185549974, + 0.0023276288993656635, + 0.002481932519003749, + 0.0029433732852339745, + 0.0024329249281436205, + 0.002312060911208391, + 0.002797442954033613, + 0.0035433154553174973, + 0.002422238700091839, + 0.002517205197364092, + 0.0021029929630458355, + 0.002065701875835657, + 0.0018986333161592484, + 0.0020386516116559505, + 0.002177879912778735, + 0.002946946071460843, + 0.001251712441444397, + 0.0028147154953330755, + 0.058211177587509155, + 0.07163266092538834, + 0.06832009553909302, + 0.09995495527982712, + 0.08415231108665466, + 0.1421298384666443, + 0.08562946319580078, + 0.08101728558540344, + 0.0015276793856173754, + 0.002963185776025057, + 0.0019955469761043787, + 0.001964355120435357, + 0.0021041668951511383, + 0.00157457054592669, + 0.0011679306626319885, + 0.0016147186979651451, + 0.0021061974111944437, + 0.0027154195122420788, + 0.0036641107872128487, + 0.002719580428674817, + 0.0020177096594125032, + 0.0036224268842488527, + 0.0016516802133992314, + 0.00240527861751616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.020585928112268448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603488028049469 + ], + "opposite_logprob_tokens": [ + -1.9103487730026245 + ], + "logprob_reward": -0.1603488028049469, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603488028049469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05383545532822609, + 0.052580513060092926, + 0.04478054866194725, + 0.03338683024048805, + 0.02463425137102604, + 0.024148013442754745, + 0.03537759557366371, + 0.6688374280929565, + 0.019982358440756798, + 0.042436983436346054 + ], + "credit_map": [ + 0.005915776360780001, + 0.006679385434836149, + 0.010290415026247501, + 0.006408660672605038, + 0.007524508982896805, + 0.006700058467686176, + 0.004055527970194817, + 0.006261121481657028, + 0.004897746257483959, + 0.006539130117744207, + 0.009942224249243736, + 0.007189607247710228, + 0.007599535398185253, + 0.006448384374380112, + 0.004531168378889561, + 0.0054327198304235935, + 0.004091455601155758, + 0.006009145639836788, + 0.007414200808852911, + 0.005534449592232704, + 0.0062425397336483, + 0.005488412454724312, + 0.005137902218848467, + 0.004862445872277021, + 0.003320106538012624, + 0.0048028151504695415, + 0.005312923341989517, + 0.003942100331187248, + 0.004073264542967081, + 0.00559300696477294, + 0.0034719982650130987, + 0.0028706130106002092, + 0.002549018943682313, + 0.00308966520242393, + 0.0034950063563883305, + 0.00295758037827909, + 0.0033857179805636406, + 0.0034236214123666286, + 0.0028184917755424976, + 0.002915148390457034, + 0.0024293470196425915, + 0.0032076144125312567, + 0.0025110661517828703, + 0.0022609923034906387, + 0.0025512916035950184, + 0.0049696071073412895, + 0.0038826465606689453, + 0.002335448283702135, + 0.004298811312764883, + 0.0031116604804992676, + 0.004274500068277121, + 0.00474165566265583, + 0.0037445032503455877, + 0.006841959897428751, + 0.0026679416187107563, + 0.005696561653167009, + 0.055247124284505844, + 0.06762807071208954, + 0.06563032418489456, + 0.09489152580499649, + 0.08541591465473175, + 0.13927485048770905, + 0.08243317902088165, + 0.07831649482250214, + 0.0018973746336996555, + 0.003091825172305107, + 0.0032632090151309967, + 0.0029455632902681828, + 0.00246544205583632, + 0.0022868530359119177, + 0.0015376509400084615, + 0.002494439482688904, + 0.004267590586096048, + 0.006999123841524124, + 0.007204622030258179, + 0.004431391600519419, + 0.004213047679513693, + 0.007825125940144062, + 0.002976631745696068, + 0.0045194500125944614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.1603488028049469 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335086883045733 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001335086883045733, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335086883045733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17548911273479462, + 0.059237778186798096, + 0.059632446616888046, + 0.12257739156484604, + 0.10972774028778076, + 0.1029033437371254, + 0.06923223286867142, + 0.12189646065235138, + 0.05782609060406685, + 0.1214773952960968 + ], + "credit_map": [ + 0.014597978442907333, + 0.015050666406750679, + 0.029174281284213066, + 0.019897300750017166, + 0.023366572335362434, + 0.03803790733218193, + 0.012255411595106125, + 0.02310899645090103, + 0.00555457966402173, + 0.007037507835775614, + 0.011646675877273083, + 0.006556717213243246, + 0.007367290090769529, + 0.010548491030931473, + 0.004509475082159042, + 0.006017046980559826, + 0.006882563699036837, + 0.007401030510663986, + 0.007145369425415993, + 0.006783860735595226, + 0.012261313386261463, + 0.0076258135959506035, + 0.005509652197360992, + 0.006022847257554531, + 0.011972145177423954, + 0.016901422291994095, + 0.01940814219415188, + 0.016760729253292084, + 0.020561467856168747, + 0.01496093813329935, + 0.010471525602042675, + 0.011541023850440979, + 0.010266249068081379, + 0.010137803852558136, + 0.024291709065437317, + 0.0123796621337533, + 0.007333696354180574, + 0.016814257949590683, + 0.013100012205541134, + 0.015404359437525272, + 0.010490926913917065, + 0.013793302699923515, + 0.01441591139882803, + 0.014108161441981792, + 0.014388100244104862, + 0.01207384467124939, + 0.010651635006070137, + 0.012981466017663479, + 0.007138066925108433, + 0.007036000955849886, + 0.006916773039847612, + 0.008374355733394623, + 0.007242568768560886, + 0.012690430507063866, + 0.010446831583976746, + 0.009387212805449963, + 0.013730865903198719, + 0.013234801590442657, + 0.013727850280702114, + 0.01575009897351265, + 0.013502695597708225, + 0.024556901305913925, + 0.01255854219198227, + 0.014834714122116566, + 0.005398438777774572, + 0.007150521967560053, + 0.009398292750120163, + 0.006716456264257431, + 0.006954292766749859, + 0.008528635837137699, + 0.009109689854085445, + 0.00456976518034935, + 0.012600834481418133, + 0.01830895245075226, + 0.026037249714136124, + 0.016658354550600052, + 0.012948589399456978, + 0.014262800104916096, + 0.009355218149721622, + 0.011305407620966434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.001335086883045733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, taken at eye level, captures a moment of intense natural drama. A vibrant, furry caterpillar, its body patterned in distinct lime green and canary yellow, streaked with velvety black, rises from a dense, emerald green plant stem. Its entire form is covered in fine, translucent white hairs that catch the bright, warm sunlight, creating a glowing, halo-like effect. The caterpillar's body is visibly tense, its back arched, poised to strike with palpable energy, while its tiny, jet-black jaws are parted just enough to reveal a gleaming sharpness. The shallow depth of field renders the background into a soft, painterly dappled light and foliage bokeh, ensuring that all focus remains on the subject's dramatic pose. The entire scene is imbued with a focused, suspended tension, a microsecond of confrontation frozen in time.", + "seed": 12034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_9/index_5345_seed_12034_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03384441938251257, + "vlm_logprob": -0.03384441938251257, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03384441938251257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022160988301038742 + ], + "opposite_logprob_tokens": [ + -4.52216100692749 + ], + "logprob_reward": -0.022160988301038742, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.022160988301038742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20031660795211792, + 0.09480689465999603, + 0.05811646953225136, + 0.08153422921895981, + 0.030126187950372696, + 0.04674306884407997, + 0.12103043496608734, + 0.06371977180242538, + 0.09696339070796967, + 0.20664295554161072 + ], + "credit_map": [ + 0.018312569707632065, + 0.02845347672700882, + 0.029003247618675232, + 0.02169392630457878, + 0.01927175559103489, + 0.030245928093791008, + 0.028676945716142654, + 0.024658741429448128, + 0.008809427730739117, + 0.02441190369427204, + 0.011272941716015339, + 0.009595589712262154, + 0.008877504616975784, + 0.01185219269245863, + 0.012281110510230064, + 0.0077062188647687435, + 0.006570132914930582, + 0.014625969342887402, + 0.006959110964089632, + 0.006779721472412348, + 0.004718651529401541, + 0.006747096311300993, + 0.005672107450664043, + 0.006043674889951944, + 0.008117154240608215, + 0.012040195055305958, + 0.013044849969446659, + 0.012620111927390099, + 0.00688154436647892, + 0.010470782406628132, + 0.011171565391123295, + 0.007188021205365658, + 0.0031603218521922827, + 0.00678655132651329, + 0.0034630452282726765, + 0.003906559199094772, + 0.0036457497626543045, + 0.003340020077303052, + 0.0030107623897492886, + 0.00281317625194788, + 0.0037861757446080446, + 0.009707105346024036, + 0.005788231268525124, + 0.0035928033757954836, + 0.005466554779559374, + 0.005765212699770927, + 0.007425455842167139, + 0.0052115232683718204, + 0.011854332871735096, + 0.015629205852746964, + 0.01583501137793064, + 0.009061063639819622, + 0.017105836421251297, + 0.019085966050624847, + 0.01137759443372488, + 0.02108141779899597, + 0.004589838907122612, + 0.008403442800045013, + 0.00713746203109622, + 0.006735747680068016, + 0.007080867420881987, + 0.011641018092632294, + 0.010308310389518738, + 0.007823076099157333, + 0.007182022091001272, + 0.012692496180534363, + 0.010364429093897343, + 0.011884640902280807, + 0.01984124258160591, + 0.0110463323071599, + 0.013113999739289284, + 0.010838218033313751, + 0.01658053696155548, + 0.02557412162423134, + 0.02958139032125473, + 0.018332213163375854, + 0.03347333148121834, + 0.028137458488345146, + 0.0353868305683136, + 0.019577063620090485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.022160988301038742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002859562635422 + ], + "opposite_logprob_tokens": [ + -2.69502854347229 + ], + "logprob_reward": -0.07002859562635422, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002859562635422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18991491198539734, + 0.07018449902534485, + 0.07624907046556473, + 0.08022081851959229, + 0.05096614733338356, + 0.041592709720134735, + 0.18888519704341888, + 0.1097768098115921, + 0.06340006738901138, + 0.12880980968475342 + ], + "credit_map": [ + 0.020113680511713028, + 0.03179900720715523, + 0.02563956193625927, + 0.01964348740875721, + 0.021492348983883858, + 0.02400209940969944, + 0.027614586055278778, + 0.01961013861000538, + 0.006510857958346605, + 0.015243316069245338, + 0.009747511707246304, + 0.009820317849516869, + 0.005336659494787455, + 0.007856282405555248, + 0.009055033326148987, + 0.00661452254280448, + 0.008906313218176365, + 0.01544440258294344, + 0.010019844397902489, + 0.008317750878632069, + 0.0056139519438147545, + 0.010581356473267078, + 0.0093015655875206, + 0.008063878864049911, + 0.007428966462612152, + 0.018720518797636032, + 0.009881149046123028, + 0.008316392078995705, + 0.007931072264909744, + 0.009027265943586826, + 0.010919240303337574, + 0.007996216416358948, + 0.004443359561264515, + 0.007144746370613575, + 0.007056922186166048, + 0.007581387646496296, + 0.007568520959466696, + 0.005723161622881889, + 0.0057569872587919235, + 0.005691060330718756, + 0.003341310890391469, + 0.0077607897110283375, + 0.004714467097073793, + 0.004110276233404875, + 0.0061694784089922905, + 0.005458478815853596, + 0.005370171274989843, + 0.004667735658586025, + 0.01744704879820347, + 0.031104551628232002, + 0.026476850733160973, + 0.014591476880013943, + 0.031078821048140526, + 0.025071535259485245, + 0.01680673286318779, + 0.02630816586315632, + 0.008737756870687008, + 0.012333426624536514, + 0.013703706674277782, + 0.011546836234629154, + 0.011183501221239567, + 0.024065082892775536, + 0.012954803183674812, + 0.015251696109771729, + 0.0045051658526062965, + 0.012510408647358418, + 0.007131658960133791, + 0.006135181989520788, + 0.010545223951339722, + 0.008232563734054565, + 0.009146856144070625, + 0.005193003918975592, + 0.009751015342772007, + 0.02208896167576313, + 0.015593674033880234, + 0.012594184838235378, + 0.018629813566803932, + 0.018824320286512375, + 0.02072674222290516, + 0.01060109119862318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.07002859562635422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581839606165886 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581839606165886, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581839606165886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1390751302242279, + 0.10243277996778488, + 0.2939734160900116, + 0.09128238260746002, + 0.02060152404010296, + 0.03140219300985336, + 0.13497638702392578, + 0.08206629753112793, + 0.0406259261071682, + 0.06356395035982132 + ], + "credit_map": [ + 0.014560943469405174, + 0.02532295137643814, + 0.019824525341391563, + 0.014047590084373951, + 0.013595664873719215, + 0.01903567649424076, + 0.019595788791775703, + 0.01309198047965765, + 0.009521140716969967, + 0.030270693823695183, + 0.012416942976415157, + 0.011919005773961544, + 0.008808097802102566, + 0.011835204437375069, + 0.00994583684951067, + 0.00771585525944829, + 0.02993379347026348, + 0.08614590018987656, + 0.035950493067502975, + 0.030248578637838364, + 0.02691880241036415, + 0.030391676351428032, + 0.031873077154159546, + 0.022511115297675133, + 0.010237160138785839, + 0.016022004187107086, + 0.01362888514995575, + 0.011942981742322445, + 0.01027188915759325, + 0.008339932188391685, + 0.012390514835715294, + 0.008449014276266098, + 0.0019316797843202949, + 0.004126504994928837, + 0.002426235005259514, + 0.00231152237392962, + 0.002814185107126832, + 0.0022470129188150167, + 0.002605301095172763, + 0.002139081247150898, + 0.0033066649921238422, + 0.01039445586502552, + 0.0037108082324266434, + 0.002754545770585537, + 0.0022537086624652147, + 0.0035847523249685764, + 0.0026834162417799234, + 0.00271384185180068, + 0.010539601556956768, + 0.02216951735317707, + 0.01569518633186817, + 0.009415979497134686, + 0.02350330725312233, + 0.02037363499403, + 0.014461581595242023, + 0.018817583099007607, + 0.005805510561913252, + 0.009238972328603268, + 0.009223256260156631, + 0.007435881532728672, + 0.013369965367019176, + 0.014758672565221786, + 0.013071035966277122, + 0.009163000620901585, + 0.0037730149924755096, + 0.006345516536384821, + 0.005460989195853472, + 0.005126936361193657, + 0.004884513095021248, + 0.004428737331181765, + 0.006731352303177118, + 0.0038748665247112513, + 0.00533943809568882, + 0.00917399674654007, + 0.007411866448819637, + 0.006677694618701935, + 0.010405185632407665, + 0.006442308891564608, + 0.012078206986188889, + 0.006035250145941973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.020581839606165886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493928864598274 + ], + "opposite_logprob_tokens": [ + -2.9299392700195312 + ], + "logprob_reward": -0.05493928864598274, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493928864598274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13735485076904297, + 0.14849673211574554, + 0.05358853563666344, + 0.1508384644985199, + 0.08092428743839264, + 0.04677430912852287, + 0.09579939395189285, + 0.0713217705488205, + 0.049678217619657516, + 0.16522344946861267 + ], + "credit_map": [ + 0.013173099607229233, + 0.0196040328592062, + 0.019065095111727715, + 0.017127424478530884, + 0.01496096607297659, + 0.017657853662967682, + 0.021364836022257805, + 0.014401549473404884, + 0.01372454036027193, + 0.04465974494814873, + 0.017844203859567642, + 0.019081532955169678, + 0.012373044155538082, + 0.01643447019159794, + 0.015321026556193829, + 0.009058183990418911, + 0.00609830766916275, + 0.012031849473714828, + 0.0061876727268099785, + 0.006512201391160488, + 0.0041352082043886185, + 0.007667556405067444, + 0.0060550072230398655, + 0.004900732543319464, + 0.012644029222428799, + 0.03593594580888748, + 0.019387779757380486, + 0.014115729369223118, + 0.014074461534619331, + 0.019000671803951263, + 0.02061927318572998, + 0.015060566365718842, + 0.008268335834145546, + 0.015628259629011154, + 0.00831076130270958, + 0.013144666329026222, + 0.009105070494115353, + 0.007924847304821014, + 0.010578623972833157, + 0.007963722571730614, + 0.004158155992627144, + 0.010462687350809574, + 0.004130004905164242, + 0.006775685120373964, + 0.007737720850855112, + 0.005105977412313223, + 0.005310405511409044, + 0.003093673614785075, + 0.009587377309799194, + 0.013642759993672371, + 0.012394102290272713, + 0.0072053829208016396, + 0.013011195696890354, + 0.015887146815657616, + 0.009359828196465969, + 0.014711600728332996, + 0.00535470200702548, + 0.011843687854707241, + 0.007974173873662949, + 0.005763035733252764, + 0.01570095866918564, + 0.011819944716989994, + 0.0069777169264853, + 0.005887554492801428, + 0.004255317151546478, + 0.008880998007953167, + 0.0055628553964197636, + 0.004371101967990398, + 0.008672550320625305, + 0.006744505371898413, + 0.0067848279140889645, + 0.004406064748764038, + 0.015332239679992199, + 0.03232559561729431, + 0.01976638287305832, + 0.01841028966009617, + 0.023746611550450325, + 0.016521567478775978, + 0.023268498480319977, + 0.015852252021431923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.05493928864598274 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015113847330212593 + ], + "opposite_logprob_tokens": [ + -6.501511573791504 + ], + "logprob_reward": -0.0015113847330212593, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015113847330212593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18995602428913116, + 0.052325762808322906, + 0.03194377198815346, + 0.08332686871290207, + 0.050896648317575455, + 0.06771387904882431, + 0.3362284302711487, + 0.059868793934583664, + 0.06181589141488075, + 0.06592393666505814 + ], + "credit_map": [ + 0.019684722647070885, + 0.027570564299821854, + 0.0274407546967268, + 0.016313603147864342, + 0.01965818740427494, + 0.029386667534708977, + 0.026194538921117783, + 0.02370697632431984, + 0.004547686781734228, + 0.01560276560485363, + 0.0064628999680280685, + 0.0055835130624473095, + 0.004425922408699989, + 0.005678028333932161, + 0.004886818118393421, + 0.005138122942298651, + 0.0036574588157236576, + 0.0062063345685601234, + 0.0045821815729141235, + 0.0038903560489416122, + 0.003141283057630062, + 0.003678260138258338, + 0.004361873958259821, + 0.0024260198697447777, + 0.0087082888931036, + 0.009280909784138203, + 0.01332964189350605, + 0.0067196195013821125, + 0.007328836712986231, + 0.016154538840055466, + 0.009550039656460285, + 0.012254992499947548, + 0.0051041957922279835, + 0.010068668983876705, + 0.0061044166795909405, + 0.004621286876499653, + 0.004021757282316685, + 0.009178876876831055, + 0.005953459534794092, + 0.005843982100486755, + 0.006121338345110416, + 0.01025459449738264, + 0.007552208844572306, + 0.009642607532441616, + 0.007714768406003714, + 0.00988379493355751, + 0.0101610803976655, + 0.006383481901139021, + 0.029581213369965553, + 0.051629748195409775, + 0.043275412172079086, + 0.024469127878546715, + 0.0504586324095726, + 0.04949990659952164, + 0.03405953571200371, + 0.0532548651099205, + 0.005325892940163612, + 0.005252741742879152, + 0.007126258220523596, + 0.006209237966686487, + 0.006558398250490427, + 0.00916577223688364, + 0.01021551713347435, + 0.010014969855546951, + 0.00632511917501688, + 0.0063737621530890465, + 0.00687410868704319, + 0.005400826223194599, + 0.00569651136174798, + 0.009436871856451035, + 0.01123362872749567, + 0.010475058108568192, + 0.005228446330875158, + 0.0066308509558439255, + 0.00830830354243517, + 0.006856460124254227, + 0.008989979512989521, + 0.010224011726677418, + 0.009981614537537098, + 0.009704260155558586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0015113847330212593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, taken at eye level, captures a moment of intense natural drama. A vibrant, furry caterpillar, its body patterned in distinct lime green and canary yellow, streaked with velvety black, rises from a dense, emerald green plant stem. Its entire form is covered in fine, translucent white hairs that catch the bright, warm sunlight, creating a glowing, halo-like effect. The caterpillar's body is visibly tense, its back arched, poised to strike with palpable energy, while its tiny, jet-black jaws are parted just enough to reveal a gleaming sharpness. The shallow depth of field renders the background into a soft, painterly dappled light and foliage bokeh, ensuring that all focus remains on the subject's dramatic pose. The entire scene is imbued with a focused, suspended tension, a microsecond of confrontation frozen in time.", + "seed": 12035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_9/index_5345_seed_12035_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023831729695666583, + "vlm_logprob": -0.023831729695666583, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.023831729695666583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01270721573382616 + ], + "opposite_logprob_tokens": [ + -5.137707233428955 + ], + "logprob_reward": -0.01270721573382616, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01270721573382616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08406048268079758, + 0.046467624604701996, + 0.06981486082077026, + 0.0542571060359478, + 0.04927418753504753, + 0.06753695756196976, + 0.03600301593542099, + 0.5388192534446716, + 0.019630223512649536, + 0.03413635119795799 + ], + "credit_map": [ + 0.010518443770706654, + 0.012566197663545609, + 0.008782478049397469, + 0.011892632581293583, + 0.008823066018521786, + 0.011842956766486168, + 0.009411249309778214, + 0.010223458521068096, + 0.005516468547284603, + 0.005136462394148111, + 0.0049278950318694115, + 0.006249828264117241, + 0.0048289112746715546, + 0.00773834902793169, + 0.0039620609022676945, + 0.00810764916241169, + 0.009590365923941135, + 0.009768133983016014, + 0.007782252971082926, + 0.008577344007790089, + 0.009495636448264122, + 0.008718378841876984, + 0.0089082857593894, + 0.006974461954087019, + 0.00817867647856474, + 0.005694843363016844, + 0.005338694900274277, + 0.008726214990019798, + 0.005291685461997986, + 0.007069278974086046, + 0.007230259478092194, + 0.006727449595928192, + 0.008909708820283413, + 0.0037806653417646885, + 0.005968259647488594, + 0.005318053532391787, + 0.0041409386321902275, + 0.007176066283136606, + 0.0038900196086615324, + 0.010090471245348454, + 0.010959316045045853, + 0.007531487848609686, + 0.006648465059697628, + 0.01058897003531456, + 0.006197096314281225, + 0.008978541009128094, + 0.007797406055033207, + 0.008835673332214355, + 0.004545732866972685, + 0.005022010300308466, + 0.0031344194430857897, + 0.0035635950043797493, + 0.004073210060596466, + 0.0053322757594287395, + 0.0032747879158705473, + 0.00705698411911726, + 0.025708256289362907, + 0.009477050974965096, + 0.01148812286555767, + 0.015004240907728672, + 0.02431059069931507, + 0.19940821826457977, + 0.015038670971989632, + 0.23838414251804352, + 0.0021202138159424067, + 0.0019324291497468948, + 0.0020494882483035326, + 0.0023389034904539585, + 0.002289609983563423, + 0.0025126445107162, + 0.0031343454029411077, + 0.0032525893766433, + 0.0031292899511754513, + 0.003634816035628319, + 0.004912541247904301, + 0.0038048706483095884, + 0.005396938417106867, + 0.00399029953405261, + 0.004794278182089329, + 0.00447331415489316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.01270721573382616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371100872755051 + ], + "opposite_logprob_tokens": [ + -3.4087109565734863 + ], + "logprob_reward": -0.03371100872755051, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371100872755051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08756659179925919, + 0.08614171296358109, + 0.0658423900604248, + 0.07258892059326172, + 0.04234487935900688, + 0.06456796079874039, + 0.04450036212801933, + 0.40850740671157837, + 0.028407273814082146, + 0.09953244030475616 + ], + "credit_map": [ + 0.013985135592520237, + 0.010524294339120388, + 0.008309055119752884, + 0.008473500609397888, + 0.010556353256106377, + 0.012917599640786648, + 0.011016588658094406, + 0.011784062720835209, + 0.007348922546952963, + 0.013243870809674263, + 0.007592663634568453, + 0.01488379668444395, + 0.010136505588889122, + 0.012070165947079659, + 0.009304612874984741, + 0.011561176739633083, + 0.006346425507217646, + 0.010723557323217392, + 0.00813207309693098, + 0.006428451277315617, + 0.010251354426145554, + 0.007947917096316814, + 0.0061730435118079185, + 0.009839573875069618, + 0.007277288939803839, + 0.010631891898810863, + 0.008643556386232376, + 0.009128966368734837, + 0.009684417396783829, + 0.00848176795989275, + 0.008733591996133327, + 0.01000744104385376, + 0.005274146795272827, + 0.004369092173874378, + 0.0061970725655555725, + 0.006429851986467838, + 0.005139836110174656, + 0.0052322521805763245, + 0.004848780110478401, + 0.004853847436606884, + 0.00757593521848321, + 0.006461372133344412, + 0.005177378188818693, + 0.011075151152908802, + 0.005961907096207142, + 0.010252658277750015, + 0.006981567479670048, + 0.011081992648541927, + 0.005431556608527899, + 0.0035642897710204124, + 0.005373905412852764, + 0.004927984904497862, + 0.005370269063860178, + 0.006347240414470434, + 0.005622351076453924, + 0.00786276999861002, + 0.021206596866250038, + 0.007712348364293575, + 0.009449043311178684, + 0.012820067815482616, + 0.017954375594854355, + 0.148297518491745, + 0.013311858288943768, + 0.17775560915470123, + 0.002889995463192463, + 0.004350570496171713, + 0.0031007498037070036, + 0.003214638214558363, + 0.0042625917121768, + 0.0028290385380387306, + 0.0037856344133615494, + 0.003974053077399731, + 0.008459635078907013, + 0.011929108761250973, + 0.01769145019352436, + 0.0119740329682827, + 0.013133617118000984, + 0.01016925647854805, + 0.01448257640004158, + 0.011692769825458527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.03371100872755051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232884231954813 + ], + "opposite_logprob_tokens": [ + -3.7732884883880615 + ], + "logprob_reward": -0.0232884231954813, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232884231954813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20914651453495026, + 0.092845618724823, + 0.07997866719961166, + 0.13750919699668884, + 0.04827600345015526, + 0.08243341743946075, + 0.03916824236512184, + 0.1708359718322754, + 0.038034457713365555, + 0.10177188366651535 + ], + "credit_map": [ + 0.03408081457018852, + 0.022980138659477234, + 0.02921847626566887, + 0.023665960878133774, + 0.02271290495991707, + 0.024622980505228043, + 0.026764534413814545, + 0.02510070614516735, + 0.010904361493885517, + 0.009663216769695282, + 0.008649971336126328, + 0.014556193724274635, + 0.01211627945303917, + 0.013828949071466923, + 0.010063174180686474, + 0.01306347455829382, + 0.008770707063376904, + 0.008641754277050495, + 0.012234887108206749, + 0.008360127918422222, + 0.010517015121877193, + 0.01171782799065113, + 0.006802285555750132, + 0.012934059835970402, + 0.016711663454771042, + 0.014917926862835884, + 0.01328569371253252, + 0.02400016412138939, + 0.017195722088217735, + 0.017635414376854897, + 0.019775589928030968, + 0.013987020589411259, + 0.007984766736626625, + 0.003959837835282087, + 0.006723896600306034, + 0.00328301009722054, + 0.005668650381267071, + 0.006173033267259598, + 0.004512659274041653, + 0.009970149956643581, + 0.009030004031956196, + 0.008636977523565292, + 0.009077134542167187, + 0.012649424374103546, + 0.011185198090970516, + 0.010102328844368458, + 0.012800149619579315, + 0.00895219761878252, + 0.004230119287967682, + 0.0039022534620016813, + 0.004095878917723894, + 0.004715032409876585, + 0.0052491300739347935, + 0.005341901909559965, + 0.004700870253145695, + 0.006933056749403477, + 0.007616550661623478, + 0.0029099672101438046, + 0.0031794996466487646, + 0.0028063973877578974, + 0.007867834530770779, + 0.06483109295368195, + 0.004272509831935167, + 0.0773521140217781, + 0.004257404711097479, + 0.005499938502907753, + 0.003625625278800726, + 0.005318880546838045, + 0.005372326821088791, + 0.004308719653636217, + 0.004516749642789364, + 0.005134813487529755, + 0.012390457093715668, + 0.00944251287728548, + 0.018128851428627968, + 0.015994111075997353, + 0.009715307503938675, + 0.011387652717530727, + 0.011794013902544975, + 0.012918977066874504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0232884231954813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863956198096275 + ], + "opposite_logprob_tokens": [ + -3.0486395359039307 + ], + "logprob_reward": -0.04863956198096275, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863956198096275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17745153605937958, + 0.046857334673404694, + 0.031125258654356003, + 0.06869477778673172, + 0.02587134577333927, + 0.05602848529815674, + 0.0366310253739357, + 0.5026845932006836, + 0.025063766166567802, + 0.029591813683509827 + ], + "credit_map": [ + 0.02974182553589344, + 0.018845919519662857, + 0.02542935684323311, + 0.020268891006708145, + 0.019198577851057053, + 0.0205526202917099, + 0.021202562376856804, + 0.02221178635954857, + 0.004164319951087236, + 0.006375573575496674, + 0.004724906757473946, + 0.0070913382805883884, + 0.005875986069440842, + 0.0076530734077095985, + 0.003959902562201023, + 0.007012233603745699, + 0.004139305092394352, + 0.0036461474373936653, + 0.004821764770895243, + 0.003923643846064806, + 0.004858736414462328, + 0.003805612912401557, + 0.003046112135052681, + 0.0028839390724897385, + 0.008249752223491669, + 0.008650281466543674, + 0.007108433172106743, + 0.01392064057290554, + 0.006707656662911177, + 0.008940126746892929, + 0.007783134002238512, + 0.007334753405302763, + 0.002781439106911421, + 0.002685343846678734, + 0.003040080424398184, + 0.003368276171386242, + 0.00378569308668375, + 0.0034726662561297417, + 0.003943742718547583, + 0.002794104628264904, + 0.005230576265603304, + 0.006924132350832224, + 0.00527458731085062, + 0.013148077763617039, + 0.005766300950199366, + 0.008422758430242538, + 0.006197783630341291, + 0.005064265802502632, + 0.004320654086768627, + 0.004608671646565199, + 0.003584979334846139, + 0.0038614284712821245, + 0.005834604613482952, + 0.004696603398770094, + 0.0034635993652045727, + 0.006260480731725693, + 0.024503808468580246, + 0.00903007760643959, + 0.009695569053292274, + 0.013116003014147282, + 0.022830713540315628, + 0.18624354898929596, + 0.014673368073999882, + 0.22259153425693512, + 0.0021987631916999817, + 0.003060793038457632, + 0.003352011553943157, + 0.0037953450810164213, + 0.002883362350985408, + 0.002686278661713004, + 0.0040066661313176155, + 0.0030805475544184446, + 0.0036698831245303154, + 0.0033240204211324453, + 0.0033548104111105204, + 0.004779701121151447, + 0.0034064273349940777, + 0.0033655462320894003, + 0.004020701628178358, + 0.003670724108815193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.04863956198096275 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008124388405121863 + ], + "opposite_logprob_tokens": [ + -7.125812530517578 + ], + "logprob_reward": -0.0008124388405121863, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008124388405121863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14986857771873474, + 0.06781419366598129, + 0.04683886095881462, + 0.06781462579965591, + 0.020036490634083748, + 0.03469986468553543, + 0.029099708423018456, + 0.5271089673042297, + 0.03103921376168728, + 0.025679538026452065 + ], + "credit_map": [ + 0.02362879179418087, + 0.015502862632274628, + 0.020455077290534973, + 0.019829167053103447, + 0.014424684457480907, + 0.018759896978735924, + 0.018949255347251892, + 0.018318846821784973, + 0.010606990195810795, + 0.005825906991958618, + 0.00555696664378047, + 0.00847044587135315, + 0.011299967765808105, + 0.009467554278671741, + 0.006933590862900019, + 0.009652769193053246, + 0.004269468132406473, + 0.00866649765521288, + 0.005139491055160761, + 0.004392432980239391, + 0.008050515316426754, + 0.005419707857072353, + 0.006244251038879156, + 0.004656497854739428, + 0.011784092523157597, + 0.007996507920324802, + 0.006278645247220993, + 0.010240619070827961, + 0.006432456895709038, + 0.008968977257609367, + 0.00746890576556325, + 0.00864441692829132, + 0.002781544579192996, + 0.0021632732823491096, + 0.002169157611206174, + 0.00266033923253417, + 0.0021471662912517786, + 0.0023959828540682793, + 0.0029012563172727823, + 0.0028177706990391016, + 0.002469463972374797, + 0.0048643797636032104, + 0.003233575727790594, + 0.005792900454252958, + 0.003386526834219694, + 0.006208458449691534, + 0.0026025609113276005, + 0.006141997873783112, + 0.002534199971705675, + 0.0035006508696824312, + 0.003622116521000862, + 0.0032966912258416414, + 0.004523822572082281, + 0.003489740425720811, + 0.004806932527571917, + 0.003325552912428975, + 0.025331690907478333, + 0.008915819227695465, + 0.009817445650696754, + 0.011886042542755604, + 0.023633670061826706, + 0.19721201062202454, + 0.014449914917349815, + 0.2358623892068863, + 0.0035467033740133047, + 0.0028135275933891535, + 0.0055052610114216805, + 0.0033031643833965063, + 0.0025203905533999205, + 0.0032912553288042545, + 0.005801781080663204, + 0.004257129970937967, + 0.002329422626644373, + 0.002791203558444977, + 0.004199597984552383, + 0.003935947548598051, + 0.002980326535180211, + 0.003216712037101388, + 0.002866290044039488, + 0.0033600383903831244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0008124388405121863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, taken at eye level, captures a moment of intense natural drama. A vibrant, furry caterpillar, its body patterned in distinct lime green and canary yellow, streaked with velvety black, rises from a dense, emerald green plant stem. Its entire form is covered in fine, translucent white hairs that catch the bright, warm sunlight, creating a glowing, halo-like effect. The caterpillar's body is visibly tense, its back arched, poised to strike with palpable energy, while its tiny, jet-black jaws are parted just enough to reveal a gleaming sharpness. The shallow depth of field renders the background into a soft, painterly dappled light and foliage bokeh, ensuring that all focus remains on the subject's dramatic pose. The entire scene is imbued with a focused, suspended tension, a microsecond of confrontation frozen in time.", + "seed": 12036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000085/group_9/index_5345_seed_12036_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07471690913662314, + "vlm_logprob": -0.07471690913662314, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.07471690913662314, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028373850509524345 + ], + "opposite_logprob_tokens": [ + -4.278373718261719 + ], + "logprob_reward": -0.028373850509524345, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.028373850509524345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1217915490269661, + 0.09109300374984741, + 0.0459575280547142, + 0.18946465849876404, + 0.04866933450102806, + 0.1609295755624771, + 0.057176489382982254, + 0.08481026440858841, + 0.12420885264873505, + 0.075898677110672 + ], + "credit_map": [ + 0.012354953214526176, + 0.01428927481174469, + 0.01626751199364662, + 0.021967779844999313, + 0.0161096453666687, + 0.011366946622729301, + 0.021532142534852028, + 0.007903306744992733, + 0.008335794322192669, + 0.010457717813551426, + 0.012506444938480854, + 0.015047451481223106, + 0.013398220762610435, + 0.00799268763512373, + 0.016880273818969727, + 0.006474426947534084, + 0.004202592186629772, + 0.005109378136694431, + 0.007594314403831959, + 0.00615333765745163, + 0.004502086900174618, + 0.005320239346474409, + 0.00856759212911129, + 0.004507992882281542, + 0.022820258513092995, + 0.024079816415905952, + 0.03264591097831726, + 0.03298380598425865, + 0.011828679591417313, + 0.020853089168667793, + 0.02950225956737995, + 0.014750849455595016, + 0.007617455441504717, + 0.005379567388445139, + 0.008372213691473007, + 0.005000023636966944, + 0.005382485222071409, + 0.007718637585639954, + 0.004735398106276989, + 0.004463557153940201, + 0.019417941570281982, + 0.016734277829527855, + 0.022009816020727158, + 0.03025810606777668, + 0.01972854696214199, + 0.015773901715874672, + 0.02342170849442482, + 0.013585302978754044, + 0.007467321585863829, + 0.005761479493230581, + 0.009590471163392067, + 0.008810342289507389, + 0.006847986485809088, + 0.006015513092279434, + 0.0065015628933906555, + 0.006181821227073669, + 0.011422043666243553, + 0.010068611241877079, + 0.012379981577396393, + 0.019136300310492516, + 0.008491544984281063, + 0.008176803588867188, + 0.008745882660150528, + 0.006389105226844549, + 0.009304282255470753, + 0.01301279105246067, + 0.013773990795016289, + 0.018750427290797234, + 0.024688925594091415, + 0.01903383620083332, + 0.01519752200692892, + 0.010447089560329914, + 0.0065762437880039215, + 0.007673060055822134, + 0.010409845039248466, + 0.011574764735996723, + 0.00935179553925991, + 0.007424228824675083, + 0.014738655649125576, + 0.008150093257427216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.028373850509524345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281275868415833 + ], + "opposite_logprob_tokens": [ + -2.017812728881836 + ], + "logprob_reward": -0.14281275868415833, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281275868415833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06294343620538712, + 0.19543662667274475, + 0.06340514123439789, + 0.040581006556749344, + 0.07873129099607468, + 0.09925586730241776, + 0.035521022975444794, + 0.1171865463256836, + 0.1705508679151535, + 0.13638819754123688 + ], + "credit_map": [ + 0.007314904127269983, + 0.00684213126078248, + 0.01232557836920023, + 0.008562376722693443, + 0.007574161514639854, + 0.007936219684779644, + 0.006068606395274401, + 0.006319460459053516, + 0.01925734430551529, + 0.022688167169690132, + 0.026546472683548927, + 0.028525101020932198, + 0.03112420067191124, + 0.01733476109802723, + 0.03657973185181618, + 0.01338085439056158, + 0.008291848935186863, + 0.005502805113792419, + 0.010601473040878773, + 0.011253925040364265, + 0.006766685750335455, + 0.005684114992618561, + 0.010975267738103867, + 0.004329018294811249, + 0.00439181923866272, + 0.004086880944669247, + 0.005372666288167238, + 0.0061325011774897575, + 0.004196062684059143, + 0.003731299424543977, + 0.007131914142519236, + 0.005537864752113819, + 0.010274541564285755, + 0.008924202062189579, + 0.013366632163524628, + 0.015427368693053722, + 0.00825020857155323, + 0.007833680137991905, + 0.009680353105068207, + 0.004974301438778639, + 0.010172062553465366, + 0.010968309827148914, + 0.013601030223071575, + 0.022440621629357338, + 0.012342768721282482, + 0.00966107752174139, + 0.012473070062696934, + 0.007596928160637617, + 0.0035397156607359648, + 0.003241903381422162, + 0.0053481971845030785, + 0.005186678841710091, + 0.0040709772147238255, + 0.006001295056194067, + 0.0040792422369122505, + 0.004053013864904642, + 0.009975776076316833, + 0.012091456912457943, + 0.01399982813745737, + 0.021267836913466454, + 0.027852199971675873, + 0.009230853989720345, + 0.015937723219394684, + 0.0068308692425489426, + 0.017506256699562073, + 0.01810060814023018, + 0.019105983898043633, + 0.02789812907576561, + 0.030592279508709908, + 0.026598375290632248, + 0.018677396699786186, + 0.012071848846971989, + 0.011915061622858047, + 0.014772670343518257, + 0.018690314143896103, + 0.02105691097676754, + 0.015629034489393234, + 0.015736648812890053, + 0.015860911458730698, + 0.02272663451731205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.14281275868415833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22582270205020905, + 0.20102012157440186, + 0.07360761612653732, + 0.040309008210897446, + 0.023716602474451065, + 0.09601027518510818, + 0.05456548184156418, + 0.06666979938745499, + 0.09637697041034698, + 0.12190140038728714 + ], + "credit_map": [ + 0.02437858283519745, + 0.025251492857933044, + 0.036505408585071564, + 0.03999257832765579, + 0.02674471214413643, + 0.023958632722496986, + 0.03303645923733711, + 0.015954844653606415, + 0.02021319605410099, + 0.022311249747872353, + 0.025068148970603943, + 0.03536181524395943, + 0.030607908964157104, + 0.018483953550457954, + 0.03514254465699196, + 0.01383130345493555, + 0.010133769363164902, + 0.006904537323862314, + 0.014856123365461826, + 0.013240964151918888, + 0.0063283261843025684, + 0.006879817694425583, + 0.010795454494655132, + 0.00446862168610096, + 0.0044625261798501015, + 0.00485926354303956, + 0.0064117517322301865, + 0.007786998990923166, + 0.004390438087284565, + 0.0037560611963272095, + 0.005571598652750254, + 0.0030703716911375523, + 0.002550173085182905, + 0.0036550231743603945, + 0.002841003006324172, + 0.0046246410347521305, + 0.0027170313987880945, + 0.002389937173575163, + 0.0030335630290210247, + 0.001905231038108468, + 0.011801161803305149, + 0.010275958105921745, + 0.014476367272436619, + 0.020612038671970367, + 0.010804434306919575, + 0.00780714163556695, + 0.01158872526139021, + 0.008644452318549156, + 0.006142825353890657, + 0.005808080546557903, + 0.008440528064966202, + 0.00962429866194725, + 0.006613192614167929, + 0.005662861280143261, + 0.007246728986501694, + 0.005026966333389282, + 0.006252730265259743, + 0.006291436962783337, + 0.008475677110254765, + 0.01200833823531866, + 0.013679456897079945, + 0.00695075374096632, + 0.009207682684063911, + 0.0038037223275750875, + 0.01262331660836935, + 0.012385829351842403, + 0.012311718426644802, + 0.015400292351841927, + 0.01816105656325817, + 0.009402656927704811, + 0.00924928393214941, + 0.0068428185768425465, + 0.010876928456127644, + 0.012245949357748032, + 0.016948651522397995, + 0.016393670812249184, + 0.012533899396657944, + 0.01412119623273611, + 0.016308706253767014, + 0.02247239463031292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02058335766196251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798762083053589 + ], + "opposite_logprob_tokens": [ + -1.8048762083053589 + ], + "logprob_reward": -0.1798762083053589, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798762083053589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1167338564991951, + 0.098773293197155, + 0.03641822934150696, + 0.06742701679468155, + 0.13990728557109833, + 0.04999009519815445, + 0.034821026027202606, + 0.11607964336872101, + 0.1296779066324234, + 0.21017161011695862 + ], + "credit_map": [ + 0.011752252466976643, + 0.011863481253385544, + 0.018052436411380768, + 0.021334243938326836, + 0.013529165647923946, + 0.012349555268883705, + 0.017884939908981323, + 0.009967793710529804, + 0.008768708445131779, + 0.01084336917847395, + 0.015654470771551132, + 0.016051603481173515, + 0.016578566282987595, + 0.008255758322775364, + 0.018001297488808632, + 0.0046195234172046185, + 0.005880801938474178, + 0.006263181567192078, + 0.004262326285243034, + 0.00537562370300293, + 0.0033827379811555147, + 0.0037023478653281927, + 0.005288111511617899, + 0.002263100119307637, + 0.00711625162512064, + 0.00782583374530077, + 0.010782630182802677, + 0.010032004676759243, + 0.00598255405202508, + 0.0074703264981508255, + 0.012465567328035831, + 0.0057518561370670795, + 0.020558565855026245, + 0.01616652123630047, + 0.02528449520468712, + 0.02563047781586647, + 0.0137333357706666, + 0.01490392629057169, + 0.014413747936487198, + 0.009216221049427986, + 0.004797769710421562, + 0.005550162401050329, + 0.006166433449834585, + 0.011335783638060093, + 0.006090937182307243, + 0.003813595976680517, + 0.00964219681918621, + 0.0025932195130735636, + 0.0033141267485916615, + 0.004005181137472391, + 0.004535292275249958, + 0.004915610421448946, + 0.0035406590905040503, + 0.007442594040185213, + 0.0037613818421959877, + 0.003306182101368904, + 0.01123098935931921, + 0.011480966582894325, + 0.017382752150297165, + 0.02028106339275837, + 0.024546988308429718, + 0.008507912047207355, + 0.015139053575694561, + 0.007509930524975061, + 0.013052906841039658, + 0.013565219938755035, + 0.015995226800441742, + 0.01812174916267395, + 0.028119340538978577, + 0.017376665025949478, + 0.01374712586402893, + 0.009699680842459202, + 0.01882062293589115, + 0.024727098643779755, + 0.02866169810295105, + 0.03224092721939087, + 0.021784048527479172, + 0.02422145940363407, + 0.025230983272194862, + 0.03448478505015373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.1798762083053589 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar's body patterned with lime green, canary yellow, and velvety black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the caterpillar covered in fine, translucent white hairs that create a glowing halo effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the caterpillar's body visibly tense with an arched back and parted jet-black jaws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background rendered with a shallow depth of field showing soft, painterly dappled light and foliage bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019383705221116543 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019383705221116543, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019383705221116543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09812521189451218, + 0.0530347116291523, + 0.13488376140594482, + 0.09687633812427521, + 0.04758331552147865, + 0.15994948148727417, + 0.0715055987238884, + 0.053087566047906876, + 0.1686849296092987, + 0.11626908928155899 + ], + "credit_map": [ + 0.013051547110080719, + 0.008916929364204407, + 0.013520784676074982, + 0.009082036092877388, + 0.010401656851172447, + 0.017308160662651062, + 0.01236726064234972, + 0.013476833701133728, + 0.005597063805907965, + 0.0060612717643380165, + 0.007211090065538883, + 0.008722388185560703, + 0.005814246367663145, + 0.0063691167160868645, + 0.007663968950510025, + 0.005595569498836994, + 0.02102608233690262, + 0.015017773024737835, + 0.028280852362513542, + 0.019252626225352287, + 0.011402017436921597, + 0.012023095972836018, + 0.020959289744496346, + 0.006922029424458742, + 0.010425718501210213, + 0.012839237228035927, + 0.012567169964313507, + 0.013982372358441353, + 0.008991198614239693, + 0.011957140639424324, + 0.018517524003982544, + 0.007595971692353487, + 0.0046528843231499195, + 0.006006928626447916, + 0.006554995197802782, + 0.007907106541097164, + 0.006930191069841385, + 0.0058573042042553425, + 0.005886928644031286, + 0.0037869762163609266, + 0.018862657248973846, + 0.016458377242088318, + 0.022088956087827682, + 0.030683714896440506, + 0.020532220602035522, + 0.017836377024650574, + 0.01957119256258011, + 0.013915984891355038, + 0.00798671506345272, + 0.009199083782732487, + 0.010164417326450348, + 0.013677341863512993, + 0.008603387512266636, + 0.007272173184901476, + 0.009423913434147835, + 0.005178571213036776, + 0.00724196108058095, + 0.007229093462228775, + 0.008177350275218487, + 0.010266363620758057, + 0.00558076798915863, + 0.00534497294574976, + 0.005596332252025604, + 0.003650725120678544, + 0.03302283585071564, + 0.030500277876853943, + 0.02128778211772442, + 0.02487134002149105, + 0.017970038577914238, + 0.014190997928380966, + 0.013292485848069191, + 0.01354916486889124, + 0.010524083860218525, + 0.010781215503811836, + 0.016319910064339638, + 0.013446510769426823, + 0.013672981411218643, + 0.012747413478791714, + 0.01910971663892269, + 0.019667254760861397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.0019383705221116543 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_0/index_tensor([8870, 8870])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_0/index_tensor([8870, 8870])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fe3a4b050c6441a7a323fc76d9ced912d80027c1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_0/index_tensor([8870, 8870])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, drone-like swooping shot smoothly descends from a high vantage point across a room illuminated by soft sunlight, ultimately stabilizing in an extreme close-up, focusing on a pair of slender, white hands. These hands delicately move across a soft cotton cloth spread across a light-colored wooden table with distinct wood grain. The cloth is a soft off-white, adorned with watercolor animal prints—a dozing bear and a rabbit chasing butterflies—in a gentle, warm hue. Gently, fingers spread out, using the pads of their fingers and the sides of their palms to methodically smooth out each subtle wrinkle. The movements are steady and loving, as if performing a sacred ritual. A ray of warm afternoon sunlight slants in from outside the frame, casting a golden halo over the scene, clearly outlining the fingers and highlighting the delicate texture of the cotton cloth. The entire image is filled with warm, soft tones, and the interplay of light and shadow permeates a profound sense of tranquility and a gentle anticipation for the arrival of new life.", + "seed": 20304, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_0/index_8870_seed_20304_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.25694922376424073, + "vlm_logprob": -0.25694922376424073, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.25694922376424073, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2315586805343628 + ], + "opposite_logprob_tokens": [ + -0.3565587103366852 + ], + "logprob_reward": -1.2315586805343628, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2315586805343628, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12270993739366531, + 0.15042439103126526, + 0.161863774061203, + 0.07146094739437103, + 0.06138363480567932, + 0.1387866884469986, + 0.12121782451868057, + 0.03993436321616173, + 0.08241736143827438, + 0.04980107769370079 + ], + "credit_map": [ + 0.015841174870729446, + 0.0132759939879179, + 0.018676722422242165, + 0.013346860185265541, + 0.014219360426068306, + 0.018236977979540825, + 0.015576492995023727, + 0.013536356389522552, + 0.01618192344903946, + 0.027783259749412537, + 0.011634589172899723, + 0.015327053144574165, + 0.018562287092208862, + 0.02313866838812828, + 0.015858115628361702, + 0.021938489750027657, + 0.011888619512319565, + 0.014401049353182316, + 0.023641327396035194, + 0.020081324502825737, + 0.011061272583901882, + 0.02814297378063202, + 0.029049241915345192, + 0.023597972467541695, + 0.005726627539843321, + 0.009042302146553993, + 0.008481304161250591, + 0.012733167968690395, + 0.00829655583947897, + 0.007502892054617405, + 0.01331905648112297, + 0.006359039805829525, + 0.004424326587468386, + 0.007085505407303572, + 0.007710622623562813, + 0.011854887939989567, + 0.006830273661762476, + 0.006107328925281763, + 0.011727425269782543, + 0.00564326997846365, + 0.011899776756763458, + 0.042603250592947006, + 0.022013138979673386, + 0.01674775779247284, + 0.01744302734732628, + 0.007246353663504124, + 0.012704718858003616, + 0.00812867283821106, + 0.016786767169833183, + 0.013622443191707134, + 0.014127095229923725, + 0.018636982887983322, + 0.014271636493504047, + 0.008362322114408016, + 0.026005426421761513, + 0.009405160322785378, + 0.0039771380834281445, + 0.006630961317569017, + 0.005579985212534666, + 0.005847910884767771, + 0.0041503943502902985, + 0.0023925676941871643, + 0.005726630799472332, + 0.005628776270896196, + 0.006579148583114147, + 0.009963084943592548, + 0.021278923377394676, + 0.008276431821286678, + 0.007731396239250898, + 0.005818203557282686, + 0.006200112402439117, + 0.01657005585730076, + 0.005186272319406271, + 0.011751302517950535, + 0.004282442387193441, + 0.00669872248545289, + 0.003486139699816704, + 0.002444274490699172, + 0.0049722036346793175, + 0.010979724116623402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -1.2315586805343628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009811155498027802 + ], + "opposite_logprob_tokens": [ + -4.634810924530029 + ], + "logprob_reward": -0.009811155498027802, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009811155498027802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24269068241119385, + 0.18922284245491028, + 0.1353992223739624, + 0.07286760956048965, + 0.04975258558988571, + 0.05970126762986183, + 0.11374542862176895, + 0.038396839052438736, + 0.05878989025950432, + 0.039433643221855164 + ], + "credit_map": [ + 0.029100770130753517, + 0.025992607697844505, + 0.03290816396474838, + 0.023622075095772743, + 0.02180090919137001, + 0.052746303379535675, + 0.026104386895895004, + 0.03041546605527401, + 0.01937447115778923, + 0.03288939222693443, + 0.012866385281085968, + 0.02049447037279606, + 0.024723947048187256, + 0.030016383156180382, + 0.025416461750864983, + 0.023441312834620476, + 0.010293981060385704, + 0.012101980857551098, + 0.018418893218040466, + 0.014700086787343025, + 0.009820427745580673, + 0.03305963799357414, + 0.018812278285622597, + 0.018191933631896973, + 0.006343992426991463, + 0.0085777398198843, + 0.008543035946786404, + 0.012236024253070354, + 0.008825073018670082, + 0.007684408687055111, + 0.012612841092050076, + 0.008044496178627014, + 0.003769845003262162, + 0.006436923984438181, + 0.006200192030519247, + 0.010026685893535614, + 0.00516207329928875, + 0.004946919158101082, + 0.009296075440943241, + 0.003913871943950653, + 0.0056715779937803745, + 0.015594457276165485, + 0.0072348532266914845, + 0.008205290883779526, + 0.007519131992012262, + 0.005141517613083124, + 0.007235256489366293, + 0.003099181456491351, + 0.015925170853734016, + 0.01417597383260727, + 0.014229843392968178, + 0.01736386865377426, + 0.010412530042231083, + 0.013036740012466908, + 0.020179638639092445, + 0.008421667851507664, + 0.0038384702056646347, + 0.005673444829881191, + 0.0052642375230789185, + 0.006610467564314604, + 0.003706866642460227, + 0.003655653214082122, + 0.005498644895851612, + 0.004149056971073151, + 0.0044629815965890884, + 0.003260623198002577, + 0.017582105472683907, + 0.005564210936427116, + 0.004616748075932264, + 0.00566020468249917, + 0.004061195068061352, + 0.013581822626292706, + 0.004049404989928007, + 0.004847581032663584, + 0.0037215433549135923, + 0.0037940407637506723, + 0.00500368420034647, + 0.004340485669672489, + 0.004441319964826107, + 0.009235585108399391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009811155498027802 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087710037827492 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.011087710037827492, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087710037827492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1760922521352768, + 0.11869993805885315, + 0.0654132291674614, + 0.05456922575831413, + 0.07090488821268082, + 0.1048804521560669, + 0.10226354748010635, + 0.05474723130464554, + 0.1948148012161255, + 0.057614512741565704 + ], + "credit_map": [ + 0.025291964411735535, + 0.024736648425459862, + 0.021344194188714027, + 0.01660204865038395, + 0.022489147260785103, + 0.019880805164575577, + 0.024810971692204475, + 0.020936476066708565, + 0.011422296054661274, + 0.022720474749803543, + 0.010957992635667324, + 0.015025019645690918, + 0.011485662311315536, + 0.017794860526919365, + 0.013430815190076828, + 0.01586281508207321, + 0.005074674729257822, + 0.010356306098401546, + 0.007173472084105015, + 0.00762707507237792, + 0.004817701410502195, + 0.008914434351027012, + 0.010972483083605766, + 0.010477082803845406, + 0.0046464488841593266, + 0.00786067545413971, + 0.006609193980693817, + 0.007621689699590206, + 0.007877211086452007, + 0.005624502897262573, + 0.009047325700521469, + 0.005282176658511162, + 0.0050803134217858315, + 0.010016467422246933, + 0.009010828100144863, + 0.012667893432080746, + 0.007678845897316933, + 0.006980824749916792, + 0.01111578568816185, + 0.008353929966688156, + 0.012168819084763527, + 0.030033456161618233, + 0.011994852684438229, + 0.012816471047699451, + 0.013384323567152023, + 0.006674811709672213, + 0.010643760673701763, + 0.007163956295698881, + 0.015120109543204308, + 0.01841265708208084, + 0.013728419318795204, + 0.012909402139484882, + 0.012216840870678425, + 0.009145609103143215, + 0.012761248275637627, + 0.007969261147081852, + 0.004868718795478344, + 0.006673947907984257, + 0.009502043016254902, + 0.011970974504947662, + 0.004076838959008455, + 0.003066033124923706, + 0.009930511936545372, + 0.004658164456486702, + 0.019381390884518623, + 0.019975736737251282, + 0.03518691658973694, + 0.01883089914917946, + 0.018654370680451393, + 0.02352643758058548, + 0.025545887649059296, + 0.03371317312121391, + 0.0051376172341406345, + 0.014498776756227016, + 0.005346311256289482, + 0.005001870449632406, + 0.003544773440808058, + 0.002957558026537299, + 0.00430639460682869, + 0.01682121306657791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.011087710037827492 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946326069533825 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946326069533825, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946326069533825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1036093458533287, + 0.07876322418451309, + 0.05403641611337662, + 0.03408488258719444, + 0.02841068245470524, + 0.16408376395702362, + 0.0850973054766655, + 0.03749319165945053, + 0.16110441088676453, + 0.2533167600631714 + ], + "credit_map": [ + 0.018647605553269386, + 0.011134627275168896, + 0.012390130199491978, + 0.013584237545728683, + 0.012448623776435852, + 0.014043527655303478, + 0.011331223882734776, + 0.010029366239905357, + 0.009145231917500496, + 0.018591294065117836, + 0.006053119897842407, + 0.007579589728266001, + 0.009353131987154484, + 0.010245831683278084, + 0.008570819161832333, + 0.009224210865795612, + 0.004034165758639574, + 0.006620464846491814, + 0.006158687639981508, + 0.006560354959219694, + 0.004089558031409979, + 0.01023093517869711, + 0.008815877139568329, + 0.007526373025029898, + 0.003133683232590556, + 0.004506511613726616, + 0.0038187180180102587, + 0.006043322384357452, + 0.005029067397117615, + 0.002846864517778158, + 0.005454841535538435, + 0.003251871792599559, + 0.00255209649913013, + 0.003791198367252946, + 0.0035850100684911013, + 0.00385499675758183, + 0.0035553646739572287, + 0.0031281509436666965, + 0.005532171577215195, + 0.0024116921704262495, + 0.011250196024775505, + 0.04057411849498749, + 0.02910803258419037, + 0.025070957839488983, + 0.018716400489211082, + 0.009061519056558609, + 0.019326889887452126, + 0.01097564771771431, + 0.010517987422645092, + 0.01087931077927351, + 0.010935061611235142, + 0.01326373778283596, + 0.0076731168664991856, + 0.007464289665222168, + 0.017372416332364082, + 0.0069913859479129314, + 0.004594218917191029, + 0.00713361706584692, + 0.005002920515835285, + 0.0056519596837460995, + 0.0037259750533849, + 0.0030651630368083715, + 0.004548895172774792, + 0.0037704436108469963, + 0.01485899742692709, + 0.012713378295302391, + 0.041989006102085114, + 0.01711444742977619, + 0.011415576562285423, + 0.013646983541548252, + 0.011763107031583786, + 0.03760291635990143, + 0.028526360169053078, + 0.04165063053369522, + 0.04115501791238785, + 0.032726410776376724, + 0.02409428358078003, + 0.022086458280682564, + 0.032057713717222214, + 0.031019890680909157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.005946326069533825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342246681451797 + ], + "opposite_logprob_tokens": [ + -3.6513421535491943 + ], + "logprob_reward": -0.026342246681451797, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342246681451797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07510564476251602, + 0.28844234347343445, + 0.09127508848905563, + 0.04105020686984062, + 0.056257281452417374, + 0.15040986239910126, + 0.10330674797296524, + 0.03850719705224037, + 0.0781121551990509, + 0.07753340899944305 + ], + "credit_map": [ + 0.010441925376653671, + 0.01381081622093916, + 0.010635662823915482, + 0.008470919914543629, + 0.010147699154913425, + 0.005043697543442249, + 0.007967083714902401, + 0.008587837219238281, + 0.03219102323055267, + 0.06443367898464203, + 0.02412889339029789, + 0.032576560974121094, + 0.03695271536707878, + 0.0365302674472332, + 0.031300581991672516, + 0.03032863885164261, + 0.007546237204223871, + 0.015375572256743908, + 0.010419316589832306, + 0.011528384871780872, + 0.007874186150729656, + 0.010696826502680779, + 0.016217028722167015, + 0.01161753199994564, + 0.0036915901582688093, + 0.004708993248641491, + 0.004856111481785774, + 0.006464520003646612, + 0.005082011688500643, + 0.0035919996444135904, + 0.008523228578269482, + 0.004131750203669071, + 0.004786467179656029, + 0.008798662573099136, + 0.008180546574294567, + 0.011145885102450848, + 0.006248335354030132, + 0.005543997045606375, + 0.007584070321172476, + 0.00396931916475296, + 0.015101165510714054, + 0.03312349691987038, + 0.01408196147531271, + 0.018334273248910904, + 0.016168782487511635, + 0.015306196175515652, + 0.02392251417040825, + 0.014371465891599655, + 0.009587668813765049, + 0.010944112204015255, + 0.01356811635196209, + 0.018116099759936333, + 0.008385353721678257, + 0.008135758340358734, + 0.026850519701838493, + 0.007719113957136869, + 0.004002604633569717, + 0.009070204570889473, + 0.005463508423417807, + 0.0044989134185016155, + 0.0045132408849895, + 0.0030143039766699076, + 0.0038962014950811863, + 0.004048219416290522, + 0.00546133890748024, + 0.013169439509510994, + 0.016512524336576462, + 0.0069631729274988174, + 0.007840190082788467, + 0.005100658629089594, + 0.008097711950540543, + 0.014967112801969051, + 0.007002126891165972, + 0.014079546555876732, + 0.006020925007760525, + 0.009231967851519585, + 0.007412567734718323, + 0.008101396262645721, + 0.008540238253772259, + 0.017144639045000076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.026342246681451797 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, drone-like swooping shot smoothly descends from a high vantage point across a room illuminated by soft sunlight, ultimately stabilizing in an extreme close-up, focusing on a pair of slender, white hands. These hands delicately move across a soft cotton cloth spread across a light-colored wooden table with distinct wood grain. The cloth is a soft off-white, adorned with watercolor animal prints—a dozing bear and a rabbit chasing butterflies—in a gentle, warm hue. Gently, fingers spread out, using the pads of their fingers and the sides of their palms to methodically smooth out each subtle wrinkle. The movements are steady and loving, as if performing a sacred ritual. A ray of warm afternoon sunlight slants in from outside the frame, casting a golden halo over the scene, clearly outlining the fingers and highlighting the delicate texture of the cotton cloth. The entire image is filled with warm, soft tones, and the interplay of light and shadow permeates a profound sense of tranquility and a gentle anticipation for the arrival of new life.", + "seed": 20305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_0/index_8870_seed_20305_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.20588188888505102, + "vlm_logprob": -0.20588188888505102, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.20588188888505102, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9815952777862549 + ], + "opposite_logprob_tokens": [ + -0.4815952479839325 + ], + "logprob_reward": -0.9815952777862549, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9815952777862549, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.40313223004341125, + 0.2637726962566376, + 0.037139903753995895, + 0.015700820833444595, + 0.06346888840198517, + 0.017251787707209587, + 0.035776425153017044, + 0.11600393801927567, + 0.02675618976354599, + 0.02099715545773506 + ], + "credit_map": [ + 0.05567122995853424, + 0.09355141967535019, + 0.08677312731742859, + 0.07513245195150375, + 0.0125938281416893, + 0.006008822936564684, + 0.03160440921783447, + 0.04179692640900612, + 0.007672748062759638, + 0.10616102814674377, + 0.008365786634385586, + 0.10853318125009537, + 0.010056569240987301, + 0.004572881851345301, + 0.012633994221687317, + 0.005776511505246162, + 0.0031572445295751095, + 0.003803463652729988, + 0.005373088177293539, + 0.007469505537301302, + 0.0032141869887709618, + 0.003410057397559285, + 0.005866564344614744, + 0.004845792893320322, + 0.001885628211311996, + 0.0035400856286287308, + 0.0018245783867314458, + 0.0022142347879707813, + 0.0015268996357917786, + 0.0016928331460803747, + 0.001746624824590981, + 0.0012699351646006107, + 0.006194060668349266, + 0.008365506306290627, + 0.006079306825995445, + 0.01173482183367014, + 0.009374207817018032, + 0.0050680553540587425, + 0.009124708361923695, + 0.007528223562985659, + 0.0019632307812571526, + 0.0033824879210442305, + 0.0023322459310293198, + 0.0024343845434486866, + 0.001550330314785242, + 0.0012684593675658107, + 0.0027240333147346973, + 0.0015966161154210567, + 0.007914531044661999, + 0.007042677141726017, + 0.005535840056836605, + 0.0021338327787816525, + 0.0031408811919391155, + 0.0024051223881542683, + 0.002357868477702141, + 0.005245668813586235, + 0.019601907581090927, + 0.009617352858185768, + 0.013087470084428787, + 0.012534460052847862, + 0.009631110355257988, + 0.00891345925629139, + 0.012673821300268173, + 0.02994435466825962, + 0.002294507110491395, + 0.002375740557909012, + 0.002912771189585328, + 0.003501043189316988, + 0.0036306052934378386, + 0.0025344963651150465, + 0.005466997157782316, + 0.004040028899908066, + 0.0018297412898391485, + 0.0013414734276011586, + 0.003718538675457239, + 0.0016727808397263288, + 0.002158485120162368, + 0.0025943350046873093, + 0.005966794211417437, + 0.0017150052590295672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.9815952777862549 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005273835267871618 + ], + "opposite_logprob_tokens": [ + -5.255273818969727 + ], + "logprob_reward": -0.005273835267871618, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005273835267871618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3161937892436981, + 0.06724254786968231, + 0.028677279129624367, + 0.01705118827521801, + 0.060424115508794785, + 0.026809385046362877, + 0.030759766697883606, + 0.4229609966278076, + 0.014476366341114044, + 0.015404587611556053 + ], + "credit_map": [ + 0.04963555559515953, + 0.06694697588682175, + 0.07523897290229797, + 0.04912026226520538, + 0.008778203278779984, + 0.005214205477386713, + 0.027844829484820366, + 0.033414795994758606, + 0.002317615319043398, + 0.025001026690006256, + 0.0032249647192656994, + 0.024616926908493042, + 0.0036670893896371126, + 0.002208002610132098, + 0.003237268654629588, + 0.002969651482999325, + 0.002451798412948847, + 0.0032714465633034706, + 0.005430842749774456, + 0.007211248390376568, + 0.0020473969634622335, + 0.0018555857241153717, + 0.0038652007933706045, + 0.00254375790245831, + 0.0020435061305761337, + 0.002779135713353753, + 0.0018991397228091955, + 0.003245837986469269, + 0.0014632911188527942, + 0.0011709023965522647, + 0.0023617290426045656, + 0.0020876452326774597, + 0.005097176413983107, + 0.009236014448106289, + 0.006226584780961275, + 0.011502724140882492, + 0.00775701692327857, + 0.005531959235668182, + 0.007745934184640646, + 0.007326704915612936, + 0.0053679379634559155, + 0.00334380054846406, + 0.003560945624485612, + 0.0034091095440089703, + 0.0019062396604567766, + 0.003229565918445587, + 0.0032498217187821865, + 0.0027419631369411945, + 0.007059652358293533, + 0.004540706984698772, + 0.004643118008971214, + 0.0032445192337036133, + 0.0025684696156531572, + 0.0018414364894852042, + 0.0020761543419212103, + 0.004785709083080292, + 0.0652039498090744, + 0.03371309116482735, + 0.049093641340732574, + 0.04893237724900246, + 0.03746588155627251, + 0.03380480408668518, + 0.04729418084025383, + 0.10745307803153992, + 0.001264039077796042, + 0.0012637015897780657, + 0.001675873063504696, + 0.0018699023639783263, + 0.0016967698466032743, + 0.0012761459220200777, + 0.0032274569384753704, + 0.002202477538958192, + 0.0018701355438679457, + 0.001678596599958837, + 0.0029007347766309977, + 0.002150429179891944, + 0.0018954711267724633, + 0.001481386017985642, + 0.0017941019032150507, + 0.0016337325796484947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.005273835267871618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647731505334377 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008647731505334377, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647731505334377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4038853943347931, + 0.28001683950424194, + 0.027627650648355484, + 0.03357361629605293, + 0.01991869881749153, + 0.037164509296417236, + 0.046128422021865845, + 0.06194787845015526, + 0.059573154896497726, + 0.030163832008838654 + ], + "credit_map": [ + 0.04300640523433685, + 0.1251235455274582, + 0.041354186832904816, + 0.027057867497205734, + 0.019720975309610367, + 0.010209041647613049, + 0.04186781123280525, + 0.09554556757211685, + 0.0065612588077783585, + 0.11110620945692062, + 0.007921610958874226, + 0.11945772916078568, + 0.010078834369778633, + 0.004962645471096039, + 0.013959148898720741, + 0.005969425663352013, + 0.002294892445206642, + 0.005724204238504171, + 0.005409852135926485, + 0.005279285367578268, + 0.0013563346583396196, + 0.0016355407424271107, + 0.0028339631389826536, + 0.003093578852713108, + 0.002409783424809575, + 0.0066246455535292625, + 0.004094068892300129, + 0.004203494172543287, + 0.002618648111820221, + 0.004605282098054886, + 0.005205425433814526, + 0.00381226628087461, + 0.0019301708089187741, + 0.0026932943146675825, + 0.002342220162972808, + 0.0033826944418251514, + 0.002514156047254801, + 0.001916303881444037, + 0.002730616135522723, + 0.002409243956208229, + 0.002953785238787532, + 0.003693363396450877, + 0.0035221518483012915, + 0.005880434066057205, + 0.0037045960780233145, + 0.005980190355330706, + 0.00642680237069726, + 0.005003183614462614, + 0.009174714796245098, + 0.0069278632290661335, + 0.006220001261681318, + 0.005381245166063309, + 0.004579752217978239, + 0.0035494682379066944, + 0.0034604487009346485, + 0.006834927946329117, + 0.010696473531425, + 0.004038343206048012, + 0.007838236168026924, + 0.006934493314474821, + 0.004143097437918186, + 0.004559056367725134, + 0.006711730267852545, + 0.017026444897055626, + 0.006874928716570139, + 0.004378446843475103, + 0.006433076225221157, + 0.008565298281610012, + 0.007046611979603767, + 0.0048672910779714584, + 0.013595069758594036, + 0.007812430616468191, + 0.0028561106882989407, + 0.0015322210965678096, + 0.004669083748012781, + 0.0032191856298595667, + 0.0025382728781551123, + 0.004313476383686066, + 0.0077860672026872635, + 0.0032494128681719303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.008647731505334377 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092888906598091 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092888906598091, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092888906598091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.473956435918808, + 0.16966475546360016, + 0.05829092860221863, + 0.036472853273153305, + 0.0657268613576889, + 0.01766214519739151, + 0.03318401426076889, + 0.09162314981222153, + 0.027990931645035744, + 0.02542785368859768 + ], + "credit_map": [ + 0.04523688927292824, + 0.12165407836437225, + 0.0747162252664566, + 0.09382575750350952, + 0.018280547112226486, + 0.012356769293546677, + 0.04410591721534729, + 0.0637803003191948, + 0.0071599241346120834, + 0.06417984515428543, + 0.008684511296451092, + 0.06291668862104416, + 0.007412414066493511, + 0.0038278321735560894, + 0.00888711865991354, + 0.0065964143723249435, + 0.004722739569842815, + 0.011964266188442707, + 0.011288641020655632, + 0.010359679348766804, + 0.0030665560625493526, + 0.00377262057736516, + 0.007223039399832487, + 0.005893385037779808, + 0.0037020074669271708, + 0.008213694207370281, + 0.0046530417166650295, + 0.005440558306872845, + 0.0025654698256403208, + 0.004118718206882477, + 0.004238905850797892, + 0.0035404565278440714, + 0.005718334577977657, + 0.012751969508826733, + 0.008485513739287853, + 0.010342329740524292, + 0.009081771597266197, + 0.006660914979875088, + 0.0059889378026127815, + 0.006697083357721567, + 0.0019952242728322744, + 0.003033633343875408, + 0.0022276914678514004, + 0.0026482397224754095, + 0.0015966530190780759, + 0.002216438762843609, + 0.002207448473200202, + 0.0017368169501423836, + 0.005955359898507595, + 0.0070474473759531975, + 0.004870500415563583, + 0.0025073965080082417, + 0.003490596078336239, + 0.0025508624967187643, + 0.002233380451798439, + 0.0045284731313586235, + 0.014474118128418922, + 0.0066047003492712975, + 0.01062744576483965, + 0.010274498723447323, + 0.007030321750789881, + 0.006921824999153614, + 0.010569943115115166, + 0.02512029930949211, + 0.0032436500769108534, + 0.002382389735430479, + 0.003647489007562399, + 0.0040870183147490025, + 0.0032682421151548624, + 0.002032980788499117, + 0.005713288206607103, + 0.0036158740986138582, + 0.00234444928355515, + 0.0015345700085163116, + 0.0038511890452355146, + 0.0021807553712278605, + 0.003026314778253436, + 0.003235420910641551, + 0.006907982751727104, + 0.002347173634916544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004092888906598091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979971095919609 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979971095919609, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979971095919609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0968267098069191, + 0.1051904484629631, + 0.1164945587515831, + 0.05769887939095497, + 0.3626539707183838, + 0.047731660306453705, + 0.04190065711736679, + 0.11195171624422073, + 0.037772610783576965, + 0.021778767928481102 + ], + "credit_map": [ + 0.005741081200540066, + 0.027828751131892204, + 0.007656990084797144, + 0.00852376688271761, + 0.00450254837051034, + 0.0032407050020992756, + 0.009843889623880386, + 0.029488984495401382, + 0.005662690382450819, + 0.03579980880022049, + 0.007233901880681515, + 0.03898994252085686, + 0.004565337207168341, + 0.002780177863314748, + 0.005421483889222145, + 0.004737115930765867, + 0.010488024912774563, + 0.013020772486925125, + 0.023514103144407272, + 0.023144856095314026, + 0.007529906462877989, + 0.009383540600538254, + 0.018009724095463753, + 0.011403623968362808, + 0.004316355567425489, + 0.007525925524532795, + 0.006639900617301464, + 0.007707517594099045, + 0.004483054857701063, + 0.004824967589229345, + 0.011418754234910011, + 0.01078240480273962, + 0.031192701309919357, + 0.06965836882591248, + 0.04848779737949371, + 0.05632390081882477, + 0.042329393327236176, + 0.038058578968048096, + 0.03336574509739876, + 0.04323749616742134, + 0.004848124925047159, + 0.0069573610089719296, + 0.004891267511993647, + 0.007377144414931536, + 0.00551601080223918, + 0.007613792549818754, + 0.005952888168394566, + 0.004575069528073072, + 0.008958486840128899, + 0.005820224527269602, + 0.00579168600961566, + 0.004060945939272642, + 0.004941211547702551, + 0.0034156807232648134, + 0.0031101524364203215, + 0.005802268628031015, + 0.0175258070230484, + 0.009139718487858772, + 0.012731138616800308, + 0.012265785597264767, + 0.009608873166143894, + 0.008877303451299667, + 0.012947137467563152, + 0.02885594591498375, + 0.0049754041247069836, + 0.0029512301553040743, + 0.0051265861839056015, + 0.005012003239244223, + 0.004565845243632793, + 0.003338196547701955, + 0.007621411234140396, + 0.0041819349862635136, + 0.0016782316379249096, + 0.001741038984619081, + 0.002350670052692294, + 0.0015992658445611596, + 0.0029917804058641195, + 0.0033183349296450615, + 0.005998315755277872, + 0.002101130783557892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.02979971095919609 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, drone-like swooping shot smoothly descends from a high vantage point across a room illuminated by soft sunlight, ultimately stabilizing in an extreme close-up, focusing on a pair of slender, white hands. These hands delicately move across a soft cotton cloth spread across a light-colored wooden table with distinct wood grain. The cloth is a soft off-white, adorned with watercolor animal prints—a dozing bear and a rabbit chasing butterflies—in a gentle, warm hue. Gently, fingers spread out, using the pads of their fingers and the sides of their palms to methodically smooth out each subtle wrinkle. The movements are steady and loving, as if performing a sacred ritual. A ray of warm afternoon sunlight slants in from outside the frame, casting a golden halo over the scene, clearly outlining the fingers and highlighting the delicate texture of the cotton cloth. The entire image is filled with warm, soft tones, and the interplay of light and shadow permeates a profound sense of tranquility and a gentle anticipation for the arrival of new life.", + "seed": 20306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_0/index_8870_seed_20306_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2077385526150465, + "vlm_logprob": -0.2077385526150465, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.2077385526150465, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9820745587348938 + ], + "opposite_logprob_tokens": [ + -0.4820745587348938 + ], + "logprob_reward": -0.9820745587348938, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9820745587348938, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.018818605691194534, + 0.015388406813144684, + 0.042779333889484406, + 0.01582310162484646, + 0.009154652245342731, + 0.8550734519958496, + 0.006913084536790848, + 0.008460074663162231, + 0.008523527532815933, + 0.019065724685788155 + ], + "credit_map": [ + 0.0021812424529343843, + 0.0024387806188315153, + 0.002042402746155858, + 0.0019539210479706526, + 0.0031021402683109045, + 0.0014767093816772103, + 0.0037516728043556213, + 0.0018717384664341807, + 0.001695981714874506, + 0.0021031000651419163, + 0.002717141294851899, + 0.0025676628574728966, + 0.0018067726632580161, + 0.0015879119746387005, + 0.0012172725982964039, + 0.0016925649251788855, + 0.002692239359021187, + 0.010891472920775414, + 0.005329323001205921, + 0.005000490695238113, + 0.006043524947017431, + 0.003042681375518441, + 0.0036128279753029346, + 0.006166775245219469, + 0.0012536203721538186, + 0.0017679206794127822, + 0.0017956830561161041, + 0.0022190625313669443, + 0.002823134185746312, + 0.0021080244332551956, + 0.0020720490720123053, + 0.0017836090410128236, + 0.0010070246644318104, + 0.001268611173145473, + 0.0008249111706390977, + 0.0012794287176802754, + 0.0012703221291303635, + 0.0011908377055078745, + 0.0011023979168385267, + 0.0012111192336305976, + 0.020654980093240738, + 0.030044473707675934, + 0.03870074823498726, + 0.01839609071612358, + 0.02990817092359066, + 0.3335038125514984, + 0.3422435522079468, + 0.04162168130278587, + 0.000696728122420609, + 0.0008860690286383033, + 0.0010496434988453984, + 0.0008169298525899649, + 0.0007717257249169052, + 0.0005709942779503763, + 0.0010420582257211208, + 0.0010789367370307446, + 0.0007320689037442207, + 0.0014365954557433724, + 0.0010741088772192597, + 0.0008073455537669361, + 0.0011437807697802782, + 0.0009938620496541262, + 0.0010423087514936924, + 0.0012300049420446157, + 0.0008639555308036506, + 0.0010893552098423243, + 0.0010191007750108838, + 0.0009234232129529119, + 0.0012950410600751638, + 0.0011762840440496802, + 0.0010591670870780945, + 0.0010972011368721724, + 0.0020506258588284254, + 0.001338234986178577, + 0.002730464329943061, + 0.0011313802096992731, + 0.0038654643576592207, + 0.003786687972024083, + 0.0020306878723204136, + 0.0021321820095181465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.9820745587348938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005965879186987877 + ], + "opposite_logprob_tokens": [ + -5.130965709686279 + ], + "logprob_reward": -0.005965879186987877, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005965879186987877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01950868032872677, + 0.011710096150636673, + 0.02904617413878441, + 0.008612149395048618, + 0.009461152367293835, + 0.8077789545059204, + 0.011621443554759026, + 0.06025199592113495, + 0.022527003660798073, + 0.0194823257625103 + ], + "credit_map": [ + 0.002783154370263219, + 0.00192740885540843, + 0.002699466422200203, + 0.002604438690468669, + 0.0027096637059003115, + 0.0016719152918085456, + 0.003042600117623806, + 0.002070032525807619, + 0.0008028633310459554, + 0.0016021092887967825, + 0.0016458450118079782, + 0.0021707701962441206, + 0.0014197631971910596, + 0.0013913591392338276, + 0.0012460829457268119, + 0.0014313024003058672, + 0.0011817027116194367, + 0.005622574128210545, + 0.002581023145467043, + 0.0059137181378901005, + 0.003224881598725915, + 0.0018287153216078877, + 0.002119100885465741, + 0.006574458908289671, + 0.0006886429619044065, + 0.0007514019962400198, + 0.0007753738318569958, + 0.0015290632145479321, + 0.0011657333234325051, + 0.0012377952225506306, + 0.0013303037267178297, + 0.0011338354088366032, + 0.0012436137767508626, + 0.0015440738061442971, + 0.0005871187895536423, + 0.001324544078670442, + 0.0015659729251638055, + 0.0010329774813726544, + 0.001134520280174911, + 0.001028332277201116, + 0.01950882561504841, + 0.028748804703354836, + 0.03672509640455246, + 0.017105910927057266, + 0.028132347390055656, + 0.31492963433265686, + 0.3232831060886383, + 0.03934521600604057, + 0.0011526652378961444, + 0.0014884108677506447, + 0.0015537728322669864, + 0.001976716099306941, + 0.0014517371309921145, + 0.0009383325814269483, + 0.0016649555182084441, + 0.001394854043610394, + 0.006244280841201544, + 0.006112398114055395, + 0.00858259666711092, + 0.00841055903583765, + 0.009287997148931026, + 0.008080899715423584, + 0.007293510716408491, + 0.006239748559892178, + 0.0016557383351027966, + 0.0022700242698192596, + 0.003039037110283971, + 0.002592877484858036, + 0.003624723991379142, + 0.002423213329166174, + 0.003308780025690794, + 0.0036126114428043365, + 0.0020891851745545864, + 0.0013837972655892372, + 0.002594436751678586, + 0.0011752151185646653, + 0.004013330675661564, + 0.0037884253542870283, + 0.0019314021337777376, + 0.002506533870473504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.005965879186987877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787074290215969 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.009787074290215969, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787074290215969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03418389707803726, + 0.050100795924663544, + 0.05441085994243622, + 0.020848726853728294, + 0.0553724467754364, + 0.690213680267334, + 0.014334318228065968, + 0.013749279081821442, + 0.015486990101635456, + 0.05129905417561531 + ], + "credit_map": [ + 0.0032417005859315395, + 0.003883672645315528, + 0.0027946755290031433, + 0.003516114316880703, + 0.007509039714932442, + 0.003286487888544798, + 0.005381157621741295, + 0.004571044351905584, + 0.00385395810008049, + 0.005503591615706682, + 0.005580353084951639, + 0.009328844025731087, + 0.005440085660666227, + 0.007187648210674524, + 0.007325636222958565, + 0.005880676209926605, + 0.0031846764031797647, + 0.013318642042577267, + 0.005682117771357298, + 0.006404535379260778, + 0.007908317260444164, + 0.004540686029940844, + 0.005848508793860674, + 0.007523372769355774, + 0.0013290578499436378, + 0.0019116358598694205, + 0.0017011241288855672, + 0.0038962140679359436, + 0.0028416900895535946, + 0.0034397125709801912, + 0.003270639106631279, + 0.002458650851622224, + 0.004686878062784672, + 0.008348489180207253, + 0.003810243448242545, + 0.007151208817958832, + 0.008761387318372726, + 0.005961019080132246, + 0.007860061712563038, + 0.008793153800070286, + 0.016693759709596634, + 0.024615691974759102, + 0.03152461349964142, + 0.01478155143558979, + 0.023954929783940315, + 0.26951736211776733, + 0.2754601836204529, + 0.03366551920771599, + 0.001382485032081604, + 0.0018845958402380347, + 0.0018010655185207725, + 0.002391533460468054, + 0.0013619456440210342, + 0.001216583070345223, + 0.002315155928954482, + 0.001980951987206936, + 0.0012424741871654987, + 0.0017133582150563598, + 0.0018160180188715458, + 0.0016907528042793274, + 0.0024951384402811527, + 0.0012952855322510004, + 0.0019151659216731787, + 0.0015810850309208035, + 0.000952394912019372, + 0.001651614555157721, + 0.0021292001474648714, + 0.0017426344566047192, + 0.0019228708697482944, + 0.0023315988946706057, + 0.002141340868547559, + 0.002615332370623946, + 0.005179948639124632, + 0.003931461367756128, + 0.006923937704414129, + 0.002989846281707287, + 0.010587010532617569, + 0.009641757234930992, + 0.005773855373263359, + 0.006271231919527054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.009787074290215969 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071558110415936 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071558110415936, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071558110415936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04705917835235596, + 0.018304575234651566, + 0.0071886684745550156, + 0.032879769802093506, + 0.033419881016016006, + 0.7758533954620361, + 0.008943204768002033, + 0.022484371438622475, + 0.009951971471309662, + 0.043915003538131714 + ], + "credit_map": [ + 0.004402440041303635, + 0.005847112741321325, + 0.003954939544200897, + 0.0056275841780006886, + 0.005672442726790905, + 0.00478933984413743, + 0.010219261050224304, + 0.006546062882989645, + 0.002349540125578642, + 0.002540240529924631, + 0.003618519753217697, + 0.0019653651397675276, + 0.0018805075669661164, + 0.002008819719776511, + 0.001896195113658905, + 0.0020453885663300753, + 0.00046727279550395906, + 0.000889191054739058, + 0.001064498326741159, + 0.0008447823347523808, + 0.0007626686128787696, + 0.001010203268378973, + 0.001152273965999484, + 0.0009977781446650624, + 0.0024785189889371395, + 0.00350262806750834, + 0.003921759780496359, + 0.0057954611256718636, + 0.004038201179355383, + 0.004035420715808868, + 0.004693536553531885, + 0.0044142454862594604, + 0.0035658772103488445, + 0.004238776862621307, + 0.002416372997686267, + 0.004977979697287083, + 0.0052071865648031235, + 0.003612047992646694, + 0.0051615177653729916, + 0.004240123555064201, + 0.01893128640949726, + 0.02808474563062191, + 0.03558387607336044, + 0.016291245818138123, + 0.02716834470629692, + 0.30239537358283997, + 0.3094562292098999, + 0.03794233128428459, + 0.001321165356785059, + 0.0009657558985054493, + 0.0010253259679302573, + 0.0012321302201598883, + 0.0009532468393445015, + 0.0007994950865395367, + 0.001602162141352892, + 0.0010439242469146848, + 0.0022566772531718016, + 0.0027545131742954254, + 0.0031672946643084288, + 0.002805953146889806, + 0.0039725275710225105, + 0.0026084783021360636, + 0.0027177021838724613, + 0.0022012279368937016, + 0.0006701555103063583, + 0.0012785415165126324, + 0.0013681144919246435, + 0.001215802039951086, + 0.0015207575634121895, + 0.0014825264224782586, + 0.0011048687156289816, + 0.0013112056767567992, + 0.0035842088982462883, + 0.0037638917565345764, + 0.00519975321367383, + 0.0033450701739639044, + 0.009306540712714195, + 0.008033878169953823, + 0.0050242142751812935, + 0.00565745122730732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.011071558110415936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793692752718925 + ], + "opposite_logprob_tokens": [ + -3.5297937393188477 + ], + "logprob_reward": -0.029793692752718925, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793692752718925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014554157853126526, + 0.03171428665518761, + 0.00929392222315073, + 0.006317369174212217, + 0.006328382063657045, + 0.8873786330223083, + 0.007137729320675135, + 0.008715304546058178, + 0.004155855625867844, + 0.024404356256127357 + ], + "credit_map": [ + 0.0011886031134054065, + 0.0024506973568350077, + 0.0010517837945371866, + 0.0014365200186148286, + 0.00363163067959249, + 0.0013075161259621382, + 0.0016966919647529721, + 0.001790714217349887, + 0.004647843539714813, + 0.003864831989631057, + 0.006310553755611181, + 0.0041419160552322865, + 0.0031317328102886677, + 0.003237220924347639, + 0.0022841226309537888, + 0.004096063319593668, + 0.0006728741573169827, + 0.0021233719307929277, + 0.0011577106779441237, + 0.0010571798775345087, + 0.0013067288091406226, + 0.0007295184768736362, + 0.0008955610683187842, + 0.0013509774580597878, + 0.0006244336254894733, + 0.0006228822167031467, + 0.000784039671998471, + 0.0009639937779866159, + 0.0008151312358677387, + 0.0008365816320292652, + 0.0009530371171422303, + 0.000717270071618259, + 0.0005726090748794377, + 0.000822636648081243, + 0.00044251699000597, + 0.0010059752967208624, + 0.0007214357028715312, + 0.000734680041205138, + 0.0008972643408924341, + 0.0011312640272080898, + 0.021307675167918205, + 0.03342448174953461, + 0.041381917893886566, + 0.019277719780802727, + 0.031043458729982376, + 0.34584885835647583, + 0.3522702753543854, + 0.042824339121580124, + 0.0008057429804466665, + 0.000748475722502917, + 0.0008065467700362206, + 0.0009019369026646018, + 0.0007047977414913476, + 0.0007381439791060984, + 0.0012731479946523905, + 0.001158936764113605, + 0.0008646253263577819, + 0.0010347767965868115, + 0.0012654219754040241, + 0.0010659833205863833, + 0.0013634758070111275, + 0.0010607854928821325, + 0.001116446335799992, + 0.0009437891421839595, + 0.00047953694593161345, + 0.0003777015954256058, + 0.0005887200823053718, + 0.0005859773955307901, + 0.0005735782324336469, + 0.00036561660817824304, + 0.00048492621863260865, + 0.0006997985183261335, + 0.002347038360312581, + 0.001944590243510902, + 0.0030479454435408115, + 0.001354680978693068, + 0.0051416014321148396, + 0.005152980796992779, + 0.002723397221416235, + 0.0026921213138848543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.029793692752718925 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, drone-like swooping shot smoothly descends from a high vantage point across a room illuminated by soft sunlight, ultimately stabilizing in an extreme close-up, focusing on a pair of slender, white hands. These hands delicately move across a soft cotton cloth spread across a light-colored wooden table with distinct wood grain. The cloth is a soft off-white, adorned with watercolor animal prints—a dozing bear and a rabbit chasing butterflies—in a gentle, warm hue. Gently, fingers spread out, using the pads of their fingers and the sides of their palms to methodically smooth out each subtle wrinkle. The movements are steady and loving, as if performing a sacred ritual. A ray of warm afternoon sunlight slants in from outside the frame, casting a golden halo over the scene, clearly outlining the fingers and highlighting the delicate texture of the cotton cloth. The entire image is filled with warm, soft tones, and the interplay of light and shadow permeates a profound sense of tranquility and a gentle anticipation for the arrival of new life.", + "seed": 20307, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_0/index_8870_seed_20307_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2132699186913669, + "vlm_logprob": -0.2132699186913669, + "vlm_yes_no_margin": 3.2500000059604646, + "vlm_token_credit": -0.2132699186913669, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9803465604782104 + ], + "opposite_logprob_tokens": [ + -0.48034659028053284 + ], + "logprob_reward": -0.9803465604782104, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9803465604782104, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3565714359283447, + 0.0855560302734375, + 0.03167043626308441, + 0.06906815618276596, + 0.058586690574884415, + 0.2551487684249878, + 0.0429309606552124, + 0.04610500484704971, + 0.028546087443828583, + 0.02581649459898472 + ], + "credit_map": [ + 0.057471323758363724, + 0.045122843235731125, + 0.03475360572338104, + 0.029736552387475967, + 0.04071422666311264, + 0.08554166555404663, + 0.03193848952651024, + 0.03129268437623978, + 0.008467826060950756, + 0.009021605364978313, + 0.007440820336341858, + 0.01215800829231739, + 0.012259985320270061, + 0.005098812747746706, + 0.012028418481349945, + 0.019080543890595436, + 0.00216102902777493, + 0.0031209331937134266, + 0.003648023586720228, + 0.0046021779999136925, + 0.0035650997888296843, + 0.0035868901759386063, + 0.006241267081350088, + 0.0047450135461986065, + 0.005246151238679886, + 0.007763019297271967, + 0.008571095764636993, + 0.010668034665286541, + 0.009525753557682037, + 0.008111346513032913, + 0.00839250162243843, + 0.010790243744850159, + 0.0050373319536447525, + 0.007295022252947092, + 0.009986120276153088, + 0.007269421126693487, + 0.006828434765338898, + 0.007842904888093472, + 0.007309672888368368, + 0.007017774041742086, + 0.010333606973290443, + 0.01456921175122261, + 0.018110129982233047, + 0.08256242424249649, + 0.015363945625722408, + 0.013015140779316425, + 0.01925073377788067, + 0.08194351941347122, + 0.005264278966933489, + 0.006626836024224758, + 0.005821137689054012, + 0.0056261601857841015, + 0.005264014936983585, + 0.004705003462731838, + 0.004247442353516817, + 0.00537608377635479, + 0.005421705078333616, + 0.008201539516448975, + 0.008431383408606052, + 0.007123743183910847, + 0.004035156685858965, + 0.003940078429877758, + 0.00449052918702364, + 0.0044608633033931255, + 0.0023511131294071674, + 0.003426320618018508, + 0.0031546608079224825, + 0.0033724536187946796, + 0.0038010189309716225, + 0.004762993194162846, + 0.00350398663431406, + 0.004173537250608206, + 0.0013848863309249282, + 0.00159447209443897, + 0.002001977525651455, + 0.004092483781278133, + 0.006696952041238546, + 0.0035127687733620405, + 0.0034012668766081333, + 0.0031316850800067186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.9803465604782104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561840936541557 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.012561840936541557, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561840936541557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04401126503944397, + 0.1193697601556778, + 0.03456824645400047, + 0.07335561513900757, + 0.08972246199846268, + 0.42687228322029114, + 0.08233649283647537, + 0.0486605279147625, + 0.04207028076052666, + 0.03903313726186752 + ], + "credit_map": [ + 0.005240078549832106, + 0.006254544947296381, + 0.004461309872567654, + 0.0028950355481356382, + 0.005256864707916975, + 0.009540311060845852, + 0.003945622127503157, + 0.006417489610612392, + 0.01256529614329338, + 0.013466540724039078, + 0.012949412688612938, + 0.012755462899804115, + 0.011671294458210468, + 0.008171173743903637, + 0.016743168234825134, + 0.031047403812408447, + 0.002221670001745224, + 0.0030850085895508528, + 0.004701264202594757, + 0.005718936212360859, + 0.004018812905997038, + 0.0025389192160218954, + 0.0081319035962224, + 0.004151728004217148, + 0.0057118539698421955, + 0.008964949287474155, + 0.010206682607531548, + 0.011174232698976994, + 0.009701807983219624, + 0.007325951475650072, + 0.008500983938574791, + 0.011769145727157593, + 0.008347888477146626, + 0.011697348207235336, + 0.012910804711282253, + 0.012361518107354641, + 0.011145045049488544, + 0.013202734291553497, + 0.010303492657840252, + 0.009753627702593803, + 0.017762843519449234, + 0.028034964576363564, + 0.03259730339050293, + 0.1342780441045761, + 0.028816942125558853, + 0.021903550252318382, + 0.0338720828294754, + 0.12960650026798248, + 0.008537127636373043, + 0.010989812202751637, + 0.009091973304748535, + 0.012101818807423115, + 0.010213837958872318, + 0.010148636996746063, + 0.009898179210722446, + 0.011355099268257618, + 0.005037376657128334, + 0.008403575979173183, + 0.006011649966239929, + 0.0054499004036188126, + 0.006572461687028408, + 0.005970591213554144, + 0.006593267433345318, + 0.004621701780706644, + 0.003196155885234475, + 0.005220979452133179, + 0.007233830634504557, + 0.006465138401836157, + 0.005085410550236702, + 0.0029254076071083546, + 0.007245732005685568, + 0.004697625059634447, + 0.004500499460846186, + 0.0036896620877087116, + 0.002832577330991626, + 0.004882607609033585, + 0.007668679114431143, + 0.004572599660605192, + 0.004450318403542042, + 0.00643619429320097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.012561840936541557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547949329018593 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547949329018593, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547949329018593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08824687451124191, + 0.11282463371753693, + 0.02858385443687439, + 0.09574294835329056, + 0.12864229083061218, + 0.377437025308609, + 0.05717289820313454, + 0.044081006199121475, + 0.04442182928323746, + 0.022846680134534836 + ], + "credit_map": [ + 0.015082440339028835, + 0.006641085259616375, + 0.008828931488096714, + 0.006835229694843292, + 0.010575166903436184, + 0.019656579941511154, + 0.009129452519118786, + 0.011497985571622849, + 0.011419132351875305, + 0.014180858619511127, + 0.013350757770240307, + 0.019174061715602875, + 0.007913299836218357, + 0.006794729270040989, + 0.01565360277891159, + 0.024338185787200928, + 0.0017175906104966998, + 0.002065850654616952, + 0.003558833384886384, + 0.004613890312612057, + 0.004079621285200119, + 0.002811756916344166, + 0.005042729899287224, + 0.004693582188338041, + 0.006805851589888334, + 0.007461387664079666, + 0.011945538222789764, + 0.016271134838461876, + 0.010574817657470703, + 0.007130394224077463, + 0.022212358191609383, + 0.013341464102268219, + 0.00964999571442604, + 0.01286101620644331, + 0.015465829521417618, + 0.018586747348308563, + 0.013793577440083027, + 0.019397534430027008, + 0.017821012064814568, + 0.0210665725171566, + 0.014820987358689308, + 0.022889461368322372, + 0.02622205764055252, + 0.12147076427936554, + 0.023243781179189682, + 0.019727781414985657, + 0.02828366495668888, + 0.12077853083610535, + 0.005468974355608225, + 0.006903632078319788, + 0.006380483042448759, + 0.010836622677743435, + 0.0065773846581578255, + 0.006218861322849989, + 0.00810480397194624, + 0.006682137027382851, + 0.0052014729008078575, + 0.006446398328989744, + 0.006455699913203716, + 0.006083393003791571, + 0.0044244080781936646, + 0.0037943145725876093, + 0.005650478880852461, + 0.0060248407535254955, + 0.004081239458173513, + 0.00598558085039258, + 0.004386878572404385, + 0.004452006425708532, + 0.005789559800177813, + 0.0066869258880615234, + 0.006376140285283327, + 0.006663499400019646, + 0.0020668841898441315, + 0.0021761369425803423, + 0.0021113716065883636, + 0.0030583268962800503, + 0.005600814241915941, + 0.002506103366613388, + 0.0024582899641245604, + 0.0028687533922493458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.012547949329018593 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945378448814154 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.005945378448814154, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945378448814154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3516268730163574, + 0.15864534676074982, + 0.036162178963422775, + 0.04181278869509697, + 0.044364508241415024, + 0.24322497844696045, + 0.04270222410559654, + 0.02818157896399498, + 0.023373642936348915, + 0.029905913397669792 + ], + "credit_map": [ + 0.05515587329864502, + 0.039335984736680984, + 0.03653955087065697, + 0.02487189695239067, + 0.037473611533641815, + 0.09595368802547455, + 0.03003750741481781, + 0.0322587713599205, + 0.01272387895733118, + 0.015087236650288105, + 0.02356140688061714, + 0.019109975546598434, + 0.016777189448475838, + 0.008795482106506824, + 0.024663304910063744, + 0.03792685642838478, + 0.002643988933414221, + 0.0026646924670785666, + 0.005184361711144447, + 0.006445621605962515, + 0.003585749538615346, + 0.00216811359860003, + 0.009540383704006672, + 0.0039292676374316216, + 0.0031081493943929672, + 0.004608743824064732, + 0.004842571448534727, + 0.008344819769263268, + 0.004830026533454657, + 0.004333240445703268, + 0.005371998529881239, + 0.006373236887156963, + 0.0035446614492684603, + 0.0027148837689310312, + 0.00516137620434165, + 0.006722721736878157, + 0.0033020793925970793, + 0.0028026916552335024, + 0.011957772076129913, + 0.008158320561051369, + 0.010703684762120247, + 0.015471071004867554, + 0.01794113963842392, + 0.07595020532608032, + 0.014609762467443943, + 0.013997135683894157, + 0.017599478363990784, + 0.0769524946808815, + 0.004684880841523409, + 0.005289806984364986, + 0.004831892903894186, + 0.0067513869144022465, + 0.005850018933415413, + 0.005810797214508057, + 0.0038530149031430483, + 0.005630426108837128, + 0.003991671837866306, + 0.0038649688940495253, + 0.0038718217983841896, + 0.0035922119859606028, + 0.003359902650117874, + 0.0023271695245057344, + 0.0038983505219221115, + 0.0032754847779870033, + 0.0017610613722354174, + 0.0028689999599009752, + 0.002843386260792613, + 0.003265486331656575, + 0.0026649388018995523, + 0.0036091862712055445, + 0.0038334811106324196, + 0.002527103293687105, + 0.003314963774755597, + 0.0034999619238078594, + 0.002314154990017414, + 0.004030826035887003, + 0.004174795467406511, + 0.0037926228251308203, + 0.00435656076297164, + 0.004422028549015522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.005945378448814154 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a drone-like shot descending from a high vantage point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene stabilize in an extreme close-up focusing on a pair of slender white hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands smoothing out wrinkles on a soft off-white cotton cloth with watercolor animal prints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the cloth spread across a light-colored wooden table with distinct wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a ray of warm afternoon sunlight slant into the frame, casting a golden halo over the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549478642642498 + ], + "opposite_logprob_tokens": [ + -2.929947853088379 + ], + "logprob_reward": -0.0549478642642498, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549478642642498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12078075110912323, + 0.1616099327802658, + 0.0821550264954567, + 0.05018354207277298, + 0.04850338026881218, + 0.26872724294662476, + 0.11066761612892151, + 0.04254288226366043, + 0.05769465118646622, + 0.05713500827550888 + ], + "credit_map": [ + 0.010907584801316261, + 0.02157772146165371, + 0.015116286464035511, + 0.01165285985916853, + 0.01641695573925972, + 0.014259503223001957, + 0.014699187129735947, + 0.01615065149962902, + 0.015310082584619522, + 0.016745373606681824, + 0.02119811624288559, + 0.01869756355881691, + 0.013554134406149387, + 0.009464330971240997, + 0.021019000560045242, + 0.04562131315469742, + 0.006393550429493189, + 0.006402656901627779, + 0.008862435817718506, + 0.016543729230761528, + 0.014301100745797157, + 0.010941783897578716, + 0.010461365804076195, + 0.008248399943113327, + 0.004538544453680515, + 0.00560561940073967, + 0.005202600732445717, + 0.007769403979182243, + 0.006782239768654108, + 0.00548351276665926, + 0.0072761354967951775, + 0.0075254845432937145, + 0.00367486453615129, + 0.004698773846030235, + 0.007095617707818747, + 0.009614686481654644, + 0.004099366720765829, + 0.00444258376955986, + 0.008770091459155083, + 0.00610739178955555, + 0.011349115520715714, + 0.015695160254836082, + 0.018494822084903717, + 0.0838841125369072, + 0.019006919115781784, + 0.014539195224642754, + 0.021856803447008133, + 0.08390111476182938, + 0.013615147210657597, + 0.015326623804867268, + 0.013150142505764961, + 0.014278996735811234, + 0.011851994320750237, + 0.015587151981890202, + 0.008887739852070808, + 0.01796981506049633, + 0.004323185887187719, + 0.008327396586537361, + 0.006564938463270664, + 0.005856949836015701, + 0.004652667790651321, + 0.004219541326165199, + 0.0047414847649633884, + 0.003856711322441697, + 0.005377030465751886, + 0.006665247492492199, + 0.008541407994925976, + 0.00978423748165369, + 0.00772440480068326, + 0.004556501749902964, + 0.008163153193891048, + 0.006882660556584597, + 0.0042879036627709866, + 0.003689706325531006, + 0.00458126375451684, + 0.011625099927186966, + 0.008862879127264023, + 0.008008069358766079, + 0.007869075052440166, + 0.008211005479097366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.0549478642642498 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_tensor([9572, 9572])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_tensor([9572, 9572])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..11f6ef8cee71acd98e424a813bb81918cc8b1426 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_tensor([9572, 9572])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, static shot, shot from a slightly upward angle, exudes the quality of highly detailed, color-saturated digital art. A middle-aged farmer, tanned and weathered, sits on the edge of a golden rice paddy. He wears a wide-brimmed straw hat, handwoven from dry straw, its brim frayed, its brim casting a shadow that partially obscures his deeply wrinkled face. He wears a simple, dark blue, coarse-sleeved shirt, sleeves casually rolled to the elbows, revealing arms sturdy from years of hard work. The twilight glow of the setting sun casts a soft, warm golden glow over the world. He holds an earthenware bowl steady in both hands, then decisively tilts his head back, his neck forming a powerful arch as he empties the clear liquor in a single, heroic gulp. His eyes are closed, his Adam's apple clearly rising and falling beneath his bronzed skin as he swallows. A few drops of wine, unswallowed, slid down the corners of his rough lips, gleaming in the setting sun. The distant ridges and sky were dyed a warm orange-red hue, and the entire scene was filled with a sense of tranquility and satisfaction after hard work, full of simple yet moving power.", + "seed": 20772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_9572_seed_20772_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.039632092136889695, + "vlm_logprob": -0.039632092136889695, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.039632092136889695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07070858776569366 + ], + "opposite_logprob_tokens": [ + -2.9457085132598877 + ], + "logprob_reward": -0.07070858776569366, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07070858776569366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0989893302321434, + 0.03826155140995979, + 0.07907088845968246, + 0.17081256210803986, + 0.03982992842793465, + 0.15501715242862701, + 0.08814151585102081, + 0.16011326014995575, + 0.11630513519048691, + 0.05345869064331055 + ], + "credit_map": [ + 0.013328970409929752, + 0.0190575048327446, + 0.010176004841923714, + 0.01077699288725853, + 0.014681641943752766, + 0.012942499481141567, + 0.007355317939072847, + 0.010670389980077744, + 0.00394994905218482, + 0.003711741417646408, + 0.006725025363266468, + 0.007068737875670195, + 0.0052511016838252544, + 0.003594143083319068, + 0.003921173978596926, + 0.0040396819822490215, + 0.009741431102156639, + 0.008836055174469948, + 0.028120215982198715, + 0.0070179239846765995, + 0.007602394092828035, + 0.005712568759918213, + 0.007656682748347521, + 0.004383617080748081, + 0.024192076176404953, + 0.011922898702323437, + 0.06758461147546768, + 0.018602192401885986, + 0.013483100570738316, + 0.009549158625304699, + 0.013054157607257366, + 0.012424353510141373, + 0.00431775813922286, + 0.0033268239349126816, + 0.0070853158831596375, + 0.004850087221711874, + 0.006132925394922495, + 0.004418440628796816, + 0.004161808639764786, + 0.005536768585443497, + 0.01897462084889412, + 0.016862472519278526, + 0.012882527895271778, + 0.015096132643520832, + 0.03449852392077446, + 0.02489248476922512, + 0.014237048104405403, + 0.01757335290312767, + 0.013826530426740646, + 0.008795849047601223, + 0.024102680385112762, + 0.007179046515375376, + 0.008188257925212383, + 0.0099295973777771, + 0.007618415635079145, + 0.008501142263412476, + 0.03142903745174408, + 0.04580188915133476, + 0.008672649040818214, + 0.014946257695555687, + 0.02941126748919487, + 0.012071446515619755, + 0.008850835263729095, + 0.008929876610636711, + 0.017427535727620125, + 0.01861697807908058, + 0.008341019973158836, + 0.008844015188515186, + 0.021837009117007256, + 0.02379857562482357, + 0.008576128631830215, + 0.008863866329193115, + 0.009222744032740593, + 0.008294967003166676, + 0.00999589916318655, + 0.005789317190647125, + 0.00781944114714861, + 0.005179962608963251, + 0.0045197391882538795, + 0.0026366221718490124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.07070858776569366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596904680132866 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020596904680132866, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596904680132866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05036311224102974, + 0.04289796203374863, + 0.05479203537106514, + 0.21690036356449127, + 0.024512508884072304, + 0.04810251295566559, + 0.14842365682125092, + 0.10877129435539246, + 0.20796038210391998, + 0.09727621078491211 + ], + "credit_map": [ + 0.005591538269072771, + 0.006744474172592163, + 0.006037986371666193, + 0.005362817086279392, + 0.009706830605864525, + 0.0052512697875499725, + 0.004833364859223366, + 0.006834831554442644, + 0.004378241021186113, + 0.005032319109886885, + 0.010046658106148243, + 0.004470186773687601, + 0.00508438004180789, + 0.0036914849188178778, + 0.004598323721438646, + 0.005596368107944727, + 0.005621991585940123, + 0.006090008188039064, + 0.016507867723703384, + 0.005983421579003334, + 0.005712100770324469, + 0.004747883882373571, + 0.006526668090373278, + 0.0036020944826304913, + 0.0299078281968832, + 0.014618434943258762, + 0.08996356278657913, + 0.023368190973997116, + 0.014072784222662449, + 0.016904287040233612, + 0.015907568857073784, + 0.012157709337770939, + 0.0023438900243490934, + 0.0025204019621014595, + 0.0030589455273002386, + 0.0027651924174278975, + 0.004516698885709047, + 0.0020978001412004232, + 0.0029005510732531548, + 0.004309030249714851, + 0.006035826168954372, + 0.008175622671842575, + 0.008088507689535618, + 0.008646810427308083, + 0.005247088614851236, + 0.004841333720833063, + 0.003464915556833148, + 0.003602407407015562, + 0.018563691526651382, + 0.013762008398771286, + 0.020523112267255783, + 0.008494231849908829, + 0.02521136775612831, + 0.03171323239803314, + 0.011898708529770374, + 0.018257299438118935, + 0.019940949976444244, + 0.028568638488650322, + 0.008904754184186459, + 0.010669280774891376, + 0.016298865899443626, + 0.01115885004401207, + 0.005564276594668627, + 0.00766567699611187, + 0.03165513649582863, + 0.028056129813194275, + 0.014704538509249687, + 0.01554829627275467, + 0.035560641437768936, + 0.04421691969037056, + 0.016542738303542137, + 0.021675968542695045, + 0.013389203697443008, + 0.010165289975702763, + 0.01261858455836773, + 0.01398409716784954, + 0.01863274537026882, + 0.011816122569143772, + 0.008674153126776218, + 0.007996013388037682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.020596904680132866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067339395172894 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067339395172894, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067339395172894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30381885170936584, + 0.060276005417108536, + 0.08655112236738205, + 0.05209222808480263, + 0.036518484354019165, + 0.1245650127530098, + 0.05110117048025131, + 0.07086372375488281, + 0.045507729053497314, + 0.16870565712451935 + ], + "credit_map": [ + 0.04551050066947937, + 0.06803536415100098, + 0.02897959016263485, + 0.02262249030172825, + 0.0414717011153698, + 0.057645589113235474, + 0.018542161211371422, + 0.021011456847190857, + 0.008325362578034401, + 0.007629197556525469, + 0.012046546675264835, + 0.00764216436073184, + 0.005830087698996067, + 0.005637010559439659, + 0.007472401484847069, + 0.005693239625543356, + 0.01202250737696886, + 0.01965942606329918, + 0.015423235483467579, + 0.008168987929821014, + 0.00784141756594181, + 0.009987149387598038, + 0.008452464826405048, + 0.004995931871235371, + 0.007407582364976406, + 0.0034024715423583984, + 0.02089555934071541, + 0.0054964423179626465, + 0.0034225373528897762, + 0.004061455838382244, + 0.003850968088954687, + 0.003555216593667865, + 0.004636481404304504, + 0.005392300896346569, + 0.007819656282663345, + 0.0068934266455471516, + 0.0038043418899178505, + 0.0022627844009548426, + 0.0031030194368213415, + 0.00260647339746356, + 0.01887362450361252, + 0.027135608717799187, + 0.020749788731336594, + 0.022828470915555954, + 0.010913463309407234, + 0.00771298399195075, + 0.008402884937822819, + 0.007948193699121475, + 0.009341194294393063, + 0.004150077700614929, + 0.01573793590068817, + 0.005515210330486298, + 0.006163932383060455, + 0.0035063661634922028, + 0.002774383407086134, + 0.003912072163075209, + 0.01306877564638853, + 0.018085483461618423, + 0.007492548320442438, + 0.006501094438135624, + 0.010595045052468777, + 0.007075525354593992, + 0.004525971598923206, + 0.003519276389852166, + 0.0061819665133953094, + 0.005869221407920122, + 0.005301863886415958, + 0.0032442903611809015, + 0.00718320719897747, + 0.006533065810799599, + 0.004710423294454813, + 0.006483692675828934, + 0.04054417461156845, + 0.04213655740022659, + 0.02490329183638096, + 0.0149366594851017, + 0.01907845214009285, + 0.013065768405795097, + 0.007474448066204786, + 0.006566295865923166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0067339395172894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808697313070297 + ], + "opposite_logprob_tokens": [ + -3.2880868911743164 + ], + "logprob_reward": -0.03808697313070297, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808697313070297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039913758635520935, + 0.019342517480254173, + 0.04750524088740349, + 0.42254796624183655, + 0.018080050125718117, + 0.05306451767683029, + 0.10765516012907028, + 0.1681327074766159, + 0.06194181367754936, + 0.061816275119781494 + ], + "credit_map": [ + 0.006921054795384407, + 0.003437420353293419, + 0.007414082996547222, + 0.00556059880182147, + 0.005047444254159927, + 0.004113292787224054, + 0.003823645878583193, + 0.003596216905862093, + 0.0024699463974684477, + 0.003125385846942663, + 0.0036541810259222984, + 0.0024142172187566757, + 0.0023299362510442734, + 0.0019252040656283498, + 0.0018420510459691286, + 0.0015815941151231527, + 0.006080093327909708, + 0.010291388258337975, + 0.008711445145308971, + 0.004109751433134079, + 0.004836864769458771, + 0.005571337416768074, + 0.0051848143339157104, + 0.0027195459697395563, + 0.06165110319852829, + 0.02822367660701275, + 0.18031412363052368, + 0.04698548838496208, + 0.02114293910562992, + 0.029588110744953156, + 0.03011210821568966, + 0.024530408903956413, + 0.0019935944583266973, + 0.001387617434374988, + 0.005032645538449287, + 0.0031336701940745115, + 0.002035521436482668, + 0.0010000901529565454, + 0.0017241220921278, + 0.0017727900994941592, + 0.007069847546517849, + 0.008978291414678097, + 0.005962077062577009, + 0.007808151189237833, + 0.0085002351552248, + 0.005596371367573738, + 0.0038056927733123302, + 0.0053438530303537846, + 0.017226818948984146, + 0.00794045627117157, + 0.031828124076128006, + 0.010214649140834808, + 0.01570267789065838, + 0.008688158355653286, + 0.007225536275655031, + 0.008828742429614067, + 0.03389478102326393, + 0.04927404597401619, + 0.010974684730172157, + 0.015753310173749924, + 0.026726296171545982, + 0.014308440499007702, + 0.010702109895646572, + 0.006499057170003653, + 0.010528665967285633, + 0.012289636768400669, + 0.009182174690067768, + 0.004635483957827091, + 0.009588380344212055, + 0.007639189716428518, + 0.0037108405958861113, + 0.004367445129901171, + 0.009397565387189388, + 0.009762408211827278, + 0.005798221565783024, + 0.010312081314623356, + 0.011288387700915337, + 0.005303735379129648, + 0.005995876155793667, + 0.003957999404519796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.03808697313070297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203405559062958 + ], + "opposite_logprob_tokens": [ + -2.8120341300964355 + ], + "logprob_reward": -0.06203405559062958, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203405559062958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08687369525432587, + 0.025866441428661346, + 0.03346762806177139, + 0.19407343864440918, + 0.042448077350854874, + 0.07584651559591293, + 0.1830522119998932, + 0.15144094824790955, + 0.08681407570838928, + 0.12011692672967911 + ], + "credit_map": [ + 0.014729883521795273, + 0.012755882926285267, + 0.013596448116004467, + 0.009250249713659286, + 0.010231764987111092, + 0.012889675796031952, + 0.006061282474547625, + 0.007358509115874767, + 0.003984520677477121, + 0.0027000822592526674, + 0.0044249277561903, + 0.0027799513190984726, + 0.004225390497595072, + 0.0027361237443983555, + 0.0027844831347465515, + 0.0022309613414108753, + 0.004511058796197176, + 0.0034869147930294275, + 0.00576619291678071, + 0.0040966058149933815, + 0.003297886112704873, + 0.004819598514586687, + 0.004903415683656931, + 0.0025859533343464136, + 0.02791518345475197, + 0.014817642979323864, + 0.08117655664682388, + 0.02134479209780693, + 0.010268696583807468, + 0.01397491991519928, + 0.013782045803964138, + 0.010793589986860752, + 0.004836926236748695, + 0.003005080157890916, + 0.013194265775382519, + 0.005992253310978413, + 0.004524651914834976, + 0.003087491961196065, + 0.0040083532221615314, + 0.003799055004492402, + 0.00997897144407034, + 0.011831887066364288, + 0.009841683320701122, + 0.010113801807165146, + 0.012148482725024223, + 0.008918200619518757, + 0.006870398763567209, + 0.006143094971776009, + 0.026645096018910408, + 0.011336026713252068, + 0.047237806022167206, + 0.013310273177921772, + 0.027636202052235603, + 0.031733568757772446, + 0.009926678612828255, + 0.015226559713482857, + 0.02525446005165577, + 0.03234723210334778, + 0.0175466425716877, + 0.013343743048608303, + 0.023485923185944557, + 0.022537022829055786, + 0.009809279814362526, + 0.007116635330021381, + 0.012739378027617931, + 0.01312354113906622, + 0.00783522892743349, + 0.0070511954836547375, + 0.014970913529396057, + 0.013318861834704876, + 0.007185937836766243, + 0.010589023120701313, + 0.017501140013337135, + 0.017377492040395737, + 0.014323403127491474, + 0.01181046012789011, + 0.02115957997739315, + 0.018194736912846565, + 0.010379809886217117, + 0.009370296262204647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.06203405559062958 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, exudes the quality of highly detailed, color-saturated digital art. A middle-aged farmer, tanned and weathered, sits on the edge of a golden rice paddy. He wears a wide-brimmed straw hat, handwoven from dry straw, its brim frayed, its brim casting a shadow that partially obscures his deeply wrinkled face. He wears a simple, dark blue, coarse-sleeved shirt, sleeves casually rolled to the elbows, revealing arms sturdy from years of hard work. The twilight glow of the setting sun casts a soft, warm golden glow over the world. He holds an earthenware bowl steady in both hands, then decisively tilts his head back, his neck forming a powerful arch as he empties the clear liquor in a single, heroic gulp. His eyes are closed, his Adam's apple clearly rising and falling beneath his bronzed skin as he swallows. A few drops of wine, unswallowed, slid down the corners of his rough lips, gleaming in the setting sun. The distant ridges and sky were dyed a warm orange-red hue, and the entire scene was filled with a sense of tranquility and satisfaction after hard work, full of simple yet moving power.", + "seed": 20773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_9572_seed_20773_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03469172865152359, + "vlm_logprob": -0.03469172865152359, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03469172865152359, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06741200387477875 + ], + "opposite_logprob_tokens": [ + -3.3174118995666504 + ], + "logprob_reward": -0.06741200387477875, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06741200387477875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1296292543411255, + 0.06802057474851608, + 0.09333758056163788, + 0.17184428870677948, + 0.07787390053272247, + 0.0679226666688919, + 0.06746993958950043, + 0.07067123055458069, + 0.042381469160318375, + 0.2108490914106369 + ], + "credit_map": [ + 0.024979103356599808, + 0.020786913111805916, + 0.024953048676252365, + 0.017044253647327423, + 0.014841686934232712, + 0.0074028391391038895, + 0.009507023729383945, + 0.01011438574641943, + 0.009013223461806774, + 0.007465126924216747, + 0.012159621343016624, + 0.010699458420276642, + 0.010040807537734509, + 0.005817020311951637, + 0.007438885048031807, + 0.005386437755078077, + 0.009484254755079746, + 0.01182590238749981, + 0.010705683380365372, + 0.014234723523259163, + 0.01635552942752838, + 0.009769314900040627, + 0.012952485121786594, + 0.00800967775285244, + 0.015921443700790405, + 0.023852920159697533, + 0.024940725415945053, + 0.04417669028043747, + 0.02327345684170723, + 0.017108285799622536, + 0.013195500709116459, + 0.009375275112688541, + 0.009925022721290588, + 0.006147578824311495, + 0.016395220533013344, + 0.01996982842683792, + 0.007451463956385851, + 0.005397140979766846, + 0.00682421401143074, + 0.005763431545346975, + 0.012003038078546524, + 0.008953174576163292, + 0.013908268883824348, + 0.008812366984784603, + 0.006863614544272423, + 0.008205853402614594, + 0.004588743206113577, + 0.004587606992572546, + 0.011483974754810333, + 0.0063658347353339195, + 0.019893227145075798, + 0.005236471071839333, + 0.008019926026463509, + 0.008424594067037106, + 0.0040319389663636684, + 0.004013971425592899, + 0.015532483346760273, + 0.006458873860538006, + 0.013221682980656624, + 0.008104371838271618, + 0.00869962852448225, + 0.007216794416308403, + 0.004502858966588974, + 0.006934531964361668, + 0.00681831082329154, + 0.003818957833573222, + 0.006092846859246492, + 0.006339774001389742, + 0.004544470459222794, + 0.005538875702768564, + 0.003506413893774152, + 0.005721820518374443, + 0.02473635971546173, + 0.01451552752405405, + 0.03873535990715027, + 0.021116971969604492, + 0.03210180997848511, + 0.023585611954331398, + 0.028308363631367683, + 0.027749085798859596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.06741200387477875 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086550429463387 + ], + "opposite_logprob_tokens": [ + -4.141086578369141 + ], + "logprob_reward": -0.016086550429463387, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086550429463387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24849581718444824, + 0.052375588566064835, + 0.13159789144992828, + 0.08028208464384079, + 0.05286179855465889, + 0.04484302178025246, + 0.24317248165607452, + 0.027232397347688675, + 0.05271455645561218, + 0.06642437726259232 + ], + "credit_map": [ + 0.044762659817934036, + 0.023785661906003952, + 0.05122318118810654, + 0.02945493906736374, + 0.030221296474337578, + 0.01709962636232376, + 0.02711271308362484, + 0.024835726246237755, + 0.004663689993321896, + 0.007685645017772913, + 0.006449797190725803, + 0.009323998354375362, + 0.009047579020261765, + 0.004008098039776087, + 0.006858798209577799, + 0.0043379818089306355, + 0.011543301865458488, + 0.01636715605854988, + 0.0230718944221735, + 0.042102932929992676, + 0.01225795317441225, + 0.007741961162537336, + 0.010169577784836292, + 0.008343112654983997, + 0.009308768436312675, + 0.014382915571331978, + 0.009766491129994392, + 0.01178676262497902, + 0.01256169006228447, + 0.007968162186443806, + 0.00753450533375144, + 0.006972786970436573, + 0.0064181918278336525, + 0.005800657439976931, + 0.008589633740484715, + 0.01168090384453535, + 0.005564277991652489, + 0.004077606834471226, + 0.0059610274620354176, + 0.004769501276314259, + 0.004009278491139412, + 0.005401464644819498, + 0.005039540119469166, + 0.006114323623478413, + 0.004659941885620356, + 0.00811629369854927, + 0.00606162752956152, + 0.005440553650259972, + 0.020973101258277893, + 0.010964312590658665, + 0.07632967084646225, + 0.013371145352721214, + 0.01862296648323536, + 0.07339179515838623, + 0.01319201197475195, + 0.01632748171687126, + 0.004231685306876898, + 0.002987699816003442, + 0.0034793694503605366, + 0.0034430234227329493, + 0.0035809155087918043, + 0.004938500467687845, + 0.0017983806319534779, + 0.0027728225104510784, + 0.008085785433650017, + 0.003966607619076967, + 0.008108231239020824, + 0.0069786980748176575, + 0.008882303722202778, + 0.006178986746817827, + 0.003506869077682495, + 0.007007075939327478, + 0.009385142475366592, + 0.005482567939907312, + 0.006333653349429369, + 0.007841914892196655, + 0.008451431058347225, + 0.011554685421288013, + 0.006929114926606417, + 0.010445864871144295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.016086550429463387 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13364586234092712, + 0.2569555342197418, + 0.09807547926902771, + 0.09401831775903702, + 0.06740080565214157, + 0.04660804569721222, + 0.015727030113339424, + 0.029305417090654373, + 0.1734103411436081, + 0.08485310524702072 + ], + "credit_map": [ + 0.022023329511284828, + 0.014556236565113068, + 0.028259724378585815, + 0.02034306712448597, + 0.01272524707019329, + 0.009728701785206795, + 0.013990695588290691, + 0.012018871493637562, + 0.025377212092280388, + 0.02649187110364437, + 0.04513729363679886, + 0.07948653399944305, + 0.018214324489235878, + 0.01358780823647976, + 0.030782796442508698, + 0.01787770725786686, + 0.01117895357310772, + 0.014019347727298737, + 0.016840096563100815, + 0.028189675882458687, + 0.006487912032753229, + 0.006938135251402855, + 0.006671822164207697, + 0.007749550044536591, + 0.010507110506296158, + 0.009315135888755322, + 0.017823578789830208, + 0.031904734671115875, + 0.007802248932421207, + 0.005279576871544123, + 0.00637269951403141, + 0.005013245157897472, + 0.01239356305450201, + 0.012851471081376076, + 0.011216551996767521, + 0.011268198490142822, + 0.007520333863794804, + 0.003867986612021923, + 0.005129784345626831, + 0.003152925753965974, + 0.009347484447062016, + 0.005606670398265123, + 0.010217281058430672, + 0.007637656759470701, + 0.0032101611141115427, + 0.003973343875259161, + 0.0035165289882570505, + 0.00309892394579947, + 0.0020415482576936483, + 0.002322291722521186, + 0.0027502630837261677, + 0.0016129123978316784, + 0.002095836214721203, + 0.0018258101772516966, + 0.001217104378156364, + 0.0018612651620060205, + 0.005240581929683685, + 0.002648033667355776, + 0.006032904144376516, + 0.0043302676640450954, + 0.003371721599251032, + 0.0025216038338840008, + 0.0024258119519799948, + 0.0027344943955540657, + 0.04583189636468887, + 0.01889977790415287, + 0.049459781497716904, + 0.011628244072198868, + 0.015277395024895668, + 0.014794299378991127, + 0.007830132730305195, + 0.00968881044536829, + 0.011726033873856068, + 0.005501810926944017, + 0.02180270105600357, + 0.00975035410374403, + 0.011067135259509087, + 0.00890720821917057, + 0.009628509171307087, + 0.0064693633466959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.004637913778424263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328341454267502 + ], + "opposite_logprob_tokens": [ + -3.7732834815979004 + ], + "logprob_reward": -0.02328341454267502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328341454267502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17112873494625092, + 0.03766992315649986, + 0.1401701271533966, + 0.08775405585765839, + 0.029857946559786797, + 0.16340991854667664, + 0.1390136182308197, + 0.05328742787241936, + 0.09825194627046585, + 0.07945631444454193 + ], + "credit_map": [ + 0.03337570279836655, + 0.016664668917655945, + 0.03929602727293968, + 0.01950324885547161, + 0.02046678401529789, + 0.010640652850270271, + 0.01673836261034012, + 0.0144432932138443, + 0.004935009870678186, + 0.0030091118533164263, + 0.006911499425768852, + 0.005937926936894655, + 0.005368655547499657, + 0.0029478592332452536, + 0.0051642213948071, + 0.0033956365659832954, + 0.01069584209471941, + 0.01637435518205166, + 0.01656273379921913, + 0.017848433926701546, + 0.027644412592053413, + 0.019843662157654762, + 0.019346440210938454, + 0.011854253709316254, + 0.009492444805800915, + 0.009417164139449596, + 0.0148922773078084, + 0.028137093409895897, + 0.00807464774698019, + 0.004527166485786438, + 0.008336972445249557, + 0.004876290913671255, + 0.004050102550536394, + 0.004127274267375469, + 0.004746844992041588, + 0.0048563131131231785, + 0.003688535653054714, + 0.0027624324429780245, + 0.003243303159251809, + 0.0023831387516111135, + 0.024776902049779892, + 0.02378043532371521, + 0.021971814334392548, + 0.029347462579607964, + 0.014899521134793758, + 0.02346857637166977, + 0.01346592791378498, + 0.011699281632900238, + 0.01161726750433445, + 0.007067161612212658, + 0.04156670346856117, + 0.011185628361999989, + 0.01404294278472662, + 0.03976159542798996, + 0.00505300797522068, + 0.008719310164451599, + 0.0071464418433606625, + 0.004213074222207069, + 0.007858360186219215, + 0.007808206137269735, + 0.008599119260907173, + 0.0077447243966162205, + 0.004096330609172583, + 0.005821173544973135, + 0.014115299098193645, + 0.008426559157669544, + 0.012871736660599709, + 0.009200595319271088, + 0.01647762395441532, + 0.011692626401782036, + 0.006709724199026823, + 0.018757784739136696, + 0.005970381665974855, + 0.005218342412263155, + 0.011897912248969078, + 0.007233149837702513, + 0.011209554970264435, + 0.012995379976928234, + 0.010113470256328583, + 0.014818127267062664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.02328341454267502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062038760632276535 + ], + "opposite_logprob_tokens": [ + -2.8120386600494385 + ], + "logprob_reward": -0.062038760632276535, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062038760632276535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14960570633411407, + 0.0649028941988945, + 0.07884082198143005, + 0.03616725280880928, + 0.06820688396692276, + 0.14835456013679504, + 0.2894991338253021, + 0.028237581253051758, + 0.05589437112212181, + 0.0802907943725586 + ], + "credit_map": [ + 0.02674974501132965, + 0.014104312285780907, + 0.035349294543266296, + 0.02340688183903694, + 0.013539551757276058, + 0.010226787067949772, + 0.013114414177834988, + 0.01311472523957491, + 0.008048604242503643, + 0.008231100626289845, + 0.011333931237459183, + 0.013850324787199497, + 0.0090226074680686, + 0.005714458879083395, + 0.005450657103210688, + 0.003251214511692524, + 0.006982683669775724, + 0.01091291569173336, + 0.009279281832277775, + 0.008987115696072578, + 0.01503113005310297, + 0.013046134263277054, + 0.008900872431695461, + 0.005700686480849981, + 0.003588037798181176, + 0.005235167220234871, + 0.0054085832089185715, + 0.009926785714924335, + 0.0037647990975528955, + 0.0038029300048947334, + 0.0021384391002357006, + 0.002302507869899273, + 0.009229626506567001, + 0.013760116882622242, + 0.009322954341769218, + 0.012575269676744938, + 0.006929539609700441, + 0.00585655402392149, + 0.007205002475529909, + 0.0033278167247772217, + 0.01996203325688839, + 0.015009861439466476, + 0.02897103875875473, + 0.03271332383155823, + 0.013673710636794567, + 0.014368331991136074, + 0.013932222500443459, + 0.00972403772175312, + 0.026904352009296417, + 0.013090264052152634, + 0.10012584924697876, + 0.01651286892592907, + 0.024911198765039444, + 0.07886479794979095, + 0.011957916431128979, + 0.0171318668872118, + 0.005436017643660307, + 0.0024780987296253443, + 0.006078397389501333, + 0.004391770344227552, + 0.002754634479060769, + 0.002338600577786565, + 0.0018708454445004463, + 0.002889217110350728, + 0.008149855770170689, + 0.0043637449853122234, + 0.00814555212855339, + 0.009322903119027615, + 0.007297107949852943, + 0.0070720394141972065, + 0.004272807389497757, + 0.007270361762493849, + 0.00715248379856348, + 0.007329975254833698, + 0.009409516118466854, + 0.008694877848029137, + 0.009686356410384178, + 0.015457655303180218, + 0.008715826086699963, + 0.013844102621078491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.062038760632276535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, exudes the quality of highly detailed, color-saturated digital art. A middle-aged farmer, tanned and weathered, sits on the edge of a golden rice paddy. He wears a wide-brimmed straw hat, handwoven from dry straw, its brim frayed, its brim casting a shadow that partially obscures his deeply wrinkled face. He wears a simple, dark blue, coarse-sleeved shirt, sleeves casually rolled to the elbows, revealing arms sturdy from years of hard work. The twilight glow of the setting sun casts a soft, warm golden glow over the world. He holds an earthenware bowl steady in both hands, then decisively tilts his head back, his neck forming a powerful arch as he empties the clear liquor in a single, heroic gulp. His eyes are closed, his Adam's apple clearly rising and falling beneath his bronzed skin as he swallows. A few drops of wine, unswallowed, slid down the corners of his rough lips, gleaming in the setting sun. The distant ridges and sky were dyed a warm orange-red hue, and the entire scene was filled with a sense of tranquility and satisfaction after hard work, full of simple yet moving power.", + "seed": 20774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_9572_seed_20774_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.034127644076943395, + "vlm_logprob": -0.034127644076943395, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.034127644076943395, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08078556507825851 + ], + "opposite_logprob_tokens": [ + -2.9557855129241943 + ], + "logprob_reward": -0.08078556507825851, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08078556507825851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06835851818323135, + 0.06694646924734116, + 0.08188169449567795, + 0.10767951607704163, + 0.07345952093601227, + 0.04605324938893318, + 0.06388846039772034, + 0.2089894711971283, + 0.1935557872056961, + 0.0891873687505722 + ], + "credit_map": [ + 0.0072190663777291775, + 0.006540063768625259, + 0.008828292600810528, + 0.013346511870622635, + 0.010256354697048664, + 0.010740666650235653, + 0.005549340508878231, + 0.005878217052668333, + 0.006920053157955408, + 0.005598573479801416, + 0.009274784475564957, + 0.011552289128303528, + 0.007972667925059795, + 0.010856508277356625, + 0.008338731713593006, + 0.0064328499138355255, + 0.010940606705844402, + 0.006164162885397673, + 0.009746680036187172, + 0.005365748889744282, + 0.014601037837564945, + 0.018820011988282204, + 0.008150836452841759, + 0.008092604577541351, + 0.010549193248152733, + 0.013507932424545288, + 0.01625681109726429, + 0.017154665663838387, + 0.01184567529708147, + 0.01451337430626154, + 0.014085122384130955, + 0.009766737930476665, + 0.007398029323667288, + 0.007620751392096281, + 0.010770386084914207, + 0.014762243255972862, + 0.009163238108158112, + 0.00861667562276125, + 0.008677290752530098, + 0.006450902670621872, + 0.005370521917939186, + 0.003843753132969141, + 0.007610403001308441, + 0.009226511232554913, + 0.00539806904271245, + 0.00565359415486455, + 0.0044458406046032906, + 0.004504550714045763, + 0.008675803430378437, + 0.00767216132953763, + 0.00786430761218071, + 0.009908832609653473, + 0.005867494270205498, + 0.007987819612026215, + 0.009332502260804176, + 0.006579533684998751, + 0.024778125807642937, + 0.05044163018465042, + 0.02723495475947857, + 0.024089820683002472, + 0.022650914266705513, + 0.021929938346147537, + 0.023664161562919617, + 0.014199912548065186, + 0.025374891236424446, + 0.02144445665180683, + 0.020090626552700996, + 0.027873970568180084, + 0.022325104102492332, + 0.03276132047176361, + 0.026201721280813217, + 0.01748369075357914, + 0.009687121026217937, + 0.00639406219124794, + 0.012171067297458649, + 0.012707470916211605, + 0.011098765768110752, + 0.017107615247368813, + 0.012212967500090599, + 0.007808295078575611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.08078556507825851 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084320843219757 + ], + "opposite_logprob_tokens": [ + -4.14108419418335 + ], + "logprob_reward": -0.016084320843219757, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084320843219757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19983625411987305, + 0.05700160562992096, + 0.053114887326955795, + 0.11401128023862839, + 0.07694076746702194, + 0.024675944820046425, + 0.09174738824367523, + 0.19924551248550415, + 0.026890533044934273, + 0.15653583407402039 + ], + "credit_map": [ + 0.02113351784646511, + 0.017659153789281845, + 0.02582181990146637, + 0.024119967594742775, + 0.029584120959043503, + 0.03759421035647392, + 0.02022930607199669, + 0.023694152012467384, + 0.0052268411964178085, + 0.007281107362359762, + 0.006864865776151419, + 0.009372142143547535, + 0.009634428657591343, + 0.007688750978559256, + 0.005956328473985195, + 0.004977143835276365, + 0.005285169929265976, + 0.004700393881648779, + 0.005730641074478626, + 0.005703622940927744, + 0.01161868218332529, + 0.009651724249124527, + 0.006251384969800711, + 0.004173270892351866, + 0.008076936937868595, + 0.011685533449053764, + 0.01388107892125845, + 0.014656545594334602, + 0.015523089095950127, + 0.023719800636172295, + 0.013476759195327759, + 0.012991541065275669, + 0.008834382519125938, + 0.004986618645489216, + 0.009452174417674541, + 0.014868236146867275, + 0.009560051374137402, + 0.012793957255780697, + 0.007685360498726368, + 0.008759989403188229, + 0.0026564293075352907, + 0.002349858870729804, + 0.002881651045754552, + 0.002384884748607874, + 0.005737919360399246, + 0.0037042759358882904, + 0.0028285894077271223, + 0.0021323366090655327, + 0.011123177595436573, + 0.00777865294367075, + 0.012984084896743298, + 0.02059820666909218, + 0.008768972009420395, + 0.0076805707067251205, + 0.015809733420610428, + 0.007003986742347479, + 0.032190192490816116, + 0.08319476991891861, + 0.04396577924489975, + 0.008456016890704632, + 0.012917282059788704, + 0.009455583989620209, + 0.0047172848135232925, + 0.004348608199506998, + 0.002613262040540576, + 0.0032346590887755156, + 0.0027292356826364994, + 0.003291604807600379, + 0.0055320351384580135, + 0.0038284570910036564, + 0.0031700702384114265, + 0.0024912094231694937, + 0.015299166552722454, + 0.014287357218563557, + 0.023634035140275955, + 0.01812790520489216, + 0.023150334134697914, + 0.03492693603038788, + 0.014059317298233509, + 0.01305078249424696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.016084320843219757 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635303281247616 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635303281247616, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635303281247616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09965194016695023, + 0.032929565757513046, + 0.04758897423744202, + 0.025197133421897888, + 0.06062313914299011, + 0.018016038462519646, + 0.16702832281589508, + 0.34685754776000977, + 0.032432228326797485, + 0.1696750819683075 + ], + "credit_map": [ + 0.013463334180414677, + 0.012364876456558704, + 0.014638619497418404, + 0.023007746785879135, + 0.011358357034623623, + 0.011166112497448921, + 0.007629028521478176, + 0.006023871712386608, + 0.0053627099841833115, + 0.003984227776527405, + 0.004093246068805456, + 0.00537892384454608, + 0.003893557470291853, + 0.00479061808437109, + 0.002511771861463785, + 0.002914510667324066, + 0.00621076812967658, + 0.005580487195402384, + 0.00505498331040144, + 0.007556527387350798, + 0.007176984567195177, + 0.007717996370047331, + 0.003949891310185194, + 0.004341338761150837, + 0.0038589537143707275, + 0.0024130609817802906, + 0.003913309890776873, + 0.005531538277864456, + 0.0017941729165613651, + 0.003469891147688031, + 0.0022544260136783123, + 0.001961782807484269, + 0.007654727902263403, + 0.006956256926059723, + 0.010151639580726624, + 0.011336623691022396, + 0.00575076462700963, + 0.006961639504879713, + 0.00715147890150547, + 0.004660011734813452, + 0.003151069628074765, + 0.002781130373477936, + 0.002384636551141739, + 0.00378727400675416, + 0.001120440079830587, + 0.0018528883811086416, + 0.0014856169000267982, + 0.0014529845211654902, + 0.025883149355649948, + 0.024857332929968834, + 0.024442389607429504, + 0.027124831452965736, + 0.013689082115888596, + 0.017813922837376595, + 0.020869538187980652, + 0.012348092161118984, + 0.054675597697496414, + 0.13899272680282593, + 0.07548069953918457, + 0.015540409833192825, + 0.022022942081093788, + 0.020509565249085426, + 0.008102966472506523, + 0.01153265405446291, + 0.0035100991372019053, + 0.0033767621498554945, + 0.0037091951817274094, + 0.0041793640702962875, + 0.004444800317287445, + 0.006232965271919966, + 0.0035549199674278498, + 0.00342412362806499, + 0.029379939660429955, + 0.024836834520101547, + 0.031376007944345474, + 0.022169090807437897, + 0.017580943182110786, + 0.016950411722064018, + 0.016795117408037186, + 0.010586736723780632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.004635303281247616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419012900441885 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.01419012900441885, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419012900441885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14053596556186676, + 0.03517116978764534, + 0.07914425432682037, + 0.0641818419098854, + 0.09284462034702301, + 0.24368691444396973, + 0.08243539184331894, + 0.06923895329236984, + 0.08900728076696396, + 0.10375359654426575 + ], + "credit_map": [ + 0.01822470873594284, + 0.013074134476482868, + 0.020534181967377663, + 0.03401753678917885, + 0.015224545262753963, + 0.01732981763780117, + 0.011403902433812618, + 0.01072714477777481, + 0.004463877994567156, + 0.0029900234658271074, + 0.004173784051090479, + 0.005387682467699051, + 0.0053224810399115086, + 0.005371248349547386, + 0.004489586688578129, + 0.0029724861960858107, + 0.010789827443659306, + 0.005984154995530844, + 0.008858227171003819, + 0.006702431011945009, + 0.01977708376944065, + 0.01211631204932928, + 0.008163698948919773, + 0.006752519402652979, + 0.006160719785839319, + 0.004728979896754026, + 0.007680073380470276, + 0.005843728315085173, + 0.009360500611364841, + 0.014627951197326183, + 0.008061735890805721, + 0.007718151900917292, + 0.009235579520463943, + 0.006323644891381264, + 0.01104740146547556, + 0.009737751446664333, + 0.01572471484541893, + 0.014320949092507362, + 0.015114997513592243, + 0.011339588090777397, + 0.01951170526444912, + 0.017890211194753647, + 0.03263213485479355, + 0.03151332214474678, + 0.049538012593984604, + 0.039983779191970825, + 0.03120175190269947, + 0.021416006609797478, + 0.006795783527195454, + 0.012859304435551167, + 0.014348635450005531, + 0.011226493865251541, + 0.007858664728701115, + 0.010583904571831226, + 0.010493100620806217, + 0.008269509300589561, + 0.008040627464652061, + 0.016548380255699158, + 0.007713689934462309, + 0.007318698801100254, + 0.009265247732400894, + 0.010301013477146626, + 0.005099568050354719, + 0.004951729439198971, + 0.007998181506991386, + 0.007599203381687403, + 0.00979226641356945, + 0.009861505590379238, + 0.02304636314511299, + 0.01454327441751957, + 0.008152192458510399, + 0.008014288730919361, + 0.011429849080741405, + 0.010036169551312923, + 0.01578986831009388, + 0.013305049389600754, + 0.014216842129826546, + 0.020519068464636803, + 0.009590704925358295, + 0.008866049349308014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.01419012900441885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494290217757225 + ], + "opposite_logprob_tokens": [ + -2.9299428462982178 + ], + "logprob_reward": -0.05494290217757225, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494290217757225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12419301271438599, + 0.0678134337067604, + 0.06442514061927795, + 0.06531310826539993, + 0.09830652922391891, + 0.08931034803390503, + 0.12224401533603668, + 0.14762549102306366, + 0.06448236107826233, + 0.1562865823507309 + ], + "credit_map": [ + 0.02008637972176075, + 0.014086002483963966, + 0.017472222447395325, + 0.017699608579277992, + 0.015848340466618538, + 0.020028190687298775, + 0.01006887387484312, + 0.008903393521904945, + 0.009176870808005333, + 0.007535390555858612, + 0.009720978327095509, + 0.010874498635530472, + 0.0070433043874800205, + 0.009578090161085129, + 0.008002576418220997, + 0.005881720222532749, + 0.008035649545490742, + 0.004200161900371313, + 0.0069689820520579815, + 0.005001790821552277, + 0.012045114301145077, + 0.013104991987347603, + 0.007751513738185167, + 0.00731693534180522, + 0.004660505335777998, + 0.005979455541819334, + 0.009569455869495869, + 0.013139267452061176, + 0.007724440190941095, + 0.013089151121675968, + 0.005424873903393745, + 0.0057259537279605865, + 0.010083945468068123, + 0.009672598913311958, + 0.014947854913771152, + 0.021396078169345856, + 0.008780539967119694, + 0.012184835970401764, + 0.012938067317008972, + 0.008302601054310799, + 0.00944964587688446, + 0.0070115807466208935, + 0.010402648709714413, + 0.008274494670331478, + 0.019078750163316727, + 0.01702355034649372, + 0.010935093276202679, + 0.007134573999792337, + 0.013881467282772064, + 0.02464202605187893, + 0.026583204045891762, + 0.014891752041876316, + 0.010920804925262928, + 0.012565694749355316, + 0.010645500384271145, + 0.008113555610179901, + 0.01627510040998459, + 0.043396417051553726, + 0.028438212350010872, + 0.012678977102041245, + 0.01545777078717947, + 0.011272911913692951, + 0.012883645482361317, + 0.007222456391900778, + 0.007855006493628025, + 0.008447249419987202, + 0.0067901769652962685, + 0.0060861059464514256, + 0.010079816915094852, + 0.010750902816653252, + 0.007743951864540577, + 0.006729142740368843, + 0.020611897110939026, + 0.023528089746832848, + 0.01681911014020443, + 0.016759006306529045, + 0.020293060690164566, + 0.02490857057273388, + 0.020475611090660095, + 0.012891232967376709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.05494290217757225 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, exudes the quality of highly detailed, color-saturated digital art. A middle-aged farmer, tanned and weathered, sits on the edge of a golden rice paddy. He wears a wide-brimmed straw hat, handwoven from dry straw, its brim frayed, its brim casting a shadow that partially obscures his deeply wrinkled face. He wears a simple, dark blue, coarse-sleeved shirt, sleeves casually rolled to the elbows, revealing arms sturdy from years of hard work. The twilight glow of the setting sun casts a soft, warm golden glow over the world. He holds an earthenware bowl steady in both hands, then decisively tilts his head back, his neck forming a powerful arch as he empties the clear liquor in a single, heroic gulp. His eyes are closed, his Adam's apple clearly rising and falling beneath his bronzed skin as he swallows. A few drops of wine, unswallowed, slid down the corners of his rough lips, gleaming in the setting sun. The distant ridges and sky were dyed a warm orange-red hue, and the entire scene was filled with a sense of tranquility and satisfaction after hard work, full of simple yet moving power.", + "seed": 20775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_1/index_9572_seed_20775_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028869427740573883, + "vlm_logprob": -0.028869427740573883, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.028869427740573883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04777950793504715 + ], + "opposite_logprob_tokens": [ + -3.4227795600891113 + ], + "logprob_reward": -0.04777950793504715, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04777950793504715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12174811214208603, + 0.05561976134777069, + 0.028299296274781227, + 0.050674792379140854, + 0.06988460570573807, + 0.12006150186061859, + 0.12419077754020691, + 0.1281118243932724, + 0.059242576360702515, + 0.24216672778129578 + ], + "credit_map": [ + 0.01161056850105524, + 0.016273237764835358, + 0.012770636938512325, + 0.016364101320505142, + 0.02223631925880909, + 0.014215807430446148, + 0.017195969820022583, + 0.011081475764513016, + 0.007538221776485443, + 0.005294489208608866, + 0.006143465172499418, + 0.006064563523977995, + 0.008756408467888832, + 0.003961754962801933, + 0.011002506129443645, + 0.00685835350304842, + 0.0029822150245308876, + 0.0031265930738300085, + 0.0028147692792117596, + 0.004850797820836306, + 0.004349933937191963, + 0.002544529503211379, + 0.004223494790494442, + 0.0034069640096277, + 0.00529236625880003, + 0.006427401211112738, + 0.007273764815181494, + 0.0065294308587908745, + 0.00786799006164074, + 0.004558898042887449, + 0.006545823998749256, + 0.006179113406687975, + 0.011582225561141968, + 0.0072588566690683365, + 0.007164096459746361, + 0.007723295595496893, + 0.014306937344372272, + 0.007611141540110111, + 0.006603724788874388, + 0.007634326349943876, + 0.0288303904235363, + 0.01094936765730381, + 0.04008673503994942, + 0.008856484666466713, + 0.00951963011175394, + 0.006767246872186661, + 0.005544212181121111, + 0.0095074363052845, + 0.01968616247177124, + 0.02785799279808998, + 0.020686272531747818, + 0.011838054284453392, + 0.013747490011155605, + 0.014823300763964653, + 0.007433554157614708, + 0.00811794400215149, + 0.013842081651091576, + 0.01679452322423458, + 0.02697557955980301, + 0.01187137421220541, + 0.016494618728756905, + 0.015967637300491333, + 0.013640007935464382, + 0.012526008300483227, + 0.010988778434693813, + 0.005906508769840002, + 0.009004997089505196, + 0.007771729025989771, + 0.006473400630056858, + 0.005504885222762823, + 0.0077079907059669495, + 0.005884286481887102, + 0.05727149546146393, + 0.015644341707229614, + 0.04603610560297966, + 0.025644151493906975, + 0.022550493478775024, + 0.044971346855163574, + 0.017104575410485268, + 0.012944231741130352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.04777950793504715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011094547808170319 + ], + "opposite_logprob_tokens": [ + -4.511094570159912 + ], + "logprob_reward": -0.011094547808170319, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011094547808170319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12418273836374283, + 0.05014714226126671, + 0.044068608433008194, + 0.11040084809064865, + 0.09688923507928848, + 0.08498531579971313, + 0.12511992454528809, + 0.0895576924085617, + 0.059368811547756195, + 0.2152797132730484 + ], + "credit_map": [ + 0.01164830569177866, + 0.017234787344932556, + 0.014581862837076187, + 0.011631439439952374, + 0.024571197107434273, + 0.012969344854354858, + 0.020574424415826797, + 0.010971372947096825, + 0.003730370430275798, + 0.0042947703041136265, + 0.00678778113797307, + 0.006465529557317495, + 0.01163401361554861, + 0.004004770889878273, + 0.00960095040500164, + 0.0036289573181420565, + 0.0035855041351169348, + 0.005077544134110212, + 0.0040579549968242645, + 0.0048796976916491985, + 0.009630409069359303, + 0.004277482628822327, + 0.0068588899448513985, + 0.005701127462089062, + 0.010993178933858871, + 0.015420620329678059, + 0.015880491584539413, + 0.012153349816799164, + 0.01640959084033966, + 0.01458815298974514, + 0.013973493129014969, + 0.01098196767270565, + 0.009374459274113178, + 0.007479189429432154, + 0.01993722841143608, + 0.0086908470839262, + 0.019784526899456978, + 0.016100818291306496, + 0.007768456358462572, + 0.007753717713057995, + 0.013559522107243538, + 0.005559122189879417, + 0.024265240877866745, + 0.008971409872174263, + 0.00904969684779644, + 0.011306402273476124, + 0.004957916680723429, + 0.007316000759601593, + 0.015693334862589836, + 0.021961720660328865, + 0.01875103823840618, + 0.018396828323602676, + 0.015904638916254044, + 0.013684424571692944, + 0.012196009047329426, + 0.008531923405826092, + 0.009510350413620472, + 0.012355799786746502, + 0.0174652561545372, + 0.00983438454568386, + 0.010472867637872696, + 0.009890620596706867, + 0.011765810661017895, + 0.008262593299150467, + 0.004864887334406376, + 0.005818796809762716, + 0.009043563157320023, + 0.009696492925286293, + 0.011735452339053154, + 0.004310240037739277, + 0.008531426079571247, + 0.00536795100197196, + 0.04003705084323883, + 0.00898277573287487, + 0.0706949308514595, + 0.04305436089634895, + 0.013237424194812775, + 0.018232714384794235, + 0.012266484089195728, + 0.00877397507429123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.011094547808170319 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052522532641887665 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.0052522532641887665, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052522532641887665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06004701182246208, + 0.04917561262845993, + 0.047645989805459976, + 0.0199002455919981, + 0.02670976147055626, + 0.09140735864639282, + 0.05026395991444588, + 0.05864220857620239, + 0.04418829828500748, + 0.5520195960998535 + ], + "credit_map": [ + 0.006564882583916187, + 0.005111237522214651, + 0.006423142738640308, + 0.008634415455162525, + 0.013382265344262123, + 0.005244905594736338, + 0.008832117542624474, + 0.0058540417812764645, + 0.00515472237020731, + 0.004992722533643246, + 0.00543504161760211, + 0.006520087830722332, + 0.010224026627838612, + 0.004369276110082865, + 0.006673384457826614, + 0.005806345026940107, + 0.0044770450331270695, + 0.007109755650162697, + 0.00529279513284564, + 0.004552777390927076, + 0.0099062854424119, + 0.006079938728362322, + 0.005781503394246101, + 0.004445885773748159, + 0.002838972955942154, + 0.001985226757824421, + 0.0030245680827647448, + 0.002076892415061593, + 0.0038043982349336147, + 0.0019168176222592592, + 0.002385181374847889, + 0.0018681855872273445, + 0.0037405870389193296, + 0.0024264308158308268, + 0.003993815742433071, + 0.0023414187598973513, + 0.006450253538787365, + 0.003580476390197873, + 0.0019485579105094075, + 0.002228217199444771, + 0.02154284343123436, + 0.008883839473128319, + 0.029563482850790024, + 0.006608922965824604, + 0.008268296718597412, + 0.005257145501673222, + 0.005330460146069527, + 0.0059523629024624825, + 0.0046494463458657265, + 0.004933768883347511, + 0.008109934628009796, + 0.00882850494235754, + 0.007718889974057674, + 0.006417237222194672, + 0.0049314722418785095, + 0.004674700088799, + 0.0059149907901883125, + 0.009499466978013515, + 0.010303482413291931, + 0.007857942953705788, + 0.007968460209667683, + 0.005450823809951544, + 0.0058138370513916016, + 0.005833199713379145, + 0.006448071449995041, + 0.004295583814382553, + 0.007159557193517685, + 0.006373309064656496, + 0.0070947082713246346, + 0.003398618660867214, + 0.005785542074590921, + 0.003632903564721346, + 0.11508829146623611, + 0.02571338415145874, + 0.1715802550315857, + 0.09831558167934418, + 0.031223798170685768, + 0.05702997371554375, + 0.028431229293346405, + 0.024637071415781975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0052522532641887665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818273961544037 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818273961544037, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818273961544037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11966931819915771, + 0.06979933381080627, + 0.07609623670578003, + 0.10416917502880096, + 0.07249914854764938, + 0.17728744447231293, + 0.059618350118398666, + 0.11595466732978821, + 0.05712663754820824, + 0.14777959883213043 + ], + "credit_map": [ + 0.010510276071727276, + 0.013110892847180367, + 0.013356595300137997, + 0.01457550935447216, + 0.0218791663646698, + 0.0179605595767498, + 0.01622617430984974, + 0.012050156481564045, + 0.007049530278891325, + 0.007797363214194775, + 0.006609347648918629, + 0.009481634944677353, + 0.012120029889047146, + 0.004796065390110016, + 0.01429172046482563, + 0.007653649430721998, + 0.006858279462903738, + 0.010308211669325829, + 0.007679391652345657, + 0.008605473674833775, + 0.014294416643679142, + 0.010152817703783512, + 0.01012839749455452, + 0.00806925818324089, + 0.01625683903694153, + 0.00971295591443777, + 0.0172880869358778, + 0.011918934993445873, + 0.012686445377767086, + 0.013239113613963127, + 0.012643414549529552, + 0.010423393920063972, + 0.008928336203098297, + 0.0065379575826227665, + 0.008036322891712189, + 0.00763863418251276, + 0.01359719131141901, + 0.012018205597996712, + 0.006446080282330513, + 0.009296424686908722, + 0.04393552243709564, + 0.017519818618893623, + 0.036863066256046295, + 0.01597742922604084, + 0.02191072516143322, + 0.01419485080987215, + 0.013378527015447617, + 0.01350754126906395, + 0.008225543424487114, + 0.007624895311892033, + 0.008815116249024868, + 0.009298973716795444, + 0.007284949533641338, + 0.006225176155567169, + 0.0062667508609592915, + 0.005876952316612005, + 0.013840966857969761, + 0.015448976308107376, + 0.01718628592789173, + 0.012431374751031399, + 0.015205347910523415, + 0.012534381821751595, + 0.014802932739257812, + 0.014504414051771164, + 0.008581368252635002, + 0.005206619389355183, + 0.009210141375660896, + 0.011396854184567928, + 0.007021022029221058, + 0.005014877766370773, + 0.0058997152373194695, + 0.004796047694981098, + 0.02839004248380661, + 0.008757142350077629, + 0.04273374378681183, + 0.025805460289120674, + 0.010887784883379936, + 0.0148256691172719, + 0.009188814088702202, + 0.007190958596765995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.01818273961544037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a middle-aged farmer with a tanned, weathered face wearing a wide-brimmed straw hat and a dark blue shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a golden rice paddy at twilight with a warm orange-red sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the farmer tilt his head back to empty a clear liquid from an earthenware bowl in a single gulp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible drops of liquid sliding down the corners of the farmer's lips while he swallows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203809008002281 + ], + "opposite_logprob_tokens": [ + -2.8120381832122803 + ], + "logprob_reward": -0.06203809008002281, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203809008002281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09421141445636749, + 0.058738648891448975, + 0.10242720693349838, + 0.18017645180225372, + 0.08794030547142029, + 0.07823096960783005, + 0.10033108294010162, + 0.15712805092334747, + 0.03408607095479965, + 0.10672983527183533 + ], + "credit_map": [ + 0.007657481357455254, + 0.011615739203989506, + 0.009793425910174847, + 0.011958113871514797, + 0.018777044489979744, + 0.007497818674892187, + 0.015250016935169697, + 0.011661767959594727, + 0.006745470222085714, + 0.005201748572289944, + 0.006855103652924299, + 0.00500012282282114, + 0.012913939543068409, + 0.00460134819149971, + 0.011736438609659672, + 0.0056844777427613735, + 0.008695371448993683, + 0.014434471726417542, + 0.010214187204837799, + 0.009763534180819988, + 0.0226970836520195, + 0.013232633471488953, + 0.014241114258766174, + 0.009148817509412766, + 0.026917796581983566, + 0.02832833305001259, + 0.020122863352298737, + 0.018423723056912422, + 0.02755410596728325, + 0.020845262333750725, + 0.019179213792085648, + 0.018805164843797684, + 0.009451819583773613, + 0.00876062922179699, + 0.013171573169529438, + 0.008472396992146969, + 0.017665119841694832, + 0.012902284972369671, + 0.008723193779587746, + 0.008793285116553307, + 0.00653947563841939, + 0.0061127482913434505, + 0.008055021986365318, + 0.007913615554571152, + 0.018997637555003166, + 0.00994714442640543, + 0.011348714120686054, + 0.009316609241068363, + 0.016458116471767426, + 0.017796382308006287, + 0.012418622151017189, + 0.01002972386777401, + 0.014182208105921745, + 0.013029877096414566, + 0.007105488330125809, + 0.009310662746429443, + 0.01816922426223755, + 0.020083237439393997, + 0.02427414245903492, + 0.016802605241537094, + 0.019958490505814552, + 0.01672106422483921, + 0.023022254928946495, + 0.018097039312124252, + 0.0047124833799898624, + 0.0038411554414778948, + 0.004209556616842747, + 0.00495154969394207, + 0.004286487121134996, + 0.003827124135568738, + 0.004577384330332279, + 0.0036803262773901224, + 0.016084885224699974, + 0.006582319270819426, + 0.03233464062213898, + 0.02022564224898815, + 0.009922313503921032, + 0.008037742227315903, + 0.007659789174795151, + 0.00588251044973731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.06203809008002281 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_tensor([4914, 4914])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_tensor([4914, 4914])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e5886ce6f9eaaafe06a21cf728bd0595e01ce760 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_tensor([4914, 4914])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle close-up depicts a minimalist chrome stapler, its polished body gleaming under a hard-key light, casting sharp, sharp shadows across the pristine surface of a clean white desk. The lighting meticulously sculpts the stapler's modern, geometric form, its mirror-like chrome finish adorned with brilliant mirror highlights. An unseen force presses the arm downward in a smooth, deliberate motion. Then, upon release, the arm returns upward in a single, swift, and decisive arc, exhibiting perfect critical decay as it instantly and perfectly comes to rest in its original resting position, devoid of any rebound or residual vibration. The entire action, set against a high-contrast, cool silver and white color palette, creates an atmosphere of refinement and focus, perfectly embodying the stylish aesthetic of corporate videos that emphasize precise engineering and cutting-edge modern design.", + "seed": 10256, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_4914_seed_10256_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17833133321255445, + "vlm_logprob": -0.17833133321255445, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.17833133321255445, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015392674133181572 + ], + "opposite_logprob_tokens": [ + -4.765392780303955 + ], + "logprob_reward": -0.015392674133181572, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015392674133181572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.127959206700325, + 0.08549477159976959, + 0.0415743850171566, + 0.04485292360186577, + 0.0851244106888771, + 0.1988631784915924, + 0.07280213385820389, + 0.13238157331943512, + 0.045428067445755005, + 0.1655193567276001 + ], + "credit_map": [ + 0.011177715845406055, + 0.012537242844700813, + 0.014769012108445168, + 0.01388018112629652, + 0.027680624276399612, + 0.014623893424868584, + 0.010698923841118813, + 0.022591618821024895, + 0.00925401784479618, + 0.012194306589663029, + 0.0111294686794281, + 0.014447523280978203, + 0.011946281418204308, + 0.009800119325518608, + 0.009540749713778496, + 0.007182298693805933, + 0.0032565651927143335, + 0.004685482010245323, + 0.008041554130613804, + 0.00660520838573575, + 0.006906226743012667, + 0.0037353443913161755, + 0.005369240418076515, + 0.0029747660737484694, + 0.004771779291331768, + 0.004874045494943857, + 0.006868738681077957, + 0.006218993104994297, + 0.005660573951900005, + 0.006318504922091961, + 0.006436827126890421, + 0.00370345963165164, + 0.006087903399020433, + 0.012929306365549564, + 0.009994424879550934, + 0.010815741494297981, + 0.006586246658116579, + 0.015351965092122555, + 0.014658101834356785, + 0.008700723759829998, + 0.017865408211946487, + 0.012591641396284103, + 0.02568170800805092, + 0.02393263764679432, + 0.00870489701628685, + 0.04202144220471382, + 0.019019123166799545, + 0.04904632270336151, + 0.006194085348397493, + 0.0032914818730205297, + 0.016333328559994698, + 0.008039075881242752, + 0.003497701371088624, + 0.011448932811617851, + 0.0049279313534498215, + 0.019069597125053406, + 0.0074962531216442585, + 0.0047434535808861256, + 0.009027861058712006, + 0.0127832917496562, + 0.0039364369586110115, + 0.044173575937747955, + 0.008936645463109016, + 0.041284047067165375, + 0.005181477405130863, + 0.0018322678515687585, + 0.013376323506236076, + 0.005769303534179926, + 0.0020365016534924507, + 0.005544756539165974, + 0.00261286785826087, + 0.009074567817151546, + 0.017224866896867752, + 0.006001987494528294, + 0.033447571098804474, + 0.015057970769703388, + 0.006229381076991558, + 0.03310877084732056, + 0.009104541502892971, + 0.0453442744910717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.015392674133181572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330845594406128 + ], + "opposite_logprob_tokens": [ + -3.773308515548706 + ], + "logprob_reward": -0.02330845594406128, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330845594406128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04264594614505768, + 0.1299986094236374, + 0.023732498288154602, + 0.022164391353726387, + 0.02033488266170025, + 0.1410408914089203, + 0.10366988182067871, + 0.04829344525933266, + 0.12276177853345871, + 0.3453577160835266 + ], + "credit_map": [ + 0.003870956366881728, + 0.005040499847382307, + 0.005512997042387724, + 0.006550297141075134, + 0.006484456360340118, + 0.005403392948210239, + 0.003577976021915674, + 0.006205367855727673, + 0.01266844104975462, + 0.02233356237411499, + 0.018091950565576553, + 0.01886753924190998, + 0.01447125431150198, + 0.010648363269865513, + 0.012543129734694958, + 0.020374363288283348, + 0.002600389998406172, + 0.0025134182069450617, + 0.003720530541613698, + 0.0054302336648106575, + 0.0030540309380739927, + 0.0018616409506648779, + 0.0023507247678935528, + 0.0022015252616256475, + 0.0022446666844189167, + 0.002370669739320874, + 0.003569119842723012, + 0.00351688452064991, + 0.0015643128426745534, + 0.001484199776314199, + 0.005396159831434488, + 0.002018377184867859, + 0.002369087655097246, + 0.001660328940488398, + 0.0024079184513539076, + 0.004172156564891338, + 0.0014494039351120591, + 0.00278024491854012, + 0.0031832209788262844, + 0.0023125202860683203, + 0.012511951848864555, + 0.008273826912045479, + 0.034194592386484146, + 0.02587735839188099, + 0.006519929971545935, + 0.01596454530954361, + 0.01567588374018669, + 0.0220227912068367, + 0.010385698638856411, + 0.006051534321159124, + 0.023589065298438072, + 0.01931425929069519, + 0.0058178068138659, + 0.01513582095503807, + 0.008614549413323402, + 0.014761138707399368, + 0.004706779029220343, + 0.002914577955380082, + 0.0122199896723032, + 0.00941824447363615, + 0.0023574698716402054, + 0.00574702862650156, + 0.00568443862721324, + 0.0052449144423007965, + 0.01500582043081522, + 0.005294522270560265, + 0.03439842537045479, + 0.014245297759771347, + 0.005219831131398678, + 0.015494833700358868, + 0.007087273057550192, + 0.026015762239694595, + 0.03991890326142311, + 0.01371585950255394, + 0.07579239457845688, + 0.03265761584043503, + 0.012643273919820786, + 0.052159905433654785, + 0.0322466678917408, + 0.0862230658531189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.02330845594406128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421763002872467 + ], + "opposite_logprob_tokens": [ + -4.264217853546143 + ], + "logprob_reward": -0.01421763002872467, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421763002872467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12928076088428497, + 0.29497385025024414, + 0.08906129002571106, + 0.07283516973257065, + 0.057368505746126175, + 0.18071097135543823, + 0.05267583578824997, + 0.05164415389299393, + 0.03526139631867409, + 0.03618812933564186 + ], + "credit_map": [ + 0.012257936410605907, + 0.018774624913930893, + 0.016033289954066277, + 0.02021983079612255, + 0.01851106435060501, + 0.013262434862554073, + 0.014320607297122478, + 0.01590096764266491, + 0.03170881047844887, + 0.044640690088272095, + 0.03970625251531601, + 0.0434282086789608, + 0.035666678100824356, + 0.027405718341469765, + 0.028178658336400986, + 0.04423883184790611, + 0.009091385640203953, + 0.013173431158065796, + 0.013126378878951073, + 0.01180961448699236, + 0.011181678622961044, + 0.0059069981798529625, + 0.014131824485957623, + 0.010639977641403675, + 0.007785456720739603, + 0.008361894637346268, + 0.008958562277257442, + 0.008823364973068237, + 0.008219107985496521, + 0.009531577117741108, + 0.014155872166156769, + 0.0069993347860872746, + 0.004703668877482414, + 0.005972796585410833, + 0.006910983007401228, + 0.006425363011658192, + 0.0074949609115719795, + 0.010774743743240833, + 0.01039652805775404, + 0.004689459688961506, + 0.019976947456598282, + 0.013321951031684875, + 0.03575973957777023, + 0.019484858959913254, + 0.00892678927630186, + 0.03969993069767952, + 0.01885128580033779, + 0.024689463898539543, + 0.006701322738081217, + 0.002532642800360918, + 0.014348412863910198, + 0.008171672001481056, + 0.0026614817325025797, + 0.007393626030534506, + 0.003087653312832117, + 0.007779024075716734, + 0.006507570389658213, + 0.002776856068521738, + 0.015339110046625137, + 0.007136377040296793, + 0.002413054695352912, + 0.008043309673666954, + 0.0036697168834507465, + 0.005758162122219801, + 0.004894907120615244, + 0.0025178391952067614, + 0.008526026271283627, + 0.0037099968176335096, + 0.0017581989523023367, + 0.006174649577587843, + 0.0038230642676353455, + 0.0038567129522562027, + 0.002192560350522399, + 0.001632586121559143, + 0.004354521632194519, + 0.002519074361771345, + 0.0012716181809082627, + 0.013186615891754627, + 0.0019112152513116598, + 0.009119939059019089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.01421763002872467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261916041374207 + ], + "opposite_logprob_tokens": [ + -0.5761916041374207 + ], + "logprob_reward": -0.8261916041374207, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261916041374207, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09462130069732666, + 0.3530045747756958, + 0.03876664489507675, + 0.03345237299799919, + 0.054241862148046494, + 0.14001576602458954, + 0.04625644162297249, + 0.0892096683382988, + 0.05569009110331535, + 0.09474126249551773 + ], + "credit_map": [ + 0.010505735874176025, + 0.009692631661891937, + 0.011357619427144527, + 0.014044591225683689, + 0.01578732579946518, + 0.009860452264547348, + 0.009605640545487404, + 0.013767304830253124, + 0.03645331412553787, + 0.059333622455596924, + 0.04869600757956505, + 0.05531448870897293, + 0.039005670696496964, + 0.03414076939225197, + 0.03462507203221321, + 0.045435648411512375, + 0.0042691887356340885, + 0.006940971594303846, + 0.004415513016283512, + 0.004956355784088373, + 0.005419349297881126, + 0.0022908139508217573, + 0.006061162333935499, + 0.004413289949297905, + 0.003435075283050537, + 0.006551372352987528, + 0.004598427563905716, + 0.003276698524132371, + 0.0035710742231458426, + 0.0026863454841077328, + 0.005856980569660664, + 0.0034764015581458807, + 0.004026738461107016, + 0.006436426192522049, + 0.006527932360768318, + 0.005516607314348221, + 0.009240394458174706, + 0.009636485949158669, + 0.007680416572839022, + 0.005176861770451069, + 0.010575016029179096, + 0.007118758279830217, + 0.014028486795723438, + 0.01699475385248661, + 0.005846492014825344, + 0.03344128280878067, + 0.01373458094894886, + 0.038276396691799164, + 0.0043635317124426365, + 0.003172748489305377, + 0.003989559132605791, + 0.0049112532287836075, + 0.0024054960813373327, + 0.011717893183231354, + 0.005363397300243378, + 0.010332562029361725, + 0.010319110937416553, + 0.004386491607874632, + 0.017468461766839027, + 0.00973476655781269, + 0.004228746052831411, + 0.017353111878037453, + 0.005534838885068893, + 0.020184138789772987, + 0.006321674212813377, + 0.0028571493458002806, + 0.013979876413941383, + 0.006530900951474905, + 0.0025661750696599483, + 0.01334475539624691, + 0.0037654635962098837, + 0.006324099842458963, + 0.008274752646684647, + 0.003817678662016988, + 0.012211029417812824, + 0.008063800632953644, + 0.004155466333031654, + 0.024671724066138268, + 0.008403506129980087, + 0.02514330856502056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.8261916041374207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546301819384098 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546301819384098, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546301819384098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07309090346097946, + 0.0345207154750824, + 0.04186395928263664, + 0.028976690024137497, + 0.09403987228870392, + 0.1297302395105362, + 0.07541143894195557, + 0.13032670319080353, + 0.07471852004528046, + 0.31732094287872314 + ], + "credit_map": [ + 0.010816234163939953, + 0.0062362211756408215, + 0.009122297167778015, + 0.009101900272071362, + 0.009827366098761559, + 0.0125731835141778, + 0.004895742516964674, + 0.010517955757677555, + 0.0033261983189731836, + 0.004852852784097195, + 0.004763148259371519, + 0.005869587883353233, + 0.004364089109003544, + 0.0034239953383803368, + 0.004120438825339079, + 0.003800404956564307, + 0.004044205415993929, + 0.005419235676527023, + 0.00838733371347189, + 0.006040130741894245, + 0.00420779176056385, + 0.0034478171728551388, + 0.005640573333948851, + 0.004676871467381716, + 0.0034177142661064863, + 0.003534504910930991, + 0.0056288037449121475, + 0.004082138650119305, + 0.0029207703191787004, + 0.0018059874419122934, + 0.005242858547717333, + 0.0023439121432602406, + 0.006680629216134548, + 0.013514700345695019, + 0.011106024496257305, + 0.011471658013761044, + 0.00858785305172205, + 0.01886279322206974, + 0.015526512637734413, + 0.0082897013053298, + 0.017115484923124313, + 0.010479713790118694, + 0.014893469400703907, + 0.0180194154381752, + 0.008915513753890991, + 0.02429545484483242, + 0.02323770709335804, + 0.012773468159139156, + 0.008111344650387764, + 0.0035318387672305107, + 0.01977129466831684, + 0.011869316920638084, + 0.003395618172362447, + 0.009963802061975002, + 0.0063650766387581825, + 0.012403142638504505, + 0.009686254896223545, + 0.00409594364464283, + 0.00856175646185875, + 0.01349403616040945, + 0.005029017571359873, + 0.044276636093854904, + 0.007117163389921188, + 0.03806590288877487, + 0.007759496569633484, + 0.0029764785431325436, + 0.017937544733285904, + 0.008021499961614609, + 0.0031772307120263577, + 0.011632007546722889, + 0.003915492445230484, + 0.019298771396279335, + 0.04084714502096176, + 0.01581800915300846, + 0.08801314979791641, + 0.03587473928928375, + 0.013949567452073097, + 0.0431273989379406, + 0.027190307155251503, + 0.05250063166022301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.012546301819384098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up depicts a minimalist chrome stapler, its polished body gleaming under a hard-key light, casting sharp, sharp shadows across the pristine surface of a clean white desk. The lighting meticulously sculpts the stapler's modern, geometric form, its mirror-like chrome finish adorned with brilliant mirror highlights. An unseen force presses the arm downward in a smooth, deliberate motion. Then, upon release, the arm returns upward in a single, swift, and decisive arc, exhibiting perfect critical decay as it instantly and perfectly comes to rest in its original resting position, devoid of any rebound or residual vibration. The entire action, set against a high-contrast, cool silver and white color palette, creates an atmosphere of refinement and focus, perfectly embodying the stylish aesthetic of corporate videos that emphasize precise engineering and cutting-edge modern design.", + "seed": 10257, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_4914_seed_10257_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16470844214782118, + "vlm_logprob": -0.16470844214782118, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.16470844214782118, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012277857400476933 + ], + "opposite_logprob_tokens": [ + -4.887278079986572 + ], + "logprob_reward": -0.012277857400476933, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012277857400476933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1750706434249878, + 0.05308331921696663, + 0.07877969741821289, + 0.21959993243217468, + 0.0973295122385025, + 0.05494798719882965, + 0.10160303860902786, + 0.11846455186605453, + 0.078831747174263, + 0.02228963002562523 + ], + "credit_map": [ + 0.020329978317022324, + 0.014095737598836422, + 0.01786908507347107, + 0.02328682504594326, + 0.03397198021411896, + 0.020102398470044136, + 0.024618789553642273, + 0.020795857533812523, + 0.008691227063536644, + 0.0047409748658537865, + 0.005729432683438063, + 0.007829401642084122, + 0.008972179144620895, + 0.0040093883872032166, + 0.0044871061109006405, + 0.008623607456684113, + 0.014656326733529568, + 0.016830451786518097, + 0.007557409815490246, + 0.013918674550950527, + 0.009712140075862408, + 0.00461984146386385, + 0.005458613391965628, + 0.006026237737387419, + 0.014709358103573322, + 0.04836538806557655, + 0.025006229057908058, + 0.062533438205719, + 0.018534550443291664, + 0.013967525213956833, + 0.022528931498527527, + 0.013954506255686283, + 0.01212821900844574, + 0.007134594488888979, + 0.010468443855643272, + 0.01910477690398693, + 0.01442035660147667, + 0.012086505070328712, + 0.011170575395226479, + 0.010816040448844433, + 0.00776040880009532, + 0.007474631071090698, + 0.006630659569054842, + 0.010448118671774864, + 0.006799373310059309, + 0.004271765239536762, + 0.006168881896883249, + 0.005394148174673319, + 0.01570405624806881, + 0.005435168277472258, + 0.012556805275380611, + 0.016265474259853363, + 0.011257561855018139, + 0.011440848931670189, + 0.01474272646009922, + 0.014200398698449135, + 0.015165911056101322, + 0.010522831231355667, + 0.014274582266807556, + 0.03059433028101921, + 0.009794912301003933, + 0.009957544505596161, + 0.014459357596933842, + 0.013695081695914268, + 0.01283290795981884, + 0.007018992677330971, + 0.009196759201586246, + 0.020597340539097786, + 0.007266262546181679, + 0.005788146052509546, + 0.007348876912146807, + 0.008782457560300827, + 0.0030461677815765142, + 0.0016144075198099017, + 0.002627698937430978, + 0.0047563351690769196, + 0.0024268864654004574, + 0.0024919488932937384, + 0.0026039520744234324, + 0.002722232835367322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.012277857400476933 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371458128094673 + ], + "opposite_logprob_tokens": [ + -3.408714532852173 + ], + "logprob_reward": -0.03371458128094673, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371458128094673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11152345687150955, + 0.28656259179115295, + 0.07423575222492218, + 0.18382734060287476, + 0.11260931193828583, + 0.055348221212625504, + 0.06664422154426575, + 0.060285232961177826, + 0.015519019216299057, + 0.0334448367357254 + ], + "credit_map": [ + 0.013039772398769855, + 0.008847653865814209, + 0.011229022406041622, + 0.024057719856500626, + 0.020954754203557968, + 0.008124190382659435, + 0.012823878787457943, + 0.012446464039385319, + 0.041019026190042496, + 0.024773407727479935, + 0.03706616908311844, + 0.07244770973920822, + 0.03296640142798424, + 0.020270710811018944, + 0.027149317786097527, + 0.030869876965880394, + 0.014680620282888412, + 0.009460676461458206, + 0.007560094352811575, + 0.01348526868969202, + 0.009518926031887531, + 0.006802867632359266, + 0.004726211074739695, + 0.008001089096069336, + 0.0178128182888031, + 0.02167779579758644, + 0.0239644106477499, + 0.052701909095048904, + 0.01435093954205513, + 0.012889905832707882, + 0.018235914409160614, + 0.022193673998117447, + 0.015416866168379784, + 0.007839290425181389, + 0.013424105942249298, + 0.01902533881366253, + 0.018912311643362045, + 0.009865128435194492, + 0.013835921883583069, + 0.01429035235196352, + 0.008604733273386955, + 0.0040868488140404224, + 0.006695762742310762, + 0.010819737799465656, + 0.004890911281108856, + 0.00532302213832736, + 0.007358410861343145, + 0.00756880221888423, + 0.012613199651241302, + 0.004041228909045458, + 0.009366896003484726, + 0.009763704612851143, + 0.007486063987016678, + 0.005088865756988525, + 0.008677576668560505, + 0.009606684558093548, + 0.009438670240342617, + 0.005071701481938362, + 0.00718093803152442, + 0.011096916161477566, + 0.00604371540248394, + 0.006151212379336357, + 0.00789286196231842, + 0.007409217301756144, + 0.0026380647905170918, + 0.0011186759220436215, + 0.0018342636758461595, + 0.0024408067110925913, + 0.001665719086304307, + 0.001879526418633759, + 0.001844752230681479, + 0.0020972106140106916, + 0.007496266160160303, + 0.002556045539677143, + 0.0043087066151201725, + 0.005442949011921883, + 0.0036926413886249065, + 0.002210676670074463, + 0.003271766472607851, + 0.004465784411877394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.03371458128094673 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016083618625998497 + ], + "opposite_logprob_tokens": [ + -4.141083717346191 + ], + "logprob_reward": -0.016083618625998497, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016083618625998497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22886012494564056, + 0.054289717227220535, + 0.08011770993471146, + 0.06814976781606674, + 0.09387046843767166, + 0.07519792020320892, + 0.1135932132601738, + 0.18463784456253052, + 0.09180639684200287, + 0.009476871229708195 + ], + "credit_map": [ + 0.03178686276078224, + 0.019734084606170654, + 0.023300651460886, + 0.04555020481348038, + 0.04034380987286568, + 0.017022520303726196, + 0.027003945782780647, + 0.024118054658174515, + 0.005928677506744862, + 0.004322707653045654, + 0.0065139289945364, + 0.013990718871355057, + 0.00911992322653532, + 0.0038819776382297277, + 0.005262913182377815, + 0.005268871318548918, + 0.005951535888016224, + 0.009384052827954292, + 0.010688980109989643, + 0.021759480237960815, + 0.00884893536567688, + 0.007455756887793541, + 0.00830700434744358, + 0.007721964735537767, + 0.0050369990058243275, + 0.007901380769908428, + 0.008020839653909206, + 0.014554976485669613, + 0.006763344630599022, + 0.006042930297553539, + 0.011693601496517658, + 0.008135695941746235, + 0.013540933839976788, + 0.008114493452012539, + 0.010055282153189182, + 0.013569840230047703, + 0.018178926780819893, + 0.009168938733637333, + 0.012250554747879505, + 0.008991497568786144, + 0.006657802499830723, + 0.005684157367795706, + 0.010180063545703888, + 0.012956767342984676, + 0.007858983241021633, + 0.008046040311455727, + 0.011534415185451508, + 0.012279695831239223, + 0.016283880919218063, + 0.00931102130562067, + 0.013135995715856552, + 0.020363803952932358, + 0.0117122707888484, + 0.011430120095610619, + 0.01705918088555336, + 0.014296943321824074, + 0.026327671483159065, + 0.013641594909131527, + 0.01991135999560356, + 0.04350239783525467, + 0.016906699165701866, + 0.01735607720911503, + 0.027554314583539963, + 0.01943773403763771, + 0.013927829451858997, + 0.007884557358920574, + 0.010518483817577362, + 0.02262629009783268, + 0.008045685477554798, + 0.007469447795301676, + 0.012261117808520794, + 0.009072981774806976, + 0.0010671436320990324, + 0.0007053371518850327, + 0.0012142184423282743, + 0.0020823634695261717, + 0.0008246967336162925, + 0.0012763675767928362, + 0.0011675552232190967, + 0.0011391888838261366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.016083618625998497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578493356704712 + ], + "opposite_logprob_tokens": [ + -0.6328493356704712 + ], + "logprob_reward": -0.7578493356704712, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578493356704712, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12334827333688736, + 0.08912525326013565, + 0.07585937529802322, + 0.25292345881462097, + 0.10292509943246841, + 0.054629676043987274, + 0.07037615776062012, + 0.12097891420125961, + 0.0761791244149208, + 0.03365473821759224 + ], + "credit_map": [ + 0.016823621466755867, + 0.00903301127254963, + 0.012067348696291447, + 0.021092230454087257, + 0.01989113539457321, + 0.013443776406347752, + 0.01514251809567213, + 0.015854611992836, + 0.014059905894100666, + 0.008719292469322681, + 0.010020026937127113, + 0.015627270564436913, + 0.014903606846928596, + 0.007472426164895296, + 0.00748828100040555, + 0.010834434069693089, + 0.01065096352249384, + 0.010695777833461761, + 0.007936843670904636, + 0.013910174369812012, + 0.008927668444812298, + 0.007011186331510544, + 0.006416557356715202, + 0.010310192592442036, + 0.014072365127503872, + 0.042069729417562485, + 0.03187195211648941, + 0.061338163912296295, + 0.02694522961974144, + 0.020008087158203125, + 0.03397328406572342, + 0.022644614800810814, + 0.014622688293457031, + 0.007777577266097069, + 0.010501580312848091, + 0.017258023843169212, + 0.01839514449238777, + 0.012946977280080318, + 0.011765928007662296, + 0.009657174348831177, + 0.0051938616670668125, + 0.004669704008847475, + 0.006367067340761423, + 0.009796337224543095, + 0.006725040730088949, + 0.007621008437126875, + 0.00630971509963274, + 0.00794693548232317, + 0.007992140017449856, + 0.00456269783899188, + 0.008317339234054089, + 0.014215877279639244, + 0.00623293174430728, + 0.007534458301961422, + 0.009864590130746365, + 0.01165611669421196, + 0.013975818641483784, + 0.008883548900485039, + 0.014110429212450981, + 0.023695291951298714, + 0.011800331063568592, + 0.01387033425271511, + 0.019130371510982513, + 0.01551277469843626, + 0.007997252978384495, + 0.0049197496846318245, + 0.00878326129168272, + 0.011432172730565071, + 0.008660915307700634, + 0.010719286277890205, + 0.013717995025217533, + 0.009948483668267727, + 0.0046883998438715935, + 0.0025919878389686346, + 0.003539986675605178, + 0.009445073083043098, + 0.00291971443220973, + 0.003361776005476713, + 0.003780958941206336, + 0.0033268341794610023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.7578493356704712 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036168177612125874 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036168177612125874, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036168177612125874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13190485537052155, + 0.10782737284898758, + 0.12781257927417755, + 0.17387999594211578, + 0.15469560027122498, + 0.11486939340829849, + 0.04101527854800224, + 0.04868343845009804, + 0.037730008363723755, + 0.06158147379755974 + ], + "credit_map": [ + 0.017887121066451073, + 0.009026351384818554, + 0.015716295689344406, + 0.030103040859103203, + 0.0133567675948143, + 0.009124718606472015, + 0.014768684282898903, + 0.021921884268522263, + 0.021156765520572662, + 0.010092873126268387, + 0.012047894299030304, + 0.01653035543859005, + 0.016912145540118217, + 0.008147403597831726, + 0.009082435630261898, + 0.013857495971024036, + 0.02842724695801735, + 0.015988696366548538, + 0.015392604283988476, + 0.023777801543474197, + 0.014340580441057682, + 0.008815567940473557, + 0.008368260227143764, + 0.012701822444796562, + 0.01864236779510975, + 0.030860407277941704, + 0.017743410542607307, + 0.024204449728131294, + 0.019875550642609596, + 0.01715691201388836, + 0.027154909446835518, + 0.018241995945572853, + 0.03864626586437225, + 0.012049093842506409, + 0.020390218123793602, + 0.019812755286693573, + 0.01775842346251011, + 0.011025305837392807, + 0.011519581079483032, + 0.02349395491182804, + 0.028152504935860634, + 0.008863192051649094, + 0.013695204630494118, + 0.016528448089957237, + 0.01296553947031498, + 0.008820099756121635, + 0.009311333298683167, + 0.01653306931257248, + 0.00717847840860486, + 0.002642004517838359, + 0.005336947739124298, + 0.007136364001780748, + 0.003623932134360075, + 0.0035935528576374054, + 0.0056376284919679165, + 0.005866370629519224, + 0.008301272988319397, + 0.0034629555884748697, + 0.005355103872716427, + 0.01143116969615221, + 0.0045600030571222305, + 0.003441764973104, + 0.006280684377998114, + 0.005850485060364008, + 0.005522295366972685, + 0.0031212985049933195, + 0.004604402929544449, + 0.007413182873278856, + 0.0036473628133535385, + 0.004136369097977877, + 0.004727991763502359, + 0.004557103384286165, + 0.016773823648691177, + 0.005400930996984243, + 0.0067616659216582775, + 0.007150884717702866, + 0.0071861776523292065, + 0.0049754781648516655, + 0.005354281049221754, + 0.007978231646120548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.0036168177612125874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up depicts a minimalist chrome stapler, its polished body gleaming under a hard-key light, casting sharp, sharp shadows across the pristine surface of a clean white desk. The lighting meticulously sculpts the stapler's modern, geometric form, its mirror-like chrome finish adorned with brilliant mirror highlights. An unseen force presses the arm downward in a smooth, deliberate motion. Then, upon release, the arm returns upward in a single, swift, and decisive arc, exhibiting perfect critical decay as it instantly and perfectly comes to rest in its original resting position, devoid of any rebound or residual vibration. The entire action, set against a high-contrast, cool silver and white color palette, creates an atmosphere of refinement and focus, perfectly embodying the stylish aesthetic of corporate videos that emphasize precise engineering and cutting-edge modern design.", + "seed": 10258, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_4914_seed_10258_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16549191288650036, + "vlm_logprob": -0.16549191288650036, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.16549191288650036, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013576318509876728 + ], + "opposite_logprob_tokens": [ + -5.138576507568359 + ], + "logprob_reward": -0.013576318509876728, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.013576318509876728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07787661999464035, + 0.14698649942874908, + 0.05713925510644913, + 0.01833650842308998, + 0.113838791847229, + 0.09375737607479095, + 0.10678789764642715, + 0.14047712087631226, + 0.04328508302569389, + 0.20151488482952118 + ], + "credit_map": [ + 0.007251088507473469, + 0.012954008765518665, + 0.006488456856459379, + 0.011838673613965511, + 0.00765796285122633, + 0.010096740908920765, + 0.012528743594884872, + 0.009060944430530071, + 0.010586974211037159, + 0.01159575954079628, + 0.011184034869074821, + 0.03052164614200592, + 0.012413627468049526, + 0.01825903356075287, + 0.032421696931123734, + 0.020003722980618477, + 0.004996945150196552, + 0.005259238183498383, + 0.006097816396504641, + 0.00810623075813055, + 0.0067564756609499454, + 0.009148850105702877, + 0.00863964669406414, + 0.008134052157402039, + 0.002199389971792698, + 0.0018629905534908175, + 0.003106417367234826, + 0.002055163262411952, + 0.0022608686704188585, + 0.002196219516918063, + 0.0025171178858727217, + 0.0021383408457040787, + 0.014150796458125114, + 0.01293623074889183, + 0.01320743653923273, + 0.01720576360821724, + 0.00954103097319603, + 0.013075582683086395, + 0.0177773367613554, + 0.01594461128115654, + 0.00901287142187357, + 0.009360711090266705, + 0.0071905264630913734, + 0.012327085249125957, + 0.0170845165848732, + 0.01214058417826891, + 0.015705788508057594, + 0.01093529723584652, + 0.008675324730575085, + 0.003918419126421213, + 0.005619820207357407, + 0.005877912510186434, + 0.004488240461796522, + 0.005476633086800575, + 0.03499810770153999, + 0.03773343935608864, + 0.026179393753409386, + 0.007203869521617889, + 0.00696689123287797, + 0.006129883229732513, + 0.007921002805233002, + 0.05740921199321747, + 0.018201148137450218, + 0.010465720668435097, + 0.008472993969917297, + 0.0017511623445898294, + 0.0021513996180146933, + 0.002286735689267516, + 0.0026206739712506533, + 0.0032783898059278727, + 0.008417443372309208, + 0.014306280761957169, + 0.06452038139104843, + 0.008307630196213722, + 0.012710092589259148, + 0.010180792771279812, + 0.0060718185268342495, + 0.013171031139791012, + 0.015060569159686565, + 0.07149256020784378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.013576318509876728 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372230380773544 + ], + "opposite_logprob_tokens": [ + -3.408722400665283 + ], + "logprob_reward": -0.03372230380773544, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372230380773544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12246957421302795, + 0.0831497311592102, + 0.03953796997666359, + 0.036339543759822845, + 0.07856892049312592, + 0.0227977242320776, + 0.030743852257728577, + 0.04272482916712761, + 0.3683152198791504, + 0.17535264790058136 + ], + "credit_map": [ + 0.01567426696419716, + 0.014357154257595539, + 0.009861252270638943, + 0.012346016243100166, + 0.014926335774362087, + 0.02391267567873001, + 0.015886574983596802, + 0.015505298972129822, + 0.00617050938308239, + 0.005874300375580788, + 0.006056647282093763, + 0.009785016998648643, + 0.00994411576539278, + 0.016269037500023842, + 0.016728833317756653, + 0.012321273796260357, + 0.00409987336024642, + 0.004148920997977257, + 0.004873438738286495, + 0.00808872189372778, + 0.0025544308591634035, + 0.005073318257927895, + 0.005738990381360054, + 0.004960274323821068, + 0.00437874486669898, + 0.004420774057507515, + 0.00435533095151186, + 0.00357505539432168, + 0.004036588594317436, + 0.007070690393447876, + 0.004531942307949066, + 0.003970417659729719, + 0.008882887661457062, + 0.01467832364141941, + 0.007700288202613592, + 0.008760234341025352, + 0.007918812334537506, + 0.009904456324875355, + 0.009251736104488373, + 0.011472183279693127, + 0.0025187579449266195, + 0.0021934511605650187, + 0.0024115818087011576, + 0.0034539399202913046, + 0.0035715969279408455, + 0.0033609403762966394, + 0.002777631161734462, + 0.0025098237674683332, + 0.0024726104456931353, + 0.002374661387875676, + 0.0029229994397610426, + 0.0035834270529448986, + 0.002390257315710187, + 0.002324672881513834, + 0.006746913772076368, + 0.00792831089347601, + 0.011783814057707787, + 0.002409768057987094, + 0.003120804438367486, + 0.0021919908467680216, + 0.0030271965079009533, + 0.012798365205526352, + 0.003881098236888647, + 0.0035117887891829014, + 0.062310364097356796, + 0.0098799429833889, + 0.011647566221654415, + 0.014258861541748047, + 0.009316056035459042, + 0.013882516883313656, + 0.10647471994161606, + 0.14054515957832336, + 0.05838903784751892, + 0.007421779911965132, + 0.01173405908048153, + 0.011378984898328781, + 0.00796180497854948, + 0.011490859091281891, + 0.031579792499542236, + 0.035396333783864975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.03372230380773544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792031720280647 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009792031720280647, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792031720280647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09541886299848557, + 0.0474310927093029, + 0.10487407445907593, + 0.03306062892079353, + 0.17131704092025757, + 0.12320080399513245, + 0.0944962278008461, + 0.03680245205760002, + 0.20045073330402374, + 0.09294802695512772 + ], + "credit_map": [ + 0.009887845255434513, + 0.01845446042716503, + 0.009967154823243618, + 0.011345542035996914, + 0.013381067663431168, + 0.009802783839404583, + 0.011135733686387539, + 0.011444282718002796, + 0.0042487140744924545, + 0.0034097826573997736, + 0.004051401745527983, + 0.008213282562792301, + 0.004301552660763264, + 0.00845455750823021, + 0.009156206622719765, + 0.005595602560788393, + 0.009672870859503746, + 0.010411074385046959, + 0.009411103092133999, + 0.02037537842988968, + 0.010153690353035927, + 0.014868698082864285, + 0.016235563904047012, + 0.013745700009167194, + 0.003240397898480296, + 0.003091663820669055, + 0.003713499754667282, + 0.005066350568085909, + 0.004310990683734417, + 0.004535098094493151, + 0.005133579019457102, + 0.003969051409512758, + 0.02338337153196335, + 0.01885003037750721, + 0.02134644240140915, + 0.023402413353323936, + 0.015347730368375778, + 0.015249883756041527, + 0.026302199810743332, + 0.027434980496764183, + 0.012853562831878662, + 0.011804590933024883, + 0.00977274402976036, + 0.012560727074742317, + 0.02961144410073757, + 0.01852782815694809, + 0.015035433694720268, + 0.013034485280513763, + 0.006269638426601887, + 0.0025107506662607193, + 0.0035691028460860252, + 0.004802256356924772, + 0.0033794704359024763, + 0.0039117299020290375, + 0.0341978520154953, + 0.035855431109666824, + 0.0053056273609399796, + 0.0011874967021867633, + 0.0011838125064969063, + 0.0019892386626452208, + 0.001376916654407978, + 0.009256841614842415, + 0.010365186259150505, + 0.0061373342759907246, + 0.03710593655705452, + 0.005689654033631086, + 0.006927062291651964, + 0.007621780503541231, + 0.0047752284444868565, + 0.007651207968592644, + 0.05432730168104172, + 0.0763525739312172, + 0.030667437240481377, + 0.0037190874572843313, + 0.005848694127053022, + 0.004259065259248018, + 0.003019643248990178, + 0.005241115111857653, + 0.006860889960080385, + 0.03333210200071335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.009792031720280647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578254342079163 + ], + "opposite_logprob_tokens": [ + -0.6328254342079163 + ], + "logprob_reward": -0.7578254342079163, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578254342079163, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0959426537156105, + 0.04721079766750336, + 0.16458754241466522, + 0.048439763486385345, + 0.1958765685558319, + 0.0530487596988678, + 0.06634771823883057, + 0.044768039137125015, + 0.12071492522954941, + 0.16306331753730774 + ], + "credit_map": [ + 0.011606533080339432, + 0.017145803198218346, + 0.00979875959455967, + 0.011241619475185871, + 0.011051658540964127, + 0.012896030209958553, + 0.010553835891187191, + 0.011648403480648994, + 0.005345752928406, + 0.0056239524856209755, + 0.005441649351269007, + 0.005799183622002602, + 0.008203289471566677, + 0.0048556262627244, + 0.006185866426676512, + 0.005755472928285599, + 0.013363802805542946, + 0.01411197055131197, + 0.018857432529330254, + 0.019476693123579025, + 0.02022589184343815, + 0.03593909367918968, + 0.021823594346642494, + 0.02078905515372753, + 0.0044126263819634914, + 0.005197982769459486, + 0.006585110444575548, + 0.004800318740308285, + 0.0062015182338654995, + 0.007357511203736067, + 0.007586588617414236, + 0.006298101507127285, + 0.02606896124780178, + 0.02703460119664669, + 0.021745454519987106, + 0.026783138513565063, + 0.016229549422860146, + 0.0176504235714674, + 0.02863086201250553, + 0.03173356130719185, + 0.004699122626334429, + 0.004489605315029621, + 0.0032717541325837374, + 0.004595677834004164, + 0.013179751113057137, + 0.008611171506345272, + 0.006333421915769577, + 0.007868249900639057, + 0.006532999686896801, + 0.004801366478204727, + 0.007314677350223064, + 0.00671247998252511, + 0.006417543161660433, + 0.005478908773511648, + 0.01573207415640354, + 0.013357661664485931, + 0.006525867152959108, + 0.0038415645249187946, + 0.0028428160585463047, + 0.0034252232871949673, + 0.003581373253837228, + 0.012776931747794151, + 0.007895039394497871, + 0.0038792185951024294, + 0.036257047206163406, + 0.005132443271577358, + 0.006990760564804077, + 0.003893687389791012, + 0.003449014388024807, + 0.00744427926838398, + 0.01311554666608572, + 0.044432129710912704, + 0.030516959726810455, + 0.004982383921742439, + 0.007817056961357594, + 0.013709686696529388, + 0.008689002133905888, + 0.008749864064157009, + 0.06404102593660355, + 0.02455732598900795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.7578254342079163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10523629188537598, + 0.14104735851287842, + 0.023886974900960922, + 0.05511041358113289, + 0.1267896145582199, + 0.08423318713903427, + 0.05134981870651245, + 0.016008291393518448, + 0.34007495641708374, + 0.05626308172941208 + ], + "credit_map": [ + 0.008009891957044601, + 0.006456288043409586, + 0.008780301548540592, + 0.013569456525146961, + 0.01357879675924778, + 0.022217420861124992, + 0.018147364258766174, + 0.014476778917014599, + 0.009198574349284172, + 0.010756208561360836, + 0.00889302883297205, + 0.02544538863003254, + 0.015112850815057755, + 0.025047697126865387, + 0.02732158824801445, + 0.019272036850452423, + 0.0023911595344543457, + 0.002160226460546255, + 0.0032211150974035263, + 0.0036100326105952263, + 0.0022007699590176344, + 0.00243597780354321, + 0.004754261113703251, + 0.00311343208886683, + 0.006192878354340792, + 0.0071573336608707905, + 0.004638770129531622, + 0.007933310233056545, + 0.0073771290481090546, + 0.0071757156401872635, + 0.00846690684556961, + 0.006168372929096222, + 0.017701711505651474, + 0.024852877482771873, + 0.011815233156085014, + 0.011395367793738842, + 0.013167528435587883, + 0.01333390548825264, + 0.014012597501277924, + 0.020510399714112282, + 0.007593765389174223, + 0.006159177049994469, + 0.00677321944385767, + 0.008511432446539402, + 0.023798217996954918, + 0.012015246786177158, + 0.009382829070091248, + 0.009999292902648449, + 0.003920799121260643, + 0.0019385326886549592, + 0.002506529912352562, + 0.0027396169025450945, + 0.0019998890347778797, + 0.0025432477705180645, + 0.01685667410492897, + 0.01884453184902668, + 0.0017389056738466024, + 0.0011632295791059732, + 0.0014256343711167574, + 0.0012539278250187635, + 0.001313324784860015, + 0.00487709091976285, + 0.003016852308064699, + 0.0012193266302347183, + 0.05450960248708725, + 0.008908511139452457, + 0.010335874743759632, + 0.013938561081886292, + 0.008926980197429657, + 0.012090898118913174, + 0.1023227795958519, + 0.12904173135757446, + 0.01341644860804081, + 0.0020538188982754946, + 0.0037724138237535954, + 0.005938970949500799, + 0.0035513064358383417, + 0.0032041347585618496, + 0.019390394911170006, + 0.004935593344271183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.012543476186692715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up depicts a minimalist chrome stapler, its polished body gleaming under a hard-key light, casting sharp, sharp shadows across the pristine surface of a clean white desk. The lighting meticulously sculpts the stapler's modern, geometric form, its mirror-like chrome finish adorned with brilliant mirror highlights. An unseen force presses the arm downward in a smooth, deliberate motion. Then, upon release, the arm returns upward in a single, swift, and decisive arc, exhibiting perfect critical decay as it instantly and perfectly comes to rest in its original resting position, devoid of any rebound or residual vibration. The entire action, set against a high-contrast, cool silver and white color palette, creates an atmosphere of refinement and focus, perfectly embodying the stylish aesthetic of corporate videos that emphasize precise engineering and cutting-edge modern design.", + "seed": 10259, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_10/index_4914_seed_10259_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16196041349321605, + "vlm_logprob": -0.16196041349321605, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.16196041349321605, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010586530901491642 + ], + "opposite_logprob_tokens": [ + -5.385586738586426 + ], + "logprob_reward": -0.010586530901491642, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010586530901491642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10583648085594177, + 0.030354492366313934, + 0.024383828043937683, + 0.0319397933781147, + 0.18292845785617828, + 0.12949632108211517, + 0.17187301814556122, + 0.033045217394828796, + 0.04472984001040459, + 0.24541255831718445 + ], + "credit_map": [ + 0.015594607219099998, + 0.010936292819678783, + 0.010961243882775307, + 0.010311744175851345, + 0.010586783289909363, + 0.024238184094429016, + 0.009499765932559967, + 0.013707856647670269, + 0.004557587672024965, + 0.005642744246870279, + 0.004512573592364788, + 0.0038447026163339615, + 0.0026191070210188627, + 0.0034826495684683323, + 0.002608703915029764, + 0.0030864241998642683, + 0.004453163594007492, + 0.004537563305348158, + 0.0027068955823779106, + 0.002320914063602686, + 0.0018930676160380244, + 0.003183037741109729, + 0.002327671507373452, + 0.002961514750495553, + 0.004748960956931114, + 0.004814560990780592, + 0.004744298756122589, + 0.003022017190232873, + 0.002782372757792473, + 0.0058810170739889145, + 0.002965242601931095, + 0.002981324912980199, + 0.016986634582281113, + 0.0683477446436882, + 0.06133643537759781, + 0.007508047856390476, + 0.011095097288489342, + 0.00655262591317296, + 0.0056731849908828735, + 0.005428678821772337, + 0.008198585361242294, + 0.0480463020503521, + 0.03819354996085167, + 0.006146931555122137, + 0.009216866455972195, + 0.010412745177745819, + 0.0045046135783195496, + 0.004776735324412584, + 0.010865956544876099, + 0.07012531161308289, + 0.05178103968501091, + 0.008060554973781109, + 0.013698695227503777, + 0.006296083331108093, + 0.004365235101431608, + 0.006680136546492577, + 0.0028162598609924316, + 0.007861586287617683, + 0.004797203466296196, + 0.0043408796191215515, + 0.0031237690709531307, + 0.0023024396505206823, + 0.003723453963175416, + 0.004079622216522694, + 0.004914965946227312, + 0.011002570390701294, + 0.011115211062133312, + 0.0032137392554432154, + 0.0037862476892769337, + 0.0045340959914028645, + 0.002965316642075777, + 0.0031976928003132343, + 0.014215761795639992, + 0.099532850086689, + 0.0924304649233818, + 0.008056450635194778, + 0.01377013698220253, + 0.007162193767726421, + 0.003795483149588108, + 0.00644923048093915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.010586530901491642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089714244008064 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089714244008064, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089714244008064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012208394706249237, + 0.009682049043476582, + 0.010092712938785553, + 0.024898711591959, + 0.5943860411643982, + 0.1520039141178131, + 0.06329954415559769, + 0.05203818529844284, + 0.04969407245516777, + 0.03169631212949753 + ], + "credit_map": [ + 0.001780712860636413, + 0.001613369444385171, + 0.0012083715992048383, + 0.0011521604610607028, + 0.0011071314802393317, + 0.0028799381107091904, + 0.0010017952881753445, + 0.0014649152290076017, + 0.0018273241585120559, + 0.0015864477027207613, + 0.0012535267742350698, + 0.0011005843989551067, + 0.0008549647754989564, + 0.0010923686204478145, + 0.0010831838008016348, + 0.0008836486958898604, + 0.0012676725164055824, + 0.0022246064618229866, + 0.0009917704155668616, + 0.0012706068810075521, + 0.0006281892419792712, + 0.0012322376715019345, + 0.000994444009847939, + 0.001483185333199799, + 0.003404785878956318, + 0.0043203444220125675, + 0.00392358610406518, + 0.002252874430269003, + 0.0025518168695271015, + 0.00401626992970705, + 0.0023931001778692007, + 0.0020359326153993607, + 0.04579576104879379, + 0.23194186389446259, + 0.20800118148326874, + 0.023915458470582962, + 0.03505747765302658, + 0.02014879323542118, + 0.014272019267082214, + 0.015253494493663311, + 0.007671065628528595, + 0.06112658977508545, + 0.055569838732481, + 0.006141616031527519, + 0.00845145434141159, + 0.00410492392256856, + 0.0038628890179097652, + 0.005075538530945778, + 0.0037942586932331324, + 0.02544616349041462, + 0.020290907472372055, + 0.003002972574904561, + 0.004608076997101307, + 0.001937953638844192, + 0.0018531647510826588, + 0.0023660489823669195, + 0.0030487379990518093, + 0.020869065076112747, + 0.01804141327738762, + 0.002109220949932933, + 0.0032337759621441364, + 0.0020304429344832897, + 0.0013678708346560597, + 0.0013376559363678098, + 0.003486485918983817, + 0.01950434222817421, + 0.015939559787511826, + 0.0021153108682483435, + 0.003456444013863802, + 0.0026523007545620203, + 0.0011252685217186809, + 0.001414362690411508, + 0.0020179576240479946, + 0.012553436681628227, + 0.011592133902013302, + 0.0011464032577350736, + 0.002051866613328457, + 0.0008324937080033123, + 0.0006541815237142146, + 0.0008478406234644353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.011089714244008064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209991320967674 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014209991320967674, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209991320967674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07680970430374146, + 0.039920005947351456, + 0.01878562569618225, + 0.03018847294151783, + 0.16844812035560608, + 0.35077232122421265, + 0.24956157803535461, + 0.03728504478931427, + 0.015359664335846901, + 0.012869534082710743 + ], + "credit_map": [ + 0.01282738521695137, + 0.009364761412143707, + 0.008092083036899567, + 0.0069226850755512714, + 0.0073371161706745625, + 0.0169360488653183, + 0.006186362821608782, + 0.009143264964222908, + 0.006705144420266151, + 0.006731074769049883, + 0.0048501938581466675, + 0.005129660479724407, + 0.0035220745485275984, + 0.004103121347725391, + 0.0053696408867836, + 0.0035090951714664698, + 0.0033022379502654076, + 0.0023378762416541576, + 0.0019605066627264023, + 0.002483826596289873, + 0.0015027476474642754, + 0.0026901692617684603, + 0.0017281918553635478, + 0.002780072158202529, + 0.003920495510101318, + 0.00581877026706934, + 0.004080508835613728, + 0.002604879206046462, + 0.004053381271660328, + 0.006059871055185795, + 0.0018468669150024652, + 0.0018036988331004977, + 0.009718178771436214, + 0.06772754341363907, + 0.06285647302865982, + 0.006489189807325602, + 0.009758823551237583, + 0.0048024337738752365, + 0.003536263480782509, + 0.0035592091735452414, + 0.018402548506855965, + 0.14203713834285736, + 0.1264822632074356, + 0.014595172367990017, + 0.020067643374204636, + 0.00894339568912983, + 0.008748926222324371, + 0.011495235376060009, + 0.01692526414990425, + 0.10463619977235794, + 0.06931246817111969, + 0.011209349147975445, + 0.022496214136481285, + 0.008757960051298141, + 0.00806213729083538, + 0.00816198531538248, + 0.0021260909270495176, + 0.014956687577068806, + 0.01347073819488287, + 0.0014980395790189505, + 0.002287940587848425, + 0.0007930394494906068, + 0.0008501706179231405, + 0.0013023394858464599, + 0.001005984959192574, + 0.0055063217878341675, + 0.005411567632108927, + 0.0007381196483038366, + 0.0009621083154343069, + 0.0005010042805224657, + 0.00048733613220974803, + 0.0007472216384485364, + 0.001197305042296648, + 0.004492776934057474, + 0.004438223782926798, + 0.0005931578925810754, + 0.0008037196821533144, + 0.0005388224381022155, + 0.0003300988464616239, + 0.0004754287947434932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.014209991320967674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578423023223877 + ], + "opposite_logprob_tokens": [ + -0.6328423023223877 + ], + "logprob_reward": -0.7578423023223877, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578423023223877, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05359813570976257, + 0.06565965712070465, + 0.03884999826550484, + 0.02786078117787838, + 0.16907961666584015, + 0.22060702741146088, + 0.20479169487953186, + 0.0375363864004612, + 0.09928236901760101, + 0.08273434638977051 + ], + "credit_map": [ + 0.009439773857593536, + 0.005963808856904507, + 0.006041715387254953, + 0.005956585053354502, + 0.005282730329781771, + 0.008349886164069176, + 0.004983519669622183, + 0.007580120116472244, + 0.010163676925003529, + 0.011052658781409264, + 0.007970071397721767, + 0.009538795799016953, + 0.0057115256786346436, + 0.006869977805763483, + 0.008146708831191063, + 0.006206246092915535, + 0.007851887494325638, + 0.008660810068249702, + 0.005626088473945856, + 0.0038564151618629694, + 0.0023018678184598684, + 0.003217449877411127, + 0.00410308875143528, + 0.003232392482459545, + 0.003402543719857931, + 0.00362911494448781, + 0.004078255500644445, + 0.0026482639368623495, + 0.0033192846458405256, + 0.00546202901750803, + 0.0029592858627438545, + 0.0023620035499334335, + 0.009556178003549576, + 0.06780708581209183, + 0.060744304209947586, + 0.00651208683848381, + 0.011035065166652203, + 0.005279191303998232, + 0.003988448064774275, + 0.004157268442213535, + 0.012357276864349842, + 0.08117377012968063, + 0.07405390590429306, + 0.012885315343737602, + 0.012545520439743996, + 0.007763997185975313, + 0.009147530421614647, + 0.010679707862436771, + 0.01538086961954832, + 0.084189772605896, + 0.0535559244453907, + 0.009704582393169403, + 0.019913800060749054, + 0.00827201921492815, + 0.0070296418853104115, + 0.0067450846545398235, + 0.0027018659748136997, + 0.011708223260939121, + 0.011869643814861774, + 0.0025946982204914093, + 0.002894223900511861, + 0.0015943149337545037, + 0.0019601413514465094, + 0.002213275758549571, + 0.005621541757136583, + 0.040519844740629196, + 0.03544054925441742, + 0.0038393642753362656, + 0.006295816041529179, + 0.0024180037435144186, + 0.0018525347113609314, + 0.0032947224099189043, + 0.00447592930868268, + 0.03336242586374283, + 0.03143804520368576, + 0.0029026593547314405, + 0.004535521846264601, + 0.0020411417353898287, + 0.001517083845101297, + 0.0024615421425551176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.7578423023223877 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist chrome stapler with a polished, mirror-like finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stapler positioned on a clean white desk surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stapler arm press downward in a smooth, deliberate motion and then return upward in a swift arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stapler arm come to rest instantly without any rebound or residual vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with hard-key light casting sharp shadows and featuring a cool silver and white color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073528677225113 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073528677225113, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073528677225113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07546573877334595, + 0.05842600017786026, + 0.021981604397296906, + 0.04382975772023201, + 0.09345455467700958, + 0.15282754600048065, + 0.13393595814704895, + 0.08549553900957108, + 0.010030342265963554, + 0.32455289363861084 + ], + "credit_map": [ + 0.011176987551152706, + 0.009889977052807808, + 0.007606076076626778, + 0.006956910714507103, + 0.007420822978019714, + 0.018297089263796806, + 0.006092301104217768, + 0.008025569841265678, + 0.009554049000144005, + 0.010399539023637772, + 0.007454290054738522, + 0.008045624941587448, + 0.005009369924664497, + 0.005969412624835968, + 0.006842060014605522, + 0.005151650868356228, + 0.0030906747560948133, + 0.003423580899834633, + 0.002814916893839836, + 0.0031503804493695498, + 0.0015316343633458018, + 0.002820637309923768, + 0.0027713763993233442, + 0.0023784032091498375, + 0.007237392943352461, + 0.01046841125935316, + 0.005880496930330992, + 0.0030306661501526833, + 0.005041102878749371, + 0.005232430528849363, + 0.003720051608979702, + 0.0032192058861255646, + 0.010651879943907261, + 0.02271423302590847, + 0.028290554881095886, + 0.007639910560101271, + 0.007474207319319248, + 0.004211253486573696, + 0.005189369898289442, + 0.007283149287104607, + 0.013238232582807541, + 0.0498819425702095, + 0.042586054652929306, + 0.01080339029431343, + 0.012378353625535965, + 0.010399309918284416, + 0.006014001090079546, + 0.007526261266320944, + 0.008177878335118294, + 0.05652892589569092, + 0.042117033153772354, + 0.005009816028177738, + 0.011201831512153149, + 0.004766740836203098, + 0.0030266749672591686, + 0.00310705229640007, + 0.004697995726019144, + 0.035006213933229446, + 0.02936422824859619, + 0.0033296942710876465, + 0.005691026337444782, + 0.002192452782765031, + 0.0020570531487464905, + 0.0031568771228194237, + 0.0011119197588413954, + 0.0018780715763568878, + 0.001326897880062461, + 0.0012051803059875965, + 0.0013999321963638067, + 0.0010177615331485868, + 0.0010756609262898564, + 0.0010149180889129639, + 0.018038371577858925, + 0.13387908041477203, + 0.12539125978946686, + 0.01040217000991106, + 0.017849452793598175, + 0.007130351848900318, + 0.004103989340364933, + 0.0077582155354321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.016073528677225113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_11/index_tensor([1373, 1373])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_11/index_tensor([1373, 1373])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e197fdee66585457b7c9da89166e2e2bcef71855 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_11/index_tensor([1373, 1373])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping aerial tracking shot, captured at a high frame rate to emphasize its exceptionally smooth motion, follows a majestic blended-wing-body airship as it executes a deliberate, low-altitude flight. This scene, rendered in a meticulously detailed artistic style, skillfully blends the ethereal romanticism of John William Waterhouse and the intricate fantasy of Vadim Kashin. The airship's slender, whale-like hull is a marvel of ornate metalwork, featuring fluid lines of polished brass and copper filigree crisscrossing its surface. Warm, golden hour light bathes the entire valley, shimmering the craft's hull as it glides silently and gracefully through delicate wisps of low-hanging cloud. Below, a tranquil river winds like a silver ribbon through an ancient forest, lushly wooded. The camera moves parallel to the airship, capturing the way it effortlessly navigates the valley's contours. Its massive shadow drifts across the deep green canopy, the leaves gently rustling, creating a tranquil, dreamlike majesty and a profound atmosphere of silent, epic journey.", + "seed": 10993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_11/index_1373_seed_10993_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03985515255481005, + "vlm_logprob": -0.03985515255481005, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03985515255481005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04783553630113602 + ], + "opposite_logprob_tokens": [ + -3.2978355884552 + ], + "logprob_reward": -0.04783553630113602, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04783553630113602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2481890469789505, + 0.1008823812007904, + 0.08439657837152481, + 0.12416987866163254, + 0.057777225971221924, + 0.044520169496536255, + 0.031232820823788643, + 0.07219275087118149, + 0.09491299092769623, + 0.14172616600990295 + ], + "credit_map": [ + 0.04917340725660324, + 0.02872323803603649, + 0.03221108764410019, + 0.023812657222151756, + 0.02215566858649254, + 0.014614671468734741, + 0.04000847786664963, + 0.03748983144760132, + 0.01617327891290188, + 0.013110162690281868, + 0.015419373288750648, + 0.012273373082280159, + 0.006442015524953604, + 0.010299356654286385, + 0.012790251523256302, + 0.014374571852385998, + 0.011240343563258648, + 0.011102862656116486, + 0.009882595390081406, + 0.011511252261698246, + 0.006595819257199764, + 0.0076759071089327335, + 0.014795586466789246, + 0.011592213995754719, + 0.010540230199694633, + 0.019808080047369003, + 0.0149609399959445, + 0.01594686135649681, + 0.013598289340734482, + 0.015156656503677368, + 0.01860983483493328, + 0.015548987314105034, + 0.004397210199385881, + 0.00474265031516552, + 0.004440292716026306, + 0.012743104249238968, + 0.0027810069732367992, + 0.00380702200345695, + 0.007280735764652491, + 0.017585203051567078, + 0.0027384930290281773, + 0.002369898371398449, + 0.004528297111392021, + 0.010568806901574135, + 0.002497695852071047, + 0.006930376868695021, + 0.004734271205961704, + 0.010152332484722137, + 0.0024732837919145823, + 0.002108106156811118, + 0.0027983675245195627, + 0.0044225482270121574, + 0.001757812686264515, + 0.008107343688607216, + 0.00575236976146698, + 0.00381299015134573, + 0.006174623034894466, + 0.008242334239184856, + 0.006318292580544949, + 0.008214552886784077, + 0.0063348691910505295, + 0.016162144020199776, + 0.008483845740556717, + 0.012262081727385521, + 0.007586807478219271, + 0.006156093440949917, + 0.010947426781058311, + 0.016505584120750427, + 0.01084897667169571, + 0.012997068464756012, + 0.017288191244006157, + 0.012582845985889435, + 0.016443297266960144, + 0.01681743748486042, + 0.014782587066292763, + 0.01975562609732151, + 0.012993131764233112, + 0.010254912078380585, + 0.022525766864418983, + 0.02815341390669346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.04783553630113602 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209791451692581 + ], + "opposite_logprob_tokens": [ + -2.8120980262756348 + ], + "logprob_reward": -0.06209791451692581, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209791451692581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.186501145362854, + 0.05069062486290932, + 0.049638617783784866, + 0.01815680041909218, + 0.06666125357151031, + 0.042857348918914795, + 0.12808270752429962, + 0.25928249955177307, + 0.0869838148355484, + 0.11114514619112015 + ], + "credit_map": [ + 0.047321636229753494, + 0.023183118551969528, + 0.019737087190151215, + 0.010151486843824387, + 0.009937796741724014, + 0.008061054162681103, + 0.03755638748407364, + 0.03055257722735405, + 0.008766638115048409, + 0.0049720932729542255, + 0.0073295049369335175, + 0.004813048057258129, + 0.004683664999902248, + 0.0074183023534715176, + 0.006501758471131325, + 0.006205613724887371, + 0.005342843942344189, + 0.008802172727882862, + 0.004791270475834608, + 0.0070992084220051765, + 0.003592392895370722, + 0.00369966309517622, + 0.008862434886395931, + 0.007448630407452583, + 0.0019491651328280568, + 0.002676525851711631, + 0.001887794118374586, + 0.003294546389952302, + 0.0018294017063453794, + 0.002042359672486782, + 0.002619211794808507, + 0.0018577943556010723, + 0.006612193305045366, + 0.0075802854262292385, + 0.00813351757824421, + 0.013784322887659073, + 0.005355347879230976, + 0.006882138084620237, + 0.00798753835260868, + 0.010325909592211246, + 0.0030432031489908695, + 0.001716730883345008, + 0.0028456321451812983, + 0.010425586253404617, + 0.002625297522172332, + 0.005980633664876223, + 0.005091002676635981, + 0.011129261925816536, + 0.0067330882884562016, + 0.008424248546361923, + 0.00975092127919197, + 0.01885542832314968, + 0.008959532715380192, + 0.03697329759597778, + 0.023281989619135857, + 0.015104193240404129, + 0.021918466314673424, + 0.030506938695907593, + 0.01939951255917549, + 0.029631292447447777, + 0.020219432190060616, + 0.060987453907728195, + 0.030978992581367493, + 0.04564039409160614, + 0.00779730686917901, + 0.0075724064372479916, + 0.009880084544420242, + 0.013771316967904568, + 0.006895990110933781, + 0.013732860796153545, + 0.013748823665082455, + 0.013585025444626808, + 0.015472923405468464, + 0.014439152553677559, + 0.010382363572716713, + 0.01769922487437725, + 0.009551208466291428, + 0.008959303610026836, + 0.01826995052397251, + 0.016371019184589386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.06209791451692581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006071507930756 + ], + "opposite_logprob_tokens": [ + -2.6950607299804688 + ], + "logprob_reward": -0.07006071507930756, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006071507930756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1602960079908371, + 0.05092117190361023, + 0.1121223047375679, + 0.06571941822767258, + 0.31775957345962524, + 0.060541607439517975, + 0.0693894475698471, + 0.06829018145799637, + 0.06617336720228195, + 0.02878689020872116 + ], + "credit_map": [ + 0.03743718937039375, + 0.019226152449846268, + 0.022754080593585968, + 0.01142151840031147, + 0.009454955346882343, + 0.007370679173618555, + 0.027377843856811523, + 0.02525360696017742, + 0.008144733496010303, + 0.006857550237327814, + 0.006067181006073952, + 0.004665921442210674, + 0.0036834976635873318, + 0.005431820638477802, + 0.00793763343244791, + 0.00813283771276474, + 0.011928369291126728, + 0.018010037019848824, + 0.011934017762541771, + 0.019893012940883636, + 0.008518727496266365, + 0.009474997408688068, + 0.01852625235915184, + 0.013836899772286415, + 0.006920547224581242, + 0.008969241753220558, + 0.008867599070072174, + 0.008370199240744114, + 0.005894917529076338, + 0.006142811384052038, + 0.009507748298346996, + 0.011046362109482288, + 0.012847243808209896, + 0.011269329115748405, + 0.014128712937235832, + 0.1087513118982315, + 0.00794932246208191, + 0.024774398654699326, + 0.026132360100746155, + 0.11190691590309143, + 0.0034157971385866404, + 0.003031821921467781, + 0.004500788636505604, + 0.01747201941907406, + 0.0019249620381742716, + 0.006104529369622469, + 0.007400535978376865, + 0.01669115386903286, + 0.004342231433838606, + 0.008482135832309723, + 0.007666693069040775, + 0.00888133142143488, + 0.005322515964508057, + 0.016629718244075775, + 0.011297418735921383, + 0.006767412647604942, + 0.004704936407506466, + 0.007135019171983004, + 0.005084066651761532, + 0.007678038906306028, + 0.006267567165195942, + 0.018155081197619438, + 0.008434596471488476, + 0.010830882005393505, + 0.005258155055344105, + 0.0042769648134708405, + 0.008208216167986393, + 0.011829259805381298, + 0.008773413486778736, + 0.007321209646761417, + 0.01205302495509386, + 0.008453127928078175, + 0.0024037309922277927, + 0.003462179098278284, + 0.004250547848641872, + 0.004930614959448576, + 0.0031846032943576574, + 0.002708838554099202, + 0.0041200630366802216, + 0.003726312192156911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.07006071507930756 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067444778978824615 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.0067444778978824615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067444778978824615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21760839223861694, + 0.10629421472549438, + 0.1096796914935112, + 0.0628201961517334, + 0.18819789588451385, + 0.05534263327717781, + 0.07341239601373672, + 0.05423770472407341, + 0.09381736069917679, + 0.03858950734138489 + ], + "credit_map": [ + 0.05404434725642204, + 0.029013218358159065, + 0.019378118216991425, + 0.014090219512581825, + 0.010945755057036877, + 0.011093845590949059, + 0.04004202410578728, + 0.03900083899497986, + 0.018587101250886917, + 0.01309192180633545, + 0.015062068589031696, + 0.011155504733324051, + 0.0074363877065479755, + 0.011020205914974213, + 0.013637888245284557, + 0.016303129494190216, + 0.01161932572722435, + 0.015658626332879066, + 0.011888522654771805, + 0.017855655401945114, + 0.0068334671668708324, + 0.00653629656881094, + 0.020168572664260864, + 0.01911921612918377, + 0.0058525679633021355, + 0.008953501470386982, + 0.009013358503580093, + 0.007550447713583708, + 0.006296093575656414, + 0.006992891896516085, + 0.00967112835496664, + 0.008490204811096191, + 0.008678787387907505, + 0.006915451027452946, + 0.007098955102264881, + 0.061665624380111694, + 0.0029824809171259403, + 0.01270405575633049, + 0.015042196027934551, + 0.07311033457517624, + 0.0031429713126271963, + 0.002957248827442527, + 0.004623370710760355, + 0.015438875183463097, + 0.002453893655911088, + 0.005437769927084446, + 0.006387327332049608, + 0.014901170507073402, + 0.005230593029409647, + 0.00881911814212799, + 0.00843428261578083, + 0.009387279860675335, + 0.005503006279468536, + 0.017352892085909843, + 0.011479271575808525, + 0.00720595009624958, + 0.0043436684645712376, + 0.005472224205732346, + 0.004499412141740322, + 0.006463638972491026, + 0.0038182430434972048, + 0.014585180208086967, + 0.005800225306302309, + 0.00925510935485363, + 0.008533160202205181, + 0.006920333951711655, + 0.011740553192794323, + 0.016791140660643578, + 0.01078331284224987, + 0.010748906061053276, + 0.015703069046139717, + 0.01259688101708889, + 0.0032047496642917395, + 0.0033734545577317476, + 0.005028669722378254, + 0.007291221059858799, + 0.004102697130292654, + 0.0035007409751415253, + 0.0066992915235459805, + 0.005388683173805475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0067444778978824615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253711897879839 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253711897879839, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253711897879839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16740106046199799, + 0.1028241291642189, + 0.09458887577056885, + 0.08348831534385681, + 0.1429697871208191, + 0.11619546264410019, + 0.05860304459929466, + 0.04184688255190849, + 0.1339285522699356, + 0.058153897523880005 + ], + "credit_map": [ + 0.03268275409936905, + 0.023332297801971436, + 0.024698438122868538, + 0.012308953329920769, + 0.010637503117322922, + 0.010149632580578327, + 0.02692151442170143, + 0.026669956743717194, + 0.015454130247235298, + 0.015993865206837654, + 0.013617041520774364, + 0.008514851331710815, + 0.007832001894712448, + 0.007957248017191887, + 0.01969217322766781, + 0.013762809336185455, + 0.011065858416259289, + 0.013828249648213387, + 0.010221264325082302, + 0.016290578991174698, + 0.005858734715729952, + 0.006415230222046375, + 0.015996357426047325, + 0.014912597835063934, + 0.007216591387987137, + 0.011356671340763569, + 0.011450416408479214, + 0.012260758318006992, + 0.006610903423279524, + 0.008776955306529999, + 0.013961301185190678, + 0.011854720301926136, + 0.007966426201164722, + 0.008936852216720581, + 0.007612295914441347, + 0.03123796172440052, + 0.008251061663031578, + 0.012384247034788132, + 0.019578220322728157, + 0.04700271785259247, + 0.007650366052985191, + 0.00669716764241457, + 0.00899219885468483, + 0.033599015325307846, + 0.005167101509869099, + 0.014428969472646713, + 0.009303023107349873, + 0.030357614159584045, + 0.0042151049710810184, + 0.008371681906282902, + 0.0059868390671908855, + 0.0074960775673389435, + 0.003206789493560791, + 0.013546506874263287, + 0.009775356389582157, + 0.0060046822763979435, + 0.0036960411816835403, + 0.004714139737188816, + 0.004268819931894541, + 0.0038368036039173603, + 0.0050493585877120495, + 0.010570909827947617, + 0.0049371314235031605, + 0.004773674998432398, + 0.012035669758915901, + 0.01043465081602335, + 0.01743202470242977, + 0.021262580528855324, + 0.016482949256896973, + 0.015618433244526386, + 0.020411334931850433, + 0.020250897854566574, + 0.00610760785639286, + 0.00776698300614953, + 0.004271546378731728, + 0.006065762136131525, + 0.004704245366156101, + 0.008573255501687527, + 0.008437425829470158, + 0.012227067723870277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.01253711897879839 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial tracking shot, captured at a high frame rate to emphasize its exceptionally smooth motion, follows a majestic blended-wing-body airship as it executes a deliberate, low-altitude flight. This scene, rendered in a meticulously detailed artistic style, skillfully blends the ethereal romanticism of John William Waterhouse and the intricate fantasy of Vadim Kashin. The airship's slender, whale-like hull is a marvel of ornate metalwork, featuring fluid lines of polished brass and copper filigree crisscrossing its surface. Warm, golden hour light bathes the entire valley, shimmering the craft's hull as it glides silently and gracefully through delicate wisps of low-hanging cloud. Below, a tranquil river winds like a silver ribbon through an ancient forest, lushly wooded. The camera moves parallel to the airship, capturing the way it effortlessly navigates the valley's contours. Its massive shadow drifts across the deep green canopy, the leaves gently rustling, creating a tranquil, dreamlike majesty and a profound atmosphere of silent, epic journey.", + "seed": 10994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_11/index_1373_seed_10994_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.057832810934633015, + "vlm_logprob": -0.057832810934633015, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.057832810934633015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05453965440392494 + ], + "opposite_logprob_tokens": [ + -3.054539680480957 + ], + "logprob_reward": -0.05453965440392494, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05453965440392494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.141246035695076, + 0.17116667330265045, + 0.0515492707490921, + 0.07460363209247589, + 0.0670202299952507, + 0.03091317042708397, + 0.057383835315704346, + 0.05256279930472374, + 0.05870906263589859, + 0.2948452830314636 + ], + "credit_map": [ + 0.027180319651961327, + 0.01918565109372139, + 0.018602337688207626, + 0.015962613746523857, + 0.019473617896437645, + 0.021857840940356255, + 0.010059273801743984, + 0.008924401365220547, + 0.018869612365961075, + 0.02725936658680439, + 0.03304944932460785, + 0.007286402862519026, + 0.04135783016681671, + 0.018008051440119743, + 0.008930571377277374, + 0.016405407339334488, + 0.00815446488559246, + 0.00801505334675312, + 0.006380363833159208, + 0.006381981074810028, + 0.006268468219786882, + 0.005549405701458454, + 0.0046058413572609425, + 0.006193696521222591, + 0.009486984461545944, + 0.01252345647662878, + 0.00874603446573019, + 0.007320948410779238, + 0.010878387838602066, + 0.00909582432359457, + 0.007520990911871195, + 0.00903100986033678, + 0.005553371738642454, + 0.0090182451531291, + 0.009789572097361088, + 0.00415034731850028, + 0.008036641404032707, + 0.01007598452270031, + 0.0095134899020195, + 0.010882587172091007, + 0.003462257795035839, + 0.003952193073928356, + 0.0050516510382294655, + 0.003901986638084054, + 0.0036786231212317944, + 0.003078540088608861, + 0.002829988719895482, + 0.0049579329788684845, + 0.00565582187846303, + 0.005699084606021643, + 0.010184929706156254, + 0.009206990711390972, + 0.006815010216087103, + 0.004998393822461367, + 0.007231567520648241, + 0.007592038717120886, + 0.004198380280286074, + 0.006831157952547073, + 0.011201412416994572, + 0.0076048667542636395, + 0.005059311166405678, + 0.006518114358186722, + 0.006012541241943836, + 0.005137020722031593, + 0.007309755776077509, + 0.0067903706803917885, + 0.006926697678864002, + 0.008613551035523415, + 0.00958534236997366, + 0.007620071992278099, + 0.007634312845766544, + 0.004228962119668722, + 0.0176792461425066, + 0.020086655393242836, + 0.020540814846754074, + 0.07960886508226395, + 0.021991996094584465, + 0.030171938240528107, + 0.07093044370412827, + 0.03383534401655197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.05453965440392494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298547685146332 + ], + "opposite_logprob_tokens": [ + -2.237985372543335 + ], + "logprob_reward": -0.11298547685146332, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298547685146332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08686760067939758, + 0.20667517185211182, + 0.09412248432636261, + 0.05414929240942001, + 0.04355527088046074, + 0.05602230504155159, + 0.025680890306830406, + 0.057975176721811295, + 0.06591679900884628, + 0.3090350329875946 + ], + "credit_map": [ + 0.016170190647244453, + 0.011492299847304821, + 0.009025681763887405, + 0.009274509735405445, + 0.010550864972174168, + 0.01324451807886362, + 0.0069891843013465405, + 0.010120359249413013, + 0.021656015887856483, + 0.03265484794974327, + 0.04030395299196243, + 0.009358430281281471, + 0.04944533854722977, + 0.022008497267961502, + 0.010475670918822289, + 0.020772403106093407, + 0.014344635419547558, + 0.012685955502092838, + 0.0055242376402020454, + 0.004900808911770582, + 0.030499689280986786, + 0.009139818139374256, + 0.006166569422930479, + 0.010860768146812916, + 0.006274158135056496, + 0.00819444376975298, + 0.007625435013324022, + 0.006267607677727938, + 0.00763077475130558, + 0.006040217820554972, + 0.005396363325417042, + 0.006720293778926134, + 0.0036270227283239365, + 0.006543173920363188, + 0.0059180534444749355, + 0.002722895937040448, + 0.004760707728564739, + 0.006971118971705437, + 0.005770910065621138, + 0.007241389714181423, + 0.006918590050190687, + 0.006795243825763464, + 0.008799831382930279, + 0.007028436288237572, + 0.0064440337009727955, + 0.005837688222527504, + 0.005076540168374777, + 0.009121941402554512, + 0.002195844892412424, + 0.002812274731695652, + 0.003903561970219016, + 0.003769227070733905, + 0.002098510041832924, + 0.0033150918316096067, + 0.0033172289840877056, + 0.004269150085747242, + 0.0035470465663820505, + 0.007330277469009161, + 0.006740989163517952, + 0.011294479481875896, + 0.011840381659567356, + 0.006917842198163271, + 0.005719728302210569, + 0.004584433976560831, + 0.003968029748648405, + 0.00333989760838449, + 0.004413960967212915, + 0.02444637380540371, + 0.002437789458781481, + 0.00194180803373456, + 0.022404467687010765, + 0.002964471699669957, + 0.013597870245575905, + 0.013129794970154762, + 0.02457311749458313, + 0.0966232419013977, + 0.018370434641838074, + 0.023453982546925545, + 0.08352328836917877, + 0.03576330095529556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.11298547685146332 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031754523515701 + ], + "opposite_logprob_tokens": [ + -2.3503174781799316 + ], + "logprob_reward": -0.10031754523515701, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031754523515701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1560739278793335, + 0.09765978157520294, + 0.1901559978723526, + 0.20018485188484192, + 0.06134895235300064, + 0.02129102312028408, + 0.033820416778326035, + 0.031454890966415405, + 0.08020750433206558, + 0.12780264019966125 + ], + "credit_map": [ + 0.013478565029799938, + 0.017327791079878807, + 0.01396522019058466, + 0.008914757519960403, + 0.04747868701815605, + 0.025449061766266823, + 0.011426004581153393, + 0.018033834174275398, + 0.01469592284411192, + 0.01975218392908573, + 0.009979866445064545, + 0.006498456001281738, + 0.015893928706645966, + 0.014198577962815762, + 0.006827289704233408, + 0.009813548997044563, + 0.02742922678589821, + 0.016265131533145905, + 0.025556335225701332, + 0.010963110253214836, + 0.042324189096689224, + 0.029032666236162186, + 0.010954118333756924, + 0.027631204575300217, + 0.03178594261407852, + 0.04061780497431755, + 0.020379213616251945, + 0.01827467978000641, + 0.02063567005097866, + 0.02221217192709446, + 0.017366446554660797, + 0.02891290932893753, + 0.00607668049633503, + 0.00863279215991497, + 0.010183095000684261, + 0.004387660883367062, + 0.006998364347964525, + 0.007185192313045263, + 0.007149239536374807, + 0.010735927149653435, + 0.0025446168147027493, + 0.0020770817063748837, + 0.0033428380265831947, + 0.002894139150157571, + 0.0026243969332426786, + 0.0023150588385760784, + 0.002018809085711837, + 0.0034740823321044445, + 0.004540291614830494, + 0.0030787198338657618, + 0.005799636244773865, + 0.0053559704683721066, + 0.0035695803817361593, + 0.0034681514371186495, + 0.0032412277068942785, + 0.004766838625073433, + 0.0021118270233273506, + 0.0037429137155413628, + 0.004798888694494963, + 0.005250413902103901, + 0.004980884958058596, + 0.004848500248044729, + 0.0028803932946175337, + 0.0028410714585334063, + 0.004704356659203768, + 0.004645261447876692, + 0.004252861253917217, + 0.024147897958755493, + 0.008834769017994404, + 0.008503638207912445, + 0.021419258788228035, + 0.0036994563415646553, + 0.006211007479578257, + 0.005148184020072222, + 0.007692038547247648, + 0.04366886243224144, + 0.005926100071519613, + 0.007473302539438009, + 0.041832178831100464, + 0.009850950911641121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.10031754523515701 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257233511656523 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005257233511656523, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257233511656523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3575645089149475, + 0.09016214311122894, + 0.039750341325998306, + 0.09311209619045258, + 0.07437040656805038, + 0.026128342375159264, + 0.04932721331715584, + 0.03852672502398491, + 0.10070425271987915, + 0.1303539127111435 + ], + "credit_map": [ + 0.04088893532752991, + 0.04205141216516495, + 0.027367526665329933, + 0.03448807820677757, + 0.07600049674510956, + 0.062431950122117996, + 0.030780043452978134, + 0.04355606809258461, + 0.013000601902604103, + 0.017320580780506134, + 0.01188497245311737, + 0.006282391957938671, + 0.015305937267839909, + 0.013343524187803268, + 0.0056114811450243, + 0.007412657607346773, + 0.0039792489260435104, + 0.006273641251027584, + 0.004411614034324884, + 0.0038918654900044203, + 0.006055104546248913, + 0.005794926546514034, + 0.0030940535943955183, + 0.006249884143471718, + 0.012922639958560467, + 0.020160850137472153, + 0.010921633802354336, + 0.007710118778049946, + 0.009888864122331142, + 0.010508120991289616, + 0.008698271587491035, + 0.012301597744226456, + 0.005992305930703878, + 0.010184894315898418, + 0.01132891420274973, + 0.004110684618353844, + 0.007855434902012348, + 0.0103292902931571, + 0.01042147446423769, + 0.014147402718663216, + 0.003308733692392707, + 0.0032840559724718332, + 0.004311345983296633, + 0.0031880110036581755, + 0.002473583910614252, + 0.002955961972475052, + 0.002177298069000244, + 0.004429350607097149, + 0.006810409482568502, + 0.00553448311984539, + 0.008177964016795158, + 0.007730221375823021, + 0.006083559710532427, + 0.004104113206267357, + 0.005577409639954567, + 0.005309055093675852, + 0.002985841128975153, + 0.005115344654768705, + 0.005552218295633793, + 0.006387708242982626, + 0.004929430317133665, + 0.005303121171891689, + 0.004538099281489849, + 0.003714960068464279, + 0.005478927865624428, + 0.008535726927220821, + 0.005485800094902515, + 0.03191053494811058, + 0.009848314337432384, + 0.007260424550622702, + 0.02820618450641632, + 0.003978343214839697, + 0.005497650243341923, + 0.005063564982265234, + 0.008181227371096611, + 0.046361371874809265, + 0.005052601918578148, + 0.006517408881336451, + 0.04336581379175186, + 0.010314266197383404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.005257233511656523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064144670963287 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064144670963287, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064144670963287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22525347769260406, + 0.09319118410348892, + 0.08429200947284698, + 0.15510718524456024, + 0.04581999033689499, + 0.06868748366832733, + 0.07317433506250381, + 0.04802872613072395, + 0.09982682019472122, + 0.10661876201629639 + ], + "credit_map": [ + 0.031093986704945564, + 0.024471139535307884, + 0.023490270599722862, + 0.021888405084609985, + 0.0485527478158474, + 0.0407884418964386, + 0.01796049252152443, + 0.017008008435368538, + 0.007986387237906456, + 0.010760652832686901, + 0.018369846045970917, + 0.003883750643581152, + 0.02624673768877983, + 0.010924863629043102, + 0.004789450671523809, + 0.010229498147964478, + 0.013878845609724522, + 0.009495420381426811, + 0.008521637879312038, + 0.004704602994024754, + 0.018408624455332756, + 0.009971940889954567, + 0.005499652121216059, + 0.013811295852065086, + 0.021603798493742943, + 0.029683595523238182, + 0.017878733575344086, + 0.014694364741444588, + 0.016488298773765564, + 0.013348394073545933, + 0.01380183082073927, + 0.027608169242739677, + 0.003265232080593705, + 0.007908280938863754, + 0.00745870266109705, + 0.0031907446682453156, + 0.00628140801563859, + 0.0043860976584255695, + 0.005065968260169029, + 0.00826355442404747, + 0.009346602484583855, + 0.007449181750416756, + 0.011520106345415115, + 0.008919108659029007, + 0.008383656851947308, + 0.005189104937016964, + 0.006537841632962227, + 0.011341884732246399, + 0.006306210067123175, + 0.006148487329483032, + 0.011869688518345356, + 0.013931644149124622, + 0.009000501595437527, + 0.008449980989098549, + 0.008244771510362625, + 0.009223049506545067, + 0.0036247922107577324, + 0.004902992397546768, + 0.00822717696428299, + 0.007271855603903532, + 0.004338210914283991, + 0.009387511759996414, + 0.005404770374298096, + 0.0048714191652834415, + 0.0047720372676849365, + 0.004959875717759132, + 0.008930092677474022, + 0.038058698177337646, + 0.003958460409194231, + 0.003330052364617586, + 0.03128337487578392, + 0.004534224048256874, + 0.007033814210444689, + 0.005498336628079414, + 0.0057777767069637775, + 0.03061056137084961, + 0.006120278965681791, + 0.007873893715441227, + 0.030918538570404053, + 0.012785562314093113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.016064144670963287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial tracking shot, captured at a high frame rate to emphasize its exceptionally smooth motion, follows a majestic blended-wing-body airship as it executes a deliberate, low-altitude flight. This scene, rendered in a meticulously detailed artistic style, skillfully blends the ethereal romanticism of John William Waterhouse and the intricate fantasy of Vadim Kashin. The airship's slender, whale-like hull is a marvel of ornate metalwork, featuring fluid lines of polished brass and copper filigree crisscrossing its surface. Warm, golden hour light bathes the entire valley, shimmering the craft's hull as it glides silently and gracefully through delicate wisps of low-hanging cloud. Below, a tranquil river winds like a silver ribbon through an ancient forest, lushly wooded. The camera moves parallel to the airship, capturing the way it effortlessly navigates the valley's contours. Its massive shadow drifts across the deep green canopy, the leaves gently rustling, creating a tranquil, dreamlike majesty and a profound atmosphere of silent, epic journey.", + "seed": 10995, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_11/index_1373_seed_10995_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03383000828325748, + "vlm_logprob": -0.03383000828325748, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03383000828325748, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03002161532640457 + ], + "opposite_logprob_tokens": [ + -3.7800216674804688 + ], + "logprob_reward": -0.03002161532640457, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03002161532640457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15745127201080322, + 0.08754193037748337, + 0.039222076535224915, + 0.02930372953414917, + 0.10392139852046967, + 0.024400554597377777, + 0.08112253993749619, + 0.03989668935537338, + 0.07136522978544235, + 0.36577460169792175 + ], + "credit_map": [ + 0.012168940156698227, + 0.02500123158097267, + 0.021221183240413666, + 0.01781955547630787, + 0.01490259263664484, + 0.021273916587233543, + 0.018901968374848366, + 0.02616187557578087, + 0.005888671614229679, + 0.006523068994283676, + 0.010600147768855095, + 0.012807524763047695, + 0.01397307775914669, + 0.008831758983433247, + 0.01580708473920822, + 0.01311060506850481, + 0.002913396805524826, + 0.0029213663656264544, + 0.003029049141332507, + 0.003839687677100301, + 0.009159669280052185, + 0.0030237510800361633, + 0.008971886709332466, + 0.0053632729686796665, + 0.0021544862538576126, + 0.004361233673989773, + 0.0028888280503451824, + 0.003226351924240589, + 0.003316867630928755, + 0.005247276276350021, + 0.004397266078740358, + 0.003711419180035591, + 0.008275127038359642, + 0.016299637034535408, + 0.00861489400267601, + 0.008972231298685074, + 0.021150939166545868, + 0.0166794303804636, + 0.010355312377214432, + 0.013573821634054184, + 0.0019551622681319714, + 0.0018705606926232576, + 0.0028080795891582966, + 0.0030890516936779022, + 0.0027563285548239946, + 0.003465198213234544, + 0.004827555269002914, + 0.00362861854955554, + 0.0087583614513278, + 0.0053560067899525166, + 0.016679655760526657, + 0.009756677784025669, + 0.006122833117842674, + 0.00846164207905531, + 0.00820019468665123, + 0.017787165939807892, + 0.005507363472133875, + 0.0029815302696079016, + 0.006348006892949343, + 0.005939871072769165, + 0.002356540644541383, + 0.004439600743353367, + 0.004412158392369747, + 0.007911616005003452, + 0.009180355817079544, + 0.003454149467870593, + 0.012320955283939838, + 0.01768694445490837, + 0.004061788320541382, + 0.008063525892794132, + 0.006703807506710291, + 0.009893699549138546, + 0.010747839696705341, + 0.0031019728630781174, + 0.008677741512656212, + 0.010573523119091988, + 0.007916547358036041, + 0.010875540785491467, + 0.16071711480617523, + 0.1531643569469452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.03002161532640457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620826780796051 + ], + "opposite_logprob_tokens": [ + -2.8120827674865723 + ], + "logprob_reward": -0.0620826780796051, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620826780796051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08716689795255661, + 0.03763153776526451, + 0.02956431731581688, + 0.010055799968540668, + 0.05569784715771675, + 0.0138406278565526, + 0.023494649678468704, + 0.010122097097337246, + 0.017330080270767212, + 0.715096116065979 + ], + "credit_map": [ + 0.005031344015151262, + 0.00425688223913312, + 0.01927141658961773, + 0.010223687626421452, + 0.00601424602791667, + 0.008754135109484196, + 0.005780236329883337, + 0.027834955602884293, + 0.002823916496708989, + 0.0024301952216774225, + 0.00432556914165616, + 0.006172407418489456, + 0.0063873822800815105, + 0.003593829460442066, + 0.0075807757675647736, + 0.004317460581660271, + 0.002169138751924038, + 0.0021399040706455708, + 0.0028389038052409887, + 0.0032528915908187628, + 0.0056584542617201805, + 0.001901838113553822, + 0.007406225427985191, + 0.004196961410343647, + 0.0008633903926238418, + 0.0020508698653429747, + 0.0010545275872573256, + 0.000967770756687969, + 0.0011168094351887703, + 0.0011379079660400748, + 0.0014944669092074037, + 0.0013700573472306132, + 0.0031239232048392296, + 0.004867821466177702, + 0.01917402818799019, + 0.006505338475108147, + 0.00236516329459846, + 0.0036092589143663645, + 0.003079885384067893, + 0.012972429394721985, + 0.0014467675937339664, + 0.001477652695029974, + 0.0015921745216473937, + 0.001747588161379099, + 0.0015502552269026637, + 0.0018619325710460544, + 0.0022786306217312813, + 0.0018856266979128122, + 0.0022693746723234653, + 0.0012944104382768273, + 0.006010607350617647, + 0.0024757820647209883, + 0.0014231610111892223, + 0.0019859373569488525, + 0.0020285167265683413, + 0.006006859242916107, + 0.0015594282886013389, + 0.0006890432559885085, + 0.0017904703272506595, + 0.0014690581010654569, + 0.0005687414086423814, + 0.0008909635944291949, + 0.0009749293676577508, + 0.002179462229833007, + 0.002673651557415724, + 0.0013450960395857692, + 0.0023342687636613846, + 0.0026307604275643826, + 0.0016288419719785452, + 0.0025254180654883385, + 0.0016772696981206536, + 0.002514774212613702, + 0.02086818404495716, + 0.005567372310906649, + 0.01725654862821102, + 0.02148839272558689, + 0.015292936936020851, + 0.021030694246292114, + 0.3147761821746826, + 0.2988158166408539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0620826780796051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865432530641556 + ], + "opposite_logprob_tokens": [ + -3.048654317855835 + ], + "logprob_reward": -0.04865432530641556, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865432530641556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03630201518535614, + 0.023329878225922585, + 0.011058586649596691, + 0.007709982804954052, + 0.02491716854274273, + 0.01933169737458229, + 0.05965592712163925, + 0.05138128995895386, + 0.015601511113345623, + 0.7507119178771973 + ], + "credit_map": [ + 0.0022395681589841843, + 0.003178390208631754, + 0.004529210738837719, + 0.004686916712671518, + 0.003907063975930214, + 0.004990854766219854, + 0.005383448675274849, + 0.007386558689177036, + 0.0020067093428224325, + 0.0016894156578928232, + 0.002417169511318207, + 0.003056113375350833, + 0.0034908135421574116, + 0.00243153003975749, + 0.004358829464763403, + 0.0038792979903519154, + 0.0011848736321553588, + 0.0009826396126300097, + 0.0008672424010001123, + 0.0014108478790149093, + 0.0017697068396955729, + 0.0014066245639696717, + 0.0020346189849078655, + 0.001402032794430852, + 0.0007228743634186685, + 0.00109397666528821, + 0.0008650451782159507, + 0.0008152342634275556, + 0.0014357343316078186, + 0.0008896943763829768, + 0.0008368359412997961, + 0.0010505885584279895, + 0.001425804803147912, + 0.002864452777430415, + 0.007160713896155357, + 0.0027166178915649652, + 0.0017162268050014973, + 0.002592684468254447, + 0.0016674777725711465, + 0.004773189779371023, + 0.002814667299389839, + 0.0012496142880991101, + 0.0017774614971131086, + 0.0029961925465613604, + 0.00199912185780704, + 0.0029368645045906305, + 0.0028847602661699057, + 0.002673014998435974, + 0.0060608056373894215, + 0.0026583841536194086, + 0.016826428472995758, + 0.005184508860111237, + 0.002537264721468091, + 0.004158685449510813, + 0.00417190371081233, + 0.01805794984102249, + 0.0065653217025101185, + 0.004055693279951811, + 0.008465253747999668, + 0.0074453498236835, + 0.0027719223871827126, + 0.0040154969319701195, + 0.006112932693213224, + 0.011949318461120129, + 0.0019344598986208439, + 0.0006426787003874779, + 0.0038605653680860996, + 0.0030054852832108736, + 0.000655030773486942, + 0.0010200521210208535, + 0.0010508217383176088, + 0.0034324179869145155, + 0.02091849595308304, + 0.005480748135596514, + 0.017649061977863312, + 0.02162528783082962, + 0.01604349911212921, + 0.02126976102590561, + 0.3317604064941406, + 0.3159646987915039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.04865432530641556 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201764017343521 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201764017343521, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201764017343521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17288094758987427, + 0.07525515556335449, + 0.15268568694591522, + 0.05561282858252525, + 0.030436014756560326, + 0.04714658111333847, + 0.12776398658752441, + 0.09003760665655136, + 0.07672359049320221, + 0.17145763337612152 + ], + "credit_map": [ + 0.010073862038552761, + 0.012766937725245953, + 0.03304923325777054, + 0.020361028611660004, + 0.011503936722874641, + 0.018197109922766685, + 0.018066421151161194, + 0.04886242747306824, + 0.00651866290718317, + 0.0063717663288116455, + 0.011724232695996761, + 0.011642049066722393, + 0.011367435567080975, + 0.007268782705068588, + 0.012376390397548676, + 0.007985832169651985, + 0.010165056213736534, + 0.017691614106297493, + 0.008815866895020008, + 0.017206167802214622, + 0.014013759791851044, + 0.018197953701019287, + 0.042511824518442154, + 0.024083437398076057, + 0.004179396666586399, + 0.008597134612500668, + 0.006689370144158602, + 0.005177579820156097, + 0.011227200739085674, + 0.0069168428890407085, + 0.007028582505881786, + 0.005796719342470169, + 0.0030999542213976383, + 0.004907304421067238, + 0.0049672238528728485, + 0.0026217603590339422, + 0.0026661138981580734, + 0.0042966557666659355, + 0.002833674196153879, + 0.005043330602347851, + 0.00476396968588233, + 0.0040284618735313416, + 0.00553694274276495, + 0.006333504803478718, + 0.005953716579824686, + 0.006103904917836189, + 0.007428187411278486, + 0.006997887976467609, + 0.0136869503185153, + 0.010587627068161964, + 0.017458530142903328, + 0.019826894626021385, + 0.012409836053848267, + 0.016378287225961685, + 0.015921415761113167, + 0.021494446322321892, + 0.014004610478878021, + 0.007977538742125034, + 0.015555372461676598, + 0.012258040718734264, + 0.005600637756288052, + 0.009550945833325386, + 0.00849313847720623, + 0.016597315669059753, + 0.008397484198212624, + 0.003836733987554908, + 0.013482832349836826, + 0.018584804609417915, + 0.004107630345970392, + 0.007480144966393709, + 0.009685891680419445, + 0.011148066259920597, + 0.006938814651221037, + 0.0026178229600191116, + 0.00681808078661561, + 0.0077401213347911835, + 0.004033554345369339, + 0.005463389214128256, + 0.07063762098550797, + 0.06720822304487228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.014201764017343521 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189658686518669 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189658686518669, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189658686518669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05337527021765709, + 0.0294855535030365, + 0.025578362867236137, + 0.04105343297123909, + 0.11566817760467529, + 0.05565193295478821, + 0.07408806681632996, + 0.04867640137672424, + 0.019106628373265266, + 0.5373161435127258 + ], + "credit_map": [ + 0.00432056887075305, + 0.005416851956397295, + 0.01084177102893591, + 0.0064971111714839935, + 0.004326983354985714, + 0.00558643089607358, + 0.005236251279711723, + 0.011149304918944836, + 0.002436739392578602, + 0.002742133801802993, + 0.0036499397829174995, + 0.004457795061171055, + 0.00494255218654871, + 0.002502894029021263, + 0.0053712148219347, + 0.003382284427061677, + 0.0011419816873967648, + 0.0015006519388407469, + 0.0016583786346018314, + 0.0026180215645581484, + 0.004230968188494444, + 0.0022234886419028044, + 0.008146967738866806, + 0.004057904705405235, + 0.004306148272007704, + 0.005329491570591927, + 0.004604294430464506, + 0.005984717048704624, + 0.004127846099436283, + 0.005702626891434193, + 0.00458927359431982, + 0.006409031804651022, + 0.008060894906520844, + 0.0073865861631929874, + 0.03512994945049286, + 0.0117298923432827, + 0.007968339137732983, + 0.005868337117135525, + 0.007833319716155529, + 0.03169086575508118, + 0.005436580162495375, + 0.004755477420985699, + 0.0055792732164263725, + 0.005376456771045923, + 0.006052052136510611, + 0.008605602197349072, + 0.010647240094840527, + 0.009199254214763641, + 0.004186434205621481, + 0.0038256337866187096, + 0.020986957475543022, + 0.007236442528665066, + 0.004148933570832014, + 0.006224524695426226, + 0.004607051610946655, + 0.022872084751725197, + 0.006971488241106272, + 0.004308957606554031, + 0.008022247813642025, + 0.007706582546234131, + 0.003333895932883024, + 0.0052059791050851345, + 0.004761472810059786, + 0.008365778252482414, + 0.0017948466120287776, + 0.001230596099048853, + 0.004057425074279308, + 0.0024674588348716497, + 0.0015406070742756128, + 0.001988830044865608, + 0.0017106982413679361, + 0.0043161665089428425, + 0.015199416317045689, + 0.0040917475707829, + 0.012875658459961414, + 0.01577604189515114, + 0.0115806357935071, + 0.01539643295109272, + 0.2369578331708908, + 0.22543840110301971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.014189658686518669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping aerial tracking shot, captured at a high frame rate to emphasize its exceptionally smooth motion, follows a majestic blended-wing-body airship as it executes a deliberate, low-altitude flight. This scene, rendered in a meticulously detailed artistic style, skillfully blends the ethereal romanticism of John William Waterhouse and the intricate fantasy of Vadim Kashin. The airship's slender, whale-like hull is a marvel of ornate metalwork, featuring fluid lines of polished brass and copper filigree crisscrossing its surface. Warm, golden hour light bathes the entire valley, shimmering the craft's hull as it glides silently and gracefully through delicate wisps of low-hanging cloud. Below, a tranquil river winds like a silver ribbon through an ancient forest, lushly wooded. The camera moves parallel to the airship, capturing the way it effortlessly navigates the valley's contours. Its massive shadow drifts across the deep green canopy, the leaves gently rustling, creating a tranquil, dreamlike majesty and a profound atmosphere of silent, epic journey.", + "seed": 10996, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_11/index_1373_seed_10996_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09333085203543305, + "vlm_logprob": -0.09333085203543305, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.09333085203543305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08226340264081955 + ], + "opposite_logprob_tokens": [ + -2.707263469696045 + ], + "logprob_reward": -0.08226340264081955, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08226340264081955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20696033537387848, + 0.057383421808481216, + 0.1466713845729828, + 0.12155928462743759, + 0.05488855391740799, + 0.06841257214546204, + 0.1709267497062683, + 0.02373107708990574, + 0.07632320374250412, + 0.07314342260360718 + ], + "credit_map": [ + 0.02667667157948017, + 0.02020922303199768, + 0.02102772332727909, + 0.017941469326615334, + 0.020306052640080452, + 0.027702821418642998, + 0.04069072753190994, + 0.03240564838051796, + 0.008044730871915817, + 0.005972869228571653, + 0.008116224780678749, + 0.005688375327736139, + 0.007007332518696785, + 0.006689735688269138, + 0.008166770450770855, + 0.007697380147874355, + 0.02300924062728882, + 0.0157224889844656, + 0.021188657730817795, + 0.015332053415477276, + 0.01140530314296484, + 0.024115175008773804, + 0.021075572818517685, + 0.014822885394096375, + 0.015304496511816978, + 0.013240745291113853, + 0.017515836283564568, + 0.011656980030238628, + 0.01454863976687193, + 0.016748003661632538, + 0.01702580600976944, + 0.015518778935074806, + 0.00636668736115098, + 0.00810878537595272, + 0.005298135336488485, + 0.00563394371420145, + 0.006101842969655991, + 0.007326032500714064, + 0.008892654441297054, + 0.0071604750119149685, + 0.007785337511450052, + 0.008819726295769215, + 0.008249509148299694, + 0.010311479680240154, + 0.010609915480017662, + 0.007689500227570534, + 0.006967431399971247, + 0.00797967053949833, + 0.02606537751853466, + 0.03259259834885597, + 0.021246302872896194, + 0.020535293966531754, + 0.018957823514938354, + 0.01614418998360634, + 0.014507738873362541, + 0.020877424627542496, + 0.0027617099694907665, + 0.002770521678030491, + 0.0045490562915802, + 0.004196858499199152, + 0.0019003951456397772, + 0.002070935443043709, + 0.0017511415062472224, + 0.003730458440259099, + 0.006084593012928963, + 0.012643566355109215, + 0.01148451678454876, + 0.01033041626214981, + 0.00858938880264759, + 0.007868759334087372, + 0.009231140837073326, + 0.010090820491313934, + 0.0071634347550570965, + 0.00848518218845129, + 0.013537323102355003, + 0.009830234572291374, + 0.009237049147486687, + 0.010200658813118935, + 0.005659077782183886, + 0.009030453860759735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.08226340264081955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814323902130127 + ], + "opposite_logprob_tokens": [ + -1.4064323902130127 + ], + "logprob_reward": -0.2814323902130127, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814323902130127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20634786784648895, + 0.05916290357708931, + 0.1456720381975174, + 0.12754137814044952, + 0.0818568617105484, + 0.04512053728103638, + 0.0754321813583374, + 0.02790660783648491, + 0.10927931219339371, + 0.121680349111557 + ], + "credit_map": [ + 0.028726959601044655, + 0.01992601901292801, + 0.020764745771884918, + 0.015677331015467644, + 0.01669877953827381, + 0.027921242639422417, + 0.0436239130795002, + 0.033008888363838196, + 0.008228717371821404, + 0.006079568061977625, + 0.008055108599364758, + 0.00719066197052598, + 0.004895832855254412, + 0.005903869867324829, + 0.010426909662783146, + 0.008382236585021019, + 0.022467050701379776, + 0.015306496061384678, + 0.02199629135429859, + 0.01605340838432312, + 0.012462088838219643, + 0.022065946832299232, + 0.020851029083132744, + 0.014469731599092484, + 0.018137365579605103, + 0.0185647364705801, + 0.013136575929820538, + 0.013482845388352871, + 0.018793119117617607, + 0.015158568508923054, + 0.01576528511941433, + 0.014502884820103645, + 0.01122228242456913, + 0.01179612148553133, + 0.007485025562345982, + 0.012317618355154991, + 0.009402023628354073, + 0.012295279651880264, + 0.00837030541151762, + 0.008968202397227287, + 0.005252378061413765, + 0.005827604793012142, + 0.004576439969241619, + 0.005329126957803965, + 0.006783134303987026, + 0.007547534070909023, + 0.0039552380330860615, + 0.00584908202290535, + 0.009620027616620064, + 0.013790509663522243, + 0.008010176941752434, + 0.009628396481275558, + 0.011523853987455368, + 0.008910324424505234, + 0.005122629459947348, + 0.008826255798339844, + 0.0029181039426475763, + 0.002772877225652337, + 0.005754603538662195, + 0.004272217862308025, + 0.002603794913738966, + 0.0028340844437479973, + 0.0022331434302031994, + 0.004517782479524612, + 0.012380504980683327, + 0.015372927300632, + 0.01502290926873684, + 0.012003551237285137, + 0.018517659977078438, + 0.012063602916896343, + 0.01139078289270401, + 0.01252736710011959, + 0.00908211711794138, + 0.007628495804965496, + 0.03990168869495392, + 0.019660864025354385, + 0.008769103325903416, + 0.01025583129376173, + 0.006121063604950905, + 0.020261183381080627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.2814323902130127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890587642788887 + ], + "opposite_logprob_tokens": [ + -2.4640588760375977 + ], + "logprob_reward": -0.0890587642788887, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890587642788887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2478334605693817, + 0.04530036821961403, + 0.10627873986959457, + 0.04399130865931511, + 0.0869697704911232, + 0.08941814303398132, + 0.12504251301288605, + 0.1729864925146103, + 0.05379166454076767, + 0.028387581929564476 + ], + "credit_map": [ + 0.03538746014237404, + 0.024614844471216202, + 0.024395348504185677, + 0.018262961879372597, + 0.023724887520074844, + 0.03204192593693733, + 0.049939561635255814, + 0.03946647793054581, + 0.005146753042936325, + 0.005268900655210018, + 0.005912748631089926, + 0.0059607066214084625, + 0.004742868710309267, + 0.006420534569770098, + 0.006146606057882309, + 0.00570125225931406, + 0.014824413694441319, + 0.007945344783365726, + 0.024597017094492912, + 0.01322893612086773, + 0.008293366990983486, + 0.009203232824802399, + 0.01138924341648817, + 0.016797183081507683, + 0.006789219565689564, + 0.004327449947595596, + 0.007381535600870848, + 0.004582705907523632, + 0.00452016294002533, + 0.004525389522314072, + 0.005774115212261677, + 0.006090730894356966, + 0.01080442126840353, + 0.013137268833816051, + 0.008399990387260914, + 0.014918774366378784, + 0.007070127408951521, + 0.011436933651566505, + 0.010980387218296528, + 0.010221865028142929, + 0.010244899429380894, + 0.011386532336473465, + 0.01350900623947382, + 0.016122760251164436, + 0.01058465801179409, + 0.007785394787788391, + 0.007868031039834023, + 0.011916856281459332, + 0.019218215718865395, + 0.028756801038980484, + 0.017188290134072304, + 0.014455199241638184, + 0.011667372658848763, + 0.009892648085951805, + 0.008242033421993256, + 0.015621949918568134, + 0.021542327478528023, + 0.018589992076158524, + 0.04725324735045433, + 0.024758178740739822, + 0.008911056444048882, + 0.012334967963397503, + 0.011264197528362274, + 0.02833251655101776, + 0.005355415400117636, + 0.008687440305948257, + 0.00865135807543993, + 0.00689911050722003, + 0.005901044234633446, + 0.004586194641888142, + 0.005977701395750046, + 0.0077334013767540455, + 0.0017103602876886725, + 0.0034624217078089714, + 0.006596868857741356, + 0.0037708920426666737, + 0.002777795074507594, + 0.0018031500512734056, + 0.0018473862437531352, + 0.006418706849217415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0890587642788887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028249616734683514 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028249616734683514, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028249616734683514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14746125042438507, + 0.09450109302997589, + 0.09568163752555847, + 0.17408207058906555, + 0.06017308309674263, + 0.15962649881839752, + 0.07608398050069809, + 0.04176121577620506, + 0.05517373979091644, + 0.09545549005270004 + ], + "credit_map": [ + 0.022689595818519592, + 0.012642346322536469, + 0.016130099073052406, + 0.011561465449631214, + 0.012289542704820633, + 0.021112686023116112, + 0.025263246148824692, + 0.025772258639335632, + 0.011517251841723919, + 0.007592562586069107, + 0.012928076088428497, + 0.009751004166901112, + 0.01046093087643385, + 0.008723437786102295, + 0.019062122330069542, + 0.014465708285570145, + 0.015405379235744476, + 0.008135879412293434, + 0.014707121066749096, + 0.0104755824431777, + 0.007165719289332628, + 0.014714007265865803, + 0.014530093409121037, + 0.01054785680025816, + 0.023062201216816902, + 0.024958670139312744, + 0.017961369827389717, + 0.018081462010741234, + 0.028159746900200844, + 0.020940620452165604, + 0.02151346392929554, + 0.019404524937272072, + 0.009429333731532097, + 0.00889589823782444, + 0.005429880693554878, + 0.008757015690207481, + 0.006484330166131258, + 0.007670282851904631, + 0.006881305947899818, + 0.006625030189752579, + 0.018888698890805244, + 0.025968030095100403, + 0.02192268893122673, + 0.02866082824766636, + 0.014085634611546993, + 0.016940129920840263, + 0.012202058918774128, + 0.020958423614501953, + 0.011226736009120941, + 0.012706086970865726, + 0.00967982318252325, + 0.009863916784524918, + 0.00935510452836752, + 0.006876726634800434, + 0.007280314341187477, + 0.009095268324017525, + 0.005622001830488443, + 0.004020993132144213, + 0.008276325650513172, + 0.006560022942721844, + 0.0034318692050874233, + 0.004194309003651142, + 0.0031850733794271946, + 0.006470619235187769, + 0.006156697869300842, + 0.007253389805555344, + 0.01276819221675396, + 0.0066728354431688786, + 0.004538581240922213, + 0.003957823384553194, + 0.0037063120398670435, + 0.010119903832674026, + 0.004443114157766104, + 0.009112834930419922, + 0.03191559761762619, + 0.010747303254902363, + 0.005682163406163454, + 0.004929587710648775, + 0.004513825289905071, + 0.02411106787621975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.0028249616734683514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blended-wing-body airship with a slender, whale-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the airship feature ornate metalwork with polished brass and copper filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a valley with a tranquil river winding through an ancient, lushly wooded forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and golden, illuminating the airship and valley during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, parallel tracking shot following the airship at a low altitude?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074741370975971 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074741370975971, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074741370975971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06796928495168686, + 0.03763362765312195, + 0.03962802141904831, + 0.1499551236629486, + 0.0797056183218956, + 0.122274249792099, + 0.30676743388175964, + 0.048626918345689774, + 0.04477273300290108, + 0.10266698151826859 + ], + "credit_map": [ + 0.009989547543227673, + 0.008929730392992496, + 0.007545782718807459, + 0.006127479951828718, + 0.0069742524065077305, + 0.010241935960948467, + 0.009520563296973705, + 0.008639988489449024, + 0.005976623389869928, + 0.0028311852365732193, + 0.005652294494211674, + 0.003375715808942914, + 0.0030972107779234648, + 0.003640685463324189, + 0.007212740369141102, + 0.005847169551998377, + 0.005474773235619068, + 0.0036282208748161793, + 0.006884311325848103, + 0.005426612216979265, + 0.003367051249369979, + 0.004011707846075296, + 0.005115262232720852, + 0.005720079876482487, + 0.015988316386938095, + 0.015288794413208961, + 0.030341384932398796, + 0.013922689482569695, + 0.018298383802175522, + 0.01827002316713333, + 0.01834266260266304, + 0.01950286328792572, + 0.011783585883677006, + 0.013663924299180508, + 0.007762673776596785, + 0.011343725956976414, + 0.007595340255647898, + 0.010798785835504532, + 0.008523624390363693, + 0.008233955129981041, + 0.012309045530855656, + 0.01740441471338272, + 0.016387229785323143, + 0.021155696362257004, + 0.013067444786429405, + 0.015455849468708038, + 0.009423322975635529, + 0.01707124151289463, + 0.047741811722517014, + 0.06956333667039871, + 0.04145463928580284, + 0.03683748468756676, + 0.02547483518719673, + 0.02608584053814411, + 0.021115902811288834, + 0.03849358484148979, + 0.0054046036675572395, + 0.005786812864243984, + 0.006645087618380785, + 0.008294491097331047, + 0.004519771784543991, + 0.005900796502828598, + 0.004858849570155144, + 0.007216502912342548, + 0.004738731309771538, + 0.008114421740174294, + 0.0068125356920063496, + 0.005294031463563442, + 0.006301168352365494, + 0.003803150961175561, + 0.004060722421854734, + 0.00564796756953001, + 0.004716942552477121, + 0.007060710806399584, + 0.03768410533666611, + 0.014762314036488533, + 0.005072843749076128, + 0.0034894237760454416, + 0.006014319136738777, + 0.0238663200289011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.011074741370975971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_12/index_tensor([1249, 1249])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_12/index_tensor([1249, 1249])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..514325651421f4fc1205d3c5b7e110b016e965cc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_12/index_tensor([1249, 1249])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot captures the hyper-accelerated, time-lapse blossoming of a Chicory 'Pumila' flower, its forms dominating the entire frame. Initially, the scene is bathed in the soft, diffused glow of natural morning light, highlighting the petals' stunningly vibrant purplish-red against a deeply out-of-focus, verdant green background. The camera remains steady as the tightly closed bud expands, its crepe-paper petals, rendered in microscopic detail to reveal every delicate wrinkle, begin to unfold with graceful and rapid vitality, revealing a cluster of bright yellow stamens at its core. Just as the blossom reaches its full, breathtaking splendor, this mesmerizing display of organic growth is interrupted. The entire screen flickers violently with digital interference; sharp, glitched horizontal scan lines race across the image, the perfectly purplish-red petals momentarily pixelate into rough blocks of color, and the frame is fractured by flickering three-color channel separation and white noise, creating a stark and unsettling contrast between the living world and abrupt technological decay.", + "seed": 32886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_12/index_1249_seed_32886_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026353324204683302, + "vlm_logprob": -0.026353324204683302, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.026353324204683302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013610069639980793 + ], + "opposite_logprob_tokens": [ + -4.888609886169434 + ], + "logprob_reward": -0.013610069639980793, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.013610069639980793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02686280757188797, + 0.012472331523895264, + 0.048733655363321304, + 0.01969251222908497, + 0.7864059209823608, + 0.006109371315687895, + 0.013145800679922104, + 0.016702787950634956, + 0.05989573150873184, + 0.009979063645005226 + ], + "credit_map": [ + 0.0033801470417529345, + 0.0027857476379722357, + 0.00359684438444674, + 0.002469859318807721, + 0.003315391018986702, + 0.0035180591512471437, + 0.0031401789747178555, + 0.004656581208109856, + 0.0010570120066404343, + 0.0009047609055414796, + 0.0014545995509251952, + 0.0020766709931194782, + 0.001047012279741466, + 0.0018540191231295466, + 0.001479737926274538, + 0.0025985201355069876, + 0.0047815409488976, + 0.005299716722220182, + 0.004160197917371988, + 0.004631580784916878, + 0.005992705933749676, + 0.008325721137225628, + 0.004783363081514835, + 0.010758832097053528, + 0.0016817228170111775, + 0.0020655461121350527, + 0.0022999681532382965, + 0.003577116411179304, + 0.002439972711727023, + 0.0025518944021314383, + 0.002922461135312915, + 0.0021538317669183016, + 0.02265768125653267, + 0.030509069561958313, + 0.021246809512376785, + 0.025914272293448448, + 0.028282830491662025, + 0.3128478527069092, + 0.02450672909617424, + 0.32044076919555664, + 0.0005723715876229107, + 0.000595593883190304, + 0.0010204032296314836, + 0.0008713503484614193, + 0.0006901844753883779, + 0.0006518804584629834, + 0.0009180917986668646, + 0.0007894962909631431, + 0.0010776353301480412, + 0.0012791361659765244, + 0.0012974656419828534, + 0.0025254073552787304, + 0.0013029462425038218, + 0.002018915954977274, + 0.0014203662285581231, + 0.002223928924649954, + 0.0019048176473006606, + 0.001481823856011033, + 0.0020143985748291016, + 0.0020739277824759483, + 0.001725942362099886, + 0.002260035602375865, + 0.0028584448155015707, + 0.002383399521932006, + 0.006177674978971481, + 0.003453323617577553, + 0.0042152805253863335, + 0.003987904638051987, + 0.004708821885287762, + 0.009177190251648426, + 0.0060636685229837894, + 0.02211187221109867, + 0.0009938207222148776, + 0.0010072427103295922, + 0.002300090389326215, + 0.00109531928319484, + 0.0012198594631627202, + 0.0015404503792524338, + 0.0008371633011847734, + 0.0009851177455857396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.013610069639980793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04315006360411644 + ], + "opposite_logprob_tokens": [ + -3.168149948120117 + ], + "logprob_reward": -0.04315006360411644, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04315006360411644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06307156383991241, + 0.05059071257710457, + 0.02016296796500683, + 0.047804590314626694, + 0.6669872403144836, + 0.016533104702830315, + 0.027281716465950012, + 0.0644993856549263, + 0.019763104617595673, + 0.02330555021762848 + ], + "credit_map": [ + 0.007928023114800453, + 0.006369691342115402, + 0.007347192149609327, + 0.006422612816095352, + 0.007362294010818005, + 0.012209863401949406, + 0.00580717995762825, + 0.009624708443880081, + 0.006615320686250925, + 0.004256669897586107, + 0.004724410362541676, + 0.005877553019672632, + 0.004037917125970125, + 0.010268672369420528, + 0.003807310713455081, + 0.011002855375409126, + 0.0020677484571933746, + 0.002174582565203309, + 0.0023661653976887465, + 0.0021695264149457216, + 0.00260936189442873, + 0.0033638724125921726, + 0.001709510455839336, + 0.003702199086546898, + 0.004994599148631096, + 0.005347625818103552, + 0.006537087727338076, + 0.008077617734670639, + 0.004968751687556505, + 0.0059187342412769794, + 0.006303598638623953, + 0.005656573921442032, + 0.02020140551030636, + 0.028665879741311073, + 0.01959778182208538, + 0.023003384470939636, + 0.026142539456486702, + 0.26067259907722473, + 0.022544635459780693, + 0.2661590874195099, + 0.0016961812507361174, + 0.0022973844315856695, + 0.0020369417034089565, + 0.0018575238063931465, + 0.0017707064980641007, + 0.0022935867309570312, + 0.001976846484467387, + 0.00260393344797194, + 0.0025742449797689915, + 0.002596657257527113, + 0.002733323024585843, + 0.005189541727304459, + 0.0025498250033706427, + 0.0036452950444072485, + 0.003247894113883376, + 0.004744935780763626, + 0.00851662177592516, + 0.006271082442253828, + 0.007020073477178812, + 0.007315289694815874, + 0.005829619709402323, + 0.010027033276855946, + 0.0085428012534976, + 0.010976864956319332, + 0.0034673400223255157, + 0.0014288455713540316, + 0.001615972607396543, + 0.001928269281052053, + 0.002435476751998067, + 0.004305352456867695, + 0.0016532232984900475, + 0.0029286243952810764, + 0.004024900030344725, + 0.0013454060535877943, + 0.002876627491787076, + 0.0024021535646170378, + 0.0026796397287398577, + 0.005628529936075211, + 0.0016564656980335712, + 0.0026918286457657814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.04315006360411644 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300418630242348 + ], + "opposite_logprob_tokens": [ + -3.7733004093170166 + ], + "logprob_reward": -0.023300418630242348, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300418630242348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014062534086406231, + 0.0068184612318873405, + 0.009594324976205826, + 0.010001826100051403, + 0.9237533807754517, + 0.00398494815453887, + 0.004498200491070747, + 0.006134317256510258, + 0.002698131836950779, + 0.01845388300716877 + ], + "credit_map": [ + 0.0014309484977275133, + 0.0016581189120188355, + 0.0017076005460694432, + 0.0013222789857536554, + 0.0015074940165504813, + 0.0030005453154444695, + 0.0014773730654269457, + 0.001958175329491496, + 0.0006058557773940265, + 0.000604672881308943, + 0.0006046471535228193, + 0.0009844694286584854, + 0.0005944586009718478, + 0.0013224853901192546, + 0.0006672106683254242, + 0.0014346615644171834, + 0.001155190053395927, + 0.001252026529982686, + 0.0007662592106498778, + 0.0008654945995658636, + 0.0010624005226418376, + 0.0015563290799036622, + 0.0008512234780937433, + 0.0020854014437645674, + 0.0010640075197443366, + 0.0012841693824157119, + 0.0012460153084248304, + 0.0012969839153811336, + 0.0016018830938264728, + 0.001219504396431148, + 0.0013538110069930553, + 0.0009354516514576972, + 0.026849891990423203, + 0.03678164258599281, + 0.02560393698513508, + 0.030582550913095474, + 0.03395793214440346, + 0.3661317229270935, + 0.02932659722864628, + 0.37451913952827454, + 0.00028826118796132505, + 0.0003283274418208748, + 0.000576615915633738, + 0.0007907602703198791, + 0.00024294070317409933, + 0.000582175562158227, + 0.0006439505377784371, + 0.0005319165647961199, + 0.0003317879163660109, + 0.0004837439628317952, + 0.0005642488831654191, + 0.00116238952614367, + 0.0004289507633075118, + 0.00043161847861483693, + 0.000615126162301749, + 0.000480335031170398, + 0.0008392962045036256, + 0.0005283581558614969, + 0.0006226147525012493, + 0.0008314672159031034, + 0.0005157962441444397, + 0.0008165804902091622, + 0.0007308977073989809, + 0.0012493063695728779, + 0.0004509222344495356, + 0.0001892619184218347, + 0.00019555160542950034, + 0.0002466539735905826, + 0.00024412997299805284, + 0.0006224574171938002, + 0.0001937875640578568, + 0.0005553671508096159, + 0.0013626676518470049, + 0.0008315583108924329, + 0.00256987358443439, + 0.006457523442804813, + 0.0013343817554414272, + 0.0033554101828485727, + 0.0012387005845084786, + 0.0013037680182605982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.023300418630242348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639813400804996 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639813400804996, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639813400804996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04512767866253853, + 0.017420196905732155, + 0.009871489368379116, + 0.015073259361088276, + 0.7804357409477234, + 0.011856989935040474, + 0.027529466897249222, + 0.02060682512819767, + 0.00988310668617487, + 0.06219525262713432 + ], + "credit_map": [ + 0.006861042231321335, + 0.003992279525846243, + 0.005319656804203987, + 0.0027718422934412956, + 0.004633982200175524, + 0.008665978908538818, + 0.003223250387236476, + 0.00965964701026678, + 0.0017709933454170823, + 0.001462736981920898, + 0.0023047931026667356, + 0.0019198048394173384, + 0.0021250159479677677, + 0.0022793493699282408, + 0.0018019434064626694, + 0.0037555592134594917, + 0.0008811490843072534, + 0.0010355579433962703, + 0.0014252131804823875, + 0.0008995499811135232, + 0.001143164816312492, + 0.001460899249650538, + 0.0008204623591154814, + 0.002205492928624153, + 0.0014247552026063204, + 0.001308655715547502, + 0.0018778827507048845, + 0.0013299866113811731, + 0.0015247834380716085, + 0.0027755536139011383, + 0.0012878372799605131, + 0.0035438055638223886, + 0.022959811612963676, + 0.03193319961428642, + 0.02214863710105419, + 0.026147624477744102, + 0.02911430038511753, + 0.3081469237804413, + 0.024797596037387848, + 0.31518763303756714, + 0.0008500486146658659, + 0.0007959262002259493, + 0.001438609091565013, + 0.0015950278611853719, + 0.0007606706931255758, + 0.002423845464363694, + 0.0016870753606781363, + 0.0023057865910232067, + 0.0037786057218909264, + 0.00230056862346828, + 0.003043357515707612, + 0.0029191526118665934, + 0.002946609864011407, + 0.004431515466421843, + 0.004069062881171703, + 0.004040593281388283, + 0.002686147578060627, + 0.0022214031778275967, + 0.0018860982963815331, + 0.0023872125893831253, + 0.0017199692083522677, + 0.002630348550155759, + 0.0023547441232949495, + 0.004720901604741812, + 0.0008763493970036507, + 0.0009201983921229839, + 0.0010854661231860518, + 0.0011298356112092733, + 0.0009152165148407221, + 0.0012150804977864027, + 0.000927439599763602, + 0.0028135201428085566, + 0.005029338877648115, + 0.0027578810695558786, + 0.008597936481237411, + 0.017529118806123734, + 0.005930972285568714, + 0.012849041260778904, + 0.003920981660485268, + 0.005579981952905655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.008639813400804996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306625574827194 + ], + "opposite_logprob_tokens": [ + -3.1680662631988525 + ], + "logprob_reward": -0.04306625574827194, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306625574827194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.004829509183764458, + 0.004202619660645723, + 0.003698912914842367, + 0.007501321379095316, + 0.77399742603302, + 0.004820503760129213, + 0.004672693088650703, + 0.004620750900357962, + 0.01258992962539196, + 0.17906630039215088 + ], + "credit_map": [ + 0.000592720927670598, + 0.0005719895125366747, + 0.0005492885829880834, + 0.0006045495974831283, + 0.0006750658503733575, + 0.0005599282449111342, + 0.000641466467641294, + 0.0006345005240291357, + 0.00044160589459352195, + 0.0004956203047186136, + 0.0004290191864129156, + 0.0005907640443183482, + 0.00041666923789307475, + 0.0006656357436440885, + 0.0004081436200067401, + 0.0007551616872660816, + 0.000379222328774631, + 0.0003687592688947916, + 0.0005214105476625264, + 0.0003835540555883199, + 0.00043809483759105206, + 0.0005480961408466101, + 0.0003614648594520986, + 0.0006983114290051162, + 0.0006680298829451203, + 0.0008845973061397672, + 0.0008314648293890059, + 0.0012630815617740154, + 0.0007534320466220379, + 0.0010180473327636719, + 0.0007217537495307624, + 0.0013609154848381877, + 0.022562921047210693, + 0.031194673851132393, + 0.021338678896427155, + 0.026253592222929, + 0.028544776141643524, + 0.30640140175819397, + 0.024727674201130867, + 0.3129737973213196, + 0.000452827982371673, + 0.000575565209146589, + 0.0006784205324947834, + 0.0007411861442960799, + 0.0004276198160368949, + 0.0005854273913428187, + 0.0007047957624308765, + 0.0006546613876707852, + 0.0005323663935996592, + 0.0003667437704280019, + 0.00042689102701842785, + 0.0007475382299162447, + 0.0003084793279413134, + 0.001061036135070026, + 0.0005722814821638167, + 0.0006573569844476879, + 0.0006410719943232834, + 0.0005608302890323102, + 0.0004216134548187256, + 0.0006231977604329586, + 0.0005839064833708107, + 0.0005063118296675384, + 0.00047112826723605394, + 0.0008126910543069243, + 0.0009867739863693714, + 0.0008479826501570642, + 0.0012654063757508993, + 0.0013113770401105285, + 0.0012259319191798568, + 0.001513599301688373, + 0.0013491488061845303, + 0.004089710768312216, + 0.014464598149061203, + 0.0072349607944488525, + 0.028542321175336838, + 0.05339375138282776, + 0.014534195885062218, + 0.03531991317868233, + 0.01074898149818182, + 0.014827614650130272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.04306625574827194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the hyper-accelerated, time-lapse blossoming of a Chicory 'Pumila' flower, its forms dominating the entire frame. Initially, the scene is bathed in the soft, diffused glow of natural morning light, highlighting the petals' stunningly vibrant purplish-red against a deeply out-of-focus, verdant green background. The camera remains steady as the tightly closed bud expands, its crepe-paper petals, rendered in microscopic detail to reveal every delicate wrinkle, begin to unfold with graceful and rapid vitality, revealing a cluster of bright yellow stamens at its core. Just as the blossom reaches its full, breathtaking splendor, this mesmerizing display of organic growth is interrupted. The entire screen flickers violently with digital interference; sharp, glitched horizontal scan lines race across the image, the perfectly purplish-red petals momentarily pixelate into rough blocks of color, and the frame is fractured by flickering three-color channel separation and white noise, creating a stark and unsettling contrast between the living world and abrupt technological decay.", + "seed": 32887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_12/index_1249_seed_32887_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014690884947776794, + "vlm_logprob": -0.014690884947776794, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.014690884947776794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021830964833498 + ], + "opposite_logprob_tokens": [ + -4.896831035614014 + ], + "logprob_reward": -0.021830964833498, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.021830964833498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1251513510942459, + 0.19074970483779907, + 0.013467089273035526, + 0.11893706023693085, + 0.057222917675971985, + 0.34221091866493225, + 0.03694240003824234, + 0.021811282262206078, + 0.05222798511385918, + 0.04127929359674454 + ], + "credit_map": [ + 0.010570988059043884, + 0.013585475273430347, + 0.03651058301329613, + 0.006999875884503126, + 0.01292787678539753, + 0.02868146263062954, + 0.004709367174655199, + 0.011165727861225605, + 0.013241547159850597, + 0.028732789680361748, + 0.07896751910448074, + 0.00713684968650341, + 0.006998326163738966, + 0.044414304196834564, + 0.004011157434433699, + 0.007247213274240494, + 0.0016116560436785221, + 0.0014045938150957227, + 0.0025506000965833664, + 0.0016266194870695472, + 0.0013179931556805968, + 0.0015995504800230265, + 0.0016285742167383432, + 0.001727502909488976, + 0.009265748783946037, + 0.00666094059124589, + 0.04520173370838165, + 0.004355441778898239, + 0.0035341286566108465, + 0.04215582087635994, + 0.002224456053227186, + 0.005538795609027147, + 0.004929457325488329, + 0.003173487726598978, + 0.012165496125817299, + 0.0032370013650506735, + 0.003510930808261037, + 0.018314993008971214, + 0.0031352792866528034, + 0.008756272494792938, + 0.039038803428411484, + 0.010378718376159668, + 0.13126489520072937, + 0.015015953220427036, + 0.010202872566878796, + 0.09166082739830017, + 0.01056349091231823, + 0.034085363149642944, + 0.0062109543941915035, + 0.0025974386371672153, + 0.006122222635895014, + 0.001962510170415044, + 0.002662091515958309, + 0.006466381251811981, + 0.00219172821380198, + 0.008729075081646442, + 0.0023471792228519917, + 0.001827741740271449, + 0.0029534369241446257, + 0.003768704365938902, + 0.0022298702970147133, + 0.0023256605491042137, + 0.002950627589598298, + 0.003408062271773815, + 0.005516732111573219, + 0.005028305575251579, + 0.006016524042934179, + 0.0056007010862231255, + 0.00626137712970376, + 0.006303447764366865, + 0.008438259363174438, + 0.009062642231583595, + 0.006094551645219326, + 0.009564673528075218, + 0.0036957585252821445, + 0.003756329882889986, + 0.0040268185548484325, + 0.004227840807288885, + 0.004689712543040514, + 0.005223608575761318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.021830964833498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026389271020889282 + ], + "opposite_logprob_tokens": [ + -3.6513893604278564 + ], + "logprob_reward": -0.026389271020889282, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026389271020889282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2757778763771057, + 0.0619686096906662, + 0.012929968535900116, + 0.03592926263809204, + 0.03110656887292862, + 0.21593251824378967, + 0.025769580155611038, + 0.037508588284254074, + 0.26226651668548584, + 0.04081042855978012 + ], + "credit_map": [ + 0.0251857228577137, + 0.030486948788166046, + 0.08962637931108475, + 0.014181689359247684, + 0.027800951153039932, + 0.056358277797698975, + 0.01155698113143444, + 0.02058093436062336, + 0.006581803783774376, + 0.009600580669939518, + 0.014952568337321281, + 0.006021812558174133, + 0.005038535688072443, + 0.005997218191623688, + 0.0068558258935809135, + 0.006920266896486282, + 0.0014670045347884297, + 0.0014430461451411247, + 0.002906721318140626, + 0.0014502800768241286, + 0.0014170475769788027, + 0.0015711031155660748, + 0.0011387845734134316, + 0.0015359815442934632, + 0.003519455436617136, + 0.0032401150092482567, + 0.009172240272164345, + 0.002411270746961236, + 0.0018427239265292883, + 0.010154616087675095, + 0.0025183777324855328, + 0.003070464823395014, + 0.0023847133852541447, + 0.004397680051624775, + 0.005048902239650488, + 0.0024258720222860575, + 0.005101359449326992, + 0.004839883651584387, + 0.0025038281455636024, + 0.00440433481708169, + 0.022991375997662544, + 0.0047613950446248055, + 0.08162056654691696, + 0.010222291573882103, + 0.007226834073662758, + 0.05823962762951851, + 0.006580815650522709, + 0.02428961731493473, + 0.0036621622275561094, + 0.001850238419137895, + 0.0045104180462658405, + 0.00204650335945189, + 0.0018824219005182385, + 0.0045858402736485004, + 0.001905940007418394, + 0.0053260535933077335, + 0.004755349829792976, + 0.003843612503260374, + 0.00455361558124423, + 0.007126644719392061, + 0.004466303624212742, + 0.003901768010109663, + 0.00440702261403203, + 0.004454272799193859, + 0.03462471440434456, + 0.03281888738274574, + 0.034520674496889114, + 0.02355794422328472, + 0.03324493393301964, + 0.02950497716665268, + 0.028059326112270355, + 0.045935068279504776, + 0.004178558010607958, + 0.008908621966838837, + 0.0036235356237739325, + 0.004178284667432308, + 0.004483472090214491, + 0.00638902373611927, + 0.0041201612912118435, + 0.004928775131702423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.026389271020889282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.123307965695858, + 0.03770866245031357, + 0.023034406825900078, + 0.28443947434425354, + 0.062426961958408356, + 0.10229501128196716, + 0.1395011693239212, + 0.06855258345603943, + 0.1465258151292801, + 0.012207962572574615 + ], + "credit_map": [ + 0.009078582748770714, + 0.014496312476694584, + 0.03708568960428238, + 0.009718723595142365, + 0.013593656942248344, + 0.009630661457777023, + 0.009169750846922398, + 0.020534584298729897, + 0.003456037025898695, + 0.0028839653823524714, + 0.002548676449805498, + 0.0027945369947701693, + 0.00230574794113636, + 0.01070113480091095, + 0.0033706030808389187, + 0.009647960774600506, + 0.002363810548558831, + 0.0017824455862864852, + 0.00455927150323987, + 0.0026207701303064823, + 0.002009735442698002, + 0.005799733567982912, + 0.0021326991263777018, + 0.0017659406876191497, + 0.020221905782818794, + 0.013019793666899204, + 0.09599583595991135, + 0.00789579376578331, + 0.004856765270233154, + 0.11285474896430969, + 0.004715555813163519, + 0.024879062548279762, + 0.006140712648630142, + 0.0037048235535621643, + 0.011662972159683704, + 0.003611481748521328, + 0.005183176137506962, + 0.017863448709249496, + 0.0039122155867516994, + 0.010348128154873848, + 0.010713960975408554, + 0.008606862276792526, + 0.022299649193882942, + 0.009127628989517689, + 0.006585624534636736, + 0.015113567933440208, + 0.007999258115887642, + 0.021848458796739578, + 0.018381718546152115, + 0.009349682368338108, + 0.019524341449141502, + 0.009464999660849571, + 0.011068485677242279, + 0.031219711527228355, + 0.008797994814813137, + 0.03169422969222069, + 0.006720780860632658, + 0.007085909601300955, + 0.008508481085300446, + 0.01306493952870369, + 0.007384535390883684, + 0.007991031743586063, + 0.007804836146533489, + 0.009992065839469433, + 0.019561732187867165, + 0.01670994609594345, + 0.01922333985567093, + 0.013480485416948795, + 0.018082812428474426, + 0.017039019614458084, + 0.01683127135038376, + 0.025597210973501205, + 0.0008657821454107761, + 0.0015922961756587029, + 0.0011025617131963372, + 0.0023321558255702257, + 0.001439290470443666, + 0.0010477127507328987, + 0.002032053656876087, + 0.0017961094854399562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951066501438618 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005951066501438618, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951066501438618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11145880073308945, + 0.1418771743774414, + 0.018339890986680984, + 0.4726622402667999, + 0.05040999874472618, + 0.10888653993606567, + 0.036134518682956696, + 0.03085055947303772, + 0.019409965723752975, + 0.009970252402126789 + ], + "credit_map": [ + 0.010348649695515633, + 0.009800545871257782, + 0.02202734351158142, + 0.007159217726439238, + 0.008183290250599384, + 0.02655772492289543, + 0.005997555796056986, + 0.02138446643948555, + 0.009100312367081642, + 0.01923256739974022, + 0.05756699666380882, + 0.004613057244569063, + 0.004925807472318411, + 0.0379069484770298, + 0.0027663239743560553, + 0.005765163339674473, + 0.002095977310091257, + 0.0016475461889058352, + 0.004724510014057159, + 0.0018359855748713017, + 0.0012762441765516996, + 0.002123952843248844, + 0.0013585792621597648, + 0.003277095500379801, + 0.03466732054948807, + 0.02286406233906746, + 0.17216505110263824, + 0.012910800985991955, + 0.008620540611445904, + 0.18638622760772705, + 0.005139041226357222, + 0.02990919165313244, + 0.006256920751184225, + 0.003341920441016555, + 0.011105201207101345, + 0.002571595599874854, + 0.002529047429561615, + 0.015112364664673805, + 0.0022589298896491528, + 0.0072340164333581924, + 0.012896126136183739, + 0.003097813343629241, + 0.03530728071928024, + 0.004674642812460661, + 0.004029562696814537, + 0.02003740705549717, + 0.003817516379058361, + 0.025026196613907814, + 0.005224959924817085, + 0.0025956793688237667, + 0.005965756252408028, + 0.002038919599726796, + 0.0024218859616667032, + 0.0075368029065430164, + 0.0018719034269452095, + 0.008478612639009953, + 0.003596676979213953, + 0.0030493352096527815, + 0.004494084976613522, + 0.0052304575219750404, + 0.0032766619697213173, + 0.003518969053402543, + 0.003636577632278204, + 0.004047793801873922, + 0.002349693328142166, + 0.0020963612478226423, + 0.0026593708898872137, + 0.0019749358762055635, + 0.002004270674660802, + 0.002491849008947611, + 0.002374633215367794, + 0.0034588519483804703, + 0.0009073489927686751, + 0.0011960368137806654, + 0.0010365599300712347, + 0.0014857532223686576, + 0.0011042177211493254, + 0.0009344522841274738, + 0.0017253148835152388, + 0.0015805686125531793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.005951066501438618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539473362267017 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539473362267017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539473362267017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18636472523212433, + 0.06566528230905533, + 0.021786890923976898, + 0.04051109775900841, + 0.039790984243154526, + 0.14734961092472076, + 0.048728570342063904, + 0.13475528359413147, + 0.0916338786482811, + 0.22341370582580566 + ], + "credit_map": [ + 0.01660400815308094, + 0.020470593124628067, + 0.017261523753404617, + 0.015363425016403198, + 0.013399791903793812, + 0.04773492366075516, + 0.015622830018401146, + 0.03990761563181877, + 0.005949169863015413, + 0.010687695816159248, + 0.02334701083600521, + 0.0046664876863360405, + 0.00393180176615715, + 0.007333104033023119, + 0.003485023509711027, + 0.006264988798648119, + 0.0024035153910517693, + 0.001925552380271256, + 0.002367147710174322, + 0.001607366488315165, + 0.0016645618015900254, + 0.004757966846227646, + 0.0017566391034051776, + 0.005304140970110893, + 0.004122927784919739, + 0.0025167420972138643, + 0.007307797204703093, + 0.0029631482902914286, + 0.003059225156903267, + 0.01117179449647665, + 0.0024076111149042845, + 0.006961849518120289, + 0.004078614991158247, + 0.0038869399577379227, + 0.007224819622933865, + 0.005003160797059536, + 0.0031463811174035072, + 0.006506340112537146, + 0.004117709118872881, + 0.005827020388096571, + 0.02128533460199833, + 0.005011402536183596, + 0.04822402819991112, + 0.007227572146803141, + 0.005190414376556873, + 0.015246853232383728, + 0.006201205309480429, + 0.038962796330451965, + 0.005855023395270109, + 0.0050831325352191925, + 0.006525573320686817, + 0.006543693132698536, + 0.0053675309754908085, + 0.00793280266225338, + 0.004926301538944244, + 0.00649451045319438, + 0.013032129034399986, + 0.013520855456590652, + 0.019602034240961075, + 0.02118632383644581, + 0.017316555604338646, + 0.014376021921634674, + 0.016079582273960114, + 0.019641762599349022, + 0.009630228392779827, + 0.011062229983508587, + 0.012745218351483345, + 0.01077816542237997, + 0.008700862526893616, + 0.010773992165923119, + 0.012889227829873562, + 0.015053949318826199, + 0.019969509914517403, + 0.0403384268283844, + 0.023634348064661026, + 0.029376041144132614, + 0.022743597626686096, + 0.025775142014026642, + 0.030243361368775368, + 0.031333256512880325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.012539473362267017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the hyper-accelerated, time-lapse blossoming of a Chicory 'Pumila' flower, its forms dominating the entire frame. Initially, the scene is bathed in the soft, diffused glow of natural morning light, highlighting the petals' stunningly vibrant purplish-red against a deeply out-of-focus, verdant green background. The camera remains steady as the tightly closed bud expands, its crepe-paper petals, rendered in microscopic detail to reveal every delicate wrinkle, begin to unfold with graceful and rapid vitality, revealing a cluster of bright yellow stamens at its core. Just as the blossom reaches its full, breathtaking splendor, this mesmerizing display of organic growth is interrupted. The entire screen flickers violently with digital interference; sharp, glitched horizontal scan lines race across the image, the perfectly purplish-red petals momentarily pixelate into rough blocks of color, and the frame is fractured by flickering three-color channel separation and white noise, creating a stark and unsettling contrast between the living world and abrupt technological decay.", + "seed": 32888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_12/index_1249_seed_32888_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13254883689805866, + "vlm_logprob": -0.13254883689805866, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.13254883689805866, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09761640429496765 + ], + "opposite_logprob_tokens": [ + -2.47261643409729 + ], + "logprob_reward": -0.09761640429496765, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09761640429496765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1780984103679657, + 0.19777673482894897, + 0.06666631251573563, + 0.09112691879272461, + 0.166044220328331, + 0.02978121116757393, + 0.08319679647684097, + 0.01840580813586712, + 0.10053126513957977, + 0.06837230175733566 + ], + "credit_map": [ + 0.016205288469791412, + 0.02157665602862835, + 0.026455091312527657, + 0.03789335489273071, + 0.017650052905082703, + 0.018992261961102486, + 0.016360586509108543, + 0.022965116426348686, + 0.02239692397415638, + 0.014593876898288727, + 0.028605977073311806, + 0.014487477019429207, + 0.016191890463232994, + 0.04633143171668053, + 0.013269874267280102, + 0.04189928621053696, + 0.01692444086074829, + 0.005066842306405306, + 0.009302400052547455, + 0.004301091656088829, + 0.005725574679672718, + 0.006255920976400375, + 0.004953079391270876, + 0.014136958867311478, + 0.009438096545636654, + 0.007020643446594477, + 0.011901285499334335, + 0.01055167056620121, + 0.01362870167940855, + 0.0059336586855351925, + 0.022438831627368927, + 0.01021402794867754, + 0.021787967532873154, + 0.012630907818675041, + 0.019381167367100716, + 0.01822322979569435, + 0.0229814350605011, + 0.01197795569896698, + 0.03421865031123161, + 0.024842893704771996, + 0.004786128178238869, + 0.002569641685113311, + 0.004927029367536306, + 0.003908485174179077, + 0.002535625360906124, + 0.0033603974152356386, + 0.0039455462247133255, + 0.0037483577616512775, + 0.024681344628334045, + 0.004039156716316938, + 0.02425229735672474, + 0.005687802564352751, + 0.0037166131660342216, + 0.004694904666393995, + 0.004503681790083647, + 0.01162099651992321, + 0.003471696749329567, + 0.0013439044123515487, + 0.003916153218597174, + 0.0022521670907735825, + 0.0018294787732884288, + 0.0015244943788275123, + 0.0016594217158854008, + 0.0024084909819066525, + 0.013336647301912308, + 0.009735146537423134, + 0.014544669538736343, + 0.01504953857511282, + 0.013524615205824375, + 0.01167750172317028, + 0.010382148437201977, + 0.012280995957553387, + 0.020870383828878403, + 0.0033259873744100332, + 0.020315546542406082, + 0.003401954658329487, + 0.0036116247065365314, + 0.008028900250792503, + 0.0032208547927439213, + 0.005597054027020931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.09761640429496765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029872257262468338 + ], + "opposite_logprob_tokens": [ + -3.529872179031372 + ], + "logprob_reward": -0.029872257262468338, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029872257262468338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11838114261627197, + 0.05899154394865036, + 0.02531035616993904, + 0.09188220649957657, + 0.2600022256374359, + 0.02246859483420849, + 0.06336129456758499, + 0.017151983454823494, + 0.07388001680374146, + 0.2685706615447998 + ], + "credit_map": [ + 0.01863035373389721, + 0.014192203059792519, + 0.020569922402501106, + 0.0128311887383461, + 0.013507911935448647, + 0.017121748998761177, + 0.010033076629042625, + 0.01149473525583744, + 0.007020626217126846, + 0.0038942229002714157, + 0.008277465589344501, + 0.004125306848436594, + 0.0042338985949754715, + 0.014304490759968758, + 0.0037451002281159163, + 0.013390432111918926, + 0.002798452042043209, + 0.0023295446299016476, + 0.0033978899009525776, + 0.0031129068229347467, + 0.004180230665951967, + 0.0024038117844611406, + 0.0041556344367563725, + 0.0029318854212760925, + 0.009289468638598919, + 0.006706883665174246, + 0.013420254923403263, + 0.011183902621269226, + 0.01376380492001772, + 0.005486448761075735, + 0.022293688729405403, + 0.009737754240632057, + 0.027891667559742928, + 0.018074851483106613, + 0.06031752750277519, + 0.019720857962965965, + 0.027732454240322113, + 0.015764446929097176, + 0.03376414254307747, + 0.05673627182841301, + 0.0031323672737926245, + 0.0021829754114151, + 0.0028128891717642546, + 0.0029838362243026495, + 0.0027317034546285868, + 0.0023860568180680275, + 0.002483972115442157, + 0.0037547943647950888, + 0.01872292160987854, + 0.0035938327200710773, + 0.018761908635497093, + 0.00415909755975008, + 0.002902272390201688, + 0.003545620245859027, + 0.003324047662317753, + 0.008351593278348446, + 0.0038158122915774584, + 0.0009821443818509579, + 0.0034161412622779608, + 0.0018290900625288486, + 0.001682200818322599, + 0.0010718166595324874, + 0.002407597843557596, + 0.0019471811829134822, + 0.008283411152660847, + 0.008278834633529186, + 0.008565816096961498, + 0.012379379011690617, + 0.007596018258482218, + 0.00761698093265295, + 0.011040898971259594, + 0.010118678212165833, + 0.09412117302417755, + 0.010129237547516823, + 0.09638593345880508, + 0.008288281038403511, + 0.00994599424302578, + 0.02354421466588974, + 0.009863034822046757, + 0.016292793676257133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.029872257262468338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989473044872284 + ], + "opposite_logprob_tokens": [ + -1.8048946857452393 + ], + "logprob_reward": -0.17989473044872284, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989473044872284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16783417761325836, + 0.25476664304733276, + 0.13509579002857208, + 0.07357349246740341, + 0.0854402631521225, + 0.0304035022854805, + 0.02989899180829525, + 0.0330677330493927, + 0.0681324228644371, + 0.12178698182106018 + ], + "credit_map": [ + 0.016984175890684128, + 0.02070474438369274, + 0.023058565333485603, + 0.030407186597585678, + 0.0206814706325531, + 0.018869945779442787, + 0.018623964861035347, + 0.01850411668419838, + 0.023294685408473015, + 0.027394458651542664, + 0.039381176233291626, + 0.02793419361114502, + 0.027974629774689674, + 0.03637203946709633, + 0.028024770319461823, + 0.04439069330692291, + 0.026497915387153625, + 0.012914735823869705, + 0.01899784244596958, + 0.012157917022705078, + 0.01626446284353733, + 0.010499434545636177, + 0.01419125497341156, + 0.023572223260998726, + 0.007752842269837856, + 0.005339756142348051, + 0.009100489318370819, + 0.009321669116616249, + 0.011374222114682198, + 0.004737882874906063, + 0.01728864014148712, + 0.008657995611429214, + 0.009912186302244663, + 0.005325167905539274, + 0.01988925226032734, + 0.005674553569406271, + 0.009746640920639038, + 0.005256983917206526, + 0.008197608403861523, + 0.021437861025333405, + 0.005490991286933422, + 0.0027820435352623463, + 0.006733080837875605, + 0.0030213131103664637, + 0.002547333249822259, + 0.0020766067318618298, + 0.003464101580902934, + 0.0042880321852862835, + 0.004675543401390314, + 0.0023833613377064466, + 0.006047414615750313, + 0.003822116646915674, + 0.00316457636654377, + 0.002792974468320608, + 0.0034947404637932777, + 0.0035182666033506393, + 0.004960065707564354, + 0.0033596977591514587, + 0.005514064803719521, + 0.003619828727096319, + 0.00293050822801888, + 0.003862968413159251, + 0.003974563907831907, + 0.004846037365496159, + 0.00681340554729104, + 0.008354980498552322, + 0.008053814992308617, + 0.012485948391258717, + 0.008542831987142563, + 0.007821034640073776, + 0.007939879782497883, + 0.008120529353618622, + 0.039862990379333496, + 0.004976144526153803, + 0.041033439338207245, + 0.005019765347242355, + 0.005653085187077522, + 0.011887977831065655, + 0.004504724405705929, + 0.00884885061532259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.17989473044872284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738202180713415 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738202180713415, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738202180713415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13540808856487274, + 0.11914777755737305, + 0.10123488306999207, + 0.08132682740688324, + 0.1811356544494629, + 0.04331834986805916, + 0.019320320338010788, + 0.040893636643886566, + 0.06623535603284836, + 0.21197906136512756 + ], + "credit_map": [ + 0.019414769485592842, + 0.014321474358439445, + 0.023546045646071434, + 0.015076708048582077, + 0.01306905597448349, + 0.019546106457710266, + 0.012269631028175354, + 0.018164299428462982, + 0.012336798943579197, + 0.01266133040189743, + 0.018985603004693985, + 0.011416669003665447, + 0.013923805207014084, + 0.01790592633187771, + 0.012564390897750854, + 0.019353259354829788, + 0.013278505764901638, + 0.011933807283639908, + 0.014075622893869877, + 0.010470304638147354, + 0.014452981762588024, + 0.008701510727405548, + 0.010346532799303532, + 0.017975619062781334, + 0.012142556719481945, + 0.004263218026608229, + 0.01692727953195572, + 0.006532594561576843, + 0.007956595160067081, + 0.005611323285847902, + 0.010106958448886871, + 0.0177862998098135, + 0.024894000962376595, + 0.007903825491666794, + 0.047347746789455414, + 0.009780027903616428, + 0.011161386966705322, + 0.012691158801317215, + 0.017556997016072273, + 0.049800530076026917, + 0.006580717861652374, + 0.00374072534032166, + 0.00826337467879057, + 0.004963481333106756, + 0.003691856749355793, + 0.0035992874763906, + 0.00534459576010704, + 0.007134310435503721, + 0.0031701894477009773, + 0.0020183303859084845, + 0.0027951952069997787, + 0.0022934922017157078, + 0.001828287960961461, + 0.0024309668224304914, + 0.0020364972297102213, + 0.0027473613154143095, + 0.010321585461497307, + 0.0018661501817405224, + 0.009379232302308083, + 0.0025575142353773117, + 0.004147731699049473, + 0.0027772511821240187, + 0.00391314597800374, + 0.00593102490529418, + 0.009334702976047993, + 0.006107694935053587, + 0.010579999536275864, + 0.008797782473266125, + 0.006976212840527296, + 0.006849785801023245, + 0.0077563864178955555, + 0.009832792915403843, + 0.07368013262748718, + 0.008951292373239994, + 0.07656051963567734, + 0.008272172883152962, + 0.008173766545951366, + 0.01601160131394863, + 0.007694731000810862, + 0.012634833343327045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.006738202180713415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.348622590303421 + ], + "opposite_logprob_tokens": [ + -1.2236225605010986 + ], + "logprob_reward": -0.348622590303421, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.348622590303421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07953722029924393, + 0.027446581050753593, + 0.01722586341202259, + 0.021256258711218834, + 0.05632452666759491, + 0.01811019517481327, + 0.04578046873211861, + 0.03506208956241608, + 0.1282588094472885, + 0.570997953414917 + ], + "credit_map": [ + 0.008022409863770008, + 0.01159795094281435, + 0.009425985626876354, + 0.01167050190269947, + 0.010643909685313702, + 0.010548912920057774, + 0.008697818964719772, + 0.00892973504960537, + 0.004031215328723192, + 0.0035887013655155897, + 0.003692039055749774, + 0.002889723517000675, + 0.003404709044843912, + 0.003472183831036091, + 0.00297134299762547, + 0.00339666660875082, + 0.0026171139907091856, + 0.0014740857295691967, + 0.002683840459212661, + 0.00191162817645818, + 0.0015694438479840755, + 0.0013630512403324246, + 0.0024159413296729326, + 0.0031907593365758657, + 0.003632522188127041, + 0.0016632800688967109, + 0.0031519075855612755, + 0.0018003903096541762, + 0.0028459017630666494, + 0.0012606448726728559, + 0.002722821431234479, + 0.004178790375590324, + 0.009800272062420845, + 0.0031605353578925133, + 0.011779595166444778, + 0.0027890594210475683, + 0.004004127811640501, + 0.004445971455425024, + 0.003209681948646903, + 0.01713528111577034, + 0.00280070211738348, + 0.001730503747239709, + 0.002486994955688715, + 0.0021505504846572876, + 0.001874457229860127, + 0.0020194307435303926, + 0.002310864394530654, + 0.0027366909198462963, + 0.008212050423026085, + 0.00479154521599412, + 0.007752740290015936, + 0.006020364351570606, + 0.0039011810440570116, + 0.0058526527136564255, + 0.003408747026696801, + 0.005841182544827461, + 0.007105562370270491, + 0.002625222085043788, + 0.009162016212940216, + 0.002918965881690383, + 0.003016362665221095, + 0.0029316090513020754, + 0.0037078573368489742, + 0.0035944953560829163, + 0.017242304980754852, + 0.013566115871071815, + 0.01692129671573639, + 0.020528072491288185, + 0.014881612733006477, + 0.013396252878010273, + 0.015468016266822815, + 0.01625513657927513, + 0.1897193044424057, + 0.027971483767032623, + 0.20126086473464966, + 0.025262128561735153, + 0.026105523109436035, + 0.044213809072971344, + 0.018940947949886322, + 0.03752387315034866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.348622590303421 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the hyper-accelerated, time-lapse blossoming of a Chicory 'Pumila' flower, its forms dominating the entire frame. Initially, the scene is bathed in the soft, diffused glow of natural morning light, highlighting the petals' stunningly vibrant purplish-red against a deeply out-of-focus, verdant green background. The camera remains steady as the tightly closed bud expands, its crepe-paper petals, rendered in microscopic detail to reveal every delicate wrinkle, begin to unfold with graceful and rapid vitality, revealing a cluster of bright yellow stamens at its core. Just as the blossom reaches its full, breathtaking splendor, this mesmerizing display of organic growth is interrupted. The entire screen flickers violently with digital interference; sharp, glitched horizontal scan lines race across the image, the perfectly purplish-red petals momentarily pixelate into rough blocks of color, and the frame is fractured by flickering three-color channel separation and white noise, creating a stark and unsettling contrast between the living world and abrupt technological decay.", + "seed": 32889, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_12/index_1249_seed_32889_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08410022691823542, + "vlm_logprob": -0.08410022691823542, + "vlm_yes_no_margin": 3.674999988079071, + "vlm_token_credit": -0.08410022691823542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020220572128891945 + ], + "opposite_logprob_tokens": [ + -4.395220756530762 + ], + "logprob_reward": -0.020220572128891945, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.020220572128891945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1360476315021515, + 0.04214385524392128, + 0.05278429016470909, + 0.10657414048910141, + 0.0471925288438797, + 0.030701979994773865, + 0.3048439919948578, + 0.14605343341827393, + 0.07397013157606125, + 0.059688035398721695 + ], + "credit_map": [ + 0.01535240188241005, + 0.015408202074468136, + 0.021957095712423325, + 0.01657271943986416, + 0.011234497651457787, + 0.02518526464700699, + 0.011297975666821003, + 0.019039472565054893, + 0.004370217211544514, + 0.009639722295105457, + 0.005890444852411747, + 0.004665149841457605, + 0.004217180889099836, + 0.005546330474317074, + 0.0035855243913829327, + 0.004229288548231125, + 0.011739702895283699, + 0.005381476134061813, + 0.008079815655946732, + 0.004819675348699093, + 0.004289369564503431, + 0.005280862096697092, + 0.0032184822484850883, + 0.009974902495741844, + 0.018500417470932007, + 0.010908112861216068, + 0.017500048503279686, + 0.013068562373518944, + 0.007483033463358879, + 0.012954320758581161, + 0.010409501381218433, + 0.015750151127576828, + 0.0057617062702775, + 0.005944924894720316, + 0.005851763300597668, + 0.006337528582662344, + 0.00532336113974452, + 0.004358576610684395, + 0.005623390898108482, + 0.007991277612745762, + 0.003744886489585042, + 0.004684096667915583, + 0.003114798804745078, + 0.004746490623801947, + 0.0033295329194515944, + 0.0032956942450255156, + 0.003367862431332469, + 0.004418617580085993, + 0.031758297234773636, + 0.04014008119702339, + 0.01779589056968689, + 0.037444040179252625, + 0.05073733255267143, + 0.03327492997050285, + 0.058389343321323395, + 0.03530407324433327, + 0.03143129497766495, + 0.012315068393945694, + 0.036083221435546875, + 0.012249224819242954, + 0.014562069438397884, + 0.013932530768215656, + 0.01202169619500637, + 0.013458327390253544, + 0.007213963661342859, + 0.020253552123904228, + 0.01043712068349123, + 0.013385730795562267, + 0.005248771049082279, + 0.00556894950568676, + 0.006910261232405901, + 0.004951782990247011, + 0.005093256942927837, + 0.008278992027044296, + 0.0093558169901371, + 0.00858362391591072, + 0.004742054268717766, + 0.0051624709740281105, + 0.009660410694777966, + 0.008811414241790771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.020220572128891945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029860801994800568 + ], + "opposite_logprob_tokens": [ + -3.529860734939575 + ], + "logprob_reward": -0.029860801994800568, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029860801994800568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06670916080474854, + 0.02473888359963894, + 0.09337391704320908, + 0.1532686948776245, + 0.050173625349998474, + 0.022046254947781563, + 0.18947052955627441, + 0.057874083518981934, + 0.2479332536458969, + 0.09441154450178146 + ], + "credit_map": [ + 0.009153016842901707, + 0.007018123287707567, + 0.010883969254791737, + 0.0068313321098685265, + 0.005485964473336935, + 0.01460291538387537, + 0.0038067202549427748, + 0.00892713014036417, + 0.0021002073772251606, + 0.0028778668493032455, + 0.003411071840673685, + 0.00258983438834548, + 0.0036020376719534397, + 0.004251481033861637, + 0.003250911831855774, + 0.002655474003404379, + 0.012976252473890781, + 0.011089649051427841, + 0.01231409516185522, + 0.013759367167949677, + 0.011411609128117561, + 0.00917917862534523, + 0.010923677124083042, + 0.011720099486410618, + 0.021153824403882027, + 0.014694162644445896, + 0.024298811331391335, + 0.018609916791319847, + 0.011605408973991871, + 0.02037258818745613, + 0.020183971151709557, + 0.022350013256072998, + 0.007023909594863653, + 0.005244349595159292, + 0.007983870804309845, + 0.005809322465211153, + 0.006229877006262541, + 0.0046148356050252914, + 0.005917624104768038, + 0.007349842227995396, + 0.002893236465752125, + 0.001960695255547762, + 0.002462871838361025, + 0.003021531505510211, + 0.0036686609964817762, + 0.0018952024402096868, + 0.0033531738445162773, + 0.002790886675938964, + 0.018467042595148087, + 0.024590926244854927, + 0.011662092991173267, + 0.024495910853147507, + 0.03346238657832146, + 0.0227985642850399, + 0.03890407085418701, + 0.015089563094079494, + 0.012143507599830627, + 0.005011436063796282, + 0.012545268051326275, + 0.006403862964361906, + 0.006576556712388992, + 0.003883691970258951, + 0.005787402391433716, + 0.005522365681827068, + 0.04167095944285393, + 0.048760946840047836, + 0.06526734679937363, + 0.025947101414203644, + 0.014542871154844761, + 0.01763807237148285, + 0.020408615469932556, + 0.013697349466383457, + 0.009953673928976059, + 0.00888083130121231, + 0.014391838572919369, + 0.012552035041153431, + 0.00641310540959239, + 0.007843264378607273, + 0.016708066686987877, + 0.01766873709857464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.029860801994800568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486190438270569 + ], + "opposite_logprob_tokens": [ + -1.223618984222412 + ], + "logprob_reward": -0.3486190438270569, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486190438270569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08000043779611588, + 0.07289212197065353, + 0.12464841455221176, + 0.2358112782239914, + 0.04524648189544678, + 0.03803382068872452, + 0.2746434807777405, + 0.02391047775745392, + 0.07503422349691391, + 0.029779264703392982 + ], + "credit_map": [ + 0.01296039018779993, + 0.008116011507809162, + 0.018241383135318756, + 0.006777615286409855, + 0.007428909186273813, + 0.010450663976371288, + 0.007913346402347088, + 0.008112114854156971, + 0.008859301917254925, + 0.008456516079604626, + 0.010680485516786575, + 0.007307150866836309, + 0.009618064388632774, + 0.010366080328822136, + 0.009385877288877964, + 0.008218645118176937, + 0.023996565490961075, + 0.015864482149481773, + 0.017794402316212654, + 0.012807386927306652, + 0.012484646402299404, + 0.012931426987051964, + 0.010375107638537884, + 0.01839440129697323, + 0.034037791192531586, + 0.026523532345891, + 0.03702665492892265, + 0.0318000391125679, + 0.01820332556962967, + 0.03008771687746048, + 0.026770049706101418, + 0.031362179666757584, + 0.0052705854177474976, + 0.0077186948619782925, + 0.006395930424332619, + 0.005501576233655214, + 0.004355631768703461, + 0.005303933285176754, + 0.003978351596742868, + 0.006721780635416508, + 0.006457820069044828, + 0.004060694947838783, + 0.005296404007822275, + 0.005719935055822134, + 0.005086857825517654, + 0.003774979617446661, + 0.0036449481267482042, + 0.003992179874330759, + 0.027842236682772636, + 0.034565191715955734, + 0.01823253743350506, + 0.03466111049056053, + 0.044665850698947906, + 0.03466656804084778, + 0.05332046002149582, + 0.026689527556300163, + 0.002999385353177786, + 0.002895544981583953, + 0.003467555856332183, + 0.0030214537400752306, + 0.0032445755787193775, + 0.0018312643514946103, + 0.00316446996293962, + 0.0032862264197319746, + 0.010938639752566814, + 0.015381942503154278, + 0.017360709607601166, + 0.009655935689806938, + 0.004974242765456438, + 0.00463027274236083, + 0.00779630895704031, + 0.004296171944588423, + 0.003205867949873209, + 0.0074543156661093235, + 0.003508505877107382, + 0.0029026817064732313, + 0.0022472618147730827, + 0.002581024309620261, + 0.004936130717396736, + 0.002943479223176837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.3486190438270569 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616105066612363 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.003616105066612363, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616105066612363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06182928383350372, + 0.13424326479434967, + 0.04023858532309532, + 0.14918427169322968, + 0.046072714030742645, + 0.05209130421280861, + 0.22489403188228607, + 0.04309186711907387, + 0.13767868280410767, + 0.11067596822977066 + ], + "credit_map": [ + 0.007446326781064272, + 0.00758113618940115, + 0.007735501974821091, + 0.008287658914923668, + 0.0070553426630795, + 0.007394143380224705, + 0.0066979872062802315, + 0.009631184861063957, + 0.014551816508173943, + 0.019858187064528465, + 0.02023477852344513, + 0.013267701491713524, + 0.014159196056425571, + 0.017322681844234467, + 0.01881743036210537, + 0.016031479462981224, + 0.004921278450638056, + 0.004517061170190573, + 0.006577071733772755, + 0.0054288520477712154, + 0.004458093084394932, + 0.004362446255981922, + 0.004625389352440834, + 0.005348394624888897, + 0.03312338888645172, + 0.012134979479014874, + 0.028484901413321495, + 0.016408689320087433, + 0.008669177070260048, + 0.015788452699780464, + 0.012344710528850555, + 0.022229958325624466, + 0.0066340649500489235, + 0.006520750001072884, + 0.006642769556492567, + 0.005405543837696314, + 0.004342882893979549, + 0.004520298447459936, + 0.004086227156221867, + 0.007920179516077042, + 0.008926453068852425, + 0.004697522148489952, + 0.008645392954349518, + 0.0076714083552360535, + 0.0052216434851288795, + 0.00395961245521903, + 0.005178452469408512, + 0.007790816482156515, + 0.023864010348916054, + 0.028179118409752846, + 0.013487352058291435, + 0.02894510328769684, + 0.03735647723078728, + 0.025051157921552658, + 0.04230227321386337, + 0.025708533823490143, + 0.00658030342310667, + 0.006299983710050583, + 0.008690621703863144, + 0.004178034141659737, + 0.004109721630811691, + 0.003935778513550758, + 0.004983020946383476, + 0.004314403049647808, + 0.021898847073316574, + 0.02913415990769863, + 0.03559678792953491, + 0.015354270115494728, + 0.008346919901669025, + 0.008488788269460201, + 0.011251700110733509, + 0.007607218809425831, + 0.031633924692869186, + 0.0107443667948246, + 0.03379524126648903, + 0.007037903182208538, + 0.004367225803434849, + 0.006349675823003054, + 0.009100216440856457, + 0.007647417951375246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.003616105066612363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the flower dominates the entire frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the flower a Chicory 'Pumila' with purplish-red petals and bright yellow stamens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the flower blooming in a hyper-accelerated, time-lapse manner with crepe-paper texture details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deeply out-of-focus verdant green scene lit by soft morning light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a violent digital glitch featuring horizontal scan lines, pixelation, and color channel separation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184611573815346 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184611573815346, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184611573815346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10633435845375061, + 0.03222811222076416, + 0.02495790459215641, + 0.08295735716819763, + 0.024495447054505348, + 0.05625384300947189, + 0.07825420796871185, + 0.043855931609869, + 0.05870066583156586, + 0.491962194442749 + ], + "credit_map": [ + 0.014329995959997177, + 0.013313797302544117, + 0.024270661175251007, + 0.010611307807266712, + 0.008650178089737892, + 0.01582590490579605, + 0.009900568053126335, + 0.00943194143474102, + 0.0030119786970317364, + 0.005508491303771734, + 0.003930376376956701, + 0.0029905387200415134, + 0.002963731763884425, + 0.004816897679120302, + 0.003915364854037762, + 0.005090734921395779, + 0.004091772250831127, + 0.003045987104997039, + 0.003512053983286023, + 0.002965793712064624, + 0.0023922789841890335, + 0.002960504963994026, + 0.002657672856003046, + 0.003331840969622135, + 0.008326984010636806, + 0.008994210511446, + 0.012613843195140362, + 0.012076926417648792, + 0.005803556181490421, + 0.011046700179576874, + 0.014112792909145355, + 0.00998234748840332, + 0.004983256570994854, + 0.002151095774024725, + 0.005324933212250471, + 0.003274454502388835, + 0.0024062301963567734, + 0.0016665582079440355, + 0.0024791413452476263, + 0.002209777943789959, + 0.008811518549919128, + 0.00563324149698019, + 0.007894625887274742, + 0.007528801914304495, + 0.008574617095291615, + 0.004444488324224949, + 0.006482129450887442, + 0.006884419359266758, + 0.00817370880395174, + 0.010066434741020203, + 0.005058621987700462, + 0.009464145638048649, + 0.012573400512337685, + 0.00864764116704464, + 0.016667751595377922, + 0.0076025077141821384, + 0.004675169009715319, + 0.008589128032326698, + 0.0042142122983932495, + 0.006312623154371977, + 0.005450206808745861, + 0.003982228692620993, + 0.005273535382002592, + 0.005358828231692314, + 0.010353912599384785, + 0.013262809254229069, + 0.007328417152166367, + 0.00903808418661356, + 0.003942461218684912, + 0.0047508166171610355, + 0.005352898500859737, + 0.004671265836805105, + 0.058786358684301376, + 0.05541004240512848, + 0.0902349129319191, + 0.05609705299139023, + 0.031197115778923035, + 0.036295436322689056, + 0.08234383165836334, + 0.08159743249416351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.018184611573815346 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_tensor([4851, 4851])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_tensor([4851, 4851])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..344c81cb62e6ec2d21030b9f6502b61e5c3d7333 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_tensor([4851, 4851])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static shot, captured under the intense glare of the stadium's floodlights, captures a moment of pure athletic power. A classic black-and-white paneled football, a violently spinning blur, is launched with explosive force into the dead center of the goal's upper crossbar. The impact is instantaneous and brutal; the ball deforms for a split second against the white-painted metal before violently rebounding onto the field. A deep, resonant tremor perceptibly shakes the entire goal frame, a powerful vibration that instantly transfers to the goal's white hexagonal grid, erupting in chaotic, wave-like vibrations. The entire structure, from crossbar to jambs and corrugated netting, shakes violently for several seconds, creating a dramatic visual echo of the powerful blow. The scene is bathed in the cold, bright lights of the stadium, casting sharp highlights on the metal and deep, dancing shadows within the trembling netting, creating an atmosphere of tense and dramatic near-misses.", + "seed": 14533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_4851_seed_14533_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.43123782947659495, + "vlm_logprob": -0.43123782947659495, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.43123782947659495, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06304614990949631 + ], + "opposite_logprob_tokens": [ + -3.5630462169647217 + ], + "logprob_reward": -0.06304614990949631, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.06304614990949631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16442668437957764, + 0.04885367304086685, + 0.07807546108961105, + 0.048468124121427536, + 0.2738211154937744, + 0.11029895395040512, + 0.13236267864704132, + 0.0662882924079895, + 0.030106427147984505, + 0.04729863256216049 + ], + "credit_map": [ + 0.02425239607691765, + 0.019647780805826187, + 0.012810524553060532, + 0.016291555017232895, + 0.021290434524416924, + 0.015471384860575199, + 0.024143891409039497, + 0.03051871433854103, + 0.008851272985339165, + 0.006814241409301758, + 0.006039017345756292, + 0.0040718065574765205, + 0.006004324182868004, + 0.005222158972173929, + 0.0056290156207978725, + 0.006221833638846874, + 0.012342086061835289, + 0.0129804452881217, + 0.010768342763185501, + 0.006704242434352636, + 0.008168607018887997, + 0.006204513367265463, + 0.008472072891891003, + 0.012435157783329487, + 0.005490189418196678, + 0.00807988177984953, + 0.006810558494180441, + 0.005478783044964075, + 0.005937807261943817, + 0.004167905077338219, + 0.005336245056241751, + 0.00716674979776144, + 0.023866694420576096, + 0.033108875155448914, + 0.01548729557543993, + 0.024564214050769806, + 0.08702245354652405, + 0.022468214854598045, + 0.02893208898603916, + 0.03837127983570099, + 0.010371183976531029, + 0.008212379179894924, + 0.0038865271490067244, + 0.00957396999001503, + 0.028727425262331963, + 0.016344193369150162, + 0.01664113625884056, + 0.016542131081223488, + 0.011325136758387089, + 0.016507035121321678, + 0.015098683536052704, + 0.01587817817926407, + 0.02209157682955265, + 0.019374247640371323, + 0.015627821907401085, + 0.016460007056593895, + 0.010451395995914936, + 0.005919487681239843, + 0.004534949082881212, + 0.010968978516757488, + 0.009248370304703712, + 0.009434145875275135, + 0.005447936709970236, + 0.010283032432198524, + 0.003986040595918894, + 0.0028807595372200012, + 0.0030941430013626814, + 0.005109283607453108, + 0.004044516943395138, + 0.004174498375505209, + 0.0031495008151978254, + 0.003667687065899372, + 0.003907293081283569, + 0.00681940047070384, + 0.004353691823780537, + 0.006599669810384512, + 0.0064756739884614944, + 0.0060522002167999744, + 0.007171666249632835, + 0.0059190369211137295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.06304614990949631 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103795289993286 + ], + "opposite_logprob_tokens": [ + -0.1603795737028122 + ], + "logprob_reward": -1.9103795289993286, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103795289993286, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03598111867904663, + 0.04106741026043892, + 0.035891905426979065, + 0.019046185538172722, + 0.527042031288147, + 0.08755777031183243, + 0.09864958375692368, + 0.08850148320198059, + 0.03407784923911095, + 0.032184623181819916 + ], + "credit_map": [ + 0.004969682544469833, + 0.004819687455892563, + 0.004079059232026339, + 0.004658723250031471, + 0.0036945671308785677, + 0.0030584216583520174, + 0.0041818986646831036, + 0.006519084796309471, + 0.008053875528275967, + 0.0037993143778294325, + 0.00357669685035944, + 0.004341975785791874, + 0.003431629156693816, + 0.004917260259389877, + 0.00450624106451869, + 0.008440421894192696, + 0.005343327764421701, + 0.005520495120435953, + 0.005115242209285498, + 0.0030190451070666313, + 0.0033919818233698606, + 0.003777913749217987, + 0.0037529312539845705, + 0.005970971658825874, + 0.0022872518748044968, + 0.002856977516785264, + 0.0017318740719929338, + 0.001605343073606491, + 0.004355479497462511, + 0.001910124788992107, + 0.00201325467787683, + 0.0022858830634504557, + 0.04726063460111618, + 0.053101930767297745, + 0.03315902128815651, + 0.04668745771050453, + 0.1623220592737198, + 0.04791444540023804, + 0.06835473328828812, + 0.06824184209108353, + 0.008712580427527428, + 0.00939662754535675, + 0.005594453774392605, + 0.012544771656394005, + 0.014181159436702728, + 0.013103068806231022, + 0.01065266877412796, + 0.013372449204325676, + 0.008837433531880379, + 0.015459067188203335, + 0.007545490749180317, + 0.013742716051638126, + 0.013307997956871986, + 0.012373792007565498, + 0.012031659483909607, + 0.015351426787674427, + 0.01026590820401907, + 0.005112086422741413, + 0.004933332558721304, + 0.009348948486149311, + 0.00962818507105112, + 0.013223258778452873, + 0.008039779029786587, + 0.027950001880526543, + 0.0036508431658148766, + 0.003822511062026024, + 0.004297648090869188, + 0.003864697180688381, + 0.005279719829559326, + 0.00532275578007102, + 0.0028918341267853975, + 0.004947844427078962, + 0.002953587332740426, + 0.0052224742248654366, + 0.0031781690195202827, + 0.004191088955849409, + 0.005060152150690556, + 0.004323494620621204, + 0.0036115918774157763, + 0.0036440682597458363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -1.9103795289993286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899990677833557 + ], + "opposite_logprob_tokens": [ + -2.5789999961853027 + ], + "logprob_reward": -0.07899990677833557, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899990677833557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049235764890909195, + 0.05660667270421982, + 0.03943014517426491, + 0.02116716094315052, + 0.38631364703178406, + 0.16932974755764008, + 0.15538015961647034, + 0.05572836473584175, + 0.03529703617095947, + 0.03151128813624382 + ], + "credit_map": [ + 0.0077301086857914925, + 0.006017901469022036, + 0.0040937927551567554, + 0.004973117727786303, + 0.004193505737930536, + 0.006456837523728609, + 0.004769421182572842, + 0.011001077480614185, + 0.012227237224578857, + 0.004056305158883333, + 0.0030637767631560564, + 0.0032911188900470734, + 0.002986654406413436, + 0.005453555844724178, + 0.007919608615338802, + 0.01760842278599739, + 0.006038331426680088, + 0.0047870720736682415, + 0.003183532739058137, + 0.0037807649932801723, + 0.003721795277670026, + 0.004541509319096804, + 0.004950911272317171, + 0.008426224812865257, + 0.0030194749124348164, + 0.0030957909766584635, + 0.0017086048610508442, + 0.0015423351433128119, + 0.00526817049831152, + 0.0018031737999990582, + 0.0022585378028452396, + 0.0024710739962756634, + 0.03345845639705658, + 0.0376897007226944, + 0.02425605058670044, + 0.03236474096775055, + 0.11990022659301758, + 0.037034183740615845, + 0.052875593304634094, + 0.04873468354344368, + 0.017165739089250565, + 0.017018955200910568, + 0.0057233283296227455, + 0.014826824888586998, + 0.04052719846367836, + 0.020954562351107597, + 0.027521979063749313, + 0.025591161102056503, + 0.015223772265017033, + 0.019739622250199318, + 0.014351691119372845, + 0.021405162289738655, + 0.023840581998229027, + 0.020031573250889778, + 0.017533879727125168, + 0.023253872990608215, + 0.005846000276505947, + 0.004934433847665787, + 0.004587940871715546, + 0.006410979200154543, + 0.006715098861604929, + 0.007562794256955385, + 0.005967707373201847, + 0.01370340958237648, + 0.005176462233066559, + 0.0035871858708560467, + 0.00434108404442668, + 0.003659590380266309, + 0.00451063783839345, + 0.005389855243265629, + 0.003248186083510518, + 0.00538403308019042, + 0.002849123440682888, + 0.005811921786516905, + 0.0033012828789651394, + 0.0033088945783674717, + 0.005461279768496752, + 0.004405459854751825, + 0.0032183018047362566, + 0.0031550258863717318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.07899990677833557 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006027549505234 + ], + "opposite_logprob_tokens": [ + -2.6950602531433105 + ], + "logprob_reward": -0.07006027549505234, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006027549505234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2648136019706726, + 0.090394526720047, + 0.08396277576684952, + 0.042271461337804794, + 0.08153504133224487, + 0.13300827145576477, + 0.08084188401699066, + 0.10062311589717865, + 0.08605412393808365, + 0.0364951528608799 + ], + "credit_map": [ + 0.02824777364730835, + 0.03038974292576313, + 0.032633379101753235, + 0.03124934621155262, + 0.024072786793112755, + 0.033064790070056915, + 0.03554614260792732, + 0.04960965737700462, + 0.017812218517065048, + 0.009929191321134567, + 0.004688194487243891, + 0.005845931358635426, + 0.007986974902451038, + 0.009614965878427029, + 0.00813642144203186, + 0.026380641385912895, + 0.0109647735953331, + 0.010205374099314213, + 0.010459315031766891, + 0.010499021038413048, + 0.009542710147798061, + 0.005894671194255352, + 0.013839966617524624, + 0.012556951493024826, + 0.006200790870934725, + 0.0054142591543495655, + 0.004564667120575905, + 0.004159340634942055, + 0.005927376449108124, + 0.004124175291508436, + 0.005289772525429726, + 0.006591083947569132, + 0.008945665322244167, + 0.011215328238904476, + 0.006172915454953909, + 0.010512139648199081, + 0.01835528202354908, + 0.006415449548512697, + 0.008459719829261303, + 0.011458544060587883, + 0.008140628226101398, + 0.01061769388616085, + 0.007778672035783529, + 0.014592457562685013, + 0.03374823182821274, + 0.014762814156711102, + 0.02681848779320717, + 0.016549300402402878, + 0.006829429883509874, + 0.008700475096702576, + 0.006147268693894148, + 0.013079330325126648, + 0.00972018763422966, + 0.010770824737846851, + 0.010095963254570961, + 0.015498410910367966, + 0.013298982754349709, + 0.008218656294047832, + 0.006468798965215683, + 0.0093298414722085, + 0.009283461607992649, + 0.014692280441522598, + 0.010703429579734802, + 0.028627678751945496, + 0.00965195894241333, + 0.009827245958149433, + 0.00917512271553278, + 0.009163904935121536, + 0.014476071111857891, + 0.012852719984948635, + 0.0076377056539058685, + 0.013269403018057346, + 0.0027675677556544542, + 0.004435581620782614, + 0.0029388496186584234, + 0.0032201355788856745, + 0.008099020458757877, + 0.007798569742590189, + 0.0038059395737946033, + 0.003429492935538292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.07006027549505234 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033703286200761795 + ], + "opposite_logprob_tokens": [ + -3.408703327178955 + ], + "logprob_reward": -0.033703286200761795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033703286200761795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.079833023250103, + 0.07785845547914505, + 0.14791157841682434, + 0.2922266721725464, + 0.14158444106578827, + 0.034060776233673096, + 0.046125710010528564, + 0.06535720825195312, + 0.06034413352608681, + 0.05469801649451256 + ], + "credit_map": [ + 0.007989413104951382, + 0.011881585232913494, + 0.009543507359921932, + 0.011844195425510406, + 0.007572214584797621, + 0.006267016287893057, + 0.011754713952541351, + 0.012980377301573753, + 0.0099002905189991, + 0.010953439399600029, + 0.007191606797277927, + 0.007899187505245209, + 0.006075771991163492, + 0.008705604821443558, + 0.011524828150868416, + 0.015607720240950584, + 0.02192912995815277, + 0.018665524199604988, + 0.01187007874250412, + 0.012743180617690086, + 0.01919071562588215, + 0.016151541844010353, + 0.01819377765059471, + 0.029167626053094864, + 0.03765697777271271, + 0.06730232387781143, + 0.025853347033262253, + 0.029295112937688828, + 0.04327695071697235, + 0.027496948838233948, + 0.031190045177936554, + 0.030154960229992867, + 0.01471103634685278, + 0.01729562319815159, + 0.013047438114881516, + 0.015144595876336098, + 0.02936391904950142, + 0.012461413629353046, + 0.01988871954381466, + 0.019671699032187462, + 0.0017110946355387568, + 0.002984452061355114, + 0.0015716480556875467, + 0.004470787476748228, + 0.009582891128957272, + 0.003464985406026244, + 0.005852807778865099, + 0.004422110505402088, + 0.003816501470282674, + 0.004869802389293909, + 0.0033740608487278223, + 0.005979260429739952, + 0.010231641121208668, + 0.0071364110335707664, + 0.005106443539261818, + 0.005611584987491369, + 0.007726721931248903, + 0.00492833461612463, + 0.003970063291490078, + 0.007584717590361834, + 0.011684294790029526, + 0.008826234377920628, + 0.005403253249824047, + 0.015233585610985756, + 0.008680411614477634, + 0.0072358399629592896, + 0.007193948607891798, + 0.005746443755924702, + 0.010140357539057732, + 0.008678215555846691, + 0.005257355980575085, + 0.007411558646708727, + 0.003936482593417168, + 0.005181756801903248, + 0.003490748815238476, + 0.00553085608407855, + 0.013951919972896576, + 0.013990108855068684, + 0.004313526209443808, + 0.004302620887756348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.033703286200761795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot, captured under the intense glare of the stadium's floodlights, captures a moment of pure athletic power. A classic black-and-white paneled football, a violently spinning blur, is launched with explosive force into the dead center of the goal's upper crossbar. The impact is instantaneous and brutal; the ball deforms for a split second against the white-painted metal before violently rebounding onto the field. A deep, resonant tremor perceptibly shakes the entire goal frame, a powerful vibration that instantly transfers to the goal's white hexagonal grid, erupting in chaotic, wave-like vibrations. The entire structure, from crossbar to jambs and corrugated netting, shakes violently for several seconds, creating a dramatic visual echo of the powerful blow. The scene is bathed in the cold, bright lights of the stadium, casting sharp highlights on the metal and deep, dancing shadows within the trembling netting, creating an atmosphere of tense and dramatic near-misses.", + "seed": 14534, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_4851_seed_14534_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5197252199053765, + "vlm_logprob": -0.5197252199053765, + "vlm_yes_no_margin": 1.250000011920929, + "vlm_token_credit": -0.5197252199053765, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1715986579656601 + ], + "opposite_logprob_tokens": [ + -2.2965986728668213 + ], + "logprob_reward": -0.1715986579656601, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1715986579656601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3607195317745209, + 0.08549194782972336, + 0.03928638994693756, + 0.039017677307128906, + 0.06718764454126358, + 0.05997094511985779, + 0.10713318735361099, + 0.07717428356409073, + 0.09268584102392197, + 0.07133258879184723 + ], + "credit_map": [ + 0.06557878851890564, + 0.030974632129073143, + 0.07241848856210709, + 0.03702399507164955, + 0.031206725165247917, + 0.03790678083896637, + 0.03751358762383461, + 0.04809652268886566, + 0.014889775775372982, + 0.00525305699557066, + 0.0168170016258955, + 0.008561459369957447, + 0.00840107724070549, + 0.010372483171522617, + 0.009939686395227909, + 0.011257399804890156, + 0.005911697633564472, + 0.002791518345475197, + 0.003762658452615142, + 0.003979452885687351, + 0.008126563392579556, + 0.004789139609783888, + 0.005656475201249123, + 0.004268883261829615, + 0.006166923325508833, + 0.002843628404662013, + 0.004023713991045952, + 0.004025057423859835, + 0.00864322204142809, + 0.005701834335923195, + 0.0034560984931886196, + 0.0041571990586817265, + 0.008664059452712536, + 0.004356138873845339, + 0.0038270854856818914, + 0.003813190385699272, + 0.0074581485241651535, + 0.010701517574489117, + 0.006972582079470158, + 0.021394919604063034, + 0.010884623043239117, + 0.002605607733130455, + 0.0033625054638832808, + 0.0035747678484767675, + 0.005056232213973999, + 0.019021252170205116, + 0.007123298943042755, + 0.008342660963535309, + 0.027991052716970444, + 0.005056075286120176, + 0.008117466233670712, + 0.005944577977061272, + 0.007946619763970375, + 0.009067758917808533, + 0.00965662021189928, + 0.033353015780448914, + 0.011522121727466583, + 0.008513567969202995, + 0.01054272148758173, + 0.0072599793784320354, + 0.005242376122623682, + 0.00730369845405221, + 0.016613127663731575, + 0.010176693089306355, + 0.014708233997225761, + 0.03427523002028465, + 0.022450413554906845, + 0.004849475808441639, + 0.0032651317305862904, + 0.00281083257868886, + 0.0036583051551133394, + 0.006668215151876211, + 0.008342632092535496, + 0.021630501374602318, + 0.01613307185471058, + 0.005567611660808325, + 0.0039914753288030624, + 0.0050573330372571945, + 0.005990049336105585, + 0.004619915504008532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.1715986579656601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.91036856174469 + ], + "opposite_logprob_tokens": [ + -0.16036860644817352 + ], + "logprob_reward": -1.91036856174469, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.91036856174469, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15717801451683044, + 0.12323857098817825, + 0.03132188692688942, + 0.061969075351953506, + 0.1254020780324936, + 0.10641269385814667, + 0.08109156787395477, + 0.11002306640148163, + 0.1115107536315918, + 0.0918523296713829 + ], + "credit_map": [ + 0.03144461289048195, + 0.007667026482522488, + 0.030931122601032257, + 0.011739526875317097, + 0.018500514328479767, + 0.01558067835867405, + 0.021323014050722122, + 0.019991517066955566, + 0.019780302420258522, + 0.009274564683437347, + 0.01676573045551777, + 0.014844207093119621, + 0.013507398776710033, + 0.01628066599369049, + 0.013312027789652348, + 0.019473664462566376, + 0.003800914390012622, + 0.001794443465769291, + 0.003991606179624796, + 0.0030557706486433744, + 0.007545359432697296, + 0.003805043175816536, + 0.0035303847398608923, + 0.003798365592956543, + 0.015674497932195663, + 0.002486596815288067, + 0.004024145193397999, + 0.003004923230037093, + 0.005539409350603819, + 0.01757805235683918, + 0.00584858562797308, + 0.007812860421836376, + 0.024828849360346794, + 0.004179623909294605, + 0.005548068322241306, + 0.005241617560386658, + 0.011483822949230671, + 0.03232654929161072, + 0.022522175684571266, + 0.019271356984972954, + 0.013691248372197151, + 0.008014040999114513, + 0.011075175367295742, + 0.010971816256642342, + 0.01251996960490942, + 0.014888300560414791, + 0.02040044404566288, + 0.014851693995296955, + 0.02276322990655899, + 0.004022225737571716, + 0.005171052645891905, + 0.004250416532158852, + 0.005735964048653841, + 0.006637032609432936, + 0.006451128050684929, + 0.026060519739985466, + 0.02036370150744915, + 0.017690924927592278, + 0.016393575817346573, + 0.009401450864970684, + 0.00844185333698988, + 0.012957786209881306, + 0.01222289726138115, + 0.012550866231322289, + 0.0161006897687912, + 0.03897401690483093, + 0.027086298912763596, + 0.009858954697847366, + 0.005706157069653273, + 0.0025066412054002285, + 0.0044271741062402725, + 0.006850815378129482, + 0.013786863535642624, + 0.02676001749932766, + 0.022301968187093735, + 0.005714226048439741, + 0.003753750817850232, + 0.006426930893212557, + 0.006671948824077845, + 0.006436619907617569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -1.91036856174469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299154907464981 + ], + "opposite_logprob_tokens": [ + -2.2379915714263916 + ], + "logprob_reward": -0.11299154907464981, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299154907464981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719556599855423, + 0.07332050055265427, + 0.03401188552379608, + 0.027106737717986107, + 0.18232189118862152, + 0.18489675223827362, + 0.1578749120235443, + 0.050233833491802216, + 0.08209868520498276, + 0.03617916628718376 + ], + "credit_map": [ + 0.034324195235967636, + 0.010258560068905354, + 0.0386790931224823, + 0.014119034633040428, + 0.016994601115584373, + 0.014186475425958633, + 0.01974179409444332, + 0.02365189976990223, + 0.011424042284488678, + 0.005521243438124657, + 0.010857847519218922, + 0.007904975675046444, + 0.010642831213772297, + 0.009738141670823097, + 0.007487590424716473, + 0.009743827395141125, + 0.004479491617530584, + 0.0025324076414108276, + 0.003111375728622079, + 0.003812731709331274, + 0.008167533203959465, + 0.004806194920092821, + 0.0041383714415133, + 0.0029637799598276615, + 0.006309357471764088, + 0.001419772976078093, + 0.0024546028580516577, + 0.002422598423436284, + 0.004258162807673216, + 0.005498368293046951, + 0.002451120177283883, + 0.002292754128575325, + 0.037527475506067276, + 0.00475342757999897, + 0.007178290747106075, + 0.00578919006511569, + 0.01091825496405363, + 0.02634095400571823, + 0.031114308163523674, + 0.05870000272989273, + 0.036672476679086685, + 0.008063462562859058, + 0.010941339656710625, + 0.010755476541817188, + 0.020894164219498634, + 0.062100667506456375, + 0.019386479631066322, + 0.016082677990198135, + 0.039031755179166794, + 0.007072858978062868, + 0.01166120357811451, + 0.009265393018722534, + 0.013261450454592705, + 0.013634719885885715, + 0.016240539029240608, + 0.04770699515938759, + 0.007249284069985151, + 0.007462493609637022, + 0.0070327650755643845, + 0.007218263112008572, + 0.0036465036682784557, + 0.003956309985369444, + 0.008734996430575848, + 0.004933217540383339, + 0.013503007590770721, + 0.027039838954806328, + 0.018128326162695885, + 0.006969242822378874, + 0.003738946979865432, + 0.003277319483458996, + 0.0032753674313426018, + 0.006166632752865553, + 0.004566681105643511, + 0.008481688797473907, + 0.007428276818245649, + 0.003729363204911351, + 0.0023010107688605785, + 0.0037138324696570635, + 0.0030276072211563587, + 0.0029307063668966293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.11299154907464981 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34869712591171265 + ], + "opposite_logprob_tokens": [ + -1.2236971855163574 + ], + "logprob_reward": -0.34869712591171265, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34869712591171265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24980364739894867, + 0.09803443402051926, + 0.06711888313293457, + 0.021993065252900124, + 0.10907892882823944, + 0.10130985826253891, + 0.11511155217885971, + 0.09736908227205276, + 0.0871528834104538, + 0.05302763357758522 + ], + "credit_map": [ + 0.04563146084547043, + 0.017338454723358154, + 0.04273458942770958, + 0.02390371635556221, + 0.022581644356250763, + 0.02328270673751831, + 0.04310377314686775, + 0.031227312982082367, + 0.017473755404353142, + 0.008289463818073273, + 0.014652489684522152, + 0.013279023580253124, + 0.010562211275100708, + 0.010531843639910221, + 0.010604065842926502, + 0.012641578912734985, + 0.0087276641279459, + 0.0040706791914999485, + 0.00589065020903945, + 0.007164740469306707, + 0.0201187152415514, + 0.006920474581420422, + 0.007760381326079369, + 0.006465577986091375, + 0.002322186017408967, + 0.0015384103171527386, + 0.0022696491796523333, + 0.003157959785312414, + 0.00333036994561553, + 0.0033023525029420853, + 0.0032855968456715345, + 0.0027865406591445208, + 0.010246303863823414, + 0.00399767654016614, + 0.0060622054152190685, + 0.0057394420728087425, + 0.011752679012715816, + 0.0295440461486578, + 0.014560963958501816, + 0.02717561274766922, + 0.013121634721755981, + 0.008620047010481358, + 0.008287019096314907, + 0.010360224172472954, + 0.013490143232047558, + 0.013810456730425358, + 0.01816636510193348, + 0.015453966334462166, + 0.026411766186356544, + 0.005199534818530083, + 0.009434583596885204, + 0.0074453940615057945, + 0.009950939565896988, + 0.00962037593126297, + 0.013808383606374264, + 0.03324057534337044, + 0.013926290906965733, + 0.02214890718460083, + 0.01846994459629059, + 0.008047884330153465, + 0.005458137020468712, + 0.006430729292333126, + 0.014950040727853775, + 0.007937155663967133, + 0.01041830237954855, + 0.022516528144478798, + 0.018939511850476265, + 0.009101585485041142, + 0.007347651291638613, + 0.004434136673808098, + 0.005994561593979597, + 0.00840060319751501, + 0.00673146964982152, + 0.015863096341490746, + 0.011657853610813618, + 0.004622464533895254, + 0.0025370419025421143, + 0.0035625810269266367, + 0.004475747235119343, + 0.003577379509806633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.34869712591171265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497020483016968 + ], + "opposite_logprob_tokens": [ + -2.9299702644348145 + ], + "logprob_reward": -0.05497020483016968, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497020483016968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49841052293777466, + 0.06918330490589142, + 0.05936020240187645, + 0.05417544022202492, + 0.039623651653528214, + 0.04959040507674217, + 0.04105527326464653, + 0.05444372072815895, + 0.05029360204935074, + 0.08386390656232834 + ], + "credit_map": [ + 0.1129457950592041, + 0.025654738768935204, + 0.13320410251617432, + 0.029526621103286743, + 0.04153238236904144, + 0.04518836736679077, + 0.053615495562553406, + 0.056743040680885315, + 0.01109554897993803, + 0.004400905221700668, + 0.0089374715462327, + 0.00745456712320447, + 0.007329581305384636, + 0.012010447680950165, + 0.008699294179677963, + 0.009255490265786648, + 0.009259503334760666, + 0.0040313503704965115, + 0.005905825644731522, + 0.006014008074998856, + 0.013870283029973507, + 0.006781505886465311, + 0.007191002834588289, + 0.006306719500571489, + 0.00876116193830967, + 0.0032354677096009254, + 0.0038793792482465506, + 0.003863988909870386, + 0.011979825794696808, + 0.012710515409708023, + 0.0043334923684597015, + 0.005411605816334486, + 0.005209088791161776, + 0.0012704347027465701, + 0.001599765382707119, + 0.0024616627488285303, + 0.004049977753311396, + 0.0043121324852108955, + 0.006473924499005079, + 0.014246665872633457, + 0.00864956434816122, + 0.002061530714854598, + 0.00251646526157856, + 0.003748285584151745, + 0.004430137574672699, + 0.016088461503386497, + 0.006019637919962406, + 0.006076321937143803, + 0.007382192183285952, + 0.0026508192531764507, + 0.004449834581464529, + 0.00280119595117867, + 0.0038630885537713766, + 0.005816721357405186, + 0.005185017362236977, + 0.008906404487788677, + 0.008180766366422176, + 0.015801556408405304, + 0.014824824407696724, + 0.004477337468415499, + 0.0024813145864754915, + 0.0018441040301695466, + 0.004237341228872538, + 0.0025964765809476376, + 0.007971115410327911, + 0.018745435401797295, + 0.011777625419199467, + 0.0030198595486581326, + 0.0020669230725616217, + 0.0013922560028731823, + 0.002528388751670718, + 0.0027919993735849857, + 0.012214954942464828, + 0.02407073974609375, + 0.019609909504652023, + 0.007049800828099251, + 0.00452665239572525, + 0.0051855770871043205, + 0.006392703857272863, + 0.004813570529222488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.05497020483016968 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot, captured under the intense glare of the stadium's floodlights, captures a moment of pure athletic power. A classic black-and-white paneled football, a violently spinning blur, is launched with explosive force into the dead center of the goal's upper crossbar. The impact is instantaneous and brutal; the ball deforms for a split second against the white-painted metal before violently rebounding onto the field. A deep, resonant tremor perceptibly shakes the entire goal frame, a powerful vibration that instantly transfers to the goal's white hexagonal grid, erupting in chaotic, wave-like vibrations. The entire structure, from crossbar to jambs and corrugated netting, shakes violently for several seconds, creating a dramatic visual echo of the powerful blow. The scene is bathed in the cold, bright lights of the stadium, casting sharp highlights on the metal and deep, dancing shadows within the trembling netting, creating an atmosphere of tense and dramatic near-misses.", + "seed": 14535, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_4851_seed_14535_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.34385817488655446, + "vlm_logprob": -0.34385817488655446, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.34385817488655446, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2240489423274994 + ], + "opposite_logprob_tokens": [ + -1.9740489721298218 + ], + "logprob_reward": -0.2240489423274994, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.2240489423274994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12969344854354858, + 0.06877314299345016, + 0.08766435831785202, + 0.15330742299556732, + 0.1321411430835724, + 0.13632601499557495, + 0.07570073753595352, + 0.11452790349721909, + 0.0664827823638916, + 0.03538304194808006 + ], + "credit_map": [ + 0.017346246168017387, + 0.01026838831603527, + 0.016141420230269432, + 0.019641010090708733, + 0.009898626245558262, + 0.013278967700898647, + 0.017397785559296608, + 0.025721000507473946, + 0.0069758216850459576, + 0.004612482152879238, + 0.014437529258430004, + 0.00944457110017538, + 0.005540024023503065, + 0.005618102382868528, + 0.012013418599963188, + 0.010131197050213814, + 0.00647687679156661, + 0.0075143431313335896, + 0.011730263009667397, + 0.022701621055603027, + 0.00538981519639492, + 0.008211075328290462, + 0.0139881307259202, + 0.011652231216430664, + 0.014349701814353466, + 0.005338022019714117, + 0.016316810622811317, + 0.024043723940849304, + 0.009381885640323162, + 0.01157792005687952, + 0.060835640877485275, + 0.011463718488812447, + 0.00926717184484005, + 0.011313388124108315, + 0.013632283546030521, + 0.016229698434472084, + 0.012487818486988544, + 0.01940740831196308, + 0.02558041363954544, + 0.024222951382398605, + 0.03385687619447708, + 0.01966303214430809, + 0.02576463110744953, + 0.019623752683401108, + 0.006602861452847719, + 0.005618304014205933, + 0.009054132737219334, + 0.016142407432198524, + 0.007674066349864006, + 0.007674706634134054, + 0.00858388189226389, + 0.011642212979495525, + 0.005499355494976044, + 0.008880790323019028, + 0.015575124882161617, + 0.010170605033636093, + 0.012010806240141392, + 0.012643828056752682, + 0.015414395369589329, + 0.020855212584137917, + 0.008121032267808914, + 0.011758856475353241, + 0.014338820241391659, + 0.019384944811463356, + 0.00897027738392353, + 0.00514694070443511, + 0.01199440285563469, + 0.006917293649166822, + 0.0057749561965465546, + 0.007378413341939449, + 0.009821205399930477, + 0.010479294694960117, + 0.0034730383194983006, + 0.004967613145709038, + 0.004254223313182592, + 0.004133166279643774, + 0.002838965505361557, + 0.004057361278682947, + 0.0051332865841686726, + 0.006525387987494469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.2240489423274994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743491411209106 + ], + "opposite_logprob_tokens": [ + -0.47434914112091064 + ], + "logprob_reward": -0.9743491411209106, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743491411209106, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12324784696102142, + 0.1337643563747406, + 0.24219606816768646, + 0.09057965874671936, + 0.09143545478582382, + 0.049877993762493134, + 0.0708712488412857, + 0.08515608310699463, + 0.02393818274140358, + 0.08893314003944397 + ], + "credit_map": [ + 0.011666715145111084, + 0.012208917178213596, + 0.018222959712147713, + 0.021929094567894936, + 0.011662086471915245, + 0.013339421711862087, + 0.01819257065653801, + 0.016026092693209648, + 0.013608409091830254, + 0.01105539035052061, + 0.029383808374404907, + 0.03398730978369713, + 0.006402990780770779, + 0.00898057222366333, + 0.018542274832725525, + 0.01180360373109579, + 0.01978042721748352, + 0.014340192079544067, + 0.06341388076543808, + 0.07086893171072006, + 0.009656920097768307, + 0.012730262242257595, + 0.024317385628819466, + 0.02708805724978447, + 0.010250221937894821, + 0.004620186984539032, + 0.01099273283034563, + 0.014659798704087734, + 0.009338002651929855, + 0.00894760899245739, + 0.022686034440994263, + 0.009085079655051231, + 0.006841674912720919, + 0.009863811545073986, + 0.009611204266548157, + 0.010381207801401615, + 0.007602166850119829, + 0.014139624312520027, + 0.016966942697763443, + 0.01602882146835327, + 0.007568050175905228, + 0.00553585821762681, + 0.0074302032589912415, + 0.0068016960285604, + 0.0041969455778598785, + 0.003974083345383406, + 0.007727384101599455, + 0.0066437735222280025, + 0.006857431028038263, + 0.006393782328814268, + 0.009358333423733711, + 0.011526710353791714, + 0.006020862143486738, + 0.007375259883701801, + 0.013166162185370922, + 0.010172710753977299, + 0.006769887637346983, + 0.012038925662636757, + 0.014392747543752193, + 0.015477562323212624, + 0.0047037359327077866, + 0.006589238531887531, + 0.009717811830341816, + 0.015466173179447651, + 0.0026758818421512842, + 0.0023177429102361202, + 0.0029276686254888773, + 0.002773328684270382, + 0.0016799132572486997, + 0.0031152856536209583, + 0.003958985675126314, + 0.004489376209676266, + 0.007789318449795246, + 0.007389185018837452, + 0.010589049197733402, + 0.013076551258563995, + 0.0067477296106517315, + 0.012193187139928341, + 0.014785847626626492, + 0.016362277790904045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.9743491411209106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38707366585731506 + ], + "opposite_logprob_tokens": [ + -1.1370736360549927 + ], + "logprob_reward": -0.38707366585731506, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38707366585731506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30366840958595276, + 0.11300931870937347, + 0.053087953478097916, + 0.14283618330955505, + 0.1627330332994461, + 0.06874749809503555, + 0.041035763919353485, + 0.061814479529857635, + 0.026651278138160706, + 0.026416076347231865 + ], + "credit_map": [ + 0.026602083817124367, + 0.026969825848937035, + 0.034162137657403946, + 0.04692118614912033, + 0.031180880963802338, + 0.03383634611964226, + 0.048643793910741806, + 0.05535215884447098, + 0.009401285089552402, + 0.007876738905906677, + 0.015098283998668194, + 0.024247869849205017, + 0.01036545354872942, + 0.012445473112165928, + 0.018613889813423157, + 0.014960326254367828, + 0.003579955780878663, + 0.005980243440717459, + 0.0058877975679934025, + 0.01003248617053032, + 0.004417214076966047, + 0.006962883751839399, + 0.008626866154372692, + 0.007600506767630577, + 0.01026735920459032, + 0.010659526102244854, + 0.015139149501919746, + 0.023516034707427025, + 0.020700203254818916, + 0.017447246238589287, + 0.026936741545796394, + 0.018169930204749107, + 0.011730161495506763, + 0.011852048337459564, + 0.015253319405019283, + 0.018615752458572388, + 0.017299754545092583, + 0.019688773900270462, + 0.04174463450908661, + 0.026548583060503006, + 0.009994101710617542, + 0.0057639023289084435, + 0.010263288393616676, + 0.011583903804421425, + 0.005045528989285231, + 0.006197045557200909, + 0.012463261373341084, + 0.0074364603497087955, + 0.0039045163430273533, + 0.006388573441654444, + 0.005171915050595999, + 0.006103537045419216, + 0.003480986226350069, + 0.004353845492005348, + 0.005749993957579136, + 0.005882398225367069, + 0.007981366477906704, + 0.01119720283895731, + 0.006756263319402933, + 0.007576454896479845, + 0.004277661442756653, + 0.006926967762410641, + 0.0075645437464118, + 0.00953401904553175, + 0.0025240620598196983, + 0.003437829902395606, + 0.004715099465101957, + 0.0048835319466888905, + 0.002150199143216014, + 0.0025503747165203094, + 0.0029065439011901617, + 0.003483637236058712, + 0.0020376830361783504, + 0.003544455859810114, + 0.003015277674421668, + 0.004251518752425909, + 0.002800657181069255, + 0.0019100146600976586, + 0.0035886282566934824, + 0.005267842207103968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.38707366585731506 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270817518234253 + ], + "opposite_logprob_tokens": [ + -2.127081871032715 + ], + "logprob_reward": -0.1270817518234253, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270817518234253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12877433001995087, + 0.1080118864774704, + 0.081624336540699, + 0.3322438895702362, + 0.14571841061115265, + 0.04007973149418831, + 0.039381444454193115, + 0.053576529026031494, + 0.020653335377573967, + 0.049936164170503616 + ], + "credit_map": [ + 0.012571284547448158, + 0.017566027119755745, + 0.020580895245075226, + 0.01873222179710865, + 0.012510162778198719, + 0.014376354403793812, + 0.014884721487760544, + 0.017552655190229416, + 0.009505699388682842, + 0.008206076920032501, + 0.014851374551653862, + 0.024131840094923973, + 0.007613442372530699, + 0.01183098554611206, + 0.018506662920117378, + 0.013365784659981728, + 0.007042160257697105, + 0.008563241921365261, + 0.015149801038205624, + 0.010443877428770065, + 0.006228504236787558, + 0.008198024705052376, + 0.010679404251277447, + 0.015319318510591984, + 0.02464132383465767, + 0.01064218021929264, + 0.027097642421722412, + 0.0831865444779396, + 0.011462618596851826, + 0.01867217756807804, + 0.13351038098335266, + 0.023031018674373627, + 0.011462211608886719, + 0.008335569873452187, + 0.01389293558895588, + 0.016048869118094444, + 0.014778560027480125, + 0.018071286380290985, + 0.04189654067158699, + 0.021232428029179573, + 0.006661815103143454, + 0.0030192830599844456, + 0.009225735440850258, + 0.005678877700120211, + 0.0020376595202833414, + 0.0038983433041721582, + 0.004559190943837166, + 0.004998823627829552, + 0.0044624838046729565, + 0.0031530840788036585, + 0.005082849878817797, + 0.005190401803702116, + 0.002949785441160202, + 0.004542406648397446, + 0.009125296957790852, + 0.004875133279711008, + 0.006289744284003973, + 0.007062643766403198, + 0.0076608541421592236, + 0.0068688965402543545, + 0.004959071055054665, + 0.0061730677261948586, + 0.007269097026437521, + 0.007293146103620529, + 0.002165355021134019, + 0.001572342123836279, + 0.003361763898283243, + 0.0032984160352498293, + 0.0013733828673139215, + 0.0018315794877707958, + 0.0037305376026779413, + 0.0033199558965861797, + 0.004309566225856543, + 0.003770619397982955, + 0.006128870882093906, + 0.006307838950306177, + 0.003982815891504288, + 0.006924822926521301, + 0.007838377729058266, + 0.010673248209059238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.1270817518234253 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737373303622007 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737373303622007, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737373303622007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16002021729946136, + 0.11183954775333405, + 0.17020046710968018, + 0.07297386229038239, + 0.09851381927728653, + 0.06000548228621483, + 0.028289705514907837, + 0.17022880911827087, + 0.06306961923837662, + 0.06485846638679504 + ], + "credit_map": [ + 0.015728246420621872, + 0.014431020244956017, + 0.01936355233192444, + 0.026370756328105927, + 0.014676030725240707, + 0.02217142842710018, + 0.021935200318694115, + 0.02534397877752781, + 0.010722088627517223, + 0.006055771838873625, + 0.02256375551223755, + 0.0177372545003891, + 0.009521887637674809, + 0.011643759906291962, + 0.01646268554031849, + 0.01713235303759575, + 0.012156931683421135, + 0.014511596411466599, + 0.021712321788072586, + 0.03817380219697952, + 0.013534589670598507, + 0.01831834204494953, + 0.027758602052927017, + 0.02403428964316845, + 0.006205640267580748, + 0.005689115729182959, + 0.006653518881648779, + 0.0116482088342309, + 0.0054855430498719215, + 0.007802801206707954, + 0.017186900600790977, + 0.01230213139206171, + 0.00651712529361248, + 0.0076089720241725445, + 0.00954779889434576, + 0.01100161112844944, + 0.009239785373210907, + 0.0126781165599823, + 0.025290638208389282, + 0.016629783436655998, + 0.013239960186183453, + 0.007848041132092476, + 0.00986776128411293, + 0.008862471207976341, + 0.002823285525664687, + 0.0033542762976139784, + 0.005172825418412685, + 0.008836858905851841, + 0.002097673714160919, + 0.001888946397230029, + 0.003812868148088455, + 0.0037030840758234262, + 0.002180765150114894, + 0.004631673917174339, + 0.005153900012373924, + 0.004820794332772493, + 0.018544849008321762, + 0.01742446981370449, + 0.031421586871147156, + 0.03135772794485092, + 0.008943433873355389, + 0.01504824310541153, + 0.02045525051653385, + 0.027033237740397453, + 0.006045500282198191, + 0.004061941523104906, + 0.008870555087924004, + 0.010348637588322163, + 0.004112069495022297, + 0.00741135748103261, + 0.010477888397872448, + 0.011741669848561287, + 0.006071261130273342, + 0.007142895832657814, + 0.007217772770673037, + 0.008645600639283657, + 0.0058986046351492405, + 0.00954508688300848, + 0.009125281125307083, + 0.01121196337044239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.006737373303622007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot, captured under the intense glare of the stadium's floodlights, captures a moment of pure athletic power. A classic black-and-white paneled football, a violently spinning blur, is launched with explosive force into the dead center of the goal's upper crossbar. The impact is instantaneous and brutal; the ball deforms for a split second against the white-painted metal before violently rebounding onto the field. A deep, resonant tremor perceptibly shakes the entire goal frame, a powerful vibration that instantly transfers to the goal's white hexagonal grid, erupting in chaotic, wave-like vibrations. The entire structure, from crossbar to jambs and corrugated netting, shakes violently for several seconds, creating a dramatic visual echo of the powerful blow. The scene is bathed in the cold, bright lights of the stadium, casting sharp highlights on the metal and deep, dancing shadows within the trembling netting, creating an atmosphere of tense and dramatic near-misses.", + "seed": 14536, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_13/index_4851_seed_14536_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4867920018732548, + "vlm_logprob": -0.4867920018732548, + "vlm_yes_no_margin": 1.55, + "vlm_token_credit": -0.4867920018732548, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09896151721477509 + ], + "opposite_logprob_tokens": [ + -3.348961591720581 + ], + "logprob_reward": -0.09896151721477509, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.09896151721477509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05646788701415062, + 0.04021597281098366, + 0.10772532224655151, + 0.048394575715065, + 0.027538185939192772, + 0.5647225379943848, + 0.03161757439374924, + 0.06169671565294266, + 0.045913904905319214, + 0.015707410871982574 + ], + "credit_map": [ + 0.008552120998501778, + 0.015158426947891712, + 0.00488834036514163, + 0.0038252382073551416, + 0.007354607805609703, + 0.004634889308363199, + 0.006663504987955093, + 0.005390757229179144, + 0.006035443861037493, + 0.007949696853756905, + 0.005752387456595898, + 0.0038049076683819294, + 0.005510909948498011, + 0.003054065862670541, + 0.0031471168622374535, + 0.004961444064974785, + 0.01413369458168745, + 0.019437434151768684, + 0.013848955743014812, + 0.01128830760717392, + 0.011224073357880116, + 0.011170477606356144, + 0.015439067967236042, + 0.01118331030011177, + 0.005046351812779903, + 0.0031993314623832703, + 0.00800585187971592, + 0.005212879274040461, + 0.0034068741369992495, + 0.004480117000639439, + 0.012156397104263306, + 0.006886770948767662, + 0.0026855000760406256, + 0.0020593355875462294, + 0.0040750205516815186, + 0.0027529906947165728, + 0.0026926789432764053, + 0.003747232025489211, + 0.0070656901225447655, + 0.002459738403558731, + 0.025849036872386932, + 0.02443832717835903, + 0.014646132476627827, + 0.22758394479751587, + 0.011124709621071815, + 0.04376440867781639, + 0.04205172508955002, + 0.17526419460773468, + 0.0032287139911204576, + 0.006764209363609552, + 0.0029562285635620356, + 0.0059928614646196365, + 0.0030953253153711557, + 0.003930093254894018, + 0.002909468486905098, + 0.0027406751178205013, + 0.003278244286775589, + 0.0038111452013254166, + 0.0034616340417414904, + 0.007912189699709415, + 0.005839701741933823, + 0.005484685767441988, + 0.015816330909729004, + 0.016092782840132713, + 0.0038206223398447037, + 0.005142203066498041, + 0.0040415432304143906, + 0.007206777110695839, + 0.005027372390031815, + 0.007059987168759108, + 0.007258730940520763, + 0.006356669124215841, + 0.0014849076978862286, + 0.001673022285103798, + 0.002701081335544586, + 0.0022974663879722357, + 0.001322139403782785, + 0.00165579654276371, + 0.002613047370687127, + 0.0019599490333348513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.09896151721477509 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103659391403198 + ], + "opposite_logprob_tokens": [ + -0.1603659689426422 + ], + "logprob_reward": -1.9103659391403198, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103659391403198, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13221855461597443, + 0.11611127108335495, + 0.13881488144397736, + 0.09396781027317047, + 0.1063409298658371, + 0.13935451209545135, + 0.06550053507089615, + 0.07378899306058884, + 0.061055805534124374, + 0.07284671813249588 + ], + "credit_map": [ + 0.01890375092625618, + 0.041426073759794235, + 0.013879889622330666, + 0.011097222566604614, + 0.011972570791840553, + 0.012600544840097427, + 0.012759304605424404, + 0.009579209610819817, + 0.01744985394179821, + 0.021631797775626183, + 0.013098274357616901, + 0.01224824134260416, + 0.014747134409844875, + 0.011547102592885494, + 0.010249147191643715, + 0.01513972133398056, + 0.01771136373281479, + 0.0226332638412714, + 0.01850159838795662, + 0.01717660017311573, + 0.01391309779137373, + 0.014261668547987938, + 0.018707161769270897, + 0.015910129994153976, + 0.010806298814713955, + 0.007981492206454277, + 0.01430540345609188, + 0.013263475149869919, + 0.005485169589519501, + 0.008293019607663155, + 0.01882239244878292, + 0.015010550618171692, + 0.011941914446651936, + 0.007140654604882002, + 0.020958518609404564, + 0.012685642577707767, + 0.009810106828808784, + 0.008276227861642838, + 0.026138868182897568, + 0.009388994425535202, + 0.01130312867462635, + 0.005268908105790615, + 0.01283168327063322, + 0.030084650963544846, + 0.00810691062361002, + 0.01582893170416355, + 0.015356728807091713, + 0.040573567152023315, + 0.005157447420060635, + 0.014435832388699055, + 0.005639777984470129, + 0.011554970405995846, + 0.008124747313559055, + 0.00847442913800478, + 0.006356325000524521, + 0.005757006350904703, + 0.0038922103121876717, + 0.003858655458316207, + 0.004968913272023201, + 0.009371461346745491, + 0.006453604903072119, + 0.006117102224379778, + 0.018609534949064255, + 0.020517509430646896, + 0.00518861273303628, + 0.00972040742635727, + 0.006922760512679815, + 0.008268984965980053, + 0.009599798358976841, + 0.007909215055406094, + 0.007244578562676907, + 0.006201446056365967, + 0.004804892465472221, + 0.005074269138276577, + 0.0060614608228206635, + 0.01723189651966095, + 0.004556078929454088, + 0.009133645333349705, + 0.018241748213768005, + 0.007742727641016245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -1.9103659391403198 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161622762680054 + ], + "opposite_logprob_tokens": [ + -1.7016162872314453 + ], + "logprob_reward": -0.20161622762680054, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161622762680054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3125247359275818, + 0.0922507718205452, + 0.07433319091796875, + 0.14548484981060028, + 0.09264271706342697, + 0.12819841504096985, + 0.03565596789121628, + 0.03108106553554535, + 0.025590447708964348, + 0.062237828969955444 + ], + "credit_map": [ + 0.042626503854990005, + 0.056698545813560486, + 0.030343646183609962, + 0.027517694979906082, + 0.058140791952610016, + 0.034469179809093475, + 0.03526735305786133, + 0.027461020275950432, + 0.01439766027033329, + 0.01731596514582634, + 0.009786670096218586, + 0.008061930537223816, + 0.009053059853613377, + 0.00936182215809822, + 0.008585295639932156, + 0.015688374638557434, + 0.008635211735963821, + 0.008444429375231266, + 0.009937750175595284, + 0.008269645273685455, + 0.006907077506184578, + 0.009326145984232426, + 0.011882299557328224, + 0.010930627584457397, + 0.015552370809018612, + 0.009465073235332966, + 0.023308463394641876, + 0.014584900811314583, + 0.009599483571946621, + 0.012477382086217403, + 0.03853242099285126, + 0.021964751183986664, + 0.008642804808914661, + 0.007160027977079153, + 0.01982184685766697, + 0.011516747064888477, + 0.010194962844252586, + 0.008695991709828377, + 0.01805410534143448, + 0.008556226268410683, + 0.007718516979366541, + 0.005509973503649235, + 0.004975917749106884, + 0.049073074012994766, + 0.004147815052419901, + 0.013502220623195171, + 0.009828805923461914, + 0.03344208002090454, + 0.0034014980774372816, + 0.007350033149123192, + 0.003960199188441038, + 0.005037656985223293, + 0.005449773743748665, + 0.0036808038130402565, + 0.003052858868613839, + 0.00372314197011292, + 0.003065637079998851, + 0.0027668869588524103, + 0.00397510826587677, + 0.005363953299820423, + 0.0025926653761416674, + 0.003655675332993269, + 0.004610616713762283, + 0.005050523206591606, + 0.0017317702295258641, + 0.003136876504868269, + 0.0029245479963719845, + 0.00402628630399704, + 0.002510953927412629, + 0.004439254757016897, + 0.0034274570643901825, + 0.0033933010417968035, + 0.0034295006189495325, + 0.004260451067239046, + 0.005511130206286907, + 0.014357509091496468, + 0.004314188379794359, + 0.007607352454215288, + 0.016229670494794846, + 0.006528027355670929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.20161622762680054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994561791419983 + ], + "opposite_logprob_tokens": [ + -1.8049455881118774 + ], + "logprob_reward": -0.17994561791419983, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994561791419983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08641523867845535, + 0.025834999978542328, + 0.06648870557546616, + 0.030039669945836067, + 0.07315634936094284, + 0.6088659167289734, + 0.01385155413299799, + 0.02636622078716755, + 0.010083191096782684, + 0.05889816954731941 + ], + "credit_map": [ + 0.012743680737912655, + 0.01924527809023857, + 0.005939611699432135, + 0.007405953481793404, + 0.01403055526316166, + 0.010707532055675983, + 0.008261811919510365, + 0.008080814965069294, + 0.0038588098250329494, + 0.005389161873608828, + 0.003838698612526059, + 0.0022988945711404085, + 0.003014937276020646, + 0.002536818152293563, + 0.002133506815880537, + 0.00276417238637805, + 0.0086511867120862, + 0.010428679175674915, + 0.008128662593662739, + 0.006489430088549852, + 0.008461372926831245, + 0.008177891373634338, + 0.007402191869914532, + 0.008749295026063919, + 0.004118677228689194, + 0.0032100751996040344, + 0.004269021097570658, + 0.003931969869881868, + 0.002125693717971444, + 0.0037631443701684475, + 0.00448576221242547, + 0.0041353278793394566, + 0.01053926907479763, + 0.006489356979727745, + 0.015183216892182827, + 0.006773706525564194, + 0.007582684047520161, + 0.006741537246853113, + 0.013153815641999245, + 0.006692767608910799, + 0.026413070037961006, + 0.0266860444098711, + 0.015962587669491768, + 0.2444632649421692, + 0.01180232409387827, + 0.043168503791093826, + 0.045808032155036926, + 0.19456209242343903, + 0.0014291747938841581, + 0.001624838332645595, + 0.0014271903783082962, + 0.0019510774873197079, + 0.002082887338474393, + 0.0017476851353421807, + 0.0019834917038679123, + 0.0016052096616476774, + 0.0011937577510252595, + 0.0013940325006842613, + 0.0014019871596246958, + 0.003132045967504382, + 0.0023311444092541933, + 0.00246898434124887, + 0.006754711270332336, + 0.007689556572586298, + 0.0010119295911863446, + 0.001961561618372798, + 0.001446855254471302, + 0.0014554030494764447, + 0.0013714595697820187, + 0.001405650400556624, + 0.0007013333961367607, + 0.0007289969944395125, + 0.004209392238408327, + 0.00321915652602911, + 0.0063704135827720165, + 0.008492236956954002, + 0.004500354640185833, + 0.010350950993597507, + 0.012564733624458313, + 0.009190931916236877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.17994561791419983 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black-and-white paneled football visible spinning violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball impact the upper crossbar and cause the goal frame to shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by intense stadium floodlights creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043070707470178604 + ], + "opposite_logprob_tokens": [ + -3.1680707931518555 + ], + "logprob_reward": -0.043070707470178604, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043070707470178604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06833634525537491, + 0.07212460041046143, + 0.2715819180011749, + 0.024825846776366234, + 0.045002639293670654, + 0.32920077443122864, + 0.09408867359161377, + 0.020635463297367096, + 0.028249189257621765, + 0.04595459625124931 + ], + "credit_map": [ + 0.009318673051893711, + 0.01655230112373829, + 0.006659518461674452, + 0.0059830318205058575, + 0.011109490878880024, + 0.0064422981813549995, + 0.006454900838434696, + 0.005816128570586443, + 0.009861237369477749, + 0.008791572414338589, + 0.00538996234536171, + 0.008111118339002132, + 0.0057230559177696705, + 0.007432973477989435, + 0.009513064287602901, + 0.017301609739661217, + 0.03234007582068443, + 0.04338810592889786, + 0.031612809747457504, + 0.032869793474674225, + 0.03678767383098602, + 0.026202023029327393, + 0.040809694677591324, + 0.02757168561220169, + 0.002229421865195036, + 0.0017433611210435629, + 0.002672470174729824, + 0.003392269602045417, + 0.0018228517146781087, + 0.002722406992688775, + 0.006124407518655062, + 0.004118653945624828, + 0.00562305748462677, + 0.003560112789273262, + 0.006260037887841463, + 0.004155529662966728, + 0.005389492027461529, + 0.00437172083184123, + 0.011798393912613392, + 0.003844292601570487, + 0.01432365458458662, + 0.013652500696480274, + 0.008796647191047668, + 0.13054223358631134, + 0.006923752371221781, + 0.02316909097135067, + 0.025908121839165688, + 0.10588475316762924, + 0.009083242155611515, + 0.019963981583714485, + 0.00862094759941101, + 0.014026648364961147, + 0.012891728430986404, + 0.012205244041979313, + 0.007270670961588621, + 0.010026196017861366, + 0.0012734236661344767, + 0.0012589883990585804, + 0.0016416251892223954, + 0.002318441402167082, + 0.0019471491686999798, + 0.0015588118694722652, + 0.00524212047457695, + 0.005394898820668459, + 0.0026855976320803165, + 0.005033586639910936, + 0.0037065313663333654, + 0.003484760643914342, + 0.00520464638248086, + 0.0030976177658885717, + 0.002523453440517187, + 0.002512994222342968, + 0.003804131643846631, + 0.0038339360617101192, + 0.006547834724187851, + 0.0053953551687300205, + 0.0033962882589548826, + 0.006622097454965115, + 0.008496406488120556, + 0.00785854086279869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.043070707470178604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_14/index_tensor([7654, 7654])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_14/index_tensor([7654, 7654])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0349b5cd327b8a50faaff5c7c7dbce7beb57af9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_14/index_tensor([7654, 7654])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth tracking shot, viewed from a slightly upward angle, glides steadily and leisurely from left to right amidst the raging typhoon. Bean-sized raindrops, almost forming a curtain of water, cascade down, lashing every surface. A howling wind sways the surrounding vegetation. The camera slowly sweeps across a futuristic cluster of intelligent greenhouses, stacked on the mountainside. Composed of translucent bio-domes and delicate metal frames, the structure embodies the signature, delicate, hand-painted quality of Hayao Miyazaki's animated films. Amidst a cold world shrouded by thick, swirling blue-gray clouds, a warm, soft orange-yellow glow emanates from within each dome. This light penetrates the damp, rain-scoured surface, casting warm patches of light on the soggy ground outside and subtly illuminating the lush, exotic flora within. The violence of the outside world forms a strong dramatic contrast with the tranquility and warmth of the inner space. This light source that persists in the storm is like a quiet and powerful oasis of life, creating a grand atmosphere full of hope and vitality in the harsh environment of doomsday.", + "seed": 14493, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_14/index_7654_seed_14493_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05249781443271786, + "vlm_logprob": -0.05249781443271786, + "vlm_yes_no_margin": 4.1999999523162845, + "vlm_token_credit": -0.05249781443271786, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02800147607922554 + ], + "opposite_logprob_tokens": [ + -4.028001308441162 + ], + "logprob_reward": -0.02800147607922554, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02800147607922554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3977857828140259, + 0.0472363717854023, + 0.047948095947504044, + 0.03328496217727661, + 0.08325468003749847, + 0.1825810968875885, + 0.1039314940571785, + 0.038916848599910736, + 0.02916759066283703, + 0.0358930379152298 + ], + "credit_map": [ + 0.02019122987985611, + 0.1002589613199234, + 0.13003341853618622, + 0.034958984702825546, + 0.04823499172925949, + 0.025682535022497177, + 0.0190962553024292, + 0.019329385831952095, + 0.004018513020128012, + 0.005209784954786301, + 0.006340982858091593, + 0.008605730719864368, + 0.005975718609988689, + 0.005292199086397886, + 0.005152835510671139, + 0.006640610750764608, + 0.003839466953650117, + 0.007192845456302166, + 0.00669369800016284, + 0.0066127716563642025, + 0.004217390436679125, + 0.004299683030694723, + 0.007070399820804596, + 0.008021838031709194, + 0.002366243861615658, + 0.0028278050012886524, + 0.005924206227064133, + 0.004313196986913681, + 0.0031644818373024464, + 0.0038325118366628885, + 0.006446730811148882, + 0.0044097877107560635, + 0.00707613117992878, + 0.010885816067457199, + 0.013057485222816467, + 0.010793669149279594, + 0.01621105894446373, + 0.007486507762223482, + 0.010388020426034927, + 0.007355990819633007, + 0.02474147081375122, + 0.04529121145606041, + 0.024941325187683105, + 0.016365481540560722, + 0.017896393314003944, + 0.01550076063722372, + 0.02456030249595642, + 0.013284164480865002, + 0.019263772293925285, + 0.026418665423989296, + 0.01318639237433672, + 0.012957192026078701, + 0.007953504100441933, + 0.006521403789520264, + 0.008514495566487312, + 0.00911607127636671, + 0.0063491384498775005, + 0.009699778631329536, + 0.005006582476198673, + 0.0055193305015563965, + 0.0036111101508140564, + 0.0019080545753240585, + 0.003424316644668579, + 0.003398534841835499, + 0.004414895549416542, + 0.008129754103720188, + 0.004919445142149925, + 0.0031215304043143988, + 0.0020037174690514803, + 0.0017858570208773017, + 0.002646504668518901, + 0.002145885955542326, + 0.003664208808913827, + 0.007693200837820768, + 0.004202508367598057, + 0.002512596780434251, + 0.0014472438488155603, + 0.003056052140891552, + 0.00357851036824286, + 0.009738720022141933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.02800147607922554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154879987239838 + ], + "opposite_logprob_tokens": [ + -1.7015488147735596 + ], + "logprob_reward": -0.20154879987239838, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154879987239838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5919399857521057, + 0.0482034869492054, + 0.06989534944295883, + 0.03640623390674591, + 0.02370571345090866, + 0.08336570113897324, + 0.06482674926519394, + 0.05326851084828377, + 0.018161635845899582, + 0.01022653840482235 + ], + "credit_map": [ + 0.028457945212721825, + 0.16745883226394653, + 0.22599990665912628, + 0.05947479233145714, + 0.032848212867975235, + 0.021557074040174484, + 0.02688726969063282, + 0.029256029054522514, + 0.005393918603658676, + 0.005944378674030304, + 0.007151640020310879, + 0.00839647464454174, + 0.005622304044663906, + 0.005056539550423622, + 0.005594581365585327, + 0.005043657962232828, + 0.006077090743929148, + 0.005806867498904467, + 0.013808918185532093, + 0.011027960106730461, + 0.00572062935680151, + 0.008973839692771435, + 0.008146327920258045, + 0.010333730839192867, + 0.002113902475684881, + 0.006371856201440096, + 0.007400156930088997, + 0.0038880323991179466, + 0.003738492727279663, + 0.004099782556295395, + 0.004908682778477669, + 0.003885332029312849, + 0.0019649581518024206, + 0.0036150161176919937, + 0.0024526570923626423, + 0.002700646175071597, + 0.005029455292969942, + 0.002989516593515873, + 0.0026803368236869574, + 0.002273129066452384, + 0.014869852922856808, + 0.020962117239832878, + 0.013770128600299358, + 0.008419071324169636, + 0.00983007624745369, + 0.004566265270113945, + 0.0076570529490709305, + 0.0032911451999098063, + 0.018305212259292603, + 0.024806853383779526, + 0.004015602637082338, + 0.004594570957124233, + 0.0054619829170405865, + 0.0022494783625006676, + 0.0027198907919228077, + 0.002673160284757614, + 0.008538597263395786, + 0.012706120498478413, + 0.0074115293100476265, + 0.008769599720835686, + 0.0039372071623802185, + 0.002468355465680361, + 0.0054814135655760765, + 0.003955697640776634, + 0.0024568692315369844, + 0.004066884983330965, + 0.003292592242360115, + 0.0031672045588493347, + 0.001613558386452496, + 0.0008060230175033212, + 0.001584562356583774, + 0.0011739457258954644, + 0.001052727340720594, + 0.0014061256079003215, + 0.0018435600213706493, + 0.0011931787012144923, + 0.0007710011559538543, + 0.0012115373974666, + 0.0009707009303383529, + 0.0017777093453332782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.20154879987239838 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257826764136553 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257826764136553, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257826764136553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2545046806335449, + 0.03475698456168175, + 0.0768105611205101, + 0.036305021494627, + 0.042021434754133224, + 0.14819210767745972, + 0.04845547303557396, + 0.13963831961154938, + 0.06300412863492966, + 0.1563113033771515 + ], + "credit_map": [ + 0.013020451180636883, + 0.06491473317146301, + 0.09292855113744736, + 0.025634683668613434, + 0.016262724995613098, + 0.010703707113862038, + 0.01444737147539854, + 0.016592446714639664, + 0.0027967633213847876, + 0.004201383329927921, + 0.0054822214879095554, + 0.005770972464233637, + 0.004537098575383425, + 0.004144591279327869, + 0.0033572011161595583, + 0.004466752987354994, + 0.006412905175238848, + 0.006130516994744539, + 0.013626638799905777, + 0.00941616389900446, + 0.004831889644265175, + 0.008858948945999146, + 0.01279551163315773, + 0.014737982302904129, + 0.0022980438079684973, + 0.0062581561505794525, + 0.007393239066004753, + 0.003707454539835453, + 0.0031826114282011986, + 0.0032264564651995897, + 0.006183460354804993, + 0.004055595491081476, + 0.003434120211750269, + 0.006189993117004633, + 0.0043492065742611885, + 0.00562739372253418, + 0.008433178998529911, + 0.00383526342920959, + 0.006166595034301281, + 0.00398567970842123, + 0.03415028378367424, + 0.03656771779060364, + 0.023934559896588326, + 0.013465974479913712, + 0.0129775982350111, + 0.005246701184660196, + 0.014920121058821678, + 0.006929140072315931, + 0.005622943863272667, + 0.005734983831644058, + 0.006172434892505407, + 0.006924609187990427, + 0.005449838470667601, + 0.005748421885073185, + 0.005052204709500074, + 0.007750031538307667, + 0.010269935242831707, + 0.014598897658288479, + 0.019402842968702316, + 0.0310067031532526, + 0.012508295476436615, + 0.01053677313029766, + 0.028719373047351837, + 0.012595485895872116, + 0.00882207602262497, + 0.01901094801723957, + 0.0125189870595932, + 0.006100747734308243, + 0.0043592569418251514, + 0.003339748363941908, + 0.005066418554633856, + 0.0037859391886740923, + 0.02391052432358265, + 0.04886481165885925, + 0.025850767269730568, + 0.008381943218410015, + 0.006254800129681826, + 0.008495216257870197, + 0.009426169097423553, + 0.025127064436674118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.005257826764136553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013365155318751931 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013365155318751931, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013365155318751931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2424398511648178, + 0.052423179149627686, + 0.030501479282975197, + 0.02784225344657898, + 0.05352918058633804, + 0.14947429299354553, + 0.12427122890949249, + 0.04418513551354408, + 0.09512042999267578, + 0.18021300435066223 + ], + "credit_map": [ + 0.014219648204743862, + 0.06639005243778229, + 0.08844999969005585, + 0.0242370143532753, + 0.01588674820959568, + 0.011118556372821331, + 0.010017610155045986, + 0.012120193801820278, + 0.003945887088775635, + 0.005436559207737446, + 0.010190749540925026, + 0.008867393247783184, + 0.007705128286033869, + 0.006278632674366236, + 0.005522246938198805, + 0.004476578440517187, + 0.003291588043794036, + 0.003292957553640008, + 0.004603626672178507, + 0.0049520148895680904, + 0.002949624788016081, + 0.00308866030536592, + 0.00422901613637805, + 0.004093992058187723, + 0.002518728142604232, + 0.0021515670232474804, + 0.003297447692602873, + 0.0035353670828044415, + 0.0030710515566170216, + 0.004415556322783232, + 0.004312371835112572, + 0.004540163092315197, + 0.003479675156995654, + 0.006187312304973602, + 0.006031478755176067, + 0.005654643289744854, + 0.008462345227599144, + 0.00479145348072052, + 0.009440720081329346, + 0.009481553919613361, + 0.01593845523893833, + 0.011777891777455807, + 0.018322540447115898, + 0.020039480179548264, + 0.02038114331662655, + 0.02139795571565628, + 0.02385338395833969, + 0.017763420939445496, + 0.02314011938869953, + 0.031838223338127136, + 0.011648228392004967, + 0.013252923265099525, + 0.009513824246823788, + 0.0074528176337480545, + 0.007378073409199715, + 0.0200470183044672, + 0.0035769569221884012, + 0.006048193667083979, + 0.006266649346798658, + 0.008302452974021435, + 0.005397733300924301, + 0.00413763290271163, + 0.0061928448267281055, + 0.004262671340256929, + 0.011243550106883049, + 0.02230226993560791, + 0.017050618305802345, + 0.013317936100065708, + 0.0069786496460437775, + 0.0068031311966478825, + 0.011120692826807499, + 0.006303580943495035, + 0.037805330008268356, + 0.06447264552116394, + 0.023010730743408203, + 0.00993432942777872, + 0.009208516217768192, + 0.008980688638985157, + 0.008346892893314362, + 0.018453877419233322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0013365155318751931 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344453915953636 + ], + "opposite_logprob_tokens": [ + -3.6513445377349854 + ], + "logprob_reward": -0.026344453915953636, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344453915953636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.253955602645874, + 0.06291940063238144, + 0.07322333008050919, + 0.04803789034485817, + 0.045977357774972916, + 0.053146328777074814, + 0.22405855357646942, + 0.06534094363451004, + 0.07247093319892883, + 0.10086963325738907 + ], + "credit_map": [ + 0.016105713322758675, + 0.06111711636185646, + 0.08476870507001877, + 0.02731795236468315, + 0.01786947064101696, + 0.014909415505826473, + 0.013730093836784363, + 0.018137119710445404, + 0.0050358036532998085, + 0.0071230074390769005, + 0.007099154405295849, + 0.007399143651127815, + 0.006734881550073624, + 0.009098697453737259, + 0.00962493009865284, + 0.010803782381117344, + 0.006387078203260899, + 0.006535056047141552, + 0.011969967745244503, + 0.008857161737978458, + 0.006030047312378883, + 0.00727575970813632, + 0.010536568239331245, + 0.015631698071956635, + 0.002836022060364485, + 0.006166434846818447, + 0.012939246371388435, + 0.004842641297727823, + 0.0031022964976727962, + 0.004875339567661285, + 0.0062146540731191635, + 0.007061254233121872, + 0.004947859328240156, + 0.007525519002228975, + 0.004590306896716356, + 0.005195418372750282, + 0.006728057283908129, + 0.006051348056644201, + 0.006386730819940567, + 0.004552118014544249, + 0.01000939030200243, + 0.011301076039671898, + 0.005481618456542492, + 0.004709483589977026, + 0.005854387301951647, + 0.004920984152704477, + 0.004679423291236162, + 0.006189961917698383, + 0.06334475427865982, + 0.0854904055595398, + 0.013534464873373508, + 0.016619788482785225, + 0.02062908373773098, + 0.0066704656928777695, + 0.008715612813830376, + 0.00905398279428482, + 0.008679891936480999, + 0.012365410104393959, + 0.009144517593085766, + 0.00912192277610302, + 0.007008392363786697, + 0.004348448943346739, + 0.008360743522644043, + 0.006311613600701094, + 0.012165261432528496, + 0.0213280338793993, + 0.012430381029844284, + 0.007453881669789553, + 0.005853991489857435, + 0.004373232834041119, + 0.005046321544796228, + 0.003819831646978855, + 0.011347283609211445, + 0.02410499006509781, + 0.008454055525362492, + 0.005784852430224419, + 0.00437811017036438, + 0.00863795168697834, + 0.009994637221097946, + 0.028167739510536194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.026344453915953636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot, viewed from a slightly upward angle, glides steadily and leisurely from left to right amidst the raging typhoon. Bean-sized raindrops, almost forming a curtain of water, cascade down, lashing every surface. A howling wind sways the surrounding vegetation. The camera slowly sweeps across a futuristic cluster of intelligent greenhouses, stacked on the mountainside. Composed of translucent bio-domes and delicate metal frames, the structure embodies the signature, delicate, hand-painted quality of Hayao Miyazaki's animated films. Amidst a cold world shrouded by thick, swirling blue-gray clouds, a warm, soft orange-yellow glow emanates from within each dome. This light penetrates the damp, rain-scoured surface, casting warm patches of light on the soggy ground outside and subtly illuminating the lush, exotic flora within. The violence of the outside world forms a strong dramatic contrast with the tranquility and warmth of the inner space. This light source that persists in the storm is like a quiet and powerful oasis of life, creating a grand atmosphere full of hope and vitality in the harsh environment of doomsday.", + "seed": 14494, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_14/index_7654_seed_14494_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07911382340826094, + "vlm_logprob": -0.07911382340826094, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.07911382340826094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18820294737815857 + ], + "opposite_logprob_tokens": [ + -1.813202977180481 + ], + "logprob_reward": -0.18820294737815857, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18820294737815857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22783631086349487, + 0.032532598823308945, + 0.02616007626056671, + 0.04555394873023033, + 0.05969613417983055, + 0.05433370918035507, + 0.09112030267715454, + 0.2969856858253479, + 0.07149343192577362, + 0.09428778290748596 + ], + "credit_map": [ + 0.020609162747859955, + 0.020876774564385414, + 0.05689971521496773, + 0.017895502969622612, + 0.054846085608005524, + 0.021939167752861977, + 0.014366853050887585, + 0.0204030591994524, + 0.002550219651311636, + 0.003878643037751317, + 0.00493711419403553, + 0.004813602194190025, + 0.0034907683730125427, + 0.0025709739420562983, + 0.0057387081906199455, + 0.00455256924033165, + 0.0019634447526186705, + 0.0025308087933808565, + 0.004183905199170113, + 0.00411744462326169, + 0.0026824388187378645, + 0.003092430764809251, + 0.004570707678794861, + 0.003018896095454693, + 0.004635212477296591, + 0.0045670862309634686, + 0.006572107318788767, + 0.008890497498214245, + 0.004526121541857719, + 0.004654137883335352, + 0.007116633467376232, + 0.004592150449752808, + 0.005173181649297476, + 0.007237199693918228, + 0.005356306210160255, + 0.007593714166432619, + 0.0066710165701806545, + 0.007938346825540066, + 0.01340141985565424, + 0.006324947811663151, + 0.007809830363839865, + 0.007435200270265341, + 0.0088886059820652, + 0.006140653509646654, + 0.007559821009635925, + 0.007548221852630377, + 0.004293461795896292, + 0.004657913465052843, + 0.011489697732031345, + 0.01224592886865139, + 0.01780853234231472, + 0.012924442999064922, + 0.011975608766078949, + 0.010629209689795971, + 0.006534749176353216, + 0.007512134965509176, + 0.05994619056582451, + 0.03363650664687157, + 0.06198674067854881, + 0.02239934168756008, + 0.03243550658226013, + 0.0431273952126503, + 0.019377047196030617, + 0.024076959118247032, + 0.006625107489526272, + 0.007040776778012514, + 0.010000676847994328, + 0.005953075364232063, + 0.012826916761696339, + 0.010371438227593899, + 0.008758455514907837, + 0.009916978888213634, + 0.004284425638616085, + 0.005838992074131966, + 0.0071243541315197945, + 0.00615654606372118, + 0.02624116651713848, + 0.027676228433847427, + 0.006740301847457886, + 0.010225774720311165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.18820294737815857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987382411956787 + ], + "opposite_logprob_tokens": [ + -1.8048738241195679 + ], + "logprob_reward": -0.17987382411956787, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987382411956787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2716890275478363, + 0.10382623970508575, + 0.08246088027954102, + 0.1357480138540268, + 0.049656063318252563, + 0.056159112602472305, + 0.07975539565086365, + 0.09443719685077667, + 0.031833380460739136, + 0.0944347009062767 + ], + "credit_map": [ + 0.02147737704217434, + 0.040538255125284195, + 0.03403874859213829, + 0.03858158737421036, + 0.022404499351978302, + 0.02936043031513691, + 0.04002298414707184, + 0.04526513069868088, + 0.007134331855922937, + 0.018020430579781532, + 0.011549313552677631, + 0.011643650010228157, + 0.011013983748853207, + 0.009965820237994194, + 0.019550368189811707, + 0.014948341995477676, + 0.005197467282414436, + 0.007804839871823788, + 0.012925521470606327, + 0.008185097016394138, + 0.008180789649486542, + 0.01532517746090889, + 0.015844125300645828, + 0.008997865952551365, + 0.012812511995434761, + 0.01541563868522644, + 0.022852221503853798, + 0.036383166909217834, + 0.011799110099673271, + 0.009480259381234646, + 0.014933358877897263, + 0.012071739882230759, + 0.003573331516236067, + 0.005105354357510805, + 0.008040384389460087, + 0.007122349925339222, + 0.004945600870996714, + 0.005596433766186237, + 0.010350720956921577, + 0.00492188660427928, + 0.00544431759044528, + 0.005564677529036999, + 0.006997444201260805, + 0.009749079123139381, + 0.006147872190922499, + 0.00913386233150959, + 0.007116144523024559, + 0.006005714647471905, + 0.010529649443924427, + 0.012652341276407242, + 0.01621699333190918, + 0.011213445104658604, + 0.00877931248396635, + 0.00922037847340107, + 0.005035612732172012, + 0.0061076669953763485, + 0.02016904577612877, + 0.012380232103168964, + 0.014720949344336987, + 0.005705761257559061, + 0.014255475252866745, + 0.0141708729788661, + 0.005600376985967159, + 0.00743449293076992, + 0.0038902226369827986, + 0.002840396948158741, + 0.0026053646579384804, + 0.0018294263863936067, + 0.007769904565066099, + 0.005324655678123236, + 0.0021958835422992706, + 0.005377527326345444, + 0.003255331888794899, + 0.004001465160399675, + 0.0039525339379906654, + 0.004799941089004278, + 0.034252192825078964, + 0.03313388302922249, + 0.0042097801342606544, + 0.006829565856605768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.17987382411956787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017169271595776081 + ], + "opposite_logprob_tokens": [ + -6.3767170906066895 + ], + "logprob_reward": -0.0017169271595776081, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017169271595776081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25517264008522034, + 0.03843376040458679, + 0.06652036309242249, + 0.09710703790187836, + 0.08795646578073502, + 0.08193743973970413, + 0.08869500458240509, + 0.11882902681827545, + 0.08615004271268845, + 0.0791981890797615 + ], + "credit_map": [ + 0.015249831601977348, + 0.023784136399626732, + 0.028579819947481155, + 0.022967958822846413, + 0.06727483868598938, + 0.04410972818732262, + 0.022384174168109894, + 0.03082217648625374, + 0.0029265948105603456, + 0.005032867193222046, + 0.006490154657512903, + 0.00445675989612937, + 0.005039024632424116, + 0.004371729679405689, + 0.005642259493470192, + 0.0044743758626282215, + 0.003680763766169548, + 0.009423659183084965, + 0.01102664414793253, + 0.0069289784878492355, + 0.00885399803519249, + 0.006789844017475843, + 0.013284772634506226, + 0.006531713996082544, + 0.006266784388571978, + 0.00988326221704483, + 0.016048798337578773, + 0.027417665347456932, + 0.007613792084157467, + 0.009199002757668495, + 0.01142205111682415, + 0.009255694225430489, + 0.01121716108173132, + 0.008936673402786255, + 0.012029812671244144, + 0.011380196548998356, + 0.010582559742033482, + 0.013442984782159328, + 0.012165266089141369, + 0.008201823569834232, + 0.011680363677442074, + 0.011872529052197933, + 0.009676635265350342, + 0.012497715651988983, + 0.009148991666734219, + 0.008972393348813057, + 0.011379677802324295, + 0.0067091407254338264, + 0.011724519543349743, + 0.011362634599208832, + 0.012980719096958637, + 0.010441380552947521, + 0.015872789546847343, + 0.009056402370333672, + 0.008197439834475517, + 0.00905913021415472, + 0.019933562725782394, + 0.012320607900619507, + 0.022794688120484352, + 0.010907476767897606, + 0.016411010175943375, + 0.015883512794971466, + 0.009116956032812595, + 0.011461226269602776, + 0.010231747291982174, + 0.010871880687773228, + 0.007800986059010029, + 0.005215905606746674, + 0.01698598451912403, + 0.012755505740642548, + 0.009580032899975777, + 0.012708007358014584, + 0.003979218192398548, + 0.004935605451464653, + 0.006380072329193354, + 0.006599174812436104, + 0.022674910724163055, + 0.025018181651830673, + 0.004164982121437788, + 0.005446053110063076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0017169271595776081 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489208010956645 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489208010956645, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489208010956645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.163002148270607, + 0.03214632719755173, + 0.03843654319643974, + 0.10647989064455032, + 0.12529802322387695, + 0.08822141587734222, + 0.1847791224718094, + 0.09328681230545044, + 0.09614334255456924, + 0.07220638543367386 + ], + "credit_map": [ + 0.01747957058250904, + 0.019591547548770905, + 0.02616793103516102, + 0.014613484963774681, + 0.03194534406065941, + 0.024356970563530922, + 0.01430415641516447, + 0.014543137513101101, + 0.0030681747011840343, + 0.003900386393070221, + 0.005749281030148268, + 0.003729593940079212, + 0.0035712660755962133, + 0.0029861193615943193, + 0.004822752438485622, + 0.004318754654377699, + 0.0024075214751064777, + 0.003109114011749625, + 0.005160344764590263, + 0.0041054971516132355, + 0.004093238618224859, + 0.006939265411347151, + 0.008048038929700851, + 0.004573527257889509, + 0.011718551628291607, + 0.012087229639291763, + 0.017071224749088287, + 0.023247666656970978, + 0.010750099085271358, + 0.010528400540351868, + 0.01042048167437315, + 0.010656235739588737, + 0.012568887323141098, + 0.012709998525679111, + 0.01897011511027813, + 0.01407863199710846, + 0.013273971155285835, + 0.021907072514295578, + 0.01839771866798401, + 0.013391628861427307, + 0.010825766250491142, + 0.010638411156833172, + 0.01422958541661501, + 0.012674622237682343, + 0.011203113943338394, + 0.011036294512450695, + 0.008599522523581982, + 0.009014103561639786, + 0.027295419946312904, + 0.024081561714410782, + 0.02703797072172165, + 0.023277223110198975, + 0.024467039853334427, + 0.020449461415410042, + 0.016020815819501877, + 0.022149641066789627, + 0.013735753484070301, + 0.011047274805605412, + 0.014737983234226704, + 0.007872170768678188, + 0.015410839579999447, + 0.012390140444040298, + 0.008370136842131615, + 0.009722528047859669, + 0.013248072005808353, + 0.011135109700262547, + 0.008864074945449829, + 0.008428866043686867, + 0.016422608867287636, + 0.013572506606578827, + 0.010367096401751041, + 0.014105013571679592, + 0.0061468705534935, + 0.005962457973510027, + 0.007062053307890892, + 0.007039680145680904, + 0.008845549076795578, + 0.01682082749903202, + 0.007491547614336014, + 0.012837406247854233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.002489208010956645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286210373044014 + ], + "opposite_logprob_tokens": [ + -3.7732861042022705 + ], + "logprob_reward": -0.023286210373044014, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286210373044014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33440959453582764, + 0.06505854427814484, + 0.03263493627309799, + 0.2754177153110504, + 0.07455682754516602, + 0.039008330553770065, + 0.03738363832235336, + 0.030796803534030914, + 0.030573680996894836, + 0.08015995472669601 + ], + "credit_map": [ + 0.01612425036728382, + 0.030618928372859955, + 0.039883874356746674, + 0.01670517586171627, + 0.12203240394592285, + 0.06745584309101105, + 0.023240163922309875, + 0.01834893971681595, + 0.004802055191248655, + 0.0060064843855798244, + 0.007375231012701988, + 0.008254282176494598, + 0.006642167456448078, + 0.005051329731941223, + 0.014194983057677746, + 0.012731998227536678, + 0.0022654507774859667, + 0.005177034065127373, + 0.004605876747518778, + 0.0024476787075400352, + 0.005582320503890514, + 0.0043884702026844025, + 0.004285342060029507, + 0.0038827552925795317, + 0.02073713392019272, + 0.030976856127381325, + 0.048163387924432755, + 0.09116655588150024, + 0.02156969904899597, + 0.013844405300915241, + 0.031932149082422256, + 0.01702750474214554, + 0.011313744820654392, + 0.007701917551457882, + 0.010544978082180023, + 0.009926403872668743, + 0.011890425346791744, + 0.0075230286456644535, + 0.008855396881699562, + 0.006800919305533171, + 0.005224797409027815, + 0.004493710584938526, + 0.0058604138903319836, + 0.008093238808214664, + 0.004715035669505596, + 0.003959936089813709, + 0.003916767425835133, + 0.0027444232255220413, + 0.005383796524256468, + 0.005137867294251919, + 0.006569846533238888, + 0.0057325647212564945, + 0.004281726665794849, + 0.0041010198183357716, + 0.0028876110445708036, + 0.003289197338744998, + 0.004875644575804472, + 0.004062635824084282, + 0.005412616766989231, + 0.002759199822321534, + 0.0030824639834463596, + 0.004059247672557831, + 0.002606729744002223, + 0.003938260953873396, + 0.003365038661286235, + 0.0025552564766258, + 0.004927380010485649, + 0.003163325833156705, + 0.003664835589006543, + 0.004171629901975393, + 0.003225756110623479, + 0.00550045445561409, + 0.005685372743755579, + 0.0063215806148946285, + 0.006960668601095676, + 0.007565614301711321, + 0.019909337162971497, + 0.021151164546608925, + 0.0066080959513783455, + 0.005958105903118849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.023286210373044014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot, viewed from a slightly upward angle, glides steadily and leisurely from left to right amidst the raging typhoon. Bean-sized raindrops, almost forming a curtain of water, cascade down, lashing every surface. A howling wind sways the surrounding vegetation. The camera slowly sweeps across a futuristic cluster of intelligent greenhouses, stacked on the mountainside. Composed of translucent bio-domes and delicate metal frames, the structure embodies the signature, delicate, hand-painted quality of Hayao Miyazaki's animated films. Amidst a cold world shrouded by thick, swirling blue-gray clouds, a warm, soft orange-yellow glow emanates from within each dome. This light penetrates the damp, rain-scoured surface, casting warm patches of light on the soggy ground outside and subtly illuminating the lush, exotic flora within. The violence of the outside world forms a strong dramatic contrast with the tranquility and warmth of the inner space. This light source that persists in the storm is like a quiet and powerful oasis of life, creating a grand atmosphere full of hope and vitality in the harsh environment of doomsday.", + "seed": 14495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_14/index_7654_seed_14495_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04816260118968785, + "vlm_logprob": -0.04816260118968785, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.04816260118968785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0887778103351593 + ], + "opposite_logprob_tokens": [ + -2.588777780532837 + ], + "logprob_reward": -0.0887778103351593, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0887778103351593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07878623157739639, + 0.08310865610837936, + 0.05525549128651619, + 0.12481462955474854, + 0.02872190810739994, + 0.14561671018600464, + 0.11442449688911438, + 0.18180377781391144, + 0.10286828875541687, + 0.0845998004078865 + ], + "credit_map": [ + 0.007290644571185112, + 0.009140031412243843, + 0.013899674639105797, + 0.010783006437122822, + 0.01444206666201353, + 0.008463844656944275, + 0.0076032462529838085, + 0.007163719739764929, + 0.009914765134453773, + 0.009040581062436104, + 0.011436078697443008, + 0.012090473435819149, + 0.007824970409274101, + 0.013364167883992195, + 0.011121710762381554, + 0.008315910585224628, + 0.005763165652751923, + 0.009405050426721573, + 0.008130981586873531, + 0.007797106169164181, + 0.005148041062057018, + 0.006507284939289093, + 0.0066397967748343945, + 0.005864064674824476, + 0.013646747916936874, + 0.015156537294387817, + 0.02377828024327755, + 0.01847168430685997, + 0.0154067138209939, + 0.010787693783640862, + 0.01856594905257225, + 0.009001016616821289, + 0.0026263128966093063, + 0.003631685394793749, + 0.006108040455728769, + 0.0040208157151937485, + 0.0035317351575940847, + 0.0028012667316943407, + 0.0026915951166301966, + 0.003310458268970251, + 0.028746189549565315, + 0.011826638132333755, + 0.022573623806238174, + 0.013222022913396358, + 0.02137124352157116, + 0.016996465623378754, + 0.015158027410507202, + 0.015722503885626793, + 0.015759510919451714, + 0.013057396747171879, + 0.014399979263544083, + 0.017111357301473618, + 0.008460468612611294, + 0.014470559544861317, + 0.0175704974681139, + 0.01359472330659628, + 0.0076883714646101, + 0.014839702285826206, + 0.007956961169838905, + 0.035565126687288284, + 0.038839977234601974, + 0.04764370992779732, + 0.01278906874358654, + 0.016480857506394386, + 0.01141628809273243, + 0.011281365528702736, + 0.013751321472227573, + 0.008062104694545269, + 0.010164081119000912, + 0.015465975739061832, + 0.014318288303911686, + 0.01840886101126671, + 0.0036178617738187313, + 0.011413440108299255, + 0.004586574621498585, + 0.003602280979976058, + 0.007387538440525532, + 0.027626801282167435, + 0.006050845142453909, + 0.020314468070864677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.0887778103351593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705036997795105 + ], + "opposite_logprob_tokens": [ + -2.1270503997802734 + ], + "logprob_reward": -0.12705036997795105, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705036997795105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16942135989665985, + 0.21205061674118042, + 0.1210891380906105, + 0.10931035131216049, + 0.06753718852996826, + 0.05711980164051056, + 0.0743178203701973, + 0.0820469781756401, + 0.059582047164440155, + 0.04752468690276146 + ], + "credit_map": [ + 0.012752101756632328, + 0.02303604781627655, + 0.021879112347960472, + 0.01586093381047249, + 0.02797766961157322, + 0.02018917351961136, + 0.025565296411514282, + 0.022161025553941727, + 0.01790654845535755, + 0.032240599393844604, + 0.023692119866609573, + 0.02462051995098591, + 0.042982954531908035, + 0.023113388568162918, + 0.01854941062629223, + 0.0289450790733099, + 0.015191682614386082, + 0.013123821467161179, + 0.016047384589910507, + 0.017028970643877983, + 0.01607588119804859, + 0.01476473268121481, + 0.01546346303075552, + 0.013393198139965534, + 0.011053984984755516, + 0.01752742938697338, + 0.01177490409463644, + 0.007598572876304388, + 0.015804218128323555, + 0.014965212903916836, + 0.014972136355936527, + 0.015613900497555733, + 0.005396069493144751, + 0.012395128607749939, + 0.004463722463697195, + 0.007327814120799303, + 0.009937298484146595, + 0.014789101667702198, + 0.004954324569553137, + 0.008273734711110592, + 0.008926727809011936, + 0.005516411270946264, + 0.0074900672771036625, + 0.006004346534609795, + 0.009564349427819252, + 0.007692456711083651, + 0.005525893997400999, + 0.006399549078196287, + 0.00890570692718029, + 0.007213688921183348, + 0.007540903519839048, + 0.009726985357701778, + 0.008142522536218166, + 0.009457756765186787, + 0.011698461137712002, + 0.0116317979991436, + 0.005231249146163464, + 0.007754776626825333, + 0.0051171076484024525, + 0.021708296611905098, + 0.015080469660460949, + 0.013620221987366676, + 0.007961849682033062, + 0.005573008209466934, + 0.006383893080055714, + 0.007526650559157133, + 0.009631119668483734, + 0.005430399905890226, + 0.0063887243159115314, + 0.00827878899872303, + 0.008706039749085903, + 0.007236427161842585, + 0.004855487961322069, + 0.005580345634371042, + 0.004875562619417906, + 0.004685691557824612, + 0.005597693845629692, + 0.007476890459656715, + 0.007308338303118944, + 0.007144675124436617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.12705036997795105 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022017541341483593 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.0022017541341483593, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022017541341483593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05174170807003975, + 0.03768841549754143, + 0.04669574648141861, + 0.06731688976287842, + 0.08664616197347641, + 0.030673399567604065, + 0.04673687741160393, + 0.08603285253047943, + 0.05750230327248573, + 0.48896557092666626 + ], + "credit_map": [ + 0.004663615021854639, + 0.004721319768577814, + 0.013384765014052391, + 0.004595302511006594, + 0.0072794207371771336, + 0.004531788174062967, + 0.005891385953873396, + 0.006674114614725113, + 0.002453354187309742, + 0.004218721762299538, + 0.0030408387538045645, + 0.004121159669011831, + 0.00502368900924921, + 0.006996546871960163, + 0.007258790545165539, + 0.004575320519506931, + 0.005170839373022318, + 0.006339830346405506, + 0.0073870145715773106, + 0.005582969170063734, + 0.003939781803637743, + 0.004429199267178774, + 0.006012605503201485, + 0.007833512499928474, + 0.0060823047533631325, + 0.00840030051767826, + 0.012239857576787472, + 0.010381540283560753, + 0.005628296639770269, + 0.003264235332608223, + 0.012369511649012566, + 0.008950847201049328, + 0.005250801797956228, + 0.010903110727667809, + 0.010931908152997494, + 0.017311519011855125, + 0.01187119260430336, + 0.009501060470938683, + 0.010831373743712902, + 0.010045204311609268, + 0.004605469759553671, + 0.0033497039694339037, + 0.004591131117194891, + 0.002942268270999193, + 0.003440381959080696, + 0.0037158040795475245, + 0.0035714698024094105, + 0.004457172006368637, + 0.008444895036518574, + 0.00448844488710165, + 0.006239206064492464, + 0.007127885241061449, + 0.0053971256129443645, + 0.005737746134400368, + 0.004792716354131699, + 0.004508857615292072, + 0.003594061126932502, + 0.003686395240947604, + 0.003357946639880538, + 0.010825054720044136, + 0.027725109830498695, + 0.02818118780851364, + 0.0037152867298573256, + 0.004947821144014597, + 0.006733261048793793, + 0.00741166016086936, + 0.006602584850043058, + 0.004749807994812727, + 0.00671407300978899, + 0.008867350406944752, + 0.007928136736154556, + 0.008495435118675232, + 0.017047099769115448, + 0.03363625705242157, + 0.01268912572413683, + 0.011028164997696877, + 0.017873596400022507, + 0.18738868832588196, + 0.023047152906656265, + 0.186255544424057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.0022017541341483593 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198661444708705 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198661444708705, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198661444708705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08711899071931839, + 0.04824870824813843, + 0.062026914209127426, + 0.03379230201244354, + 0.043166689574718475, + 0.04917169362306595, + 0.14373110234737396, + 0.15872816741466522, + 0.0680597648024559, + 0.3059556186199188 + ], + "credit_map": [ + 0.0054602543823421, + 0.009706314653158188, + 0.016446566209197044, + 0.009622596204280853, + 0.01319065596908331, + 0.0101506682112813, + 0.010176608338952065, + 0.01236533559858799, + 0.0029720983002334833, + 0.005814034026116133, + 0.003689569653943181, + 0.006899374071508646, + 0.007164555601775646, + 0.007260721642524004, + 0.008699634112417698, + 0.00574872363358736, + 0.006805421318858862, + 0.007761360611766577, + 0.008508127182722092, + 0.007819327525794506, + 0.0077784559689462185, + 0.008555700071156025, + 0.007554835174232721, + 0.007243689615279436, + 0.003263182006776333, + 0.004195894580334425, + 0.00403045117855072, + 0.004292374476790428, + 0.004503078293055296, + 0.0026191165670752525, + 0.006759499665349722, + 0.004128701984882355, + 0.0035884135868400335, + 0.00583296176046133, + 0.008323965594172478, + 0.005725059192627668, + 0.006296841893345118, + 0.005087214522063732, + 0.0036719683557748795, + 0.004640266764909029, + 0.008873498067259789, + 0.004902807995676994, + 0.008202014490962029, + 0.004920316394418478, + 0.007201394997537136, + 0.006499954499304295, + 0.004042863845825195, + 0.0045288424007594585, + 0.021045317873358727, + 0.019168181344866753, + 0.015765640884637833, + 0.024357104673981667, + 0.014269116334617138, + 0.017663536593317986, + 0.01724567823112011, + 0.014216536656022072, + 0.00784516055136919, + 0.006600784137845039, + 0.00439923070371151, + 0.010341341607272625, + 0.053861021995544434, + 0.05289027467370033, + 0.010285264812409878, + 0.012505078688263893, + 0.006843287032097578, + 0.01236794888973236, + 0.008261432871222496, + 0.006665909197181463, + 0.007074806373566389, + 0.008438640274107456, + 0.0098195755854249, + 0.008588171564042568, + 0.02619081176817417, + 0.04129253327846527, + 0.03465672582387924, + 0.030065160244703293, + 0.036553021520376205, + 0.05081936717033386, + 0.04777934029698372, + 0.038598667830228806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.002198661444708705 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584410056471825 + ], + "opposite_logprob_tokens": [ + -3.8955843448638916 + ], + "logprob_reward": -0.020584410056471825, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584410056471825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1551022082567215, + 0.07827046513557434, + 0.05308792367577553, + 0.05104873701930046, + 0.04426247626543045, + 0.042537543922662735, + 0.09100852161645889, + 0.07840417325496674, + 0.038772691041231155, + 0.36750528216362 + ], + "credit_map": [ + 0.010800281539559364, + 0.022246532142162323, + 0.01798788458108902, + 0.017855875194072723, + 0.029541676864027977, + 0.015684010460972786, + 0.020021893084049225, + 0.020964067429304123, + 0.0049036480486392975, + 0.009531466290354729, + 0.006161114200949669, + 0.010214242152869701, + 0.01466362550854683, + 0.010479677468538284, + 0.012424749322235584, + 0.009891948662698269, + 0.006433779839426279, + 0.007470667362213135, + 0.007339806295931339, + 0.007122691720724106, + 0.005904905963689089, + 0.004932420328259468, + 0.007557551842182875, + 0.006326095666736364, + 0.0038646417669951916, + 0.006203635595738888, + 0.006895515136420727, + 0.007681997958570719, + 0.005708454176783562, + 0.004523213487118483, + 0.008410483598709106, + 0.007760798092931509, + 0.0026043117977678776, + 0.0072328788228333, + 0.004304087720811367, + 0.009041543118655682, + 0.006447480991482735, + 0.005787112284451723, + 0.004310729447752237, + 0.004534331616014242, + 0.0051055168733000755, + 0.004414671566337347, + 0.0063142962753772736, + 0.006500301416963339, + 0.005723246838897467, + 0.005565755069255829, + 0.00485004810616374, + 0.00406370684504509, + 0.009144827723503113, + 0.00758656021207571, + 0.008670649491250515, + 0.009718724526464939, + 0.013558955863118172, + 0.01385977491736412, + 0.016113590449094772, + 0.012355439364910126, + 0.006587281823158264, + 0.007309495471417904, + 0.006011790595948696, + 0.008960511535406113, + 0.011431485414505005, + 0.012986072339117527, + 0.0098594781011343, + 0.015258057974278927, + 0.0040272600017488, + 0.005475252866744995, + 0.004010916221886873, + 0.003096651518717408, + 0.005148187279701233, + 0.004870437551289797, + 0.007158222608268261, + 0.004985760897397995, + 0.013718890957534313, + 0.028066381812095642, + 0.012070199474692345, + 0.010675711557269096, + 0.01603049598634243, + 0.13607636094093323, + 0.023130131885409355, + 0.12773708999156952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.020584410056471825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot, viewed from a slightly upward angle, glides steadily and leisurely from left to right amidst the raging typhoon. Bean-sized raindrops, almost forming a curtain of water, cascade down, lashing every surface. A howling wind sways the surrounding vegetation. The camera slowly sweeps across a futuristic cluster of intelligent greenhouses, stacked on the mountainside. Composed of translucent bio-domes and delicate metal frames, the structure embodies the signature, delicate, hand-painted quality of Hayao Miyazaki's animated films. Amidst a cold world shrouded by thick, swirling blue-gray clouds, a warm, soft orange-yellow glow emanates from within each dome. This light penetrates the damp, rain-scoured surface, casting warm patches of light on the soggy ground outside and subtly illuminating the lush, exotic flora within. The violence of the outside world forms a strong dramatic contrast with the tranquility and warmth of the inner space. This light source that persists in the storm is like a quiet and powerful oasis of life, creating a grand atmosphere full of hope and vitality in the harsh environment of doomsday.", + "seed": 14496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_14/index_7654_seed_14496_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11311064932961017, + "vlm_logprob": -0.11311064932961017, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.11311064932961017, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06796113401651382 + ], + "opposite_logprob_tokens": [ + -2.8179612159729004 + ], + "logprob_reward": -0.06796113401651382, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06796113401651382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11785995960235596, + 0.08720965683460236, + 0.03638198599219322, + 0.10247781127691269, + 0.28577274084091187, + 0.06316549330949783, + 0.08161818981170654, + 0.10230612009763718, + 0.09186194092035294, + 0.03134612739086151 + ], + "credit_map": [ + 0.011691407300531864, + 0.011211548931896687, + 0.016636228188872337, + 0.020250767469406128, + 0.009716291911900043, + 0.020846907049417496, + 0.01360281277447939, + 0.013903980143368244, + 0.008473261259496212, + 0.00897968839854002, + 0.018832171335816383, + 0.005644570104777813, + 0.016237860545516014, + 0.012920188717544079, + 0.008585356175899506, + 0.007536549121141434, + 0.00462486082687974, + 0.005588062107563019, + 0.0061703091487288475, + 0.0024967121426016092, + 0.004370018839836121, + 0.005811729002743959, + 0.003967659082263708, + 0.0033526290208101273, + 0.012282554060220718, + 0.011633931659162045, + 0.01190175674855709, + 0.012133787386119366, + 0.01318458653986454, + 0.01614796184003353, + 0.01115594245493412, + 0.014037273824214935, + 0.03846805915236473, + 0.038686953485012054, + 0.02184019796550274, + 0.02357158437371254, + 0.051640260964632034, + 0.05630309507250786, + 0.029222380369901657, + 0.026040181517601013, + 0.008508924394845963, + 0.007876436226069927, + 0.007105544209480286, + 0.007422144524753094, + 0.005855237599462271, + 0.009921985678374767, + 0.010205305181443691, + 0.006269907578825951, + 0.01067629549652338, + 0.010268686339259148, + 0.006941884756088257, + 0.010669411160051823, + 0.012353715486824512, + 0.01881680265069008, + 0.0065315281972289085, + 0.005359861999750137, + 0.014005834236741066, + 0.013230050913989544, + 0.008533221669495106, + 0.010138288140296936, + 0.013629967346787453, + 0.016794590279459953, + 0.014456309378147125, + 0.011517851613461971, + 0.01295384019613266, + 0.01260860450565815, + 0.01326288003474474, + 0.008794733323156834, + 0.01198496762663126, + 0.01037683431059122, + 0.01073837373405695, + 0.011141694150865078, + 0.005034771282225847, + 0.002407040446996689, + 0.008003460243344307, + 0.002961262594908476, + 0.0040639303624629974, + 0.003910487983375788, + 0.0023507478181272745, + 0.0026144245639443398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.06796113401651382 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742511212825775 + ], + "opposite_logprob_tokens": [ + -0.9742511510848999 + ], + "logprob_reward": -0.4742511212825775, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742511212825775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1529141366481781, + 0.09325549006462097, + 0.14337632060050964, + 0.20453493297100067, + 0.13205088675022125, + 0.09550199657678604, + 0.03303524851799011, + 0.05679429695010185, + 0.04665963351726532, + 0.04187706857919693 + ], + "credit_map": [ + 0.01807321049273014, + 0.015477675013244152, + 0.01281739491969347, + 0.014156355522572994, + 0.020029401406645775, + 0.04233243316411972, + 0.01311490312218666, + 0.016912750899791718, + 0.010438177734613419, + 0.011617003008723259, + 0.016823073849081993, + 0.007647612597793341, + 0.013708333484828472, + 0.014039081521332264, + 0.009699990972876549, + 0.009282215498387814, + 0.022485952824354172, + 0.02099831961095333, + 0.019001103937625885, + 0.01491155568510294, + 0.012819170951843262, + 0.018011242151260376, + 0.019463596865534782, + 0.01568537764251232, + 0.026734402403235435, + 0.02600642293691635, + 0.030358638614416122, + 0.022001542150974274, + 0.025346849113702774, + 0.027615422382950783, + 0.022076476365327835, + 0.024395177140831947, + 0.01774532161653042, + 0.017674000933766365, + 0.019560692831873894, + 0.017572326585650444, + 0.014821097254753113, + 0.01250944659113884, + 0.013967657461762428, + 0.01820034720003605, + 0.011880050413310528, + 0.014277291484177113, + 0.01023871824145317, + 0.010437089018523693, + 0.01145174540579319, + 0.012534292414784431, + 0.012538367882370949, + 0.01214444637298584, + 0.004365714266896248, + 0.004836930427700281, + 0.005638145841658115, + 0.003455545986071229, + 0.003040996380150318, + 0.00499724643304944, + 0.002764578443020582, + 0.003936091437935829, + 0.007269388530403376, + 0.009910304099321365, + 0.005507836118340492, + 0.007146029733121395, + 0.007575748488306999, + 0.007793614640831947, + 0.005515675991773605, + 0.006075701210647821, + 0.007125353440642357, + 0.005055586341768503, + 0.007070300634950399, + 0.00533272372558713, + 0.006290001329034567, + 0.006533415988087654, + 0.004213761072605848, + 0.0050384895876049995, + 0.00584250595420599, + 0.004142363090068102, + 0.004437555093318224, + 0.005600005853921175, + 0.0047353035770356655, + 0.005968568380922079, + 0.006936169229447842, + 0.004214594140648842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.4742511212825775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019455092260614038 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019455092260614038, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019455092260614038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1438080370426178, + 0.060068026185035706, + 0.12350738048553467, + 0.0833790972828865, + 0.16781695187091827, + 0.05832022801041603, + 0.07169808447360992, + 0.08476641029119492, + 0.10402145236730576, + 0.10261424630880356 + ], + "credit_map": [ + 0.013765254989266396, + 0.014318565838038921, + 0.020092487335205078, + 0.021488677710294724, + 0.017464345321059227, + 0.027515482157468796, + 0.013750610873103142, + 0.01541261374950409, + 0.007129831239581108, + 0.00619518244639039, + 0.01017939392477274, + 0.004506859462708235, + 0.0059120324440300465, + 0.007534409873187542, + 0.006893395446240902, + 0.011716930195689201, + 0.013587105087935925, + 0.014426294714212418, + 0.02855168841779232, + 0.015608706511557102, + 0.011553612537682056, + 0.015334091149270535, + 0.014674261212348938, + 0.00977163203060627, + 0.01116868294775486, + 0.010799210518598557, + 0.010563453659415245, + 0.011399360373616219, + 0.012504479847848415, + 0.010663114488124847, + 0.008085348643362522, + 0.008195450529456139, + 0.02464534156024456, + 0.02436968870460987, + 0.020449435338377953, + 0.016864577308297157, + 0.023139135912060738, + 0.021933574229478836, + 0.015975212678313255, + 0.020440004765987396, + 0.006874202750623226, + 0.009529615752398968, + 0.006437317002564669, + 0.007396656554192305, + 0.006703467108309269, + 0.006858518347144127, + 0.008271301165223122, + 0.006249156314879656, + 0.010902433656156063, + 0.011303573846817017, + 0.007219579070806503, + 0.0076250494457781315, + 0.009758696891367435, + 0.015821652486920357, + 0.0045090410858392715, + 0.0045580631121993065, + 0.011907090432941914, + 0.012082093395292759, + 0.008349206298589706, + 0.007953809574246407, + 0.014236659742891788, + 0.016247304156422615, + 0.006408435292541981, + 0.007581816054880619, + 0.015346823260188103, + 0.014304435811936855, + 0.016662828624248505, + 0.008956553414463997, + 0.01737823337316513, + 0.015377030707895756, + 0.008289816789329052, + 0.007705741096287966, + 0.013813277706503868, + 0.009698883630335331, + 0.015677116811275482, + 0.010791885666549206, + 0.011495355516672134, + 0.014176655560731888, + 0.01716669835150242, + 0.009794380515813828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0019455092260614038 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008142255246639252 + ], + "opposite_logprob_tokens": [ + -7.125814437866211 + ], + "logprob_reward": -0.0008142255246639252, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008142255246639252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.206743985414505, + 0.0846569612622261, + 0.1874380260705948, + 0.17229656875133514, + 0.099382184445858, + 0.07733470946550369, + 0.03839235007762909, + 0.038457468152046204, + 0.05227721482515335, + 0.04302046075463295 + ], + "credit_map": [ + 0.020035767927765846, + 0.02309134230017662, + 0.021212531253695488, + 0.03016875684261322, + 0.022710978984832764, + 0.04486702010035515, + 0.022001411765813828, + 0.022656183689832687, + 0.01059330627322197, + 0.009805653244256973, + 0.01348640862852335, + 0.008032410405576229, + 0.006453015375882387, + 0.013183158822357655, + 0.007819416001439095, + 0.015283596701920033, + 0.028635533526539803, + 0.02196415886282921, + 0.026241451501846313, + 0.014800208620727062, + 0.02268626168370247, + 0.04202626645565033, + 0.017516592517495155, + 0.013567546382546425, + 0.02285635657608509, + 0.02335602417588234, + 0.020362991839647293, + 0.018846796825528145, + 0.02310137078166008, + 0.024248911067843437, + 0.01740751601755619, + 0.02211659960448742, + 0.012182289734482765, + 0.014551959000527859, + 0.01090828888118267, + 0.013401953503489494, + 0.012113019824028015, + 0.012479015626013279, + 0.008826804347336292, + 0.014918860048055649, + 0.009420682676136494, + 0.009996959008276463, + 0.00881173461675644, + 0.008251798339188099, + 0.011887330561876297, + 0.013108422048389912, + 0.009215815924108028, + 0.006641968619078398, + 0.005660787224769592, + 0.005430165212601423, + 0.004490951541811228, + 0.0042225392535328865, + 0.004394151270389557, + 0.006303330883383751, + 0.0036383364349603653, + 0.004252090584486723, + 0.005286683794111013, + 0.005527050234377384, + 0.004861051216721535, + 0.005383671727031469, + 0.004594271536916494, + 0.0061558024026453495, + 0.0031881125178188086, + 0.003460824256762862, + 0.008204275742173195, + 0.0060549164190888405, + 0.008447648026049137, + 0.004106645472347736, + 0.007216694299131632, + 0.008241256698966026, + 0.0044858879409730434, + 0.00551989022642374, + 0.006222971249371767, + 0.006666653789579868, + 0.004787581041455269, + 0.003831292036920786, + 0.004869177006185055, + 0.0062117851339280605, + 0.00571091566234827, + 0.004720084834843874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0008142255246639252 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth tracking shot that glides steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a raging typhoon with bean-sized raindrops forming a curtain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there futuristic intelligent greenhouses stacked on a mountainside composed of translucent bio-domes and metal frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a warm orange-yellow glow emanating from within the domes contrasting with the cold blue-gray storm outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by a delicate, hand-painted quality reminiscent of Hayao Miyazaki's animated films?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581256598234177 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581256598234177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581256598234177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2102365344762802, + 0.06213945522904396, + 0.03527012839913368, + 0.057810742408037186, + 0.10310077667236328, + 0.1776130348443985, + 0.09904660284519196, + 0.08426208794116974, + 0.0938357263803482, + 0.07668489217758179 + ], + "credit_map": [ + 0.020108511671423912, + 0.019977165386080742, + 0.023557521402835846, + 0.043086327612400055, + 0.02924024686217308, + 0.029943067580461502, + 0.016828885301947594, + 0.027494831010699272, + 0.006733998656272888, + 0.007735803723335266, + 0.009350409731268883, + 0.0048404885455966, + 0.008371577598154545, + 0.009358554147183895, + 0.0056532700546085835, + 0.010095354169607162, + 0.005168925505131483, + 0.004623113665729761, + 0.00758717255666852, + 0.0027560240123420954, + 0.002744306344538927, + 0.004237210378050804, + 0.004718408919870853, + 0.0034349707420915365, + 0.008638106286525726, + 0.007445068098604679, + 0.0069336299784481525, + 0.008217164315283298, + 0.007563714403659105, + 0.007230543997138739, + 0.004740216303616762, + 0.007042307872325182, + 0.010225687175989151, + 0.0117422454059124, + 0.012687706388533115, + 0.014310245402157307, + 0.013234849087893963, + 0.016433320939540863, + 0.01008941326290369, + 0.014377313666045666, + 0.02193780243396759, + 0.028818698599934578, + 0.015444444492459297, + 0.014791570603847504, + 0.020548230037093163, + 0.02423817850649357, + 0.025407731533050537, + 0.026426376774907112, + 0.013580892235040665, + 0.013830389827489853, + 0.009106690064072609, + 0.00965573638677597, + 0.013767614960670471, + 0.022580625489354134, + 0.010064872913062572, + 0.00645978469401598, + 0.012420631013810635, + 0.0119003402069211, + 0.008436760865151882, + 0.0065467627719044685, + 0.00993929523974657, + 0.01829618588089943, + 0.008243170566856861, + 0.008478944189846516, + 0.014330081641674042, + 0.012943578884005547, + 0.012701988220214844, + 0.008594718761742115, + 0.017830444499850273, + 0.013846661895513535, + 0.006839539855718613, + 0.0067487116903066635, + 0.011049770750105381, + 0.01220277976244688, + 0.010413823649287224, + 0.00859073456376791, + 0.007703840266913176, + 0.010476754978299141, + 0.008273892104625702, + 0.00797329843044281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.020581256598234177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_15/index_tensor([1244, 1244])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_15/index_tensor([1244, 1244])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..704cbd11adbe603b6f33aad269077537500a9684 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_15/index_tensor([1244, 1244])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow and graceful tracking shot follows a translucent jellyfish as it glides effortlessly through the abyssal depths of the deep ocean. The creature's bell, a nearly transparent, gelatinous dome, pulsates gently and rhythmically, propelling it forward with a tranquil, contemplative spirit. From beneath the bell, a string of delicate, luminous tendrils trails, each a long, ethereal ray of light, emitting a vibrant, soft bioluminescence in ever-shifting hues of electric blue and soft purple. The surrounding water is an inky, profound black, a perfect canvas, making the jellyfish the sole source of light. This inner glow creates a soft, ethereal shimmer nearby, highlighting the mesmerizing, hypnotic dance of the trailing tentacles as they swirl and undulate with the subtle currents of the water. The entire scene, captured in silent, continuous motion, creates an atmosphere of profound tranquility and otherworldliness, one of solitary beauty and quiet wonder.", + "seed": 26395, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_15/index_1244_seed_26395_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05038688536733389, + "vlm_logprob": -0.05038688536733389, + "vlm_yes_no_margin": 3.725000023841858, + "vlm_token_credit": -0.05038688536733389, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006025008857250214 + ], + "opposite_logprob_tokens": [ + -5.7560248374938965 + ], + "logprob_reward": -0.006025008857250214, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.006025008857250214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06696092337369919, + 0.03083045780658722, + 0.09988951683044434, + 0.04892563447356224, + 0.04922056198120117, + 0.03548162803053856, + 0.22267553210258484, + 0.09117884933948517, + 0.15082550048828125, + 0.20401138067245483 + ], + "credit_map": [ + 0.006096669938415289, + 0.006053345277905464, + 0.011668330989778042, + 0.005516516976058483, + 0.010993732139468193, + 0.010531289502978325, + 0.0049663991667330265, + 0.011134634725749493, + 0.0032781576737761497, + 0.0035166903398931026, + 0.0043060071766376495, + 0.002973838010802865, + 0.005029421765357256, + 0.004790870938450098, + 0.001840066397562623, + 0.0050954013131558895, + 0.013591176830232143, + 0.008452063426375389, + 0.017032111063599586, + 0.008372428826987743, + 0.014400618150830269, + 0.01601613685488701, + 0.007503510918468237, + 0.014521466568112373, + 0.003963562659919262, + 0.0032110444735735655, + 0.007205578964203596, + 0.0037285741418600082, + 0.005351364612579346, + 0.008227361366152763, + 0.005403708200901747, + 0.011834436096251011, + 0.004655545577406883, + 0.0047557661309838295, + 0.004229527898132801, + 0.0060448660515248775, + 0.00818417128175497, + 0.006378249730914831, + 0.009314161725342274, + 0.0056582726538181305, + 0.0026519685052335262, + 0.0039059012196958065, + 0.0039220587350428104, + 0.006295728962868452, + 0.007188914809376001, + 0.004250481259077787, + 0.004407270811498165, + 0.0028592990711331367, + 0.024816932156682014, + 0.020239373669028282, + 0.04409955069422722, + 0.033413417637348175, + 0.018422456458210945, + 0.03217719495296478, + 0.02003161981701851, + 0.029474973678588867, + 0.01330845057964325, + 0.005452199839055538, + 0.0075875502079725266, + 0.00485266325995326, + 0.00765528529882431, + 0.020485764369368553, + 0.005865512415766716, + 0.025971418246626854, + 0.01492506917566061, + 0.012154250405728817, + 0.012034319341182709, + 0.032149869948625565, + 0.007538629695773125, + 0.027840357273817062, + 0.012964406050741673, + 0.031218601390719414, + 0.018973836675286293, + 0.018185842782258987, + 0.02185913920402527, + 0.03304186463356018, + 0.016928445547819138, + 0.02983991988003254, + 0.034389108419418335, + 0.03079320676624775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.006025008857250214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330484427511692 + ], + "opposite_logprob_tokens": [ + -3.7733049392700195 + ], + "logprob_reward": -0.02330484427511692, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330484427511692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034029021859169006, + 0.08676713705062866, + 0.06816888600587845, + 0.04377739503979683, + 0.04780390486121178, + 0.1380987912416458, + 0.4125049412250519, + 0.043694447726011276, + 0.05075570568442345, + 0.07439979165792465 + ], + "credit_map": [ + 0.0042287432588636875, + 0.002579407300800085, + 0.0030423833522945642, + 0.002820781199261546, + 0.008325845003128052, + 0.005866541992872953, + 0.0021370667964220047, + 0.005028255749493837, + 0.010450261645019054, + 0.00742271076887846, + 0.008944006636738777, + 0.0056662834249436855, + 0.021671826019883156, + 0.013859890401363373, + 0.006503848824650049, + 0.012248304672539234, + 0.006060751620680094, + 0.005154951475560665, + 0.017783712595701218, + 0.005970792844891548, + 0.007262195460498333, + 0.010866913013160229, + 0.0059469910338521, + 0.009122575633227825, + 0.004209844395518303, + 0.003448155941441655, + 0.007742374669760466, + 0.004866621922701597, + 0.005348799284547567, + 0.006111495662480593, + 0.006057634949684143, + 0.005992467049509287, + 0.0067321243695914745, + 0.003471494885161519, + 0.009525107219815254, + 0.00525125116109848, + 0.00334907160140574, + 0.007902304641902447, + 0.004355316050350666, + 0.007217236794531345, + 0.012236441485583782, + 0.011348670348525047, + 0.022860445082187653, + 0.023058269172906876, + 0.01933109574019909, + 0.01844623312354088, + 0.010836132802069187, + 0.0199815072119236, + 0.04596603661775589, + 0.02594728395342827, + 0.12829847633838654, + 0.05692443624138832, + 0.02311667613685131, + 0.06101025268435478, + 0.024894626811146736, + 0.04634712636470795, + 0.004588013980537653, + 0.005674482323229313, + 0.003571457928046584, + 0.0056049153208732605, + 0.00980833824723959, + 0.005214724224060774, + 0.004828701261430979, + 0.0044038137421011925, + 0.005156279541552067, + 0.003674431936815381, + 0.004570389166474342, + 0.010670187883079052, + 0.002728682244196534, + 0.009614820592105389, + 0.003473415272310376, + 0.010867497883737087, + 0.007561816368252039, + 0.006715908646583557, + 0.008201800286769867, + 0.010441047139465809, + 0.005894617177546024, + 0.011578449048101902, + 0.011635389178991318, + 0.012370764277875423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.02330484427511692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786837734282017 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786837734282017, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786837734282017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10558167845010757, + 0.08122358471155167, + 0.08840674161911011, + 0.2024155557155609, + 0.037962350994348526, + 0.11520248651504517, + 0.14514385163784027, + 0.04648839682340622, + 0.05554347112774849, + 0.12203192710876465 + ], + "credit_map": [ + 0.014233320951461792, + 0.007704597432166338, + 0.010978654026985168, + 0.008339566178619862, + 0.020871667191386223, + 0.017956141382455826, + 0.008525936864316463, + 0.016971789300441742, + 0.009523970074951649, + 0.00903576985001564, + 0.008507654070854187, + 0.005664578173309565, + 0.016330722719430923, + 0.012376558035612106, + 0.008676647208631039, + 0.011107680387794971, + 0.012671947479248047, + 0.008482878096401691, + 0.014174135401844978, + 0.005997843109071255, + 0.010204236954450607, + 0.015646198764443398, + 0.00497685419395566, + 0.01625264436006546, + 0.018082531169056892, + 0.013568484224379063, + 0.053883299231529236, + 0.015996862202882767, + 0.01738622412085533, + 0.025778481736779213, + 0.033446524292230606, + 0.02427314966917038, + 0.005602399818599224, + 0.002731641987338662, + 0.004950256086885929, + 0.0034537306055426598, + 0.003093710634857416, + 0.0067816562950611115, + 0.0034904249478131533, + 0.00785853061825037, + 0.010277851484715939, + 0.008735395036637783, + 0.017796892672777176, + 0.017997683957219124, + 0.017937585711479187, + 0.01597418636083603, + 0.008986293338239193, + 0.01749659702181816, + 0.017119329422712326, + 0.010136193595826626, + 0.039194315671920776, + 0.020702514797449112, + 0.008786176331341267, + 0.020979033783078194, + 0.010247820988297462, + 0.017978472635149956, + 0.005149506032466888, + 0.003674071282148361, + 0.004220847971737385, + 0.0034280559048056602, + 0.005371804349124432, + 0.009507428854703903, + 0.0031990478746593, + 0.011937635019421577, + 0.005281612742692232, + 0.0038115361239761114, + 0.006215105764567852, + 0.012896968983113766, + 0.0036732463631778955, + 0.00919459480792284, + 0.005993736442178488, + 0.00847667083144188, + 0.01364914607256651, + 0.009388486854732037, + 0.012563887983560562, + 0.015442335046827793, + 0.008719154633581638, + 0.020736785605549812, + 0.016978872939944267, + 0.02455325610935688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.009786837734282017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281192421913147 + ], + "opposite_logprob_tokens": [ + -2.0178120136260986 + ], + "logprob_reward": -0.14281192421913147, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14281192421913147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1187451183795929, + 0.09729117900133133, + 0.2942657470703125, + 0.026680905371904373, + 0.050585437566041946, + 0.07296597212553024, + 0.13609176874160767, + 0.029146142303943634, + 0.06782696396112442, + 0.1064007356762886 + ], + "credit_map": [ + 0.007126329466700554, + 0.006865801755338907, + 0.007666145451366901, + 0.00760132959112525, + 0.02242395281791687, + 0.03395073115825653, + 0.007872821763157845, + 0.025238005444407463, + 0.011755392886698246, + 0.010152612812817097, + 0.011981661431491375, + 0.007229058537632227, + 0.015525764785706997, + 0.015492202714085579, + 0.010976959951221943, + 0.014177531935274601, + 0.04762551188468933, + 0.02429240569472313, + 0.04034808278083801, + 0.022540276870131493, + 0.04547161981463432, + 0.05081735551357269, + 0.019121596589684486, + 0.044048894196748734, + 0.0023648194037377834, + 0.002112150890752673, + 0.004442596808075905, + 0.00285701104439795, + 0.0034003674518316984, + 0.0033010670449584723, + 0.004404369276016951, + 0.003798523684963584, + 0.009540416300296783, + 0.003521844046190381, + 0.006148632150143385, + 0.0040896679274737835, + 0.0035492887254804373, + 0.009166372939944267, + 0.003695053281262517, + 0.01087416335940361, + 0.006577442865818739, + 0.007008457090705633, + 0.009933248162269592, + 0.011922765523195267, + 0.011251368559896946, + 0.01001485250890255, + 0.006834964267909527, + 0.009422879666090012, + 0.016610752791166306, + 0.013623882085084915, + 0.014686344191432, + 0.01781163178384304, + 0.014099178835749626, + 0.021692294627428055, + 0.014543990604579449, + 0.0230237003415823, + 0.0023863427340984344, + 0.003361823270097375, + 0.0034298067912459373, + 0.003263866063207388, + 0.004576152190566063, + 0.004869974683970213, + 0.00270187109708786, + 0.004556305706501007, + 0.006952745374292135, + 0.004527611657977104, + 0.006240861490368843, + 0.015561266802251339, + 0.003565537976101041, + 0.012465260922908783, + 0.0050898343324661255, + 0.013423845171928406, + 0.01518516056239605, + 0.006566271185874939, + 0.008059763349592686, + 0.00975791085511446, + 0.0070956735871732235, + 0.021480606868863106, + 0.00956337433308363, + 0.028691979125142097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.14281192421913147 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000581175088882 + ], + "opposite_logprob_tokens": [ + -2.6950058937072754 + ], + "logprob_reward": -0.07000581175088882, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000581175088882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10910088568925858, + 0.08664988726377487, + 0.23265384137630463, + 0.03805520012974739, + 0.03789687529206276, + 0.06893587112426758, + 0.12569303810596466, + 0.03670049086213112, + 0.06783034652471542, + 0.1964835673570633 + ], + "credit_map": [ + 0.008342727087438107, + 0.006266717333346605, + 0.0065660541877150536, + 0.006863237824290991, + 0.017402466386556625, + 0.0315929651260376, + 0.0068392762914299965, + 0.025227436795830727, + 0.008073419332504272, + 0.012669338844716549, + 0.013886570930480957, + 0.008643153123557568, + 0.011789456009864807, + 0.011922337114810944, + 0.008946464397013187, + 0.01071914378553629, + 0.038200344890356064, + 0.024205122143030167, + 0.032311949878931046, + 0.018339427188038826, + 0.03063287027180195, + 0.04005851969122887, + 0.015680359676480293, + 0.03322526812553406, + 0.0038017528131604195, + 0.0024580692406743765, + 0.009838945232331753, + 0.0028676665388047695, + 0.00324808806180954, + 0.006137664895504713, + 0.004423674661666155, + 0.0052793375216424465, + 0.006616200786083937, + 0.0027911635115742683, + 0.005120600573718548, + 0.0032718030270189047, + 0.0021347024012356997, + 0.007115255109965801, + 0.0025612458121031523, + 0.008285904303193092, + 0.0058610220439732075, + 0.005617876071482897, + 0.010219788178801537, + 0.010186702944338322, + 0.012097114697098732, + 0.009415857493877411, + 0.005972337909042835, + 0.009565170854330063, + 0.0157561544328928, + 0.011689143255352974, + 0.014592256397008896, + 0.01603683829307556, + 0.01280666422098875, + 0.01986517198383808, + 0.013590790331363678, + 0.021356016397476196, + 0.002752788597717881, + 0.00328085757791996, + 0.0055736880749464035, + 0.0032927272841334343, + 0.002632993506267667, + 0.007938914000988007, + 0.0030311408918350935, + 0.008197383023798466, + 0.005673467181622982, + 0.004933485761284828, + 0.0065583293326199055, + 0.016253840178251266, + 0.004713831469416618, + 0.0118716424331069, + 0.005974655505269766, + 0.011851092800498009, + 0.025997545570135117, + 0.012957037426531315, + 0.0161893367767334, + 0.01999775692820549, + 0.013685926795005798, + 0.037574443966150284, + 0.021012360230088234, + 0.04906916618347168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.07000581175088882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful tracking shot follows a translucent jellyfish as it glides effortlessly through the abyssal depths of the deep ocean. The creature's bell, a nearly transparent, gelatinous dome, pulsates gently and rhythmically, propelling it forward with a tranquil, contemplative spirit. From beneath the bell, a string of delicate, luminous tendrils trails, each a long, ethereal ray of light, emitting a vibrant, soft bioluminescence in ever-shifting hues of electric blue and soft purple. The surrounding water is an inky, profound black, a perfect canvas, making the jellyfish the sole source of light. This inner glow creates a soft, ethereal shimmer nearby, highlighting the mesmerizing, hypnotic dance of the trailing tentacles as they swirl and undulate with the subtle currents of the water. The entire scene, captured in silent, continuous motion, creates an atmosphere of profound tranquility and otherworldliness, one of solitary beauty and quiet wonder.", + "seed": 26396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_15/index_1244_seed_26396_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07482000421732664, + "vlm_logprob": -0.07482000421732664, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.07482000421732664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005405571311712265 + ], + "opposite_logprob_tokens": [ + -6.130405426025391 + ], + "logprob_reward": -0.005405571311712265, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005405571311712265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11455388367176056, + 0.060160256922245026, + 0.07332286983728409, + 0.07842489331960678, + 0.08836306631565094, + 0.10617974400520325, + 0.149957075715065, + 0.07703769207000732, + 0.19943109154701233, + 0.05256936326622963 + ], + "credit_map": [ + 0.009471474215388298, + 0.013098761439323425, + 0.014914674684405327, + 0.012475525960326195, + 0.014791915193200111, + 0.021426035091280937, + 0.015904176980257034, + 0.012471320107579231, + 0.006372541654855013, + 0.006621335633099079, + 0.0074701267294585705, + 0.007259022910147905, + 0.006801587995141745, + 0.010532064363360405, + 0.008104328997433186, + 0.006999252829700708, + 0.00950957927852869, + 0.005523500498384237, + 0.007907039485871792, + 0.006784623954445124, + 0.007029934320598841, + 0.011153168976306915, + 0.00915396399796009, + 0.016261069104075432, + 0.007715665735304356, + 0.007976890541613102, + 0.00909170601516962, + 0.007433305960148573, + 0.01097677182406187, + 0.011963841505348682, + 0.00969731342047453, + 0.013569398783147335, + 0.006340268068015575, + 0.008941537700593472, + 0.012045387178659439, + 0.009524412453174591, + 0.009195464663207531, + 0.011930552311241627, + 0.011852094903588295, + 0.018533363938331604, + 0.008459026925265789, + 0.0068430546671152115, + 0.014746306464076042, + 0.016078395769000053, + 0.008381504565477371, + 0.014064716175198555, + 0.010527643375098705, + 0.027079110965132713, + 0.012034722603857517, + 0.014052730053663254, + 0.016178306192159653, + 0.015247192233800888, + 0.018123164772987366, + 0.021801795810461044, + 0.020315425470471382, + 0.03220376372337341, + 0.006766404490917921, + 0.0094891507178545, + 0.008740042336285114, + 0.01256103627383709, + 0.011870048008859158, + 0.010029604658484459, + 0.009127688594162464, + 0.008453726768493652, + 0.012404635548591614, + 0.017080508172512054, + 0.029625531286001205, + 0.027435559779405594, + 0.026565851643681526, + 0.03607529029250145, + 0.022916361689567566, + 0.02732737548649311, + 0.005215901415795088, + 0.0062503754161298275, + 0.007195594254881144, + 0.006047429051250219, + 0.005790427327156067, + 0.005837061908096075, + 0.008300777524709702, + 0.007931798696517944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.005405571311712265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096787638962269 + ], + "opposite_logprob_tokens": [ + -4.511096954345703 + ], + "logprob_reward": -0.011096787638962269, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096787638962269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06348571926355362, + 0.16768266260623932, + 0.10678727179765701, + 0.07096695899963379, + 0.025223873555660248, + 0.08250084519386292, + 0.09119632095098495, + 0.04924861341714859, + 0.29763880372047424, + 0.04526898264884949 + ], + "credit_map": [ + 0.004630108829587698, + 0.005400560330599546, + 0.009381520561873913, + 0.003926048055291176, + 0.005794868338853121, + 0.010861532762646675, + 0.009655113331973553, + 0.013835964724421501, + 0.019399812445044518, + 0.013068793341517448, + 0.023609206080436707, + 0.019048497080802917, + 0.01719311997294426, + 0.02749481424689293, + 0.026864880695939064, + 0.021003536880016327, + 0.012878402136266232, + 0.007168964017182589, + 0.010674173012375832, + 0.009179561398923397, + 0.014676759019494057, + 0.01694643311202526, + 0.013572677038609982, + 0.02169029973447323, + 0.008810214698314667, + 0.006314640864729881, + 0.00824431236833334, + 0.004689523950219154, + 0.006877881940454245, + 0.012244905345141888, + 0.007589463610202074, + 0.01619601808488369, + 0.0021364660933613777, + 0.004125537816435099, + 0.0032871216535568237, + 0.0028472798876464367, + 0.002062723273411393, + 0.003869294188916683, + 0.0025968251284211874, + 0.00429862504824996, + 0.008277841843664646, + 0.01158098503947258, + 0.011057655327022076, + 0.013624973595142365, + 0.006662291940301657, + 0.008684542030096054, + 0.008076246827840805, + 0.01453631091862917, + 0.00823861826211214, + 0.008090094663202763, + 0.010538741014897823, + 0.011528544127941132, + 0.010912737809121609, + 0.012163715437054634, + 0.011670305393636227, + 0.018053561449050903, + 0.004592393524944782, + 0.005208881571888924, + 0.004974527284502983, + 0.007062537129968405, + 0.007882487960159779, + 0.0070579987950623035, + 0.006105028558522463, + 0.006364756729453802, + 0.018310043960809708, + 0.015624859370291233, + 0.09011869132518768, + 0.02327560819685459, + 0.02336161769926548, + 0.029261747375130653, + 0.021237336099147797, + 0.07644888758659363, + 0.0062932660803198814, + 0.004755782894790173, + 0.005801477935165167, + 0.008478161878883839, + 0.004351363051682711, + 0.0037900700699537992, + 0.005384538322687149, + 0.006414319854229689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.011096787638962269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202704653143883 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202704653143883, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202704653143883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08796526491641998, + 0.07798711955547333, + 0.055917803198099136, + 0.04692935198545456, + 0.11246593296527863, + 0.18531621992588043, + 0.15362627804279327, + 0.047725025564432144, + 0.15650437772274017, + 0.07556255161762238 + ], + "credit_map": [ + 0.009911296889185905, + 0.00795320700854063, + 0.010989928618073463, + 0.00879163108766079, + 0.011048346757888794, + 0.018069235607981682, + 0.01017644815146923, + 0.011025166139006615, + 0.01114730630069971, + 0.005418790969997644, + 0.014005996286869049, + 0.008054240606725216, + 0.005737412720918655, + 0.010504071600735188, + 0.009956344030797482, + 0.01316295750439167, + 0.006918194238096476, + 0.005876793060451746, + 0.006874564569443464, + 0.00506380433216691, + 0.00504085049033165, + 0.007857178337872028, + 0.005927606485784054, + 0.01235880982130766, + 0.005405120551586151, + 0.004509160295128822, + 0.005332897882908583, + 0.0044039469212293625, + 0.005281232297420502, + 0.008980073034763336, + 0.0053525990806519985, + 0.007664320524781942, + 0.008666761219501495, + 0.012322383932769299, + 0.0148147102445364, + 0.011179275810718536, + 0.01100433524698019, + 0.014621301554143429, + 0.017760002985596657, + 0.022097164765000343, + 0.016000816598534584, + 0.02306864783167839, + 0.025599230080842972, + 0.03767014294862747, + 0.016682658344507217, + 0.01880667358636856, + 0.018858617171645164, + 0.028629444539546967, + 0.014525840058922768, + 0.014503325335681438, + 0.018975209444761276, + 0.02062126249074936, + 0.01740327663719654, + 0.019836384803056717, + 0.018389366567134857, + 0.029371606186032295, + 0.005155931226909161, + 0.004160831682384014, + 0.004770815372467041, + 0.006794249173253775, + 0.007436197251081467, + 0.006832466460764408, + 0.006311364471912384, + 0.006263175513595343, + 0.009228014387190342, + 0.009911914356052876, + 0.03951900079846382, + 0.013153137639164925, + 0.015253757126629353, + 0.019695885479450226, + 0.013827730901539326, + 0.035914935171604156, + 0.011891472153365612, + 0.010527691803872585, + 0.008429709821939468, + 0.013822464272379875, + 0.0065850308164954185, + 0.006660927087068558, + 0.0086143147200346, + 0.00903093907982111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.014202704653143883 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813638150691986 + ], + "opposite_logprob_tokens": [ + -1.406363844871521 + ], + "logprob_reward": -0.2813638150691986, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813638150691986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025088321417570114, + 0.04774376377463341, + 0.043333444744348526, + 0.031360264867544174, + 0.048125721514225006, + 0.1174735501408577, + 0.1105228140950203, + 0.042320091277360916, + 0.46195289492607117, + 0.0720791295170784 + ], + "credit_map": [ + 0.0026915750931948423, + 0.002705677878111601, + 0.0033130012452602386, + 0.002800591755658388, + 0.002138944575563073, + 0.004152381327003241, + 0.0028767145704478025, + 0.0044094352051615715, + 0.005627473350614309, + 0.005133896134793758, + 0.008020424284040928, + 0.005928333383053541, + 0.004737855400890112, + 0.006842673756182194, + 0.005605387035757303, + 0.0058477213606238365, + 0.006258297245949507, + 0.003657727036625147, + 0.005443222355097532, + 0.003472536336630583, + 0.003874208079650998, + 0.006191117689013481, + 0.004103115759789944, + 0.010333222337067127, + 0.002968247514218092, + 0.002958565019071102, + 0.003971153404563665, + 0.00267283758148551, + 0.00242660753428936, + 0.00458524888381362, + 0.003674825420603156, + 0.00810278207063675, + 0.0037636258639395237, + 0.004376032389700413, + 0.0058889202773571014, + 0.005266399588435888, + 0.005297557916492224, + 0.006493424065411091, + 0.007614173926413059, + 0.00942558329552412, + 0.011679626069962978, + 0.012063181959092617, + 0.014996020123362541, + 0.014767107553780079, + 0.008505292236804962, + 0.014835252426564693, + 0.011344190686941147, + 0.02928287908434868, + 0.00972387008368969, + 0.00945301353931427, + 0.012048042379319668, + 0.010817990638315678, + 0.013272542506456375, + 0.016613736748695374, + 0.014933435246348381, + 0.023660175502300262, + 0.0036889524199068546, + 0.006024877540767193, + 0.004734639544039965, + 0.005913922097533941, + 0.006692894734442234, + 0.005842447280883789, + 0.004912158008664846, + 0.00451020197942853, + 0.028919916599988937, + 0.01836698316037655, + 0.1573268622159958, + 0.025432124733924866, + 0.028208214789628983, + 0.036868251860141754, + 0.03456846624612808, + 0.13226209580898285, + 0.010382852517068386, + 0.007279761601239443, + 0.009596412070095539, + 0.01161117758601904, + 0.006197358015924692, + 0.005447886884212494, + 0.010217227973043919, + 0.011346453800797462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.2813638150691986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203114241361618 + ], + "opposite_logprob_tokens": [ + -2.8120310306549072 + ], + "logprob_reward": -0.06203114241361618, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203114241361618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02946474589407444, + 0.021258842200040817, + 0.02498694695532322, + 0.03289167210459709, + 0.0872993916273117, + 0.1499878317117691, + 0.2319350689649582, + 0.05268171429634094, + 0.3155224919319153, + 0.05397123098373413 + ], + "credit_map": [ + 0.0027712907176464796, + 0.00397602329030633, + 0.003623665077611804, + 0.002666130429133773, + 0.002462374744936824, + 0.003789584618061781, + 0.0039726970717310905, + 0.006202982272952795, + 0.0026604626327753067, + 0.0024112167302519083, + 0.003013904904946685, + 0.0025630746968090534, + 0.0022391213569790125, + 0.0032507157884538174, + 0.0032625370658934116, + 0.0018578098388388753, + 0.003747639711946249, + 0.002609910909086466, + 0.003588216146454215, + 0.0025214701890945435, + 0.002611930714920163, + 0.0034247443545609713, + 0.002711258828639984, + 0.0037717779632657766, + 0.004044016357511282, + 0.0025784580502659082, + 0.0039827218279242516, + 0.0023917134385555983, + 0.003077529603615403, + 0.005949554964900017, + 0.003444203408434987, + 0.0074234772473573685, + 0.00865656416863203, + 0.007837562821805477, + 0.010272701270878315, + 0.009315604344010353, + 0.008786410093307495, + 0.012287789955735207, + 0.01364002376794815, + 0.016502736136317253, + 0.01565014198422432, + 0.017939520999789238, + 0.01943035237491131, + 0.021472349762916565, + 0.009927323088049889, + 0.016818787902593613, + 0.013807449489831924, + 0.03494191914796829, + 0.020988643169403076, + 0.019018495455384254, + 0.025155480951070786, + 0.02200457826256752, + 0.02886972203850746, + 0.03604273870587349, + 0.030221836641430855, + 0.04963360354304314, + 0.004942378494888544, + 0.007223706226795912, + 0.005554397124797106, + 0.007736723404377699, + 0.008072525262832642, + 0.007230117917060852, + 0.005863145925104618, + 0.006058724131435156, + 0.01933097094297409, + 0.014865376986563206, + 0.09838763624429703, + 0.018778618425130844, + 0.023423610255122185, + 0.032279472798109055, + 0.02447098307311535, + 0.08398586511611938, + 0.0073521314188838005, + 0.004964154213666916, + 0.006940280552953482, + 0.008428394794464111, + 0.005181999411433935, + 0.003995565231889486, + 0.008811438456177711, + 0.00829726830124855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.06203114241361618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful tracking shot follows a translucent jellyfish as it glides effortlessly through the abyssal depths of the deep ocean. The creature's bell, a nearly transparent, gelatinous dome, pulsates gently and rhythmically, propelling it forward with a tranquil, contemplative spirit. From beneath the bell, a string of delicate, luminous tendrils trails, each a long, ethereal ray of light, emitting a vibrant, soft bioluminescence in ever-shifting hues of electric blue and soft purple. The surrounding water is an inky, profound black, a perfect canvas, making the jellyfish the sole source of light. This inner glow creates a soft, ethereal shimmer nearby, highlighting the mesmerizing, hypnotic dance of the trailing tentacles as they swirl and undulate with the subtle currents of the water. The entire scene, captured in silent, continuous motion, creates an atmosphere of profound tranquility and otherworldliness, one of solitary beauty and quiet wonder.", + "seed": 26397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_15/index_1244_seed_26397_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03385834982618689, + "vlm_logprob": -0.03385834982618689, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03385834982618689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0056226435117423534 + ], + "opposite_logprob_tokens": [ + -5.505622863769531 + ], + "logprob_reward": -0.0056226435117423534, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0056226435117423534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06947075575590134, + 0.5710851550102234, + 0.03275427222251892, + 0.1619977504014969, + 0.02381184510886669, + 0.06868544965982437, + 0.018261652439832687, + 0.021574610844254494, + 0.010703294537961483, + 0.02165522612631321 + ], + "credit_map": [ + 0.009299946948885918, + 0.006615453399717808, + 0.01094940397888422, + 0.01027527917176485, + 0.010044807568192482, + 0.009284518659114838, + 0.006874062120914459, + 0.0061272825114429, + 0.16822399199008942, + 0.037341609597206116, + 0.24896763265132904, + 0.01661919429898262, + 0.010271013714373112, + 0.06325545907020569, + 0.013862128369510174, + 0.012544121593236923, + 0.0070674847811460495, + 0.00403154082596302, + 0.005405434872955084, + 0.00281737488694489, + 0.0026372785214334726, + 0.003570153843611479, + 0.0032534387428313494, + 0.003971566446125507, + 0.03568265587091446, + 0.03276074305176735, + 0.06715545058250427, + 0.0056172925978899, + 0.0044128973968327045, + 0.0035862973891198635, + 0.005292678251862526, + 0.007489729672670364, + 0.0037859610747545958, + 0.002527328673750162, + 0.003762106178328395, + 0.0026633793022483587, + 0.0021477830596268177, + 0.0021467984188348055, + 0.002426484599709511, + 0.004352004267275333, + 0.007821600884199142, + 0.00548494653776288, + 0.009260923601686954, + 0.008314590901136398, + 0.008929637260735035, + 0.00745813176035881, + 0.007989251986145973, + 0.01342636626213789, + 0.001951637794263661, + 0.002278795000165701, + 0.0023237469140440226, + 0.002565389731898904, + 0.0026326605584472418, + 0.0014498577220365405, + 0.0028429569210857153, + 0.0022166077978909016, + 0.0027196838054805994, + 0.0021115774288773537, + 0.0031436968129128218, + 0.0029906274285167456, + 0.0027894345112144947, + 0.0024067589547485113, + 0.003326311009004712, + 0.00208652182482183, + 0.0009393733344040811, + 0.0011591329239308834, + 0.0014428662834689021, + 0.0011876376811414957, + 0.0010125208646059036, + 0.0012315113563090563, + 0.0022307864855974913, + 0.0014994657831266522, + 0.0029498771764338017, + 0.0027422739658504725, + 0.0029554974753409624, + 0.0020824919920414686, + 0.0027337227948009968, + 0.0016876376466825604, + 0.0036260816268622875, + 0.002877644496038556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0056226435117423534 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794510900974274 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794510900974274, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794510900974274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062059950083494186, + 0.3165549337863922, + 0.019741736352443695, + 0.48245230317115784, + 0.01487414725124836, + 0.05984124913811684, + 0.010576089844107628, + 0.010073654353618622, + 0.010524658486247063, + 0.013301227241754532 + ], + "credit_map": [ + 0.007105195429176092, + 0.006518899463117123, + 0.01280344370752573, + 0.005508927628397942, + 0.006608566734939814, + 0.008362850174307823, + 0.007532298564910889, + 0.0076197669841349125, + 0.09212316572666168, + 0.02049247920513153, + 0.13693998754024506, + 0.009508187882602215, + 0.005848562810570002, + 0.03479110077023506, + 0.008839373476803303, + 0.00801206473261118, + 0.004459971096366644, + 0.0021168789826333523, + 0.004516401793807745, + 0.0017152271466329694, + 0.0014040210517123342, + 0.0017368484986945987, + 0.002015140140429139, + 0.0017772489227354527, + 0.10834411531686783, + 0.09688039869070053, + 0.20477940142154694, + 0.015306067653000355, + 0.012403768487274647, + 0.00931003037840128, + 0.014162887819111347, + 0.021265625953674316, + 0.0027289094869047403, + 0.0013704585144296288, + 0.0021241172216832638, + 0.0018978717271238565, + 0.001982838846743107, + 0.0014692704426124692, + 0.0018786214059218764, + 0.001422058790922165, + 0.007376676890999079, + 0.00516145396977663, + 0.008159360848367214, + 0.007568722125142813, + 0.007336228620260954, + 0.006000046618282795, + 0.00735055934637785, + 0.01088819932192564, + 0.0014425532426685095, + 0.0011027862783521414, + 0.001586430473253131, + 0.0018207441316917539, + 0.00108537042979151, + 0.0008613672107458115, + 0.0010339404689148068, + 0.0016428977251052856, + 0.0012641188222914934, + 0.0009434456005692482, + 0.0014735714066773653, + 0.0013955007307231426, + 0.0013517505722120404, + 0.0009878691053017974, + 0.0016099631320685148, + 0.0010474344016984105, + 0.0010105890687555075, + 0.0009554385906085372, + 0.0014279201859608293, + 0.0014065206050872803, + 0.0007142421673052013, + 0.000980221084319055, + 0.0027030818164348602, + 0.0013266452588140965, + 0.0019360230071470141, + 0.0017221214948222041, + 0.0019638382364064455, + 0.0013217348605394363, + 0.0019749777857214212, + 0.0012229151325300336, + 0.0018475418910384178, + 0.001312074949964881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.009794510900974274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062039993703365326 + ], + "opposite_logprob_tokens": [ + -2.812040090560913 + ], + "logprob_reward": -0.062039993703365326, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062039993703365326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11513889580965042, + 0.3672451376914978, + 0.05480968952178955, + 0.10852517932653427, + 0.09681367874145508, + 0.05491628497838974, + 0.07862387597560883, + 0.015856433659791946, + 0.017593028023838997, + 0.09047776460647583 + ], + "credit_map": [ + 0.015707122161984444, + 0.009770133532583714, + 0.01955501362681389, + 0.01404307596385479, + 0.01842614635825157, + 0.016800321638584137, + 0.0117033077403903, + 0.009133781306445599, + 0.10655514150857925, + 0.023232340812683105, + 0.1566084623336792, + 0.011388852261006832, + 0.007771953474730253, + 0.03891414403915405, + 0.011538483202457428, + 0.011235778219997883, + 0.010711501352488995, + 0.006390835624188185, + 0.009408323094248772, + 0.00487402081489563, + 0.00535960029810667, + 0.005366540513932705, + 0.006374744698405266, + 0.0063241315074265, + 0.010448607616126537, + 0.03105444461107254, + 0.038438580930233, + 0.005031315144151449, + 0.00431931484490633, + 0.004118724260479212, + 0.007646683603525162, + 0.007467519957572222, + 0.01762976124882698, + 0.010666726157069206, + 0.015199572779238224, + 0.009084773249924183, + 0.007234958000481129, + 0.007486816495656967, + 0.010219250805675983, + 0.0192918349057436, + 0.008014571852982044, + 0.005270624533295631, + 0.007639924995601177, + 0.005635260138660669, + 0.007578963413834572, + 0.004742101766169071, + 0.00562952971085906, + 0.010405316017568111, + 0.02159593068063259, + 0.00931069627404213, + 0.01413960661739111, + 0.007798369042575359, + 0.0056409514509141445, + 0.004246460273861885, + 0.0059229424223303795, + 0.009968925267457962, + 0.0024186414666473866, + 0.0017356316093355417, + 0.0025913286954164505, + 0.00202274601906538, + 0.0015996216097846627, + 0.0013163178227841854, + 0.002105583203956485, + 0.0020665638148784637, + 0.003406227333471179, + 0.0018682003719732165, + 0.002792359096929431, + 0.0019485889934003353, + 0.0013245699228718877, + 0.0015432171057909727, + 0.0019402162870392203, + 0.002769650425761938, + 0.015668757259845734, + 0.012309401296079159, + 0.014564228244125843, + 0.01034564059227705, + 0.009931681677699089, + 0.006377927027642727, + 0.011823633685708046, + 0.009456508792936802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.062039993703365326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062043916434049606 + ], + "opposite_logprob_tokens": [ + -2.8120439052581787 + ], + "logprob_reward": -0.062043916434049606, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062043916434049606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07600756734609604, + 0.47953304648399353, + 0.02607916295528412, + 0.16672441363334656, + 0.015761539340019226, + 0.13628773391246796, + 0.01768319122493267, + 0.016278482973575592, + 0.016569308936595917, + 0.04907551035284996 + ], + "credit_map": [ + 0.009301314130425453, + 0.006632399745285511, + 0.01477204728871584, + 0.009191632270812988, + 0.00971909984946251, + 0.01133622694760561, + 0.007960721850395203, + 0.007094122935086489, + 0.14024394750595093, + 0.03186386451125145, + 0.20936615765094757, + 0.01395122054964304, + 0.009249115362763405, + 0.05321303382515907, + 0.011719221249222755, + 0.009926497004926205, + 0.004490462597459555, + 0.002748970640823245, + 0.004309230949729681, + 0.0028847474604845047, + 0.002582998713478446, + 0.0029894006438553333, + 0.0032367082312703133, + 0.0028366416227072477, + 0.04010026901960373, + 0.018151605501770973, + 0.029151806607842445, + 0.018883591517806053, + 0.012683723121881485, + 0.009600474499166012, + 0.021768970414996147, + 0.01638396456837654, + 0.0021425168961286545, + 0.001862147357314825, + 0.0020830484572798014, + 0.0025611796882003546, + 0.0017092182533815503, + 0.0012020551366731524, + 0.001987743191421032, + 0.0022136324550956488, + 0.0161128006875515, + 0.01075184065848589, + 0.01832011342048645, + 0.015301736071705818, + 0.01694498583674431, + 0.014325149357318878, + 0.01636536791920662, + 0.028165742754936218, + 0.0021673564333468676, + 0.0029616772662848234, + 0.002445654245093465, + 0.001577814924530685, + 0.002209973055869341, + 0.001574606285430491, + 0.0019198706140741706, + 0.002826238516718149, + 0.0022150794975459576, + 0.0013344865292310715, + 0.0025083483196794987, + 0.002031635260209441, + 0.0018499164143577218, + 0.001614012522622943, + 0.002472926629707217, + 0.002252077916637063, + 0.0016422668704763055, + 0.0013642748817801476, + 0.0022543196100741625, + 0.0022679681424051523, + 0.0016072646249085665, + 0.0015665412647649646, + 0.0029213684611022472, + 0.0029453050810843706, + 0.004476082511246204, + 0.007803924847394228, + 0.00556586729362607, + 0.005955350119620562, + 0.007606801576912403, + 0.00501786544919014, + 0.007016932126134634, + 0.005632685497403145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.062043916434049606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790684580802917 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790684580802917, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790684580802917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051180969923734665, + 0.6021248698234558, + 0.019826238974928856, + 0.18514499068260193, + 0.021598149091005325, + 0.047410767525434494, + 0.03105309046804905, + 0.012909699231386185, + 0.008777650073170662, + 0.019973574206233025 + ], + "credit_map": [ + 0.006262892857193947, + 0.005272654350847006, + 0.008912647143006325, + 0.0060142953880131245, + 0.00388336181640625, + 0.008667010813951492, + 0.005038087721914053, + 0.00713002122938633, + 0.17679812014102936, + 0.03950963541865349, + 0.2621678411960602, + 0.017709461972117424, + 0.011230537667870522, + 0.06516893953084946, + 0.015370153822004795, + 0.014170198701322079, + 0.004222228191792965, + 0.0021540559828281403, + 0.005666755139827728, + 0.0017771752318367362, + 0.0011182440211996436, + 0.0012467767810449004, + 0.0017697021830826998, + 0.0018713034223765135, + 0.01720438152551651, + 0.04747909680008888, + 0.05932379886507988, + 0.01293522771447897, + 0.01029195450246334, + 0.007836643606424332, + 0.016996679827570915, + 0.013077232986688614, + 0.0027319113723933697, + 0.0024059347342699766, + 0.0031662958208471537, + 0.0028014967683702707, + 0.002011355943977833, + 0.0016697418177500367, + 0.0022349832579493523, + 0.0045764329843223095, + 0.005823106970638037, + 0.004093589261174202, + 0.0064520216546952724, + 0.005633196793496609, + 0.006507465150207281, + 0.0052252281457185745, + 0.005715042352676392, + 0.007961123250424862, + 0.004220318049192429, + 0.0044515919871628284, + 0.004511422012001276, + 0.0037515570875257254, + 0.0029185377061367035, + 0.002903522690758109, + 0.003260564524680376, + 0.005035579204559326, + 0.0018973620608448982, + 0.0010331558296456933, + 0.002446951810270548, + 0.0013729966012760997, + 0.001489030197262764, + 0.001322152791544795, + 0.0017357070464640856, + 0.0016123446403071284, + 0.0010314587270841002, + 0.001093497616238892, + 0.001037382404319942, + 0.0012103389017283916, + 0.0007027295650914311, + 0.0008623431785963476, + 0.0018048034980893135, + 0.0010350963566452265, + 0.0017276614671573043, + 0.0032295731361955404, + 0.0020769136026501656, + 0.002626651432365179, + 0.0024600389879196882, + 0.0018961571622639894, + 0.003005208680406213, + 0.002951371483504772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.029790684580802917 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful tracking shot follows a translucent jellyfish as it glides effortlessly through the abyssal depths of the deep ocean. The creature's bell, a nearly transparent, gelatinous dome, pulsates gently and rhythmically, propelling it forward with a tranquil, contemplative spirit. From beneath the bell, a string of delicate, luminous tendrils trails, each a long, ethereal ray of light, emitting a vibrant, soft bioluminescence in ever-shifting hues of electric blue and soft purple. The surrounding water is an inky, profound black, a perfect canvas, making the jellyfish the sole source of light. This inner glow creates a soft, ethereal shimmer nearby, highlighting the mesmerizing, hypnotic dance of the trailing tentacles as they swirl and undulate with the subtle currents of the water. The entire scene, captured in silent, continuous motion, creates an atmosphere of profound tranquility and otherworldliness, one of solitary beauty and quiet wonder.", + "seed": 26398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_15/index_1244_seed_26398_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10918700378388166, + "vlm_logprob": -0.10918700378388166, + "vlm_yes_no_margin": 3.849999988079071, + "vlm_token_credit": -0.10918700378388166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004697477910667658 + ], + "opposite_logprob_tokens": [ + -6.129697322845459 + ], + "logprob_reward": -0.004697477910667658, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.004697477910667658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16686326265335083, + 0.058190327137708664, + 0.05767224729061127, + 0.07102347165346146, + 0.19078902900218964, + 0.0801975280046463, + 0.08367624878883362, + 0.13139823079109192, + 0.07209393382072449, + 0.08809571713209152 + ], + "credit_map": [ + 0.020147128030657768, + 0.01772749423980713, + 0.016793187707662582, + 0.017362531274557114, + 0.019723694771528244, + 0.027909457683563232, + 0.02531176060438156, + 0.021888013929128647, + 0.005188529845327139, + 0.007656860165297985, + 0.00550249544903636, + 0.007486478425562382, + 0.006577186286449432, + 0.008093157783150673, + 0.010119503363966942, + 0.007566115818917751, + 0.005020490847527981, + 0.006241573952138424, + 0.008358122780919075, + 0.007233225740492344, + 0.005269620101898909, + 0.0061409431509673595, + 0.008311506360769272, + 0.011096763424575329, + 0.006544377189129591, + 0.008208457380533218, + 0.009231927804648876, + 0.017139801755547523, + 0.004789700265973806, + 0.007092438172549009, + 0.00892525352537632, + 0.009091510437428951, + 0.018878767266869545, + 0.024150414392352104, + 0.024378497153520584, + 0.02829478122293949, + 0.01933867484331131, + 0.022328093647956848, + 0.02439439482986927, + 0.02902538701891899, + 0.006698790937662125, + 0.007742548361420631, + 0.015916604548692703, + 0.007187427021563053, + 0.007577256765216589, + 0.010748268105089664, + 0.006354105658829212, + 0.01797252707183361, + 0.00609596585854888, + 0.006803920492529869, + 0.01977427303791046, + 0.0076755341142416, + 0.006790263578295708, + 0.00801799725741148, + 0.007490187883377075, + 0.02102811448276043, + 0.015165431424975395, + 0.00780520448461175, + 0.03932369500398636, + 0.009877912700176239, + 0.005447437521070242, + 0.008685585111379623, + 0.007978334091603756, + 0.03711462765932083, + 0.011669565923511982, + 0.004493529908359051, + 0.015638187527656555, + 0.008503087796270847, + 0.006603803485631943, + 0.00710124708712101, + 0.005557404365390539, + 0.012527108192443848, + 0.008166184648871422, + 0.009706968441605568, + 0.012323184870183468, + 0.00902020838111639, + 0.01685979589819908, + 0.01002287957817316, + 0.008289615623652935, + 0.013706882484257221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.004697477910667658 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751937326043844 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.006751937326043844, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751937326043844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11302352696657181, + 0.0717068761587143, + 0.11003345996141434, + 0.07650236040353775, + 0.14718568325042725, + 0.14806711673736572, + 0.07085200399160385, + 0.16982001066207886, + 0.03742412477731705, + 0.05538482218980789 + ], + "credit_map": [ + 0.015014326199889183, + 0.009334921836853027, + 0.011670648120343685, + 0.013391127809882164, + 0.012192747555673122, + 0.02072715573012829, + 0.012101290747523308, + 0.01859131082892418, + 0.0072434269823133945, + 0.007244788110256195, + 0.007294934708625078, + 0.0075219725258648396, + 0.009174498729407787, + 0.007846735417842865, + 0.01287816371768713, + 0.012502350844442844, + 0.00989426951855421, + 0.009127654135227203, + 0.018751034513115883, + 0.012916656211018562, + 0.008499682880938053, + 0.01997707597911358, + 0.008320986293256283, + 0.022546101361513138, + 0.00758372712880373, + 0.0071229781024158, + 0.009575224481523037, + 0.0116572929546237, + 0.006470596883445978, + 0.015529552474617958, + 0.0059947362169623375, + 0.012568247504532337, + 0.022127900272607803, + 0.013903013430535793, + 0.02086389809846878, + 0.019110674038529396, + 0.013157402165234089, + 0.017983585596084595, + 0.014039495028555393, + 0.02599971368908882, + 0.013314437121152878, + 0.014783523976802826, + 0.028900431469082832, + 0.013807191513478756, + 0.013554614968597889, + 0.018190298229455948, + 0.01800929196178913, + 0.02750733122229576, + 0.005324610974639654, + 0.003774830140173435, + 0.02065325528383255, + 0.004124799743294716, + 0.0036599852610379457, + 0.005630315747112036, + 0.004927924368530512, + 0.02275628224015236, + 0.016198931261897087, + 0.007087016478180885, + 0.05490715056657791, + 0.010604924522340298, + 0.007282037287950516, + 0.011314308270812035, + 0.009124325588345528, + 0.05330131575465202, + 0.004517340566962957, + 0.0028574224561452866, + 0.005866611376404762, + 0.006543850060552359, + 0.0033380889799445868, + 0.004827443510293961, + 0.002840154804289341, + 0.006633214186877012, + 0.005746911279857159, + 0.006905430927872658, + 0.006755649112164974, + 0.005836489610373974, + 0.010886738076806068, + 0.006763545796275139, + 0.003413126105442643, + 0.009076928719878197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.006751937326043844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052580637857317924 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.0052580637857317924, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052580637857317924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16127897799015045, + 0.05302014201879501, + 0.06287085264921188, + 0.0859404131770134, + 0.08430373668670654, + 0.10445807129144669, + 0.17092524468898773, + 0.14185959100723267, + 0.08998523652553558, + 0.04535774141550064 + ], + "credit_map": [ + 0.027942685410380363, + 0.01619860716164112, + 0.0187416709959507, + 0.01827847585082054, + 0.01663682796061039, + 0.019599223509430885, + 0.02309001423418522, + 0.020791465416550636, + 0.004791627172380686, + 0.004238183610141277, + 0.0059300740249454975, + 0.005542804021388292, + 0.00612650578841567, + 0.012810621410608292, + 0.006268492434173822, + 0.007311830297112465, + 0.007395068183541298, + 0.006133509334176779, + 0.00988724734634161, + 0.011309315450489521, + 0.005782767664641142, + 0.00847436860203743, + 0.006554732099175453, + 0.007333841640502214, + 0.010191895999014378, + 0.007409538142383099, + 0.013320917263627052, + 0.012157692573964596, + 0.009158458560705185, + 0.012090185657143593, + 0.007068798877298832, + 0.014542920514941216, + 0.008362065069377422, + 0.0098494878038764, + 0.011356810107827187, + 0.009995345026254654, + 0.00920984148979187, + 0.012860308401286602, + 0.009751969017088413, + 0.012917901389300823, + 0.008351430296897888, + 0.011551130563020706, + 0.017924588173627853, + 0.010129379108548164, + 0.009602965787053108, + 0.013933051377534866, + 0.01304519735276699, + 0.01992032863199711, + 0.012425166554749012, + 0.00980018638074398, + 0.048843175172805786, + 0.013373138383030891, + 0.009444587863981724, + 0.012200475670397282, + 0.013509589247405529, + 0.051328908652067184, + 0.015059745870530605, + 0.008077006787061691, + 0.039960701018571854, + 0.01610121689736843, + 0.006512885447591543, + 0.00809475127607584, + 0.009946390055119991, + 0.03810688480734825, + 0.009554855525493622, + 0.0076319254003465176, + 0.018916532397270203, + 0.010091221891343594, + 0.009268070571124554, + 0.012264984659850597, + 0.00703045167028904, + 0.015227190218865871, + 0.004614565055817366, + 0.0035881786607205868, + 0.006948333699256182, + 0.006644755601882935, + 0.004277805797755718, + 0.00650416687130928, + 0.004012955352663994, + 0.008766980841755867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0052580637857317924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289317727088928 + ], + "opposite_logprob_tokens": [ + -1.053931713104248 + ], + "logprob_reward": -0.4289317727088928, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4289317727088928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08353591710329056, + 0.04508571699261665, + 0.06649556010961533, + 0.08228334039449692, + 0.08389488607645035, + 0.23636996746063232, + 0.07203928381204605, + 0.1651550829410553, + 0.07461295276880264, + 0.09052733331918716 + ], + "credit_map": [ + 0.010649128817021847, + 0.00930185429751873, + 0.009987006895244122, + 0.012636998668313026, + 0.008072094060480595, + 0.009109266102313995, + 0.011845380067825317, + 0.011934181675314903, + 0.004747782368212938, + 0.004488329868763685, + 0.004325898829847574, + 0.004858960397541523, + 0.006308312527835369, + 0.007122385781258345, + 0.006393424700945616, + 0.00684062484651804, + 0.006452198605984449, + 0.0062784841284155846, + 0.008544078096747398, + 0.009689434431493282, + 0.005948559381067753, + 0.0074127740226686, + 0.009929058142006397, + 0.012240973301231861, + 0.009350820444524288, + 0.007724754046648741, + 0.011154137551784515, + 0.012258213944733143, + 0.008355224505066872, + 0.009983963333070278, + 0.010643137618899345, + 0.012813091278076172, + 0.008160985074937344, + 0.010169468820095062, + 0.01103918720036745, + 0.011786401271820068, + 0.009864678606390953, + 0.010818594135344028, + 0.009542138315737247, + 0.012513433583080769, + 0.02102683298289776, + 0.019289378076791763, + 0.05850192531943321, + 0.01689261570572853, + 0.017682109028100967, + 0.02377203479409218, + 0.019222360104322433, + 0.059982702136039734, + 0.004525204189121723, + 0.0030467098113149405, + 0.023745223879814148, + 0.004571468103677034, + 0.0031972143333405256, + 0.004126390907913446, + 0.0037298419047147036, + 0.025097226724028587, + 0.018151454627513885, + 0.007262379862368107, + 0.052380360662937164, + 0.009446161799132824, + 0.008078268729150295, + 0.010122448205947876, + 0.009645218029618263, + 0.05006880313158035, + 0.014017685316503048, + 0.005810257513076067, + 0.01308448426425457, + 0.008262009359896183, + 0.008121551014482975, + 0.007925530895590782, + 0.005177364684641361, + 0.012214066460728645, + 0.011158262379467487, + 0.010852208361029625, + 0.011992997489869595, + 0.010167793370783329, + 0.009710135869681835, + 0.01567702926695347, + 0.008222784847021103, + 0.01274612732231617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.4289317727088928 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent jellyfish with a gelatinous dome bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's trailing tendrils emitting bioluminescence in electric blue and soft purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the surrounding water depicted as inky black, making the jellyfish the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jellyfish move by pulsating its bell to glide forward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow and graceful tracking shot following the jellyfish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002957671880722 + ], + "opposite_logprob_tokens": [ + -2.3502957820892334 + ], + "logprob_reward": -0.1002957671880722, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002957671880722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06938903778791428, + 0.09281573444604874, + 0.06375285238027573, + 0.12079380452632904, + 0.26373040676116943, + 0.098342664539814, + 0.05087663605809212, + 0.1018301472067833, + 0.07291343063116074, + 0.06555532664060593 + ], + "credit_map": [ + 0.013795320875942707, + 0.005257719662040472, + 0.010553660802543163, + 0.009485018439590931, + 0.005342963617295027, + 0.0076638562604784966, + 0.007924762554466724, + 0.00936573650687933, + 0.008563307113945484, + 0.010404687374830246, + 0.008245035074651241, + 0.014921457506716251, + 0.01017814502120018, + 0.014889035373926163, + 0.013112775050103664, + 0.012501287274062634, + 0.005719667766243219, + 0.006946873385459185, + 0.008452962152659893, + 0.008371924981474876, + 0.006431115325540304, + 0.006942941807210445, + 0.009110379964113235, + 0.011776987463235855, + 0.015842342749238014, + 0.009884950704872608, + 0.0174171831458807, + 0.017485905438661575, + 0.01111020240932703, + 0.017611967399716377, + 0.010906269773840904, + 0.020534982904791832, + 0.023828452453017235, + 0.03409581631422043, + 0.03204509615898132, + 0.03809510916471481, + 0.0282362699508667, + 0.035704776644706726, + 0.031401533633470535, + 0.04032332822680473, + 0.00904884748160839, + 0.01060508657246828, + 0.013253167271614075, + 0.011112457141280174, + 0.009960288181900978, + 0.01451090071350336, + 0.014071035198867321, + 0.01578088104724884, + 0.004027651157230139, + 0.0033526818733662367, + 0.012117399834096432, + 0.005607192870229483, + 0.0034675488714128733, + 0.00480924965813756, + 0.00458540441468358, + 0.012909507378935814, + 0.008969396352767944, + 0.006450944114476442, + 0.02991933375597, + 0.008393842726945877, + 0.0047753299586474895, + 0.0075523885898292065, + 0.007241582497954369, + 0.028527334332466125, + 0.014034716412425041, + 0.004721731413155794, + 0.0164254829287529, + 0.006653464399278164, + 0.0072860922664403915, + 0.005656915251165628, + 0.004358948208391666, + 0.013776087202131748, + 0.007821079343557358, + 0.007263313513249159, + 0.008428285829722881, + 0.00810947921127081, + 0.007137499749660492, + 0.011359217576682568, + 0.005973696243017912, + 0.00946275144815445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.1002957671880722 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_tensor([4397, 4397])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_tensor([4397, 4397])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4ede36f2ee0448f2617a6adfeab4fd3e51ffbbff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_tensor([4397, 4397])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a close-up shot from a horizontal angle, the camera begins a remarkably smooth, slow 360-degree pan, its focus fixed on a mithril longsword, hovering silently in the center of the dark forge. The sword itself is a source of light, its slender, double-edged blade radiating a soft, cool, moonlit glow, gently pushing back the surrounding darkness. As the camera steadily pans, exquisite detail is captured: ancient runes, deeply etched into the metal, shimmer with a rhythmic, almost imperceptible aura of pale blue, as if flowing slowly across the blade with a lifelike quality. The camera precisely captures the cold texture of every inch of the blade, etched by countless hammering processes, and the tiny, almost imperceptible scratches, clearly visible in the mysterious light, revealing its ancient history. The background is the blurred and deep outline of a forging room. The cold anvil and the tools hanging on the wall cast long, dancing shadows under the reflection of the sword. The air is filled with the smell of ancient stone and metal. The whole scene is rendered in cold colors, creating a solemn, mysterious and tranquil atmosphere full of ancient power.", + "seed": 24354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_4397_seed_24354_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1447896372526884, + "vlm_logprob": -0.1447896372526884, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.1447896372526884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21635231375694275 + ], + "opposite_logprob_tokens": [ + -1.7163523435592651 + ], + "logprob_reward": -0.21635231375694275, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21635231375694275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02184172533452511, + 0.01604347862303257, + 0.012340408749878407, + 0.008696351200342178, + 0.0062652816995978355, + 0.02512018010020256, + 0.016555313020944595, + 0.018232382833957672, + 0.018911559134721756, + 0.8559933304786682 + ], + "credit_map": [ + 0.0025876518338918686, + 0.002531749429181218, + 0.0036663704086095095, + 0.004663561470806599, + 0.0023054645862430334, + 0.0012896825792267919, + 0.00276187714189291, + 0.0020353682339191437, + 0.0025463474448770285, + 0.0021471395157277584, + 0.002701994962990284, + 0.00268365116789937, + 0.002080631908029318, + 0.0011307153617963195, + 0.0018694467144086957, + 0.0008835516637191176, + 0.0020865928381681442, + 0.0011717560701072216, + 0.0018679070053622127, + 0.002063601277768612, + 0.0010597379878163338, + 0.001493610441684723, + 0.0012277931673452258, + 0.001369409030303359, + 0.001006917911581695, + 0.000764746917411685, + 0.0013034502044320107, + 0.0014803787926211953, + 0.000959271565079689, + 0.0008373251766897738, + 0.0011262765619903803, + 0.0012179847108200192, + 0.0009248194983229041, + 0.0004826118238270283, + 0.0009172161808237433, + 0.000804872193839401, + 0.001037811511196196, + 0.0007640810217708349, + 0.000628911831881851, + 0.000704957521520555, + 0.0045602270402014256, + 0.003846146399155259, + 0.0025109099224209785, + 0.003353940788656473, + 0.0037003550678491592, + 0.0022197116632014513, + 0.002302631037309766, + 0.0026262567844241858, + 0.002536744810640812, + 0.0019362851744517684, + 0.002688910812139511, + 0.0016385975759476423, + 0.0020096017979085445, + 0.001970060868188739, + 0.001634219428524375, + 0.002140892669558525, + 0.0027936792466789484, + 0.004196309484541416, + 0.001284359605051577, + 0.0015837730607017875, + 0.0020387759432196617, + 0.0018141688778996468, + 0.0013259360566735268, + 0.0031953807920217514, + 0.002504132455214858, + 0.0026546150911599398, + 0.0015528060030192137, + 0.0017509455792605877, + 0.003478516358882189, + 0.002852539997547865, + 0.0017273162957280874, + 0.0023906882852315903, + 0.3602197468280792, + 0.16014058887958527, + 0.01964673213660717, + 0.021915165707468987, + 0.02192159742116928, + 0.04089103266596794, + 0.020856590941548347, + 0.21040189266204834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.21635231375694275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799325793981552 + ], + "opposite_logprob_tokens": [ + -1.8049325942993164 + ], + "logprob_reward": -0.1799325793981552, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799325793981552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04207906499505043, + 0.02757430635392666, + 0.028940070420503616, + 0.031145278364419937, + 0.011880057863891125, + 0.04750055819749832, + 0.08645150810480118, + 0.029896555468440056, + 0.028024353086948395, + 0.6665082573890686 + ], + "credit_map": [ + 0.005193133372813463, + 0.0042907847091555595, + 0.006814342457801104, + 0.008201969787478447, + 0.004773333668708801, + 0.0043514687567949295, + 0.004567607305943966, + 0.0038864233065396547, + 0.00484512047842145, + 0.003813904942944646, + 0.004634760320186615, + 0.004033950157463551, + 0.003615828463807702, + 0.0023744881618767977, + 0.0030292742885649204, + 0.0012269782600924373, + 0.004355994518846273, + 0.003331235609948635, + 0.004460971336811781, + 0.0050042904913425446, + 0.0023593902587890625, + 0.003594351699575782, + 0.0028501939959824085, + 0.002983640879392624, + 0.003971830941736698, + 0.0032680928707122803, + 0.005328040570020676, + 0.004933642689138651, + 0.0031984581146389246, + 0.0035156954545527697, + 0.003319960320368409, + 0.0036095597315579653, + 0.0016930009005591273, + 0.000731755921151489, + 0.0019446202786639333, + 0.0024296422488987446, + 0.0016503341030329466, + 0.0012103932676836848, + 0.0011779104825109243, + 0.0010424013016745448, + 0.005909213330596685, + 0.0066669899970293045, + 0.004189024213701487, + 0.0048020207323133945, + 0.006789781618863344, + 0.010154147632420063, + 0.004266118165105581, + 0.004723262507468462, + 0.013516764156520367, + 0.0206422358751297, + 0.013481180183589458, + 0.010050162672996521, + 0.009915067814290524, + 0.007196242455393076, + 0.0041755023412406445, + 0.007474355865269899, + 0.004460823722183704, + 0.005987845361232758, + 0.002473728498443961, + 0.00299325049854815, + 0.0030376978684216738, + 0.003047599457204342, + 0.002031126292422414, + 0.005864483304321766, + 0.00363574898801744, + 0.003627390367910266, + 0.002462405012920499, + 0.002744139637798071, + 0.004487245809286833, + 0.004215996712446213, + 0.002952743088826537, + 0.0038986830040812492, + 0.27219995856285095, + 0.12086743116378784, + 0.017870981246232986, + 0.02067367359995842, + 0.021924162283539772, + 0.03359198197722435, + 0.018067387863993645, + 0.16131268441677094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.1799325793981552 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207850337028503 + ], + "opposite_logprob_tokens": [ + -1.5020785331726074 + ], + "logprob_reward": -0.25207850337028503, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207850337028503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06452219933271408, + 0.05574611574411392, + 0.04619713872671127, + 0.04732149466872215, + 0.042042504996061325, + 0.12977539002895355, + 0.07198161631822586, + 0.13697175681591034, + 0.03259654343128204, + 0.37284529209136963 + ], + "credit_map": [ + 0.006948980037122965, + 0.006547237746417522, + 0.01060321182012558, + 0.016257207840681076, + 0.005969224963337183, + 0.005183897912502289, + 0.007546279113739729, + 0.005466154310852289, + 0.007922819815576077, + 0.006992269307374954, + 0.009241987019777298, + 0.00926179438829422, + 0.008153107017278671, + 0.0054986849427223206, + 0.006078507285565138, + 0.002596943872049451, + 0.0067518348805606365, + 0.004103038925677538, + 0.006773064844310284, + 0.009346846491098404, + 0.004532260820269585, + 0.005706076510250568, + 0.00502016581594944, + 0.003963854163885117, + 0.0059056938625872135, + 0.00483402144163847, + 0.007252246141433716, + 0.00756662804633379, + 0.005684746894985437, + 0.0054253339767456055, + 0.006121645215898752, + 0.004531180020421743, + 0.0050092642195522785, + 0.0037622973322868347, + 0.006416114512830973, + 0.008173191919922829, + 0.0073249018751084805, + 0.005300989840179682, + 0.0035448281560093164, + 0.0025109166745096445, + 0.02008216828107834, + 0.020995955914258957, + 0.013829248957335949, + 0.01368569117039442, + 0.015926839783787727, + 0.023581916466355324, + 0.010311078280210495, + 0.011362484656274319, + 0.011759507469832897, + 0.009280300699174404, + 0.013029088266193867, + 0.007296004798263311, + 0.009104909375309944, + 0.008199097588658333, + 0.005945356097072363, + 0.007367347367107868, + 0.02280668541789055, + 0.03746947646141052, + 0.01140981912612915, + 0.012616859748959541, + 0.01613498106598854, + 0.0157260000705719, + 0.00804141629487276, + 0.012766510248184204, + 0.003829243592917919, + 0.006600308232009411, + 0.002246421528980136, + 0.004036779049783945, + 0.0032559121027588844, + 0.00307911797426641, + 0.004520043730735779, + 0.005028719548135996, + 0.1571180373430252, + 0.06761789321899414, + 0.008207340724766254, + 0.008990530855953693, + 0.008688320405781269, + 0.01762198656797409, + 0.009198213927447796, + 0.09540298581123352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.25207850337028503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054998978972435 + ], + "opposite_logprob_tokens": [ + -2.9299988746643066 + ], + "logprob_reward": -0.054998978972435, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054998978972435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09993964433670044, + 0.06307400017976761, + 0.11517056822776794, + 0.054678499698638916, + 0.026370281353592873, + 0.09198319911956787, + 0.0838940218091011, + 0.07081809639930725, + 0.0833343118429184, + 0.3107373118400574 + ], + "credit_map": [ + 0.010858112946152687, + 0.010649751871824265, + 0.01712372712790966, + 0.02528589591383934, + 0.010245485231280327, + 0.004794062580913305, + 0.012169674970209599, + 0.008812929503619671, + 0.009381867945194244, + 0.009258832782506943, + 0.010654372163116932, + 0.010361061431467533, + 0.007694154977798462, + 0.005217628087848425, + 0.006824358832091093, + 0.0036817227955907583, + 0.022979753091931343, + 0.017483286559581757, + 0.01803819090127945, + 0.020351070910692215, + 0.009453345090150833, + 0.010497739538550377, + 0.008852529339492321, + 0.007514650467783213, + 0.0086527643725276, + 0.005989766679704189, + 0.008608690463006496, + 0.009301946498453617, + 0.006464511156082153, + 0.004231617320328951, + 0.005161806475371122, + 0.0062673962675035, + 0.0036305866669863462, + 0.0014990102499723434, + 0.004628338385373354, + 0.004864350892603397, + 0.004385864827781916, + 0.002534610917791724, + 0.0029105262365192175, + 0.0019169918959960341, + 0.011858156882226467, + 0.01152268797159195, + 0.007593074813485146, + 0.009969562292098999, + 0.014425003901124, + 0.01991930790245533, + 0.008163592778146267, + 0.008531809784471989, + 0.012735254131257534, + 0.01578632742166519, + 0.018239175900816917, + 0.006501107942312956, + 0.009185591712594032, + 0.006863773800432682, + 0.005313579458743334, + 0.00926920771598816, + 0.012248694896697998, + 0.009520293213427067, + 0.005679515190422535, + 0.006248157005757093, + 0.0056085470132529736, + 0.006744626443833113, + 0.0056326137855648994, + 0.019135648384690285, + 0.01863563247025013, + 0.01165674813091755, + 0.0080892788246274, + 0.008982072584331036, + 0.009192648343741894, + 0.0084980558604002, + 0.008501826785504818, + 0.009778053499758244, + 0.10738088190555573, + 0.04981637001037598, + 0.017519166693091393, + 0.01796078495681286, + 0.0185310747474432, + 0.019449036568403244, + 0.016577521339058876, + 0.06350245326757431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.054998978972435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585810765624046 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585810765624046, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585810765624046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1628846973180771, + 0.03375673294067383, + 0.043520137667655945, + 0.03107970766723156, + 0.022495733574032784, + 0.06133053079247475, + 0.036984000355005264, + 0.07221576571464539, + 0.02184172533452511, + 0.5138909220695496 + ], + "credit_map": [ + 0.02144385129213333, + 0.01852543279528618, + 0.02435295842587948, + 0.028370792046189308, + 0.015857985243201256, + 0.01813996396958828, + 0.021211763843894005, + 0.014981946907937527, + 0.004636377561837435, + 0.0032416796311736107, + 0.006181810982525349, + 0.005208417307585478, + 0.003821491962298751, + 0.004776254296302795, + 0.00361018069088459, + 0.0022805186454206705, + 0.00792551226913929, + 0.005299592390656471, + 0.006614708341658115, + 0.00813233107328415, + 0.0031044287607073784, + 0.004819076508283615, + 0.0035636175889521837, + 0.004060869570821524, + 0.003194845747202635, + 0.0028174363542348146, + 0.004638778045773506, + 0.0043518636375665665, + 0.004085851833224297, + 0.002990331267938018, + 0.0043728966265916824, + 0.0046277036890387535, + 0.0033623231574892998, + 0.0014543121214956045, + 0.003241612110286951, + 0.0038197736721485853, + 0.003478999948129058, + 0.0031712062191218138, + 0.001862070639617741, + 0.002105435822159052, + 0.00955876987427473, + 0.00921988021582365, + 0.005640714894980192, + 0.008434112183749676, + 0.008076690137386322, + 0.005524099338799715, + 0.00852508470416069, + 0.006351179908961058, + 0.005833649076521397, + 0.00685194181278348, + 0.006034673657268286, + 0.0032704162877053022, + 0.004309982992708683, + 0.003710641758516431, + 0.002464283723384142, + 0.0045084115117788315, + 0.010479777120053768, + 0.013973681256175041, + 0.006575617473572493, + 0.008026535622775555, + 0.005935858003795147, + 0.007855666801333427, + 0.006035019177943468, + 0.013333613984286785, + 0.003930870443582535, + 0.0029371348209679127, + 0.002188122132793069, + 0.0026557163801044226, + 0.0020910610910505056, + 0.002113192342221737, + 0.0027434411458671093, + 0.0031821872107684612, + 0.21658514440059662, + 0.09461228549480438, + 0.011286615394055843, + 0.01275048591196537, + 0.012555455788969994, + 0.02406453527510166, + 0.012780396267771721, + 0.129256010055542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.020585810765624046 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot from a horizontal angle, the camera begins a remarkably smooth, slow 360-degree pan, its focus fixed on a mithril longsword, hovering silently in the center of the dark forge. The sword itself is a source of light, its slender, double-edged blade radiating a soft, cool, moonlit glow, gently pushing back the surrounding darkness. As the camera steadily pans, exquisite detail is captured: ancient runes, deeply etched into the metal, shimmer with a rhythmic, almost imperceptible aura of pale blue, as if flowing slowly across the blade with a lifelike quality. The camera precisely captures the cold texture of every inch of the blade, etched by countless hammering processes, and the tiny, almost imperceptible scratches, clearly visible in the mysterious light, revealing its ancient history. The background is the blurred and deep outline of a forging room. The cold anvil and the tools hanging on the wall cast long, dancing shadows under the reflection of the sword. The air is filled with the smell of ancient stone and metal. The whole scene is rendered in cold colors, creating a solemn, mysterious and tranquil atmosphere full of ancient power.", + "seed": 24355, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_4397_seed_24355_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12409333139657974, + "vlm_logprob": -0.12409333139657974, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.12409333139657974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12282762676477432 + ], + "opposite_logprob_tokens": [ + -2.3728275299072266 + ], + "logprob_reward": -0.12282762676477432, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12282762676477432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.309002548456192, + 0.17069917917251587, + 0.02887771464884281, + 0.0451224111020565, + 0.04290327429771423, + 0.04763071984052658, + 0.11819986253976822, + 0.055607181042432785, + 0.05610010027885437, + 0.12585695087909698 + ], + "credit_map": [ + 0.03573815897107124, + 0.03566300496459007, + 0.04704144969582558, + 0.059921812266111374, + 0.020691316574811935, + 0.035898517817258835, + 0.03506447374820709, + 0.03898382931947708, + 0.02257261984050274, + 0.028122935444116592, + 0.02601674385368824, + 0.022724580019712448, + 0.012923120521008968, + 0.01991230621933937, + 0.016065476462244987, + 0.022361423820257187, + 0.005371640436351299, + 0.004845202900469303, + 0.005264840554445982, + 0.003096655709668994, + 0.002067726803943515, + 0.0026162457652390003, + 0.002729839412495494, + 0.0028855667915195227, + 0.008122941479086876, + 0.006271599791944027, + 0.011187335476279259, + 0.0038681072182953358, + 0.002912376541644335, + 0.004517068155109882, + 0.003703195368871093, + 0.004539791960269213, + 0.007785568479448557, + 0.005804228596389294, + 0.007218909449875355, + 0.003973885904997587, + 0.0035394758451730013, + 0.0034747212193906307, + 0.006202174816280603, + 0.004904312081634998, + 0.007692172657698393, + 0.00702956086024642, + 0.006225185934454203, + 0.00449035782366991, + 0.006614708807319403, + 0.0045209345407783985, + 0.004252742510288954, + 0.0068050590343773365, + 0.029284706339240074, + 0.017642946913838387, + 0.023075805976986885, + 0.0076125795021653175, + 0.008030171506106853, + 0.007882006466388702, + 0.009077293798327446, + 0.015594367869198322, + 0.013631606474518776, + 0.00863165594637394, + 0.010647977702319622, + 0.004398449789732695, + 0.0046363151632249355, + 0.0036813141778111458, + 0.0039403908886015415, + 0.006039471365511417, + 0.015747787430882454, + 0.013603978790342808, + 0.009187329560518265, + 0.0036720791831612587, + 0.004478754475712776, + 0.002677831333130598, + 0.002093407791107893, + 0.004638936370611191, + 0.03577391430735588, + 0.01640208624303341, + 0.034209027886390686, + 0.01040008943527937, + 0.007805175613611937, + 0.006310643628239632, + 0.00519984494894743, + 0.009756170213222504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.12282762676477432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298409104347229 + ], + "opposite_logprob_tokens": [ + -2.2379841804504395 + ], + "logprob_reward": -0.11298409104347229, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298409104347229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2704286575317383, + 0.1410740166902542, + 0.04662621021270752, + 0.10074905306100845, + 0.08704999089241028, + 0.03431307151913643, + 0.0893387421965599, + 0.02658366784453392, + 0.03987174108624458, + 0.16396479308605194 + ], + "credit_map": [ + 0.03311654552817345, + 0.0359744094312191, + 0.035078227519989014, + 0.042112406343221664, + 0.020085403695702553, + 0.03173046186566353, + 0.031245799735188484, + 0.04108542948961258, + 0.016858847811818123, + 0.023330379277467728, + 0.02022409625351429, + 0.01561223715543747, + 0.01193937286734581, + 0.01585419662296772, + 0.015564690344035625, + 0.021690230816602707, + 0.009331273846328259, + 0.00920066051185131, + 0.00759862270206213, + 0.004289972595870495, + 0.003370452905073762, + 0.004117792937904596, + 0.0042954073287546635, + 0.004422036465257406, + 0.018219176679849625, + 0.013902143575251102, + 0.018560655415058136, + 0.008327273651957512, + 0.007636932656168938, + 0.00817774049937725, + 0.014005405828356743, + 0.011919740587472916, + 0.013607963919639587, + 0.014190088026225567, + 0.01842396892607212, + 0.007858348079025745, + 0.008257434703409672, + 0.007322126068174839, + 0.00899824034422636, + 0.008391834795475006, + 0.004429647698998451, + 0.00509684020653367, + 0.005077851936221123, + 0.003858931828290224, + 0.0034096469171345234, + 0.0032979168463498354, + 0.004687277600169182, + 0.004454963840544224, + 0.010867193341255188, + 0.01828654482960701, + 0.019021932035684586, + 0.007174970582127571, + 0.00771315535530448, + 0.006180854979902506, + 0.006398394238203764, + 0.013695714063942432, + 0.004265835974365473, + 0.0042459554970264435, + 0.00436356058344245, + 0.003396597458049655, + 0.002212864113971591, + 0.0026019273791462183, + 0.0017403593519702554, + 0.0037565736565738916, + 0.008077469654381275, + 0.00820882711559534, + 0.00593134481459856, + 0.004092758055776358, + 0.00333404541015625, + 0.002392180496826768, + 0.0022952065337449312, + 0.0055399187840521336, + 0.04418564215302467, + 0.02045990526676178, + 0.05471067130565643, + 0.012764147482812405, + 0.008918162435293198, + 0.009463122114539146, + 0.006799714639782906, + 0.006663450971245766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.11298409104347229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255602926015854 + ], + "opposite_logprob_tokens": [ + -1.6005603075027466 + ], + "logprob_reward": -0.2255602926015854, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255602926015854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13233831524848938, + 0.060114625841379166, + 0.08167120069265366, + 0.05766706541180611, + 0.1035631075501442, + 0.04042652249336243, + 0.10184135288000107, + 0.047210562974214554, + 0.15640787780284882, + 0.21875937283039093 + ], + "credit_map": [ + 0.01992291584610939, + 0.014323354698717594, + 0.016642296686768532, + 0.024480048567056656, + 0.011797119863331318, + 0.012813886627554893, + 0.012008643709123135, + 0.020350050181150436, + 0.010833329521119595, + 0.009464557282626629, + 0.006888759322464466, + 0.005713536404073238, + 0.0044041466899216175, + 0.0062405080534517765, + 0.00842132605612278, + 0.008148464374244213, + 0.014072810299694538, + 0.013628636486828327, + 0.013245655223727226, + 0.008675238117575645, + 0.007284705992788076, + 0.0058650728315114975, + 0.008230828680098057, + 0.010668253526091576, + 0.011043863371014595, + 0.007893840782344341, + 0.010727908462285995, + 0.004864669404923916, + 0.00536132650449872, + 0.0038955016061663628, + 0.007175812032073736, + 0.006704142317175865, + 0.015365214087069035, + 0.015548906289041042, + 0.0330863855779171, + 0.004364119842648506, + 0.00991011131554842, + 0.00767645426094532, + 0.009088939987123013, + 0.008522971533238888, + 0.008041366934776306, + 0.0057394905015826225, + 0.006602599751204252, + 0.003255582181736827, + 0.004653462674468756, + 0.0031023137271404266, + 0.004224340431392193, + 0.004807364195585251, + 0.02121732197701931, + 0.021959418430924416, + 0.022490939125418663, + 0.006617722101509571, + 0.007167478557676077, + 0.006512660998851061, + 0.005835610907524824, + 0.010040194727480412, + 0.00960465520620346, + 0.0049864319153130054, + 0.006866497453302145, + 0.0033412075135856867, + 0.0028718614485114813, + 0.0031911677215248346, + 0.004762246739119291, + 0.011586496606469154, + 0.05139363929629326, + 0.02548791840672493, + 0.029400406405329704, + 0.01216554082930088, + 0.012173008173704147, + 0.009343358688056469, + 0.006299654021859169, + 0.010144351981580257, + 0.06356596201658249, + 0.03219583258032799, + 0.06288431584835052, + 0.015529368072748184, + 0.012048429809510708, + 0.010198542848229408, + 0.008643370121717453, + 0.01369355246424675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.2255602926015854 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005638629198074 + ], + "opposite_logprob_tokens": [ + -2.695056438446045 + ], + "logprob_reward": -0.07005638629198074, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005638629198074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13061434030532837, + 0.09425047785043716, + 0.036437325179576874, + 0.05080137401819229, + 0.07148008048534393, + 0.06569363176822662, + 0.07354528456926346, + 0.07570908963680267, + 0.11760702729225159, + 0.28386127948760986 + ], + "credit_map": [ + 0.019940918311476707, + 0.017284614965319633, + 0.013220961205661297, + 0.01762755960226059, + 0.011788700707256794, + 0.019366761669516563, + 0.013340918347239494, + 0.01804392598569393, + 0.01879917085170746, + 0.01177531573921442, + 0.015961498022079468, + 0.014552824199199677, + 0.005189576651901007, + 0.008889173157513142, + 0.007786431349813938, + 0.01129649206995964, + 0.00804640632122755, + 0.005325499456375837, + 0.0049082497134804726, + 0.0043414924293756485, + 0.0022571769077330828, + 0.002664367202669382, + 0.0037763791624456644, + 0.005117759108543396, + 0.007813443429768085, + 0.00780476164072752, + 0.012669390998780727, + 0.0037939739413559437, + 0.003867625491693616, + 0.004441729746758938, + 0.005554714240133762, + 0.004855739884078503, + 0.018217774108052254, + 0.011090432293713093, + 0.012014180421829224, + 0.004147497471421957, + 0.005147451534867287, + 0.005346698220819235, + 0.007673453073948622, + 0.007842605002224445, + 0.014939671382308006, + 0.009115457534790039, + 0.011364965699613094, + 0.0038377237506210804, + 0.005654053296893835, + 0.004312105942517519, + 0.005909179337322712, + 0.010560480877757072, + 0.019046681001782417, + 0.0163153987377882, + 0.01328973937779665, + 0.0031418921425938606, + 0.005793765187263489, + 0.004155631642788649, + 0.004844221286475658, + 0.006957965902984142, + 0.01874646171927452, + 0.013727941550314426, + 0.013723735697567463, + 0.00479270750656724, + 0.006553340703248978, + 0.004619934130460024, + 0.0057261111214756966, + 0.007818861864507198, + 0.026911132037639618, + 0.029545685276389122, + 0.017734110355377197, + 0.009648825041949749, + 0.009578924626111984, + 0.006520427763462067, + 0.006485180463641882, + 0.011182746849954128, + 0.07906977832317352, + 0.03934177756309509, + 0.07579355686903, + 0.02814008668065071, + 0.015342759899795055, + 0.012975183315575123, + 0.013502212241292, + 0.019695943221449852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.07005638629198074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903826028108597 + ], + "opposite_logprob_tokens": [ + -2.464038372039795 + ], + "logprob_reward": -0.08903826028108597, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903826028108597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12065493315458298, + 0.07479348033666611, + 0.05765123292803764, + 0.05185141786932945, + 0.04600735753774643, + 0.09183017909526825, + 0.06766121089458466, + 0.21405543386936188, + 0.09801541268825531, + 0.17747926712036133 + ], + "credit_map": [ + 0.020483998581767082, + 0.016392942517995834, + 0.015064941719174385, + 0.018025167286396027, + 0.012297648005187511, + 0.012902543880045414, + 0.01061939261853695, + 0.014868306927382946, + 0.018216457217931747, + 0.013048436492681503, + 0.008443955332040787, + 0.009267265908420086, + 0.004547519609332085, + 0.004609064664691687, + 0.006206783931702375, + 0.010453999042510986, + 0.013102507218718529, + 0.011553121730685234, + 0.01007066573947668, + 0.0053357891738414764, + 0.0037388706114143133, + 0.005113244988024235, + 0.0036897414829581976, + 0.005047296639531851, + 0.008340033702552319, + 0.00765403313562274, + 0.011964777484536171, + 0.004661845043301582, + 0.004084153100848198, + 0.004334261640906334, + 0.005024970509111881, + 0.00578734464943409, + 0.010064146481454372, + 0.006525993812829256, + 0.010505201295018196, + 0.0036606513895094395, + 0.0032268234062939882, + 0.0036608953960239887, + 0.004249977879226208, + 0.004113671835511923, + 0.009238261729478836, + 0.017838677391409874, + 0.013573084957897663, + 0.012244485318660736, + 0.009346157312393188, + 0.007657770533114672, + 0.00914114248007536, + 0.012790611013770103, + 0.013705517165362835, + 0.014327255077660084, + 0.01207041461020708, + 0.004815332125872374, + 0.005553777329623699, + 0.0047025191597640514, + 0.004687073640525341, + 0.00779932364821434, + 0.0521528422832489, + 0.040819693356752396, + 0.039507534354925156, + 0.015620124526321888, + 0.01859821192920208, + 0.013939936645328999, + 0.013962069526314735, + 0.01945502683520317, + 0.03208475559949875, + 0.020311687141656876, + 0.018116112798452377, + 0.006032056175172329, + 0.006905338726937771, + 0.005615510977804661, + 0.0034528723917901516, + 0.0054970793426036835, + 0.0491299070417881, + 0.020217755809426308, + 0.06345757097005844, + 0.014489579014480114, + 0.0075289844535291195, + 0.008357848040759563, + 0.006546081975102425, + 0.007751574739813805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.08903826028108597 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot from a horizontal angle, the camera begins a remarkably smooth, slow 360-degree pan, its focus fixed on a mithril longsword, hovering silently in the center of the dark forge. The sword itself is a source of light, its slender, double-edged blade radiating a soft, cool, moonlit glow, gently pushing back the surrounding darkness. As the camera steadily pans, exquisite detail is captured: ancient runes, deeply etched into the metal, shimmer with a rhythmic, almost imperceptible aura of pale blue, as if flowing slowly across the blade with a lifelike quality. The camera precisely captures the cold texture of every inch of the blade, etched by countless hammering processes, and the tiny, almost imperceptible scratches, clearly visible in the mysterious light, revealing its ancient history. The background is the blurred and deep outline of a forging room. The cold anvil and the tools hanging on the wall cast long, dancing shadows under the reflection of the sword. The air is filled with the smell of ancient stone and metal. The whole scene is rendered in cold colors, creating a solemn, mysterious and tranquil atmosphere full of ancient power.", + "seed": 24356, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_4397_seed_24356_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11372008323669433, + "vlm_logprob": -0.11372008323669433, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.11372008323669433, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09127964079380035 + ], + "opposite_logprob_tokens": [ + -2.5912797451019287 + ], + "logprob_reward": -0.09127964079380035, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09127964079380035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1895083636045456, + 0.049611158668994904, + 0.07673798501491547, + 0.021177329123020172, + 0.08655423671007156, + 0.06972397863864899, + 0.07119912654161453, + 0.1934439092874527, + 0.21382099390029907, + 0.028222931548953056 + ], + "credit_map": [ + 0.019067231565713882, + 0.035232312977313995, + 0.01690700650215149, + 0.01782490499317646, + 0.039737578481435776, + 0.025335313752293587, + 0.018591124564409256, + 0.016812898218631744, + 0.003916791174560785, + 0.005647928919643164, + 0.003336479654535651, + 0.0059926146641373634, + 0.007598682772368193, + 0.008828197605907917, + 0.008546985685825348, + 0.005743476562201977, + 0.0032741441391408443, + 0.005155179649591446, + 0.0048103006556630135, + 0.005960715003311634, + 0.02797119878232479, + 0.017774073407053947, + 0.005999132990837097, + 0.005793244577944279, + 0.0012411943171173334, + 0.001489285845309496, + 0.002390395849943161, + 0.002161787124350667, + 0.0024684180971235037, + 0.00491879228502512, + 0.0026026975829154253, + 0.0039047584868967533, + 0.004028283525258303, + 0.005471188575029373, + 0.005313804838806391, + 0.005965169053524733, + 0.007219988852739334, + 0.011956732720136642, + 0.014232192188501358, + 0.03236686810851097, + 0.006436995230615139, + 0.006177250295877457, + 0.00878294836729765, + 0.008736107498407364, + 0.008553748019039631, + 0.007575367111712694, + 0.013093621470034122, + 0.010367943905293941, + 0.007128614000976086, + 0.007397323381155729, + 0.007570417132228613, + 0.004784178454428911, + 0.008158148266375065, + 0.015020512044429779, + 0.00522850314155221, + 0.01591142825782299, + 0.013269045390188694, + 0.008355235680937767, + 0.07783951610326767, + 0.017457885667681694, + 0.008068862371146679, + 0.007818372920155525, + 0.01331247203052044, + 0.04732251167297363, + 0.016222907230257988, + 0.01638278365135193, + 0.07257181406021118, + 0.03006504476070404, + 0.013362095691263676, + 0.022116512060165405, + 0.021045999601483345, + 0.022053822875022888, + 0.0024755278136581182, + 0.00423234049230814, + 0.004345416557043791, + 0.004573299549520016, + 0.004109004512429237, + 0.002547303680330515, + 0.002442540368065238, + 0.003497498109936714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.09127964079380035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034547746181488 + ], + "opposite_logprob_tokens": [ + -2.3503453731536865 + ], + "logprob_reward": -0.10034547746181488, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034547746181488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16399449110031128, + 0.04181573912501335, + 0.1616857349872589, + 0.041905276477336884, + 0.06278806179761887, + 0.06282033026218414, + 0.05516405776143074, + 0.17268741130828857, + 0.22073104977607727, + 0.01640789955854416 + ], + "credit_map": [ + 0.014833171851933002, + 0.025204848498106003, + 0.012218054383993149, + 0.010674367658793926, + 0.042132530361413956, + 0.01971626840531826, + 0.02275637909770012, + 0.01645885966718197, + 0.00461795786395669, + 0.005158096086233854, + 0.0030409761238843203, + 0.004166209604591131, + 0.005097117740660906, + 0.006612739525735378, + 0.005633522756397724, + 0.007489118259400129, + 0.005562735255807638, + 0.005387218203395605, + 0.009357075206935406, + 0.014785373583436012, + 0.010924756526947021, + 0.050516046583652496, + 0.008234820328652859, + 0.056917715817689896, + 0.002739870920777321, + 0.0036336176563054323, + 0.007949710823595524, + 0.006541142240166664, + 0.0044930120930075645, + 0.004604710731655359, + 0.004738557618111372, + 0.007204651832580566, + 0.0033827568404376507, + 0.0067159999161958694, + 0.007897504605352879, + 0.00979059562087059, + 0.006893000565469265, + 0.0065107448026537895, + 0.010757348500192165, + 0.01084010861814022, + 0.008495100773870945, + 0.006216864101588726, + 0.007745381910353899, + 0.006691712886095047, + 0.009322099387645721, + 0.006727848667651415, + 0.00988721288740635, + 0.0077341035939753056, + 0.00460446160286665, + 0.007102231029421091, + 0.007871701382100582, + 0.004852453246712685, + 0.008430452086031437, + 0.009682620875537395, + 0.004050160758197308, + 0.008569976314902306, + 0.009950933046638966, + 0.004522818140685558, + 0.07068463414907455, + 0.013362687081098557, + 0.006377454847097397, + 0.007289068773388863, + 0.011930443346500397, + 0.048569366335868835, + 0.01686164177954197, + 0.018739119172096252, + 0.07243135571479797, + 0.03162387013435364, + 0.014757632277905941, + 0.02394050732254982, + 0.020600324496626854, + 0.02177661471068859, + 0.001378490007482469, + 0.002006505150347948, + 0.002891938667744398, + 0.0020591262727975845, + 0.0018680618377402425, + 0.0024654760491102934, + 0.0014189532957971096, + 0.00231934804469347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.10034547746181488 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703388929367065 + ], + "opposite_logprob_tokens": [ + -2.1270339488983154 + ], + "logprob_reward": -0.12703388929367065, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703388929367065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27265116572380066, + 0.07145391404628754, + 0.17895807325839996, + 0.061873406171798706, + 0.10345710068941116, + 0.06698092073202133, + 0.037775151431560516, + 0.09137315303087234, + 0.07545973360538483, + 0.04001741483807564 + ], + "credit_map": [ + 0.017550718039274216, + 0.019665110856294632, + 0.023973817005753517, + 0.023989049717783928, + 0.05341384932398796, + 0.03639591857790947, + 0.06791280955076218, + 0.029749920591711998, + 0.006421111989766359, + 0.006963346153497696, + 0.005093859974294901, + 0.00792687851935625, + 0.008696123026311398, + 0.012405808083713055, + 0.012409677729010582, + 0.011537105776369572, + 0.006687965709716082, + 0.01126379705965519, + 0.010134914889931679, + 0.023525863885879517, + 0.044924236834049225, + 0.042345039546489716, + 0.011555921286344528, + 0.028520341962575912, + 0.004365607164800167, + 0.004666078835725784, + 0.004794722888618708, + 0.009502877481281757, + 0.006943867541849613, + 0.013624576851725578, + 0.007594012655317783, + 0.010381662286818027, + 0.00504532502964139, + 0.0062434496358036995, + 0.006731946486979723, + 0.009480591863393784, + 0.007064808625727892, + 0.012494921684265137, + 0.019237762317061424, + 0.0371582917869091, + 0.006306029390543699, + 0.009044270031154156, + 0.01028345711529255, + 0.006666752975434065, + 0.009876764379441738, + 0.010077329352498055, + 0.006804622709751129, + 0.00792169664055109, + 0.003036840120330453, + 0.005005214363336563, + 0.006089908070862293, + 0.004841932561248541, + 0.003628774080425501, + 0.0048983413726091385, + 0.003626040183007717, + 0.0066480995155870914, + 0.004956559278070927, + 0.003550770226866007, + 0.0378710962831974, + 0.005276848562061787, + 0.003047107020393014, + 0.0042733484879136086, + 0.005290694069117308, + 0.02710673026740551, + 0.006796388421207666, + 0.006705195177346468, + 0.022274702787399292, + 0.008611568249762058, + 0.006240868475288153, + 0.007444383576512337, + 0.010148555040359497, + 0.007238068617880344, + 0.003894433844834566, + 0.00594038050621748, + 0.006827418692409992, + 0.005256103351712227, + 0.005667990539222956, + 0.0036185518838465214, + 0.004112653434276581, + 0.004699880722910166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.12703388929367065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005871832370758 + ], + "opposite_logprob_tokens": [ + -2.695058822631836 + ], + "logprob_reward": -0.07005871832370758, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005871832370758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20417918264865875, + 0.09811651706695557, + 0.06714505702257156, + 0.03040502779185772, + 0.10820217430591583, + 0.12159647792577744, + 0.08439407497644424, + 0.14489306509494781, + 0.09769587218761444, + 0.04337258264422417 + ], + "credit_map": [ + 0.02229403518140316, + 0.028643270954489708, + 0.01863962970674038, + 0.020315177738666534, + 0.042485203593969345, + 0.026206962764263153, + 0.02594268135726452, + 0.019652212038636208, + 0.006621900945901871, + 0.010715519078075886, + 0.005621365271508694, + 0.010308542288839817, + 0.01786336489021778, + 0.020645326003432274, + 0.012503407895565033, + 0.01383709255605936, + 0.0030666522216051817, + 0.0041855378076434135, + 0.003374943044036627, + 0.004904760047793388, + 0.015826648101210594, + 0.009680337272584438, + 0.005819166079163551, + 0.020287008956074715, + 0.001874056993983686, + 0.0019615718629211187, + 0.002882099710404873, + 0.002641719300299883, + 0.004035370424389839, + 0.009026703424751759, + 0.0031658431980758905, + 0.0048176622949540615, + 0.005362689960747957, + 0.011848165653645992, + 0.00777150085195899, + 0.011311567388474941, + 0.013461454771459103, + 0.013824515976011753, + 0.01855023391544819, + 0.02607204020023346, + 0.00987330637872219, + 0.014096793718636036, + 0.0154519472271204, + 0.01400244515389204, + 0.018583502620458603, + 0.016996176913380623, + 0.017707563936710358, + 0.014884736388921738, + 0.009179450571537018, + 0.009874400682747364, + 0.012418153695762157, + 0.009535886347293854, + 0.012940558604896069, + 0.008100694976747036, + 0.006167793180793524, + 0.016177136451005936, + 0.019743209704756737, + 0.023576537147164345, + 0.02286401391029358, + 0.020669862627983093, + 0.019497262313961983, + 0.011398316361010075, + 0.010538832284510136, + 0.016605036333203316, + 0.00869094580411911, + 0.009435852989554405, + 0.03192918747663498, + 0.012966644018888474, + 0.0069635347463190556, + 0.00919661670923233, + 0.008943363092839718, + 0.009569726884365082, + 0.0038317018188536167, + 0.008207844570279121, + 0.0073922467418015, + 0.0062425751239061356, + 0.00670950673520565, + 0.003659741720184684, + 0.0027729091234505177, + 0.004556053783744574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.07005871832370758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798826903104782 + ], + "opposite_logprob_tokens": [ + -1.8048826456069946 + ], + "logprob_reward": -0.1798826903104782, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798826903104782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5077527165412903, + 0.031056294217705727, + 0.15956701338291168, + 0.04828263074159622, + 0.07278982549905777, + 0.031031431630253792, + 0.046011991798877716, + 0.06572035700082779, + 0.025841865688562393, + 0.011945881880819798 + ], + "credit_map": [ + 0.032867807894945145, + 0.03353644162416458, + 0.04213618114590645, + 0.04268793761730194, + 0.11744901537895203, + 0.06427516043186188, + 0.11731775104999542, + 0.05748240649700165, + 0.002321261679753661, + 0.0031294231303036213, + 0.002726882230490446, + 0.0029868886340409517, + 0.004451408516615629, + 0.006547663826495409, + 0.0037984398659318686, + 0.005094327498227358, + 0.005550919566303492, + 0.005438092164695263, + 0.006482354365289211, + 0.022855743765830994, + 0.021274879574775696, + 0.044901106506586075, + 0.007574576418846846, + 0.04548933357000351, + 0.003702188841998577, + 0.0021008355543017387, + 0.00845723319798708, + 0.0025467548985034227, + 0.004926791414618492, + 0.01698247343301773, + 0.003232142888009548, + 0.006334207020699978, + 0.0035746130160987377, + 0.006367350462824106, + 0.005322613753378391, + 0.005290163215249777, + 0.007234626915305853, + 0.009658577851951122, + 0.011504504829645157, + 0.0238373763859272, + 0.002090209163725376, + 0.00199231319129467, + 0.003954135812819004, + 0.004007968585938215, + 0.004722693003714085, + 0.006101417355239391, + 0.0038282456807792187, + 0.004334450699388981, + 0.00371390744112432, + 0.00641015637665987, + 0.007026900537312031, + 0.005494203418493271, + 0.003956078551709652, + 0.005393897648900747, + 0.004862730856984854, + 0.009154117666184902, + 0.003041887655854225, + 0.0020327300298959017, + 0.028212329372763634, + 0.004027160350233316, + 0.001937673194333911, + 0.0022925985977053642, + 0.0040028998628258705, + 0.020173076540231705, + 0.002698408206924796, + 0.00301354075782001, + 0.007389482576400042, + 0.003269004402682185, + 0.0017539963591843843, + 0.0027460786513984203, + 0.0022157819475978613, + 0.002755574882030487, + 0.0012537508737295866, + 0.001820790464989841, + 0.001912430627271533, + 0.0015653299633413553, + 0.0016882646596059203, + 0.0011015828931704164, + 0.0008907734882086515, + 0.001712958561256528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.1798826903104782 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot from a horizontal angle, the camera begins a remarkably smooth, slow 360-degree pan, its focus fixed on a mithril longsword, hovering silently in the center of the dark forge. The sword itself is a source of light, its slender, double-edged blade radiating a soft, cool, moonlit glow, gently pushing back the surrounding darkness. As the camera steadily pans, exquisite detail is captured: ancient runes, deeply etched into the metal, shimmer with a rhythmic, almost imperceptible aura of pale blue, as if flowing slowly across the blade with a lifelike quality. The camera precisely captures the cold texture of every inch of the blade, etched by countless hammering processes, and the tiny, almost imperceptible scratches, clearly visible in the mysterious light, revealing its ancient history. The background is the blurred and deep outline of a forging room. The cold anvil and the tools hanging on the wall cast long, dancing shadows under the reflection of the sword. The air is filled with the smell of ancient stone and metal. The whole scene is rendered in cold colors, creating a solemn, mysterious and tranquil atmosphere full of ancient power.", + "seed": 24357, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_2/index_4397_seed_24357_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07020130828022957, + "vlm_logprob": -0.07020130828022957, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.07020130828022957, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10188256204128265 + ], + "opposite_logprob_tokens": [ + -2.6018824577331543 + ], + "logprob_reward": -0.10188256204128265, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.10188256204128265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12849190831184387, + 0.07479456812143326, + 0.057333484292030334, + 0.06384644657373428, + 0.09731456637382507, + 0.2258511483669281, + 0.12624764442443848, + 0.16069811582565308, + 0.029687903821468353, + 0.03573423996567726 + ], + "credit_map": [ + 0.014465409331023693, + 0.023359645158052444, + 0.01955106481909752, + 0.01231515035033226, + 0.011401400901377201, + 0.011567408218979836, + 0.017740709707140923, + 0.018091121688485146, + 0.01200762391090393, + 0.011049880646169186, + 0.011056059040129185, + 0.007132039871066809, + 0.008618100546300411, + 0.008697012439370155, + 0.007947183214128017, + 0.00828667264431715, + 0.007638776209205389, + 0.007789261639118195, + 0.00823177769780159, + 0.0057786088436841965, + 0.0048094079829752445, + 0.009865517728030682, + 0.005906115751713514, + 0.007314017508178949, + 0.008640533313155174, + 0.008430028334259987, + 0.007376238703727722, + 0.005944808479398489, + 0.006645527668297291, + 0.009524528868496418, + 0.009552265517413616, + 0.007732513826340437, + 0.014104369096457958, + 0.015342513099312782, + 0.014163662679493427, + 0.00845508836209774, + 0.011746348813176155, + 0.014009405858814716, + 0.007933679968118668, + 0.011559497565031052, + 0.014483100734651089, + 0.02684756927192211, + 0.01481181662529707, + 0.015407407656311989, + 0.02465542033314705, + 0.05908473953604698, + 0.02096167765557766, + 0.049599409103393555, + 0.013575760647654533, + 0.014340559020638466, + 0.017425168305635452, + 0.013804705813527107, + 0.02450641244649887, + 0.020571239292621613, + 0.010323313064873219, + 0.01170048676431179, + 0.03340943530201912, + 0.03524191305041313, + 0.021467532962560654, + 0.01330491155385971, + 0.01968870311975479, + 0.026289643719792366, + 0.0041056531481444836, + 0.007190321106463671, + 0.006649804767221212, + 0.003759383223950863, + 0.003998601343482733, + 0.004124258644878864, + 0.0034080934710800648, + 0.00475702341645956, + 0.0012436069082468748, + 0.0017471342580392957, + 0.0031532973516732454, + 0.0016033070860430598, + 0.008520259521901608, + 0.002999916672706604, + 0.007958374917507172, + 0.007508453447371721, + 0.0015413436340168118, + 0.002449288032948971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.10188256204128265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297877132892609 + ], + "opposite_logprob_tokens": [ + -2.23797869682312 + ], + "logprob_reward": -0.11297877132892609, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297877132892609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10410892218351364, + 0.0697421208024025, + 0.07880157977342606, + 0.10164609551429749, + 0.1615447700023651, + 0.11518926173448563, + 0.16026663780212402, + 0.1305663138628006, + 0.052501190453767776, + 0.025633184239268303 + ], + "credit_map": [ + 0.011586847715079784, + 0.017496120184659958, + 0.0179996769875288, + 0.009796098805963993, + 0.0071982149966061115, + 0.00781896710395813, + 0.013837069272994995, + 0.01837591454386711, + 0.01079662423580885, + 0.015019218437373638, + 0.008164509199559689, + 0.005293778609484434, + 0.008556130342185497, + 0.007112679071724415, + 0.00920949224382639, + 0.005589678883552551, + 0.009258837439119816, + 0.010538987815380096, + 0.008476663380861282, + 0.007425026036798954, + 0.008729280903935432, + 0.013880765996873379, + 0.01048796996474266, + 0.01000403705984354, + 0.009544006548821926, + 0.013006011955440044, + 0.017188139259815216, + 0.009759576991200447, + 0.0092486422508955, + 0.014450492337346077, + 0.015744199976325035, + 0.012705015949904919, + 0.02129342406988144, + 0.030803339555859566, + 0.030263882130384445, + 0.021253552287817, + 0.012900011613965034, + 0.02185564488172531, + 0.00938482116907835, + 0.01379009336233139, + 0.010454157367348671, + 0.01826988160610199, + 0.00822767335921526, + 0.010456463322043419, + 0.005322809796780348, + 0.02981795184314251, + 0.007224229630082846, + 0.025416085496544838, + 0.01890452951192856, + 0.01969585195183754, + 0.016067437827587128, + 0.016002004966139793, + 0.0345199853181839, + 0.02836889773607254, + 0.014653949998319149, + 0.012053980492055416, + 0.025599436834454536, + 0.023464323952794075, + 0.018331944942474365, + 0.012776071205735207, + 0.013483639806509018, + 0.01958651654422283, + 0.0038189785555005074, + 0.013505382463335991, + 0.006409178487956524, + 0.005440601613372564, + 0.009629473090171814, + 0.006395624950528145, + 0.008732126094400883, + 0.008207231760025024, + 0.0031107619870454073, + 0.004576181992888451, + 0.0029484594706445932, + 0.0015670036664232612, + 0.006150151137262583, + 0.00282908882945776, + 0.0046972609125077724, + 0.004485870711505413, + 0.0010589637095108628, + 0.0018963824259117246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.11297877132892609 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054958585649728775 + ], + "opposite_logprob_tokens": [ + -2.9299585819244385 + ], + "logprob_reward": -0.054958585649728775, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054958585649728775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19733481109142303, + 0.08348826318979263, + 0.06292794644832611, + 0.09946165978908539, + 0.09505841881036758, + 0.2249637097120285, + 0.06274279952049255, + 0.09009592235088348, + 0.04802051931619644, + 0.035905949771404266 + ], + "credit_map": [ + 0.02100621908903122, + 0.034309569746255875, + 0.03735346347093582, + 0.015898339450359344, + 0.016205379739403725, + 0.019407330080866814, + 0.02743961103260517, + 0.02571490779519081, + 0.022978896275162697, + 0.0166190005838871, + 0.009368699043989182, + 0.006738102529197931, + 0.007850252091884613, + 0.0063658105209469795, + 0.007680394221097231, + 0.005887110251933336, + 0.007300870958715677, + 0.010994607582688332, + 0.007270467933267355, + 0.005420047324150801, + 0.007919302210211754, + 0.008194651454687119, + 0.007728853262960911, + 0.008099152706563473, + 0.012783042155206203, + 0.011315785348415375, + 0.011628788895905018, + 0.011893745511770248, + 0.010977616533637047, + 0.011821504682302475, + 0.01591736078262329, + 0.013123818673193455, + 0.01113904919475317, + 0.01360254269093275, + 0.013044735416769981, + 0.012179461307823658, + 0.010040130466222763, + 0.012686766684055328, + 0.012080764397978783, + 0.010284967720508575, + 0.02693687193095684, + 0.03592362627387047, + 0.035676851868629456, + 0.038674939423799515, + 0.018510377034544945, + 0.027639416977763176, + 0.017891572788357735, + 0.02371005341410637, + 0.008065900765359402, + 0.008860655128955841, + 0.010539150796830654, + 0.006011191755533218, + 0.010077672079205513, + 0.009896758943796158, + 0.004806316923350096, + 0.004485154990106821, + 0.018776144832372665, + 0.017675569280982018, + 0.01302207913249731, + 0.008770286105573177, + 0.009105315431952477, + 0.01380923017859459, + 0.00297494325786829, + 0.0059623573906719685, + 0.008025886490941048, + 0.004296593368053436, + 0.009473301470279694, + 0.006206158082932234, + 0.006179745774716139, + 0.007781425956636667, + 0.0016534337773919106, + 0.0044039771892130375, + 0.0032673252280801535, + 0.0016267028404399753, + 0.009617111645638943, + 0.003635719884186983, + 0.008087961003184319, + 0.006642041262239218, + 0.0012777383672073483, + 0.0017513505881652236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.054958585649728775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308886453509331 + ], + "opposite_logprob_tokens": [ + -3.168088912963867 + ], + "logprob_reward": -0.04308886453509331, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308886453509331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12707383930683136, + 0.08263776451349258, + 0.17186132073402405, + 0.055631641298532486, + 0.09826208651065826, + 0.15248754620552063, + 0.049806661903858185, + 0.19460748136043549, + 0.050899356603622437, + 0.01673230715095997 + ], + "credit_map": [ + 0.013850885443389416, + 0.02229257859289646, + 0.016814814880490303, + 0.011929435655474663, + 0.013379447162151337, + 0.015168191865086555, + 0.017028095200657845, + 0.016610389575362206, + 0.014573686756193638, + 0.013922998681664467, + 0.009643377736210823, + 0.006363958120346069, + 0.009453845210373402, + 0.008464143611490726, + 0.010642708279192448, + 0.009573056362569332, + 0.02914363145828247, + 0.0202411450445652, + 0.01912682130932808, + 0.013631465844810009, + 0.017326626926660538, + 0.02458202838897705, + 0.021107276901602745, + 0.026702335104346275, + 0.007268694229424, + 0.00936933420598507, + 0.01044033095240593, + 0.0040101828053593636, + 0.004401451908051968, + 0.008085468783974648, + 0.007051271852105856, + 0.005004907492548227, + 0.012685089372098446, + 0.014347100630402565, + 0.014982870779931545, + 0.010029849596321583, + 0.012898026034235954, + 0.012730092741549015, + 0.009950648061931133, + 0.01063841488212347, + 0.011206929571926594, + 0.024876054376363754, + 0.01378706656396389, + 0.017489921301603317, + 0.015533030033111572, + 0.031439341604709625, + 0.011274401098489761, + 0.026880798861384392, + 0.007328701671212912, + 0.005008299369364977, + 0.009307699277997017, + 0.0053133657202124596, + 0.007461298257112503, + 0.008011884056031704, + 0.0026361607015132904, + 0.00473925331607461, + 0.042163800448179245, + 0.04415496066212654, + 0.026373395696282387, + 0.01502495538443327, + 0.023125890642404556, + 0.031316667795181274, + 0.004158514551818371, + 0.008289281278848648, + 0.009397194720804691, + 0.005827384535223246, + 0.01092152576893568, + 0.0059659043326973915, + 0.005694281775504351, + 0.00749457161873579, + 0.002127215266227722, + 0.0034712757915258408, + 0.0012656438630074263, + 0.0010528169805184007, + 0.004382376093417406, + 0.0019169618608430028, + 0.0025666949804872274, + 0.0030292163137346506, + 0.000774138665292412, + 0.0017444585682824254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.04308886453509331 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mithril longsword with a slender, double-edged blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sword hovering silently in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a smooth, slow 360-degree pan around the sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are ancient runes deeply etched into the blade, shimmering with a pale blue aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred and deep outline of a forging room with a cold anvil and hanging tools?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809775784611702 + ], + "opposite_logprob_tokens": [ + -3.288097858428955 + ], + "logprob_reward": -0.03809775784611702, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809775784611702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20900112390518188, + 0.06303741782903671, + 0.07163175195455551, + 0.12080604583024979, + 0.19720879197120667, + 0.08601929247379303, + 0.05863621085882187, + 0.14316579699516296, + 0.034165844321250916, + 0.01632768101990223 + ], + "credit_map": [ + 0.023275066167116165, + 0.042374175041913986, + 0.025211505591869354, + 0.01591397263109684, + 0.017519647255539894, + 0.02010085992515087, + 0.027068890631198883, + 0.03753703460097313, + 0.009327325038611889, + 0.011918100528419018, + 0.008103812113404274, + 0.005282022524625063, + 0.006191297899931669, + 0.006291716825217009, + 0.009114736691117287, + 0.006808415055274963, + 0.012080100364983082, + 0.009926775470376015, + 0.008049912750720978, + 0.007283092476427555, + 0.006985404994338751, + 0.009547640569508076, + 0.009313201531767845, + 0.00844563078135252, + 0.014822177588939667, + 0.01850399374961853, + 0.02650780975818634, + 0.00925848726183176, + 0.009166019037365913, + 0.01576501503586769, + 0.01615695096552372, + 0.010625608265399933, + 0.032303016632795334, + 0.03757552057504654, + 0.0238140057772398, + 0.019373180344700813, + 0.01652226597070694, + 0.031169425696134567, + 0.01622815430164337, + 0.020223233848810196, + 0.008777105249464512, + 0.011486312374472618, + 0.008448069915175438, + 0.008895074017345905, + 0.007560414727777243, + 0.017661189660429955, + 0.007783223874866962, + 0.015407906845211983, + 0.006668996065855026, + 0.010420995764434338, + 0.008259542286396027, + 0.006153188645839691, + 0.0077752587385475636, + 0.00929975975304842, + 0.004618822131305933, + 0.005439654923975468, + 0.03273170441389084, + 0.03201215714216232, + 0.018344324082136154, + 0.010601602494716644, + 0.017736753448843956, + 0.022669188678264618, + 0.003099946305155754, + 0.0059701260179281235, + 0.00591622106730938, + 0.0023929686285555363, + 0.006413704715669155, + 0.0032181066926568747, + 0.005099998787045479, + 0.0058899130672216415, + 0.001450894633308053, + 0.003784039756283164, + 0.0013412958942353725, + 0.0009951948886737227, + 0.004425466060638428, + 0.0016316173132508993, + 0.002417602576315403, + 0.0031522721983492374, + 0.0007862086058594286, + 0.0015780245885252953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.03809775784611702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_tensor([3176, 3176])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_tensor([3176, 3176])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ad119c9183276756ed8c75e6690add85dbc9d462 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_tensor([3176, 3176])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro still shot, captured from an almost reverent, horizontal perspective, captures a dramatic scene of light and shadow, steeped in the sensibility of Old Master paintings. In the image, a slender finger, clad in a deep black velvet glove, its tip stained with a few large, irregularly shaped grains of shimmering sea salt, delicately touches a thinly sliced ham, as crimson as the finest ruby. A single, warm, focused light source from the side creates a precise, intense contrast between light and dark (chiaroscuro), making the fine pile of the velvet appear even darker in the shadows, while the illuminated side radiates a soft, restrained luster. As the fingers gently and steadily rubbed to one side, the sea salt particles left a shiny trail on the moist surface of the ham. The rich white fat texture between the ham's texture was instantly illuminated by the light, reflecting a rich and attractive luster like liquid gold, forming a sharp contrast with the deep red lean meat part. The whole picture, in tranquility and concentration, created an extremely luxurious and ritualistic culinary aesthetic atmosphere.", + "seed": 13265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_3176_seed_13265_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012317304033786058, + "vlm_logprob": -0.012317304033786058, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.012317304033786058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01341930776834488 + ], + "opposite_logprob_tokens": [ + -4.638419151306152 + ], + "logprob_reward": -0.01341930776834488, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01341930776834488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14571313560009003, + 0.09689392149448395, + 0.0985436886548996, + 0.062143631279468536, + 0.04934241622686386, + 0.03734440356492996, + 0.1273827850818634, + 0.10229651629924774, + 0.2146744579076767, + 0.06566500663757324 + ], + "credit_map": [ + 0.01791343279182911, + 0.023013904690742493, + 0.01733260229229927, + 0.013960675336420536, + 0.021067427471280098, + 0.018422964960336685, + 0.01888684742152691, + 0.015115280635654926, + 0.009721299633383751, + 0.009712550789117813, + 0.01923506148159504, + 0.00859169103205204, + 0.016316130757331848, + 0.016239017248153687, + 0.009539132937788963, + 0.007539036218076944, + 0.01342022605240345, + 0.013789521530270576, + 0.013878910802304745, + 0.009094789624214172, + 0.015084115788340569, + 0.008224638178944588, + 0.01061434019356966, + 0.014437150210142136, + 0.007345723453909159, + 0.008852778933942318, + 0.010360888205468655, + 0.005252622999250889, + 0.010114407166838646, + 0.005359770730137825, + 0.00886122789233923, + 0.0059962146915495396, + 0.007223581895232201, + 0.005789545830339193, + 0.006827837321907282, + 0.005168679635971785, + 0.007961702533066273, + 0.006821252405643463, + 0.0036545246839523315, + 0.005895293317735195, + 0.005455302074551582, + 0.005347828846424818, + 0.005000811070203781, + 0.0038427344989031553, + 0.0051006656140089035, + 0.004606548231095076, + 0.004312227480113506, + 0.0036782866809517145, + 0.010571942664682865, + 0.023408828303217888, + 0.017543975263834, + 0.007843011058866978, + 0.03205551952123642, + 0.01139934454113245, + 0.01180323213338852, + 0.012756936252117157, + 0.015444571152329445, + 0.01476590521633625, + 0.018931962549686432, + 0.010033637285232544, + 0.013498067855834961, + 0.007483914494514465, + 0.009961897507309914, + 0.012176552787423134, + 0.029148999601602554, + 0.028560258448123932, + 0.03577933460474014, + 0.017635442316532135, + 0.02291644737124443, + 0.01857498101890087, + 0.024331089109182358, + 0.03772791847586632, + 0.008560637943446636, + 0.008743424899876118, + 0.007477995008230209, + 0.011886391788721085, + 0.004963739775121212, + 0.007065732032060623, + 0.00630927924066782, + 0.010657809674739838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.01341930776834488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554306536912918 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012554306536912918, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554306536912918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11460880190134048, + 0.09005319327116013, + 0.06723194569349289, + 0.026431050151586533, + 0.05798871070146561, + 0.049815546721220016, + 0.11805277317762375, + 0.048085570335388184, + 0.33403480052948, + 0.09369764477014542 + ], + "credit_map": [ + 0.0141218276694417, + 0.018381627276539803, + 0.015329615212976933, + 0.013693024404346943, + 0.016067948192358017, + 0.011256344616413116, + 0.014229848980903625, + 0.011528565548360348, + 0.012135087512433529, + 0.008002875372767448, + 0.01761987991631031, + 0.011716114357113838, + 0.013715601526200771, + 0.0068008434027433395, + 0.008366684429347515, + 0.011696100234985352, + 0.010051030665636063, + 0.006666230503469706, + 0.012805217877030373, + 0.007327346596866846, + 0.010341744869947433, + 0.004901393782347441, + 0.005845075007528067, + 0.009293906390666962, + 0.0031743000727146864, + 0.0036417320370674133, + 0.004125426057726145, + 0.0027724679093807936, + 0.0039055808447301388, + 0.003439837135374546, + 0.0024342010729014874, + 0.0029375050216913223, + 0.008026497438549995, + 0.010326298885047436, + 0.0062083108350634575, + 0.0052821519784629345, + 0.009952995926141739, + 0.007929440587759018, + 0.004554872866719961, + 0.005708144977688789, + 0.007389924954622984, + 0.008160091005265713, + 0.005895812064409256, + 0.004937984514981508, + 0.007407959084957838, + 0.005604095291346312, + 0.0054228720255196095, + 0.004996808711439371, + 0.01020994782447815, + 0.03304407745599747, + 0.008946290239691734, + 0.00877168495208025, + 0.03158116713166237, + 0.008007030002772808, + 0.009269329719245434, + 0.008223241195082664, + 0.006124706473201513, + 0.00636456161737442, + 0.010046055540442467, + 0.00458298297598958, + 0.0058721923269331455, + 0.004792709369212389, + 0.0044765775091946125, + 0.005825783591717482, + 0.046725932508707047, + 0.04075736179947853, + 0.06330448389053345, + 0.03765687346458435, + 0.02738790214061737, + 0.02669796161353588, + 0.03694430738687515, + 0.05455997586250305, + 0.01059316098690033, + 0.012771048583090305, + 0.0147246690467, + 0.02627761848270893, + 0.005531054921448231, + 0.00833970122039318, + 0.007509912829846144, + 0.007950480096042156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.012554306536912918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490753773599863 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.002490753773599863, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490753773599863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13782218098640442, + 0.23614348471164703, + 0.1282447725534439, + 0.04722285270690918, + 0.04130376875400543, + 0.03644035756587982, + 0.045673541724681854, + 0.11361449956893921, + 0.11574241518974304, + 0.09779217839241028 + ], + "credit_map": [ + 0.016509782522916794, + 0.015193340368568897, + 0.027857419103384018, + 0.014696724712848663, + 0.018664633855223656, + 0.01568767800927162, + 0.015029175207018852, + 0.014183420687913895, + 0.029105113819241524, + 0.027354301884770393, + 0.04633242264389992, + 0.027052205055952072, + 0.035996872931718826, + 0.025098424404859543, + 0.022619042545557022, + 0.022585077211260796, + 0.015236414968967438, + 0.016739612445235252, + 0.02766934223473072, + 0.017235025763511658, + 0.01943245343863964, + 0.008773989975452423, + 0.011378808878362179, + 0.011779120191931725, + 0.005979279987514019, + 0.006937957368791103, + 0.00923681166023016, + 0.00592651916667819, + 0.00763827795162797, + 0.004139302764087915, + 0.003898728871718049, + 0.0034659705124795437, + 0.005920535884797573, + 0.005347518250346184, + 0.004838070832192898, + 0.0035926157142966986, + 0.007629276718944311, + 0.0061679743230342865, + 0.0031507874373346567, + 0.0046569835394620895, + 0.004472809843719006, + 0.0035823637153953314, + 0.004942216910421848, + 0.0030834809876978397, + 0.004913905635476112, + 0.004632896743714809, + 0.0051460848189890385, + 0.00566659402102232, + 0.005554335657507181, + 0.006614716723561287, + 0.006767678540199995, + 0.003816454205662012, + 0.008074019104242325, + 0.003336577210575342, + 0.005529231391847134, + 0.0059805260971188545, + 0.014121739193797112, + 0.015077916905283928, + 0.022456685081124306, + 0.012153181247413158, + 0.013718921691179276, + 0.008680050261318684, + 0.01138276420533657, + 0.01602323167026043, + 0.014415194280445576, + 0.014177125878632069, + 0.029339110478758812, + 0.01790633425116539, + 0.011310279369354248, + 0.009689055383205414, + 0.008094999007880688, + 0.010810311883687973, + 0.011998976580798626, + 0.011795984581112862, + 0.010976055637001991, + 0.01525317132472992, + 0.007891258224844933, + 0.011625662446022034, + 0.010509969666600227, + 0.017741087824106216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.002490753773599863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058008871972561 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.02058008871972561, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058008871972561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16362643241882324, + 0.05111304298043251, + 0.05111284181475639, + 0.025507397949695587, + 0.023553522303700447, + 0.02548561431467533, + 0.05778764560818672, + 0.0487140491604805, + 0.4647238552570343, + 0.08837561309337616 + ], + "credit_map": [ + 0.02192533388733864, + 0.01849028468132019, + 0.011272457428276539, + 0.007996789179742336, + 0.029918065294623375, + 0.04265810176730156, + 0.012449458241462708, + 0.018915930762887, + 0.00627489946782589, + 0.007907147519290447, + 0.0060757859610021114, + 0.004812897648662329, + 0.007842383347451687, + 0.005804672371596098, + 0.007364185526967049, + 0.005031071603298187, + 0.008530831895768642, + 0.005532361567020416, + 0.008180666714906693, + 0.004772478714585304, + 0.008230164647102356, + 0.005014863796532154, + 0.004561112727969885, + 0.006290361285209656, + 0.0025929890107363462, + 0.00329766352660954, + 0.0033619541209191084, + 0.0018031841609627008, + 0.004208937752991915, + 0.002251947531476617, + 0.005474373698234558, + 0.0025163472164422274, + 0.003718297928571701, + 0.0030786553397774696, + 0.002949304645881057, + 0.002093695104122162, + 0.003762258216738701, + 0.003644084557890892, + 0.0017873645992949605, + 0.0025198603980243206, + 0.003482165979221463, + 0.0038800151087343693, + 0.0025735856033861637, + 0.00274087255820632, + 0.0037914817221462727, + 0.0038898109924048185, + 0.0022172683384269476, + 0.0029104147106409073, + 0.0067539578303694725, + 0.008031818084418774, + 0.00927191786468029, + 0.005926490295678377, + 0.010154570452868938, + 0.005052602384239435, + 0.004951666574925184, + 0.007644619792699814, + 0.005216533783823252, + 0.005941614508628845, + 0.008725358173251152, + 0.004053239710628986, + 0.006104252766817808, + 0.005157582927495241, + 0.005357579328119755, + 0.008157889358699322, + 0.05178727209568024, + 0.055335428565740585, + 0.08922293782234192, + 0.07984255999326706, + 0.05796363204717636, + 0.04155053570866585, + 0.03674614056944847, + 0.052275363355875015, + 0.014181166887283325, + 0.011624753475189209, + 0.01762772537767887, + 0.012087346985936165, + 0.007369583938270807, + 0.006919523701071739, + 0.0074816555716097355, + 0.011083859018981457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.02058008871972561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542063370347023 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542063370347023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542063370347023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11233329772949219, + 0.08826015889644623, + 0.028926735743880272, + 0.026803363114595413, + 0.04278731718659401, + 0.03855568915605545, + 0.035004422068595886, + 0.041687704622745514, + 0.1295686662197113, + 0.456072598695755 + ], + "credit_map": [ + 0.011518798768520355, + 0.010489322245121002, + 0.013937488198280334, + 0.014546196907758713, + 0.015188256278634071, + 0.014286922290921211, + 0.018232246860861778, + 0.01413408201187849, + 0.010424882173538208, + 0.010744565166532993, + 0.012794949114322662, + 0.008896139450371265, + 0.013523368164896965, + 0.009839615784585476, + 0.012442799285054207, + 0.009593845345079899, + 0.0036057617980986834, + 0.00323180528357625, + 0.0039835404604673386, + 0.0035101391840726137, + 0.004012184217572212, + 0.0033989844378083944, + 0.0029820185154676437, + 0.0042023067362606525, + 0.0026619783602654934, + 0.0035446693655103445, + 0.0038142597768455744, + 0.0029011345468461514, + 0.004201818723231554, + 0.003157487604767084, + 0.003933860454708338, + 0.0025881563778966665, + 0.006370482500642538, + 0.004445549566298723, + 0.0050109210424125195, + 0.0035201087594032288, + 0.006533187814056873, + 0.008123406209051609, + 0.0031986276153475046, + 0.00558503670617938, + 0.00470737786963582, + 0.004476705566048622, + 0.00367445545271039, + 0.00393208721652627, + 0.005312619265168905, + 0.006330902222543955, + 0.005023897625505924, + 0.005097647197544575, + 0.0036722372751682997, + 0.006526615936309099, + 0.004852441139519215, + 0.0033496455289423466, + 0.006628174334764481, + 0.0034319607075303793, + 0.0032795520965009928, + 0.0032637978438287973, + 0.005397321190685034, + 0.005497243721038103, + 0.006228373385965824, + 0.0032755311112850904, + 0.005090424790978432, + 0.003447243943810463, + 0.00558282108977437, + 0.007168748416006565, + 0.015192274935543537, + 0.016117408871650696, + 0.019699035212397575, + 0.023441994562745094, + 0.01672009378671646, + 0.011944383382797241, + 0.011286385357379913, + 0.015167108736932278, + 0.06490936875343323, + 0.05875740945339203, + 0.06201908364892006, + 0.063597671687603, + 0.03327149525284767, + 0.052252478897571564, + 0.046982839703559875, + 0.07428231090307236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.012542063370347023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro still shot, captured from an almost reverent, horizontal perspective, captures a dramatic scene of light and shadow, steeped in the sensibility of Old Master paintings. In the image, a slender finger, clad in a deep black velvet glove, its tip stained with a few large, irregularly shaped grains of shimmering sea salt, delicately touches a thinly sliced ham, as crimson as the finest ruby. A single, warm, focused light source from the side creates a precise, intense contrast between light and dark (chiaroscuro), making the fine pile of the velvet appear even darker in the shadows, while the illuminated side radiates a soft, restrained luster. As the fingers gently and steadily rubbed to one side, the sea salt particles left a shiny trail on the moist surface of the ham. The rich white fat texture between the ham's texture was instantly illuminated by the light, reflecting a rich and attractive luster like liquid gold, forming a sharp contrast with the deep red lean meat part. The whole picture, in tranquility and concentration, created an extremely luxurious and ritualistic culinary aesthetic atmosphere.", + "seed": 13266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_3176_seed_13266_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014556532818824052, + "vlm_logprob": -0.014556532818824052, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.014556532818824052, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016797764226794243 + ], + "opposite_logprob_tokens": [ + -4.1417975425720215 + ], + "logprob_reward": -0.016797764226794243, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016797764226794243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23657162487506866, + 0.06969892233610153, + 0.10445674508810043, + 0.12403672188520432, + 0.0606522411108017, + 0.03281528875231743, + 0.1749601513147354, + 0.03517452999949455, + 0.09303481876850128, + 0.06859901547431946 + ], + "credit_map": [ + 0.02510753460228443, + 0.04300998896360397, + 0.030606284737586975, + 0.03009583242237568, + 0.03388286009430885, + 0.024144457653164864, + 0.019645055755972862, + 0.030079586431384087, + 0.007851329632103443, + 0.014080420136451721, + 0.008629572577774525, + 0.006028966978192329, + 0.010026573203504086, + 0.005127230193465948, + 0.007317230571061373, + 0.010637594386935234, + 0.009942491538822651, + 0.025475801900029182, + 0.009622881188988686, + 0.00863166805356741, + 0.023328186944127083, + 0.007575851399451494, + 0.0078063467517495155, + 0.01207350380718708, + 0.020279884338378906, + 0.022478505969047546, + 0.025119418278336525, + 0.008859187364578247, + 0.01649656891822815, + 0.012695652432739735, + 0.008221949450671673, + 0.009885543957352638, + 0.016142994165420532, + 0.010340566746890545, + 0.006629747338593006, + 0.005511534865945578, + 0.004565479699522257, + 0.0053492216393351555, + 0.005244583357125521, + 0.006868104450404644, + 0.004136330913752317, + 0.0038012112490832806, + 0.008371897973120213, + 0.0031177387572824955, + 0.0033823878038674593, + 0.002950608730316162, + 0.0027773394249379635, + 0.004277769476175308, + 0.05336971953511238, + 0.02454753778874874, + 0.007109728641808033, + 0.0050984700210392475, + 0.010220831260085106, + 0.056916914880275726, + 0.009447973221540451, + 0.008248955942690372, + 0.00878186896443367, + 0.006323544774204493, + 0.006232939660549164, + 0.003132800804451108, + 0.002567273797467351, + 0.0029335361905395985, + 0.0022248204331845045, + 0.0029777404852211475, + 0.011193303391337395, + 0.014047477394342422, + 0.019492803141474724, + 0.012675994075834751, + 0.012205928564071655, + 0.006260204594582319, + 0.007879795506596565, + 0.009279304184019566, + 0.006916118785738945, + 0.012374305166304111, + 0.006222052499651909, + 0.013189788907766342, + 0.010007650591433048, + 0.005532259587198496, + 0.008309612050652504, + 0.006047226022928953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.016797764226794243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644304238259792 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644304238259792, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644304238259792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09376587718725204, + 0.10151543468236923, + 0.38730165362358093, + 0.05727101489901543, + 0.07361374795436859, + 0.01869182102382183, + 0.1638396978378296, + 0.020190546289086342, + 0.03816860914230347, + 0.04564155265688896 + ], + "credit_map": [ + 0.009657276794314384, + 0.016670340672135353, + 0.013848294503986835, + 0.01450483500957489, + 0.012719819322228432, + 0.010527048259973526, + 0.006290779449045658, + 0.009547482244670391, + 0.011066565290093422, + 0.015131687745451927, + 0.014432581141591072, + 0.008323865942656994, + 0.014364904724061489, + 0.013554399833083153, + 0.011047564446926117, + 0.013593871146440506, + 0.03161388635635376, + 0.07367930561304092, + 0.04652715101838112, + 0.03628244996070862, + 0.05527367815375328, + 0.04677393659949303, + 0.04168034344911575, + 0.05547090992331505, + 0.009223820641636848, + 0.009817101992666721, + 0.013497089967131615, + 0.004423275589942932, + 0.0050514391623437405, + 0.004457481671124697, + 0.00498151546344161, + 0.005819292273372412, + 0.014107681810855865, + 0.010398107580840588, + 0.011955111287534237, + 0.005412114784121513, + 0.006602908950299025, + 0.011012960225343704, + 0.006203304976224899, + 0.007921560667455196, + 0.002854346763342619, + 0.003005909500643611, + 0.0026998892426490784, + 0.001449142349883914, + 0.0020998537074774504, + 0.0021289854776114225, + 0.0019005048088729382, + 0.002553186845034361, + 0.0499337874352932, + 0.022961292415857315, + 0.005919191054999828, + 0.005337330978363752, + 0.009895135648548603, + 0.05249231308698654, + 0.009116173721849918, + 0.008184479549527168, + 0.003192776581272483, + 0.002643825486302376, + 0.003318679053336382, + 0.002251558005809784, + 0.0026374615263193846, + 0.003019016934558749, + 0.0014127979520708323, + 0.0017144321464002132, + 0.00574384955689311, + 0.005798432044684887, + 0.005696961190551519, + 0.004189011640846729, + 0.0049374718219041824, + 0.0038201373536139727, + 0.0036409865133464336, + 0.004341758321970701, + 0.004005377646535635, + 0.009418795816600323, + 0.004014467354863882, + 0.010509289801120758, + 0.007181683089584112, + 0.003655377309769392, + 0.004185980651527643, + 0.0026705816853791475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.008644304238259792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816402819007635 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816402819007635, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816402819007635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13752973079681396, + 0.050342921167612076, + 0.18779818713665009, + 0.11035934090614319, + 0.05400853231549263, + 0.09578805416822433, + 0.22369612753391266, + 0.03282509371638298, + 0.07099488377571106, + 0.03665710985660553 + ], + "credit_map": [ + 0.012149578891694546, + 0.02585063874721527, + 0.017904385924339294, + 0.016270991414785385, + 0.021277979016304016, + 0.0167371965944767, + 0.012032466009259224, + 0.01530649233609438, + 0.005198631435632706, + 0.008958510123193264, + 0.005256493575870991, + 0.004204259254038334, + 0.0072409966960549355, + 0.0068533215671777725, + 0.004751317203044891, + 0.007879388518631458, + 0.016686294227838516, + 0.02388431318104267, + 0.021501172333955765, + 0.016408998519182205, + 0.02906680479645729, + 0.025223419070243835, + 0.023674415424466133, + 0.03135277330875397, + 0.01789606176316738, + 0.014633036218583584, + 0.016934694722294807, + 0.010059045627713203, + 0.010314228013157845, + 0.0135505311191082, + 0.011645715683698654, + 0.015326030552387238, + 0.015129090286791325, + 0.009251139126718044, + 0.008152427151799202, + 0.0034049402456730604, + 0.004231376573443413, + 0.005142514128237963, + 0.003931052051484585, + 0.004765992518514395, + 0.026772242039442062, + 0.010427397675812244, + 0.004569647833704948, + 0.004002187866717577, + 0.0061743976548314095, + 0.033076051622629166, + 0.005672670900821686, + 0.005093451123684645, + 0.06754644215106964, + 0.03089074045419693, + 0.007883548736572266, + 0.008683224208652973, + 0.013749316334724426, + 0.07103817909955978, + 0.012525578029453754, + 0.011379092000424862, + 0.0056897415779531, + 0.005196726880967617, + 0.00830068439245224, + 0.00396542577072978, + 0.0022105607204139233, + 0.00233157305046916, + 0.002575463382527232, + 0.002554920269176364, + 0.007018137723207474, + 0.010370996780693531, + 0.0119174774736166, + 0.010357673279941082, + 0.008914720267057419, + 0.005617913790047169, + 0.008763744495809078, + 0.008034217171370983, + 0.003049031598493457, + 0.006334210280328989, + 0.003172690747305751, + 0.007455866783857346, + 0.005373999010771513, + 0.0032271207310259342, + 0.004479989875108004, + 0.003564203390851617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.002816402819007635 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633899636566639 + ], + "opposite_logprob_tokens": [ + -3.651339054107666 + ], + "logprob_reward": -0.02633899636566639, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633899636566639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13831844925880432, + 0.10237067937850952, + 0.17218410968780518, + 0.06449411064386368, + 0.093391552567482, + 0.0592125728726387, + 0.18807025253772736, + 0.06345907598733902, + 0.08056077361106873, + 0.03793847933411598 + ], + "credit_map": [ + 0.01687643676996231, + 0.023996109142899513, + 0.01820852980017662, + 0.02129911445081234, + 0.020099151879549026, + 0.014536816626787186, + 0.011010956019163132, + 0.012291338294744492, + 0.01204511895775795, + 0.017503393813967705, + 0.013427803292870522, + 0.00905653741210699, + 0.013471651822328568, + 0.0072149611078202724, + 0.01072994340211153, + 0.018921270966529846, + 0.01593693159520626, + 0.022107724100351334, + 0.030884115025401115, + 0.017709463834762573, + 0.02029760181903839, + 0.014710416086018085, + 0.02074403315782547, + 0.02979382500052452, + 0.01296447217464447, + 0.00884921383112669, + 0.008474357426166534, + 0.0045143733732402325, + 0.005207246169447899, + 0.006830343045294285, + 0.00919280108064413, + 0.008461298421025276, + 0.020829981192946434, + 0.014199440367519855, + 0.009928284212946892, + 0.009111879393458366, + 0.009229233488440514, + 0.010192064568400383, + 0.00864007044583559, + 0.011260596103966236, + 0.011350136250257492, + 0.00937949400395155, + 0.004871565382927656, + 0.004465503618121147, + 0.006382918916642666, + 0.011273789219558239, + 0.005033047404140234, + 0.006456112023442984, + 0.05691301450133324, + 0.026432933285832405, + 0.006986214779317379, + 0.006180647294968367, + 0.011170101352036, + 0.06040465831756592, + 0.010535374283790588, + 0.009447317570447922, + 0.01318405196070671, + 0.009906978346407413, + 0.012111196294426918, + 0.006583104841411114, + 0.00390858668833971, + 0.004450824111700058, + 0.00732788722962141, + 0.005986441392451525, + 0.005642082542181015, + 0.012528911232948303, + 0.01623416692018509, + 0.011324029415845871, + 0.01102382317185402, + 0.007115377578884363, + 0.00863413792103529, + 0.008058244362473488, + 0.003284861333668232, + 0.005379568785429001, + 0.0034969481639564037, + 0.007146647199988365, + 0.005886898376047611, + 0.0033664731308817863, + 0.004827647935599089, + 0.004549437202513218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.02633899636566639 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185196444392204 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185196444392204, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185196444392204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23526667058467865, + 0.06394986808300018, + 0.12550492584705353, + 0.03525037690997124, + 0.1846146285533905, + 0.10822286456823349, + 0.08656676113605499, + 0.08207665383815765, + 0.04853089526295662, + 0.030016329139471054 + ], + "credit_map": [ + 0.029014183208346367, + 0.046055231243371964, + 0.030577106401324272, + 0.02658979780972004, + 0.034860026091337204, + 0.01959940232336521, + 0.01957816630601883, + 0.02899276092648506, + 0.006487261038273573, + 0.008925475180149078, + 0.007938231341540813, + 0.006157482974231243, + 0.008691132068634033, + 0.007120785769075155, + 0.005905091296881437, + 0.012724405154585838, + 0.009472317062318325, + 0.027096249163150787, + 0.01337609812617302, + 0.007688196375966072, + 0.03728044033050537, + 0.012816209346055984, + 0.008380088955163956, + 0.009395306929945946, + 0.006893450394272804, + 0.005295796785503626, + 0.006547877099364996, + 0.003058903617784381, + 0.003306616796180606, + 0.0026884921826422215, + 0.0027188521344214678, + 0.004740385804325342, + 0.04031405970454216, + 0.02943243272602558, + 0.03512364253401756, + 0.012566500343382359, + 0.015533753670752048, + 0.023483946919441223, + 0.012009821832180023, + 0.01615045964717865, + 0.030055901035666466, + 0.01241085771471262, + 0.004919086582958698, + 0.0045931097120046616, + 0.007678655441850424, + 0.03474920988082886, + 0.007239312864840031, + 0.0065767355263233185, + 0.023721924051642418, + 0.011838110163807869, + 0.00532859330996871, + 0.004517829976975918, + 0.005371610634028912, + 0.023750972002744675, + 0.00626806914806366, + 0.005769649054855108, + 0.01464633084833622, + 0.0131950993090868, + 0.011725139804184437, + 0.007429070770740509, + 0.008597515523433685, + 0.007688914891332388, + 0.006553891580551863, + 0.012240692041814327, + 0.004992351867258549, + 0.008485191501677036, + 0.008317661471664906, + 0.0068026259541511536, + 0.006953501608222723, + 0.004067079164087772, + 0.004958541132509708, + 0.003953941166400909, + 0.003268470289185643, + 0.004488835111260414, + 0.0026135104708373547, + 0.0049883718602359295, + 0.004256160464137793, + 0.00353492540307343, + 0.003671162063255906, + 0.00319489068351686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.018185196444392204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro still shot, captured from an almost reverent, horizontal perspective, captures a dramatic scene of light and shadow, steeped in the sensibility of Old Master paintings. In the image, a slender finger, clad in a deep black velvet glove, its tip stained with a few large, irregularly shaped grains of shimmering sea salt, delicately touches a thinly sliced ham, as crimson as the finest ruby. A single, warm, focused light source from the side creates a precise, intense contrast between light and dark (chiaroscuro), making the fine pile of the velvet appear even darker in the shadows, while the illuminated side radiates a soft, restrained luster. As the fingers gently and steadily rubbed to one side, the sea salt particles left a shiny trail on the moist surface of the ham. The rich white fat texture between the ham's texture was instantly illuminated by the light, reflecting a rich and attractive luster like liquid gold, forming a sharp contrast with the deep red lean meat part. The whole picture, in tranquility and concentration, created an extremely luxurious and ritualistic culinary aesthetic atmosphere.", + "seed": 13267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_3176_seed_13267_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011155515210703015, + "vlm_logprob": -0.011155515210703015, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.011155515210703015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005441852379590273 + ], + "opposite_logprob_tokens": [ + -5.505441665649414 + ], + "logprob_reward": -0.005441852379590273, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005441852379590273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17341655492782593, + 0.20472504198551178, + 0.03805556148290634, + 0.07111141085624695, + 0.20566008985042572, + 0.06436769664287567, + 0.07140038162469864, + 0.09615025669336319, + 0.03832527622580528, + 0.036787718534469604 + ], + "credit_map": [ + 0.01843375898897648, + 0.018082797527313232, + 0.032235391438007355, + 0.019702831283211708, + 0.02423742599785328, + 0.020553164184093475, + 0.02320851758122444, + 0.016962673515081406, + 0.06787177175283432, + 0.04353627562522888, + 0.012367730028927326, + 0.00895641464740038, + 0.014265510253608227, + 0.034497637301683426, + 0.007894203066825867, + 0.01533550675958395, + 0.009311230853199959, + 0.007871665060520172, + 0.0051107765175402164, + 0.0038526749704033136, + 0.004353491123765707, + 0.0023257650900632143, + 0.0024219737388193607, + 0.002807981101796031, + 0.008652220480144024, + 0.008094323799014091, + 0.0204329751431942, + 0.007190369535237551, + 0.009900190867483616, + 0.0068639544770121574, + 0.005410916171967983, + 0.0045664627104997635, + 0.017254818230867386, + 0.02611553855240345, + 0.024536579847335815, + 0.014327283017337322, + 0.044153232127428055, + 0.03440002724528313, + 0.032982319593429565, + 0.011890292167663574, + 0.00558639969676733, + 0.006720494478940964, + 0.0068795764818787575, + 0.00689903786405921, + 0.011373492889106274, + 0.011011862196028233, + 0.009274267591536045, + 0.006622568238526583, + 0.007726349402219057, + 0.00968415942043066, + 0.007205563131719828, + 0.007055593188852072, + 0.010570609010756016, + 0.010974983684718609, + 0.009628449566662312, + 0.008554675616323948, + 0.01199619472026825, + 0.012429879046976566, + 0.008914951235055923, + 0.007177747320383787, + 0.013811301440000534, + 0.013108249753713608, + 0.014540965668857098, + 0.01417096983641386, + 0.004172178916633129, + 0.005277827847748995, + 0.004309572279453278, + 0.004415605682879686, + 0.0059112124145030975, + 0.005000915378332138, + 0.0040440973825752735, + 0.0051938691176474094, + 0.004872930236160755, + 0.005713419057428837, + 0.0051692151464521885, + 0.0029947201255708933, + 0.006895276717841625, + 0.005859640426933765, + 0.002950301393866539, + 0.0023322117049247026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.005441852379590273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003622637828812003 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.003622637828812003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003622637828812003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1595812737941742, + 0.03673918917775154, + 0.02900553122162819, + 0.024415457621216774, + 0.5664474368095398, + 0.028700275346636772, + 0.02281280606985092, + 0.02469615265727043, + 0.026501432061195374, + 0.08110042661428452 + ], + "credit_map": [ + 0.016433821991086006, + 0.017288489267230034, + 0.016058262437582016, + 0.016963627189397812, + 0.023800646886229515, + 0.02217891626060009, + 0.022896554321050644, + 0.02396094612777233, + 0.0063545252196490765, + 0.006198400165885687, + 0.0047594341449439526, + 0.0036495511885732412, + 0.0052136993035674095, + 0.003729234216734767, + 0.003273662878200412, + 0.0035606827586889267, + 0.004459903575479984, + 0.004868445452302694, + 0.0068301428109407425, + 0.003556878538802266, + 0.002746511483564973, + 0.0021156384609639645, + 0.00213811406865716, + 0.0022898977622389793, + 0.0035484579857438803, + 0.003270508022978902, + 0.0027302992530167103, + 0.0017805204261094332, + 0.005415745545178652, + 0.003499985672533512, + 0.0021535775158554316, + 0.002016362501308322, + 0.051846928894519806, + 0.06477780640125275, + 0.0572914220392704, + 0.040782198309898376, + 0.12534576654434204, + 0.09614058583974838, + 0.09636399894952774, + 0.0338987298309803, + 0.003094626124948263, + 0.0030462276190519333, + 0.005728777963668108, + 0.0022133784368634224, + 0.0039678155444562435, + 0.004642692860215902, + 0.0029260125011205673, + 0.0030807459261268377, + 0.0023128206375986338, + 0.0038448057603091, + 0.0019387907814234495, + 0.002399296034127474, + 0.0034855110570788383, + 0.0034607676789164543, + 0.0022034908179193735, + 0.0031673226039856672, + 0.0019394331611692905, + 0.00294130458496511, + 0.0018260418437421322, + 0.0028804319445043802, + 0.003411286510527134, + 0.003394857980310917, + 0.0040880595333874226, + 0.004214736633002758, + 0.0022572227753698826, + 0.003454412566497922, + 0.0017850880976766348, + 0.0018510115332901478, + 0.005863251630216837, + 0.004008325282484293, + 0.00291480403393507, + 0.004367315210402012, + 0.010001753456890583, + 0.016484424471855164, + 0.010405070148408413, + 0.009559912607073784, + 0.010777724906802177, + 0.010547783225774765, + 0.00795710738748312, + 0.005366647150367498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.003622637828812003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940036192536354 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.001940036192536354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940036192536354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07573976367712021, + 0.0406714491546154, + 0.13513246178627014, + 0.03884866088628769, + 0.49316221475601196, + 0.03968347981572151, + 0.03694886341691017, + 0.025839481502771378, + 0.014981529675424099, + 0.09899204969406128 + ], + "credit_map": [ + 0.010148261673748493, + 0.010257788933813572, + 0.012711365707218647, + 0.010530545376241207, + 0.008584090508520603, + 0.00837032776325941, + 0.007348184939473867, + 0.0077891964465379715, + 0.008741054683923721, + 0.005344257690012455, + 0.004743380006402731, + 0.0037074906285852194, + 0.004652405623346567, + 0.007146864198148251, + 0.0024684674572199583, + 0.0038675316609442234, + 0.01944347657263279, + 0.026969674974679947, + 0.03233206644654274, + 0.018374929204583168, + 0.012512064538896084, + 0.00681311683729291, + 0.00911407545208931, + 0.009573055431246758, + 0.005192258860915899, + 0.004848779644817114, + 0.012762564234435558, + 0.0038817913737148046, + 0.0033363699913024902, + 0.0035594264045357704, + 0.0028840783052146435, + 0.0023833916056901217, + 0.03966657817363739, + 0.04981033876538277, + 0.044650059193372726, + 0.03485298529267311, + 0.1094963550567627, + 0.08761174976825714, + 0.09610328823328018, + 0.0309708621352911, + 0.0033004400320351124, + 0.003878877265378833, + 0.0069182198494672775, + 0.004799753427505493, + 0.004694630391895771, + 0.0055872718803584576, + 0.005183686967939138, + 0.005320602096617222, + 0.0038028843700885773, + 0.005151964258402586, + 0.00331232906319201, + 0.004847727715969086, + 0.005443209782242775, + 0.0049056787975132465, + 0.0037671762984246016, + 0.005717894993722439, + 0.0021464473102241755, + 0.0027238414622843266, + 0.002135787857696414, + 0.003415097948163748, + 0.0034528232645243406, + 0.00363363791257143, + 0.00380544806830585, + 0.004526397213339806, + 0.001557689975015819, + 0.0027442651335150003, + 0.001154711819253862, + 0.0015095563139766455, + 0.0023910398595035076, + 0.0022259806282818317, + 0.00143674504943192, + 0.0019615411292761564, + 0.015795590355992317, + 0.006630266085267067, + 0.022133024409413338, + 0.010975254699587822, + 0.011152823455631733, + 0.011231525801122189, + 0.013068925589323044, + 0.008004635572433472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.001940036192536354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033694181591272354 + ], + "opposite_logprob_tokens": [ + -3.408694267272949 + ], + "logprob_reward": -0.033694181591272354, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033694181591272354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07182873040437698, + 0.08900249749422073, + 0.04802298545837402, + 0.03531131520867348, + 0.29705560207366943, + 0.057369109243154526, + 0.10976268351078033, + 0.14861960709095, + 0.05175771191716194, + 0.09126970916986465 + ], + "credit_map": [ + 0.008060451596975327, + 0.007028745952993631, + 0.016457481309771538, + 0.006877063307911158, + 0.011232188902795315, + 0.0070289503782987595, + 0.008015761151909828, + 0.007128097116947174, + 0.027584480121731758, + 0.017687665298581123, + 0.0074756755493581295, + 0.004513312131166458, + 0.006107477471232414, + 0.01407465897500515, + 0.004258016124367714, + 0.007301210425794125, + 0.012180834077298641, + 0.010388628579676151, + 0.008155275136232376, + 0.004677384626120329, + 0.004211554769426584, + 0.0023582205176353455, + 0.0027052476070821285, + 0.0033458415418863297, + 0.00478352652862668, + 0.004313459619879723, + 0.00942116416990757, + 0.0035216431133449078, + 0.004653030540794134, + 0.003923790063709021, + 0.002375683980062604, + 0.002319016959518194, + 0.022006476297974586, + 0.035752926021814346, + 0.017001107335090637, + 0.02296692319214344, + 0.06814893335103989, + 0.05415631830692291, + 0.058296091854572296, + 0.018726833164691925, + 0.006095647346228361, + 0.006975616794079542, + 0.01011998113244772, + 0.006841935683041811, + 0.005781945772469044, + 0.007845086045563221, + 0.006202777847647667, + 0.007506118156015873, + 0.009907090105116367, + 0.015371366403996944, + 0.007353934925049543, + 0.013420486822724342, + 0.016085123643279076, + 0.016434000805020332, + 0.013706855475902557, + 0.017483819276094437, + 0.011434657499194145, + 0.0173091571778059, + 0.013751685619354248, + 0.015335410833358765, + 0.023655207827687263, + 0.01974128931760788, + 0.02357158064842224, + 0.02382062003016472, + 0.0038384285289794207, + 0.00836310163140297, + 0.005018678493797779, + 0.0048199924640357494, + 0.010045699775218964, + 0.007158784661442041, + 0.0064261737279593945, + 0.0060868533328175545, + 0.018296703696250916, + 0.014764390885829926, + 0.012914505787193775, + 0.009001321159303188, + 0.01109190471470356, + 0.01161772757768631, + 0.008069871924817562, + 0.005513285286724567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.033694181591272354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078868061304092 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078868061304092, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078868061304092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27390193939208984, + 0.1075238287448883, + 0.05976232886314392, + 0.08223141729831696, + 0.11920091509819031, + 0.06903766095638275, + 0.05198987200856209, + 0.08803129941225052, + 0.06594985723495483, + 0.08237087726593018 + ], + "credit_map": [ + 0.03177414834499359, + 0.02847284823656082, + 0.058165863156318665, + 0.024996422231197357, + 0.028051190078258514, + 0.034638963639736176, + 0.030812745913863182, + 0.03698975220322609, + 0.022356370463967323, + 0.015925470739603043, + 0.008757509291172028, + 0.009333704598248005, + 0.012501238845288754, + 0.013703808188438416, + 0.009204013273119926, + 0.01574171893298626, + 0.008248838596045971, + 0.009735210798680782, + 0.0075938100926578045, + 0.0067347576841712, + 0.010644149966537952, + 0.005087623838335276, + 0.005072105675935745, + 0.006645834539085627, + 0.01123946625739336, + 0.007800368592143059, + 0.02463887259364128, + 0.006778367329388857, + 0.011214234866201878, + 0.010183224454522133, + 0.005649147555232048, + 0.004727729596197605, + 0.008484294638037682, + 0.009132054634392262, + 0.013661772012710571, + 0.008629210293293, + 0.02422110177576542, + 0.02076658047735691, + 0.02527005225419998, + 0.009035848081111908, + 0.008854465559124947, + 0.008134556002914906, + 0.009279665537178516, + 0.006143604405224323, + 0.010877362452447414, + 0.011519454419612885, + 0.007601926103234291, + 0.00662662647664547, + 0.006358668673783541, + 0.008480675518512726, + 0.006170941982418299, + 0.004121663514524698, + 0.005931447260081768, + 0.007457989268004894, + 0.007004359271377325, + 0.006464126519858837, + 0.006563371978700161, + 0.012516583316028118, + 0.00600218353793025, + 0.012357323430478573, + 0.015026205219328403, + 0.013529414311051369, + 0.010171405971050262, + 0.011864821426570415, + 0.005759793799370527, + 0.007421625778079033, + 0.005662580020725727, + 0.007279639597982168, + 0.011440480127930641, + 0.010229547508060932, + 0.007690039929002523, + 0.01046614907681942, + 0.013875486329197884, + 0.016571206972002983, + 0.01057079154998064, + 0.00789584219455719, + 0.011373613029718399, + 0.01117380429059267, + 0.00685860263183713, + 0.004051530733704567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.011078868061304092 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro still shot, captured from an almost reverent, horizontal perspective, captures a dramatic scene of light and shadow, steeped in the sensibility of Old Master paintings. In the image, a slender finger, clad in a deep black velvet glove, its tip stained with a few large, irregularly shaped grains of shimmering sea salt, delicately touches a thinly sliced ham, as crimson as the finest ruby. A single, warm, focused light source from the side creates a precise, intense contrast between light and dark (chiaroscuro), making the fine pile of the velvet appear even darker in the shadows, while the illuminated side radiates a soft, restrained luster. As the fingers gently and steadily rubbed to one side, the sea salt particles left a shiny trail on the moist surface of the ham. The rich white fat texture between the ham's texture was instantly illuminated by the light, reflecting a rich and attractive luster like liquid gold, forming a sharp contrast with the deep red lean meat part. The whole picture, in tranquility and concentration, created an extremely luxurious and ritualistic culinary aesthetic atmosphere.", + "seed": 13268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_3/index_3176_seed_13268_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012634108681231737, + "vlm_logprob": -0.012634108681231737, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.012634108681231737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013477410189807415 + ], + "opposite_logprob_tokens": [ + -4.388477325439453 + ], + "logprob_reward": -0.013477410189807415, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013477410189807415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07814916968345642, + 0.03009282425045967, + 0.7154552936553955, + 0.01952289044857025, + 0.02519618533551693, + 0.01637473702430725, + 0.02571820467710495, + 0.028559017926454544, + 0.04506628215312958, + 0.015865422785282135 + ], + "credit_map": [ + 0.009059318341314793, + 0.010983438231050968, + 0.014976534061133862, + 0.01323666051030159, + 0.009671573527157307, + 0.007327534724026918, + 0.0071099246852099895, + 0.005784171633422375, + 0.004718797747045755, + 0.004111311864107847, + 0.003372733248397708, + 0.004270981997251511, + 0.003348535392433405, + 0.004494380671530962, + 0.003662769217044115, + 0.0021133108530193567, + 0.30900442600250244, + 0.28886422514915466, + 0.015088406391441822, + 0.02292585000395775, + 0.028313880786299706, + 0.01931355707347393, + 0.012029407545924187, + 0.019915474578738213, + 0.002858805237337947, + 0.002305738627910614, + 0.002515022875741124, + 0.0031719820108264685, + 0.0018860043492168188, + 0.0021446608006954193, + 0.003412855789065361, + 0.0012278190115466714, + 0.003927161451429129, + 0.002621287479996681, + 0.004086809698492289, + 0.004334647208452225, + 0.002248104428872466, + 0.0024543057661503553, + 0.0036048537585884333, + 0.0019190105376765132, + 0.0016150230076164007, + 0.002306812908500433, + 0.002582007087767124, + 0.002923897234722972, + 0.00260819005779922, + 0.0013676992384716868, + 0.002000163309276104, + 0.0009709425503388047, + 0.002754649380221963, + 0.002744708675891161, + 0.005223842337727547, + 0.003902195021510124, + 0.0029108095914125443, + 0.0034776460379362106, + 0.002351057017222047, + 0.0023532940540462732, + 0.0029701993335038424, + 0.0035467585548758507, + 0.004850967787206173, + 0.004824477713555098, + 0.0030460390262305737, + 0.00574997765943408, + 0.0022983334492892027, + 0.0012722591636702418, + 0.005107408855110407, + 0.005388377234339714, + 0.004359533544629812, + 0.005510486196726561, + 0.007651322986930609, + 0.007315465249121189, + 0.005523915868252516, + 0.004209768492728472, + 0.002441986696794629, + 0.0024355738423764706, + 0.0013807734940201044, + 0.0022101877257227898, + 0.002227507997304201, + 0.0023618785198777914, + 0.0017012133030220866, + 0.0011062996927648783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.013477410189807415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204819686710835 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014204819686710835, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204819686710835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07531784474849701, + 0.029416896402835846, + 0.5138973593711853, + 0.044096820056438446, + 0.023178212344646454, + 0.017660262063145638, + 0.013786875642836094, + 0.06531889736652374, + 0.1649860143661499, + 0.05234086513519287 + ], + "credit_map": [ + 0.009715636260807514, + 0.011035582050681114, + 0.010675218887627125, + 0.011952557601034641, + 0.013387574814260006, + 0.008054625242948532, + 0.007157621905207634, + 0.0033390242606401443, + 0.006224486045539379, + 0.0030003206338733435, + 0.002633690135553479, + 0.0028214030899107456, + 0.00274452636949718, + 0.006782449781894684, + 0.003011289518326521, + 0.002198730828240514, + 0.2247406542301178, + 0.2101590633392334, + 0.009448044933378696, + 0.016026955097913742, + 0.019629186019301414, + 0.013156475499272346, + 0.006559900939464569, + 0.014177064411342144, + 0.005333289969712496, + 0.004449465312063694, + 0.005033466033637524, + 0.006595572456717491, + 0.004738131538033485, + 0.004079202190041542, + 0.010551013052463531, + 0.0033166795037686825, + 0.00309529691003263, + 0.0026650168001651764, + 0.0044735753908753395, + 0.004970984999090433, + 0.0022488238755613565, + 0.0019397608703002334, + 0.001947023905813694, + 0.0018377287779003382, + 0.002345116576179862, + 0.002260225359350443, + 0.0026379311457276344, + 0.002506702207028866, + 0.0027800938114523888, + 0.0018031090730801225, + 0.0020269895903766155, + 0.0013000941835343838, + 0.0015631854766979814, + 0.001262143487110734, + 0.002746087731793523, + 0.002031502081081271, + 0.0014725382206961513, + 0.0020356886088848114, + 0.0014427665155380964, + 0.001232963870279491, + 0.006235043052583933, + 0.00568836135789752, + 0.012944009155035019, + 0.012018462643027306, + 0.005123422015458345, + 0.014111212454736233, + 0.006150827277451754, + 0.003047560341656208, + 0.009117115288972855, + 0.05477611720561981, + 0.016013128682971, + 0.023402269929647446, + 0.02800379879772663, + 0.015125886537134647, + 0.01150173507630825, + 0.0070459600538015366, + 0.007412292063236237, + 0.005289839114993811, + 0.0052080596797168255, + 0.0059062461368739605, + 0.0072339423932135105, + 0.010530904866755009, + 0.005521612241864204, + 0.005237967241555452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.014204819686710835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193043638020754 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.003193043638020754, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193043638020754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029921989887952805, + 0.012207630090415478, + 0.7252574563026428, + 0.014206388033926487, + 0.0229861531406641, + 0.005636286921799183, + 0.015924058854579926, + 0.0812951922416687, + 0.04483544081449509, + 0.0477294884622097 + ], + "credit_map": [ + 0.004421363119035959, + 0.003457847284153104, + 0.004245534539222717, + 0.00496009923517704, + 0.0038400425110012293, + 0.0035635873209685087, + 0.0038951875176280737, + 0.001538323238492012, + 0.001582272001542151, + 0.002068144269287586, + 0.0012479439610615373, + 0.0014851454179733992, + 0.002398435492068529, + 0.0011743305949494243, + 0.001505394233390689, + 0.0007459635380655527, + 0.31787341833114624, + 0.29858681559562683, + 0.01128834392875433, + 0.022143200039863586, + 0.02769368514418602, + 0.01883085072040558, + 0.009003094397485256, + 0.01983794942498207, + 0.001505084102973342, + 0.0015960937598720193, + 0.0014406370464712381, + 0.0021903414744883776, + 0.0011530985357239842, + 0.0012869089841842651, + 0.0038298622239381075, + 0.001204360043630004, + 0.003138104686513543, + 0.0031731599010527134, + 0.0034273669589310884, + 0.0031803061719983816, + 0.0021561747416853905, + 0.0025562697555869818, + 0.00326314358972013, + 0.002091626636683941, + 0.0006637363112531602, + 0.0006359117687679827, + 0.0009256243356503546, + 0.0010194649221375585, + 0.0008571453508920968, + 0.00047662470024079084, + 0.0007284603198058903, + 0.0003293182817287743, + 0.0021543342154473066, + 0.002729499712586403, + 0.002274641767144203, + 0.002084189560264349, + 0.0015914472751319408, + 0.0021905191242694855, + 0.0014825323596596718, + 0.00141689192969352, + 0.007124989293515682, + 0.006820914801210165, + 0.016149498522281647, + 0.01805707812309265, + 0.006284430623054504, + 0.01543136965483427, + 0.007943913340568542, + 0.003482992062345147, + 0.004203456919640303, + 0.010058972984552383, + 0.004502234514802694, + 0.005217429716140032, + 0.005978395696729422, + 0.0063627250492572784, + 0.004874987993389368, + 0.003637229558080435, + 0.005932641681283712, + 0.0075641898438334465, + 0.004913894459605217, + 0.006303213071078062, + 0.005999618675559759, + 0.0067940400913357735, + 0.0055679031647741795, + 0.0046539814211428165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.003193043638020754 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345033198595047 + ], + "opposite_logprob_tokens": [ + -3.6513450145721436 + ], + "logprob_reward": -0.026345033198595047, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345033198595047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024825898930430412, + 0.013825545087456703, + 0.7111725807189941, + 0.009976770728826523, + 0.03447650000452995, + 0.015223746187984943, + 0.014435669407248497, + 0.10736216604709625, + 0.029340973123908043, + 0.03936009481549263 + ], + "credit_map": [ + 0.003151309210807085, + 0.003540178295224905, + 0.004690266214311123, + 0.0036355811171233654, + 0.0026513973716646433, + 0.0025648539885878563, + 0.002579793566837907, + 0.002012518933042884, + 0.0017328793182969093, + 0.0019739503040909767, + 0.0022799824364483356, + 0.0018339878879487514, + 0.0016767297638580203, + 0.0015263246605172753, + 0.0017453563632443547, + 0.0010563349351286888, + 0.31182846426963806, + 0.29255232214927673, + 0.01097037736326456, + 0.021652553230524063, + 0.02714967355132103, + 0.018640218302607536, + 0.00896584801375866, + 0.019413132220506668, + 0.0014819056959822774, + 0.001413250109180808, + 0.0013978449860587716, + 0.0016336668049916625, + 0.0011867270804941654, + 0.001025960547849536, + 0.001245183520950377, + 0.0005922313430346549, + 0.004464926663786173, + 0.004051445052027702, + 0.006231691222637892, + 0.005542435683310032, + 0.0039047058671712875, + 0.0038029891438782215, + 0.003859017277136445, + 0.0026192867662757635, + 0.0017486291471868753, + 0.0022619005758315325, + 0.0023231557570397854, + 0.002204588148742914, + 0.0024548377841711044, + 0.0014401922235265374, + 0.0016897444147616625, + 0.00110069930087775, + 0.0019440297037363052, + 0.001781147439032793, + 0.002082731109112501, + 0.002123636892065406, + 0.0014566396130248904, + 0.0019180268282070756, + 0.0017156755784526467, + 0.0014137830585241318, + 0.009295857511460781, + 0.008239995688199997, + 0.021416090428829193, + 0.021556027233600616, + 0.008305739611387253, + 0.023363862186670303, + 0.010064215399324894, + 0.005120376590639353, + 0.003099981462582946, + 0.0037233466282486916, + 0.0032686018384993076, + 0.005345760378986597, + 0.0037856947164982557, + 0.004901095759123564, + 0.0030088312923908234, + 0.0022076612804085016, + 0.005906442180275917, + 0.005278789903968573, + 0.005921475123614073, + 0.005714877042919397, + 0.005202790722250938, + 0.005455699749290943, + 0.0037179996725171804, + 0.0021620248444378376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.026345033198595047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a macro still shot featuring a slender finger wearing a deep black velvet glove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tip of the finger stained with large, irregularly shaped grains of shimmering sea salt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the finger touching a thinly sliced ham that is deep red with visible white fat texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the finger gently rubbing to one side, leaving a shiny trail of sea salt on the ham?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting a single warm source from the side creating high contrast (chiaroscuro) with deep shadows and bright highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950236693024635 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950236693024635, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950236693024635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14083166420459747, + 0.07467440515756607, + 0.33023226261138916, + 0.036471132189035416, + 0.07504969090223312, + 0.039643920958042145, + 0.020851440727710724, + 0.09846147149801254, + 0.1108294278383255, + 0.07295462489128113 + ], + "credit_map": [ + 0.018668657168745995, + 0.018558382987976074, + 0.02204793691635132, + 0.021345918998122215, + 0.022646451368927956, + 0.014804638922214508, + 0.01513535063713789, + 0.0076243216171860695, + 0.011504264548420906, + 0.009142696857452393, + 0.008776038885116577, + 0.007841954939067364, + 0.010448111221194267, + 0.015236271545290947, + 0.007446371950209141, + 0.004278688225895166, + 0.1382783055305481, + 0.13027359545230865, + 0.007763295900076628, + 0.011800942942500114, + 0.01341040525585413, + 0.01151843462139368, + 0.00770685775205493, + 0.009480401873588562, + 0.005531836301088333, + 0.004090378992259502, + 0.004457504488527775, + 0.005851083435118198, + 0.004978110082447529, + 0.0036399788223206997, + 0.005447270814329386, + 0.0024749638978391886, + 0.011808245442807674, + 0.009250367991626263, + 0.012409528717398643, + 0.012532171793282032, + 0.006770556326955557, + 0.007137971930205822, + 0.009293650276958942, + 0.00584719143807888, + 0.004480625037103891, + 0.005640021059662104, + 0.006044753361493349, + 0.005965905264019966, + 0.0065651871263980865, + 0.003274309216067195, + 0.005185686983168125, + 0.002487428719177842, + 0.0026486744172871113, + 0.001857950584962964, + 0.004296952858567238, + 0.002857688581570983, + 0.001645142794586718, + 0.0038949495647102594, + 0.0020715552382171154, + 0.001578523893840611, + 0.00938121136277914, + 0.009993776679039001, + 0.01863250695168972, + 0.021766897290945053, + 0.00807125959545374, + 0.01622469164431095, + 0.009810788556933403, + 0.004580329637974501, + 0.01006853487342596, + 0.030302057042717934, + 0.012496670708060265, + 0.013179472647607327, + 0.017796743661165237, + 0.013724167831242085, + 0.00852811336517334, + 0.004733662120997906, + 0.013158855959773064, + 0.011708264239132404, + 0.009243430569767952, + 0.010505661368370056, + 0.00945544894784689, + 0.007657770533114672, + 0.006637748796492815, + 0.004587441682815552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.005950236693024635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_4/index_tensor([3380, 3380])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_4/index_tensor([3380, 3380])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0cafe6c86a4c0eae781170a6a5d9b6b5717ba988 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_4/index_tensor([3380, 3380])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot, slowly and steadily tilting upward and panning forward, draws the viewer deep into an ancient and tranquil forest. In the dappled twilight filtering through the dense canopy, thousands of translucent crystal fragments, radiating an inner glow, slowly rise from the cracks in the trees and from the leaf-covered ground like living fireflies. These fragments, shimmering in shades of warm gold, soft pink, and delicate blue, trace graceful paths through the air, drawn by an invisible gravitational force, spiraling toward the center of the forest clearing. As the fragments merge, a vast crystal lotus composed of pure light energy steadily and gracefully unfolds from a tiny core. Each translucent petal is woven from countless tiny points of light, revealing delicate textures and veins, while the center of the pistil radiates the brightest, warmest golden light. The entire image is imbued with the warm, hand-painted quality that is signature to Hayao Miyazaki's animations, with saturated, soft colors and delicate, brushstroke-like outlines. The soft glow of this flower of light casts a dreamy halo on the surrounding moss-covered ancient trees and the ground, creating an atmosphere of extreme tranquility, magic, and healing power.", + "seed": 25648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_4/index_3380_seed_25648_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.034401114797219634, + "vlm_logprob": -0.034401114797219634, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.034401114797219634, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10603667795658112 + ], + "opposite_logprob_tokens": [ + -2.35603666305542 + ], + "logprob_reward": -0.10603667795658112, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10603667795658112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5561402440071106, + 0.03330530971288681, + 0.03337651863694191, + 0.03587695211172104, + 0.02156422846019268, + 0.06297671049833298, + 0.02568509243428707, + 0.1447269171476364, + 0.04193814471364021, + 0.04440983757376671 + ], + "credit_map": [ + 0.02544114924967289, + 0.013519038446247578, + 0.21665357053279877, + 0.03134649991989136, + 0.016514623537659645, + 0.021549029275774956, + 0.01297676283866167, + 0.21813960373401642, + 0.0026129407342523336, + 0.0032573279459029436, + 0.004918635357171297, + 0.004974566400051117, + 0.002579297637566924, + 0.003400057554244995, + 0.006925227120518684, + 0.0046372562646865845, + 0.003532630857080221, + 0.004571501165628433, + 0.0051784333772957325, + 0.0027638135943561792, + 0.003378179157152772, + 0.004606843926012516, + 0.0032686784397810698, + 0.0060764397494494915, + 0.0024772435426712036, + 0.004920331761240959, + 0.005792197305709124, + 0.0033546939957886934, + 0.003847497282549739, + 0.004209351725876331, + 0.003836908843368292, + 0.00743872718885541, + 0.0024177348241209984, + 0.002495722845196724, + 0.0032279256265610456, + 0.002157087903469801, + 0.002106680069118738, + 0.004125051666051149, + 0.002217911882326007, + 0.002816114341840148, + 0.007003418169915676, + 0.008212701417505741, + 0.006541809067130089, + 0.007985759526491165, + 0.006934050004929304, + 0.007478170096874237, + 0.010781610384583473, + 0.008039193227887154, + 0.004058661870658398, + 0.0037231796886771917, + 0.004028117749840021, + 0.0023313399869948626, + 0.0037005385383963585, + 0.0026090589817613363, + 0.0025931466370821, + 0.002641049213707447, + 0.00873245857656002, + 0.00891247857362032, + 0.011378642171621323, + 0.03999081626534462, + 0.004638912621885538, + 0.013682854361832142, + 0.01457550935447216, + 0.04281523823738098, + 0.00480647012591362, + 0.0043278527446091175, + 0.00635684235021472, + 0.004910815507173538, + 0.0042962972074747086, + 0.005094505846500397, + 0.00602986803278327, + 0.006115492433309555, + 0.004215956199914217, + 0.004501001909375191, + 0.0036503656301647425, + 0.0071331774815917015, + 0.004339495208114386, + 0.006921570748090744, + 0.00803637970238924, + 0.005611891858279705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.10603667795658112 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635142020881176 + ], + "opposite_logprob_tokens": [ + -3.6513514518737793 + ], + "logprob_reward": -0.02635142020881176, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635142020881176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09485601633787155, + 0.040810804814100266, + 0.09978026151657104, + 0.06605438143014908, + 0.020871469751000404, + 0.0660160556435585, + 0.030805183574557304, + 0.5026859045028687, + 0.0488433875143528, + 0.029276587069034576 + ], + "credit_map": [ + 0.004897645674645901, + 0.004508602432906628, + 0.03178299963474274, + 0.00552656315267086, + 0.004440949764102697, + 0.006612827070057392, + 0.004234224557876587, + 0.032852206379175186, + 0.0032706817146390676, + 0.003604928031563759, + 0.006874210201203823, + 0.006387399043887854, + 0.0034069668035954237, + 0.004236680455505848, + 0.007706818636506796, + 0.005323120858520269, + 0.00941039901226759, + 0.016084207221865654, + 0.015792250633239746, + 0.00860289391130209, + 0.011836162768304348, + 0.011637388728559017, + 0.009111694060266018, + 0.017305269837379456, + 0.004956778138875961, + 0.006930420640856028, + 0.009242719039320946, + 0.008459364995360374, + 0.007519994862377644, + 0.007068591192364693, + 0.010995839722454548, + 0.010880676098167896, + 0.0021154063288122416, + 0.001929845311678946, + 0.0035593006759881973, + 0.0025127383414655924, + 0.0017733762506395578, + 0.003502293024212122, + 0.0022210271563380957, + 0.0032574827782809734, + 0.0069969031028449535, + 0.010438002645969391, + 0.008431149646639824, + 0.008327864110469818, + 0.008145855739712715, + 0.006999909412115812, + 0.008741114288568497, + 0.00793524831533432, + 0.003316976595669985, + 0.0023784891236573458, + 0.003859012620523572, + 0.004499561153352261, + 0.0028220766689628363, + 0.004470168612897396, + 0.004922355990856886, + 0.004536543041467667, + 0.02824844792485237, + 0.053408924490213394, + 0.031167520210146904, + 0.165176659822464, + 0.01895289123058319, + 0.03978443518280983, + 0.05956687033176422, + 0.10638011246919632, + 0.004515386186540127, + 0.004620232153683901, + 0.006064951419830322, + 0.008444299921393394, + 0.004171063657850027, + 0.00628290930762887, + 0.006588118616491556, + 0.008156424388289452, + 0.0009118999005295336, + 0.0009682806557975709, + 0.0006907470524311066, + 0.011383705772459507, + 0.0010877751046791673, + 0.0011472753249108791, + 0.011942340061068535, + 0.0011445637792348862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.02635142020881176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631315384060144 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631315384060144, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631315384060144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4247191548347473, + 0.01672857254743576, + 0.04457971081137657, + 0.0726405680179596, + 0.02484707720577717, + 0.024188373237848282, + 0.00800947193056345, + 0.35294729471206665, + 0.013800250366330147, + 0.01753954403102398 + ], + "credit_map": [ + 0.018977424129843712, + 0.00838138535618782, + 0.16998353600502014, + 0.02114296332001686, + 0.010051572695374489, + 0.015003197826445103, + 0.008654279634356499, + 0.17252478003501892, + 0.0016080059576779604, + 0.0014292526757344604, + 0.002652885625138879, + 0.002301724860444665, + 0.001405494986101985, + 0.0020937342196702957, + 0.0025699366815388203, + 0.0026675371918827295, + 0.005011068191379309, + 0.007281701546162367, + 0.00707683339715004, + 0.0038765461649745703, + 0.005349648650735617, + 0.005855272524058819, + 0.002643120940774679, + 0.007485516835004091, + 0.0057205636985599995, + 0.011501433327794075, + 0.007500461768358946, + 0.0076943663880229, + 0.011297892779111862, + 0.00867723673582077, + 0.01006218884140253, + 0.010186427272856236, + 0.002485225209966302, + 0.0019539447966963053, + 0.003029129933565855, + 0.0028580946382135153, + 0.002295269165188074, + 0.005552658345550299, + 0.003015529364347458, + 0.003657226450741291, + 0.0029625054448843002, + 0.004114719107747078, + 0.0018682788359001279, + 0.0021810249891132116, + 0.0033393166959285736, + 0.0030638810712844133, + 0.003476269543170929, + 0.0031823788303881884, + 0.0009915176779031754, + 0.0009985953802242875, + 0.0009940159507095814, + 0.0008164030150510371, + 0.0010908673284575343, + 0.0009828528854995966, + 0.000955760944634676, + 0.0011794586898759007, + 0.01947684958577156, + 0.033830221742391586, + 0.022389298304915428, + 0.11413415521383286, + 0.012597772292792797, + 0.02836282178759575, + 0.04124991595745087, + 0.08090626448392868, + 0.0015451365616172552, + 0.0013437868328765035, + 0.0024367100559175014, + 0.0019895564764738083, + 0.0011692331172525883, + 0.0012406441383063793, + 0.0016793084796518087, + 0.0023958745878189802, + 0.0010717901168391109, + 0.0011629724176600575, + 0.0011962008429691195, + 0.005329883191734552, + 0.0013792809331789613, + 0.0013272009091451764, + 0.005248278379440308, + 0.000823936949018389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.007631315384060144 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021964015904814005 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021964015904814005, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021964015904814005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.132791206240654, + 0.040625788271427155, + 0.023344306275248528, + 0.05227617546916008, + 0.031846050173044205, + 0.04759703576564789, + 0.022259384393692017, + 0.5842556953430176, + 0.04224899411201477, + 0.02275538444519043 + ], + "credit_map": [ + 0.007184751331806183, + 0.008571561425924301, + 0.039484523236751556, + 0.013100802898406982, + 0.009410378523170948, + 0.006919571664184332, + 0.009846819564700127, + 0.03827279433608055, + 0.003270903369411826, + 0.004041033796966076, + 0.0055097476579248905, + 0.006511088460683823, + 0.003162556793540716, + 0.004038963932543993, + 0.008792337030172348, + 0.005299157928675413, + 0.0017715709982439876, + 0.0025080314371734858, + 0.005741749424487352, + 0.0021769131999462843, + 0.0018958726432174444, + 0.0019431135151535273, + 0.002559378044679761, + 0.004747677128762007, + 0.0033516953699290752, + 0.009263352490961552, + 0.004936938174068928, + 0.005454220809042454, + 0.010122292675077915, + 0.0049232072196900845, + 0.006248675752431154, + 0.00797579251229763, + 0.002602607710286975, + 0.0028588136192411184, + 0.0037717975210398436, + 0.0054443273693323135, + 0.003907413221895695, + 0.0025309862103313208, + 0.006954571697860956, + 0.003775532590225339, + 0.003917624708265066, + 0.004238068126142025, + 0.005874976050108671, + 0.007146273273974657, + 0.0052781254053115845, + 0.006544917821884155, + 0.00792276207357645, + 0.00667429156601429, + 0.002425030106678605, + 0.004780549090355635, + 0.0019876034930348396, + 0.001979669788852334, + 0.004806021228432655, + 0.001560069853439927, + 0.0028308145701885223, + 0.001889625913463533, + 0.03159981966018677, + 0.0546782985329628, + 0.03704805672168732, + 0.18692660331726074, + 0.020636877045035362, + 0.04746276140213013, + 0.06747706979513168, + 0.13842616975307465, + 0.004745970480144024, + 0.004821069538593292, + 0.005991184152662754, + 0.007188005372881889, + 0.0032587053719908, + 0.004143108148127794, + 0.005915225949138403, + 0.006185723468661308, + 0.0014090840704739094, + 0.0013905622763559222, + 0.0014717656886205077, + 0.00586797995492816, + 0.001547935069538653, + 0.002805054420605302, + 0.006048946641385555, + 0.0022140569053590298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0021964015904814005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978975884616375 + ], + "opposite_logprob_tokens": [ + -3.529789686203003 + ], + "logprob_reward": -0.02978975884616375, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978975884616375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1746339648962021, + 0.062422338873147964, + 0.12070494145154953, + 0.07195191830396652, + 0.044156234711408615, + 0.05490764603018761, + 0.06618832796812057, + 0.30379804968833923, + 0.03251011669635773, + 0.0687265545129776 + ], + "credit_map": [ + 0.01673770509660244, + 0.00794160645455122, + 0.047975145280361176, + 0.01610582135617733, + 0.008329092524945736, + 0.022482607513666153, + 0.010467947460711002, + 0.044594015926122665, + 0.005018362309783697, + 0.006676110439002514, + 0.008950629271566868, + 0.009618940763175488, + 0.005127506796270609, + 0.006752441171556711, + 0.012533504515886307, + 0.007744837552309036, + 0.00682913837954402, + 0.012097081169486046, + 0.03282458335161209, + 0.009348787367343903, + 0.00890403613448143, + 0.00960233248770237, + 0.006931371986865997, + 0.034167591482400894, + 0.00759042939171195, + 0.007769737858325243, + 0.00839630700647831, + 0.009403162635862827, + 0.009443865157663822, + 0.008557385765016079, + 0.010610183700919151, + 0.010180833749473095, + 0.005616785027086735, + 0.003282781457528472, + 0.005913563072681427, + 0.006323442794382572, + 0.0033653993159532547, + 0.007816500030457973, + 0.005674673244357109, + 0.006163081619888544, + 0.006390345748513937, + 0.00809873640537262, + 0.005961655639111996, + 0.0049004629254341125, + 0.008241142146289349, + 0.007196022663265467, + 0.0077181244269013405, + 0.0064011490903794765, + 0.006592414807528257, + 0.008151700720191002, + 0.0098109794780612, + 0.007820440456271172, + 0.008857744745910168, + 0.007851055823266506, + 0.006904530804604292, + 0.010199455544352531, + 0.016816070303320885, + 0.03245691582560539, + 0.019131822511553764, + 0.09953727573156357, + 0.011180299334228039, + 0.024369124323129654, + 0.035465460270643234, + 0.06484104692935944, + 0.0028509621042758226, + 0.003156512277200818, + 0.004739298019558191, + 0.003615833120420575, + 0.0026405793614685535, + 0.0033906539902091026, + 0.004365469794720411, + 0.007750803604722023, + 0.003625720739364624, + 0.003810955211520195, + 0.003771228948608041, + 0.022556545212864876, + 0.00408145459368825, + 0.004579068161547184, + 0.022509776055812836, + 0.003791798371821642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.02978975884616375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, slowly and steadily tilting upward and panning forward, draws the viewer deep into an ancient and tranquil forest. In the dappled twilight filtering through the dense canopy, thousands of translucent crystal fragments, radiating an inner glow, slowly rise from the cracks in the trees and from the leaf-covered ground like living fireflies. These fragments, shimmering in shades of warm gold, soft pink, and delicate blue, trace graceful paths through the air, drawn by an invisible gravitational force, spiraling toward the center of the forest clearing. As the fragments merge, a vast crystal lotus composed of pure light energy steadily and gracefully unfolds from a tiny core. Each translucent petal is woven from countless tiny points of light, revealing delicate textures and veins, while the center of the pistil radiates the brightest, warmest golden light. The entire image is imbued with the warm, hand-painted quality that is signature to Hayao Miyazaki's animations, with saturated, soft colors and delicate, brushstroke-like outlines. The soft glow of this flower of light casts a dreamy halo on the surrounding moss-covered ancient trees and the ground, creating an atmosphere of extreme tranquility, magic, and healing power.", + "seed": 25649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_4/index_3380_seed_25649_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0294888709904626, + "vlm_logprob": -0.0294888709904626, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.0294888709904626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09357517957687378 + ], + "opposite_logprob_tokens": [ + -2.4685752391815186 + ], + "logprob_reward": -0.09357517957687378, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09357517957687378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20770396292209625, + 0.06727848947048187, + 0.042231813073158264, + 0.08509386330842972, + 0.04496549442410469, + 0.28876033425331116, + 0.0301740113645792, + 0.07953323423862457, + 0.14441807568073273, + 0.009840663522481918 + ], + "credit_map": [ + 0.027701320126652718, + 0.01853039301931858, + 0.042294781655073166, + 0.02054508589208126, + 0.022678488865494728, + 0.018358904868364334, + 0.017112215980887413, + 0.0404827706515789, + 0.008700349368155003, + 0.010235901921987534, + 0.008707595989108086, + 0.006226932629942894, + 0.007552505470812321, + 0.007331913337111473, + 0.00697264913469553, + 0.011550646275281906, + 0.005956974811851978, + 0.005383764859288931, + 0.0065046921372413635, + 0.006186973303556442, + 0.003727884963154793, + 0.003679367946460843, + 0.005113949999213219, + 0.005678208079189062, + 0.00639532133936882, + 0.014978364109992981, + 0.010422170162200928, + 0.012600419111549854, + 0.008685139007866383, + 0.0049721612595021725, + 0.010372421704232693, + 0.01666787452995777, + 0.00303401960991323, + 0.003939023707062006, + 0.012542709708213806, + 0.0041058361530303955, + 0.003780246712267399, + 0.003108406439423561, + 0.008683319203555584, + 0.005771937780082226, + 0.009812644682824612, + 0.06606736034154892, + 0.013594874180853367, + 0.1040583997964859, + 0.023642461746931076, + 0.010837499983608723, + 0.051162950694561005, + 0.009584162384271622, + 0.0026771638076752424, + 0.0030932165682315826, + 0.002671827794983983, + 0.0035892468877136707, + 0.0022033939603716135, + 0.006078888662159443, + 0.003671179758384824, + 0.0061890967190265656, + 0.0028644015546888113, + 0.005739131011068821, + 0.0027813741471618414, + 0.031917184591293335, + 0.00328090600669384, + 0.00214753788895905, + 0.02803170494735241, + 0.002771001076325774, + 0.011789235286414623, + 0.019123917445540428, + 0.02140052057802677, + 0.017715608701109886, + 0.017088517546653748, + 0.021370455622673035, + 0.01802174746990204, + 0.017908088862895966, + 0.0011580674909055233, + 0.0010802921606227756, + 0.001085422351025045, + 0.0013782447203993797, + 0.0007874279981479049, + 0.0010129837319254875, + 0.0017869309522211552, + 0.0015512957470491529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.09357517957687378 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350026950240135 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026350026950240135, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350026950240135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25591400265693665, + 0.06233315169811249, + 0.042396947741508484, + 0.036995988339185715, + 0.05719269439578056, + 0.19274310767650604, + 0.11380551755428314, + 0.15566924214363098, + 0.05764845013618469, + 0.02530086413025856 + ], + "credit_map": [ + 0.022351598367094994, + 0.028170503675937653, + 0.0521119199693203, + 0.022383728995919228, + 0.03057766519486904, + 0.01769600249826908, + 0.023242752999067307, + 0.059379834681749344, + 0.005859780590981245, + 0.00827283225953579, + 0.011531230993568897, + 0.0075170304626226425, + 0.006989679764956236, + 0.005874680355191231, + 0.0057260687462985516, + 0.010561849921941757, + 0.0034334594383835793, + 0.0033450003247708082, + 0.004800416529178619, + 0.008340489119291306, + 0.005350667517632246, + 0.0032993985805660486, + 0.007701252121478319, + 0.006126266438513994, + 0.004307782743126154, + 0.004814043641090393, + 0.00575272785499692, + 0.004236150067299604, + 0.004497717600315809, + 0.003795105032622814, + 0.00431316252797842, + 0.005279298406094313, + 0.002992079360410571, + 0.0077423639595508575, + 0.016443395987153053, + 0.0044816890731453896, + 0.006590884644538164, + 0.0023062056861817837, + 0.010056895203888416, + 0.006579179782420397, + 0.006552303209900856, + 0.043046753853559494, + 0.009216690436005592, + 0.06994283944368362, + 0.015208554454147816, + 0.007041953969746828, + 0.03513190895318985, + 0.006602121517062187, + 0.005378330126404762, + 0.012858211062848568, + 0.006816359236836433, + 0.03271067515015602, + 0.006667032837867737, + 0.016054902225732803, + 0.019471585750579834, + 0.013848419301211834, + 0.003129989840090275, + 0.004850238561630249, + 0.004913989454507828, + 0.06539993733167648, + 0.004042032174766064, + 0.0030804681591689587, + 0.06445598602294922, + 0.00579660153016448, + 0.004445868544280529, + 0.007237858138978481, + 0.008486769162118435, + 0.0075470805168151855, + 0.007218559272587299, + 0.007262387778609991, + 0.008185159415006638, + 0.007264762185513973, + 0.0023937865626066923, + 0.0035661724396049976, + 0.00291969021782279, + 0.0028040946926921606, + 0.00385088543407619, + 0.0031701456755399704, + 0.003602268872782588, + 0.0029938232619315386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.026350026950240135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627766579389572 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627766579389572, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627766579389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11682262271642685, + 0.02909683622419834, + 0.023610107600688934, + 0.01535730343312025, + 0.035274162888526917, + 0.04363488405942917, + 0.016684027388691902, + 0.6628773808479309, + 0.025303004309535027, + 0.03133971244096756 + ], + "credit_map": [ + 0.006252069026231766, + 0.008573940023779869, + 0.033844441175460815, + 0.009135067462921143, + 0.00974500086158514, + 0.005944566801190376, + 0.007729370146989822, + 0.035598162561655045, + 0.004286384209990501, + 0.0027727321721613407, + 0.005052412860095501, + 0.00290047493763268, + 0.0037952412385493517, + 0.0031053994316607714, + 0.0026172648649662733, + 0.004566927906125784, + 0.0020294543355703354, + 0.0022828762885183096, + 0.0028286732267588377, + 0.004974566865712404, + 0.0028169373981654644, + 0.002034970559179783, + 0.003588158404454589, + 0.0030544716864824295, + 0.0015360817778855562, + 0.0025285626761615276, + 0.0015112345572561026, + 0.001869774772785604, + 0.001986048649996519, + 0.001361870439723134, + 0.0020170893985778093, + 0.002546641742810607, + 0.001674350816756487, + 0.001792202820070088, + 0.0066377511247992516, + 0.005524272099137306, + 0.00281671155244112, + 0.0020322559867054224, + 0.012085124850273132, + 0.0027114921249449253, + 0.0019735468085855246, + 0.007748006377369165, + 0.002205793047323823, + 0.016449956223368645, + 0.002539408626034856, + 0.0022701278794556856, + 0.008576598949730396, + 0.0018714454490691423, + 0.0008206497877836227, + 0.002752945525571704, + 0.001297258073464036, + 0.002626581583172083, + 0.002042796229943633, + 0.001207683002576232, + 0.003679068526253104, + 0.0022570432629436255, + 0.011303076520562172, + 0.018977416679263115, + 0.01726382039487362, + 0.2866719961166382, + 0.015503271482884884, + 0.008380998857319355, + 0.2837417423725128, + 0.021035049110651016, + 0.0018455757526680827, + 0.0031238028313964605, + 0.003517612349241972, + 0.0032343019265681505, + 0.00308181787841022, + 0.003488731337711215, + 0.004296176601201296, + 0.0027149850502610207, + 0.002494889311492443, + 0.006333953235298395, + 0.0026203137822449207, + 0.003248924855142832, + 0.004713351372629404, + 0.003074375446885824, + 0.00376109778881073, + 0.005092807579785585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.007627766579389572 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017114529618993402 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017114529618993402, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017114529618993402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23511691391468048, + 0.05220692604780197, + 0.13191929459571838, + 0.06836920231580734, + 0.07713641226291656, + 0.014232059940695763, + 0.038324885070323944, + 0.32118362188339233, + 0.029224587604403496, + 0.03228602185845375 + ], + "credit_map": [ + 0.017040515318512917, + 0.013944377191364765, + 0.07171366363763809, + 0.014874424785375595, + 0.014087884686887264, + 0.013098780065774918, + 0.012743796221911907, + 0.07761348783969879, + 0.006160094868391752, + 0.0065408069640398026, + 0.007408210542052984, + 0.005574329290539026, + 0.005598295480012894, + 0.006143484730273485, + 0.0057885246351361275, + 0.008993181400001049, + 0.00823181588202715, + 0.01182405836880207, + 0.0146608492359519, + 0.029955005273222923, + 0.014860298484563828, + 0.009732063859701157, + 0.02460223250091076, + 0.018052978441119194, + 0.0068083410151302814, + 0.007855442352592945, + 0.008709832094609737, + 0.011574659496545792, + 0.006701909005641937, + 0.004535713233053684, + 0.012487782165408134, + 0.009695529006421566, + 0.003177147125825286, + 0.004058544524013996, + 0.01459711603820324, + 0.010733610950410366, + 0.008686477318406105, + 0.005283897276967764, + 0.025096485391259193, + 0.005503146909177303, + 0.0010765742044895887, + 0.0038379847537726164, + 0.0012171990238130093, + 0.0020898578222841024, + 0.002183395903557539, + 0.0009244116372428834, + 0.0018677646294236183, + 0.0010348728392273188, + 0.003214366501197219, + 0.004659564234316349, + 0.0031276477966457605, + 0.007678537163883448, + 0.002997193718329072, + 0.004788521211594343, + 0.006216179579496384, + 0.005642877891659737, + 0.005266671534627676, + 0.009225376881659031, + 0.007894118316471577, + 0.13747139275074005, + 0.0072420379146933556, + 0.00384985632263124, + 0.13958558440208435, + 0.010648624040186405, + 0.0020588627085089684, + 0.0034165228717029095, + 0.003980141598731279, + 0.003890398656949401, + 0.003475266508758068, + 0.004139594733715057, + 0.005334150977432728, + 0.0029296528082340956, + 0.003382523311302066, + 0.0041504716500639915, + 0.004087085835635662, + 0.0037459651939570904, + 0.004329303279519081, + 0.004639299586415291, + 0.004053276963531971, + 0.0038980955723673105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0017114529618993402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817992888391018 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.01817992888391018, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817992888391018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6595408320426941, + 0.01735895499587059, + 0.006931205280125141, + 0.019678687676787376, + 0.013474845327436924, + 0.026962215080857277, + 0.029802128672599792, + 0.1929672360420227, + 0.02823823131620884, + 0.005045594181865454 + ], + "credit_map": [ + 0.034088969230651855, + 0.017782673239707947, + 0.26244980096817017, + 0.0308847539126873, + 0.012018772773444653, + 0.02133357711136341, + 0.012963654473423958, + 0.2680186629295349, + 0.002416822127997875, + 0.0016975634498521686, + 0.004310538060963154, + 0.001257826341316104, + 0.0016269962070509791, + 0.0018404697766527534, + 0.0014246748760342598, + 0.0027840631082654, + 0.0005599881405942142, + 0.0007084333919920027, + 0.0009618549956940114, + 0.0012218703050166368, + 0.0006529887323267758, + 0.0008225188939832151, + 0.0010844434145838022, + 0.0009191073477268219, + 0.002199796726927161, + 0.004480425268411636, + 0.002138451673090458, + 0.0025143837556242943, + 0.0028313258662819862, + 0.0017310258699581027, + 0.0018790592439472675, + 0.0019042198546230793, + 0.001035673194564879, + 0.0018074580002576113, + 0.002639158396050334, + 0.0020377084147185087, + 0.0013908204855397344, + 0.0010602731490507722, + 0.0022047623060643673, + 0.0012989923125132918, + 0.0017910946626216173, + 0.005983220413327217, + 0.0014895767671987414, + 0.0058957794681191444, + 0.003673058934509754, + 0.002660614438354969, + 0.003906734753400087, + 0.001562135643325746, + 0.0019275282975286245, + 0.0038775659631937742, + 0.0020195923279970884, + 0.010037855245172977, + 0.0017551223281770945, + 0.0017447334248572588, + 0.0055858236737549305, + 0.0028539085760712624, + 0.004223660100251436, + 0.007069176994264126, + 0.0055497377179563046, + 0.08391578495502472, + 0.004399633966386318, + 0.0029148522298783064, + 0.07939756661653519, + 0.005496830213814974, + 0.0019037446472793818, + 0.0037110571283847094, + 0.003758925013244152, + 0.0038077030330896378, + 0.0034536125604063272, + 0.003833272960036993, + 0.004999105352908373, + 0.0027708120178431273, + 0.0003520693862810731, + 0.0008962384890764952, + 0.0004768547078128904, + 0.0006409560446627438, + 0.0006133865681476891, + 0.0003702511894516647, + 0.0008119621779769659, + 0.0008838754147291183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.01817992888391018 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, slowly and steadily tilting upward and panning forward, draws the viewer deep into an ancient and tranquil forest. In the dappled twilight filtering through the dense canopy, thousands of translucent crystal fragments, radiating an inner glow, slowly rise from the cracks in the trees and from the leaf-covered ground like living fireflies. These fragments, shimmering in shades of warm gold, soft pink, and delicate blue, trace graceful paths through the air, drawn by an invisible gravitational force, spiraling toward the center of the forest clearing. As the fragments merge, a vast crystal lotus composed of pure light energy steadily and gracefully unfolds from a tiny core. Each translucent petal is woven from countless tiny points of light, revealing delicate textures and veins, while the center of the pistil radiates the brightest, warmest golden light. The entire image is imbued with the warm, hand-painted quality that is signature to Hayao Miyazaki's animations, with saturated, soft colors and delicate, brushstroke-like outlines. The soft glow of this flower of light casts a dreamy halo on the surrounding moss-covered ancient trees and the ground, creating an atmosphere of extreme tranquility, magic, and healing power.", + "seed": 25650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_4/index_3380_seed_25650_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.042404714459553364, + "vlm_logprob": -0.042404714459553364, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.042404714459553364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10772155225276947 + ], + "opposite_logprob_tokens": [ + -2.3577215671539307 + ], + "logprob_reward": -0.10772155225276947, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10772155225276947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45572829246520996, + 0.1534481644630432, + 0.10444781184196472, + 0.11124251782894135, + 0.028401797637343407, + 0.0568012073636055, + 0.04820714518427849, + 0.016741083934903145, + 0.012551096267998219, + 0.012430844828486443 + ], + "credit_map": [ + 0.033857863396406174, + 0.03165792301297188, + 0.02013842575252056, + 0.023570114746689796, + 0.03924822434782982, + 0.028809960931539536, + 0.14008381962776184, + 0.13836197555065155, + 0.019805943593382835, + 0.013195957988500595, + 0.01891983486711979, + 0.013836288824677467, + 0.016726680099964142, + 0.02870527096092701, + 0.02078644186258316, + 0.021471749991178513, + 0.012219827622175217, + 0.008943489752709866, + 0.009235549718141556, + 0.013744779862463474, + 0.010180065408349037, + 0.024592475965619087, + 0.011032880283892155, + 0.014498740434646606, + 0.009079760871827602, + 0.015247795730829239, + 0.01590724103152752, + 0.011188484728336334, + 0.02158144675195217, + 0.014785832725465298, + 0.013793236576020718, + 0.009658720344305038, + 0.002918137004598975, + 0.002173904562368989, + 0.0051447805017232895, + 0.0028425371274352074, + 0.0024623004719614983, + 0.004014020785689354, + 0.004818891640752554, + 0.0040272255428135395, + 0.006631376687437296, + 0.005539880134165287, + 0.009468186646699905, + 0.006188330706208944, + 0.005213025491684675, + 0.0064455787651240826, + 0.009010029025375843, + 0.008304801769554615, + 0.00439600320532918, + 0.006838025525212288, + 0.00332287373021245, + 0.004938923288136721, + 0.005847469437867403, + 0.009416579268872738, + 0.007142993155866861, + 0.006304276641458273, + 0.0017791260033845901, + 0.0016673686914145947, + 0.0011609832290560007, + 0.0023636252153664827, + 0.001750525669194758, + 0.0031449527014046907, + 0.0025587568525224924, + 0.002315744524821639, + 0.0018093588296324015, + 0.002005435526371002, + 0.0009500557207502425, + 0.0012038730783388019, + 0.0022312162909656763, + 0.0014777113683521748, + 0.0014176755212247372, + 0.0014557699905708432, + 0.002650773385539651, + 0.0012184490915387869, + 0.0010328127536922693, + 0.0014004093827679753, + 0.0009305724524892867, + 0.001842982368543744, + 0.0017874451586976647, + 0.0015673998277634382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.10772155225276947 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810946270823479 + ], + "opposite_logprob_tokens": [ + -3.288109540939331 + ], + "logprob_reward": -0.03810946270823479, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810946270823479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47878116369247437, + 0.14425109326839447, + 0.08724497258663177, + 0.09447447210550308, + 0.038045305758714676, + 0.05730302631855011, + 0.05397442355751991, + 0.029230065643787384, + 0.007048391737043858, + 0.009647087194025517 + ], + "credit_map": [ + 0.03742902725934982, + 0.022125134244561195, + 0.020791655406355858, + 0.025301311165094376, + 0.03156270086765289, + 0.038489021360874176, + 0.1437341570854187, + 0.15934817492961884, + 0.01902400329709053, + 0.014319149777293205, + 0.01640959456562996, + 0.013043145649135113, + 0.01516149751842022, + 0.026164207607507706, + 0.022514155134558678, + 0.01761532761156559, + 0.010283651761710644, + 0.005416668485850096, + 0.007250863593071699, + 0.008041393011808395, + 0.008836586028337479, + 0.021420661360025406, + 0.014022844843566418, + 0.011972308158874512, + 0.00892842561006546, + 0.01058470644056797, + 0.012966054491698742, + 0.011272883974015713, + 0.01703045517206192, + 0.012051064521074295, + 0.012378974817693233, + 0.0092619014903903, + 0.004131000488996506, + 0.00434234831482172, + 0.006716609001159668, + 0.00322351255454123, + 0.005256309639662504, + 0.004905372392386198, + 0.004909528885036707, + 0.004560626577585936, + 0.007451100740581751, + 0.005088211502879858, + 0.008821418508887291, + 0.00594565412029624, + 0.004912663251161575, + 0.006943775340914726, + 0.009193062782287598, + 0.008947140537202358, + 0.004972364287823439, + 0.006746371742337942, + 0.003969490062445402, + 0.006270520389080048, + 0.005562913604080677, + 0.010290340520441532, + 0.008559873327612877, + 0.007602548226714134, + 0.004434227477759123, + 0.0023926093708723783, + 0.0027057251427322626, + 0.0034159754868596792, + 0.0025478273164480925, + 0.006522639654576778, + 0.004013285972177982, + 0.0031977754551917315, + 0.001211521215736866, + 0.0005103069706819952, + 0.0008146247128024697, + 0.0008420135709457099, + 0.0005863427650183439, + 0.001222224091179669, + 0.0007972579915076494, + 0.0010641004191711545, + 0.0017243406036868691, + 0.0007882934878580272, + 0.000837838975712657, + 0.0012091813841834664, + 0.0010143015533685684, + 0.0017934530042111874, + 0.0010432879207655787, + 0.0012363909045234323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03810946270823479 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631906773895025 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631906773895025, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631906773895025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8988441824913025, + 0.018531223759055138, + 0.023823626339435577, + 0.018869059160351753, + 0.008278836496174335, + 0.007073710206896067, + 0.010055653750896454, + 0.007994761690497398, + 0.0029344887007027864, + 0.0035943281836807728 + ], + "credit_map": [ + 0.059087805449962616, + 0.02816012129187584, + 0.027514241635799408, + 0.03227589651942253, + 0.03008040226995945, + 0.054803308099508286, + 0.3299529552459717, + 0.3369695842266083, + 0.002840595319867134, + 0.0025544785894453526, + 0.0020551769994199276, + 0.0016805698396638036, + 0.0016726333415135741, + 0.002473618136718869, + 0.0029636071994900703, + 0.002290545031428337, + 0.0031723978463560343, + 0.0016635096399113536, + 0.0022024225909262896, + 0.001992373028770089, + 0.001719042775221169, + 0.006234745495021343, + 0.0034076934680342674, + 0.0034314424265176058, + 0.0020198056008666754, + 0.0015391448978334665, + 0.0040845489129424095, + 0.002797007327899337, + 0.0017268323572352529, + 0.0023668950889259577, + 0.0022223666310310364, + 0.0021124579943716526, + 0.0010393913835287094, + 0.001064803684130311, + 0.0011714273132383823, + 0.0008623662870377302, + 0.0010606198338791728, + 0.0009828998008742929, + 0.0012068331707268953, + 0.000890494731720537, + 0.0008845461416058242, + 0.0005362434894777834, + 0.0009419815614819527, + 0.0007596072973683476, + 0.0006412062211893499, + 0.0010425490327179432, + 0.001005043857730925, + 0.001262533012777567, + 0.0008333896985277534, + 0.001245652325451374, + 0.0006430188077501953, + 0.0010206227889284492, + 0.0010177291696891189, + 0.0018471527146175504, + 0.001731249736621976, + 0.0017168388003483415, + 0.001345399534329772, + 0.0010282878065481782, + 0.0007527191773988307, + 0.0008293755236081779, + 0.0008834327454678714, + 0.0010062220972031355, + 0.0013563648099079728, + 0.0007929602870717645, + 0.00041823129868134856, + 0.0003113681741524488, + 0.00019538108608685434, + 0.0003479792212601751, + 0.00044731973321177065, + 0.000522337039001286, + 0.00037820529541932046, + 0.00031366682378575206, + 0.0006585608934983611, + 0.00040053139673545957, + 0.00028232362819835544, + 0.0005004118429496884, + 0.00033811412868089974, + 0.0005104378215037286, + 0.0004909004201181233, + 0.0004130482266191393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.007631906773895025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036128980573266745 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.0036128980573266745, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036128980573266745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5155001878738403, + 0.09046722203493118, + 0.04442731291055679, + 0.09725069999694824, + 0.046622056514024734, + 0.042757511138916016, + 0.07936757057905197, + 0.02832244522869587, + 0.04207345470786095, + 0.013211594894528389 + ], + "credit_map": [ + 0.03568146005272865, + 0.017894091084599495, + 0.029236670583486557, + 0.026452315971255302, + 0.024135295301675797, + 0.030924392864108086, + 0.18092778325080872, + 0.1702481210231781, + 0.013168377801775932, + 0.008729842491447926, + 0.011268639005720615, + 0.0075053200125694275, + 0.00969219021499157, + 0.01594739966094494, + 0.01125014666467905, + 0.012905295938253403, + 0.003595205722376704, + 0.004613105207681656, + 0.0050644041039049625, + 0.005907576996833086, + 0.004914325661957264, + 0.006638905964791775, + 0.00678268913179636, + 0.006911095231771469, + 0.008970489725470543, + 0.01004820130765438, + 0.010036149993538857, + 0.012552046217024326, + 0.016384683549404144, + 0.012525811791419983, + 0.014048579148948193, + 0.012684736400842667, + 0.005986015312373638, + 0.0047009289264678955, + 0.008418671786785126, + 0.005259276833385229, + 0.004515404812991619, + 0.00571146234869957, + 0.00571014266461134, + 0.006320151500403881, + 0.004957395605742931, + 0.0045100487768650055, + 0.006094204727560282, + 0.0046464791521430016, + 0.004063407890498638, + 0.00519072450697422, + 0.007459566928446293, + 0.00583568075671792, + 0.007035503163933754, + 0.011560468003153801, + 0.0049004447646439075, + 0.008242755196988583, + 0.009833063930273056, + 0.01574467122554779, + 0.011347129009664059, + 0.010703534819185734, + 0.005183124914765358, + 0.0023010470904409885, + 0.002203070092946291, + 0.003377889283001423, + 0.0025543610099703074, + 0.007163006346672773, + 0.0027489461936056614, + 0.002790997037664056, + 0.005057601258158684, + 0.0036428968887776136, + 0.00276915798895061, + 0.005671523977071047, + 0.004866261500865221, + 0.00835239328444004, + 0.0060766711831092834, + 0.0056369430385529995, + 0.003021067474037409, + 0.0012375486548990011, + 0.0010181268444284797, + 0.0017784358933568, + 0.0013437125599011779, + 0.002130214823409915, + 0.0011774740414693952, + 0.001505014137364924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0036128980573266745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494775250554085 + ], + "opposite_logprob_tokens": [ + -2.929947853088379 + ], + "logprob_reward": -0.05494775250554085, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494775250554085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5591232776641846, + 0.027594318613409996, + 0.050471894443035126, + 0.05076271668076515, + 0.05390559881925583, + 0.0726286917924881, + 0.1268644481897354, + 0.03751372545957565, + 0.0060150814242661, + 0.015120369382202625 + ], + "credit_map": [ + 0.040664542466402054, + 0.022290384396910667, + 0.021536948159337044, + 0.024684639647603035, + 0.02738833986222744, + 0.041661880910396576, + 0.20507125556468964, + 0.17582516372203827, + 0.0032001675572246313, + 0.0036045718006789684, + 0.004077568184584379, + 0.002293242374435067, + 0.0024878617841750383, + 0.004768710117787123, + 0.00395298283547163, + 0.0032092120964080095, + 0.00674473587423563, + 0.004550817422568798, + 0.004684076178818941, + 0.005363198462873697, + 0.004173371009528637, + 0.010932366363704205, + 0.007450805511325598, + 0.0065725212916731834, + 0.005338168237358332, + 0.0044924537651240826, + 0.010873674415051937, + 0.007338084280490875, + 0.003901605261489749, + 0.006174798589199781, + 0.007098005153238773, + 0.005545922555029392, + 0.0059787314385175705, + 0.0069991848431527615, + 0.008003847673535347, + 0.0051358845084905624, + 0.006986377295106649, + 0.0062698666006326675, + 0.008093702606856823, + 0.006437998265028, + 0.008302520029246807, + 0.007681073155254126, + 0.010437487624585629, + 0.008980909362435341, + 0.009319332428276539, + 0.008399559184908867, + 0.009920674376189709, + 0.009587127715349197, + 0.010897916741669178, + 0.016625531017780304, + 0.008893257938325405, + 0.013551036827266216, + 0.014457502402365208, + 0.024000616744160652, + 0.020458707585930824, + 0.017979862168431282, + 0.0068178740330040455, + 0.002375292358919978, + 0.0036660803016275167, + 0.0037558048497885466, + 0.0026841480284929276, + 0.007590361405164003, + 0.004448159132152796, + 0.006175999063998461, + 0.0009273295290768147, + 0.00042012089397758245, + 0.000621616025455296, + 0.0005703220376744866, + 0.0005210794042795897, + 0.0009162573260255158, + 0.000620776554569602, + 0.0014175797114148736, + 0.002609296003356576, + 0.001740272156894207, + 0.001352751743979752, + 0.002124535385519266, + 0.0012861730065196753, + 0.002065539127215743, + 0.0021437399554997683, + 0.0017980597913265228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.05494775250554085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, slowly and steadily tilting upward and panning forward, draws the viewer deep into an ancient and tranquil forest. In the dappled twilight filtering through the dense canopy, thousands of translucent crystal fragments, radiating an inner glow, slowly rise from the cracks in the trees and from the leaf-covered ground like living fireflies. These fragments, shimmering in shades of warm gold, soft pink, and delicate blue, trace graceful paths through the air, drawn by an invisible gravitational force, spiraling toward the center of the forest clearing. As the fragments merge, a vast crystal lotus composed of pure light energy steadily and gracefully unfolds from a tiny core. Each translucent petal is woven from countless tiny points of light, revealing delicate textures and veins, while the center of the pistil radiates the brightest, warmest golden light. The entire image is imbued with the warm, hand-painted quality that is signature to Hayao Miyazaki's animations, with saturated, soft colors and delicate, brushstroke-like outlines. The soft glow of this flower of light casts a dreamy halo on the surrounding moss-covered ancient trees and the ground, creating an atmosphere of extreme tranquility, magic, and healing power.", + "seed": 25651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_4/index_3380_seed_25651_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15594217693433166, + "vlm_logprob": -0.15594217693433166, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.15594217693433166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4345066249370575 + ], + "opposite_logprob_tokens": [ + -1.0595066547393799 + ], + "logprob_reward": -0.4345066249370575, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4345066249370575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07395706325769424, + 0.2113306224346161, + 0.2834623456001282, + 0.13519516587257385, + 0.035726334899663925, + 0.05679658427834511, + 0.06028814613819122, + 0.043455999344587326, + 0.029220636934041977, + 0.07056716829538345 + ], + "credit_map": [ + 0.007507206406444311, + 0.011133584193885326, + 0.007937143556773663, + 0.00840156339108944, + 0.012148695066571236, + 0.010055668652057648, + 0.00801839679479599, + 0.00875480193644762, + 0.01979944109916687, + 0.04864713177084923, + 0.012191605754196644, + 0.012036263011395931, + 0.025449883192777634, + 0.02603308856487274, + 0.012722964398562908, + 0.05445024371147156, + 0.031186457723379135, + 0.052310358732938766, + 0.01677996665239334, + 0.022748148068785667, + 0.029173141345381737, + 0.045475445687770844, + 0.01895230822265148, + 0.06683650612831116, + 0.011087601073086262, + 0.022353395819664, + 0.008417255245149136, + 0.015905875712633133, + 0.03825237601995468, + 0.011474003084003925, + 0.017134709283709526, + 0.010569950565695763, + 0.0034792216029018164, + 0.00451138848438859, + 0.003097957232967019, + 0.006590857170522213, + 0.006913339253515005, + 0.0033514995593577623, + 0.004088270477950573, + 0.003693799953907728, + 0.005054026376456022, + 0.008127822540700436, + 0.007234546821564436, + 0.010226952843368053, + 0.007156377658247948, + 0.0056059835478663445, + 0.008026805706322193, + 0.005364063661545515, + 0.0038382175844162703, + 0.0074194855988025665, + 0.005775371100753546, + 0.007531650364398956, + 0.00921041052788496, + 0.010866090655326843, + 0.006141385063529015, + 0.009505534544587135, + 0.004127781838178635, + 0.0064222426153719425, + 0.003057063790038228, + 0.005999355111271143, + 0.006787976250052452, + 0.006617032922804356, + 0.0056422119960188866, + 0.004802333656698465, + 0.001995712984353304, + 0.0080935750156641, + 0.001605297438800335, + 0.005771208088845015, + 0.0025886220391839743, + 0.0035173557698726654, + 0.003267888445407152, + 0.0023809769190847874, + 0.006165266502648592, + 0.007252340205013752, + 0.007274641655385494, + 0.0107483621686697, + 0.007176184561103582, + 0.012711944989860058, + 0.010770061984658241, + 0.008468367159366608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.4345066249370575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134445548057556 + ], + "opposite_logprob_tokens": [ + -1.3134446144104004 + ], + "logprob_reward": -0.3134445548057556, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134445548057556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09808548539876938, + 0.1205974593758583, + 0.2975355088710785, + 0.10086844116449356, + 0.0631972998380661, + 0.053745612502098083, + 0.07244456559419632, + 0.0230341088026762, + 0.051896627992391586, + 0.11859486252069473 + ], + "credit_map": [ + 0.011355439200997353, + 0.012821289710700512, + 0.012032526545226574, + 0.011501161381602287, + 0.010330072604119778, + 0.013026341795921326, + 0.015171044506132603, + 0.011847607791423798, + 0.01177914533764124, + 0.018706301227211952, + 0.013283716514706612, + 0.010022281669080257, + 0.01896645314991474, + 0.014100891537964344, + 0.016856985166668892, + 0.016881683841347694, + 0.057968027889728546, + 0.024660304188728333, + 0.01613069698214531, + 0.014169339090585709, + 0.022642195224761963, + 0.04208601638674736, + 0.015067515894770622, + 0.1048114225268364, + 0.008586649782955647, + 0.01544744148850441, + 0.008638117462396622, + 0.014163666404783726, + 0.021303653717041016, + 0.009490851312875748, + 0.01387664396315813, + 0.009361420758068562, + 0.00523223914206028, + 0.007897384464740753, + 0.009318139404058456, + 0.011144275777041912, + 0.007026533130556345, + 0.007211365737020969, + 0.007520945742726326, + 0.007846414111554623, + 0.00509409187361598, + 0.007689401041716337, + 0.005845461506396532, + 0.007831706665456295, + 0.007889789529144764, + 0.006654747761785984, + 0.006228402256965637, + 0.006512010935693979, + 0.00473444489762187, + 0.009184571914374828, + 0.005610268097370863, + 0.00966694951057434, + 0.011536701582372189, + 0.013788389042019844, + 0.007438285741955042, + 0.010484956204891205, + 0.0022179281804710627, + 0.0030456078238785267, + 0.0017201974987983704, + 0.0036743946839123964, + 0.003634188324213028, + 0.0029331648256629705, + 0.0035207513719797134, + 0.0022878763265907764, + 0.003825592575594783, + 0.017475254833698273, + 0.0029107006266713142, + 0.011052949354052544, + 0.004322989843785763, + 0.004107141401618719, + 0.005201192572712898, + 0.003000806085765362, + 0.00930505059659481, + 0.012641387060284615, + 0.011226986534893513, + 0.018043993040919304, + 0.010992120951414108, + 0.02206774801015854, + 0.018963968381285667, + 0.015353597700595856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.3134445548057556 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951303523033857 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951303523033857, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951303523033857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14228597283363342, + 0.15523776412010193, + 0.3045177757740021, + 0.09240428358316422, + 0.04395420849323273, + 0.031346410512924194, + 0.08004774898290634, + 0.03655128926038742, + 0.06055883318185806, + 0.053095750510692596 + ], + "credit_map": [ + 0.012662808410823345, + 0.01823475770652294, + 0.01620599254965782, + 0.019163547083735466, + 0.016301173716783524, + 0.01858247071504593, + 0.024421628564596176, + 0.016713595017790794, + 0.01223230641335249, + 0.03428305312991142, + 0.01003329548984766, + 0.008827288635075092, + 0.019506510347127914, + 0.02015935443341732, + 0.014419730752706528, + 0.03577620908617973, + 0.05830466374754906, + 0.03472903370857239, + 0.020983748137950897, + 0.02483321540057659, + 0.03043765015900135, + 0.038527779281139374, + 0.02267380990087986, + 0.0740278884768486, + 0.008608957752585411, + 0.0119948610663414, + 0.009568346664309502, + 0.013855255208909512, + 0.012783411890268326, + 0.00960366427898407, + 0.012918014079332352, + 0.013071774505078793, + 0.003111151745542884, + 0.006646256428211927, + 0.0048529659397900105, + 0.011461597867310047, + 0.006116894539445639, + 0.003159736515954137, + 0.004629980772733688, + 0.003975625615566969, + 0.0027608214877545834, + 0.0031358434353023767, + 0.0033143735490739346, + 0.0046966527588665485, + 0.0045349677093327045, + 0.00470587657764554, + 0.0038939770311117172, + 0.004303898196667433, + 0.005122968461364508, + 0.009536057710647583, + 0.006180379074066877, + 0.010813095606863499, + 0.01325615681707859, + 0.015819482505321503, + 0.008689229376614094, + 0.010630381293594837, + 0.003379402682185173, + 0.005373445805162191, + 0.0036503311712294817, + 0.006294691935181618, + 0.0042055887170135975, + 0.006036314181983471, + 0.0034579974599182606, + 0.004153518937528133, + 0.004261418245732784, + 0.02101162075996399, + 0.0035241148434579372, + 0.01336253434419632, + 0.005075124558061361, + 0.004053695127367973, + 0.006230375729501247, + 0.0030399481765925884, + 0.005733133293688297, + 0.006285598035901785, + 0.0037683609407395124, + 0.006966522894799709, + 0.00522210355848074, + 0.011024938896298409, + 0.00773455249145627, + 0.006360543891787529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.005951303523033857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076280031353235245 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.0076280031353235245, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076280031353235245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12533757090568542, + 0.15758033096790314, + 0.18386982381343842, + 0.18392665684223175, + 0.05191204696893692, + 0.08909203857183456, + 0.06222686171531677, + 0.037254348397254944, + 0.04395646974444389, + 0.06484382599592209 + ], + "credit_map": [ + 0.009691541083157063, + 0.02181878872215748, + 0.014362279325723648, + 0.015283860266208649, + 0.0181211419403553, + 0.01264733262360096, + 0.01752232201397419, + 0.015890315175056458, + 0.018288850784301758, + 0.021285798400640488, + 0.016708407551050186, + 0.014755843207240105, + 0.015698449686169624, + 0.018872899934649467, + 0.026491498574614525, + 0.02547859586775303, + 0.031341083347797394, + 0.04062435030937195, + 0.01085746381431818, + 0.017501838505268097, + 0.021686790511012077, + 0.024737872183322906, + 0.01412095595151186, + 0.022999482229351997, + 0.018988972529768944, + 0.026682980358600616, + 0.012182479724287987, + 0.025348082184791565, + 0.044271066784858704, + 0.021448800340294838, + 0.01868310011923313, + 0.016321182250976562, + 0.004856474697589874, + 0.00673008942976594, + 0.006448694039136171, + 0.011520289815962315, + 0.00561072863638401, + 0.006144657265394926, + 0.005540771409869194, + 0.005060349125415087, + 0.010046535171568394, + 0.009546062909066677, + 0.01057473849505186, + 0.011703522875905037, + 0.008545280434191227, + 0.01881805807352066, + 0.009803134948015213, + 0.010054710321128368, + 0.0050888326950371265, + 0.008335908874869347, + 0.004263944923877716, + 0.01008851733058691, + 0.010331480763852596, + 0.009643735364079475, + 0.007510168012231588, + 0.006964278407394886, + 0.003499454353004694, + 0.004714257083833218, + 0.003190749092027545, + 0.005329668056219816, + 0.004887978080660105, + 0.007748063188046217, + 0.00435855844989419, + 0.003525623120367527, + 0.0030456429813057184, + 0.013938154093921185, + 0.003266911720857024, + 0.008590133860707283, + 0.0041113197803497314, + 0.0035094902850687504, + 0.004792667925357819, + 0.0027021521236747503, + 0.005271805915981531, + 0.007598846685141325, + 0.0060834866017103195, + 0.010815493762493134, + 0.0066490513272583485, + 0.011323521845042706, + 0.009746897965669632, + 0.0073547265492379665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.0076280031353235245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle shot that tilts upward and pans forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent crystal fragments visible rising from the ground and tree cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments shimmer in shades of warm gold, soft pink, and delicate blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vast crystal lotus composed of pure light energy unfold from a tiny core in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a warm, hand-painted quality with saturated colors and delicate, brushstroke-like outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180398270487785 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180398270487785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180398270487785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07082445174455643, + 0.05886853486299515, + 0.23112568259239197, + 0.07923558354377747, + 0.054566726088523865, + 0.06376835703849792, + 0.1295042783021927, + 0.17438435554504395, + 0.10064474493265152, + 0.037077270448207855 + ], + "credit_map": [ + 0.007747537922114134, + 0.00856254156678915, + 0.008196039125323296, + 0.008660062216222286, + 0.008884883485734463, + 0.008300228975713253, + 0.011205552145838737, + 0.009267606772482395, + 0.007437476422637701, + 0.009358575567603111, + 0.006429342553019524, + 0.005491003394126892, + 0.008562523871660233, + 0.007101510651409626, + 0.0066431378945708275, + 0.007844967767596245, + 0.03345510736107826, + 0.05237709730863571, + 0.015417803078889847, + 0.019114267081022263, + 0.03206659108400345, + 0.028866995126008987, + 0.01971190609037876, + 0.03011590801179409, + 0.008068163879215717, + 0.011602778919041157, + 0.005345268175005913, + 0.00942744966596365, + 0.01994379051029682, + 0.0099926283583045, + 0.00849531777203083, + 0.006360180675983429, + 0.003906834404915571, + 0.006589582189917564, + 0.005367683712393045, + 0.010571680963039398, + 0.010465729981660843, + 0.005697326269000769, + 0.007003169506788254, + 0.0049647181294858456, + 0.005527992732822895, + 0.008330133743584156, + 0.005192003212869167, + 0.008088734932243824, + 0.009715038351714611, + 0.0071884384378790855, + 0.010349256917834282, + 0.00937675591558218, + 0.010282802395522594, + 0.01561171468347311, + 0.00854453258216381, + 0.018945446237921715, + 0.023169318214058876, + 0.023432010784745216, + 0.014972777105867863, + 0.01454567164182663, + 0.01775999739766121, + 0.0254525039345026, + 0.013691558502614498, + 0.02523370087146759, + 0.025746017694473267, + 0.025761935859918594, + 0.021197760477662086, + 0.01954088918864727, + 0.008767603896558285, + 0.026977092027664185, + 0.008655356243252754, + 0.021654566749930382, + 0.006566236261278391, + 0.010581630282104015, + 0.009518916718661785, + 0.007923346944153309, + 0.003423477755859494, + 0.004602025728672743, + 0.004472971893846989, + 0.004818125627934933, + 0.004029361996799707, + 0.005984207149595022, + 0.005540401674807072, + 0.004206699784845114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.018180398270487785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_5/index_tensor([7830, 7830])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_5/index_tensor([7830, 7830])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4abb8125ad17d10c53ca302d6864534eebb1abe4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_5/index_tensor([7830, 7830])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, spiraling crane shot begins at a low angle, immersing itself in the crackling, burning undergrowth of the forest floor, before rapidly spinning upward to reveal the silhouette of the massive Terrorbeak against a raging, infernal apocalyptic backdrop. In the style of a highly detailed digital illustration, each element is rendered with crisp clarity; the creature's shadowy plumage is not a flat black, but a complex texture composed of individually defined obsidian-like plumes, catching the infernal backlight along sharp edges. Its head, dotted with several pairs of blazing red eyes, burns like malevolent coals, and its head tilts back as its massive, hooked beak opens. Embers dance and swirl in the superheated air like a chaotic blizzard, illuminating its immense scale within the wall of fire. The creature lets out a wordless, terrifying scream as its throat muscles contort in a purely visual display of power, unleashing visible waves of heat distortion into the shimmering, ash-filled atmosphere, creating a vast and terrifying spectacle.", + "seed": 31947, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_5/index_7830_seed_31947_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033735603420063855, + "vlm_logprob": -0.033735603420063855, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.033735603420063855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14467783272266388 + ], + "opposite_logprob_tokens": [ + -2.0196778774261475 + ], + "logprob_reward": -0.14467783272266388, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14467783272266388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08134519308805466, + 0.06217551231384277, + 0.2141951620578766, + 0.4121864438056946, + 0.06400275975465775, + 0.038127392530441284, + 0.037606943398714066, + 0.027396468445658684, + 0.03262174502015114, + 0.03034232370555401 + ], + "credit_map": [ + 0.00976213812828064, + 0.012505444698035717, + 0.007577352225780487, + 0.010905418545007706, + 0.00837426632642746, + 0.009388022124767303, + 0.013586034998297691, + 0.009246516972780228, + 0.004960232879966497, + 0.00932333990931511, + 0.006056281737983227, + 0.004984652623534203, + 0.018629975616931915, + 0.00629406189545989, + 0.007109614089131355, + 0.004817358683794737, + 0.04026437923312187, + 0.058278024196624756, + 0.020031649619340897, + 0.0207746010273695, + 0.014998632483184338, + 0.01782926917076111, + 0.029881570488214493, + 0.012137031182646751, + 0.03091491386294365, + 0.07165268808603287, + 0.16611577570438385, + 0.053026147186756134, + 0.018294520676136017, + 0.02620292641222477, + 0.018941836431622505, + 0.02703762799501419, + 0.004074611235409975, + 0.004018217790871859, + 0.010983367450535297, + 0.019613562151789665, + 0.003000530879944563, + 0.0029302083421498537, + 0.004224277567118406, + 0.01515798456966877, + 0.00477965222671628, + 0.004082073457539082, + 0.004012478515505791, + 0.0036807297728955746, + 0.006074910517781973, + 0.005337203852832317, + 0.0060786716639995575, + 0.004081673454493284, + 0.004993000999093056, + 0.0036831991747021675, + 0.003225679975003004, + 0.0030934209935367107, + 0.005977289751172066, + 0.008394204080104828, + 0.0036156780552119017, + 0.004624471068382263, + 0.0021116789430379868, + 0.002756178844720125, + 0.002627471461892128, + 0.002850367920473218, + 0.003683855291455984, + 0.005734667647629976, + 0.003308892250061035, + 0.004323357250541449, + 0.004416072275489569, + 0.004459792282432318, + 0.0026286186184734106, + 0.003317941213026643, + 0.006085747387260199, + 0.00448353448882699, + 0.003892612410709262, + 0.003337428206577897, + 0.004493452608585358, + 0.0036411245819181204, + 0.004455971997231245, + 0.0023158739786595106, + 0.006152487825602293, + 0.003388953860849142, + 0.0026417053304612637, + 0.003252755617722869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.14467783272266388 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005968722980469465 + ], + "opposite_logprob_tokens": [ + -5.1309685707092285 + ], + "logprob_reward": -0.005968722980469465, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005968722980469465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07893543690443039, + 0.03601672872900963, + 0.3189104199409485, + 0.21089580655097961, + 0.10934484750032425, + 0.07015024870634079, + 0.033370211720466614, + 0.06363476067781448, + 0.047506183385849, + 0.031235352158546448 + ], + "credit_map": [ + 0.01135221403092146, + 0.017565617337822914, + 0.009022166952490807, + 0.008176418021321297, + 0.007836000993847847, + 0.00711406534537673, + 0.009155930019915104, + 0.008713027462363243, + 0.00485832616686821, + 0.004247967153787613, + 0.003245533676818013, + 0.0033322314266115427, + 0.004388674162328243, + 0.005649462807923555, + 0.0056060124188661575, + 0.0046885209158062935, + 0.0538114532828331, + 0.06386294215917587, + 0.024093259125947952, + 0.026902751997113228, + 0.02150576561689377, + 0.029258497059345245, + 0.07075167447328568, + 0.02872408553957939, + 0.018074780702590942, + 0.031010499224066734, + 0.0826929435133934, + 0.02847922407090664, + 0.011093929409980774, + 0.01380582619458437, + 0.011895238421857357, + 0.013843348249793053, + 0.007983589544892311, + 0.007412993814796209, + 0.018005313351750374, + 0.0324142687022686, + 0.004077174235135317, + 0.008039282634854317, + 0.006598751060664654, + 0.024813475087285042, + 0.009682878851890564, + 0.009779072366654873, + 0.00633030291646719, + 0.006715934257954359, + 0.011303873732686043, + 0.01065046340227127, + 0.008181849494576454, + 0.007505869492888451, + 0.004127153195440769, + 0.00499954167753458, + 0.004208780825138092, + 0.005068851634860039, + 0.003509742673486471, + 0.004906541667878628, + 0.0027945153415203094, + 0.003755084937438369, + 0.006620124448090792, + 0.009428075514733791, + 0.0069213868118822575, + 0.0065955352038145065, + 0.013682716526091099, + 0.009678821079432964, + 0.005232327152043581, + 0.005475769750773907, + 0.004818306770175695, + 0.003102333750575781, + 0.003632465610280633, + 0.002784772776067257, + 0.0038839783519506454, + 0.012308616191148758, + 0.005390486214309931, + 0.01158522441983223, + 0.003914246801286936, + 0.003998741507530212, + 0.003928691148757935, + 0.0033148876391351223, + 0.005721015390008688, + 0.005035765003412962, + 0.002642854582518339, + 0.002679151017218828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.005968722980469465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978435855358839 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.00978435855358839, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978435855358839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08350057899951935, + 0.1335805207490921, + 0.05052882805466652, + 0.0763285905122757, + 0.08583548665046692, + 0.14688141644001007, + 0.12248499691486359, + 0.12418980151414871, + 0.08352641761302948, + 0.09314335882663727 + ], + "credit_map": [ + 0.008967249654233456, + 0.018857644870877266, + 0.009323608130216599, + 0.010885011404752731, + 0.008682407438755035, + 0.010174810886383057, + 0.009301682934165001, + 0.007308159954845905, + 0.012935747392475605, + 0.021981244906783104, + 0.009995275177061558, + 0.013197699561715126, + 0.020858095958828926, + 0.01926254667341709, + 0.022818109020590782, + 0.012531799264252186, + 0.010529885068535805, + 0.0116481464356184, + 0.005389458034187555, + 0.0038428374100476503, + 0.0037091649137437344, + 0.004294825252145529, + 0.007397843524813652, + 0.0037166653200984, + 0.007674302440136671, + 0.009104447439312935, + 0.0150713250041008, + 0.012828219681978226, + 0.007623331621289253, + 0.010317199863493443, + 0.007263480219990015, + 0.006446287035942078, + 0.00961607601493597, + 0.008800292387604713, + 0.015050515532493591, + 0.013537172228097916, + 0.007850146852433681, + 0.01191153284162283, + 0.009652996435761452, + 0.009416753426194191, + 0.014258279465138912, + 0.02553538605570793, + 0.01861339807510376, + 0.013130128383636475, + 0.01744973286986351, + 0.020769689232110977, + 0.02006666362285614, + 0.017058130353689194, + 0.015114568173885345, + 0.014735013246536255, + 0.017384236678481102, + 0.013423236086964607, + 0.015462050214409828, + 0.021833354607224464, + 0.01006144005805254, + 0.014471101574599743, + 0.01388942264020443, + 0.016755830496549606, + 0.01216053869575262, + 0.013693866319954395, + 0.0244621392339468, + 0.017911028116941452, + 0.012449225410819054, + 0.012867755256593227, + 0.011882364749908447, + 0.008255579508841038, + 0.008651764132082462, + 0.010581741109490395, + 0.009123492985963821, + 0.016719909384846687, + 0.007367251440882683, + 0.010944318026304245, + 0.008373475633561611, + 0.014117840677499771, + 0.010800562798976898, + 0.009742354974150658, + 0.021234510466456413, + 0.014581073075532913, + 0.006987245287746191, + 0.007306301966309547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.00978435855358839 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673524197191 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.00673524197191, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673524197191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12089399993419647, + 0.15895788371562958, + 0.03360779210925102, + 0.0613703727722168, + 0.14898937940597534, + 0.0920855775475502, + 0.07691443711519241, + 0.0968756228685379, + 0.065266452729702, + 0.14503848552703857 + ], + "credit_map": [ + 0.018012266606092453, + 0.030001280829310417, + 0.013430994004011154, + 0.015122022479772568, + 0.011143618263304234, + 0.00843749474734068, + 0.01218405645340681, + 0.012562270276248455, + 0.014597734436392784, + 0.01901494711637497, + 0.013595739379525185, + 0.017252828925848007, + 0.022490408271551132, + 0.024168336763978004, + 0.030118120834231377, + 0.017719758674502373, + 0.005404999013990164, + 0.004576933104544878, + 0.00416817469522357, + 0.00475380988791585, + 0.003167342161759734, + 0.0031560570932924747, + 0.0050302245654165745, + 0.0033502550795674324, + 0.005583216901868582, + 0.007231937721371651, + 0.016932494938373566, + 0.0074196127243340015, + 0.004084370564669371, + 0.006819559261202812, + 0.004555856343358755, + 0.008743326179683208, + 0.010586823336780071, + 0.011256037279963493, + 0.02415521629154682, + 0.042344555258750916, + 0.00660948920994997, + 0.01393247302621603, + 0.009050018154084682, + 0.031054764986038208, + 0.017138943076133728, + 0.010733159258961678, + 0.006461175158619881, + 0.005927490536123514, + 0.011487570591270924, + 0.016313141211867332, + 0.012656371109187603, + 0.01136772334575653, + 0.011743823997676373, + 0.009400474838912487, + 0.009988832287490368, + 0.007954145781695843, + 0.008480203337967396, + 0.012027780525386333, + 0.008412504568696022, + 0.00890667550265789, + 0.008948493748903275, + 0.011646847240626812, + 0.0077671753242611885, + 0.007836742326617241, + 0.016560109332203865, + 0.018899671733379364, + 0.008571101352572441, + 0.016645487397909164, + 0.006360014900565147, + 0.003453803714364767, + 0.002863011322915554, + 0.00305065861903131, + 0.004950590431690216, + 0.017264649271965027, + 0.009094461798667908, + 0.01822926290333271, + 0.01593869738280773, + 0.018950389698147774, + 0.01693265326321125, + 0.01501499768346548, + 0.03465054929256439, + 0.0206377562135458, + 0.01133286114782095, + 0.011580577120184898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.00673524197191 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015118608716875315 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015118608716875315, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015118608716875315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07570090889930725, + 0.08206207305192947, + 0.355269193649292, + 0.16973531246185303, + 0.10575821995735168, + 0.025026259943842888, + 0.04923334717750549, + 0.03665309026837349, + 0.0698409453034401, + 0.030720675364136696 + ], + "credit_map": [ + 0.00855712778866291, + 0.010817757807672024, + 0.008933699689805508, + 0.007352760992944241, + 0.008192839100956917, + 0.01207063253968954, + 0.009477199986577034, + 0.0102988937869668, + 0.00949781946837902, + 0.015006277710199356, + 0.007945137098431587, + 0.009369372390210629, + 0.010706299915909767, + 0.008977043442428112, + 0.011773188598453999, + 0.008786933496594429, + 0.08097727596759796, + 0.10076530277729034, + 0.02326568216085434, + 0.03181220218539238, + 0.023270782083272934, + 0.02429140731692314, + 0.04923825338482857, + 0.021648257970809937, + 0.014531325548887253, + 0.026547852903604507, + 0.06364746391773224, + 0.020688433200120926, + 0.007926292717456818, + 0.011476758867502213, + 0.008401074446737766, + 0.016516102477908134, + 0.010028956457972527, + 0.008961821906268597, + 0.01677481271326542, + 0.023914895951747894, + 0.006177943665534258, + 0.010686446912586689, + 0.009546912275254726, + 0.019666427746415138, + 0.002897492377087474, + 0.003290607826784253, + 0.0027451671194285154, + 0.0022887431550771, + 0.003985947463661432, + 0.004142714664340019, + 0.0030275993049144745, + 0.0026479882653802633, + 0.005853882525116205, + 0.003584975842386484, + 0.004877117928117514, + 0.00427550682798028, + 0.004929869435727596, + 0.01354218926280737, + 0.004953937605023384, + 0.00721586961299181, + 0.004020372871309519, + 0.0050766183994710445, + 0.004545860458165407, + 0.00301319919526577, + 0.006798972375690937, + 0.005348941311240196, + 0.0034791752696037292, + 0.004369952250272036, + 0.007824198342859745, + 0.009249935857951641, + 0.005523358471691608, + 0.006644516251981258, + 0.007450114469975233, + 0.015056042931973934, + 0.005964021664112806, + 0.012128762900829315, + 0.004213045816868544, + 0.004149946849793196, + 0.004180512856692076, + 0.0021710724104195833, + 0.004588783718645573, + 0.004366139881312847, + 0.0027017618995159864, + 0.004349410068243742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.0015118608716875315 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, spiraling crane shot begins at a low angle, immersing itself in the crackling, burning undergrowth of the forest floor, before rapidly spinning upward to reveal the silhouette of the massive Terrorbeak against a raging, infernal apocalyptic backdrop. In the style of a highly detailed digital illustration, each element is rendered with crisp clarity; the creature's shadowy plumage is not a flat black, but a complex texture composed of individually defined obsidian-like plumes, catching the infernal backlight along sharp edges. Its head, dotted with several pairs of blazing red eyes, burns like malevolent coals, and its head tilts back as its massive, hooked beak opens. Embers dance and swirl in the superheated air like a chaotic blizzard, illuminating its immense scale within the wall of fire. The creature lets out a wordless, terrifying scream as its throat muscles contort in a purely visual display of power, unleashing visible waves of heat distortion into the shimmering, ash-filled atmosphere, creating a vast and terrifying spectacle.", + "seed": 31948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_5/index_7830_seed_31948_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.005040518147870899, + "vlm_logprob": -0.005040518147870899, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.005040518147870899, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00729646161198616 + ], + "opposite_logprob_tokens": [ + -5.132296562194824 + ], + "logprob_reward": -0.00729646161198616, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00729646161198616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21443457901477814, + 0.10866604745388031, + 0.05833565443754196, + 0.18840481340885162, + 0.103835828602314, + 0.08216845244169235, + 0.08578536659479141, + 0.0561571829020977, + 0.04299706965684891, + 0.05921496823430061 + ], + "credit_map": [ + 0.022692278027534485, + 0.016778122633695602, + 0.038811568170785904, + 0.028032803907990456, + 0.02952677384018898, + 0.02081868425011635, + 0.03029477782547474, + 0.027479566633701324, + 0.008674138225615025, + 0.006014432292431593, + 0.01874891109764576, + 0.017254378646612167, + 0.007670300547033548, + 0.007815416902303696, + 0.016874393448233604, + 0.02561408095061779, + 0.008410179987549782, + 0.007361777126789093, + 0.007743827998638153, + 0.006820290349423885, + 0.00585744995623827, + 0.006219572387635708, + 0.006202744320034981, + 0.009719815105199814, + 0.031175196170806885, + 0.021196719259023666, + 0.026225954294204712, + 0.02352837286889553, + 0.0205848291516304, + 0.027617167681455612, + 0.021271027624607086, + 0.016805538907647133, + 0.012833114713430405, + 0.015215283259749413, + 0.010803092271089554, + 0.011239307001233101, + 0.01803305558860302, + 0.013434640131890774, + 0.009809623472392559, + 0.012467710301280022, + 0.00441142451018095, + 0.008921915665268898, + 0.005631971172988415, + 0.006029428448528051, + 0.006994590163230896, + 0.022548427805304527, + 0.0045539503917098045, + 0.023076748475432396, + 0.0071420371532440186, + 0.01006749365478754, + 0.006517034024000168, + 0.006203798111528158, + 0.007459550630301237, + 0.023463068529963493, + 0.006667090114206076, + 0.018265293911099434, + 0.010382686741650105, + 0.006018460262566805, + 0.003594781272113323, + 0.008827444165945053, + 0.004433815833181143, + 0.014373907819390297, + 0.004208102356642485, + 0.004317987244576216, + 0.007557358592748642, + 0.007005288731306791, + 0.00391573878005147, + 0.005482104606926441, + 0.00737459771335125, + 0.005816569086164236, + 0.0033951587975025177, + 0.002450257772579789, + 0.009590182453393936, + 0.004335666075348854, + 0.008243857882916927, + 0.0058534215204417706, + 0.006261902395635843, + 0.011695760302245617, + 0.006598804611712694, + 0.0066353739239275455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.00729646161198616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028351848013699055 + ], + "opposite_logprob_tokens": [ + -5.877835273742676 + ], + "logprob_reward": -0.0028351848013699055, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028351848013699055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13381095230579376, + 0.05379951372742653, + 0.047213904559612274, + 0.05375257506966591, + 0.05883653834462166, + 0.06421922892332077, + 0.19211427867412567, + 0.07575354725122452, + 0.08444902300834656, + 0.23605044186115265 + ], + "credit_map": [ + 0.012560099363327026, + 0.010241677053272724, + 0.022314831614494324, + 0.023361245170235634, + 0.013930209912359715, + 0.013047107495367527, + 0.020484043285250664, + 0.017871737480163574, + 0.0053783287294209, + 0.006338535342365503, + 0.007025078870356083, + 0.0067436727695167065, + 0.009293858893215656, + 0.0053323316387832165, + 0.005884739104658365, + 0.007802967447787523, + 0.004411397501826286, + 0.006574939005076885, + 0.005675788037478924, + 0.0069343820214271545, + 0.005374243017286062, + 0.0055976021103560925, + 0.005116095766425133, + 0.007529459893703461, + 0.007615421898663044, + 0.009997071698307991, + 0.0067131295800209045, + 0.008316028863191605, + 0.006060430314391851, + 0.005227337125688791, + 0.0054498412646353245, + 0.004373316187411547, + 0.006655137985944748, + 0.006225680001080036, + 0.005633245687931776, + 0.013643861748278141, + 0.008040168322622776, + 0.00824780110269785, + 0.0062261116690933704, + 0.0041645304299890995, + 0.005007797386497259, + 0.004329781047999859, + 0.005070909857749939, + 0.00452419463545084, + 0.0038421167992055416, + 0.019628815352916718, + 0.005184592679142952, + 0.01663101837038994, + 0.024097835645079613, + 0.012132897041738033, + 0.00992873590439558, + 0.01003300677984953, + 0.01113953161984682, + 0.06826651841402054, + 0.010182943195104599, + 0.04633280262351036, + 0.01748070865869522, + 0.005415545776486397, + 0.005201822612434626, + 0.007969207130372524, + 0.005148961208760738, + 0.02067977748811245, + 0.007042651530355215, + 0.00681487238034606, + 0.018416326493024826, + 0.011944976635277271, + 0.008241146802902222, + 0.008133705705404282, + 0.013458299450576305, + 0.0111269261687994, + 0.007525070104748011, + 0.005602569319307804, + 0.03843889757990837, + 0.010633041150867939, + 0.043826427310705185, + 0.01634260080754757, + 0.012542537413537502, + 0.06009242683649063, + 0.013978335075080395, + 0.040196191519498825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0028351848013699055 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640995249152184 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640995249152184, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640995249152184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11208674311637878, + 0.05318243056535721, + 0.039176732301712036, + 0.11581839621067047, + 0.0746142789721489, + 0.06052285432815552, + 0.20776717364788055, + 0.10774020105600357, + 0.12536519765853882, + 0.10372599214315414 + ], + "credit_map": [ + 0.011089853011071682, + 0.013970619067549706, + 0.015962863340973854, + 0.012289566919207573, + 0.018372338265180588, + 0.015326151624321938, + 0.011135084554553032, + 0.013940269127488136, + 0.005887368693947792, + 0.008454808965325356, + 0.0055268146097660065, + 0.0056284102611243725, + 0.009396308101713657, + 0.005735359620302916, + 0.008437969721853733, + 0.004115396644920111, + 0.003238550154492259, + 0.004151978064328432, + 0.003427039133384824, + 0.006460840813815594, + 0.0026660799048841, + 0.0061987945809960365, + 0.005464506801217794, + 0.007568945176899433, + 0.01974075846374035, + 0.0236513614654541, + 0.009256811812520027, + 0.016739163547754288, + 0.014789151027798653, + 0.013546599075198174, + 0.008205034770071507, + 0.009889530949294567, + 0.012647903524339199, + 0.012699499726295471, + 0.004980625584721565, + 0.007646495476365089, + 0.015093114227056503, + 0.010549628175795078, + 0.004686134401708841, + 0.006310889031738043, + 0.006816316395998001, + 0.008228589780628681, + 0.003944250289350748, + 0.006736511830240488, + 0.0059234811924397945, + 0.012761233374476433, + 0.0032422321382910013, + 0.01287024561315775, + 0.0237518809735775, + 0.01901957020163536, + 0.011641982942819595, + 0.012462492100894451, + 0.015588001348078251, + 0.0665799155831337, + 0.013098972849547863, + 0.045624371618032455, + 0.022043850272893906, + 0.010653901845216751, + 0.004689376335591078, + 0.008811012841761112, + 0.006547397468239069, + 0.036911170929670334, + 0.004777985624969006, + 0.013305512256920338, + 0.040591977536678314, + 0.012862811796367168, + 0.01120702177286148, + 0.010883309878408909, + 0.012463873252272606, + 0.018363872542977333, + 0.011006031185388565, + 0.007986312732100487, + 0.012239636853337288, + 0.011021521873772144, + 0.015478263609111309, + 0.00879065040498972, + 0.01616871915757656, + 0.016452161595225334, + 0.011905435472726822, + 0.011669620871543884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.008640995249152184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250830203294754 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005250830203294754, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250830203294754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1211070865392685, + 0.04331371933221817, + 0.05117110162973404, + 0.09380389004945755, + 0.05575047805905342, + 0.08050649613142014, + 0.1163204237818718, + 0.13836443424224854, + 0.1174202412366867, + 0.18224212527275085 + ], + "credit_map": [ + 0.011928570456802845, + 0.014065778814256191, + 0.02018633484840393, + 0.01548541709780693, + 0.018453188240528107, + 0.011677943170070648, + 0.014985657297074795, + 0.01432418916374445, + 0.004278238397091627, + 0.0033503989689052105, + 0.005872596055269241, + 0.00654236227273941, + 0.00419626897200942, + 0.0035613214131444693, + 0.009447567164897919, + 0.006064961664378643, + 0.005001498851925135, + 0.0043354094959795475, + 0.014651454985141754, + 0.007582400925457478, + 0.0029343541245907545, + 0.0038483627140522003, + 0.005968536250293255, + 0.006849079392850399, + 0.013725685887038708, + 0.012950243428349495, + 0.012338905595242977, + 0.010845646262168884, + 0.008466039784252644, + 0.013725543394684792, + 0.01113042701035738, + 0.010621392168104649, + 0.008015654981136322, + 0.008750571869313717, + 0.004801094997674227, + 0.007273867726325989, + 0.010831368155777454, + 0.0067923059687018394, + 0.004584412090480328, + 0.004701201803982258, + 0.007284447085112333, + 0.01237922441214323, + 0.008357241749763489, + 0.012753238901495934, + 0.013463223353028297, + 0.009528388269245625, + 0.008712639100849628, + 0.008028088137507439, + 0.010193546302616596, + 0.00843062810599804, + 0.00943036749958992, + 0.009470601566135883, + 0.00783584825694561, + 0.03161395341157913, + 0.011086951941251755, + 0.028258521109819412, + 0.02077408879995346, + 0.017980920150876045, + 0.008529242128133774, + 0.02467680163681507, + 0.010328729636967182, + 0.02782173827290535, + 0.016874952241778374, + 0.011377956718206406, + 0.02907397784292698, + 0.018317848443984985, + 0.009369803592562675, + 0.011290140450000763, + 0.02006387710571289, + 0.016894835978746414, + 0.007023460231721401, + 0.0053862882778048515, + 0.02992132678627968, + 0.018260709941387177, + 0.021942906081676483, + 0.01392285991460085, + 0.015325279906392097, + 0.0447823740541935, + 0.0142072644084692, + 0.0238793957978487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.005250830203294754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001179118873551488 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.001179118873551488, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001179118873551488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17780284583568573, + 0.06517362594604492, + 0.06316140294075012, + 0.05054538697004318, + 0.06031319871544838, + 0.09528800845146179, + 0.15553687512874603, + 0.0890008807182312, + 0.021832959726452827, + 0.22134478390216827 + ], + "credit_map": [ + 0.02084384486079216, + 0.015238214284181595, + 0.030002739280462265, + 0.025424322113394737, + 0.02399115078151226, + 0.015350143425166607, + 0.021918777376413345, + 0.025033656507730484, + 0.00496240658685565, + 0.005827170331031084, + 0.011711877770721912, + 0.009975917637348175, + 0.006409518886357546, + 0.004794278647750616, + 0.010568289086222649, + 0.010924169793725014, + 0.005736551247537136, + 0.014898165129125118, + 0.005992877297103405, + 0.00965123437345028, + 0.0066717141307890415, + 0.006850170437246561, + 0.006396948359906673, + 0.006963747087866068, + 0.005068302154541016, + 0.006091012619435787, + 0.008746401406824589, + 0.007896161638200283, + 0.005531857255846262, + 0.006394351366907358, + 0.0056738839484751225, + 0.005143416114151478, + 0.009446614421904087, + 0.006744095589965582, + 0.0048398347571492195, + 0.01710093766450882, + 0.0074743605218827724, + 0.007065969053655863, + 0.003908779937773943, + 0.0037326086312532425, + 0.008307360112667084, + 0.006998047232627869, + 0.006764239631593227, + 0.008524013683199883, + 0.008567150682210922, + 0.024968691170215607, + 0.0056887525133788586, + 0.02546975202858448, + 0.0156532172113657, + 0.009448474273085594, + 0.016818026080727577, + 0.014964264817535877, + 0.009074973873794079, + 0.03849659860134125, + 0.020526841282844543, + 0.030554480850696564, + 0.01932898350059986, + 0.006167637184262276, + 0.0043820166029036045, + 0.006888391450047493, + 0.005347015336155891, + 0.03129281476140022, + 0.004802820738404989, + 0.010791203007102013, + 0.00584798026829958, + 0.001822659862227738, + 0.002036906545981765, + 0.0031297975219786167, + 0.002101570600643754, + 0.0030696382746100426, + 0.002191480016335845, + 0.0016329274512827396, + 0.040022049099206924, + 0.008607421070337296, + 0.04110635071992874, + 0.015534372068941593, + 0.007970674894750118, + 0.06075553968548775, + 0.009477118961513042, + 0.03787125647068024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.001179118873551488 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, spiraling crane shot begins at a low angle, immersing itself in the crackling, burning undergrowth of the forest floor, before rapidly spinning upward to reveal the silhouette of the massive Terrorbeak against a raging, infernal apocalyptic backdrop. In the style of a highly detailed digital illustration, each element is rendered with crisp clarity; the creature's shadowy plumage is not a flat black, but a complex texture composed of individually defined obsidian-like plumes, catching the infernal backlight along sharp edges. Its head, dotted with several pairs of blazing red eyes, burns like malevolent coals, and its head tilts back as its massive, hooked beak opens. Embers dance and swirl in the superheated air like a chaotic blizzard, illuminating its immense scale within the wall of fire. The creature lets out a wordless, terrifying scream as its throat muscles contort in a purely visual display of power, unleashing visible waves of heat distortion into the shimmering, ash-filled atmosphere, creating a vast and terrifying spectacle.", + "seed": 31949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_5/index_7830_seed_31949_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.032850328646600244, + "vlm_logprob": -0.032850328646600244, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.032850328646600244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12819185853004456 + ], + "opposite_logprob_tokens": [ + -2.1281919479370117 + ], + "logprob_reward": -0.12819185853004456, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12819185853004456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6828150749206543, + 0.035170335322618484, + 0.016420284286141396, + 0.02321198210120201, + 0.020157581195235252, + 0.012986868619918823, + 0.04749802500009537, + 0.07482670247554779, + 0.0753418579697609, + 0.01157123502343893 + ], + "credit_map": [ + 0.10902270674705505, + 0.07876549661159515, + 0.10901536792516708, + 0.05137239396572113, + 0.04263022541999817, + 0.19919797778129578, + 0.04410684108734131, + 0.04870409145951271, + 0.003292870707809925, + 0.0032782491762191057, + 0.010019619017839432, + 0.005031641107052565, + 0.003949849866330624, + 0.0035530172754079103, + 0.002417386043816805, + 0.003627703059464693, + 0.001847525592893362, + 0.0019570705480873585, + 0.002027014968916774, + 0.0026002246886491776, + 0.002399750053882599, + 0.0019772383384406567, + 0.0019343281164765358, + 0.0016771337250247598, + 0.00221789488568902, + 0.0032301603350788355, + 0.002671721624210477, + 0.004295666702091694, + 0.0027774758636951447, + 0.0023145382292568684, + 0.003331490559503436, + 0.002373036928474903, + 0.0033889063633978367, + 0.0024912157095968723, + 0.002635582583025098, + 0.003206738969311118, + 0.0018338083755224943, + 0.002157951006665826, + 0.0020995535887777805, + 0.002343826461583376, + 0.0017981037963181734, + 0.0011642659083008766, + 0.001806863583624363, + 0.002246291609480977, + 0.001258110860362649, + 0.0017841705121099949, + 0.0019516523461788893, + 0.0009774109348654747, + 0.006462265271693468, + 0.0030373765621334314, + 0.0036735853645950556, + 0.007939946837723255, + 0.0038022473454475403, + 0.010071737691760063, + 0.0032637235708534718, + 0.009247149340808392, + 0.022443190217018127, + 0.006108283065259457, + 0.0062469900585711, + 0.012486700899899006, + 0.006339625921100378, + 0.009311060421168804, + 0.005731751210987568, + 0.0061591085977852345, + 0.007812952622771263, + 0.005224114749580622, + 0.004774922970682383, + 0.01677241176366806, + 0.005280155222862959, + 0.02542811632156372, + 0.0053007290698587894, + 0.0047484636306762695, + 0.0014339589979499578, + 0.0012309536105021834, + 0.00127416814211756, + 0.0020267621148377657, + 0.0011169026838615537, + 0.0018848193576559424, + 0.0013608218869194388, + 0.001242849393747747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.12819185853004456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006754186935722828 + ], + "opposite_logprob_tokens": [ + -5.006754398345947 + ], + "logprob_reward": -0.006754186935722828, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006754186935722828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20873379707336426, + 0.08132278174161911, + 0.02373959869146347, + 0.05662883073091507, + 0.10957220196723938, + 0.06537871062755585, + 0.039268895983695984, + 0.03545680269598961, + 0.35301706194877625, + 0.026881355792284012 + ], + "credit_map": [ + 0.012446089647710323, + 0.06824222207069397, + 0.022993309423327446, + 0.015742408111691475, + 0.01601598411798477, + 0.04596923291683197, + 0.016868365928530693, + 0.010456169955432415, + 0.005972645711153746, + 0.019369713962078094, + 0.004297497682273388, + 0.005957083776593208, + 0.025833528488874435, + 0.006571317557245493, + 0.006957370322197676, + 0.0063636149279773235, + 0.0023234994150698185, + 0.005385772790759802, + 0.0018976933788508177, + 0.0030854535289108753, + 0.004253737162798643, + 0.00173585070297122, + 0.0024224105291068554, + 0.0026351772248744965, + 0.005709311459213495, + 0.00925921555608511, + 0.006705336272716522, + 0.00891900435090065, + 0.009185681119561195, + 0.005675342865288258, + 0.0053916210308671, + 0.005783310160040855, + 0.012861542403697968, + 0.015073774382472038, + 0.01737564615905285, + 0.016319528222084045, + 0.0144997239112854, + 0.012750542722642422, + 0.012637801468372345, + 0.008053625002503395, + 0.005846615880727768, + 0.0034686941653490067, + 0.017899280413985252, + 0.016179420053958893, + 0.004813179839402437, + 0.005752826575189829, + 0.00782188680022955, + 0.0035968010779470205, + 0.004593525547534227, + 0.002689545974135399, + 0.0028455164283514023, + 0.007317426148802042, + 0.0028584448155015707, + 0.012271231971681118, + 0.003770106704905629, + 0.0029230937361717224, + 0.004856836516410112, + 0.003044610610231757, + 0.0031477026641368866, + 0.006551042664796114, + 0.005034285597503185, + 0.005898989737033844, + 0.003272855654358864, + 0.0036504785530269146, + 0.033197373151779175, + 0.026349712163209915, + 0.020504241809248924, + 0.07852593809366226, + 0.025724057108163834, + 0.12424873560667038, + 0.02320469170808792, + 0.02126230113208294, + 0.0039013982750475407, + 0.003959403373301029, + 0.004236673470586538, + 0.004186364822089672, + 0.0022671055048704147, + 0.0030676890164613724, + 0.0027415468357503414, + 0.002521170536056161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.006754186935722828 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818285696208477 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818285696208477, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818285696208477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15424105525016785, + 0.036070045083761215, + 0.02062021940946579, + 0.05941881984472275, + 0.21078260242938995, + 0.06512060016393661, + 0.062249768525362015, + 0.2398640662431717, + 0.08967031538486481, + 0.061962444335222244 + ], + "credit_map": [ + 0.029470110312104225, + 0.01545457262545824, + 0.010422855615615845, + 0.007670827209949493, + 0.013975966721773148, + 0.05441237613558769, + 0.011347370222210884, + 0.011486979201436043, + 0.002753313397988677, + 0.007968094199895859, + 0.004386219196021557, + 0.0021786908619105816, + 0.010596430860459805, + 0.002732510445639491, + 0.0022341753356158733, + 0.003220608690753579, + 0.0021176135633140802, + 0.0035177881363779306, + 0.0020838610362261534, + 0.0024841518606990576, + 0.0034143913071602583, + 0.0022721176501363516, + 0.0023154495283961296, + 0.0024148463271558285, + 0.005990181118249893, + 0.010103270411491394, + 0.0061907218769192696, + 0.009303370490670204, + 0.00873990822583437, + 0.005731081124395132, + 0.007243001833558083, + 0.006117281503975391, + 0.033958036452531815, + 0.026542238891124725, + 0.03596355766057968, + 0.0333499051630497, + 0.021987663581967354, + 0.024136770516633987, + 0.022492798045277596, + 0.012351639568805695, + 0.007025294005870819, + 0.005685041192919016, + 0.010499083437025547, + 0.010787892155349255, + 0.005378433968871832, + 0.00731063773855567, + 0.008246393874287605, + 0.01018782239407301, + 0.008205829188227654, + 0.00553902518004179, + 0.005529386457055807, + 0.008357876911759377, + 0.005194840952754021, + 0.010792375542223454, + 0.008569411933422089, + 0.010061021894216537, + 0.06711463630199432, + 0.019620578736066818, + 0.021214885637164116, + 0.03970176354050636, + 0.023930877447128296, + 0.03260521590709686, + 0.018364327028393745, + 0.017311779782176018, + 0.009371815249323845, + 0.006134144961833954, + 0.00476068165153265, + 0.018890686333179474, + 0.00670014601200819, + 0.03257519379258156, + 0.005500343162566423, + 0.005737309344112873, + 0.008856172673404217, + 0.007063029333949089, + 0.01092239934951067, + 0.009320403449237347, + 0.00503906887024641, + 0.011970004066824913, + 0.005142986308783293, + 0.003648380283266306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.01818285696208477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636267855763435 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636267855763435, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636267855763435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053992919623851776, + 0.031614162027835846, + 0.005822774954140186, + 0.012469039298593998, + 0.02064201980829239, + 0.04013299196958542, + 0.04539669305086136, + 0.04942185431718826, + 0.6801403760910034, + 0.06036717817187309 + ], + "credit_map": [ + 0.010868343524634838, + 0.00405708746984601, + 0.006467247381806374, + 0.0041885701939463615, + 0.004581415094435215, + 0.011171560734510422, + 0.007336119655519724, + 0.005322571843862534, + 0.0035105461720377207, + 0.0033459076657891273, + 0.009149864315986633, + 0.0036951135843992233, + 0.004694984294474125, + 0.00251428852789104, + 0.0021755925845354795, + 0.0025278653483837843, + 0.0006345831789076328, + 0.0010337791172787547, + 0.0007634953362867236, + 0.0005590053042396903, + 0.0012326983269304037, + 0.0005687669035978615, + 0.0005593631649389863, + 0.0004710840876214206, + 0.0012731761671602726, + 0.001637364155612886, + 0.0015897626290097833, + 0.002444311510771513, + 0.0011477337684482336, + 0.0010687330504879355, + 0.001773857162334025, + 0.0015341013204306364, + 0.003068895312026143, + 0.0027880712877959013, + 0.003173465607687831, + 0.002964240964502096, + 0.0024116213899105787, + 0.0016670114127919078, + 0.002501274459064007, + 0.0020674406550824642, + 0.00506999297067523, + 0.0031795830000191927, + 0.00801658071577549, + 0.008159317076206207, + 0.002873563440516591, + 0.004379577934741974, + 0.0057576498948037624, + 0.0026967281010001898, + 0.006788159720599651, + 0.002131979214027524, + 0.002991798333823681, + 0.006801753304898739, + 0.0034541285131126642, + 0.01203175913542509, + 0.004098495468497276, + 0.007098618429154158, + 0.015046991407871246, + 0.0038045006804168224, + 0.0038714429829269648, + 0.007942330092191696, + 0.004490313585847616, + 0.0066566127352416515, + 0.0037320840638130903, + 0.003877576207742095, + 0.060886599123477936, + 0.05043741315603256, + 0.03511866554617882, + 0.14745254814624786, + 0.052467431873083115, + 0.24722206592559814, + 0.04566681385040283, + 0.040888845920562744, + 0.008844250813126564, + 0.010363641194999218, + 0.007111290469765663, + 0.00933300144970417, + 0.0052775247022509575, + 0.008006248623132706, + 0.006434164941310883, + 0.004997059237211943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.008636267855763435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486472949385643 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.002486472949385643, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486472949385643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23977161943912506, + 0.053132228553295135, + 0.02129022404551506, + 0.0314570851624012, + 0.2096196711063385, + 0.08613967895507812, + 0.09333691745996475, + 0.07805022597312927, + 0.16995418071746826, + 0.017248190939426422 + ], + "credit_map": [ + 0.04067397117614746, + 0.048223163932561874, + 0.017440302297472954, + 0.010280454531311989, + 0.01322186179459095, + 0.08322617411613464, + 0.011288476176559925, + 0.015417216345667839, + 0.00451504485681653, + 0.01096905954182148, + 0.004940069280564785, + 0.003397747641429305, + 0.01581455208361149, + 0.004015558399260044, + 0.0029540846589952707, + 0.0065261139534413815, + 0.002164636505767703, + 0.0024391314946115017, + 0.0021740563679486513, + 0.0023277117870748043, + 0.003427712945267558, + 0.0029372258577495813, + 0.0022814252879470587, + 0.00353832496330142, + 0.003366859396919608, + 0.004507619421929121, + 0.004140270408242941, + 0.004459083545953035, + 0.003998511470854282, + 0.0025791465304791927, + 0.004009482450783253, + 0.0043961117044091225, + 0.022150767967104912, + 0.024658067151904106, + 0.03150441497564316, + 0.034290749579668045, + 0.024045132100582123, + 0.02874981239438057, + 0.025636261329054832, + 0.01858445815742016, + 0.009821244515478611, + 0.006543860770761967, + 0.02000420354306698, + 0.0130604337900877, + 0.00823955237865448, + 0.007139347493648529, + 0.012663685716688633, + 0.008667347021400928, + 0.010726931504905224, + 0.0061958846636116505, + 0.007640764117240906, + 0.016148190945386887, + 0.008432728238403797, + 0.0257552620023489, + 0.009678312577307224, + 0.008758844807744026, + 0.01821637712419033, + 0.007241027895361185, + 0.00823884829878807, + 0.012526918202638626, + 0.007216797675937414, + 0.008347805589437485, + 0.008222764357924461, + 0.0080396868288517, + 0.016240864992141724, + 0.011572065763175488, + 0.010641882196068764, + 0.037850454449653625, + 0.012282988987863064, + 0.05997742339968681, + 0.01101013831794262, + 0.010378353297710419, + 0.0027850971091538668, + 0.0020345200318843126, + 0.0018939842702820897, + 0.002551164710894227, + 0.001216977834701538, + 0.0034440848976373672, + 0.0014234203845262527, + 0.001898941001854837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.002486472949385643 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, spiraling crane shot begins at a low angle, immersing itself in the crackling, burning undergrowth of the forest floor, before rapidly spinning upward to reveal the silhouette of the massive Terrorbeak against a raging, infernal apocalyptic backdrop. In the style of a highly detailed digital illustration, each element is rendered with crisp clarity; the creature's shadowy plumage is not a flat black, but a complex texture composed of individually defined obsidian-like plumes, catching the infernal backlight along sharp edges. Its head, dotted with several pairs of blazing red eyes, burns like malevolent coals, and its head tilts back as its massive, hooked beak opens. Embers dance and swirl in the superheated air like a chaotic blizzard, illuminating its immense scale within the wall of fire. The creature lets out a wordless, terrifying scream as its throat muscles contort in a purely visual display of power, unleashing visible waves of heat distortion into the shimmering, ash-filled atmosphere, creating a vast and terrifying spectacle.", + "seed": 31950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_5/index_7830_seed_31950_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03843995165079832, + "vlm_logprob": -0.03843995165079832, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.03843995165079832, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16378355026245117 + ], + "opposite_logprob_tokens": [ + -1.9137835502624512 + ], + "logprob_reward": -0.16378355026245117, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16378355026245117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4534083604812622, + 0.09691055864095688, + 0.08319102227687836, + 0.03627170994877815, + 0.05847221985459328, + 0.02659444510936737, + 0.022156493738293648, + 0.03328867256641388, + 0.04669913277029991, + 0.14300741255283356 + ], + "credit_map": [ + 0.052230365574359894, + 0.04542243853211403, + 0.06481992453336716, + 0.06419841945171356, + 0.03811684623360634, + 0.05384964495897293, + 0.07530298829078674, + 0.05946771055459976, + 0.0072547695599496365, + 0.0129671860486269, + 0.010947134345769882, + 0.017392335459589958, + 0.007527059875428677, + 0.010686573572456837, + 0.015351912938058376, + 0.014783583581447601, + 0.006009757984429598, + 0.012940705753862858, + 0.012484163977205753, + 0.007769768591970205, + 0.015099998563528061, + 0.012335321865975857, + 0.010382908396422863, + 0.006168392486870289, + 0.004166872706264257, + 0.003426899900659919, + 0.005372351501137018, + 0.0036332218442112207, + 0.004946267697960138, + 0.0043785469606518745, + 0.005357780493795872, + 0.004989768844097853, + 0.004381230566650629, + 0.003071480430662632, + 0.02062170021235943, + 0.0024605100043118, + 0.004243241157382727, + 0.020200109109282494, + 0.0018048887141048908, + 0.0016890575643628836, + 0.0026843841187655926, + 0.003122930880635977, + 0.0029891140293329954, + 0.0058594876900315285, + 0.002169265178963542, + 0.005296190269291401, + 0.0023400152567774057, + 0.002133056754246354, + 0.0032013694290071726, + 0.0016221755649894476, + 0.0029454876203089952, + 0.003434938844293356, + 0.003296472132205963, + 0.0026763046626001596, + 0.0025995164178311825, + 0.002380229067057371, + 0.004339410923421383, + 0.003478279570117593, + 0.005583835765719414, + 0.005525569897145033, + 0.003346184967085719, + 0.00401623360812664, + 0.0026805519592016935, + 0.00431860564276576, + 0.005625314079225063, + 0.006376625970005989, + 0.006577775347977877, + 0.005596500355750322, + 0.005008479114621878, + 0.006727853789925575, + 0.0036917575635015965, + 0.007094826083630323, + 0.016368092969059944, + 0.004702288191765547, + 0.055777404457330704, + 0.00502272043377161, + 0.002399392193183303, + 0.009661604650318623, + 0.0031728914473205805, + 0.045903027057647705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.16378355026245117 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006761291529983282 + ], + "opposite_logprob_tokens": [ + -5.006761074066162 + ], + "logprob_reward": -0.006761291529983282, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006761291529983282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24934184551239014, + 0.09752151370048523, + 0.07415121048688889, + 0.147354394197464, + 0.11652294546365738, + 0.04581539332866669, + 0.04878581315279007, + 0.14716383814811707, + 0.036901991814374924, + 0.03644103184342384 + ], + "credit_map": [ + 0.031902290880680084, + 0.028430478647351265, + 0.030454685911536217, + 0.03504928946495056, + 0.024774719029664993, + 0.02774093672633171, + 0.038542136549949646, + 0.03244732692837715, + 0.010682888329029083, + 0.01080553513020277, + 0.011123775504529476, + 0.009373081848025322, + 0.016776395961642265, + 0.01640143245458603, + 0.010243597440421581, + 0.012114807032048702, + 0.005065550561994314, + 0.01061277650296688, + 0.007038015406578779, + 0.0083842221647501, + 0.01579863764345646, + 0.011942421086132526, + 0.010124093852937222, + 0.005185500252991915, + 0.013136196881532669, + 0.014222177676856518, + 0.017553869634866714, + 0.015039349906146526, + 0.022423816844820976, + 0.023248063400387764, + 0.022811291739344597, + 0.018919650465250015, + 0.010299726389348507, + 0.006519733462482691, + 0.033846233040094376, + 0.005564708728343248, + 0.009002474136650562, + 0.04194945469498634, + 0.003831705544143915, + 0.005508911795914173, + 0.004553597886115313, + 0.00644240016117692, + 0.005581685807555914, + 0.0051869177259504795, + 0.005241236183792353, + 0.010407201945781708, + 0.003518469166010618, + 0.0048838891088962555, + 0.007269410882145166, + 0.00429868558421731, + 0.00763632683083415, + 0.006318866740912199, + 0.005574880633503199, + 0.008123707957565784, + 0.004757096525281668, + 0.004806841723620892, + 0.01587950438261032, + 0.01847243309020996, + 0.018962813541293144, + 0.018796036019921303, + 0.0221120435744524, + 0.022445203736424446, + 0.013508378528058529, + 0.016987431794404984, + 0.0040553538128733635, + 0.0057584373280406, + 0.004065249115228653, + 0.005181583110243082, + 0.008551685139536858, + 0.003687225980684161, + 0.002329101786017418, + 0.0032733608968555927, + 0.00341287674382329, + 0.004152363166213036, + 0.0028232440818101168, + 0.0029172569047659636, + 0.003222392639145255, + 0.010811748914420605, + 0.0015865507302805781, + 0.007514599710702896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.006761291529983282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199530705809593 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199530705809593, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199530705809593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322660744190216, + 0.1680905520915985, + 0.05371619015932083, + 0.07361961901187897, + 0.06779270619153976, + 0.07206901907920837, + 0.04363791644573212, + 0.145083487033844, + 0.13411764800548553, + 0.10960680991411209 + ], + "credit_map": [ + 0.011494385078549385, + 0.015096291899681091, + 0.017165806144475937, + 0.01932084746658802, + 0.013269157148897648, + 0.015418833121657372, + 0.023562993854284286, + 0.016937756910920143, + 0.014852911233901978, + 0.015988664701581, + 0.019167553633451462, + 0.020552484318614006, + 0.027099566534161568, + 0.02964342013001442, + 0.017300300300121307, + 0.02348564751446247, + 0.004734727554023266, + 0.006596492603421211, + 0.008256002329289913, + 0.004481973126530647, + 0.009543923661112785, + 0.007251477800309658, + 0.0071125878021121025, + 0.005739009939134121, + 0.005326761398464441, + 0.0067891343496739864, + 0.0067911879159510136, + 0.008443661965429783, + 0.011927343904972076, + 0.008604319766163826, + 0.015190895646810532, + 0.010546320118010044, + 0.006379854865372181, + 0.005496694706380367, + 0.013996672816574574, + 0.00781608559191227, + 0.008317794650793076, + 0.012790457345545292, + 0.007594468537718058, + 0.00540067907422781, + 0.008183306083083153, + 0.007650257553905249, + 0.011197539046406746, + 0.013333757407963276, + 0.007664323318749666, + 0.013580075465142727, + 0.004713514354079962, + 0.005746240261942148, + 0.005562945734709501, + 0.004018789157271385, + 0.006694847717881203, + 0.006380113773047924, + 0.006451901979744434, + 0.005662826821208, + 0.0046042585745453835, + 0.004262230359017849, + 0.014077961444854736, + 0.019237764179706573, + 0.020472487434744835, + 0.02268102392554283, + 0.019785476848483086, + 0.020565401762723923, + 0.012687381356954575, + 0.015575984492897987, + 0.01605495624244213, + 0.01951620727777481, + 0.016269207000732422, + 0.01424639392644167, + 0.02613167278468609, + 0.019277408719062805, + 0.009454682469367981, + 0.013167121447622776, + 0.011579354293644428, + 0.004717919509857893, + 0.045243874192237854, + 0.005284477025270462, + 0.0028267973102629185, + 0.015367312356829643, + 0.0019211425678804517, + 0.022665925323963165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.014199530705809593 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638150800019503 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638150800019503, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638150800019503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11033128201961517, + 0.05262397602200508, + 0.03724845126271248, + 0.03745299205183983, + 0.12321710586547852, + 0.036038000136613846, + 0.08008348941802979, + 0.13464568555355072, + 0.28481149673461914, + 0.10354749858379364 + ], + "credit_map": [ + 0.011142315343022346, + 0.012096588499844074, + 0.013968326151371002, + 0.017160436138510704, + 0.010993259027600288, + 0.011644706130027771, + 0.01906978152692318, + 0.014255874790251255, + 0.006383595056831837, + 0.006034843623638153, + 0.006236129906028509, + 0.005797581281512976, + 0.009743841364979744, + 0.0075757685117423534, + 0.00557877030223608, + 0.00527344923466444, + 0.002474025124683976, + 0.004291010554879904, + 0.006403558887541294, + 0.004350216127932072, + 0.005497389007359743, + 0.005729138385504484, + 0.004025217145681381, + 0.004477894399315119, + 0.004194191657006741, + 0.004016377497464418, + 0.006051541306078434, + 0.0037783747538924217, + 0.005465376656502485, + 0.0046664047986269, + 0.004938606172800064, + 0.0043421173468232155, + 0.00839313492178917, + 0.0047877575270831585, + 0.044990479946136475, + 0.0051241060718894005, + 0.005499337334185839, + 0.04489737004041672, + 0.004820012953132391, + 0.004704911261796951, + 0.004139050841331482, + 0.005606898106634617, + 0.004066278226673603, + 0.006934620440006256, + 0.0032391378190368414, + 0.00590520491823554, + 0.0021371114999055862, + 0.004009697586297989, + 0.011573108844459057, + 0.009187295101583004, + 0.012285888195037842, + 0.007971611805260181, + 0.014957624487578869, + 0.00827751588076353, + 0.0073363399133086205, + 0.008494099602103233, + 0.02044518105685711, + 0.01493677869439125, + 0.016820179298520088, + 0.011874803341925144, + 0.01905977353453636, + 0.01983223482966423, + 0.013296387158334255, + 0.018380340188741684, + 0.043244797736406326, + 0.035058218985795975, + 0.037762463092803955, + 0.028659014031291008, + 0.04008402302861214, + 0.04100817069411278, + 0.01958988420665264, + 0.039404939860105515, + 0.010206952691078186, + 0.005228741094470024, + 0.04228048399090767, + 0.004340461455285549, + 0.002543011447414756, + 0.019504785537719727, + 0.0018554439302533865, + 0.017587613314390182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.004638150800019503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle shot of a forest floor with burning undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a dynamic spiraling motion that moves upward to reveal a massive creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central creature a large bird-like entity with obsidian-like plumage and multiple pairs of blazing red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature tilt its head back and open a massive hooked beak while unleashing visible waves of heat distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a raging, infernal apocalyptic scene filled with swirling embers and ash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817234955728054 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817234955728054, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817234955728054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22188785672187805, + 0.2618785500526428, + 0.03712795674800873, + 0.028293566778302193, + 0.06873412430286407, + 0.03587982431054115, + 0.035217516124248505, + 0.08410174399614334, + 0.058494530618190765, + 0.16838426887989044 + ], + "credit_map": [ + 0.027354979887604713, + 0.03614186495542526, + 0.027290930971503258, + 0.02862675115466118, + 0.017198042944073677, + 0.027577681466937065, + 0.029325086623430252, + 0.028372516855597496, + 0.024424536153674126, + 0.030879613012075424, + 0.03021829202771187, + 0.027695821598172188, + 0.03184743970632553, + 0.04523768648505211, + 0.02989153563976288, + 0.0416836254298687, + 0.003014222951605916, + 0.0034087705425918102, + 0.00771225243806839, + 0.004623706918209791, + 0.0062197414226830006, + 0.005008524749428034, + 0.0032157183159142733, + 0.003925020806491375, + 0.002853733953088522, + 0.0030980354640632868, + 0.005061805713921785, + 0.002905910601839423, + 0.0034362757578492165, + 0.0030076690018177032, + 0.004017662722617388, + 0.003912473097443581, + 0.004905181936919689, + 0.004879902116954327, + 0.020255250856280327, + 0.0035176295787096024, + 0.006947915535420179, + 0.021825917065143585, + 0.0032056565396487713, + 0.0031966683454811573, + 0.005022643133997917, + 0.0038188600447028875, + 0.0053466628305613995, + 0.006659283302724361, + 0.002763371681794524, + 0.003755784360691905, + 0.0023959202226251364, + 0.006117299664765596, + 0.006420941557735205, + 0.003636991372331977, + 0.005296406801789999, + 0.0036969203501939774, + 0.00461944006383419, + 0.004442883655428886, + 0.0021387108135968447, + 0.004965220578014851, + 0.008041134104132652, + 0.011545741930603981, + 0.012373167090117931, + 0.015517453663051128, + 0.011272364296019077, + 0.011680761352181435, + 0.006602826993912458, + 0.007068288046866655, + 0.007909916341304779, + 0.007489590905606747, + 0.00686577707529068, + 0.006554082967340946, + 0.010231513530015945, + 0.00913584791123867, + 0.004189523868262768, + 0.0061182789504528046, + 0.020762648433446884, + 0.007650179788470268, + 0.0690387487411499, + 0.0073850881308317184, + 0.0040243761613965034, + 0.02002534084022045, + 0.0036096409894526005, + 0.03588826581835747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.002817234955728054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_tensor([8179, 8179])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_tensor([8179, 8179])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b1b6f61318c26f99dcd26bcb466f94ec940d66c3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_tensor([8179, 8179])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot, captured from a slightly overhead, is rendered entirely in the style of a black-and-white pencil sketch. At the very center, a mother sits quietly in a dark wooden chair, her back to the camera. She wears a cozy home sweater, her hair loosely tied in a bun. An old-fashioned desk lamp provides the sole source of light, casting a warm, focused beam. This light, depicted in the sketch as a delicate field of white space and the lightest grays, precisely illuminates her focused back and the movements of her hands, while the rest of the room is shrouded in a deep, cross-hatched darkness. Her head is slightly lowered, her shoulders hunched forward, and her hands fly up and down with steady, deliberate movements. Two slender knitting needles intertwine between her deft fingers, a forming scarf dangling from them, its delicate texture sharply outlined in the light. As she gently tugged, the large, fluffy ball of yarn on the table beside her began to roll at an almost imperceptible pace, extremely slowly and irregularly. Each tiny twitch tugged at the lifeline connecting her hands. The entire scene, devoid of a single trace of superfluous color, consisted entirely of shifting lines and blocks of varying grayscale, creating an atmosphere of profound tranquility, warmth, and nostalgia.", + "seed": 27126, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_8179_seed_27126_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23521797228604555, + "vlm_logprob": -0.23521797228604555, + "vlm_yes_no_margin": 2.799999952316284, + "vlm_token_credit": -0.23521797228604555, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026803677901625633 + ], + "opposite_logprob_tokens": [ + -4.026803493499756 + ], + "logprob_reward": -0.026803677901625633, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.026803677901625633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1859855204820633, + 0.09012862294912338, + 0.567008376121521, + 0.051271531730890274, + 0.01951293833553791, + 0.011645101010799408, + 0.03266869857907295, + 0.018158404156565666, + 0.010788405314087868, + 0.01283238548785448 + ], + "credit_map": [ + 0.013372438959777355, + 0.007964267395436764, + 0.013213623315095901, + 0.010942029766738415, + 0.009283400140702724, + 0.06902341544628143, + 0.010441213846206665, + 0.05174515023827553, + 0.006988393608480692, + 0.004740441683679819, + 0.006618889514356852, + 0.005180394276976585, + 0.00517674395814538, + 0.028674615547060966, + 0.006356136407703161, + 0.026393016800284386, + 0.024268871173262596, + 0.009353282861411572, + 0.02093707025051117, + 0.009412329643964767, + 0.01269493717700243, + 0.2537679970264435, + 0.0111354049295187, + 0.22543849050998688, + 0.003248903201892972, + 0.002538349712267518, + 0.0037706512957811356, + 0.003797280602157116, + 0.0035964734852313995, + 0.015738651156425476, + 0.006571395788341761, + 0.012009830214083195, + 0.002200375311076641, + 0.0020500447135418653, + 0.0025791360531002283, + 0.0027895388193428516, + 0.002162926597520709, + 0.0026773514691740274, + 0.0029285994824022055, + 0.002124966587871313, + 0.0010398938320577145, + 0.0010576527565717697, + 0.0014220509910956025, + 0.0019685185980051756, + 0.0017012469470500946, + 0.0018019478302448988, + 0.001471786992624402, + 0.0011820037616416812, + 0.0032415559981018305, + 0.0031322911381721497, + 0.0036988791543990374, + 0.00628433795645833, + 0.004225017037242651, + 0.003704176051542163, + 0.004557644948363304, + 0.003824802115559578, + 0.001728832139633596, + 0.0015331129543483257, + 0.0019201029790565372, + 0.002608404029160738, + 0.0022767039481550455, + 0.003602576209232211, + 0.0022103269584476948, + 0.002278346801176667, + 0.0014237032737582922, + 0.0012880967697128654, + 0.0012819203548133373, + 0.0013473219005391002, + 0.0012759292731061578, + 0.0013200310058891773, + 0.001562007237225771, + 0.0012893968960270286, + 0.00136419793125242, + 0.0012097725411877036, + 0.001307884813286364, + 0.001405979972332716, + 0.0015573883429169655, + 0.002053706906735897, + 0.0013758477289229631, + 0.002557608066126704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.026803677901625633 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307173565030098 + ], + "opposite_logprob_tokens": [ + -3.168071746826172 + ], + "logprob_reward": -0.04307173565030098, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307173565030098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28905680775642395, + 0.0599348284304142, + 0.45832785964012146, + 0.10225776582956314, + 0.017014069482684135, + 0.011580330319702625, + 0.02070612832903862, + 0.012438619509339333, + 0.009545905515551567, + 0.019137652590870857 + ], + "credit_map": [ + 0.015130816958844662, + 0.008900659158825874, + 0.016786012798547745, + 0.01206150371581316, + 0.01109861396253109, + 0.11806288361549377, + 0.011078941635787487, + 0.09593736380338669, + 0.004422798752784729, + 0.004013844765722752, + 0.005124139133840799, + 0.006532476283609867, + 0.0057816882617771626, + 0.014619824476540089, + 0.006611411459743977, + 0.012828643433749676, + 0.018931150436401367, + 0.007227021269500256, + 0.017206786200404167, + 0.008256743662059307, + 0.009546693414449692, + 0.20540019869804382, + 0.00920762401074171, + 0.18255169689655304, + 0.0056320056319236755, + 0.00365475844591856, + 0.006728536449372768, + 0.00929566752165556, + 0.005365137942135334, + 0.03725642338395119, + 0.004351711831986904, + 0.029973527416586876, + 0.00173214974347502, + 0.0013313202653080225, + 0.0017020385712385178, + 0.002674130257219076, + 0.002605788642540574, + 0.0017602117732167244, + 0.003161768428981304, + 0.002046662848442793, + 0.0010297775734215975, + 0.0010202039266005158, + 0.0010526427067816257, + 0.0020330375991761684, + 0.0013719252310693264, + 0.0019434895366430283, + 0.001852498622611165, + 0.001276755123399198, + 0.0022729607298970222, + 0.002144315978512168, + 0.002175400499254465, + 0.004436011891812086, + 0.002611503005027771, + 0.002130677690729499, + 0.0027800812385976315, + 0.0021551779936999083, + 0.0013851134572178125, + 0.0011363174999132752, + 0.0015781793044880033, + 0.0015611962880939245, + 0.0017147031612694263, + 0.0020530761685222387, + 0.0015500069130212069, + 0.0014600277645513415, + 0.0007667533354833722, + 0.0007472545839846134, + 0.00089973583817482, + 0.0018637782195582986, + 0.0010889944387599826, + 0.0012897858396172523, + 0.0013750859070569277, + 0.0015145170036703348, + 0.0024809599854052067, + 0.0015928185312077403, + 0.0027869204059243202, + 0.0029201272409409285, + 0.0019219181267544627, + 0.0032598734833300114, + 0.002209738362580538, + 0.0019652952905744314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.04307173565030098 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0142021169885993 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.0142021169885993, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0142021169885993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017871873453259468, + 0.03175143525004387, + 0.8788628578186035, + 0.02966955490410328, + 0.020231401547789574, + 0.0017652194947004318, + 0.007688866462558508, + 0.00590268149971962, + 0.0027129855006933212, + 0.003543033730238676 + ], + "credit_map": [ + 0.0021484976168721914, + 0.0018877239199355245, + 0.0020958203822374344, + 0.0025388479698449373, + 0.00217272131703794, + 0.0019443851197138429, + 0.0026290970854461193, + 0.0024547805078327656, + 0.0017234007827937603, + 0.0010009774705395103, + 0.0018653101287782192, + 0.0011793539160862565, + 0.0009707663557492197, + 0.013405837118625641, + 0.0010900492779910564, + 0.010515740141272545, + 0.038445279002189636, + 0.014487797394394875, + 0.03310949727892876, + 0.014001872390508652, + 0.01877678371965885, + 0.3944019377231598, + 0.01741904206573963, + 0.3482206463813782, + 0.001624377560801804, + 0.0008394618635065854, + 0.0016873686108738184, + 0.0020798412151634693, + 0.001663275877945125, + 0.011090016923844814, + 0.0016577157657593489, + 0.009027496911585331, + 0.0026637224946171045, + 0.0027032862417399883, + 0.0025367424823343754, + 0.0024295304901897907, + 0.0020735044963657856, + 0.0025111266877502203, + 0.002723804907873273, + 0.0025896832812577486, + 0.00022977846674621105, + 0.00016074614541139454, + 0.00017434931942261755, + 0.00029155868105590343, + 0.00022823143808636814, + 0.0002424404228804633, + 0.0002467547601554543, + 0.00019136029004584998, + 0.0006763018318451941, + 0.0006949997623451054, + 0.0008755645249038935, + 0.0015619788318872452, + 0.0011310857953503728, + 0.0008374752942472696, + 0.0010969366412609816, + 0.0008145238389261067, + 0.0004856599844060838, + 0.0004175205249339342, + 0.0006344254361465573, + 0.000799795554485172, + 0.0007626874139532447, + 0.0012422320432960987, + 0.0007765095797367394, + 0.0007838511373847723, + 0.00027031818171963096, + 0.0002373509923927486, + 0.00029325459036044776, + 0.0003485430497676134, + 0.0003909944207407534, + 0.0003605135134421289, + 0.00043771654600277543, + 0.0003742941771633923, + 0.0004934778553433716, + 0.0003267576976213604, + 0.0004725770268123597, + 0.0005467602168209851, + 0.0003488665970508009, + 0.0004826157237403095, + 0.0003325703728478402, + 0.0005394082982093096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.0142021169885993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038090988993644714 + ], + "opposite_logprob_tokens": [ + -3.288090944290161 + ], + "logprob_reward": -0.038090988993644714, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038090988993644714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26828286051750183, + 0.1514291763305664, + 0.14465050399303436, + 0.16683952510356903, + 0.07543767988681793, + 0.024698860943317413, + 0.06619609892368317, + 0.04317227005958557, + 0.02909543178975582, + 0.030197568237781525 + ], + "credit_map": [ + 0.020905645564198494, + 0.017499713227152824, + 0.019833317026495934, + 0.028976382687687874, + 0.01394664030522108, + 0.08393549919128418, + 0.01785830222070217, + 0.06532736867666245, + 0.013753385283052921, + 0.008459864184260368, + 0.012034277431666851, + 0.013726496137678623, + 0.011594000272452831, + 0.04536516219377518, + 0.014085662551224232, + 0.032410312443971634, + 0.009946883656084538, + 0.00722713815048337, + 0.008724522776901722, + 0.013365832157433033, + 0.0121738500893116, + 0.040345095098018646, + 0.011333266273140907, + 0.04153391346335411, + 0.012358389794826508, + 0.007503202650696039, + 0.011756868101656437, + 0.018441092222929, + 0.008617601357400417, + 0.05326532945036888, + 0.010441986843943596, + 0.044455040246248245, + 0.006132107228040695, + 0.005327081307768822, + 0.00780476164072752, + 0.011067469604313374, + 0.00950221624225378, + 0.0109232934191823, + 0.015942810103297234, + 0.008737931959331036, + 0.001809213776141405, + 0.001874184119515121, + 0.002783536445349455, + 0.006338614504784346, + 0.002462817821651697, + 0.002920129569247365, + 0.0036405560094863176, + 0.002869807183742523, + 0.005725548602640629, + 0.0054499367251992226, + 0.007232446223497391, + 0.014419740997254848, + 0.009015708230435848, + 0.007981800474226475, + 0.00985498446971178, + 0.006515926215797663, + 0.0031755976378917694, + 0.002744971076026559, + 0.0046272058971226215, + 0.005920235067605972, + 0.004901987966150045, + 0.010004673153162003, + 0.005966785363852978, + 0.005830811802297831, + 0.0031825725454837084, + 0.0029051746241748333, + 0.0032502824906259775, + 0.004048608709126711, + 0.003766514826565981, + 0.0038633234798908234, + 0.004320935811847448, + 0.0037580179050564766, + 0.0031580780632793903, + 0.001771853887476027, + 0.00325048272497952, + 0.003948441706597805, + 0.003440785687416792, + 0.005564884282648563, + 0.0037766785826534033, + 0.0052863615565001965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.038090988993644714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539213418960571 + ], + "opposite_logprob_tokens": [ + -0.42892131209373474 + ], + "logprob_reward": -1.0539213418960571, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539213418960571, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10071396082639694, + 0.2213970571756363, + 0.34092941880226135, + 0.19709831476211548, + 0.02452530339360237, + 0.020853202790021896, + 0.03818501532077789, + 0.028066547587513924, + 0.012925623916089535, + 0.015305559150874615 + ], + "credit_map": [ + 0.008300382643938065, + 0.007416178938001394, + 0.007759281899780035, + 0.010657969862222672, + 0.008591986261308193, + 0.02222936972975731, + 0.010514650493860245, + 0.025244126096367836, + 0.011299597099423409, + 0.004386398941278458, + 0.012047773227095604, + 0.004865370225161314, + 0.006250719074159861, + 0.09727485477924347, + 0.006328013259917498, + 0.07894432544708252, + 0.016127608716487885, + 0.006909859366714954, + 0.01337154395878315, + 0.00765067245811224, + 0.008205276913940907, + 0.15060880780220032, + 0.008182584308087826, + 0.12987300753593445, + 0.01140169333666563, + 0.005962540861219168, + 0.01286587305366993, + 0.01828034594655037, + 0.00967785157263279, + 0.07297414541244507, + 0.007756222039461136, + 0.058179620653390884, + 0.003070661099627614, + 0.001839079661294818, + 0.0029478957876563072, + 0.003409150056540966, + 0.003133729100227356, + 0.002742116106674075, + 0.004037352278828621, + 0.003345317440107465, + 0.001615347689948976, + 0.0018511801026761532, + 0.0019018766470253468, + 0.003201447194442153, + 0.002652121940627694, + 0.0038209997583180666, + 0.0036094903480261564, + 0.0022007364314049482, + 0.00387473008595407, + 0.003616265719756484, + 0.00405422318726778, + 0.00740321958437562, + 0.00551867438480258, + 0.004033337812870741, + 0.006332995370030403, + 0.003351564286276698, + 0.0033572479151189327, + 0.0029616104438900948, + 0.0032757767476141453, + 0.00466894218698144, + 0.0035112195182591677, + 0.0034567697439342737, + 0.003770291805267334, + 0.0030646841041743755, + 0.0013744571479037404, + 0.0012190660927444696, + 0.0015816434752196074, + 0.0014982452848926187, + 0.0015110826352611184, + 0.0026433393359184265, + 0.001461937092244625, + 0.0016358522698283195, + 0.002216137945652008, + 0.0014306470984593034, + 0.0014839873183518648, + 0.0025683119893074036, + 0.0021901768632233143, + 0.0017780335620045662, + 0.00203066342510283, + 0.0016075989697128534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -1.0539213418960571 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "steady medium shot, captured from a slightly overhead, is rendered entirely in the style of a black-and-white pencil sketch. At the very center, a mother sits quietly in a dark wooden chair, her back to the camera. She wears a cozy home sweater, her hair loosely tied in a bun. An old-fashioned desk lamp provides the sole source of light, casting a warm, focused beam. This light, depicted in the sketch as a delicate field of white space and the lightest grays, precisely illuminates her focused back and the movements of her hands, while the rest of the room is shrouded in a deep, cross-hatched darkness. Her head is slightly lowered, her shoulders hunched forward, and her hands fly up and down with steady, deliberate movements. Two slender knitting needles intertwine between her deft fingers, a forming scarf dangling from them, its delicate texture sharply outlined in the light. As she gently tugged, the large, fluffy ball of yarn on the table beside her began to roll at an almost imperceptible pace, extremely slowly and irregularly. Each tiny twitch tugged at the lifeline connecting her hands. The entire scene, devoid of a single trace of superfluous color, consisted entirely of shifting lines and blocks of varying grayscale, creating an atmosphere of profound tranquility, warmth, and nostalgia.", + "seed": 27127, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_8179_seed_27127_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24363884776830674, + "vlm_logprob": -0.24363884776830674, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.24363884776830674, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04114016517996788 + ], + "opposite_logprob_tokens": [ + -3.916140079498291 + ], + "logprob_reward": -0.04114016517996788, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.04114016517996788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25067779421806335, + 0.16719168424606323, + 0.09301339834928513, + 0.0821123793721199, + 0.07455027848482132, + 0.05526452139019966, + 0.09892091900110245, + 0.04682065546512604, + 0.057185132056474686, + 0.07426324486732483 + ], + "credit_map": [ + 0.03708668053150177, + 0.019331352785229683, + 0.03786088526248932, + 0.02261963300406933, + 0.03694530948996544, + 0.02916005253791809, + 0.03074979968369007, + 0.0369240902364254, + 0.035207998007535934, + 0.013964894227683544, + 0.022583073005080223, + 0.009845406748354435, + 0.020715724676847458, + 0.0186698567122221, + 0.024053392931818962, + 0.02215135470032692, + 0.01497071422636509, + 0.006140129640698433, + 0.014866107143461704, + 0.008333290927112103, + 0.013213164173066616, + 0.010334797203540802, + 0.01223591435700655, + 0.012919284403324127, + 0.010668104514479637, + 0.00606779707595706, + 0.010454444214701653, + 0.00454525975510478, + 0.010339003056287766, + 0.010898522101342678, + 0.014565928839147091, + 0.014573323540389538, + 0.012002423405647278, + 0.007405018899589777, + 0.007360952440649271, + 0.0055945334024727345, + 0.01203975174576044, + 0.008765853941440582, + 0.010117627680301666, + 0.011264113709330559, + 0.007417480926960707, + 0.004387464374303818, + 0.007949506863951683, + 0.0057331882417202, + 0.006767887156456709, + 0.00614706939086318, + 0.008088496513664722, + 0.00877342838793993, + 0.014626062475144863, + 0.008086622692644596, + 0.012590756639838219, + 0.0060454970225691795, + 0.018067510798573494, + 0.011260319501161575, + 0.013188709504902363, + 0.015055437572300434, + 0.007009529508650303, + 0.00448479363694787, + 0.005583911202847958, + 0.0034305357839912176, + 0.007558614946901798, + 0.0052369628101587296, + 0.006628261413425207, + 0.006888046395033598, + 0.010281862691044807, + 0.004407224711030722, + 0.007230522111058235, + 0.003595986170694232, + 0.009944083169102669, + 0.005981413647532463, + 0.007430646102875471, + 0.008313393220305443, + 0.010777311399579048, + 0.007316167466342449, + 0.008744820021092892, + 0.003986291587352753, + 0.0115636782720685, + 0.010943526402115822, + 0.010053199715912342, + 0.010878247208893299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.04114016517996788 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003426551818848 + ], + "opposite_logprob_tokens": [ + -2.6950342655181885 + ], + "logprob_reward": -0.07003426551818848, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003426551818848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20088553428649902, + 0.20374073088169098, + 0.051229700446128845, + 0.07573040574789047, + 0.1357870101928711, + 0.08233755081892014, + 0.09360560774803162, + 0.0621037594974041, + 0.028610428795218468, + 0.06596917659044266 + ], + "credit_map": [ + 0.022204505279660225, + 0.018796902149915695, + 0.025453047826886177, + 0.02183530665934086, + 0.03581688180565834, + 0.0175503920763731, + 0.028946971520781517, + 0.030281564220786095, + 0.02699034847319126, + 0.021607480943202972, + 0.025985155254602432, + 0.012614909559488297, + 0.02941303141415119, + 0.026860475540161133, + 0.027881858870387077, + 0.032387491315603256, + 0.005780038423836231, + 0.0037164129316806793, + 0.007023555226624012, + 0.004251144826412201, + 0.008553370833396912, + 0.005210470873862505, + 0.00853678211569786, + 0.008157927542924881, + 0.01112507563084364, + 0.006828377954661846, + 0.008771620690822601, + 0.0048592062667012215, + 0.00896299909800291, + 0.010465885512530804, + 0.012476712465286255, + 0.012240533716976643, + 0.012428670190274715, + 0.014645974151790142, + 0.014291733503341675, + 0.009697122499346733, + 0.020561937242746353, + 0.022570857778191566, + 0.020612459629774094, + 0.020978271961212158, + 0.0151417450979352, + 0.006681906525045633, + 0.009871003217995167, + 0.0050799003802239895, + 0.011323397047817707, + 0.008019804954528809, + 0.012431161478161812, + 0.013788641430437565, + 0.012537315487861633, + 0.007512714248150587, + 0.010120871476829052, + 0.006739206146448851, + 0.01496048178523779, + 0.010824025608599186, + 0.016176383942365646, + 0.014734617434442043, + 0.00996489729732275, + 0.005067907273769379, + 0.009117837995290756, + 0.006011659745126963, + 0.009923427365720272, + 0.0059101274237036705, + 0.007756214123219252, + 0.008351699449121952, + 0.00485181575641036, + 0.0024900808930397034, + 0.0043036784045398235, + 0.0022281911224126816, + 0.00425859447568655, + 0.0030376508366316557, + 0.003215250326320529, + 0.004225168377161026, + 0.006670708302408457, + 0.009384862147271633, + 0.00902553927153349, + 0.004220090806484222, + 0.009455054067075253, + 0.007596345618367195, + 0.009426233358681202, + 0.010190350003540516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.07003426551818848 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299720138311386 + ], + "opposite_logprob_tokens": [ + -3.7732996940612793 + ], + "logprob_reward": -0.023299720138311386, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299720138311386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18463170528411865, + 0.0619344487786293, + 0.046489499509334564, + 0.021626194939017296, + 0.035028502345085144, + 0.07636464387178421, + 0.35488295555114746, + 0.180153489112854, + 0.014924897812306881, + 0.023963648825883865 + ], + "credit_map": [ + 0.021845709532499313, + 0.017123298719525337, + 0.02222347818315029, + 0.017362503334879875, + 0.030246306210756302, + 0.021331222727894783, + 0.02497604675590992, + 0.02952313795685768, + 0.007556689903140068, + 0.005308418534696102, + 0.0068886117078363895, + 0.005578072275966406, + 0.009760493412613869, + 0.00604837154969573, + 0.009854649193584919, + 0.010939141735434532, + 0.005654154345393181, + 0.004188133403658867, + 0.005047278944402933, + 0.00608178460970521, + 0.006911080796271563, + 0.004359336104243994, + 0.006822754628956318, + 0.007424972951412201, + 0.0026894956827163696, + 0.0019383274484425783, + 0.0031914019491523504, + 0.0017384886741638184, + 0.0030338759534060955, + 0.002816391410306096, + 0.0031703594140708447, + 0.00304785487242043, + 0.0039470805786550045, + 0.0032422859221696854, + 0.0043769399635493755, + 0.002142667304724455, + 0.006269758101552725, + 0.004827572498470545, + 0.0047965701669454575, + 0.005425631534308195, + 0.009274967014789581, + 0.006168121472001076, + 0.008754469454288483, + 0.010280138812959194, + 0.010222967714071274, + 0.006835669279098511, + 0.011198733933269978, + 0.01362957339733839, + 0.045695286244153976, + 0.031328458338975906, + 0.04565976932644844, + 0.023729322478175163, + 0.04913203418254852, + 0.04161100834608078, + 0.054419293999671936, + 0.06330776959657669, + 0.024367978796362877, + 0.014172668568789959, + 0.024407872930169106, + 0.013129128143191338, + 0.02625456266105175, + 0.021980581805109978, + 0.026058558374643326, + 0.029782136902213097, + 0.0017924089916050434, + 0.0013974168105050921, + 0.0021610211115330458, + 0.0013402178883552551, + 0.002633258467540145, + 0.0014465234708040953, + 0.0018291978631168604, + 0.0023248528596013784, + 0.0043350886553525925, + 0.0023053502663969994, + 0.0029650882352143526, + 0.0018275866750627756, + 0.003153534373268485, + 0.002454134402796626, + 0.003407982876524329, + 0.0035148828756064177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.023299720138311386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792189598083496 + ], + "opposite_logprob_tokens": [ + -3.529792308807373 + ], + "logprob_reward": -0.029792189598083496, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792189598083496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1936439871788025, + 0.08730388432741165, + 0.09255775064229965, + 0.04954967275261879, + 0.05571909248828888, + 0.07356665283441544, + 0.25294798612594604, + 0.1247912123799324, + 0.025390522554516792, + 0.044529207050800323 + ], + "credit_map": [ + 0.02636067569255829, + 0.015543382614850998, + 0.02477097511291504, + 0.014969009906053543, + 0.026960700750350952, + 0.02410949021577835, + 0.03481582552194595, + 0.026113921776413918, + 0.010350743308663368, + 0.00829971581697464, + 0.009900923818349838, + 0.006289129611104727, + 0.014571783132851124, + 0.011838792823255062, + 0.01159868948161602, + 0.014454108662903309, + 0.01490981038659811, + 0.007978023961186409, + 0.014249248430132866, + 0.007493455428630114, + 0.012576117180287838, + 0.010245444253087044, + 0.013683082535862923, + 0.01142257172614336, + 0.008396603167057037, + 0.004038029350340366, + 0.005720027722418308, + 0.002780415816232562, + 0.007107732351869345, + 0.006220746785402298, + 0.007650065701454878, + 0.007636050693690777, + 0.004912919830530882, + 0.004701193422079086, + 0.007040554657578468, + 0.004578695632517338, + 0.010124512016773224, + 0.007156210485845804, + 0.008788689970970154, + 0.008416314609348774, + 0.009922761470079422, + 0.006335902493447065, + 0.007476203143596649, + 0.008194491267204285, + 0.009723776951432228, + 0.007368352264165878, + 0.009571058675646782, + 0.014974109828472137, + 0.03147837892174721, + 0.02271193265914917, + 0.032000403851270676, + 0.01635161228477955, + 0.034978270530700684, + 0.030574271455407143, + 0.040337782353162766, + 0.04451533779501915, + 0.017133696004748344, + 0.00963153038173914, + 0.016716858372092247, + 0.009034160524606705, + 0.017947446554899216, + 0.015381787903606892, + 0.018131490796804428, + 0.02081424556672573, + 0.003935071174055338, + 0.002189934952184558, + 0.0040575782768428326, + 0.0019567382987588644, + 0.004008642863482237, + 0.002719693351536989, + 0.0029168743640184402, + 0.003605988807976246, + 0.00736231217160821, + 0.004700850695371628, + 0.00623015221208334, + 0.0024803944397717714, + 0.005597345530986786, + 0.005548688117414713, + 0.006137473043054342, + 0.0064719924703240395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.029792189598083496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539278984069824 + ], + "opposite_logprob_tokens": [ + -0.4289278984069824 + ], + "logprob_reward": -1.0539278984069824, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539278984069824, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2333269566297531, + 0.08913634717464447, + 0.05712776631116867, + 0.058486610651016235, + 0.09147395193576813, + 0.1011415421962738, + 0.1874133050441742, + 0.08247225731611252, + 0.05052981898188591, + 0.04889136552810669 + ], + "credit_map": [ + 0.026548614725470543, + 0.020873576402664185, + 0.026365265250205994, + 0.018809495493769646, + 0.0426466166973114, + 0.03592855855822563, + 0.026677284389734268, + 0.03547757863998413, + 0.014769173227250576, + 0.009215771220624447, + 0.01128318253904581, + 0.007030325476080179, + 0.014134895987808704, + 0.01084611751139164, + 0.009927067905664444, + 0.011929829604923725, + 0.00827409140765667, + 0.003925848752260208, + 0.00740988552570343, + 0.0033790755551308393, + 0.011003480292856693, + 0.008201949298381805, + 0.006531621795147657, + 0.008401824161410332, + 0.011247350834310055, + 0.004646021407097578, + 0.006490131840109825, + 0.0042218598537147045, + 0.007529675494879484, + 0.005553926806896925, + 0.00917130522429943, + 0.009626349434256554, + 0.016233965754508972, + 0.008600878529250622, + 0.014171475544571877, + 0.006684734486043453, + 0.012465844862163067, + 0.009536138735711575, + 0.010485024191439152, + 0.01329590193927288, + 0.012349422089755535, + 0.00758858909830451, + 0.010348738171160221, + 0.006476502399891615, + 0.02049216255545616, + 0.01572079211473465, + 0.011481997556984425, + 0.016683349385857582, + 0.02347872033715248, + 0.017003057524561882, + 0.0226710494607687, + 0.012232980690896511, + 0.025253960862755775, + 0.023095840588212013, + 0.030918996781110764, + 0.03275873884558678, + 0.011207561008632183, + 0.0067626843228936195, + 0.010060260072350502, + 0.006121998652815819, + 0.012218313291668892, + 0.008980713784694672, + 0.012583278119564056, + 0.01453746110200882, + 0.008268384262919426, + 0.0029778045136481524, + 0.006484458222985268, + 0.003619761671870947, + 0.009228547103703022, + 0.0057105980813503265, + 0.006556898355484009, + 0.007683373987674713, + 0.006175411865115166, + 0.005117641296237707, + 0.005004031118005514, + 0.003146519185975194, + 0.008856277912855148, + 0.0075341397896409035, + 0.005490658339112997, + 0.007566696032881737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -1.0539278984069824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "steady medium shot, captured from a slightly overhead, is rendered entirely in the style of a black-and-white pencil sketch. At the very center, a mother sits quietly in a dark wooden chair, her back to the camera. She wears a cozy home sweater, her hair loosely tied in a bun. An old-fashioned desk lamp provides the sole source of light, casting a warm, focused beam. This light, depicted in the sketch as a delicate field of white space and the lightest grays, precisely illuminates her focused back and the movements of her hands, while the rest of the room is shrouded in a deep, cross-hatched darkness. Her head is slightly lowered, her shoulders hunched forward, and her hands fly up and down with steady, deliberate movements. Two slender knitting needles intertwine between her deft fingers, a forming scarf dangling from them, its delicate texture sharply outlined in the light. As she gently tugged, the large, fluffy ball of yarn on the table beside her began to roll at an almost imperceptible pace, extremely slowly and irregularly. Each tiny twitch tugged at the lifeline connecting her hands. The entire scene, devoid of a single trace of superfluous color, consisted entirely of shifting lines and blocks of varying grayscale, creating an atmosphere of profound tranquility, warmth, and nostalgia.", + "seed": 27128, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_8179_seed_27128_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23585597798228264, + "vlm_logprob": -0.23585597798228264, + "vlm_yes_no_margin": 1.875, + "vlm_token_credit": -0.23585597798228264, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1187269389629364 + ], + "opposite_logprob_tokens": [ + -2.243726968765259 + ], + "logprob_reward": -0.1187269389629364, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1187269389629364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27243494987487793, + 0.1659543663263321, + 0.09895016252994537, + 0.040526002645492554, + 0.024098142981529236, + 0.05940792337059975, + 0.03143078088760376, + 0.024907255545258522, + 0.04949003458023071, + 0.23280037939548492 + ], + "credit_map": [ + 0.04161858931183815, + 0.035313475877046585, + 0.03022676892578602, + 0.04967778921127319, + 0.037298593670129776, + 0.031011050567030907, + 0.019266212359070778, + 0.028022481128573418, + 0.01828845776617527, + 0.01750705949962139, + 0.027251435443758965, + 0.030238807201385498, + 0.022170621901750565, + 0.018238244578242302, + 0.01740981452167034, + 0.014849931932985783, + 0.012818594463169575, + 0.009595215320587158, + 0.015732742846012115, + 0.016199586912989616, + 0.010135142132639885, + 0.015438081696629524, + 0.007385615725070238, + 0.011645183898508549, + 0.004061475396156311, + 0.0038032508455216885, + 0.007872482761740685, + 0.0077014220878481865, + 0.0044000945053994656, + 0.005164234898984432, + 0.004032233264297247, + 0.003490808419883251, + 0.003999603912234306, + 0.0023360936902463436, + 0.0028328518383204937, + 0.004423167090862989, + 0.002494476968422532, + 0.002972512971609831, + 0.0026164944283664227, + 0.00242294161580503, + 0.007690452504903078, + 0.006491338834166527, + 0.00754532078281045, + 0.010968138463795185, + 0.00811635423451662, + 0.006173430010676384, + 0.005549844354391098, + 0.006873048841953278, + 0.003585715312510729, + 0.0038289676886051893, + 0.003825787454843521, + 0.004347586072981358, + 0.005704905837774277, + 0.0037965632509440184, + 0.002677912823855877, + 0.003663340350612998, + 0.002856160281226039, + 0.0031428562942892313, + 0.003572075627744198, + 0.004021692089736462, + 0.00297925085760653, + 0.002667589345946908, + 0.0025574860628694296, + 0.0031101449858397245, + 0.005547575652599335, + 0.004471153952181339, + 0.010615624487400055, + 0.0063895550556480885, + 0.00722999032586813, + 0.005865495651960373, + 0.003141480963677168, + 0.006229162681847811, + 0.023969104513525963, + 0.01945379376411438, + 0.06469761580228806, + 0.03295207396149635, + 0.03319164738059044, + 0.02729443646967411, + 0.01986873522400856, + 0.011372978799045086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.1187269389629364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993506789207458 + ], + "opposite_logprob_tokens": [ + -1.804935097694397 + ], + "logprob_reward": -0.17993506789207458, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993506789207458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16556183993816376, + 0.10840760916471481, + 0.08154713362455368, + 0.12078683823347092, + 0.04295102134346962, + 0.04723166674375534, + 0.048204559832811356, + 0.04281371831893921, + 0.07562830299139023, + 0.2668673098087311 + ], + "credit_map": [ + 0.01826915703713894, + 0.025010988116264343, + 0.021095504984259605, + 0.025952056050300598, + 0.026500822976231575, + 0.02215827815234661, + 0.011556648649275303, + 0.015018383972346783, + 0.010313484817743301, + 0.010171346366405487, + 0.023331832140684128, + 0.016198212280869484, + 0.015200111083686352, + 0.013943028636276722, + 0.00852534081786871, + 0.010724260471761227, + 0.006746532861143351, + 0.00539774727076292, + 0.0215067770332098, + 0.012709403410553932, + 0.00837523490190506, + 0.01328397449105978, + 0.005674635525792837, + 0.007852831855416298, + 0.012385450303554535, + 0.010664228349924088, + 0.02690798044204712, + 0.01997399516403675, + 0.013238529674708843, + 0.017068685963749886, + 0.008893503807485104, + 0.011654462665319443, + 0.005919543094933033, + 0.0037494415882974863, + 0.009032616391777992, + 0.007276891730725765, + 0.004790837876498699, + 0.004937062971293926, + 0.003949507139623165, + 0.0032951177563518286, + 0.005824284162372351, + 0.003753525670617819, + 0.011195597238838673, + 0.00838482566177845, + 0.0051044621504843235, + 0.005305848550051451, + 0.004392419941723347, + 0.0032707054633647203, + 0.005076191853731871, + 0.005725434981286526, + 0.005961897782981396, + 0.006896139122545719, + 0.008573384955525398, + 0.005863169673830271, + 0.003563289064913988, + 0.006545047741383314, + 0.004426559433341026, + 0.005248095374554396, + 0.006731866858899593, + 0.007665639277547598, + 0.00779175478965044, + 0.004817245528101921, + 0.0030502229928970337, + 0.0030823282431811094, + 0.007804859429597855, + 0.009993288666009903, + 0.01168679166585207, + 0.010809719562530518, + 0.015063067898154259, + 0.009254645556211472, + 0.005592159926891327, + 0.005423770751804113, + 0.025726690888404846, + 0.020883528515696526, + 0.08165119588375092, + 0.03732699900865555, + 0.03546789661049843, + 0.0324985533952713, + 0.01909215748310089, + 0.014220267534255981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.17993506789207458 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371826931834221 + ], + "opposite_logprob_tokens": [ + -3.4087183475494385 + ], + "logprob_reward": -0.03371826931834221, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371826931834221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17185451090335846, + 0.09537766128778458, + 0.07308042049407959, + 0.29662850499153137, + 0.03071073442697525, + 0.014866583049297333, + 0.04485860466957092, + 0.02378935180604458, + 0.09043613821268082, + 0.15839746594429016 + ], + "credit_map": [ + 0.01924995891749859, + 0.02288558892905712, + 0.02364960126578808, + 0.02703293226659298, + 0.027702925726771355, + 0.021688327193260193, + 0.013301389291882515, + 0.016343798488378525, + 0.007917914539575577, + 0.00813272688537836, + 0.022062702104449272, + 0.014844241552054882, + 0.012527456507086754, + 0.015034643933176994, + 0.006109057925641537, + 0.008748921565711498, + 0.005972496699541807, + 0.005205837078392506, + 0.014533886685967445, + 0.013734661042690277, + 0.00848514772951603, + 0.011228831484913826, + 0.007258262950927019, + 0.006661296356469393, + 0.02729441039264202, + 0.024766651913523674, + 0.06640131771564484, + 0.04894666746258736, + 0.0370943546295166, + 0.039386529475450516, + 0.028280310332775116, + 0.02445826306939125, + 0.004035611636936665, + 0.002528650686144829, + 0.0063914088532328606, + 0.0053499662317335606, + 0.003125599352642894, + 0.00365012651309371, + 0.0029386160895228386, + 0.002690754598006606, + 0.001949033234268427, + 0.0012461685109883547, + 0.002351642120629549, + 0.0024510701186954975, + 0.0021256706677377224, + 0.0015942720929160714, + 0.0016662750858813524, + 0.0014824505196884274, + 0.004506247118115425, + 0.0048608072102069855, + 0.007891514338552952, + 0.007501645013689995, + 0.007921839132905006, + 0.005974609404802322, + 0.002880063373595476, + 0.0033218765165656805, + 0.002793762367218733, + 0.002657400444149971, + 0.004921645857393742, + 0.0035172621719539165, + 0.00332944979891181, + 0.0027759938966482878, + 0.0018163688946515322, + 0.001977466745302081, + 0.008370071649551392, + 0.009010654874145985, + 0.021107681095600128, + 0.01217852532863617, + 0.016102904453873634, + 0.010858595371246338, + 0.0058022551238536835, + 0.007005450781434774, + 0.015027410350739956, + 0.013320437632501125, + 0.04570285603404045, + 0.021777993068099022, + 0.023814363405108452, + 0.01932113617658615, + 0.011873890645802021, + 0.007559370249509811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.03371826931834221 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904175460338593 + ], + "opposite_logprob_tokens": [ + -2.4640417098999023 + ], + "logprob_reward": -0.08904175460338593, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904175460338593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18090617656707764, + 0.12676207721233368, + 0.09457307308912277, + 0.0945107489824295, + 0.05718519166111946, + 0.06239372864365578, + 0.04166471213102341, + 0.033038731664419174, + 0.06175798550248146, + 0.24720752239227295 + ], + "credit_map": [ + 0.018570007756352425, + 0.026767387986183167, + 0.022938238456845284, + 0.03475755453109741, + 0.030364317819476128, + 0.019426245242357254, + 0.012856690213084221, + 0.015225769020617008, + 0.012798052281141281, + 0.010774090886116028, + 0.02338016964495182, + 0.019188126549124718, + 0.013597805984318256, + 0.01866155117750168, + 0.013496724888682365, + 0.014865560457110405, + 0.010531187057495117, + 0.00980033352971077, + 0.015164471231400967, + 0.015376185066998005, + 0.01163532491773367, + 0.011092238128185272, + 0.009933189488947392, + 0.011040152981877327, + 0.008743860758841038, + 0.010085567831993103, + 0.018091624602675438, + 0.01598072238266468, + 0.013831756077706814, + 0.012802829965949059, + 0.006048405542969704, + 0.008925992995500565, + 0.00639347406104207, + 0.007855040021240711, + 0.007303422316908836, + 0.010801419615745544, + 0.008532650768756866, + 0.006194621790200472, + 0.006271444261074066, + 0.003833124879747629, + 0.006720645818859339, + 0.007756183389574289, + 0.008161061443388462, + 0.011614422313869, + 0.0101176081225276, + 0.006550423800945282, + 0.0065952786244452, + 0.004878109320998192, + 0.005220138002187014, + 0.004816959612071514, + 0.005870729684829712, + 0.006474127061665058, + 0.005668977275490761, + 0.004783600103110075, + 0.0034989151172339916, + 0.0053312708623707294, + 0.0035006029065698385, + 0.004142969846725464, + 0.005519367754459381, + 0.005452786572277546, + 0.005778414662927389, + 0.0035994956269860268, + 0.00244659511372447, + 0.0025985026732087135, + 0.005845288746058941, + 0.007165363524109125, + 0.012181238271296024, + 0.008574691601097584, + 0.011418696492910385, + 0.007535412907600403, + 0.004251671023666859, + 0.004785625264048576, + 0.022949445992708206, + 0.023971157148480415, + 0.06789032369852066, + 0.030573712661862373, + 0.0410151444375515, + 0.030203884467482567, + 0.01816665381193161, + 0.012437218800187111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.08904175460338593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578578591346741 + ], + "opposite_logprob_tokens": [ + -0.6328578591346741 + ], + "logprob_reward": -0.7578578591346741, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578578591346741, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14160218834877014, + 0.1345427930355072, + 0.14340068399906158, + 0.08180340379476547, + 0.15048211812973022, + 0.07250116765499115, + 0.024696117267012596, + 0.06337658315896988, + 0.043782200664281845, + 0.14381273090839386 + ], + "credit_map": [ + 0.015890896320343018, + 0.01940501481294632, + 0.0157777089625597, + 0.02424459531903267, + 0.024001164361834526, + 0.016252197325229645, + 0.012273897416889668, + 0.013756710104644299, + 0.013076814822852612, + 0.013329779729247093, + 0.029667416587471962, + 0.019505489617586136, + 0.017293445765972137, + 0.019225934520363808, + 0.009363947436213493, + 0.013079958967864513, + 0.013036011718213558, + 0.011310100555419922, + 0.03971444442868233, + 0.02226734533905983, + 0.015505418181419373, + 0.019854912534356117, + 0.010403654538094997, + 0.011308783665299416, + 0.007357487455010414, + 0.009277804754674435, + 0.013454845175147057, + 0.013299272395670414, + 0.014133026823401451, + 0.010407431051135063, + 0.006852009333670139, + 0.0070215133018791676, + 0.016307640820741653, + 0.01630701683461666, + 0.03800104558467865, + 0.02354714833199978, + 0.017166832461953163, + 0.015001745894551277, + 0.014471102505922318, + 0.009679572656750679, + 0.007236804813146591, + 0.006592951249331236, + 0.017901645973324776, + 0.010722366161644459, + 0.008216938003897667, + 0.007637695409357548, + 0.007807435002177954, + 0.006385322194546461, + 0.003008815925568342, + 0.0029845822136849165, + 0.0032476303167641163, + 0.004401483107358217, + 0.0033335420303046703, + 0.0027158083394169807, + 0.00261806882917881, + 0.002386184176430106, + 0.007174656260758638, + 0.0057603539898991585, + 0.016599304974079132, + 0.00931322667747736, + 0.007954935543239117, + 0.00798949133604765, + 0.004922814201563597, + 0.003661789931356907, + 0.004238353110849857, + 0.0045191943645477295, + 0.008344888687133789, + 0.007807986810803413, + 0.006125589367002249, + 0.004711485002189875, + 0.0037948961835354567, + 0.004239807836711407, + 0.015018180944025517, + 0.009386682882905006, + 0.04806758463382721, + 0.021520672366023064, + 0.01411695871502161, + 0.018071182072162628, + 0.009980953298509121, + 0.007650509476661682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.7578578591346741 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "steady medium shot, captured from a slightly overhead, is rendered entirely in the style of a black-and-white pencil sketch. At the very center, a mother sits quietly in a dark wooden chair, her back to the camera. She wears a cozy home sweater, her hair loosely tied in a bun. An old-fashioned desk lamp provides the sole source of light, casting a warm, focused beam. This light, depicted in the sketch as a delicate field of white space and the lightest grays, precisely illuminates her focused back and the movements of her hands, while the rest of the room is shrouded in a deep, cross-hatched darkness. Her head is slightly lowered, her shoulders hunched forward, and her hands fly up and down with steady, deliberate movements. Two slender knitting needles intertwine between her deft fingers, a forming scarf dangling from them, its delicate texture sharply outlined in the light. As she gently tugged, the large, fluffy ball of yarn on the table beside her began to roll at an almost imperceptible pace, extremely slowly and irregularly. Each tiny twitch tugged at the lifeline connecting her hands. The entire scene, devoid of a single trace of superfluous color, consisted entirely of shifting lines and blocks of varying grayscale, creating an atmosphere of profound tranquility, warmth, and nostalgia.", + "seed": 27129, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_6/index_8179_seed_27129_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1931217972189188, + "vlm_logprob": -0.1931217972189188, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.1931217972189188, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01729150488972664 + ], + "opposite_logprob_tokens": [ + -4.76729154586792 + ], + "logprob_reward": -0.01729150488972664, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01729150488972664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21980559825897217, + 0.14238150417804718, + 0.1077602282166481, + 0.06100493296980858, + 0.10552935302257538, + 0.07362695783376694, + 0.04095404967665672, + 0.06295979022979736, + 0.08920654654502869, + 0.09677101671695709 + ], + "credit_map": [ + 0.022836165502667427, + 0.024737030267715454, + 0.020287679508328438, + 0.02718990296125412, + 0.03481429070234299, + 0.03306052088737488, + 0.03132166713476181, + 0.025558333843946457, + 0.015156840905547142, + 0.016359692439436913, + 0.013919907622039318, + 0.019612397998571396, + 0.016940215602517128, + 0.01802990213036537, + 0.026789212599396706, + 0.015573319979012012, + 0.008361641317605972, + 0.017076000571250916, + 0.011891421861946583, + 0.011912287212908268, + 0.016070595011115074, + 0.013280786573886871, + 0.01734575442969799, + 0.011821734718978405, + 0.006766767241060734, + 0.006205408833920956, + 0.007290083449333906, + 0.006885012611746788, + 0.008204723708331585, + 0.00907046813517809, + 0.006810346152633429, + 0.009772120974957943, + 0.009067813865840435, + 0.014349116943776608, + 0.009952877648174763, + 0.014349615201354027, + 0.013688714243471622, + 0.01512098591774702, + 0.01571779139339924, + 0.013282446190714836, + 0.007098350673913956, + 0.007982034236192703, + 0.007827220484614372, + 0.011226826347410679, + 0.0102096451446414, + 0.008873289451003075, + 0.011668132618069649, + 0.008741458877921104, + 0.0046314760111272335, + 0.004290234763175249, + 0.0070014758966863155, + 0.004421737976372242, + 0.005414172075688839, + 0.005037045106291771, + 0.005018644034862518, + 0.005139262415468693, + 0.00482243113219738, + 0.007852908223867416, + 0.006499032024294138, + 0.006791484542191029, + 0.008868650533258915, + 0.007990731857717037, + 0.011206860654056072, + 0.008927691727876663, + 0.013594405725598335, + 0.010970239527523518, + 0.011059565469622612, + 0.00884256511926651, + 0.012387226335704327, + 0.01055725198239088, + 0.012439326383173466, + 0.009355973452329636, + 0.009471532888710499, + 0.009586936794221401, + 0.011541716754436493, + 0.01343468762934208, + 0.014134486205875874, + 0.010389616712927818, + 0.018254738301038742, + 0.00995730236172676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01729150488972664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255018636584282 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.01255018636584282, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255018636584282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11006604135036469, + 0.11568447947502136, + 0.07583613693714142, + 0.06278608739376068, + 0.3922196626663208, + 0.09910839051008224, + 0.03635436296463013, + 0.06641244143247604, + 0.01355535164475441, + 0.02797696925699711 + ], + "credit_map": [ + 0.008142098784446716, + 0.014077977277338505, + 0.00958298146724701, + 0.01622891053557396, + 0.017669226974248886, + 0.012033243663609028, + 0.019790541380643845, + 0.012541056610643864, + 0.010287403129041195, + 0.015534235164523125, + 0.010993441566824913, + 0.014626101590692997, + 0.019003529101610184, + 0.01548779010772705, + 0.01619434915482998, + 0.013557630591094494, + 0.00726506020873785, + 0.012394115328788757, + 0.007988782599568367, + 0.00681640300899744, + 0.010170824825763702, + 0.010053952224552631, + 0.010670771822333336, + 0.010476228781044483, + 0.005413717590272427, + 0.007118830922991037, + 0.007757669314742088, + 0.0069844708777964115, + 0.009264979511499405, + 0.008736012503504753, + 0.009458296000957489, + 0.008052114397287369, + 0.036491502076387405, + 0.052774928510189056, + 0.03649628907442093, + 0.055668000131845474, + 0.061472222208976746, + 0.056377705186605453, + 0.05635904148221016, + 0.03657998889684677, + 0.008404738269746304, + 0.011092793196439743, + 0.009678027592599392, + 0.013904460705816746, + 0.012661888264119625, + 0.011860202066600323, + 0.020042099058628082, + 0.011464179493486881, + 0.0029192103538662195, + 0.00617603724822402, + 0.0031530382111668587, + 0.0035168251488357782, + 0.005835919175297022, + 0.004404849372804165, + 0.005711859557777643, + 0.004636622034013271, + 0.005508083384484053, + 0.007457902189344168, + 0.004705311730504036, + 0.009846478700637817, + 0.009599552489817142, + 0.008184090256690979, + 0.012549426406621933, + 0.008561590686440468, + 0.001151195028796792, + 0.0021002686116844416, + 0.0010628511663526297, + 0.00137383583933115, + 0.002195080742239952, + 0.00181127293035388, + 0.0024133631959557533, + 0.0014474847121164203, + 0.002804576186463237, + 0.003282577730715275, + 0.0024010050110518932, + 0.00310880527831614, + 0.0031652310863137245, + 0.0035158719401806593, + 0.005258990917354822, + 0.004439911805093288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.01255018636584282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079810559749603 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079810559749603, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079810559749603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24274596571922302, + 0.26203101873397827, + 0.06744606792926788, + 0.04264130815863609, + 0.11044564843177795, + 0.051228977739810944, + 0.1125275194644928, + 0.05452917143702507, + 0.02401752769947052, + 0.032386817038059235 + ], + "credit_map": [ + 0.021783675998449326, + 0.02935042604804039, + 0.024176783859729767, + 0.03146246820688248, + 0.039380475878715515, + 0.029284849762916565, + 0.04011938348412514, + 0.0271878931671381, + 0.025973215699195862, + 0.03405393287539482, + 0.022335274145007133, + 0.027207976207137108, + 0.039089519530534744, + 0.03194024786353111, + 0.05351600423455238, + 0.0279148668050766, + 0.005193355493247509, + 0.007656638044863939, + 0.005780933890491724, + 0.008474443107843399, + 0.009200217202305794, + 0.008875318802893162, + 0.014418333768844604, + 0.007846825756132603, + 0.0030259101185947657, + 0.006478249095380306, + 0.003828178159892559, + 0.00562048377469182, + 0.00505073880776763, + 0.005568494088947773, + 0.008778692223131657, + 0.0042905667796730995, + 0.011519006453454494, + 0.01087408047169447, + 0.011152583174407482, + 0.01268150843679905, + 0.016159117221832275, + 0.015021152794361115, + 0.019966892898082733, + 0.013071303255856037, + 0.004377118777483702, + 0.006501630414277315, + 0.006665537599474192, + 0.006995969917625189, + 0.006763903424143791, + 0.006612133234739304, + 0.006314747501164675, + 0.0069979410618543625, + 0.00943064596503973, + 0.009683025069534779, + 0.012308834120631218, + 0.016205159947276115, + 0.011171613819897175, + 0.01598098874092102, + 0.020187167450785637, + 0.017560092732310295, + 0.0075930031016469, + 0.005569823086261749, + 0.005857509560883045, + 0.005887850187718868, + 0.008479166775941849, + 0.006650952156633139, + 0.007644303608685732, + 0.006846564821898937, + 0.0032746868673712015, + 0.0025963641237467527, + 0.0023357190657407045, + 0.0033712247386574745, + 0.0033765726257115602, + 0.002680350560694933, + 0.004037071950733662, + 0.0023455375339835882, + 0.002873634919524193, + 0.0035019987262785435, + 0.0029407325200736523, + 0.0055420841090381145, + 0.003520156256854534, + 0.004250061232596636, + 0.006567423697561026, + 0.0031907260417938232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.011079810559749603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337603107094765 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337603107094765, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337603107094765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30970460176467896, + 0.17897722125053406, + 0.09958342462778091, + 0.09236416965723038, + 0.05553990229964256, + 0.08562462031841278, + 0.036191631108522415, + 0.06278771162033081, + 0.03549836203455925, + 0.04372834041714668 + ], + "credit_map": [ + 0.03193755820393562, + 0.03215174376964569, + 0.02831675298511982, + 0.03593968600034714, + 0.045923955738544464, + 0.042110592126846313, + 0.06347598135471344, + 0.029848312959074974, + 0.01107937190681696, + 0.025889676064252853, + 0.01747126132249832, + 0.026479586958885193, + 0.024290792644023895, + 0.02268810011446476, + 0.03247084096074104, + 0.018607590347528458, + 0.006925572641193867, + 0.012248179875314236, + 0.011534973047673702, + 0.0089856693521142, + 0.012257315218448639, + 0.014588774181902409, + 0.022630048915743828, + 0.010412892326712608, + 0.005189531482756138, + 0.012526357546448708, + 0.006576245184987783, + 0.011699710041284561, + 0.013709011487662792, + 0.012149553745985031, + 0.019724708050489426, + 0.010789057239890099, + 0.005903392098844051, + 0.0073857358656823635, + 0.006363854277879, + 0.0053637088276445866, + 0.0072949957102537155, + 0.0073750875890254974, + 0.00848411861807108, + 0.007369013968855143, + 0.010266628116369247, + 0.011105325073003769, + 0.009462830610573292, + 0.009561457671225071, + 0.01255002524703741, + 0.010663420893251896, + 0.013347442261874676, + 0.008667493239045143, + 0.002676135627552867, + 0.006028004456311464, + 0.00256918091326952, + 0.00493152579292655, + 0.005033622961491346, + 0.004777554422616959, + 0.005874402821063995, + 0.00430120388045907, + 0.0065161301754415035, + 0.006524473894387484, + 0.007589056622236967, + 0.00593918189406395, + 0.00903067085891962, + 0.007642792072147131, + 0.010236335918307304, + 0.00930907391011715, + 0.0032342798076570034, + 0.004144554492086172, + 0.0032186678145080805, + 0.004151568282395601, + 0.005680518224835396, + 0.004732825793325901, + 0.006876050960272551, + 0.003459897357970476, + 0.00282629719004035, + 0.005128164775669575, + 0.0037746622692793608, + 0.00751933129504323, + 0.005282989703118801, + 0.0056956373155117035, + 0.009828276932239532, + 0.0036729807034134865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026337603107094765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered entirely in a black-and-white pencil sketch style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a mother sitting in a dark wooden chair with her back to the camera at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old-fashioned desk lamp visible as the sole light source casting a focused beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are knitting needles and a forming scarf visible in the woman's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a large fluffy ball of yarn on a table beside the woman rolling extremely slowly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983498811721802 + ], + "opposite_logprob_tokens": [ + -0.5233498811721802 + ], + "logprob_reward": -0.8983498811721802, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983498811721802, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23180878162384033, + 0.09031610935926437, + 0.09819984436035156, + 0.07700227200984955, + 0.14848867058753967, + 0.10061760246753693, + 0.04665328189730644, + 0.1207987442612648, + 0.05503059923648834, + 0.031084109097719193 + ], + "credit_map": [ + 0.016072459518909454, + 0.027147898450493813, + 0.018017586320638657, + 0.026964304968714714, + 0.04176861047744751, + 0.030042091384530067, + 0.0466914102435112, + 0.02510441467165947, + 0.008864767849445343, + 0.011920950375497341, + 0.010609665885567665, + 0.009714088402688503, + 0.014435313642024994, + 0.010421999730169773, + 0.014765290543437004, + 0.009584021754562855, + 0.006690540350973606, + 0.01625930517911911, + 0.008156749419867992, + 0.01019922737032175, + 0.010753793641924858, + 0.014723834581673145, + 0.019390979781746864, + 0.012025398202240467, + 0.007947014644742012, + 0.00996891874819994, + 0.008421879261732101, + 0.00797945261001587, + 0.011126254685223103, + 0.01028334442526102, + 0.009420446120202541, + 0.011854952201247215, + 0.012043972499668598, + 0.021282171830534935, + 0.011528169736266136, + 0.01741981692612171, + 0.02425263449549675, + 0.01974356546998024, + 0.026766538619995117, + 0.015451771207153797, + 0.01302535180002451, + 0.010960510931909084, + 0.011137156747281551, + 0.011944211088120937, + 0.013385746628046036, + 0.01231506746262312, + 0.01605607382953167, + 0.011793473735451698, + 0.005220036953687668, + 0.0059114727191627026, + 0.005178871098905802, + 0.006131897214800119, + 0.006838319357484579, + 0.00563665060326457, + 0.007068661507219076, + 0.004667368251830339, + 0.012240085750818253, + 0.015021990053355694, + 0.013889498077332973, + 0.011001519858837128, + 0.02044070139527321, + 0.015083069913089275, + 0.018207760527729988, + 0.014914121478796005, + 0.005520819220691919, + 0.006761024706065655, + 0.00818564835935831, + 0.005840136669576168, + 0.007481341250240803, + 0.007017891388386488, + 0.008662187494337559, + 0.005561542231589556, + 0.0029730997048318386, + 0.003662499599158764, + 0.0036511817015707493, + 0.004314338322728872, + 0.0032419145572930574, + 0.003915193025022745, + 0.0056288898922502995, + 0.003696990665048361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.8983498811721802 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_tensor([4075, 4075])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_tensor([4075, 4075])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3e07711a5109178699f9b5be847d87ffd595ba40 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_tensor([4075, 4075])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward zoom shot begins just below the water's surface, then tilts upward and zooms forward, focusing on the center of a fairy-like pond, where a vast, radiant lotus blossoms, gleaming with divine light. The petals of this lotus, a translucent jade-pink hue, are rimmed with a soft golden halo, each one enormous and full. As the camera approaches, the outermost petals gracefully and silently unfold first, followed by the inner petals, layer by layer, in a rhythmic cadence. From the depths of the flower, a warm, golden light emanates, increasingly intense, tinging the mirror-like surface, reflecting a myriad of auspicious clouds, and the surrounding ethereal atmosphere with a sacred hue. When the final layer of petals fully unfolds, the radiance reaches its peak, revealing a delicate child in the center of the lotus seat. His hair is tied in two lovely buns, secured with red ribbons. A touch of vermilion adorns his brow. He wears a bright red silk bellyband, and gleaming gold bracelets adorn his wrists and ankles. He stood firmly with bare feet, holding a flaming spear taller than himself in one hand. The tip of the spear burned with an eternal golden-red flame, casting dynamic light and shadow on his focused and innocent face. The whole picture was shrouded in a solemn, fantastic and vital atmosphere, as if witnessing the birth of a god.", + "seed": 22434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_4075_seed_22434_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09517090762965381, + "vlm_logprob": -0.09517090762965381, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.09517090762965381, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43734249472618103 + ], + "opposite_logprob_tokens": [ + -1.0623425245285034 + ], + "logprob_reward": -0.43734249472618103, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43734249472618103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2930600047111511, + 0.0952378660440445, + 0.20566557347774506, + 0.07066241651773453, + 0.11227940022945404, + 0.07201851159334183, + 0.06798397749662399, + 0.022702720016241074, + 0.026347743347287178, + 0.034041814506053925 + ], + "credit_map": [ + 0.04139281064271927, + 0.042649220675230026, + 0.025152569636702538, + 0.04846368730068207, + 0.03172951936721802, + 0.02879088558256626, + 0.028658650815486908, + 0.04622264578938484, + 0.008994714356958866, + 0.022177347913384438, + 0.010287733748555183, + 0.014899917878210545, + 0.012915941886603832, + 0.007352491375058889, + 0.008391330018639565, + 0.010218387469649315, + 0.03189942240715027, + 0.04083357751369476, + 0.02626762166619301, + 0.01676188036799431, + 0.02464514598250389, + 0.03231469541788101, + 0.017017263919115067, + 0.01592596434056759, + 0.011036211624741554, + 0.006119665689766407, + 0.008520436473190784, + 0.008238366805016994, + 0.00645404402166605, + 0.015620529651641846, + 0.006365112029016018, + 0.008308051154017448, + 0.016224455088377, + 0.015423917211592197, + 0.018539784476161003, + 0.01166140753775835, + 0.013870243914425373, + 0.01474707294255495, + 0.009928924031555653, + 0.01188358198851347, + 0.008719922974705696, + 0.009480084292590618, + 0.011239071376621723, + 0.007358746603131294, + 0.011984044685959816, + 0.0070857275277376175, + 0.00987374596297741, + 0.006277169566601515, + 0.007686156779527664, + 0.008837386965751648, + 0.0061225611716508865, + 0.007024261634796858, + 0.016195176169276237, + 0.008786918595433235, + 0.007666018325835466, + 0.005665498785674572, + 0.0037479354068636894, + 0.0029082822147756815, + 0.003089344361796975, + 0.0020161194261163473, + 0.002242948627099395, + 0.003224759129807353, + 0.0027752157766371965, + 0.0026981146074831486, + 0.003339430782943964, + 0.0032306357752531767, + 0.0030919474083930254, + 0.002596752718091011, + 0.0031585993710905313, + 0.003690923796966672, + 0.002934170188382268, + 0.004305284004658461, + 0.0038255583494901657, + 0.003966589458286762, + 0.004490033257752657, + 0.0030714268796145916, + 0.0031418711878359318, + 0.00593526428565383, + 0.003621523967012763, + 0.005989544093608856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.43734249472618103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003633565502241254 + ], + "opposite_logprob_tokens": [ + -5.628633499145508 + ], + "logprob_reward": -0.003633565502241254, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003633565502241254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16052095592021942, + 0.17802807688713074, + 0.18158915638923645, + 0.08685515820980072, + 0.17456230521202087, + 0.04698546230792999, + 0.08355835825204849, + 0.029763657599687576, + 0.027966508641839027, + 0.030170349404215813 + ], + "credit_map": [ + 0.016241153702139854, + 0.018678652122616768, + 0.010958555154502392, + 0.031867556273937225, + 0.021527903154492378, + 0.01939738728106022, + 0.012228267267346382, + 0.029621465131640434, + 0.01500877644866705, + 0.046206846833229065, + 0.011649660766124725, + 0.014449249021708965, + 0.03074042685329914, + 0.013980166055262089, + 0.013018686324357986, + 0.03297426551580429, + 0.022912204265594482, + 0.03420814126729965, + 0.023040778934955597, + 0.014930563978850842, + 0.018575746566057205, + 0.021781403571367264, + 0.0180318932980299, + 0.028108416125178337, + 0.012945183552801609, + 0.011953353881835938, + 0.011330329813063145, + 0.011993751861155033, + 0.011566048488020897, + 0.012139571830630302, + 0.0068974304012954235, + 0.00802948884665966, + 0.03383506461977959, + 0.01204152125865221, + 0.05224684625864029, + 0.014746845699846745, + 0.014202995225787163, + 0.016778893768787384, + 0.012251007370650768, + 0.0184591393917799, + 0.005920958239585161, + 0.006254424341022968, + 0.010147325694561005, + 0.004260238725692034, + 0.006084946449846029, + 0.005356629844754934, + 0.004973896313458681, + 0.0039870403707027435, + 0.012991977855563164, + 0.008677137084305286, + 0.01180510688573122, + 0.006727481260895729, + 0.013721674680709839, + 0.012734970077872276, + 0.008550625294446945, + 0.008349387906491756, + 0.0038881951477378607, + 0.0033302553929388523, + 0.004347425885498524, + 0.002732724417001009, + 0.003058485686779022, + 0.005680683068931103, + 0.002681621117517352, + 0.00404426921159029, + 0.0033860185649245977, + 0.0031048099044710398, + 0.003362723160535097, + 0.0035188987385481596, + 0.003177360165864229, + 0.003269022097811103, + 0.003221135353669524, + 0.004926539491862059, + 0.004271587356925011, + 0.003662991337478161, + 0.003740716725587845, + 0.0023370757699012756, + 0.003147795330733061, + 0.003719392465427518, + 0.004363808315247297, + 0.004926982335746288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.003633565502241254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181988924741745 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.0181988924741745, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181988924741745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20132744312286377, + 0.08428020030260086, + 0.23858757317066193, + 0.1092613935470581, + 0.12500612437725067, + 0.06680551916360855, + 0.08829981088638306, + 0.03592732176184654, + 0.025684170424938202, + 0.024820419028401375 + ], + "credit_map": [ + 0.0295624528080225, + 0.027453109622001648, + 0.015987584367394447, + 0.03599444031715393, + 0.02695854753255844, + 0.024114331230521202, + 0.019222039729356766, + 0.022034939378499985, + 0.013798677362501621, + 0.012762052938342094, + 0.0070940288715064526, + 0.010882210917770863, + 0.011100652627646923, + 0.009362207725644112, + 0.008320449851453304, + 0.010959921404719353, + 0.03725964576005936, + 0.04943374544382095, + 0.03517431020736694, + 0.020200125873088837, + 0.0290156789124012, + 0.030263932421803474, + 0.018639031797647476, + 0.018601108342409134, + 0.01693250797688961, + 0.011633841320872307, + 0.012690551578998566, + 0.013794731348752975, + 0.012881129048764706, + 0.02277812920510769, + 0.008931401185691357, + 0.00961910467594862, + 0.022028649225831032, + 0.010570633225142956, + 0.04075285419821739, + 0.01240164041519165, + 0.009483764879405499, + 0.011327842250466347, + 0.007655577268451452, + 0.01078516710549593, + 0.008101305924355984, + 0.009384037926793098, + 0.010956759564578533, + 0.005577030126005411, + 0.011144365184009075, + 0.007741083391010761, + 0.007251022383570671, + 0.006649915594607592, + 0.014204766601324081, + 0.01197227742522955, + 0.010285504162311554, + 0.008408146910369396, + 0.014619182795286179, + 0.009093941189348698, + 0.010746470652520657, + 0.008969524875283241, + 0.006822572555392981, + 0.00465159909799695, + 0.005248196888715029, + 0.0037981418427079916, + 0.0028300208505243063, + 0.0037561049684882164, + 0.004587891977280378, + 0.004232794512063265, + 0.0032339277677237988, + 0.004919795785099268, + 0.0027850603219121695, + 0.0034262409899383783, + 0.002912182593718171, + 0.0021559535525739193, + 0.003650539554655552, + 0.002600470557808876, + 0.0031242570839822292, + 0.0038729964289814234, + 0.002772795269265771, + 0.003297507530078292, + 0.0024355093482881784, + 0.002477345522493124, + 0.0034470963291823864, + 0.003392911748960614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0181988924741745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024903970770537853 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024903970770537853, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024903970770537853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20225878059864044, + 0.06302192807197571, + 0.07820398360490799, + 0.04388882964849472, + 0.1518644243478775, + 0.08506816625595093, + 0.15073759853839874, + 0.07085838913917542, + 0.0777881070971489, + 0.07630982249975204 + ], + "credit_map": [ + 0.032528236508369446, + 0.01828918047249317, + 0.01033839862793684, + 0.04288965091109276, + 0.021938256919384003, + 0.022119065746665, + 0.01925957389175892, + 0.034896403551101685, + 0.005591202061623335, + 0.010460290126502514, + 0.004702845122665167, + 0.006900300737470388, + 0.008185572922229767, + 0.007426724769175053, + 0.006005750969052315, + 0.013749241828918457, + 0.010948467068374157, + 0.012243395671248436, + 0.015418789349496365, + 0.006324686110019684, + 0.009291626513004303, + 0.008000879548490047, + 0.006845489609986544, + 0.009130651131272316, + 0.006527247838675976, + 0.005881341639906168, + 0.0054664891213178635, + 0.004448622465133667, + 0.006010759621858597, + 0.006418412085622549, + 0.00388339557684958, + 0.005252559203654528, + 0.02892799861729145, + 0.012029100209474564, + 0.04088161140680313, + 0.012842985801398754, + 0.01266578771173954, + 0.01527739129960537, + 0.011870495043694973, + 0.017369037494063377, + 0.017859918996691704, + 0.005961469374597073, + 0.022647539153695107, + 0.006032236386090517, + 0.00607127184048295, + 0.010978779755532742, + 0.005943162366747856, + 0.00957378651946783, + 0.03002350591123104, + 0.012722683139145374, + 0.018216775730252266, + 0.008826927281916142, + 0.015303976833820343, + 0.03399922326207161, + 0.01163238100707531, + 0.020012110471725464, + 0.007672339212149382, + 0.008919443935155869, + 0.008630461059510708, + 0.008971363306045532, + 0.0074203298427164555, + 0.008296866901218891, + 0.009991168044507504, + 0.01095641776919365, + 0.015520931221544743, + 0.009742191061377525, + 0.009647863917052746, + 0.010299239307641983, + 0.010883545503020287, + 0.007201973348855972, + 0.007881787605583668, + 0.006610572803765535, + 0.011004971340298653, + 0.010124292224645615, + 0.009809998795390129, + 0.005681594368070364, + 0.00633613346144557, + 0.01136180479079485, + 0.008705760352313519, + 0.013285260647535324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0024903970770537853 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189188368618488 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189188368618488, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189188368618488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08227752149105072, + 0.035712964832782745, + 0.22200316190719604, + 0.03172700107097626, + 0.13697871565818787, + 0.035857442766427994, + 0.05165089666843414, + 0.10093051940202713, + 0.11883479356765747, + 0.18402695655822754 + ], + "credit_map": [ + 0.010807090438902378, + 0.007531055714935064, + 0.005376179702579975, + 0.018611108884215355, + 0.008588213473558426, + 0.01047548558562994, + 0.008188551291823387, + 0.012699836865067482, + 0.004417166579514742, + 0.005723381415009499, + 0.0030913574155420065, + 0.005065295845270157, + 0.005363973788917065, + 0.004220110829919577, + 0.003610159270465374, + 0.004221518523991108, + 0.03032750077545643, + 0.031697969883680344, + 0.045262984931468964, + 0.018452182412147522, + 0.026621874421834946, + 0.02580908127129078, + 0.019045181572437286, + 0.024786366149783134, + 0.005160721950232983, + 0.0027582996990531683, + 0.0029298067092895508, + 0.0035588343162089586, + 0.002656833501532674, + 0.007866855710744858, + 0.0029463304672390223, + 0.003849320113658905, + 0.02518101967871189, + 0.013441251590847969, + 0.033828504383563995, + 0.010842720046639442, + 0.01325594075024128, + 0.014049731194972992, + 0.00990109983831644, + 0.016478458419442177, + 0.007089919410645962, + 0.004092467948794365, + 0.006982238031923771, + 0.0026551487389951944, + 0.0048982142470777035, + 0.004382019862532616, + 0.002242671325802803, + 0.003514760173857212, + 0.006594907492399216, + 0.007305950857698917, + 0.004393165931105614, + 0.003537493059411645, + 0.011913108639419079, + 0.008849265985190868, + 0.005735830403864384, + 0.003321175230666995, + 0.01814349554479122, + 0.014108926989138126, + 0.0177452452480793, + 0.008230522274971008, + 0.008164335042238235, + 0.011535140685737133, + 0.009773206897079945, + 0.013229651376605034, + 0.01812470704317093, + 0.013852301053702831, + 0.0130548644810915, + 0.018432632088661194, + 0.02066473849117756, + 0.00994893442839384, + 0.014990489929914474, + 0.009766126051545143, + 0.024859676137566566, + 0.035332560539245605, + 0.02128715068101883, + 0.01867486909031868, + 0.019343998283147812, + 0.021898752078413963, + 0.021699383854866028, + 0.020930571481585503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.014189188368618488 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward zoom shot begins just below the water's surface, then tilts upward and zooms forward, focusing on the center of a fairy-like pond, where a vast, radiant lotus blossoms, gleaming with divine light. The petals of this lotus, a translucent jade-pink hue, are rimmed with a soft golden halo, each one enormous and full. As the camera approaches, the outermost petals gracefully and silently unfold first, followed by the inner petals, layer by layer, in a rhythmic cadence. From the depths of the flower, a warm, golden light emanates, increasingly intense, tinging the mirror-like surface, reflecting a myriad of auspicious clouds, and the surrounding ethereal atmosphere with a sacred hue. When the final layer of petals fully unfolds, the radiance reaches its peak, revealing a delicate child in the center of the lotus seat. His hair is tied in two lovely buns, secured with red ribbons. A touch of vermilion adorns his brow. He wears a bright red silk bellyband, and gleaming gold bracelets adorn his wrists and ankles. He stood firmly with bare feet, holding a flaming spear taller than himself in one hand. The tip of the spear burned with an eternal golden-red flame, casting dynamic light and shadow on his focused and innocent face. The whole picture was shrouded in a solemn, fantastic and vital atmosphere, as if witnessing the birth of a god.", + "seed": 22435, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_4075_seed_22435_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14835559725761413, + "vlm_logprob": -0.14835559725761413, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.14835559725761413, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6991982460021973 + ], + "opposite_logprob_tokens": [ + -0.6991982460021973 + ], + "logprob_reward": -0.6991982460021973, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6991982460021973, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09868171811103821, + 0.07861479371786118, + 0.5247662663459778, + 0.14803923666477203, + 0.022803114727139473, + 0.016183335334062576, + 0.05655097961425781, + 0.042429059743881226, + 0.00494510168209672, + 0.006986487656831741 + ], + "credit_map": [ + 0.011028925888240337, + 0.00867470819503069, + 0.021051792427897453, + 0.01541710365563631, + 0.006238595582544804, + 0.008618245832622051, + 0.008919527754187584, + 0.01873280107975006, + 0.01105749886482954, + 0.009144063107669353, + 0.012159999459981918, + 0.007631620857864618, + 0.009786552749574184, + 0.010135279037058353, + 0.006992056500166655, + 0.011707708239555359, + 0.042566195130348206, + 0.01136526558548212, + 0.17714907228946686, + 0.02660761959850788, + 0.01403872948139906, + 0.2056189626455307, + 0.015362318605184555, + 0.03205801919102669, + 0.010142204351723194, + 0.0038914703764021397, + 0.052688922733068466, + 0.013095281086862087, + 0.005935703404247761, + 0.04812072217464447, + 0.0063663022592663765, + 0.007798612117767334, + 0.0014141311403363943, + 0.0050907861441373825, + 0.005131706595420837, + 0.002526658121496439, + 0.0023802313953638077, + 0.0016497924225404859, + 0.0019804048351943493, + 0.002629401395097375, + 0.0014452580362558365, + 0.0016500061610713601, + 0.002388560911640525, + 0.0022117053158581257, + 0.0017340766498818994, + 0.0018822746351361275, + 0.001735224504955113, + 0.0031362269073724747, + 0.004644142929464579, + 0.013902956619858742, + 0.0057142348960042, + 0.005345537792891264, + 0.009411873295903206, + 0.006368228700011969, + 0.005315819755196571, + 0.005848180502653122, + 0.004189319908618927, + 0.010264390148222446, + 0.004752301145344973, + 0.007908432744443417, + 0.004632141906768084, + 0.002331904601305723, + 0.004171101842075586, + 0.004179462790489197, + 0.0005206629866734147, + 0.0006592621211893857, + 0.0006444748141802847, + 0.0007372082909569144, + 0.00045438477536663413, + 0.0006015421822667122, + 0.0004876232414972037, + 0.0008399426005780697, + 0.0007221559644676745, + 0.001437660655938089, + 0.0006686949054710567, + 0.000825185445137322, + 0.0008936534868553281, + 0.0007649109466001391, + 0.0007975142798386514, + 0.000876711099408567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.6991982460021973 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003629883285611868 + ], + "opposite_logprob_tokens": [ + -5.628629684448242 + ], + "logprob_reward": -0.003629883285611868, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003629883285611868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0667576938867569, + 0.0532015897333622, + 0.454337477684021, + 0.1555081009864807, + 0.06082065403461456, + 0.05755624175071716, + 0.025808487087488174, + 0.056443434208631516, + 0.05046803131699562, + 0.019098253920674324 + ], + "credit_map": [ + 0.010534986853599548, + 0.008538094349205494, + 0.00897982344031334, + 0.008141854777932167, + 0.0059694284573197365, + 0.006712672300636768, + 0.0067519256845116615, + 0.011128900572657585, + 0.008634218946099281, + 0.007783536799252033, + 0.006799571216106415, + 0.006476191338151693, + 0.006616292055696249, + 0.0060351635329425335, + 0.005562067497521639, + 0.0052945492789149284, + 0.03628865256905556, + 0.01135631836950779, + 0.15497390925884247, + 0.025782084092497826, + 0.012368588708341122, + 0.1734842211008072, + 0.013574461452662945, + 0.0265092384070158, + 0.010744393803179264, + 0.0037077998276799917, + 0.05380311235785484, + 0.008969327434897423, + 0.005799584090709686, + 0.05566877871751785, + 0.00751758087426424, + 0.009297529235482216, + 0.005519659724086523, + 0.009669138118624687, + 0.009592597372829914, + 0.007207079324871302, + 0.006444169208407402, + 0.006132467649877071, + 0.007269533351063728, + 0.00898601021617651, + 0.005642826668918133, + 0.009683765470981598, + 0.009674466215074062, + 0.007725996430963278, + 0.005049355328083038, + 0.005225494038313627, + 0.005782974883913994, + 0.008771364577114582, + 0.001963977934792638, + 0.004215696360915899, + 0.004338650032877922, + 0.0026613955851644278, + 0.002500060945749283, + 0.0021523311734199524, + 0.003181047970429063, + 0.004795326851308346, + 0.006934983655810356, + 0.010702421888709068, + 0.005228525027632713, + 0.00977180153131485, + 0.006638977210968733, + 0.004100167658179998, + 0.006462861318141222, + 0.006603698246181011, + 0.0036446647718548775, + 0.011014222167432308, + 0.005983099341392517, + 0.007798663806170225, + 0.007928344421088696, + 0.004430576227605343, + 0.004910973832011223, + 0.004757489077746868, + 0.0013825620990246534, + 0.0028214233461767435, + 0.0023024098481982946, + 0.0021545225754380226, + 0.0019391168607398868, + 0.003092096885666251, + 0.0025905524380505085, + 0.0028155699837952852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.003629883285611868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356877759099007 + ], + "opposite_logprob_tokens": [ + -3.6513569355010986 + ], + "logprob_reward": -0.026356877759099007, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356877759099007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04522789269685745, + 0.039059773087501526, + 0.5841039419174194, + 0.1699029952287674, + 0.03206842392683029, + 0.024104004725813866, + 0.03445615619421005, + 0.03059045784175396, + 0.025179509073495865, + 0.01530686765909195 + ], + "credit_map": [ + 0.004657725803554058, + 0.005594039801508188, + 0.007076984737068415, + 0.005948756355792284, + 0.004460679832845926, + 0.007436350453644991, + 0.003952373284846544, + 0.006100986618548632, + 0.004193066619336605, + 0.005085598211735487, + 0.005881470162421465, + 0.004437945783138275, + 0.004395151976495981, + 0.005220641847699881, + 0.004350225906819105, + 0.005495673045516014, + 0.04734661430120468, + 0.01219437550753355, + 0.19642402231693268, + 0.02956230379641056, + 0.015446734614670277, + 0.23080018162727356, + 0.016354981809854507, + 0.035974759608507156, + 0.01145076658576727, + 0.004813682287931442, + 0.05532339587807655, + 0.01582755520939827, + 0.007244447711855173, + 0.05585610494017601, + 0.006978198885917664, + 0.012408860959112644, + 0.002997919451445341, + 0.005026292521506548, + 0.003568080021068454, + 0.005221198312938213, + 0.0031103771179914474, + 0.003314135828986764, + 0.004263665061444044, + 0.004566757939755917, + 0.002108885906636715, + 0.0021079762373119593, + 0.003834191942587495, + 0.0031177953351289034, + 0.001921588322147727, + 0.0025494976434856653, + 0.0034866707865148783, + 0.004977400880306959, + 0.003184097120538354, + 0.007910769432783127, + 0.003581652883440256, + 0.003390605328604579, + 0.004892469383776188, + 0.003556100884452462, + 0.003956106025725603, + 0.003984356299042702, + 0.002148773754015565, + 0.007266402710229158, + 0.0032110412139445543, + 0.004390899091959, + 0.004078424535691738, + 0.0021273428574204445, + 0.0033759656362235546, + 0.003991610836237669, + 0.0025765872560441494, + 0.00354590336792171, + 0.0034791447687894106, + 0.004342598374933004, + 0.0025424049235880375, + 0.0024141487665474415, + 0.0026226937770843506, + 0.003656029235571623, + 0.0018809299217537045, + 0.0029191041830927134, + 0.0017417838098481297, + 0.00186753342859447, + 0.0016293905209749937, + 0.001681690220721066, + 0.0016435921424999833, + 0.001942845294252038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.026356877759099007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028177103959023952 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028177103959023952, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028177103959023952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0526864267885685, + 0.033203739672899246, + 0.40383508801460266, + 0.21605446934700012, + 0.03923886641860008, + 0.03128688409924507, + 0.06212778761982918, + 0.07244831323623657, + 0.03666131943464279, + 0.0524570532143116 + ], + "credit_map": [ + 0.007443632930517197, + 0.004565895535051823, + 0.011680510826408863, + 0.004916450474411249, + 0.002854571444913745, + 0.008300241082906723, + 0.003481825115159154, + 0.009443300776183605, + 0.0060021718963980675, + 0.0037181286606937647, + 0.004178476054221392, + 0.0031768670305609703, + 0.004263291135430336, + 0.00423006433993578, + 0.002650526352226734, + 0.004984218627214432, + 0.032962214201688766, + 0.009114299900829792, + 0.13439658284187317, + 0.02056743949651718, + 0.011146650649607182, + 0.15947876870632172, + 0.011382844299077988, + 0.024786299094557762, + 0.014291957952082157, + 0.005841855425387621, + 0.06752507388591766, + 0.0214831605553627, + 0.008148297667503357, + 0.07082197815179825, + 0.00879459735006094, + 0.019147558137774467, + 0.0031957656610757113, + 0.007979940623044968, + 0.003503075335174799, + 0.003954473417252302, + 0.0037065634969621897, + 0.003162272274494171, + 0.0036969794891774654, + 0.01003979705274105, + 0.0037815070245414972, + 0.006392461247742176, + 0.003187506692484021, + 0.00339105655439198, + 0.00339863752014935, + 0.0036976721603423357, + 0.0026530760806053877, + 0.004784967750310898, + 0.005816118326038122, + 0.011545369401574135, + 0.008512166328728199, + 0.007019322365522385, + 0.007309939246624708, + 0.006715560331940651, + 0.007917512208223343, + 0.007291802670806646, + 0.004836123436689377, + 0.021991945803165436, + 0.006751737557351589, + 0.009346812032163143, + 0.010460156947374344, + 0.0054933479987084866, + 0.007977168075740337, + 0.005591024179011583, + 0.0029241645243018866, + 0.005449654534459114, + 0.004792907275259495, + 0.008178350515663624, + 0.004087719600647688, + 0.0037697902880609035, + 0.003750308882445097, + 0.0037084240466356277, + 0.006270640529692173, + 0.010200027376413345, + 0.006340370513498783, + 0.005208230577409267, + 0.006321065127849579, + 0.005219451151788235, + 0.00437057763338089, + 0.008526692166924477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.0028177103959023952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775268845260143 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775268845260143, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775268845260143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05469939485192299, + 0.059303466230630875, + 0.5780116319656372, + 0.022019565105438232, + 0.037150781601667404, + 0.01783151924610138, + 0.049637384712696075, + 0.08756577968597412, + 0.05859745666384697, + 0.03518299385905266 + ], + "credit_map": [ + 0.006089663598686457, + 0.007412780076265335, + 0.00951346755027771, + 0.005296436604112387, + 0.006241186987608671, + 0.00545453280210495, + 0.004685929510742426, + 0.010005400516092777, + 0.007585473824292421, + 0.007703687064349651, + 0.008180949836969376, + 0.006372989621013403, + 0.008161576464772224, + 0.007807154208421707, + 0.0061871763318777084, + 0.007304466795176268, + 0.0467040091753006, + 0.013234064914286137, + 0.1933022290468216, + 0.03173597902059555, + 0.01494065672159195, + 0.22688204050064087, + 0.016509942710399628, + 0.03470275178551674, + 0.0016162018291652203, + 0.0013073096051812172, + 0.005919830407947302, + 0.0022046007215976715, + 0.001475393190048635, + 0.0035187839530408382, + 0.002087472239509225, + 0.003889975370839238, + 0.002541568595916033, + 0.0061000860296189785, + 0.012001285329461098, + 0.002912969794124365, + 0.0027849439065903425, + 0.0024211308918893337, + 0.0037483007181435823, + 0.004640497732907534, + 0.0016671346966177225, + 0.0018555548740550876, + 0.003375105559825897, + 0.002366220811381936, + 0.0015566821675747633, + 0.001775902695953846, + 0.0028792961966246367, + 0.0023556228261440992, + 0.003525325795635581, + 0.01199561171233654, + 0.0078002638183534145, + 0.0042574722319841385, + 0.005946587771177292, + 0.003241774160414934, + 0.004995857365429401, + 0.007874496281147003, + 0.0060209305956959724, + 0.021344268694519997, + 0.006487565580755472, + 0.018026359379291534, + 0.01230651419609785, + 0.005582356359809637, + 0.010709268972277641, + 0.007088521029800177, + 0.005197535734623671, + 0.010778346098959446, + 0.0062661669217050076, + 0.013276012614369392, + 0.005932348780333996, + 0.0045369346626102924, + 0.005603109020739794, + 0.007007007487118244, + 0.003506693057715893, + 0.005906412843614817, + 0.0036079734563827515, + 0.003549174638465047, + 0.004349640570580959, + 0.0037447065114974976, + 0.00460868701338768, + 0.00590970553457737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.009775268845260143 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward zoom shot begins just below the water's surface, then tilts upward and zooms forward, focusing on the center of a fairy-like pond, where a vast, radiant lotus blossoms, gleaming with divine light. The petals of this lotus, a translucent jade-pink hue, are rimmed with a soft golden halo, each one enormous and full. As the camera approaches, the outermost petals gracefully and silently unfold first, followed by the inner petals, layer by layer, in a rhythmic cadence. From the depths of the flower, a warm, golden light emanates, increasingly intense, tinging the mirror-like surface, reflecting a myriad of auspicious clouds, and the surrounding ethereal atmosphere with a sacred hue. When the final layer of petals fully unfolds, the radiance reaches its peak, revealing a delicate child in the center of the lotus seat. His hair is tied in two lovely buns, secured with red ribbons. A touch of vermilion adorns his brow. He wears a bright red silk bellyband, and gleaming gold bracelets adorn his wrists and ankles. He stood firmly with bare feet, holding a flaming spear taller than himself in one hand. The tip of the spear burned with an eternal golden-red flame, casting dynamic light and shadow on his focused and innocent face. The whole picture was shrouded in a solemn, fantastic and vital atmosphere, as if witnessing the birth of a god.", + "seed": 22436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_4075_seed_22436_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0992148614488542, + "vlm_logprob": -0.0992148614488542, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.0992148614488542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47810977697372437 + ], + "opposite_logprob_tokens": [ + -0.9781097769737244 + ], + "logprob_reward": -0.47810977697372437, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47810977697372437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05603852868080139, + 0.16989445686340332, + 0.4693005681037903, + 0.14543354511260986, + 0.030082030221819878, + 0.045568592846393585, + 0.019643371924757957, + 0.03605397790670395, + 0.01699109561741352, + 0.010993804782629013 + ], + "credit_map": [ + 0.0062477621249854565, + 0.006860002875328064, + 0.015697894617915154, + 0.005250708665698767, + 0.006747079081833363, + 0.005751791875809431, + 0.004100766032934189, + 0.0053825220093131065, + 0.02734510414302349, + 0.02405690960586071, + 0.015283930115401745, + 0.021255191415548325, + 0.019965562969446182, + 0.015338893048465252, + 0.015622577629983425, + 0.031026283279061317, + 0.03386566415429115, + 0.0518980547785759, + 0.03893962875008583, + 0.07787300646305084, + 0.03879593685269356, + 0.050421811640262604, + 0.14216449856758118, + 0.03534197434782982, + 0.019408635795116425, + 0.022262992337346077, + 0.018992464989423752, + 0.01865866221487522, + 0.019917719066143036, + 0.015223563648760319, + 0.01238246913999319, + 0.018587037920951843, + 0.004817300476133823, + 0.0025179057847708464, + 0.003526468761265278, + 0.003562931902706623, + 0.002296103397384286, + 0.003788100788369775, + 0.0045718541368842125, + 0.005001362878829241, + 0.008717089891433716, + 0.004141169134527445, + 0.004226613324135542, + 0.006436992436647415, + 0.005030015483498573, + 0.005507319234311581, + 0.0035959461238235235, + 0.007913446053862572, + 0.002917576814070344, + 0.0019391566747799516, + 0.0027589586097747087, + 0.002018040046095848, + 0.0015917844139039516, + 0.002828099997714162, + 0.0025575843174010515, + 0.003032171633094549, + 0.004287736490368843, + 0.0026380415074527264, + 0.006194197107106447, + 0.005974260158836842, + 0.002242449438199401, + 0.003276066854596138, + 0.00543338805437088, + 0.00600783945992589, + 0.0019151584710925817, + 0.0016902161296457052, + 0.002471158280968666, + 0.0019445415819063783, + 0.0013728064950555563, + 0.0031513452995568514, + 0.0022534055169671774, + 0.0021924632601439953, + 0.0009767188457772136, + 0.0014746702509000897, + 0.0016228175954893231, + 0.0013496035244315863, + 0.001528229215182364, + 0.0012048259377479553, + 0.001331974403001368, + 0.0015049652429297566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.47810977697372437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032062334939837456 + ], + "opposite_logprob_tokens": [ + -5.753206253051758 + ], + "logprob_reward": -0.0032062334939837456, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032062334939837456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0879754051566124, + 0.095028355717659, + 0.5948148369789124, + 0.08103587478399277, + 0.03624686598777771, + 0.03294195234775543, + 0.01565762422978878, + 0.021852558478713036, + 0.021552512422204018, + 0.012893942184746265 + ], + "credit_map": [ + 0.013952071778476238, + 0.008546462282538414, + 0.01594105362892151, + 0.006060012150555849, + 0.007357701193541288, + 0.008321652188897133, + 0.004863333888351917, + 0.022933121770620346, + 0.009902095422148705, + 0.01453800406306982, + 0.007897162809967995, + 0.01853475160896778, + 0.01166261825710535, + 0.008709668181836605, + 0.008336911909282207, + 0.015447140671312809, + 0.04590923711657524, + 0.06214078888297081, + 0.044239919632673264, + 0.10241112858057022, + 0.0479961633682251, + 0.06784673780202866, + 0.1797971874475479, + 0.04447369650006294, + 0.005847037769854069, + 0.02221776358783245, + 0.009088621474802494, + 0.010699959471821785, + 0.010790822096168995, + 0.00736549636349082, + 0.0074612791649997234, + 0.007564895786345005, + 0.0037781603168696165, + 0.004121606238186359, + 0.004239261616021395, + 0.004967999644577503, + 0.004073436837643385, + 0.0030629406683146954, + 0.005458892323076725, + 0.0065445676445961, + 0.006319512613117695, + 0.0016541631193831563, + 0.004691343288868666, + 0.0036968698259443045, + 0.002474575536325574, + 0.005706549622118473, + 0.0031358010601252317, + 0.005263139959424734, + 0.002496246248483658, + 0.0016255375230684876, + 0.002598606050014496, + 0.0014590071514248848, + 0.0013404851779341698, + 0.002133181318640709, + 0.002113793743774295, + 0.0018907663179561496, + 0.0031491408590227365, + 0.0022678314708173275, + 0.00414876826107502, + 0.0027481606230139732, + 0.0016787484055384994, + 0.0021846480667591095, + 0.003135901875793934, + 0.0025393585674464703, + 0.0022020735777914524, + 0.0019517778418958187, + 0.0023543427232652903, + 0.0035418926272541285, + 0.0019455092260614038, + 0.0038820381741970778, + 0.003181099658831954, + 0.002493777545168996, + 0.001205942127853632, + 0.001130873104557395, + 0.0021738698706030846, + 0.0015508546493947506, + 0.001113080303184688, + 0.001481951680034399, + 0.0014668686781078577, + 0.002770502818748355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0032062334939837456 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634746376425028 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634746376425028, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634746376425028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030798470601439476, + 0.04778911545872688, + 0.5812953114509583, + 0.1291489154100418, + 0.023400435224175453, + 0.1051083654165268, + 0.027948228642344475, + 0.022605163976550102, + 0.012525022029876709, + 0.01938096433877945 + ], + "credit_map": [ + 0.003181398380547762, + 0.004063088446855545, + 0.004515223670750856, + 0.0036731474101543427, + 0.004722595680505037, + 0.0041593704372644424, + 0.0035849520936608315, + 0.002898693084716797, + 0.005248718429356813, + 0.006576525047421455, + 0.004619041457772255, + 0.010934179648756981, + 0.0038162816781550646, + 0.004800414200872183, + 0.003920665942132473, + 0.007873286493122578, + 0.04137883335351944, + 0.06913743913173676, + 0.0403454452753067, + 0.09918598085641861, + 0.04434183984994888, + 0.053744278848171234, + 0.19100305438041687, + 0.04215845838189125, + 0.009028389118611813, + 0.029393093660473824, + 0.019469885155558586, + 0.018087953329086304, + 0.016152886673808098, + 0.00904818531125784, + 0.013433286920189857, + 0.014535226859152317, + 0.0021456771064549685, + 0.003498026868328452, + 0.0026252628304064274, + 0.0033830490428954363, + 0.0030596312135457993, + 0.001745952875353396, + 0.0038680844008922577, + 0.003074751468375325, + 0.018462494015693665, + 0.0036950823850929737, + 0.012443041428923607, + 0.042872779071331024, + 0.0038807829841971397, + 0.004248351790010929, + 0.00543298851698637, + 0.014072845689952374, + 0.0048812273889780045, + 0.003316479502245784, + 0.003986224066466093, + 0.003180865664035082, + 0.0019302923465147614, + 0.0036551980301737785, + 0.002650211798027158, + 0.00434773089364171, + 0.0026481507811695337, + 0.0019822968170046806, + 0.004269815515726805, + 0.003345602424815297, + 0.0016399123705923557, + 0.0018007552716881037, + 0.00402867654338479, + 0.0028899535536766052, + 0.0016796063864603639, + 0.0013377689756453037, + 0.001493894960731268, + 0.0021125697530806065, + 0.0012893342645838857, + 0.001953299855813384, + 0.001613881322555244, + 0.0010446665110066533, + 0.0020554226357489824, + 0.0027053735684603453, + 0.002733686938881874, + 0.002601365791633725, + 0.002487892284989357, + 0.0016753535019233823, + 0.002277917228639126, + 0.002843952737748623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.007634746376425028 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011793570592999458 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.0011793570592999458, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011793570592999458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13205713033676147, + 0.04503992944955826, + 0.4030567407608032, + 0.0562431775033474, + 0.059973087161779404, + 0.16930295526981354, + 0.024958867579698563, + 0.02294880710542202, + 0.02186465449631214, + 0.06455463171005249 + ], + "credit_map": [ + 0.022069497033953667, + 0.012048195116221905, + 0.026471000164747238, + 0.007698237430304289, + 0.010276454500854015, + 0.009386073797941208, + 0.005864416249096394, + 0.03824325278401375, + 0.00491638109087944, + 0.005618898663669825, + 0.0049550593830645084, + 0.008195382542908192, + 0.0049737184308469296, + 0.006613948382437229, + 0.004400553181767464, + 0.005365987773984671, + 0.027561629191040993, + 0.04805157333612442, + 0.027493610978126526, + 0.06851807981729507, + 0.029878752306103706, + 0.03745734691619873, + 0.1357962191104889, + 0.028299545869231224, + 0.0052445740438997746, + 0.007444234099239111, + 0.008333341218531132, + 0.008106337860226631, + 0.006409379653632641, + 0.004781597293913364, + 0.0064520095475018024, + 0.009471699595451355, + 0.005031336564570665, + 0.006753457710146904, + 0.007471241522580385, + 0.008459101431071758, + 0.007625071331858635, + 0.007841160520911217, + 0.006784671917557716, + 0.010007047094404697, + 0.03046759031713009, + 0.007814411073923111, + 0.01882769539952278, + 0.06758071482181549, + 0.005381506402045488, + 0.0084215197712183, + 0.009638110175728798, + 0.02117140218615532, + 0.003976404666900635, + 0.0031911958940327168, + 0.0040044053457677364, + 0.0023681840393692255, + 0.0017971096094697714, + 0.0034499294124543667, + 0.0029223812744021416, + 0.003249255008995533, + 0.002235210034996271, + 0.00282350298948586, + 0.0032993415370583534, + 0.0035546226426959038, + 0.001761826453730464, + 0.002530444646254182, + 0.004039385821670294, + 0.002704472979530692, + 0.0024348220322281122, + 0.0024485564790666103, + 0.0023555923253297806, + 0.003735761623829603, + 0.002108623506501317, + 0.0029965469148010015, + 0.003099861554801464, + 0.0026848907582461834, + 0.005372419022023678, + 0.009179452434182167, + 0.009100671857595444, + 0.00990831945091486, + 0.0066534229554235935, + 0.00823838822543621, + 0.008584972470998764, + 0.007516984362155199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.0011793570592999458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059441933408379555 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.0059441933408379555, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059441933408379555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10386285930871964, + 0.036575980484485626, + 0.15791048109531403, + 0.07181928306818008, + 0.03264528512954712, + 0.3718821108341217, + 0.06082496792078018, + 0.07026739418506622, + 0.0390104316174984, + 0.0552012100815773 + ], + "credit_map": [ + 0.019082173705101013, + 0.009215202182531357, + 0.01976141147315502, + 0.005269540473818779, + 0.007528891321271658, + 0.007871926762163639, + 0.004129951354116201, + 0.031003767624497414, + 0.0039395615458488464, + 0.005045884754508734, + 0.0038139335811138153, + 0.007657913491129875, + 0.0032485327683389187, + 0.004434324335306883, + 0.0031848554499447346, + 0.005250977352261543, + 0.011599384248256683, + 0.019172517582774162, + 0.010850599966943264, + 0.027288267388939857, + 0.012357235886156559, + 0.01493841502815485, + 0.05009967088699341, + 0.011604384519159794, + 0.005294851493090391, + 0.007967671379446983, + 0.011798792518675327, + 0.013059854507446289, + 0.007017661817371845, + 0.005772003438323736, + 0.010420478880405426, + 0.010487966239452362, + 0.004261847585439682, + 0.004103058483451605, + 0.0038047146517783403, + 0.004971824120730162, + 0.004597312770783901, + 0.003401047782972455, + 0.0036917794495821, + 0.003813697025179863, + 0.07181469351053238, + 0.013268250040709972, + 0.04199235886335373, + 0.15420803427696228, + 0.007361769210547209, + 0.015589527785778046, + 0.018966196104884148, + 0.048681288957595825, + 0.007724154274910688, + 0.0086296945810318, + 0.006925809197127819, + 0.008397617377340794, + 0.005961350630968809, + 0.007484446279704571, + 0.01050599105656147, + 0.005195907317101955, + 0.01184176653623581, + 0.004252033773809671, + 0.021321948617696762, + 0.006869081873446703, + 0.0032758712768554688, + 0.005458652041852474, + 0.008905809372663498, + 0.008342228829860687, + 0.0035898827482014894, + 0.005492928437888622, + 0.004048105329275131, + 0.007371870800852776, + 0.004390772897750139, + 0.005362248979508877, + 0.005879831500351429, + 0.002874793019145727, + 0.005599255207926035, + 0.005969819147139788, + 0.009113607928156853, + 0.005905388854444027, + 0.007873483933508396, + 0.004728313069790602, + 0.006166695151478052, + 0.009844646789133549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.0059441933408379555 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward zoom shot begins just below the water's surface, then tilts upward and zooms forward, focusing on the center of a fairy-like pond, where a vast, radiant lotus blossoms, gleaming with divine light. The petals of this lotus, a translucent jade-pink hue, are rimmed with a soft golden halo, each one enormous and full. As the camera approaches, the outermost petals gracefully and silently unfold first, followed by the inner petals, layer by layer, in a rhythmic cadence. From the depths of the flower, a warm, golden light emanates, increasingly intense, tinging the mirror-like surface, reflecting a myriad of auspicious clouds, and the surrounding ethereal atmosphere with a sacred hue. When the final layer of petals fully unfolds, the radiance reaches its peak, revealing a delicate child in the center of the lotus seat. His hair is tied in two lovely buns, secured with red ribbons. A touch of vermilion adorns his brow. He wears a bright red silk bellyband, and gleaming gold bracelets adorn his wrists and ankles. He stood firmly with bare feet, holding a flaming spear taller than himself in one hand. The tip of the spear burned with an eternal golden-red flame, casting dynamic light and shadow on his focused and innocent face. The whole picture was shrouded in a solemn, fantastic and vital atmosphere, as if witnessing the birth of a god.", + "seed": 22437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_7/index_4075_seed_22437_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10192872202023864, + "vlm_logprob": -0.10192872202023864, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.10192872202023864, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4788968563079834 + ], + "opposite_logprob_tokens": [ + -0.9788968563079834 + ], + "logprob_reward": -0.4788968563079834, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4788968563079834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12085861712694168, + 0.07999278604984283, + 0.029140407219529152, + 0.1098211333155632, + 0.3143560290336609, + 0.057491105049848557, + 0.06690499186515808, + 0.09548139572143555, + 0.08131293207406998, + 0.044640593230724335 + ], + "credit_map": [ + 0.014920772984623909, + 0.02739638276398182, + 0.010228447616100311, + 0.009460795670747757, + 0.028873348608613014, + 0.011633995920419693, + 0.01022484339773655, + 0.008120051585137844, + 0.010461805388331413, + 0.010690913535654545, + 0.00810688640922308, + 0.007684154901653528, + 0.015088859014213085, + 0.010396147146821022, + 0.009531057439744473, + 0.008032966405153275, + 0.0038887534756213427, + 0.004793297965079546, + 0.002848271280527115, + 0.0028896222356706858, + 0.005337828304618597, + 0.0032786743249744177, + 0.003835271345451474, + 0.0022686924785375595, + 0.007303119637072086, + 0.024536745622754097, + 0.007433858700096607, + 0.015338052064180374, + 0.025198573246598244, + 0.011999744921922684, + 0.010490461252629757, + 0.007520584389567375, + 0.017870604991912842, + 0.022335901856422424, + 0.02132452093064785, + 0.11874669790267944, + 0.09028404206037521, + 0.013665971346199512, + 0.017173923552036285, + 0.01295437291264534, + 0.010495824739336967, + 0.006885663140565157, + 0.007245969492942095, + 0.007947398349642754, + 0.006715541239827871, + 0.00746344355866313, + 0.006237495224922895, + 0.004499770235270262, + 0.010653221048414707, + 0.00622576754540205, + 0.009353699162602425, + 0.00981263816356659, + 0.008502164855599403, + 0.007611307315528393, + 0.007728537078946829, + 0.007017667405307293, + 0.016423402354121208, + 0.006695236545056105, + 0.030682824552059174, + 0.009915506467223167, + 0.009723099879920483, + 0.00541362538933754, + 0.007321858778595924, + 0.009305855259299278, + 0.01302108634263277, + 0.008206945843994617, + 0.01044122688472271, + 0.007472331169992685, + 0.008085394278168678, + 0.01483163982629776, + 0.006498008966445923, + 0.012756302021443844, + 0.008049710653722286, + 0.005990127567201853, + 0.0046523031778633595, + 0.004384393338114023, + 0.0051675341092050076, + 0.005656552035361528, + 0.005854307673871517, + 0.004885668400675058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.4788968563079834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108916502445936 + ], + "opposite_logprob_tokens": [ + -5.504108905792236 + ], + "logprob_reward": -0.004108916502445936, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108916502445936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1916625201702118, + 0.09818430244922638, + 0.07535485178232193, + 0.07559839636087418, + 0.08200426399707794, + 0.04868115484714508, + 0.061336446553468704, + 0.18620876967906952, + 0.09138727188110352, + 0.08958204090595245 + ], + "credit_map": [ + 0.01709696277976036, + 0.04182566702365875, + 0.015144824050366879, + 0.01392494048923254, + 0.054336968809366226, + 0.019475486129522324, + 0.01714230887591839, + 0.0127153554931283, + 0.010111909359693527, + 0.015826933085918427, + 0.0077933091670274734, + 0.007108899299055338, + 0.02711007185280323, + 0.011708181351423264, + 0.009949494153261185, + 0.008575503714382648, + 0.006021724548190832, + 0.016697734594345093, + 0.006038003135472536, + 0.0052266051061451435, + 0.02038651891052723, + 0.00822195690125227, + 0.0076514664106070995, + 0.005110833793878555, + 0.00732378987595439, + 0.012072792276740074, + 0.011509167961776257, + 0.007424567360430956, + 0.009639497846364975, + 0.009403900243341923, + 0.00953278224915266, + 0.008691897615790367, + 0.012120605446398258, + 0.006526906508952379, + 0.006283365655690432, + 0.012555460445582867, + 0.02014736272394657, + 0.008080258034169674, + 0.007592229638248682, + 0.00869806483387947, + 0.008216125890612602, + 0.006999046076089144, + 0.0045449500903487206, + 0.0055993953719735146, + 0.005004578270018101, + 0.005878683179616928, + 0.00725026149302721, + 0.005188112612813711, + 0.008141213096678257, + 0.006861353758722544, + 0.006870789919048548, + 0.00807715393602848, + 0.008050420321524143, + 0.007341489661484957, + 0.008688163943588734, + 0.007305858191102743, + 0.03778769075870514, + 0.011809728108346462, + 0.07592710107564926, + 0.017353126779198647, + 0.012808994390070438, + 0.009220872074365616, + 0.010219866409897804, + 0.011081394739449024, + 0.020350199192762375, + 0.009720658883452415, + 0.010444076731801033, + 0.009309831075370312, + 0.007872526533901691, + 0.014319595880806446, + 0.005622842814773321, + 0.013747531920671463, + 0.013893391005694866, + 0.008361984975636005, + 0.010043863207101822, + 0.010518461465835571, + 0.00997498631477356, + 0.013095663860440254, + 0.01434920821338892, + 0.009344466961920261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.004108916502445936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193624913692474 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193624913692474, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193624913692474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18840999901294708, + 0.04184826835989952, + 0.05620653182268143, + 0.09190423786640167, + 0.2365865260362625, + 0.06345263123512268, + 0.11008168756961823, + 0.04850218445062637, + 0.0626162439584732, + 0.1003917008638382 + ], + "credit_map": [ + 0.011574270203709602, + 0.05209749937057495, + 0.009935591369867325, + 0.010765597224235535, + 0.06260108202695847, + 0.013933421112596989, + 0.01893489807844162, + 0.0085676284506917, + 0.0049557797610759735, + 0.006313658785074949, + 0.003837972180917859, + 0.003094572341069579, + 0.01092851348221302, + 0.0045052203349769115, + 0.004747013561427593, + 0.0034655341878533363, + 0.004857721272855997, + 0.010870899073779583, + 0.0037845310289412737, + 0.005147858988493681, + 0.01295984536409378, + 0.007669952232390642, + 0.005621341988444328, + 0.005294383503496647, + 0.00926556158810854, + 0.01362849585711956, + 0.011988183483481407, + 0.012836209498345852, + 0.018516846001148224, + 0.00872949417680502, + 0.01030353270471096, + 0.006635913159698248, + 0.010602708905935287, + 0.015789857134222984, + 0.022069856524467468, + 0.09071028977632523, + 0.06562595814466476, + 0.01001344807446003, + 0.012964503839612007, + 0.008809907361865044, + 0.008790658786892891, + 0.0098385876044631, + 0.007182369474321604, + 0.008721724152565002, + 0.00965932197868824, + 0.005835540127009153, + 0.007445184048265219, + 0.00597924692556262, + 0.010137766599655151, + 0.012235253117978573, + 0.008915577083826065, + 0.013890729285776615, + 0.016699837520718575, + 0.009775460697710514, + 0.025658486410975456, + 0.012768581509590149, + 0.007250315975397825, + 0.006493519991636276, + 0.0062472401186823845, + 0.005831614602357149, + 0.007341482676565647, + 0.005283868405967951, + 0.004937467165291309, + 0.005116675514727831, + 0.012978709302842617, + 0.007037503179162741, + 0.009123282507061958, + 0.007038467563688755, + 0.006433909293264151, + 0.007396071217954159, + 0.00487796263769269, + 0.007730334065854549, + 0.0210956409573555, + 0.011012520641088486, + 0.01140571665018797, + 0.007658017799258232, + 0.012876833789050579, + 0.012571940198540688, + 0.011613568291068077, + 0.012157469056546688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.018193624913692474 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017122859135270119 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017122859135270119, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017122859135270119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10617393255233765, + 0.04141929745674133, + 0.059000689536333084, + 0.06666303426027298, + 0.049482353031635284, + 0.08996039628982544, + 0.06603869795799255, + 0.12647929787635803, + 0.19929037988185883, + 0.19549189507961273 + ], + "credit_map": [ + 0.010120074264705181, + 0.017714466899633408, + 0.01291807834059, + 0.007878483273088932, + 0.029147574678063393, + 0.00790103618055582, + 0.013117114081978798, + 0.007377105299383402, + 0.004198424983769655, + 0.005750251933932304, + 0.003361131064593792, + 0.003935260698199272, + 0.010650194250047207, + 0.00480670714750886, + 0.005120847374200821, + 0.003596477210521698, + 0.0049863397143781185, + 0.013028040528297424, + 0.005166105926036835, + 0.006567253265529871, + 0.012331830337643623, + 0.005550707690417767, + 0.006908905692398548, + 0.004461507312953472, + 0.005308449734002352, + 0.01503151748329401, + 0.006177309900522232, + 0.008278626017272472, + 0.014590197242796421, + 0.006806943099945784, + 0.005493479780852795, + 0.0049765151925385, + 0.007867248728871346, + 0.006159260403364897, + 0.006294566672295332, + 0.008291336707770824, + 0.004391361027956009, + 0.006637497339397669, + 0.005121780559420586, + 0.004719301126897335, + 0.011958246119320393, + 0.011731373146176338, + 0.009008323773741722, + 0.011732648126780987, + 0.010507171973586082, + 0.009513789787888527, + 0.016257012262940407, + 0.009251829236745834, + 0.008167311549186707, + 0.006181677337735891, + 0.007528130896389484, + 0.009685710072517395, + 0.008391485549509525, + 0.008071674965322018, + 0.013270514085888863, + 0.004742187913507223, + 0.022025087848305702, + 0.007374791894108057, + 0.05242176726460457, + 0.01119088102132082, + 0.011453526094555855, + 0.0061217001639306545, + 0.008976297453045845, + 0.0069152419455349445, + 0.04383660852909088, + 0.022523043677210808, + 0.024922991171479225, + 0.017177047207951546, + 0.024558061733841896, + 0.03413977473974228, + 0.013128235004842281, + 0.019004613161087036, + 0.03490463271737099, + 0.02183203212916851, + 0.02116236835718155, + 0.019074581563472748, + 0.029049348086118698, + 0.023406125605106354, + 0.02441420778632164, + 0.021648602560162544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0017122859135270119 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start below the water's surface and tilt upward while zooming forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large lotus with translucent jade-pink petals rimmed with a soft golden halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the lotus petals unfold sequentially from the outermost layer to the inner layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a child with hair in two red-ribboned buns, wearing a red silk bellyband and gold bracelets, visible in the center of the lotus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the child holding a flaming spear with an eternal golden-red flame that casts dynamic light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731926463544369 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006731926463544369, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731926463544369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0756121575832367, + 0.042633116245269775, + 0.03287099301815033, + 0.06068391725420952, + 0.05921623855829239, + 0.07178454101085663, + 0.11601035296916962, + 0.19186057150363922, + 0.12843482196331024, + 0.2208932340145111 + ], + "credit_map": [ + 0.005342644173651934, + 0.02220837213099003, + 0.007444592192769051, + 0.004612858407199383, + 0.02080974169075489, + 0.004896785598248243, + 0.007072837557643652, + 0.0032243290916085243, + 0.004890482872724533, + 0.006605552043765783, + 0.0044656977988779545, + 0.0030149181839078665, + 0.01332786399871111, + 0.003379634814336896, + 0.0038452432490885258, + 0.003103723283857107, + 0.0035723820328712463, + 0.004201643168926239, + 0.0058129015378654, + 0.003546037245541811, + 0.005069033242762089, + 0.003520495956763625, + 0.003933703992515802, + 0.0032147979363799095, + 0.005871164612472057, + 0.011766026727855206, + 0.00963224470615387, + 0.006995271425694227, + 0.008960754610598087, + 0.0064557702280581, + 0.0056677707470953465, + 0.005334916990250349, + 0.004425175953656435, + 0.003472427139058709, + 0.008028198033571243, + 0.019537895917892456, + 0.012367594987154007, + 0.003910922445356846, + 0.004007579758763313, + 0.003466447815299034, + 0.006016048602759838, + 0.006444592960178852, + 0.008347488939762115, + 0.014682361856102943, + 0.007494522258639336, + 0.005445336923003197, + 0.016382403671741486, + 0.00697178952395916, + 0.010618271306157112, + 0.016833940520882607, + 0.010488404892385006, + 0.01639487035572529, + 0.016814857721328735, + 0.011119477450847626, + 0.01943914406001568, + 0.014301394112408161, + 0.027274517342448235, + 0.02312336303293705, + 0.034426622092723846, + 0.0178072489798069, + 0.03466688096523285, + 0.012706129811704159, + 0.026682889088988304, + 0.015172932296991348, + 0.01864045485854149, + 0.01702706515789032, + 0.01566980965435505, + 0.014008251950144768, + 0.018334757536649704, + 0.0124247120693326, + 0.018503444269299507, + 0.013826332986354828, + 0.04185720160603523, + 0.025907082483172417, + 0.02591363526880741, + 0.023703593760728836, + 0.025445980951189995, + 0.02289949730038643, + 0.029136989265680313, + 0.026029249653220177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.006731926463544369 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_8/index_tensor([849, 849])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_8/index_tensor([849, 849])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e712230f8af2023a28e1d82a50f9296216883a70 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_8/index_tensor([849, 849])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely close-up, static shot depicts the lower half of a giant panda's face, sharply focused on its powerful jaws. The camera remains perfectly still, held at eye level, capturing every minute detail. The panda's iconic fur is a dense tapestry of pure white around its muzzle and deep, light-absorbing black along its jawline, giving it a realistic texture. Its powerful jaws move in a steady, methodical rhythm, closing with deliberate force to crush a thick, vibrant green bamboo stalk held firmly in its claws just below the frame. With each bite, the bamboo crackles audibly, revealing a cross-section of its fibrous interior. The bright green, slender fibers, mixed with saliva, cling to the panda's black lips and the broken end of the stalk, shimmering in the soft, unfocused sunlight filtering from above. The background, a completely blurred palette of greens and browns, suggests a dense bamboo forest, ensuring the viewer's full attention remains focused on the primal, mechanical power of the animal's feeding. The atmosphere is one of concentrated, primal activity, a captivating display of nature's simple yet immense power.", + "seed": 29048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_8/index_849_seed_29048_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08228586195036769, + "vlm_logprob": -0.08228586195036769, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.08228586195036769, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03308646380901337 + ], + "opposite_logprob_tokens": [ + -4.40808629989624 + ], + "logprob_reward": -0.03308646380901337, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03308646380901337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12303031235933304, + 0.10115191340446472, + 0.10081163048744202, + 0.07900919765233994, + 0.1507192999124527, + 0.07586155086755753, + 0.11427545547485352, + 0.044218190014362335, + 0.07022175192832947, + 0.14070065319538116 + ], + "credit_map": [ + 0.017279721796512604, + 0.012114861980080605, + 0.02149173803627491, + 0.012564348056912422, + 0.01235582772642374, + 0.011403459124267101, + 0.01854228973388672, + 0.017278077080845833, + 0.0114522660151124, + 0.016952626407146454, + 0.010317034088075161, + 0.011064600199460983, + 0.008378256112337112, + 0.02037089318037033, + 0.009928574785590172, + 0.012687663547694683, + 0.013415036723017693, + 0.008433017879724503, + 0.013213484548032284, + 0.009303824976086617, + 0.011872301809489727, + 0.017116066068410873, + 0.013079526834189892, + 0.014378366060554981, + 0.009838970378041267, + 0.006314550060778856, + 0.013612869195640087, + 0.00889352336525917, + 0.005490897223353386, + 0.008789722807705402, + 0.015820030122995377, + 0.0102486377581954, + 0.01865999959409237, + 0.01716991513967514, + 0.022106502205133438, + 0.033534273505210876, + 0.009879508055746555, + 0.013481518253684044, + 0.020752914249897003, + 0.015134667046368122, + 0.012118375860154629, + 0.009581157006323338, + 0.01209007017314434, + 0.00820313673466444, + 0.0075175086967647076, + 0.009245375171303749, + 0.0066886181011796, + 0.010417308658361435, + 0.01507272943854332, + 0.007941138930618763, + 0.01436561532318592, + 0.017061486840248108, + 0.0075262486934661865, + 0.02131608873605728, + 0.014619479887187481, + 0.016372669488191605, + 0.006658048834651709, + 0.004717483650892973, + 0.004692941438406706, + 0.0027394562494009733, + 0.004839870613068342, + 0.007830118760466576, + 0.0044176010414958, + 0.008322671055793762, + 0.009254125878214836, + 0.009445944800972939, + 0.010647682473063469, + 0.007891587913036346, + 0.005017398856580257, + 0.0089222751557827, + 0.009056084789335728, + 0.009986651130020618, + 0.025541536509990692, + 0.009279558435082436, + 0.020715603604912758, + 0.021188849583268166, + 0.018166622146964073, + 0.014989478513598442, + 0.012899714522063732, + 0.017919296398758888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.03308646380901337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005995148327201605 + ], + "opposite_logprob_tokens": [ + -5.130995273590088 + ], + "logprob_reward": -0.005995148327201605, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005995148327201605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10823018103837967, + 0.11275765299797058, + 0.08098214864730835, + 0.09183008223772049, + 0.1519087255001068, + 0.06131840869784355, + 0.0650593638420105, + 0.07332269847393036, + 0.05502958968281746, + 0.19956116378307343 + ], + "credit_map": [ + 0.017688767984509468, + 0.01648304983973503, + 0.017582595348358154, + 0.008522103540599346, + 0.011653165332973003, + 0.013459998182952404, + 0.010229164734482765, + 0.01261135097593069, + 0.014375911094248295, + 0.012533191591501236, + 0.023460768163204193, + 0.008570639416575432, + 0.01057526282966137, + 0.016963934525847435, + 0.011675430461764336, + 0.014602524228394032, + 0.013150333426892757, + 0.0053455582819879055, + 0.01631559617817402, + 0.012719242833554745, + 0.006449911743402481, + 0.007527966983616352, + 0.008994591422379017, + 0.01047895010560751, + 0.013373777270317078, + 0.008449381217360497, + 0.011995458975434303, + 0.012479678727686405, + 0.006626996677368879, + 0.009782785549759865, + 0.015970071777701378, + 0.013151929713785648, + 0.01825007051229477, + 0.019667483866214752, + 0.016380874440073967, + 0.017177682369947433, + 0.013403329998254776, + 0.022774165496230125, + 0.02280445769429207, + 0.02145066298544407, + 0.008716853335499763, + 0.009143011644482613, + 0.008595562539994717, + 0.005545034538954496, + 0.006897481624037027, + 0.008807293139398098, + 0.005108226556330919, + 0.008504954166710377, + 0.008990487083792686, + 0.00498270383104682, + 0.0067566330544650555, + 0.006218154914677143, + 0.006176096852868795, + 0.013049532659351826, + 0.0069727106019854546, + 0.011913050897419453, + 0.009310027584433556, + 0.009007859043776989, + 0.006845318712294102, + 0.004712696652859449, + 0.010577874258160591, + 0.012438522651791573, + 0.00742160901427269, + 0.013008796609938145, + 0.006993907503783703, + 0.00478101521730423, + 0.005249840207397938, + 0.00691467197611928, + 0.0033266847021877766, + 0.0066957734525203705, + 0.008038213476538658, + 0.013029483146965504, + 0.03432684764266014, + 0.016337474808096886, + 0.024029558524489403, + 0.018704531714320183, + 0.026734737679362297, + 0.019965460523962975, + 0.031162336468696594, + 0.028300223872065544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.005995148327201605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085941456258297 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.011085941456258297, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085941456258297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10980749875307083, + 0.09496195614337921, + 0.09467659145593643, + 0.0880032554268837, + 0.07693803310394287, + 0.05480732396245003, + 0.19495606422424316, + 0.05692276358604431, + 0.1440991312265396, + 0.08482737094163895 + ], + "credit_map": [ + 0.01419750414788723, + 0.01600544899702072, + 0.012635466642677784, + 0.011928102001547813, + 0.011503626592457294, + 0.012948879040777683, + 0.018331658095121384, + 0.012256814166903496, + 0.011566621251404285, + 0.01638052426278591, + 0.010318100452423096, + 0.016719529405236244, + 0.006707025226205587, + 0.01055814977735281, + 0.012491652742028236, + 0.01022034976631403, + 0.011634635739028454, + 0.009358201175928116, + 0.010752481408417225, + 0.01293576043099165, + 0.013290759176015854, + 0.01566155068576336, + 0.010679658502340317, + 0.010363542474806309, + 0.010551037266850471, + 0.008586129173636436, + 0.010571716353297234, + 0.015346803702414036, + 0.008966063149273396, + 0.010979906655848026, + 0.012283205054700375, + 0.01071839313954115, + 0.00881054438650608, + 0.008506190963089466, + 0.005855597089976072, + 0.011558600701391697, + 0.008570075035095215, + 0.013210749253630638, + 0.008709290064871311, + 0.011716987006366253, + 0.007134521845728159, + 0.008501091040670872, + 0.006970420479774475, + 0.006941921543329954, + 0.005887986160814762, + 0.006599463988095522, + 0.006016611121594906, + 0.006755306385457516, + 0.0229592714458704, + 0.02089434489607811, + 0.021273333579301834, + 0.03704655170440674, + 0.018311895430088043, + 0.026735208928585052, + 0.02360575459897518, + 0.02412969060242176, + 0.007052943576127291, + 0.008826551958918571, + 0.00634424714371562, + 0.010620526038110256, + 0.005284278187900782, + 0.005199055187404156, + 0.006123615428805351, + 0.007471547927707434, + 0.020676610991358757, + 0.010527795180678368, + 0.029226133599877357, + 0.017937298864126205, + 0.009398923255503178, + 0.01700575277209282, + 0.015513707883656025, + 0.02381291799247265, + 0.011833898723125458, + 0.007310369051992893, + 0.011030498892068863, + 0.013601751998066902, + 0.007545593194663525, + 0.010761442594230175, + 0.007472105324268341, + 0.015271713957190514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.011085941456258297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546301819384098 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546301819384098, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546301819384098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09659053385257721, + 0.11660075187683105, + 0.1295992136001587, + 0.1065119281411171, + 0.0705932080745697, + 0.06705101579427719, + 0.0677841305732727, + 0.05323711782693863, + 0.1266520470380783, + 0.16538001596927643 + ], + "credit_map": [ + 0.010823434218764305, + 0.01231027115136385, + 0.011826898902654648, + 0.009237490594387054, + 0.010960846208035946, + 0.013240370899438858, + 0.01471585500985384, + 0.013475366868078709, + 0.015556402504444122, + 0.02057887800037861, + 0.014281938783824444, + 0.016943613067269325, + 0.008796806447207928, + 0.012913207523524761, + 0.015414169989526272, + 0.012115737423300743, + 0.01587957888841629, + 0.01024625264108181, + 0.01725347898900509, + 0.02205878496170044, + 0.01060998160392046, + 0.01335348654538393, + 0.0197460800409317, + 0.02045156992971897, + 0.015614318661391735, + 0.008695772849023342, + 0.016832605004310608, + 0.014938398264348507, + 0.006736340466886759, + 0.011615709401667118, + 0.014724683947861195, + 0.017354099079966545, + 0.008103652857244015, + 0.012985908426344395, + 0.0077786571346223354, + 0.010220656171441078, + 0.008660895749926567, + 0.0073879072442650795, + 0.006815818604081869, + 0.008639714680612087, + 0.006919369101524353, + 0.005739698186516762, + 0.010695798322558403, + 0.008429658599197865, + 0.007911842316389084, + 0.009823865257203579, + 0.007552238181233406, + 0.00997854582965374, + 0.008819554932415485, + 0.005881467368453741, + 0.009501716122031212, + 0.01096261478960514, + 0.004565914161503315, + 0.006787778809666634, + 0.010075216181576252, + 0.011189869605004787, + 0.006469841115176678, + 0.00481384014710784, + 0.007231585681438446, + 0.0038895250763744116, + 0.005071181338280439, + 0.0080104423686862, + 0.008359462022781372, + 0.009391237050294876, + 0.016152556985616684, + 0.008694414980709553, + 0.024997282773256302, + 0.008940760977566242, + 0.010541853494942188, + 0.019020268693566322, + 0.017959194257855415, + 0.020345700904726982, + 0.023994827643036842, + 0.009637810289859772, + 0.044630423188209534, + 0.02348795160651207, + 0.009667727164924145, + 0.017885534092783928, + 0.015565069392323494, + 0.020510679110884666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.012546301819384098 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487154543399811 + ], + "opposite_logprob_tokens": [ + -1.2237154245376587 + ], + "logprob_reward": -0.3487154543399811, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3487154543399811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10364886373281479, + 0.05138799920678139, + 0.10385872423648834, + 0.08752573281526566, + 0.2218930870294571, + 0.18355251848697662, + 0.058331169188022614, + 0.0468614362180233, + 0.030023759230971336, + 0.11291669309139252 + ], + "credit_map": [ + 0.013055197894573212, + 0.014001689851284027, + 0.016205552965402603, + 0.009645441547036171, + 0.012240253388881683, + 0.013097316958010197, + 0.012687945738434792, + 0.012715471908450127, + 0.007647624704986811, + 0.004407814703881741, + 0.009916338138282299, + 0.004226888995617628, + 0.003777413396164775, + 0.007931623607873917, + 0.005944213829934597, + 0.007536089047789574, + 0.01323621068149805, + 0.009083271026611328, + 0.013636929914355278, + 0.010575708001852036, + 0.012581930495798588, + 0.017057795077562332, + 0.014566374011337757, + 0.013120513409376144, + 0.011553787626326084, + 0.008218356408178806, + 0.012013820931315422, + 0.009527438320219517, + 0.009041371755301952, + 0.010563424788415432, + 0.015754062682390213, + 0.010853474028408527, + 0.02655651792883873, + 0.03882179781794548, + 0.02778877690434456, + 0.042116425931453705, + 0.015394974499940872, + 0.02471558190882206, + 0.02484462782740593, + 0.0216544046998024, + 0.020651206374168396, + 0.021283259615302086, + 0.025173505768179893, + 0.01620997115969658, + 0.02262711152434349, + 0.031519610434770584, + 0.01790555939078331, + 0.02818230353295803, + 0.008679160848259926, + 0.006109639070928097, + 0.00598800228908658, + 0.0059282537549734116, + 0.005960914306342602, + 0.011149522848427296, + 0.00657097902148962, + 0.007944700308144093, + 0.00563610065728426, + 0.0071242982521653175, + 0.0031736164819449186, + 0.0029750384856015444, + 0.007121728267520666, + 0.008526860736310482, + 0.0052044205367565155, + 0.007099375128746033, + 0.004013092257082462, + 0.0025079725310206413, + 0.005696817301213741, + 0.0036087343469262123, + 0.0024315316695719957, + 0.0033368775621056557, + 0.003952597267925739, + 0.004476138856261969, + 0.016681352630257607, + 0.008190877735614777, + 0.022260593250393867, + 0.01581941358745098, + 0.009978645481169224, + 0.013630765490233898, + 0.011790845543146133, + 0.01456420123577118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.3487154543399811 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close-up, static shot depicts the lower half of a giant panda's face, sharply focused on its powerful jaws. The camera remains perfectly still, held at eye level, capturing every minute detail. The panda's iconic fur is a dense tapestry of pure white around its muzzle and deep, light-absorbing black along its jawline, giving it a realistic texture. Its powerful jaws move in a steady, methodical rhythm, closing with deliberate force to crush a thick, vibrant green bamboo stalk held firmly in its claws just below the frame. With each bite, the bamboo crackles audibly, revealing a cross-section of its fibrous interior. The bright green, slender fibers, mixed with saliva, cling to the panda's black lips and the broken end of the stalk, shimmering in the soft, unfocused sunlight filtering from above. The background, a completely blurred palette of greens and browns, suggests a dense bamboo forest, ensuring the viewer's full attention remains focused on the primal, mechanical power of the animal's feeding. The atmosphere is one of concentrated, primal activity, a captivating display of nature's simple yet immense power.", + "seed": 29049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_8/index_849_seed_29049_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1802665673196316, + "vlm_logprob": -0.1802665673196316, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.1802665673196316, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04442651942372322 + ], + "opposite_logprob_tokens": [ + -3.669426441192627 + ], + "logprob_reward": -0.04442651942372322, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04442651942372322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22950556874275208, + 0.05047418549656868, + 0.11243069916963577, + 0.09396643191576004, + 0.1443861424922943, + 0.06695374101400375, + 0.08984911441802979, + 0.15999309718608856, + 0.026578979566693306, + 0.025862105190753937 + ], + "credit_map": [ + 0.015484753996133804, + 0.025119254365563393, + 0.017587047070264816, + 0.024167202413082123, + 0.025749273598194122, + 0.06122129037976265, + 0.03179007023572922, + 0.028386680409312248, + 0.0058927275240421295, + 0.004122554324567318, + 0.00452003488317132, + 0.00910993106663227, + 0.008363163098692894, + 0.007799127139151096, + 0.0059022787027060986, + 0.004764368757605553, + 0.010054175741970539, + 0.023104820400476456, + 0.011288751848042011, + 0.00760797131806612, + 0.0159115232527256, + 0.018079448491334915, + 0.012170271016657352, + 0.014213735237717628, + 0.008355684578418732, + 0.008416563272476196, + 0.011118454858660698, + 0.009048865176737309, + 0.015275850892066956, + 0.019492626190185547, + 0.010572598315775394, + 0.011685790494084358, + 0.01518795546144247, + 0.011241975240409374, + 0.01766885258257389, + 0.019297989085316658, + 0.019373921677470207, + 0.026628928259015083, + 0.014753265306353569, + 0.02023325487971306, + 0.006202986463904381, + 0.00969783030450344, + 0.010054357349872589, + 0.008349855430424213, + 0.010938183404505253, + 0.007705141790211201, + 0.006611215882003307, + 0.007394168991595507, + 0.008732096292078495, + 0.017097022384405136, + 0.01804720051586628, + 0.007275009993463755, + 0.010486618615686893, + 0.009375854395329952, + 0.006354534067213535, + 0.012480768375098705, + 0.011238149367272854, + 0.012793835252523422, + 0.025665612891316414, + 0.02240801230072975, + 0.01927337795495987, + 0.021696779876947403, + 0.019757909700274467, + 0.027159418910741806, + 0.0024856014642864466, + 0.0022660959511995316, + 0.004428021609783173, + 0.002801242982968688, + 0.0034371039364486933, + 0.004353885073214769, + 0.0029625440947711468, + 0.0038444832898676395, + 0.001947769895195961, + 0.002676191972568631, + 0.003388812765479088, + 0.002951556583866477, + 0.0035926068667322397, + 0.0036367452703416348, + 0.004463270306587219, + 0.003205152926966548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.04442651942372322 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009826619178056717 + ], + "opposite_logprob_tokens": [ + -4.63482666015625 + ], + "logprob_reward": -0.009826619178056717, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009826619178056717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3427896797657013, + 0.10004211962223053, + 0.11981446295976639, + 0.11036340892314911, + 0.07227334380149841, + 0.02735072933137417, + 0.033144187182188034, + 0.04781379923224449, + 0.09547220915555954, + 0.05093611404299736 + ], + "credit_map": [ + 0.022566260769963264, + 0.046164851635694504, + 0.015795288607478142, + 0.02329426445066929, + 0.04118609055876732, + 0.10849267244338989, + 0.04088209196925163, + 0.04440813511610031, + 0.006794849876314402, + 0.010483477264642715, + 0.007113161962479353, + 0.009917129762470722, + 0.021192362532019615, + 0.013043896295130253, + 0.01723429374396801, + 0.014262942597270012, + 0.01333719864487648, + 0.02030896581709385, + 0.017574582248926163, + 0.009587948210537434, + 0.010982624255120754, + 0.018413960933685303, + 0.010700508952140808, + 0.018908662721514702, + 0.016396615654230118, + 0.010159404017031193, + 0.010546774603426456, + 0.007063169032335281, + 0.02293633110821247, + 0.01793120615184307, + 0.01066863164305687, + 0.01466126088052988, + 0.006244678515940905, + 0.006766543723642826, + 0.010378453880548477, + 0.011305508203804493, + 0.007112778723239899, + 0.012515767477452755, + 0.00881880335509777, + 0.009130805730819702, + 0.0023439598735421896, + 0.002817912260070443, + 0.003699067747220397, + 0.0042526014149188995, + 0.003518319223076105, + 0.003964985720813274, + 0.003791897092014551, + 0.0029619846027344465, + 0.0027900729328393936, + 0.004442553035914898, + 0.006629244890064001, + 0.0035851015709340572, + 0.0036037242971360683, + 0.0039533087983727455, + 0.0031089349649846554, + 0.005031247157603502, + 0.004512018524110317, + 0.0049832724034786224, + 0.007498904597014189, + 0.006363525055348873, + 0.006859865039587021, + 0.006981987506151199, + 0.004531884100288153, + 0.006082335952669382, + 0.007030556909739971, + 0.012566613964736462, + 0.011837761849164963, + 0.008123906329274178, + 0.016136089339852333, + 0.016797535121440887, + 0.007631272077560425, + 0.015348470769822598, + 0.004299028776586056, + 0.0034228444565087557, + 0.006113589275628328, + 0.004742163699120283, + 0.004608313553035259, + 0.011404987424612045, + 0.00784085039049387, + 0.008504332043230534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.009826619178056717 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979179609566927 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.00979179609566927, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979179609566927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2262265384197235, + 0.10162334144115448, + 0.06408338248729706, + 0.09244727343320847, + 0.21471063792705536, + 0.03375277295708656, + 0.039162777364254, + 0.1269644945859909, + 0.03448975458741188, + 0.0665389895439148 + ], + "credit_map": [ + 0.01911030523478985, + 0.02269970066845417, + 0.017663389444351196, + 0.018526092171669006, + 0.023697221651673317, + 0.0608573779463768, + 0.026143252849578857, + 0.037529200315475464, + 0.009643247351050377, + 0.008846231736242771, + 0.0088500389829278, + 0.015629608184099197, + 0.015610981732606888, + 0.011460421606898308, + 0.01826712116599083, + 0.01331568043678999, + 0.005624860059469938, + 0.012281115166842937, + 0.00855510588735342, + 0.00553445890545845, + 0.007099183276295662, + 0.008826603181660175, + 0.006382326129823923, + 0.009779730811715126, + 0.011619091033935547, + 0.014591970480978489, + 0.007024385966360569, + 0.0066094123758375645, + 0.01326039619743824, + 0.015817439183592796, + 0.010719583369791508, + 0.012804989702999592, + 0.024249030277132988, + 0.01995832473039627, + 0.027133163064718246, + 0.027369195595383644, + 0.022022217512130737, + 0.039835646748542786, + 0.029122326523065567, + 0.025020746514201164, + 0.0030474141240119934, + 0.004004510585218668, + 0.005021261516958475, + 0.005112209357321262, + 0.00338310981169343, + 0.005083594471216202, + 0.004048183094710112, + 0.004052488598972559, + 0.0036458177492022514, + 0.0066019343212246895, + 0.0073491972871124744, + 0.004167068284004927, + 0.005919917486608028, + 0.003941831644624472, + 0.0022525014355778694, + 0.005284509621560574, + 0.011031582951545715, + 0.012050256133079529, + 0.01891167089343071, + 0.011948560364544392, + 0.02233797498047352, + 0.018968120217323303, + 0.012966644950211048, + 0.01874968782067299, + 0.0026368708349764347, + 0.0035286734346300364, + 0.006102324463427067, + 0.0045333136804401875, + 0.003980135545134544, + 0.005231463350355625, + 0.00350857712328434, + 0.004968396853655577, + 0.005634556990116835, + 0.00488670077174902, + 0.010436245240271091, + 0.006332899909466505, + 0.007395096123218536, + 0.010038908571004868, + 0.011536418460309505, + 0.010278156958520412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.00979179609566927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081933043897152 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081933043897152, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081933043897152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1446581929922104, + 0.04536733403801918, + 0.0879666730761528, + 0.10798964649438858, + 0.08366777002811432, + 0.06681355088949203, + 0.04877757653594017, + 0.13970662653446198, + 0.16022728383541107, + 0.11482537537813187 + ], + "credit_map": [ + 0.013597927987575531, + 0.018707165494561195, + 0.012180649675428867, + 0.011111492291092873, + 0.02215312421321869, + 0.021976158022880554, + 0.028474412858486176, + 0.01645726151764393, + 0.004656338132917881, + 0.002949072513729334, + 0.003306644270196557, + 0.008632668294012547, + 0.004826027899980545, + 0.006546616554260254, + 0.007152274250984192, + 0.007297692820429802, + 0.008651274256408215, + 0.015488438308238983, + 0.015325489453971386, + 0.007024809718132019, + 0.009015923365950584, + 0.010706408880650997, + 0.008097423240542412, + 0.013656909577548504, + 0.012505059130489826, + 0.011333541013300419, + 0.009888803586363792, + 0.008904747664928436, + 0.022345980629324913, + 0.019969088956713676, + 0.011224769055843353, + 0.01181765180081129, + 0.012900169938802719, + 0.005596661940217018, + 0.012593644671142101, + 0.009651546366512775, + 0.007763106841593981, + 0.01294920314103365, + 0.010444512590765953, + 0.011768922209739685, + 0.004992018453776836, + 0.007746844552457333, + 0.009280303493142128, + 0.009361226111650467, + 0.012769728899002075, + 0.007232089061290026, + 0.007869771681725979, + 0.0075615644454956055, + 0.004225538577884436, + 0.005111917853355408, + 0.009321079589426517, + 0.005888947751373053, + 0.005392399616539478, + 0.006565961521118879, + 0.004952214192599058, + 0.007319519761949778, + 0.012582742609083652, + 0.014917983673512936, + 0.018013792112469673, + 0.01559472642838955, + 0.025701159611344337, + 0.017539849504828453, + 0.014147555455565453, + 0.02120882086455822, + 0.014074545353651047, + 0.016077864915132523, + 0.029333848506212234, + 0.016048818826675415, + 0.022215228527784348, + 0.024797679856419563, + 0.01434362307190895, + 0.023335671052336693, + 0.008570129051804543, + 0.007900254800915718, + 0.015234897844493389, + 0.013182305730879307, + 0.010592695325613022, + 0.019855402410030365, + 0.022771570831537247, + 0.016718123108148575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.011081933043897152 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262059688568115 + ], + "opposite_logprob_tokens": [ + -0.5762059688568115 + ], + "logprob_reward": -0.8262059688568115, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262059688568115, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4020931124687195, + 0.04295649006962776, + 0.09663839638233185, + 0.12798239290714264, + 0.0934280976653099, + 0.03827930986881256, + 0.033962611109018326, + 0.09011989086866379, + 0.057621538639068604, + 0.016918111592531204 + ], + "credit_map": [ + 0.0290085319429636, + 0.05815194174647331, + 0.01913437619805336, + 0.024454263970255852, + 0.05050649121403694, + 0.12081115692853928, + 0.04973932355642319, + 0.05028700456023216, + 0.004528924822807312, + 0.0029602537397295237, + 0.003655629465356469, + 0.008049625903367996, + 0.004721466917544603, + 0.004838576540350914, + 0.006391739938408136, + 0.0078102718107402325, + 0.00875579472631216, + 0.017125757411122322, + 0.01389092206954956, + 0.007953636348247528, + 0.014278835617005825, + 0.011487852782011032, + 0.00959804467856884, + 0.013547551818192005, + 0.02250068262219429, + 0.016597678884863853, + 0.01096264272928238, + 0.00803388375788927, + 0.02138334885239601, + 0.0192074254155159, + 0.012278570793569088, + 0.017018156126141548, + 0.009031777270138264, + 0.00885665975511074, + 0.010568402707576752, + 0.01214703917503357, + 0.011399917304515839, + 0.01720576547086239, + 0.013217187486588955, + 0.0110013447701931, + 0.004329102113842964, + 0.005955379921942949, + 0.004807981662452221, + 0.0051933941431343555, + 0.006858422886580229, + 0.004246823955327272, + 0.0030391609761863947, + 0.0038490428123623133, + 0.0021989475935697556, + 0.0034888440277427435, + 0.010026929900050163, + 0.003305603750050068, + 0.0030075376853346825, + 0.003872462548315525, + 0.002150313463062048, + 0.005911971442401409, + 0.008161170408129692, + 0.009287378750741482, + 0.010303704999387264, + 0.013754896819591522, + 0.014540927484631538, + 0.013253728859126568, + 0.010107350535690784, + 0.01071073580533266, + 0.004447809886187315, + 0.007464354392141104, + 0.007901041768491268, + 0.004912001546472311, + 0.009996669366955757, + 0.008693819865584373, + 0.0052401041612029076, + 0.008965742774307728, + 0.0013694369699805975, + 0.0014281354378908873, + 0.003642862429842353, + 0.001492321491241455, + 0.00151727139018476, + 0.0025754428934305906, + 0.0023417766205966473, + 0.002550864592194557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.8262059688568115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extremely close-up, static shot depicts the lower half of a giant panda's face, sharply focused on its powerful jaws. The camera remains perfectly still, held at eye level, capturing every minute detail. The panda's iconic fur is a dense tapestry of pure white around its muzzle and deep, light-absorbing black along its jawline, giving it a realistic texture. Its powerful jaws move in a steady, methodical rhythm, closing with deliberate force to crush a thick, vibrant green bamboo stalk held firmly in its claws just below the frame. With each bite, the bamboo crackles audibly, revealing a cross-section of its fibrous interior. The bright green, slender fibers, mixed with saliva, cling to the panda's black lips and the broken end of the stalk, shimmering in the soft, unfocused sunlight filtering from above. The background, a completely blurred palette of greens and browns, suggests a dense bamboo forest, ensuring the viewer's full attention remains focused on the primal, mechanical power of the animal's feeding. The atmosphere is one of concentrated, primal activity, a captivating display of nature's simple yet immense power.", + "seed": 29050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_8/index_849_seed_29050_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1472606903873384, + "vlm_logprob": -0.1472606903873384, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.1472606903873384, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01198978815227747 + ], + "opposite_logprob_tokens": [ + -6.136989593505859 + ], + "logprob_reward": -0.01198978815227747, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.01198978815227747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2692818343639374, + 0.13085447251796722, + 0.06170346215367317, + 0.04301152750849724, + 0.044828321784734726, + 0.11261696368455887, + 0.12089155614376068, + 0.09705474972724915, + 0.05754023790359497, + 0.062216829508543015 + ], + "credit_map": [ + 0.041968222707509995, + 0.024639366194605827, + 0.04782745987176895, + 0.035227954387664795, + 0.027001570910215378, + 0.025842919945716858, + 0.033638302236795425, + 0.03313606604933739, + 0.04082475230097771, + 0.010852774605154991, + 0.008203988894820213, + 0.008891547098755836, + 0.0348348468542099, + 0.010466440580785275, + 0.00960709061473608, + 0.007173036690801382, + 0.00685037299990654, + 0.0059220255352556705, + 0.007631498388946056, + 0.011814696714282036, + 0.00518723065033555, + 0.006741279736161232, + 0.007772969547659159, + 0.009783397428691387, + 0.004355632234364748, + 0.004514438100159168, + 0.0040126065723598, + 0.009245125576853752, + 0.005547068081796169, + 0.00484956381842494, + 0.0063230968080461025, + 0.0041640023700892925, + 0.008331949822604656, + 0.005319427698850632, + 0.006422677077353001, + 0.004112159833312035, + 0.00888676568865776, + 0.0043403105810284615, + 0.003623375901952386, + 0.003791661001741886, + 0.01242678053677082, + 0.00824492983520031, + 0.00819909293204546, + 0.008599299937486649, + 0.025606192648410797, + 0.011747786775231361, + 0.009057353250682354, + 0.028735538944602013, + 0.012064577080309391, + 0.013936537317931652, + 0.008106392808258533, + 0.017488950863480568, + 0.014365625567734241, + 0.03136398643255234, + 0.014338122680783272, + 0.00922737829387188, + 0.013208076357841492, + 0.010267105884850025, + 0.013758430257439613, + 0.014108339324593544, + 0.013779138214886189, + 0.007365576922893524, + 0.015490653924643993, + 0.009077445603907108, + 0.009209821932017803, + 0.008505800738930702, + 0.004642945248633623, + 0.006038238760083914, + 0.007223485037684441, + 0.006835366599261761, + 0.007578319404274225, + 0.007506270427256823, + 0.0075479247607290745, + 0.006781607400625944, + 0.008190780878067017, + 0.012093277648091316, + 0.004407649394124746, + 0.007277291733771563, + 0.008533087559044361, + 0.007385217119008303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01198978815227747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00596955232322216 + ], + "opposite_logprob_tokens": [ + -5.130969524383545 + ], + "logprob_reward": -0.00596955232322216, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00596955232322216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12041416019201279, + 0.14044851064682007, + 0.16688279807567596, + 0.08857600390911102, + 0.10004027932882309, + 0.0447467640042305, + 0.028254197910428047, + 0.11564420908689499, + 0.13073191046714783, + 0.06426113843917847 + ], + "credit_map": [ + 0.018847744911909103, + 0.01570066809654236, + 0.00947274174541235, + 0.01020471565425396, + 0.016578208655118942, + 0.019051462411880493, + 0.012202258221805096, + 0.018356366083025932, + 0.03819947689771652, + 0.011264693923294544, + 0.010638969950377941, + 0.01228269562125206, + 0.031564705073833466, + 0.013063330203294754, + 0.00944616086781025, + 0.01398846972733736, + 0.020204050466418266, + 0.016617553308606148, + 0.015987243503332138, + 0.03903967887163162, + 0.01887822523713112, + 0.014655969105660915, + 0.017835058271884918, + 0.02366502769291401, + 0.009149688296020031, + 0.012194476090371609, + 0.0058401962742209435, + 0.01810465380549431, + 0.009463710710406303, + 0.011125177145004272, + 0.010652115568518639, + 0.012045986950397491, + 0.0173126682639122, + 0.009876363910734653, + 0.013548345305025578, + 0.010376308113336563, + 0.017913825809955597, + 0.01095314510166645, + 0.00846292544156313, + 0.011596699245274067, + 0.004968070890754461, + 0.0048671565018594265, + 0.005328826606273651, + 0.005803769454360008, + 0.004872587509453297, + 0.0061516305431723595, + 0.0070519354194402695, + 0.005702787544578314, + 0.0032961599063128233, + 0.0032333252020180225, + 0.004727776162326336, + 0.0027947535272687674, + 0.0038280184380710125, + 0.0035139701794832945, + 0.0037936896551400423, + 0.003066505305469036, + 0.011403585784137249, + 0.009515706449747086, + 0.025351429358124733, + 0.011606487445533276, + 0.028359098359942436, + 0.009024555794894695, + 0.012982439249753952, + 0.007400910835713148, + 0.02806774526834488, + 0.010797141119837761, + 0.008330060169100761, + 0.010354684665799141, + 0.011528137139976025, + 0.03495553508400917, + 0.014426340349018574, + 0.012272270396351814, + 0.007871473208069801, + 0.006175694055855274, + 0.007309622596949339, + 0.007353559136390686, + 0.00434977887198329, + 0.014532721601426601, + 0.007303471677005291, + 0.009364813566207886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.00596955232322216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191752955317497 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191752955317497, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191752955317497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3039906620979309, + 0.148360475897789, + 0.11953600496053696, + 0.034200072288513184, + 0.05318392068147659, + 0.05039459466934204, + 0.037669338285923004, + 0.08809869736433029, + 0.0857756957411766, + 0.078790582716465 + ], + "credit_map": [ + 0.04107154533267021, + 0.027415018528699875, + 0.024043474346399307, + 0.02659408003091812, + 0.03157844394445419, + 0.06789363920688629, + 0.035213712602853775, + 0.05018073320388794, + 0.04590005427598953, + 0.010057881474494934, + 0.010778357274830341, + 0.011895304545760155, + 0.03644593060016632, + 0.01246584765613079, + 0.00902694184333086, + 0.01179015077650547, + 0.013572249561548233, + 0.008878658525645733, + 0.014188367873430252, + 0.021245216950774193, + 0.009835061617195606, + 0.011983262374997139, + 0.016789468005299568, + 0.023043718189001083, + 0.003422129899263382, + 0.0050304546020925045, + 0.0025744945742189884, + 0.007440946064889431, + 0.0043632048182189465, + 0.0038047991693019867, + 0.004037119913846254, + 0.003526922781020403, + 0.007106599397957325, + 0.0054870410822331905, + 0.006938467733561993, + 0.007075438275933266, + 0.011522328481078148, + 0.00575857562944293, + 0.004473426379263401, + 0.004822040442377329, + 0.006609808653593063, + 0.0050050425343215466, + 0.003761165775358677, + 0.005093073006719351, + 0.010779106989502907, + 0.004480780567973852, + 0.009463422000408173, + 0.005202192347496748, + 0.004215306136757135, + 0.0028528422117233276, + 0.0030527228955179453, + 0.00368926371447742, + 0.005059621762484312, + 0.008085415698587894, + 0.004217476584017277, + 0.006496689282357693, + 0.013877540826797485, + 0.0074443393386900425, + 0.012049159966409206, + 0.017539624124765396, + 0.008133665658533573, + 0.007486449554562569, + 0.014082114212214947, + 0.00748580414801836, + 0.014109475538134575, + 0.009407092817127705, + 0.006439910735934973, + 0.008787054568529129, + 0.008634570986032486, + 0.016125859692692757, + 0.01278756558895111, + 0.00948416255414486, + 0.013306970708072186, + 0.004966552834957838, + 0.007708215620368719, + 0.006281540263444185, + 0.00477816304191947, + 0.02488948591053486, + 0.007759396918118, + 0.009100254625082016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018191752955317497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067374915815889835 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.0067374915815889835, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067374915815889835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27839815616607666, + 0.31139203906059265, + 0.04865089803934097, + 0.042932622134685516, + 0.02892472967505455, + 0.04610259085893631, + 0.03876625746488571, + 0.06197519227862358, + 0.04067685455083847, + 0.10218067467212677 + ], + "credit_map": [ + 0.03953464701771736, + 0.023695090785622597, + 0.04096740484237671, + 0.037117477506399155, + 0.023547129705548286, + 0.02347886562347412, + 0.040357206016778946, + 0.04970034211874008, + 0.10028740018606186, + 0.022115539759397507, + 0.0189386997371912, + 0.024233423173427582, + 0.07852861285209656, + 0.024866392835974693, + 0.021432409062981606, + 0.020989568904042244, + 0.00613639410585165, + 0.006217059679329395, + 0.004348441027104855, + 0.009350932203233242, + 0.005318270996212959, + 0.0045357272028923035, + 0.006762847304344177, + 0.005981226917356253, + 0.004199660383164883, + 0.005032552871853113, + 0.0030951935332268476, + 0.009171687066555023, + 0.00485563650727272, + 0.005323995836079121, + 0.005480441730469465, + 0.0057734539732337, + 0.004486513789743185, + 0.001985348993912339, + 0.004416702315211296, + 0.0037527652457356453, + 0.003519483609125018, + 0.005015810485929251, + 0.0032834727317094803, + 0.002464631339535117, + 0.00523618096485734, + 0.004883520770817995, + 0.004445779137313366, + 0.004426616244018078, + 0.009717049077153206, + 0.003937056753784418, + 0.0034560866188257933, + 0.010000301524996758, + 0.005349515471607447, + 0.003482439322397113, + 0.004313562996685505, + 0.005193311255425215, + 0.0047239153645932674, + 0.005045391619205475, + 0.005979279987514019, + 0.004678840283304453, + 0.008706576190888882, + 0.004430265165865421, + 0.011617899872362614, + 0.008278277702629566, + 0.013110784813761711, + 0.0038230803329497576, + 0.006736834067851305, + 0.005271478556096554, + 0.006379116792231798, + 0.004010397009551525, + 0.003633947344496846, + 0.004990903660655022, + 0.0035211401991546154, + 0.007673854939639568, + 0.006186644546687603, + 0.004280851222574711, + 0.01088301744312048, + 0.005196652375161648, + 0.007832483388483524, + 0.01198312547057867, + 0.005786710884422064, + 0.03374924510717392, + 0.008359444327652454, + 0.018390001729130745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0067374915815889835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934148669242859 + ], + "opposite_logprob_tokens": [ + -0.6934148669242859 + ], + "logprob_reward": -0.6934148669242859, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934148669242859, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16675083339214325, + 0.2810073494911194, + 0.1259320229291916, + 0.026906728744506836, + 0.04102720692753792, + 0.1096961572766304, + 0.04852381348609924, + 0.0482109971344471, + 0.03621295839548111, + 0.11573190242052078 + ], + "credit_map": [ + 0.019100673496723175, + 0.01209211815148592, + 0.023336781188845634, + 0.026708411052823067, + 0.013576599769294262, + 0.022445637732744217, + 0.027920665219426155, + 0.02156994864344597, + 0.08609955757856369, + 0.019679050892591476, + 0.01964251883327961, + 0.0225722324103117, + 0.06727811694145203, + 0.022861557081341743, + 0.018754271790385246, + 0.02412007376551628, + 0.013406672514975071, + 0.023069581016898155, + 0.012372592464089394, + 0.022532207891345024, + 0.014335103332996368, + 0.015606002882122993, + 0.012417115271091461, + 0.012192760594189167, + 0.0028453273698687553, + 0.0031527909450232983, + 0.00260147824883461, + 0.003380919573828578, + 0.004329226911067963, + 0.003439940745010972, + 0.0032728551886975765, + 0.0038841895293444395, + 0.004980219062417746, + 0.00434504309669137, + 0.005798387806862593, + 0.005860469304025173, + 0.007032988127321005, + 0.005047381389886141, + 0.0047216396778821945, + 0.0032410789281129837, + 0.013016588985919952, + 0.010823013260960579, + 0.0060430667363107204, + 0.0057121641002595425, + 0.029229510575532913, + 0.009835065342485905, + 0.012173761613667011, + 0.022862983867526054, + 0.005847839172929525, + 0.005092347506433725, + 0.004484895151108503, + 0.006052049808204174, + 0.005333967972546816, + 0.013214272446930408, + 0.0047049918211996555, + 0.003793448908254504, + 0.0051225279457867146, + 0.005272905342280865, + 0.010562664829194546, + 0.0071334876120090485, + 0.007732355035841465, + 0.003013168927282095, + 0.006081326398998499, + 0.003292562672868371, + 0.005847957916557789, + 0.005669800564646721, + 0.00257932860404253, + 0.00444559333845973, + 0.004519794601947069, + 0.003678955603390932, + 0.004532569088041782, + 0.004938956815749407, + 0.02057991735637188, + 0.008909843862056732, + 0.010085845366120338, + 0.008544511161744595, + 0.006063310895115137, + 0.04039493203163147, + 0.008927013725042343, + 0.012226521037518978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.6934148669242859 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extremely close-up, static shot depicts the lower half of a giant panda's face, sharply focused on its powerful jaws. The camera remains perfectly still, held at eye level, capturing every minute detail. The panda's iconic fur is a dense tapestry of pure white around its muzzle and deep, light-absorbing black along its jawline, giving it a realistic texture. Its powerful jaws move in a steady, methodical rhythm, closing with deliberate force to crush a thick, vibrant green bamboo stalk held firmly in its claws just below the frame. With each bite, the bamboo crackles audibly, revealing a cross-section of its fibrous interior. The bright green, slender fibers, mixed with saliva, cling to the panda's black lips and the broken end of the stalk, shimmering in the soft, unfocused sunlight filtering from above. The background, a completely blurred palette of greens and browns, suggests a dense bamboo forest, ensuring the viewer's full attention remains focused on the primal, mechanical power of the animal's feeding. The atmosphere is one of concentrated, primal activity, a captivating display of nature's simple yet immense power.", + "seed": 29051, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_8/index_849_seed_29051_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17865899382159114, + "vlm_logprob": -0.17865899382159114, + "vlm_yes_no_margin": 3.5249999523162843, + "vlm_token_credit": -0.17865899382159114, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03424381837248802 + ], + "opposite_logprob_tokens": [ + -4.034243583679199 + ], + "logprob_reward": -0.03424381837248802, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03424381837248802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35595232248306274, + 0.046631380915641785, + 0.1856013685464859, + 0.07745236158370972, + 0.0333147794008255, + 0.029756680130958557, + 0.12916940450668335, + 0.06251168996095657, + 0.024778803810477257, + 0.0548311248421669 + ], + "credit_map": [ + 0.09960425645112991, + 0.04633697122335434, + 0.11588120460510254, + 0.02411959134042263, + 0.016111230477690697, + 0.014460764825344086, + 0.020544443279504776, + 0.01889386773109436, + 0.012961715459823608, + 0.0042415340431034565, + 0.012149336747825146, + 0.0043734158389270306, + 0.0028296876698732376, + 0.0034303367137908936, + 0.002880726009607315, + 0.003764629364013672, + 0.021537447348237038, + 0.03215862065553665, + 0.03577513247728348, + 0.018076108768582344, + 0.018994474783539772, + 0.014508629217743874, + 0.025547103956341743, + 0.019003858789801598, + 0.011991425417363644, + 0.010813560336828232, + 0.016434723511338234, + 0.011151661165058613, + 0.007099950686097145, + 0.005317275878041983, + 0.006306621246039867, + 0.008337144739925861, + 0.00442994711920619, + 0.004883557558059692, + 0.007192458026111126, + 0.0035647053737193346, + 0.00276984297670424, + 0.003868690226227045, + 0.00403670035302639, + 0.002568879397585988, + 0.00514250760897994, + 0.004484796430915594, + 0.0037377341650426388, + 0.004077553749084473, + 0.004079005680978298, + 0.0022645029239356518, + 0.002273849444463849, + 0.003696730127558112, + 0.015333249233663082, + 0.01710793375968933, + 0.017984826117753983, + 0.012569093145430088, + 0.01517532393336296, + 0.017179366201162338, + 0.018280187621712685, + 0.01553941797465086, + 0.006576171610504389, + 0.010727612301707268, + 0.009191721677780151, + 0.007566518150269985, + 0.00900224968791008, + 0.005330609157681465, + 0.009580212645232677, + 0.004536593332886696, + 0.002342818770557642, + 0.0040559787303209305, + 0.004653383512049913, + 0.002947198925539851, + 0.002787103410810232, + 0.0024977531284093857, + 0.0020874529145658016, + 0.003407114651054144, + 0.009322180412709713, + 0.008688550442457199, + 0.010747095569968224, + 0.006138003896921873, + 0.004507554695010185, + 0.00480201980099082, + 0.0050556533969938755, + 0.005570065230131149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.03424381837248802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005982587579637766 + ], + "opposite_logprob_tokens": [ + -5.130982398986816 + ], + "logprob_reward": -0.005982587579637766, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005982587579637766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11637810617685318, + 0.06365197151899338, + 0.04521630331873894, + 0.08719301968812943, + 0.058192748576402664, + 0.04216495901346207, + 0.04830190911889076, + 0.4078989326953888, + 0.08118385076522827, + 0.04981820657849312 + ], + "credit_map": [ + 0.023644892498850822, + 0.014057490043342113, + 0.020093968138098717, + 0.012071505188941956, + 0.013059462420642376, + 0.009728863835334778, + 0.015297425910830498, + 0.008424500934779644, + 0.011342119425535202, + 0.00579189695417881, + 0.009976772591471672, + 0.008521183393895626, + 0.007382991258054972, + 0.006405523046851158, + 0.006470834836363792, + 0.007760651875287294, + 0.005337177775800228, + 0.005887337028980255, + 0.006959849502891302, + 0.0044655450619757175, + 0.008162735030055046, + 0.007493882440030575, + 0.003211798844859004, + 0.0036979797296226025, + 0.015233270823955536, + 0.011358232237398624, + 0.018880173563957214, + 0.012656320817768574, + 0.008990682661533356, + 0.006411953829228878, + 0.005783401429653168, + 0.007878982461988926, + 0.007300014607608318, + 0.007467429153621197, + 0.009906642138957977, + 0.004343806300312281, + 0.010954796336591244, + 0.005612080916762352, + 0.008256092667579651, + 0.004351886920630932, + 0.008858065120875835, + 0.009572707116603851, + 0.005599491763859987, + 0.004881965462118387, + 0.0034200670197606087, + 0.0031422912143170834, + 0.0026059886440634727, + 0.004084383603185415, + 0.006753043737262487, + 0.0060856225900352, + 0.003956628032028675, + 0.0037765975575894117, + 0.00889988336712122, + 0.006091209128499031, + 0.006410978734493256, + 0.006327945273369551, + 0.015618929639458656, + 0.13365285098552704, + 0.1314786970615387, + 0.023569848388433456, + 0.019350912421941757, + 0.021013008430600166, + 0.042555756866931915, + 0.0206589438021183, + 0.009273133240640163, + 0.015247603878378868, + 0.011266926303505898, + 0.008787712082266808, + 0.008961073122918606, + 0.007937905378639698, + 0.007981522008776665, + 0.011727980338037014, + 0.008419458754360676, + 0.005512713920325041, + 0.0069219013676047325, + 0.006170188542455435, + 0.004393103066831827, + 0.005025844089686871, + 0.00753876194357872, + 0.005836233962327242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.005982587579637766 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195031210780144 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018195031210780144, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195031210780144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20607423782348633, + 0.12074451148509979, + 0.06844274699687958, + 0.14751657843589783, + 0.03363092988729477, + 0.055513713508844376, + 0.05087656155228615, + 0.15449663996696472, + 0.03943141922354698, + 0.12327263504266739 + ], + "credit_map": [ + 0.045514971017837524, + 0.028806541115045547, + 0.06264040619134903, + 0.012566882185637951, + 0.016081225126981735, + 0.0088193379342556, + 0.018852166831493378, + 0.012792732566595078, + 0.03457285836338997, + 0.011446358636021614, + 0.03862764313817024, + 0.011650786735117435, + 0.005907097365707159, + 0.005752748344093561, + 0.006372540257871151, + 0.006414483301341534, + 0.008743208833038807, + 0.010353544726967812, + 0.014720276929438114, + 0.008874151855707169, + 0.005160083994269371, + 0.005388966761529446, + 0.008931498043239117, + 0.006271026097238064, + 0.019718443974852562, + 0.01919017918407917, + 0.022978466004133224, + 0.026145007461309433, + 0.01634962484240532, + 0.01429581455886364, + 0.0120907723903656, + 0.016748279333114624, + 0.0037134874146431684, + 0.005028917454183102, + 0.007178900297731161, + 0.004139514174312353, + 0.0030631537083536386, + 0.002575460821390152, + 0.004363874904811382, + 0.0035676269326359034, + 0.0074445526115596294, + 0.007470425684005022, + 0.008734936825931072, + 0.006054303143173456, + 0.00898903701454401, + 0.0039726621471345425, + 0.004252430982887745, + 0.008595369756221771, + 0.007456446066498756, + 0.007124592550098896, + 0.00792698934674263, + 0.004384583327919245, + 0.0061110383830964565, + 0.0051971483044326305, + 0.006001277826726437, + 0.006674489937722683, + 0.012497549876570702, + 0.04064932093024254, + 0.03590673208236694, + 0.01374287623912096, + 0.017516063526272774, + 0.011056510731577873, + 0.014613739214837551, + 0.008513865061104298, + 0.004055733792483807, + 0.006163397803902626, + 0.008813705295324326, + 0.0057989927008748055, + 0.003563899779692292, + 0.0024556200951337814, + 0.0036467695608735085, + 0.004933303687721491, + 0.0191609226167202, + 0.013723192736506462, + 0.03661666065454483, + 0.013685050420463085, + 0.005810344126075506, + 0.0072058867663145065, + 0.013833041302859783, + 0.01323755457997322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.018195031210780144 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642767556011677 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642767556011677, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642767556011677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37408578395843506, + 0.05370010808110237, + 0.10250747203826904, + 0.11439696699380875, + 0.038447123020887375, + 0.07999265193939209, + 0.03779440373182297, + 0.07090287655591965, + 0.059463340789079666, + 0.06870932877063751 + ], + "credit_map": [ + 0.0683714896440506, + 0.06352417916059494, + 0.11314468830823898, + 0.02422829158604145, + 0.021818596869707108, + 0.017904488369822502, + 0.03680871054530144, + 0.028285307809710503, + 0.010532137006521225, + 0.0073897577822208405, + 0.012669486925005913, + 0.005019407253712416, + 0.004687212407588959, + 0.004453089088201523, + 0.004204767756164074, + 0.0047442493960261345, + 0.01384010910987854, + 0.0157073512673378, + 0.023623600602149963, + 0.011120441369712353, + 0.006404569838196039, + 0.00878904014825821, + 0.012803086079657078, + 0.010219266638159752, + 0.018884679302573204, + 0.014589198864996433, + 0.01999768801033497, + 0.01919461414217949, + 0.0097713153809309, + 0.008920756168663502, + 0.009909329004585743, + 0.013129379600286484, + 0.0054479637183249, + 0.0048043131828308105, + 0.007609286345541477, + 0.0042497399263083935, + 0.0035224019084125757, + 0.004447386134415865, + 0.00463536661118269, + 0.0037306633312255144, + 0.011369641870260239, + 0.013545187190175056, + 0.016496019437909126, + 0.009369698353111744, + 0.008800307288765907, + 0.007199653424322605, + 0.0054427157156169415, + 0.007769430056214333, + 0.004869766999036074, + 0.004852060228586197, + 0.004424316808581352, + 0.003961421083658934, + 0.0033155286218971014, + 0.004984904546290636, + 0.004733107518404722, + 0.006653294898569584, + 0.007661629933863878, + 0.010692639276385307, + 0.009086096659302711, + 0.01238893810659647, + 0.008825649507343769, + 0.00782782956957817, + 0.00860278494656086, + 0.005817298777401447, + 0.007134119048714638, + 0.009704756550490856, + 0.009375077672302723, + 0.014060935005545616, + 0.0036225169897079468, + 0.005829897243529558, + 0.003369716927409172, + 0.006366315297782421, + 0.005329452455043793, + 0.009891299530863762, + 0.010105657391250134, + 0.009387821890413761, + 0.0059116012416779995, + 0.00767367472872138, + 0.012281469069421291, + 0.008128350600600243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.008642767556011677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely close-up view of the lower half of a giant panda's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panda's fur depicted as dense white around the muzzle and deep black along the jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the panda's jaw moving in a steady rhythm to crush a thick, vibrant green bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright green bamboo fibers mixed with saliva visible clinging to the panda's black lips and the broken bamboo end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still and held at eye level with a completely blurred background of greens and browns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262307643890381 + ], + "opposite_logprob_tokens": [ + -0.5762307643890381 + ], + "logprob_reward": -0.8262307643890381, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262307643890381, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13982897996902466, + 0.20527765154838562, + 0.059402596205472946, + 0.04102558642625809, + 0.05883845314383507, + 0.04828009381890297, + 0.04306786507368088, + 0.3104678690433502, + 0.052404675632715225, + 0.04140622913837433 + ], + "credit_map": [ + 0.02963992767035961, + 0.01774612069129944, + 0.0386018231511116, + 0.009450402110815048, + 0.012104349210858345, + 0.00911884568631649, + 0.01600535586476326, + 0.007162163034081459, + 0.06806530058383942, + 0.018041234463453293, + 0.07312748581171036, + 0.014408707618713379, + 0.006533223669975996, + 0.008130179718136787, + 0.007993937470018864, + 0.00897758174687624, + 0.009703675284981728, + 0.009060447104275227, + 0.009736215695738792, + 0.00658729812130332, + 0.0046605621464550495, + 0.00778039637953043, + 0.00598902627825737, + 0.005884977523237467, + 0.0074701509438455105, + 0.006897146347910166, + 0.007804816588759422, + 0.005244776140898466, + 0.00425147358328104, + 0.0029577144887298346, + 0.0034192234743386507, + 0.0029802832286804914, + 0.006281297188252211, + 0.007467873860150576, + 0.010102233849465847, + 0.004366034176200628, + 0.00902223028242588, + 0.006513399537652731, + 0.010492482222616673, + 0.00459290063008666, + 0.005413503386080265, + 0.008544252254068851, + 0.007989596575498581, + 0.0045075202360749245, + 0.006727686617523432, + 0.004338832106441259, + 0.0035842948127537966, + 0.007174411788582802, + 0.0054231747053563595, + 0.005063415039330721, + 0.0038213052321225405, + 0.004858072847127914, + 0.007243158295750618, + 0.006513868924230337, + 0.0036826853174716234, + 0.006462185177952051, + 0.012585078366100788, + 0.11633020639419556, + 0.1173684149980545, + 0.020414255559444427, + 0.014914234168827534, + 0.006588938646018505, + 0.015290869399905205, + 0.006975862663239241, + 0.005329775158315897, + 0.007280468940734863, + 0.00784354843199253, + 0.006107992492616177, + 0.007108665537089109, + 0.005707897711545229, + 0.006650615483522415, + 0.0063757095485925674, + 0.007918345741927624, + 0.006155811250209808, + 0.005532633513212204, + 0.0051422929391264915, + 0.004068609792739153, + 0.0035192735958844423, + 0.0037890232633799314, + 0.005280239041894674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.8262307643890381 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_9/index_tensor([5443, 5443])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_9/index_tensor([5443, 5443])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2a2540597767ab90a808d778a5032e03a773d6af --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_9/index_tensor([5443, 5443])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle, long shot captured from an underwater eye-level perspective captures a panoramic view of the vast coral reef. The camera pans slowly and steadily to the right, following the subject's movements. The clear tropical waters, filtered by direct overhead sunlight, create countless dancing, flickering rays of light, illuminating the complex and colorful seabed topography of staghorn corals, brain corals, and soft sea fans in the background. In the foreground center of the frame, a small damselfish initially hovers in the water, its scales shimmering in vibrant hues of electric blue and lemon yellow, its translucent fins edged with delicate lines of bright blue. Suddenly, its body tenses, its tail fin swung sharply, transforming into a blur of blue and yellow lightning. With astonishing force, it lunges forward. In a fraction of a second, its tiny mouth opens with precise precision, swallowing a tiny, almost invisible plankton with lightning speed, before quickly resuming its hovering position. The entire picture is highly saturated in color, with azure as the base tone, dotted with the colorful corals and bright colors of the damselfish, perfectly capturing the energetic sense of life surging beneath the tranquility of the ocean world.", + "seed": 8449, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_9/index_5443_seed_8449_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1440482484176755, + "vlm_logprob": -0.1440482484176755, + "vlm_yes_no_margin": 2.875000011920929, + "vlm_token_credit": -0.1440482484176755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011307324282824993 + ], + "opposite_logprob_tokens": [ + -4.636307239532471 + ], + "logprob_reward": -0.011307324282824993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011307324282824993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.775301992893219, + 0.059664200991392136, + 0.014517980627715588, + 0.0250848401337862, + 0.01185393426567316, + 0.03589499369263649, + 0.017118185758590698, + 0.01559971459209919, + 0.020648006349802017, + 0.02431616187095642 + ], + "credit_map": [ + 0.11153152585029602, + 0.3375864624977112, + 0.19602181017398834, + 0.019941076636314392, + 0.038367316126823425, + 0.03116425685584545, + 0.0171186700463295, + 0.023570837453007698, + 0.0056938850320875645, + 0.013313417322933674, + 0.01311990711838007, + 0.005449699703603983, + 0.005923598073422909, + 0.005616804119199514, + 0.005316112190485001, + 0.005230775102972984, + 0.002082852413877845, + 0.0026252479292452335, + 0.0018098310101777315, + 0.0019109563436359167, + 0.0015710886800661683, + 0.0015795518411323428, + 0.0015088643413037062, + 0.001429587951861322, + 0.003523305058479309, + 0.0036864664871245623, + 0.0031467487569898367, + 0.003040449693799019, + 0.0033180825412273407, + 0.0017697203438729048, + 0.0021331841126084328, + 0.004466882906854153, + 0.003318103961646557, + 0.0009595553274266422, + 0.0034712504129856825, + 0.000990107306279242, + 0.0009997915476560593, + 0.0007276469841599464, + 0.0006481930031441152, + 0.0007392870611511171, + 0.011168969795107841, + 0.004732940811663866, + 0.007760209497064352, + 0.0024513602256774902, + 0.0038562200497835875, + 0.002139269607141614, + 0.0018964445916935802, + 0.0018895786488428712, + 0.004027678165584803, + 0.002368774265050888, + 0.002580172149464488, + 0.0013429350219666958, + 0.0015755019849166274, + 0.002508916426450014, + 0.0011648698709905148, + 0.0015493378741666675, + 0.0015895861433818936, + 0.0031383365858346224, + 0.0018934397958219051, + 0.0012471042573451996, + 0.0031928548123687506, + 0.002244718372821808, + 0.001079962239600718, + 0.001213712152093649, + 0.0022275173105299473, + 0.001942016533575952, + 0.0028582820668816566, + 0.0024660336785018444, + 0.0024731324519962072, + 0.0040046293288469315, + 0.0021542098838835955, + 0.002522184746339917, + 0.0020146954338997602, + 0.002839634194970131, + 0.002073566662147641, + 0.002689950866624713, + 0.003760774154216051, + 0.0067497980780899525, + 0.0025031298864632845, + 0.0016846128273755312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.011307324282824993 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348749175667763 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348749175667763, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348749175667763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8562007546424866, + 0.010128561407327652, + 0.006206186022609472, + 0.019961323589086533, + 0.004799370188266039, + 0.005719766952097416, + 0.008067307993769646, + 0.005066742189228535, + 0.046190619468688965, + 0.03765931725502014 + ], + "credit_map": [ + 0.10759354382753372, + 0.36165499687194824, + 0.2215147614479065, + 0.02779826894402504, + 0.0414583757519722, + 0.038961321115493774, + 0.02111189067363739, + 0.03610767424106598, + 0.0010472906287759542, + 0.0012755580246448517, + 0.0015580195467919111, + 0.0011205410119146109, + 0.0008628859650343657, + 0.0008812687010504305, + 0.0016419016756117344, + 0.0017410966102033854, + 0.0009137223823927343, + 0.0011497706873342395, + 0.0008413273608312011, + 0.0009129105019383132, + 0.0005570740322582424, + 0.0007126731215976179, + 0.0006176289753057063, + 0.000501079426612705, + 0.0025434724520891905, + 0.003731263568624854, + 0.0031023474875837564, + 0.003121773712337017, + 0.0023841983638703823, + 0.0015076041454449296, + 0.0015119552845135331, + 0.002058710204437375, + 0.0007045804522931576, + 0.0006587175885215402, + 0.0007521911174990237, + 0.0006610105629079044, + 0.0005402720416896045, + 0.00055689801229164, + 0.0005044103018008173, + 0.0004212906933389604, + 0.0012617679312825203, + 0.0006278297514654696, + 0.0010355872800573707, + 0.0006823790026828647, + 0.0006229404825717211, + 0.0004896176978945732, + 0.0004652129136957228, + 0.0005344326491467655, + 0.0018991527613252401, + 0.0007347944192588329, + 0.0014519748510792851, + 0.0005341783980838954, + 0.0007092239102348685, + 0.0011359913041815162, + 0.0006744709680788219, + 0.0009275225456804037, + 0.00047493723104707897, + 0.0010244919685646892, + 0.0005977226537652314, + 0.00038529877201654017, + 0.0006523263291455805, + 0.0011554309166967869, + 0.0003035130212083459, + 0.00047302187886089087, + 0.0034298596438020468, + 0.004800435621291399, + 0.005615399684756994, + 0.0051271612755954266, + 0.005990029778331518, + 0.012166944332420826, + 0.005223045591264963, + 0.0038377451710402966, + 0.0027696043252944946, + 0.004029107745736837, + 0.002624921500682831, + 0.0034897252917289734, + 0.006007161922752857, + 0.01178409717977047, + 0.004144032020121813, + 0.002810669830068946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.026348749175667763 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133177757263184 + ], + "opposite_logprob_tokens": [ + -1.4063317775726318 + ], + "logprob_reward": -0.28133177757263184, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133177757263184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8235222101211548, + 0.007160102017223835, + 0.008204936049878597, + 0.013614804483950138, + 0.00928383506834507, + 0.008810046128928661, + 0.006983540486544371, + 0.007917456328868866, + 0.07242981344461441, + 0.04207315668463707 + ], + "credit_map": [ + 0.12285204976797104, + 0.3689945936203003, + 0.20915429294109344, + 0.019468527287244797, + 0.035534098744392395, + 0.03474580496549606, + 0.010633480735123158, + 0.022139469161629677, + 0.0006874474929645658, + 0.0010654921643435955, + 0.0008878776570782065, + 0.0007577738142572343, + 0.0008011927711777389, + 0.0007811462855897844, + 0.0006662461091764271, + 0.0015129270032048225, + 0.0011064711725339293, + 0.0013255039229989052, + 0.0009225242538377643, + 0.0010912143625319004, + 0.0009498888393864036, + 0.0007422710186801851, + 0.0011803191155195236, + 0.0008867441792972386, + 0.0018255015602335334, + 0.0025376181583851576, + 0.002138837706297636, + 0.0018555602291598916, + 0.0016108436975628138, + 0.0012231223518028855, + 0.0012440672144293785, + 0.001179255312308669, + 0.0020172488875687122, + 0.0011298522586002946, + 0.002311503514647484, + 0.0012311649043112993, + 0.0005789059214293957, + 0.0006542575429193676, + 0.0006852486403658986, + 0.0006756556103937328, + 0.0024366965517401695, + 0.0008820744697004557, + 0.00203348184004426, + 0.0008023003465496004, + 0.0008466895087622106, + 0.000717295566573739, + 0.0005572242080233991, + 0.0005342849181033671, + 0.0015450535574927926, + 0.0007691925275139511, + 0.0011587300105020404, + 0.0005981682916171849, + 0.0009228343842551112, + 0.0009724461706355214, + 0.0005243977066129446, + 0.0004927187110297382, + 0.0006611886201426387, + 0.0015895531978458166, + 0.0006776831578463316, + 0.0005549004417844117, + 0.0011231637327000499, + 0.0019115988397970796, + 0.0006804676377214491, + 0.0007189018069766462, + 0.0055421399883925915, + 0.007762746419757605, + 0.008197742514312267, + 0.007502950727939606, + 0.009881946258246899, + 0.01984645053744316, + 0.007963860407471657, + 0.005731988698244095, + 0.0033379215747117996, + 0.004349717870354652, + 0.0031967717222869396, + 0.0036471528001129627, + 0.0063567697070539, + 0.01305360160768032, + 0.0047833966091275215, + 0.003347828285768628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.28133177757263184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202821999788284 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202821999788284, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202821999788284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7147635817527771, + 0.025887910276651382, + 0.009472338482737541, + 0.04752590134739876, + 0.05087444186210632, + 0.06509378552436829, + 0.018250465393066406, + 0.009744993411004543, + 0.029299145564436913, + 0.029087549075484276 + ], + "credit_map": [ + 0.11458311229944229, + 0.3251204788684845, + 0.17646466195583344, + 0.015301682986319065, + 0.03125074505805969, + 0.028597654774785042, + 0.008060324937105179, + 0.015384849160909653, + 0.0035653701052069664, + 0.005315129645168781, + 0.004185102414339781, + 0.0028400185983628035, + 0.0030533226672559977, + 0.0025281168054789305, + 0.001917968736961484, + 0.0024828799068927765, + 0.0016782335005700588, + 0.0013095742324367166, + 0.0011732826242223382, + 0.0012993811396881938, + 0.0009660493233241141, + 0.0009312236215919256, + 0.0010087547125294805, + 0.00110583973582834, + 0.005784268490970135, + 0.008692106232047081, + 0.0060145496390759945, + 0.006317358464002609, + 0.006418504752218723, + 0.003892514854669571, + 0.004844915121793747, + 0.005561682395637035, + 0.010500050149857998, + 0.0071542649529874325, + 0.008385196328163147, + 0.00640435004606843, + 0.007503089960664511, + 0.0041402108035981655, + 0.0038508025463670492, + 0.002936475444585085, + 0.022427871823310852, + 0.006960587576031685, + 0.015257831662893295, + 0.005520267877727747, + 0.007223089691251516, + 0.0024971936363726854, + 0.002475283807143569, + 0.002731654094532132, + 0.00469942158088088, + 0.0031594904139637947, + 0.0027441305574029684, + 0.00129717867821455, + 0.001423985930159688, + 0.0013712227810174227, + 0.0013386488426476717, + 0.0022163873072713614, + 0.0009325748542323709, + 0.00180715078022331, + 0.0008115675300359726, + 0.0006091385730542243, + 0.002529827645048499, + 0.001753227785229683, + 0.0007017562165856361, + 0.000599749677348882, + 0.002581421984359622, + 0.0031790423672646284, + 0.0030600770842283964, + 0.0033958624117076397, + 0.003925916738808155, + 0.007520688697695732, + 0.003124559298157692, + 0.0025115772150456905, + 0.002351268893107772, + 0.003095831722021103, + 0.0025913126301020384, + 0.002221642294898629, + 0.005029893014580011, + 0.0071030971594154835, + 0.004024811554700136, + 0.0026696904096752405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.014202821999788284 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870505690574646 + ], + "opposite_logprob_tokens": [ + -1.1370506286621094 + ], + "logprob_reward": -0.3870505690574646, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870505690574646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9614667296409607, + 0.005074517801403999, + 0.002133423462510109, + 0.0033770424779504538, + 0.006720658857375383, + 0.008706668391823769, + 0.0025573663879185915, + 0.0022997481282800436, + 0.0024173154961317778, + 0.005246548913419247 + ], + "credit_map": [ + 0.14830435812473297, + 0.43646490573883057, + 0.24450933933258057, + 0.02081369049847126, + 0.04122903570532799, + 0.03901169076561928, + 0.00995059683918953, + 0.021183105185627937, + 0.0004963370738551021, + 0.0007488718838430941, + 0.0010939870262518525, + 0.0003661819500848651, + 0.0004262664879206568, + 0.0005266137886792421, + 0.00043016529525630176, + 0.0009860944701358676, + 0.0002869176387321204, + 0.00034054488060064614, + 0.00021829035540577024, + 0.00029819755582138896, + 0.000301794963888824, + 0.00024109298828989267, + 0.0002208403020631522, + 0.00022574463218916208, + 0.0004990386078134179, + 0.00044864421943202615, + 0.0004399410099722445, + 0.0005215484998188913, + 0.00034581447835080326, + 0.00032499077497050166, + 0.00033506922773085535, + 0.0004619958344846964, + 0.0019119562348350883, + 0.0005001359386369586, + 0.001921535236760974, + 0.0007673510699532926, + 0.0004141298995818943, + 0.0004452046996448189, + 0.0003632621082942933, + 0.000397083698771894, + 0.0031275376677513123, + 0.0008841473027132452, + 0.0020024324767291546, + 0.0008442050311714411, + 0.0008674372802488506, + 0.00034538141335360706, + 0.00034613447496667504, + 0.0002893925702665001, + 0.0006557804299518466, + 0.0003452939272392541, + 0.0003944859199691564, + 0.00020677896100096405, + 0.00022380590962711722, + 0.00030348222935572267, + 0.00021443719742819667, + 0.0002133018570020795, + 0.0002010720781981945, + 0.0002985781757161021, + 0.00025895913131535053, + 0.00025606239796616137, + 0.00037056024302728474, + 0.0004939275677315891, + 0.0002170746010961011, + 0.00020351393322926015, + 0.00022668461315333843, + 0.0002699453034438193, + 0.00032168361940421164, + 0.0002854892809409648, + 0.00027532674721442163, + 0.00038218076224438846, + 0.00028681999538093805, + 0.0003691852034535259, + 0.0003944148193113506, + 0.0005417062202468514, + 0.0010479040211066604, + 0.0007455167360603809, + 0.0005299564218148589, + 0.0009176530293188989, + 0.0005216242861934006, + 0.0005477735539898276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.3870505690574646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot captured from an underwater eye-level perspective captures a panoramic view of the vast coral reef. The camera pans slowly and steadily to the right, following the subject's movements. The clear tropical waters, filtered by direct overhead sunlight, create countless dancing, flickering rays of light, illuminating the complex and colorful seabed topography of staghorn corals, brain corals, and soft sea fans in the background. In the foreground center of the frame, a small damselfish initially hovers in the water, its scales shimmering in vibrant hues of electric blue and lemon yellow, its translucent fins edged with delicate lines of bright blue. Suddenly, its body tenses, its tail fin swung sharply, transforming into a blur of blue and yellow lightning. With astonishing force, it lunges forward. In a fraction of a second, its tiny mouth opens with precise precision, swallowing a tiny, almost invisible plankton with lightning speed, before quickly resuming its hovering position. The entire picture is highly saturated in color, with azure as the base tone, dotted with the colorful corals and bright colors of the damselfish, perfectly capturing the energetic sense of life surging beneath the tranquility of the ocean world.", + "seed": 8450, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_9/index_5443_seed_8450_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16494989339262248, + "vlm_logprob": -0.16494989339262248, + "vlm_yes_no_margin": 2.6000000119209288, + "vlm_token_credit": -0.16494989339262248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010591485537588596 + ], + "opposite_logprob_tokens": [ + -4.760591506958008 + ], + "logprob_reward": -0.010591485537588596, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010591485537588596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056298576295375824, + 0.032522622495889664, + 0.6257120370864868, + 0.024883899837732315, + 0.024152370169758797, + 0.016022436320781708, + 0.013692405074834824, + 0.022947201505303383, + 0.09460355341434479, + 0.0891648679971695 + ], + "credit_map": [ + 0.005450340919196606, + 0.006748754531145096, + 0.006539097987115383, + 0.006104588508605957, + 0.0071423896588385105, + 0.005404915194958448, + 0.013703728094696999, + 0.005204763729125261, + 0.006840539164841175, + 0.004822524264454842, + 0.003508553374558687, + 0.0030068722553551197, + 0.004005770199000835, + 0.0022199549712240696, + 0.0054661487229168415, + 0.002652258612215519, + 0.1760578751564026, + 0.24386818706989288, + 0.03241419419646263, + 0.023048462346196175, + 0.09176651388406754, + 0.01969337835907936, + 0.023829249665141106, + 0.015034167096018791, + 0.0030584854539483786, + 0.003573445603251457, + 0.00724114291369915, + 0.0026337914168834686, + 0.0018852341454476118, + 0.0013431275729089975, + 0.0036787590943276882, + 0.0014699131716042757, + 0.0034291078336536884, + 0.00279512838460505, + 0.003043902339413762, + 0.002459761220961809, + 0.0028085564263164997, + 0.0029083334375172853, + 0.00347901857458055, + 0.00322856311686337, + 0.00353018450550735, + 0.001918263267725706, + 0.0010941182263195515, + 0.0014957275707274675, + 0.002378105651587248, + 0.0035160111729055643, + 0.0012947889044880867, + 0.0007952370797283947, + 0.002060619881376624, + 0.0016618866939097643, + 0.0017388195265084505, + 0.0016031988197937608, + 0.002033894183114171, + 0.0016706408932805061, + 0.0018150805262848735, + 0.0011082642013207078, + 0.003860888071358204, + 0.0027083735913038254, + 0.0035351805854588747, + 0.001989081036299467, + 0.004103338345885277, + 0.001960471738129854, + 0.002812101738527417, + 0.0019777670968323946, + 0.017374912276864052, + 0.013726393692195415, + 0.0072482796385884285, + 0.008396279066801071, + 0.021603699773550034, + 0.01136799156665802, + 0.008471875451505184, + 0.0064141168259084225, + 0.01487111859023571, + 0.016719048842787743, + 0.008847279474139214, + 0.00560712069272995, + 0.02415834367275238, + 0.00953332707285881, + 0.004256847780197859, + 0.005171779543161392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.010591485537588596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810567781329155 + ], + "opposite_logprob_tokens": [ + -3.2881057262420654 + ], + "logprob_reward": -0.03810567781329155, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810567781329155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05710529163479805, + 0.02963760308921337, + 0.306692898273468, + 0.018381915986537933, + 0.029942970722913742, + 0.0168320219963789, + 0.01425728015601635, + 0.0359276607632637, + 0.2545681893825531, + 0.236654132604599 + ], + "credit_map": [ + 0.0069513386115431786, + 0.007855966687202454, + 0.009358750656247139, + 0.0055589573457837105, + 0.00753453653305769, + 0.006252639926970005, + 0.008445242419838905, + 0.005147858522832394, + 0.004941405262798071, + 0.005106117110699415, + 0.0034281639382243156, + 0.0021822787821292877, + 0.004279620014131069, + 0.0022521503269672394, + 0.0053107962012290955, + 0.002137073315680027, + 0.08594920486211777, + 0.12125858664512634, + 0.014618132263422012, + 0.010948829352855682, + 0.04468296840786934, + 0.009856479242444038, + 0.012097804807126522, + 0.007280874066054821, + 0.0024218950420618057, + 0.002727638930082321, + 0.004103066865354776, + 0.0017452631145715714, + 0.002013091929256916, + 0.00106211774982512, + 0.0029481041710823774, + 0.0013607371365651488, + 0.004007133189588785, + 0.0033436864614486694, + 0.004229940474033356, + 0.003533839713782072, + 0.003184174420312047, + 0.0038057712372392416, + 0.003998254891484976, + 0.003840169869363308, + 0.004775793757289648, + 0.002102500759065151, + 0.001120225409977138, + 0.001531459973193705, + 0.0018727747956290841, + 0.0033854602370411158, + 0.0012237275950610638, + 0.0008200800511986017, + 0.0017517980886623263, + 0.0020520908292382956, + 0.002092901151627302, + 0.0013248256873339415, + 0.0021841456182301044, + 0.0015305845299735665, + 0.0018917485140264034, + 0.0014291860861703753, + 0.0075270794332027435, + 0.0033698061015456915, + 0.00666813226416707, + 0.0025150002911686897, + 0.0069649373181164265, + 0.00290950876660645, + 0.003112023463472724, + 0.0028611719608306885, + 0.04129599407315254, + 0.03676915913820267, + 0.017292682081460953, + 0.026012884452939034, + 0.05290531367063522, + 0.02749873511493206, + 0.03350905328989029, + 0.01928437501192093, + 0.04338991641998291, + 0.04635420814156532, + 0.023625880479812622, + 0.012774825096130371, + 0.060437820851802826, + 0.025082461535930634, + 0.011864743195474148, + 0.013124284334480762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.03810567781329155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289235472679138 + ], + "opposite_logprob_tokens": [ + -1.0539236068725586 + ], + "logprob_reward": -0.4289235472679138, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4289235472679138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027211470529437065, + 0.014356442727148533, + 0.536431074142456, + 0.02107512392103672, + 0.02240055613219738, + 0.015483885072171688, + 0.011467565782368183, + 0.0233005303889513, + 0.08034119755029678, + 0.24793221056461334 + ], + "credit_map": [ + 0.0039467900060117245, + 0.004282727371901274, + 0.004067873116582632, + 0.0023541743867099285, + 0.004586922004818916, + 0.0028601365629583597, + 0.002660423517227173, + 0.0024524221662431955, + 0.002826375886797905, + 0.0023701756726950407, + 0.0016190876485779881, + 0.001250128261744976, + 0.0020734926220029593, + 0.0013490078272297978, + 0.0017909014131873846, + 0.001077271648682654, + 0.1474926471710205, + 0.21068087220191956, + 0.02674606442451477, + 0.019817769527435303, + 0.08098351210355759, + 0.017305761575698853, + 0.02044089324772358, + 0.012963502667844296, + 0.0028101876378059387, + 0.0029137444216758013, + 0.005472607910633087, + 0.0020215145777910948, + 0.0019633378833532333, + 0.0013265126617625356, + 0.0028302730061113834, + 0.0017369443085044622, + 0.002886852715164423, + 0.0024892380461096764, + 0.0028801592998206615, + 0.002649793867021799, + 0.002704392187297344, + 0.0029038088396191597, + 0.0031417948193848133, + 0.0027445140294730663, + 0.004155140835791826, + 0.0019039349863305688, + 0.0009192148572765291, + 0.0010762992314994335, + 0.002076659584417939, + 0.0034334713127464056, + 0.001051032799296081, + 0.0008681293111294508, + 0.0015236095059663057, + 0.0015190399717539549, + 0.001432572491466999, + 0.0017762846546247602, + 0.0018317523645237088, + 0.0010160983074456453, + 0.0010657061357051134, + 0.0013025013031437993, + 0.005154785700142384, + 0.0016981561202555895, + 0.004946679342538118, + 0.0026948507875204086, + 0.003982783295214176, + 0.0012284525437280536, + 0.0018422366119921207, + 0.0017525835428386927, + 0.013277919963002205, + 0.010898198932409286, + 0.005076147615909576, + 0.008109007962048054, + 0.018444761633872986, + 0.008735228329896927, + 0.010088094510138035, + 0.00571183767169714, + 0.04523714631795883, + 0.04832081496715546, + 0.023619025945663452, + 0.013080372475087643, + 0.0648503229022026, + 0.02718612365424633, + 0.012131834402680397, + 0.013506576418876648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.4289235472679138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370974212884903 + ], + "opposite_logprob_tokens": [ + -3.408709764480591 + ], + "logprob_reward": -0.03370974212884903, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370974212884903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055172547698020935, + 0.05847148224711418, + 0.42701756954193115, + 0.026296036317944527, + 0.02479369379580021, + 0.1362820565700531, + 0.05053376778960228, + 0.02048088051378727, + 0.1218339204788208, + 0.07911806553602219 + ], + "credit_map": [ + 0.013387009501457214, + 0.010281834751367569, + 0.008146695792675018, + 0.003526661777868867, + 0.009072378277778625, + 0.003229668829590082, + 0.004442577250301838, + 0.0030857205856591463, + 0.014616822823882103, + 0.011622000485658646, + 0.006307673174887896, + 0.004184095188975334, + 0.008314976468682289, + 0.003973500337451696, + 0.005564607214182615, + 0.003887803526595235, + 0.11495748907327652, + 0.16859908401966095, + 0.020294606685638428, + 0.01657164841890335, + 0.06504657119512558, + 0.014076702296733856, + 0.01696780137717724, + 0.010503618977963924, + 0.0031701531261205673, + 0.003566217143088579, + 0.008928002789616585, + 0.0024384709540754557, + 0.0016448834212496877, + 0.00135713757481426, + 0.0035786309745162725, + 0.0016125383554026484, + 0.00569329597055912, + 0.003957646898925304, + 0.0032791970297694206, + 0.002061655046418309, + 0.002632411662489176, + 0.003219081787392497, + 0.00220426800660789, + 0.001746134483255446, + 0.049503471702337265, + 0.02172250673174858, + 0.005993502214550972, + 0.0059808590449392796, + 0.005615887697786093, + 0.04070667922496796, + 0.0038261045701801777, + 0.0029330430552363396, + 0.008345158770680428, + 0.01012355461716652, + 0.006643941160291433, + 0.006216633599251509, + 0.00737143075093627, + 0.004495540168136358, + 0.004406062886118889, + 0.0029314435087144375, + 0.003948885016143322, + 0.0022898733150213957, + 0.003971878904849291, + 0.0014806095277890563, + 0.004047370981425047, + 0.0016456317389383912, + 0.0015539685264229774, + 0.0015426600584760308, + 0.018762484192848206, + 0.01652497611939907, + 0.008495420217514038, + 0.014555762521922588, + 0.025975381955504417, + 0.011915693990886211, + 0.016628490760922432, + 0.008975714445114136, + 0.008621401153504848, + 0.011990036815404892, + 0.0058586779050529, + 0.004143513273447752, + 0.024935143068432808, + 0.015416900627315044, + 0.0038635472301393747, + 0.004288843832910061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.03370974212884903 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341901421546936 + ], + "opposite_logprob_tokens": [ + -1.313418984413147 + ], + "logprob_reward": -0.31341901421546936, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341901421546936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10850613564252853, + 0.0515328124165535, + 0.35308074951171875, + 0.0867665633559227, + 0.0564841628074646, + 0.042322613298892975, + 0.01961062289774418, + 0.07707222551107407, + 0.0709632933139801, + 0.13366089761257172 + ], + "credit_map": [ + 0.01539409440010786, + 0.01601121947169304, + 0.01490175910294056, + 0.01049176324158907, + 0.01641884259879589, + 0.010421675629913807, + 0.014490797184407711, + 0.01037598866969347, + 0.0107616251334548, + 0.007850192487239838, + 0.005526361055672169, + 0.00456283101812005, + 0.0054460554383695126, + 0.0035039992071688175, + 0.010598722845315933, + 0.003283025464043021, + 0.09777410328388214, + 0.13773465156555176, + 0.018575282767415047, + 0.012835116125643253, + 0.052264224737882614, + 0.011378292925655842, + 0.013980453833937645, + 0.008538639172911644, + 0.009428535588085651, + 0.0106272017583251, + 0.025956088677048683, + 0.009441009722650051, + 0.007031388580799103, + 0.006010010372847319, + 0.011570664122700691, + 0.006701665464788675, + 0.010002008639276028, + 0.006998055148869753, + 0.006899969186633825, + 0.006624322850257158, + 0.004990970250219107, + 0.006802645977586508, + 0.00788487121462822, + 0.006281319539994001, + 0.00888324249535799, + 0.004300258122384548, + 0.0032283542677760124, + 0.0059865317307412624, + 0.006407714448869228, + 0.007817444391548634, + 0.003497517202049494, + 0.002201552502810955, + 0.002897784346714616, + 0.0035056755878031254, + 0.0023797075264155865, + 0.002640762133523822, + 0.002402207814157009, + 0.0019575220067054033, + 0.0022925445809960365, + 0.0015344178536906838, + 0.011749000288546085, + 0.01570850983262062, + 0.008111387491226196, + 0.009436918422579765, + 0.012697829864919186, + 0.006381784565746784, + 0.007528571877628565, + 0.005458221770823002, + 0.009663917124271393, + 0.01104198582470417, + 0.004081141669303179, + 0.0084567591547966, + 0.01723722741007805, + 0.006525591481477022, + 0.008836294524371624, + 0.005120373796671629, + 0.0178117286413908, + 0.019518109038472176, + 0.014394332654774189, + 0.010006344877183437, + 0.028961529955267906, + 0.02362293004989624, + 0.008651679381728172, + 0.01069424394518137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.31341901421546936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot captured from an underwater eye-level perspective captures a panoramic view of the vast coral reef. The camera pans slowly and steadily to the right, following the subject's movements. The clear tropical waters, filtered by direct overhead sunlight, create countless dancing, flickering rays of light, illuminating the complex and colorful seabed topography of staghorn corals, brain corals, and soft sea fans in the background. In the foreground center of the frame, a small damselfish initially hovers in the water, its scales shimmering in vibrant hues of electric blue and lemon yellow, its translucent fins edged with delicate lines of bright blue. Suddenly, its body tenses, its tail fin swung sharply, transforming into a blur of blue and yellow lightning. With astonishing force, it lunges forward. In a fraction of a second, its tiny mouth opens with precise precision, swallowing a tiny, almost invisible plankton with lightning speed, before quickly resuming its hovering position. The entire picture is highly saturated in color, with azure as the base tone, dotted with the colorful corals and bright colors of the damselfish, perfectly capturing the energetic sense of life surging beneath the tranquility of the ocean world.", + "seed": 8451, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_9/index_5443_seed_8451_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18885718006640673, + "vlm_logprob": -0.18885718006640673, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.18885718006640673, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007210659794509411 + ], + "opposite_logprob_tokens": [ + -5.257210731506348 + ], + "logprob_reward": -0.007210659794509411, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007210659794509411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05634842440485954, + 0.1121920570731163, + 0.0568138062953949, + 0.03935700282454491, + 0.04613569378852844, + 0.018501969054341316, + 0.05288452282547951, + 0.03761942684650421, + 0.5339716076850891, + 0.04617543891072273 + ], + "credit_map": [ + 0.006444808095693588, + 0.006818037945777178, + 0.009198666550219059, + 0.006125248968601227, + 0.008448612876236439, + 0.007239483762532473, + 0.005800582468509674, + 0.0062729897908866405, + 0.033981803804636, + 0.00908102560788393, + 0.006061682011932135, + 0.0034312058705836535, + 0.008180340752005577, + 0.040867555886507034, + 0.004166523925960064, + 0.0064219278283417225, + 0.006506826728582382, + 0.007192500866949558, + 0.006412881892174482, + 0.004775799810886383, + 0.007310574874281883, + 0.005677750799804926, + 0.004403311759233475, + 0.014534165151417255, + 0.007281207945197821, + 0.0036588250659406185, + 0.004826305899769068, + 0.003868357976898551, + 0.0048220171593129635, + 0.0045874398201704025, + 0.004708035383373499, + 0.0056048170663416386, + 0.003358351532369852, + 0.003451719880104065, + 0.003139097010716796, + 0.007251839153468609, + 0.003853960195556283, + 0.0060025835409760475, + 0.014438515529036522, + 0.004639630205929279, + 0.0022812415845692158, + 0.0012899743160232902, + 0.002345516113564372, + 0.0029162156861275434, + 0.002462108852341771, + 0.0017391069559380412, + 0.003121832152828574, + 0.00234597478993237, + 0.003724828362464905, + 0.003593150759115815, + 0.00421922979876399, + 0.0038976208306849003, + 0.018467070534825325, + 0.009040409699082375, + 0.0051839011721313, + 0.004758320748806, + 0.003088731784373522, + 0.003777130274102092, + 0.004142015241086483, + 0.003958489745855331, + 0.006978021468967199, + 0.006395494099706411, + 0.004242167342454195, + 0.0050373803824186325, + 0.01948380470275879, + 0.024844223633408546, + 0.14034685492515564, + 0.05560948699712753, + 0.17726831138134003, + 0.07969311624765396, + 0.02011500485241413, + 0.01661081612110138, + 0.0062344069592654705, + 0.005793309770524502, + 0.0030140953604131937, + 0.002894452540203929, + 0.008012109436094761, + 0.011415204033255577, + 0.0037386163603514433, + 0.005073246546089649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.007210659794509411 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981220744550228 + ], + "opposite_logprob_tokens": [ + -3.5298120975494385 + ], + "logprob_reward": -0.02981220744550228, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981220744550228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08960375189781189, + 0.2976149320602417, + 0.08595405519008636, + 0.016828196123242378, + 0.019788242876529694, + 0.014784643426537514, + 0.032140620052814484, + 0.08705055713653564, + 0.1073407232761383, + 0.24889428913593292 + ], + "credit_map": [ + 0.010677038691937923, + 0.009167238138616085, + 0.012498794123530388, + 0.01215305831283331, + 0.016915293410420418, + 0.009468705393373966, + 0.008926971815526485, + 0.00979664921760559, + 0.03951447829604149, + 0.06360255926847458, + 0.028252974152565002, + 0.022406097501516342, + 0.05593773350119591, + 0.04197324067354202, + 0.021203696727752686, + 0.024724172428250313, + 0.006912658456712961, + 0.005386583507061005, + 0.008868345059454441, + 0.0068756588734686375, + 0.01079338975250721, + 0.01964840292930603, + 0.009694064036011696, + 0.017774958163499832, + 0.0018492029048502445, + 0.0026288600638508797, + 0.0014299158938229084, + 0.001835713512264192, + 0.003161061555147171, + 0.0022620721720159054, + 0.0018221280770376325, + 0.0018392420606687665, + 0.002428553532809019, + 0.002151389140635729, + 0.0014690158423036337, + 0.002194690052419901, + 0.0018642456270754337, + 0.0024059659335762262, + 0.004480915609747171, + 0.002793466206640005, + 0.0015356536023318768, + 0.000822601024992764, + 0.001369819394312799, + 0.0022284812293946743, + 0.0019174079643562436, + 0.0017109920736402273, + 0.0028746570460498333, + 0.002325031440705061, + 0.003011835739016533, + 0.00248632300645113, + 0.005022033117711544, + 0.003852536203339696, + 0.00482561532407999, + 0.0035298734437674284, + 0.005890142638236284, + 0.003522259648889303, + 0.006658968981355429, + 0.008557578548789024, + 0.008395085111260414, + 0.008679445832967758, + 0.01651121862232685, + 0.01476324163377285, + 0.011640461161732674, + 0.011844562366604805, + 0.005953985266387463, + 0.009982006624341011, + 0.017615079879760742, + 0.011047878302633762, + 0.023802785202860832, + 0.01794431544840336, + 0.007898776791989803, + 0.01309590507298708, + 0.018115127459168434, + 0.02981545962393284, + 0.011837298050522804, + 0.014303524978458881, + 0.057511232793331146, + 0.0664355605840683, + 0.02041996270418167, + 0.03045611083507538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.02981220744550228 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761827826499939 + ], + "opposite_logprob_tokens": [ + -0.8261827826499939 + ], + "logprob_reward": -0.5761827826499939, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761827826499939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04903552308678627, + 0.24369016289710999, + 0.25852882862091064, + 0.029086628928780556, + 0.08125902712345123, + 0.024257943034172058, + 0.03290475904941559, + 0.03973354026675224, + 0.053258031606674194, + 0.18824560940265656 + ], + "credit_map": [ + 0.005305835045874119, + 0.007721468806266785, + 0.006097886711359024, + 0.005763276945799589, + 0.00661853002384305, + 0.004587579984217882, + 0.005939438007771969, + 0.007001502439379692, + 0.08092936873435974, + 0.013166788034141064, + 0.010932723991572857, + 0.005691779311746359, + 0.01310854684561491, + 0.10286843031644821, + 0.006270130630582571, + 0.01072237454354763, + 0.015929054468870163, + 0.012171591632068157, + 0.02206754870712757, + 0.016659153625369072, + 0.04009596258401871, + 0.0764516219496727, + 0.019804013893008232, + 0.055349864065647125, + 0.007926237769424915, + 0.0024361400865018368, + 0.003779105143621564, + 0.0025005461648106575, + 0.0022996445186436176, + 0.003524451283738017, + 0.0030093227978795767, + 0.00361118046566844, + 0.00915821548551321, + 0.01844007521867752, + 0.007826450280845165, + 0.007927332073450089, + 0.013640896417200565, + 0.009002869948744774, + 0.007073074579238892, + 0.008190102875232697, + 0.0020647274795919657, + 0.0013090969296172261, + 0.002520654583349824, + 0.003763331100344658, + 0.0033783407416194677, + 0.0027182118501514196, + 0.0055611771531403065, + 0.002942400285974145, + 0.0027102807071059942, + 0.0022686959709972143, + 0.005161850713193417, + 0.004369884729385376, + 0.005165059585124254, + 0.004371441435068846, + 0.005625755526125431, + 0.0032317861914634705, + 0.0040526376105844975, + 0.0036742102820426226, + 0.0047920867800712585, + 0.003476001787930727, + 0.005047169979661703, + 0.00487742992118001, + 0.006931759417057037, + 0.006882242858409882, + 0.003560475306585431, + 0.003889557672664523, + 0.011672147549688816, + 0.004615000914782286, + 0.013983923010528088, + 0.007823524065315723, + 0.003066103905439377, + 0.004647293593734503, + 0.01948823779821396, + 0.023524751886725426, + 0.011367167346179485, + 0.013091029599308968, + 0.03292284160852432, + 0.050913307815790176, + 0.015233704820275307, + 0.021704532206058502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.5761827826499939 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900398224592209 + ], + "opposite_logprob_tokens": [ + -2.5790040493011475 + ], + "logprob_reward": -0.07900398224592209, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900398224592209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07842748612165451, + 0.15150253474712372, + 0.06035497039556503, + 0.0850759893655777, + 0.04334070906043053, + 0.032478440552949905, + 0.2759974002838135, + 0.03210989758372307, + 0.10749822854995728, + 0.13321436941623688 + ], + "credit_map": [ + 0.013097623363137245, + 0.005669747944921255, + 0.012833835557103157, + 0.006281339097768068, + 0.010724205523729324, + 0.012452698312699795, + 0.007239147555083036, + 0.010128888301551342, + 0.04163626581430435, + 0.013592884875833988, + 0.008232616819441319, + 0.005270105786621571, + 0.01170742604881525, + 0.0564144141972065, + 0.006314788479357958, + 0.008334038779139519, + 0.005359503906220198, + 0.003498263657093048, + 0.008144434541463852, + 0.005615717731416225, + 0.004055624827742577, + 0.01010097749531269, + 0.006435414310544729, + 0.017145033925771713, + 0.012903926894068718, + 0.00756551418453455, + 0.00920926220715046, + 0.007350444328039885, + 0.014752844348549843, + 0.010693671181797981, + 0.009704864583909512, + 0.012895457446575165, + 0.0044550467282533646, + 0.009125814773142338, + 0.0035889833234250546, + 0.004492190200835466, + 0.006393227726221085, + 0.004267472308129072, + 0.006865692790597677, + 0.004152282141149044, + 0.0037229338195174932, + 0.001953242579475045, + 0.004345821216702461, + 0.004825036507099867, + 0.00553096504881978, + 0.0035381284542381763, + 0.005074041895568371, + 0.003488269867375493, + 0.014179893769323826, + 0.019525909796357155, + 0.02731401100754738, + 0.022812705487012863, + 0.09314026683568954, + 0.045566145330667496, + 0.02753485180437565, + 0.025923607870936394, + 0.002798971254378557, + 0.0039426484145224094, + 0.0032123983837664127, + 0.004036995116621256, + 0.005183091852813959, + 0.004972696304321289, + 0.004098957870155573, + 0.003864139085635543, + 0.004198900889605284, + 0.0060711996629834175, + 0.02577703632414341, + 0.011041102930903435, + 0.032906655222177505, + 0.016232501715421677, + 0.0039846524596214294, + 0.007286174688488245, + 0.009875074960291386, + 0.016931885853409767, + 0.009063228033483028, + 0.008795862086117268, + 0.028359225019812584, + 0.03227889537811279, + 0.013562744483351707, + 0.014347455464303493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.07900398224592209 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207626819610596 + ], + "opposite_logprob_tokens": [ + -1.502076268196106 + ], + "logprob_reward": -0.25207626819610596, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207626819610596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024553222581744194, + 0.03068992681801319, + 0.08643098920583725, + 0.04100548103451729, + 0.05763094127178192, + 0.05937373638153076, + 0.0962492823600769, + 0.35568326711654663, + 0.21293309330940247, + 0.03545014560222626 + ], + "credit_map": [ + 0.004226485267281532, + 0.003257193136960268, + 0.003944158088415861, + 0.0024057519622147083, + 0.0034311991184949875, + 0.002200409770011902, + 0.0017996082315221429, + 0.0032884120009839535, + 0.004659579135477543, + 0.00510003324598074, + 0.0037977383472025394, + 0.0036221127957105637, + 0.004469596780836582, + 0.00303448224440217, + 0.00195751478895545, + 0.00404886482283473, + 0.004699741955846548, + 0.004173435736447573, + 0.00419866107404232, + 0.0036890399642288685, + 0.02582557685673237, + 0.03188326954841614, + 0.004393486306071281, + 0.007567760534584522, + 0.005340089090168476, + 0.0031873120460659266, + 0.00469342852011323, + 0.006303641013801098, + 0.005325899925082922, + 0.005651555489748716, + 0.00462157279253006, + 0.0058819749392569065, + 0.0077383811585605145, + 0.010029852390289307, + 0.005601026583462954, + 0.00783811416476965, + 0.007343264762312174, + 0.006054671946913004, + 0.008067048154771328, + 0.004958571400493383, + 0.007869642227888107, + 0.004492790438234806, + 0.005894722882658243, + 0.009359056130051613, + 0.00633906340226531, + 0.00522661255672574, + 0.014109494164586067, + 0.006082344334572554, + 0.008215874433517456, + 0.008158923126757145, + 0.012283164076507092, + 0.010199119336903095, + 0.021194644272327423, + 0.01328323408961296, + 0.013885188847780228, + 0.009029115550220013, + 0.03210938721895218, + 0.03520343825221062, + 0.04889252781867981, + 0.03908504918217659, + 0.06147097796201706, + 0.05011525750160217, + 0.041135434061288834, + 0.04767110198736191, + 0.008295760490000248, + 0.01297806017100811, + 0.04856256768107414, + 0.02293344959616661, + 0.0649913027882576, + 0.029744746163487434, + 0.007555055897682905, + 0.017872115597128868, + 0.005002439022064209, + 0.004992136266082525, + 0.002828780096024275, + 0.002956244396045804, + 0.005614519119262695, + 0.005491971038281918, + 0.003852201160043478, + 0.004711846821010113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.25207626819610596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot captured from an underwater eye-level perspective captures a panoramic view of the vast coral reef. The camera pans slowly and steadily to the right, following the subject's movements. The clear tropical waters, filtered by direct overhead sunlight, create countless dancing, flickering rays of light, illuminating the complex and colorful seabed topography of staghorn corals, brain corals, and soft sea fans in the background. In the foreground center of the frame, a small damselfish initially hovers in the water, its scales shimmering in vibrant hues of electric blue and lemon yellow, its translucent fins edged with delicate lines of bright blue. Suddenly, its body tenses, its tail fin swung sharply, transforming into a blur of blue and yellow lightning. With astonishing force, it lunges forward. In a fraction of a second, its tiny mouth opens with precise precision, swallowing a tiny, almost invisible plankton with lightning speed, before quickly resuming its hovering position. The entire picture is highly saturated in color, with azure as the base tone, dotted with the colorful corals and bright colors of the damselfish, perfectly capturing the energetic sense of life surging beneath the tranquility of the ocean world.", + "seed": 8452, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000088/group_9/index_5443_seed_8452_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11806288864463568, + "vlm_logprob": -0.11806288864463568, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.11806288864463568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00935929361730814 + ], + "opposite_logprob_tokens": [ + -4.884359359741211 + ], + "logprob_reward": -0.00935929361730814, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00935929361730814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06278740614652634, + 0.0784088522195816, + 0.03267429769039154, + 0.034991826862096786, + 0.040524426847696304, + 0.019177164882421494, + 0.030038105323910713, + 0.05212899297475815, + 0.04789109528064728, + 0.6013778448104858 + ], + "credit_map": [ + 0.007834848016500473, + 0.013611903414130211, + 0.011774408631026745, + 0.009600736200809479, + 0.0047327387146651745, + 0.0056298463605344296, + 0.005185029935091734, + 0.004417890682816505, + 0.007617691997438669, + 0.00893960427492857, + 0.013704990968108177, + 0.010958763770759106, + 0.009799343533813953, + 0.010568865574896336, + 0.01062074862420559, + 0.006198842544108629, + 0.00405071722343564, + 0.004167122766375542, + 0.0036200019530951977, + 0.003948291763663292, + 0.006231118459254503, + 0.0030444238800555468, + 0.003872374538332224, + 0.003740245709195733, + 0.0051588695496320724, + 0.005501710344105959, + 0.0036244334187358618, + 0.003294053953140974, + 0.004188381601125002, + 0.0036683580838143826, + 0.005731790326535702, + 0.003824228886514902, + 0.005781948566436768, + 0.003968036267906427, + 0.0040801409631967545, + 0.00420367531478405, + 0.0049155340529978275, + 0.006005334667861462, + 0.006512696389108896, + 0.0050570606254041195, + 0.00228193961083889, + 0.0019392817048355937, + 0.0028867616783827543, + 0.001693469239398837, + 0.0024472179356962442, + 0.0030129135120660067, + 0.0024137927684932947, + 0.002501789480447769, + 0.003687771735712886, + 0.00289072934538126, + 0.0024386520963162184, + 0.002204124117270112, + 0.005869951564818621, + 0.003074740758165717, + 0.003364227944985032, + 0.006507907994091511, + 0.004640041850507259, + 0.005416002590209246, + 0.003706652671098709, + 0.003438152838498354, + 0.008937978185713291, + 0.008931504562497139, + 0.006919401697814465, + 0.010139258578419685, + 0.0047158366069197655, + 0.005502840969711542, + 0.006016947329044342, + 0.005856714677065611, + 0.0054282150231301785, + 0.0068156300112605095, + 0.008080778643488884, + 0.005474131088703871, + 0.049257151782512665, + 0.04512816667556763, + 0.02372678741812706, + 0.0506260022521019, + 0.04572338983416557, + 0.12193285673856735, + 0.2229696810245514, + 0.042013783007860184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.00935929361730814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746964529156685 + ], + "opposite_logprob_tokens": [ + -5.006746768951416 + ], + "logprob_reward": -0.006746964529156685, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746964529156685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07865116000175476, + 0.09478594362735748, + 0.02459784783422947, + 0.2286008596420288, + 0.014789527282118797, + 0.024107487872242928, + 0.022026479244232178, + 0.01654038205742836, + 0.02346227504312992, + 0.4724380075931549 + ], + "credit_map": [ + 0.011455315165221691, + 0.014066617004573345, + 0.009215334430336952, + 0.008902130648493767, + 0.007547864690423012, + 0.010529971681535244, + 0.009091904386878014, + 0.007842034101486206, + 0.011604111641645432, + 0.0118306465446949, + 0.013769272714853287, + 0.011610939167439938, + 0.010834288783371449, + 0.014135953970253468, + 0.012848175130784512, + 0.008152562193572521, + 0.002633836818858981, + 0.004342408385127783, + 0.003465968184173107, + 0.0024320802185684443, + 0.004308301489800215, + 0.003059163223952055, + 0.0021139178425073624, + 0.0022421744652092457, + 0.03427901864051819, + 0.02561507746577263, + 0.017058705911040306, + 0.015569117851555347, + 0.042498163878917694, + 0.03131670132279396, + 0.04043620079755783, + 0.02182789146900177, + 0.002281961729750037, + 0.0014184813480824232, + 0.001403065980412066, + 0.0014827349223196507, + 0.0023394592572003603, + 0.0023052154574543238, + 0.002178385853767395, + 0.001380224246531725, + 0.002340777311474085, + 0.0017775677843019366, + 0.0036689876578748226, + 0.0016118427738547325, + 0.005044161342084408, + 0.004840615671128035, + 0.002254453720524907, + 0.00256908405572176, + 0.0022117283660918474, + 0.0013665652368217707, + 0.0013564897235482931, + 0.0013929876731708646, + 0.004937771707773209, + 0.005441660992801189, + 0.0029884418472647667, + 0.002330835210159421, + 0.0011290605179965496, + 0.003012024099007249, + 0.0008834722102619708, + 0.0011020783567801118, + 0.0026749838143587112, + 0.0039620911702513695, + 0.0012680677464231849, + 0.0025086072273552418, + 0.002303806599229574, + 0.002181567484512925, + 0.0033690237905830145, + 0.002321474486961961, + 0.0025587966665625572, + 0.003744286485016346, + 0.005126220174133778, + 0.0018571007531136274, + 0.03711654990911484, + 0.019450906664133072, + 0.016627373173832893, + 0.04243861138820648, + 0.019852392375469208, + 0.11145588755607605, + 0.1953124701976776, + 0.03018382005393505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.006746964529156685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210249423980713 + ], + "opposite_logprob_tokens": [ + -1.5021024942398071 + ], + "logprob_reward": -0.25210249423980713, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210249423980713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07133238762617111, + 0.051676854491233826, + 0.12024154514074326, + 0.07680000364780426, + 0.012883255258202553, + 0.018257562071084976, + 0.008365575224161148, + 0.011534282006323338, + 0.013156902976334095, + 0.6157516241073608 + ], + "credit_map": [ + 0.011861949227750301, + 0.015123284421861172, + 0.00862735416740179, + 0.009208646602928638, + 0.006246956530958414, + 0.010543959215283394, + 0.005162514746189117, + 0.004557727370411158, + 0.004504359792917967, + 0.007362234406173229, + 0.007193566299974918, + 0.006201677490025759, + 0.007483540568500757, + 0.0072044194675982, + 0.007648651022464037, + 0.004078416619449854, + 0.017289048060774803, + 0.01574011892080307, + 0.017595522105693817, + 0.013487173244357109, + 0.017269274219870567, + 0.014849001541733742, + 0.013092595152556896, + 0.010918828658759594, + 0.010538740083575249, + 0.009885505773127079, + 0.007639160379767418, + 0.005270102061331272, + 0.012730497866868973, + 0.010097980499267578, + 0.012499338015913963, + 0.008138688281178474, + 0.0015055832918733358, + 0.0008906345465220511, + 0.0010015321895480156, + 0.001458651153370738, + 0.0026112727355211973, + 0.0018537852447479963, + 0.0019497578032314777, + 0.0016120390500873327, + 0.002781625371426344, + 0.0014256986323744059, + 0.003824022365733981, + 0.0017768513644114137, + 0.0023491487372666597, + 0.002903421875089407, + 0.0013432441046461463, + 0.0018535519484430552, + 0.0008513598004356027, + 0.0007027366082184017, + 0.000384922866942361, + 0.0005068142199888825, + 0.0023313802666962147, + 0.0010174006456509233, + 0.0008904156857170165, + 0.001680545974522829, + 0.0015681312652304769, + 0.0015167015371844172, + 0.0012191329151391983, + 0.0011277501471340656, + 0.0013507052790373564, + 0.0015140934847295284, + 0.0015881191939115524, + 0.001649649115279317, + 0.0019435288850218058, + 0.0015413154615089297, + 0.0014497662195935845, + 0.0012396126985549927, + 0.001245267572812736, + 0.0016266762977465987, + 0.002403676277026534, + 0.0017070607282221317, + 0.050858862698078156, + 0.01892651990056038, + 0.023323196917772293, + 0.05610888823866844, + 0.024222804233431816, + 0.14379696547985077, + 0.25946739315986633, + 0.039047036319971085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.25210249423980713 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002915441989899 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.07002915441989899, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002915441989899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05597049370408058, + 0.12037868797779083, + 0.014837904833257198, + 0.02333296276628971, + 0.03327728807926178, + 0.02130286954343319, + 0.024912256747484207, + 0.07322274148464203, + 0.05237842723727226, + 0.5803864002227783 + ], + "credit_map": [ + 0.007059201598167419, + 0.010022114962339401, + 0.008510323241353035, + 0.008716081269085407, + 0.005435264203697443, + 0.004722751211374998, + 0.007291417568922043, + 0.004213330801576376, + 0.01318419724702835, + 0.014331440441310406, + 0.018880149349570274, + 0.01597418636083603, + 0.01390397734940052, + 0.01638772152364254, + 0.017702559009194374, + 0.010014435276389122, + 0.0019281944260001183, + 0.002644562628120184, + 0.0018416410312056541, + 0.0018327382858842611, + 0.0024662327487021685, + 0.0016612065955996513, + 0.0012928888900205493, + 0.001170439994893968, + 0.0030067337211221457, + 0.0030291348230093718, + 0.0016710873460397124, + 0.0018093630205839872, + 0.004067989531904459, + 0.002796982182189822, + 0.0044472855515778065, + 0.0025043839123100042, + 0.003367679426446557, + 0.0023469615262001753, + 0.0026831950526684523, + 0.003499479265883565, + 0.004681667312979698, + 0.005018333438783884, + 0.006263162009418011, + 0.00541680259630084, + 0.00254830508492887, + 0.0013327441411092877, + 0.004484936129301786, + 0.0015194640727713704, + 0.004471414256840944, + 0.0034579457715153694, + 0.0013931103749200702, + 0.002094947500154376, + 0.0024482477456331253, + 0.0019613055046647787, + 0.0014094681246206164, + 0.0018385559087619185, + 0.0039916737005114555, + 0.005847956519573927, + 0.003048770595341921, + 0.004366277251392603, + 0.004797835368663073, + 0.008699416182935238, + 0.004299727734178305, + 0.003223301377147436, + 0.008746339939534664, + 0.02221258543431759, + 0.006033534649759531, + 0.015209990553557873, + 0.004714682698249817, + 0.007538153324276209, + 0.009247209876775742, + 0.005495131481438875, + 0.006665213964879513, + 0.007316626608371735, + 0.005913928616791964, + 0.005487474147230387, + 0.04803688824176788, + 0.025638120248913765, + 0.02182893455028534, + 0.051965825259685516, + 0.029146354645490646, + 0.12995344400405884, + 0.2358783781528473, + 0.037938397377729416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.07002915441989899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a clear tropical environment featuring staghorn corals, brain corals, and soft sea fans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small damselfish visible in the foreground center of the frame with scales shimmering in electric blue and lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the damselfish lunge forward to swallow a tiny plankton with lightning speed before resuming its hovering position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing, flickering rays of light visible in the water caused by direct overhead sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily to the right following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207653641700745 + ], + "opposite_logprob_tokens": [ + -1.502076506614685 + ], + "logprob_reward": -0.25207653641700745, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207653641700745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05471137538552284, + 0.05432650074362755, + 0.026066338643431664, + 0.16273127496242523, + 0.050425540655851364, + 0.049203187227249146, + 0.10167475044727325, + 0.025407535955309868, + 0.07810869812965393, + 0.39734479784965515 + ], + "credit_map": [ + 0.011318967677652836, + 0.01021150778979063, + 0.006213260814547539, + 0.007222624961286783, + 0.005087785888463259, + 0.006101554725319147, + 0.005081256851553917, + 0.0034744145814329386, + 0.011399421840906143, + 0.006301840767264366, + 0.006668569054454565, + 0.005976962856948376, + 0.0049494169652462006, + 0.009615838527679443, + 0.004992891568690538, + 0.0044215586967766285, + 0.002968972083181143, + 0.004295438062399626, + 0.0031764835584908724, + 0.0025665832217782736, + 0.00433248421177268, + 0.0028626129496842623, + 0.0029081604443490505, + 0.002955603413283825, + 0.0237595047801733, + 0.018323557451367378, + 0.014709011651575565, + 0.011412642896175385, + 0.027850188314914703, + 0.021763542667031288, + 0.026553954929113388, + 0.018358876928687096, + 0.007477675098925829, + 0.003574610920622945, + 0.004736152943223715, + 0.005616991315037012, + 0.006950917188078165, + 0.0063708471134305, + 0.00862149614840746, + 0.007076853420585394, + 0.007136583793908358, + 0.005742542911320925, + 0.007456519175320864, + 0.005350553430616856, + 0.005028704646974802, + 0.006858460139483213, + 0.005478262435644865, + 0.006151556968688965, + 0.0060999453999102116, + 0.006359306164085865, + 0.004082067403942347, + 0.004127750638872385, + 0.023503005504608154, + 0.032366883009672165, + 0.013169636018574238, + 0.011966154910624027, + 0.0028596362099051476, + 0.0023321767803281546, + 0.0026341951452195644, + 0.0019127518171444535, + 0.0030875937081873417, + 0.004791300743818283, + 0.001990601187571883, + 0.005799279548227787, + 0.006101562175899744, + 0.008407726883888245, + 0.00807665754109621, + 0.008046464994549751, + 0.007659580558538437, + 0.010895863175392151, + 0.019338039681315422, + 0.009582800790667534, + 0.03180713579058647, + 0.013904321938753128, + 0.014632878825068474, + 0.0364161916077137, + 0.01713886670768261, + 0.0925365537405014, + 0.16581785678863525, + 0.025091012939810753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.25207653641700745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_0/index_tensor([1325, 1325])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_0/index_tensor([1325, 1325])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f4869b5e975d824b6dba0c215a5895dd8dae6d76 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_0/index_tensor([1325, 1325])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle shot, tracking and rapidly panning sideways, captures the full force of a heavy main battle tank, its olive-green armor streaked with dried mud and fresh, wet dirt. It lies on a vast, muddy field, beneath a dark brown, sticky mud riddled with waterlogged ruts, reflecting the gloomy gray sky. With a dull roar, the steel behemoth swerves sharply, its massive hull tilting violently. One massive metal track sinks deeply into the saturated soil, practically stagnating as a fulcrum, while the other track spins at breakneck speed. This movement instantly transforms the track into a brutal excavator, hurling huge chunks of wet mud and water into the air in chaotic arcs, forming a fleeting, dark curtain of mud. The flying particles of earth are clearly visible in the diffuse, lifeless sunlight, imbuing the entire image with a raw, mechanical aesthetic of violence, creating an atmosphere of tension, unstoppable force, and immense power.", + "seed": 9876, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_0/index_1325_seed_9876_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13050601342692972, + "vlm_logprob": -0.13050601342692972, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.13050601342692972, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10262537002563477 + ], + "opposite_logprob_tokens": [ + -3.1026253700256348 + ], + "logprob_reward": -0.10262537002563477, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.10262537002563477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08884933590888977, + 0.041012875735759735, + 0.015110486187040806, + 0.3578498661518097, + 0.1096896156668663, + 0.03834579885005951, + 0.023334980010986328, + 0.28699612617492676, + 0.017711160704493523, + 0.021099768579006195 + ], + "credit_map": [ + 0.008332216180860996, + 0.0102988351136446, + 0.009945843368768692, + 0.012205302715301514, + 0.010053537786006927, + 0.015117377042770386, + 0.016308458521962166, + 0.0065877679735422134, + 0.004796668421477079, + 0.007501750253140926, + 0.0030140592716634274, + 0.003602302400395274, + 0.005610862746834755, + 0.004020671360194683, + 0.005482887849211693, + 0.006983674596995115, + 0.0015221075154840946, + 0.002063389867544174, + 0.0014650895027443767, + 0.001588050159625709, + 0.0024077941197901964, + 0.0022197195794433355, + 0.002179775619879365, + 0.0016645599389448762, + 0.06420808285474777, + 0.01966121606528759, + 0.01332686748355627, + 0.018396027386188507, + 0.14186224341392517, + 0.029623402282595634, + 0.013740137219429016, + 0.05703188106417656, + 0.03342420607805252, + 0.008645215071737766, + 0.03978033736348152, + 0.006992923561483622, + 0.0071213506162166595, + 0.004975629970431328, + 0.004165262449532747, + 0.004584696143865585, + 0.004379607737064362, + 0.004693890456110239, + 0.007306966464966536, + 0.005320535507053137, + 0.00667392136529088, + 0.005045072175562382, + 0.0024857758544385433, + 0.0024400288239121437, + 0.0032662982121109962, + 0.0024322194512933493, + 0.007696007378399372, + 0.0021765155252069235, + 0.002612913027405739, + 0.0028365771286189556, + 0.001223160419613123, + 0.0010912880534306169, + 0.03257054090499878, + 0.02870929427444935, + 0.1204243078827858, + 0.028454141691327095, + 0.01775028370320797, + 0.030355624854564667, + 0.013327503576874733, + 0.015404434874653816, + 0.003120273584499955, + 0.0018427143804728985, + 0.0019431214313954115, + 0.0030079311691224575, + 0.001777303870767355, + 0.0014282154152169824, + 0.002428425708785653, + 0.0021631759591400623, + 0.0018737572245299816, + 0.00417231535539031, + 0.0018951039528474212, + 0.0022901135962456465, + 0.0026430082507431507, + 0.002314530313014984, + 0.002951287431642413, + 0.0029596518725156784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.10262537002563477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016110599040985107 + ], + "opposite_logprob_tokens": [ + -4.141110420227051 + ], + "logprob_reward": -0.016110599040985107, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016110599040985107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4016335904598236, + 0.10625810921192169, + 0.0334034226834774, + 0.10766441375017166, + 0.1994597166776657, + 0.03737269714474678, + 0.04086681827902794, + 0.03251628205180168, + 0.02087162807583809, + 0.019953351467847824 + ], + "credit_map": [ + 0.023847855627536774, + 0.020636960864067078, + 0.018577557057142258, + 0.04498990625143051, + 0.04010836035013199, + 0.10568570345640182, + 0.12087515741586685, + 0.02691212110221386, + 0.013108343817293644, + 0.011682858690619469, + 0.006753555964678526, + 0.008323602378368378, + 0.011812284588813782, + 0.01619604043662548, + 0.017207341268658638, + 0.021174082532525063, + 0.0021846892777830362, + 0.0035422833170741796, + 0.003288698848336935, + 0.00458632642403245, + 0.005791941192001104, + 0.0059206727892160416, + 0.004357526078820229, + 0.0037312868516892195, + 0.024763714522123337, + 0.008769367821514606, + 0.0075008682906627655, + 0.007517208345234394, + 0.029246211051940918, + 0.010295923799276352, + 0.008936614729464054, + 0.010634498670697212, + 0.05591684579849243, + 0.01500125415623188, + 0.07236140221357346, + 0.013677152805030346, + 0.018627755343914032, + 0.008156148716807365, + 0.00781280267983675, + 0.007906338199973106, + 0.004285061731934547, + 0.0051405406557023525, + 0.006868417840451002, + 0.00446948641911149, + 0.004073598422110081, + 0.0040584406815469265, + 0.005212103947997093, + 0.0032650481443852186, + 0.003780442988499999, + 0.004934098571538925, + 0.004156908951699734, + 0.0035412688739597797, + 0.008792825974524021, + 0.008649391122162342, + 0.0032128721941262484, + 0.003799008671194315, + 0.0027809571474790573, + 0.0034578151535242796, + 0.009408526122570038, + 0.003769068280234933, + 0.00288513395935297, + 0.004434989299625158, + 0.002554817358031869, + 0.003224974265322089, + 0.0027345933485776186, + 0.002832086756825447, + 0.0028890126850456, + 0.0035516105126589537, + 0.001858157804235816, + 0.0016042239731177688, + 0.0029630970675498247, + 0.0024388465099036694, + 0.0021146407816559076, + 0.0030932303052395582, + 0.003847991582006216, + 0.0033291869331151247, + 0.001797547098249197, + 0.0017522487323731184, + 0.002054499462246895, + 0.001964007969945669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.016110599040985107 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486400842666626 + ], + "opposite_logprob_tokens": [ + -1.2236400842666626 + ], + "logprob_reward": -0.3486400842666626, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486400842666626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044583760201931, + 0.06863804161548615, + 0.012987161055207253, + 0.5805980563163757, + 0.120266392827034, + 0.03386523202061653, + 0.03550611436367035, + 0.06583671271800995, + 0.028319407254457474, + 0.00939920824021101 + ], + "credit_map": [ + 0.0034939139150083065, + 0.004757001530379057, + 0.008227922022342682, + 0.008279907517135143, + 0.004780429415404797, + 0.0038911844603717327, + 0.006439578719437122, + 0.004713814239948988, + 0.005873623304069042, + 0.007603617385029793, + 0.004031337331980467, + 0.004403078928589821, + 0.011975914239883423, + 0.015718093141913414, + 0.008714783005416393, + 0.010317591018974781, + 0.0012110959505662322, + 0.0017530216136947274, + 0.001225483836606145, + 0.0014465468702837825, + 0.002358189783990383, + 0.0021092304959893227, + 0.0015466990880668163, + 0.0013368910877034068, + 0.11787626892328262, + 0.03404003754258156, + 0.017984913662075996, + 0.02987297624349594, + 0.231650248169899, + 0.04823257029056549, + 0.023673512041568756, + 0.07726745307445526, + 0.03578794002532959, + 0.008373971097171307, + 0.04568390175700188, + 0.008209938183426857, + 0.008933409117162228, + 0.004790538921952248, + 0.0039024802390486, + 0.004584200214594603, + 0.004034515004605055, + 0.0044520203955471516, + 0.00577120203524828, + 0.004687247332185507, + 0.0038408739492297173, + 0.003187838476151228, + 0.0047779446467757225, + 0.003113584825769067, + 0.004669732879847288, + 0.0035957330837845802, + 0.008418841287493706, + 0.004385448060929775, + 0.004999037832021713, + 0.004993785638362169, + 0.0021168615203350782, + 0.002326671266928315, + 0.007608842104673386, + 0.005499665159732103, + 0.03166630119085312, + 0.006357218604534864, + 0.0029455553740262985, + 0.005391150712966919, + 0.003108220873400569, + 0.003259755438193679, + 0.0028530547861009836, + 0.0025442931801080704, + 0.005913713481277227, + 0.006044904701411724, + 0.002334302756935358, + 0.0025172214955091476, + 0.003830638714134693, + 0.0022812746465206146, + 0.0011153897503390908, + 0.0010188821470364928, + 0.0009576444863341749, + 0.0012739230878651142, + 0.000843735586386174, + 0.001024403260089457, + 0.0015316292410716414, + 0.001633599866181612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.3486400842666626 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255692172795534 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255692172795534, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255692172795534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05076140537858009, + 0.25054827332496643, + 0.04758940264582634, + 0.21489165723323822, + 0.04442033916711807, + 0.05648749694228172, + 0.1038021370768547, + 0.15808966755867004, + 0.06125061586499214, + 0.012159016914665699 + ], + "credit_map": [ + 0.006416916381567717, + 0.005095351953059435, + 0.008728484623134136, + 0.00625046668574214, + 0.004097358323633671, + 0.005473189055919647, + 0.008718358352780342, + 0.00598127581179142, + 0.030662454664707184, + 0.032756853848695755, + 0.017189309000968933, + 0.019613875076174736, + 0.03464360162615776, + 0.04049498587846756, + 0.032509695738554, + 0.04267748445272446, + 0.005812860559672117, + 0.007216648198664188, + 0.004713237751275301, + 0.004841047804802656, + 0.008891288191080093, + 0.005904395133256912, + 0.004829138517379761, + 0.005380785092711449, + 0.0455876849591732, + 0.013040915131568909, + 0.010963167995214462, + 0.01233084499835968, + 0.08018600940704346, + 0.01829979009926319, + 0.009969199076294899, + 0.024514025077223778, + 0.005295832175761461, + 0.0055586048401892185, + 0.008152194321155548, + 0.005529622547328472, + 0.0085744084790349, + 0.0035463464446365833, + 0.00451451912522316, + 0.003248808439821005, + 0.006717320065945387, + 0.006944471504539251, + 0.008704021573066711, + 0.006850941572338343, + 0.007415220607072115, + 0.006005290430039167, + 0.008423201739788055, + 0.005427026655524969, + 0.011094728484749794, + 0.012091516517102718, + 0.01634920760989189, + 0.011019708588719368, + 0.01875336468219757, + 0.018861640244722366, + 0.0073255570605397224, + 0.00830639898777008, + 0.018558643758296967, + 0.016191253438591957, + 0.07332275807857513, + 0.014784667640924454, + 0.007951401174068451, + 0.012406818568706512, + 0.006682521663606167, + 0.008191583678126335, + 0.00562310079112649, + 0.006273210979998112, + 0.009862218052148819, + 0.011752685531973839, + 0.005918701179325581, + 0.0062162200920283794, + 0.009459728375077248, + 0.006144746206700802, + 0.0012912672245875, + 0.0014722279738634825, + 0.002155192196369171, + 0.0017232882091775537, + 0.001103713526390493, + 0.0013656930532306433, + 0.0014434403274208307, + 0.0016041938215494156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.005255692172795534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989832162857056 + ], + "opposite_logprob_tokens": [ + -1.8048982620239258 + ], + "logprob_reward": -0.17989832162857056, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989832162857056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29312142729759216, + 0.05549969896674156, + 0.011164647527039051, + 0.40202364325523376, + 0.10553023219108582, + 0.03234999254345894, + 0.03036496601998806, + 0.04491252452135086, + 0.01763877645134926, + 0.00739410100504756 + ], + "credit_map": [ + 0.01529638934880495, + 0.010913395322859287, + 0.010923586785793304, + 0.034299179911613464, + 0.029365452006459236, + 0.08141792565584183, + 0.09137053787708282, + 0.019534971565008163, + 0.007266812026500702, + 0.00867763813585043, + 0.004352446645498276, + 0.005254702642560005, + 0.007619473617523909, + 0.0051852744072675705, + 0.007228356786072254, + 0.009914991445839405, + 0.0013517130864784122, + 0.0020055891945958138, + 0.0009985892102122307, + 0.0010118702193722129, + 0.0020702811889350414, + 0.0012295965570956469, + 0.0010853138519451022, + 0.001411694218404591, + 0.07648303359746933, + 0.022776827216148376, + 0.012801655568182468, + 0.01908014342188835, + 0.1627662628889084, + 0.033078864216804504, + 0.015762921422719955, + 0.059273939579725266, + 0.03324796259403229, + 0.007939722388982773, + 0.04047349840402603, + 0.006519699934870005, + 0.006814667955040932, + 0.004092693794518709, + 0.0028751154895871878, + 0.0035668748896569014, + 0.004857588559389114, + 0.004039124585688114, + 0.007058953400701284, + 0.003699531313031912, + 0.0040626307018101215, + 0.003348312806338072, + 0.0027605793438851833, + 0.0025232720654457808, + 0.0034224886912852526, + 0.0025985995307564735, + 0.005062374752014875, + 0.0029207628685981035, + 0.006069499533623457, + 0.005705223418772221, + 0.002135075628757477, + 0.002450939966365695, + 0.005046837497502565, + 0.0037044875789433718, + 0.02042466588318348, + 0.004649011883884668, + 0.002818369772285223, + 0.003941111266613007, + 0.0023782551288604736, + 0.0019497855100780725, + 0.0018107573268935084, + 0.0015726085985079408, + 0.0017583428416401148, + 0.002699505304917693, + 0.0021523842588067055, + 0.0029179081320762634, + 0.002651070011779666, + 0.0020761992782354355, + 0.0007109933067113161, + 0.0007237285026349127, + 0.0008103371364995837, + 0.0013445778749883175, + 0.000744938850402832, + 0.0008883194532245398, + 0.0010279130656272173, + 0.0011432928731665015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.17989832162857056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot, tracking and rapidly panning sideways, captures the full force of a heavy main battle tank, its olive-green armor streaked with dried mud and fresh, wet dirt. It lies on a vast, muddy field, beneath a dark brown, sticky mud riddled with waterlogged ruts, reflecting the gloomy gray sky. With a dull roar, the steel behemoth swerves sharply, its massive hull tilting violently. One massive metal track sinks deeply into the saturated soil, practically stagnating as a fulcrum, while the other track spins at breakneck speed. This movement instantly transforms the track into a brutal excavator, hurling huge chunks of wet mud and water into the air in chaotic arcs, forming a fleeting, dark curtain of mud. The flying particles of earth are clearly visible in the diffuse, lifeless sunlight, imbuing the entire image with a raw, mechanical aesthetic of violence, creating an atmosphere of tension, unstoppable force, and immense power.", + "seed": 9877, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_0/index_1325_seed_9877_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15897784898988904, + "vlm_logprob": -0.15897784898988904, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.15897784898988904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39865899085998535 + ], + "opposite_logprob_tokens": [ + -1.3986589908599854 + ], + "logprob_reward": -0.39865899085998535, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.39865899085998535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12281408905982971, + 0.05086500570178032, + 0.07345207780599594, + 0.05724804103374481, + 0.08710930496454239, + 0.15514670312404633, + 0.07619890570640564, + 0.02751493640244007, + 0.2057746946811676, + 0.14387628436088562 + ], + "credit_map": [ + 0.015364598482847214, + 0.014507059007883072, + 0.017367269843816757, + 0.01593136228621006, + 0.013548505492508411, + 0.014639845117926598, + 0.02342088706791401, + 0.008034558035433292, + 0.007734025362879038, + 0.0063796802423894405, + 0.0060580139979720116, + 0.006512730848044157, + 0.0051575773395597935, + 0.005177736282348633, + 0.007295537739992142, + 0.006549696903675795, + 0.009558593854308128, + 0.008190743625164032, + 0.005532942712306976, + 0.009139087982475758, + 0.012485094368457794, + 0.010346966795623302, + 0.009461541660130024, + 0.008737098425626755, + 0.0071687581948935986, + 0.011706041172146797, + 0.006401225924491882, + 0.007443862967193127, + 0.011475873179733753, + 0.005264013074338436, + 0.004614199046045542, + 0.0031740625854581594, + 0.0077990018762648106, + 0.018398750573396683, + 0.009472026489675045, + 0.012603189796209335, + 0.018461080268025398, + 0.009634703397750854, + 0.00636441120877862, + 0.0043761273846030235, + 0.01731792651116848, + 0.007592024281620979, + 0.06203174218535423, + 0.04261322692036629, + 0.007435697130858898, + 0.0048818993382155895, + 0.006533113308250904, + 0.006741060875356197, + 0.014397277496755123, + 0.00886803213506937, + 0.010450857691466808, + 0.008766653016209602, + 0.009303922764956951, + 0.010299213230609894, + 0.005130220204591751, + 0.00898272916674614, + 0.004130805842578411, + 0.0026793736033141613, + 0.004951599985361099, + 0.003440034342929721, + 0.003381410613656044, + 0.0028108772821724415, + 0.002540559507906437, + 0.003580273361876607, + 0.017438869923353195, + 0.01665329374372959, + 0.028851067647337914, + 0.01635395735502243, + 0.020957034081220627, + 0.047961704432964325, + 0.027899986132979393, + 0.02965875156223774, + 0.004334231372922659, + 0.005624322220683098, + 0.009698322974145412, + 0.008989713154733181, + 0.0071049886755645275, + 0.006096577271819115, + 0.04916273429989815, + 0.052865367382764816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.39865899085998535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011120719835162163 + ], + "opposite_logprob_tokens": [ + -4.511120796203613 + ], + "logprob_reward": -0.011120719835162163, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011120719835162163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10777211934328079, + 0.03650166466832161, + 0.059116873890161514, + 0.07409162819385529, + 0.11306318640708923, + 0.05025959014892578, + 0.03855285421013832, + 0.07773978263139725, + 0.35345590114593506, + 0.08944645524024963 + ], + "credit_map": [ + 0.011113760061562061, + 0.011822887696325779, + 0.015733400359749794, + 0.012432007119059563, + 0.013401338830590248, + 0.01072488259524107, + 0.025439327582716942, + 0.0071045164950191975, + 0.004951132461428642, + 0.003955972380936146, + 0.005338452756404877, + 0.008054404519498348, + 0.0028134698513895273, + 0.005214544013142586, + 0.003796659642830491, + 0.0023770274128764868, + 0.007978111505508423, + 0.008849959820508957, + 0.006260300986468792, + 0.008311551064252853, + 0.009303758852183819, + 0.006278537213802338, + 0.007690293248742819, + 0.004444360267370939, + 0.009665477089583874, + 0.012949501164257526, + 0.006694845389574766, + 0.010211971588432789, + 0.014083776623010635, + 0.007409658282995224, + 0.007691432721912861, + 0.005384963471442461, + 0.010548950172960758, + 0.025247380137443542, + 0.012627668678760529, + 0.014804569073021412, + 0.021738843992352486, + 0.013225702568888664, + 0.007940244860947132, + 0.006929825060069561, + 0.005538141820579767, + 0.00477007869631052, + 0.015549885109066963, + 0.011061484925448895, + 0.005220634397119284, + 0.0021537705324590206, + 0.003239184617996216, + 0.002726413309574127, + 0.005026852712035179, + 0.005012715235352516, + 0.004984704777598381, + 0.004432285204529762, + 0.006460490636527538, + 0.004903161898255348, + 0.0039328308776021, + 0.0037998126354068518, + 0.013822765089571476, + 0.005970230791717768, + 0.011651509441435337, + 0.0098109170794487, + 0.013916036114096642, + 0.007259246427565813, + 0.0056512802839279175, + 0.009657793678343296, + 0.018814651295542717, + 0.016097547486424446, + 0.018970800563693047, + 0.02010350488126278, + 0.0291825570166111, + 0.127672478556633, + 0.03867267817258835, + 0.08394167572259903, + 0.004846858326345682, + 0.006115743890404701, + 0.01242287177592516, + 0.00946921668946743, + 0.007186209782958031, + 0.006334749050438404, + 0.019585030153393745, + 0.02348577044904232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.011120719835162163 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991484701633453 + ], + "opposite_logprob_tokens": [ + -1.8049148321151733 + ], + "logprob_reward": -0.17991484701633453, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991484701633453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05832548066973686, + 0.033941835165023804, + 0.05075918138027191, + 0.07115868479013443, + 0.21004760265350342, + 0.2626815140247345, + 0.0410105399787426, + 0.029078422114253044, + 0.15882247686386108, + 0.0841742679476738 + ], + "credit_map": [ + 0.009167848154902458, + 0.006208691745996475, + 0.007516908459365368, + 0.008029048331081867, + 0.00897679291665554, + 0.006772753316909075, + 0.00696221087127924, + 0.0046912250109016895, + 0.004595452919602394, + 0.0035196172539144754, + 0.00443742610514164, + 0.005118310451507568, + 0.003704276867210865, + 0.005301380529999733, + 0.003534977324306965, + 0.003730391850695014, + 0.0072989435866475105, + 0.006997609045356512, + 0.004753596615046263, + 0.005211417563259602, + 0.00801064446568489, + 0.007107277400791645, + 0.0061528319492936134, + 0.005226853769272566, + 0.007405863143503666, + 0.008257267996668816, + 0.00952859316021204, + 0.010657736100256443, + 0.01279421802610159, + 0.007824622094631195, + 0.008596841245889664, + 0.0060935430228710175, + 0.02423859015107155, + 0.020946098491549492, + 0.05232411250472069, + 0.036823105067014694, + 0.022705186158418655, + 0.020602986216545105, + 0.01761866733431816, + 0.014788839966058731, + 0.03042927011847496, + 0.013933266513049603, + 0.10691627860069275, + 0.07392091304063797, + 0.009815458208322525, + 0.007295283954590559, + 0.009931759908795357, + 0.010439254343509674, + 0.009416462853550911, + 0.004542574752122164, + 0.005975027568638325, + 0.004093032330274582, + 0.004374653100967407, + 0.004397635813802481, + 0.003748958930373192, + 0.004462189506739378, + 0.002969088265672326, + 0.002324553206562996, + 0.006084691733121872, + 0.004202691372483969, + 0.0032469260040670633, + 0.003288209205493331, + 0.002765131648629904, + 0.004197127651423216, + 0.015080788172781467, + 0.010053069330751896, + 0.019489841535687447, + 0.010062595829367638, + 0.013413702137768269, + 0.0369793064892292, + 0.018627066165208817, + 0.035116106271743774, + 0.003546618390828371, + 0.003997019492089748, + 0.005677003879100084, + 0.00540523836389184, + 0.004565102979540825, + 0.004711995832622051, + 0.026750905439257622, + 0.029520375654101372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.17991484701633453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617530455812812 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.003617530455812812, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617530455812812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1326581835746765, + 0.07309973239898682, + 0.0973707064986229, + 0.04266003146767616, + 0.17268358170986176, + 0.10769785195589066, + 0.04941819608211517, + 0.09057893604040146, + 0.1891007423400879, + 0.04473210126161575 + ], + "credit_map": [ + 0.01251950953155756, + 0.014705203473567963, + 0.017176244407892227, + 0.014440534636378288, + 0.015124465338885784, + 0.018623165786266327, + 0.030609961599111557, + 0.00945910532027483, + 0.008103186264634132, + 0.007728300988674164, + 0.008041507564485073, + 0.013868995010852814, + 0.007459723390638828, + 0.014263112097978592, + 0.007187649141997099, + 0.006447259336709976, + 0.012636076658964157, + 0.012935548089444637, + 0.01296988409012556, + 0.014396383427083492, + 0.010005885735154152, + 0.009109139442443848, + 0.01555582694709301, + 0.009761962108314037, + 0.004645919892936945, + 0.005118739791214466, + 0.005236176308244467, + 0.005718550179153681, + 0.00806506723165512, + 0.004657049663364887, + 0.005551988258957863, + 0.003666541539132595, + 0.019384166225790977, + 0.017758920788764954, + 0.045970771461725235, + 0.02590116113424301, + 0.020017731934785843, + 0.018034471198916435, + 0.014760072343051434, + 0.010856295935809612, + 0.011703998781740665, + 0.011652643792331219, + 0.029132110998034477, + 0.023590723052620888, + 0.010112300515174866, + 0.007305770646780729, + 0.007384413853287697, + 0.006815888918936253, + 0.007220751140266657, + 0.005595025140792131, + 0.0073737939819693565, + 0.004826809745281935, + 0.0056135207414627075, + 0.006966445595026016, + 0.0038801582995802164, + 0.00794169120490551, + 0.015259451232850552, + 0.00905677955597639, + 0.01485509518533945, + 0.011283713392913342, + 0.01235046423971653, + 0.008319571614265442, + 0.007898119278252125, + 0.011555740609765053, + 0.013471818529069424, + 0.010381440632045269, + 0.018376003950834274, + 0.01048985030502081, + 0.012595439329743385, + 0.059210602194070816, + 0.015773838385939598, + 0.04880175366997719, + 0.0017581551801413298, + 0.0030510530341416597, + 0.0021846110466867685, + 0.0037661322858184576, + 0.003557848045602441, + 0.0029988514725118876, + 0.012894589453935623, + 0.014520860277116299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.003617530455812812 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157715678215027 + ], + "opposite_logprob_tokens": [ + -1.7015771865844727 + ], + "logprob_reward": -0.20157715678215027, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157715678215027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06986463814973831, + 0.042922161519527435, + 0.04181230068206787, + 0.06522832065820694, + 0.09300241619348526, + 0.16819241642951965, + 0.06466938555240631, + 0.06578931957483292, + 0.3448788523674011, + 0.04364024102687836 + ], + "credit_map": [ + 0.008053405210375786, + 0.008209518156945705, + 0.008439240977168083, + 0.009884676896035671, + 0.008576702326536179, + 0.008812200278043747, + 0.011735699139535427, + 0.006153191439807415, + 0.006969957146793604, + 0.0051788343116641045, + 0.00528260413557291, + 0.005309202708303928, + 0.003879050724208355, + 0.005546418484300375, + 0.005395327229052782, + 0.005360766779631376, + 0.005001547280699015, + 0.005313368514180183, + 0.0025911161210387945, + 0.004299746360629797, + 0.009428064338862896, + 0.00599611084908247, + 0.004425796680152416, + 0.0047565484419465065, + 0.009207640774548054, + 0.014467178843915462, + 0.00818732101470232, + 0.007723751477897167, + 0.010879532434046268, + 0.005882736761122942, + 0.004334381781518459, + 0.004545767791569233, + 0.00710033904761076, + 0.014251708984375, + 0.017381001263856888, + 0.017761964350938797, + 0.011582023464143276, + 0.01076765637844801, + 0.00791750568896532, + 0.006240206770598888, + 0.018953092396259308, + 0.012785691767930984, + 0.06596481055021286, + 0.04417097941040993, + 0.007046001497656107, + 0.006493002641946077, + 0.006661222316324711, + 0.006117584649473429, + 0.01173450518399477, + 0.004879220388829708, + 0.010572913102805614, + 0.007821368984878063, + 0.006874911021441221, + 0.008228618651628494, + 0.004309204872697592, + 0.010248640552163124, + 0.010305988602340221, + 0.010490949265658855, + 0.00872548297047615, + 0.007976379245519638, + 0.007639724761247635, + 0.007450893055647612, + 0.006760019809007645, + 0.006439871154725552, + 0.04433058202266693, + 0.03607947751879692, + 0.05258877947926521, + 0.042984042316675186, + 0.04559476301074028, + 0.04001360759139061, + 0.04843614995479584, + 0.034851450473070145, + 0.0021479830611497164, + 0.002899145008996129, + 0.0028370823711156845, + 0.0024720507208257914, + 0.00410045450553298, + 0.003467447590082884, + 0.011868992820382118, + 0.01384708285331726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.20157715678215027 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot, tracking and rapidly panning sideways, captures the full force of a heavy main battle tank, its olive-green armor streaked with dried mud and fresh, wet dirt. It lies on a vast, muddy field, beneath a dark brown, sticky mud riddled with waterlogged ruts, reflecting the gloomy gray sky. With a dull roar, the steel behemoth swerves sharply, its massive hull tilting violently. One massive metal track sinks deeply into the saturated soil, practically stagnating as a fulcrum, while the other track spins at breakneck speed. This movement instantly transforms the track into a brutal excavator, hurling huge chunks of wet mud and water into the air in chaotic arcs, forming a fleeting, dark curtain of mud. The flying particles of earth are clearly visible in the diffuse, lifeless sunlight, imbuing the entire image with a raw, mechanical aesthetic of violence, creating an atmosphere of tension, unstoppable force, and immense power.", + "seed": 9878, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_0/index_1325_seed_9878_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11549116745591163, + "vlm_logprob": -0.11549116745591163, + "vlm_yes_no_margin": 3.024999976158142, + "vlm_token_credit": -0.11549116745591163, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18763551115989685 + ], + "opposite_logprob_tokens": [ + -2.3126354217529297 + ], + "logprob_reward": -0.18763551115989685, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.18763551115989685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18480533361434937, + 0.06165351718664169, + 0.08633051812648773, + 0.2115786373615265, + 0.034347016364336014, + 0.07744759321212769, + 0.08336374908685684, + 0.08851760625839233, + 0.13882224261760712, + 0.03313373774290085 + ], + "credit_map": [ + 0.017749980092048645, + 0.015903055667877197, + 0.018681811168789864, + 0.022589631378650665, + 0.017750950530171394, + 0.020220277830958366, + 0.04136470705270767, + 0.030544942244887352, + 0.006212594453245401, + 0.005670605227351189, + 0.007215541787445545, + 0.009531560353934765, + 0.011947226710617542, + 0.0047246720641851425, + 0.008595844730734825, + 0.007755478844046593, + 0.009190703742206097, + 0.008617045357823372, + 0.01305620837956667, + 0.015463266521692276, + 0.007445089053362608, + 0.006152460351586342, + 0.015108208172023296, + 0.01129754539579153, + 0.018969841301441193, + 0.029169803485274315, + 0.02523965947329998, + 0.027951709926128387, + 0.020877180621027946, + 0.038616228848695755, + 0.036995235830545425, + 0.013758990913629532, + 0.005340916570276022, + 0.004501847084611654, + 0.004954570438712835, + 0.005903054494410753, + 0.0036108680069446564, + 0.002934162737801671, + 0.0039024674333631992, + 0.003199130529537797, + 0.007325988728553057, + 0.016262881457805634, + 0.010702674277126789, + 0.010646611452102661, + 0.009229744784533978, + 0.004847445525228977, + 0.008371282368898392, + 0.010060969740152359, + 0.008040696382522583, + 0.007888776250183582, + 0.018958959728479385, + 0.011455303058028221, + 0.011705063283443451, + 0.007610687520354986, + 0.007735173683613539, + 0.009969096630811691, + 0.00837163906544447, + 0.015583056025207043, + 0.013356264680624008, + 0.014700697734951973, + 0.010547720827162266, + 0.006001253612339497, + 0.010345050133764744, + 0.009611930698156357, + 0.015613595955073833, + 0.025964343920350075, + 0.015636615455150604, + 0.021268635988235474, + 0.015241444110870361, + 0.009948525577783585, + 0.020458374172449112, + 0.01469071488827467, + 0.005862305872142315, + 0.005199267528951168, + 0.0033450014889240265, + 0.0029360561165958643, + 0.0046321856789290905, + 0.0028512785211205482, + 0.004798160400241613, + 0.003509487258270383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.18763551115989685 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018229680135846138 + ], + "opposite_logprob_tokens": [ + -4.0182294845581055 + ], + "logprob_reward": -0.018229680135846138, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018229680135846138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0763191506266594, + 0.0456291101872921, + 0.04079782962799072, + 0.14219911396503448, + 0.07078900188207626, + 0.05214319005608559, + 0.09671741724014282, + 0.27102184295654297, + 0.03739975392818451, + 0.16698364913463593 + ], + "credit_map": [ + 0.009253541938960552, + 0.007438940927386284, + 0.014280982315540314, + 0.009038213640451431, + 0.01201443001627922, + 0.008426304906606674, + 0.008505218662321568, + 0.007361516356468201, + 0.005560300312936306, + 0.0038176733069121838, + 0.006708759348839521, + 0.008589240722358227, + 0.003977244719862938, + 0.003599863499403, + 0.007978654466569424, + 0.005397371482104063, + 0.006267137359827757, + 0.0029828583355993032, + 0.007368903141468763, + 0.007908620871603489, + 0.004282096400856972, + 0.0036899331025779247, + 0.004654477816075087, + 0.0036438005045056343, + 0.010260525159537792, + 0.011111673898994923, + 0.007467829156666994, + 0.0063657560385763645, + 0.04701749235391617, + 0.046225082129240036, + 0.009078417904675007, + 0.0046723331324756145, + 0.013966972939670086, + 0.007470143958926201, + 0.008021043613553047, + 0.00864140409976244, + 0.008207087405025959, + 0.007397972978651524, + 0.00921508390456438, + 0.007869290187954903, + 0.005167824681848288, + 0.01019404549151659, + 0.0081890057772398, + 0.008061119355261326, + 0.006301608402282, + 0.003310959320515394, + 0.006128144450485706, + 0.004790478851646185, + 0.011286327615380287, + 0.010850179009139538, + 0.020923012867569923, + 0.015431503765285015, + 0.01278071478009224, + 0.00768917053937912, + 0.009117319248616695, + 0.008639195002615452, + 0.021024154499173164, + 0.04093720018863678, + 0.048245690762996674, + 0.057709429413080215, + 0.030254509299993515, + 0.021598557010293007, + 0.03155480697751045, + 0.01969747245311737, + 0.0034854356199502945, + 0.005542499478906393, + 0.005122167989611626, + 0.006440217141062021, + 0.004250118974596262, + 0.0028847004286944866, + 0.0058126794174313545, + 0.0038619351107627153, + 0.020683102309703827, + 0.020895663648843765, + 0.02283155545592308, + 0.024335598573088646, + 0.010998955927789211, + 0.014277906157076359, + 0.025416698306798935, + 0.027544157579541206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.018229680135846138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282360672950745 + ], + "opposite_logprob_tokens": [ + -2.0178236961364746 + ], + "logprob_reward": -0.14282360672950745, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14282360672950745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08224037289619446, + 0.059778615832328796, + 0.055405180901288986, + 0.14951908588409424, + 0.07914017140865326, + 0.2462049424648285, + 0.15501563251018524, + 0.07558943331241608, + 0.04504772648215294, + 0.052058856934309006 + ], + "credit_map": [ + 0.011165597476065159, + 0.006317312363535166, + 0.02002030424773693, + 0.009516052901744843, + 0.0065871537663042545, + 0.005527333356440067, + 0.01315484382212162, + 0.009951773099601269, + 0.005104529671370983, + 0.004660245031118393, + 0.0062592364847660065, + 0.011416706256568432, + 0.010049809701740742, + 0.005107991863042116, + 0.00949096865952015, + 0.007689128629863262, + 0.0069340807385742664, + 0.008881177753210068, + 0.0070131318643689156, + 0.009320233017206192, + 0.005435310769826174, + 0.00479354802519083, + 0.007480188272893429, + 0.005547513719648123, + 0.012107315473258495, + 0.018110031262040138, + 0.01601114124059677, + 0.020609542727470398, + 0.0171282310038805, + 0.03244692459702492, + 0.02297532558441162, + 0.010130566544830799, + 0.014052338898181915, + 0.009314124472439289, + 0.011161032132804394, + 0.013026595115661621, + 0.007352141197770834, + 0.00650222972035408, + 0.009176298975944519, + 0.008555402979254723, + 0.019302692264318466, + 0.041248586028814316, + 0.04526501148939133, + 0.056390125304460526, + 0.027351832017302513, + 0.011945712380111217, + 0.02918274886906147, + 0.015518250875175, + 0.016561688855290413, + 0.018125683069229126, + 0.026538047939538956, + 0.028167815878987312, + 0.014724652282893658, + 0.013074815273284912, + 0.021997569128870964, + 0.015825355425477028, + 0.0066377874463796616, + 0.009344147518277168, + 0.013372772373259068, + 0.010689359158277512, + 0.009328192099928856, + 0.00781488511711359, + 0.01041181106120348, + 0.007990473881363869, + 0.004053839016705751, + 0.007244978565722704, + 0.005837408825755119, + 0.007131056394428015, + 0.004771008156239986, + 0.003159135580062866, + 0.008395190350711346, + 0.004455110058188438, + 0.0036414526402950287, + 0.007816397584974766, + 0.005433700047433376, + 0.009503696113824844, + 0.004869282711297274, + 0.005214795470237732, + 0.007335783448070288, + 0.008243747986853123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.14282360672950745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031949449330568314 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031949449330568314, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031949449330568314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07530827820301056, + 0.04866792634129524, + 0.022761758416891098, + 0.220838263630867, + 0.06314137578010559, + 0.17644459009170532, + 0.1875983029603958, + 0.07983174175024033, + 0.04782499000430107, + 0.07758276164531708 + ], + "credit_map": [ + 0.009305067360401154, + 0.006517598405480385, + 0.014569269493222237, + 0.008367271162569523, + 0.010589939542114735, + 0.005879455711692572, + 0.011259978637099266, + 0.008819691836833954, + 0.005047745071351528, + 0.004339094273746014, + 0.005482392851263285, + 0.007874357514083385, + 0.008858035318553448, + 0.0037062945775687695, + 0.007221763487905264, + 0.0061382390558719635, + 0.002782918978482485, + 0.0023351688869297504, + 0.003526080632582307, + 0.0032761781476438046, + 0.002782992320135236, + 0.0021642916835844517, + 0.0031422653701156378, + 0.002751860534772277, + 0.016744589433073997, + 0.008105059154331684, + 0.01136657502502203, + 0.015008077956736088, + 0.06838280707597733, + 0.07869719713926315, + 0.015057831071317196, + 0.007476107683032751, + 0.00997946597635746, + 0.006600706372410059, + 0.008816665038466454, + 0.01125979796051979, + 0.006118358112871647, + 0.00507589103654027, + 0.007512988056987524, + 0.0077774981036782265, + 0.015433506108820438, + 0.023477530106902122, + 0.032658252865076065, + 0.0403190553188324, + 0.017782533541321754, + 0.009240402840077877, + 0.020270511507987976, + 0.017262784764170647, + 0.020193306729197502, + 0.02589748241007328, + 0.032747965306043625, + 0.032079823315143585, + 0.01893952116370201, + 0.015093056485056877, + 0.023235145956277847, + 0.019411982968449593, + 0.007861445657908916, + 0.00968747679144144, + 0.013640541583299637, + 0.013223545625805855, + 0.00938358623534441, + 0.0060730851255357265, + 0.010534964501857758, + 0.009427088312804699, + 0.004451671615242958, + 0.008281801827251911, + 0.0063358498737216, + 0.007276267744600773, + 0.004848693497478962, + 0.004008854739367962, + 0.007115605752915144, + 0.005506242159754038, + 0.009297876618802547, + 0.008389362134039402, + 0.009082140401005745, + 0.01197904534637928, + 0.008617108687758446, + 0.007253007031977177, + 0.010383092798292637, + 0.01258112769573927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.0031949449330568314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557209432125092 + ], + "opposite_logprob_tokens": [ + -1.600572109222412 + ], + "logprob_reward": -0.22557209432125092, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557209432125092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05293160676956177, + 0.12893080711364746, + 0.06192096695303917, + 0.17939917743206024, + 0.12331482768058777, + 0.1280750334262848, + 0.17175796627998352, + 0.08465579897165298, + 0.04498838260769844, + 0.02402549982070923 + ], + "credit_map": [ + 0.007084682118147612, + 0.005045895464718342, + 0.008029935881495476, + 0.005283270496875048, + 0.009638885967433453, + 0.004920832812786102, + 0.007021486293524504, + 0.005906618200242519, + 0.014288675040006638, + 0.0076832701452076435, + 0.014583099633455276, + 0.026920881122350693, + 0.009093469940125942, + 0.007591577712446451, + 0.029247375205159187, + 0.01952245458960533, + 0.006882938090711832, + 0.009021460078656673, + 0.01011872198432684, + 0.010481296107172966, + 0.005138819105923176, + 0.004209738224744797, + 0.008940252475440502, + 0.007127733435481787, + 0.01279847789555788, + 0.01635172963142395, + 0.010553997941315174, + 0.012004983611404896, + 0.0567481592297554, + 0.05028815567493439, + 0.014064459130167961, + 0.006589217577129602, + 0.014060920104384422, + 0.012409473769366741, + 0.019705910235643387, + 0.026278994977474213, + 0.011042753234505653, + 0.008459101431071758, + 0.018937354907393456, + 0.012420319952070713, + 0.0109251094982028, + 0.011822319589555264, + 0.025189703330397606, + 0.03737513720989227, + 0.00987753365188837, + 0.006645460147410631, + 0.016233274713158607, + 0.010006486438214779, + 0.01518908329308033, + 0.020266542211174965, + 0.028617747128009796, + 0.022829405963420868, + 0.0223220344632864, + 0.016135800629854202, + 0.01574750430881977, + 0.03064984641969204, + 0.007973170839250088, + 0.011731999926269054, + 0.013812592253088951, + 0.011728032492101192, + 0.009009236469864845, + 0.006080418359488249, + 0.014076374471187592, + 0.010243971832096577, + 0.004473949782550335, + 0.00905897282063961, + 0.005840398371219635, + 0.006687010172754526, + 0.004826663993299007, + 0.0034789610654115677, + 0.006379425060003996, + 0.004243001341819763, + 0.0017729009268805385, + 0.0036767972633242607, + 0.002044816268607974, + 0.0041770716197788715, + 0.0028215176425874233, + 0.0023849718272686005, + 0.003427556250244379, + 0.003719867905601859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.22557209432125092 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot, tracking and rapidly panning sideways, captures the full force of a heavy main battle tank, its olive-green armor streaked with dried mud and fresh, wet dirt. It lies on a vast, muddy field, beneath a dark brown, sticky mud riddled with waterlogged ruts, reflecting the gloomy gray sky. With a dull roar, the steel behemoth swerves sharply, its massive hull tilting violently. One massive metal track sinks deeply into the saturated soil, practically stagnating as a fulcrum, while the other track spins at breakneck speed. This movement instantly transforms the track into a brutal excavator, hurling huge chunks of wet mud and water into the air in chaotic arcs, forming a fleeting, dark curtain of mud. The flying particles of earth are clearly visible in the diffuse, lifeless sunlight, imbuing the entire image with a raw, mechanical aesthetic of violence, creating an atmosphere of tension, unstoppable force, and immense power.", + "seed": 9879, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_0/index_1325_seed_9879_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10269146356731654, + "vlm_logprob": -0.10269146356731654, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.10269146356731654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18616653978824615 + ], + "opposite_logprob_tokens": [ + -2.436166524887085 + ], + "logprob_reward": -0.18616653978824615, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.18616653978824615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12464004009962082, + 0.06376899778842926, + 0.04715057462453842, + 0.07659393548965454, + 0.06178992614150047, + 0.06750252097845078, + 0.169612854719162, + 0.24214482307434082, + 0.07012899965047836, + 0.07666727900505066 + ], + "credit_map": [ + 0.009911034256219864, + 0.01064313855022192, + 0.011255388148128986, + 0.009059715084731579, + 0.017349617555737495, + 0.021775631234049797, + 0.021430358290672302, + 0.023215169087052345, + 0.012765910476446152, + 0.007882037200033665, + 0.005653676111251116, + 0.004834441002458334, + 0.007804228458553553, + 0.00966150127351284, + 0.00815669260919094, + 0.00701051764190197, + 0.00728549575433135, + 0.005340708419680595, + 0.00704637635499239, + 0.004729935899376869, + 0.004972746595740318, + 0.0056791300885379314, + 0.006758787669241428, + 0.005337398033589125, + 0.011851035989820957, + 0.013735990040004253, + 0.008049165830016136, + 0.009160950779914856, + 0.010483619756996632, + 0.008125515654683113, + 0.007851368747651577, + 0.007336292415857315, + 0.009191693738102913, + 0.006844762712717056, + 0.012158320285379887, + 0.008035961538553238, + 0.00524537917226553, + 0.004535578191280365, + 0.0072262282483279705, + 0.0085520064458251, + 0.004563195630908012, + 0.007227700669318438, + 0.01054274383932352, + 0.010967421345412731, + 0.005822791717946529, + 0.006894031073898077, + 0.010808343067765236, + 0.010676302015781403, + 0.01260424219071865, + 0.01927732676267624, + 0.019207226112484932, + 0.012758257798850536, + 0.0394810326397419, + 0.024072928354144096, + 0.017757169902324677, + 0.02445468120276928, + 0.01079533901065588, + 0.012163189239799976, + 0.07119536399841309, + 0.013564012013375759, + 0.017380909994244576, + 0.07920429110527039, + 0.01472757663577795, + 0.023114167153835297, + 0.005342116113752127, + 0.007992719300091267, + 0.014173761010169983, + 0.004703943617641926, + 0.008139876648783684, + 0.021956367418169975, + 0.0027722904924303293, + 0.005047930870205164, + 0.010276729241013527, + 0.009355793707072735, + 0.011257543228566647, + 0.005454710219055414, + 0.011673624627292156, + 0.01764487661421299, + 0.004207598976790905, + 0.0067964098416268826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.18616653978824615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016119513660669327 + ], + "opposite_logprob_tokens": [ + -4.141119480133057 + ], + "logprob_reward": -0.016119513660669327, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016119513660669327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06448478251695633, + 0.08384747803211212, + 0.04284578561782837, + 0.04326708987355232, + 0.034741081297397614, + 0.035380225628614426, + 0.04273777827620506, + 0.13405713438987732, + 0.3888300359249115, + 0.12980864942073822 + ], + "credit_map": [ + 0.005040591582655907, + 0.005431159865111113, + 0.007409121375530958, + 0.006889247335493565, + 0.008600407280027866, + 0.011525243520736694, + 0.010407245717942715, + 0.009181763976812363, + 0.011731823906302452, + 0.010524963960051537, + 0.013107717968523502, + 0.010518206283450127, + 0.008738013915717602, + 0.010938361287117004, + 0.009914848022162914, + 0.008373533375561237, + 0.005593386013060808, + 0.0038840207271277905, + 0.004276246763765812, + 0.0036504941526800394, + 0.008829153142869473, + 0.005606291349977255, + 0.006221661809831858, + 0.004784529563039541, + 0.005821135826408863, + 0.006714607123285532, + 0.006504395976662636, + 0.0038159240502864122, + 0.003582881297916174, + 0.004700926598161459, + 0.006981848739087582, + 0.005145368631929159, + 0.0035784461069852114, + 0.005194215103983879, + 0.005935323424637318, + 0.00450093625113368, + 0.002931486116722226, + 0.002360096899792552, + 0.004805048927664757, + 0.005435526836663485, + 0.002572893863543868, + 0.005116267595440149, + 0.00524158077314496, + 0.003991481848061085, + 0.003390103578567505, + 0.0052507854998111725, + 0.004825664218515158, + 0.004991449415683746, + 0.0053199781104922295, + 0.004263224545866251, + 0.006910625379532576, + 0.004127061925828457, + 0.004735371097922325, + 0.0061923665925860405, + 0.004237964749336243, + 0.00695118447765708, + 0.006155192852020264, + 0.005155778490006924, + 0.04422377794981003, + 0.009168867021799088, + 0.007191890850663185, + 0.03777139633893967, + 0.009679444134235382, + 0.014710781164467335, + 0.021457280963659286, + 0.012649452313780785, + 0.1049332320690155, + 0.019656339660286903, + 0.04291744530200958, + 0.15479949116706848, + 0.008547944016754627, + 0.023868832737207413, + 0.014986546710133553, + 0.01910005323588848, + 0.025078780949115753, + 0.012333571910858154, + 0.016449378803372383, + 0.02119494043290615, + 0.00782347097992897, + 0.012841899879276752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.016119513660669327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497223883867264 + ], + "opposite_logprob_tokens": [ + -2.9299721717834473 + ], + "logprob_reward": -0.05497223883867264, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497223883867264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06226002052426338, + 0.18311148881912231, + 0.21801790595054626, + 0.12916366755962372, + 0.03812912851572037, + 0.04794779419898987, + 0.027631551027297974, + 0.20230089128017426, + 0.06876378506422043, + 0.022673742845654488 + ], + "credit_map": [ + 0.009960561990737915, + 0.006892123259603977, + 0.005174858961254358, + 0.005564472172409296, + 0.009220643900334835, + 0.011853444389998913, + 0.006086844019591808, + 0.007507070433348417, + 0.0346536785364151, + 0.022450460121035576, + 0.02287893556058407, + 0.020269066095352173, + 0.018149754032492638, + 0.02302180603146553, + 0.025501249358057976, + 0.016186531633138657, + 0.02564547397196293, + 0.01934315450489521, + 0.030680181458592415, + 0.02269718050956726, + 0.03173636645078659, + 0.026786889880895615, + 0.033474840223789215, + 0.027653813362121582, + 0.016340970993041992, + 0.026337945833802223, + 0.014995599165558815, + 0.013373168185353279, + 0.012811566703021526, + 0.013169202953577042, + 0.02206697314977646, + 0.010068234987556934, + 0.003968244418501854, + 0.004448056221008301, + 0.006856623571366072, + 0.005948358215391636, + 0.0031332923099398613, + 0.0027676341123878956, + 0.005771629046648741, + 0.00523529015481472, + 0.00363948754966259, + 0.003977422136813402, + 0.00923794787377119, + 0.005780104082077742, + 0.005974208004772663, + 0.00729875173419714, + 0.005925258155912161, + 0.006114608142524958, + 0.003942886367440224, + 0.002236913191154599, + 0.0034810069482773542, + 0.0028810834046453238, + 0.0041507696732878685, + 0.003921860363334417, + 0.0031761019490659237, + 0.0038409268017858267, + 0.008799700066447258, + 0.007587447762489319, + 0.0678575187921524, + 0.014181377366185188, + 0.012458240613341331, + 0.05444290116429329, + 0.014505183324217796, + 0.02246851474046707, + 0.003986922558397055, + 0.004035112913697958, + 0.01609734073281288, + 0.0036966593470424414, + 0.008914605714380741, + 0.026557298377156258, + 0.0017436979105696082, + 0.003732141572982073, + 0.0016853391425684094, + 0.003122929250821471, + 0.004227842669934034, + 0.002821065951138735, + 0.0038812204729765654, + 0.0026422985829412937, + 0.0016152143944054842, + 0.002677832031622529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.05497223883867264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098587669432163 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098587669432163, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098587669432163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05909034237265587, + 0.11831163614988327, + 0.08917734026908875, + 0.1690492331981659, + 0.03334284946322441, + 0.10196047276258469, + 0.07693032920360565, + 0.08767976611852646, + 0.22568656504154205, + 0.03877147659659386 + ], + "credit_map": [ + 0.005937662906944752, + 0.004404701758176088, + 0.005387526471167803, + 0.0061127375811338425, + 0.007479235529899597, + 0.010116882622241974, + 0.009761962108314037, + 0.009889631532132626, + 0.01860054023563862, + 0.013463366776704788, + 0.014180868864059448, + 0.014676801860332489, + 0.013724293559789658, + 0.015994593501091003, + 0.016489336267113686, + 0.011181818321347237, + 0.011886137537658215, + 0.010228144004940987, + 0.012277974747121334, + 0.007722568232566118, + 0.011910224333405495, + 0.011870474554598331, + 0.012134486809372902, + 0.011147325858473778, + 0.016966909170150757, + 0.03190365061163902, + 0.02250867523252964, + 0.019256949424743652, + 0.01417907327413559, + 0.015690594911575317, + 0.03212399408221245, + 0.016419386491179466, + 0.004064283333718777, + 0.00500232120975852, + 0.005670186132192612, + 0.004427996464073658, + 0.0028111073188483715, + 0.0029713932890444994, + 0.004414167255163193, + 0.003981390502303839, + 0.008018016815185547, + 0.009744558483362198, + 0.02022978477180004, + 0.012964594177901745, + 0.008044934831559658, + 0.009635756723582745, + 0.015579412691295147, + 0.017743408679962158, + 0.00840978603810072, + 0.008454060181975365, + 0.009306224063038826, + 0.008300365880131721, + 0.009532930329442024, + 0.011664928868412971, + 0.00847520213574171, + 0.012786827981472015, + 0.006410904228687286, + 0.006412203423678875, + 0.018560191616415977, + 0.010728553868830204, + 0.008639724925160408, + 0.011392385698854923, + 0.012625634670257568, + 0.01291015837341547, + 0.012442111037671566, + 0.007505413610488176, + 0.0628955289721489, + 0.011787669733166695, + 0.025023803114891052, + 0.0895276814699173, + 0.004865072667598724, + 0.01163927186280489, + 0.006085634231567383, + 0.004690599162131548, + 0.005247253924608231, + 0.003139028325676918, + 0.004973555915057659, + 0.008694605901837349, + 0.0027354531921446323, + 0.003205343149602413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.004098587669432163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavy battle tank with olive-green armor streaked with dried mud and fresh wet dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank swerve sharply with one track sinking deeply into the soil while the other spins rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are huge chunks of wet mud and water being hurled into the air in chaotic arcs by the spinning track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking and rapidly panning sideways throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521004378795624 + ], + "opposite_logprob_tokens": [ + -1.5021004676818848 + ], + "logprob_reward": -0.2521004378795624, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521004378795624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07879796624183655, + 0.10186412930488586, + 0.13929690420627594, + 0.08209425956010818, + 0.045727141201496124, + 0.05427001044154167, + 0.05605415254831314, + 0.1609836369752884, + 0.23742355406284332, + 0.0434882715344429 + ], + "credit_map": [ + 0.004808593541383743, + 0.005123404320329428, + 0.0075958794914186, + 0.008096454665064812, + 0.01107086893171072, + 0.017465779557824135, + 0.010867794044315815, + 0.013769191689789295, + 0.01824343204498291, + 0.012182731181383133, + 0.014658286236226559, + 0.011001178994774818, + 0.010337510146200657, + 0.01212229486554861, + 0.014745308086276054, + 0.008573383092880249, + 0.021516313776373863, + 0.019183048978447914, + 0.01621108129620552, + 0.011215652339160442, + 0.022888043895363808, + 0.021504057571291924, + 0.015021082013845444, + 0.01175762340426445, + 0.008600801229476929, + 0.011497755534946918, + 0.009886207990348339, + 0.007128582336008549, + 0.01335254218429327, + 0.009598512202501297, + 0.014344364404678345, + 0.007685495540499687, + 0.005673663225024939, + 0.004914356395602226, + 0.008528654463589191, + 0.007326408289372921, + 0.0031169948633760214, + 0.0033585571218281984, + 0.006589149124920368, + 0.006219359580427408, + 0.003948213532567024, + 0.006686116568744183, + 0.010009311139583588, + 0.005837887525558472, + 0.004142682068049908, + 0.005723288748413324, + 0.009164858609437943, + 0.008757649920880795, + 0.003917057998478413, + 0.006411936599761248, + 0.01025834959000349, + 0.007506383582949638, + 0.007263634353876114, + 0.006486391648650169, + 0.0061157518066465855, + 0.008094646967947483, + 0.012085183523595333, + 0.011549606919288635, + 0.022656917572021484, + 0.011379768140614033, + 0.017066366970539093, + 0.043147701770067215, + 0.020488450303673744, + 0.022609656676650047, + 0.014932314865291119, + 0.008609560318291187, + 0.06769188493490219, + 0.012669241055846214, + 0.025286424905061722, + 0.0904269739985466, + 0.005044588819146156, + 0.012762568891048431, + 0.0042458572424948215, + 0.004890150390565395, + 0.006693684495985508, + 0.003579587209969759, + 0.007537295576184988, + 0.0090873371809721, + 0.004029134288430214, + 0.003425226081162691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.2521004378795624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_tensor([1171, 1171])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_tensor([1171, 1171])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..603350964b404db26100fd86a8f00b10ba55978a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_tensor([1171, 1171])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "grand, upward-looking rotating camera smoothly and majestically orbits a massive dragon, soaring above a boundless sea of clouds. The dragon's body is long and sinuous, covered in layers of crisscrossing, crisscrossing cyan scales, each one seemingly carved from the purest jade. Under the warm, golden light of the low sun, it gleams with an iridescent pearlescent sheen. It possesses towering antlers and a flowing, silver-white beard. Its eyes, blazing with golden light, radiate ancient wisdom and majesty. As it gracefully coils and unfurls, a pale white mane of pure energy undulates along its back. Surrounding it is a thick, fluffy sea of white clouds, tinged warmly by the golden sunlight. Each movement of the dragon carves gentle swirls and trails across the soft cloudtops. As the camera moves steadily 360 degrees, we can appreciate its huge and elegant figure from different angles. The light keeps flowing on its scales, sometimes bright as a mirror, sometimes deep as the sea. The whole picture is full of sacred, tranquil and awe-inspiring epic feeling.", + "seed": 15415, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_1171_seed_15415_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5318168871104717, + "vlm_logprob": -0.5318168871104717, + "vlm_yes_no_margin": 1.7, + "vlm_token_credit": -0.5318168871104717, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4067119359970093 + ], + "opposite_logprob_tokens": [ + -0.28171196579933167 + ], + "logprob_reward": -1.4067119359970093, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4067119359970093, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20066924393177032, + 0.05222207307815552, + 0.03581686690449715, + 0.04443444311618805, + 0.0667780414223671, + 0.06699816882610321, + 0.04289223998785019, + 0.039436668157577515, + 0.24525749683380127, + 0.205494686961174 + ], + "credit_map": [ + 0.019078928977251053, + 0.02613600343465805, + 0.035448793321847916, + 0.027170510962605476, + 0.020611876621842384, + 0.01795372925698757, + 0.022422093898057938, + 0.03184732049703598, + 0.003712696023285389, + 0.0054441457614302635, + 0.009472282603383064, + 0.005382858216762543, + 0.005829792004078627, + 0.006392649374902248, + 0.011409304104745388, + 0.004578346852213144, + 0.004595092963427305, + 0.0033113507088273764, + 0.005807762965559959, + 0.004210637416690588, + 0.0030915027018636465, + 0.005466431379318237, + 0.004556263796985149, + 0.004777827300131321, + 0.005868087988346815, + 0.00536296097561717, + 0.008194485679268837, + 0.0052519324235618114, + 0.0035655370447784662, + 0.0063816215842962265, + 0.005312376655638218, + 0.0044974456541240215, + 0.006064703222364187, + 0.0073051149956882, + 0.007573365233838558, + 0.010062743909657001, + 0.008290579542517662, + 0.01571151427924633, + 0.006754449103027582, + 0.0050155785866081715, + 0.006475473754107952, + 0.006833937019109726, + 0.009592731483280659, + 0.009020638652145863, + 0.006744044367223978, + 0.012243511155247688, + 0.009637460112571716, + 0.006450370419770479, + 0.006642800755798817, + 0.006403160281479359, + 0.004104376304894686, + 0.0033926148898899555, + 0.005475462879985571, + 0.0035571265034377575, + 0.00709361070767045, + 0.006223087199032307, + 0.005093984771519899, + 0.004308043979108334, + 0.005897468887269497, + 0.0072382171638309956, + 0.0034634002950042486, + 0.0030106857884675264, + 0.00663452222943306, + 0.0037903490010648966, + 0.030918847769498825, + 0.03051568754017353, + 0.04593025892972946, + 0.03238550201058388, + 0.02305399812757969, + 0.019633624702692032, + 0.03412872925400734, + 0.028690870851278305, + 0.023195412009954453, + 0.029103493317961693, + 0.03126473352313042, + 0.01766842044889927, + 0.021132221445441246, + 0.038799408823251724, + 0.023550141602754593, + 0.020780863240361214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -1.4067119359970093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018228041008114815 + ], + "opposite_logprob_tokens": [ + -4.018228054046631 + ], + "logprob_reward": -0.018228041008114815, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018228041008114815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1589530110359192, + 0.053219206631183624, + 0.026581523939967155, + 0.05586293712258339, + 0.06684267520904541, + 0.06655892729759216, + 0.06594939529895782, + 0.03790894150733948, + 0.23460344970226288, + 0.23351995646953583 + ], + "credit_map": [ + 0.016845716163516045, + 0.023584242910146713, + 0.025705896317958832, + 0.0168803371489048, + 0.02047405019402504, + 0.018887557089328766, + 0.015923090279102325, + 0.02065212093293667, + 0.0040409257635474205, + 0.005674157757312059, + 0.009704730473458767, + 0.0053905765525996685, + 0.004947050474584103, + 0.007197996135801077, + 0.0104195736348629, + 0.005844197701662779, + 0.004249914083629847, + 0.0031006583012640476, + 0.0037129377014935017, + 0.003246456850320101, + 0.0016706963069736958, + 0.002980618504807353, + 0.003674222156405449, + 0.003946018870919943, + 0.009626545011997223, + 0.007954602129757404, + 0.007669531740248203, + 0.007240576203912497, + 0.004400951787829399, + 0.006143630016595125, + 0.006753217428922653, + 0.0060738809406757355, + 0.005599158350378275, + 0.006977400742471218, + 0.007423293311148882, + 0.00789471436291933, + 0.009624401107430458, + 0.017566028982400894, + 0.006826777011156082, + 0.004930903669446707, + 0.00824352540075779, + 0.0070472895167768, + 0.009102928452193737, + 0.008111553266644478, + 0.00861327163875103, + 0.010530108585953712, + 0.009535287506878376, + 0.00537496292963624, + 0.011468629352748394, + 0.011616107076406479, + 0.011101579293608665, + 0.006393909454345703, + 0.006545399781316519, + 0.005856635048985481, + 0.006236682645976543, + 0.006730454042553902, + 0.005926798563450575, + 0.00373335974290967, + 0.00604753103107214, + 0.00709401722997427, + 0.003167581744492054, + 0.002566603012382984, + 0.006047932896763086, + 0.003325121244415641, + 0.02940915897488594, + 0.02889358438551426, + 0.045133449137210846, + 0.027132513001561165, + 0.02176116406917572, + 0.019887538626790047, + 0.03319653123617172, + 0.029189500957727432, + 0.025143520906567574, + 0.03372287377715111, + 0.03250635415315628, + 0.019795868545770645, + 0.025275884196162224, + 0.04737835377454758, + 0.024748262017965317, + 0.0249488428235054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.018228041008114815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2522346079349518 + ], + "opposite_logprob_tokens": [ + -1.5022345781326294 + ], + "logprob_reward": -0.2522346079349518, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2522346079349518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3692184090614319, + 0.04479395970702171, + 0.027618426829576492, + 0.059830959886312485, + 0.07644018530845642, + 0.12767937779426575, + 0.05881384015083313, + 0.033069733530282974, + 0.13712318241596222, + 0.06541188061237335 + ], + "credit_map": [ + 0.03012692928314209, + 0.04857252910733223, + 0.06227602809667587, + 0.04844793677330017, + 0.049973879009485245, + 0.05282357335090637, + 0.037323158234357834, + 0.03967437893152237, + 0.004907865542918444, + 0.005898808129131794, + 0.0063618822023272514, + 0.003788062371313572, + 0.004357681144028902, + 0.0055698794312775135, + 0.008926719427108765, + 0.0049830591306090355, + 0.003031868953257799, + 0.0026023760437965393, + 0.0030469780322164297, + 0.00271527748554945, + 0.0024143375921994448, + 0.006786678452044725, + 0.0031182458624243736, + 0.003902664640918374, + 0.006448627449572086, + 0.007598129101097584, + 0.011613299138844013, + 0.00734951626509428, + 0.005780278705060482, + 0.008403856307268143, + 0.006672834511846304, + 0.005964415613561869, + 0.00994101446121931, + 0.015395943075418472, + 0.007191137410700321, + 0.010539861395955086, + 0.009514533914625645, + 0.010504447855055332, + 0.0073535628616809845, + 0.005999691318720579, + 0.018420783802866936, + 0.024675549939274788, + 0.011058599688112736, + 0.02216789871454239, + 0.016924163326621056, + 0.009025374427437782, + 0.013845090754330158, + 0.011561914347112179, + 0.009773102588951588, + 0.013418522663414478, + 0.004630648531019688, + 0.005080004222691059, + 0.005409514997154474, + 0.004842220339924097, + 0.00624806247651577, + 0.009411763399839401, + 0.004510883241891861, + 0.00393398804590106, + 0.005142600741237402, + 0.006301500368863344, + 0.002776332199573517, + 0.0020509129390120506, + 0.005644472315907478, + 0.002709046471863985, + 0.016519246622920036, + 0.016770625486969948, + 0.032574206590652466, + 0.01696177013218403, + 0.011349517852067947, + 0.008960721082985401, + 0.017964988946914673, + 0.016022104769945145, + 0.00803772546350956, + 0.008587312884628773, + 0.01358843594789505, + 0.007957505993545055, + 0.007747394498437643, + 0.007300690282136202, + 0.0066917468793690205, + 0.005501067731529474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.2522346079349518 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076280031353235245 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.0076280031353235245, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076280031353235245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23860804736614227, + 0.13361728191375732, + 0.031348079442977905, + 0.03809308633208275, + 0.06912323832511902, + 0.058901097625494, + 0.06368574500083923, + 0.02395607903599739, + 0.22030135989189148, + 0.12236598879098892 + ], + "credit_map": [ + 0.013209765776991844, + 0.01814606972038746, + 0.04092736169695854, + 0.050304047763347626, + 0.022763988003134727, + 0.018342310562729836, + 0.02943962998688221, + 0.04547487571835518, + 0.008971252478659153, + 0.009911078959703445, + 0.021604225039482117, + 0.017978062853217125, + 0.01347153540700674, + 0.017088832333683968, + 0.02928079105913639, + 0.015311505645513535, + 0.003959996160119772, + 0.0027273076120764017, + 0.004858081694692373, + 0.0041348254308104515, + 0.0022230593021959066, + 0.004639537073671818, + 0.0035633412189781666, + 0.00524193374440074, + 0.006008853670209646, + 0.004370690789073706, + 0.005435279104858637, + 0.0048104566521942616, + 0.002867465140298009, + 0.005296697374433279, + 0.005032863933593035, + 0.004270779900252819, + 0.00514268409460783, + 0.007265869993716478, + 0.007599193602800369, + 0.008176099509000778, + 0.010064173489809036, + 0.018758567050099373, + 0.007002027705311775, + 0.005114625673741102, + 0.006396799813956022, + 0.007275329437106848, + 0.010723631829023361, + 0.007133007049560547, + 0.006060929968953133, + 0.00805694330483675, + 0.007914355956017971, + 0.005340097937732935, + 0.012144237756729126, + 0.010084234178066254, + 0.009757502935826778, + 0.006685301195830107, + 0.006994801107794046, + 0.005520632490515709, + 0.006739927921444178, + 0.005759104620665312, + 0.0032303028274327517, + 0.002662121318280697, + 0.003649720922112465, + 0.004618083126842976, + 0.0023572049103677273, + 0.0016939984634518623, + 0.0033885871525853872, + 0.002356060314923525, + 0.02701397053897381, + 0.027036111801862717, + 0.04226509481668472, + 0.030401693657040596, + 0.02052108570933342, + 0.016956821084022522, + 0.029977543279528618, + 0.02612905390560627, + 0.014574116095900536, + 0.01765698753297329, + 0.01822124607861042, + 0.01015383843332529, + 0.012775099836289883, + 0.023265404626727104, + 0.013988344930112362, + 0.011730947531759739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0076280031353235245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742818474769592 + ], + "opposite_logprob_tokens": [ + -0.47428184747695923 + ], + "logprob_reward": -0.9742818474769592, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742818474769592, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.152313694357872, + 0.07683780044317245, + 0.10390915721654892, + 0.0822853296995163, + 0.11409498751163483, + 0.08455485105514526, + 0.06030350178480148, + 0.049232445657253265, + 0.15752187371253967, + 0.11894632875919342 + ], + "credit_map": [ + 0.014829664491117, + 0.018932251259684563, + 0.025130119174718857, + 0.017449554055929184, + 0.018644098192453384, + 0.01986047439277172, + 0.016567636281251907, + 0.020899903029203415, + 0.006286615040153265, + 0.007211944554001093, + 0.015037531033158302, + 0.00839936826378107, + 0.009363150224089622, + 0.009938442148268223, + 0.014428691938519478, + 0.006172057706862688, + 0.008557910099625587, + 0.006703820079565048, + 0.01413098443299532, + 0.01365489698946476, + 0.008003349415957928, + 0.023891422897577286, + 0.015364943072199821, + 0.01360183022916317, + 0.014719047583639622, + 0.009802097454667091, + 0.012837864458560944, + 0.011936275288462639, + 0.006645950023084879, + 0.008360090665519238, + 0.00821820180863142, + 0.009765802882611752, + 0.014860620722174644, + 0.015171411447227001, + 0.014531552791595459, + 0.025017349049448967, + 0.01180256437510252, + 0.011896952986717224, + 0.01065110880881548, + 0.010163423605263233, + 0.007834427990019321, + 0.007807189133018255, + 0.01343134418129921, + 0.009835992939770222, + 0.008538995869457722, + 0.016865992918610573, + 0.011901402845978737, + 0.00833949912339449, + 0.01504087820649147, + 0.008143783546984196, + 0.00795973464846611, + 0.00413311505690217, + 0.007462165784090757, + 0.004665756598114967, + 0.006377536337822676, + 0.006520530674606562, + 0.006492764689028263, + 0.006367075722664595, + 0.005145554896444082, + 0.005442350637167692, + 0.007814357057213783, + 0.006737555377185345, + 0.005941316019743681, + 0.005291472654789686, + 0.021654628217220306, + 0.020246202126145363, + 0.03344652056694031, + 0.018735401332378387, + 0.016636531800031662, + 0.012576354667544365, + 0.01873328723013401, + 0.015492950566112995, + 0.015223046764731407, + 0.016360189765691757, + 0.025772133842110634, + 0.00991669949144125, + 0.01091168262064457, + 0.018223954364657402, + 0.01287524588406086, + 0.009663369506597519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.9742818474769592 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "grand, upward-looking rotating camera smoothly and majestically orbits a massive dragon, soaring above a boundless sea of clouds. The dragon's body is long and sinuous, covered in layers of crisscrossing, crisscrossing cyan scales, each one seemingly carved from the purest jade. Under the warm, golden light of the low sun, it gleams with an iridescent pearlescent sheen. It possesses towering antlers and a flowing, silver-white beard. Its eyes, blazing with golden light, radiate ancient wisdom and majesty. As it gracefully coils and unfurls, a pale white mane of pure energy undulates along its back. Surrounding it is a thick, fluffy sea of white clouds, tinged warmly by the golden sunlight. Each movement of the dragon carves gentle swirls and trails across the soft cloudtops. As the camera moves steadily 360 degrees, we can appreciate its huge and elegant figure from different angles. The light keeps flowing on its scales, sometimes bright as a mirror, sometimes deep as the sea. The whole picture is full of sacred, tranquil and awe-inspiring epic feeling.", + "seed": 15416, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_1171_seed_15416_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.22547482941299676, + "vlm_logprob": -0.22547482941299676, + "vlm_yes_no_margin": 2.574999952316284, + "vlm_token_credit": -0.22547482941299676, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08926975727081299 + ], + "opposite_logprob_tokens": [ + -2.4642696380615234 + ], + "logprob_reward": -0.08926975727081299, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08926975727081299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017187895253300667, + 0.011371796019375324, + 0.043648216873407364, + 0.06164097785949707, + 0.03527136519551277, + 0.014687484130263329, + 0.0367765910923481, + 0.07146227359771729, + 0.04250146448612213, + 0.6654520034790039 + ], + "credit_map": [ + 0.0025283920112997293, + 0.001560583827085793, + 0.003035136265680194, + 0.0015435743844136596, + 0.001648968318477273, + 0.001947716693393886, + 0.001845477381721139, + 0.003078044857829809, + 0.0007349654915742576, + 0.0008160079014487565, + 0.0022600425872951746, + 0.0014791736612096429, + 0.0007743688183836639, + 0.001957036554813385, + 0.0011296746088191867, + 0.002220525173470378, + 0.002128711435943842, + 0.0017654674593359232, + 0.005259144119918346, + 0.0030396394431591034, + 0.0018926169723272324, + 0.010997927747666836, + 0.005907769314944744, + 0.01265693549066782, + 0.008544012904167175, + 0.004385932348668575, + 0.004753696266561747, + 0.00822074431926012, + 0.004078636411577463, + 0.0137552535161376, + 0.005254257936030626, + 0.012648443691432476, + 0.0041236658580601215, + 0.003899955889210105, + 0.007042214274406433, + 0.0034919227473437786, + 0.003147416515275836, + 0.004249129444360733, + 0.003904033452272415, + 0.005413023289293051, + 0.002211189828813076, + 0.0017404773971065879, + 0.002209458965808153, + 0.0015650782734155655, + 0.0013559485087171197, + 0.0018011961365118623, + 0.0013688126346096396, + 0.0024353202898055315, + 0.008113655261695385, + 0.003486522240564227, + 0.0034431619569659233, + 0.0031206547282636166, + 0.002876824699342251, + 0.004366712644696236, + 0.0037364806048572063, + 0.007632575463503599, + 0.013056439347565174, + 0.008813652209937572, + 0.008968615904450417, + 0.008839890360832214, + 0.006703867577016354, + 0.008722667582333088, + 0.0045730965211987495, + 0.011784037575125694, + 0.007276623044162989, + 0.00675075501203537, + 0.005677283741533756, + 0.004461785778403282, + 0.004490059334784746, + 0.0045310514979064465, + 0.0031925879884511232, + 0.006121314596384764, + 0.07091197371482849, + 0.07873699069023132, + 0.08812148869037628, + 0.07664471864700317, + 0.08727730065584183, + 0.05485204607248306, + 0.1354387402534485, + 0.0734686627984047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.08926975727081299 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0492507740855217, + 0.03482377529144287, + 0.08906518667936325, + 0.047356512397527695, + 0.06659793853759766, + 0.0471835657954216, + 0.08138757199048996, + 0.1837082952260971, + 0.08283223956823349, + 0.3177941143512726 + ], + "credit_map": [ + 0.00589453661814332, + 0.004660002421587706, + 0.007401288487017155, + 0.004153468180447817, + 0.0050745741464197636, + 0.009730014950037003, + 0.006326465401798487, + 0.006010422017425299, + 0.003797938348725438, + 0.0030250917188823223, + 0.005147254094481468, + 0.0073095401749014854, + 0.0025408752262592316, + 0.005036451853811741, + 0.0038118348456919193, + 0.004154786933213472, + 0.005219955928623676, + 0.005765228997915983, + 0.014425816014409065, + 0.008122137747704983, + 0.004734563175588846, + 0.018737923353910446, + 0.011853596195578575, + 0.020205961540341377, + 0.005780267994850874, + 0.005350486841052771, + 0.004576425999403, + 0.006562197580933571, + 0.004069391638040543, + 0.007264632731676102, + 0.006432280410081148, + 0.007320824079215527, + 0.007951653562486172, + 0.005949893966317177, + 0.008787048980593681, + 0.005929834675043821, + 0.006557885091751814, + 0.009006643667817116, + 0.006189410109072924, + 0.01622556336224079, + 0.005683265160769224, + 0.00458680372685194, + 0.006703582126647234, + 0.006659509148448706, + 0.005029614083468914, + 0.005330124404281378, + 0.005710135214030743, + 0.0074805305339396, + 0.010239405557513237, + 0.006815735716372728, + 0.00883344653993845, + 0.0069958786480128765, + 0.007306321524083614, + 0.013255497440695763, + 0.007633116561919451, + 0.02030816860496998, + 0.03373352810740471, + 0.02170722186565399, + 0.0224035382270813, + 0.022244438529014587, + 0.017232108861207962, + 0.022586673498153687, + 0.011888491921126842, + 0.031912289559841156, + 0.014154180884361267, + 0.008514093235135078, + 0.011900695040822029, + 0.006595497950911522, + 0.0057614827528595924, + 0.019069457426667213, + 0.005995764397084713, + 0.01084106881171465, + 0.03323521092534065, + 0.041009992361068726, + 0.04454965889453888, + 0.03585436940193176, + 0.041448965668678284, + 0.027732444927096367, + 0.060068342834711075, + 0.03389514237642288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.01819819025695324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295992136001587 + ], + "opposite_logprob_tokens": [ + -2.237959861755371 + ], + "logprob_reward": -0.11295992136001587, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295992136001587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11307855695486069, + 0.05176294967532158, + 0.09161291271448135, + 0.046982210129499435, + 0.07407954335212708, + 0.03379391133785248, + 0.1517447978258133, + 0.21036438643932343, + 0.052683696150779724, + 0.17389702796936035 + ], + "credit_map": [ + 0.013647854328155518, + 0.008674177341163158, + 0.018973514437675476, + 0.008273539133369923, + 0.010480505414307117, + 0.022663848474621773, + 0.015505579300224781, + 0.014859541319310665, + 0.004586558323353529, + 0.005595771130174398, + 0.00907122902572155, + 0.007985038682818413, + 0.003686056938022375, + 0.006728965323418379, + 0.005064106080681086, + 0.009045223705470562, + 0.008061147294938564, + 0.007568072993308306, + 0.01552548911422491, + 0.008789809420704842, + 0.004724970553070307, + 0.01768088899552822, + 0.00962012354284525, + 0.019642414525151253, + 0.005931387189775705, + 0.006387272384017706, + 0.004012556746602058, + 0.006207400932908058, + 0.006261252332478762, + 0.0072340574115514755, + 0.005570004228502512, + 0.005378279369324446, + 0.007267717272043228, + 0.0075311437249183655, + 0.011897394433617592, + 0.01035671029239893, + 0.009379717521369457, + 0.010086188092827797, + 0.006270779296755791, + 0.011289898306131363, + 0.003975613042712212, + 0.003108019009232521, + 0.0030899441335350275, + 0.0025178578216582537, + 0.0033220231998711824, + 0.005836966447532177, + 0.0030605450738221407, + 0.008882940746843815, + 0.02361452765762806, + 0.010892163962125778, + 0.011489290744066238, + 0.010771471075713634, + 0.012158713303506374, + 0.02700054459273815, + 0.014874657616019249, + 0.04094342142343521, + 0.041954223066568375, + 0.02682206593453884, + 0.02628576010465622, + 0.02595079317688942, + 0.020661192014813423, + 0.02639125846326351, + 0.013352849520742893, + 0.02894623950123787, + 0.011164954863488674, + 0.0028523108921945095, + 0.004328805021941662, + 0.0030930954962968826, + 0.0038647956680506468, + 0.01500822976231575, + 0.0035588303580880165, + 0.008812673389911652, + 0.020490601658821106, + 0.018803177401423454, + 0.02096652053296566, + 0.02134125679731369, + 0.02309054136276245, + 0.01449872087687254, + 0.033448994159698486, + 0.02125720866024494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.11295992136001587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635676465928555 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635676465928555, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635676465928555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07161759585142136, + 0.009433925151824951, + 0.032122306525707245, + 0.032228678464889526, + 0.024468282237648964, + 0.01952498033642769, + 0.06647663563489914, + 0.05435819923877716, + 0.1139884665608406, + 0.5757809281349182 + ], + "credit_map": [ + 0.008516432717442513, + 0.0036440668627619743, + 0.016202958300709724, + 0.004392446484416723, + 0.005076576955616474, + 0.012981795705854893, + 0.005208902060985565, + 0.015594419091939926, + 0.0007296415278688073, + 0.0006891829543747008, + 0.0012982001062482595, + 0.0013287793844938278, + 0.0007892888388596475, + 0.0021098044235259295, + 0.0009226257097907364, + 0.0015664021484553814, + 0.002217660890892148, + 0.0017614680109545588, + 0.004610114730894566, + 0.00274601555429399, + 0.0014492602786049247, + 0.006834397092461586, + 0.004098370671272278, + 0.008405018597841263, + 0.0025085462257266045, + 0.002225367119535804, + 0.00202712113969028, + 0.00405562249943614, + 0.002593579236418009, + 0.006548826117068529, + 0.0033070493955165148, + 0.008962566964328289, + 0.0027274948079138994, + 0.0023905092384666204, + 0.003437015926465392, + 0.00326459645293653, + 0.0028039589524269104, + 0.0033761425875127316, + 0.002240340691059828, + 0.004228223580867052, + 0.0023616915568709373, + 0.002000137697905302, + 0.0017523495480418205, + 0.0021085154730826616, + 0.002159327967092395, + 0.0031820840667933226, + 0.00196307641454041, + 0.003997797146439552, + 0.014724924229085445, + 0.006473566871136427, + 0.006902433466166258, + 0.005889664404094219, + 0.005051886197179556, + 0.00834931805729866, + 0.006227185949683189, + 0.012857655994594097, + 0.009870619513094425, + 0.006071351934224367, + 0.008008969947695732, + 0.00719921849668026, + 0.004074168857187033, + 0.006620919331908226, + 0.0034935527946799994, + 0.009019396267831326, + 0.02016168273985386, + 0.009166093543171883, + 0.010390156880021095, + 0.009025301784276962, + 0.011360534466803074, + 0.026149962097406387, + 0.012296900153160095, + 0.015437830239534378, + 0.06292621046304703, + 0.06501888483762741, + 0.07373880594968796, + 0.06550846993923187, + 0.07856140285730362, + 0.045328136533498764, + 0.12047216296195984, + 0.06422682851552963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.008635676465928555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983106017112732 + ], + "opposite_logprob_tokens": [ + -0.5233106017112732 + ], + "logprob_reward": -0.8983106017112732, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983106017112732, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.055266544222831726, + 0.02622602880001068, + 0.057397179305553436, + 0.048476096242666245, + 0.029065178707242012, + 0.050831735134124756, + 0.31965693831443787, + 0.16858404874801636, + 0.13824893534183502, + 0.10624735802412033 + ], + "credit_map": [ + 0.006215935107320547, + 0.005048487335443497, + 0.006527516525238752, + 0.005155791994184256, + 0.004770205356180668, + 0.010014298371970654, + 0.01093426626175642, + 0.006600036285817623, + 0.0027732891030609608, + 0.0036107187625020742, + 0.003744729794561863, + 0.004144655540585518, + 0.0026169007178395987, + 0.003433875273913145, + 0.0029069927986711264, + 0.002994864247739315, + 0.004726043902337551, + 0.0028242706321179867, + 0.0089363818988204, + 0.0042588962242007256, + 0.0030141212046146393, + 0.010512497276067734, + 0.008417350240051746, + 0.014707610942423344, + 0.00556225748732686, + 0.005271438974887133, + 0.003121117828413844, + 0.005812875460833311, + 0.005350826773792505, + 0.009002413600683212, + 0.005027406383305788, + 0.00932775717228651, + 0.0034895981661975384, + 0.002340613165870309, + 0.002922387793660164, + 0.005574349779635668, + 0.0030011411290615797, + 0.003996111452579498, + 0.0034827094059437513, + 0.004258266184478998, + 0.006225386634469032, + 0.0053969090804457664, + 0.004653627052903175, + 0.006094459909945726, + 0.0039781229570508, + 0.00786098837852478, + 0.004910064861178398, + 0.01171217393130064, + 0.060750048607587814, + 0.02810489386320114, + 0.02544364519417286, + 0.02260768786072731, + 0.022842690348625183, + 0.04467759281396866, + 0.031585924327373505, + 0.08364444971084595, + 0.036408066749572754, + 0.0241293553262949, + 0.017065634950995445, + 0.020808342844247818, + 0.019679773598909378, + 0.02048356831073761, + 0.010751050896942616, + 0.019258253276348114, + 0.021975606679916382, + 0.01879238709807396, + 0.016408251598477364, + 0.012542452663183212, + 0.01521294191479683, + 0.02394138276576996, + 0.01098589412868023, + 0.018390001729130745, + 0.013668181374669075, + 0.011597814038395882, + 0.010672623291611671, + 0.01225325558334589, + 0.011181456968188286, + 0.008756953291594982, + 0.021969128400087357, + 0.01614793762564659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.8983106017112732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "grand, upward-looking rotating camera smoothly and majestically orbits a massive dragon, soaring above a boundless sea of clouds. The dragon's body is long and sinuous, covered in layers of crisscrossing, crisscrossing cyan scales, each one seemingly carved from the purest jade. Under the warm, golden light of the low sun, it gleams with an iridescent pearlescent sheen. It possesses towering antlers and a flowing, silver-white beard. Its eyes, blazing with golden light, radiate ancient wisdom and majesty. As it gracefully coils and unfurls, a pale white mane of pure energy undulates along its back. Surrounding it is a thick, fluffy sea of white clouds, tinged warmly by the golden sunlight. Each movement of the dragon carves gentle swirls and trails across the soft cloudtops. As the camera moves steadily 360 degrees, we can appreciate its huge and elegant figure from different angles. The light keeps flowing on its scales, sometimes bright as a mirror, sometimes deep as the sea. The whole picture is full of sacred, tranquil and awe-inspiring epic feeling.", + "seed": 15417, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_1171_seed_15417_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.28835852432530373, + "vlm_logprob": -0.28835852432530373, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.28835852432530373, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8263486623764038 + ], + "opposite_logprob_tokens": [ + -0.5763486623764038 + ], + "logprob_reward": -0.8263486623764038, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8263486623764038, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09145820140838623, + 0.1491856873035431, + 0.03697894886136055, + 0.15741150081157684, + 0.05701814964413643, + 0.047021135687828064, + 0.18745087087154388, + 0.06694156676530838, + 0.12460941076278687, + 0.08192452788352966 + ], + "credit_map": [ + 0.01482997927814722, + 0.008091913536190987, + 0.021074088290333748, + 0.00973371509462595, + 0.007974308915436268, + 0.009034021757543087, + 0.010029117576777935, + 0.01069105789065361, + 0.01865202933549881, + 0.0139561602845788, + 0.027571314945816994, + 0.0197747815400362, + 0.014103957451879978, + 0.012380233965814114, + 0.022703252732753754, + 0.020043957978487015, + 0.003989896271377802, + 0.004337055142968893, + 0.006877811625599861, + 0.0037597508635371923, + 0.004671094007790089, + 0.004649486392736435, + 0.004031111020594835, + 0.004662745166569948, + 0.014505947940051556, + 0.019695231691002846, + 0.017357539385557175, + 0.015161893330514431, + 0.014954909682273865, + 0.03366812318563461, + 0.014406049624085426, + 0.027661798521876335, + 0.0049063665792346, + 0.004263824783265591, + 0.011859126389026642, + 0.0059718843549489975, + 0.0037236937787383795, + 0.011722469702363014, + 0.006390816997736692, + 0.008179968222975731, + 0.0034347185865044594, + 0.0012105789501219988, + 0.004207480698823929, + 0.001934334635734558, + 0.0013203175039961934, + 0.01738547720015049, + 0.0027536165434867144, + 0.014774610288441181, + 0.014981250278651714, + 0.02279464527964592, + 0.02210989221930504, + 0.021461401134729385, + 0.02557673119008541, + 0.02322782762348652, + 0.02869976870715618, + 0.02859935350716114, + 0.011656208895146847, + 0.006171572022140026, + 0.006273861508816481, + 0.008199126459658146, + 0.00773104652762413, + 0.013265937566757202, + 0.006101503502577543, + 0.0075423093512654305, + 0.018572885543107986, + 0.013307785615324974, + 0.02599279396235943, + 0.012078455649316311, + 0.01290056761354208, + 0.012798197567462921, + 0.013561645522713661, + 0.01539708487689495, + 0.014256423339247704, + 0.010452640242874622, + 0.010060997679829597, + 0.012125817127525806, + 0.0059737833216786385, + 0.008403006009757519, + 0.009573817253112793, + 0.011078041978180408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.8263486623764038 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596086978912354 + ], + "opposite_logprob_tokens": [ + -3.8955960273742676 + ], + "logprob_reward": -0.020596086978912354, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596086978912354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06556567549705505, + 0.056723158806562424, + 0.08330279588699341, + 0.07172445207834244, + 0.04559575021266937, + 0.4100291132926941, + 0.044655464589595795, + 0.03613495081663132, + 0.08101744949817657, + 0.10525114089250565 + ], + "credit_map": [ + 0.018874898552894592, + 0.006311697419732809, + 0.01143475342541933, + 0.005322852171957493, + 0.004709980916231871, + 0.00504909735172987, + 0.007043012883514166, + 0.0068193827755749226, + 0.006868832278996706, + 0.006216426845639944, + 0.011862467043101788, + 0.0049459137953817844, + 0.0055237277410924435, + 0.005753019824624062, + 0.007811840623617172, + 0.007740927394479513, + 0.008498696610331535, + 0.006036595441401005, + 0.010742180980741978, + 0.010794519446790218, + 0.012683781795203686, + 0.010161375626921654, + 0.012410479597747326, + 0.011975168250501156, + 0.006279022898525, + 0.00568086514249444, + 0.014305801130831242, + 0.0042814090847969055, + 0.004925672896206379, + 0.013535921461880207, + 0.00554472953081131, + 0.017171036452054977, + 0.004558524116873741, + 0.0030152585823088884, + 0.011502429842948914, + 0.004717308096587658, + 0.0029168210458010435, + 0.008676284924149513, + 0.003152844263240695, + 0.007056276313960552, + 0.02704933099448681, + 0.004494077991694212, + 0.027029139921069145, + 0.008158266544342041, + 0.010702041909098625, + 0.16576902568340302, + 0.015450882725417614, + 0.15137630701065063, + 0.004180170129984617, + 0.0045180488377809525, + 0.005402872338891029, + 0.005220436956733465, + 0.005604642443358898, + 0.007481672335416079, + 0.0066506327129900455, + 0.005596987437456846, + 0.004737221635878086, + 0.004448471590876579, + 0.005727639887481928, + 0.005055148154497147, + 0.0034129221457988024, + 0.003586005186662078, + 0.005068393424153328, + 0.0040991525165736675, + 0.011630920693278313, + 0.008885650895535946, + 0.01795077882707119, + 0.008516207337379456, + 0.00841417908668518, + 0.007250775583088398, + 0.008490120992064476, + 0.009878816083073616, + 0.016524771228432655, + 0.01732676289975643, + 0.011370059102773666, + 0.014272794127464294, + 0.009860253892838955, + 0.012591026723384857, + 0.010199416428804398, + 0.013106059283018112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.020596086978912354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002904266119003 + ], + "opposite_logprob_tokens": [ + -2.6950290203094482 + ], + "logprob_reward": -0.07002904266119003, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002904266119003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07922634482383728, + 0.023284679278731346, + 0.03944848105311394, + 0.10050541907548904, + 0.02922850102186203, + 0.32245662808418274, + 0.07666980475187302, + 0.032330844551324844, + 0.18510502576828003, + 0.11174427717924118 + ], + "credit_map": [ + 0.023810474202036858, + 0.0070261238142848015, + 0.014727325178682804, + 0.005500552244484425, + 0.006801940500736237, + 0.00646441662684083, + 0.00700182281434536, + 0.007893692702054977, + 0.002480638911947608, + 0.0025282923597842455, + 0.0033594046253710985, + 0.0027852461207658052, + 0.002596217440441251, + 0.0023778502363711596, + 0.003495765384286642, + 0.0036612653639167547, + 0.004347160458564758, + 0.005627690348774195, + 0.004845099989324808, + 0.003883874509483576, + 0.006259720306843519, + 0.004848932847380638, + 0.004459615796804428, + 0.005176388658583164, + 0.012538114562630653, + 0.00984092429280281, + 0.019121797755360603, + 0.007477771490812302, + 0.00865244772285223, + 0.01899269036948681, + 0.006332165561616421, + 0.017549514770507812, + 0.0035211106296628714, + 0.002474023262038827, + 0.004017828498035669, + 0.0031700804829597473, + 0.0023755389265716076, + 0.005828985944390297, + 0.0025973638985306025, + 0.005243570078164339, + 0.021435867995023727, + 0.0056009553372859955, + 0.0218221265822649, + 0.007650621235370636, + 0.008922837674617767, + 0.1309196799993515, + 0.011550625786185265, + 0.11455392092466354, + 0.005884486250579357, + 0.008329812437295914, + 0.009472093544900417, + 0.00937308743596077, + 0.008121497929096222, + 0.011103936471045017, + 0.012491786852478981, + 0.011893102899193764, + 0.0031505015213042498, + 0.004493118729442358, + 0.0036472666542977095, + 0.00465835677459836, + 0.0035371652338653803, + 0.004546206910163164, + 0.0034294782672077417, + 0.0048687513917684555, + 0.03550619259476662, + 0.02104705199599266, + 0.033918336033821106, + 0.01917934976518154, + 0.016913603991270065, + 0.02060088701546192, + 0.015857914462685585, + 0.022081702947616577, + 0.015646114945411682, + 0.016121523454785347, + 0.012016749009490013, + 0.014728917740285397, + 0.009389060549438, + 0.018500331789255142, + 0.010239433497190475, + 0.015102146193385124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.07002904266119003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015099564334377646 + ], + "opposite_logprob_tokens": [ + -6.501510143280029 + ], + "logprob_reward": -0.0015099564334377646, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015099564334377646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04916461184620857, + 0.11855212599039078, + 0.042417291551828384, + 0.17403458058834076, + 0.027074404060840607, + 0.4232729971408844, + 0.05409475415945053, + 0.04243738576769829, + 0.03594883158802986, + 0.033003069460392 + ], + "credit_map": [ + 0.007784927263855934, + 0.0043279374949634075, + 0.011428620666265488, + 0.005761769134551287, + 0.004734877962619066, + 0.0037573869340121746, + 0.006269423291087151, + 0.005099670961499214, + 0.017930425703525543, + 0.0059293536469340324, + 0.03736674785614014, + 0.009043578058481216, + 0.006155665963888168, + 0.012016716413199902, + 0.00874479953199625, + 0.021364841610193253, + 0.005122052039951086, + 0.005750966724008322, + 0.007717205677181482, + 0.004105071071535349, + 0.005664009600877762, + 0.005308488849550486, + 0.003923910669982433, + 0.004825584590435028, + 0.017715902999043465, + 0.01888561062514782, + 0.028697848320007324, + 0.01435727160423994, + 0.014774002134799957, + 0.03486202284693718, + 0.012935849837958813, + 0.03180607408285141, + 0.003318403149023652, + 0.0021479157730937004, + 0.004279329441487789, + 0.002916615689173341, + 0.0022856814321130514, + 0.005613834597170353, + 0.0024064702447503805, + 0.00410615187138319, + 0.027532141655683517, + 0.0045406105928123, + 0.02791685424745083, + 0.008268489502370358, + 0.010900026187300682, + 0.1745898723602295, + 0.015700360760092735, + 0.15382465720176697, + 0.004876012913882732, + 0.00676098745316267, + 0.006220356095582247, + 0.00567630073055625, + 0.008114711381494999, + 0.006607957649976015, + 0.00730226980522275, + 0.008536155335605145, + 0.007793846540153027, + 0.003381539136171341, + 0.0042240009643137455, + 0.003202470252290368, + 0.004729496780782938, + 0.010180847719311714, + 0.0027515392284840345, + 0.006173643749207258, + 0.0035248873755335808, + 0.0032522922847419977, + 0.009095313027501106, + 0.004457705654203892, + 0.0030589974485337734, + 0.003491774434223771, + 0.004409754648804665, + 0.004658107180148363, + 0.005251139402389526, + 0.004388065543025732, + 0.003800220787525177, + 0.0032737855799496174, + 0.003492301795631647, + 0.004131075460463762, + 0.00476866727694869, + 0.0038978152442723513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.0015099564334377646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233088731765747 + ], + "opposite_logprob_tokens": [ + -0.8983088731765747 + ], + "logprob_reward": -0.5233088731765747, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233088731765747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08980640769004822, + 0.16453337669372559, + 0.10032778978347778, + 0.0946415513753891, + 0.11333013325929642, + 0.18944139778614044, + 0.06468746066093445, + 0.035453684628009796, + 0.10556217283010483, + 0.04221605509519577 + ], + "credit_map": [ + 0.014579872600734234, + 0.007642048876732588, + 0.013607132248580456, + 0.01022854819893837, + 0.00633475836366415, + 0.01138025987893343, + 0.009488947689533234, + 0.016544828191399574, + 0.018488090485334396, + 0.014376300387084484, + 0.027017369866371155, + 0.02224302478134632, + 0.015929214656352997, + 0.013726525008678436, + 0.03069423884153366, + 0.022058596834540367, + 0.010140408761799335, + 0.012005859985947609, + 0.01294003613293171, + 0.010189748369157314, + 0.021252261474728584, + 0.01362176239490509, + 0.008417666889727116, + 0.011760043911635876, + 0.00907859206199646, + 0.008271398022770882, + 0.01568537950515747, + 0.011776532977819443, + 0.012477259151637554, + 0.01043673511594534, + 0.01449656207114458, + 0.012419083155691624, + 0.009592931717634201, + 0.008754848502576351, + 0.026828903704881668, + 0.009544267319142818, + 0.006950465962290764, + 0.022970059886574745, + 0.007540101185441017, + 0.021148549392819405, + 0.013856795616447926, + 0.010462623089551926, + 0.030498124659061432, + 0.009190876968204975, + 0.008515715599060059, + 0.06748788058757782, + 0.01190871186554432, + 0.037520673125982285, + 0.0041129677556455135, + 0.006692752707749605, + 0.006622313521802425, + 0.0068468572571873665, + 0.009497849270701408, + 0.008517308160662651, + 0.01133700180798769, + 0.011060407385230064, + 0.004731638822704554, + 0.003586472710594535, + 0.004300074186176062, + 0.004063749685883522, + 0.004337105434387922, + 0.005766306072473526, + 0.0040351771749556065, + 0.004633156582713127, + 0.013340220786631107, + 0.009639748372137547, + 0.014475604519248009, + 0.00815136544406414, + 0.011040267534554005, + 0.023848935961723328, + 0.009596118703484535, + 0.01546990405768156, + 0.006357255857437849, + 0.007611915003508329, + 0.0029225219041109085, + 0.006547538563609123, + 0.003985492046922445, + 0.005839918274432421, + 0.003712276229634881, + 0.0052391341887414455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.5233088731765747 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand, upward-looking rotating camera smoothly and majestically orbits a massive dragon, soaring above a boundless sea of clouds. The dragon's body is long and sinuous, covered in layers of crisscrossing, crisscrossing cyan scales, each one seemingly carved from the purest jade. Under the warm, golden light of the low sun, it gleams with an iridescent pearlescent sheen. It possesses towering antlers and a flowing, silver-white beard. Its eyes, blazing with golden light, radiate ancient wisdom and majesty. As it gracefully coils and unfurls, a pale white mane of pure energy undulates along its back. Surrounding it is a thick, fluffy sea of white clouds, tinged warmly by the golden sunlight. Each movement of the dragon carves gentle swirls and trails across the soft cloudtops. As the camera moves steadily 360 degrees, we can appreciate its huge and elegant figure from different angles. The light keeps flowing on its scales, sometimes bright as a mirror, sometimes deep as the sea. The whole picture is full of sacred, tranquil and awe-inspiring epic feeling.", + "seed": 15418, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_1/index_1171_seed_15418_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.266406367206946, + "vlm_logprob": -0.266406367206946, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.266406367206946, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8986086845397949 + ], + "opposite_logprob_tokens": [ + -0.5236086845397949 + ], + "logprob_reward": -0.8986086845397949, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8986086845397949, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.27974873781204224, + 0.025120224803686142, + 0.01558198593556881, + 0.013692280277609825, + 0.39694932103157043, + 0.03074938803911209, + 0.04389355704188347, + 0.0362081341445446, + 0.06448978185653687, + 0.09356656670570374 + ], + "credit_map": [ + 0.030084939673542976, + 0.03561822324991226, + 0.03094758838415146, + 0.02228998951613903, + 0.02649538964033127, + 0.04343395307660103, + 0.059596773236989975, + 0.031281884759664536, + 0.003571325447410345, + 0.0027374967467039824, + 0.003642924828454852, + 0.002141641452908516, + 0.0025275524239987135, + 0.003067953744903207, + 0.004883536137640476, + 0.0025477944873273373, + 0.0030009194742888212, + 0.0013093572342768312, + 0.0021717362105846405, + 0.001018643262796104, + 0.0013876029988750815, + 0.001752892043441534, + 0.0011371204163879156, + 0.003803713945671916, + 0.0020238838624209166, + 0.0012559079332277179, + 0.00296282977797091, + 0.0010431648697704077, + 0.001339204260148108, + 0.0014731893315911293, + 0.001372440718114376, + 0.0022216595243662596, + 0.0546802394092083, + 0.036067359149456024, + 0.06434720754623413, + 0.028784485533833504, + 0.043677762150764465, + 0.03770172595977783, + 0.0361001081764698, + 0.09559041261672974, + 0.0050243292935192585, + 0.0018967029172927141, + 0.007877081632614136, + 0.0020885898265987635, + 0.002019123639911413, + 0.003718304680660367, + 0.0019137101480737329, + 0.006211547646671534, + 0.010099305771291256, + 0.005404919385910034, + 0.005614610854536295, + 0.004861067980527878, + 0.003718071384355426, + 0.004245239309966564, + 0.004508963320404291, + 0.00544138066470623, + 0.004232414066791534, + 0.004220270551741123, + 0.004188386723399162, + 0.003945745062083006, + 0.005068210884928703, + 0.004748516250401735, + 0.0053275167010724545, + 0.004477072972804308, + 0.007681326009333134, + 0.009492593817412853, + 0.009386077523231506, + 0.007103240117430687, + 0.007585976738482714, + 0.008462836965918541, + 0.007997605949640274, + 0.006780119147151709, + 0.017756495624780655, + 0.007968884892761707, + 0.008953175507485867, + 0.008403168991208076, + 0.015124757774174213, + 0.01801859773695469, + 0.006214212626218796, + 0.011127274483442307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.8986086845397949 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078924760222435 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016078924760222435, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078924760222435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1286548227071762, + 0.05989385023713112, + 0.0897606760263443, + 0.0755816102027893, + 0.19222553074359894, + 0.06961967051029205, + 0.08727920800447464, + 0.11526399105787277, + 0.07274419069290161, + 0.10897640883922577 + ], + "credit_map": [ + 0.012515509501099586, + 0.023877311497926712, + 0.01491039153188467, + 0.014565429650247097, + 0.01446565706282854, + 0.015823351219296455, + 0.02014307864010334, + 0.012354097329080105, + 0.005608293693512678, + 0.005997930187731981, + 0.014732343144714832, + 0.005006014369428158, + 0.005692394450306892, + 0.005297352094203234, + 0.009930423460900784, + 0.007629100698977709, + 0.017024239525198936, + 0.008090019226074219, + 0.018010258674621582, + 0.005582133308053017, + 0.007597347721457481, + 0.007175093051046133, + 0.007438817527145147, + 0.018842769786715508, + 0.009363503195345402, + 0.006566637195646763, + 0.017549637705087662, + 0.005984289571642876, + 0.007266949862241745, + 0.008957453072071075, + 0.007196740247309208, + 0.012696395628154278, + 0.02320031449198723, + 0.014514967799186707, + 0.025570344179868698, + 0.014532564207911491, + 0.016880786046385765, + 0.024364542216062546, + 0.0191658865660429, + 0.053996141999959946, + 0.009899606928229332, + 0.004529482685029507, + 0.012174582108855247, + 0.005467616021633148, + 0.006537953391671181, + 0.008548344485461712, + 0.006872688885778189, + 0.015589398331940174, + 0.01489636767655611, + 0.01059294119477272, + 0.009356300346553326, + 0.007185813505202532, + 0.012236770242452621, + 0.013481366448104382, + 0.010249907150864601, + 0.009279741905629635, + 0.020729556679725647, + 0.011719627305865288, + 0.013439053669571877, + 0.009279766120016575, + 0.016506023705005646, + 0.016594234853982925, + 0.013969429768621922, + 0.013026301749050617, + 0.01164851151406765, + 0.009307712316513062, + 0.011800522916018963, + 0.006194601766765118, + 0.007743961177766323, + 0.011333328671753407, + 0.006863235495984554, + 0.007852320559322834, + 0.0264290701597929, + 0.01107647642493248, + 0.009322256781160831, + 0.009885800071060658, + 0.01660507544875145, + 0.016469907015562057, + 0.007786984089761972, + 0.011400837451219559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.016078924760222435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032142698764801 + ], + "opposite_logprob_tokens": [ + -2.3503215312957764 + ], + "logprob_reward": -0.10032142698764801, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032142698764801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47322413325309753, + 0.039091337472200394, + 0.04339832440018654, + 0.01499134674668312, + 0.11829295009374619, + 0.22180965542793274, + 0.011441715061664581, + 0.029128490015864372, + 0.02718278020620346, + 0.02143923006951809 + ], + "credit_map": [ + 0.050808656960725784, + 0.06167484447360039, + 0.05277901515364647, + 0.03732077404856682, + 0.04564850032329559, + 0.07365825027227402, + 0.09901586920022964, + 0.052318211644887924, + 0.005079685710370541, + 0.004224967211484909, + 0.006321843713521957, + 0.0034891576506197453, + 0.004542632959783077, + 0.004640176892280579, + 0.0067367954179644585, + 0.004056080710142851, + 0.010448670014739037, + 0.003174449549987912, + 0.006518106907606125, + 0.0021623880602419376, + 0.0020218202844262123, + 0.0029062400572001934, + 0.0037516953889280558, + 0.012414955534040928, + 0.0029172366484999657, + 0.0008473226916976273, + 0.003025110810995102, + 0.0008507468155585229, + 0.000983507721684873, + 0.0010555461049079895, + 0.0015691436128690839, + 0.0037427321076393127, + 0.017521066591143608, + 0.010714312084019184, + 0.019835103303194046, + 0.008347169496119022, + 0.012254707515239716, + 0.011278442107141018, + 0.010501498356461525, + 0.02784064970910549, + 0.03568471595644951, + 0.004719250835478306, + 0.07468578964471817, + 0.00876153539866209, + 0.005685111973434687, + 0.022656068205833435, + 0.006945160683244467, + 0.06267202645540237, + 0.002800716320052743, + 0.0014355994062498212, + 0.0013517345068976283, + 0.0009756987565197051, + 0.0011485520517453551, + 0.0011186591582372785, + 0.0010767506901174784, + 0.0015340031823143363, + 0.006527741905301809, + 0.003934596199542284, + 0.002909009577706456, + 0.0024396348744630814, + 0.0032545686699450016, + 0.00289219175465405, + 0.0033210208639502525, + 0.0038497254718095064, + 0.0037284018471837044, + 0.004083434585481882, + 0.003739664563909173, + 0.0025474203284829855, + 0.002962685190141201, + 0.0037522593047469854, + 0.0033444114960730076, + 0.0030245031230151653, + 0.004481294192373753, + 0.002328711561858654, + 0.0019750145729631186, + 0.0020278366282582283, + 0.003309044521301985, + 0.0035919896326959133, + 0.001561066834256053, + 0.002164273988455534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.10032142698764801 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036109976936131716 + ], + "opposite_logprob_tokens": [ + -5.628611087799072 + ], + "logprob_reward": -0.0036109976936131716, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036109976936131716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4566665589809418, + 0.051409702748060226, + 0.06251620501279831, + 0.016263095661997795, + 0.2026868611574173, + 0.10954982787370682, + 0.013173379004001617, + 0.03231637552380562, + 0.04143232852220535, + 0.013985654339194298 + ], + "credit_map": [ + 0.046847373247146606, + 0.06220507249236107, + 0.0492112934589386, + 0.03739340603351593, + 0.04318048804998398, + 0.06945168226957321, + 0.09701543301343918, + 0.0513618178665638, + 0.006599775515496731, + 0.0022058645263314247, + 0.008103043772280216, + 0.0024884752929210663, + 0.0026608770713210106, + 0.003065020078793168, + 0.017293503507971764, + 0.008993142284452915, + 0.01574363373219967, + 0.004622722510248423, + 0.010038395412266254, + 0.0027684629894793034, + 0.0021586953662335873, + 0.004669537767767906, + 0.005669739563018084, + 0.016845017671585083, + 0.002626738976687193, + 0.0010392828844487667, + 0.004198999144136906, + 0.0010180628160014749, + 0.0009180012275464833, + 0.001605343073606491, + 0.0016045199008658528, + 0.003252147464081645, + 0.028275780379772186, + 0.017357146367430687, + 0.03275270760059357, + 0.014449902810156345, + 0.0205629114061594, + 0.019840240478515625, + 0.01870555989444256, + 0.05074261501431465, + 0.018603075295686722, + 0.0026700953021645546, + 0.0365493968129158, + 0.0047195847146213055, + 0.0034688273444771767, + 0.01123177818953991, + 0.0033679651096463203, + 0.028939101845026016, + 0.00233161891810596, + 0.0015826846938580275, + 0.001605710480362177, + 0.0010771164670586586, + 0.0015094942646101117, + 0.0019402654143050313, + 0.0015983610646799207, + 0.00152812828309834, + 0.0065424093045294285, + 0.003516171593219042, + 0.004159761592745781, + 0.002295845188200474, + 0.003942226991057396, + 0.004245171323418617, + 0.003744114190340042, + 0.0038706737104803324, + 0.012034392915666103, + 0.00454285554587841, + 0.0041226488538086414, + 0.002776721492409706, + 0.004400814417749643, + 0.00451702531427145, + 0.0035568629391491413, + 0.0054810065776109695, + 0.0039018462412059307, + 0.0011205683695152402, + 0.0012271505547687411, + 0.0014351786812767386, + 0.0016891310224309564, + 0.00203904346562922, + 0.0009621639037504792, + 0.0016105714021250606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0036109976936131716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with towering antlers and a silver-white beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan and arranged in crisscrossing layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dragon's eyes glowing with golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless sea of white clouds tinged with warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth 360-degree orbit around the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341180205345154 + ], + "opposite_logprob_tokens": [ + -1.313411831855774 + ], + "logprob_reward": -0.31341180205345154, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341180205345154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13094763457775116, + 0.023989493027329445, + 0.01892145723104477, + 0.07939282804727554, + 0.45078718662261963, + 0.05118719860911369, + 0.03470741957426071, + 0.052333418279886246, + 0.09248322248458862, + 0.06525015085935593 + ], + "credit_map": [ + 0.015611630864441395, + 0.016421029344201088, + 0.015719374641776085, + 0.010363194160163403, + 0.013389989733695984, + 0.020518245175480843, + 0.02359168231487274, + 0.01533249020576477, + 0.004404873587191105, + 0.002350674942135811, + 0.003645261051133275, + 0.001983953407034278, + 0.002132641151547432, + 0.002619646256789565, + 0.004429486580193043, + 0.0024229567497968674, + 0.0026057937648147345, + 0.0020354469306766987, + 0.0024359223898500204, + 0.001358034205622971, + 0.0032040050718933344, + 0.0028734190855175257, + 0.001498436089605093, + 0.002910399343818426, + 0.01523467618972063, + 0.004508527461439371, + 0.018238049000501633, + 0.004234074614942074, + 0.005972872022539377, + 0.0060860952362418175, + 0.00729771563783288, + 0.0178208127617836, + 0.06127822771668434, + 0.0413065105676651, + 0.07188960909843445, + 0.03304964303970337, + 0.05024657025933266, + 0.04349296912550926, + 0.04118198901414871, + 0.10834167152643204, + 0.011931050568819046, + 0.0036449385806918144, + 0.009600075893104076, + 0.002869123825803399, + 0.005275736562907696, + 0.005418051034212112, + 0.00449880538508296, + 0.00794941745698452, + 0.006757252383977175, + 0.005391738843172789, + 0.004801148548722267, + 0.0035648925695568323, + 0.003097792388871312, + 0.0033274716697633266, + 0.004010382108390331, + 0.0037567398976534605, + 0.007354522589594126, + 0.007348204497247934, + 0.00709757162258029, + 0.00474271085113287, + 0.008251537568867207, + 0.006340664345771074, + 0.004955611191689968, + 0.006242597941309214, + 0.017858242616057396, + 0.010705976746976376, + 0.011154537089169025, + 0.007297975476831198, + 0.010524554178118706, + 0.01221530418843031, + 0.012102238833904266, + 0.010624400340020657, + 0.015570012852549553, + 0.005972640123218298, + 0.005694739520549774, + 0.005437989719212055, + 0.009759215638041496, + 0.010872796177864075, + 0.004933560732752085, + 0.007009197026491165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.31341180205345154 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_tensor([8509, 8509])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_tensor([8509, 8509])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5dbe89e2c12ee61f808ef72f27a5ff40b71f5794 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_tensor([8509, 8509])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium close-up captures the figure from the chest up as they deliberately don a crisp, light-blue cotton shirt. Set in a minimalist bedroom, the scene is bathed in the warm, golden light of early morning, filtered in from the side by an unseen window. As the figure slips their arm into the sleeve and places the shirt over their shoulder, the starched fabric rustles softly, satisfyingly. The strong sidelight delicately accentuates the fine weave of the material and illuminates motes dancing in the air. Their hands move down their torso, carefully and gently smoothing the fabric across their chest and abdomen, pressing out any slight creases. This focused, deliberate movement, combined with the soft, directional light that casts gentle shadows, creates an atmosphere of tranquility, intimacy, and serenity, a quiet preparation for the day ahead.", + "seed": 17523, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_8509_seed_17523_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0629954107105732, + "vlm_logprob": -0.0629954107105732, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.0629954107105732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1424035280942917 + ], + "opposite_logprob_tokens": [ + -2.1424036026000977 + ], + "logprob_reward": -0.1424035280942917, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1424035280942917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04491540044546127, + 0.03440648689866066, + 0.07130465656518936, + 0.044971201568841934, + 0.07662550359964371, + 0.10070997476577759, + 0.2626698613166809, + 0.1233428493142128, + 0.21185354888439178, + 0.029200518503785133 + ], + "credit_map": [ + 0.005041955038905144, + 0.005543674807995558, + 0.005713801831007004, + 0.003756885649636388, + 0.007359888404607773, + 0.007623965851962566, + 0.005011208821088076, + 0.004864021670073271, + 0.003941848408430815, + 0.003039315575733781, + 0.007499619387090206, + 0.0031798859126865864, + 0.004055945202708244, + 0.0036286944523453712, + 0.004003775771707296, + 0.005057401955127716, + 0.005241768900305033, + 0.006604223512113094, + 0.006847590208053589, + 0.0055931489914655685, + 0.017745763063430786, + 0.006562106776982546, + 0.013843084685504436, + 0.008866976015269756, + 0.0036276322789490223, + 0.005408531054854393, + 0.00405771704390645, + 0.004090613219887018, + 0.009393488056957722, + 0.007371822372078896, + 0.006218369118869305, + 0.004803028889000416, + 0.006198759190738201, + 0.010729454457759857, + 0.010781008750200272, + 0.008142802864313126, + 0.016922321170568466, + 0.008680091239511967, + 0.010098041966557503, + 0.0050730276852846146, + 0.005622441880404949, + 0.006199424620717764, + 0.01026193331927061, + 0.008937719278037548, + 0.012465634383261204, + 0.010125864297151566, + 0.02287629060447216, + 0.024220671504735947, + 0.022784944623708725, + 0.017942113801836967, + 0.013761433772742748, + 0.020706845447421074, + 0.0734081119298935, + 0.04289846867322922, + 0.04228214919567108, + 0.028885791078209877, + 0.010169290006160736, + 0.008350435644388199, + 0.01884775049984455, + 0.01565255969762802, + 0.012956158258020878, + 0.01387454941868782, + 0.017452297732234, + 0.026039814576506615, + 0.019765304401516914, + 0.013989794999361038, + 0.022310946136713028, + 0.018973231315612793, + 0.027195144444704056, + 0.02349073812365532, + 0.02639334835112095, + 0.059735044836997986, + 0.0027963619213551283, + 0.0018852045759558678, + 0.004410755820572376, + 0.005205376539379358, + 0.0025546124670654535, + 0.0027789140585809946, + 0.004751984961330891, + 0.004817306995391846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.1424035280942917 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10040251165628433 + ], + "opposite_logprob_tokens": [ + -2.350402593612671 + ], + "logprob_reward": -0.10040251165628433, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10040251165628433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06860598176717758, + 0.038537025451660156, + 0.07803253084421158, + 0.04571082442998886, + 0.03119312785565853, + 0.03555303066968918, + 0.2059064656496048, + 0.1768411099910736, + 0.06195961683988571, + 0.25766029953956604 + ], + "credit_map": [ + 0.006816087290644646, + 0.005508025176823139, + 0.0074429744854569435, + 0.005468457471579313, + 0.011907877400517464, + 0.014714346267282963, + 0.007678454741835594, + 0.009069752879440784, + 0.004009092692285776, + 0.003915283363312483, + 0.004547270480543375, + 0.0029918858781456947, + 0.00595359830185771, + 0.004796276334673166, + 0.0055297925136983395, + 0.006793823093175888, + 0.007104977499693632, + 0.008754904381930828, + 0.003988285548985004, + 0.0031706481240689754, + 0.023929569870233536, + 0.007501428946852684, + 0.014535143040120602, + 0.009047575294971466, + 0.0039340839721262455, + 0.0051254089921712875, + 0.005783085711300373, + 0.002898054663091898, + 0.006862656679004431, + 0.006051905453205109, + 0.005428570322692394, + 0.009627052582800388, + 0.002469123573973775, + 0.003790693823248148, + 0.004888969473540783, + 0.0021223104558885098, + 0.005503385327756405, + 0.0032305559143424034, + 0.006239711306989193, + 0.002948376815766096, + 0.0033940833527594805, + 0.004455206915736198, + 0.005224649794399738, + 0.0031813776586204767, + 0.006947459653019905, + 0.0037716603837907314, + 0.005196467507630587, + 0.003382124239578843, + 0.020265992730855942, + 0.014633222483098507, + 0.017791587859392166, + 0.015581128187477589, + 0.05307489633560181, + 0.03178098797798157, + 0.028136983513832092, + 0.024641644209623337, + 0.018482372164726257, + 0.016996033489704132, + 0.013237589038908482, + 0.019916333258152008, + 0.025171879678964615, + 0.02345367521047592, + 0.02441849187016487, + 0.035164736211299896, + 0.005192338488996029, + 0.0029012279119342566, + 0.006711388938128948, + 0.004601605236530304, + 0.006473577581346035, + 0.010798690840601921, + 0.010598254390060902, + 0.014682531356811523, + 0.02477392926812172, + 0.020301496610045433, + 0.015843085944652557, + 0.034878674894571304, + 0.037591706961393356, + 0.03575475141406059, + 0.04407884180545807, + 0.044437818229198456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.10040251165628433 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054974719882011414 + ], + "opposite_logprob_tokens": [ + -2.9299747943878174 + ], + "logprob_reward": -0.054974719882011414, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054974719882011414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11621762067079544, + 0.07352326810359955, + 0.07250884920358658, + 0.16378577053546906, + 0.12574777007102966, + 0.08256428688764572, + 0.15733443200588226, + 0.03541120886802673, + 0.06062488630414009, + 0.11228185147047043 + ], + "credit_map": [ + 0.013272264041006565, + 0.01027713529765606, + 0.02848602645099163, + 0.00989473331719637, + 0.009174642153084278, + 0.013105415739119053, + 0.014054388739168644, + 0.017953021451830864, + 0.007703494280576706, + 0.005233905278146267, + 0.01946130394935608, + 0.005847591441124678, + 0.007589597720652819, + 0.00906873494386673, + 0.009117563255131245, + 0.009501075372099876, + 0.009212463162839413, + 0.009291493333876133, + 0.006135878618806601, + 0.004285482689738274, + 0.014824219979345798, + 0.00661430275067687, + 0.012731975875794888, + 0.009413032792508602, + 0.01574704982340336, + 0.01615188643336296, + 0.021631011739373207, + 0.011706727556884289, + 0.025334781035780907, + 0.018039869144558907, + 0.03336499631404877, + 0.02180945873260498, + 0.010050613433122635, + 0.018535196781158447, + 0.02085101418197155, + 0.015442519448697567, + 0.023548981174826622, + 0.013314599171280861, + 0.01621301658451557, + 0.007791819050908089, + 0.007683708798140287, + 0.009389640763401985, + 0.013147194869816303, + 0.008613740094006062, + 0.011321119032800198, + 0.011000750586390495, + 0.009283746592700481, + 0.01212439127266407, + 0.01096660178154707, + 0.0095022888854146, + 0.008282825350761414, + 0.01177360862493515, + 0.05055712163448334, + 0.027276137843728065, + 0.023163579404354095, + 0.015812264755368233, + 0.003662734292447567, + 0.004163358360528946, + 0.0037281601689755917, + 0.003926504869014025, + 0.004491996951401234, + 0.0047162543050944805, + 0.005605537444353104, + 0.005116659682244062, + 0.0058691962622106075, + 0.004909530282020569, + 0.009604140184819698, + 0.0047370269894599915, + 0.005046203266829252, + 0.008619535714387894, + 0.0084474952891469, + 0.01339175645262003, + 0.00857995543628931, + 0.007845639251172543, + 0.027652045711874962, + 0.01802237331867218, + 0.009642066434025764, + 0.012052702717483044, + 0.013731221668422222, + 0.014755846932530403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.054974719882011414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549362145364285 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549362145364285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549362145364285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053235627710819244, + 0.03712581843137741, + 0.04774205759167671, + 0.049767982214689255, + 0.07318120449781418, + 0.11701449751853943, + 0.1864018589258194, + 0.21461373567581177, + 0.08554556965827942, + 0.13537165522575378 + ], + "credit_map": [ + 0.005883251782506704, + 0.004732189700007439, + 0.007413762155920267, + 0.006138899363577366, + 0.005620368290692568, + 0.007846426218748093, + 0.005676107481122017, + 0.009924624115228653, + 0.0032610795460641384, + 0.004407903645187616, + 0.00441029854118824, + 0.0033246709499508142, + 0.005571215879172087, + 0.004274759441614151, + 0.005513390526175499, + 0.006362497806549072, + 0.0050390856340527534, + 0.0076450081542134285, + 0.004073883406817913, + 0.0027107996866106987, + 0.010962148196995258, + 0.003400814952328801, + 0.006635603494942188, + 0.007274711038917303, + 0.0037104070652276278, + 0.005872723646461964, + 0.004559749271720648, + 0.006012023892253637, + 0.008847699500620365, + 0.00543993292376399, + 0.00639353459700942, + 0.00893191248178482, + 0.0056909737177193165, + 0.010889130644500256, + 0.007655399385839701, + 0.00664140610024333, + 0.01721162162721157, + 0.008537018671631813, + 0.01022813469171524, + 0.006327521987259388, + 0.009878914803266525, + 0.010844457894563675, + 0.01296814065426588, + 0.010828975588083267, + 0.012738389894366264, + 0.012631881982088089, + 0.023787762969732285, + 0.023335974663496017, + 0.01750629022717476, + 0.01267671212553978, + 0.010946580208837986, + 0.015449521131813526, + 0.04546979069709778, + 0.029869426041841507, + 0.031827475875616074, + 0.02265605516731739, + 0.012707971967756748, + 0.009995006956160069, + 0.017252668738365173, + 0.023314546793699265, + 0.034703634679317474, + 0.02954459749162197, + 0.02975078858435154, + 0.05734451860189438, + 0.007069161627441645, + 0.005384359508752823, + 0.006858862470835447, + 0.008689962327480316, + 0.010083438828587532, + 0.014403258450329304, + 0.013593485578894615, + 0.01946304552257061, + 0.011255458928644657, + 0.008041982538998127, + 0.011419146321713924, + 0.02015192061662674, + 0.019651472568511963, + 0.01786034367978573, + 0.019115692004561424, + 0.027875635772943497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.012549362145364285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646931774914265 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646931774914265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646931774914265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12052839249372482, + 0.05717400461435318, + 0.07191072404384613, + 0.057575762271881104, + 0.0892205610871315, + 0.08197067677974701, + 0.17726916074752808, + 0.06462384760379791, + 0.07984373718500137, + 0.19988317787647247 + ], + "credit_map": [ + 0.012640326283872128, + 0.011007127352058887, + 0.025816703215241432, + 0.012853243388235569, + 0.011794287711381912, + 0.016719045117497444, + 0.012114464305341244, + 0.017583172768354416, + 0.005223626270890236, + 0.005160776898264885, + 0.011595004238188267, + 0.005450521130114794, + 0.007557154167443514, + 0.007371732499450445, + 0.007311580237001181, + 0.007503601722419262, + 0.005928847938776016, + 0.008896767161786556, + 0.0066485111601650715, + 0.005451234057545662, + 0.012440246529877186, + 0.008840824477374554, + 0.009176730178296566, + 0.014527558349072933, + 0.004342495463788509, + 0.006277932319790125, + 0.006978320423513651, + 0.004853552673012018, + 0.013002727180719376, + 0.007904273457825184, + 0.008311024866998196, + 0.005905428435653448, + 0.007680343464016914, + 0.014860312454402447, + 0.011128094047307968, + 0.010559391230344772, + 0.01541983988136053, + 0.010090406984090805, + 0.011969671584665775, + 0.007512490730732679, + 0.013705683872103691, + 0.006461302749812603, + 0.01419930998235941, + 0.009638477116823196, + 0.011259722523391247, + 0.006933888420462608, + 0.01187085174024105, + 0.007901428267359734, + 0.012734994292259216, + 0.011710141785442829, + 0.010365964844822884, + 0.01406374853104353, + 0.05395010486245155, + 0.029275482520461082, + 0.026578962802886963, + 0.018589746206998825, + 0.0041065094992518425, + 0.005824824795126915, + 0.008311305195093155, + 0.00670635187998414, + 0.009243332780897617, + 0.010135295800864697, + 0.007179809268563986, + 0.013116414658725262, + 0.004860586952418089, + 0.0044724768958985806, + 0.011424785479903221, + 0.005708287935703993, + 0.006564101204276085, + 0.009033279493451118, + 0.014859279617667198, + 0.022920936346054077, + 0.013689645566046238, + 0.013261936604976654, + 0.055056266486644745, + 0.034243013709783554, + 0.012639344669878483, + 0.020408745855093002, + 0.023460056632757187, + 0.02712414041161537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.004646931774914265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures the figure from the chest up as they deliberately don a crisp, light-blue cotton shirt. Set in a minimalist bedroom, the scene is bathed in the warm, golden light of early morning, filtered in from the side by an unseen window. As the figure slips their arm into the sleeve and places the shirt over their shoulder, the starched fabric rustles softly, satisfyingly. The strong sidelight delicately accentuates the fine weave of the material and illuminates motes dancing in the air. Their hands move down their torso, carefully and gently smoothing the fabric across their chest and abdomen, pressing out any slight creases. This focused, deliberate movement, combined with the soft, directional light that casts gentle shadows, creates an atmosphere of tranquility, intimacy, and serenity, a quiet preparation for the day ahead.", + "seed": 17524, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_8509_seed_17524_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027486139582470058, + "vlm_logprob": -0.027486139582470058, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.027486139582470058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08002172410488129 + ], + "opposite_logprob_tokens": [ + -2.830021619796753 + ], + "logprob_reward": -0.08002172410488129, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08002172410488129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1779542714357376, + 0.04380645230412483, + 0.1507643312215805, + 0.017800742760300636, + 0.03764846920967102, + 0.45978662371635437, + 0.023765088990330696, + 0.018331024795770645, + 0.039668288081884384, + 0.030474765226244926 + ], + "credit_map": [ + 0.010423596017062664, + 0.044268228113651276, + 0.018932964652776718, + 0.0180483590811491, + 0.02407771907746792, + 0.019767185673117638, + 0.021478015929460526, + 0.020958183333277702, + 0.004876941908150911, + 0.004513686057180166, + 0.004739128518849611, + 0.0037991595454514027, + 0.005779328756034374, + 0.007594861090183258, + 0.006359159480780363, + 0.006144182290881872, + 0.006966808810830116, + 0.005594590213149786, + 0.00584823451936245, + 0.009562999941408634, + 0.010619197972118855, + 0.0532241053879261, + 0.0315871387720108, + 0.027361245825886726, + 0.0016186329303309321, + 0.0008514457731507719, + 0.0012426635948941112, + 0.0013120442163199186, + 0.003114180639386177, + 0.003151060314849019, + 0.0026691604871302843, + 0.0038415510207414627, + 0.0029661457519978285, + 0.005057457368820906, + 0.002954408060759306, + 0.003210405819118023, + 0.0040217055939137936, + 0.006437009200453758, + 0.005078558344393969, + 0.00792277604341507, + 0.03038012981414795, + 0.04012994468212128, + 0.031003816053271294, + 0.06138958781957626, + 0.078790582716465, + 0.08277685940265656, + 0.0818982869386673, + 0.0534173846244812, + 0.0017897184006869793, + 0.0028506203088909388, + 0.0021736870985478163, + 0.0031317290849983692, + 0.003576077753677964, + 0.0028623798862099648, + 0.004732158500701189, + 0.002648715628311038, + 0.0022757432889193296, + 0.0017620727885514498, + 0.0014977551763877273, + 0.001840636832639575, + 0.0033095472026616335, + 0.0020260484889149666, + 0.00326910731382668, + 0.002350113121792674, + 0.0023665421176701784, + 0.0034183149691671133, + 0.0020898582879453897, + 0.0031317833345383406, + 0.00506358128041029, + 0.006155614275485277, + 0.012419993057847023, + 0.005022597033530474, + 0.002080240985378623, + 0.002711792243644595, + 0.0019443496130406857, + 0.0013390540843829513, + 0.004611663985997438, + 0.008379846811294556, + 0.00551884900778532, + 0.0038889648858457804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.08002172410488129 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029861845076084137 + ], + "opposite_logprob_tokens": [ + -3.5298619270324707 + ], + "logprob_reward": -0.029861845076084137, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029861845076084137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15181347727775574, + 0.08598894625902176, + 0.04988381266593933, + 0.02384389378130436, + 0.055503156036138535, + 0.38910290598869324, + 0.05248647183179855, + 0.04791466146707535, + 0.08090995252132416, + 0.06255272030830383 + ], + "credit_map": [ + 0.008500497788190842, + 0.024126775562763214, + 0.021941879764199257, + 0.01170913502573967, + 0.016809027642011642, + 0.027042165398597717, + 0.01829511485993862, + 0.023388898000121117, + 0.006954874377697706, + 0.007938908413052559, + 0.0075210644863545895, + 0.006409755442291498, + 0.015190979465842247, + 0.01428666990250349, + 0.015000701881945133, + 0.012685999274253845, + 0.003851833753287792, + 0.0033744287211447954, + 0.003227578243240714, + 0.004194287583231926, + 0.00547925615683198, + 0.008129950612783432, + 0.009852920658886433, + 0.01177356019616127, + 0.003250035922974348, + 0.0014141100691631436, + 0.0022864951752126217, + 0.0018593877321109176, + 0.0035414155572652817, + 0.0043655396439135075, + 0.003348797094076872, + 0.003778114216402173, + 0.006891890428960323, + 0.004271761514246464, + 0.005921624600887299, + 0.0034614787437021732, + 0.014303934760391712, + 0.007238820660859346, + 0.00749677699059248, + 0.005916873924434185, + 0.024075614288449287, + 0.03370174765586853, + 0.025957824662327766, + 0.0550321601331234, + 0.06809103488922119, + 0.0707339271903038, + 0.0673876702785492, + 0.044122982770204544, + 0.003364365315064788, + 0.006082760635763407, + 0.004790465347468853, + 0.00679393345490098, + 0.010230052284896374, + 0.005805533844977617, + 0.010217693634331226, + 0.0052016740664839745, + 0.005381293129175901, + 0.005587142426520586, + 0.004678022116422653, + 0.005978582892566919, + 0.007097922265529633, + 0.004895580001175404, + 0.007979311980307102, + 0.006316811311990023, + 0.005016130395233631, + 0.00679825060069561, + 0.0038446024991571903, + 0.005500818602740765, + 0.013017741031944752, + 0.01192158367484808, + 0.021808311343193054, + 0.013002526946365833, + 0.0030489189084619284, + 0.0032284476328641176, + 0.0026168229524046183, + 0.002474847249686718, + 0.011550582014024258, + 0.01828604005277157, + 0.009870282374322414, + 0.011476779356598854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.029861845076084137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644304238259792 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644304238259792, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644304238259792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19035093486309052, + 0.08437974005937576, + 0.1494603008031845, + 0.05866363272070885, + 0.0775688886642456, + 0.10491375625133514, + 0.11251342296600342, + 0.08588296175003052, + 0.05710515007376671, + 0.0791611447930336 + ], + "credit_map": [ + 0.019085928797721863, + 0.013623178005218506, + 0.04485442489385605, + 0.013789213262498379, + 0.01643759198486805, + 0.021375054493546486, + 0.01308794692158699, + 0.048097603023052216, + 0.008586931973695755, + 0.006255148909986019, + 0.008500305935740471, + 0.00523257115855813, + 0.015840986743569374, + 0.012449895031750202, + 0.01460382528603077, + 0.012910072691738605, + 0.01108403317630291, + 0.010275594890117645, + 0.008864214643836021, + 0.009004256688058376, + 0.012440463528037071, + 0.044673871248960495, + 0.027876928448677063, + 0.0252409428358078, + 0.005958986468613148, + 0.004532020539045334, + 0.005949413403868675, + 0.004196096211671829, + 0.00815491285175085, + 0.012572385370731354, + 0.009292625822126865, + 0.00800719391554594, + 0.00628237659111619, + 0.008599615655839443, + 0.007989520207047462, + 0.006159495562314987, + 0.01834188960492611, + 0.0085353534668684, + 0.011917203664779663, + 0.009743438102304935, + 0.009723734110593796, + 0.015123516321182251, + 0.011200636625289917, + 0.02794780395925045, + 0.007769415620714426, + 0.012977516278624535, + 0.01079539954662323, + 0.0093757389113307, + 0.008970797061920166, + 0.012039519846439362, + 0.011001680046319962, + 0.01605195924639702, + 0.01745450496673584, + 0.01287909597158432, + 0.020947672426700592, + 0.013168183155357838, + 0.009357539005577564, + 0.008355728350579739, + 0.008079868741333485, + 0.007146112620830536, + 0.011350298300385475, + 0.010938542895019054, + 0.014163690619170666, + 0.016491184011101723, + 0.004547986201941967, + 0.007469881791621447, + 0.004213966429233551, + 0.0052180541679263115, + 0.008965065702795982, + 0.007832948118448257, + 0.010555668734014034, + 0.00830157846212387, + 0.005398797336965799, + 0.0070709348656237125, + 0.004465318750590086, + 0.004944442771375179, + 0.013245428912341595, + 0.02117721177637577, + 0.013316458091139793, + 0.009542553685605526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.008644304238259792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078457236289978 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.016078457236289978, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078457236289978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054439518600702286, + 0.04847901314496994, + 0.12177015095949173, + 0.03860848397016525, + 0.0407312735915184, + 0.09141720831394196, + 0.10299140959978104, + 0.042134568095207214, + 0.26617157459259033, + 0.19325684010982513 + ], + "credit_map": [ + 0.004720903467386961, + 0.006574305240064859, + 0.006399987265467644, + 0.00578122166916728, + 0.006357872858643532, + 0.011817962862551212, + 0.006060488987714052, + 0.0067267753183841705, + 0.00397340627387166, + 0.007848258130252361, + 0.0040167528204619884, + 0.0036483926232904196, + 0.007717766799032688, + 0.007671374827623367, + 0.00741450022906065, + 0.006188562605530024, + 0.004947746638208628, + 0.004168570041656494, + 0.004428145941346884, + 0.006699430290609598, + 0.008324326947331429, + 0.04442507028579712, + 0.024776499718427658, + 0.02400035224854946, + 0.003316759131848812, + 0.0016800117446109653, + 0.003220032900571823, + 0.0023106010630726814, + 0.0058600869961082935, + 0.008459877222776413, + 0.004526404198259115, + 0.009234712459146976, + 0.004922965541481972, + 0.005469538737088442, + 0.004679754376411438, + 0.003557417541742325, + 0.004780849441885948, + 0.0050287507474422455, + 0.004882004577666521, + 0.0074099949561059475, + 0.0075004869140684605, + 0.008587750606238842, + 0.006690863519906998, + 0.011843075975775719, + 0.015064508654177189, + 0.015318083576858044, + 0.016226833686232567, + 0.010185606777668, + 0.00847602728754282, + 0.012517227791249752, + 0.009274952113628387, + 0.010023741982877254, + 0.023447955027222633, + 0.011609643697738647, + 0.018319247290492058, + 0.00932261161506176, + 0.004181756172329187, + 0.00479862792417407, + 0.0036201272159814835, + 0.003712622681632638, + 0.006599462125450373, + 0.005782699678093195, + 0.00734183844178915, + 0.006097432225942612, + 0.015501842834055424, + 0.021865779533982277, + 0.012057163752615452, + 0.02392495423555374, + 0.0368313193321228, + 0.039338961243629456, + 0.07266915589570999, + 0.04398238658905029, + 0.0056565324775874615, + 0.0065871430560946465, + 0.0033030586782842875, + 0.004906758200377226, + 0.04690035805106163, + 0.0800975039601326, + 0.03412802144885063, + 0.011677454225718975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.016078457236289978 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002824367256835103 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.002824367256835103, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002824367256835103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18917135894298553, + 0.06018554046750069, + 0.07899720221757889, + 0.043093133717775345, + 0.13607560098171234, + 0.13970261812210083, + 0.1184685081243515, + 0.03439050167798996, + 0.07113231718540192, + 0.12878325581550598 + ], + "credit_map": [ + 0.013341246172785759, + 0.03147568181157112, + 0.03491000458598137, + 0.013718458823859692, + 0.01759975031018257, + 0.01832769624888897, + 0.01801377721130848, + 0.04178473353385925, + 0.006933532189577818, + 0.007203471846878529, + 0.005586143583059311, + 0.00416537607088685, + 0.011186575517058372, + 0.007810788229107857, + 0.009189708158373833, + 0.008109943009912968, + 0.0059434701688587666, + 0.0070552946999669075, + 0.004969276022166014, + 0.006071701645851135, + 0.008731327950954437, + 0.009544563479721546, + 0.0199180506169796, + 0.016763517633080482, + 0.0051308549009263515, + 0.004712102003395557, + 0.004626378882676363, + 0.003029881976544857, + 0.006760213989764452, + 0.006631676107645035, + 0.006717834621667862, + 0.00548419076949358, + 0.0159403495490551, + 0.011056391522288322, + 0.014955504797399044, + 0.00887422263622284, + 0.03673245385289192, + 0.014506700448691845, + 0.02035621926188469, + 0.013653756119310856, + 0.010139263235032558, + 0.016320204362273216, + 0.011418095789849758, + 0.02952301688492298, + 0.01880124770104885, + 0.01898047886788845, + 0.01892799511551857, + 0.01559231523424387, + 0.00897816102951765, + 0.014031426049768925, + 0.010623589158058167, + 0.013599821366369724, + 0.024766983464360237, + 0.013975990004837513, + 0.021582884714007378, + 0.010909659788012505, + 0.0042128232307732105, + 0.0038806614466011524, + 0.003397044027224183, + 0.0031472654081881046, + 0.006237495690584183, + 0.004057301674038172, + 0.004213510546833277, + 0.005244397558271885, + 0.005499576684087515, + 0.006797396577894688, + 0.006405130960047245, + 0.006972894538193941, + 0.011606399901211262, + 0.009076828137040138, + 0.013361004181206226, + 0.011413080617785454, + 0.00352933956310153, + 0.006997779477387667, + 0.003311475971713662, + 0.0043316674418747425, + 0.03126984089612961, + 0.0489203967154026, + 0.020774807780981064, + 0.009647935628890991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.002824367256835103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures the figure from the chest up as they deliberately don a crisp, light-blue cotton shirt. Set in a minimalist bedroom, the scene is bathed in the warm, golden light of early morning, filtered in from the side by an unseen window. As the figure slips their arm into the sleeve and places the shirt over their shoulder, the starched fabric rustles softly, satisfyingly. The strong sidelight delicately accentuates the fine weave of the material and illuminates motes dancing in the air. Their hands move down their torso, carefully and gently smoothing the fabric across their chest and abdomen, pressing out any slight creases. This focused, deliberate movement, combined with the soft, directional light that casts gentle shadows, creates an atmosphere of tranquility, intimacy, and serenity, a quiet preparation for the day ahead.", + "seed": 17525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_8509_seed_17525_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.057533306535333395, + "vlm_logprob": -0.057533306535333395, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.057533306535333395, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19514502584934235 + ], + "opposite_logprob_tokens": [ + -1.8201450109481812 + ], + "logprob_reward": -0.19514502584934235, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19514502584934235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10358511656522751, + 0.104364313185215, + 0.0251533892005682, + 0.1434182971715927, + 0.07255963981151581, + 0.045166417956352234, + 0.16049496829509735, + 0.11302569508552551, + 0.11358074098825455, + 0.11865140497684479 + ], + "credit_map": [ + 0.020725658163428307, + 0.010304374620318413, + 0.01248486153781414, + 0.012152706272900105, + 0.011672777123749256, + 0.009158714674413204, + 0.014117962680757046, + 0.012968061491847038, + 0.013218160718679428, + 0.011048845946788788, + 0.007627943530678749, + 0.009376109577715397, + 0.03614320978522301, + 0.010113307274878025, + 0.009228017181158066, + 0.007608715910464525, + 0.001970506040379405, + 0.00295410118997097, + 0.0016712615033611655, + 0.002548957010731101, + 0.004262582398951054, + 0.004541868809610605, + 0.003066543722525239, + 0.004137568175792694, + 0.00778234051540494, + 0.009297676384449005, + 0.007291921880096197, + 0.007996953092515469, + 0.014295383356511593, + 0.03734961524605751, + 0.01791347563266754, + 0.04149091616272926, + 0.003493456868454814, + 0.0035748989321291447, + 0.004027107264846563, + 0.003281522309407592, + 0.027467072010040283, + 0.011631669476628304, + 0.013013424351811409, + 0.006070483010262251, + 0.00260109337978065, + 0.004129122011363506, + 0.00537695549428463, + 0.003956668544560671, + 0.008125567808747292, + 0.007719789631664753, + 0.006757836788892746, + 0.006499384064227343, + 0.018039030954241753, + 0.014272038824856281, + 0.017453782260417938, + 0.015518482774496078, + 0.019427599385380745, + 0.02310100942850113, + 0.020932067185640335, + 0.03175097331404686, + 0.005473548546433449, + 0.007463337853550911, + 0.005915447138249874, + 0.006607804913073778, + 0.035952091217041016, + 0.02640545554459095, + 0.010173480957746506, + 0.01503453217446804, + 0.012443150393664837, + 0.013006658293306828, + 0.012028600089251995, + 0.017238538712263107, + 0.01659417897462845, + 0.01566338911652565, + 0.013669583015143871, + 0.012936647050082684, + 0.010952086187899113, + 0.009917653165757656, + 0.008865063078701496, + 0.007624886464327574, + 0.011136245913803577, + 0.022108281031250954, + 0.016128696501255035, + 0.03191849961876869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.19514502584934235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062102507799863815 + ], + "opposite_logprob_tokens": [ + -2.8121025562286377 + ], + "logprob_reward": -0.062102507799863815, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062102507799863815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0771486833691597, + 0.30901777744293213, + 0.033533401787281036, + 0.0591532364487648, + 0.08659665286540985, + 0.02166593447327614, + 0.051666777580976486, + 0.1714644581079483, + 0.05060943216085434, + 0.13914361596107483 + ], + "credit_map": [ + 0.010816543363034725, + 0.006642226129770279, + 0.006344923749566078, + 0.011179386638104916, + 0.010082559660077095, + 0.010663759894669056, + 0.011334949173033237, + 0.010084333829581738, + 0.02030295878648758, + 0.03625534847378731, + 0.0638110563158989, + 0.07509376108646393, + 0.02905925363302231, + 0.031230224296450615, + 0.02631847932934761, + 0.026946688070893288, + 0.0023201049771159887, + 0.004707439336925745, + 0.0023758779279887676, + 0.0025752538349479437, + 0.008370185270905495, + 0.005096340551972389, + 0.0031948985997587442, + 0.004893300589174032, + 0.0028521420899778605, + 0.00325627438724041, + 0.002125480445101857, + 0.0038994266651570797, + 0.007638902403414249, + 0.015206864103674889, + 0.007592226844280958, + 0.01658191904425621, + 0.00491212448105216, + 0.005600957665592432, + 0.006486599799245596, + 0.004197284113615751, + 0.02625214122235775, + 0.013693340122699738, + 0.014515276066958904, + 0.010938931256532669, + 0.0017136131646111608, + 0.0016006353544071317, + 0.002751626307144761, + 0.002232710598036647, + 0.002521730028092861, + 0.004071188159286976, + 0.0030430196784436703, + 0.0037314118817448616, + 0.0038208882324397564, + 0.005284961778670549, + 0.006081436760723591, + 0.006336707156151533, + 0.009511887095868587, + 0.0077011603862047195, + 0.005880994256585836, + 0.007048744708299637, + 0.008266562595963478, + 0.009895475581288338, + 0.00771317258477211, + 0.010530488565564156, + 0.05971735715866089, + 0.04284341260790825, + 0.013444321230053902, + 0.01905367337167263, + 0.00370037741959095, + 0.0051283021457493305, + 0.0037936419248580933, + 0.0051289028488099575, + 0.009432719089090824, + 0.009468460455536842, + 0.005118373781442642, + 0.008838656358420849, + 0.010996644385159016, + 0.009594195522367954, + 0.008646627888083458, + 0.008455507457256317, + 0.011794663034379482, + 0.029320845380425453, + 0.014362637884914875, + 0.045972488820552826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.062102507799863815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076326169073581696 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.0076326169073581696, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076326169073581696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1581479161977768, + 0.2845989465713501, + 0.10803396999835968, + 0.06184014678001404, + 0.0673375055193901, + 0.057590849697589874, + 0.04208143427968025, + 0.12979604303836823, + 0.0381508506834507, + 0.05242229625582695 + ], + "credit_map": [ + 0.017505165189504623, + 0.019881227985024452, + 0.015764741227030754, + 0.020340776070952415, + 0.022237520664930344, + 0.02133553847670555, + 0.024208886548876762, + 0.01687406376004219, + 0.02110712230205536, + 0.03171930089592934, + 0.046034201979637146, + 0.06669339537620544, + 0.041321057826280594, + 0.029100222513079643, + 0.023302046582102776, + 0.025321636348962784, + 0.01094876043498516, + 0.019809680059552193, + 0.018923791125416756, + 0.019078725948929787, + 0.008084679953753948, + 0.0143404146656394, + 0.007864519953727722, + 0.008983395993709564, + 0.004669727757573128, + 0.005376824643462896, + 0.006762370001524687, + 0.009705102071166039, + 0.007862369529902935, + 0.009820862673223019, + 0.007375848479568958, + 0.010267042554914951, + 0.0038785680662840605, + 0.0032700425945222378, + 0.002945652697235346, + 0.0022293690126389265, + 0.023731622844934464, + 0.011892283335328102, + 0.011853800155222416, + 0.007536166813224554, + 0.004617337603121996, + 0.006495616864413023, + 0.007041711360216141, + 0.006770486012101173, + 0.009152049198746681, + 0.010102461092174053, + 0.006196391303092241, + 0.007214796729385853, + 0.003963013645261526, + 0.0025370323564857244, + 0.004940716549754143, + 0.003952374216169119, + 0.005007602274417877, + 0.006015278864651918, + 0.00536237470805645, + 0.010303042829036713, + 0.009374749846756458, + 0.015428595244884491, + 0.007367231417447329, + 0.010483317077159882, + 0.028367498889565468, + 0.026718316599726677, + 0.01656246930360794, + 0.015493867918848991, + 0.00459880568087101, + 0.00420541362836957, + 0.006306730210781097, + 0.004421878140419722, + 0.004861474968492985, + 0.004486448131501675, + 0.003824417246505618, + 0.0054456815123558044, + 0.005368315614759922, + 0.005095497239381075, + 0.005545984022319317, + 0.004456233233213425, + 0.004525765776634216, + 0.008441362529993057, + 0.005669896025210619, + 0.013319244608283043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0076326169073581696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06660329550504684, + 0.18926887214183807, + 0.035225819796323776, + 0.0452796071767807, + 0.05422548949718475, + 0.09408381581306458, + 0.19230583310127258, + 0.0947495847940445, + 0.10452031344175339, + 0.12373729795217514 + ], + "credit_map": [ + 0.009418843314051628, + 0.0070090522058308125, + 0.0069615477696061134, + 0.009818263351917267, + 0.009340435266494751, + 0.0065317959524691105, + 0.009443406946957111, + 0.008079946041107178, + 0.014345131814479828, + 0.020609784871339798, + 0.03102056309580803, + 0.03965580090880394, + 0.03917602077126503, + 0.015841592103242874, + 0.014977666549384594, + 0.013642310164868832, + 0.0022664836142212152, + 0.003721678163856268, + 0.002910755341872573, + 0.004238458815962076, + 0.006018494721502066, + 0.005359064321964979, + 0.005105157848447561, + 0.005605724640190601, + 0.0034552121069282293, + 0.00410287082195282, + 0.003401511814445257, + 0.004517673049122095, + 0.005188182462006807, + 0.009336395189166069, + 0.0044757407158613205, + 0.010802021250128746, + 0.0034836449194699526, + 0.00353867351077497, + 0.00473744235932827, + 0.003451570635661483, + 0.015277068130671978, + 0.00880781002342701, + 0.007906915619969368, + 0.007022366859018803, + 0.0053613148629665375, + 0.007608402520418167, + 0.01009715348482132, + 0.005896680522710085, + 0.01720871403813362, + 0.024182619526982307, + 0.010556979104876518, + 0.013171946629881859, + 0.015346693806350231, + 0.022713838145136833, + 0.01815986819565296, + 0.01844402588903904, + 0.03887759894132614, + 0.03263036906719208, + 0.02423223853111267, + 0.021901195868849754, + 0.0074140517972409725, + 0.0058788456954061985, + 0.0046922555193305016, + 0.004994097165763378, + 0.026272002607584, + 0.019508862867951393, + 0.012750170193612576, + 0.01323929987847805, + 0.01105445995926857, + 0.011581401340663433, + 0.010202187113463879, + 0.011084990575909615, + 0.017429398372769356, + 0.013760991394519806, + 0.015492637641727924, + 0.013914247043430805, + 0.008133775554597378, + 0.007728425785899162, + 0.008641033433377743, + 0.00749622005969286, + 0.012062140740454197, + 0.026830632239580154, + 0.012743485160171986, + 0.04010159522294998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.02058546058833599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200921531766653 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.002200921531766653, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200921531766653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17369577288627625, + 0.08647964149713516, + 0.04995008558034897, + 0.08331682533025742, + 0.1839657574892044, + 0.032866835594177246, + 0.09845048934221268, + 0.15883587300777435, + 0.07110735774040222, + 0.061331287026405334 + ], + "credit_map": [ + 0.023496409878134727, + 0.01918518729507923, + 0.01617424562573433, + 0.026272816583514214, + 0.02172752656042576, + 0.02412368729710579, + 0.02039225585758686, + 0.022323647513985634, + 0.010822275653481483, + 0.009418192319571972, + 0.007785479072481394, + 0.008444669656455517, + 0.03021286055445671, + 0.00785106886178255, + 0.007366201374679804, + 0.004578894004225731, + 0.005292579997330904, + 0.010149194858968258, + 0.005834511946886778, + 0.006008852273225784, + 0.008164276368916035, + 0.006534752901643515, + 0.003218825673684478, + 0.004747090395539999, + 0.006407209672033787, + 0.006864253897219896, + 0.008150750771164894, + 0.006699331104755402, + 0.00927149597555399, + 0.015907160937786102, + 0.010826751589775085, + 0.019189877435564995, + 0.011949588544666767, + 0.00850591715425253, + 0.019347861409187317, + 0.011787465773522854, + 0.06306041777133942, + 0.02706107497215271, + 0.028298087418079376, + 0.013955341652035713, + 0.002655648859217763, + 0.0032934993505477905, + 0.004604915156960487, + 0.003436266677454114, + 0.005744489375501871, + 0.004265045281499624, + 0.004641596227884293, + 0.004225374665111303, + 0.009225621819496155, + 0.011027786880731583, + 0.009546704590320587, + 0.011335479095578194, + 0.01235981099307537, + 0.013399983756244183, + 0.013128205202519894, + 0.018426895141601562, + 0.009922485798597336, + 0.011230605654418468, + 0.009045337326824665, + 0.009121439419686794, + 0.052696265280246735, + 0.036427758634090424, + 0.013376107439398766, + 0.017015883699059486, + 0.00992319080978632, + 0.008229917846620083, + 0.010978803969919682, + 0.007627381477504969, + 0.008672048337757587, + 0.010129350237548351, + 0.006555136293172836, + 0.008991527371108532, + 0.007706977892667055, + 0.004503275267779827, + 0.00849082786589861, + 0.0054833125323057175, + 0.003929360304027796, + 0.010479896329343319, + 0.005672696977853775, + 0.01506494078785181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.002200921531766653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures the figure from the chest up as they deliberately don a crisp, light-blue cotton shirt. Set in a minimalist bedroom, the scene is bathed in the warm, golden light of early morning, filtered in from the side by an unseen window. As the figure slips their arm into the sleeve and places the shirt over their shoulder, the starched fabric rustles softly, satisfyingly. The strong sidelight delicately accentuates the fine weave of the material and illuminates motes dancing in the air. Their hands move down their torso, carefully and gently smoothing the fabric across their chest and abdomen, pressing out any slight creases. This focused, deliberate movement, combined with the soft, directional light that casts gentle shadows, creates an atmosphere of tranquility, intimacy, and serenity, a quiet preparation for the day ahead.", + "seed": 17526, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_10/index_8509_seed_17526_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0470038334839046, + "vlm_logprob": -0.0470038334839046, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.0470038334839046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1468980759382248 + ], + "opposite_logprob_tokens": [ + -2.146898031234741 + ], + "logprob_reward": -0.1468980759382248, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1468980759382248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13853321969509125, + 0.17428171634674072, + 0.14815984666347504, + 0.1382894068956375, + 0.12937328219413757, + 0.054273784160614014, + 0.029279367998242378, + 0.042102232575416565, + 0.0675857663154602, + 0.0781213566660881 + ], + "credit_map": [ + 0.013845445588231087, + 0.01567537523806095, + 0.016353392973542213, + 0.022624721750617027, + 0.012861872091889381, + 0.02094632014632225, + 0.015617924742400646, + 0.020608164370059967, + 0.012960469350218773, + 0.027131838724017143, + 0.023592276498675346, + 0.02097289264202118, + 0.01896275207400322, + 0.023841436952352524, + 0.02032480761408806, + 0.02649524435400963, + 0.015762189403176308, + 0.016942568123340607, + 0.023453878238797188, + 0.025814203545451164, + 0.01852862723171711, + 0.01648702472448349, + 0.015207148157060146, + 0.015964210033416748, + 0.01742798089981079, + 0.00880370568484068, + 0.011870197020471096, + 0.013744178228080273, + 0.01530090905725956, + 0.031214885413646698, + 0.019406165927648544, + 0.020521383732557297, + 0.01476847380399704, + 0.013952517881989479, + 0.015127847902476788, + 0.01188214123249054, + 0.018698792904615402, + 0.019147738814353943, + 0.015645386651158333, + 0.020150374621152878, + 0.00823331531137228, + 0.003991073928773403, + 0.005275818053632975, + 0.004373549018055201, + 0.008137830533087254, + 0.008098841644823551, + 0.006013748701661825, + 0.010149604640901089, + 0.0030682249926030636, + 0.002819197252392769, + 0.003479788778349757, + 0.005303787067532539, + 0.0031725524459034204, + 0.004125821404159069, + 0.004375371616333723, + 0.0029346239753067493, + 0.003411650424823165, + 0.00343515002168715, + 0.004677874967455864, + 0.005530316848307848, + 0.006315878592431545, + 0.005845985841006041, + 0.007986940443515778, + 0.004898433107882738, + 0.004867248702794313, + 0.005589187145233154, + 0.007395383436232805, + 0.012767118401825428, + 0.006793103646486998, + 0.008349472656846046, + 0.015877213329076767, + 0.00594703946262598, + 0.007370148319751024, + 0.0038179210387170315, + 0.007263312116265297, + 0.007194723468273878, + 0.011666345410048962, + 0.015707997605204582, + 0.016512563452124596, + 0.008588344790041447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.1468980759382248 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501296743750572 + ], + "opposite_logprob_tokens": [ + -2.9300129413604736 + ], + "logprob_reward": -0.05501296743750572, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501296743750572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1549593210220337, + 0.09975291788578033, + 0.1403350681066513, + 0.1335163712501526, + 0.06789393723011017, + 0.10879473388195038, + 0.03431893140077591, + 0.1285945177078247, + 0.05780477821826935, + 0.07402949035167694 + ], + "credit_map": [ + 0.011535517871379852, + 0.012185596860945225, + 0.012885970994830132, + 0.014441626146435738, + 0.019505703821778297, + 0.028153706341981888, + 0.03632247447967529, + 0.019928714260458946, + 0.008730239234864712, + 0.018178772181272507, + 0.015510640107095242, + 0.01707272231578827, + 0.011194135062396526, + 0.010349559597671032, + 0.007536775432527065, + 0.011180064640939236, + 0.009679972194135189, + 0.011339119635522366, + 0.010682182386517525, + 0.02488044463098049, + 0.01861659437417984, + 0.02074604108929634, + 0.02689887396991253, + 0.017491843551397324, + 0.0197139885276556, + 0.008866842836141586, + 0.011036340147256851, + 0.01241605170071125, + 0.011980459094047546, + 0.031950514763593674, + 0.01586870476603508, + 0.021683460101485252, + 0.006598388310521841, + 0.008838814683258533, + 0.006412692368030548, + 0.00786383543163538, + 0.00792378094047308, + 0.011522499844431877, + 0.010427099652588367, + 0.00830682273954153, + 0.014701832085847855, + 0.011257153935730457, + 0.010096144862473011, + 0.016432134434580803, + 0.011909532360732555, + 0.01386974286288023, + 0.01266440749168396, + 0.017863769084215164, + 0.0039826068095862865, + 0.003423727350309491, + 0.003134822938591242, + 0.005859933327883482, + 0.0053583793342113495, + 0.00525888754054904, + 0.004154481925070286, + 0.0031460891477763653, + 0.008308321237564087, + 0.015271900221705437, + 0.00888372678309679, + 0.01587587594985962, + 0.023332828655838966, + 0.019600655883550644, + 0.01948701962828636, + 0.017834166064858437, + 0.00477816304191947, + 0.007452131249010563, + 0.005061820149421692, + 0.007265930995345116, + 0.007736291736364365, + 0.0076393866911530495, + 0.011392846703529358, + 0.006478207651525736, + 0.007222793530672789, + 0.004710440058261156, + 0.006135025527328253, + 0.009712872095406055, + 0.010383167304098606, + 0.00904172845184803, + 0.01657634787261486, + 0.01024710200726986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.05501296743750572 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206582680344582 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206582680344582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206582680344582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3350500166416168, + 0.04846680909395218, + 0.08162302523851395, + 0.2564031779766083, + 0.06738714873790741, + 0.04343107342720032, + 0.013996993191540241, + 0.07177029550075531, + 0.027732491493225098, + 0.054138947278261185 + ], + "credit_map": [ + 0.0146942725405097, + 0.020252853631973267, + 0.019832853227853775, + 0.021628497168421745, + 0.03998275473713875, + 0.07121865451335907, + 0.09281400591135025, + 0.05462613329291344, + 0.004459724295884371, + 0.007337109185755253, + 0.005639644339680672, + 0.009821583516895771, + 0.005869907792657614, + 0.004880182445049286, + 0.005019559524953365, + 0.0054391007870435715, + 0.006164884660393, + 0.00727153429761529, + 0.008095900528132915, + 0.010305657982826233, + 0.009786222130060196, + 0.015636658295989037, + 0.014695405028760433, + 0.00966676790267229, + 0.03107897937297821, + 0.02449086681008339, + 0.043870773166418076, + 0.031841009855270386, + 0.03308434039354324, + 0.032888468354940414, + 0.029098914936184883, + 0.030049800872802734, + 0.006055804435163736, + 0.0062139299698174, + 0.007218550890684128, + 0.006116581615060568, + 0.008235620334744453, + 0.009426736272871494, + 0.011769180186092854, + 0.012350738979876041, + 0.007734275888651609, + 0.0034497312735766172, + 0.005830252077430487, + 0.004227181896567345, + 0.004204802680760622, + 0.006418758537620306, + 0.004664510488510132, + 0.006901560816913843, + 0.000992368906736374, + 0.0013153874315321445, + 0.0013331410009413958, + 0.0027193648274987936, + 0.0011005096603184938, + 0.002340194070711732, + 0.0018642069771885872, + 0.002331820083782077, + 0.008440644480288029, + 0.008171677589416504, + 0.007994519546627998, + 0.01100531779229641, + 0.010241863317787647, + 0.009457387961447239, + 0.008998584933578968, + 0.007460302673280239, + 0.0019047219539061189, + 0.00298004481010139, + 0.002586888847872615, + 0.003842582693323493, + 0.0037120922934263945, + 0.0037750541232526302, + 0.006526229437440634, + 0.0024048786144703627, + 0.006229490041732788, + 0.003076218068599701, + 0.006275529973208904, + 0.00494461739435792, + 0.011604106053709984, + 0.011414680629968643, + 0.0070342496037483215, + 0.0035600594710558653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.014206582680344582 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076579689979553 + ], + "opposite_logprob_tokens": [ + -4.141076564788818 + ], + "logprob_reward": -0.016076579689979553, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076579689979553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29686468839645386, + 0.10788146406412125, + 0.058819763362407684, + 0.0822133794426918, + 0.04652081057429314, + 0.062184888869524, + 0.035833317786455154, + 0.05218680202960968, + 0.15246455371379852, + 0.10503033548593521 + ], + "credit_map": [ + 0.025010230019688606, + 0.03359563276171684, + 0.031095143407583237, + 0.038158223032951355, + 0.03649260848760605, + 0.04327112063765526, + 0.05352519080042839, + 0.035716522485017776, + 0.009497001767158508, + 0.024306058883666992, + 0.010024528950452805, + 0.01450627762824297, + 0.012553244829177856, + 0.011411840096116066, + 0.015464129857718945, + 0.010118382051587105, + 0.007386122364550829, + 0.004895446822047234, + 0.008564868941903114, + 0.007340355776250362, + 0.006213196087628603, + 0.006923438981175423, + 0.009226804599165916, + 0.008269529789686203, + 0.0068948762491345406, + 0.005795700009912252, + 0.009244916960597038, + 0.007953507825732231, + 0.016094237565994263, + 0.012262583710253239, + 0.011329608969390392, + 0.012637940235435963, + 0.005568887107074261, + 0.005003944039344788, + 0.004805444739758968, + 0.004484068136662245, + 0.006569637451320887, + 0.006011431571096182, + 0.006921157706528902, + 0.00715623889118433, + 0.00951052363961935, + 0.006025408860296011, + 0.006283808499574661, + 0.008635776117444038, + 0.00742596248164773, + 0.007972434163093567, + 0.006897917483001947, + 0.009433051571249962, + 0.006516861729323864, + 0.0032273216638714075, + 0.004227644298225641, + 0.005903757642954588, + 0.00361477374099195, + 0.00534615945070982, + 0.003755256999284029, + 0.0032415450550615788, + 0.0064729549922049046, + 0.006054850295186043, + 0.006496603135019541, + 0.007152367848902941, + 0.005900015588849783, + 0.006703768391162157, + 0.008864343166351318, + 0.004541896283626556, + 0.01076483353972435, + 0.010059897787868977, + 0.015060348436236382, + 0.022136112675070763, + 0.018169205635786057, + 0.023685218766331673, + 0.04270624369382858, + 0.009882702492177486, + 0.010894791223108768, + 0.0088705625385046, + 0.00882036704570055, + 0.009429492987692356, + 0.014390077441930771, + 0.014919404871761799, + 0.0261306744068861, + 0.011574963107705116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.016076579689979553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up showing the figure from the chest up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing or putting on a crisp, light-blue cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a minimalist bedroom with warm, golden morning light filtering from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure perform the specific actions of slipping an arm into a sleeve, placing the shirt over their shoulder, and smoothing the fabric across their chest and abdomen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting strong and directional (sidelight) that accentuates the fabric texture and illuminates motes dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028249616734683514 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028249616734683514, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028249616734683514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3184155225753784, + 0.06272656470537186, + 0.07273290306329727, + 0.10454850643873215, + 0.1140521764755249, + 0.0411263071000576, + 0.028925837948918343, + 0.07129047811031342, + 0.019171522930264473, + 0.16701015830039978 + ], + "credit_map": [ + 0.02502124011516571, + 0.03223545476794243, + 0.03131459653377533, + 0.03963432088494301, + 0.033785682171583176, + 0.05213722959160805, + 0.06455383449792862, + 0.0397331640124321, + 0.006024952046573162, + 0.008227742277085781, + 0.00974472239613533, + 0.009243288077414036, + 0.008376086130738258, + 0.0071481564082205296, + 0.007105060387402773, + 0.0068565597757697105, + 0.00561944954097271, + 0.008486953563988209, + 0.009835426695644855, + 0.016845282167196274, + 0.009830666705965996, + 0.007312120404094458, + 0.006952803581953049, + 0.00785020086914301, + 0.017406437546014786, + 0.007565265521407127, + 0.01643722504377365, + 0.011698908172547817, + 0.012184232473373413, + 0.0142365712672472, + 0.013012049719691277, + 0.012007823213934898, + 0.015541129745543003, + 0.010335046797990799, + 0.013888788409531116, + 0.011208056472241879, + 0.01195661909878254, + 0.017445463687181473, + 0.01569647714495659, + 0.017980601638555527, + 0.00896458700299263, + 0.0029491803143173456, + 0.005822611507028341, + 0.005176500417292118, + 0.004402569495141506, + 0.005670512095093727, + 0.0036834441125392914, + 0.004456901457160711, + 0.005491942632943392, + 0.003079324960708618, + 0.0030584088526666164, + 0.004118825308978558, + 0.002470726612955332, + 0.004200482740998268, + 0.003253173315897584, + 0.003252956550568342, + 0.006978699006140232, + 0.006664265878498554, + 0.007780843880027533, + 0.014009754173457623, + 0.012146984227001667, + 0.01056537963449955, + 0.008204275742173195, + 0.0049402788281440735, + 0.0016854314599186182, + 0.0013966246042400599, + 0.0019096164032816887, + 0.0021654507145285606, + 0.0025590008590370417, + 0.002762965392321348, + 0.004916887264698744, + 0.0017755477456375957, + 0.017034564167261124, + 0.00862450897693634, + 0.013985441997647285, + 0.011307226493954659, + 0.04638377204537392, + 0.04808202385902405, + 0.011559667065739632, + 0.010032953694462776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0028249616734683514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_11/index_tensor([7082, 7082])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_11/index_tensor([7082, 7082])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cd85dbf68987aea07f1e78429aee0f2565d27dc9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_11/index_tensor([7082, 7082])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle macro shot remains static, focusing with exquisite precision on a vibrant red crab poised to strike. Its hard chitin shell, a brilliant fiery red, gleams in direct, bright sunlight. Its massive right pincer, a powerful weapon, stretches high, revealing its sharp, serrated inner edge, frozen in a tense moment of preparation. The camera employs a very shallow depth of field, completely blurring the surroundings into a soft, painterly, out-of-focus aquamarine ocean and beige sand, creating a beautiful bokeh effect. The crab's beady black eyes, perched on a stalk, gaze forward with aggressive intent, its body tense, its other leg braced firmly on the dark, damp sand below. The entire scene is imbued with a tense, confrontational atmosphere, capturing the raw, dramatic beauty of a creature in a defensive, hair-trigger stance.", + "seed": 5584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_11/index_7082_seed_5584_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01011209994321689, + "vlm_logprob": -0.01011209994321689, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.01011209994321689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00845850259065628 + ], + "opposite_logprob_tokens": [ + -5.008458614349365 + ], + "logprob_reward": -0.00845850259065628, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00845850259065628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0578315295279026, + 0.03219572827219963, + 0.10683200508356094, + 0.03704081103205681, + 0.049069587141275406, + 0.01820535585284233, + 0.08730904012918472, + 0.1393686980009079, + 0.2176872193813324, + 0.25446003675460815 + ], + "credit_map": [ + 0.008392393589019775, + 0.009920203126966953, + 0.008638311177492142, + 0.006695856805890799, + 0.006131232716143131, + 0.0066328272223472595, + 0.0061006261967122555, + 0.005320071708410978, + 0.003558393567800522, + 0.002970324596390128, + 0.0035784693900495768, + 0.004001770168542862, + 0.006194864399731159, + 0.003966968040913343, + 0.0033442694693803787, + 0.004580664914101362, + 0.01131418440490961, + 0.010619116015732288, + 0.01494629867374897, + 0.011828294955193996, + 0.02030719630420208, + 0.013769295997917652, + 0.009989882819354534, + 0.01405772939324379, + 0.005056349094957113, + 0.004761576186865568, + 0.004818280693143606, + 0.004221230745315552, + 0.004352652933448553, + 0.004347134847193956, + 0.003435540711507201, + 0.006048043258488178, + 0.008528000675141811, + 0.005748953204602003, + 0.005951606668531895, + 0.00619323318824172, + 0.007011198904365301, + 0.004505310207605362, + 0.005147062242031097, + 0.005984218325465918, + 0.0031414746772497892, + 0.0019640023820102215, + 0.0017286977963522077, + 0.0023353497963398695, + 0.0027082646265625954, + 0.0022847731597721577, + 0.0018785730935633183, + 0.0021642178762704134, + 0.01031801663339138, + 0.010028833523392677, + 0.01187088992446661, + 0.013166598975658417, + 0.011824602261185646, + 0.012367975898087025, + 0.008641383610665798, + 0.009090720675885677, + 0.021884838119149208, + 0.030636053532361984, + 0.022925719618797302, + 0.013708041980862617, + 0.016406269744038582, + 0.012842397205531597, + 0.012271481566131115, + 0.008693890646100044, + 0.05448499694466591, + 0.02136891335248947, + 0.01883038878440857, + 0.01880858652293682, + 0.05008134990930557, + 0.01770060881972313, + 0.01462465152144432, + 0.021787697449326515, + 0.03691119700670242, + 0.041403546929359436, + 0.054448265582323074, + 0.026575716212391853, + 0.032698798924684525, + 0.024490442126989365, + 0.022030731663107872, + 0.01590133085846901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.00845850259065628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059071511030197 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.02059071511030197, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059071511030197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07717481255531311, + 0.03151274099946022, + 0.08455990999937057, + 0.08555296808481216, + 0.026906518265604973, + 0.022401858121156693, + 0.0667804703116417, + 0.0969521701335907, + 0.4508604109287262, + 0.05729816108942032 + ], + "credit_map": [ + 0.012001515366137028, + 0.014660592190921307, + 0.015428034588694572, + 0.0086781932041049, + 0.008230911567807198, + 0.005904074292629957, + 0.007214167155325413, + 0.005057327449321747, + 0.003489200957119465, + 0.004920802544802427, + 0.003868794534355402, + 0.003606135956943035, + 0.006315904203802347, + 0.003644431009888649, + 0.0024790302850306034, + 0.003188445931300521, + 0.015812939032912254, + 0.008887707255780697, + 0.014111231081187725, + 0.008280618116259575, + 0.008410067297518253, + 0.008307266049087048, + 0.007137784268707037, + 0.013612302020192146, + 0.011705800890922546, + 0.012164167128503323, + 0.01300253439694643, + 0.008592206984758377, + 0.014637093059718609, + 0.008750412613153458, + 0.007733786478638649, + 0.008966974914073944, + 0.0049566831439733505, + 0.0033514322713017464, + 0.003132773097604513, + 0.0027959782164543867, + 0.004056819714605808, + 0.0029135646764189005, + 0.0023745261132717133, + 0.003324743825942278, + 0.005110591650009155, + 0.00350126251578331, + 0.0024944348260760307, + 0.0023257387802004814, + 0.0025931813288480043, + 0.0024035878013819456, + 0.00199340982362628, + 0.00197965232655406, + 0.005937377922236919, + 0.007793963886797428, + 0.008364822715520859, + 0.010338143445551395, + 0.01034762803465128, + 0.009643396362662315, + 0.005920715630054474, + 0.008434430696070194, + 0.017097823321819305, + 0.01714153029024601, + 0.014863638207316399, + 0.008449792861938477, + 0.01600397378206253, + 0.008550986647605896, + 0.008923953399062157, + 0.005920483265072107, + 0.10625658184289932, + 0.04449817165732384, + 0.06487100571393967, + 0.032669804990291595, + 0.08538276702165604, + 0.03611482307314873, + 0.028258228674530983, + 0.052809055894613266, + 0.007542541716247797, + 0.01031553652137518, + 0.009200645610690117, + 0.004347293172031641, + 0.00918588601052761, + 0.006805516313761473, + 0.005744673777371645, + 0.004156068898737431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.02059071511030197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064731404185295 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064731404185295, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064731404185295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06215869262814522, + 0.03561057522892952, + 0.07258470356464386, + 0.07914598286151886, + 0.051650431007146835, + 0.06567856669425964, + 0.06438882648944855, + 0.07213470339775085, + 0.4356100857257843, + 0.06103745847940445 + ], + "credit_map": [ + 0.008571538142859936, + 0.011722964234650135, + 0.009492699056863785, + 0.007202264852821827, + 0.0073349191807210445, + 0.006286067422479391, + 0.00715101882815361, + 0.004397211596369743, + 0.005840256344527006, + 0.006933663506060839, + 0.006716831587255001, + 0.0027203813660889864, + 0.0032960327807813883, + 0.002832353115081787, + 0.002944283653050661, + 0.004326772876083851, + 0.015075359493494034, + 0.005294963251799345, + 0.012364672496914864, + 0.0076333750039339066, + 0.00854436680674553, + 0.008401216939091682, + 0.006104288622736931, + 0.00916645210236311, + 0.01410374604165554, + 0.013103598728775978, + 0.012679754756391048, + 0.006675698794424534, + 0.010338615626096725, + 0.008509614504873753, + 0.006435762159526348, + 0.007299180142581463, + 0.005734156351536512, + 0.00512726092711091, + 0.00555519200861454, + 0.006998918019235134, + 0.007899018004536629, + 0.006102660205215216, + 0.0057406495325267315, + 0.00849257130175829, + 0.008260779082775116, + 0.008800542913377285, + 0.007080461829900742, + 0.009843200445175171, + 0.009297436103224754, + 0.007030990906059742, + 0.007165960501879454, + 0.008199190720915794, + 0.008370482362806797, + 0.007271341513842344, + 0.007901263423264027, + 0.008031328208744526, + 0.012707475572824478, + 0.0073057059198617935, + 0.006556190550327301, + 0.006245036143809557, + 0.009245868772268295, + 0.014512507244944572, + 0.008308262564241886, + 0.007095996290445328, + 0.014772354625165462, + 0.0074843475595116615, + 0.005643127951771021, + 0.005072230473160744, + 0.11261609196662903, + 0.03727201744914055, + 0.040156472474336624, + 0.018303461372852325, + 0.13399697840213776, + 0.04497583210468292, + 0.028902199119329453, + 0.019387008622288704, + 0.008687525056302547, + 0.014162750914692879, + 0.008581256493926048, + 0.004897380247712135, + 0.010950093157589436, + 0.005065058823674917, + 0.0048278821632266045, + 0.0038655041716992855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.016064731404185295 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008111285860650241 + ], + "opposite_logprob_tokens": [ + -7.1258111000061035 + ], + "logprob_reward": -0.0008111285860650241, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008111285860650241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0868612751364708, + 0.09466226398944855, + 0.0922093316912651, + 0.20017480850219727, + 0.05072924122214317, + 0.0515187568962574, + 0.08025197684764862, + 0.10965573787689209, + 0.1493048220872879, + 0.0846317857503891 + ], + "credit_map": [ + 0.011887067928910255, + 0.015054232440888882, + 0.013070916756987572, + 0.011274426244199276, + 0.008468459360301495, + 0.00970379263162613, + 0.009575961157679558, + 0.007826416753232479, + 0.015317551791667938, + 0.02405427023768425, + 0.020940596237778664, + 0.008141748607158661, + 0.009171818383038044, + 0.006436684634536505, + 0.004625000059604645, + 0.005974593572318554, + 0.0115661034360528, + 0.015216575935482979, + 0.01179664395749569, + 0.008764212019741535, + 0.011203047819435596, + 0.009735379368066788, + 0.008915206417441368, + 0.015012168325483799, + 0.050324320793151855, + 0.04005848243832588, + 0.03403391316533089, + 0.015927806496620178, + 0.015305077657103539, + 0.017547963187098503, + 0.013824439607560635, + 0.013152812607586384, + 0.011992842890322208, + 0.008357609622180462, + 0.007670024875551462, + 0.004644065164029598, + 0.004779384471476078, + 0.004391238559037447, + 0.004078068770468235, + 0.004816008266061544, + 0.008989262394607067, + 0.006377812474966049, + 0.006001296453177929, + 0.006427757907658815, + 0.005577895324677229, + 0.004520429763942957, + 0.0050520761869847775, + 0.008572230115532875, + 0.011342471465468407, + 0.008440138772130013, + 0.013796309940516949, + 0.01189187727868557, + 0.010640256106853485, + 0.009769530035555363, + 0.007429012097418308, + 0.00694238068535924, + 0.0181839931756258, + 0.024942774325609207, + 0.016186710447072983, + 0.011731158010661602, + 0.011558863334357738, + 0.009881035424768925, + 0.009834985248744488, + 0.007336214650422335, + 0.02428312599658966, + 0.016065387055277824, + 0.01949874311685562, + 0.012311375699937344, + 0.03400573506951332, + 0.016833942383527756, + 0.011182130314409733, + 0.015124376863241196, + 0.010310971178114414, + 0.013458345085382462, + 0.011238792911171913, + 0.01016569510102272, + 0.011437257751822472, + 0.01150493323802948, + 0.007363744080066681, + 0.009152039885520935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.0008111285860650241 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635422024875879 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635422024875879, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635422024875879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036149777472019196, + 0.052048634737730026, + 0.07181108742952347, + 0.10808622092008591, + 0.02751520462334156, + 0.06523492187261581, + 0.05341393128037453, + 0.034034986048936844, + 0.520348846912384, + 0.03135636821389198 + ], + "credit_map": [ + 0.004441129509359598, + 0.005242682062089443, + 0.004692143760621548, + 0.005566853564232588, + 0.003337971633300185, + 0.0034251431934535503, + 0.004705165047198534, + 0.004738684743642807, + 0.005960403010249138, + 0.008450583554804325, + 0.008075366728007793, + 0.006290655583143234, + 0.0091621745377779, + 0.005279831122606993, + 0.004706536885350943, + 0.004123080987483263, + 0.012367939576506615, + 0.00794016569852829, + 0.010901447385549545, + 0.005794920492917299, + 0.011096552014350891, + 0.008419281803071499, + 0.006190086714923382, + 0.00910069141536951, + 0.017340993508696556, + 0.013463574461638927, + 0.01371873077005148, + 0.010769505053758621, + 0.01708725094795227, + 0.010378829203546047, + 0.007597438991069794, + 0.01772988960146904, + 0.005171298515051603, + 0.0031981775537133217, + 0.003980507608503103, + 0.003107592696323991, + 0.0033171987161040306, + 0.0028597579803317785, + 0.002077509183436632, + 0.003803158411756158, + 0.012404454872012138, + 0.010773923248052597, + 0.006964319385588169, + 0.007419210392981768, + 0.008036679588258266, + 0.005990355741232634, + 0.007111184764653444, + 0.006534790154546499, + 0.01040246058255434, + 0.00736995181068778, + 0.013346638530492783, + 0.005561797879636288, + 0.005860153120011091, + 0.0037768573965877295, + 0.0032196699175983667, + 0.003876397619023919, + 0.006819371599704027, + 0.006939087063074112, + 0.005334469489753246, + 0.002741684904322028, + 0.0035574825014919043, + 0.0032800985500216484, + 0.0030295776668936014, + 0.002333213109523058, + 0.14251109957695007, + 0.044351592659950256, + 0.0348467193543911, + 0.029855215921998024, + 0.14439110457897186, + 0.04305198788642883, + 0.03295481950044632, + 0.04838629439473152, + 0.0036143662873655558, + 0.003634021384641528, + 0.005236313212662935, + 0.002551570301875472, + 0.004202213138341904, + 0.004336340352892876, + 0.0033041296992450953, + 0.004477411508560181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.004635422024875879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle macro shot remains static, focusing with exquisite precision on a vibrant red crab poised to strike. Its hard chitin shell, a brilliant fiery red, gleams in direct, bright sunlight. Its massive right pincer, a powerful weapon, stretches high, revealing its sharp, serrated inner edge, frozen in a tense moment of preparation. The camera employs a very shallow depth of field, completely blurring the surroundings into a soft, painterly, out-of-focus aquamarine ocean and beige sand, creating a beautiful bokeh effect. The crab's beady black eyes, perched on a stalk, gaze forward with aggressive intent, its body tense, its other leg braced firmly on the dark, damp sand below. The entire scene is imbued with a tense, confrontational atmosphere, capturing the raw, dramatic beauty of a creature in a defensive, hair-trigger stance.", + "seed": 5585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_11/index_7082_seed_5585_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01162543788086623, + "vlm_logprob": -0.01162543788086623, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.01162543788086623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009280282072722912 + ], + "opposite_logprob_tokens": [ + -5.384280204772949 + ], + "logprob_reward": -0.009280282072722912, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.009280282072722912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12542171776294708, + 0.20075875520706177, + 0.3164258599281311, + 0.02871798165142536, + 0.09386232495307922, + 0.021898044273257256, + 0.06532810628414154, + 0.025466734543442726, + 0.02771204337477684, + 0.09440845251083374 + ], + "credit_map": [ + 0.02015027590095997, + 0.025711936876177788, + 0.011944474652409554, + 0.010481268167495728, + 0.013624956831336021, + 0.015824129804968834, + 0.013059627264738083, + 0.014625041745603085, + 0.043868958950042725, + 0.04983217641711235, + 0.02263576351106167, + 0.022296590730547905, + 0.01522551104426384, + 0.017159396782517433, + 0.015388942323625088, + 0.014351414516568184, + 0.12052085250616074, + 0.1266319304704666, + 0.008267054334282875, + 0.010961056686937809, + 0.015713559463620186, + 0.015112951397895813, + 0.01085776835680008, + 0.00836071465164423, + 0.004630746319890022, + 0.003662878880277276, + 0.0025976616889238358, + 0.0029933673795312643, + 0.003463855478912592, + 0.005494853015989065, + 0.003336431225761771, + 0.0025381878949701786, + 0.02356022596359253, + 0.0147425327450037, + 0.006394806317985058, + 0.008148100227117538, + 0.016984619200229645, + 0.00790907721966505, + 0.008942506276071072, + 0.0071804579347372055, + 0.005289399530738592, + 0.0032085198909044266, + 0.001816718839108944, + 0.0017498605884611607, + 0.0029439288191497326, + 0.00265938020311296, + 0.002741391770541668, + 0.0014888442819938064, + 0.014391209930181503, + 0.010698492638766766, + 0.005003775469958782, + 0.005319629795849323, + 0.011251777410507202, + 0.00664858240634203, + 0.006881168112158775, + 0.005133467726409435, + 0.006644171196967363, + 0.006055929232388735, + 0.0026464711409062147, + 0.0017754043219611049, + 0.0025051431730389595, + 0.002903325716033578, + 0.0014830147847533226, + 0.0014532760251313448, + 0.005006430204957724, + 0.004711323417723179, + 0.002093259245157242, + 0.0039370059967041016, + 0.004375912714749575, + 0.003234842326492071, + 0.0025060579646378756, + 0.0018472116207703948, + 0.019062884151935577, + 0.02419009618461132, + 0.01015234924852848, + 0.012540522962808609, + 0.007719100918620825, + 0.009235586039721966, + 0.005839390680193901, + 0.005668523255735636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.009280282072722912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608009822666645 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.01608009822666645, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608009822666645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08134341984987259, + 0.081736259162426, + 0.4963471591472626, + 0.03184478357434273, + 0.07975955307483673, + 0.018612375482916832, + 0.03488036245107651, + 0.036763232201337814, + 0.093828484416008, + 0.044884320348501205 + ], + "credit_map": [ + 0.0151832839474082, + 0.019034862518310547, + 0.009662852622568607, + 0.008502352982759476, + 0.006810558028519154, + 0.008198537863790989, + 0.008086695335805416, + 0.005864281672984362, + 0.010371202602982521, + 0.00858363602310419, + 0.013233751989901066, + 0.013489632867276669, + 0.008443587459623814, + 0.00905870832502842, + 0.009229537099599838, + 0.00932620745152235, + 0.19360841810703278, + 0.20139236748218536, + 0.013255787082016468, + 0.017687520012259483, + 0.022777194157242775, + 0.022987214848399162, + 0.013229542411863804, + 0.01140914298593998, + 0.005430900026112795, + 0.004160335753113031, + 0.0029438715428113937, + 0.004185267724096775, + 0.005199573468416929, + 0.004588021896779537, + 0.002926962450146675, + 0.0024098558351397514, + 0.02215566299855709, + 0.01278696395456791, + 0.005440124776214361, + 0.005794985685497522, + 0.0122373066842556, + 0.008492751978337765, + 0.006672482006251812, + 0.006179278250783682, + 0.00384035287424922, + 0.003346378216519952, + 0.0017210098449140787, + 0.0018527084030210972, + 0.002767161000519991, + 0.0016287496546283364, + 0.001979215070605278, + 0.0014768014661967754, + 0.007720252964645624, + 0.006262035574764013, + 0.002788874553516507, + 0.0036455730441957712, + 0.004863048437982798, + 0.0037778785917907953, + 0.0030937434639781713, + 0.002728957450017333, + 0.007573412265628576, + 0.008733338676393032, + 0.0029959676321595907, + 0.0033595585264265537, + 0.004104022867977619, + 0.004523357842117548, + 0.0026633532252162695, + 0.002810220466926694, + 0.02698693983256817, + 0.02853146754205227, + 0.004661035258322954, + 0.0066993809305131435, + 0.01021658256649971, + 0.007633361965417862, + 0.005006404127925634, + 0.004093312658369541, + 0.008831717073917389, + 0.007435671519488096, + 0.003946959972381592, + 0.0052276249043643475, + 0.006200138479471207, + 0.0038472991436719894, + 0.006137179676443338, + 0.0032577330712229013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.01608009822666645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633899636566639 + ], + "opposite_logprob_tokens": [ + -3.651339054107666 + ], + "logprob_reward": -0.02633899636566639, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633899636566639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08415628224611282, + 0.20101206004619598, + 0.3375641405582428, + 0.05399565398693085, + 0.08842860162258148, + 0.027803000062704086, + 0.04140856862068176, + 0.03596242517232895, + 0.055474646389484406, + 0.07419466227293015 + ], + "credit_map": [ + 0.014072034507989883, + 0.017899461090564728, + 0.01089680939912796, + 0.009734668768942356, + 0.008947014808654785, + 0.008218403905630112, + 0.007818404585123062, + 0.006569479126483202, + 0.074814073741436, + 0.0754353255033493, + 0.00909404642879963, + 0.010782523080706596, + 0.009387281723320484, + 0.006700040772557259, + 0.00781956035643816, + 0.006979207508265972, + 0.12181879580020905, + 0.1273052841424942, + 0.008927484974265099, + 0.01402666512876749, + 0.025997448712587357, + 0.015833334997296333, + 0.013827369548380375, + 0.00982775166630745, + 0.013833538629114628, + 0.009375092573463917, + 0.004988144617527723, + 0.005548934917896986, + 0.00833893846720457, + 0.004265147261321545, + 0.004265575669705868, + 0.0033802820835262537, + 0.017184533178806305, + 0.012180659919977188, + 0.008335075341165066, + 0.007608328014612198, + 0.016648277640342712, + 0.010666577145457268, + 0.008188032545149326, + 0.007617124356329441, + 0.008289330638945103, + 0.004673874005675316, + 0.002376360585913062, + 0.002686875406652689, + 0.003710606135427952, + 0.0021134146954864264, + 0.0021879973355680704, + 0.001764541259035468, + 0.006215160246938467, + 0.005413745995610952, + 0.004457275383174419, + 0.003601828357204795, + 0.006837891880422831, + 0.005661090835928917, + 0.005442013964056969, + 0.003779561957344413, + 0.007567035034298897, + 0.00586328050121665, + 0.0032778787426650524, + 0.003572196466848254, + 0.005341910757124424, + 0.0031056092120707035, + 0.00412283418700099, + 0.0031116786412894726, + 0.011561710387468338, + 0.011816438287496567, + 0.0064185322262346745, + 0.005968266166746616, + 0.005197376944124699, + 0.00690839160233736, + 0.0032898166682571173, + 0.004314113408327103, + 0.012033797800540924, + 0.01270148903131485, + 0.010140150785446167, + 0.00970609299838543, + 0.007328804116696119, + 0.010694937780499458, + 0.006530002225190401, + 0.005059388000518084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.02633899636566639 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178642618469894 + ], + "opposite_logprob_tokens": [ + -6.751178741455078 + ], + "logprob_reward": -0.001178642618469894, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178642618469894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20826108753681183, + 0.07278215885162354, + 0.2512023448944092, + 0.08234845101833344, + 0.08218947798013687, + 0.05197323113679886, + 0.07343658804893494, + 0.04375224933028221, + 0.06511019170284271, + 0.06894417107105255 + ], + "credit_map": [ + 0.0416075736284256, + 0.048648785799741745, + 0.02537420205771923, + 0.019261740148067474, + 0.019485481083393097, + 0.021061161532998085, + 0.01818373240530491, + 0.01463839691132307, + 0.017055446282029152, + 0.016591480001807213, + 0.005861373618245125, + 0.007829273119568825, + 0.006862684153020382, + 0.008275668136775494, + 0.005716834682971239, + 0.004589402582496405, + 0.09189947694540024, + 0.0952238142490387, + 0.009821118786931038, + 0.009995686821639538, + 0.012229599058628082, + 0.011220863088965416, + 0.010489853098988533, + 0.010321949608623981, + 0.02789454162120819, + 0.01576521247625351, + 0.0051040383987128735, + 0.00749060045927763, + 0.008583493530750275, + 0.007087209261953831, + 0.005834069103002548, + 0.004589286167174578, + 0.028263533487915993, + 0.017615338787436485, + 0.0048050121404230595, + 0.0053522964008152485, + 0.008481190539896488, + 0.006835645530372858, + 0.005041603464633226, + 0.005794861353933811, + 0.014557271264493465, + 0.00979850348085165, + 0.0034785938914865255, + 0.004681645426899195, + 0.005614187568426132, + 0.005143736023455858, + 0.004146907478570938, + 0.004552388098090887, + 0.014344020746648312, + 0.012100407853722572, + 0.006880353204905987, + 0.00852952804416418, + 0.00956546887755394, + 0.009513606317341328, + 0.006854823790490627, + 0.005648381542414427, + 0.011272786185145378, + 0.006989666726440191, + 0.003740462241694331, + 0.005227637477219105, + 0.005093064159154892, + 0.004782195668667555, + 0.0036495353560894728, + 0.0029969022143632174, + 0.021282797679305077, + 0.02290738932788372, + 0.0029940903186798096, + 0.0032964213751256466, + 0.004846594762057066, + 0.003878663992509246, + 0.0030186548829078674, + 0.0028855777345597744, + 0.0144475307315588, + 0.015425064601004124, + 0.005866097752004862, + 0.007301943376660347, + 0.008884083479642868, + 0.0062351166270673275, + 0.005914196837693453, + 0.0048701344057917595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.001178642618469894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249170120805502 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005249170120805502, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249170120805502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20047876238822937, + 0.15439032018184662, + 0.2537398040294647, + 0.0726216584444046, + 0.052575018256902695, + 0.0500180646777153, + 0.06343282014131546, + 0.032968416810035706, + 0.039920687675476074, + 0.07985451072454453 + ], + "credit_map": [ + 0.03772241622209549, + 0.056763503700494766, + 0.015614987351000309, + 0.018271375447511673, + 0.017702870070934296, + 0.0213776808232069, + 0.01811359077692032, + 0.014912295155227184, + 0.04564642161130905, + 0.048781149089336395, + 0.011561380699276924, + 0.012326817959547043, + 0.01007305271923542, + 0.010724147781729698, + 0.008170959539711475, + 0.007106370758265257, + 0.09470675140619278, + 0.09777010977268219, + 0.0082537941634655, + 0.010606301948428154, + 0.014671232551336288, + 0.012348107062280178, + 0.00836555939167738, + 0.0070179118774831295, + 0.0221764724701643, + 0.0128589803352952, + 0.004424255806952715, + 0.006764797959476709, + 0.008489131927490234, + 0.005115494132041931, + 0.005916251800954342, + 0.006876267958432436, + 0.017138050869107246, + 0.010444765910506248, + 0.003333642613142729, + 0.003524527885019779, + 0.0063593508675694466, + 0.004792099352926016, + 0.00342333922162652, + 0.0035592354834079742, + 0.014410867355763912, + 0.009112873114645481, + 0.0034767738543450832, + 0.005373036023229361, + 0.006115819327533245, + 0.004530181642621756, + 0.003748897463083267, + 0.0032496063504368067, + 0.012211518362164497, + 0.009699374437332153, + 0.005322699900716543, + 0.007606932893395424, + 0.009789547882974148, + 0.008237778209149837, + 0.0060032508336007595, + 0.004561707843095064, + 0.006015863735228777, + 0.005492054857313633, + 0.0028454086277633905, + 0.0031658445950597525, + 0.005705663934350014, + 0.003142384346574545, + 0.003624544944614172, + 0.002976648975163698, + 0.010693453252315521, + 0.011549540795385838, + 0.0027638552710413933, + 0.0023769536055624485, + 0.004666682332754135, + 0.002767224796116352, + 0.002800755202770233, + 0.0023022182285785675, + 0.017032964155077934, + 0.01750314235687256, + 0.005262593738734722, + 0.008138839155435562, + 0.013723371550440788, + 0.005531640723347664, + 0.008615917526185513, + 0.004046033602207899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.005249170120805502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle macro shot remains static, focusing with exquisite precision on a vibrant red crab poised to strike. Its hard chitin shell, a brilliant fiery red, gleams in direct, bright sunlight. Its massive right pincer, a powerful weapon, stretches high, revealing its sharp, serrated inner edge, frozen in a tense moment of preparation. The camera employs a very shallow depth of field, completely blurring the surroundings into a soft, painterly, out-of-focus aquamarine ocean and beige sand, creating a beautiful bokeh effect. The crab's beady black eyes, perched on a stalk, gaze forward with aggressive intent, its body tense, its other leg braced firmly on the dark, damp sand below. The entire scene is imbued with a tense, confrontational atmosphere, capturing the raw, dramatic beauty of a creature in a defensive, hair-trigger stance.", + "seed": 5586, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_11/index_7082_seed_5586_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012463662121444941, + "vlm_logprob": -0.012463662121444941, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.012463662121444941, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01450212113559246 + ], + "opposite_logprob_tokens": [ + -4.389502048492432 + ], + "logprob_reward": -0.01450212113559246, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01450212113559246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12223094701766968, + 0.09842457622289658, + 0.1240193098783493, + 0.0347970686852932, + 0.17353011667728424, + 0.0723513513803482, + 0.0430767796933651, + 0.2579181492328644, + 0.03944918140769005, + 0.03420251980423927 + ], + "credit_map": [ + 0.018891368061304092, + 0.011085322126746178, + 0.01872020587325096, + 0.015041634440422058, + 0.014130893163383007, + 0.01302260346710682, + 0.01297991257160902, + 0.01835901290178299, + 0.01482196245342493, + 0.010973067954182625, + 0.012540011666715145, + 0.011092721484601498, + 0.011088671162724495, + 0.01985933445394039, + 0.00662437966093421, + 0.01142442412674427, + 0.01815386302769184, + 0.019934970885515213, + 0.011559292674064636, + 0.010399609804153442, + 0.01437470968812704, + 0.027023082599043846, + 0.00937427394092083, + 0.01319950819015503, + 0.004935221280902624, + 0.005226812791079283, + 0.0029423804953694344, + 0.0033979674335569143, + 0.004866834729909897, + 0.006194243673235178, + 0.002797323279082775, + 0.004436283838003874, + 0.027506357058882713, + 0.02382933720946312, + 0.017318330705165863, + 0.032376814633607864, + 0.012735037133097649, + 0.02827254869043827, + 0.014973707497119904, + 0.01651800610125065, + 0.011124596931040287, + 0.01004309207201004, + 0.007310694083571434, + 0.01229004841297865, + 0.005237234756350517, + 0.010695740580558777, + 0.006608687341213226, + 0.009041251614689827, + 0.007173146586865187, + 0.005505461245775223, + 0.0035363854840397835, + 0.006262805312871933, + 0.003928584046661854, + 0.00893049780279398, + 0.004046211019158363, + 0.003693689126521349, + 0.03990260511636734, + 0.049093104898929596, + 0.012838457711040974, + 0.02192803844809532, + 0.02833096869289875, + 0.0625583603978157, + 0.013156400062143803, + 0.03011021763086319, + 0.006273097358644009, + 0.006939963437616825, + 0.0027279674541205168, + 0.004299943335354328, + 0.004184451419860125, + 0.0074001215398311615, + 0.004214039538055658, + 0.0034095984883606434, + 0.005013362504541874, + 0.005372957326471806, + 0.003003676189109683, + 0.0038709882646799088, + 0.0029352353885769844, + 0.006596988067030907, + 0.0031324997544288635, + 0.004276815336197615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.01450212113559246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302515968680382 + ], + "opposite_logprob_tokens": [ + -3.7733025550842285 + ], + "logprob_reward": -0.023302515968680382, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302515968680382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10925425589084625, + 0.06692962348461151, + 0.1102987602353096, + 0.07000339031219482, + 0.08574352413415909, + 0.06785087287425995, + 0.05576702207326889, + 0.31226760149002075, + 0.050020020455121994, + 0.07186494022607803 + ], + "credit_map": [ + 0.01499588880687952, + 0.010423034429550171, + 0.022712239995598793, + 0.010642614215612411, + 0.011659183539450169, + 0.016834808513522148, + 0.008646959438920021, + 0.01333952322602272, + 0.009438960812985897, + 0.009092644788324833, + 0.01007724367082119, + 0.008675937540829182, + 0.00799239706248045, + 0.00864744558930397, + 0.00643977802246809, + 0.006565210409462452, + 0.014228977262973785, + 0.02007959596812725, + 0.009850134141743183, + 0.01239616796374321, + 0.014928611926734447, + 0.015527359209954739, + 0.010873708873987198, + 0.012414209544658661, + 0.009113174863159657, + 0.008856547065079212, + 0.008342603221535683, + 0.007898760959506035, + 0.006063166540116072, + 0.013099719770252705, + 0.005896205548197031, + 0.010733210481703281, + 0.011700604110956192, + 0.013207295909523964, + 0.007458143401890993, + 0.012058338150382042, + 0.010735098272562027, + 0.015431320294737816, + 0.007493000011891127, + 0.007659726310521364, + 0.010351317003369331, + 0.008133724331855774, + 0.008629652671515942, + 0.009080257266759872, + 0.007645945530384779, + 0.010361443273723125, + 0.006201474461704493, + 0.007447054143995047, + 0.00833834894001484, + 0.005238340236246586, + 0.00462309829890728, + 0.008066197857260704, + 0.005030755884945393, + 0.011690138839185238, + 0.007348334416747093, + 0.005431803409010172, + 0.04912498965859413, + 0.061065178364515305, + 0.01576605811715126, + 0.02304619550704956, + 0.03367280215024948, + 0.07277216762304306, + 0.015918849036097527, + 0.04090135172009468, + 0.007887990213930607, + 0.006923730485141277, + 0.0037970177363604307, + 0.0048595876432955265, + 0.0051119388081133366, + 0.010732418857514858, + 0.005528692621737719, + 0.005178641062229872, + 0.009173490107059479, + 0.011139359325170517, + 0.007181718945503235, + 0.0073425606824457645, + 0.00928623229265213, + 0.010773645713925362, + 0.007123867515474558, + 0.009844060055911541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.023302515968680382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065435484051704 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065435484051704, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065435484051704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15066976845264435, + 0.05540326237678528, + 0.10199681669473648, + 0.036237433552742004, + 0.09622491896152496, + 0.054264239966869354, + 0.07049142569303513, + 0.34190911054611206, + 0.04376678168773651, + 0.04903620854020119 + ], + "credit_map": [ + 0.020045317709445953, + 0.016453512012958527, + 0.014003448188304901, + 0.019648097455501556, + 0.023715930059552193, + 0.021414872258901596, + 0.017322741448879242, + 0.018065856769680977, + 0.008425407111644745, + 0.008197000250220299, + 0.004911101423203945, + 0.005881601944565773, + 0.008231919258832932, + 0.008579099550843239, + 0.005803623702377081, + 0.005373513326048851, + 0.021057551726698875, + 0.011205713264644146, + 0.007374077569693327, + 0.00856942031532526, + 0.008783498778939247, + 0.027225933969020844, + 0.006397180259227753, + 0.011383436620235443, + 0.004778517410159111, + 0.0064695836044847965, + 0.0033990084193646908, + 0.0044924491085112095, + 0.004405774176120758, + 0.00619171978905797, + 0.0030993216205388308, + 0.003401057794690132, + 0.016644001007080078, + 0.01239763107150793, + 0.009076705202460289, + 0.01111393328756094, + 0.007826592773199081, + 0.02431267872452736, + 0.008781490847468376, + 0.006071895826607943, + 0.007805010303854942, + 0.010019788518548012, + 0.004239486064761877, + 0.004849475342780352, + 0.007121945731341839, + 0.012190704233944416, + 0.0039846021682024, + 0.00405323039740324, + 0.01417582668364048, + 0.0075127724558115005, + 0.005332206841558218, + 0.007724151015281677, + 0.0063978577964007854, + 0.016561342403292656, + 0.0071252030320465565, + 0.005662070121616125, + 0.054158058017492294, + 0.07483562082052231, + 0.016695093363523483, + 0.028767192736268044, + 0.039200685918331146, + 0.07575791329145432, + 0.01653123088181019, + 0.03596333786845207, + 0.005862819496542215, + 0.00913237314671278, + 0.0031070157419890165, + 0.006029347889125347, + 0.004573386162519455, + 0.004678148310631514, + 0.0050908299162983894, + 0.0052928682416677475, + 0.009056666865944862, + 0.011895766481757164, + 0.0037906402722001076, + 0.0050415340811014175, + 0.00592332286760211, + 0.0042925369925796986, + 0.005717508029192686, + 0.003318238537758589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.016065435484051704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017122859135270119 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017122859135270119, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017122859135270119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21665871143341064, + 0.1230580136179924, + 0.09059017151594162, + 0.09377547353506088, + 0.08890888839960098, + 0.07580716907978058, + 0.049285221844911575, + 0.10318521410226822, + 0.039163973182439804, + 0.11956718564033508 + ], + "credit_map": [ + 0.033931177109479904, + 0.02305101975798607, + 0.02092975191771984, + 0.020847748965024948, + 0.021930048242211342, + 0.050982262939214706, + 0.021893974393606186, + 0.023092713207006454, + 0.024193156510591507, + 0.014465991407632828, + 0.008847614750266075, + 0.011542560532689095, + 0.011648902669548988, + 0.028810637071728706, + 0.007585524115711451, + 0.015963613986968994, + 0.014613328501582146, + 0.010799447074532509, + 0.007066538091748953, + 0.008113154210150242, + 0.007661403622478247, + 0.023019805550575256, + 0.009789630770683289, + 0.00952685996890068, + 0.017194759100675583, + 0.009684186428785324, + 0.005920246243476868, + 0.007102250121533871, + 0.008056273683905602, + 0.03178536519408226, + 0.005978711880743504, + 0.008053682744503021, + 0.012297241017222404, + 0.011914982460439205, + 0.007792462594807148, + 0.010511772707104683, + 0.009148105047643185, + 0.0217723585665226, + 0.009841836988925934, + 0.005630120635032654, + 0.010319355875253677, + 0.01434736791998148, + 0.00703577371314168, + 0.008143719285726547, + 0.007045572157949209, + 0.017518965527415276, + 0.004544617608189583, + 0.006851786281913519, + 0.006899374071508646, + 0.004965171217918396, + 0.004917491227388382, + 0.007884475402534008, + 0.004417666234076023, + 0.009081934578716755, + 0.007162449415773153, + 0.003956656903028488, + 0.014623476192355156, + 0.023074135184288025, + 0.005369211547076702, + 0.01004431489855051, + 0.01243226695805788, + 0.020179802551865578, + 0.006730442866683006, + 0.010731547139585018, + 0.006630730349570513, + 0.009064817801117897, + 0.0028976649045944214, + 0.004668803419917822, + 0.00383636262267828, + 0.004318392835557461, + 0.004762512166053057, + 0.0029846851248294115, + 0.020277520641684532, + 0.03187081962823868, + 0.007750188000500202, + 0.00992680061608553, + 0.013209891505539417, + 0.011682826094329357, + 0.015192057006061077, + 0.00965708028525114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0017122859135270119 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735952105373144 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735952105373144, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735952105373144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10215795785188675, + 0.12727612257003784, + 0.13962626457214355, + 0.044673673808574677, + 0.0968298390507698, + 0.11521878838539124, + 0.08924562484025955, + 0.09917648881673813, + 0.12216644734144211, + 0.06362883001565933 + ], + "credit_map": [ + 0.015963975340127945, + 0.01193313580006361, + 0.010669663548469543, + 0.011706055141985416, + 0.012694244273006916, + 0.02034202218055725, + 0.007395619992166758, + 0.01145323459059, + 0.02477477863430977, + 0.012390125542879105, + 0.008965947665274143, + 0.009395059198141098, + 0.013620392419397831, + 0.03658737987279892, + 0.007384230382740498, + 0.014158208854496479, + 0.02909073419868946, + 0.01504780724644661, + 0.009476588107645512, + 0.011510409414768219, + 0.01204638835042715, + 0.03640768676996231, + 0.010471650399267673, + 0.015575005672872066, + 0.006827360950410366, + 0.004335510078817606, + 0.00422658771276474, + 0.0046219895593822, + 0.003815407631918788, + 0.009999348782002926, + 0.004016941878944635, + 0.0068305316381156445, + 0.013016692362725735, + 0.012715499848127365, + 0.01053815707564354, + 0.013569596223533154, + 0.011257665231823921, + 0.016839521005749702, + 0.012631187215447426, + 0.006261525675654411, + 0.01591401919722557, + 0.02233230508863926, + 0.013033442199230194, + 0.016347376629710197, + 0.011203763075172901, + 0.015440287999808788, + 0.008371573872864246, + 0.012576020322740078, + 0.013498066924512386, + 0.009440992027521133, + 0.008017954416573048, + 0.012075502425432205, + 0.008266206830739975, + 0.021152043715119362, + 0.010174158960580826, + 0.0066207037307322025, + 0.015914807096123695, + 0.024786915630102158, + 0.006391191389411688, + 0.009028528816998005, + 0.011658676899969578, + 0.014574519358575344, + 0.008091499097645283, + 0.008730348199605942, + 0.019527433440089226, + 0.03180393949151039, + 0.006421591620892286, + 0.01194631028920412, + 0.015521791763603687, + 0.019607853144407272, + 0.0071503897197544575, + 0.010187133215367794, + 0.009751349687576294, + 0.015370099805295467, + 0.004044622182846069, + 0.006410625297576189, + 0.00713350810110569, + 0.010841422714293003, + 0.004620529245585203, + 0.005456676706671715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.006735952105373144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle macro shot remains static, focusing with exquisite precision on a vibrant red crab poised to strike. Its hard chitin shell, a brilliant fiery red, gleams in direct, bright sunlight. Its massive right pincer, a powerful weapon, stretches high, revealing its sharp, serrated inner edge, frozen in a tense moment of preparation. The camera employs a very shallow depth of field, completely blurring the surroundings into a soft, painterly, out-of-focus aquamarine ocean and beige sand, creating a beautiful bokeh effect. The crab's beady black eyes, perched on a stalk, gaze forward with aggressive intent, its body tense, its other leg braced firmly on the dark, damp sand below. The entire scene is imbued with a tense, confrontational atmosphere, capturing the raw, dramatic beauty of a creature in a defensive, hair-trigger stance.", + "seed": 5587, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_11/index_7082_seed_5587_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013847611891105772, + "vlm_logprob": -0.013847611891105772, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.013847611891105772, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02827291749417782 + ], + "opposite_logprob_tokens": [ + -3.653272867202759 + ], + "logprob_reward": -0.02827291749417782, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02827291749417782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1112355962395668, + 0.11224229633808136, + 0.1172807589173317, + 0.16287200152873993, + 0.11291023343801498, + 0.16270634531974792, + 0.06398989260196686, + 0.017315158620476723, + 0.06936538964509964, + 0.07008232921361923 + ], + "credit_map": [ + 0.012163817882537842, + 0.012979867868125439, + 0.010136197321116924, + 0.016375837847590446, + 0.010592829436063766, + 0.014339612796902657, + 0.01728629320859909, + 0.017361147329211235, + 0.01567447930574417, + 0.03552118316292763, + 0.027739115059375763, + 0.009322956204414368, + 0.006618359591811895, + 0.005312665831297636, + 0.006667500827461481, + 0.00538603076711297, + 0.023929957300424576, + 0.04328855499625206, + 0.026472484692931175, + 0.006730853579938412, + 0.0048493132926523685, + 0.003423180663958192, + 0.004468225408345461, + 0.004118185956031084, + 0.04657033458352089, + 0.05528200417757034, + 0.01635604165494442, + 0.009727475233376026, + 0.008962191641330719, + 0.009363549761474133, + 0.008701193146407604, + 0.007909207604825497, + 0.018290944397449493, + 0.017731333151459694, + 0.012420246377587318, + 0.018148986622691154, + 0.010710801929235458, + 0.011800236068665981, + 0.013522772118449211, + 0.010284919291734695, + 0.022467857226729393, + 0.06168344244360924, + 0.0508401058614254, + 0.006427512038499117, + 0.006274496205151081, + 0.00461970642209053, + 0.005074101034551859, + 0.005319133400917053, + 0.006313502322882414, + 0.0059029776602983475, + 0.007692997343838215, + 0.007435759529471397, + 0.008020235225558281, + 0.007499257568269968, + 0.012456236407160759, + 0.008668922819197178, + 0.0029705106280744076, + 0.00212724506855011, + 0.003019093070179224, + 0.0016844882629811764, + 0.0019295673118904233, + 0.001798209734261036, + 0.0016604142729192972, + 0.0021256310865283012, + 0.009786160662770271, + 0.009925364516675472, + 0.007913084700703621, + 0.011622270569205284, + 0.005797707941383123, + 0.008173669688403606, + 0.007631534244865179, + 0.008515596389770508, + 0.009251563809812069, + 0.008792257867753506, + 0.012437456287443638, + 0.007108891382813454, + 0.006853118538856506, + 0.008307195268571377, + 0.007882005535066128, + 0.009449840523302555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.02827291749417782 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197137862443924 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197137862443924, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197137862443924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19090847671031952, + 0.2770407199859619, + 0.04841959476470947, + 0.09792695939540863, + 0.06290500611066818, + 0.032619915902614594, + 0.09254835546016693, + 0.033246658742427826, + 0.029276054352521896, + 0.13510820269584656 + ], + "credit_map": [ + 0.0210241861641407, + 0.019149793311953545, + 0.017975537106394768, + 0.03420749679207802, + 0.01696058176457882, + 0.019186964258551598, + 0.03417036682367325, + 0.02823357656598091, + 0.03696153685450554, + 0.10412847250699997, + 0.08451186120510101, + 0.014678947627544403, + 0.010786906816065311, + 0.00872095674276352, + 0.008788267150521278, + 0.00846378318965435, + 0.006114725489169359, + 0.009110759012401104, + 0.013747993856668472, + 0.004380807280540466, + 0.003906618803739548, + 0.0034161400981247425, + 0.0034402243327349424, + 0.004302330315113068, + 0.02339596301317215, + 0.03152775764465332, + 0.01547954510897398, + 0.006445913575589657, + 0.005837826523929834, + 0.0047543845139443874, + 0.004807288758456707, + 0.005678290035575628, + 0.00843240786343813, + 0.009102980606257915, + 0.008110709488391876, + 0.007109937258064747, + 0.007507624104619026, + 0.00849095731973648, + 0.007560758385807276, + 0.006589632946997881, + 0.00896584801375866, + 0.008052135817706585, + 0.002483668038621545, + 0.0029891887679696083, + 0.002134328940883279, + 0.0022055883891880512, + 0.0030683306977152824, + 0.0027208318933844566, + 0.015487585216760635, + 0.00943413283675909, + 0.016454525291919708, + 0.009085472673177719, + 0.01121283508837223, + 0.00955742783844471, + 0.01131343748420477, + 0.01000293716788292, + 0.005920701660215855, + 0.004019654355943203, + 0.005298728588968515, + 0.0031549446284770966, + 0.0032570709008723497, + 0.005380741320550442, + 0.0028594702016562223, + 0.0033553503453731537, + 0.004834367893636227, + 0.004068555776029825, + 0.0048740883357822895, + 0.0031361279543489218, + 0.002800718182697892, + 0.004225465934723616, + 0.002702072961255908, + 0.002634662203490734, + 0.022233230993151665, + 0.017139343544840813, + 0.02612774260342121, + 0.01176473032683134, + 0.013484437018632889, + 0.01666662096977234, + 0.012018081732094288, + 0.015674041584134102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.018197137862443924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180163577198982 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180163577198982, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180163577198982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2851991653442383, + 0.19492608308792114, + 0.062456365674734116, + 0.05100144073367119, + 0.08851040154695511, + 0.15453766286373138, + 0.06421660631895065, + 0.022784164175391197, + 0.0361117459833622, + 0.040256306529045105 + ], + "credit_map": [ + 0.02584519051015377, + 0.030017171055078506, + 0.030125416815280914, + 0.049579039216041565, + 0.021049397066235542, + 0.038449354469776154, + 0.050836797803640366, + 0.03929683938622475, + 0.020937755703926086, + 0.06422095745801926, + 0.0576191209256649, + 0.016240166500210762, + 0.011287893168628216, + 0.007173018530011177, + 0.009494110941886902, + 0.007953069172799587, + 0.010412005707621574, + 0.017361395061016083, + 0.011705920100212097, + 0.005620614159852266, + 0.0043199313804507256, + 0.004267060663551092, + 0.004303880967199802, + 0.004465562291443348, + 0.008500173687934875, + 0.006723768077790737, + 0.008960499428212643, + 0.008286661468446255, + 0.0036705653183162212, + 0.0037825503386557102, + 0.00494480412453413, + 0.006132421083748341, + 0.013710485771298409, + 0.010692441835999489, + 0.012925609014928341, + 0.010696903802454472, + 0.010684815235435963, + 0.010697477497160435, + 0.008481288328766823, + 0.010621383786201477, + 0.02561625838279724, + 0.057731278240680695, + 0.04254572466015816, + 0.007405989803373814, + 0.005988544784486294, + 0.0041930400766432285, + 0.005962062627077103, + 0.005094789434224367, + 0.013642434030771255, + 0.005978780798614025, + 0.01537930779159069, + 0.007892895489931107, + 0.0034354454837739468, + 0.0050857048481702805, + 0.0077828578650951385, + 0.005019188858568668, + 0.0029290856327861547, + 0.002581618959084153, + 0.0030025371816009283, + 0.0032596206292510033, + 0.0022646004799753428, + 0.0031964650843292475, + 0.0028710761107504368, + 0.00267916196025908, + 0.004484662786126137, + 0.00430637551471591, + 0.004688288550823927, + 0.006782634649425745, + 0.003174601122736931, + 0.005159609019756317, + 0.0035380034241825342, + 0.0039775739423930645, + 0.005402832757681608, + 0.005677924025803804, + 0.004805870819836855, + 0.004528080113232136, + 0.004794460255652666, + 0.006350955925881863, + 0.0039454312063753605, + 0.004750754218548536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.018180163577198982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004943578969687223 + ], + "opposite_logprob_tokens": [ + -7.625494480133057 + ], + "logprob_reward": -0.0004943578969687223, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004943578969687223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11846520751714706, + 0.16975700855255127, + 0.3051052391529083, + 0.09056555479764938, + 0.035977378487586975, + 0.10464207082986832, + 0.04540492594242096, + 0.03903139755129814, + 0.0667259469628334, + 0.024325236678123474 + ], + "credit_map": [ + 0.01606745272874832, + 0.02344612590968609, + 0.013227777555584908, + 0.013625575229525566, + 0.009635663591325283, + 0.012906888499855995, + 0.013563341461122036, + 0.015992386266589165, + 0.03052503988146782, + 0.06590158492326736, + 0.04236188530921936, + 0.0076877400279045105, + 0.006783133838325739, + 0.005577967967838049, + 0.005483143497258425, + 0.00543651357293129, + 0.048469364643096924, + 0.12069261074066162, + 0.09510082751512527, + 0.010375694371759892, + 0.010741797275841236, + 0.007170313969254494, + 0.006697876378893852, + 0.005856750998646021, + 0.022661827504634857, + 0.01309012621641159, + 0.020631885156035423, + 0.009760379791259766, + 0.005728596355766058, + 0.0056803105399012566, + 0.0056214602664113045, + 0.007390968035906553, + 0.006445360369980335, + 0.004504511132836342, + 0.005801133345812559, + 0.004351729527115822, + 0.003028203034773469, + 0.004427777137607336, + 0.0033166634384542704, + 0.004102002829313278, + 0.020473795011639595, + 0.02646784484386444, + 0.03842640668153763, + 0.005655778106302023, + 0.0034382804296910763, + 0.0031543546356260777, + 0.003740407060831785, + 0.003285201732069254, + 0.01209063921123743, + 0.0058250268921256065, + 0.009548497386276722, + 0.0035839383490383625, + 0.0033125008922070265, + 0.0032445495016872883, + 0.0040954952128231525, + 0.0037042777985334396, + 0.009991944767534733, + 0.005258284509181976, + 0.009008158929646015, + 0.0030137801077216864, + 0.00220667477697134, + 0.004083177540451288, + 0.0026884018443524837, + 0.002780975541099906, + 0.01401037909090519, + 0.009864185005426407, + 0.011636829003691673, + 0.004896059166640043, + 0.0061371042393147945, + 0.00788791198283434, + 0.005826854612678289, + 0.006466631777584553, + 0.003921708557754755, + 0.003037489950656891, + 0.003931808285415173, + 0.0036808087024837732, + 0.002752850065007806, + 0.0026105577126145363, + 0.0024184836074709892, + 0.0019715302623808384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0004943578969687223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant red crab with hard chitin shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab's massive right pincer stretched high with a visible sharp, serrated inner edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the crab's beady black eyes visible on stalks gazing forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background blurred into a soft, painterly mix of aquamarine ocean and beige sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static low-angle macro view with a very shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040934826247394085 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040934826247394085, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040934826247394085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1504073143005371, + 0.07465683668851852, + 0.2228282243013382, + 0.19239240884780884, + 0.05814822390675545, + 0.1013534739613533, + 0.07568909972906113, + 0.03234194591641426, + 0.04246341437101364, + 0.049718961119651794 + ], + "credit_map": [ + 0.019156495109200478, + 0.024410318583250046, + 0.015383971855044365, + 0.0167078897356987, + 0.01275006402283907, + 0.017600826919078827, + 0.017411423847079277, + 0.02698633447289467, + 0.017039187252521515, + 0.014829814434051514, + 0.008890441618859768, + 0.009120709262788296, + 0.006697637960314751, + 0.006494104862213135, + 0.005563860293477774, + 0.006021095905452967, + 0.035669032484292984, + 0.08687446266412735, + 0.06823515892028809, + 0.00725843058899045, + 0.008335031569004059, + 0.005612003616988659, + 0.005130270961672068, + 0.005713861435651779, + 0.06211421638727188, + 0.041381556540727615, + 0.02795948088169098, + 0.015513450838625431, + 0.009517761878669262, + 0.011239683255553246, + 0.010504193603992462, + 0.014162089675664902, + 0.009569885209202766, + 0.008034481666982174, + 0.009581525810062885, + 0.006089541129767895, + 0.0057257250882685184, + 0.006544643547385931, + 0.00479059386998415, + 0.007811831776052713, + 0.016527781262993813, + 0.028796276077628136, + 0.024851815775036812, + 0.008671269752085209, + 0.0050786021165549755, + 0.005999073851853609, + 0.006152596790343523, + 0.005276070907711983, + 0.02116338349878788, + 0.010499461553990841, + 0.016090404242277145, + 0.007204732391983271, + 0.004579927772283554, + 0.004798635840415955, + 0.005602882243692875, + 0.005749677773565054, + 0.005923452787101269, + 0.005417444743216038, + 0.004751489032059908, + 0.002719490323215723, + 0.0027684695087373257, + 0.004822819493710995, + 0.0026784222573041916, + 0.003260362194851041, + 0.008532784879207611, + 0.005717047955840826, + 0.005318577867001295, + 0.0035949661396443844, + 0.004446355625987053, + 0.004553316626697779, + 0.0038896796759217978, + 0.006410690024495125, + 0.006857891567051411, + 0.005538822151720524, + 0.0057336390018463135, + 0.005888777319341898, + 0.007448225282132626, + 0.009625965729355812, + 0.004170148633420467, + 0.004455495625734329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0040934826247394085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_12/index_tensor([9620, 9620])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_12/index_tensor([9620, 9620])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..049a47ac18fd466952c434fb57cedc2e4928d6db --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_12/index_tensor([9620, 9620])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up tracking shot moves slowly and steadily with an elderly woman, her face a beautiful map of time, kind, wrinkled eyes, and a deep smile etched across her warm brown skin. Her silver-gray hair is neatly pulled back in a simple bun. She is draped in a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta, with a simple embroidered border. She stands in a sun-drenched room, brimming with rustic charm, its uneven, plastered earth walls gleaming in the golden light of late afternoon. A bright ray of sunlight filters through a small, unglazed window, illuminating countless motes of dust dancing in the air. With a practiced, deliberate grace, she holds a thin, dark incense stick between her thumb and index finger. Her other hand holds a lit match to the tip, the small flame casting a fleeting, flickering warmth across her focused features. She held it until the tip glowed a vibrant orange, then blew gently, extinguishing the flame and releasing a thin, fragrant stream of gray-white smoke that curled gracefully upward, catching the light as it rose into the sunlight, creating a serene and deeply spiritual atmosphere of quiet reverence.", + "seed": 15359, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_12/index_9620_seed_15359_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017925035394728183, + "vlm_logprob": -0.017925035394728183, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.017925035394728183, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009589205496013165 + ], + "opposite_logprob_tokens": [ + -4.884589195251465 + ], + "logprob_reward": -0.009589205496013165, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009589205496013165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21106897294521332, + 0.07172799855470657, + 0.09509952366352081, + 0.07222151756286621, + 0.07542560994625092, + 0.09610351175069809, + 0.1046772375702858, + 0.05863068997859955, + 0.06390233337879181, + 0.15114262700080872 + ], + "credit_map": [ + 0.02395232766866684, + 0.030946137383580208, + 0.024770814925432205, + 0.023058773949742317, + 0.0407695546746254, + 0.02244933694601059, + 0.01912655495107174, + 0.025995472446084023, + 0.011486434377729893, + 0.009119157679378986, + 0.00665978854522109, + 0.005344347096979618, + 0.009192809462547302, + 0.015824533998966217, + 0.007512391544878483, + 0.006588535849004984, + 0.012262224219739437, + 0.013590559363365173, + 0.021125881001353264, + 0.006501478608697653, + 0.011087693274021149, + 0.012609174475073814, + 0.007236455101519823, + 0.010686059482395649, + 0.010361714288592339, + 0.0060462444089353085, + 0.01945996657013893, + 0.006602051202207804, + 0.008912187069654465, + 0.007691355422139168, + 0.006587405223399401, + 0.006560593377798796, + 0.007762519642710686, + 0.00756486039608717, + 0.01270817406475544, + 0.009951786138117313, + 0.010164031758904457, + 0.006952264346182346, + 0.008936538361012936, + 0.01138543151319027, + 0.018119750544428825, + 0.007815566845238209, + 0.015259049832820892, + 0.015434181317687035, + 0.009290257468819618, + 0.006339802406728268, + 0.012200474739074707, + 0.011644427664577961, + 0.025640465319156647, + 0.01484015490859747, + 0.018254850059747696, + 0.00836045853793621, + 0.010749912820756435, + 0.008038878440856934, + 0.00738275982439518, + 0.011409763246774673, + 0.007556874305009842, + 0.009542806074023247, + 0.007465952076017857, + 0.006157468538731337, + 0.006639882456511259, + 0.005777389742434025, + 0.009295713156461716, + 0.006194604095071554, + 0.007946663536131382, + 0.009798328392207623, + 0.007278944831341505, + 0.008088215254247189, + 0.008715239353477955, + 0.008326518349349499, + 0.008130732923746109, + 0.0056176879443228245, + 0.016048628836870193, + 0.021128006279468536, + 0.014881200157105923, + 0.014951035380363464, + 0.031569723039865494, + 0.018383126705884933, + 0.01895064488053322, + 0.015230270102620125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.009589205496013165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603911951184273 + ], + "opposite_logprob_tokens": [ + -3.895603895187378 + ], + "logprob_reward": -0.020603911951184273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603911951184273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21519498527050018, + 0.03808623552322388, + 0.08547535538673401, + 0.03340107575058937, + 0.0467861033976078, + 0.20950883626937866, + 0.07115425914525986, + 0.04110744968056679, + 0.1167575791478157, + 0.14252817630767822 + ], + "credit_map": [ + 0.029852325096726418, + 0.023164181038737297, + 0.031678490340709686, + 0.020733477547764778, + 0.031270742416381836, + 0.03242921084165573, + 0.02246449701488018, + 0.02360205538570881, + 0.005708306562155485, + 0.005599269177764654, + 0.0057380665093660355, + 0.0027766672428697348, + 0.00543516268953681, + 0.004453796427696943, + 0.0035242345184087753, + 0.004850731231272221, + 0.013196351006627083, + 0.013469863682985306, + 0.016290320083498955, + 0.0061754146590828896, + 0.008743378333747387, + 0.01018782053142786, + 0.00860370509326458, + 0.008808499202132225, + 0.005500317085534334, + 0.0024386760778725147, + 0.006412738934159279, + 0.0033976947888731956, + 0.004495431203395128, + 0.004576585721224546, + 0.003138106781989336, + 0.0034415265545248985, + 0.005873466841876507, + 0.004702264908701181, + 0.007668335922062397, + 0.004482490010559559, + 0.006692436523735523, + 0.0042672050185501575, + 0.008125384338200092, + 0.004974521696567535, + 0.04184524714946747, + 0.018685905262827873, + 0.03137267753481865, + 0.030314834788441658, + 0.022262906655669212, + 0.013275230303406715, + 0.02747967466711998, + 0.024272361770272255, + 0.019089266657829285, + 0.008361120708286762, + 0.012012621387839317, + 0.005550905596464872, + 0.007609807886183262, + 0.005276599898934364, + 0.005445255432277918, + 0.007808685768395662, + 0.00727339182049036, + 0.007334718480706215, + 0.005346445832401514, + 0.0038932356983423233, + 0.0047846343368291855, + 0.0038969716988503933, + 0.004509603604674339, + 0.004068451467901468, + 0.016425184905529022, + 0.017959151417016983, + 0.013352624140679836, + 0.011206603609025478, + 0.017935648560523987, + 0.014514918439090252, + 0.01221433375030756, + 0.013149106875061989, + 0.016124291345477104, + 0.019676774740219116, + 0.015546108596026897, + 0.014502747915685177, + 0.02735680155456066, + 0.016587363556027412, + 0.016525931656360626, + 0.016208147630095482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.020603911951184273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741991266608238 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.006741991266608238, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741991266608238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10960958898067474, + 0.2797060012817383, + 0.14969968795776367, + 0.03268997371196747, + 0.03648439794778824, + 0.07104548066854477, + 0.10890968888998032, + 0.07650555670261383, + 0.04242314398288727, + 0.092926524579525 + ], + "credit_map": [ + 0.012798751704394817, + 0.01674947515130043, + 0.013927219435572624, + 0.011571265757083893, + 0.01595485582947731, + 0.014610125683248043, + 0.012479408644139767, + 0.01151848305016756, + 0.05984555184841156, + 0.06041669845581055, + 0.033201221376657486, + 0.017264388501644135, + 0.01589779369533062, + 0.05818180739879608, + 0.015163738280534744, + 0.019734784960746765, + 0.024198133498430252, + 0.02062905579805374, + 0.03660380467772484, + 0.007712011691182852, + 0.010753442533314228, + 0.022511936724185944, + 0.012571939267218113, + 0.014719363301992416, + 0.00675298972055316, + 0.0029229989740997553, + 0.007956484332680702, + 0.0019627620931714773, + 0.0039027624297887087, + 0.004154737573117018, + 0.0020288771484047174, + 0.0030083609744906425, + 0.003416758729144931, + 0.004165154416114092, + 0.004988938104361296, + 0.0038140129763633013, + 0.0053457776084542274, + 0.003354752669110894, + 0.006255228538066149, + 0.005143773276358843, + 0.013576299883425236, + 0.005979130044579506, + 0.010877625085413456, + 0.010514499619603157, + 0.007423346396535635, + 0.0043165418319404125, + 0.009511382319033146, + 0.008846653625369072, + 0.03817032277584076, + 0.011191402561962605, + 0.022929172962903976, + 0.005912806373089552, + 0.009132836014032364, + 0.008956734091043472, + 0.005902678705751896, + 0.0067137302830815315, + 0.011764171533286572, + 0.01927502639591694, + 0.008310331963002682, + 0.008394206874072552, + 0.007487195543944836, + 0.00838468223810196, + 0.008325682021677494, + 0.004564258269965649, + 0.005935871973633766, + 0.008910033851861954, + 0.00475065503269434, + 0.002885664813220501, + 0.007157334126532078, + 0.005587364546954632, + 0.003441202687099576, + 0.0037550157867372036, + 0.013601218350231647, + 0.01713406853377819, + 0.013535709120333195, + 0.007680708076804876, + 0.016701141372323036, + 0.010390021838247776, + 0.006876826751977205, + 0.007006833329796791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.006741991266608238 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347588747739792 + ], + "opposite_logprob_tokens": [ + -3.6513476371765137 + ], + "logprob_reward": -0.026347588747739792, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347588747739792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.088679738342762, + 0.09346608817577362, + 0.03317423537373543, + 0.04595418646931648, + 0.03586948662996292, + 0.03469888120889664, + 0.02929842285811901, + 0.3303084969520569, + 0.2498534470796585, + 0.058696944266557693 + ], + "credit_map": [ + 0.007983984425663948, + 0.00899288710206747, + 0.01062839012593031, + 0.008529831655323505, + 0.01837645284831524, + 0.010424066334962845, + 0.012417154386639595, + 0.011326972395181656, + 0.01976361684501171, + 0.011145653203129768, + 0.008028190582990646, + 0.006311654578894377, + 0.00812473427504301, + 0.027269288897514343, + 0.006744593381881714, + 0.006078353617340326, + 0.0034472451079636812, + 0.006001691799610853, + 0.007002617698162794, + 0.0025250911712646484, + 0.00361711741425097, + 0.00581870973110199, + 0.0022834937553852797, + 0.002478271024301648, + 0.008899491280317307, + 0.0028942504432052374, + 0.013621400110423565, + 0.00285511277616024, + 0.0037444704212248325, + 0.0062940968200564384, + 0.003385880496352911, + 0.00425948528572917, + 0.002788166282698512, + 0.0027265746612101793, + 0.004617196973413229, + 0.005833630915731192, + 0.0038157880771905184, + 0.0033541417215019464, + 0.005999118089675903, + 0.006734867580235004, + 0.007677797228097916, + 0.003944252151995897, + 0.006197087001055479, + 0.002617310034111142, + 0.004156281240284443, + 0.0022556332405656576, + 0.003485775087028742, + 0.004364747554063797, + 0.008171960711479187, + 0.0024341975804418325, + 0.005909434985369444, + 0.0020140649285167456, + 0.0024930492509156466, + 0.0023660294245928526, + 0.003617035923525691, + 0.0022926493547856808, + 0.04463370144367218, + 0.08291833102703094, + 0.031117206439375877, + 0.03771471977233887, + 0.03345559909939766, + 0.03833547979593277, + 0.03792647272348404, + 0.024206988513469696, + 0.041351113468408585, + 0.04050246626138687, + 0.03811280429363251, + 0.023062199354171753, + 0.029076535254716873, + 0.03326616808772087, + 0.02445964515209198, + 0.020022515207529068, + 0.007235756143927574, + 0.011817436665296555, + 0.0076994397677481174, + 0.006420953664928675, + 0.009470340795814991, + 0.00620475597679615, + 0.00493946997448802, + 0.004908793605864048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.026347588747739792 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634247951209545 + ], + "opposite_logprob_tokens": [ + -3.6513423919677734 + ], + "logprob_reward": -0.02634247951209545, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634247951209545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12429401278495789, + 0.1877405047416687, + 0.11072523891925812, + 0.05864924564957619, + 0.053680144250392914, + 0.2007579356431961, + 0.10444227606058121, + 0.05921550840139389, + 0.06974263489246368, + 0.03075251169502735 + ], + "credit_map": [ + 0.01689378172159195, + 0.017225084826350212, + 0.018667031079530716, + 0.010700634680688381, + 0.015005730092525482, + 0.021659892052412033, + 0.011472074314951897, + 0.012669782154262066, + 0.04631044343113899, + 0.020505394786596298, + 0.019330354407429695, + 0.011768114753067493, + 0.013193659484386444, + 0.05514809861779213, + 0.010211191140115261, + 0.011273253709077835, + 0.015601658262312412, + 0.014985497109591961, + 0.029025593772530556, + 0.006116003729403019, + 0.01044474821537733, + 0.017742471769452095, + 0.006715231575071812, + 0.010094037279486656, + 0.010431481525301933, + 0.0035140630789101124, + 0.014404581859707832, + 0.005459378007799387, + 0.005758346524089575, + 0.007053954526782036, + 0.0061285183764994144, + 0.005898922681808472, + 0.007303723134100437, + 0.004841281566768885, + 0.010509318672120571, + 0.006280557252466679, + 0.007994926534593105, + 0.005247192922979593, + 0.006438404321670532, + 0.005064745433628559, + 0.02441655844449997, + 0.021929804235696793, + 0.026372071355581284, + 0.03249940276145935, + 0.028163211420178413, + 0.01251722127199173, + 0.02906632237136364, + 0.025793354958295822, + 0.019672272726893425, + 0.015385467559099197, + 0.016774334013462067, + 0.00857463013380766, + 0.01149718090891838, + 0.008961814455688, + 0.0093235457316041, + 0.014253035187721252, + 0.009174865670502186, + 0.01085847057402134, + 0.006259080953896046, + 0.007009764201939106, + 0.007136509288102388, + 0.005910372361540794, + 0.007934345863759518, + 0.004932102747261524, + 0.009735113009810448, + 0.010053936392068863, + 0.010947585105895996, + 0.00668718945235014, + 0.008381430059671402, + 0.008092748001217842, + 0.00797185953706503, + 0.00787277426570654, + 0.004239142406731844, + 0.0055226958356797695, + 0.0034002759493887424, + 0.003307341132313013, + 0.004853839986026287, + 0.0032154452055692673, + 0.003083835355937481, + 0.0031299355905503035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.02634247951209545 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up tracking shot moves slowly and steadily with an elderly woman, her face a beautiful map of time, kind, wrinkled eyes, and a deep smile etched across her warm brown skin. Her silver-gray hair is neatly pulled back in a simple bun. She is draped in a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta, with a simple embroidered border. She stands in a sun-drenched room, brimming with rustic charm, its uneven, plastered earth walls gleaming in the golden light of late afternoon. A bright ray of sunlight filters through a small, unglazed window, illuminating countless motes of dust dancing in the air. With a practiced, deliberate grace, she holds a thin, dark incense stick between her thumb and index finger. Her other hand holds a lit match to the tip, the small flame casting a fleeting, flickering warmth across her focused features. She held it until the tip glowed a vibrant orange, then blew gently, extinguishing the flame and releasing a thin, fragrant stream of gray-white smoke that curled gracefully upward, catching the light as it rose into the sunlight, creating a serene and deeply spiritual atmosphere of quiet reverence.", + "seed": 15360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_12/index_9620_seed_15360_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03664282504469156, + "vlm_logprob": -0.03664282504469156, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03664282504469156, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01592911221086979 + ], + "opposite_logprob_tokens": [ + -4.265929222106934 + ], + "logprob_reward": -0.01592911221086979, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01592911221086979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1332443803548813, + 0.07146311551332474, + 0.0361117348074913, + 0.03472569212317467, + 0.03579007834196091, + 0.2652197480201721, + 0.25816255807876587, + 0.035237718373537064, + 0.056441839784383774, + 0.07360310107469559 + ], + "credit_map": [ + 0.033366765826940536, + 0.04302535206079483, + 0.019224172458052635, + 0.00840296596288681, + 0.008521470241248608, + 0.007027277257293463, + 0.006572490092366934, + 0.007103890646249056, + 0.018037935718894005, + 0.010004266165196896, + 0.018758093938231468, + 0.0033898469991981983, + 0.0070448946207761765, + 0.004817608743906021, + 0.00604987982660532, + 0.0033605843782424927, + 0.0036174808628857136, + 0.0043998598121106625, + 0.003951352555304766, + 0.0024789636954665184, + 0.007842487655580044, + 0.004229939542710781, + 0.006862394977360964, + 0.002729256870225072, + 0.00400833273306489, + 0.004845932591706514, + 0.005814330652356148, + 0.0033622472546994686, + 0.005561926402151585, + 0.0030054058879613876, + 0.004671432077884674, + 0.00345608196221292, + 0.004420274868607521, + 0.0040448871441185474, + 0.006155859213322401, + 0.0024519609287381172, + 0.0033647422678768635, + 0.007697656285017729, + 0.005395608022809029, + 0.002259093103930354, + 0.09000331163406372, + 0.015625352039933205, + 0.06614458560943604, + 0.020170120522379875, + 0.02104213275015354, + 0.019480908289551735, + 0.017199650406837463, + 0.015553701668977737, + 0.10241086781024933, + 0.015196460299193859, + 0.08208820968866348, + 0.014863944612443447, + 0.012260109186172485, + 0.01338441763073206, + 0.012052183039486408, + 0.00590637419372797, + 0.006404650863260031, + 0.0075652990490198135, + 0.0048725116066634655, + 0.0023307183291763067, + 0.004644482396543026, + 0.00480402959510684, + 0.002361488761380315, + 0.002254538470879197, + 0.007858323864638805, + 0.003984999377280474, + 0.00638870382681489, + 0.005610316526144743, + 0.012206906452775002, + 0.004461643751710653, + 0.00810026004910469, + 0.007830691523849964, + 0.01104476023465395, + 0.009680753573775291, + 0.01441987045109272, + 0.005859877914190292, + 0.011832762509584427, + 0.0073296804912388325, + 0.0076217553578317165, + 0.0058136447332799435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.01592911221086979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026374757289886475 + ], + "opposite_logprob_tokens": [ + -3.6513748168945312 + ], + "logprob_reward": -0.026374757289886475, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026374757289886475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20086082816123962, + 0.06755221635103226, + 0.06869123876094818, + 0.08935823291540146, + 0.050442203879356384, + 0.04347466677427292, + 0.32267194986343384, + 0.03528805077075958, + 0.05923556908965111, + 0.06242503225803375 + ], + "credit_map": [ + 0.038576625287532806, + 0.052285224199295044, + 0.031182395294308662, + 0.01858588308095932, + 0.01563701592385769, + 0.011788271367549896, + 0.019211724400520325, + 0.013593699783086777, + 0.014615016058087349, + 0.007418415043503046, + 0.013055125251412392, + 0.0038568030577152967, + 0.008308613672852516, + 0.006876252591609955, + 0.00903239380568266, + 0.004389592446386814, + 0.007837015204131603, + 0.007524688728153706, + 0.006612948141992092, + 0.0072868806309998035, + 0.012536587193608284, + 0.0065877484157681465, + 0.015167523175477982, + 0.00513784633949399, + 0.00915574375540018, + 0.011589575558900833, + 0.010221930220723152, + 0.015492906793951988, + 0.012893560342490673, + 0.008396930061280727, + 0.01608859747648239, + 0.005518991965800524, + 0.019117340445518494, + 0.004178991541266441, + 0.011358105577528477, + 0.0016833023400977254, + 0.0027429775800555944, + 0.006227163132280111, + 0.0031134251039475203, + 0.0020208940841257572, + 0.004405885003507137, + 0.005480584222823381, + 0.008785564452409744, + 0.004217828158289194, + 0.006511667277663946, + 0.005191636737436056, + 0.005217274650931358, + 0.0036642260383814573, + 0.13208112120628357, + 0.017497491091489792, + 0.10634057968854904, + 0.018064331263303757, + 0.010454073548316956, + 0.01743181049823761, + 0.013313885778188705, + 0.007488669361919165, + 0.005485183093696833, + 0.0060930801555514336, + 0.004389066249132156, + 0.0032926297280937433, + 0.005663196090608835, + 0.004176581744104624, + 0.0035908843856304884, + 0.002597429323941469, + 0.009008730761706829, + 0.009267177432775497, + 0.011642294935882092, + 0.004778070375323296, + 0.007856368087232113, + 0.0061721294187009335, + 0.005735091399401426, + 0.004775703884661198, + 0.012135605327785015, + 0.005538390949368477, + 0.01749890297651291, + 0.005259606521576643, + 0.0071212537586688995, + 0.005686074495315552, + 0.005091905128210783, + 0.004093291237950325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.026374757289886475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978813599795103 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.00978813599795103, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978813599795103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06607294827699661, + 0.21983207762241364, + 0.015852004289627075, + 0.020231066271662712, + 0.14430300891399384, + 0.03998437523841858, + 0.19685056805610657, + 0.1424015909433365, + 0.06617654860019684, + 0.08829578757286072 + ], + "credit_map": [ + 0.006169250700622797, + 0.013631426729261875, + 0.02021198906004429, + 0.00472621014341712, + 0.006061806809157133, + 0.006001120898872614, + 0.004762858152389526, + 0.004508279263973236, + 0.06678707897663116, + 0.03031783737242222, + 0.07490324974060059, + 0.006438940763473511, + 0.01188097707927227, + 0.012435573153197765, + 0.01023098174482584, + 0.006837426219135523, + 0.002400698373094201, + 0.0022371537052094936, + 0.002689426066353917, + 0.0011452321195974946, + 0.0021490503568202257, + 0.0019571695011109114, + 0.0019358807476237416, + 0.0013373919064179063, + 0.0023317928425967693, + 0.002469986444339156, + 0.002651250921189785, + 0.001773323048837483, + 0.003053501946851611, + 0.0015067997155711055, + 0.00462442496791482, + 0.0018199831247329712, + 0.014530380256474018, + 0.02334410510957241, + 0.042993441224098206, + 0.008054491132497787, + 0.011419299989938736, + 0.02669728919863701, + 0.009946883656084538, + 0.007317099720239639, + 0.010419230908155441, + 0.003073963103815913, + 0.008204271085560322, + 0.0033308882266283035, + 0.003488982794806361, + 0.0067047784104943275, + 0.0028452076949179173, + 0.0019170460291206837, + 0.07936736196279526, + 0.010755731724202633, + 0.06558922678232193, + 0.011553983204066753, + 0.006012663710862398, + 0.010479976423084736, + 0.008710861206054688, + 0.004380750469863415, + 0.015851661562919617, + 0.03236799314618111, + 0.013816412538290024, + 0.011199112050235271, + 0.029987098649144173, + 0.01733946055173874, + 0.012796127237379551, + 0.009043714962899685, + 0.01442051026970148, + 0.010682330466806889, + 0.011989855207502842, + 0.004042260814458132, + 0.005583076737821102, + 0.00947913620620966, + 0.004882712848484516, + 0.005096661392599344, + 0.016073275357484818, + 0.007521231193095446, + 0.029631074517965317, + 0.0064526633359491825, + 0.006359614431858063, + 0.01251782476902008, + 0.005093706771731377, + 0.004646382760256529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.00978813599795103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293458193540573 + ], + "opposite_logprob_tokens": [ + -2.2379345893859863 + ], + "logprob_reward": -0.11293458193540573, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293458193540573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061667270958423615, + 0.1824856549501419, + 0.020669903606176376, + 0.01780228316783905, + 0.042460326105356216, + 0.05917457863688469, + 0.05898623168468475, + 0.42922401428222656, + 0.04356253892183304, + 0.0839671939611435 + ], + "credit_map": [ + 0.015629874542355537, + 0.020767131820321083, + 0.009460694156587124, + 0.0032168577890843153, + 0.0029944349080324173, + 0.003537603188306093, + 0.003216692479327321, + 0.002843985566869378, + 0.04668636620044708, + 0.031158898025751114, + 0.06974505633115768, + 0.0050548287108540535, + 0.007569608278572559, + 0.01161972340196371, + 0.006614094600081444, + 0.004037083592265844, + 0.002992852358147502, + 0.003140044864267111, + 0.0033495647367089987, + 0.0017430498264729977, + 0.0026750739198178053, + 0.002058029640465975, + 0.0027112201787531376, + 0.0020000687800347805, + 0.002355745295062661, + 0.002604815410450101, + 0.0033096123952418566, + 0.001466952497139573, + 0.0030197259038686752, + 0.0014494493370875716, + 0.002477237954735756, + 0.0011187430936843157, + 0.003608784871175885, + 0.004696513991802931, + 0.006864861585199833, + 0.003261975245550275, + 0.004410248249769211, + 0.006095608696341515, + 0.009954208508133888, + 0.003568127751350403, + 0.011174498125910759, + 0.006259637884795666, + 0.008032710291445255, + 0.008660796098411083, + 0.00627188989892602, + 0.007770742289721966, + 0.006776121910661459, + 0.004228181671351194, + 0.019344858825206757, + 0.004464244470000267, + 0.0168754979968071, + 0.0034789859782904387, + 0.0046641337685287, + 0.002969152992591262, + 0.004317838698625565, + 0.0028715208172798157, + 0.07381677627563477, + 0.11016780138015747, + 0.053827110677957535, + 0.027520323172211647, + 0.06363797932863235, + 0.0488724410533905, + 0.028213003650307655, + 0.02316858433187008, + 0.004912434611469507, + 0.005635831505060196, + 0.010661733336746693, + 0.002982453675940633, + 0.0058290413580834866, + 0.005233295261859894, + 0.0037210723385214806, + 0.004586678463965654, + 0.017479771748185158, + 0.014786171726882458, + 0.012881804257631302, + 0.007630782667547464, + 0.009614663198590279, + 0.006590520963072777, + 0.007091565523296595, + 0.007891914807260036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.11293458193540573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187537789344788 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187537789344788, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187537789344788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06422848999500275, + 0.06680755317211151, + 0.02871854603290558, + 0.03619972988963127, + 0.07806479185819626, + 0.17399358749389648, + 0.3480582535266876, + 0.12251723557710648, + 0.029941467568278313, + 0.0514703132212162 + ], + "credit_map": [ + 0.006983302067965269, + 0.015380535274744034, + 0.018567269667983055, + 0.0035465636756271124, + 0.006518849171698093, + 0.0054230582900345325, + 0.003931919578462839, + 0.003876995062455535, + 0.0070471689105033875, + 0.009679729118943214, + 0.01452930923551321, + 0.00431705079972744, + 0.010089396499097347, + 0.007248118054121733, + 0.008562140166759491, + 0.005334641318768263, + 0.0035700697917491198, + 0.004230410326272249, + 0.00552765466272831, + 0.0022742468863725662, + 0.003947870340198278, + 0.0034122455399483442, + 0.0033045795280486345, + 0.0024514677934348583, + 0.004609130322933197, + 0.0057395221665501595, + 0.005840052384883165, + 0.0033641443587839603, + 0.005576764233410358, + 0.003936427179723978, + 0.004412000998854637, + 0.002721689408645034, + 0.029240760952234268, + 0.0066285692155361176, + 0.009056326001882553, + 0.0023101523984223604, + 0.003445640904828906, + 0.018828699365258217, + 0.0057567716576159, + 0.0027978690341115, + 0.05788450688123703, + 0.011730076745152473, + 0.0401192232966423, + 0.015472388826310635, + 0.012847357429564, + 0.013344290666282177, + 0.011377722956240177, + 0.011218022555112839, + 0.12934187054634094, + 0.02030123397707939, + 0.10918854176998138, + 0.01960861124098301, + 0.01910746470093727, + 0.019087625667452812, + 0.020031852647662163, + 0.011391044594347477, + 0.023461198434233665, + 0.0308750681579113, + 0.01555113960057497, + 0.007579978555440903, + 0.016131294891238213, + 0.013307150453329086, + 0.008846932090818882, + 0.006764472462236881, + 0.003453637706115842, + 0.005200356710702181, + 0.0041631232015788555, + 0.002672821981832385, + 0.004522261675447226, + 0.003050424624234438, + 0.0038460525684058666, + 0.003032790729776025, + 0.0055065336637198925, + 0.0045714606530964375, + 0.0067053791135549545, + 0.005161370150744915, + 0.0100425248965621, + 0.007820517756044865, + 0.007056597154587507, + 0.004605927970260382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.018187537789344788 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up tracking shot moves slowly and steadily with an elderly woman, her face a beautiful map of time, kind, wrinkled eyes, and a deep smile etched across her warm brown skin. Her silver-gray hair is neatly pulled back in a simple bun. She is draped in a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta, with a simple embroidered border. She stands in a sun-drenched room, brimming with rustic charm, its uneven, plastered earth walls gleaming in the golden light of late afternoon. A bright ray of sunlight filters through a small, unglazed window, illuminating countless motes of dust dancing in the air. With a practiced, deliberate grace, she holds a thin, dark incense stick between her thumb and index finger. Her other hand holds a lit match to the tip, the small flame casting a fleeting, flickering warmth across her focused features. She held it until the tip glowed a vibrant orange, then blew gently, extinguishing the flame and releasing a thin, fragrant stream of gray-white smoke that curled gracefully upward, catching the light as it rose into the sunlight, creating a serene and deeply spiritual atmosphere of quiet reverence.", + "seed": 15361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_12/index_9620_seed_15361_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0378532350063324, + "vlm_logprob": -0.0378532350063324, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.0378532350063324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014265462756156921 + ], + "opposite_logprob_tokens": [ + -4.389265537261963 + ], + "logprob_reward": -0.014265462756156921, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014265462756156921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34795886278152466, + 0.04144826903939247, + 0.037601206451654434, + 0.08806277066469193, + 0.03550799563527107, + 0.07862430810928345, + 0.1599920392036438, + 0.0702102780342102, + 0.09219489246606827, + 0.0483994223177433 + ], + "credit_map": [ + 0.10346114635467529, + 0.021556658670306206, + 0.10127232223749161, + 0.03555426746606827, + 0.025416813790798187, + 0.01129225268959999, + 0.029357748106122017, + 0.020047660917043686, + 0.008520452305674553, + 0.004144046455621719, + 0.007858402095735073, + 0.005020275246351957, + 0.0037326968740671873, + 0.002645036904141307, + 0.0051461742259562016, + 0.004381184466183186, + 0.0066838436760008335, + 0.0036573195829987526, + 0.006392028648406267, + 0.005079980939626694, + 0.0032790338154882193, + 0.004435475915670395, + 0.004980860743671656, + 0.0030926603358238935, + 0.013453155755996704, + 0.012101614847779274, + 0.01352986041456461, + 0.016559604555368423, + 0.007431482896208763, + 0.007390868384391069, + 0.009216801263391972, + 0.008379381150007248, + 0.006137828808277845, + 0.00243722484447062, + 0.007435801904648542, + 0.005117988213896751, + 0.003262067912146449, + 0.0045869601890444756, + 0.0038748823571950197, + 0.002655241871252656, + 0.02137414738535881, + 0.004957013763487339, + 0.02445054054260254, + 0.007145811803638935, + 0.003915705718100071, + 0.00681693060323596, + 0.00569168059155345, + 0.0042724767699837685, + 0.01991330273449421, + 0.020244508981704712, + 0.020713815465569496, + 0.030916815623641014, + 0.014403502456843853, + 0.015228056348860264, + 0.0217883437871933, + 0.016783680766820908, + 0.01072745956480503, + 0.013086442835628986, + 0.008134115487337112, + 0.008673977106809616, + 0.005821459926664829, + 0.011375860311090946, + 0.006127545144408941, + 0.006263419054448605, + 0.015342158265411854, + 0.008652478456497192, + 0.017656320706009865, + 0.009012211114168167, + 0.0065634120255708694, + 0.015641558915376663, + 0.009159221313893795, + 0.010167536325752735, + 0.007799758575856686, + 0.0030796851497143507, + 0.009865946136415005, + 0.0044224439188838005, + 0.0027930389624089003, + 0.011138560250401497, + 0.004273097962141037, + 0.005026891361922026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.014265462756156921 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594919100403786 + ], + "opposite_logprob_tokens": [ + -3.895594835281372 + ], + "logprob_reward": -0.020594919100403786, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594919100403786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.270648330450058, + 0.14202360808849335, + 0.03230404853820801, + 0.11894160509109497, + 0.017521869391202927, + 0.026273705065250397, + 0.03233535215258598, + 0.023375622928142548, + 0.06460419297218323, + 0.2719716727733612 + ], + "credit_map": [ + 0.09006786346435547, + 0.01566884107887745, + 0.08857020735740662, + 0.0266373660415411, + 0.01412669476121664, + 0.0075639584101736546, + 0.015752363950014114, + 0.012261027470231056, + 0.029403269290924072, + 0.01500985212624073, + 0.034329526126384735, + 0.014916355721652508, + 0.010088124312460423, + 0.011093024164438248, + 0.013235853984951973, + 0.013947594910860062, + 0.005304650869220495, + 0.0038739433512091637, + 0.004640624392777681, + 0.004013604484498501, + 0.0021017349790781736, + 0.004687745124101639, + 0.004381160251796246, + 0.003300586948171258, + 0.032406553626060486, + 0.00849465373903513, + 0.030586879700422287, + 0.018727023154497147, + 0.005924151744693518, + 0.009150947444140911, + 0.008608249016106129, + 0.005043147597461939, + 0.004784329328685999, + 0.0009381304844282568, + 0.004339767619967461, + 0.0024851071648299694, + 0.0013207142474129796, + 0.0015608135145157576, + 0.001190696842968464, + 0.000902310130186379, + 0.007466230541467667, + 0.0014830309664830565, + 0.008246132172644138, + 0.002099316567182541, + 0.0017120888223871589, + 0.0018802769482135773, + 0.0020475939381867647, + 0.0013390358071774244, + 0.004700300749391317, + 0.002828337252140045, + 0.005314756650477648, + 0.003141105640679598, + 0.002843090333044529, + 0.00343907973729074, + 0.0039207651279866695, + 0.006147916428744793, + 0.003239182522520423, + 0.0026559235993772745, + 0.003447228344157338, + 0.0032861216459423304, + 0.0015722746029496193, + 0.003801607061177492, + 0.002396564232185483, + 0.0029767199885100126, + 0.011491653509438038, + 0.005460962653160095, + 0.01162539329379797, + 0.00568811409175396, + 0.0032656071707606316, + 0.00959774013608694, + 0.006405156571418047, + 0.011069568805396557, + 0.014545811340212822, + 0.010104955174028873, + 0.09385910630226135, + 0.013833113946020603, + 0.008475951850414276, + 0.10862879455089569, + 0.006621403153985739, + 0.01590253971517086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.020594919100403786 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094313830137253 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094313830137253, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094313830137253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13463576138019562, + 0.1405084878206253, + 0.07201272249221802, + 0.08358044177293777, + 0.058436665683984756, + 0.045342933386564255, + 0.04199429228901863, + 0.02661176770925522, + 0.03556960076093674, + 0.3613073229789734 + ], + "credit_map": [ + 0.03437351435422897, + 0.013489674776792526, + 0.035967033356428146, + 0.012348313815891743, + 0.009055816568434238, + 0.007440103683620691, + 0.013072051107883453, + 0.008889244869351387, + 0.022327670827507973, + 0.017724260687828064, + 0.027126001194119453, + 0.018502742052078247, + 0.011606888845562935, + 0.011924061924219131, + 0.01893194578588009, + 0.01236492209136486, + 0.01236727274954319, + 0.008119110949337482, + 0.013122571632266045, + 0.006484233308583498, + 0.005219509359449148, + 0.01656542532145977, + 0.005650700069963932, + 0.004483897238969803, + 0.02478311024606228, + 0.00605992553755641, + 0.026210837066173553, + 0.006792523432523012, + 0.0028644041158258915, + 0.007100716233253479, + 0.0051210178062319756, + 0.004647902213037014, + 0.012966640293598175, + 0.0036015049554407597, + 0.013069785200059414, + 0.00579878780990839, + 0.003428416093811393, + 0.0109034962952137, + 0.003693315898999572, + 0.004974720533937216, + 0.015239554457366467, + 0.0017635071417316794, + 0.01680643856525421, + 0.0035877451300621033, + 0.001698991283774376, + 0.0023740974720567465, + 0.0025232466869056225, + 0.0013493502046912909, + 0.00503290630877018, + 0.0050530508160591125, + 0.005120419897139072, + 0.009560954757034779, + 0.0033751209266483784, + 0.003957278095185757, + 0.006532603409141302, + 0.0033619573805481195, + 0.003159712301567197, + 0.004089020658284426, + 0.0034541585482656956, + 0.003752070479094982, + 0.0027449820190668106, + 0.00266138999722898, + 0.0033524129539728165, + 0.003398021450266242, + 0.006567256059497595, + 0.002922294894233346, + 0.0065478673204779625, + 0.00352674862369895, + 0.0021495018154382706, + 0.0057417950592935085, + 0.0033027944155037403, + 0.004811342339962721, + 0.017970670014619827, + 0.012502165511250496, + 0.1270674467086792, + 0.018310746178030968, + 0.009360837750136852, + 0.14823926985263824, + 0.006586723495274782, + 0.021269479766488075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.004094313830137253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270306259393692 + ], + "opposite_logprob_tokens": [ + -2.127030611038208 + ], + "logprob_reward": -0.1270306259393692, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270306259393692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18182122707366943, + 0.05349012091755867, + 0.06373585760593414, + 0.04323941096663475, + 0.031495917588472366, + 0.048352766782045364, + 0.14895522594451904, + 0.05249474197626114, + 0.18205049633979797, + 0.19436421990394592 + ], + "credit_map": [ + 0.05319574102759361, + 0.012006931938230991, + 0.05078823119401932, + 0.016944579780101776, + 0.013263104483485222, + 0.008414539508521557, + 0.015427611768245697, + 0.011780506931245327, + 0.010813084430992603, + 0.004948379937559366, + 0.009993797168135643, + 0.007390919607132673, + 0.005019754637032747, + 0.0034134851302951574, + 0.005344420205801725, + 0.006566285155713558, + 0.01022227481007576, + 0.00721701979637146, + 0.009650129824876785, + 0.007994200102984905, + 0.005810435861349106, + 0.011781918816268444, + 0.0055836522951722145, + 0.005476227030158043, + 0.0070444499142467976, + 0.003254575189203024, + 0.008405477739870548, + 0.008224211633205414, + 0.0020735892467200756, + 0.005401212722063065, + 0.003583533689379692, + 0.005252365488559008, + 0.003945524338632822, + 0.002910267561674118, + 0.0040361215360462666, + 0.00609350623562932, + 0.002138815587386489, + 0.004507595673203468, + 0.003120951820164919, + 0.004743137396872044, + 0.010167934000492096, + 0.0055025783367455006, + 0.00973964761942625, + 0.005976361688226461, + 0.0034961693454533815, + 0.0050677089020609856, + 0.004781989846378565, + 0.0036203772760927677, + 0.01900516264140606, + 0.018447011709213257, + 0.020473772659897804, + 0.030134325847029686, + 0.010908540338277817, + 0.01763172075152397, + 0.0206771157681942, + 0.011677578091621399, + 0.007027165964245796, + 0.008284063078463078, + 0.008854295127093792, + 0.005424498580396175, + 0.004863039590418339, + 0.006246442440897226, + 0.005290737841278315, + 0.006504503078758717, + 0.03837719187140465, + 0.016736937686800957, + 0.032630741596221924, + 0.015101350843906403, + 0.013616514392197132, + 0.030206283554434776, + 0.017789341509342194, + 0.017592154443264008, + 0.010546635836362839, + 0.0077189248986542225, + 0.06414563953876495, + 0.01071242056787014, + 0.008202261291444302, + 0.07488668709993362, + 0.006824430078268051, + 0.011327225714921951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.1270306259393692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280853405594826 + ], + "opposite_logprob_tokens": [ + -3.7732808589935303 + ], + "logprob_reward": -0.023280853405594826, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280853405594826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19592152535915375, + 0.12018448114395142, + 0.06349215656518936, + 0.09140404313802719, + 0.07101261615753174, + 0.0742136612534523, + 0.11477173864841461, + 0.11581123620271683, + 0.061770517379045486, + 0.09141796082258224 + ], + "credit_map": [ + 0.04136791080236435, + 0.022116977721452713, + 0.04226624220609665, + 0.027709774672985077, + 0.01934337429702282, + 0.011090307496488094, + 0.018783748149871826, + 0.013243199326097965, + 0.02518574520945549, + 0.008362889289855957, + 0.021637296304106712, + 0.013021932914853096, + 0.010514778085052967, + 0.007920414209365845, + 0.014344296418130398, + 0.01919715665280819, + 0.010805701836943626, + 0.005372549872845411, + 0.01165216788649559, + 0.007104632444679737, + 0.004891447722911835, + 0.011664771474897861, + 0.0059518455527722836, + 0.006049044895917177, + 0.01742505095899105, + 0.0100522106513381, + 0.017251206561923027, + 0.016327975317835808, + 0.00559629499912262, + 0.00720971217378974, + 0.008764286525547504, + 0.008777313865721226, + 0.02322390489280224, + 0.003795402590185404, + 0.021926138550043106, + 0.006948035676032305, + 0.0029393709264695644, + 0.003282773308455944, + 0.004074421711266041, + 0.004822573158890009, + 0.013994523324072361, + 0.005738015752285719, + 0.014303814619779587, + 0.010693894699215889, + 0.005571292247623205, + 0.010694876313209534, + 0.008452187292277813, + 0.004765065386891365, + 0.01390008069574833, + 0.015619173645973206, + 0.012872503139078617, + 0.02360091358423233, + 0.009572582319378853, + 0.0116569260135293, + 0.014781522564589977, + 0.01276804693043232, + 0.014116787351667881, + 0.023985154926776886, + 0.0158100426197052, + 0.011174346320331097, + 0.012187555432319641, + 0.014609993435442448, + 0.010966363362967968, + 0.012961006723344326, + 0.007251909002661705, + 0.004717941861599684, + 0.00806775875389576, + 0.007413029205054045, + 0.004157166928052902, + 0.010062675923109055, + 0.010170869529247284, + 0.009929174557328224, + 0.005938234273344278, + 0.007102189585566521, + 0.024533824995160103, + 0.007210753858089447, + 0.005259133875370026, + 0.02817326784133911, + 0.0065636951476335526, + 0.006636868696659803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.023280853405594826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up tracking shot moves slowly and steadily with an elderly woman, her face a beautiful map of time, kind, wrinkled eyes, and a deep smile etched across her warm brown skin. Her silver-gray hair is neatly pulled back in a simple bun. She is draped in a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta, with a simple embroidered border. She stands in a sun-drenched room, brimming with rustic charm, its uneven, plastered earth walls gleaming in the golden light of late afternoon. A bright ray of sunlight filters through a small, unglazed window, illuminating countless motes of dust dancing in the air. With a practiced, deliberate grace, she holds a thin, dark incense stick between her thumb and index finger. Her other hand holds a lit match to the tip, the small flame casting a fleeting, flickering warmth across her focused features. She held it until the tip glowed a vibrant orange, then blew gently, extinguishing the flame and releasing a thin, fragrant stream of gray-white smoke that curled gracefully upward, catching the light as it rose into the sunlight, creating a serene and deeply spiritual atmosphere of quiet reverence.", + "seed": 15362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_12/index_9620_seed_15362_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027652565762400626, + "vlm_logprob": -0.027652565762400626, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.027652565762400626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019272571429610252 + ], + "opposite_logprob_tokens": [ + -4.144272804260254 + ], + "logprob_reward": -0.019272571429610252, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019272571429610252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08813442289829254, + 0.11901775747537613, + 0.07030041515827179, + 0.02509867213666439, + 0.19739539921283722, + 0.030603546649217606, + 0.08433274179697037, + 0.07410111278295517, + 0.057025808840990067, + 0.253990113735199 + ], + "credit_map": [ + 0.01496822852641344, + 0.014470056630671024, + 0.020012201741337776, + 0.009366300888359547, + 0.006182767450809479, + 0.008481274358928204, + 0.008537836372852325, + 0.006115758791565895, + 0.022303489968180656, + 0.021982844918966293, + 0.042799532413482666, + 0.007115926127880812, + 0.004078240133821964, + 0.009690947830677032, + 0.006055292673408985, + 0.004991491325199604, + 0.015791095793247223, + 0.012349645607173443, + 0.011809619143605232, + 0.007635756395757198, + 0.004953185096383095, + 0.005765576846897602, + 0.004543945658951998, + 0.007451592944562435, + 0.0036781029775738716, + 0.003394881496205926, + 0.00497090257704258, + 0.003564173821359873, + 0.0017702126642689109, + 0.002590407617390156, + 0.0030503906309604645, + 0.002079601399600506, + 0.03584306687116623, + 0.029093747958540916, + 0.03388110548257828, + 0.014325258322060108, + 0.01613033562898636, + 0.021700652316212654, + 0.023176493123173714, + 0.023244742304086685, + 0.004735779017210007, + 0.004001501016318798, + 0.004598614759743214, + 0.0029520904645323753, + 0.0035513315815478563, + 0.004162911791354418, + 0.003127546049654484, + 0.0034737749956548214, + 0.02987140789628029, + 0.006203783210366964, + 0.027965890243649483, + 0.005355094559490681, + 0.005402710754424334, + 0.003220152575522661, + 0.003141235327348113, + 0.0031724798027426004, + 0.018342643976211548, + 0.005543771665543318, + 0.01954178512096405, + 0.006275559309870005, + 0.008102529682219028, + 0.006238466128706932, + 0.005383460782468319, + 0.0046729035675525665, + 0.005430775694549084, + 0.005621351767331362, + 0.0115435142070055, + 0.016557103022933006, + 0.006191449239850044, + 0.0030887348111718893, + 0.004732965957373381, + 0.0038599150720983744, + 0.026032039895653725, + 0.023688046261668205, + 0.04497424513101578, + 0.0809006541967392, + 0.020705696195364, + 0.014894068241119385, + 0.020905256271362305, + 0.021890124306082726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.019272571429610252 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421563234180212 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.01421563234180212, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421563234180212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09269099682569504, + 0.12217153608798981, + 0.2773817479610443, + 0.026735944673419, + 0.21137207746505737, + 0.026792267337441444, + 0.058156080543994904, + 0.07470033317804337, + 0.056305862963199615, + 0.05369316413998604 + ], + "credit_map": [ + 0.013145741075277328, + 0.006587882060557604, + 0.015503599308431149, + 0.012753760442137718, + 0.008294740691781044, + 0.015778690576553345, + 0.010943838395178318, + 0.009682743810117245, + 0.034556057304143906, + 0.010609517805278301, + 0.03918469697237015, + 0.0074464804492890835, + 0.005668822675943375, + 0.00872263964265585, + 0.007923319935798645, + 0.008059998974204063, + 0.04929447919130325, + 0.05310268700122833, + 0.03873281180858612, + 0.023156436160206795, + 0.021234866231679916, + 0.03019297681748867, + 0.021404189988970757, + 0.040263280272483826, + 0.004578175023198128, + 0.0026900088414549828, + 0.004739394411444664, + 0.002422747202217579, + 0.0021754386834800243, + 0.0027061982546001673, + 0.0042379917576909065, + 0.0031859902665019035, + 0.028251994401216507, + 0.030020469799637794, + 0.03123880922794342, + 0.01627877913415432, + 0.02111239545047283, + 0.02799205854535103, + 0.029704108834266663, + 0.026773467659950256, + 0.006181622389703989, + 0.004703064914792776, + 0.004639602266252041, + 0.002040920313447714, + 0.0024456342216581106, + 0.0028495807200670242, + 0.0017388890264555812, + 0.0021929542999714613, + 0.011644916608929634, + 0.00802494678646326, + 0.010239452123641968, + 0.005761323031038046, + 0.007522988598793745, + 0.004870009142905474, + 0.005000852979719639, + 0.005091588478535414, + 0.017211003229022026, + 0.007224376313388348, + 0.017982177436351776, + 0.0062576774507761, + 0.009882759302854538, + 0.006496316287666559, + 0.005430016200989485, + 0.004216007888317108, + 0.0077298469841480255, + 0.007859610952436924, + 0.009549169801175594, + 0.01326711755245924, + 0.005112300626933575, + 0.0039004343561828136, + 0.005737292114645243, + 0.0031500901095569134, + 0.006124184932559729, + 0.008798525668680668, + 0.007366510108113289, + 0.010546673089265823, + 0.007495449390262365, + 0.003528703935444355, + 0.005070451647043228, + 0.004762662574648857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.01421563234180212 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639693535864353 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639693535864353, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639693535864353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07841257750988007, + 0.1704278588294983, + 0.035116102546453476, + 0.11016561836004257, + 0.08788977563381195, + 0.05422181636095047, + 0.19790956377983093, + 0.06549005210399628, + 0.05119047686457634, + 0.1491761952638626 + ], + "credit_map": [ + 0.012042048387229443, + 0.008022964932024479, + 0.013905346393585205, + 0.010972718708217144, + 0.007063510827720165, + 0.007934559136629105, + 0.011915682815015316, + 0.006555747706443071, + 0.04805278405547142, + 0.022519925609230995, + 0.06532244384288788, + 0.008627751842141151, + 0.004253422375768423, + 0.007498393300920725, + 0.007113985251635313, + 0.007039157208055258, + 0.005721915513277054, + 0.006624042056500912, + 0.0056016393937170506, + 0.003782296786084771, + 0.00272421445697546, + 0.004712880123406649, + 0.0026320170145481825, + 0.003317098366096616, + 0.017525838688015938, + 0.01944112777709961, + 0.024391217157244682, + 0.01307913102209568, + 0.005432270001620054, + 0.008940144442021847, + 0.01139545813202858, + 0.009960432536900043, + 0.019394414499402046, + 0.005928174592554569, + 0.02111937291920185, + 0.005437770858407021, + 0.0065415105782449245, + 0.0097255390137434, + 0.01142310444265604, + 0.008319887332618237, + 0.012246369384229183, + 0.008115934208035469, + 0.007839454337954521, + 0.004226576071232557, + 0.006145175080746412, + 0.005304975900799036, + 0.004959526471793652, + 0.005383803974837065, + 0.03543220832943916, + 0.04896976426243782, + 0.029691200703382492, + 0.02815796434879303, + 0.01921725459396839, + 0.013393350876867771, + 0.013461161404848099, + 0.009586657397449017, + 0.015988260507583618, + 0.006692452356219292, + 0.017850102856755257, + 0.005753386300057173, + 0.007519447710365057, + 0.005231406074017286, + 0.003378709778189659, + 0.003076276509091258, + 0.004532391671091318, + 0.006769675761461258, + 0.007786438334733248, + 0.013963415287435055, + 0.006520652677863836, + 0.002499900059774518, + 0.006218672730028629, + 0.0028993289452046156, + 0.01670323684811592, + 0.01428279746323824, + 0.031137101352214813, + 0.05368080735206604, + 0.011350431479513645, + 0.0064262873493134975, + 0.008703883737325668, + 0.006891641300171614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.004639693535864353 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204279512166977 + ], + "opposite_logprob_tokens": [ + -2.812042713165283 + ], + "logprob_reward": -0.06204279512166977, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204279512166977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05076194927096367, + 0.022698061540722847, + 0.01558964978903532, + 0.010618274100124836, + 0.05545063316822052, + 0.020285259932279587, + 0.1508067548274994, + 0.025319503620266914, + 0.3658781945705414, + 0.28259170055389404 + ], + "credit_map": [ + 0.007698673754930496, + 0.007321920245885849, + 0.008745360188186169, + 0.007363626733422279, + 0.0035623384173959494, + 0.00536205992102623, + 0.006106532644480467, + 0.004601434804499149, + 0.00639194343239069, + 0.002720956224948168, + 0.0039274971932172775, + 0.0018378921085968614, + 0.0014198600547388196, + 0.0021811332553625107, + 0.002130551263689995, + 0.002088225446641445, + 0.002369200810790062, + 0.0017241021851077676, + 0.0028051238041371107, + 0.0015674205496907234, + 0.001632337924093008, + 0.0019520993810147047, + 0.001015331712551415, + 0.00252403412014246, + 0.0017501416150480509, + 0.001257433439604938, + 0.0018119652522727847, + 0.0009651452419348061, + 0.0006973379640839994, + 0.0013257088139653206, + 0.0014073565835133195, + 0.0014031848404556513, + 0.015917371958494186, + 0.004341118037700653, + 0.016458667814731598, + 0.0029055108316242695, + 0.0035122856497764587, + 0.0039723594672977924, + 0.004260836634784937, + 0.004082482773810625, + 0.002710366854444146, + 0.004638957791030407, + 0.0031582298688590527, + 0.0017936653457581997, + 0.002507004654034972, + 0.002033781725913286, + 0.0016680284170433879, + 0.0017752260901033878, + 0.06116948276758194, + 0.007621826604008675, + 0.05636919289827347, + 0.007551493588835001, + 0.0053107161074876785, + 0.004198975395411253, + 0.004080093465745449, + 0.004504974465817213, + 0.006824427749961615, + 0.0021254881285130978, + 0.006626923102885485, + 0.0022177959326654673, + 0.0023273355327546597, + 0.0015357485972344875, + 0.001942754490301013, + 0.001719032065011561, + 0.03826597332954407, + 0.051194366067647934, + 0.06168448552489281, + 0.11646940559148788, + 0.028052182868123055, + 0.015774695202708244, + 0.03724876046180725, + 0.017188293859362602, + 0.031898725777864456, + 0.028999267145991325, + 0.05805717036128044, + 0.10122004896402359, + 0.019366273656487465, + 0.01218858640640974, + 0.01664290577173233, + 0.014218706637620926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.06204279512166977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with wrinkled eyes and warm brown skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a traditional handwoven cotton sari in earthy tones of ochre and deep terracotta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a sun-drenched room with uneven plastered earth walls and a small unglazed window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman light an incense stick using a match, causing the tip to glow orange before she blows it out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up that tracks slowly and steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809213638305664 + ], + "opposite_logprob_tokens": [ + -3.2880921363830566 + ], + "logprob_reward": -0.03809213638305664, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809213638305664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07314883172512054, + 0.065147764980793, + 0.05308970808982849, + 0.035664841532707214, + 0.2245049774646759, + 0.05711248889565468, + 0.05353168025612831, + 0.158071830868721, + 0.12186507135629654, + 0.15786276757717133 + ], + "credit_map": [ + 0.011431539431214333, + 0.008614758029580116, + 0.013984930701553822, + 0.008044390007853508, + 0.004908346571028233, + 0.010066483169794083, + 0.008983995765447617, + 0.0071143945679068565, + 0.016759654507040977, + 0.006769166793674231, + 0.015946893021464348, + 0.0060133361257612705, + 0.004328302573412657, + 0.005977923516184092, + 0.003993653226643801, + 0.005358838476240635, + 0.009576517157256603, + 0.005591945257037878, + 0.009615818969905376, + 0.004281728528439999, + 0.0030949555803090334, + 0.005951459985226393, + 0.004656641278415918, + 0.010320642031729221, + 0.006683937273919582, + 0.007822114042937756, + 0.005782180931419134, + 0.0021813628263771534, + 0.003140836488455534, + 0.002886989386752248, + 0.0036028658505529165, + 0.0035645586904138327, + 0.07067925482988358, + 0.02055336721241474, + 0.06812537461519241, + 0.012111321091651917, + 0.012249244377017021, + 0.012598652392625809, + 0.015673363581299782, + 0.012514404952526093, + 0.011481396853923798, + 0.011067403480410576, + 0.007965399883687496, + 0.004802620969712734, + 0.006806007586419582, + 0.005803677253425121, + 0.004497583024203777, + 0.00468840217217803, + 0.018568377941846848, + 0.0033068424090743065, + 0.017147039994597435, + 0.002478949958458543, + 0.004662001505494118, + 0.0021191344130784273, + 0.0030144064221531153, + 0.00223493087105453, + 0.05467892438173294, + 0.009512671269476414, + 0.054543301463127136, + 0.009225725196301937, + 0.01111817080527544, + 0.00635350588709116, + 0.007704436779022217, + 0.004935114178806543, + 0.01279506366699934, + 0.015689613297581673, + 0.018085794523358345, + 0.0383271723985672, + 0.01109254453331232, + 0.0074355402030050755, + 0.012285253964364529, + 0.006154093891382217, + 0.016324171796441078, + 0.015264968387782574, + 0.03284400328993797, + 0.05856714025139809, + 0.01259331963956356, + 0.006209393031895161, + 0.00903223268687725, + 0.007027546409517527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.03809213638305664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_tensor([635, 635])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_tensor([635, 635])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b5a663284672036a10d06493c775bb97a26005d6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_tensor([635, 635])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot focuses on a calloused, weathered hand, firmly gripping a log of dry, cracked pinewood. The skin is dry and rough, with the knuckles prominent from years of toil. The palm and fingertips are covered in thick, yellowed calluses, and even tiny sawdust is embedded in the skin's texture. With a careful, steady movement, the hand slowly pushes the log toward the open door of a heavy, black cast-iron stove, its surface mottled with age. Within the hearth, a fierce orange-red flame leaps and tumbles, and the end of the log, upon contact with the heat, bursts into crackling golden sparks. The only illumination in the scene comes from the dancing flames within, its warm glow casting flickering highlights on the rough back of the hand, contrasting sharply with the cool, metallic texture of the cast iron. The image, dominated by deep black and fiery orange-red, creates an atmosphere of primal, warm, and focused tranquility.", + "seed": 32240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_635_seed_32240_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06609907438978553, + "vlm_logprob": -0.06609907438978553, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.06609907438978553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012267612852156162 + ], + "opposite_logprob_tokens": [ + -5.387267589569092 + ], + "logprob_reward": -0.012267612852156162, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012267612852156162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13231176137924194, + 0.07429542392492294, + 0.08800139278173447, + 0.07024523615837097, + 0.021754959598183632, + 0.03838665410876274, + 0.16611942648887634, + 0.09862327575683594, + 0.0884472206234932, + 0.22181464731693268 + ], + "credit_map": [ + 0.011856279335916042, + 0.012619616463780403, + 0.02700473740696907, + 0.020323598757386208, + 0.00794715341180563, + 0.022237315773963928, + 0.016726892441511154, + 0.013596169650554657, + 0.0065285395830869675, + 0.01504215132445097, + 0.011985723860561848, + 0.008607701398432255, + 0.006155698094516993, + 0.01423157099634409, + 0.006338991690427065, + 0.005405055359005928, + 0.00953566562384367, + 0.010687155649065971, + 0.01521974802017212, + 0.007779182866215706, + 0.007394186221063137, + 0.021725991740822792, + 0.006706301122903824, + 0.008953168988227844, + 0.006726862397044897, + 0.00832777377218008, + 0.010472571477293968, + 0.011831930838525295, + 0.00527995964512229, + 0.009541969746351242, + 0.010031722486019135, + 0.008032452315092087, + 0.0034884223714470863, + 0.001978762913495302, + 0.0036668232642114162, + 0.003117151092737913, + 0.0012561233015730977, + 0.0034368231426924467, + 0.002366208005696535, + 0.002444647718220949, + 0.003757536644116044, + 0.005776825360953808, + 0.00472087599337101, + 0.005234584677964449, + 0.002982644597068429, + 0.0050941105000674725, + 0.005332657601684332, + 0.00548742339015007, + 0.019909881055355072, + 0.0318928025662899, + 0.01997942104935646, + 0.016842879354953766, + 0.012239157222211361, + 0.029282057657837868, + 0.018731530755758286, + 0.017241720110177994, + 0.00975835882127285, + 0.006073824129998684, + 0.02089879848062992, + 0.0192740336060524, + 0.007019113749265671, + 0.014027541503310204, + 0.012397670187056065, + 0.009173939004540443, + 0.0072276839055120945, + 0.010887173004448414, + 0.0072677889838814735, + 0.010653430595993996, + 0.020300384610891342, + 0.012761580757796764, + 0.012197308242321014, + 0.007151876110583544, + 0.02669060230255127, + 0.07283060997724533, + 0.03823641687631607, + 0.020027296617627144, + 0.027211463078856468, + 0.01598336361348629, + 0.012623000890016556, + 0.008211915381252766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.012267612852156162 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041907668113708 + ], + "opposite_logprob_tokens": [ + -1.9104191064834595 + ], + "logprob_reward": -0.16041907668113708, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041907668113708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1446307897567749, + 0.0741405114531517, + 0.11282554268836975, + 0.056444354355335236, + 0.040703028440475464, + 0.030139299109578133, + 0.09312612563371658, + 0.08194125443696976, + 0.04192943498492241, + 0.3241196870803833 + ], + "credit_map": [ + 0.013493862003087997, + 0.010303537361323833, + 0.03058902733027935, + 0.018804552033543587, + 0.010261849500238895, + 0.03698788583278656, + 0.013215777464210987, + 0.01097430381923914, + 0.007235886994749308, + 0.012847923673689365, + 0.012612080201506615, + 0.007025543600320816, + 0.006230541504919529, + 0.013325673528015614, + 0.006803234573453665, + 0.008059623651206493, + 0.008238568902015686, + 0.013145633973181248, + 0.021875521168112755, + 0.005545035935938358, + 0.010127292014658451, + 0.03480825200676918, + 0.0070085604675114155, + 0.01207667589187622, + 0.005135559476912022, + 0.0060266354121267796, + 0.008540400303900242, + 0.009358823299407959, + 0.0035511215683072805, + 0.008263840340077877, + 0.00680353119969368, + 0.008764439262449741, + 0.0039281584322452545, + 0.007022161968052387, + 0.006265752483159304, + 0.004544843919575214, + 0.0029498932417482138, + 0.005603964440524578, + 0.005354992114007473, + 0.005033263936638832, + 0.0022595382761210203, + 0.0032263013999909163, + 0.004416261799633503, + 0.005146190524101257, + 0.0024197527673095465, + 0.004528218414634466, + 0.003371299710124731, + 0.004771735053509474, + 0.012001617811620235, + 0.0181935615837574, + 0.011966808699071407, + 0.007916860282421112, + 0.006447074003517628, + 0.016889965161681175, + 0.010801916010677814, + 0.008908323012292385, + 0.008199453353881836, + 0.0059711202047765255, + 0.011714460328221321, + 0.014497438445687294, + 0.007852073758840561, + 0.015881305560469627, + 0.008622382767498493, + 0.009203022345900536, + 0.0041303979232907295, + 0.003724846290424466, + 0.004599327687174082, + 0.008071349002420902, + 0.005736515391618013, + 0.004170728847384453, + 0.007767301518470049, + 0.0037289725150913, + 0.06343366950750351, + 0.11528278142213821, + 0.02928529493510723, + 0.019445199519395828, + 0.037556033581495285, + 0.02603132091462612, + 0.021640073508024216, + 0.011445321142673492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.16041907668113708 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745188031345606 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745188031345606, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745188031345606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11458311975002289, + 0.0384439118206501, + 0.12225998938083649, + 0.02911497838795185, + 0.028613630682229996, + 0.020460113883018494, + 0.16855773329734802, + 0.08487071841955185, + 0.039816442877054214, + 0.3532794117927551 + ], + "credit_map": [ + 0.007766499184072018, + 0.0075515266507864, + 0.03471267595887184, + 0.007889049127697945, + 0.005699054803699255, + 0.033880557864904404, + 0.006842229049652815, + 0.010241517797112465, + 0.004054914228618145, + 0.005893256049603224, + 0.005046105943620205, + 0.004444714169949293, + 0.003244636580348015, + 0.006316123530268669, + 0.004418049938976765, + 0.005026108119636774, + 0.010245541110634804, + 0.016211379319429398, + 0.02271275967359543, + 0.009055230766534805, + 0.01055937446653843, + 0.03702547401189804, + 0.0077225700952112675, + 0.008727652952075005, + 0.004582778550684452, + 0.0018517166608944535, + 0.004555628634989262, + 0.005040150135755539, + 0.001863794052042067, + 0.0023963055573403835, + 0.0041430010460317135, + 0.004681601654738188, + 0.004942823201417923, + 0.0032842829823493958, + 0.004188730381429195, + 0.0031945453956723213, + 0.001738797640427947, + 0.0043490855023264885, + 0.0034369747154414654, + 0.0034783894661813974, + 0.0022003052290529013, + 0.0035557684022933245, + 0.0024642611388117075, + 0.0019679975230246782, + 0.0015396653907373548, + 0.003756170626729727, + 0.002482496900483966, + 0.002493445761501789, + 0.01808716170489788, + 0.035312335938215256, + 0.02103065699338913, + 0.018890835344791412, + 0.012891406193375587, + 0.030555211007595062, + 0.01664091646671295, + 0.015149209648370743, + 0.010385394096374512, + 0.00923111941665411, + 0.01585049368441105, + 0.009434058330953121, + 0.006792060099542141, + 0.01702912338078022, + 0.006846801843494177, + 0.009301661513745785, + 0.004049214068800211, + 0.0045755766332149506, + 0.0048076133243739605, + 0.008092920295894146, + 0.0046436539851129055, + 0.0039903526194393635, + 0.0064232926815748215, + 0.003233816707506776, + 0.07175803929567337, + 0.12355740368366241, + 0.037223584949970245, + 0.023106075823307037, + 0.039066337049007416, + 0.02503471076488495, + 0.0215798057615757, + 0.011953438632190228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.006745188031345606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038120709359645844 + ], + "opposite_logprob_tokens": [ + -3.288120746612549 + ], + "logprob_reward": -0.038120709359645844, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038120709359645844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11896573752164841, + 0.044631585478782654, + 0.037962689995765686, + 0.08096924424171448, + 0.03145292401313782, + 0.017652690410614014, + 0.05426862835884094, + 0.07415486872196198, + 0.04670187830924988, + 0.49323976039886475 + ], + "credit_map": [ + 0.012230191379785538, + 0.010832373052835464, + 0.024538755416870117, + 0.013484067283570766, + 0.008501734584569931, + 0.02876998670399189, + 0.010950947180390358, + 0.009657681919634342, + 0.005179276689887047, + 0.006105742882937193, + 0.00694677047431469, + 0.004480706062167883, + 0.0035562587436288595, + 0.007089152466505766, + 0.004697201307862997, + 0.006576478946954012, + 0.0031503706704825163, + 0.004933240357786417, + 0.005697962362319231, + 0.003299663309007883, + 0.004075396340340376, + 0.009140712209045887, + 0.0041849929839372635, + 0.003480349201709032, + 0.008395690470933914, + 0.006515747867524624, + 0.0134124755859375, + 0.01686258427798748, + 0.0049469489604234695, + 0.0065078274346888065, + 0.01168543566018343, + 0.01264253444969654, + 0.00491985073313117, + 0.004272349178791046, + 0.004348574671894312, + 0.005364059470593929, + 0.0016415308928117156, + 0.004086540546268225, + 0.0047694058157503605, + 0.002050613285973668, + 0.003149953205138445, + 0.0022140147630125284, + 0.0026569960173219442, + 0.0023055551573634148, + 0.0011256247526034713, + 0.0021300206426531076, + 0.001621503266505897, + 0.0024490216746926308, + 0.006251807324588299, + 0.009831149131059647, + 0.007037491537630558, + 0.00614267960190773, + 0.004768757149577141, + 0.0075159533880651, + 0.008381840772926807, + 0.004338948056101799, + 0.006872949190437794, + 0.005566280800849199, + 0.010173712857067585, + 0.01563522219657898, + 0.006099293008446693, + 0.01100009586662054, + 0.01239820383489132, + 0.006409111898392439, + 0.003706053365021944, + 0.004395287949591875, + 0.004592590965330601, + 0.007876386865973473, + 0.009729653596878052, + 0.007716818246990442, + 0.005688344594091177, + 0.0029967420268803835, + 0.12607264518737793, + 0.17724454402923584, + 0.0352979451417923, + 0.025180254131555557, + 0.05280471220612526, + 0.036069296300411224, + 0.02755846083164215, + 0.013011896051466465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.038120709359645844 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294278502464294 + ], + "opposite_logprob_tokens": [ + -2.237942695617676 + ], + "logprob_reward": -0.11294278502464294, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294278502464294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128676176071167, + 0.06677347421646118, + 0.03499293327331543, + 0.025487728416919708, + 0.040790021419525146, + 0.039378322660923004, + 0.09849466383457184, + 0.15064594149589539, + 0.06827930361032486, + 0.34648144245147705 + ], + "credit_map": [ + 0.010964959859848022, + 0.013190386816859245, + 0.03314342349767685, + 0.009677847847342491, + 0.0060457200743258, + 0.035274188965559006, + 0.008391582407057285, + 0.011988050304353237, + 0.006040279753506184, + 0.00815450120717287, + 0.008255183696746826, + 0.005468941759318113, + 0.005047260317951441, + 0.0105664087459445, + 0.007479901425540447, + 0.015760991722345352, + 0.004292171914130449, + 0.004087100271135569, + 0.005344693548977375, + 0.005056843161582947, + 0.002955963835120201, + 0.004911049269139767, + 0.003713733982294798, + 0.0046313744969666, + 0.002628987655043602, + 0.002124717691913247, + 0.003808865323662758, + 0.003961892798542976, + 0.002004054607823491, + 0.002790549537166953, + 0.004080587066709995, + 0.0040880730375647545, + 0.0050224740989506245, + 0.004199151881039143, + 0.006485716439783573, + 0.0032877579797059298, + 0.0027027837932109833, + 0.008615178056061268, + 0.0033123919274657965, + 0.007164565846323967, + 0.0040296027436852455, + 0.0035995934158563614, + 0.005342971067875624, + 0.003768791211768985, + 0.002913291798904538, + 0.004831164609640837, + 0.004843897186219692, + 0.010049006901681423, + 0.009510992094874382, + 0.019240908324718475, + 0.01323067769408226, + 0.010674264281988144, + 0.006726542022079229, + 0.018119599670171738, + 0.013213756494224072, + 0.0077779246494174, + 0.01559281162917614, + 0.018147656694054604, + 0.020737016573548317, + 0.013000712729990482, + 0.015342620201408863, + 0.03672302886843681, + 0.011547497473657131, + 0.019554588943719864, + 0.005931217223405838, + 0.0066527314484119415, + 0.007948667742311954, + 0.010536438785493374, + 0.011738196946680546, + 0.010427126660943031, + 0.0081174336373806, + 0.00692748511210084, + 0.049557317048311234, + 0.09771336615085602, + 0.04179328680038452, + 0.02584090828895569, + 0.03411467373371124, + 0.040586695075035095, + 0.03363446146249771, + 0.023240691050887108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.11294278502464294 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot focuses on a calloused, weathered hand, firmly gripping a log of dry, cracked pinewood. The skin is dry and rough, with the knuckles prominent from years of toil. The palm and fingertips are covered in thick, yellowed calluses, and even tiny sawdust is embedded in the skin's texture. With a careful, steady movement, the hand slowly pushes the log toward the open door of a heavy, black cast-iron stove, its surface mottled with age. Within the hearth, a fierce orange-red flame leaps and tumbles, and the end of the log, upon contact with the heat, bursts into crackling golden sparks. The only illumination in the scene comes from the dancing flames within, its warm glow casting flickering highlights on the rough back of the hand, contrasting sharply with the cool, metallic texture of the cast iron. The image, dominated by deep black and fiery orange-red, creates an atmosphere of primal, warm, and focused tranquility.", + "seed": 32241, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_635_seed_32241_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0696045153774321, + "vlm_logprob": -0.0696045153774321, + "vlm_yes_no_margin": 3.3499999761581423, + "vlm_token_credit": -0.0696045153774321, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013286976143717766 + ], + "opposite_logprob_tokens": [ + -4.76328706741333 + ], + "logprob_reward": -0.013286976143717766, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013286976143717766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08339964598417282, + 0.0332648865878582, + 0.10493328422307968, + 0.3167749047279358, + 0.11170399188995361, + 0.0706448033452034, + 0.024462757632136345, + 0.14161765575408936, + 0.048902109265327454, + 0.06429598480463028 + ], + "credit_map": [ + 0.011204333044588566, + 0.011906172148883343, + 0.01198670081794262, + 0.008385792374610901, + 0.011635729111731052, + 0.009957941249012947, + 0.010148279368877411, + 0.00817468948662281, + 0.004464681260287762, + 0.0036722507793456316, + 0.004410620778799057, + 0.006320783402770758, + 0.0033072163350880146, + 0.003600762691348791, + 0.003700408386066556, + 0.0037881608586758375, + 0.00841217115521431, + 0.03437649831175804, + 0.03474189713597298, + 0.005678852088749409, + 0.004934664350003004, + 0.007502514403313398, + 0.004556466359645128, + 0.0047302101738750935, + 0.05332057550549507, + 0.05740540474653244, + 0.05606207251548767, + 0.03887772932648659, + 0.027424201369285583, + 0.029747245833277702, + 0.027587678283452988, + 0.026349974796175957, + 0.01626541279256344, + 0.011682500131428242, + 0.01640333980321884, + 0.020346619188785553, + 0.009045902639627457, + 0.010533839464187622, + 0.01579030603170395, + 0.011636067181825638, + 0.010578486137092113, + 0.007863304577767849, + 0.009386155754327774, + 0.009190382435917854, + 0.008246304467320442, + 0.00974009744822979, + 0.008454701863229275, + 0.0071853697299957275, + 0.007621604949235916, + 0.007035819813609123, + 0.0028024008497595787, + 0.001573659130372107, + 0.0012020060094073415, + 0.0017889955779537559, + 0.0011664453195407987, + 0.0012718247016891837, + 0.04401477053761482, + 0.044067177921533585, + 0.010297618806362152, + 0.012004316784441471, + 0.007212184835225344, + 0.0057845404371619225, + 0.012810287065804005, + 0.005426754709333181, + 0.009697343222796917, + 0.010268867947161198, + 0.005346212070435286, + 0.00677220756188035, + 0.0038995768409222364, + 0.002567427000030875, + 0.007046916056424379, + 0.003303560893982649, + 0.009027399122714996, + 0.008328540250658989, + 0.01596003584563732, + 0.00943500641733408, + 0.005589811597019434, + 0.0036831507459282875, + 0.007288219872862101, + 0.00498381070792675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.013286976143717766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428891271352768 + ], + "opposite_logprob_tokens": [ + -2.0178890228271484 + ], + "logprob_reward": -0.1428891271352768, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428891271352768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10486139357089996, + 0.027359988540410995, + 0.08160040527582169, + 0.07621561735868454, + 0.07483362406492233, + 0.03382885083556175, + 0.036756716668605804, + 0.37161511182785034, + 0.06533852964639664, + 0.12758970260620117 + ], + "credit_map": [ + 0.011801128275692463, + 0.009120105765759945, + 0.0157875195145607, + 0.009834532625973225, + 0.01495223119854927, + 0.012095831334590912, + 0.02105594426393509, + 0.010214108042418957, + 0.002969577442854643, + 0.003293556859716773, + 0.004162646364420652, + 0.004988703411072493, + 0.002650116104632616, + 0.0030001418199390173, + 0.0038564817514270544, + 0.002438767347484827, + 0.005657839123159647, + 0.025312112644314766, + 0.029181472957134247, + 0.004628179129213095, + 0.0035286564379930496, + 0.005766659509390593, + 0.0035058713983744383, + 0.004019616637378931, + 0.01756804622709751, + 0.022079719230532646, + 0.006332698278129101, + 0.006386928725987673, + 0.00586606515571475, + 0.00745390122756362, + 0.005213501397520304, + 0.005314767826348543, + 0.02249540388584137, + 0.013797830790281296, + 0.006472393404692411, + 0.0067004733718931675, + 0.003923358395695686, + 0.005999833345413208, + 0.0045715803280472755, + 0.010872755199670792, + 0.005323061719536781, + 0.004632985219359398, + 0.0032400935888290405, + 0.004299886990338564, + 0.0028363780584186316, + 0.00473724864423275, + 0.004492174834012985, + 0.004267022479325533, + 0.010295897722244263, + 0.011766180396080017, + 0.003676437074318528, + 0.0016379178268834949, + 0.0020576410461217165, + 0.003107391996309161, + 0.0016360192093998194, + 0.0025792319793254137, + 0.13595274090766907, + 0.13334964215755463, + 0.017558777704834938, + 0.02002822980284691, + 0.014910055324435234, + 0.014490917325019836, + 0.02455245889723301, + 0.010772320441901684, + 0.012314420193433762, + 0.01315181516110897, + 0.007401672191917896, + 0.009814110584557056, + 0.005563562735915184, + 0.0034874912817031145, + 0.009156410582363605, + 0.004449053667485714, + 0.019809577614068985, + 0.023972056806087494, + 0.015499924309551716, + 0.01739957369863987, + 0.014287005178630352, + 0.013411229476332664, + 0.013584264554083347, + 0.009626073762774467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.1428891271352768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595770264044404 + ], + "opposite_logprob_tokens": [ + -5.13095760345459 + ], + "logprob_reward": -0.00595770264044404, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595770264044404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09853710979223251, + 0.04649447649717331, + 0.052969805896282196, + 0.04661112278699875, + 0.061694685369729996, + 0.18887458741664886, + 0.025725003331899643, + 0.17950944602489471, + 0.1896253526210785, + 0.10995849221944809 + ], + "credit_map": [ + 0.007742066401988268, + 0.013915373012423515, + 0.013832056894898415, + 0.010293308645486832, + 0.010765346698462963, + 0.0083667216822505, + 0.024100514128804207, + 0.00952170416712761, + 0.005823517683893442, + 0.0051178522408008575, + 0.0063239894807338715, + 0.010391872376203537, + 0.003900617128238082, + 0.004249237943440676, + 0.006561545189470053, + 0.004125839564949274, + 0.005371958017349243, + 0.016868775710463524, + 0.01455671712756157, + 0.002998103853315115, + 0.0028312457725405693, + 0.004397634416818619, + 0.002286280505359173, + 0.0036590832751244307, + 0.007872099056839943, + 0.005483343731611967, + 0.007771724835038185, + 0.006025173235684633, + 0.0037441581953316927, + 0.0047766962088644505, + 0.005741160828620195, + 0.0051967608742415905, + 0.007902109064161777, + 0.008492129854857922, + 0.008135193958878517, + 0.008811905048787594, + 0.005194448865950108, + 0.00654886569827795, + 0.0064399950206279755, + 0.010170032270252705, + 0.030048590153455734, + 0.02162204310297966, + 0.021272292360663414, + 0.015129003673791885, + 0.019646795466542244, + 0.03796371445059776, + 0.017101192846894264, + 0.026090919971466064, + 0.006369417067617178, + 0.005543040111660957, + 0.003457080340012908, + 0.0019785650074481964, + 0.0020027190912514925, + 0.0025740978308022022, + 0.0018176307203248143, + 0.0019824502523988485, + 0.07063563168048859, + 0.0679604560136795, + 0.0068831415846943855, + 0.006713809911161661, + 0.0059571401216089725, + 0.0051256087608635426, + 0.01152574922889471, + 0.004707895219326019, + 0.03641296923160553, + 0.0468936525285244, + 0.025907503440976143, + 0.017528416588902473, + 0.015699302777647972, + 0.012138047255575657, + 0.02158319763839245, + 0.013462233357131481, + 0.016813043504953384, + 0.020004361867904663, + 0.019427547231316566, + 0.01326716411858797, + 0.009512514807283878, + 0.010720741003751755, + 0.012485632672905922, + 0.00772747490555048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.00595770264044404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308087006211281 + ], + "opposite_logprob_tokens": [ + -3.1680808067321777 + ], + "logprob_reward": -0.04308087006211281, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308087006211281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08467043191194534, + 0.04869047552347183, + 0.02975320629775524, + 0.018474675714969635, + 0.0906173586845398, + 0.1674157977104187, + 0.06778085976839066, + 0.2971961200237274, + 0.0897509977221489, + 0.10565003007650375 + ], + "credit_map": [ + 0.006355892866849899, + 0.009815157391130924, + 0.01448288094252348, + 0.010416942648589611, + 0.008885405026376247, + 0.007536889519542456, + 0.016566649079322815, + 0.01061060931533575, + 0.008660396561026573, + 0.004644426982849836, + 0.008001084439456463, + 0.005702413618564606, + 0.0034575534518808126, + 0.006793138571083546, + 0.00528468145057559, + 0.006146781612187624, + 0.003455555997788906, + 0.003414673265069723, + 0.004737585783004761, + 0.004043645691126585, + 0.003074485808610916, + 0.0038661123253405094, + 0.003102766117081046, + 0.004058381076902151, + 0.002242865739390254, + 0.002846811432391405, + 0.0030419223476201296, + 0.0020273656118661165, + 0.001915816799737513, + 0.0022504995577037334, + 0.0017168862977996469, + 0.00243250816129148, + 0.016903212293982506, + 0.012523295357823372, + 0.014514600858092308, + 0.008377671241760254, + 0.005047535989433527, + 0.009235018864274025, + 0.006298389285802841, + 0.01771763153374195, + 0.0249013788998127, + 0.018944712355732918, + 0.018741238862276077, + 0.014237497001886368, + 0.0186501145362854, + 0.0332796610891819, + 0.015212427824735641, + 0.02344876527786255, + 0.015172336250543594, + 0.012105281464755535, + 0.007298638112843037, + 0.00439106160774827, + 0.009377734735608101, + 0.006734155118465424, + 0.006086185108870268, + 0.0066154696978628635, + 0.12498310953378677, + 0.1188468411564827, + 0.010082507506012917, + 0.008704745210707188, + 0.008385786786675453, + 0.007467268500477076, + 0.012278785929083824, + 0.006447074003517628, + 0.019456051290035248, + 0.025342602282762527, + 0.009878498502075672, + 0.0071118674241006374, + 0.008028402924537659, + 0.00517015065997839, + 0.008771074935793877, + 0.005992346443235874, + 0.013760513626039028, + 0.012004561722278595, + 0.020730052143335342, + 0.01755935698747635, + 0.010245774872601032, + 0.00909147597849369, + 0.014154771342873573, + 0.008103528060019016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.04308087006211281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280790090560913 + ], + "opposite_logprob_tokens": [ + -2.017807960510254 + ], + "logprob_reward": -0.14280790090560913, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280790090560913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0563342422246933, + 0.04971260204911232, + 0.06167050451040268, + 0.03429817780852318, + 0.030508579686284065, + 0.05854754522442818, + 0.025482328608632088, + 0.2818647623062134, + 0.27207034826278687, + 0.12951093912124634 + ], + "credit_map": [ + 0.005252995062619448, + 0.006956254132091999, + 0.008418485522270203, + 0.006933332420885563, + 0.005347941070795059, + 0.005789122544229031, + 0.01210747566074133, + 0.005528641398996115, + 0.009054707363247871, + 0.005575698800384998, + 0.007055181544274092, + 0.009003300219774246, + 0.0035440835636109114, + 0.004480651114135981, + 0.006738715805113316, + 0.004260263871401548, + 0.005911567714065313, + 0.020261872559785843, + 0.016349395737051964, + 0.004215550608932972, + 0.003324544057250023, + 0.0042169541120529175, + 0.004144151229411364, + 0.0032464684918522835, + 0.004217996262013912, + 0.004374031908810139, + 0.005357113666832447, + 0.0048207128420472145, + 0.004159677308052778, + 0.0035027004778385162, + 0.004281400237232447, + 0.0035845432430505753, + 0.005374036263674498, + 0.004846158437430859, + 0.004907792899757624, + 0.0038927621208131313, + 0.0015988103114068508, + 0.0029380901250988245, + 0.00270345457829535, + 0.004247474484145641, + 0.009093697182834148, + 0.00644529378041625, + 0.008036921732127666, + 0.006231260951608419, + 0.0064229764975607395, + 0.009204323403537273, + 0.006018149666488171, + 0.0070949215441942215, + 0.004977870732545853, + 0.005294429138302803, + 0.0031100651249289513, + 0.0023772718850523233, + 0.0022027436643838882, + 0.0023938382510095835, + 0.0029999082908034325, + 0.0021262026857584715, + 0.1145903691649437, + 0.11163517087697983, + 0.010463524609804153, + 0.010360031388700008, + 0.00881613977253437, + 0.008561638183891773, + 0.011249554343521595, + 0.006188331637531519, + 0.053904060274362564, + 0.06973949074745178, + 0.035728879272937775, + 0.02396167628467083, + 0.023068316280841827, + 0.016019990667700768, + 0.03095182776451111, + 0.018696101382374763, + 0.019510392099618912, + 0.022551462054252625, + 0.0224774070084095, + 0.0137640330940485, + 0.013956441543996334, + 0.011473390273749828, + 0.01470669824630022, + 0.011071118526160717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.14280790090560913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot focuses on a calloused, weathered hand, firmly gripping a log of dry, cracked pinewood. The skin is dry and rough, with the knuckles prominent from years of toil. The palm and fingertips are covered in thick, yellowed calluses, and even tiny sawdust is embedded in the skin's texture. With a careful, steady movement, the hand slowly pushes the log toward the open door of a heavy, black cast-iron stove, its surface mottled with age. Within the hearth, a fierce orange-red flame leaps and tumbles, and the end of the log, upon contact with the heat, bursts into crackling golden sparks. The only illumination in the scene comes from the dancing flames within, its warm glow casting flickering highlights on the rough back of the hand, contrasting sharply with the cool, metallic texture of the cast iron. The image, dominated by deep black and fiery orange-red, creates an atmosphere of primal, warm, and focused tranquility.", + "seed": 32242, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_635_seed_32242_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12866975935176014, + "vlm_logprob": -0.12866975935176014, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.12866975935176014, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01189661305397749 + ], + "opposite_logprob_tokens": [ + -5.13689661026001 + ], + "logprob_reward": -0.01189661305397749, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01189661305397749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17918965220451355, + 0.14320489764213562, + 0.040402982383966446, + 0.06543287634849548, + 0.1327044814825058, + 0.112437404692173, + 0.08712350577116013, + 0.07611895352602005, + 0.07222241163253784, + 0.0911627784371376 + ], + "credit_map": [ + 0.015467074699699879, + 0.04129479080438614, + 0.02082730084657669, + 0.02709796279668808, + 0.03401242569088936, + 0.013682429678738117, + 0.014508092775940895, + 0.012299577705562115, + 0.013539697974920273, + 0.007755237631499767, + 0.045851629227399826, + 0.015917304903268814, + 0.014583964832127094, + 0.009680120274424553, + 0.010704240761697292, + 0.025172708556056023, + 0.007326560560613871, + 0.0033906525932252407, + 0.00622748676687479, + 0.003843865590170026, + 0.004141037352383137, + 0.008877257816493511, + 0.0029492187313735485, + 0.003646902274340391, + 0.010177615098655224, + 0.006682208739221096, + 0.009677313268184662, + 0.010830986313521862, + 0.004105929750949144, + 0.005393077619373798, + 0.005338830873370171, + 0.013226907700300217, + 0.014167308807373047, + 0.014052726328372955, + 0.0139091107994318, + 0.01048205979168415, + 0.01650889962911606, + 0.03165743499994278, + 0.01395934633910656, + 0.0179675854742527, + 0.008844890631735325, + 0.018291378393769264, + 0.011940537951886654, + 0.008879787288606167, + 0.025103412568569183, + 0.01715504750609398, + 0.009242333471775055, + 0.012980015017092228, + 0.012309692800045013, + 0.013935832306742668, + 0.012045174837112427, + 0.01531911175698042, + 0.012225156649947166, + 0.00879256334155798, + 0.006577343214303255, + 0.005918627604842186, + 0.0067254723981022835, + 0.010260735638439655, + 0.009632635861635208, + 0.008581751957535744, + 0.007195999380201101, + 0.015619232319295406, + 0.00889850128442049, + 0.0092046232894063, + 0.006730739027261734, + 0.007404012605547905, + 0.012557395733892918, + 0.008273342624306679, + 0.007606816478073597, + 0.01075706910341978, + 0.010340049862861633, + 0.00855298712849617, + 0.010153263807296753, + 0.014925308525562286, + 0.010104605928063393, + 0.014412726275622845, + 0.014623904600739479, + 0.01246610190719366, + 0.008143031038343906, + 0.006333836354315281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01189661305397749 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300358176231384 + ], + "opposite_logprob_tokens": [ + -2.2380034923553467 + ], + "logprob_reward": -0.11300358176231384, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300358176231384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12312990427017212, + 0.18901826441287994, + 0.09563923627138138, + 0.08313052356243134, + 0.09173191338777542, + 0.0776238739490509, + 0.06384029239416122, + 0.037250373512506485, + 0.1108211949467659, + 0.1278144270181656 + ], + "credit_map": [ + 0.009602288715541363, + 0.01929572783410549, + 0.020297953858971596, + 0.01736212894320488, + 0.02149067260324955, + 0.012105369940400124, + 0.010332440957427025, + 0.012643318623304367, + 0.018575191497802734, + 0.00963286217302084, + 0.06404317170381546, + 0.023917192593216896, + 0.017696144059300423, + 0.010683111846446991, + 0.01242594700306654, + 0.032044630497694016, + 0.017454368993639946, + 0.005172939971089363, + 0.004588284995406866, + 0.005218754056841135, + 0.0058182911016047, + 0.021003153175115585, + 0.003515693126246333, + 0.03286774829030037, + 0.012774546630680561, + 0.007413662038743496, + 0.012130659073591232, + 0.011885909363627434, + 0.0060368929989635944, + 0.007365748751908541, + 0.008217179216444492, + 0.01730591617524624, + 0.01115497574210167, + 0.005891392007470131, + 0.011234934441745281, + 0.005897582974284887, + 0.010567194782197475, + 0.02478569746017456, + 0.008256743662059307, + 0.01394339744001627, + 0.006359777878969908, + 0.012253550812602043, + 0.010790574364364147, + 0.007559026125818491, + 0.015287617221474648, + 0.011564471758902073, + 0.005653725937008858, + 0.00815513078123331, + 0.006212224252521992, + 0.0070662181824445724, + 0.008256768807768822, + 0.01786881871521473, + 0.005005516577512026, + 0.006866170559078455, + 0.006502902135252953, + 0.006061670370399952, + 0.00396932614967227, + 0.006495674140751362, + 0.003937313333153725, + 0.005110286641865969, + 0.0039034062065184116, + 0.005048304796218872, + 0.004390151239931583, + 0.0043959105387330055, + 0.007831806316971779, + 0.006394206080585718, + 0.011431536637246609, + 0.00716793816536665, + 0.013471763581037521, + 0.01663648895919323, + 0.014820837415754795, + 0.03306661918759346, + 0.021067574620246887, + 0.015069553628563881, + 0.011997169815003872, + 0.014174230396747589, + 0.019307870417833328, + 0.0161475520581007, + 0.01830892078578472, + 0.011741564609110355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.11300358176231384 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960665177553892 + ], + "opposite_logprob_tokens": [ + -5.130960464477539 + ], + "logprob_reward": -0.005960665177553892, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960665177553892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14100408554077148, + 0.06904663145542145, + 0.036036062985658646, + 0.10986253619194031, + 0.12041665613651276, + 0.20412558317184448, + 0.11448007076978683, + 0.05411362275481224, + 0.0772247165441513, + 0.07369004189968109 + ], + "credit_map": [ + 0.009257236495614052, + 0.024320095777511597, + 0.024369850754737854, + 0.018018359318375587, + 0.01920895278453827, + 0.017123844474554062, + 0.015384502708911896, + 0.013321250677108765, + 0.007730027660727501, + 0.0050820051692426205, + 0.011841032654047012, + 0.011878515593707561, + 0.006421657744795084, + 0.0047945305705070496, + 0.00786686409264803, + 0.013431999832391739, + 0.004087583627551794, + 0.0028029808308929205, + 0.004296648316085339, + 0.0038571034092456102, + 0.0036435036454349756, + 0.008236770518124104, + 0.0046425168402493, + 0.00446896068751812, + 0.011683076620101929, + 0.012065079063177109, + 0.021746771410107613, + 0.011839201673865318, + 0.013092524372041225, + 0.008612844161689281, + 0.017614511772990227, + 0.0132085420191288, + 0.013730771839618683, + 0.016157932579517365, + 0.015786971896886826, + 0.007761823944747448, + 0.013652177527546883, + 0.03125361353158951, + 0.008034128695726395, + 0.014039251953363419, + 0.01719553954899311, + 0.0395653210580349, + 0.01648380421102047, + 0.019058816134929657, + 0.038363829255104065, + 0.028968045487999916, + 0.019734961912035942, + 0.02475527860224247, + 0.013234893791377544, + 0.01800190843641758, + 0.012779002077877522, + 0.019398951902985573, + 0.019321665167808533, + 0.011465848423540592, + 0.007253627292811871, + 0.01302417553961277, + 0.004628576338291168, + 0.004697645548731089, + 0.010091309435665607, + 0.005810387432575226, + 0.008033686317503452, + 0.01064880657941103, + 0.004946050699800253, + 0.005257162265479565, + 0.00415525957942009, + 0.004224205389618874, + 0.00748582324013114, + 0.00452398182824254, + 0.009709526784718037, + 0.014130864292383194, + 0.00926551129668951, + 0.02372955158352852, + 0.011198882013559341, + 0.009873129427433014, + 0.007080006878823042, + 0.009066040627658367, + 0.011604518629610538, + 0.011957468464970589, + 0.006609025876969099, + 0.006300973240286112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.005960665177553892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038130808621644974 + ], + "opposite_logprob_tokens": [ + -3.288130760192871 + ], + "logprob_reward": -0.038130808621644974, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038130808621644974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15718276798725128, + 0.048517387360334396, + 0.3390485942363739, + 0.06735022366046906, + 0.06267419457435608, + 0.12222131341695786, + 0.0438447967171669, + 0.04394972696900368, + 0.06430292129516602, + 0.05090806260704994 + ], + "credit_map": [ + 0.008673391304910183, + 0.01872839592397213, + 0.01825094223022461, + 0.02017855830490589, + 0.016916049644351006, + 0.025867711752653122, + 0.013101303018629551, + 0.03546639904379845, + 0.005053072236478329, + 0.003668826771900058, + 0.01246156170964241, + 0.005453737452626228, + 0.004858857486397028, + 0.005368047393858433, + 0.006363637745380402, + 0.005289650522172451, + 0.08238541334867477, + 0.01947791315615177, + 0.01635889708995819, + 0.016614381223917007, + 0.014545667916536331, + 0.05833970755338669, + 0.010187771171331406, + 0.12113883346319199, + 0.006737175863236189, + 0.004833046812564135, + 0.016644151881337166, + 0.0060489121824502945, + 0.005320277996361256, + 0.008188044652342796, + 0.006394593510776758, + 0.013184025883674622, + 0.006693634670227766, + 0.00420571630820632, + 0.007242986932396889, + 0.0038612252101302147, + 0.0071570673026144505, + 0.01753338985145092, + 0.0054587870836257935, + 0.010521383956074715, + 0.010547701269388199, + 0.017196141183376312, + 0.013016901910305023, + 0.011499580927193165, + 0.020538240671157837, + 0.019489435479044914, + 0.011230669915676117, + 0.018702642992138863, + 0.004368673078715801, + 0.0070432210341095924, + 0.005385131109505892, + 0.006537538021802902, + 0.0064310673624277115, + 0.005105453077703714, + 0.003514412324875593, + 0.00545930303633213, + 0.0032856145408004522, + 0.00426182895898819, + 0.006302235648036003, + 0.003381627844646573, + 0.008052688091993332, + 0.007405534852296114, + 0.00416747760027647, + 0.0070927198976278305, + 0.005304576363414526, + 0.00444631464779377, + 0.007113185245543718, + 0.004481568466871977, + 0.008448231033980846, + 0.00996885634958744, + 0.008752684108912945, + 0.015787508338689804, + 0.006527050398290157, + 0.005792886950075626, + 0.003954623825848103, + 0.005083036608994007, + 0.008262522518634796, + 0.008762849494814873, + 0.008154856972396374, + 0.004370233044028282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.038130808621644974 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47435712814331055 + ], + "opposite_logprob_tokens": [ + -0.9743571281433105 + ], + "logprob_reward": -0.47435712814331055, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47435712814331055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1484392136335373, + 0.07454735040664673, + 0.14143377542495728, + 0.05213657394051552, + 0.059972066432237625, + 0.03952489420771599, + 0.036349210888147354, + 0.031980857253074646, + 0.3044007420539856, + 0.1112152710556984 + ], + "credit_map": [ + 0.009768864139914513, + 0.018937962129712105, + 0.017065564170479774, + 0.02166561968624592, + 0.016370445489883423, + 0.017741531133651733, + 0.028966590762138367, + 0.01792263798415661, + 0.00738055445253849, + 0.0048108408227562904, + 0.01018153503537178, + 0.015484550967812538, + 0.008237458765506744, + 0.00556674599647522, + 0.007313993759453297, + 0.01557168085128069, + 0.029971450567245483, + 0.00878229085355997, + 0.00842112023383379, + 0.006687127519398928, + 0.00903919618576765, + 0.026425059884786606, + 0.005053795874118805, + 0.0470537431538105, + 0.003936379682272673, + 0.005250381771475077, + 0.01255276519805193, + 0.0062715536914765835, + 0.004138355143368244, + 0.004194183275103569, + 0.006077231373637915, + 0.009715725667774677, + 0.005926912650465965, + 0.0035399035550653934, + 0.00982788298279047, + 0.005019348114728928, + 0.0054948232136666775, + 0.018350709229707718, + 0.004816719330847263, + 0.006995772011578083, + 0.00373627245426178, + 0.008538699708878994, + 0.003402578877285123, + 0.003623278345912695, + 0.0072252159006893635, + 0.005572974681854248, + 0.0028184005059301853, + 0.004607473034411669, + 0.0030530057847499847, + 0.004420049022883177, + 0.006307696923613548, + 0.0021106491331011057, + 0.004867066629230976, + 0.007088990416377783, + 0.0021319217048585415, + 0.006369833834469318, + 0.0024053046945482492, + 0.00275438092648983, + 0.0048588234931230545, + 0.0021954576950520277, + 0.004217714071273804, + 0.007976057007908821, + 0.002896651392802596, + 0.004676472395658493, + 0.02002359740436077, + 0.015408743172883987, + 0.025557467713952065, + 0.013988660648465157, + 0.03211755305528641, + 0.049753349274396896, + 0.03983385115861893, + 0.10771751403808594, + 0.015138142742216587, + 0.011117393150925636, + 0.010425472632050514, + 0.015567569993436337, + 0.016254795715212822, + 0.016445547342300415, + 0.01603609137237072, + 0.010230259038507938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.47435712814331055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot focuses on a calloused, weathered hand, firmly gripping a log of dry, cracked pinewood. The skin is dry and rough, with the knuckles prominent from years of toil. The palm and fingertips are covered in thick, yellowed calluses, and even tiny sawdust is embedded in the skin's texture. With a careful, steady movement, the hand slowly pushes the log toward the open door of a heavy, black cast-iron stove, its surface mottled with age. Within the hearth, a fierce orange-red flame leaps and tumbles, and the end of the log, upon contact with the heat, bursts into crackling golden sparks. The only illumination in the scene comes from the dancing flames within, its warm glow casting flickering highlights on the rough back of the hand, contrasting sharply with the cool, metallic texture of the cast iron. The image, dominated by deep black and fiery orange-red, creates an atmosphere of primal, warm, and focused tranquility.", + "seed": 32243, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_13/index_635_seed_32243_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.058514222223311665, + "vlm_logprob": -0.058514222223311665, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.058514222223311665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008068585768342018 + ], + "opposite_logprob_tokens": [ + -5.883068561553955 + ], + "logprob_reward": -0.008068585768342018, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.008068585768342018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14550596475601196, + 0.09691209346055984, + 0.11579897999763489, + 0.07404237240552902, + 0.0827755331993103, + 0.09677723795175552, + 0.07315199077129364, + 0.15201643109321594, + 0.09299657493829727, + 0.07002287358045578 + ], + "credit_map": [ + 0.0364668145775795, + 0.03738582134246826, + 0.018267741426825523, + 0.0121486596763134, + 0.01042687613517046, + 0.011433440260589123, + 0.009993710555136204, + 0.009382880292832851, + 0.015269833616912365, + 0.01016533188521862, + 0.02265099063515663, + 0.010372938588261604, + 0.009182324633002281, + 0.009703770279884338, + 0.0102279894053936, + 0.009338905103504658, + 0.018334608525037766, + 0.020587310194969177, + 0.012171079404652119, + 0.01071898452937603, + 0.020031701773405075, + 0.009137149900197983, + 0.013453862629830837, + 0.011364271864295006, + 0.010413751937448978, + 0.00971626490354538, + 0.013803661800920963, + 0.01337396539747715, + 0.006960609927773476, + 0.005331453401595354, + 0.00751887820661068, + 0.0069237807765603065, + 0.00812437478452921, + 0.008816252462565899, + 0.012911972589790821, + 0.008725794032216072, + 0.01700686477124691, + 0.013116493821144104, + 0.009026728570461273, + 0.005047037731856108, + 0.01053888164460659, + 0.00917260255664587, + 0.016415124759078026, + 0.018408313393592834, + 0.011192784644663334, + 0.008042540401220322, + 0.008937408216297626, + 0.014069575816392899, + 0.011397819966077805, + 0.009838085621595383, + 0.010568597353994846, + 0.010291626676917076, + 0.008092103525996208, + 0.008172135800123215, + 0.009653652086853981, + 0.005137958098202944, + 0.019985811784863472, + 0.02434159815311432, + 0.029520466923713684, + 0.040511518716812134, + 0.009101752191781998, + 0.009242245927453041, + 0.010155786760151386, + 0.00915723480284214, + 0.00885373167693615, + 0.01820455677807331, + 0.01786796934902668, + 0.028408922255039215, + 0.005176225211471319, + 0.0035939745139330626, + 0.006862344220280647, + 0.004028834402561188, + 0.006918067112565041, + 0.01198072824627161, + 0.01417043711990118, + 0.016176357865333557, + 0.00655380031093955, + 0.004917883779853582, + 0.00471741147339344, + 0.004588174633681774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.008068585768342018 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906563371419907 + ], + "opposite_logprob_tokens": [ + -2.4640655517578125 + ], + "logprob_reward": -0.08906563371419907, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906563371419907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10265745967626572, + 0.16437023878097534, + 0.07553911209106445, + 0.10000328719615936, + 0.05666758492588997, + 0.08186972141265869, + 0.10654718428850174, + 0.13143740594387054, + 0.05851958692073822, + 0.12238843739032745 + ], + "credit_map": [ + 0.02025543339550495, + 0.020087534561753273, + 0.011861833743751049, + 0.011020460166037083, + 0.009242940694093704, + 0.010363593697547913, + 0.010372217744588852, + 0.00945344753563404, + 0.0240645594894886, + 0.0193241685628891, + 0.04394970461726189, + 0.0228969007730484, + 0.012579955160617828, + 0.009554466232657433, + 0.021164940670132637, + 0.010835550725460052, + 0.01065461803227663, + 0.010164469480514526, + 0.012238585390150547, + 0.00901107955724001, + 0.012448498047888279, + 0.006566574797034264, + 0.006490856409072876, + 0.00796443410217762, + 0.01581082120537758, + 0.012120497412979603, + 0.01905834674835205, + 0.011373095214366913, + 0.010767950676381588, + 0.00936287920922041, + 0.010243658907711506, + 0.011266039684414864, + 0.005362479481846094, + 0.006022202782332897, + 0.009350281208753586, + 0.007680369075387716, + 0.008915913291275501, + 0.007631715387105942, + 0.008289771154522896, + 0.0034148520790040493, + 0.013287884183228016, + 0.011851723305881023, + 0.013750345446169376, + 0.014032787643373013, + 0.005937391892075539, + 0.006170121021568775, + 0.006840541958808899, + 0.009998931549489498, + 0.011878148652613163, + 0.017771415412425995, + 0.013624163344502449, + 0.01372437085956335, + 0.010276543907821178, + 0.01439552754163742, + 0.011286727152764797, + 0.013590291142463684, + 0.017813293263316154, + 0.017849009484052658, + 0.011191966943442822, + 0.03213466703891754, + 0.015790503472089767, + 0.01145679410547018, + 0.012724227271974087, + 0.01247693132609129, + 0.006089888047426939, + 0.011359273456037045, + 0.00922432541847229, + 0.01549261249601841, + 0.005408916622400284, + 0.0027495769318193197, + 0.005093562416732311, + 0.0031014333944767714, + 0.008926814422011375, + 0.020647792145609856, + 0.02614278718829155, + 0.022548694163560867, + 0.012639430351555347, + 0.009183105081319809, + 0.012522151693701744, + 0.009777667000889778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.08906563371419907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257826764136553 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257826764136553, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257826764136553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18138737976551056, + 0.04930052533745766, + 0.04824187979102135, + 0.12846790254116058, + 0.04320928826928139, + 0.1201719269156456, + 0.02970719337463379, + 0.15257011353969574, + 0.13519158959388733, + 0.1117522269487381 + ], + "credit_map": [ + 0.06369685381650925, + 0.061182208359241486, + 0.014074834063649178, + 0.01004264410585165, + 0.009852454997599125, + 0.006737872958183289, + 0.008501778356730938, + 0.0072987317107617855, + 0.007475235499441624, + 0.004348262213170528, + 0.008263231255114079, + 0.005094068590551615, + 0.005173877812922001, + 0.005907129961997271, + 0.003932377323508263, + 0.009106339886784554, + 0.007716184016317129, + 0.007466085720807314, + 0.0073286606930196285, + 0.004993773531168699, + 0.006844861898571253, + 0.0033465558663010597, + 0.005185679066926241, + 0.005360079929232597, + 0.015104315243661404, + 0.016456149518489838, + 0.020370587706565857, + 0.012753286398947239, + 0.01921592466533184, + 0.013710246421396732, + 0.014945710077881813, + 0.015911685302853584, + 0.0038418560288846493, + 0.00501885823905468, + 0.0066544488072395325, + 0.0063387369737029076, + 0.006275486201047897, + 0.00676358537748456, + 0.004755551926791668, + 0.0035607621539384127, + 0.010882568545639515, + 0.010544336400926113, + 0.01753033511340618, + 0.018438568338751793, + 0.018008815124630928, + 0.015299464575946331, + 0.008271495811641216, + 0.02119634672999382, + 0.004908449482172728, + 0.0055431644432246685, + 0.0034800174180418253, + 0.0040715718641877174, + 0.003059835871681571, + 0.0030303436797112226, + 0.0031083484645932913, + 0.0025054612196981907, + 0.020086564123630524, + 0.025135139003396034, + 0.013112653978168964, + 0.034618038684129715, + 0.0172980148345232, + 0.016064554452896118, + 0.012996356002986431, + 0.013258789665997028, + 0.012471867725253105, + 0.029311111196875572, + 0.030605848878622055, + 0.030560167506337166, + 0.008683479391038418, + 0.006430732551962137, + 0.010100792162120342, + 0.007027584593743086, + 0.007894925773143768, + 0.02041255682706833, + 0.02710910141468048, + 0.019119810312986374, + 0.009614254347980022, + 0.0075495000928640366, + 0.011270303279161453, + 0.008781776763498783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.005257826764136553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981775999069214 + ], + "opposite_logprob_tokens": [ + -3.529817819595337 + ], + "logprob_reward": -0.02981775999069214, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981775999069214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13889974355697632, + 0.0948696881532669, + 0.02726166322827339, + 0.029676901176571846, + 0.05760602653026581, + 0.04030809551477432, + 0.1055765450000763, + 0.30048245191574097, + 0.13182225823402405, + 0.07349666953086853 + ], + "credit_map": [ + 0.05063239485025406, + 0.04874412342905998, + 0.006607738323509693, + 0.005351087544113398, + 0.008284036070108414, + 0.006664145737886429, + 0.005518793128430843, + 0.007097442634403706, + 0.01171843707561493, + 0.012621404603123665, + 0.017904607579112053, + 0.010785411112010479, + 0.010125554166734219, + 0.006868057418614626, + 0.01172291673719883, + 0.013123303651809692, + 0.00436487328261137, + 0.003117467276751995, + 0.004658924415707588, + 0.003787305671721697, + 0.0023194164969027042, + 0.0027819990646094084, + 0.002393857343122363, + 0.0038378206081688404, + 0.0040327622555196285, + 0.004015305545181036, + 0.0046549346297979355, + 0.003457701066508889, + 0.003566981293261051, + 0.003008760279044509, + 0.0035137617960572243, + 0.0034266968723386526, + 0.004183594137430191, + 0.004983209539204836, + 0.008976321667432785, + 0.005987241864204407, + 0.008398998528718948, + 0.009113242849707603, + 0.008537521585822105, + 0.007425902411341667, + 0.004585564602166414, + 0.005455812904983759, + 0.006660527549684048, + 0.007903223857283592, + 0.003718836000189185, + 0.00406741863116622, + 0.0033284921664744616, + 0.00458822026848793, + 0.015275079756975174, + 0.015078315511345863, + 0.016663625836372375, + 0.01474347896873951, + 0.010101856663823128, + 0.009269473142921925, + 0.014750421047210693, + 0.009694304317235947, + 0.036925554275512695, + 0.041570715606212616, + 0.028246836736798286, + 0.06968682259321213, + 0.03463037312030792, + 0.030301691964268684, + 0.029148908331990242, + 0.029971562325954437, + 0.014888192526996136, + 0.02976691909134388, + 0.024438848719000816, + 0.03744710236787796, + 0.008061626926064491, + 0.004672373179346323, + 0.006973932031542063, + 0.005573267117142677, + 0.005412436556071043, + 0.01373072899878025, + 0.01502720545977354, + 0.02601027861237526, + 0.004466773010790348, + 0.0024959600996226072, + 0.004064349923282862, + 0.002288942923769355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02981775999069214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a calloused, weathered hand gripping a log?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the log made of dry, cracked pinewood with visible sawdust embedded in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily to push the log toward the open door of a heavy, black cast-iron stove?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated only by the warm glow of fierce orange-red flames inside the hearth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the log burst into crackling golden sparks upon contact with the heat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036130487918854 + ], + "opposite_logprob_tokens": [ + -1.9103612899780273 + ], + "logprob_reward": -0.16036130487918854, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036130487918854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1055140495300293, + 0.1068429946899414, + 0.06808752566576004, + 0.10730593651533127, + 0.03231006860733032, + 0.22413945198059082, + 0.055153489112854004, + 0.15656603872776031, + 0.09259390830993652, + 0.051486507058143616 + ], + "credit_map": [ + 0.03400241583585739, + 0.031945306807756424, + 0.012676225043833256, + 0.006637244485318661, + 0.005727960262447596, + 0.0037495417054742575, + 0.005950465798377991, + 0.004824893083423376, + 0.014745849184691906, + 0.009431055746972561, + 0.014935123734176159, + 0.011834977194666862, + 0.0131154153496027, + 0.012476460076868534, + 0.010708441957831383, + 0.019595665857195854, + 0.009565531276166439, + 0.011123532429337502, + 0.007529298309236765, + 0.0073400381952524185, + 0.010699219070374966, + 0.005475134588778019, + 0.008534322492778301, + 0.007820450700819492, + 0.012758657336235046, + 0.012964258901774883, + 0.016148753464221954, + 0.011457983404397964, + 0.015957720577716827, + 0.011775137856602669, + 0.012561438605189323, + 0.013681990094482899, + 0.003466432448476553, + 0.0029766231309622526, + 0.005591203458607197, + 0.0042898645624518394, + 0.004771191161125898, + 0.0053019109182059765, + 0.0031644327100366354, + 0.0027484141755849123, + 0.01819434203207493, + 0.019619803875684738, + 0.027602894231677055, + 0.022711170837283134, + 0.03365302085876465, + 0.039402637630701065, + 0.015285336412489414, + 0.047670260071754456, + 0.008192827925086021, + 0.0064788805320858955, + 0.0068733226507902145, + 0.0069355047307908535, + 0.005722719710320234, + 0.007524407934397459, + 0.006179335992783308, + 0.007246487308293581, + 0.021804392337799072, + 0.02731901779770851, + 0.0211165901273489, + 0.03265349566936493, + 0.017602410167455673, + 0.011113808490335941, + 0.014144261367619038, + 0.010812059044837952, + 0.009335999377071857, + 0.01561758667230606, + 0.020581703633069992, + 0.02032947726547718, + 0.006429466884583235, + 0.0049766674637794495, + 0.008511273190379143, + 0.006811730097979307, + 0.0037984561640769243, + 0.008409899659454823, + 0.010489803738892078, + 0.010047934018075466, + 0.004347738344222307, + 0.003805869724601507, + 0.006511407438665628, + 0.004075397737324238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.16036130487918854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_14/index_tensor([3292, 3292])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_14/index_tensor([3292, 3292])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..75a6b8babc9fa11eda400539e7c1baacb0c02f5f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_14/index_tensor([3292, 3292])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, fixed shot, shot from a horizontal angle, begins in pure black, then is instantly ripped apart by a violent volcanic eruption. Viscous, incandescent purplish-red lava, mingled with deep magenta and fiery scarlet, rises from a vast fissure in the earth's surface, bubbles bubbling and emitting searing sparks. Immediately behind, a massive column of dense volcanic ash surges upward with unstoppable force, like an ancient beast awakened from the earth's core. Its dark gray and black edges expand dramatically, driven by the energy within. Within the thick ash cloud, numerous razor-sharp, piercing purple lightning bolts crisscross and intertwine, instantly forming a vast electric grid covering the sky. Accompanied by the imagined violent low-frequency roar of the earth and the sharp sound of electricity like shattering glass, the entire dynamic disaster scene was stripped of all sound and movement in an instant. The picture suddenly became silent, and all movement came to an abrupt end. The violent power grid solidified in an instant, fixed into a complex and symmetrical tree-like totem, its branches flickering with cold purple light, quietly suspended in the center of the dead volcanic ash cloud. The whole scene presented a doomsday-like film texture full of dramatic light and shadow contrasts.", + "seed": 15616, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_14/index_3292_seed_15616_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.27743940660730004, + "vlm_logprob": -0.27743940660730004, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.27743940660730004, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3145016431808472 + ], + "opposite_logprob_tokens": [ + -0.3145016133785248 + ], + "logprob_reward": -1.3145016431808472, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3145016431808472, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2369457334280014, + 0.09196186065673828, + 0.1356891393661499, + 0.030236613005399704, + 0.06751781702041626, + 0.05315089598298073, + 0.29000476002693176, + 0.02888513170182705, + 0.030809203162789345, + 0.03479885682463646 + ], + "credit_map": [ + 0.023928571492433548, + 0.025973517447710037, + 0.0369512215256691, + 0.03193996474146843, + 0.023594917729496956, + 0.02162417583167553, + 0.028043711557984352, + 0.044889651238918304, + 0.005123174283653498, + 0.007175143342465162, + 0.006993509363383055, + 0.011950941756367683, + 0.01699526607990265, + 0.020056137815117836, + 0.015919817611575127, + 0.00774788111448288, + 0.011748375371098518, + 0.02271246537566185, + 0.013373241759836674, + 0.024898428469896317, + 0.009144305251538754, + 0.013083692640066147, + 0.027978111058473587, + 0.012750527821481228, + 0.004635209683328867, + 0.005696303676813841, + 0.003712560748681426, + 0.002641013590618968, + 0.004017150029540062, + 0.0044004712253808975, + 0.00242978404276073, + 0.0027041241992264986, + 0.011768288910388947, + 0.008862578310072422, + 0.009607033804059029, + 0.012147961184382439, + 0.0075872186571359634, + 0.006022583227604628, + 0.006529077421873808, + 0.004993076901882887, + 0.005249121226370335, + 0.005161160137504339, + 0.0051374537870287895, + 0.008325325325131416, + 0.007255718111991882, + 0.008963114582002163, + 0.004949596710503101, + 0.008109410293400288, + 0.032630082219839096, + 0.020719218999147415, + 0.023199740797281265, + 0.01896504871547222, + 0.10632731020450592, + 0.055284883826971054, + 0.021521827206015587, + 0.011356670409440994, + 0.0027822984848171473, + 0.0033127502538263798, + 0.004820475820451975, + 0.0026278106961399317, + 0.0038577194791287184, + 0.004569033160805702, + 0.0025913480203598738, + 0.004323696717619896, + 0.0028976881876587868, + 0.002684617880731821, + 0.0046262675896286964, + 0.0027135279960930347, + 0.002650669775903225, + 0.00541253387928009, + 0.0036161423195153475, + 0.006207758095115423, + 0.0039236159063875675, + 0.0038592179771512747, + 0.005967373959720135, + 0.0035195990931242704, + 0.0031708076130598783, + 0.00689290277659893, + 0.0030991353560239077, + 0.0043662115931510925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -1.3145016431808472 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004692020360380411 + ], + "opposite_logprob_tokens": [ + -5.379692077636719 + ], + "logprob_reward": -0.004692020360380411, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004692020360380411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19663836061954498, + 0.08667535334825516, + 0.14684288203716278, + 0.060195352882146835, + 0.05948290601372719, + 0.13066665828227997, + 0.127691850066185, + 0.07522722333669662, + 0.054896365851163864, + 0.06168306991457939 + ], + "credit_map": [ + 0.02022070251405239, + 0.022324562072753906, + 0.02929389849305153, + 0.023506127297878265, + 0.018256738781929016, + 0.020749490708112717, + 0.020916959270834923, + 0.04136987030506134, + 0.005017757881432772, + 0.007508773822337389, + 0.00900314748287201, + 0.012435564771294594, + 0.02231261506676674, + 0.008745845407247543, + 0.015164805576205254, + 0.006486841011792421, + 0.013661680743098259, + 0.02261725440621376, + 0.014281886629760265, + 0.029033808037638664, + 0.013434495776891708, + 0.015347289852797985, + 0.02640138566493988, + 0.012065072543919086, + 0.00946731772273779, + 0.01056467555463314, + 0.007167452480643988, + 0.00540977343916893, + 0.007237415760755539, + 0.007760791573673487, + 0.005529065616428852, + 0.007058857474476099, + 0.00904113333672285, + 0.008680247701704502, + 0.008012779988348484, + 0.0060515799559652805, + 0.008768330328166485, + 0.006164737045764923, + 0.007269605062901974, + 0.005494491197168827, + 0.01442711241543293, + 0.016256418079137802, + 0.009124702773988247, + 0.011853422969579697, + 0.018181955441832542, + 0.029577843844890594, + 0.01317437645047903, + 0.01807081513106823, + 0.012185637839138508, + 0.016585681587457657, + 0.021558159962296486, + 0.01509485300630331, + 0.01960555836558342, + 0.01679195836186409, + 0.016768909990787506, + 0.009101089090108871, + 0.008099325932562351, + 0.008162163197994232, + 0.015394899994134903, + 0.006552752573043108, + 0.00794711522758007, + 0.009697701781988144, + 0.00910861510783434, + 0.010264645330607891, + 0.006176823750138283, + 0.006371104158461094, + 0.008866828866302967, + 0.007305623963475227, + 0.005405839066952467, + 0.007432675454765558, + 0.005690485239028931, + 0.007646979298442602, + 0.007040880154818296, + 0.0070557924918830395, + 0.011054130271077156, + 0.0057201688177883625, + 0.004940190352499485, + 0.012411437928676605, + 0.005792094394564629, + 0.007668371312320232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.004692020360380411 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609206385910511 + ], + "opposite_logprob_tokens": [ + -4.141092300415039 + ], + "logprob_reward": -0.01609206385910511, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609206385910511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09411309659481049, + 0.08263074606657028, + 0.12114151567220688, + 0.04243674874305725, + 0.03261317312717438, + 0.051803573966026306, + 0.11984075605869293, + 0.0631769597530365, + 0.2352275401353836, + 0.15701590478420258 + ], + "credit_map": [ + 0.007757925428450108, + 0.020358368754386902, + 0.01041778177022934, + 0.009336824528872967, + 0.0166585985571146, + 0.0074187954887747765, + 0.00898756179958582, + 0.01317723747342825, + 0.005471737124025822, + 0.008243264630436897, + 0.009427223354578018, + 0.013476448133587837, + 0.01399767491966486, + 0.015497124753892422, + 0.010867675766348839, + 0.005649595521390438, + 0.011527898721396923, + 0.018745893612504005, + 0.012823297642171383, + 0.02421610616147518, + 0.008790254592895508, + 0.010595505125820637, + 0.024722032248973846, + 0.009720520116388798, + 0.006139849312603474, + 0.006768105085939169, + 0.005036074202507734, + 0.0036316104233264923, + 0.005220310762524605, + 0.005827818997204304, + 0.004026440437883139, + 0.005786539986729622, + 0.004101924132555723, + 0.005957425106316805, + 0.004888666328042746, + 0.003226066241040826, + 0.0044275084510445595, + 0.00433065602555871, + 0.002893264638260007, + 0.0027876629028469324, + 0.004564553964883089, + 0.005874307826161385, + 0.004909641575068235, + 0.010922283865511417, + 0.007296415511518717, + 0.005486393813043833, + 0.005206505302339792, + 0.007543469779193401, + 0.011250673793256283, + 0.012593298219144344, + 0.01868639886379242, + 0.013872086070477962, + 0.023934468626976013, + 0.017119312658905983, + 0.015017992816865444, + 0.0073665170930325985, + 0.006286469753831625, + 0.007546567358076572, + 0.010288719087839127, + 0.006455212831497192, + 0.010550059378147125, + 0.00852007232606411, + 0.006562340538948774, + 0.0069675203412771225, + 0.019655289128422737, + 0.026140648871660233, + 0.025069762021303177, + 0.026131493970751762, + 0.02860995940864086, + 0.033090461045503616, + 0.039468683302402496, + 0.03706122934818268, + 0.019555099308490753, + 0.01409129612147808, + 0.02455374039709568, + 0.01337096095085144, + 0.01262669824063778, + 0.03729402646422386, + 0.015375131741166115, + 0.020148945972323418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.01609206385910511 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192805349826813 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.018192805349826813, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192805349826813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062411416321992874, + 0.05763426050543785, + 0.05875192582607269, + 0.01781880483031273, + 0.03330454230308533, + 0.037999726831912994, + 0.19602809846401215, + 0.04786180332303047, + 0.20823119580745697, + 0.27995824813842773 + ], + "credit_map": [ + 0.006280751898884773, + 0.007662318646907806, + 0.008623586036264896, + 0.009617055766284466, + 0.007376706227660179, + 0.006237720139324665, + 0.008730867877602577, + 0.007882409729063511, + 0.004126247018575668, + 0.007262177299708128, + 0.009201519191265106, + 0.006480467971414328, + 0.009583602659404278, + 0.007108038291335106, + 0.0072378660552203655, + 0.0066343448124825954, + 0.005254928022623062, + 0.0085840355604887, + 0.005457466933876276, + 0.011205253191292286, + 0.008074051700532436, + 0.006261595990508795, + 0.007704165764153004, + 0.006210428662598133, + 0.0019778856076300144, + 0.002671643393114209, + 0.0023783096112310886, + 0.0020889451261609793, + 0.002108881250023842, + 0.003004716709256172, + 0.0019541815854609013, + 0.0016342417802661657, + 0.0038885364774614573, + 0.005335518158972263, + 0.006610575132071972, + 0.005828217603266239, + 0.0033399818930774927, + 0.003957738168537617, + 0.0020818121265619993, + 0.002262163907289505, + 0.003786963876336813, + 0.004832175560295582, + 0.0028259996324777603, + 0.003554300172254443, + 0.005305274855345488, + 0.008475104346871376, + 0.003676642896607518, + 0.005543265491724014, + 0.022628353908658028, + 0.014938237145543098, + 0.013788207434117794, + 0.012047569267451763, + 0.0729936808347702, + 0.036847785115242004, + 0.014366558752954006, + 0.008417706936597824, + 0.004919619765132666, + 0.008585783652961254, + 0.006844545714557171, + 0.004665942396968603, + 0.006534247193485498, + 0.0059777130372822285, + 0.005423923954367638, + 0.0049100290052592754, + 0.017111219465732574, + 0.027070732787251472, + 0.01925479993224144, + 0.025909528136253357, + 0.024194849655032158, + 0.028184428811073303, + 0.03270997107028961, + 0.03379566967487335, + 0.03448190167546272, + 0.028345180675387383, + 0.0322379395365715, + 0.023900063708424568, + 0.026008205488324165, + 0.06631714850664139, + 0.02874678559601307, + 0.039921004325151443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.018192805349826813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033718500286340714 + ], + "opposite_logprob_tokens": [ + -3.4087185859680176 + ], + "logprob_reward": -0.033718500286340714, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033718500286340714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19045411050319672, + 0.1010846272110939, + 0.15889161825180054, + 0.07083876430988312, + 0.07756201177835464, + 0.06627758592367172, + 0.11860734969377518, + 0.11255849897861481, + 0.04882822930812836, + 0.0548972487449646 + ], + "credit_map": [ + 0.015957530587911606, + 0.03092360869050026, + 0.023317990824580193, + 0.023701315745711327, + 0.0316460095345974, + 0.016346655786037445, + 0.01780509576201439, + 0.030755875632166862, + 0.0070877340622246265, + 0.009781207889318466, + 0.009434469044208527, + 0.009552775882184505, + 0.016465485095977783, + 0.02542221173644066, + 0.015426275320351124, + 0.007914459332823753, + 0.014259670861065388, + 0.024376656860113144, + 0.015362092293798923, + 0.03281231224536896, + 0.019810931757092476, + 0.015787651762366295, + 0.01891240105032921, + 0.017569905146956444, + 0.007398414425551891, + 0.008466875180602074, + 0.007494824007153511, + 0.006316919345408678, + 0.012088695541024208, + 0.01623181626200676, + 0.005561253055930138, + 0.007279969286173582, + 0.015306896530091763, + 0.011388406157493591, + 0.009232140146195889, + 0.008615855127573013, + 0.010747038759291172, + 0.009115607477724552, + 0.006744201295077801, + 0.006411856971681118, + 0.007075123488903046, + 0.01026251818984747, + 0.00673566572368145, + 0.010561395436525345, + 0.009288042783737183, + 0.006681322120130062, + 0.006132937502115965, + 0.009540576487779617, + 0.010325562208890915, + 0.011897062882781029, + 0.014368112199008465, + 0.014287727884948254, + 0.017895381897687912, + 0.021140191704034805, + 0.013689316809177399, + 0.01500399224460125, + 0.012095062993466854, + 0.018064463511109352, + 0.012054817751049995, + 0.012868096120655537, + 0.017134079709649086, + 0.01461398508399725, + 0.015321259386837482, + 0.010406727902591228, + 0.004281234927475452, + 0.0059470124542713165, + 0.008769175037741661, + 0.0050020283088088036, + 0.004702089354395866, + 0.007002169266343117, + 0.005663029849529266, + 0.007461486384272575, + 0.006884048692882061, + 0.004558670800179243, + 0.0071544647216796875, + 0.004851552192121744, + 0.004863387439399958, + 0.011244858615100384, + 0.006790204904973507, + 0.00855005532503128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.033718500286340714 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, fixed shot, shot from a horizontal angle, begins in pure black, then is instantly ripped apart by a violent volcanic eruption. Viscous, incandescent purplish-red lava, mingled with deep magenta and fiery scarlet, rises from a vast fissure in the earth's surface, bubbles bubbling and emitting searing sparks. Immediately behind, a massive column of dense volcanic ash surges upward with unstoppable force, like an ancient beast awakened from the earth's core. Its dark gray and black edges expand dramatically, driven by the energy within. Within the thick ash cloud, numerous razor-sharp, piercing purple lightning bolts crisscross and intertwine, instantly forming a vast electric grid covering the sky. Accompanied by the imagined violent low-frequency roar of the earth and the sharp sound of electricity like shattering glass, the entire dynamic disaster scene was stripped of all sound and movement in an instant. The picture suddenly became silent, and all movement came to an abrupt end. The violent power grid solidified in an instant, fixed into a complex and symmetrical tree-like totem, its branches flickering with cold purple light, quietly suspended in the center of the dead volcanic ash cloud. The whole scene presented a doomsday-like film texture full of dramatic light and shadow contrasts.", + "seed": 15617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_14/index_3292_seed_15617_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3162215269170702, + "vlm_logprob": -0.3162215269170702, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.3162215269170702, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5030025243759155 + ], + "opposite_logprob_tokens": [ + -0.25300249457359314 + ], + "logprob_reward": -1.5030025243759155, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5030025243759155, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3097541034221649, + 0.09827426820993423, + 0.04255128651857376, + 0.13984552025794983, + 0.14253194630146027, + 0.07495661824941635, + 0.0365196093916893, + 0.06782177090644836, + 0.039952218532562256, + 0.04779267683625221 + ], + "credit_map": [ + 0.031074265018105507, + 0.06732004880905151, + 0.033568453043699265, + 0.03662106394767761, + 0.039085421711206436, + 0.044321462512016296, + 0.03131674975156784, + 0.02644662745296955, + 0.009911179542541504, + 0.01577385887503624, + 0.012942059896886349, + 0.012173728086054325, + 0.008917693980038166, + 0.012542692944407463, + 0.012548261322081089, + 0.013464796356856823, + 0.003303209785372019, + 0.00664962362498045, + 0.004190789069980383, + 0.005998300388455391, + 0.007019034121185541, + 0.007116246037185192, + 0.0037902265321463346, + 0.004483856726437807, + 0.01367601752281189, + 0.02534782513976097, + 0.01650286093354225, + 0.027010129764676094, + 0.020970888435840607, + 0.014946166425943375, + 0.011075976304709911, + 0.01031565759330988, + 0.019993523135781288, + 0.01985209248960018, + 0.011882872320711613, + 0.014193850569427013, + 0.02534463442862034, + 0.015556554310023785, + 0.02124830149114132, + 0.014460104517638683, + 0.008264988660812378, + 0.019045718014240265, + 0.004125360865145922, + 0.004622167907655239, + 0.005103470757603645, + 0.021726712584495544, + 0.004486063960939646, + 0.007582131773233414, + 0.00365230324678123, + 0.005347047466784716, + 0.004534638486802578, + 0.006303997244685888, + 0.0038883546367287636, + 0.0046417005360126495, + 0.003658842295408249, + 0.004492725245654583, + 0.014970729127526283, + 0.004470376297831535, + 0.007817576639354229, + 0.0056635532528162, + 0.003964252769947052, + 0.01747874915599823, + 0.0038402739446610212, + 0.009616264142096043, + 0.005212077405303717, + 0.007083187345415354, + 0.003868765663355589, + 0.005254962015897036, + 0.004689514171332121, + 0.003391146194189787, + 0.006239171139895916, + 0.004213396459817886, + 0.005801605526357889, + 0.007150022312998772, + 0.007073560729622841, + 0.006163551937788725, + 0.003994084428995848, + 0.006304522044956684, + 0.006442442070692778, + 0.004862884525209665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -1.5030025243759155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004148805979639292 + ], + "opposite_logprob_tokens": [ + -5.504148960113525 + ], + "logprob_reward": -0.004148805979639292, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004148805979639292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28067049384117126, + 0.07828247547149658, + 0.03152932599186897, + 0.12888099253177643, + 0.1371004581451416, + 0.05876925215125084, + 0.05251795053482056, + 0.04718540608882904, + 0.03306000307202339, + 0.152003675699234 + ], + "credit_map": [ + 0.025736166164278984, + 0.058572180569171906, + 0.02886234037578106, + 0.03440229594707489, + 0.0417846255004406, + 0.03683452680706978, + 0.03272027522325516, + 0.02175804227590561, + 0.007243127562105656, + 0.012382172979414463, + 0.010466537438333035, + 0.010317927226424217, + 0.008919944055378437, + 0.011190828867256641, + 0.009330135770142078, + 0.008431793190538883, + 0.0029312116093933582, + 0.005359893664717674, + 0.0029184476006776094, + 0.004862388130277395, + 0.004902774002403021, + 0.004418764729052782, + 0.002853726502507925, + 0.003282114863395691, + 0.012989413924515247, + 0.02378409169614315, + 0.011603524908423424, + 0.02137867361307144, + 0.017045937478542328, + 0.016726907342672348, + 0.012142307125031948, + 0.013210132718086243, + 0.01853216625750065, + 0.024319328367710114, + 0.010534342378377914, + 0.014114187099039555, + 0.020316921174526215, + 0.016111090779304504, + 0.012788432650268078, + 0.020383965224027634, + 0.00651210080832243, + 0.012660011649131775, + 0.0038459848146885633, + 0.004868834279477596, + 0.005273290444165468, + 0.014234279282391071, + 0.005157997366040945, + 0.006216748151928186, + 0.005939934868365526, + 0.0079275444149971, + 0.006999021861702204, + 0.008848210796713829, + 0.005306256003677845, + 0.006628872826695442, + 0.004545866511762142, + 0.006322240456938744, + 0.009210837073624134, + 0.004300774075090885, + 0.0068251886405050755, + 0.004147746134549379, + 0.0035088290460407734, + 0.007495011668652296, + 0.004150347784161568, + 0.007546667940914631, + 0.005280112847685814, + 0.003614897606894374, + 0.00644296407699585, + 0.0031523045618087053, + 0.003998648840934038, + 0.003671678714454174, + 0.0033119276631623507, + 0.0035874659661203623, + 0.028900254517793655, + 0.011026068590581417, + 0.03924909234046936, + 0.016512293368577957, + 0.009396805427968502, + 0.014342522248625755, + 0.02029595710337162, + 0.01228067185729742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.004148805979639292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420846302062273 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.01420846302062273, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420846302062273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09275981038808823, + 0.030951712280511856, + 0.01976146548986435, + 0.10350831598043442, + 0.0583464615046978, + 0.03426539525389671, + 0.029936350882053375, + 0.30907773971557617, + 0.022736888378858566, + 0.2986558675765991 + ], + "credit_map": [ + 0.008426317945122719, + 0.019531793892383575, + 0.010144438594579697, + 0.011583826504647732, + 0.01352350041270256, + 0.010990082286298275, + 0.010906027629971504, + 0.007653822656720877, + 0.0023749906104058027, + 0.0052165980450809, + 0.0030754252802580595, + 0.004315215162932873, + 0.0033955962862819433, + 0.004644081927835941, + 0.0039133853279054165, + 0.004016419872641563, + 0.002519373781979084, + 0.003397079184651375, + 0.0020968825556337833, + 0.00266633159480989, + 0.0024979107547551394, + 0.0024178207386285067, + 0.002570508047938347, + 0.0015955603448674083, + 0.009754776023328304, + 0.02103462629020214, + 0.012747643515467644, + 0.014696755446493626, + 0.01424581091850996, + 0.010640151798725128, + 0.00944057572633028, + 0.01094797346740961, + 0.007832634262740612, + 0.0084692956879735, + 0.0048859016969799995, + 0.0065752239897847176, + 0.009770726785063744, + 0.006865392439067364, + 0.006928851827979088, + 0.007018436212092638, + 0.0034052219707518816, + 0.009089837782084942, + 0.0027329926379024982, + 0.00232647149823606, + 0.0026527810841798782, + 0.008401673287153244, + 0.0020474062766879797, + 0.003609010949730873, + 0.0033577438443899155, + 0.004102330654859543, + 0.004519295413047075, + 0.005251671187579632, + 0.0029090098105371, + 0.0035777143202722073, + 0.0026290060486644506, + 0.003589578904211521, + 0.10083341598510742, + 0.014357431791722775, + 0.01873546838760376, + 0.010613068006932735, + 0.016043011099100113, + 0.10893114656209946, + 0.014987547881901264, + 0.02457665652036667, + 0.002556668361648917, + 0.0024589160457253456, + 0.004059942904859781, + 0.00322192395105958, + 0.0024490864016115665, + 0.0032567039597779512, + 0.0021938520949333906, + 0.0025397944264113903, + 0.0782398208975792, + 0.01842101290822029, + 0.10403832048177719, + 0.022372014820575714, + 0.014820104464888573, + 0.016131948679685593, + 0.011701731942594051, + 0.032930925488471985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.01420846302062273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075802147388458 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075802147388458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075802147388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07444192469120026, + 0.0526222325861454, + 0.02182905003428459, + 0.039805226027965546, + 0.0549447201192379, + 0.02011049911379814, + 0.032789330929517746, + 0.0582709014415741, + 0.041281308978796005, + 0.6039047241210938 + ], + "credit_map": [ + 0.008068819530308247, + 0.016027484089136124, + 0.009303795173764229, + 0.009648654609918594, + 0.007591518107801676, + 0.0089411037042737, + 0.0076514799147844315, + 0.007209067232906818, + 0.006148905958980322, + 0.007742318790405989, + 0.00719573674723506, + 0.006412215996533632, + 0.006685194093734026, + 0.007324529811739922, + 0.004789174534380436, + 0.006324156653136015, + 0.002243352821096778, + 0.002655078424140811, + 0.0022710193879902363, + 0.004107002634555101, + 0.0024280850775539875, + 0.002615791279822588, + 0.0026073683984577656, + 0.0029013510793447495, + 0.00372151006013155, + 0.009174728766083717, + 0.003401530208066106, + 0.00531688891351223, + 0.005486814305186272, + 0.00447294395416975, + 0.004822532646358013, + 0.0034082788042724133, + 0.006034309510141611, + 0.009264927357435226, + 0.0044410331174731255, + 0.005557179916650057, + 0.007652024272829294, + 0.004524525720626116, + 0.009741594083607197, + 0.007729128934442997, + 0.0031737040262669325, + 0.003414493054151535, + 0.002111959271132946, + 0.0016302247531712055, + 0.002433750545606017, + 0.002904264722019434, + 0.002152720233425498, + 0.002289382042363286, + 0.00466566439718008, + 0.004615400452166796, + 0.0029816904570907354, + 0.0042685107327997684, + 0.004647077061235905, + 0.004450530745089054, + 0.002814071485772729, + 0.0043463860638439655, + 0.009670587256550789, + 0.003733645658940077, + 0.0062356614507734776, + 0.004426680970937014, + 0.0044130319729447365, + 0.011181443929672241, + 0.005831404589116573, + 0.012778445146977901, + 0.0056851087138056755, + 0.00475533539429307, + 0.0059452904388308525, + 0.005086057353764772, + 0.004454195033758879, + 0.005347198341041803, + 0.004359265323728323, + 0.005648861173540354, + 0.15105722844600677, + 0.039390891790390015, + 0.20750154554843903, + 0.058603864163160324, + 0.027153879404067993, + 0.036839570850133896, + 0.03294144198298454, + 0.05041632056236267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.011075802147388458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048672039061784744 + ], + "opposite_logprob_tokens": [ + -3.0486719608306885 + ], + "logprob_reward": -0.048672039061784744, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048672039061784744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09417785704135895, + 0.08100937306880951, + 0.051117636263370514, + 0.09346859157085419, + 0.10933529585599899, + 0.08154161274433136, + 0.05758783221244812, + 0.1436193734407425, + 0.07606448978185654, + 0.21207794547080994 + ], + "credit_map": [ + 0.010698731057345867, + 0.01709001138806343, + 0.01144795399159193, + 0.0137058449909091, + 0.011612307280302048, + 0.009784383699297905, + 0.010792745277285576, + 0.009045875631272793, + 0.007641713600605726, + 0.014493479393422604, + 0.010359508916735649, + 0.009914921596646309, + 0.01089425664395094, + 0.011172556318342686, + 0.007161631248891354, + 0.009371306747198105, + 0.005058338865637779, + 0.0055672787129879, + 0.005411762278527021, + 0.00646166643127799, + 0.009269999340176582, + 0.006564377341419458, + 0.0056902458891272545, + 0.007093964610248804, + 0.009652492590248585, + 0.018089530989527702, + 0.011374467052519321, + 0.011844140477478504, + 0.00991873163729906, + 0.010348259471356869, + 0.00976297352463007, + 0.012477998621761799, + 0.014750759117305279, + 0.01815241016447544, + 0.010044591501355171, + 0.010206787846982479, + 0.017997227609157562, + 0.011229989118874073, + 0.015742765739560127, + 0.011210769414901733, + 0.009649374522268772, + 0.014082226902246475, + 0.008726845495402813, + 0.0070257424376904964, + 0.009075183421373367, + 0.01694878563284874, + 0.006636646576225758, + 0.009396813809871674, + 0.00825214758515358, + 0.008563642390072346, + 0.006691393442451954, + 0.004729185253381729, + 0.005757619626820087, + 0.008502205833792686, + 0.004412482958287001, + 0.010679156519472599, + 0.03779236227273941, + 0.007260765414685011, + 0.013395694084465504, + 0.009501062333583832, + 0.0069923861883580685, + 0.037128202617168427, + 0.007399177178740501, + 0.02414972521364689, + 0.014939665794372559, + 0.009816024452447891, + 0.011171759106218815, + 0.009251865558326244, + 0.005682007409632206, + 0.008019668981432915, + 0.007507179398089647, + 0.009676319546997547, + 0.055137570947408676, + 0.014546405524015427, + 0.07107355445623398, + 0.019609952345490456, + 0.009345629252493382, + 0.011021728627383709, + 0.014029690064489841, + 0.017313417047262192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.048672039061784744 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, fixed shot, shot from a horizontal angle, begins in pure black, then is instantly ripped apart by a violent volcanic eruption. Viscous, incandescent purplish-red lava, mingled with deep magenta and fiery scarlet, rises from a vast fissure in the earth's surface, bubbles bubbling and emitting searing sparks. Immediately behind, a massive column of dense volcanic ash surges upward with unstoppable force, like an ancient beast awakened from the earth's core. Its dark gray and black edges expand dramatically, driven by the energy within. Within the thick ash cloud, numerous razor-sharp, piercing purple lightning bolts crisscross and intertwine, instantly forming a vast electric grid covering the sky. Accompanied by the imagined violent low-frequency roar of the earth and the sharp sound of electricity like shattering glass, the entire dynamic disaster scene was stripped of all sound and movement in an instant. The picture suddenly became silent, and all movement came to an abrupt end. The violent power grid solidified in an instant, fixed into a complex and symmetrical tree-like totem, its branches flickering with cold purple light, quietly suspended in the center of the dead volcanic ash cloud. The whole scene presented a doomsday-like film texture full of dramatic light and shadow contrasts.", + "seed": 15618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_14/index_3292_seed_15618_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2790659603662789, + "vlm_logprob": -0.2790659603662789, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.2790659603662789, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3146281242370605 + ], + "opposite_logprob_tokens": [ + -0.31462809443473816 + ], + "logprob_reward": -1.3146281242370605, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3146281242370605, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.434516578912735, + 0.018808487802743912, + 0.031226180493831635, + 0.1492282748222351, + 0.03693530708551407, + 0.03147060424089432, + 0.15534581243991852, + 0.07318945229053497, + 0.02657611295580864, + 0.04270317777991295 + ], + "credit_map": [ + 0.04505522921681404, + 0.08643939346075058, + 0.023494185879826546, + 0.18474014103412628, + 0.022728702053427696, + 0.02513984777033329, + 0.032906848937273026, + 0.01401222962886095, + 0.0023484062403440475, + 0.003108402481302619, + 0.002353498712182045, + 0.0023463270626962185, + 0.0018903188174590468, + 0.0014890595339238644, + 0.00270105735398829, + 0.0025714170187711716, + 0.0030758348293602467, + 0.0036100700963288546, + 0.005130095407366753, + 0.004623252898454666, + 0.004056031350046396, + 0.003709194017574191, + 0.003926061559468508, + 0.00309563847258687, + 0.011778127402067184, + 0.005863751284778118, + 0.04423181340098381, + 0.006301071960479021, + 0.009781556203961372, + 0.008360421285033226, + 0.006728151347488165, + 0.056183382868766785, + 0.0037119470071047544, + 0.005329266656190157, + 0.0037535647861659527, + 0.004828474018722773, + 0.005966660566627979, + 0.004893582779914141, + 0.004616903141140938, + 0.003834907431155443, + 0.0033450277987867594, + 0.0052439989522099495, + 0.0029191607609391212, + 0.003773786360397935, + 0.00655590882524848, + 0.004552995320409536, + 0.0025448251981288195, + 0.0025348984636366367, + 0.01745588518679142, + 0.013387855142354965, + 0.04487806186079979, + 0.008121496066451073, + 0.013731908984482288, + 0.012729266658425331, + 0.006833108142018318, + 0.038208216428756714, + 0.006492064334452152, + 0.00943716336041689, + 0.010402357205748558, + 0.007566867861896753, + 0.00913288164883852, + 0.01047667209059, + 0.008812416344881058, + 0.010869030840694904, + 0.0018544845515862107, + 0.002433669986203313, + 0.004209231119602919, + 0.0034734001383185387, + 0.0026481840759515762, + 0.003669461701065302, + 0.004989669658243656, + 0.0032980116084218025, + 0.004495757166296244, + 0.003906400874257088, + 0.00623822258785367, + 0.007152817212045193, + 0.004183748736977577, + 0.0053966473788022995, + 0.004315163474529982, + 0.007014422211796045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -1.3146281242370605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032420000061392784 + ], + "opposite_logprob_tokens": [ + -5.753242015838623 + ], + "logprob_reward": -0.0032420000061392784, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032420000061392784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3503875732421875, + 0.029777690768241882, + 0.05459247902035713, + 0.1615750938653946, + 0.041058532893657684, + 0.024156859144568443, + 0.08977060765028, + 0.16502876579761505, + 0.04857295751571655, + 0.035079438239336014 + ], + "credit_map": [ + 0.043037984520196915, + 0.06370032578706741, + 0.019701287150382996, + 0.14171268045902252, + 0.01896078884601593, + 0.015879901126027107, + 0.0291032362729311, + 0.018291359767317772, + 0.0027105314657092094, + 0.003825660329312086, + 0.0035931053571403027, + 0.004145922604948282, + 0.003960650879889727, + 0.0038911160081624985, + 0.003763360669836402, + 0.0038873443845659494, + 0.005862249527126551, + 0.007568837143480778, + 0.006167583167552948, + 0.005618451628834009, + 0.006642233580350876, + 0.008757860399782658, + 0.006819271016865969, + 0.007155992556363344, + 0.012364797294139862, + 0.007633967325091362, + 0.04767695069313049, + 0.007688311394304037, + 0.010892441496253014, + 0.009091604501008987, + 0.0077624875120818615, + 0.05846453830599785, + 0.004120062105357647, + 0.006474479101598263, + 0.003894094144925475, + 0.005879763513803482, + 0.006982249673455954, + 0.005759557709097862, + 0.0039028010796755552, + 0.004045526497066021, + 0.0021356125362217426, + 0.0027840640395879745, + 0.003968231845647097, + 0.002660812344402075, + 0.0031691426411271095, + 0.004317285027354956, + 0.002761228010058403, + 0.002360483631491661, + 0.009549540467560291, + 0.009705225005745888, + 0.02314390242099762, + 0.0075399624183773994, + 0.009070281870663166, + 0.0076349517330527306, + 0.00677885115146637, + 0.016347892582416534, + 0.012243256904184818, + 0.01519470103085041, + 0.02408011443912983, + 0.014946846291422844, + 0.01935005374252796, + 0.03216487541794777, + 0.0182309802621603, + 0.02881794236600399, + 0.003505855333060026, + 0.004216367844492197, + 0.008618072606623173, + 0.0043141962960362434, + 0.00653818529099226, + 0.011664198711514473, + 0.006020817440003157, + 0.0036952595692127943, + 0.0038983244448900223, + 0.0032698919530957937, + 0.0046647232957184315, + 0.006058737635612488, + 0.0032871144358068705, + 0.0048029241152107716, + 0.0037669301964342594, + 0.0053307898342609406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0032420000061392784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008654704317450523 + ], + "opposite_logprob_tokens": [ + -4.758654594421387 + ], + "logprob_reward": -0.008654704317450523, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008654704317450523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09129630774259567, + 0.04009231925010681, + 0.017849169671535492, + 0.14355535805225372, + 0.05216642841696739, + 0.025330036878585815, + 0.3058018088340759, + 0.18549561500549316, + 0.10913863778114319, + 0.029274320229887962 + ], + "credit_map": [ + 0.009675232693552971, + 0.01468589622527361, + 0.010322017595171928, + 0.01752239651978016, + 0.011634091846644878, + 0.007589016575366259, + 0.01252844836562872, + 0.007339205592870712, + 0.003624460194259882, + 0.004338396713137627, + 0.0030633241403847933, + 0.0034246023278683424, + 0.005521825049072504, + 0.0062130470760166645, + 0.010306615382432938, + 0.003600045805796981, + 0.001961646368727088, + 0.002486451994627714, + 0.0016001566546037793, + 0.0022856665309518576, + 0.0023233923129737377, + 0.002308407099917531, + 0.0022170692682266235, + 0.0026663795579224825, + 0.011260692961513996, + 0.00696615781635046, + 0.041382595896720886, + 0.006158838048577309, + 0.01046373788267374, + 0.008371950127184391, + 0.0069457110948860645, + 0.05200566351413727, + 0.004988478031009436, + 0.009517380967736244, + 0.005910968873649836, + 0.008208019658923149, + 0.00916366744786501, + 0.005577064584940672, + 0.004225808195769787, + 0.004575040191411972, + 0.0033802271354943514, + 0.003909837454557419, + 0.00213203695602715, + 0.00234076171182096, + 0.005033672787249088, + 0.005031063221395016, + 0.001546414801850915, + 0.001956024905666709, + 0.03696460649371147, + 0.023920178413391113, + 0.10107948631048203, + 0.017816897481679916, + 0.028863422572612762, + 0.024958301335573196, + 0.01551763154566288, + 0.05668128281831741, + 0.013386364094913006, + 0.01966385915875435, + 0.026198318228125572, + 0.018133016303181648, + 0.021904421970248222, + 0.035275839269161224, + 0.021031763404607773, + 0.029902027919888496, + 0.008771694265305996, + 0.006548652425408363, + 0.03158925101161003, + 0.009167968295514584, + 0.006346122827380896, + 0.028112584725022316, + 0.009224922396242619, + 0.009377442300319672, + 0.0028027226217091084, + 0.0029707676731050014, + 0.0037307061720639467, + 0.0037866393104195595, + 0.002910647075623274, + 0.004164321813732386, + 0.0036597116850316525, + 0.005248803645372391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.008654704317450523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738202180713415 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738202180713415, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738202180713415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25679895281791687, + 0.03925417736172676, + 0.02990148961544037, + 0.09145666658878326, + 0.03871787711977959, + 0.02018893137574196, + 0.1490919440984726, + 0.05915219709277153, + 0.20677350461483002, + 0.10866427421569824 + ], + "credit_map": [ + 0.03560388460755348, + 0.032195497304201126, + 0.026224683970212936, + 0.0931924432516098, + 0.020817242562770844, + 0.010418773628771305, + 0.024452561512589455, + 0.013893872499465942, + 0.004023505374789238, + 0.006251085549592972, + 0.004425979219377041, + 0.0042493282817304134, + 0.00370414019562304, + 0.005830803420394659, + 0.00705467164516449, + 0.0037146625109016895, + 0.003074366133660078, + 0.00328350393101573, + 0.0028825884219259024, + 0.0022627969738096, + 0.004778984934091568, + 0.005983940325677395, + 0.0037138639017939568, + 0.003921446390450001, + 0.007214874494820833, + 0.0064104716293513775, + 0.02609681710600853, + 0.004217629786580801, + 0.007051342632621527, + 0.005902084056288004, + 0.004956888034939766, + 0.02960655465722084, + 0.0036622080951929092, + 0.005563892889767885, + 0.0037302637938410044, + 0.006010265555232763, + 0.006133178249001503, + 0.0049327970482409, + 0.003917951602488756, + 0.0047673191875219345, + 0.0021461572032421827, + 0.004987653810530901, + 0.002249497454613447, + 0.0020550296176224947, + 0.0032527209259569645, + 0.002038931706920266, + 0.0019414291018620133, + 0.0015175121370702982, + 0.0180402509868145, + 0.010864491574466228, + 0.052538711577653885, + 0.007890363223850727, + 0.012570436112582684, + 0.010729086585342884, + 0.008065436035394669, + 0.028393175452947617, + 0.004038211889564991, + 0.006285754963755608, + 0.009233186021447182, + 0.005924594588577747, + 0.0061108809895813465, + 0.007453149184584618, + 0.009180260822176933, + 0.010926156304776669, + 0.012791642919182777, + 0.01128351129591465, + 0.05253849923610687, + 0.019096821546554565, + 0.01718221604824066, + 0.05999189242720604, + 0.01925850659608841, + 0.014630409888923168, + 0.011556976474821568, + 0.009855566546320915, + 0.015865884721279144, + 0.012854059226810932, + 0.010377958416938782, + 0.014237196184694767, + 0.01561837736517191, + 0.018298255279660225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.006738202180713415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206677109003067 + ], + "opposite_logprob_tokens": [ + -2.8120667934417725 + ], + "logprob_reward": -0.06206677109003067, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206677109003067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2933443784713745, + 0.05347968637943268, + 0.04330930486321449, + 0.09472207725048065, + 0.04089611768722534, + 0.05542312562465668, + 0.07976865768432617, + 0.06045043095946312, + 0.08442708104848862, + 0.1941792219877243 + ], + "credit_map": [ + 0.019276978448033333, + 0.0690048411488533, + 0.01878724806010723, + 0.12057431787252426, + 0.016087207943201065, + 0.01856091246008873, + 0.022386230528354645, + 0.00866664107888937, + 0.005488908849656582, + 0.008367029950022697, + 0.005488489754498005, + 0.006533603183925152, + 0.005397558677941561, + 0.005991990212351084, + 0.01177277509123087, + 0.004439325071871281, + 0.004224689211696386, + 0.005354168359190226, + 0.005359056405723095, + 0.0038365640211850405, + 0.007282908074557781, + 0.006721028126776218, + 0.0057429163716733456, + 0.004787974525243044, + 0.007510922849178314, + 0.007356991060078144, + 0.0258583165705204, + 0.0045600892044603825, + 0.007529758848249912, + 0.00576826324686408, + 0.00530461547896266, + 0.030833115801215172, + 0.003400257555767894, + 0.007724613416939974, + 0.005460548214614391, + 0.005748153664171696, + 0.006213576532900333, + 0.005229689180850983, + 0.0032235924154520035, + 0.003895685076713562, + 0.006381307262927294, + 0.007865354418754578, + 0.004409948829561472, + 0.007678929716348648, + 0.010925241746008396, + 0.008367117494344711, + 0.004592582583427429, + 0.005202639847993851, + 0.007984447292983532, + 0.008331653662025928, + 0.020003702491521835, + 0.007178172003477812, + 0.012588397599756718, + 0.007032923400402069, + 0.00733909010887146, + 0.00931027065962553, + 0.004739062860608101, + 0.007819198071956635, + 0.00974845141172409, + 0.007266452070325613, + 0.0073604462668299675, + 0.008435928262770176, + 0.007576396688818932, + 0.007504499051719904, + 0.005875561386346817, + 0.004520129412412643, + 0.024009497836232185, + 0.005781575106084347, + 0.006179538555443287, + 0.02247503213584423, + 0.007726500742137432, + 0.007859247736632824, + 0.02406405657529831, + 0.01212534960359335, + 0.031119493767619133, + 0.018796419724822044, + 0.015643835067749023, + 0.026467222720384598, + 0.025959433987736702, + 0.04000340774655342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.06206677109003067 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, fixed shot, shot from a horizontal angle, begins in pure black, then is instantly ripped apart by a violent volcanic eruption. Viscous, incandescent purplish-red lava, mingled with deep magenta and fiery scarlet, rises from a vast fissure in the earth's surface, bubbles bubbling and emitting searing sparks. Immediately behind, a massive column of dense volcanic ash surges upward with unstoppable force, like an ancient beast awakened from the earth's core. Its dark gray and black edges expand dramatically, driven by the energy within. Within the thick ash cloud, numerous razor-sharp, piercing purple lightning bolts crisscross and intertwine, instantly forming a vast electric grid covering the sky. Accompanied by the imagined violent low-frequency roar of the earth and the sharp sound of electricity like shattering glass, the entire dynamic disaster scene was stripped of all sound and movement in an instant. The picture suddenly became silent, and all movement came to an abrupt end. The violent power grid solidified in an instant, fixed into a complex and symmetrical tree-like totem, its branches flickering with cold purple light, quietly suspended in the center of the dead volcanic ash cloud. The whole scene presented a doomsday-like film texture full of dramatic light and shadow contrasts.", + "seed": 15619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_14/index_3292_seed_15619_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3161669167689979, + "vlm_logprob": -0.3161669167689979, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.3161669167689979, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5030362606048584 + ], + "opposite_logprob_tokens": [ + -0.2530362010002136 + ], + "logprob_reward": -1.5030362606048584, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5030362606048584, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07621067017316818, + 0.026739196851849556, + 0.04429842159152031, + 0.03400038927793503, + 0.6863347291946411, + 0.04423213750123978, + 0.016539184376597404, + 0.014076834544539452, + 0.009964431636035442, + 0.0476040318608284 + ], + "credit_map": [ + 0.010083679109811783, + 0.008416431955993176, + 0.013237474486231804, + 0.007867545820772648, + 0.009839469566941261, + 0.009806020185351372, + 0.00892371404916048, + 0.008036323823034763, + 0.0028325452003628016, + 0.004321172833442688, + 0.002648181514814496, + 0.0039442540146410465, + 0.0036137206479907036, + 0.0025703178253024817, + 0.003466273657977581, + 0.003342727664858103, + 0.006278313230723143, + 0.01056823693215847, + 0.004320801235735416, + 0.006854897364974022, + 0.00513945147395134, + 0.0037655639462172985, + 0.004323488567024469, + 0.003047667210921645, + 0.004166123922914267, + 0.005404542665928602, + 0.003125367220491171, + 0.004530780483037233, + 0.00579843670129776, + 0.005027099512517452, + 0.0036612069234251976, + 0.0022868269588798285, + 0.04342930391430855, + 0.2685893476009369, + 0.009775853715837002, + 0.01786920055747032, + 0.2702546715736389, + 0.046115975826978683, + 0.01549521367996931, + 0.014805158600211143, + 0.0028424838092178106, + 0.014370685443282127, + 0.0029515724163502455, + 0.003368947422131896, + 0.010431241244077682, + 0.004935396369546652, + 0.0027952552773058414, + 0.0025365487672388554, + 0.0017410472501069307, + 0.0015376951778307557, + 0.002440664917230606, + 0.0014424022519961, + 0.0037828036583960056, + 0.0017535388469696045, + 0.0020517692901194096, + 0.0017892606556415558, + 0.0018376946682110429, + 0.0017500297399237752, + 0.0019836989231407642, + 0.0014293825952336192, + 0.002152531873434782, + 0.0015224796952679753, + 0.0017849195282906294, + 0.0016160978702828288, + 0.0015193452127277851, + 0.0009663745295256376, + 0.0017952275229617953, + 0.0009719936642795801, + 0.0010744305327534676, + 0.0009171883575618267, + 0.0009782409761101007, + 0.0017416296759620309, + 0.0038770753890275955, + 0.0028850194066762924, + 0.007004099432379007, + 0.002724692225456238, + 0.002464932855218649, + 0.012877821922302246, + 0.002955254167318344, + 0.012815133668482304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -1.5030362606048584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004702461417764425 + ], + "opposite_logprob_tokens": [ + -5.379702568054199 + ], + "logprob_reward": -0.004702461417764425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004702461417764425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08295874297618866, + 0.02650747075676918, + 0.043616581708192825, + 0.037994809448719025, + 0.581017792224884, + 0.04292867332696915, + 0.03391396999359131, + 0.02987891621887684, + 0.03250860795378685, + 0.0886744037270546 + ], + "credit_map": [ + 0.010482569225132465, + 0.008928983472287655, + 0.01476775761693716, + 0.009605873376131058, + 0.009833110496401787, + 0.011361704207956791, + 0.009763491339981556, + 0.008215246722102165, + 0.0029914178885519505, + 0.0043457429856061935, + 0.002640442457050085, + 0.0039794063195586205, + 0.0032785600051283836, + 0.002602970227599144, + 0.0033300791401416063, + 0.0033388505689799786, + 0.005739882122725248, + 0.006508291233330965, + 0.0038371889386326075, + 0.006684044376015663, + 0.004197278991341591, + 0.005376656539738178, + 0.007113964296877384, + 0.004159275442361832, + 0.0042559560388326645, + 0.005765215493738651, + 0.0036417716182768345, + 0.0048587750643491745, + 0.007073205430060625, + 0.005470622330904007, + 0.004003637004643679, + 0.0029256250709295273, + 0.03636036068201065, + 0.22351665794849396, + 0.009285409934818745, + 0.015899278223514557, + 0.22527174651622772, + 0.04024123027920723, + 0.015713488683104515, + 0.014729599468410015, + 0.007853512652218342, + 0.006644126959145069, + 0.0050937458872795105, + 0.003683769376948476, + 0.004992982372641563, + 0.006402975879609585, + 0.0036016504745930433, + 0.004655910190194845, + 0.003913757391273975, + 0.003972282633185387, + 0.0034330713097006083, + 0.0027098837308585644, + 0.007029879372566938, + 0.003997845575213432, + 0.005299387965351343, + 0.003557860851287842, + 0.003621167503297329, + 0.00363758928142488, + 0.0040289536118507385, + 0.003656459040939808, + 0.0031950015109032393, + 0.0027679982595145702, + 0.005435328930616379, + 0.003536418080329895, + 0.004053832031786442, + 0.0018988378578796983, + 0.004555932711809874, + 0.004739252384752035, + 0.0029814366716891527, + 0.0022277708631008863, + 0.007363322656601667, + 0.0046882242895662785, + 0.008267827332019806, + 0.010326039977371693, + 0.008794530294835567, + 0.009559519588947296, + 0.011866048909723759, + 0.017995517700910568, + 0.010961928404867649, + 0.010902993381023407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.004702461417764425 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016089249402284622 + ], + "opposite_logprob_tokens": [ + -4.14108943939209 + ], + "logprob_reward": -0.016089249402284622, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016089249402284622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06686844676733017, + 0.023140359669923782, + 0.026663154363632202, + 0.028971510007977486, + 0.4165615737438202, + 0.022049006074666977, + 0.032322049140930176, + 0.0508529394865036, + 0.02771086059510708, + 0.30486011505126953 + ], + "credit_map": [ + 0.0076437401585280895, + 0.007292271591722965, + 0.01189977303147316, + 0.00784498080611229, + 0.007699486333876848, + 0.009548679925501347, + 0.008119892328977585, + 0.006819627247750759, + 0.0025978886988013983, + 0.0027747051790356636, + 0.0029666456393897533, + 0.0026976654771715403, + 0.0033150585368275642, + 0.0024168447125703096, + 0.0029709350783377886, + 0.0034006168134510517, + 0.005191433243453503, + 0.003294136840850115, + 0.002351341536268592, + 0.0027226032689213753, + 0.0030768103897571564, + 0.0024952166713774204, + 0.004565581679344177, + 0.0029660307336598635, + 0.004247283097356558, + 0.006186584942042828, + 0.0022169046569615602, + 0.003661206690594554, + 0.004133164882659912, + 0.004616952501237392, + 0.0022902335040271282, + 0.001619180548004806, + 0.02613876760005951, + 0.16325202584266663, + 0.006037913262844086, + 0.010491668246686459, + 0.16453000903129578, + 0.027597660198807716, + 0.009328141808509827, + 0.009185376577079296, + 0.0024065079633146524, + 0.00469907745718956, + 0.0020353207364678383, + 0.0026433274615556, + 0.003909105900675058, + 0.002417518524453044, + 0.0020008718129247427, + 0.001937275636009872, + 0.0043502128683030605, + 0.00305238738656044, + 0.004725456703454256, + 0.0028488803654909134, + 0.005872348323464394, + 0.003732677549123764, + 0.004976381082087755, + 0.0027637043967843056, + 0.00662199268117547, + 0.009396357461810112, + 0.006874445825815201, + 0.005394818726927042, + 0.004883443471044302, + 0.0032671864610165358, + 0.0073831635527312756, + 0.007031531073153019, + 0.0031714322976768017, + 0.0037480604369193316, + 0.0028484431095421314, + 0.004976392723619938, + 0.0034810907673090696, + 0.0024748980067670345, + 0.004064966458827257, + 0.002945577958598733, + 0.021374067291617393, + 0.027452602982521057, + 0.04623182490468025, + 0.027140570804476738, + 0.03239764645695686, + 0.04521840065717697, + 0.030296754091978073, + 0.07474823296070099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.016089249402284622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291800171136856 + ], + "opposite_logprob_tokens": [ + -3.773291826248169 + ], + "logprob_reward": -0.023291800171136856, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291800171136856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06567403674125671, + 0.02104235626757145, + 0.031147481873631477, + 0.03531573340296745, + 0.20785097777843475, + 0.021409772336483, + 0.0590977780520916, + 0.035911913961172104, + 0.061978779733181, + 0.46057116985321045 + ], + "credit_map": [ + 0.007696132175624371, + 0.007041417062282562, + 0.011768031865358353, + 0.00918290764093399, + 0.007068203296512365, + 0.007939458824694157, + 0.009265061467885971, + 0.0057128313928842545, + 0.0024875234812498093, + 0.0027042904403060675, + 0.002718248637393117, + 0.0023723815102130175, + 0.0024047179613262415, + 0.002533065155148506, + 0.0029589603655040264, + 0.0028631696477532387, + 0.005629257299005985, + 0.003752289805561304, + 0.0025012141559273005, + 0.003914642613381147, + 0.0026534334756433964, + 0.003147957380861044, + 0.0057677640579640865, + 0.003780925879254937, + 0.004775085020810366, + 0.005494866985827684, + 0.003414242761209607, + 0.004556003957986832, + 0.006503791082650423, + 0.005124631337821484, + 0.0032659803982824087, + 0.0021811313927173615, + 0.012658110819756985, + 0.07949386537075043, + 0.0034536023158580065, + 0.005457209888845682, + 0.08185778558254242, + 0.014155805110931396, + 0.005542310886085033, + 0.005232301075011492, + 0.002901531057432294, + 0.002577052917331457, + 0.0027815436478704214, + 0.0027705451939255, + 0.002495634136721492, + 0.0027690203860402107, + 0.0028716239612549543, + 0.0022428210359066725, + 0.006458387244492769, + 0.006688058376312256, + 0.005453015211969614, + 0.004380912985652685, + 0.017822088673710823, + 0.007289628032594919, + 0.006578338798135519, + 0.0044273510575294495, + 0.0046462407335639, + 0.005235826130956411, + 0.004941887222230434, + 0.003304640529677272, + 0.0055378093384206295, + 0.0036331256851553917, + 0.0038836002349853516, + 0.004728786647319794, + 0.006551394239068031, + 0.005820728838443756, + 0.0059022847563028336, + 0.008801900781691074, + 0.006729677319526672, + 0.004488195758312941, + 0.014203853905200958, + 0.009480749256908894, + 0.042006269097328186, + 0.05666312202811241, + 0.05450005829334259, + 0.048806533217430115, + 0.06330017000436783, + 0.07845466583967209, + 0.051653921604156494, + 0.06518646329641342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.023291800171136856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a pure black screen before the eruption occurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lava visible as viscous and incandescent with purplish-red, deep magenta, and fiery scarlet colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are razor-sharp purple lightning bolts visible crisscrossing within the ash cloud to form a grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from dynamic movement to a static, silent state where the power grid solidifies into a symmetrical tree-like totem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed horizontally without any movement or tracking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033714812248945236 + ], + "opposite_logprob_tokens": [ + -3.408714771270752 + ], + "logprob_reward": -0.033714812248945236, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033714812248945236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07567692548036575, + 0.027730271220207214, + 0.10354455560445786, + 0.039456844329833984, + 0.22987334430217743, + 0.08572208136320114, + 0.08564971387386322, + 0.07612751424312592, + 0.16017135977745056, + 0.11604733765125275 + ], + "credit_map": [ + 0.010049247182905674, + 0.008309578523039818, + 0.00895099900662899, + 0.00906707625836134, + 0.00804088357836008, + 0.0126201082020998, + 0.010096590034663677, + 0.008542447350919247, + 0.002922455547377467, + 0.0038097992073744535, + 0.003104663919657469, + 0.0030903334263712168, + 0.0033541603479534388, + 0.002913592616096139, + 0.004069210961461067, + 0.004466053564101458, + 0.011618983000516891, + 0.017337938770651817, + 0.01121250819414854, + 0.018144885078072548, + 0.010422968305647373, + 0.009576024487614632, + 0.014900777488946915, + 0.010330465622246265, + 0.005370238795876503, + 0.00678505701944232, + 0.0036604925990104675, + 0.005556541960686445, + 0.007017261814326048, + 0.004420856013894081, + 0.003660457208752632, + 0.002985939849168062, + 0.015636179596185684, + 0.08746886998414993, + 0.0046656932681798935, + 0.006348700262606144, + 0.08652845770120621, + 0.016129791736602783, + 0.006152680609375238, + 0.006942966021597385, + 0.013064400292932987, + 0.015257098712027073, + 0.009448675438761711, + 0.010147799737751484, + 0.010231392458081245, + 0.01047653891146183, + 0.0069734747521579266, + 0.010122696869075298, + 0.009117478504776955, + 0.00971175730228424, + 0.009956469759345055, + 0.007096994202584028, + 0.021413221955299377, + 0.009985135868191719, + 0.01097156386822462, + 0.007397090084850788, + 0.010806117206811905, + 0.010510401800274849, + 0.013162744231522083, + 0.0070199305191636086, + 0.006454679649323225, + 0.0065511916764080524, + 0.008671330288052559, + 0.012951117940247059, + 0.01601983793079853, + 0.01108697522431612, + 0.01431512925773859, + 0.027192478999495506, + 0.014419603161513805, + 0.010180787183344364, + 0.04370850324630737, + 0.023248033598065376, + 0.010924732312560081, + 0.014678657986223698, + 0.015128366649150848, + 0.011341906152665615, + 0.015339882113039494, + 0.01970844715833664, + 0.012293694540858269, + 0.016631651669740677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.033714812248945236 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_15/index_tensor([4843, 4843])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_15/index_tensor([4843, 4843])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c971f6ae864f688dfcb17f320a45c127067da159 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_15/index_tensor([4843, 4843])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static medium shot captures a majestic, dark brown bull, its muscular frame defined by rippling sinew beneath a gleaming hide, standing in the center of a sun-baked amphitheater. The ground is a fine layer of ochre dust, piled atop the dirt. In a sudden burst of primal power, the bull violently throws its massive head to the side, its thick, curved horns slicing the still air. This twitching motion shifts its immense weight, causing its front hooves to dig into the ground, sending a spectacular spray of golden-brown dust and fine grit into the air. In the harsh, direct rays of the midday sun, every particle of air gleams like gold, momentarily obscuring its powerful legs. The atmosphere is rich, filled with palpable tension and the raw, untamed energy of the animal, captured in pure kinetic energy.", + "seed": 32834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_15/index_4843_seed_32834_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.27249422515742483, + "vlm_logprob": -0.27249422515742483, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.27249422515742483, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10242126882076263 + ], + "opposite_logprob_tokens": [ + -2.977421283721924 + ], + "logprob_reward": -0.10242126882076263, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.10242126882076263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07171672582626343, + 0.02853526547551155, + 0.016792401671409607, + 0.09574149549007416, + 0.04725261777639389, + 0.467955619096756, + 0.02138100378215313, + 0.09982278943061829, + 0.09608647227287292, + 0.0547155998647213 + ], + "credit_map": [ + 0.009960256516933441, + 0.008614608086645603, + 0.007978689856827259, + 0.013622955419123173, + 0.010069419629871845, + 0.009473655372858047, + 0.005896733608096838, + 0.0061004082672297955, + 0.0035075198393315077, + 0.003642067313194275, + 0.003126844996586442, + 0.003633334767073393, + 0.003593039931729436, + 0.0037198520731180906, + 0.003656130749732256, + 0.0036564746405929327, + 0.0019069899572059512, + 0.001806870917789638, + 0.0016029452672228217, + 0.0020466845016926527, + 0.002653613220900297, + 0.002227489138022065, + 0.0021597270388156176, + 0.0023880815133452415, + 0.01767789199948311, + 0.015821699053049088, + 0.011081143282353878, + 0.005804951768368483, + 0.01580161415040493, + 0.013760648667812347, + 0.006774555426090956, + 0.009018988348543644, + 0.004453560803085566, + 0.005518781021237373, + 0.011038487777113914, + 0.004374213516712189, + 0.005598263815045357, + 0.005346561782062054, + 0.004590179305523634, + 0.006332571618258953, + 0.02097184769809246, + 0.018046895042061806, + 0.194828063249588, + 0.020975911989808083, + 0.00740766292437911, + 0.018016725778579712, + 0.007862532511353493, + 0.17984598875045776, + 0.0023268212098628283, + 0.002486910903826356, + 0.004235470201820135, + 0.002017563208937645, + 0.0021666523534804583, + 0.003394786734133959, + 0.002091775182634592, + 0.002661023521795869, + 0.007344390731304884, + 0.007234117481857538, + 0.004920963663607836, + 0.002653145231306553, + 0.005547272972762585, + 0.037959568202495575, + 0.004524645861238241, + 0.029638679698109627, + 0.00885138101875782, + 0.012315697968006134, + 0.014502826146781445, + 0.00604917062446475, + 0.007905235514044762, + 0.015470228157937527, + 0.006721030920743942, + 0.024270901456475258, + 0.00931217148900032, + 0.009519994258880615, + 0.00396632868796587, + 0.006672397255897522, + 0.00789646990597248, + 0.007591996807605028, + 0.004117449279874563, + 0.005638790782541037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.10242126882076263 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743412733078003 + ], + "opposite_logprob_tokens": [ + -0.4743412733078003 + ], + "logprob_reward": -0.9743412733078003, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743412733078003, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04031822085380554, + 0.0454614982008934, + 0.04626892879605293, + 0.04264427348971367, + 0.032832153141498566, + 0.6715406179428101, + 0.017838403582572937, + 0.039952460676431656, + 0.01613832823932171, + 0.047005172818899155 + ], + "credit_map": [ + 0.00704881502315402, + 0.006272749975323677, + 0.004838364664465189, + 0.004488338716328144, + 0.0043982272036373615, + 0.003271030029281974, + 0.0043215062469244, + 0.005679187830537558, + 0.006369976792484522, + 0.006436220835894346, + 0.005696497391909361, + 0.0057608988136053085, + 0.006465104874223471, + 0.004543033894151449, + 0.006125677842646837, + 0.004064083099365234, + 0.0044897631742060184, + 0.00559210404753685, + 0.006616321858018637, + 0.008199421688914299, + 0.005436117760837078, + 0.004828828852623701, + 0.003909442573785782, + 0.007196924649178982, + 0.006998124532401562, + 0.006466259248554707, + 0.00414107134565711, + 0.004564952105283737, + 0.00651160255074501, + 0.006413987837731838, + 0.003216076409444213, + 0.0043321936391294, + 0.003269696142524481, + 0.004955690819770098, + 0.0029227526392787695, + 0.003261595033109188, + 0.006698915734887123, + 0.0032298986334353685, + 0.005620396230369806, + 0.002873206976801157, + 0.030265387147665024, + 0.024796586483716965, + 0.2793625295162201, + 0.02952888421714306, + 0.010715973563492298, + 0.02516544610261917, + 0.011985060758888721, + 0.25972068309783936, + 0.00249904184602201, + 0.002252943115308881, + 0.0031574300955981016, + 0.001753116142936051, + 0.0019221155671402812, + 0.001979771303012967, + 0.0020640622824430466, + 0.0022099220659583807, + 0.0036664726212620735, + 0.0029930605087429285, + 0.003121405141428113, + 0.0016778167337179184, + 0.002723008394241333, + 0.013211550191044807, + 0.0020114376675337553, + 0.010547705926001072, + 0.0019433001289144158, + 0.0027115384582430124, + 0.0015727170975878835, + 0.0010894829174503684, + 0.001883527496829629, + 0.0035160016268491745, + 0.0015617732424288988, + 0.0018599863396957517, + 0.0049443026073277, + 0.0038798153400421143, + 0.006432782858610153, + 0.0049226959235966206, + 0.00380580872297287, + 0.011603636667132378, + 0.005565619561821222, + 0.005850506015121937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.9743412733078003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001191859133541584 + ], + "opposite_logprob_tokens": [ + -6.751192092895508 + ], + "logprob_reward": -0.001191859133541584, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001191859133541584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03080613538622856, + 0.021299105137586594, + 0.052298665046691895, + 0.04831969365477562, + 0.04172828048467636, + 0.6965480446815491, + 0.03554147854447365, + 0.016863761469721794, + 0.019644027575850487, + 0.036950744688510895 + ], + "credit_map": [ + 0.004475259687751532, + 0.005678126588463783, + 0.00375431589782238, + 0.00412821164354682, + 0.0036687441170215607, + 0.002741788513958454, + 0.003142444882541895, + 0.003217242890968919, + 0.002010524971410632, + 0.0019867189694195986, + 0.002286817179992795, + 0.0023394038435071707, + 0.003036011243239045, + 0.0020942804403603077, + 0.004342788364738226, + 0.0032025582622736692, + 0.00569277023896575, + 0.006325436290353537, + 0.006657083053141832, + 0.007237449754029512, + 0.008824951015412807, + 0.006111999042332172, + 0.006371849216520786, + 0.005077126435935497, + 0.00925600714981556, + 0.005673850420862436, + 0.005118648521602154, + 0.004476289730519056, + 0.004560393281280994, + 0.008592764846980572, + 0.004234309308230877, + 0.006407431326806545, + 0.0060969931073486805, + 0.004729361739009619, + 0.007188087794929743, + 0.004142836667597294, + 0.0070730699226260185, + 0.004036714788526297, + 0.0038477289490401745, + 0.0046134875155985355, + 0.03101898916065693, + 0.027486182749271393, + 0.29095226526260376, + 0.030069740489125252, + 0.00976479984819889, + 0.026150986552238464, + 0.01133949775248766, + 0.26976561546325684, + 0.004981014411896467, + 0.0037646132986992598, + 0.006386048160493374, + 0.0036330483853816986, + 0.002968279179185629, + 0.005443404894322157, + 0.0025662509724497795, + 0.005798818543553352, + 0.0012953169643878937, + 0.0015397596871480346, + 0.001818283461034298, + 0.0009059152216650546, + 0.0015153288841247559, + 0.004028184339404106, + 0.0018722624517977238, + 0.0038887113332748413, + 0.0029624763410538435, + 0.003088389988988638, + 0.0022137071937322617, + 0.0011779359774664044, + 0.0023678659927099943, + 0.0038422022480517626, + 0.0014829394640401006, + 0.0025085117667913437, + 0.005005608778446913, + 0.004861945752054453, + 0.003115024184808135, + 0.0042134434916079044, + 0.0050616334192454815, + 0.006487797014415264, + 0.004149640444666147, + 0.004055652767419815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.001191859133541584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132665157318115 + ], + "opposite_logprob_tokens": [ + -1.4063266515731812 + ], + "logprob_reward": -0.28132665157318115, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132665157318115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04010338708758354, + 0.014456383883953094, + 0.12465324997901917, + 0.0626031681895256, + 0.03052261658012867, + 0.62421053647995, + 0.011948708444833755, + 0.022614585235714912, + 0.03144313767552376, + 0.03744422644376755 + ], + "credit_map": [ + 0.004607590846717358, + 0.005639995448291302, + 0.004516443237662315, + 0.006113396026194096, + 0.005957318004220724, + 0.004874170757830143, + 0.004395459778606892, + 0.003999010194092989, + 0.0015980714233592153, + 0.001885526580736041, + 0.0018594046123325825, + 0.0019159000366926193, + 0.0021820776164531708, + 0.0014178248820826411, + 0.002167709171772003, + 0.0014298700261861086, + 0.01362866722047329, + 0.016804883256554604, + 0.016704782843589783, + 0.0179864764213562, + 0.01762359030544758, + 0.01393367350101471, + 0.012683743610978127, + 0.015287429094314575, + 0.009920907206833363, + 0.011350033804774284, + 0.007046664133667946, + 0.004246743395924568, + 0.010040614753961563, + 0.009318544529378414, + 0.004569590091705322, + 0.006110064685344696, + 0.0030951742082834244, + 0.004137016367167234, + 0.006696625147014856, + 0.00297652673907578, + 0.005412211176007986, + 0.0027051034849137068, + 0.0017320627812296152, + 0.003767898539081216, + 0.02784695103764534, + 0.024404028430581093, + 0.26194435358047485, + 0.02722221054136753, + 0.007958879694342613, + 0.023028697818517685, + 0.010524429380893707, + 0.24128098785877228, + 0.0016471404815092683, + 0.0014438555808737874, + 0.0024619349278509617, + 0.0010810540989041328, + 0.0012088930234313011, + 0.00178568868432194, + 0.0008149559143930674, + 0.0015051851514726877, + 0.0017896988429129124, + 0.0018824810395017266, + 0.001586367143318057, + 0.0007428362732753158, + 0.0016986178234219551, + 0.007257341407239437, + 0.0016056810272857547, + 0.006051562260836363, + 0.0033179267775267363, + 0.003655771492049098, + 0.0038685889448970556, + 0.0018716411432251334, + 0.003179586958140135, + 0.006373641546815634, + 0.0027901344001293182, + 0.00638584652915597, + 0.0043554762378335, + 0.0033839105162769556, + 0.003941388800740242, + 0.004572774283587933, + 0.0036592113319784403, + 0.010043583810329437, + 0.0034631798043847084, + 0.004024701192975044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.28132665157318115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190072951838374 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003190072951838374, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190072951838374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07713696360588074, + 0.0269045177847147, + 0.08702582865953445, + 0.07815203815698624, + 0.04447907209396362, + 0.5353432297706604, + 0.02744278870522976, + 0.03900999203324318, + 0.035917017608881, + 0.048588503152132034 + ], + "credit_map": [ + 0.00926174409687519, + 0.0112835131585598, + 0.009620673954486847, + 0.010689216665923595, + 0.012078242376446724, + 0.006754435133188963, + 0.00865574274212122, + 0.008793394081294537, + 0.004269767086952925, + 0.00384767004288733, + 0.003077239263802767, + 0.003697324777022004, + 0.0027709268033504486, + 0.0036918548867106438, + 0.0021725851111114025, + 0.003377150511369109, + 0.015151758678257465, + 0.011114739812910557, + 0.013589550741016865, + 0.018825192004442215, + 0.00818399153649807, + 0.007429547607898712, + 0.005642898380756378, + 0.007088145241141319, + 0.01643982157111168, + 0.008710423484444618, + 0.009493726305663586, + 0.00491240993142128, + 0.010122018866240978, + 0.013934371061623096, + 0.005612918175756931, + 0.00892634503543377, + 0.005177751183509827, + 0.005240520462393761, + 0.008592714555561543, + 0.0045842877589166164, + 0.008727583102881908, + 0.004285489674657583, + 0.0029877126216888428, + 0.004883010871708393, + 0.024130702018737793, + 0.019442282617092133, + 0.22345468401908875, + 0.02339799888432026, + 0.006516915280371904, + 0.02010260336101055, + 0.008876473642885685, + 0.2094215303659439, + 0.0030276915058493614, + 0.003222196828573942, + 0.003603007411584258, + 0.003287490224465728, + 0.003047033678740263, + 0.00553128169849515, + 0.0025547256227582693, + 0.003169360337778926, + 0.002861589193344116, + 0.0027069211937487125, + 0.002034611999988556, + 0.0013115918263792992, + 0.0030257334001362324, + 0.01409170962870121, + 0.0022776538971811533, + 0.010700185783207417, + 0.004393097944557667, + 0.004482166841626167, + 0.004369679372757673, + 0.001987595111131668, + 0.004081137012690306, + 0.007232004776597023, + 0.005008351057767868, + 0.004362985957413912, + 0.0047439387999475, + 0.003889723215252161, + 0.004477331414818764, + 0.004545207601040602, + 0.00434297788888216, + 0.014891520142555237, + 0.004486944526433945, + 0.007210865616798401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.003190072951838374 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures a majestic, dark brown bull, its muscular frame defined by rippling sinew beneath a gleaming hide, standing in the center of a sun-baked amphitheater. The ground is a fine layer of ochre dust, piled atop the dirt. In a sudden burst of primal power, the bull violently throws its massive head to the side, its thick, curved horns slicing the still air. This twitching motion shifts its immense weight, causing its front hooves to dig into the ground, sending a spectacular spray of golden-brown dust and fine grit into the air. In the harsh, direct rays of the midday sun, every particle of air gleams like gold, momentarily obscuring its powerful legs. The atmosphere is rich, filled with palpable tension and the raw, untamed energy of the animal, captured in pure kinetic energy.", + "seed": 32835, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_15/index_4843_seed_32835_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.30651565492153166, + "vlm_logprob": -0.30651565492153166, + "vlm_yes_no_margin": 2.924999988079071, + "vlm_token_credit": -0.30651565492153166, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20110958814620972 + ], + "opposite_logprob_tokens": [ + -2.0761096477508545 + ], + "logprob_reward": -0.20110958814620972, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.20110958814620972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2183540314435959, + 0.33319950103759766, + 0.05699226260185242, + 0.11988764256238937, + 0.10713937878608704, + 0.03828078880906105, + 0.05331162363290787, + 0.03697369992733002, + 0.024685485288500786, + 0.011175557039678097 + ], + "credit_map": [ + 0.02947322465479374, + 0.027742784470319748, + 0.021449722349643707, + 0.03042764775454998, + 0.024839866906404495, + 0.02926875464618206, + 0.03270583599805832, + 0.022446172311902046, + 0.050831809639930725, + 0.04053642600774765, + 0.03703790158033371, + 0.023101748898625374, + 0.06244393065571785, + 0.0678972527384758, + 0.02777361311018467, + 0.023576801642775536, + 0.006956410128623247, + 0.00653081526979804, + 0.007202210370451212, + 0.006203416269272566, + 0.007670259103178978, + 0.009721669368445873, + 0.0063881222158670425, + 0.006319359410554171, + 0.014385156333446503, + 0.008159964345395565, + 0.02317316085100174, + 0.012933433055877686, + 0.012321657501161098, + 0.013643525540828705, + 0.012706904672086239, + 0.022563843056559563, + 0.011463306844234467, + 0.012325586751103401, + 0.013354132883250713, + 0.011220772750675678, + 0.013166528195142746, + 0.014358771964907646, + 0.013620864599943161, + 0.017629411071538925, + 0.006220609415322542, + 0.004192730877548456, + 0.003723764792084694, + 0.007433855440467596, + 0.005083309020847082, + 0.0036056849639862776, + 0.003446161514148116, + 0.00457467045634985, + 0.009237405844032764, + 0.005826909095048904, + 0.006936811842024326, + 0.006098048761487007, + 0.0061130765825510025, + 0.007829919457435608, + 0.003938374109566212, + 0.007331077475100756, + 0.0032259912695735693, + 0.003559026401489973, + 0.006973010487854481, + 0.007997285574674606, + 0.0037457456346601248, + 0.004153162706643343, + 0.0034271024633198977, + 0.003892376320436597, + 0.002311010379344225, + 0.0027762476820498705, + 0.004799315705895424, + 0.0034188353456556797, + 0.002581088338047266, + 0.002680203877389431, + 0.003135527018457651, + 0.0029832555446773767, + 0.001157062011770904, + 0.002156143542379141, + 0.0009398224647156894, + 0.0017096663359552622, + 0.0019937625620514154, + 0.0012146409135311842, + 0.0010545694967731833, + 0.0009498900035396218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.20110958814620972 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236831188201904 + ], + "opposite_logprob_tokens": [ + -0.34868305921554565 + ], + "logprob_reward": -1.2236831188201904, + "yes_no_margin_reward": -0.8750000596046448, + "token_credit_reward": -1.2236831188201904, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06278903782367706, + 0.4510144591331482, + 0.07332859933376312, + 0.09268863499164581, + 0.1669055074453354, + 0.01629668101668358, + 0.020859109237790108, + 0.01917439140379429, + 0.08080609142780304, + 0.016137482598423958 + ], + "credit_map": [ + 0.007540031801909208, + 0.006807302590459585, + 0.008023006841540337, + 0.00847917515784502, + 0.007262681610882282, + 0.009365600533783436, + 0.008173836395144463, + 0.007137401029467583, + 0.06813860684633255, + 0.05835988372564316, + 0.05134871229529381, + 0.03308575600385666, + 0.08791898936033249, + 0.08275679498910904, + 0.037197768688201904, + 0.03220794349908829, + 0.009624437429010868, + 0.007649715524166822, + 0.008490674197673798, + 0.008111357688903809, + 0.010912944562733173, + 0.01002909429371357, + 0.010159308090806007, + 0.008351064287126064, + 0.014390666037797928, + 0.006623485125601292, + 0.01652359962463379, + 0.009476454928517342, + 0.008701923303306103, + 0.012690011411905289, + 0.008950995281338692, + 0.015331503003835678, + 0.019764170050621033, + 0.011353718116879463, + 0.02589777670800686, + 0.017093239352107048, + 0.013025732710957527, + 0.021402351558208466, + 0.024184947833418846, + 0.0341835655272007, + 0.0020106860902160406, + 0.0017595381941646338, + 0.0014737627934664488, + 0.002059916965663433, + 0.003949476405978203, + 0.0015330136520788074, + 0.0019663695711642504, + 0.0015439167618751526, + 0.00374426762573421, + 0.002450102474540472, + 0.002170755062252283, + 0.0028867963701486588, + 0.0024083007592707872, + 0.0028182649984955788, + 0.0014880243688821793, + 0.0028925975784659386, + 0.0022830339148640633, + 0.002531705889850855, + 0.003263090504333377, + 0.0031057680025696754, + 0.002606719732284546, + 0.001653105253353715, + 0.0016757837729528546, + 0.0020551844500005245, + 0.005154315382242203, + 0.0040736026130616665, + 0.029592445120215416, + 0.006905713118612766, + 0.0038497166242450476, + 0.00540811475366354, + 0.007163750473409891, + 0.018658434972167015, + 0.0013905699597671628, + 0.0019715414382517338, + 0.003102450165897608, + 0.0028570215217769146, + 0.001538537093438208, + 0.0018543758196756244, + 0.002349336165934801, + 0.0010736502008512616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -1.2236831188201904 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002211269922554493 + ], + "opposite_logprob_tokens": [ + -6.127211093902588 + ], + "logprob_reward": -0.002211269922554493, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002211269922554493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08069110661745071, + 0.12185262888669968, + 0.0839972272515297, + 0.4138323664665222, + 0.11868762969970703, + 0.07581811398267746, + 0.031116440892219543, + 0.02460876852273941, + 0.03997647389769554, + 0.009419248439371586 + ], + "credit_map": [ + 0.008816979825496674, + 0.010731484740972519, + 0.0106730405241251, + 0.012484284117817879, + 0.009951673448085785, + 0.011539405211806297, + 0.009230246767401695, + 0.007263996172696352, + 0.018971331417560577, + 0.015652785077691078, + 0.01372467540204525, + 0.009701699949800968, + 0.022211426869034767, + 0.022848820313811302, + 0.01000123843550682, + 0.008740651421248913, + 0.007522242609411478, + 0.008228993974626064, + 0.010199415497481823, + 0.009425446391105652, + 0.01787823811173439, + 0.007180371321737766, + 0.013802092522382736, + 0.00976043101400137, + 0.04474635049700737, + 0.028561275452375412, + 0.06415113806724548, + 0.04267056658864021, + 0.03348991274833679, + 0.05356520414352417, + 0.05254927650094032, + 0.09409864991903305, + 0.009844706393778324, + 0.010125765576958656, + 0.01621444895863533, + 0.014161515049636364, + 0.013608994893729687, + 0.012124396860599518, + 0.015991292893886566, + 0.026616493239998817, + 0.009484006091952324, + 0.01027403213083744, + 0.007256159093230963, + 0.009445661678910255, + 0.017710762098431587, + 0.00743100605905056, + 0.006989717949181795, + 0.007226769812405109, + 0.004015565384179354, + 0.004866548348218203, + 0.0029638074338436127, + 0.004483535420149565, + 0.0034397714771330357, + 0.004638336598873138, + 0.0038205981254577637, + 0.002888278104364872, + 0.0023074133787304163, + 0.0032179001718759537, + 0.005460389889776707, + 0.0035585195291787386, + 0.002702658763155341, + 0.001977087464183569, + 0.002638330450281501, + 0.002746469108387828, + 0.004151675850152969, + 0.003789047244936228, + 0.009999946691095829, + 0.0034340398851782084, + 0.0037843252066522837, + 0.004847332369536161, + 0.0034293043427169323, + 0.006540799047797918, + 0.0007618906674906611, + 0.0010870912810787559, + 0.0014657080173492432, + 0.0012639183551073074, + 0.001258685952052474, + 0.0013707653852179646, + 0.0014858823269605637, + 0.000725307036191225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.002211269922554493 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032477229833603 + ], + "opposite_logprob_tokens": [ + -2.350324869155884 + ], + "logprob_reward": -0.10032477229833603, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032477229833603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18038585782051086, + 0.057519007474184036, + 0.02628953754901886, + 0.07353474944829941, + 0.1123201996088028, + 0.06831803917884827, + 0.40176430344581604, + 0.022836264222860336, + 0.043325357139110565, + 0.013706632889807224 + ], + "credit_map": [ + 0.025121988728642464, + 0.020255018025636673, + 0.018860451877117157, + 0.02332332730293274, + 0.019952869042754173, + 0.02646319940686226, + 0.024291928857564926, + 0.02211708202958107, + 0.007731150835752487, + 0.008421721868216991, + 0.006406809203326702, + 0.004573846235871315, + 0.013426155783236027, + 0.007022179197520018, + 0.004955611657351255, + 0.004981541074812412, + 0.002824073191732168, + 0.0025669375900179148, + 0.0034488472156226635, + 0.003654572181403637, + 0.004151185508817434, + 0.0032727017533034086, + 0.003201429732143879, + 0.0031697938684374094, + 0.006959083955734968, + 0.006628495641052723, + 0.010349027812480927, + 0.010130181908607483, + 0.00837198831140995, + 0.008174724876880646, + 0.009738651104271412, + 0.013182600028812885, + 0.012093042023479939, + 0.008428385481238365, + 0.015819579362869263, + 0.01183403842151165, + 0.014865946024656296, + 0.010425601154565811, + 0.018538091331720352, + 0.02031552419066429, + 0.012074047699570656, + 0.00800985749810934, + 0.006092973053455353, + 0.009777523577213287, + 0.010255965404212475, + 0.009129753336310387, + 0.006476124748587608, + 0.006501794792711735, + 0.03522752970457077, + 0.060346826910972595, + 0.05890965834259987, + 0.06174253672361374, + 0.06650885194540024, + 0.038422517478466034, + 0.051591016352176666, + 0.029015421867370605, + 0.0031558778136968613, + 0.003394875442609191, + 0.002928870264440775, + 0.0027132669929414988, + 0.0038948848377913237, + 0.002310496987774968, + 0.0023778523318469524, + 0.0020601411815732718, + 0.0036279107443988323, + 0.0059727998450398445, + 0.007124625612050295, + 0.00842707883566618, + 0.0049915434792637825, + 0.0041029080748558044, + 0.005777069833129644, + 0.003301421646028757, + 0.0012897639535367489, + 0.0028925701044499874, + 0.002242839429527521, + 0.0016214853385463357, + 0.0014175542164593935, + 0.0011552373180165887, + 0.00151001603808254, + 0.0015771678881719708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.10032477229833603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249525420367718 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249525420367718, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249525420367718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37795937061309814, + 0.3701861500740051, + 0.050889648497104645, + 0.013353216461837292, + 0.07097721099853516, + 0.027212565764784813, + 0.03816620633006096, + 0.015296757221221924, + 0.022758521139621735, + 0.013200443238019943 + ], + "credit_map": [ + 0.0553346611559391, + 0.044298864901065826, + 0.03978899493813515, + 0.045760903507471085, + 0.038933802396059036, + 0.058136824518442154, + 0.04926534742116928, + 0.046439941972494125, + 0.0575304850935936, + 0.04467897489666939, + 0.04401032626628876, + 0.027325840666890144, + 0.06461474299430847, + 0.07431696355342865, + 0.030696509405970573, + 0.02701226994395256, + 0.006118472665548325, + 0.004718969110399485, + 0.006577919702976942, + 0.005102234892547131, + 0.007278566714376211, + 0.007879170589148998, + 0.006127685774117708, + 0.007086627185344696, + 0.002071560826152563, + 0.0016033351421356201, + 0.0019556174520403147, + 0.00139686930924654, + 0.001598179922439158, + 0.001929859397932887, + 0.001177620142698288, + 0.0016201739199459553, + 0.0079566715285182, + 0.004497797694057226, + 0.011363085359334946, + 0.006832227110862732, + 0.005871123168617487, + 0.00831784401088953, + 0.009420450776815414, + 0.016717996448278427, + 0.004730082582682371, + 0.0026927965227514505, + 0.0034140688367187977, + 0.0037581278011202812, + 0.004102202598005533, + 0.0029220038559287786, + 0.0022218439262360334, + 0.0033714366145431995, + 0.004150559660047293, + 0.004965764004737139, + 0.005540855228900909, + 0.00526896957308054, + 0.0065958318300545216, + 0.004293793346732855, + 0.004207111895084381, + 0.003143319161608815, + 0.002397105097770691, + 0.001384616713039577, + 0.002685019513592124, + 0.002286254893988371, + 0.0015723251271992922, + 0.0017001060768961906, + 0.001306304824538529, + 0.001965024508535862, + 0.0014830004656687379, + 0.0011720703914761543, + 0.007800576277077198, + 0.0018747872672975063, + 0.0015330788446590304, + 0.0013361976016312838, + 0.0020684567280113697, + 0.0054903533309698105, + 0.001443103072233498, + 0.0016195733333006501, + 0.0012937103165313601, + 0.0017047073924914002, + 0.0014626821503043175, + 0.0026638819836080074, + 0.0015196509193629026, + 0.001493132789619267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005249525420367718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures a majestic, dark brown bull, its muscular frame defined by rippling sinew beneath a gleaming hide, standing in the center of a sun-baked amphitheater. The ground is a fine layer of ochre dust, piled atop the dirt. In a sudden burst of primal power, the bull violently throws its massive head to the side, its thick, curved horns slicing the still air. This twitching motion shifts its immense weight, causing its front hooves to dig into the ground, sending a spectacular spray of golden-brown dust and fine grit into the air. In the harsh, direct rays of the midday sun, every particle of air gleams like gold, momentarily obscuring its powerful legs. The atmosphere is rich, filled with palpable tension and the raw, untamed energy of the animal, captured in pure kinetic energy.", + "seed": 32836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_15/index_4843_seed_32836_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3217138899490237, + "vlm_logprob": -0.3217138899490237, + "vlm_yes_no_margin": 2.7500000059604646, + "vlm_token_credit": -0.3217138899490237, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15226897597312927 + ], + "opposite_logprob_tokens": [ + -2.402268886566162 + ], + "logprob_reward": -0.15226897597312927, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.15226897597312927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14409708976745605, + 0.0894499197602272, + 0.10802248120307922, + 0.05103011801838875, + 0.07791125029325485, + 0.05643019825220108, + 0.11192796379327774, + 0.1713179051876068, + 0.13568845391273499, + 0.054124582558870316 + ], + "credit_map": [ + 0.023951783776283264, + 0.018793664872646332, + 0.01978691667318344, + 0.021508697420358658, + 0.02300320565700531, + 0.016123764216899872, + 0.010835999622941017, + 0.01009306963533163, + 0.013785515911877155, + 0.008962691761553288, + 0.01527717337012291, + 0.02378387190401554, + 0.010187817737460136, + 0.007971562445163727, + 0.0037199684884399176, + 0.005761308595538139, + 0.011470873840153217, + 0.014549114741384983, + 0.026581350713968277, + 0.015225795097649097, + 0.012838531285524368, + 0.012141377665102482, + 0.007291965652257204, + 0.007923468947410583, + 0.008422617800533772, + 0.005753656849265099, + 0.007923628203570843, + 0.00538780028000474, + 0.005967289209365845, + 0.009366435930132866, + 0.0034178895875811577, + 0.004790802486240864, + 0.006685602944344282, + 0.01826535351574421, + 0.01449904777109623, + 0.009521616622805595, + 0.009494385682046413, + 0.00957068894058466, + 0.005797571036964655, + 0.004076987039297819, + 0.006195638794451952, + 0.013079973869025707, + 0.006654075812548399, + 0.006174766458570957, + 0.008719542995095253, + 0.007925843819975853, + 0.0047004334628582, + 0.0029799197800457478, + 0.009203450754284859, + 0.011047220788896084, + 0.027447165921330452, + 0.012091033160686493, + 0.0254049189388752, + 0.011403168551623821, + 0.008594580926001072, + 0.006736428011208773, + 0.02242056094110012, + 0.03172365948557854, + 0.01597815379500389, + 0.014528992585837841, + 0.03620192036032677, + 0.024523096159100533, + 0.01413294393569231, + 0.011808579787611961, + 0.021303759887814522, + 0.013363420031964779, + 0.020840920507907867, + 0.022956250235438347, + 0.01837380975484848, + 0.014481120742857456, + 0.014838243834674358, + 0.009530930779874325, + 0.007524190470576286, + 0.006466287653893232, + 0.006145010236650705, + 0.0061820619739592075, + 0.008617325685918331, + 0.011467398144304752, + 0.00361626292578876, + 0.004106040578335524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.15226897597312927 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743244647979736 + ], + "opposite_logprob_tokens": [ + -0.474324494600296 + ], + "logprob_reward": -0.9743244647979736, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9743244647979736, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07294043153524399, + 0.032195448875427246, + 0.20038576424121857, + 0.036978304386138916, + 0.03327270224690437, + 0.14595341682434082, + 0.17434601485729218, + 0.1551484614610672, + 0.06327321380376816, + 0.08550629019737244 + ], + "credit_map": [ + 0.00933300331234932, + 0.00791232194751501, + 0.011683148331940174, + 0.011566830798983574, + 0.009930044412612915, + 0.010691944509744644, + 0.004895453806966543, + 0.006927679758518934, + 0.0037914207205176353, + 0.004657093435525894, + 0.004603641107678413, + 0.006721172481775284, + 0.00432866346091032, + 0.003838198957964778, + 0.0017682899488136172, + 0.0024869712069630623, + 0.024768579751253128, + 0.04393411800265312, + 0.02795010805130005, + 0.02685159258544445, + 0.023522067815065384, + 0.02140766941010952, + 0.015193717554211617, + 0.01675790548324585, + 0.004177950322628021, + 0.003897215938195586, + 0.004692069720476866, + 0.004693285562098026, + 0.006151712965220213, + 0.007003952283412218, + 0.0026471444871276617, + 0.003714973572641611, + 0.0032221938017755747, + 0.006398126017302275, + 0.004808187484741211, + 0.003635321743786335, + 0.004612016957253218, + 0.004916714504361153, + 0.003341342555359006, + 0.002338798949494958, + 0.017491720616817474, + 0.031163355335593224, + 0.014358623884618282, + 0.015096690505743027, + 0.02542683109641075, + 0.020459694787859917, + 0.012356767430901527, + 0.009599732235074043, + 0.014675966463983059, + 0.03796137869358063, + 0.011384464800357819, + 0.012723519466817379, + 0.050944577902555466, + 0.01959856040775776, + 0.01606471836566925, + 0.010992836207151413, + 0.015365827828645706, + 0.030829189345240593, + 0.025177055969834328, + 0.01374576985836029, + 0.023901021108031273, + 0.019457383081316948, + 0.01641691103577614, + 0.0102553004398942, + 0.009189171716570854, + 0.007274305447936058, + 0.009794903919100761, + 0.005342309363186359, + 0.012285253964364529, + 0.009937520138919353, + 0.003567752195522189, + 0.005881997756659985, + 0.007953902706503868, + 0.009621389210224152, + 0.010279327630996704, + 0.010919778607785702, + 0.011233802884817123, + 0.015246291644871235, + 0.010781452059745789, + 0.009470347315073013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.9743244647979736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017241863533854485 + ], + "opposite_logprob_tokens": [ + -6.3767242431640625 + ], + "logprob_reward": -0.0017241863533854485, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017241863533854485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15352389216423035, + 0.07250960916280746, + 0.10501360148191452, + 0.055732596665620804, + 0.0676351860165596, + 0.07057618349790573, + 0.3314550817012787, + 0.05596202611923218, + 0.04156378656625748, + 0.04602808505296707 + ], + "credit_map": [ + 0.026162128895521164, + 0.020522255450487137, + 0.021016476675868034, + 0.02475317008793354, + 0.020251860842108727, + 0.015739044174551964, + 0.01055670902132988, + 0.014522242359817028, + 0.008758202195167542, + 0.00938094686716795, + 0.009121130220592022, + 0.012498773634433746, + 0.012667229399085045, + 0.008435997180640697, + 0.0061887940391898155, + 0.005458530504256487, + 0.011583041399717331, + 0.013519905507564545, + 0.01827569305896759, + 0.014219177886843681, + 0.015569178387522697, + 0.014783136546611786, + 0.008773316629230976, + 0.008290164172649384, + 0.007717869244515896, + 0.005713840946555138, + 0.007230221293866634, + 0.005720726680010557, + 0.007289893925189972, + 0.010746343061327934, + 0.005496589001268148, + 0.005817110650241375, + 0.006167167332023382, + 0.01481297891587019, + 0.01532220933586359, + 0.008392729796469212, + 0.007415907457470894, + 0.007072771433740854, + 0.004851645324379206, + 0.0035997836384922266, + 0.007816774770617485, + 0.017326679080724716, + 0.008272952400147915, + 0.006659238133579493, + 0.010646121576428413, + 0.01085730455815792, + 0.005084263160824776, + 0.003912845626473427, + 0.0253955889493227, + 0.07540420442819595, + 0.022213751450181007, + 0.023567035794258118, + 0.097167007625103, + 0.036546740680933, + 0.02939870022237301, + 0.021762065589427948, + 0.005358309019356966, + 0.008763091638684273, + 0.004933346528559923, + 0.004536980763077736, + 0.012376547791063786, + 0.010133247822523117, + 0.005123557522892952, + 0.004736941307783127, + 0.004088479094207287, + 0.004731259774416685, + 0.008416884578764439, + 0.005970739293843508, + 0.005833338014781475, + 0.0064178407192230225, + 0.002991367829963565, + 0.0031138767953962088, + 0.0053329081274569035, + 0.004450525157153606, + 0.004896052647382021, + 0.006197941955178976, + 0.006205180659890175, + 0.008390354923903942, + 0.005885126069188118, + 0.004669996444135904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.0017241863533854485 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47430312633514404 + ], + "opposite_logprob_tokens": [ + -0.974303126335144 + ], + "logprob_reward": -0.47430312633514404, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47430312633514404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15975792706012726, + 0.15865659713745117, + 0.06289844959974289, + 0.06114905700087547, + 0.10865039378404617, + 0.08513213694095612, + 0.13230149447917938, + 0.0723164826631546, + 0.07080832123756409, + 0.08832911401987076 + ], + "credit_map": [ + 0.014768688008189201, + 0.020250268280506134, + 0.02292955480515957, + 0.019201621413230896, + 0.028731366619467735, + 0.02288142219185829, + 0.014882855117321014, + 0.016112154349684715, + 0.022463258355855942, + 0.015764234587550163, + 0.03489253297448158, + 0.040360841900110245, + 0.016420647501945496, + 0.011282408609986305, + 0.007512497249990702, + 0.009960171766579151, + 0.008991027250885963, + 0.0093118567019701, + 0.01028958149254322, + 0.006785104051232338, + 0.008997122757136822, + 0.007097890134900808, + 0.006150383967906237, + 0.005275479052215815, + 0.008113802410662174, + 0.007592410314828157, + 0.009024732746183872, + 0.008408824913203716, + 0.007096674293279648, + 0.008815031498670578, + 0.00578673928976059, + 0.006310839671641588, + 0.008087429217994213, + 0.0260219257324934, + 0.028295954689383507, + 0.012313631363213062, + 0.010789459571242332, + 0.01091513317078352, + 0.00573259312659502, + 0.006494270171970129, + 0.009634673595428467, + 0.01974792405962944, + 0.008595509454607964, + 0.007858888246119022, + 0.013228322379291058, + 0.012525003403425217, + 0.007294645067304373, + 0.006247170269489288, + 0.008765734732151031, + 0.032617416232824326, + 0.017273331061005592, + 0.010350230149924755, + 0.03135426715016365, + 0.014195398427546024, + 0.009730217978358269, + 0.008014899678528309, + 0.00948136206716299, + 0.011319256387650967, + 0.007406786549836397, + 0.008510409854352474, + 0.014053303748369217, + 0.007216028403490782, + 0.007720006629824638, + 0.006609335541725159, + 0.00815620832145214, + 0.0073068528436124325, + 0.011015928350389004, + 0.014731379225850105, + 0.008911817334592342, + 0.00749348197132349, + 0.007616837043315172, + 0.005575814750045538, + 0.01096602063626051, + 0.008967545814812183, + 0.013486888259649277, + 0.011166063137352467, + 0.01088386494666338, + 0.011793731711804867, + 0.010286098346114159, + 0.010778901167213917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.47430312633514404 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948696285486221 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948696285486221, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948696285486221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11805591732263565, + 0.1607152670621872, + 0.12451084703207016, + 0.03569325804710388, + 0.24471428990364075, + 0.04463145509362221, + 0.06155741587281227, + 0.05200790613889694, + 0.08280722796916962, + 0.07530642300844193 + ], + "credit_map": [ + 0.014099973253905773, + 0.01244810875505209, + 0.01864832080900669, + 0.01916503719985485, + 0.017493201419711113, + 0.01901020109653473, + 0.007953854277729988, + 0.009237216785550117, + 0.02563653700053692, + 0.013652618043124676, + 0.02612234279513359, + 0.05501653999090195, + 0.012062525376677513, + 0.011408538557589054, + 0.006098312791436911, + 0.010717843659222126, + 0.014176796190440655, + 0.024006454274058342, + 0.02177485078573227, + 0.017237987369298935, + 0.013986058533191681, + 0.01307755708694458, + 0.009385840967297554, + 0.010865302756428719, + 0.0052061909809708595, + 0.003948320634663105, + 0.005684045143425465, + 0.0038727654609829187, + 0.003960401751101017, + 0.006455673836171627, + 0.0032501036766916513, + 0.0033157560974359512, + 0.019863761961460114, + 0.06562337279319763, + 0.06590152531862259, + 0.027760004624724388, + 0.023516958579421043, + 0.021667152643203735, + 0.011256993748247623, + 0.009124524891376495, + 0.004153026267886162, + 0.0075050340965390205, + 0.005096288397908211, + 0.004160038661211729, + 0.008024618029594421, + 0.007741679437458515, + 0.004804402124136686, + 0.003146369941532612, + 0.006129037123173475, + 0.013086747378110886, + 0.005619881674647331, + 0.00415478041395545, + 0.017566462978720665, + 0.005850365851074457, + 0.005950757302343845, + 0.0031993798911571503, + 0.0038829182740300894, + 0.009190605022013187, + 0.007924450561404228, + 0.004624677821993828, + 0.008880816400051117, + 0.007606263272464275, + 0.006104296073317528, + 0.0037938780151307583, + 0.006951254326850176, + 0.009940416552126408, + 0.014301075600087643, + 0.016675138846039772, + 0.011354454793035984, + 0.007895689457654953, + 0.008803707547485828, + 0.006885488983243704, + 0.014641287736594677, + 0.006452229805290699, + 0.013563156127929688, + 0.014028611592948437, + 0.007066851481795311, + 0.006440158002078533, + 0.006905652582645416, + 0.0062084742821753025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.005948696285486221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures a majestic, dark brown bull, its muscular frame defined by rippling sinew beneath a gleaming hide, standing in the center of a sun-baked amphitheater. The ground is a fine layer of ochre dust, piled atop the dirt. In a sudden burst of primal power, the bull violently throws its massive head to the side, its thick, curved horns slicing the still air. This twitching motion shifts its immense weight, causing its front hooves to dig into the ground, sending a spectacular spray of golden-brown dust and fine grit into the air. In the harsh, direct rays of the midday sun, every particle of air gleams like gold, momentarily obscuring its powerful legs. The atmosphere is rich, filled with palpable tension and the raw, untamed energy of the animal, captured in pure kinetic energy.", + "seed": 32837, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_15/index_4843_seed_32837_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24059877559775486, + "vlm_logprob": -0.24059877559775486, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.24059877559775486, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09092072397470474 + ], + "opposite_logprob_tokens": [ + -3.2159206867218018 + ], + "logprob_reward": -0.09092072397470474, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.09092072397470474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31787440180778503, + 0.059165630489587784, + 0.07731993496417999, + 0.04473283514380455, + 0.038920894265174866, + 0.0332987979054451, + 0.18487510085105896, + 0.0644216239452362, + 0.05982322245836258, + 0.11956755816936493 + ], + "credit_map": [ + 0.05558960139751434, + 0.03865145146846771, + 0.03930382430553436, + 0.019148578867316246, + 0.026372570544481277, + 0.08804450929164886, + 0.01843295618891716, + 0.03233092650771141, + 0.011706862598657608, + 0.0077536567114293575, + 0.0067004491575062275, + 0.0078008538112044334, + 0.007236808072775602, + 0.006855982821434736, + 0.005573790520429611, + 0.0055372221395373344, + 0.01176839042454958, + 0.007361020427197218, + 0.009535820223391056, + 0.008178235962986946, + 0.009204639121890068, + 0.007757311686873436, + 0.016570616513490677, + 0.006943893618881702, + 0.008183605037629604, + 0.008407717570662498, + 0.004446870181709528, + 0.004850142169743776, + 0.0057628341019153595, + 0.0060274917632341385, + 0.0033864551223814487, + 0.0036677164025604725, + 0.005811637733131647, + 0.004168808925896883, + 0.00434380816295743, + 0.004133247304707766, + 0.006651323288679123, + 0.005006170365959406, + 0.005220481660217047, + 0.0035854200832545757, + 0.003977420274168253, + 0.006874089129269123, + 0.0038262561429291964, + 0.005782831460237503, + 0.0031867618672549725, + 0.0029509880114346743, + 0.004043462220579386, + 0.0026569871697574854, + 0.03948062285780907, + 0.018866071477532387, + 0.02372773364186287, + 0.0215600598603487, + 0.016348158940672874, + 0.02582027204334736, + 0.011217954568564892, + 0.027854226529598236, + 0.016233714297413826, + 0.006813235580921173, + 0.007418307010084391, + 0.008082796819508076, + 0.005270928610116243, + 0.005997197236865759, + 0.0066578141413629055, + 0.007947632111608982, + 0.007369064725935459, + 0.005948049481958151, + 0.00933136697858572, + 0.008994543924927711, + 0.00931746605783701, + 0.0034176420886069536, + 0.010501650162041187, + 0.004943439271301031, + 0.00820047315210104, + 0.006038037594407797, + 0.04055709391832352, + 0.03901959955692291, + 0.01094431709498167, + 0.004499643575400114, + 0.007083448581397533, + 0.0032249512150883675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.09092072397470474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261997699737549 + ], + "opposite_logprob_tokens": [ + -0.5761997699737549 + ], + "logprob_reward": -0.8261997699737549, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261997699737549, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1252812147140503, + 0.05336020141839981, + 0.07843709737062454, + 0.11305948346853256, + 0.04217679053544998, + 0.08758869767189026, + 0.13039742410182953, + 0.07171066850423813, + 0.12850329279899597, + 0.16948512196540833 + ], + "credit_map": [ + 0.022251522168517113, + 0.017157340422272682, + 0.014636991545557976, + 0.008828643709421158, + 0.010746931657195091, + 0.03262710198760033, + 0.007452851627022028, + 0.011579831130802631, + 0.008533422835171223, + 0.006053054705262184, + 0.005933966021984816, + 0.0073804608546197414, + 0.005229907110333443, + 0.008864505216479301, + 0.005482293199747801, + 0.0058825877495110035, + 0.010203538462519646, + 0.009152418933808804, + 0.009916767477989197, + 0.007960937917232513, + 0.008656522259116173, + 0.009656122885644436, + 0.0152884004637599, + 0.0076023973524570465, + 0.012114333920180798, + 0.013375325128436089, + 0.011909115128219128, + 0.015869200229644775, + 0.022327827289700508, + 0.013873658142983913, + 0.015133964829146862, + 0.008456066250801086, + 0.005539018660783768, + 0.0046485792845487595, + 0.004326009191572666, + 0.0047089941799640656, + 0.007869397290050983, + 0.004827867262065411, + 0.006206141319125891, + 0.0040507847443223, + 0.016739195212721825, + 0.010248912498354912, + 0.00971711240708828, + 0.012591749429702759, + 0.00920792669057846, + 0.013476448133587837, + 0.00709953298792243, + 0.008507823571562767, + 0.02426888234913349, + 0.011641750112175941, + 0.017164360731840134, + 0.0191706120967865, + 0.00860022846609354, + 0.015156115405261517, + 0.010567517019808292, + 0.023827962577342987, + 0.016053372994065285, + 0.008474147878587246, + 0.009857043623924255, + 0.008846951648592949, + 0.007276211865246296, + 0.006305744405835867, + 0.006655131001025438, + 0.008242069743573666, + 0.009973793290555477, + 0.012072899378836155, + 0.018272465094923973, + 0.017535679042339325, + 0.027061548084020615, + 0.006819252856075764, + 0.02675982192158699, + 0.010007825680077076, + 0.012254077941179276, + 0.012263082899153233, + 0.04924967885017395, + 0.04798676818609238, + 0.01719103381037712, + 0.009964932687580585, + 0.014425836503505707, + 0.006149701774120331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.8261997699737549 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009314250783063471 + ], + "opposite_logprob_tokens": [ + -7.000931262969971 + ], + "logprob_reward": -0.0009314250783063471, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009314250783063471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13508935272693634, + 0.0919143483042717, + 0.07403223216533661, + 0.08005008101463318, + 0.07618115097284317, + 0.0944438949227333, + 0.11969660222530365, + 0.04694836586713791, + 0.19737252593040466, + 0.0842713862657547 + ], + "credit_map": [ + 0.018825553357601166, + 0.016661090776324272, + 0.014657747931778431, + 0.010644079186022282, + 0.019241221249103546, + 0.029281405732035637, + 0.01126509252935648, + 0.014513175003230572, + 0.018900584429502487, + 0.010503968223929405, + 0.011766425333917141, + 0.010237228125333786, + 0.012646992690861225, + 0.010052154771983624, + 0.010305887088179588, + 0.007501118816435337, + 0.011485446244478226, + 0.00800353940576315, + 0.00706839095801115, + 0.007073366083204746, + 0.008847733028233051, + 0.009459960274398327, + 0.013708577491343021, + 0.00838522333651781, + 0.006866075564175844, + 0.00901306513696909, + 0.008968731388449669, + 0.011736229062080383, + 0.014829106628894806, + 0.01130475103855133, + 0.011157490313053131, + 0.006174631882458925, + 0.008762654848396778, + 0.01015145517885685, + 0.009046847000718117, + 0.006320955231785774, + 0.015770461410284042, + 0.007883090525865555, + 0.012460394762456417, + 0.005785297602415085, + 0.010398007929325104, + 0.011153585277497768, + 0.009109027683734894, + 0.013690854422748089, + 0.015192869119346142, + 0.010466846637427807, + 0.01290199812501669, + 0.011530717834830284, + 0.0264024268835783, + 0.011875208467245102, + 0.015128045342862606, + 0.015097652561962605, + 0.008968858979642391, + 0.016540173441171646, + 0.008074907585978508, + 0.01760934665799141, + 0.005282669328153133, + 0.0069341352209448814, + 0.0046145617961883545, + 0.005698950495570898, + 0.008016606792807579, + 0.004190354607999325, + 0.008107380010187626, + 0.004103711806237698, + 0.018160441890358925, + 0.0228872187435627, + 0.027090925723314285, + 0.024731092154979706, + 0.04492760822176933, + 0.010340005159378052, + 0.038953088223934174, + 0.010282167233526707, + 0.005973259452730417, + 0.008939551189541817, + 0.02386140450835228, + 0.02049216628074646, + 0.011031108908355236, + 0.004004352726042271, + 0.00649384455755353, + 0.0034757046960294247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0009314250783063471 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813277542591095 + ], + "opposite_logprob_tokens": [ + -1.406327724456787 + ], + "logprob_reward": -0.2813277542591095, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813277542591095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05834659934043884, + 0.09506063908338547, + 0.08205607533454895, + 0.22588005661964417, + 0.1069549173116684, + 0.15193809568881989, + 0.034615274518728256, + 0.04862162098288536, + 0.14249856770038605, + 0.05402818322181702 + ], + "credit_map": [ + 0.009713287465274334, + 0.00683960784226656, + 0.00692321639508009, + 0.004140963312238455, + 0.008520478382706642, + 0.01109225582331419, + 0.005264847539365292, + 0.005851942114531994, + 0.015034575946629047, + 0.01179131306707859, + 0.01300525851547718, + 0.013116707094013691, + 0.011162840761244297, + 0.00938874389976263, + 0.010100291110575199, + 0.011460904031991959, + 0.010442987084388733, + 0.011031979694962502, + 0.008804751560091972, + 0.008814170025289059, + 0.014052380807697773, + 0.007466953247785568, + 0.014298082329332829, + 0.007144770585000515, + 0.022268982604146004, + 0.029817871749401093, + 0.020703807473182678, + 0.029058631509542465, + 0.05176782235503197, + 0.02232343889772892, + 0.032655019313097, + 0.017284467816352844, + 0.008597773499786854, + 0.011737992987036705, + 0.014372768811881542, + 0.012396900914609432, + 0.019677110016345978, + 0.00800560973584652, + 0.024485651403665543, + 0.007681112736463547, + 0.01619618758559227, + 0.02300974912941456, + 0.017618484795093536, + 0.02910996787250042, + 0.01855430379509926, + 0.011844637803733349, + 0.019400043413043022, + 0.01620471477508545, + 0.005340938922017813, + 0.00453681405633688, + 0.0033833577763289213, + 0.0035300152376294136, + 0.006158010568469763, + 0.0028332415968179703, + 0.004568201024085283, + 0.004264696501195431, + 0.008065162226557732, + 0.006436705589294434, + 0.005528157111257315, + 0.0070784008130431175, + 0.005245028529316187, + 0.0048355343751609325, + 0.004961898550391197, + 0.006470732856541872, + 0.017566988244652748, + 0.019018808379769325, + 0.014640100300312042, + 0.016522083431482315, + 0.02630593813955784, + 0.008528072386980057, + 0.02956463396549225, + 0.010351932607591152, + 0.004997461102902889, + 0.004478713497519493, + 0.01470101810991764, + 0.013503581285476685, + 0.0041804201900959015, + 0.004281923174858093, + 0.005071392748504877, + 0.002813671948388219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.2813277542591095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static (not moving)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown bull with a muscular frame visible in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bull violently throw its head to the side, causing dust to spray from its front hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a sun-baked amphitheater with ochre dust on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.00361420470289886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361420470289886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12384631484746933, + 0.0387951098382473, + 0.06322216242551804, + 0.04620947316288948, + 0.05121900141239166, + 0.02547389641404152, + 0.07985590398311615, + 0.17811720073223114, + 0.10113206505775452, + 0.2921288013458252 + ], + "credit_map": [ + 0.025832399725914, + 0.013982103206217289, + 0.013908561319112778, + 0.008335192687809467, + 0.010069677606225014, + 0.03251279145479202, + 0.007483688183128834, + 0.011721903458237648, + 0.009265630505979061, + 0.004373730625957251, + 0.004777003079652786, + 0.0039124502800405025, + 0.003162636887282133, + 0.00539395771920681, + 0.0024801266845315695, + 0.005429572891443968, + 0.011292844079434872, + 0.0060561057180166245, + 0.0073203700594604015, + 0.0063318125903606415, + 0.007254997733980417, + 0.007079422473907471, + 0.011139215901494026, + 0.006747390143573284, + 0.0060640135779976845, + 0.005424175411462784, + 0.0043716225773096085, + 0.0054802061058580875, + 0.008318735286593437, + 0.007759392727166414, + 0.005256550386548042, + 0.0035347777884453535, + 0.010387248359620571, + 0.007354798261076212, + 0.005991441197693348, + 0.004982215818017721, + 0.006313350051641464, + 0.007308597676455975, + 0.00480946758762002, + 0.004071882925927639, + 0.0053587304428219795, + 0.0031338846310973167, + 0.0026164972223341465, + 0.0033647988457232714, + 0.0023947882000356913, + 0.004016720224171877, + 0.0025382039602845907, + 0.0020502733532339334, + 0.01484302431344986, + 0.008939595893025398, + 0.010204039514064789, + 0.01375083439052105, + 0.0064623490907251835, + 0.00763506768271327, + 0.006651786621659994, + 0.011369205079972744, + 0.04688943549990654, + 0.019826635718345642, + 0.026931099593639374, + 0.02833344042301178, + 0.011528389528393745, + 0.017050962895154953, + 0.012170141562819481, + 0.015387088060379028, + 0.014158674515783787, + 0.013981364667415619, + 0.011466260068118572, + 0.013376197777688503, + 0.01734740287065506, + 0.005987887736409903, + 0.01830805093050003, + 0.006506227422505617, + 0.017455872148275375, + 0.017778566107153893, + 0.11490971595048904, + 0.10875900834798813, + 0.011325116269290447, + 0.005268456880003214, + 0.010620025917887688, + 0.006012059282511473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.00361420470289886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_tensor([1600, 1600])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_tensor([1600, 1600])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d4ea56d12b5c4df22a2a53f5e0c43c1190b53320 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_tensor([1600, 1600])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, horizontal, medium shot brings us into an ancient military tent, lit by flickering candlelight. The strategist Zhuge Liang, clad in a loose, dark blue silk robe intricately embroidered with subtle patterns, his hair pulled high in a bun. His face is thin and resolute, a neatly trimmed black beard cascading down his chest. He sits upright on a low couch covered with straw mats, leaning slightly forward, his brow furrowed, his gaze fixed intently on a scroll of pale yellow bamboo slips spread out before him. Steady and deliberate, his right hand holds a wolf-hair brush, the tip saturated with thick ink, as he carefully traces the narrow strips of bamboo. Each stroke is a testament to deep deliberation, leaving behind powerful, ancient seal characters that reveal the secrets of the battle. The only illumination comes from a single candle in a bronze candlestick beside him. The flickering flame casts shifting shadows on his focused face and the sleeves of his wide robe, illuminating a military map hanging from the heavy canvas tent wall in the background and an ancient sword leaning in the corner. The entire painting is warm and deep in color, and in the extreme silence, it exudes a sense of the tension of strategic planning and the tranquil atmosphere of a wise man in solitude.", + "seed": 31904, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_1600_seed_31904_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0314574396237731, + "vlm_logprob": -0.0314574396237731, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.0314574396237731, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02793099544942379 + ], + "opposite_logprob_tokens": [ + -3.777930974960327 + ], + "logprob_reward": -0.02793099544942379, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02793099544942379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3406323194503784, + 0.13158905506134033, + 0.0633290633559227, + 0.04981483891606331, + 0.10070336610078812, + 0.13396646082401276, + 0.0646771565079689, + 0.05409981682896614, + 0.03701726347208023, + 0.024170679971575737 + ], + "credit_map": [ + 0.03672252222895622, + 0.06262201815843582, + 0.04091208428144455, + 0.03527436405420303, + 0.05677134916186333, + 0.031447745859622955, + 0.030860571190714836, + 0.046021632850170135, + 0.013983537442982197, + 0.03377692401409149, + 0.00966249406337738, + 0.012402190826833248, + 0.016390502452850342, + 0.020675135776400566, + 0.014555557630956173, + 0.010142683982849121, + 0.006749008782207966, + 0.012789279222488403, + 0.005549875553697348, + 0.009007124230265617, + 0.007121902424842119, + 0.00935192871838808, + 0.0053021665662527084, + 0.00745777552947402, + 0.00526362843811512, + 0.008909515105187893, + 0.00447522709146142, + 0.008627526462078094, + 0.0061223856173455715, + 0.005597779527306557, + 0.005048542749136686, + 0.005770228803157806, + 0.012764626182615757, + 0.02360469289124012, + 0.00792626477777958, + 0.008634164929389954, + 0.009622334502637386, + 0.015645671635866165, + 0.007254909723997116, + 0.01525069773197174, + 0.017561553046107292, + 0.040328674018383026, + 0.011652241460978985, + 0.008890128694474697, + 0.01305475365370512, + 0.02030465565621853, + 0.008485957980155945, + 0.013688490726053715, + 0.0075025842525064945, + 0.018671449273824692, + 0.0046421438455581665, + 0.003852304071187973, + 0.007462304085493088, + 0.009439866989850998, + 0.00542031554505229, + 0.007686182856559753, + 0.006080698687583208, + 0.011503132991492748, + 0.004374500829726458, + 0.005306196864694357, + 0.0068819220177829266, + 0.006602171808481216, + 0.0034597371704876423, + 0.009891449473798275, + 0.003656734712421894, + 0.005930636078119278, + 0.0026573899667710066, + 0.007043277844786644, + 0.005364191252738237, + 0.00405347952619195, + 0.0038077523931860924, + 0.004503798205405474, + 0.002713570138439536, + 0.007418433204293251, + 0.0013979101786389947, + 0.0018886711914092302, + 0.0031779620330780745, + 0.0033193135168403387, + 0.0019490552367642522, + 0.002305760281160474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02793099544942379 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005060464143753 + ], + "opposite_logprob_tokens": [ + -2.6950507164001465 + ], + "logprob_reward": -0.07005060464143753, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005060464143753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32526853680610657, + 0.12822780013084412, + 0.14670884609222412, + 0.06890352815389633, + 0.06199690327048302, + 0.04766666889190674, + 0.04484627768397331, + 0.06146745756268501, + 0.057314448058605194, + 0.05759960040450096 + ], + "credit_map": [ + 0.0335707925260067, + 0.06602823734283447, + 0.0352601557970047, + 0.029438596218824387, + 0.05314447358250618, + 0.037757668644189835, + 0.03401104733347893, + 0.03605754300951958, + 0.013352354988455772, + 0.02764296531677246, + 0.011030186899006367, + 0.015500884503126144, + 0.01956162042915821, + 0.018653448671102524, + 0.011890427209436893, + 0.010595900006592274, + 0.013234147801995277, + 0.019105179235339165, + 0.011281616985797882, + 0.02036382630467415, + 0.03226970136165619, + 0.016536669805645943, + 0.019329557195305824, + 0.014588117599487305, + 0.008671600371599197, + 0.011133800260722637, + 0.006611257791519165, + 0.005586034618318081, + 0.01239021122455597, + 0.012844729237258434, + 0.005304201506078243, + 0.006361683830618858, + 0.006015540100634098, + 0.007035281043499708, + 0.005254770163446665, + 0.008617597632110119, + 0.012065358459949493, + 0.006867826916277409, + 0.008242588490247726, + 0.007897933945059776, + 0.004860059358179569, + 0.00778996292501688, + 0.0033354356419295073, + 0.006839464884251356, + 0.008806006982922554, + 0.005095589905977249, + 0.004811062477529049, + 0.006129080895334482, + 0.004209045320749283, + 0.01055954396724701, + 0.0027175198774784803, + 0.006099661346524954, + 0.0058467634953558445, + 0.005385512486100197, + 0.003637127112597227, + 0.006391093600541353, + 0.006504475139081478, + 0.011762279085814953, + 0.0051205772906541824, + 0.007110878825187683, + 0.00977861974388361, + 0.006618807557970285, + 0.008657988160848618, + 0.005913822446018457, + 0.006335296668112278, + 0.00891569908708334, + 0.004686952102929354, + 0.004701909609138966, + 0.011122668161988258, + 0.008766824379563332, + 0.0052459402941167355, + 0.007539150770753622, + 0.005240766331553459, + 0.00922622624784708, + 0.004444359801709652, + 0.00571468286216259, + 0.010656699538230896, + 0.007037511095404625, + 0.007246325723826885, + 0.008033022284507751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07005060464143753 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549362145364285 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549362145364285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549362145364285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3056160807609558, + 0.10945198684930801, + 0.08634672313928604, + 0.11235304921865463, + 0.06065503880381584, + 0.07359825819730759, + 0.05321991443634033, + 0.07279153168201447, + 0.04490445554256439, + 0.08106301724910736 + ], + "credit_map": [ + 0.03011215105652809, + 0.06098282337188721, + 0.029121175408363342, + 0.0351947657763958, + 0.038167260587215424, + 0.037508998066186905, + 0.034019917249679565, + 0.04050898179411888, + 0.010081362910568714, + 0.02359135076403618, + 0.007928834296762943, + 0.01032104343175888, + 0.017875418066978455, + 0.014557584188878536, + 0.00918078888207674, + 0.015915602445602417, + 0.009940534830093384, + 0.019085966050624847, + 0.008098118007183075, + 0.008910072036087513, + 0.013136095367372036, + 0.014138872735202312, + 0.0063356501050293446, + 0.0067014144733548164, + 0.01475706696510315, + 0.014279204420745373, + 0.010432001203298569, + 0.016695551574230194, + 0.013415664434432983, + 0.012516146525740623, + 0.010473624803125858, + 0.01978379115462303, + 0.007062674965709448, + 0.010394103825092316, + 0.00662187160924077, + 0.0059575471095740795, + 0.009122584015130997, + 0.006709889508783817, + 0.005861782934516668, + 0.008924586698412895, + 0.010399810038506985, + 0.017117008566856384, + 0.009071076288819313, + 0.005936911329627037, + 0.00857885554432869, + 0.00933061819523573, + 0.0057638827711343765, + 0.0074000912718474865, + 0.0065947603434324265, + 0.008601958863437176, + 0.004767260048538446, + 0.006428355351090431, + 0.00820296723395586, + 0.00626591220498085, + 0.0050088195130229, + 0.007349876221269369, + 0.00898985005915165, + 0.013094326481223106, + 0.005549962166696787, + 0.008276136592030525, + 0.011368392035365105, + 0.00803606491535902, + 0.006442335434257984, + 0.01103446539491415, + 0.005239209625869989, + 0.008573543280363083, + 0.003954222425818443, + 0.005563558544963598, + 0.0054854741320014, + 0.006386349909007549, + 0.003992523532360792, + 0.005709572695195675, + 0.010078869760036469, + 0.01981162466108799, + 0.005515813361853361, + 0.005684794392436743, + 0.009038617834448814, + 0.01228060107678175, + 0.007058796938508749, + 0.011593894101679325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012549362145364285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811394050717354 + ], + "opposite_logprob_tokens": [ + -3.288113832473755 + ], + "logprob_reward": -0.03811394050717354, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811394050717354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34050238132476807, + 0.11926761269569397, + 0.10553564876317978, + 0.07170224189758301, + 0.0730116218328476, + 0.07022290676832199, + 0.03608870133757591, + 0.1041904091835022, + 0.033892057836055756, + 0.04558643698692322 + ], + "credit_map": [ + 0.03261512517929077, + 0.040253810584545135, + 0.03224977105855942, + 0.046932850033044815, + 0.0564713291823864, + 0.04329420626163483, + 0.033052969723939896, + 0.055632323026657104, + 0.011261777952313423, + 0.016250956803560257, + 0.008395293727517128, + 0.014051631093025208, + 0.026606405153870583, + 0.015762709081172943, + 0.011235258542001247, + 0.015703579410910606, + 0.010208696126937866, + 0.020377153530716896, + 0.010135644115507603, + 0.0166307520121336, + 0.013806367293000221, + 0.012970521114766598, + 0.00978839211165905, + 0.011618117801845074, + 0.008379187434911728, + 0.01131610106676817, + 0.004711550660431385, + 0.010668197646737099, + 0.009384910576045513, + 0.00985861849039793, + 0.006967250257730484, + 0.01041642390191555, + 0.00789663102477789, + 0.014515613205730915, + 0.006401735357940197, + 0.00935677532106638, + 0.010919659398496151, + 0.008481097407639027, + 0.008181571029126644, + 0.007258536759763956, + 0.007913727313280106, + 0.017531363293528557, + 0.00720722833648324, + 0.007878843694925308, + 0.009613960050046444, + 0.009315485134720802, + 0.00488382950425148, + 0.005878466181457043, + 0.004378765355795622, + 0.008889996446669102, + 0.0033263075165450573, + 0.004440435208380222, + 0.00453632976859808, + 0.004511925857514143, + 0.0021848182659596205, + 0.0038201226852834225, + 0.010922943241894245, + 0.015117151662707329, + 0.006727570202201605, + 0.013943743892014027, + 0.02084234543144703, + 0.012424052692949772, + 0.010155904106795788, + 0.014056694693863392, + 0.0038973866030573845, + 0.006671769544482231, + 0.0028692723717540503, + 0.004047074355185032, + 0.005975270178169012, + 0.004666021093726158, + 0.0026487272698432207, + 0.0031165354885160923, + 0.00467694504186511, + 0.00773428613319993, + 0.004687211941927671, + 0.0038420793134719133, + 0.0055862125009298325, + 0.005925402510911226, + 0.004675426986068487, + 0.008458872325718403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03811394050717354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642295375466347 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642295375466347, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642295375466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3223899304866791, + 0.11010706424713135, + 0.14713378250598907, + 0.09469009935855865, + 0.03682730719447136, + 0.08629999309778214, + 0.07126541435718536, + 0.05233166739344597, + 0.05189727246761322, + 0.027057461440563202 + ], + "credit_map": [ + 0.03849717602133751, + 0.0766960009932518, + 0.02791621722280979, + 0.0321974903345108, + 0.030175795778632164, + 0.03748945891857147, + 0.021489618346095085, + 0.057928189635276794, + 0.011162964627146721, + 0.021040527150034904, + 0.010410057380795479, + 0.018887760117650032, + 0.012185193598270416, + 0.011044560000300407, + 0.007826164364814758, + 0.01754983700811863, + 0.014155334793031216, + 0.03514507785439491, + 0.01561113353818655, + 0.02098853886127472, + 0.012781421653926373, + 0.018239369615912437, + 0.010732893832027912, + 0.0194800216704607, + 0.01032072864472866, + 0.014327543787658215, + 0.008951614610850811, + 0.016099441796541214, + 0.011047416366636753, + 0.009421711787581444, + 0.00892580859363079, + 0.01559582818299532, + 0.0038447920233011246, + 0.00880575180053711, + 0.0029972188640385866, + 0.0038578633684664965, + 0.0053483578376472, + 0.004575931467115879, + 0.0035089252050966024, + 0.0038884642999619246, + 0.010346220806241035, + 0.024827225133776665, + 0.006614811718463898, + 0.006908957846462727, + 0.008181829005479813, + 0.012931359000504017, + 0.006770596839487553, + 0.009718989953398705, + 0.008696628734469414, + 0.021374110132455826, + 0.005569092463701963, + 0.005714078899472952, + 0.008928568102419376, + 0.009516925550997257, + 0.004753234330564737, + 0.006712774746119976, + 0.005435000639408827, + 0.010975204408168793, + 0.0041647437028586864, + 0.006110318936407566, + 0.00675057852640748, + 0.00754774734377861, + 0.003002436365932226, + 0.008345638401806355, + 0.005586916580796242, + 0.007070001680403948, + 0.0044079492799937725, + 0.010234683752059937, + 0.006634874735027552, + 0.004727581050246954, + 0.005570517852902412, + 0.00766474474221468, + 0.003089947160333395, + 0.006244396325200796, + 0.002705212449654937, + 0.0024386350996792316, + 0.0034734581131488085, + 0.0037236963398754597, + 0.002461867406964302, + 0.00292024714872241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008642295375466347 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium shot brings us into an ancient military tent, lit by flickering candlelight. The strategist Zhuge Liang, clad in a loose, dark blue silk robe intricately embroidered with subtle patterns, his hair pulled high in a bun. His face is thin and resolute, a neatly trimmed black beard cascading down his chest. He sits upright on a low couch covered with straw mats, leaning slightly forward, his brow furrowed, his gaze fixed intently on a scroll of pale yellow bamboo slips spread out before him. Steady and deliberate, his right hand holds a wolf-hair brush, the tip saturated with thick ink, as he carefully traces the narrow strips of bamboo. Each stroke is a testament to deep deliberation, leaving behind powerful, ancient seal characters that reveal the secrets of the battle. The only illumination comes from a single candle in a bronze candlestick beside him. The flickering flame casts shifting shadows on his focused face and the sleeves of his wide robe, illuminating a military map hanging from the heavy canvas tent wall in the background and an ancient sword leaning in the corner. The entire painting is warm and deep in color, and in the extreme silence, it exudes a sense of the tension of strategic planning and the tranquil atmosphere of a wise man in solitude.", + "seed": 31905, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_1600_seed_31905_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04496600902639329, + "vlm_logprob": -0.04496600902639329, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04496600902639329, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1501137763261795 + ], + "opposite_logprob_tokens": [ + -2.025113821029663 + ], + "logprob_reward": -0.1501137763261795, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1501137763261795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25584718585014343, + 0.07582957297563553, + 0.04461529478430748, + 0.045944541692733765, + 0.06798548996448517, + 0.08036849647760391, + 0.10128407180309296, + 0.09748879820108414, + 0.03219936415553093, + 0.19843721389770508 + ], + "credit_map": [ + 0.033665210008621216, + 0.044076915830373764, + 0.03893084451556206, + 0.03038867749273777, + 0.031164780259132385, + 0.02791401371359825, + 0.021295659244060516, + 0.02841106802225113, + 0.009575948119163513, + 0.012654959224164486, + 0.011416622437536716, + 0.00868458952754736, + 0.009743831120431423, + 0.008110200986266136, + 0.007803854066878557, + 0.007839560508728027, + 0.005984459072351456, + 0.006185311824083328, + 0.006763729266822338, + 0.005853543058037758, + 0.00436636945232749, + 0.0066782208159565926, + 0.004796668887138367, + 0.003986990079283714, + 0.004862143658101559, + 0.007982006296515465, + 0.006448688916862011, + 0.004931037314236164, + 0.0042015365324914455, + 0.006254108622670174, + 0.004628820810467005, + 0.006636195816099644, + 0.007307649124413729, + 0.01163224782794714, + 0.011092234402894974, + 0.0075516956858336926, + 0.008718895725905895, + 0.009046883322298527, + 0.00563646899536252, + 0.006999405100941658, + 0.00960320234298706, + 0.009751650504767895, + 0.011603634804487228, + 0.0069703394547104836, + 0.012745007872581482, + 0.012134921737015247, + 0.009016785770654678, + 0.00854294653981924, + 0.017690608277916908, + 0.013151640072464943, + 0.012501363642513752, + 0.007902119308710098, + 0.016655240207910538, + 0.014187132939696312, + 0.008764723315834999, + 0.01043123658746481, + 0.02018372341990471, + 0.013663514517247677, + 0.012558570131659508, + 0.006873807869851589, + 0.008634666912257671, + 0.01639314368367195, + 0.008160211145877838, + 0.011021146550774574, + 0.006033001933246851, + 0.005622767377644777, + 0.00467380415648222, + 0.00221230904571712, + 0.0031537520699203014, + 0.0042058564722537994, + 0.002774460008367896, + 0.003523406572639942, + 0.036731064319610596, + 0.04958896338939667, + 0.02074448950588703, + 0.017073504626750946, + 0.01919315755367279, + 0.022258907556533813, + 0.015897484496235847, + 0.016949618235230446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.1501137763261795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043108731508255005 + ], + "opposite_logprob_tokens": [ + -3.1681087017059326 + ], + "logprob_reward": -0.043108731508255005, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043108731508255005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18039555847644806, + 0.09201415628194809, + 0.12713727355003357, + 0.09207224100828171, + 0.12074272334575653, + 0.14158561825752258, + 0.05506809428334236, + 0.05711939185857773, + 0.03963928297162056, + 0.09422563016414642 + ], + "credit_map": [ + 0.016460373997688293, + 0.01620878465473652, + 0.02672792598605156, + 0.02547532133758068, + 0.030256720259785652, + 0.020846039056777954, + 0.026256605982780457, + 0.018163815140724182, + 0.009221785701811314, + 0.008476940914988518, + 0.008313000202178955, + 0.009892751462757587, + 0.012501501478254795, + 0.020000074058771133, + 0.012872887775301933, + 0.010735214687883854, + 0.012467513792216778, + 0.011046167463064194, + 0.007247466593980789, + 0.009024500846862793, + 0.024644160643219948, + 0.021598637104034424, + 0.02451697736978531, + 0.016591854393482208, + 0.010466130450367928, + 0.01060081459581852, + 0.007338857743889093, + 0.008133684284985065, + 0.011173024773597717, + 0.019306836649775505, + 0.012739989906549454, + 0.01231290865689516, + 0.018933873623609543, + 0.010583365336060524, + 0.01137552596628666, + 0.010584664531052113, + 0.024960804730653763, + 0.015342770144343376, + 0.01473170705139637, + 0.014230018481612206, + 0.015095745213329792, + 0.011344537138938904, + 0.011608277447521687, + 0.013508232310414314, + 0.023507822304964066, + 0.03252793848514557, + 0.014468870125710964, + 0.01952420547604561, + 0.007848942652344704, + 0.007915287278592587, + 0.00870862789452076, + 0.005427117925137281, + 0.007315334863960743, + 0.007590588182210922, + 0.005234844051301479, + 0.005027354694902897, + 0.00881165824830532, + 0.007319144904613495, + 0.007890904322266579, + 0.006079565733671188, + 0.00586360041052103, + 0.008556131273508072, + 0.006818248424679041, + 0.005780140403658152, + 0.005030187778174877, + 0.004862879868596792, + 0.00329565885476768, + 0.002760903211310506, + 0.007109520491212606, + 0.0073652109131217, + 0.004624481312930584, + 0.0045904433354735374, + 0.01135544665157795, + 0.015662528574466705, + 0.009913772344589233, + 0.008053853176534176, + 0.014984632842242718, + 0.011721891351044178, + 0.013325373642146587, + 0.009208139963448048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.043108731508255005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792031720280647 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009792031720280647, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792031720280647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.261710524559021, + 0.09498216211795807, + 0.07416603714227676, + 0.03911728784441948, + 0.0587988905608654, + 0.06580966711044312, + 0.09684821963310242, + 0.05667342618107796, + 0.07314103096723557, + 0.17875273525714874 + ], + "credit_map": [ + 0.02984754368662834, + 0.04115503281354904, + 0.04456062242388725, + 0.026981577277183533, + 0.038176655769348145, + 0.02901989407837391, + 0.027651770040392876, + 0.024317419156432152, + 0.009612389840185642, + 0.013309264555573463, + 0.011135511100292206, + 0.010748811066150665, + 0.013523026369512081, + 0.0125200767070055, + 0.013722633011639118, + 0.010410452261567116, + 0.0065141585655510426, + 0.0076670655980706215, + 0.0074111889116466045, + 0.007379168178886175, + 0.01476930920034647, + 0.013827257789671421, + 0.007783310022205114, + 0.008814578875899315, + 0.00471843546256423, + 0.003976382780820131, + 0.0041510420851409435, + 0.003117089392617345, + 0.006895862985402346, + 0.004731759428977966, + 0.00608872203156352, + 0.005437992978841066, + 0.008342425338923931, + 0.00881833117455244, + 0.004573998972773552, + 0.005558152217417955, + 0.008461562916636467, + 0.008602362126111984, + 0.007619718089699745, + 0.006822338327765465, + 0.005706769414246082, + 0.007809278089553118, + 0.009595652110874653, + 0.006273283623158932, + 0.01002754271030426, + 0.013061149045825005, + 0.006286371033638716, + 0.007049620151519775, + 0.018605461344122887, + 0.012731037102639675, + 0.010263810865581036, + 0.008676528930664062, + 0.014963885769248009, + 0.012087118811905384, + 0.009277447126805782, + 0.010242925025522709, + 0.009863539598882198, + 0.006125332321971655, + 0.007122601382434368, + 0.003409123979508877, + 0.005050802603363991, + 0.01255620364099741, + 0.00509069487452507, + 0.007455125916749239, + 0.014456458389759064, + 0.011609702371060848, + 0.006664710585027933, + 0.004303224850445986, + 0.009171494282782078, + 0.013233758509159088, + 0.005670555867254734, + 0.008031129837036133, + 0.03423633053898811, + 0.0382644385099411, + 0.022050481289625168, + 0.012217472307384014, + 0.0169177558273077, + 0.025002876296639442, + 0.014379603788256645, + 0.01568378135561943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.009792031720280647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196668475866318 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196668475866318, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196668475866318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3265122175216675, + 0.0894000455737114, + 0.14925241470336914, + 0.05510522797703743, + 0.09160241484642029, + 0.0719221755862236, + 0.05948492884635925, + 0.06483474373817444, + 0.024587582796812057, + 0.06729824095964432 + ], + "credit_map": [ + 0.036509040743112564, + 0.04569743201136589, + 0.028605204075574875, + 0.03224511817097664, + 0.05455440282821655, + 0.039322853088378906, + 0.0527542382478714, + 0.03682391718029976, + 0.009702803567051888, + 0.011962862685322762, + 0.01009245216846466, + 0.010641312226653099, + 0.014923041686415672, + 0.012386040762066841, + 0.010228436440229416, + 0.009463099762797356, + 0.01532533299177885, + 0.011233717203140259, + 0.015628404915332794, + 0.01135020051151514, + 0.02580110915005207, + 0.029409734532237053, + 0.021139368414878845, + 0.019364548847079277, + 0.00623994879424572, + 0.00937356986105442, + 0.00879516452550888, + 0.005482862237840891, + 0.0062749022617936134, + 0.00783967599272728, + 0.00554615119472146, + 0.005552954971790314, + 0.009063501842319965, + 0.013974468223750591, + 0.012685636058449745, + 0.008046437986195087, + 0.007830789312720299, + 0.016291853040456772, + 0.009160341694951057, + 0.01454939041286707, + 0.007837925106287003, + 0.011654653586447239, + 0.010241937823593616, + 0.00831404235213995, + 0.009080933406949043, + 0.008421912789344788, + 0.009265130385756493, + 0.007105641532689333, + 0.0067215776070952415, + 0.005499458871781826, + 0.006820456590503454, + 0.0056026228703558445, + 0.01280948519706726, + 0.008903518319129944, + 0.006767534650862217, + 0.006360278930515051, + 0.01262314710766077, + 0.008200881071388721, + 0.007840795442461967, + 0.004380316007882357, + 0.006664309650659561, + 0.011687683872878551, + 0.005554546136409044, + 0.007883062586188316, + 0.0027077691629529, + 0.0030599606689065695, + 0.002117870142683387, + 0.0022969443816691637, + 0.0039141057059168816, + 0.004189922474324703, + 0.002819115063175559, + 0.003481896361336112, + 0.01081879436969757, + 0.007729745469987392, + 0.008256536908447742, + 0.007221058942377567, + 0.008918266743421555, + 0.010797916911542416, + 0.006501162890344858, + 0.0070547545328736305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.018196668475866318 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036188371013849974 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036188371013849974, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036188371013849974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2754475176334381, + 0.10401816666126251, + 0.06233908608555794, + 0.06517159193754196, + 0.07457030564546585, + 0.07911249250173569, + 0.09547856450080872, + 0.1435527503490448, + 0.04447619616985321, + 0.055833324790000916 + ], + "credit_map": [ + 0.030316373333334923, + 0.03676195442676544, + 0.054900866001844406, + 0.030117005109786987, + 0.03250619396567345, + 0.032828010618686676, + 0.026981081813573837, + 0.03103601559996605, + 0.009856064803898335, + 0.013126387260854244, + 0.018084120005369186, + 0.011763776652514935, + 0.017145777121186256, + 0.01352782268077135, + 0.010735063813626766, + 0.009779156185686588, + 0.006434004753828049, + 0.006995068863034248, + 0.007993466220796108, + 0.006742466241121292, + 0.007012953050434589, + 0.008770996704697609, + 0.011448313482105732, + 0.006941814906895161, + 0.008415470831096172, + 0.012698709964752197, + 0.009168318472802639, + 0.005688242148607969, + 0.007159387227147818, + 0.008713506162166595, + 0.005912180989980698, + 0.007415782194584608, + 0.007472244556993246, + 0.010983324609696865, + 0.010584652423858643, + 0.008591340854763985, + 0.010898353531956673, + 0.01087011769413948, + 0.006894981954246759, + 0.008275290951132774, + 0.00946020893752575, + 0.010232843458652496, + 0.013201122172176838, + 0.008127081207931042, + 0.010734166018664837, + 0.00973387062549591, + 0.008276224136352539, + 0.009346971288323402, + 0.016518324613571167, + 0.01264327671378851, + 0.009673448279500008, + 0.009810385294258595, + 0.017784273251891136, + 0.011259790509939194, + 0.008750176057219505, + 0.009038891643285751, + 0.021264638751745224, + 0.017321525141596794, + 0.02300855703651905, + 0.014307397417724133, + 0.01581285521388054, + 0.022260114550590515, + 0.014492173679172993, + 0.01508548203855753, + 0.007384969387203455, + 0.007027551997452974, + 0.006767558865249157, + 0.00418434850871563, + 0.0048573631793260574, + 0.005556329153478146, + 0.004397877957671881, + 0.004300197120755911, + 0.006816711742430925, + 0.005902013275772333, + 0.00939207337796688, + 0.005310881417244673, + 0.006734958849847317, + 0.008834155276417732, + 0.006529954262077808, + 0.006312577519565821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0036188371013849974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium shot brings us into an ancient military tent, lit by flickering candlelight. The strategist Zhuge Liang, clad in a loose, dark blue silk robe intricately embroidered with subtle patterns, his hair pulled high in a bun. His face is thin and resolute, a neatly trimmed black beard cascading down his chest. He sits upright on a low couch covered with straw mats, leaning slightly forward, his brow furrowed, his gaze fixed intently on a scroll of pale yellow bamboo slips spread out before him. Steady and deliberate, his right hand holds a wolf-hair brush, the tip saturated with thick ink, as he carefully traces the narrow strips of bamboo. Each stroke is a testament to deep deliberation, leaving behind powerful, ancient seal characters that reveal the secrets of the battle. The only illumination comes from a single candle in a bronze candlestick beside him. The flickering flame casts shifting shadows on his focused face and the sleeves of his wide robe, illuminating a military map hanging from the heavy canvas tent wall in the background and an ancient sword leaning in the corner. The entire painting is warm and deep in color, and in the extreme silence, it exudes a sense of the tension of strategic planning and the tranquil atmosphere of a wise man in solitude.", + "seed": 31906, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_1600_seed_31906_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02405178975313902, + "vlm_logprob": -0.02405178975313902, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.02405178975313902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029989108443260193 + ], + "opposite_logprob_tokens": [ + -3.779989004135132 + ], + "logprob_reward": -0.029989108443260193, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.029989108443260193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.274288535118103, + 0.1278502643108368, + 0.09948107600212097, + 0.03824254125356674, + 0.08583670109510422, + 0.057731546461582184, + 0.046772900968790054, + 0.1765664517879486, + 0.04586723446846008, + 0.04736277088522911 + ], + "credit_map": [ + 0.05242515727877617, + 0.05459496006369591, + 0.04270358011126518, + 0.014436602592468262, + 0.04139941930770874, + 0.030967624858021736, + 0.01774623990058899, + 0.020014934241771698, + 0.028644030913710594, + 0.02033981867134571, + 0.024474021047353745, + 0.00811660848557949, + 0.01104625966399908, + 0.017626462504267693, + 0.006720595993101597, + 0.010882462374866009, + 0.020117374137043953, + 0.012867827899754047, + 0.016875330358743668, + 0.00827985629439354, + 0.011112149804830551, + 0.012886475771665573, + 0.006791545078158379, + 0.010550500825047493, + 0.005868175532668829, + 0.007537490222603083, + 0.0066516101360321045, + 0.0017477093497291207, + 0.0041593932546675205, + 0.004196279216557741, + 0.0049194106832146645, + 0.003162466688081622, + 0.017097139731049538, + 0.011673538014292717, + 0.01590620167553425, + 0.007168964948505163, + 0.008836725726723671, + 0.009575875476002693, + 0.006583212874829769, + 0.008995039388537407, + 0.008468396961688995, + 0.0083177974447608, + 0.009729016572237015, + 0.004239152651280165, + 0.005407445132732391, + 0.006107207387685776, + 0.009520703926682472, + 0.005941823590546846, + 0.006823667325079441, + 0.006500974763184786, + 0.006570682395249605, + 0.003376574721187353, + 0.0052228933200240135, + 0.0048355828039348125, + 0.008656134828925133, + 0.004786391276866198, + 0.028932364657521248, + 0.04429080709815025, + 0.033860232681035995, + 0.0068777333945035934, + 0.015843937173485756, + 0.024018175899982452, + 0.013406765647232533, + 0.009336412884294987, + 0.010527714155614376, + 0.009917213581502438, + 0.0065941233187913895, + 0.002365041757002473, + 0.00483124703168869, + 0.00481985742226243, + 0.0030269171111285686, + 0.0037851149681955576, + 0.005293722730129957, + 0.005905602592974901, + 0.008310484699904919, + 0.003954916261136532, + 0.007515268865972757, + 0.0046937852166593075, + 0.007452444639056921, + 0.004236539825797081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.029989108443260193 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038131725043058395 + ], + "opposite_logprob_tokens": [ + -3.2881317138671875 + ], + "logprob_reward": -0.038131725043058395, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038131725043058395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21732813119888306, + 0.0748135894536972, + 0.10369547456502914, + 0.12882116436958313, + 0.04110834747552872, + 0.10820005089044571, + 0.046894751489162445, + 0.045679833739995956, + 0.07173717021942139, + 0.16172145307064056 + ], + "credit_map": [ + 0.03858819231390953, + 0.03550070524215698, + 0.032206617295742035, + 0.014810043387115002, + 0.025502000004053116, + 0.020277725532650948, + 0.025760529562830925, + 0.024682311341166496, + 0.010500187985599041, + 0.009899687953293324, + 0.010853290557861328, + 0.005459799896925688, + 0.010305888950824738, + 0.006813642103224993, + 0.011274478398263454, + 0.00970661174505949, + 0.021837420761585236, + 0.012536684051156044, + 0.016467725858092308, + 0.007038837298750877, + 0.010462017729878426, + 0.012186350300908089, + 0.011226765811443329, + 0.011939682066440582, + 0.027479136362671852, + 0.019361693412065506, + 0.022210121154785156, + 0.008543492294847965, + 0.012874701991677284, + 0.01536690816283226, + 0.01045691967010498, + 0.012528198771178722, + 0.007415545638650656, + 0.008599492721259594, + 0.004647921770811081, + 0.001971373101696372, + 0.005975605454295874, + 0.0032777225133031607, + 0.0050417170859873295, + 0.004178968723863363, + 0.015494766645133495, + 0.013100684620440006, + 0.017479432746767998, + 0.005795290227979422, + 0.018130945041775703, + 0.012394838966429234, + 0.017014216631650925, + 0.008789877407252789, + 0.005117163062095642, + 0.0069880299270153046, + 0.006288241595029831, + 0.002631488488987088, + 0.008130206726491451, + 0.004518537782132626, + 0.009470634162425995, + 0.003750448115170002, + 0.00794160831719637, + 0.007328961510211229, + 0.005978028755635023, + 0.0024131028912961483, + 0.006208792328834534, + 0.006030441727489233, + 0.005960019305348396, + 0.0038188807666301727, + 0.006501566618680954, + 0.007906608283519745, + 0.008761732839047909, + 0.004862038418650627, + 0.011602072045207024, + 0.00918830931186676, + 0.014627651311457157, + 0.00828719325363636, + 0.012444534339010715, + 0.01443591620773077, + 0.021301638334989548, + 0.013015066273510456, + 0.040064919739961624, + 0.01729412190616131, + 0.016503185033798218, + 0.02666207030415535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.038131725043058395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786365553736687 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.009786365553736687, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786365553736687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2557491362094879, + 0.11117701977491379, + 0.08047425001859665, + 0.07613404095172882, + 0.029049130156636238, + 0.04288554936647415, + 0.08257576078176498, + 0.10040409117937088, + 0.14120452105998993, + 0.08034653961658478 + ], + "credit_map": [ + 0.056662142276763916, + 0.04012863710522652, + 0.04625042527914047, + 0.014692997559905052, + 0.03324805945158005, + 0.030712919309735298, + 0.016323484480381012, + 0.01773046888411045, + 0.020736563950777054, + 0.014803009107708931, + 0.018900126218795776, + 0.006192860193550587, + 0.019047778099775314, + 0.011304669082164764, + 0.009710597805678844, + 0.010481400415301323, + 0.01712646149098873, + 0.011624975129961967, + 0.013638176023960114, + 0.003632789012044668, + 0.013448061421513557, + 0.008032542653381824, + 0.006125553045421839, + 0.006845681928098202, + 0.0181083083152771, + 0.009281463921070099, + 0.01392934937030077, + 0.004381703212857246, + 0.009109016507863998, + 0.009611688554286957, + 0.005163708236068487, + 0.00654880004003644, + 0.005556842777878046, + 0.004183461423963308, + 0.0050808037631213665, + 0.001663730014115572, + 0.003366683842614293, + 0.0035436302423477173, + 0.0029668209608644247, + 0.002687155967578292, + 0.007689828518778086, + 0.0062227374874055386, + 0.006637118756771088, + 0.002484686207026243, + 0.006216786801815033, + 0.00478314608335495, + 0.005199067760258913, + 0.0036521698348224163, + 0.015189077705144882, + 0.017860734835267067, + 0.014311941340565681, + 0.00372072821483016, + 0.009518704377114773, + 0.011610273271799088, + 0.005211112089455128, + 0.005153178703039885, + 0.016981523483991623, + 0.01966344192624092, + 0.01670057512819767, + 0.005215522833168507, + 0.009234453551471233, + 0.014616209082305431, + 0.011816889978945255, + 0.0061754644848406315, + 0.016319260001182556, + 0.02078242227435112, + 0.01474608015269041, + 0.009614632464945316, + 0.01811523176729679, + 0.018239427357912064, + 0.026882508769631386, + 0.01650494895875454, + 0.015326276421546936, + 0.016113441437482834, + 0.012632619589567184, + 0.003851215122267604, + 0.012145030312240124, + 0.007065058685839176, + 0.004430618602782488, + 0.008782275021076202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.009786365553736687 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808273538947105 + ], + "opposite_logprob_tokens": [ + -3.529808282852173 + ], + "logprob_reward": -0.029808273538947105, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808273538947105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3016997277736664, + 0.1254301816225052, + 0.0971926599740982, + 0.057037170976400375, + 0.04142754152417183, + 0.0540374331176281, + 0.06173031032085419, + 0.07694496959447861, + 0.0656784400343895, + 0.11882153153419495 + ], + "credit_map": [ + 0.03713154047727585, + 0.03660346567630768, + 0.042125482112169266, + 0.020516548305749893, + 0.03864755854010582, + 0.03534946218132973, + 0.06479893624782562, + 0.02652675472199917, + 0.022873178124427795, + 0.02373347617685795, + 0.021832292899489403, + 0.008493823930621147, + 0.013757586479187012, + 0.015664581209421158, + 0.01052250899374485, + 0.008552728220820427, + 0.01282519195228815, + 0.012711089104413986, + 0.014824451878666878, + 0.009247196838259697, + 0.016026081517338753, + 0.01189503539353609, + 0.009307421743869781, + 0.010356184095144272, + 0.004980419296771288, + 0.007573103066533804, + 0.0073677124455571175, + 0.004260072484612465, + 0.009919336996972561, + 0.006442641839385033, + 0.010207906365394592, + 0.006285981275141239, + 0.004189054016023874, + 0.005752492696046829, + 0.005213845055550337, + 0.002554212464019656, + 0.009500899352133274, + 0.004676626529544592, + 0.0058071608655154705, + 0.003733250079676509, + 0.005559053737670183, + 0.007196298334747553, + 0.006928554270416498, + 0.00330184493213892, + 0.010072915814816952, + 0.006247033830732107, + 0.009290394373238087, + 0.005441341083496809, + 0.008290141820907593, + 0.011440937407314777, + 0.0093181561678648, + 0.005285017192363739, + 0.007487665396183729, + 0.005738206207752228, + 0.008227493613958359, + 0.005942693445831537, + 0.01060258224606514, + 0.009661904536187649, + 0.011973348446190357, + 0.0051458305679261684, + 0.007707799784839153, + 0.007231888361275196, + 0.017331944778561592, + 0.007289673667401075, + 0.009712628088891506, + 0.014294102787971497, + 0.009531821124255657, + 0.0051766191609203815, + 0.007351183332502842, + 0.006444587372243404, + 0.0059034996666014194, + 0.007263995707035065, + 0.021428830921649933, + 0.015531117096543312, + 0.021238096058368683, + 0.009284434840083122, + 0.010228459723293781, + 0.015948936343193054, + 0.016942718997597694, + 0.008218931034207344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.029808273538947105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19675907492637634, + 0.07101822644472122, + 0.134249746799469, + 0.0798589214682579, + 0.10122586041688919, + 0.11819495260715485, + 0.09739062935113907, + 0.07412221282720566, + 0.04464492201805115, + 0.0825355052947998 + ], + "credit_map": [ + 0.04124990105628967, + 0.03449501842260361, + 0.03904513269662857, + 0.011299173347651958, + 0.01800365000963211, + 0.02338278666138649, + 0.011445964686572552, + 0.017837440595030785, + 0.010623948648571968, + 0.015980318188667297, + 0.012715463526546955, + 0.004531465005129576, + 0.006994038820266724, + 0.007745556998997927, + 0.004362443462014198, + 0.008064983412623405, + 0.017683982849121094, + 0.02216683141887188, + 0.022131511941552162, + 0.010769516229629517, + 0.017161676660180092, + 0.016058756038546562, + 0.011407867074012756, + 0.016869591549038887, + 0.010049507953226566, + 0.011259716004133224, + 0.012364043854176998, + 0.006246467586606741, + 0.012086247093975544, + 0.008886557072401047, + 0.00966599676758051, + 0.009300378151237965, + 0.020363368093967438, + 0.01731288805603981, + 0.01838690973818302, + 0.006369732320308685, + 0.011006619781255722, + 0.010829858481884003, + 0.007054613437503576, + 0.00990186259150505, + 0.02126247063279152, + 0.013311567716300488, + 0.019623111933469772, + 0.00787858385592699, + 0.012609700672328472, + 0.014086338691413403, + 0.018673250451683998, + 0.010749922133982182, + 0.016881436109542847, + 0.015904046595096588, + 0.018062911927700043, + 0.006219169124960899, + 0.011238055303692818, + 0.01520167849957943, + 0.006821265909820795, + 0.007062060292810202, + 0.011332668364048004, + 0.014763298444449902, + 0.01233531441539526, + 0.0032354830764234066, + 0.006993350572884083, + 0.008460862562060356, + 0.01200435683131218, + 0.004996870644390583, + 0.006802503950893879, + 0.011914527043700218, + 0.0067724185064435005, + 0.002104154322296381, + 0.005410438869148493, + 0.004855663049966097, + 0.0026899538934230804, + 0.004095257259905338, + 0.018750401213765144, + 0.01312077883630991, + 0.015182625502347946, + 0.004294343758374453, + 0.006621787324547768, + 0.011198892258107662, + 0.007975000888109207, + 0.005391671787947416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.012543476186692715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium shot brings us into an ancient military tent, lit by flickering candlelight. The strategist Zhuge Liang, clad in a loose, dark blue silk robe intricately embroidered with subtle patterns, his hair pulled high in a bun. His face is thin and resolute, a neatly trimmed black beard cascading down his chest. He sits upright on a low couch covered with straw mats, leaning slightly forward, his brow furrowed, his gaze fixed intently on a scroll of pale yellow bamboo slips spread out before him. Steady and deliberate, his right hand holds a wolf-hair brush, the tip saturated with thick ink, as he carefully traces the narrow strips of bamboo. Each stroke is a testament to deep deliberation, leaving behind powerful, ancient seal characters that reveal the secrets of the battle. The only illumination comes from a single candle in a bronze candlestick beside him. The flickering flame casts shifting shadows on his focused face and the sleeves of his wide robe, illuminating a military map hanging from the heavy canvas tent wall in the background and an ancient sword leaning in the corner. The entire painting is warm and deep in color, and in the extreme silence, it exudes a sense of the tension of strategic planning and the tranquil atmosphere of a wise man in solitude.", + "seed": 31907, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_2/index_1600_seed_31907_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02208004631102085, + "vlm_logprob": -0.02208004631102085, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.02208004631102085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014053785242140293 + ], + "opposite_logprob_tokens": [ + -4.389053821563721 + ], + "logprob_reward": -0.014053785242140293, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014053785242140293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24732963740825653, + 0.04430820420384407, + 0.02758162096142769, + 0.02424749918282032, + 0.015331336297094822, + 0.02197832241654396, + 0.3495679497718811, + 0.08305956423282623, + 0.04597390815615654, + 0.14062194526195526 + ], + "credit_map": [ + 0.05103645846247673, + 0.05344189703464508, + 0.03194303810596466, + 0.017056046053767204, + 0.014336664229631424, + 0.060238029807806015, + 0.007530317176133394, + 0.011747204698622227, + 0.008484994061291218, + 0.008830039761960506, + 0.00445607490837574, + 0.00227119168266654, + 0.0022782068699598312, + 0.0143462298437953, + 0.0013193809427320957, + 0.0023220903240144253, + 0.005256510339677334, + 0.005834783427417278, + 0.0028513045981526375, + 0.0019802809692919254, + 0.0020324739161878824, + 0.006837846245616674, + 0.0010456644231453538, + 0.0017427578568458557, + 0.0049146926030516624, + 0.0036668682005256414, + 0.001861810451373458, + 0.001415098668076098, + 0.0014171355869621038, + 0.00789011362940073, + 0.0009660087525844574, + 0.0021157709416002035, + 0.002435466507449746, + 0.0015232485020533204, + 0.0012244676472619176, + 0.0012858568225055933, + 0.003165531437844038, + 0.0036372856702655554, + 0.0010563040850684047, + 0.0010031735291704535, + 0.004014998208731413, + 0.00526973232626915, + 0.0024264368694275618, + 0.0018554871203377843, + 0.0020363042131066322, + 0.0033115071710199118, + 0.0014880984090268612, + 0.0015757569344714284, + 0.02618473954498768, + 0.12704578042030334, + 0.01211399957537651, + 0.01776355318725109, + 0.011309529654681683, + 0.1368880569934845, + 0.006422887556254864, + 0.011839395388960838, + 0.016703426837921143, + 0.014943857677280903, + 0.0059370663948357105, + 0.0029459483921527863, + 0.003679154673591256, + 0.032024335116147995, + 0.0022051511332392693, + 0.004620620980858803, + 0.007270991336554289, + 0.008681871928274632, + 0.00685337046161294, + 0.004709124099463224, + 0.004254689905792475, + 0.007846406660974026, + 0.003199611557647586, + 0.003157839411869645, + 0.03821974992752075, + 0.012180853635072708, + 0.008670898154377937, + 0.00630166195333004, + 0.0057441494427621365, + 0.04959052801132202, + 0.00620791781693697, + 0.01370619609951973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.014053785242140293 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635594829916954 + ], + "opposite_logprob_tokens": [ + -3.6513559818267822 + ], + "logprob_reward": -0.02635594829916954, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635594829916954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03817017003893852, + 0.015306808985769749, + 0.0781860426068306, + 0.014934741891920567, + 0.01836637780070305, + 0.026080913841724396, + 0.6550926566123962, + 0.017584629356861115, + 0.010230633430182934, + 0.1260470449924469 + ], + "credit_map": [ + 0.005914868321269751, + 0.011554050259292126, + 0.0037137882318347692, + 0.002994176698848605, + 0.0038191082421690226, + 0.006897705607116222, + 0.0012845147866755724, + 0.0019919578917324543, + 0.0015413712244480848, + 0.004122321493923664, + 0.0009871950605884194, + 0.0011327523970976472, + 0.0019142951350659132, + 0.004205598961561918, + 0.0007114362670108676, + 0.0006918389117345214, + 0.010778847150504589, + 0.020472148433327675, + 0.006259588059037924, + 0.004400728270411491, + 0.004892763216048479, + 0.025615030899643898, + 0.002209142781794071, + 0.003557791467756033, + 0.002174351830035448, + 0.002931449795141816, + 0.0010335681727156043, + 0.0009425238822586834, + 0.0012329020537436008, + 0.004558381624519825, + 0.0007748391944915056, + 0.0012867243494838476, + 0.0021706153638660908, + 0.003864060854539275, + 0.001128345145843923, + 0.0009437216212972999, + 0.0018508202629163861, + 0.006214810535311699, + 0.0009459827560931444, + 0.0012480218429118395, + 0.004027579445391893, + 0.004777808208018541, + 0.001464533037506044, + 0.0011034348281100392, + 0.002159280702471733, + 0.00959610752761364, + 0.0013666101731359959, + 0.0015855588717386127, + 0.05213481932878494, + 0.2478417009115219, + 0.02068163827061653, + 0.03276680037379265, + 0.02161710150539875, + 0.24752920866012573, + 0.011539541184902191, + 0.020981835201382637, + 0.0031447261571884155, + 0.0030219010077416897, + 0.00116171024274081, + 0.0007526170811615884, + 0.0014024542178958654, + 0.006407253444194794, + 0.0006378198158927262, + 0.0010561475064605474, + 0.0017489673336967826, + 0.0014472872717306018, + 0.000771057908423245, + 0.0007019528420642018, + 0.0015566294314339757, + 0.002186649478971958, + 0.0010757531272247434, + 0.000742335629183799, + 0.0349096804857254, + 0.012595883570611477, + 0.008245238102972507, + 0.008053159341216087, + 0.007971511222422123, + 0.030215345323085785, + 0.0073151239193975925, + 0.016741091385483742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.02635594829916954 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052552176639437675 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.0052552176639437675, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052552176639437675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07424070686101913, + 0.02757158689200878, + 0.06497856974601746, + 0.04076829180121422, + 0.05507189407944679, + 0.012595249339938164, + 0.5693240761756897, + 0.033887214958667755, + 0.03497893363237381, + 0.08658339083194733 + ], + "credit_map": [ + 0.010133594274520874, + 0.018186338245868683, + 0.0068218400701880455, + 0.008128934539854527, + 0.009126318618655205, + 0.012060978449881077, + 0.003978364635258913, + 0.005804345011711121, + 0.005852592643350363, + 0.005659933667629957, + 0.003445843467488885, + 0.001807175693102181, + 0.002214777050539851, + 0.005459080450236797, + 0.0013785980409011245, + 0.0017535891383886337, + 0.00944042019546032, + 0.01589191146194935, + 0.004695478361099958, + 0.0033395474310964346, + 0.0038091829046607018, + 0.023050403222441673, + 0.0017319561447948217, + 0.003019673516973853, + 0.004777359776198864, + 0.010026516392827034, + 0.0034439407754689455, + 0.002776979934424162, + 0.002378676552325487, + 0.012460337951779366, + 0.001621249713934958, + 0.0032832305878400803, + 0.007494103629142046, + 0.01205941941589117, + 0.0034759847912937403, + 0.0027837976813316345, + 0.0033493689261376858, + 0.01997482217848301, + 0.00219533359631896, + 0.003739065257832408, + 0.001952774589881301, + 0.002779023488983512, + 0.0014218302676454186, + 0.0010003701318055391, + 0.0017474398482590914, + 0.002079521305859089, + 0.0007735108956694603, + 0.0008407795685343444, + 0.04323309287428856, + 0.2106374055147171, + 0.019206510856747627, + 0.028172938153147697, + 0.01844906248152256, + 0.21999073028564453, + 0.010383043438196182, + 0.01925135776400566, + 0.006088446360081434, + 0.0064438218250870705, + 0.0024227446410804987, + 0.001687371637672186, + 0.002295086858794093, + 0.011828347109258175, + 0.0012583360075950623, + 0.0018630614504218102, + 0.006822153925895691, + 0.0054005286656320095, + 0.002429464366286993, + 0.0020339388865977526, + 0.003896813839673996, + 0.010395599529147148, + 0.0021631051786243916, + 0.0018373287748545408, + 0.009882530197501183, + 0.02269763872027397, + 0.0059906840324401855, + 0.00394380884245038, + 0.005098313093185425, + 0.0264377873390913, + 0.0033983811736106873, + 0.009134254418313503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0052552176639437675 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866386204957962 + ], + "opposite_logprob_tokens": [ + -3.048663854598999 + ], + "logprob_reward": -0.04866386204957962, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866386204957962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07157522439956665, + 0.05292243883013725, + 0.04718678817152977, + 0.0184372179210186, + 0.02028115838766098, + 0.03267310559749603, + 0.4747087061405182, + 0.0916917696595192, + 0.05062518268823624, + 0.13989843428134918 + ], + "credit_map": [ + 0.016588114202022552, + 0.012212062254548073, + 0.008311253041028976, + 0.004709230270236731, + 0.01104417908936739, + 0.010302246548235416, + 0.003662663511931896, + 0.004745475482195616, + 0.011496498249471188, + 0.009597128257155418, + 0.004752692766487598, + 0.0030269885901361704, + 0.0031775394454598427, + 0.0166917871683836, + 0.0016311511863023043, + 0.0025486501399427652, + 0.006100824568420649, + 0.012636682018637657, + 0.0036531875375658274, + 0.0025945210363715887, + 0.002771221799775958, + 0.016114845871925354, + 0.001313938177190721, + 0.002001569140702486, + 0.0022515032906085253, + 0.0035972113255411386, + 0.0022850201930850744, + 0.0017822469817474484, + 0.0019138611387461424, + 0.0025514448061585426, + 0.0011487493757158518, + 0.0029071811586618423, + 0.0018514138646423817, + 0.004691745154559612, + 0.0012106623034924269, + 0.001247015898115933, + 0.0026892609894275665, + 0.00557744363322854, + 0.0013136019697412848, + 0.0017000138759613037, + 0.004525885451585054, + 0.007243546657264233, + 0.0019579557701945305, + 0.001647224766202271, + 0.0020637251436710358, + 0.01217438094317913, + 0.0013541539665311575, + 0.0017062335973605514, + 0.03782812878489494, + 0.17898719012737274, + 0.015293768607079983, + 0.023758845403790474, + 0.015684088692069054, + 0.17955423891544342, + 0.008518903516232967, + 0.015083533711731434, + 0.017811615020036697, + 0.01676122657954693, + 0.006053806282579899, + 0.003244437277317047, + 0.0048750159330666065, + 0.03570755571126938, + 0.002365224761888385, + 0.004872897174209356, + 0.007471121847629547, + 0.0077527533285319805, + 0.003340262221172452, + 0.0030956300906836987, + 0.008088130503892899, + 0.013620113022625446, + 0.004138435702770948, + 0.0031187329441308975, + 0.029294369742274284, + 0.022883277386426926, + 0.008138222619891167, + 0.004249420017004013, + 0.00519533222541213, + 0.0576365701854229, + 0.004653782118111849, + 0.007847461849451065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.04866386204957962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside an ancient military tent lit by flickering candlelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang wearing a loose dark blue silk robe with intricate embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character sitting upright on a low couch covered with straw mats while holding a wolf-hair brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character tracing pale yellow bamboo slips with a brush saturated in thick ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single candle in a bronze candlestick casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071418300271034 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071418300271034, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071418300271034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5943699479103088, + 0.10612387210130692, + 0.07665196061134338, + 0.02443203143775463, + 0.032182104885578156, + 0.01058820728212595, + 0.09007541835308075, + 0.01922917366027832, + 0.022210057824850082, + 0.024137184023857117 + ], + "credit_map": [ + 0.12074758857488632, + 0.16819098591804504, + 0.0839623510837555, + 0.05143066495656967, + 0.051509711891412735, + 0.07191276550292969, + 0.019720347598195076, + 0.02689550817012787, + 0.0159768033772707, + 0.02628314308822155, + 0.00804148055613041, + 0.005629534833133221, + 0.005574728827923536, + 0.038344062864780426, + 0.002301448257640004, + 0.00397266261279583, + 0.009994232095777988, + 0.02128988690674305, + 0.006241957191377878, + 0.004657164216041565, + 0.0038336729630827904, + 0.024691548198461533, + 0.0022241866681724787, + 0.0037193086463958025, + 0.003508694237098098, + 0.004940195474773645, + 0.0016047295648604631, + 0.0010637403465807438, + 0.0025341641157865524, + 0.008462972939014435, + 0.0010692777577787638, + 0.0012482572346925735, + 0.0031149520073086023, + 0.006975908763706684, + 0.0021944346372038126, + 0.0021562210749834776, + 0.002840060507878661, + 0.009844533167779446, + 0.0017960523255169392, + 0.003259941004216671, + 0.001534932292997837, + 0.0032995250076055527, + 0.0008556885295547545, + 0.000626963737886399, + 0.0007185841677710414, + 0.002288806950673461, + 0.0004576549108605832, + 0.0008060519467107952, + 0.01812167465686798, + 0.03702128306031227, + 0.004994072951376438, + 0.0038498733192682266, + 0.0038092907052487135, + 0.01768484339118004, + 0.0021806079894304276, + 0.002413771115243435, + 0.00373995047993958, + 0.003386803437024355, + 0.0012847351608797908, + 0.0009235086035914719, + 0.0010742621961981058, + 0.006962026935070753, + 0.0007201501284725964, + 0.0011377368355169892, + 0.004642408806830645, + 0.0031474113930016756, + 0.001975709106773138, + 0.0013169399462640285, + 0.0018489216454327106, + 0.0070093125104904175, + 0.0010651573538780212, + 0.001204197877086699, + 0.007762671448290348, + 0.0029785861261188984, + 0.0017046923749148846, + 0.0009639386553317308, + 0.0015258559724316, + 0.006806344725191593, + 0.0008223099866881967, + 0.0015727832214906812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.016071418300271034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_tensor([2012, 2012])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_tensor([2012, 2012])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e555e5b1884a7a35d1cecd7dd4d3dcdc1818b77f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_tensor([2012, 2012])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, low-angle, overhead shot draws the viewer's gaze toward an ancient forest shrouded in thick morning mist. The first rays of morning light struggle to penetrate the dense canopy, casting soft, sacred shafts of light in the humid air. At the center of the frame lies a burly, heavily bearded woodcutter, dressed in a classic red and black flannel plaid shirt. His sturdy arms bulge with muscle, his face weathered, his expression focused and resolute. He grips a felling axe, its handle smooth from constant use, and raises it with all his might above his head. The heavy steel blade gleams coldly in the thinning light, his body leaning back slightly as he gathers strength, like a fully drawn bow, caught in the brief, tense moment before the final blow. Around him, tall pine trees are looming in the mist, and the ground is covered with wet moss and rotten leaves. The whole scene is based on cold green and dark brown. Only the red on his shirt is a fresh embellishment, creating a dramatic atmosphere that is primitive, quiet yet contains explosive power.", + "seed": 3243, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_2012_seed_3243_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016755386255681515, + "vlm_logprob": -0.016755386255681515, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.016755386255681515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017811352387070656 + ], + "opposite_logprob_tokens": [ + -4.392811298370361 + ], + "logprob_reward": -0.017811352387070656, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017811352387070656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21784210205078125, + 0.04189275577664375, + 0.07222027331590652, + 0.11401858180761337, + 0.07882540673017502, + 0.12107371538877487, + 0.05157170444726944, + 0.09349773824214935, + 0.07087602466344833, + 0.13818177580833435 + ], + "credit_map": [ + 0.024332864210009575, + 0.024358196184039116, + 0.03230878710746765, + 0.04597384110093117, + 0.028536295518279076, + 0.02016601338982582, + 0.01579185761511326, + 0.026374267414212227, + 0.00918560940772295, + 0.002501944312825799, + 0.004273327998816967, + 0.004277809988707304, + 0.0025523218791931868, + 0.003458969993516803, + 0.0037455870769917965, + 0.01189718209207058, + 0.007996417582035065, + 0.006852738559246063, + 0.013150259852409363, + 0.012243522331118584, + 0.004679424222558737, + 0.010879216715693474, + 0.009334920905530453, + 0.00708377081900835, + 0.007564958650618792, + 0.008731545880436897, + 0.017260108143091202, + 0.02647763304412365, + 0.007173428311944008, + 0.006284224335104227, + 0.018642447888851166, + 0.021884234622120857, + 0.012035489082336426, + 0.006874161306768656, + 0.011797423474490643, + 0.011454329825937748, + 0.0051040323451161385, + 0.0070050302892923355, + 0.00918017327785492, + 0.01537476945668459, + 0.016246655955910683, + 0.019641200080513954, + 0.018477221950888634, + 0.02034851722419262, + 0.007584352511912584, + 0.011892806738615036, + 0.014227237552404404, + 0.01265572290867567, + 0.003814181312918663, + 0.002930204849690199, + 0.0067991395480930805, + 0.00679496256634593, + 0.005584217607975006, + 0.0033369327429682016, + 0.011174611747264862, + 0.011137456633150578, + 0.010449466295540333, + 0.0086445864289999, + 0.013983936980366707, + 0.01573685184121132, + 0.013040234334766865, + 0.008087308146059513, + 0.012624042108654976, + 0.01093131210654974, + 0.006342966575175524, + 0.006411046255379915, + 0.00897084642201662, + 0.010902900248765945, + 0.0114121800288558, + 0.012166463769972324, + 0.008264696225523949, + 0.006404925603419542, + 0.00945392344146967, + 0.0064561269246041775, + 0.020946821197867393, + 0.037752725183963776, + 0.011438827030360699, + 0.009003774262964725, + 0.02273324877023697, + 0.02039632759988308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.017811352387070656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00363107118755579 + ], + "opposite_logprob_tokens": [ + -5.628631114959717 + ], + "logprob_reward": -0.00363107118755579, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00363107118755579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2924290597438812, + 0.07909062504768372, + 0.03831072151660919, + 0.09282352030277252, + 0.07075314223766327, + 0.04849982634186745, + 0.05350867286324501, + 0.10293841361999512, + 0.060971081256866455, + 0.16067500412464142 + ], + "credit_map": [ + 0.02031627483665943, + 0.020966120064258575, + 0.029122034087777138, + 0.089998759329319, + 0.023450082167983055, + 0.025027479976415634, + 0.020700331777334213, + 0.06284794956445694, + 0.007903322577476501, + 0.004629834555089474, + 0.010467947460711002, + 0.011359252035617828, + 0.012446984648704529, + 0.006635936442762613, + 0.015297165140509605, + 0.010350174270570278, + 0.0024255458265542984, + 0.002704562386497855, + 0.0052842204459011555, + 0.005925494711846113, + 0.002809708472341299, + 0.0031618529465049505, + 0.006830237805843353, + 0.009169097058475018, + 0.00609592255204916, + 0.006377826444804668, + 0.015303568914532661, + 0.013719350099563599, + 0.01088455319404602, + 0.00565736461430788, + 0.021119317039847374, + 0.013665609061717987, + 0.007590821478515863, + 0.005342358257621527, + 0.0077645182609558105, + 0.009211655706167221, + 0.013044662773609161, + 0.005056927911937237, + 0.014403545297682285, + 0.008338644169270992, + 0.0029303275514394045, + 0.002791706472635269, + 0.005983079317957163, + 0.0048881578259170055, + 0.007629090454429388, + 0.005916522815823555, + 0.011669022031128407, + 0.006691914517432451, + 0.005895241629332304, + 0.0043289209716022015, + 0.006195245310664177, + 0.007505312096327543, + 0.006837296299636364, + 0.004658282734453678, + 0.009994002990424633, + 0.008094362914562225, + 0.006442364305257797, + 0.00599881075322628, + 0.009721406735479832, + 0.015415957197546959, + 0.01596692204475403, + 0.008225894533097744, + 0.02759317122399807, + 0.01357387937605381, + 0.0050925458781421185, + 0.005736581515520811, + 0.008733035065233707, + 0.009790191426873207, + 0.008771106600761414, + 0.008552180603146553, + 0.006767802406102419, + 0.0075276317074894905, + 0.011126021854579449, + 0.008821510709822178, + 0.022395599633455276, + 0.04644189774990082, + 0.01272363681346178, + 0.01150583103299141, + 0.025762638077139854, + 0.02189784310758114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.00363107118755579 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307413473725319 + ], + "opposite_logprob_tokens": [ + -3.168074131011963 + ], + "logprob_reward": -0.04307413473725319, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307413473725319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45792028307914734, + 0.05309993773698807, + 0.04605454206466675, + 0.07015939056873322, + 0.06281739473342896, + 0.04109538719058037, + 0.03316793590784073, + 0.07074257731437683, + 0.07458610087633133, + 0.09035647660493851 + ], + "credit_map": [ + 0.026647046208381653, + 0.021541811525821686, + 0.0351431742310524, + 0.15111924707889557, + 0.014467003755271435, + 0.03329019248485565, + 0.02574690245091915, + 0.149964839220047, + 0.0061934515833854675, + 0.0034009097144007683, + 0.007673857733607292, + 0.010645073838531971, + 0.004046121146529913, + 0.003671701066195965, + 0.008484553545713425, + 0.008984268642961979, + 0.004726327955722809, + 0.004620429128408432, + 0.00639072060585022, + 0.008692666888237, + 0.002454716246575117, + 0.005744035821408033, + 0.00726547883823514, + 0.006160165648907423, + 0.005889237392693758, + 0.004151694010943174, + 0.010643298737704754, + 0.011207829229533672, + 0.006372939329594374, + 0.007974264211952686, + 0.013446155935525894, + 0.010473966598510742, + 0.006466051563620567, + 0.003833914641290903, + 0.007359737996011972, + 0.009863262996077538, + 0.009656615555286407, + 0.004727848805487156, + 0.012478508055210114, + 0.00843144953250885, + 0.004598685540258884, + 0.003894903464242816, + 0.0044132135808467865, + 0.006044457666575909, + 0.00484358100220561, + 0.005814427975565195, + 0.005038028582930565, + 0.006448087748140097, + 0.0021972048562020063, + 0.002358443569391966, + 0.0027183201164007187, + 0.0038463070522993803, + 0.006592424586415291, + 0.002539809327572584, + 0.008426418527960777, + 0.004489004611968994, + 0.0037409854121506214, + 0.0031506880186498165, + 0.006450402084738016, + 0.008211507461965084, + 0.01194207463413477, + 0.005434954538941383, + 0.020181870087981224, + 0.011630086228251457, + 0.005288217216730118, + 0.005049778148531914, + 0.007105500902980566, + 0.014103110879659653, + 0.013083252124488354, + 0.005991647951304913, + 0.011445541866123676, + 0.012519053183495998, + 0.007548192050307989, + 0.005720057524740696, + 0.013166159391403198, + 0.019212830811738968, + 0.011447090655565262, + 0.007690518628805876, + 0.011949785053730011, + 0.013621827587485313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.04307413473725319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066022217273712 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066022217273712, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066022217273712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27395889163017273, + 0.07710510492324829, + 0.04449892044067383, + 0.09091091901063919, + 0.04004574567079544, + 0.058073531836271286, + 0.0860748142004013, + 0.06492578238248825, + 0.08122317492961884, + 0.18318311870098114 + ], + "credit_map": [ + 0.018578961491584778, + 0.01854936219751835, + 0.02296624891459942, + 0.07680512964725494, + 0.032121285796165466, + 0.020441120490431786, + 0.023859819397330284, + 0.06063694506883621, + 0.005566219799220562, + 0.00521607231348753, + 0.0107351616024971, + 0.017315497621893883, + 0.007250053808093071, + 0.00614243233576417, + 0.015128422528505325, + 0.009751240722835064, + 0.003443661844357848, + 0.003166552633047104, + 0.008970960043370724, + 0.007782275788486004, + 0.0036238941829651594, + 0.005360142793506384, + 0.007034734822809696, + 0.0051166946068406105, + 0.005717009771615267, + 0.004922174382954836, + 0.01196239609271288, + 0.01716758869588375, + 0.007207220885902643, + 0.0055354624055325985, + 0.024056080728769302, + 0.014342980459332466, + 0.0030702108051627874, + 0.0032464147079735994, + 0.004639174323529005, + 0.007317518349736929, + 0.004740067292004824, + 0.003300948766991496, + 0.00797924306243658, + 0.005752168130129576, + 0.0063823191449046135, + 0.008080370724201202, + 0.0064535788260400295, + 0.006881137378513813, + 0.010145759209990501, + 0.00455824239179492, + 0.00890219397842884, + 0.006669923197478056, + 0.006703448016196489, + 0.00564569653943181, + 0.008154542185366154, + 0.010722460225224495, + 0.01655740663409233, + 0.0066952151246368885, + 0.018430300056934357, + 0.013165739364922047, + 0.0039358967915177345, + 0.004055034834891558, + 0.0067485906183719635, + 0.007599775679409504, + 0.010697898454964161, + 0.006616001017391682, + 0.017277106642723083, + 0.007995469495654106, + 0.0066541400738060474, + 0.0051960707642138, + 0.0075223371386528015, + 0.01626855507493019, + 0.010945027694106102, + 0.007994416169822216, + 0.013151825405657291, + 0.013490799814462662, + 0.013831726275384426, + 0.007983645424246788, + 0.030137749388813972, + 0.061467498540878296, + 0.016354762017726898, + 0.010645695962011814, + 0.017965219914913177, + 0.02479681558907032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.016066022217273712 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031943507492542267 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.0031943507492542267, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031943507492542267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3281048536300659, + 0.12017118185758591, + 0.2084871381521225, + 0.040250156074762344, + 0.033631615340709686, + 0.026457497850060463, + 0.021732507273554802, + 0.028409946709871292, + 0.09996280819177628, + 0.09279226511716843 + ], + "credit_map": [ + 0.020472388714551926, + 0.015872439369559288, + 0.028230946511030197, + 0.1091826781630516, + 0.01754866912961006, + 0.02538044936954975, + 0.017588796094059944, + 0.09382849931716919, + 0.020963286980986595, + 0.008289935067296028, + 0.015429710038006306, + 0.020258881151676178, + 0.007212834432721138, + 0.008101469837129116, + 0.009493700228631496, + 0.030421357601881027, + 0.020410196855664253, + 0.020925290882587433, + 0.04176335781812668, + 0.034934524446725845, + 0.012977899983525276, + 0.029690051451325417, + 0.02436276711523533, + 0.02342304214835167, + 0.0038495513144880533, + 0.0026859992649406195, + 0.005838336423039436, + 0.009050819091498852, + 0.004403078928589821, + 0.004076744895428419, + 0.004574704449623823, + 0.0057709189131855965, + 0.004189503379166126, + 0.002314128214493394, + 0.005040914285928011, + 0.005031130276620388, + 0.003976250998675823, + 0.0025498990435153246, + 0.0045061795972287655, + 0.006023605819791555, + 0.0035598971880972385, + 0.0030740948859602213, + 0.003435421735048294, + 0.003638496855273843, + 0.0023043062537908554, + 0.0017575827660039067, + 0.004218253772705793, + 0.004469442181289196, + 0.002207190962508321, + 0.0015271323500201106, + 0.0030724992975592613, + 0.003707039635628462, + 0.0025189665611833334, + 0.0016779567813500762, + 0.004144905135035515, + 0.002876815851777792, + 0.0019448961829766631, + 0.0019191736355423927, + 0.003028332022950053, + 0.0045412452891469, + 0.004567352589219809, + 0.001871163840405643, + 0.006868724711239338, + 0.0036690563429147005, + 0.01098337396979332, + 0.010029083117842674, + 0.013785140588879585, + 0.017449652776122093, + 0.014698388986289501, + 0.015642110258340836, + 0.009121528826653957, + 0.008253523148596287, + 0.006593632511794567, + 0.004497469402849674, + 0.014388008043169975, + 0.025653760880231857, + 0.007466407958418131, + 0.005552837625145912, + 0.014834224246442318, + 0.013805918395519257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0031943507492542267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle, overhead shot draws the viewer's gaze toward an ancient forest shrouded in thick morning mist. The first rays of morning light struggle to penetrate the dense canopy, casting soft, sacred shafts of light in the humid air. At the center of the frame lies a burly, heavily bearded woodcutter, dressed in a classic red and black flannel plaid shirt. His sturdy arms bulge with muscle, his face weathered, his expression focused and resolute. He grips a felling axe, its handle smooth from constant use, and raises it with all his might above his head. The heavy steel blade gleams coldly in the thinning light, his body leaning back slightly as he gathers strength, like a fully drawn bow, caught in the brief, tense moment before the final blow. Around him, tall pine trees are looming in the mist, and the ground is covered with wet moss and rotten leaves. The whole scene is based on cold green and dark brown. Only the red on his shirt is a fresh embellishment, creating a dramatic atmosphere that is primitive, quiet yet contains explosive power.", + "seed": 3244, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_2012_seed_3244_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01741547998972237, + "vlm_logprob": -0.01741547998972237, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.01741547998972237, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021828515455126762 + ], + "opposite_logprob_tokens": [ + -4.271828651428223 + ], + "logprob_reward": -0.021828515455126762, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.021828515455126762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12402088940143585, + 0.0259099118411541, + 0.12616488337516785, + 0.20220805704593658, + 0.08973275125026703, + 0.03758610785007477, + 0.3257097899913788, + 0.01803727075457573, + 0.03651376813650131, + 0.014116635546088219 + ], + "credit_map": [ + 0.008554940111935139, + 0.008860365487635136, + 0.027781957760453224, + 0.017963847145438194, + 0.006477325689047575, + 0.007958873175084591, + 0.011995388194918633, + 0.0344281941652298, + 0.0034172371961176395, + 0.002305150730535388, + 0.004871046636253595, + 0.0032760570757091045, + 0.001522194710560143, + 0.002672553528100252, + 0.004571059253066778, + 0.0032746109645813704, + 0.009363415651023388, + 0.01464137528091669, + 0.01257192250341177, + 0.02083897776901722, + 0.013586590066552162, + 0.015365361236035824, + 0.025256557390093803, + 0.014540682546794415, + 0.023837288841605186, + 0.019538190215826035, + 0.01747880131006241, + 0.031031686812639236, + 0.02814541384577751, + 0.033916205167770386, + 0.020688848569989204, + 0.027571627870202065, + 0.006440624129027128, + 0.01070625800639391, + 0.013859208673238754, + 0.020198384299874306, + 0.007137537468224764, + 0.007714059203863144, + 0.01131160743534565, + 0.012365071102976799, + 0.006227404344826937, + 0.0029444126412272453, + 0.004081137478351593, + 0.003925406374037266, + 0.0039705210365355015, + 0.00782333966344595, + 0.004081452265381813, + 0.004532436840236187, + 0.025133008137345314, + 0.010352200828492641, + 0.13213062286376953, + 0.011893054470419884, + 0.009341123513877392, + 0.017832785844802856, + 0.011619784869253635, + 0.10740723460912704, + 0.0024452803190797567, + 0.002727256389334798, + 0.0024444665759801865, + 0.0019155702320858836, + 0.001965021248906851, + 0.0024918101262301207, + 0.0013118705246597528, + 0.002735994989052415, + 0.004055697936564684, + 0.003951862920075655, + 0.004186066333204508, + 0.007950243540108204, + 0.0030873676296323538, + 0.004618707112967968, + 0.004219844006001949, + 0.004443976562470198, + 0.002028771908953786, + 0.001698578940704465, + 0.0014094016514718533, + 0.0013484945520758629, + 0.002218661829829216, + 0.0018212731229141355, + 0.0014468872686848044, + 0.0021445665042847395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.021828515455126762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022142434027045965 + ], + "opposite_logprob_tokens": [ + -6.127214431762695 + ], + "logprob_reward": -0.0022142434027045965, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022142434027045965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4344337582588196, + 0.025901317596435547, + 0.031130734831094742, + 0.11354224383831024, + 0.046815771609544754, + 0.023988787084817886, + 0.2675153613090515, + 0.015590287744998932, + 0.021823249757289886, + 0.019258499145507812 + ], + "credit_map": [ + 0.027275986969470978, + 0.013303513638675213, + 0.15281902253627777, + 0.02720062807202339, + 0.013083023019134998, + 0.026113634929060936, + 0.028681430965662003, + 0.14595650136470795, + 0.004187260754406452, + 0.002343075117096305, + 0.004370295442640781, + 0.0030053488444536924, + 0.0023049877490848303, + 0.0033016528468579054, + 0.0031450935639441013, + 0.0032436035107821226, + 0.0036062500439584255, + 0.0025740396231412888, + 0.003954705316573381, + 0.004879574757069349, + 0.001993077341467142, + 0.00453084334731102, + 0.005738219246268272, + 0.003854023991152644, + 0.012001599185168743, + 0.014107261784374714, + 0.00924842618405819, + 0.013523450121283531, + 0.02155894786119461, + 0.019343528896570206, + 0.009339778684079647, + 0.01441925298422575, + 0.0052723209373652935, + 0.004269465338438749, + 0.006745772436261177, + 0.009294950403273106, + 0.005081756040453911, + 0.0048246076330542564, + 0.0050650774501264095, + 0.006261823233217001, + 0.0025761574506759644, + 0.0021176550071686506, + 0.0023105493746697903, + 0.0047274366952478886, + 0.002269098302349448, + 0.0031002340838313103, + 0.003773314878344536, + 0.00311434012837708, + 0.02069137617945671, + 0.00843504723161459, + 0.10848414152860641, + 0.008979521691799164, + 0.008563614450395107, + 0.014556585811078548, + 0.009970183484256268, + 0.0878349021077156, + 0.0017398035852238536, + 0.002059957245364785, + 0.002076206961646676, + 0.002622771542519331, + 0.0019473553402349353, + 0.0019038054160773754, + 0.0014403766253963113, + 0.0018000123091042042, + 0.0018977961735799909, + 0.002726122736930847, + 0.0033219740726053715, + 0.004773089662194252, + 0.0020786700770258904, + 0.002678505377843976, + 0.002448956947773695, + 0.001898135058581829, + 0.0019165129633620381, + 0.002183391712605953, + 0.002225336618721485, + 0.003369103418663144, + 0.0020252414979040623, + 0.002629833994433284, + 0.0022019955795258284, + 0.0027070818468928337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0022142434027045965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113366812467575 + ], + "opposite_logprob_tokens": [ + -3.2881133556365967 + ], + "logprob_reward": -0.038113366812467575, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113366812467575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2116178572177887, + 0.03428834676742554, + 0.057561613619327545, + 0.2544649541378021, + 0.06305758655071259, + 0.033613964915275574, + 0.2603702247142792, + 0.02151544578373432, + 0.02220270037651062, + 0.04130735620856285 + ], + "credit_map": [ + 0.02028028480708599, + 0.009556415490806103, + 0.0704479068517685, + 0.020424332469701767, + 0.006611992605030537, + 0.011060735210776329, + 0.014091931283473969, + 0.05914423614740372, + 0.005286616273224354, + 0.0027849720790982246, + 0.004364244174212217, + 0.0045473906211555, + 0.002116374671459198, + 0.0033316551707684994, + 0.006751994602382183, + 0.005105096381157637, + 0.003889672923833132, + 0.007617489900439978, + 0.00478092348203063, + 0.009002657607197762, + 0.0072487653233110905, + 0.005739836487919092, + 0.012203492224216461, + 0.007078771013766527, + 0.024578450247645378, + 0.03353952616453171, + 0.019843360409140587, + 0.030617082491517067, + 0.04844863712787628, + 0.04216613993048668, + 0.02311251312494278, + 0.03215921297669411, + 0.004536911845207214, + 0.009835182689130306, + 0.006486108060926199, + 0.013367155566811562, + 0.005267089232802391, + 0.005288203712552786, + 0.010069720447063446, + 0.008207205682992935, + 0.002527126343920827, + 0.003323604818433523, + 0.003507194109261036, + 0.0062194932252168655, + 0.0024316366761922836, + 0.005387859884649515, + 0.0056606195867061615, + 0.004556429572403431, + 0.0203676950186491, + 0.009149745106697083, + 0.10368729382753372, + 0.010676039382815361, + 0.007976924069225788, + 0.014315987937152386, + 0.010096292942762375, + 0.08410022407770157, + 0.00265032472088933, + 0.002109762281179428, + 0.0031549192499369383, + 0.0029396156314760447, + 0.0019297641701996326, + 0.0030841869302093983, + 0.002010330790653825, + 0.0036365387495607138, + 0.0021952532697468996, + 0.002043971559032798, + 0.0025931764394044876, + 0.00370490993373096, + 0.0026673993561416864, + 0.002380464458838105, + 0.0027985768392682076, + 0.003818944562226534, + 0.0036960269790142775, + 0.0027698855847120285, + 0.005791086237877607, + 0.004561541136354208, + 0.0034275527577847242, + 0.006039146799594164, + 0.004832027945667505, + 0.010190082713961601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.038113366812467575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184611573815346 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184611573815346, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184611573815346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28922078013420105, + 0.06428848206996918, + 0.05579373240470886, + 0.03727521002292633, + 0.061588168144226074, + 0.04453803226351738, + 0.3927029073238373, + 0.017390312626957893, + 0.014891061931848526, + 0.02231118641793728 + ], + "credit_map": [ + 0.02191889099776745, + 0.013083608821034431, + 0.0924692451953888, + 0.01708369515836239, + 0.01003536581993103, + 0.01874251663684845, + 0.01951812580227852, + 0.09636940062046051, + 0.004056315403431654, + 0.0071699414402246475, + 0.007061667274683714, + 0.008685785345733166, + 0.0047990987077355385, + 0.007424238603562117, + 0.014719498343765736, + 0.010371943935751915, + 0.005342857912182808, + 0.008642684668302536, + 0.004686323925852776, + 0.008162692189216614, + 0.008138473145663738, + 0.006163970101624727, + 0.007898510433733463, + 0.006758229807019234, + 0.004192745313048363, + 0.003509494010359049, + 0.003247726708650589, + 0.00520647969096899, + 0.006023601163178682, + 0.0066681294701993465, + 0.003670655656605959, + 0.004756386391818523, + 0.00569943618029356, + 0.010539034381508827, + 0.006638180464506149, + 0.010271544568240643, + 0.005357686895877123, + 0.007216686382889748, + 0.007576724514365196, + 0.008288887329399586, + 0.004278927575796843, + 0.004122018348425627, + 0.004749604035168886, + 0.006917223799973726, + 0.004969016183167696, + 0.006450646091252565, + 0.007292898837476969, + 0.00575770391151309, + 0.02984321489930153, + 0.012806284241378307, + 0.15888912975788116, + 0.015669872984290123, + 0.011438528075814247, + 0.021033072844147682, + 0.014918860979378223, + 0.12810403108596802, + 0.0025508154649287462, + 0.0023075933568179607, + 0.0025480748154222965, + 0.0021413173526525497, + 0.0015418317634612322, + 0.0023098159581422806, + 0.0013430247781798244, + 0.0026478422805666924, + 0.0016090619610622525, + 0.002908094320446253, + 0.0023486895952373743, + 0.001666752970777452, + 0.0015223242808133364, + 0.0015709573635831475, + 0.0014170180074870586, + 0.0018481668084859848, + 0.002413902897387743, + 0.0020568938925862312, + 0.0023060862440615892, + 0.0034066992811858654, + 0.0029487174469977617, + 0.002170838415622711, + 0.004357502330094576, + 0.0026505496352910995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.018184611573815346 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736662704497576 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736662704497576, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736662704497576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3085216283798218, + 0.043641529977321625, + 0.19650083780288696, + 0.037451379001140594, + 0.032708872109651566, + 0.07120620459318161, + 0.20283950865268707, + 0.016708167269825935, + 0.015451147221028805, + 0.07497070729732513 + ], + "credit_map": [ + 0.019687999039888382, + 0.015780659392476082, + 0.1040523424744606, + 0.02242306061089039, + 0.010848387144505978, + 0.01840599812567234, + 0.020309166982769966, + 0.09701403230428696, + 0.006198032293468714, + 0.0027882906142622232, + 0.0070023261941969395, + 0.005451771430671215, + 0.002597078448161483, + 0.004926198627799749, + 0.008359348401427269, + 0.0063184816390275955, + 0.016228536143898964, + 0.025698164477944374, + 0.017333509400486946, + 0.03133614361286163, + 0.023208806291222572, + 0.019317060708999634, + 0.03921085596084595, + 0.02416776865720749, + 0.0037316044326871634, + 0.00528879277408123, + 0.004595267586410046, + 0.004687138367444277, + 0.006367390509694815, + 0.004593772813677788, + 0.003361958311870694, + 0.004825452342629433, + 0.0047362749464809895, + 0.0031262359116226435, + 0.0061806547455489635, + 0.0034749212209135294, + 0.003021164098754525, + 0.005839631427079439, + 0.0029074675403535366, + 0.00342252547852695, + 0.006132865324616432, + 0.004475264344364405, + 0.005136758089065552, + 0.006839048583060503, + 0.007318795658648014, + 0.022270729765295982, + 0.00985630415380001, + 0.009176437743008137, + 0.015550104901194572, + 0.006510757375508547, + 0.08233852684497833, + 0.00714817363768816, + 0.005912030581384897, + 0.010801821015775204, + 0.007732663303613663, + 0.06684542447328568, + 0.002504410920664668, + 0.0019387849606573582, + 0.0024393422063440084, + 0.001938426517881453, + 0.002361085033044219, + 0.0022759330458939075, + 0.001307708560489118, + 0.0019424748606979847, + 0.0015918525168672204, + 0.0015669639687985182, + 0.0023968808818608522, + 0.0018287209095433354, + 0.002170780673623085, + 0.0017076770309358835, + 0.0022721269633620977, + 0.0019161441596224904, + 0.00720257917419076, + 0.005135067272931337, + 0.007872781716287136, + 0.01026543602347374, + 0.007359730079770088, + 0.008518528193235397, + 0.015802742913365364, + 0.012813840061426163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.006736662704497576 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle, overhead shot draws the viewer's gaze toward an ancient forest shrouded in thick morning mist. The first rays of morning light struggle to penetrate the dense canopy, casting soft, sacred shafts of light in the humid air. At the center of the frame lies a burly, heavily bearded woodcutter, dressed in a classic red and black flannel plaid shirt. His sturdy arms bulge with muscle, his face weathered, his expression focused and resolute. He grips a felling axe, its handle smooth from constant use, and raises it with all his might above his head. The heavy steel blade gleams coldly in the thinning light, his body leaning back slightly as he gathers strength, like a fully drawn bow, caught in the brief, tense moment before the final blow. Around him, tall pine trees are looming in the mist, and the ground is covered with wet moss and rotten leaves. The whole scene is based on cold green and dark brown. Only the red on his shirt is a fresh embellishment, creating a dramatic atmosphere that is primitive, quiet yet contains explosive power.", + "seed": 3245, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_2012_seed_3245_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010883833467960357, + "vlm_logprob": -0.010883833467960357, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.010883833467960357, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012756414711475372 + ], + "opposite_logprob_tokens": [ + -4.76275634765625 + ], + "logprob_reward": -0.012756414711475372, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012756414711475372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3436444401741028, + 0.07070724666118622, + 0.10384095460176468, + 0.048149894922971725, + 0.06209363415837288, + 0.0439179465174675, + 0.09507614374160767, + 0.11951343715190887, + 0.052517976611852646, + 0.060538385063409805 + ], + "credit_map": [ + 0.03104175440967083, + 0.024848587810993195, + 0.05293605104088783, + 0.05823906883597374, + 0.036163147538900375, + 0.04516780003905296, + 0.05480031296610832, + 0.040447723120450974, + 0.0056022717617452145, + 0.006243384908884764, + 0.01001929584890604, + 0.010765735991299152, + 0.007586271967738867, + 0.008609716780483723, + 0.013049843721091747, + 0.008830728940665722, + 0.012561238370835781, + 0.007113352417945862, + 0.016102468594908714, + 0.01480916328728199, + 0.013596322387456894, + 0.010502736084163189, + 0.015765588730573654, + 0.013390090316534042, + 0.005723102018237114, + 0.0033708575647324324, + 0.008745416067540646, + 0.005038006696850061, + 0.004219016060233116, + 0.009195227175951004, + 0.006259848829358816, + 0.005598417017608881, + 0.007790577132254839, + 0.011089740321040154, + 0.005838386714458466, + 0.0075317141599953175, + 0.006868466734886169, + 0.007660293485969305, + 0.00800296850502491, + 0.007311491295695305, + 0.00671979459002614, + 0.004435967653989792, + 0.007745471782982349, + 0.004523089621216059, + 0.00403551384806633, + 0.007424577139317989, + 0.0043698870576918125, + 0.004663643892854452, + 0.009879501536488533, + 0.009431960992515087, + 0.012783624231815338, + 0.011072084307670593, + 0.00660916930064559, + 0.014017593115568161, + 0.01865275204181671, + 0.012629454024136066, + 0.009277787990868092, + 0.008275323547422886, + 0.012630238197743893, + 0.024597402662038803, + 0.015465793199837208, + 0.011600899510085583, + 0.02113574557006359, + 0.016530238091945648, + 0.005122385453432798, + 0.0035896070767194033, + 0.008399808779358864, + 0.00999892596155405, + 0.004896099679172039, + 0.0051788692362606525, + 0.00798596628010273, + 0.007346317172050476, + 0.0037640880327671766, + 0.0037058054003864527, + 0.006300530396401882, + 0.008716369979083538, + 0.007775315083563328, + 0.01407960057258606, + 0.007953363470733166, + 0.008243316784501076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.012756414711475372 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002209128811955452 + ], + "opposite_logprob_tokens": [ + -6.127209186553955 + ], + "logprob_reward": -0.002209128811955452, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002209128811955452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20887340605258942, + 0.027524229139089584, + 0.0427633561193943, + 0.042006298899650574, + 0.02347603254020214, + 0.029577994719147682, + 0.018746275454759598, + 0.3747836649417877, + 0.03543664515018463, + 0.19681210815906525 + ], + "credit_map": [ + 0.015318363904953003, + 0.012077344581484795, + 0.03672961890697479, + 0.02734394371509552, + 0.02726757898926735, + 0.024679718539118767, + 0.03624630346894264, + 0.029210524633526802, + 0.003142650704830885, + 0.002535203704610467, + 0.003662318689748645, + 0.0036410498432815075, + 0.0024384153075516224, + 0.004866306204348803, + 0.0036524771712720394, + 0.003585806814953685, + 0.004820266738533974, + 0.00223785568960011, + 0.0064069489017128944, + 0.008859928697347641, + 0.004349001217633486, + 0.0039383661933243275, + 0.007292632013559341, + 0.004858354572206736, + 0.004986253567039967, + 0.003462344640865922, + 0.005603522527962923, + 0.004587637260556221, + 0.0036730896681547165, + 0.00720415823161602, + 0.007730487734079361, + 0.004758807364851236, + 0.0027653458528220654, + 0.004113795235753059, + 0.0024173015262931585, + 0.0031652802135795355, + 0.0032729669474065304, + 0.003027182538062334, + 0.0020770670380443335, + 0.002637092722579837, + 0.0030085109174251556, + 0.004159585572779179, + 0.004509694408625364, + 0.0033016069792211056, + 0.004428064450621605, + 0.0025784156750887632, + 0.004213711712509394, + 0.0033784022089093924, + 0.0025925124064087868, + 0.0020068828016519547, + 0.0026194648817181587, + 0.0018197288736701012, + 0.0022031019907444715, + 0.0023164507001638412, + 0.0031188717111945152, + 0.0020692634861916304, + 0.018392927944660187, + 0.01849120669066906, + 0.026298725977540016, + 0.10931988805532455, + 0.032214853912591934, + 0.030307842418551445, + 0.08143952488899231, + 0.058318693190813065, + 0.00400248309597373, + 0.0029930826276540756, + 0.0036212096456438303, + 0.006109464447945356, + 0.003927194979041815, + 0.005831212271004915, + 0.00419667549431324, + 0.004755320958793163, + 0.009779652580618858, + 0.009159044362604618, + 0.017321685329079628, + 0.015872137621045113, + 0.025842439383268356, + 0.05995990335941315, + 0.025998491793870926, + 0.03287876024842262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.002209128811955452 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302631452679634 + ], + "opposite_logprob_tokens": [ + -3.7733025550842285 + ], + "logprob_reward": -0.023302631452679634, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302631452679634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20695488154888153, + 0.07396548986434937, + 0.06857368350028992, + 0.03169295936822891, + 0.1572190821170807, + 0.03875967487692833, + 0.023000171408057213, + 0.19794677197933197, + 0.11321654915809631, + 0.08867070823907852 + ], + "credit_map": [ + 0.023337893187999725, + 0.013402163982391357, + 0.022471679374575615, + 0.033210426568984985, + 0.024115808308124542, + 0.02723444066941738, + 0.03744129091501236, + 0.025741172954440117, + 0.007165217772126198, + 0.005691940896213055, + 0.010332015343010426, + 0.012588909827172756, + 0.009574961848556995, + 0.0062993667088449, + 0.013230175711214542, + 0.00908290408551693, + 0.00914060976356268, + 0.0044182646088302135, + 0.010243816301226616, + 0.009766530245542526, + 0.008309978991746902, + 0.00710342638194561, + 0.010949806310236454, + 0.008641249500215054, + 0.006764829158782959, + 0.0022728622425347567, + 0.004220563918352127, + 0.003246447304263711, + 0.002209780504927039, + 0.004851844161748886, + 0.004293425008654594, + 0.0038332086987793446, + 0.022197261452674866, + 0.030211325734853745, + 0.01634431816637516, + 0.017725694924592972, + 0.014439279213547707, + 0.021621160209178925, + 0.016106590628623962, + 0.018573449924588203, + 0.006903880275785923, + 0.004074219614267349, + 0.005078171379864216, + 0.003776860423386097, + 0.003979843109846115, + 0.005996249616146088, + 0.004411805421113968, + 0.004538645036518574, + 0.0036047070752829313, + 0.00190811010543257, + 0.002907868940383196, + 0.002468549180775881, + 0.002101113088428974, + 0.0031926212832331657, + 0.0034748537000268698, + 0.0033423490822315216, + 0.011107818223536015, + 0.010609233751893044, + 0.01355830579996109, + 0.0567200593650341, + 0.015368160791695118, + 0.014669119380414486, + 0.045298147946596146, + 0.030615925788879395, + 0.010046063922345638, + 0.009985069744288921, + 0.011649183928966522, + 0.009525464847683907, + 0.022743161767721176, + 0.02068454399704933, + 0.013275875709950924, + 0.015307185240089893, + 0.007516294252127409, + 0.00591679010540247, + 0.008225314319133759, + 0.011063373647630215, + 0.01283231656998396, + 0.02089373767375946, + 0.009304734878242016, + 0.012918139807879925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.023302631452679634 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534293346107006 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012534293346107006, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534293346107006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12141828238964081, + 0.023174762725830078, + 0.02846967801451683, + 0.030722714960575104, + 0.020770728588104248, + 0.02059011533856392, + 0.025497062131762505, + 0.4719148874282837, + 0.06941212713718414, + 0.18802964687347412 + ], + "credit_map": [ + 0.014183775521814823, + 0.009757916443049908, + 0.016644610092043877, + 0.015641657635569572, + 0.016777528449892998, + 0.015429291874170303, + 0.018812241032719612, + 0.014171256683766842, + 0.002301425440236926, + 0.002550616627559066, + 0.00334506225772202, + 0.00355816469527781, + 0.0019614172633737326, + 0.0028878282755613327, + 0.0038534202612936497, + 0.002716827904805541, + 0.004674002528190613, + 0.001643374445848167, + 0.003429792355746031, + 0.005047617480158806, + 0.0023933190386742353, + 0.0029874376486986876, + 0.00510838208720088, + 0.003185752546414733, + 0.004235123284161091, + 0.0026373318396508694, + 0.0033041960559785366, + 0.0034875846467912197, + 0.003952030558139086, + 0.004622898530215025, + 0.004722632467746735, + 0.003760917577892542, + 0.0032770768739283085, + 0.0031346753239631653, + 0.002090360736474395, + 0.002878122730180621, + 0.002064055297523737, + 0.002446563448756933, + 0.0025198219809681177, + 0.0023600521963089705, + 0.00360691174864769, + 0.002588429022580385, + 0.0025000697933137417, + 0.002012475859373808, + 0.0020160945132374763, + 0.0028565668035298586, + 0.0027499108109623194, + 0.002259658183902502, + 0.0029557845555245876, + 0.0029263165779411793, + 0.003111812984570861, + 0.0032556147780269384, + 0.0024415007792413235, + 0.0021071666851639748, + 0.005513899028301239, + 0.00318496604450047, + 0.024078700691461563, + 0.02408638969063759, + 0.033767376095056534, + 0.13340319693088531, + 0.041777800768613815, + 0.03641478717327118, + 0.10524675250053406, + 0.07313987612724304, + 0.00713465828448534, + 0.0051967063918709755, + 0.006597793661057949, + 0.007958419620990753, + 0.006895974278450012, + 0.018483681604266167, + 0.006939259823411703, + 0.01020563393831253, + 0.010735943913459778, + 0.010478389449417591, + 0.020101694390177727, + 0.017748361453413963, + 0.024720000103116035, + 0.046842943876981735, + 0.02683367021381855, + 0.030568644404411316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.012534293346107006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616699017584324 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616699017584324, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616699017584324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19765940308570862, + 0.18913869559764862, + 0.03869989514350891, + 0.05114145204424858, + 0.12544800341129303, + 0.03845042362809181, + 0.05821364372968674, + 0.165871724486351, + 0.06234237924218178, + 0.07303434610366821 + ], + "credit_map": [ + 0.017132142558693886, + 0.018628163263201714, + 0.037179216742515564, + 0.028713207691907883, + 0.018707087263464928, + 0.022181501612067223, + 0.030836626887321472, + 0.024281466379761696, + 0.018314575776457787, + 0.017681073397397995, + 0.025484049692749977, + 0.029680728912353516, + 0.02118346467614174, + 0.024857448413968086, + 0.027998745441436768, + 0.023938609287142754, + 0.004301162902265787, + 0.002478719921782613, + 0.005244022700935602, + 0.007168013136833906, + 0.0026877792552113533, + 0.0035971635952591896, + 0.007626707665622234, + 0.0055963266640901566, + 0.01001724787056446, + 0.0033850602339953184, + 0.0072643267922103405, + 0.004701737780123949, + 0.0042721894569695, + 0.008086021058261395, + 0.006553002633154392, + 0.0068618664517998695, + 0.017702113837003708, + 0.025471478700637817, + 0.01250632107257843, + 0.014180092141032219, + 0.010960490442812443, + 0.01767064817249775, + 0.012988710775971413, + 0.013968153856694698, + 0.004332154057919979, + 0.006125250831246376, + 0.005107457283884287, + 0.003661639057099819, + 0.005441540386527777, + 0.0034390429500490427, + 0.006102737970650196, + 0.004240601789206266, + 0.007220713421702385, + 0.006602982059121132, + 0.006818890105932951, + 0.006728257518261671, + 0.0046936096623539925, + 0.007428914774209261, + 0.011599047109484673, + 0.007121229078620672, + 0.010954614728689194, + 0.010547874495387077, + 0.013085455633699894, + 0.041634753346443176, + 0.017066102474927902, + 0.017932366579771042, + 0.03171057626605034, + 0.022939974442124367, + 0.007198363076895475, + 0.004919837228953838, + 0.007415503729134798, + 0.00920019019395113, + 0.005467577837407589, + 0.010209612548351288, + 0.008801684714853764, + 0.00912961084395647, + 0.0060194372199475765, + 0.005419316701591015, + 0.009374966844916344, + 0.009951437823474407, + 0.009596250019967556, + 0.013179684989154339, + 0.009093020111322403, + 0.010400226339697838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.003616699017584324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, low-angle, overhead shot draws the viewer's gaze toward an ancient forest shrouded in thick morning mist. The first rays of morning light struggle to penetrate the dense canopy, casting soft, sacred shafts of light in the humid air. At the center of the frame lies a burly, heavily bearded woodcutter, dressed in a classic red and black flannel plaid shirt. His sturdy arms bulge with muscle, his face weathered, his expression focused and resolute. He grips a felling axe, its handle smooth from constant use, and raises it with all his might above his head. The heavy steel blade gleams coldly in the thinning light, his body leaning back slightly as he gathers strength, like a fully drawn bow, caught in the brief, tense moment before the final blow. Around him, tall pine trees are looming in the mist, and the ground is covered with wet moss and rotten leaves. The whole scene is based on cold green and dark brown. Only the red on his shirt is a fresh embellishment, creating a dramatic atmosphere that is primitive, quiet yet contains explosive power.", + "seed": 3246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_3/index_2012_seed_3246_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011857288749888539, + "vlm_logprob": -0.011857288749888539, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.011857288749888539, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0118456045165658 + ], + "opposite_logprob_tokens": [ + -4.761845588684082 + ], + "logprob_reward": -0.0118456045165658, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0118456045165658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3552839457988739, + 0.13809417188167572, + 0.044189296662807465, + 0.04227502644062042, + 0.06442880630493164, + 0.05758025124669075, + 0.06346260756254196, + 0.056393083184957504, + 0.03444690257310867, + 0.14384591579437256 + ], + "credit_map": [ + 0.09208333492279053, + 0.032192397862672806, + 0.056109342724084854, + 0.027048546820878983, + 0.021689601242542267, + 0.043445248156785965, + 0.03899708390235901, + 0.04371839016675949, + 0.024269195273518562, + 0.011436599306762218, + 0.021122746169567108, + 0.01768215000629425, + 0.011355812661349773, + 0.017959684133529663, + 0.01302260160446167, + 0.021245362237095833, + 0.003008256433531642, + 0.005813347175717354, + 0.0070916954427957535, + 0.0064077358692884445, + 0.0037260260432958603, + 0.004067057743668556, + 0.005786382593214512, + 0.008288792334496975, + 0.003971229773014784, + 0.0040230234153568745, + 0.005750036798417568, + 0.00569511903449893, + 0.0027179603930562735, + 0.0036006695590913296, + 0.008968201465904713, + 0.007548783905804157, + 0.010437184944748878, + 0.004989049397408962, + 0.011088614352047443, + 0.008143439888954163, + 0.004778042435646057, + 0.009330873377621174, + 0.007213687989860773, + 0.00844790879637003, + 0.008544704876840115, + 0.004365675617009401, + 0.005494373384863138, + 0.010278628207743168, + 0.0042473142966628075, + 0.0073600998148322105, + 0.01013774424791336, + 0.007151707075536251, + 0.004997702315449715, + 0.0033691804856061935, + 0.013703047297894955, + 0.00777072599157691, + 0.0032961273100227118, + 0.004724794067442417, + 0.0074913096614181995, + 0.018109718337655067, + 0.005771850235760212, + 0.0038411500863730907, + 0.01074800081551075, + 0.010853463783860207, + 0.0033908900804817677, + 0.0063989427872002125, + 0.005423617549240589, + 0.00996516365557909, + 0.0031352564692497253, + 0.0037881580647081137, + 0.005828728433698416, + 0.004403846804052591, + 0.0044156573712825775, + 0.0038895478937774897, + 0.004692052491009235, + 0.004293651320040226, + 0.014896438457071781, + 0.010638600215315819, + 0.025155972689390182, + 0.03778614476323128, + 0.008616538718342781, + 0.016047971323132515, + 0.017537377774715424, + 0.013166859745979309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0118456045165658 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028299542609602213 + ], + "opposite_logprob_tokens": [ + -5.8778300285339355 + ], + "logprob_reward": -0.0028299542609602213, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028299542609602213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25999975204467773, + 0.12000510096549988, + 0.09075499325990677, + 0.15052039921283722, + 0.056948453187942505, + 0.03171504661440849, + 0.06559465080499649, + 0.036888495087623596, + 0.04369919374585152, + 0.143873929977417 + ], + "credit_map": [ + 0.05540704354643822, + 0.023011835291981697, + 0.03491612523794174, + 0.02749931626021862, + 0.023303385823965073, + 0.03483124449849129, + 0.03178894519805908, + 0.02924187481403351, + 0.008952593430876732, + 0.007597866468131542, + 0.01623784936964512, + 0.019661113619804382, + 0.011513447389006615, + 0.00966294202953577, + 0.028365755453705788, + 0.01801353693008423, + 0.006717415526509285, + 0.008153468370437622, + 0.01089058630168438, + 0.012809962034225464, + 0.010834500193595886, + 0.009230249561369419, + 0.0191686749458313, + 0.012950140982866287, + 0.009362403303384781, + 0.007883286103606224, + 0.020394254475831985, + 0.022727856412529945, + 0.007130214478820562, + 0.00972986314445734, + 0.04131856560707092, + 0.031973958015441895, + 0.007536491844803095, + 0.004544644616544247, + 0.005808827932924032, + 0.008522880263626575, + 0.003932218998670578, + 0.00641722371801734, + 0.011250671930611134, + 0.00893549993634224, + 0.003580073593184352, + 0.002656945725902915, + 0.004174740519374609, + 0.004076257813721895, + 0.003738913917914033, + 0.004016540478914976, + 0.0046771978959441185, + 0.004794378764927387, + 0.003127750940620899, + 0.005236565135419369, + 0.008460906334221363, + 0.006541498005390167, + 0.004364641383290291, + 0.0034090231638401747, + 0.01906445622444153, + 0.015389809384942055, + 0.0031813131645321846, + 0.0024701107759028673, + 0.006077613681554794, + 0.006419711280614138, + 0.0028345808386802673, + 0.0024233218282461166, + 0.007779014762490988, + 0.005702828522771597, + 0.0032359228935092688, + 0.004534323234111071, + 0.008460571058094501, + 0.005059848073869944, + 0.003979269415140152, + 0.004456092603504658, + 0.006230915430933237, + 0.00774225452914834, + 0.010682492516934872, + 0.010813403874635696, + 0.023515336215496063, + 0.024802083149552345, + 0.017365077510476112, + 0.01499206479638815, + 0.02670414000749588, + 0.014999346807599068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0028299542609602213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329447865486145 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.02329447865486145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329447865486145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23413005471229553, + 0.07444816082715988, + 0.2150012105703354, + 0.0809796079993248, + 0.02831352688372135, + 0.03012007288634777, + 0.20502333343029022, + 0.03190373629331589, + 0.03666618466377258, + 0.06341414153575897 + ], + "credit_map": [ + 0.059122852981090546, + 0.02476009912788868, + 0.033120278269052505, + 0.01460238453000784, + 0.018587009981274605, + 0.02942119725048542, + 0.026018723845481873, + 0.02849750779569149, + 0.010853474028408527, + 0.0077985175885260105, + 0.01024243701249361, + 0.008808838203549385, + 0.0061080120503902435, + 0.009377988055348396, + 0.009687473066151142, + 0.011571415700018406, + 0.015173530206084251, + 0.015550883486866951, + 0.038473065942525864, + 0.04529472813010216, + 0.015761084854602814, + 0.01341883186250925, + 0.03384209796786308, + 0.0374869778752327, + 0.0077819195576012135, + 0.005489948205649853, + 0.014369509182870388, + 0.00937875360250473, + 0.007446514442563057, + 0.004951855633407831, + 0.016857922077178955, + 0.014703178778290749, + 0.0029272190295159817, + 0.002409690525382757, + 0.0035019025672227144, + 0.002993521746248007, + 0.0018751784227788448, + 0.002569373929873109, + 0.006581309251487255, + 0.005455330945551395, + 0.0034640729427337646, + 0.0025962574873119593, + 0.0035074178595095873, + 0.006232406944036484, + 0.0030157912988215685, + 0.003550491528585553, + 0.003670805599540472, + 0.004082829691469669, + 0.009501253254711628, + 0.012466498650610447, + 0.06821609288454056, + 0.012834778055548668, + 0.005849121138453484, + 0.006323087029159069, + 0.02157154493033886, + 0.06826096028089523, + 0.0028323614969849586, + 0.002528759650886059, + 0.00470482325181365, + 0.005026545375585556, + 0.0024201988708227873, + 0.004474632441997528, + 0.004281256813555956, + 0.005635153967887163, + 0.002897299826145172, + 0.0024356732610613108, + 0.009431934915482998, + 0.0037373241502791643, + 0.002766107441857457, + 0.0031720101833343506, + 0.0038586517330259085, + 0.008367184549570084, + 0.004756583366543055, + 0.005223453510552645, + 0.015303753316402435, + 0.014193310402333736, + 0.004890142008662224, + 0.006193671841174364, + 0.007514814846217632, + 0.005338409449905157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.02329447865486145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063323244452477 + ], + "opposite_logprob_tokens": [ + -4.141063213348389 + ], + "logprob_reward": -0.016063323244452477, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063323244452477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16787876188755035, + 0.07426665723323822, + 0.10175087302923203, + 0.061335597187280655, + 0.16616225242614746, + 0.07196971029043198, + 0.2414385825395584, + 0.03310507908463478, + 0.027705619111657143, + 0.05438687279820442 + ], + "credit_map": [ + 0.029577666893601418, + 0.02350614406168461, + 0.018284905701875687, + 0.020582884550094604, + 0.017407739534974098, + 0.0185096338391304, + 0.020429285243153572, + 0.01958048902451992, + 0.007420899346470833, + 0.006662949454039335, + 0.010275098495185375, + 0.009906060993671417, + 0.008553103543817997, + 0.007039705757051706, + 0.014253989793360233, + 0.010154856368899345, + 0.008979421108961105, + 0.013058981858193874, + 0.015615087933838367, + 0.014790366403758526, + 0.008105353452265263, + 0.010925591923296452, + 0.014503809623420238, + 0.0157722570002079, + 0.004310678690671921, + 0.006675370503216982, + 0.007725761272013187, + 0.009068828076124191, + 0.008128331042826176, + 0.006240797229111195, + 0.008806088007986546, + 0.010379742830991745, + 0.03264864161610603, + 0.010337728075683117, + 0.03149738535284996, + 0.019576817750930786, + 0.007776549085974693, + 0.0246581993997097, + 0.019169222563505173, + 0.020497716963291168, + 0.004882140550762415, + 0.005435364320874214, + 0.006325194146484137, + 0.010633638128638268, + 0.005493502598255873, + 0.0075316657312214375, + 0.019316179677844048, + 0.012352031655609608, + 0.011037953197956085, + 0.015269920229911804, + 0.07653939723968506, + 0.013166175223886967, + 0.009780565276741982, + 0.007615020032972097, + 0.029352394863963127, + 0.07867717742919922, + 0.002215214539319277, + 0.0030973097309470177, + 0.004743341356515884, + 0.00439224299043417, + 0.0028131238650530577, + 0.0030300808139145374, + 0.007969435304403305, + 0.004844333045184612, + 0.0024442397989332676, + 0.0025250392500311136, + 0.005107196979224682, + 0.0031347693875432014, + 0.0027641900815069675, + 0.0039957184344530106, + 0.003471534000709653, + 0.004262931179255247, + 0.006361815147101879, + 0.004708482418209314, + 0.012637630105018616, + 0.007572368253022432, + 0.005043322686105967, + 0.006595696322619915, + 0.006210977211594582, + 0.005256579723209143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.016063323244452477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient forest with tall pine trees covered in wet moss and rotten leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a burly, heavily bearded woodcutter wearing a red and black flannel plaid shirt visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter holding a felling axe with a smooth handle and a gleaming steel blade raised above his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter's body leaning back slightly in a tense pose, resembling a fully drawn bow before a strike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft shafts of morning light penetrating a dense canopy, creating a misty atmosphere with cold green and dark brown tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253083072602749 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253083072602749, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253083072602749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1937299519777298, + 0.15408429503440857, + 0.06691265106201172, + 0.03381168097257614, + 0.03315288573503494, + 0.022748848423361778, + 0.15177473425865173, + 0.04062226042151451, + 0.1627340465784073, + 0.14042861759662628 + ], + "credit_map": [ + 0.03753829374909401, + 0.024937355890870094, + 0.025948071852326393, + 0.016090651974081993, + 0.017677325755357742, + 0.02778095379471779, + 0.019059138372540474, + 0.0246981643140316, + 0.020959606394171715, + 0.011118214577436447, + 0.024502703920006752, + 0.020325258374214172, + 0.016393832862377167, + 0.015339442528784275, + 0.018166558817029, + 0.027278674766421318, + 0.006011627148836851, + 0.005939891561865807, + 0.009294040501117706, + 0.00933874398469925, + 0.006602624896913767, + 0.00820200890302658, + 0.01150527037680149, + 0.010018443688750267, + 0.0027453729417175055, + 0.003747885813936591, + 0.004666462074965239, + 0.004961516708135605, + 0.0028736842796206474, + 0.002641177736222744, + 0.005398212466388941, + 0.006777368485927582, + 0.004418148659169674, + 0.002522881142795086, + 0.005012891720980406, + 0.005139210261404514, + 0.0018588296370580792, + 0.0029053620528429747, + 0.006492737680673599, + 0.004802825395017862, + 0.0024153725244104862, + 0.002416017698124051, + 0.0027524176985025406, + 0.003556105075404048, + 0.0019053544383496046, + 0.0026417062617838383, + 0.003620817791670561, + 0.003441058099269867, + 0.007321828510612249, + 0.008712207898497581, + 0.050180207937955856, + 0.009329366497695446, + 0.0042478833347558975, + 0.005598057061433792, + 0.01501825638115406, + 0.051366932690143585, + 0.0033801309764385223, + 0.003778647631406784, + 0.008935636840760708, + 0.004520564340054989, + 0.0026427197735756636, + 0.004392917267978191, + 0.003802219871431589, + 0.009169422090053558, + 0.01728469505906105, + 0.01530464831739664, + 0.03715554252266884, + 0.016077829524874687, + 0.010474438779056072, + 0.01795637421309948, + 0.011251435615122318, + 0.037229087203741074, + 0.01221462432295084, + 0.009100325405597687, + 0.027686936780810356, + 0.03970770537853241, + 0.006130168214440346, + 0.011596528813242912, + 0.014271616004407406, + 0.01972072571516037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005253083072602749 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_4/index_tensor([9925, 9925])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_4/index_tensor([9925, 9925])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0002254e239be0133d9a4259da0bdafb37d395cc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_4/index_tensor([9925, 9925])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, ultra-wide shot reveals a breathtaking and ominous spectacle, showcasing the crumbling, sun-baked ruins of the ancient city of Gaochang, its eroded earthen walls, and the remains of skeletal buildings spread across the vast, arid desert floor. From the right side of the frame, a massive sandstorm roars toward the city, a towering, surging wall of ochre and brown dust that fills the sky. The entire scene is captured with a hyperreal, high-shutter-speed aesthetic, freezing each grain of sand in mid-air with incredible clarity. As the storm's leading edge makes contact, it washes over the ancient, cracked walls in a chaotic, violent swirl, with grains visibly rushing against the terracotta-colored surface in a frozen moment of destruction. The light is completely split: brilliant, harsh sunlight still illuminates parts of the ruins, casting deep, sharp shadows, while the encroaching storm casts a suffocating, yellowish haze over the rest, creating a powerful atmosphere of impending doom and the overwhelming power of nature.", + "seed": 21830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_4/index_9925_seed_21830_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3136056136805564, + "vlm_logprob": -0.3136056136805564, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.3136056136805564, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13232333958148956 + ], + "opposite_logprob_tokens": [ + -2.3823232650756836 + ], + "logprob_reward": -0.13232333958148956, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.13232333958148956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15961281955242157, + 0.11898944526910782, + 0.04890554025769234, + 0.17272718250751495, + 0.04874631389975548, + 0.05725039169192314, + 0.07407215982675552, + 0.06611374765634537, + 0.176680326461792, + 0.07690205425024033 + ], + "credit_map": [ + 0.01353231631219387, + 0.016578618437051773, + 0.02112315036356449, + 0.026713170111179352, + 0.016022764146327972, + 0.013858240097761154, + 0.031854815781116486, + 0.019929759204387665, + 0.007223791908472776, + 0.007536017801612616, + 0.017546243965625763, + 0.00784213189035654, + 0.04519171640276909, + 0.02189130336046219, + 0.005955010652542114, + 0.005803242325782776, + 0.004554307088255882, + 0.00578850694000721, + 0.00612232182174921, + 0.0061050234362483025, + 0.010422226041555405, + 0.005973257124423981, + 0.004314621444791555, + 0.005625282414257526, + 0.01968737505376339, + 0.03341593220829964, + 0.019394870847463608, + 0.015270359814167023, + 0.021617578342556953, + 0.02510152943432331, + 0.01848141849040985, + 0.01975814253091812, + 0.00640143034979701, + 0.005667389370501041, + 0.007067919243127108, + 0.005388680379837751, + 0.005057874601334333, + 0.004810249898582697, + 0.008741837926208973, + 0.005610935855656862, + 0.005907363258302212, + 0.006472096778452396, + 0.0050520580261945724, + 0.008211676962673664, + 0.008576398715376854, + 0.01056569442152977, + 0.005671125836670399, + 0.0067939856089651585, + 0.006871914956718683, + 0.011835620738565922, + 0.011278456076979637, + 0.009275520220398903, + 0.009362369775772095, + 0.009265475906431675, + 0.006710785441100597, + 0.00947202555835247, + 0.005738490726798773, + 0.006381291430443525, + 0.012500504031777382, + 0.009985227137804031, + 0.007977345027029514, + 0.008974920958280563, + 0.007051452063024044, + 0.0075045241974294186, + 0.016321692615747452, + 0.015695542097091675, + 0.023388871923089027, + 0.028473639860749245, + 0.02213730290532112, + 0.03145667538046837, + 0.017935030162334442, + 0.021271592006087303, + 0.005769756156951189, + 0.01813420280814171, + 0.015947476029396057, + 0.006554782390594482, + 0.006353077013045549, + 0.008053316734731197, + 0.009239712730050087, + 0.006849735975265503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.13232333958148956 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020618624985218048 + ], + "opposite_logprob_tokens": [ + -3.8956186771392822 + ], + "logprob_reward": -0.020618624985218048, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020618624985218048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09416373819112778, + 0.09485518932342529, + 0.27473974227905273, + 0.26157069206237793, + 0.08831961452960968, + 0.02076023630797863, + 0.03072936087846756, + 0.028017820790410042, + 0.08813194930553436, + 0.018711648881435394 + ], + "credit_map": [ + 0.012051866389811039, + 0.010243265889585018, + 0.010388373397290707, + 0.009585303254425526, + 0.022185057401657104, + 0.008377967402338982, + 0.00870381947606802, + 0.012628087773919106, + 0.0063968985341489315, + 0.006381104700267315, + 0.01228692289441824, + 0.004819212481379509, + 0.03664841130375862, + 0.017482206225395203, + 0.005972662940621376, + 0.004867766052484512, + 0.028200605884194374, + 0.03221500292420387, + 0.029327278956770897, + 0.0234437994658947, + 0.03369229659438133, + 0.040934816002845764, + 0.028263825923204422, + 0.058662138879299164, + 0.02783035859465599, + 0.04349915683269501, + 0.02993936836719513, + 0.022130845114588737, + 0.03432649374008179, + 0.04472377896308899, + 0.026920260861516, + 0.032200444489717484, + 0.009445201605558395, + 0.009222276508808136, + 0.011319904588162899, + 0.016395915299654007, + 0.00883558765053749, + 0.00926006305962801, + 0.015644919127225876, + 0.008195741102099419, + 0.0023005190305411816, + 0.0028502661734819412, + 0.002317772014066577, + 0.0037224760744720697, + 0.002566482173278928, + 0.0017839723732322454, + 0.003028843319043517, + 0.0021899044513702393, + 0.0025106961838901043, + 0.004024716094136238, + 0.005000611301511526, + 0.005113308317959309, + 0.002943733474239707, + 0.003823574399575591, + 0.003648800076916814, + 0.0036639219615608454, + 0.0027308627031743526, + 0.003491475945338607, + 0.005383545067161322, + 0.003484992543235421, + 0.0029536669608205557, + 0.002605017274618149, + 0.00254426384344697, + 0.0048239948228001595, + 0.008556703105568886, + 0.008282234892249107, + 0.01140637043863535, + 0.010382330045104027, + 0.00959054846316576, + 0.017017580568790436, + 0.00952992495149374, + 0.01336626149713993, + 0.0017160737188532948, + 0.00354633666574955, + 0.0036846413277089596, + 0.0015664923703297973, + 0.0018649355042725801, + 0.0019446128280833364, + 0.002534003695473075, + 0.001854553585872054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.020618624985218048 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001955979270860553 + ], + "opposite_logprob_tokens": [ + -6.251955986022949 + ], + "logprob_reward": -0.001955979270860553, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001955979270860553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1560937464237213, + 0.08528677374124527, + 0.07506616413593292, + 0.06330211460590363, + 0.08441173285245895, + 0.04109549894928932, + 0.07999370992183685, + 0.1596362590789795, + 0.12209966033697128, + 0.13301436603069305 + ], + "credit_map": [ + 0.015138804912567139, + 0.01635569892823696, + 0.01701444759964943, + 0.028920786455273628, + 0.02319948747754097, + 0.014906806871294975, + 0.013648252934217453, + 0.026909461244940758, + 0.004106409382075071, + 0.004235915839672089, + 0.005472682882100344, + 0.00540901580825448, + 0.02766076661646366, + 0.022695211693644524, + 0.006851566955447197, + 0.008855207823216915, + 0.006249029189348221, + 0.008162829093635082, + 0.007666848599910736, + 0.0074686515145003796, + 0.010635467246174812, + 0.010523078963160515, + 0.010758234187960625, + 0.013602024875581264, + 0.00826728343963623, + 0.009519033133983612, + 0.006780908443033695, + 0.007843881845474243, + 0.0068788849748671055, + 0.00829097256064415, + 0.0071150693111121655, + 0.008606080897152424, + 0.01198294572532177, + 0.013067377731204033, + 0.009810718707740307, + 0.009608171880245209, + 0.00968456082046032, + 0.012233668006956577, + 0.010371774435043335, + 0.007652512285858393, + 0.005462297238409519, + 0.005120405927300453, + 0.00472438195720315, + 0.0075265406630933285, + 0.004859507083892822, + 0.003710861084982753, + 0.004575587809085846, + 0.005115917883813381, + 0.0061016990803182125, + 0.012813040986657143, + 0.008944535627961159, + 0.0069148000329732895, + 0.015018995851278305, + 0.008830665610730648, + 0.009078455157577991, + 0.012291516177356243, + 0.015239875763654709, + 0.01969286985695362, + 0.020336799323558807, + 0.049408670514822006, + 0.012348499149084091, + 0.012366394512355328, + 0.012358213774859905, + 0.01788492128252983, + 0.011074335314333439, + 0.0118327047675848, + 0.015188856050372124, + 0.013607406057417393, + 0.010638497769832611, + 0.027299970388412476, + 0.016228890046477318, + 0.016228999942541122, + 0.00756460428237915, + 0.03742348402738571, + 0.04487781599164009, + 0.010349521413445473, + 0.012405886314809322, + 0.005621008574962616, + 0.00872747041285038, + 0.006044563837349415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.001955979270860553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063661098480225 + ], + "opposite_logprob_tokens": [ + -0.28136616945266724 + ], + "logprob_reward": -1.4063661098480225, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063661098480225, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13673135638237, + 0.06293225288391113, + 0.09343940764665604, + 0.12825508415699005, + 0.08473501354455948, + 0.10122133046388626, + 0.03364825248718262, + 0.11032377183437347, + 0.08411549031734467, + 0.1645980179309845 + ], + "credit_map": [ + 0.014113958925008774, + 0.012945609167218208, + 0.026761189103126526, + 0.01670842431485653, + 0.024430111050605774, + 0.012020505964756012, + 0.013328108005225658, + 0.016423439607024193, + 0.0042897495441138744, + 0.004187984857708216, + 0.01778537593781948, + 0.006097775883972645, + 0.013375626876950264, + 0.0059576318599283695, + 0.0047859931364655495, + 0.006452113855630159, + 0.006748044863343239, + 0.008630652911961079, + 0.00990514736622572, + 0.009956334717571735, + 0.012496116571128368, + 0.014039902947843075, + 0.00894357729703188, + 0.02271963469684124, + 0.012176107615232468, + 0.01635412871837616, + 0.01765705645084381, + 0.013798417523503304, + 0.011637595482170582, + 0.02296367660164833, + 0.01499777752906084, + 0.01867031492292881, + 0.008650882169604301, + 0.01210982259362936, + 0.009924905374646187, + 0.01107487641274929, + 0.00992302130907774, + 0.011915110982954502, + 0.012158425524830818, + 0.008977970108389854, + 0.00967501848936081, + 0.018345128744840622, + 0.009943797253072262, + 0.014098726212978363, + 0.012683110311627388, + 0.009314724244177341, + 0.016806233674287796, + 0.010354600846767426, + 0.0029270872473716736, + 0.0037248553708195686, + 0.006897705141454935, + 0.004214182496070862, + 0.0033036707900464535, + 0.0043001421727240086, + 0.004331623669713736, + 0.003948982339352369, + 0.012362532317638397, + 0.010417385958135128, + 0.015416059643030167, + 0.015555035322904587, + 0.015550816431641579, + 0.011740469373762608, + 0.012042257934808731, + 0.017239222303032875, + 0.008822987787425518, + 0.009474470280110836, + 0.010703575797379017, + 0.010115924291312695, + 0.008961385115981102, + 0.013931114226579666, + 0.009778155013918877, + 0.012327877804636955, + 0.014891949482262135, + 0.03315035626292229, + 0.02528419904410839, + 0.025214867666363716, + 0.01685149036347866, + 0.01435838732868433, + 0.02238105610013008, + 0.012465720064938068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -1.4063661098480225 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006764014717191458 + ], + "opposite_logprob_tokens": [ + -5.006763935089111 + ], + "logprob_reward": -0.006764014717191458, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006764014717191458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15065792202949524, + 0.06287821382284164, + 0.057658977806568146, + 0.15452353656291962, + 0.0741664245724678, + 0.0532979853451252, + 0.030683554708957672, + 0.1870659440755844, + 0.07218148559331894, + 0.15688596665859222 + ], + "credit_map": [ + 0.016801847144961357, + 0.016148487105965614, + 0.028651079162955284, + 0.020484624430537224, + 0.016535231843590736, + 0.010757378302514553, + 0.016053106635808945, + 0.02522616647183895, + 0.004246694501489401, + 0.00332379131577909, + 0.0036508096382021904, + 0.0036041084676980972, + 0.02332780696451664, + 0.015318651683628559, + 0.0050962213426828384, + 0.004310126416385174, + 0.006352083757519722, + 0.006321366410702467, + 0.008049391210079193, + 0.006106623448431492, + 0.006658274680376053, + 0.006915028672665358, + 0.007198326755315065, + 0.010057879611849785, + 0.021597137674689293, + 0.028355510905385017, + 0.018323736265301704, + 0.012931501492857933, + 0.020199619233608246, + 0.024880968034267426, + 0.012976144440472126, + 0.015258921310305595, + 0.006878968793898821, + 0.008185973390936852, + 0.008584242314100266, + 0.01087181456387043, + 0.007618463598191738, + 0.011222712695598602, + 0.011603237129747868, + 0.009201014414429665, + 0.006120252422988415, + 0.008484638296067715, + 0.005277832504361868, + 0.007376220542937517, + 0.005620956886559725, + 0.006268253084272146, + 0.007881751284003258, + 0.00626808125525713, + 0.00257299211807549, + 0.005573325790464878, + 0.0034391956869512796, + 0.0037612735759466887, + 0.003735512727871537, + 0.0038828968536108732, + 0.004146344494074583, + 0.0035720132291316986, + 0.016160903498530388, + 0.027898145839571953, + 0.024870922788977623, + 0.05789799615740776, + 0.013956185430288315, + 0.010031282901763916, + 0.015149998478591442, + 0.021100498735904694, + 0.005571553949266672, + 0.007739742752164602, + 0.007980846799910069, + 0.01075808797031641, + 0.005818686913698912, + 0.012478838674724102, + 0.010622543282806873, + 0.011211183853447437, + 0.01250346377491951, + 0.03866434842348099, + 0.033557865768671036, + 0.016520531848073006, + 0.014735805802047253, + 0.012732959352433681, + 0.017079155892133713, + 0.011091833002865314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.006764014717191458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, high-angle, ultra-wide shot reveals a breathtaking and ominous spectacle, showcasing the crumbling, sun-baked ruins of the ancient city of Gaochang, its eroded earthen walls, and the remains of skeletal buildings spread across the vast, arid desert floor. From the right side of the frame, a massive sandstorm roars toward the city, a towering, surging wall of ochre and brown dust that fills the sky. The entire scene is captured with a hyperreal, high-shutter-speed aesthetic, freezing each grain of sand in mid-air with incredible clarity. As the storm's leading edge makes contact, it washes over the ancient, cracked walls in a chaotic, violent swirl, with grains visibly rushing against the terracotta-colored surface in a frozen moment of destruction. The light is completely split: brilliant, harsh sunlight still illuminates parts of the ruins, casting deep, sharp shadows, while the encroaching storm casts a suffocating, yellowish haze over the rest, creating a powerful atmosphere of impending doom and the overwhelming power of nature.", + "seed": 21831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_4/index_9925_seed_21831_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5240323069971055, + "vlm_logprob": -0.5240323069971055, + "vlm_yes_no_margin": 2.324999988079071, + "vlm_token_credit": -0.5240323069971055, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1503479480743408 + ], + "opposite_logprob_tokens": [ + -0.40034788846969604 + ], + "logprob_reward": -1.1503479480743408, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.1503479480743408, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.26043304800987244, + 0.08276636153459549, + 0.17750589549541473, + 0.05546941980719566, + 0.11239496618509293, + 0.10230344533920288, + 0.03382681682705879, + 0.08393785357475281, + 0.04975493252277374, + 0.041607216000556946 + ], + "credit_map": [ + 0.01609215699136257, + 0.027600565925240517, + 0.027660928666591644, + 0.051179468631744385, + 0.07572776079177856, + 0.026501886546611786, + 0.01567651331424713, + 0.019993793219327927, + 0.007397494278848171, + 0.017919940873980522, + 0.008402382954955101, + 0.009105263277888298, + 0.013079247437417507, + 0.009925873018801212, + 0.007833212614059448, + 0.009102948009967804, + 0.01786753162741661, + 0.0230500977486372, + 0.018343009054660797, + 0.01877058856189251, + 0.026277851313352585, + 0.0351654477417469, + 0.017144227400422096, + 0.020887142047286034, + 0.002038766862824559, + 0.004282234236598015, + 0.0029773819260299206, + 0.004320127889513969, + 0.014890575781464577, + 0.019804388284683228, + 0.0026492776814848185, + 0.004506674129515886, + 0.008436357602477074, + 0.01597650535404682, + 0.015026981011033058, + 0.014889365993440151, + 0.01134535949677229, + 0.01799406111240387, + 0.010225865989923477, + 0.01850048080086708, + 0.007926378399133682, + 0.008054791949689388, + 0.02560306526720524, + 0.019077520817518234, + 0.0057461014948785305, + 0.008295099250972271, + 0.009027956984937191, + 0.018572544679045677, + 0.002488793572410941, + 0.0030875462107360363, + 0.004321014508605003, + 0.004769500810652971, + 0.003725634654983878, + 0.006172490771859884, + 0.00433120783418417, + 0.004930632188916206, + 0.009413905441761017, + 0.01201098132878542, + 0.00814798939973116, + 0.01003475021570921, + 0.013569324277341366, + 0.010115280747413635, + 0.010423339903354645, + 0.010222290642559528, + 0.003943602088838816, + 0.006361683364957571, + 0.007562247104942799, + 0.009884490631520748, + 0.006497288588434458, + 0.004217907320708036, + 0.006097833625972271, + 0.0051898835226893425, + 0.004553376231342554, + 0.008370528928935528, + 0.005507586058229208, + 0.005382447503507137, + 0.006269421894103289, + 0.00307348626665771, + 0.004358823876827955, + 0.00409154687076807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -1.1503479480743408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486886166036129 + ], + "opposite_logprob_tokens": [ + -3.0486886501312256 + ], + "logprob_reward": -0.0486886166036129, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486886166036129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31472623348236084, + 0.030901826918125153, + 0.17689910531044006, + 0.04659898579120636, + 0.12808431684970856, + 0.19355499744415283, + 0.06817846745252609, + 0.013844325207173824, + 0.017811842262744904, + 0.009399871341884136 + ], + "credit_map": [ + 0.03763624280691147, + 0.06880811601877213, + 0.027444658800959587, + 0.025199677795171738, + 0.07136733829975128, + 0.047901879996061325, + 0.01520871464163065, + 0.021159587427973747, + 0.0027800037059932947, + 0.005715991370379925, + 0.0035878431517630816, + 0.0035729887895286083, + 0.003538415301591158, + 0.005389421712607145, + 0.002354297786951065, + 0.003962865564972162, + 0.015812337398529053, + 0.038073863834142685, + 0.015114525333046913, + 0.013451659120619297, + 0.02975299395620823, + 0.0320436991751194, + 0.018016941845417023, + 0.014633077196776867, + 0.0029190927743911743, + 0.004612560849636793, + 0.0031334205996245146, + 0.005439419765025377, + 0.008548605255782604, + 0.0138214360922575, + 0.00319312303327024, + 0.004931326489895582, + 0.007790738716721535, + 0.028883252292871475, + 0.00811667088419199, + 0.02983425371348858, + 0.010739819146692753, + 0.021778972819447517, + 0.008219531737267971, + 0.012721089646220207, + 0.010492607951164246, + 0.048259034752845764, + 0.04860544949769974, + 0.024114638566970825, + 0.02050396054983139, + 0.012703062035143375, + 0.01111617498099804, + 0.017760083079338074, + 0.006428764667361975, + 0.006357140839099884, + 0.007818610407412052, + 0.008300402201712132, + 0.010079160332679749, + 0.012755626812577248, + 0.008505689911544323, + 0.007933068089187145, + 0.0014960492262616754, + 0.0024149958044290543, + 0.0015970089007169008, + 0.001593993161804974, + 0.0023952769115567207, + 0.0018155645811930299, + 0.001272870576940477, + 0.0012585660442709923, + 0.0016433617565780878, + 0.0022004174534231424, + 0.0030088128987699747, + 0.0043101077899336815, + 0.0016375218983739614, + 0.00119600014295429, + 0.0021591924596577883, + 0.0016564278630539775, + 0.0014319641049951315, + 0.0020040974486619234, + 0.0012730711605399847, + 0.0010582783725112677, + 0.0013982933014631271, + 0.0005468993913382292, + 0.0007664667209610343, + 0.0009208011324517429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0486886166036129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002209247788414359 + ], + "opposite_logprob_tokens": [ + -6.127209186553955 + ], + "logprob_reward": -0.002209247788414359, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002209247788414359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10081121325492859, + 0.039439719170331955, + 0.25176677107810974, + 0.08743056654930115, + 0.163119375705719, + 0.17619378864765167, + 0.032062023878097534, + 0.08017897605895996, + 0.03918367624282837, + 0.029813922941684723 + ], + "credit_map": [ + 0.010336500592529774, + 0.02082817442715168, + 0.008829604834318161, + 0.00855266023427248, + 0.02334294281899929, + 0.015453243628144264, + 0.005110171157866716, + 0.008357909508049488, + 0.0031562047079205513, + 0.007772611454129219, + 0.00425644451752305, + 0.004275162238627672, + 0.005100028123706579, + 0.0069111790508031845, + 0.0032800217159092426, + 0.004688060376793146, + 0.02711920067667961, + 0.0386209711432457, + 0.022494837641716003, + 0.020063165575265884, + 0.04286250099539757, + 0.05144514515995979, + 0.024870045483112335, + 0.02429090067744255, + 0.005644532851874828, + 0.010835452005267143, + 0.004512405022978783, + 0.005715509410947561, + 0.0256798192858696, + 0.020098574459552765, + 0.0066596330143511295, + 0.008284642361104488, + 0.009199150837957859, + 0.04370395094156265, + 0.014819307252764702, + 0.03701523691415787, + 0.013523070141673088, + 0.021470308303833008, + 0.008403219282627106, + 0.01498512364923954, + 0.009918000549077988, + 0.04057695344090462, + 0.049581874161958694, + 0.024570446461439133, + 0.017668480053544044, + 0.008928008377552032, + 0.009977232664823532, + 0.014972786419093609, + 0.004986296407878399, + 0.0031831308733671904, + 0.0032638604752719402, + 0.004128803964704275, + 0.004304490517824888, + 0.004415149334818125, + 0.003908819984644651, + 0.003871468361467123, + 0.009950241073966026, + 0.01428343914449215, + 0.008441707119345665, + 0.008961951360106468, + 0.014161949045956135, + 0.007355041336268187, + 0.00882806908339262, + 0.008196568116545677, + 0.003783821826800704, + 0.004818073008209467, + 0.00743373017758131, + 0.008262194693088531, + 0.004874458070844412, + 0.0020407955162227154, + 0.004439405631273985, + 0.003531194757670164, + 0.0037019255105406046, + 0.005565233528614044, + 0.004152999725192785, + 0.00371612049639225, + 0.004706010688096285, + 0.0019680391997098923, + 0.0032034346368163824, + 0.0028001577593386173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.002209247788414359 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.406348466873169 + ], + "opposite_logprob_tokens": [ + -0.28134843707084656 + ], + "logprob_reward": -1.406348466873169, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.406348466873169, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21386942267417908, + 0.05840039253234863, + 0.06717082858085632, + 0.07349774241447449, + 0.07333149760961533, + 0.23713557422161102, + 0.15830525755882263, + 0.07006322592496872, + 0.02953793667256832, + 0.018688151612877846 + ], + "credit_map": [ + 0.016092225909233093, + 0.02736647054553032, + 0.02765044942498207, + 0.04851677268743515, + 0.04040615260601044, + 0.016574041917920113, + 0.016429061070084572, + 0.02083425223827362, + 0.006374084856361151, + 0.01054922305047512, + 0.006494809407740831, + 0.006842928472906351, + 0.006397987715899944, + 0.009195472113788128, + 0.005048767663538456, + 0.00749711599200964, + 0.008072961121797562, + 0.011737536638975143, + 0.00687195360660553, + 0.005059557501226664, + 0.011966324411332607, + 0.011618784628808498, + 0.0056500122882425785, + 0.0061936997808516026, + 0.007567237131297588, + 0.012177755124866962, + 0.0050847455859184265, + 0.008679743856191635, + 0.011164197698235512, + 0.00919119082391262, + 0.006669027730822563, + 0.012963843531906605, + 0.0038924450054764748, + 0.026208261027932167, + 0.0038606273010373116, + 0.012390723451972008, + 0.00573854660615325, + 0.011863759718835354, + 0.002970443107187748, + 0.006406697444617748, + 0.012991269119083881, + 0.05657045915722847, + 0.04638386890292168, + 0.036612749099731445, + 0.02640136517584324, + 0.015079936012625694, + 0.014353185892105103, + 0.02874273806810379, + 0.016249561682343483, + 0.01591450907289982, + 0.023139748722314835, + 0.020269092172384262, + 0.020034845918416977, + 0.02245377190411091, + 0.0231105238199234, + 0.017133206129074097, + 0.005528409965336323, + 0.011198141612112522, + 0.008901708759367466, + 0.008865159004926682, + 0.008910741657018661, + 0.008766126818954945, + 0.008425293490290642, + 0.00946764275431633, + 0.003589621977880597, + 0.0035887686535716057, + 0.004672711715102196, + 0.0069809164851903915, + 0.003077434143051505, + 0.0019628386944532394, + 0.003171563381329179, + 0.0024940818548202515, + 0.001994933234527707, + 0.002825947478413582, + 0.002909040078520775, + 0.0037719239480793476, + 0.001842057448811829, + 0.0019435278372839093, + 0.0015932162059471011, + 0.0018075064290314913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -1.406348466873169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012567255645990372 + ], + "opposite_logprob_tokens": [ + -4.387567043304443 + ], + "logprob_reward": -0.012567255645990372, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012567255645990372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16746008396148682, + 0.06034763157367706, + 0.19074010848999023, + 0.0713423416018486, + 0.06764480471611023, + 0.18712158501148224, + 0.08095075190067291, + 0.10844231396913528, + 0.043869197368621826, + 0.022081173956394196 + ], + "credit_map": [ + 0.015927962958812714, + 0.029499610885977745, + 0.01637539453804493, + 0.014448028057813644, + 0.04770136997103691, + 0.02343497984111309, + 0.008624096401035786, + 0.011448651552200317, + 0.005616897717118263, + 0.014101647771894932, + 0.005763787776231766, + 0.006157878786325455, + 0.008329705335199833, + 0.00941614992916584, + 0.005103193689137697, + 0.005858373362571001, + 0.01710401475429535, + 0.02863285131752491, + 0.019508332014083862, + 0.017751419916749, + 0.029970381408929825, + 0.03847570717334747, + 0.0195627324283123, + 0.019734656438231468, + 0.0027130802627652884, + 0.007004235405474901, + 0.0031313623767346144, + 0.002426153514534235, + 0.022732526063919067, + 0.02513304352760315, + 0.0029396426398307085, + 0.005262293852865696, + 0.0033622481860220432, + 0.018025383353233337, + 0.0052762445993721485, + 0.014321571215987206, + 0.007054587360471487, + 0.007502846885472536, + 0.005763149820268154, + 0.006338770966976881, + 0.011590334586799145, + 0.0409206748008728, + 0.054887667298316956, + 0.027598654851317406, + 0.017039066180586815, + 0.00943691935390234, + 0.010475732386112213, + 0.015172538347542286, + 0.0118176294490695, + 0.008012652397155762, + 0.009434056468307972, + 0.010234171524643898, + 0.01096613984555006, + 0.011245837435126305, + 0.00949994008988142, + 0.009740324690937996, + 0.01354341208934784, + 0.024996252730488777, + 0.012674080207943916, + 0.012362215667963028, + 0.017591381445527077, + 0.008262364193797112, + 0.010432220064103603, + 0.008580389432609081, + 0.004917140584439039, + 0.0056995912455022335, + 0.007974764332175255, + 0.00763360969722271, + 0.0061029489152133465, + 0.0027387163136154413, + 0.005219805985689163, + 0.0035826207604259253, + 0.00287845847196877, + 0.004513895139098167, + 0.0025566047988831997, + 0.0023588293697685003, + 0.003588005667552352, + 0.0017527641030028462, + 0.0022004535421729088, + 0.002232165075838566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.012567255645990372 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, high-angle, ultra-wide shot reveals a breathtaking and ominous spectacle, showcasing the crumbling, sun-baked ruins of the ancient city of Gaochang, its eroded earthen walls, and the remains of skeletal buildings spread across the vast, arid desert floor. From the right side of the frame, a massive sandstorm roars toward the city, a towering, surging wall of ochre and brown dust that fills the sky. The entire scene is captured with a hyperreal, high-shutter-speed aesthetic, freezing each grain of sand in mid-air with incredible clarity. As the storm's leading edge makes contact, it washes over the ancient, cracked walls in a chaotic, violent swirl, with grains visibly rushing against the terracotta-colored surface in a frozen moment of destruction. The light is completely split: brilliant, harsh sunlight still illuminates parts of the ruins, casting deep, sharp shadows, while the encroaching storm casts a suffocating, yellowish haze over the rest, creating a powerful atmosphere of impending doom and the overwhelming power of nature.", + "seed": 21832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_4/index_9925_seed_21832_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.39328538784757255, + "vlm_logprob": -0.39328538784757255, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.39328538784757255, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21477361023426056 + ], + "opposite_logprob_tokens": [ + -1.8397736549377441 + ], + "logprob_reward": -0.21477361023426056, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21477361023426056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16587507724761963, + 0.10064137727022171, + 0.09700949490070343, + 0.07213551551103592, + 0.252772718667984, + 0.08770131319761276, + 0.10328204929828644, + 0.03137199208140373, + 0.04358447343111038, + 0.045625951141119 + ], + "credit_map": [ + 0.019983431324362755, + 0.04366505891084671, + 0.019768983125686646, + 0.012848823331296444, + 0.018384840339422226, + 0.02469446323812008, + 0.014051557518541813, + 0.012477926909923553, + 0.015054385177791119, + 0.01289690937846899, + 0.014423549175262451, + 0.008485697209835052, + 0.013851583935320377, + 0.015331742353737354, + 0.013379688374698162, + 0.007217816077172756, + 0.009005927480757236, + 0.01174844428896904, + 0.009730063378810883, + 0.012917671352624893, + 0.01214287057518959, + 0.02447090856730938, + 0.009175422601401806, + 0.0078181903809309, + 0.009086470119655132, + 0.006530787795782089, + 0.013113019056618214, + 0.011224398389458656, + 0.007557247765362263, + 0.008589636534452438, + 0.010414822027087212, + 0.005619133822619915, + 0.008716290816664696, + 0.047480564564466476, + 0.01161113940179348, + 0.013690398074686527, + 0.10275173932313919, + 0.05076722055673599, + 0.011005135253071785, + 0.006750242318958044, + 0.0076094274409115314, + 0.011326444335281849, + 0.011827046982944012, + 0.01772671565413475, + 0.01266573928296566, + 0.0071861641481518745, + 0.010774217545986176, + 0.008585555478930473, + 0.010195375420153141, + 0.01078434381633997, + 0.018532099202275276, + 0.02547215111553669, + 0.012463264167308807, + 0.006890955846756697, + 0.010857811197638512, + 0.008086048066616058, + 0.0044021084904670715, + 0.003596757771447301, + 0.00492181396111846, + 0.004363679327070713, + 0.0034725589212030172, + 0.002543875714763999, + 0.00429580919444561, + 0.0037753875367343426, + 0.0031730523332953453, + 0.005689934361726046, + 0.005321220029145479, + 0.005775531753897667, + 0.009624735452234745, + 0.004440585616976023, + 0.004619787912815809, + 0.004939620848745108, + 0.0032096432987600565, + 0.006092751398682594, + 0.008950155228376389, + 0.0063950372859835625, + 0.006969664245843887, + 0.0062356446869671345, + 0.0050315456464886665, + 0.002741509350016713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.21477361023426056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038150086998939514 + ], + "opposite_logprob_tokens": [ + -3.2881500720977783 + ], + "logprob_reward": -0.038150086998939514, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038150086998939514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7802363038063049, + 0.03468777611851692, + 0.017180660739541054, + 0.03489146754145622, + 0.03740217536687851, + 0.021720098331570625, + 0.012566294521093369, + 0.02042318880558014, + 0.023502638563513756, + 0.01738935150206089 + ], + "credit_map": [ + 0.13442489504814148, + 0.3030734956264496, + 0.04246445372700691, + 0.045179177075624466, + 0.08724554628133774, + 0.10210423916578293, + 0.03504816070199013, + 0.030696334317326546, + 0.003913091961294413, + 0.003075861604884267, + 0.004073028452694416, + 0.002638692269101739, + 0.005180148407816887, + 0.0077348011545836926, + 0.004647162742912769, + 0.003424988826736808, + 0.002818227047100663, + 0.002018408849835396, + 0.0020342227071523666, + 0.001971290912479162, + 0.001780891790986061, + 0.0040581244975328445, + 0.0011995763052254915, + 0.0012999203754588962, + 0.00277732009999454, + 0.002950191730633378, + 0.005765946581959724, + 0.0066812653094530106, + 0.003462871303781867, + 0.004968428052961826, + 0.0052998680621385574, + 0.0029855738393962383, + 0.0043518091551959515, + 0.007714266423135996, + 0.008471894077956676, + 0.0033026377204805613, + 0.004237757530063391, + 0.0036112400703132153, + 0.00270669674500823, + 0.00300587248057127, + 0.002115934155881405, + 0.002659877296537161, + 0.004797676112502813, + 0.002441416261717677, + 0.00262862304225564, + 0.0027964983601123095, + 0.0023248072247952223, + 0.0019552672747522593, + 0.0010697012767195702, + 0.0014239826705306768, + 0.002120977034792304, + 0.0019901504274457693, + 0.001631828024983406, + 0.0014065512223169208, + 0.0016373363323509693, + 0.001285766949877143, + 0.0022130950819700956, + 0.002135802758857608, + 0.0029913068283349276, + 0.004055033437907696, + 0.002983895828947425, + 0.0017638606950640678, + 0.0022738410625606775, + 0.002006353810429573, + 0.0019875115249305964, + 0.0031376397237181664, + 0.002994009992107749, + 0.003222789615392685, + 0.005058609880506992, + 0.002298420760780573, + 0.00233068922534585, + 0.0024729701690375805, + 0.0012048248900100589, + 0.0022002069745212793, + 0.003747259732335806, + 0.002073970390483737, + 0.0028810196090489626, + 0.0018486370099708438, + 0.002218266949057579, + 0.0012151667615398765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.038150086998939514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032133632339537144 + ], + "opposite_logprob_tokens": [ + -5.753213405609131 + ], + "logprob_reward": -0.0032133632339537144, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032133632339537144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09334470331668854, + 0.0585484653711319, + 0.04730026051402092, + 0.0951927900314331, + 0.4478459060192108, + 0.07828713953495026, + 0.04482341557741165, + 0.018098188564181328, + 0.06811878085136414, + 0.048440322279930115 + ], + "credit_map": [ + 0.009929770603775978, + 0.012137490324676037, + 0.020963476970791817, + 0.00868011824786663, + 0.010732337832450867, + 0.012633070349693298, + 0.011158199980854988, + 0.007110241334885359, + 0.006546116899698973, + 0.005435732658952475, + 0.009888325817883015, + 0.00423905486240983, + 0.0076774912886321545, + 0.011262409389019012, + 0.008348314091563225, + 0.005151024553924799, + 0.004473268985748291, + 0.004312166478484869, + 0.004998082295060158, + 0.005999386310577393, + 0.009668580256402493, + 0.007571721915155649, + 0.006624308414757252, + 0.0036527470219880342, + 0.009065593592822552, + 0.008828705176711082, + 0.014896281063556671, + 0.013717742636799812, + 0.009462342597544193, + 0.01859418675303459, + 0.014585197903215885, + 0.006042744033038616, + 0.014712056145071983, + 0.07932105660438538, + 0.03734647482633591, + 0.0189365204423666, + 0.1794472336769104, + 0.08794108033180237, + 0.01723397895693779, + 0.012907516211271286, + 0.006666063796728849, + 0.010236567817628384, + 0.011524745263159275, + 0.013783629052340984, + 0.011872006580233574, + 0.007898146286606789, + 0.009011199697852135, + 0.007294795010238886, + 0.0052520050667226315, + 0.005248688161373138, + 0.006396626587957144, + 0.006136859767138958, + 0.00834739487618208, + 0.0044260406866669655, + 0.0052746860310435295, + 0.0037411220837384462, + 0.0021739748772233725, + 0.0020312946289777756, + 0.0029568932950496674, + 0.0022945315577089787, + 0.0029272648971527815, + 0.0019054211443290114, + 0.00214019906707108, + 0.0016686104936525226, + 0.005487850867211819, + 0.009646021761000156, + 0.006663353648036718, + 0.007393790874630213, + 0.017737315967679024, + 0.006678273901343346, + 0.007045063655823469, + 0.007467109709978104, + 0.0032148046884685755, + 0.00583623768761754, + 0.009587928652763367, + 0.005960523150861263, + 0.007344846613705158, + 0.00725146709010005, + 0.005291061941534281, + 0.003953458741307259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0032133632339537144 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7016137838363647 + ], + "opposite_logprob_tokens": [ + -0.20161379873752594 + ], + "logprob_reward": -1.7016137838363647, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7016137838363647, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18101781606674194, + 0.054198917001485825, + 0.07611266523599625, + 0.035999659448862076, + 0.41983288526535034, + 0.048936739563941956, + 0.08308020979166031, + 0.039661891758441925, + 0.023412639275193214, + 0.037746500223875046 + ], + "credit_map": [ + 0.029240807518363, + 0.060359105467796326, + 0.01755872555077076, + 0.011667465791106224, + 0.02122884802520275, + 0.01638379879295826, + 0.014630753546953201, + 0.00994834303855896, + 0.009322477504611015, + 0.008348595350980759, + 0.004476658068597317, + 0.004422609228640795, + 0.007394063752144575, + 0.010121415369212627, + 0.005102731287479401, + 0.005010376684367657, + 0.007368434686213732, + 0.0070616318844258785, + 0.007943155243992805, + 0.00842514168471098, + 0.013828221708536148, + 0.01678517460823059, + 0.009333288297057152, + 0.005367625504732132, + 0.002779799047857523, + 0.003946034237742424, + 0.0034298752434551716, + 0.003794236807152629, + 0.004959631711244583, + 0.0054334597662091255, + 0.007787923328578472, + 0.0038687039632350206, + 0.01773333176970482, + 0.043703705072402954, + 0.041429080069065094, + 0.02249021641910076, + 0.15909597277641296, + 0.10169629752635956, + 0.01933036372065544, + 0.01435400266200304, + 0.0027808239683508873, + 0.00974653847515583, + 0.0065495287999510765, + 0.005766076501458883, + 0.012849966995418072, + 0.003761846572160721, + 0.004309691488742828, + 0.003172277007251978, + 0.008999970741569996, + 0.009881141595542431, + 0.009965907782316208, + 0.0156859140843153, + 0.0146497692912817, + 0.008794345892965794, + 0.008748326450586319, + 0.006354846991598606, + 0.004961258266121149, + 0.0036126100458204746, + 0.007179911248385906, + 0.007647200953215361, + 0.00481728371232748, + 0.0029897112399339676, + 0.004989891778677702, + 0.0034640340600162745, + 0.0023545371368527412, + 0.0030697560869157314, + 0.0030998336151242256, + 0.0023609520867466927, + 0.005278983153402805, + 0.002241774694994092, + 0.0024942171294242144, + 0.0025125909596681595, + 0.0030135533306747675, + 0.005035300739109516, + 0.006540708709508181, + 0.0059730918146669865, + 0.006397284101694822, + 0.0034249837044626474, + 0.004555072635412216, + 0.0028065091464668512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -1.7016137838363647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008676094934344292 + ], + "opposite_logprob_tokens": [ + -4.758676052093506 + ], + "logprob_reward": -0.008676094934344292, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008676094934344292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12407287955284119, + 0.05470139533281326, + 0.058942969888448715, + 0.05261583626270294, + 0.4670470654964447, + 0.06714430451393127, + 0.06263570487499237, + 0.02561231516301632, + 0.05233965814113617, + 0.03488790616393089 + ], + "credit_map": [ + 0.016962412744760513, + 0.02555454708635807, + 0.01459827832877636, + 0.009427422657608986, + 0.014533511362969875, + 0.018027616664767265, + 0.013951030559837818, + 0.011018048971891403, + 0.005793171469122171, + 0.006258486770093441, + 0.005871741566807032, + 0.003947250545024872, + 0.008445344865322113, + 0.013832243159413338, + 0.006335229147225618, + 0.004217926412820816, + 0.006145995110273361, + 0.005095807369798422, + 0.006017796229571104, + 0.006504571530967951, + 0.009709478355944157, + 0.014605330303311348, + 0.0065168640576303005, + 0.004347122274339199, + 0.005105446558445692, + 0.00498009379953146, + 0.008398366160690784, + 0.008683999069035053, + 0.0060103898867964745, + 0.0063460469245910645, + 0.008154617622494698, + 0.004936873447149992, + 0.01670786924660206, + 0.0646890252828598, + 0.0424220971763134, + 0.022730745375156403, + 0.18327239155769348, + 0.1018810123205185, + 0.019848739728331566, + 0.01549513079226017, + 0.004725406877696514, + 0.011064363643527031, + 0.011147305369377136, + 0.010743886232376099, + 0.013232069090008736, + 0.0044153006747365, + 0.0077467807568609715, + 0.004069187678396702, + 0.010259760543704033, + 0.007574848830699921, + 0.008668073453009129, + 0.009545138105750084, + 0.010144258849322796, + 0.0049233511090278625, + 0.00687258318066597, + 0.004647684749215841, + 0.002733896253630519, + 0.0027134784031659365, + 0.004155794158577919, + 0.0034605388063937426, + 0.0030644878279417753, + 0.0026351292617619038, + 0.0034876407589763403, + 0.0033613471314311028, + 0.00418686680495739, + 0.007275613956153393, + 0.0049746278673410416, + 0.005484410561621189, + 0.01357684563845396, + 0.005145053844898939, + 0.005742608569562435, + 0.00595362763851881, + 0.003013537498190999, + 0.005286769010126591, + 0.0050023761577904224, + 0.004510244354605675, + 0.007422954775393009, + 0.003632006933912635, + 0.0033042083960026503, + 0.0027158032171428204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.008676094934344292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, high-angle, ultra-wide shot reveals a breathtaking and ominous spectacle, showcasing the crumbling, sun-baked ruins of the ancient city of Gaochang, its eroded earthen walls, and the remains of skeletal buildings spread across the vast, arid desert floor. From the right side of the frame, a massive sandstorm roars toward the city, a towering, surging wall of ochre and brown dust that fills the sky. The entire scene is captured with a hyperreal, high-shutter-speed aesthetic, freezing each grain of sand in mid-air with incredible clarity. As the storm's leading edge makes contact, it washes over the ancient, cracked walls in a chaotic, violent swirl, with grains visibly rushing against the terracotta-colored surface in a frozen moment of destruction. The light is completely split: brilliant, harsh sunlight still illuminates parts of the ruins, casting deep, sharp shadows, while the encroaching storm casts a suffocating, yellowish haze over the rest, creating a powerful atmosphere of impending doom and the overwhelming power of nature.", + "seed": 21833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_4/index_9925_seed_21833_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3918749089119956, + "vlm_logprob": -0.3918749089119956, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.3918749089119956, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23195528984069824 + ], + "opposite_logprob_tokens": [ + -1.7319552898406982 + ], + "logprob_reward": -0.23195528984069824, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.23195528984069824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2076462060213089, + 0.15374669432640076, + 0.07155971229076385, + 0.03961572051048279, + 0.04592696949839592, + 0.041249457746744156, + 0.161298006772995, + 0.041774701327085495, + 0.09042686969041824, + 0.1467556655406952 + ], + "credit_map": [ + 0.017338475212454796, + 0.024676421657204628, + 0.01661999151110649, + 0.05432373657822609, + 0.01887773536145687, + 0.02563633769750595, + 0.02564394474029541, + 0.024529561400413513, + 0.013457255437970161, + 0.01361596118658781, + 0.0160776786506176, + 0.03367944434285164, + 0.01831270195543766, + 0.022438205778598785, + 0.020306754857301712, + 0.015858696773648262, + 0.004163659177720547, + 0.005444933660328388, + 0.005814183037728071, + 0.011420613154768944, + 0.01289403811097145, + 0.010010741651058197, + 0.00697910925373435, + 0.014832433313131332, + 0.0025345413014292717, + 0.0028015021234750748, + 0.0037520413752645254, + 0.004645125940442085, + 0.01031809113919735, + 0.003906994592398405, + 0.00413730600848794, + 0.007520114071667194, + 0.0038217073306441307, + 0.00443925941362977, + 0.004036182537674904, + 0.007625287398695946, + 0.008191794157028198, + 0.007972967810928822, + 0.0043257661163806915, + 0.005514005199074745, + 0.003330472158268094, + 0.004363550804555416, + 0.004113530274480581, + 0.0072313472628593445, + 0.007385403849184513, + 0.005228049121797085, + 0.005957567133009434, + 0.003639538073912263, + 0.0361744649708271, + 0.0646342784166336, + 0.006841376889497042, + 0.009783177636563778, + 0.015031097456812859, + 0.01795193739235401, + 0.006707450374960899, + 0.004174211993813515, + 0.004582112655043602, + 0.007117822766304016, + 0.004447423852980137, + 0.007850815542042255, + 0.005747226998209953, + 0.0043107070960104465, + 0.004220671486109495, + 0.0034979230258613825, + 0.006725942716002464, + 0.01228808332234621, + 0.012899807654321194, + 0.0173215139657259, + 0.008855619467794895, + 0.006901659537106752, + 0.013284201733767986, + 0.012150043621659279, + 0.01941685937345028, + 0.015543626621365547, + 0.010586751624941826, + 0.01118855644017458, + 0.04158106446266174, + 0.02279537171125412, + 0.014637383632361889, + 0.011006064713001251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.23195528984069824 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014256061054766178 + ], + "opposite_logprob_tokens": [ + -4.264256000518799 + ], + "logprob_reward": -0.014256061054766178, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014256061054766178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16237927973270416, + 0.14560578763484955, + 0.09306864440441132, + 0.2142900973558426, + 0.04685705155134201, + 0.04726043716073036, + 0.0752074345946312, + 0.060889020562171936, + 0.052472684532403946, + 0.10196951031684875 + ], + "credit_map": [ + 0.016527477651834488, + 0.019284401088953018, + 0.009584138169884682, + 0.014663168229162693, + 0.024795064702630043, + 0.0425461083650589, + 0.01762513257563114, + 0.017353802919387817, + 0.013639938086271286, + 0.01553571131080389, + 0.009868892841041088, + 0.012823778204619884, + 0.027928514406085014, + 0.039233505725860596, + 0.01433342881500721, + 0.012242013588547707, + 0.009352104738354683, + 0.01127670705318451, + 0.007536065299063921, + 0.013044675812125206, + 0.011255050078034401, + 0.018906405195593834, + 0.010717646218836308, + 0.010979989543557167, + 0.012743341736495495, + 0.025344112887978554, + 0.00906726811081171, + 0.014512515626847744, + 0.04610646516084671, + 0.04627358913421631, + 0.015814460813999176, + 0.04442833736538887, + 0.0043168202973902225, + 0.0049113682471215725, + 0.004071391653269529, + 0.008695631287992, + 0.0070939138531684875, + 0.006863004993647337, + 0.005369354039430618, + 0.005535570438951254, + 0.003813191084191203, + 0.004943439736962318, + 0.0031717729289084673, + 0.0066760266199707985, + 0.007900072261691093, + 0.012660625390708447, + 0.003639473579823971, + 0.0044558364897966385, + 0.014943589456379414, + 0.02525661140680313, + 0.0036851016338914633, + 0.005080457776784897, + 0.008359095081686974, + 0.00963533390313387, + 0.003913844469934702, + 0.004333399701863527, + 0.00574052520096302, + 0.012299635447561741, + 0.004621293395757675, + 0.0065683056600391865, + 0.006704315543174744, + 0.009642357006669044, + 0.007934609428048134, + 0.007377974223345518, + 0.005509506911039352, + 0.007039179559797049, + 0.0064913067035377026, + 0.007030060514807701, + 0.0070668114349246025, + 0.004333542659878731, + 0.00825960747897625, + 0.006742667406797409, + 0.010045897215604782, + 0.01654215157032013, + 0.0131219532340765, + 0.010602926835417747, + 0.014199226163327694, + 0.008599283173680305, + 0.014035362750291824, + 0.0148227009922266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.014256061054766178 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017326356610283256 + ], + "opposite_logprob_tokens": [ + -6.37673282623291 + ], + "logprob_reward": -0.0017326356610283256, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017326356610283256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10424505919218063, + 0.11534512042999268, + 0.14172683656215668, + 0.15477490425109863, + 0.05301598832011223, + 0.04331051930785179, + 0.09099308401346207, + 0.05311603099107742, + 0.06887631863355637, + 0.1745961606502533 + ], + "credit_map": [ + 0.011143767274916172, + 0.01274220459163189, + 0.009429688565433025, + 0.011953047476708889, + 0.016307123005390167, + 0.017483791336417198, + 0.012679021805524826, + 0.012506420724093914, + 0.012035658583045006, + 0.011628365144133568, + 0.005895169451832771, + 0.008089919574558735, + 0.01792784407734871, + 0.03904099389910698, + 0.010126136243343353, + 0.010601037181913853, + 0.007854068651795387, + 0.009005171246826649, + 0.004784141667187214, + 0.005400361493229866, + 0.047060295939445496, + 0.01865820586681366, + 0.007577212527394295, + 0.04138737544417381, + 0.006645485293120146, + 0.012555285356938839, + 0.0032145879231393337, + 0.006165725644677877, + 0.05632153898477554, + 0.01732543483376503, + 0.00811214279383421, + 0.04443470016121864, + 0.004140146542340517, + 0.004926357418298721, + 0.004949222784489393, + 0.009374009445309639, + 0.007724177557975054, + 0.01010633073747158, + 0.006028387229889631, + 0.005767359398305416, + 0.0036039354745298624, + 0.004820279311388731, + 0.00328820594586432, + 0.00507571455091238, + 0.007254639640450478, + 0.010978186503052711, + 0.0046487608924508095, + 0.003640797222033143, + 0.015836574137210846, + 0.030177783221006393, + 0.005055652465671301, + 0.007244096603244543, + 0.009177975356578827, + 0.01253634225577116, + 0.006497323513031006, + 0.004467328079044819, + 0.004426731262356043, + 0.012682783417403698, + 0.004725005477666855, + 0.006481522228568792, + 0.006278900429606438, + 0.008714119903743267, + 0.004249304533004761, + 0.005557663273066282, + 0.0053284806199371815, + 0.009511246345937252, + 0.009414184838533401, + 0.013784950599074364, + 0.007928822189569473, + 0.004189677536487579, + 0.009381521493196487, + 0.009337438270449638, + 0.014305343851447105, + 0.020789412781596184, + 0.011277036741375923, + 0.011889660730957985, + 0.05800946429371834, + 0.034185655415058136, + 0.013160040602087975, + 0.010979550890624523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.0017326356610283256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7016088962554932 + ], + "opposite_logprob_tokens": [ + -0.20160892605781555 + ], + "logprob_reward": -1.7016088962554932, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7016088962554932, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08745893836021423, + 0.03158797696232796, + 0.0493122860789299, + 0.027144886553287506, + 0.03959893062710762, + 0.027191756293177605, + 0.08073590695858002, + 0.07960274070501328, + 0.05039611831307411, + 0.5269704461097717 + ], + "credit_map": [ + 0.0066724056378006935, + 0.00934534426778555, + 0.007329858373850584, + 0.01407622080296278, + 0.014818902127444744, + 0.013969560153782368, + 0.010893095284700394, + 0.010353551246225834, + 0.0021475160028785467, + 0.002732996828854084, + 0.0031092504505068064, + 0.0052602943032979965, + 0.003934024833142757, + 0.006926564034074545, + 0.00432136096060276, + 0.0031559704802930355, + 0.0026285452768206596, + 0.0026764310896396637, + 0.002045827219262719, + 0.0039367275312542915, + 0.015170587226748466, + 0.006369009148329496, + 0.0030131321400403976, + 0.013472024351358414, + 0.00197285576723516, + 0.003584821941331029, + 0.0016709804767742753, + 0.0034209832083433867, + 0.0033107453491538763, + 0.00806064810603857, + 0.002345855813473463, + 0.0027779971715062857, + 0.0029522443655878305, + 0.003930582664906979, + 0.0028592399321496487, + 0.0042664650827646255, + 0.0071920789778232574, + 0.011343504302203655, + 0.0030799349769949913, + 0.00397487822920084, + 0.002919065533205867, + 0.002842794870957732, + 0.0025855021085590124, + 0.0038522102404385805, + 0.004481926094740629, + 0.004246025811880827, + 0.003944514784961939, + 0.0023197177797555923, + 0.011743624694645405, + 0.01675410009920597, + 0.0073381830006837845, + 0.01550457626581192, + 0.006488850340247154, + 0.008509174920618534, + 0.009063647128641605, + 0.00533374585211277, + 0.009179217740893364, + 0.01065967045724392, + 0.009157772175967693, + 0.009671593084931374, + 0.011986155062913895, + 0.011046884581446648, + 0.008198464289307594, + 0.009702984243631363, + 0.004713967442512512, + 0.0077600725926458836, + 0.008059336803853512, + 0.005922055337578058, + 0.007907706312835217, + 0.0036676099989563227, + 0.007005852181464434, + 0.005359516944736242, + 0.04481041803956032, + 0.05449143424630165, + 0.017824584618210793, + 0.02939879149198532, + 0.19952036440372467, + 0.11950252950191498, + 0.03010299988090992, + 0.03131934627890587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -1.7016088962554932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the ruins of the ancient city of Gaochang with eroded earthen walls and skeletal buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive sandstorm visible approaching from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sandstorm depicted as a towering wall of ochre and brown dust filling the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style hyperreal with high-shutter-speed clarity freezing individual grains of sand in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting split between brilliant harsh sunlight on some ruins and a yellowish haze from the storm on others?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009821661747992039 + ], + "opposite_logprob_tokens": [ + -4.634821891784668 + ], + "logprob_reward": -0.009821661747992039, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009821661747992039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11979296803474426, + 0.030192136764526367, + 0.029702940955758095, + 0.05840378999710083, + 0.03954692929983139, + 0.0215089563280344, + 0.11896955966949463, + 0.08662215620279312, + 0.0595836415886879, + 0.4356769919395447 + ], + "credit_map": [ + 0.01372146513313055, + 0.011660137213766575, + 0.011010154150426388, + 0.017833871766924858, + 0.016107380390167236, + 0.020089317113161087, + 0.015642108395695686, + 0.013728526420891285, + 0.0026149977929890156, + 0.0026179449632763863, + 0.0028456714935600758, + 0.004144332837313414, + 0.004907600115984678, + 0.006941954605281353, + 0.00351741723716259, + 0.002602218184620142, + 0.0022606824059039354, + 0.002469235798344016, + 0.0014261420583352447, + 0.0026159961707890034, + 0.007931855507194996, + 0.003472259733825922, + 0.002461312571540475, + 0.007065456826239824, + 0.00283294846303761, + 0.003255540505051613, + 0.0018195456359535456, + 0.0025759325362741947, + 0.020736046135425568, + 0.006346389651298523, + 0.0027104546315968037, + 0.01812693662941456, + 0.002946152351796627, + 0.003737884806469083, + 0.002954592928290367, + 0.0058939796872437, + 0.006648967042565346, + 0.0098933856934309, + 0.003342155134305358, + 0.004129811190068722, + 0.00267027597874403, + 0.0024840286932885647, + 0.0021962327882647514, + 0.0031340690329670906, + 0.0030357209034264088, + 0.0024381550028920174, + 0.003576909191906452, + 0.001973564736545086, + 0.024438321590423584, + 0.03861158341169357, + 0.007329230196774006, + 0.013168021105229855, + 0.009986910969018936, + 0.013108087703585625, + 0.007145886309444904, + 0.005181517917662859, + 0.011537827551364899, + 0.011825544759631157, + 0.008963974192738533, + 0.009005366824567318, + 0.011899500153958797, + 0.01291729137301445, + 0.009905757382512093, + 0.010566892102360725, + 0.004804406780749559, + 0.008581775240600109, + 0.005236827302724123, + 0.010935928672552109, + 0.009200076572597027, + 0.006029045674949884, + 0.007483263965696096, + 0.007312320172786713, + 0.03918398544192314, + 0.04463288187980652, + 0.013170489110052586, + 0.021826546639204025, + 0.16835902631282806, + 0.09952229261398315, + 0.023634614422917366, + 0.02534712851047516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.009821661747992039 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_5/index_tensor([1339, 1339])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_5/index_tensor([1339, 1339])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fe67e6cafac256f855e8e9c119c7947d32357ba1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_5/index_tensor([1339, 1339])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, medium shot depicts the dynamic action with clinical precision. A single, spherical jawbreaker, its surface covered in a high-gloss shell, reveals vibrant, swirling layers of electric blue, canary yellow, and fuchsia, bouncing across the smooth, reflective surface of a polished, dark gray concrete floor. Set within a minimalist studio, precisely calibrated lighting from a large, soft overhead source provides perfect, even illumination, creating perfectly crisp highlights on the candy's curvaceous form. The action begins with the jawbreaker's initial impact, its rigid, resilient structure causing it to spring back into the air, spinning gracefully. It then executes a rhythmic series of successive, smaller bounces, losing perceptible energy with each diminishing arc, until its vertical motion seamlessly transitions into a slow, silent roll. As it glides, the mirror-like floor perfectly reflects its swirling colors, creating a liquid play of light. The jaw-dropping object finally sways for a moment before coming to a complete still, its journey complete. The work’s rigorously minimalist approach and cool, neutral color palette isolates the candy’s kinetic behavior, creating an atmosphere of captivation and focus that elevates a simple physical event into a visual poem.", + "seed": 8595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_5/index_1339_seed_8595_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022670373506844043, + "vlm_logprob": -0.022670373506844043, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.022670373506844043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0147604551166296 + ], + "opposite_logprob_tokens": [ + -4.389760494232178 + ], + "logprob_reward": -0.0147604551166296, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0147604551166296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22900599241256714, + 0.16298726201057434, + 0.06264030188322067, + 0.11597287654876709, + 0.1114598736166954, + 0.06869149953126907, + 0.044544655829668045, + 0.06362992525100708, + 0.09082521498203278, + 0.050242383033037186 + ], + "credit_map": [ + 0.017898766323924065, + 0.022726383060216904, + 0.0279550701379776, + 0.03535596653819084, + 0.03897740691900253, + 0.020325804129242897, + 0.03698800876736641, + 0.028778577223420143, + 0.012952198274433613, + 0.03290939703583717, + 0.012264612130820751, + 0.018815940245985985, + 0.034463074058294296, + 0.022038739174604416, + 0.01270056702196598, + 0.016842735931277275, + 0.006998169235885143, + 0.009818106889724731, + 0.008189952000975609, + 0.008892925456166267, + 0.008334019221365452, + 0.0053591057658195496, + 0.00812650565057993, + 0.006921519059687853, + 0.008617098443210125, + 0.012026399374008179, + 0.015026011504232883, + 0.026495499536395073, + 0.014791848137974739, + 0.009379087947309017, + 0.01995673030614853, + 0.009680194780230522, + 0.01012418419122696, + 0.010212006978690624, + 0.011100037023425102, + 0.015265806578099728, + 0.011898037046194077, + 0.01120541151612997, + 0.02605557255446911, + 0.015598822385072708, + 0.006146818865090609, + 0.010847451165318489, + 0.008733936585485935, + 0.010839946568012238, + 0.008045099675655365, + 0.006901761516928673, + 0.00851019099354744, + 0.008666290901601315, + 0.004477090202271938, + 0.009365086443722248, + 0.004656950943171978, + 0.005286580882966518, + 0.0053979139775037766, + 0.004204108379781246, + 0.005375879351049662, + 0.005781044717878103, + 0.005615727510303259, + 0.00731588015332818, + 0.009400468319654465, + 0.011815842241048813, + 0.0072386376559734344, + 0.005780842155218124, + 0.00933846179395914, + 0.00712406262755394, + 0.010593865066766739, + 0.018192311748862267, + 0.008830942213535309, + 0.012745567597448826, + 0.009811103343963623, + 0.009226194582879543, + 0.012147118337452412, + 0.00927810464054346, + 0.005143805872648954, + 0.005586175248026848, + 0.007325001526623964, + 0.00846657995134592, + 0.005753560923039913, + 0.004688989836722612, + 0.006256082095205784, + 0.00702218571677804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0147604551166296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559839524328709 + ], + "opposite_logprob_tokens": [ + -4.38755989074707 + ], + "logprob_reward": -0.012559839524328709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559839524328709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1394546627998352, + 0.09612074494361877, + 0.04149903357028961, + 0.10479777306318283, + 0.4090096056461334, + 0.029371093958616257, + 0.033016081899404526, + 0.10371676832437515, + 0.02232578955590725, + 0.020688507705926895 + ], + "credit_map": [ + 0.010489401407539845, + 0.012820950709283352, + 0.016307035461068153, + 0.020250467583537102, + 0.026792045682668686, + 0.013600618578493595, + 0.02247142419219017, + 0.016722718253731728, + 0.008272222243249416, + 0.014577471651136875, + 0.01049117837101221, + 0.007814928889274597, + 0.025932420045137405, + 0.01056689489632845, + 0.008006196469068527, + 0.010459437035024166, + 0.003968334756791592, + 0.0038225201424211264, + 0.00541042722761631, + 0.006460193544626236, + 0.005466369912028313, + 0.0038233469240367413, + 0.008407707326114178, + 0.0041401321068406105, + 0.006797571666538715, + 0.009880281053483486, + 0.011080457828938961, + 0.01427204068750143, + 0.019411593675613403, + 0.010618535801768303, + 0.024836495518684387, + 0.00790080614387989, + 0.041836757212877274, + 0.043053824454545975, + 0.04152018949389458, + 0.07393687963485718, + 0.03730341047048569, + 0.03329179063439369, + 0.0956944078207016, + 0.042372338473796844, + 0.0032414747402071953, + 0.005051042884588242, + 0.0036028081085532904, + 0.0032598478719592094, + 0.0029926470015197992, + 0.0028936886228621006, + 0.0046464018523693085, + 0.00368318147957325, + 0.0030322305392473936, + 0.0033196385484188795, + 0.0048738992772996426, + 0.0038823175709694624, + 0.004746855702251196, + 0.0029885154217481613, + 0.004264291375875473, + 0.005908332299441099, + 0.01059463620185852, + 0.018873445689678192, + 0.009844889864325523, + 0.01666560396552086, + 0.010078475810587406, + 0.009488803334534168, + 0.018142059445381165, + 0.01002885214984417, + 0.0016333136009052396, + 0.003426554147154093, + 0.002464069053530693, + 0.003278308315202594, + 0.0034068988170474768, + 0.0021511707454919815, + 0.00284699909389019, + 0.0031184775289148092, + 0.002000179374590516, + 0.0024087559431791306, + 0.002093253657221794, + 0.004856923595070839, + 0.0023628328926861286, + 0.0020548696629703045, + 0.0026180625427514315, + 0.0022936302702873945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.012559839524328709 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784241206943989 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784241206943989, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784241206943989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18141205608844757, + 0.1070307195186615, + 0.06100235506892204, + 0.06439419835805893, + 0.33891719579696655, + 0.08882830291986465, + 0.05352745205163956, + 0.042744942009449005, + 0.03501202166080475, + 0.027130722999572754 + ], + "credit_map": [ + 0.014122230932116508, + 0.019070278853178024, + 0.026203615590929985, + 0.02917063608765602, + 0.030723391100764275, + 0.017460908740758896, + 0.02293672412633896, + 0.021724257618188858, + 0.013920314610004425, + 0.017451662570238113, + 0.017822137102484703, + 0.01076650433242321, + 0.009658722206950188, + 0.00951976515352726, + 0.010802596807479858, + 0.017089011147618294, + 0.008001933805644512, + 0.006303310859948397, + 0.008453386835753918, + 0.008383340202271938, + 0.005214703734964132, + 0.005394794512540102, + 0.008888404816389084, + 0.010362480767071247, + 0.005418766289949417, + 0.00926906242966652, + 0.007643112912774086, + 0.014265096746385098, + 0.007246225140988827, + 0.005840970668941736, + 0.007700398098677397, + 0.007010567467659712, + 0.03286641836166382, + 0.03578736633062363, + 0.04238970950245857, + 0.05860314145684242, + 0.03383732959628105, + 0.025943294167518616, + 0.06900179386138916, + 0.04048813506960869, + 0.006508907303214073, + 0.008833271451294422, + 0.01226736605167389, + 0.015053356997668743, + 0.009386580437421799, + 0.009115856140851974, + 0.014047243632376194, + 0.013615716248750687, + 0.003921711351722479, + 0.007052507717162371, + 0.007543692830950022, + 0.007598903030157089, + 0.006719101686030626, + 0.0055552516132593155, + 0.007340571377426386, + 0.00779571570456028, + 0.004302591551095247, + 0.0041345031931996346, + 0.004328831564635038, + 0.007088156882673502, + 0.005537663120776415, + 0.004676265642046928, + 0.007010476663708687, + 0.0056664529256522655, + 0.0026589236222207546, + 0.004345062654465437, + 0.004101581405848265, + 0.004632353316992521, + 0.005365622695535421, + 0.0034925781656056643, + 0.004740103613585234, + 0.005675795022398233, + 0.0027704820968210697, + 0.0035912266466766596, + 0.0035191404167562723, + 0.0033969630021601915, + 0.0030958871357142925, + 0.00307456380687654, + 0.0037244190461933613, + 0.003958041779696941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.009784241206943989 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203527010977268 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203527010977268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203527010977268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12461835145950317, + 0.09922347962856293, + 0.04118764400482178, + 0.04323110356926918, + 0.27942174673080444, + 0.11908125877380371, + 0.07996253669261932, + 0.0457492396235466, + 0.032875292003154755, + 0.13464942574501038 + ], + "credit_map": [ + 0.012694966979324818, + 0.016196388751268387, + 0.017071224749088287, + 0.02088366635143757, + 0.01472618617117405, + 0.011207599192857742, + 0.018246110528707504, + 0.013592198491096497, + 0.011522072367370129, + 0.008765882812440395, + 0.010832649655640125, + 0.007206310983747244, + 0.022234292700886726, + 0.01730908453464508, + 0.007071036379784346, + 0.01428215391933918, + 0.00412810780107975, + 0.003706905525177717, + 0.003925036173313856, + 0.005282392259687185, + 0.007083330303430557, + 0.004352539777755737, + 0.006555776111781597, + 0.006153556518256664, + 0.004338917788118124, + 0.004322460852563381, + 0.005289663095027208, + 0.004699615761637688, + 0.006621064618229866, + 0.004472317174077034, + 0.008535793051123619, + 0.004951272159814835, + 0.027324974536895752, + 0.028746260330080986, + 0.027332201600074768, + 0.04880053550004959, + 0.02911679819226265, + 0.02379603125154972, + 0.06985297799110413, + 0.02445194125175476, + 0.009593134745955467, + 0.01986020803451538, + 0.012017695233225822, + 0.017505627125501633, + 0.012881642207503319, + 0.011488720774650574, + 0.02242564968764782, + 0.01330858375877142, + 0.008600384928286076, + 0.011852007359266281, + 0.011217732913792133, + 0.01017033215612173, + 0.010389387607574463, + 0.0070470659993588924, + 0.008331303484737873, + 0.012354322709143162, + 0.003693033941090107, + 0.006253310013562441, + 0.005669786594808102, + 0.012451281771063805, + 0.005028929095715284, + 0.0038614135701209307, + 0.004922710359096527, + 0.0038687726482748985, + 0.0036676479503512383, + 0.003286330960690975, + 0.003224849933758378, + 0.007041964214295149, + 0.003477236023172736, + 0.0031509986147284508, + 0.00544871436432004, + 0.003577551804482937, + 0.012287707068026066, + 0.01859130524098873, + 0.013353223912417889, + 0.028521597385406494, + 0.01454020943492651, + 0.01668417826294899, + 0.01696951314806938, + 0.013701696880161762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.014203527010977268 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204380467534065 + ], + "opposite_logprob_tokens": [ + -2.8120439052581787 + ], + "logprob_reward": -0.06204380467534065, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204380467534065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18449187278747559, + 0.07830401510000229, + 0.0710143968462944, + 0.09551957249641418, + 0.2673695385456085, + 0.12601891160011292, + 0.07176624238491058, + 0.04568053409457207, + 0.030829600989818573, + 0.02900535613298416 + ], + "credit_map": [ + 0.013881882652640343, + 0.017855720594525337, + 0.021402906626462936, + 0.02574164792895317, + 0.033245865255594254, + 0.018915729597210884, + 0.0291305985301733, + 0.02431751973927021, + 0.010765289887785912, + 0.015989935025572777, + 0.009139709174633026, + 0.006733911577612162, + 0.011647244915366173, + 0.007705613039433956, + 0.0061535825952887535, + 0.01016872376203537, + 0.007858507335186005, + 0.012984913773834705, + 0.007700478658080101, + 0.010459239594638348, + 0.00868438370525837, + 0.0066295163705945015, + 0.010494956746697426, + 0.006202390417456627, + 0.008013938553631306, + 0.01675223372876644, + 0.01121217105537653, + 0.016563648357987404, + 0.011936154216527939, + 0.009308280423283577, + 0.012720393016934395, + 0.009012742899358273, + 0.02663031965494156, + 0.026294566690921783, + 0.026535116136074066, + 0.04543743282556534, + 0.026600051671266556, + 0.0222899429500103, + 0.05888480693101883, + 0.034697286784648895, + 0.00798154715448618, + 0.015271761454641819, + 0.015282927080988884, + 0.020798031240701675, + 0.013679301366209984, + 0.012497982010245323, + 0.02311558462679386, + 0.017391758039593697, + 0.006180556025356054, + 0.008049814030528069, + 0.01006301585584879, + 0.008035587146878242, + 0.010042928159236908, + 0.007825287990272045, + 0.007846823893487453, + 0.013722224161028862, + 0.004952211398631334, + 0.004342753905802965, + 0.005137920379638672, + 0.006573413498699665, + 0.006531907711178064, + 0.004938554484397173, + 0.007834689691662788, + 0.005369082558900118, + 0.0032047966960817575, + 0.007750973105430603, + 0.002864573383703828, + 0.0027433368377387524, + 0.0031743841245770454, + 0.00431805057451129, + 0.002327836351469159, + 0.0044456468895077705, + 0.003427153918892145, + 0.0038651868235319853, + 0.0036533670499920845, + 0.004075999371707439, + 0.003760351100936532, + 0.0036686803214251995, + 0.0033879964612424374, + 0.003166617825627327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.06204380467534065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts the dynamic action with clinical precision. A single, spherical jawbreaker, its surface covered in a high-gloss shell, reveals vibrant, swirling layers of electric blue, canary yellow, and fuchsia, bouncing across the smooth, reflective surface of a polished, dark gray concrete floor. Set within a minimalist studio, precisely calibrated lighting from a large, soft overhead source provides perfect, even illumination, creating perfectly crisp highlights on the candy's curvaceous form. The action begins with the jawbreaker's initial impact, its rigid, resilient structure causing it to spring back into the air, spinning gracefully. It then executes a rhythmic series of successive, smaller bounces, losing perceptible energy with each diminishing arc, until its vertical motion seamlessly transitions into a slow, silent roll. As it glides, the mirror-like floor perfectly reflects its swirling colors, creating a liquid play of light. The jaw-dropping object finally sways for a moment before coming to a complete still, its journey complete. The work’s rigorously minimalist approach and cool, neutral color palette isolates the candy’s kinetic behavior, creating an atmosphere of captivation and focus that elevates a simple physical event into a visual poem.", + "seed": 8596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_5/index_1339_seed_8596_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.20755764842033386, + "vlm_logprob": -0.20755764842033386, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.20755764842033386, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005691631697118282 + ], + "opposite_logprob_tokens": [ + -5.3806915283203125 + ], + "logprob_reward": -0.005691631697118282, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005691631697118282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13609296083450317, + 0.03643595799803734, + 0.04614631086587906, + 0.2219575196504593, + 0.06402157247066498, + 0.1803349107503891, + 0.10358498990535736, + 0.08424767851829529, + 0.04263719543814659, + 0.08454088121652603 + ], + "credit_map": [ + 0.012874062173068523, + 0.01400507241487503, + 0.02683335170149803, + 0.009904151782393456, + 0.008650057949125767, + 0.014433181844651699, + 0.008984862826764584, + 0.04040822386741638, + 0.002969622379168868, + 0.0042744786478579044, + 0.005201637279242277, + 0.009371514432132244, + 0.003010546788573265, + 0.0032453113235533237, + 0.002880207961425185, + 0.005482638254761696, + 0.005524226929992437, + 0.0088245440274477, + 0.008586928248405457, + 0.004206840414553881, + 0.0053978897631168365, + 0.004077347926795483, + 0.003420670749619603, + 0.006107859313488007, + 0.0229774359613657, + 0.0108274444937706, + 0.08296170830726624, + 0.014847545884549618, + 0.009414802305400372, + 0.04059881344437599, + 0.008246931247413158, + 0.03208282217383385, + 0.007246214896440506, + 0.013319515623152256, + 0.011886787600815296, + 0.004917117301374674, + 0.009126691147685051, + 0.004978652112185955, + 0.00585944252088666, + 0.0066871484741568565, + 0.01366719976067543, + 0.04969223216176033, + 0.010491828434169292, + 0.017396925017237663, + 0.05268239974975586, + 0.01283507701009512, + 0.013189026154577732, + 0.010380215011537075, + 0.007295059971511364, + 0.03343645855784416, + 0.017831997945904732, + 0.004614474717527628, + 0.02187575213611126, + 0.005363902542740107, + 0.00565835228189826, + 0.007508996874094009, + 0.003838290460407734, + 0.03331763297319412, + 0.005311568267643452, + 0.003212320152670145, + 0.02691715396940708, + 0.004754770081490278, + 0.003479005303233862, + 0.0034169384744018316, + 0.007610374595969915, + 0.002872090321034193, + 0.011334031820297241, + 0.004303213208913803, + 0.0026803852524608374, + 0.004018319305032492, + 0.004303699359297752, + 0.005515080876648426, + 0.013515361584722996, + 0.006933433935046196, + 0.017481593415141106, + 0.010443096980452538, + 0.006558238063007593, + 0.00822659581899643, + 0.012820698320865631, + 0.008561857044696808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.005691631697118282 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089006438851357 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011089006438851357, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089006438851357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06071130558848381, + 0.023672867566347122, + 0.03654545173048973, + 0.022848015651106834, + 0.015506613068282604, + 0.30724453926086426, + 0.03143000975251198, + 0.4067199230194092, + 0.034393634647130966, + 0.060927655547857285 + ], + "credit_map": [ + 0.004382535815238953, + 0.012276927940547466, + 0.010222122073173523, + 0.005731946788728237, + 0.009922843426465988, + 0.005996108520776033, + 0.005000382196158171, + 0.007178437430411577, + 0.002472425578162074, + 0.003398995380848646, + 0.0020203625317662954, + 0.00373586593195796, + 0.004014384932816029, + 0.0017780052730813622, + 0.0022826134227216244, + 0.003970214631408453, + 0.0033117374405264854, + 0.00550209591165185, + 0.008091709576547146, + 0.0040006074123084545, + 0.003768360475078225, + 0.00404956191778183, + 0.0031382679007947445, + 0.004683113656938076, + 0.0020879802759736776, + 0.0038346420042216778, + 0.00560351274907589, + 0.0027186409570276737, + 0.002179492264986038, + 0.0024682716466486454, + 0.0013496045721694827, + 0.0026058724615722895, + 0.0013732177903875709, + 0.0018483522580936551, + 0.0036523831076920033, + 0.0013985437108203769, + 0.0010980813531205058, + 0.0017178956186398864, + 0.0019210611935704947, + 0.0024970790836960077, + 0.013539716601371765, + 0.12304497510194778, + 0.0074849785305559635, + 0.012698222883045673, + 0.11044365912675858, + 0.013835226185619831, + 0.019785553216934204, + 0.00641221646219492, + 0.0021795209031552076, + 0.00960464309900999, + 0.006039887201040983, + 0.0013120099902153015, + 0.006124213803559542, + 0.0014552201610058546, + 0.0022092540748417377, + 0.0025052602868527174, + 0.022047819569706917, + 0.166853666305542, + 0.03958439454436302, + 0.013902955688536167, + 0.12528309226036072, + 0.01857304573059082, + 0.00684876274317503, + 0.01362619549036026, + 0.002401200355961919, + 0.003488890826702118, + 0.004747169557958841, + 0.005478817503899336, + 0.0032881188672035933, + 0.0027155266143381596, + 0.008443254977464676, + 0.003830655012279749, + 0.007013935130089521, + 0.005922969430685043, + 0.010680211707949638, + 0.008276422508060932, + 0.005543810315430164, + 0.0061429147608578205, + 0.010406145825982094, + 0.006941245403140783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.011089006438851357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086353225633502 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.0086353225633502, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086353225633502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13379864394664764, + 0.053518980741500854, + 0.045930344611406326, + 0.09855564683675766, + 0.035751089453697205, + 0.08109360933303833, + 0.11738615483045578, + 0.3928614556789398, + 0.02174750156700611, + 0.019356535747647285 + ], + "credit_map": [ + 0.010758647695183754, + 0.028496578335762024, + 0.02705960161983967, + 0.00893913209438324, + 0.020684104412794113, + 0.0146830128505826, + 0.007946210913360119, + 0.015231366269290447, + 0.005438018590211868, + 0.007456133142113686, + 0.0054801334626972675, + 0.010701576247811317, + 0.006842877250164747, + 0.005995196755975485, + 0.006715756841003895, + 0.004889289382845163, + 0.0048232171684503555, + 0.006278787273913622, + 0.005912794265896082, + 0.00359150767326355, + 0.0029369299300014973, + 0.0050070337019860744, + 0.003904610639438033, + 0.013475462794303894, + 0.008497913368046284, + 0.008952727541327477, + 0.025889281183481216, + 0.008265684358775616, + 0.006540640722960234, + 0.01572335511445999, + 0.006082777865231037, + 0.018603263422846794, + 0.006205754820257425, + 0.0031611823942512274, + 0.00612106267362833, + 0.005289125721901655, + 0.0033164038322865963, + 0.003781538922339678, + 0.00459717120975256, + 0.0032788505777716637, + 0.0050685047172009945, + 0.023793712258338928, + 0.008454500697553158, + 0.00686073349788785, + 0.018934139981865883, + 0.004441298544406891, + 0.004595302045345306, + 0.00894541572779417, + 0.007216054014861584, + 0.0394083671271801, + 0.020336756482720375, + 0.006399187725037336, + 0.025575386360287666, + 0.005235057789832354, + 0.007008220534771681, + 0.006207122001796961, + 0.030893469229340553, + 0.16278718411922455, + 0.011117673479020596, + 0.01658600941300392, + 0.12814536690711975, + 0.025227447971701622, + 0.009218230843544006, + 0.008886056020855904, + 0.0029409590642899275, + 0.0016628874000161886, + 0.003464137203991413, + 0.0021601265762001276, + 0.001418373198248446, + 0.005032688844949007, + 0.0026928619481623173, + 0.002375468611717224, + 0.0019242791458964348, + 0.0021995394490659237, + 0.002496860222890973, + 0.002834717743098736, + 0.0016374748665839434, + 0.0021593691781163216, + 0.002963816514238715, + 0.003140477929264307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0086353225633502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097646087408066 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097646087408066, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097646087408066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06081871688365936, + 0.04091135412454605, + 0.018716761842370033, + 0.06870628148317337, + 0.08184576034545898, + 0.08053351938724518, + 0.08329281955957413, + 0.48769232630729675, + 0.035877157002687454, + 0.04160520061850548 + ], + "credit_map": [ + 0.004943202715367079, + 0.010278106667101383, + 0.010289963334798813, + 0.0038082432001829147, + 0.010010267607867718, + 0.005642953794449568, + 0.0037954975850880146, + 0.012050498276948929, + 0.003304883372038603, + 0.005040414165705442, + 0.006392872426658869, + 0.009696295484900475, + 0.0036084370221942663, + 0.003959265071898699, + 0.0041060238145291805, + 0.0048031737096607685, + 0.002433954505249858, + 0.002665905049070716, + 0.00267076981253922, + 0.0017935973592102528, + 0.001676678191870451, + 0.0015419184928759933, + 0.0016256151720881462, + 0.004308327101171017, + 0.005925843026489019, + 0.009662012569606304, + 0.014679941348731518, + 0.01059713214635849, + 0.005767669063061476, + 0.008916632272303104, + 0.005410333164036274, + 0.00774672906845808, + 0.0072410451248288155, + 0.020944006741046906, + 0.014789278618991375, + 0.006109099369496107, + 0.01452605426311493, + 0.0052901809103786945, + 0.005822477862238884, + 0.007123630028218031, + 0.007234153337776661, + 0.008264673873782158, + 0.014251938089728355, + 0.011204561218619347, + 0.007939541712403297, + 0.007754781283438206, + 0.010148315690457821, + 0.013735569082200527, + 0.004684668965637684, + 0.027238333597779274, + 0.009943244978785515, + 0.005976121872663498, + 0.0186277087777853, + 0.004265979863703251, + 0.007481473032385111, + 0.00507530989125371, + 0.033683255314826965, + 0.16003838181495667, + 0.10586553812026978, + 0.014477930963039398, + 0.11222753673791885, + 0.018753977492451668, + 0.007177228573709726, + 0.03546854108572006, + 0.004806265700608492, + 0.0038645281456410885, + 0.005711809732019901, + 0.005189116578549147, + 0.0029181584250181913, + 0.0033672787249088287, + 0.005107284989207983, + 0.004912723787128925, + 0.005380123388022184, + 0.005368752870708704, + 0.005394189152866602, + 0.005126845091581345, + 0.003999661188572645, + 0.0049043018370866776, + 0.005225789733231068, + 0.006205544341355562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.038097646087408066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742746353149414 + ], + "opposite_logprob_tokens": [ + -0.4742746353149414 + ], + "logprob_reward": -0.9742746353149414, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742746353149414, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04019497334957123, + 0.1678488850593567, + 0.013189810328185558, + 0.02491661347448826, + 0.04304875433444977, + 0.1535276621580124, + 0.062197111546993256, + 0.477448433637619, + 0.009440954774618149, + 0.00818688329309225 + ], + "credit_map": [ + 0.0032644225284457207, + 0.007761362008750439, + 0.005083292257040739, + 0.003975809086114168, + 0.006490770727396011, + 0.0034418501891195774, + 0.002565257018432021, + 0.0076122102327644825, + 0.018087144941091537, + 0.026453664526343346, + 0.01753244549036026, + 0.02218659780919552, + 0.040257398039102554, + 0.015438269823789597, + 0.016162676736712456, + 0.011730688624083996, + 0.001700639957562089, + 0.0023245890624821186, + 0.001874596462585032, + 0.0011351910652592778, + 0.0010697359684854746, + 0.0013665646547451615, + 0.0011473445920273662, + 0.0025711485650390387, + 0.002146832412108779, + 0.0036434473004192114, + 0.005283349193632603, + 0.002629912458360195, + 0.00245231157168746, + 0.0034482532646507025, + 0.0017745831282809377, + 0.0035379237961024046, + 0.005027066450566053, + 0.008559309877455235, + 0.004529588855803013, + 0.0037434622645378113, + 0.0058075133711099625, + 0.005777565762400627, + 0.004072233103215694, + 0.005532016512006521, + 0.005973296705633402, + 0.061139144003391266, + 0.007106958422809839, + 0.006641109939664602, + 0.054281238466501236, + 0.007504432927817106, + 0.0065987021662294865, + 0.004282788373529911, + 0.00444380659610033, + 0.019255414605140686, + 0.013181476853787899, + 0.0021867589093744755, + 0.012066619470715523, + 0.003023463301360607, + 0.002756210044026375, + 0.005283367354422808, + 0.03636989742517471, + 0.1507805436849594, + 0.11374837905168533, + 0.014117861166596413, + 0.10150463134050369, + 0.01926969178020954, + 0.0053377351723611355, + 0.036319710314273834, + 0.0008901773253455758, + 0.0010106649715453386, + 0.0015334316994994879, + 0.0010704189771786332, + 0.0008912636549212039, + 0.0009952899999916553, + 0.0013827539514750242, + 0.0016669542528688908, + 0.0010144838597625494, + 0.0011266652727499604, + 0.001438140869140625, + 0.0007910660933703184, + 0.0008521938580088317, + 0.0007103584939613938, + 0.0009846435859799385, + 0.0012693313183262944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.9742746353149414 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts the dynamic action with clinical precision. A single, spherical jawbreaker, its surface covered in a high-gloss shell, reveals vibrant, swirling layers of electric blue, canary yellow, and fuchsia, bouncing across the smooth, reflective surface of a polished, dark gray concrete floor. Set within a minimalist studio, precisely calibrated lighting from a large, soft overhead source provides perfect, even illumination, creating perfectly crisp highlights on the candy's curvaceous form. The action begins with the jawbreaker's initial impact, its rigid, resilient structure causing it to spring back into the air, spinning gracefully. It then executes a rhythmic series of successive, smaller bounces, losing perceptible energy with each diminishing arc, until its vertical motion seamlessly transitions into a slow, silent roll. As it glides, the mirror-like floor perfectly reflects its swirling colors, creating a liquid play of light. The jaw-dropping object finally sways for a moment before coming to a complete still, its journey complete. The work’s rigorously minimalist approach and cool, neutral color palette isolates the candy’s kinetic behavior, creating an atmosphere of captivation and focus that elevates a simple physical event into a visual poem.", + "seed": 8597, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_5/index_1339_seed_8597_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.044598168972879645, + "vlm_logprob": -0.044598168972879645, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.044598168972879645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004008236341178417 + ], + "opposite_logprob_tokens": [ + -5.8790082931518555 + ], + "logprob_reward": -0.004008236341178417, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004008236341178417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18977566063404083, + 0.12101645767688751, + 0.0649501159787178, + 0.16572165489196777, + 0.026399845257401466, + 0.06158161908388138, + 0.14034311473369598, + 0.09726502001285553, + 0.08051878213882446, + 0.052427712827920914 + ], + "credit_map": [ + 0.031185105443000793, + 0.011287209577858448, + 0.025985363870859146, + 0.010749822482466698, + 0.029055580496788025, + 0.04338855668902397, + 0.009051511995494366, + 0.029072515666484833, + 0.0147035988047719, + 0.0076516130939126015, + 0.018287835642695427, + 0.011066553182899952, + 0.010467341169714928, + 0.023406656458973885, + 0.017987336963415146, + 0.017445527017116547, + 0.0077164932154119015, + 0.005670695565640926, + 0.009863377548754215, + 0.00837303139269352, + 0.004583335947245359, + 0.0072328695096075535, + 0.010471636429429054, + 0.011038681492209435, + 0.021081551909446716, + 0.01649068109691143, + 0.017457634210586548, + 0.022169820964336395, + 0.015192806720733643, + 0.0307680182158947, + 0.03014017641544342, + 0.01242096722126007, + 0.0034539508633315563, + 0.0016592956380918622, + 0.005069086793810129, + 0.003044029464945197, + 0.0013704755110666156, + 0.003467714414000511, + 0.0033962002489715815, + 0.00493909465149045, + 0.006972858216613531, + 0.006074483972042799, + 0.009540947154164314, + 0.011645214632153511, + 0.004231131169945002, + 0.00701249111443758, + 0.008503066375851631, + 0.007601429242640734, + 0.01695381850004196, + 0.009342024102807045, + 0.021618111059069633, + 0.022240448743104935, + 0.007033290341496468, + 0.018101880326867104, + 0.015373588539659977, + 0.02967996522784233, + 0.017218999564647675, + 0.006317518185824156, + 0.01048437412828207, + 0.016945578157901764, + 0.00398535281419754, + 0.024079665541648865, + 0.008183106780052185, + 0.010050436481833458, + 0.011493314988911152, + 0.005624307319521904, + 0.009483863599598408, + 0.015984181314706802, + 0.0038347935769706964, + 0.011440463364124298, + 0.008100624196231365, + 0.014557241462171078, + 0.005397642496973276, + 0.0031037465669214725, + 0.007337808143347502, + 0.010946978814899921, + 0.0031817033886909485, + 0.005566212348639965, + 0.0065676746889948845, + 0.010325945913791656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.004008236341178417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01109018549323082 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.01109018549323082, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01109018549323082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24999293684959412, + 0.068643718957901, + 0.04781996086239815, + 0.2681696116924286, + 0.014232371002435684, + 0.12246975302696228, + 0.031944580376148224, + 0.04426579922437668, + 0.06957216560840607, + 0.082889124751091 + ], + "credit_map": [ + 0.04771805182099342, + 0.01843411475419998, + 0.02862967737019062, + 0.01164165697991848, + 0.04315631091594696, + 0.061608802527189255, + 0.013731497339904308, + 0.025072850286960602, + 0.006809948477894068, + 0.006706169340759516, + 0.008144328370690346, + 0.0056722876615822315, + 0.009624496102333069, + 0.0068910266272723675, + 0.013998365961015224, + 0.010797095485031605, + 0.005672055762261152, + 0.004480492789298296, + 0.006642171647399664, + 0.006147292442619801, + 0.003765486180782318, + 0.005059052258729935, + 0.008182241581380367, + 0.007871168665587902, + 0.03560461848974228, + 0.02562073990702629, + 0.02703368104994297, + 0.0341842919588089, + 0.024417638778686523, + 0.05184529721736908, + 0.049253348261117935, + 0.02020999602973461, + 0.0016971636796370149, + 0.0010298910783603787, + 0.001821447047404945, + 0.0018738694489002228, + 0.0009893630631268024, + 0.002425424987450242, + 0.0026753395795822144, + 0.001719872234389186, + 0.01417532842606306, + 0.008966109715402126, + 0.013954391703009605, + 0.015351141802966595, + 0.006986383348703384, + 0.020266860723495483, + 0.015207255259156227, + 0.027562296018004417, + 0.004788440652191639, + 0.0023078101221472025, + 0.003942215349525213, + 0.004557669162750244, + 0.001439181505702436, + 0.004573140759021044, + 0.0034015104174613953, + 0.006934609264135361, + 0.00668686255812645, + 0.003227599197998643, + 0.007172968238592148, + 0.006405093241482973, + 0.0020763895008713007, + 0.0058481115847826, + 0.004001519177109003, + 0.008847259916365147, + 0.010599175468087196, + 0.005863054655492306, + 0.008210603147745132, + 0.014314280822873116, + 0.0037627897690981627, + 0.010243306867778301, + 0.006554492749273777, + 0.010024472139775753, + 0.010887427255511284, + 0.0066831293515861034, + 0.009530961513519287, + 0.017807690426707268, + 0.005527232773602009, + 0.010909343138337135, + 0.009525911882519722, + 0.01201743632555008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.01109018549323082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948459263890982 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948459263890982, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948459263890982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1957414150238037, + 0.07407946139574051, + 0.11493610590696335, + 0.05924845114350319, + 0.027353806421160698, + 0.24778808653354645, + 0.08308477699756622, + 0.07702295482158661, + 0.03421829268336296, + 0.08652669191360474 + ], + "credit_map": [ + 0.025534944608807564, + 0.016680220142006874, + 0.022433428093791008, + 0.014594909735023975, + 0.04789755120873451, + 0.034400761127471924, + 0.013249240815639496, + 0.020950352773070335, + 0.007119039539247751, + 0.0067419144324958324, + 0.008276870474219322, + 0.006822880823165178, + 0.00911747571080923, + 0.008380892686545849, + 0.013612354174256325, + 0.014008031226694584, + 0.01360282301902771, + 0.011162961833178997, + 0.016976341605186462, + 0.01491463277488947, + 0.008645560592412949, + 0.012773189693689346, + 0.018889298662543297, + 0.01797129586338997, + 0.009233451448380947, + 0.005125254392623901, + 0.0063017066568136215, + 0.007202102802693844, + 0.00458881352096796, + 0.012823124416172504, + 0.008623719215393066, + 0.005350279621779919, + 0.0033946745097637177, + 0.0021117059513926506, + 0.004288491327315569, + 0.0026374494191259146, + 0.0014122374122962356, + 0.004183829762041569, + 0.0026189424097537994, + 0.006706474348902702, + 0.027389029040932655, + 0.021121405065059662, + 0.03378104791045189, + 0.03167514130473137, + 0.015073439106345177, + 0.02905890718102455, + 0.02892678789794445, + 0.06076231226325035, + 0.011622854508459568, + 0.005045154131948948, + 0.013765561394393444, + 0.008857131004333496, + 0.004352681338787079, + 0.011347427032887936, + 0.007380926515907049, + 0.020713036879897118, + 0.011825847439467907, + 0.003912964835762978, + 0.012872704304754734, + 0.005919273477047682, + 0.002968129003420472, + 0.013627936132252216, + 0.0039822859689593315, + 0.021913813427090645, + 0.0035852957516908646, + 0.00253119645640254, + 0.005246770568192005, + 0.004952546674758196, + 0.0015978000592440367, + 0.005107946693897247, + 0.0030978708527982235, + 0.008098866790533066, + 0.015654906630516052, + 0.004363431595265865, + 0.014428145252168179, + 0.006565408315509558, + 0.005484686233103275, + 0.015101444907486439, + 0.005686771124601364, + 0.01924189366400242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.005948459263890982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902833610773087 + ], + "opposite_logprob_tokens": [ + -2.4640283584594727 + ], + "logprob_reward": -0.08902833610773087, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902833610773087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15073028206825256, + 0.1338585764169693, + 0.08413807302713394, + 0.20996259152889252, + 0.03847932815551758, + 0.10451390594244003, + 0.09619449824094772, + 0.07143101096153259, + 0.07480521500110626, + 0.035886552184820175 + ], + "credit_map": [ + 0.024272577837109566, + 0.013245537877082825, + 0.01486838236451149, + 0.007123949006199837, + 0.0328109934926033, + 0.0324384868144989, + 0.009080794639885426, + 0.0168895460665226, + 0.016477877274155617, + 0.012047000229358673, + 0.017749980092048645, + 0.010286972858011723, + 0.01665339805185795, + 0.014326606877148151, + 0.025165822356939316, + 0.02115090750157833, + 0.010118186473846436, + 0.006538459099829197, + 0.009127721190452576, + 0.00984412245452404, + 0.006616363301873207, + 0.011992644518613815, + 0.01885589212179184, + 0.011044683866202831, + 0.027291806414723396, + 0.020803239196538925, + 0.02034398913383484, + 0.027784176170825958, + 0.019117658957839012, + 0.03978773579001427, + 0.0388847216963768, + 0.015949254855513573, + 0.006115063559263945, + 0.00258855358697474, + 0.007227132562547922, + 0.0042845020070672035, + 0.0017240121960639954, + 0.00564591633155942, + 0.004388032481074333, + 0.006506112404167652, + 0.013477581553161144, + 0.008878623135387897, + 0.014071796089410782, + 0.01882813684642315, + 0.0067712184973061085, + 0.01473942119628191, + 0.01420340035110712, + 0.013543721288442612, + 0.009955157525837421, + 0.007045235950499773, + 0.012226817198097706, + 0.019921842962503433, + 0.006424315739423037, + 0.014384807087481022, + 0.012521224096417427, + 0.013715093955397606, + 0.012482733465731144, + 0.005098415073007345, + 0.008733885362744331, + 0.008408959023654461, + 0.0025709616020321846, + 0.015774153172969818, + 0.0052576311863958836, + 0.013104264624416828, + 0.013412531465291977, + 0.004151554312556982, + 0.008282801136374474, + 0.011273197829723358, + 0.003373414045199752, + 0.018071718513965607, + 0.005895715206861496, + 0.010344269685447216, + 0.0055101546458899975, + 0.0020617851987481117, + 0.005725283175706863, + 0.0039612967520952225, + 0.0023679782170802355, + 0.005755754187703133, + 0.0024501648731529713, + 0.008054131641983986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.08902833610773087 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291562765836716 + ], + "opposite_logprob_tokens": [ + -2.237915515899658 + ], + "logprob_reward": -0.11291562765836716, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291562765836716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2370389997959137, + 0.11135256290435791, + 0.09108627587556839, + 0.033977922052145004, + 0.03288695216178894, + 0.3608948290348053, + 0.03217891603708267, + 0.030267341062426567, + 0.019825510680675507, + 0.050490643829107285 + ], + "credit_map": [ + 0.03422076255083084, + 0.02186962217092514, + 0.024067861959338188, + 0.020627707242965698, + 0.05259353294968605, + 0.04372232407331467, + 0.01671696826815605, + 0.023220224305987358, + 0.012675725854933262, + 0.009064143523573875, + 0.012287414632737637, + 0.01107940636575222, + 0.009117730893194675, + 0.014062965288758278, + 0.019190937280654907, + 0.02387424372136593, + 0.010279574431478977, + 0.010038108564913273, + 0.01335267536342144, + 0.011698964051902294, + 0.007223658729344606, + 0.009608965367078781, + 0.014843503944575787, + 0.014040828682482243, + 0.004946017637848854, + 0.003079943126067519, + 0.004835761617869139, + 0.003287107218056917, + 0.0023208933416754007, + 0.005801060236990452, + 0.004973423667252064, + 0.004733715206384659, + 0.004809602629393339, + 0.0021169811952859163, + 0.003663236740976572, + 0.0032485800329595804, + 0.0014931457117199898, + 0.005922410171478987, + 0.003353692824020982, + 0.008279303088784218, + 0.038838181644678116, + 0.029632512480020523, + 0.04898908734321594, + 0.04734710231423378, + 0.021784707903862, + 0.04223340377211571, + 0.04121842607855797, + 0.09085144847631454, + 0.00425950950011611, + 0.0021094230469316244, + 0.0054426975548267365, + 0.003567836945876479, + 0.001367792603559792, + 0.003938073758035898, + 0.0023509736638516188, + 0.009142608381807804, + 0.005185337271541357, + 0.0020609546918421984, + 0.0035771490074694157, + 0.004155890084803104, + 0.0012667677365243435, + 0.005609613377600908, + 0.0025311328936368227, + 0.005880493670701981, + 0.0022781880106776953, + 0.001066631288267672, + 0.0037923234049230814, + 0.002514839405193925, + 0.001088864984922111, + 0.0026741630863398314, + 0.0018224166706204414, + 0.0045880828984081745, + 0.0051961494609713554, + 0.0031299598049372435, + 0.010011928156018257, + 0.005407013464719057, + 0.004313880577683449, + 0.005841391626745462, + 0.0043934667482972145, + 0.012196853756904602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.11291562765836716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot depicts the dynamic action with clinical precision. A single, spherical jawbreaker, its surface covered in a high-gloss shell, reveals vibrant, swirling layers of electric blue, canary yellow, and fuchsia, bouncing across the smooth, reflective surface of a polished, dark gray concrete floor. Set within a minimalist studio, precisely calibrated lighting from a large, soft overhead source provides perfect, even illumination, creating perfectly crisp highlights on the candy's curvaceous form. The action begins with the jawbreaker's initial impact, its rigid, resilient structure causing it to spring back into the air, spinning gracefully. It then executes a rhythmic series of successive, smaller bounces, losing perceptible energy with each diminishing arc, until its vertical motion seamlessly transitions into a slow, silent roll. As it glides, the mirror-like floor perfectly reflects its swirling colors, creating a liquid play of light. The jaw-dropping object finally sways for a moment before coming to a complete still, its journey complete. The work’s rigorously minimalist approach and cool, neutral color palette isolates the candy’s kinetic behavior, creating an atmosphere of captivation and focus that elevates a simple physical event into a visual poem.", + "seed": 8598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_5/index_1339_seed_8598_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013413283322006465, + "vlm_logprob": -0.013413283322006465, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.013413283322006465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015731986612081528 + ], + "opposite_logprob_tokens": [ + -4.2657318115234375 + ], + "logprob_reward": -0.015731986612081528, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015731986612081528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04736936837434769, + 0.027659837156534195, + 0.1324825882911682, + 0.04931211099028587, + 0.2622920572757721, + 0.2916885018348694, + 0.11292513459920883, + 0.025915972888469696, + 0.03858828544616699, + 0.011766087263822556 + ], + "credit_map": [ + 0.0037832262460142374, + 0.0052630966529250145, + 0.007018565200269222, + 0.007665857672691345, + 0.007059811148792505, + 0.005373183637857437, + 0.004408516455441713, + 0.006797118112444878, + 0.0024290853179991245, + 0.005209966562688351, + 0.0034974883310496807, + 0.002258880762383342, + 0.0036235086154192686, + 0.003817508462816477, + 0.003150452161207795, + 0.0036729490384459496, + 0.007733279373496771, + 0.0516790896654129, + 0.003616586560383439, + 0.0031869246158748865, + 0.008430072106420994, + 0.050116680562496185, + 0.00439553102478385, + 0.0033244409132748842, + 0.0034885164350271225, + 0.013930734246969223, + 0.0031790847424417734, + 0.0031102560460567474, + 0.005541670136153698, + 0.011541824787855148, + 0.005686256103217602, + 0.002833776408806443, + 0.014306674711406231, + 0.10658805072307587, + 0.004906932357698679, + 0.005602753721177578, + 0.017948830500245094, + 0.10446155071258545, + 0.003127950942143798, + 0.00534930732101202, + 0.01621531881392002, + 0.11804825812578201, + 0.004964147228747606, + 0.005422590766102076, + 0.018050141632556915, + 0.1200690045952797, + 0.0032035864423960447, + 0.005715467967092991, + 0.006186006125062704, + 0.038254085928201675, + 0.007020145654678345, + 0.005135230254381895, + 0.005880820564925671, + 0.03841753676533699, + 0.004880334250628948, + 0.007150984834879637, + 0.0018279600189998746, + 0.008458691649138927, + 0.0016730751376599073, + 0.001360670430585742, + 0.002801370108500123, + 0.005457725841552019, + 0.002178356982767582, + 0.0021581253968179226, + 0.0021997028961777687, + 0.014589953236281872, + 0.001427495968528092, + 0.0011194392573088408, + 0.0025120959617197514, + 0.013804062269628048, + 0.0016613614279776812, + 0.0012741769896820188, + 0.0007174769416451454, + 0.0030400280375033617, + 0.0013959819916635752, + 0.0007866070372983813, + 0.0008309194236062467, + 0.0034374981187283993, + 0.0009130067774094641, + 0.0006445695762522519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.015731986612081528 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086597865447402 + ], + "opposite_logprob_tokens": [ + -4.758659839630127 + ], + "logprob_reward": -0.0086597865447402, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086597865447402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14155307412147522, + 0.34106844663619995, + 0.05348238721489906, + 0.04919356852769852, + 0.05888237804174423, + 0.12024588137865067, + 0.0935860425233841, + 0.0497467927634716, + 0.05416083335876465, + 0.03808055818080902 + ], + "credit_map": [ + 0.012994665652513504, + 0.016125036403536797, + 0.015910258516669273, + 0.014158893376588821, + 0.03254464268684387, + 0.017951250076293945, + 0.01585126854479313, + 0.016017062589526176, + 0.022739896550774574, + 0.1276213675737381, + 0.005276188254356384, + 0.007776212878525257, + 0.017358234152197838, + 0.14728669822216034, + 0.005899103358387947, + 0.00711077731102705, + 0.004939997103065252, + 0.008997212164103985, + 0.004980496130883694, + 0.004315528552979231, + 0.007182654924690723, + 0.006378220394253731, + 0.010962770320475101, + 0.005725513212382793, + 0.0035786114167422056, + 0.006496037822216749, + 0.006287051364779472, + 0.006651291158050299, + 0.007878748700022697, + 0.005605053156614304, + 0.006442440208047628, + 0.006254339590668678, + 0.0036497495602816343, + 0.01285356841981411, + 0.006499055307358503, + 0.006176041904836893, + 0.006611973978579044, + 0.01029799785465002, + 0.007880454882979393, + 0.004913536366075277, + 0.007178816478699446, + 0.04402432590723038, + 0.0029311662074178457, + 0.003358092624694109, + 0.00688022468239069, + 0.04820254445075989, + 0.004274959675967693, + 0.003395753912627697, + 0.004967657383531332, + 0.03420533612370491, + 0.004877780564129353, + 0.003911993466317654, + 0.0065631503239274025, + 0.03023742511868477, + 0.005497639067471027, + 0.003325068624690175, + 0.004245702177286148, + 0.011023939587175846, + 0.006046751979738474, + 0.004732613451778889, + 0.005737700499594212, + 0.006587750744074583, + 0.006151666399091482, + 0.005220675840973854, + 0.003377871122211218, + 0.0186140239238739, + 0.003607796272262931, + 0.002598193706944585, + 0.004201236646622419, + 0.015630237758159637, + 0.0035287667997181416, + 0.0026027136482298374, + 0.0019736764952540398, + 0.012406662106513977, + 0.0020659510046243668, + 0.0025046963710337877, + 0.001968831056728959, + 0.013320869766175747, + 0.001793408882804215, + 0.0020464668050408363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0086597865447402 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08920542895793915, + 0.048893269151449203, + 0.0677105113863945, + 0.017605243250727654, + 0.27227282524108887, + 0.19745101034641266, + 0.12343446910381317, + 0.07661563158035278, + 0.04012633487582207, + 0.06668532639741898 + ], + "credit_map": [ + 0.00895170122385025, + 0.0118075767531991, + 0.014012075960636139, + 0.014061853289604187, + 0.009436501190066338, + 0.011593630537390709, + 0.00802569929510355, + 0.01131637766957283, + 0.004517186898738146, + 0.0054153939709067345, + 0.009085399098694324, + 0.006881450302898884, + 0.004127351101487875, + 0.004958664067089558, + 0.005145098548382521, + 0.008762720040977001, + 0.004325328394770622, + 0.02066951058804989, + 0.006062154658138752, + 0.005718523170799017, + 0.004142813850194216, + 0.01916191168129444, + 0.002951079746708274, + 0.004679173696786165, + 0.0012582483468577266, + 0.0027097624260932207, + 0.0025626104325056076, + 0.0024689962156116962, + 0.001999563304707408, + 0.0026164883747696877, + 0.0021630318369716406, + 0.0018265395192429423, + 0.014757340773940086, + 0.11267854273319244, + 0.004248935263603926, + 0.005363315809518099, + 0.020008355379104614, + 0.10777605324983597, + 0.0023088939487934113, + 0.005131364334374666, + 0.010525941848754883, + 0.07824999839067459, + 0.005656342022120953, + 0.004931941628456116, + 0.013098420575261116, + 0.07708344608545303, + 0.0032021314837038517, + 0.004702769685536623, + 0.007273449562489986, + 0.042056672275066376, + 0.007799533195793629, + 0.007484801113605499, + 0.006029326934367418, + 0.0428658090531826, + 0.0035441434010863304, + 0.0063807303085923195, + 0.003952423110604286, + 0.027383089065551758, + 0.004969618283212185, + 0.004338144324719906, + 0.006485871970653534, + 0.02227470837533474, + 0.0030086154583841562, + 0.0042031570337712765, + 0.00252897129394114, + 0.012969017960131168, + 0.00309661403298378, + 0.0028172670863568783, + 0.0024738935753703117, + 0.012097804807126522, + 0.0015483885072171688, + 0.0025943713262677193, + 0.00359504297375679, + 0.025422988459467888, + 0.0024078392889350653, + 0.002372932154685259, + 0.003662313800305128, + 0.025151163339614868, + 0.0015056643169373274, + 0.002567366696894169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.005252490285784006 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078395880758762 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078395880758762, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078395880758762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04782700911164284, + 0.13071101903915405, + 0.3552241027355194, + 0.0717012882232666, + 0.2257126122713089, + 0.047638557851314545, + 0.026028256863355637, + 0.02077656425535679, + 0.04495266079902649, + 0.02942791022360325 + ], + "credit_map": [ + 0.003926577512174845, + 0.005162458401173353, + 0.006423832383006811, + 0.00549590028822422, + 0.008378018625080585, + 0.006730629131197929, + 0.004415051080286503, + 0.00729454355314374, + 0.007681136019527912, + 0.04663361608982086, + 0.003760276362299919, + 0.0046033295802772045, + 0.0077525717206299305, + 0.04985242709517479, + 0.006253684870898724, + 0.00417398102581501, + 0.021121202036738396, + 0.141198992729187, + 0.008538197726011276, + 0.0069967168383300304, + 0.02145947515964508, + 0.14085881412029266, + 0.006572945509105921, + 0.008477780967950821, + 0.003434901125729084, + 0.023744769394397736, + 0.0029072288889437914, + 0.004413314629346132, + 0.006877756677567959, + 0.0216692965477705, + 0.005574923940002918, + 0.003079097019508481, + 0.012324816547334194, + 0.09056398272514343, + 0.003990089986473322, + 0.004380861297249794, + 0.016377020627260208, + 0.08884473145008087, + 0.004733873996883631, + 0.004497244022786617, + 0.002368127228692174, + 0.012793129310011864, + 0.004169303923845291, + 0.0029860157519578934, + 0.0033358170185238123, + 0.013814523816108704, + 0.00374894542619586, + 0.004422695375978947, + 0.002028780523687601, + 0.0033360691741108894, + 0.0038448635023087263, + 0.0038181960117071867, + 0.0029400561470538378, + 0.0026206406764686108, + 0.004239857196807861, + 0.003199794562533498, + 0.0012457467382773757, + 0.005123062524944544, + 0.002107014413923025, + 0.002255542203783989, + 0.0015745264245197177, + 0.005257225129753351, + 0.001094577950425446, + 0.002118870848789811, + 0.003261942183598876, + 0.004181277938187122, + 0.008173051290214062, + 0.00582424271851778, + 0.005493097938597202, + 0.0046190437860786915, + 0.006526344455778599, + 0.006873661186546087, + 0.002009997144341469, + 0.005226500332355499, + 0.004299993626773357, + 0.0033842602279037237, + 0.003450024640187621, + 0.004715730436146259, + 0.003934621810913086, + 0.002406782004982233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.011078395880758762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single spherical jawbreaker with a high-gloss shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jawbreaker display vibrant swirling layers of electric blue, canary yellow, and fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a minimalist studio with a polished dark gray concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the jawbreaker bounce, spin, and roll across the floor before coming to a complete still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, eye-level medium shot with even overhead lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343757286667824 + ], + "opposite_logprob_tokens": [ + -3.651343822479248 + ], + "logprob_reward": -0.026343757286667824, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343757286667824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10628453642129898, + 0.07309360802173615, + 0.12791532278060913, + 0.03342790901660919, + 0.18870024383068085, + 0.12591023743152618, + 0.15519225597381592, + 0.043494101613759995, + 0.11554504930973053, + 0.030436785891652107 + ], + "credit_map": [ + 0.008777813985943794, + 0.009625210426747799, + 0.013607828877866268, + 0.016155777499079704, + 0.013856093399226665, + 0.011581283994019032, + 0.012551922351121902, + 0.020128611475229263, + 0.004839231260120869, + 0.020741397514939308, + 0.004955415613949299, + 0.004204381722956896, + 0.004017751198261976, + 0.023854441940784454, + 0.0043986281380057335, + 0.006082357373088598, + 0.007585680577903986, + 0.04114287719130516, + 0.008647026494145393, + 0.006743183825165033, + 0.0074446494691073895, + 0.04409283027052879, + 0.006714430637657642, + 0.00554464990273118, + 0.0033761586528271437, + 0.007623402867466211, + 0.003115040482953191, + 0.003353139152750373, + 0.0038587055169045925, + 0.004559697583317757, + 0.004489519167691469, + 0.003052247455343604, + 0.010254982858896255, + 0.07706394791603088, + 0.0028526494279503822, + 0.003697074018418789, + 0.013417316600680351, + 0.0751061737537384, + 0.0026711318641901016, + 0.003636955749243498, + 0.006362255197018385, + 0.04609113186597824, + 0.006260951980948448, + 0.005359718110412359, + 0.008419910445809364, + 0.04568249359726906, + 0.003171690506860614, + 0.004562077112495899, + 0.009094043634831905, + 0.06137344241142273, + 0.005900688003748655, + 0.004577615298330784, + 0.009954490698873997, + 0.05662064254283905, + 0.003225240157917142, + 0.004446091130375862, + 0.00314935389906168, + 0.00988597422838211, + 0.00494424207136035, + 0.004024274181574583, + 0.004909606650471687, + 0.007016641087830067, + 0.004939625971019268, + 0.0046243807300925255, + 0.007222000043839216, + 0.02402542158961296, + 0.014641446992754936, + 0.015301854349672794, + 0.007674247957766056, + 0.026200633496046066, + 0.007548701949417591, + 0.012930751778185368, + 0.001634195214137435, + 0.007952725514769554, + 0.0019126617116853595, + 0.0025180757511407137, + 0.0020876831840723753, + 0.008519243448972702, + 0.0015094229020178318, + 0.004302782006561756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.026343757286667824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_tensor([214, 214])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_tensor([214, 214])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9c3a46fa1c486d4f324ec232e8024391965978ec --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_tensor([214, 214])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, forward tracking shot begins, moving steadily at eye level toward the massive, ancient stone arch that dominates the frame. Constructed of massive, weathered blocks of gray granite, its rough surface covered in dark green moss and creepers, intricate, faded carvings of forgotten symbols can be seen along its curved summit. As the camera silently glides through the cool, shadowy passage beneath the arch, the frame is briefly shrouded in deep, cool shadows before emerging in a breathtaking, panoramic, extreme wide shot, revealing a vast, sun-drenched valley. A pristine, emerald-green river meanders through lush, emerald-green meadows dotted with vibrant yellow and purple wildflowers. Vast tracts of dense, dark pine forests climb the gentle slopes of the surrounding hills, which recede into a distant, hazy mountain range with snow-capped peaks. The entire landscape is bathed in the warm golden glow of the afternoon sun, casting long, soft shadows that make the lake shimmer and creating an atmosphere of profound tranquility and epic proportions beneath a vast, cloudless, azure sky.", + "seed": 11698, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_214_seed_11698_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033813066128641364, + "vlm_logprob": -0.033813066128641364, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.033813066128641364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032362740486860275 + ], + "opposite_logprob_tokens": [ + -4.032362937927246 + ], + "logprob_reward": -0.032362740486860275, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.032362740486860275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18850670754909515, + 0.31107109785079956, + 0.10354528576135635, + 0.06395361572504044, + 0.09140501916408539, + 0.033749546855688095, + 0.07783148437738419, + 0.04577457159757614, + 0.03511451184749603, + 0.04904811456799507 + ], + "credit_map": [ + 0.01296902447938919, + 0.01866881549358368, + 0.029491186141967773, + 0.024547569453716278, + 0.018026962876319885, + 0.026095015928149223, + 0.027689537033438683, + 0.031018607318401337, + 0.014823511242866516, + 0.03079834394156933, + 0.06849675625562668, + 0.019054843112826347, + 0.06194228306412697, + 0.08192148804664612, + 0.014255974441766739, + 0.019777905195951462, + 0.013678301125764847, + 0.013249865733087063, + 0.01586264558136463, + 0.012258828617632389, + 0.010626443661749363, + 0.014432095922529697, + 0.012453808449208736, + 0.010983304120600224, + 0.007799831684678793, + 0.006869746837764978, + 0.007988812401890755, + 0.009345942176878452, + 0.011416943743824959, + 0.011686091311275959, + 0.005287299398332834, + 0.0035589567851275206, + 0.006909466814249754, + 0.015208815224468708, + 0.018923094496130943, + 0.015329812653362751, + 0.008180069737136364, + 0.011318334378302097, + 0.0066962698474526405, + 0.008839157409965992, + 0.0032508710864931345, + 0.007378174923360348, + 0.0056763640604913235, + 0.007579613011330366, + 0.002077886601909995, + 0.0018846028251573443, + 0.0024451708886772394, + 0.003456868464127183, + 0.015497826971113682, + 0.007742635905742645, + 0.013899329118430614, + 0.011561521328985691, + 0.003960954491049051, + 0.006638703402131796, + 0.005881950259208679, + 0.012648568488657475, + 0.004343830514699221, + 0.0036672207061201334, + 0.00521059101447463, + 0.005139742977917194, + 0.004250518511980772, + 0.006326931994408369, + 0.00982925109565258, + 0.007006490137428045, + 0.003183179534971714, + 0.004339071922004223, + 0.008049974218010902, + 0.0029566825833171606, + 0.003474303986877203, + 0.004371696151793003, + 0.003909298684448004, + 0.00483030965551734, + 0.007720584515482187, + 0.006776062771677971, + 0.004061987157911062, + 0.004060694482177496, + 0.00632472475990653, + 0.007494774181395769, + 0.007262714672833681, + 0.0053465780802071095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.032362740486860275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0891047865152359 + ], + "opposite_logprob_tokens": [ + -2.4641048908233643 + ], + "logprob_reward": -0.0891047865152359, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0891047865152359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15564709901809692, + 0.45956918597221375, + 0.10446283221244812, + 0.048766665160655975, + 0.06418096274137497, + 0.023652249947190285, + 0.04933315888047218, + 0.03436800092458725, + 0.02401568368077278, + 0.036004144698381424 + ], + "credit_map": [ + 0.012408004142343998, + 0.017851144075393677, + 0.021965818479657173, + 0.020365063101053238, + 0.022200394421815872, + 0.01797727681696415, + 0.019953252747654915, + 0.02292613312602043, + 0.02512342669069767, + 0.0411309152841568, + 0.09508267790079117, + 0.030953453853726387, + 0.09002269059419632, + 0.12422303110361099, + 0.021684911102056503, + 0.0313480868935585, + 0.011296031065285206, + 0.01682843454182148, + 0.01894652098417282, + 0.011329950764775276, + 0.010739462450146675, + 0.014054221101105213, + 0.009774797596037388, + 0.011493419297039509, + 0.005518198013305664, + 0.006135425064712763, + 0.0043911500833928585, + 0.007440587040036917, + 0.008015584200620651, + 0.009323897771537304, + 0.004639089107513428, + 0.0033027357421815395, + 0.007855195552110672, + 0.01026515569537878, + 0.009074537083506584, + 0.01027136854827404, + 0.0068866233341395855, + 0.009165672585368156, + 0.003950117155909538, + 0.006712295580655336, + 0.0033083464950323105, + 0.005412161350250244, + 0.003800294827669859, + 0.003528862725943327, + 0.0013844892382621765, + 0.0016321028815582395, + 0.002132558962330222, + 0.002453435445204377, + 0.009072192013263702, + 0.0051482440903782845, + 0.007702144328504801, + 0.007443499285727739, + 0.0029516846407204866, + 0.004511124454438686, + 0.0051283217035233974, + 0.007375949062407017, + 0.0033013049978762865, + 0.003670970443636179, + 0.00413002073764801, + 0.00587838189676404, + 0.003157360479235649, + 0.003994320053607225, + 0.0050132847391068935, + 0.005222358740866184, + 0.0029809456318616867, + 0.0038068636786192656, + 0.00439281715080142, + 0.0019671949557960033, + 0.002694438910111785, + 0.002248214092105627, + 0.0027262435760349035, + 0.003198966383934021, + 0.005007946398109198, + 0.0062662335112690926, + 0.0022827419452369213, + 0.0034807706251740456, + 0.004628004506230354, + 0.004345635883510113, + 0.006424478720873594, + 0.003568333340808749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0891047865152359 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041068983264267445 + ], + "opposite_logprob_tokens": [ + -5.5041069984436035 + ], + "logprob_reward": -0.0041068983264267445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041068983264267445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16816674172878265, + 0.15448150038719177, + 0.103246308863163, + 0.19992050528526306, + 0.0869782343506813, + 0.10427911579608917, + 0.07969964295625687, + 0.04162225499749184, + 0.02692514844238758, + 0.0346805602312088 + ], + "credit_map": [ + 0.01478960458189249, + 0.021721279248595238, + 0.024006107822060585, + 0.02184506133198738, + 0.013671367429196835, + 0.019932696595788002, + 0.03135772421956062, + 0.02084289863705635, + 0.008225827477872372, + 0.01599232107400894, + 0.03465557098388672, + 0.010618502274155617, + 0.022379402071237564, + 0.03452320396900177, + 0.012185130268335342, + 0.015901530161499977, + 0.013122809119522572, + 0.01481130812317133, + 0.015610086731612682, + 0.011068053543567657, + 0.008841819129884243, + 0.012552609667181969, + 0.014996754936873913, + 0.012242866680026054, + 0.009235686622560024, + 0.008277717046439648, + 0.006983558181673288, + 0.013924446888267994, + 0.07774172723293304, + 0.07091958075761795, + 0.007257037330418825, + 0.005580758675932884, + 0.008321351371705532, + 0.011545504443347454, + 0.01405428722500801, + 0.01728280447423458, + 0.0065859537571668625, + 0.010059734806418419, + 0.00805631559342146, + 0.011072279885411263, + 0.01124632265418768, + 0.02158072404563427, + 0.017580680549144745, + 0.02738623507320881, + 0.006542320363223553, + 0.005523798521608114, + 0.0067324270494282246, + 0.00768660381436348, + 0.016302848234772682, + 0.007371447514742613, + 0.013014841824769974, + 0.007915305905044079, + 0.0035513099282979965, + 0.007441636174917221, + 0.008433283306658268, + 0.015668969601392746, + 0.003664212068542838, + 0.004536585882306099, + 0.004070654045790434, + 0.0026089001912623644, + 0.004740480333566666, + 0.007866173051297665, + 0.005605749785900116, + 0.008529498241841793, + 0.0019867352675646544, + 0.0047959997318685055, + 0.003150202101096511, + 0.0013746784534305334, + 0.0035365282092243433, + 0.006146501284092665, + 0.0030785214621573687, + 0.0028559817001223564, + 0.004552244674414396, + 0.004742278717458248, + 0.0031485874205827713, + 0.0032787092495709658, + 0.004655288532376289, + 0.003082362236455083, + 0.005930664949119091, + 0.0052904230542480946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0041068983264267445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978577509522438 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.00978577509522438, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978577509522438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0927102342247963, + 0.13416320085525513, + 0.05652802065014839, + 0.09800940006971359, + 0.1048274114727974, + 0.029502619057893753, + 0.1410590559244156, + 0.09696149080991745, + 0.18148618936538696, + 0.06475242227315903 + ], + "credit_map": [ + 0.006861571222543716, + 0.008106091059744358, + 0.015198145061731339, + 0.011484207585453987, + 0.007337327580899, + 0.011219973675906658, + 0.015184911899268627, + 0.017318010330200195, + 0.007105011958628893, + 0.010798883624374866, + 0.029885325580835342, + 0.007815927267074585, + 0.024980925023555756, + 0.03895339369773865, + 0.006148600485175848, + 0.008475136011838913, + 0.006354355253279209, + 0.0074610295705497265, + 0.010827494785189629, + 0.006144660525023937, + 0.004338450264185667, + 0.008284639567136765, + 0.006614114157855511, + 0.006503279320895672, + 0.01081959530711174, + 0.008812670595943928, + 0.008230343461036682, + 0.01524308044463396, + 0.01998281478881836, + 0.020933421328663826, + 0.007310995366424322, + 0.006676478311419487, + 0.01152073498815298, + 0.015595306642353535, + 0.012742855586111546, + 0.01954500563442707, + 0.011010744608938694, + 0.014973816461861134, + 0.0075587984174489975, + 0.011880150996148586, + 0.0034631690941751003, + 0.006045416928827763, + 0.004050861112773418, + 0.004481129348278046, + 0.002255475614219904, + 0.002519640140235424, + 0.002812162274494767, + 0.0038747633807361126, + 0.01851056143641472, + 0.00970282033085823, + 0.019714370369911194, + 0.011343889869749546, + 0.009480214677751064, + 0.01881292462348938, + 0.01617194339632988, + 0.03732232749462128, + 0.007697336375713348, + 0.004541567526757717, + 0.009587508626282215, + 0.005436936859041452, + 0.009864418767392635, + 0.0105991680175066, + 0.015594538301229477, + 0.03364001587033272, + 0.017601842060685158, + 0.020709237083792686, + 0.041613657027482986, + 0.01631159707903862, + 0.022733405232429504, + 0.019755154848098755, + 0.016070866957306862, + 0.026690427213907242, + 0.00826683547347784, + 0.01139522809535265, + 0.0055915480479598045, + 0.006674089934676886, + 0.010622511617839336, + 0.007079096976667643, + 0.007741358131170273, + 0.007381750270724297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.00978577509522438 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033705130219459534 + ], + "opposite_logprob_tokens": [ + -3.408705234527588 + ], + "logprob_reward": -0.033705130219459534, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033705130219459534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10863528400659561, + 0.09403175115585327, + 0.06454509496688843, + 0.29308339953422546, + 0.043770015239715576, + 0.0575738288462162, + 0.0998535081744194, + 0.0571429543197155, + 0.048959821462631226, + 0.13240434229373932 + ], + "credit_map": [ + 0.008546516299247742, + 0.00850163958966732, + 0.017293045297265053, + 0.014727042987942696, + 0.008263010531663895, + 0.016552574932575226, + 0.015120083466172218, + 0.01963137835264206, + 0.005767697934061289, + 0.009081280790269375, + 0.01827187091112137, + 0.009981496259570122, + 0.015145068056881428, + 0.015809115022420883, + 0.006517739035189152, + 0.013457484543323517, + 0.0070124659687280655, + 0.007301583886146545, + 0.007099360693246126, + 0.00576403271406889, + 0.006099049933254719, + 0.010736411437392235, + 0.012084515765309334, + 0.008447671309113503, + 0.014266695827245712, + 0.020251305773854256, + 0.011722954921424389, + 0.020645014941692352, + 0.09868934750556946, + 0.09828831255435944, + 0.017414743080735207, + 0.011805020272731781, + 0.005042305216193199, + 0.006772756576538086, + 0.006188912782818079, + 0.006823497824370861, + 0.003912808373570442, + 0.0065498375333845615, + 0.00388862076215446, + 0.004591276869177818, + 0.005869025830179453, + 0.01164303533732891, + 0.008875682950019836, + 0.01219173427671194, + 0.003459746250882745, + 0.0037482238840311766, + 0.005100067239254713, + 0.0066863130778074265, + 0.015977555885910988, + 0.007599985226988792, + 0.014566264115273952, + 0.010516034439206123, + 0.006614976562559605, + 0.013395297341048717, + 0.010949678719043732, + 0.020233726128935814, + 0.005367682781070471, + 0.005442936439067125, + 0.0068250237964093685, + 0.01013893075287342, + 0.005059865768998861, + 0.006587932351976633, + 0.006885226350277662, + 0.010835357941687107, + 0.004093213006854057, + 0.00816468894481659, + 0.0074494085274636745, + 0.004271811805665493, + 0.006431609392166138, + 0.0061343153938651085, + 0.004957777447998524, + 0.007456998340785503, + 0.019538700580596924, + 0.031751301139593124, + 0.008579185232520103, + 0.010913936421275139, + 0.017416460439562798, + 0.018095582723617554, + 0.014539027586579323, + 0.011570150963962078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.033705130219459534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, forward tracking shot begins, moving steadily at eye level toward the massive, ancient stone arch that dominates the frame. Constructed of massive, weathered blocks of gray granite, its rough surface covered in dark green moss and creepers, intricate, faded carvings of forgotten symbols can be seen along its curved summit. As the camera silently glides through the cool, shadowy passage beneath the arch, the frame is briefly shrouded in deep, cool shadows before emerging in a breathtaking, panoramic, extreme wide shot, revealing a vast, sun-drenched valley. A pristine, emerald-green river meanders through lush, emerald-green meadows dotted with vibrant yellow and purple wildflowers. Vast tracts of dense, dark pine forests climb the gentle slopes of the surrounding hills, which recede into a distant, hazy mountain range with snow-capped peaks. The entire landscape is bathed in the warm golden glow of the afternoon sun, casting long, soft shadows that make the lake shimmer and creating an atmosphere of profound tranquility and epic proportions beneath a vast, cloudless, azure sky.", + "seed": 11699, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_214_seed_11699_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.063371289242059, + "vlm_logprob": -0.063371289242059, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.063371289242059, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03611205145716667 + ], + "opposite_logprob_tokens": [ + -3.661112070083618 + ], + "logprob_reward": -0.03611205145716667, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03611205145716667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14973075687885284, + 0.08140609413385391, + 0.08649799972772598, + 0.07648894935846329, + 0.1700495183467865, + 0.10190339386463165, + 0.07729386538267136, + 0.11843731254339218, + 0.08109357953071594, + 0.057098593562841415 + ], + "credit_map": [ + 0.021839449182152748, + 0.020407428964972496, + 0.021236272528767586, + 0.01652025803923607, + 0.010690308175981045, + 0.025154629722237587, + 0.015794647857546806, + 0.01808774843811989, + 0.009401481598615646, + 0.01482328586280346, + 0.01150398701429367, + 0.013345948420464993, + 0.00832234788686037, + 0.01280310656875372, + 0.006378366611897945, + 0.004827562253922224, + 0.010398932732641697, + 0.013004041276872158, + 0.011590268462896347, + 0.013630920089781284, + 0.010155386291444302, + 0.011081594042479992, + 0.009209265932440758, + 0.007427585311233997, + 0.009530599229037762, + 0.012961618602275848, + 0.012237699702382088, + 0.009851738810539246, + 0.007231947500258684, + 0.009048838168382645, + 0.00776314502581954, + 0.007863359525799751, + 0.02647421695291996, + 0.017618762329220772, + 0.02203160710632801, + 0.021695183590054512, + 0.02017105370759964, + 0.01906818337738514, + 0.023512806743383408, + 0.019477704539895058, + 0.017570897936820984, + 0.009664124809205532, + 0.016896452754735947, + 0.012493783608078957, + 0.010270473547279835, + 0.008293811231851578, + 0.014810612425208092, + 0.01190322544425726, + 0.021599439904093742, + 0.006728834472596645, + 0.014409713447093964, + 0.008203743025660515, + 0.00402152631431818, + 0.005093840882182121, + 0.008483667857944965, + 0.008753097616136074, + 0.0238819420337677, + 0.01155106071382761, + 0.036519233137369156, + 0.011222689412534237, + 0.006465327925980091, + 0.008767399936914444, + 0.009224919602274895, + 0.010804736986756325, + 0.015706198289990425, + 0.0064880335703492165, + 0.022870058193802834, + 0.00606262544170022, + 0.008939520455896854, + 0.006621151231229305, + 0.006814361549913883, + 0.007591621018946171, + 0.00788276270031929, + 0.007812146097421646, + 0.006531867664307356, + 0.005248840898275375, + 0.008985788561403751, + 0.011356929317116737, + 0.004152644891291857, + 0.005127607844769955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.03611205145716667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992720007896423 + ], + "opposite_logprob_tokens": [ + -1.8049272298812866 + ], + "logprob_reward": -0.17992720007896423, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992720007896423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11024727672338486, + 0.1042594239115715, + 0.08927064388990402, + 0.2365047037601471, + 0.08745643496513367, + 0.030085792765021324, + 0.07619477063417435, + 0.13099734485149384, + 0.10676036775112152, + 0.02822326496243477 + ], + "credit_map": [ + 0.016761401668190956, + 0.018991151824593544, + 0.01569243147969246, + 0.013377024792134762, + 0.011436721310019493, + 0.015151661820709705, + 0.008527442812919617, + 0.010309438221156597, + 0.010508856736123562, + 0.02374303713440895, + 0.015154119580984116, + 0.011714191175997257, + 0.008559989742934704, + 0.016989652067422867, + 0.00745050422847271, + 0.010139070451259613, + 0.011794939637184143, + 0.01606503129005432, + 0.012291881255805492, + 0.010026431642472744, + 0.009310062043368816, + 0.01650926284492016, + 0.007409767713397741, + 0.005863271187990904, + 0.03633009269833565, + 0.048401884734630585, + 0.04296034947037697, + 0.019970400258898735, + 0.02264079637825489, + 0.02575157769024372, + 0.026493025943636894, + 0.013956579379737377, + 0.010881587862968445, + 0.010304181836545467, + 0.01518544927239418, + 0.00994714256376028, + 0.009429353289306164, + 0.009715456515550613, + 0.012369805946946144, + 0.009623457677662373, + 0.006041263230144978, + 0.0025125653482973576, + 0.004282737150788307, + 0.0028808899223804474, + 0.0029452801682054996, + 0.00231513986364007, + 0.004505218472331762, + 0.004602697677910328, + 0.01628216728568077, + 0.008280578069388866, + 0.011515301652252674, + 0.007490833289921284, + 0.005534558091312647, + 0.008650511503219604, + 0.00899224542081356, + 0.009448569267988205, + 0.0290624238550663, + 0.011963313445448875, + 0.04257291555404663, + 0.010765064507722855, + 0.00505002262070775, + 0.01143154501914978, + 0.009304438717663288, + 0.010847631841897964, + 0.01901526004076004, + 0.010779340751469135, + 0.02919863723218441, + 0.008636433631181717, + 0.011188220232725143, + 0.010932747274637222, + 0.007781030144542456, + 0.009228700771927834, + 0.004245383199304342, + 0.003485877765342593, + 0.003900542389601469, + 0.003290108172222972, + 0.003456281730905175, + 0.003955460619181395, + 0.0030948591884225607, + 0.0027947514317929745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.17992720007896423 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763498293235898 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.00763498293235898, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763498293235898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1756572425365448, + 0.11362200975418091, + 0.0732942521572113, + 0.1013142317533493, + 0.25814202427864075, + 0.048101864755153656, + 0.07354686409235, + 0.061541419476270676, + 0.04087292402982712, + 0.053907208144664764 + ], + "credit_map": [ + 0.026783568784594536, + 0.021730920299887657, + 0.025127803906798363, + 0.02076687105000019, + 0.014095488004386425, + 0.03315919637680054, + 0.01690218597650528, + 0.017091207206249237, + 0.010296150110661983, + 0.019119523465633392, + 0.01438151579350233, + 0.019692564383149147, + 0.016181008890271187, + 0.015101396478712559, + 0.012188724242150784, + 0.0066611324436962605, + 0.00985937099903822, + 0.01438526064157486, + 0.011201038025319576, + 0.010374647565186024, + 0.007148868404328823, + 0.010041764937341213, + 0.005992278456687927, + 0.0042910268530249596, + 0.013203732669353485, + 0.01614849828183651, + 0.016291668638586998, + 0.013158481568098068, + 0.010311121121048927, + 0.011277016252279282, + 0.011786073446273804, + 0.009137636050581932, + 0.03212250769138336, + 0.02787371166050434, + 0.039622388780117035, + 0.02895393781363964, + 0.029160907492041588, + 0.03398878872394562, + 0.03744713217020035, + 0.028972657397389412, + 0.006460931617766619, + 0.004813577979803085, + 0.0074501633644104, + 0.0057018715888261795, + 0.004805121570825577, + 0.00479981629177928, + 0.007105145137757063, + 0.00696523254737258, + 0.014060941524803638, + 0.007703261915594339, + 0.010685299523174763, + 0.009007735177874565, + 0.004988303408026695, + 0.0057464055716991425, + 0.011113431304693222, + 0.010241487063467503, + 0.011690737679600716, + 0.0053833965212106705, + 0.014384019188582897, + 0.006326646544039249, + 0.003650369355455041, + 0.0057746474631130695, + 0.006665420718491077, + 0.007666182238608599, + 0.005638417787849903, + 0.004138869233429432, + 0.0075511871837079525, + 0.0033670607954263687, + 0.004502754658460617, + 0.0055231391452252865, + 0.004509514197707176, + 0.005641980562359095, + 0.008972756564617157, + 0.005698396824300289, + 0.0075311665423214436, + 0.007264346349984407, + 0.006279779598116875, + 0.006011603865772486, + 0.005698853638023138, + 0.006450302433222532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.00763498293235898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200588688254356 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200588688254356, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200588688254356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09257137775421143, + 0.12505578994750977, + 0.11046071350574493, + 0.1594250202178955, + 0.0800565779209137, + 0.06212737411260605, + 0.0701904371380806, + 0.11267627030611038, + 0.10934542864561081, + 0.07809102535247803 + ], + "credit_map": [ + 0.012403802014887333, + 0.013092832639813423, + 0.013813276775181293, + 0.01021897979080677, + 0.00858289934694767, + 0.01562399324029684, + 0.00880431942641735, + 0.010031274519860744, + 0.011501124128699303, + 0.021663768216967583, + 0.01958795078098774, + 0.017384734004735947, + 0.007035950198769569, + 0.022711150348186493, + 0.00814095325767994, + 0.017030149698257446, + 0.012177783995866776, + 0.01879982277750969, + 0.01563776656985283, + 0.010835385881364346, + 0.014800732024013996, + 0.016469575464725494, + 0.012148212641477585, + 0.009591438807547092, + 0.018440864980220795, + 0.02300599031150341, + 0.0225667804479599, + 0.018414543941617012, + 0.017751488834619522, + 0.017857156693935394, + 0.026444504037499428, + 0.014943689107894897, + 0.009535140357911587, + 0.010140182450413704, + 0.013188033364713192, + 0.01043744571506977, + 0.006233204156160355, + 0.009355242364108562, + 0.012519563548266888, + 0.008647767826914787, + 0.011381899937987328, + 0.005649460945278406, + 0.009029858745634556, + 0.008817339316010475, + 0.006556213833391666, + 0.004965795204043388, + 0.008668538182973862, + 0.007058264687657356, + 0.010943994857370853, + 0.010092446580529213, + 0.009938366711139679, + 0.007814436219632626, + 0.004748992621898651, + 0.006869776174426079, + 0.00789372343569994, + 0.01188870333135128, + 0.023235032334923744, + 0.008523783646523952, + 0.03430435061454773, + 0.009139160625636578, + 0.004858570639044046, + 0.008250490762293339, + 0.01412193849682808, + 0.010242939926683903, + 0.014951478689908981, + 0.009411956183612347, + 0.018774673342704773, + 0.00717806164175272, + 0.011438785120844841, + 0.018411759287118912, + 0.013469035737216473, + 0.015709683299064636, + 0.011984717100858688, + 0.009860944002866745, + 0.009530790150165558, + 0.006698769051581621, + 0.010769794695079327, + 0.01398078165948391, + 0.006782170385122299, + 0.008483060635626316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.014200588688254356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898162305355072 + ], + "opposite_logprob_tokens": [ + -2.578981637954712 + ], + "logprob_reward": -0.07898162305355072, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898162305355072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21804089844226837, + 0.10289505869150162, + 0.0714336484670639, + 0.11735755950212479, + 0.07165978848934174, + 0.07433319836854935, + 0.07013766467571259, + 0.09644830971956253, + 0.08798848092556, + 0.08970537781715393 + ], + "credit_map": [ + 0.027692463248968124, + 0.031051049008965492, + 0.0333317294716835, + 0.024490242823958397, + 0.014933647587895393, + 0.04597776383161545, + 0.01744367554783821, + 0.023120321333408356, + 0.013053429313004017, + 0.012552333995699883, + 0.012608883902430534, + 0.019353309646248817, + 0.014858107082545757, + 0.014289764687418938, + 0.00842397753149271, + 0.007755252066999674, + 0.008495626039803028, + 0.011552834883332253, + 0.01029475312680006, + 0.010225965641438961, + 0.008122945204377174, + 0.008979558013379574, + 0.006397812161594629, + 0.007364151068031788, + 0.016060061752796173, + 0.014975816011428833, + 0.01825122907757759, + 0.016088079661130905, + 0.009935644455254078, + 0.014589188620448112, + 0.013487353920936584, + 0.013970191590487957, + 0.008677443489432335, + 0.008828600868582726, + 0.012232635170221329, + 0.00976667832583189, + 0.006522631738334894, + 0.00749154994264245, + 0.011252490803599358, + 0.00688775721937418, + 0.010070376098155975, + 0.007074839901179075, + 0.01280282624065876, + 0.008868257515132427, + 0.00804852694272995, + 0.007599939126521349, + 0.009946215897798538, + 0.009922215715050697, + 0.012960324995219707, + 0.009648151695728302, + 0.010976327583193779, + 0.007853304967284203, + 0.005365025717765093, + 0.0061646089889109135, + 0.006868969649076462, + 0.010300946421921253, + 0.020300934091210365, + 0.007888570427894592, + 0.029022764414548874, + 0.007971457205712795, + 0.004121418111026287, + 0.009540053084492683, + 0.009358124807476997, + 0.008244981057941914, + 0.012325163930654526, + 0.007505503948777914, + 0.016961945220828056, + 0.005895060487091541, + 0.007610606960952282, + 0.011192666366696358, + 0.01212554331868887, + 0.014371990226209164, + 0.011863618157804012, + 0.011467942036688328, + 0.011635053902864456, + 0.010201412253081799, + 0.010223450139164925, + 0.011609106324613094, + 0.01162692066282034, + 0.01107787061482668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.07898162305355072 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, forward tracking shot begins, moving steadily at eye level toward the massive, ancient stone arch that dominates the frame. Constructed of massive, weathered blocks of gray granite, its rough surface covered in dark green moss and creepers, intricate, faded carvings of forgotten symbols can be seen along its curved summit. As the camera silently glides through the cool, shadowy passage beneath the arch, the frame is briefly shrouded in deep, cool shadows before emerging in a breathtaking, panoramic, extreme wide shot, revealing a vast, sun-drenched valley. A pristine, emerald-green river meanders through lush, emerald-green meadows dotted with vibrant yellow and purple wildflowers. Vast tracts of dense, dark pine forests climb the gentle slopes of the surrounding hills, which recede into a distant, hazy mountain range with snow-capped peaks. The entire landscape is bathed in the warm golden glow of the afternoon sun, casting long, soft shadows that make the lake shimmer and creating an atmosphere of profound tranquility and epic proportions beneath a vast, cloudless, azure sky.", + "seed": 11700, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_214_seed_11700_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1380546586588025, + "vlm_logprob": -0.1380546586588025, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.1380546586588025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35443115234375 + ], + "opposite_logprob_tokens": [ + -1.22943115234375 + ], + "logprob_reward": -0.35443115234375, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35443115234375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2651855945587158, + 0.2281738519668579, + 0.10884084552526474, + 0.08037541806697845, + 0.04946949705481529, + 0.08243777602910995, + 0.09516242146492004, + 0.0541122741997242, + 0.01588580571115017, + 0.02035650797188282 + ], + "credit_map": [ + 0.03300555422902107, + 0.021545127034187317, + 0.038459017872810364, + 0.014239994809031487, + 0.06850831210613251, + 0.021628079935908318, + 0.014253354631364346, + 0.05354614555835724, + 0.031963225454092026, + 0.01191941648721695, + 0.04369282349944115, + 0.012750335969030857, + 0.014341138303279877, + 0.05494223162531853, + 0.011243331246078014, + 0.047321345657110214, + 0.015537148341536522, + 0.007016005460172892, + 0.01205823291093111, + 0.012588884681463242, + 0.006728010717779398, + 0.01937703602015972, + 0.01496116816997528, + 0.020574353635311127, + 0.010953634046018124, + 0.005121727008372545, + 0.013543305918574333, + 0.014003114774823189, + 0.005649851635098457, + 0.010753253474831581, + 0.009468551725149155, + 0.010881975293159485, + 0.00512557290494442, + 0.004581949673593044, + 0.00640131114050746, + 0.007141058798879385, + 0.005459561478346586, + 0.00644873408600688, + 0.006397441029548645, + 0.007913868874311447, + 0.016467677429318428, + 0.006341126747429371, + 0.008635005913674831, + 0.01674535684287548, + 0.005524320062249899, + 0.004748502746224403, + 0.009497757069766521, + 0.014478032477200031, + 0.008046605624258518, + 0.007887995801866055, + 0.014711673371493816, + 0.014570336788892746, + 0.012261084280908108, + 0.009553845040500164, + 0.013500206172466278, + 0.014630675315856934, + 0.006195759866386652, + 0.006259117275476456, + 0.009300860576331615, + 0.008212539367377758, + 0.006414261646568775, + 0.004174445290118456, + 0.005109729710966349, + 0.008445559069514275, + 0.0014513052301481366, + 0.001817429787479341, + 0.0024231858551502228, + 0.0021367287263274193, + 0.0011267111403867602, + 0.0009635472088120878, + 0.003153091762214899, + 0.0028138072229921818, + 0.0021604301873594522, + 0.0027190607506781816, + 0.003745415946468711, + 0.0034199163783341646, + 0.0021854082588106394, + 0.0018603960052132607, + 0.0022380154114216566, + 0.002027865033596754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.35443115234375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25223180651664734 + ], + "opposite_logprob_tokens": [ + -1.5022318363189697 + ], + "logprob_reward": -0.25223180651664734, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25223180651664734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32163599133491516, + 0.2289007604122162, + 0.13057628273963928, + 0.06907938420772552, + 0.03335185721516609, + 0.05620977655053139, + 0.07354730367660522, + 0.04005812108516693, + 0.019550127908587456, + 0.02709035947918892 + ], + "credit_map": [ + 0.03905557841062546, + 0.02992587350308895, + 0.046187855303287506, + 0.017435038462281227, + 0.0885503739118576, + 0.030418403446674347, + 0.01725168712437153, + 0.05281119793653488, + 0.03251262009143829, + 0.01160451304167509, + 0.04083061218261719, + 0.014026544988155365, + 0.013745383359491825, + 0.06373696774244308, + 0.009717142209410667, + 0.04272698983550072, + 0.02076892927289009, + 0.00905729178339243, + 0.014862112700939178, + 0.015857789665460587, + 0.00885708350688219, + 0.023280970752239227, + 0.016292542219161987, + 0.021599577739834785, + 0.009502077475190163, + 0.006291689816862345, + 0.010895895771682262, + 0.009210123680531979, + 0.006047986913472414, + 0.011914083734154701, + 0.005080150905996561, + 0.01013738289475441, + 0.0034705118741840124, + 0.0028927510138601065, + 0.0049791838973760605, + 0.004383554216474295, + 0.0039784568361938, + 0.004355936776846647, + 0.0041439589112997055, + 0.005147506017237902, + 0.004023162182420492, + 0.0037504173815250397, + 0.006146343424916267, + 0.016685374081134796, + 0.00345295830629766, + 0.004971791058778763, + 0.009042571298778057, + 0.008137163706123829, + 0.00634417962282896, + 0.0061735170893371105, + 0.011497407220304012, + 0.011788595467805862, + 0.009433009661734104, + 0.00749571667984128, + 0.00991697609424591, + 0.010897907428443432, + 0.004563109949231148, + 0.004520618822425604, + 0.007567161228507757, + 0.006431062705814838, + 0.004402920603752136, + 0.003300753654912114, + 0.003652590326964855, + 0.005619905423372984, + 0.0013864390784874558, + 0.0010276883840560913, + 0.002808663295581937, + 0.0024661612696945667, + 0.0013965265825390816, + 0.0009326401632279158, + 0.00495279161259532, + 0.004579219501465559, + 0.002949378453195095, + 0.003953380975872278, + 0.0050473385490477085, + 0.004234367981553078, + 0.0032167695462703705, + 0.00274960370734334, + 0.002479088492691517, + 0.0024604322388768196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.25223180651664734 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255748514086008 + ], + "opposite_logprob_tokens": [ + -4.387557506561279 + ], + "logprob_reward": -0.01255748514086008, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255748514086008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21783792972564697, + 0.20124174654483795, + 0.05141588672995567, + 0.06209650635719299, + 0.07404633611440659, + 0.1138545349240303, + 0.110896997153759, + 0.09906765818595886, + 0.03229161351919174, + 0.037250816822052 + ], + "credit_map": [ + 0.02542686089873314, + 0.01696779392659664, + 0.032524224370718, + 0.013523649424314499, + 0.05644147843122482, + 0.015449808910489082, + 0.01183544285595417, + 0.04566869139671326, + 0.02774979919195175, + 0.008605841547250748, + 0.031435512006282806, + 0.01088652666658163, + 0.00855419784784317, + 0.06182113662362099, + 0.008821375668048859, + 0.043367356061935425, + 0.008645869791507721, + 0.003454555757343769, + 0.005655934102833271, + 0.005754152312874794, + 0.0033385350834578276, + 0.009028390981256962, + 0.005875444505363703, + 0.009663008153438568, + 0.005117384716868401, + 0.0032892453018575907, + 0.007195454090833664, + 0.015001291409134865, + 0.0036283545196056366, + 0.011351647786796093, + 0.007340955547988415, + 0.009172176942229271, + 0.006156277842819691, + 0.006199518218636513, + 0.00971672497689724, + 0.01098901592195034, + 0.00872888881713152, + 0.009822738356888294, + 0.010008106008172035, + 0.01242506317794323, + 0.025646381080150604, + 0.00884690135717392, + 0.01166557613760233, + 0.01940210536122322, + 0.007229642942547798, + 0.0076576704159379005, + 0.011552720330655575, + 0.02185353823006153, + 0.008725288324058056, + 0.008939146064221859, + 0.0187593474984169, + 0.015665406361222267, + 0.013859821483492851, + 0.010703637264668941, + 0.015895862132310867, + 0.018348487094044685, + 0.009835504926741123, + 0.008101869374513626, + 0.018353171646595, + 0.013985215686261654, + 0.010394081473350525, + 0.008779244497418404, + 0.013973692432045937, + 0.01564488373696804, + 0.002991538494825363, + 0.003636846551671624, + 0.004741151817142963, + 0.004236558917909861, + 0.0025854797568172216, + 0.002607734641060233, + 0.006129848305135965, + 0.005362453404814005, + 0.004147191997617483, + 0.004719702061265707, + 0.007730350363999605, + 0.00689186155796051, + 0.003571367124095559, + 0.002698138589039445, + 0.003756958991289139, + 0.003735243808478117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.01255748514086008 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078807413578033 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016078807413578033, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078807413578033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059423238039016724, + 0.08805649727582932, + 0.04718286171555519, + 0.11903800815343857, + 0.09646910429000854, + 0.21304599940776825, + 0.1944786012172699, + 0.04442242160439491, + 0.026993747800588608, + 0.11088953912258148 + ], + "credit_map": [ + 0.007740058470517397, + 0.004634220153093338, + 0.013538516126573086, + 0.006439730525016785, + 0.007876106537878513, + 0.006084835156798363, + 0.004917807411402464, + 0.008191959001123905, + 0.011911694891750813, + 0.004197738599032164, + 0.014655998907983303, + 0.00676856329664588, + 0.005517894402146339, + 0.023177599534392357, + 0.005080766975879669, + 0.016746236011385918, + 0.007977425120770931, + 0.003589491592720151, + 0.005952777806669474, + 0.003759410697966814, + 0.004317354876548052, + 0.007675048895180225, + 0.003718384774401784, + 0.010192971676588058, + 0.008541077375411987, + 0.009523076005280018, + 0.015090934000909328, + 0.015451200306415558, + 0.01322074607014656, + 0.023414064198732376, + 0.01643483340740204, + 0.017362086102366447, + 0.00860501453280449, + 0.006504071410745382, + 0.013388605788350105, + 0.018441524356603622, + 0.009287345223128796, + 0.011354236863553524, + 0.013342342339456081, + 0.015545963309705257, + 0.0528264082968235, + 0.01792498305439949, + 0.021444285288453102, + 0.026917364448308945, + 0.012194429524242878, + 0.014102060347795486, + 0.017622124403715134, + 0.05001433566212654, + 0.01575622148811817, + 0.01471033040434122, + 0.049332112073898315, + 0.028876548632979393, + 0.018073823302984238, + 0.01304557640105486, + 0.023864634335041046, + 0.030819347128272057, + 0.005131971091032028, + 0.006098512560129166, + 0.01098654605448246, + 0.005174059886485338, + 0.004275494255125523, + 0.002856355858966708, + 0.004779767245054245, + 0.005119717214256525, + 0.002731628715991974, + 0.0035434209275990725, + 0.004250896163284779, + 0.003568756626918912, + 0.002256962703540921, + 0.00209019985049963, + 0.004725676961243153, + 0.0038262067828327417, + 0.013082744553685188, + 0.01663675531744957, + 0.021886147558689117, + 0.021459313109517097, + 0.010036850348114967, + 0.00790029764175415, + 0.009522490203380585, + 0.010364936664700508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.016078807413578033 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054974041879177094 + ], + "opposite_logprob_tokens": [ + -2.92997407913208 + ], + "logprob_reward": -0.054974041879177094, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054974041879177094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306064873933792, + 0.1288081556558609, + 0.044709231704473495, + 0.09942050278186798, + 0.11251123994588852, + 0.28168582916259766, + 0.0439814031124115, + 0.03776180371642113, + 0.030344996601343155, + 0.09017026424407959 + ], + "credit_map": [ + 0.014258245006203651, + 0.012273162603378296, + 0.016301142051815987, + 0.009318184107542038, + 0.04069136083126068, + 0.011198527179658413, + 0.009402467869222164, + 0.017163405194878578, + 0.018428342416882515, + 0.005702576134353876, + 0.02287139557301998, + 0.007029127795249224, + 0.007260223850607872, + 0.03651466220617294, + 0.0076070851646363735, + 0.02339474856853485, + 0.0074958219192922115, + 0.003216944169253111, + 0.006816614884883165, + 0.003990707919001579, + 0.002946738852187991, + 0.007737615145742893, + 0.0052313911728560925, + 0.007273400668054819, + 0.01227732002735138, + 0.009040814824402332, + 0.01623614691197872, + 0.01253492385149002, + 0.010473690927028656, + 0.015554513782262802, + 0.007908917963504791, + 0.015394180081784725, + 0.008251248858869076, + 0.010769080370664597, + 0.018265537917613983, + 0.012463719584047794, + 0.015664851292967796, + 0.017981046810746193, + 0.012453527189791203, + 0.016662223264575005, + 0.07899774610996246, + 0.025584466755390167, + 0.02843807265162468, + 0.01739332638680935, + 0.017088737338781357, + 0.023166559636592865, + 0.017614057287573814, + 0.07340285927057266, + 0.003973086830228567, + 0.003120977897197008, + 0.009606304578483105, + 0.007488314062356949, + 0.0038271022494882345, + 0.0036281824577599764, + 0.006273230537772179, + 0.006064203567802906, + 0.003811096539720893, + 0.0031829560175538063, + 0.007050292566418648, + 0.006087671499699354, + 0.0038901804946362972, + 0.0026130173355340958, + 0.004080129787325859, + 0.007046461571007967, + 0.0023979272227734327, + 0.0029183784499764442, + 0.006110373418778181, + 0.0034000363666564226, + 0.002252192934975028, + 0.0020211220253258944, + 0.0060806660912930965, + 0.005164298228919506, + 0.011239736340939999, + 0.017897002398967743, + 0.013172861188650131, + 0.00985279493033886, + 0.010431830771267414, + 0.007282823324203491, + 0.01060751173645258, + 0.009685694240033627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.054974041879177094 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, forward tracking shot begins, moving steadily at eye level toward the massive, ancient stone arch that dominates the frame. Constructed of massive, weathered blocks of gray granite, its rough surface covered in dark green moss and creepers, intricate, faded carvings of forgotten symbols can be seen along its curved summit. As the camera silently glides through the cool, shadowy passage beneath the arch, the frame is briefly shrouded in deep, cool shadows before emerging in a breathtaking, panoramic, extreme wide shot, revealing a vast, sun-drenched valley. A pristine, emerald-green river meanders through lush, emerald-green meadows dotted with vibrant yellow and purple wildflowers. Vast tracts of dense, dark pine forests climb the gentle slopes of the surrounding hills, which recede into a distant, hazy mountain range with snow-capped peaks. The entire landscape is bathed in the warm golden glow of the afternoon sun, casting long, soft shadows that make the lake shimmer and creating an atmosphere of profound tranquility and epic proportions beneath a vast, cloudless, azure sky.", + "seed": 11701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_6/index_214_seed_11701_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02876369662117213, + "vlm_logprob": -0.02876369662117213, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02876369662117213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020162859931588173 + ], + "opposite_logprob_tokens": [ + -4.395163059234619 + ], + "logprob_reward": -0.020162859931588173, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.020162859931588173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15174178779125214, + 0.10468339920043945, + 0.06335575878620148, + 0.12813082337379456, + 0.12197117507457733, + 0.05619613826274872, + 0.17830125987529755, + 0.06969141960144043, + 0.03533998504281044, + 0.09058824181556702 + ], + "credit_map": [ + 0.014594157226383686, + 0.008550262078642845, + 0.026390990242362022, + 0.013148599304258823, + 0.011646303348243237, + 0.0316188707947731, + 0.019022008404135704, + 0.026770589873194695, + 0.008234525099396706, + 0.007015440613031387, + 0.018979866057634354, + 0.012257277965545654, + 0.003843075828626752, + 0.013881920836865902, + 0.009733017534017563, + 0.030738281086087227, + 0.0053606582805514336, + 0.007061001844704151, + 0.01058889925479889, + 0.00816445890814066, + 0.005055293906480074, + 0.007599843665957451, + 0.008092504926025867, + 0.011433100327849388, + 0.016919676214456558, + 0.021143442019820213, + 0.02290516532957554, + 0.010725163854658604, + 0.011904335580766201, + 0.011528451927006245, + 0.009849470108747482, + 0.023155126720666885, + 0.01796797290444374, + 0.015203221701085567, + 0.017077675089240074, + 0.00861623976379633, + 0.023690978065133095, + 0.017511604353785515, + 0.008975047618150711, + 0.01292843371629715, + 0.007457895204424858, + 0.006373078096657991, + 0.007261432707309723, + 0.004161534830927849, + 0.005813104100525379, + 0.009243885986506939, + 0.0051981741562485695, + 0.010687031783163548, + 0.01568002812564373, + 0.009288476780056953, + 0.014902818016707897, + 0.011351416818797588, + 0.014285750687122345, + 0.039750855416059494, + 0.011817625723779202, + 0.06122430041432381, + 0.008386436849832535, + 0.0066347792744636536, + 0.007266628555953503, + 0.008310907520353794, + 0.006256937514990568, + 0.01579713635146618, + 0.006065616849809885, + 0.01097297202795744, + 0.0035586003214120865, + 0.003932063467800617, + 0.004939285106956959, + 0.003904364537447691, + 0.006123985629528761, + 0.004609144292771816, + 0.0037735686637461185, + 0.004498972557485104, + 0.011760573834180832, + 0.02638026513159275, + 0.007658584509044886, + 0.00827282015234232, + 0.009746924042701721, + 0.004648242611438036, + 0.005536198616027832, + 0.016584640368819237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.020162859931588173 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909126371145248 + ], + "opposite_logprob_tokens": [ + -2.4640913009643555 + ], + "logprob_reward": -0.08909126371145248, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909126371145248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1992853581905365, + 0.11521093547344208, + 0.06233592703938484, + 0.1747041642665863, + 0.09889668226242065, + 0.0989886149764061, + 0.08856243640184402, + 0.07508301734924316, + 0.04438455030322075, + 0.042548347264528275 + ], + "credit_map": [ + 0.017204107716679573, + 0.012244751676917076, + 0.03948688507080078, + 0.024644451215863228, + 0.013628384098410606, + 0.027827005833387375, + 0.032078854739665985, + 0.03217092156410217, + 0.00849836878478527, + 0.009903430938720703, + 0.022074613720178604, + 0.009494543075561523, + 0.005445974413305521, + 0.019754953682422638, + 0.008178257383406162, + 0.03186077997088432, + 0.0050964378751814365, + 0.00455559091642499, + 0.01329232007265091, + 0.004208305850625038, + 0.0029346561059355736, + 0.006054041441529989, + 0.004705801140516996, + 0.021488772705197334, + 0.020442768931388855, + 0.03005002811551094, + 0.023081066086888313, + 0.01799086481332779, + 0.014954975806176662, + 0.021125460043549538, + 0.017496474087238312, + 0.029562542214989662, + 0.016442574560642242, + 0.01103377714753151, + 0.016084523871541023, + 0.006918066181242466, + 0.014175891876220703, + 0.015474028885364532, + 0.008431162685155869, + 0.010336649604141712, + 0.013083435595035553, + 0.008432500995695591, + 0.014760726131498814, + 0.008275751955807209, + 0.011479291133582592, + 0.021557439118623734, + 0.008286157622933388, + 0.013113313354551792, + 0.009313094429671764, + 0.0043809674680233, + 0.007361653260886669, + 0.0057188705541193485, + 0.004661163315176964, + 0.01932474598288536, + 0.006418694742023945, + 0.031383246183395386, + 0.008513235487043858, + 0.006808574311435223, + 0.010584036819636822, + 0.007879712618887424, + 0.01336254645138979, + 0.010787428356707096, + 0.006884997244924307, + 0.010262482799589634, + 0.0037316519301384687, + 0.004515530075877905, + 0.007781538646668196, + 0.0041428483091294765, + 0.005401213653385639, + 0.00666439114138484, + 0.004289686679840088, + 0.007857687771320343, + 0.0034794537350535393, + 0.003189603565260768, + 0.00799169298261404, + 0.007643809076398611, + 0.0035447091795504093, + 0.004309431649744511, + 0.005472857505083084, + 0.00691679073497653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.08909126371145248 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019464610377326608 + ], + "opposite_logprob_tokens": [ + -6.251946449279785 + ], + "logprob_reward": -0.0019464610377326608, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019464610377326608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19496512413024902, + 0.10693888366222382, + 0.043694283813238144, + 0.15096531808376312, + 0.11459188908338547, + 0.07252456992864609, + 0.11453133821487427, + 0.08575010299682617, + 0.04807170853018761, + 0.06796680390834808 + ], + "credit_map": [ + 0.017868321388959885, + 0.013154618442058563, + 0.049804192036390305, + 0.016713159158825874, + 0.006744583137333393, + 0.018779058009386063, + 0.014550184831023216, + 0.057350996881723404, + 0.01077623013406992, + 0.007756766397505999, + 0.020024234429001808, + 0.008562898263335228, + 0.005732689518481493, + 0.015134538523852825, + 0.009196329861879349, + 0.029755203053355217, + 0.004311901517212391, + 0.005324491299688816, + 0.006592150311917067, + 0.005346185527741909, + 0.00474232854321599, + 0.004681456834077835, + 0.003766037756577134, + 0.008929729461669922, + 0.014040506444871426, + 0.034180618822574615, + 0.027530133724212646, + 0.00840761698782444, + 0.01871371455490589, + 0.01879740133881569, + 0.009030554443597794, + 0.02026476338505745, + 0.014260121621191502, + 0.014866773970425129, + 0.017327792942523956, + 0.01043867040425539, + 0.01769590936601162, + 0.01602291874587536, + 0.011915531940758228, + 0.012064165435731411, + 0.008543168194591999, + 0.007442978210747242, + 0.00999282207340002, + 0.007495779078453779, + 0.00704182218760252, + 0.013174917548894882, + 0.0062974123284220695, + 0.012535667978227139, + 0.00831908080726862, + 0.006091640330851078, + 0.015335474163293839, + 0.008064801804721355, + 0.007383192423731089, + 0.0251759085804224, + 0.007383220363408327, + 0.03677802160382271, + 0.008944842033088207, + 0.009831893257796764, + 0.01385345309972763, + 0.011341925710439682, + 0.01155854295939207, + 0.011113299056887627, + 0.00769104203209281, + 0.011415108107030392, + 0.004300376400351524, + 0.005576287396252155, + 0.006750674918293953, + 0.0043184501118958, + 0.008953286334872246, + 0.006766659673303366, + 0.0048830085434019566, + 0.006522961892187595, + 0.007917316630482674, + 0.017627811059355736, + 0.006438770331442356, + 0.007759052328765392, + 0.006624130066484213, + 0.004600847605615854, + 0.004654465243220329, + 0.012344412505626678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0019464610377326608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028183048125356436 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.0028183048125356436, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028183048125356436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09317661821842194, + 0.043258197605609894, + 0.017942393198609352, + 0.034328531473875046, + 0.06751154363155365, + 0.04516053944826126, + 0.09512697905302048, + 0.13055850565433502, + 0.15242932736873627, + 0.32050737738609314 + ], + "credit_map": [ + 0.006042119115591049, + 0.005965370684862137, + 0.019644727930426598, + 0.012436422519385815, + 0.004124494735151529, + 0.009737679734826088, + 0.014372682198882103, + 0.02085312083363533, + 0.005146470852196217, + 0.004595262464135885, + 0.006631792057305574, + 0.004103416111320257, + 0.003231455571949482, + 0.00825350172817707, + 0.003191499738022685, + 0.00810480210930109, + 0.0016185465501621366, + 0.002184995450079441, + 0.003218988422304392, + 0.0013834034325554967, + 0.0018565775826573372, + 0.0019777552224695683, + 0.0014908243902027607, + 0.004211302846670151, + 0.0036565803457051516, + 0.009305104613304138, + 0.004468266852200031, + 0.002165604615584016, + 0.003928538877516985, + 0.0031689954921603203, + 0.001937240594998002, + 0.005698200780898333, + 0.01047468464821577, + 0.008846105076372623, + 0.011213399469852448, + 0.004541608504951, + 0.011079946532845497, + 0.009726065210998058, + 0.004581759218126535, + 0.007047982886433601, + 0.004996783100068569, + 0.004749423824250698, + 0.006998921278864145, + 0.0046052876859903336, + 0.0056145284324884415, + 0.006319984328001738, + 0.004413385409861803, + 0.007462226320058107, + 0.0068362061865627766, + 0.005326013080775738, + 0.008603177964687347, + 0.006982007529586554, + 0.010993248783051968, + 0.02038894221186638, + 0.00705779530107975, + 0.028939586132764816, + 0.01338714174926281, + 0.015074452385306358, + 0.016950925812125206, + 0.01148135680705309, + 0.018515395000576973, + 0.0225296039134264, + 0.013343164697289467, + 0.019276469945907593, + 0.012328920885920525, + 0.013104351237416267, + 0.022591982036828995, + 0.0220170971006155, + 0.02065972238779068, + 0.02083064801990986, + 0.016108138486742973, + 0.02478845976293087, + 0.04536440223455429, + 0.10411784797906876, + 0.01744341105222702, + 0.021439079195261, + 0.033424463123083115, + 0.020665915682911873, + 0.015940207988023758, + 0.062112025916576385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0028183048125356436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a steady forward tracking shot moving at eye level toward a massive ancient stone arch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the stone arch constructed of weathered gray granite blocks covered in dark green moss and creepers with faded carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera glide through a cool, shadowy passage beneath the arch before emerging into a panoramic extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed landscape a sun-drenched valley featuring an emerald-green river, yellow and purple wildflowers, and dark pine forests on hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with a vast cloudless azure sky and distant snow-capped mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979959361255169 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979959361255169, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979959361255169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10337814688682556, + 0.07410099357366562, + 0.029750986024737358, + 0.05781536549329758, + 0.08731110394001007, + 0.043224629014730453, + 0.10439317673444748, + 0.19109252095222473, + 0.16638752818107605, + 0.14254553616046906 + ], + "credit_map": [ + 0.008374138735234737, + 0.006339950952678919, + 0.022246185690164566, + 0.009146034717559814, + 0.007560496684163809, + 0.019168123602867126, + 0.0098955687135458, + 0.020647644996643066, + 0.00377679243683815, + 0.0054517206735908985, + 0.014405225403606892, + 0.005621320102363825, + 0.005154560785740614, + 0.014553764835000038, + 0.0058595892041921616, + 0.019278021529316902, + 0.003146412316709757, + 0.005096531473100185, + 0.004502647090703249, + 0.0025451977271586657, + 0.002278286963701248, + 0.004024477209895849, + 0.0021998437587171793, + 0.0059575894847512245, + 0.00507719861343503, + 0.00661173090338707, + 0.01413580309599638, + 0.0035461867228150368, + 0.005244569852948189, + 0.003282369114458561, + 0.004011232405900955, + 0.01590627431869507, + 0.013836230151355267, + 0.010908620432019234, + 0.017083263024687767, + 0.006709104869514704, + 0.012174575589597225, + 0.012103362940251827, + 0.005675360560417175, + 0.00882058497518301, + 0.005699429661035538, + 0.0037344396114349365, + 0.005016161594539881, + 0.004878116771578789, + 0.005258676130324602, + 0.008271599188446999, + 0.004599467851221561, + 0.0057667396031320095, + 0.008620277978479862, + 0.008184731006622314, + 0.00824426393955946, + 0.007057106122374535, + 0.013134895823895931, + 0.01958475261926651, + 0.00816473737359047, + 0.03140241280198097, + 0.01846790686249733, + 0.0266120545566082, + 0.02508203499019146, + 0.01662311516702175, + 0.020201053470373154, + 0.03677508234977722, + 0.018159829080104828, + 0.029171444475650787, + 0.015280680730938911, + 0.021280033513903618, + 0.02097136527299881, + 0.013366860337555408, + 0.0342441126704216, + 0.02342217043042183, + 0.015980862081050873, + 0.021841444075107574, + 0.019456887617707253, + 0.044616058468818665, + 0.008442397229373455, + 0.007244745269417763, + 0.01787528395652771, + 0.011839841492474079, + 0.0069921198301017284, + 0.026078196242451668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.02979959361255169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_tensor([8494, 8494])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_tensor([8494, 8494])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a0a61d4f1169030917ac74de43abdf2ec9b6713 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_tensor([8494, 8494])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a low angle, the camera slowly panning to the right, captures a scene in a deep, narrow alley. The frame focuses on a wall constructed of ancient blue bricks, moisture seeping through the cracks and dappled with patches of velvety moss of varying shades. A faint beam of natural light slants in from the end of the alley, illuminating the wall while the surroundings remain shrouded in deep shadow. A delicate hand, rounded knuckles, clad in a slightly oversized off-white sweater with slightly rolled-up sleeves, cautiously reaches into the frame. The child's fingers first gently touch the moist moss, then, with a touch of curiosity and wonder, the palm of the hand steadily and deliberately slides across the rough wall and soft moss, feeling the unique, cool, yet vibrant touch. Against the cool background, this light and this innocent gesture imbue the scene with a remarkable tranquility, concentration, and poetry.", + "seed": 29814, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_8494_seed_29814_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08653866499662399, + "vlm_logprob": -0.08653866499662399, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.08653866499662399, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0234310831874609 + ], + "opposite_logprob_tokens": [ + -4.023431301116943 + ], + "logprob_reward": -0.0234310831874609, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0234310831874609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08488088101148605, + 0.055355504155159, + 0.09828084707260132, + 0.14698992669582367, + 0.2856360077857971, + 0.06751543283462524, + 0.10562341660261154, + 0.03689299523830414, + 0.024381045252084732, + 0.0944439247250557 + ], + "credit_map": [ + 0.006169015076011419, + 0.0054929256439208984, + 0.010862642899155617, + 0.006745067425072193, + 0.010812623426318169, + 0.014522276818752289, + 0.016008395701646805, + 0.014267939142882824, + 0.004454786889255047, + 0.003788608592003584, + 0.00799607578665018, + 0.006567313801497221, + 0.008873905055224895, + 0.007773225195705891, + 0.008351352997124195, + 0.007550235837697983, + 0.007363035809248686, + 0.0046620601788163185, + 0.026963867247104645, + 0.007825827226042747, + 0.005689220502972603, + 0.008986161090433598, + 0.009648483246564865, + 0.02714218944311142, + 0.009165509603917599, + 0.011990567669272423, + 0.01687944121658802, + 0.018903067335486412, + 0.026195967569947243, + 0.02054404839873314, + 0.023439811542630196, + 0.019871516153216362, + 0.08413082361221313, + 0.010534199886023998, + 0.014584465883672237, + 0.010068627074360847, + 0.023740459233522415, + 0.02233038656413555, + 0.03095223568379879, + 0.08929479867219925, + 0.007895676419138908, + 0.005743582732975483, + 0.008153301663696766, + 0.012201418168842793, + 0.009084245190024376, + 0.007777948398143053, + 0.009154441766440868, + 0.007504819426685572, + 0.010413728654384613, + 0.010584588162600994, + 0.006597347557544708, + 0.013634013012051582, + 0.018155191093683243, + 0.012697260826826096, + 0.02160695381462574, + 0.011934327892959118, + 0.007255333475768566, + 0.0022642610128968954, + 0.0019041141495108604, + 0.0023492283653467894, + 0.002322133630514145, + 0.003791064489632845, + 0.005242565181106329, + 0.011764292605221272, + 0.00230701407417655, + 0.0016178442165255547, + 0.002311025746166706, + 0.0025620830710977316, + 0.0033256050664931536, + 0.00528579019010067, + 0.0035289847292006016, + 0.0034426988568156958, + 0.007124876603484154, + 0.00971694104373455, + 0.014111589640378952, + 0.013603019528090954, + 0.006813391111791134, + 0.02101651020348072, + 0.0071353367529809475, + 0.014922259375452995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.0234310831874609 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195733428001404 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195733428001404, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195733428001404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0371563546359539, + 0.01739591173827648, + 0.03477058559656143, + 0.024565456435084343, + 0.1446123719215393, + 0.09733043611049652, + 0.11986766010522842, + 0.04892051964998245, + 0.3318835496902466, + 0.14349721372127533 + ], + "credit_map": [ + 0.0034619413781911135, + 0.0036383760161697865, + 0.0050714630633592606, + 0.004281116183847189, + 0.005477158352732658, + 0.004284382332116365, + 0.005864686798304319, + 0.005077227018773556, + 0.0015900165308266878, + 0.0012936904095113277, + 0.002594881923869252, + 0.001553708454594016, + 0.0028166400734335184, + 0.002472085179761052, + 0.002820087131112814, + 0.0022548004053533077, + 0.002823425456881523, + 0.0013183680130168796, + 0.009320959448814392, + 0.002526897005736828, + 0.002680898644030094, + 0.002731274114921689, + 0.004728158004581928, + 0.008640600368380547, + 0.0013095388421788812, + 0.0018492508679628372, + 0.0027891830541193485, + 0.00331213534809649, + 0.004042067099362612, + 0.004023795481771231, + 0.004412839654833078, + 0.002826644340530038, + 0.04255128651857376, + 0.005923772230744362, + 0.0070845321752130985, + 0.005610582884401083, + 0.011948822066187859, + 0.012012111023068428, + 0.014216592535376549, + 0.04526466131210327, + 0.018229234963655472, + 0.008741529658436775, + 0.01152640487998724, + 0.016637571156024933, + 0.012227454222738743, + 0.005889889318495989, + 0.011712469160556793, + 0.012365869246423244, + 0.01160387322306633, + 0.010988425463438034, + 0.007834955118596554, + 0.015533888712525368, + 0.021494440734386444, + 0.013637596741318703, + 0.025796057656407356, + 0.01297841127961874, + 0.00789569690823555, + 0.002265048446133733, + 0.0023902542889118195, + 0.00430276058614254, + 0.005488233640789986, + 0.006260068621486425, + 0.007208004593849182, + 0.01311044953763485, + 0.07485820353031158, + 0.011410258710384369, + 0.03296177089214325, + 0.010636589489877224, + 0.009111692197620869, + 0.043503910303115845, + 0.015454021282494068, + 0.13394707441329956, + 0.014529976062476635, + 0.014323532581329346, + 0.020539062097668648, + 0.016520045697689056, + 0.010726600885391235, + 0.03232109919190407, + 0.012113434262573719, + 0.022423435002565384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.018195733428001404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069306060671806 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069306060671806, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069306060671806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051808878779411316, + 0.04650282859802246, + 0.05887003242969513, + 0.035448044538497925, + 0.11471350491046906, + 0.14955033361911774, + 0.06524988263845444, + 0.06195094808936119, + 0.26153087615966797, + 0.15437465906143188 + ], + "credit_map": [ + 0.0046910629607737064, + 0.0036124237813055515, + 0.007294678129255772, + 0.0041943867690861225, + 0.005045714788138866, + 0.008559726178646088, + 0.009780444204807281, + 0.008630440570414066, + 0.004413359332829714, + 0.002186518395319581, + 0.007632600609213114, + 0.0042442139238119125, + 0.006655210163444281, + 0.0063547720201313496, + 0.00823312345892191, + 0.006783034652471542, + 0.004998630844056606, + 0.0019895844161510468, + 0.01700187660753727, + 0.004387070424854755, + 0.003799069905653596, + 0.0054730079136788845, + 0.004950399976223707, + 0.016270389780402184, + 0.0023368990514427423, + 0.0034903110936284065, + 0.0036114268004894257, + 0.005079018883407116, + 0.0056954179890453815, + 0.00454701017588377, + 0.007354334462434053, + 0.003333623521029949, + 0.03125276789069176, + 0.004705085884779692, + 0.0064279246143996716, + 0.006841159891337156, + 0.009232558310031891, + 0.008950765244662762, + 0.014263900928199291, + 0.033039338886737823, + 0.036265913397073746, + 0.00857167225331068, + 0.01461733691394329, + 0.011136835440993309, + 0.01437631156295538, + 0.012136711739003658, + 0.01905081979930401, + 0.033394742757081985, + 0.007115297485142946, + 0.005558560136705637, + 0.007108621299266815, + 0.010354242287576199, + 0.010914895683526993, + 0.005798947531729937, + 0.010167001746594906, + 0.008232321590185165, + 0.0089956009760499, + 0.003122500143945217, + 0.0032350695692002773, + 0.005083551164716482, + 0.007244591135531664, + 0.008775076828897, + 0.009666241705417633, + 0.015828315168619156, + 0.06079719588160515, + 0.00824281107634306, + 0.026170939207077026, + 0.007694590371102095, + 0.00606005871668458, + 0.03514990583062172, + 0.00909591093659401, + 0.1083194687962532, + 0.012189533561468124, + 0.0128748444840312, + 0.02525145933032036, + 0.02640192024409771, + 0.01371443085372448, + 0.03041122853755951, + 0.00986754521727562, + 0.02366369590163231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.016069306060671806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346776634454727 + ], + "opposite_logprob_tokens": [ + -3.6513466835021973 + ], + "logprob_reward": -0.026346776634454727, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346776634454727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06785020977258682, + 0.059714511036872864, + 0.05914762616157532, + 0.01456114836037159, + 0.1492224931716919, + 0.031065301969647408, + 0.16002342104911804, + 0.03267884626984596, + 0.15971492230892181, + 0.2660214900970459 + ], + "credit_map": [ + 0.003553948597982526, + 0.003923103678971529, + 0.009360424242913723, + 0.006548093631863594, + 0.008851185441017151, + 0.011354067362844944, + 0.015490048564970493, + 0.008769339881837368, + 0.003897777758538723, + 0.002350081456825137, + 0.010119106620550156, + 0.004562385380268097, + 0.009578692726790905, + 0.008990787900984287, + 0.012242491357028484, + 0.007973186671733856, + 0.003717265324667096, + 0.0029003408271819353, + 0.013340859673917294, + 0.004969848319888115, + 0.005690462421625853, + 0.006606188137084246, + 0.00830927025526762, + 0.013613392598927021, + 0.0011332363355904818, + 0.0017666745698079467, + 0.0015778784872964025, + 0.0018694205209612846, + 0.002200676593929529, + 0.0020878827199339867, + 0.00233807903714478, + 0.0015872997464612126, + 0.03632407262921333, + 0.00985379982739687, + 0.014556236565113068, + 0.021284004673361778, + 0.012055735103785992, + 0.011096901260316372, + 0.013647926971316338, + 0.030403822660446167, + 0.004325410351157188, + 0.0030194318387657404, + 0.002687643049284816, + 0.004616080783307552, + 0.00400174967944622, + 0.0028114155866205692, + 0.0038403794169425964, + 0.00576319033280015, + 0.014288920909166336, + 0.017373044043779373, + 0.010853752493858337, + 0.01676982082426548, + 0.03237937390804291, + 0.02070465125143528, + 0.02792072296142578, + 0.019733142107725143, + 0.004430677276104689, + 0.002398926764726639, + 0.0018446509493514895, + 0.0034550002310425043, + 0.0039010511245578527, + 0.0032031890004873276, + 0.005398781504482031, + 0.008046567440032959, + 0.034359000623226166, + 0.004677974618971348, + 0.015216716565191746, + 0.005859448108822107, + 0.0076770600862801075, + 0.02172708883881569, + 0.008358065038919449, + 0.061839569360017776, + 0.016489481553435326, + 0.02635282650589943, + 0.03606439381837845, + 0.042699120938777924, + 0.028781214728951454, + 0.046614453196525574, + 0.03357778862118721, + 0.03544221445918083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.026346776634454727 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865042567253113 + ], + "opposite_logprob_tokens": [ + -1.2236504554748535 + ], + "logprob_reward": -0.34865042567253113, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34865042567253113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08503713458776474, + 0.09636571258306503, + 0.14503830671310425, + 0.060842812061309814, + 0.11197437345981598, + 0.06532831490039825, + 0.12891487777233124, + 0.13021701574325562, + 0.13753098249435425, + 0.03875049203634262 + ], + "credit_map": [ + 0.007937449961900711, + 0.00875051785260439, + 0.011613667942583561, + 0.008334360085427761, + 0.012001559138298035, + 0.012372620403766632, + 0.014432515949010849, + 0.009594446048140526, + 0.007612030953168869, + 0.004416017793118954, + 0.01586874946951866, + 0.010582070797681808, + 0.013900969177484512, + 0.01455507893115282, + 0.018408766016364098, + 0.011022020131349564, + 0.010706947185099125, + 0.006986953318119049, + 0.03414391353726387, + 0.012656522914767265, + 0.011515040881931782, + 0.01604786142706871, + 0.020129628479480743, + 0.032851431518793106, + 0.0044417548924684525, + 0.005899267736822367, + 0.007425643969327211, + 0.007621135096997023, + 0.011379046365618706, + 0.010594694875180721, + 0.0074216523207724094, + 0.006059618201106787, + 0.03115779533982277, + 0.004292295780032873, + 0.005374654661864042, + 0.006551502272486687, + 0.00846199132502079, + 0.010051651857793331, + 0.013966375961899757, + 0.03211810439825058, + 0.010974539443850517, + 0.0029128810856491327, + 0.004137199837714434, + 0.006007080897688866, + 0.006214782129973173, + 0.006773380097001791, + 0.011128625832498074, + 0.01717982441186905, + 0.01158462930470705, + 0.010771653614938259, + 0.008428952656686306, + 0.014485948719084263, + 0.02659643441438675, + 0.017866123467683792, + 0.02406265400350094, + 0.015118484385311604, + 0.021827735006809235, + 0.00791096780449152, + 0.006612875498831272, + 0.009128209203481674, + 0.009575781412422657, + 0.014437687583267689, + 0.020750846713781357, + 0.039972927421331406, + 0.031742554157972336, + 0.004231960512697697, + 0.013614724390208721, + 0.004305203910917044, + 0.003113268408924341, + 0.019062532112002373, + 0.004799824208021164, + 0.056660912930965424, + 0.0036822278052568436, + 0.003449775045737624, + 0.006001985631883144, + 0.0044624293223023415, + 0.004185373894870281, + 0.006543260533362627, + 0.004252526443451643, + 0.0061729117296636105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.34865042567253113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera slowly panning to the right, captures a scene in a deep, narrow alley. The frame focuses on a wall constructed of ancient blue bricks, moisture seeping through the cracks and dappled with patches of velvety moss of varying shades. A faint beam of natural light slants in from the end of the alley, illuminating the wall while the surroundings remain shrouded in deep shadow. A delicate hand, rounded knuckles, clad in a slightly oversized off-white sweater with slightly rolled-up sleeves, cautiously reaches into the frame. The child's fingers first gently touch the moist moss, then, with a touch of curiosity and wonder, the palm of the hand steadily and deliberately slides across the rough wall and soft moss, feeling the unique, cool, yet vibrant touch. Against the cool background, this light and this innocent gesture imbue the scene with a remarkable tranquility, concentration, and poetry.", + "seed": 29815, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_8494_seed_29815_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18882344532757997, + "vlm_logprob": -0.18882344532757997, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.18882344532757997, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31790781021118164 + ], + "opposite_logprob_tokens": [ + -1.3179078102111816 + ], + "logprob_reward": -0.31790781021118164, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31790781021118164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15412533283233643, + 0.050303198397159576, + 0.0841076448559761, + 0.060547277331352234, + 0.07016583532094955, + 0.05621366947889328, + 0.02606535330414772, + 0.040197696536779404, + 0.40282541513442993, + 0.0554484948515892 + ], + "credit_map": [ + 0.011366637423634529, + 0.027253489941358566, + 0.008747875690460205, + 0.012012192979454994, + 0.023306531831622124, + 0.026428135111927986, + 0.020134013146162033, + 0.024876467883586884, + 0.004455864895135164, + 0.008256935514509678, + 0.00351297901943326, + 0.004589082673192024, + 0.007254159078001976, + 0.007139985915273428, + 0.009273312985897064, + 0.005820886697620153, + 0.0056285252794623375, + 0.01139665488153696, + 0.005931788589805365, + 0.006899005733430386, + 0.011328831315040588, + 0.017285969108343124, + 0.012462496757507324, + 0.013174386695027351, + 0.0035680874716490507, + 0.005765540059655905, + 0.004116583615541458, + 0.0038068853318691254, + 0.008541095070540905, + 0.008959907107055187, + 0.016375310719013214, + 0.009413880296051502, + 0.005720712244510651, + 0.006349262781441212, + 0.005049547646194696, + 0.005708659533411264, + 0.014856086112558842, + 0.009268557652831078, + 0.013027398847043514, + 0.01018561888486147, + 0.0038084927946329117, + 0.0044117034412920475, + 0.003871783846989274, + 0.005268472712486982, + 0.010359270498156548, + 0.006296234205365181, + 0.013423039577901363, + 0.008774681948125362, + 0.0030174909625202417, + 0.0019944794476032257, + 0.001717758597806096, + 0.00221847603097558, + 0.004043274093419313, + 0.00557769276201725, + 0.003388351295143366, + 0.004107832908630371, + 0.0053262305445969105, + 0.0021383846178650856, + 0.0028922369237989187, + 0.0024343966506421566, + 0.006491761654615402, + 0.008091717958450317, + 0.00589738879352808, + 0.006925584282726049, + 0.04871387779712677, + 0.019183939322829247, + 0.020425500348210335, + 0.02524503320455551, + 0.057872746139764786, + 0.09992414712905884, + 0.0752372294664383, + 0.05622295290231705, + 0.007822058163583279, + 0.003576093353331089, + 0.0036156370770186186, + 0.0037142899818718433, + 0.008352505043148994, + 0.0071606822311878204, + 0.00774439238011837, + 0.013462846167385578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.31790781021118164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205407351255417 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.014205407351255417, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205407351255417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23945830762386322, + 0.06766906380653381, + 0.13278856873512268, + 0.061078667640686035, + 0.07407240569591522, + 0.045826926827430725, + 0.037411000579595566, + 0.06832406669855118, + 0.13970346748828888, + 0.13366751372814178 + ], + "credit_map": [ + 0.015152950771152973, + 0.03305233642458916, + 0.013988322578370571, + 0.03239265829324722, + 0.03177354484796524, + 0.04181353747844696, + 0.037217024713754654, + 0.03406795859336853, + 0.004603981506079435, + 0.009999769739806652, + 0.007194299716502428, + 0.007658739108592272, + 0.011463193222880363, + 0.009403223171830177, + 0.008654478006064892, + 0.008691376075148582, + 0.009198579005897045, + 0.019546054303646088, + 0.017491284757852554, + 0.01732635870575905, + 0.015569562092423439, + 0.020410751923918724, + 0.018466126173734665, + 0.014779852703213692, + 0.004476774018257856, + 0.007188854273408651, + 0.005037762224674225, + 0.00412657018750906, + 0.007954113185405731, + 0.010883418843150139, + 0.011061911471188068, + 0.010349265299737453, + 0.005361320450901985, + 0.008271485567092896, + 0.005739058367908001, + 0.005644230637699366, + 0.010836395435035229, + 0.011351251043379307, + 0.015214776620268822, + 0.011653888039290905, + 0.004258908797055483, + 0.004511378705501556, + 0.003919028677046299, + 0.0037232637405395508, + 0.0065292962826788425, + 0.007260117679834366, + 0.0076583269983530045, + 0.007966606877744198, + 0.0036255044396966696, + 0.005600436124950647, + 0.004734338726848364, + 0.00461576646193862, + 0.0054719760082662106, + 0.0038957593496888876, + 0.005010148975998163, + 0.00445706769824028, + 0.008768417872488499, + 0.006964527536183596, + 0.007026568055152893, + 0.0059979381039738655, + 0.00797949731349945, + 0.01110262330621481, + 0.0098567521199584, + 0.010627739131450653, + 0.019358672201633453, + 0.011179616674780846, + 0.010808966122567654, + 0.014395204372704029, + 0.01746409945189953, + 0.019639790058135986, + 0.023918798193335533, + 0.022938314825296402, + 0.011094613932073116, + 0.009782698936760426, + 0.016170425340533257, + 0.015078688971698284, + 0.0215672105550766, + 0.01512435358017683, + 0.02525663562119007, + 0.0195928867906332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.014205407351255417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539827264845371 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539827264845371, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539827264845371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1160987913608551, + 0.06804976612329483, + 0.0641404390335083, + 0.0820436030626297, + 0.16721628606319427, + 0.0616363063454628, + 0.09046284109354019, + 0.15349642932415009, + 0.08220706880092621, + 0.11464843899011612 + ], + "credit_map": [ + 0.007367634680122137, + 0.01864219643175602, + 0.0068210032768547535, + 0.012950929813086987, + 0.019358588382601738, + 0.021782245486974716, + 0.016465360298752785, + 0.012710834853351116, + 0.005634903442114592, + 0.012286565266549587, + 0.005438642576336861, + 0.005080622620880604, + 0.012214018031954765, + 0.012394939549267292, + 0.006668516900390387, + 0.008331564255058765, + 0.00486656092107296, + 0.007959055714309216, + 0.005758564453572035, + 0.007175164297223091, + 0.005464242305606604, + 0.012711100280284882, + 0.009532407857477665, + 0.010673345066606998, + 0.006741078104823828, + 0.006559337023645639, + 0.005718621890991926, + 0.006714131683111191, + 0.01142389327287674, + 0.010348991490900517, + 0.022756284102797508, + 0.011781269684433937, + 0.017490200698375702, + 0.017613127827644348, + 0.014676865190267563, + 0.012824375182390213, + 0.024258816614747047, + 0.021033942699432373, + 0.03336066007614136, + 0.02595829963684082, + 0.005462951492518187, + 0.006937862373888493, + 0.006639782804995775, + 0.005989962723106146, + 0.008860459551215172, + 0.007292387541383505, + 0.012085465714335442, + 0.008367438800632954, + 0.012040745466947556, + 0.009914631024003029, + 0.009453651495277882, + 0.008231351152062416, + 0.011328859254717827, + 0.016886120662093163, + 0.010840412229299545, + 0.011767080053687096, + 0.017421158030629158, + 0.015273071825504303, + 0.018923183903098106, + 0.014507615007460117, + 0.018141644075512886, + 0.019538508728146553, + 0.026391159743070602, + 0.02330010011792183, + 0.00880344957113266, + 0.007761312182992697, + 0.00847339816391468, + 0.009774976409971714, + 0.011391955427825451, + 0.010267609730362892, + 0.011439266614615917, + 0.01429509837180376, + 0.0169150922447443, + 0.006079566199332476, + 0.010374022647738457, + 0.00793046411126852, + 0.010374893434345722, + 0.015017268247902393, + 0.016081273555755615, + 0.031875867396593094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.012539827264845371 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288074880838394 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023288074880838394, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288074880838394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17459283769130707, + 0.028309499844908714, + 0.05548793822526932, + 0.08636793494224548, + 0.06774340569972992, + 0.046309664845466614, + 0.03887174278497696, + 0.09391160309314728, + 0.3502565324306488, + 0.058148860931396484 + ], + "credit_map": [ + 0.014630137011408806, + 0.021768584847450256, + 0.008920731022953987, + 0.016242366284132004, + 0.016879256814718246, + 0.04825463891029358, + 0.03405207395553589, + 0.013845040462911129, + 0.002437627175822854, + 0.003745600115507841, + 0.00250760349445045, + 0.003030194900929928, + 0.004045320674777031, + 0.004500080831348896, + 0.004252964630723, + 0.0037901082541793585, + 0.004823558963835239, + 0.009500120766460896, + 0.005462506785988808, + 0.005849851295351982, + 0.005847910884767771, + 0.010095284320414066, + 0.005818650592118502, + 0.00809005182236433, + 0.004607103765010834, + 0.0076636336743831635, + 0.005414652172476053, + 0.00603330647572875, + 0.01249451283365488, + 0.011207235977053642, + 0.02497374452650547, + 0.013973749242722988, + 0.004482457879930735, + 0.006121253129094839, + 0.005211733281612396, + 0.004877968225628138, + 0.0123618608340621, + 0.008220668882131577, + 0.015554902143776417, + 0.010912559926509857, + 0.003960041329264641, + 0.0059896428138017654, + 0.003579688025638461, + 0.003341267816722393, + 0.00680087273940444, + 0.008387415669858456, + 0.007868767715990543, + 0.006381968967616558, + 0.005599888972938061, + 0.0021120491437613964, + 0.0019535920582711697, + 0.0018574981950223446, + 0.005542504135519266, + 0.011341295205056667, + 0.004270308651030064, + 0.006194609683007002, + 0.013577945530414581, + 0.005691624712198973, + 0.005521898623555899, + 0.004958524368703365, + 0.01646391674876213, + 0.019634373486042023, + 0.013612286187708378, + 0.01445102971047163, + 0.053054071962833405, + 0.03653281182050705, + 0.021779924631118774, + 0.03339502587914467, + 0.044285714626312256, + 0.056633561849594116, + 0.05558376386761665, + 0.048991668969392776, + 0.01064655464142561, + 0.005447670817375183, + 0.0031555735040456057, + 0.005332151427865028, + 0.010549737140536308, + 0.007845007814466953, + 0.009996802546083927, + 0.0051753646694123745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.023288074880838394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576176106929779 + ], + "opposite_logprob_tokens": [ + -0.826176106929779 + ], + "logprob_reward": -0.576176106929779, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576176106929779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1455385982990265, + 0.07635083794593811, + 0.05796985700726509, + 0.06762459129095078, + 0.08568733185529709, + 0.09348583966493607, + 0.1546349674463272, + 0.09603848308324814, + 0.11562532186508179, + 0.10704418271780014 + ], + "credit_map": [ + 0.00801864918321371, + 0.022279635071754456, + 0.009002895094454288, + 0.016061555594205856, + 0.019258767366409302, + 0.03214523196220398, + 0.02487245202064514, + 0.013899422250688076, + 0.0060720257461071014, + 0.014536348171532154, + 0.005671892315149307, + 0.004810416605323553, + 0.015477922745049, + 0.012397324666380882, + 0.007962176576256752, + 0.009422730654478073, + 0.006042758002877235, + 0.010126815177500248, + 0.004142825026065111, + 0.005183778237551451, + 0.008778098039329052, + 0.00843728892505169, + 0.0067888544872403145, + 0.00846943911164999, + 0.006202033255249262, + 0.010532157495617867, + 0.006505310535430908, + 0.006249706726521254, + 0.007652623113244772, + 0.012839260511100292, + 0.008564290590584278, + 0.009079212322831154, + 0.00583742605522275, + 0.008358035236597061, + 0.006637145299464464, + 0.007296373136341572, + 0.010725130327045918, + 0.021466050297021866, + 0.011577500030398369, + 0.013789672404527664, + 0.006335960701107979, + 0.007776457816362381, + 0.006332256365567446, + 0.00754941301420331, + 0.01651913858950138, + 0.01276046596467495, + 0.021192079409956932, + 0.015020068734884262, + 0.016017088666558266, + 0.009840104728937149, + 0.007764627691358328, + 0.008427754044532776, + 0.02243161015212536, + 0.04310838133096695, + 0.01638716831803322, + 0.03065822646021843, + 0.011397001333534718, + 0.004962420091032982, + 0.004675137810409069, + 0.004613642580807209, + 0.017170144245028496, + 0.021455418318510056, + 0.015805652365088463, + 0.01595906727015972, + 0.016330841928720474, + 0.009595273993909359, + 0.005919097922742367, + 0.009399808943271637, + 0.015256138518452644, + 0.021707218140363693, + 0.020946068689227104, + 0.01647086814045906, + 0.015435577370226383, + 0.005517560988664627, + 0.006257610861212015, + 0.0084458589553833, + 0.008821950294077396, + 0.01711495779454708, + 0.018390877172350883, + 0.027059802785515785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.576176106929779 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera slowly panning to the right, captures a scene in a deep, narrow alley. The frame focuses on a wall constructed of ancient blue bricks, moisture seeping through the cracks and dappled with patches of velvety moss of varying shades. A faint beam of natural light slants in from the end of the alley, illuminating the wall while the surroundings remain shrouded in deep shadow. A delicate hand, rounded knuckles, clad in a slightly oversized off-white sweater with slightly rolled-up sleeves, cautiously reaches into the frame. The child's fingers first gently touch the moist moss, then, with a touch of curiosity and wonder, the palm of the hand steadily and deliberately slides across the rough wall and soft moss, feeling the unique, cool, yet vibrant touch. Against the cool background, this light and this innocent gesture imbue the scene with a remarkable tranquility, concentration, and poetry.", + "seed": 29816, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_8494_seed_29816_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.060377922095358375, + "vlm_logprob": -0.060377922095358375, + "vlm_yes_no_margin": 3.174999976158142, + "vlm_token_credit": -0.060377922095358375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1327572762966156 + ], + "opposite_logprob_tokens": [ + -2.1327571868896484 + ], + "logprob_reward": -0.1327572762966156, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1327572762966156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15672525763511658, + 0.4820461571216583, + 0.05151096358895302, + 0.04101758822798729, + 0.01903904601931572, + 0.017536696046590805, + 0.0810253918170929, + 0.054967667907476425, + 0.07203824818134308, + 0.024092933163046837 + ], + "credit_map": [ + 0.014258157461881638, + 0.014230120927095413, + 0.01832721009850502, + 0.015829529613256454, + 0.034100405871868134, + 0.02209841087460518, + 0.019654009491205215, + 0.018227411434054375, + 0.013795102946460247, + 0.007675154134631157, + 0.020942188799381256, + 0.01857212744653225, + 0.022059325128793716, + 0.21152901649475098, + 0.015904709696769714, + 0.17156851291656494, + 0.0040033250115811825, + 0.003688226919621229, + 0.003520719474181533, + 0.004629336763173342, + 0.009811794385313988, + 0.009000645019114017, + 0.007023588288575411, + 0.00983333121985197, + 0.0026579275727272034, + 0.002637853380292654, + 0.0019025802612304688, + 0.002831270918250084, + 0.008312167599797249, + 0.006858080625534058, + 0.00950636900961399, + 0.006311336532235146, + 0.0010066991671919823, + 0.002708571497350931, + 0.0008717644959688187, + 0.0013450713595375419, + 0.003548414446413517, + 0.0033594551496207714, + 0.003655597334727645, + 0.002543474081903696, + 0.0017144334269687533, + 0.0013263893779367208, + 0.001995242200791836, + 0.00161117990501225, + 0.002530298661440611, + 0.0022222106344997883, + 0.004009951371699572, + 0.002126990584656596, + 0.00703316368162632, + 0.005210644565522671, + 0.005528467241674662, + 0.004119295626878738, + 0.00908141303807497, + 0.015437690541148186, + 0.021959729492664337, + 0.012654984369874, + 0.003202267689630389, + 0.005005256272852421, + 0.0023325530346482992, + 0.004258575849235058, + 0.011357645504176617, + 0.009711170569062233, + 0.011718072928488255, + 0.007382125593721867, + 0.003619779134169221, + 0.005120080430060625, + 0.0037444839254021645, + 0.005160307511687279, + 0.014474285766482353, + 0.010768839158117771, + 0.017492424696683884, + 0.011658052913844585, + 0.0014875000342726707, + 0.0024126635398715734, + 0.0019464835058897734, + 0.0050680129788815975, + 0.0029372943099588156, + 0.0036123809404671192, + 0.0023955225478857756, + 0.004233075305819511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.1327572762966156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125464191660285 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.0125464191660285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125464191660285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0248368289321661, + 0.4974256157875061, + 0.03314319625496864, + 0.012018651701509953, + 0.038122545927762985, + 0.01820461079478264, + 0.025911515578627586, + 0.05379517376422882, + 0.16707251965999603, + 0.1294693499803543 + ], + "credit_map": [ + 0.0025172457098960876, + 0.0020402574446052313, + 0.0029111760668456554, + 0.002245857147499919, + 0.0044670505449175835, + 0.00427605165168643, + 0.0032963287085294724, + 0.0030828628223389387, + 0.013234071433544159, + 0.006411887239664793, + 0.019109707325696945, + 0.01840117573738098, + 0.021545594558119774, + 0.2221294641494751, + 0.014931845478713512, + 0.18166187405586243, + 0.0030599518213421106, + 0.002742384560406208, + 0.002652327064424753, + 0.0025190222077071667, + 0.006512049585580826, + 0.005265833344310522, + 0.0057589216157794, + 0.0046327076852321625, + 0.0007422862108796835, + 0.0009945984929800034, + 0.0007642019190825522, + 0.001080983318388462, + 0.0026867904234677553, + 0.0018067362252622843, + 0.0019512304570525885, + 0.0019918240141123533, + 0.0036516820546239614, + 0.006546524818986654, + 0.003188993548974395, + 0.0034768646582961082, + 0.0059642549604177475, + 0.004929963033646345, + 0.00499122915789485, + 0.005373033229261637, + 0.0013892634306102991, + 0.00128556985873729, + 0.0014645849587395787, + 0.0014952534111216664, + 0.002968126442283392, + 0.002240069443359971, + 0.005344750359654427, + 0.002016992773860693, + 0.0027733203023672104, + 0.002408761763945222, + 0.0028955277521163225, + 0.0016555130714550614, + 0.0037481477484107018, + 0.003377713495865464, + 0.005036229733377695, + 0.0040163011290133, + 0.004054676275700331, + 0.005781103391200304, + 0.004639816470444202, + 0.0044647566974163055, + 0.008706970140337944, + 0.008924799039959908, + 0.009269781410694122, + 0.007953270338475704, + 0.012383764609694481, + 0.01674320362508297, + 0.012177841737866402, + 0.012853792868554592, + 0.03117392398416996, + 0.019533410668373108, + 0.036069054156541824, + 0.026137514039874077, + 0.009325223974883556, + 0.011121070012450218, + 0.012920654378831387, + 0.020800111815333366, + 0.023053061217069626, + 0.01857689395546913, + 0.01191666442900896, + 0.02175566740334034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.0125464191660285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818707026541233 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818707026541233, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818707026541233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05128125101327896, + 0.49966347217559814, + 0.023036981001496315, + 0.033615048974752426, + 0.03167532756924629, + 0.04763413593173027, + 0.03505316004157066, + 0.0646166205406189, + 0.0565619021654129, + 0.15686212480068207 + ], + "credit_map": [ + 0.005579679738730192, + 0.006734353490173817, + 0.005053780972957611, + 0.004682763479650021, + 0.009679730981588364, + 0.0064415824599564075, + 0.006491594947874546, + 0.006617771461606026, + 0.013089165091514587, + 0.006869858596473932, + 0.01907181553542614, + 0.01819596253335476, + 0.02222038060426712, + 0.2215803861618042, + 0.015954861417412758, + 0.18268105387687683, + 0.0015639212215319276, + 0.0027321302331984043, + 0.002197026973590255, + 0.002172474516555667, + 0.004243351984769106, + 0.0033845840953290462, + 0.0031336965039372444, + 0.003609797917306423, + 0.0018627216340973973, + 0.0029555605724453926, + 0.0018474686658009887, + 0.002642052248120308, + 0.0063337539322674274, + 0.006305707152932882, + 0.006111221853643656, + 0.005556564312428236, + 0.0026016118936240673, + 0.003309110412374139, + 0.0020569772459566593, + 0.00247580511495471, + 0.004622816573828459, + 0.006919455714523792, + 0.004509893711656332, + 0.005179658532142639, + 0.0046895951963961124, + 0.004976197611540556, + 0.006697927601635456, + 0.003954344894737005, + 0.009258083999156952, + 0.005280940793454647, + 0.008744223974645138, + 0.004032825585454702, + 0.001982214394956827, + 0.002156272530555725, + 0.0029998866375535727, + 0.0022450978867709637, + 0.004384834319353104, + 0.005509235430508852, + 0.009443433955311775, + 0.0063321893103420734, + 0.007205714471638203, + 0.010281025432050228, + 0.00616841297596693, + 0.005520286969840527, + 0.006837000139057636, + 0.008531038649380207, + 0.00828560907393694, + 0.0117875374853611, + 0.0050759753212332726, + 0.004892443772405386, + 0.003665252821519971, + 0.0050433422438800335, + 0.010377875529229641, + 0.008271683007478714, + 0.010222605429589748, + 0.009012729860842228, + 0.01079125888645649, + 0.01593633182346821, + 0.013128512538969517, + 0.020351160317659378, + 0.031488705426454544, + 0.023419464007019997, + 0.01679324358701706, + 0.024953460320830345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.01818707026541233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809477761387825 + ], + "opposite_logprob_tokens": [ + -3.2880947589874268 + ], + "logprob_reward": -0.03809477761387825, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809477761387825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04795286804437637, + 0.5359521508216858, + 0.03247502073645592, + 0.03727097436785698, + 0.019663969054818153, + 0.01973615400493145, + 0.04083257541060448, + 0.04705483093857765, + 0.05532039329409599, + 0.1637410670518875 + ], + "credit_map": [ + 0.004598691128194332, + 0.004223659168928862, + 0.005391902290284634, + 0.004098393488675356, + 0.008968779817223549, + 0.006084385793656111, + 0.008556238375604153, + 0.0060308147221803665, + 0.013780375011265278, + 0.011047332547605038, + 0.024056602269411087, + 0.0233988668769598, + 0.029337773099541664, + 0.22819717228412628, + 0.020504143089056015, + 0.18562984466552734, + 0.002035684883594513, + 0.002597005805000663, + 0.0024411596823483706, + 0.004604209214448929, + 0.0050876508466899395, + 0.005621830467134714, + 0.005007966887205839, + 0.005079514347016811, + 0.0025507083628326654, + 0.0023416271433234215, + 0.0029249864164739847, + 0.0026913173496723175, + 0.005931551568210125, + 0.005873478949069977, + 0.008126599714159966, + 0.006830702535808086, + 0.0016633628401905298, + 0.003918921574950218, + 0.001641100854612887, + 0.0018400661647319794, + 0.002431347733363509, + 0.002479241229593754, + 0.002714728470891714, + 0.002975200768560171, + 0.0018972846446558833, + 0.001924987998791039, + 0.003166366834193468, + 0.0016206527361646295, + 0.003820163430646062, + 0.002424017759039998, + 0.002666870364919305, + 0.0022158101201057434, + 0.0036736028268933296, + 0.002702210796996951, + 0.003014851361513138, + 0.0020868629217147827, + 0.00535516906529665, + 0.006986909080296755, + 0.012054501101374626, + 0.0049584656953811646, + 0.0031074380967766047, + 0.00395496841520071, + 0.0021273326128721237, + 0.003504664869979024, + 0.009947100654244423, + 0.007918321527540684, + 0.009779301472008228, + 0.006715702824294567, + 0.003920378163456917, + 0.010177262127399445, + 0.004842543043196201, + 0.006610126234591007, + 0.009734153747558594, + 0.005662337876856327, + 0.006753259338438511, + 0.007620332296937704, + 0.008503913879394531, + 0.008752540685236454, + 0.011176824569702148, + 0.015309343114495277, + 0.027535459026694298, + 0.03408733382821083, + 0.02023780532181263, + 0.03813783824443817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.03809477761387825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030406713485718 + ], + "opposite_logprob_tokens": [ + -2.350304126739502 + ], + "logprob_reward": -0.10030406713485718, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030406713485718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2711298167705536, + 0.3008795380592346, + 0.05569653585553169, + 0.04327557608485222, + 0.09975262731313705, + 0.0339721255004406, + 0.04720199853181839, + 0.043893709778785706, + 0.052460916340351105, + 0.05173718184232712 + ], + "credit_map": [ + 0.030950253829360008, + 0.029563799500465393, + 0.02626379393041134, + 0.027587061747908592, + 0.05897079035639763, + 0.029716428369283676, + 0.032464247196912766, + 0.03561343997716904, + 0.013636314310133457, + 0.009334716014564037, + 0.01644090563058853, + 0.013164057396352291, + 0.02567298151552677, + 0.11246803402900696, + 0.015294347889721394, + 0.09486819058656693, + 0.004659418947994709, + 0.004885219968855381, + 0.005697939079254866, + 0.0042109061032533646, + 0.011325616389513016, + 0.01109551452100277, + 0.006086534354835749, + 0.007735384162515402, + 0.004073410760611296, + 0.003925859462469816, + 0.004129758570343256, + 0.0033012644853442907, + 0.007257065735757351, + 0.007811008486896753, + 0.00555818947032094, + 0.0072190165519714355, + 0.007060648873448372, + 0.010508464649319649, + 0.005042712204158306, + 0.007104022428393364, + 0.016241637989878654, + 0.023211441934108734, + 0.016326958313584328, + 0.01425673346966505, + 0.0037718394305557013, + 0.0036147302016615868, + 0.0031380949076265097, + 0.002626317087560892, + 0.0059584625996649265, + 0.0038593406789004803, + 0.006824932526797056, + 0.004178409930318594, + 0.004488754086196423, + 0.005796876270323992, + 0.004456829745322466, + 0.003220696235075593, + 0.006267975550144911, + 0.007305077277123928, + 0.007852507755160332, + 0.007813281379640102, + 0.004690987057983875, + 0.004041503649204969, + 0.0034096096642315388, + 0.004320214502513409, + 0.008288474753499031, + 0.0063088987953960896, + 0.006558530032634735, + 0.006275493185967207, + 0.004903413355350494, + 0.006365600973367691, + 0.004532920196652412, + 0.005290256347507238, + 0.009994295425713062, + 0.007478363811969757, + 0.006397855933755636, + 0.0074982051737606525, + 0.0029746664222329855, + 0.004349563270807266, + 0.0026188495103269815, + 0.00513792410492897, + 0.00942892674356699, + 0.010948480106890202, + 0.0051299212500452995, + 0.0111488476395607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.10030406713485718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera slowly panning to the right, captures a scene in a deep, narrow alley. The frame focuses on a wall constructed of ancient blue bricks, moisture seeping through the cracks and dappled with patches of velvety moss of varying shades. A faint beam of natural light slants in from the end of the alley, illuminating the wall while the surroundings remain shrouded in deep shadow. A delicate hand, rounded knuckles, clad in a slightly oversized off-white sweater with slightly rolled-up sleeves, cautiously reaches into the frame. The child's fingers first gently touch the moist moss, then, with a touch of curiosity and wonder, the palm of the hand steadily and deliberately slides across the rough wall and soft moss, feeling the unique, cool, yet vibrant touch. Against the cool background, this light and this innocent gesture imbue the scene with a remarkable tranquility, concentration, and poetry.", + "seed": 29817, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_7/index_8494_seed_29817_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.052088703401386736, + "vlm_logprob": -0.052088703401386736, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.052088703401386736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0171623807400465 + ], + "opposite_logprob_tokens": [ + -4.392162322998047 + ], + "logprob_reward": -0.0171623807400465, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0171623807400465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16159749031066895, + 0.2311108410358429, + 0.24318724870681763, + 0.04829759523272514, + 0.04319482669234276, + 0.045209240168333054, + 0.023974556475877762, + 0.09911053627729416, + 0.04447140544652939, + 0.05984630435705185 + ], + "credit_map": [ + 0.025774411857128143, + 0.01434088684618473, + 0.03461521118879318, + 0.016851523891091347, + 0.01929515227675438, + 0.01970081590116024, + 0.015284677967429161, + 0.015734808519482613, + 0.023037808015942574, + 0.03554658219218254, + 0.022385675460100174, + 0.038867153227329254, + 0.019868288189172745, + 0.023062746971845627, + 0.0432429164648056, + 0.025099677965044975, + 0.056184276938438416, + 0.013598070479929447, + 0.01592315174639225, + 0.01317370031028986, + 0.03043297864496708, + 0.02399238385260105, + 0.023627249523997307, + 0.06625542044639587, + 0.004269575234502554, + 0.005867017433047295, + 0.006218902301043272, + 0.004731167107820511, + 0.007478026207536459, + 0.0062002213671803474, + 0.0057836067862808704, + 0.0077490778639912605, + 0.003196657868102193, + 0.004678911063820124, + 0.004143547732383013, + 0.004177025053650141, + 0.006891398690640926, + 0.007974812760949135, + 0.008165323175489902, + 0.003967153374105692, + 0.004160748329013586, + 0.003919300157576799, + 0.008319997228682041, + 0.0041416906751692295, + 0.005499235820025206, + 0.00683206832036376, + 0.006274000275880098, + 0.0060621993616223335, + 0.0017322339117527008, + 0.002148666884750128, + 0.001619843882508576, + 0.0023777703754603863, + 0.004287975374609232, + 0.0034714003559201956, + 0.004333650693297386, + 0.004003015346825123, + 0.02035549469292164, + 0.020575150847434998, + 0.004940668120980263, + 0.008431830443441868, + 0.011821769177913666, + 0.012976568192243576, + 0.01065219659358263, + 0.009356856346130371, + 0.005193063989281654, + 0.006930135656148195, + 0.0036133727990090847, + 0.006783950142562389, + 0.0031038932502269745, + 0.005982043221592903, + 0.005497644189745188, + 0.007367303594946861, + 0.00899796187877655, + 0.013009948655962944, + 0.005143031012266874, + 0.006171959452331066, + 0.005291355308145285, + 0.006933489814400673, + 0.00537828728556633, + 0.008920270949602127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0171623807400465 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790615178644657 + ], + "opposite_logprob_tokens": [ + -4.634790420532227 + ], + "logprob_reward": -0.009790615178644657, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790615178644657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10904379934072495, + 0.054593008011579514, + 0.22200480103492737, + 0.05326473340392113, + 0.04747890308499336, + 0.08989261090755463, + 0.06565500050783157, + 0.185944065451622, + 0.08908749371767044, + 0.0830356702208519 + ], + "credit_map": [ + 0.019848138093948364, + 0.0058294665068387985, + 0.027274418622255325, + 0.009646627120673656, + 0.0082539152354002, + 0.017872212454676628, + 0.011148025281727314, + 0.009170989505946636, + 0.009664556011557579, + 0.005856309086084366, + 0.00549505976960063, + 0.006613218225538731, + 0.00534582044929266, + 0.005835914984345436, + 0.005726877134293318, + 0.010055248625576496, + 0.06066610664129257, + 0.01200418546795845, + 0.01580957882106304, + 0.009423265233635902, + 0.016262536868453026, + 0.023877456784248352, + 0.011201026849448681, + 0.07276064157485962, + 0.0035209173802286386, + 0.007603021338582039, + 0.0038320126477628946, + 0.007507944013923407, + 0.008590519428253174, + 0.006201139185577631, + 0.008986314758658409, + 0.007022862788289785, + 0.004457742441445589, + 0.004445955157279968, + 0.0063234297558665276, + 0.004403047263622284, + 0.006123727187514305, + 0.00941424909979105, + 0.007156759034842253, + 0.005153991281986237, + 0.007304754573851824, + 0.011892200447618961, + 0.009001113474369049, + 0.017702972516417503, + 0.010072030127048492, + 0.008712279610335827, + 0.014777147211134434, + 0.010430107824504375, + 0.004980788566172123, + 0.005407134536653757, + 0.006439688615500927, + 0.00891084410250187, + 0.011422297917306423, + 0.01074613630771637, + 0.009628547355532646, + 0.008119554258883, + 0.028586165979504585, + 0.03605521842837334, + 0.014806708320975304, + 0.01995857246220112, + 0.021754585206508636, + 0.024750344455242157, + 0.01972133293747902, + 0.020311109721660614, + 0.007615013513714075, + 0.013019578531384468, + 0.012377477250993252, + 0.009057712741196156, + 0.007711210288107395, + 0.014927213080227375, + 0.013909663073718548, + 0.010469620116055012, + 0.009065854363143444, + 0.01064907293766737, + 0.010332217440009117, + 0.0090316291898489, + 0.010942935012280941, + 0.00867967028170824, + 0.014294094406068325, + 0.01004018820822239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.009790615178644657 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637804538011551 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637804538011551, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637804538011551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057016633450984955, + 0.07036510109901428, + 0.11035320162773132, + 0.05766316130757332, + 0.06532567739486694, + 0.040749017149209976, + 0.0742366760969162, + 0.1149652749300003, + 0.15932048857212067, + 0.25000473856925964 + ], + "credit_map": [ + 0.0059301285073161125, + 0.0048592546954751015, + 0.007933392189443111, + 0.007767696864902973, + 0.007513027172535658, + 0.008455580100417137, + 0.007096372079104185, + 0.007461181376129389, + 0.006321759894490242, + 0.009309480898082256, + 0.006358201149851084, + 0.009783447720110416, + 0.006106038112193346, + 0.008118572644889355, + 0.014122464694082737, + 0.01024513691663742, + 0.016576314345002174, + 0.009687392972409725, + 0.008837328292429447, + 0.012596912682056427, + 0.018132014200091362, + 0.01175797637552023, + 0.012878023087978363, + 0.01988723687827587, + 0.003828753950074315, + 0.010238571092486382, + 0.005346783436834812, + 0.007947683334350586, + 0.007863703183829784, + 0.00693347817286849, + 0.007939418777823448, + 0.007564767263829708, + 0.010889314115047455, + 0.007148974109441042, + 0.010568387806415558, + 0.006087619811296463, + 0.007184888701885939, + 0.00761176273226738, + 0.008630645461380482, + 0.00720408046618104, + 0.005012608598917723, + 0.004051876254379749, + 0.004939588252454996, + 0.0049804397858679295, + 0.005058949813246727, + 0.004638248588889837, + 0.005136494059115648, + 0.006930812261998653, + 0.0058175199665129185, + 0.0052599539048969746, + 0.0045658303424716, + 0.008485237136483192, + 0.014965943992137909, + 0.013801051303744316, + 0.01261213980615139, + 0.008729001507163048, + 0.01840951293706894, + 0.022179575636982918, + 0.010426747612655163, + 0.01314311008900404, + 0.014332355000078678, + 0.01321880891919136, + 0.011391968466341496, + 0.01186319813132286, + 0.018837634474039078, + 0.02708788774907589, + 0.012185986153781414, + 0.01786164380609989, + 0.013866713270545006, + 0.021213313564658165, + 0.0209585539996624, + 0.027308763936161995, + 0.03472910076379776, + 0.045567240566015244, + 0.03920556977391243, + 0.02972530946135521, + 0.016148334369063377, + 0.026658788323402405, + 0.024968529120087624, + 0.03300188481807709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.008637804538011551 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328481338918209 + ], + "opposite_logprob_tokens": [ + -3.773284912109375 + ], + "logprob_reward": -0.02328481338918209, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328481338918209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14612382650375366, + 0.20070089399814606, + 0.18131694197654724, + 0.03652067482471466, + 0.03859047219157219, + 0.035376064479351044, + 0.06038933992385864, + 0.09025367349386215, + 0.0784582793712616, + 0.13226987421512604 + ], + "credit_map": [ + 0.02670147269964218, + 0.012965263798832893, + 0.04649556428194046, + 0.011071713641285896, + 0.010660565458238125, + 0.021436940878629684, + 0.008748862892389297, + 0.00804342795163393, + 0.018436700105667114, + 0.04437777027487755, + 0.01904185675084591, + 0.03983495384454727, + 0.01725327968597412, + 0.016730954870581627, + 0.019477054476737976, + 0.02554832212626934, + 0.05043718218803406, + 0.010655500926077366, + 0.01354606356471777, + 0.007141782436519861, + 0.013295818120241165, + 0.01926792599260807, + 0.007754880469292402, + 0.05921778827905655, + 0.002744750352576375, + 0.004235657397657633, + 0.0031847816426306963, + 0.005876758135855198, + 0.004403389524668455, + 0.004011854529380798, + 0.007030677516013384, + 0.005032806657254696, + 0.005337818060070276, + 0.004502009134739637, + 0.00534607982262969, + 0.005240349564701319, + 0.003983793314546347, + 0.004879890941083431, + 0.005745795089751482, + 0.0035547357983887196, + 0.0031245630234479904, + 0.0039393287152051926, + 0.003176385536789894, + 0.007538631558418274, + 0.0046668993309140205, + 0.0033181272447109222, + 0.005530700087547302, + 0.00408143037930131, + 0.00405088160187006, + 0.0033632335253059864, + 0.006284706294536591, + 0.0049629914574325085, + 0.019031960517168045, + 0.008311569690704346, + 0.00758796650916338, + 0.006796029396355152, + 0.008301003836095333, + 0.008664719760417938, + 0.0050618769600987434, + 0.010892357677221298, + 0.01785956509411335, + 0.016371173784136772, + 0.012893587350845337, + 0.010209389962255955, + 0.004533364903181791, + 0.005390997976064682, + 0.004628611262887716, + 0.006044727750122547, + 0.018260912969708443, + 0.008285115472972393, + 0.015331599861383438, + 0.01598294824361801, + 0.03008885867893696, + 0.0384998582303524, + 0.007732834666967392, + 0.006556464359164238, + 0.009175012819468975, + 0.010975721292197704, + 0.011063295416533947, + 0.018177825957536697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.02328481338918209 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a deep, narrow alley with a wall made of ancient blue bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a delicate child's hand with rounded knuckles visible reaching into the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand wearing a slightly oversized off-white sweater with slightly rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slide its palm across the rough wall and soft moss after touching it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right while a faint beam of natural light slants in from the end of the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015679031610489 + ], + "opposite_logprob_tokens": [ + -1.7015678882598877 + ], + "logprob_reward": -0.2015679031610489, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015679031610489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21967899799346924, + 0.156759113073349, + 0.19220811128616333, + 0.07923591881990433, + 0.10939587652683258, + 0.06711768358945847, + 0.0367429256439209, + 0.08542399853467941, + 0.03470267727971077, + 0.018734732642769814 + ], + "credit_map": [ + 0.03461839258670807, + 0.017181413248181343, + 0.05602505803108215, + 0.021469810977578163, + 0.019094474613666534, + 0.03105561062693596, + 0.020270176231861115, + 0.019964048638939857, + 0.018196281045675278, + 0.02377222292125225, + 0.015282956883311272, + 0.027420632541179657, + 0.017228273674845695, + 0.01428490411490202, + 0.01667861081659794, + 0.023895222693681717, + 0.016065634787082672, + 0.022815559059381485, + 0.012638472020626068, + 0.033678777515888214, + 0.024590151384472847, + 0.020055171102285385, + 0.03623216971755028, + 0.02613217756152153, + 0.007419746834784746, + 0.010538616217672825, + 0.010621583089232445, + 0.008846478536725044, + 0.010258369147777557, + 0.010894943960011005, + 0.010080165229737759, + 0.01057601347565651, + 0.007624203804880381, + 0.010261633433401585, + 0.011733991093933582, + 0.012145419605076313, + 0.01355009339749813, + 0.017496468499302864, + 0.024779843166470528, + 0.011804216541349888, + 0.007549106143414974, + 0.007265492342412472, + 0.007101719733327627, + 0.009208263829350471, + 0.010118708945810795, + 0.009036531671881676, + 0.009588533081114292, + 0.007249326445162296, + 0.002727671293541789, + 0.0021929284557700157, + 0.0032633927185088396, + 0.004038677085191011, + 0.008277715183794498, + 0.006959710270166397, + 0.005114903673529625, + 0.004167927894741297, + 0.015149331651628017, + 0.017859630286693573, + 0.004843704402446747, + 0.008560460060834885, + 0.011595477350056171, + 0.009311317466199398, + 0.00979030691087246, + 0.00831377413123846, + 0.003190758405253291, + 0.0030594272539019585, + 0.0033626791555434465, + 0.007277717813849449, + 0.0037313648499548435, + 0.0031461180187761784, + 0.004703665152192116, + 0.006230945698916912, + 0.0017519474495202303, + 0.0021182734053581953, + 0.0013611983740702271, + 0.0026637353003025055, + 0.002676293021067977, + 0.0025622353423386812, + 0.002935838419944048, + 0.002665212145075202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.2015679031610489 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_8/index_tensor([8498, 8498])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_8/index_tensor([8498, 8498])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..467313261d4a9019d60c41c32ad153b33897214e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_8/index_tensor([8498, 8498])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stark black-and-white, slightly low-angle close-up, held static, captures a classic portrait of a large American bullfrog (Lithobates catesbeianus) in the midst of its intense bark. High-contrast lighting from a single source dramatically illuminates the scene, sculpting the amphibian from a dark, completely out-of-focus background. As the frog calls, its two vocal sacs beneath its jaws inflate steadily and symmetrically, expanding like pale, taut balloons. The harsh light meticulously renders the shimmering, bumpy texture of the frog's moist skin with brilliant specular highlights, while casting deep, velvety shadows that accentuate its form. The sacs themselves become almost pure white, their thin, stretched skin contrasting sharply against the darker hues of the frog's body. The entire work emphasizes the primal nature of the performance, creating an atmosphere of visceral, focused concentration that is both beautiful and intense.", + "seed": 7359, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_8/index_8498_seed_7359_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.29110081465914844, + "vlm_logprob": -0.29110081465914844, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.29110081465914844, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4396217167377472 + ], + "opposite_logprob_tokens": [ + -1.0646216869354248 + ], + "logprob_reward": -0.4396217167377472, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4396217167377472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14111773669719696, + 0.03922554478049278, + 0.2535265386104584, + 0.08085979521274567, + 0.09133148938417435, + 0.02523757703602314, + 0.14808215200901031, + 0.11144974082708359, + 0.051349155604839325, + 0.05782027915120125 + ], + "credit_map": [ + 0.027085402980446815, + 0.01143240462988615, + 0.008300689049065113, + 0.01686173677444458, + 0.010691898874938488, + 0.02301156148314476, + 0.018835896626114845, + 0.024898147210478783, + 0.005645412020385265, + 0.00429240707308054, + 0.003071341197937727, + 0.0059974282048642635, + 0.0039720963686704636, + 0.004989135079085827, + 0.0049054487608373165, + 0.0063522798009216785, + 0.03166981041431427, + 0.02063627354800701, + 0.01719619147479534, + 0.045045215636491776, + 0.019655155017971992, + 0.037946105003356934, + 0.0529150664806366, + 0.028462715446949005, + 0.012589499354362488, + 0.005063909105956554, + 0.0023452420718967915, + 0.013089282438158989, + 0.005648055579513311, + 0.02117978222668171, + 0.007551170885562897, + 0.01339285634458065, + 0.006618191488087177, + 0.0057152556255459785, + 0.004725305363535881, + 0.02417495660483837, + 0.005817394237965345, + 0.00649825157597661, + 0.01096704974770546, + 0.026815081015229225, + 0.002496992703527212, + 0.0033435779623687267, + 0.0022434864658862352, + 0.0037592912558466196, + 0.0030187189113348722, + 0.0022330007050186396, + 0.004643309861421585, + 0.0034991998691111803, + 0.006965096108615398, + 0.008187578059732914, + 0.004511851351708174, + 0.05460001155734062, + 0.0036944018211215734, + 0.008858971297740936, + 0.008893108926713467, + 0.052371133118867874, + 0.006881444249302149, + 0.008822866715490818, + 0.007341149728745222, + 0.02753555029630661, + 0.008799362927675247, + 0.010983786545693874, + 0.017143158242106438, + 0.02394242212176323, + 0.004031027667224407, + 0.0032612497452646494, + 0.0025357266422361135, + 0.014904591254889965, + 0.0029541654512286186, + 0.004048740025609732, + 0.005795588716864586, + 0.013818062841892242, + 0.0061794789507985115, + 0.008015346713364124, + 0.00491132028400898, + 0.011637518182396889, + 0.007297253701835871, + 0.00689557334408164, + 0.006786041893064976, + 0.006097747478634119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.4396217167377472 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006021216977387667 + ], + "opposite_logprob_tokens": [ + -5.131021022796631 + ], + "logprob_reward": -0.006021216977387667, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006021216977387667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14785438776016235, + 0.04906737431883812, + 0.26166439056396484, + 0.02757561206817627, + 0.024666035547852516, + 0.030200332403182983, + 0.21436341106891632, + 0.18156199157238007, + 0.03000837191939354, + 0.033038027584552765 + ], + "credit_map": [ + 0.026673370972275734, + 0.008658579550683498, + 0.008522325195372105, + 0.021270619705319405, + 0.011518755927681923, + 0.025905072689056396, + 0.019779905676841736, + 0.0255257710814476, + 0.007089601829648018, + 0.0027169822715222836, + 0.0022521375212818384, + 0.008868474513292313, + 0.004160355310887098, + 0.00820101983845234, + 0.0063628507778048515, + 0.009415953420102596, + 0.02692006155848503, + 0.02080680988729, + 0.016849471256136894, + 0.04466846585273743, + 0.023486381396651268, + 0.03615899384021759, + 0.056018490344285965, + 0.0367557592689991, + 0.005213832017034292, + 0.0017156715039163828, + 0.0012800516560673714, + 0.0028795271646231413, + 0.001412825658917427, + 0.008848476223647594, + 0.0021970318630337715, + 0.004028198774904013, + 0.0025100144557654858, + 0.0022402778267860413, + 0.001493070856668055, + 0.005631712265312672, + 0.002553621307015419, + 0.0022223880514502525, + 0.003968674689531326, + 0.004046276677399874, + 0.0027538875583559275, + 0.002318258862942457, + 0.0017114680958911777, + 0.007755522150546312, + 0.0018175584264099598, + 0.0020752372220158577, + 0.003972793463617563, + 0.007795608602464199, + 0.009210716001689434, + 0.012322665192186832, + 0.006829192861914635, + 0.0819588154554367, + 0.004222295247018337, + 0.01250191405415535, + 0.013307680375874043, + 0.07401016354560852, + 0.010119720362126827, + 0.0163620226085186, + 0.013659434393048286, + 0.046355340629816055, + 0.020172541961073875, + 0.016756152734160423, + 0.02367221564054489, + 0.03446458280086517, + 0.0035835441667586565, + 0.0026177209801971912, + 0.0026772678829729557, + 0.007661616429686546, + 0.0031468782108277082, + 0.00231951498426497, + 0.0034907611552625895, + 0.00451107369735837, + 0.004271445795893669, + 0.0030764632392674685, + 0.002815169980749488, + 0.007785118650645018, + 0.00335017591714859, + 0.00439446372911334, + 0.003567035309970379, + 0.0037781596183776855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.006021216977387667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898403525352478 + ], + "opposite_logprob_tokens": [ + -0.523403525352478 + ], + "logprob_reward": -0.898403525352478, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898403525352478, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03159051015973091, + 0.033449165523052216, + 0.041545286774635315, + 0.026015629991889, + 0.2045348882675171, + 0.10910621285438538, + 0.08066317439079285, + 0.2785600423812866, + 0.17799687385559082, + 0.01653820276260376 + ], + "credit_map": [ + 0.004678930621594191, + 0.0021760568488389254, + 0.0018534483388066292, + 0.004103461746126413, + 0.0028061969205737114, + 0.005503442604094744, + 0.00568176107481122, + 0.004787209909409285, + 0.005637911148369312, + 0.00188852206338197, + 0.002348527079448104, + 0.004403325263410807, + 0.0020577511750161648, + 0.006085539702326059, + 0.004055934026837349, + 0.0069716558791697025, + 0.004898055922240019, + 0.0035807981621474028, + 0.0027526565827429295, + 0.007028607185930014, + 0.003422355744987726, + 0.005570864770561457, + 0.009870960377156734, + 0.004420990124344826, + 0.003307464998215437, + 0.002792773535475135, + 0.001936594839207828, + 0.0028447099030017853, + 0.002819652436301112, + 0.004549663048237562, + 0.003429171396419406, + 0.004335599485784769, + 0.010976252146065235, + 0.012013141997158527, + 0.008290325291454792, + 0.07057756185531616, + 0.011657610535621643, + 0.011450088582932949, + 0.02135232463479042, + 0.05821758136153221, + 0.009806722402572632, + 0.012483608908951283, + 0.009280389174818993, + 0.015651315450668335, + 0.011866673827171326, + 0.010774864815175533, + 0.02291363663971424, + 0.016329003497958183, + 0.005571295041590929, + 0.0063647134229540825, + 0.005058647599071264, + 0.024608895182609558, + 0.005546283908188343, + 0.004931952804327011, + 0.006513748783618212, + 0.022067641839385033, + 0.017053555697202682, + 0.02110319957137108, + 0.021073153242468834, + 0.06815030425786972, + 0.03345249220728874, + 0.022647716104984283, + 0.04297754168510437, + 0.052102092653512955, + 0.00675536971539259, + 0.009964827448129654, + 0.007836868055164814, + 0.06847360730171204, + 0.005248012486845255, + 0.010523244738578796, + 0.009902950376272202, + 0.05929197371006012, + 0.0014171890215948224, + 0.0012480414006859064, + 0.0012788869207724929, + 0.0045601241290569305, + 0.0020944522693753242, + 0.0018996966537088156, + 0.002350283320993185, + 0.0016895284643396735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.898403525352478 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003742665052414 + ], + "opposite_logprob_tokens": [ + -2.350374221801758 + ], + "logprob_reward": -0.1003742665052414, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003742665052414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07748435437679291, + 0.024119790643453598, + 0.03225722536444664, + 0.02864394150674343, + 0.09615880250930786, + 0.06577062606811523, + 0.0791008248925209, + 0.03050103969871998, + 0.5431673526763916, + 0.02279605343937874 + ], + "credit_map": [ + 0.01071335468441248, + 0.0052109514363110065, + 0.004198119509965181, + 0.015509682707488537, + 0.00458931690081954, + 0.012935194186866283, + 0.014339135028421879, + 0.00998859852552414, + 0.0028286983724683523, + 0.0017269381787627935, + 0.0016187389846891165, + 0.005246559623628855, + 0.001999145606532693, + 0.004074697382748127, + 0.0035964264534413815, + 0.003028586506843567, + 0.003737099701538682, + 0.0025622998364269733, + 0.002144942991435528, + 0.005566612351685762, + 0.0029380363412201405, + 0.0042763203382492065, + 0.007069785613566637, + 0.003962124232202768, + 0.004262984264642, + 0.003227440407499671, + 0.0019550244323909283, + 0.005188229493796825, + 0.0037042214535176754, + 0.0036761390510946512, + 0.0027991628739982843, + 0.0038307399954646826, + 0.0065823765471577644, + 0.007535097189247608, + 0.005500142928212881, + 0.03000384010374546, + 0.006938057951629162, + 0.00657672481611371, + 0.012678289785981178, + 0.020344270393252373, + 0.00529526500031352, + 0.005538958590477705, + 0.004891596734523773, + 0.015756139531731606, + 0.005006223451346159, + 0.004863722715526819, + 0.011299656704068184, + 0.013119066134095192, + 0.0037273496855050325, + 0.004448430147022009, + 0.002641817322000861, + 0.028633035719394684, + 0.0035368110984563828, + 0.004467809107154608, + 0.005541997496038675, + 0.02610357664525509, + 0.002635313430801034, + 0.0032864343374967575, + 0.002236793050542474, + 0.005726379342377186, + 0.004444377031177282, + 0.0024048311170190573, + 0.00566639332100749, + 0.004100518766790628, + 0.02021111734211445, + 0.02856392227113247, + 0.017406418919563293, + 0.21027100086212158, + 0.014554312452673912, + 0.03312021121382713, + 0.0337555818259716, + 0.1852847784757614, + 0.003770082024857402, + 0.001991427270695567, + 0.0014659017324447632, + 0.005708335433155298, + 0.0018512848764657974, + 0.0025050134863704443, + 0.0026797691825777292, + 0.002824239432811737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.1003742665052414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083347722887993 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083347722887993, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083347722887993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10869205743074417, + 0.06374939531087875, + 0.1426965445280075, + 0.02895473688840866, + 0.016549037769436836, + 0.019674230366945267, + 0.23060305416584015, + 0.16548708081245422, + 0.16832561790943146, + 0.055268190801143646 + ], + "credit_map": [ + 0.012762202881276608, + 0.00750915938988328, + 0.005956672132015228, + 0.022568294778466225, + 0.007928401231765747, + 0.017568588256835938, + 0.019467169418931007, + 0.014931578189134598, + 0.010095682926476002, + 0.005114869214594364, + 0.004869016353040934, + 0.010807894170284271, + 0.00482403626665473, + 0.010750223882496357, + 0.008004168048501015, + 0.009283512830734253, + 0.019475005567073822, + 0.011823625303804874, + 0.008479191921651363, + 0.02961936965584755, + 0.015866758301854134, + 0.02368283085525036, + 0.016689511016011238, + 0.017060264945030212, + 0.005150271113961935, + 0.002279613632708788, + 0.0015650015557184815, + 0.003872999455779791, + 0.00241460045799613, + 0.006827984470874071, + 0.0023860414512455463, + 0.004458227660506964, + 0.0020413403399288654, + 0.0017692457186058164, + 0.001677494845353067, + 0.0022531880531460047, + 0.0017488501034677029, + 0.001734539750032127, + 0.002994365757331252, + 0.0023300140164792538, + 0.00223491620272398, + 0.002049492672085762, + 0.001608458231203258, + 0.0037211235612630844, + 0.0014285234501585364, + 0.0017556162783876061, + 0.0035062276292592287, + 0.0033698740880936384, + 0.009693923406302929, + 0.013267213478684425, + 0.007326414808630943, + 0.0879015177488327, + 0.0043758200481534, + 0.01363682933151722, + 0.01423501968383789, + 0.08016635477542877, + 0.011960883624851704, + 0.015293731354176998, + 0.012123086489737034, + 0.04166324436664581, + 0.018434923142194748, + 0.015399647876620293, + 0.020108018070459366, + 0.03050355799496174, + 0.0059683904983103275, + 0.009082251228392124, + 0.00478731095790863, + 0.06661372631788254, + 0.0034476667642593384, + 0.010157152079045773, + 0.00980629213154316, + 0.05846284329891205, + 0.005720991175621748, + 0.007941006682813168, + 0.005302944220602512, + 0.00915458146482706, + 0.005872378125786781, + 0.006451793015003204, + 0.006316336803138256, + 0.0085081672295928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.011083347722887993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "stark black-and-white, slightly low-angle close-up, held static, captures a classic portrait of a large American bullfrog (Lithobates catesbeianus) in the midst of its intense bark. High-contrast lighting from a single source dramatically illuminates the scene, sculpting the amphibian from a dark, completely out-of-focus background. As the frog calls, its two vocal sacs beneath its jaws inflate steadily and symmetrically, expanding like pale, taut balloons. The harsh light meticulously renders the shimmering, bumpy texture of the frog's moist skin with brilliant specular highlights, while casting deep, velvety shadows that accentuate its form. The sacs themselves become almost pure white, their thin, stretched skin contrasting sharply against the darker hues of the frog's body. The entire work emphasizes the primal nature of the performance, creating an atmosphere of visceral, focused concentration that is both beautiful and intense.", + "seed": 7360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_8/index_8498_seed_7360_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07483382541686297, + "vlm_logprob": -0.07483382541686297, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.07483382541686297, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2708823084831238 + ], + "opposite_logprob_tokens": [ + -1.5208823680877686 + ], + "logprob_reward": -0.2708823084831238, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2708823084831238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11327103525400162, + 0.10930079966783524, + 0.047023165971040726, + 0.06552407145500183, + 0.11742641776800156, + 0.1252647042274475, + 0.2330205738544464, + 0.04935427010059357, + 0.05611327663064003, + 0.08370165526866913 + ], + "credit_map": [ + 0.012870711274445057, + 0.013105593621730804, + 0.010763793252408504, + 0.009865627624094486, + 0.021927332505583763, + 0.01496900711208582, + 0.01600593887269497, + 0.013763041235506535, + 0.0108887804672122, + 0.009192856028676033, + 0.017069581896066666, + 0.009917562827467918, + 0.015586192719638348, + 0.016646306961774826, + 0.015217075124382973, + 0.014782450161874294, + 0.009419766254723072, + 0.004229056648910046, + 0.0050261057913303375, + 0.004216345492750406, + 0.004790221806615591, + 0.007940773852169514, + 0.004469596315175295, + 0.006931306328624487, + 0.00935740489512682, + 0.00515523087233305, + 0.00721904868260026, + 0.004867881536483765, + 0.010038657113909721, + 0.009907969273626804, + 0.010420581325888634, + 0.008557304739952087, + 0.00840387586504221, + 0.0165305994451046, + 0.008129854686558247, + 0.012923257425427437, + 0.025090182200074196, + 0.013996879570186138, + 0.01801416464149952, + 0.014337616041302681, + 0.017880069091916084, + 0.009418563917279243, + 0.00928918644785881, + 0.012138649821281433, + 0.01969389244914055, + 0.021498609334230423, + 0.0168890580534935, + 0.018456680700182915, + 0.03857593238353729, + 0.013064631260931492, + 0.01603669673204422, + 0.02994893118739128, + 0.01841612718999386, + 0.03879038617014885, + 0.04186871647834778, + 0.03631915524601936, + 0.005075895227491856, + 0.003290509572252631, + 0.005279377102851868, + 0.004985345061868429, + 0.008772415108978748, + 0.006432716269046068, + 0.00912008062005043, + 0.006397936027497053, + 0.007532288320362568, + 0.005249409936368465, + 0.007245820946991444, + 0.006061311345547438, + 0.00668391352519393, + 0.0070744953118264675, + 0.008148576132953167, + 0.008117467164993286, + 0.008401434868574142, + 0.007790536619722843, + 0.017520548775792122, + 0.007354443892836571, + 0.009815988130867481, + 0.009833014570176601, + 0.01301613450050354, + 0.00996956042945385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.2708823084831238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002300951164215803 + ], + "opposite_logprob_tokens": [ + -6.12730073928833 + ], + "logprob_reward": -0.002300951164215803, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002300951164215803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.160738006234169, + 0.05544281750917435, + 0.10324707627296448, + 0.2825146019458771, + 0.0399702712893486, + 0.11744377017021179, + 0.09400270134210587, + 0.04686272144317627, + 0.028652969747781754, + 0.07112514227628708 + ], + "credit_map": [ + 0.028545521199703217, + 0.01471715047955513, + 0.013605617918074131, + 0.009797319769859314, + 0.02447841875255108, + 0.025549160316586494, + 0.01921788975596428, + 0.024826908484101295, + 0.0051203640177845955, + 0.006247553043067455, + 0.008948019705712795, + 0.004641542676836252, + 0.00805829931050539, + 0.008213519118726254, + 0.006093986798077822, + 0.008119525387883186, + 0.01005050353705883, + 0.013188564218580723, + 0.008958392776548862, + 0.016035806387662888, + 0.015559977851808071, + 0.010908144526183605, + 0.01657991297543049, + 0.011965766549110413, + 0.020801233127713203, + 0.022421488538384438, + 0.027684614062309265, + 0.03408467769622803, + 0.05667334049940109, + 0.02764408476650715, + 0.05751452222466469, + 0.03569059073925018, + 0.0032371601555496454, + 0.0032994470093399286, + 0.003635525470599532, + 0.004171522334218025, + 0.007297059055417776, + 0.005364402662962675, + 0.007905583828687668, + 0.0050595710054039955, + 0.013393141329288483, + 0.00795106403529644, + 0.011929649859666824, + 0.013828017748892307, + 0.013838838785886765, + 0.019566837698221207, + 0.018200572580099106, + 0.01873563602566719, + 0.017240585759282112, + 0.007942247204482555, + 0.008092718198895454, + 0.008922522887587547, + 0.012389584444463253, + 0.0160208810120821, + 0.010368223302066326, + 0.01302593108266592, + 0.0072282468900084496, + 0.00311309308744967, + 0.008356726728379726, + 0.003556116484105587, + 0.004413836169987917, + 0.0073182652704417706, + 0.006105841137468815, + 0.006770591251552105, + 0.0030087337363511324, + 0.0020695265848189592, + 0.006017095409333706, + 0.003094212617725134, + 0.0030253897421061993, + 0.0038625490851700306, + 0.0035746870562434196, + 0.004000773187726736, + 0.011317674070596695, + 0.00523713743314147, + 0.011982626281678677, + 0.006020642351359129, + 0.006656567100435495, + 0.01233326829969883, + 0.008673193864524364, + 0.008904027752578259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.002300951164215803 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003893703222275 + ], + "opposite_logprob_tokens": [ + -2.6950390338897705 + ], + "logprob_reward": -0.07003893703222275, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003893703222275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0871383398771286, + 0.03552853316068649, + 0.07295311242341995, + 0.3765629827976227, + 0.071617990732193, + 0.052752695977687836, + 0.12280405312776566, + 0.1000794768333435, + 0.05585719272494316, + 0.02470564842224121 + ], + "credit_map": [ + 0.010231719352304935, + 0.006814781576395035, + 0.009509881027042866, + 0.007448576856404543, + 0.014109843410551548, + 0.012680686078965664, + 0.01440823171287775, + 0.01193461287766695, + 0.0055467900820076466, + 0.0027187331579625607, + 0.0066604153253138065, + 0.0029243342578411102, + 0.0035115210339426994, + 0.004840793553739786, + 0.004165703896433115, + 0.0051602390594780445, + 0.0065567390993237495, + 0.009523110464215279, + 0.00782149750739336, + 0.009765167720615864, + 0.010907208546996117, + 0.008203227072954178, + 0.011233839206397533, + 0.008942321874201298, + 0.03001752495765686, + 0.031079476699233055, + 0.03593287989497185, + 0.05340952426195145, + 0.05714360997080803, + 0.04002871736884117, + 0.0807126984000206, + 0.04823852330446243, + 0.00549356872215867, + 0.007057491224259138, + 0.011388706974685192, + 0.0062285130843520164, + 0.012738971039652824, + 0.008307777345180511, + 0.010498655028641224, + 0.009904303587973118, + 0.0067922500893473625, + 0.0037451847456395626, + 0.00571970222517848, + 0.006901389919221401, + 0.006245273165404797, + 0.008760587312281132, + 0.0071793911047279835, + 0.007408913224935532, + 0.024552511051297188, + 0.007075134199112654, + 0.011776583269238472, + 0.010290313512086868, + 0.012172500602900982, + 0.024565985426306725, + 0.013546700589358807, + 0.01882430911064148, + 0.012194892391562462, + 0.011541576124727726, + 0.008775326423346996, + 0.01303157303482294, + 0.015244201757013798, + 0.01156697142869234, + 0.015345371328294277, + 0.012379561550915241, + 0.005223322659730911, + 0.004371063783764839, + 0.00677265040576458, + 0.007344074547290802, + 0.009407573379576206, + 0.006944663822650909, + 0.00816648080945015, + 0.007627359591424465, + 0.004034473095089197, + 0.0019940852653235197, + 0.0038858847692608833, + 0.0024585481733083725, + 0.002541316905990243, + 0.004098293371498585, + 0.0024871029891073704, + 0.0032059422228485346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.07003893703222275 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331206575036049 + ], + "opposite_logprob_tokens": [ + -3.7733120918273926 + ], + "logprob_reward": -0.02331206575036049, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331206575036049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050445087254047394, + 0.04740041866898537, + 0.10811852663755417, + 0.16516515612602234, + 0.1598396599292755, + 0.056077778339385986, + 0.1515655666589737, + 0.10931088775396347, + 0.12301099300384521, + 0.02906595915555954 + ], + "credit_map": [ + 0.005217546131461859, + 0.005983355455100536, + 0.007134546060115099, + 0.0036812899634242058, + 0.007732845377177, + 0.006692094728350639, + 0.007419796660542488, + 0.006583611946552992, + 0.005493958946317434, + 0.0038951097521930933, + 0.005667596589773893, + 0.004991840571165085, + 0.006183881312608719, + 0.007099939044564962, + 0.007752048317342997, + 0.006316041573882103, + 0.014470291323959827, + 0.010742951184511185, + 0.01704232208430767, + 0.008749629370868206, + 0.014133574441075325, + 0.011854917742311954, + 0.01750222221016884, + 0.013622621074318886, + 0.010702650062739849, + 0.012966388836503029, + 0.013777313753962517, + 0.017286093905568123, + 0.03885394707322121, + 0.017885293811559677, + 0.031823765486478806, + 0.021869691088795662, + 0.010483172722160816, + 0.02010148949921131, + 0.013211563229560852, + 0.02184596285223961, + 0.03451445698738098, + 0.01673506386578083, + 0.021919945254921913, + 0.021027998998761177, + 0.0035155012737959623, + 0.004707904998213053, + 0.0062140510417521, + 0.006856538355350494, + 0.010170567780733109, + 0.006438601296395063, + 0.01054258644580841, + 0.007632026448845863, + 0.023596322163939476, + 0.009152243845164776, + 0.010790527798235416, + 0.021159306168556213, + 0.01566348597407341, + 0.024398107081651688, + 0.022945910692214966, + 0.02385966293513775, + 0.010142188519239426, + 0.011557408608496189, + 0.016862470656633377, + 0.013640877790749073, + 0.016479983925819397, + 0.012091591022908688, + 0.016141897067427635, + 0.012394476681947708, + 0.013412919826805592, + 0.011803271248936653, + 0.011178646236658096, + 0.01668657176196575, + 0.022252215072512627, + 0.014127342961728573, + 0.018744952976703644, + 0.014805079437792301, + 0.005728192161768675, + 0.0024784428533166647, + 0.0030661269556730986, + 0.002220290247350931, + 0.003565577557310462, + 0.005081890616565943, + 0.002670247806236148, + 0.00425519235432148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.02331206575036049 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634864654392004 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634864654392004, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634864654392004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10226868093013763, + 0.09593524038791656, + 0.0803113579750061, + 0.13254804909229279, + 0.02453581988811493, + 0.11490979790687561, + 0.1700402945280075, + 0.049856822937726974, + 0.025251463055610657, + 0.20434245467185974 + ], + "credit_map": [ + 0.011589165776968002, + 0.011907177045941353, + 0.019344858825206757, + 0.005123233888298273, + 0.011173108592629433, + 0.015678146854043007, + 0.014161649160087109, + 0.013291341252624989, + 0.009170476347208023, + 0.005829467438161373, + 0.02933814376592636, + 0.006864082999527454, + 0.006568417884409428, + 0.01326275896281004, + 0.01113701518625021, + 0.013764875009655952, + 0.01598602533340454, + 0.004634998273104429, + 0.014609932899475098, + 0.004083751235157251, + 0.006696269381791353, + 0.012398390099406242, + 0.010123440995812416, + 0.01177855022251606, + 0.011704635806381702, + 0.011284193955361843, + 0.013716246001422405, + 0.018580447882413864, + 0.018966784700751305, + 0.0146295465528965, + 0.026745570823550224, + 0.01692062057554722, + 0.003687732620164752, + 0.0016857037553563714, + 0.001599118928425014, + 0.0023242533206939697, + 0.0030455433297902346, + 0.004753998480737209, + 0.0035707983188331127, + 0.003868670668452978, + 0.015036209486424923, + 0.007568554952740669, + 0.010996447876095772, + 0.014967825263738632, + 0.013255295343697071, + 0.017968274652957916, + 0.01824912615120411, + 0.01686806231737137, + 0.04324967414140701, + 0.010355674661695957, + 0.01070933137089014, + 0.010442916303873062, + 0.012617225758731365, + 0.035773973912000656, + 0.016955668106675148, + 0.02993583492934704, + 0.00954117625951767, + 0.0025226864963769913, + 0.006143698003143072, + 0.00361026544123888, + 0.005186406895518303, + 0.009119740687310696, + 0.006067684851586819, + 0.007665163837373257, + 0.002430944237858057, + 0.0025028588715940714, + 0.0027293674647808075, + 0.0030140019953250885, + 0.004070112481713295, + 0.0031089424155652523, + 0.004484705626964569, + 0.002910530660301447, + 0.03624040260910988, + 0.009878387674689293, + 0.057564061135053635, + 0.00970975961536169, + 0.01376272551715374, + 0.03476109355688095, + 0.012555723078548908, + 0.029870320111513138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.007634864654392004 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stark black-and-white, slightly low-angle close-up, held static, captures a classic portrait of a large American bullfrog (Lithobates catesbeianus) in the midst of its intense bark. High-contrast lighting from a single source dramatically illuminates the scene, sculpting the amphibian from a dark, completely out-of-focus background. As the frog calls, its two vocal sacs beneath its jaws inflate steadily and symmetrically, expanding like pale, taut balloons. The harsh light meticulously renders the shimmering, bumpy texture of the frog's moist skin with brilliant specular highlights, while casting deep, velvety shadows that accentuate its form. The sacs themselves become almost pure white, their thin, stretched skin contrasting sharply against the darker hues of the frog's body. The entire work emphasizes the primal nature of the performance, creating an atmosphere of visceral, focused concentration that is both beautiful and intense.", + "seed": 7361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_8/index_8498_seed_7361_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15693529825657607, + "vlm_logprob": -0.15693529825657607, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.15693529825657607, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3681708872318268 + ], + "opposite_logprob_tokens": [ + -1.2431708574295044 + ], + "logprob_reward": -0.3681708872318268, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3681708872318268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18135365843772888, + 0.08585791289806366, + 0.039359282702207565, + 0.18708284199237823, + 0.036679379642009735, + 0.20462392270565033, + 0.027382448315620422, + 0.034992191940546036, + 0.0332106314599514, + 0.1694577932357788 + ], + "credit_map": [ + 0.019265471026301384, + 0.01948496513068676, + 0.017825646325945854, + 0.024033881723880768, + 0.023593440651893616, + 0.021869737654924393, + 0.02769109047949314, + 0.027589423581957817, + 0.010885403491556644, + 0.005167175084352493, + 0.006992311216890812, + 0.007621513679623604, + 0.010077462531626225, + 0.010439169593155384, + 0.021938329562544823, + 0.012736545875668526, + 0.004341739229857922, + 0.0028304625302553177, + 0.0039924620650708675, + 0.0031567488331347704, + 0.004903215914964676, + 0.005313041154295206, + 0.007927299477159977, + 0.006894312333315611, + 0.028457773849368095, + 0.009560858830809593, + 0.01755465380847454, + 0.013905744068324566, + 0.015093165449798107, + 0.021679434925317764, + 0.05378787964582443, + 0.027043331414461136, + 0.004731966648250818, + 0.0021839526016265154, + 0.002738943789154291, + 0.003974535036832094, + 0.0033201409969478846, + 0.004737844690680504, + 0.009283367544412613, + 0.005708626005798578, + 0.020382845774292946, + 0.023394979536533356, + 0.02064630389213562, + 0.020321901887655258, + 0.02657538466155529, + 0.023457372561097145, + 0.04307657480239868, + 0.026768557727336884, + 0.004122327081859112, + 0.0016877768794074655, + 0.0019513575825840235, + 0.0023896668571978807, + 0.003434358397498727, + 0.003839442739263177, + 0.0060915048234164715, + 0.003866012440994382, + 0.004766346421092749, + 0.0024503048043698072, + 0.003815285861492157, + 0.0028516086749732494, + 0.0029562253039330244, + 0.0047818515449762344, + 0.007848745211958885, + 0.005521825514733791, + 0.004301676992326975, + 0.002592738252133131, + 0.0030671649146825075, + 0.0037476809229701757, + 0.0033098163548856974, + 0.003736641025170684, + 0.006829520221799612, + 0.005625393241643906, + 0.016100017353892326, + 0.01245149690657854, + 0.015516191720962524, + 0.013665400445461273, + 0.023243272677063942, + 0.02048887498676777, + 0.04456864297389984, + 0.023423900827765465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.3681708872318268 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00417480431497097 + ], + "opposite_logprob_tokens": [ + -5.504174709320068 + ], + "logprob_reward": -0.00417480431497097, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00417480431497097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24808533489704132, + 0.1184144914150238, + 0.06200626492500305, + 0.034703221172094345, + 0.0653802827000618, + 0.31410038471221924, + 0.02813384309411049, + 0.03586381673812866, + 0.027626393362879753, + 0.0656859502196312 + ], + "credit_map": [ + 0.031239133328199387, + 0.029569925740361214, + 0.026896750554442406, + 0.021400758996605873, + 0.030511226505041122, + 0.033097874373197556, + 0.036657948046922684, + 0.03871171176433563, + 0.012898409739136696, + 0.008052160032093525, + 0.011309877969324589, + 0.012692661955952644, + 0.014857728034257889, + 0.014421166852116585, + 0.026746846735477448, + 0.017435641959309578, + 0.008899074047803879, + 0.004998709540814161, + 0.007774441037327051, + 0.0061788843013346195, + 0.007181649096310139, + 0.007998419925570488, + 0.009055562317371368, + 0.009919526986777782, + 0.004078959580510855, + 0.0022472566924989223, + 0.0026635625399649143, + 0.0028587353881448507, + 0.0039521376602351665, + 0.004772126208990812, + 0.009030080400407314, + 0.005100365728139877, + 0.007545060478150845, + 0.0037500793114304543, + 0.005416258703917265, + 0.006546441465616226, + 0.007627243176102638, + 0.008908584713935852, + 0.01657397486269474, + 0.009012633934617043, + 0.03726275637745857, + 0.0303945429623127, + 0.03269554302096367, + 0.0312645360827446, + 0.03740274906158447, + 0.037411391735076904, + 0.07017417252063751, + 0.03749470412731171, + 0.0030967877246439457, + 0.0014363540103659034, + 0.001690297620370984, + 0.0022953141015022993, + 0.0030861063860356808, + 0.0038575828075408936, + 0.007857841439545155, + 0.004813558422029018, + 0.004985778592526913, + 0.004197485279291868, + 0.005181408487260342, + 0.0032934255432337523, + 0.002728018444031477, + 0.0043172454461455345, + 0.006846040952950716, + 0.004314415622502565, + 0.0032503753900527954, + 0.001721713924780488, + 0.002155184978619218, + 0.0028406078927218914, + 0.0031448386143893003, + 0.003389789257198572, + 0.006962040439248085, + 0.004161844030022621, + 0.006861253175884485, + 0.006351803429424763, + 0.006192638073116541, + 0.005537835881114006, + 0.010758239775896072, + 0.008471415378153324, + 0.011799512431025505, + 0.0097132483497262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.00417480431497097 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31347593665122986 + ], + "opposite_logprob_tokens": [ + -1.3134759664535522 + ], + "logprob_reward": -0.31347593665122986, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31347593665122986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13817119598388672, + 0.215160071849823, + 0.22874146699905396, + 0.08695331960916519, + 0.055033013224601746, + 0.08996044099330902, + 0.057508476078510284, + 0.06032732501626015, + 0.043439216911792755, + 0.024705450981855392 + ], + "credit_map": [ + 0.01669154316186905, + 0.012769712135195732, + 0.016081979498267174, + 0.010509775951504707, + 0.0166233628988266, + 0.01939334347844124, + 0.02259667031466961, + 0.023504819720983505, + 0.0293282400816679, + 0.014794685877859592, + 0.022189296782016754, + 0.02321457490324974, + 0.024827394634485245, + 0.027586089447140694, + 0.0367530956864357, + 0.03646670654416084, + 0.03086811490356922, + 0.020851951092481613, + 0.028229612857103348, + 0.023707997053861618, + 0.030333302915096283, + 0.0269336998462677, + 0.038204651325941086, + 0.029612140730023384, + 0.010857169516384602, + 0.0045508285984396935, + 0.008122527040541172, + 0.007119964342564344, + 0.007022954057902098, + 0.010703317821025848, + 0.025948628783226013, + 0.01262793131172657, + 0.006723640486598015, + 0.002286344300955534, + 0.0036226639058440924, + 0.00371408904902637, + 0.004936517681926489, + 0.008693180978298187, + 0.0128242839127779, + 0.012232295237481594, + 0.01272877398878336, + 0.01196560263633728, + 0.0063039883971214294, + 0.007025981787592173, + 0.007338306866586208, + 0.01208599004894495, + 0.017027640715241432, + 0.015484160743653774, + 0.0071262940764427185, + 0.003243363695219159, + 0.004476143512874842, + 0.005168267525732517, + 0.006070426665246487, + 0.007636228576302528, + 0.01379293855279684, + 0.0099948113784194, + 0.006884952541440725, + 0.006621603388339281, + 0.006925696041435003, + 0.005608406849205494, + 0.0046694413758814335, + 0.008106724359095097, + 0.01418600045144558, + 0.007324499078094959, + 0.005081280134618282, + 0.003078783629462123, + 0.003708382835611701, + 0.004784818738698959, + 0.0059669967740774155, + 0.005417939275503159, + 0.008413096889853477, + 0.006987920496612787, + 0.003919924609363079, + 0.0015214696759358048, + 0.0018722664099186659, + 0.002121076686307788, + 0.003234640695154667, + 0.00335542275570333, + 0.004850058350712061, + 0.0038305914495140314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.31347593665122986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906956017017365 + ], + "opposite_logprob_tokens": [ + -2.4640696048736572 + ], + "logprob_reward": -0.08906956017017365, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906956017017365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22617171704769135, + 0.05275961011648178, + 0.12381211668252945, + 0.05132637917995453, + 0.07000716030597687, + 0.18943695724010468, + 0.051933083683252335, + 0.07717131823301315, + 0.09814684092998505, + 0.05923484265804291 + ], + "credit_map": [ + 0.028489651158452034, + 0.01658720336854458, + 0.027109187096357346, + 0.0165327750146389, + 0.02674633078277111, + 0.028983628377318382, + 0.04433921352028847, + 0.037383727729320526, + 0.004550308454781771, + 0.0036229153629392385, + 0.004420066252350807, + 0.004634563811123371, + 0.006776125635951757, + 0.007377268746495247, + 0.013085871934890747, + 0.008292486891150475, + 0.016683943569660187, + 0.010113700293004513, + 0.018167858943343163, + 0.013293672353029251, + 0.015076584182679653, + 0.014342039823532104, + 0.019375471398234367, + 0.016758838668465614, + 0.008212024345993996, + 0.0029256099369376898, + 0.004686621017754078, + 0.0040664589032530785, + 0.006369984708726406, + 0.005704943556338549, + 0.0118998559191823, + 0.007460875902324915, + 0.007931475527584553, + 0.004341516178101301, + 0.007347763516008854, + 0.005665614735335112, + 0.008506359532475471, + 0.00925751868635416, + 0.015956105664372444, + 0.011000799015164375, + 0.019200162962079048, + 0.021724848076701164, + 0.015866680070757866, + 0.017181526869535446, + 0.024511609226465225, + 0.02557739056646824, + 0.035888370126485825, + 0.02948635257780552, + 0.006622903514653444, + 0.003243780927732587, + 0.0070982747711241245, + 0.004731867928057909, + 0.004408638458698988, + 0.006651231087744236, + 0.01068468764424324, + 0.008491696789860725, + 0.00900937244296074, + 0.005584692116826773, + 0.006859087850898504, + 0.0068159871734678745, + 0.009105660021305084, + 0.009672615677118301, + 0.0205263402312994, + 0.009597561322152615, + 0.012771104462444782, + 0.005053977482020855, + 0.006315722595900297, + 0.0070269908756017685, + 0.00884728692471981, + 0.013392607681453228, + 0.027296975255012512, + 0.017442170530557632, + 0.004954228643327951, + 0.004937243182212114, + 0.005196030717343092, + 0.005153284873813391, + 0.008985953405499458, + 0.0072918543592095375, + 0.013726149685680866, + 0.008990091271698475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.08906956017017365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978530291467905 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.00978530291467905, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978530291467905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18970336019992828, + 0.1340802162885666, + 0.11462417244911194, + 0.06187796592712402, + 0.0823071077466011, + 0.1189170628786087, + 0.06677320599555969, + 0.07446987181901932, + 0.019348494708538055, + 0.13789859414100647 + ], + "credit_map": [ + 0.021917100995779037, + 0.02281988225877285, + 0.018394263461232185, + 0.019834743812680244, + 0.02704755775630474, + 0.02337276004254818, + 0.03204641118645668, + 0.02427063323557377, + 0.013918145559728146, + 0.010645000264048576, + 0.014978640712797642, + 0.01525605283677578, + 0.017699304968118668, + 0.015373149886727333, + 0.026726696640253067, + 0.01948322169482708, + 0.013887973502278328, + 0.008782858960330486, + 0.012970196083188057, + 0.010044115595519543, + 0.01439070887863636, + 0.014964494854211807, + 0.01900526136159897, + 0.02057856321334839, + 0.012065834365785122, + 0.003685052040964365, + 0.005061598960310221, + 0.004997828975319862, + 0.005432541482150555, + 0.007990083657205105, + 0.01326034963130951, + 0.009384673088788986, + 0.012914977967739105, + 0.004481074400246143, + 0.005024215206503868, + 0.006112388335168362, + 0.006835192441940308, + 0.013068830594420433, + 0.016642453148961067, + 0.017227966338396072, + 0.01548230554908514, + 0.010304171591997147, + 0.00896941777318716, + 0.009936473332345486, + 0.01342038344591856, + 0.017666205763816833, + 0.0175318606197834, + 0.02560623735189438, + 0.007538558449596167, + 0.005061494652181864, + 0.010679331608116627, + 0.006068920250982046, + 0.006477650720626116, + 0.008672687225043774, + 0.01393772941082716, + 0.008336832746863365, + 0.012334795668721199, + 0.0062743136659264565, + 0.011754780076444149, + 0.006985037121921778, + 0.005958758294582367, + 0.00893105287104845, + 0.012016602791845798, + 0.010214525274932384, + 0.0033870902843773365, + 0.0012065292103216052, + 0.0020239981822669506, + 0.0019255579682067037, + 0.002163141965866089, + 0.0024948331993073225, + 0.0035690253134816885, + 0.0025783167220652103, + 0.014918021857738495, + 0.01077767089009285, + 0.01140406634658575, + 0.01061961892992258, + 0.020832940936088562, + 0.018176766112446785, + 0.03011779673397541, + 0.02105170302093029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.00978530291467905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stark black-and-white, slightly low-angle close-up, held static, captures a classic portrait of a large American bullfrog (Lithobates catesbeianus) in the midst of its intense bark. High-contrast lighting from a single source dramatically illuminates the scene, sculpting the amphibian from a dark, completely out-of-focus background. As the frog calls, its two vocal sacs beneath its jaws inflate steadily and symmetrically, expanding like pale, taut balloons. The harsh light meticulously renders the shimmering, bumpy texture of the frog's moist skin with brilliant specular highlights, while casting deep, velvety shadows that accentuate its form. The sacs themselves become almost pure white, their thin, stretched skin contrasting sharply against the darker hues of the frog's body. The entire work emphasizes the primal nature of the performance, creating an atmosphere of visceral, focused concentration that is both beautiful and intense.", + "seed": 7362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_8/index_8498_seed_7362_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0726342959329486, + "vlm_logprob": -0.0726342959329486, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.0726342959329486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1013631746172905 + ], + "opposite_logprob_tokens": [ + -2.476363182067871 + ], + "logprob_reward": -0.1013631746172905, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1013631746172905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2197647988796234, + 0.11407806724309921, + 0.1682967096567154, + 0.06700869649648666, + 0.05507296696305275, + 0.04044423624873161, + 0.1888146698474884, + 0.03797968849539757, + 0.05374322459101677, + 0.05479690432548523 + ], + "credit_map": [ + 0.02197040058672428, + 0.03144347667694092, + 0.02901783399283886, + 0.019146818667650223, + 0.022794321179389954, + 0.024991575628519058, + 0.04406699538230896, + 0.02633337676525116, + 0.010526050813496113, + 0.007373815402388573, + 0.00795090850442648, + 0.005892615765333176, + 0.013494187034666538, + 0.018167778849601746, + 0.033535197377204895, + 0.01713751070201397, + 0.01631416194140911, + 0.019031649455428123, + 0.01589697226881981, + 0.014139040373265743, + 0.025917069986462593, + 0.021398819983005524, + 0.02302461490035057, + 0.0325743705034256, + 0.00614310335367918, + 0.010130464099347591, + 0.006611521355807781, + 0.004955758340656757, + 0.012198694981634617, + 0.00777121027931571, + 0.00997540820389986, + 0.009222544729709625, + 0.0055567421950399876, + 0.00377328391186893, + 0.005264362785965204, + 0.003799742553383112, + 0.004961708560585976, + 0.008137727156281471, + 0.013497523963451385, + 0.010081876069307327, + 0.0037110925186425447, + 0.004608780611306429, + 0.0034075924195349216, + 0.0024558743461966515, + 0.0055536916479468346, + 0.0051355380564928055, + 0.008087260648608208, + 0.007484403904527426, + 0.020244454964995384, + 0.015001869760453701, + 0.022281065583229065, + 0.019414115697145462, + 0.03189856931567192, + 0.020206689834594727, + 0.03026684746146202, + 0.02950107306241989, + 0.003497277619317174, + 0.0033701045904308558, + 0.0039018187671899796, + 0.0029916169587522745, + 0.005045180208981037, + 0.004824561532586813, + 0.007147567346692085, + 0.007201561704277992, + 0.005009111482650042, + 0.004158271010965109, + 0.004005929920822382, + 0.0030935630202293396, + 0.0070049152709543705, + 0.007342171855270863, + 0.013062744401395321, + 0.010066516697406769, + 0.0056008268147706985, + 0.005991902202367783, + 0.0060751005075871944, + 0.005005392245948315, + 0.006988513749092817, + 0.00697704404592514, + 0.008839928545057774, + 0.009318194352090359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.1013631746172905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0037003159523010254 + ], + "opposite_logprob_tokens": [ + -5.628700256347656 + ], + "logprob_reward": -0.0037003159523010254, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0037003159523010254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1809474229812622, + 0.15824058651924133, + 0.11127783358097076, + 0.08405129611492157, + 0.0722513273358345, + 0.08545063436031342, + 0.08106967061758041, + 0.10451361536979675, + 0.050393518060445786, + 0.07180411368608475 + ], + "credit_map": [ + 0.01837800443172455, + 0.017350563779473305, + 0.01203125435858965, + 0.010143868625164032, + 0.02378285489976406, + 0.02735373191535473, + 0.031034814193844795, + 0.040872324258089066, + 0.01613426022231579, + 0.014910304918885231, + 0.012360566295683384, + 0.011606081388890743, + 0.02243029512465, + 0.021164724603295326, + 0.031530868262052536, + 0.02810349501669407, + 0.013595851138234138, + 0.007757219485938549, + 0.007628866005688906, + 0.006260109134018421, + 0.017396744340658188, + 0.016577599570155144, + 0.019358225166797638, + 0.02270321361720562, + 0.00833493284881115, + 0.010420762933790684, + 0.008260134607553482, + 0.00844532810151577, + 0.008747204206883907, + 0.010588623583316803, + 0.014885216020047665, + 0.01436909381300211, + 0.00792932603508234, + 0.006076049990952015, + 0.007488439325243235, + 0.0045183878391981125, + 0.009706848300993443, + 0.010211406275629997, + 0.012105325236916542, + 0.014215542003512383, + 0.008464933373034, + 0.007658606395125389, + 0.008621915243566036, + 0.0054925777949392796, + 0.012118355371057987, + 0.011996396817266941, + 0.01764577627182007, + 0.013452067039906979, + 0.008278326131403446, + 0.0072817010805010796, + 0.007267784792929888, + 0.005338647402822971, + 0.00992516241967678, + 0.010841227136552334, + 0.017703643068671227, + 0.014433188363909721, + 0.009984559379518032, + 0.008622518740594387, + 0.009623895399272442, + 0.006372763309627771, + 0.011878518387675285, + 0.013752417638897896, + 0.02079024910926819, + 0.023488685488700867, + 0.0048003001138567924, + 0.007516028359532356, + 0.00390605884604156, + 0.0038826202508062124, + 0.00788269005715847, + 0.006131854839622974, + 0.010763467289507389, + 0.0055104950442910194, + 0.008049217984080315, + 0.008530774153769016, + 0.0076569439843297005, + 0.005218823906034231, + 0.008705607615411282, + 0.010330636985599995, + 0.010826432146131992, + 0.012485681101679802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0037003159523010254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906574547290802 + ], + "opposite_logprob_tokens": [ + -2.4640657901763916 + ], + "logprob_reward": -0.08906574547290802, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906574547290802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17871148884296417, + 0.21929587423801422, + 0.0801670029759407, + 0.20389780402183533, + 0.02410505898296833, + 0.022262580692768097, + 0.04655957221984863, + 0.09375013411045074, + 0.032560113817453384, + 0.09869039803743362 + ], + "credit_map": [ + 0.015613614581525326, + 0.027923518791794777, + 0.021191870793700218, + 0.01186523586511612, + 0.022073861211538315, + 0.020238516852259636, + 0.03471478074789047, + 0.025090094655752182, + 0.020934976637363434, + 0.014718293212354183, + 0.026485970243811607, + 0.020699162036180496, + 0.029885515570640564, + 0.0280937310308218, + 0.04276612028479576, + 0.035712093114852905, + 0.007439166773110628, + 0.010071644559502602, + 0.01338900439441204, + 0.005734586156904697, + 0.008095871657133102, + 0.01059021521359682, + 0.010292839258909225, + 0.0145536744967103, + 0.02015683799982071, + 0.028345871716737747, + 0.019871246069669724, + 0.01724998466670513, + 0.03647226467728615, + 0.024934981018304825, + 0.030042504891753197, + 0.026824114844202995, + 0.002470980864018202, + 0.0026798020116984844, + 0.0023814362939447165, + 0.0017198673449456692, + 0.003111304482445121, + 0.0033403229899704456, + 0.004215294029563665, + 0.004186052363365889, + 0.002226497745141387, + 0.0017104133730754256, + 0.0022424079943448305, + 0.0017696534050628543, + 0.0029983981512486935, + 0.003219680394977331, + 0.004982109181582928, + 0.003113420680165291, + 0.0048096138052642345, + 0.003990841563791037, + 0.004170619882643223, + 0.0037536758463829756, + 0.005443796515464783, + 0.006403661798685789, + 0.00926960539072752, + 0.008717757649719715, + 0.009696605615317822, + 0.006952963769435883, + 0.010179072618484497, + 0.007650045212358236, + 0.01635720580816269, + 0.01114107109606266, + 0.01994919218122959, + 0.011823986656963825, + 0.004251017235219479, + 0.005675686988979578, + 0.0027061428409069777, + 0.0023605101741850376, + 0.0032474042382091284, + 0.0043222056701779366, + 0.005139095243066549, + 0.00485805282369256, + 0.010738297365605831, + 0.014007261022925377, + 0.01156615186482668, + 0.011280996724963188, + 0.013042558915913105, + 0.012157382443547249, + 0.015256217680871487, + 0.010641532018780708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.08906574547290802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039876639842987 + ], + "opposite_logprob_tokens": [ + -1.9103987216949463 + ], + "logprob_reward": -0.16039876639842987, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039876639842987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2415497601032257, + 0.11722792685031891, + 0.1850896179676056, + 0.055293306708335876, + 0.08863776922225952, + 0.057284120470285416, + 0.04073774442076683, + 0.09794557094573975, + 0.021303381770849228, + 0.09493076801300049 + ], + "credit_map": [ + 0.027333369478583336, + 0.0284098032861948, + 0.020641813054680824, + 0.01831679604947567, + 0.02566227689385414, + 0.03169426694512367, + 0.04281926155090332, + 0.04667217656970024, + 0.012046484276652336, + 0.010392636992037296, + 0.016950082033872604, + 0.011594508774578571, + 0.015284599736332893, + 0.013329615816473961, + 0.022807126864790916, + 0.014822871424257755, + 0.020307309925556183, + 0.025470662862062454, + 0.016126064583659172, + 0.01125694066286087, + 0.02448323369026184, + 0.02443934790790081, + 0.026014557108283043, + 0.036991506814956665, + 0.005309480708092451, + 0.009167581796646118, + 0.005384639371186495, + 0.0038329451344907284, + 0.011156144551932812, + 0.006240465212613344, + 0.007757348008453846, + 0.006444703787565231, + 0.01007708627730608, + 0.009182027541100979, + 0.008072172291576862, + 0.0051832739263772964, + 0.011163217015564442, + 0.01334954984486103, + 0.012797982431948185, + 0.01881246082484722, + 0.005668054800480604, + 0.00652664341032505, + 0.006440965458750725, + 0.004237920977175236, + 0.00848218984901905, + 0.007370933890342712, + 0.009945843368768692, + 0.008611570112407207, + 0.004090195521712303, + 0.0036765222903341055, + 0.002538346452638507, + 0.0038509941659867764, + 0.004940548911690712, + 0.006121897138655186, + 0.009319424629211426, + 0.006199815776199102, + 0.010859988629817963, + 0.008241626434028149, + 0.011859691701829433, + 0.006262618117034435, + 0.014719553291797638, + 0.011417008005082607, + 0.018502868711948395, + 0.016082223504781723, + 0.0020959489047527313, + 0.004181778058409691, + 0.0013999476796016097, + 0.0018129057716578245, + 0.0023233036044985056, + 0.0026434632018208504, + 0.0025449334643781185, + 0.004301100503653288, + 0.009973786771297455, + 0.00993195828050375, + 0.009768356569111347, + 0.00800432451069355, + 0.010659100487828255, + 0.012743453495204449, + 0.01774498261511326, + 0.016104806214571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.16039876639842987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a large American bullfrog (Lithobates catesbeianus)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a close-up taken from a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera held static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the frog's two vocal sacs beneath its jaws inflate steadily and symmetrically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and completely out-of-focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643477223813534 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643477223813534, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643477223813534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16868193447589874, + 0.06345038115978241, + 0.3250046372413635, + 0.05034109950065613, + 0.13275140523910522, + 0.05438409745693207, + 0.05941799655556679, + 0.051425863057374954, + 0.04357998073101044, + 0.05096264183521271 + ], + "credit_map": [ + 0.018768370151519775, + 0.008618404157459736, + 0.014134951867163181, + 0.008029323071241379, + 0.017601897940039635, + 0.026035316288471222, + 0.038188837468624115, + 0.0373048260807991, + 0.00702374754473567, + 0.003539855359122157, + 0.0050854794681072235, + 0.0037746296729892492, + 0.006059675011783838, + 0.009801910258829594, + 0.013576233759522438, + 0.01458884496241808, + 0.03146296367049217, + 0.025087833404541016, + 0.016523001715540886, + 0.017991723492741585, + 0.034569188952445984, + 0.053286537528038025, + 0.06009134650230408, + 0.0859920084476471, + 0.005653477739542723, + 0.005909026134759188, + 0.005040663760155439, + 0.004240702837705612, + 0.005161957815289497, + 0.006551904138177633, + 0.007278665900230408, + 0.010504700243473053, + 0.01220471877604723, + 0.014695034362375736, + 0.008752213791012764, + 0.006549759302288294, + 0.015076623298227787, + 0.020343536511063576, + 0.020783690735697746, + 0.034345827996730804, + 0.005492448806762695, + 0.003880490781739354, + 0.0038168926257640123, + 0.0035252231173217297, + 0.006341511849313974, + 0.007697520777583122, + 0.015414934605360031, + 0.008215071633458138, + 0.0067876409739255905, + 0.0053728800266981125, + 0.005466230679303408, + 0.004874331410974264, + 0.006241397466510534, + 0.008267851546406746, + 0.009989985264837742, + 0.012417672201991081, + 0.006060359533876181, + 0.003906295634806156, + 0.0030251718126237392, + 0.0028708442114293575, + 0.006065890658646822, + 0.007766678463667631, + 0.0073462361469864845, + 0.014384382404386997, + 0.0036827686708420515, + 0.004967569839209318, + 0.0033360167872160673, + 0.0032182319555431604, + 0.006233121268451214, + 0.0052624973468482494, + 0.010496937669813633, + 0.006382835563272238, + 0.0050234319642186165, + 0.006013305392116308, + 0.0029820732306689024, + 0.005737464874982834, + 0.005551212467253208, + 0.007009847555309534, + 0.006737269461154938, + 0.011908031068742275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.008643477223813534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_9/index_tensor([1741, 1741])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_9/index_tensor([1741, 1741])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5f55ac42fb58a9ca9490ad4c4f8ace74f9cc0084 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_9/index_tensor([1741, 1741])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "horizontal tracking medium shot, filmed from the side, follows a traditional Chinese red sedan chair, exquisitely constructed and vibrantly colored. Constructed of vermilion lacquered wood, it reflects a warm, rich luster in the bright sunlight. Its walls are intricately carved with auspicious dragon and phoenix motifs and embellished with gleaming gold ornaments. Its roof features an ornate peaked roof, adorned with long golden tassels at the corners. Its curtains, tightly drawn from thick red silk, add a touch of mystery. Several bearers, dressed in uniform red uniforms and felt hats, move with steady, coordinated steps, carrying two sturdy red poles. With each step, the sedan chair rises and falls rhythmically, gently and steadily, along the quaint cobblestone streets. The golden tassels sway in sync. Against the backdrop of lively voices and the traditional Chinese architecture, the entire scene is imbued with a rich, festive, and solemn atmosphere of traditional celebration.", + "seed": 4195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_9/index_1741_seed_4195_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06999868918210268, + "vlm_logprob": -0.06999868918210268, + "vlm_yes_no_margin": 3.299999952316284, + "vlm_token_credit": -0.06999868918210268, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0222741961479187 + ], + "opposite_logprob_tokens": [ + -4.022274017333984 + ], + "logprob_reward": -0.0222741961479187, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0222741961479187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.164536714553833, + 0.035826098173856735, + 0.10020675510168076, + 0.02162495069205761, + 0.17206794023513794, + 0.13186568021774292, + 0.14556357264518738, + 0.053240954875946045, + 0.1504482626914978, + 0.024618983268737793 + ], + "credit_map": [ + 0.01837138645350933, + 0.013163615018129349, + 0.01789485290646553, + 0.01709936000406742, + 0.011877285316586494, + 0.022262729704380035, + 0.02782711200416088, + 0.03604039177298546, + 0.004763784818351269, + 0.004789670463651419, + 0.005782836116850376, + 0.005533924326300621, + 0.0031254375353455544, + 0.0036258187610656023, + 0.0031207012943923473, + 0.005083929281681776, + 0.00746991066262126, + 0.01104989554733038, + 0.006785687990486622, + 0.032544493675231934, + 0.0057101184502244, + 0.015317951329052448, + 0.015610468573868275, + 0.005718242842704058, + 0.001980028348043561, + 0.0021460596472024918, + 0.0018867859616875648, + 0.0030492295045405626, + 0.003343745367601514, + 0.003553686197847128, + 0.002845184877514839, + 0.0028202347457408905, + 0.01160370372235775, + 0.06285460293292999, + 0.005981705617159605, + 0.010213177651166916, + 0.010061697103083134, + 0.012601558119058609, + 0.008756095543503761, + 0.049995433539152145, + 0.006943647284060717, + 0.015968576073646545, + 0.00792214646935463, + 0.0191289521753788, + 0.014355476945638657, + 0.02384093590080738, + 0.017624694854021072, + 0.026081250980496407, + 0.024196943268179893, + 0.0148994205519557, + 0.00892234593629837, + 0.051126878708601, + 0.009883781895041466, + 0.00660521537065506, + 0.019488634541630745, + 0.010440384969115257, + 0.00803492870181799, + 0.005825970321893692, + 0.010453666560351849, + 0.005806348752230406, + 0.0073503307066857815, + 0.004566413816064596, + 0.0070872316136956215, + 0.004116069991141558, + 0.028318343684077263, + 0.045119598507881165, + 0.012689091265201569, + 0.021658848971128464, + 0.014126553200185299, + 0.008049890398979187, + 0.011669205501675606, + 0.008816744200885296, + 0.0017767463577911258, + 0.0023710792884230614, + 0.0023204146418720484, + 0.002766459248960018, + 0.004448752850294113, + 0.0038014641031622887, + 0.003400354413315654, + 0.00373371341265738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0222741961479187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087254509329796 + ], + "opposite_logprob_tokens": [ + -4.141087055206299 + ], + "logprob_reward": -0.016087254509329796, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087254509329796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10037887841463089, + 0.02028621919453144, + 0.4713703691959381, + 0.008319812826812267, + 0.022834543138742447, + 0.03568916395306587, + 0.056339897215366364, + 0.04944298043847084, + 0.21935304999351501, + 0.01598503254354 + ], + "credit_map": [ + 0.005755900405347347, + 0.005388712976127863, + 0.010675737634301186, + 0.012252358719706535, + 0.006635132245719433, + 0.011944549158215523, + 0.020968448370695114, + 0.02675805054605007, + 0.0017976333620026708, + 0.002898174338042736, + 0.0030450497288256884, + 0.002646662760525942, + 0.002208468271419406, + 0.002669052453711629, + 0.0021588571835309267, + 0.002862323075532913, + 0.031677912920713425, + 0.055807363241910934, + 0.0283312126994133, + 0.1753058284521103, + 0.015135352499783039, + 0.06386344879865646, + 0.08348479866981506, + 0.01776447519659996, + 0.0007313024834729731, + 0.0006471832748502493, + 0.0007532204617746174, + 0.0014666671631857753, + 0.0008884586277417839, + 0.0012919883010908961, + 0.0012316269567236304, + 0.0013093671295791864, + 0.0012987898662686348, + 0.005805485416203737, + 0.001484700827859342, + 0.0019214387284591794, + 0.0023865506518632174, + 0.0024534647818654776, + 0.0025192624889314175, + 0.004964851308614016, + 0.0026086699217557907, + 0.0060533685609698296, + 0.002491776365786791, + 0.004915582947432995, + 0.004218658432364464, + 0.0070259650237858295, + 0.003696287516504526, + 0.004678857978433371, + 0.010451127775013447, + 0.006121532525867224, + 0.004390936344861984, + 0.014573831111192703, + 0.006130393128842115, + 0.003631020663306117, + 0.006594338919967413, + 0.004446722101420164, + 0.008175283670425415, + 0.006083288695663214, + 0.008466445840895176, + 0.004459785297513008, + 0.008677950128912926, + 0.0036978209391236305, + 0.006368300411850214, + 0.003514108946546912, + 0.04964124411344528, + 0.07385189831256866, + 0.015367944724857807, + 0.03288425877690315, + 0.016366107389330864, + 0.006255921442061663, + 0.015309973619878292, + 0.009675727225840092, + 0.002491247607395053, + 0.0014460858656093478, + 0.0022734608501195908, + 0.0019053348805755377, + 0.0020649300422519445, + 0.002193639287725091, + 0.0017714434070512652, + 0.001838892581872642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.016087254509329796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979274045675993 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.00979274045675993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979274045675993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03894081711769104, + 0.010417982935905457, + 0.6197816729545593, + 0.004231410101056099, + 0.24994030594825745, + 0.02622811496257782, + 0.014575758948922157, + 0.009641573764383793, + 0.015089127235114574, + 0.011153224855661392 + ], + "credit_map": [ + 0.00393284996971488, + 0.004437815397977829, + 0.004101704340428114, + 0.0042752791196107864, + 0.002602226799353957, + 0.004323670174926519, + 0.006346068810671568, + 0.008921205997467041, + 0.0009682374657131732, + 0.0011974055087193847, + 0.0018645115196704865, + 0.0012997729936614633, + 0.0009251770097762346, + 0.0017568464390933514, + 0.0010730673093348742, + 0.0013329649809747934, + 0.04120415449142456, + 0.07294047623872757, + 0.03703856095671654, + 0.22980338335037231, + 0.020101457834243774, + 0.08532309532165527, + 0.1094810739159584, + 0.02388950064778328, + 0.00037289594183675945, + 0.0004963281680829823, + 0.00043331910273991525, + 0.00047437427565455437, + 0.0006134049035608768, + 0.0006646557594649494, + 0.0005439284141175449, + 0.0006325037102214992, + 0.01598896086215973, + 0.09205059707164764, + 0.006125136744230986, + 0.014376970008015633, + 0.014991275034844875, + 0.019648225978016853, + 0.013390794396400452, + 0.07336835563182831, + 0.0015194832812994719, + 0.00332580110989511, + 0.0015882431762292981, + 0.004598930478096008, + 0.0026944365818053484, + 0.003673322731629014, + 0.004961454775184393, + 0.0038664431776851416, + 0.0021392954513430595, + 0.0015239063650369644, + 0.0010706493631005287, + 0.004022901877760887, + 0.0013573250034824014, + 0.0010991741437464952, + 0.0018338109366595745, + 0.0015286961570382118, + 0.001303609344176948, + 0.0012618382461369038, + 0.001406781724654138, + 0.0008760761702433228, + 0.001491373055614531, + 0.0008850504527799785, + 0.001689957920461893, + 0.0007268873741850257, + 0.0025288185570389032, + 0.0036795272026211023, + 0.0008759781485423446, + 0.0015769933816045523, + 0.0013947738334536552, + 0.001451933174394071, + 0.0018079517176374793, + 0.0017731504049152136, + 0.0008126910543069243, + 0.0015468044439330697, + 0.0006963458727113903, + 0.0012413747608661652, + 0.0013785995543003082, + 0.0017100708791986108, + 0.002190327038988471, + 0.0015770109603181481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.00979274045675993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028412193059921 + ], + "opposite_logprob_tokens": [ + -2.3502840995788574 + ], + "logprob_reward": -0.10028412193059921, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028412193059921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033833812922239304, + 0.028880376368761063, + 0.5343521237373352, + 0.018039846792817116, + 0.1655309498310089, + 0.07477999478578568, + 0.0414716973900795, + 0.02206960879266262, + 0.060105256736278534, + 0.02093628980219364 + ], + "credit_map": [ + 0.0025534098967909813, + 0.0027816041838377714, + 0.004097200930118561, + 0.003707075724378228, + 0.0027507110498845577, + 0.003522627055644989, + 0.00626309122890234, + 0.008158093318343163, + 0.0032632125075906515, + 0.0035562696866691113, + 0.004361126106232405, + 0.004056940320879221, + 0.0032227232586592436, + 0.0029168801847845316, + 0.0027823608834296465, + 0.0047208648175001144, + 0.03661542013287544, + 0.06259436905384064, + 0.03206907585263252, + 0.1981135904788971, + 0.016577167436480522, + 0.07319629192352295, + 0.09450090676546097, + 0.020685281604528427, + 0.0023388960398733616, + 0.0016176768112927675, + 0.00185807840898633, + 0.0016026761149987578, + 0.002513405168429017, + 0.0032095913775265217, + 0.0021628057584166527, + 0.002736717462539673, + 0.011248704977333546, + 0.05659184232354164, + 0.007417237386107445, + 0.010919902473688126, + 0.009544625878334045, + 0.014178029261529446, + 0.009673075750470161, + 0.0459575280547142, + 0.004101034253835678, + 0.009706848300993443, + 0.004460989963263273, + 0.02099619060754776, + 0.007371013052761555, + 0.007831758819520473, + 0.006934169214218855, + 0.013377993367612362, + 0.004904523026198149, + 0.00422065332531929, + 0.0031079244799911976, + 0.011936785653233528, + 0.0029418303165584803, + 0.0028428351506590843, + 0.00464006420224905, + 0.006877081468701363, + 0.0027396364603191614, + 0.0027107070200145245, + 0.0030045579187572002, + 0.0018490799702703953, + 0.004248416516929865, + 0.0027567821089178324, + 0.0031756481621414423, + 0.0015847813338041306, + 0.013169551268219948, + 0.02011496014893055, + 0.003744221990928054, + 0.00897559616714716, + 0.005228434223681688, + 0.002028286922723055, + 0.004356055520474911, + 0.0024881537538021803, + 0.0017101753037422895, + 0.002641037805005908, + 0.0015183049254119396, + 0.0023971013724803925, + 0.002897851634770632, + 0.0034374988172203302, + 0.002529213670641184, + 0.0038051074370741844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.10028412193059921 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155513286590576 + ], + "opposite_logprob_tokens": [ + -1.7015551328659058 + ], + "logprob_reward": -0.20155513286590576, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155513286590576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05906877666711807, + 0.011566913686692715, + 0.6178900003433228, + 0.0086809191852808, + 0.14683225750923157, + 0.047122567892074585, + 0.04543289914727211, + 0.009307832457125187, + 0.022284645587205887, + 0.031813137233257294 + ], + "credit_map": [ + 0.004685969091951847, + 0.004923942033201456, + 0.0058716521598398685, + 0.007932709529995918, + 0.004214270506054163, + 0.007428076583892107, + 0.009972790256142616, + 0.014039368368685246, + 0.001582882134243846, + 0.0013258688850328326, + 0.0018373556667938828, + 0.0015398088144138455, + 0.0010078097693622112, + 0.0014191368827596307, + 0.0012914633844047785, + 0.0015625879168510437, + 0.03992241621017456, + 0.07499722391366959, + 0.03711734712123871, + 0.23290015757083893, + 0.020927924662828445, + 0.0779561996459961, + 0.11180467903614044, + 0.02226412668824196, + 0.0009918099967762828, + 0.0007407712982967496, + 0.0007266026805154979, + 0.0010133981704711914, + 0.0010428385576233268, + 0.0012476224219426513, + 0.0016845810459926724, + 0.0012332953047007322, + 0.009489670395851135, + 0.05512106791138649, + 0.0030435887165367603, + 0.008297988213598728, + 0.00828629545867443, + 0.011068018153309822, + 0.0072685061022639275, + 0.04425712674856186, + 0.001898953807540238, + 0.007289521861821413, + 0.002091058064252138, + 0.005239643156528473, + 0.007335998583585024, + 0.012013262137770653, + 0.004816753789782524, + 0.006437378935515881, + 0.004974639508873224, + 0.004039905034005642, + 0.0028812673408538103, + 0.013252588920295238, + 0.0032269260846078396, + 0.0029581065755337477, + 0.007123539689928293, + 0.00697592506185174, + 0.0012617646716535091, + 0.0010116351768374443, + 0.0010129621950909495, + 0.0008275866857729852, + 0.0015396297676488757, + 0.0011526751331984997, + 0.0016893494175747037, + 0.0008122294675558805, + 0.0043343971483409405, + 0.0067009879276156425, + 0.0014034685445949435, + 0.0029616206884384155, + 0.002155026886612177, + 0.0011994059896096587, + 0.001884964294731617, + 0.0016447744565084577, + 0.0019188453443348408, + 0.004248898942023516, + 0.002571662189438939, + 0.002777241403236985, + 0.0045239427126944065, + 0.006585844326764345, + 0.005664818920195103, + 0.003521884558722377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.20155513286590576 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking medium shot, filmed from the side, follows a traditional Chinese red sedan chair, exquisitely constructed and vibrantly colored. Constructed of vermilion lacquered wood, it reflects a warm, rich luster in the bright sunlight. Its walls are intricately carved with auspicious dragon and phoenix motifs and embellished with gleaming gold ornaments. Its roof features an ornate peaked roof, adorned with long golden tassels at the corners. Its curtains, tightly drawn from thick red silk, add a touch of mystery. Several bearers, dressed in uniform red uniforms and felt hats, move with steady, coordinated steps, carrying two sturdy red poles. With each step, the sedan chair rises and falls rhythmically, gently and steadily, along the quaint cobblestone streets. The golden tassels sway in sync. Against the backdrop of lively voices and the traditional Chinese architecture, the entire scene is imbued with a rich, festive, and solemn atmosphere of traditional celebration.", + "seed": 4196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_9/index_1741_seed_4196_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09929756447672844, + "vlm_logprob": -0.09929756447672844, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.09929756447672844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008151477202773094 + ], + "opposite_logprob_tokens": [ + -5.133151531219482 + ], + "logprob_reward": -0.008151477202773094, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008151477202773094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1223912388086319, + 0.12132935971021652, + 0.10650769621133804, + 0.15416356921195984, + 0.09697002917528152, + 0.059543538838624954, + 0.17837156355381012, + 0.049436528235673904, + 0.040640056133270264, + 0.07064641267061234 + ], + "credit_map": [ + 0.012339332140982151, + 0.011888434179127216, + 0.009966632351279259, + 0.0173872709274292, + 0.01392240822315216, + 0.020295139402151108, + 0.019580749794840813, + 0.01701127365231514, + 0.006275885738432407, + 0.00820912979543209, + 0.005330841522663832, + 0.013135538436472416, + 0.009542002342641354, + 0.03506796807050705, + 0.03212473914027214, + 0.011643252335488796, + 0.010189203545451164, + 0.0061378637328743935, + 0.003860458265990019, + 0.019900141283869743, + 0.00949215330183506, + 0.009747997857630253, + 0.03493914753198624, + 0.012240730226039886, + 0.008666235022246838, + 0.011323139071464539, + 0.0059926570393145084, + 0.017868615686893463, + 0.012643092311918736, + 0.008099578320980072, + 0.05205674469470978, + 0.03751349449157715, + 0.016035666689276695, + 0.008601863868534565, + 0.00931884441524744, + 0.02400204911828041, + 0.011950554326176643, + 0.005509553011506796, + 0.015560447238385677, + 0.005991047713905573, + 0.0059574986808001995, + 0.0060889581218361855, + 0.005647164769470692, + 0.00783644150942564, + 0.011899054981768131, + 0.006689993664622307, + 0.008239326067268848, + 0.0071850987151265144, + 0.014631633646786213, + 0.017249615862965584, + 0.03954396769404411, + 0.03929010406136513, + 0.016874827444553375, + 0.013673200272023678, + 0.022805582731962204, + 0.014302637428045273, + 0.004324836656451225, + 0.0048505268059670925, + 0.0051873475313186646, + 0.008240355178713799, + 0.006160615012049675, + 0.0055889347568154335, + 0.010492091067135334, + 0.004591821227222681, + 0.004362821113318205, + 0.004632740747183561, + 0.0035749278031289577, + 0.006273434963077307, + 0.0050674197264015675, + 0.004119881894439459, + 0.007723438553512096, + 0.004885390866547823, + 0.005734165199100971, + 0.011808659881353378, + 0.0076754456385970116, + 0.009593532420694828, + 0.007810473907738924, + 0.008175506256520748, + 0.009184081107378006, + 0.010664541274309158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.008151477202773094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038137808442115784 + ], + "opposite_logprob_tokens": [ + -3.288137912750244 + ], + "logprob_reward": -0.038137808442115784, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038137808442115784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15444384515285492, + 0.26136720180511475, + 0.08171892911195755, + 0.0628226175904274, + 0.06387428194284439, + 0.061937328428030014, + 0.13972215354442596, + 0.08402509242296219, + 0.051975980401039124, + 0.03811260312795639 + ], + "credit_map": [ + 0.009392820298671722, + 0.010664727538824081, + 0.009207663126289845, + 0.024796929210424423, + 0.016649091616272926, + 0.02226313203573227, + 0.024378476664423943, + 0.03709098696708679, + 0.014112863689661026, + 0.02668183669447899, + 0.012859487906098366, + 0.03854040801525116, + 0.021873144432902336, + 0.05126284435391426, + 0.05754396691918373, + 0.03849264234304428, + 0.0087276054546237, + 0.004616059362888336, + 0.0038175215013325214, + 0.016962923109531403, + 0.007938934490084648, + 0.005719887092709541, + 0.025762343779206276, + 0.008173646405339241, + 0.00428044842556119, + 0.004890196956694126, + 0.004146830644458532, + 0.007127340417355299, + 0.007241308223456144, + 0.0055641429498791695, + 0.016725750640034676, + 0.012846594676375389, + 0.008267399854958057, + 0.009073768742382526, + 0.004438302014023066, + 0.015443075448274612, + 0.008134718984365463, + 0.0052762567065656185, + 0.008667953312397003, + 0.0045728012919425964, + 0.007055456284433603, + 0.009271802380681038, + 0.004604448098689318, + 0.006598470266908407, + 0.013028508983552456, + 0.006992951966822147, + 0.007799380924552679, + 0.006586301606148481, + 0.012158794328570366, + 0.013453623279929161, + 0.029866274446249008, + 0.029944054782390594, + 0.01121864840388298, + 0.012837436981499195, + 0.018298814073204994, + 0.011944504454731941, + 0.009554857388138771, + 0.010706035420298576, + 0.01002202183008194, + 0.009310434572398663, + 0.011315210722386837, + 0.01292480994015932, + 0.011393352411687374, + 0.00879836268723011, + 0.004534901585429907, + 0.0064232307486236095, + 0.00452684611082077, + 0.008514156565070152, + 0.007105245720595121, + 0.004674275405704975, + 0.010946814902126789, + 0.005250511225312948, + 0.0027466267347335815, + 0.006768004037439823, + 0.005756029859185219, + 0.005344267003238201, + 0.0038901588413864374, + 0.003929594065994024, + 0.0063934167847037315, + 0.0032845023088157177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.038137808442115784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018208375200629234 + ], + "opposite_logprob_tokens": [ + -4.0182085037231445 + ], + "logprob_reward": -0.018208375200629234, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018208375200629234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10017555952072144, + 0.11351916939020157, + 0.18561261892318726, + 0.07984849810600281, + 0.14769801497459412, + 0.08908719569444656, + 0.08625505119562149, + 0.08307007700204849, + 0.06109469756484032, + 0.05363912135362625 + ], + "credit_map": [ + 0.009444267489016056, + 0.008754025213420391, + 0.007725215516984463, + 0.01677839457988739, + 0.012704373337328434, + 0.01306710671633482, + 0.017668819054961205, + 0.014033360406756401, + 0.007620122749358416, + 0.0073485588654875755, + 0.0059059313498437405, + 0.02629348635673523, + 0.00848542433232069, + 0.020061081275343895, + 0.02360471524298191, + 0.014199835248291492, + 0.017909934744238853, + 0.012523842975497246, + 0.01087016612291336, + 0.04883389174938202, + 0.016513608396053314, + 0.011606131680309772, + 0.0513637512922287, + 0.01599128358066082, + 0.006171473767608404, + 0.006215205416083336, + 0.0036972539965063334, + 0.010547112673521042, + 0.007125788368284702, + 0.009632566943764687, + 0.02098052203655243, + 0.015478572808206081, + 0.02618606761097908, + 0.013840590603649616, + 0.01185902301222086, + 0.04783352464437485, + 0.009779428131878376, + 0.006760166492313147, + 0.022014861926436424, + 0.009424356743693352, + 0.009838040918111801, + 0.014640378765761852, + 0.0061897942796349525, + 0.008776597678661346, + 0.018387269228696823, + 0.010331853292882442, + 0.011843817308545113, + 0.009079434908926487, + 0.00746164983138442, + 0.008473987691104412, + 0.016785411164164543, + 0.01738717593252659, + 0.00907100923359394, + 0.007275858893990517, + 0.012225217185914516, + 0.007574736140668392, + 0.008252344094216824, + 0.011443848721683025, + 0.008725670166313648, + 0.010927846655249596, + 0.010581486858427525, + 0.01015846710652113, + 0.013761764392256737, + 0.009218648076057434, + 0.00528355548158288, + 0.011268538422882557, + 0.0036922935396432877, + 0.009740122593939304, + 0.007995487190783024, + 0.005812361836433411, + 0.011733793653547764, + 0.005568536929786205, + 0.0038972257170826197, + 0.008824639022350311, + 0.004872295074164867, + 0.008861699141561985, + 0.005206608213484287, + 0.006586009170860052, + 0.0068185594864189625, + 0.008572079241275787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.018208375200629234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798679381608963 + ], + "opposite_logprob_tokens": [ + -1.8048679828643799 + ], + "logprob_reward": -0.1798679381608963, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798679381608963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06776732951402664, + 0.09356085956096649, + 0.06854648143053055, + 0.1800459921360016, + 0.11942330002784729, + 0.1616181880235672, + 0.036659155040979385, + 0.1691196709871292, + 0.06892126798629761, + 0.03433782607316971 + ], + "credit_map": [ + 0.00592613872140646, + 0.008389449678361416, + 0.006243079900741577, + 0.01069106999784708, + 0.008625839836895466, + 0.009837653487920761, + 0.009604247286915779, + 0.008449843153357506, + 0.005759669467806816, + 0.008165525272488594, + 0.005636577028781176, + 0.024252258241176605, + 0.008049121126532555, + 0.015028807334601879, + 0.0175174493342638, + 0.009151450358331203, + 0.006172480061650276, + 0.005315036978572607, + 0.004407054744660854, + 0.00972160417586565, + 0.006802945863455534, + 0.006878783460706472, + 0.017938047647476196, + 0.01131052989512682, + 0.010175671428442001, + 0.006312730256468058, + 0.0053979684598743916, + 0.013982409611344337, + 0.010858248919248581, + 0.009147600270807743, + 0.07144387811422348, + 0.05272747576236725, + 0.017767157405614853, + 0.014521009288728237, + 0.009205980226397514, + 0.02222723700106144, + 0.02121979370713234, + 0.011159119196236134, + 0.015593263320624828, + 0.007729729637503624, + 0.020177047699689865, + 0.01443140022456646, + 0.015306169167160988, + 0.018400151282548904, + 0.028364144265651703, + 0.026608219370245934, + 0.018099330365657806, + 0.020231718197464943, + 0.003567057428881526, + 0.0046422164887189865, + 0.006460024043917656, + 0.004859435837715864, + 0.0035960176028311253, + 0.004626584704965353, + 0.005078290123492479, + 0.0038295285776257515, + 0.017492270097136497, + 0.021718597039580345, + 0.01900213025510311, + 0.023928260430693626, + 0.01860320009291172, + 0.021821927279233932, + 0.030086345970630646, + 0.016466941684484482, + 0.006810750346630812, + 0.015277465805411339, + 0.004620616789907217, + 0.011904682964086533, + 0.009585118852555752, + 0.004554517567157745, + 0.011508267372846603, + 0.004659854341298342, + 0.004108866211026907, + 0.006734630558639765, + 0.004281255882233381, + 0.00367023516446352, + 0.0034592740703374147, + 0.0038835909217596054, + 0.003949606791138649, + 0.004250362049788237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.1798679381608963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521222233772278 + ], + "opposite_logprob_tokens": [ + -1.502122163772583 + ], + "logprob_reward": -0.2521222233772278, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521222233772278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09499043971300125, + 0.04985164850950241, + 0.13815142214298248, + 0.38852134346961975, + 0.05380069836974144, + 0.07449065893888474, + 0.06141182780265808, + 0.0697113499045372, + 0.0488600917160511, + 0.020210539922118187 + ], + "credit_map": [ + 0.005446935072541237, + 0.007733976934105158, + 0.006417720578610897, + 0.02262592315673828, + 0.01057584397494793, + 0.015455248765647411, + 0.0116572305560112, + 0.01507754810154438, + 0.0038414872251451015, + 0.005291253328323364, + 0.0034992217551916838, + 0.01171665359288454, + 0.005798596888780594, + 0.007772332988679409, + 0.008323639631271362, + 0.0036084603052586317, + 0.015102026984095573, + 0.01082407496869564, + 0.007598477881401777, + 0.022552579641342163, + 0.013433577492833138, + 0.008761338889598846, + 0.04346894100308418, + 0.01641038991510868, + 0.022134942933917046, + 0.013851162046194077, + 0.014402469620108604, + 0.029901182278990746, + 0.02459375560283661, + 0.01978622004389763, + 0.15336881577968597, + 0.11048275232315063, + 0.005996011663228273, + 0.005023005418479443, + 0.0051515232771635056, + 0.014569796621799469, + 0.004780731629580259, + 0.003594414098188281, + 0.008394959382712841, + 0.006290253251791, + 0.007147848606109619, + 0.012452167458832264, + 0.006570062134414911, + 0.008012069389224052, + 0.01310544554144144, + 0.012686481699347496, + 0.008627040311694145, + 0.005889532621949911, + 0.0051206620410084724, + 0.005289861466735601, + 0.008008868433535099, + 0.008311244659125805, + 0.006072327960282564, + 0.0056794751435518265, + 0.015212885104119778, + 0.00771650206297636, + 0.006355012767016888, + 0.01076722051948309, + 0.006186629645526409, + 0.010712679475545883, + 0.008708561770617962, + 0.009184286929666996, + 0.011416194029152393, + 0.006380751729011536, + 0.004685370251536369, + 0.010908223688602448, + 0.0027699382044374943, + 0.008693483658134937, + 0.006349631585180759, + 0.003584897378459573, + 0.008150882087647915, + 0.0037176613695919514, + 0.0018429853953421116, + 0.0026361439377069473, + 0.003647168865427375, + 0.0025318048428744078, + 0.0020609113853424788, + 0.00188790587708354, + 0.0029202867299318314, + 0.002683330327272415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.2521222233772278 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking medium shot, filmed from the side, follows a traditional Chinese red sedan chair, exquisitely constructed and vibrantly colored. Constructed of vermilion lacquered wood, it reflects a warm, rich luster in the bright sunlight. Its walls are intricately carved with auspicious dragon and phoenix motifs and embellished with gleaming gold ornaments. Its roof features an ornate peaked roof, adorned with long golden tassels at the corners. Its curtains, tightly drawn from thick red silk, add a touch of mystery. Several bearers, dressed in uniform red uniforms and felt hats, move with steady, coordinated steps, carrying two sturdy red poles. With each step, the sedan chair rises and falls rhythmically, gently and steadily, along the quaint cobblestone streets. The golden tassels sway in sync. Against the backdrop of lively voices and the traditional Chinese architecture, the entire scene is imbued with a rich, festive, and solemn atmosphere of traditional celebration.", + "seed": 4197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_9/index_1741_seed_4197_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13834357671439648, + "vlm_logprob": -0.13834357671439648, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.13834357671439648, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01780455932021141 + ], + "opposite_logprob_tokens": [ + -4.2678046226501465 + ], + "logprob_reward": -0.01780455932021141, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01780455932021141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16194026172161102, + 0.07202879339456558, + 0.09311958402395248, + 0.12151600420475006, + 0.02907641790807247, + 0.030231522396206856, + 0.08481631428003311, + 0.2895733416080475, + 0.055105894804000854, + 0.06259187310934067 + ], + "credit_map": [ + 0.01330758910626173, + 0.011319044977426529, + 0.02551468461751938, + 0.035771626979112625, + 0.01768716238439083, + 0.011917047202587128, + 0.02897510677576065, + 0.017448002472519875, + 0.006484244018793106, + 0.006130919326096773, + 0.007600743789225817, + 0.01119440421462059, + 0.007598034106194973, + 0.009073494002223015, + 0.009017465636134148, + 0.01492949016392231, + 0.008028361946344376, + 0.007234831806272268, + 0.008530084043741226, + 0.018718313425779343, + 0.010357958264648914, + 0.007212861441075802, + 0.023177647963166237, + 0.009859520010650158, + 0.006178547628223896, + 0.006449813023209572, + 0.036646414548158646, + 0.041349463164806366, + 0.005463192705065012, + 0.004056294448673725, + 0.01683160290122032, + 0.0045406813733279705, + 0.0024732353631407022, + 0.0019450924592092633, + 0.0032581938430666924, + 0.004754700232297182, + 0.0029209943022578955, + 0.003490723203867674, + 0.00580997858196497, + 0.00442349910736084, + 0.0024943272583186626, + 0.0021761953830718994, + 0.0029959063977003098, + 0.005183815024793148, + 0.003002622164785862, + 0.00379512389190495, + 0.006018619053065777, + 0.004564912524074316, + 0.008842483162879944, + 0.009198999963700771, + 0.005774638615548611, + 0.012271950021386147, + 0.011018546298146248, + 0.0108671635389328, + 0.012681504711508751, + 0.014161028899252415, + 0.02179509587585926, + 0.037536706775426865, + 0.02283416874706745, + 0.03951248899102211, + 0.033749498426914215, + 0.029683981090784073, + 0.04604094475507736, + 0.058420486748218536, + 0.0025940698105841875, + 0.005001236218959093, + 0.0028797555714845657, + 0.01690383441746235, + 0.0022656102664768696, + 0.01216091774404049, + 0.006293430458754301, + 0.007007040549069643, + 0.006145711522549391, + 0.003357408568263054, + 0.0047386204823851585, + 0.0036076377145946026, + 0.005042238626629114, + 0.020475490018725395, + 0.006715334020555019, + 0.012509433552622795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.01780455932021141 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210777536034584 + ], + "opposite_logprob_tokens": [ + -2.812107801437378 + ], + "logprob_reward": -0.06210777536034584, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210777536034584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07910583913326263, + 0.16677135229110718, + 0.17864890396595, + 0.23473061621189117, + 0.020844899117946625, + 0.024809451773762703, + 0.040201686322689056, + 0.07774606347084045, + 0.06037108972668648, + 0.11677005887031555 + ], + "credit_map": [ + 0.006502327974885702, + 0.00398356793448329, + 0.008425538428127766, + 0.017234420403838158, + 0.00515380036085844, + 0.011997793801128864, + 0.01284948829561472, + 0.012958903796970844, + 0.01377780456095934, + 0.007006027270108461, + 0.01485329307615757, + 0.025647709146142006, + 0.01118352822959423, + 0.03240448981523514, + 0.023583348840475082, + 0.03831516206264496, + 0.01450798287987709, + 0.012172766961157322, + 0.01752069592475891, + 0.029510099440813065, + 0.02085237205028534, + 0.0180039219558239, + 0.0427231527864933, + 0.023357925936579704, + 0.011351491324603558, + 0.010350576601922512, + 0.07667740434408188, + 0.0850682407617569, + 0.006633492652326822, + 0.006230881903320551, + 0.03065684251487255, + 0.007761698681861162, + 0.0024198677856475115, + 0.0015273416647687554, + 0.0016645743744447827, + 0.0036513314116746187, + 0.0020790714770555496, + 0.002179438713937998, + 0.004046368878334761, + 0.003276906907558441, + 0.0018306473502889276, + 0.0016181760001927614, + 0.0033132622484117746, + 0.004101220518350601, + 0.0023965532891452312, + 0.0028726125601679087, + 0.004920430481433868, + 0.0037565501406788826, + 0.0029520585667341948, + 0.00468130549415946, + 0.00418425677344203, + 0.0049841804429888725, + 0.005660923197865486, + 0.005110419820994139, + 0.00664289528504014, + 0.005985649302601814, + 0.0067536486312747, + 0.01033044047653675, + 0.006360333412885666, + 0.010095098055899143, + 0.009083989076316357, + 0.007863929495215416, + 0.012834264896810055, + 0.014424357563257217, + 0.003013615496456623, + 0.00621408224105835, + 0.005205707158893347, + 0.015575319528579712, + 0.0037938926834613085, + 0.00845258217304945, + 0.008935471065342426, + 0.009180421940982342, + 0.00926175806671381, + 0.005178052932024002, + 0.015956956893205643, + 0.009622368030250072, + 0.008474401198327541, + 0.03214583918452263, + 0.013438010588288307, + 0.02269267849624157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.06210777536034584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637499012053013 + ], + "opposite_logprob_tokens": [ + -3.6513750553131104 + ], + "logprob_reward": -0.02637499012053013, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637499012053013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22075504064559937, + 0.07866302877664566, + 0.0775732696056366, + 0.1969982087612152, + 0.047628115862607956, + 0.07273493707180023, + 0.10656009614467621, + 0.12943272292613983, + 0.025293130427598953, + 0.04436146840453148 + ], + "credit_map": [ + 0.018932310864329338, + 0.012712685391306877, + 0.03245743736624718, + 0.04758270084857941, + 0.017794819548726082, + 0.02891334518790245, + 0.03571060672402382, + 0.026651129126548767, + 0.008077542297542095, + 0.006794307380914688, + 0.008667992427945137, + 0.015081910416483879, + 0.008931964635848999, + 0.012426257133483887, + 0.008921691216528416, + 0.009761366061866283, + 0.006160448305308819, + 0.005374528467655182, + 0.008919809944927692, + 0.015109981410205364, + 0.00861374568194151, + 0.0072235180996358395, + 0.017315823584794998, + 0.008855417370796204, + 0.008668210357427597, + 0.008031307719647884, + 0.054486874490976334, + 0.06396850198507309, + 0.021861519664525986, + 0.0051415180787444115, + 0.02768273651599884, + 0.00715754646807909, + 0.005163225345313549, + 0.0036189022939652205, + 0.004811613354831934, + 0.007270422298461199, + 0.004669227171689272, + 0.006703054532408714, + 0.009574158117175102, + 0.0058175125159323215, + 0.005689593963325024, + 0.004408103413879871, + 0.007155724335461855, + 0.01205395720899105, + 0.007127978373318911, + 0.010148432105779648, + 0.0126032754778862, + 0.013547873124480247, + 0.007849941961467266, + 0.012540124356746674, + 0.007867795415222645, + 0.013596556149423122, + 0.01911330223083496, + 0.018141834065318108, + 0.013602722436189651, + 0.013847816735506058, + 0.013543701730668545, + 0.010620816610753536, + 0.00898685958236456, + 0.018738504499197006, + 0.010720109567046165, + 0.014538331888616085, + 0.023090826347470284, + 0.029193563386797905, + 0.001980804605409503, + 0.0028201090171933174, + 0.003545478917658329, + 0.0034299076069146395, + 0.00203258590772748, + 0.003155242186039686, + 0.0036897321697324514, + 0.0046392702497541904, + 0.004132423549890518, + 0.0027746488340198994, + 0.006685828790068626, + 0.004069105722010136, + 0.003340217750519514, + 0.01081026904284954, + 0.004455423913896084, + 0.008093548007309437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.02637499012053013 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203853711485863 + ], + "opposite_logprob_tokens": [ + -2.8120384216308594 + ], + "logprob_reward": -0.06203853711485863, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203853711485863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0845131129026413, + 0.036847494542598724, + 0.06857622414827347, + 0.2859432101249695, + 0.027281805872917175, + 0.06285462528467178, + 0.19931763410568237, + 0.09052681922912598, + 0.12685127556324005, + 0.017287829890847206 + ], + "credit_map": [ + 0.007046874612569809, + 0.006646028719842434, + 0.02022388018667698, + 0.018101109191775322, + 0.0050870259292423725, + 0.00546037033200264, + 0.01080014556646347, + 0.011147686280310154, + 0.003760933643206954, + 0.0027319954242557287, + 0.006807192228734493, + 0.005357428919523954, + 0.003582791658118367, + 0.0032974262721836567, + 0.006456488277763128, + 0.004853235557675362, + 0.003934095613658428, + 0.0042310720309615135, + 0.007814797572791576, + 0.018404385074973106, + 0.005130327772349119, + 0.004996618255972862, + 0.017723795026540756, + 0.006341134663671255, + 0.014967259019613266, + 0.012825983576476574, + 0.0925249308347702, + 0.09719236940145493, + 0.013438782654702663, + 0.00809417013078928, + 0.03652341291308403, + 0.010376314632594585, + 0.0025072831194847822, + 0.0019421089673414826, + 0.0029600972775369883, + 0.004023075103759766, + 0.0035253155510872602, + 0.0038282484747469425, + 0.00550349336117506, + 0.0029921827372163534, + 0.005392927210777998, + 0.005098831374198198, + 0.008963660337030888, + 0.009833387099206448, + 0.005431751720607281, + 0.007783498149365187, + 0.010968087241053581, + 0.009382475167512894, + 0.018390148878097534, + 0.016394225880503654, + 0.01707790419459343, + 0.0334792360663414, + 0.024797767400741577, + 0.02347624860703945, + 0.031470153480768204, + 0.03423194959759712, + 0.007300569675862789, + 0.014133014716207981, + 0.008797472342848778, + 0.010717442259192467, + 0.01204525213688612, + 0.009768753312528133, + 0.010894495993852615, + 0.016869820654392242, + 0.006904703564941883, + 0.014744451269507408, + 0.014087466523051262, + 0.031098337844014168, + 0.006768169812858105, + 0.013747685588896275, + 0.02112346701323986, + 0.01837698370218277, + 0.0009449149947613478, + 0.0020379365887492895, + 0.0019036013400182128, + 0.001526199164800346, + 0.0017869073199108243, + 0.0027688988484442234, + 0.002580806612968445, + 0.003738565370440483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.06203853711485863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233920216560364 + ], + "opposite_logprob_tokens": [ + -0.8983920216560364 + ], + "logprob_reward": -0.5233920216560364, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233920216560364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17020905017852783, + 0.1636883020401001, + 0.04925236105918884, + 0.0776134580373764, + 0.03290579840540886, + 0.056792907416820526, + 0.06553782522678375, + 0.25795066356658936, + 0.07063917815685272, + 0.055410441011190414 + ], + "credit_map": [ + 0.012905827723443508, + 0.011080742813646793, + 0.023614415898919106, + 0.04074104130268097, + 0.015832174569368362, + 0.022055689245462418, + 0.02176954783499241, + 0.022209594026207924, + 0.014125206507742405, + 0.010470183566212654, + 0.016880644485354424, + 0.030765507370233536, + 0.014134419150650501, + 0.029311787337064743, + 0.02212907187640667, + 0.025871489197015762, + 0.003572918241843581, + 0.003314042929559946, + 0.0041667926125228405, + 0.01351421419531107, + 0.0047506713308393955, + 0.0033164257183670998, + 0.012272978201508522, + 0.004344320390373468, + 0.004467520862817764, + 0.0038571131881326437, + 0.022709280252456665, + 0.023585818707942963, + 0.004166129976511002, + 0.003344128839671612, + 0.01130533404648304, + 0.004178134724497795, + 0.003668057732284069, + 0.0020438102073967457, + 0.004242343828082085, + 0.004757029470056295, + 0.002352282637730241, + 0.005702697206288576, + 0.005725501570850611, + 0.004414075519889593, + 0.00396468210965395, + 0.0032580255065113306, + 0.0060265217907726765, + 0.009945781901478767, + 0.006085371132940054, + 0.00855772290378809, + 0.009285825304687023, + 0.00966897513717413, + 0.006401063408702612, + 0.00719514861702919, + 0.00825952272862196, + 0.010486474260687828, + 0.007087766192853451, + 0.008235187269747257, + 0.007687161676585674, + 0.010185496881604195, + 0.040445681661367416, + 0.045415494590997696, + 0.029509330168366432, + 0.026168638840317726, + 0.03722159191966057, + 0.019984479993581772, + 0.028302226215600967, + 0.03090321086347103, + 0.005383682437241077, + 0.0071749333292245865, + 0.011140349321067333, + 0.012353547848761082, + 0.005208997987210751, + 0.012535320594906807, + 0.008115970529615879, + 0.008726373314857483, + 0.005832410417497158, + 0.0027791340835392475, + 0.0035870117135345936, + 0.0021583614870905876, + 0.004351359326392412, + 0.020061112940311432, + 0.006154377944767475, + 0.010486675426363945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.5233920216560364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking medium shot, filmed from the side, follows a traditional Chinese red sedan chair, exquisitely constructed and vibrantly colored. Constructed of vermilion lacquered wood, it reflects a warm, rich luster in the bright sunlight. Its walls are intricately carved with auspicious dragon and phoenix motifs and embellished with gleaming gold ornaments. Its roof features an ornate peaked roof, adorned with long golden tassels at the corners. Its curtains, tightly drawn from thick red silk, add a touch of mystery. Several bearers, dressed in uniform red uniforms and felt hats, move with steady, coordinated steps, carrying two sturdy red poles. With each step, the sedan chair rises and falls rhythmically, gently and steadily, along the quaint cobblestone streets. The golden tassels sway in sync. Against the backdrop of lively voices and the traditional Chinese architecture, the entire scene is imbued with a rich, festive, and solemn atmosphere of traditional celebration.", + "seed": 4198, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000111/group_9/index_1741_seed_4198_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08723290544003248, + "vlm_logprob": -0.08723290544003248, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.08723290544003248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009133817628026009 + ], + "opposite_logprob_tokens": [ + -4.884133815765381 + ], + "logprob_reward": -0.009133817628026009, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009133817628026009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15690556168556213, + 0.10808615386486053, + 0.039879344403743744, + 0.05675937607884407, + 0.06442258507013321, + 0.13749468326568604, + 0.11345072835683823, + 0.19353190064430237, + 0.06818720698356628, + 0.061282433569431305 + ], + "credit_map": [ + 0.014034036546945572, + 0.03311393782496452, + 0.016986850649118423, + 0.02034677192568779, + 0.022009767591953278, + 0.013365470804274082, + 0.023935172706842422, + 0.013113554567098618, + 0.007945926859974861, + 0.012182541191577911, + 0.016264455392956734, + 0.019043631851673126, + 0.010763452388346195, + 0.009205324575304985, + 0.023495489731431007, + 0.009185331873595715, + 0.003267585998401046, + 0.004061485640704632, + 0.006968147121369839, + 0.007220973260700703, + 0.004524573218077421, + 0.003932533320039511, + 0.006010256242007017, + 0.0038937872741371393, + 0.004828372038900852, + 0.006602725014090538, + 0.004328885581344366, + 0.012963675893843174, + 0.0051152887754142284, + 0.004418613389134407, + 0.011041399091482162, + 0.007460411172360182, + 0.0041686175391077995, + 0.006547591648995876, + 0.007923562079668045, + 0.01545916311442852, + 0.00733962282538414, + 0.005513327661901712, + 0.01200788002461195, + 0.005462817847728729, + 0.007417058572173119, + 0.016406655311584473, + 0.010896185413002968, + 0.02896110713481903, + 0.015404519625008106, + 0.011359119787812233, + 0.023906804621219635, + 0.02314324676990509, + 0.006132970098406076, + 0.014617018401622772, + 0.011178769171237946, + 0.017272252589464188, + 0.020503446459770203, + 0.019785428419709206, + 0.01110231876373291, + 0.012858533300459385, + 0.013677174225449562, + 0.02115253172814846, + 0.008776187896728516, + 0.016649380326271057, + 0.04162268340587616, + 0.037126585841178894, + 0.02609035186469555, + 0.028437012806534767, + 0.003994021099060774, + 0.0032912895549088717, + 0.0031846819911152124, + 0.002608306473121047, + 0.020282773301005363, + 0.020635388791561127, + 0.007409596350044012, + 0.006781158968806267, + 0.01119212806224823, + 0.012055197730660439, + 0.005992519203573465, + 0.007373637519776821, + 0.005205759312957525, + 0.006504669785499573, + 0.005984900519251823, + 0.006973625160753727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.009133817628026009 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812346234917641 + ], + "opposite_logprob_tokens": [ + -3.288123369216919 + ], + "logprob_reward": -0.03812346234917641, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812346234917641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11494293063879013, + 0.058782417327165604, + 0.050088319927453995, + 0.06162014603614807, + 0.05081731826066971, + 0.07476560771465302, + 0.16162702441215515, + 0.18636628985404968, + 0.0648294985294342, + 0.17616043984889984 + ], + "credit_map": [ + 0.00829287152737379, + 0.016993677243590355, + 0.011014331132173538, + 0.012977872975170612, + 0.0181229617446661, + 0.011717240326106548, + 0.021012892946600914, + 0.014811079017817974, + 0.0035479748621582985, + 0.005701270420104265, + 0.0054902201518416405, + 0.008226417936384678, + 0.005721777677536011, + 0.010541019961237907, + 0.010723482817411423, + 0.008830258622765541, + 0.002991805551573634, + 0.006910127121955156, + 0.008301053196191788, + 0.010698524303734303, + 0.003887240309268236, + 0.006124489009380341, + 0.005400876980274916, + 0.005774203222244978, + 0.004771459847688675, + 0.009141789749264717, + 0.003734822617843747, + 0.010536289773881435, + 0.012554313987493515, + 0.004874363075941801, + 0.010438713245093822, + 0.0055683935061097145, + 0.0034661602694541216, + 0.005992052610963583, + 0.006304136943072081, + 0.008213351480662823, + 0.009438886307179928, + 0.004676751792430878, + 0.00840773992240429, + 0.004318237770348787, + 0.0036487376783043146, + 0.008064203895628452, + 0.004866929724812508, + 0.016312777996063232, + 0.0082326028496027, + 0.006805088836699724, + 0.016043931245803833, + 0.010791338048875332, + 0.015025492757558823, + 0.027674319222569466, + 0.015361707657575607, + 0.02226422354578972, + 0.030101129785180092, + 0.021139420568943024, + 0.01434787455946207, + 0.015712833032011986, + 0.011985625140368938, + 0.02172168716788292, + 0.010329033248126507, + 0.017866145819425583, + 0.0404127836227417, + 0.0356471911072731, + 0.0258803628385067, + 0.022523459047079086, + 0.00548973074182868, + 0.0064065721817314625, + 0.005877955816686153, + 0.006026553921401501, + 0.009763342328369617, + 0.00882643274962902, + 0.011164559982717037, + 0.011274349875748158, + 0.0313778892159462, + 0.04208139330148697, + 0.0185142382979393, + 0.02556484192609787, + 0.014936906285583973, + 0.016568420454859734, + 0.014247267507016659, + 0.012869457714259624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.03812346234917641 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791442193090916 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.009791442193090916, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791442193090916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13212314248085022, + 0.026891248300671577, + 0.05125578120350838, + 0.03313525393605232, + 0.04019808769226074, + 0.23302879929542542, + 0.1768127679824829, + 0.21416190266609192, + 0.06326949596405029, + 0.02912352792918682 + ], + "credit_map": [ + 0.01302359439432621, + 0.023898061364889145, + 0.013456817716360092, + 0.015913555398583412, + 0.016841480508446693, + 0.012141489423811436, + 0.020360508933663368, + 0.01648763008415699, + 0.002177231712266803, + 0.0033035355154424906, + 0.003403083886951208, + 0.0038042382802814245, + 0.0029665546026080847, + 0.0029518601950258017, + 0.003927013371139765, + 0.004357730969786644, + 0.003419824643060565, + 0.007449941709637642, + 0.008416828699409962, + 0.012022376991808414, + 0.004396576900035143, + 0.005140363704413176, + 0.0053566498681902885, + 0.005053217988461256, + 0.002495791530236602, + 0.0035830310080200434, + 0.002360857091844082, + 0.0074431211687624454, + 0.005114294122904539, + 0.003271512221544981, + 0.005820603109896183, + 0.0030460446141660213, + 0.0033873619977384806, + 0.004804303403943777, + 0.00357377459295094, + 0.00688943499699235, + 0.007197890896350145, + 0.0030656694434583187, + 0.007930141873657703, + 0.0033495118841528893, + 0.010952005162835121, + 0.02640797570347786, + 0.01746726967394352, + 0.048372477293014526, + 0.02873518131673336, + 0.017672860994935036, + 0.043212391436100006, + 0.04020863398909569, + 0.010290202684700489, + 0.030823376029729843, + 0.01425692904740572, + 0.02366410568356514, + 0.034581802785396576, + 0.03606349229812622, + 0.012498307041823864, + 0.014634539373219013, + 0.011999428272247314, + 0.027163902297616005, + 0.009690570645034313, + 0.02155766636133194, + 0.04593071714043617, + 0.03642025217413902, + 0.033508289605379105, + 0.02789108268916607, + 0.0039027256425470114, + 0.003072792664170265, + 0.0032094235066324472, + 0.002643159357830882, + 0.0179822389036417, + 0.018231498077511787, + 0.007259370759129524, + 0.006968286819756031, + 0.00334856822155416, + 0.005472101271152496, + 0.0028782340232282877, + 0.003545950399711728, + 0.0027728325221687555, + 0.0030053602531552315, + 0.0034872619435191154, + 0.004613220226019621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.009791442193090916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702055275440216 + ], + "opposite_logprob_tokens": [ + -2.1270205974578857 + ], + "logprob_reward": -0.12702055275440216, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702055275440216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10277511179447174, + 0.049345895648002625, + 0.09192728251218796, + 0.1095358282327652, + 0.10607732832431793, + 0.16545391082763672, + 0.17940127849578857, + 0.12034539878368378, + 0.04929807037115097, + 0.025839930400252342 + ], + "credit_map": [ + 0.011429146863520145, + 0.02467774972319603, + 0.011614018119871616, + 0.014155462384223938, + 0.010420667007565498, + 0.01025361754000187, + 0.011966832913458347, + 0.008257613517343998, + 0.004269499331712723, + 0.0052030328661203384, + 0.006162552163004875, + 0.008876638486981392, + 0.005488994996994734, + 0.007257669232785702, + 0.00470865610986948, + 0.007378848735243082, + 0.006724835839122534, + 0.013268657959997654, + 0.013399296440184116, + 0.021245447918772697, + 0.00877392292022705, + 0.00864875316619873, + 0.01215344201773405, + 0.007712929975241423, + 0.007754483260214329, + 0.011877276003360748, + 0.012449473142623901, + 0.02607521042227745, + 0.01619604416191578, + 0.008073853328824043, + 0.018159111961722374, + 0.008950367569923401, + 0.008566959761083126, + 0.014049489051103592, + 0.006390346679836512, + 0.019088225439190865, + 0.010800599120557308, + 0.007025879807770252, + 0.030665405094623566, + 0.00949042197316885, + 0.00936143659055233, + 0.01862691342830658, + 0.01277869287878275, + 0.03152814880013466, + 0.01776900328695774, + 0.01923554763197899, + 0.028079872950911522, + 0.028074296191334724, + 0.010120315477252007, + 0.0343133918941021, + 0.006016709841787815, + 0.022307049483060837, + 0.037622641772031784, + 0.038508541882038116, + 0.014448328875005245, + 0.01606430485844612, + 0.022543281316757202, + 0.013442225754261017, + 0.006429798901081085, + 0.011301334947347641, + 0.021882381290197372, + 0.023452650755643845, + 0.011387691833078861, + 0.009906038641929626, + 0.006640677340328693, + 0.007434059400111437, + 0.004680723883211613, + 0.004822760354727507, + 0.007092800922691822, + 0.007036732509732246, + 0.006122891791164875, + 0.00546742370352149, + 0.002912285039201379, + 0.003647181438282132, + 0.002505453070625663, + 0.0033814150374382734, + 0.001852881396189332, + 0.0036040854174643755, + 0.003637473564594984, + 0.0042991554364562035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.12702055275440216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a traditional Chinese red sedan chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bearers wearing uniform red uniforms and felt hats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair constructed of vermilion lacquered wood with intricate dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a horizontal tracking shot filmed from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a quaint cobblestone street with traditional Chinese architecture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520952522754669 + ], + "opposite_logprob_tokens": [ + -1.5020952224731445 + ], + "logprob_reward": -0.2520952522754669, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520952522754669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22242547571659088, + 0.049785930663347244, + 0.07859578728675842, + 0.2859766483306885, + 0.049876194447278976, + 0.07103250920772552, + 0.06598566472530365, + 0.07778896391391754, + 0.059838347136974335, + 0.038694530725479126 + ], + "credit_map": [ + 0.019631005823612213, + 0.03717732056975365, + 0.025767607614398003, + 0.02738552913069725, + 0.020153388381004333, + 0.022379036992788315, + 0.03942791000008583, + 0.03050367347896099, + 0.004400582984089851, + 0.007155850995332003, + 0.005570304114371538, + 0.006615079939365387, + 0.006798712071031332, + 0.006085491739213467, + 0.007405674085021019, + 0.00575423426926136, + 0.004416998941451311, + 0.010648174211382866, + 0.013758246786892414, + 0.018075156956911087, + 0.006101007107645273, + 0.008490514010190964, + 0.008701879531145096, + 0.008403806947171688, + 0.019138412550091743, + 0.04476486146450043, + 0.01834164746105671, + 0.04889442399144173, + 0.06571440398693085, + 0.021141991019248962, + 0.047250617295503616, + 0.020730290561914444, + 0.003237619064748287, + 0.004931629169732332, + 0.004494020715355873, + 0.013343242928385735, + 0.006917402148246765, + 0.003225171472877264, + 0.008852195926010609, + 0.004874910227954388, + 0.004689797759056091, + 0.019605373963713646, + 0.004586060997098684, + 0.00886626448482275, + 0.015218596905469894, + 0.005410605575889349, + 0.008065675385296345, + 0.004590126685798168, + 0.006092284806072712, + 0.009458237327635288, + 0.0053240349516272545, + 0.009140740148723125, + 0.011962116695940495, + 0.008358386345207691, + 0.008798127062618732, + 0.006851738318800926, + 0.006473642308264971, + 0.011051888577640057, + 0.004427682142704725, + 0.008105532266199589, + 0.014786883257329464, + 0.012826548889279366, + 0.01020493358373642, + 0.009911859408020973, + 0.005271872039884329, + 0.005224773660302162, + 0.0034204847179353237, + 0.004374914802610874, + 0.014677107334136963, + 0.01506812684237957, + 0.0073176282458007336, + 0.004483439959585667, + 0.004690074827522039, + 0.005423570517450571, + 0.0037575350143015385, + 0.004520733840763569, + 0.003583734156563878, + 0.00486898934468627, + 0.005652421619743109, + 0.006197470240294933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.2520952522754669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_0/index_tensor([6284, 6284])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_0/index_tensor([6284, 6284])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7a60b914c2d3b223251a7bc41d89117b94476456 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_0/index_tensor([6284, 6284])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up macro lens is firmly locked on the top of a large bowl of steaming pickled fish. A shriveled and bright red dried chili pepper lies quietly on the golden pickled fish soup and a few tender white fish fillets. Its surface is covered with fine wrinkles, foreshadowing the coming storm. Suddenly, a stream of boiling, clear oil is poured down from outside the frame. At the moment the hot oil contacts the chili peppers with a sharp sizzling sound, the camera captures in slow motion the chili peppers’ skins rapidly shrinking and bursting due to the high temperature, the color becoming deeper and glossier, releasing a rich spicy aroma. At the same time, the camera begins to perform a classic Hitchcock zoom (Dolly Zoom), that is, the camera pulls back while the lens focus simultaneously moves forward, causing the blurred Chinese restaurant environment in the background - including wooden tables and chairs under warm lights and the figures of diners - to be visually distorted and rapidly recede backwards, while the size of the main body of the exploding chili peppers in the foreground strangely remains unchanged, creating a strong sense of visual dizziness and highly dramatic tension. The entire picture is full of shocking sensory impact, and the atmosphere is tense and fascinating.", + "seed": 23811, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_0/index_6284_seed_23811_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2199761825148016, + "vlm_logprob": -0.2199761825148016, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.2199761825148016, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7676981687545776 + ], + "opposite_logprob_tokens": [ + -0.6426981687545776 + ], + "logprob_reward": -0.7676981687545776, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7676981687545776, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1461460441350937, + 0.04700431972742081, + 0.02551407366991043, + 0.06753126531839371, + 0.0500260628759861, + 0.055528271943330765, + 0.047258615493774414, + 0.2451716959476471, + 0.10706725716590881, + 0.20875243842601776 + ], + "credit_map": [ + 0.015846753492951393, + 0.028248896822333336, + 0.015282928943634033, + 0.01889290288090706, + 0.01934739015996456, + 0.016856160014867783, + 0.015455414541065693, + 0.01621561124920845, + 0.006007782649248838, + 0.010706892237067223, + 0.004693006630986929, + 0.006118750665336847, + 0.004748322069644928, + 0.004119924269616604, + 0.004606373608112335, + 0.006003264803439379, + 0.002952061826363206, + 0.004950773436576128, + 0.0029936523642390966, + 0.0035100444220006466, + 0.0029659487772732973, + 0.0025572124868631363, + 0.002873821184039116, + 0.002710560569539666, + 0.004803015850484371, + 0.007844967767596245, + 0.00831062626093626, + 0.011565257795155048, + 0.010900476016104221, + 0.00844187755137682, + 0.00904250517487526, + 0.006622538901865482, + 0.004084589891135693, + 0.004447792656719685, + 0.00535466568544507, + 0.006862633861601353, + 0.005507158115506172, + 0.005056355148553848, + 0.01042984239757061, + 0.008283027447760105, + 0.006241097580641508, + 0.007292398251593113, + 0.00763352494686842, + 0.005944790784269571, + 0.005617402959614992, + 0.007316985167562962, + 0.00737655209377408, + 0.008105522952973843, + 0.003859687363728881, + 0.00818186067044735, + 0.004834549035876989, + 0.006332546938210726, + 0.004578289110213518, + 0.005485584493726492, + 0.008581118658185005, + 0.005404979456216097, + 0.023893114179372787, + 0.020657990127801895, + 0.03198718652129173, + 0.04053090140223503, + 0.03167346864938736, + 0.032703567296266556, + 0.03387857973575592, + 0.029846878722310066, + 0.0067545282654464245, + 0.013322694227099419, + 0.015173438936471939, + 0.026185275986790657, + 0.012133137322962284, + 0.010765367187559605, + 0.014497949741780758, + 0.008234870620071888, + 0.012547940947115421, + 0.01454389188438654, + 0.061235781759023666, + 0.06489554792642593, + 0.011809317395091057, + 0.009770944714546204, + 0.020651519298553467, + 0.013297474943101406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.7676981687545776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032414060551673174 + ], + "opposite_logprob_tokens": [ + -5.753241539001465 + ], + "logprob_reward": -0.0032414060551673174, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032414060551673174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0850319191813469, + 0.02432326413691044, + 0.01940070278942585, + 0.03102421946823597, + 0.04163159430027008, + 0.06506457179784775, + 0.05084924399852753, + 0.17294541001319885, + 0.12896238267421722, + 0.3807666599750519 + ], + "credit_map": [ + 0.012754169292747974, + 0.012120253406465054, + 0.011282801628112793, + 0.01059282012283802, + 0.008490660227835178, + 0.010934360325336456, + 0.008125695399940014, + 0.010731150396168232, + 0.0027071526274085045, + 0.002871372504159808, + 0.002666399348527193, + 0.003640128765255213, + 0.003332866355776787, + 0.0023600778076797724, + 0.003642129246145487, + 0.0031031332910060883, + 0.0017109104664996266, + 0.001871366985142231, + 0.0029284628108143806, + 0.004050672985613346, + 0.0022177917417138815, + 0.0015743717085570097, + 0.002664379309862852, + 0.0023827445693314075, + 0.0033464482985436916, + 0.0065069496631622314, + 0.002782099647447467, + 0.003149462165310979, + 0.004399897065013647, + 0.003366705495864153, + 0.003922656178474426, + 0.003549998626112938, + 0.003917189314961433, + 0.004023822024464607, + 0.006070576608181, + 0.0047256010584533215, + 0.005368038546293974, + 0.00473914435133338, + 0.006362807005643845, + 0.006424411199986935, + 0.007196734193712473, + 0.009665272198617458, + 0.008315961807966232, + 0.006784883327782154, + 0.00791135523468256, + 0.012846133671700954, + 0.005339997820556164, + 0.007004233077168465, + 0.004284325521439314, + 0.006674288772046566, + 0.0044880216009914875, + 0.004534069914370775, + 0.008550615049898624, + 0.005900547839701176, + 0.008812758140265942, + 0.007604616228491068, + 0.01793469488620758, + 0.016387591138482094, + 0.018951302394270897, + 0.030434969812631607, + 0.020735956728458405, + 0.02509421482682228, + 0.024490801617503166, + 0.01891586184501648, + 0.011769360862672329, + 0.013907979242503643, + 0.01479716319590807, + 0.016365068033337593, + 0.01645190827548504, + 0.014491424895823002, + 0.02081933803856373, + 0.020360130816698074, + 0.04184601828455925, + 0.04284435510635376, + 0.0421583354473114, + 0.05550045892596245, + 0.048158906400203705, + 0.053509194403886795, + 0.05608795955777168, + 0.04066138714551926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.0032414060551673174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867850989103317 + ], + "opposite_logprob_tokens": [ + -3.048678398132324 + ], + "logprob_reward": -0.04867850989103317, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867850989103317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08120863884687424, + 0.023748893290758133, + 0.015553116798400879, + 0.046213939785957336, + 0.027154717594385147, + 0.061372265219688416, + 0.059133101254701614, + 0.1974373310804367, + 0.08233040571212769, + 0.4058476388454437 + ], + "credit_map": [ + 0.009391415864229202, + 0.017021967098116875, + 0.009732484817504883, + 0.012505004182457924, + 0.008236915804445744, + 0.008192145265638828, + 0.008016898296773434, + 0.008111805655062199, + 0.0027368126902729273, + 0.005026382859796286, + 0.0024857448879629374, + 0.003475963370874524, + 0.0022010966204106808, + 0.002011805772781372, + 0.002705093938857317, + 0.003105992218479514, + 0.002095541451126337, + 0.0025785258039832115, + 0.0017810903955250978, + 0.0021017626859247684, + 0.001735233818180859, + 0.0017658793367445469, + 0.0018320600502192974, + 0.0016630211612209678, + 0.005327369086444378, + 0.01351915393024683, + 0.0031571073923259974, + 0.005603031255304813, + 0.005186445079743862, + 0.0044338214211165905, + 0.00419193459674716, + 0.004795071668922901, + 0.0030891699716448784, + 0.0024415396619588137, + 0.002786516910418868, + 0.0025339818093925714, + 0.004020867869257927, + 0.0028193227481096983, + 0.006166515406221151, + 0.0032968027517199516, + 0.0034626233391463757, + 0.005357156973332167, + 0.0060311476700007915, + 0.0033378051593899727, + 0.010935553349554539, + 0.01539636217057705, + 0.006330982781946659, + 0.010520623996853828, + 0.003757466794922948, + 0.004686066880822182, + 0.004461213946342468, + 0.004667441360652447, + 0.004241455812007189, + 0.004289557226002216, + 0.015620849095284939, + 0.017409048974514008, + 0.020329242572188377, + 0.018702274188399315, + 0.024500641971826553, + 0.026197155937552452, + 0.02765716426074505, + 0.02835380658507347, + 0.027081605046987534, + 0.024615414440631866, + 0.006019942928105593, + 0.009772354736924171, + 0.00889454036951065, + 0.011032165959477425, + 0.011892012320458889, + 0.008747813291847706, + 0.014176229014992714, + 0.011795341968536377, + 0.03476894646883011, + 0.053624384105205536, + 0.07497995346784592, + 0.08109499514102936, + 0.048326872289180756, + 0.0300997756421566, + 0.047360509634017944, + 0.03559218347072601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.04867850989103317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799263060092926 + ], + "opposite_logprob_tokens": [ + -1.8049262762069702 + ], + "logprob_reward": -0.1799263060092926, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799263060092926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1721392273902893, + 0.03340863808989525, + 0.015641268342733383, + 0.034581396728754044, + 0.03713687136769295, + 0.05059459060430527, + 0.05618792399764061, + 0.18769031763076782, + 0.14604154229164124, + 0.26657822728157043 + ], + "credit_map": [ + 0.015858473256230354, + 0.028228580951690674, + 0.023365091532468796, + 0.026449715718626976, + 0.012499870732426643, + 0.027158286422491074, + 0.016773249953985214, + 0.021805958822369576, + 0.0028242103289812803, + 0.0033291704021394253, + 0.003775096032768488, + 0.005201300140470266, + 0.004543759394437075, + 0.005937263369560242, + 0.004025185015052557, + 0.003772653406485915, + 0.0015196524327620864, + 0.0017695153364911675, + 0.002009097719565034, + 0.002765863435342908, + 0.002155457390472293, + 0.0015851043863222003, + 0.0019828230142593384, + 0.001853754511103034, + 0.0038248489145189524, + 0.008171088993549347, + 0.0033861203119158745, + 0.004592785611748695, + 0.004560013301670551, + 0.002923123072832823, + 0.0032533223275095224, + 0.0038700937293469906, + 0.003958979621529579, + 0.004839547909796238, + 0.004250525496900082, + 0.00461336737498641, + 0.004128337372094393, + 0.0025611405726522207, + 0.008062951266765594, + 0.004722020588815212, + 0.005484064109623432, + 0.011104785837233067, + 0.005197169724851847, + 0.004089747089892626, + 0.005431720055639744, + 0.008807261474430561, + 0.004795487970113754, + 0.005684358533471823, + 0.004028061404824257, + 0.004206740763038397, + 0.006278802640736103, + 0.005353380460292101, + 0.003934696316719055, + 0.007867558859288692, + 0.009720268659293652, + 0.014798412099480629, + 0.027461819350719452, + 0.02520393580198288, + 0.021085062995553017, + 0.020092284306883812, + 0.022547312080860138, + 0.025903383269906044, + 0.02487940341234207, + 0.020517118275165558, + 0.009911478497087955, + 0.017769917845726013, + 0.01810278370976448, + 0.036718886345624924, + 0.010309197939932346, + 0.012230266816914082, + 0.02682870253920555, + 0.014170310460031033, + 0.01703020930290222, + 0.01879778504371643, + 0.07633668184280396, + 0.07648350298404694, + 0.018719354644417763, + 0.019933292642235756, + 0.019082065671682358, + 0.02019532211124897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.1799263060092926 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033652186393738 + ], + "opposite_logprob_tokens": [ + -2.3503365516662598 + ], + "logprob_reward": -0.10033652186393738, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033652186393738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1621202826499939, + 0.03611372783780098, + 0.04060106724500656, + 0.025460010394454002, + 0.05507886782288551, + 0.08853743225336075, + 0.0505303256213665, + 0.4186091125011444, + 0.0630907267332077, + 0.05985843017697334 + ], + "credit_map": [ + 0.015010599978268147, + 0.03287804126739502, + 0.01795918680727482, + 0.022106828168034554, + 0.018015485256910324, + 0.023510854691267014, + 0.014563038945198059, + 0.01807624101638794, + 0.004664616193622351, + 0.004031711257994175, + 0.0037029655650258064, + 0.005890598986297846, + 0.004714546725153923, + 0.004086881875991821, + 0.005550154950469732, + 0.003472250420600176, + 0.004715152084827423, + 0.005379179026931524, + 0.005206170491874218, + 0.005774661432951689, + 0.005989816039800644, + 0.004837662912905216, + 0.004718891344964504, + 0.003979532513767481, + 0.002155944472178817, + 0.004248499870300293, + 0.0029189197812229395, + 0.003614322282373905, + 0.0032439129427075386, + 0.002794738858938217, + 0.0032331368420273066, + 0.003250535111874342, + 0.006469192914664745, + 0.005993632599711418, + 0.0082775903865695, + 0.006226797588169575, + 0.005269119516015053, + 0.007924878038465977, + 0.007765596266835928, + 0.007152059115469456, + 0.010971905663609505, + 0.01281675137579441, + 0.013715400360524654, + 0.010738956741988659, + 0.009551558643579483, + 0.008336066268384457, + 0.010099520906805992, + 0.012307269498705864, + 0.004250515718013048, + 0.005041488911956549, + 0.00568767124786973, + 0.006318818312138319, + 0.0075411079451441765, + 0.00635613314807415, + 0.007599513046443462, + 0.007735075894743204, + 0.06471861153841019, + 0.04253255948424339, + 0.05408599227666855, + 0.08043299615383148, + 0.03692026808857918, + 0.026157423853874207, + 0.07592006027698517, + 0.037841200828552246, + 0.004717073403298855, + 0.005889443214982748, + 0.006818696856498718, + 0.013867470435798168, + 0.005482385400682688, + 0.007283378858119249, + 0.00962104368954897, + 0.009411236271262169, + 0.0048222774639725685, + 0.0065804277546703815, + 0.010847263969480991, + 0.016301985830068588, + 0.003666891250759363, + 0.006786302663385868, + 0.00668618967756629, + 0.0041670906357467175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.10033652186393738 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro lens is firmly locked on the top of a large bowl of steaming pickled fish. A shriveled and bright red dried chili pepper lies quietly on the golden pickled fish soup and a few tender white fish fillets. Its surface is covered with fine wrinkles, foreshadowing the coming storm. Suddenly, a stream of boiling, clear oil is poured down from outside the frame. At the moment the hot oil contacts the chili peppers with a sharp sizzling sound, the camera captures in slow motion the chili peppers’ skins rapidly shrinking and bursting due to the high temperature, the color becoming deeper and glossier, releasing a rich spicy aroma. At the same time, the camera begins to perform a classic Hitchcock zoom (Dolly Zoom), that is, the camera pulls back while the lens focus simultaneously moves forward, causing the blurred Chinese restaurant environment in the background - including wooden tables and chairs under warm lights and the figures of diners - to be visually distorted and rapidly recede backwards, while the size of the main body of the exploding chili peppers in the foreground strangely remains unchanged, creating a strong sense of visual dizziness and highly dramatic tension. The entire picture is full of shocking sensory impact, and the atmosphere is tense and fascinating.", + "seed": 23812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_0/index_6284_seed_23812_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24941157083958387, + "vlm_logprob": -0.24941157083958387, + "vlm_yes_no_margin": 2.674999988079071, + "vlm_token_credit": -0.24941157083958387, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28820133209228516 + ], + "opposite_logprob_tokens": [ + -1.4132013320922852 + ], + "logprob_reward": -0.28820133209228516, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28820133209228516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22521834075450897, + 0.38409459590911865, + 0.04014812037348747, + 0.0329710952937603, + 0.05604131147265434, + 0.028416765853762627, + 0.058563001453876495, + 0.04293433949351311, + 0.06528574973344803, + 0.06632673740386963 + ], + "credit_map": [ + 0.020069705322384834, + 0.06885188072919846, + 0.013096162118017673, + 0.00731293810531497, + 0.018902458250522614, + 0.07131997495889664, + 0.012057511135935783, + 0.0136076919734478, + 0.022868093103170395, + 0.14266033470630646, + 0.012909319251775742, + 0.012109477072954178, + 0.02321922779083252, + 0.1286945939064026, + 0.01986972615122795, + 0.021763775497674942, + 0.004494699649512768, + 0.0035152356140315533, + 0.004228735342621803, + 0.0037452252581715584, + 0.006247718818485737, + 0.005755222402513027, + 0.0072685773484408855, + 0.004892699886113405, + 0.005877965595573187, + 0.006304444745182991, + 0.002976241521537304, + 0.0027110716328024864, + 0.004656341392546892, + 0.003287467174232006, + 0.003066190518438816, + 0.004091368988156319, + 0.0078213419765234, + 0.004795243963599205, + 0.0042938650585711, + 0.005496999714523554, + 0.004118750803172588, + 0.005973307881504297, + 0.016009598970413208, + 0.007532197516411543, + 0.0037202793173491955, + 0.0034759428817778826, + 0.0036270529963076115, + 0.00326802721247077, + 0.004003097303211689, + 0.0032011058647185564, + 0.0041814642027020454, + 0.0029397939797490835, + 0.005265893880277872, + 0.006290634162724018, + 0.01464073546230793, + 0.005219394341111183, + 0.005834604613482952, + 0.013112113811075687, + 0.003843107260763645, + 0.004356509540230036, + 0.004944134969264269, + 0.004681218881160021, + 0.004916589241474867, + 0.00526281027123332, + 0.007284964434802532, + 0.005606117658317089, + 0.0042311809957027435, + 0.006007314659655094, + 0.006388314533978701, + 0.008756543509662151, + 0.008242116309702396, + 0.010195276699960232, + 0.008430452086031437, + 0.005366257857531309, + 0.009434430859982967, + 0.008472352288663387, + 0.010418926365673542, + 0.007959987968206406, + 0.008099888451397419, + 0.0081506771966815, + 0.005510196555405855, + 0.00412745913490653, + 0.006953225005418062, + 0.015106369741261005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.28820133209228516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013658013194799423 + ], + "opposite_logprob_tokens": [ + -6.626365661621094 + ], + "logprob_reward": -0.0013658013194799423, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013658013194799423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08602529019117355, + 0.26836585998535156, + 0.04765065386891365, + 0.05064242333173752, + 0.080118328332901, + 0.0555100180208683, + 0.09441222995519638, + 0.11273466795682907, + 0.08435633778572083, + 0.12018416821956635 + ], + "credit_map": [ + 0.010093439370393753, + 0.011931084096431732, + 0.007763833738863468, + 0.00638926774263382, + 0.019830424338579178, + 0.011931456625461578, + 0.011063444428145885, + 0.00702234311029315, + 0.01883479207754135, + 0.07919371128082275, + 0.02309950441122055, + 0.013454287312924862, + 0.022946156561374664, + 0.07499716430902481, + 0.018490688875317574, + 0.017349569126963615, + 0.006086558569222689, + 0.006014095153659582, + 0.005530324764549732, + 0.0031954271253198385, + 0.006280855741351843, + 0.007506024092435837, + 0.007481762673705816, + 0.005555604584515095, + 0.005752504337579012, + 0.008409634232521057, + 0.005228008143603802, + 0.0035485895350575447, + 0.006005482282489538, + 0.008296792395412922, + 0.007132512051612139, + 0.006268901284784079, + 0.008096475154161453, + 0.00636456860229373, + 0.005816304124891758, + 0.006489121355116367, + 0.006812776904553175, + 0.015798034146428108, + 0.015784449875354767, + 0.014956606552004814, + 0.007876656018197536, + 0.008530051447451115, + 0.006571784615516663, + 0.004824908450245857, + 0.008726958185434341, + 0.006798400077968836, + 0.006467398256063461, + 0.005713863298296928, + 0.010768402367830276, + 0.012495117262005806, + 0.018479421734809875, + 0.010192891582846642, + 0.009363768622279167, + 0.00818126741796732, + 0.009266904555261135, + 0.01566445641219616, + 0.00797059666365385, + 0.014414210803806782, + 0.012441683560609818, + 0.01866207644343376, + 0.024541161954402924, + 0.011476358398795128, + 0.01358634140342474, + 0.009642243385314941, + 0.01007669884711504, + 0.011187674477696419, + 0.00952883530408144, + 0.01092590019106865, + 0.012020542286336422, + 0.009469385258853436, + 0.011526873335242271, + 0.009620431810617447, + 0.011159910820424557, + 0.013663612306118011, + 0.012100226245820522, + 0.02709020860493183, + 0.012484416365623474, + 0.005827827379107475, + 0.022842831909656525, + 0.015015147626399994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.0013658013194799423 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089124716818333 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.011089124716818333, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089124716818333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08728691935539246, + 0.22150567173957825, + 0.05725305899977684, + 0.05719626694917679, + 0.23553083837032318, + 0.17184293270111084, + 0.052478037774562836, + 0.048288699239492416, + 0.03620901331305504, + 0.03240850940346718 + ], + "credit_map": [ + 0.009654900059103966, + 0.008544824086129665, + 0.009626688435673714, + 0.00783604197204113, + 0.011302647180855274, + 0.018051939085125923, + 0.013678696937859058, + 0.00859118439257145, + 0.012242279946804047, + 0.07991445064544678, + 0.007473453413695097, + 0.007294224575161934, + 0.014072205871343613, + 0.07752939313650131, + 0.010565029457211494, + 0.012414633296430111, + 0.006955390330404043, + 0.006497367285192013, + 0.005888587329536676, + 0.005665039177983999, + 0.008136780001223087, + 0.006955402437597513, + 0.01148901879787445, + 0.005665471777319908, + 0.007595133036375046, + 0.004847892094403505, + 0.0062154680490493774, + 0.004335487727075815, + 0.0063818576745688915, + 0.014109723269939423, + 0.007518716622143984, + 0.006191989406943321, + 0.04352763667702675, + 0.029175328090786934, + 0.015959210693836212, + 0.01270325854420662, + 0.02192116528749466, + 0.04154631495475769, + 0.0256474781781435, + 0.04505045339465141, + 0.017633982002735138, + 0.031347911804914474, + 0.014262109063565731, + 0.012880637310445309, + 0.0365990549325943, + 0.022969001904129982, + 0.015482236631214619, + 0.02066798508167267, + 0.005279902834445238, + 0.007832827977836132, + 0.008816146291792393, + 0.005489783361554146, + 0.008523576892912388, + 0.004681975115090609, + 0.005122737493366003, + 0.006731089670211077, + 0.0047294944524765015, + 0.0053702848963439465, + 0.00522612314671278, + 0.009879994206130505, + 0.006230803206562996, + 0.005480295047163963, + 0.0055173481814563274, + 0.005854358430951834, + 0.002815171144902706, + 0.005827727261930704, + 0.004933286923915148, + 0.005419505760073662, + 0.004839417524635792, + 0.002385338069871068, + 0.004870915785431862, + 0.005117650143802166, + 0.002879201900213957, + 0.003438859712332487, + 0.0036767038982361555, + 0.0075245206244289875, + 0.0034829839132726192, + 0.002788545098155737, + 0.0038404404185712337, + 0.004777253605425358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.011089124716818333 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134947419166565 + ], + "opposite_logprob_tokens": [ + -1.3134946823120117 + ], + "logprob_reward": -0.3134947419166565, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134947419166565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19755390286445618, + 0.09092528373003006, + 0.0456245020031929, + 0.03165401890873909, + 0.2523338198661804, + 0.14491093158721924, + 0.07859374582767487, + 0.036770109087228775, + 0.06492458283901215, + 0.056709107011556625 + ], + "credit_map": [ + 0.015715476125478745, + 0.05468432605266571, + 0.01294989325106144, + 0.010129177942872047, + 0.016757389530539513, + 0.05932501703500748, + 0.01506388932466507, + 0.012928737327456474, + 0.0077032363042235374, + 0.031397562474012375, + 0.005309590138494968, + 0.003364832606166601, + 0.007912572473287582, + 0.02252020500600338, + 0.00703207403421402, + 0.005685212556272745, + 0.0053306701593101025, + 0.00577562814578414, + 0.004786958452314138, + 0.003723836038261652, + 0.006518176291137934, + 0.009140601381659508, + 0.005692839156836271, + 0.004655790515244007, + 0.004764374811202288, + 0.004771871957927942, + 0.0024676050525158644, + 0.002471245126798749, + 0.0037812800146639347, + 0.005755169317126274, + 0.003946322947740555, + 0.0036961492151021957, + 0.01955939456820488, + 0.014470315538346767, + 0.015911636874079704, + 0.021833308041095734, + 0.015601592138409615, + 0.026086581870913506, + 0.07509060949087143, + 0.06378038972616196, + 0.013821710832417011, + 0.02100098319351673, + 0.01358489878475666, + 0.012588039971888065, + 0.02975589968264103, + 0.019736822694540024, + 0.014776275493204594, + 0.0196462981402874, + 0.006822495721280575, + 0.011117943562567234, + 0.009973406791687012, + 0.007362431846559048, + 0.012325454503297806, + 0.01416272297501564, + 0.0066333007998764515, + 0.010195991024374962, + 0.003430949291214347, + 0.004509935155510902, + 0.0034499657340347767, + 0.0040648458525538445, + 0.008767877705395222, + 0.004751686006784439, + 0.004579486791044474, + 0.0032153641805052757, + 0.00696798088029027, + 0.007595007307827473, + 0.0068912687711417675, + 0.008690224960446358, + 0.007087232079356909, + 0.006828548852354288, + 0.008730332367122173, + 0.012133989483118057, + 0.005632296670228243, + 0.005610037129372358, + 0.005942126736044884, + 0.0066115944646298885, + 0.0050127278082072735, + 0.0035054965410381556, + 0.013235758058726788, + 0.011159070767462254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.3134947419166565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329068541526794 + ], + "opposite_logprob_tokens": [ + -0.7579068541526794 + ], + "logprob_reward": -0.6329068541526794, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329068541526794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1607072353363037, + 0.1690797358751297, + 0.11222987622022629, + 0.13867615163326263, + 0.06161743402481079, + 0.10282424837350845, + 0.10127680003643036, + 0.05749429389834404, + 0.031578756868839264, + 0.06451553851366043 + ], + "credit_map": [ + 0.010369396768510342, + 0.03476784750819206, + 0.01340547762811184, + 0.007172879297286272, + 0.01915745623409748, + 0.05120323225855827, + 0.012879190035164356, + 0.011751741170883179, + 0.01032654382288456, + 0.060519564896821976, + 0.006456983275711536, + 0.0062066372483968735, + 0.01142579223960638, + 0.05376141518354416, + 0.010521484538912773, + 0.009861270897090435, + 0.015447383746504784, + 0.010932527482509613, + 0.01130705513060093, + 0.01076506357640028, + 0.016109993681311607, + 0.013477138243615627, + 0.020839441567659378, + 0.013351257890462875, + 0.029240911826491356, + 0.018510205671191216, + 0.015844209119677544, + 0.010409076698124409, + 0.014500096440315247, + 0.016721416264772415, + 0.016519570723176003, + 0.01693064160645008, + 0.017583372071385384, + 0.006449065171182156, + 0.00443313829600811, + 0.0027364264242351055, + 0.0039962236769497395, + 0.004723134450614452, + 0.006009898614138365, + 0.015686169266700745, + 0.008164086379110813, + 0.01997247338294983, + 0.007758636958897114, + 0.005963074043393135, + 0.01921876147389412, + 0.014876674860715866, + 0.011009767651557922, + 0.01586076244711876, + 0.013026600703597069, + 0.01281842403113842, + 0.021228617057204247, + 0.011512227356433868, + 0.010344498790800571, + 0.00819351151585579, + 0.008287275210022926, + 0.015865635126829147, + 0.005838544573634863, + 0.006281606387346983, + 0.006004635244607925, + 0.00894902553409338, + 0.010584934614598751, + 0.007232761010527611, + 0.005792850628495216, + 0.006809927523136139, + 0.0033194010611623526, + 0.003369138343259692, + 0.0042520975694060326, + 0.003974654711782932, + 0.0034434504341334105, + 0.003070145146921277, + 0.004554349929094315, + 0.005595513619482517, + 0.0073752193711698055, + 0.008014691062271595, + 0.008191409520804882, + 0.009501866064965725, + 0.006164973136037588, + 0.004600004758685827, + 0.009085779078304768, + 0.011581587605178356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.6329068541526794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro lens is firmly locked on the top of a large bowl of steaming pickled fish. A shriveled and bright red dried chili pepper lies quietly on the golden pickled fish soup and a few tender white fish fillets. Its surface is covered with fine wrinkles, foreshadowing the coming storm. Suddenly, a stream of boiling, clear oil is poured down from outside the frame. At the moment the hot oil contacts the chili peppers with a sharp sizzling sound, the camera captures in slow motion the chili peppers’ skins rapidly shrinking and bursting due to the high temperature, the color becoming deeper and glossier, releasing a rich spicy aroma. At the same time, the camera begins to perform a classic Hitchcock zoom (Dolly Zoom), that is, the camera pulls back while the lens focus simultaneously moves forward, causing the blurred Chinese restaurant environment in the background - including wooden tables and chairs under warm lights and the figures of diners - to be visually distorted and rapidly recede backwards, while the size of the main body of the exploding chili peppers in the foreground strangely remains unchanged, creating a strong sense of visual dizziness and highly dramatic tension. The entire picture is full of shocking sensory impact, and the atmosphere is tense and fascinating.", + "seed": 23813, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_0/index_6284_seed_23813_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18691815822385252, + "vlm_logprob": -0.18691815822385252, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.18691815822385252, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6999267339706421 + ], + "opposite_logprob_tokens": [ + -0.6999267339706421 + ], + "logprob_reward": -0.6999267339706421, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6999267339706421, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06808555126190186, + 0.07470642030239105, + 0.18311242759227753, + 0.05646214261651039, + 0.40766966342926025, + 0.030595628544688225, + 0.03687627613544464, + 0.02097817324101925, + 0.07714378833770752, + 0.044369954615831375 + ], + "credit_map": [ + 0.009990816935896873, + 0.012295483611524105, + 0.010147571563720703, + 0.007984328083693981, + 0.006278385873883963, + 0.008703120052814484, + 0.005203324370086193, + 0.0074825226329267025, + 0.012578362599015236, + 0.005523652769625187, + 0.016373001039028168, + 0.006888037081807852, + 0.009521259926259518, + 0.010003997944295406, + 0.006115633528679609, + 0.0077024768106639385, + 0.026235518977046013, + 0.01724289171397686, + 0.0283976998180151, + 0.019197294488549232, + 0.023766040802001953, + 0.022293828427791595, + 0.012358731590211391, + 0.03362041711807251, + 0.004362283740192652, + 0.00320932874456048, + 0.00534675782546401, + 0.0068451534025371075, + 0.009477975778281689, + 0.005793984048068523, + 0.004820877220481634, + 0.016605786979198456, + 0.022324344143271446, + 0.011080978438258171, + 0.16749708354473114, + 0.023656684905290604, + 0.01629527285695076, + 0.020195532590150833, + 0.027040686458349228, + 0.11957907676696777, + 0.0021602967754006386, + 0.002478530863299966, + 0.00528189679607749, + 0.004769544582813978, + 0.005289918277412653, + 0.0025909272953867912, + 0.003956154454499483, + 0.004068359266966581, + 0.004855875391513109, + 0.003339282935485244, + 0.005082502029836178, + 0.003711555851623416, + 0.005726149305701256, + 0.0037575785536319017, + 0.005140275694429874, + 0.005263058003038168, + 0.0033527170307934284, + 0.0019229749450460076, + 0.0024852175265550613, + 0.002679205732420087, + 0.0029066584538668394, + 0.0019532833248376846, + 0.0024306192062795162, + 0.003247495973482728, + 0.009391688741743565, + 0.007748040370643139, + 0.01126181986182928, + 0.008016576059162617, + 0.008142855949699879, + 0.008515359833836555, + 0.011641948483884335, + 0.01242550928145647, + 0.004580606706440449, + 0.004134005401283503, + 0.00692098680883646, + 0.004803777672350407, + 0.006061222869902849, + 0.004260441288352013, + 0.007250381633639336, + 0.006358531303703785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.6999267339706421 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022381513845175505 + ], + "opposite_logprob_tokens": [ + -6.1272382736206055 + ], + "logprob_reward": -0.0022381513845175505, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022381513845175505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06524703651666641, + 0.05708543211221695, + 0.08351261168718338, + 0.09713394939899445, + 0.42512184381484985, + 0.056987110525369644, + 0.04851646348834038, + 0.03165966272354126, + 0.08695343881845474, + 0.04778245836496353 + ], + "credit_map": [ + 0.008446676656603813, + 0.010358351282775402, + 0.008437657728791237, + 0.009523622691631317, + 0.0055472818203270435, + 0.006871107965707779, + 0.006452434696257114, + 0.009609902277588844, + 0.007010873407125473, + 0.007737858220934868, + 0.013398438692092896, + 0.00657649477943778, + 0.0054770405404269695, + 0.005427363328635693, + 0.005448955576866865, + 0.006008402910083532, + 0.009901009500026703, + 0.005175382364541292, + 0.013882193714380264, + 0.00774226151406765, + 0.008533096872270107, + 0.026077615097165108, + 0.0048101507127285, + 0.007390893064439297, + 0.006226398516446352, + 0.004863006062805653, + 0.018255194649100304, + 0.010350608266890049, + 0.008586488664150238, + 0.010120547376573086, + 0.006787790916860104, + 0.0319439172744751, + 0.023429684340953827, + 0.01244444027543068, + 0.17326532304286957, + 0.025667572394013405, + 0.018064916133880615, + 0.021124981343746185, + 0.028381887823343277, + 0.12274301052093506, + 0.005513815674930811, + 0.00558669026941061, + 0.006571116391569376, + 0.010218283161520958, + 0.009761343710124493, + 0.004725162405520678, + 0.006046730559319258, + 0.008563966490328312, + 0.009626148268580437, + 0.00405089883133769, + 0.007474726065993309, + 0.003647361183539033, + 0.005400741007179022, + 0.006221877411007881, + 0.004677510820329189, + 0.007417196407914162, + 0.003871362889185548, + 0.003658801317214966, + 0.003361908718943596, + 0.004349798429757357, + 0.004638821352273226, + 0.0028642872348427773, + 0.004239434376358986, + 0.004675248637795448, + 0.012108393013477325, + 0.008784451521933079, + 0.012586270458996296, + 0.009048990905284882, + 0.008645921014249325, + 0.011821331456303596, + 0.013697875663638115, + 0.010260200127959251, + 0.004961855243891478, + 0.005065194331109524, + 0.007599303033202887, + 0.004885717760771513, + 0.007078345399349928, + 0.005469278898090124, + 0.007401429582387209, + 0.00532133225351572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.0022381513845175505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206643581390381 + ], + "opposite_logprob_tokens": [ + -2.8120665550231934 + ], + "logprob_reward": -0.06206643581390381, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206643581390381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04364105686545372, + 0.02591579593718052, + 0.4029044210910797, + 0.09651069343090057, + 0.22079592943191528, + 0.06633709371089935, + 0.01920255459845066, + 0.0412827767431736, + 0.06541305780410767, + 0.017996640875935555 + ], + "credit_map": [ + 0.0065140435472130775, + 0.007101239636540413, + 0.005785400979220867, + 0.005603248719125986, + 0.004757866263389587, + 0.004523843992501497, + 0.0036501153372228146, + 0.005705294664949179, + 0.0025362621527165174, + 0.003442424116656184, + 0.0045504882000386715, + 0.003489141119644046, + 0.003594169393181801, + 0.0022227757144719362, + 0.0022880772594362497, + 0.0037924551870673895, + 0.05199788138270378, + 0.029824893921613693, + 0.08462349325418472, + 0.039967674762010574, + 0.04633306711912155, + 0.05283963307738304, + 0.021732131019234657, + 0.07558560371398926, + 0.00604330375790596, + 0.005164909176528454, + 0.0072371698915958405, + 0.01223840657621622, + 0.023211346939206123, + 0.006955383345484734, + 0.007614675909280777, + 0.02804548479616642, + 0.01224205270409584, + 0.006872514728456736, + 0.08893131464719772, + 0.013726189732551575, + 0.010055888444185257, + 0.010908110998570919, + 0.014954712241888046, + 0.06310512870550156, + 0.006403086706995964, + 0.005030070897191763, + 0.009906706400215626, + 0.009304293431341648, + 0.009688597172498703, + 0.006331217475235462, + 0.006932706106454134, + 0.01274040900170803, + 0.0031954653095453978, + 0.0014419946819543839, + 0.0034115053713321686, + 0.0017618633573874831, + 0.0020210640504956245, + 0.002695279661566019, + 0.0017026293789967895, + 0.002972751623019576, + 0.007370924577116966, + 0.0034932566341012716, + 0.005891819018870592, + 0.004823341500014067, + 0.003750844392925501, + 0.005046206992119551, + 0.004215504042804241, + 0.006690874695777893, + 0.007627693470567465, + 0.005991467274725437, + 0.00913054309785366, + 0.0067165931686758995, + 0.005361064802855253, + 0.009097672998905182, + 0.009350951761007309, + 0.012137065641582012, + 0.0020097107626497746, + 0.0014764101943001151, + 0.0032738451845943928, + 0.002078237710520625, + 0.0021375014912337065, + 0.0018832468194887042, + 0.0030138855800032616, + 0.00212379964068532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.06206643581390381 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033976286649704 + ], + "opposite_logprob_tokens": [ + -2.350339651107788 + ], + "logprob_reward": -0.10033976286649704, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033976286649704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0910433679819107, + 0.07293131202459335, + 0.33597084879875183, + 0.06257448345422745, + 0.22173520922660828, + 0.07754969596862793, + 0.030797256156802177, + 0.018288658931851387, + 0.05176033452153206, + 0.03734888136386871 + ], + "credit_map": [ + 0.01093465369194746, + 0.01414650771766901, + 0.014328476041555405, + 0.011333795264363289, + 0.007905710488557816, + 0.009512323886156082, + 0.007727694232016802, + 0.015154197812080383, + 0.00872308574616909, + 0.009972179308533669, + 0.011362649500370026, + 0.008045175112783909, + 0.01285008154809475, + 0.009695935994386673, + 0.004901621025055647, + 0.0073805805295705795, + 0.037825047969818115, + 0.035744380205869675, + 0.061322711408138275, + 0.04253719747066498, + 0.03825636953115463, + 0.05266532301902771, + 0.030621422454714775, + 0.03699835389852524, + 0.00615134509280324, + 0.005168896168470383, + 0.007207524497061968, + 0.010981663130223751, + 0.006535742431879044, + 0.007002364844083786, + 0.006710562855005264, + 0.012816378846764565, + 0.011897776275873184, + 0.005675026681274176, + 0.09037500619888306, + 0.01322515495121479, + 0.008068684488534927, + 0.011146781966090202, + 0.014455867931246758, + 0.0668908953666687, + 0.007251648232340813, + 0.008419215679168701, + 0.00955041591078043, + 0.0153783168643713, + 0.012302062474191189, + 0.007804858032613993, + 0.0075553711503744125, + 0.009287799708545208, + 0.0025091776624321938, + 0.002934185555204749, + 0.006376252975314856, + 0.0031045451760292053, + 0.004800461698323488, + 0.0027016091626137495, + 0.004776400048285723, + 0.0035946201533079147, + 0.0019189597805961967, + 0.0018291675951331854, + 0.0028887796215713024, + 0.0027295525651425123, + 0.001676732674241066, + 0.0012638282496482134, + 0.002223115647211671, + 0.003758520819246769, + 0.005152617581188679, + 0.004738609306514263, + 0.007041709963232279, + 0.007958099246025085, + 0.005413088481873274, + 0.005367419216781855, + 0.006492333021014929, + 0.009596455842256546, + 0.003778867656365037, + 0.004841020796447992, + 0.004236905369907618, + 0.003964606672525406, + 0.006035586819052696, + 0.004123511724174023, + 0.006124221719801426, + 0.004244160372763872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.10033976286649704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001970708370209 + ], + "opposite_logprob_tokens": [ + -2.6950197219848633 + ], + "logprob_reward": -0.07001970708370209, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001970708370209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08061204850673676, + 0.12334092706441879, + 0.16935350000858307, + 0.17375724017620087, + 0.22645889222621918, + 0.04569530859589577, + 0.10332146286964417, + 0.016516199335455894, + 0.046190932393074036, + 0.014753544703125954 + ], + "credit_map": [ + 0.012481288984417915, + 0.012792549096047878, + 0.010484134778380394, + 0.008237916976213455, + 0.006861304398626089, + 0.014145724475383759, + 0.004926903638988733, + 0.010682220570743084, + 0.019835738465189934, + 0.022232739254832268, + 0.02378193847835064, + 0.013189638033509254, + 0.014984088018536568, + 0.01211731880903244, + 0.007663615047931671, + 0.0095358369871974, + 0.019654959440231323, + 0.015006800182163715, + 0.02901950664818287, + 0.014913536608219147, + 0.017478136345744133, + 0.02051960863173008, + 0.011131450533866882, + 0.04162949323654175, + 0.012018933892250061, + 0.00888968538492918, + 0.025202037766575813, + 0.014327188953757286, + 0.015281911939382553, + 0.02085304819047451, + 0.012177972123026848, + 0.06500644236803055, + 0.012518436647951603, + 0.006787587888538837, + 0.09147419035434723, + 0.013176805339753628, + 0.009986869059503078, + 0.01153476256877184, + 0.015255848877131939, + 0.06572437286376953, + 0.004079999402165413, + 0.003663463518023491, + 0.004920615814626217, + 0.00552182924002409, + 0.004503252450376749, + 0.005484610330313444, + 0.0052511789835989475, + 0.012270352803170681, + 0.02053261362016201, + 0.009422224014997482, + 0.018114224076271057, + 0.0070254011079669, + 0.007645157631486654, + 0.015133189037442207, + 0.008229793980717659, + 0.01721884310245514, + 0.0032135150395333767, + 0.0014064223505556583, + 0.0019960859790444374, + 0.001926940050907433, + 0.0018835739465430379, + 0.0017004278488457203, + 0.0018238569609820843, + 0.002565376926213503, + 0.0046991310082376, + 0.004570762626826763, + 0.007365524303168058, + 0.006375342607498169, + 0.00391449173912406, + 0.005780926439911127, + 0.00524457823485136, + 0.008240170776844025, + 0.0014171515358611941, + 0.0015754272462800145, + 0.0027519986033439636, + 0.0017647044733166695, + 0.001789349946193397, + 0.0015222072834149003, + 0.0022495898883789778, + 0.0016831137472763658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.07001970708370209 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro lens is firmly locked on the top of a large bowl of steaming pickled fish. A shriveled and bright red dried chili pepper lies quietly on the golden pickled fish soup and a few tender white fish fillets. Its surface is covered with fine wrinkles, foreshadowing the coming storm. Suddenly, a stream of boiling, clear oil is poured down from outside the frame. At the moment the hot oil contacts the chili peppers with a sharp sizzling sound, the camera captures in slow motion the chili peppers’ skins rapidly shrinking and bursting due to the high temperature, the color becoming deeper and glossier, releasing a rich spicy aroma. At the same time, the camera begins to perform a classic Hitchcock zoom (Dolly Zoom), that is, the camera pulls back while the lens focus simultaneously moves forward, causing the blurred Chinese restaurant environment in the background - including wooden tables and chairs under warm lights and the figures of diners - to be visually distorted and rapidly recede backwards, while the size of the main body of the exploding chili peppers in the foreground strangely remains unchanged, creating a strong sense of visual dizziness and highly dramatic tension. The entire picture is full of shocking sensory impact, and the atmosphere is tense and fascinating.", + "seed": 23814, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_0/index_6284_seed_23814_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19845053516328334, + "vlm_logprob": -0.19845053516328334, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.19845053516328334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5869424343109131 + ], + "opposite_logprob_tokens": [ + -0.8369424343109131 + ], + "logprob_reward": -0.5869424343109131, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5869424343109131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2004377543926239, + 0.12268088012933731, + 0.06938795745372772, + 0.037846144288778305, + 0.05989482253789902, + 0.09684962034225464, + 0.12082669883966446, + 0.06475615501403809, + 0.12449746578931808, + 0.1028224527835846 + ], + "credit_map": [ + 0.014363089576363564, + 0.014931605197489262, + 0.025950705632567406, + 0.03184434399008751, + 0.028908148407936096, + 0.021739862859249115, + 0.03693816438317299, + 0.025761842727661133, + 0.010458781383931637, + 0.007050447631627321, + 0.014848627150058746, + 0.011796267703175545, + 0.013259238563477993, + 0.0252529364079237, + 0.013976899906992912, + 0.026037678122520447, + 0.004470124375075102, + 0.006385206710547209, + 0.005726587492972612, + 0.006935029290616512, + 0.007596686948090792, + 0.012530399486422539, + 0.01794847659766674, + 0.007795443292707205, + 0.0035641840659081936, + 0.003396906889975071, + 0.005625800229609013, + 0.0058790710754692554, + 0.0033152331598103046, + 0.003924023360013962, + 0.006569789256900549, + 0.005571139045059681, + 0.004019617103040218, + 0.006832374259829521, + 0.005687241442501545, + 0.009265015833079815, + 0.011354642920196056, + 0.008933563716709614, + 0.007074578665196896, + 0.006727788131684065, + 0.008623447269201279, + 0.009014727547764778, + 0.011781290173530579, + 0.0171455517411232, + 0.011622067540884018, + 0.008416330441832542, + 0.015962770208716393, + 0.014283428899943829, + 0.007210160605609417, + 0.017394429072737694, + 0.013926968909800053, + 0.024823719635605812, + 0.015166804194450378, + 0.008813508786261082, + 0.017339415848255157, + 0.01615169458091259, + 0.005227290093898773, + 0.007179863750934601, + 0.007953149266541004, + 0.009743032976984978, + 0.006606581155210733, + 0.005499314051121473, + 0.012439539656043053, + 0.010107382200658321, + 0.009313641116023064, + 0.01757132075726986, + 0.01579098589718342, + 0.020976219326257706, + 0.011895605362951756, + 0.011717920191586018, + 0.017854701727628708, + 0.019377076998353004, + 0.010498643852770329, + 0.007888798601925373, + 0.01654868759214878, + 0.019229060038924217, + 0.011334202252328396, + 0.008171645924448967, + 0.018773656338453293, + 0.010377757251262665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.5869424343109131 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00414346344769001 + ], + "opposite_logprob_tokens": [ + -5.504143238067627 + ], + "logprob_reward": -0.00414346344769001, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00414346344769001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13473637402057648, + 0.09834372997283936, + 0.06426180154085159, + 0.050763070583343506, + 0.0789809301495552, + 0.037863872945308685, + 0.11374278366565704, + 0.050541892647743225, + 0.09484578669071198, + 0.27591976523399353 + ], + "credit_map": [ + 0.00829220935702324, + 0.010897544212639332, + 0.01768685318529606, + 0.016338573768734932, + 0.01806912198662758, + 0.017871107906103134, + 0.023811450228095055, + 0.02176949754357338, + 0.009270353242754936, + 0.009209051728248596, + 0.013309730216860771, + 0.013614547438919544, + 0.011478417553007603, + 0.010001366026699543, + 0.019412711262702942, + 0.012047546915709972, + 0.004343784414231777, + 0.004112987779080868, + 0.005150217097252607, + 0.007727314252406359, + 0.006628257222473621, + 0.008585973642766476, + 0.020683003589510918, + 0.007030257023870945, + 0.004745751153677702, + 0.003716939827427268, + 0.008988354355096817, + 0.005848281551152468, + 0.005333382170647383, + 0.006946869194507599, + 0.007061338052153587, + 0.008122146129608154, + 0.00723354471847415, + 0.00837162509560585, + 0.00848788395524025, + 0.013039398938417435, + 0.010241768322885036, + 0.011807925067842007, + 0.009682857431471348, + 0.010115922428667545, + 0.0034888985101133585, + 0.003375946544110775, + 0.00615968881174922, + 0.006680381018668413, + 0.004543925169855356, + 0.003319639014080167, + 0.005323583260178566, + 0.004971807822585106, + 0.009832426905632019, + 0.015139502473175526, + 0.013083362020552158, + 0.023130398243665695, + 0.01650952361524105, + 0.011183090507984161, + 0.012611371465027332, + 0.012253103777766228, + 0.004403402097523212, + 0.00805569626390934, + 0.005536054261028767, + 0.006976335775107145, + 0.006287689786404371, + 0.004893349949270487, + 0.00752642285078764, + 0.006862934213131666, + 0.007028470281511545, + 0.012103906832635403, + 0.014683088287711143, + 0.013170617632567883, + 0.008725480176508427, + 0.009818147867918015, + 0.013832655735313892, + 0.015483412891626358, + 0.015408269129693508, + 0.01613575592637062, + 0.04307061806321144, + 0.036257293075323105, + 0.033561382442712784, + 0.017560403794050217, + 0.048685457557439804, + 0.06524054706096649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.00414346344769001 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868634417653084 + ], + "opposite_logprob_tokens": [ + -3.0486862659454346 + ], + "logprob_reward": -0.04868634417653084, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868634417653084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15291689336299896, + 0.0819004476070404, + 0.0730816051363945, + 0.02385239489376545, + 0.08460463583469391, + 0.0337383858859539, + 0.08845889568328857, + 0.0697365552186966, + 0.1741008162498474, + 0.21760936081409454 + ], + "credit_map": [ + 0.011102828197181225, + 0.009730582125484943, + 0.024038657546043396, + 0.018525509163737297, + 0.032349299639463425, + 0.015294678509235382, + 0.0190811138600111, + 0.022794216871261597, + 0.007222848013043404, + 0.006022366229444742, + 0.01411512866616249, + 0.007881066761910915, + 0.007953435182571411, + 0.013346327468752861, + 0.010217251256108284, + 0.015142017975449562, + 0.005389682948589325, + 0.010034020990133286, + 0.00801847968250513, + 0.00883155595511198, + 0.008817083202302456, + 0.009382457472383976, + 0.014935106970369816, + 0.007673207204788923, + 0.002512095496058464, + 0.002107845852151513, + 0.004091830924153328, + 0.003439202904701233, + 0.002455818699672818, + 0.0023370080161839724, + 0.003423222340643406, + 0.003485369961708784, + 0.006785038858652115, + 0.0072619118727743626, + 0.012888734228909016, + 0.017643572762608528, + 0.007523282431066036, + 0.006850627716630697, + 0.010385533794760704, + 0.015265925787389278, + 0.002314154990017414, + 0.003182528307661414, + 0.00611551059409976, + 0.005436894949525595, + 0.003095293417572975, + 0.003708754200488329, + 0.005226251669228077, + 0.004658994730561972, + 0.008115116506814957, + 0.011442355811595917, + 0.01081932708621025, + 0.016093598678708076, + 0.012592926621437073, + 0.009135873056948185, + 0.010305535979568958, + 0.009954160079360008, + 0.004966513719409704, + 0.009901488199830055, + 0.01012646034359932, + 0.011034238152205944, + 0.009627954103052616, + 0.0065553756430745125, + 0.008646499365568161, + 0.008878018707036972, + 0.012403076514601707, + 0.020982149988412857, + 0.022168094292283058, + 0.02589796856045723, + 0.015924399718642235, + 0.026907773688435555, + 0.02509443089365959, + 0.024722900241613388, + 0.015590458177030087, + 0.014075114391744137, + 0.05162756145000458, + 0.031153244897723198, + 0.02552451193332672, + 0.017957419157028198, + 0.032231900840997696, + 0.029449138790369034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.04868634417653084 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25214046239852905 + ], + "opposite_logprob_tokens": [ + -1.5021405220031738 + ], + "logprob_reward": -0.25214046239852905, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25214046239852905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.124286949634552, + 0.06904374063014984, + 0.04392915219068527, + 0.024352896958589554, + 0.06983708590269089, + 0.04114199057221413, + 0.05630199983716011, + 0.0469566285610199, + 0.1530405431985855, + 0.3711090087890625 + ], + "credit_map": [ + 0.007563356310129166, + 0.006359267048537731, + 0.014903390780091286, + 0.01187045592814684, + 0.01655357889831066, + 0.02612498216331005, + 0.01233366597443819, + 0.0285782590508461, + 0.005950877908617258, + 0.003919703420251608, + 0.010751516558229923, + 0.005059791728854179, + 0.00655338866636157, + 0.01609615609049797, + 0.006024380214512348, + 0.014687923714518547, + 0.0031857499852776527, + 0.003566770115867257, + 0.006174459122121334, + 0.005022699944674969, + 0.005255242809653282, + 0.006516181863844395, + 0.008715561591088772, + 0.005492486525326967, + 0.0021496708504855633, + 0.003246149979531765, + 0.0035291637759655714, + 0.002749723382294178, + 0.003569337772205472, + 0.0026918607763946056, + 0.003196640405803919, + 0.003220350481569767, + 0.004602351691573858, + 0.007599404081702232, + 0.005691035650670528, + 0.00752206752076745, + 0.016057835891842842, + 0.011446314863860607, + 0.010052804835140705, + 0.00686526857316494, + 0.0038572032935917377, + 0.004184624645859003, + 0.005726184230297804, + 0.007091644685715437, + 0.004956213291734457, + 0.004111023154109716, + 0.005736136343330145, + 0.005478959996253252, + 0.004778193775564432, + 0.006473513785749674, + 0.006718427408486605, + 0.012359085492789745, + 0.006299242377281189, + 0.005737292114645243, + 0.00796223059296608, + 0.005974011495709419, + 0.003264325438067317, + 0.0042873951606452465, + 0.008261003531515598, + 0.005130954552441835, + 0.008007767610251904, + 0.005495922639966011, + 0.006498269736766815, + 0.006010988261550665, + 0.010684213601052761, + 0.020370081067085266, + 0.018676891922950745, + 0.02580513432621956, + 0.013918586075305939, + 0.019207553938031197, + 0.022217756137251854, + 0.022160323336720467, + 0.022241344675421715, + 0.026729056611657143, + 0.057180020958185196, + 0.05815241485834122, + 0.0590345524251461, + 0.02472212165594101, + 0.0836963802576065, + 0.039353128522634506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.25214046239852905 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view focused on a large bowl of steaming pickled fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a shriveled, bright red dried chili pepper visible on top of the golden soup and white fish fillets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stream of boiling clear oil poured down from outside the frame onto the chili pepper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the chili pepper skin shrinking and bursting in slow motion while the background distorts via a Dolly Zoom effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a Chinese restaurant with wooden tables, chairs, warm lights, and diner figures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033997148275375 + ], + "opposite_logprob_tokens": [ + -2.350339889526367 + ], + "logprob_reward": -0.10033997148275375, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033997148275375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07273449003696442, + 0.15208971500396729, + 0.05712784826755524, + 0.036021314561367035, + 0.05514173582196236, + 0.03439857065677643, + 0.03998146206140518, + 0.2372855842113495, + 0.05042007192969322, + 0.26479923725128174 + ], + "credit_map": [ + 0.006402619648724794, + 0.006869377568364143, + 0.01579562947154045, + 0.011132167652249336, + 0.008372393436729908, + 0.006811539176851511, + 0.008393569849431515, + 0.008957195095717907, + 0.014704505912959576, + 0.012160619720816612, + 0.02741473726928234, + 0.01387847401201725, + 0.012949801981449127, + 0.026112819090485573, + 0.014125414192676544, + 0.030743354931473732, + 0.004606815055012703, + 0.00852967705577612, + 0.008010629564523697, + 0.00695284828543663, + 0.006855602841824293, + 0.007713927887380123, + 0.008059692569077015, + 0.006398655939847231, + 0.003715331433340907, + 0.00344039686024189, + 0.00584236579015851, + 0.003891518572345376, + 0.005017661023885012, + 0.0056365844793617725, + 0.0038642575964331627, + 0.004613196477293968, + 0.004914093296974897, + 0.0067153917625546455, + 0.006875153165310621, + 0.005827572196722031, + 0.009058886207640171, + 0.007673302199691534, + 0.007920381613075733, + 0.006156955845654011, + 0.0033995627891272306, + 0.003090575570240617, + 0.004001766908913851, + 0.0043854559771716595, + 0.00438587786629796, + 0.0034573287703096867, + 0.004867563955485821, + 0.006810441613197327, + 0.0035826554521918297, + 0.006454796995967627, + 0.005020988639444113, + 0.008059773594141006, + 0.0036584939807653427, + 0.0034388978965580463, + 0.005373784806579351, + 0.0043920716270804405, + 0.03129145875573158, + 0.03223871439695358, + 0.02617277391254902, + 0.03031049855053425, + 0.030880816280841827, + 0.024671301245689392, + 0.03590459004044533, + 0.025815434753894806, + 0.0035564920399338007, + 0.00594852352514863, + 0.006071237847208977, + 0.005179998930543661, + 0.006603349465876818, + 0.008155487477779388, + 0.006503002252429724, + 0.008401976898312569, + 0.013670329935848713, + 0.009144493378698826, + 0.08212487399578094, + 0.02216103859245777, + 0.0154944509267807, + 0.011427823454141617, + 0.02620910108089447, + 0.0845671221613884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.10033997148275375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_tensor([3099, 3099])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_tensor([3099, 3099])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bb688fc98fa2d89917ea1c5921e5baf46323fe3b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_tensor([3099, 3099])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot surveys a long, narrow, decaying concrete corridor, its sole source of illumination being a bare fluorescent tube light suspended from the low ceiling by a pair of frayed wires. The fixture oscillates gently and unevenly in slow, hypnotic arcs, emitting a cold, sickly greenish-white light that flickers intermittently, casting a harsh, glaring glow accompanied by a low, constant electrical hum. As the light flickers, deep, inky shadows cast by exposed ceiling ducts and imperfections in the grimy, water-stained concrete walls waver and stretch precariously; they run along the length of the corridor in twisted, elongated patterns before violently rebounding and contorting into grotesque, ever-shifting shapes. The rhythmic yet unpredictable dance of high-contrast light and dark within the stark, colorless environment creates a deeply unsettling and tense atmosphere, transforming the desolate corridor into a space of pure psychological terror.", + "seed": 26802, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_3099_seed_26802_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012329087522812188, + "vlm_logprob": -0.012329087522812188, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.012329087522812188, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022320711985230446 + ], + "opposite_logprob_tokens": [ + -4.397320747375488 + ], + "logprob_reward": -0.022320711985230446, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.022320711985230446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3837733268737793, + 0.057473234832286835, + 0.027807079255580902, + 0.044741272926330566, + 0.05075814947485924, + 0.1517392247915268, + 0.011237830854952335, + 0.1935541033744812, + 0.046319033950567245, + 0.03259668126702309 + ], + "credit_map": [ + 0.017521554604172707, + 0.013085674494504929, + 0.10213164240121841, + 0.17280857264995575, + 0.007424441631883383, + 0.007831829600036144, + 0.037153031677007675, + 0.025816593319177628, + 0.0029411378782242537, + 0.0021736174821853638, + 0.006945089902728796, + 0.02406078204512596, + 0.0015177527675405145, + 0.002054102485999465, + 0.005878682713955641, + 0.01190207339823246, + 0.0019547375850379467, + 0.001212201314046979, + 0.001838512485846877, + 0.010111173614859581, + 0.0008495667134411633, + 0.0013013689313083887, + 0.002569348318502307, + 0.007970171980559826, + 0.0029319056775420904, + 0.002794455038383603, + 0.004208012018352747, + 0.008378217928111553, + 0.003427566960453987, + 0.0049741072580218315, + 0.006864591501653194, + 0.011162417940795422, + 0.002993818372488022, + 0.004480088595300913, + 0.004014985635876656, + 0.009515942074358463, + 0.007003351580351591, + 0.0040837968699634075, + 0.005873943213373423, + 0.012792231515049934, + 0.009669813327491283, + 0.013250349089503288, + 0.013799794018268585, + 0.029436757788062096, + 0.018359076231718063, + 0.021016286686062813, + 0.030683143064379692, + 0.015524025075137615, + 0.0008144529419951141, + 0.001004100893624127, + 0.001185195054858923, + 0.002051824936643243, + 0.0013878780882805586, + 0.0012684906832873821, + 0.002406095154583454, + 0.0011197946732863784, + 0.004541646223515272, + 0.010400756262242794, + 0.022312749177217484, + 0.07767906039953232, + 0.00927769485861063, + 0.008553022518754005, + 0.018764669075608253, + 0.04202451556921005, + 0.0024990292731672525, + 0.003800884587690234, + 0.00590856559574604, + 0.008695400319993496, + 0.004507224541157484, + 0.0030270388815551996, + 0.014349489472806454, + 0.0035314064007252455, + 0.0024337598588317633, + 0.003140500281006098, + 0.0038921358063817024, + 0.005652058869600296, + 0.004243473522365093, + 0.003124136943370104, + 0.006521949078887701, + 0.0035886710975319147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.022320711985230446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012564901262521744 + ], + "opposite_logprob_tokens": [ + -4.3875651359558105 + ], + "logprob_reward": -0.012564901262521744, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012564901262521744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.512651801109314, + 0.06919882446527481, + 0.0961206704378128, + 0.03082013688981533, + 0.02743123285472393, + 0.060594841837882996, + 0.059190381318330765, + 0.06351035088300705, + 0.04531621187925339, + 0.035165488719940186 + ], + "credit_map": [ + 0.023119570687413216, + 0.01772061362862587, + 0.13832826912403107, + 0.2331484705209732, + 0.009250066243112087, + 0.008287663571536541, + 0.04932396486401558, + 0.03347324579954147, + 0.004154795315116644, + 0.0018043237505480647, + 0.012844880111515522, + 0.015648864209651947, + 0.002108721062541008, + 0.0030176262371242046, + 0.004678418394178152, + 0.024941202253103256, + 0.0035934075713157654, + 0.0033510446082800627, + 0.00643544178456068, + 0.03555501997470856, + 0.003819597652181983, + 0.0037149647250771523, + 0.007691094186156988, + 0.03196011483669281, + 0.0026556169614195824, + 0.0020710364915430546, + 0.0025310562923550606, + 0.007403665687888861, + 0.002366327214986086, + 0.002105936175212264, + 0.0034507992677390575, + 0.008235704153776169, + 0.0030869622714817524, + 0.0030000025872141123, + 0.003198373131453991, + 0.002966090105473995, + 0.0050438097678124905, + 0.002926182933151722, + 0.0038314496632665396, + 0.003378366120159626, + 0.004099342040717602, + 0.0056404671631753445, + 0.0059077441692352295, + 0.012064507231116295, + 0.009969407692551613, + 0.006419888231903315, + 0.009299885481595993, + 0.007193607743829489, + 0.0013758114073425531, + 0.0035938373766839504, + 0.005320529453456402, + 0.024545839056372643, + 0.0029835698660463095, + 0.002357207704335451, + 0.006106212269514799, + 0.01290738396346569, + 0.0022705746814608574, + 0.0037423772737383842, + 0.009805742651224136, + 0.024037275463342667, + 0.004048701375722885, + 0.0033303468953818083, + 0.007303845603018999, + 0.008971496485173702, + 0.003575436072424054, + 0.004328530747443438, + 0.005610041320323944, + 0.00760791776701808, + 0.004148093983530998, + 0.004353045951575041, + 0.011692097410559654, + 0.0040010553784668446, + 0.004109866451472044, + 0.0036418777890503407, + 0.009103089570999146, + 0.005066745914518833, + 0.0029456086922436953, + 0.003028774866834283, + 0.004189818166196346, + 0.0030797107610851526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012564901262521744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649423528462648 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.004649423528462648, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649423528462648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16377943754196167, + 0.17322489619255066, + 0.04066300392150879, + 0.03690634295344353, + 0.043735187500715256, + 0.021792089566588402, + 0.14645549654960632, + 0.2127646952867508, + 0.040653444826602936, + 0.1200254037976265 + ], + "credit_map": [ + 0.007415838539600372, + 0.0054914988577365875, + 0.039490956813097, + 0.07298437505960464, + 0.004698570817708969, + 0.003835326759144664, + 0.01608979143202305, + 0.013773085549473763, + 0.012987044639885426, + 0.006973321549594402, + 0.020377449691295624, + 0.05018806830048561, + 0.00786530040204525, + 0.007411749102175236, + 0.016494864597916603, + 0.05092711001634598, + 0.0020946988370269537, + 0.0015067741042003036, + 0.003213406540453434, + 0.013753222301602364, + 0.0021894099190831184, + 0.0016935328021645546, + 0.003155942540615797, + 0.013056016527116299, + 0.0032121858093887568, + 0.003520926460623741, + 0.004498814232647419, + 0.003973755519837141, + 0.006804346106946468, + 0.0028123250231146812, + 0.006906569469720125, + 0.005177419167011976, + 0.003814834402874112, + 0.0032212408259510994, + 0.004369177855551243, + 0.006486399099230766, + 0.005950698629021645, + 0.0031481243204325438, + 0.00856495276093483, + 0.008179761469364166, + 0.0014481695834547281, + 0.0027967803180217743, + 0.0017514403443783522, + 0.003278930438682437, + 0.0026900148950517178, + 0.0024046702310442924, + 0.003142713801935315, + 0.004279369954019785, + 0.003934752661734819, + 0.009674257598817348, + 0.011616012081503868, + 0.05550414323806763, + 0.010571733117103577, + 0.006006785202771425, + 0.01746547967195511, + 0.031682342290878296, + 0.008159983903169632, + 0.014332580380141735, + 0.03156822919845581, + 0.07282960414886475, + 0.018337368965148926, + 0.011604543775320053, + 0.026503171771764755, + 0.02942921221256256, + 0.0040402477607131, + 0.0026357006281614304, + 0.012607753276824951, + 0.007123193237930536, + 0.0032773588318377733, + 0.0024711419828236103, + 0.0045376489870250225, + 0.0039603994227945805, + 0.010596775449812412, + 0.009174874983727932, + 0.027954155579209328, + 0.02208660915493965, + 0.013107741251587868, + 0.010011793114244938, + 0.01834001950919628, + 0.00875343568623066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.004649423528462648 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015180503251031041 + ], + "opposite_logprob_tokens": [ + -6.501518249511719 + ], + "logprob_reward": -0.0015180503251031041, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015180503251031041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.451201856136322, + 0.03986566141247749, + 0.032670456916093826, + 0.04017217084765434, + 0.02505761943757534, + 0.04335564374923706, + 0.038252249360084534, + 0.04934326559305191, + 0.13562631607055664, + 0.14445482194423676 + ], + "credit_map": [ + 0.023048782721161842, + 0.017702296376228333, + 0.11361230909824371, + 0.19932128489017487, + 0.009912736713886261, + 0.009684083983302116, + 0.04574689269065857, + 0.03217340633273125, + 0.003067332785576582, + 0.002136130118742585, + 0.003812763374298811, + 0.011861118488013744, + 0.003174219746142626, + 0.003588946769014001, + 0.004887206479907036, + 0.0073379394598305225, + 0.00438989931717515, + 0.0019620240200310946, + 0.006037911865860224, + 0.006877002771943808, + 0.002113496186211705, + 0.0015500342706218362, + 0.004502571187913418, + 0.005237516015768051, + 0.003240650985389948, + 0.0035713459365069866, + 0.004783665295690298, + 0.005448138806968927, + 0.004494418855756521, + 0.004536835011094809, + 0.009531143121421337, + 0.004565968178212643, + 0.0020309556275606155, + 0.0017451029270887375, + 0.0018439542036503553, + 0.0051705921068787575, + 0.0027044799644500017, + 0.0018788325833156705, + 0.0028919975738972425, + 0.006791702006012201, + 0.0020577460527420044, + 0.0029268779326230288, + 0.003528833854943514, + 0.009900279343128204, + 0.004784321412444115, + 0.004466372076421976, + 0.008981731720268726, + 0.00670947739854455, + 0.0022726247552782297, + 0.0033464201260358095, + 0.0034176353365182877, + 0.009228669106960297, + 0.005029422231018543, + 0.003960100002586842, + 0.005690832156687975, + 0.0053065381944179535, + 0.003871953347697854, + 0.0061505879275500774, + 0.0052689798176288605, + 0.005791550036519766, + 0.007806123699992895, + 0.006176994647830725, + 0.006926117930561304, + 0.00735095702111721, + 0.008420072495937347, + 0.011758416891098022, + 0.01690066233277321, + 0.0244836937636137, + 0.013119387440383434, + 0.009315368719398975, + 0.040618401020765305, + 0.011010306887328625, + 0.01994137279689312, + 0.014307924546301365, + 0.01934090070426464, + 0.01921440288424492, + 0.021382354199886322, + 0.012788316234946251, + 0.02405194193124771, + 0.013427591882646084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0015180503251031041 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592350512742996 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592350512742996, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592350512742996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43459439277648926, + 0.0671180933713913, + 0.06236914172768593, + 0.03739874064922333, + 0.03348581865429878, + 0.016841094940900803, + 0.02144794352352619, + 0.08199498057365417, + 0.09992890805006027, + 0.144820898771286 + ], + "credit_map": [ + 0.021736014634370804, + 0.01498967595398426, + 0.11527650058269501, + 0.1952756643295288, + 0.008161879144608974, + 0.008176389150321484, + 0.041613221168518066, + 0.029365072026848793, + 0.005775281228125095, + 0.0019915953744202852, + 0.017991693690419197, + 0.01732938177883625, + 0.0019611665047705173, + 0.0032634835224598646, + 0.007174694910645485, + 0.011630800552666187, + 0.0022287580650299788, + 0.0020873392932116985, + 0.005692507140338421, + 0.019832130521535873, + 0.002009246964007616, + 0.002981817349791527, + 0.004888953175395727, + 0.022648392245173454, + 0.003030787454918027, + 0.0024128970690071583, + 0.004312240518629551, + 0.00604208093136549, + 0.0036542362067848444, + 0.002927296096459031, + 0.007302622310817242, + 0.007716580294072628, + 0.002665753709152341, + 0.002308955881744623, + 0.002834344282746315, + 0.006660486105829477, + 0.0027799627278000116, + 0.002223265590146184, + 0.005513042211532593, + 0.00850000698119402, + 0.0010248005855828524, + 0.0011533027281984687, + 0.0013314825482666492, + 0.003959928639233112, + 0.0013038392644375563, + 0.0012058847350999713, + 0.003281638491898775, + 0.0035802177153527737, + 0.00112869031727314, + 0.0017382389632984996, + 0.0018161603948101401, + 0.0045080180279910564, + 0.002812377642840147, + 0.0015787392621859908, + 0.0038247578777372837, + 0.004040959291160107, + 0.003540165489539504, + 0.005759959574788809, + 0.012721597217023373, + 0.02986185811460018, + 0.005559749435633421, + 0.00444935355335474, + 0.009452039375901222, + 0.010650256648659706, + 0.008526009507477283, + 0.005258665885776281, + 0.031433213502168655, + 0.016774829477071762, + 0.006601573899388313, + 0.004760662093758583, + 0.01593581773340702, + 0.010638141073286533, + 0.01750725694000721, + 0.013791568577289581, + 0.01770690083503723, + 0.022680405527353287, + 0.01705850660800934, + 0.011941040866076946, + 0.027656208723783493, + 0.016479022800922394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.020592350512742996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot surveys a long, narrow, decaying concrete corridor, its sole source of illumination being a bare fluorescent tube light suspended from the low ceiling by a pair of frayed wires. The fixture oscillates gently and unevenly in slow, hypnotic arcs, emitting a cold, sickly greenish-white light that flickers intermittently, casting a harsh, glaring glow accompanied by a low, constant electrical hum. As the light flickers, deep, inky shadows cast by exposed ceiling ducts and imperfections in the grimy, water-stained concrete walls waver and stretch precariously; they run along the length of the corridor in twisted, elongated patterns before violently rebounding and contorting into grotesque, ever-shifting shapes. The rhythmic yet unpredictable dance of high-contrast light and dark within the stark, colorless environment creates a deeply unsettling and tense atmosphere, transforming the desolate corridor into a space of pure psychological terror.", + "seed": 26803, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_3099_seed_26803_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013944055046886205, + "vlm_logprob": -0.013944055046886205, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.013944055046886205, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018847426399588585 + ], + "opposite_logprob_tokens": [ + -4.268847465515137 + ], + "logprob_reward": -0.018847426399588585, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018847426399588585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09642556309700012, + 0.10720455646514893, + 0.029977669939398766, + 0.13675324618816376, + 0.18291111290454865, + 0.3007958233356476, + 0.011278640478849411, + 0.03606753796339035, + 0.016051016747951508, + 0.08253490924835205 + ], + "credit_map": [ + 0.0068090567365288734, + 0.012720870785415173, + 0.022326156497001648, + 0.008762212470173836, + 0.0055512934923172, + 0.008114628493785858, + 0.010279187932610512, + 0.021862143650650978, + 0.00874963030219078, + 0.022713718935847282, + 0.01820526458323002, + 0.01011434942483902, + 0.009488939307630062, + 0.0127343088388443, + 0.008831200189888477, + 0.01636713370680809, + 0.002341312123462558, + 0.0025170850567519665, + 0.00805302057415247, + 0.0037300516851246357, + 0.0024023724254220724, + 0.0019698278047144413, + 0.0017110930057242513, + 0.007252904586493969, + 0.012466991320252419, + 0.013047768734395504, + 0.038230545818805695, + 0.010428251698613167, + 0.011569448746740818, + 0.013475914485752583, + 0.006865141913294792, + 0.03066917695105076, + 0.01547548733651638, + 0.007457979954779148, + 0.06514706462621689, + 0.011763661168515682, + 0.0037258623633533716, + 0.010787430219352245, + 0.00624304311349988, + 0.06231057271361351, + 0.02593608759343624, + 0.051854457706213, + 0.05332354083657265, + 0.032083310186862946, + 0.036545101553201675, + 0.04095744714140892, + 0.020241063088178635, + 0.03985476493835449, + 0.001047612284310162, + 0.0007716713007539511, + 0.003204141277819872, + 0.0006621659849770367, + 0.0007554978365078568, + 0.0009620923083275557, + 0.0007073680753819644, + 0.0031680897809565067, + 0.003049139864742756, + 0.0043674674816429615, + 0.009235160425305367, + 0.0026820984203368425, + 0.0021498939022421837, + 0.0032656679395586252, + 0.0028166226111352444, + 0.00850148405879736, + 0.0013870043912902474, + 0.001269654487259686, + 0.004038688726723194, + 0.002124790335074067, + 0.0008523239521309733, + 0.0019205319695174694, + 0.0009753706981427968, + 0.0034826495684683323, + 0.007881931960582733, + 0.002484539756551385, + 0.032594215124845505, + 0.003091616788879037, + 0.001111472025513649, + 0.004024466034024954, + 0.0016359485452994704, + 0.029710713773965836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.018847426399588585 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219275675714016 + ], + "opposite_logprob_tokens": [ + -4.264219284057617 + ], + "logprob_reward": -0.014219275675714016, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219275675714016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1269606053829193, + 0.26326316595077515, + 0.06249498948454857, + 0.058919526636600494, + 0.13163529336452484, + 0.08783061802387238, + 0.04867009073495865, + 0.04163182154297829, + 0.05824456736445427, + 0.12034930288791656 + ], + "credit_map": [ + 0.01155892014503479, + 0.011257488280534744, + 0.039241500198841095, + 0.006832269486039877, + 0.006806827150285244, + 0.009574778378009796, + 0.005457850638777018, + 0.03623097389936447, + 0.02266687899827957, + 0.020193764939904213, + 0.09079224616289139, + 0.010497857816517353, + 0.007608579006046057, + 0.01638338714838028, + 0.007822536863386631, + 0.08729793131351471, + 0.005174870602786541, + 0.0066357264295220375, + 0.013565020635724068, + 0.011026321910321712, + 0.005458728410303593, + 0.004936976358294487, + 0.004113962408155203, + 0.011583385989069939, + 0.00523710111156106, + 0.004832756239920855, + 0.016235478222370148, + 0.005518399644643068, + 0.004849653225392103, + 0.004799336660653353, + 0.0041478597559034824, + 0.013298941776156425, + 0.01091082114726305, + 0.006152409594506025, + 0.04544384777545929, + 0.008104568347334862, + 0.0034146306570619345, + 0.008078428916633129, + 0.005754347424954176, + 0.043776243925094604, + 0.007549224887043238, + 0.016559768468141556, + 0.013693212531507015, + 0.011001022532582283, + 0.01075827982276678, + 0.011744099669158459, + 0.006406223867088556, + 0.010118787176907063, + 0.004769586026668549, + 0.0037269755266606808, + 0.014876969158649445, + 0.0041876910254359245, + 0.0016308381455019116, + 0.004390768241137266, + 0.0025650239549577236, + 0.012522239238023758, + 0.0038181955460458994, + 0.003416277002543211, + 0.01152640487998724, + 0.0031920981127768755, + 0.0039802296087145805, + 0.0032360090408474207, + 0.002701119054108858, + 0.009761487133800983, + 0.004209262318909168, + 0.01068403571844101, + 0.011260386556386948, + 0.008108533918857574, + 0.0046804542653262615, + 0.005644313059747219, + 0.004044128581881523, + 0.009613452479243279, + 0.011262773536145687, + 0.0034107286483049393, + 0.04795432463288307, + 0.004253057297319174, + 0.0017506618751212955, + 0.005770484916865826, + 0.0018765029963105917, + 0.044070757925510406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.014219275675714016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644202534109354 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.004644202534109354, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644202534109354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14811167120933533, + 0.14175105094909668, + 0.08946020156145096, + 0.06000867486000061, + 0.06467272341251373, + 0.05947258323431015, + 0.04738467559218407, + 0.05339832603931427, + 0.046261321753263474, + 0.28947877883911133 + ], + "credit_map": [ + 0.013105378486216068, + 0.011214039288461208, + 0.04319160059094429, + 0.010065928101539612, + 0.009584501385688782, + 0.01079218927770853, + 0.009500069543719292, + 0.04065795987844467, + 0.012262469157576561, + 0.0069921561516821384, + 0.05086981877684593, + 0.005196071695536375, + 0.004165866412222385, + 0.009157822467386723, + 0.0035091349855065346, + 0.049597710371017456, + 0.007734628859907389, + 0.009968557395040989, + 0.012660672888159752, + 0.00924323033541441, + 0.01753273420035839, + 0.011548811569809914, + 0.008815595880150795, + 0.011955964379012585, + 0.0053754947148263454, + 0.010559980757534504, + 0.00730266235768795, + 0.009820375591516495, + 0.007382860407233238, + 0.007068200036883354, + 0.006284545175731182, + 0.0062145511619746685, + 0.004558913875371218, + 0.00392314326018095, + 0.019167808815836906, + 0.005000453442335129, + 0.0038404902443289757, + 0.004148886539041996, + 0.004567401483654976, + 0.019465621560811996, + 0.005119639914482832, + 0.01024238858371973, + 0.0087281484156847, + 0.008397125639021397, + 0.007255136966705322, + 0.006295579019933939, + 0.006757816299796104, + 0.006676743272691965, + 0.0043823374435305595, + 0.0026950258761644363, + 0.01563934236764908, + 0.004298673011362553, + 0.0014364812523126602, + 0.0024891230277717113, + 0.002345883287489414, + 0.014097806066274643, + 0.005262240767478943, + 0.0051887910813093185, + 0.014318644069135189, + 0.0036528287455439568, + 0.003749527968466282, + 0.005084275733679533, + 0.004144963342696428, + 0.011997050605714321, + 0.00404537795111537, + 0.004188841208815575, + 0.012795298360288143, + 0.002346560824662447, + 0.004530536942183971, + 0.003443170106038451, + 0.0031811015214771032, + 0.011730431579053402, + 0.026480525732040405, + 0.01030438207089901, + 0.10958917438983917, + 0.010799083858728409, + 0.010427982546389103, + 0.014564060606062412, + 0.006283697672188282, + 0.10102985799312592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.004644202534109354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022031813859939575 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.0022031813859939575, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022031813859939575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16243264079093933, + 0.1875344067811966, + 0.03873112052679062, + 0.06534583121538162, + 0.07268918305635452, + 0.16107353568077087, + 0.04105226695537567, + 0.14623136818408966, + 0.07094849646091461, + 0.053961124271154404 + ], + "credit_map": [ + 0.011543449945747852, + 0.03120623715221882, + 0.024590007960796356, + 0.02157731167972088, + 0.014226756989955902, + 0.021584514528512955, + 0.0186209324747324, + 0.019083421677350998, + 0.014005222357809544, + 0.04073016718029976, + 0.03520650416612625, + 0.01577124372124672, + 0.013989926315844059, + 0.021798966452479362, + 0.012888253666460514, + 0.033144112676382065, + 0.003464106936007738, + 0.005557708907872438, + 0.004067073576152325, + 0.006266339682042599, + 0.004774635192006826, + 0.005808194167912006, + 0.004418442957103252, + 0.004374618176370859, + 0.005532915238291025, + 0.010540781542658806, + 0.008020865730941296, + 0.013354934751987457, + 0.006211480591446161, + 0.00842988956719637, + 0.006636178120970726, + 0.006618782877922058, + 0.005421278532594442, + 0.004077705554664135, + 0.022878944873809814, + 0.006330479867756367, + 0.003056684974581003, + 0.004303098656237125, + 0.0037073271814733744, + 0.02291366085410118, + 0.013706695288419724, + 0.02764343097805977, + 0.028751814737915993, + 0.01881319284439087, + 0.016605746001005173, + 0.021608632057905197, + 0.011007142253220081, + 0.022936886176466942, + 0.003582131350412965, + 0.002148574450984597, + 0.01270974799990654, + 0.0038021779619157314, + 0.0019415352726355195, + 0.0029191365465521812, + 0.002022922970354557, + 0.01192604098469019, + 0.013209483586251736, + 0.011317144148051739, + 0.041375841945409775, + 0.010089204646646976, + 0.008586762472987175, + 0.013644433580338955, + 0.009646493941545486, + 0.03836200758814812, + 0.005938041023910046, + 0.018171001225709915, + 0.0084321154281497, + 0.009552695788443089, + 0.00787699967622757, + 0.007366891019046307, + 0.007382696028798819, + 0.006228057201951742, + 0.005100923124700785, + 0.0014024025294929743, + 0.02056664600968361, + 0.0024359163362532854, + 0.0011924522696062922, + 0.0032150072511285543, + 0.0011081192642450333, + 0.018939655274152756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0022031813859939575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806189239025116 + ], + "opposite_logprob_tokens": [ + -3.529806137084961 + ], + "logprob_reward": -0.029806189239025116, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806189239025116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17497198283672333, + 0.23894661664962769, + 0.025210576131939888, + 0.04006239026784897, + 0.11599911004304886, + 0.09163834154605865, + 0.03850466385483742, + 0.10984507203102112, + 0.1281912922859192, + 0.03662993013858795 + ], + "credit_map": [ + 0.012466894462704659, + 0.024873526766896248, + 0.034658145159482956, + 0.02294263429939747, + 0.01444977056235075, + 0.0161035917699337, + 0.018868276849389076, + 0.030609125271439552, + 0.022926300764083862, + 0.06095319613814354, + 0.03046431764960289, + 0.02440831996500492, + 0.02233058586716652, + 0.03209766373038292, + 0.02070217952132225, + 0.02506406418979168, + 0.001720155356451869, + 0.003035064088180661, + 0.004791948013007641, + 0.0027802670374512672, + 0.0023825529497116804, + 0.002992934314534068, + 0.002310248091816902, + 0.005197404883801937, + 0.003252738621085882, + 0.0034191941376775503, + 0.009914546273648739, + 0.004334700759500265, + 0.0037462525069713593, + 0.0036054844968020916, + 0.0036270322743803263, + 0.008162440732121468, + 0.00928228348493576, + 0.004044310189783573, + 0.04116705432534218, + 0.007152306381613016, + 0.00284535507671535, + 0.006635427940636873, + 0.004348521586507559, + 0.04052385687828064, + 0.007773743476718664, + 0.011238588020205498, + 0.021989665925502777, + 0.008318631909787655, + 0.007727880030870438, + 0.010312709026038647, + 0.005565354600548744, + 0.018711764365434647, + 0.003503164742141962, + 0.002587141003459692, + 0.012164892628788948, + 0.003557938849553466, + 0.0018116189166903496, + 0.0021538645960390568, + 0.0020122977439314127, + 0.010713741183280945, + 0.010544787161052227, + 0.0046348413452506065, + 0.037200115621089935, + 0.007153212558478117, + 0.003472709096968174, + 0.007151070050895214, + 0.005280102137476206, + 0.03440823778510094, + 0.010978183709084988, + 0.00936854351311922, + 0.03960675001144409, + 0.01028131041675806, + 0.004517143126577139, + 0.010147307068109512, + 0.006811268161982298, + 0.03648078814148903, + 0.0028064411599189043, + 0.0027916489634662867, + 0.011549766175448895, + 0.0028509716503322124, + 0.0017182492883875966, + 0.002474633976817131, + 0.0014580875867977738, + 0.01098012924194336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.029806189239025116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot surveys a long, narrow, decaying concrete corridor, its sole source of illumination being a bare fluorescent tube light suspended from the low ceiling by a pair of frayed wires. The fixture oscillates gently and unevenly in slow, hypnotic arcs, emitting a cold, sickly greenish-white light that flickers intermittently, casting a harsh, glaring glow accompanied by a low, constant electrical hum. As the light flickers, deep, inky shadows cast by exposed ceiling ducts and imperfections in the grimy, water-stained concrete walls waver and stretch precariously; they run along the length of the corridor in twisted, elongated patterns before violently rebounding and contorting into grotesque, ever-shifting shapes. The rhythmic yet unpredictable dance of high-contrast light and dark within the stark, colorless environment creates a deeply unsettling and tense atmosphere, transforming the desolate corridor into a space of pure psychological terror.", + "seed": 26804, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_3099_seed_26804_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013432538334745914, + "vlm_logprob": -0.013432538334745914, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.013432538334745914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02016659826040268 + ], + "opposite_logprob_tokens": [ + -4.895166397094727 + ], + "logprob_reward": -0.02016659826040268, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.02016659826040268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02310086041688919, + 0.02832047827541828, + 0.08986702561378479, + 0.036387305706739426, + 0.12412626296281815, + 0.011938911862671375, + 0.6316324472427368, + 0.015250999480485916, + 0.029397226870059967, + 0.009978484362363815 + ], + "credit_map": [ + 0.0027685053646564484, + 0.002847821218892932, + 0.002974460367113352, + 0.0028519381303340197, + 0.003692633705213666, + 0.0020742809865623713, + 0.0034062075428664684, + 0.002485014032572508, + 0.0021856462117284536, + 0.005451612640172243, + 0.006361008156090975, + 0.003772963769733906, + 0.0015548371011391282, + 0.002024684101343155, + 0.002037754049524665, + 0.004931972362101078, + 0.008776242844760418, + 0.002901641186326742, + 0.03714035078883171, + 0.006413524504750967, + 0.0018712219316512346, + 0.0039810845628380775, + 0.0022394550032913685, + 0.026543499901890755, + 0.0032573286443948746, + 0.003340393304824829, + 0.004355994984507561, + 0.007341481279581785, + 0.002757500158622861, + 0.006334993056952953, + 0.004931229632347822, + 0.004068384878337383, + 0.0037761034909635782, + 0.0029002567753195763, + 0.006869758479297161, + 0.0028118786867707968, + 0.0029505230486392975, + 0.0529714860022068, + 0.0028407745994627476, + 0.04900548234581947, + 0.0017357709584757686, + 0.0012613580329343677, + 0.0016429920215159655, + 0.0012259942013770342, + 0.0018394585931673646, + 0.0018584791105240583, + 0.0012961758766323328, + 0.0010786834172904491, + 0.0654543861746788, + 0.039547886699438095, + 0.24630463123321533, + 0.026996877044439316, + 0.012167216278612614, + 0.18221043050289154, + 0.01388703752309084, + 0.04506399482488632, + 0.0013300772989168763, + 0.0009979527676478028, + 0.0028696057852357626, + 0.0017555105732753873, + 0.001742613734677434, + 0.0026651921216398478, + 0.0013648206368088722, + 0.0025252255145460367, + 0.002743928460404277, + 0.001302580814808607, + 0.010717390105128288, + 0.002047243295237422, + 0.001770671340636909, + 0.005865008104592562, + 0.0018140344182029366, + 0.003136369865387678, + 0.0010021206689998507, + 0.000577583909034729, + 0.001668018870986998, + 0.0010278359986841679, + 0.0009777328232303262, + 0.002324209315702319, + 0.0011341252829879522, + 0.0012668576091527939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.02016659826040268 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822289079427719 + ], + "opposite_logprob_tokens": [ + -4.018222808837891 + ], + "logprob_reward": -0.01822289079427719, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822289079427719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040773700922727585, + 0.059202421456575394, + 0.19355285167694092, + 0.03517637774348259, + 0.2446300983428955, + 0.01471660751849413, + 0.23138687014579773, + 0.043169766664505005, + 0.08655218034982681, + 0.0508391298353672 + ], + "credit_map": [ + 0.0037078673485666513, + 0.004693009424954653, + 0.009595348499715328, + 0.003968041390180588, + 0.0034708750899881124, + 0.0034747018944472075, + 0.0041625057347118855, + 0.007701354566961527, + 0.003901712130755186, + 0.009147435426712036, + 0.012954972684383392, + 0.008564452640712261, + 0.004402318969368935, + 0.004438588861376047, + 0.005234805401414633, + 0.010558134876191616, + 0.017398247495293617, + 0.007546317297965288, + 0.07729831337928772, + 0.015105919912457466, + 0.004209619481116533, + 0.00922117568552494, + 0.005593943875283003, + 0.05717930942773819, + 0.0031039882451295853, + 0.0036182079929858446, + 0.004139486700296402, + 0.007031049579381943, + 0.00329535361379385, + 0.004042407497763634, + 0.005871197674423456, + 0.004074686672538519, + 0.007389926817268133, + 0.0050718821585178375, + 0.013646991923451424, + 0.00489113200455904, + 0.005275932140648365, + 0.10521338880062103, + 0.005640411749482155, + 0.09750042110681534, + 0.001922552939504385, + 0.0013655817601829767, + 0.001711467164568603, + 0.0014377287589013577, + 0.0020046576391905546, + 0.002628293354064226, + 0.0015217062318697572, + 0.002124619670212269, + 0.034285660833120346, + 0.02466011419892311, + 0.07418810576200485, + 0.010365142486989498, + 0.005455065984278917, + 0.02055974304676056, + 0.0076929847709834576, + 0.05418005958199501, + 0.0035553984344005585, + 0.0019749710336327553, + 0.014248592779040337, + 0.0030064727179706097, + 0.0020941600669175386, + 0.005484346766024828, + 0.00230897543951869, + 0.010496852919459343, + 0.0079492824152112, + 0.0022923105861991644, + 0.035003408789634705, + 0.003290591761469841, + 0.004033113829791546, + 0.02572156861424446, + 0.002630776958540082, + 0.005631122272461653, + 0.00310091907158494, + 0.0026190956123173237, + 0.00708924001082778, + 0.0038366876542568207, + 0.003552780020982027, + 0.014088848605751991, + 0.008195034228265285, + 0.008356527425348759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.01822289079427719 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004652389790862799 + ], + "opposite_logprob_tokens": [ + -5.379652500152588 + ], + "logprob_reward": -0.004652389790862799, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004652389790862799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06290913373231888, + 0.08756665140390396, + 0.21025094389915466, + 0.011122332885861397, + 0.15304569900035858, + 0.025189200416207314, + 0.0595499724149704, + 0.07794389873743057, + 0.2501925230026245, + 0.06222965940833092 + ], + "credit_map": [ + 0.007492281962186098, + 0.008611387573182583, + 0.009934243746101856, + 0.00687452731654048, + 0.005845495965331793, + 0.006324118003249168, + 0.010200505144894123, + 0.007626563776284456, + 0.009559639729559422, + 0.011147131212055683, + 0.013170394115149975, + 0.010301964357495308, + 0.009697074070572853, + 0.013742130249738693, + 0.008394688367843628, + 0.011553612537682056, + 0.018944667652249336, + 0.012459424324333668, + 0.08203138411045074, + 0.016345268115401268, + 0.004366685636341572, + 0.008255014196038246, + 0.00721452571451664, + 0.06063397228717804, + 0.0012718960642814636, + 0.0018756582867354155, + 0.00120711256749928, + 0.0018395009683445096, + 0.0008596953121013939, + 0.0016109864227473736, + 0.0011106261517852545, + 0.0013468553079292178, + 0.003866461804136634, + 0.003045689547434449, + 0.008662346750497818, + 0.002864688169211149, + 0.0030983316246420145, + 0.06649765372276306, + 0.003610792104154825, + 0.06139972433447838, + 0.0024388639722019434, + 0.0021161064505577087, + 0.0038107901345938444, + 0.001887248596176505, + 0.0020957975648343563, + 0.007646072190254927, + 0.0014684440102428198, + 0.0037258747033774853, + 0.005041892174631357, + 0.005019838456064463, + 0.006526455283164978, + 0.00201802933588624, + 0.004350528120994568, + 0.0210469588637352, + 0.0024820887483656406, + 0.013064175844192505, + 0.00646641431376338, + 0.0036766170524060726, + 0.02644340507686138, + 0.005832050461322069, + 0.002959001576527953, + 0.005116797983646393, + 0.003930992446839809, + 0.02351861633360386, + 0.023978980258107185, + 0.006365923676639795, + 0.10295610129833221, + 0.009266271255910397, + 0.009915949776768684, + 0.07333049923181534, + 0.006929817143827677, + 0.01744896173477173, + 0.004243331961333752, + 0.002985812956467271, + 0.005741561762988567, + 0.00478589441627264, + 0.0037539785262197256, + 0.01696799136698246, + 0.0076240478083491325, + 0.016127033159136772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.004652389790862799 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008192281820811331 + ], + "opposite_logprob_tokens": [ + -7.125819206237793 + ], + "logprob_reward": -0.0008192281820811331, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008192281820811331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03753097727894783, + 0.021243389695882797, + 0.2721659541130066, + 0.017910504713654518, + 0.15215714275836945, + 0.016668152064085007, + 0.12422403693199158, + 0.09058498591184616, + 0.1731094866991043, + 0.09440542012453079 + ], + "credit_map": [ + 0.00514257000759244, + 0.004477430135011673, + 0.00704393582418561, + 0.0041091227903962135, + 0.0036043033469468355, + 0.0025511300191283226, + 0.0052628894336521626, + 0.00533959548920393, + 0.0015649943379685283, + 0.0022980691865086555, + 0.004982187412679195, + 0.0029922490939497948, + 0.0019507159013301134, + 0.00198656739667058, + 0.0016385763883590698, + 0.0038300268352031708, + 0.026208534836769104, + 0.009988962672650814, + 0.11246540397405624, + 0.021215301007032394, + 0.004835925530642271, + 0.010698546655476093, + 0.00542493537068367, + 0.08132831007242203, + 0.0014796936884522438, + 0.0028980232309550047, + 0.0020797932520508766, + 0.0026626940816640854, + 0.00214573135599494, + 0.002918155863881111, + 0.0016914448933675885, + 0.0020349666010588408, + 0.003988835960626602, + 0.0028837278950959444, + 0.00806237943470478, + 0.0030733849853277206, + 0.002709396416321397, + 0.06618838757276535, + 0.0038278435822576284, + 0.0614231638610363, + 0.002916917437687516, + 0.002024249406531453, + 0.00276187970302999, + 0.0015237414045259356, + 0.00140151078812778, + 0.003174762474372983, + 0.001484240172430873, + 0.001380849163979292, + 0.013134808279573917, + 0.007356335874646902, + 0.047538578510284424, + 0.005215992219746113, + 0.002677819924429059, + 0.0377982072532177, + 0.0026730394456535578, + 0.007829239591956139, + 0.007661445997655392, + 0.004015457350760698, + 0.03209363669157028, + 0.005562990438193083, + 0.0028743629809468985, + 0.0067107719369232655, + 0.003916545771062374, + 0.027749763801693916, + 0.015551907010376453, + 0.00449834018945694, + 0.06868097186088562, + 0.006602322682738304, + 0.008314471691846848, + 0.054778311401605606, + 0.005500722676515579, + 0.009182408452033997, + 0.004257295746356249, + 0.0019755668472498655, + 0.008025023154914379, + 0.0035896149929612875, + 0.0035066334530711174, + 0.03761526569724083, + 0.005576015915721655, + 0.02985999919474125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.0008192281820811331 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301584646105766 + ], + "opposite_logprob_tokens": [ + -3.773301601409912 + ], + "logprob_reward": -0.023301584646105766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301584646105766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03240710869431496, + 0.03736912086606026, + 0.4366619288921356, + 0.012996060773730278, + 0.056335292756557465, + 0.007727429736405611, + 0.23842468857765198, + 0.032651882618665695, + 0.124427929520607, + 0.020998520776629448 + ], + "credit_map": [ + 0.00307465810328722, + 0.005467762239277363, + 0.00738114956766367, + 0.003491029841825366, + 0.0018777652876451612, + 0.0017354871379211545, + 0.0031661130487918854, + 0.0062131439335644245, + 0.003852623049169779, + 0.002823533723130822, + 0.009209712035953999, + 0.0038356406148523092, + 0.0036494294181466103, + 0.004352952353656292, + 0.0032602199353277683, + 0.006385010201483965, + 0.041312191635370255, + 0.023229707032442093, + 0.17543925344944, + 0.034192994236946106, + 0.008581654168665409, + 0.0181819386780262, + 0.010372363030910492, + 0.12535180151462555, + 0.00146260776091367, + 0.0027513066306710243, + 0.001816910458728671, + 0.0018466007895767689, + 0.0010840269969776273, + 0.001312214182689786, + 0.000901696621440351, + 0.0018206977983936667, + 0.004576513543725014, + 0.00361003540456295, + 0.004061284940689802, + 0.0034455079585313797, + 0.0037996950559318066, + 0.016976922750473022, + 0.004427918698638678, + 0.01543741300702095, + 0.0010495905298739672, + 0.0007973985630087554, + 0.0009563093190081418, + 0.000949651759583503, + 0.0010543727548792958, + 0.001237194170244038, + 0.0009073459659703076, + 0.0007755669066682458, + 0.02606857195496559, + 0.016424113884568214, + 0.0929836630821228, + 0.009875872172415257, + 0.004315838683396578, + 0.06158590316772461, + 0.00527737895026803, + 0.021893354132771492, + 0.002883219625800848, + 0.0013847935479134321, + 0.011075036600232124, + 0.0018803960410878062, + 0.0013360861921682954, + 0.0035474407486617565, + 0.0015824651345610619, + 0.00896244402974844, + 0.01149259228259325, + 0.00290556438267231, + 0.051135022193193436, + 0.004606106784194708, + 0.004718177020549774, + 0.039206601679325104, + 0.0029268001671880484, + 0.007437068969011307, + 0.0014408327406272292, + 0.001202364219352603, + 0.0021091988310217857, + 0.0019648168236017227, + 0.0012772154295817018, + 0.00523335998877883, + 0.0023235452827066183, + 0.0054471855983138084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.023301584646105766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot surveys a long, narrow, decaying concrete corridor, its sole source of illumination being a bare fluorescent tube light suspended from the low ceiling by a pair of frayed wires. The fixture oscillates gently and unevenly in slow, hypnotic arcs, emitting a cold, sickly greenish-white light that flickers intermittently, casting a harsh, glaring glow accompanied by a low, constant electrical hum. As the light flickers, deep, inky shadows cast by exposed ceiling ducts and imperfections in the grimy, water-stained concrete walls waver and stretch precariously; they run along the length of the corridor in twisted, elongated patterns before violently rebounding and contorting into grotesque, ever-shifting shapes. The rhythmic yet unpredictable dance of high-contrast light and dark within the stark, colorless environment creates a deeply unsettling and tense atmosphere, transforming the desolate corridor into a space of pure psychological terror.", + "seed": 26805, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_1/index_3099_seed_26805_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011980699584819376, + "vlm_logprob": -0.011980699584819376, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.011980699584819376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013904843479394913 + ], + "opposite_logprob_tokens": [ + -4.763905048370361 + ], + "logprob_reward": -0.013904843479394913, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013904843479394913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15757079422473907, + 0.05074252933263779, + 0.03481854125857353, + 0.03661217913031578, + 0.07713573426008224, + 0.06997350603342056, + 0.04843859374523163, + 0.053035374730825424, + 0.39893534779548645, + 0.07273745536804199 + ], + "credit_map": [ + 0.019591690972447395, + 0.016833050176501274, + 0.01666010171175003, + 0.026864705607295036, + 0.013238868676126003, + 0.019397521391510963, + 0.027526477351784706, + 0.017458369955420494, + 0.005659176968038082, + 0.005898874718695879, + 0.005920650903135538, + 0.007598250638693571, + 0.004216771572828293, + 0.006404043175280094, + 0.008614289574325085, + 0.006430470384657383, + 0.0032641568686813116, + 0.004376668948680162, + 0.004296886269003153, + 0.005303950048983097, + 0.0029202396981418133, + 0.003624160075560212, + 0.005653154570609331, + 0.00537932338193059, + 0.0036648986861109734, + 0.005577499978244305, + 0.004285398405045271, + 0.005871965549886227, + 0.003942497074604034, + 0.0034587145783007145, + 0.005474945064634085, + 0.0043362583965063095, + 0.008279796689748764, + 0.013076006434857845, + 0.008734497241675854, + 0.00873166136443615, + 0.00871712900698185, + 0.00808154046535492, + 0.009447002783417702, + 0.012068093754351139, + 0.007247268222272396, + 0.013326049782335758, + 0.00815477967262268, + 0.008981666527688503, + 0.008782156743109226, + 0.0072481390088796616, + 0.007520484272390604, + 0.008712955750524998, + 0.005490508396178484, + 0.004203749354928732, + 0.005790432915091515, + 0.007054748013615608, + 0.0052738068625330925, + 0.00711803138256073, + 0.005946781020611525, + 0.007560531608760357, + 0.006256886292248964, + 0.0068114930763840675, + 0.006574924569576979, + 0.00589656550437212, + 0.003982954192906618, + 0.011644124984741211, + 0.005532501265406609, + 0.006335919257253408, + 0.05955948308110237, + 0.025836024433374405, + 0.05797021463513374, + 0.05122203379869461, + 0.0324573777616024, + 0.06608568131923676, + 0.04743470996618271, + 0.05836978927254677, + 0.006075446493923664, + 0.010942314751446247, + 0.012041819281876087, + 0.00724590802565217, + 0.007313805632293224, + 0.00813956931233406, + 0.010965065099298954, + 0.010013522580265999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.013904843479394913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610027626156807 + ], + "opposite_logprob_tokens": [ + -4.1411004066467285 + ], + "logprob_reward": -0.01610027626156807, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610027626156807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3090222477912903, + 0.09626796841621399, + 0.048541489988565445, + 0.049957212060689926, + 0.04232028126716614, + 0.09847426414489746, + 0.03313053399324417, + 0.05789633095264435, + 0.21866846084594727, + 0.04572124779224396 + ], + "credit_map": [ + 0.032775312662124634, + 0.034492265433073044, + 0.03353419154882431, + 0.05395038053393364, + 0.028025588020682335, + 0.037526801228523254, + 0.05157741159200668, + 0.03714029863476753, + 0.01022243220359087, + 0.009317290037870407, + 0.011366861872375011, + 0.014175362884998322, + 0.007601745892316103, + 0.015423940494656563, + 0.014924436807632446, + 0.01323589775711298, + 0.00467710243538022, + 0.004524726886302233, + 0.006013369653373957, + 0.00684816250577569, + 0.0038147801533341408, + 0.007026679813861847, + 0.007837408222258091, + 0.007799262180924416, + 0.003967578988522291, + 0.008585724048316479, + 0.006223691161721945, + 0.0061216154135763645, + 0.005456092767417431, + 0.006303426343947649, + 0.006738272961229086, + 0.006560812704265118, + 0.004698880948126316, + 0.005513541400432587, + 0.004818377085030079, + 0.005456536542624235, + 0.0038174763321876526, + 0.006016559898853302, + 0.006279389839619398, + 0.005719518754631281, + 0.010110531933605671, + 0.021421924233436584, + 0.01067840214818716, + 0.01172795332968235, + 0.012586326338350773, + 0.009654765948653221, + 0.009057648479938507, + 0.013236711733043194, + 0.004248758312314749, + 0.0030268686823546886, + 0.0042361048981547356, + 0.00472993915900588, + 0.0022138773929327726, + 0.0035393876023590565, + 0.004026154056191444, + 0.00710944551974535, + 0.0078103384003043175, + 0.006464449688792229, + 0.006954040843993425, + 0.006984390318393707, + 0.004336472600698471, + 0.010629103519022465, + 0.007348746992647648, + 0.007368789054453373, + 0.02939288131892681, + 0.03503451123833656, + 0.02928307093679905, + 0.02009769342839718, + 0.02378593385219574, + 0.025944851338863373, + 0.02730667218565941, + 0.02782284840941429, + 0.005348479840904474, + 0.005076311994343996, + 0.0063835615292191505, + 0.00465448247268796, + 0.004370305221527815, + 0.006886068731546402, + 0.006798388436436653, + 0.006203650962561369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.01610027626156807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648237023502588 + ], + "opposite_logprob_tokens": [ + -5.379648208618164 + ], + "logprob_reward": -0.004648237023502588, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648237023502588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15115684270858765, + 0.09993959218263626, + 0.11337761580944061, + 0.1923711597919464, + 0.036340873688459396, + 0.22605550289154053, + 0.01755797676742077, + 0.02624990977346897, + 0.0860312357544899, + 0.05091923847794533 + ], + "credit_map": [ + 0.01487416960299015, + 0.023907838389277458, + 0.017412973567843437, + 0.022385215386748314, + 0.016372719779610634, + 0.017917782068252563, + 0.021116169169545174, + 0.017169993370771408, + 0.010879864916205406, + 0.02094411849975586, + 0.01044938713312149, + 0.014453427866101265, + 0.011702142655849457, + 0.008341903798282146, + 0.010461210273206234, + 0.012707540765404701, + 0.011946679092943668, + 0.02175241895020008, + 0.012052076868712902, + 0.016757236793637276, + 0.013369855470955372, + 0.009393728338181973, + 0.01336405985057354, + 0.0147415641695261, + 0.01953979581594467, + 0.044883888214826584, + 0.020620308816432953, + 0.023251337930560112, + 0.02644916996359825, + 0.017014481127262115, + 0.01793941855430603, + 0.02267276495695114, + 0.0033245482482016087, + 0.0088232746347785, + 0.00432461267337203, + 0.0036630514077842236, + 0.004645912442356348, + 0.003346106968820095, + 0.0041113924235105515, + 0.004101980943232775, + 0.023409441113471985, + 0.05332307517528534, + 0.02449900098145008, + 0.026806386187672615, + 0.030372830107808113, + 0.01953291706740856, + 0.018834726884961128, + 0.029277127236127853, + 0.0017343465005978942, + 0.002012114506214857, + 0.002227487275376916, + 0.0019044075161218643, + 0.002372969873249531, + 0.002373018302023411, + 0.001987348310649395, + 0.002946284366771579, + 0.0027989265508949757, + 0.0027464714366942644, + 0.003379624569788575, + 0.002778098452836275, + 0.0020238279830664396, + 0.005513262003660202, + 0.003218964906409383, + 0.003790736896917224, + 0.01345091313123703, + 0.012638232670724392, + 0.01178884319961071, + 0.009558720514178276, + 0.007444140035659075, + 0.011472143232822418, + 0.008771635591983795, + 0.010906605049967766, + 0.00658143125474453, + 0.005182793363928795, + 0.006959537044167519, + 0.006003451533615589, + 0.0038381742779165506, + 0.0077679879032075405, + 0.006398443132638931, + 0.008187422528862953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.004648237023502588 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019464610377326608 + ], + "opposite_logprob_tokens": [ + -6.251946449279785 + ], + "logprob_reward": -0.0019464610377326608, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019464610377326608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16459207236766815, + 0.05066881328821182, + 0.04386483132839203, + 0.1009700745344162, + 0.04932539537549019, + 0.14706383645534515, + 0.032104890793561935, + 0.05283857882022858, + 0.15966644883155823, + 0.1989051103591919 + ], + "credit_map": [ + 0.024681273847818375, + 0.01408330723643303, + 0.019958224147558212, + 0.013318256475031376, + 0.016734683886170387, + 0.036934077739715576, + 0.01827147975564003, + 0.020610757172107697, + 0.007041869685053825, + 0.004929845687001944, + 0.006671422626823187, + 0.005386242642998695, + 0.004356886725872755, + 0.009618265554308891, + 0.005857121665030718, + 0.006807155907154083, + 0.004683483392000198, + 0.00649172393605113, + 0.004888484720140696, + 0.005292360670864582, + 0.004408248234540224, + 0.006364097353070974, + 0.0052504511550068855, + 0.00648597814142704, + 0.011290447786450386, + 0.014568525366485119, + 0.0110986502841115, + 0.012009741738438606, + 0.011650354601442814, + 0.01323140412569046, + 0.012767927721142769, + 0.014353012666106224, + 0.005820324644446373, + 0.006483509670943022, + 0.005803205538541079, + 0.005147475749254227, + 0.0056329770013689995, + 0.008448043838143349, + 0.005174180492758751, + 0.006815672852098942, + 0.015979282557964325, + 0.02573285438120365, + 0.01614230126142502, + 0.019095327705144882, + 0.017135145142674446, + 0.014771632850170135, + 0.01594376377761364, + 0.022263525053858757, + 0.0037028249353170395, + 0.003837205935269594, + 0.004630618263036013, + 0.004006384871900082, + 0.0035015929024666548, + 0.0037024307530373335, + 0.0035558268427848816, + 0.005168006289750338, + 0.005629133433103561, + 0.005371623672544956, + 0.007231212221086025, + 0.006452052854001522, + 0.0038264128379523754, + 0.011738772504031658, + 0.006037589628249407, + 0.006551777012646198, + 0.021389065310359, + 0.015177970752120018, + 0.022387122735381126, + 0.019932078197598457, + 0.01291748695075512, + 0.02461707964539528, + 0.017349116504192352, + 0.025896504521369934, + 0.022194920107722282, + 0.01671884022653103, + 0.03723730146884918, + 0.024976806715130806, + 0.013727949932217598, + 0.023945845663547516, + 0.026196368038654327, + 0.03390707075595856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0019464610377326608 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long, narrow, decaying concrete corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sole light source a bare fluorescent tube suspended by frayed wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light emit a cold, sickly greenish-white glow and flicker intermittently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do deep shadows cast by ceiling ducts and wall imperfections waver and contort into shifting shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330368012189865 + ], + "opposite_logprob_tokens": [ + -3.773303747177124 + ], + "logprob_reward": -0.02330368012189865, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330368012189865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11784889549016953, + 0.08503658324480057, + 0.07071730494499207, + 0.0916227400302887, + 0.07783358544111252, + 0.06409330666065216, + 0.08183588087558746, + 0.037996452301740646, + 0.15326300263404846, + 0.2197522670030594 + ], + "credit_map": [ + 0.014750214293599129, + 0.01170975249260664, + 0.01603570580482483, + 0.01699019968509674, + 0.010582751594483852, + 0.01571516878902912, + 0.014755815267562866, + 0.017309293150901794, + 0.010253793559968472, + 0.009870711714029312, + 0.009228607639670372, + 0.011629719287157059, + 0.007828856818377972, + 0.012024563737213612, + 0.011339949443936348, + 0.012860380113124847, + 0.008786773309111595, + 0.009563757106661797, + 0.0072562978602945805, + 0.009873472154140472, + 0.0066952998749911785, + 0.009219256229698658, + 0.009896865114569664, + 0.009425589814782143, + 0.0118942242115736, + 0.012508390471339226, + 0.008720096200704575, + 0.011886236257851124, + 0.00945728924125433, + 0.013330858200788498, + 0.012949378229677677, + 0.01087627001106739, + 0.00994472298771143, + 0.010253059677779675, + 0.008697966113686562, + 0.009681196883320808, + 0.007938825525343418, + 0.0085108932107687, + 0.010193530470132828, + 0.012613391503691673, + 0.006980318110436201, + 0.012052125297486782, + 0.006801133044064045, + 0.00827404297888279, + 0.007097144145518541, + 0.007200432009994984, + 0.007301710080355406, + 0.00838639959692955, + 0.010310818441212177, + 0.007005145773291588, + 0.011395871639251709, + 0.010050793178379536, + 0.007475578226149082, + 0.009715668857097626, + 0.008629592135548592, + 0.017252415418624878, + 0.0032761087641119957, + 0.005068030674010515, + 0.005655411630868912, + 0.004728641360998154, + 0.0027535564731806517, + 0.0081785898655653, + 0.0029024037066847086, + 0.005433710291981697, + 0.022567329928278923, + 0.01413425337523222, + 0.0245041586458683, + 0.017625588923692703, + 0.012766469269990921, + 0.022249888628721237, + 0.017182650044560432, + 0.02223266288638115, + 0.02586817555129528, + 0.018695125356316566, + 0.038400813937187195, + 0.024929236620664597, + 0.016138488426804543, + 0.0314200334250927, + 0.02726445533335209, + 0.03703593090176582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.02330368012189865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_tensor([4551, 4551])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_tensor([4551, 4551])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f26e1630598a234d40c2603c2392a7c079bbc83c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_tensor([4551, 4551])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "horizontal tracking shot, shot from a slightly upward angle, follows General Zhao Yun, clad in a full suit of intricately carved silver armor, as he charges forward unstoppably across the dusty battlefield. The armor gleams with a cold metallic sheen in the dim sunlight, the intricately carved dragons on his breastplate, pauldrons, and gauntlets clearly visible. The scratches and dents of battle bear witness to the brutality of fierce combat. His heroic face is matted with dust and sweat, his eyes sharp and resolute, his lips pursed with a resolute resolve. His long, bound black hair falls loose beneath his helmet as he violently moves. He holds a spear steady in one hand, the tip gleaming with a cold gleam, pointed forward. The red tassel on the spearhead, combined with the flowing white cape behind him, dances wildly in the wind, creating a dramatic arc. His powerful legs stamped against the churning scorched earth, each step stirring up vast clouds of yellow sand, leaving behind a blur of chaotic battlefield filled with the clash of weapons and shouts. The entire image is imbued with a strong sense of movement and epic heroism, with the color palette leaning towards a warm and gritty brown, highlighting the general's unwavering focus and bravery as he charges forward amidst the chaos.", + "seed": 2406, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_4551_seed_2406_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029768912866711617, + "vlm_logprob": -0.029768912866711617, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.029768912866711617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05221644788980484 + ], + "opposite_logprob_tokens": [ + -3.3022165298461914 + ], + "logprob_reward": -0.05221644788980484, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05221644788980484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13565640151500702, + 0.10234446823596954, + 0.09330430626869202, + 0.1184314712882042, + 0.10237178206443787, + 0.03169720619916916, + 0.07607538998126984, + 0.059707000851631165, + 0.12176358699798584, + 0.1586483269929886 + ], + "credit_map": [ + 0.014252428896725178, + 0.01819061115384102, + 0.01779627613723278, + 0.022043339908123016, + 0.015684884041547775, + 0.016499115154147148, + 0.01437319815158844, + 0.01681656390428543, + 0.010567685589194298, + 0.012170207686722279, + 0.009209501557052135, + 0.021798191592097282, + 0.013834320940077305, + 0.009739281609654427, + 0.01373581774532795, + 0.011289465241134167, + 0.0111876605078578, + 0.007777216844260693, + 0.018807154148817062, + 0.007417605724185705, + 0.012015377171337605, + 0.007767567411065102, + 0.015130025334656239, + 0.01320169772952795, + 0.013102896511554718, + 0.016032708808779716, + 0.015880269929766655, + 0.013822183944284916, + 0.013352357782423496, + 0.01642576791346073, + 0.014317736960947514, + 0.015497539192438126, + 0.013242813758552074, + 0.01710321009159088, + 0.016761627048254013, + 0.016044506803154945, + 0.011873295530676842, + 0.008129597641527653, + 0.011776993051171303, + 0.007439739070832729, + 0.0035790270194411278, + 0.0037619583308696747, + 0.006281769834458828, + 0.0029287636280059814, + 0.0021491285879164934, + 0.005063240882009268, + 0.0023541105911135674, + 0.0055792066268622875, + 0.007136570289731026, + 0.008768302388489246, + 0.009265477769076824, + 0.008155443705618382, + 0.008668148890137672, + 0.007129242643713951, + 0.01866041123867035, + 0.008291790261864662, + 0.005226525943726301, + 0.006294627673923969, + 0.005602382589131594, + 0.008506725542247295, + 0.006437249481678009, + 0.009309440851211548, + 0.010947490110993385, + 0.007382560055702925, + 0.012643408961594105, + 0.019736656919121742, + 0.010981027968227863, + 0.01456764992326498, + 0.016947191208600998, + 0.012967408634722233, + 0.01670842617750168, + 0.017211824655532837, + 0.010671705938875675, + 0.0173470638692379, + 0.022345485165715218, + 0.020920582115650177, + 0.025348206982016563, + 0.01705322042107582, + 0.01739204302430153, + 0.027570027858018875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.05221644788980484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085494309663773 + ], + "opposite_logprob_tokens": [ + -4.141085624694824 + ], + "logprob_reward": -0.016085494309663773, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085494309663773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24274751543998718, + 0.1395491659641266, + 0.046972282230854034, + 0.17777089774608612, + 0.06487268209457397, + 0.045943137258291245, + 0.06391504406929016, + 0.034890949726104736, + 0.10847022384405136, + 0.07486813515424728 + ], + "credit_map": [ + 0.024883300065994263, + 0.037344224750995636, + 0.03797781094908714, + 0.035373255610466, + 0.02293161302804947, + 0.016439292579889297, + 0.026231057941913605, + 0.04156695678830147, + 0.011278465390205383, + 0.013949407264590263, + 0.015174577943980694, + 0.019122157245874405, + 0.0323801152408123, + 0.016620177775621414, + 0.017123272642493248, + 0.01390099711716175, + 0.004903282504528761, + 0.0030587222427129745, + 0.010661272332072258, + 0.0037530383560806513, + 0.006357481237500906, + 0.00481155002489686, + 0.008862437680363655, + 0.004564496222883463, + 0.022328976541757584, + 0.019258733838796616, + 0.028429821133613586, + 0.017455851659178734, + 0.023800233379006386, + 0.024584181606769562, + 0.0197147186845541, + 0.02219838835299015, + 0.011306317523121834, + 0.00740449596196413, + 0.0125703951343894, + 0.007623056415468454, + 0.0084524629637599, + 0.004949153400957584, + 0.007054557092487812, + 0.005512242205440998, + 0.0056359898298978806, + 0.006153585854917765, + 0.007368204649537802, + 0.004693569615483284, + 0.005028544459491968, + 0.0059671238996088505, + 0.004615103360265493, + 0.006481016520410776, + 0.00593271479010582, + 0.007792080752551556, + 0.007428184617310762, + 0.006554886698722839, + 0.012984247878193855, + 0.005414309445768595, + 0.012536199763417244, + 0.0052724177949130535, + 0.005099060479551554, + 0.0031732821371406317, + 0.0022792911622673273, + 0.0037045120261609554, + 0.004868543706834316, + 0.005451240576803684, + 0.004729392472654581, + 0.005585630424320698, + 0.013057230971753597, + 0.02311762049794197, + 0.010185051709413528, + 0.013856037519872189, + 0.01465655118227005, + 0.009930955246090889, + 0.01469049509614706, + 0.008976273238658905, + 0.007343736011534929, + 0.006988625042140484, + 0.008564436808228493, + 0.010875623673200607, + 0.010389595292508602, + 0.007800636813044548, + 0.012829097919166088, + 0.010076375678181648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.016085494309663773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961494520306587 + ], + "opposite_logprob_tokens": [ + -5.1309614181518555 + ], + "logprob_reward": -0.005961494520306587, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961494520306587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08950711786746979, + 0.03903438523411751, + 0.05591869726777077, + 0.08588650077581406, + 0.04704700782895088, + 0.053811702877283096, + 0.30287277698516846, + 0.03714972734451294, + 0.0952361598610878, + 0.19353602826595306 + ], + "credit_map": [ + 0.007591062691062689, + 0.016440337523818016, + 0.011667069047689438, + 0.016275251284241676, + 0.010231077671051025, + 0.007884559221565723, + 0.010041367262601852, + 0.009376379661262035, + 0.003927181474864483, + 0.005189135670661926, + 0.0033214956056326628, + 0.004471779800951481, + 0.007232381962239742, + 0.004347017966210842, + 0.005151568911969662, + 0.005393821280449629, + 0.006707833614200354, + 0.004292226862162352, + 0.012428035959601402, + 0.0067943399772048, + 0.004918606486171484, + 0.00450145173817873, + 0.008871978148818016, + 0.00740421749651432, + 0.012869499623775482, + 0.0075691924430429935, + 0.021316178143024445, + 0.006849449593573809, + 0.013508265838027, + 0.0075301192700862885, + 0.0070288535207509995, + 0.009214930236339569, + 0.005580165423452854, + 0.008974861353635788, + 0.005911021493375301, + 0.007631573360413313, + 0.005688304081559181, + 0.0028104546945542097, + 0.005670100916177034, + 0.004780523478984833, + 0.005156878847628832, + 0.00592609541490674, + 0.009527292102575302, + 0.005771620664745569, + 0.005174705293029547, + 0.0055677215568721294, + 0.003620587522163987, + 0.013066795654594898, + 0.029966937378048897, + 0.05421147495508194, + 0.040014222264289856, + 0.0389290377497673, + 0.04088573902845383, + 0.029245100915431976, + 0.04523497074842453, + 0.02438526228070259, + 0.0037455642595887184, + 0.004031304735690355, + 0.0029405669774860144, + 0.006307065021246672, + 0.003537885146215558, + 0.004782759118825197, + 0.006110963877290487, + 0.005693613085895777, + 0.012079854495823383, + 0.027001960203051567, + 0.009236115962266922, + 0.012380599044263363, + 0.010561538860201836, + 0.007332969456911087, + 0.009656145237386227, + 0.006986966356635094, + 0.044158004224300385, + 0.02517281100153923, + 0.02421329915523529, + 0.033200714737176895, + 0.018966330215334892, + 0.014117790386080742, + 0.018040578812360764, + 0.015666475519537926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.005961494520306587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543711811304092 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543711811304092, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543711811304092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2281607687473297, + 0.06246519461274147, + 0.11626898497343063, + 0.3012424409389496, + 0.045429982244968414, + 0.05815200135111809, + 0.048373542726039886, + 0.044408030807971954, + 0.04951624944806099, + 0.04598282650113106 + ], + "credit_map": [ + 0.027530904859304428, + 0.02438211441040039, + 0.031666602939367294, + 0.03045937418937683, + 0.023806480690836906, + 0.022603385150432587, + 0.02540481463074684, + 0.04230710119009018, + 0.004607825074344873, + 0.006959224585443735, + 0.00560518354177475, + 0.008433018811047077, + 0.011597900651395321, + 0.0065783108584582806, + 0.011017357930541039, + 0.007666369900107384, + 0.013995320536196232, + 0.009193869307637215, + 0.020119227468967438, + 0.009621767327189445, + 0.020793233066797256, + 0.01357439998537302, + 0.015721149742603302, + 0.013250015676021576, + 0.03959031403064728, + 0.0329970121383667, + 0.05903002619743347, + 0.027995724231004715, + 0.040248218923807144, + 0.03538171574473381, + 0.032775044441223145, + 0.033224377781152725, + 0.006489909719675779, + 0.0058348593302071095, + 0.008302882313728333, + 0.005232424009591341, + 0.005549181718379259, + 0.004346300847828388, + 0.005137478467077017, + 0.004536948632448912, + 0.004790619947016239, + 0.005296849645674229, + 0.009634209796786308, + 0.008182788267731667, + 0.006078153382986784, + 0.008355064317584038, + 0.006190921179950237, + 0.009623397141695023, + 0.004426448605954647, + 0.0063524809665977955, + 0.006829055491834879, + 0.00491893058642745, + 0.005437325220555067, + 0.005861299578100443, + 0.009279061108827591, + 0.0052689420990645885, + 0.006371730472892523, + 0.00453199353069067, + 0.003553335089236498, + 0.004916958510875702, + 0.004491893108934164, + 0.00748008769005537, + 0.004661891609430313, + 0.008400142192840576, + 0.007296927273273468, + 0.006550088990479708, + 0.0053436667658388615, + 0.0058400691486895084, + 0.006015802267938852, + 0.0056894272565841675, + 0.007150192279368639, + 0.00563007639721036, + 0.003588132793083787, + 0.003993723075836897, + 0.005793375428766012, + 0.005926457699388266, + 0.006626472808420658, + 0.006126676686108112, + 0.006398260593414307, + 0.007529727183282375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012543711811304092 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203741580247879 + ], + "opposite_logprob_tokens": [ + -2.812037467956543 + ], + "logprob_reward": -0.06203741580247879, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203741580247879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20463639497756958, + 0.08684606105089188, + 0.06473973393440247, + 0.10087142884731293, + 0.049328453838825226, + 0.05487727373838425, + 0.08962289988994598, + 0.0735224038362503, + 0.12882162630558014, + 0.14673376083374023 + ], + "credit_map": [ + 0.022737085819244385, + 0.038481105118989944, + 0.02614007517695427, + 0.037511903792619705, + 0.023710977286100388, + 0.01559280976653099, + 0.02053816430270672, + 0.019924266263842583, + 0.008077463135123253, + 0.010254137217998505, + 0.00819031335413456, + 0.010293056257069111, + 0.01800605095922947, + 0.00950706098228693, + 0.012492590583860874, + 0.0100253876298666, + 0.007157441694289446, + 0.004101195838302374, + 0.009880004450678825, + 0.00660095876082778, + 0.00726894848048687, + 0.006792941130697727, + 0.012626071460545063, + 0.01031216885894537, + 0.01198433805257082, + 0.013331218622624874, + 0.016233516857028008, + 0.010084059089422226, + 0.01618383452296257, + 0.012327326461672783, + 0.012389934621751308, + 0.008337191306054592, + 0.005361134186387062, + 0.00723701436072588, + 0.006621708627790213, + 0.007628877647221088, + 0.006578052882105112, + 0.0036948486231267452, + 0.007102626375854015, + 0.00510418601334095, + 0.005553734488785267, + 0.007274433970451355, + 0.009281815961003304, + 0.0064104096964001656, + 0.0051239263266325, + 0.008578477427363396, + 0.004817910026758909, + 0.007836567237973213, + 0.009810843504965305, + 0.006816039327532053, + 0.011247129179537296, + 0.006639695726335049, + 0.015447656624019146, + 0.009756617248058319, + 0.015775129199028015, + 0.014129784889519215, + 0.006540928967297077, + 0.009426647797226906, + 0.007783690467476845, + 0.009900454431772232, + 0.008315731771290302, + 0.011149251833558083, + 0.011902783997356892, + 0.008502914570271969, + 0.012784862890839577, + 0.023755310103297234, + 0.01126704178750515, + 0.012366997078061104, + 0.019378986209630966, + 0.012774735689163208, + 0.02101939171552658, + 0.015474295243620872, + 0.011771941557526588, + 0.016783133149147034, + 0.017786763608455658, + 0.0174751877784729, + 0.027510667219758034, + 0.015089181251823902, + 0.024390719830989838, + 0.015926165506243706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.06203741580247879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking shot, shot from a slightly upward angle, follows General Zhao Yun, clad in a full suit of intricately carved silver armor, as he charges forward unstoppably across the dusty battlefield. The armor gleams with a cold metallic sheen in the dim sunlight, the intricately carved dragons on his breastplate, pauldrons, and gauntlets clearly visible. The scratches and dents of battle bear witness to the brutality of fierce combat. His heroic face is matted with dust and sweat, his eyes sharp and resolute, his lips pursed with a resolute resolve. His long, bound black hair falls loose beneath his helmet as he violently moves. He holds a spear steady in one hand, the tip gleaming with a cold gleam, pointed forward. The red tassel on the spearhead, combined with the flowing white cape behind him, dances wildly in the wind, creating a dramatic arc. His powerful legs stamped against the churning scorched earth, each step stirring up vast clouds of yellow sand, leaving behind a blur of chaotic battlefield filled with the clash of weapons and shouts. The entire image is imbued with a strong sense of movement and epic heroism, with the color palette leaning towards a warm and gritty brown, highlighting the general's unwavering focus and bravery as he charges forward amidst the chaos.", + "seed": 2407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_4551_seed_2407_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029730059579014777, + "vlm_logprob": -0.029730059579014777, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.029730059579014777, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04779212176799774 + ], + "opposite_logprob_tokens": [ + -3.2977921962738037 + ], + "logprob_reward": -0.04779212176799774, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04779212176799774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09372352808713913, + 0.18627989292144775, + 0.043634988367557526, + 0.038365986198186874, + 0.299666166305542, + 0.12953831255435944, + 0.08457867801189423, + 0.06576026976108551, + 0.0299533698707819, + 0.028498772531747818 + ], + "credit_map": [ + 0.007212020922452211, + 0.0091641154140234, + 0.011034577153623104, + 0.017683589830994606, + 0.01598585955798626, + 0.00666461419314146, + 0.016696352511644363, + 0.009282395243644714, + 0.030764581635594368, + 0.010756513103842735, + 0.012012163177132607, + 0.020746605470776558, + 0.009662730619311333, + 0.043902650475502014, + 0.012747452594339848, + 0.04568719118833542, + 0.006321408785879612, + 0.004625148605555296, + 0.00820519495755434, + 0.007301395293325186, + 0.004352442920207977, + 0.004686362575739622, + 0.004369730595499277, + 0.003773306030780077, + 0.0034898340236395597, + 0.004165029618889093, + 0.00461560906842351, + 0.003742012195289135, + 0.007955112494528294, + 0.004589827731251717, + 0.0042604985646903515, + 0.0055480641312897205, + 0.016102084890007973, + 0.018932512030005455, + 0.11003038287162781, + 0.09203397482633591, + 0.02094852179288864, + 0.010229178704321384, + 0.023545434698462486, + 0.007844049483537674, + 0.01813899353146553, + 0.019531944766640663, + 0.013135670684278011, + 0.021127747371792793, + 0.01204141229391098, + 0.012348424643278122, + 0.014294219203293324, + 0.01891990751028061, + 0.010503776371479034, + 0.011800532229244709, + 0.014724603854119778, + 0.01067205611616373, + 0.010505005717277527, + 0.011074368841946125, + 0.009203909896314144, + 0.006094420328736305, + 0.007745618466287851, + 0.006121910642832518, + 0.008005669340491295, + 0.0053958590142428875, + 0.014071936719119549, + 0.008894377388060093, + 0.009775680489838123, + 0.00574921490624547, + 0.002919526072219014, + 0.0044620949774980545, + 0.004801069851964712, + 0.003212973475456238, + 0.0036679732147604227, + 0.004262325819581747, + 0.0029213824309408665, + 0.0037060237955302, + 0.003410629229620099, + 0.0036234701983630657, + 0.002811400219798088, + 0.0024069841019809246, + 0.005264980252832174, + 0.0034909944515675306, + 0.0037995728198438883, + 0.003690742887556553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.04779212176799774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004113071598112583 + ], + "opposite_logprob_tokens": [ + -5.50411319732666 + ], + "logprob_reward": -0.004113071598112583, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004113071598112583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09960553795099258, + 0.0931566059589386, + 0.0850137323141098, + 0.024536162614822388, + 0.11914230138063431, + 0.038930825889110565, + 0.05386433005332947, + 0.3353613018989563, + 0.11521806567907333, + 0.03517111390829086 + ], + "credit_map": [ + 0.010969043709337711, + 0.007832398638129234, + 0.011634971015155315, + 0.017336871474981308, + 0.012931425124406815, + 0.009172720834612846, + 0.013881037943065166, + 0.015847062692046165, + 0.007004020269960165, + 0.0034825103357434273, + 0.004762289579957724, + 0.0093945087864995, + 0.004074855707585812, + 0.028714748099446297, + 0.0068315062671899796, + 0.028892161324620247, + 0.008860099129378796, + 0.007922043092548847, + 0.016485093161463737, + 0.010426973924040794, + 0.006791128311306238, + 0.01624239981174469, + 0.010528833605349064, + 0.007757163606584072, + 0.002420759294182062, + 0.003498578444123268, + 0.0026627767365425825, + 0.0027004340663552284, + 0.004496903624385595, + 0.0029098561499267817, + 0.002630462171509862, + 0.003216391196474433, + 0.006192400120198727, + 0.01217928621917963, + 0.047989945858716965, + 0.033397190272808075, + 0.005815343931317329, + 0.0034880770836025476, + 0.007770856376737356, + 0.0023091973271220922, + 0.004312543664127588, + 0.0045748017728328705, + 0.004084098618477583, + 0.006034189835190773, + 0.005371738225221634, + 0.004816901870071888, + 0.005362413357943296, + 0.004374138545244932, + 0.008440577425062656, + 0.004232659470289946, + 0.0112756984308362, + 0.006426399108022451, + 0.005475993733853102, + 0.006992418318986893, + 0.0053812968544662, + 0.005639287177473307, + 0.045793693512678146, + 0.03231700882315636, + 0.06682626157999039, + 0.029786942526698112, + 0.03675173595547676, + 0.048657454550266266, + 0.04318498820066452, + 0.0320432148873806, + 0.011369631625711918, + 0.015178985893726349, + 0.01699589565396309, + 0.010899161919951439, + 0.025858495384454727, + 0.01362519059330225, + 0.011306514963507652, + 0.009984192438423634, + 0.004752539098262787, + 0.0028717077802866697, + 0.00509906280785799, + 0.0033414119388908148, + 0.0049180807545781136, + 0.004359339363873005, + 0.0037243994884192944, + 0.0061045740731060505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.004113071598112583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205760322511196 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205760322511196, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205760322511196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2273695468902588, + 0.16900351643562317, + 0.09646253287792206, + 0.03732747212052345, + 0.07740035653114319, + 0.05610308051109314, + 0.07381223887205124, + 0.143812358379364, + 0.0352620929479599, + 0.08344686776399612 + ], + "credit_map": [ + 0.02245722897350788, + 0.02168275974690914, + 0.025915663689374924, + 0.042691342532634735, + 0.024904968217015266, + 0.02220192365348339, + 0.03311491757631302, + 0.03440072014927864, + 0.014474883675575256, + 0.0166099201887846, + 0.014443627558648586, + 0.018654923886060715, + 0.016073454171419144, + 0.03715579956769943, + 0.015515651553869247, + 0.036075253039598465, + 0.008568404242396355, + 0.008876658976078033, + 0.012463543564081192, + 0.010876297950744629, + 0.02191043272614479, + 0.013086406514048576, + 0.01432235911488533, + 0.006358415819704533, + 0.0033060654532164335, + 0.0048302458599209785, + 0.005480347201228142, + 0.004208265338093042, + 0.00564152654260397, + 0.006942423991858959, + 0.0034095013979822397, + 0.003509093774482608, + 0.010816339403390884, + 0.014113250188529491, + 0.012311178259551525, + 0.01767576113343239, + 0.006484716199338436, + 0.004844821058213711, + 0.007299966178834438, + 0.003854320151731372, + 0.00508604571223259, + 0.00516083836555481, + 0.0058299824595451355, + 0.009172686375677586, + 0.00908527709543705, + 0.006476554088294506, + 0.007795505691319704, + 0.007496184203773737, + 0.009172887541353703, + 0.008650396019220352, + 0.009922010824084282, + 0.011023108847439289, + 0.010208993218839169, + 0.01240435242652893, + 0.006952209398150444, + 0.005478275939822197, + 0.019819676876068115, + 0.013446908444166183, + 0.02656395547091961, + 0.011293644085526466, + 0.01987476646900177, + 0.019227085635066032, + 0.019333599135279655, + 0.014252718538045883, + 0.004404253792017698, + 0.005328082013875246, + 0.004710646811872721, + 0.003947890363633633, + 0.006180792115628719, + 0.0038884845562279224, + 0.003911002539098263, + 0.00289093772880733, + 0.009663047268986702, + 0.008756907656788826, + 0.010917888954281807, + 0.006788826547563076, + 0.008798250928521156, + 0.01172632910311222, + 0.010293534956872463, + 0.016502078622579575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.014205760322511196 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538531795144081 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538531795144081, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538531795144081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08666391670703888, + 0.1569742113351822, + 0.06656995415687561, + 0.05866639316082001, + 0.19715221226215363, + 0.07858263701200485, + 0.12453962862491608, + 0.05078345164656639, + 0.0738159567117691, + 0.10625161975622177 + ], + "credit_map": [ + 0.014934578910470009, + 0.009542609564960003, + 0.012299924157559872, + 0.014755931682884693, + 0.008585184812545776, + 0.006808662787079811, + 0.011530977673828602, + 0.008206035941839218, + 0.031224464997649193, + 0.012052403762936592, + 0.011022064834833145, + 0.01832125522196293, + 0.012255377136170864, + 0.029845835641026497, + 0.013383702374994755, + 0.02886909991502762, + 0.007295163348317146, + 0.00960058905184269, + 0.015452619642019272, + 0.008740792982280254, + 0.005683150142431259, + 0.007031891494989395, + 0.00785597413778305, + 0.004909766372293234, + 0.008314711041748524, + 0.01087598130106926, + 0.006229300517588854, + 0.007934670895338058, + 0.008693812415003777, + 0.007846285589039326, + 0.004123962949961424, + 0.0046476642601192, + 0.014961767010390759, + 0.02645718865096569, + 0.07388805598020554, + 0.047447457909584045, + 0.00923179741948843, + 0.00776748126372695, + 0.01253003440797329, + 0.004868429619818926, + 0.00877625122666359, + 0.009712119586765766, + 0.008635143749415874, + 0.01664849929511547, + 0.006860325578600168, + 0.00799216516315937, + 0.010100620798766613, + 0.00985750649124384, + 0.0186337660998106, + 0.014544688165187836, + 0.025510406121611595, + 0.010875669308006763, + 0.010957418009638786, + 0.021161798387765884, + 0.012277040630578995, + 0.010578829795122147, + 0.005842697806656361, + 0.005514536984264851, + 0.012646013870835304, + 0.00414042454212904, + 0.004904404282569885, + 0.00725116953253746, + 0.004597524646669626, + 0.005886679049581289, + 0.006723148748278618, + 0.008068790659308434, + 0.017232147976756096, + 0.008489146828651428, + 0.009603305719792843, + 0.010384604334831238, + 0.005492158699780703, + 0.007822645828127861, + 0.014355326071381569, + 0.011101724579930305, + 0.015187355689704418, + 0.009824320673942566, + 0.010340403765439987, + 0.013554275035858154, + 0.011897006072103977, + 0.019991189241409302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.012538531795144081 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000081241130829 + ], + "opposite_logprob_tokens": [ + -2.6950008869171143 + ], + "logprob_reward": -0.07000081241130829, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000081241130829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10587861388921738, + 0.13527433574199677, + 0.05764206126332283, + 0.06256710737943649, + 0.16219562292099, + 0.061035268008708954, + 0.10128923505544662, + 0.17810691893100739, + 0.050276268273591995, + 0.08573458343744278 + ], + "credit_map": [ + 0.015367616899311543, + 0.013459036126732826, + 0.011626407504081726, + 0.01783348061144352, + 0.012835396453738213, + 0.0084468899294734, + 0.016150105744600296, + 0.010159675031900406, + 0.016462573781609535, + 0.009771574288606644, + 0.011514863930642605, + 0.015256993472576141, + 0.00618519214913249, + 0.03427286818623543, + 0.009667857550084591, + 0.03214240446686745, + 0.0046880510635674, + 0.005875186063349247, + 0.01187993586063385, + 0.007934018969535828, + 0.008704218082129955, + 0.007820669561624527, + 0.005319130141288042, + 0.005420852452516556, + 0.008347515016794205, + 0.0068336776457726955, + 0.00773917231708765, + 0.0065324753522872925, + 0.00895442720502615, + 0.009448657743632793, + 0.006549290847033262, + 0.008161886595189571, + 0.009953096508979797, + 0.009697191417217255, + 0.06444857269525528, + 0.05040993168950081, + 0.008696706965565681, + 0.004678517580032349, + 0.01057012565433979, + 0.003741477383300662, + 0.006604630500078201, + 0.006697115488350391, + 0.005345093552023172, + 0.009534486569464207, + 0.00805780291557312, + 0.0070012095384299755, + 0.01004241406917572, + 0.007752513978630304, + 0.01252787560224533, + 0.011667484417557716, + 0.01600082404911518, + 0.010885940864682198, + 0.015705998986959457, + 0.01768401451408863, + 0.008910233154892921, + 0.007906860671937466, + 0.023100096732378006, + 0.01677730865776539, + 0.027061864733695984, + 0.014901885762810707, + 0.026005038991570473, + 0.02497631311416626, + 0.028203172609210014, + 0.01708124205470085, + 0.006791772320866585, + 0.006743657868355513, + 0.007271007169038057, + 0.005723173264414072, + 0.004734603688120842, + 0.008263915777206421, + 0.004946693312376738, + 0.005801443941891193, + 0.009201752953231335, + 0.008806323632597923, + 0.010485007427632809, + 0.006935226731002331, + 0.015035023912787437, + 0.010735099203884602, + 0.013123284094035625, + 0.011412868276238441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.07000081241130829 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking shot, shot from a slightly upward angle, follows General Zhao Yun, clad in a full suit of intricately carved silver armor, as he charges forward unstoppably across the dusty battlefield. The armor gleams with a cold metallic sheen in the dim sunlight, the intricately carved dragons on his breastplate, pauldrons, and gauntlets clearly visible. The scratches and dents of battle bear witness to the brutality of fierce combat. His heroic face is matted with dust and sweat, his eyes sharp and resolute, his lips pursed with a resolute resolve. His long, bound black hair falls loose beneath his helmet as he violently moves. He holds a spear steady in one hand, the tip gleaming with a cold gleam, pointed forward. The red tassel on the spearhead, combined with the flowing white cape behind him, dances wildly in the wind, creating a dramatic arc. His powerful legs stamped against the churning scorched earth, each step stirring up vast clouds of yellow sand, leaving behind a blur of chaotic battlefield filled with the clash of weapons and shouts. The entire image is imbued with a strong sense of movement and epic heroism, with the color palette leaning towards a warm and gritty brown, highlighting the general's unwavering focus and bravery as he charges forward amidst the chaos.", + "seed": 2408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_4551_seed_2408_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03705412577837706, + "vlm_logprob": -0.03705412577837706, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.03705412577837706, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03957594931125641 + ], + "opposite_logprob_tokens": [ + -3.5395760536193848 + ], + "logprob_reward": -0.03957594931125641, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03957594931125641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15457242727279663, + 0.0934881940484047, + 0.07971605658531189, + 0.29559633135795593, + 0.03215538337826729, + 0.017207372933626175, + 0.04702683910727501, + 0.06186588108539581, + 0.06872405111789703, + 0.14964747428894043 + ], + "credit_map": [ + 0.029668966308236122, + 0.01362825557589531, + 0.016252411529421806, + 0.022679204121232033, + 0.01345362514257431, + 0.02019151858985424, + 0.01691572181880474, + 0.021782727912068367, + 0.014798315241932869, + 0.009899088181555271, + 0.010873369872570038, + 0.011538835242390633, + 0.009635642170906067, + 0.01459436770528555, + 0.010340370237827301, + 0.011808197945356369, + 0.008276060223579407, + 0.010193368420004845, + 0.010702238418161869, + 0.009210655465722084, + 0.011067179962992668, + 0.00946393795311451, + 0.012435166165232658, + 0.008367444388568401, + 0.02650950849056244, + 0.02863316424190998, + 0.036731693893671036, + 0.03755507618188858, + 0.0265357568860054, + 0.060349445790052414, + 0.019341379404067993, + 0.05994030460715294, + 0.004079597536474466, + 0.003090513637289405, + 0.007978135719895363, + 0.0025629661977291107, + 0.003228340297937393, + 0.004606172908097506, + 0.0021687892731279135, + 0.004440869204699993, + 0.0015359235694631934, + 0.0025801516603678465, + 0.0014415220357477665, + 0.002215260174125433, + 0.0025839144363999367, + 0.0022653653286397457, + 0.0025367203634232283, + 0.0020485147833824158, + 0.007873677648603916, + 0.00411073025316, + 0.00930977426469326, + 0.0038105854764580727, + 0.00456126406788826, + 0.006218574475497007, + 0.003772109281271696, + 0.007370122708380222, + 0.00910207536071539, + 0.004746926482766867, + 0.010401647537946701, + 0.006073791068047285, + 0.00928895641118288, + 0.009748351760208607, + 0.005692312493920326, + 0.006811817176640034, + 0.005508827045559883, + 0.0078707505017519, + 0.004248400684446096, + 0.004688709508627653, + 0.004067461937665939, + 0.020168377086520195, + 0.0037594230379909277, + 0.01841210387647152, + 0.018187228590250015, + 0.014106337912380695, + 0.012551852501928806, + 0.013176255859434605, + 0.021993305534124374, + 0.01482051145285368, + 0.02646777778863907, + 0.028344199061393738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.03957594931125641 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866274069994688 + ], + "opposite_logprob_tokens": [ + -4.758662700653076 + ], + "logprob_reward": -0.00866274069994688, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866274069994688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18045474588871002, + 0.03805830329656601, + 0.048510897904634476, + 0.4908669888973236, + 0.042999714612960815, + 0.04705135151743889, + 0.06021834537386894, + 0.03144916892051697, + 0.04032650589942932, + 0.020064007490873337 + ], + "credit_map": [ + 0.03410997614264488, + 0.011414910666644573, + 0.0338846817612648, + 0.01520639006048441, + 0.010408657602965832, + 0.03745715320110321, + 0.012397249229252338, + 0.02557571977376938, + 0.005836746655404568, + 0.003914827946573496, + 0.0036385038401931524, + 0.0038696613628417253, + 0.0043583642691373825, + 0.004841715097427368, + 0.00571074802428484, + 0.005887735169380903, + 0.005055848974734545, + 0.004736803472042084, + 0.010210311971604824, + 0.00455496646463871, + 0.005279454868286848, + 0.006200198549777269, + 0.007340623065829277, + 0.005132691469043493, + 0.02777564711868763, + 0.014945262111723423, + 0.03640398010611534, + 0.023256687447428703, + 0.014532556757330894, + 0.17781108617782593, + 0.01718039996922016, + 0.1789613664150238, + 0.004397060256451368, + 0.004815364722162485, + 0.008776649832725525, + 0.0046434588730335236, + 0.004941999912261963, + 0.006220525596290827, + 0.0031875434797257185, + 0.00601711543276906, + 0.004191137850284576, + 0.006234090309590101, + 0.0058022900484502316, + 0.005004466511309147, + 0.005377860739827156, + 0.00856584683060646, + 0.0054772840812802315, + 0.006398376543074846, + 0.006054256577044725, + 0.0041706194169819355, + 0.009088415652513504, + 0.0059247384779155254, + 0.0045035649091005325, + 0.01294697169214487, + 0.00468981871381402, + 0.01283995807170868, + 0.006168386898934841, + 0.00205039675347507, + 0.005970527417957783, + 0.002424062229692936, + 0.003111997852101922, + 0.005602155812084675, + 0.002537276828661561, + 0.0035843653604388237, + 0.007399383932352066, + 0.006802932359278202, + 0.004393421579152346, + 0.0035018008202314377, + 0.003788554109632969, + 0.007468407973647118, + 0.0022235922515392303, + 0.004748414270579815, + 0.00222615129314363, + 0.0034154707100242376, + 0.0019429171225056052, + 0.0018461301224306226, + 0.002733616391196847, + 0.0020778425969183445, + 0.0029269810765981674, + 0.002894898410886526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.00866274069994688 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981533110141754 + ], + "opposite_logprob_tokens": [ + -3.529815435409546 + ], + "logprob_reward": -0.02981533110141754, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981533110141754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12277728319168091, + 0.1011483445763588, + 0.051333192735910416, + 0.44424960017204285, + 0.03880813717842102, + 0.021815622225403786, + 0.023947540670633316, + 0.02076750062406063, + 0.048045482486486435, + 0.12710727751255035 + ], + "credit_map": [ + 0.01968994550406933, + 0.012609262019395828, + 0.01416847575455904, + 0.016954852268099785, + 0.010433418676257133, + 0.016030732542276382, + 0.014722035266458988, + 0.018168559297919273, + 0.01613868959248066, + 0.011146538890898228, + 0.01139277033507824, + 0.012584990821778774, + 0.01162994746118784, + 0.015569188632071018, + 0.012961968779563904, + 0.009724244475364685, + 0.005905504338443279, + 0.006670942530035973, + 0.009628131985664368, + 0.006668532267212868, + 0.00536553468555212, + 0.004980237688869238, + 0.006728534586727619, + 0.005385770928114653, + 0.02128031477332115, + 0.015042733401060104, + 0.02335507422685623, + 0.012636316008865833, + 0.013380657881498337, + 0.17249636352062225, + 0.011809894815087318, + 0.17424826323986053, + 0.0025844883639365435, + 0.006855650804936886, + 0.00462532602250576, + 0.006431019399315119, + 0.005201449617743492, + 0.004237283952534199, + 0.00361408619210124, + 0.005258830729871988, + 0.0021692104637622833, + 0.0030713104642927647, + 0.0014880598755553365, + 0.002177929040044546, + 0.002240311587229371, + 0.004672436974942684, + 0.0028060113545507193, + 0.0031903504859656096, + 0.002564320107921958, + 0.004029436502605677, + 0.0034730341285467148, + 0.0031725959852337837, + 0.0028055505827069283, + 0.002772795967757702, + 0.0019138571806252003, + 0.003215949749574065, + 0.002620883285999298, + 0.0018576121656224132, + 0.00417747488245368, + 0.0022290521301329136, + 0.002589776646345854, + 0.0028924448415637016, + 0.0021388791501522064, + 0.002261376939713955, + 0.005344206001609564, + 0.010575886815786362, + 0.004397800657898188, + 0.005958776921033859, + 0.004487754311412573, + 0.009183444082736969, + 0.003057429101318121, + 0.005040185991674662, + 0.01509844046086073, + 0.011981208808720112, + 0.010711638256907463, + 0.011277865618467331, + 0.022539040073752403, + 0.011978584341704845, + 0.021785534918308258, + 0.02173495851457119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.02981533110141754 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190231174230576 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.018190231174230576, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190231174230576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14725792407989502, + 0.0403391495347023, + 0.11598751693964005, + 0.4992939531803131, + 0.05676078423857689, + 0.01838139444589615, + 0.011379212141036987, + 0.01886913552880287, + 0.061989594250917435, + 0.029741395264863968 + ], + "credit_map": [ + 0.012965850532054901, + 0.010421006940305233, + 0.0452897734940052, + 0.014921881258487701, + 0.006495037581771612, + 0.03690175339579582, + 0.007507797796279192, + 0.012754809111356735, + 0.007699454203248024, + 0.0037845175247639418, + 0.005054340232163668, + 0.005200792569667101, + 0.004108133260160685, + 0.005917303264141083, + 0.0028808326460421085, + 0.005693772342056036, + 0.014430133625864983, + 0.005523086059838533, + 0.04151394963264465, + 0.008914433419704437, + 0.004922701511532068, + 0.014971127733588219, + 0.00858731847256422, + 0.017124755308032036, + 0.024938540533185005, + 0.007757917512208223, + 0.02665012888610363, + 0.008648731745779514, + 0.0083934860303998, + 0.2024184912443161, + 0.010299090296030045, + 0.21018752455711365, + 0.006101237144321203, + 0.004278253298252821, + 0.014573909342288971, + 0.004843809176236391, + 0.004439721815288067, + 0.008448581211268902, + 0.004732079803943634, + 0.009343190118670464, + 0.0019441661424934864, + 0.0021101157180964947, + 0.0017444138647988439, + 0.0020520638208836317, + 0.0024104390759021044, + 0.0034442690666764975, + 0.002121932338923216, + 0.0025539924390614033, + 0.0016705499729141593, + 0.001167632988654077, + 0.0015376077499240637, + 0.0012689162977039814, + 0.0012970417737960815, + 0.0016288763144984841, + 0.0012256928021088243, + 0.001582892844453454, + 0.0035749957896769047, + 0.0008487777668051422, + 0.005172680132091045, + 0.0010879128240048885, + 0.001651893020607531, + 0.0028719576075673103, + 0.0009794499492272735, + 0.0026814669836312532, + 0.005568898748606443, + 0.005923466756939888, + 0.0031211026944220066, + 0.0033691166900098324, + 0.003811437403783202, + 0.021615421399474144, + 0.002005920745432377, + 0.01657422073185444, + 0.004027461167424917, + 0.0026733747217804193, + 0.0022238895762711763, + 0.002569371135905385, + 0.004293059464544058, + 0.0032333426643162966, + 0.0045251427218317986, + 0.006195751950144768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.018190231174230576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902637660503387 + ], + "opposite_logprob_tokens": [ + -2.46402645111084 + ], + "logprob_reward": -0.08902637660503387, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902637660503387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1584164947271347, + 0.10164572298526764, + 0.13834653794765472, + 0.17146432399749756, + 0.061089228838682175, + 0.04434452950954437, + 0.07785739749670029, + 0.058988507837057114, + 0.10452968627214432, + 0.08331765234470367 + ], + "credit_map": [ + 0.01827409490942955, + 0.01985006406903267, + 0.022958001121878624, + 0.02238832600414753, + 0.015390350483357906, + 0.022643839940428734, + 0.015661226585507393, + 0.02125057950615883, + 0.015797728672623634, + 0.0099437041208148, + 0.010848479345440865, + 0.011912152171134949, + 0.013364359736442566, + 0.0129489004611969, + 0.014030102640390396, + 0.012800285592675209, + 0.019289543852210045, + 0.017478862777352333, + 0.021335074678063393, + 0.0150655098259449, + 0.01907428167760372, + 0.013559728860855103, + 0.019181426614522934, + 0.013362082652747631, + 0.023367421701550484, + 0.015452956780791283, + 0.035043198615312576, + 0.02515404298901558, + 0.01615515723824501, + 0.019640449434518814, + 0.015993757173419, + 0.020657340064644814, + 0.007911249995231628, + 0.008429919369518757, + 0.010098216123878956, + 0.008086027577519417, + 0.009439146146178246, + 0.005618731025606394, + 0.00405782600864768, + 0.007448101881891489, + 0.0038679176941514015, + 0.0051721869967877865, + 0.0051180352456867695, + 0.0046567609533667564, + 0.004633558914065361, + 0.00893380492925644, + 0.005433970130980015, + 0.006528289057314396, + 0.005388892721384764, + 0.016755327582359314, + 0.010115017183125019, + 0.011197770945727825, + 0.012077388353645802, + 0.008028163574635983, + 0.007468772120773792, + 0.006826052442193031, + 0.007554190699011087, + 0.005343538708984852, + 0.011880174279212952, + 0.006504875607788563, + 0.008656774647533894, + 0.007416342850774527, + 0.005263833329081535, + 0.00636876979842782, + 0.021155133843421936, + 0.014730353839695454, + 0.010449589230120182, + 0.01027363259345293, + 0.008619528263807297, + 0.019847918301820755, + 0.005408088210970163, + 0.014045423828065395, + 0.009059188887476921, + 0.00895918719470501, + 0.007508625742048025, + 0.005883426871150732, + 0.011856531724333763, + 0.008507007732987404, + 0.014214898459613323, + 0.017328783869743347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.08902637660503387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking shot, shot from a slightly upward angle, follows General Zhao Yun, clad in a full suit of intricately carved silver armor, as he charges forward unstoppably across the dusty battlefield. The armor gleams with a cold metallic sheen in the dim sunlight, the intricately carved dragons on his breastplate, pauldrons, and gauntlets clearly visible. The scratches and dents of battle bear witness to the brutality of fierce combat. His heroic face is matted with dust and sweat, his eyes sharp and resolute, his lips pursed with a resolute resolve. His long, bound black hair falls loose beneath his helmet as he violently moves. He holds a spear steady in one hand, the tip gleaming with a cold gleam, pointed forward. The red tassel on the spearhead, combined with the flowing white cape behind him, dances wildly in the wind, creating a dramatic arc. His powerful legs stamped against the churning scorched earth, each step stirring up vast clouds of yellow sand, leaving behind a blur of chaotic battlefield filled with the clash of weapons and shouts. The entire image is imbued with a strong sense of movement and epic heroism, with the color palette leaning towards a warm and gritty brown, highlighting the general's unwavering focus and bravery as he charges forward amidst the chaos.", + "seed": 2409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_10/index_4551_seed_2409_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033072861330583694, + "vlm_logprob": -0.033072861330583694, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.033072861330583694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04409007728099823 + ], + "opposite_logprob_tokens": [ + -3.2940900325775146 + ], + "logprob_reward": -0.04409007728099823, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04409007728099823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27286407351493835, + 0.1358678638935089, + 0.21413367986679077, + 0.10054131597280502, + 0.04644578695297241, + 0.07753734290599823, + 0.027566805481910706, + 0.022690435871481895, + 0.05966399610042572, + 0.04268866404891014 + ], + "credit_map": [ + 0.03509135544300079, + 0.034722402691841125, + 0.03634975478053093, + 0.03615712374448776, + 0.03796612098813057, + 0.02372903935611248, + 0.04053569957613945, + 0.028312593698501587, + 0.015533575788140297, + 0.024841386824846268, + 0.012831429950892925, + 0.020161233842372894, + 0.013556340709328651, + 0.015382025390863419, + 0.01544850692152977, + 0.01811336912214756, + 0.01704297959804535, + 0.03671954944729805, + 0.022254901006817818, + 0.033958520740270615, + 0.02751326560974121, + 0.033771272748708725, + 0.018984105437994003, + 0.023889077827334404, + 0.010473117232322693, + 0.011801383458077908, + 0.01216635201126337, + 0.009512241929769516, + 0.024349674582481384, + 0.011980608105659485, + 0.013917414471507072, + 0.006340520456433296, + 0.004418347030878067, + 0.005321283359080553, + 0.004825138486921787, + 0.006127915810793638, + 0.0069413078017532825, + 0.007144747767597437, + 0.00674588093534112, + 0.0049211629666388035, + 0.007363961078226566, + 0.004790882579982281, + 0.01739000529050827, + 0.0076944115571677685, + 0.004255631007254124, + 0.021625323221087456, + 0.005980935879051685, + 0.008436187170445919, + 0.004161720629781485, + 0.0029462063685059547, + 0.004349710885435343, + 0.003637254238128662, + 0.003101291134953499, + 0.0034333053044974804, + 0.0029852990992367268, + 0.002952016657218337, + 0.003511708229780197, + 0.002945138141512871, + 0.003325430676341057, + 0.003479549428448081, + 0.0023145771119743586, + 0.0019339185673743486, + 0.002969565102830529, + 0.002210546750575304, + 0.00786644034087658, + 0.005510853603482246, + 0.00701648835092783, + 0.007242660969495773, + 0.013680391013622284, + 0.0055575501173734665, + 0.009457829408347607, + 0.0033317813649773598, + 0.0059228455647826195, + 0.005448915064334869, + 0.006985832005739212, + 0.004340189974755049, + 0.0070998333394527435, + 0.004937944933772087, + 0.003927492070943117, + 0.004025608766824007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04409007728099823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036351096350699663 + ], + "opposite_logprob_tokens": [ + -5.628634929656982 + ], + "logprob_reward": -0.0036351096350699663, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036351096350699663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21843627095222473, + 0.06321872770786285, + 0.12073814868927002, + 0.060041483491659164, + 0.06909600645303726, + 0.1254955679178238, + 0.08653276413679123, + 0.08373241871595383, + 0.10291632264852524, + 0.06979229301214218 + ], + "credit_map": [ + 0.028188597410917282, + 0.030547022819519043, + 0.026742372661828995, + 0.030739901587367058, + 0.0305718295276165, + 0.018736129626631737, + 0.029869338497519493, + 0.023041071370244026, + 0.007166450843214989, + 0.016652902588248253, + 0.009975358843803406, + 0.008184145204722881, + 0.005925371777266264, + 0.004884498659521341, + 0.005420462694019079, + 0.005009538494050503, + 0.012510607950389385, + 0.021957937628030777, + 0.012836653739213943, + 0.016573715955018997, + 0.015463574789464474, + 0.018955988809466362, + 0.012288405559957027, + 0.010151256807148457, + 0.0069252499379217625, + 0.005414274055510759, + 0.008367079310119152, + 0.006096572615206242, + 0.012023295275866985, + 0.006999420467764139, + 0.009394630789756775, + 0.004820963367819786, + 0.00802136491984129, + 0.00753058260306716, + 0.0062874676659703255, + 0.010280911810696125, + 0.009055223315954208, + 0.009932732209563255, + 0.00888241920620203, + 0.009105302393436432, + 0.006119161378592253, + 0.00416851881891489, + 0.04427167400717735, + 0.005948779638856649, + 0.004114195238798857, + 0.0491221621632576, + 0.0035963000264018774, + 0.008154772222042084, + 0.010214514099061489, + 0.0068589686416089535, + 0.010279099456965923, + 0.007665147073566914, + 0.01756473071873188, + 0.010415096767246723, + 0.01254984736442566, + 0.010985364206135273, + 0.009198062121868134, + 0.015416266396641731, + 0.016308533027768135, + 0.013821490108966827, + 0.007007243111729622, + 0.006079654674977064, + 0.009637884795665741, + 0.006263282615691423, + 0.0125044547021389, + 0.010733195580542088, + 0.01839900016784668, + 0.01264827698469162, + 0.011928620748221874, + 0.011324290186166763, + 0.01688435859978199, + 0.008494127541780472, + 0.012729390524327755, + 0.00650256359949708, + 0.01359627302736044, + 0.009407841600477695, + 0.006902497727423906, + 0.008762558922171593, + 0.006102811079472303, + 0.005788358859717846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0036351096350699663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016083383932709694 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.016083383932709694, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016083383932709694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15559110045433044, + 0.07145211845636368, + 0.3638925552368164, + 0.12991374731063843, + 0.01602325774729252, + 0.07516027241945267, + 0.0754442885518074, + 0.03221232444047928, + 0.044282201677560806, + 0.0360281877219677 + ], + "credit_map": [ + 0.02213512361049652, + 0.021494433283805847, + 0.02481372281908989, + 0.023448629304766655, + 0.01424017921090126, + 0.013348009437322617, + 0.021125737577676773, + 0.014985268004238605, + 0.007094680331647396, + 0.010933246463537216, + 0.007290613371878862, + 0.007599119562655687, + 0.007502274122089148, + 0.013657690957188606, + 0.008741081692278385, + 0.008633410558104515, + 0.036079198122024536, + 0.06718669831752777, + 0.035760704427957535, + 0.035281870514154434, + 0.051603980362415314, + 0.07233703881502151, + 0.03578980639576912, + 0.029853252694010735, + 0.012192207388579845, + 0.01366790197789669, + 0.015345321036875248, + 0.013132498599588871, + 0.03506249561905861, + 0.015552373602986336, + 0.016774628311395645, + 0.008186318911612034, + 0.0017247899668291211, + 0.0016191822942346334, + 0.0014781802892684937, + 0.002153739333152771, + 0.003740631742402911, + 0.0014703253982588649, + 0.0022189693991094828, + 0.001617440371774137, + 0.003435081336647272, + 0.0023411791771650314, + 0.026629794389009476, + 0.0033125795889645815, + 0.002729515079408884, + 0.0299965962767601, + 0.0023657558485865593, + 0.004349765833467245, + 0.008300752379000187, + 0.008073889650404453, + 0.010212555527687073, + 0.009069262072443962, + 0.015520872548222542, + 0.010685677640140057, + 0.0059843421913683414, + 0.007596938405185938, + 0.0035780039615929127, + 0.006734030786901712, + 0.005392646417021751, + 0.0031549374107271433, + 0.003025206970050931, + 0.0030159475281834602, + 0.004204180557280779, + 0.0031073749996721745, + 0.006660424638539553, + 0.005228545982390642, + 0.007272336632013321, + 0.006442112848162651, + 0.004667219705879688, + 0.005072937812656164, + 0.005861375946551561, + 0.0030772460158914328, + 0.005258370190858841, + 0.003694316139444709, + 0.005996988154947758, + 0.004567021504044533, + 0.005211840849369764, + 0.003939506132155657, + 0.0035716709680855274, + 0.0037884691264480352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.016083383932709694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539827264845371 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539827264845371, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539827264845371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18472450971603394, + 0.08123929798603058, + 0.22483398020267487, + 0.052463483065366745, + 0.04647644981741905, + 0.06105296313762665, + 0.09556544572114944, + 0.07369976490736008, + 0.03657102212309837, + 0.1433730572462082 + ], + "credit_map": [ + 0.02680651843547821, + 0.026169976219534874, + 0.021161671727895737, + 0.028091968968510628, + 0.019275451079010963, + 0.018384795635938644, + 0.02017351984977722, + 0.02466060034930706, + 0.009355533868074417, + 0.00992034561932087, + 0.012150418013334274, + 0.011311273090541363, + 0.009099557995796204, + 0.009782872162759304, + 0.009006068110466003, + 0.010613231919705868, + 0.028001094236969948, + 0.040137603878974915, + 0.019654082134366035, + 0.02018052339553833, + 0.02996828220784664, + 0.04581387713551521, + 0.0237936582416296, + 0.01728484407067299, + 0.006084497086703777, + 0.0046310825273394585, + 0.007635252550244331, + 0.0073517970740795135, + 0.008254881016910076, + 0.008754745125770569, + 0.005663690157234669, + 0.0040875389240682125, + 0.005923104472458363, + 0.005087102763354778, + 0.005529905669391155, + 0.005175572820007801, + 0.005778211168944836, + 0.006756865419447422, + 0.007266475819051266, + 0.004959210753440857, + 0.007911557331681252, + 0.005044236779212952, + 0.012438124977052212, + 0.0058471751399338245, + 0.00423670606687665, + 0.01716766133904457, + 0.004305184818804264, + 0.004102314356714487, + 0.010793948546051979, + 0.0084729865193367, + 0.011005178093910217, + 0.0134131433442235, + 0.014384288340806961, + 0.011329935863614082, + 0.011633694171905518, + 0.01453226525336504, + 0.015066955238580704, + 0.012138438411056995, + 0.010440107434988022, + 0.005670969374477863, + 0.006740314420312643, + 0.007791837211698294, + 0.008820667862892151, + 0.007030472159385681, + 0.00482060294598341, + 0.005719815846532583, + 0.005574549548327923, + 0.00446654949337244, + 0.0046586692333221436, + 0.004608834162354469, + 0.0037712620105594397, + 0.0029507381841540337, + 0.023246435448527336, + 0.016659148037433624, + 0.02625114470720291, + 0.014966821298003197, + 0.017339035868644714, + 0.01952558383345604, + 0.012534837238490582, + 0.01285005547106266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.012539827264845371 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character identified as General Zhao Yun wearing a full suit of intricately carved silver armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character performing a charging motion forward across a dusty battlefield while stirring up clouds of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a spear with a red tassel on the tip, pointed forward in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a horizontal tracking shot from a slightly upward angle following the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the character's long black hair visible falling loose beneath his helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890159085392952 + ], + "opposite_logprob_tokens": [ + -2.4640159606933594 + ], + "logprob_reward": -0.0890159085392952, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890159085392952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08462110161781311, + 0.04084910452365875, + 0.14647388458251953, + 0.047587085515260696, + 0.05014761909842491, + 0.3723613917827606, + 0.06043142452836037, + 0.041432879865169525, + 0.03118613176047802, + 0.12490930408239365 + ], + "credit_map": [ + 0.01194235123693943, + 0.010901990346610546, + 0.011702186428010464, + 0.01314051728695631, + 0.009247085079550743, + 0.007790674921125174, + 0.010048775933682919, + 0.009847529232501984, + 0.004513418767601252, + 0.00438733072951436, + 0.004431162029504776, + 0.004286978393793106, + 0.0050541008822619915, + 0.007445194758474827, + 0.004425336141139269, + 0.006305585615336895, + 0.016449779272079468, + 0.02002742700278759, + 0.014490276575088501, + 0.017935605719685555, + 0.014776668511331081, + 0.025326889008283615, + 0.017379876226186752, + 0.020087379962205887, + 0.006157648749649525, + 0.005203266628086567, + 0.005681266542524099, + 0.0045573897659778595, + 0.011693616397678852, + 0.005350181367248297, + 0.00599578907713294, + 0.0029479293152689934, + 0.004377949517220259, + 0.00567223597317934, + 0.004452758468687534, + 0.00887820590287447, + 0.008428698405623436, + 0.0070426748134195805, + 0.007506667170673609, + 0.0037884358316659927, + 0.015049837529659271, + 0.007456094026565552, + 0.14141809940338135, + 0.014441479928791523, + 0.008711956441402435, + 0.1586732119321823, + 0.007899913936853409, + 0.01871083304286003, + 0.006738404743373394, + 0.006054737605154514, + 0.00727469427511096, + 0.006724901031702757, + 0.010768067091703415, + 0.00836208276450634, + 0.009187587536871433, + 0.005320955067873001, + 0.005922087002545595, + 0.005183975212275982, + 0.005165490787476301, + 0.004886296112090349, + 0.00505019398406148, + 0.004561369773000479, + 0.0065187555737793446, + 0.004144715610891581, + 0.0038517722859978676, + 0.0035601980052888393, + 0.005224776919931173, + 0.0034584642853587866, + 0.004710494074970484, + 0.0037974652368575335, + 0.0038869609124958515, + 0.002696002833545208, + 0.012353465892374516, + 0.018855229020118713, + 0.0189578365534544, + 0.01346862968057394, + 0.021888764575123787, + 0.016873765736818314, + 0.010248596780002117, + 0.012263028882443905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.0890159085392952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_11/index_tensor([5799, 5799])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_11/index_tensor([5799, 5799])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..60642b448777cd89b817c81fc9c5fc9d44d2f490 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_11/index_tensor([5799, 5799])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking fisheye lens close-up, shot from a low, upward angle, the camera remains stationary, completely capturing the viewer's gaze. A bony, fair-skinned male hand, steady and powerful, grips a transparent, modern glass. The glass is filled with a distinctly layered liquid, gradating from a deep purple at the bottom to a neon pink in the middle, and finally to a bright cyan at the top. Tiny bubbles linger in the liquid, shimmering with a strange luster. The background blurs into a profound darkness, with only a few blurred, color-distorted streaks of light distorting at the edges. As the hand raises the glass to his lips, the figure's taut jawline and the prominent Adam's apple at the bottom of the frame come into sharp focus. As he tilts his head back to drink, the apple begins to roll violently and rapidly, each movement imbued with power. At the highest point of the Adam's apple rolling, the entire picture was suddenly torn apart by horizontal and vertical digital noise and pixel blocks, the colors shifted drastically, the bright liquid colors were intertwined with the dazzling white light and the misplaced RGB channels, forming a visual instant collapse and explosion, creating a tense, psychedelic and strong atmosphere full of the aesthetics of collapse in the digital age.", + "seed": 31075, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_11/index_5799_seed_31075_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10732413735240698, + "vlm_logprob": -0.10732413735240698, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.10732413735240698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033045750111341476 + ], + "opposite_logprob_tokens": [ + -3.533045768737793 + ], + "logprob_reward": -0.033045750111341476, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.033045750111341476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09278249740600586, + 0.0826897993683815, + 0.1014980599284172, + 0.09974809736013412, + 0.03630917891860008, + 0.3652990758419037, + 0.10066155344247818, + 0.07180284708738327, + 0.01991286315023899, + 0.029296092689037323 + ], + "credit_map": [ + 0.016368040814995766, + 0.020465455949306488, + 0.016542721539735794, + 0.015670377761125565, + 0.008599335327744484, + 0.00493073184043169, + 0.004474985413253307, + 0.005730840377509594, + 0.01342415064573288, + 0.009791115298867226, + 0.01251369621604681, + 0.009344294667243958, + 0.008058435283601284, + 0.008044708520174026, + 0.013293450698256493, + 0.008219935931265354, + 0.01964353211224079, + 0.021884392946958542, + 0.01494477316737175, + 0.01879219338297844, + 0.005528502631932497, + 0.003362700343132019, + 0.011417219415307045, + 0.005924741737544537, + 0.017461810261011124, + 0.021593216806650162, + 0.01307723019272089, + 0.010605632327497005, + 0.010557969100773335, + 0.006140769924968481, + 0.012417064048349857, + 0.007894393056631088, + 0.004016251303255558, + 0.0035154863726347685, + 0.00513275945559144, + 0.003140845336019993, + 0.002766462741419673, + 0.004552207421511412, + 0.005415074992924929, + 0.00777008943259716, + 0.03260498121380806, + 0.0413711778819561, + 0.029216891154646873, + 0.017868896946310997, + 0.018220240250229836, + 0.016378222033381462, + 0.09423442929983139, + 0.11540421098470688, + 0.010825123637914658, + 0.008703569881618023, + 0.01238613948225975, + 0.0074252774938941, + 0.0076520973816514015, + 0.00938020832836628, + 0.021274028345942497, + 0.023015107959508896, + 0.008784892968833447, + 0.00405164947733283, + 0.005387857556343079, + 0.0038473622407764196, + 0.004744089674204588, + 0.0030257035978138447, + 0.018653521314263344, + 0.02330777235329151, + 0.002540070563554764, + 0.0030434769578278065, + 0.0027381079271435738, + 0.0029477730859071016, + 0.0017581675201654434, + 0.0018655534368008375, + 0.002700947457924485, + 0.0023187638726085424, + 0.004959974437952042, + 0.005233602598309517, + 0.004548665136098862, + 0.004298301879316568, + 0.0018502498278394341, + 0.0038704725448042154, + 0.0014638947322964668, + 0.0030709283892065287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.033045750111341476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371654078364372 + ], + "opposite_logprob_tokens": [ + -3.4087164402008057 + ], + "logprob_reward": -0.03371654078364372, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371654078364372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08638112246990204, + 0.12646561861038208, + 0.04101058095693588, + 0.10292636603116989, + 0.16529271006584167, + 0.08150584995746613, + 0.12187552452087402, + 0.1490572690963745, + 0.028611144050955772, + 0.09687379002571106 + ], + "credit_map": [ + 0.013817574828863144, + 0.013853144831955433, + 0.014567487873136997, + 0.013351304456591606, + 0.008301839232444763, + 0.007908408530056477, + 0.0065688518807291985, + 0.00801251083612442, + 0.017260262742638588, + 0.01888127066195011, + 0.017906585708260536, + 0.021735765039920807, + 0.010731787420809269, + 0.010786684229969978, + 0.016198180615901947, + 0.012965094298124313, + 0.005901318043470383, + 0.004455682821571827, + 0.0051778098568320274, + 0.0040780967101454735, + 0.005119973793625832, + 0.005523436237126589, + 0.005456941667944193, + 0.005297320894896984, + 0.014616550877690315, + 0.015123230405151844, + 0.013946167193353176, + 0.010104408487677574, + 0.014116759411990643, + 0.00981186330318451, + 0.011201193556189537, + 0.014006199315190315, + 0.017208877950906754, + 0.01677733287215233, + 0.03817471116781235, + 0.010213634930551052, + 0.01274834293872118, + 0.022794105112552643, + 0.030084306374192238, + 0.017291411757469177, + 0.00860687717795372, + 0.009998949244618416, + 0.008661859668791294, + 0.005676766857504845, + 0.014719161204993725, + 0.007033254019916058, + 0.012770338915288448, + 0.014038649387657642, + 0.009447461925446987, + 0.0076792421750724316, + 0.011054908856749535, + 0.006133018992841244, + 0.006543423980474472, + 0.005975036881864071, + 0.04127907007932663, + 0.033763375133275986, + 0.008235711604356766, + 0.00570835592225194, + 0.009332956746220589, + 0.006334660574793816, + 0.006442173384130001, + 0.0060656010173261166, + 0.05237073451280594, + 0.054567087441682816, + 0.004963472485542297, + 0.0028144291136413813, + 0.004016890190541744, + 0.005069378763437271, + 0.003020322183147073, + 0.0020434001926332712, + 0.0035414339508861303, + 0.003141819266602397, + 0.0102378623560071, + 0.01744672656059265, + 0.007638586685061455, + 0.010196438990533352, + 0.007337570656090975, + 0.013937841169536114, + 0.00588812492787838, + 0.0241906326264143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.03371654078364372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097646087408066 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097646087408066, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097646087408066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13839368522167206, + 0.07118789106607437, + 0.10401210933923721, + 0.3138808310031891, + 0.07916691899299622, + 0.06303409487009048, + 0.05480219051241875, + 0.11592978984117508, + 0.03283890709280968, + 0.026753531768918037 + ], + "credit_map": [ + 0.020366964861750603, + 0.03193211555480957, + 0.025005806237459183, + 0.02855646423995495, + 0.012950073927640915, + 0.005616862792521715, + 0.008617889136075974, + 0.0053475224412977695, + 0.011216184124350548, + 0.01154125202447176, + 0.010799482464790344, + 0.008905141614377499, + 0.0064577446319162846, + 0.0053317854180932045, + 0.010893026366829872, + 0.006043281406164169, + 0.022934162989258766, + 0.017979130148887634, + 0.018240144476294518, + 0.017248841002583504, + 0.004685318097472191, + 0.003558454103767872, + 0.01242784783244133, + 0.006938219536095858, + 0.057003311812877655, + 0.081309974193573, + 0.04307932034134865, + 0.02748146466910839, + 0.030816806480288506, + 0.016210919246077538, + 0.037920426577329636, + 0.02005864307284355, + 0.00796041265130043, + 0.006640574894845486, + 0.013728199526667595, + 0.008458348922431469, + 0.005241421051323414, + 0.005540806800127029, + 0.014133816584944725, + 0.01746334880590439, + 0.005202018190175295, + 0.00661417655646801, + 0.0060669188387691975, + 0.004804073832929134, + 0.006378634832799435, + 0.003598567796871066, + 0.0153766805306077, + 0.014993038028478622, + 0.0073825703002512455, + 0.006723059806972742, + 0.00962285790592432, + 0.004260017070919275, + 0.0037355702370405197, + 0.0035618962720036507, + 0.009929344989359379, + 0.009586884640157223, + 0.010158024728298187, + 0.004607167560607195, + 0.007274504750967026, + 0.002628563204780221, + 0.004363570362329483, + 0.0031311821658164263, + 0.03822976350784302, + 0.04553702846169472, + 0.0044247787445783615, + 0.004942185245454311, + 0.004371962510049343, + 0.0038018240593373775, + 0.0035900757648050785, + 0.0033442992717027664, + 0.004360625986009836, + 0.004003160633146763, + 0.003809628076851368, + 0.004421234596520662, + 0.0031225408893078566, + 0.004345778375864029, + 0.0021866809111088514, + 0.004168255720287561, + 0.001949930563569069, + 0.0027494856622070074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.038097646087408066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289435148239136 + ], + "opposite_logprob_tokens": [ + -1.0539435148239136 + ], + "logprob_reward": -0.4289435148239136, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289435148239136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12469211220741272, + 0.044479914009571075, + 0.06178078427910805, + 0.10451050847768784, + 0.1624123454093933, + 0.13957752287387848, + 0.1426430642604828, + 0.10910917818546295, + 0.061361439526081085, + 0.04943321645259857 + ], + "credit_map": [ + 0.019772617146372795, + 0.027371782809495926, + 0.018450124189257622, + 0.02244548313319683, + 0.012604585848748684, + 0.00678210286423564, + 0.008410482667386532, + 0.00885492842644453, + 0.009456584230065346, + 0.006909102667123079, + 0.008307469077408314, + 0.006105718202888966, + 0.003489263355731964, + 0.002476682187989354, + 0.0035067577846348286, + 0.004228333942592144, + 0.010605293326079845, + 0.01499791257083416, + 0.007992925122380257, + 0.007132307160645723, + 0.005003463011234999, + 0.003952931147068739, + 0.006313679274171591, + 0.005782267544418573, + 0.015660269185900688, + 0.017247522249817848, + 0.01417152862995863, + 0.012062576599419117, + 0.00963110476732254, + 0.008799589239060879, + 0.014476624317467213, + 0.012461286038160324, + 0.014799177646636963, + 0.014352750033140182, + 0.042422208935022354, + 0.010098373517394066, + 0.010465607978403568, + 0.018748529255390167, + 0.03109149821102619, + 0.020434172824025154, + 0.017199411988258362, + 0.02318308688700199, + 0.02172235958278179, + 0.010645853355526924, + 0.020638395100831985, + 0.009244341403245926, + 0.018832441419363022, + 0.01811162382364273, + 0.008582841604948044, + 0.01391849759966135, + 0.011773696169257164, + 0.012787201441824436, + 0.008127682842314243, + 0.0065605915151536465, + 0.042730290442705154, + 0.03816225379705429, + 0.013151547871530056, + 0.005752888508141041, + 0.008209836669266224, + 0.003361339448019862, + 0.005044966470450163, + 0.0037324607837945223, + 0.03076687641441822, + 0.03908924013376236, + 0.009778947569429874, + 0.008666133508086205, + 0.008770559914410114, + 0.007948591373860836, + 0.004711127374321222, + 0.007312048692256212, + 0.008256166242063046, + 0.00591785553842783, + 0.005989451427012682, + 0.008797334507107735, + 0.0038486807607114315, + 0.005939226597547531, + 0.003814882831647992, + 0.00843161903321743, + 0.0030628303065896034, + 0.009549185633659363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.4289435148239136 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817234955728054 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817234955728054, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817234955728054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0694141834974289, + 0.035110462456941605, + 0.028534259647130966, + 0.07268927991390228, + 0.05050373822450638, + 0.17391258478164673, + 0.1165628582239151, + 0.26085174083709717, + 0.05331188812851906, + 0.13910911977291107 + ], + "credit_map": [ + 0.01034647785127163, + 0.014470712281763554, + 0.01065894030034542, + 0.01392173022031784, + 0.005694721359759569, + 0.004552169237285852, + 0.005102161783725023, + 0.004667268600314856, + 0.0044440762139856815, + 0.004252146929502487, + 0.004557418636977673, + 0.006018621847033501, + 0.003602843265980482, + 0.003545737126842141, + 0.0051292069256305695, + 0.0035604112781584263, + 0.005858645308762789, + 0.005666018929332495, + 0.004218958783894777, + 0.002500898903235793, + 0.0023896056227385998, + 0.0019180120434612036, + 0.0034787943586707115, + 0.002503324067220092, + 0.010258493013679981, + 0.015525070019066334, + 0.01008988544344902, + 0.007795419543981552, + 0.006062348373234272, + 0.005278187338262796, + 0.010718090459704399, + 0.006961774546653032, + 0.004622485488653183, + 0.005512303672730923, + 0.006109335925430059, + 0.0046824621967971325, + 0.0034730772022157907, + 0.0035343540366739035, + 0.00923529826104641, + 0.013334415853023529, + 0.010938923805952072, + 0.01083348598331213, + 0.01236781943589449, + 0.007121649105101824, + 0.006266510114073753, + 0.004213034175336361, + 0.056928958743810654, + 0.06524217128753662, + 0.008650379255414009, + 0.012236304581165314, + 0.012983331456780434, + 0.01480154413729906, + 0.008664911612868309, + 0.007314298301935196, + 0.02748882584273815, + 0.02442324534058571, + 0.011937670409679413, + 0.010221357457339764, + 0.015614297240972519, + 0.009686946868896484, + 0.010738843120634556, + 0.007966380566358566, + 0.09086252003908157, + 0.1038237139582634, + 0.005947883240878582, + 0.005769030191004276, + 0.007001189049333334, + 0.007853865623474121, + 0.005651839543133974, + 0.0047494107857346535, + 0.007552238181233406, + 0.008786424994468689, + 0.025535158812999725, + 0.02344844490289688, + 0.02112964354455471, + 0.020030807703733444, + 0.012945050373673439, + 0.013114621862769127, + 0.007370804902166128, + 0.015534568578004837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.002817234955728054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking fisheye lens close-up, shot from a low, upward angle, the camera remains stationary, completely capturing the viewer's gaze. A bony, fair-skinned male hand, steady and powerful, grips a transparent, modern glass. The glass is filled with a distinctly layered liquid, gradating from a deep purple at the bottom to a neon pink in the middle, and finally to a bright cyan at the top. Tiny bubbles linger in the liquid, shimmering with a strange luster. The background blurs into a profound darkness, with only a few blurred, color-distorted streaks of light distorting at the edges. As the hand raises the glass to his lips, the figure's taut jawline and the prominent Adam's apple at the bottom of the frame come into sharp focus. As he tilts his head back to drink, the apple begins to roll violently and rapidly, each movement imbued with power. At the highest point of the Adam's apple rolling, the entire picture was suddenly torn apart by horizontal and vertical digital noise and pixel blocks, the colors shifted drastically, the bright liquid colors were intertwined with the dazzling white light and the misplaced RGB channels, forming a visual instant collapse and explosion, creating a tense, psychedelic and strong atmosphere full of the aesthetics of collapse in the digital age.", + "seed": 31076, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_11/index_5799_seed_31076_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08108714167028666, + "vlm_logprob": -0.08108714167028666, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.08108714167028666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013411309570074081 + ], + "opposite_logprob_tokens": [ + -4.638411521911621 + ], + "logprob_reward": -0.013411309570074081, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013411309570074081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40582025051116943, + 0.026682337746024132, + 0.021284928545355797, + 0.03375615179538727, + 0.025326771661639214, + 0.040110643953084946, + 0.34377628564834595, + 0.04181576520204544, + 0.031031183898448944, + 0.030395716428756714 + ], + "credit_map": [ + 0.10339374840259552, + 0.019760562106966972, + 0.01350520458072424, + 0.019862612709403038, + 0.03827638551592827, + 0.05302627384662628, + 0.019870832562446594, + 0.13812461495399475, + 0.0021043659653514624, + 0.0019109846325591207, + 0.0041952007450163364, + 0.0044534397311508656, + 0.002295165788382292, + 0.005757512059062719, + 0.0026632247027009726, + 0.0033024440053850412, + 0.0025259475223720074, + 0.0022562602534890175, + 0.003055074019357562, + 0.003741280408576131, + 0.002350233029574156, + 0.002849331358447671, + 0.0020770810078829527, + 0.0024297211784869432, + 0.0026803696528077126, + 0.0025532920844852924, + 0.0031705733854323626, + 0.003396035637706518, + 0.0030533387325704098, + 0.00461361650377512, + 0.008788091130554676, + 0.005500833969563246, + 0.0018242441583424807, + 0.0034144248347729445, + 0.0028493390418589115, + 0.004372288007289171, + 0.0025822825264185667, + 0.0030539603903889656, + 0.004072637762874365, + 0.003157595172524452, + 0.0045561883598566055, + 0.003566593863070011, + 0.0026410669088363647, + 0.0064161610789597034, + 0.0036038781981915236, + 0.006191176362335682, + 0.00859842449426651, + 0.004537154920399189, + 0.017770571634173393, + 0.012583121657371521, + 0.01250909548252821, + 0.13497157394886017, + 0.010375476442277431, + 0.009709575213491917, + 0.13201904296875, + 0.013837818056344986, + 0.0033124336041510105, + 0.004483463242650032, + 0.004393226001411676, + 0.008618599735200405, + 0.004543702583760023, + 0.004909345414489508, + 0.007690172642469406, + 0.003864822443574667, + 0.0032923459075391293, + 0.0028938523028045893, + 0.0023455903865396976, + 0.006776083260774612, + 0.001748753129504621, + 0.0022782182786613703, + 0.006259780377149582, + 0.0054365587420761585, + 0.0031763631850481033, + 0.004016992636024952, + 0.003465045476332307, + 0.005550759844481945, + 0.004467593971639872, + 0.004228741861879826, + 0.0031054262071847916, + 0.0023847934789955616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.013411309570074081 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901455461978912 + ], + "opposite_logprob_tokens": [ + -2.579014539718628 + ], + "logprob_reward": -0.07901455461978912, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901455461978912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4954758584499359, + 0.05000782385468483, + 0.042066335678100586, + 0.06906615942716599, + 0.037614334374666214, + 0.01899007335305214, + 0.13147488236427307, + 0.03284924849867821, + 0.09588992595672607, + 0.02656533196568489 + ], + "credit_map": [ + 0.15979519486427307, + 0.02744784951210022, + 0.011926488019526005, + 0.013025491498410702, + 0.014082914218306541, + 0.04726274684071541, + 0.018456853926181793, + 0.20347833633422852, + 0.0040682642720639706, + 0.004239351022988558, + 0.00871898140758276, + 0.005419875029474497, + 0.005128889344632626, + 0.010452540591359138, + 0.005281371064484119, + 0.006698551122099161, + 0.006405917927622795, + 0.003228334477171302, + 0.005929302889853716, + 0.0050579034723341465, + 0.004563339985907078, + 0.005959127098321915, + 0.005488456226885319, + 0.005433950107544661, + 0.004235096275806427, + 0.007750564254820347, + 0.0061676534824073315, + 0.0096536073833704, + 0.005112709477543831, + 0.013417408801615238, + 0.013282503001391888, + 0.00944661907851696, + 0.0030916251707822084, + 0.005139089189469814, + 0.0038446474354714155, + 0.005912459455430508, + 0.005348070524632931, + 0.005728245712816715, + 0.005205899011343718, + 0.003344297409057617, + 0.0018293092725798488, + 0.004114212933927774, + 0.0013443577336147428, + 0.0023411924485117197, + 0.0019236485240980983, + 0.002292043063789606, + 0.003092193976044655, + 0.002053115051239729, + 0.006676347926259041, + 0.004525637254118919, + 0.0047936150804162025, + 0.052479058504104614, + 0.003959170542657375, + 0.0029864904936403036, + 0.05101694539189339, + 0.005037615541368723, + 0.0020726940128952265, + 0.004013375379145145, + 0.004273708909749985, + 0.00605294993147254, + 0.0018546757055446506, + 0.005808777641505003, + 0.005431498400866985, + 0.003341569099575281, + 0.012022601440548897, + 0.014413092285394669, + 0.007322131656110287, + 0.01953091472387314, + 0.006688770372420549, + 0.00893080048263073, + 0.016411760821938515, + 0.010569849982857704, + 0.003017989918589592, + 0.005238024052232504, + 0.003930012695491314, + 0.002890299540013075, + 0.002765156328678131, + 0.003735495265573263, + 0.0026256474666297436, + 0.0023627064656466246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.07901455461978912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107639167457819 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.01107639167457819, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107639167457819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4490054249763489, + 0.02648243121802807, + 0.014423900283873081, + 0.012137878686189651, + 0.02897648885846138, + 0.05406035855412483, + 0.33538275957107544, + 0.04631616920232773, + 0.021292248740792274, + 0.011922290548682213 + ], + "credit_map": [ + 0.14770187437534332, + 0.019756240770220757, + 0.014188282191753387, + 0.014607792720198631, + 0.037706159055233, + 0.035409729927778244, + 0.013697623275220394, + 0.16593773663043976, + 0.0020474325865507126, + 0.0027762241661548615, + 0.0044796946458518505, + 0.003322351025417447, + 0.0032750668469816446, + 0.0042470768094062805, + 0.00302480929531157, + 0.0033097760751843452, + 0.002266090363264084, + 0.0011766598327085376, + 0.0025008930824697018, + 0.0021614376455545425, + 0.0018819792894646525, + 0.0013939013006165624, + 0.0014339355984702706, + 0.0016090036369860172, + 0.0009174871956929564, + 0.001459170482121408, + 0.0008790498832240701, + 0.001253850874491036, + 0.0010159244993701577, + 0.0021065385080873966, + 0.00302319903858006, + 0.0014826580882072449, + 0.0019404564518481493, + 0.0035970003809779882, + 0.003949007950723171, + 0.00435857567936182, + 0.003514837706461549, + 0.00355419865809381, + 0.004571927711367607, + 0.0034904854837805033, + 0.00574575224891305, + 0.009955691173672676, + 0.0032088838052004576, + 0.007571842055767775, + 0.004080314189195633, + 0.006223773583769798, + 0.010043755173683167, + 0.007230347488075495, + 0.02031865157186985, + 0.012443581596016884, + 0.013100770302116871, + 0.12987983226776123, + 0.010522985830903053, + 0.0081302085891366, + 0.12718762457370758, + 0.01379912719130516, + 0.0029132263734936714, + 0.004077213350683451, + 0.005160504020750523, + 0.008555733598768711, + 0.003054872853681445, + 0.009421133436262608, + 0.00805310532450676, + 0.00508038280531764, + 0.0016494740266352892, + 0.0016516209580004215, + 0.0022839608136564493, + 0.0039461953565478325, + 0.0016198783414438367, + 0.002755330642685294, + 0.00489333039149642, + 0.0024924566969275475, + 0.0018791384063661098, + 0.0017302444903180003, + 0.001809627516195178, + 0.0013375014532357454, + 0.0017036683857440948, + 0.0013883109204471111, + 0.001110218232497573, + 0.0009635809110477567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.01107639167457819 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813609540462494 + ], + "opposite_logprob_tokens": [ + -1.4063609838485718 + ], + "logprob_reward": -0.2813609540462494, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813609540462494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6123954653739929, + 0.07312511652708054, + 0.025634609162807465, + 0.026779599487781525, + 0.029174530878663063, + 0.021190524101257324, + 0.11082248389720917, + 0.04317617043852806, + 0.034319594502449036, + 0.023381849750876427 + ], + "credit_map": [ + 0.21173815429210663, + 0.02419705130159855, + 0.01462388876825571, + 0.01108229998499155, + 0.01171802170574665, + 0.05796388164162636, + 0.019596366211771965, + 0.26147589087486267, + 0.005049377214163542, + 0.005860865581780672, + 0.012991066090762615, + 0.008675186894834042, + 0.006924780085682869, + 0.015367562882602215, + 0.00816628523170948, + 0.010090005584061146, + 0.003776160767301917, + 0.0020901747047901154, + 0.004958627745509148, + 0.004900189582258463, + 0.0024300457444041967, + 0.002726730890572071, + 0.0017106038285419345, + 0.0030420804396271706, + 0.002173139015212655, + 0.002236177446320653, + 0.002734828507527709, + 0.003949772100895643, + 0.0020145573653280735, + 0.005276877898722887, + 0.004094766918569803, + 0.004299487452954054, + 0.0018852005014196038, + 0.003678844543173909, + 0.002598921302706003, + 0.004440174903720617, + 0.0032404311932623386, + 0.006204896606504917, + 0.0045091756619513035, + 0.002616890938952565, + 0.0029575899243354797, + 0.002067711902782321, + 0.0013603824190795422, + 0.003117903834208846, + 0.0017544656293466687, + 0.002017171820625663, + 0.005231948569417, + 0.0026833550073206425, + 0.003292080946266651, + 0.0035182563588023186, + 0.0032596406526863575, + 0.04566076025366783, + 0.0032717359717935324, + 0.003557944903150201, + 0.04404135048389435, + 0.004220734816044569, + 0.0036173437256366014, + 0.004329835996031761, + 0.005278870463371277, + 0.006559803616255522, + 0.003487334121018648, + 0.008828257210552692, + 0.006264916155487299, + 0.004809819161891937, + 0.005571743939071894, + 0.004722436424344778, + 0.0030345043633133173, + 0.0048834593035280704, + 0.00210572499781847, + 0.003041383810341358, + 0.006974810268729925, + 0.003985537216067314, + 0.002888720016926527, + 0.0032459315843880177, + 0.00231373799033463, + 0.003732029115781188, + 0.002762001706287265, + 0.0027227315586060286, + 0.0034418595023453236, + 0.002274842234328389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.2813609540462494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572498440742493 + ], + "opposite_logprob_tokens": [ + -3.8955724239349365 + ], + "logprob_reward": -0.020572498440742493, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572498440742493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2679348886013031, + 0.02616574428975582, + 0.044282328337430954, + 0.03254785016179085, + 0.034118786454200745, + 0.06468471139669418, + 0.17273420095443726, + 0.03164105489850044, + 0.1725294142961502, + 0.15336103737354279 + ], + "credit_map": [ + 0.04056048020720482, + 0.026013949885964394, + 0.012712760828435421, + 0.016818251460790634, + 0.04288671538233757, + 0.04587133973836899, + 0.021421726793050766, + 0.06164965033531189, + 0.002352304058149457, + 0.002066054381430149, + 0.005096046254038811, + 0.0036977638956159353, + 0.0034507422242313623, + 0.0033790634479373693, + 0.0031646310817450285, + 0.002959140343591571, + 0.0030485461466014385, + 0.0029869761783629656, + 0.007575031835585833, + 0.007887267507612705, + 0.006123258266597986, + 0.005701619200408459, + 0.005849748384207487, + 0.005109880119562149, + 0.0031552372965961695, + 0.0030531277880072594, + 0.004259608685970306, + 0.004410277586430311, + 0.0027857383247464895, + 0.004620003048330545, + 0.005589793436229229, + 0.004674064926803112, + 0.0022821410093456507, + 0.0033112710807472467, + 0.00503931799903512, + 0.0052750976756215096, + 0.003776989644393325, + 0.006592853926122189, + 0.0050071366131305695, + 0.002833981066942215, + 0.006928786635398865, + 0.011689613573253155, + 0.00380482105538249, + 0.00894362572580576, + 0.005529849324375391, + 0.010754076763987541, + 0.009200848639011383, + 0.007833092473447323, + 0.016399897634983063, + 0.008763925172388554, + 0.009022258222103119, + 0.06291905045509338, + 0.004788463935256004, + 0.006797282490879297, + 0.057524342089891434, + 0.006518959067761898, + 0.0023254717234522104, + 0.0028224859852343798, + 0.003062845440581441, + 0.006012934260070324, + 0.0031894289422780275, + 0.004440606571733952, + 0.006161874160170555, + 0.003625406650826335, + 0.0129069983959198, + 0.019038453698158264, + 0.014059728942811489, + 0.039253171533346176, + 0.011409384198486805, + 0.01623583398759365, + 0.038510121405124664, + 0.021115731447935104, + 0.02074747532606125, + 0.024297406896948814, + 0.023056356236338615, + 0.014102903194725513, + 0.023915953934192657, + 0.015573538839817047, + 0.014998120255768299, + 0.016669301316142082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.020572498440742493 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking fisheye lens close-up, shot from a low, upward angle, the camera remains stationary, completely capturing the viewer's gaze. A bony, fair-skinned male hand, steady and powerful, grips a transparent, modern glass. The glass is filled with a distinctly layered liquid, gradating from a deep purple at the bottom to a neon pink in the middle, and finally to a bright cyan at the top. Tiny bubbles linger in the liquid, shimmering with a strange luster. The background blurs into a profound darkness, with only a few blurred, color-distorted streaks of light distorting at the edges. As the hand raises the glass to his lips, the figure's taut jawline and the prominent Adam's apple at the bottom of the frame come into sharp focus. As he tilts his head back to drink, the apple begins to roll violently and rapidly, each movement imbued with power. At the highest point of the Adam's apple rolling, the entire picture was suddenly torn apart by horizontal and vertical digital noise and pixel blocks, the colors shifted drastically, the bright liquid colors were intertwined with the dazzling white light and the misplaced RGB channels, forming a visual instant collapse and explosion, creating a tense, psychedelic and strong atmosphere full of the aesthetics of collapse in the digital age.", + "seed": 31077, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_11/index_5799_seed_31077_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08717964496463537, + "vlm_logprob": -0.08717964496463537, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.08717964496463537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03264280781149864 + ], + "opposite_logprob_tokens": [ + -3.5326428413391113 + ], + "logprob_reward": -0.03264280781149864, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03264280781149864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26387646794319153, + 0.10653633624315262, + 0.04778597131371498, + 0.02687108889222145, + 0.0584382563829422, + 0.024103056639432907, + 0.04721058905124664, + 0.024217821657657623, + 0.3452720642089844, + 0.05568838492035866 + ], + "credit_map": [ + 0.028113754466176033, + 0.03229720890522003, + 0.03817302733659744, + 0.02813347987830639, + 0.0569000318646431, + 0.02214931510388851, + 0.026047654449939728, + 0.03206198662519455, + 0.010384004563093185, + 0.010970023460686207, + 0.019549883902072906, + 0.01752166450023651, + 0.015288924798369408, + 0.007258330937474966, + 0.01267269067466259, + 0.01289080735296011, + 0.0096186688169837, + 0.005262742750346661, + 0.005379537586122751, + 0.004428309854120016, + 0.0066232020035386086, + 0.004306663293391466, + 0.00684995437040925, + 0.005316890310496092, + 0.002696811920031905, + 0.003004593774676323, + 0.0021266036201268435, + 0.004585620481520891, + 0.00465727224946022, + 0.0029975862707942724, + 0.004186831880360842, + 0.0026157679967582226, + 0.00775027833878994, + 0.008960713632404804, + 0.005342807620763779, + 0.007420372683554888, + 0.008714666590094566, + 0.005444927141070366, + 0.00672058155760169, + 0.008083907887339592, + 0.0024515388067811728, + 0.0043371147476136684, + 0.002703168196603656, + 0.004363151267170906, + 0.0029789835680276155, + 0.002515628933906555, + 0.0026667106430977583, + 0.0020867602434009314, + 0.006817016284912825, + 0.0054211993701756, + 0.00308211799710989, + 0.006690237205475569, + 0.010546023957431316, + 0.004252395126968622, + 0.005906060338020325, + 0.004495540168136358, + 0.001997732324525714, + 0.002756768139079213, + 0.002893335185945034, + 0.0037569585256278515, + 0.0032307871151715517, + 0.0021225647069513798, + 0.002983194077387452, + 0.004476482514292002, + 0.025328686460852623, + 0.012271055951714516, + 0.02753446437418461, + 0.026162339374423027, + 0.030723480507731438, + 0.13226275146007538, + 0.07628685981035233, + 0.014702429063618183, + 0.005462345201522112, + 0.01599767804145813, + 0.004207991063594818, + 0.005007868632674217, + 0.004439221229404211, + 0.014921937137842178, + 0.0030205748043954372, + 0.002630770904943347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.03264280781149864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867919161915779 + ], + "opposite_logprob_tokens": [ + -3.0486791133880615 + ], + "logprob_reward": -0.04867919161915779, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867919161915779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09635085612535477, + 0.02590274065732956, + 0.032063018530607224, + 0.02381162904202938, + 0.015314335934817791, + 0.041294049471616745, + 0.020720116794109344, + 0.011677399277687073, + 0.44243675470352173, + 0.290429025888443 + ], + "credit_map": [ + 0.01059794332832098, + 0.01387815922498703, + 0.01452463585883379, + 0.009727728553116322, + 0.01876308023929596, + 0.008753081783652306, + 0.008897862397134304, + 0.011208371259272099, + 0.0035601365379989147, + 0.0028506373055279255, + 0.0036309100687503815, + 0.0035830019041895866, + 0.003031681990250945, + 0.0021970306988805532, + 0.003354522166773677, + 0.003694819752126932, + 0.005451418925076723, + 0.0031883753836154938, + 0.0035653922241181135, + 0.0034054992720484734, + 0.005661443341523409, + 0.0024815986398607492, + 0.005182940047234297, + 0.003126350697129965, + 0.002687676576897502, + 0.003253673203289509, + 0.0036098812706768513, + 0.003259543562307954, + 0.0037053569685667753, + 0.002004558453336358, + 0.0031766111496835947, + 0.0021143287885934114, + 0.0015830503543838859, + 0.0018048365600407124, + 0.001889161067083478, + 0.001666405936703086, + 0.002084295731037855, + 0.0029723020270466805, + 0.0017324769869446754, + 0.0015818073879927397, + 0.005311564542353153, + 0.006333936471492052, + 0.00511695072054863, + 0.00690267700701952, + 0.005344895180314779, + 0.004209824837744236, + 0.004340402316302061, + 0.0037338011898100376, + 0.0026044140104204416, + 0.0024770700838416815, + 0.0013517007464542985, + 0.003369358368217945, + 0.004010295961052179, + 0.0029574173968285322, + 0.0024166216608136892, + 0.0015332379844039679, + 0.0012705529807135463, + 0.001739606144838035, + 0.0011356215691193938, + 0.0021581577602773905, + 0.0019141648663207889, + 0.000938637531362474, + 0.0014490344328805804, + 0.0010716244578361511, + 0.03218097984790802, + 0.015860632061958313, + 0.03474017605185509, + 0.03430333733558655, + 0.03825884684920311, + 0.16879194974899292, + 0.09956292808055878, + 0.018737925216555595, + 0.028427019715309143, + 0.1089116632938385, + 0.013558885082602501, + 0.0242274422198534, + 0.013455938547849655, + 0.0838170275092125, + 0.007786386180669069, + 0.010244655422866344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.04867919161915779 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001670449972153 + ], + "opposite_logprob_tokens": [ + -2.695016622543335 + ], + "logprob_reward": -0.07001670449972153, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001670449972153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2824638783931732, + 0.04849624261260033, + 0.017868783324956894, + 0.018339822068810463, + 0.02304999716579914, + 0.028659656643867493, + 0.012803188525140285, + 0.030650608241558075, + 0.4786983132362366, + 0.058969512581825256 + ], + "credit_map": [ + 0.029987966641783714, + 0.03449385613203049, + 0.03197095915675163, + 0.024108845740556717, + 0.0748358741402626, + 0.02438700757920742, + 0.028152981773018837, + 0.034526366740465164, + 0.005465375259518623, + 0.004906910005956888, + 0.009141803719103336, + 0.00843263790011406, + 0.0056105731055140495, + 0.003516940865665674, + 0.005797159392386675, + 0.005624843295663595, + 0.0028875130228698254, + 0.0022337634582072496, + 0.0025720917619764805, + 0.002354369265958667, + 0.00258607673458755, + 0.0013363526668399572, + 0.0024618422612547874, + 0.0014367752010002732, + 0.0020939467940479517, + 0.002515499945729971, + 0.0016693308716639876, + 0.002015575533732772, + 0.004138929303735495, + 0.0015359247336164117, + 0.0026714273262768984, + 0.0016991886077448726, + 0.0027348038274794817, + 0.001993791200220585, + 0.0025483325589448214, + 0.0027349989395588636, + 0.001996670151129365, + 0.005808863788843155, + 0.002727189799770713, + 0.002505346667021513, + 0.0037762343417853117, + 0.003963220864534378, + 0.0037494138814508915, + 0.006517453584820032, + 0.0027347225695848465, + 0.0025608125142753124, + 0.0033947499468922615, + 0.0019630491733551025, + 0.0023785075172781944, + 0.00125692889560014, + 0.0009730638703331351, + 0.002351087052375078, + 0.002044625813141465, + 0.0013245255686342716, + 0.0015325628919526935, + 0.000941887847147882, + 0.002932744100689888, + 0.003515461925417185, + 0.003458175575360656, + 0.004339548293501139, + 0.004994722083210945, + 0.0025826608762145042, + 0.0035747895017266273, + 0.005252505652606487, + 0.03471893072128296, + 0.017422091215848923, + 0.037324950098991394, + 0.03736979514360428, + 0.04093481972813606, + 0.1815817505121231, + 0.10903066396713257, + 0.020315323024988174, + 0.005091140046715736, + 0.01794547215104103, + 0.0045745233073830605, + 0.004911152645945549, + 0.004428710788488388, + 0.0159775298088789, + 0.003349201288074255, + 0.0026917834766209126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.07001670449972153 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28136956691741943 + ], + "opposite_logprob_tokens": [ + -1.4063695669174194 + ], + "logprob_reward": -0.28136956691741943, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28136956691741943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0600924976170063, + 0.022017203271389008, + 0.011015270836651325, + 0.028800128027796745, + 0.031117944046854973, + 0.036499276757240295, + 0.036840517073869705, + 0.013144991360604763, + 0.6697217226028442, + 0.09075044095516205 + ], + "credit_map": [ + 0.008030259981751442, + 0.011667287908494473, + 0.006763141136616468, + 0.006171019282191992, + 0.0073366230353713036, + 0.0047220876440405846, + 0.008149068802595139, + 0.007253013085573912, + 0.0024033798836171627, + 0.0020299057941883802, + 0.004205252975225449, + 0.0031125929672271013, + 0.002604709705337882, + 0.0015627738321200013, + 0.002648161957040429, + 0.0034504265058785677, + 0.0011029121233150363, + 0.0019003946799784899, + 0.0013372539542615414, + 0.0016110559226945043, + 0.0016289101913571358, + 0.0007590888999402523, + 0.0014419456711038947, + 0.001233707764185965, + 0.00386393116787076, + 0.0044251796789467335, + 0.0040280562825500965, + 0.002788785845041275, + 0.0038232675287872553, + 0.002591280732303858, + 0.004297977313399315, + 0.0029816480819135904, + 0.004679925739765167, + 0.004855223000049591, + 0.0030234726145863533, + 0.003555938834324479, + 0.004224142525345087, + 0.004327022470533848, + 0.0036965892650187016, + 0.0027556302957236767, + 0.004419946577399969, + 0.0053347814828157425, + 0.004246479365974665, + 0.0065182410180568695, + 0.0035923346877098083, + 0.0034888610243797302, + 0.004157321993261576, + 0.004741312470287085, + 0.004839360248297453, + 0.004635958466678858, + 0.0023718089796602726, + 0.005940055008977652, + 0.006978088058531284, + 0.005279941018670797, + 0.00415036641061306, + 0.0026449342258274555, + 0.0014992416836321354, + 0.0013793004909530282, + 0.001391701982356608, + 0.002957206219434738, + 0.001532437396235764, + 0.0011515258811414242, + 0.0016271838685497642, + 0.0016063940711319447, + 0.048766233026981354, + 0.024010583758354187, + 0.052714940160512924, + 0.05143857002258301, + 0.05760132521390915, + 0.255188524723053, + 0.15153098106384277, + 0.028470585122704506, + 0.008438903838396072, + 0.03263182193040848, + 0.004802217707037926, + 0.006917564198374748, + 0.004353981930762529, + 0.02712216041982174, + 0.002663742983713746, + 0.0038200481794774532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.28136956691741943 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189953975379467 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003189953975379467, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189953975379467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0645417720079422, + 0.040229398757219315, + 0.02672063745558262, + 0.05290579795837402, + 0.01980455406010151, + 0.027037961408495903, + 0.1056067943572998, + 0.0247188750654459, + 0.17751987278461456, + 0.46091440320014954 + ], + "credit_map": [ + 0.0080668730661273, + 0.00786084495484829, + 0.00957410130649805, + 0.006806466728448868, + 0.013211624696850777, + 0.005777559243142605, + 0.006714564748108387, + 0.006529729347676039, + 0.005695994943380356, + 0.004825875628739595, + 0.0064880880527198315, + 0.005995672661811113, + 0.00506196916103363, + 0.003247893415391445, + 0.005848659668117762, + 0.0030652405694127083, + 0.0024125780910253525, + 0.004623575136065483, + 0.003333641681820154, + 0.0031435175333172083, + 0.004172428045421839, + 0.0025397678837180138, + 0.0035791611298918724, + 0.002915966557338834, + 0.006448167376220226, + 0.006860607303678989, + 0.0053981090895831585, + 0.006524166092276573, + 0.010129124857485294, + 0.004532022401690483, + 0.007344281766563654, + 0.005669318605214357, + 0.002307872986420989, + 0.0023111337795853615, + 0.0023879706859588623, + 0.0023255730047822, + 0.0028498906176537275, + 0.0025168273132294416, + 0.0024262266233563423, + 0.0026790592819452286, + 0.0032086092978715897, + 0.0033918262924999, + 0.002897734520956874, + 0.005552054848521948, + 0.0028002008330076933, + 0.0031815734691917896, + 0.00303534884005785, + 0.0029706107452511787, + 0.012636986561119556, + 0.013130602426826954, + 0.006502776872366667, + 0.017662053927779198, + 0.02136450633406639, + 0.01543397456407547, + 0.011828997172415257, + 0.00704688485711813, + 0.002118093194440007, + 0.0033953869715332985, + 0.0020606117323040962, + 0.0062919845804572105, + 0.0032642055302858353, + 0.0019534139428287745, + 0.0032346583902835846, + 0.0024005218874663115, + 0.012671753764152527, + 0.007130627520382404, + 0.014059977605938911, + 0.013821813277900219, + 0.01463842298835516, + 0.06591036170721054, + 0.04114899784326553, + 0.0081378985196352, + 0.04230817034840584, + 0.14463983476161957, + 0.05885132774710655, + 0.03676103428006172, + 0.028488734737038612, + 0.10172388702630997, + 0.02370132878422737, + 0.024440079927444458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.003189953975379467 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking fisheye lens close-up, shot from a low, upward angle, the camera remains stationary, completely capturing the viewer's gaze. A bony, fair-skinned male hand, steady and powerful, grips a transparent, modern glass. The glass is filled with a distinctly layered liquid, gradating from a deep purple at the bottom to a neon pink in the middle, and finally to a bright cyan at the top. Tiny bubbles linger in the liquid, shimmering with a strange luster. The background blurs into a profound darkness, with only a few blurred, color-distorted streaks of light distorting at the edges. As the hand raises the glass to his lips, the figure's taut jawline and the prominent Adam's apple at the bottom of the frame come into sharp focus. As he tilts his head back to drink, the apple begins to roll violently and rapidly, each movement imbued with power. At the highest point of the Adam's apple rolling, the entire picture was suddenly torn apart by horizontal and vertical digital noise and pixel blocks, the colors shifted drastically, the bright liquid colors were intertwined with the dazzling white light and the misplaced RGB channels, forming a visual instant collapse and explosion, creating a tense, psychedelic and strong atmosphere full of the aesthetics of collapse in the digital age.", + "seed": 31078, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_11/index_5799_seed_31078_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08909537941217423, + "vlm_logprob": -0.08909537941217423, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.08909537941217423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02917885221540928 + ], + "opposite_logprob_tokens": [ + -3.6541788578033447 + ], + "logprob_reward": -0.02917885221540928, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02917885221540928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1933819204568863, + 0.022494126111268997, + 0.05021870508790016, + 0.031276244670152664, + 0.03554851561784744, + 0.038614384829998016, + 0.06290874630212784, + 0.5175567865371704, + 0.02312462590634823, + 0.024875910952687263 + ], + "credit_map": [ + 0.0382034033536911, + 0.029330244287848473, + 0.044374980032444, + 0.030277973040938377, + 0.016640054062008858, + 0.010876161977648735, + 0.010071471333503723, + 0.013607650995254517, + 0.0032601123675704002, + 0.0030526225455105305, + 0.003544457955285907, + 0.0041863322257995605, + 0.0020552377682179213, + 0.0017656388226896524, + 0.0025645471177995205, + 0.002065180568024516, + 0.008946068584918976, + 0.008602515794336796, + 0.009853586554527283, + 0.011168682016432285, + 0.0027639104519039392, + 0.0021252455189824104, + 0.0031308496836572886, + 0.0036278509069234133, + 0.005963723640888929, + 0.0040186685509979725, + 0.004292555619031191, + 0.00685153529047966, + 0.002194045577198267, + 0.002825715346261859, + 0.0023174253292381763, + 0.0028125771787017584, + 0.005096998997032642, + 0.0037856444250792265, + 0.004462699871510267, + 0.007327220868319273, + 0.0029662749730050564, + 0.003071182407438755, + 0.0031175969634205103, + 0.005720898974686861, + 0.005366146564483643, + 0.004771384876221418, + 0.005112534388899803, + 0.00649029016494751, + 0.0028159278444945812, + 0.0030443635769188404, + 0.0045248279348015785, + 0.006488913204520941, + 0.009137876331806183, + 0.007511664181947708, + 0.01662553660571575, + 0.007360564544796944, + 0.005048444960266352, + 0.0052973199635744095, + 0.007413594052195549, + 0.004513746593147516, + 0.12987478077411652, + 0.012957898899912834, + 0.22260093688964844, + 0.02150704897940159, + 0.010124988853931427, + 0.015364419668912888, + 0.08815761655569077, + 0.016969164833426476, + 0.0030622705817222595, + 0.002894484670832753, + 0.0036451625637710094, + 0.0021445725578814745, + 0.0030857869423925877, + 0.0023890454322099686, + 0.0025533202569931746, + 0.0033499840646982193, + 0.0030724096577614546, + 0.003623873693868518, + 0.0037816280964761972, + 0.0031757040414959192, + 0.00294277211651206, + 0.002835614373907447, + 0.002759072929620743, + 0.0026848362758755684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.02917885221540928 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372172638773918 + ], + "opposite_logprob_tokens": [ + -3.408721685409546 + ], + "logprob_reward": -0.03372172638773918, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372172638773918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07306241989135742, + 0.029368948191404343, + 0.01973654143512249, + 0.01567666232585907, + 0.010868764482438564, + 0.04084746912121773, + 0.02331801876425743, + 0.7346862554550171, + 0.032971564680337906, + 0.01946338266134262 + ], + "credit_map": [ + 0.02131308615207672, + 0.004201231990009546, + 0.022246817126870155, + 0.006085669621825218, + 0.004185732454061508, + 0.0036007934249937534, + 0.005006289109587669, + 0.006422796286642551, + 0.00461177434772253, + 0.00302842422388494, + 0.005419332068413496, + 0.003387862117961049, + 0.002706090221181512, + 0.0020158598199486732, + 0.002848425880074501, + 0.00535117881372571, + 0.002109875902533531, + 0.002873818390071392, + 0.002562243491411209, + 0.004892588593065739, + 0.001167324255220592, + 0.0016872161068022251, + 0.0020035135094076395, + 0.0024399615358561277, + 0.0017073929775506258, + 0.0019946927204728127, + 0.0020820440258830786, + 0.003278942545875907, + 0.0012066204799339175, + 0.0015055780531838536, + 0.001735968398861587, + 0.0021654239390045404, + 0.0012484254548326135, + 0.0012978847371414304, + 0.0014254080597311258, + 0.0011551921488717198, + 0.0010978553909808397, + 0.00138082227203995, + 0.0009707085555419326, + 0.002292467514052987, + 0.004912674427032471, + 0.004664483014494181, + 0.005762492772191763, + 0.004627023823559284, + 0.0039008324965834618, + 0.005576185882091522, + 0.004149566404521465, + 0.007254208903759718, + 0.0027307793498039246, + 0.002632445190101862, + 0.003529980545863509, + 0.0029041501693427563, + 0.003529901849105954, + 0.0020729878451675177, + 0.0029664908070117235, + 0.002951282076537609, + 0.1808157116174698, + 0.01726408489048481, + 0.3227872848510742, + 0.02882791869342327, + 0.011920735239982605, + 0.01958967000246048, + 0.13178877532482147, + 0.02169208973646164, + 0.004778135102242231, + 0.0038275958504527807, + 0.005312651861459017, + 0.003317398950457573, + 0.0043328264728188515, + 0.0030234886799007654, + 0.003947784658521414, + 0.004431681241840124, + 0.0022429227828979492, + 0.005328168626874685, + 0.0018233851296827197, + 0.0020105026196688414, + 0.0023235073313117027, + 0.002164674224331975, + 0.0017003659158945084, + 0.0018698556814342737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.03372172638773918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980063669383526 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980063669383526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980063669383526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05138442665338516, + 0.024717921391129494, + 0.017740853130817413, + 0.10260652750730515, + 0.12793757021427155, + 0.05879775062203407, + 0.02659877948462963, + 0.5681419372558594, + 0.00709009263664484, + 0.014984237030148506 + ], + "credit_map": [ + 0.006934327073395252, + 0.009111059829592705, + 0.00859830528497696, + 0.00962797086685896, + 0.0060478681698441505, + 0.002758961170911789, + 0.0035803355276584625, + 0.004725594539195299, + 0.005230021197348833, + 0.0032931286841630936, + 0.006210703868418932, + 0.0037207871209830046, + 0.001886140089482069, + 0.0011537842219695449, + 0.0015213642036542296, + 0.0017019894439727068, + 0.0021074996329844, + 0.0030869825277477503, + 0.00236756750382483, + 0.004776301793754101, + 0.0010541235096752644, + 0.001129917218349874, + 0.0015803829301148653, + 0.0016380760353058577, + 0.014980895444750786, + 0.014295061118900776, + 0.014798152260482311, + 0.02660837583243847, + 0.0056281196884810925, + 0.006607771385461092, + 0.009043412283062935, + 0.010644728317856789, + 0.018526427447795868, + 0.01547903474420309, + 0.017523400485515594, + 0.0303502157330513, + 0.007630986161530018, + 0.009496631100773811, + 0.012452643364667892, + 0.016478216275572777, + 0.008661474101245403, + 0.006288261152803898, + 0.010110295377671719, + 0.011797802522778511, + 0.003736353013664484, + 0.0032086283899843693, + 0.006810287479311228, + 0.0081846434623003, + 0.004120240919291973, + 0.0020768223330378532, + 0.006960837636142969, + 0.003878391347825527, + 0.0016129695577546954, + 0.002056204481050372, + 0.0037300812546163797, + 0.0021632304415106773, + 0.14056482911109924, + 0.014312006533145905, + 0.24885769188404083, + 0.022757239639759064, + 0.008829757571220398, + 0.015675121918320656, + 0.10027924180030823, + 0.01686595380306244, + 0.0009056120179593563, + 0.0010142590617761016, + 0.0009112407569773495, + 0.0006837034597992897, + 0.001267978805117309, + 0.0007760022999718785, + 0.0007254214724525809, + 0.0008058735402300954, + 0.002420437056571245, + 0.0028576382901519537, + 0.0018951970851048827, + 0.0015999410534277558, + 0.0014487181324511766, + 0.0019261385314166546, + 0.0014242117758840322, + 0.0014119524275884032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.02980063669383526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486831486225128 + ], + "opposite_logprob_tokens": [ + -1.2236831188201904 + ], + "logprob_reward": -0.3486831486225128, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486831486225128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40739205479621887, + 0.04670019820332527, + 0.034067247062921524, + 0.07345221936702728, + 0.03741402551531792, + 0.06783391535282135, + 0.08701001852750778, + 0.16553400456905365, + 0.0351896770298481, + 0.04540657252073288 + ], + "credit_map": [ + 0.1388474702835083, + 0.030025508254766464, + 0.14360453188419342, + 0.035034097731113434, + 0.018445979803800583, + 0.01182837225496769, + 0.011463728733360767, + 0.018142446875572205, + 0.008734483271837234, + 0.007676577195525169, + 0.007382919546216726, + 0.007439861539751291, + 0.004684777464717627, + 0.002521947491914034, + 0.003914501518011093, + 0.004345137160271406, + 0.005564108956605196, + 0.00755059439688921, + 0.0055010137148201466, + 0.006423186976462603, + 0.001957629807293415, + 0.0019975723698735237, + 0.0024220843333750963, + 0.0026510588359087706, + 0.011070436798036098, + 0.009482302702963352, + 0.011340880766510963, + 0.01818208396434784, + 0.004366532899439335, + 0.006714412476867437, + 0.005744149908423424, + 0.006551430560648441, + 0.005987994372844696, + 0.004275172017514706, + 0.006463501136749983, + 0.009244667366147041, + 0.0020667072385549545, + 0.001850105356425047, + 0.0033755390904843807, + 0.004150343593209982, + 0.009281937964260578, + 0.008485859259963036, + 0.008880402892827988, + 0.015192776918411255, + 0.004473151173442602, + 0.003936644643545151, + 0.00827473308891058, + 0.00930841639637947, + 0.013102725148200989, + 0.007935463450849056, + 0.01678045094013214, + 0.012262672185897827, + 0.0071248565800487995, + 0.006345235742628574, + 0.012690696865320206, + 0.010767929255962372, + 0.031908366829156876, + 0.007085941266268492, + 0.06718889623880386, + 0.009942198172211647, + 0.004513435531407595, + 0.005846031475812197, + 0.031090181320905685, + 0.00795895978808403, + 0.004149789456278086, + 0.004231175407767296, + 0.007997593842446804, + 0.0042697973549366, + 0.00352908531203866, + 0.0028848578222095966, + 0.00459305988624692, + 0.003534319344907999, + 0.005753822159022093, + 0.009744871407747269, + 0.0058702402748167515, + 0.005972859915345907, + 0.004326962400227785, + 0.005338112358003855, + 0.004152853973209858, + 0.004246853291988373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.3486831486225128 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up featuring a bony, fair-skinned male hand gripping a transparent glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the liquid in the glass display a layered gradient from deep purple at the bottom to neon pink in the middle and bright cyan at the top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the hand raising the glass to the lips, revealing a taut jawline and a prominent Adam's apple in sharp focus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Adam's apple roll violently and rapidly as the figure tilts their head back to drink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the image being torn apart by horizontal and vertical digital noise, pixel blocks, and misplaced RGB channels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092533141374588 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092533141374588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092533141374588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31168943643569946, + 0.04917183145880699, + 0.029333632439374924, + 0.054778311401605606, + 0.030160170048475266, + 0.019658561795949936, + 0.0453159362077713, + 0.14837655425071716, + 0.07740723341703415, + 0.2341083139181137 + ], + "credit_map": [ + 0.10158850997686386, + 0.02696976810693741, + 0.10799239575862885, + 0.031557437032461166, + 0.014561926014721394, + 0.007044762372970581, + 0.00829459447413683, + 0.01368002314120531, + 0.010570270009338856, + 0.006397012621164322, + 0.009705924428999424, + 0.0075562079437077045, + 0.003888396080583334, + 0.0026299452874809504, + 0.003314120462164283, + 0.0051099564880132675, + 0.003920851740986109, + 0.004731056280434132, + 0.004891171585768461, + 0.0075624058954417706, + 0.001821294310502708, + 0.0018145182402804494, + 0.0022848064545542, + 0.002307530026882887, + 0.007324279751628637, + 0.006507501471787691, + 0.006473625544458628, + 0.010529501363635063, + 0.004393388982862234, + 0.005753739736974239, + 0.004403319675475359, + 0.009392958134412766, + 0.004158295225352049, + 0.003783636959269643, + 0.00395562220364809, + 0.006559555418789387, + 0.0020767010282725096, + 0.0030049665365368128, + 0.0028738174587488174, + 0.0037475754506886005, + 0.003944190219044685, + 0.001464676228351891, + 0.004959455691277981, + 0.0021005389280617237, + 0.0017298924503847957, + 0.0020681770984083414, + 0.0015812920173630118, + 0.0018103384645655751, + 0.006817155051976442, + 0.004717710893601179, + 0.009039386175572872, + 0.0064654420129954815, + 0.0036512790247797966, + 0.004220847971737385, + 0.005042640026658773, + 0.005361475050449371, + 0.027971666306257248, + 0.006713544949889183, + 0.05969573184847832, + 0.01039062812924385, + 0.004072565585374832, + 0.004693291150033474, + 0.02857290208339691, + 0.0062662274576723576, + 0.015610291622579098, + 0.01209264900535345, + 0.014010894112288952, + 0.007186028175055981, + 0.007437100168317556, + 0.008441634476184845, + 0.007604443468153477, + 0.0050241886638104916, + 0.030997449532151222, + 0.04981384053826332, + 0.027851969003677368, + 0.026285320520401, + 0.026838786900043488, + 0.02911391109228134, + 0.020440932363271713, + 0.02276611328125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.004092533141374588 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_12/index_tensor([5940, 5940])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_12/index_tensor([5940, 5940])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5495a04d76171129715d56518b91893d765140f5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_12/index_tensor([5940, 5940])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a vibrant, '90s Disney-style animation, a dynamic medium shot centers on a dramatic aerial standoff between two male blue Niltava birds, the camera panning smoothly to follow their frantic circular movements. Their feathers are a riot of bold, diaphanous colors: a brilliant, saturated cobalt blue covers their backs and wings, while their faces and throats are a deep, velvety black. Their large eyes narrow in pure aggression. They engage in a chaotic aerial fight, tumbling and spinning in a choreographed ballet of violence, viciously clawing at each other's breasts with sharp claws. As they struggle, individual blue and black feathers are shed and gracefully float downward in dreamy slow motion, catching the light. The entire conflict is set against a lush, hand-painted jungle backdrop, filled with oversized, stylized ferns and dreamy, softly glowing flowers, emitting an ethereal glow. Dramatic golden sunlight penetrates the dense foliage from above, illuminating the frantic movements and floating feathers, creating a tense, high-energy atmosphere. The birds' aggressive, shrill chirping seems to echo through the beautiful yet dangerous surroundings.", + "seed": 12275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_12/index_5940_seed_12275_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03129807082004845, + "vlm_logprob": -0.03129807082004845, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.03129807082004845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05962882563471794 + ], + "opposite_logprob_tokens": [ + -3.05962872505188 + ], + "logprob_reward": -0.05962882563471794, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05962882563471794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09613125026226044, + 0.05098579823970795, + 0.051828209310770035, + 0.25416338443756104, + 0.03049461357295513, + 0.03777909651398659, + 0.351955771446228, + 0.02267695590853691, + 0.0583423487842083, + 0.04564256966114044 + ], + "credit_map": [ + 0.010277901776134968, + 0.010918054729700089, + 0.016561096534132957, + 0.011677694506943226, + 0.01596679352223873, + 0.011560560204088688, + 0.008901378139853477, + 0.01026776060461998, + 0.005974830128252506, + 0.007785027381032705, + 0.005085465032607317, + 0.004734969697892666, + 0.006383930332958698, + 0.007805283647030592, + 0.006298442836850882, + 0.006917845457792282, + 0.0039058872498571873, + 0.012189527973532677, + 0.0030144641641527414, + 0.002869363408535719, + 0.003218748839572072, + 0.020151833072304726, + 0.0023793205618858337, + 0.004099062643945217, + 0.022473908960819244, + 0.10356223583221436, + 0.011026211082935333, + 0.011124121025204659, + 0.016585081815719604, + 0.0650021955370903, + 0.00916399247944355, + 0.01522565633058548, + 0.0031755431555211544, + 0.004293342586606741, + 0.0033380254171788692, + 0.002671350259333849, + 0.004036396276205778, + 0.006044060457497835, + 0.0031323896255344152, + 0.0038035025354474783, + 0.004216295201331377, + 0.006169689819216728, + 0.0031515539158135653, + 0.004301662091165781, + 0.006653120741248131, + 0.005065820645540953, + 0.00458360742777586, + 0.003637342480942607, + 0.06746627390384674, + 0.02306751348078251, + 0.03642375022172928, + 0.021735891699790955, + 0.022405320778489113, + 0.07216835767030716, + 0.02664789743721485, + 0.08204072713851929, + 0.0027967386413365602, + 0.0025365122128278017, + 0.0028546501416713, + 0.0025619177613407373, + 0.0036617317236959934, + 0.0030106003396213055, + 0.0022482757922261953, + 0.0030065278988331556, + 0.009315353818237782, + 0.004573864862322807, + 0.01305860374122858, + 0.006325097754597664, + 0.006237695924937725, + 0.006151960697025061, + 0.0063392529264092445, + 0.006340514402836561, + 0.007827315479516983, + 0.0036624076310545206, + 0.004771091975271702, + 0.0037037497386336327, + 0.0023120183032006025, + 0.011210405267775059, + 0.004033830016851425, + 0.008121751248836517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.05962882563471794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010550415609031916 + ], + "opposite_logprob_tokens": [ + -6.8760552406311035 + ], + "logprob_reward": -0.0010550415609031916, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010550415609031916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13363081216812134, + 0.0726633071899414, + 0.07672890275716782, + 0.295310914516449, + 0.0331420823931694, + 0.07083725184202194, + 0.06694231182336807, + 0.03549322113394737, + 0.07323497533798218, + 0.1420162469148636 + ], + "credit_map": [ + 0.017454389482736588, + 0.015085866674780846, + 0.027544967830181122, + 0.01718977838754654, + 0.011372219771146774, + 0.013787216506898403, + 0.014210516586899757, + 0.016985854133963585, + 0.00845442432910204, + 0.012101448141038418, + 0.009312078356742859, + 0.007325633428990841, + 0.009642272256314754, + 0.009034725837409496, + 0.008045893162488937, + 0.008746830746531487, + 0.006513289641588926, + 0.012028802186250687, + 0.004426825791597366, + 0.004250457976013422, + 0.004852082580327988, + 0.02947136200964451, + 0.003762203501537442, + 0.011423875577747822, + 0.022730372846126556, + 0.12323363870382309, + 0.011779519729316235, + 0.012701521627604961, + 0.020239075645804405, + 0.07524451613426208, + 0.010349354706704617, + 0.019032904878258705, + 0.00396479107439518, + 0.005383204203099012, + 0.004117196425795555, + 0.0029573445208370686, + 0.0054552904330194, + 0.004479933530092239, + 0.0037027536891400814, + 0.0030815701466053724, + 0.006333961151540279, + 0.008972669020295143, + 0.005901407916098833, + 0.008493514731526375, + 0.0066706459037959576, + 0.016235115006566048, + 0.009567025117576122, + 0.008662908338010311, + 0.009964694269001484, + 0.005523480009287596, + 0.008970963768661022, + 0.004510067403316498, + 0.005408980883657932, + 0.012639221735298634, + 0.006252011749893427, + 0.013672889210283756, + 0.00326105416752398, + 0.004655708558857441, + 0.004812672734260559, + 0.005526111461222172, + 0.003935857210308313, + 0.00467727379873395, + 0.004740772303193808, + 0.003883768804371357, + 0.01039850153028965, + 0.006608542054891586, + 0.01354626938700676, + 0.007780636195093393, + 0.010732145048677921, + 0.009629279375076294, + 0.008589725010097027, + 0.005949873011559248, + 0.024678383022546768, + 0.01004628837108612, + 0.019433042034506798, + 0.009253326803445816, + 0.00706192571669817, + 0.03762995824217796, + 0.009789672680199146, + 0.024123648181557655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0010550415609031916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571095868945122 + ], + "opposite_logprob_tokens": [ + -3.895570993423462 + ], + "logprob_reward": -0.020571095868945122, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571095868945122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1818232387304306, + 0.20091207325458527, + 0.037341032177209854, + 0.06808176636695862, + 0.04717494174838066, + 0.029592974111437798, + 0.33490264415740967, + 0.0337393581867218, + 0.0305462796241045, + 0.035885751247406006 + ], + "credit_map": [ + 0.015092543326318264, + 0.014967613853514194, + 0.02341063693165779, + 0.019231142476201057, + 0.0424109548330307, + 0.026834633201360703, + 0.020556652918457985, + 0.01931905560195446, + 0.02049289271235466, + 0.029925450682640076, + 0.026051262393593788, + 0.025975294411182404, + 0.02733803540468216, + 0.02275933139026165, + 0.020339632406830788, + 0.028030164539813995, + 0.005088946782052517, + 0.004365629982203245, + 0.0038110404275357723, + 0.0034317735116928816, + 0.002485142322257161, + 0.007316501345485449, + 0.0031677267979830503, + 0.007674267515540123, + 0.00666276179254055, + 0.028592074289917946, + 0.002850742544978857, + 0.0035482875537127256, + 0.006562028545886278, + 0.012552758678793907, + 0.0032231651712208986, + 0.004089936148375273, + 0.006650430615991354, + 0.00679191155359149, + 0.004814796149730682, + 0.004153704736381769, + 0.0038992336485534906, + 0.009636388160288334, + 0.006257456727325916, + 0.004971012938767672, + 0.0023960969410836697, + 0.002868998795747757, + 0.003415292128920555, + 0.0030113235116004944, + 0.0021877381950616837, + 0.006558412220329046, + 0.003308188868686557, + 0.0058469208888709545, + 0.06402754038572311, + 0.020208602771162987, + 0.03358298912644386, + 0.020147433504462242, + 0.018970802426338196, + 0.06496979296207428, + 0.026678619906306267, + 0.08631683140993118, + 0.005219145677983761, + 0.002776332199573517, + 0.004046252463012934, + 0.004104520194232464, + 0.004518698435276747, + 0.003996925428509712, + 0.003755878424271941, + 0.00532160047441721, + 0.0037395127583295107, + 0.003086855635046959, + 0.005172111093997955, + 0.004365415778011084, + 0.004926568828523159, + 0.003140862099826336, + 0.0028124151285737753, + 0.003302535507827997, + 0.003944092895835638, + 0.0030269587878137827, + 0.003592546796426177, + 0.003746194066479802, + 0.002536707790568471, + 0.00994840543717146, + 0.004634837154299021, + 0.004456004593521357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.020571095868945122 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00524691678583622 + ], + "opposite_logprob_tokens": [ + -5.255247116088867 + ], + "logprob_reward": -0.00524691678583622, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00524691678583622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08901267498731613, + 0.05320173501968384, + 0.046214405447244644, + 0.2609345614910126, + 0.08378944545984268, + 0.012009694240987301, + 0.3021044433116913, + 0.07084452360868454, + 0.021328801289200783, + 0.06055968254804611 + ], + "credit_map": [ + 0.00882338173687458, + 0.013994903303682804, + 0.010968280024826527, + 0.014198198914527893, + 0.008128028362989426, + 0.008206364698708057, + 0.013453716412186623, + 0.01123981736600399, + 0.005814475007355213, + 0.007213410921394825, + 0.006639108061790466, + 0.005885500926524401, + 0.009268611669540405, + 0.006511783227324486, + 0.004465518519282341, + 0.007403332740068436, + 0.004504321608692408, + 0.007577578537166119, + 0.004882576875388622, + 0.0035516198258847, + 0.007225611712783575, + 0.010512487031519413, + 0.003437051549553871, + 0.004523162264376879, + 0.022626809775829315, + 0.10149428248405457, + 0.013567431829869747, + 0.015739815309643745, + 0.01789121888577938, + 0.05480681732296944, + 0.013140886090695858, + 0.02166731283068657, + 0.011991595849394798, + 0.012543855234980583, + 0.009179781191051006, + 0.009934644214808941, + 0.008721048943698406, + 0.012652797624468803, + 0.006509172264486551, + 0.012256565503776073, + 0.00100433686748147, + 0.0012347846059128642, + 0.0010015881853178144, + 0.0012868692865595222, + 0.0010530040599405766, + 0.0029094116762280464, + 0.0014200498117133975, + 0.0020996511448174715, + 0.05920173600316048, + 0.018720753490924835, + 0.03003673441708088, + 0.016360851004719734, + 0.016023483127355576, + 0.06399989128112793, + 0.02415136806666851, + 0.07360965758562088, + 0.006296322215348482, + 0.012936966493725777, + 0.0055919671431183815, + 0.0071555692702531815, + 0.011669086292386055, + 0.011226003989577293, + 0.007096006069332361, + 0.008872604928910732, + 0.003123166272416711, + 0.001578861498273909, + 0.004447148181498051, + 0.0023878279607743025, + 0.003310839645564556, + 0.0025186913553625345, + 0.0019617066718637943, + 0.002000561449676752, + 0.01197119802236557, + 0.0041786557994782925, + 0.006956976372748613, + 0.0041889529675245285, + 0.00308855134062469, + 0.014233752153813839, + 0.004851446487009525, + 0.011090157553553581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.00524691678583622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06998847424983978 + ], + "opposite_logprob_tokens": [ + -2.694988489151001 + ], + "logprob_reward": -0.06998847424983978, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06998847424983978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17734520137310028, + 0.1026030033826828, + 0.11363236606121063, + 0.23405112326145172, + 0.05150260403752327, + 0.0405539833009243, + 0.06366923451423645, + 0.03547285869717598, + 0.04639275744557381, + 0.13477692008018494 + ], + "credit_map": [ + 0.01723846234381199, + 0.02011234126985073, + 0.030295997858047485, + 0.031185543164610863, + 0.017462149262428284, + 0.017463594675064087, + 0.01930476725101471, + 0.024282339960336685, + 0.013862970285117626, + 0.02158772014081478, + 0.009273584932088852, + 0.007988694123923779, + 0.017707200720906258, + 0.013601746410131454, + 0.008545590564608574, + 0.010035498067736626, + 0.010160847567021847, + 0.023300332948565483, + 0.006889639887958765, + 0.007361388299614191, + 0.0071431780233979225, + 0.04133019223809242, + 0.005472057964652777, + 0.011974731460213661, + 0.021410604938864708, + 0.09261126816272736, + 0.010315348394215107, + 0.012021089904010296, + 0.015408482402563095, + 0.05281754210591316, + 0.010478967800736427, + 0.018987812101840973, + 0.008399466052651405, + 0.007421558722853661, + 0.005251054186373949, + 0.0048540676943957806, + 0.006538313813507557, + 0.006929290946573019, + 0.005112078972160816, + 0.0069967713207006454, + 0.004758926574140787, + 0.006762030068784952, + 0.003718918189406395, + 0.0037932684645056725, + 0.0070235393941402435, + 0.003962858580052853, + 0.005977580323815346, + 0.004556863568723202, + 0.010015519335865974, + 0.0048687574453651905, + 0.007379205897450447, + 0.00579923577606678, + 0.006080894730985165, + 0.008746909908950329, + 0.006573236547410488, + 0.014205480925738811, + 0.0062225330621004105, + 0.003934589680284262, + 0.003916408866643906, + 0.003918602131307125, + 0.004770922474563122, + 0.004334249533712864, + 0.003925873897969723, + 0.004449679050594568, + 0.005255457013845444, + 0.005470305681228638, + 0.007357217371463776, + 0.0046106427907943726, + 0.007444832008332014, + 0.005730625241994858, + 0.005096225533634424, + 0.005427452735602856, + 0.020194312557578087, + 0.012361257337033749, + 0.01555850449949503, + 0.010452816262841225, + 0.011565225198864937, + 0.018593616783618927, + 0.012983711436390877, + 0.033067476004362106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.06998847424983978 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a vibrant, '90s Disney-style animation, a dynamic medium shot centers on a dramatic aerial standoff between two male blue Niltava birds, the camera panning smoothly to follow their frantic circular movements. Their feathers are a riot of bold, diaphanous colors: a brilliant, saturated cobalt blue covers their backs and wings, while their faces and throats are a deep, velvety black. Their large eyes narrow in pure aggression. They engage in a chaotic aerial fight, tumbling and spinning in a choreographed ballet of violence, viciously clawing at each other's breasts with sharp claws. As they struggle, individual blue and black feathers are shed and gracefully float downward in dreamy slow motion, catching the light. The entire conflict is set against a lush, hand-painted jungle backdrop, filled with oversized, stylized ferns and dreamy, softly glowing flowers, emitting an ethereal glow. Dramatic golden sunlight penetrates the dense foliage from above, illuminating the frantic movements and floating feathers, creating a tense, high-energy atmosphere. The birds' aggressive, shrill chirping seems to echo through the beautiful yet dangerous surroundings.", + "seed": 12276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_12/index_5940_seed_12276_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.052432540676090864, + "vlm_logprob": -0.052432540676090864, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.052432540676090864, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09869877249002457 + ], + "opposite_logprob_tokens": [ + -2.473698854446411 + ], + "logprob_reward": -0.09869877249002457, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09869877249002457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34223422408103943, + 0.07185149192810059, + 0.11443877965211868, + 0.13700099289417267, + 0.032763414084911346, + 0.031108275055885315, + 0.06180055812001228, + 0.09099370241165161, + 0.04818398132920265, + 0.0696246474981308 + ], + "credit_map": [ + 0.10265275090932846, + 0.02977830357849598, + 0.024821564555168152, + 0.03270024433732033, + 0.024797553196549416, + 0.07710777223110199, + 0.0282033272087574, + 0.022172706201672554, + 0.00690783467143774, + 0.007174494210630655, + 0.00595462042838335, + 0.008839340880513191, + 0.006906665395945311, + 0.01237082201987505, + 0.01715107448399067, + 0.006546638440340757, + 0.014611712656915188, + 0.02070179022848606, + 0.013441222719848156, + 0.01786578632891178, + 0.01121713686734438, + 0.013108388520777225, + 0.014412416145205498, + 0.00908032525330782, + 0.015295563265681267, + 0.011371192522346973, + 0.03341401368379593, + 0.020450858399271965, + 0.011638825759291649, + 0.012546541169285774, + 0.015602572821080685, + 0.016681423410773277, + 0.0037625448312610388, + 0.0030916689429432154, + 0.004102237056940794, + 0.004525840748101473, + 0.0040474263951182365, + 0.003656551707535982, + 0.005392656195908785, + 0.004184487741440535, + 0.004893377888947725, + 0.003785498905926943, + 0.003147256327793002, + 0.00444922037422657, + 0.003811671631410718, + 0.004831197205930948, + 0.003335156012326479, + 0.0028548967093229294, + 0.009701545350253582, + 0.005461858119815588, + 0.0050535607151687145, + 0.01003858633339405, + 0.008036934770643711, + 0.010381260886788368, + 0.008344017900526524, + 0.0047827898524701595, + 0.011724326759576797, + 0.01483042910695076, + 0.02624557353556156, + 0.005872820038348436, + 0.005298303440213203, + 0.010173745453357697, + 0.011317135766148567, + 0.0055313692428171635, + 0.011392167769372463, + 0.005324045196175575, + 0.004763271193951368, + 0.004347192123532295, + 0.004936077632009983, + 0.004658014513552189, + 0.0070983669720590115, + 0.00566484360024333, + 0.010975198820233345, + 0.007274335715919733, + 0.008712690323591232, + 0.008267121389508247, + 0.009372210130095482, + 0.010074881836771965, + 0.008969730697572231, + 0.005978476721793413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.09869877249002457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009357126546092331 + ], + "opposite_logprob_tokens": [ + -7.0009355545043945 + ], + "logprob_reward": -0.0009357126546092331, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009357126546092331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2988354563713074, + 0.05000436305999756, + 0.04828256741166115, + 0.2736196219921112, + 0.07267212122678757, + 0.027317581698298454, + 0.02713698521256447, + 0.0679914727807045, + 0.042671967297792435, + 0.09146783500909805 + ], + "credit_map": [ + 0.11073037981987, + 0.022574901580810547, + 0.01695679873228073, + 0.01346108689904213, + 0.013791797682642937, + 0.09245521575212479, + 0.011352616362273693, + 0.017512677237391472, + 0.0051087490282952785, + 0.0037075113505125046, + 0.007660472299903631, + 0.006257351487874985, + 0.005226724315434694, + 0.008030124008655548, + 0.009777244180440903, + 0.004236187785863876, + 0.006962667219340801, + 0.007288830354809761, + 0.006392559502273798, + 0.005024600774049759, + 0.0037961620837450027, + 0.0071936906315386295, + 0.0070643057115375996, + 0.0045597548596560955, + 0.02407471463084221, + 0.02229308895766735, + 0.055398788303136826, + 0.043014589697122574, + 0.023749899119138718, + 0.02095806784927845, + 0.062358271330595016, + 0.021772189065814018, + 0.010308328084647655, + 0.008590271696448326, + 0.007205955218523741, + 0.0074122208170592785, + 0.012017103843390942, + 0.006754586938768625, + 0.011276615783572197, + 0.009107041172683239, + 0.0045751952566206455, + 0.0033643741626292467, + 0.0026291401591151953, + 0.0030688075348734856, + 0.0039950525388121605, + 0.0029396440368145704, + 0.0032983054406940937, + 0.003447065595537424, + 0.004217634908854961, + 0.0023236554116010666, + 0.002246994525194168, + 0.005551625974476337, + 0.004719128832221031, + 0.0030974410474300385, + 0.0029872956220060587, + 0.0019932123832404613, + 0.009767552837729454, + 0.008731184527277946, + 0.01753493770956993, + 0.0048236544243991375, + 0.004434657283127308, + 0.007685273885726929, + 0.009659351781010628, + 0.00535485940054059, + 0.006752180401235819, + 0.004192032385617495, + 0.00407353276386857, + 0.00499327015131712, + 0.005967068951576948, + 0.006927613168954849, + 0.004323132336139679, + 0.005443141330033541, + 0.0123867467045784, + 0.00792775396257639, + 0.012575129978358746, + 0.006780659314244986, + 0.009410837665200233, + 0.011576167307794094, + 0.020336152985692024, + 0.01047439407557249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0009357126546092331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895011454820633 + ], + "opposite_logprob_tokens": [ + -2.5789501667022705 + ], + "logprob_reward": -0.07895011454820633, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895011454820633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14375878870487213, + 0.06046820431947708, + 0.11218446493148804, + 0.20275038480758667, + 0.03849441185593605, + 0.02897014282643795, + 0.10691773891448975, + 0.13321201503276825, + 0.12300486862659454, + 0.050238899886608124 + ], + "credit_map": [ + 0.03166614845395088, + 0.012809443287551403, + 0.017493579536676407, + 0.01615435630083084, + 0.011857731267809868, + 0.024709638208150864, + 0.01232459582388401, + 0.016743313521146774, + 0.005133919883519411, + 0.006381339393556118, + 0.00679041538387537, + 0.00600019097328186, + 0.0055401511490345, + 0.01124168187379837, + 0.015083052217960358, + 0.004297464154660702, + 0.014171754941344261, + 0.02210870571434498, + 0.0119563527405262, + 0.01595979556441307, + 0.011687543243169785, + 0.0127333365380764, + 0.014512792229652405, + 0.009054195135831833, + 0.023722557350993156, + 0.017712309956550598, + 0.0411921925842762, + 0.030176039785146713, + 0.012332254089415073, + 0.02159842848777771, + 0.027344543486833572, + 0.02867208607494831, + 0.0051008565351367, + 0.00413742009550333, + 0.006915791891515255, + 0.00500844232738018, + 0.005197777412831783, + 0.0032036120537668467, + 0.004623418673872948, + 0.004307096358388662, + 0.003779534948989749, + 0.0045930007472634315, + 0.002046512207016349, + 0.00291352323256433, + 0.0055891708470880985, + 0.003977234475314617, + 0.002994189504534006, + 0.0030769789591431618, + 0.013512350618839264, + 0.008412478491663933, + 0.008669844828546047, + 0.021115876734256744, + 0.018121054396033287, + 0.013456843793392181, + 0.014915548264980316, + 0.008713745512068272, + 0.020421043038368225, + 0.024303311482071877, + 0.04000398516654968, + 0.007784574292600155, + 0.006474592722952366, + 0.012095590122044086, + 0.016243478283286095, + 0.005885444115847349, + 0.015703151002526283, + 0.009103509597480297, + 0.013383250683546066, + 0.01627361588180065, + 0.01918429136276245, + 0.01357737835496664, + 0.022578716278076172, + 0.013200955465435982, + 0.009543056599795818, + 0.006268009543418884, + 0.007016636896878481, + 0.008203361183404922, + 0.004630278795957565, + 0.005019140895456076, + 0.005447797942906618, + 0.004110627342015505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.07895011454820633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633167292922735 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633167292922735, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633167292922735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17463278770446777, + 0.08802899718284607, + 0.04410196840763092, + 0.2948647141456604, + 0.043059419840574265, + 0.04286203905940056, + 0.12347738444805145, + 0.07936153560876846, + 0.051949258893728256, + 0.05766184628009796 + ], + "credit_map": [ + 0.03713627904653549, + 0.017988864332437515, + 0.018338125199079514, + 0.017984868958592415, + 0.011466794647276402, + 0.03511030599474907, + 0.018071042373776436, + 0.018536517396569252, + 0.007123006973415613, + 0.008559854701161385, + 0.009941045194864273, + 0.007945354096591473, + 0.00702603068202734, + 0.015906620770692825, + 0.02231338992714882, + 0.0092137036845088, + 0.006225505378097296, + 0.008943728171288967, + 0.005078988149762154, + 0.005283030215650797, + 0.003517038654536009, + 0.0039020071271806955, + 0.006196553353220224, + 0.0049551245756447315, + 0.02524050511419773, + 0.023332105949521065, + 0.05314457044005394, + 0.05212150514125824, + 0.024984609335660934, + 0.023658012971282005, + 0.07472450286149979, + 0.017658911645412445, + 0.00498999236151576, + 0.0044493586756289005, + 0.005045418161898851, + 0.006300716195255518, + 0.0061336406506598, + 0.003736334852874279, + 0.006003987975418568, + 0.006399972829967737, + 0.008328811265528202, + 0.004840935580432415, + 0.0037836087867617607, + 0.0054854401387274265, + 0.0044951168820261955, + 0.005720197223126888, + 0.004201232455670834, + 0.006006703246384859, + 0.014889688231050968, + 0.010830090381205082, + 0.010138310492038727, + 0.023630404844880104, + 0.017971960827708244, + 0.018641700968146324, + 0.01774294488132, + 0.009632300585508347, + 0.016826791688799858, + 0.009504564106464386, + 0.010439560748636723, + 0.007373312022536993, + 0.005949233192950487, + 0.014217705465853214, + 0.007616219576448202, + 0.0074341511353850365, + 0.010580767877399921, + 0.005651205312460661, + 0.005609211977571249, + 0.004128697328269482, + 0.005400034599006176, + 0.009797492064535618, + 0.005436101462692022, + 0.005345754325389862, + 0.011142459698021412, + 0.006582764908671379, + 0.007972031831741333, + 0.0071211326867341995, + 0.0045138197019696236, + 0.007121677976101637, + 0.008185701444745064, + 0.005022263154387474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.004633167292922735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894493639469147 + ], + "opposite_logprob_tokens": [ + -2.5789449214935303 + ], + "logprob_reward": -0.07894493639469147, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894493639469147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09502700716257095, + 0.060783229768276215, + 0.06455594301223755, + 0.3967917263507843, + 0.03260292857885361, + 0.03160598874092102, + 0.1229923814535141, + 0.08073629438877106, + 0.08894827216863632, + 0.02595621533691883 + ], + "credit_map": [ + 0.01766514778137207, + 0.008493006229400635, + 0.011344153434038162, + 0.011763636954128742, + 0.007977264001965523, + 0.015847520902752876, + 0.010839734226465225, + 0.01109654363244772, + 0.003967497497797012, + 0.005230078939348459, + 0.00592413917183876, + 0.004848148673772812, + 0.004524661228060722, + 0.01541975699365139, + 0.01550442073494196, + 0.005364527460187674, + 0.00860312394797802, + 0.010670991614460945, + 0.007235948462039232, + 0.009123161435127258, + 0.005067001562565565, + 0.005928489845246077, + 0.0119389109313488, + 0.0059883128851652145, + 0.03462459519505501, + 0.030796771869063377, + 0.06067968159914017, + 0.07593172788619995, + 0.032076235860586166, + 0.03471401333808899, + 0.10305217653512955, + 0.0249165166169405, + 0.003989154007285833, + 0.002998758340254426, + 0.0060531687922775745, + 0.0036557039711624384, + 0.0038377351593226194, + 0.0035110791213810444, + 0.004575057420879602, + 0.003982271999120712, + 0.007948774844408035, + 0.0035882701631635427, + 0.002552125370129943, + 0.0024868121836334467, + 0.004129644483327866, + 0.003997214138507843, + 0.0033956251572817564, + 0.0035075205378234386, + 0.018632560968399048, + 0.011932618916034698, + 0.012935670092701912, + 0.01832057721912861, + 0.0204777754843235, + 0.017772750928997993, + 0.01471866387873888, + 0.008201765827834606, + 0.00842499453574419, + 0.016228094696998596, + 0.02265971153974533, + 0.007521970197558403, + 0.005482376087456942, + 0.006545576732605696, + 0.009728427976369858, + 0.004145139828324318, + 0.025915078818798065, + 0.010024388320744038, + 0.009309862740337849, + 0.010329625569283962, + 0.008618702180683613, + 0.009639300405979156, + 0.00935168657451868, + 0.005759629886597395, + 0.004854853264987469, + 0.0034664717968553305, + 0.003567921230569482, + 0.002835307503119111, + 0.002084560925140977, + 0.002671266905963421, + 0.004067962523549795, + 0.0024078732822090387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.07894493639469147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a vibrant, '90s Disney-style animation, a dynamic medium shot centers on a dramatic aerial standoff between two male blue Niltava birds, the camera panning smoothly to follow their frantic circular movements. Their feathers are a riot of bold, diaphanous colors: a brilliant, saturated cobalt blue covers their backs and wings, while their faces and throats are a deep, velvety black. Their large eyes narrow in pure aggression. They engage in a chaotic aerial fight, tumbling and spinning in a choreographed ballet of violence, viciously clawing at each other's breasts with sharp claws. As they struggle, individual blue and black feathers are shed and gracefully float downward in dreamy slow motion, catching the light. The entire conflict is set against a lush, hand-painted jungle backdrop, filled with oversized, stylized ferns and dreamy, softly glowing flowers, emitting an ethereal glow. Dramatic golden sunlight penetrates the dense foliage from above, illuminating the frantic movements and floating feathers, creating a tense, high-energy atmosphere. The birds' aggressive, shrill chirping seems to echo through the beautiful yet dangerous surroundings.", + "seed": 12277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_12/index_5940_seed_12277_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09922886658459902, + "vlm_logprob": -0.09922886658459902, + "vlm_yes_no_margin": 3.8500000119209288, + "vlm_token_credit": -0.09922886658459902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055405762046575546 + ], + "opposite_logprob_tokens": [ + -3.055405855178833 + ], + "logprob_reward": -0.055405762046575546, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.055405762046575546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.587573766708374, + 0.09810016304254532, + 0.08000560849905014, + 0.028650516644120216, + 0.028915461152791977, + 0.039134759455919266, + 0.028839576989412308, + 0.06310389935970306, + 0.020881567150354385, + 0.02479470893740654 + ], + "credit_map": [ + 0.14921967685222626, + 0.18931974470615387, + 0.03911313787102699, + 0.03946100175380707, + 0.04012571647763252, + 0.055912625044584274, + 0.03403046354651451, + 0.04039134830236435, + 0.011940741911530495, + 0.006831401959061623, + 0.01345668826252222, + 0.008434844203293324, + 0.011319899000227451, + 0.01911097578704357, + 0.014811987988650799, + 0.012193625792860985, + 0.008693529292941093, + 0.010221285745501518, + 0.01145633589476347, + 0.007159031927585602, + 0.014844618737697601, + 0.00876497384160757, + 0.009007478132843971, + 0.00985835399478674, + 0.003485222812741995, + 0.0033182455226778984, + 0.0033679218031466007, + 0.00281217391602695, + 0.002714661182835698, + 0.005070617888122797, + 0.003885360201820731, + 0.003996314946562052, + 0.0038897511549293995, + 0.0028717885725200176, + 0.004052476957440376, + 0.004271680023521185, + 0.003621304640546441, + 0.0030940494034439325, + 0.0036033024080097675, + 0.003511107061058283, + 0.006924676243215799, + 0.005997076164931059, + 0.0031985812820494175, + 0.0035002771764993668, + 0.003527830122038722, + 0.007080711890012026, + 0.00389568111859262, + 0.005009925924241543, + 0.0035427999682724476, + 0.002623661421239376, + 0.00395348621532321, + 0.00376606872305274, + 0.00358554907143116, + 0.0035165948793292046, + 0.00389352859929204, + 0.0039578890427947044, + 0.009657107293605804, + 0.007729706354439259, + 0.00849121529608965, + 0.011044973507523537, + 0.007615549024194479, + 0.00820585060864687, + 0.005681992042809725, + 0.004677511751651764, + 0.0029172964859753847, + 0.002524601062759757, + 0.002096671611070633, + 0.002160398056730628, + 0.0027428241446614265, + 0.0038766965735703707, + 0.002146693877875805, + 0.0024163860362023115, + 0.0026986643206328154, + 0.0025984288658946753, + 0.0028212873730808496, + 0.0038134667556732893, + 0.003393240040168166, + 0.0020084939897060394, + 0.004455534275621176, + 0.003005592618137598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.055405762046575546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001190906623378396 + ], + "opposite_logprob_tokens": [ + -6.751191139221191 + ], + "logprob_reward": -0.001190906623378396, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001190906623378396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39298534393310547, + 0.26762455701828003, + 0.08647466450929642, + 0.026060858741402626, + 0.016484210267663002, + 0.020726662129163742, + 0.02286408096551895, + 0.08562033623456955, + 0.042305655777454376, + 0.03885359689593315 + ], + "credit_map": [ + 0.042813174426555634, + 0.12563002109527588, + 0.02451866678893566, + 0.027613960206508636, + 0.023666633293032646, + 0.03213940188288689, + 0.021799776703119278, + 0.09480371326208115, + 0.04095030575990677, + 0.015883807092905045, + 0.04633312299847603, + 0.021887771785259247, + 0.021321462467312813, + 0.060319237411022186, + 0.028349431231617928, + 0.03257942572236061, + 0.011904893442988396, + 0.011827641166746616, + 0.012026133015751839, + 0.007246057502925396, + 0.01230772864073515, + 0.01204198319464922, + 0.006106606684625149, + 0.013013620860874653, + 0.002707291627302766, + 0.003009006381034851, + 0.0059873489663004875, + 0.003426790004596114, + 0.002054968848824501, + 0.0032422111835330725, + 0.0028862294275313616, + 0.0027470123022794724, + 0.0028605638071894646, + 0.00179950799793005, + 0.0022049511317163706, + 0.0018618344329297543, + 0.001826621824875474, + 0.0018010366475209594, + 0.0020318771712481976, + 0.002097817836329341, + 0.002219913760200143, + 0.0035017221234738827, + 0.0027612901758402586, + 0.002000778680667281, + 0.002583602210506797, + 0.002671626163646579, + 0.0026682750321924686, + 0.002319454215466976, + 0.0033949878998100758, + 0.002885300200432539, + 0.0027948766946792603, + 0.0022905683144927025, + 0.003399851731956005, + 0.00309762847609818, + 0.0023025337141007185, + 0.002698333701118827, + 0.007714303210377693, + 0.010389384813606739, + 0.010788509622216225, + 0.01790029928088188, + 0.011073495261371136, + 0.011300677433609962, + 0.009555249474942684, + 0.006898412015289068, + 0.004094650503247976, + 0.005796089302748442, + 0.00338156009092927, + 0.0042091598734259605, + 0.009014825336635113, + 0.005879898089915514, + 0.0044815135188400745, + 0.005447959061712027, + 0.00787911843508482, + 0.004135648719966412, + 0.005122373811900616, + 0.0036657885648310184, + 0.004552772734314203, + 0.003679389599710703, + 0.004973551258444786, + 0.004844953306019306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.001190906623378396 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34860390424728394 + ], + "opposite_logprob_tokens": [ + -1.2236039638519287 + ], + "logprob_reward": -0.34860390424728394, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34860390424728394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4117400646209717, + 0.16453489661216736, + 0.08883274346590042, + 0.02705286256968975, + 0.05627359077334404, + 0.029550490900874138, + 0.04753672331571579, + 0.0515982024371624, + 0.0482809878885746, + 0.07459942996501923 + ], + "credit_map": [ + 0.12001092731952667, + 0.11041785776615143, + 0.019588202238082886, + 0.024242127314209938, + 0.022147348150610924, + 0.046724334359169006, + 0.022048402577638626, + 0.04656089469790459, + 0.018844004720449448, + 0.012548104859888554, + 0.03168309107422829, + 0.01564565673470497, + 0.015804395079612732, + 0.030224375426769257, + 0.02135387994349003, + 0.01843138411641121, + 0.012791831977665424, + 0.011720746755599976, + 0.009078525938093662, + 0.009021982550621033, + 0.013865779154002666, + 0.011445350013673306, + 0.007698156405240297, + 0.013210379518568516, + 0.0034957067109644413, + 0.002482935320585966, + 0.003650949103757739, + 0.0025553316809237003, + 0.002543138340115547, + 0.005293524824082851, + 0.0036764871329069138, + 0.0033547913189977407, + 0.008787947706878185, + 0.006680049933493137, + 0.011110063642263412, + 0.005006269086152315, + 0.005363367963582277, + 0.007951107807457447, + 0.005775017198175192, + 0.0055997720919549465, + 0.004251338541507721, + 0.004167596343904734, + 0.0024872166104614735, + 0.0034419717267155647, + 0.004281540401279926, + 0.00445911567658186, + 0.0031513383146375418, + 0.0033103746827691793, + 0.007128980942070484, + 0.005447188392281532, + 0.004553753416985273, + 0.003848088439553976, + 0.006042076274752617, + 0.007251670118421316, + 0.004855480510741472, + 0.008409486152231693, + 0.00542291346937418, + 0.008617912419140339, + 0.004558646120131016, + 0.008526244200766087, + 0.008048892952501774, + 0.004813224542886019, + 0.005331945605576038, + 0.006278424058109522, + 0.007427750620990992, + 0.007290822919458151, + 0.006118902936577797, + 0.003441627835854888, + 0.004195664077997208, + 0.010641082189977169, + 0.004804830998182297, + 0.00436030887067318, + 0.010453338734805584, + 0.009653622284531593, + 0.006749327760189772, + 0.007403254508972168, + 0.013586344197392464, + 0.007952065207064152, + 0.011158415116369724, + 0.007643073797225952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.34860390424728394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019377756398171186 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.0019377756398171186, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019377756398171186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31171536445617676, + 0.0726538822054863, + 0.06453625857830048, + 0.03478287160396576, + 0.03645247593522072, + 0.12222069501876831, + 0.02544769085943699, + 0.17452695965766907, + 0.09620983898639679, + 0.061453964561223984 + ], + "credit_map": [ + 0.07490459829568863, + 0.018187254667282104, + 0.029472902417182922, + 0.017606917768716812, + 0.01629376783967018, + 0.04151095822453499, + 0.015097931027412415, + 0.0986410453915596, + 0.011531618423759937, + 0.005408399272710085, + 0.010981806553900242, + 0.004863422829657793, + 0.007597370073199272, + 0.016595786437392235, + 0.007149840239435434, + 0.008525640703737736, + 0.009439881891012192, + 0.008375667966902256, + 0.016345402225852013, + 0.005540503654628992, + 0.007546219974756241, + 0.006204410921782255, + 0.005461037624627352, + 0.0056231324560940266, + 0.006925816182047129, + 0.0037710764445364475, + 0.0045551154762506485, + 0.0028496524319052696, + 0.003668559715151787, + 0.00466116750612855, + 0.004496519919484854, + 0.0038549648597836494, + 0.004542248323559761, + 0.003306429833173752, + 0.005406943149864674, + 0.003895993111655116, + 0.0037497072480618954, + 0.004910585004836321, + 0.005178132560104132, + 0.005462440196424723, + 0.014097193256020546, + 0.018119102343916893, + 0.019669286906719208, + 0.014260033145546913, + 0.010819106362760067, + 0.01830250397324562, + 0.012488230131566525, + 0.014465236105024815, + 0.0045143854804337025, + 0.0030126823112368584, + 0.0024727897252887487, + 0.0026658244896680117, + 0.003207372035831213, + 0.0030238116160035133, + 0.0025715469382703304, + 0.0039792791940271854, + 0.03903514891862869, + 0.02224595472216606, + 0.027348391711711884, + 0.02641860581934452, + 0.01423096563667059, + 0.019676128402352333, + 0.013088519684970379, + 0.012483255006372929, + 0.01282505877315998, + 0.01417749933898449, + 0.011599170044064522, + 0.007577099371701479, + 0.012965643778443336, + 0.014835976995527744, + 0.012524033896625042, + 0.009705360978841782, + 0.009847494773566723, + 0.005041897762566805, + 0.009691966697573662, + 0.006467249244451523, + 0.0071785044856369495, + 0.005764776840806007, + 0.009505771100521088, + 0.007956303656101227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0019377756398171186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890059843659401 + ], + "opposite_logprob_tokens": [ + -2.464005947113037 + ], + "logprob_reward": -0.0890059843659401, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890059843659401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25198832154273987, + 0.19311398267745972, + 0.06658202409744263, + 0.08428037911653519, + 0.04105083644390106, + 0.06353478133678436, + 0.04816335439682007, + 0.12587232887744904, + 0.06286546587944031, + 0.06254854053258896 + ], + "credit_map": [ + 0.058954399079084396, + 0.054452575743198395, + 0.03444455936551094, + 0.018137503415346146, + 0.013700257986783981, + 0.03015078417956829, + 0.016054755076766014, + 0.026093464344739914, + 0.022193174809217453, + 0.014808318577706814, + 0.03631483390927315, + 0.01782107725739479, + 0.015483276918530464, + 0.038937222212553024, + 0.021506749093532562, + 0.026049340143799782, + 0.007215982303023338, + 0.0054410467855632305, + 0.018061477690935135, + 0.005618348252028227, + 0.007464239373803139, + 0.008952987380325794, + 0.006396581884473562, + 0.007431360427290201, + 0.009686801582574844, + 0.010158396326005459, + 0.018452370539307594, + 0.009113852865993977, + 0.00818557571619749, + 0.010701973922550678, + 0.009789333678781986, + 0.008192071691155434, + 0.006239317357540131, + 0.00471543101593852, + 0.006403312552720308, + 0.004269976168870926, + 0.0042155408300459385, + 0.0049721598625183105, + 0.005389256402850151, + 0.004845840390771627, + 0.007454833015799522, + 0.008820682764053345, + 0.009485196322202682, + 0.006286995951086283, + 0.007155509665608406, + 0.008497403003275394, + 0.008038940839469433, + 0.007795225828886032, + 0.008040301501750946, + 0.005483946762979031, + 0.004343581385910511, + 0.005808891262859106, + 0.0041075097396969795, + 0.006283697206526995, + 0.006934853736311197, + 0.007160574663430452, + 0.015324610285460949, + 0.015324743464589119, + 0.014459135942161083, + 0.020640354603528976, + 0.02146465703845024, + 0.015016759745776653, + 0.01087452843785286, + 0.012767545878887177, + 0.0101016229018569, + 0.007944304496049881, + 0.007285261061042547, + 0.0073433793149888515, + 0.0070290942676365376, + 0.010065551847219467, + 0.006934286095201969, + 0.006161967758089304, + 0.007321970071643591, + 0.007107763085514307, + 0.004484980832785368, + 0.005299154203385115, + 0.008997590281069279, + 0.010016254149377346, + 0.00856652483344078, + 0.0107542984187603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0890059843659401 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a vibrant, '90s Disney-style animation, a dynamic medium shot centers on a dramatic aerial standoff between two male blue Niltava birds, the camera panning smoothly to follow their frantic circular movements. Their feathers are a riot of bold, diaphanous colors: a brilliant, saturated cobalt blue covers their backs and wings, while their faces and throats are a deep, velvety black. Their large eyes narrow in pure aggression. They engage in a chaotic aerial fight, tumbling and spinning in a choreographed ballet of violence, viciously clawing at each other's breasts with sharp claws. As they struggle, individual blue and black feathers are shed and gracefully float downward in dreamy slow motion, catching the light. The entire conflict is set against a lush, hand-painted jungle backdrop, filled with oversized, stylized ferns and dreamy, softly glowing flowers, emitting an ethereal glow. Dramatic golden sunlight penetrates the dense foliage from above, illuminating the frantic movements and floating feathers, creating a tense, high-energy atmosphere. The birds' aggressive, shrill chirping seems to echo through the beautiful yet dangerous surroundings.", + "seed": 12278, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_12/index_5940_seed_12278_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04690555273555219, + "vlm_logprob": -0.04690555273555219, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.04690555273555219, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07764244079589844 + ], + "opposite_logprob_tokens": [ + -2.7026424407958984 + ], + "logprob_reward": -0.07764244079589844, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07764244079589844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18979793787002563, + 0.07842131704092026, + 0.05531960725784302, + 0.05046353489160538, + 0.06438317149877548, + 0.0959094986319542, + 0.08040756732225418, + 0.06815732270479202, + 0.12742461264133453, + 0.18971553444862366 + ], + "credit_map": [ + 0.019638188183307648, + 0.02455814741551876, + 0.016934817656874657, + 0.028161130845546722, + 0.03263581544160843, + 0.014685669913887978, + 0.024658851325511932, + 0.028525300323963165, + 0.008860995061695576, + 0.007746333256363869, + 0.007094507571309805, + 0.012480303645133972, + 0.008964748121798038, + 0.008958247490227222, + 0.013141575269401073, + 0.011174609884619713, + 0.006354242563247681, + 0.010176214389503002, + 0.005880380515009165, + 0.005619295872747898, + 0.010404307395219803, + 0.005248157773166895, + 0.0049795834347605705, + 0.006657424848526716, + 0.005952254869043827, + 0.004290045239031315, + 0.005968706216663122, + 0.0056607648730278015, + 0.004567261319607496, + 0.004677094519138336, + 0.012166318483650684, + 0.007181087043136358, + 0.008048280142247677, + 0.00891463365405798, + 0.007643161807209253, + 0.012280761264264584, + 0.0041876984760165215, + 0.004733012989163399, + 0.008997933007776737, + 0.009577685967087746, + 0.013608829118311405, + 0.012343033216893673, + 0.01207818929105997, + 0.015855485573410988, + 0.007027449086308479, + 0.010140822269022465, + 0.01334996335208416, + 0.0115057323127985, + 0.007028760388493538, + 0.0040400209836661816, + 0.006695420015603304, + 0.011280756443738937, + 0.01964045874774456, + 0.008363606408238411, + 0.0084677180275321, + 0.014890821650624275, + 0.008338620886206627, + 0.004890044219791889, + 0.005195566453039646, + 0.008498799987137318, + 0.011154825799167156, + 0.0079856151714921, + 0.013399140909314156, + 0.008694714866578579, + 0.02039525844156742, + 0.008473541587591171, + 0.008894497528672218, + 0.009166701696813107, + 0.014252662658691406, + 0.02622854709625244, + 0.012191572226583958, + 0.027821822091937065, + 0.06033261865377426, + 0.012705940753221512, + 0.011093588545918465, + 0.020365677773952484, + 0.01132810115814209, + 0.00823872908949852, + 0.03477153182029724, + 0.03087933175265789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.07764244079589844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001053493469953537 + ], + "opposite_logprob_tokens": [ + -6.876053333282471 + ], + "logprob_reward": -0.001053493469953537, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001053493469953537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11664172261953354, + 0.0759185329079628, + 0.08231997489929199, + 0.05991450697183609, + 0.041776739060878754, + 0.057716287672519684, + 0.05344630405306816, + 0.015478702262043953, + 0.032421208918094635, + 0.46436598896980286 + ], + "credit_map": [ + 0.011336765252053738, + 0.017017081379890442, + 0.01123090274631977, + 0.01651659794151783, + 0.01896299421787262, + 0.009626797400414944, + 0.015366589650511742, + 0.016583995893597603, + 0.007219206541776657, + 0.006118517369031906, + 0.006263094488531351, + 0.022557590156793594, + 0.006312881130725145, + 0.01300155557692051, + 0.0072760917246341705, + 0.007169603835791349, + 0.01016097329556942, + 0.015345703810453415, + 0.009721627458930016, + 0.009061703458428383, + 0.010363152250647545, + 0.008294272236526012, + 0.009401868097484112, + 0.009970690123736858, + 0.007847410626709461, + 0.0068014757707715034, + 0.0060461838729679585, + 0.007140619680285454, + 0.0076076434925198555, + 0.00868288055062294, + 0.009001955389976501, + 0.006786345969885588, + 0.006015637423843145, + 0.006906690075993538, + 0.004146061837673187, + 0.006094549782574177, + 0.0030172362457960844, + 0.0031949901022017, + 0.005598223302513361, + 0.006803350523114204, + 0.006802695337682962, + 0.008033310063183308, + 0.009092302061617374, + 0.00947759672999382, + 0.006435445509850979, + 0.003755287267267704, + 0.008066345006227493, + 0.006053305231034756, + 0.006065642926841974, + 0.0028270455077290535, + 0.005439257249236107, + 0.007165819406509399, + 0.012806660495698452, + 0.004704830702394247, + 0.00523673091083765, + 0.009200320579111576, + 0.0014803747180849314, + 0.0008049967000260949, + 0.0017340928316116333, + 0.0022669488098472357, + 0.001768376911059022, + 0.002929223235696554, + 0.0017468496225774288, + 0.0027478414122015238, + 0.00527117820456624, + 0.0028538592159748077, + 0.002238650806248188, + 0.002644775900989771, + 0.0042518205009400845, + 0.00299241510219872, + 0.003615479450672865, + 0.0085530336946249, + 0.12271137535572052, + 0.01909494213759899, + 0.02212296798825264, + 0.029348282143473625, + 0.015682833269238472, + 0.017858795821666718, + 0.06382723897695541, + 0.1737195998430252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.001053493469953537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338880881667137 + ], + "opposite_logprob_tokens": [ + -3.651338815689087 + ], + "logprob_reward": -0.026338880881667137, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338880881667137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07067601382732391, + 0.06797077506780624, + 0.07572006434202194, + 0.129493847489357, + 0.03877473622560501, + 0.06979285925626755, + 0.2585582137107849, + 0.056660111993551254, + 0.06678971648216248, + 0.165563702583313 + ], + "credit_map": [ + 0.008712335489690304, + 0.008708380162715912, + 0.005000874865800142, + 0.009698634967207909, + 0.006205462384968996, + 0.0053786300122737885, + 0.010558230802416801, + 0.01641346886754036, + 0.009224704466760159, + 0.010112878866493702, + 0.00615153182297945, + 0.011335081420838833, + 0.006662290543317795, + 0.00929426122456789, + 0.008891301229596138, + 0.006298729218542576, + 0.010320954956114292, + 0.01037166453897953, + 0.011642520315945148, + 0.005674354266375303, + 0.007981918752193451, + 0.012206406332552433, + 0.007637673523277044, + 0.009884570725262165, + 0.01369625236839056, + 0.014235205948352814, + 0.013572669588029385, + 0.015818994492292404, + 0.015072863548994064, + 0.021079853177070618, + 0.01824839971959591, + 0.017769595608115196, + 0.006633501034229994, + 0.00479960348457098, + 0.004040875472128391, + 0.008731921203434467, + 0.0028569300193339586, + 0.002657044678926468, + 0.00421871105208993, + 0.00483615230768919, + 0.009477915242314339, + 0.008477390743792057, + 0.010879010893404484, + 0.009359229356050491, + 0.007869884371757507, + 0.005749042145907879, + 0.008884880691766739, + 0.009095503017306328, + 0.027231628075242043, + 0.015554969199001789, + 0.02284846641123295, + 0.03911052644252777, + 0.06178227439522743, + 0.025682561099529266, + 0.027678707614541054, + 0.038669075816869736, + 0.004985169507563114, + 0.003199916332960129, + 0.00961978454142809, + 0.010376734659075737, + 0.004867877345532179, + 0.007435867562890053, + 0.009480654262006283, + 0.006694109179079533, + 0.009092963300645351, + 0.004811891354620457, + 0.003905749646946788, + 0.005769392475485802, + 0.009234760887920856, + 0.014134630560874939, + 0.006128826178610325, + 0.01371149905025959, + 0.03659968823194504, + 0.009745625779032707, + 0.01665910892188549, + 0.02298044227063656, + 0.0052767288871109486, + 0.007664896547794342, + 0.019291386008262634, + 0.047345828264951706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.026338880881667137 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024846892338246107 + ], + "opposite_logprob_tokens": [ + -6.0024847984313965 + ], + "logprob_reward": -0.0024846892338246107, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024846892338246107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05292212590575218, + 0.08366121351718903, + 0.12041006237268448, + 0.03300701454281807, + 0.07015016674995422, + 0.03995485231280327, + 0.06643598526716232, + 0.09187249094247818, + 0.05382055416703224, + 0.3877655565738678 + ], + "credit_map": [ + 0.004843304865062237, + 0.00713507691398263, + 0.0047015962190926075, + 0.007679591421037912, + 0.009114730171859264, + 0.004714937414973974, + 0.006163901183754206, + 0.00856898631900549, + 0.012188798747956753, + 0.010128582827746868, + 0.00922667421400547, + 0.007539261598140001, + 0.008625374175608158, + 0.01590675488114357, + 0.011736813001334667, + 0.0083089554682374, + 0.0141688147559762, + 0.021128060296177864, + 0.018191587179899216, + 0.012166137807071209, + 0.015332354232668877, + 0.015797434374690056, + 0.010054431855678558, + 0.013571247458457947, + 0.003265767591074109, + 0.0035809099208563566, + 0.0037372612860053778, + 0.0034944924991577864, + 0.0032868930138647556, + 0.0046523441560566425, + 0.006309310905635357, + 0.004680035635828972, + 0.011855591088533401, + 0.011713151820003986, + 0.0063104890286922455, + 0.0102634122595191, + 0.004650567192584276, + 0.004007298499345779, + 0.011583292856812477, + 0.009766362607479095, + 0.00549494381994009, + 0.0051327235996723175, + 0.0043725199066102505, + 0.006380927283316851, + 0.003828420303761959, + 0.003507177112624049, + 0.006513011176139116, + 0.004725129343569279, + 0.00783009734004736, + 0.0028289593756198883, + 0.006018167827278376, + 0.010194330476224422, + 0.006349001545459032, + 0.008863757364451885, + 0.004891746211796999, + 0.019459929317235947, + 0.010763135738670826, + 0.0049352520145475864, + 0.022589296102523804, + 0.015231591649353504, + 0.007882123813033104, + 0.008653081022202969, + 0.0070629604160785675, + 0.014755051583051682, + 0.007882400415837765, + 0.00584775535389781, + 0.006989439018070698, + 0.007747569587081671, + 0.004754003137350082, + 0.00746567640453577, + 0.004252893850207329, + 0.008880814537405968, + 0.11448575556278229, + 0.016083629801869392, + 0.016161305829882622, + 0.01484609954059124, + 0.009275838732719421, + 0.020647965371608734, + 0.044101499021053314, + 0.15216346085071564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.0024846892338246107 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects male blue Niltava birds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the birds have cobalt blue backs/wings and deep black faces/throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the birds engaged in a chaotic aerial fight involving tumbling, spinning, and clawing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a lush jungle with oversized stylized ferns and softly glowing flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to follow the birds' circular movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12700825929641724 + ], + "opposite_logprob_tokens": [ + -2.1270081996917725 + ], + "logprob_reward": -0.12700825929641724, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12700825929641724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049537837505340576, + 0.06894078105688095, + 0.23024709522724152, + 0.08314038813114166, + 0.0507671982049942, + 0.06787111610174179, + 0.15588359534740448, + 0.08518797904253006, + 0.0667572021484375, + 0.14166682958602905 + ], + "credit_map": [ + 0.005860562901943922, + 0.005319579038769007, + 0.003968806937336922, + 0.005760681349784136, + 0.006031618919223547, + 0.003334593726322055, + 0.007368519902229309, + 0.011893476359546185, + 0.01311496365815401, + 0.008450944907963276, + 0.008738414384424686, + 0.005857289303094149, + 0.00628372048959136, + 0.01441198866814375, + 0.005458638072013855, + 0.006624813657253981, + 0.026426425203680992, + 0.03892238438129425, + 0.03182576596736908, + 0.022362109273672104, + 0.028345610946416855, + 0.033910661935806274, + 0.020363502204418182, + 0.02809063345193863, + 0.008607146330177784, + 0.00844679307192564, + 0.008207576349377632, + 0.009539881721138954, + 0.011134970001876354, + 0.013153157196938992, + 0.011675289832055569, + 0.012375568971037865, + 0.006902479100972414, + 0.0059760683216154575, + 0.0038813643623143435, + 0.007495288737118244, + 0.0034849122166633606, + 0.005715057719498873, + 0.008676629513502121, + 0.008635400794446468, + 0.0077684600837528706, + 0.009418395347893238, + 0.008870575577020645, + 0.010306007228791714, + 0.008164992555975914, + 0.005595744587481022, + 0.009712708182632923, + 0.008034227415919304, + 0.013059962540864944, + 0.008058805018663406, + 0.011508025228977203, + 0.02372138202190399, + 0.037264011800289154, + 0.017501432448625565, + 0.017314326018095016, + 0.0274556502699852, + 0.009506603702902794, + 0.005145510192960501, + 0.01957421936094761, + 0.01411386113613844, + 0.010112176649272442, + 0.007421643007546663, + 0.007823382504284382, + 0.011490577831864357, + 0.00904541090130806, + 0.0047091105952858925, + 0.008132322691380978, + 0.008345209062099457, + 0.006374672055244446, + 0.008667566813528538, + 0.007460152730345726, + 0.014022755436599255, + 0.04673248156905174, + 0.011009152047336102, + 0.008187911473214626, + 0.009064015932381153, + 0.006008821073919535, + 0.007030806038528681, + 0.005844683386385441, + 0.04778895154595375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.12700825929641724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_tensor([2676, 2676])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_tensor([2676, 2676])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..39146fb140424f497192f66842bd41a90929d3b1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_tensor([2676, 2676])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, low-angle, extreme close-up tracking shot synchronizes with the knobby tire of a mountain bike as it spins at breakneck speed, dominating the frame. The deep, aggressive tread of the black rubber is caked with a thick, dark brown slurry, which violently sprays a chaotic stream of mud and dirt toward the camera as the wheel spins at a dizzying pace. Individual droplets escape the tire, catching the cool, diffuse light filtering through the dense forest canopy. Some appear as sharp, gleaming projectiles, while others form a fine, mist-like spray. A soft, artistic blur of damp soil, decaying ochre leaves, and dark tree trunks in the background creates an extremely shallow depth of field, isolating the raw action. The entire scene is rendered with a subtle film grain, enhancing the gritty textures and conveying a powerful, vibrant off-road atmosphere.", + "seed": 21919, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_2676_seed_21919_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02825056177098304, + "vlm_logprob": -0.02825056177098304, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.02825056177098304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021613650023937225 + ], + "opposite_logprob_tokens": [ + -6.896613597869873 + ], + "logprob_reward": -0.021613650023937225, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.021613650023937225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27996331453323364, + 0.07314673066139221, + 0.04311024770140648, + 0.19310659170150757, + 0.060669682919979095, + 0.053108733147382736, + 0.05467965081334114, + 0.050599440932273865, + 0.06829610466957092, + 0.12331952154636383 + ], + "credit_map": [ + 0.02661033719778061, + 0.05765954777598381, + 0.0364568792283535, + 0.018175844103097916, + 0.06523016840219498, + 0.031114373356103897, + 0.021170470863580704, + 0.023545674979686737, + 0.004983108025044203, + 0.01042831502854824, + 0.007446363102644682, + 0.008016090840101242, + 0.014542247168719769, + 0.009868254885077477, + 0.007212015334516764, + 0.010650337673723698, + 0.0039693256840109825, + 0.00584843335673213, + 0.005672393832355738, + 0.004065010230988264, + 0.003556167706847191, + 0.007712755352258682, + 0.0059477416798472404, + 0.006338418927043676, + 0.01561865583062172, + 0.01637106016278267, + 0.021592117846012115, + 0.037438128143548965, + 0.021416980773210526, + 0.016996363177895546, + 0.037951018661260605, + 0.025722268968820572, + 0.006204523146152496, + 0.00792920496314764, + 0.00658371252939105, + 0.006173555739223957, + 0.010676401667296886, + 0.006217069458216429, + 0.010636888444423676, + 0.006248326972126961, + 0.006610213313251734, + 0.006191721651703119, + 0.008588713593780994, + 0.005312612280249596, + 0.00821507629007101, + 0.0033399323001503944, + 0.006637673359364271, + 0.00821278989315033, + 0.006371939554810524, + 0.005674025043845177, + 0.008829155936837196, + 0.006654034834355116, + 0.006287627387791872, + 0.008578067645430565, + 0.005823860410600901, + 0.006460939534008503, + 0.013861946761608124, + 0.005626523867249489, + 0.009113491512835026, + 0.004370535258203745, + 0.004694181960076094, + 0.0055876136757433414, + 0.004422928672283888, + 0.002922220854088664, + 0.007366926874965429, + 0.008880455046892166, + 0.008669722825288773, + 0.006093149073421955, + 0.013827670365571976, + 0.009991390630602837, + 0.006779130548238754, + 0.006687657907605171, + 0.00850916188210249, + 0.01791791059076786, + 0.013676726259291172, + 0.009006660431623459, + 0.02877992019057274, + 0.018058201298117638, + 0.00916027557104826, + 0.01821066066622734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.021613650023937225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048683051019907 + ], + "opposite_logprob_tokens": [ + -3.0486831665039062 + ], + "logprob_reward": -0.048683051019907, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048683051019907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23375920951366425, + 0.11667414754629135, + 0.10343935340642929, + 0.10781541466712952, + 0.07565387338399887, + 0.056707993149757385, + 0.0409412607550621, + 0.04551440477371216, + 0.14328835904598236, + 0.07620593160390854 + ], + "credit_map": [ + 0.02827594429254532, + 0.03410201147198677, + 0.0596206933259964, + 0.015514595434069633, + 0.029393158853054047, + 0.020971011370420456, + 0.02688196673989296, + 0.018999842926859856, + 0.008310669101774693, + 0.019468799233436584, + 0.009980721399188042, + 0.013348405249416828, + 0.022666022181510925, + 0.017268026247620583, + 0.012819832190871239, + 0.012811676599085331, + 0.012988453730940819, + 0.011758307926356792, + 0.01184675469994545, + 0.008401497267186642, + 0.011975332163274288, + 0.02150552161037922, + 0.014758632518351078, + 0.010204863734543324, + 0.009380844421684742, + 0.012249247170984745, + 0.01148420199751854, + 0.019650859758257866, + 0.012182535603642464, + 0.011521301232278347, + 0.019616426900029182, + 0.011730006895959377, + 0.005800772923976183, + 0.010838647373020649, + 0.009307809174060822, + 0.008304833434522152, + 0.01552280131727457, + 0.008540989831089973, + 0.0104233892634511, + 0.006914636120200157, + 0.006758814677596092, + 0.009052176959812641, + 0.010303334333002567, + 0.005804018117487431, + 0.005311412271112204, + 0.004514491185545921, + 0.0054443287663161755, + 0.009519421495497227, + 0.005296709015965462, + 0.0030830600298941135, + 0.00436185160651803, + 0.005774328485131264, + 0.005149089265614748, + 0.008379883132874966, + 0.00294517888687551, + 0.005951159168034792, + 0.008876214735209942, + 0.00822497345507145, + 0.006094881799072027, + 0.004636105615645647, + 0.004849080462008715, + 0.005423138849437237, + 0.0046425070613622665, + 0.002767506754025817, + 0.01700529083609581, + 0.016616832464933395, + 0.01806074008345604, + 0.00949128344655037, + 0.025465110316872597, + 0.016261251643300056, + 0.015597610734403133, + 0.024790244176983833, + 0.007937563583254814, + 0.011741137132048607, + 0.008460585959255695, + 0.006116949021816254, + 0.015425399877130985, + 0.009128418751060963, + 0.006770151201635599, + 0.010625733062624931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.048683051019907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048654209822416306 + ], + "opposite_logprob_tokens": [ + -3.048654317855835 + ], + "logprob_reward": -0.048654209822416306, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654209822416306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14597776532173157, + 0.06854895502328873, + 0.23848889768123627, + 0.10163280367851257, + 0.06506817787885666, + 0.03856239840388298, + 0.05766424536705017, + 0.06314384192228317, + 0.1013878583908081, + 0.11952503770589828 + ], + "credit_map": [ + 0.013645150698721409, + 0.014192177914083004, + 0.047383394092321396, + 0.010106470435857773, + 0.0173588115721941, + 0.014367913827300072, + 0.01646505296230316, + 0.012458797544240952, + 0.006103534251451492, + 0.011625888757407665, + 0.005887435749173164, + 0.007386296521872282, + 0.013398238457739353, + 0.009265145286917686, + 0.007297624368220568, + 0.007584796752780676, + 0.025385988876223564, + 0.038541000336408615, + 0.0224557314068079, + 0.018863286823034286, + 0.02696738764643669, + 0.04329230263829231, + 0.029911315068602562, + 0.03307189419865608, + 0.010608070529997349, + 0.012811589054763317, + 0.011650080792605877, + 0.01624361239373684, + 0.012529482133686543, + 0.013337157666683197, + 0.013174330815672874, + 0.011278479360044003, + 0.006975897122174501, + 0.008348473347723484, + 0.007434305269271135, + 0.010000815615057945, + 0.006473669782280922, + 0.008736185729503632, + 0.011979554779827595, + 0.005119276233017445, + 0.004890217445790768, + 0.005126013420522213, + 0.0057976809330284595, + 0.003856187453493476, + 0.005383126903325319, + 0.00356496125459671, + 0.00484336307272315, + 0.005100849084556103, + 0.007043509744107723, + 0.006676475051790476, + 0.007267506327480078, + 0.0051995981484651566, + 0.008685613051056862, + 0.008727834559977055, + 0.006813675630837679, + 0.007250036578625441, + 0.011651437729597092, + 0.011802264489233494, + 0.007525877095758915, + 0.00764407217502594, + 0.008289552293717861, + 0.00618220679461956, + 0.0068980976939201355, + 0.0031503387726843357, + 0.011299334466457367, + 0.014512860216200352, + 0.011633826419711113, + 0.0076972185634076595, + 0.017195099964737892, + 0.016400771215558052, + 0.01105933915823698, + 0.011589410714805126, + 0.009691638872027397, + 0.017920605838298798, + 0.014763200655579567, + 0.00802908930927515, + 0.02331123687326908, + 0.021120350807905197, + 0.008616136386990547, + 0.016072792932391167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.048654209822416306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017157370457425714 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.0017157370457425714, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017157370457425714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21498064696788788, + 0.0464668869972229, + 0.22123172879219055, + 0.04505803436040878, + 0.033706191927194595, + 0.05280851572751999, + 0.03143058344721794, + 0.01720236800611019, + 0.02412111684679985, + 0.3129939138889313 + ], + "credit_map": [ + 0.017320692539215088, + 0.02516166679561138, + 0.04333191365003586, + 0.01720324717462063, + 0.03567865118384361, + 0.023878883570432663, + 0.0358993299305439, + 0.016506291925907135, + 0.005240398459136486, + 0.006516984198242426, + 0.004865357186645269, + 0.004311795812100172, + 0.009433473460376263, + 0.00560163427144289, + 0.0053389472886919975, + 0.00515830097720027, + 0.018094675615429878, + 0.07109054177999496, + 0.01673094742000103, + 0.0077507649548351765, + 0.01567736640572548, + 0.012045067735016346, + 0.007799840997904539, + 0.0720425546169281, + 0.0034083316568285227, + 0.003443666035309434, + 0.005246206186711788, + 0.005411945749074221, + 0.0090194595977664, + 0.004781015682965517, + 0.007771576289087534, + 0.005975836422294378, + 0.002712148241698742, + 0.005485938861966133, + 0.003998628351837397, + 0.0034370983485132456, + 0.007559069432318211, + 0.004145358689129353, + 0.003543615574017167, + 0.002824336290359497, + 0.005871512461453676, + 0.007046214770525694, + 0.008920715190470219, + 0.005877178627997637, + 0.006235154345631599, + 0.005763346329331398, + 0.007712608203291893, + 0.005381792783737183, + 0.0035924625117331743, + 0.00279581593349576, + 0.0034667355939745903, + 0.002649203408509493, + 0.004713891074061394, + 0.007529016118496656, + 0.003373793326318264, + 0.0033096675761044025, + 0.003548557171598077, + 0.0021622099447995424, + 0.0021985273342579603, + 0.0016801252495497465, + 0.001602075295522809, + 0.0023259378504008055, + 0.0027254570741206408, + 0.0009594815201126039, + 0.001980528701096773, + 0.00364287500269711, + 0.0036628595553338528, + 0.0023438541684299707, + 0.0040992991998791695, + 0.0034178062342107296, + 0.0028153040912002325, + 0.0021585917565971613, + 0.022234516218304634, + 0.05167368799448013, + 0.033450450748205185, + 0.020671948790550232, + 0.07223233580589294, + 0.04640117287635803, + 0.02427808567881584, + 0.042051736265420914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.0017157370457425714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586160942912102 + ], + "opposite_logprob_tokens": [ + -3.8955862522125244 + ], + "logprob_reward": -0.020586160942912102, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586160942912102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4399136006832123, + 0.060047946870326996, + 0.1377374231815338, + 0.030792970210313797, + 0.03555603697896004, + 0.03555242344737053, + 0.020744575187563896, + 0.023543192073702812, + 0.04159431532025337, + 0.1745174527168274 + ], + "credit_map": [ + 0.03808319941163063, + 0.07869567722082138, + 0.05401704087853432, + 0.030839314684271812, + 0.11174197494983673, + 0.03642193228006363, + 0.05495133623480797, + 0.03516312316060066, + 0.005999213084578514, + 0.011366656050086021, + 0.00629384582862258, + 0.004031049553304911, + 0.015711070969700813, + 0.005642677191644907, + 0.005045812111347914, + 0.005957623012363911, + 0.013292104005813599, + 0.02741270139813423, + 0.01636139117181301, + 0.009360278956592083, + 0.011158302426338196, + 0.014840036630630493, + 0.012304378673434258, + 0.03300821781158447, + 0.002405633917078376, + 0.0022168285213410854, + 0.0034526207018643618, + 0.003569432068616152, + 0.0058044851757586, + 0.003174755023792386, + 0.005645882338285446, + 0.004523332696408033, + 0.0037287366576492786, + 0.004742102697491646, + 0.0045377532951533794, + 0.0033133781980723143, + 0.005068827886134386, + 0.003902462311089039, + 0.006500511895865202, + 0.003762264037504792, + 0.0049036298878490925, + 0.003868120489642024, + 0.005220555700361729, + 0.004087889101356268, + 0.004618922248482704, + 0.0034921930637210608, + 0.004426212515681982, + 0.00493489857763052, + 0.003300779964774847, + 0.0018186138477176428, + 0.0030349623411893845, + 0.0023959067184478045, + 0.002044958993792534, + 0.002902749925851822, + 0.002226276323199272, + 0.0030203277710825205, + 0.004856754094362259, + 0.0026185137685388327, + 0.005438711494207382, + 0.0029803933575749397, + 0.0019166108686476946, + 0.002560424618422985, + 0.0019032369600608945, + 0.001268549356609583, + 0.003970450256019831, + 0.006359733175486326, + 0.0059548537246882915, + 0.003903093049302697, + 0.00787428393959999, + 0.005057003814727068, + 0.004339816980063915, + 0.0041350810788571835, + 0.01207904051989317, + 0.02873201295733452, + 0.016702672466635704, + 0.01034382451325655, + 0.041726063936948776, + 0.02802382968366146, + 0.009610709734261036, + 0.027299314737319946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.020586160942912102 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, extreme close-up tracking shot synchronizes with the knobby tire of a mountain bike as it spins at breakneck speed, dominating the frame. The deep, aggressive tread of the black rubber is caked with a thick, dark brown slurry, which violently sprays a chaotic stream of mud and dirt toward the camera as the wheel spins at a dizzying pace. Individual droplets escape the tire, catching the cool, diffuse light filtering through the dense forest canopy. Some appear as sharp, gleaming projectiles, while others form a fine, mist-like spray. A soft, artistic blur of damp soil, decaying ochre leaves, and dark tree trunks in the background creates an extremely shallow depth of field, isolating the raw action. The entire scene is rendered with a subtle film grain, enhancing the gritty textures and conveying a powerful, vibrant off-road atmosphere.", + "seed": 21920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_2676_seed_21920_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023670858005061746, + "vlm_logprob": -0.023670858005061746, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.023670858005061746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016784166917204857 + ], + "opposite_logprob_tokens": [ + -7.266784191131592 + ], + "logprob_reward": -0.016784166917204857, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.016784166917204857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061115194112062454, + 0.08351520448923111, + 0.02517266944050789, + 0.11935652047395706, + 0.30149346590042114, + 0.0774712860584259, + 0.044176239520311356, + 0.04772396385669708, + 0.18753907084465027, + 0.05243648216128349 + ], + "credit_map": [ + 0.005334686487913132, + 0.005797217600047588, + 0.010578802786767483, + 0.009009066969156265, + 0.008119228295981884, + 0.011120093055069447, + 0.005057990085333586, + 0.006098106969147921, + 0.011003552004694939, + 0.015744797885417938, + 0.011359339579939842, + 0.010101335123181343, + 0.009315895847976208, + 0.008084436878561974, + 0.011635924689471722, + 0.006269922945648432, + 0.0029461286030709743, + 0.003064094576984644, + 0.003285213140770793, + 0.003735495964065194, + 0.002708540065214038, + 0.002354071009904146, + 0.00430346978828311, + 0.0027756562922149897, + 0.015118921175599098, + 0.01044059731066227, + 0.02179829776287079, + 0.012469779700040817, + 0.01992768980562687, + 0.015579119324684143, + 0.005047401878982782, + 0.018974710255861282, + 0.01869744062423706, + 0.01211774256080389, + 0.10126254707574844, + 0.013157729990780354, + 0.11755938827991486, + 0.018827855587005615, + 0.01118036825209856, + 0.008690407499670982, + 0.007544280961155891, + 0.003986159339547157, + 0.006123299710452557, + 0.0047951131127774715, + 0.012412606738507748, + 0.014859924092888832, + 0.006096620112657547, + 0.021653279662132263, + 0.006940912920981646, + 0.007905660197138786, + 0.0038243066519498825, + 0.006908678449690342, + 0.0037037055008113384, + 0.005320678930729628, + 0.004357429686933756, + 0.005214866250753403, + 0.00620874622836709, + 0.004961702506989241, + 0.005678942427039146, + 0.006124922540038824, + 0.006119218654930592, + 0.007330721244215965, + 0.00468086451292038, + 0.006618845276534557, + 0.02994808927178383, + 0.018813123926520348, + 0.016507577151060104, + 0.02734854258596897, + 0.02126227132976055, + 0.02811567857861519, + 0.017694346606731415, + 0.027849428355693817, + 0.010315478779375553, + 0.0038064660038799047, + 0.008172557689249516, + 0.005226670298725367, + 0.009227585978806019, + 0.005925524514168501, + 0.004194579087197781, + 0.005567618645727634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.016784166917204857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305660113692284 + ], + "opposite_logprob_tokens": [ + -3.773305654525757 + ], + "logprob_reward": -0.023305660113692284, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305660113692284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17069609463214874, + 0.10342475026845932, + 0.04522756114602089, + 0.06732895225286484, + 0.3291252851486206, + 0.04037090390920639, + 0.016005560755729675, + 0.055450741201639175, + 0.15474185347557068, + 0.017628369852900505 + ], + "credit_map": [ + 0.012417742051184177, + 0.022540606558322906, + 0.018134556710720062, + 0.026054581627249718, + 0.028979500755667686, + 0.021713264286518097, + 0.020827004685997963, + 0.02002882771193981, + 0.013260903768241405, + 0.015038225799798965, + 0.013743242248892784, + 0.012756811454892159, + 0.010562579147517681, + 0.007968680001795292, + 0.02090909704566002, + 0.009185206145048141, + 0.006746827624738216, + 0.003963545896112919, + 0.005492616444826126, + 0.004561825655400753, + 0.0052254986949265, + 0.005881131626665592, + 0.004953495226800442, + 0.008402625098824501, + 0.007065073121339083, + 0.005796446464955807, + 0.017416151240468025, + 0.00771092576906085, + 0.006979533936828375, + 0.010742003098130226, + 0.0050806268118321896, + 0.0065381890162825584, + 0.01925054006278515, + 0.011820686049759388, + 0.10655544698238373, + 0.016560647636651993, + 0.12279589474201202, + 0.019303690642118454, + 0.021610939875245094, + 0.01122741773724556, + 0.009174232371151447, + 0.00324341282248497, + 0.003940653055906296, + 0.005235864315181971, + 0.006291060242801905, + 0.0031451110262423754, + 0.002786529017612338, + 0.006554038729518652, + 0.0021454289089888334, + 0.002728479914367199, + 0.0016439452301710844, + 0.0015926227206364274, + 0.0015132029075175524, + 0.0014766021631658077, + 0.0020726409275084734, + 0.002832637866958976, + 0.006146383937448263, + 0.005379516165703535, + 0.005578096956014633, + 0.00542772701010108, + 0.01037240494042635, + 0.008037867955863476, + 0.0037848115898668766, + 0.010723935440182686, + 0.02683544158935547, + 0.014172517694532871, + 0.013676449656486511, + 0.03265443071722984, + 0.017181770876049995, + 0.01638314314186573, + 0.014005718752741814, + 0.019832374528050423, + 0.0027348652947694063, + 0.0014414226170629263, + 0.0019958249758929014, + 0.0019445179495960474, + 0.003262906102463603, + 0.0020020685624331236, + 0.001834617112763226, + 0.0024121482856571674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.023305660113692284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305163770914078 + ], + "opposite_logprob_tokens": [ + -3.1680517196655273 + ], + "logprob_reward": -0.04305163770914078, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305163770914078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05155276507139206, + 0.08195987343788147, + 0.06294846534729004, + 0.07248332351446152, + 0.3871992230415344, + 0.06099964678287506, + 0.01798051781952381, + 0.036831334233284, + 0.2168266475200653, + 0.011218223720788956 + ], + "credit_map": [ + 0.004528738558292389, + 0.004882083740085363, + 0.008922677487134933, + 0.01015904638916254, + 0.004915970843285322, + 0.006839750334620476, + 0.00638936460018158, + 0.004915136378258467, + 0.011440434493124485, + 0.009182403795421124, + 0.008241066709160805, + 0.008807417936623096, + 0.011739729903638363, + 0.009277611039578915, + 0.014698411338031292, + 0.008572804741561413, + 0.008300703018903732, + 0.005955039989203215, + 0.013395752757787704, + 0.007943974807858467, + 0.006464286707341671, + 0.004622411448508501, + 0.009608743712306023, + 0.006657554768025875, + 0.007889228872954845, + 0.008868416771292686, + 0.01296780165284872, + 0.010119685903191566, + 0.01608484797179699, + 0.008483115583658218, + 0.003990867640823126, + 0.004079360980540514, + 0.024709191173315048, + 0.01350132655352354, + 0.13148818910121918, + 0.01573590375483036, + 0.14867058396339417, + 0.022205904126167297, + 0.019561540335416794, + 0.011326607316732407, + 0.012701530940830708, + 0.0049888151697814465, + 0.006698677781969309, + 0.007519632112234831, + 0.007849914021790028, + 0.007486050948500633, + 0.0036583354230970144, + 0.010096690617501736, + 0.0026745290961116552, + 0.0031255038920789957, + 0.0017686274368315935, + 0.002132742665708065, + 0.001713309669867158, + 0.0017769169062376022, + 0.0021692842710763216, + 0.0026196036487817764, + 0.004635861609131098, + 0.0030910579953342676, + 0.0037721097469329834, + 0.0032303279731422663, + 0.006562252528965473, + 0.0068009174428880215, + 0.004702110309153795, + 0.004036696162074804, + 0.03554149344563484, + 0.022259924560785294, + 0.01834855042397976, + 0.03600604459643364, + 0.026541465893387794, + 0.030113155022263527, + 0.01701316423714161, + 0.031002825126051903, + 0.001996314385905862, + 0.0011564160231500864, + 0.0018152524717152119, + 0.0013739061541855335, + 0.0015666437102481723, + 0.001112892758101225, + 0.001112684840336442, + 0.0010841137263923883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.04305163770914078 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015149556566029787 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015149556566029787, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015149556566029787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14160864055156708, + 0.08349891006946564, + 0.024360807612538338, + 0.2431017905473709, + 0.25366321206092834, + 0.027825308963656425, + 0.02823491580784321, + 0.02987624704837799, + 0.0824725553393364, + 0.08535757660865784 + ], + "credit_map": [ + 0.011793128214776516, + 0.013175778090953827, + 0.025876041501760483, + 0.03183742240071297, + 0.012853951193392277, + 0.01563098281621933, + 0.015197702683508396, + 0.015243636444211006, + 0.012437469325959682, + 0.01150733046233654, + 0.007559801917523146, + 0.009690848179161549, + 0.013246743008494377, + 0.008854909799993038, + 0.013049070723354816, + 0.007152742240577936, + 0.0028461625333875418, + 0.002850499004125595, + 0.0018999899039044976, + 0.0026341313496232033, + 0.004069739021360874, + 0.003790733404457569, + 0.0035297235008329153, + 0.0027398280799388885, + 0.02331448160111904, + 0.021633761003613472, + 0.056143440306186676, + 0.04080590605735779, + 0.039868567138910294, + 0.02836494892835617, + 0.011395481415092945, + 0.021575191989541054, + 0.01571037620306015, + 0.01735350489616394, + 0.07172252237796783, + 0.013855748809874058, + 0.07951053977012634, + 0.023503869771957397, + 0.01866554468870163, + 0.013341126032173634, + 0.005419403314590454, + 0.0027689130511134863, + 0.0027392262127250433, + 0.0032023184467107058, + 0.003377889283001423, + 0.003900395007804036, + 0.002325850771740079, + 0.004091312643140554, + 0.004577523097395897, + 0.006847284268587828, + 0.0026131547056138515, + 0.0038899474311619997, + 0.0027390047907829285, + 0.002500280737876892, + 0.0021567970979958773, + 0.002910922048613429, + 0.0035847770050168037, + 0.0036191970575600863, + 0.003676603315398097, + 0.003185584442690015, + 0.002610479947179556, + 0.005542956758290529, + 0.0031780945137143135, + 0.0044785537756979465, + 0.015266794711351395, + 0.0072584147565066814, + 0.008031667210161686, + 0.02234802022576332, + 0.00771161587908864, + 0.005357416812330484, + 0.007770287338644266, + 0.00872834026813507, + 0.010764394886791706, + 0.006888014264404774, + 0.013267760165035725, + 0.008110652677714825, + 0.01768607273697853, + 0.012773144990205765, + 0.009726263582706451, + 0.006141272839158773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0015149556566029787 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369786962866783 + ], + "opposite_logprob_tokens": [ + -3.4086978435516357 + ], + "logprob_reward": -0.03369786962866783, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369786962866783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13521869480609894, + 0.10338319838047028, + 0.04345747455954552, + 0.1698950231075287, + 0.31323665380477905, + 0.03729987516999245, + 0.020779797807335854, + 0.030246347188949585, + 0.04038756713271141, + 0.10609535872936249 + ], + "credit_map": [ + 0.012609974481165409, + 0.010516527108848095, + 0.025577886030077934, + 0.030297884717583656, + 0.01115141250193119, + 0.018632272258400917, + 0.01106969453394413, + 0.015363034792244434, + 0.014126006513834, + 0.016664909198880196, + 0.008225884288549423, + 0.01099376194179058, + 0.018292248249053955, + 0.010271507315337658, + 0.01677715964615345, + 0.008031724952161312, + 0.0068267169408500195, + 0.0060246773064136505, + 0.004738000687211752, + 0.004306657239794731, + 0.004779295530170202, + 0.004788560327142477, + 0.004963223822414875, + 0.00703034084290266, + 0.01918395794928074, + 0.009263661690056324, + 0.047220297157764435, + 0.02597174048423767, + 0.027187705039978027, + 0.017744150012731552, + 0.005496789235621691, + 0.01782671920955181, + 0.02007756568491459, + 0.013734659180045128, + 0.10542070120573044, + 0.013554156757891178, + 0.118024542927742, + 0.020042387768626213, + 0.011339765973389149, + 0.011042860336601734, + 0.007299786899238825, + 0.0033907503820955753, + 0.0035041377414017916, + 0.0039785983972251415, + 0.005939235910773277, + 0.003923533484339714, + 0.002481333678588271, + 0.006782500073313713, + 0.0029715965501964092, + 0.004660856910049915, + 0.0019976457115262747, + 0.002815880347043276, + 0.0022352784872055054, + 0.0017836886690929532, + 0.001945607247762382, + 0.0023692448157817125, + 0.0029253384564071894, + 0.0033007836900651455, + 0.003004755126312375, + 0.004459758754819632, + 0.002433352405205369, + 0.0053697614930570126, + 0.0026501507963985205, + 0.006102446932345629, + 0.005958840250968933, + 0.0037241545505821705, + 0.003512825584039092, + 0.006920419633388519, + 0.005523973610252142, + 0.0035172358620911837, + 0.006721370853483677, + 0.004508750978857279, + 0.022600017488002777, + 0.008902403526008129, + 0.008767740800976753, + 0.010134324431419373, + 0.02047584392130375, + 0.011729463934898376, + 0.012371612712740898, + 0.01111395750194788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.03369786962866783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, extreme close-up tracking shot synchronizes with the knobby tire of a mountain bike as it spins at breakneck speed, dominating the frame. The deep, aggressive tread of the black rubber is caked with a thick, dark brown slurry, which violently sprays a chaotic stream of mud and dirt toward the camera as the wheel spins at a dizzying pace. Individual droplets escape the tire, catching the cool, diffuse light filtering through the dense forest canopy. Some appear as sharp, gleaming projectiles, while others form a fine, mist-like spray. A soft, artistic blur of damp soil, decaying ochre leaves, and dark tree trunks in the background creates an extremely shallow depth of field, isolating the raw action. The entire scene is rendered with a subtle film grain, enhancing the gritty textures and conveying a powerful, vibrant off-road atmosphere.", + "seed": 21921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_2676_seed_21921_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023327162582427265, + "vlm_logprob": -0.023327162582427265, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.023327162582427265, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01698588952422142 + ], + "opposite_logprob_tokens": [ + -7.016985893249512 + ], + "logprob_reward": -0.01698588952422142, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.01698588952422142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17904159426689148, + 0.058087196201086044, + 0.23285417258739471, + 0.06669624149799347, + 0.04173387214541435, + 0.03923835977911949, + 0.08785603195428848, + 0.08585142344236374, + 0.07782524079084396, + 0.1308158040046692 + ], + "credit_map": [ + 0.021451381966471672, + 0.02515716850757599, + 0.022904690355062485, + 0.033023323863744736, + 0.014262553304433823, + 0.01453438587486744, + 0.02969380095601082, + 0.018014298751950264, + 0.005911804269999266, + 0.01120847463607788, + 0.00804722961038351, + 0.004942023660987616, + 0.008105889894068241, + 0.0050371079705655575, + 0.005816012155264616, + 0.00901865866035223, + 0.023945583030581474, + 0.012201005592942238, + 0.011064017191529274, + 0.011574430391192436, + 0.012392790988087654, + 0.0844574049115181, + 0.009190019220113754, + 0.06802894175052643, + 0.006129239685833454, + 0.011341194622218609, + 0.007593206129968166, + 0.0071627115830779076, + 0.0073740482330322266, + 0.0072149960324168205, + 0.01117037981748581, + 0.008710465393960476, + 0.002869023708626628, + 0.008105138316750526, + 0.004804401658475399, + 0.005300888791680336, + 0.0066457889042794704, + 0.005612735170871019, + 0.0056470055133104324, + 0.0027488924097269773, + 0.004954480566084385, + 0.005043397657573223, + 0.005597326438874006, + 0.006054610945284367, + 0.0039451452903449535, + 0.003797149285674095, + 0.004752952139824629, + 0.005093304440379143, + 0.009964179247617722, + 0.010709132999181747, + 0.0154848238453269, + 0.014537937007844448, + 0.007795772515237331, + 0.006722570862621069, + 0.0136862862855196, + 0.008955338038504124, + 0.014449767768383026, + 0.01156788319349289, + 0.010793752036988735, + 0.009045958518981934, + 0.012479862198233604, + 0.011156166903674603, + 0.008786863647401333, + 0.007571175694465637, + 0.006750568747520447, + 0.016820887103676796, + 0.015244340524077415, + 0.009201067499816418, + 0.009565853513777256, + 0.0071985311806201935, + 0.007102676201611757, + 0.005941327195614576, + 0.021087748929858208, + 0.028006862848997116, + 0.021860066801309586, + 0.013593015260994434, + 0.013330450281500816, + 0.015100139193236828, + 0.007919340394437313, + 0.009918204508721828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.01698588952422142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331264689564705 + ], + "opposite_logprob_tokens": [ + -3.773312568664551 + ], + "logprob_reward": -0.02331264689564705, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331264689564705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08402872085571289, + 0.024864383041858673, + 0.3684059977531433, + 0.10098972171545029, + 0.06603164225816727, + 0.02465130016207695, + 0.08931885659694672, + 0.06845379620790482, + 0.04209213703870773, + 0.13116338849067688 + ], + "credit_map": [ + 0.009608713909983635, + 0.017895707860589027, + 0.009684828110039234, + 0.012779874727129936, + 0.00706862099468708, + 0.006903008092194796, + 0.011201945133507252, + 0.008886026218533516, + 0.0024925635661929846, + 0.0042024762369692326, + 0.0034641542006284, + 0.00283258524723351, + 0.00291773141361773, + 0.002203898271545768, + 0.00305978418327868, + 0.00369118875823915, + 0.03887663036584854, + 0.017652247101068497, + 0.01641063578426838, + 0.01610233634710312, + 0.022031206637620926, + 0.13494661450386047, + 0.014902227558195591, + 0.10748409479856491, + 0.01547123584896326, + 0.01872299425303936, + 0.008317979983985424, + 0.010315109975636005, + 0.01028641127049923, + 0.011208525858819485, + 0.011624488048255444, + 0.015042973682284355, + 0.004530488979071379, + 0.015553341247141361, + 0.00700768269598484, + 0.010744457133114338, + 0.0055043925531208515, + 0.01249915361404419, + 0.006390837952494621, + 0.0038012873847037554, + 0.002828555414453149, + 0.0024915251415222883, + 0.003228355897590518, + 0.0029049979057163, + 0.00192496576346457, + 0.0022834541741758585, + 0.004939648322761059, + 0.004049796145409346, + 0.013373682275414467, + 0.010699496604502201, + 0.024299707263708115, + 0.010096464306116104, + 0.009489602409303188, + 0.006075710989534855, + 0.008559918031096458, + 0.006724273320287466, + 0.00712739210575819, + 0.008827068842947483, + 0.0069153145886957645, + 0.0072862980887293816, + 0.007301864214241505, + 0.00833764299750328, + 0.013062385842204094, + 0.009595828130841255, + 0.003792900824919343, + 0.004941874649375677, + 0.011958939023315907, + 0.004239754285663366, + 0.003827969776466489, + 0.004494499880820513, + 0.004868815187364817, + 0.003967382479459047, + 0.013775801286101341, + 0.017941800877451897, + 0.022902796044945717, + 0.021541375666856766, + 0.014822269789874554, + 0.01642375998198986, + 0.012186386622488499, + 0.011569200083613396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.02331264689564705 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865023493766785 + ], + "opposite_logprob_tokens": [ + -3.0486502647399902 + ], + "logprob_reward": -0.04865023493766785, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865023493766785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16775891184806824, + 0.04358083754777908, + 0.2332000583410263, + 0.16403156518936157, + 0.06483025848865509, + 0.04269612208008766, + 0.06135612353682518, + 0.042717356234788895, + 0.07648541033267975, + 0.1033434048295021 + ], + "credit_map": [ + 0.016703477129340172, + 0.038065917789936066, + 0.02441057376563549, + 0.015994874760508537, + 0.015858706086874008, + 0.016978023573756218, + 0.02133399248123169, + 0.018413331359624863, + 0.004050500225275755, + 0.008890134282410145, + 0.005191120319068432, + 0.004365952219814062, + 0.007404386531561613, + 0.004116385243833065, + 0.005766029469668865, + 0.003796327393501997, + 0.023558087646961212, + 0.012605829164385796, + 0.01340873260051012, + 0.011687430553138256, + 0.022811034694314003, + 0.0747511014342308, + 0.014197775162756443, + 0.06018006429076195, + 0.02250768058001995, + 0.012370020151138306, + 0.023661505430936813, + 0.02363874576985836, + 0.017607999965548515, + 0.018973397091031075, + 0.022958479821681976, + 0.022313736379146576, + 0.004509311635047197, + 0.008246331475675106, + 0.009986449033021927, + 0.010833918116986752, + 0.007555101532489061, + 0.010066024027764797, + 0.008762001991271973, + 0.004871114157140255, + 0.004830814432352781, + 0.004618180450052023, + 0.005689367186278105, + 0.004630052484571934, + 0.004173224326223135, + 0.004181017633527517, + 0.007583638187497854, + 0.006989826448261738, + 0.008465049788355827, + 0.007849941961467266, + 0.011604315601289272, + 0.010188030079007149, + 0.00462589506059885, + 0.004168509040027857, + 0.0076606180518865585, + 0.006793765351176262, + 0.003953495528548956, + 0.0058687361888587475, + 0.004892588593065739, + 0.005996967200189829, + 0.005289382766932249, + 0.0057722278870642185, + 0.005620438139885664, + 0.005323520861566067, + 0.007071762345731258, + 0.00834093987941742, + 0.022349631413817406, + 0.00768804457038641, + 0.007013274822384119, + 0.008234835229814053, + 0.008109391666948795, + 0.0076775336638092995, + 0.013360042124986649, + 0.014579558745026588, + 0.016118252649903297, + 0.013773199170827866, + 0.012375856749713421, + 0.01261261384934187, + 0.01237430702894926, + 0.008149570785462856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.04865023493766785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013379440642893314 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013379440642893314, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013379440642893314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12301850318908691, + 0.042317550629377365, + 0.13934341073036194, + 0.1390160471200943, + 0.08795630931854248, + 0.030843013897538185, + 0.22128674387931824, + 0.120073601603508, + 0.03447878733277321, + 0.06166597828269005 + ], + "credit_map": [ + 0.015299901366233826, + 0.03264399617910385, + 0.016945598646998405, + 0.00954541377723217, + 0.011773201636970043, + 0.010593635961413383, + 0.00986151397228241, + 0.016355235129594803, + 0.0037744042929261923, + 0.007341690827161074, + 0.00640180753543973, + 0.004115561489015818, + 0.006901816930621862, + 0.0036667571403086185, + 0.00415072962641716, + 0.005964784882962704, + 0.01442014891654253, + 0.007039655931293964, + 0.005632044281810522, + 0.006516970694065094, + 0.008442109450697899, + 0.049719229340553284, + 0.005794204771518707, + 0.0417790561914444, + 0.014990496449172497, + 0.014424593187868595, + 0.01319137867540121, + 0.020968856289982796, + 0.021181240677833557, + 0.01434808038175106, + 0.02309349551796913, + 0.016817912459373474, + 0.00834939256310463, + 0.018116461113095284, + 0.011854477226734161, + 0.013946603052318096, + 0.008605091832578182, + 0.009972781874239445, + 0.010121668688952923, + 0.006989836227148771, + 0.003290823195129633, + 0.004231920465826988, + 0.0037652330938726664, + 0.00438675656914711, + 0.0037515300791710615, + 0.003497207770124078, + 0.004010908305644989, + 0.003908636514097452, + 0.029845962300896645, + 0.03353355824947357, + 0.04402308166027069, + 0.03509505093097687, + 0.01802983507514, + 0.012198016978800297, + 0.02287762612104416, + 0.02568361908197403, + 0.013568976894021034, + 0.010496295057237148, + 0.014861700125038624, + 0.012863654643297195, + 0.015892526134848595, + 0.015215080231428146, + 0.016918111592531204, + 0.0202572513371706, + 0.003884686157107353, + 0.004736834205687046, + 0.007629901636391878, + 0.0036914017982780933, + 0.0035622613504529, + 0.003995642997324467, + 0.004051887895911932, + 0.002926170825958252, + 0.00928841345012188, + 0.010243282653391361, + 0.008310474455356598, + 0.005693141836673021, + 0.008867119438946247, + 0.006564196664839983, + 0.008727772161364555, + 0.003971575293689966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.0013379440642893314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634909749031067 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.02634909749031067, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634909749031067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08185437321662903, + 0.056113313883543015, + 0.1498005986213684, + 0.10366843640804291, + 0.06861617416143417, + 0.11807192862033844, + 0.15301501750946045, + 0.09151912480592728, + 0.08607599884271622, + 0.09126501530408859 + ], + "credit_map": [ + 0.008863707073032856, + 0.01597759500145912, + 0.014163230545818806, + 0.014175516553223133, + 0.007120742928236723, + 0.0061409990303218365, + 0.008438603952527046, + 0.0069739739410579205, + 0.0050416141748428345, + 0.009133725427091122, + 0.012315653264522552, + 0.004871607758104801, + 0.006833391264081001, + 0.006337631028145552, + 0.006107096094638109, + 0.005472593009471893, + 0.015398644842207432, + 0.007629509083926678, + 0.006625938229262829, + 0.007290819194167852, + 0.007901465520262718, + 0.054693568497896194, + 0.005178206134587526, + 0.0450824573636055, + 0.013366359286010265, + 0.010186215862631798, + 0.014581047929823399, + 0.01511972863227129, + 0.012996508739888668, + 0.010874603874981403, + 0.014334411360323429, + 0.012209558859467506, + 0.00911903940141201, + 0.00900403130799532, + 0.011709985323250294, + 0.011347302235662937, + 0.006991049740463495, + 0.007905279286205769, + 0.007131958846002817, + 0.005407527554780245, + 0.01384907029569149, + 0.013304642401635647, + 0.016213655471801758, + 0.01613445207476616, + 0.012709500268101692, + 0.01749502681195736, + 0.013682572171092033, + 0.014683004468679428, + 0.02405698038637638, + 0.018204426392912865, + 0.037583112716674805, + 0.024142680689692497, + 0.013021964579820633, + 0.01001701783388853, + 0.015212828293442726, + 0.010776006616652012, + 0.00936242938041687, + 0.012955648824572563, + 0.009692816995084286, + 0.0087243327870965, + 0.012360279448330402, + 0.010367187671363354, + 0.015875769779086113, + 0.012180660851299763, + 0.012494601309299469, + 0.01321623008698225, + 0.022350819781422615, + 0.007766394875943661, + 0.007365303114056587, + 0.00734963221475482, + 0.008820565417408943, + 0.006712455302476883, + 0.009413134306669235, + 0.011375723406672478, + 0.014091452583670616, + 0.016438394784927368, + 0.012885360978543758, + 0.011042439378798008, + 0.008664501830935478, + 0.007354008033871651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.02634909749031067 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, extreme close-up tracking shot synchronizes with the knobby tire of a mountain bike as it spins at breakneck speed, dominating the frame. The deep, aggressive tread of the black rubber is caked with a thick, dark brown slurry, which violently sprays a chaotic stream of mud and dirt toward the camera as the wheel spins at a dizzying pace. Individual droplets escape the tire, catching the cool, diffuse light filtering through the dense forest canopy. Some appear as sharp, gleaming projectiles, while others form a fine, mist-like spray. A soft, artistic blur of damp soil, decaying ochre leaves, and dark tree trunks in the background creates an extremely shallow depth of field, isolating the raw action. The entire scene is rendered with a subtle film grain, enhancing the gritty textures and conveying a powerful, vibrant off-road atmosphere.", + "seed": 21922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_13/index_2676_seed_21922_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022551801730878653, + "vlm_logprob": -0.022551801730878653, + "vlm_yes_no_margin": 4.974999952316284, + "vlm_token_credit": -0.022551801730878653, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0271486546844244 + ], + "opposite_logprob_tokens": [ + -7.152148723602295 + ], + "logprob_reward": -0.0271486546844244, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0271486546844244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10557744652032852, + 0.048349812626838684, + 0.05525172874331474, + 0.11921794712543488, + 0.09926514327526093, + 0.21795159578323364, + 0.05959559977054596, + 0.07219138741493225, + 0.1274077445268631, + 0.09519163519144058 + ], + "credit_map": [ + 0.009154685772955418, + 0.008197613060474396, + 0.012089243158698082, + 0.007844840176403522, + 0.008751532062888145, + 0.016036944463849068, + 0.017170237377285957, + 0.026332352310419083, + 0.0032226063776761293, + 0.006321537774056196, + 0.004987305961549282, + 0.005606363527476788, + 0.005642733536660671, + 0.005258258432149887, + 0.011394078843295574, + 0.005916933063417673, + 0.00513999629765749, + 0.010505089536309242, + 0.006591236684471369, + 0.007478019222617149, + 0.0068987892009317875, + 0.007545481435954571, + 0.007386195007711649, + 0.0037069215904921293, + 0.013559715822339058, + 0.018733855336904526, + 0.00467621348798275, + 0.007632293272763491, + 0.024268733337521553, + 0.021336177363991737, + 0.018708480522036552, + 0.01030246913433075, + 0.009010332636535168, + 0.016902653500437737, + 0.011331003159284592, + 0.014702114276587963, + 0.010005257092416286, + 0.012512348592281342, + 0.015474158339202404, + 0.00932727288454771, + 0.012845896184444427, + 0.08474092185497284, + 0.005729999393224716, + 0.07725382596254349, + 0.010573735460639, + 0.014433766715228558, + 0.006573186255991459, + 0.005800260230898857, + 0.012324707582592964, + 0.007710123434662819, + 0.010623965412378311, + 0.010470744222402573, + 0.005621264688670635, + 0.0037349893245846033, + 0.0051182713359594345, + 0.003991534002125263, + 0.017127003520727158, + 0.006087585352361202, + 0.019900521263480186, + 0.00909615308046341, + 0.004302870482206345, + 0.004195630084723234, + 0.007912272587418556, + 0.003569353139027953, + 0.014293445274233818, + 0.00952670257538557, + 0.02844521962106228, + 0.008427915163338184, + 0.011357898823916912, + 0.007052039727568626, + 0.039290931075811386, + 0.009013586677610874, + 0.011888384819030762, + 0.009705123491585255, + 0.010862348601222038, + 0.01345837488770485, + 0.015985660254955292, + 0.010310808196663857, + 0.013195566833019257, + 0.009785360656678677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.0271486546844244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332126535475254 + ], + "opposite_logprob_tokens": [ + -3.7733211517333984 + ], + "logprob_reward": -0.02332126535475254, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332126535475254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17136968672275543, + 0.06628100574016571, + 0.05471939593553543, + 0.19621245563030243, + 0.030660638585686684, + 0.1486867517232895, + 0.12618714570999146, + 0.12315835803747177, + 0.04761920124292374, + 0.03510534018278122 + ], + "credit_map": [ + 0.021698327735066414, + 0.014738673344254494, + 0.025704437866806984, + 0.013751701451838017, + 0.013313178904354572, + 0.017309749498963356, + 0.02760489657521248, + 0.037248726934194565, + 0.005781135521829128, + 0.011633914895355701, + 0.006236214656382799, + 0.009270415641367435, + 0.006976105272769928, + 0.006989168468862772, + 0.012982835061848164, + 0.0064112162217497826, + 0.0062567065469920635, + 0.006578330881893635, + 0.007328948006033897, + 0.005346899386495352, + 0.0105068888515234, + 0.00736180879175663, + 0.006843734532594681, + 0.004496075678616762, + 0.017445597797632217, + 0.016306068748235703, + 0.01088856253772974, + 0.00789300911128521, + 0.017309661954641342, + 0.06721607595682144, + 0.049072712659835815, + 0.010080788284540176, + 0.003557902295142412, + 0.0037192802410572767, + 0.005577784962952137, + 0.004473437089473009, + 0.0034754585940390825, + 0.004413173533976078, + 0.002965756691992283, + 0.002477844012901187, + 0.010587221011519432, + 0.0513114407658577, + 0.005204471293836832, + 0.05083170905709267, + 0.009396220557391644, + 0.012104647234082222, + 0.0048319390043616295, + 0.004419103730469942, + 0.017033737152814865, + 0.022337494418025017, + 0.03644707798957825, + 0.013944745995104313, + 0.010662826709449291, + 0.009471572935581207, + 0.009727259166538715, + 0.006562438793480396, + 0.036345914006233215, + 0.006969817914068699, + 0.03786538913846016, + 0.009609263390302658, + 0.004635943565517664, + 0.0057341535575687885, + 0.00915452279150486, + 0.012843353673815727, + 0.006292791571468115, + 0.0045397924259305, + 0.00922714825719595, + 0.004924085922539234, + 0.0037072263658046722, + 0.002896101213991642, + 0.012116334401071072, + 0.003915723878890276, + 0.004199544433504343, + 0.003854710143059492, + 0.004175960086286068, + 0.006030909717082977, + 0.0050124311819672585, + 0.003344267373904586, + 0.004507056903094053, + 0.003980461973696947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.02332126535475254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057918548583984 + ], + "opposite_logprob_tokens": [ + -3.168057918548584 + ], + "logprob_reward": -0.043057918548583984, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057918548583984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08069644123315811, + 0.028723012655973434, + 0.06258759647607803, + 0.0591491274535656, + 0.040314678102731705, + 0.2896777391433716, + 0.1937086284160614, + 0.14060771465301514, + 0.03417893871665001, + 0.07035612314939499 + ], + "credit_map": [ + 0.010661986656486988, + 0.0069351475685834885, + 0.016820736229419708, + 0.006406525149941444, + 0.007906590588390827, + 0.007332020439207554, + 0.011673850007355213, + 0.012959575280547142, + 0.002959229052066803, + 0.005647304002195597, + 0.0029006258118897676, + 0.0039117964915931225, + 0.002667234046384692, + 0.0027911376673728228, + 0.005339138675481081, + 0.0025065457448363304, + 0.005945520009845495, + 0.013035106472671032, + 0.005886860657483339, + 0.005328160710632801, + 0.009682937525212765, + 0.007221202831715345, + 0.009273813106119633, + 0.006213988643139601, + 0.006104189436882734, + 0.008068273775279522, + 0.0031109130941331387, + 0.004765165038406849, + 0.013630945235490799, + 0.008156752213835716, + 0.010772871784865856, + 0.004540008492767811, + 0.0034873115364462137, + 0.005047400016337633, + 0.005903357639908791, + 0.005709580145776272, + 0.0044202134013175964, + 0.008068078197538853, + 0.004318036139011383, + 0.0033606979995965958, + 0.019198505207896233, + 0.1045113205909729, + 0.008832152001559734, + 0.09961746633052826, + 0.01788783259689808, + 0.022523438557982445, + 0.009160074405372143, + 0.007946931757032871, + 0.025213999673724174, + 0.033075783401727676, + 0.055267393589019775, + 0.02079232782125473, + 0.01993032917380333, + 0.014874621294438839, + 0.014107441529631615, + 0.010446726344525814, + 0.0408182255923748, + 0.0072121452540159225, + 0.039642881602048874, + 0.010870818980038166, + 0.004283080343157053, + 0.007150185294449329, + 0.011947821825742722, + 0.018682541325688362, + 0.003686486976221204, + 0.004310642369091511, + 0.005499925930052996, + 0.003017507726326585, + 0.003249727189540863, + 0.0026431791484355927, + 0.007785502355545759, + 0.003985963761806488, + 0.008636127226054668, + 0.005485708825290203, + 0.00869709812104702, + 0.010593559592962265, + 0.010747863911092281, + 0.009643277153372765, + 0.008737041614949703, + 0.007815436460077763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.043057918548583984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010449193650856614 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010449193650856614, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010449193650856614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0734483003616333, + 0.04768960922956467, + 0.08568155765533447, + 0.07964994013309479, + 0.0761796161532402, + 0.07840806245803833, + 0.032260920852422714, + 0.3444864749908447, + 0.13357281684875488, + 0.04862270876765251 + ], + "credit_map": [ + 0.009893649257719517, + 0.006589301861822605, + 0.014727533794939518, + 0.00646662013605237, + 0.009378277696669102, + 0.009967107325792313, + 0.008853067643940449, + 0.007572739385068417, + 0.004703173413872719, + 0.0073716891929507256, + 0.005649395752698183, + 0.006193618755787611, + 0.0055353897623717785, + 0.0069718630984425545, + 0.006570771802216768, + 0.004693706054240465, + 0.00810287520289421, + 0.008197087794542313, + 0.01164800301194191, + 0.015154127962887287, + 0.0077366637997329235, + 0.01066063903272152, + 0.017448974773287773, + 0.006733184214681387, + 0.007755055092275143, + 0.010300145484507084, + 0.006061178166419268, + 0.007037819363176823, + 0.011311851441860199, + 0.01506014447659254, + 0.01387742254883051, + 0.008246325887739658, + 0.009592375718057156, + 0.013478693552315235, + 0.007489206735044718, + 0.008526995778083801, + 0.007404383737593889, + 0.014157785102725029, + 0.009493081830441952, + 0.006037088576704264, + 0.005170261021703482, + 0.02679169550538063, + 0.0039734686724841595, + 0.023013170808553696, + 0.004816497676074505, + 0.007515722420066595, + 0.00406769709661603, + 0.003059548791497946, + 0.004813062958419323, + 0.003700229339301586, + 0.006424556951969862, + 0.00445715244859457, + 0.003198350081220269, + 0.0030961802694946527, + 0.003286966122686863, + 0.0032844224479049444, + 0.12454687058925629, + 0.012638119980692863, + 0.14020869135856628, + 0.030213717371225357, + 0.007059785537421703, + 0.007592914160341024, + 0.013541307300329208, + 0.00868503749370575, + 0.017797768115997314, + 0.014629198238253593, + 0.030004356056451797, + 0.009727033786475658, + 0.010917969048023224, + 0.008084932342171669, + 0.03232859820127487, + 0.010082961991429329, + 0.006071497220546007, + 0.003876063274219632, + 0.004738499876111746, + 0.0080944849178195, + 0.007522789761424065, + 0.006357961799949408, + 0.007364416960626841, + 0.004596995655447245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.0010449193650856614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on the knobby tire of a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tire spinning at a very high speed while being caked with thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mud spray violently toward the camera as a chaotic stream of droplets and mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a shallow depth of field blur showing damp soil, ochre leaves, and dark tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cool, diffuse light filtering through a dense forest canopy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818625070154667 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.01818625070154667, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818625070154667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1415242999792099, + 0.07420384138822556, + 0.1570814698934555, + 0.044432345777750015, + 0.05405852571129799, + 0.18990887701511383, + 0.04575613886117935, + 0.11515780538320541, + 0.13135866820812225, + 0.046518076211214066 + ], + "credit_map": [ + 0.01660795882344246, + 0.013407054357230663, + 0.017307380214333534, + 0.009073429740965366, + 0.01547869574278593, + 0.018726466223597527, + 0.02276952564716339, + 0.028153786435723305, + 0.009465145878493786, + 0.008346333168447018, + 0.010954662226140499, + 0.010863637551665306, + 0.007232570089399815, + 0.009767350740730762, + 0.009331791661679745, + 0.0082423510029912, + 0.014395470730960369, + 0.02291210927069187, + 0.013133962638676167, + 0.021806878969073296, + 0.0227104052901268, + 0.021565500646829605, + 0.023214833810925484, + 0.017342302948236465, + 0.006539041642099619, + 0.004668072331696749, + 0.0029811139684170485, + 0.003100705798715353, + 0.006942431442439556, + 0.008923281915485859, + 0.007736695930361748, + 0.003541001584380865, + 0.0065099201165139675, + 0.008087466470897198, + 0.007382208481431007, + 0.006420442368835211, + 0.005143358372151852, + 0.010186244733631611, + 0.0057325283996760845, + 0.004596361890435219, + 0.011435423046350479, + 0.07100339978933334, + 0.004926811903715134, + 0.06485298275947571, + 0.01042828056961298, + 0.014065852388739586, + 0.007878034375607967, + 0.005318101495504379, + 0.00717152701690793, + 0.00438203988596797, + 0.007972324267029762, + 0.005824477411806583, + 0.0050112409517169, + 0.005047523882240057, + 0.00579338101670146, + 0.004553622100502253, + 0.016343645751476288, + 0.015537156723439693, + 0.012149330228567123, + 0.014094647951424122, + 0.00955993589013815, + 0.00912550836801529, + 0.022116677835583687, + 0.016230901703238487, + 0.011784805916249752, + 0.014934109523892403, + 0.023762894794344902, + 0.008975735865533352, + 0.011995390057563782, + 0.00804443284869194, + 0.04197874292731285, + 0.009882546029984951, + 0.005237220320850611, + 0.003953287843614817, + 0.005345736630260944, + 0.007035546936094761, + 0.008068268187344074, + 0.005381790455430746, + 0.0069503472186625, + 0.004545879550278187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.01818625070154667 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_14/index_tensor([3652, 3652])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_14/index_tensor([3652, 3652])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..058847c8d9c9c182030aea84eccd37d0f2eb7fcd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_14/index_tensor([3652, 3652])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot, the camera slowly and steadily pans forward, focusing on the summit of a beacon tower on the ancient Great Wall. A massive crimson battle flag flutters wildly in the wind. Crafted from heavy linen, its edges frayed by age, the flag is embroidered with a vivid dragon totem in golden thread, its scales shimmering in the light. The relentless wind tears at the flag, sometimes unfurling with a snapping sound, fully revealing the unyielding dragon, and sometimes curling tightly around the weathered, dark-wood pole. The composition of light and shadow possesses Annie Leibovitz's signature quality: the evening sun hangs low in the sky, its golden rays penetrating thick, swirling clouds at a sharp angle, precisely illuminating the windward side of the flag, making its deep red appear exceptionally saturated and the gold embroidery gleam with a metallic sheen. Meanwhile, the other side of the flag and the massive stone brick walls of the beacon tower are shrouded in deep, cold shadows, creating a strong contrast between light and dark, highlighting the rough texture of the stone and the vicissitudes of history. The backdrop is a rolling, rugged mountain range, gradually merging into the gray-blue twilight in the distance. The entire scene, in this extreme contrast of movement and stillness, is permeated with an epic atmosphere of solemnity, loneliness, and historical power.", + "seed": 16929, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_14/index_3652_seed_16929_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0367110344581306, + "vlm_logprob": -0.0367110344581306, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0367110344581306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09068449586629868 + ], + "opposite_logprob_tokens": [ + -3.090684413909912 + ], + "logprob_reward": -0.09068449586629868, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.09068449586629868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29763174057006836, + 0.11264633387327194, + 0.04824058338999748, + 0.05367382988333702, + 0.049441296607255936, + 0.04470447078347206, + 0.0676053836941719, + 0.10484519600868225, + 0.16402582824230194, + 0.0571853369474411 + ], + "credit_map": [ + 0.026022056117653847, + 0.023416509851813316, + 0.04487679898738861, + 0.07356324046850204, + 0.014735700562596321, + 0.012137433513998985, + 0.04129188880324364, + 0.061588119715452194, + 0.009616510011255741, + 0.009280259720981121, + 0.016403721645474434, + 0.029162153601646423, + 0.007698965724557638, + 0.00471285218372941, + 0.016088297590613365, + 0.019683577120304108, + 0.0047307866625487804, + 0.003951951861381531, + 0.007510435301810503, + 0.008041756227612495, + 0.005378506146371365, + 0.0031753084622323513, + 0.007724415510892868, + 0.007727423682808876, + 0.0067210388369858265, + 0.00497583718970418, + 0.012989262118935585, + 0.007343715988099575, + 0.0049843210726976395, + 0.0035545898135751486, + 0.0061385491862893105, + 0.006966518238186836, + 0.003961237147450447, + 0.0045923469588160515, + 0.006132060196250677, + 0.008217631839215755, + 0.0076786791905760765, + 0.0037692247424274683, + 0.007119040936231613, + 0.007971074432134628, + 0.0061629959382116795, + 0.004206936340779066, + 0.004959038458764553, + 0.0063401502557098866, + 0.00767479557543993, + 0.0036990256048738956, + 0.0056573692709207535, + 0.006004159804433584, + 0.008150621317327023, + 0.006898818537592888, + 0.007723268587142229, + 0.00937801506370306, + 0.013334574177861214, + 0.003753288649022579, + 0.008285021409392357, + 0.01008177176117897, + 0.012530846521258354, + 0.012303533963859081, + 0.016638247296214104, + 0.013823096640408039, + 0.020509658381342888, + 0.007417536806315184, + 0.007123726885765791, + 0.014498546719551086, + 0.02059340849518776, + 0.025467950850725174, + 0.021428178995847702, + 0.017097070813179016, + 0.026421304792165756, + 0.016043415293097496, + 0.014942198991775513, + 0.022032301872968674, + 0.008684451691806316, + 0.00891937781125307, + 0.008379545994102955, + 0.005609039682894945, + 0.010940850712358952, + 0.003870265558362007, + 0.005043490324169397, + 0.005738317500799894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.09068449586629868 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207595765590668 + ], + "opposite_logprob_tokens": [ + -2.8120758533477783 + ], + "logprob_reward": -0.06207595765590668, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207595765590668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10126736015081406, + 0.09015597403049469, + 0.06476736813783646, + 0.04336028918623924, + 0.07951647788286209, + 0.07170488685369492, + 0.10517924278974533, + 0.15189287066459656, + 0.13247577846050262, + 0.15967977046966553 + ], + "credit_map": [ + 0.0081849480047822, + 0.009775778278708458, + 0.011893552727997303, + 0.017440669238567352, + 0.01564439944922924, + 0.006614083889871836, + 0.015590480528771877, + 0.016123443841934204, + 0.010717072524130344, + 0.01009244192391634, + 0.013530814088881016, + 0.017978742718696594, + 0.008801322430372238, + 0.0058083301410079, + 0.010593684390187263, + 0.012633568607270718, + 0.006134201306849718, + 0.006268181838095188, + 0.008759391494095325, + 0.00771753815934062, + 0.01026388630270958, + 0.004540432710200548, + 0.01315616350620985, + 0.007927574217319489, + 0.00574718602001667, + 0.004088412504643202, + 0.010151649825274944, + 0.007303276099264622, + 0.002388270106166601, + 0.002267012372612953, + 0.004498307593166828, + 0.006916174199432135, + 0.010796567425131798, + 0.007734083104878664, + 0.0157617274671793, + 0.011862136423587799, + 0.007687677163630724, + 0.0038455112371593714, + 0.01072999369353056, + 0.01109878346323967, + 0.006578052882105112, + 0.005984662100672722, + 0.010537121444940567, + 0.01292586699128151, + 0.010297928936779499, + 0.0055895838886499405, + 0.008428698405623436, + 0.01136296708136797, + 0.011736413463950157, + 0.012368805706501007, + 0.012100224383175373, + 0.011216586455702782, + 0.02675165981054306, + 0.00694821123033762, + 0.009297945536673069, + 0.014759398996829987, + 0.01914726011455059, + 0.01787038892507553, + 0.022707168012857437, + 0.027745962142944336, + 0.02714141272008419, + 0.008906624279916286, + 0.010593157261610031, + 0.01778089441359043, + 0.0168311670422554, + 0.02809191681444645, + 0.016655772924423218, + 0.015712760388851166, + 0.020645465701818466, + 0.009873054921627045, + 0.009330738335847855, + 0.0153349032625556, + 0.027046088129281998, + 0.027436310425400734, + 0.028634129092097282, + 0.013746366836130619, + 0.029649240896105766, + 0.010482321493327618, + 0.010262643918395042, + 0.012422679923474789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.06207595765590668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194561824202538 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.018194561824202538, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194561824202538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17067787051200867, + 0.08131258189678192, + 0.049247171729803085, + 0.03356365114450455, + 0.02210674062371254, + 0.07586027681827545, + 0.05387702211737633, + 0.041492536664009094, + 0.05812814459204674, + 0.4137340188026428 + ], + "credit_map": [ + 0.016033580526709557, + 0.016515113413333893, + 0.02278437465429306, + 0.025059910491108894, + 0.025327609851956367, + 0.014153740368783474, + 0.020708652213215828, + 0.030094895511865616, + 0.00720570795238018, + 0.008757579140365124, + 0.01071687787771225, + 0.011600024066865444, + 0.012350929901003838, + 0.005387214478105307, + 0.013997020199894905, + 0.011297229677438736, + 0.006792299449443817, + 0.0057064238935709, + 0.0066695366986095905, + 0.006134968250989914, + 0.006749561522156, + 0.002891266020014882, + 0.007954679429531097, + 0.006348438560962677, + 0.0036760771181434393, + 0.003216654760763049, + 0.00662469957023859, + 0.00447655376046896, + 0.003180474042892456, + 0.0017565343296155334, + 0.005231022834777832, + 0.005401637405157089, + 0.0029794820584356785, + 0.0026108413003385067, + 0.003303695935755968, + 0.0029220746364444494, + 0.002908503171056509, + 0.0019783859606832266, + 0.00263409037142992, + 0.0027696709148585796, + 0.007795125246047974, + 0.009033863432705402, + 0.007697693537920713, + 0.007880263961851597, + 0.019160814583301544, + 0.009042835794389248, + 0.007904564030468464, + 0.007345122750848532, + 0.006937412545084953, + 0.007437505759298801, + 0.008235440589487553, + 0.010347056202590466, + 0.0076206703670322895, + 0.003162522567436099, + 0.0046334294602274895, + 0.005502993706613779, + 0.00478425482288003, + 0.004175799433141947, + 0.006768676917999983, + 0.008160560391843319, + 0.006973397918045521, + 0.0025540811475366354, + 0.002873565535992384, + 0.005202207248657942, + 0.006315130274742842, + 0.013850064016878605, + 0.008433024398982525, + 0.00640585832297802, + 0.008715854026377201, + 0.0033611857797950506, + 0.0047656092792749405, + 0.006281420588493347, + 0.05496842414140701, + 0.07728835940361023, + 0.06168307736515999, + 0.0350213348865509, + 0.10362351685762405, + 0.03124476596713066, + 0.022253431379795074, + 0.02765113115310669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.018194561824202538 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978577509522438 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.00978577509522438, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978577509522438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18838298320770264, + 0.06182890385389328, + 0.14304019510746002, + 0.0340757742524147, + 0.1959487646818161, + 0.061904892325401306, + 0.07073662430047989, + 0.04159948229789734, + 0.14984607696533203, + 0.05263632908463478 + ], + "credit_map": [ + 0.02606596238911152, + 0.019374700263142586, + 0.038317520171403885, + 0.02924540266394615, + 0.01612313836812973, + 0.017926078289747238, + 0.0186696145683527, + 0.02266058139503002, + 0.007632282096892595, + 0.0076233320869505405, + 0.008216740563511848, + 0.009610707871615887, + 0.00800755713135004, + 0.0026692303363233805, + 0.007442183792591095, + 0.010626870207488537, + 0.022106187418103218, + 0.015103285200893879, + 0.02493646740913391, + 0.028093229979276657, + 0.010857658460736275, + 0.005776572041213512, + 0.01809600368142128, + 0.01807078905403614, + 0.0045724473893642426, + 0.0034971616696566343, + 0.005243363324552774, + 0.004651444498449564, + 0.0035027838312089443, + 0.0020904247649013996, + 0.004956472665071487, + 0.005561677739024162, + 0.03044465184211731, + 0.019334955140948296, + 0.06171368062496185, + 0.03463831916451454, + 0.010055096819996834, + 0.005650550127029419, + 0.012647728435695171, + 0.021463792771100998, + 0.007029353175312281, + 0.0064225089736282825, + 0.011882358230650425, + 0.008690944872796535, + 0.010839691385626793, + 0.004113467875868082, + 0.0059800781309604645, + 0.006946492940187454, + 0.00817616656422615, + 0.008947673253715038, + 0.011213364079594612, + 0.008185386657714844, + 0.014213989488780499, + 0.004730605520308018, + 0.005858335178345442, + 0.009411109611392021, + 0.004841457586735487, + 0.0064020659774541855, + 0.006868144031614065, + 0.004686002153903246, + 0.006054254714399576, + 0.0039952523075044155, + 0.002886800328269601, + 0.005865507759153843, + 0.01887117326259613, + 0.02121187560260296, + 0.02647460624575615, + 0.014626939781010151, + 0.024780340492725372, + 0.01634032651782036, + 0.011383056640625, + 0.016157755628228188, + 0.00932207703590393, + 0.008650420233607292, + 0.007424493320286274, + 0.005804015789180994, + 0.009186767973005772, + 0.0032939082011580467, + 0.003086426528170705, + 0.0058682202361524105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.00978577509522438 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028143818490207195 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.0028143818490207195, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028143818490207195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2983747124671936, + 0.1308375597000122, + 0.08432209491729736, + 0.05330829694867134, + 0.0560707189142704, + 0.07216887921094894, + 0.057374149560928345, + 0.0837470144033432, + 0.0634675920009613, + 0.1003289669752121 + ], + "credit_map": [ + 0.02853027731180191, + 0.02442309446632862, + 0.043721798807382584, + 0.06117212027311325, + 0.02910066768527031, + 0.017319731414318085, + 0.03548260033130646, + 0.058624427765607834, + 0.015186447650194168, + 0.011803953908383846, + 0.021558331325650215, + 0.02394823357462883, + 0.01596909947693348, + 0.007860306650400162, + 0.015384349972009659, + 0.019126834347844124, + 0.010564981028437614, + 0.00782287772744894, + 0.020743392407894135, + 0.012798141688108444, + 0.007165883202105761, + 0.0043515232391655445, + 0.010228793136775494, + 0.010646499693393707, + 0.005900902673602104, + 0.005643075332045555, + 0.007662246935069561, + 0.004440381191670895, + 0.008368042297661304, + 0.004872446414083242, + 0.009614873677492142, + 0.006806325167417526, + 0.008135590702295303, + 0.0060422420501708984, + 0.008693244308233261, + 0.011110452003777027, + 0.004757275804877281, + 0.003494696691632271, + 0.006013663951307535, + 0.007823551073670387, + 0.0086573651060462, + 0.0070807067677378654, + 0.011930166743695736, + 0.009358154609799385, + 0.01476163137704134, + 0.004583042114973068, + 0.0070982943288981915, + 0.008699520491063595, + 0.007020918186753988, + 0.006375390104949474, + 0.0078046685084700584, + 0.011985821649432182, + 0.0070057702250778675, + 0.003216927871108055, + 0.00586364883929491, + 0.008101006038486958, + 0.009453578852117062, + 0.013954782858490944, + 0.014174899086356163, + 0.011241539381444454, + 0.011380854994058609, + 0.006572995334863663, + 0.0068323081359267235, + 0.010136047378182411, + 0.008396449498832226, + 0.008313717320561409, + 0.010691342875361443, + 0.007890853099524975, + 0.00904526375234127, + 0.005628008395433426, + 0.005790697876363993, + 0.0077112531289458275, + 0.019138673320412636, + 0.012471416965126991, + 0.020574303343892097, + 0.00796907301992178, + 0.020367540419101715, + 0.006468529347330332, + 0.005985933821648359, + 0.007353493943810463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0028143818490207195 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly and steadily pans forward, focusing on the summit of a beacon tower on the ancient Great Wall. A massive crimson battle flag flutters wildly in the wind. Crafted from heavy linen, its edges frayed by age, the flag is embroidered with a vivid dragon totem in golden thread, its scales shimmering in the light. The relentless wind tears at the flag, sometimes unfurling with a snapping sound, fully revealing the unyielding dragon, and sometimes curling tightly around the weathered, dark-wood pole. The composition of light and shadow possesses Annie Leibovitz's signature quality: the evening sun hangs low in the sky, its golden rays penetrating thick, swirling clouds at a sharp angle, precisely illuminating the windward side of the flag, making its deep red appear exceptionally saturated and the gold embroidery gleam with a metallic sheen. Meanwhile, the other side of the flag and the massive stone brick walls of the beacon tower are shrouded in deep, cold shadows, creating a strong contrast between light and dark, highlighting the rough texture of the stone and the vicissitudes of history. The backdrop is a rolling, rugged mountain range, gradually merging into the gray-blue twilight in the distance. The entire scene, in this extreme contrast of movement and stillness, is permeated with an epic atmosphere of solemnity, loneliness, and historical power.", + "seed": 16930, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_14/index_3652_seed_16930_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04553960575722158, + "vlm_logprob": -0.04553960575722158, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.04553960575722158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09555672109127045 + ], + "opposite_logprob_tokens": [ + -2.7205567359924316 + ], + "logprob_reward": -0.09555672109127045, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09555672109127045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1876315027475357, + 0.08628033846616745, + 0.026214713230729103, + 0.03043173812329769, + 0.11954108625650406, + 0.06302405893802643, + 0.27013206481933594, + 0.048901528120040894, + 0.07310422509908676, + 0.09473869949579239 + ], + "credit_map": [ + 0.017401784658432007, + 0.020219122990965843, + 0.02539980784058571, + 0.035693444311618805, + 0.012556497938930988, + 0.022174648940563202, + 0.022636402398347855, + 0.031549807637929916, + 0.006850159261375666, + 0.007429606281220913, + 0.009472577832639217, + 0.01347533892840147, + 0.008695750497281551, + 0.005453481804579496, + 0.021138563752174377, + 0.013764869421720505, + 0.0024998821318149567, + 0.0027920533902943134, + 0.0036712046712636948, + 0.004071109928190708, + 0.0033337105996906757, + 0.0018030876526609063, + 0.0045571099035441875, + 0.0034865569323301315, + 0.0021195136941969395, + 0.002359632169827819, + 0.003941219765692949, + 0.0041632018983364105, + 0.0035758214071393013, + 0.0025859952438622713, + 0.0063484543934464455, + 0.005337903276085854, + 0.04289295896887779, + 0.040787965059280396, + 0.006561897229403257, + 0.007007362321019173, + 0.006733023561537266, + 0.003412457648664713, + 0.006451265420764685, + 0.005694162100553513, + 0.006718376651406288, + 0.006398944184184074, + 0.006778567563742399, + 0.008314840495586395, + 0.010358603671193123, + 0.0035715564154088497, + 0.012093720957636833, + 0.008789455518126488, + 0.02812071703374386, + 0.03941302001476288, + 0.0331948921084404, + 0.029946761205792427, + 0.04861241206526756, + 0.018746547400951385, + 0.03516361489892006, + 0.03693413361907005, + 0.0060415854677557945, + 0.003862025449052453, + 0.009786004200577736, + 0.008323960937559605, + 0.004118192475289106, + 0.0019821557216346264, + 0.00619331793859601, + 0.008594292216002941, + 0.007999103516340256, + 0.00427614338696003, + 0.009654602967202663, + 0.01015468779951334, + 0.005203627981245518, + 0.006013612262904644, + 0.009453481994569302, + 0.02034897357225418, + 0.007336597889661789, + 0.008834844455122948, + 0.014230525121092796, + 0.011870769783854485, + 0.010728521272540092, + 0.017157718539237976, + 0.010660858824849129, + 0.013918867334723473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.09555672109127045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034223645925522 + ], + "opposite_logprob_tokens": [ + -2.350342273712158 + ], + "logprob_reward": -0.10034223645925522, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034223645925522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12949034571647644, + 0.055373236536979675, + 0.023627489805221558, + 0.06441798806190491, + 0.07609706372022629, + 0.1183154284954071, + 0.14187383651733398, + 0.2180452048778534, + 0.0759190022945404, + 0.09684038162231445 + ], + "credit_map": [ + 0.018048951402306557, + 0.016387782990932465, + 0.014770974405109882, + 0.017461508512496948, + 0.014992259442806244, + 0.009154721163213253, + 0.02031368762254715, + 0.01836046576499939, + 0.004752133507281542, + 0.004905293695628643, + 0.006983646657317877, + 0.006821365095674992, + 0.008262768387794495, + 0.0045824493281543255, + 0.011064951308071613, + 0.008000630885362625, + 0.00241939933039248, + 0.002220363821834326, + 0.0025229640305042267, + 0.003607510356232524, + 0.0044639743864536285, + 0.0019627525471150875, + 0.0033670966513454914, + 0.0030634268186986446, + 0.004791432525962591, + 0.0051565030589699745, + 0.00833708606660366, + 0.008619002997875214, + 0.007368799299001694, + 0.005425602663308382, + 0.01499950047582388, + 0.009720059111714363, + 0.01944747567176819, + 0.018318798393011093, + 0.005970173515379429, + 0.007138664834201336, + 0.006737438961863518, + 0.003287451108917594, + 0.008683944121003151, + 0.00651311781257391, + 0.009913163259625435, + 0.009502589702606201, + 0.01644676737487316, + 0.016159500926733017, + 0.016767285764217377, + 0.007527157198637724, + 0.024618729948997498, + 0.01738024316728115, + 0.012317836284637451, + 0.011637961491942406, + 0.017326679080724716, + 0.019181109964847565, + 0.017070678994059563, + 0.010452689602971077, + 0.031191328540444374, + 0.022695554420351982, + 0.02453748509287834, + 0.012838541530072689, + 0.05854833498597145, + 0.023654785007238388, + 0.016009042039513588, + 0.00935311894863844, + 0.015320280566811562, + 0.05778359994292259, + 0.0081913061439991, + 0.004766461439430714, + 0.013324390165507793, + 0.006486006546765566, + 0.006330123636871576, + 0.00415335176512599, + 0.008131338283419609, + 0.024536022916436195, + 0.007887735031545162, + 0.006254614796489477, + 0.012885217554867268, + 0.011295348405838013, + 0.00829270202666521, + 0.013276499696075916, + 0.02251387946307659, + 0.014434383250772953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.10034223645925522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543711811304092 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543711811304092, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543711811304092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2028958946466446, + 0.06276100128889084, + 0.030047966167330742, + 0.02295502834022045, + 0.03960220888257027, + 0.03786221146583557, + 0.08102863281965256, + 0.3222677409648895, + 0.13802197575569153, + 0.0625574067234993 + ], + "credit_map": [ + 0.014631128869950771, + 0.017954159528017044, + 0.03078547865152359, + 0.02777428925037384, + 0.025065666064620018, + 0.019199389964342117, + 0.04005410894751549, + 0.02743164636194706, + 0.005808605346828699, + 0.005502647720277309, + 0.006034293211996555, + 0.007169338408857584, + 0.010258260183036327, + 0.004586954601109028, + 0.013188056647777557, + 0.010212840512394905, + 0.0032838191837072372, + 0.002989034866914153, + 0.0024232289288192987, + 0.00312443426810205, + 0.006658817175775766, + 0.002700117416679859, + 0.005004059057682753, + 0.0038644527085125446, + 0.0020467345602810383, + 0.0020920478273183107, + 0.0025471877306699753, + 0.0030901676509529352, + 0.0029688540380448103, + 0.0021022246219217777, + 0.004819558933377266, + 0.0032882513478398323, + 0.004490848630666733, + 0.0033637708984315395, + 0.004941270686686039, + 0.00533977709710598, + 0.0058715930208563805, + 0.004355647135525942, + 0.005768067203462124, + 0.005471231881529093, + 0.004292308818548918, + 0.003927267622202635, + 0.004001610912382603, + 0.003757916856557131, + 0.00588731886819005, + 0.003911321982741356, + 0.006959886755794287, + 0.005124578718096018, + 0.007053140550851822, + 0.010346353985369205, + 0.01138029620051384, + 0.009505112655460835, + 0.008868524804711342, + 0.006382642779499292, + 0.014261542819440365, + 0.013231002725660801, + 0.03136540949344635, + 0.016711756587028503, + 0.0839075893163681, + 0.034494463354349136, + 0.02057485654950142, + 0.015210398472845554, + 0.03171377256512642, + 0.08828946948051453, + 0.013316315598785877, + 0.011581726372241974, + 0.017305025830864906, + 0.014344296418130398, + 0.013893460854887962, + 0.008661172352731228, + 0.022418312728405, + 0.036501653492450714, + 0.006108473055064678, + 0.0045301723293960094, + 0.009674851782619953, + 0.007220773957669735, + 0.006606393493711948, + 0.009340789169073105, + 0.009038235992193222, + 0.010037703439593315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.012543711811304092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067663207650185 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067663207650185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067663207650185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11409827321767807, + 0.0502055287361145, + 0.025334198027849197, + 0.040915194898843765, + 0.12158768624067307, + 0.03690674155950546, + 0.17440977692604065, + 0.07998824864625931, + 0.09310070425271988, + 0.2634536325931549 + ], + "credit_map": [ + 0.016344603151082993, + 0.014248695224523544, + 0.017036791890859604, + 0.015502428635954857, + 0.010219804011285305, + 0.00989286694675684, + 0.013629656285047531, + 0.017223432660102844, + 0.006330297328531742, + 0.0052679842337965965, + 0.007881121709942818, + 0.005982804112136364, + 0.0048048351891338825, + 0.003971692640334368, + 0.008454853668808937, + 0.00751193892210722, + 0.002182862488552928, + 0.0019205702701583505, + 0.002885795896872878, + 0.0030954487156122923, + 0.003852523397654295, + 0.00231100688688457, + 0.005360081326216459, + 0.0037259082309901714, + 0.006141177378594875, + 0.0064248135313391685, + 0.004728342406451702, + 0.0054253628477454185, + 0.003753525670617819, + 0.004229888319969177, + 0.005530332215130329, + 0.0046817525289952755, + 0.04648137837648392, + 0.042751409113407135, + 0.006136791780591011, + 0.005685938056558371, + 0.005687884520739317, + 0.0045550004579126835, + 0.005182216409593821, + 0.005107061937451363, + 0.004538113716989756, + 0.0027613844722509384, + 0.00555945560336113, + 0.004598578438162804, + 0.005080648232251406, + 0.0038621036801487207, + 0.005242778919637203, + 0.0052636777982115746, + 0.020414164289832115, + 0.02933100238442421, + 0.021545669063925743, + 0.018958749249577522, + 0.028256753459572792, + 0.01444127969443798, + 0.020672187209129333, + 0.020789973437786102, + 0.006189591716974974, + 0.013064474798738956, + 0.011029343120753765, + 0.011443624272942543, + 0.01073089987039566, + 0.005157105159014463, + 0.011913464404642582, + 0.010459741577506065, + 0.006848202086985111, + 0.007932922802865505, + 0.01289922185242176, + 0.014291267842054367, + 0.009794341400265694, + 0.009443153627216816, + 0.013898037374019623, + 0.017993558198213577, + 0.029788032174110413, + 0.024523615837097168, + 0.03941059857606888, + 0.027945252135396004, + 0.03239735960960388, + 0.04845646023750305, + 0.027302589267492294, + 0.03362973406910896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.016067663207650185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031876962166279554 + ], + "opposite_logprob_tokens": [ + -5.753187656402588 + ], + "logprob_reward": -0.0031876962166279554, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031876962166279554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2013167291879654, + 0.0704423263669014, + 0.03730284795165062, + 0.03478819876909256, + 0.0717020258307457, + 0.06239961460232735, + 0.1899029016494751, + 0.05593499168753624, + 0.07930638641119003, + 0.1969040334224701 + ], + "credit_map": [ + 0.01634807698428631, + 0.020379288122057915, + 0.030143357813358307, + 0.02873888611793518, + 0.01715882308781147, + 0.018960440531373024, + 0.03627166897058487, + 0.03331618010997772, + 0.006669566500931978, + 0.006586579140275717, + 0.01244798582047224, + 0.010654161684215069, + 0.0051909140311181545, + 0.00800149142742157, + 0.009111021645367146, + 0.011780612170696259, + 0.0038689556531608105, + 0.0035790118854492903, + 0.004945914726704359, + 0.005622101482003927, + 0.0044296360574662685, + 0.0031375004909932613, + 0.005964227952063084, + 0.005755499936640263, + 0.005378222558647394, + 0.005178236402571201, + 0.0031636799685657024, + 0.004701678641140461, + 0.0038829415570944548, + 0.003440903965383768, + 0.004682945553213358, + 0.004359589423984289, + 0.016859691590070724, + 0.01655675284564495, + 0.006104225292801857, + 0.00860015582293272, + 0.00570938503369689, + 0.0030154066625982523, + 0.008567122742533684, + 0.006289286073297262, + 0.005108796525746584, + 0.004252732265740633, + 0.006515734829008579, + 0.008763923309743404, + 0.007965322583913803, + 0.0031439990270882845, + 0.016641680151224136, + 0.010007424280047417, + 0.01919391006231308, + 0.028545279055833817, + 0.025907794013619423, + 0.0208139605820179, + 0.02837877906858921, + 0.015561160631477833, + 0.02383042685687542, + 0.027671590447425842, + 0.00357803120277822, + 0.0063211023807525635, + 0.007908756844699383, + 0.006700324825942516, + 0.005383935756981373, + 0.003400829154998064, + 0.012027214281260967, + 0.010614795610308647, + 0.008021705783903599, + 0.005650872830301523, + 0.009592658840119839, + 0.00986064225435257, + 0.0068239495158195496, + 0.006319177802652121, + 0.010437903925776482, + 0.022599469870328903, + 0.028106462210416794, + 0.017954673618078232, + 0.043478988111019135, + 0.02309650182723999, + 0.015448055230081081, + 0.02184952050447464, + 0.016617432236671448, + 0.030352408066391945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.0031876962166279554 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly and steadily pans forward, focusing on the summit of a beacon tower on the ancient Great Wall. A massive crimson battle flag flutters wildly in the wind. Crafted from heavy linen, its edges frayed by age, the flag is embroidered with a vivid dragon totem in golden thread, its scales shimmering in the light. The relentless wind tears at the flag, sometimes unfurling with a snapping sound, fully revealing the unyielding dragon, and sometimes curling tightly around the weathered, dark-wood pole. The composition of light and shadow possesses Annie Leibovitz's signature quality: the evening sun hangs low in the sky, its golden rays penetrating thick, swirling clouds at a sharp angle, precisely illuminating the windward side of the flag, making its deep red appear exceptionally saturated and the gold embroidery gleam with a metallic sheen. Meanwhile, the other side of the flag and the massive stone brick walls of the beacon tower are shrouded in deep, cold shadows, creating a strong contrast between light and dark, highlighting the rough texture of the stone and the vicissitudes of history. The backdrop is a rolling, rugged mountain range, gradually merging into the gray-blue twilight in the distance. The entire scene, in this extreme contrast of movement and stillness, is permeated with an epic atmosphere of solemnity, loneliness, and historical power.", + "seed": 16931, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_14/index_3652_seed_16931_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04680063761770725, + "vlm_logprob": -0.04680063761770725, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.04680063761770725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11175825446844101 + ], + "opposite_logprob_tokens": [ + -2.736758232116699 + ], + "logprob_reward": -0.11175825446844101, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.11175825446844101, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17650960385799408, + 0.05397823080420494, + 0.17986619472503662, + 0.12226416915655136, + 0.06158117577433586, + 0.041038427501916885, + 0.04949271306395531, + 0.06758810579776764, + 0.06942737102508545, + 0.17825400829315186 + ], + "credit_map": [ + 0.017909616231918335, + 0.019103579223155975, + 0.019534507766366005, + 0.02569614164531231, + 0.01966826803982258, + 0.020277179777622223, + 0.03385724499821663, + 0.020463062450289726, + 0.005323528777807951, + 0.006061350926756859, + 0.006417217198759317, + 0.008746083825826645, + 0.004325441084802151, + 0.005781452637165785, + 0.009598051197826862, + 0.007725103292614222, + 0.019842365756630898, + 0.018361413851380348, + 0.016136091202497482, + 0.022624684497714043, + 0.02242329530417919, + 0.010022643022239208, + 0.04071575775742531, + 0.02973994053900242, + 0.027837185189127922, + 0.03096562996506691, + 0.008922239765524864, + 0.011046992614865303, + 0.009111753664910793, + 0.014235356822609901, + 0.01079846266657114, + 0.009346555918455124, + 0.006186643615365028, + 0.004922769032418728, + 0.0066351923160254955, + 0.00740723917260766, + 0.00826368760317564, + 0.0036946493200957775, + 0.013065372593700886, + 0.011405623517930508, + 0.005260881502181292, + 0.005122481845319271, + 0.007597489282488823, + 0.0051896884106099606, + 0.003958639223128557, + 0.0040421076118946075, + 0.005671423859894276, + 0.004195712506771088, + 0.004400074016302824, + 0.005885814782232046, + 0.008042271248996258, + 0.006250160280615091, + 0.005851588677614927, + 0.0035023896489292383, + 0.004980872850865126, + 0.010579542256891727, + 0.006170667242258787, + 0.010084317065775394, + 0.008563178591430187, + 0.00969192385673523, + 0.007151005789637566, + 0.008300690911710262, + 0.0073058269917964935, + 0.010320489294826984, + 0.004418237134814262, + 0.009218675084412098, + 0.005925700534135103, + 0.017316123470664024, + 0.003909381106495857, + 0.006924537010490894, + 0.005511286202818155, + 0.01620342954993248, + 0.03024127520620823, + 0.05241682380437851, + 0.014162447303533554, + 0.019516678526997566, + 0.010916599072515965, + 0.010794397443532944, + 0.010364336892962456, + 0.02984144724905491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.11175825446844101 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003398671746254 + ], + "opposite_logprob_tokens": [ + -2.350339889526367 + ], + "logprob_reward": -0.1003398671746254, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003398671746254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14899227023124695, + 0.05298163741827011, + 0.06093832105398178, + 0.06983625143766403, + 0.07424354553222656, + 0.06680602580308914, + 0.06487739831209183, + 0.15279164910316467, + 0.05283745378255844, + 0.2556954622268677 + ], + "credit_map": [ + 0.017060929909348488, + 0.01725531928241253, + 0.01128628384321928, + 0.016204580664634705, + 0.017873134464025497, + 0.010550996288657188, + 0.03865028917789459, + 0.020110726356506348, + 0.006179135758429766, + 0.0047790794633328915, + 0.006829751655459404, + 0.005710499361157417, + 0.005680502858012915, + 0.0036096328403800726, + 0.012097029946744442, + 0.008096004836261272, + 0.0071268873289227486, + 0.006286300253123045, + 0.007012168876826763, + 0.007169793359935284, + 0.005201021209359169, + 0.0052748615853488445, + 0.009191075339913368, + 0.013676213100552559, + 0.00716715632006526, + 0.007496856153011322, + 0.009021163918077946, + 0.011888936161994934, + 0.0048856353387236595, + 0.005075494758784771, + 0.008663532324135303, + 0.015637481585144997, + 0.010597112588584423, + 0.009431280195713043, + 0.009226429276168346, + 0.008099623024463654, + 0.00685035390779376, + 0.003433023579418659, + 0.0157148540019989, + 0.010890872217714787, + 0.009145202115178108, + 0.009217717684805393, + 0.010351910255849361, + 0.008198498748242855, + 0.007276756688952446, + 0.006871578749269247, + 0.008387685753405094, + 0.007356680929660797, + 0.0063442750833928585, + 0.005276212003082037, + 0.010156686417758465, + 0.007685206830501556, + 0.00933132879436016, + 0.004790512379258871, + 0.006789047736674547, + 0.014504131861031055, + 0.01269739679992199, + 0.026326116174459457, + 0.018411193042993546, + 0.02428690530359745, + 0.019229277968406677, + 0.016597794368863106, + 0.012529981322586536, + 0.022712990641593933, + 0.004056028556078672, + 0.007132296916097403, + 0.005037115886807442, + 0.00997970998287201, + 0.0038383724167943, + 0.006250061560422182, + 0.004835302010178566, + 0.011708571575582027, + 0.04534738138318062, + 0.07363317906856537, + 0.022015376016497612, + 0.028566312044858932, + 0.012958250939846039, + 0.012994090095162392, + 0.014332552440464497, + 0.045848313719034195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.1003398671746254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087591759860516 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087591759860516, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087591759860516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23735034465789795, + 0.10360786318778992, + 0.09858836233615875, + 0.03976055979728699, + 0.06405839323997498, + 0.02216709591448307, + 0.030577030032873154, + 0.18714381754398346, + 0.029879331588745117, + 0.18686719238758087 + ], + "credit_map": [ + 0.0335175022482872, + 0.030898215249180794, + 0.01843170076608658, + 0.026748811826109886, + 0.023224936798214912, + 0.01998070999979973, + 0.05729583650827408, + 0.027252618223428726, + 0.009751793928444386, + 0.01051583606749773, + 0.01117483526468277, + 0.013447845354676247, + 0.010868332348763943, + 0.011828012764453888, + 0.01956939324736595, + 0.016451817005872726, + 0.011611822061240673, + 0.008093326352536678, + 0.011323017068207264, + 0.014633427374064922, + 0.01000091154128313, + 0.0067430948838591576, + 0.01851397193968296, + 0.017668798565864563, + 0.006371662486344576, + 0.005733800586313009, + 0.003993838559836149, + 0.005820506252348423, + 0.0034197717905044556, + 0.002677287207916379, + 0.005912903696298599, + 0.0058307889848947525, + 0.0071845026686787605, + 0.0060539040714502335, + 0.006487999577075243, + 0.00854324921965599, + 0.009433452039957047, + 0.004909840412437916, + 0.014677603729069233, + 0.006767843384295702, + 0.0026987288147211075, + 0.002408181084319949, + 0.002241478767246008, + 0.002150215907022357, + 0.0024769650772213936, + 0.0024116679560393095, + 0.0030831231269985437, + 0.004696735646575689, + 0.0031505683436989784, + 0.002734747249633074, + 0.003424650989472866, + 0.003698682878166437, + 0.0029187272302806377, + 0.002647670218721032, + 0.004190896637737751, + 0.007811087183654308, + 0.016841813921928406, + 0.05160650238394737, + 0.020023522898554802, + 0.03425360098481178, + 0.01584768481552601, + 0.009871665388345718, + 0.014590772800147533, + 0.024108244106173515, + 0.0034688066225498915, + 0.004514907021075487, + 0.004315115045756102, + 0.007164445705711842, + 0.002193793887272477, + 0.0019554123282432556, + 0.002182047115638852, + 0.0040848031640052795, + 0.031263940036296844, + 0.05444977432489395, + 0.014445372857153416, + 0.02089744061231613, + 0.009858652018010616, + 0.009220295585691929, + 0.012001916766166687, + 0.034729793667793274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.011087591759860516 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629540748894215 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629540748894215, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629540748894215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1287343055009842, + 0.06580311059951782, + 0.13146796822547913, + 0.04470856115221977, + 0.03801985830068588, + 0.05742693319916725, + 0.05451721325516701, + 0.10847626626491547, + 0.0843878760933876, + 0.2864578366279602 + ], + "credit_map": [ + 0.015089435502886772, + 0.013577125035226345, + 0.01809588633477688, + 0.01897803694009781, + 0.015347744338214397, + 0.012343738228082657, + 0.017956048250198364, + 0.017346298322081566, + 0.006801865994930267, + 0.006586427800357342, + 0.007267770357429981, + 0.007733286824077368, + 0.007117077708244324, + 0.009259865619242191, + 0.012447378598153591, + 0.008589440025389194, + 0.01969469152390957, + 0.03103894367814064, + 0.01805601269006729, + 0.013599268160760403, + 0.00724890036508441, + 0.012703766115009785, + 0.01133488118648529, + 0.01779152825474739, + 0.005997870117425919, + 0.006439568474888802, + 0.0056923748925328255, + 0.004490886814892292, + 0.004295704420655966, + 0.006018956191837788, + 0.006491899956017733, + 0.005281301215291023, + 0.0052214073948562145, + 0.0043752179481089115, + 0.006520367227494717, + 0.0047271158546209335, + 0.0037138427142053843, + 0.001959162298589945, + 0.006764785852283239, + 0.004737961571663618, + 0.00966673344373703, + 0.008698117919266224, + 0.01094247866421938, + 0.006185002159327269, + 0.005228166468441486, + 0.004455238580703735, + 0.006156467366963625, + 0.006094733253121376, + 0.004955185577273369, + 0.0063614435493946075, + 0.005761414300650358, + 0.008039931766688824, + 0.00893896259367466, + 0.0037671958561986685, + 0.006033682730048895, + 0.010659401305019855, + 0.009208894334733486, + 0.022248808294534683, + 0.010946130380034447, + 0.01589762233197689, + 0.01651119440793991, + 0.00987219251692295, + 0.008220850490033627, + 0.015570580959320068, + 0.008394873701035976, + 0.014395322650671005, + 0.011539166793227196, + 0.017317447811365128, + 0.006393884774297476, + 0.0057397158816456795, + 0.006459164898842573, + 0.014148305170238018, + 0.05141763016581535, + 0.09056632965803146, + 0.020924989134073257, + 0.032607559114694595, + 0.015386426821351051, + 0.013344504870474339, + 0.015566184185445309, + 0.046644240617752075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.007629540748894215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003187933936715126 + ], + "opposite_logprob_tokens": [ + -5.753188133239746 + ], + "logprob_reward": -0.003187933936715126, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003187933936715126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18213742971420288, + 0.10412083566188812, + 0.10709623247385025, + 0.0701441764831543, + 0.03452588990330696, + 0.03400043398141861, + 0.034176576882600784, + 0.08496516197919846, + 0.07343567162752151, + 0.2753975987434387 + ], + "credit_map": [ + 0.021758699789643288, + 0.022218182682991028, + 0.030174342915415764, + 0.025279713794589043, + 0.01460825465619564, + 0.012015335261821747, + 0.0318751260638237, + 0.024207770824432373, + 0.014854415319859982, + 0.013705270364880562, + 0.012005061842501163, + 0.013715486973524094, + 0.008234716020524502, + 0.009538061916828156, + 0.01713624782860279, + 0.014931575395166874, + 0.014190138317644596, + 0.013856291770935059, + 0.012832015752792358, + 0.011467520147562027, + 0.010078460909426212, + 0.008211620151996613, + 0.012662671506404877, + 0.023797515779733658, + 0.008233783766627312, + 0.007882270961999893, + 0.008578651584684849, + 0.011545584537088871, + 0.005910546984523535, + 0.005684273783117533, + 0.009413385763764381, + 0.012895677238702774, + 0.004029656760394573, + 0.0033548050560057163, + 0.004025864414870739, + 0.004134181421250105, + 0.0037354864180088043, + 0.0025879009626805782, + 0.0060674273408949375, + 0.006590566132217646, + 0.005273934453725815, + 0.004593380261212587, + 0.004374867770820856, + 0.003889137879014015, + 0.0026195738464593887, + 0.005704717710614204, + 0.003659590147435665, + 0.003885231912136078, + 0.003284430131316185, + 0.003822463098913431, + 0.004358459264039993, + 0.004622186534106731, + 0.005640772171318531, + 0.0020691766403615475, + 0.004242005292326212, + 0.00613708421587944, + 0.00804175902158022, + 0.015702826902270317, + 0.010639108717441559, + 0.012052137404680252, + 0.011204016394913197, + 0.0073136333376169205, + 0.007788107730448246, + 0.012223578989505768, + 0.006916049402207136, + 0.011889759451150894, + 0.007619782350957394, + 0.019588856026530266, + 0.0054292525164783, + 0.007940801791846752, + 0.004707858897745609, + 0.009343313053250313, + 0.048621390014886856, + 0.08251111209392548, + 0.022574273869395256, + 0.03166003152728081, + 0.014226572588086128, + 0.013391050510108471, + 0.014938490465283394, + 0.04747467115521431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.003187933936715126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly and steadily pans forward, focusing on the summit of a beacon tower on the ancient Great Wall. A massive crimson battle flag flutters wildly in the wind. Crafted from heavy linen, its edges frayed by age, the flag is embroidered with a vivid dragon totem in golden thread, its scales shimmering in the light. The relentless wind tears at the flag, sometimes unfurling with a snapping sound, fully revealing the unyielding dragon, and sometimes curling tightly around the weathered, dark-wood pole. The composition of light and shadow possesses Annie Leibovitz's signature quality: the evening sun hangs low in the sky, its golden rays penetrating thick, swirling clouds at a sharp angle, precisely illuminating the windward side of the flag, making its deep red appear exceptionally saturated and the gold embroidery gleam with a metallic sheen. Meanwhile, the other side of the flag and the massive stone brick walls of the beacon tower are shrouded in deep, cold shadows, creating a strong contrast between light and dark, highlighting the rough texture of the stone and the vicissitudes of history. The backdrop is a rolling, rugged mountain range, gradually merging into the gray-blue twilight in the distance. The entire scene, in this extreme contrast of movement and stillness, is permeated with an epic atmosphere of solemnity, loneliness, and historical power.", + "seed": 16932, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_14/index_3652_seed_16932_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13501533716917039, + "vlm_logprob": -0.13501533716917039, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.13501533716917039, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1787315309047699 + ], + "opposite_logprob_tokens": [ + -2.3037314414978027 + ], + "logprob_reward": -0.1787315309047699, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1787315309047699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13004812598228455, + 0.04243708401918411, + 0.03848608583211899, + 0.019547022879123688, + 0.028067931532859802, + 0.06143779680132866, + 0.04835738241672516, + 0.046831727027893066, + 0.5269798040390015, + 0.057807061821222305 + ], + "credit_map": [ + 0.014878746122121811, + 0.014132041484117508, + 0.022063827142119408, + 0.021617067977786064, + 0.01041414961218834, + 0.009831862524151802, + 0.01849035546183586, + 0.018620073795318604, + 0.0063240001909434795, + 0.005981093738228083, + 0.004918496590107679, + 0.006933257915079594, + 0.0031778598204255104, + 0.005194223485887051, + 0.005407283548265696, + 0.004500867333263159, + 0.006893132347613573, + 0.007521898485720158, + 0.0026926074642688036, + 0.004585504997521639, + 0.003204707521945238, + 0.0024976369459182024, + 0.006505997385829687, + 0.0045846011489629745, + 0.003263550577685237, + 0.0029439758509397507, + 0.0023526418954133987, + 0.0024313621688634157, + 0.00209773238748312, + 0.001361801172606647, + 0.0028982774820178747, + 0.0021976830903440714, + 0.0034197636414319277, + 0.0036747902631759644, + 0.0033255266025662422, + 0.004374364856630564, + 0.002708992687985301, + 0.0035080513916909695, + 0.003949571400880814, + 0.0031068730168044567, + 0.008364081382751465, + 0.0066530960611999035, + 0.007569197565317154, + 0.00842502061277628, + 0.005704756360501051, + 0.006154504604637623, + 0.011826957575976849, + 0.006740184035152197, + 0.006041120737791061, + 0.005180475767701864, + 0.0056031933054327965, + 0.010392753407359123, + 0.004378849174827337, + 0.004430086817592382, + 0.007688563782721758, + 0.004642339888960123, + 0.008709834888577461, + 0.004534690175205469, + 0.005791746079921722, + 0.00636876467615366, + 0.004529062658548355, + 0.006789145991206169, + 0.005783680360764265, + 0.004324797075241804, + 0.10627978295087814, + 0.06714778393507004, + 0.0847105160355568, + 0.09307003766298294, + 0.031418003141880035, + 0.06336820870637894, + 0.03386954590678215, + 0.04711589962244034, + 0.010389824397861958, + 0.006620847620069981, + 0.010798745788633823, + 0.008305667899549007, + 0.0035245739854872227, + 0.006162623409181833, + 0.005993119906634092, + 0.006011658813804388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.1787315309047699 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742887616157532 + ], + "opposite_logprob_tokens": [ + -0.9742887616157532 + ], + "logprob_reward": -0.4742887616157532, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742887616157532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09025935083627701, + 0.041893120855093, + 0.029568782076239586, + 0.032729730010032654, + 0.09679371863603592, + 0.16162443161010742, + 0.18353275954723358, + 0.09676734358072281, + 0.21791012585163116, + 0.04892059043049812 + ], + "credit_map": [ + 0.011979261413216591, + 0.010739442892372608, + 0.012082060799002647, + 0.012347787618637085, + 0.009457405656576157, + 0.008543344214558601, + 0.012578463181853294, + 0.01253159623593092, + 0.006197978742420673, + 0.005400871392339468, + 0.004508365411311388, + 0.005628084298223257, + 0.005851511377841234, + 0.004155667964369059, + 0.0062298704870045185, + 0.003920774906873703, + 0.0027980380691587925, + 0.00341885257512331, + 0.0037101528141647577, + 0.004341618157923222, + 0.0034302857238799334, + 0.004275184124708176, + 0.004468814469873905, + 0.003125835908576846, + 0.004895939491689205, + 0.004924206528812647, + 0.003246121807023883, + 0.0043687596917152405, + 0.0037020500749349594, + 0.00234914175234735, + 0.005731502082198858, + 0.003512011608108878, + 0.011951422318816185, + 0.010792357847094536, + 0.00945765059441328, + 0.014222929254174232, + 0.009537412784993649, + 0.012550476007163525, + 0.013020615093410015, + 0.015260867774486542, + 0.023325325921177864, + 0.026210136711597443, + 0.023010818287730217, + 0.02605319395661354, + 0.008886203169822693, + 0.017077680677175522, + 0.018009141087532043, + 0.019051939249038696, + 0.026059959083795547, + 0.01882171630859375, + 0.026845011860132217, + 0.027876367792487144, + 0.017344944179058075, + 0.019637981429696083, + 0.02832084707915783, + 0.018625935539603233, + 0.014289350248873234, + 0.014145107939839363, + 0.014286265708506107, + 0.013081693090498447, + 0.008396472781896591, + 0.011577174067497253, + 0.007946118712425232, + 0.013045170344412327, + 0.03076268546283245, + 0.032597579061985016, + 0.036511003971099854, + 0.028362680226564407, + 0.015720777213573456, + 0.022540368139743805, + 0.028948428109288216, + 0.022466618567705154, + 0.00835942942649126, + 0.004576537292450666, + 0.008946999907493591, + 0.007055781316012144, + 0.003400207031518221, + 0.005089977290481329, + 0.004579107277095318, + 0.006912552285939455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.4742887616157532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14429107308387756, + 0.07226591557264328, + 0.05044685676693916, + 0.035321224480867386, + 0.05137182027101517, + 0.060460470616817474, + 0.06662281602621078, + 0.030646946281194687, + 0.22390629351139069, + 0.26466652750968933 + ], + "credit_map": [ + 0.016722336411476135, + 0.015654562041163445, + 0.01880306378006935, + 0.022020533680915833, + 0.021570954471826553, + 0.011819765903055668, + 0.02102334424853325, + 0.016676519066095352, + 0.00857620220631361, + 0.00779248820617795, + 0.008816806599497795, + 0.01174858957529068, + 0.010489319451153278, + 0.006510849576443434, + 0.010888989083468914, + 0.007442671339958906, + 0.005382193718105555, + 0.005036866758018732, + 0.005184231325984001, + 0.008415602147579193, + 0.005802466068416834, + 0.004647265654057264, + 0.00888771191239357, + 0.007090519182384014, + 0.004485169891268015, + 0.005277849733829498, + 0.0040702116675674915, + 0.004540587309747934, + 0.0030884742736816406, + 0.0036791153252124786, + 0.004882744513452053, + 0.005297074094414711, + 0.005748338531702757, + 0.006683217827230692, + 0.006365219596773386, + 0.008444723673164845, + 0.005288949701935053, + 0.005775284953415394, + 0.007012555375695229, + 0.006053528748452663, + 0.007083151489496231, + 0.011423284187912941, + 0.006087173707783222, + 0.012645878829061985, + 0.00481646740809083, + 0.005091954488307238, + 0.0075960648246109486, + 0.005716494284570217, + 0.011000237427651882, + 0.00843873806297779, + 0.007001870311796665, + 0.011064882390201092, + 0.005968213547021151, + 0.008072283118963242, + 0.010163852944970131, + 0.004912738688290119, + 0.0063422322273254395, + 0.004587455652654171, + 0.004708245396614075, + 0.004167321603745222, + 0.001490170368924737, + 0.004681653808802366, + 0.0016961749643087387, + 0.0029736938886344433, + 0.0314481183886528, + 0.03639141097664833, + 0.0303685050457716, + 0.04371696710586548, + 0.014247210696339607, + 0.024891356006264687, + 0.02575955167412758, + 0.017083188518881798, + 0.041440121829509735, + 0.0249570831656456, + 0.06868048012256622, + 0.05200131982564926, + 0.011578075587749481, + 0.027253631502389908, + 0.016070310026407242, + 0.022685524076223373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0097922682762146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631788495928049 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631788495928049, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631788495928049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21343757212162018, + 0.08850781619548798, + 0.05537046864628792, + 0.03971448168158531, + 0.07067648321390152, + 0.07781549543142319, + 0.07531020045280457, + 0.09114769101142883, + 0.23364503681659698, + 0.054374754428863525 + ], + "credit_map": [ + 0.022371850907802582, + 0.020867042243480682, + 0.029002327471971512, + 0.02869405969977379, + 0.018772905692458153, + 0.013622278347611427, + 0.031761907041072845, + 0.04834518954157829, + 0.016628403216600418, + 0.015114134177565575, + 0.013508244417607784, + 0.010232268832623959, + 0.005378200206905603, + 0.00803471077233553, + 0.010075822472572327, + 0.009536032564938068, + 0.008548663929104805, + 0.0069764102809131145, + 0.00980138685554266, + 0.007881230674684048, + 0.0034151871223002672, + 0.007719296962022781, + 0.005739609245210886, + 0.005288680549710989, + 0.0065515064634382725, + 0.004610002040863037, + 0.005189546383917332, + 0.0039551397785544395, + 0.0032583491411060095, + 0.004779395181685686, + 0.004687951877713203, + 0.00668259197846055, + 0.009831331670284271, + 0.005580459721386433, + 0.0077356272377073765, + 0.0066773006692528725, + 0.006662413477897644, + 0.011006442829966545, + 0.00847433041781187, + 0.01470857858657837, + 0.009935488924384117, + 0.012038596905767918, + 0.014288823120296001, + 0.01073510106652975, + 0.0058373138308525085, + 0.008316164836287498, + 0.007871179841458797, + 0.008792826905846596, + 0.010110392235219479, + 0.00621554022654891, + 0.014285113662481308, + 0.012200052849948406, + 0.00548062426969409, + 0.006797309964895248, + 0.010444572195410728, + 0.009776599705219269, + 0.019274597987532616, + 0.011567668989300728, + 0.015086914412677288, + 0.010969968512654305, + 0.006153697147965431, + 0.01157190278172493, + 0.007534468080848455, + 0.008988478220999241, + 0.04492179676890373, + 0.02105158567428589, + 0.043181002140045166, + 0.032850272953510284, + 0.01654762588441372, + 0.02714593894779682, + 0.024115467444062233, + 0.023831341415643692, + 0.00837745238095522, + 0.005917643196880817, + 0.013267531991004944, + 0.01037546806037426, + 0.002949624555185437, + 0.00541334692388773, + 0.003423198126256466, + 0.00465049035847187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.007631788495928049 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan forward towards the summit of a beacon tower on the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crimson battle flag visible, featuring a golden dragon totem embroidered on heavy linen with frayed edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by low evening sun casting strong golden highlights on the windward side of the flag while the rest of the scene is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rolling, rugged mountain range merging into a gray-blue twilight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632336553186178 + ], + "opposite_logprob_tokens": [ + -5.379632472991943 + ], + "logprob_reward": -0.004632336553186178, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632336553186178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2156757265329361, + 0.09369014203548431, + 0.036046307533979416, + 0.05824701115489006, + 0.04500505328178406, + 0.05333666875958443, + 0.06882500648498535, + 0.041696947067976, + 0.31656932830810547, + 0.070907823741436 + ], + "credit_map": [ + 0.02810121886432171, + 0.024610096588730812, + 0.031017445027828217, + 0.02941124513745308, + 0.019862011075019836, + 0.015556322410702705, + 0.031720876693725586, + 0.0353965200483799, + 0.01243000291287899, + 0.010535079054534435, + 0.012872439809143543, + 0.012683344073593616, + 0.006439455784857273, + 0.01216114591807127, + 0.012709512375295162, + 0.013859154656529427, + 0.005020169075578451, + 0.004132292699068785, + 0.005447054747492075, + 0.00434177927672863, + 0.0023967314045876265, + 0.004416822921484709, + 0.004008505493402481, + 0.006282949820160866, + 0.012163439765572548, + 0.011816325597465038, + 0.006749662570655346, + 0.005820443853735924, + 0.0026406855322420597, + 0.007981214672327042, + 0.004469220992177725, + 0.006606015842407942, + 0.004935816861689091, + 0.004943571053445339, + 0.005610360763967037, + 0.006507381331175566, + 0.003844565711915493, + 0.0071633197367191315, + 0.00651855394244194, + 0.005481485743075609, + 0.006660865154117346, + 0.004940721672028303, + 0.00778678385540843, + 0.006013514474034309, + 0.0038788318634033203, + 0.005842973943799734, + 0.007933837361633778, + 0.010279142297804356, + 0.010062644258141518, + 0.007937517948448658, + 0.009155675768852234, + 0.01119846198707819, + 0.00484853982925415, + 0.007509045768529177, + 0.008958442136645317, + 0.00915467832237482, + 0.009072991088032722, + 0.004481238778680563, + 0.005733808968216181, + 0.005095916334539652, + 0.0031819010619074106, + 0.006718574091792107, + 0.0035317293368279934, + 0.003880788804963231, + 0.06123770773410797, + 0.03237810358405113, + 0.05720622465014458, + 0.05056430771946907, + 0.01999935321509838, + 0.03667213395237923, + 0.02781866304576397, + 0.03069283999502659, + 0.009000560268759727, + 0.008158516138792038, + 0.017774909734725952, + 0.014705462381243706, + 0.004239737521857023, + 0.006523035001009703, + 0.005207875277847052, + 0.005297726485878229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.004632336553186178 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_15/index_tensor([2226, 2226])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_15/index_tensor([2226, 2226])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c8bb0af374e7e7a950074d5826b5c43bad6da564 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_15/index_tensor([2226, 2226])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In Hirohiko Araki's signature high-contrast manga style, the camera, in a sharp, dizzying Hitchcockian dolly zoom, zooms in on a massive, brown jellyfish suspended in an abstract mid-air. The creature's umbrella-like body, translucent like smoky quartz, is covered in intricate geometric patterns and biological veins, as clear as an anatomical diagram, its edges defined by sharp, hard black lines. In an instant, it arches its body, its umbrella-like structure contracting with a sense of monstrous biological power, as if accumulating a massive surge of energy. Immediately, countless tentacles beneath it, previously lithe and floating, instantly straighten and explode in all directions, as if struck by a high-voltage electrical current. Each one becomes as hard and erect as a needle of poisoned glass, its tip gleaming with an ominous, cold light. As the camera moves forward and the lens pulls away simultaneously, the background space is extremely compressed and distorted, creating a strong visual impact and a suffocating sense of oppression. The entire picture is filled with an imminent and unimaginable sense of danger in the harsh light and shadow and surreal dynamics.", + "seed": 28012, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_15/index_2226_seed_28012_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2233199064619839, + "vlm_logprob": -0.2233199064619839, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.2233199064619839, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015710510313510895 + ], + "opposite_logprob_tokens": [ + -4.265710353851318 + ], + "logprob_reward": -0.015710510313510895, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015710510313510895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25413674116134644, + 0.10687978565692902, + 0.06379445642232895, + 0.09006260335445404, + 0.10488792508840561, + 0.22069358825683594, + 0.06276363134384155, + 0.0488252267241478, + 0.03923404961824417, + 0.008721965365111828 + ], + "credit_map": [ + 0.032450903207063675, + 0.017422163859009743, + 0.05899569392204285, + 0.017829300835728645, + 0.019527044147253036, + 0.030676325783133507, + 0.031956855207681656, + 0.04527845233678818, + 0.012418936006724834, + 0.013248137198388577, + 0.014711299911141396, + 0.005077363923192024, + 0.01061985082924366, + 0.016893597319722176, + 0.008989796042442322, + 0.02492079697549343, + 0.012814274989068508, + 0.0071609667502343655, + 0.008213129825890064, + 0.005162558052688837, + 0.008664983324706554, + 0.007401998620480299, + 0.005245527718216181, + 0.009131016209721565, + 0.020725486800074577, + 0.01031477376818657, + 0.012894464656710625, + 0.004585142247378826, + 0.0102907195687294, + 0.013231162913143635, + 0.006463143974542618, + 0.01155771128833294, + 0.009928274899721146, + 0.014707268215715885, + 0.009502670727670193, + 0.013511399738490582, + 0.012400214560329914, + 0.015984706580638885, + 0.01668054237961769, + 0.012172860093414783, + 0.011133600026369095, + 0.06621837615966797, + 0.006500332150608301, + 0.00741119310259819, + 0.027339061722159386, + 0.021365530788898468, + 0.020676862448453903, + 0.06004863977432251, + 0.005943599157035351, + 0.007623116951435804, + 0.006214960478246212, + 0.005194496363401413, + 0.016999052837491035, + 0.009267902001738548, + 0.0040956237353384495, + 0.007424885407090187, + 0.002264777896925807, + 0.01870044693350792, + 0.0015598178142681718, + 0.0010293133091181517, + 0.019295141100883484, + 0.002219072775915265, + 0.0011452873004600406, + 0.0026113714557141066, + 0.003718300722539425, + 0.004532902501523495, + 0.0012517690192908049, + 0.003814717987552285, + 0.010456639342010021, + 0.0021894017700105906, + 0.010592181235551834, + 0.0026781363412737846, + 0.0010401642648503184, + 0.0013711489737033844, + 0.0005265329382382333, + 0.0009386042365804315, + 0.002433907939121127, + 0.000938546028919518, + 0.0007968046702444553, + 0.0006762570119462907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.015710510313510895 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023306243121623993 + ], + "opposite_logprob_tokens": [ + -3.773306131362915 + ], + "logprob_reward": -0.023306243121623993, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306243121623993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12397561967372894, + 0.16325904428958893, + 0.055285073816776276, + 0.06048291549086571, + 0.11841703206300735, + 0.04013741761445999, + 0.030488355085253716, + 0.36677607893943787, + 0.01641257293522358, + 0.024765899404883385 + ], + "credit_map": [ + 0.016879934817552567, + 0.013020806945860386, + 0.017627889290452003, + 0.007795165292918682, + 0.019138621166348457, + 0.014889928512275219, + 0.016490302979946136, + 0.018132967874407768, + 0.02036461979150772, + 0.014599325135350227, + 0.024332301691174507, + 0.010007242672145367, + 0.011964828707277775, + 0.026373330503702164, + 0.021214554086327553, + 0.034402839839458466, + 0.00844337698072195, + 0.009248259477317333, + 0.008762181736528873, + 0.004107012879103422, + 0.0059612952172756195, + 0.0076356856152415276, + 0.005701014306396246, + 0.005426247138530016, + 0.008106376975774765, + 0.006330639123916626, + 0.007499929051846266, + 0.003635939210653305, + 0.007001865655183792, + 0.010950865224003792, + 0.006329216528683901, + 0.010628084652125835, + 0.009949666447937489, + 0.015589275397360325, + 0.009930459782481194, + 0.01824158988893032, + 0.014979933388531208, + 0.015202580019831657, + 0.022067470476031303, + 0.012456060387194157, + 0.002593138488009572, + 0.007996116764843464, + 0.0029701765161007643, + 0.0030561599414795637, + 0.005942595191299915, + 0.004202706273645163, + 0.004055787809193134, + 0.009320736862719059, + 0.004200783092528582, + 0.0041413926519453526, + 0.0023339116014540195, + 0.0022159686777740717, + 0.006810743827372789, + 0.00397976441308856, + 0.0027295288164168596, + 0.004076261539012194, + 0.013844632543623447, + 0.1489458531141281, + 0.009697487577795982, + 0.005744368303567171, + 0.15318956971168518, + 0.010383223183453083, + 0.007474624086171389, + 0.017496313899755478, + 0.001892339438199997, + 0.0037714603822678328, + 0.0013015224831178784, + 0.001520543941296637, + 0.004280650522559881, + 0.000961217621807009, + 0.001477466314099729, + 0.0012073707766830921, + 0.0031720013357698917, + 0.002844954840838909, + 0.002605430083349347, + 0.004592398647218943, + 0.0032587586902081966, + 0.002332117175683379, + 0.0027953479439020157, + 0.00316489115357399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.023306243121623993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606719382107258 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.01606719382107258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606719382107258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1151401475071907, + 0.0850893035531044, + 0.03341694548726082, + 0.04606521502137184, + 0.10227527469396591, + 0.23510287702083588, + 0.02477993443608284, + 0.2811889052391052, + 0.05837619677186012, + 0.01856524683535099 + ], + "credit_map": [ + 0.011632432229816914, + 0.01276788767427206, + 0.0215697530657053, + 0.007663240656256676, + 0.015603565610945225, + 0.011810200288891792, + 0.013620482757687569, + 0.020472578704357147, + 0.010750962421298027, + 0.011204972863197327, + 0.011597597040235996, + 0.0052101281471550465, + 0.009113584645092487, + 0.012953003868460655, + 0.006863280665129423, + 0.01739577017724514, + 0.0037991127464920282, + 0.0038764611817896366, + 0.0047593750059604645, + 0.0031878699082881212, + 0.003246884560212493, + 0.006095822900533676, + 0.0031916534062474966, + 0.005259765312075615, + 0.0066429669968783855, + 0.00603129155933857, + 0.0076177180744707584, + 0.002696059178560972, + 0.006193982902914286, + 0.006623246241360903, + 0.0038629984483122826, + 0.006396949291229248, + 0.007444432470947504, + 0.015367249958217144, + 0.008045216090977192, + 0.012044861912727356, + 0.015456335619091988, + 0.014240816235542297, + 0.017560211941599846, + 0.012116143479943275, + 0.012289531528949738, + 0.06932368129491806, + 0.008205143734812737, + 0.008857582695782185, + 0.028726739808917046, + 0.02211720123887062, + 0.024205833673477173, + 0.061377156525850296, + 0.002156586153432727, + 0.002756464760750532, + 0.002713119611144066, + 0.0025438580196350813, + 0.0061138952150940895, + 0.0036545328330248594, + 0.0018893973901867867, + 0.002952081384137273, + 0.01050314400345087, + 0.11456354707479477, + 0.00717884860932827, + 0.004158095456659794, + 0.11739238351583481, + 0.007586270570755005, + 0.005650561768561602, + 0.014156054705381393, + 0.005677737761288881, + 0.007534857839345932, + 0.0017083456041291356, + 0.0041628647595644, + 0.017038729041814804, + 0.0032728842925280333, + 0.015313887037336826, + 0.003666888689622283, + 0.0019323205342516303, + 0.003167380578815937, + 0.0014557993272319436, + 0.0028020921163260937, + 0.0035039016511291265, + 0.0015495623229071498, + 0.0018887127516791224, + 0.0022654773201793432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.01606719382107258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631078828126192 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631078828126192, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631078828126192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12723009288311005, + 0.10163016617298126, + 0.0513770692050457, + 0.0946318507194519, + 0.09825011342763901, + 0.17765696346759796, + 0.032597433775663376, + 0.1989295482635498, + 0.0680549368262291, + 0.04964178428053856 + ], + "credit_map": [ + 0.011504491791129112, + 0.017019832506775856, + 0.022500399500131607, + 0.009492503479123116, + 0.014910993166267872, + 0.014217155054211617, + 0.0171087346971035, + 0.020475979894399643, + 0.012150353752076626, + 0.006881834007799625, + 0.015112318098545074, + 0.010189685970544815, + 0.00776969688013196, + 0.015464642085134983, + 0.016526903957128525, + 0.017534729093313217, + 0.006880346219986677, + 0.007237140089273453, + 0.007961701601743698, + 0.00654026260599494, + 0.004622326698154211, + 0.006646119058132172, + 0.006391344591975212, + 0.00509782787412405, + 0.014596677385270596, + 0.009021867997944355, + 0.01328448299318552, + 0.006597492843866348, + 0.010210897773504257, + 0.015528679825365543, + 0.010221417061984539, + 0.01517033576965332, + 0.011017520911991596, + 0.016362551599740982, + 0.010014588944613934, + 0.008664483204483986, + 0.016619760543107986, + 0.009945966303348541, + 0.014928520657122135, + 0.010696717537939548, + 0.007916170172393322, + 0.052695490419864655, + 0.007347973063588142, + 0.0065967123955488205, + 0.01812007836997509, + 0.01779080368578434, + 0.015755081549286842, + 0.05143466219305992, + 0.0036631939001381397, + 0.0038298950530588627, + 0.003252046648412943, + 0.003996244631707668, + 0.007500619627535343, + 0.004486975725740194, + 0.002261352725327015, + 0.00360710802488029, + 0.009086810983717442, + 0.07825621962547302, + 0.005623376928269863, + 0.005385223310440779, + 0.07928896695375443, + 0.00731379259377718, + 0.004553136415779591, + 0.00942203402519226, + 0.006478456314653158, + 0.006449599750339985, + 0.0024945386685431004, + 0.0069037918001413345, + 0.01795366033911705, + 0.0033535815309733152, + 0.020222997292876244, + 0.004198309034109116, + 0.0053164297714829445, + 0.008493395522236824, + 0.0028805206529796124, + 0.008468550629913807, + 0.010554947890341282, + 0.0036089783534407616, + 0.004634374286979437, + 0.005684587638825178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.007631078828126192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053884506225586 + ], + "opposite_logprob_tokens": [ + -0.42888450622558594 + ], + "logprob_reward": -1.053884506225586, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.053884506225586, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.26931560039520264, + 0.04251381754875183, + 0.016072938218712807, + 0.017561325803399086, + 0.02514226920902729, + 0.2625279724597931, + 0.010571320541203022, + 0.1335824728012085, + 0.21583792567253113, + 0.006874397397041321 + ], + "credit_map": [ + 0.0327533520758152, + 0.01662507653236389, + 0.061860620975494385, + 0.020166251808404922, + 0.021108971908688545, + 0.032487671822309494, + 0.033436767756938934, + 0.05087687447667122, + 0.005788072478026152, + 0.0045083677396178246, + 0.008874370716512203, + 0.0020280296448618174, + 0.004623008891940117, + 0.005535793025046587, + 0.0035755173303186893, + 0.0075806560926139355, + 0.0021643268410116434, + 0.0013968400890007615, + 0.0022294525988399982, + 0.0012650468852370977, + 0.0012828613398596644, + 0.003157398197799921, + 0.0012558395974338055, + 0.003321171971037984, + 0.0032113806810230017, + 0.002673480426892638, + 0.00217851297929883, + 0.0008392782765440643, + 0.0024264100939035416, + 0.002530698198825121, + 0.0010620623361319304, + 0.002639500889927149, + 0.002261475194245577, + 0.0035705280024558306, + 0.0022049075923860073, + 0.003086848184466362, + 0.003073841566219926, + 0.004039084538817406, + 0.003652120940387249, + 0.0032534634228795767, + 0.012586986646056175, + 0.08007671684026718, + 0.007390941493213177, + 0.00875299796462059, + 0.028668716549873352, + 0.026253199204802513, + 0.023698441684246063, + 0.07509996742010117, + 0.0008508323808200657, + 0.0011186606716364622, + 0.0011816498590633273, + 0.0012368520256131887, + 0.002821187721565366, + 0.0013905385276302695, + 0.0006904234760440886, + 0.001281175296753645, + 0.00503937155008316, + 0.05431486293673515, + 0.0033266821410506964, + 0.0020540570840239525, + 0.055712491273880005, + 0.003819874022156, + 0.002715763635933399, + 0.0065993815660476685, + 0.02144360914826393, + 0.03506649658083916, + 0.005972042214125395, + 0.011831655167043209, + 0.07052706927061081, + 0.011771549470722675, + 0.048290904611349106, + 0.010934589430689812, + 0.001007190439850092, + 0.0008110917406156659, + 0.0005818200297653675, + 0.0008670731913298368, + 0.0010091420263051987, + 0.0006645031971856952, + 0.001113455044105649, + 0.0008201219607144594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -1.053884506225586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In Hirohiko Araki's signature high-contrast manga style, the camera, in a sharp, dizzying Hitchcockian dolly zoom, zooms in on a massive, brown jellyfish suspended in an abstract mid-air. The creature's umbrella-like body, translucent like smoky quartz, is covered in intricate geometric patterns and biological veins, as clear as an anatomical diagram, its edges defined by sharp, hard black lines. In an instant, it arches its body, its umbrella-like structure contracting with a sense of monstrous biological power, as if accumulating a massive surge of energy. Immediately, countless tentacles beneath it, previously lithe and floating, instantly straighten and explode in all directions, as if struck by a high-voltage electrical current. Each one becomes as hard and erect as a needle of poisoned glass, its tip gleaming with an ominous, cold light. As the camera moves forward and the lens pulls away simultaneously, the background space is extremely compressed and distorted, creating a strong visual impact and a suffocating sense of oppression. The entire picture is filled with an imminent and unimaginable sense of danger in the harsh light and shadow and surreal dynamics.", + "seed": 28013, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_15/index_2226_seed_28013_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.170056289434433, + "vlm_logprob": -0.170056289434433, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.170056289434433, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013459768146276474 + ], + "opposite_logprob_tokens": [ + -4.3884596824646 + ], + "logprob_reward": -0.013459768146276474, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013459768146276474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17515674233436584, + 0.09483059495687485, + 0.331459641456604, + 0.11220856010913849, + 0.017191872000694275, + 0.021157681941986084, + 0.01906563527882099, + 0.13262802362442017, + 0.019148772582411766, + 0.07715250551700592 + ], + "credit_map": [ + 0.018808752298355103, + 0.015800118446350098, + 0.029791392385959625, + 0.016614539548754692, + 0.018531665205955505, + 0.027218792587518692, + 0.01632317528128624, + 0.032068315893411636, + 0.009391214698553085, + 0.005631639622151852, + 0.019955240190029144, + 0.006633375771343708, + 0.0059547205455601215, + 0.01819610968232155, + 0.006937625352293253, + 0.02213067188858986, + 0.03668292984366417, + 0.007986962795257568, + 0.032075293362140656, + 0.01495411153882742, + 0.009448494762182236, + 0.09691890329122543, + 0.013977425172924995, + 0.1194155141711235, + 0.02267405390739441, + 0.00929213222116232, + 0.016324343159794807, + 0.016859007999300957, + 0.010863388888537884, + 0.010179571807384491, + 0.011438140645623207, + 0.014577914960682392, + 0.002046359935775399, + 0.0019540064968168736, + 0.003929843660444021, + 0.001902631251141429, + 0.0017653349786996841, + 0.0012321972753852606, + 0.00202821078710258, + 0.002333289012312889, + 0.002851912984624505, + 0.00250546564348042, + 0.003356142668053508, + 0.0031964543741196394, + 0.002120929304510355, + 0.002075484488159418, + 0.002649246482178569, + 0.002402045764029026, + 0.002832610160112381, + 0.0030767812859266996, + 0.0017540393164381385, + 0.0012748857261613011, + 0.0028031456749886274, + 0.0030002063140273094, + 0.0022658156231045723, + 0.0020581509452313185, + 0.008385001681745052, + 0.04980092495679855, + 0.0062226769514381886, + 0.0047645713202655315, + 0.007684377953410149, + 0.0060470569878816605, + 0.0032194582745432854, + 0.04650396481156349, + 0.0022962831426411867, + 0.003890151623636484, + 0.002132267225533724, + 0.0022654824424535036, + 0.0021981478203088045, + 0.0022128901910036802, + 0.002870996482670307, + 0.0012825526064261794, + 0.005124582443386316, + 0.004150932654738426, + 0.025191400200128555, + 0.003422326408326626, + 0.0022507624235004187, + 0.004651512019336224, + 0.003771701594814658, + 0.028589289635419846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.013459768146276474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867555946111679 + ], + "opposite_logprob_tokens": [ + -3.048675537109375 + ], + "logprob_reward": -0.04867555946111679, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867555946111679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11844778805971146, + 0.06530317664146423, + 0.25084319710731506, + 0.06520190834999084, + 0.11056695878505707, + 0.06553962826728821, + 0.047771554440259933, + 0.19984619319438934, + 0.01596767082810402, + 0.06051189452409744 + ], + "credit_map": [ + 0.01305757649242878, + 0.007630992215126753, + 0.026818424463272095, + 0.009484514594078064, + 0.009885329753160477, + 0.01890714094042778, + 0.010532835498452187, + 0.022130979225039482, + 0.00605703704059124, + 0.0085222152993083, + 0.008362977765500546, + 0.007330683525651693, + 0.008218267001211643, + 0.009751852601766586, + 0.007451048120856285, + 0.009609098546206951, + 0.02764054387807846, + 0.006980597507208586, + 0.027904115617275238, + 0.010174079798161983, + 0.006751334760338068, + 0.06897784769535065, + 0.008842979557812214, + 0.09357170760631561, + 0.00827073585242033, + 0.004872521385550499, + 0.012978572398424149, + 0.010030874982476234, + 0.0055971909314394, + 0.006062770262360573, + 0.007703147362917662, + 0.00968609843403101, + 0.013646257109940052, + 0.008794120512902737, + 0.016559701412916183, + 0.018193120136857033, + 0.01052266824990511, + 0.010614017024636269, + 0.015594174154102802, + 0.01664290577173233, + 0.007886108011007309, + 0.007961300201714039, + 0.012504064477980137, + 0.007035135757178068, + 0.007233538664877415, + 0.0071418373845517635, + 0.008535237982869148, + 0.007242408115416765, + 0.006086600013077259, + 0.008690780028700829, + 0.004339062608778477, + 0.004763831850141287, + 0.007684530224651098, + 0.006222718395292759, + 0.004582470748573542, + 0.005401568952947855, + 0.012828608974814415, + 0.07438531517982483, + 0.009326785802841187, + 0.006473998539149761, + 0.011911878362298012, + 0.009855146519839764, + 0.0048148492351174355, + 0.07024962455034256, + 0.0017598180565983057, + 0.0028752677608281374, + 0.0015426439931616187, + 0.001883054501377046, + 0.0018493407405912876, + 0.002206823555752635, + 0.002554627601057291, + 0.0012960955500602722, + 0.004080052021890879, + 0.0030458925757557154, + 0.019222276285290718, + 0.0026390370912849903, + 0.0017942232079803944, + 0.004392876289784908, + 0.0031827425118535757, + 0.022154802456498146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.04867555946111679 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578453317284584 + ], + "opposite_logprob_tokens": [ + -3.895578384399414 + ], + "logprob_reward": -0.020578453317284584, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578453317284584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06114228069782257, + 0.022223129868507385, + 0.24736151099205017, + 0.03876904398202896, + 0.019626710563898087, + 0.022466836497187614, + 0.14529035985469818, + 0.2138039916753769, + 0.02425810880959034, + 0.20505811274051666 + ], + "credit_map": [ + 0.006382793188095093, + 0.005017512943595648, + 0.009764250367879868, + 0.007746134884655476, + 0.005769216921180487, + 0.011339960619807243, + 0.005389830097556114, + 0.009732571430504322, + 0.002145518781617284, + 0.00177117926068604, + 0.0026087393052875996, + 0.0029346062801778316, + 0.0017872407333925366, + 0.0035350003745406866, + 0.0023250428494066, + 0.005115799605846405, + 0.030440641567111015, + 0.006806076038628817, + 0.023167550563812256, + 0.012747845612466335, + 0.007958369329571724, + 0.06803396344184875, + 0.012074356898665428, + 0.08613269776105881, + 0.0037641702219843864, + 0.003060712944716215, + 0.007824943400919437, + 0.006178185343742371, + 0.004126674495637417, + 0.0033003075513988733, + 0.005005574319511652, + 0.005508474074304104, + 0.0019765303004533052, + 0.002329405164346099, + 0.004586225375533104, + 0.0027770688757300377, + 0.0021700779907405376, + 0.0017630839720368385, + 0.0018748498987406492, + 0.002149467822164297, + 0.0034480555914342403, + 0.002607874106615782, + 0.0035492440219968557, + 0.003325866302475333, + 0.0025827805511653423, + 0.002342596184462309, + 0.0023872810415923595, + 0.002223138464614749, + 0.015057752840220928, + 0.02717486582696438, + 0.013084912672638893, + 0.0113131795078516, + 0.028687262907624245, + 0.0195902269333601, + 0.017820170149207115, + 0.012561975978314877, + 0.013472910970449448, + 0.08109194040298462, + 0.00994135346263647, + 0.0070366938598454, + 0.012811977416276932, + 0.0090145543217659, + 0.005178573541343212, + 0.07525595277547836, + 0.0035009088460355997, + 0.0031923914793878794, + 0.0027425484731793404, + 0.0023820435162633657, + 0.0022035043220967054, + 0.0040840706788003445, + 0.004461376927793026, + 0.0016912624705582857, + 0.013494960032403469, + 0.00989536289125681, + 0.06813054531812668, + 0.008198290131986141, + 0.005334058776497841, + 0.01203872635960579, + 0.00942204613238573, + 0.07854410260915756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.020578453317284584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097829420119524 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.0097829420119524, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097829420119524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.091400146484375, + 0.035643868148326874, + 0.17644168436527252, + 0.051473964005708694, + 0.04708041623234749, + 0.04474016651511192, + 0.22647947072982788, + 0.19869396090507507, + 0.014918393455445766, + 0.1131279245018959 + ], + "credit_map": [ + 0.00959853082895279, + 0.006366555579006672, + 0.018341561779379845, + 0.008912786841392517, + 0.0065992120653390884, + 0.014659211970865726, + 0.008645678870379925, + 0.01827661134302616, + 0.0036889088805764914, + 0.0026122217532247305, + 0.00746727641671896, + 0.0024985414929687977, + 0.0025452193804085255, + 0.0052758194506168365, + 0.0024860810954123735, + 0.00906980037689209, + 0.020343223586678505, + 0.003840038087219, + 0.01848364621400833, + 0.006920084357261658, + 0.0040108985267579556, + 0.05081920698285103, + 0.006994251627475023, + 0.06503032892942429, + 0.007708532270044088, + 0.004042590036988258, + 0.009241564199328423, + 0.007355016190558672, + 0.005148354917764664, + 0.004543389659374952, + 0.006172165274620056, + 0.007262354250997305, + 0.00602760910987854, + 0.004720880649983883, + 0.006179433315992355, + 0.007233717478811741, + 0.005404211580753326, + 0.004303772933781147, + 0.006344901863485575, + 0.006865889299660921, + 0.003582810750231147, + 0.006426698993891478, + 0.008440385572612286, + 0.006227919366210699, + 0.004894843325018883, + 0.004365669097751379, + 0.005667642690241337, + 0.005134195555001497, + 0.02035188116133213, + 0.0372765026986599, + 0.017000310122966766, + 0.029552573338150978, + 0.03938574716448784, + 0.025638746097683907, + 0.03799273073673248, + 0.019280970096588135, + 0.01249217800796032, + 0.07352343946695328, + 0.00932939350605011, + 0.006848204880952835, + 0.012286461889743805, + 0.010184748098254204, + 0.005086788907647133, + 0.06894274055957794, + 0.0015886857872828841, + 0.0019616871140897274, + 0.0016158558428287506, + 0.001858469913713634, + 0.0016762128798291087, + 0.002065118635073304, + 0.0030576572753489017, + 0.001094707055017352, + 0.007202249951660633, + 0.004979738499969244, + 0.03791700303554535, + 0.004342344123870134, + 0.0030513282399624586, + 0.0065354639664292336, + 0.005431557539850473, + 0.0436682403087616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.0097829420119524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577847242355347 + ], + "opposite_logprob_tokens": [ + -0.6327847242355347 + ], + "logprob_reward": -0.7577847242355347, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577847242355347, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04779001325368881, + 0.03995736688375473, + 0.14232391119003296, + 0.0175604410469532, + 0.041701361536979675, + 0.05369747057557106, + 0.15563172101974487, + 0.14952655136585236, + 0.07367666810750961, + 0.2781345546245575 + ], + "credit_map": [ + 0.005046915262937546, + 0.004695792682468891, + 0.00902695581316948, + 0.0046456907875835896, + 0.003845038590952754, + 0.007655280642211437, + 0.0037005534395575523, + 0.009173776023089886, + 0.0031916832085698843, + 0.002967990003526211, + 0.00801412295550108, + 0.0035372916609048843, + 0.003056763205677271, + 0.007695546839386225, + 0.004225661512464285, + 0.007268303539603949, + 0.016351984813809395, + 0.004083752166479826, + 0.013920913450419903, + 0.007103638723492622, + 0.0042727054096758366, + 0.04053902626037598, + 0.006761826574802399, + 0.04929004982113838, + 0.002871020929887891, + 0.0016015752917155623, + 0.003333217464387417, + 0.002319858642295003, + 0.0017218609573319554, + 0.0016061781207099557, + 0.0017095875227823853, + 0.0023971409536898136, + 0.005448960233479738, + 0.003389632562175393, + 0.011492284946143627, + 0.005226144101470709, + 0.0034002545289695263, + 0.0030408846214413643, + 0.003343594027683139, + 0.006359603721648455, + 0.0047982302494347095, + 0.0065187448635697365, + 0.010358712635934353, + 0.006994969677180052, + 0.006114231422543526, + 0.005151586141437292, + 0.0076472763903439045, + 0.006113715935498476, + 0.017091354355216026, + 0.02999144420027733, + 0.01280522532761097, + 0.011870545335114002, + 0.031248072162270546, + 0.02165062353014946, + 0.01720687933266163, + 0.013767548836767673, + 0.009031049907207489, + 0.05611308291554451, + 0.0064625367522239685, + 0.006307616364210844, + 0.010032232850790024, + 0.0063605112954974174, + 0.004322994500398636, + 0.05089651793241501, + 0.005495743826031685, + 0.009013863280415535, + 0.006855546031147242, + 0.006755285896360874, + 0.009476928040385246, + 0.006263324525207281, + 0.02243690937757492, + 0.007379056885838509, + 0.017734382301568985, + 0.015428110957145691, + 0.08246566355228424, + 0.014729350805282593, + 0.01034126803278923, + 0.02223816141486168, + 0.01750115118920803, + 0.0976964607834816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.7577847242355347 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In Hirohiko Araki's signature high-contrast manga style, the camera, in a sharp, dizzying Hitchcockian dolly zoom, zooms in on a massive, brown jellyfish suspended in an abstract mid-air. The creature's umbrella-like body, translucent like smoky quartz, is covered in intricate geometric patterns and biological veins, as clear as an anatomical diagram, its edges defined by sharp, hard black lines. In an instant, it arches its body, its umbrella-like structure contracting with a sense of monstrous biological power, as if accumulating a massive surge of energy. Immediately, countless tentacles beneath it, previously lithe and floating, instantly straighten and explode in all directions, as if struck by a high-voltage electrical current. Each one becomes as hard and erect as a needle of poisoned glass, its tip gleaming with an ominous, cold light. As the camera moves forward and the lens pulls away simultaneously, the background space is extremely compressed and distorted, creating a strong visual impact and a suffocating sense of oppression. The entire picture is filled with an imminent and unimaginable sense of danger in the harsh light and shadow and surreal dynamics.", + "seed": 28014, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_15/index_2226_seed_28014_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1685481172055006, + "vlm_logprob": -0.1685481172055006, + "vlm_yes_no_margin": 2.699999952316284, + "vlm_token_credit": -0.1685481172055006, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027527956292033195 + ], + "opposite_logprob_tokens": [ + -3.6525280475616455 + ], + "logprob_reward": -0.027527956292033195, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027527956292033195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13706329464912415, + 0.08987725526094437, + 0.06624814122915268, + 0.2623938024044037, + 0.08230291306972504, + 0.11742306500673294, + 0.08621945977210999, + 0.08517024666070938, + 0.03604574501514435, + 0.03725605458021164 + ], + "credit_map": [ + 0.013680914416909218, + 0.023975003510713577, + 0.013772264122962952, + 0.017238827422261238, + 0.021237283945083618, + 0.017523949965834618, + 0.01271555945277214, + 0.016919493675231934, + 0.007327512837946415, + 0.011828383430838585, + 0.022310219705104828, + 0.006380031816661358, + 0.009149249643087387, + 0.019128495827317238, + 0.0061516305431723595, + 0.007601727265864611, + 0.010260343551635742, + 0.007510334253311157, + 0.007745875511318445, + 0.008669637143611908, + 0.009220636449754238, + 0.008147403597831726, + 0.004733667708933353, + 0.009960244409739971, + 0.013355030678212643, + 0.01199108362197876, + 0.024486416950821877, + 0.010895779356360435, + 0.01304372027516365, + 0.08388125151395798, + 0.01122320257127285, + 0.09351733326911926, + 0.009880492463707924, + 0.012455889955163002, + 0.008045753464102745, + 0.0064881332218647, + 0.017087340354919434, + 0.00856257788836956, + 0.00735563924536109, + 0.012427084147930145, + 0.008451011963188648, + 0.014584471471607685, + 0.008496391586959362, + 0.005565322935581207, + 0.014727918431162834, + 0.03356993570923805, + 0.006468083243817091, + 0.025559931993484497, + 0.012803511694073677, + 0.009198364801704884, + 0.009141923859715462, + 0.011242657899856567, + 0.0088952062651515, + 0.009688655845820904, + 0.01483199093490839, + 0.010417147539556026, + 0.0074641406536102295, + 0.004142141435295343, + 0.0071844118647277355, + 0.005100732669234276, + 0.005690154153853655, + 0.02532963454723358, + 0.005925269797444344, + 0.02433377131819725, + 0.003698004875332117, + 0.004034904297441244, + 0.00618734722957015, + 0.004984613042324781, + 0.00333365797996521, + 0.005196285434067249, + 0.004609535448253155, + 0.004001397639513016, + 0.0031526582315564156, + 0.00528997927904129, + 0.006103024818003178, + 0.006112176459282637, + 0.004383038729429245, + 0.0037409455981105566, + 0.003857417032122612, + 0.004616813268512487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.027527956292033195 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868929833173752 + ], + "opposite_logprob_tokens": [ + -3.048689365386963 + ], + "logprob_reward": -0.04868929833173752, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868929833173752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2067326158285141, + 0.15520808100700378, + 0.09376301616430283, + 0.16800571978092194, + 0.048187319189310074, + 0.09240569174289703, + 0.03825758770108223, + 0.11290806531906128, + 0.0355154350399971, + 0.04901637136936188 + ], + "credit_map": [ + 0.017942624166607857, + 0.03324686363339424, + 0.02069251984357834, + 0.03589661419391632, + 0.033531490713357925, + 0.02146436646580696, + 0.021343033760786057, + 0.022615112364292145, + 0.01881425641477108, + 0.019889816641807556, + 0.025841252878308296, + 0.011912216432392597, + 0.01842515729367733, + 0.026941949501633644, + 0.01114810910075903, + 0.022235332056879997, + 0.011114723980426788, + 0.012028524652123451, + 0.010682468302547932, + 0.008731381967663765, + 0.01637415960431099, + 0.013265803456306458, + 0.008971741423010826, + 0.012594211846590042, + 0.011876394972205162, + 0.010229573585093021, + 0.010314987041056156, + 0.00876337755471468, + 0.009793389588594437, + 0.05768553912639618, + 0.007425505202263594, + 0.05191695690155029, + 0.006653072778135538, + 0.007309150416404009, + 0.005355779081583023, + 0.00398468691855669, + 0.007873939350247383, + 0.005861138924956322, + 0.003839966608211398, + 0.007309583481401205, + 0.0061857751570641994, + 0.008590564131736755, + 0.0068824575282633305, + 0.004299437627196312, + 0.009567088447511196, + 0.027463702484965324, + 0.005455374717712402, + 0.023961292579770088, + 0.00508069945499301, + 0.005087410099804401, + 0.00412358995527029, + 0.004107617307454348, + 0.004638440441340208, + 0.004727024585008621, + 0.005044975318014622, + 0.005447831004858017, + 0.008719922974705696, + 0.006376839708536863, + 0.009404528886079788, + 0.009826436638832092, + 0.007988961413502693, + 0.03606129810214043, + 0.007352252025157213, + 0.027177829295396805, + 0.003183922031894326, + 0.004610504489392042, + 0.004329472314566374, + 0.005606872029602528, + 0.003215308766812086, + 0.0052121360786259174, + 0.0030327681452035904, + 0.006324448622763157, + 0.006137283984571695, + 0.006663357838988304, + 0.006745720747858286, + 0.00652441056445241, + 0.005069894250482321, + 0.007072269916534424, + 0.0031748933251947165, + 0.007628538645803928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.04868929833173752 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495880916714668 + ], + "opposite_logprob_tokens": [ + -2.9299588203430176 + ], + "logprob_reward": -0.05495880916714668, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495880916714668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08859303593635559, + 0.05298984795808792, + 0.0561746321618557, + 0.38162028789520264, + 0.03573557361960411, + 0.273581862449646, + 0.03619733825325966, + 0.024003716185688972, + 0.019783636555075645, + 0.031320005655288696 + ], + "credit_map": [ + 0.009700428694486618, + 0.011417320929467678, + 0.009342009201645851, + 0.01533681619912386, + 0.011001953855156898, + 0.013522057794034481, + 0.009783665649592876, + 0.00848879013210535, + 0.006448286585509777, + 0.005141378380358219, + 0.006619175896048546, + 0.003072712104767561, + 0.005832479801028967, + 0.00975492037832737, + 0.003784756874665618, + 0.012336138635873795, + 0.009449115954339504, + 0.006198727525770664, + 0.006175489164888859, + 0.005692973267287016, + 0.0090284813195467, + 0.00785026978701353, + 0.004933512303978205, + 0.006846064236015081, + 0.022564135491847992, + 0.01127799041569233, + 0.02934603951871395, + 0.011705854907631874, + 0.02097364328801632, + 0.13578802347183228, + 0.01194859016686678, + 0.13801604509353638, + 0.0047627901658415794, + 0.004366736859083176, + 0.00486824382096529, + 0.0030699644703418016, + 0.006783488672226667, + 0.0036740435753017664, + 0.003388924291357398, + 0.004821384325623512, + 0.023814529180526733, + 0.04010888189077377, + 0.02262003719806671, + 0.017948787659406662, + 0.035744402557611465, + 0.0678764209151268, + 0.01725604571402073, + 0.04821276292204857, + 0.004324852954596281, + 0.0038265029434114695, + 0.005158334970474243, + 0.0035759801976382732, + 0.005410342011600733, + 0.00357397785410285, + 0.004093526862561703, + 0.006233818829059601, + 0.002638174220919609, + 0.002769716316834092, + 0.002316988306120038, + 0.0018927669152617455, + 0.0031093659345060587, + 0.004822048358619213, + 0.0029026586562395096, + 0.00355199771001935, + 0.001787408720701933, + 0.002277855761349201, + 0.0026499712839722633, + 0.003633327316492796, + 0.0017702438635751605, + 0.0027134527917951345, + 0.0018948741490021348, + 0.0030565017368644476, + 0.00316965882666409, + 0.0034428888466209173, + 0.005041362252086401, + 0.005043901037424803, + 0.003376305103302002, + 0.004943015985190868, + 0.0033213351853191853, + 0.0029815388843417168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.05495880916714668 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195848912000656 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195848912000656, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018195848912000656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16429801285266876, + 0.12663057446479797, + 0.09325850009918213, + 0.18519669771194458, + 0.09488788992166519, + 0.08242756873369217, + 0.03773662447929382, + 0.1068769171833992, + 0.06683226674795151, + 0.04185496270656586 + ], + "credit_map": [ + 0.01790432445704937, + 0.023426230996847153, + 0.01808529533445835, + 0.026139114052057266, + 0.022181663662195206, + 0.01963096484541893, + 0.017916541546583176, + 0.019013866782188416, + 0.012565093114972115, + 0.010262120515108109, + 0.038891732692718506, + 0.006001012399792671, + 0.010766681283712387, + 0.01641179621219635, + 0.007854270748794079, + 0.023877857252955437, + 0.012850272469222546, + 0.012166239321231842, + 0.011197172105312347, + 0.009727179072797298, + 0.017073987051844597, + 0.010176382027566433, + 0.00783088244497776, + 0.01223637256771326, + 0.011085986159741879, + 0.006807803642004728, + 0.018252337351441383, + 0.006836047861725092, + 0.008533832617104053, + 0.059665508568286896, + 0.007382226642221212, + 0.0666329488158226, + 0.010216102935373783, + 0.015333482064306736, + 0.011217853985726833, + 0.007526040077209473, + 0.01956925168633461, + 0.009175636805593967, + 0.006296616513282061, + 0.015552902594208717, + 0.009408305399119854, + 0.012858252972364426, + 0.00894614402204752, + 0.009407193399965763, + 0.010295797139406204, + 0.012446423061192036, + 0.006870865821838379, + 0.012194577604532242, + 0.0076021673157811165, + 0.003554968861863017, + 0.004665357992053032, + 0.0026745102368295193, + 0.0051240031607449055, + 0.0052771433256566525, + 0.0036174466367810965, + 0.005221022758632898, + 0.009913574904203415, + 0.008659915067255497, + 0.012301499024033546, + 0.013663352467119694, + 0.011020544916391373, + 0.020317574962973595, + 0.013812055811285973, + 0.01718839257955551, + 0.0064347959123551846, + 0.00610191747546196, + 0.009687022306025028, + 0.010371970012784004, + 0.005978861358016729, + 0.008606069721281528, + 0.006689167581498623, + 0.012962453998625278, + 0.005437491927295923, + 0.007351740729063749, + 0.007037883624434471, + 0.007649386767297983, + 0.0031996648758649826, + 0.004009021446108818, + 0.003373498562723398, + 0.003796269651502371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.018195848912000656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693368673324585 + ], + "opposite_logprob_tokens": [ + -0.693368673324585 + ], + "logprob_reward": -0.693368673324585, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693368673324585, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07813115417957306, + 0.06467734277248383, + 0.04462388530373573, + 0.5987442135810852, + 0.038635317236185074, + 0.06619077920913696, + 0.021263210102915764, + 0.021032992750406265, + 0.0461232028901577, + 0.020577916875481606 + ], + "credit_map": [ + 0.007666829042136669, + 0.01566602662205696, + 0.007139530032873154, + 0.009452346712350845, + 0.012558743357658386, + 0.010891295038163662, + 0.006737761665135622, + 0.008018609136343002, + 0.005378916393965483, + 0.005287863779813051, + 0.012331977486610413, + 0.0038384031504392624, + 0.00435949070379138, + 0.010936703532934189, + 0.0032675922848284245, + 0.01927638240158558, + 0.005439767614006996, + 0.006703678984194994, + 0.003480655374005437, + 0.006740420591086149, + 0.007629772182554007, + 0.004467035178095102, + 0.005677600856870413, + 0.004484946373850107, + 0.03363581374287605, + 0.01915530301630497, + 0.052765339612960815, + 0.0220696609467268, + 0.03241048753261566, + 0.20165503025054932, + 0.021250147372484207, + 0.2158023715019226, + 0.004824111238121986, + 0.005196794401854277, + 0.0049922349862754345, + 0.0034098520409315825, + 0.006865989416837692, + 0.0042108152993023396, + 0.0038290396332740784, + 0.005306470673531294, + 0.006967051420360804, + 0.008557651191949844, + 0.00649732630699873, + 0.005025206133723259, + 0.010685701854526997, + 0.014654676429927349, + 0.004341240040957928, + 0.009461916983127594, + 0.0033600227907299995, + 0.0016901979688555002, + 0.003955814987421036, + 0.0012557702139019966, + 0.0012666787952184677, + 0.0025446703657507896, + 0.001335887354798615, + 0.005854163784533739, + 0.0030862872954458, + 0.0017528771422803402, + 0.0026960193645209074, + 0.002693119691684842, + 0.0014729838585481048, + 0.002705702092498541, + 0.0021721443627029657, + 0.004453855566680431, + 0.005572200752794743, + 0.005749067757278681, + 0.00679005729034543, + 0.007725744973868132, + 0.0042676376178860664, + 0.006196802482008934, + 0.003307036589831114, + 0.006514651235193014, + 0.0022082217037677765, + 0.002767853904515505, + 0.002901487285271287, + 0.0025277722161263227, + 0.0023921942338347435, + 0.0033989606890827417, + 0.001502346363849938, + 0.0028790768701583147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.693368673324585 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In Hirohiko Araki's signature high-contrast manga style, the camera, in a sharp, dizzying Hitchcockian dolly zoom, zooms in on a massive, brown jellyfish suspended in an abstract mid-air. The creature's umbrella-like body, translucent like smoky quartz, is covered in intricate geometric patterns and biological veins, as clear as an anatomical diagram, its edges defined by sharp, hard black lines. In an instant, it arches its body, its umbrella-like structure contracting with a sense of monstrous biological power, as if accumulating a massive surge of energy. Immediately, countless tentacles beneath it, previously lithe and floating, instantly straighten and explode in all directions, as if struck by a high-voltage electrical current. Each one becomes as hard and erect as a needle of poisoned glass, its tip gleaming with an ominous, cold light. As the camera moves forward and the lens pulls away simultaneously, the background space is extremely compressed and distorted, creating a strong visual impact and a suffocating sense of oppression. The entire picture is filled with an imminent and unimaginable sense of danger in the harsh light and shadow and surreal dynamics.", + "seed": 28015, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_15/index_2226_seed_28015_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17698450833559037, + "vlm_logprob": -0.17698450833559037, + "vlm_yes_no_margin": 2.4999999523162844, + "vlm_token_credit": -0.17698450833559037, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05628339946269989 + ], + "opposite_logprob_tokens": [ + -2.931283473968506 + ], + "logprob_reward": -0.05628339946269989, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05628339946269989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10082528740167618, + 0.20089563727378845, + 0.09810349345207214, + 0.07097497582435608, + 0.0369182713329792, + 0.4028979241847992, + 0.03143426403403282, + 0.010718070901930332, + 0.03219141811132431, + 0.015040609054267406 + ], + "credit_map": [ + 0.013431147672235966, + 0.012416410259902477, + 0.01450889278203249, + 0.009330579079687595, + 0.01020677387714386, + 0.016986608505249023, + 0.009441913105547428, + 0.01450296863913536, + 0.02183379977941513, + 0.018222657963633537, + 0.03393247723579407, + 0.02072007954120636, + 0.01868857443332672, + 0.034819912165403366, + 0.023921318352222443, + 0.028756842017173767, + 0.019114593043923378, + 0.009695304557681084, + 0.007580919191241264, + 0.010084696114063263, + 0.008772707544267178, + 0.009901946410536766, + 0.012346858158707619, + 0.02060648240149021, + 0.005645413417369127, + 0.008165518753230572, + 0.009910875000059605, + 0.008382686413824558, + 0.009959015995264053, + 0.008494340814650059, + 0.013368476182222366, + 0.007048657163977623, + 0.002793297404423356, + 0.005305913742631674, + 0.008603966794908047, + 0.004385239910334349, + 0.002650283509865403, + 0.003180944360792637, + 0.003452366916462779, + 0.0065462626516819, + 0.03592550754547119, + 0.014162416569888592, + 0.14780277013778687, + 0.013034197501838207, + 0.017144791781902313, + 0.14192183315753937, + 0.013292082585394382, + 0.019614342600107193, + 0.003899773582816124, + 0.00471479119732976, + 0.0033994026016443968, + 0.005215019453316927, + 0.003953450359404087, + 0.004390835762023926, + 0.0028509569820016623, + 0.003010036190971732, + 0.0013174342457205057, + 0.0011674455599859357, + 0.0012550598476082087, + 0.0008532186620868742, + 0.0022164613474160433, + 0.0014521200209856033, + 0.001056324690580368, + 0.001400007284246385, + 0.0038306256756186485, + 0.004851410165429115, + 0.003776507917791605, + 0.0026485901325941086, + 0.005842990707606077, + 0.0041620172560215, + 0.0039001228287816048, + 0.0031791524961590767, + 0.0017276507569476962, + 0.0011138508561998606, + 0.0009492310346104205, + 0.0013777369167655706, + 0.0016118347411975265, + 0.003191277850419283, + 0.004093960393220186, + 0.000975067145191133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.05628339946269989 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210598349571228 + ], + "opposite_logprob_tokens": [ + -2.812105894088745 + ], + "logprob_reward": -0.06210598349571228, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210598349571228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07809451967477798, + 0.07383083552122116, + 0.06924907118082047, + 0.054480548948049545, + 0.08997836709022522, + 0.5067498087882996, + 0.04102369025349617, + 0.04171809181571007, + 0.01603664457798004, + 0.028838371858000755 + ], + "credit_map": [ + 0.008998352102935314, + 0.009322687052190304, + 0.011906459927558899, + 0.007421874441206455, + 0.01040714979171753, + 0.012850831262767315, + 0.006251234095543623, + 0.010935939848423004, + 0.011360845528542995, + 0.004867444280534983, + 0.02072172798216343, + 0.004025597590953112, + 0.004742338787764311, + 0.012756864540278912, + 0.0047189886681735516, + 0.010637032799422741, + 0.006472983863204718, + 0.007881068624556065, + 0.0076899901032447815, + 0.01076136901974678, + 0.007813903503119946, + 0.006863685324788094, + 0.013052811846137047, + 0.008713267743587494, + 0.004647016525268555, + 0.00759548507630825, + 0.007620790973305702, + 0.00550793157890439, + 0.007142229005694389, + 0.0070213377475738525, + 0.008432899601757526, + 0.006512867286801338, + 0.007647509686648846, + 0.01807238720357418, + 0.018789252266287804, + 0.012690259143710136, + 0.007733058650046587, + 0.007251123897731304, + 0.007084627635776997, + 0.010710155591368675, + 0.04469842463731766, + 0.017336931079626083, + 0.18635329604148865, + 0.015272527933120728, + 0.022135818377137184, + 0.18004028499126434, + 0.01684528775513172, + 0.024067282676696777, + 0.003090531099587679, + 0.005049774888902903, + 0.005895825568586588, + 0.008511709980666637, + 0.0053719813004136086, + 0.005121748894453049, + 0.004282000008970499, + 0.0037001220043748617, + 0.007927105762064457, + 0.008683714084327221, + 0.003890241961926222, + 0.002851787256076932, + 0.00614762119948864, + 0.005480811465531588, + 0.0033778147771954536, + 0.0033589976374059916, + 0.0021891938522458076, + 0.0021616509184241295, + 0.0030691714491695166, + 0.0012187360553070903, + 0.001833082758821547, + 0.002081462414935231, + 0.0016431334661319852, + 0.0018402156420052052, + 0.003956279717385769, + 0.0023524814750999212, + 0.0021215546876192093, + 0.003080049529671669, + 0.0035649677738547325, + 0.006554746069014072, + 0.005422335118055344, + 0.0017859585350379348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.06210598349571228 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495982617139816 + ], + "opposite_logprob_tokens": [ + -2.929959774017334 + ], + "logprob_reward": -0.05495982617139816, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495982617139816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10017244517803192, + 0.050531528890132904, + 0.07187387347221375, + 0.03851817175745964, + 0.263526052236557, + 0.22666175663471222, + 0.17654474079608917, + 0.021621987223625183, + 0.033673908561468124, + 0.01687558926641941 + ], + "credit_map": [ + 0.007859203964471817, + 0.01128822285681963, + 0.02692514657974243, + 0.0063320426270365715, + 0.006394859403371811, + 0.024587448686361313, + 0.007162168156355619, + 0.009623355232179165, + 0.005652274005115032, + 0.0053750923834741116, + 0.006496952380985022, + 0.00540441507473588, + 0.00571193965151906, + 0.009487459436058998, + 0.00566577073186636, + 0.006737620569765568, + 0.007944434881210327, + 0.006387665867805481, + 0.006390159483999014, + 0.01244912575930357, + 0.009414132684469223, + 0.006189095322042704, + 0.014582610689103603, + 0.008516651578247547, + 0.004223009571433067, + 0.004620970692485571, + 0.00599122466519475, + 0.004291013348847628, + 0.005177909508347511, + 0.005920108407735825, + 0.004686750005930662, + 0.0036071871872991323, + 0.020548656582832336, + 0.04055137187242508, + 0.03443215787410736, + 0.041803669184446335, + 0.01993723399937153, + 0.015906358137726784, + 0.03174690157175064, + 0.05859970673918724, + 0.020019957795739174, + 0.008936538361012936, + 0.08183671534061432, + 0.006956972181797028, + 0.011717474088072777, + 0.07851430773735046, + 0.007758669555187225, + 0.0109211141243577, + 0.01470453292131424, + 0.02124682441353798, + 0.02406894788146019, + 0.03568026423454285, + 0.024032773450016975, + 0.022336071357131004, + 0.019155915826559067, + 0.015319406986236572, + 0.004299303982406855, + 0.004114337265491486, + 0.00221728696487844, + 0.0014267382211983204, + 0.0023622075095772743, + 0.0031741359271109104, + 0.0017728954553604126, + 0.0022550830617547035, + 0.005633983761072159, + 0.004212623927742243, + 0.0072646308690309525, + 0.0031561150681227446, + 0.002641793107613921, + 0.004239111673086882, + 0.002893460914492607, + 0.003632191801443696, + 0.003404794028028846, + 0.0013887559762224555, + 0.00118648714851588, + 0.002137140603736043, + 0.0016262650024145842, + 0.0027656324673444033, + 0.003381936112418771, + 0.000984576647169888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.05495982617139816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202990293502808 + ], + "opposite_logprob_tokens": [ + -4.018202781677246 + ], + "logprob_reward": -0.018202990293502808, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018202990293502808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07233887910842896, + 0.05598146468400955, + 0.04821344465017319, + 0.04180260747671127, + 0.1173325851559639, + 0.5750224590301514, + 0.03037916123867035, + 0.017114786431193352, + 0.03472890704870224, + 0.007085730321705341 + ], + "credit_map": [ + 0.004977202042937279, + 0.005597526673227549, + 0.02231968194246292, + 0.004281802102923393, + 0.005545003339648247, + 0.02018335647881031, + 0.003219994017854333, + 0.006214306689798832, + 0.008128086104989052, + 0.0045404438860714436, + 0.010994814336299896, + 0.0049003069289028645, + 0.00459513533860445, + 0.008100217208266258, + 0.0063630156219005585, + 0.008359435945749283, + 0.008555668406188488, + 0.003477667225524783, + 0.003860988887026906, + 0.006215345114469528, + 0.004463829565793276, + 0.003126909025013447, + 0.007731451652944088, + 0.010781575925648212, + 0.0032846832182258368, + 0.004124488681554794, + 0.0047396766021847725, + 0.005361873656511307, + 0.005367946345359087, + 0.004760468378663063, + 0.009505643509328365, + 0.004657823592424393, + 0.011008458212018013, + 0.011316387914121151, + 0.02027871273458004, + 0.02027890644967556, + 0.009730638936161995, + 0.008817190304398537, + 0.01597781851887703, + 0.01992446929216385, + 0.05076222121715546, + 0.019546639174222946, + 0.21166624128818512, + 0.01758394204080105, + 0.024881284683942795, + 0.20413565635681152, + 0.019038952887058258, + 0.0274074487388134, + 0.003410514211282134, + 0.004787919111549854, + 0.003284418024122715, + 0.005550843197852373, + 0.004176393151283264, + 0.004118142183870077, + 0.0024784135166555643, + 0.002572514582425356, + 0.0021064942702651024, + 0.0024241232313215733, + 0.0022480976767838, + 0.0016356109408661723, + 0.002105181571096182, + 0.0027594168204814196, + 0.0017285941867157817, + 0.0021072672680020332, + 0.0062731122598052025, + 0.003321439493447542, + 0.008366474881768227, + 0.00377465458586812, + 0.002748314756900072, + 0.00377993774600327, + 0.003201225772500038, + 0.003263743594288826, + 0.0010566562414169312, + 0.0009423592127859592, + 0.0011582346633076668, + 0.0008028766606003046, + 0.0008961489074863493, + 0.0007849257090128958, + 0.0008600125438533723, + 0.0005845158593729138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.018202990293502808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brown jellyfish with an umbrella-like body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's body translucent like smoky quartz and covered in intricate geometric patterns and biological veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tentacles beneath the jellyfish instantly straighten and explode in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the straightened tentacles appear hard and erect like needles of poisoned glass with tips gleaming with cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a dolly zoom (zooming in while pulling away) causing the background to compress and distort?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933703422546387 + ], + "opposite_logprob_tokens": [ + -0.6933703422546387 + ], + "logprob_reward": -0.6933703422546387, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933703422546387, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14528633654117584, + 0.08705496042966843, + 0.05752944573760033, + 0.05254270136356354, + 0.06341870874166489, + 0.38737404346466064, + 0.10032299906015396, + 0.018841946497559547, + 0.06521731615066528, + 0.022411542013287544 + ], + "credit_map": [ + 0.01367854978889227, + 0.01008867658674717, + 0.04164552688598633, + 0.00949468556791544, + 0.008224464021623135, + 0.03913969174027443, + 0.00927957333624363, + 0.013735145330429077, + 0.013528473675251007, + 0.0042429473251104355, + 0.02537630870938301, + 0.005302068777382374, + 0.004685491789132357, + 0.016717925667762756, + 0.004547827877104282, + 0.012653913348913193, + 0.011299975216388702, + 0.005646446254104376, + 0.00406317925080657, + 0.005097455345094204, + 0.00520425196737051, + 0.005158633459359407, + 0.006032944656908512, + 0.015026559121906757, + 0.004900929983705282, + 0.006620093714445829, + 0.007304246071726084, + 0.005688294302672148, + 0.00654254574328661, + 0.005731118377298117, + 0.009739749133586884, + 0.006015720311552286, + 0.0054852948524057865, + 0.020744433626532555, + 0.012732965871691704, + 0.0044607133604586124, + 0.005804242100566626, + 0.008185145445168018, + 0.0024795662611722946, + 0.003526340937241912, + 0.034952662885189056, + 0.013585885055363178, + 0.14197185635566711, + 0.012180392630398273, + 0.016629861667752266, + 0.13711805641651154, + 0.012790929526090622, + 0.01814437285065651, + 0.008454260416328907, + 0.011901157908141613, + 0.01115840021520853, + 0.02202295884490013, + 0.015040948987007141, + 0.012157680466771126, + 0.011284667998552322, + 0.008302927948534489, + 0.003128757467493415, + 0.0023722313344478607, + 0.0023047118447721004, + 0.0013906080275774002, + 0.003338678041473031, + 0.0025551067665219307, + 0.0016556348418816924, + 0.0020962159615010023, + 0.01282202173024416, + 0.004713064059615135, + 0.018900418654084206, + 0.0071712639182806015, + 0.00504986010491848, + 0.005734848324209452, + 0.006257505156099796, + 0.004568323493003845, + 0.004273576196283102, + 0.0019028347451239824, + 0.0020879367366433144, + 0.002776197623461485, + 0.002146610291674733, + 0.002803894691169262, + 0.004980497993528843, + 0.0014399932697415352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.6933703422546387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_tensor([7778, 7778])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_tensor([7778, 7778])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..13418b6b3599ddd6958c90b35cf0ea6979fdfd3e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_tensor([7778, 7778])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, low-angle, static shot frames a quiet, tense scene on the surface of a dark, polished wooden table. A heavy, rectangular block of black ink, with a faint matte, dried-ink residue, looms like a monolith. From the deep shadow cast by this object, a tiny wild mouse cautiously peers out. Its fur is a soft, earthy brown, its beady black eyes wide and flickering with fear, while its delicate whiskers droop. The mouse's head tentatively emerges into the light, its tiny pink nose twitching almost imperceptibly as it sniffs the air. The entire scene is bisected by a powerful beam of light, creating a dramatic black-and-white contrast that transforms the ink into a majestic silhouette while brightly illuminating the mouse's inquisitive face and the rich, dark textures of the table. The atmosphere is quiet and suspenseful, completely focused on the creature's cautious exploration and vulnerability.", + "seed": 16418, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_7778_seed_16418_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1845004478469491, + "vlm_logprob": -0.1845004478469491, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.1845004478469491, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5782026648521423 + ], + "opposite_logprob_tokens": [ + -0.8282026648521423 + ], + "logprob_reward": -0.5782026648521423, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5782026648521423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41659611463546753, + 0.09725558757781982, + 0.04774646833539009, + 0.12334813922643661, + 0.16885869204998016, + 0.043087974190711975, + 0.03208906948566437, + 0.017862536013126373, + 0.013425920158624649, + 0.0397295206785202 + ], + "credit_map": [ + 0.07977218925952911, + 0.06414366513490677, + 0.040151096880435944, + 0.037080004811286926, + 0.051780521869659424, + 0.07361549884080887, + 0.0428951233625412, + 0.027158034965395927, + 0.017467258498072624, + 0.014994855970144272, + 0.00963982567191124, + 0.005964073818176985, + 0.013811358250677586, + 0.01502115000039339, + 0.012903132475912571, + 0.007453933358192444, + 0.007011881098151207, + 0.007644393015652895, + 0.006135123781859875, + 0.003983053378760815, + 0.008228398859500885, + 0.0061155217699706554, + 0.005115856416523457, + 0.0035122432745993137, + 0.017747148871421814, + 0.01615043357014656, + 0.013044660910964012, + 0.018397178500890732, + 0.01652059331536293, + 0.011421137489378452, + 0.020864374935626984, + 0.009202612563967705, + 0.025812903419137, + 0.015085435472428799, + 0.028233734890818596, + 0.026202159002423286, + 0.018597645685076714, + 0.020614128559827805, + 0.0185858067125082, + 0.015726886689662933, + 0.006453768350183964, + 0.0049649751745164394, + 0.004341652151197195, + 0.0046462551690638065, + 0.005583200603723526, + 0.006043451838195324, + 0.007463595364242792, + 0.0035910767037421465, + 0.00557728623971343, + 0.003916902467608452, + 0.004231327213346958, + 0.003964449744671583, + 0.002797509077936411, + 0.005683281924575567, + 0.003234689123928547, + 0.0026836220640689135, + 0.0029443022795021534, + 0.0024943421594798565, + 0.0018072586972266436, + 0.001704006688669324, + 0.0020211052615195513, + 0.0026970573235303164, + 0.002208148129284382, + 0.0019863161724060774, + 0.0017094970680773258, + 0.0017016975907608867, + 0.0013811459066346288, + 0.0024381207767874002, + 0.001407216303050518, + 0.0018745503621175885, + 0.0016459724865853786, + 0.0012677195481956005, + 0.006395476870238781, + 0.0036761935334652662, + 0.003926320932805538, + 0.004657575394958258, + 0.0052874148823320866, + 0.005313154775649309, + 0.007302474696189165, + 0.0031709095928817987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.5782026648521423 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25212621688842773 + ], + "opposite_logprob_tokens": [ + -1.5021262168884277 + ], + "logprob_reward": -0.25212621688842773, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25212621688842773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3281264007091522, + 0.10636226087808609, + 0.05317745730280876, + 0.07493322342634201, + 0.08293136209249496, + 0.06615311652421951, + 0.060326896607875824, + 0.031734589487314224, + 0.09602203220129013, + 0.10023269802331924 + ], + "credit_map": [ + 0.06210573762655258, + 0.05557942017912865, + 0.030155861750245094, + 0.02853349968791008, + 0.04810439422726631, + 0.0473223440349102, + 0.03736412152647972, + 0.01896100677549839, + 0.01521318405866623, + 0.019789373502135277, + 0.009586709551513195, + 0.007691333070397377, + 0.017928246408700943, + 0.010391738265752792, + 0.01685342751443386, + 0.008908255025744438, + 0.009346626698970795, + 0.010065920650959015, + 0.004710059612989426, + 0.004373148083686829, + 0.007082012481987476, + 0.00621917936950922, + 0.00692156795412302, + 0.004458942916244268, + 0.011153538711369038, + 0.009229800663888454, + 0.010521975345909595, + 0.011182831600308418, + 0.0075938524678349495, + 0.009059535339474678, + 0.008547594770789146, + 0.007644096855074167, + 0.014727527275681496, + 0.012538405135273933, + 0.008984878659248352, + 0.007502379361540079, + 0.011979923583567142, + 0.013297397643327713, + 0.009038291871547699, + 0.004862558096647263, + 0.009120265021920204, + 0.006784952245652676, + 0.008053095079958439, + 0.009413723833858967, + 0.008154621347784996, + 0.009822092019021511, + 0.00972037948668003, + 0.005083987023681402, + 0.010142028331756592, + 0.00701595563441515, + 0.005934794899076223, + 0.006664694286882877, + 0.008508292958140373, + 0.009004013612866402, + 0.0072768982499837875, + 0.00578021677210927, + 0.005071387626230717, + 0.0035392853897064924, + 0.0027752204332500696, + 0.0028167213313281536, + 0.0040581184439361095, + 0.004441035445779562, + 0.005593506619334221, + 0.0034393160603940487, + 0.013425678946077824, + 0.012180694378912449, + 0.011695929802954197, + 0.014645345509052277, + 0.012587957084178925, + 0.012161580845713615, + 0.011516240425407887, + 0.00780859962105751, + 0.015772171318531036, + 0.010984228923916817, + 0.014207794331014156, + 0.009508563205599785, + 0.013699539005756378, + 0.014146652072668076, + 0.013659028336405754, + 0.00825471244752407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.25212621688842773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033743396401405334 + ], + "opposite_logprob_tokens": [ + -3.408743381500244 + ], + "logprob_reward": -0.033743396401405334, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033743396401405334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4340149164199829, + 0.10866051912307739, + 0.048897795379161835, + 0.12322597950696945, + 0.15045273303985596, + 0.04809439554810524, + 0.021938350051641464, + 0.01367954071611166, + 0.01320386491715908, + 0.03783191367983818 + ], + "credit_map": [ + 0.08200427144765854, + 0.06389117240905762, + 0.04515685886144638, + 0.03965866565704346, + 0.05110343545675278, + 0.07810269296169281, + 0.04463823512196541, + 0.029459573328495026, + 0.018476663157343864, + 0.01637565717101097, + 0.012299305759370327, + 0.007204544264823198, + 0.015171515755355358, + 0.015755891799926758, + 0.014514180831611156, + 0.008862755261361599, + 0.007536664139479399, + 0.007003140170127153, + 0.0065734232775866985, + 0.004506487865000963, + 0.007704615592956543, + 0.005965336691588163, + 0.00605156272649765, + 0.0035565663129091263, + 0.019196270033717155, + 0.01612287014722824, + 0.012137927114963531, + 0.016361864283680916, + 0.016476722434163094, + 0.014269952662289143, + 0.02038334682583809, + 0.008277024142444134, + 0.025553321465849876, + 0.012543454766273499, + 0.02597237192094326, + 0.02361479587852955, + 0.013342123478651047, + 0.02122812531888485, + 0.014274157583713531, + 0.01392438355833292, + 0.008064419962465763, + 0.004454148001968861, + 0.0048771388828754425, + 0.005589968990534544, + 0.005318851675838232, + 0.008324598893523216, + 0.0073232692666351795, + 0.004142002668231726, + 0.003309184918180108, + 0.0023132816422730684, + 0.0032363110221922398, + 0.0038707898929715157, + 0.0023287103977054358, + 0.002711720997467637, + 0.0020572987850755453, + 0.002111053327098489, + 0.0021523863542824984, + 0.0010879321489483118, + 0.002253481652587652, + 0.0017133993096649647, + 0.0014569145860150456, + 0.0020095515064895153, + 0.0015782249392941594, + 0.0014276497531682253, + 0.0018152020638808608, + 0.002119982149451971, + 0.0012200295459479094, + 0.0023222146555781364, + 0.0013367879437282681, + 0.002042921259999275, + 0.00148424890358001, + 0.0008624779293313622, + 0.006976924370974302, + 0.003282542573288083, + 0.004122484941035509, + 0.003808416659012437, + 0.004533754196017981, + 0.00611763633787632, + 0.00626028748229146, + 0.002729865023866296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.033743396401405334 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865103214979172 + ], + "opposite_logprob_tokens": [ + -3.0486509799957275 + ], + "logprob_reward": -0.04865103214979172, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865103214979172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38281041383743286, + 0.05730048567056656, + 0.09922599047422409, + 0.22306975722312927, + 0.06959235668182373, + 0.049036819487810135, + 0.03169626742601395, + 0.035915572196245193, + 0.016046108677983284, + 0.035306256264448166 + ], + "credit_map": [ + 0.058754369616508484, + 0.0523591972887516, + 0.05606042966246605, + 0.041052307933568954, + 0.042392078787088394, + 0.060510020703077316, + 0.035505786538124084, + 0.03617621213197708, + 0.008493796922266483, + 0.01106408890336752, + 0.008808891288936138, + 0.004665265791118145, + 0.008467961102724075, + 0.006983237806707621, + 0.005099743604660034, + 0.003717498853802681, + 0.014272278174757957, + 0.013162254355847836, + 0.018501976504921913, + 0.008354442194104195, + 0.010627664625644684, + 0.01074558962136507, + 0.00997935514897108, + 0.013582435436546803, + 0.048771120607852936, + 0.026722952723503113, + 0.013739940710365772, + 0.020001251250505447, + 0.026244524866342545, + 0.04741010442376137, + 0.028329098597168922, + 0.011850758455693722, + 0.010696537792682648, + 0.009802955202758312, + 0.009030723012983799, + 0.009716848842799664, + 0.008076492697000504, + 0.008990294300019741, + 0.0073686037212610245, + 0.005909896921366453, + 0.007042160723358393, + 0.0047052958980202675, + 0.006762902718037367, + 0.0074694519862532616, + 0.005172775126993656, + 0.007839196361601353, + 0.005187544040381908, + 0.004857494030147791, + 0.005181430373340845, + 0.002895406447350979, + 0.0043744854629039764, + 0.0028687489684671164, + 0.004037527833133936, + 0.0062929196283221245, + 0.003156021237373352, + 0.002889726310968399, + 0.005441066808998585, + 0.0035782933700829744, + 0.0035956103820353746, + 0.0044774701818823814, + 0.004114377778023481, + 0.005564915016293526, + 0.004661913495510817, + 0.004481922835111618, + 0.0021392174530774355, + 0.00245351018384099, + 0.0019154964247718453, + 0.0027438406832516193, + 0.001821108628064394, + 0.0015896701952442527, + 0.0016673121135681868, + 0.0017159528797492385, + 0.006023922935128212, + 0.004256440792232752, + 0.004379988182336092, + 0.003204754553735256, + 0.006044155918061733, + 0.005782672669738531, + 0.0033696680329740047, + 0.002244652481749654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.04865103214979172 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778928942978382 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778928942978382, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778928942978382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2725121080875397, + 0.07875112444162369, + 0.10127129405736923, + 0.11570549011230469, + 0.12094510346651077, + 0.03234897181391716, + 0.025072263553738594, + 0.031194545328617096, + 0.019468611106276512, + 0.20273049175739288 + ], + "credit_map": [ + 0.04463231563568115, + 0.04834478721022606, + 0.032056476920843124, + 0.02720625326037407, + 0.03727088123559952, + 0.029813703149557114, + 0.031839631497859955, + 0.021348072215914726, + 0.013305718079209328, + 0.01674472913146019, + 0.006341640371829271, + 0.006611388176679611, + 0.011471904814243317, + 0.008430582471191883, + 0.011451800353825092, + 0.004393376875668764, + 0.01809658482670784, + 0.02107859216630459, + 0.006260913331061602, + 0.008119828999042511, + 0.014657649211585522, + 0.012103892862796783, + 0.013835051096975803, + 0.007118783425539732, + 0.026102405041456223, + 0.013032940216362476, + 0.007331675849854946, + 0.010515768080949783, + 0.012481853365898132, + 0.025437505915760994, + 0.015367529354989529, + 0.005435817874968052, + 0.021245243027806282, + 0.010594471357762814, + 0.013335948809981346, + 0.014063709415495396, + 0.014325699768960476, + 0.022686617448925972, + 0.015390394255518913, + 0.009303024969995022, + 0.004711450077593327, + 0.004105171654373407, + 0.002797190099954605, + 0.003901833901181817, + 0.0049359966069459915, + 0.005084608215838671, + 0.004356981720775366, + 0.002455740701407194, + 0.003939953166991472, + 0.002660837722942233, + 0.0018784572603181005, + 0.0019073189469054341, + 0.004274498671293259, + 0.0035132227931171656, + 0.004857025574892759, + 0.0020409519784152508, + 0.005311830434948206, + 0.0028517625760287046, + 0.0030528074130415916, + 0.0029696358833462, + 0.004025894682854414, + 0.00451514171436429, + 0.005497870966792107, + 0.0029696032870560884, + 0.003306965110823512, + 0.0027677372563630342, + 0.0022438575979322195, + 0.002563504735007882, + 0.002146677812561393, + 0.0028718411922454834, + 0.00203282805159688, + 0.0015352022601291537, + 0.026270253583788872, + 0.019640544429421425, + 0.03267650678753853, + 0.020211944356560707, + 0.024659626185894012, + 0.02736547961831093, + 0.028603430837392807, + 0.023302700370550156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.009778928942978382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot frames a quiet, tense scene on the surface of a dark, polished wooden table. A heavy, rectangular block of black ink, with a faint matte, dried-ink residue, looms like a monolith. From the deep shadow cast by this object, a tiny wild mouse cautiously peers out. Its fur is a soft, earthy brown, its beady black eyes wide and flickering with fear, while its delicate whiskers droop. The mouse's head tentatively emerges into the light, its tiny pink nose twitching almost imperceptibly as it sniffs the air. The entire scene is bisected by a powerful beam of light, creating a dramatic black-and-white contrast that transforms the ink into a majestic silhouette while brightly illuminating the mouse's inquisitive face and the rich, dark textures of the table. The atmosphere is quiet and suspenseful, completely focused on the creature's cautious exploration and vulnerability.", + "seed": 16419, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_7778_seed_16419_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09930368475615978, + "vlm_logprob": -0.09930368475615978, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.09930368475615978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12129654735326767 + ], + "opposite_logprob_tokens": [ + -2.2462966442108154 + ], + "logprob_reward": -0.12129654735326767, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12129654735326767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24137617647647858, + 0.07571303099393845, + 0.05471612513065338, + 0.032324329018592834, + 0.03700219467282295, + 0.07154444605112076, + 0.11665818840265274, + 0.0774693563580513, + 0.27022793889045715, + 0.02296825312077999 + ], + "credit_map": [ + 0.02035408467054367, + 0.036489296704530716, + 0.023163767531514168, + 0.0225725956261158, + 0.03296248987317085, + 0.041599929332733154, + 0.02874046191573143, + 0.03549357131123543, + 0.004977639298886061, + 0.01255616545677185, + 0.006357718259096146, + 0.00905627477914095, + 0.013258855789899826, + 0.01103869266808033, + 0.010384132154285908, + 0.00808355025947094, + 0.0038054233882576227, + 0.00666192639619112, + 0.003929676953703165, + 0.004948603454977274, + 0.005487948656082153, + 0.011418260633945465, + 0.007331942208111286, + 0.011132343672215939, + 0.0030221552588045597, + 0.0039011570625007153, + 0.0036117732524871826, + 0.0039571127854287624, + 0.0048976922407746315, + 0.005000051110982895, + 0.004305663052946329, + 0.0036287226248532534, + 0.0028045070357620716, + 0.004134968854486942, + 0.004733247682452202, + 0.0035657049156725407, + 0.006929189898073673, + 0.0054070777259767056, + 0.004982628393918276, + 0.004444871097803116, + 0.006524241995066404, + 0.00948640238493681, + 0.008978491649031639, + 0.007154266349971294, + 0.010381639003753662, + 0.008972720243036747, + 0.012843040749430656, + 0.00720364460721612, + 0.009440899826586246, + 0.02346305549144745, + 0.008131878450512886, + 0.01110552903264761, + 0.01735733635723591, + 0.016423609107732773, + 0.018077751621603966, + 0.01265812385827303, + 0.006950291804969311, + 0.010316048748791218, + 0.009157109074294567, + 0.00876943301409483, + 0.011790052056312561, + 0.011275951750576496, + 0.010345714166760445, + 0.008864754810929298, + 0.020109044387936592, + 0.03419433906674385, + 0.025947803631424904, + 0.03184330463409424, + 0.04481213912367821, + 0.03996959328651428, + 0.04289202764630318, + 0.030459696426987648, + 0.002726386534050107, + 0.0029662298038601875, + 0.002709168242290616, + 0.0026695197448134422, + 0.003079388989135623, + 0.0032838843762874603, + 0.0022578828502446413, + 0.0032757928129285574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.12129654735326767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256171852350235 + ], + "opposite_logprob_tokens": [ + -1.6006171703338623 + ], + "logprob_reward": -0.2256171852350235, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256171852350235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39325958490371704, + 0.06466730684041977, + 0.04694252833724022, + 0.07305353879928589, + 0.034702710807323456, + 0.034680772572755814, + 0.13138529658317566, + 0.07305137068033218, + 0.121026411652565, + 0.02723044715821743 + ], + "credit_map": [ + 0.027269165962934494, + 0.0770379900932312, + 0.03817227482795715, + 0.03304969146847725, + 0.04571952298283577, + 0.05998234823346138, + 0.049311451613903046, + 0.06271716207265854, + 0.005933508276939392, + 0.007532046176493168, + 0.007047364953905344, + 0.008503721095621586, + 0.008752267807722092, + 0.008953248150646687, + 0.009188156574964523, + 0.008756990544497967, + 0.003673041006550193, + 0.007765313610434532, + 0.0024115638807415962, + 0.00367601215839386, + 0.006595173850655556, + 0.008645293302834034, + 0.006014531012624502, + 0.008161596022546291, + 0.004618490114808083, + 0.010180165059864521, + 0.005001738667488098, + 0.006439441815018654, + 0.010422762483358383, + 0.014721010811626911, + 0.009950540959835052, + 0.011719388887286186, + 0.003730165306478739, + 0.005462391301989555, + 0.0032539917156100273, + 0.002945367246866226, + 0.006634808145463467, + 0.004446282051503658, + 0.004777710419148207, + 0.00345199229195714, + 0.00328183569945395, + 0.004021360073238611, + 0.004345329478383064, + 0.0033622703049331903, + 0.0038577388040721416, + 0.005968844518065453, + 0.0048585012555122375, + 0.004984895698726177, + 0.010066988877952099, + 0.0282821673899889, + 0.008631148375570774, + 0.011172113008797169, + 0.019823390990495682, + 0.018449882045388222, + 0.01892903447151184, + 0.0160305667668581, + 0.005764035973697901, + 0.009439880959689617, + 0.007903984747827053, + 0.006943856831640005, + 0.011396437883377075, + 0.011020648293197155, + 0.00958382897078991, + 0.010998694226145744, + 0.009753908962011337, + 0.014780093915760517, + 0.013717427849769592, + 0.015348074957728386, + 0.01754078082740307, + 0.017314905300736427, + 0.0188883189111948, + 0.013682904653251171, + 0.002094215713441372, + 0.002698486903682351, + 0.0030430746264755726, + 0.0030579124577343464, + 0.004464279860258102, + 0.004103899467736483, + 0.0033615920692682266, + 0.0044069853611290455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.2256171852350235 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608901470899582 + ], + "opposite_logprob_tokens": [ + -4.141088962554932 + ], + "logprob_reward": -0.01608901470899582, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608901470899582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22725407779216766, + 0.08878354728221893, + 0.028900496661663055, + 0.038034986704587936, + 0.049893077462911606, + 0.08698391914367676, + 0.18903334438800812, + 0.14966854453086853, + 0.10671530663967133, + 0.03473270311951637 + ], + "credit_map": [ + 0.02163286693394184, + 0.031110726296901703, + 0.024482103064656258, + 0.016074996441602707, + 0.03543185815215111, + 0.0356467179954052, + 0.025810377672314644, + 0.03706441447138786, + 0.007016108836978674, + 0.018103061243891716, + 0.0075483256950974464, + 0.010180586948990822, + 0.013015919364988804, + 0.013227628543972969, + 0.011054137721657753, + 0.008637772873044014, + 0.0022191423922777176, + 0.0047044395469129086, + 0.002292211167514324, + 0.0032490163575857878, + 0.003972518257796764, + 0.004987145308405161, + 0.0032914706971496344, + 0.004184553399682045, + 0.0033648707903921604, + 0.003742608241736889, + 0.002799379173666239, + 0.0029674898833036423, + 0.0037411372177302837, + 0.008773349225521088, + 0.00492392061278224, + 0.00772223062813282, + 0.004460029769688845, + 0.0077740163542330265, + 0.0041398522444069386, + 0.004985372070223093, + 0.011283825151622295, + 0.006639512255787849, + 0.006098298355937004, + 0.0045121717266738415, + 0.007147351745516062, + 0.012435640208423138, + 0.008793233893811703, + 0.008768578059971333, + 0.014475339092314243, + 0.011072424240410328, + 0.014027128927409649, + 0.010264227166771889, + 0.01582328788936138, + 0.04336472600698471, + 0.011879656463861465, + 0.019087443128228188, + 0.02699669823050499, + 0.02598024532198906, + 0.026448704302310944, + 0.019452590495347977, + 0.007927616126835346, + 0.019916485995054245, + 0.01203207392245531, + 0.01463744230568409, + 0.03455955162644386, + 0.022188205271959305, + 0.01865719072520733, + 0.019749967381358147, + 0.007179814390838146, + 0.018483305349946022, + 0.009493151679635048, + 0.014000162482261658, + 0.014196488074958324, + 0.014949670061469078, + 0.014601205475628376, + 0.013811508193612099, + 0.0026834215968847275, + 0.005309120286256075, + 0.004601005464792252, + 0.004268599208444357, + 0.0035740577150136232, + 0.005801502149552107, + 0.0042747072875499725, + 0.004220293834805489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.01608901470899582 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293628811836243 + ], + "opposite_logprob_tokens": [ + -2.23793625831604 + ], + "logprob_reward": -0.11293628811836243, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293628811836243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2638597786426544, + 0.04062900319695473, + 0.057696543633937836, + 0.053789425641298294, + 0.059945859014987946, + 0.24154877662658691, + 0.05985641852021217, + 0.050110895186662674, + 0.11060591042041779, + 0.06195732578635216 + ], + "credit_map": [ + 0.022595403715968132, + 0.04470222815871239, + 0.02046370692551136, + 0.025242922827601433, + 0.03619546443223953, + 0.039452530443668365, + 0.04464128240942955, + 0.030566278845071793, + 0.002844913862645626, + 0.005404253955930471, + 0.00417165644466877, + 0.005097980611026287, + 0.005637151654809713, + 0.006269283127039671, + 0.005979013163596392, + 0.005224753171205521, + 0.004389852285385132, + 0.007507553789764643, + 0.006108259316533804, + 0.00549690006300807, + 0.008868174627423286, + 0.007515942677855492, + 0.010501131415367126, + 0.007308736443519592, + 0.0038367793895304203, + 0.0061326236464083195, + 0.006141085177659988, + 0.005589492153376341, + 0.006345010828226805, + 0.009251349605619907, + 0.008507461287081242, + 0.007985628210008144, + 0.006253623869270086, + 0.008205761201679707, + 0.00813066866248846, + 0.009070309810340405, + 0.0060530006885528564, + 0.00814637541770935, + 0.00518275611102581, + 0.008903367444872856, + 0.024029798805713654, + 0.04227207601070404, + 0.027050215750932693, + 0.02139417454600334, + 0.02690272592008114, + 0.032956261187791824, + 0.03464224562048912, + 0.032301295548677444, + 0.006121004465967417, + 0.010875449515879154, + 0.005891414824873209, + 0.005426270887255669, + 0.006727138534188271, + 0.008447302505373955, + 0.008299854584038258, + 0.008067985065281391, + 0.0031721789855509996, + 0.005887693725526333, + 0.004310770891606808, + 0.005528164096176624, + 0.012434468604624271, + 0.007210078649222851, + 0.006319724023342133, + 0.005247818771749735, + 0.010644471272826195, + 0.011806073598563671, + 0.01823831908404827, + 0.014855791814625263, + 0.012407055124640465, + 0.017028771340847015, + 0.015285955742001534, + 0.010339477099478245, + 0.005255990661680698, + 0.007749519776552916, + 0.005554424598813057, + 0.0057275486178696156, + 0.01387822162359953, + 0.009982393123209476, + 0.006588437594473362, + 0.007220793049782515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.11293628811836243 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579388365149498 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579388365149498, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579388365149498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09311045706272125, + 0.059767335653305054, + 0.026146242395043373, + 0.03919534012675285, + 0.05607045069336891, + 0.06302312761545181, + 0.20135779678821564, + 0.32298797369003296, + 0.046515654772520065, + 0.09182561933994293 + ], + "credit_map": [ + 0.01008357759565115, + 0.01382707990705967, + 0.015810197219252586, + 0.00725387129932642, + 0.007237518206238747, + 0.01717848889529705, + 0.007327367551624775, + 0.014392358250916004, + 0.00400947593152523, + 0.009118105284869671, + 0.006137731485068798, + 0.008568376302719116, + 0.008466597646474838, + 0.008765138685703278, + 0.008144375868141651, + 0.006557539105415344, + 0.0017989762127399445, + 0.0031191417947411537, + 0.0021302986424416304, + 0.0018236698815599084, + 0.002179860370233655, + 0.006424045655876398, + 0.0026864183600991964, + 0.0059838308952748775, + 0.0027478961274027824, + 0.0032360961195081472, + 0.003477781778201461, + 0.0031757624819874763, + 0.003078556153923273, + 0.009515752084553242, + 0.0051165251061320305, + 0.008846974931657314, + 0.00508462218567729, + 0.008879364468157291, + 0.007083836942911148, + 0.006820933893322945, + 0.006598606705665588, + 0.007403185591101646, + 0.007841195911169052, + 0.0063587068580091, + 0.004265631549060345, + 0.008785927668213844, + 0.005708632525056601, + 0.007375661283731461, + 0.009600307792425156, + 0.007768816314637661, + 0.011483797803521156, + 0.008034352213144302, + 0.020056135952472687, + 0.0404033362865448, + 0.018453184515237808, + 0.0210711769759655, + 0.02544317953288555, + 0.027910277247428894, + 0.025013331323862076, + 0.023007182404398918, + 0.017461445182561874, + 0.044851288199424744, + 0.028330372646450996, + 0.029418181627988815, + 0.0746922567486763, + 0.04765964299440384, + 0.03983209654688835, + 0.04074268788099289, + 0.004449397325515747, + 0.004288729280233383, + 0.006273008417338133, + 0.005934475921094418, + 0.005303398706018925, + 0.0074824681505560875, + 0.005986799020320177, + 0.006797379348427057, + 0.011142716743052006, + 0.011974213644862175, + 0.0156828835606575, + 0.008192372508347034, + 0.00943718757480383, + 0.01526299212127924, + 0.008917359635233879, + 0.011215893551707268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.020579388365149498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot frames a quiet, tense scene on the surface of a dark, polished wooden table. A heavy, rectangular block of black ink, with a faint matte, dried-ink residue, looms like a monolith. From the deep shadow cast by this object, a tiny wild mouse cautiously peers out. Its fur is a soft, earthy brown, its beady black eyes wide and flickering with fear, while its delicate whiskers droop. The mouse's head tentatively emerges into the light, its tiny pink nose twitching almost imperceptibly as it sniffs the air. The entire scene is bisected by a powerful beam of light, creating a dramatic black-and-white contrast that transforms the ink into a majestic silhouette while brightly illuminating the mouse's inquisitive face and the rich, dark textures of the table. The atmosphere is quiet and suspenseful, completely focused on the creature's cautious exploration and vulnerability.", + "seed": 16420, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_7778_seed_16420_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03943524919450283, + "vlm_logprob": -0.03943524919450283, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.03943524919450283, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11871254444122314 + ], + "opposite_logprob_tokens": [ + -2.2437124252319336 + ], + "logprob_reward": -0.11871254444122314, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11871254444122314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2980025112628937, + 0.2082740068435669, + 0.06674835830926895, + 0.04649285972118378, + 0.07620953023433685, + 0.026822742074728012, + 0.09394223988056183, + 0.07086385041475296, + 0.10222331434488297, + 0.010420586913824081 + ], + "credit_map": [ + 0.043601248413324356, + 0.04235934838652611, + 0.048695191740989685, + 0.030115358531475067, + 0.03577347472310066, + 0.02628936991095543, + 0.03470558673143387, + 0.0364629402756691, + 0.030029255896806717, + 0.032554708421230316, + 0.03288484364748001, + 0.017230266705155373, + 0.01980789378285408, + 0.02268010936677456, + 0.02418951503932476, + 0.028897419571876526, + 0.005538886412978172, + 0.008966674096882343, + 0.01133735291659832, + 0.010877915658056736, + 0.007558576762676239, + 0.005183473229408264, + 0.00955435074865818, + 0.007731121964752674, + 0.00529193039983511, + 0.005697691347450018, + 0.006975159514695406, + 0.005066275596618652, + 0.006596129387617111, + 0.004394305869936943, + 0.00660764379426837, + 0.00586372334510088, + 0.006629612762480974, + 0.011099474504590034, + 0.015411926433444023, + 0.006631310097873211, + 0.007685260381549597, + 0.009631060063838959, + 0.00933865737169981, + 0.009782232344150543, + 0.002520485082641244, + 0.003172362456098199, + 0.005696430336683989, + 0.0028365477919578552, + 0.0026893680915236473, + 0.0027676343452185392, + 0.003924929071217775, + 0.0032149851322174072, + 0.009680823422968388, + 0.009586617350578308, + 0.0213130135089159, + 0.010091335512697697, + 0.010240091942250729, + 0.008461996912956238, + 0.011018752120435238, + 0.013549604453146458, + 0.0100849699229002, + 0.006878843531012535, + 0.01745617389678955, + 0.005933751817792654, + 0.00782482884824276, + 0.006374317687004805, + 0.006441653706133366, + 0.009869311936199665, + 0.00920841470360756, + 0.008126677945256233, + 0.016185324639081955, + 0.009654604829847813, + 0.014294297434389591, + 0.009423993527889252, + 0.02018589712679386, + 0.015144099481403828, + 0.001033762819133699, + 0.002405373612418771, + 0.001150579540990293, + 0.0011203192407265306, + 0.0012231391156092286, + 0.0010058743646368384, + 0.0012481168378144503, + 0.0012334223138168454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.11871254444122314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813459351658821 + ], + "opposite_logprob_tokens": [ + -3.2881345748901367 + ], + "logprob_reward": -0.03813459351658821, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813459351658821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31560152769088745, + 0.2810659110546112, + 0.05659194663167, + 0.05960828438401222, + 0.027095839381217957, + 0.022991375997662544, + 0.0451715923845768, + 0.06537900120019913, + 0.10963092744350433, + 0.016863549128174782 + ], + "credit_map": [ + 0.05169907212257385, + 0.04179762303829193, + 0.04978916049003601, + 0.023445047438144684, + 0.042770884931087494, + 0.0310165137052536, + 0.03531642630696297, + 0.0397668294608593, + 0.029025929048657417, + 0.030268171802163124, + 0.04148038849234581, + 0.026301437988877296, + 0.03667121380567551, + 0.03750312700867653, + 0.04409494996070862, + 0.035720717161893845, + 0.003722584806382656, + 0.006307991687208414, + 0.005856510251760483, + 0.005089606624096632, + 0.013440366834402084, + 0.009170501492917538, + 0.007215043064206839, + 0.005789346527308226, + 0.005443306174129248, + 0.004686904605478048, + 0.009912526234984398, + 0.005887225270271301, + 0.005891670007258654, + 0.006347389426082373, + 0.012287665158510208, + 0.00915160309523344, + 0.002314745681360364, + 0.00309468200430274, + 0.006377957761287689, + 0.0025445702485740185, + 0.0027641095221042633, + 0.002580268308520317, + 0.003519464982673526, + 0.0039000418037176132, + 0.001780782244168222, + 0.002331844996660948, + 0.005900867749005556, + 0.0027268563862890005, + 0.0014387103728950024, + 0.0018053255043923855, + 0.003028137143701315, + 0.003978853113949299, + 0.004494040738791227, + 0.0043581826612353325, + 0.010502568446099758, + 0.005182321183383465, + 0.004978485871106386, + 0.004116131458431482, + 0.005247274413704872, + 0.006292592268437147, + 0.004596867132931948, + 0.004988594446331263, + 0.00926645752042532, + 0.005513165146112442, + 0.009208595380187035, + 0.009287302382290363, + 0.013012927956879139, + 0.009505094029009342, + 0.010192979127168655, + 0.008602381683886051, + 0.016463911160826683, + 0.009791620075702667, + 0.015610894188284874, + 0.010165481828153133, + 0.022294001653790474, + 0.01650966890156269, + 0.001958171371370554, + 0.0014751767739653587, + 0.003309947671368718, + 0.0014042883412912488, + 0.002073026727885008, + 0.0017646363703534007, + 0.0023694373667240143, + 0.002508866833522916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03813459351658821 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008669240400195122 + ], + "opposite_logprob_tokens": [ + -4.758669376373291 + ], + "logprob_reward": -0.008669240400195122, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008669240400195122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26495301723480225, + 0.13325464725494385, + 0.08367861062288284, + 0.0662383884191513, + 0.04736320674419403, + 0.05085792392492294, + 0.06666997075080872, + 0.08073833584785461, + 0.18971654772758484, + 0.01652933843433857 + ], + "credit_map": [ + 0.043313879519701004, + 0.03166089579463005, + 0.046582382172346115, + 0.02920028194785118, + 0.03591400757431984, + 0.01971902884542942, + 0.027843419462442398, + 0.030719127506017685, + 0.017340557649731636, + 0.02199496515095234, + 0.019327066838741302, + 0.010778888128697872, + 0.015696071088314056, + 0.015892280265688896, + 0.015507953241467476, + 0.016716869547963142, + 0.007956276647746563, + 0.012186230160295963, + 0.013396207243204117, + 0.013024529442191124, + 0.008640247397124767, + 0.005985909607261419, + 0.013117320835590363, + 0.009371895343065262, + 0.007756074890494347, + 0.007588375825434923, + 0.009557584300637245, + 0.006718946620821953, + 0.01015408430248499, + 0.007339296396821737, + 0.008715721778571606, + 0.008408303372561932, + 0.003585262456908822, + 0.008078839629888535, + 0.009190043434500694, + 0.004183468408882618, + 0.0044278111308813095, + 0.0039679924957454205, + 0.0078097134828567505, + 0.006120078731328249, + 0.003997867461293936, + 0.005620566662400961, + 0.012544170022010803, + 0.004825851880013943, + 0.004371044225990772, + 0.004787779878824949, + 0.00751988822594285, + 0.007190754637122154, + 0.00627126544713974, + 0.007492770440876484, + 0.014445208013057709, + 0.007001080550253391, + 0.00791297759860754, + 0.005923707503825426, + 0.007853255607187748, + 0.009769702330231667, + 0.015593374148011208, + 0.009643955156207085, + 0.01571427471935749, + 0.0076431832276284695, + 0.007625953294336796, + 0.006355091463774443, + 0.00855726283043623, + 0.009605231694877148, + 0.017100922763347626, + 0.013461635448038578, + 0.029260722920298576, + 0.017477624118328094, + 0.02601877599954605, + 0.017198698595166206, + 0.04005907475948334, + 0.029139084741473198, + 0.0015244975220412016, + 0.0036174990236759186, + 0.0021833484061062336, + 0.0018667903495952487, + 0.0017164509044960141, + 0.001309819403104484, + 0.002318423008546233, + 0.0019925087690353394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.008669240400195122 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08612726628780365, + 0.07668142020702362, + 0.059768471866846085, + 0.02038639411330223, + 0.08482857793569565, + 0.014082892797887325, + 0.01289679016917944, + 0.024233529344201088, + 0.6152070164680481, + 0.005787618458271027 + ], + "credit_map": [ + 0.013146720826625824, + 0.011089432053267956, + 0.0132679995149374, + 0.005071369465440512, + 0.011649450287222862, + 0.008654276840388775, + 0.011464221403002739, + 0.011783810332417488, + 0.01148422434926033, + 0.013570389710366726, + 0.01225756760686636, + 0.006706056650727987, + 0.006123100873082876, + 0.007387702353298664, + 0.009044650942087173, + 0.0101077351719141, + 0.0052680266089737415, + 0.007769717834889889, + 0.010605420917272568, + 0.010107080452144146, + 0.006586702540516853, + 0.004562441725283861, + 0.008856317028403282, + 0.006012774538248777, + 0.0016488286200910807, + 0.0017359601333737373, + 0.0031990266870707273, + 0.002933872863650322, + 0.002104777842760086, + 0.002132509369403124, + 0.003857817267999053, + 0.0027736020274460316, + 0.0061799599789083, + 0.00538997445255518, + 0.010343941859900951, + 0.00968879833817482, + 0.008186311461031437, + 0.012019524350762367, + 0.019489483907818794, + 0.013530593365430832, + 0.0012666151160374284, + 0.001993840793147683, + 0.0029143148567527533, + 0.0016010937979444861, + 0.0014162653824314475, + 0.0012540611205622554, + 0.001807546359486878, + 0.0018291573505848646, + 0.0013927293475717306, + 0.0018465378088876605, + 0.0028745844028890133, + 0.001101478235796094, + 0.0012497506104409695, + 0.0012525934725999832, + 0.001422006986103952, + 0.0017571112839505076, + 0.0023814933374524117, + 0.0024193020071834326, + 0.004809799138456583, + 0.0019390130182728171, + 0.00222397455945611, + 0.0028898483142256737, + 0.004046434536576271, + 0.0035236673429608345, + 0.05103636160492897, + 0.046650439500808716, + 0.09141770005226135, + 0.0574117936193943, + 0.0799468383193016, + 0.054410338401794434, + 0.13575010001659393, + 0.09858349710702896, + 0.0004926106776110828, + 0.0005878756637685001, + 0.001116301049478352, + 0.0007046416285447776, + 0.0006045568734407425, + 0.0005134632810950279, + 0.0009386977762915194, + 0.000829471624456346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.02058335766196251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36410433053970337, + 0.12849466502666473, + 0.07122895121574402, + 0.14731045067310333, + 0.10190222412347794, + 0.03374173492193222, + 0.03781299293041229, + 0.03541231527924538, + 0.06322203576564789, + 0.016770176589488983 + ], + "credit_map": [ + 0.0612017959356308, + 0.04693756252527237, + 0.06000988557934761, + 0.0341583676636219, + 0.04884447902441025, + 0.030438408255577087, + 0.03552848473191261, + 0.04698538780212402, + 0.020021332427859306, + 0.01999770477414131, + 0.022176874801516533, + 0.011522240936756134, + 0.008472458459436893, + 0.011822324246168137, + 0.015844566747546196, + 0.018637193366885185, + 0.008040683344006538, + 0.012106070294976234, + 0.011183642782270908, + 0.010795315727591515, + 0.006918879691511393, + 0.006021718494594097, + 0.009042500518262386, + 0.007120151538401842, + 0.02253350429236889, + 0.01935824193060398, + 0.021265512332320213, + 0.01702457293868065, + 0.02222762629389763, + 0.014020556584000587, + 0.01573091372847557, + 0.015149525366723537, + 0.008026079274713993, + 0.012091167271137238, + 0.019040554761886597, + 0.009265242144465446, + 0.016692524775862694, + 0.009378738701343536, + 0.01593756303191185, + 0.011470350436866283, + 0.0036735772155225277, + 0.003527092980220914, + 0.005986790172755718, + 0.003449377603828907, + 0.003982526250183582, + 0.00409682234749198, + 0.004757310263812542, + 0.004268243908882141, + 0.0030855564400553703, + 0.005962886847555637, + 0.006379632279276848, + 0.004077184479683638, + 0.004491087049245834, + 0.0038982450496405363, + 0.00541317742317915, + 0.004505227319896221, + 0.005206326954066753, + 0.005350948311388493, + 0.006908463779836893, + 0.0026658419519662857, + 0.0033794662449508905, + 0.0028095412999391556, + 0.004320028703659773, + 0.004771699197590351, + 0.008845110423862934, + 0.008695484139025211, + 0.009847375564277172, + 0.00459530483931303, + 0.009384760633111, + 0.006254187785089016, + 0.00767179112881422, + 0.007928033359348774, + 0.0019720618147403, + 0.0025743795558810234, + 0.0023529245518147945, + 0.0015503823524340987, + 0.0021185672376304865, + 0.0017912184121087193, + 0.002096370095387101, + 0.002314274897798896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.011076509952545166 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot frames a quiet, tense scene on the surface of a dark, polished wooden table. A heavy, rectangular block of black ink, with a faint matte, dried-ink residue, looms like a monolith. From the deep shadow cast by this object, a tiny wild mouse cautiously peers out. Its fur is a soft, earthy brown, its beady black eyes wide and flickering with fear, while its delicate whiskers droop. The mouse's head tentatively emerges into the light, its tiny pink nose twitching almost imperceptibly as it sniffs the air. The entire scene is bisected by a powerful beam of light, creating a dramatic black-and-white contrast that transforms the ink into a majestic silhouette while brightly illuminating the mouse's inquisitive face and the rich, dark textures of the table. The atmosphere is quiet and suspenseful, completely focused on the creature's cautious exploration and vulnerability.", + "seed": 16421, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_2/index_7778_seed_16421_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0473899319767952, + "vlm_logprob": -0.0473899319767952, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.0473899319767952, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05273055657744408 + ], + "opposite_logprob_tokens": [ + -3.0527305603027344 + ], + "logprob_reward": -0.05273055657744408, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05273055657744408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10879715532064438, + 0.020540785044431686, + 0.016202939674258232, + 0.21282048523426056, + 0.11024092882871628, + 0.04517657682299614, + 0.09011027216911316, + 0.16011276841163635, + 0.13448020815849304, + 0.10151787847280502 + ], + "credit_map": [ + 0.009151754900813103, + 0.01550743356347084, + 0.008015386760234833, + 0.019445208832621574, + 0.012276927009224892, + 0.01784563809633255, + 0.009398658759891987, + 0.017156146466732025, + 0.0016733643133193254, + 0.003118823980912566, + 0.0016060384223237634, + 0.003541294252499938, + 0.0023660387378185987, + 0.0031775280367583036, + 0.0018869917839765549, + 0.0031707065645605326, + 0.002562741283327341, + 0.0013086505932733417, + 0.0017460539238527417, + 0.0013775744009763002, + 0.00183212966658175, + 0.002557566622272134, + 0.0018537279684096575, + 0.002964495215564966, + 0.02013767510652542, + 0.02446436882019043, + 0.017848672345280647, + 0.021579844877123833, + 0.023364678025245667, + 0.042868293821811676, + 0.015365386381745338, + 0.047191571444272995, + 0.010614961385726929, + 0.014402694068849087, + 0.010956289246678352, + 0.010818964801728725, + 0.0105460649356246, + 0.01979397051036358, + 0.011039366014301777, + 0.02206861600279808, + 0.004212474450469017, + 0.007346751634031534, + 0.004006410017609596, + 0.00463353656232357, + 0.007817496545612812, + 0.006502190604805946, + 0.004807692486792803, + 0.005850024987012148, + 0.006983655039221048, + 0.010223117657005787, + 0.004967294167727232, + 0.005804115440696478, + 0.015571667812764645, + 0.018839862197637558, + 0.010856087319552898, + 0.01686447486281395, + 0.024112973362207413, + 0.014988125301897526, + 0.015346839092671871, + 0.016806546598672867, + 0.027877923101186752, + 0.028444096446037292, + 0.020198572427034378, + 0.012337693944573402, + 0.011494450271129608, + 0.01289860624819994, + 0.010369213297963142, + 0.01583273336291313, + 0.018008070066571236, + 0.02391372062265873, + 0.020250288769602776, + 0.021713128313422203, + 0.006998767610639334, + 0.007646831683814526, + 0.00645355274900794, + 0.0071512009017169476, + 0.024025041610002518, + 0.018550526350736618, + 0.012882579118013382, + 0.017809385433793068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.05273055657744408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908101171255112 + ], + "opposite_logprob_tokens": [ + -2.464081048965454 + ], + "logprob_reward": -0.08908101171255112, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908101171255112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07924376428127289, + 0.045564670115709305, + 0.02364811673760414, + 0.16500519216060638, + 0.07046133279800415, + 0.06284149736166, + 0.07199162244796753, + 0.18194617331027985, + 0.26346784830093384, + 0.03582979366183281 + ], + "credit_map": [ + 0.005466432776302099, + 0.013735387474298477, + 0.004998550750315189, + 0.014770547859370708, + 0.010154972784221172, + 0.011879491619765759, + 0.0068282256834208965, + 0.01141015812754631, + 0.003471796866506338, + 0.0072233122773468494, + 0.0028525148518383503, + 0.00871884822845459, + 0.005907280836254358, + 0.006703562568873167, + 0.004448724444955587, + 0.006238630041480064, + 0.0022482455242425203, + 0.0030542691238224506, + 0.0019387948559597135, + 0.0032478387001901865, + 0.003229610389098525, + 0.003660672577098012, + 0.002917293692007661, + 0.003351391525939107, + 0.015555360354483128, + 0.014336236752569675, + 0.014795441180467606, + 0.013834555633366108, + 0.0203402116894722, + 0.03488421067595482, + 0.011096134781837463, + 0.0401630625128746, + 0.0068451156839728355, + 0.008689522743225098, + 0.006399770267307758, + 0.00706575158983469, + 0.007697483990341425, + 0.011651989072561264, + 0.007998855784535408, + 0.014112857170403004, + 0.00481249438598752, + 0.010541105642914772, + 0.00372343510389328, + 0.005150259472429752, + 0.012841006740927696, + 0.009722028858959675, + 0.00780280027538538, + 0.008248364552855492, + 0.008940048515796661, + 0.010556144639849663, + 0.0062192403711378574, + 0.007315509952604771, + 0.010334537364542484, + 0.010393091477453709, + 0.009892339818179607, + 0.008340714499354362, + 0.024546802043914795, + 0.01712372899055481, + 0.014675004407763481, + 0.01997317001223564, + 0.025973688811063766, + 0.03278067708015442, + 0.025602953508496284, + 0.02127014473080635, + 0.02124716341495514, + 0.029903624206781387, + 0.029206877574324608, + 0.028537370264530182, + 0.0447024405002594, + 0.03728584200143814, + 0.03659982234239578, + 0.035984694957733154, + 0.0026232486125081778, + 0.002831509802490473, + 0.0027589891105890274, + 0.003149946453049779, + 0.007265039719641209, + 0.0059458306059241295, + 0.0056719304993748665, + 0.0055832997895777225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.08908101171255112 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009813398122787476 + ], + "opposite_logprob_tokens": [ + -4.63481330871582 + ], + "logprob_reward": -0.009813398122787476, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009813398122787476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05226396024227142, + 0.03355589136481285, + 0.018073908984661102, + 0.14235998690128326, + 0.18809117376804352, + 0.18805694580078125, + 0.0675848200917244, + 0.05579325184226036, + 0.1710646152496338, + 0.08315543830394745 + ], + "credit_map": [ + 0.005276028532534838, + 0.004948913585394621, + 0.006829769816249609, + 0.006474831607192755, + 0.006367390044033527, + 0.00799165666103363, + 0.005848429165780544, + 0.008526941761374474, + 0.002728208666667342, + 0.004995761439204216, + 0.002564110094681382, + 0.0062487986870110035, + 0.004408352077007294, + 0.004966886714100838, + 0.003171450924128294, + 0.00447232648730278, + 0.0018666730029508471, + 0.0017389944987371564, + 0.0019323260057717562, + 0.0020738905295729637, + 0.002386930398643017, + 0.002838796703144908, + 0.00253415503539145, + 0.0027021451387554407, + 0.01340573187917471, + 0.010767495259642601, + 0.01151792798191309, + 0.01656840741634369, + 0.015130720101296902, + 0.029450271278619766, + 0.010811896063387394, + 0.03470754250884056, + 0.016778966411948204, + 0.02035103738307953, + 0.024099178612232208, + 0.018088294193148613, + 0.014761295169591904, + 0.037325479090213776, + 0.015657754614949226, + 0.041029177606105804, + 0.017321042716503143, + 0.040227726101875305, + 0.01399275939911604, + 0.018845506012439728, + 0.031511299312114716, + 0.02540053240954876, + 0.020722972229123116, + 0.020035116001963615, + 0.004780797753483057, + 0.008825976401567459, + 0.004247234668582678, + 0.004458459094166756, + 0.011904453858733177, + 0.013801521621644497, + 0.00724489102140069, + 0.012321490794420242, + 0.005220410879701376, + 0.004299452994018793, + 0.004679776728153229, + 0.004215535242110491, + 0.011242998763918877, + 0.009551583789288998, + 0.00984535738825798, + 0.006738137919455767, + 0.016469212248921394, + 0.01761489547789097, + 0.018819602206349373, + 0.015533404424786568, + 0.03150671720504761, + 0.02704734168946743, + 0.021860022097826004, + 0.02221343293786049, + 0.0054274932481348515, + 0.008342184126377106, + 0.00619765929877758, + 0.007339944131672382, + 0.015695268288254738, + 0.015949660912156105, + 0.008714021183550358, + 0.01548920851200819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.009813398122787476 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328772470355034 + ], + "opposite_logprob_tokens": [ + -3.773287773132324 + ], + "logprob_reward": -0.02328772470355034, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328772470355034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046091627329587936, + 0.01226353831589222, + 0.0183322224766016, + 0.16720925271511078, + 0.1020275205373764, + 0.06721528619527817, + 0.10305581986904144, + 0.24347318708896637, + 0.15224304795265198, + 0.08808846771717072 + ], + "credit_map": [ + 0.004331119824200869, + 0.0037380133289843798, + 0.0034653055481612682, + 0.00597959803417325, + 0.004343154840171337, + 0.009896516799926758, + 0.004329281393438578, + 0.01000863965600729, + 0.0015164250507950783, + 0.0012468061177060008, + 0.0018510674126446247, + 0.0012553384294733405, + 0.0010554760228842497, + 0.001983222085982561, + 0.0012008770136162639, + 0.0021543269976973534, + 0.001530895009636879, + 0.0014677137369289994, + 0.0025683026760816574, + 0.001271866844035685, + 0.0017176871187984943, + 0.003658677451312542, + 0.0016618187073618174, + 0.004455261398106813, + 0.013708212412893772, + 0.022363847121596336, + 0.014967312105000019, + 0.01533526461571455, + 0.01871749944984913, + 0.0323963388800621, + 0.012837931513786316, + 0.036882854998111725, + 0.008993241004645824, + 0.01226190198212862, + 0.01049495767802, + 0.012723059393465519, + 0.010592898353934288, + 0.01670915074646473, + 0.012726662680506706, + 0.017525654286146164, + 0.006616412196308374, + 0.006840587127953768, + 0.009926656261086464, + 0.005537144839763641, + 0.010097325779497623, + 0.009179205633699894, + 0.009676688350737095, + 0.009341271594166756, + 0.00972677767276764, + 0.009105059318244457, + 0.008977998048067093, + 0.010357336141169071, + 0.015842141583561897, + 0.01769072934985161, + 0.015865320339798927, + 0.015490452758967876, + 0.035591572523117065, + 0.022503528743982315, + 0.021219441667199135, + 0.026505883783102036, + 0.03945361450314522, + 0.04527165740728378, + 0.03233400732278824, + 0.02059345692396164, + 0.010015378706157207, + 0.017413388937711716, + 0.011911021545529366, + 0.01452967245131731, + 0.025359604507684708, + 0.026796704158186913, + 0.021234285086393356, + 0.0249829962849617, + 0.0060731940902769566, + 0.006615299731492996, + 0.006369099952280521, + 0.00620981166139245, + 0.023810958489775658, + 0.016135070472955704, + 0.009449292905628681, + 0.01342573668807745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.02328772470355034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny wild mouse on a dark, polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, rectangular block of black ink with a faint matte, dried-ink residue visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mouse's fur soft and earthy brown with beady black eyes and drooping whiskers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bisected by a powerful beam of light creating a high-contrast black-and-white effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle low and static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062036968767642975 + ], + "opposite_logprob_tokens": [ + -2.8120369911193848 + ], + "logprob_reward": -0.062036968767642975, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062036968767642975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04551752284169197, + 0.02861187979578972, + 0.02674931101500988, + 0.11433643847703934, + 0.058074433356523514, + 0.1786167174577713, + 0.12626950442790985, + 0.10259661078453064, + 0.1299348622560501, + 0.18929271399974823 + ], + "credit_map": [ + 0.004344610497355461, + 0.004579866770654917, + 0.00550040090456605, + 0.00519306818023324, + 0.004929745569825172, + 0.007313546724617481, + 0.006723689381033182, + 0.0069326008670032024, + 0.0018702580127865076, + 0.003123209113255143, + 0.002908082911744714, + 0.004417486023157835, + 0.0032110114116221666, + 0.004576726816594601, + 0.0038515652995556593, + 0.004653544165194035, + 0.0027574694249778986, + 0.003146484261378646, + 0.0020095098298043013, + 0.004087927285581827, + 0.004187337122857571, + 0.003791181603446603, + 0.0031278233509510756, + 0.0036415783688426018, + 0.010831460356712341, + 0.011255540885031223, + 0.010406063869595528, + 0.008993538096547127, + 0.016628483310341835, + 0.022865455597639084, + 0.007857704535126686, + 0.025498202070593834, + 0.006157995667308569, + 0.006004288326948881, + 0.006841466296464205, + 0.008603204041719437, + 0.006347425747662783, + 0.008266400545835495, + 0.00783601589500904, + 0.0080176442861557, + 0.015375089831650257, + 0.030383992940187454, + 0.015486947260797024, + 0.024750901386141777, + 0.02758242003619671, + 0.024981439113616943, + 0.021154366433620453, + 0.018901558592915535, + 0.011369570158421993, + 0.013845912180840969, + 0.006232353392988443, + 0.00961932074278593, + 0.023719042539596558, + 0.02519194222986698, + 0.014246552251279354, + 0.022044816985726357, + 0.009935461916029453, + 0.011932148598134518, + 0.0074384636245667934, + 0.012440960854291916, + 0.01714673824608326, + 0.0177916269749403, + 0.01666576787829399, + 0.009245449677109718, + 0.01137528382241726, + 0.012840913608670235, + 0.012058500200510025, + 0.0138764139264822, + 0.01936488039791584, + 0.023644667118787766, + 0.017271747812628746, + 0.01950247213244438, + 0.015592273324728012, + 0.019660478457808495, + 0.02128014527261257, + 0.024487361311912537, + 0.03605668991804123, + 0.03166085109114647, + 0.019162975251674652, + 0.02139194868505001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.062036968767642975 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_tensor([1064, 1064])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_tensor([1064, 1064])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e6651074984c9e7d542b62be3c67fbcbfcde912 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_tensor([1064, 1064])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, tilting slowly downward from a horizontal angle, focuses on a woman clad in an elegant, deep lake blue silk cheongsam. Her raven hair is pulled back in a delicate bun, her face delicate and graceful, her expression serene yet expectant. She sits at a dark, polished rosewood table. A soft afternoon light streams in through a carved wooden window, framing the soft contours of her features and illuminating the delicate orchid pattern embroidered on her cheongsam. On the table, an antique rosewood box rests quietly, its surface patinated by the passage of time. Carefully, she reaches out with her slender, white hands, gently tracing the box. She finds and unfastens a slightly tarnished brass clasp with a soft, crisp click. Then, with steady and deliberate movements, she slowly lifts the lid upwards, a gesture imbued with a ritual, as if unveiling a forgotten memory. Under the warm light and shadow, the whole scene presents a classical atmosphere that is quiet, focused and full of stories.", + "seed": 6598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_1064_seed_6598_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03391406629234552, + "vlm_logprob": -0.03391406629234552, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03391406629234552, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01654628850519657 + ], + "opposite_logprob_tokens": [ + -4.641546249389648 + ], + "logprob_reward": -0.01654628850519657, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01654628850519657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18357039988040924, + 0.050297897309064865, + 0.11147883534431458, + 0.07401340454816818, + 0.05760256573557854, + 0.05547260120511055, + 0.04948422685265541, + 0.20164962112903595, + 0.12222017347812653, + 0.09421029686927795 + ], + "credit_map": [ + 0.027510881423950195, + 0.019154516980051994, + 0.031659550964832306, + 0.02078450471162796, + 0.025009406730532646, + 0.019138460978865623, + 0.01712876930832863, + 0.023184318095445633, + 0.004404820967465639, + 0.007133862469345331, + 0.006223657168447971, + 0.005979233421385288, + 0.009762031957507133, + 0.003197643905878067, + 0.007821070030331612, + 0.005775575526058674, + 0.012008198536932468, + 0.011299577541649342, + 0.012161904945969582, + 0.02164633758366108, + 0.011338419280946255, + 0.007250074762851, + 0.02219771221280098, + 0.013576611876487732, + 0.007009176537394524, + 0.009999359957873821, + 0.006877278909087181, + 0.008794008754193783, + 0.020425422117114067, + 0.006026623770594597, + 0.007265214808285236, + 0.007616324350237846, + 0.006598392501473427, + 0.006912051234394312, + 0.010683180764317513, + 0.007469859905540943, + 0.006661653518676758, + 0.008652491495013237, + 0.004742070566862822, + 0.0058828662149608135, + 0.006118027959018946, + 0.009000686928629875, + 0.007543077226728201, + 0.0073097748681902885, + 0.004753345623612404, + 0.008106771856546402, + 0.005525656044483185, + 0.007115264423191547, + 0.005542484577745199, + 0.007549233268946409, + 0.00739790964871645, + 0.0049741086550056934, + 0.006639409810304642, + 0.005619535688310862, + 0.005791184492409229, + 0.0059703621082007885, + 0.01791953854262829, + 0.044295135885477066, + 0.022807443514466286, + 0.033037468791007996, + 0.027964040637016296, + 0.01917455345392227, + 0.022963160648941994, + 0.01348827499896288, + 0.01062496192753315, + 0.01736471988260746, + 0.011950983665883541, + 0.013412500731647015, + 0.012045667506754398, + 0.01320138480514288, + 0.026351645588874817, + 0.017268307507038116, + 0.012758320197463036, + 0.008285745978355408, + 0.014813536778092384, + 0.00978480651974678, + 0.010319523513317108, + 0.014076301828026772, + 0.01378176175057888, + 0.01039030123502016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.01654628850519657 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011105275712907314 + ], + "opposite_logprob_tokens": [ + -4.511105060577393 + ], + "logprob_reward": -0.011105275712907314, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011105275712907314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2028816193342209, + 0.15954925119876862, + 0.09451832622289658, + 0.10917623341083527, + 0.046328045427799225, + 0.02938660979270935, + 0.043303634971380234, + 0.08921076357364655, + 0.1570405811071396, + 0.06860493868589401 + ], + "credit_map": [ + 0.019681716337800026, + 0.023085419088602066, + 0.04203701391816139, + 0.02377541922032833, + 0.024649204686284065, + 0.015493029728531837, + 0.026890812441706657, + 0.02726900577545166, + 0.011657635681331158, + 0.023925090208649635, + 0.034849103540182114, + 0.01823895424604416, + 0.023434357717633247, + 0.011362149380147457, + 0.01370663195848465, + 0.022375326603651047, + 0.009861989878118038, + 0.007285188417881727, + 0.010509397834539413, + 0.015652064234018326, + 0.012051626108586788, + 0.007077000103890896, + 0.020536791533231735, + 0.011544267646968365, + 0.008236967958509922, + 0.0089311134070158, + 0.008850342594087124, + 0.01717483624815941, + 0.024375909939408302, + 0.009436804801225662, + 0.011797119863331318, + 0.020373141393065453, + 0.005424935836344957, + 0.006389225367456675, + 0.00772452587261796, + 0.0041940282098948956, + 0.0074715507216751575, + 0.006436278112232685, + 0.003927980083972216, + 0.004759519826620817, + 0.0034459903836250305, + 0.004110663663595915, + 0.0035575793590396643, + 0.0033332163002341986, + 0.004335991106927395, + 0.004224889911711216, + 0.003255194518715143, + 0.0031230843160301447, + 0.005501491017639637, + 0.006232683081179857, + 0.0077178520150482655, + 0.004054947290569544, + 0.004661403596401215, + 0.004846240859478712, + 0.006112391594797373, + 0.004176625516265631, + 0.009646188467741013, + 0.012654588557779789, + 0.013800380751490593, + 0.012069372460246086, + 0.00885104387998581, + 0.0084982393309474, + 0.012517021037638187, + 0.011173929087817669, + 0.012580332346260548, + 0.023225244134664536, + 0.010299213230609894, + 0.012569563463330269, + 0.01639677584171295, + 0.01745973713696003, + 0.03682207688689232, + 0.027687635272741318, + 0.006861790083348751, + 0.007011807523667812, + 0.009210444986820221, + 0.006823220755904913, + 0.006570558529347181, + 0.008138937875628471, + 0.015396180562675, + 0.008591996505856514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.011105275712907314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978577509522438 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.00978577509522438, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978577509522438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11895644664764404, + 0.06959397345781326, + 0.05256114900112152, + 0.04457440227270126, + 0.09985974431037903, + 0.05183800682425499, + 0.04442467540502548, + 0.15462613105773926, + 0.18326404690742493, + 0.18030144274234772 + ], + "credit_map": [ + 0.010702903382480145, + 0.008492744527757168, + 0.03626967594027519, + 0.013316627591848373, + 0.011185337789356709, + 0.012814694084227085, + 0.009462602436542511, + 0.016711857169866562, + 0.008826752193272114, + 0.005752735771238804, + 0.012955513782799244, + 0.008756201714277267, + 0.007973816245794296, + 0.007276509888470173, + 0.008295097388327122, + 0.009757344610989094, + 0.003818446770310402, + 0.004927692003548145, + 0.007384616881608963, + 0.010929528623819351, + 0.00555532006546855, + 0.003201168030500412, + 0.009610474109649658, + 0.007133902050554752, + 0.005287726875394583, + 0.005708124488592148, + 0.005671873223036528, + 0.006967389024794102, + 0.006167958024889231, + 0.004160613752901554, + 0.005048647057265043, + 0.005562065169215202, + 0.010015136562287807, + 0.009683304466307163, + 0.01975022442638874, + 0.01457533985376358, + 0.0119539275765419, + 0.011916883289813995, + 0.008873176760971546, + 0.013091749511659145, + 0.005540297366678715, + 0.008663443848490715, + 0.006732372101396322, + 0.00668663065880537, + 0.004999796859920025, + 0.00705649284645915, + 0.006212306674569845, + 0.005946666933596134, + 0.00506330793723464, + 0.009163131937384605, + 0.0056823138147592545, + 0.006017121020704508, + 0.003512218128889799, + 0.004402125254273415, + 0.005698382388800383, + 0.004886075388640165, + 0.013890646398067474, + 0.029861511662602425, + 0.01968730241060257, + 0.023240668699145317, + 0.014976981095969677, + 0.01431060116738081, + 0.022052405402064323, + 0.016606008633971214, + 0.017344649881124496, + 0.02029644511640072, + 0.013619561679661274, + 0.015134896151721478, + 0.017123403027653694, + 0.021309271454811096, + 0.03850654140114784, + 0.039929281920194626, + 0.012480334378778934, + 0.009486752562224865, + 0.015476642176508904, + 0.014820681884884834, + 0.014446042478084564, + 0.028561044484376907, + 0.05089380219578743, + 0.03413613513112068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.00978577509522438 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908112347126007 + ], + "opposite_logprob_tokens": [ + -2.464081048965454 + ], + "logprob_reward": -0.08908112347126007, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908112347126007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053023528307676315, + 0.0402839295566082, + 0.20819376409053802, + 0.057426635175943375, + 0.055045176297426224, + 0.06878578662872314, + 0.05377425253391266, + 0.21126067638397217, + 0.172226220369339, + 0.07998006045818329 + ], + "credit_map": [ + 0.005173785146325827, + 0.0060355523601174355, + 0.01102348044514656, + 0.00627332367002964, + 0.008028813637793064, + 0.005533027462661266, + 0.00486806919798255, + 0.006087472662329674, + 0.004091306589543819, + 0.00448804022744298, + 0.006932316347956657, + 0.00549920042976737, + 0.00415613641962409, + 0.0028256443329155445, + 0.006653326563537121, + 0.005637955851852894, + 0.014765780419111252, + 0.0198165625333786, + 0.02283387817442417, + 0.04790147766470909, + 0.018041783943772316, + 0.011415984481573105, + 0.04413805529475212, + 0.02928023971617222, + 0.0047945547848939896, + 0.007300140801817179, + 0.0068021551705896854, + 0.010681423358619213, + 0.010137436911463737, + 0.0057084630243480206, + 0.005697146523743868, + 0.006305315066128969, + 0.005417677573859692, + 0.004780169110745192, + 0.010656366124749184, + 0.008923615328967571, + 0.0071610575541853905, + 0.006305615417659283, + 0.005959611386060715, + 0.0058410633355379105, + 0.007546512875705957, + 0.01267896220088005, + 0.007858692668378353, + 0.007105089258402586, + 0.008234890177845955, + 0.009756085462868214, + 0.006928449962288141, + 0.008677099831402302, + 0.007286050822585821, + 0.010772431269288063, + 0.007339106872677803, + 0.005469353403896093, + 0.005638171918690205, + 0.006020544096827507, + 0.005817285738885403, + 0.00543130561709404, + 0.018390927463769913, + 0.040805213153362274, + 0.02690666913986206, + 0.034738689661026, + 0.025392595678567886, + 0.019700296223163605, + 0.026618747040629387, + 0.01870754174888134, + 0.015594515018165112, + 0.035443536937236786, + 0.012396184727549553, + 0.01828758418560028, + 0.021053923293948174, + 0.019099043682217598, + 0.031122146174311638, + 0.019229277968406677, + 0.010516512207686901, + 0.006869477219879627, + 0.011011823080480099, + 0.00661213556304574, + 0.007311167661100626, + 0.009248527698218822, + 0.015625612810254097, + 0.012784801423549652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.08908112347126007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305186867713928 + ], + "opposite_logprob_tokens": [ + -3.1680519580841064 + ], + "logprob_reward": -0.04305186867713928, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305186867713928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08330091834068298, + 0.07644128054380417, + 0.05005438253283501, + 0.09660361707210541, + 0.11073484271764755, + 0.04236409068107605, + 0.07011264562606812, + 0.38477766513824463, + 0.03710727393627167, + 0.04850330948829651 + ], + "credit_map": [ + 0.012977704405784607, + 0.0063704862259328365, + 0.019087176769971848, + 0.009592858143150806, + 0.009455799125134945, + 0.007664536125957966, + 0.008664831519126892, + 0.00948751624673605, + 0.0070015438832342625, + 0.009473939426243305, + 0.014068343676626682, + 0.01139311958104372, + 0.007912710309028625, + 0.007121020462363958, + 0.00885553378611803, + 0.010615062899887562, + 0.003625631332397461, + 0.0037204534746706486, + 0.006988064385950565, + 0.012350792065262794, + 0.0038517776411026716, + 0.0024807709269225597, + 0.00980082992464304, + 0.00723605789244175, + 0.011970840394496918, + 0.012587040662765503, + 0.013562656007707119, + 0.011571560986340046, + 0.014057629741728306, + 0.006959299091249704, + 0.009514386765658855, + 0.016380196437239647, + 0.012395046651363373, + 0.01239665225148201, + 0.017107302322983742, + 0.015382185578346252, + 0.013699078001081944, + 0.014950567856431007, + 0.01147192157804966, + 0.013332086615264416, + 0.004840575624257326, + 0.0064614019356667995, + 0.005218096077442169, + 0.0055536385625600815, + 0.0044095865450799465, + 0.005783207714557648, + 0.005036050453782082, + 0.005061531439423561, + 0.008338328450918198, + 0.01478539127856493, + 0.007053038105368614, + 0.009726123884320259, + 0.008244398981332779, + 0.006783643737435341, + 0.00926999468356371, + 0.0059117223136126995, + 0.03406685218214989, + 0.09414538741111755, + 0.047336988151073456, + 0.06777198612689972, + 0.04708638787269592, + 0.0352957583963871, + 0.0386977419257164, + 0.020376525819301605, + 0.0034225170966237783, + 0.004389502108097076, + 0.002983824349939823, + 0.004392985720187426, + 0.0024488354101777077, + 0.0033787398133426905, + 0.009731805883347988, + 0.006359058432281017, + 0.006716991774737835, + 0.0046576219610869884, + 0.006869750563055277, + 0.004909627139568329, + 0.005615359637886286, + 0.005607160739600658, + 0.008015326224267483, + 0.006111468654125929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.04305186867713928 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, tilting slowly downward from a horizontal angle, focuses on a woman clad in an elegant, deep lake blue silk cheongsam. Her raven hair is pulled back in a delicate bun, her face delicate and graceful, her expression serene yet expectant. She sits at a dark, polished rosewood table. A soft afternoon light streams in through a carved wooden window, framing the soft contours of her features and illuminating the delicate orchid pattern embroidered on her cheongsam. On the table, an antique rosewood box rests quietly, its surface patinated by the passage of time. Carefully, she reaches out with her slender, white hands, gently tracing the box. She finds and unfastens a slightly tarnished brass clasp with a soft, crisp click. Then, with steady and deliberate movements, she slowly lifts the lid upwards, a gesture imbued with a ritual, as if unveiling a forgotten memory. Under the warm light and shadow, the whole scene presents a classical atmosphere that is quiet, focused and full of stories.", + "seed": 6599, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_1064_seed_6599_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0364104013890028, + "vlm_logprob": -0.0364104013890028, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.0364104013890028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0413186214864254 + ], + "opposite_logprob_tokens": [ + -3.291318655014038 + ], + "logprob_reward": -0.0413186214864254, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0413186214864254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12445520609617233, + 0.2744280993938446, + 0.18585561215877533, + 0.05336998775601387, + 0.1830192655324936, + 0.0928540974855423, + 0.026864096522331238, + 0.025319155305624008, + 0.006822485942393541, + 0.02701202593743801 + ], + "credit_map": [ + 0.011602288112044334, + 0.021443843841552734, + 0.010837874375283718, + 0.018235856667160988, + 0.019132087007164955, + 0.017140867188572884, + 0.012094837613403797, + 0.01396754290908575, + 0.011813781224191189, + 0.031010359525680542, + 0.014308620244264603, + 0.009450075216591358, + 0.021098872646689415, + 0.08658774197101593, + 0.009352837689220905, + 0.090805783867836, + 0.01856132782995701, + 0.0324832983314991, + 0.01873701438307762, + 0.029790116474032402, + 0.016910601407289505, + 0.02425001561641693, + 0.022130034863948822, + 0.022993190214037895, + 0.00765306456014514, + 0.006616361439228058, + 0.004908304661512375, + 0.0065396856516599655, + 0.005622632335871458, + 0.008297735825181007, + 0.006281777285039425, + 0.007450420875102282, + 0.045983199030160904, + 0.01061919704079628, + 0.012099625542759895, + 0.007679325062781572, + 0.008117309771478176, + 0.08027968555688858, + 0.006324979476630688, + 0.011915944516658783, + 0.01475425623357296, + 0.012826075777411461, + 0.009938209317624569, + 0.008951740339398384, + 0.009500371292233467, + 0.01585560105741024, + 0.008007197640836239, + 0.013020641170442104, + 0.00330971647053957, + 0.003668438410386443, + 0.004389243666082621, + 0.002642708597704768, + 0.0036556327249854803, + 0.002955061150714755, + 0.003673776751384139, + 0.0025695154909044504, + 0.003265406470745802, + 0.0030793813057243824, + 0.004552192986011505, + 0.002563140820711851, + 0.0028523679357022047, + 0.0039544361643493176, + 0.0020937470253556967, + 0.0029584814328700304, + 0.0008109005284495652, + 0.0006350945914164186, + 0.0007880951161496341, + 0.0006734902854077518, + 0.0011981969000771642, + 0.0011123159201815724, + 0.0006299832020886242, + 0.0009744087583385408, + 0.002782829338684678, + 0.003188101574778557, + 0.0038821224588900805, + 0.002716355724260211, + 0.002555473241955042, + 0.0026869375724345446, + 0.005062231793999672, + 0.004137971438467503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.0413186214864254 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026409240439534187 + ], + "opposite_logprob_tokens": [ + -3.651409149169922 + ], + "logprob_reward": -0.026409240439534187, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026409240439534187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09405088424682617, + 0.05552442744374275, + 0.0651627704501152, + 0.04912522807717323, + 0.41017404198646545, + 0.1943792700767517, + 0.03774517774581909, + 0.03334443271160126, + 0.013631436042487621, + 0.04686235636472702 + ], + "credit_map": [ + 0.008615330792963505, + 0.015062913298606873, + 0.00758407823741436, + 0.011618520133197308, + 0.01598794013261795, + 0.017263369634747505, + 0.006846853066235781, + 0.011071872897446156, + 0.004531973972916603, + 0.00814252533018589, + 0.013643952086567879, + 0.005804078653454781, + 0.005259405821561813, + 0.006090269424021244, + 0.005290611647069454, + 0.006761608645319939, + 0.004366644658148289, + 0.008254729211330414, + 0.004106034524738789, + 0.008267958648502827, + 0.00291146757081151, + 0.006226959638297558, + 0.01618536002933979, + 0.01484361570328474, + 0.007799814455211163, + 0.00510459765791893, + 0.004920576233416796, + 0.006264021620154381, + 0.004267688374966383, + 0.007402903866022825, + 0.0053055863827466965, + 0.008060037158429623, + 0.10815808176994324, + 0.0227406807243824, + 0.018362419679760933, + 0.01219707541167736, + 0.018619854003190994, + 0.19167426228523254, + 0.010077142156660557, + 0.02834446169435978, + 0.04766806215047836, + 0.044650182127952576, + 0.014453296549618244, + 0.0089778583496809, + 0.0170847550034523, + 0.032563719898462296, + 0.009129729121923447, + 0.019851671531796455, + 0.005041105672717094, + 0.005164097063243389, + 0.007467005867511034, + 0.003852987661957741, + 0.004293287172913551, + 0.003687863703817129, + 0.004571394994854927, + 0.0036674346774816513, + 0.0048812273889780045, + 0.003538459539413452, + 0.004836915992200375, + 0.003110451390966773, + 0.004106793086975813, + 0.006858339533209801, + 0.002576957456767559, + 0.0034352883230894804, + 0.0013238780666142702, + 0.0015101130120456219, + 0.0031182346865534782, + 0.001239461125805974, + 0.0013488742988556623, + 0.0015682134544476867, + 0.0016783548053354025, + 0.0018443063599988818, + 0.004336011596024036, + 0.006819357629865408, + 0.007166931871324778, + 0.0033775835763663054, + 0.003198092570528388, + 0.005048444960266352, + 0.009871192276477814, + 0.007044744677841663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.026409240439534187 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108889013528824 + ], + "opposite_logprob_tokens": [ + -3.2881088256835938 + ], + "logprob_reward": -0.038108889013528824, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108889013528824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05564671754837036, + 0.023743238300085068, + 0.09738828241825104, + 0.03549879044294357, + 0.22543098032474518, + 0.43522167205810547, + 0.02768530882894993, + 0.026389213278889656, + 0.00564724812284112, + 0.06734856963157654 + ], + "credit_map": [ + 0.009149305522441864, + 0.008002779446542263, + 0.004124920349568129, + 0.004401919897645712, + 0.009174778126180172, + 0.00926528126001358, + 0.004061213228851557, + 0.007466516457498074, + 0.0015861004358157516, + 0.002710969652980566, + 0.0016557802446186543, + 0.000993791501969099, + 0.0016351445810869336, + 0.006708588916808367, + 0.0009188901749439538, + 0.0075339749455451965, + 0.006084946915507317, + 0.01045112032443285, + 0.003941184841096401, + 0.01173905935138464, + 0.004945006687194109, + 0.01299726590514183, + 0.024319982156157494, + 0.022909721359610558, + 0.005355771631002426, + 0.004056174773722887, + 0.0033936025574803352, + 0.0043971240520477295, + 0.0036443148273974657, + 0.005130533128976822, + 0.00439994502812624, + 0.005121326074004173, + 0.061094269156455994, + 0.01125751156359911, + 0.01081538014113903, + 0.006509686820209026, + 0.01065503153949976, + 0.10322252660989761, + 0.005639452952891588, + 0.016237134113907814, + 0.10271617770195007, + 0.09530620276927948, + 0.033190976828336716, + 0.024701254442334175, + 0.040392402559518814, + 0.07613547891378403, + 0.021851796656847, + 0.04092739149928093, + 0.003913175780326128, + 0.003154846141114831, + 0.004628503229469061, + 0.003653810592368245, + 0.0027324287220835686, + 0.00371945439837873, + 0.003081363858655095, + 0.002801725873723626, + 0.0035962641704827547, + 0.00288031087256968, + 0.003969048149883747, + 0.0019985793624073267, + 0.0036571980454027653, + 0.005701771005988121, + 0.0018504946492612362, + 0.002735547022894025, + 0.0009178338223136961, + 0.0005856028874404728, + 0.0009286265703849494, + 0.000651274633128196, + 0.0005503247957676649, + 0.0005700015462934971, + 0.000589202216360718, + 0.0008543817093595862, + 0.005601792596280575, + 0.00811203382909298, + 0.006993429735302925, + 0.005267872009426355, + 0.005084800999611616, + 0.008579670451581478, + 0.015703704208135605, + 0.01200526300817728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.038108889013528824 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038120366632938385 + ], + "opposite_logprob_tokens": [ + -3.2881202697753906 + ], + "logprob_reward": -0.038120366632938385, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038120366632938385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03171646222472191, + 0.19112172722816467, + 0.025889765471220016, + 0.04502105340361595, + 0.34412893652915955, + 0.18096297979354858, + 0.06358762085437775, + 0.02587013877928257, + 0.016199825331568718, + 0.07550151646137238 + ], + "credit_map": [ + 0.002451511099934578, + 0.004158798139542341, + 0.00281026610173285, + 0.0035320192109793425, + 0.005187025759369135, + 0.006733392830938101, + 0.003125353017821908, + 0.003718097461387515, + 0.007345412392169237, + 0.022129761055111885, + 0.007927208207547665, + 0.005606659688055515, + 0.014499135315418243, + 0.06227502226829529, + 0.005951295141130686, + 0.06538722664117813, + 0.0022776166442781687, + 0.003915008623152971, + 0.0028070714324712753, + 0.005827992223203182, + 0.0025155034381896257, + 0.002844761125743389, + 0.0031672248151153326, + 0.002534586936235428, + 0.006977920886129141, + 0.005488571710884571, + 0.004125163424760103, + 0.005970297381281853, + 0.004583842121064663, + 0.0055725183337926865, + 0.006142355035990477, + 0.006160382647067308, + 0.09145443886518478, + 0.018160730600357056, + 0.0157285388559103, + 0.013570462353527546, + 0.01641683839261532, + 0.15713028609752655, + 0.009909061715006828, + 0.021758578717708588, + 0.04223887249827385, + 0.03919331729412079, + 0.014853683300316334, + 0.01081682089716196, + 0.016078077256679535, + 0.030512508004903793, + 0.009348472580313683, + 0.01792122609913349, + 0.008848876692354679, + 0.0066234213300049305, + 0.009545705281198025, + 0.00919745396822691, + 0.0071740783751010895, + 0.007891391403973103, + 0.007955013774335384, + 0.006351681426167488, + 0.0027516603004187346, + 0.0024355235509574413, + 0.0037138438783586025, + 0.0033295440953224897, + 0.0031357924453914165, + 0.004327744245529175, + 0.0029712067916989326, + 0.0032048227731138468, + 0.0019126998959109187, + 0.0015832541976124048, + 0.0028083468787372112, + 0.0021683787927031517, + 0.0017615477554500103, + 0.0014845996629446745, + 0.0021520727314054966, + 0.00232892413623631, + 0.006341645494103432, + 0.012931150384247303, + 0.015467108227312565, + 0.0047570341266691685, + 0.005687609780579805, + 0.006720433011651039, + 0.013961996883153915, + 0.009634535759687424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.038120366632938385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038094889372587204 + ], + "opposite_logprob_tokens": [ + -3.288094997406006 + ], + "logprob_reward": -0.038094889372587204, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038094889372587204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02249540016055107, + 0.06034408137202263, + 0.02240452542901039, + 0.019749537110328674, + 0.3493838906288147, + 0.28785744309425354, + 0.14653411507606506, + 0.05398942530155182, + 0.008844657801091671, + 0.02839697152376175 + ], + "credit_map": [ + 0.0032420370262116194, + 0.0031051672995090485, + 0.0024111424572765827, + 0.0026688517536967993, + 0.0024456665851175785, + 0.0026828451082110405, + 0.0022774082608520985, + 0.003662283066660166, + 0.0027161631733179092, + 0.00670469319447875, + 0.002627406734973192, + 0.002187546109780669, + 0.004549108445644379, + 0.019213376566767693, + 0.002202275674790144, + 0.02014351077377796, + 0.0028474279679358006, + 0.004435143433511257, + 0.0020676348358392715, + 0.004892735742032528, + 0.001550710410811007, + 0.0019959646742790937, + 0.002496801782399416, + 0.0021181076299399137, + 0.003681506495922804, + 0.0018668073462322354, + 0.0027267627883702517, + 0.002933527110144496, + 0.0016112725716084242, + 0.0022546625696122646, + 0.0021085289772599936, + 0.002566467970609665, + 0.09604299813508987, + 0.018933238461613655, + 0.015235024504363537, + 0.009645646438002586, + 0.016551801934838295, + 0.1608634740114212, + 0.00939874816685915, + 0.022712960839271545, + 0.07235780358314514, + 0.0676925927400589, + 0.02112579345703125, + 0.012804527767002583, + 0.024598799645900726, + 0.047716233879327774, + 0.012994538992643356, + 0.028567129746079445, + 0.019681503996253014, + 0.016767233610153198, + 0.025764277204871178, + 0.020341962575912476, + 0.0162137933075428, + 0.015401987358927727, + 0.017982356250286102, + 0.01438099518418312, + 0.00794441718608141, + 0.0055403639562428, + 0.009026557207107544, + 0.0032415653113275766, + 0.0069559370167553425, + 0.010552874766290188, + 0.003991696517914534, + 0.006736012641340494, + 0.0014958889223635197, + 0.0008582397713325918, + 0.0010567358694970608, + 0.0013312107184901834, + 0.0009163447539322078, + 0.0011104060104116797, + 0.0008318981272168458, + 0.001243932987563312, + 0.0026914067566394806, + 0.007556040305644274, + 0.00846165046095848, + 0.0016614665510132909, + 0.0018690360011532903, + 0.0018548594089224935, + 0.0024904594756662846, + 0.0018120543099939823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.038094889372587204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, tilting slowly downward from a horizontal angle, focuses on a woman clad in an elegant, deep lake blue silk cheongsam. Her raven hair is pulled back in a delicate bun, her face delicate and graceful, her expression serene yet expectant. She sits at a dark, polished rosewood table. A soft afternoon light streams in through a carved wooden window, framing the soft contours of her features and illuminating the delicate orchid pattern embroidered on her cheongsam. On the table, an antique rosewood box rests quietly, its surface patinated by the passage of time. Carefully, she reaches out with her slender, white hands, gently tracing the box. She finds and unfastens a slightly tarnished brass clasp with a soft, crisp click. Then, with steady and deliberate movements, she slowly lifts the lid upwards, a gesture imbued with a ritual, as if unveiling a forgotten memory. Under the warm light and shadow, the whole scene presents a classical atmosphere that is quiet, focused and full of stories.", + "seed": 6600, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_1064_seed_6600_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01780496258288622, + "vlm_logprob": -0.01780496258288622, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.01780496258288622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009572203271090984 + ], + "opposite_logprob_tokens": [ + -5.0095720291137695 + ], + "logprob_reward": -0.009572203271090984, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009572203271090984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2707681655883789, + 0.12076234072446823, + 0.07262515276670456, + 0.1401449590921402, + 0.10396728664636612, + 0.08708827942609787, + 0.060173604637384415, + 0.06102289631962776, + 0.04842015355825424, + 0.03502713888883591 + ], + "credit_map": [ + 0.03719642385840416, + 0.021929142996668816, + 0.0366668738424778, + 0.03732011467218399, + 0.04514515772461891, + 0.03291020542383194, + 0.03697390481829643, + 0.022626331076025963, + 0.010026817210018635, + 0.011504393070936203, + 0.013933063484728336, + 0.016120079904794693, + 0.017678966745734215, + 0.010967150330543518, + 0.023388024419546127, + 0.017143845558166504, + 0.006750152446329594, + 0.007850654423236847, + 0.00820216629654169, + 0.012634490616619587, + 0.007983000949025154, + 0.006803926546126604, + 0.011324934661388397, + 0.011075817979872227, + 0.010321502573788166, + 0.025644205510616302, + 0.012497247196733952, + 0.013769647106528282, + 0.031753554940223694, + 0.019190840423107147, + 0.016364265233278275, + 0.010603705421090126, + 0.008567987009882927, + 0.02367343008518219, + 0.010184106417000294, + 0.010202384553849697, + 0.01561112143099308, + 0.0174343753606081, + 0.010893560945987701, + 0.00740031385794282, + 0.008151041343808174, + 0.011063702404499054, + 0.015531509183347225, + 0.012042840011417866, + 0.012167909182608128, + 0.010292560793459415, + 0.010691197589039803, + 0.007147522177547216, + 0.006508321966975927, + 0.005887551233172417, + 0.01032338384538889, + 0.007410178426653147, + 0.0073976642452180386, + 0.006657273508608341, + 0.009046375751495361, + 0.006942854728549719, + 0.006261730100959539, + 0.00924994982779026, + 0.011473297141492367, + 0.007362231612205505, + 0.008081069216132164, + 0.007269870024174452, + 0.00552357267588377, + 0.005801173858344555, + 0.004235397092998028, + 0.005482977721840143, + 0.0077788918279111385, + 0.00807421375066042, + 0.0045386552810668945, + 0.005720893852412701, + 0.008050494827330112, + 0.004538629669696093, + 0.002782396972179413, + 0.003474842058494687, + 0.010000394657254219, + 0.007257773540914059, + 0.002625270513817668, + 0.00324110290966928, + 0.002959530334919691, + 0.0026858269702643156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009572203271090984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658958598971367 + ], + "opposite_logprob_tokens": [ + -4.7586588859558105 + ], + "logprob_reward": -0.008658958598971367, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658958598971367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3614911735057831, + 0.07963014394044876, + 0.051049839705228806, + 0.07479961961507797, + 0.10010311752557755, + 0.04671066999435425, + 0.15076269209384918, + 0.057090874761343, + 0.045451655983924866, + 0.03291017934679985 + ], + "credit_map": [ + 0.048555273562669754, + 0.03914414346218109, + 0.04981685057282448, + 0.04642310366034508, + 0.057161226868629456, + 0.046546127647161484, + 0.04978136718273163, + 0.024063073098659515, + 0.007070477120578289, + 0.01045852992683649, + 0.013737881556153297, + 0.008827025070786476, + 0.012198792770504951, + 0.008545346558094025, + 0.00963674671947956, + 0.009155345149338245, + 0.004206453915685415, + 0.005786251742392778, + 0.004649550653994083, + 0.010970448143780231, + 0.0058458405546844006, + 0.0039539639838039875, + 0.008557556197047234, + 0.007079774979501963, + 0.0044075981713831425, + 0.01368788443505764, + 0.007287527900189161, + 0.007740892004221678, + 0.017713744193315506, + 0.010614512488245964, + 0.00889952015131712, + 0.0044479407370090485, + 0.01110563613474369, + 0.018901823088526726, + 0.00958264246582985, + 0.010053317062556744, + 0.014039410278201103, + 0.014524859376251698, + 0.011339547112584114, + 0.010555884800851345, + 0.004980824887752533, + 0.0066952453926205635, + 0.007451250683516264, + 0.006524499971419573, + 0.006958105135709047, + 0.005197382997721434, + 0.0049927495419979095, + 0.003910612780600786, + 0.013429353944957256, + 0.01674540340900421, + 0.023339910432696342, + 0.016148552298545837, + 0.020796341821551323, + 0.024829663336277008, + 0.019808128476142883, + 0.015665341168642044, + 0.005444346461445093, + 0.007827251218259335, + 0.010680408217012882, + 0.006511288695037365, + 0.007580978330224752, + 0.006382443010807037, + 0.008114936761558056, + 0.004549222998321056, + 0.004268321208655834, + 0.004865352530032396, + 0.008108124136924744, + 0.0074933418072760105, + 0.005740978289395571, + 0.003448596689850092, + 0.007353284861892462, + 0.004173653665930033, + 0.0034916435834020376, + 0.0040368810296058655, + 0.00885005109012127, + 0.005396977998316288, + 0.003332301741465926, + 0.0032233616802841425, + 0.002269524149596691, + 0.0023094373755156994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.008658958598971367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075212620198727 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075212620198727, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075212620198727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13399678468704224, + 0.04386741667985916, + 0.13988636434078217, + 0.2035219520330429, + 0.05536027252674103, + 0.07359817624092102, + 0.14622311294078827, + 0.11981267482042313, + 0.026084288954734802, + 0.0576489120721817 + ], + "credit_map": [ + 0.01642504520714283, + 0.014042113907635212, + 0.01705847680568695, + 0.02106701210141182, + 0.022700704634189606, + 0.014443832449615002, + 0.019534824416041374, + 0.00872478075325489, + 0.004019132815301418, + 0.004617353901267052, + 0.00500395055860281, + 0.004927429836243391, + 0.006089456379413605, + 0.004688101354986429, + 0.007340730633586645, + 0.007181261666119099, + 0.0117040378972888, + 0.013761945068836212, + 0.013294786214828491, + 0.029252197593450546, + 0.013084453530609608, + 0.010426917113363743, + 0.024884149432182312, + 0.023477889597415924, + 0.01263868622481823, + 0.03975513577461243, + 0.016766663640737534, + 0.020185362547636032, + 0.047908999025821686, + 0.02893463708460331, + 0.024392595514655113, + 0.012939870357513428, + 0.004925269167870283, + 0.008938298560678959, + 0.008071002550423145, + 0.005665146745741367, + 0.00800473801791668, + 0.009008008055388927, + 0.00633719377219677, + 0.004410613793879747, + 0.005093362648040056, + 0.009542887099087238, + 0.009569654241204262, + 0.010295739397406578, + 0.009777877479791641, + 0.008759784512221813, + 0.01346528995782137, + 0.00709358137100935, + 0.01575755700469017, + 0.016022659838199615, + 0.022976068779826164, + 0.015850724652409554, + 0.02117760293185711, + 0.019073670729994774, + 0.019732272252440453, + 0.01563256047666073, + 0.01273948885500431, + 0.015450437553226948, + 0.02847651205956936, + 0.011796274222433567, + 0.019803324714303017, + 0.010934005491435528, + 0.011100033298134804, + 0.00951259583234787, + 0.0022725507151335478, + 0.003154775593429804, + 0.004038586746901274, + 0.004188532941043377, + 0.002916011493653059, + 0.0028124265372753143, + 0.0035982239060103893, + 0.00310318055562675, + 0.004086923785507679, + 0.004788417834788561, + 0.018513504415750504, + 0.014307002536952496, + 0.003998848609626293, + 0.004062403459101915, + 0.004386909306049347, + 0.003504903521388769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.011075212620198727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107910368591547 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.01107910368591547, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107910368591547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1450805515050888, + 0.0634009912610054, + 0.05161615461111069, + 0.22880706191062927, + 0.08704017847776413, + 0.10484996438026428, + 0.12903955578804016, + 0.1092703565955162, + 0.045471277087926865, + 0.03542385995388031 + ], + "credit_map": [ + 0.017369581386446953, + 0.021297115832567215, + 0.01756594143807888, + 0.022426635026931763, + 0.023111023008823395, + 0.01623256504535675, + 0.017860304564237595, + 0.009217384271323681, + 0.005547990091145039, + 0.005444835405796766, + 0.00927874818444252, + 0.008701631799340248, + 0.011363948695361614, + 0.0069532012566924095, + 0.010117319412529469, + 0.005993315484374762, + 0.005951594561338425, + 0.005158147308975458, + 0.005034406203776598, + 0.006344400346279144, + 0.0064504449255764484, + 0.006418077275156975, + 0.006837035994976759, + 0.009422048926353455, + 0.012012445367872715, + 0.04395262524485588, + 0.022470124065876007, + 0.022906137630343437, + 0.0559823252260685, + 0.0298069566488266, + 0.02714235708117485, + 0.014534084126353264, + 0.00791583675891161, + 0.01762913353741169, + 0.008921481668949127, + 0.009265696629881859, + 0.011875858530402184, + 0.016072401776909828, + 0.009211668744683266, + 0.006148099899291992, + 0.008838300593197346, + 0.014304162934422493, + 0.01427507121115923, + 0.015424098819494247, + 0.013135792687535286, + 0.013640408404171467, + 0.012239635922014713, + 0.012992491945624352, + 0.009318643249571323, + 0.015412524342536926, + 0.017426220700144768, + 0.014037610031664371, + 0.016865208745002747, + 0.025621837005019188, + 0.017721783369779587, + 0.012635729275643826, + 0.01076622772961855, + 0.01637439616024494, + 0.013280975632369518, + 0.015074123628437519, + 0.015220613218843937, + 0.016956890001893044, + 0.011802827939391136, + 0.009794311597943306, + 0.004696480929851532, + 0.003625402692705393, + 0.009316105395555496, + 0.005831800866872072, + 0.0056653753854334354, + 0.004356496501713991, + 0.008012213744223118, + 0.003967401571571827, + 0.0032635568641126156, + 0.0036706123501062393, + 0.00841718539595604, + 0.007536475080996752, + 0.0031706986483186483, + 0.0030369043815881014, + 0.0038777454756200314, + 0.00245068222284317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.01107910368591547 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863933473825455 + ], + "opposite_logprob_tokens": [ + -3.0486392974853516 + ], + "logprob_reward": -0.04863933473825455, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863933473825455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08321279287338257, + 0.03809260576963425, + 0.025277461856603622, + 0.4996851980686188, + 0.033451274037361145, + 0.038463447242975235, + 0.21436944603919983, + 0.02479376085102558, + 0.028550568968057632, + 0.01410339493304491 + ], + "credit_map": [ + 0.012514828704297543, + 0.004744690842926502, + 0.013033228926360607, + 0.00895426794886589, + 0.014373268000781536, + 0.010233039036393166, + 0.007746920455247164, + 0.011612555012106895, + 0.0054802424274384975, + 0.00478974636644125, + 0.004342437721788883, + 0.004536425694823265, + 0.005431118421256542, + 0.004677229560911655, + 0.006287955678999424, + 0.002547451527789235, + 0.003377477405592799, + 0.002331195631995797, + 0.0036050276830792427, + 0.0030494683887809515, + 0.0030777466017752886, + 0.001830561552196741, + 0.0035431385040283203, + 0.004462846554815769, + 0.030999688431620598, + 0.09406685829162598, + 0.04522910341620445, + 0.048498619347810745, + 0.11737295985221863, + 0.06597358733415604, + 0.058833006769418716, + 0.038711417466402054, + 0.00355503405444324, + 0.004693073220551014, + 0.005510083865374327, + 0.003503868356347084, + 0.004441730212420225, + 0.004737845156341791, + 0.003935746382921934, + 0.00307389535009861, + 0.00324417557567358, + 0.0057979486882686615, + 0.004636777099221945, + 0.003939975518733263, + 0.00538119999691844, + 0.00652632350102067, + 0.004313231911510229, + 0.004623813554644585, + 0.02163649909198284, + 0.023850146681070328, + 0.03253193572163582, + 0.020278465002775192, + 0.030214419588446617, + 0.03820420429110527, + 0.026168297976255417, + 0.021485472097992897, + 0.0027313989121466875, + 0.0035606962628662586, + 0.0033403760753571987, + 0.003190012648701668, + 0.003999410197138786, + 0.0026297306176275015, + 0.003213424002751708, + 0.00212871260009706, + 0.002004667418077588, + 0.0032478596549481153, + 0.0052607194520533085, + 0.0051386370323598385, + 0.0027487913612276316, + 0.003035715315490961, + 0.004510028287768364, + 0.002604148117825389, + 0.0011593499220907688, + 0.001185971312224865, + 0.005001180339604616, + 0.0030219072941690683, + 0.0012204392114654183, + 0.0009417427936568856, + 0.0009189943084493279, + 0.0006538103916682303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.04863933473825455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, tilting slowly downward from a horizontal angle, focuses on a woman clad in an elegant, deep lake blue silk cheongsam. Her raven hair is pulled back in a delicate bun, her face delicate and graceful, her expression serene yet expectant. She sits at a dark, polished rosewood table. A soft afternoon light streams in through a carved wooden window, framing the soft contours of her features and illuminating the delicate orchid pattern embroidered on her cheongsam. On the table, an antique rosewood box rests quietly, its surface patinated by the passage of time. Carefully, she reaches out with her slender, white hands, gently tracing the box. She finds and unfastens a slightly tarnished brass clasp with a soft, crisp click. Then, with steady and deliberate movements, she slowly lifts the lid upwards, a gesture imbued with a ritual, as if unveiling a forgotten memory. Under the warm light and shadow, the whole scene presents a classical atmosphere that is quiet, focused and full of stories.", + "seed": 6601, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_3/index_1064_seed_6601_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03460680730640888, + "vlm_logprob": -0.03460680730640888, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03460680730640888, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015381873585283756 + ], + "opposite_logprob_tokens": [ + -4.390381813049316 + ], + "logprob_reward": -0.015381873585283756, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015381873585283756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1861768364906311, + 0.07732522487640381, + 0.0819053053855896, + 0.0451764278113842, + 0.18492771685123444, + 0.08331210911273956, + 0.05978142470121384, + 0.0839482843875885, + 0.0974590852856636, + 0.09998759627342224 + ], + "credit_map": [ + 0.015101483091711998, + 0.029175635427236557, + 0.020953943952918053, + 0.04788993299007416, + 0.024621430784463882, + 0.009867235086858273, + 0.024958455935120583, + 0.013608709909021854, + 0.007312939036637545, + 0.01050951424986124, + 0.009179436601698399, + 0.010109979659318924, + 0.015560160391032696, + 0.00901866052299738, + 0.008737375028431416, + 0.0068971579894423485, + 0.010264324955642223, + 0.012100020423531532, + 0.00768861873075366, + 0.014820894226431847, + 0.006750010885298252, + 0.008762257173657417, + 0.01201645564287901, + 0.009502722881734371, + 0.003951898775994778, + 0.0052829766646027565, + 0.00580490892753005, + 0.0070612626150250435, + 0.008976300247013569, + 0.005473536904901266, + 0.004486325196921825, + 0.004139219410717487, + 0.03288370370864868, + 0.030173515900969505, + 0.02478509396314621, + 0.02031942456960678, + 0.017596431076526642, + 0.03230296075344086, + 0.013949266634881496, + 0.012917309999465942, + 0.008436530828475952, + 0.01615973934531212, + 0.01028947439044714, + 0.00816398486495018, + 0.01018243283033371, + 0.012966207228600979, + 0.008963873609900475, + 0.008149865083396435, + 0.007437161169946194, + 0.013306703418493271, + 0.007912738248705864, + 0.007531530223786831, + 0.004594521597027779, + 0.008828436955809593, + 0.0044553885236382484, + 0.00571494409814477, + 0.007725165691226721, + 0.008373505435883999, + 0.01106602419167757, + 0.009322882629930973, + 0.008393770083785057, + 0.009302538819611073, + 0.022376367822289467, + 0.007388037629425526, + 0.012507959268987179, + 0.01152776274830103, + 0.01717195473611355, + 0.012881798669695854, + 0.009388186037540436, + 0.012446616776287556, + 0.01089917030185461, + 0.010635643266141415, + 0.015250779688358307, + 0.01824253983795643, + 0.009795053862035275, + 0.011433180421590805, + 0.008132126182317734, + 0.012248305603861809, + 0.013634677045047283, + 0.011250939220190048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.015381873585283756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012572435662150383 + ], + "opposite_logprob_tokens": [ + -4.387572288513184 + ], + "logprob_reward": -0.012572435662150383, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012572435662150383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16683711111545563, + 0.14607186615467072, + 0.0387076772749424, + 0.049670711159706116, + 0.05148744210600853, + 0.05489065498113632, + 0.11541440337896347, + 0.13203418254852295, + 0.12692047655582428, + 0.11796547472476959 + ], + "credit_map": [ + 0.011181866750121117, + 0.034570932388305664, + 0.020123010501265526, + 0.02816273458302021, + 0.026466868817806244, + 0.011414357461035252, + 0.02110995166003704, + 0.013807385228574276, + 0.01294559147208929, + 0.012648694217205048, + 0.04213007166981697, + 0.02383456565439701, + 0.01743033155798912, + 0.01679358258843422, + 0.00837208516895771, + 0.011916927993297577, + 0.0055275363847613335, + 0.0077062505297362804, + 0.0035036311019212008, + 0.0052726129069924355, + 0.002819304820150137, + 0.0035690467339009047, + 0.003477622289210558, + 0.006831676699221134, + 0.005103972274810076, + 0.006259952671825886, + 0.005511161405593157, + 0.0081807691603899, + 0.012241356074810028, + 0.003688353579491377, + 0.0056535922922194, + 0.0030315560288727283, + 0.007142082788050175, + 0.00815457385033369, + 0.005073783919215202, + 0.006931827403604984, + 0.0053199054673314095, + 0.007573843002319336, + 0.004946110304445028, + 0.006345314905047417, + 0.005594249349087477, + 0.006441981997340918, + 0.00840107910335064, + 0.005723722279071808, + 0.005828059744089842, + 0.00861917994916439, + 0.007290551904588938, + 0.006991828791797161, + 0.012184510007500648, + 0.011570082977414131, + 0.016464998945593834, + 0.01564129814505577, + 0.011160160414874554, + 0.013455485925078392, + 0.023311160504817963, + 0.011626717634499073, + 0.010154925286769867, + 0.01771901547908783, + 0.015528815798461437, + 0.01942596025764942, + 0.013823946937918663, + 0.012545650824904442, + 0.030650142580270767, + 0.012185728177428246, + 0.015296845696866512, + 0.016963498666882515, + 0.021570146083831787, + 0.01569206826388836, + 0.014919600449502468, + 0.012338264845311642, + 0.01814064383506775, + 0.01199941523373127, + 0.018377289175987244, + 0.024463240057229996, + 0.013733548112213612, + 0.013840299099683762, + 0.010363978333771229, + 0.014864420518279076, + 0.011264155618846416, + 0.011058551259338856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012572435662150383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081933043897152 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081933043897152, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081933043897152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04487702250480652, + 0.13243718445301056, + 0.0345139242708683, + 0.06353309750556946, + 0.035809386521577835, + 0.09288447350263596, + 0.06722778081893921, + 0.08608744293451309, + 0.1449154019355774, + 0.2977142930030823 + ], + "credit_map": [ + 0.0038885597605258226, + 0.006056042388081551, + 0.005027568433433771, + 0.00991368480026722, + 0.0077011603862047195, + 0.003022895660251379, + 0.00469988165423274, + 0.004567230585962534, + 0.010533743537962437, + 0.004994433838874102, + 0.04177258536219597, + 0.02081366255879402, + 0.01290189754217863, + 0.028938181698322296, + 0.0034602524247020483, + 0.009022418409585953, + 0.00698615750297904, + 0.007074899040162563, + 0.0029167651664465666, + 0.004250004421919584, + 0.002002427587285638, + 0.0036838522646576166, + 0.002434763126075268, + 0.005165052600204945, + 0.004888146184384823, + 0.009229639545083046, + 0.007988677360117435, + 0.01190057024359703, + 0.010869981721043587, + 0.005387269891798496, + 0.006566707510501146, + 0.006702104117721319, + 0.004430672153830528, + 0.004114716313779354, + 0.004927785601466894, + 0.0057374113239347935, + 0.0027283551171422005, + 0.005229495000094175, + 0.003955398220568895, + 0.004685551393777132, + 0.008263356983661652, + 0.00974017009139061, + 0.018411707133054733, + 0.009526049718260765, + 0.008213099092245102, + 0.014389601536095142, + 0.00848222617059946, + 0.015858257189393044, + 0.008441515266895294, + 0.006979791913181543, + 0.010258079506456852, + 0.008424484170973301, + 0.005670094396919012, + 0.008714385330677032, + 0.01159613486379385, + 0.007143289782106876, + 0.006759162060916424, + 0.011204550042748451, + 0.012257309630513191, + 0.010730326175689697, + 0.00942421518266201, + 0.007621003780514002, + 0.017868585884571075, + 0.010222284123301506, + 0.01636124774813652, + 0.02082447148859501, + 0.022198380902409554, + 0.01783195696771145, + 0.01599355787038803, + 0.013146796263754368, + 0.025792747735977173, + 0.012766224332153797, + 0.044911351054906845, + 0.0454697459936142, + 0.03146228939294815, + 0.03762993961572647, + 0.028844071552157402, + 0.03309788927435875, + 0.037681788206100464, + 0.03861723840236664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.011081933043897152 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370720520615578 + ], + "opposite_logprob_tokens": [ + -3.4087071418762207 + ], + "logprob_reward": -0.03370720520615578, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370720520615578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056218162178993225, + 0.07285188138484955, + 0.09163768589496613, + 0.10226075351238251, + 0.10506732761859894, + 0.07294628769159317, + 0.13342227041721344, + 0.1254812330007553, + 0.1507900357246399, + 0.08932431787252426 + ], + "credit_map": [ + 0.004274375736713409, + 0.009390433318912983, + 0.007057682611048222, + 0.013352750800549984, + 0.006581099703907967, + 0.003009458538144827, + 0.006563992705196142, + 0.005988372955471277, + 0.006110306829214096, + 0.005923507269471884, + 0.018942106515169144, + 0.009702092967927456, + 0.00871825497597456, + 0.014946005307137966, + 0.003818364115431905, + 0.004691253881901503, + 0.024855056777596474, + 0.014707916416227818, + 0.006347113288938999, + 0.00937001220881939, + 0.004708938300609589, + 0.013504001311957836, + 0.005008619278669357, + 0.01313603576272726, + 0.00840269960463047, + 0.014623665250837803, + 0.012950184755027294, + 0.020300811156630516, + 0.016101622954010963, + 0.009540755301713943, + 0.010825199075043201, + 0.009515819139778614, + 0.01096690446138382, + 0.01063965167850256, + 0.017006661742925644, + 0.013594269752502441, + 0.008564645424485207, + 0.017555255442857742, + 0.015372458845376968, + 0.011367497965693474, + 0.00539134256541729, + 0.010937562212347984, + 0.010280044749379158, + 0.00883700605481863, + 0.005033810157328844, + 0.011017759330570698, + 0.0099367406219244, + 0.011512027122080326, + 0.01785588264465332, + 0.014678281731903553, + 0.021607235074043274, + 0.018733995035290718, + 0.0111535070464015, + 0.016370229423046112, + 0.0218143779784441, + 0.011208764277398586, + 0.008550397120416164, + 0.017078211531043053, + 0.015659397467970848, + 0.019342688843607903, + 0.016138503327965736, + 0.011553461663424969, + 0.02495640702545643, + 0.012202183715999126, + 0.018894510343670845, + 0.0230268444865942, + 0.024461379274725914, + 0.0169083084911108, + 0.018128933385014534, + 0.01460417453199625, + 0.02100408636033535, + 0.013761808164417744, + 0.010194585658609867, + 0.011592968367040157, + 0.009472621604800224, + 0.010013571940362453, + 0.008864435367286205, + 0.011765807867050171, + 0.01498111430555582, + 0.01243921834975481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.03370720520615578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a woman wearing a deep lake blue silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair styled in a delicate bun and is her expression serene yet expectant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include a dark polished rosewood table with an antique rosewood box on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman unfasten a tarnished brass clasp and lift the box lid upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slowly downward from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029058903455734 + ], + "opposite_logprob_tokens": [ + -2.350290536880493 + ], + "logprob_reward": -0.10029058903455734, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029058903455734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035481855273246765, + 0.16305415332317352, + 0.0683853030204773, + 0.05884091928601265, + 0.09184115380048752, + 0.0903535708785057, + 0.16707110404968262, + 0.1648361086845398, + 0.14594869315624237, + 0.014187158085405827 + ], + "credit_map": [ + 0.0023652210365980864, + 0.0057108779437839985, + 0.0047238473780453205, + 0.007935822941362858, + 0.00427651684731245, + 0.002470292616635561, + 0.005072492174804211, + 0.002926779678091407, + 0.013124417513608932, + 0.006578610744327307, + 0.05027081444859505, + 0.02492588385939598, + 0.015986349433660507, + 0.036465536803007126, + 0.004534921143203974, + 0.011167601682245731, + 0.012140724807977676, + 0.013422510586678982, + 0.008825786411762238, + 0.008244465105235577, + 0.003931296057999134, + 0.006053554825484753, + 0.005739081185311079, + 0.010027878917753696, + 0.006136336829513311, + 0.007758148014545441, + 0.007948165759444237, + 0.008298691362142563, + 0.007884898222982883, + 0.007583869155496359, + 0.0055147078819572926, + 0.0077160997316241264, + 0.013084784150123596, + 0.013889037072658539, + 0.009520321153104305, + 0.012578579597175121, + 0.008451631292700768, + 0.013103934936225414, + 0.010157796554267406, + 0.011055056937038898, + 0.009672663174569607, + 0.012319717556238174, + 0.013017972931265831, + 0.011987952515482903, + 0.006501166149973869, + 0.015250393189489841, + 0.009609771892428398, + 0.011993924155831337, + 0.018235862255096436, + 0.016381604596972466, + 0.0264801774173975, + 0.022640027105808258, + 0.015537871047854424, + 0.01966283656656742, + 0.032617878168821335, + 0.01551482081413269, + 0.013833396136760712, + 0.022758042439818382, + 0.020160552114248276, + 0.023565033450722694, + 0.018572254106402397, + 0.01595664583146572, + 0.03393176943063736, + 0.01605840027332306, + 0.017029674723744392, + 0.01936817727982998, + 0.02453281730413437, + 0.018017087131738663, + 0.016804559156298637, + 0.013670514337718487, + 0.022471606731414795, + 0.014054239727556705, + 0.0028467634692788124, + 0.0025632428005337715, + 0.001488946028985083, + 0.001580559997819364, + 0.001345394761301577, + 0.0014320174232125282, + 0.0016695702215656638, + 0.0012606619857251644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.10029058903455734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_4/index_tensor([5581, 5581])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_4/index_tensor([5581, 5581])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dbb7a534504ab65aa11f8c3a6df6f76a7d311807 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_4/index_tensor([5581, 5581])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up meticulously depicts the face of a young woman, her expression filled with pure, silent awe. Her long, lustrous black hair, elegantly pulled back and tied with a vibrant red satin ribbon, creates a warm and striking contrast to the cool tones of the scene. The sole source of illumination is the spectacular meteor shower, whose ethereal, shimmering light delicately dances across her features, casting a series of fleeting silver highlights across her smooth skin. Her large, dark eyes, sparkling with unshed tears of wonder, act like perfect mirrors of the celestial display above; within their liquid depths, each meteor is reflected as a bright, fleeting flash of light, flaring and fading in an instant. Her lips, slightly parted as if in a silent gasp, her brows gently raised, capture a profoundly intimate and magical moment of connection with the universe, all rendered with an elegant and refined style that emphasizes the scene's serene and wondrous atmosphere.", + "seed": 33175, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_4/index_5581_seed_33175_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.270272378064692, + "vlm_logprob": -0.270272378064692, + "vlm_yes_no_margin": 3.7250000119209288, + "vlm_token_credit": -0.270272378064692, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0044165924191474915 + ], + "opposite_logprob_tokens": [ + -6.129416465759277 + ], + "logprob_reward": -0.0044165924191474915, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0044165924191474915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18535253405570984, + 0.055229347199201584, + 0.1209869384765625, + 0.10285844653844833, + 0.06693284213542938, + 0.13698332011699677, + 0.12883087992668152, + 0.06229246407747269, + 0.08863679319620132, + 0.05189639702439308 + ], + "credit_map": [ + 0.019495252519845963, + 0.02036456950008869, + 0.024582291021943092, + 0.017548372969031334, + 0.018443405628204346, + 0.02632349729537964, + 0.03220881149172783, + 0.026386339217424393, + 0.005444574635475874, + 0.006082721520215273, + 0.004030838143080473, + 0.006248558405786753, + 0.009095199406147003, + 0.008377657271921635, + 0.0071559990756213665, + 0.008793801069259644, + 0.008512154221534729, + 0.013166652992367744, + 0.009857716038823128, + 0.010733427479863167, + 0.02619108557701111, + 0.023447023704648018, + 0.013298902660608292, + 0.015779972076416016, + 0.009080166928470135, + 0.02235100045800209, + 0.008263524621725082, + 0.01047773752361536, + 0.008885715156793594, + 0.015410900115966797, + 0.01808345317840576, + 0.010305950418114662, + 0.006551900412887335, + 0.0096477922052145, + 0.0059351492673158646, + 0.007226430810987949, + 0.00676629738882184, + 0.00864449329674244, + 0.011639147065579891, + 0.010521633550524712, + 0.02079015038907528, + 0.016313543543219566, + 0.01620832458138466, + 0.009390560910105705, + 0.009157623164355755, + 0.011867611669003963, + 0.03354422748088837, + 0.01971127837896347, + 0.02297889068722725, + 0.017988549545407295, + 0.020261302590370178, + 0.018011080101132393, + 0.011574367992579937, + 0.012926661409437656, + 0.012923942878842354, + 0.012166089378297329, + 0.00995795801281929, + 0.006917830556631088, + 0.015245486050844193, + 0.005253270268440247, + 0.004555401857942343, + 0.0043022544123232365, + 0.01068112812936306, + 0.005379138048738241, + 0.01435005385428667, + 0.017414337024092674, + 0.010791660286486149, + 0.019292311742901802, + 0.005756542552262545, + 0.0046017961576581, + 0.010429245419800282, + 0.006000852212309837, + 0.006007514428347349, + 0.005520862527191639, + 0.005928751081228256, + 0.005588008090853691, + 0.007161783054471016, + 0.006248843390494585, + 0.007240692153573036, + 0.008199946023523808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0044165924191474915 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011101149953901768 + ], + "opposite_logprob_tokens": [ + -4.511101245880127 + ], + "logprob_reward": -0.011101149953901768, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011101149953901768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.148490771651268, + 0.23005788028240204, + 0.1203925833106041, + 0.03832944482564926, + 0.07421506941318512, + 0.045304857194423676, + 0.037322331219911575, + 0.09808235615491867, + 0.040281280875205994, + 0.16752351820468903 + ], + "credit_map": [ + 0.019703855738043785, + 0.023121925070881844, + 0.019343487918376923, + 0.019507277756929398, + 0.014073261059820652, + 0.015003311447799206, + 0.01911493018269539, + 0.018622705712914467, + 0.04858098924160004, + 0.03365487605333328, + 0.03071623668074608, + 0.025727372616529465, + 0.017647957429289818, + 0.020567385479807854, + 0.023932112380862236, + 0.02923094481229782, + 0.026694655418395996, + 0.018954986706376076, + 0.02118726074695587, + 0.012958300299942493, + 0.011048443615436554, + 0.007911225780844688, + 0.011839289218187332, + 0.009798411279916763, + 0.0035518743097782135, + 0.004628054331988096, + 0.0036525516770780087, + 0.004652763716876507, + 0.003790432121604681, + 0.007161438465118408, + 0.006244652904570103, + 0.004647670779377222, + 0.010104960761964321, + 0.016222236678004265, + 0.009079289622604847, + 0.007012646645307541, + 0.004250735510140657, + 0.003560901852324605, + 0.016535719856619835, + 0.007448574993759394, + 0.007189325522631407, + 0.0058999573811888695, + 0.013027381151914597, + 0.003592120483517647, + 0.003198923310264945, + 0.0029861314687877893, + 0.004914497025310993, + 0.004496516659855843, + 0.006469923071563244, + 0.005367887206375599, + 0.00466547254472971, + 0.003749496303498745, + 0.003002382814884186, + 0.0031094509176909924, + 0.006139534525573254, + 0.00481817964464426, + 0.01198672130703926, + 0.014035573229193687, + 0.020155103877186775, + 0.017682848498225212, + 0.005723400507122278, + 0.005333659704774618, + 0.016584156081080437, + 0.006580879911780357, + 0.0050016725435853004, + 0.005396118853241205, + 0.00603258702903986, + 0.004354973789304495, + 0.0038599565159529448, + 0.004545118659734726, + 0.005361639894545078, + 0.005729209166020155, + 0.020100204274058342, + 0.018293816596269608, + 0.034276947379112244, + 0.017329568043351173, + 0.015756351873278618, + 0.0128584960475564, + 0.033786289393901825, + 0.015121822245419025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.011101149953901768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3136794567108154 + ], + "opposite_logprob_tokens": [ + -0.3136795163154602 + ], + "logprob_reward": -1.3136794567108154, + "yes_no_margin_reward": -0.9999999403953552, + "token_credit_reward": -1.3136794567108154, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21577981114387512, + 0.054810572415590286, + 0.06547395884990692, + 0.056739985942840576, + 0.07046554982662201, + 0.3134991228580475, + 0.04574345052242279, + 0.07207459211349487, + 0.06123962253332138, + 0.04417336732149124 + ], + "credit_map": [ + 0.02299422211945057, + 0.021859830245375633, + 0.0283712986856699, + 0.021621061488986015, + 0.023601418361067772, + 0.024126961827278137, + 0.04185149073600769, + 0.03135352581739426, + 0.011484278365969658, + 0.006920207291841507, + 0.006569691002368927, + 0.006393968593329191, + 0.005225400906056166, + 0.006088783498853445, + 0.005618680268526077, + 0.006509557832032442, + 0.007211309392005205, + 0.011107704602181911, + 0.011916179209947586, + 0.0072586266323924065, + 0.007917104288935661, + 0.006793285254389048, + 0.006267624907195568, + 0.007002122234553099, + 0.0051776994951069355, + 0.007178202271461487, + 0.004760220181196928, + 0.007165648974478245, + 0.00840787310153246, + 0.01188258733600378, + 0.005315794609487057, + 0.00685195904225111, + 0.011427727527916431, + 0.02090739831328392, + 0.006231792736798525, + 0.008545957505702972, + 0.004453643690794706, + 0.0033787512220442295, + 0.010811933316290379, + 0.0047083464451134205, + 0.06583239138126373, + 0.03803035616874695, + 0.031885601580142975, + 0.022119874134659767, + 0.018725048750638962, + 0.01827886700630188, + 0.09146089106798172, + 0.027166057378053665, + 0.00574162183329463, + 0.005554012954235077, + 0.004770995583385229, + 0.004993772134184837, + 0.005812510382384062, + 0.004692096263170242, + 0.006818058900535107, + 0.007360380608588457, + 0.007066934835165739, + 0.007826854474842548, + 0.019834011793136597, + 0.007330691441893578, + 0.005361814051866531, + 0.00426845857873559, + 0.014079596847295761, + 0.006306225433945656, + 0.004960383288562298, + 0.013826465234160423, + 0.006076537072658539, + 0.019518185406923294, + 0.004305933136492968, + 0.0030095865949988365, + 0.005052399821579456, + 0.004490124527364969, + 0.006154353730380535, + 0.004459263291209936, + 0.005281843710690737, + 0.0034091549459844828, + 0.0052412874065339565, + 0.005847361404448748, + 0.008078503422439098, + 0.0057015977799892426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -1.3136794567108154 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089595966041088 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089595966041088, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089595966041088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1614798605442047, + 0.05224371701478958, + 0.06614109873771667, + 0.05423136427998543, + 0.0813368707895279, + 0.2892765700817108, + 0.057303640991449356, + 0.12643344700336456, + 0.04122452437877655, + 0.07032893598079681 + ], + "credit_map": [ + 0.019858865067362785, + 0.016002371907234192, + 0.024711957201361656, + 0.015109137631952763, + 0.01600532978773117, + 0.021665941923856735, + 0.0269161406904459, + 0.021210111677646637, + 0.010324148461222649, + 0.005693687126040459, + 0.006646702066063881, + 0.00550806475803256, + 0.005808837711811066, + 0.006290316116064787, + 0.00509361969307065, + 0.006878336425870657, + 0.00752608198672533, + 0.009380357339978218, + 0.010919014923274517, + 0.00625868234783411, + 0.00998606439679861, + 0.00798861775547266, + 0.006493655499070883, + 0.007588623091578484, + 0.005975448060780764, + 0.013709727674722672, + 0.005244768224656582, + 0.006364624015986919, + 0.005433908198028803, + 0.005314012058079243, + 0.0069196708500385284, + 0.005269200075417757, + 0.020649367943406105, + 0.013421954587101936, + 0.010121245868504047, + 0.006444641854614019, + 0.006259756628423929, + 0.0058105639182031155, + 0.010188923217356205, + 0.008440416306257248, + 0.061242520809173584, + 0.03326164558529854, + 0.03177747130393982, + 0.01854855567216873, + 0.01777992956340313, + 0.01777888648211956, + 0.08158442378044128, + 0.027303116396069527, + 0.005823051556944847, + 0.007106558885425329, + 0.00714934803545475, + 0.005317311268299818, + 0.0070481193251907825, + 0.007408128120005131, + 0.009372892789542675, + 0.008078228682279587, + 0.012532005086541176, + 0.0160722266882658, + 0.03672746196389198, + 0.015234206803143024, + 0.00892034824937582, + 0.006233980879187584, + 0.020697813481092453, + 0.010015387088060379, + 0.0036920872516930103, + 0.008883722126483917, + 0.004986695479601622, + 0.012553923763334751, + 0.002965303836390376, + 0.0022198527585715055, + 0.003447561291977763, + 0.0024753743782639503, + 0.009265274740755558, + 0.006379329599440098, + 0.012555082328617573, + 0.005447235889732838, + 0.007318622432649136, + 0.007315654307603836, + 0.012852529995143414, + 0.009195199236273766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.011089595966041088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075095273554325 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075095273554325, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075095273554325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14050854742527008, + 0.08482878655195236, + 0.07239490747451782, + 0.10119574517011642, + 0.042550377547740936, + 0.20987045764923096, + 0.04048589617013931, + 0.09248865395784378, + 0.08687285333871841, + 0.1288038194179535 + ], + "credit_map": [ + 0.01375300157815218, + 0.015667058527469635, + 0.01514064148068428, + 0.01387608889490366, + 0.01261149998754263, + 0.019400088116526604, + 0.03179017826914787, + 0.018269995227456093, + 0.011114303022623062, + 0.008130816742777824, + 0.009504931047558784, + 0.010583152994513512, + 0.009456325322389603, + 0.01403353363275528, + 0.009031584486365318, + 0.012974141165614128, + 0.007314037531614304, + 0.0075235869735479355, + 0.009565239772200584, + 0.007323477417230606, + 0.013081327080726624, + 0.012102662585675716, + 0.007274903357028961, + 0.008209672756493092, + 0.01489782240241766, + 0.02430438995361328, + 0.006870381999760866, + 0.009619663469493389, + 0.01077039260417223, + 0.011581609025597572, + 0.009965360164642334, + 0.01318612415343523, + 0.0036257319152355194, + 0.009691823273897171, + 0.004764864686876535, + 0.004387736786156893, + 0.004340990912169218, + 0.0037671546451747417, + 0.006843023933470249, + 0.005129050929099321, + 0.0344671830534935, + 0.02011498250067234, + 0.0162434633821249, + 0.013325873762369156, + 0.022790951654314995, + 0.024799857288599014, + 0.035876039415597916, + 0.04225209727883339, + 0.003930386155843735, + 0.004106271546334028, + 0.0037851606030017138, + 0.004542458802461624, + 0.005585187114775181, + 0.0058347745798528194, + 0.006127705331891775, + 0.006573953665792942, + 0.009732324630022049, + 0.010573158971965313, + 0.025160223245620728, + 0.008692667819559574, + 0.005717107094824314, + 0.006899525411427021, + 0.014882453717291355, + 0.010831189341843128, + 0.006464940495789051, + 0.013690090738236904, + 0.0059041595086455345, + 0.017731623724102974, + 0.010525589808821678, + 0.011060998775064945, + 0.00959527026861906, + 0.011900177225470543, + 0.015233097597956657, + 0.01169519405812025, + 0.02521045133471489, + 0.00978195108473301, + 0.014702987857162952, + 0.013299738988280296, + 0.023329395800828934, + 0.015551000833511353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.011075095273554325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up meticulously depicts the face of a young woman, her expression filled with pure, silent awe. Her long, lustrous black hair, elegantly pulled back and tied with a vibrant red satin ribbon, creates a warm and striking contrast to the cool tones of the scene. The sole source of illumination is the spectacular meteor shower, whose ethereal, shimmering light delicately dances across her features, casting a series of fleeting silver highlights across her smooth skin. Her large, dark eyes, sparkling with unshed tears of wonder, act like perfect mirrors of the celestial display above; within their liquid depths, each meteor is reflected as a bright, fleeting flash of light, flaring and fading in an instant. Her lips, slightly parted as if in a silent gasp, her brows gently raised, capture a profoundly intimate and magical moment of connection with the universe, all rendered with an elegant and refined style that emphasizes the scene's serene and wondrous atmosphere.", + "seed": 33176, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_4/index_5581_seed_33176_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.28756466126069424, + "vlm_logprob": -0.28756466126069424, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.28756466126069424, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004928108304738998 + ], + "opposite_logprob_tokens": [ + -5.754928112030029 + ], + "logprob_reward": -0.004928108304738998, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004928108304738998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17394039034843445, + 0.16744282841682434, + 0.08911509066820145, + 0.05226365104317665, + 0.13526952266693115, + 0.03786978870630264, + 0.07234842330217361, + 0.059818290174007416, + 0.11167746782302856, + 0.10025449842214584 + ], + "credit_map": [ + 0.02334371581673622, + 0.019129157066345215, + 0.023884406313300133, + 0.03008110634982586, + 0.014255148358643055, + 0.01450420543551445, + 0.02190673165023327, + 0.02683592401444912, + 0.03649956360459328, + 0.01438917126506567, + 0.010484511032700539, + 0.01486597117036581, + 0.01166848000138998, + 0.013138928450644016, + 0.015482883900403976, + 0.05091333016753197, + 0.012895516119897366, + 0.008928587660193443, + 0.005330357700586319, + 0.007302196696400642, + 0.012332084588706493, + 0.011787028051912785, + 0.014882330782711506, + 0.015656985342502594, + 0.003805091604590416, + 0.005138379521667957, + 0.006114313378930092, + 0.0035500978119671345, + 0.008159616030752659, + 0.00806018989533186, + 0.008058033883571625, + 0.009377928450703621, + 0.01750931888818741, + 0.024589819833636284, + 0.01991974376142025, + 0.03246775269508362, + 0.009241166524589062, + 0.009098384529352188, + 0.011873375624418259, + 0.010569967329502106, + 0.006718277931213379, + 0.0058645992539823055, + 0.00425010034814477, + 0.006774993147701025, + 0.003054369008168578, + 0.004214698448777199, + 0.0031502291094511747, + 0.0038425209932029247, + 0.015026326291263103, + 0.01688852161169052, + 0.005632072687149048, + 0.007769698277115822, + 0.007604978512972593, + 0.006294257938861847, + 0.0059176255017519, + 0.00721493735909462, + 0.010663267225027084, + 0.006317412946373224, + 0.006515336222946644, + 0.005000757984817028, + 0.008190913125872612, + 0.005925411824136972, + 0.008638562634587288, + 0.008566629141569138, + 0.024452656507492065, + 0.01653401367366314, + 0.01417130883783102, + 0.008016603998839855, + 0.011956719681620598, + 0.008835331536829472, + 0.01333628874272108, + 0.01437454205006361, + 0.01997332088649273, + 0.016355473548173904, + 0.009112294763326645, + 0.008783332072198391, + 0.01402081735432148, + 0.00896440539509058, + 0.01164233312010765, + 0.011402525007724762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.004928108304738998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004652627278119326 + ], + "opposite_logprob_tokens": [ + -5.379652500152588 + ], + "logprob_reward": -0.004652627278119326, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004652627278119326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17168276011943817, + 0.11497831344604492, + 0.05863209068775177, + 0.0489623099565506, + 0.09765279293060303, + 0.15848517417907715, + 0.1407955139875412, + 0.1348637342453003, + 0.03846099600195885, + 0.03548639640212059 + ], + "credit_map": [ + 0.017396865412592888, + 0.03066171333193779, + 0.02734280377626419, + 0.03497890383005142, + 0.013431125320494175, + 0.011704700998961926, + 0.017717424780130386, + 0.01844920590519905, + 0.015120773576200008, + 0.014444067142903805, + 0.01972719468176365, + 0.018127011135220528, + 0.007023098412901163, + 0.008812392130494118, + 0.009700040332973003, + 0.022023729979991913, + 0.008196918293833733, + 0.012175384908914566, + 0.006661832798272371, + 0.008430802263319492, + 0.005525895394384861, + 0.006287918891757727, + 0.0053217001259326935, + 0.006031633820384741, + 0.00644485279917717, + 0.007585062179714441, + 0.007802089210599661, + 0.009124753065407276, + 0.004396399483084679, + 0.003835058771073818, + 0.005193253047764301, + 0.0045808362774550915, + 0.013062899932265282, + 0.015676511451601982, + 0.014759053476154804, + 0.022876178845763206, + 0.00897303782403469, + 0.006032946053892374, + 0.00900775846093893, + 0.007264390122145414, + 0.03188380226492882, + 0.03732950612902641, + 0.027086356654763222, + 0.02029365673661232, + 0.011783571913838387, + 0.010937806218862534, + 0.009457563050091267, + 0.009712898172438145, + 0.027865281328558922, + 0.03157459944486618, + 0.016533495858311653, + 0.013515755534172058, + 0.014494593255221844, + 0.010400158353149891, + 0.012389415875077248, + 0.014022203162312508, + 0.02583741582930088, + 0.023747114464640617, + 0.016080882400274277, + 0.010709528811275959, + 0.01458693016320467, + 0.010704481042921543, + 0.019749267026782036, + 0.013448113575577736, + 0.0073971995152533054, + 0.004156069830060005, + 0.005213577765971422, + 0.003684501862153411, + 0.005143847316503525, + 0.0030555224511772394, + 0.004668165929615498, + 0.005142110399901867, + 0.00614058505743742, + 0.006256180349737406, + 0.0036721837241202593, + 0.003312155371531844, + 0.004665686283260584, + 0.003432768862694502, + 0.004101431462913752, + 0.00390540505759418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.004652627278119326 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4064040184020996 + ], + "opposite_logprob_tokens": [ + -0.281404048204422 + ], + "logprob_reward": -1.4064040184020996, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4064040184020996, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1663137525320053, + 0.11574720591306686, + 0.09297893941402435, + 0.09220600128173828, + 0.1971483677625656, + 0.06432980298995972, + 0.11873719096183777, + 0.058303527534008026, + 0.0672118291258812, + 0.027023382484912872 + ], + "credit_map": [ + 0.014704701490700245, + 0.0222074706107378, + 0.017407147213816643, + 0.015057583339512348, + 0.021817127242684364, + 0.019186463207006454, + 0.029552659019827843, + 0.026380596682429314, + 0.01894463784992695, + 0.01950841397047043, + 0.02316485159099102, + 0.0157338734716177, + 0.0075999717228114605, + 0.006224345415830612, + 0.011031695641577244, + 0.01353942509740591, + 0.0154730174690485, + 0.016145044937729836, + 0.01785922423005104, + 0.009123644791543484, + 0.008782034739851952, + 0.008708958514034748, + 0.008375824429094791, + 0.008511191233992577, + 0.009286544285714626, + 0.018033361062407494, + 0.014284894801676273, + 0.013202579692006111, + 0.008958052843809128, + 0.01011534035205841, + 0.009030306711792946, + 0.009294918738305569, + 0.02645452506840229, + 0.03030332364141941, + 0.03174953907728195, + 0.05224952474236488, + 0.01358792744576931, + 0.011555477045476437, + 0.016625940799713135, + 0.014622095972299576, + 0.014000331051647663, + 0.011692054569721222, + 0.011180710047483444, + 0.009763834066689014, + 0.004064613953232765, + 0.004329183604568243, + 0.004361066967248917, + 0.004938011523336172, + 0.019631503149867058, + 0.029155269265174866, + 0.019362574443221092, + 0.008780657313764095, + 0.012671115808188915, + 0.01016222219914198, + 0.008275190368294716, + 0.010698650032281876, + 0.010953196324408054, + 0.00755988759920001, + 0.007188207469880581, + 0.004066732712090015, + 0.0073594823479652405, + 0.005970547907054424, + 0.008823574520647526, + 0.006381900515407324, + 0.012465229257941246, + 0.008501123636960983, + 0.009989344514906406, + 0.005175399128347635, + 0.008950881659984589, + 0.006098547950387001, + 0.008799819275736809, + 0.007231487892568111, + 0.004929894581437111, + 0.0038853974547237158, + 0.003649412654340267, + 0.002286650240421295, + 0.00399166252464056, + 0.0023799529299139977, + 0.0029538755770772696, + 0.0029465367551892996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -1.4064040184020996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641489617526531 + ], + "opposite_logprob_tokens": [ + -4.882641315460205 + ], + "logprob_reward": -0.007641489617526531, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641489617526531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14463381469249725, + 0.11339203268289566, + 0.06918361783027649, + 0.11248154193162918, + 0.22075672447681427, + 0.04790528863668442, + 0.07374953478574753, + 0.08322715014219284, + 0.06955373287200928, + 0.06511660665273666 + ], + "credit_map": [ + 0.01538239698857069, + 0.022514961659908295, + 0.0268380269408226, + 0.015281575731933117, + 0.014748256653547287, + 0.011225340887904167, + 0.02025904320180416, + 0.018384207040071487, + 0.01569547690451145, + 0.016949567943811417, + 0.015561267733573914, + 0.01264235656708479, + 0.007047673221677542, + 0.007410962600260973, + 0.012635448016226292, + 0.025449275970458984, + 0.010068864561617374, + 0.008700709789991379, + 0.007416626904159784, + 0.00431783078238368, + 0.009738133288919926, + 0.010295244865119457, + 0.009177991189062595, + 0.009468214586377144, + 0.01057680044323206, + 0.02561153657734394, + 0.017626408487558365, + 0.014994734898209572, + 0.01004653051495552, + 0.01161598414182663, + 0.00954469759017229, + 0.012464848347008228, + 0.03215916082262993, + 0.030056549236178398, + 0.034887075424194336, + 0.0556970089673996, + 0.016855688765645027, + 0.012802811339497566, + 0.020147711038589478, + 0.01815071515738964, + 0.00918863620609045, + 0.006388994865119457, + 0.009439803659915924, + 0.008169046603143215, + 0.0038872656878083944, + 0.003933195490390062, + 0.003177496837452054, + 0.0037208511494100094, + 0.010812309570610523, + 0.008134821429848671, + 0.014425416477024555, + 0.005784591659903526, + 0.008191153407096863, + 0.00754398200660944, + 0.0071661751717329025, + 0.0116910794749856, + 0.015707068145275116, + 0.013114296831190586, + 0.010387344285845757, + 0.0068278685212135315, + 0.008366219699382782, + 0.007293650880455971, + 0.012139527127146721, + 0.009391170926392078, + 0.01418796181678772, + 0.012068809941411018, + 0.010662265121936798, + 0.007036386523395777, + 0.006253224331885576, + 0.004923383705317974, + 0.007004574406892061, + 0.007417119573801756, + 0.012024170719087124, + 0.011485878378152847, + 0.00652492418885231, + 0.006508472841233015, + 0.008011076599359512, + 0.004894258454442024, + 0.007739587686955929, + 0.007928240112960339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.007641489617526531 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197062700986862 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014197062700986862, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197062700986862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20929476618766785, + 0.06546467542648315, + 0.0437101349234581, + 0.08698810636997223, + 0.13690312206745148, + 0.06238330900669098, + 0.10079494118690491, + 0.04881776496767998, + 0.07114855200052261, + 0.1744946390390396 + ], + "credit_map": [ + 0.01981540210545063, + 0.024678809568285942, + 0.02116853930056095, + 0.02597145549952984, + 0.02801882103085518, + 0.023502573370933533, + 0.0380159355700016, + 0.028123224154114723, + 0.016056807711720467, + 0.006330045405775309, + 0.013435494154691696, + 0.0059731584042310715, + 0.004322294145822525, + 0.004621212370693684, + 0.006439663004130125, + 0.008285998366773129, + 0.006040644366294146, + 0.006723447702825069, + 0.004790680482983589, + 0.0034769990015774965, + 0.00429878756403923, + 0.006674049887806177, + 0.004812507890164852, + 0.006893019657582045, + 0.00751888332888484, + 0.007067906204611063, + 0.01006233785301447, + 0.007571143563836813, + 0.012921953573822975, + 0.011865219101309776, + 0.015040125697851181, + 0.014940542168915272, + 0.013957390561699867, + 0.015316462144255638, + 0.015130213461816311, + 0.013252150267362595, + 0.023274585604667664, + 0.013923880644142628, + 0.02172049507498741, + 0.020327942445874214, + 0.011754358187317848, + 0.011617947369813919, + 0.011387249454855919, + 0.008957401849329472, + 0.004403010476380587, + 0.005079077556729317, + 0.004373559728264809, + 0.004810703452676535, + 0.01397297065705061, + 0.009372268803417683, + 0.016025040298700333, + 0.009084854274988174, + 0.01215424295514822, + 0.011155636981129646, + 0.0119663430377841, + 0.017063584178686142, + 0.00874093547463417, + 0.0050177015364170074, + 0.0067736865021288395, + 0.0032469909638166428, + 0.006022769957780838, + 0.005086041986942291, + 0.006691993214190006, + 0.007237649988383055, + 0.014341788366436958, + 0.00989406369626522, + 0.010725339874625206, + 0.0062575251795351505, + 0.007631306536495686, + 0.00605262815952301, + 0.008907808922231197, + 0.007338087074458599, + 0.022491704672574997, + 0.03165621683001518, + 0.015324816107749939, + 0.016644874587655067, + 0.026705831289291382, + 0.01386478915810585, + 0.024875478819012642, + 0.02293093316257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.014197062700986862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up meticulously depicts the face of a young woman, her expression filled with pure, silent awe. Her long, lustrous black hair, elegantly pulled back and tied with a vibrant red satin ribbon, creates a warm and striking contrast to the cool tones of the scene. The sole source of illumination is the spectacular meteor shower, whose ethereal, shimmering light delicately dances across her features, casting a series of fleeting silver highlights across her smooth skin. Her large, dark eyes, sparkling with unshed tears of wonder, act like perfect mirrors of the celestial display above; within their liquid depths, each meteor is reflected as a bright, fleeting flash of light, flaring and fading in an instant. Her lips, slightly parted as if in a silent gasp, her brows gently raised, capture a profoundly intimate and magical moment of connection with the universe, all rendered with an elegant and refined style that emphasizes the scene's serene and wondrous atmosphere.", + "seed": 33177, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_4/index_5581_seed_33177_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2192304695956409, + "vlm_logprob": -0.2192304695956409, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.2192304695956409, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00646949652582407 + ], + "opposite_logprob_tokens": [ + -5.6314697265625 + ], + "logprob_reward": -0.00646949652582407, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00646949652582407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09353938698768616, + 0.0949936956167221, + 0.03195970878005028, + 0.07152826339006424, + 0.08452770859003067, + 0.4131345748901367, + 0.023120444267988205, + 0.05160113424062729, + 0.05385440215468407, + 0.08174058794975281 + ], + "credit_map": [ + 0.011561159044504166, + 0.011552512645721436, + 0.021278027445077896, + 0.011307605542242527, + 0.009492216631770134, + 0.010867896489799023, + 0.007039216812700033, + 0.01044076681137085, + 0.012172458693385124, + 0.0062670474871993065, + 0.019388923421502113, + 0.011938942596316338, + 0.009982909075915813, + 0.010322474874556065, + 0.009935213252902031, + 0.014985735528171062, + 0.00738091254606843, + 0.003032274544239044, + 0.004981805104762316, + 0.0028190934099256992, + 0.0030254230368882418, + 0.00523363659158349, + 0.002394578419625759, + 0.003091990714892745, + 0.008982377126812935, + 0.005464030895382166, + 0.007727726362645626, + 0.008370824158191681, + 0.00734947994351387, + 0.011169375851750374, + 0.00989450141787529, + 0.012569955550134182, + 0.010103977285325527, + 0.007805532310158014, + 0.01464436948299408, + 0.011046063154935837, + 0.012495755217969418, + 0.00957602821290493, + 0.009087453596293926, + 0.009768535383045673, + 0.05894361808896065, + 0.026788387447595596, + 0.08977054804563522, + 0.0340869203209877, + 0.03341824188828468, + 0.035991404205560684, + 0.08502518385648727, + 0.04911031201481819, + 0.002308967523276806, + 0.0018770013703033328, + 0.0032398842740803957, + 0.0022694526705890894, + 0.0027537900023162365, + 0.003433534409850836, + 0.003615802386775613, + 0.003622015006840229, + 0.010511158034205437, + 0.0037449277006089687, + 0.013059757649898529, + 0.005038348492234945, + 0.003983878064900637, + 0.005085782613605261, + 0.0029979324899613857, + 0.007179354317486286, + 0.008494469337165356, + 0.004690294153988361, + 0.0064752656035125256, + 0.006353867705911398, + 0.004946244414895773, + 0.010994226671755314, + 0.00423045689240098, + 0.0076695834286510944, + 0.01432974636554718, + 0.007803593296557665, + 0.014125448651611805, + 0.010319133289158344, + 0.008500154130160809, + 0.008775906637310982, + 0.007776462007313967, + 0.010110151953995228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.00646949652582407 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005280001554638147 + ], + "opposite_logprob_tokens": [ + -5.255280017852783 + ], + "logprob_reward": -0.005280001554638147, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005280001554638147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3729113042354584, + 0.08350062370300293, + 0.08296021819114685, + 0.09510623663663864, + 0.04660230875015259, + 0.02773665264248848, + 0.015531002543866634, + 0.03184967860579491, + 0.22203679382801056, + 0.0217652078717947 + ], + "credit_map": [ + 0.04212677851319313, + 0.02623969502747059, + 0.07114090770483017, + 0.039650194346904755, + 0.0281680840998888, + 0.08030329644680023, + 0.04032674804329872, + 0.04495559632778168, + 0.010551758110523224, + 0.004802610259503126, + 0.019580120220780373, + 0.01000171061605215, + 0.007215138990432024, + 0.007610219065099955, + 0.008516625501215458, + 0.015222448855638504, + 0.009090662933886051, + 0.005202492233365774, + 0.0059013706631958485, + 0.008263654075562954, + 0.007712800987064838, + 0.016621222719550133, + 0.021305689588189125, + 0.008862326852977276, + 0.01944451965391636, + 0.005765087902545929, + 0.021704889833927155, + 0.006284305825829506, + 0.0061732204630970955, + 0.009480035863816738, + 0.006812586449086666, + 0.01944158785045147, + 0.004904371686279774, + 0.005230364855378866, + 0.008358933962881565, + 0.006982766091823578, + 0.005088326055556536, + 0.005006859544664621, + 0.004275968298316002, + 0.006754718720912933, + 0.004630496259778738, + 0.0024613437708467245, + 0.004551547579467297, + 0.002405357314273715, + 0.002583349123597145, + 0.00319909630343318, + 0.004142172168940306, + 0.003763289423659444, + 0.0023447037674486637, + 0.0013463885989040136, + 0.0027217837050557137, + 0.001455035642720759, + 0.0018767319852486253, + 0.001733013428747654, + 0.001891225459985435, + 0.002162119373679161, + 0.0074703325517475605, + 0.002949716756120324, + 0.006354082841426134, + 0.0030690592247992754, + 0.002937519922852516, + 0.002828805474564433, + 0.002049346687272191, + 0.004190814681351185, + 0.05556060001254082, + 0.013726195320487022, + 0.021792905405163765, + 0.013326619751751423, + 0.014975927770137787, + 0.04081982001662254, + 0.01343864481896162, + 0.04839608818292618, + 0.003343278309330344, + 0.002503043506294489, + 0.004542757757008076, + 0.0031347607728093863, + 0.002708318643271923, + 0.0016942087095230818, + 0.0015655414899811149, + 0.002273298567160964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.005280001554638147 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0540112257003784 + ], + "opposite_logprob_tokens": [ + -0.42901119589805603 + ], + "logprob_reward": -1.0540112257003784, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0540112257003784, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20170702040195465, + 0.21094608306884766, + 0.03690486028790474, + 0.1710515320301056, + 0.054331183433532715, + 0.08696167916059494, + 0.046277400106191635, + 0.1165916919708252, + 0.05013449490070343, + 0.025094104930758476 + ], + "credit_map": [ + 0.028245655819773674, + 0.026547793298959732, + 0.03158063814043999, + 0.028038540855050087, + 0.023421887308359146, + 0.02688993699848652, + 0.019293643534183502, + 0.017688928171992302, + 0.026053421199321747, + 0.018871663138270378, + 0.03364628925919533, + 0.028292112052440643, + 0.024453850463032722, + 0.024087371304631233, + 0.025003155693411827, + 0.030538232997059822, + 0.005697821732610464, + 0.002644074847921729, + 0.004499703645706177, + 0.0032110088504850864, + 0.0038435726892203093, + 0.005947334226220846, + 0.00748189352452755, + 0.003579450771212578, + 0.038143400102853775, + 0.013886342756450176, + 0.034429386258125305, + 0.013760171830654144, + 0.01568022556602955, + 0.016823021695017815, + 0.012205809354782104, + 0.026123175397515297, + 0.005711168050765991, + 0.0042007723823189735, + 0.01420128345489502, + 0.0072547695599496365, + 0.005626881495118141, + 0.004772920627146959, + 0.0056203799322247505, + 0.006943007931113243, + 0.011523827910423279, + 0.006402955390512943, + 0.008782611228525639, + 0.006854448467493057, + 0.00834560301154852, + 0.009174416773021221, + 0.023297147825360298, + 0.01258067786693573, + 0.007594633381813765, + 0.0043257432989776134, + 0.009350349195301533, + 0.004921183921396732, + 0.005760529078543186, + 0.0042861937545239925, + 0.003654082305729389, + 0.0063846823759377, + 0.026170434430241585, + 0.007480944041162729, + 0.03103126771748066, + 0.009304512292146683, + 0.009392179548740387, + 0.009975776076316833, + 0.00680872006341815, + 0.016427861526608467, + 0.007303094025701284, + 0.003800412639975548, + 0.009266681969165802, + 0.0036310518626123667, + 0.004692167043685913, + 0.010425355285406113, + 0.003609804902225733, + 0.007405925542116165, + 0.0040594711899757385, + 0.0027061766013503075, + 0.00442969985306263, + 0.0026621229480952024, + 0.0035146232694387436, + 0.002626169938594103, + 0.0029550837352871895, + 0.0021407546009868383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -1.0540112257003784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009803010150790215 + ], + "opposite_logprob_tokens": [ + -4.63480281829834 + ], + "logprob_reward": -0.009803010150790215, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009803010150790215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14496570825576782, + 0.15698392689228058, + 0.0563378743827343, + 0.2031877636909485, + 0.07852791994810104, + 0.07146043330430984, + 0.042674433439970016, + 0.14345668256282806, + 0.03177669644355774, + 0.07062862813472748 + ], + "credit_map": [ + 0.02012608014047146, + 0.021893678233027458, + 0.01767691597342491, + 0.02259884588420391, + 0.021718228235840797, + 0.012909245677292347, + 0.01660575158894062, + 0.011436937376856804, + 0.0185527503490448, + 0.015039857476949692, + 0.021340185776352882, + 0.02244621329009533, + 0.021568631753325462, + 0.01718120090663433, + 0.01975538767874241, + 0.02109968103468418, + 0.006709659937769175, + 0.004194710869342089, + 0.004123316612094641, + 0.006202041171491146, + 0.006086757406592369, + 0.009973205626010895, + 0.013470939360558987, + 0.005577238742262125, + 0.0432884618639946, + 0.012226527556777, + 0.03065895289182663, + 0.013690633699297905, + 0.011747402139008045, + 0.01999729499220848, + 0.020649384707212448, + 0.05092908442020416, + 0.012405597604811192, + 0.00884447991847992, + 0.010193025693297386, + 0.010471411049365997, + 0.010652097873389721, + 0.007246131543070078, + 0.008056732825934887, + 0.010658434592187405, + 0.010103996843099594, + 0.0054094549268484116, + 0.00840688031166792, + 0.005892605986446142, + 0.00720705883577466, + 0.0069732884876430035, + 0.018390482291579247, + 0.009076662361621857, + 0.006226876750588417, + 0.0035504379775375128, + 0.010785570368170738, + 0.0038800747133791447, + 0.004478720016777515, + 0.004083107691258192, + 0.003172786207869649, + 0.006496857386082411, + 0.026457082480192184, + 0.010079706087708473, + 0.03509622812271118, + 0.016391078010201454, + 0.013125848025083542, + 0.012483610771596432, + 0.009267531335353851, + 0.020555593073368073, + 0.005850506946444511, + 0.0025634625926613808, + 0.004967886954545975, + 0.0035474447067826986, + 0.002937556244432926, + 0.004934235010296106, + 0.00247276877053082, + 0.00450283195823431, + 0.014456713572144508, + 0.004930997733026743, + 0.02142089046537876, + 0.006477150600403547, + 0.0050215646624565125, + 0.00646185502409935, + 0.00503435218706727, + 0.006825098302215338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.009803010150790215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058861404657364 + ], + "opposite_logprob_tokens": [ + -3.8955886363983154 + ], + "logprob_reward": -0.02058861404657364, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058861404657364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09804602712392807, + 0.08541956543922424, + 0.05815279111266136, + 0.11393709480762482, + 0.39450711011886597, + 0.037810735404491425, + 0.03842860087752342, + 0.06440255045890808, + 0.06083502992987633, + 0.048460572957992554 + ], + "credit_map": [ + 0.015318168327212334, + 0.01660999283194542, + 0.012979420833289623, + 0.01332335826009512, + 0.009474221616983414, + 0.013197589665651321, + 0.007964310236275196, + 0.009178954176604748, + 0.006370192859321833, + 0.008835523389279842, + 0.012232597917318344, + 0.011232562363147736, + 0.012469612061977386, + 0.01139395497739315, + 0.010825557634234428, + 0.012059547938406467, + 0.011303178034722805, + 0.0051544709131121635, + 0.0064169918186962605, + 0.0058547561056911945, + 0.008217304944992065, + 0.005091143772006035, + 0.01066832710057497, + 0.005446611903607845, + 0.022448169067502022, + 0.006063101347535849, + 0.017397455871105194, + 0.00708974152803421, + 0.007225272711366415, + 0.012030008248984814, + 0.0121827507391572, + 0.029500581324100494, + 0.04488983005285263, + 0.050581060349941254, + 0.06547336280345917, + 0.07581696659326553, + 0.04477059096097946, + 0.028537068516016006, + 0.0427551195025444, + 0.04168306291103363, + 0.003735146950930357, + 0.003186612855643034, + 0.010934224352240562, + 0.004707239102572203, + 0.003213286865502596, + 0.003678911831229925, + 0.003944806754589081, + 0.004410500638186932, + 0.004991351161152124, + 0.004005314316600561, + 0.007267406675964594, + 0.0056246560998260975, + 0.0036585002671927214, + 0.0033174287527799606, + 0.003470649477094412, + 0.006093290634453297, + 0.01945151388645172, + 0.004263945389539003, + 0.011623422615230083, + 0.004955548793077469, + 0.004302188288420439, + 0.004577968269586563, + 0.003809854155406356, + 0.011418097652494907, + 0.015545075759291649, + 0.004289205651730299, + 0.006802714429795742, + 0.003458531340584159, + 0.0036038716789335012, + 0.010320684872567654, + 0.0034107782412320375, + 0.013404155150055885, + 0.010601808317005634, + 0.003884481731802225, + 0.014097940176725388, + 0.00436901580542326, + 0.003679240122437477, + 0.004267118405550718, + 0.003482484957203269, + 0.004078477155417204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.02058861404657364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up meticulously depicts the face of a young woman, her expression filled with pure, silent awe. Her long, lustrous black hair, elegantly pulled back and tied with a vibrant red satin ribbon, creates a warm and striking contrast to the cool tones of the scene. The sole source of illumination is the spectacular meteor shower, whose ethereal, shimmering light delicately dances across her features, casting a series of fleeting silver highlights across her smooth skin. Her large, dark eyes, sparkling with unshed tears of wonder, act like perfect mirrors of the celestial display above; within their liquid depths, each meteor is reflected as a bright, fleeting flash of light, flaring and fading in an instant. Her lips, slightly parted as if in a silent gasp, her brows gently raised, capture a profoundly intimate and magical moment of connection with the universe, all rendered with an elegant and refined style that emphasizes the scene's serene and wondrous atmosphere.", + "seed": 33178, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_4/index_5581_seed_33178_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.35079908631742, + "vlm_logprob": -0.35079908631742, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.35079908631742, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005159278400242329 + ], + "opposite_logprob_tokens": [ + -5.755159378051758 + ], + "logprob_reward": -0.005159278400242329, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.005159278400242329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10710995644330978, + 0.01966504380106926, + 0.07094268500804901, + 0.1260610818862915, + 0.036733318120241165, + 0.4776223301887512, + 0.06545965373516083, + 0.03018055483698845, + 0.038883548229932785, + 0.027341848239302635 + ], + "credit_map": [ + 0.015456996858119965, + 0.011351642198860645, + 0.015860792249441147, + 0.012456175871193409, + 0.014998383820056915, + 0.010302913375198841, + 0.014833974651992321, + 0.011849075555801392, + 0.0024243886582553387, + 0.00267173838801682, + 0.002278605243191123, + 0.0025460566394031048, + 0.0025656879879534245, + 0.0021500196307897568, + 0.0027805541176348925, + 0.0022479952313005924, + 0.008842346258461475, + 0.014711081050336361, + 0.010240625590085983, + 0.007478829473257065, + 0.00801306962966919, + 0.006726514082401991, + 0.008631842210888863, + 0.006298375315964222, + 0.01081076916307211, + 0.012566995806992054, + 0.01005877647548914, + 0.02161203883588314, + 0.039775438606739044, + 0.009181469678878784, + 0.012789014726877213, + 0.009266584180295467, + 0.003763454034924507, + 0.0049717905931174755, + 0.0044699786230921745, + 0.0053124697878956795, + 0.004423697013407946, + 0.0034020536113530397, + 0.006595526356250048, + 0.0037943460047245026, + 0.1144050806760788, + 0.0351124033331871, + 0.03811018541455269, + 0.1721602827310562, + 0.05229565128684044, + 0.012593898922204971, + 0.028637871146202087, + 0.02430695667862892, + 0.01324815209954977, + 0.005763628054410219, + 0.014659023843705654, + 0.008537896908819675, + 0.004800719674676657, + 0.004501798655837774, + 0.0076352194882929325, + 0.006313213612884283, + 0.002893612952902913, + 0.0036719373892992735, + 0.001974283019080758, + 0.003983396105468273, + 0.004235409200191498, + 0.003647758159786463, + 0.005723466631025076, + 0.004050690680742264, + 0.0026504811830818653, + 0.00363992340862751, + 0.009783589281141758, + 0.010590185411274433, + 0.0041350387036800385, + 0.0027078399434685707, + 0.003720225067809224, + 0.0016562652308493853, + 0.00659015029668808, + 0.003079769667237997, + 0.005832756869494915, + 0.003139972221106291, + 0.002122066915035248, + 0.002235597698017955, + 0.0025328071787953377, + 0.0018087277421727777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.005159278400242329 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005973462946712971 + ], + "opposite_logprob_tokens": [ + -5.1309733390808105 + ], + "logprob_reward": -0.005973462946712971, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005973462946712971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03656934201717377, + 0.025084862485527992, + 0.06451759487390518, + 0.131698340177536, + 0.06182948872447014, + 0.5873150825500488, + 0.03390606492757797, + 0.01777482032775879, + 0.012449369765818119, + 0.028855061158537865 + ], + "credit_map": [ + 0.005431004334241152, + 0.004735044669359922, + 0.005346909165382385, + 0.004501710645854473, + 0.004138546530157328, + 0.003849739907309413, + 0.004908256698399782, + 0.0036581314634531736, + 0.0041390107944607735, + 0.003487143898382783, + 0.0017496179789304733, + 0.002549091586843133, + 0.0033906837925314903, + 0.002567318733781576, + 0.003495903220027685, + 0.003706093179062009, + 0.0075463359244167805, + 0.012340797111392021, + 0.01258991751819849, + 0.007978716865181923, + 0.007553130388259888, + 0.005877248477190733, + 0.005669392179697752, + 0.004962053149938583, + 0.013968564569950104, + 0.010431384667754173, + 0.010598964989185333, + 0.015071901492774487, + 0.045624639838933945, + 0.010196799412369728, + 0.015879079699516296, + 0.009927020408213139, + 0.008927004411816597, + 0.006157746538519859, + 0.008481367491185665, + 0.007451503537595272, + 0.0077719539403915405, + 0.005262583028525114, + 0.012188469991087914, + 0.005588860716670752, + 0.1413031816482544, + 0.04598459601402283, + 0.045314282178878784, + 0.21175706386566162, + 0.06203838810324669, + 0.015514095313847065, + 0.035090722143650055, + 0.030312711372971535, + 0.006772859487682581, + 0.0024836466182023287, + 0.005146102048456669, + 0.002856222679838538, + 0.0035270985681563616, + 0.0025444473139941692, + 0.004635203164070845, + 0.005940488539636135, + 0.0024049945641309023, + 0.002002224326133728, + 0.0013971470762044191, + 0.0016773056704550982, + 0.0027255096938461065, + 0.0024340059608221054, + 0.0030733696185052395, + 0.0020602636504918337, + 0.0015830665361136198, + 0.0011584419989958405, + 0.0036747376434504986, + 0.0029569531325250864, + 0.0008599918219260871, + 0.0006851719808764756, + 0.0009681045776233077, + 0.0005629016668535769, + 0.004565653391182423, + 0.0039855800569057465, + 0.0019669653847813606, + 0.011198543012142181, + 0.001291873399168253, + 0.0016233802307397127, + 0.003006384940817952, + 0.0012166813248768449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.005973462946712971 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7016527652740479 + ], + "opposite_logprob_tokens": [ + -0.20165276527404785 + ], + "logprob_reward": -1.7016527652740479, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7016527652740479, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07575704902410507, + 0.04109915718436241, + 0.02602250501513481, + 0.027393320575356483, + 0.027674010023474693, + 0.4934302568435669, + 0.11411917209625244, + 0.07141777873039246, + 0.08781789243221283, + 0.03526889532804489 + ], + "credit_map": [ + 0.010577984154224396, + 0.010077991522848606, + 0.008859653025865555, + 0.008986413478851318, + 0.010593518614768982, + 0.006117960903793573, + 0.011916841380298138, + 0.008626678958535194, + 0.007342902477830648, + 0.00849202461540699, + 0.005989742465317249, + 0.005963400937616825, + 0.0034382080193609, + 0.0029504557605832815, + 0.0033581287134438753, + 0.0035642923321574926, + 0.0037041965406388044, + 0.0068674818612635136, + 0.0032566979061812162, + 0.0028555935714393854, + 0.00322526041418314, + 0.002272052690386772, + 0.0024070474319159985, + 0.0014341726200655103, + 0.003522069426253438, + 0.003053999738767743, + 0.0023946636356413364, + 0.0025256334338337183, + 0.00748283788561821, + 0.0019824891351163387, + 0.0038334201090037823, + 0.002598205581307411, + 0.0038051961455494165, + 0.00271230167709291, + 0.005068512633442879, + 0.003318506060168147, + 0.004147784784436226, + 0.002404823899269104, + 0.004277376923710108, + 0.0019395045237615705, + 0.11682649701833725, + 0.03697691485285759, + 0.03864816576242447, + 0.17694039642810822, + 0.0560501292347908, + 0.013185885734856129, + 0.030131718143820763, + 0.024670515209436417, + 0.020625369623303413, + 0.009207669645547867, + 0.014651785604655743, + 0.014406868256628513, + 0.010869503952562809, + 0.0089771319180727, + 0.014040590263903141, + 0.021340249106287956, + 0.010218343697488308, + 0.006021149456501007, + 0.006091631483286619, + 0.007586691528558731, + 0.008727816864848137, + 0.009806038811802864, + 0.01254340074956417, + 0.010422704741358757, + 0.005932276137173176, + 0.006326883099973202, + 0.025883283466100693, + 0.02809380367398262, + 0.007757469546049833, + 0.0046001519076526165, + 0.006189534906297922, + 0.003034478286281228, + 0.00802923459559679, + 0.004443231504410505, + 0.006229416932910681, + 0.00822446122765541, + 0.0020973440259695053, + 0.0015627748798578978, + 0.003123700153082609, + 0.0015587278176099062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -1.7016527652740479 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020620377734303474 + ], + "opposite_logprob_tokens": [ + -3.895620346069336 + ], + "logprob_reward": -0.020620377734303474, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020620377734303474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19564968347549438, + 0.11126574873924255, + 0.07668359577655792, + 0.10957502573728561, + 0.05980649217963219, + 0.1309369057416916, + 0.11833778768777847, + 0.08158372342586517, + 0.06240091845393181, + 0.05376005545258522 + ], + "credit_map": [ + 0.03025575913488865, + 0.030780352652072906, + 0.028175977990031242, + 0.030778106302022934, + 0.02170487679541111, + 0.013812065124511719, + 0.021931419149041176, + 0.018211131915450096, + 0.019620027393102646, + 0.01904255524277687, + 0.0176165122538805, + 0.014477438293397427, + 0.009607351385056973, + 0.008420775644481182, + 0.010802480392158031, + 0.011678609997034073, + 0.007638989016413689, + 0.019441770389676094, + 0.012090504169464111, + 0.00787544809281826, + 0.009000873193144798, + 0.006002968177199364, + 0.009013697504997253, + 0.0056193494237959385, + 0.0119977667927742, + 0.008606367744505405, + 0.00851665809750557, + 0.010953563265502453, + 0.03704572468996048, + 0.008707406930625439, + 0.014509127475321293, + 0.00923840794712305, + 0.007295250426977873, + 0.005065761506557465, + 0.010332188569009304, + 0.006748577579855919, + 0.008207504637539387, + 0.006356154102832079, + 0.009363147430121899, + 0.0064379065297544, + 0.03293994814157486, + 0.010846228338778019, + 0.010765301063656807, + 0.04562680423259735, + 0.012006756849586964, + 0.0034215389750897884, + 0.008909815922379494, + 0.006420513615012169, + 0.02168324775993824, + 0.009288525208830833, + 0.015497532673180103, + 0.016483427956700325, + 0.01066944282501936, + 0.009296182543039322, + 0.013636508956551552, + 0.02178291231393814, + 0.011083407327532768, + 0.006839026231318712, + 0.006605165544897318, + 0.009660723619163036, + 0.01005930732935667, + 0.010775281116366386, + 0.015259401872754097, + 0.011301415972411633, + 0.0049519226886332035, + 0.005272688809782267, + 0.011927205137908459, + 0.01066553220152855, + 0.010428276844322681, + 0.0063743977807462215, + 0.00851038470864296, + 0.004270508885383606, + 0.00973581988364458, + 0.006427696906030178, + 0.007771416567265987, + 0.014171813614666462, + 0.004499521106481552, + 0.0028591894078999758, + 0.005651972256600857, + 0.002642623847350478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.020620377734303474 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static extreme close-up of a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's hair long, black, and tied back with a vibrant red satin ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated solely by the light of a meteor shower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the woman's large dark eyes reflecting the meteors as bright flashes of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the woman's lips slightly parted and her brows gently raised?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589547231793404 + ], + "opposite_logprob_tokens": [ + -3.895589590072632 + ], + "logprob_reward": -0.020589547231793404, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589547231793404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13575607538223267, + 0.05920236557722092, + 0.05805337429046631, + 0.20057842135429382, + 0.11359763890504837, + 0.2996935546398163, + 0.03414859622716904, + 0.037166837602853775, + 0.03376714885234833, + 0.02803598903119564 + ], + "credit_map": [ + 0.01873919926583767, + 0.024926304817199707, + 0.023576557636260986, + 0.01583496853709221, + 0.019434327259659767, + 0.00941425934433937, + 0.014210650697350502, + 0.009619807824492455, + 0.009999063797295094, + 0.0084235155954957, + 0.00689642783254385, + 0.008549955673515797, + 0.007440201472491026, + 0.00478338310495019, + 0.007172488607466221, + 0.005937330424785614, + 0.0077809980139136314, + 0.010958905331790447, + 0.008302523754537106, + 0.005258177872747183, + 0.008305277675390244, + 0.005150584038347006, + 0.007582833059132099, + 0.00471407501026988, + 0.018074195832014084, + 0.01566975936293602, + 0.01486666314303875, + 0.030217798426747322, + 0.07064191997051239, + 0.015281579457223415, + 0.02292284183204174, + 0.01290365681052208, + 0.01110148523002863, + 0.012816374190151691, + 0.014431600458920002, + 0.010019353590905666, + 0.016692059114575386, + 0.011761680245399475, + 0.024877704679965973, + 0.011897384189069271, + 0.052938636392354965, + 0.04444006457924843, + 0.02518722228705883, + 0.084504134953022, + 0.032872274518013, + 0.015180005691945553, + 0.02817421220242977, + 0.01639700122177601, + 0.0067541035823524, + 0.0026128708850592375, + 0.00449140090495348, + 0.003210209310054779, + 0.0033066985197365284, + 0.002523500472307205, + 0.005156447645276785, + 0.0060933660715818405, + 0.0030900160782039165, + 0.004103235434740782, + 0.0023343018256127834, + 0.0035973878111690283, + 0.007154475431889296, + 0.003753341268748045, + 0.008058053441345692, + 0.0050760251469910145, + 0.0042441003024578094, + 0.00270385155454278, + 0.008336558006703854, + 0.005156382918357849, + 0.004490512423217297, + 0.002453847788274288, + 0.004669488407671452, + 0.0017124067526310682, + 0.003641459159553051, + 0.0033578025177121162, + 0.0024272522423416376, + 0.008144537918269634, + 0.0030288242269307375, + 0.002070370828732848, + 0.003943103831261396, + 0.0014226390048861504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.020589547231793404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_5/index_tensor([5366, 5366])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_5/index_tensor([5366, 5366])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b75a0c6ab7dbc2ae5e6cd1395d5ea315fff31203 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_5/index_tensor([5366, 5366])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, shot from a level angle, quietly focuses on a corner of an antique, dark-wood bookshelf. On the central shelf, a thick, ancient tome lies flat. Its dark brown, hard leather cover, frayed at the edges and adorned with faded gilt patterns on the spine, reveals a sense of age and wear. A warm, concentrated beam of light slants down from the right side of the frame, illuminating the book perfectly, making even the finest dust particles in the air visible while casting deep shadows across the depths of the shelf. Without any external force, the book's yellowed, slightly rough-edged parchment pages turn one by one in a steady and leisurely rhythm. Each leaf traces a light arc through the air before silently dropping to the other side, as if gently manipulated by an invisible hand. The entire scene exudes a profound tranquility. The warm tones of light and shadow combine with the uncanny, autonomous turning of the pages to create a magical atmosphere imbued with mystery and a sense of time.", + "seed": 25979, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_5/index_5366_seed_25979_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11915789577178656, + "vlm_logprob": -0.11915789577178656, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.11915789577178656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.44047465920448303 + ], + "opposite_logprob_tokens": [ + -1.0654746294021606 + ], + "logprob_reward": -0.44047465920448303, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.44047465920448303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12572605907917023, + 0.042744118720293045, + 0.03747114539146423, + 0.03281250596046448, + 0.125910684466362, + 0.33089131116867065, + 0.09906426817178726, + 0.04363586753606796, + 0.10325153917074203, + 0.05849248543381691 + ], + "credit_map": [ + 0.015194742009043694, + 0.017298894003033638, + 0.011087884195148945, + 0.008688217028975487, + 0.03146766126155853, + 0.018506836146116257, + 0.009147273376584053, + 0.014334551058709621, + 0.004431867506355047, + 0.004346830770373344, + 0.004418822471052408, + 0.004685514606535435, + 0.010978945530951023, + 0.005666063632816076, + 0.003224287647753954, + 0.004991786554455757, + 0.004870728589594364, + 0.004891292192041874, + 0.003177727572619915, + 0.00359854637645185, + 0.007073502987623215, + 0.0055848825722932816, + 0.002422051504254341, + 0.0058524152263998985, + 0.003781083272770047, + 0.0038691763766109943, + 0.0030532884411513805, + 0.006283292546868324, + 0.0032541905529797077, + 0.0038393521681427956, + 0.004503998439759016, + 0.004228122532367706, + 0.01258204784244299, + 0.013409044593572617, + 0.009606015868484974, + 0.01199721172451973, + 0.015545511618256569, + 0.012695365585386753, + 0.031373754143714905, + 0.018701735883951187, + 0.039933908730745316, + 0.05210213363170624, + 0.01781998760998249, + 0.022441796958446503, + 0.060462236404418945, + 0.04449400305747986, + 0.0626172199845314, + 0.031020021066069603, + 0.015976455062627792, + 0.015788057819008827, + 0.008411808870732784, + 0.008774520829319954, + 0.018995780497789383, + 0.013993515633046627, + 0.008444168604910374, + 0.00867996085435152, + 0.003438184969127178, + 0.004614170175045729, + 0.004440045915544033, + 0.005015185568481684, + 0.007534559816122055, + 0.007248479872941971, + 0.003645754186436534, + 0.007699485868215561, + 0.00830864068120718, + 0.014320517890155315, + 0.006970751564949751, + 0.010039295069873333, + 0.027139626443386078, + 0.018483411520719528, + 0.006728826090693474, + 0.011260473169386387, + 0.004482789430767298, + 0.011464701034128666, + 0.005511830095201731, + 0.004593373276293278, + 0.013997063972055912, + 0.011177213862538338, + 0.002860583830624819, + 0.004404930397868156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.44047465920448303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04873301088809967 + ], + "opposite_logprob_tokens": [ + -3.0487329959869385 + ], + "logprob_reward": -0.04873301088809967, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04873301088809967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11225182563066483, + 0.058401837944984436, + 0.025262273848056793, + 0.07872236520051956, + 0.030585335567593575, + 0.3741052448749542, + 0.08531282097101212, + 0.06558085978031158, + 0.12278729677200317, + 0.046990059316158295 + ], + "credit_map": [ + 0.014299409464001656, + 0.01421264186501503, + 0.01039781142026186, + 0.008912519551813602, + 0.02161247283220291, + 0.015362605452537537, + 0.012324001640081406, + 0.015130380168557167, + 0.005409835372120142, + 0.0041243270970880985, + 0.007415103260427713, + 0.0053372979164123535, + 0.008328528143465519, + 0.007613821886479855, + 0.008101191371679306, + 0.012071734294295311, + 0.0027292051818221807, + 0.001836598850786686, + 0.0024695575702935457, + 0.003324156394228339, + 0.0029202948790043592, + 0.0031688869930803776, + 0.002954694675281644, + 0.005858881864696741, + 0.008735470473766327, + 0.008649989031255245, + 0.00971797201782465, + 0.013765345327556133, + 0.010764465667307377, + 0.009559803642332554, + 0.0060037048533558846, + 0.011525620706379414, + 0.0029474275652319193, + 0.0030657616443932056, + 0.0022381606977432966, + 0.0027005255687981844, + 0.004094637930393219, + 0.0033444277942180634, + 0.007372260093688965, + 0.004822138696908951, + 0.042485423386096954, + 0.05989205837249756, + 0.020673299208283424, + 0.026151947677135468, + 0.07327007502317429, + 0.049773793667554855, + 0.06719569116830826, + 0.03466300666332245, + 0.013036897405982018, + 0.013874652795493603, + 0.006756629329174757, + 0.006808570120483637, + 0.01689847558736801, + 0.0127446623519063, + 0.007603758480399847, + 0.007589186541736126, + 0.0068414038978517056, + 0.012188093736767769, + 0.006821852643042803, + 0.007780473679304123, + 0.010186277329921722, + 0.009272796101868153, + 0.004642665386199951, + 0.007847301661968231, + 0.009410863742232323, + 0.0157758928835392, + 0.006821443326771259, + 0.011250477284193039, + 0.03463045135140419, + 0.022718889638781548, + 0.007635681889951229, + 0.014543608762323856, + 0.0037592684384435415, + 0.00713095860555768, + 0.00591128459200263, + 0.005255640018731356, + 0.00807206705212593, + 0.0075495271012187, + 0.003368753707036376, + 0.005942562595009804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.04873301088809967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028347091283649206 + ], + "opposite_logprob_tokens": [ + -5.877834796905518 + ], + "logprob_reward": -0.0028347091283649206, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028347091283649206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16563574969768524, + 0.07474216818809509, + 0.03377823904156685, + 0.07856457680463791, + 0.11340969055891037, + 0.3099896013736725, + 0.07043062895536423, + 0.05957672372460365, + 0.06484463065862656, + 0.029028020799160004 + ], + "credit_map": [ + 0.016184763982892036, + 0.01897943764925003, + 0.01808936335146427, + 0.01130363717675209, + 0.039276666939258575, + 0.02309209480881691, + 0.01777118258178234, + 0.02093859761953354, + 0.007219510618597269, + 0.00836690329015255, + 0.0069826748222112656, + 0.005264832638204098, + 0.022485552355647087, + 0.010907778516411781, + 0.005590964108705521, + 0.007923946715891361, + 0.004035665187984705, + 0.003998680505901575, + 0.0035113650374114513, + 0.0032264208421111107, + 0.006569845601916313, + 0.004638525657355785, + 0.0023197175469249487, + 0.005478017032146454, + 0.008149566128849983, + 0.008149024099111557, + 0.006271112244576216, + 0.007638151291757822, + 0.011429219506680965, + 0.009894740767776966, + 0.015645915642380714, + 0.011386851780116558, + 0.011831933632493019, + 0.012088588438928127, + 0.008211422711610794, + 0.012356282211840153, + 0.01566915027797222, + 0.011016521602869034, + 0.025892293080687523, + 0.016343502327799797, + 0.035792745649814606, + 0.04942667856812477, + 0.017081936821341515, + 0.02030189335346222, + 0.05562751367688179, + 0.04250234737992287, + 0.059883758425712585, + 0.029372718185186386, + 0.009560205973684788, + 0.015897678211331367, + 0.005632949061691761, + 0.006067353300750256, + 0.011950703337788582, + 0.008402896113693714, + 0.006541750393807888, + 0.006377092096954584, + 0.007829480804502964, + 0.006553926970809698, + 0.00488587049767375, + 0.005767300724983215, + 0.010960492305457592, + 0.009112134575843811, + 0.0040291291661560535, + 0.010438387282192707, + 0.006848229561001062, + 0.015199250541627407, + 0.007207768969237804, + 0.004583724308758974, + 0.012267092242836952, + 0.009694141335785389, + 0.0034750818740576506, + 0.005569343455135822, + 0.0024988718796521425, + 0.004489480052143335, + 0.003543191123753786, + 0.0028394663240760565, + 0.004983006045222282, + 0.004748144652694464, + 0.0018471715738996863, + 0.004078690428286791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0028347091283649206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033712275326251984 + ], + "opposite_logprob_tokens": [ + -3.408712387084961 + ], + "logprob_reward": -0.033712275326251984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033712275326251984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1111784428358078, + 0.06512661278247833, + 0.031171640381217003, + 0.17887580394744873, + 0.04832062870264053, + 0.14243875443935394, + 0.061882779002189636, + 0.12419858574867249, + 0.14269527792930603, + 0.09411147981882095 + ], + "credit_map": [ + 0.01187806110829115, + 0.01569327525794506, + 0.009997580200433731, + 0.016749992966651917, + 0.018887726590037346, + 0.015894096344709396, + 0.007350664585828781, + 0.014727042987942696, + 0.006856139283627272, + 0.009334983304142952, + 0.007143903058022261, + 0.010697848163545132, + 0.010656693018972874, + 0.008882110938429832, + 0.004671595059335232, + 0.006883335765451193, + 0.0035531839821487665, + 0.003089021658524871, + 0.004527461249381304, + 0.003770302515476942, + 0.00518055260181427, + 0.00469199288636446, + 0.003261949634179473, + 0.0030971739906817675, + 0.018905047327280045, + 0.017584161832928658, + 0.023497169837355614, + 0.024441992864012718, + 0.025253644213080406, + 0.021761002019047737, + 0.016097404062747955, + 0.03133537247776985, + 0.007384112570434809, + 0.007771049626171589, + 0.0025007447693496943, + 0.0033844681456685066, + 0.0063968924805521965, + 0.00617544911801815, + 0.008010605350136757, + 0.006697306875139475, + 0.016396865248680115, + 0.022404059767723083, + 0.008402188308537006, + 0.014951630495488644, + 0.03175324574112892, + 0.0160280279815197, + 0.016132988035678864, + 0.016369739547371864, + 0.010399159975349903, + 0.007724668364971876, + 0.006905543152242899, + 0.004140573553740978, + 0.012343925423920155, + 0.008227336220443249, + 0.005031455308198929, + 0.007110112346708775, + 0.02012581005692482, + 0.019947756081819534, + 0.011539840139448643, + 0.010918243788182735, + 0.021728793159127235, + 0.01885341666638851, + 0.007957499474287033, + 0.013127218000590801, + 0.014363539405167103, + 0.037686318159103394, + 0.010966228321194649, + 0.011791404336690903, + 0.0272457804530859, + 0.019566480070352554, + 0.01141548901796341, + 0.009660033509135246, + 0.008833312429487705, + 0.015183752402663231, + 0.00994592159986496, + 0.010559176094830036, + 0.018978524953126907, + 0.016165638342499733, + 0.005929217673838139, + 0.008515926077961922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.033712275326251984 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003482431173325 + ], + "opposite_logprob_tokens": [ + -2.6950347423553467 + ], + "logprob_reward": -0.07003482431173325, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003482431173325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07898690551519394, + 0.0716068223118782, + 0.044892407953739166, + 0.07343113422393799, + 0.0732010081410408, + 0.08990907669067383, + 0.14033687114715576, + 0.09726127237081528, + 0.24608367681503296, + 0.08429090678691864 + ], + "credit_map": [ + 0.008431139402091503, + 0.010006406344473362, + 0.010203024372458458, + 0.009108302183449268, + 0.011877769604325294, + 0.009759997949004173, + 0.008214583620429039, + 0.011385674588382244, + 0.006712432485073805, + 0.007517388090491295, + 0.010556305758655071, + 0.008433219976723194, + 0.013494513928890228, + 0.008253587409853935, + 0.007847513072192669, + 0.008791852742433548, + 0.003614902961999178, + 0.004291384480893612, + 0.00524990214034915, + 0.009863907471299171, + 0.006832188460975885, + 0.005778212565928698, + 0.004864704329520464, + 0.004397199489176273, + 0.008319360204041004, + 0.00892120972275734, + 0.006325853988528252, + 0.0055700927041471004, + 0.013447524979710579, + 0.010040462017059326, + 0.010027100332081318, + 0.010779522359371185, + 0.0068948958069086075, + 0.006765448488295078, + 0.0075782244093716145, + 0.006839359644800425, + 0.009805889800190926, + 0.008163626305758953, + 0.014603442512452602, + 0.012550114654004574, + 0.012115702033042908, + 0.012603670358657837, + 0.005069836508482695, + 0.007592272479087114, + 0.017503779381513596, + 0.012926016934216022, + 0.014191878028213978, + 0.00790590699762106, + 0.02097032405436039, + 0.02494162879884243, + 0.011685517616569996, + 0.010845004580914974, + 0.028006982058286667, + 0.019674208015203476, + 0.01215699128806591, + 0.012056204490363598, + 0.011134915053844452, + 0.012041865848004818, + 0.008967950008809566, + 0.00692348275333643, + 0.018484821543097496, + 0.015935145318508148, + 0.006890688557177782, + 0.016882386058568954, + 0.019663112238049507, + 0.05392829701304436, + 0.019498171284794807, + 0.014114169403910637, + 0.061254967004060745, + 0.04004474729299545, + 0.012509805150330067, + 0.02507035806775093, + 0.00637098029255867, + 0.01577695831656456, + 0.0072867004200816154, + 0.005802926141768694, + 0.02244366519153118, + 0.01580260694026947, + 0.004509730730205774, + 0.0062973289750516415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.07003482431173325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot from a level angle, quietly focuses on a corner of an antique, dark-wood bookshelf. On the central shelf, a thick, ancient tome lies flat. Its dark brown, hard leather cover, frayed at the edges and adorned with faded gilt patterns on the spine, reveals a sense of age and wear. A warm, concentrated beam of light slants down from the right side of the frame, illuminating the book perfectly, making even the finest dust particles in the air visible while casting deep shadows across the depths of the shelf. Without any external force, the book's yellowed, slightly rough-edged parchment pages turn one by one in a steady and leisurely rhythm. Each leaf traces a light arc through the air before silently dropping to the other side, as if gently manipulated by an invisible hand. The entire scene exudes a profound tranquility. The warm tones of light and shadow combine with the uncanny, autonomous turning of the pages to create a magical atmosphere imbued with mystery and a sense of time.", + "seed": 25980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_5/index_5366_seed_25980_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026967624109238385, + "vlm_logprob": -0.026967624109238385, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.026967624109238385, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025816014036536217 + ], + "opposite_logprob_tokens": [ + -3.900815963745117 + ], + "logprob_reward": -0.025816014036536217, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025816014036536217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1621778905391693, + 0.06004136800765991, + 0.09997668117284775, + 0.06599846482276917, + 0.07365014404058456, + 0.17185992002487183, + 0.07188305258750916, + 0.08074015378952026, + 0.04475875943899155, + 0.1689135581254959 + ], + "credit_map": [ + 0.022038308903574944, + 0.018184028565883636, + 0.013423966243863106, + 0.028895383700728416, + 0.018688084557652473, + 0.026867682114243507, + 0.01973159983754158, + 0.014348841272294521, + 0.007657212670892477, + 0.012147915549576283, + 0.0051752799190580845, + 0.005964371841400862, + 0.010753091424703598, + 0.006484212353825569, + 0.006469234358519316, + 0.0053900484926998615, + 0.014249157160520554, + 0.016942329704761505, + 0.011903525330126286, + 0.009966832585632801, + 0.013977797701954842, + 0.008691865019500256, + 0.009316292591392994, + 0.014928879216313362, + 0.010984214022755623, + 0.01598869077861309, + 0.006400200072675943, + 0.009714355692267418, + 0.0038259276188910007, + 0.005221507977694273, + 0.008273419924080372, + 0.0055901529267430305, + 0.00952425878494978, + 0.013907080516219139, + 0.006259652320295572, + 0.008865399286150932, + 0.005245652981102467, + 0.013165593147277832, + 0.010384238325059414, + 0.00629827193915844, + 0.01650533638894558, + 0.02721821516752243, + 0.021729111671447754, + 0.02843690663576126, + 0.01764119602739811, + 0.019267337396740913, + 0.0269822608679533, + 0.014079570770263672, + 0.007279780227690935, + 0.011077545583248138, + 0.008571745827794075, + 0.012879468500614166, + 0.006525388918817043, + 0.007893840782344341, + 0.010980400256812572, + 0.006674878299236298, + 0.009415949694812298, + 0.01125787291675806, + 0.015054362826049328, + 0.012291177175939083, + 0.007825320586562157, + 0.005883765406906605, + 0.011489751748740673, + 0.007521952968090773, + 0.005602086428552866, + 0.007207463029772043, + 0.005701199173927307, + 0.0065057845786213875, + 0.004212397150695324, + 0.004787068348377943, + 0.004762052558362484, + 0.00598070677369833, + 0.014796653762459755, + 0.027810297906398773, + 0.017956644296646118, + 0.020229971036314964, + 0.025090189650654793, + 0.020486023277044296, + 0.031512632966041565, + 0.011031143367290497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.025816014036536217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018233541399240494 + ], + "opposite_logprob_tokens": [ + -4.018233776092529 + ], + "logprob_reward": -0.018233541399240494, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018233541399240494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1732606291770935, + 0.06703957170248032, + 0.10802962630987167, + 0.10903573036193848, + 0.10823230445384979, + 0.10592070966959, + 0.09562600404024124, + 0.15915684401988983, + 0.03999250382184982, + 0.033706046640872955 + ], + "credit_map": [ + 0.025087404996156693, + 0.028705431148409843, + 0.012056536041200161, + 0.040001481771469116, + 0.013623827137053013, + 0.01656324975192547, + 0.023897608742117882, + 0.013325091451406479, + 0.009657380171120167, + 0.015076575800776482, + 0.004176618065685034, + 0.010294962674379349, + 0.007595761679112911, + 0.005309500731527805, + 0.008286542259156704, + 0.0066422270610928535, + 0.023770881816744804, + 0.029391376301646233, + 0.009591815993189812, + 0.011449252255260944, + 0.011320688761770725, + 0.005836326163262129, + 0.008482742123305798, + 0.00818653590977192, + 0.027254050597548485, + 0.03785547986626625, + 0.0075739724561572075, + 0.010505994781851768, + 0.005592004861682653, + 0.005755700170993805, + 0.0080043263733387, + 0.006494208239018917, + 0.010517485439777374, + 0.017599303275346756, + 0.012229980900883675, + 0.01993129774928093, + 0.008677397854626179, + 0.01212387066334486, + 0.01791897416114807, + 0.009233998134732246, + 0.00881071574985981, + 0.012564224191009998, + 0.007745074573904276, + 0.020321695134043694, + 0.01051744818687439, + 0.011482873000204563, + 0.025612276047468185, + 0.008866393007338047, + 0.008187812753021717, + 0.014779952354729176, + 0.014736435376107693, + 0.018786685541272163, + 0.0075386823154985905, + 0.008941737934947014, + 0.014810938388109207, + 0.007843762636184692, + 0.02047165296971798, + 0.021898843348026276, + 0.03021649830043316, + 0.023908406496047974, + 0.016609707847237587, + 0.011922841891646385, + 0.019670676440000534, + 0.014458219520747662, + 0.003790493356063962, + 0.004398358520120382, + 0.0035480065271258354, + 0.004162666853517294, + 0.005756164900958538, + 0.008778909221291542, + 0.005632491782307625, + 0.003925411030650139, + 0.0032563582062721252, + 0.005172609351575375, + 0.004458046518266201, + 0.004010685253888369, + 0.005087184254080057, + 0.004062663298100233, + 0.005111856386065483, + 0.002546645002439618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.018233541399240494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004652745556086302 + ], + "opposite_logprob_tokens": [ + -5.379652976989746 + ], + "logprob_reward": -0.004652745556086302, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004652745556086302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10111667960882187, + 0.07384917140007019, + 0.08087515830993652, + 0.12462297081947327, + 0.13725896179676056, + 0.10069833695888519, + 0.11385499686002731, + 0.2006704956293106, + 0.046689629554748535, + 0.02036365680396557 + ], + "credit_map": [ + 0.013134108856320381, + 0.0150022953748703, + 0.006150688044726849, + 0.012058638967573643, + 0.014730232767760754, + 0.020069818943738937, + 0.009033707901835442, + 0.010937176644802094, + 0.011651579290628433, + 0.01833726465702057, + 0.006450658198446035, + 0.010584106668829918, + 0.007475563790649176, + 0.005331151653081179, + 0.007118778768926859, + 0.006900065112859011, + 0.01737653650343418, + 0.018913118168711662, + 0.008024614304304123, + 0.007007901091128588, + 0.00814660545438528, + 0.006076807156205177, + 0.006220260169357061, + 0.00910930335521698, + 0.037036314606666565, + 0.04898100346326828, + 0.006004599388688803, + 0.007051785010844469, + 0.0063850171864032745, + 0.005409725941717625, + 0.007269425317645073, + 0.006485088728368282, + 0.014009799808263779, + 0.023511946201324463, + 0.016167744994163513, + 0.023437298834323883, + 0.010785265825688839, + 0.018343806266784668, + 0.01926913857460022, + 0.011733941733837128, + 0.00784219615161419, + 0.011621722020208836, + 0.006939723156392574, + 0.01988711580634117, + 0.00953563954681158, + 0.01006794162094593, + 0.02574835903942585, + 0.009055627509951591, + 0.010281183756887913, + 0.01879170723259449, + 0.017556549981236458, + 0.021204445511102676, + 0.008724992163479328, + 0.010318232700228691, + 0.01783117651939392, + 0.009146706201136112, + 0.0246612299233675, + 0.02684532105922699, + 0.04082845523953438, + 0.030459651723504066, + 0.01960589922964573, + 0.014767689630389214, + 0.026289569213986397, + 0.017212659120559692, + 0.0037373353261500597, + 0.006134694907814264, + 0.0049482593312859535, + 0.007071048486977816, + 0.004740349482744932, + 0.005273033399134874, + 0.00995360966771841, + 0.004831294529139996, + 0.0016867968952283263, + 0.003937887493520975, + 0.0016630166210234165, + 0.0028861421160399914, + 0.0024439143016934395, + 0.001982462592422962, + 0.004257249180227518, + 0.0015061857411637902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.004652745556086302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043066710233688354 + ], + "opposite_logprob_tokens": [ + -3.1680667400360107 + ], + "logprob_reward": -0.043066710233688354, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043066710233688354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11241945624351501, + 0.1009025126695633, + 0.18360267579555511, + 0.16093644499778748, + 0.06199679896235466, + 0.12771350145339966, + 0.025545459240674973, + 0.0733635425567627, + 0.13055583834648132, + 0.022963784635066986 + ], + "credit_map": [ + 0.01447166595607996, + 0.012344549410045147, + 0.007167806848883629, + 0.028667621314525604, + 0.01185799203813076, + 0.012984644621610641, + 0.0165568757802248, + 0.008368297480046749, + 0.011936705559492111, + 0.01557797472923994, + 0.007439886685460806, + 0.02101459912955761, + 0.012470882385969162, + 0.011746739968657494, + 0.013257844373583794, + 0.007457870524376631, + 0.0358395017683506, + 0.03560291975736618, + 0.022068174555897713, + 0.014359746128320694, + 0.018941348418593407, + 0.01907181367278099, + 0.015448693186044693, + 0.022270480170845985, + 0.04745979234576225, + 0.06205739080905914, + 0.00887210015207529, + 0.007878503762185574, + 0.008727282285690308, + 0.009407680481672287, + 0.008224030956625938, + 0.008309655822813511, + 0.0066946763545274734, + 0.008470120839774609, + 0.008705093525350094, + 0.0075798314064741135, + 0.00691598653793335, + 0.009411267936229706, + 0.00898263230919838, + 0.005237189121544361, + 0.009716653265058994, + 0.017030028626322746, + 0.020306594669818878, + 0.015416301786899567, + 0.016721785068511963, + 0.02337263897061348, + 0.012962916865944862, + 0.012186583131551743, + 0.002003842731937766, + 0.0029726815409958363, + 0.0021267568226903677, + 0.004072306212037802, + 0.003096476662904024, + 0.006584436632692814, + 0.0024124591145664454, + 0.00227649905718863, + 0.007503921631723642, + 0.011291952803730965, + 0.00791169609874487, + 0.010721043683588505, + 0.00797637365758419, + 0.0045685176737606525, + 0.014283818192780018, + 0.00910621602088213, + 0.012832794338464737, + 0.015263817273080349, + 0.009717948734760284, + 0.012572822161018848, + 0.014745096676051617, + 0.02612202800810337, + 0.021462630480527878, + 0.017838701605796814, + 0.001773978117853403, + 0.003719631116837263, + 0.0020428309217095375, + 0.0036259829066693783, + 0.002344537992030382, + 0.001513038412667811, + 0.005709872115403414, + 0.00223391386680305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.043066710233688354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043069109320640564 + ], + "opposite_logprob_tokens": [ + -3.1680691242218018 + ], + "logprob_reward": -0.043069109320640564, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043069109320640564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12170039862394333, + 0.09196054935455322, + 0.12128407508134842, + 0.19678640365600586, + 0.07274642586708069, + 0.11965582519769669, + 0.0714796632528305, + 0.06257278472185135, + 0.08095204830169678, + 0.060861747711896896 + ], + "credit_map": [ + 0.015311681665480137, + 0.013011297211050987, + 0.007207774557173252, + 0.022923480719327927, + 0.015453631989657879, + 0.016107119619846344, + 0.019271917641162872, + 0.012413501739501953, + 0.011268235743045807, + 0.014853774569928646, + 0.006513637490570545, + 0.013630016706883907, + 0.013563737273216248, + 0.012772963382303715, + 0.011963761411607265, + 0.007394431624561548, + 0.02903841622173786, + 0.031069882214069366, + 0.008404147811233997, + 0.007998423650860786, + 0.011702317744493484, + 0.009925221092998981, + 0.00930815190076828, + 0.013837517239153385, + 0.055952735245227814, + 0.0734684094786644, + 0.011933357454836369, + 0.009938659146428108, + 0.011286787688732147, + 0.01212471816688776, + 0.011178245767951012, + 0.01090351864695549, + 0.007902591489255428, + 0.010104081593453884, + 0.00817768182605505, + 0.008609462529420853, + 0.008032767102122307, + 0.017061814665794373, + 0.006680164020508528, + 0.006177871022373438, + 0.009412324987351894, + 0.017067277804017067, + 0.019776515662670135, + 0.0147608183324337, + 0.013283299282193184, + 0.018656285479664803, + 0.014871071092784405, + 0.011828247457742691, + 0.006524595431983471, + 0.012945398688316345, + 0.01152205839753151, + 0.013657732866704464, + 0.00539267435669899, + 0.00546465115621686, + 0.00954261515289545, + 0.006429947447031736, + 0.008294705301523209, + 0.006995900999754667, + 0.01094820722937584, + 0.009508701972663403, + 0.008954958990216255, + 0.0046280077658593655, + 0.0074072168208658695, + 0.005835089832544327, + 0.007823020219802856, + 0.010316423140466213, + 0.009611447341740131, + 0.007724152412265539, + 0.010980788618326187, + 0.014835418201982975, + 0.008166744373738766, + 0.011494064703583717, + 0.005595839582383633, + 0.011775878258049488, + 0.006035347934812307, + 0.008069921284914017, + 0.009353792294859886, + 0.007300083059817553, + 0.007798198144882917, + 0.004932694137096405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.043069109320640564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot from a level angle, quietly focuses on a corner of an antique, dark-wood bookshelf. On the central shelf, a thick, ancient tome lies flat. Its dark brown, hard leather cover, frayed at the edges and adorned with faded gilt patterns on the spine, reveals a sense of age and wear. A warm, concentrated beam of light slants down from the right side of the frame, illuminating the book perfectly, making even the finest dust particles in the air visible while casting deep shadows across the depths of the shelf. Without any external force, the book's yellowed, slightly rough-edged parchment pages turn one by one in a steady and leisurely rhythm. Each leaf traces a light arc through the air before silently dropping to the other side, as if gently manipulated by an invisible hand. The entire scene exudes a profound tranquility. The warm tones of light and shadow combine with the uncanny, autonomous turning of the pages to create a magical atmosphere imbued with mystery and a sense of time.", + "seed": 25981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_5/index_5366_seed_25981_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.058765841834247115, + "vlm_logprob": -0.058765841834247115, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.058765841834247115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10966385155916214 + ], + "opposite_logprob_tokens": [ + -2.359663963317871 + ], + "logprob_reward": -0.10966385155916214, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10966385155916214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28710228204727173, + 0.25055012106895447, + 0.057354822754859924, + 0.03392564877867699, + 0.04796479642391205, + 0.02432253770530224, + 0.04306534677743912, + 0.09907346218824387, + 0.029535265639424324, + 0.1271056830883026 + ], + "credit_map": [ + 0.019624978303909302, + 0.021323464810848236, + 0.09747476130723953, + 0.025908298790454865, + 0.030158929526805878, + 0.029091039672493935, + 0.02051425352692604, + 0.04300655797123909, + 0.01125540304929018, + 0.010744215920567513, + 0.09022068977355957, + 0.013158042915165424, + 0.014299142174422741, + 0.016588222235441208, + 0.01832897588610649, + 0.07595543563365936, + 0.0037846234627068043, + 0.008426649495959282, + 0.006821437738835812, + 0.005046718753874302, + 0.009053187444806099, + 0.006749291904270649, + 0.00508329551666975, + 0.012389620766043663, + 0.002522739116102457, + 0.0038054680917412043, + 0.006083906162530184, + 0.0031097193714231253, + 0.0037582111544907093, + 0.006781756412237883, + 0.003811303758993745, + 0.004052542150020599, + 0.003963821567595005, + 0.006427194457501173, + 0.010967005975544453, + 0.0041893115267157555, + 0.006776753347367048, + 0.006062167696654797, + 0.00434883451089263, + 0.005229707807302475, + 0.001985270529985428, + 0.0030074056703597307, + 0.004479193594306707, + 0.0041367048397660255, + 0.0027143533807247877, + 0.0022743651643395424, + 0.0024681456852704287, + 0.003257099539041519, + 0.0030453461222350597, + 0.0013562911190092564, + 0.016063997521996498, + 0.0015642775688320398, + 0.0011023480910807848, + 0.00413409573957324, + 0.001956766238436103, + 0.013842226006090641, + 0.004203473217785358, + 0.0036760088987648487, + 0.04067572206258774, + 0.00421956367790699, + 0.005264469422399998, + 0.004495685454457998, + 0.004723977297544479, + 0.031814564019441605, + 0.0026529706083238125, + 0.0031121261417865753, + 0.004606745205819607, + 0.003141438588500023, + 0.0038723659235984087, + 0.0035466921981424093, + 0.0031178256031125784, + 0.005485100671648979, + 0.006660774815827608, + 0.005757911596447229, + 0.055033743381500244, + 0.006811945233494043, + 0.006801370065659285, + 0.004842396359890699, + 0.007733052130788565, + 0.03346448391675949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.10966385155916214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029854901134967804 + ], + "opposite_logprob_tokens": [ + -3.5298550128936768 + ], + "logprob_reward": -0.029854901134967804, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029854901134967804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29011794924736023, + 0.33336904644966125, + 0.0380183644592762, + 0.021285003051161766, + 0.02693360671401024, + 0.019456205889582634, + 0.01341470330953598, + 0.09461470693349838, + 0.024002578109502792, + 0.13878785073757172 + ], + "credit_map": [ + 0.020326560363173485, + 0.02026437781751156, + 0.10397551208734512, + 0.018710143864154816, + 0.03045671246945858, + 0.028019586578011513, + 0.02203165367245674, + 0.046333398669958115, + 0.014350411482155323, + 0.014793649315834045, + 0.13241855800151825, + 0.014567617326974869, + 0.015560584142804146, + 0.01629164069890976, + 0.020960351452231407, + 0.10442619770765305, + 0.0026332330889999866, + 0.006106209475547075, + 0.0037301830016076565, + 0.003756860736757517, + 0.006735804490745068, + 0.005296255461871624, + 0.0037485910579562187, + 0.0060112266801297665, + 0.0015923368046060205, + 0.003026141319423914, + 0.0032180389389395714, + 0.0020938869565725327, + 0.0033666975796222687, + 0.002738978946581483, + 0.002607049886137247, + 0.0026418748311698437, + 0.0024211311247199774, + 0.003653591265901923, + 0.004500082693994045, + 0.002484032651409507, + 0.003726002760231495, + 0.0046285586431622505, + 0.0024981233291327953, + 0.0030220872722566128, + 0.0013691739877685905, + 0.0021221833303570747, + 0.004097457509487867, + 0.002771601779386401, + 0.0024151632096618414, + 0.0019408484222367406, + 0.0017837464110925794, + 0.002956030424684286, + 0.0010978356003761292, + 0.0011808180715888739, + 0.0038407461252063513, + 0.0010259776609018445, + 0.0011245873756706715, + 0.0010128646390512586, + 0.0009719988447614014, + 0.0031598745845258236, + 0.003951770719140768, + 0.004012830555438995, + 0.03870679438114166, + 0.004191666375845671, + 0.005624666344374418, + 0.004083891864866018, + 0.004268817603588104, + 0.029774267226457596, + 0.002222280018031597, + 0.0021716284099966288, + 0.004675473086535931, + 0.0017300881445407867, + 0.002715924521908164, + 0.0038458751514554024, + 0.0019795119296759367, + 0.004661795683205128, + 0.00756034255027771, + 0.006493702530860901, + 0.0608522891998291, + 0.007760822772979736, + 0.008298884145915508, + 0.005135520361363888, + 0.00913600530475378, + 0.03355029225349426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.029854901134967804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032039759680628777 + ], + "opposite_logprob_tokens": [ + -5.753203868865967 + ], + "logprob_reward": -0.0032039759680628777, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032039759680628777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19620442390441895, + 0.26324203610420227, + 0.057656180113554, + 0.02291448600590229, + 0.020045943558216095, + 0.013177248649299145, + 0.013359473086893559, + 0.042147908359766006, + 0.13506954908370972, + 0.23618271946907043 + ], + "credit_map": [ + 0.01226290687918663, + 0.013434121385216713, + 0.0663677453994751, + 0.01599060744047165, + 0.020861128345131874, + 0.017958100885152817, + 0.018469780683517456, + 0.03086004965007305, + 0.011003587394952774, + 0.010940337553620338, + 0.10071732848882675, + 0.012978610582649708, + 0.012816176749765873, + 0.01417845580726862, + 0.019914941862225533, + 0.0806925967335701, + 0.004888394381850958, + 0.007652945816516876, + 0.006562612019479275, + 0.005806958302855492, + 0.0095206992700696, + 0.006291482597589493, + 0.005377078894525766, + 0.01155601255595684, + 0.001603933284059167, + 0.0030908051412552595, + 0.003685524221509695, + 0.0020815355237573385, + 0.003171407151967287, + 0.004513886291533709, + 0.0021816366352140903, + 0.00258575938642025, + 0.0014617249835282564, + 0.0018034683307632804, + 0.00449726264923811, + 0.0016686177114024758, + 0.0032904348336160183, + 0.0031161103397607803, + 0.00166199984960258, + 0.0025463267229497433, + 0.0011017329525202513, + 0.0018717384664341807, + 0.0020258757285773754, + 0.0017184587195515633, + 0.0012234841706231236, + 0.0017800962086766958, + 0.0014097649836912751, + 0.0020460993982851505, + 0.0008410349837504327, + 0.0012494149850681424, + 0.00273760873824358, + 0.0012891149381175637, + 0.0014954013749957085, + 0.0017181361326947808, + 0.0009192347642965615, + 0.003109527751803398, + 0.001783334999345243, + 0.00152319495100528, + 0.0174447912722826, + 0.0018215636955574155, + 0.0019895918667316437, + 0.001945589785464108, + 0.0022276961244642735, + 0.01341214869171381, + 0.008998344652354717, + 0.010463681071996689, + 0.03656776621937752, + 0.00786777026951313, + 0.011383287608623505, + 0.014552267268300056, + 0.009110212326049805, + 0.03612622991204262, + 0.01286336313933134, + 0.010824994184076786, + 0.10486043244600296, + 0.012002650648355484, + 0.014900772832334042, + 0.008127987384796143, + 0.01622653380036354, + 0.05637598782777786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0032039759680628777 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205982342362404 + ], + "opposite_logprob_tokens": [ + -2.8120598793029785 + ], + "logprob_reward": -0.06205982342362404, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205982342362404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08031408488750458, + 0.3264051377773285, + 0.09730567038059235, + 0.019506337121129036, + 0.018922338262200356, + 0.02257777750492096, + 0.0416581854224205, + 0.07384965568780899, + 0.19939330220222473, + 0.12006760388612747 + ], + "credit_map": [ + 0.007240150589495897, + 0.00771256722509861, + 0.008518957532942295, + 0.011191063560545444, + 0.007018491625785828, + 0.011992305517196655, + 0.010553653351962566, + 0.016086893156170845, + 0.013712712563574314, + 0.011370536871254444, + 0.1244148537516594, + 0.018492400646209717, + 0.015324754640460014, + 0.01745562069118023, + 0.027714911848306656, + 0.0979192852973938, + 0.00797309260815382, + 0.009854519739747047, + 0.008850350975990295, + 0.01119331456720829, + 0.012761498801410198, + 0.02232426218688488, + 0.009472076781094074, + 0.014876545406877995, + 0.0014905232237651944, + 0.002233669627457857, + 0.00299794040620327, + 0.0018652345752343535, + 0.00291662965901196, + 0.0030271285213530064, + 0.0021800848189741373, + 0.0027951255906373262, + 0.0019297933904454112, + 0.0027319611981511116, + 0.003347889520227909, + 0.002278353087604046, + 0.002644779859110713, + 0.0022918335162103176, + 0.0016287608304992318, + 0.002068963833153248, + 0.0022672214545309544, + 0.0034853711258620024, + 0.003371157217770815, + 0.002515436615794897, + 0.002710203407332301, + 0.0025477646850049496, + 0.0024902885779738426, + 0.0031903341878205538, + 0.0027293721213936806, + 0.0017389258136972785, + 0.013567234389483929, + 0.0020729468669742346, + 0.0021676579490303993, + 0.004379308316856623, + 0.002020817017182708, + 0.012981919571757317, + 0.0029735027346760035, + 0.0022578511852771044, + 0.03126184642314911, + 0.002613968448713422, + 0.0019164675613865256, + 0.0033419127576053143, + 0.0033469267655164003, + 0.026137176901102066, + 0.014408349990844727, + 0.019526084885001183, + 0.042612962424755096, + 0.014298414811491966, + 0.019808737561106682, + 0.027209145948290825, + 0.015130107291042805, + 0.04639947786927223, + 0.006722505670040846, + 0.00567833473905921, + 0.05354035273194313, + 0.006106339395046234, + 0.005408786237239838, + 0.00390577781945467, + 0.008083546534180641, + 0.03062194585800171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.06205982342362404 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890466570854187 + ], + "opposite_logprob_tokens": [ + -2.4640467166900635 + ], + "logprob_reward": -0.0890466570854187, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890466570854187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09645092487335205, + 0.13090963661670685, + 0.021605590358376503, + 0.01589393801987171, + 0.019904613494873047, + 0.010824793018400669, + 0.014096124097704887, + 0.03551196679472923, + 0.18006670475006104, + 0.47473570704460144 + ], + "credit_map": [ + 0.006781921721994877, + 0.007234807591885328, + 0.03408726304769516, + 0.006436422932893038, + 0.0053926738910377026, + 0.0063981954008340836, + 0.00931132584810257, + 0.02080831490457058, + 0.005810339469462633, + 0.004481234587728977, + 0.04960689693689346, + 0.008378724567592144, + 0.0065421718172729015, + 0.006496830377727747, + 0.01150087546557188, + 0.03809257596731186, + 0.001639666617847979, + 0.002515403786674142, + 0.0023378527257591486, + 0.002306088339537382, + 0.0027372196782380342, + 0.004526147618889809, + 0.0016941010253503919, + 0.0038491091690957546, + 0.0012765625724568963, + 0.0015428898623213172, + 0.0027484335005283356, + 0.001395671279169619, + 0.0015941617311909795, + 0.003336576046422124, + 0.0016159976366907358, + 0.0023836444597691298, + 0.0015342426486313343, + 0.0022782369051128626, + 0.0025305424351245165, + 0.00219542789272964, + 0.003216183977201581, + 0.003391427919268608, + 0.0018989935051649809, + 0.00285955797880888, + 0.0009645897662267089, + 0.0015987507067620754, + 0.0016865136567503214, + 0.001063350006006658, + 0.0018696162151172757, + 0.000826773582957685, + 0.0009257421479560435, + 0.0018894568784162402, + 0.0010513191809877753, + 0.0010419735917821527, + 0.0030073646921664476, + 0.0012723777908831835, + 0.0015120834577828646, + 0.0016992316814139485, + 0.0012694057077169418, + 0.003242367645725608, + 0.0015324291307479143, + 0.0014552919892594218, + 0.014203452505171299, + 0.001460870262235403, + 0.001568953855894506, + 0.0017429208382964134, + 0.0016346656484529376, + 0.011913380585610867, + 0.012173393741250038, + 0.014841828495264053, + 0.04885249584913254, + 0.011085203848779202, + 0.014493380673229694, + 0.020248571410775185, + 0.011979527771472931, + 0.04639231786131859, + 0.02535262145102024, + 0.021616794168949127, + 0.21174246072769165, + 0.023396247997879982, + 0.023587653413414955, + 0.016717780381441116, + 0.030473219230771065, + 0.1218489333987236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.0890466570854187 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot from a level angle, quietly focuses on a corner of an antique, dark-wood bookshelf. On the central shelf, a thick, ancient tome lies flat. Its dark brown, hard leather cover, frayed at the edges and adorned with faded gilt patterns on the spine, reveals a sense of age and wear. A warm, concentrated beam of light slants down from the right side of the frame, illuminating the book perfectly, making even the finest dust particles in the air visible while casting deep shadows across the depths of the shelf. Without any external force, the book's yellowed, slightly rough-edged parchment pages turn one by one in a steady and leisurely rhythm. Each leaf traces a light arc through the air before silently dropping to the other side, as if gently manipulated by an invisible hand. The entire scene exudes a profound tranquility. The warm tones of light and shadow combine with the uncanny, autonomous turning of the pages to create a magical atmosphere imbued with mystery and a sense of time.", + "seed": 25982, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_5/index_5366_seed_25982_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.034369297325611115, + "vlm_logprob": -0.034369297325611115, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.034369297325611115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0209150779992342 + ], + "opposite_logprob_tokens": [ + -4.2709150314331055 + ], + "logprob_reward": -0.0209150779992342, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0209150779992342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7264779210090637, + 0.045724596828222275, + 0.021614018827676773, + 0.020430974662303925, + 0.03536861017346382, + 0.013097203336656094, + 0.032179661095142365, + 0.08087953925132751, + 0.019466711208224297, + 0.004760751035064459 + ], + "credit_map": [ + 0.10924414545297623, + 0.11701253801584244, + 0.025541458278894424, + 0.028147432953119278, + 0.2818787395954132, + 0.0654991865158081, + 0.04372534528374672, + 0.0554291196167469, + 0.006707565858960152, + 0.003644898533821106, + 0.0032989229075610638, + 0.0051489402540028095, + 0.003092121798545122, + 0.008090997114777565, + 0.00540670333430171, + 0.01033445168286562, + 0.003031695494428277, + 0.003182691289111972, + 0.0020638606511056423, + 0.001925936434417963, + 0.0018862629076465964, + 0.002883243840187788, + 0.0025439332239329815, + 0.004096397664397955, + 0.0029940211679786444, + 0.003229177091270685, + 0.0023002817761152983, + 0.0021949182264506817, + 0.0017229221994057298, + 0.003010212443768978, + 0.0026604861486703157, + 0.0023189580533653498, + 0.005372591316699982, + 0.004523212555795908, + 0.0026168187614530325, + 0.005013559479266405, + 0.0025065720546990633, + 0.005477437283843756, + 0.003520253114402294, + 0.006338168401271105, + 0.0018289451254531741, + 0.001213584328070283, + 0.0016863433411344886, + 0.0014872633619233966, + 0.001196785713545978, + 0.00183321931399405, + 0.001958082662895322, + 0.001892981119453907, + 0.0057095494121313095, + 0.0029979501850903034, + 0.0025936432648450136, + 0.0037189600989222527, + 0.00349821662530303, + 0.004480935167521238, + 0.0033569596707820892, + 0.005823451094329357, + 0.008574876002967358, + 0.007457949686795473, + 0.006914735306054354, + 0.007101931609213352, + 0.005070853047072887, + 0.017629709094762802, + 0.01338136475533247, + 0.014748124405741692, + 0.0024853094946593046, + 0.002401513047516346, + 0.001968495547771454, + 0.0018776997458189726, + 0.0012723931577056646, + 0.0028897367883473635, + 0.003244842169806361, + 0.003326722653582692, + 0.0006086748326197267, + 0.0005418525543063879, + 0.0008850349113345146, + 0.0005273870774544775, + 0.0005039545358158648, + 0.0005859981756657362, + 0.00047159005771391094, + 0.0006362596177496016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0209150779992342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026388922706246376 + ], + "opposite_logprob_tokens": [ + -3.6513888835906982 + ], + "logprob_reward": -0.026388922706246376, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026388922706246376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5106642842292786, + 0.1447739452123642, + 0.05954613536596298, + 0.04219162464141846, + 0.07309306412935257, + 0.035181161016225815, + 0.05517475679516792, + 0.037621039897203445, + 0.033054180443286896, + 0.008699815720319748 + ], + "credit_map": [ + 0.06862031668424606, + 0.11794468015432358, + 0.04362459108233452, + 0.042421456426382065, + 0.07537149637937546, + 0.046345796436071396, + 0.03626430034637451, + 0.08007160574197769, + 0.019902346655726433, + 0.014516450464725494, + 0.0094714080914855, + 0.013685574755072594, + 0.015565208159387112, + 0.03617190197110176, + 0.01207891758531332, + 0.023382127285003662, + 0.008278806693851948, + 0.009739595465362072, + 0.005783445667475462, + 0.005081248935312033, + 0.005029161460697651, + 0.010154343210160732, + 0.0067717949859797955, + 0.00870773196220398, + 0.005804288666695356, + 0.005569037981331348, + 0.004030137322843075, + 0.004349066875874996, + 0.003619280643761158, + 0.006557800341397524, + 0.005620582029223442, + 0.006641428917646408, + 0.012975234538316727, + 0.008237645961344242, + 0.009910479187965393, + 0.006684558466076851, + 0.005125603172928095, + 0.010261911898851395, + 0.007673634681850672, + 0.012223992496728897, + 0.004969223868101835, + 0.004266056697815657, + 0.005608590319752693, + 0.004234632942825556, + 0.0031026096548885107, + 0.004667485132813454, + 0.004874518606811762, + 0.0034580391366034746, + 0.012298725545406342, + 0.00527532771229744, + 0.004788446240127087, + 0.004364212043583393, + 0.005753096658736467, + 0.008319026790559292, + 0.0056463382206857204, + 0.008729577995836735, + 0.0034336340613663197, + 0.004168751183897257, + 0.003566273022443056, + 0.0036979783326387405, + 0.0024222314823418856, + 0.006472958251833916, + 0.007144855335354805, + 0.0067143552005290985, + 0.0025550518184900284, + 0.0025396982673555613, + 0.0020097626838833094, + 0.0022407544311136007, + 0.0027505136094987392, + 0.005107892211526632, + 0.006199950352311134, + 0.00965055450797081, + 0.0007914079469628632, + 0.001375588239170611, + 0.0008594910032115877, + 0.0009766569128260016, + 0.0011931051267310977, + 0.0011279340833425522, + 0.0011118261609226465, + 0.0012638060143217444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.026388922706246376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003628220409154892 + ], + "opposite_logprob_tokens": [ + -5.628628253936768 + ], + "logprob_reward": -0.003628220409154892, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003628220409154892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30281782150268555, + 0.12523697316646576, + 0.12475928664207458, + 0.14985401928424835, + 0.05451542139053345, + 0.07381385564804077, + 0.04912466183304787, + 0.04159572720527649, + 0.04557625576853752, + 0.032705970108509064 + ], + "credit_map": [ + 0.04474049434065819, + 0.05161541327834129, + 0.021114399656653404, + 0.021527251228690147, + 0.038253359496593475, + 0.031112972646951675, + 0.022121161222457886, + 0.07233278453350067, + 0.016638904809951782, + 0.01581457257270813, + 0.0075155096128582954, + 0.009281887672841549, + 0.009629026055335999, + 0.03769250214099884, + 0.011227124370634556, + 0.017437448725104332, + 0.013963587582111359, + 0.024020060896873474, + 0.010203319601714611, + 0.01098232064396143, + 0.007182760164141655, + 0.015879573300480843, + 0.0148633336648345, + 0.02766435034573078, + 0.016644541174173355, + 0.017290780320763588, + 0.015124339610338211, + 0.016652856022119522, + 0.010290376842021942, + 0.026988795027136803, + 0.02253815531730652, + 0.024324165657162666, + 0.010948955081403255, + 0.005954510532319546, + 0.005700917914509773, + 0.005847900174558163, + 0.0034599080681800842, + 0.00701482780277729, + 0.00574573315680027, + 0.009842677041888237, + 0.008876041509211063, + 0.00833701342344284, + 0.008968234993517399, + 0.008460016921162605, + 0.006083001382648945, + 0.009520330466330051, + 0.011061971075832844, + 0.012507252395153046, + 0.007047396618872881, + 0.0055147381499409676, + 0.0054939668625593185, + 0.004008802585303783, + 0.00374882435426116, + 0.00831972062587738, + 0.006539377849549055, + 0.008451838046312332, + 0.005923238582909107, + 0.004214473068714142, + 0.0030486576724797487, + 0.0028155038598924875, + 0.0024848978500813246, + 0.0070933932438492775, + 0.004381502512842417, + 0.011634059250354767, + 0.0058722952380776405, + 0.003468676470220089, + 0.0034374778624624014, + 0.004132441710680723, + 0.004719165153801441, + 0.006301376037299633, + 0.00577295757830143, + 0.011871867813169956, + 0.0038654112722724676, + 0.004806701093912125, + 0.0046799443662166595, + 0.004241029731929302, + 0.0034435628913342953, + 0.003709356766194105, + 0.004189041443169117, + 0.003770922077819705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.003628220409154892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032962262630463 + ], + "opposite_logprob_tokens": [ + -2.350329637527466 + ], + "logprob_reward": -0.10032962262630463, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032962262630463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2362706959247589, + 0.17446403205394745, + 0.3272593319416046, + 0.057418059557676315, + 0.0257823895663023, + 0.046905726194381714, + 0.06849969923496246, + 0.033169735223054886, + 0.02479621022939682, + 0.005434129852801561 + ], + "credit_map": [ + 0.025976376608014107, + 0.036844782531261444, + 0.014216133393347263, + 0.019917910918593407, + 0.07150977849960327, + 0.019002221524715424, + 0.022173551842570305, + 0.02662995271384716, + 0.022523101419210434, + 0.017954494804143906, + 0.011298972181975842, + 0.017963234335184097, + 0.02343757636845112, + 0.03339533507823944, + 0.019251050427556038, + 0.028640277683734894, + 0.03603063151240349, + 0.05739431455731392, + 0.018817748874425888, + 0.030617868527770042, + 0.028914615511894226, + 0.0531083382666111, + 0.03490364924073219, + 0.0674721747636795, + 0.007639868650585413, + 0.006803578231483698, + 0.0039003235287964344, + 0.005008871667087078, + 0.004161039832979441, + 0.0101421307772398, + 0.007975779473781586, + 0.01178647018969059, + 0.0023039409425109625, + 0.0033152964897453785, + 0.0023171768989413977, + 0.00286844908259809, + 0.00314799090847373, + 0.005298094358295202, + 0.0032369205728173256, + 0.0032945217099040747, + 0.005485839676111937, + 0.005837493576109409, + 0.0060534048825502396, + 0.006812346167862415, + 0.005503423977643251, + 0.0052684517577290535, + 0.006592727266252041, + 0.00535203842446208, + 0.008066157810389996, + 0.00631756242364645, + 0.006440147291868925, + 0.0073476326651871204, + 0.006961903069168329, + 0.015497853048145771, + 0.0086745023727417, + 0.009193941950798035, + 0.003592073218896985, + 0.0032851712312549353, + 0.0028147962875664234, + 0.0034353670198470354, + 0.002770337974652648, + 0.006599918939173222, + 0.004932889714837074, + 0.005739184096455574, + 0.004105838946998119, + 0.001957397675141692, + 0.00200263992883265, + 0.0022188767325133085, + 0.0028094190638512373, + 0.0036440626718103886, + 0.00266479910351336, + 0.005393178202211857, + 0.0007411575643345714, + 0.0006195169407874346, + 0.000878930208273232, + 0.0006311891484074295, + 0.0007004047511145473, + 0.0006387570756487548, + 0.0005813250900246203, + 0.0006428495398722589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.10032962262630463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot focused on a corner of an antique dark-wood bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick ancient tome with a dark brown hard leather cover, frayed edges, and faded gilt patterns visible on the central shelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a warm concentrated beam of light slanting down from the right side of the frame, illuminating dust particles and casting deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the book's yellowed rough-edged parchment pages turn one by one in a steady leisurely rhythm without external force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle level and does the scene remain static without camera movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058464288711548 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.02058464288711548, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058464288711548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4736314117908478, + 0.1053367629647255, + 0.05655451864004135, + 0.0749361589550972, + 0.059358153492212296, + 0.02750185690820217, + 0.04198811203241348, + 0.0916176289319992, + 0.04862833395600319, + 0.020447151735424995 + ], + "credit_map": [ + 0.06826275587081909, + 0.06900221109390259, + 0.024050777778029442, + 0.03239273279905319, + 0.15389490127563477, + 0.043709468096494675, + 0.03379100561141968, + 0.0485275499522686, + 0.01300755888223648, + 0.014648506417870522, + 0.013916733674705029, + 0.014949941076338291, + 0.007334847468882799, + 0.015011471696197987, + 0.009059945121407509, + 0.017407748848199844, + 0.00872314814478159, + 0.010558015666902065, + 0.004098464269191027, + 0.005298037081956863, + 0.0031881798058748245, + 0.007533628959208727, + 0.006257655564695597, + 0.010897384956479073, + 0.00778789771720767, + 0.008453496731817722, + 0.009496107697486877, + 0.008319650776684284, + 0.00567908538505435, + 0.012365050613880157, + 0.008862549439072609, + 0.013972310349345207, + 0.0114855682477355, + 0.00706233736127615, + 0.005426866002380848, + 0.004863362293690443, + 0.006017613224685192, + 0.011486813426017761, + 0.005580314435064793, + 0.007435267791152, + 0.0032761688344180584, + 0.002916329074651003, + 0.0028774868696928024, + 0.0038725188933312893, + 0.0025062114000320435, + 0.003169433679431677, + 0.0034798861015588045, + 0.005403819959610701, + 0.006091771647334099, + 0.003908522427082062, + 0.004725960083305836, + 0.0036341422237455845, + 0.003925221040844917, + 0.00666988268494606, + 0.004929168149828911, + 0.008103442378342152, + 0.013115176931023598, + 0.007930717431008816, + 0.007480347529053688, + 0.009607505984604359, + 0.004492641892284155, + 0.013773752376437187, + 0.012218121439218521, + 0.022999359294772148, + 0.005712671671062708, + 0.0033996200654655695, + 0.0032710025552660227, + 0.003950322978198528, + 0.004324507433921099, + 0.0077278828248381615, + 0.005759828258305788, + 0.014482488855719566, + 0.0028068632818758488, + 0.001984706614166498, + 0.0037828099448233843, + 0.0019608871079981327, + 0.002248456934466958, + 0.0025353156961500645, + 0.002205935539677739, + 0.002922175917774439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02058464288711548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_tensor([6496, 6496])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_tensor([6496, 6496])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d9a72083a80ca1df12a13a64a821e24c95fd1343 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_tensor([6496, 6496])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow and graceful shot, beginning from a slightly upward angle, tilts slightly upward and pans to the right, slowly sweeping across an open pavilion imbued with a fantastical oriental aesthetic. The pavilion is constructed of dark polished sandalwood, with arched eaves and carved pillars adorned with lifelike dragon motifs. A middle-aged craftsman in a deep teal linen robe bends over a rustic wooden workbench. His long, jet-black hair is tied back with a simple wooden hairpin, a few strands of silver dangling at his temples, and his brows betray a look of intense concentration. As the camera pans, the focus first falls on his focused profile, then shifts to his right hand, which steadily grips a fine wolf-hair brush, the tip saturated with ink, carefully moving it across a rectangular pillow made of moon-white brocade. Here, the camera slows down to capture a close-up of the pillow's details: With a slight twist of his wrist, the outline of a precipitous mountain gradually emerges, bringing the pine trees and drifting clouds within to life. The still-wet ink creates a delicate sheen on the silky surface. A soft, golden morning light slants in through the openwork window on the pavilion's east side, casting dappled shadows in the air, illuminating the dancing dust particles and rendering the mist-shrouded, suspended mountains and the emerald lotus pond outside the pavilion dreamlike. The entire scene, bathed in warm light, is permeated with an almost sacred sense of tranquility and creative concentration, imbued with the poetic and magical qualities of classical Eastern art.", + "seed": 7877, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_6496_seed_7877_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1011797159910202, + "vlm_logprob": -0.1011797159910202, + "vlm_yes_no_margin": 3.424999952316284, + "vlm_token_credit": -0.1011797159910202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01924731396138668 + ], + "opposite_logprob_tokens": [ + -4.394247531890869 + ], + "logprob_reward": -0.01924731396138668, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01924731396138668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06845390796661377, + 0.11840642988681793, + 0.11219193041324615, + 0.1236509308218956, + 0.08760162442922592, + 0.1604834496974945, + 0.14078807830810547, + 0.04516952484846115, + 0.0698225274682045, + 0.07343167066574097 + ], + "credit_map": [ + 0.012717652134597301, + 0.008079687133431435, + 0.006053516175597906, + 0.004211012274026871, + 0.010191946290433407, + 0.005728798918426037, + 0.008101156912744045, + 0.013370133005082607, + 0.016584767028689384, + 0.017184866592288017, + 0.009167329408228397, + 0.019866636022925377, + 0.02144116349518299, + 0.008887521922588348, + 0.011696048080921173, + 0.0135780805721879, + 0.017284546047449112, + 0.015314054675400257, + 0.013634736649692059, + 0.017817897722125053, + 0.01244425680488348, + 0.010398641228675842, + 0.012938768602907658, + 0.01235901564359665, + 0.01223702821880579, + 0.011660301126539707, + 0.010386588051915169, + 0.017448047176003456, + 0.012456521391868591, + 0.01740204356610775, + 0.02271692454814911, + 0.019343461841344833, + 0.004828579258173704, + 0.026176748797297478, + 0.004744786769151688, + 0.007226159796118736, + 0.022733842954039574, + 0.009255053475499153, + 0.006075919605791569, + 0.006560533307492733, + 0.011135894805192947, + 0.011428321711719036, + 0.013929733075201511, + 0.04525494948029518, + 0.018062513321638107, + 0.029845574870705605, + 0.0147116519510746, + 0.01611480303108692, + 0.009545420296490192, + 0.007693022955209017, + 0.011851269751787186, + 0.03302905708551407, + 0.009234477765858173, + 0.02191670425236225, + 0.008708064444363117, + 0.03881004825234413, + 0.00433700205758214, + 0.007440491113811731, + 0.0035320825409144163, + 0.006502267438918352, + 0.009419548325240612, + 0.006161590572446585, + 0.003127065021544695, + 0.004649474285542965, + 0.008087129332125187, + 0.008295111358165741, + 0.011050918139517307, + 0.0068688117899000645, + 0.011156102642416954, + 0.007972602732479572, + 0.010375483892858028, + 0.006016356870532036, + 0.006381373852491379, + 0.007194069214165211, + 0.010948093608021736, + 0.0193693358451128, + 0.011123700998723507, + 0.006620639003813267, + 0.004548235796391964, + 0.007246214896440506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.01924731396138668 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792976081371307 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792976081371307, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792976081371307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09795940667390823, + 0.12628304958343506, + 0.09151306748390198, + 0.0352148674428463, + 0.16631877422332764, + 0.07532073557376862, + 0.14645391702651978, + 0.04661078378558159, + 0.18253932893276215, + 0.031786080449819565 + ], + "credit_map": [ + 0.01695305109024048, + 0.01204056665301323, + 0.012865900993347168, + 0.006411117501556873, + 0.011109020560979843, + 0.008222133852541447, + 0.01028926856815815, + 0.020068349316716194, + 0.016752691939473152, + 0.013749999925494194, + 0.015983035787940025, + 0.01717366650700569, + 0.018415436148643494, + 0.015655018389225006, + 0.018349293619394302, + 0.01020391471683979, + 0.010854875668883324, + 0.008829046972095966, + 0.013574673794209957, + 0.016030898317694664, + 0.008434917777776718, + 0.009551739320158958, + 0.012334287166595459, + 0.011902634054422379, + 0.0040417881682515144, + 0.004332353826612234, + 0.00346454163081944, + 0.004998012911528349, + 0.003230075351893902, + 0.006009756587445736, + 0.004892384167760611, + 0.004245956428349018, + 0.009341088123619556, + 0.046027231961488724, + 0.007847861386835575, + 0.015802642330527306, + 0.0405258946120739, + 0.016140883788466454, + 0.01734921894967556, + 0.013283955864608288, + 0.006936912424862385, + 0.00606254069134593, + 0.006445285398513079, + 0.019955703988671303, + 0.010721653699874878, + 0.010911942459642887, + 0.00817576888948679, + 0.006110921502113342, + 0.010249508544802666, + 0.008560985326766968, + 0.010614441707730293, + 0.02861792780458927, + 0.012470156885683537, + 0.025189191102981567, + 0.008378568105399609, + 0.042373135685920715, + 0.004320208914577961, + 0.007655169814825058, + 0.0035428174305707216, + 0.005248773377388716, + 0.009567410685122013, + 0.006297852843999863, + 0.004612375050783157, + 0.005366173572838306, + 0.011081499978899956, + 0.015609889291226864, + 0.033624064177274704, + 0.01648125611245632, + 0.012112945318222046, + 0.06245507299900055, + 0.008088628761470318, + 0.02308596484363079, + 0.0030286849942058325, + 0.0046701012179255486, + 0.0038247748743742704, + 0.005272786598652601, + 0.005540614947676659, + 0.004251427482813597, + 0.0018447812180966139, + 0.0033529086504131556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.009792976081371307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981313318014145 + ], + "opposite_logprob_tokens": [ + -3.529813051223755 + ], + "logprob_reward": -0.02981313318014145, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981313318014145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06713731586933136, + 0.09474937617778778, + 0.07817781716585159, + 0.058394767343997955, + 0.1476837694644928, + 0.08367221802473068, + 0.06553930789232254, + 0.05534035712480545, + 0.2596568167209625, + 0.08964824676513672 + ], + "credit_map": [ + 0.014614456333220005, + 0.007341829594224691, + 0.007024016231298447, + 0.005114168860018253, + 0.006880508735775948, + 0.004736158065497875, + 0.006353194359689951, + 0.015072986483573914, + 0.010764749720692635, + 0.009119568392634392, + 0.010422823019325733, + 0.01540655642747879, + 0.016759376972913742, + 0.007336010225117207, + 0.013439581729471684, + 0.011500711552798748, + 0.010071973316371441, + 0.010099767707288265, + 0.008805378340184689, + 0.013838563114404678, + 0.009521091356873512, + 0.006841019261628389, + 0.010967252776026726, + 0.00803276989609003, + 0.007679053582251072, + 0.004157917574048042, + 0.00627369387075305, + 0.007152517791837454, + 0.0041105058044195175, + 0.012983754277229309, + 0.00671246787533164, + 0.009324861690402031, + 0.008452417328953743, + 0.042008452117443085, + 0.0070298342034220695, + 0.013911289162933826, + 0.037788692861795425, + 0.013502251356840134, + 0.013784443959593773, + 0.011206391267478466, + 0.0045904433354735374, + 0.00910705141723156, + 0.007356638088822365, + 0.015694405883550644, + 0.015480958856642246, + 0.009874668903648853, + 0.012768043205142021, + 0.008800005540251732, + 0.005100183188915253, + 0.007200765889137983, + 0.005950444377958775, + 0.008167864754796028, + 0.012914602644741535, + 0.009751779027283192, + 0.00466912193223834, + 0.011784547939896584, + 0.004887656308710575, + 0.007694037165492773, + 0.004882194101810455, + 0.009237729012966156, + 0.007979259826242924, + 0.007556910626590252, + 0.0042989738285541534, + 0.00880359299480915, + 0.012538229115307331, + 0.01818135753273964, + 0.07469165325164795, + 0.03313219174742699, + 0.01305537298321724, + 0.07761881500482559, + 0.010836419649422169, + 0.019602762535214424, + 0.006905198097229004, + 0.009889746084809303, + 0.019438039511442184, + 0.014531174674630165, + 0.014336581341922283, + 0.008675230666995049, + 0.008835776709020138, + 0.007036498282104731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.02981313318014145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288589060306549 + ], + "opposite_logprob_tokens": [ + -1.0538588762283325 + ], + "logprob_reward": -0.4288589060306549, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288589060306549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10926202684640884, + 0.13464657962322235, + 0.12601953744888306, + 0.05473358556628227, + 0.14069385826587677, + 0.16378161311149597, + 0.11453328281641006, + 0.06223393604159355, + 0.08362939953804016, + 0.010466191917657852 + ], + "credit_map": [ + 0.01870626211166382, + 0.015683183446526527, + 0.013849643059074879, + 0.010144948028028011, + 0.011242246255278587, + 0.01067651342600584, + 0.009436978958547115, + 0.019522257149219513, + 0.022254714742302895, + 0.022309552878141403, + 0.01245072577148676, + 0.014620526693761349, + 0.03213363513350487, + 0.008003097958862782, + 0.013150863349437714, + 0.009723451919853687, + 0.016928084194660187, + 0.014002220705151558, + 0.0186784528195858, + 0.021859657019376755, + 0.012296848930418491, + 0.0119132399559021, + 0.01580956019461155, + 0.014531474560499191, + 0.006327989511191845, + 0.00507686473429203, + 0.006104184314608574, + 0.009237109683454037, + 0.006021848414093256, + 0.0057936436496675014, + 0.00866785366088152, + 0.007504092063754797, + 0.0071764299646019936, + 0.037839312106370926, + 0.006578786764293909, + 0.013985702767968178, + 0.03242858126759529, + 0.015173915773630142, + 0.016078554093837738, + 0.011432589031755924, + 0.013891594484448433, + 0.009287774562835693, + 0.014729450456798077, + 0.05185551941394806, + 0.014449301175773144, + 0.03223679959774017, + 0.014284921810030937, + 0.013046245090663433, + 0.009476819075644016, + 0.008665061555802822, + 0.011446772143244743, + 0.026077842339873314, + 0.008827652782201767, + 0.016101034358143806, + 0.007461555767804384, + 0.026476550847291946, + 0.007509918417781591, + 0.011726027354598045, + 0.0036763795651495457, + 0.006387683097273111, + 0.014142947271466255, + 0.008653671480715275, + 0.00552779994904995, + 0.004609506577253342, + 0.007038119249045849, + 0.009312322363257408, + 0.017515266314148903, + 0.007918394170701504, + 0.009635373018682003, + 0.01834515482187271, + 0.006441065575927496, + 0.00742371054366231, + 0.0008588358759880066, + 0.0011355194728821516, + 0.002176914131268859, + 0.001961132511496544, + 0.0009673022432252765, + 0.001090632169507444, + 0.001006786827929318, + 0.0012690686853602529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.4288589060306549 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818625070154667 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.01818625070154667, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818625070154667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26792657375335693, + 0.13906244933605194, + 0.10108742862939835, + 0.09700823575258255, + 0.04581966996192932, + 0.04923354461789131, + 0.05643770843744278, + 0.053219329565763474, + 0.12662962079048157, + 0.06357544660568237 + ], + "credit_map": [ + 0.06225460022687912, + 0.037767428904771805, + 0.017311088740825653, + 0.015746308490633965, + 0.023740625008940697, + 0.022305365651845932, + 0.01711229421198368, + 0.07168886065483093, + 0.02141292206943035, + 0.023603739216923714, + 0.013686676509678364, + 0.014339907094836235, + 0.014641810208559036, + 0.012265917845070362, + 0.023103537037968636, + 0.016007941216230392, + 0.01761062629520893, + 0.014458958990871906, + 0.01269484031945467, + 0.015457016415894032, + 0.011332587338984013, + 0.007935418747365475, + 0.012972078286111355, + 0.008625898510217667, + 0.00986091885715723, + 0.007856262847781181, + 0.008300990797579288, + 0.01364125031977892, + 0.008479315787553787, + 0.018126249313354492, + 0.01441400870680809, + 0.016329241916537285, + 0.0027875297237187624, + 0.010310192592442036, + 0.0037485440261662006, + 0.004374663811177015, + 0.008802943862974644, + 0.004640275612473488, + 0.006304079666733742, + 0.004851444158703089, + 0.00427992083132267, + 0.004807200748473406, + 0.004880887921899557, + 0.009156394749879837, + 0.0067186118103563786, + 0.004758724011480808, + 0.006773580331355333, + 0.00785822793841362, + 0.006852275226265192, + 0.007246068213135004, + 0.00548630440607667, + 0.0061759548261761665, + 0.006876836065202951, + 0.008705159649252892, + 0.004926680121570826, + 0.01016843132674694, + 0.005041694268584251, + 0.008809435181319714, + 0.004795771557837725, + 0.00802549161016941, + 0.0059552318416535854, + 0.008645488880574703, + 0.0038902279920876026, + 0.008055989630520344, + 0.008341371081769466, + 0.009941636584699154, + 0.029302984476089478, + 0.013478657230734825, + 0.010694555006921291, + 0.03439094126224518, + 0.010352418757975101, + 0.010127048008143902, + 0.0053609623573720455, + 0.005870176944881678, + 0.008453757502138615, + 0.015216769650578499, + 0.00987718440592289, + 0.006469594780355692, + 0.0034166094847023487, + 0.008910391479730606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01818625070154667 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful shot, beginning from a slightly upward angle, tilts slightly upward and pans to the right, slowly sweeping across an open pavilion imbued with a fantastical oriental aesthetic. The pavilion is constructed of dark polished sandalwood, with arched eaves and carved pillars adorned with lifelike dragon motifs. A middle-aged craftsman in a deep teal linen robe bends over a rustic wooden workbench. His long, jet-black hair is tied back with a simple wooden hairpin, a few strands of silver dangling at his temples, and his brows betray a look of intense concentration. As the camera pans, the focus first falls on his focused profile, then shifts to his right hand, which steadily grips a fine wolf-hair brush, the tip saturated with ink, carefully moving it across a rectangular pillow made of moon-white brocade. Here, the camera slows down to capture a close-up of the pillow's details: With a slight twist of his wrist, the outline of a precipitous mountain gradually emerges, bringing the pine trees and drifting clouds within to life. The still-wet ink creates a delicate sheen on the silky surface. A soft, golden morning light slants in through the openwork window on the pavilion's east side, casting dappled shadows in the air, illuminating the dancing dust particles and rendering the mist-shrouded, suspended mountains and the emerald lotus pond outside the pavilion dreamlike. The entire scene, bathed in warm light, is permeated with an almost sacred sense of tranquility and creative concentration, imbued with the poetic and magical qualities of classical Eastern art.", + "seed": 7878, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_6496_seed_7878_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02765864273533225, + "vlm_logprob": -0.02765864273533225, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02765864273533225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033902086317539215 + ], + "opposite_logprob_tokens": [ + -3.658902168273926 + ], + "logprob_reward": -0.033902086317539215, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.033902086317539215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4886016845703125, + 0.05234530568122864, + 0.03312775492668152, + 0.03761047124862671, + 0.0674019530415535, + 0.1269819140434265, + 0.07821226119995117, + 0.045537032186985016, + 0.038416605442762375, + 0.03176501765847206 + ], + "credit_map": [ + 0.0636921375989914, + 0.05705147236585617, + 0.08667045831680298, + 0.05034362152218819, + 0.06780160218477249, + 0.06281248480081558, + 0.051756445318460464, + 0.048473481088876724, + 0.006024952512234449, + 0.009693710133433342, + 0.005056560505181551, + 0.00519151147454977, + 0.006993475370109081, + 0.007036903873085976, + 0.006880143191665411, + 0.005468050483614206, + 0.0026339164469391108, + 0.002131188055500388, + 0.006281890906393528, + 0.003195781959220767, + 0.002956951968371868, + 0.01020984910428524, + 0.0027638261672109365, + 0.002954353578388691, + 0.00570174865424633, + 0.004032201133668423, + 0.006680425256490707, + 0.003402926027774811, + 0.005054691806435585, + 0.003511978778988123, + 0.0037922922056168318, + 0.005434209946542978, + 0.008934791199862957, + 0.005522484425455332, + 0.009044690988957882, + 0.0070740277878940105, + 0.009598782286047935, + 0.00787355937063694, + 0.008069473318755627, + 0.011284156702458858, + 0.01645001396536827, + 0.015823686495423317, + 0.014896439388394356, + 0.015327422879636288, + 0.023899594321846962, + 0.016086174175143242, + 0.013628232292830944, + 0.010870352387428284, + 0.007168470416218042, + 0.007995263673365116, + 0.012242083437740803, + 0.010440906509757042, + 0.014330766163766384, + 0.010971548967063427, + 0.008646764792501926, + 0.0064164698123931885, + 0.006506140809506178, + 0.004364015534520149, + 0.006456403061747551, + 0.00622066343203187, + 0.005445222835987806, + 0.003813444869592786, + 0.008588956668972969, + 0.004142190329730511, + 0.0034241988323628902, + 0.0027771855238825083, + 0.003766819601878524, + 0.004630461800843477, + 0.0031071172561496496, + 0.0035591330379247665, + 0.009097332134842873, + 0.00805436260998249, + 0.003081244183704257, + 0.003343326272442937, + 0.00362576125189662, + 0.004335472360253334, + 0.003260029014199972, + 0.004073698073625565, + 0.006921912543475628, + 0.003123576520010829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.033902086317539215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059562805108726025 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.0059562805108726025, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059562805108726025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3135014772415161, + 0.2331446260213852, + 0.08966749161481857, + 0.06659973412752151, + 0.02900323085486889, + 0.06476887315511703, + 0.04110696539282799, + 0.05723815783858299, + 0.059009309858083725, + 0.045960038900375366 + ], + "credit_map": [ + 0.04291686788201332, + 0.032321542501449585, + 0.05155864730477333, + 0.04211210831999779, + 0.04042920842766762, + 0.03734838590025902, + 0.03443986177444458, + 0.032374896109104156, + 0.035838473588228226, + 0.01929658278822899, + 0.0205617006868124, + 0.02608056552708149, + 0.020575640723109245, + 0.022794967517256737, + 0.01847914606332779, + 0.0695175752043724, + 0.005906608887016773, + 0.005836462136358023, + 0.009701922535896301, + 0.01109134592115879, + 0.011008713394403458, + 0.025546228513121605, + 0.009280321188271046, + 0.011295896954834461, + 0.011400824412703514, + 0.008075561374425888, + 0.011085416190326214, + 0.006963781546801329, + 0.0076729184947907925, + 0.006624216679483652, + 0.007481732405722141, + 0.00729528721421957, + 0.002901425352320075, + 0.0035043356474488974, + 0.003368543926626444, + 0.002945736050605774, + 0.004730996675789356, + 0.0036714058369398117, + 0.0029161046259105206, + 0.004964683670550585, + 0.006679248996078968, + 0.0063811191357672215, + 0.007061608135700226, + 0.008425251580774784, + 0.011298464611172676, + 0.011472884565591812, + 0.006472995970398188, + 0.006977308541536331, + 0.0034604405518621206, + 0.004493115935474634, + 0.0049691335298120975, + 0.00503145158290863, + 0.008865883573889732, + 0.005756174214184284, + 0.0047213248908519745, + 0.003809447633102536, + 0.00904512032866478, + 0.005844299215823412, + 0.006547628436237574, + 0.0068893395364284515, + 0.007485909853130579, + 0.0036446675658226013, + 0.012428686022758484, + 0.0053525157272815704, + 0.007913106121122837, + 0.013313835486769676, + 0.009442940354347229, + 0.0061402153223752975, + 0.004686326254159212, + 0.006435409188270569, + 0.005662870593369007, + 0.005414612125605345, + 0.004795696120709181, + 0.008193142712116241, + 0.005340316332876682, + 0.008954595774412155, + 0.003678663168102503, + 0.003101225709542632, + 0.007184902206063271, + 0.004711500834673643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0059562805108726025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371170163154602 + ], + "opposite_logprob_tokens": [ + -3.4087116718292236 + ], + "logprob_reward": -0.03371170163154602, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371170163154602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13135820627212524, + 0.2775459587574005, + 0.044601429253816605, + 0.07099246233701706, + 0.06405598670244217, + 0.08763601630926132, + 0.06976954638957977, + 0.0786156877875328, + 0.0578220896422863, + 0.11760261654853821 + ], + "credit_map": [ + 0.01627308502793312, + 0.018265487626194954, + 0.01671704277396202, + 0.012404348701238632, + 0.021337635815143585, + 0.01785869337618351, + 0.017515383660793304, + 0.010986534878611565, + 0.04863468557596207, + 0.022932885214686394, + 0.02672826126217842, + 0.018275141716003418, + 0.023414384573698044, + 0.028071310371160507, + 0.019172821193933487, + 0.09031646698713303, + 0.0029012076556682587, + 0.003644568845629692, + 0.004804802592843771, + 0.00510728545486927, + 0.007201665081083775, + 0.010413079522550106, + 0.005904329940676689, + 0.004624487832188606, + 0.007877477444708347, + 0.006875853985548019, + 0.00871539581567049, + 0.006390464957803488, + 0.010499713011085987, + 0.007744086440652609, + 0.006484472192823887, + 0.016405001282691956, + 0.008166501298546791, + 0.00751537224277854, + 0.007308684755116701, + 0.007839364930987358, + 0.008612364530563354, + 0.007900012657046318, + 0.006529523525387049, + 0.010184157639741898, + 0.009398255497217178, + 0.010150579735636711, + 0.00891053956001997, + 0.010491813533008099, + 0.015548549592494965, + 0.014629088342189789, + 0.0095160948112607, + 0.008991092443466187, + 0.006615373305976391, + 0.004990580957382917, + 0.006778748240321875, + 0.005930541083216667, + 0.0197425689548254, + 0.011110223829746246, + 0.006289344746619463, + 0.008312166668474674, + 0.01171849761158228, + 0.009416230954229832, + 0.010663709603250027, + 0.011684239841997623, + 0.008519968949258327, + 0.004936710465699434, + 0.013533225283026695, + 0.008143104612827301, + 0.00802975706756115, + 0.011383107863366604, + 0.009091111831367016, + 0.006770568434149027, + 0.00484802108258009, + 0.006426176521927118, + 0.0063532497733831406, + 0.004920096602290869, + 0.01091855764389038, + 0.016672693192958832, + 0.014321905560791492, + 0.017924470826983452, + 0.010924765840172768, + 0.008332595229148865, + 0.026843169704079628, + 0.011664454825222492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.03371170163154602 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493996664881706 + ], + "opposite_logprob_tokens": [ + -2.9299399852752686 + ], + "logprob_reward": -0.05493996664881706, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493996664881706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1356615126132965, + 0.3492079973220825, + 0.08337190002202988, + 0.1542118340730667, + 0.06923485547304153, + 0.07642373442649841, + 0.04523801803588867, + 0.03524213656783104, + 0.03683291748166084, + 0.01457507535815239 + ], + "credit_map": [ + 0.017426323145627975, + 0.01701805181801319, + 0.01996336691081524, + 0.014885932207107544, + 0.018761269748210907, + 0.020687425509095192, + 0.015228254720568657, + 0.011690887622535229, + 0.0613282136619091, + 0.027512868866324425, + 0.03494171425700188, + 0.03042670153081417, + 0.030153077095746994, + 0.031371645629405975, + 0.025691017508506775, + 0.1077827587723732, + 0.007856406271457672, + 0.0052779195830225945, + 0.018331708386540413, + 0.0070103988982737064, + 0.0078279385343194, + 0.023332538083195686, + 0.006568029057234526, + 0.007166959811002016, + 0.029334284365177155, + 0.01531943492591381, + 0.033533014357089996, + 0.017858752980828285, + 0.013167453929781914, + 0.014547322876751423, + 0.011196752078831196, + 0.019254816696047783, + 0.008315016515552998, + 0.008213195018470287, + 0.007771797012537718, + 0.008641690947115421, + 0.010587861761450768, + 0.009321690537035465, + 0.007151832338422537, + 0.009231766685843468, + 0.007150520570576191, + 0.006346289999783039, + 0.006495405454188585, + 0.008901729248464108, + 0.014202399179339409, + 0.011754610575735569, + 0.009090298786759377, + 0.012482480145990849, + 0.004868284799158573, + 0.0032839039340615273, + 0.00603210274130106, + 0.004946214612573385, + 0.00928674265742302, + 0.005474322009831667, + 0.006034059915691614, + 0.005312387831509113, + 0.0035172796342521906, + 0.003960794769227505, + 0.005045897327363491, + 0.004036708269268274, + 0.004679958801716566, + 0.0037284677382558584, + 0.006211721803992987, + 0.004061306826770306, + 0.00395641103386879, + 0.0025363005697727203, + 0.00357859511859715, + 0.0029476797208189964, + 0.003921446390450001, + 0.005166278686374426, + 0.007428233977407217, + 0.007297974545508623, + 0.00117991934530437, + 0.0019351186929270625, + 0.00208657537586987, + 0.0021588278468698263, + 0.0015888884663581848, + 0.0015338328666985035, + 0.0021945680491626263, + 0.0018973443657159805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.05493996664881706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783178567886353 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783178567886353, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783178567886353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2060822695493698, + 0.239160418510437, + 0.06578756868839264, + 0.17557267844676971, + 0.06405630707740784, + 0.06964852660894394, + 0.04885612055659294, + 0.022761782631278038, + 0.06478839367628098, + 0.043285880237817764 + ], + "credit_map": [ + 0.03063177317380905, + 0.02400372177362442, + 0.034098971635103226, + 0.019147440791130066, + 0.025208789855241776, + 0.02422245591878891, + 0.02843249775469303, + 0.020336633548140526, + 0.040453653782606125, + 0.019924309104681015, + 0.023288767784833908, + 0.017135456204414368, + 0.023155497387051582, + 0.023708822205662727, + 0.018758142367005348, + 0.07273577898740768, + 0.008053482510149479, + 0.004342221189290285, + 0.014842928387224674, + 0.005726804956793785, + 0.006074549164623022, + 0.012862135656177998, + 0.005398854613304138, + 0.008486590348184109, + 0.02924931049346924, + 0.01900973729789257, + 0.03053230233490467, + 0.015895487740635872, + 0.028405657038092613, + 0.017168520018458366, + 0.017293745651841164, + 0.018017930909991264, + 0.006347754970192909, + 0.004669283516705036, + 0.007113135885447264, + 0.006725840270519257, + 0.014096151106059551, + 0.00933821965008974, + 0.008083345368504524, + 0.00768257537856698, + 0.006705272011458874, + 0.010544809512794018, + 0.006182968150824308, + 0.008589561097323895, + 0.011511114425957203, + 0.00900778267532587, + 0.008230676874518394, + 0.00887633953243494, + 0.005895029753446579, + 0.004599541891366243, + 0.004329549614340067, + 0.0041430192068219185, + 0.012135895900428295, + 0.006563497707247734, + 0.005570658016949892, + 0.005618925206363201, + 0.0024511166848242283, + 0.0019811626989394426, + 0.0022112135775387287, + 0.0027275467291474342, + 0.004075323697179556, + 0.004092388786375523, + 0.002516453620046377, + 0.002706576604396105, + 0.004593213088810444, + 0.007461965084075928, + 0.007912535220384598, + 0.0068604773841798306, + 0.007596006616950035, + 0.009716294705867767, + 0.010911324992775917, + 0.009736575186252594, + 0.005197616294026375, + 0.007600564509630203, + 0.004702960606664419, + 0.005613469518721104, + 0.0041003040969371796, + 0.003990902099758387, + 0.007425627671182156, + 0.004654430318623781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.009783178567886353 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful shot, beginning from a slightly upward angle, tilts slightly upward and pans to the right, slowly sweeping across an open pavilion imbued with a fantastical oriental aesthetic. The pavilion is constructed of dark polished sandalwood, with arched eaves and carved pillars adorned with lifelike dragon motifs. A middle-aged craftsman in a deep teal linen robe bends over a rustic wooden workbench. His long, jet-black hair is tied back with a simple wooden hairpin, a few strands of silver dangling at his temples, and his brows betray a look of intense concentration. As the camera pans, the focus first falls on his focused profile, then shifts to his right hand, which steadily grips a fine wolf-hair brush, the tip saturated with ink, carefully moving it across a rectangular pillow made of moon-white brocade. Here, the camera slows down to capture a close-up of the pillow's details: With a slight twist of his wrist, the outline of a precipitous mountain gradually emerges, bringing the pine trees and drifting clouds within to life. The still-wet ink creates a delicate sheen on the silky surface. A soft, golden morning light slants in through the openwork window on the pavilion's east side, casting dappled shadows in the air, illuminating the dancing dust particles and rendering the mist-shrouded, suspended mountains and the emerald lotus pond outside the pavilion dreamlike. The entire scene, bathed in warm light, is permeated with an almost sacred sense of tranquility and creative concentration, imbued with the poetic and magical qualities of classical Eastern art.", + "seed": 7879, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_6496_seed_7879_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0323649825528264, + "vlm_logprob": -0.0323649825528264, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0323649825528264, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021998438984155655 + ], + "opposite_logprob_tokens": [ + -4.146998405456543 + ], + "logprob_reward": -0.021998438984155655, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.021998438984155655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12124679982662201, + 0.03900860622525215, + 0.034099481999874115, + 0.10743322968482971, + 0.07734918594360352, + 0.24441029131412506, + 0.10212977975606918, + 0.07173953205347061, + 0.08089761435985565, + 0.1216854602098465 + ], + "credit_map": [ + 0.01063678227365017, + 0.026435302570462227, + 0.00822845846414566, + 0.014080185443162918, + 0.008128284476697445, + 0.010261599905788898, + 0.012412290088832378, + 0.031063897535204887, + 0.00349222170189023, + 0.002541041700169444, + 0.003985714167356491, + 0.005741840694099665, + 0.006423673592507839, + 0.003635641885921359, + 0.007926193997263908, + 0.005262278951704502, + 0.00333373318426311, + 0.002741573378443718, + 0.00455454783514142, + 0.005904329475015402, + 0.005602714139968157, + 0.0032906020060181618, + 0.0035843919031322002, + 0.005087588913738728, + 0.00933155883103609, + 0.01693934202194214, + 0.01484055444598198, + 0.026439404115080833, + 0.00958858523517847, + 0.00764834601432085, + 0.014245389960706234, + 0.008400055579841137, + 0.01035128254443407, + 0.01085034292191267, + 0.012888279743492603, + 0.009204622358083725, + 0.010842317715287209, + 0.005407765507698059, + 0.010069765150547028, + 0.007734808139503002, + 0.021303201094269753, + 0.031114008277654648, + 0.03823151811957359, + 0.047244053333997726, + 0.035076823085546494, + 0.0189548097550869, + 0.028598936274647713, + 0.023886920884251595, + 0.010278988629579544, + 0.016277125105261803, + 0.011759376153349876, + 0.014342514798045158, + 0.020259931683540344, + 0.008447576314210892, + 0.010613083839416504, + 0.010151187889277935, + 0.007992901839315891, + 0.009225734509527683, + 0.01091584749519825, + 0.008854436688125134, + 0.009182038716971874, + 0.011326844803988934, + 0.007148691453039646, + 0.0070930360816419125, + 0.010001071728765965, + 0.010218671523034573, + 0.01257466059178114, + 0.013529660180211067, + 0.0063134110532701015, + 0.008299645967781544, + 0.008767847903072834, + 0.011192641220986843, + 0.010775175876915455, + 0.01338571310043335, + 0.016548974439501762, + 0.02459516003727913, + 0.011095786467194557, + 0.010527394711971283, + 0.017714954912662506, + 0.017042292281985283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.021998438984155655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084291152656078 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011084291152656078, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084291152656078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0745266005396843, + 0.05446997284889221, + 0.07463537901639938, + 0.0437716580927372, + 0.08456215262413025, + 0.15928058326244354, + 0.09502913802862167, + 0.10629373788833618, + 0.0660979151725769, + 0.24133293330669403 + ], + "credit_map": [ + 0.004451965447515249, + 0.015058868564665318, + 0.005154918413609266, + 0.010305719450116158, + 0.007586305029690266, + 0.0051368167623877525, + 0.009113054722547531, + 0.017718950286507607, + 0.00494809914380312, + 0.003967163618654013, + 0.005271794740110636, + 0.0077132247388362885, + 0.008450237102806568, + 0.0053070723079144955, + 0.012080146931111813, + 0.006732237990945578, + 0.006007264833897352, + 0.0054116458632051945, + 0.005241538397967815, + 0.007656443398445845, + 0.017392892390489578, + 0.0078100115060806274, + 0.013881788589060307, + 0.011233792640268803, + 0.005044642370194197, + 0.0055683208629488945, + 0.007441035471856594, + 0.006183447781950235, + 0.005582510028034449, + 0.004660519305616617, + 0.004291999619454145, + 0.004999180790036917, + 0.008078954182565212, + 0.009835314936935902, + 0.01104158628731966, + 0.009687447920441628, + 0.009431281127035618, + 0.010529604740440845, + 0.013052314519882202, + 0.012905648909509182, + 0.012902887538075447, + 0.016419604420661926, + 0.03230147808790207, + 0.01917390525341034, + 0.022237248718738556, + 0.02022080309689045, + 0.016612136736512184, + 0.019412506371736526, + 0.005977371707558632, + 0.01175490953028202, + 0.01479532290250063, + 0.011593609116971493, + 0.017092691734433174, + 0.013768277131021023, + 0.010875585488975048, + 0.009171368554234505, + 0.011498408392071724, + 0.012968580238521099, + 0.01457058172672987, + 0.015031307935714722, + 0.014428284019231796, + 0.009885862469673157, + 0.016599327325820923, + 0.011311382986605167, + 0.008074112236499786, + 0.008876562118530273, + 0.006753017194569111, + 0.010220782831311226, + 0.008217980153858662, + 0.006033120211213827, + 0.007403930649161339, + 0.010518410243093967, + 0.016491733491420746, + 0.022973136976361275, + 0.035094164311885834, + 0.04651254788041115, + 0.022371752187609673, + 0.014616063795983791, + 0.0552976094186306, + 0.02797590382397175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.011084291152656078 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370685875415802 + ], + "opposite_logprob_tokens": [ + -3.4087069034576416 + ], + "logprob_reward": -0.03370685875415802, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370685875415802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04960261285305023, + 0.01389259472489357, + 0.02269645594060421, + 0.033597495406866074, + 0.055784922093153, + 0.09027926623821259, + 0.0613790862262249, + 0.08332346379756927, + 0.08605322986841202, + 0.5033907890319824 + ], + "credit_map": [ + 0.003956938162446022, + 0.01139879785478115, + 0.0036282672081142664, + 0.0056571331806480885, + 0.0035145683214068413, + 0.003407236421480775, + 0.006877007894217968, + 0.011162670329213142, + 0.0019210493192076683, + 0.001193869044072926, + 0.001414819504134357, + 0.0018450494389981031, + 0.0019211386097595096, + 0.0012681197840720415, + 0.0025247831363230944, + 0.0018037670524790883, + 0.0027443096041679382, + 0.0014064583228901029, + 0.001944175222888589, + 0.002816089428961277, + 0.004769706167280674, + 0.0028300804551690817, + 0.0026434974279254675, + 0.003542142454534769, + 0.0036292204167693853, + 0.0050461068749427795, + 0.0029963459819555283, + 0.006791731808334589, + 0.004156762268394232, + 0.003132111858576536, + 0.005054631270468235, + 0.0027905863244086504, + 0.006122850347310305, + 0.00838507991284132, + 0.006279943510890007, + 0.006564596202224493, + 0.008489253930747509, + 0.005780805367976427, + 0.007400927133858204, + 0.00676147174090147, + 0.00882650911808014, + 0.010306846350431442, + 0.01680140756070614, + 0.016036756336688995, + 0.01302340067923069, + 0.008513069711625576, + 0.008675428107380867, + 0.008095857687294483, + 0.005567737855017185, + 0.009958761744201183, + 0.007336594630032778, + 0.006528797559440136, + 0.012951563112437725, + 0.006146957166492939, + 0.00838075764477253, + 0.004507920704782009, + 0.008003131486475468, + 0.012690100818872452, + 0.011963879689574242, + 0.011494932696223259, + 0.011904134415090084, + 0.008708635345101357, + 0.0115675562992692, + 0.006991101894527674, + 0.00873999297618866, + 0.012910246849060059, + 0.00935253594070673, + 0.013182886876165867, + 0.009078974835574627, + 0.008130398578941822, + 0.010015946812927723, + 0.014642257243394852, + 0.032673608511686325, + 0.046164821833372116, + 0.07930373400449753, + 0.11077331006526947, + 0.04337022453546524, + 0.02334441803395748, + 0.12466882169246674, + 0.043091874569654465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.03370685875415802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896719127893448 + ], + "opposite_logprob_tokens": [ + -2.5789670944213867 + ], + "logprob_reward": -0.07896719127893448, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896719127893448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11448554694652557, + 0.036358751356601715, + 0.1252257525920868, + 0.09557990729808807, + 0.13532976806163788, + 0.10744773596525192, + 0.08054128289222717, + 0.11127842962741852, + 0.059952061623334885, + 0.13380077481269836 + ], + "credit_map": [ + 0.009650757536292076, + 0.02464846894145012, + 0.00777812022715807, + 0.013595274649560452, + 0.01383658405393362, + 0.008942493237555027, + 0.017243247479200363, + 0.01879059337079525, + 0.0036610665265470743, + 0.002396194264292717, + 0.004511111881583929, + 0.005595657974481583, + 0.005578724201768637, + 0.0032044847030192614, + 0.0069156247191131115, + 0.0044958824291825294, + 0.009989013895392418, + 0.009536916390061378, + 0.008417747914791107, + 0.015873074531555176, + 0.02976255863904953, + 0.012875745072960854, + 0.01660492643713951, + 0.022165751084685326, + 0.007822688668966293, + 0.013604925945401192, + 0.013808193616569042, + 0.018839044496417046, + 0.009118294343352318, + 0.010327599011361599, + 0.0115004638209939, + 0.010558699257671833, + 0.016471417620778084, + 0.01256866380572319, + 0.017766550183296204, + 0.0147326048463583, + 0.013398093171417713, + 0.016724659129977226, + 0.024164846166968346, + 0.019502924755215645, + 0.009715372696518898, + 0.011331495828926563, + 0.022065909579396248, + 0.010728567838668823, + 0.012018577195703983, + 0.010740063153207302, + 0.01627798192203045, + 0.014569750986993313, + 0.009615062735974789, + 0.01049876306205988, + 0.011207066476345062, + 0.010588385164737701, + 0.009656155481934547, + 0.009338969364762306, + 0.010726536624133587, + 0.008910342119634151, + 0.008919292129576206, + 0.013173900544643402, + 0.020327745005488396, + 0.013551882468163967, + 0.015284529887139797, + 0.01711922511458397, + 0.011477709747850895, + 0.011424122378230095, + 0.005110583268105984, + 0.008729047141969204, + 0.006492491811513901, + 0.009115726687014103, + 0.006219882983714342, + 0.0054499078541994095, + 0.008637514896690845, + 0.010196899063885212, + 0.012352866120636463, + 0.018161939457058907, + 0.017691336572170258, + 0.02799963392317295, + 0.013333707116544247, + 0.010316037572920322, + 0.018852220848202705, + 0.015093031339347363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.07896719127893448 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606813259422779 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.01606813259422779, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606813259422779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0786832645535469, + 0.027014685794711113, + 0.05124577134847641, + 0.04511449858546257, + 0.05457308515906334, + 0.19015057384967804, + 0.13355526328086853, + 0.10106617212295532, + 0.04849692061543465, + 0.2700996994972229 + ], + "credit_map": [ + 0.00849020667374134, + 0.006618833635002375, + 0.01153766643255949, + 0.009626336395740509, + 0.009008688852190971, + 0.00552171329036355, + 0.016523830592632294, + 0.011355994269251823, + 0.0028380858711898327, + 0.0022885215003043413, + 0.003159538609907031, + 0.0032476691994816065, + 0.0045000649988651276, + 0.0032924998085945845, + 0.0040482813492417336, + 0.003640025621280074, + 0.007114503998309374, + 0.003567951265722513, + 0.005791832692921162, + 0.005754666868597269, + 0.009840519167482853, + 0.006495638284832239, + 0.005360775161534548, + 0.007319883443415165, + 0.004114179871976376, + 0.005850487854331732, + 0.0058912415988743305, + 0.008534234948456287, + 0.0059217228554189205, + 0.004845675081014633, + 0.005094708874821663, + 0.00486224889755249, + 0.006724323611706495, + 0.006883608642965555, + 0.007452582940459251, + 0.008726274594664574, + 0.007031062617897987, + 0.005745034199208021, + 0.005993697326630354, + 0.0060165030881762505, + 0.015796782448887825, + 0.02809233032166958, + 0.029948202893137932, + 0.03429635241627693, + 0.020154492929577827, + 0.01746898889541626, + 0.023080920800566673, + 0.021312512457370758, + 0.00905588362365961, + 0.011747224256396294, + 0.03176098316907883, + 0.023304695263504982, + 0.01721012219786644, + 0.015595040284097195, + 0.013656549155712128, + 0.011224782094359398, + 0.010935819707810879, + 0.013248499482870102, + 0.016028890386223793, + 0.014145900495350361, + 0.013879043981432915, + 0.01650375686585903, + 0.008653123863041401, + 0.007671149913221598, + 0.006334339268505573, + 0.010460438206791878, + 0.004611871670931578, + 0.005870561581104994, + 0.005240443628281355, + 0.0031879746820777655, + 0.006302414927631617, + 0.0064888810738921165, + 0.021033544093370438, + 0.02528030052781105, + 0.042137205600738525, + 0.06069144234061241, + 0.021147742867469788, + 0.01403865311294794, + 0.06379710137844086, + 0.021973730996251106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.01606813259422779 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful shot, beginning from a slightly upward angle, tilts slightly upward and pans to the right, slowly sweeping across an open pavilion imbued with a fantastical oriental aesthetic. The pavilion is constructed of dark polished sandalwood, with arched eaves and carved pillars adorned with lifelike dragon motifs. A middle-aged craftsman in a deep teal linen robe bends over a rustic wooden workbench. His long, jet-black hair is tied back with a simple wooden hairpin, a few strands of silver dangling at his temples, and his brows betray a look of intense concentration. As the camera pans, the focus first falls on his focused profile, then shifts to his right hand, which steadily grips a fine wolf-hair brush, the tip saturated with ink, carefully moving it across a rectangular pillow made of moon-white brocade. Here, the camera slows down to capture a close-up of the pillow's details: With a slight twist of his wrist, the outline of a precipitous mountain gradually emerges, bringing the pine trees and drifting clouds within to life. The still-wet ink creates a delicate sheen on the silky surface. A soft, golden morning light slants in through the openwork window on the pavilion's east side, casting dappled shadows in the air, illuminating the dancing dust particles and rendering the mist-shrouded, suspended mountains and the emerald lotus pond outside the pavilion dreamlike. The entire scene, bathed in warm light, is permeated with an almost sacred sense of tranquility and creative concentration, imbued with the poetic and magical qualities of classical Eastern art.", + "seed": 7880, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_6/index_6496_seed_7880_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10997801637277008, + "vlm_logprob": -0.10997801637277008, + "vlm_yes_no_margin": 3.299999988079071, + "vlm_token_credit": -0.10997801637277008, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016257451847195625 + ], + "opposite_logprob_tokens": [ + -4.891257286071777 + ], + "logprob_reward": -0.016257451847195625, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.016257451847195625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07584050297737122, + 0.08399095386266708, + 0.06126474216580391, + 0.18001317977905273, + 0.03917204961180687, + 0.039578258991241455, + 0.040247123688459396, + 0.05300327017903328, + 0.3949405550956726, + 0.03194940462708473 + ], + "credit_map": [ + 0.006486439611762762, + 0.010412766598165035, + 0.007681594230234623, + 0.008606335148215294, + 0.011624252423644066, + 0.013580158352851868, + 0.00697429059073329, + 0.010474666021764278, + 0.005816001445055008, + 0.009894217364490032, + 0.007073080632835627, + 0.006220379378646612, + 0.019597943872213364, + 0.021005181595683098, + 0.003644813783466816, + 0.010739332064986229, + 0.005132125224918127, + 0.006270331796258688, + 0.004960366990417242, + 0.00842486135661602, + 0.0053887455724179745, + 0.011842460371553898, + 0.008389757946133614, + 0.010856092907488346, + 0.012934772297739983, + 0.018971694633364677, + 0.016041114926338196, + 0.01751156710088253, + 0.024806324392557144, + 0.036992866545915604, + 0.016211295500397682, + 0.03654354438185692, + 0.003188652452081442, + 0.003820949699729681, + 0.003970441874116659, + 0.0032583302818238735, + 0.004460118245333433, + 0.007234978023916483, + 0.004872356075793505, + 0.00836622342467308, + 0.00439269095659256, + 0.004518098663538694, + 0.0053499010391533375, + 0.0047780596651136875, + 0.004237419459968805, + 0.005529353395104408, + 0.004802211187779903, + 0.0059705255553126335, + 0.004117139149457216, + 0.003051798092201352, + 0.0035412972792983055, + 0.003229686291888356, + 0.0054877763614058495, + 0.009856531396508217, + 0.00454094260931015, + 0.006421949714422226, + 0.008064747788012028, + 0.004819477908313274, + 0.004087168723344803, + 0.008088599890470505, + 0.005282651167362928, + 0.008237783797085285, + 0.005338000599294901, + 0.009084840305149555, + 0.057530228048563004, + 0.03082432970404625, + 0.03928207606077194, + 0.045543037354946136, + 0.030740806832909584, + 0.05666416510939598, + 0.0407632477581501, + 0.09359265863895416, + 0.0049109808169305325, + 0.0024848703760653734, + 0.0033650235272943974, + 0.0024309526197612286, + 0.002691368106752634, + 0.005431490018963814, + 0.005963542498648167, + 0.004671176895499229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.016257451847195625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306088760495186 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.04306088760495186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306088760495186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07781378924846649, + 0.08456151932477951, + 0.18249760568141937, + 0.18075966835021973, + 0.10397151857614517, + 0.0689065158367157, + 0.08802619576454163, + 0.09986263513565063, + 0.08116485178470612, + 0.032435689121484756 + ], + "credit_map": [ + 0.006650126539170742, + 0.010682107880711555, + 0.0054504466243088245, + 0.00788077898323536, + 0.009898936375975609, + 0.020662881433963776, + 0.008746565319597721, + 0.007841943763196468, + 0.009239494800567627, + 0.008990785107016563, + 0.007478618063032627, + 0.008991764858365059, + 0.012444915249943733, + 0.014531954191625118, + 0.009844142943620682, + 0.013039840385317802, + 0.015604491345584393, + 0.01821112632751465, + 0.014329185709357262, + 0.029743505641818047, + 0.024026120081543922, + 0.01963566243648529, + 0.033548660576343536, + 0.027398861944675446, + 0.010022505186498165, + 0.018604939803481102, + 0.018487799912691116, + 0.014914385043084621, + 0.026242738589644432, + 0.03736121952533722, + 0.01769721321761608, + 0.03742887079715729, + 0.008594272658228874, + 0.008783228695392609, + 0.008921998552978039, + 0.008260234259068966, + 0.01203449722379446, + 0.020410647615790367, + 0.011635280214250088, + 0.025331363081932068, + 0.007850469090044498, + 0.009168603457510471, + 0.0069190701469779015, + 0.007901889272034168, + 0.009539393708109856, + 0.008081289008259773, + 0.010857260785996914, + 0.008588540367782116, + 0.007347042206674814, + 0.007591343484818935, + 0.008652161806821823, + 0.014045991003513336, + 0.01198046188801527, + 0.017956193536520004, + 0.008366350084543228, + 0.0120866559445858, + 0.011729096993803978, + 0.006887666881084442, + 0.007706431206315756, + 0.017562977969646454, + 0.008167722262442112, + 0.01982796937227249, + 0.008410092443227768, + 0.01957067847251892, + 0.016413280740380287, + 0.004555804654955864, + 0.008390403352677822, + 0.004795523826032877, + 0.006296434905380011, + 0.012683767825365067, + 0.0047839791513979435, + 0.023245666176080704, + 0.004091218113899231, + 0.0032689678482711315, + 0.002802856033667922, + 0.003760463558137417, + 0.0037080158945173025, + 0.005868479609489441, + 0.004025987349450588, + 0.0049096993170678616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.04306088760495186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054966483265161514 + ], + "opposite_logprob_tokens": [ + -2.929966449737549 + ], + "logprob_reward": -0.054966483265161514, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054966483265161514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1793971210718155, + 0.2451087087392807, + 0.09784083068370819, + 0.12511618435382843, + 0.059574395418167114, + 0.047930750995874405, + 0.04147394001483917, + 0.09183261543512344, + 0.06978096067905426, + 0.04194451496005058 + ], + "credit_map": [ + 0.015853535383939743, + 0.035336725413799286, + 0.01093199010938406, + 0.017380837351083755, + 0.0201631560921669, + 0.0332128182053566, + 0.020493965595960617, + 0.02602408081293106, + 0.016293294727802277, + 0.02058936282992363, + 0.024833079427480698, + 0.012004446238279343, + 0.061763469129800797, + 0.04307328537106514, + 0.014432798139750957, + 0.05211896449327469, + 0.006902224384248257, + 0.007420761976391077, + 0.007839289493858814, + 0.01194711308926344, + 0.012656854465603828, + 0.019247762858867645, + 0.01773497462272644, + 0.014091846533119678, + 0.010178746655583382, + 0.014021561481058598, + 0.011005999520421028, + 0.01203075796365738, + 0.019668156281113625, + 0.02565896324813366, + 0.011985166929662228, + 0.020566822960972786, + 0.007495320402085781, + 0.004893995355814695, + 0.0077446578070521355, + 0.006743237376213074, + 0.01069726049900055, + 0.008050878532230854, + 0.006915738340467215, + 0.007033301517367363, + 0.00487106665968895, + 0.004606956150382757, + 0.006666338536888361, + 0.005659039597958326, + 0.004914496093988419, + 0.006565913092344999, + 0.006714836228638887, + 0.007932103238999844, + 0.0038129789754748344, + 0.0054705762304365635, + 0.004876296501606703, + 0.004363339859992266, + 0.0060594589449465275, + 0.007276726420968771, + 0.0039264908991754055, + 0.0056880684569478035, + 0.00942489318549633, + 0.005416919942945242, + 0.005893571302294731, + 0.015230286866426468, + 0.005251988768577576, + 0.01838577352464199, + 0.005233567673712969, + 0.02699561044573784, + 0.016115702688694, + 0.00727804796770215, + 0.00833115167915821, + 0.009059526026248932, + 0.005005978047847748, + 0.007577177137136459, + 0.006576045416295528, + 0.009837325662374496, + 0.006597525905817747, + 0.002481758827343583, + 0.0017981610726565123, + 0.0015501435846090317, + 0.0035928068682551384, + 0.012998472899198532, + 0.0029190261848270893, + 0.010006614029407501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.054966483265161514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288693070411682 + ], + "opposite_logprob_tokens": [ + -1.0538692474365234 + ], + "logprob_reward": -0.4288693070411682, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4288693070411682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13333137333393097, + 0.07731505483388901, + 0.17100967466831207, + 0.057939331978559494, + 0.03450929746031761, + 0.09180235862731934, + 0.06258197128772736, + 0.10761774331331253, + 0.22839248180389404, + 0.035500671714544296 + ], + "credit_map": [ + 0.00997550692409277, + 0.025695912539958954, + 0.008853385224938393, + 0.016975853592157364, + 0.01465473510324955, + 0.018828757107257843, + 0.019568534567952156, + 0.018778711557388306, + 0.010238227434456348, + 0.008805777877569199, + 0.007307379972189665, + 0.008978691883385181, + 0.009274905547499657, + 0.016629710793495178, + 0.00882564764469862, + 0.007254719268530607, + 0.017910296097397804, + 0.0162239670753479, + 0.015289342030882835, + 0.03544691950082779, + 0.01737387664616108, + 0.028888465836644173, + 0.016266556456685066, + 0.02361026406288147, + 0.003620830597355962, + 0.006310793571174145, + 0.004427962005138397, + 0.004844042472541332, + 0.007127719931304455, + 0.013444310985505581, + 0.006968468427658081, + 0.011195210739970207, + 0.004082958679646254, + 0.002736804075539112, + 0.004952031187713146, + 0.0034334573429077864, + 0.005743568297475576, + 0.004793632309883833, + 0.0038942410610616207, + 0.0048726084642112255, + 0.0078088161535561085, + 0.010945343412458897, + 0.010076127015054226, + 0.01098104752600193, + 0.008684331551194191, + 0.013317346572875977, + 0.018355049192905426, + 0.011634312570095062, + 0.009839998558163643, + 0.006254736799746752, + 0.008827959187328815, + 0.005896198097616434, + 0.008561690337955952, + 0.00738923205062747, + 0.008878473192453384, + 0.0069336919113993645, + 0.011003527790307999, + 0.010977867059409618, + 0.009821268729865551, + 0.016650183126330376, + 0.006320704706013203, + 0.01812504604458809, + 0.006758998613804579, + 0.02796016074717045, + 0.047335706651210785, + 0.012859788723289967, + 0.011403911747038364, + 0.009229697287082672, + 0.017833324149250984, + 0.04330063238739967, + 0.01497361995279789, + 0.07145582884550095, + 0.0112950773909688, + 0.0024323025718331337, + 0.0020513106137514114, + 0.002752257976680994, + 0.003261046251282096, + 0.0030009078327566385, + 0.0036923084408044815, + 0.007015465758740902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.4288693070411682 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an open pavilion with a fantastical oriental aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a middle-aged craftsman wearing a deep teal linen robe visible working at a rustic wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a fine wolf-hair brush with ink to draw a mountain on a rectangular pillow made of moon-white brocade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin from a slightly upward angle, tilt upward, and pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light casting dappled shadows and illuminating dust particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735952105373144 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735952105373144, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735952105373144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08475467562675476, + 0.06802508234977722, + 0.12104717642068863, + 0.03913908824324608, + 0.04719537869095802, + 0.06446447968482971, + 0.11063650250434875, + 0.16517803072929382, + 0.1306106001138687, + 0.16894899308681488 + ], + "credit_map": [ + 0.006425611674785614, + 0.011042027734220028, + 0.004571957513689995, + 0.006860589142888784, + 0.01305098831653595, + 0.021574711427092552, + 0.01394625473767519, + 0.007282533682882786, + 0.008253215812146664, + 0.00876545999199152, + 0.005706109572201967, + 0.005682311486452818, + 0.012481060810387135, + 0.01372227631509304, + 0.007101233117282391, + 0.0063134147785604, + 0.012511267326772213, + 0.012246728874742985, + 0.01032476220279932, + 0.01770627126097679, + 0.01623551733791828, + 0.018862059339880943, + 0.016153406351804733, + 0.01700715906918049, + 0.004881636705249548, + 0.004207007121294737, + 0.003780724946409464, + 0.002869007410481572, + 0.00436406908556819, + 0.007849507965147495, + 0.00353407533839345, + 0.007653059437870979, + 0.007630479987710714, + 0.00469569256529212, + 0.0060462262481451035, + 0.004637337289750576, + 0.0062052784487605095, + 0.006213515531271696, + 0.00622841902077198, + 0.005538430064916611, + 0.009241688065230846, + 0.00801701471209526, + 0.011474966071546078, + 0.005136512219905853, + 0.0052351574413478374, + 0.00836004689335823, + 0.008055238053202629, + 0.008943852968513966, + 0.010083796456456184, + 0.009701255708932877, + 0.01438816823065281, + 0.02147558517754078, + 0.01951705478131771, + 0.012215289287269115, + 0.011107353493571281, + 0.012148001231253147, + 0.015638569369912148, + 0.009891731664538383, + 0.011580584570765495, + 0.029513126239180565, + 0.010095938108861446, + 0.03231802582740784, + 0.00843923818320036, + 0.04770081862807274, + 0.01942446455359459, + 0.01101669855415821, + 0.015035981312394142, + 0.02117275446653366, + 0.009565950371325016, + 0.0181232038885355, + 0.012389889918267727, + 0.02388165146112442, + 0.024279188364744186, + 0.01450731698423624, + 0.01617964170873165, + 0.016977308318018913, + 0.01656755618751049, + 0.02589666284620762, + 0.030352922156453133, + 0.024188388139009476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.006735952105373144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_tensor([530, 530])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_tensor([530, 530])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..24167d83c4f28040672d87ea880ef9d788ef5c19 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_tensor([530, 530])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An oppressive high-angle Dutch angle shot locks the viewer's gaze upon the murky, swampy waters of a nearby swamp. Rendered with hyper-realism using a next-generation game engine, a prehistoric giant, a true alligator snapping turtle, lurks within. Its mountainous, dark brown shell, adorned with three signature jagged ridges, scarred by age and covered in moss, resembles a dormant volcano. Surrounded by the murky water, only a few faint beams of light penetrate the surface and decaying vegetation, casting a volumetric light effect that precisely outlines its rough, warty hide, every inch of its slippery surface reflecting a cold sheen. At first, it's as still as a rock on a riverbed, then, its primal power is instantly awakened: its massive triangular head thrusts forward, its jaws snapping open with terrifying force, revealing its vast, pale mouth. The entire movement is swift and decisive, stirring up a small patch of turbid sand around it. This static yet tense shot, through exquisite detail and menacing composition, perfectly captures the moment this ancient creature goes from stillness to explosion, creating a primal, wild, and awe-inspiring atmosphere of tension.", + "seed": 9290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_530_seed_9290_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10769259147346019, + "vlm_logprob": -0.10769259147346019, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.10769259147346019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.40298399329185486 + ], + "opposite_logprob_tokens": [ + -1.1529840230941772 + ], + "logprob_reward": -0.40298399329185486, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.40298399329185486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13078030943870544, + 0.047732509672641754, + 0.07504440099000931, + 0.05818089842796326, + 0.045198000967502594, + 0.05992948263883591, + 0.09273999184370041, + 0.1264529675245285, + 0.04817596822977066, + 0.31576552987098694 + ], + "credit_map": [ + 0.010683459229767323, + 0.008993261493742466, + 0.008131815120577812, + 0.014199978671967983, + 0.03005252219736576, + 0.026368319988250732, + 0.015582570806145668, + 0.01676836982369423, + 0.005509182345122099, + 0.0065671824850142, + 0.005279881879687309, + 0.006404681596904993, + 0.006302713416516781, + 0.006293651647865772, + 0.006776614114642143, + 0.004598598927259445, + 0.0050072818994522095, + 0.011205180548131466, + 0.008560135960578918, + 0.010845544748008251, + 0.011205943301320076, + 0.011503282934427261, + 0.009507345966994762, + 0.007209678180515766, + 0.005317889153957367, + 0.006798091344535351, + 0.005062832962721586, + 0.007533953990787268, + 0.005131808575242758, + 0.008144953288137913, + 0.010875736363232136, + 0.00931562390178442, + 0.002727009356021881, + 0.0032032437156885862, + 0.0037509058602154255, + 0.006208387669175863, + 0.008954848162829876, + 0.004746376536786556, + 0.005984258372336626, + 0.0096229687333107, + 0.005357787478715181, + 0.009158734232187271, + 0.005828073713928461, + 0.007815289311110973, + 0.005315545480698347, + 0.009294189512729645, + 0.007746122777462006, + 0.009413735941052437, + 0.007437467575073242, + 0.008819268085062504, + 0.010791560634970665, + 0.011810285970568657, + 0.009928706102073193, + 0.01833938993513584, + 0.014932678081095219, + 0.010680624283850193, + 0.009906021878123283, + 0.013232597149908543, + 0.00909722875803709, + 0.017272891476750374, + 0.013764141127467155, + 0.03276321291923523, + 0.015622885897755623, + 0.014793982729315758, + 0.0038560086395591497, + 0.004923641681671143, + 0.0044299885630607605, + 0.00612421240657568, + 0.005678132176399231, + 0.009179091081023216, + 0.0071075535379350185, + 0.006877339445054531, + 0.011519637890160084, + 0.00996225792914629, + 0.01527431607246399, + 0.03508384898304939, + 0.01459790114313364, + 0.02122163027524948, + 0.09757965058088303, + 0.11052626371383667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.40298399329185486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608443818986416 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.01608443818986416, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608443818986416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23878467082977295, + 0.06528973579406738, + 0.0679207369685173, + 0.05906368046998978, + 0.03093945048749447, + 0.053401123732328415, + 0.08772580325603485, + 0.0714203342795372, + 0.10826551914215088, + 0.21718892455101013 + ], + "credit_map": [ + 0.023236330598592758, + 0.02485569566488266, + 0.02638307400047779, + 0.0296743493527174, + 0.04523063823580742, + 0.03984126076102257, + 0.02167213335633278, + 0.027891192585229874, + 0.005037500988692045, + 0.007482617162168026, + 0.006003368180245161, + 0.009309685789048672, + 0.009014062583446503, + 0.008441158570349216, + 0.010752562433481216, + 0.009248781017959118, + 0.008032034151256084, + 0.008681369945406914, + 0.007710309233516455, + 0.008889477699995041, + 0.007569713052362204, + 0.009172406978905201, + 0.009193740785121918, + 0.008671684190630913, + 0.006976342760026455, + 0.005352563690394163, + 0.006145297549664974, + 0.007224410306662321, + 0.007798643317073584, + 0.00907029490917921, + 0.010181260295212269, + 0.006314871832728386, + 0.00478036655113101, + 0.002909562550485134, + 0.0029792902059853077, + 0.003848278895020485, + 0.003234926611185074, + 0.004707628861069679, + 0.003719316329807043, + 0.004760082345455885, + 0.009375225752592087, + 0.0038805962540209293, + 0.006973951123654842, + 0.006276863161474466, + 0.006080915220081806, + 0.0072255246341228485, + 0.006405045744031668, + 0.007183005101978779, + 0.010397329926490784, + 0.00946037657558918, + 0.010684206150472164, + 0.011399713344871998, + 0.014346089214086533, + 0.011561566963791847, + 0.009591958485543728, + 0.01028455886989832, + 0.007431436330080032, + 0.005550215020775795, + 0.006087689194828272, + 0.00960029847919941, + 0.007488492876291275, + 0.019298981875181198, + 0.008587734773755074, + 0.007375485263764858, + 0.007874014787375927, + 0.005528821609914303, + 0.0052645341493189335, + 0.009903313592076302, + 0.0066101038828492165, + 0.015405846759676933, + 0.02408348210155964, + 0.03359540179371834, + 0.012134217657148838, + 0.0189512986689806, + 0.013791319914162159, + 0.021602390334010124, + 0.016390278935432434, + 0.023100649937987328, + 0.04406431317329407, + 0.06715445965528488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.01608443818986416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635536715388298 + ], + "opposite_logprob_tokens": [ + -3.651355266571045 + ], + "logprob_reward": -0.02635536715388298, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635536715388298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09480105340480804, + 0.08510851114988327, + 0.18186123669147491, + 0.2142481803894043, + 0.029723379760980606, + 0.05641107261180878, + 0.07055529952049255, + 0.057355791330337524, + 0.04397542029619217, + 0.1659601479768753 + ], + "credit_map": [ + 0.01208439003676176, + 0.009737830609083176, + 0.00982563104480505, + 0.013004113920032978, + 0.01313738338649273, + 0.01603940688073635, + 0.008320066146552563, + 0.01265221368521452, + 0.007456840947270393, + 0.010886100120842457, + 0.008701947517693043, + 0.010325321927666664, + 0.010728014633059502, + 0.01082037203013897, + 0.01554764062166214, + 0.010642262175679207, + 0.020559217780828476, + 0.02260827086865902, + 0.01873980462551117, + 0.025793395936489105, + 0.019034620374441147, + 0.025077184662222862, + 0.02502555213868618, + 0.025023186579346657, + 0.02155589871108532, + 0.01640995405614376, + 0.01977759413421154, + 0.025606118142604828, + 0.02414197474718094, + 0.03248274698853493, + 0.03830279782414436, + 0.03597107157111168, + 0.0020760546904057264, + 0.002070119371637702, + 0.0030595245771110058, + 0.00415315805003047, + 0.0029024770483374596, + 0.0038727368228137493, + 0.00346513488329947, + 0.008124171756207943, + 0.008118070662021637, + 0.005741354543715715, + 0.006627554073929787, + 0.0062116472981870174, + 0.004306481219828129, + 0.012101193889975548, + 0.005231861490756273, + 0.008072902448475361, + 0.0060574645176529884, + 0.006631345022469759, + 0.0086111631244421, + 0.009558109566569328, + 0.008644119836390018, + 0.011628634296357632, + 0.009750623255968094, + 0.009673829190433025, + 0.00529131107032299, + 0.005442042835056782, + 0.004411383531987667, + 0.006770401727408171, + 0.010926215909421444, + 0.011997056193649769, + 0.00658258656039834, + 0.00593478512018919, + 0.0026947734877467155, + 0.0026455428451299667, + 0.002457129769027233, + 0.004531020298600197, + 0.0028294469229876995, + 0.0053198556415736675, + 0.00955292023718357, + 0.013944726437330246, + 0.006086309440433979, + 0.009168542921543121, + 0.009661062620580196, + 0.017928684130311012, + 0.009031759575009346, + 0.013977260328829288, + 0.044579919427633286, + 0.055526599287986755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.02635536715388298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054950349032878876 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.054950349032878876, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054950349032878876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045677170157432556, + 0.051020730286836624, + 0.11441417783498764, + 0.08236914128065109, + 0.022304372861981392, + 0.04876871034502983, + 0.08623485267162323, + 0.2821411192417145, + 0.06334435939788818, + 0.20372530817985535 + ], + "credit_map": [ + 0.0033503654412925243, + 0.006531590595841408, + 0.003926645033061504, + 0.006520821712911129, + 0.007138635031878948, + 0.006751558743417263, + 0.005874268710613251, + 0.005583287216722965, + 0.004080497194081545, + 0.006420889403671026, + 0.005538737867027521, + 0.006938066799193621, + 0.007572679314762354, + 0.007279978133738041, + 0.007188324350863695, + 0.006001556757837534, + 0.00883427169173956, + 0.01418022345751524, + 0.012212014757096767, + 0.015841612592339516, + 0.016664231196045876, + 0.017048005014657974, + 0.01692475937306881, + 0.012709063477814198, + 0.010228665545582771, + 0.005943452473729849, + 0.006537471432238817, + 0.009382404386997223, + 0.00997882429510355, + 0.012290575541555882, + 0.012116380035877228, + 0.01589137315750122, + 0.0013981105294078588, + 0.0017076455987989902, + 0.0016927322139963508, + 0.002956515410915017, + 0.0038777312729507685, + 0.002994571113958955, + 0.003482158062979579, + 0.004194908309727907, + 0.007222889922559261, + 0.0035878459457308054, + 0.004565492272377014, + 0.004928962793201208, + 0.004058415070176125, + 0.014481483027338982, + 0.005146046169102192, + 0.004777577705681324, + 0.010605111718177795, + 0.008257781155407429, + 0.007257145829498768, + 0.012898117303848267, + 0.018227573484182358, + 0.013159404508769512, + 0.009132539853453636, + 0.006697180215269327, + 0.027334803715348244, + 0.025963040068745613, + 0.017663050442934036, + 0.030863745138049126, + 0.056471094489097595, + 0.0599583201110363, + 0.038424666970968246, + 0.025462405756115913, + 0.006855467334389687, + 0.004772030748426914, + 0.003751370357349515, + 0.006537093315273523, + 0.004759218078106642, + 0.011854544281959534, + 0.011476479470729828, + 0.013338151387870312, + 0.009534927085042, + 0.013360053300857544, + 0.011648211628198624, + 0.02153119631111622, + 0.010865321382880211, + 0.013764066621661186, + 0.057509180158376694, + 0.06551235914230347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.054950349032878876 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038088809698820114 + ], + "opposite_logprob_tokens": [ + -3.288088798522949 + ], + "logprob_reward": -0.038088809698820114, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038088809698820114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10211031138896942, + 0.050677355378866196, + 0.15889528393745422, + 0.07675755769014359, + 0.043148815631866455, + 0.07217039912939072, + 0.05030263960361481, + 0.060018330812454224, + 0.061641618609428406, + 0.3242776095867157 + ], + "credit_map": [ + 0.011915341950953007, + 0.009359916672110558, + 0.009909925982356071, + 0.009167918935418129, + 0.016116242855787277, + 0.014086027629673481, + 0.01250268705189228, + 0.01905224844813347, + 0.00436369189992547, + 0.007909691892564297, + 0.003936447203159332, + 0.00670963991433382, + 0.006300992798060179, + 0.006211486645042896, + 0.008453651331365108, + 0.0067917536944150925, + 0.015473444014787674, + 0.021719759330153465, + 0.016178429126739502, + 0.021297138184309006, + 0.023964062333106995, + 0.021600812673568726, + 0.021466905251145363, + 0.01719474047422409, + 0.008143458515405655, + 0.005424329545348883, + 0.008370095863938332, + 0.009035620838403702, + 0.008464144542813301, + 0.010912412777543068, + 0.012620427645742893, + 0.013787065632641315, + 0.0032902047969400883, + 0.005049544852226973, + 0.003239218145608902, + 0.004464949015527964, + 0.007733314763754606, + 0.004866213072091341, + 0.005823882762342691, + 0.008681489154696465, + 0.006703540217131376, + 0.009461800567805767, + 0.007722641807049513, + 0.00923227146267891, + 0.00790227297693491, + 0.01391131617128849, + 0.00841579120606184, + 0.008820759132504463, + 0.00487171858549118, + 0.00435244757682085, + 0.005913723260164261, + 0.005128621589392424, + 0.004409914370626211, + 0.009308100678026676, + 0.00861174426972866, + 0.007706366945058107, + 0.007642477285116911, + 0.004485195968300104, + 0.004889987874776125, + 0.006388186477124691, + 0.00846016313880682, + 0.014029789716005325, + 0.00797331053763628, + 0.0061492156237363815, + 0.0026852383743971586, + 0.0033598574809730053, + 0.002983382437378168, + 0.005807832349091768, + 0.0027411761693656445, + 0.00442689610645175, + 0.01690201833844185, + 0.022735217586159706, + 0.01363921258598566, + 0.02223513089120388, + 0.02023434452712536, + 0.03905736282467842, + 0.02146797440946102, + 0.02168792299926281, + 0.09192927181720734, + 0.09402638673782349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.038088809698820114 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An oppressive high-angle Dutch angle shot locks the viewer's gaze upon the murky, swampy waters of a nearby swamp. Rendered with hyper-realism using a next-generation game engine, a prehistoric giant, a true alligator snapping turtle, lurks within. Its mountainous, dark brown shell, adorned with three signature jagged ridges, scarred by age and covered in moss, resembles a dormant volcano. Surrounded by the murky water, only a few faint beams of light penetrate the surface and decaying vegetation, casting a volumetric light effect that precisely outlines its rough, warty hide, every inch of its slippery surface reflecting a cold sheen. At first, it's as still as a rock on a riverbed, then, its primal power is instantly awakened: its massive triangular head thrusts forward, its jaws snapping open with terrifying force, revealing its vast, pale mouth. The entire movement is swift and decisive, stirring up a small patch of turbid sand around it. This static yet tense shot, through exquisite detail and menacing composition, perfectly captures the moment this ancient creature goes from stillness to explosion, creating a primal, wild, and awe-inspiring atmosphere of tension.", + "seed": 9291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_530_seed_9291_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2134838156402111, + "vlm_logprob": -0.2134838156402111, + "vlm_yes_no_margin": 1.975000023841858, + "vlm_token_credit": -0.2134838156402111, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.707997739315033 + ], + "opposite_logprob_tokens": [ + -0.707997739315033 + ], + "logprob_reward": -0.707997739315033, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.707997739315033, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16034919023513794, + 0.07692151516675949, + 0.06245815381407738, + 0.05582944303750992, + 0.07193052023649216, + 0.15342624485492706, + 0.05950302258133888, + 0.07015050947666168, + 0.1684465855360031, + 0.12098485231399536 + ], + "credit_map": [ + 0.026955971494317055, + 0.028111765161156654, + 0.021367942914366722, + 0.01655665785074234, + 0.020792623981833458, + 0.02049178257584572, + 0.013162790797650814, + 0.012909661047160625, + 0.00931362621486187, + 0.009928585030138493, + 0.005932781379669905, + 0.011168967001140118, + 0.012239161878824234, + 0.010077979415655136, + 0.01038625929504633, + 0.007874160073697567, + 0.00873579178005457, + 0.009377569891512394, + 0.004723303020000458, + 0.007328900508582592, + 0.0103578120470047, + 0.007036061026155949, + 0.007443388924002647, + 0.0074553280137479305, + 0.007038149517029524, + 0.007867562584578991, + 0.004917819518595934, + 0.00693301809951663, + 0.007148063741624355, + 0.007393099833279848, + 0.007627806160598993, + 0.006903927773237228, + 0.008260650560259819, + 0.008790982887148857, + 0.0157945416867733, + 0.00552778597921133, + 0.007177361287176609, + 0.008580672554671764, + 0.00865219160914421, + 0.009146328084170818, + 0.024731647223234177, + 0.019421402364969254, + 0.01938273012638092, + 0.008798672817647457, + 0.015858028084039688, + 0.025211816653609276, + 0.01678580604493618, + 0.023236144334077835, + 0.00699823210015893, + 0.008825036697089672, + 0.004764942917972803, + 0.007357743568718433, + 0.007602522149682045, + 0.00856130849570036, + 0.007676518987864256, + 0.007716718129813671, + 0.010012250393629074, + 0.007704697549343109, + 0.005943071562796831, + 0.007982403039932251, + 0.00752935279160738, + 0.007689040154218674, + 0.010824781842529774, + 0.012464910745620728, + 0.03453435003757477, + 0.013405542820692062, + 0.016223464161157608, + 0.0132150175049901, + 0.014857111498713493, + 0.013269923627376556, + 0.02572132647037506, + 0.03721984848380089, + 0.014712144620716572, + 0.011210035532712936, + 0.008377129212021828, + 0.01738019660115242, + 0.017997803166508675, + 0.008870335295796394, + 0.023346085101366043, + 0.019091123715043068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.707997739315033 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486886166036129 + ], + "opposite_logprob_tokens": [ + -3.0486886501312256 + ], + "logprob_reward": -0.0486886166036129, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486886166036129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20404404401779175, + 0.09886050224304199, + 0.08572107553482056, + 0.04340076446533203, + 0.03482833877205849, + 0.08205360919237137, + 0.09250887483358383, + 0.05681246519088745, + 0.19272319972515106, + 0.10904717445373535 + ], + "credit_map": [ + 0.03285447880625725, + 0.026801515370607376, + 0.021243605762720108, + 0.015786563977599144, + 0.032762255519628525, + 0.02439122460782528, + 0.020058900117874146, + 0.03014550358057022, + 0.010923536494374275, + 0.01733475923538208, + 0.010528534650802612, + 0.011368080973625183, + 0.012378547340631485, + 0.013534008525311947, + 0.012034471146762371, + 0.010758562944829464, + 0.009445015341043472, + 0.015051797032356262, + 0.008375120349228382, + 0.009276483207941055, + 0.011576871387660503, + 0.010135283693671227, + 0.014522495679557323, + 0.007338006515055895, + 0.005575024988502264, + 0.007279426325112581, + 0.006294336169958115, + 0.0042054844088852406, + 0.005038441624492407, + 0.005380278453230858, + 0.005318398587405682, + 0.0043093739077448845, + 0.003666504519060254, + 0.003953168634325266, + 0.00870704185217619, + 0.0029534590430557728, + 0.0031345458701252937, + 0.004400158766657114, + 0.004044343251734972, + 0.003969118464738131, + 0.012645105831325054, + 0.010410314425826073, + 0.01205213088542223, + 0.007666940800845623, + 0.00782907847315073, + 0.011903229169547558, + 0.009172474965453148, + 0.010374335572123528, + 0.011451293714344501, + 0.010561494156718254, + 0.0087275430560112, + 0.009007291868329048, + 0.010461818426847458, + 0.016828058287501335, + 0.011116446927189827, + 0.014354925602674484, + 0.00710758613422513, + 0.005965060088783503, + 0.005361225921660662, + 0.005378305912017822, + 0.005618029739707708, + 0.006000475026667118, + 0.010930940508842468, + 0.010450840927660465, + 0.040884729474782944, + 0.016577940434217453, + 0.017892159521579742, + 0.012631500139832497, + 0.014424622990190983, + 0.015357634983956814, + 0.028751902282238007, + 0.046202704310417175, + 0.013162830844521523, + 0.010133186355233192, + 0.011750309728085995, + 0.015057327225804329, + 0.01671731099486351, + 0.008610806427896023, + 0.020487332716584206, + 0.013128064572811127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0486886166036129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498160421848297 + ], + "opposite_logprob_tokens": [ + -2.9299817085266113 + ], + "logprob_reward": -0.05498160421848297, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498160421848297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13092119991779327, + 0.18183894455432892, + 0.15326525270938873, + 0.03551989793777466, + 0.04016930237412453, + 0.09940262138843536, + 0.039582960307598114, + 0.044158145785331726, + 0.1756320595741272, + 0.09950952976942062 + ], + "credit_map": [ + 0.015583573840558529, + 0.020290076732635498, + 0.01265578344464302, + 0.013380235061049461, + 0.0221250019967556, + 0.01772773638367653, + 0.013119992800056934, + 0.0160388071089983, + 0.037763312458992004, + 0.02717449888586998, + 0.01915885880589485, + 0.018167097121477127, + 0.018545275554060936, + 0.025327365845441818, + 0.01808815263211727, + 0.017614390701055527, + 0.02567456290125847, + 0.024043096229434013, + 0.012404028326272964, + 0.014582950621843338, + 0.025528213009238243, + 0.018798429518938065, + 0.016594815999269485, + 0.0156391728669405, + 0.005192686803638935, + 0.006165729369968176, + 0.004972814116626978, + 0.003524166066199541, + 0.004361042287200689, + 0.004347668960690498, + 0.0036188792437314987, + 0.0033369152806699276, + 0.0063750287517905235, + 0.006490188185125589, + 0.008284972049295902, + 0.0029968188609927893, + 0.0030699106864631176, + 0.005446764174848795, + 0.0031396611593663692, + 0.004365964327007532, + 0.014620945788919926, + 0.012224113568663597, + 0.014091217890381813, + 0.010905942879617214, + 0.0095729511231184, + 0.015901178121566772, + 0.010753887705504894, + 0.011332396417856216, + 0.006634939927607775, + 0.003668449353426695, + 0.0038935288321226835, + 0.003639867063611746, + 0.004033344332128763, + 0.004531468264758587, + 0.007489060517400503, + 0.005692305043339729, + 0.006155496928840876, + 0.004943421576172113, + 0.004139645956456661, + 0.004541231319308281, + 0.005364765413105488, + 0.004722265992313623, + 0.007015221752226353, + 0.0072761038318276405, + 0.038571797311306, + 0.016183413565158844, + 0.01832568645477295, + 0.012105001136660576, + 0.012120592407882214, + 0.01086617261171341, + 0.02723475731909275, + 0.04022465646266937, + 0.011482764035463333, + 0.008372681215405464, + 0.010633722878992558, + 0.0188678540289402, + 0.011803066357970238, + 0.0074544306844472885, + 0.019335251301527023, + 0.011559763923287392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.05498160421848297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294373869895935 + ], + "opposite_logprob_tokens": [ + -2.237943649291992 + ], + "logprob_reward": -0.11294373869895935, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294373869895935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0790279358625412, + 0.10066929459571838, + 0.02895784191787243, + 0.047836605459451675, + 0.029673317447304726, + 0.17282482981681824, + 0.055427201092243195, + 0.14057087898254395, + 0.1901247203350067, + 0.1548873782157898 + ], + "credit_map": [ + 0.009476118721067905, + 0.00965218897908926, + 0.009686794131994247, + 0.007733341306447983, + 0.012667429633438587, + 0.0107400668784976, + 0.00941501185297966, + 0.009656976908445358, + 0.0123499296605587, + 0.014020697213709354, + 0.009663121774792671, + 0.011059912852942944, + 0.01270209439098835, + 0.012793850153684616, + 0.016315598040819168, + 0.01176407653838396, + 0.0033766934648156166, + 0.005645317956805229, + 0.003586144419386983, + 0.0029688600916415453, + 0.002771867671981454, + 0.0037185417022556067, + 0.004054111894220114, + 0.0028363021556288004, + 0.0065481653437018394, + 0.006562188733369112, + 0.006444375030696392, + 0.00427748030051589, + 0.005481967702507973, + 0.006320582237094641, + 0.007143533322960138, + 0.0050583104602992535, + 0.003572822082787752, + 0.004020757973194122, + 0.005643417127430439, + 0.002228924771770835, + 0.0037292116321623325, + 0.0031457117293030024, + 0.004323924891650677, + 0.0030085428152233362, + 0.024812959134578705, + 0.020546352490782738, + 0.0219354797154665, + 0.0106684984639287, + 0.01773163489997387, + 0.030269816517829895, + 0.020653359591960907, + 0.026206713169813156, + 0.008173464797437191, + 0.005674057174474001, + 0.004942639730870724, + 0.006795368157327175, + 0.0057740346528589725, + 0.008977517485618591, + 0.008301115594804287, + 0.006788997445255518, + 0.020665286108851433, + 0.017425987869501114, + 0.015404687263071537, + 0.018557317554950714, + 0.016287608072161674, + 0.01781463250517845, + 0.01664930023252964, + 0.01776604913175106, + 0.02916508913040161, + 0.020762359723448753, + 0.022626370191574097, + 0.02079649083316326, + 0.01890094392001629, + 0.02391287125647068, + 0.028266234323382378, + 0.0256943441927433, + 0.01706753671169281, + 0.014269844628870487, + 0.014254373498260975, + 0.023103823885321617, + 0.02138020284473896, + 0.02007794938981533, + 0.027896005660295486, + 0.01683761551976204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.11294373869895935 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280737936496735 + ], + "opposite_logprob_tokens": [ + -2.0178074836730957 + ], + "logprob_reward": -0.14280737936496735, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280737936496735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1281728893518448, + 0.04116189479827881, + 0.03476262837648392, + 0.03068925440311432, + 0.02905748412013054, + 0.2760622203350067, + 0.024562787264585495, + 0.0717378631234169, + 0.24575580656528473, + 0.11803718656301498 + ], + "credit_map": [ + 0.024389423429965973, + 0.01687379740178585, + 0.018538357689976692, + 0.010479643940925598, + 0.018957428634166718, + 0.014710680581629276, + 0.008444820530712605, + 0.015778737142682076, + 0.007215048652142286, + 0.005826023407280445, + 0.004721238277852535, + 0.004367221612483263, + 0.00544203445315361, + 0.00514586316421628, + 0.0038756572175771, + 0.004568806383758783, + 0.0056100282818078995, + 0.005565881263464689, + 0.003016894683241844, + 0.0032928704749792814, + 0.005954499822109938, + 0.004339542239904404, + 0.0037714121863245964, + 0.003211500123143196, + 0.0045799147337675095, + 0.0046518039889633656, + 0.003527879947796464, + 0.002894602483138442, + 0.0038480660878121853, + 0.004136213567107916, + 0.0037205994594842196, + 0.0033301732037216425, + 0.0038390897680073977, + 0.0027403593994677067, + 0.006642960477620363, + 0.0025074747391045094, + 0.003240772522985935, + 0.0034759240224957466, + 0.003072733525186777, + 0.0035381698980927467, + 0.041113678365945816, + 0.02833600714802742, + 0.04378571733832359, + 0.025986693799495697, + 0.026287544518709183, + 0.03986527770757675, + 0.034395914524793625, + 0.03629138320684433, + 0.0034556423779577017, + 0.002505525713786483, + 0.0026817424222826958, + 0.0022270504850894213, + 0.002522998722270131, + 0.004806941840797663, + 0.003016611561179161, + 0.003346274606883526, + 0.012028088793158531, + 0.008014287799596786, + 0.007229419890791178, + 0.007450276054441929, + 0.009531578980386257, + 0.006787578575313091, + 0.010066055692732334, + 0.01063057966530323, + 0.04477027803659439, + 0.024034209549427032, + 0.03422579914331436, + 0.024379640817642212, + 0.02012319304049015, + 0.025477800518274307, + 0.03960553929209709, + 0.033139344304800034, + 0.009253522381186485, + 0.010624261572957039, + 0.013041836209595203, + 0.02348128892481327, + 0.013370168395340443, + 0.01177881471812725, + 0.022676659747958183, + 0.013810621574521065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.14280737936496735 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An oppressive high-angle Dutch angle shot locks the viewer's gaze upon the murky, swampy waters of a nearby swamp. Rendered with hyper-realism using a next-generation game engine, a prehistoric giant, a true alligator snapping turtle, lurks within. Its mountainous, dark brown shell, adorned with three signature jagged ridges, scarred by age and covered in moss, resembles a dormant volcano. Surrounded by the murky water, only a few faint beams of light penetrate the surface and decaying vegetation, casting a volumetric light effect that precisely outlines its rough, warty hide, every inch of its slippery surface reflecting a cold sheen. At first, it's as still as a rock on a riverbed, then, its primal power is instantly awakened: its massive triangular head thrusts forward, its jaws snapping open with terrifying force, revealing its vast, pale mouth. The entire movement is swift and decisive, stirring up a small patch of turbid sand around it. This static yet tense shot, through exquisite detail and menacing composition, perfectly captures the moment this ancient creature goes from stillness to explosion, creating a primal, wild, and awe-inspiring atmosphere of tension.", + "seed": 9292, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_530_seed_9292_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12238235305994749, + "vlm_logprob": -0.12238235305994749, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.12238235305994749, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3401608169078827 + ], + "opposite_logprob_tokens": [ + -1.340160846710205 + ], + "logprob_reward": -0.3401608169078827, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3401608169078827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17874543368816376, + 0.04615602269768715, + 0.04868509620428085, + 0.05286834016442299, + 0.1491474062204361, + 0.14193235337734222, + 0.16089585423469543, + 0.12920092046260834, + 0.057684607803821564, + 0.03468392416834831 + ], + "credit_map": [ + 0.014024454168975353, + 0.010294409468770027, + 0.030313601717352867, + 0.019901344552636147, + 0.016253285109996796, + 0.029340559616684914, + 0.009503282606601715, + 0.049114517867565155, + 0.0025980074424296618, + 0.0026377409230917692, + 0.009764035232365131, + 0.006775726564228535, + 0.007661169394850731, + 0.006813364569097757, + 0.004653562791645527, + 0.005252421833574772, + 0.00420780573040247, + 0.0038844712544232607, + 0.008677957579493523, + 0.005995283834636211, + 0.007036481983959675, + 0.007138632703572512, + 0.003860591445118189, + 0.00788387656211853, + 0.008767643943428993, + 0.0035142202395945787, + 0.007676310837268829, + 0.005807778332382441, + 0.0031620694790035486, + 0.007424547802656889, + 0.004631371237337589, + 0.011884400621056557, + 0.020417189225554466, + 0.010339598171412945, + 0.027209529653191566, + 0.017566747963428497, + 0.013418981805443764, + 0.020553622394800186, + 0.017009202390909195, + 0.022632550448179245, + 0.011640193872153759, + 0.009691066108644009, + 0.015777086839079857, + 0.013515586964786053, + 0.010412652976810932, + 0.037965595722198486, + 0.024992506951093674, + 0.0179376769810915, + 0.007068872917443514, + 0.00830763764679432, + 0.008701203390955925, + 0.013251994736492634, + 0.00750731909647584, + 0.010472911410033703, + 0.05716661736369133, + 0.04841931164264679, + 0.011008555069565773, + 0.008140242658555508, + 0.010518938302993774, + 0.031342849135398865, + 0.009647337719798088, + 0.019455693662166595, + 0.020623087882995605, + 0.01846422255039215, + 0.0066914912313222885, + 0.003999294713139534, + 0.005871792323887348, + 0.009963934309780598, + 0.00614609569311142, + 0.006694262847304344, + 0.010053631849586964, + 0.008264112286269665, + 0.003294329158961773, + 0.003428183263167739, + 0.003376168664544821, + 0.006793080363422632, + 0.005693104118108749, + 0.003051555249840021, + 0.004758207127451897, + 0.004289301577955484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.3401608169078827 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214574359357357 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214574359357357, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214574359357357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24576540291309357, + 0.08371128141880035, + 0.0327468104660511, + 0.09298139065504074, + 0.13335350155830383, + 0.0667453184723854, + 0.05553362891077995, + 0.23153629899024963, + 0.028773687779903412, + 0.02885262295603752 + ], + "credit_map": [ + 0.017608042806386948, + 0.01453312300145626, + 0.06196865066885948, + 0.023820504546165466, + 0.024217715486884117, + 0.03326672688126564, + 0.013140707276761532, + 0.05720992758870125, + 0.007255617529153824, + 0.004786008037626743, + 0.018444016575813293, + 0.010153882205486298, + 0.01356975082308054, + 0.01024444680660963, + 0.00632080901414156, + 0.012936750426888466, + 0.0031746819149702787, + 0.0027991682291030884, + 0.004083664156496525, + 0.0028709699399769306, + 0.004357484634965658, + 0.0064123161137104034, + 0.0023044641129672527, + 0.006744062062352896, + 0.0066312928684055805, + 0.0075938100926578045, + 0.013895386829972267, + 0.027309343218803406, + 0.005572665948420763, + 0.01275468710809946, + 0.008518771268427372, + 0.0107054328545928, + 0.011148354038596153, + 0.009339177049696445, + 0.0166762787848711, + 0.0157392006367445, + 0.009034788236021996, + 0.02092060260474682, + 0.01848871447145939, + 0.03200637549161911, + 0.005942982155829668, + 0.006496650632470846, + 0.00802795309573412, + 0.008652149699628353, + 0.005728202871978283, + 0.013340497389435768, + 0.011505890637636185, + 0.007050992455333471, + 0.004364524036645889, + 0.0029848385602235794, + 0.009335262700915337, + 0.006285896059125662, + 0.0033243864309042692, + 0.011952408589422703, + 0.009444168768823147, + 0.00784214586019516, + 0.02162044495344162, + 0.015989655628800392, + 0.02445095032453537, + 0.058432791382074356, + 0.01360092218965292, + 0.034066092222929, + 0.03757603466510773, + 0.025799419730901718, + 0.003077229019254446, + 0.0023413740564137697, + 0.0027090520597994328, + 0.005236663389950991, + 0.003021703567355871, + 0.003063544863834977, + 0.004096778575330973, + 0.005227343644946814, + 0.00370509154163301, + 0.0031379451975226402, + 0.003085282864049077, + 0.005441849119961262, + 0.0023312086705118418, + 0.001980464207008481, + 0.004367870278656483, + 0.00480291061103344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.014214574359357357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200181424617767 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.018200181424617767, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018200181424617767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10859614610671997, + 0.17826952040195465, + 0.04557759687304497, + 0.11802179366350174, + 0.13257595896720886, + 0.06529954075813293, + 0.061959560960531235, + 0.09275179356336594, + 0.14033471047878265, + 0.05661334469914436 + ], + "credit_map": [ + 0.008446001447737217, + 0.01053609885275364, + 0.018699917942285538, + 0.015534837730228901, + 0.011209682561457157, + 0.013992189429700375, + 0.007396494038403034, + 0.022780925035476685, + 0.01610150747001171, + 0.010031537152826786, + 0.05193300545215607, + 0.015481401234865189, + 0.0230752844363451, + 0.023261884227395058, + 0.012941005639731884, + 0.0254439115524292, + 0.004323846194893122, + 0.003445906564593315, + 0.008632287383079529, + 0.00939160492271185, + 0.004367331508547068, + 0.006210724823176861, + 0.003401060588657856, + 0.0058048381470143795, + 0.010338844731450081, + 0.011482569389045238, + 0.01656397618353367, + 0.028483951464295387, + 0.007546755485236645, + 0.015185310505330563, + 0.012052723206579685, + 0.016367673873901367, + 0.01970609277486801, + 0.010196182876825333, + 0.019689712673425674, + 0.019257990643382072, + 0.009026397950947285, + 0.01638609729707241, + 0.015463226474821568, + 0.02285025455057621, + 0.007048097904771566, + 0.003627439495176077, + 0.005353459622710943, + 0.004632510244846344, + 0.003938603214919567, + 0.019719980657100677, + 0.01202991884201765, + 0.00894953589886427, + 0.003821010235697031, + 0.003132660873234272, + 0.00823415257036686, + 0.006069501396268606, + 0.003224447136744857, + 0.009375004097819328, + 0.01784680411219597, + 0.01025599054992199, + 0.008574038743972778, + 0.006136192008852959, + 0.008735650219023228, + 0.012885474599897861, + 0.005766073241829872, + 0.018080106005072594, + 0.017537102103233337, + 0.015037166886031628, + 0.017200620844960213, + 0.010851518251001835, + 0.013272250071167946, + 0.019135871902108192, + 0.015668107196688652, + 0.020795630291104317, + 0.02188047207891941, + 0.02153025008738041, + 0.00528762536123395, + 0.007706596981734037, + 0.004612181801348925, + 0.012615037150681019, + 0.004247372969985008, + 0.0032849148847162724, + 0.008642708882689476, + 0.01021691132336855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.018200181424617767 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898636162281036 + ], + "opposite_logprob_tokens": [ + -2.578986406326294 + ], + "logprob_reward": -0.07898636162281036, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898636162281036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1231393963098526, + 0.135118767619133, + 0.08113791048526764, + 0.05114579200744629, + 0.11460249125957489, + 0.04804835841059685, + 0.07279003411531448, + 0.1718575358390808, + 0.09861722588539124, + 0.10354244709014893 + ], + "credit_map": [ + 0.009875195100903511, + 0.008422103710472584, + 0.016222432255744934, + 0.024668682366609573, + 0.01277097687125206, + 0.017165059223771095, + 0.008438033983111382, + 0.025576919317245483, + 0.010760544799268246, + 0.007485499605536461, + 0.03577946498990059, + 0.010648246854543686, + 0.014711551368236542, + 0.022193878889083862, + 0.00793298427015543, + 0.025606602430343628, + 0.005169131327420473, + 0.00576108880341053, + 0.009107805788516998, + 0.013030050322413445, + 0.010517289862036705, + 0.015605838969349861, + 0.007040169555693865, + 0.014906538650393486, + 0.004473625682294369, + 0.002812335966154933, + 0.00801621563732624, + 0.009103944525122643, + 0.004739767871797085, + 0.0068990918807685375, + 0.00515454588457942, + 0.009946268051862717, + 0.009935090318322182, + 0.010465717874467373, + 0.014930661767721176, + 0.013801859691739082, + 0.011071627959609032, + 0.017560064792633057, + 0.017281746491789818, + 0.019555730745196342, + 0.0043135290034115314, + 0.004185376688838005, + 0.0038491073064506054, + 0.0038539916276931763, + 0.003995969425886869, + 0.014008053578436375, + 0.007735319435596466, + 0.006107013672590256, + 0.003472615499049425, + 0.0023785012308508158, + 0.00670769764110446, + 0.006174044217914343, + 0.002960568992421031, + 0.008040662854909897, + 0.021751683205366135, + 0.021304257214069366, + 0.014829263091087341, + 0.010589275509119034, + 0.014978863298892975, + 0.05001883953809738, + 0.0097476402297616, + 0.019413048401474953, + 0.0321742445230484, + 0.020106375217437744, + 0.016124580055475235, + 0.007701538503170013, + 0.0077675036154687405, + 0.014296496286988258, + 0.01109845656901598, + 0.009937621653079987, + 0.015698609873652458, + 0.01599242351949215, + 0.012684577144682407, + 0.013496771454811096, + 0.007818353362381458, + 0.019338319078087807, + 0.00954002607613802, + 0.007828482426702976, + 0.014851272106170654, + 0.017984645441174507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.07898636162281036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034983098506927 + ], + "opposite_logprob_tokens": [ + -1.9103498458862305 + ], + "logprob_reward": -0.16034983098506927, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034983098506927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27152109146118164, + 0.03559847176074982, + 0.09566520899534225, + 0.10518947243690491, + 0.12254409492015839, + 0.06890515983104706, + 0.05040879175066948, + 0.10546065121889114, + 0.044987089931964874, + 0.09971990436315536 + ], + "credit_map": [ + 0.018078533932566643, + 0.013727441430091858, + 0.06764354556798935, + 0.026550929993391037, + 0.024931298568844795, + 0.03997721150517464, + 0.013855328783392906, + 0.06675682961940765, + 0.0032273810356855392, + 0.0027723200619220734, + 0.0046079447492957115, + 0.005481880158185959, + 0.005080942530184984, + 0.005797139368951321, + 0.002609625458717346, + 0.006021242588758469, + 0.009888066910207272, + 0.007716309744864702, + 0.01388472504913807, + 0.01053523737937212, + 0.00971759483218193, + 0.018215244635939598, + 0.0066660610027611256, + 0.019041985273361206, + 0.015318877063691616, + 0.006604696623980999, + 0.01731259562075138, + 0.017598265781998634, + 0.006767382845282555, + 0.013299426063895226, + 0.008208200335502625, + 0.02008003741502762, + 0.01613778993487358, + 0.010422098450362682, + 0.018015172332525253, + 0.016574300825595856, + 0.009755515493452549, + 0.016681017354130745, + 0.015282042324543, + 0.019676173105835915, + 0.0059129162691533566, + 0.005218226462602615, + 0.00553150475025177, + 0.0054635656997561455, + 0.004944131709635258, + 0.021670415997505188, + 0.013844698667526245, + 0.006319709587842226, + 0.0032279430888593197, + 0.0020044001284986734, + 0.004970768932253122, + 0.004195318557322025, + 0.002240018220618367, + 0.00518705602735281, + 0.01764296367764473, + 0.010940329171717167, + 0.011955725960433483, + 0.006098143756389618, + 0.008926017209887505, + 0.012068008072674274, + 0.00710498820990324, + 0.025811735540628433, + 0.01909981109201908, + 0.01439623162150383, + 0.004772543907165527, + 0.003041245276108384, + 0.004559820517897606, + 0.007756777573376894, + 0.0036885656882077456, + 0.0062593198381364346, + 0.008295916952192783, + 0.006612904369831085, + 0.008460713550448418, + 0.012744002044200897, + 0.012121839448809624, + 0.02205752022564411, + 0.007956755347549915, + 0.007561867590993643, + 0.014144840650260448, + 0.014672374352812767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.16034983098506927 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An oppressive high-angle Dutch angle shot locks the viewer's gaze upon the murky, swampy waters of a nearby swamp. Rendered with hyper-realism using a next-generation game engine, a prehistoric giant, a true alligator snapping turtle, lurks within. Its mountainous, dark brown shell, adorned with three signature jagged ridges, scarred by age and covered in moss, resembles a dormant volcano. Surrounded by the murky water, only a few faint beams of light penetrate the surface and decaying vegetation, casting a volumetric light effect that precisely outlines its rough, warty hide, every inch of its slippery surface reflecting a cold sheen. At first, it's as still as a rock on a riverbed, then, its primal power is instantly awakened: its massive triangular head thrusts forward, its jaws snapping open with terrifying force, revealing its vast, pale mouth. The entire movement is swift and decisive, stirring up a small patch of turbid sand around it. This static yet tense shot, through exquisite detail and menacing composition, perfectly captures the moment this ancient creature goes from stillness to explosion, creating a primal, wild, and awe-inspiring atmosphere of tension.", + "seed": 9293, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_7/index_530_seed_9293_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.072871158644557, + "vlm_logprob": -0.072871158644557, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.072871158644557, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24454396963119507 + ], + "opposite_logprob_tokens": [ + -1.6195440292358398 + ], + "logprob_reward": -0.24454396963119507, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24454396963119507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16649749875068665, + 0.0709327980875969, + 0.16051273047924042, + 0.11972536891698837, + 0.052900608628988266, + 0.06348951160907745, + 0.15287448465824127, + 0.08426275849342346, + 0.06126224994659424, + 0.06754201650619507 + ], + "credit_map": [ + 0.030646398663520813, + 0.017490696161985397, + 0.025294756516814232, + 0.012785759754478931, + 0.021546075120568275, + 0.023304009810090065, + 0.01665041409432888, + 0.018779387697577477, + 0.007937931455671787, + 0.005172297824174166, + 0.012376893311738968, + 0.00557396374642849, + 0.008473493158817291, + 0.018835904076695442, + 0.006206464022397995, + 0.006355848163366318, + 0.017980000004172325, + 0.016839155927300453, + 0.018889229744672775, + 0.017494475468993187, + 0.0253578033298254, + 0.031405363231897354, + 0.01449610386043787, + 0.018050601705908775, + 0.01384058315306902, + 0.016109906136989594, + 0.01558653637766838, + 0.007501750718802214, + 0.01728213205933571, + 0.019237956032156944, + 0.014964047819375992, + 0.015202462673187256, + 0.005802970379590988, + 0.004512062296271324, + 0.010014576837420464, + 0.0034972091671079397, + 0.003921636380255222, + 0.01283333171159029, + 0.005178444087505341, + 0.007140378467738628, + 0.00822392851114273, + 0.009028750471770763, + 0.0061943442560732365, + 0.004459739662706852, + 0.008183859288692474, + 0.009164605289697647, + 0.008478558622300625, + 0.009755725972354412, + 0.025987328961491585, + 0.019816752523183823, + 0.025816263630986214, + 0.012763122096657753, + 0.018698127940297127, + 0.020328357815742493, + 0.014463580213487148, + 0.015000949613749981, + 0.01327210757881403, + 0.004685652442276478, + 0.020975708961486816, + 0.00576100405305624, + 0.006305286195129156, + 0.011171814054250717, + 0.01505134254693985, + 0.007039835210889578, + 0.0067659527994692326, + 0.004281799308955669, + 0.006658319383859634, + 0.0074256407096982, + 0.005251317750662565, + 0.01464923843741417, + 0.008600336499512196, + 0.007629646919667721, + 0.00860607624053955, + 0.0035013360902667046, + 0.006975740194320679, + 0.008325598202645779, + 0.004543547984212637, + 0.014346275478601456, + 0.01393522135913372, + 0.007308223284780979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.24454396963119507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598655566573143 + ], + "opposite_logprob_tokens": [ + -3.8955986499786377 + ], + "logprob_reward": -0.020598655566573143, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598655566573143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12759478390216827, + 0.28611788153648376, + 0.10657087713479996, + 0.05770586431026459, + 0.04796338081359863, + 0.03727981075644493, + 0.18994712829589844, + 0.044353581964969635, + 0.053150374442338943, + 0.04931630194187164 + ], + "credit_map": [ + 0.023093629628419876, + 0.010790973901748657, + 0.01924833469092846, + 0.010705358348786831, + 0.016906311735510826, + 0.01912582851946354, + 0.013113625347614288, + 0.014610710553824902, + 0.030065156519412994, + 0.024028226733207703, + 0.050885140895843506, + 0.020549479871988297, + 0.029077934101223946, + 0.07680054754018784, + 0.03223418444395065, + 0.02247719280421734, + 0.01934490166604519, + 0.01058957539498806, + 0.011983511969447136, + 0.008684646338224411, + 0.011712200939655304, + 0.024107899516820908, + 0.009670702740550041, + 0.010477439500391483, + 0.009798441082239151, + 0.006427358835935593, + 0.005184872075915337, + 0.003379313973709941, + 0.0068601081147789955, + 0.01242571696639061, + 0.0069727166555821896, + 0.0066573345102369785, + 0.005530268419533968, + 0.0029698279686272144, + 0.005483461078256369, + 0.00346161425113678, + 0.0037972554564476013, + 0.013007101602852345, + 0.007484095171093941, + 0.006229754537343979, + 0.003923310432583094, + 0.0040270439349114895, + 0.0030310433357954025, + 0.0033817908260971308, + 0.005009987857192755, + 0.006563267204910517, + 0.005895519163459539, + 0.005447844509035349, + 0.029988979920744896, + 0.018981561064720154, + 0.025538761168718338, + 0.015604664571583271, + 0.032642994076013565, + 0.027134040370583534, + 0.026800105348229408, + 0.013256021775305271, + 0.007378870155662298, + 0.004528264980763197, + 0.004122271668165922, + 0.0032799600157886744, + 0.004741994198411703, + 0.009436309337615967, + 0.006964258383959532, + 0.003901650896295905, + 0.005313735920935869, + 0.005163068883121014, + 0.0054632993414998055, + 0.00549284927546978, + 0.004114398267120123, + 0.013800876215100288, + 0.008561099879443645, + 0.005241044797003269, + 0.006664665415883064, + 0.0026433041784912348, + 0.009018134325742722, + 0.003203798085451126, + 0.003925611265003681, + 0.009463898837566376, + 0.008980008773505688, + 0.005416875705122948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.020598655566573143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635931596159935 + ], + "opposite_logprob_tokens": [ + -3.6513593196868896 + ], + "logprob_reward": -0.02635931596159935, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635931596159935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14671504497528076, + 0.20945122838020325, + 0.09134276956319809, + 0.060423027724027634, + 0.07055388391017914, + 0.06380660086870193, + 0.08360613137483597, + 0.033742260187864304, + 0.06444521993398666, + 0.17591385543346405 + ], + "credit_map": [ + 0.02878098003566265, + 0.01174097042530775, + 0.02061370015144348, + 0.00984830129891634, + 0.0229035671800375, + 0.02185576595366001, + 0.013982622884213924, + 0.01698913797736168, + 0.0243073757737875, + 0.01919732056558132, + 0.024104641750454903, + 0.013667948544025421, + 0.04156101495027542, + 0.049597010016441345, + 0.020441634580492973, + 0.016574285924434662, + 0.012861109338700771, + 0.011078580282628536, + 0.009700216352939606, + 0.0073607685044407845, + 0.009084395132958889, + 0.017617475241422653, + 0.009911163710057735, + 0.013729061000049114, + 0.00849924050271511, + 0.007304063532501459, + 0.0062172906473279, + 0.005099534057080746, + 0.008974092081189156, + 0.00917130522429943, + 0.009024950675666332, + 0.006132552400231361, + 0.006095969583839178, + 0.0038300100713968277, + 0.010922408662736416, + 0.0038816984742879868, + 0.004830474033951759, + 0.025449059903621674, + 0.008986672386527061, + 0.0065575879998505116, + 0.006645041052252054, + 0.006536773405969143, + 0.010860687121748924, + 0.0036090626381337643, + 0.006242081057280302, + 0.01563282124698162, + 0.007372311316430569, + 0.006907821632921696, + 0.011861522682011127, + 0.012057989835739136, + 0.010126223787665367, + 0.007098481059074402, + 0.01482890360057354, + 0.01033929456025362, + 0.00935312733054161, + 0.00794058758765459, + 0.005094121675938368, + 0.005287892185151577, + 0.003440363798290491, + 0.0029599617701023817, + 0.0036080549471080303, + 0.00528302974998951, + 0.003883067751303315, + 0.004185768775641918, + 0.007104653399437666, + 0.006186329759657383, + 0.007588246371597052, + 0.006586072966456413, + 0.005209817085415125, + 0.014353727921843529, + 0.01139652356505394, + 0.006019847933202982, + 0.023759309202432632, + 0.005890378728508949, + 0.043141987174749374, + 0.009780896827578545, + 0.009620356373488903, + 0.027435049414634705, + 0.04071661829948425, + 0.015569259412586689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.02635931596159935 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063513934612274 + ], + "opposite_logprob_tokens": [ + -3.1680634021759033 + ], + "logprob_reward": -0.043063513934612274, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063513934612274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14738939702510834, + 0.14652054011821747, + 0.22248995304107666, + 0.06752877682447433, + 0.07124567776918411, + 0.052370619028806686, + 0.08714652061462402, + 0.07429308444261551, + 0.08161284774541855, + 0.049402546137571335 + ], + "credit_map": [ + 0.019027123227715492, + 0.01429951936006546, + 0.024339010939002037, + 0.011469775810837746, + 0.019415689632296562, + 0.020952394232153893, + 0.019452158361673355, + 0.018433719873428345, + 0.016991762444376945, + 0.019948596134781837, + 0.012982169166207314, + 0.011813284829258919, + 0.019670238718390465, + 0.025082554668188095, + 0.02275083027780056, + 0.017281098291277885, + 0.0357978492975235, + 0.020396564155817032, + 0.022504162043333054, + 0.018343111500144005, + 0.02879839390516281, + 0.05444073677062988, + 0.022135810926556587, + 0.020073333755135536, + 0.00851563923060894, + 0.010643752291798592, + 0.006750843953341246, + 0.005302135832607746, + 0.011954140849411488, + 0.009988933801651001, + 0.006956466473639011, + 0.007416862063109875, + 0.008458300493657589, + 0.007329796906560659, + 0.007349131628870964, + 0.006079185754060745, + 0.006710784509778023, + 0.01876809261739254, + 0.009226434864103794, + 0.007323953788727522, + 0.006528795231133699, + 0.007286163978278637, + 0.005547743756324053, + 0.003918736707419157, + 0.009197392500936985, + 0.007348943036049604, + 0.005820121616125107, + 0.0067227245308458805, + 0.014845133759081364, + 0.00976178701967001, + 0.01601158268749714, + 0.005550232715904713, + 0.013056511990725994, + 0.011462869122624397, + 0.007308962754905224, + 0.009149441495537758, + 0.01015644520521164, + 0.009225311689078808, + 0.009521979838609695, + 0.00755362818017602, + 0.00759227154776454, + 0.010306179523468018, + 0.01237722672522068, + 0.00756004499271512, + 0.010759330354630947, + 0.007218227256089449, + 0.0075966548174619675, + 0.007386409677565098, + 0.006805613171309233, + 0.022849466651678085, + 0.0107529042288661, + 0.008244238793849945, + 0.007837158627808094, + 0.003744811285287142, + 0.0040998151525855064, + 0.007336060982197523, + 0.004564016126096249, + 0.00780714163556695, + 0.007954258471727371, + 0.006059280596673489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.043063513934612274 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle with a Dutch angle tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a prehistoric giant alligator snapping turtle with a dark brown shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle's shell feature three signature jagged ridges and appear covered in moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the turtle thrust its massive triangular head forward and snap its jaws open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action stir up a small patch of turbid sand around the turtle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979033812880516 + ], + "opposite_logprob_tokens": [ + -3.5297904014587402 + ], + "logprob_reward": -0.02979033812880516, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979033812880516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10814358294010162, + 0.043684717267751694, + 0.10586199909448624, + 0.04999888315796852, + 0.2917627990245819, + 0.13006965816020966, + 0.0487670861184597, + 0.130642369389534, + 0.06444907933473587, + 0.02661987394094467 + ], + "credit_map": [ + 0.01474510133266449, + 0.009952807798981667, + 0.02045051008462906, + 0.008285000920295715, + 0.012539519928395748, + 0.014914898201823235, + 0.010354934260249138, + 0.016900813207030296, + 0.004575113765895367, + 0.002142790937796235, + 0.009351094253361225, + 0.002151581458747387, + 0.004071240313351154, + 0.015146088786423206, + 0.0032451017759740353, + 0.0030017015524208546, + 0.015285137109458447, + 0.010310881771147251, + 0.014825264923274517, + 0.008701897226274014, + 0.00997329130768776, + 0.023318910971283913, + 0.00937729049474001, + 0.014069332741200924, + 0.006579315755516291, + 0.005540899466723204, + 0.003795687109231949, + 0.0031661451794207096, + 0.013810747303068638, + 0.005873392336070538, + 0.006107147783041, + 0.005125549156218767, + 0.023137381300330162, + 0.012402571737766266, + 0.05394672229886055, + 0.008847678080201149, + 0.011808897368609905, + 0.12091133743524551, + 0.03443416953086853, + 0.026274045929312706, + 0.011371650733053684, + 0.007317929994314909, + 0.027438560500741005, + 0.004554694052785635, + 0.006351931486278772, + 0.04841611906886101, + 0.013324690982699394, + 0.011294067837297916, + 0.007873030379414558, + 0.006662111263722181, + 0.004578474909067154, + 0.003665032796561718, + 0.007026246283203363, + 0.008221086114645004, + 0.005323362536728382, + 0.005417740438133478, + 0.018122779205441475, + 0.007482070475816727, + 0.025616580620408058, + 0.010992266237735748, + 0.008772204630076885, + 0.013957757502794266, + 0.031361449509859085, + 0.014337247237563133, + 0.0058242036029696465, + 0.004559918772429228, + 0.00980729516595602, + 0.005578851792961359, + 0.004766401834785938, + 0.01610998436808586, + 0.012093904428184032, + 0.005708519835025072, + 0.005528260488063097, + 0.001499043544754386, + 0.0035733014810830355, + 0.002780380193144083, + 0.0017274789279326797, + 0.004381588194519281, + 0.0038858649786561728, + 0.0032439548522233963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.02979033812880516 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_8/index_tensor([7003, 7003])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_8/index_tensor([7003, 7003])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9536eee914cd84661073df3cf8416d72e1e5fa0b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_8/index_tensor([7003, 7003])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, downward shot from a high vantage point captures an elderly man, meticulously combed and white-haired, immersed in Tai Chi practice beneath a twisted, vigorous ancient pine tree. Clad in a loose, pure white silk training suit, his face is serene, his eyes slightly closed, his cheeks lined with the wrinkles of time. Soft morning sunlight filters through the sparse pine needles, casting dappled shadows on the dark green ground and softly outlining the white sleeves of his robe. His arms trace smooth, flowing arcs as he executes a slow and steady \"cloud hand\" movement, his center of gravity shifting effortlessly between his legs. Each turn, hand lift, and palm push is imbued with an inner sense of power and harmonious rhythm. The dark brown bark of the ancient pine tree cracks like dragon scales, and its thick branches stretch out in all directions, forming a natural canopy. The air is filled with a faint pine scent and morning mist. The whole picture presents an extreme tranquility and concentration, full of the harmonious beauty and the artistic conception of unity between man and nature in Eastern philosophy.", + "seed": 27289, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_8/index_7003_seed_27289_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1708155658096075, + "vlm_logprob": -0.1708155658096075, + "vlm_yes_no_margin": 2.424999952316284, + "vlm_token_credit": -0.1708155658096075, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6083731055259705 + ], + "opposite_logprob_tokens": [ + -0.8583731055259705 + ], + "logprob_reward": -0.6083731055259705, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.6083731055259705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29101991653442383, + 0.15458618104457855, + 0.15518736839294434, + 0.0389217846095562, + 0.04640870913863182, + 0.08033379167318344, + 0.10376573354005814, + 0.03817877545952797, + 0.015641991049051285, + 0.07595579326152802 + ], + "credit_map": [ + 0.03290541097521782, + 0.03161568567156792, + 0.03761790692806244, + 0.0596834160387516, + 0.024223310872912407, + 0.03231963887810707, + 0.03659074380993843, + 0.03606380894780159, + 0.016862990334630013, + 0.013336929492652416, + 0.014165989123284817, + 0.011294749565422535, + 0.023269349709153175, + 0.035572949796915054, + 0.01325151789933443, + 0.02683171071112156, + 0.01892683655023575, + 0.010531889274716377, + 0.020619500428438187, + 0.010266577824950218, + 0.01666414737701416, + 0.0374671071767807, + 0.01120327040553093, + 0.02950802817940712, + 0.00458077248185873, + 0.007166234776377678, + 0.008483944460749626, + 0.0028295039664953947, + 0.004935124423354864, + 0.005551918409764767, + 0.0024029670748859644, + 0.002971319016069174, + 0.006888698786497116, + 0.008345713838934898, + 0.006892799399793148, + 0.003906467463821173, + 0.007182886824011803, + 0.006075861398130655, + 0.003517228178679943, + 0.0035990525502711535, + 0.008633822202682495, + 0.010457620024681091, + 0.023356638848781586, + 0.010771972127258778, + 0.005861630663275719, + 0.01291685737669468, + 0.004579117521643639, + 0.003756121965125203, + 0.015412183478474617, + 0.021511156111955643, + 0.01824704185128212, + 0.014338339678943157, + 0.008962871506810188, + 0.013659304939210415, + 0.006390758790075779, + 0.005244079511612654, + 0.005425161216408014, + 0.007631367538124323, + 0.006103899329900742, + 0.0031343833543360233, + 0.003852889407426119, + 0.006425623781979084, + 0.0028039049357175827, + 0.0028015465941280127, + 0.0022435036953538656, + 0.0016315735410898924, + 0.001715035643428564, + 0.0015548361698165536, + 0.00278439000248909, + 0.0022699085529893637, + 0.00135924038477242, + 0.0020835029426962137, + 0.005326504819095135, + 0.003450211137533188, + 0.011606053449213505, + 0.004230646416544914, + 0.004887846764177084, + 0.01647951826453209, + 0.004863467533141375, + 0.025111548602581024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.6083731055259705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909213542938232 + ], + "opposite_logprob_tokens": [ + -2.464092254638672 + ], + "logprob_reward": -0.08909213542938232, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909213542938232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25602635741233826, + 0.11683829128742218, + 0.11488346755504608, + 0.13116350769996643, + 0.04027227312326431, + 0.07098731398582458, + 0.09551484882831573, + 0.060165975242853165, + 0.04342721402645111, + 0.07072078436613083 + ], + "credit_map": [ + 0.0338730625808239, + 0.027515092864632607, + 0.025660358369350433, + 0.0463135652244091, + 0.029023174196481705, + 0.030954143032431602, + 0.032805830240249634, + 0.02988111600279808, + 0.013404999859631062, + 0.009166114032268524, + 0.01422900427132845, + 0.012122537940740585, + 0.01923372969031334, + 0.021085107699036598, + 0.010927733033895493, + 0.01666906662285328, + 0.013105272315442562, + 0.00919036753475666, + 0.014428814873099327, + 0.008782931603491306, + 0.01506937388330698, + 0.02494402788579464, + 0.010522911325097084, + 0.01883975975215435, + 0.014448936097323895, + 0.017902476713061333, + 0.01994483172893524, + 0.012344472110271454, + 0.02799006924033165, + 0.014499526470899582, + 0.012977530248463154, + 0.011055676266551018, + 0.0063149346970021725, + 0.005570242181420326, + 0.004520065616816282, + 0.003616690170019865, + 0.005280185025185347, + 0.005496324971318245, + 0.005569960456341505, + 0.0039038723334670067, + 0.007022167090326548, + 0.010472389869391918, + 0.013517862185835838, + 0.009682687930762768, + 0.008176695555448532, + 0.009654415771365166, + 0.0065107098780572414, + 0.00595038291066885, + 0.014806882478296757, + 0.008953412994742393, + 0.012551944702863693, + 0.014751635491847992, + 0.0156292412430048, + 0.012288554571568966, + 0.008652187883853912, + 0.007880990393459797, + 0.005015543196350336, + 0.008279508911073208, + 0.00780855817720294, + 0.009204003028571606, + 0.007382300216704607, + 0.008572828024625778, + 0.00632676063105464, + 0.00757646793499589, + 0.0048208292573690414, + 0.0034262214321643114, + 0.005369425751268864, + 0.00503766443580389, + 0.009175161831080914, + 0.0054296511225402355, + 0.004223623313009739, + 0.005944632925093174, + 0.007618846837431192, + 0.0045228092931210995, + 0.011811795644462109, + 0.008263521827757359, + 0.0048422194086015224, + 0.013027802109718323, + 0.005805899854749441, + 0.014827881008386612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.08909213542938232 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811015188694 + ], + "opposite_logprob_tokens": [ + -3.2881102561950684 + ], + "logprob_reward": -0.03811015188694, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811015188694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22232340276241302, + 0.20005017518997192, + 0.08814188092947006, + 0.0974043533205986, + 0.03715898469090462, + 0.05117349326610565, + 0.0687875747680664, + 0.05619209632277489, + 0.03389212489128113, + 0.1448759138584137 + ], + "credit_map": [ + 0.032964374870061874, + 0.028830111026763916, + 0.026255788281559944, + 0.024753840640187263, + 0.02383912168443203, + 0.03413552790880203, + 0.018184851855039597, + 0.03335979953408241, + 0.02156611531972885, + 0.02213376946747303, + 0.02152475342154503, + 0.017325349152088165, + 0.025892402976751328, + 0.0438193678855896, + 0.015296410769224167, + 0.03249200060963631, + 0.010232475586235523, + 0.010430581867694855, + 0.009241597726941109, + 0.005476223770529032, + 0.008934053592383862, + 0.018890436738729477, + 0.0067362855188548565, + 0.018200218677520752, + 0.009814174845814705, + 0.00944602582603693, + 0.01795501820743084, + 0.00921133253723383, + 0.02319733239710331, + 0.012302378192543983, + 0.008416097611188889, + 0.007061995565891266, + 0.005229508504271507, + 0.00396270165219903, + 0.007980777882039547, + 0.003486338071525097, + 0.005604161415249109, + 0.004880730994045734, + 0.003675619140267372, + 0.002339146798476577, + 0.004289770498871803, + 0.006551519501954317, + 0.010090967640280724, + 0.0075322482734918594, + 0.007337348069995642, + 0.006620893254876137, + 0.004236492794007063, + 0.004514255095273256, + 0.00863146036863327, + 0.008504490368068218, + 0.00898569542914629, + 0.009029817767441273, + 0.012579716742038727, + 0.008397568948566914, + 0.0071544526144862175, + 0.005504371598362923, + 0.0057311635464429855, + 0.006160069722682238, + 0.007470444310456514, + 0.005734368693083525, + 0.006999324541538954, + 0.007930980995297432, + 0.006809252314269543, + 0.009356493130326271, + 0.004649838898330927, + 0.002256350591778755, + 0.003948652185499668, + 0.003086991608142853, + 0.006331436801701784, + 0.00491164019331336, + 0.0024767047725617886, + 0.006230510771274567, + 0.01519183162599802, + 0.008761676028370857, + 0.023586655035614967, + 0.011831048876047134, + 0.009839443489909172, + 0.02745160646736622, + 0.012200692668557167, + 0.036012958735227585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03811015188694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188124522566795 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018188124522566795, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018188124522566795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1947661191225052, + 0.09143149852752686, + 0.2933277487754822, + 0.06298772990703583, + 0.14982913434505463, + 0.054483648389577866, + 0.05262408405542374, + 0.03770666569471359, + 0.029369009658694267, + 0.03347441181540489 + ], + "credit_map": [ + 0.0331251434981823, + 0.02170187048614025, + 0.025213757529854774, + 0.02133423276245594, + 0.02231530100107193, + 0.02683618851006031, + 0.014726110734045506, + 0.029513496905565262, + 0.009635543450713158, + 0.0069727604277431965, + 0.009482225403189659, + 0.00570888165384531, + 0.01470961607992649, + 0.021430522203445435, + 0.0069403275847435, + 0.016551613807678223, + 0.035196997225284576, + 0.023587139323353767, + 0.036981210112571716, + 0.01947936974465847, + 0.03720145300030708, + 0.06764890998601913, + 0.025180183351039886, + 0.04805246740579605, + 0.00527446111664176, + 0.010152338072657585, + 0.018097082152962685, + 0.004549714736640453, + 0.006980083882808685, + 0.009269515983760357, + 0.003716430626809597, + 0.00494809728115797, + 0.022761685773730278, + 0.021799737587571144, + 0.03974820300936699, + 0.010677391663193703, + 0.01126120612025261, + 0.02536868117749691, + 0.008764040656387806, + 0.009448183700442314, + 0.005000206176191568, + 0.010675514116883278, + 0.011184130795300007, + 0.008165369741618633, + 0.004816337488591671, + 0.007910792715847492, + 0.0036000271793454885, + 0.003131261793896556, + 0.007649728562682867, + 0.008124111220240593, + 0.007419412024319172, + 0.011342411860823631, + 0.0046719214878976345, + 0.006989326793700457, + 0.003688785945996642, + 0.002738382900133729, + 0.006348633673042059, + 0.005501766223460436, + 0.006447524763643742, + 0.004256554413586855, + 0.0042275781743228436, + 0.006130177062004805, + 0.0026591175701469183, + 0.002135310787707567, + 0.0050838422030210495, + 0.0038273106329143047, + 0.00433536060154438, + 0.0026729588862508535, + 0.004053962882608175, + 0.003995210397988558, + 0.0025549677666276693, + 0.002845392096787691, + 0.004457499831914902, + 0.0033629995305091143, + 0.006437336560338736, + 0.0027137366123497486, + 0.00306361704133451, + 0.005415122956037521, + 0.003439292311668396, + 0.004584805108606815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.018188124522566795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031431168317795 + ], + "opposite_logprob_tokens": [ + -2.3503143787384033 + ], + "logprob_reward": -0.10031431168317795, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031431168317795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17714449763298035, + 0.07987585663795471, + 0.36924731731414795, + 0.04742944985628128, + 0.0562276691198349, + 0.10051961988210678, + 0.086824432015419, + 0.0237346850335598, + 0.022612564265727997, + 0.036383870989084244 + ], + "credit_map": [ + 0.023494109511375427, + 0.019928958266973495, + 0.02240617573261261, + 0.032483384013175964, + 0.017304545268416405, + 0.023629069328308105, + 0.015601703897118568, + 0.022296566516160965, + 0.007093513384461403, + 0.00501194829121232, + 0.007014948409050703, + 0.005335595924407244, + 0.013734634965658188, + 0.01928526908159256, + 0.006428999826312065, + 0.015970943495631218, + 0.045334670692682266, + 0.02591574192047119, + 0.039555300027132034, + 0.01944907195866108, + 0.03888127952814102, + 0.0987464115023613, + 0.023515230044722557, + 0.07784963399171829, + 0.0036955238319933414, + 0.008493120782077312, + 0.012258431874215603, + 0.003631332889199257, + 0.006216684356331825, + 0.006437208037823439, + 0.0029599065892398357, + 0.0037372419610619545, + 0.008369079791009426, + 0.010717861354351044, + 0.011257897131145, + 0.005504919216036797, + 0.00472658546641469, + 0.007859276607632637, + 0.003029746003448963, + 0.004762301221489906, + 0.009116830304265022, + 0.01994728483259678, + 0.015834804624319077, + 0.017216291278600693, + 0.008407213725149632, + 0.01653125137090683, + 0.006446948740631342, + 0.007018999196588993, + 0.012286427430808544, + 0.013972357846796513, + 0.014249155297875404, + 0.016982607543468475, + 0.006356917787343264, + 0.011818641796708107, + 0.0060164956375956535, + 0.005141828209161758, + 0.004044007044285536, + 0.004660417325794697, + 0.003753198543563485, + 0.0024114069528877735, + 0.0018021747237071395, + 0.004544410854578018, + 0.0013578948564827442, + 0.0011611748486757278, + 0.004097368102520704, + 0.0025443993508815765, + 0.0037664356641471386, + 0.0023597260005772114, + 0.002345495158806443, + 0.0036721813958138227, + 0.0017607402987778187, + 0.002066217828541994, + 0.005633038468658924, + 0.002901374828070402, + 0.00676351273432374, + 0.003067225916311145, + 0.0032360637560486794, + 0.005548396147787571, + 0.003212015377357602, + 0.006022244226187468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.10031431168317795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, downward shot from a high vantage point captures an elderly man, meticulously combed and white-haired, immersed in Tai Chi practice beneath a twisted, vigorous ancient pine tree. Clad in a loose, pure white silk training suit, his face is serene, his eyes slightly closed, his cheeks lined with the wrinkles of time. Soft morning sunlight filters through the sparse pine needles, casting dappled shadows on the dark green ground and softly outlining the white sleeves of his robe. His arms trace smooth, flowing arcs as he executes a slow and steady \"cloud hand\" movement, his center of gravity shifting effortlessly between his legs. Each turn, hand lift, and palm push is imbued with an inner sense of power and harmonious rhythm. The dark brown bark of the ancient pine tree cracks like dragon scales, and its thick branches stretch out in all directions, forming a natural canopy. The air is filled with a faint pine scent and morning mist. The whole picture presents an extreme tranquility and concentration, full of the harmonious beauty and the artistic conception of unity between man and nature in Eastern philosophy.", + "seed": 27290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_8/index_7003_seed_27290_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09727572472766041, + "vlm_logprob": -0.09727572472766041, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.09727572472766041, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3390711843967438 + ], + "opposite_logprob_tokens": [ + -1.3390711545944214 + ], + "logprob_reward": -0.3390711843967438, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3390711843967438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2892388105392456, + 0.30913716554641724, + 0.02727574110031128, + 0.053691793233156204, + 0.0466194786131382, + 0.026995735242962837, + 0.04493431746959686, + 0.03909261152148247, + 0.06432666629552841, + 0.09868776053190231 + ], + "credit_map": [ + 0.03420466557145119, + 0.04195639491081238, + 0.03309930860996246, + 0.02368343062698841, + 0.02500314824283123, + 0.041056204587221146, + 0.0420103445649147, + 0.0482252798974514, + 0.024315882474184036, + 0.019352730363607407, + 0.01549603696912527, + 0.01808861270546913, + 0.015055772848427296, + 0.11065815389156342, + 0.08336801826953888, + 0.02280193567276001, + 0.003944339696317911, + 0.0063097476959228516, + 0.003108095610514283, + 0.0029001384973526, + 0.0033984659239649773, + 0.0028813243843615055, + 0.002594840247184038, + 0.0021387862507253885, + 0.006371129769831896, + 0.004965850617736578, + 0.007342774886637926, + 0.003790489863604307, + 0.009864982217550278, + 0.00596275832504034, + 0.0062749916687607765, + 0.00911881122738123, + 0.004996750969439745, + 0.004413574002683163, + 0.006059086415916681, + 0.007334851194173098, + 0.009075025096535683, + 0.004963529296219349, + 0.005469838157296181, + 0.0043068211525678635, + 0.004607241600751877, + 0.002362856874242425, + 0.0038236977998167276, + 0.003243151120841503, + 0.004336354322731495, + 0.003264947095885873, + 0.002293204888701439, + 0.003064278746023774, + 0.006129500921815634, + 0.008485190570354462, + 0.006360306870192289, + 0.008271287195384502, + 0.0034378189593553543, + 0.003723221831023693, + 0.005431796424090862, + 0.0030951891094446182, + 0.005254736635833979, + 0.005600686650723219, + 0.004825143609195948, + 0.005217623896896839, + 0.0038770262617617846, + 0.005183265078812838, + 0.005550378933548927, + 0.0035837446339428425, + 0.011552435345947742, + 0.006134675350040197, + 0.011857601813971996, + 0.005827673710882664, + 0.005041175987571478, + 0.01025801058858633, + 0.004193695727735758, + 0.00946139544248581, + 0.011978515423834324, + 0.015731099992990494, + 0.015447869896888733, + 0.009143860079348087, + 0.011970385909080505, + 0.01241823099553585, + 0.011018402874469757, + 0.01097938884049654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.3390711843967438 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430946871638298 + ], + "opposite_logprob_tokens": [ + -3.1680946350097656 + ], + "logprob_reward": -0.0430946871638298, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430946871638298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16392123699188232, + 0.14501243829727173, + 0.026268338784575462, + 0.02577749826014042, + 0.021095041185617447, + 0.026001237332820892, + 0.07018781453371048, + 0.06077701225876808, + 0.050653133541345596, + 0.4103062152862549 + ], + "credit_map": [ + 0.01524907536804676, + 0.01901705004274845, + 0.017389625310897827, + 0.020119475200772285, + 0.022383974865078926, + 0.01848052255809307, + 0.033498138189315796, + 0.017783373594284058, + 0.01211143471300602, + 0.010898962616920471, + 0.009433863684535027, + 0.009916692972183228, + 0.010334525257349014, + 0.045296560972929, + 0.03682196885347366, + 0.010198424570262432, + 0.0032991115003824234, + 0.003180562984198332, + 0.002600037958472967, + 0.0022667590528726578, + 0.006199212744832039, + 0.0025373767130076885, + 0.00307907210662961, + 0.003106203628703952, + 0.005662784446030855, + 0.0023663563188165426, + 0.004191095940768719, + 0.0020760332699865103, + 0.0026925019919872284, + 0.0031998860649764538, + 0.002082621678709984, + 0.003506216686218977, + 0.0021639750339090824, + 0.0034136346075683832, + 0.0030625679064542055, + 0.0026737949810922146, + 0.003507268149405718, + 0.002405512845143676, + 0.0018446727190166712, + 0.002023613080382347, + 0.0035291288513690233, + 0.0022668929304927588, + 0.0035907491110265255, + 0.004414056893438101, + 0.003818521974608302, + 0.0030949015635997057, + 0.002849109936505556, + 0.002437877468764782, + 0.004248943645507097, + 0.023139620199799538, + 0.0053867134265601635, + 0.02442355826497078, + 0.0020661598537117243, + 0.0024818223901093006, + 0.00605762517079711, + 0.002383374609053135, + 0.007443523965775967, + 0.008725516498088837, + 0.009021845646202564, + 0.0089497622102499, + 0.005284193903207779, + 0.008218424394726753, + 0.008297014981508255, + 0.004836731590330601, + 0.005367623642086983, + 0.006003942806273699, + 0.0075979819521307945, + 0.004833446349948645, + 0.00822637602686882, + 0.006563304923474789, + 0.005530777852982283, + 0.006529677659273148, + 0.03397700935602188, + 0.05651412904262543, + 0.04048488289117813, + 0.043978262692689896, + 0.08957650512456894, + 0.06002587080001831, + 0.0514078252017498, + 0.03434176370501518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0430946871638298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591415464878082 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591415464878082, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591415464878082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10616035014390945, + 0.2621535658836365, + 0.07695404440164566, + 0.02862912230193615, + 0.0483475998044014, + 0.019225435331463814, + 0.054126497358083725, + 0.13425345718860626, + 0.07042766362428665, + 0.19972220063209534 + ], + "credit_map": [ + 0.010484986938536167, + 0.02097443863749504, + 0.009220006875693798, + 0.008983120322227478, + 0.010746976360678673, + 0.01454742718487978, + 0.017798181623220444, + 0.013405213132500648, + 0.02256411872804165, + 0.024087142199277878, + 0.02565968781709671, + 0.017487119883298874, + 0.015449938364326954, + 0.08050497621297836, + 0.060481440275907516, + 0.015919120982289314, + 0.012462688609957695, + 0.009313051588833332, + 0.010309086181223392, + 0.006586472969502211, + 0.014773356728255749, + 0.007746244315057993, + 0.007321350742131472, + 0.008441788144409657, + 0.005793862044811249, + 0.004075661767274141, + 0.005016086157411337, + 0.0023654394317418337, + 0.002661816542968154, + 0.003881770884618163, + 0.0018954337574541569, + 0.0029390505515038967, + 0.007062570191919804, + 0.013506842777132988, + 0.007673504296690226, + 0.004006123170256615, + 0.0046037836000323296, + 0.005426777061074972, + 0.0029129046015441418, + 0.0031550906132906675, + 0.0024872850626707077, + 0.001676082843914628, + 0.0024714593309909105, + 0.0029314165003597736, + 0.0023894086480140686, + 0.0023197843693196774, + 0.002348576905205846, + 0.0026014214381575584, + 0.004866140894591808, + 0.015165709890425205, + 0.005732591263949871, + 0.015728047117590904, + 0.0023251234088093042, + 0.00357593665830791, + 0.003947356715798378, + 0.0027855951339006424, + 0.020230868831276894, + 0.015477144159376621, + 0.02967291697859764, + 0.015992864966392517, + 0.010938910767436028, + 0.01654626615345478, + 0.013408548198640347, + 0.01198593433946371, + 0.009615125134587288, + 0.005554489325731993, + 0.014334413222968578, + 0.007491989526897669, + 0.00857741292566061, + 0.01011652871966362, + 0.0052969492971897125, + 0.009440756402909756, + 0.014992226846516132, + 0.025754323229193687, + 0.01979932188987732, + 0.021148277446627617, + 0.0476081408560276, + 0.029041172936558723, + 0.023089637979865074, + 0.018289100378751755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.020591415464878082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046369642950594425 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.0046369642950594425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046369642950594425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10300279408693314, + 0.48725536465644836, + 0.02217722497880459, + 0.10876629501581192, + 0.0390920527279377, + 0.030490899458527565, + 0.06141390651464462, + 0.023587921634316444, + 0.023240454494953156, + 0.10097299516201019 + ], + "credit_map": [ + 0.009389501065015793, + 0.016003146767616272, + 0.013070924207568169, + 0.006966190412640572, + 0.015602489933371544, + 0.01567789725959301, + 0.013170965947210789, + 0.01312168873846531, + 0.06790315359830856, + 0.07267677038908005, + 0.08580534160137177, + 0.029066525399684906, + 0.03373266011476517, + 0.10293933749198914, + 0.06864325702190399, + 0.026488373056054115, + 0.002183812903240323, + 0.0020424623508006334, + 0.0023957369849085808, + 0.0020397778134793043, + 0.005246989894658327, + 0.0020881090313196182, + 0.002911964664235711, + 0.003268374130129814, + 0.026178408414125443, + 0.016990911215543747, + 0.019779467955231667, + 0.008498096838593483, + 0.007139171008020639, + 0.01725912280380726, + 0.00662067299708724, + 0.006300461944192648, + 0.003753786673769355, + 0.009910093620419502, + 0.0055518546141684055, + 0.004526558797806501, + 0.0045685842633247375, + 0.004889749456197023, + 0.0026783307548612356, + 0.0032130987383425236, + 0.003931333310902119, + 0.002607346512377262, + 0.0056435526348650455, + 0.002590808318927884, + 0.0022961413487792015, + 0.006991864647716284, + 0.002300554420799017, + 0.004129303619265556, + 0.0036273598670959473, + 0.01934001035988331, + 0.002941441722214222, + 0.022527774795889854, + 0.0018043919699266553, + 0.002292085438966751, + 0.006465822458267212, + 0.002415024908259511, + 0.0025581982918083668, + 0.004586203955113888, + 0.0030166644137352705, + 0.0032050921581685543, + 0.001950356992892921, + 0.003252709284424782, + 0.002996681723743677, + 0.002022017491981387, + 0.0036643927451223135, + 0.0023691102396696806, + 0.00370008940808475, + 0.002103848848491907, + 0.002380883786827326, + 0.0032111997716128826, + 0.002180421259254217, + 0.0036305116955190897, + 0.010624002665281296, + 0.013614687137305737, + 0.010402562096714973, + 0.01022351998835802, + 0.02084978111088276, + 0.013884673826396465, + 0.012166998349130154, + 0.009206784889101982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0046369642950594425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898437231779099 + ], + "opposite_logprob_tokens": [ + -2.578984260559082 + ], + "logprob_reward": -0.07898437231779099, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898437231779099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4674517810344696, + 0.26552078127861023, + 0.019700992852449417, + 0.02799590677022934, + 0.04936351627111435, + 0.05386793240904808, + 0.04601951315999031, + 0.015578391961753368, + 0.02552580088376999, + 0.02897541970014572 + ], + "credit_map": [ + 0.048198144882917404, + 0.05046326294541359, + 0.04682696610689163, + 0.046741586178541183, + 0.04594989866018295, + 0.06151869520545006, + 0.08243657648563385, + 0.08531661331653595, + 0.02302296832203865, + 0.020194977521896362, + 0.01594795472919941, + 0.01489600446075201, + 0.013263815082609653, + 0.09301815927028656, + 0.06752959638834, + 0.01764732599258423, + 0.0030811873730272055, + 0.0022783251479268074, + 0.0023888174910098314, + 0.0019600626546889544, + 0.0022448853123933077, + 0.0023385551758110523, + 0.0019411796238273382, + 0.0034679805394262075, + 0.0024492626544088125, + 0.005526288878172636, + 0.005235833115875721, + 0.001940548187121749, + 0.0038892896845936775, + 0.0037838248535990715, + 0.0021435567177832127, + 0.0030273026786744595, + 0.00492100790143013, + 0.0054756756871938705, + 0.006394901778548956, + 0.006557374726980925, + 0.007287520449608564, + 0.006066251080483198, + 0.0048430017195641994, + 0.007817782461643219, + 0.01006720494478941, + 0.0046806796453893185, + 0.011206052266061306, + 0.004046099726110697, + 0.004691217094659805, + 0.009261730127036572, + 0.0033561387099325657, + 0.00655880942940712, + 0.0030885753221809864, + 0.014318888075649738, + 0.002689255168661475, + 0.016562003642320633, + 0.0014763899380341172, + 0.0018765911227092147, + 0.004591548349708319, + 0.0014162599109113216, + 0.0020287332590669394, + 0.002939817262813449, + 0.003344115102663636, + 0.0014446941204369068, + 0.001477781101129949, + 0.00141389190685004, + 0.001302755088545382, + 0.0016266044694930315, + 0.0033341769594699144, + 0.0024373552296310663, + 0.00391771923750639, + 0.0016751858638599515, + 0.0019663339480757713, + 0.004911066964268684, + 0.0017333000432699919, + 0.005550660192966461, + 0.003042335854843259, + 0.004333280958235264, + 0.004549902398139238, + 0.002657894976437092, + 0.002669115085154772, + 0.004006111994385719, + 0.004427216947078705, + 0.003289562650024891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07898437231779099 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, downward shot from a high vantage point captures an elderly man, meticulously combed and white-haired, immersed in Tai Chi practice beneath a twisted, vigorous ancient pine tree. Clad in a loose, pure white silk training suit, his face is serene, his eyes slightly closed, his cheeks lined with the wrinkles of time. Soft morning sunlight filters through the sparse pine needles, casting dappled shadows on the dark green ground and softly outlining the white sleeves of his robe. His arms trace smooth, flowing arcs as he executes a slow and steady \"cloud hand\" movement, his center of gravity shifting effortlessly between his legs. Each turn, hand lift, and palm push is imbued with an inner sense of power and harmonious rhythm. The dark brown bark of the ancient pine tree cracks like dragon scales, and its thick branches stretch out in all directions, forming a natural canopy. The air is filled with a faint pine scent and morning mist. The whole picture presents an extreme tranquility and concentration, full of the harmonious beauty and the artistic conception of unity between man and nature in Eastern philosophy.", + "seed": 27291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_8/index_7003_seed_27291_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.33004971873015165, + "vlm_logprob": -0.33004971873015165, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.33004971873015165, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.512353777885437 + ], + "opposite_logprob_tokens": [ + -0.262353777885437 + ], + "logprob_reward": -1.512353777885437, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.512353777885437, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10679756104946136, + 0.14083296060562134, + 0.0407312773168087, + 0.16459371149539948, + 0.14428181946277618, + 0.07480476051568985, + 0.06461033225059509, + 0.08110298961400986, + 0.12167427688837051, + 0.060570280998945236 + ], + "credit_map": [ + 0.01012973953038454, + 0.011225543916225433, + 0.01301524043083191, + 0.011843130923807621, + 0.009021655656397343, + 0.01942511834204197, + 0.015438827686011791, + 0.016698308289051056, + 0.021476389840245247, + 0.016632389277219772, + 0.021100882440805435, + 0.02165883779525757, + 0.014570532366633415, + 0.016475943848490715, + 0.016309408470988274, + 0.012608584947884083, + 0.0031138109043240547, + 0.003177163191139698, + 0.0043356819078326225, + 0.005509532988071442, + 0.007219030987471342, + 0.007627390790730715, + 0.005815555807203054, + 0.003933114465326071, + 0.01120767556130886, + 0.007300843019038439, + 0.013465418480336666, + 0.024898668751120567, + 0.0320902094244957, + 0.030221417546272278, + 0.02959347702562809, + 0.01581602171063423, + 0.007792639080435038, + 0.007875890471041203, + 0.029807917773723602, + 0.018598277121782303, + 0.03111637756228447, + 0.02284182421863079, + 0.014060422778129578, + 0.012188472785055637, + 0.009348111227154732, + 0.010159685276448727, + 0.010800634510815144, + 0.00737992487847805, + 0.011135836131870747, + 0.01260604802519083, + 0.004340736195445061, + 0.00903378613293171, + 0.007359758950769901, + 0.01009218581020832, + 0.008992061950266361, + 0.009718349203467369, + 0.007369727361947298, + 0.01024330873042345, + 0.003657769178971648, + 0.0071771759539842606, + 0.01276305876672268, + 0.007848424836993217, + 0.014812084846198559, + 0.006718493066728115, + 0.006279815919697285, + 0.015029000118374825, + 0.005018309690058231, + 0.012633809819817543, + 0.010408570058643818, + 0.014745616354048252, + 0.013870053924620152, + 0.01128717977553606, + 0.010517645627260208, + 0.016451099887490273, + 0.009624670259654522, + 0.0347694493830204, + 0.005891619250178337, + 0.007725575938820839, + 0.004308326635509729, + 0.005641466937959194, + 0.007069166749715805, + 0.006729452405124903, + 0.011944078840315342, + 0.011260597966611385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -1.512353777885437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038138497620821 + ], + "opposite_logprob_tokens": [ + -3.2881383895874023 + ], + "logprob_reward": -0.038138497620821, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038138497620821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09395216405391693, + 0.11986921727657318, + 0.2241193652153015, + 0.04339318722486496, + 0.1547488123178482, + 0.036124784499406815, + 0.053960975259542465, + 0.15769606828689575, + 0.06507007032632828, + 0.051065389066934586 + ], + "credit_map": [ + 0.007117991801351309, + 0.007082770112901926, + 0.009604599326848984, + 0.013269427232444286, + 0.007721803151071072, + 0.02573179267346859, + 0.011214769445359707, + 0.012209009379148483, + 0.013891520909965038, + 0.012698598206043243, + 0.017401952296495438, + 0.021787989884614944, + 0.014913346618413925, + 0.011979705654084682, + 0.015754515305161476, + 0.011441593058407307, + 0.018907558172941208, + 0.013020485639572144, + 0.022563138976693153, + 0.03839358687400818, + 0.03220587596297264, + 0.04270802438259125, + 0.03280743211507797, + 0.023513250052928925, + 0.003070435021072626, + 0.0023539282847195864, + 0.003733929945155978, + 0.005973090883344412, + 0.008029980584979057, + 0.008730972185730934, + 0.007100042421370745, + 0.004400807432830334, + 0.0070822094567120075, + 0.008647595532238483, + 0.023366384208202362, + 0.020764760673046112, + 0.037477217614650726, + 0.022681772708892822, + 0.02093617245554924, + 0.013792683370411396, + 0.004349044989794493, + 0.0035491648595780134, + 0.006932248827069998, + 0.004611562937498093, + 0.0044237831607460976, + 0.006422223057597876, + 0.0019103483064100146, + 0.00392640708014369, + 0.00765244010835886, + 0.007496893405914307, + 0.0067067695781588554, + 0.006461767014116049, + 0.005877367686480284, + 0.0065572215244174, + 0.005775490775704384, + 0.007433023769408464, + 0.020848672837018967, + 0.019172988831996918, + 0.021745184436440468, + 0.01607579179108143, + 0.016941513866186142, + 0.03524086996912956, + 0.011860430240631104, + 0.015810614451766014, + 0.0067048571072518826, + 0.0109656797721982, + 0.006811457220464945, + 0.006346657872200012, + 0.006467961706221104, + 0.010129628702998161, + 0.005540426820516586, + 0.012103402987122536, + 0.00629472266882658, + 0.005827337969094515, + 0.004978615790605545, + 0.0029107811860740185, + 0.008309785276651382, + 0.005981448106467724, + 0.006858533248305321, + 0.009904162026941776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.038138497620821 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371792286634445 + ], + "opposite_logprob_tokens": [ + -3.4087178707122803 + ], + "logprob_reward": -0.03371792286634445, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371792286634445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11020733416080475, + 0.09479665011167526, + 0.09833651781082153, + 0.12454497814178467, + 0.0626264438033104, + 0.04976838827133179, + 0.08356353640556335, + 0.08067499101161957, + 0.1658972203731537, + 0.12958396971225739 + ], + "credit_map": [ + 0.0112015251070261, + 0.00789816863834858, + 0.01326719205826521, + 0.011166644282639027, + 0.009993602521717548, + 0.022597258910536766, + 0.011826622299849987, + 0.02225632034242153, + 0.008109125308692455, + 0.01025259681046009, + 0.014324131421744823, + 0.01744926907122135, + 0.009979126043617725, + 0.011754224076867104, + 0.012122256681323051, + 0.010805911384522915, + 0.009526179172098637, + 0.008911436423659325, + 0.01125011034309864, + 0.007503490429371595, + 0.018865594640374184, + 0.011646725237369537, + 0.01692540943622589, + 0.013707569800317287, + 0.00783702451735735, + 0.006264394614845514, + 0.009554554708302021, + 0.01589595153927803, + 0.026033340021967888, + 0.022595694288611412, + 0.023701732978224754, + 0.012662290595471859, + 0.00405900226905942, + 0.004292507190257311, + 0.0118101192638278, + 0.008067145943641663, + 0.014434974640607834, + 0.00910897459834814, + 0.005729460157454014, + 0.005124263931065798, + 0.005221187137067318, + 0.006853954866528511, + 0.007251609582453966, + 0.005132521968334913, + 0.007694676518440247, + 0.00948998425155878, + 0.0023461151868104935, + 0.005778340622782707, + 0.017713354900479317, + 0.010181517340242863, + 0.011313393712043762, + 0.013066616840660572, + 0.008625640533864498, + 0.008952738717198372, + 0.00612210389226675, + 0.007588176056742668, + 0.008710497990250587, + 0.008051943965256214, + 0.01500336080789566, + 0.006864186376333237, + 0.005982527043670416, + 0.018627936020493507, + 0.005041250493377447, + 0.012393281795084476, + 0.023993220180273056, + 0.02345035783946514, + 0.02200985886156559, + 0.015778297558426857, + 0.01718498207628727, + 0.025858404114842415, + 0.01250810269266367, + 0.025113975629210472, + 0.0191816296428442, + 0.010770224034786224, + 0.012458577752113342, + 0.009497471153736115, + 0.00930455420166254, + 0.014818239025771618, + 0.020853063091635704, + 0.03270020708441734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.03371792286634445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14460937678813934, + 0.2447815090417862, + 0.11318983882665634, + 0.05791323259472847, + 0.21773867309093475, + 0.06253091990947723, + 0.02203705534338951, + 0.021154042333364487, + 0.06846894323825836, + 0.04757647216320038 + ], + "credit_map": [ + 0.01436716876924038, + 0.020182408392429352, + 0.019014770165085793, + 0.02066996693611145, + 0.013273603282868862, + 0.023051749914884567, + 0.019084610044956207, + 0.014965096488595009, + 0.026739714667201042, + 0.03972785174846649, + 0.04958903044462204, + 0.03839629888534546, + 0.02652260847389698, + 0.021542605012655258, + 0.025947624817490578, + 0.016315756365656853, + 0.015185622498393059, + 0.017398586496710777, + 0.024335073307156563, + 0.011799965985119343, + 0.011356305330991745, + 0.016875099390745163, + 0.007845069281756878, + 0.008394111879169941, + 0.004247131757438183, + 0.003998214844614267, + 0.005179310217499733, + 0.007740722503513098, + 0.011119667440652847, + 0.010187673382461071, + 0.009416719898581505, + 0.006023786496371031, + 0.01304392796009779, + 0.01615021750330925, + 0.04883619025349617, + 0.025513319298624992, + 0.041379138827323914, + 0.03823580592870712, + 0.018260309472680092, + 0.016319748014211655, + 0.009584391489624977, + 0.004425722640007734, + 0.012286601588129997, + 0.004109266214072704, + 0.004587468225508928, + 0.013650264590978622, + 0.003614321118220687, + 0.010272872634232044, + 0.002585382666438818, + 0.0031318441033363342, + 0.002822431968525052, + 0.0024671838618814945, + 0.0027808661106973886, + 0.003012646920979023, + 0.0025151753798127174, + 0.0027215213049203157, + 0.0052862465381622314, + 0.002212927443906665, + 0.002540884306654334, + 0.0018494119867682457, + 0.001711846562102437, + 0.0029541957192122936, + 0.0016225373838096857, + 0.0029759930912405252, + 0.006723571568727493, + 0.009906194172799587, + 0.0059370375238358974, + 0.00490155816078186, + 0.006707814987748861, + 0.010312482714653015, + 0.005498930346220732, + 0.01848134957253933, + 0.008239722810685635, + 0.007338607683777809, + 0.002554570557549596, + 0.003128547454252839, + 0.00755226518958807, + 0.008222658187150955, + 0.005285380873829126, + 0.005254717078059912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0110772168263793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054961178451776505 + ], + "opposite_logprob_tokens": [ + -2.9299612045288086 + ], + "logprob_reward": -0.054961178451776505, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054961178451776505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13599619269371033, + 0.11212701350450516, + 0.0723273754119873, + 0.08949495851993561, + 0.16072441637516022, + 0.19841407239437103, + 0.04085485637187958, + 0.035712745040655136, + 0.07828580588102341, + 0.07606251537799835 + ], + "credit_map": [ + 0.017722440883517265, + 0.016250159591436386, + 0.021600445732474327, + 0.01386543270200491, + 0.014580790884792805, + 0.017097417265176773, + 0.02342778630554676, + 0.011451726779341698, + 0.009680851362645626, + 0.014399443753063679, + 0.02043907716870308, + 0.01843477226793766, + 0.01439700648188591, + 0.011364929378032684, + 0.014027641154825687, + 0.009383289143443108, + 0.008748789317905903, + 0.007914789021015167, + 0.014113212004303932, + 0.008336126804351807, + 0.008840223774313927, + 0.011899663135409355, + 0.006603075657039881, + 0.005871501285582781, + 0.0067978110164403915, + 0.007985946722328663, + 0.009981021285057068, + 0.010907755233347416, + 0.01567276567220688, + 0.016256287693977356, + 0.013163991272449493, + 0.008729369379580021, + 0.007780388463288546, + 0.010322690941393375, + 0.020172415301203728, + 0.021382896229624748, + 0.039732132107019424, + 0.024177169427275658, + 0.02155732735991478, + 0.015599394217133522, + 0.023632211610674858, + 0.01756802387535572, + 0.04264051094651222, + 0.014618603512644768, + 0.01180200558155775, + 0.05620642006397247, + 0.007726637180894613, + 0.024219656363129616, + 0.006771822459995747, + 0.0045920624397695065, + 0.006186157930642366, + 0.0037409141659736633, + 0.0038405617233365774, + 0.004951375536620617, + 0.0038286675699055195, + 0.006943290587514639, + 0.007710048463195562, + 0.005036284681409597, + 0.004353500436991453, + 0.0028532110154628754, + 0.002712822984904051, + 0.006614151876419783, + 0.0019336924888193607, + 0.004499032627791166, + 0.008371256291866302, + 0.010041692294180393, + 0.0067129069939255714, + 0.005566698499023914, + 0.008646976202726364, + 0.013231286779046059, + 0.007074851077049971, + 0.01864014007151127, + 0.008536715991795063, + 0.010541628114879131, + 0.0052621690556406975, + 0.006218181923031807, + 0.011592981405556202, + 0.01662103645503521, + 0.007453990634530783, + 0.009835811331868172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.054961178451776505 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, downward shot from a high vantage point captures an elderly man, meticulously combed and white-haired, immersed in Tai Chi practice beneath a twisted, vigorous ancient pine tree. Clad in a loose, pure white silk training suit, his face is serene, his eyes slightly closed, his cheeks lined with the wrinkles of time. Soft morning sunlight filters through the sparse pine needles, casting dappled shadows on the dark green ground and softly outlining the white sleeves of his robe. His arms trace smooth, flowing arcs as he executes a slow and steady \"cloud hand\" movement, his center of gravity shifting effortlessly between his legs. Each turn, hand lift, and palm push is imbued with an inner sense of power and harmonious rhythm. The dark brown bark of the ancient pine tree cracks like dragon scales, and its thick branches stretch out in all directions, forming a natural canopy. The air is filled with a faint pine scent and morning mist. The whole picture presents an extreme tranquility and concentration, full of the harmonious beauty and the artistic conception of unity between man and nature in Eastern philosophy.", + "seed": 27292, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_8/index_7003_seed_27292_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.32524140663444995, + "vlm_logprob": -0.32524140663444995, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.32524140663444995, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4203579425811768 + ], + "opposite_logprob_tokens": [ + -0.29535794258117676 + ], + "logprob_reward": -1.4203579425811768, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4203579425811768, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11959423869848251, + 0.2018706202507019, + 0.08793144673109055, + 0.08466777205467224, + 0.07699718326330185, + 0.059968411922454834, + 0.03760421648621559, + 0.0557943657040596, + 0.2148537039756775, + 0.06071805581450462 + ], + "credit_map": [ + 0.010619220323860645, + 0.023080984130501747, + 0.014000830240547657, + 0.015833625569939613, + 0.01266677025705576, + 0.013265491463243961, + 0.015131779946386814, + 0.014995529316365719, + 0.016917448490858078, + 0.026337282732129097, + 0.021716034039855003, + 0.02227291278541088, + 0.03134147450327873, + 0.022137142717838287, + 0.035801272839307785, + 0.025347057729959488, + 0.008611399680376053, + 0.006455474998801947, + 0.015744712203741074, + 0.02483002468943596, + 0.006754016969352961, + 0.009051918983459473, + 0.011814039200544357, + 0.0046698590740561485, + 0.005889711435884237, + 0.008158489130437374, + 0.013244344852864742, + 0.006767273880541325, + 0.0128472363576293, + 0.009572313167154789, + 0.01018657349050045, + 0.018001828342676163, + 0.009942922741174698, + 0.006011792924255133, + 0.01128037367016077, + 0.006094696931540966, + 0.00925449375063181, + 0.009496589191257954, + 0.007089436519891024, + 0.017826883122324944, + 0.00955132395029068, + 0.005913150496780872, + 0.00823235884308815, + 0.004339021630585194, + 0.00788425374776125, + 0.005692957900464535, + 0.00554182194173336, + 0.012813521549105644, + 0.0048932526260614395, + 0.007475678343325853, + 0.004756189417093992, + 0.004518055357038975, + 0.004464478697627783, + 0.0034759475383907557, + 0.003376200795173645, + 0.004644414409995079, + 0.006746127735823393, + 0.004992760717868805, + 0.009329505264759064, + 0.005200773943215609, + 0.005251438356935978, + 0.005312580615282059, + 0.010437024757266045, + 0.008524157106876373, + 0.02303115651011467, + 0.011623392812907696, + 0.07515311986207962, + 0.027805335819721222, + 0.012072693556547165, + 0.010715086944401264, + 0.009570532478392124, + 0.0448824018239975, + 0.00585902389138937, + 0.008031056262552738, + 0.00757541274651885, + 0.008681810460984707, + 0.0045828986912965775, + 0.00942012295126915, + 0.007076485548168421, + 0.009491242468357086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -1.4203579425811768 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790407732129097 + ], + "opposite_logprob_tokens": [ + -2.579040765762329 + ], + "logprob_reward": -0.0790407732129097, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790407732129097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09960275888442993, + 0.09148552268743515, + 0.16468197107315063, + 0.03634761646389961, + 0.11919239908456802, + 0.030875056982040405, + 0.02381242625415325, + 0.023974573239684105, + 0.29006361961364746, + 0.11996404081583023 + ], + "credit_map": [ + 0.011192284524440765, + 0.024887286126613617, + 0.011538577266037464, + 0.008127530105412006, + 0.011731648817658424, + 0.008764890022575855, + 0.010691607370972633, + 0.012668930925428867, + 0.011879204772412777, + 0.013167206197977066, + 0.010624521411955357, + 0.00756089948117733, + 0.016334855929017067, + 0.012461032718420029, + 0.009673784486949444, + 0.009784016758203506, + 0.01034645363688469, + 0.009945138357579708, + 0.04066471382975578, + 0.05697189271450043, + 0.006225061137229204, + 0.014594082720577717, + 0.0215512216091156, + 0.004383412655442953, + 0.0042394609190523624, + 0.004067259840667248, + 0.004986318293958902, + 0.003677526256069541, + 0.005372883751988411, + 0.0037644102703779936, + 0.005298726260662079, + 0.004941028542816639, + 0.013160983100533485, + 0.00892590917646885, + 0.018314959481358528, + 0.009970246814191341, + 0.017589358612895012, + 0.013773526065051556, + 0.009394864551723003, + 0.028062550351023674, + 0.005047498736530542, + 0.003646103898063302, + 0.0045529925264418125, + 0.0022404061164706945, + 0.004836658015847206, + 0.0036918336991220713, + 0.0024468915071338415, + 0.004412672948092222, + 0.0032631331123411655, + 0.0021300960797816515, + 0.0033313476014882326, + 0.002244556788355112, + 0.0037827519699931145, + 0.0029154340736567974, + 0.0030405402649194, + 0.0031045640353113413, + 0.00210502278059721, + 0.0023111356422305107, + 0.004141905345022678, + 0.0026481521781533957, + 0.0019982783123850822, + 0.002281242050230503, + 0.0050254035741090775, + 0.0034634326584637165, + 0.01817389950156212, + 0.016405705362558365, + 0.10740474611520767, + 0.04758305475115776, + 0.011774628423154354, + 0.012694032862782478, + 0.016591032966971397, + 0.05943651497364044, + 0.011179095134139061, + 0.015669776126742363, + 0.02175837755203247, + 0.013411741703748703, + 0.014664720743894577, + 0.010804058983922005, + 0.018191006034612656, + 0.01428526733070612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.0790407732129097 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867839813232422 + ], + "opposite_logprob_tokens": [ + -3.048678398132324 + ], + "logprob_reward": -0.04867839813232422, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867839813232422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11339011788368225, + 0.30642226338386536, + 0.11741158366203308, + 0.07781420648097992, + 0.026451395824551582, + 0.0568358413875103, + 0.020102370530366898, + 0.04606008529663086, + 0.15925508737564087, + 0.07625703513622284 + ], + "credit_map": [ + 0.010753074660897255, + 0.020332353189587593, + 0.014978567138314247, + 0.01657181978225708, + 0.008008784614503384, + 0.01192215085029602, + 0.015662511810660362, + 0.015160856768488884, + 0.020086031407117844, + 0.02874636836349964, + 0.03724926710128784, + 0.03985167667269707, + 0.0362900085747242, + 0.03419345244765282, + 0.06316189467906952, + 0.04684358462691307, + 0.008085659705102444, + 0.007423610892146826, + 0.02709093689918518, + 0.03992851823568344, + 0.005364115349948406, + 0.010221543721854687, + 0.015281536616384983, + 0.004015661776065826, + 0.007846028544008732, + 0.008936134167015553, + 0.009140951558947563, + 0.010091726668179035, + 0.012124134227633476, + 0.01013604924082756, + 0.00983906164765358, + 0.009700126014649868, + 0.0027945530600845814, + 0.0028588189743459225, + 0.0038568631280213594, + 0.0023621220607310534, + 0.003654340747743845, + 0.0033510003704577684, + 0.002646618289873004, + 0.004927077796310186, + 0.00486882496625185, + 0.005264637526124716, + 0.008364533074200153, + 0.0032374109141528606, + 0.009549465030431747, + 0.006872517056763172, + 0.005994773004204035, + 0.012683677487075329, + 0.0026276528369635344, + 0.0020702912006527185, + 0.002699602395296097, + 0.0020958795212209225, + 0.0031398599967360497, + 0.002955253701657057, + 0.0020325491204857826, + 0.002481283387169242, + 0.0041403938084840775, + 0.003930575679987669, + 0.007371160201728344, + 0.004166172817349434, + 0.003617832437157631, + 0.006180155090987682, + 0.008969388902187347, + 0.007684408221393824, + 0.00891063455492258, + 0.014082868583500385, + 0.05577601119875908, + 0.021733921021223068, + 0.008467287756502628, + 0.008372866548597813, + 0.00878818053752184, + 0.03312331438064575, + 0.007856824435293674, + 0.010475138202309608, + 0.008162948302924633, + 0.01098178792744875, + 0.010461094789206982, + 0.009030349552631378, + 0.00987200252711773, + 0.009416885673999786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.04867839813232422 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607188768684864 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607188768684864, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607188768684864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11650055646896362, + 0.04321030527353287, + 0.43885740637779236, + 0.04919612780213356, + 0.03775058686733246, + 0.0163063183426857, + 0.033923324197530746, + 0.011453836224973202, + 0.21689391136169434, + 0.03590761870145798 + ], + "credit_map": [ + 0.01227572001516819, + 0.018150174990296364, + 0.01657676324248314, + 0.009421306662261486, + 0.015102175064384937, + 0.016066642478108406, + 0.011731990613043308, + 0.017175788059830666, + 0.003688086988404393, + 0.005896484944969416, + 0.00567718455567956, + 0.004377737641334534, + 0.007238255813717842, + 0.004929801914840937, + 0.006677620112895966, + 0.004725134000182152, + 0.029219532385468483, + 0.029358794912695885, + 0.1094498485326767, + 0.1514855921268463, + 0.01487790048122406, + 0.03909868374466896, + 0.055129919201135635, + 0.01023710984736681, + 0.0033931389916688204, + 0.0050318120047450066, + 0.007688093464821577, + 0.004045721609145403, + 0.007006498984992504, + 0.007345004007220268, + 0.0046760872937738895, + 0.010009771212935448, + 0.004333651624619961, + 0.004263454116880894, + 0.005907298065721989, + 0.003104372648522258, + 0.00558260316029191, + 0.004717348609119654, + 0.003782630432397127, + 0.006059227045625448, + 0.002088231733068824, + 0.0015273246681317687, + 0.0024082050658762455, + 0.0012725660344585776, + 0.0023311518598347902, + 0.0024679750204086304, + 0.0015008841874077916, + 0.0027099798899143934, + 0.0050035445019602776, + 0.003893940942361951, + 0.005097812972962856, + 0.002664652420207858, + 0.003977519925683737, + 0.005196115933358669, + 0.003063425188884139, + 0.005026316270232201, + 0.0010439021280035377, + 0.0014411397278308868, + 0.0014364472590386868, + 0.000928127090446651, + 0.0019079404883086681, + 0.0014447550056502223, + 0.0016546710394322872, + 0.0015968538355082273, + 0.015208985656499863, + 0.010681577026844025, + 0.08406542986631393, + 0.03160667046904564, + 0.008846353739500046, + 0.009253187105059624, + 0.009878442622721195, + 0.04735326021909714, + 0.004021576140075922, + 0.0042341602966189384, + 0.002905169501900673, + 0.008130991831421852, + 0.003582486417144537, + 0.004098915494978428, + 0.0030964333564043045, + 0.0058378903195261955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.01607188768684864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a high vantage point looking downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, pure white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing a slow 'cloud hand' movement with smooth, flowing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background dominated by a twisted ancient pine tree with dark brown bark and thick branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning sunlight filtering through pine needles to cast dappled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205803155899048 + ], + "opposite_logprob_tokens": [ + -2.8120579719543457 + ], + "logprob_reward": -0.06205803155899048, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205803155899048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1474463939666748, + 0.12746278941631317, + 0.28054916858673096, + 0.04974030703306198, + 0.054097630083560944, + 0.026010850444436073, + 0.04454740881919861, + 0.044794730842113495, + 0.17816826701164246, + 0.04718247056007385 + ], + "credit_map": [ + 0.01731215976178646, + 0.017919350415468216, + 0.01597486063838005, + 0.015315234661102295, + 0.012944931164383888, + 0.027400566264986992, + 0.021016085520386696, + 0.019563211128115654, + 0.009818612597882748, + 0.015685906633734703, + 0.015123355202376842, + 0.015575047582387924, + 0.01711956411600113, + 0.013118370436131954, + 0.024755652993917465, + 0.01626628264784813, + 0.019022755324840546, + 0.018911555409431458, + 0.0692136138677597, + 0.09648411720991135, + 0.009714016690850258, + 0.025305207818746567, + 0.03529297187924385, + 0.006604932714253664, + 0.004188705701380968, + 0.005389221478253603, + 0.007477565668523312, + 0.004532745573669672, + 0.006624358706176281, + 0.006482651457190514, + 0.005611451342701912, + 0.009433604776859283, + 0.005888718646019697, + 0.005086646415293217, + 0.009798561222851276, + 0.0037512604612857103, + 0.0058299703523516655, + 0.006654879078269005, + 0.004798186011612415, + 0.012289407663047314, + 0.0032197677064687014, + 0.0030186905059963465, + 0.003274518297985196, + 0.00187200796790421, + 0.003954318817704916, + 0.004007450770586729, + 0.0020986422896385193, + 0.0045654550194740295, + 0.0076200696639716625, + 0.007209464907646179, + 0.0064071957021951675, + 0.003379786852747202, + 0.005544360261410475, + 0.005954664200544357, + 0.0028218617662787437, + 0.005610005930066109, + 0.004343297798186541, + 0.003819733392447233, + 0.008679419755935669, + 0.004314478486776352, + 0.004444059915840626, + 0.004317650571465492, + 0.008858500979840755, + 0.006017590407282114, + 0.013804171234369278, + 0.01022532768547535, + 0.06635838001966476, + 0.025736121460795403, + 0.008000808767974377, + 0.008806617930531502, + 0.008039475418627262, + 0.03719736635684967, + 0.004208337049931288, + 0.006270124111324549, + 0.0051100109703838825, + 0.007535731419920921, + 0.005009355954825878, + 0.008276051841676235, + 0.004022084642201662, + 0.006750772707164288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.06205803155899048 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_9/index_tensor([8014, 8014])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_9/index_tensor([8014, 8014])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ace42ae146696dce28781bb49e4a6d467e8e0b6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_9/index_tensor([8014, 8014])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium-length shot tilts smoothly to the right, tracking a modern, stylish city bus as it glides to a stop at a bustling city center under the warm, golden light of afternoon. Painted in vibrant electric blue and crisp white, the bus's exterior gleams with polished surfaces, and expansive, tinted, panoramic windows reflect the cityscape. Along its sides, bright, high-resolution LED screens display dynamic, colorful abstract graphics that move and flow. With a quiet aerodynamic hiss, its stylish double-glazed doors slide open, releasing a diverse group of passengers onto the crowded sidewalk. Inside, the clean, brightly lit interior reveals rows of modern, molded seats, occupied by commuters engrossed in their phones and books. The driver, a woman in a crisp, dark blue uniform, greets the new passengers aboard with a friendly, professional greeting. As the doors close seamlessly, the bus pulls away from the curb and effortlessly blends into the bustling street traffic. The camera then widens to capture the bus navigating a vibrant urban canyon, its forms contrasting past ornate historic buildings and towering glass and steel skyscrapers, perfectly encapsulating the vibrant, pulsating nature of modern city life.", + "seed": 20997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_9/index_8014_seed_20997_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3790958564728498, + "vlm_logprob": -0.3790958564728498, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.3790958564728498, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007835840806365013 + ], + "opposite_logprob_tokens": [ + -5.757835865020752 + ], + "logprob_reward": -0.007835840806365013, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007835840806365013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061094731092453, + 0.1251089870929718, + 0.34762847423553467, + 0.02124658040702343, + 0.03192364051938057, + 0.022812049835920334, + 0.07483649253845215, + 0.160050168633461, + 0.03837791830301285, + 0.11692091822624207 + ], + "credit_map": [ + 0.010018613189458847, + 0.008215785957872868, + 0.0092717744410038, + 0.006705424748361111, + 0.009263262152671814, + 0.006889767944812775, + 0.0057176025584340096, + 0.005012499634176493, + 0.030024563893675804, + 0.01005770917981863, + 0.011732426472008228, + 0.010253796353936195, + 0.013789152726531029, + 0.012419958598911762, + 0.015283669345080853, + 0.021547721698880196, + 0.013029134832322598, + 0.012890554033219814, + 0.007780398242175579, + 0.017521310597658157, + 0.020528754219412804, + 0.02205354906618595, + 0.13710038363933563, + 0.11672438681125641, + 0.0025086619425565004, + 0.002306029200553894, + 0.0019737135153263807, + 0.002110937377437949, + 0.0019987847190350294, + 0.002368736546486616, + 0.004737404175102711, + 0.0032423140946775675, + 0.004207868594676256, + 0.002140713855624199, + 0.006671519484370947, + 0.004404999781399965, + 0.0036743017844855785, + 0.0025714715011417866, + 0.003962328657507896, + 0.004290436394512653, + 0.003302693832665682, + 0.0020877046044915915, + 0.0033287378028035164, + 0.003529617330059409, + 0.0017591232899576426, + 0.0025448359083384275, + 0.002943510189652443, + 0.003315826179459691, + 0.010331857018172741, + 0.009565548971295357, + 0.00872835423797369, + 0.010787746869027615, + 0.005866517778486013, + 0.005666847340762615, + 0.014434006065130234, + 0.009455613791942596, + 0.013707233592867851, + 0.017328131943941116, + 0.02225102111697197, + 0.018362028524279594, + 0.022228574380278587, + 0.018857231363654137, + 0.028058897703886032, + 0.01925705000758171, + 0.0034851664677262306, + 0.0032552452757954597, + 0.004529598634690046, + 0.004196712747216225, + 0.0053902133367955685, + 0.006872970145195723, + 0.004445888567715883, + 0.006202122662216425, + 0.008001817390322685, + 0.009022068232297897, + 0.008447310887277126, + 0.005497206002473831, + 0.037939660251140594, + 0.01013906393200159, + 0.0062340181320905685, + 0.03163976967334747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.007835840806365013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0111062191426754 + ], + "opposite_logprob_tokens": [ + -4.511106014251709 + ], + "logprob_reward": -0.0111062191426754, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0111062191426754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04382236674427986, + 0.043582234531641006, + 0.45141664147377014, + 0.0332484096288681, + 0.02036711387336254, + 0.023943880572915077, + 0.033681467175483704, + 0.0646538957953453, + 0.01771107316017151, + 0.2675729990005493 + ], + "credit_map": [ + 0.008166016079485416, + 0.0076010217890143394, + 0.004720985423773527, + 0.003432883881032467, + 0.0075636450201272964, + 0.004857728257775307, + 0.003673447296023369, + 0.0038066324777901173, + 0.013391897082328796, + 0.004535072948783636, + 0.0032944295089691877, + 0.0018961760215461254, + 0.0034364212770015, + 0.0036428167950361967, + 0.003583402605727315, + 0.009802015498280525, + 0.015919366851449013, + 0.013173721730709076, + 0.009205265901982784, + 0.022477243095636368, + 0.025942571461200714, + 0.0279932152479887, + 0.1824764460325241, + 0.15422877669334412, + 0.0032827649265527725, + 0.003977560438215733, + 0.003524367930367589, + 0.0029621142894029617, + 0.0035797380842268467, + 0.004171634092926979, + 0.006790857762098312, + 0.0049593690782785416, + 0.003517553210258484, + 0.0020720665343105793, + 0.003183771390467882, + 0.0020207385532557964, + 0.0029332763515412807, + 0.002007403178140521, + 0.0022542139049619436, + 0.0023780895862728357, + 0.0026939173694700003, + 0.0033245719969272614, + 0.0033840558025985956, + 0.003471827134490013, + 0.0026324924547225237, + 0.00247701327316463, + 0.003089835401624441, + 0.002870165044441819, + 0.0022322330623865128, + 0.003527844324707985, + 0.0029320337343961, + 0.003225487656891346, + 0.003998255357146263, + 0.0020109533797949553, + 0.009328040294349194, + 0.0064266156405210495, + 0.008868849836289883, + 0.007043363526463509, + 0.007838192395865917, + 0.006658546160906553, + 0.0071611106395721436, + 0.007702071685343981, + 0.011781920678913593, + 0.007599830161780119, + 0.001849265885539353, + 0.002092784270644188, + 0.002417218405753374, + 0.0018036473775282502, + 0.0023619586136192083, + 0.002936596516519785, + 0.0019253408536314964, + 0.0023242589086294174, + 0.01003174390643835, + 0.014691537246108055, + 0.011549342423677444, + 0.009918241761624813, + 0.10471093654632568, + 0.02222045138478279, + 0.014212805777788162, + 0.08023791015148163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.0111062191426754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015070915222168 + ], + "opposite_logprob_tokens": [ + -0.2015070915222168 + ], + "logprob_reward": -1.7015070915222168, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015070915222168, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06590141355991364, + 0.07580526173114777, + 0.23972760140895844, + 0.04678722843527794, + 0.045010145753622055, + 0.028769899159669876, + 0.33104798197746277, + 0.05817032977938652, + 0.021408507600426674, + 0.08737161010503769 + ], + "credit_map": [ + 0.011008392088115215, + 0.010609187185764313, + 0.007057800889015198, + 0.0046389782801270485, + 0.013007561676204205, + 0.008848805911839008, + 0.006112111732363701, + 0.0046185702085494995, + 0.025622187182307243, + 0.006411836948245764, + 0.004867523908615112, + 0.003742125118151307, + 0.005806306377053261, + 0.00549470866099, + 0.006318097468465567, + 0.01754247583448887, + 0.013388183899223804, + 0.017576826736330986, + 0.0069681513123214245, + 0.01382149662822485, + 0.0173125471919775, + 0.015248673036694527, + 0.08324220031499863, + 0.07216951251029968, + 0.006338317878544331, + 0.0048318528570234776, + 0.005407205317169428, + 0.003870191052556038, + 0.0044373441487550735, + 0.003830658970400691, + 0.009885251522064209, + 0.008186406455934048, + 0.006129388231784105, + 0.0034902128390967846, + 0.007810854818671942, + 0.005199229344725609, + 0.00437605706974864, + 0.0036417676601558924, + 0.007592627312988043, + 0.006770011968910694, + 0.003033475251868367, + 0.0032554809004068375, + 0.0032615778036415577, + 0.0034438122529536486, + 0.00377165456302464, + 0.0029014102183282375, + 0.004793452098965645, + 0.004309036303311586, + 0.022506697103381157, + 0.030380895361304283, + 0.01731765642762184, + 0.032302021980285645, + 0.04208856448531151, + 0.01645236648619175, + 0.09637322276830673, + 0.0736265629529953, + 0.010494306683540344, + 0.006434763316065073, + 0.007315005175769329, + 0.006332762539386749, + 0.0036290506832301617, + 0.005196114536374807, + 0.010433842428028584, + 0.008334487676620483, + 0.0024300117511302233, + 0.0026429062709212303, + 0.002244764706119895, + 0.0025239763781428337, + 0.0035619852133095264, + 0.0033773069735616446, + 0.002540304558351636, + 0.0020872524473816156, + 0.0037550872657448053, + 0.004730252083390951, + 0.0040928092785179615, + 0.004289875738322735, + 0.032678622752428055, + 0.0070268092676997185, + 0.005710132420063019, + 0.025088025256991386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -1.7015070915222168 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129799410700798 + ], + "opposite_logprob_tokens": [ + -2.2379798889160156 + ], + "logprob_reward": -0.1129799410700798, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129799410700798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038982510566711426, + 0.06932571530342102, + 0.28628575801849365, + 0.1039448007941246, + 0.05833985283970833, + 0.018158400431275368, + 0.2707064151763916, + 0.05174211785197258, + 0.03953388333320618, + 0.06298049539327621 + ], + "credit_map": [ + 0.006380475126206875, + 0.005166272632777691, + 0.004996674135327339, + 0.0032574301585555077, + 0.005089419428259134, + 0.0061663175001740456, + 0.004163680598139763, + 0.0037622430827468634, + 0.02336759679019451, + 0.005413467064499855, + 0.00486557325348258, + 0.004240698646754026, + 0.004721915349364281, + 0.005566975101828575, + 0.00626347353681922, + 0.014886017888784409, + 0.012050312012434006, + 0.01308111660182476, + 0.0067089637741446495, + 0.015234744176268578, + 0.018291331827640533, + 0.017687244340777397, + 0.10934729874134064, + 0.09388477355241776, + 0.0069174026139080524, + 0.008657937869429588, + 0.009641624055802822, + 0.01204897090792656, + 0.0070816162042319775, + 0.00923121441155672, + 0.03146377578377724, + 0.01890227012336254, + 0.007108744699507952, + 0.004007380921393633, + 0.00939245242625475, + 0.007476027589291334, + 0.005712810438126326, + 0.004387755878269672, + 0.01088931504637003, + 0.009365366771817207, + 0.0023711323738098145, + 0.0020332452841103077, + 0.002701961435377598, + 0.0021465087775141, + 0.0016808733344078064, + 0.0017204043688252568, + 0.003175841411575675, + 0.002328435890376568, + 0.019342442974448204, + 0.02517944946885109, + 0.014416003599762917, + 0.024659112095832825, + 0.03367091715335846, + 0.013759126886725426, + 0.07919549196958542, + 0.0604839026927948, + 0.009268242865800858, + 0.006132727488875389, + 0.006172761786729097, + 0.004981308709830046, + 0.0036393278278410435, + 0.004915823694318533, + 0.008980835787951946, + 0.007651091553270817, + 0.004366937559098005, + 0.0054221125319600105, + 0.005668254103511572, + 0.005758027080446482, + 0.0046119242906570435, + 0.005512302741408348, + 0.004651992581784725, + 0.003542337566614151, + 0.0032221523579210043, + 0.004487851168960333, + 0.0043655866757035255, + 0.00424495805054903, + 0.021339600905776024, + 0.0054396772757172585, + 0.004557139705866575, + 0.015323529951274395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.1129799410700798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062050189822912216 + ], + "opposite_logprob_tokens": [ + -2.8120501041412354 + ], + "logprob_reward": -0.062050189822912216, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062050189822912216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10262509435415268, + 0.11723385006189346, + 0.1476721167564392, + 0.09604660421609879, + 0.07861945778131485, + 0.037447646260261536, + 0.15637843310832977, + 0.16512683033943176, + 0.04261806234717369, + 0.05623192340135574 + ], + "credit_map": [ + 0.02004273235797882, + 0.018559804186224937, + 0.014090674929320812, + 0.008403829298913479, + 0.010830532759428024, + 0.013081175275146961, + 0.009701133705675602, + 0.007915211841464043, + 0.013965887017548084, + 0.014049378223717213, + 0.014634666033089161, + 0.011746526695787907, + 0.018894964829087257, + 0.013123426586389542, + 0.017310764640569687, + 0.013508232310414314, + 0.01000527385622263, + 0.01113611739128828, + 0.005872128531336784, + 0.00933819729834795, + 0.010565372183918953, + 0.01106139738112688, + 0.05016079545021057, + 0.03953284025192261, + 0.010348492302000523, + 0.008212794549763203, + 0.008665500208735466, + 0.01236205268651247, + 0.007937605492770672, + 0.008439072407782078, + 0.026257464662194252, + 0.013823620975017548, + 0.010775603353977203, + 0.006237778812646866, + 0.02217290550470352, + 0.006549155805259943, + 0.0063796197064220905, + 0.0053350054658949375, + 0.010297066532075405, + 0.010872319340705872, + 0.005427529104053974, + 0.0041150315664708614, + 0.0054516978561878204, + 0.005214101169258356, + 0.0035839760676026344, + 0.003878106130287051, + 0.005564545746892691, + 0.004212657455354929, + 0.007854696363210678, + 0.013701792806386948, + 0.008398400619626045, + 0.015515338629484177, + 0.01804288849234581, + 0.0075899288058280945, + 0.04718298465013504, + 0.03809241205453873, + 0.02527616359293461, + 0.0179689172655344, + 0.020134223625063896, + 0.016490379348397255, + 0.016414141282439232, + 0.019294722005724907, + 0.029074272140860558, + 0.020474009215831757, + 0.005683050490915775, + 0.004261888563632965, + 0.005395549815148115, + 0.004953123163431883, + 0.005223163869231939, + 0.0055757188238203526, + 0.005627802107483149, + 0.00589776411652565, + 0.007838424295186996, + 0.007193139288574457, + 0.007105643395334482, + 0.006348856724798679, + 0.008148400112986565, + 0.007171149365603924, + 0.00574179133400321, + 0.006684518419206142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.062050189822912216 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot tilts smoothly to the right, tracking a modern, stylish city bus as it glides to a stop at a bustling city center under the warm, golden light of afternoon. Painted in vibrant electric blue and crisp white, the bus's exterior gleams with polished surfaces, and expansive, tinted, panoramic windows reflect the cityscape. Along its sides, bright, high-resolution LED screens display dynamic, colorful abstract graphics that move and flow. With a quiet aerodynamic hiss, its stylish double-glazed doors slide open, releasing a diverse group of passengers onto the crowded sidewalk. Inside, the clean, brightly lit interior reveals rows of modern, molded seats, occupied by commuters engrossed in their phones and books. The driver, a woman in a crisp, dark blue uniform, greets the new passengers aboard with a friendly, professional greeting. As the doors close seamlessly, the bus pulls away from the curb and effortlessly blends into the bustling street traffic. The camera then widens to capture the bus navigating a vibrant urban canyon, its forms contrasting past ornate historic buildings and towering glass and steel skyscrapers, perfectly encapsulating the vibrant, pulsating nature of modern city life.", + "seed": 20998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_9/index_8014_seed_20998_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2962004886008799, + "vlm_logprob": -0.2962004886008799, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.2962004886008799, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006578692700713873 + ], + "opposite_logprob_tokens": [ + -6.2565789222717285 + ], + "logprob_reward": -0.006578692700713873, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.006578692700713873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1244308203458786, + 0.0720444843173027, + 0.09029872715473175, + 0.09553839266300201, + 0.14416621625423431, + 0.0966372862458229, + 0.12488782405853271, + 0.10362257808446884, + 0.0787155032157898, + 0.06965816020965576 + ], + "credit_map": [ + 0.016318434849381447, + 0.012512268498539925, + 0.022580213844776154, + 0.009475736878812313, + 0.012666978873312473, + 0.01301466766744852, + 0.014263167977333069, + 0.023599348962306976, + 0.008148830384016037, + 0.014385105110704899, + 0.00845889002084732, + 0.012232040986418724, + 0.008002422749996185, + 0.006476431619375944, + 0.007240035571157932, + 0.007100733928382397, + 0.019602077081799507, + 0.01093288417905569, + 0.007097918540239334, + 0.00765304546803236, + 0.010078751482069492, + 0.010743062943220139, + 0.005050953011959791, + 0.01914004236459732, + 0.010817499831318855, + 0.014012468047440052, + 0.009634914807975292, + 0.0155951464548707, + 0.013709306716918945, + 0.006705369334667921, + 0.013338811695575714, + 0.011724886484444141, + 0.019722087308764458, + 0.012644888833165169, + 0.022247832268476486, + 0.013477576896548271, + 0.012805894017219543, + 0.025859368965029716, + 0.022699421271681786, + 0.014709154143929482, + 0.01077752374112606, + 0.008846664801239967, + 0.010065347887575626, + 0.011507570743560791, + 0.01443543378263712, + 0.008718413300812244, + 0.019383156672120094, + 0.0129031827673316, + 0.018951354548335075, + 0.013987761922180653, + 0.01708964817225933, + 0.0160761047154665, + 0.013093937188386917, + 0.01455760933458805, + 0.01707504503428936, + 0.014056370593607426, + 0.010085271671414375, + 0.02403745800256729, + 0.007917876355350018, + 0.009674401953816414, + 0.006872803438454866, + 0.008698662742972374, + 0.014045292511582375, + 0.02229081466794014, + 0.007698439992964268, + 0.011727976612746716, + 0.009127851575613022, + 0.010791588574647903, + 0.015075779519975185, + 0.008504246361553669, + 0.008588377386331558, + 0.00720125762745738, + 0.0055121732875704765, + 0.009175771847367287, + 0.0037046237848699093, + 0.005777394399046898, + 0.009451349265873432, + 0.01342838816344738, + 0.012063022702932358, + 0.010545444674789906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.006578692700713873 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067633045837283134 + ], + "opposite_logprob_tokens": [ + -5.006763458251953 + ], + "logprob_reward": -0.0067633045837283134, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067633045837283134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13029569387435913, + 0.04857126250863075, + 0.1308036893606186, + 0.0810680016875267, + 0.10513867437839508, + 0.040502700954675674, + 0.11695640534162521, + 0.2413526326417923, + 0.047304220497608185, + 0.05800676718354225 + ], + "credit_map": [ + 0.020236941054463387, + 0.016650604084134102, + 0.017542079091072083, + 0.012389401905238628, + 0.011567470617592335, + 0.011042043566703796, + 0.012924672104418278, + 0.027942489832639694, + 0.009466677904129028, + 0.005869806744158268, + 0.0031663328409194946, + 0.005421046633273363, + 0.01146791223436594, + 0.004683104809373617, + 0.004844166804105043, + 0.003652219194918871, + 0.022266460582613945, + 0.02129979059100151, + 0.008974180556833744, + 0.007884609512984753, + 0.010549969971179962, + 0.017857536673545837, + 0.008036899380385876, + 0.03393423929810524, + 0.008992535062134266, + 0.011824578046798706, + 0.009991299360990524, + 0.012622411362826824, + 0.013545474968850613, + 0.00646609952673316, + 0.01024160347878933, + 0.007384008262306452, + 0.012387024238705635, + 0.012052911333739758, + 0.016841880977153778, + 0.010637531988322735, + 0.008154262788593769, + 0.017693523317575455, + 0.01748659834265709, + 0.009884933941066265, + 0.005709432531148195, + 0.0038000077474862337, + 0.005271314177662134, + 0.005312648601830006, + 0.00546443834900856, + 0.003358133602887392, + 0.006391608156263828, + 0.005195118952542543, + 0.015087400563061237, + 0.014318710193037987, + 0.016643710434436798, + 0.014899928122758865, + 0.01180552039295435, + 0.010682545602321625, + 0.019896572455763817, + 0.01362202875316143, + 0.024544229730963707, + 0.053591445088386536, + 0.01956040784716606, + 0.03073856793344021, + 0.022283188998699188, + 0.01724361628293991, + 0.03150610998272896, + 0.041885048151016235, + 0.005234545562416315, + 0.0049370392225682735, + 0.004469489213079214, + 0.006475236266851425, + 0.007458058651536703, + 0.0065301996655762196, + 0.006635414436459541, + 0.005564235150814056, + 0.004325548652559519, + 0.004634710494428873, + 0.003976756241172552, + 0.006176351569592953, + 0.007135212421417236, + 0.010837552137672901, + 0.011481331661343575, + 0.00943930633366108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0067633045837283134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4062690734863281 + ], + "opposite_logprob_tokens": [ + -0.2812690734863281 + ], + "logprob_reward": -1.4062690734863281, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4062690734863281, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04405510425567627, + 0.06728119403123856, + 0.10142096877098083, + 0.09674204885959625, + 0.20493851602077484, + 0.22006776928901672, + 0.04944760724902153, + 0.131929412484169, + 0.04271404817700386, + 0.04140333831310272 + ], + "credit_map": [ + 0.007825877517461777, + 0.0065426998771727085, + 0.00664861686527729, + 0.0034711910411715508, + 0.00489496486261487, + 0.00650758808478713, + 0.0042390418238937855, + 0.003925122320652008, + 0.017072433605790138, + 0.00735160056501627, + 0.004566912539303303, + 0.005213343538343906, + 0.020737141370773315, + 0.006415811832994223, + 0.0033508550841361284, + 0.0025730920024216175, + 0.015140103176236153, + 0.018686525523662567, + 0.006500073708593845, + 0.007055757101625204, + 0.010499240830540657, + 0.014575078152120113, + 0.010945185087621212, + 0.01801900751888752, + 0.010687445290386677, + 0.015606838278472424, + 0.01045707706362009, + 0.018286291509866714, + 0.012485860846936703, + 0.00814279355108738, + 0.010717383585870266, + 0.010358357802033424, + 0.024899983778595924, + 0.02352633886039257, + 0.03519173711538315, + 0.020858250558376312, + 0.015908878296613693, + 0.035905420780181885, + 0.031237682327628136, + 0.01741022802889347, + 0.023060712963342667, + 0.018458940088748932, + 0.021536873653531075, + 0.023077508434653282, + 0.035576753318309784, + 0.022656874731183052, + 0.046040330082178116, + 0.029659774154424667, + 0.005222687032073736, + 0.0057848189026117325, + 0.010267658159136772, + 0.005578317679464817, + 0.0054670218378305435, + 0.0053505729883909225, + 0.005799361504614353, + 0.005977167747914791, + 0.01771322265267372, + 0.028944186866283417, + 0.012346670031547546, + 0.0164237841963768, + 0.01459894236177206, + 0.012242238968610764, + 0.01704910583794117, + 0.012611271813511848, + 0.00427927216514945, + 0.005754040088504553, + 0.006014583166688681, + 0.005284723360091448, + 0.007889888249337673, + 0.0047849928960204124, + 0.004452908411622047, + 0.00425364077091217, + 0.004357845522463322, + 0.004204440396279097, + 0.0029515959322452545, + 0.0036830308381468058, + 0.005551232490688562, + 0.008190618827939034, + 0.005896717309951782, + 0.006567856762558222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -1.4062690734863281 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038109004497528076 + ], + "opposite_logprob_tokens": [ + -3.288109064102173 + ], + "logprob_reward": -0.038109004497528076, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038109004497528076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08668938279151917, + 0.0570770688354969, + 0.07310645282268524, + 0.08749180287122726, + 0.11019733548164368, + 0.11050917953252792, + 0.08718489855527878, + 0.3269365131855011, + 0.030882611870765686, + 0.02992480807006359 + ], + "credit_map": [ + 0.011139382608234882, + 0.008729883469641209, + 0.009326279163360596, + 0.005589151289314032, + 0.006419881712645292, + 0.007490235380828381, + 0.012553182430565357, + 0.02544138953089714, + 0.013860637322068214, + 0.007112082559615374, + 0.003908534534275532, + 0.004387116990983486, + 0.016482539474964142, + 0.005392489023506641, + 0.003488034475594759, + 0.0024456335231661797, + 0.0149956364184618, + 0.011487691663205624, + 0.004439129028469324, + 0.00392589345574379, + 0.005862761288881302, + 0.009343595243990421, + 0.0054541220888495445, + 0.01759762316942215, + 0.010252312757074833, + 0.013226574286818504, + 0.009674901142716408, + 0.01422154437750578, + 0.012567779049277306, + 0.006699076388031244, + 0.014469385147094727, + 0.006380223669111729, + 0.012826653197407722, + 0.013128109276294708, + 0.018046794459223747, + 0.011196702718734741, + 0.009231636300683022, + 0.01897306554019451, + 0.01649882271885872, + 0.010295556858181953, + 0.012217127718031406, + 0.009220821782946587, + 0.011041293852031231, + 0.010645600967109203, + 0.019326074048876762, + 0.010735916905105114, + 0.02289283275604248, + 0.014429505914449692, + 0.008992270566523075, + 0.010859628207981586, + 0.014524917118251324, + 0.01197335310280323, + 0.008337150327861309, + 0.00897056981921196, + 0.013521255925297737, + 0.010005751624703407, + 0.022480955347418785, + 0.05517992749810219, + 0.01741921156644821, + 0.05345173552632332, + 0.02339465171098709, + 0.025023138150572777, + 0.04692468047142029, + 0.08306223154067993, + 0.00401382427662611, + 0.003947898279875517, + 0.004542567767202854, + 0.0036308718845248222, + 0.0056009734980762005, + 0.003300773212686181, + 0.0028272622730582952, + 0.0030184404458850622, + 0.0026223482564091682, + 0.004028309602290392, + 0.0018870164640247822, + 0.0038005211390554905, + 0.004494816530495882, + 0.0035995584912598133, + 0.005792455747723579, + 0.0036997811403125525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.038109004497528076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328236773610115 + ], + "opposite_logprob_tokens": [ + -3.773282289505005 + ], + "logprob_reward": -0.02328236773610115, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328236773610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12775124609470367, + 0.062390003353357315, + 0.03506510332226753, + 0.12442449480295181, + 0.35590195655822754, + 0.07099258154630661, + 0.03616294637322426, + 0.06881985068321228, + 0.02868286892771721, + 0.08980897814035416 + ], + "credit_map": [ + 0.024082334712147713, + 0.016309523954987526, + 0.011339168064296246, + 0.008603076450526714, + 0.009725936688482761, + 0.008598999120295048, + 0.013614256866276264, + 0.03547793626785278, + 0.013970506377518177, + 0.007894539274275303, + 0.005103862378746271, + 0.0057642171159386635, + 0.01675347052514553, + 0.0071515413001179695, + 0.003283235477283597, + 0.002468627644702792, + 0.006565883755683899, + 0.007602686528116465, + 0.0031711363699287176, + 0.0031016943976283073, + 0.0049115135334432125, + 0.003355581546202302, + 0.003387370379641652, + 0.0029692319221794605, + 0.014138120226562023, + 0.016521556302905083, + 0.012757080607116222, + 0.017691226676106453, + 0.019335592165589333, + 0.011860791593790054, + 0.014698789454996586, + 0.017421327531337738, + 0.045275960117578506, + 0.03937007114291191, + 0.06234622374176979, + 0.03669201210141182, + 0.02711593545973301, + 0.056216634809970856, + 0.058205507695674896, + 0.03067958541214466, + 0.007810113951563835, + 0.0058431969955563545, + 0.00840864609926939, + 0.009271662682294846, + 0.010678378865122795, + 0.005523597355931997, + 0.012119539082050323, + 0.01133743952959776, + 0.003919569309800863, + 0.0037066228687763214, + 0.007608381099998951, + 0.004962180275470018, + 0.003441727254539728, + 0.00512759480625391, + 0.003444961505010724, + 0.003951903898268938, + 0.0072723194025456905, + 0.012098045088350773, + 0.006622167304158211, + 0.011456521227955818, + 0.007481012027710676, + 0.005064958240836859, + 0.008428339846432209, + 0.010396476835012436, + 0.002942034276202321, + 0.0035632075741887093, + 0.0035615768283605576, + 0.002755391178652644, + 0.005272912327200174, + 0.004007752053439617, + 0.0037064882926642895, + 0.0028735061641782522, + 0.012118535116314888, + 0.011177723295986652, + 0.006597411818802357, + 0.010359209030866623, + 0.01158144325017929, + 0.010007794946432114, + 0.013311479240655899, + 0.01465537492185831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.02328236773610115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot tilts smoothly to the right, tracking a modern, stylish city bus as it glides to a stop at a bustling city center under the warm, golden light of afternoon. Painted in vibrant electric blue and crisp white, the bus's exterior gleams with polished surfaces, and expansive, tinted, panoramic windows reflect the cityscape. Along its sides, bright, high-resolution LED screens display dynamic, colorful abstract graphics that move and flow. With a quiet aerodynamic hiss, its stylish double-glazed doors slide open, releasing a diverse group of passengers onto the crowded sidewalk. Inside, the clean, brightly lit interior reveals rows of modern, molded seats, occupied by commuters engrossed in their phones and books. The driver, a woman in a crisp, dark blue uniform, greets the new passengers aboard with a friendly, professional greeting. As the doors close seamlessly, the bus pulls away from the curb and effortlessly blends into the bustling street traffic. The camera then widens to capture the bus navigating a vibrant urban canyon, its forms contrasting past ornate historic buildings and towering glass and steel skyscrapers, perfectly encapsulating the vibrant, pulsating nature of modern city life.", + "seed": 20999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_9/index_8014_seed_20999_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.42875698693096637, + "vlm_logprob": -0.42875698693096637, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.42875698693096637, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008252331987023354 + ], + "opposite_logprob_tokens": [ + -5.633252143859863 + ], + "logprob_reward": -0.008252331987023354, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.008252331987023354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06907342374324799, + 0.0438961461186409, + 0.06295956671237946, + 0.03805936127901077, + 0.45726439356803894, + 0.052238933742046356, + 0.05175894871354103, + 0.12772515416145325, + 0.07012186199426651, + 0.026902245357632637 + ], + "credit_map": [ + 0.014733821153640747, + 0.009723268449306488, + 0.008345561102032661, + 0.00646447017788887, + 0.007283161394298077, + 0.009162621572613716, + 0.007187952753156424, + 0.006172569934278727, + 0.005051954183727503, + 0.004921088460832834, + 0.005787687376141548, + 0.0052331723272800446, + 0.005958664696663618, + 0.004748376552015543, + 0.005776707082986832, + 0.006418497301638126, + 0.007679522968828678, + 0.011266720481216908, + 0.007915872149169445, + 0.011473039165139198, + 0.008546817116439342, + 0.006697341334074736, + 0.0039535146206617355, + 0.0054267393425107, + 0.004243799485266209, + 0.0038872382137924433, + 0.0071350811049342155, + 0.004938448313623667, + 0.0063682254403829575, + 0.005607491824775934, + 0.0030067397747188807, + 0.0028723382856696844, + 0.05695243552327156, + 0.036296404898166656, + 0.03567308932542801, + 0.0800880491733551, + 0.023857446387410164, + 0.05729426443576813, + 0.02104462869465351, + 0.14605805277824402, + 0.006907330360263586, + 0.006108955480158329, + 0.008299109525978565, + 0.0036789088044315577, + 0.006602682173252106, + 0.009267644956707954, + 0.005094919819384813, + 0.006279380526393652, + 0.006884297356009483, + 0.00559442862868309, + 0.006297747138887644, + 0.006955649703741074, + 0.0065423306077718735, + 0.005265486426651478, + 0.005613206420093775, + 0.008605805225670338, + 0.015241664834320545, + 0.012781145982444286, + 0.01909688673913479, + 0.023799754679203033, + 0.013261564075946808, + 0.012086090631783009, + 0.016205746680498123, + 0.015252300538122654, + 0.006704368628561497, + 0.009515921585261822, + 0.0065005007199943066, + 0.006922461092472076, + 0.00446084700524807, + 0.019307367503643036, + 0.009659698233008385, + 0.0070506976917386055, + 0.0029959797393530607, + 0.0027957765851169825, + 0.0027884782757610083, + 0.0032712589018046856, + 0.0032646802719682455, + 0.003639898030087352, + 0.003988896496593952, + 0.004157276824116707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.008252331987023354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609816402196884 + ], + "opposite_logprob_tokens": [ + -4.1410980224609375 + ], + "logprob_reward": -0.01609816402196884, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609816402196884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11278844624757767, + 0.048009682446718216, + 0.07586086541414261, + 0.034017413854599, + 0.38419994711875916, + 0.04300092160701752, + 0.13305144011974335, + 0.05602510645985603, + 0.07824470847845078, + 0.034801557660102844 + ], + "credit_map": [ + 0.01681501790881157, + 0.019632387906312943, + 0.011550672352313995, + 0.007902261801064014, + 0.012623126618564129, + 0.016972726210951805, + 0.011930698528885841, + 0.015361558645963669, + 0.00604661088436842, + 0.00656950194388628, + 0.007937675341963768, + 0.005471258889883757, + 0.007194149773567915, + 0.005176114849746227, + 0.0038254489190876484, + 0.005788922775536776, + 0.010588401928544044, + 0.016016343608498573, + 0.0082609374076128, + 0.01118012610822916, + 0.010723278857767582, + 0.007940689101815224, + 0.004938100464642048, + 0.00621298560872674, + 0.0036880162078887224, + 0.004450320731848478, + 0.003418226959183812, + 0.003472673473879695, + 0.005116089712828398, + 0.007646418176591396, + 0.002461580326780677, + 0.0037640873342752457, + 0.051726143807172775, + 0.03416973352432251, + 0.028834449127316475, + 0.06783954799175262, + 0.023033419623970985, + 0.0457732193171978, + 0.018135763704776764, + 0.11468765139579773, + 0.00554649205878377, + 0.003954341635107994, + 0.008965264074504375, + 0.004549065604805946, + 0.00406013336032629, + 0.005884912796318531, + 0.004343193955719471, + 0.005697517655789852, + 0.015976423397660255, + 0.014895197004079819, + 0.017457008361816406, + 0.02179216779768467, + 0.013926595449447632, + 0.01681128703057766, + 0.012122259475290775, + 0.020070496946573257, + 0.007654412649571896, + 0.006005704868584871, + 0.010248973034322262, + 0.00834340788424015, + 0.006287808995693922, + 0.005064118187874556, + 0.007767468690872192, + 0.004653208889067173, + 0.008996529504656792, + 0.012033765204250813, + 0.011388206854462624, + 0.011009655892848969, + 0.006310386583209038, + 0.008690642192959785, + 0.008747954852879047, + 0.011067567393183708, + 0.003016975475475192, + 0.0034814404789358377, + 0.004003757145255804, + 0.0036188752856105566, + 0.003773862263187766, + 0.005875053349882364, + 0.005894089583307505, + 0.005137505941092968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.01609816402196884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048319816589355 + ], + "opposite_logprob_tokens": [ + -0.17983199656009674 + ], + "logprob_reward": -1.8048319816589355, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048319816589355, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08672409504652023, + 0.03991410508751869, + 0.13184159994125366, + 0.05965365096926689, + 0.09656929969787598, + 0.07676534354686737, + 0.056985042989254, + 0.20318716764450073, + 0.16325348615646362, + 0.08510621637105942 + ], + "credit_map": [ + 0.01668919064104557, + 0.01643436960875988, + 0.008718948811292648, + 0.007334980648010969, + 0.008895458653569221, + 0.013272343203425407, + 0.007033369969576597, + 0.008345438167452812, + 0.005198866594582796, + 0.004307747818529606, + 0.006551248021423817, + 0.004051826428622007, + 0.0062254611402750015, + 0.004257266875356436, + 0.003689597826451063, + 0.005632095970213413, + 0.02090359479188919, + 0.03301651030778885, + 0.015187854878604412, + 0.016630450263619423, + 0.015249696560204029, + 0.012714355252683163, + 0.005678507033735514, + 0.012460635043680668, + 0.006851299200206995, + 0.007202610373497009, + 0.008893207646906376, + 0.005909394938498735, + 0.0074209729209542274, + 0.011288302019238472, + 0.004601458087563515, + 0.007486408576369286, + 0.00999364722520113, + 0.0074316211976110935, + 0.007191758137196302, + 0.0142520098015666, + 0.007251040544360876, + 0.012761569581925869, + 0.006163164507597685, + 0.03152449429035187, + 0.007132446393370628, + 0.011035814881324768, + 0.012116702273488045, + 0.0073258657939732075, + 0.008339757099747658, + 0.012747726403176785, + 0.01131255179643631, + 0.006754481233656406, + 0.008692104369401932, + 0.007543584331870079, + 0.007564971223473549, + 0.00804993137717247, + 0.00804122630506754, + 0.005588828586041927, + 0.00533208018168807, + 0.006172318011522293, + 0.026594843715429306, + 0.03050149977207184, + 0.02111896313726902, + 0.03372647613286972, + 0.02110549807548523, + 0.021323032677173615, + 0.027842069044709206, + 0.020974798128008842, + 0.01832638494670391, + 0.02488907054066658, + 0.018672047182917595, + 0.018162038177251816, + 0.015003560110926628, + 0.03944533318281174, + 0.012139576487243176, + 0.016615482047200203, + 0.00909130647778511, + 0.008288195356726646, + 0.009711580350995064, + 0.009418017230927944, + 0.013153456151485443, + 0.009603535756468773, + 0.015611977316439152, + 0.010228155180811882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -1.8048319816589355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20165403187274933 + ], + "opposite_logprob_tokens": [ + -1.701654076576233 + ], + "logprob_reward": -0.20165403187274933, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20165403187274933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027332047000527382, + 0.02503420040011406, + 0.05875852704048157, + 0.018983136862516403, + 0.6104555726051331, + 0.03650185093283653, + 0.10000801831483841, + 0.06369937211275101, + 0.022289317101240158, + 0.03693794831633568 + ], + "credit_map": [ + 0.0046165501698851585, + 0.0037989539559930563, + 0.0029273200780153275, + 0.0030830774921923876, + 0.003064349526539445, + 0.004024512134492397, + 0.0036165572237223387, + 0.0022007266525179148, + 0.002926354529336095, + 0.003990544006228447, + 0.003359039779752493, + 0.0034032061230391264, + 0.004166565835475922, + 0.0025750952772796154, + 0.0021022583823651075, + 0.002511139027774334, + 0.010609266348183155, + 0.012676541693508625, + 0.0077535673044621944, + 0.006728885695338249, + 0.006100069265812635, + 0.005284534767270088, + 0.0031185115221887827, + 0.0064871544018387794, + 0.00265177502296865, + 0.0021750975865870714, + 0.0032200394198298454, + 0.0024225651286542416, + 0.0028311097994446754, + 0.0023961402475833893, + 0.0017900991952046752, + 0.0014963122084736824, + 0.08543319255113602, + 0.052779294550418854, + 0.04386754706501961, + 0.10947755724191666, + 0.032305825501680374, + 0.07326159626245499, + 0.028625503182411194, + 0.18470507860183716, + 0.0038975365459918976, + 0.004456618335098028, + 0.006031726486980915, + 0.0036132310051470995, + 0.00434533879160881, + 0.006163183134049177, + 0.0035033393651247025, + 0.004490881692618132, + 0.01478221733123064, + 0.017746364697813988, + 0.011981522664427757, + 0.011888978071510792, + 0.0120324045419693, + 0.012856053188443184, + 0.006956321187317371, + 0.011764157563447952, + 0.007886281237006187, + 0.008373646065592766, + 0.008127463050186634, + 0.010729031637310982, + 0.006566422991454601, + 0.00785147026181221, + 0.007485543843358755, + 0.006679513957351446, + 0.0018817635718733072, + 0.003928049933165312, + 0.002252918668091297, + 0.0025606690905988216, + 0.0020981975831091404, + 0.003961960319429636, + 0.002153116511180997, + 0.0034526425879448652, + 0.003609089646488428, + 0.0035386839881539345, + 0.004243064671754837, + 0.004467644728720188, + 0.004404490347951651, + 0.004743605386465788, + 0.006812329404056072, + 0.005119044333696365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.20165403187274933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294842511415482 + ], + "opposite_logprob_tokens": [ + -2.237948417663574 + ], + "logprob_reward": -0.11294842511415482, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294842511415482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1872800886631012, + 0.024799255654215813, + 0.08228182047605515, + 0.03630179539322853, + 0.08840560913085938, + 0.08208570629358292, + 0.03673217073082924, + 0.14203643798828125, + 0.2766297459602356, + 0.04344739764928818 + ], + "credit_map": [ + 0.04135692119598389, + 0.02870956063270569, + 0.01537906564772129, + 0.012747681699693203, + 0.024553334340453148, + 0.03456643223762512, + 0.01567327417433262, + 0.01429381500929594, + 0.002182555152103305, + 0.003888694103807211, + 0.0030692140571773052, + 0.0039296443574130535, + 0.0037893387489020824, + 0.002986591076478362, + 0.002114882692694664, + 0.0028383347671478987, + 0.010491140186786652, + 0.009810996241867542, + 0.020189853385090828, + 0.012256002984941006, + 0.007314884103834629, + 0.008768811821937561, + 0.004490331746637821, + 0.008959797210991383, + 0.006352582480758429, + 0.004183195531368256, + 0.007164597045630217, + 0.0042502074502408504, + 0.004302821587771177, + 0.0033217468298971653, + 0.0034977507311850786, + 0.0032288909424096346, + 0.013867032714188099, + 0.009459285996854305, + 0.007565957494080067, + 0.015195210464298725, + 0.005862903781235218, + 0.009834693744778633, + 0.005308868363499641, + 0.021311646327376366, + 0.012755520641803741, + 0.011849112808704376, + 0.00821103435009718, + 0.005140521563589573, + 0.009927582927048206, + 0.019833121448755264, + 0.006428815424442291, + 0.007939997129142284, + 0.005360541865229607, + 0.003476575016975403, + 0.004287002608180046, + 0.005144961178302765, + 0.004076736979186535, + 0.0035707736387848854, + 0.0028827660717070103, + 0.007932813838124275, + 0.015734117478132248, + 0.015037215314805508, + 0.01930946670472622, + 0.022408701479434967, + 0.015330919064581394, + 0.02042466029524803, + 0.016500897705554962, + 0.017290454357862473, + 0.030444994568824768, + 0.035672035068273544, + 0.03442196175456047, + 0.04284520074725151, + 0.021595217287540436, + 0.05463593453168869, + 0.0264681875705719, + 0.03054620511829853, + 0.004424897488206625, + 0.004392426926642656, + 0.005692780017852783, + 0.004595313686877489, + 0.0040043010376393795, + 0.005861559882760048, + 0.008598431013524532, + 0.005877687130123377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.11294842511415482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot tilts smoothly to the right, tracking a modern, stylish city bus as it glides to a stop at a bustling city center under the warm, golden light of afternoon. Painted in vibrant electric blue and crisp white, the bus's exterior gleams with polished surfaces, and expansive, tinted, panoramic windows reflect the cityscape. Along its sides, bright, high-resolution LED screens display dynamic, colorful abstract graphics that move and flow. With a quiet aerodynamic hiss, its stylish double-glazed doors slide open, releasing a diverse group of passengers onto the crowded sidewalk. Inside, the clean, brightly lit interior reveals rows of modern, molded seats, occupied by commuters engrossed in their phones and books. The driver, a woman in a crisp, dark blue uniform, greets the new passengers aboard with a friendly, professional greeting. As the doors close seamlessly, the bus pulls away from the curb and effortlessly blends into the bustling street traffic. The camera then widens to capture the bus navigating a vibrant urban canyon, its forms contrasting past ornate historic buildings and towering glass and steel skyscrapers, perfectly encapsulating the vibrant, pulsating nature of modern city life.", + "seed": 21000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000116/group_9/index_8014_seed_21000_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3573897870257497, + "vlm_logprob": -0.3573897870257497, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.3573897870257497, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00873801950365305 + ], + "opposite_logprob_tokens": [ + -5.383738040924072 + ], + "logprob_reward": -0.00873801950365305, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00873801950365305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10212381184101105, + 0.16226668655872345, + 0.16461783647537231, + 0.09705682098865509, + 0.0638519823551178, + 0.16734257340431213, + 0.05819050967693329, + 0.0823521837592125, + 0.05310805141925812, + 0.049089543521404266 + ], + "credit_map": [ + 0.015044313855469227, + 0.011800122447311878, + 0.013786520808935165, + 0.01159127987921238, + 0.009095381014049053, + 0.010373355820775032, + 0.020042961463332176, + 0.010389887727797031, + 0.03814220800995827, + 0.018926559016108513, + 0.014055905863642693, + 0.0255728829652071, + 0.011387785896658897, + 0.01112330798059702, + 0.013825921341776848, + 0.02923213504254818, + 0.03410881385207176, + 0.010898719541728497, + 0.007834101095795631, + 0.014938912354409695, + 0.010277925990521908, + 0.017940934747457504, + 0.011972864158451557, + 0.056645575910806656, + 0.015251191332936287, + 0.010745396837592125, + 0.016397006809711456, + 0.011993606574833393, + 0.007483361754566431, + 0.010936257429420948, + 0.012395729310810566, + 0.011854282580316067, + 0.008325029164552689, + 0.010623274371027946, + 0.01008873712271452, + 0.008917577564716339, + 0.005631745792925358, + 0.006438164506107569, + 0.007710708305239677, + 0.006116754375398159, + 0.01979260891675949, + 0.011958249844610691, + 0.02886115200817585, + 0.016492683440446854, + 0.010109445080161095, + 0.04722759872674942, + 0.011938566341996193, + 0.020962290465831757, + 0.004734061658382416, + 0.005794244352728128, + 0.005317178089171648, + 0.004727254621684551, + 0.012937730178236961, + 0.008846232667565346, + 0.007403860334306955, + 0.008429952897131443, + 0.011276339180767536, + 0.009555280208587646, + 0.01009714137762785, + 0.010337219573557377, + 0.011904072016477585, + 0.011048657819628716, + 0.010994840413331985, + 0.007138644810765982, + 0.007057070732116699, + 0.009720969945192337, + 0.006168506573885679, + 0.006916923448443413, + 0.0065539018251001835, + 0.005214666947722435, + 0.005889341235160828, + 0.005586676299571991, + 0.005513181909918785, + 0.007250001188367605, + 0.006344698369503021, + 0.006250154227018356, + 0.008968023583292961, + 0.006307743024080992, + 0.004720143973827362, + 0.0037356021348387003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.00873801950365305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609429344534874 + ], + "opposite_logprob_tokens": [ + -4.141094207763672 + ], + "logprob_reward": -0.01609429344534874, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609429344534874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10861917585134506, + 0.07623185962438583, + 0.11986953020095825, + 0.15067771077156067, + 0.07190118730068207, + 0.03785613924264908, + 0.2519100308418274, + 0.0819537565112114, + 0.04409950226545334, + 0.056881166994571686 + ], + "credit_map": [ + 0.01412274967879057, + 0.017047766596078873, + 0.011611411347985268, + 0.013018707744777203, + 0.013289982452988625, + 0.014724227599799633, + 0.01292172446846962, + 0.011882607825100422, + 0.009100755676627159, + 0.012652575969696045, + 0.007684739772230387, + 0.008778702467679977, + 0.010847999714314938, + 0.010287859477102757, + 0.00876305066049099, + 0.008116178214550018, + 0.012517619878053665, + 0.010905799455940723, + 0.006734342314302921, + 0.012881511822342873, + 0.010702154599130154, + 0.017698287963867188, + 0.022309161722660065, + 0.02612065151333809, + 0.015939762815833092, + 0.014712879434227943, + 0.015304258093237877, + 0.018601546064019203, + 0.017591409385204315, + 0.01795668713748455, + 0.025388585403561592, + 0.025182584300637245, + 0.0073038553819060326, + 0.011768794618546963, + 0.011237450875341892, + 0.008265667594969273, + 0.00644196942448616, + 0.00824460107833147, + 0.007928011938929558, + 0.010710831731557846, + 0.004481731913983822, + 0.002723871497437358, + 0.005899414885789156, + 0.004889831878244877, + 0.003910013474524021, + 0.006945928558707237, + 0.004293008707463741, + 0.004712337162345648, + 0.02324380725622177, + 0.024657871574163437, + 0.023993661627173424, + 0.012244459241628647, + 0.03144663944840431, + 0.10058348625898361, + 0.01588108018040657, + 0.01985905133187771, + 0.010027562268078327, + 0.01119293738156557, + 0.011270146816968918, + 0.009815662167966366, + 0.012704188004136086, + 0.011362153105437756, + 0.008313518948853016, + 0.00726758548989892, + 0.004514307249337435, + 0.0073653762228786945, + 0.005710020195692778, + 0.0045829834416508675, + 0.008832594379782677, + 0.004237630404531956, + 0.003984689246863127, + 0.00487190205603838, + 0.007474851794540882, + 0.009200804866850376, + 0.008190304040908813, + 0.008347079157829285, + 0.0066552236676216125, + 0.005715947598218918, + 0.005243848077952862, + 0.006053104996681213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.01609429344534874 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005091667175293 + ], + "opposite_logprob_tokens": [ + -0.22550910711288452 + ], + "logprob_reward": -1.6005091667175293, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005091667175293, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06772411614656448, + 0.44294530153274536, + 0.047855667769908905, + 0.05060352012515068, + 0.07652788609266281, + 0.06310120224952698, + 0.12408314645290375, + 0.050173282623291016, + 0.057334788143634796, + 0.019651105627417564 + ], + "credit_map": [ + 0.008994153700768948, + 0.009630977176129818, + 0.006139643024653196, + 0.007865076884627342, + 0.008581899106502533, + 0.010182731784880161, + 0.00811200961470604, + 0.008217624388635159, + 0.1280887871980667, + 0.047285888344049454, + 0.03190293163061142, + 0.05569256842136383, + 0.025130780413746834, + 0.032727330923080444, + 0.02235189452767372, + 0.09976513683795929, + 0.005145554430782795, + 0.004570175893604755, + 0.003096714848652482, + 0.0046671065501868725, + 0.003872216446325183, + 0.007748770527541637, + 0.008734408766031265, + 0.010020717978477478, + 0.00471817422658205, + 0.0050018769688904285, + 0.005877444054931402, + 0.011037176474928856, + 0.00392404617741704, + 0.0063902572728693485, + 0.0069394949823617935, + 0.006715045776218176, + 0.008632940240204334, + 0.011890536174178123, + 0.01780730113387108, + 0.008841807954013348, + 0.005748575087636709, + 0.00605172710493207, + 0.007584800478070974, + 0.009970204904675484, + 0.007905182428658009, + 0.005595079157501459, + 0.007844536565244198, + 0.007448277436196804, + 0.0043843029998242855, + 0.014880315400660038, + 0.006750676315277815, + 0.008292832411825657, + 0.012573455460369587, + 0.011610422283411026, + 0.009182787500321865, + 0.005587659776210785, + 0.026534870266914368, + 0.03969006612896919, + 0.009882239624857903, + 0.009021639823913574, + 0.007602758705615997, + 0.00474179582670331, + 0.00748335151001811, + 0.005390989128500223, + 0.005896511487662792, + 0.007971003651618958, + 0.006254518870264292, + 0.004832355305552483, + 0.005571635905653238, + 0.010708030313253403, + 0.010597153566777706, + 0.006282576825469732, + 0.007865812629461288, + 0.005566764157265425, + 0.005665421485900879, + 0.005077393725514412, + 0.002255834173411131, + 0.0026506823487579823, + 0.0032583463471382856, + 0.0023930519819259644, + 0.002491179620847106, + 0.0025555267930030823, + 0.0022548572160303593, + 0.0017916260985657573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -1.6005091667175293 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129613071680069 + ], + "opposite_logprob_tokens": [ + -2.2379612922668457 + ], + "logprob_reward": -0.1129613071680069, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129613071680069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1021471843123436, + 0.2639086842536926, + 0.10146904736757278, + 0.2188108265399933, + 0.10147573798894882, + 0.030808793380856514, + 0.07896953821182251, + 0.032878223806619644, + 0.04903886839747429, + 0.020493078976869583 + ], + "credit_map": [ + 0.01419241912662983, + 0.013360164128243923, + 0.010126634500920773, + 0.010859389789402485, + 0.011190621182322502, + 0.011581906117498875, + 0.020086703822016716, + 0.010749347507953644, + 0.051729992032051086, + 0.03620557487010956, + 0.02556128241121769, + 0.05005984380841255, + 0.019758880138397217, + 0.025191722437739372, + 0.02019808068871498, + 0.035203300416469574, + 0.019063210114836693, + 0.007565805222839117, + 0.005320331547409296, + 0.007124157156795263, + 0.0062281060963869095, + 0.013010586611926556, + 0.008658620528876781, + 0.03449823334813118, + 0.0209029633551836, + 0.023719701915979385, + 0.019552120938897133, + 0.026738157495856285, + 0.043351802974939346, + 0.025987878441810608, + 0.03196524456143379, + 0.026592962443828583, + 0.010302183218300343, + 0.017306098714470863, + 0.024536358192563057, + 0.011522449553012848, + 0.007834871299564838, + 0.008218812756240368, + 0.009056064300239086, + 0.012698902748525143, + 0.0035399608314037323, + 0.0033172115217894316, + 0.004252016544342041, + 0.0037461670581251383, + 0.0033380240201950073, + 0.005188444629311562, + 0.00329824211075902, + 0.004128727596253157, + 0.007723366376012564, + 0.007265607360750437, + 0.005015858449041843, + 0.004512538202106953, + 0.01244436576962471, + 0.0306007768958807, + 0.005515899043530226, + 0.0058911265805363655, + 0.004195202142000198, + 0.0035739410668611526, + 0.004475273657590151, + 0.0028763574082404375, + 0.004828140139579773, + 0.004830431658774614, + 0.00442107766866684, + 0.0036777975037693977, + 0.005591186229139566, + 0.009513359516859055, + 0.008168362081050873, + 0.004928386304527521, + 0.006127751432359219, + 0.004963552579283714, + 0.003814744995906949, + 0.00593152642250061, + 0.0019474056316539645, + 0.0030084438621997833, + 0.003162624780088663, + 0.0027117752470076084, + 0.0027493699453771114, + 0.0028392341919243336, + 0.002232142724096775, + 0.0018420823616907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.1129613071680069 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern city bus painted in vibrant electric blue and crisp white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bus feature expansive tinted panoramic windows and bright LED screens displaying dynamic abstract graphics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the bus's double-glazed doors sliding open to release passengers onto a crowded sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the interior visible showing a woman driver in a dark blue uniform greeting passengers while commuters sit in molded seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium-length shot that tilts smoothly to the right and later widens to show the bus navigating between historic buildings and skyscrapers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646148294210434 + ], + "opposite_logprob_tokens": [ + -3.0486462116241455 + ], + "logprob_reward": -0.048646148294210434, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646148294210434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.063181571662426, + 0.4095648229122162, + 0.051233936101198196, + 0.04077793285250664, + 0.034848276525735855, + 0.1320229023694992, + 0.13091646134853363, + 0.06618960946798325, + 0.041738204658031464, + 0.029526246711611748 + ], + "credit_map": [ + 0.00801905058324337, + 0.009070253930985928, + 0.006912468932569027, + 0.0057090590707957745, + 0.007616464979946613, + 0.00898690428584814, + 0.009551617316901684, + 0.00731574697420001, + 0.1467023640871048, + 0.028235873207449913, + 0.0272264014929533, + 0.0322684645652771, + 0.01662510074675083, + 0.028602274134755135, + 0.01693630777299404, + 0.11296802759170532, + 0.00891924649477005, + 0.004576442297548056, + 0.0044061182998120785, + 0.005514811258763075, + 0.003837340045720339, + 0.005847868509590626, + 0.004655486438423395, + 0.013476617634296417, + 0.005121724680066109, + 0.004172161221504211, + 0.005234917625784874, + 0.005310953129082918, + 0.005004294216632843, + 0.00477487500756979, + 0.0062901945784688, + 0.004868814721703529, + 0.00405176030471921, + 0.007271550130099058, + 0.0055192443542182446, + 0.003721690969541669, + 0.0032428810372948647, + 0.004090350586920977, + 0.0033692694269120693, + 0.003581529948860407, + 0.015852665528655052, + 0.010164499282836914, + 0.022141750901937485, + 0.01354240532964468, + 0.008600442670285702, + 0.03741365671157837, + 0.009812870062887669, + 0.014494623057544231, + 0.015125608071684837, + 0.017175301909446716, + 0.009465404786169529, + 0.006228214595466852, + 0.017339345067739487, + 0.04546985775232315, + 0.010030174627900124, + 0.010082554072141647, + 0.007076967507600784, + 0.006714144255965948, + 0.008168715983629227, + 0.007579923141747713, + 0.009912258014082909, + 0.008766680024564266, + 0.00907648541033268, + 0.008894432336091995, + 0.0059291040524840355, + 0.008970748633146286, + 0.004115592688322067, + 0.0032516515348106623, + 0.0064494493417441845, + 0.003939905669540167, + 0.004338953178375959, + 0.0047427997924387455, + 0.004451869986951351, + 0.003649735590443015, + 0.0035498826764523983, + 0.0050046406686306, + 0.00380727113224566, + 0.002994607435539365, + 0.003007187508046627, + 0.003061052178964019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.048646148294210434 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_0/index_tensor([3437, 3437])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_0/index_tensor([3437, 3437])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f083d55410e4d1a0039304a3e2817a8920bf5be0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_0/index_tensor([3437, 3437])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "superb macro lens, shot using high-speed photography, captures a moment of destructive beauty in extreme slow motion, while the camera performs a slow, smooth 360-degree orbital rotation around the central point of the shattering event. A flawless, crystalline crystal ball, its smooth surface reflecting a sharply defined spotlight from above, silently dents inward upon contact with a dark, highly polished, reflective obsidian floor. This explosion is not a chaotic spattering, but rather a silent blossoming of glass, with thousands of shimmering, sharp-edged fragments of varying sizes floating, tumbling, and spreading outward with grace and ease. Each fragment, like a single diamond, illuminated by a single, intense point light source, swirls and refracts a dazzling, iridescent spectrum, trailing fleeting, luminous trails against the pure black background, creating a surreal atmosphere that freezes the violent moment into a tranquil, radiant, and captivating visual spectacle.", + "seed": 7978, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_0/index_3437_seed_7978_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.28003592491149903, + "vlm_logprob": -0.28003592491149903, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.28003592491149903, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0583255290985107 + ], + "opposite_logprob_tokens": [ + -0.43332549929618835 + ], + "logprob_reward": -1.0583255290985107, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0583255290985107, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09314478188753128, + 0.35555654764175415, + 0.04616285115480423, + 0.030354339629411697, + 0.029088210314512253, + 0.013440273702144623, + 0.16823676228523254, + 0.09733336418867111, + 0.04463699832558632, + 0.12204594165086746 + ], + "credit_map": [ + 0.02040255442261696, + 0.00782096665352583, + 0.012113535776734352, + 0.006773530039936304, + 0.0069652884267270565, + 0.01563333161175251, + 0.008762583136558533, + 0.014672991819679737, + 0.12794305384159088, + 0.010021504946053028, + 0.027471227571368217, + 0.00644500320777297, + 0.011581397615373135, + 0.13163574039936066, + 0.006034743972122669, + 0.03442388027906418, + 0.0069405981339514256, + 0.004089036025106907, + 0.007658447604626417, + 0.007176581304520369, + 0.002908251713961363, + 0.004279000218957663, + 0.006980749312788248, + 0.006130191031843424, + 0.006336770486086607, + 0.0019611609168350697, + 0.004617319442331791, + 0.0021415872033685446, + 0.0017623493913561106, + 0.005429066717624664, + 0.0022642011754214764, + 0.005841883830726147, + 0.0036134887486696243, + 0.0024317053612321615, + 0.00497831217944622, + 0.004007560200989246, + 0.0014858492650091648, + 0.004909686744213104, + 0.0028889288660138845, + 0.004772679414600134, + 0.0018192899879068136, + 0.0018512126989662647, + 0.0022285631857812405, + 0.001507471315562725, + 0.0010874997824430466, + 0.0011896603973582387, + 0.0016079164342954755, + 0.002148660598322749, + 0.01662949100136757, + 0.0033209803514182568, + 0.060051869601011276, + 0.005692875478416681, + 0.005340314004570246, + 0.0071420264430344105, + 0.00984366424381733, + 0.060215529054403305, + 0.010472068563103676, + 0.0109831802546978, + 0.010644366033375263, + 0.015213839709758759, + 0.008815945126116276, + 0.016406606882810593, + 0.010338260792195797, + 0.0144590949639678, + 0.005784125998616219, + 0.004010975360870361, + 0.004719747696071863, + 0.0051091983914375305, + 0.004997984040528536, + 0.0069978381507098675, + 0.007201723288744688, + 0.005815406795591116, + 0.02837809920310974, + 0.005115583539009094, + 0.006953778676688671, + 0.008088513277471066, + 0.006698183249682188, + 0.008252053521573544, + 0.04219610616564751, + 0.01636362075805664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -1.0583255290985107 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016130540519952774 + ], + "opposite_logprob_tokens": [ + -4.141130447387695 + ], + "logprob_reward": -0.016130540519952774, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016130540519952774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07643067836761475, + 0.527108907699585, + 0.07492777705192566, + 0.030648743733763695, + 0.02328219637274742, + 0.008713031187653542, + 0.052027519792318344, + 0.05925903841853142, + 0.026915542781352997, + 0.12068656086921692 + ], + "credit_map": [ + 0.012422203086316586, + 0.007502228487282991, + 0.011130589991807938, + 0.0055245873518288136, + 0.007571992930024862, + 0.011310875415802002, + 0.008308778516948223, + 0.012659420259296894, + 0.20384804904460907, + 0.01692870818078518, + 0.04403359442949295, + 0.009062434546649456, + 0.016968926414847374, + 0.19021573662757874, + 0.0077043515630066395, + 0.03834714740514755, + 0.006758861243724823, + 0.0059007457457482815, + 0.010181003250181675, + 0.014722106978297234, + 0.006603054236620665, + 0.006426410283893347, + 0.013669691048562527, + 0.010665904730558395, + 0.006195941008627415, + 0.0033366905990988016, + 0.004921319428831339, + 0.002679733792319894, + 0.0019230765756219625, + 0.004160678014159203, + 0.0024607062805444, + 0.00497059803456068, + 0.002589860465377569, + 0.0018168686656281352, + 0.003953232895582914, + 0.003365395125001669, + 0.0012252803426235914, + 0.0030997593421489, + 0.0018316622590646148, + 0.005400138441473246, + 0.001056567532941699, + 0.0014383895322680473, + 0.0014336254680529237, + 0.0009327339939773083, + 0.0006861294968985021, + 0.000834306119941175, + 0.0011394981993362308, + 0.0011917807860299945, + 0.00510716438293457, + 0.0015105531783774495, + 0.0171711016446352, + 0.002278832718729973, + 0.0017599110724404454, + 0.002524738200008869, + 0.004082621540874243, + 0.017592595890164375, + 0.0082189179956913, + 0.006430246401578188, + 0.0054700057953596115, + 0.006395117845386267, + 0.005219745449721813, + 0.011966567486524582, + 0.005949532613158226, + 0.00960890669375658, + 0.003802143968641758, + 0.0025972493458539248, + 0.0030771344900131226, + 0.002886540489271283, + 0.0033538942225277424, + 0.004333093762397766, + 0.0036576546262949705, + 0.00320783001370728, + 0.027981193736195564, + 0.004815402440726757, + 0.00556890619918704, + 0.008637906983494759, + 0.006946847308427095, + 0.007310341112315655, + 0.04319705814123154, + 0.01622891053557396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.016130540519952774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647880792617798 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647880792617798, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647880792617798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12211433053016663, + 0.1592869609594345, + 0.1032707542181015, + 0.0955619290471077, + 0.04971667379140854, + 0.028328485786914825, + 0.14370307326316833, + 0.06517240405082703, + 0.06367014348506927, + 0.16917528212070465 + ], + "credit_map": [ + 0.02005116268992424, + 0.009299125522375107, + 0.01777525432407856, + 0.00932378601282835, + 0.01014221366494894, + 0.018833231180906296, + 0.011604509316384792, + 0.025085046887397766, + 0.04473891481757164, + 0.005944198928773403, + 0.011665722355246544, + 0.006013518199324608, + 0.005686148535460234, + 0.05586862191557884, + 0.0059229303151369095, + 0.02344689890742302, + 0.01131279394030571, + 0.0060349698178470135, + 0.024096466600894928, + 0.012179186567664146, + 0.006010616198182106, + 0.008722562342882156, + 0.010981103405356407, + 0.023933054879307747, + 0.017415540292859077, + 0.008759037591516972, + 0.012891565449535847, + 0.00992652028799057, + 0.009809950366616249, + 0.014373335987329483, + 0.009603166952729225, + 0.012782808393239975, + 0.005253770854324102, + 0.0025445104110985994, + 0.010869856923818588, + 0.004161326214671135, + 0.00203182315453887, + 0.008978679776191711, + 0.003167596645653248, + 0.012709107249975204, + 0.0033389218151569366, + 0.004115310497581959, + 0.004911676049232483, + 0.003155594691634178, + 0.0024934308603405952, + 0.0029066859278827906, + 0.0033752811141312122, + 0.004031584598124027, + 0.024142712354660034, + 0.00747166620567441, + 0.03766554966568947, + 0.00640767440199852, + 0.005966763943433762, + 0.010036153718829155, + 0.010736615397036076, + 0.041275933384895325, + 0.009616685099899769, + 0.007081546820700169, + 0.007225639186799526, + 0.009758864529430866, + 0.005826173350214958, + 0.007495956029742956, + 0.007599672768265009, + 0.010567866265773773, + 0.009625600650906563, + 0.005893551278859377, + 0.007008098065853119, + 0.005644272547215223, + 0.007665548007935286, + 0.014392419718205929, + 0.004694956820458174, + 0.008745690807700157, + 0.033955950289964676, + 0.00529848225414753, + 0.007334237918257713, + 0.011267095804214478, + 0.009763390757143497, + 0.010419911704957485, + 0.06380713731050491, + 0.02732907049357891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.004647880792617798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637467235326767 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637467235326767, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637467235326767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16326549649238586, + 0.1612628698348999, + 0.2572990953922272, + 0.034075990319252014, + 0.1316341757774353, + 0.03491734713315964, + 0.07970928400754929, + 0.02364787831902504, + 0.07869504392147064, + 0.03549287095665932 + ], + "credit_map": [ + 0.019952522590756416, + 0.00948166474699974, + 0.03081907331943512, + 0.01358370017260313, + 0.009732996113598347, + 0.0262921079993248, + 0.0160251185297966, + 0.03737831488251686, + 0.05342715233564377, + 0.004906312562525272, + 0.016111692413687706, + 0.004454713314771652, + 0.004447826649993658, + 0.016350043937563896, + 0.0060716355219483376, + 0.05549350008368492, + 0.028213003650307655, + 0.017317865043878555, + 0.043131258338689804, + 0.0467139333486557, + 0.016737567260861397, + 0.02017853781580925, + 0.045158252120018005, + 0.03984867036342621, + 0.005957770626991987, + 0.003068308113142848, + 0.005321298260241747, + 0.003580341814085841, + 0.002312679309397936, + 0.004874986130744219, + 0.004298424813896418, + 0.0046621812507510185, + 0.01567242294549942, + 0.006230332423001528, + 0.02882210537791252, + 0.013154863379895687, + 0.004419445060193539, + 0.027356352657079697, + 0.008269693702459335, + 0.027708949521183968, + 0.004409519024193287, + 0.004498826805502176, + 0.005730656906962395, + 0.004233554936945438, + 0.0030920756980776787, + 0.0034122460056096315, + 0.0038338485173881054, + 0.005706618074327707, + 0.015989987179636955, + 0.004558340180665255, + 0.01713273674249649, + 0.0036646793596446514, + 0.0037085912190377712, + 0.007420482579618692, + 0.0070481798611581326, + 0.020186280831694603, + 0.004097771365195513, + 0.0026394256856292486, + 0.0030491219367831945, + 0.002900205086916685, + 0.0020053889602422714, + 0.0028111811261624098, + 0.0028320197016000748, + 0.0033127623610198498, + 0.012695690616965294, + 0.007153928745537996, + 0.008389611728489399, + 0.006069251801818609, + 0.008246678858995438, + 0.018213113769888878, + 0.007715391460806131, + 0.010211367160081863, + 0.007115560118108988, + 0.00271081761457026, + 0.0035843015648424625, + 0.0029517801012843847, + 0.0027129394002258778, + 0.003254538169130683, + 0.00543238827958703, + 0.007730542216449976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007637467235326767 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343820691108704 + ], + "opposite_logprob_tokens": [ + -1.3134381771087646 + ], + "logprob_reward": -0.31343820691108704, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343820691108704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05973058566451073, + 0.5617759227752686, + 0.06446780264377594, + 0.029650026932358742, + 0.03129787743091583, + 0.031568195670843124, + 0.08283523470163345, + 0.04206131398677826, + 0.03352254629135132, + 0.06309057772159576 + ], + "credit_map": [ + 0.010947221890091896, + 0.00489114411175251, + 0.008337910287082195, + 0.004367088433355093, + 0.0065392120741307735, + 0.009011290036141872, + 0.007099189795553684, + 0.008537515997886658, + 0.20500127971172333, + 0.01575707085430622, + 0.043911226093769073, + 0.00971817784011364, + 0.0179611649364233, + 0.20761379599571228, + 0.008973547257483006, + 0.05283955857157707, + 0.005035589449107647, + 0.005037394352257252, + 0.013570060022175312, + 0.008515559136867523, + 0.005497068632394075, + 0.005254879128187895, + 0.009325266815721989, + 0.012231971137225628, + 0.00649447413161397, + 0.0023177836555987597, + 0.004367087502032518, + 0.0023383060470223427, + 0.00198099366389215, + 0.004995438270270824, + 0.0030036272946745157, + 0.004152311012148857, + 0.0044041406363248825, + 0.002009524032473564, + 0.005850767716765404, + 0.0027883604634553194, + 0.0012154997093603015, + 0.004813929554075003, + 0.002275483449921012, + 0.007940168492496014, + 0.0030762965325266123, + 0.004147678148001432, + 0.00551288528367877, + 0.0036053601652383804, + 0.0028513658326119184, + 0.0034900852479040623, + 0.0035156304948031902, + 0.0053688883781433105, + 0.012236136011779308, + 0.002648566849529743, + 0.026500999927520752, + 0.0030016559176146984, + 0.0033471849747002125, + 0.004940282553434372, + 0.005000378470867872, + 0.025160014629364014, + 0.006354154087603092, + 0.004460746422410011, + 0.004439953248947859, + 0.0044226073659956455, + 0.0032066761050373316, + 0.008196070790290833, + 0.00372545188292861, + 0.007255646865814924, + 0.004604781977832317, + 0.003481228370219469, + 0.002963258884847164, + 0.0029727707151323557, + 0.004278101492673159, + 0.0076565612107515335, + 0.0029154440853744745, + 0.004650392569601536, + 0.020586801692843437, + 0.0031754167284816504, + 0.003831540932878852, + 0.006268535740673542, + 0.0037628731224685907, + 0.005350656341761351, + 0.008219356648623943, + 0.01189538836479187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.31343820691108704 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "superb macro lens, shot using high-speed photography, captures a moment of destructive beauty in extreme slow motion, while the camera performs a slow, smooth 360-degree orbital rotation around the central point of the shattering event. A flawless, crystalline crystal ball, its smooth surface reflecting a sharply defined spotlight from above, silently dents inward upon contact with a dark, highly polished, reflective obsidian floor. This explosion is not a chaotic spattering, but rather a silent blossoming of glass, with thousands of shimmering, sharp-edged fragments of varying sizes floating, tumbling, and spreading outward with grace and ease. Each fragment, like a single diamond, illuminated by a single, intense point light source, swirls and refracts a dazzling, iridescent spectrum, trailing fleeting, luminous trails against the pure black background, creating a surreal atmosphere that freezes the violent moment into a tranquil, radiant, and captivating visual spectacle.", + "seed": 7979, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_0/index_3437_seed_7979_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15800851141102612, + "vlm_logprob": -0.15800851141102612, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.15800851141102612, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5815073251724243 + ], + "opposite_logprob_tokens": [ + -0.8315073251724243 + ], + "logprob_reward": -0.5815073251724243, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5815073251724243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35125312209129333, + 0.15180444717407227, + 0.10392971336841583, + 0.05076573044061661, + 0.043134767562150955, + 0.05462087690830231, + 0.150532066822052, + 0.03665769100189209, + 0.02887868322432041, + 0.028422869741916656 + ], + "credit_map": [ + 0.03933590650558472, + 0.025488262996077538, + 0.07354890555143356, + 0.02929561957716942, + 0.02909895032644272, + 0.04547981172800064, + 0.04487397149205208, + 0.06413168460130692, + 0.015939833596348763, + 0.01420347299426794, + 0.03750788792967796, + 0.00950529146939516, + 0.016319047659635544, + 0.02177038975059986, + 0.011320657096803188, + 0.025237876921892166, + 0.01093323901295662, + 0.009827274829149246, + 0.015745507553219795, + 0.011509341187775135, + 0.008891727775335312, + 0.013561172410845757, + 0.012239831499755383, + 0.021221619099378586, + 0.0076026879251003265, + 0.004987065214663744, + 0.007358490489423275, + 0.0074912323616445065, + 0.0037979725748300552, + 0.00825526937842369, + 0.0034809105563908815, + 0.0077921003103256226, + 0.005812433082610369, + 0.008958786725997925, + 0.006629603449255228, + 0.005918711423873901, + 0.003584115533158183, + 0.004497853573411703, + 0.0035505483392626047, + 0.004182714503258467, + 0.006658127997070551, + 0.01122590433806181, + 0.005984319373965263, + 0.006038339342921972, + 0.006869097240269184, + 0.007028714753687382, + 0.005152854137122631, + 0.0056635187938809395, + 0.021399157121777534, + 0.02452678605914116, + 0.01412574015557766, + 0.029008645564317703, + 0.015404659323394299, + 0.020744353532791138, + 0.013378827832639217, + 0.011943895369768143, + 0.004338786471635103, + 0.0033825323916971684, + 0.0064155105501413345, + 0.004634444136172533, + 0.0037184073589742184, + 0.004991891328245401, + 0.003986191935837269, + 0.005189924035221338, + 0.003984303213655949, + 0.003609965555369854, + 0.0027312610764056444, + 0.0019575469195842743, + 0.0032964942511171103, + 0.004669769201427698, + 0.005126780364662409, + 0.003502564039081335, + 0.003350485349074006, + 0.0033159954473376274, + 0.0035272708628326654, + 0.0022877731826156378, + 0.004688336048275232, + 0.0063394345343112946, + 0.0023856814950704575, + 0.002527892356738448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.5815073251724243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02334688790142536 + ], + "opposite_logprob_tokens": [ + -3.7733469009399414 + ], + "logprob_reward": -0.02334688790142536, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02334688790142536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2308095246553421, + 0.277383953332901, + 0.09951110184192657, + 0.07281360775232315, + 0.02536839246749878, + 0.037290308624506, + 0.08829297125339508, + 0.07048404961824417, + 0.06389520317316055, + 0.0341508686542511 + ], + "credit_map": [ + 0.0325479656457901, + 0.013982795178890228, + 0.038457226008176804, + 0.014895487576723099, + 0.016578281298279762, + 0.031990863382816315, + 0.016074281185865402, + 0.06628262996673584, + 0.030567334964871407, + 0.032854944467544556, + 0.04880117252469063, + 0.025804970413446426, + 0.035348400473594666, + 0.036431215703487396, + 0.028646018356084824, + 0.03892989456653595, + 0.015840670093894005, + 0.008378393948078156, + 0.015605971217155457, + 0.008593715727329254, + 0.007058642338961363, + 0.015428316779434681, + 0.010805084370076656, + 0.017800303176045418, + 0.015037446282804012, + 0.007573661860078573, + 0.01178015023469925, + 0.00821004994213581, + 0.005794531665742397, + 0.012397682294249535, + 0.0046536847949028015, + 0.007366398349404335, + 0.003963608294725418, + 0.0033501528669148684, + 0.002979092998430133, + 0.003560046898201108, + 0.002397674135863781, + 0.0024272138252854347, + 0.0029118352103978395, + 0.003778766840696335, + 0.006190615706145763, + 0.006906813941895962, + 0.0038355186115950346, + 0.003222221042960882, + 0.00519120879471302, + 0.0048790969885885715, + 0.0027963905595242977, + 0.004268443677574396, + 0.01171522680670023, + 0.01642131246626377, + 0.00884901825338602, + 0.015139245428144932, + 0.009105577133595943, + 0.012160413898527622, + 0.007330870721489191, + 0.007571306545287371, + 0.017883220687508583, + 0.003975869156420231, + 0.0062182131223380566, + 0.009080550633370876, + 0.005649906117469072, + 0.01795389875769615, + 0.005585154052823782, + 0.004137240815907717, + 0.006051911041140556, + 0.006792508997023106, + 0.007173809688538313, + 0.006017582956701517, + 0.007202595937997103, + 0.009979627095162868, + 0.010131662711501122, + 0.010545507073402405, + 0.003461097599938512, + 0.003846449311822653, + 0.003315080190077424, + 0.0033062943257391453, + 0.005524469539523125, + 0.007838976569473743, + 0.0036076472606509924, + 0.003250853158533573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02334688790142536 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024984830524772406 + ], + "opposite_logprob_tokens": [ + -6.002498626708984 + ], + "logprob_reward": -0.0024984830524772406, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024984830524772406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10431968420743942, + 0.056783612817525864, + 0.03639242425560951, + 0.04126632213592529, + 0.023670002818107605, + 0.0508689321577549, + 0.06076626107096672, + 0.5698314905166626, + 0.029546307399868965, + 0.02655496634542942 + ], + "credit_map": [ + 0.011385328136384487, + 0.008697801269590855, + 0.020131567493081093, + 0.00869639590382576, + 0.00903013814240694, + 0.014163125306367874, + 0.013001683168113232, + 0.01921365037560463, + 0.005436671432107687, + 0.006425985135138035, + 0.009614122100174427, + 0.004876977298408747, + 0.008345311507582664, + 0.007892419584095478, + 0.006224931683391333, + 0.007967201992869377, + 0.0059120142832398415, + 0.003512898925691843, + 0.005816563498228788, + 0.0032101753167808056, + 0.0024784947745501995, + 0.005440825130790472, + 0.0035738248843699694, + 0.0064476290717720985, + 0.0068143014796078205, + 0.003491015871986747, + 0.006408780813217163, + 0.00510152243077755, + 0.0038237206172198057, + 0.006628945004194975, + 0.003709490643814206, + 0.0052885510958731174, + 0.003390291705727577, + 0.004044916480779648, + 0.002578666200861335, + 0.003220590064302087, + 0.0025576944462955, + 0.0027931605000048876, + 0.0022827615030109882, + 0.002801925642415881, + 0.005807536654174328, + 0.008729699067771435, + 0.005559769459068775, + 0.005913358647376299, + 0.008016427047550678, + 0.0066381492651999, + 0.004193857777863741, + 0.006010144948959351, + 0.008472778834402561, + 0.010279899463057518, + 0.005702788010239601, + 0.01009585801512003, + 0.0067237443290650845, + 0.009267634712159634, + 0.005129045341163874, + 0.00509451562538743, + 0.20843669772148132, + 0.024660823866724968, + 0.026635820046067238, + 0.017871228978037834, + 0.028448395431041718, + 0.23003602027893066, + 0.01290190126746893, + 0.020840616896748543, + 0.003767200745642185, + 0.002888589398935437, + 0.0027653840370476246, + 0.0024509201757609844, + 0.0035521395038813353, + 0.004729145206511021, + 0.005695601459592581, + 0.003697330132126808, + 0.0031397314742207527, + 0.0037996384780853987, + 0.0034180639777332544, + 0.002116215182468295, + 0.0043921624310314655, + 0.00535682262852788, + 0.00214221328496933, + 0.0021901216823607683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0024984830524772406 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028246049769222736 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028246049769222736, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028246049769222736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11239723116159439, + 0.29265233874320984, + 0.08156780898571014, + 0.07857981324195862, + 0.03392953425645828, + 0.05822008103132248, + 0.09113668650388718, + 0.17541609704494476, + 0.02468935027718544, + 0.05141106992959976 + ], + "credit_map": [ + 0.011863356456160545, + 0.009350860491394997, + 0.020293015986680984, + 0.010283774696290493, + 0.011418460868299007, + 0.015271449461579323, + 0.0108705535531044, + 0.02304576151072979, + 0.029884878545999527, + 0.04226483777165413, + 0.04860628396272659, + 0.026146432384848595, + 0.04339643567800522, + 0.040379445999860764, + 0.029536642134189606, + 0.032437391579151154, + 0.014187989756464958, + 0.007964788004755974, + 0.012312299571931362, + 0.008590998128056526, + 0.008463998325169086, + 0.010423341765999794, + 0.006992666516453028, + 0.012631727382540703, + 0.012864791788160801, + 0.007948040962219238, + 0.0122208371758461, + 0.01078311912715435, + 0.007032278925180435, + 0.012598201632499695, + 0.005810091737657785, + 0.0093224560841918, + 0.004888912662863731, + 0.0034645418636500835, + 0.00591828441247344, + 0.004450739361345768, + 0.0033866199664771557, + 0.003257612930610776, + 0.004340664483606815, + 0.004222157411277294, + 0.008149895817041397, + 0.005777058657258749, + 0.0053367651998996735, + 0.010295223444700241, + 0.0071034738793969154, + 0.007920149713754654, + 0.006783889140933752, + 0.006853623315691948, + 0.01142802182585001, + 0.016330938786268234, + 0.008434712886810303, + 0.01910748891532421, + 0.008827450685203075, + 0.01163069624453783, + 0.008127929642796516, + 0.007249441929161549, + 0.06075840815901756, + 0.009886528365314007, + 0.009429246187210083, + 0.0076437778770923615, + 0.008387588895857334, + 0.06333427131175995, + 0.00903609860688448, + 0.006940172053873539, + 0.002558587584644556, + 0.0024995070416480303, + 0.0029252655804157257, + 0.00209141056984663, + 0.003158217528834939, + 0.004038380458950996, + 0.004662216641008854, + 0.0027557637076824903, + 0.004576345440000296, + 0.006580055225640535, + 0.005347992293536663, + 0.003636811161413789, + 0.008675642311573029, + 0.011660708114504814, + 0.006738274358212948, + 0.004195236600935459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0028246049769222736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798652559518814 + ], + "opposite_logprob_tokens": [ + -1.8048652410507202 + ], + "logprob_reward": -0.1798652559518814, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798652559518814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22027595341205597, + 0.10717469453811646, + 0.07218312472105026, + 0.0909799188375473, + 0.06552927196025848, + 0.1304320991039276, + 0.1424807608127594, + 0.06891913712024689, + 0.0578032024204731, + 0.04422181472182274 + ], + "credit_map": [ + 0.022489873692393303, + 0.012548576109111309, + 0.047768618911504745, + 0.016396543011069298, + 0.01956907846033573, + 0.030325405299663544, + 0.025219351053237915, + 0.04595852643251419, + 0.011258771643042564, + 0.009573612362146378, + 0.021727511659264565, + 0.009181641042232513, + 0.011527040041983128, + 0.014228037558495998, + 0.013374590314924717, + 0.01630348712205887, + 0.012884840369224548, + 0.006605465896427631, + 0.009669042192399502, + 0.006162263918668032, + 0.0063581811264157295, + 0.011365147307515144, + 0.005139930173754692, + 0.013998256996273994, + 0.015534724108874798, + 0.010443762876093388, + 0.013415801338851452, + 0.013161196373403072, + 0.006875460501760244, + 0.01589902490377426, + 0.00625590467825532, + 0.00939403660595417, + 0.008747382089495659, + 0.008952760137617588, + 0.00925503484904766, + 0.008196152746677399, + 0.007652088068425655, + 0.005771683994680643, + 0.008253742009401321, + 0.008700428530573845, + 0.021040763705968857, + 0.03617367520928383, + 0.010865963064134121, + 0.008290396071970463, + 0.014149901457130909, + 0.017683541402220726, + 0.010080170817673206, + 0.012147693894803524, + 0.018517807126045227, + 0.028652019798755646, + 0.014673294499516487, + 0.020323842763900757, + 0.015127990394830704, + 0.0202399380505085, + 0.011423158459365368, + 0.01352271530777216, + 0.011804197914898396, + 0.00817417073994875, + 0.010352088138461113, + 0.005081444978713989, + 0.005723814945667982, + 0.013229584321379662, + 0.005465442314743996, + 0.009088391438126564, + 0.012345187366008759, + 0.007809034548699856, + 0.004633849952369928, + 0.0035857174079865217, + 0.006280788220465183, + 0.008261153474450111, + 0.008817543275654316, + 0.00606992794200778, + 0.00565209798514843, + 0.005279524717479944, + 0.0063560232520103455, + 0.005135726183652878, + 0.007068426348268986, + 0.005612984299659729, + 0.0051911259070038795, + 0.003925906959921122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.1798652559518814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "superb macro lens, shot using high-speed photography, captures a moment of destructive beauty in extreme slow motion, while the camera performs a slow, smooth 360-degree orbital rotation around the central point of the shattering event. A flawless, crystalline crystal ball, its smooth surface reflecting a sharply defined spotlight from above, silently dents inward upon contact with a dark, highly polished, reflective obsidian floor. This explosion is not a chaotic spattering, but rather a silent blossoming of glass, with thousands of shimmering, sharp-edged fragments of varying sizes floating, tumbling, and spreading outward with grace and ease. Each fragment, like a single diamond, illuminated by a single, intense point light source, swirls and refracts a dazzling, iridescent spectrum, trailing fleeting, luminous trails against the pure black background, creating a surreal atmosphere that freezes the violent moment into a tranquil, radiant, and captivating visual spectacle.", + "seed": 7980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_0/index_3437_seed_7980_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3220352653414011, + "vlm_logprob": -0.3220352653414011, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.3220352653414011, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1430206298828125 + ], + "opposite_logprob_tokens": [ + -0.3930206000804901 + ], + "logprob_reward": -1.1430206298828125, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1430206298828125, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0916796326637268, + 0.08706727623939514, + 0.08927170187234879, + 0.041912976652383804, + 0.04458915814757347, + 0.038562048226594925, + 0.048751816153526306, + 0.09299445152282715, + 0.41268765926361084, + 0.0524832159280777 + ], + "credit_map": [ + 0.011574587784707546, + 0.014014970511198044, + 0.011982140131294727, + 0.014204521663486958, + 0.014509809203445911, + 0.008815414272248745, + 0.009634220972657204, + 0.0069439830258488655, + 0.01372250821441412, + 0.009134496562182903, + 0.009914200752973557, + 0.009614556096494198, + 0.007959975861012936, + 0.013226794078946114, + 0.01253579743206501, + 0.010958955623209476, + 0.029082849621772766, + 0.005174730904400349, + 0.010147015564143658, + 0.005700566340237856, + 0.007380410563200712, + 0.009868040680885315, + 0.006636013742536306, + 0.01528208702802658, + 0.005541898310184479, + 0.005739032290875912, + 0.004211262334138155, + 0.006042572669684887, + 0.005374362226575613, + 0.0051504941657185555, + 0.005646592006087303, + 0.004206767771393061, + 0.004685582593083382, + 0.0075467294082045555, + 0.0047614132054150105, + 0.005843643099069595, + 0.004978088196367025, + 0.004868702031672001, + 0.007553809322416782, + 0.004351194016635418, + 0.0037930079270154238, + 0.006326553411781788, + 0.004228207282721996, + 0.006131056230515242, + 0.003961476497352123, + 0.004822489805519581, + 0.004957885015755892, + 0.004341378808021545, + 0.005264272913336754, + 0.00828174501657486, + 0.004652740433812141, + 0.005490738898515701, + 0.0038846980314701796, + 0.008581507951021194, + 0.00650026835501194, + 0.006095848046243191, + 0.01701473630964756, + 0.008320928551256657, + 0.012636816129088402, + 0.011035493575036526, + 0.0071632107719779015, + 0.012973828241229057, + 0.009423763491213322, + 0.014425677247345448, + 0.04100416600704193, + 0.02701284922659397, + 0.14846672117710114, + 0.021485833451151848, + 0.02031022496521473, + 0.11866004765033722, + 0.01593414880335331, + 0.019813699647784233, + 0.005167212802916765, + 0.0058415960520505905, + 0.008128183893859386, + 0.004325452260673046, + 0.005115218460559845, + 0.008887144736945629, + 0.004874133039265871, + 0.010144281201064587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -1.1430206298828125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023387065157294273 + ], + "opposite_logprob_tokens": [ + -3.7733869552612305 + ], + "logprob_reward": -0.023387065157294273, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023387065157294273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27296873927116394, + 0.11576999723911285, + 0.15727637708187103, + 0.059310801327228546, + 0.03665747120976448, + 0.0850270465016365, + 0.101570725440979, + 0.06520073115825653, + 0.04101930186152458, + 0.06519874185323715 + ], + "credit_map": [ + 0.024374866858124733, + 0.036206018179655075, + 0.03209376335144043, + 0.03827103227376938, + 0.0450935997068882, + 0.0307915098965168, + 0.03284790739417076, + 0.033290062099695206, + 0.016346387565135956, + 0.01202467456459999, + 0.014344080351293087, + 0.012105715461075306, + 0.011015316471457481, + 0.016540946438908577, + 0.015752436593174934, + 0.01764043979346752, + 0.06199606508016586, + 0.008975788950920105, + 0.021945934742689133, + 0.00609713327139616, + 0.005676919128745794, + 0.014912310056388378, + 0.005975425709038973, + 0.03169681131839752, + 0.01065628882497549, + 0.009058964438736439, + 0.007178725674748421, + 0.0057954732328653336, + 0.009082330390810966, + 0.00578547827899456, + 0.006451497320085764, + 0.005302043166011572, + 0.003936307039111853, + 0.0044520022347569466, + 0.003409582655876875, + 0.004023135639727116, + 0.0036295349709689617, + 0.005246065091341734, + 0.007408772595226765, + 0.004552067257463932, + 0.01591521129012108, + 0.011365422047674656, + 0.009307235479354858, + 0.008410358801484108, + 0.006385657005012035, + 0.019892048090696335, + 0.006579427979886532, + 0.007171689998358488, + 0.009612273424863815, + 0.010750989429652691, + 0.010224300436675549, + 0.009088542312383652, + 0.017653940245509148, + 0.016031306236982346, + 0.013872935436666012, + 0.014336437918245792, + 0.01099343691021204, + 0.007374153006821871, + 0.00842535775154829, + 0.009801708161830902, + 0.005331399850547314, + 0.009331277571618557, + 0.00761563191190362, + 0.006327769719064236, + 0.007710047997534275, + 0.005768537987023592, + 0.004807574674487114, + 0.0036967636551707983, + 0.003804722335189581, + 0.004732124973088503, + 0.0038111500907689333, + 0.006688380613923073, + 0.006178131327033043, + 0.005404748022556305, + 0.011456120759248734, + 0.005455802194774151, + 0.006255574058741331, + 0.009820807725191116, + 0.007693950552493334, + 0.012933610007166862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.023387065157294273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609429344534874 + ], + "opposite_logprob_tokens": [ + -4.141094207763672 + ], + "logprob_reward": -0.01609429344534874, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609429344534874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14313572645187378, + 0.10415489226579666, + 0.262892484664917, + 0.05990402400493622, + 0.022860223427414894, + 0.03101308085024357, + 0.11244578659534454, + 0.0660058930516243, + 0.1604294627904892, + 0.037158407270908356 + ], + "credit_map": [ + 0.01416021678596735, + 0.017865516245365143, + 0.016851656138896942, + 0.01610092632472515, + 0.020877737551927567, + 0.016412662342190742, + 0.020015275105834007, + 0.0208517387509346, + 0.015295097604393959, + 0.011131876148283482, + 0.012183773331344128, + 0.010357263498008251, + 0.010216043330729008, + 0.01688488945364952, + 0.014319273643195629, + 0.013766669668257236, + 0.11106697469949722, + 0.014140459708869457, + 0.03687269613146782, + 0.00522959278896451, + 0.006228366866707802, + 0.023019440472126007, + 0.007399960421025753, + 0.05893498286604881, + 0.010870026424527168, + 0.008281152695417404, + 0.007812083698809147, + 0.006519467104226351, + 0.008526569232344627, + 0.006549197249114513, + 0.0053816367872059345, + 0.005963892675936222, + 0.0034462830517441034, + 0.0032420463394373655, + 0.0023179161362349987, + 0.0026631273794919252, + 0.0022493915166705847, + 0.003265350591391325, + 0.0030774609185755253, + 0.002598647028207779, + 0.005918772891163826, + 0.0030458930414170027, + 0.00440219696611166, + 0.003674795851111412, + 0.0022358170244842768, + 0.006170459557324648, + 0.0024906801991164684, + 0.0030744667164981365, + 0.013714378699660301, + 0.013483098708093166, + 0.007942753843963146, + 0.011782757006585598, + 0.01582472212612629, + 0.018777688965201378, + 0.016343001276254654, + 0.014577384106814861, + 0.010978288017213345, + 0.007572180591523647, + 0.008221540600061417, + 0.00980002898722887, + 0.005381752736866474, + 0.009825732558965683, + 0.0075742038898169994, + 0.006652161478996277, + 0.016746120527386665, + 0.014524951577186584, + 0.05500638857483864, + 0.007345902733504772, + 0.008013536222279072, + 0.04595155268907547, + 0.005815862212330103, + 0.007025151513516903, + 0.0038748218212276697, + 0.0038810553960502148, + 0.0065878527238965034, + 0.0030253110453486443, + 0.003459265222772956, + 0.005785779096186161, + 0.0042938473634421825, + 0.0062504718080163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.01609429344534874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901675999164581 + ], + "opposite_logprob_tokens": [ + -2.57901668548584 + ], + "logprob_reward": -0.07901675999164581, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901675999164581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11573193222284317, + 0.1081925481557846, + 0.32885196805000305, + 0.08877129852771759, + 0.011342590674757957, + 0.02416136860847473, + 0.10997141897678375, + 0.11253143846988678, + 0.08119870722293854, + 0.019246747717261314 + ], + "credit_map": [ + 0.010764681734144688, + 0.014368447475135326, + 0.012385988608002663, + 0.012856699526309967, + 0.01773439720273018, + 0.013041755184531212, + 0.017914585769176483, + 0.016665369272232056, + 0.016743535175919533, + 0.010962282307446003, + 0.013532704673707485, + 0.010221141390502453, + 0.01020125113427639, + 0.01712370291352272, + 0.014808137901127338, + 0.01459979172796011, + 0.13969282805919647, + 0.0179336778819561, + 0.04662003740668297, + 0.006317444145679474, + 0.0068010734394192696, + 0.02913324348628521, + 0.008656534366309643, + 0.07369711995124817, + 0.016531281173229218, + 0.01274990476667881, + 0.013682677410542965, + 0.010355556383728981, + 0.009300771169364452, + 0.008747115731239319, + 0.009233622811734676, + 0.008170368149876595, + 0.002183903008699417, + 0.001890311250463128, + 0.0016055505257099867, + 0.001037960173562169, + 0.0011896378127858043, + 0.001187445828691125, + 0.0011898467782884836, + 0.0010579341324046254, + 0.0034493415150791407, + 0.002777255140244961, + 0.0041344682686030865, + 0.003751882119104266, + 0.001900858711451292, + 0.0038283015601336956, + 0.002127492567524314, + 0.00219176709651947, + 0.013461507856845856, + 0.011673429049551487, + 0.008836658671498299, + 0.01441928744316101, + 0.017300916835665703, + 0.016113916411995888, + 0.015097368508577347, + 0.013068324886262417, + 0.016911745071411133, + 0.014102481305599213, + 0.012926477007567883, + 0.018343055620789528, + 0.009117210283875465, + 0.014761237427592278, + 0.012622465379536152, + 0.013746768236160278, + 0.009204871021211147, + 0.007704111281782389, + 0.02726009674370289, + 0.0035570727195590734, + 0.004262048285454512, + 0.022387463599443436, + 0.00295159756205976, + 0.0038714406546205282, + 0.0020321032498031855, + 0.002442448865622282, + 0.0028721706476062536, + 0.0018972345860674977, + 0.0019243983551859856, + 0.0030520290601998568, + 0.0021412645000964403, + 0.002885098336264491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.07901675999164581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486575782299042 + ], + "opposite_logprob_tokens": [ + -1.2236576080322266 + ], + "logprob_reward": -0.3486575782299042, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486575782299042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1138448715209961, + 0.08701547980308533, + 0.36401471495628357, + 0.06356163322925568, + 0.0372127965092659, + 0.0318094901740551, + 0.06884416937828064, + 0.022326355800032616, + 0.1071457639336586, + 0.10422469675540924 + ], + "credit_map": [ + 0.016004253178834915, + 0.009461096487939358, + 0.01265143696218729, + 0.012120503932237625, + 0.012647220864892006, + 0.015275350771844387, + 0.011693423613905907, + 0.02399159036576748, + 0.009827873669564724, + 0.010307212360203266, + 0.009698214940726757, + 0.014150300063192844, + 0.008467058651149273, + 0.012629653327167034, + 0.012994916178286076, + 0.008940250612795353, + 0.15105192363262177, + 0.020251231268048286, + 0.05089179053902626, + 0.008338837884366512, + 0.009612346068024635, + 0.03316204622387886, + 0.0106180589646101, + 0.08008851110935211, + 0.01000067126005888, + 0.0067137423902750015, + 0.010694107972085476, + 0.007801447995007038, + 0.005624501034617424, + 0.0075701260939240456, + 0.009198869578540325, + 0.005958171095699072, + 0.005714484490454197, + 0.0037439297884702682, + 0.004061727784574032, + 0.0038978324737399817, + 0.004457804840058088, + 0.004876737017184496, + 0.005922520533204079, + 0.004537760280072689, + 0.003920877352356911, + 0.0043137893080711365, + 0.00372372311539948, + 0.0040075210854411125, + 0.003548447974026203, + 0.004166937433183193, + 0.0037353886291384697, + 0.004392808303236961, + 0.00731594068929553, + 0.012091386131942272, + 0.006717131473124027, + 0.007656998001039028, + 0.006350789684802294, + 0.01179001946002245, + 0.009074586443603039, + 0.007847324013710022, + 0.004540255758911371, + 0.0014225187478587031, + 0.004418416414409876, + 0.001917115063406527, + 0.002026108792051673, + 0.002831224352121353, + 0.002777732675895095, + 0.0023929844610393047, + 0.015555675141513348, + 0.014390428550541401, + 0.02724521793425083, + 0.006090336479246616, + 0.007202640641480684, + 0.0253166314214468, + 0.005073231179267168, + 0.006271602585911751, + 0.009692519903182983, + 0.011442859657108784, + 0.01639198325574398, + 0.010352659039199352, + 0.010803606361150742, + 0.012689817696809769, + 0.016574332490563393, + 0.016276925802230835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.3486575782299042 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "superb macro lens, shot using high-speed photography, captures a moment of destructive beauty in extreme slow motion, while the camera performs a slow, smooth 360-degree orbital rotation around the central point of the shattering event. A flawless, crystalline crystal ball, its smooth surface reflecting a sharply defined spotlight from above, silently dents inward upon contact with a dark, highly polished, reflective obsidian floor. This explosion is not a chaotic spattering, but rather a silent blossoming of glass, with thousands of shimmering, sharp-edged fragments of varying sizes floating, tumbling, and spreading outward with grace and ease. Each fragment, like a single diamond, illuminated by a single, intense point light source, swirls and refracts a dazzling, iridescent spectrum, trailing fleeting, luminous trails against the pure black background, creating a surreal atmosphere that freezes the violent moment into a tranquil, radiant, and captivating visual spectacle.", + "seed": 7981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_0/index_3437_seed_7981_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14735350059345365, + "vlm_logprob": -0.14735350059345365, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.14735350059345365, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5289717316627502 + ], + "opposite_logprob_tokens": [ + -0.9039717316627502 + ], + "logprob_reward": -0.5289717316627502, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5289717316627502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1618008315563202, + 0.3096139132976532, + 0.11454475671052933, + 0.03131355717778206, + 0.04132029041647911, + 0.07248982787132263, + 0.10673917084932327, + 0.05008191242814064, + 0.02803424373269081, + 0.08406145870685577 + ], + "credit_map": [ + 0.02046203799545765, + 0.015467757359147072, + 0.021035868674516678, + 0.021559709683060646, + 0.024132216349244118, + 0.018202131614089012, + 0.020598875358700752, + 0.02034224383533001, + 0.04395623877644539, + 0.029488857835531235, + 0.05031387507915497, + 0.030404837802052498, + 0.025473980233073235, + 0.027555324137210846, + 0.05977358669042587, + 0.04264719784259796, + 0.027957675978541374, + 0.0047380151227116585, + 0.030851157382130623, + 0.008385768160223961, + 0.007067241705954075, + 0.018234150484204292, + 0.008457744494080544, + 0.008853004314005375, + 0.0036915824748575687, + 0.0027589858509600163, + 0.004943264182657003, + 0.004087942652404308, + 0.0032262883614748716, + 0.0035509022418409586, + 0.005381390918046236, + 0.0036731965374201536, + 0.004573151934891939, + 0.0054489439353346825, + 0.006457883398979902, + 0.005527293775230646, + 0.00605348963290453, + 0.003703358815982938, + 0.004220052156597376, + 0.005336117930710316, + 0.007816563360393047, + 0.008209972642362118, + 0.010873355902731419, + 0.009345382452011108, + 0.009380199015140533, + 0.008758042007684708, + 0.00876281876116991, + 0.009343501180410385, + 0.010465795174241066, + 0.012400216422975063, + 0.027999522164463997, + 0.011942736804485321, + 0.009722759015858173, + 0.012346665374934673, + 0.011990467086434364, + 0.009871005080640316, + 0.008563009090721607, + 0.005109082441776991, + 0.00714195566251874, + 0.004745268262922764, + 0.005347380414605141, + 0.00591494794934988, + 0.006068500690162182, + 0.00719177071005106, + 0.004825115669518709, + 0.004513758700340986, + 0.006957515608519316, + 0.00206441106274724, + 0.001671710517257452, + 0.0027191198896616697, + 0.0031100986525416374, + 0.0021725131664425135, + 0.015500896610319614, + 0.01182712521404028, + 0.017625460401177406, + 0.00813685730099678, + 0.006873895414173603, + 0.00810168031603098, + 0.009721705690026283, + 0.006273835897445679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.5289717316627502 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020633572712540627 + ], + "opposite_logprob_tokens": [ + -3.8956334590911865 + ], + "logprob_reward": -0.020633572712540627, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020633572712540627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24333453178405762, + 0.2662689685821533, + 0.09949421137571335, + 0.03337050974369049, + 0.04770516976714134, + 0.09056950360536575, + 0.03248947486281395, + 0.04966744780540466, + 0.060311365872621536, + 0.07678885757923126 + ], + "credit_map": [ + 0.03598704934120178, + 0.03174050524830818, + 0.03371662274003029, + 0.025568461045622826, + 0.02961956150829792, + 0.030241288244724274, + 0.028104273602366447, + 0.02835673652589321, + 0.03962951898574829, + 0.024501759558916092, + 0.04708067327737808, + 0.028567880392074585, + 0.025225622579455376, + 0.021837709471583366, + 0.04333170875906944, + 0.03609407693147659, + 0.020238913595676422, + 0.010112016461789608, + 0.02003926783800125, + 0.0077875470742583275, + 0.011243256740272045, + 0.013466757722198963, + 0.0071710762567818165, + 0.009435364976525307, + 0.00351839791983366, + 0.0026424722746014595, + 0.004946947563439608, + 0.004640615079551935, + 0.002502571325749159, + 0.003949016332626343, + 0.007362069096416235, + 0.0038084189873188734, + 0.0050542172975838184, + 0.007470899261534214, + 0.009055404923856258, + 0.0059071979485452175, + 0.003882987191900611, + 0.005493021570146084, + 0.004487216006964445, + 0.006354218814522028, + 0.009129876270890236, + 0.009789526462554932, + 0.014188306406140327, + 0.011098522692918777, + 0.013262533582746983, + 0.009350135922431946, + 0.010590228252112865, + 0.013160365633666515, + 0.0034361323341727257, + 0.0036539665888994932, + 0.00807393342256546, + 0.0035515467170625925, + 0.0035475636832416058, + 0.004064721055328846, + 0.0032370032276958227, + 0.002924602944403887, + 0.0076444619335234165, + 0.006810114253312349, + 0.012682849541306496, + 0.004141686949878931, + 0.0039294990710914135, + 0.004004669841378927, + 0.00517741683870554, + 0.00527674425393343, + 0.009078015573322773, + 0.006701882462948561, + 0.014095645397901535, + 0.005476030986756086, + 0.0036717175971716642, + 0.007748512551188469, + 0.007320116274058819, + 0.006219436880201101, + 0.013578435406088829, + 0.00893503986299038, + 0.01936272159218788, + 0.008418044075369835, + 0.005749975331127644, + 0.00687782745808363, + 0.007980241440236568, + 0.005886564496904612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.020633572712540627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032021934166550636 + ], + "opposite_logprob_tokens": [ + -5.753201961517334 + ], + "logprob_reward": -0.0032021934166550636, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032021934166550636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.155457004904747, + 0.10248488187789917, + 0.18206647038459778, + 0.04359598085284233, + 0.06872431188821793, + 0.10955730080604553, + 0.037489429116249084, + 0.050001248717308044, + 0.036793917417526245, + 0.21382944285869598 + ], + "credit_map": [ + 0.021804412826895714, + 0.018711544573307037, + 0.019696945324540138, + 0.01667623594403267, + 0.022225217893719673, + 0.01364865805953741, + 0.022605370730161667, + 0.020088614895939827, + 0.018733346834778786, + 0.006526994053274393, + 0.02378002554178238, + 0.007796373218297958, + 0.011159180663526058, + 0.008076063357293606, + 0.015414582565426826, + 0.010998307727277279, + 0.06518325209617615, + 0.00649303337559104, + 0.0696076974272728, + 0.00895104929804802, + 0.008115729317069054, + 0.008283818140625954, + 0.00810960866510868, + 0.007322295103222132, + 0.006368494126945734, + 0.005239447578787804, + 0.007121043745428324, + 0.004943547770380974, + 0.00357793178409338, + 0.006081458646804094, + 0.005246010143309832, + 0.005018042866140604, + 0.007781248539686203, + 0.00933900661766529, + 0.010094523429870605, + 0.008587496355175972, + 0.008987322449684143, + 0.006587073672562838, + 0.006971872411668301, + 0.010375766083598137, + 0.011697431094944477, + 0.011825460940599442, + 0.01804809644818306, + 0.01433502696454525, + 0.011852089315652847, + 0.01167948916554451, + 0.014666954055428505, + 0.015452759340405464, + 0.004084298852831125, + 0.004400205798447132, + 0.004763452801853418, + 0.004728781990706921, + 0.005167246330529451, + 0.0033140501473098993, + 0.007265783846378326, + 0.0037656049244105816, + 0.0073613012209534645, + 0.006073088850826025, + 0.01586206629872322, + 0.0042473990470170975, + 0.004087388515472412, + 0.004317800980061293, + 0.004308873787522316, + 0.0037433290854096413, + 0.005631426814943552, + 0.005286841653287411, + 0.006558489054441452, + 0.0038545846473425627, + 0.0024976753629744053, + 0.0031482756603509188, + 0.005877336487174034, + 0.0039392877370119095, + 0.02322555147111416, + 0.018597513437271118, + 0.04927363619208336, + 0.013793155550956726, + 0.019669312983751297, + 0.014601875096559525, + 0.0598672516644001, + 0.014801151119172573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.0032021934166550636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041007245890796185 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.0041007245890796185, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041007245890796185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15161341428756714, + 0.1440148949623108, + 0.1010136678814888, + 0.043388061225414276, + 0.03368786722421646, + 0.041127387434244156, + 0.03776051104068756, + 0.04423803091049194, + 0.03289685770869255, + 0.37025925517082214 + ], + "credit_map": [ + 0.021679433062672615, + 0.01905433088541031, + 0.01908022165298462, + 0.016254466027021408, + 0.023293359205126762, + 0.014471790753304958, + 0.020545192062854767, + 0.01723462902009487, + 0.024884728714823723, + 0.011044027283787727, + 0.03622759133577347, + 0.01255306787788868, + 0.01116176787763834, + 0.009261546656489372, + 0.018606062978506088, + 0.020276103168725967, + 0.018123792484402657, + 0.007742900867015123, + 0.020792189985513687, + 0.007918030023574829, + 0.009655470959842205, + 0.017841210588812828, + 0.009426696226000786, + 0.009513379074633121, + 0.006262874696403742, + 0.0033702498767524958, + 0.00810740701854229, + 0.005301734432578087, + 0.0034691309556365013, + 0.005790898576378822, + 0.006138541270047426, + 0.004947228357195854, + 0.004205649718642235, + 0.004717670381069183, + 0.0066776759922504425, + 0.003361702896654606, + 0.0028902089688926935, + 0.003777440171688795, + 0.0040657613426446915, + 0.003991757053881884, + 0.004448901396244764, + 0.0048556081019341946, + 0.0072869062423706055, + 0.005007900297641754, + 0.0062689390033483505, + 0.004540930036455393, + 0.004337969236075878, + 0.004380234517157078, + 0.003966078627854586, + 0.0044336081482470036, + 0.0042595942504704, + 0.005826106294989586, + 0.0057977051474153996, + 0.0041943807154893875, + 0.005561893805861473, + 0.0037211449816823006, + 0.00666588032618165, + 0.00543105648830533, + 0.01402628980576992, + 0.003799943020567298, + 0.0034085570368915796, + 0.004218542017042637, + 0.0030610687099397182, + 0.003626692108809948, + 0.00579549465328455, + 0.004923906642943621, + 0.006119093392044306, + 0.0024550785310566425, + 0.002231098711490631, + 0.004782684147357941, + 0.002958324970677495, + 0.0036311775911599398, + 0.03714125603437424, + 0.034554414451122284, + 0.08598961681127548, + 0.025912892073392868, + 0.03433056175708771, + 0.026897244155406952, + 0.10008420795202255, + 0.025349058210849762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0041007245890796185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a flawless, crystalline crystal ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crystal ball dent inward upon contact with a dark, highly polished, reflective obsidian floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thousands of sharp-edged glass fragments float, tumble, and spread outward in a silent blossoming motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments refract a dazzling, iridescent spectrum and trail luminous light trails against a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbital rotation around the central shattering event?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985928058624268 + ], + "opposite_logprob_tokens": [ + -1.8048592805862427 + ], + "logprob_reward": -0.17985928058624268, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985928058624268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11496727913618088, + 0.15698063373565674, + 0.0483713336288929, + 0.08350282907485962, + 0.09815357625484467, + 0.049390923231840134, + 0.053402215242385864, + 0.1584502011537552, + 0.1742406040430069, + 0.06254038214683533 + ], + "credit_map": [ + 0.013989897444844246, + 0.013321551494300365, + 0.01609380543231964, + 0.010460840538144112, + 0.018829913809895515, + 0.014360649511218071, + 0.013952585868537426, + 0.013958039693534374, + 0.025302931666374207, + 0.013946243561804295, + 0.023543281480669975, + 0.016719432547688484, + 0.014357196167111397, + 0.013674642890691757, + 0.031740497797727585, + 0.017696399241685867, + 0.0077780731953680515, + 0.0036847705487161875, + 0.006705856416374445, + 0.006106324959546328, + 0.004878590814769268, + 0.006259221583604813, + 0.005836036056280136, + 0.007122458424419165, + 0.010108232498168945, + 0.006641189102083445, + 0.012039883993566036, + 0.01157888863235712, + 0.008190077729523182, + 0.010087570175528526, + 0.013253643177449703, + 0.011603339575231075, + 0.010836739093065262, + 0.014167157001793385, + 0.017466912046074867, + 0.010236194357275963, + 0.00798540748655796, + 0.012864070944488049, + 0.010779964737594128, + 0.013817128725349903, + 0.005715737119317055, + 0.005770854186266661, + 0.007286713924258947, + 0.005681834649294615, + 0.005360993556678295, + 0.00658243615180254, + 0.005617258604615927, + 0.007375089917331934, + 0.0064514996483922005, + 0.005885136313736439, + 0.014754563570022583, + 0.005514245945960283, + 0.005844964180141687, + 0.005556296557188034, + 0.005364459473639727, + 0.004031049087643623, + 0.024166058748960495, + 0.01470898650586605, + 0.03456258773803711, + 0.016438262537121773, + 0.014188209548592567, + 0.020561834797263145, + 0.013517393730580807, + 0.020306862890720367, + 0.02831726334989071, + 0.018575593829154968, + 0.043304771184921265, + 0.012589548714458942, + 0.010113576427102089, + 0.02261369302868843, + 0.01900411583483219, + 0.019722040742635727, + 0.0069260187447071075, + 0.008024021983146667, + 0.015356994234025478, + 0.004483076743781567, + 0.006441804114729166, + 0.004031872842460871, + 0.013387570157647133, + 0.0038890226278454065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.17985928058624268 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_tensor([4725, 4725])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_tensor([4725, 4725])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..599977ae7b8e7db30e18e0218a755f60d5089ce5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_tensor([4725, 4725])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion arc shot from an upward angle closely follows the movements of the mythical figure Nezha. He appears as the classic demon child from the film \"Nezha: The Devil Child Comes into the World,\" sporting his signature twin buns, heavy smoky makeup, and a defiant expression. Clad in an exquisite red bellyband, he thrusts his flaming spear with all his might amidst a dark void filled with swirling, ink-style smoke. The spear's tip, blazing like molten lava, spins through the air at high speed, its tip precisely grazing a suspended pool of sticky cinnabar and a cloud of shimmering, icy ink, instantly enveloping them in a cacophony of colors. Then, like a giant brush, the flaming spear strikes with thunderous force across a vast, blank scroll unfolding from thin air. With each slashing and stabbing of the gun's tip, it carved the four characters \"我命著我\" (My destiny is my own) with a wild and powerful brushstroke. Its calligraphic style perfectly replicated the raw and commanding power of the title character from \"Ne Zha.\" Each stroke ignited countless golden sparks and embers, leaving trails of burning against the dark background. The entire scene, illuminated by the blazing gunfire and golden sparks, was imbued with the majestic momentum of defiance against fate and a powerful visual impact.", + "seed": 18322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_4725_seed_18322_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13696659225970506, + "vlm_logprob": -0.13696659225970506, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.13696659225970506, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010359208099544048 + ], + "opposite_logprob_tokens": [ + -4.760359287261963 + ], + "logprob_reward": -0.010359208099544048, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010359208099544048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15154092013835907, + 0.06822092086076736, + 0.14843672513961792, + 0.24473296105861664, + 0.05281088873744011, + 0.08236844092607498, + 0.053856104612350464, + 0.04597454518079758, + 0.10582011938095093, + 0.04623832553625107 + ], + "credit_map": [ + 0.024255314841866493, + 0.016564466059207916, + 0.019525805488228798, + 0.017983056604862213, + 0.0217933040112257, + 0.022876448929309845, + 0.01609673723578453, + 0.012445804663002491, + 0.004798124078661203, + 0.007297176867723465, + 0.00981504749506712, + 0.009973026812076569, + 0.012245481833815575, + 0.0079304538667202, + 0.01020406000316143, + 0.005957548972219229, + 0.011070306412875652, + 0.01348982285708189, + 0.02865622751414776, + 0.015080323442816734, + 0.027664368972182274, + 0.015724873170256615, + 0.016606640070676804, + 0.02014416828751564, + 0.016886882483959198, + 0.022324418649077415, + 0.03929218277335167, + 0.017311042174696922, + 0.06404661387205124, + 0.03244253993034363, + 0.03113948181271553, + 0.021289803087711334, + 0.006999361794441938, + 0.006297216285020113, + 0.006645468529313803, + 0.004269124008715153, + 0.013459134846925735, + 0.007343362085521221, + 0.005689450539648533, + 0.002107773208990693, + 0.009371348656713963, + 0.014238817617297173, + 0.013500742614269257, + 0.009936158545315266, + 0.010830205865204334, + 0.012502790428698063, + 0.005920550785958767, + 0.006067831069231033, + 0.008528066799044609, + 0.004805165808647871, + 0.005840881261974573, + 0.00814066268503666, + 0.009030661545693874, + 0.007801101077347994, + 0.005554026458412409, + 0.004155545029789209, + 0.006951178889721632, + 0.007925420068204403, + 0.0040604849345982075, + 0.005232946947216988, + 0.0047271219082176685, + 0.005323137156665325, + 0.005465083755552769, + 0.006289178505539894, + 0.012110484763979912, + 0.019526394084095955, + 0.010152732953429222, + 0.012836676090955734, + 0.023835785686969757, + 0.008433876559138298, + 0.009830540046095848, + 0.009093628264963627, + 0.005602592136710882, + 0.006513601168990135, + 0.003962568938732147, + 0.005054935347288847, + 0.010278654284775257, + 0.004319305997341871, + 0.004441431723535061, + 0.0060652391985058784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.010359208099544048 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026363611221313477 + ], + "opposite_logprob_tokens": [ + -3.6513636112213135 + ], + "logprob_reward": -0.026363611221313477, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026363611221313477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20322592556476593, + 0.14268244802951813, + 0.055999867618083954, + 0.13592985272407532, + 0.047651901841163635, + 0.10196365416049957, + 0.1499081254005432, + 0.031492654234170914, + 0.08730947226285934, + 0.04383612796664238 + ], + "credit_map": [ + 0.03502407297492027, + 0.02313247323036194, + 0.025329099968075752, + 0.024027056992053986, + 0.030080966651439667, + 0.029676664620637894, + 0.0207226425409317, + 0.015232939273118973, + 0.01263225357979536, + 0.010403218679130077, + 0.01981733739376068, + 0.03252851590514183, + 0.019833771511912346, + 0.01572062075138092, + 0.020082412287592888, + 0.0116643151268363, + 0.004540052730590105, + 0.005228720139712095, + 0.009311547502875328, + 0.00504707545042038, + 0.010931036435067654, + 0.0067754704505205154, + 0.007426688447594643, + 0.006739272736012936, + 0.007782342378050089, + 0.010877290740609169, + 0.025554118677973747, + 0.011421353556215763, + 0.029848838225007057, + 0.02848082222044468, + 0.01104933675378561, + 0.010915739461779594, + 0.004402452614158392, + 0.0041506607085466385, + 0.007333098910748959, + 0.0078787412494421, + 0.007857823744416237, + 0.006630913820117712, + 0.006384875625371933, + 0.0030133328400552273, + 0.010135412216186523, + 0.022328568622469902, + 0.010104178451001644, + 0.009914853610098362, + 0.018857140094041824, + 0.013995626010000706, + 0.008239693008363247, + 0.00838816910982132, + 0.014065098017454147, + 0.014615869149565697, + 0.019196728244423866, + 0.02641594596207142, + 0.030489135533571243, + 0.017370309680700302, + 0.01685125194489956, + 0.010903767310082912, + 0.003705760231241584, + 0.0058810776099562645, + 0.002551774028688669, + 0.003912117797881365, + 0.005461652297526598, + 0.0030526022892445326, + 0.0030595934949815273, + 0.003868074854835868, + 0.010402802377939224, + 0.016253357753157616, + 0.008562656119465828, + 0.00620223255828023, + 0.02014564350247383, + 0.0074296677485108376, + 0.008426014333963394, + 0.009887095540761948, + 0.005129334516823292, + 0.0055140117183327675, + 0.003508007386699319, + 0.00452899606898427, + 0.009973183274269104, + 0.005920615047216415, + 0.004702413454651833, + 0.004559563472867012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.026363611221313477 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329995296895504 + ], + "opposite_logprob_tokens": [ + -3.7732999324798584 + ], + "logprob_reward": -0.02329995296895504, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329995296895504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15714165568351746, + 0.056643880903720856, + 0.04919084161520004, + 0.10233272612094879, + 0.03654560446739197, + 0.32355043292045593, + 0.09300599992275238, + 0.043244317173957825, + 0.08756532520055771, + 0.05077921599149704 + ], + "credit_map": [ + 0.017552470788359642, + 0.017425084486603737, + 0.015197803266346455, + 0.02232586406171322, + 0.03496706113219261, + 0.019274456426501274, + 0.02147398702800274, + 0.008924932219088078, + 0.0069318003952503204, + 0.0067772893235087395, + 0.006484977435320616, + 0.008689808659255505, + 0.010042065754532814, + 0.007681813091039658, + 0.006099502090364695, + 0.003936626948416233, + 0.003781811101362109, + 0.005370097700506449, + 0.00789547897875309, + 0.005092487670481205, + 0.010188386775553226, + 0.00561130465939641, + 0.006435679737478495, + 0.004815594293177128, + 0.006967970170080662, + 0.005790893454104662, + 0.026967255398631096, + 0.006426907144486904, + 0.012099534273147583, + 0.030733825638890266, + 0.006982527673244476, + 0.006363811902701855, + 0.0026386105455458164, + 0.003031752072274685, + 0.007718084380030632, + 0.0061089592054486275, + 0.004905093461275101, + 0.005774617660790682, + 0.003989985212683678, + 0.0023785000666975975, + 0.0311314407736063, + 0.077247254550457, + 0.03218100965023041, + 0.0320177860558033, + 0.0536319874227047, + 0.04360773786902428, + 0.021876376122236252, + 0.031856853514909744, + 0.008444445207715034, + 0.00853328499943018, + 0.0113278953358531, + 0.019068485125899315, + 0.017116939648985863, + 0.010457276366651058, + 0.010205411352217197, + 0.007852254435420036, + 0.006802419200539589, + 0.007396613247692585, + 0.003558763535693288, + 0.006022480316460133, + 0.005743006244301796, + 0.004811587743461132, + 0.004570827353745699, + 0.004338620696216822, + 0.008028444834053516, + 0.01171890925616026, + 0.008892596699297428, + 0.01490421686321497, + 0.01935591734945774, + 0.007545358967036009, + 0.008209104649722576, + 0.008910778909921646, + 0.005110620055347681, + 0.006094136741012335, + 0.005036712624132633, + 0.0044337850995361805, + 0.009628724306821823, + 0.00967060960829258, + 0.004384737927466631, + 0.006419891957193613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.02329995296895504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865625500679016 + ], + "opposite_logprob_tokens": [ + -3.0486562252044678 + ], + "logprob_reward": -0.04865625500679016, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865625500679016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10874313861131668, + 0.03661637008190155, + 0.0539645217359066, + 0.06113563850522041, + 0.026638610288500786, + 0.05333343893289566, + 0.04782796651124954, + 0.06359528005123138, + 0.3261762857437134, + 0.22196877002716064 + ], + "credit_map": [ + 0.01712319813668728, + 0.011709709651768208, + 0.013879438862204552, + 0.011908939108252525, + 0.013518783263862133, + 0.018686076626181602, + 0.009970041923224926, + 0.011946949176490307, + 0.0031369486823678017, + 0.003735332516953349, + 0.004357635043561459, + 0.004840461537241936, + 0.007887212559580803, + 0.004148491192609072, + 0.006459930911660194, + 0.0020503592677414417, + 0.0034023490734398365, + 0.004861378576606512, + 0.01006957795470953, + 0.005114686209708452, + 0.013350991532206535, + 0.005472224671393633, + 0.0056806630454957485, + 0.006012652534991503, + 0.003535019000992179, + 0.004288495518267155, + 0.011631142348051071, + 0.005081277806311846, + 0.013034147210419178, + 0.014226531609892845, + 0.0053390320390462875, + 0.003999992739409208, + 0.004067424219101667, + 0.0019178857328370214, + 0.004744435660541058, + 0.002926061861217022, + 0.0043687946163117886, + 0.003958065062761307, + 0.0025097907055169344, + 0.002146152313798666, + 0.00988313090056181, + 0.0095393480733037, + 0.007104464806616306, + 0.005011372268199921, + 0.006037916522473097, + 0.008169261738657951, + 0.0037719213869422674, + 0.0038160202093422413, + 0.005379274487495422, + 0.00479406351223588, + 0.005013498477637768, + 0.006904096342623234, + 0.008290702477097511, + 0.005405832547694445, + 0.006930013652890921, + 0.005110483150929213, + 0.008812928572297096, + 0.012139206752181053, + 0.00525453919544816, + 0.008754119277000427, + 0.007611087057739496, + 0.0061790733598172665, + 0.007112671155482531, + 0.007731659337878227, + 0.030268991366028786, + 0.04158122092485428, + 0.03186133876442909, + 0.048104189336299896, + 0.07993952184915543, + 0.027662744745612144, + 0.03560424596071243, + 0.031154010444879532, + 0.02312169037759304, + 0.023878127336502075, + 0.018189387395977974, + 0.01815655641257763, + 0.041891783475875854, + 0.04662810266017914, + 0.02047191932797432, + 0.029631201177835464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.04865625500679016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761539340019226 + ], + "opposite_logprob_tokens": [ + -0.8261539340019226 + ], + "logprob_reward": -0.5761539340019226, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761539340019226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06108909472823143, + 0.06544310599565506, + 0.047631680965423584, + 0.21877503395080566, + 0.053170666098594666, + 0.13714194297790527, + 0.215934619307518, + 0.07988401502370834, + 0.07722888886928558, + 0.043700963258743286 + ], + "credit_map": [ + 0.009493235498666763, + 0.008603977039456367, + 0.0069436198100447655, + 0.007247066590934992, + 0.010518621653318405, + 0.006364104337990284, + 0.006350478623062372, + 0.005567988380789757, + 0.005692599341273308, + 0.005114932544529438, + 0.008830619044601917, + 0.016394298523664474, + 0.00825864914804697, + 0.006703501101583242, + 0.009583920240402222, + 0.004864581860601902, + 0.004997153300791979, + 0.0038598510436713696, + 0.007674369961023331, + 0.0062094987370073795, + 0.006496702786535025, + 0.006885682698339224, + 0.0063791838474571705, + 0.005129235330969095, + 0.011538136750459671, + 0.011953187175095081, + 0.05761704966425896, + 0.013681245967745781, + 0.02327975071966648, + 0.07079894095659256, + 0.016695110127329826, + 0.013211616314947605, + 0.005613476037979126, + 0.005182422231882811, + 0.011817592196166515, + 0.005572849418967962, + 0.00677170604467392, + 0.008011210709810257, + 0.005805447231978178, + 0.004395961295813322, + 0.009773564524948597, + 0.01572982594370842, + 0.01805230975151062, + 0.02091081812977791, + 0.028714219108223915, + 0.022450540214776993, + 0.01113528199493885, + 0.010375379584729671, + 0.04722847789525986, + 0.024196332320570946, + 0.02436898835003376, + 0.022410161793231964, + 0.030849792063236237, + 0.027438998222351074, + 0.020166996866464615, + 0.019274871796369553, + 0.009678907692432404, + 0.014193098992109299, + 0.007157632615417242, + 0.011291940696537495, + 0.010175922885537148, + 0.010426533408463001, + 0.007278124336153269, + 0.00968185719102621, + 0.011214989237487316, + 0.01416831649839878, + 0.00741442060098052, + 0.007968571968376637, + 0.013907858170568943, + 0.007555444724857807, + 0.007623737677931786, + 0.0073755537159740925, + 0.005265730898827314, + 0.005928000435233116, + 0.0050646900199353695, + 0.003330841427668929, + 0.009590265341103077, + 0.005109444726258516, + 0.004653370939195156, + 0.004758617375046015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.5761539340019226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion arc shot from an upward angle closely follows the movements of the mythical figure Nezha. He appears as the classic demon child from the film \"Nezha: The Devil Child Comes into the World,\" sporting his signature twin buns, heavy smoky makeup, and a defiant expression. Clad in an exquisite red bellyband, he thrusts his flaming spear with all his might amidst a dark void filled with swirling, ink-style smoke. The spear's tip, blazing like molten lava, spins through the air at high speed, its tip precisely grazing a suspended pool of sticky cinnabar and a cloud of shimmering, icy ink, instantly enveloping them in a cacophony of colors. Then, like a giant brush, the flaming spear strikes with thunderous force across a vast, blank scroll unfolding from thin air. With each slashing and stabbing of the gun's tip, it carved the four characters \"我命著我\" (My destiny is my own) with a wild and powerful brushstroke. Its calligraphic style perfectly replicated the raw and commanding power of the title character from \"Ne Zha.\" Each stroke ignited countless golden sparks and embers, leaving trails of burning against the dark background. The entire scene, illuminated by the blazing gunfire and golden sparks, was imbued with the majestic momentum of defiance against fate and a powerful visual impact.", + "seed": 18323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_4725_seed_18323_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16974389739334583, + "vlm_logprob": -0.16974389739334583, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.16974389739334583, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03083893284201622 + ], + "opposite_logprob_tokens": [ + -3.530838966369629 + ], + "logprob_reward": -0.03083893284201622, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03083893284201622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17336033284664154, + 0.10171908885240555, + 0.09239933639764786, + 0.041864074766635895, + 0.05750258266925812, + 0.1851157695055008, + 0.07898499816656113, + 0.11506160348653793, + 0.05407686159014702, + 0.09991534799337387 + ], + "credit_map": [ + 0.01858251541852951, + 0.02028859220445156, + 0.03582915663719177, + 0.01970546692609787, + 0.014543796889483929, + 0.022775618359446526, + 0.016025790944695473, + 0.02560938149690628, + 0.01519333478063345, + 0.010960760526359081, + 0.01716684363782406, + 0.010903285816311836, + 0.009371625259518623, + 0.01515272632241249, + 0.009039082564413548, + 0.013931430876255035, + 0.012569916434586048, + 0.006674318574368954, + 0.015672430396080017, + 0.008036880753934383, + 0.008413366042077541, + 0.018429730087518692, + 0.008344894275069237, + 0.01425779890269041, + 0.005319633986800909, + 0.005413183942437172, + 0.005739736836403608, + 0.005313507281243801, + 0.0025186503771692514, + 0.00816755648702383, + 0.0026246339548379183, + 0.0067671723663806915, + 0.0056079234927892685, + 0.004454747773706913, + 0.014702202752232552, + 0.008082670159637928, + 0.005501715000718832, + 0.006059303879737854, + 0.005721186753362417, + 0.007372832391411066, + 0.025788171216845512, + 0.0065905447117984295, + 0.05788911134004593, + 0.011138944886624813, + 0.009354003705084324, + 0.020141758024692535, + 0.007931957952678204, + 0.04628128558397293, + 0.014124473556876183, + 0.004795557353645563, + 0.020870184525847435, + 0.009604176506400108, + 0.008200444281101227, + 0.004984667524695396, + 0.003430441953241825, + 0.012975054793059826, + 0.019074583426117897, + 0.009410622529685497, + 0.014821719378232956, + 0.017288757488131523, + 0.01474631205201149, + 0.0122199272736907, + 0.013689440675079823, + 0.013810243457555771, + 0.003922491334378719, + 0.009469304233789444, + 0.005840129218995571, + 0.004694779869168997, + 0.01588681899011135, + 0.004288000054657459, + 0.004522690549492836, + 0.0054526496678590775, + 0.012055025435984135, + 0.016342325136065483, + 0.01248753722757101, + 0.010878450237214565, + 0.012919058091938496, + 0.006707021966576576, + 0.01041934359818697, + 0.01810658909380436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.03083893284201622 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211636036634445 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211636036634445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211636036634445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10522082448005676, + 0.18926677107810974, + 0.07563157379627228, + 0.03850312530994415, + 0.049169693142175674, + 0.07552734017372131, + 0.10817258805036545, + 0.24766826629638672, + 0.04254990071058273, + 0.06828988343477249 + ], + "credit_map": [ + 0.00873833429068327, + 0.009885641746222973, + 0.016614239662885666, + 0.010568832978606224, + 0.010717198252677917, + 0.015605545602738857, + 0.014965314418077469, + 0.01812572404742241, + 0.02413336932659149, + 0.019890690222382545, + 0.03333691135048866, + 0.02210645005106926, + 0.017205379903316498, + 0.029182514175772667, + 0.019993294030427933, + 0.023418189957737923, + 0.008971679024398327, + 0.0049787466414272785, + 0.01650811918079853, + 0.00702329259365797, + 0.008615843951702118, + 0.012715788558125496, + 0.0061274305917322636, + 0.010690678842365742, + 0.00450595747679472, + 0.004399633035063744, + 0.007137151900678873, + 0.004166828468441963, + 0.003968794364482164, + 0.006294683553278446, + 0.0030307003762573004, + 0.0049993786960840225, + 0.003881870536133647, + 0.0038618380203843117, + 0.008531336672604084, + 0.007770756725221872, + 0.005338466726243496, + 0.005023730453103781, + 0.007381003815680742, + 0.00738069461658597, + 0.008744362741708755, + 0.0031830682419240475, + 0.022456681355834007, + 0.0060340361669659615, + 0.005984535440802574, + 0.008637557737529278, + 0.004953891038894653, + 0.015533213503658772, + 0.016893191263079643, + 0.011040930636227131, + 0.019216326996684074, + 0.011930911801755428, + 0.009843312203884125, + 0.013139148242771626, + 0.007350739557296038, + 0.01875804364681244, + 0.03257554769515991, + 0.02752089872956276, + 0.029510699212551117, + 0.029252709820866585, + 0.06161454692482948, + 0.033442556858062744, + 0.015363424085080624, + 0.018387900665402412, + 0.0028781213331967592, + 0.007352203130722046, + 0.004204025026410818, + 0.004571089521050453, + 0.009324142709374428, + 0.0036397818475961685, + 0.003794066607952118, + 0.006786474492400885, + 0.010172943584620953, + 0.013234289363026619, + 0.011040921323001385, + 0.007937205955386162, + 0.008406400680541992, + 0.004006708972156048, + 0.005391906481236219, + 0.008099508471786976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.014211636036634445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607470214366913 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.01607470214366913, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607470214366913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09555087983608246, + 0.12121555209159851, + 0.09188730269670486, + 0.05433924123644829, + 0.07281522452831268, + 0.19183950126171112, + 0.06574463099241257, + 0.18751035630702972, + 0.05564962700009346, + 0.06344766169786453 + ], + "credit_map": [ + 0.009093087166547775, + 0.008302937261760235, + 0.020362695679068565, + 0.009343711659312248, + 0.007853426970541477, + 0.019267600029706955, + 0.008414464071393013, + 0.012912965379655361, + 0.013915044255554676, + 0.011439988389611244, + 0.017199741676449776, + 0.013107731007039547, + 0.010249221697449684, + 0.022330347448587418, + 0.017112312838435173, + 0.01586117222905159, + 0.011807375587522984, + 0.0063393572345376015, + 0.020693382248282433, + 0.009950862266123295, + 0.009389406070113182, + 0.012636329047381878, + 0.007705383934080601, + 0.013365210965275764, + 0.009975463151931763, + 0.006636031437665224, + 0.008294298313558102, + 0.0061702183447778225, + 0.0042709168046712875, + 0.0064603714272379875, + 0.004385154228657484, + 0.008146785199642181, + 0.008193721063435078, + 0.006822890602052212, + 0.015608441084623337, + 0.008600901812314987, + 0.006744930520653725, + 0.007936139591038227, + 0.007634190376847982, + 0.011274012736976147, + 0.022765208035707474, + 0.010913880541920662, + 0.055788639932870865, + 0.015625976026058197, + 0.015492933802306652, + 0.022739117965102196, + 0.009403804317116737, + 0.03910992667078972, + 0.011567833833396435, + 0.006077720783650875, + 0.013189841993153095, + 0.008996180258691311, + 0.0046846503391861916, + 0.005080144386738539, + 0.005051081068813801, + 0.011097179725766182, + 0.027850119397044182, + 0.016089139506220818, + 0.025923844426870346, + 0.025741174817085266, + 0.02873675711452961, + 0.023253243416547775, + 0.02031850628554821, + 0.01959756761789322, + 0.0035796642769128084, + 0.00821866374462843, + 0.0059190839529037476, + 0.006351697724312544, + 0.01108468696475029, + 0.004382701124995947, + 0.005970321595668793, + 0.010142807848751545, + 0.008300978690385818, + 0.013918116688728333, + 0.008470898494124413, + 0.009021184407174587, + 0.007290099747478962, + 0.004137855488806963, + 0.004602157976478338, + 0.007706371136009693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.01607470214366913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801445081830025 + ], + "opposite_logprob_tokens": [ + -3.529801368713379 + ], + "logprob_reward": -0.029801445081830025, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801445081830025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08254578709602356, + 0.04446939006447792, + 0.028007933869957924, + 0.02986091561615467, + 0.0279011782258749, + 0.053937338292598724, + 0.06337063014507294, + 0.32278791069984436, + 0.11009301245212555, + 0.23702587187290192 + ], + "credit_map": [ + 0.006906790193170309, + 0.0073388866148889065, + 0.018976161256432533, + 0.006325382273644209, + 0.007939575240015984, + 0.013749432750046253, + 0.008063515648245811, + 0.01324604731053114, + 0.004667484201490879, + 0.004063721280544996, + 0.006821576971560717, + 0.005018545314669609, + 0.003785329172387719, + 0.007711556740105152, + 0.005861405748873949, + 0.0065397717989981174, + 0.0041951825842261314, + 0.0015675394097343087, + 0.0067832088097929955, + 0.0023092792835086584, + 0.0028710816986858845, + 0.0036095387768000364, + 0.0028179138898849487, + 0.0038541904650628567, + 0.004495072644203901, + 0.003043180098757148, + 0.006700529251247644, + 0.0038454635068774223, + 0.0019926647655665874, + 0.0037047038786113262, + 0.002047482877969742, + 0.004031818825751543, + 0.0034954308066517115, + 0.0022003811318427324, + 0.005430299788713455, + 0.003903834382072091, + 0.0024878615513443947, + 0.003966611810028553, + 0.0027050247881561518, + 0.003711734199896455, + 0.007199245039373636, + 0.0056828767992556095, + 0.008591723628342152, + 0.007400535978376865, + 0.0044569154269993305, + 0.005803721025586128, + 0.006913064979016781, + 0.00788925215601921, + 0.009716685861349106, + 0.006030327174812555, + 0.009527282789349556, + 0.008462592959403992, + 0.006971840281039476, + 0.00594007782638073, + 0.00583985261619091, + 0.01088197436183691, + 0.036779019981622696, + 0.02986474707722664, + 0.04200883209705353, + 0.03555475175380707, + 0.06862371414899826, + 0.04885866120457649, + 0.030573274940252304, + 0.030524905771017075, + 0.009157056920230389, + 0.01833484135568142, + 0.01389996986836195, + 0.010887190699577332, + 0.03288495913147926, + 0.007456029765307903, + 0.010274633765220642, + 0.007198328152298927, + 0.03490316867828369, + 0.046697571873664856, + 0.03273801878094673, + 0.028132794424891472, + 0.023944374173879623, + 0.013531344011425972, + 0.01920321024954319, + 0.03787538409233093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.029801445081830025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577927708625793 + ], + "opposite_logprob_tokens": [ + -0.6327927708625793 + ], + "logprob_reward": -0.7577927708625793, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577927708625793, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13420405983924866, + 0.052956901490688324, + 0.08700855076313019, + 0.07588104903697968, + 0.06852643191814423, + 0.06778118759393692, + 0.10934319347143173, + 0.28954797983169556, + 0.042411766946315765, + 0.07233884930610657 + ], + "credit_map": [ + 0.016289716586470604, + 0.011561422608792782, + 0.022668300196528435, + 0.010421962477266788, + 0.01470818743109703, + 0.02846791222691536, + 0.01633951999247074, + 0.013747042044997215, + 0.006662080995738506, + 0.0036597149446606636, + 0.010333188809454441, + 0.005403222981840372, + 0.005056452006101608, + 0.008913370780646801, + 0.005505682900547981, + 0.007423189468681812, + 0.014007996767759323, + 0.008487526327371597, + 0.013416154310107231, + 0.00802549161016941, + 0.009859852492809296, + 0.013385393656790257, + 0.006832425016909838, + 0.012993715703487396, + 0.012497983872890472, + 0.008395553566515446, + 0.010284910909831524, + 0.008939895778894424, + 0.006790950894355774, + 0.011548266746103764, + 0.005355974193662405, + 0.012067509815096855, + 0.00962838251143694, + 0.007233366835862398, + 0.016023214906454086, + 0.00753728486597538, + 0.006149728782474995, + 0.007340140640735626, + 0.005326657090336084, + 0.009287659078836441, + 0.007084448356181383, + 0.006441439501941204, + 0.01451407466083765, + 0.00869771745055914, + 0.007977109402418137, + 0.0074878050945699215, + 0.0068271891213953495, + 0.008751403540372849, + 0.012933737598359585, + 0.009997857734560966, + 0.023044388741254807, + 0.01697087660431862, + 0.012519080191850662, + 0.010198385454714298, + 0.0076409378089010715, + 0.016037922352552414, + 0.03398875892162323, + 0.032381340861320496, + 0.037522826343774796, + 0.032600510865449905, + 0.06871835142374039, + 0.04074806347489357, + 0.02279731072485447, + 0.020790832117199898, + 0.004823631141334772, + 0.0054601943120360374, + 0.006152752786874771, + 0.006492218468338251, + 0.004621237050741911, + 0.003993477206677198, + 0.004148358013480902, + 0.006719901692122221, + 0.007620606571435928, + 0.010368406772613525, + 0.011091374792158604, + 0.010673414915800095, + 0.01053265854716301, + 0.004529120866209269, + 0.007235201541334391, + 0.01028806809335947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.7577927708625793 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow-motion arc shot from an upward angle closely follows the movements of the mythical figure Nezha. He appears as the classic demon child from the film \"Nezha: The Devil Child Comes into the World,\" sporting his signature twin buns, heavy smoky makeup, and a defiant expression. Clad in an exquisite red bellyband, he thrusts his flaming spear with all his might amidst a dark void filled with swirling, ink-style smoke. The spear's tip, blazing like molten lava, spins through the air at high speed, its tip precisely grazing a suspended pool of sticky cinnabar and a cloud of shimmering, icy ink, instantly enveloping them in a cacophony of colors. Then, like a giant brush, the flaming spear strikes with thunderous force across a vast, blank scroll unfolding from thin air. With each slashing and stabbing of the gun's tip, it carved the four characters \"我命著我\" (My destiny is my own) with a wild and powerful brushstroke. Its calligraphic style perfectly replicated the raw and commanding power of the title character from \"Ne Zha.\" Each stroke ignited countless golden sparks and embers, leaving trails of burning against the dark background. The entire scene, illuminated by the blazing gunfire and golden sparks, was imbued with the majestic momentum of defiance against fate and a powerful visual impact.", + "seed": 18324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_4725_seed_18324_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17623738050460816, + "vlm_logprob": -0.17623738050460816, + "vlm_yes_no_margin": 2.924999952316284, + "vlm_token_credit": -0.17623738050460816, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01911504939198494 + ], + "opposite_logprob_tokens": [ + -4.019114971160889 + ], + "logprob_reward": -0.01911504939198494, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01911504939198494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20898807048797607, + 0.40525034070014954, + 0.07544255256652832, + 0.071722611784935, + 0.02770339883863926, + 0.05351530760526657, + 0.04431294649839401, + 0.04727747291326523, + 0.027537601068615913, + 0.038249723613262177 + ], + "credit_map": [ + 0.019037391990423203, + 0.021214092150330544, + 0.04784631356596947, + 0.022212645038962364, + 0.019841184839606285, + 0.017827408388257027, + 0.026360949501395226, + 0.03464806079864502, + 0.05637340992689133, + 0.060550328344106674, + 0.05694160237908363, + 0.03381462022662163, + 0.03743232041597366, + 0.07552748173475266, + 0.04869433119893074, + 0.035916224122047424, + 0.007519606500864029, + 0.00857511069625616, + 0.014650085009634495, + 0.014321746304631233, + 0.00788759533315897, + 0.006327199749648571, + 0.010199230164289474, + 0.0059619783423841, + 0.007899579592049122, + 0.010232857428491116, + 0.010904456488788128, + 0.010159761644899845, + 0.008316376246511936, + 0.01025934237986803, + 0.006443527527153492, + 0.00750671187415719, + 0.005025909282267094, + 0.007159578148275614, + 0.004014180041849613, + 0.003360845847055316, + 0.0022037054877728224, + 0.0020509727764874697, + 0.0021850024349987507, + 0.0017032045871019363, + 0.008884770795702934, + 0.011052869260311127, + 0.006578553002327681, + 0.007374078966677189, + 0.004312450997531414, + 0.005537185352295637, + 0.004930988419800997, + 0.0048444075509905815, + 0.007864194922149181, + 0.005238363053649664, + 0.005564799997955561, + 0.006565005052834749, + 0.0037977765314280987, + 0.004870622418820858, + 0.00443640211597085, + 0.005975783336907625, + 0.007302075158804655, + 0.003654379863291979, + 0.005079502239823341, + 0.005113019607961178, + 0.0045794532634317875, + 0.007940889336168766, + 0.005748136900365353, + 0.007860014215111732, + 0.006027617491781712, + 0.0024569006636738777, + 0.002907475922256708, + 0.004010566510260105, + 0.002184997545555234, + 0.0024711780715733767, + 0.0026780522894114256, + 0.0048008086159825325, + 0.0050020781345665455, + 0.0023832498118281364, + 0.004697985481470823, + 0.0029763004276901484, + 0.0032918311189860106, + 0.007735017221421003, + 0.0038678674027323723, + 0.008295392617583275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01911504939198494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209194764494896 + ], + "opposite_logprob_tokens": [ + -4.018208980560303 + ], + "logprob_reward": -0.018209194764494896, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018209194764494896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2709147036075592, + 0.18395830690860748, + 0.23253467679023743, + 0.04559839516878128, + 0.036496978253126144, + 0.04726236313581467, + 0.023393450304865837, + 0.02480468340218067, + 0.0830797478556633, + 0.051956649869680405 + ], + "credit_map": [ + 0.023087957873940468, + 0.024263307452201843, + 0.0649004876613617, + 0.029192449524998665, + 0.02365725114941597, + 0.02377288043498993, + 0.029051678255200386, + 0.052988700568675995, + 0.01911271922290325, + 0.02139154076576233, + 0.027891647070646286, + 0.017688926309347153, + 0.027163969352841377, + 0.026263169944286346, + 0.02403067611157894, + 0.02041565626859665, + 0.027816403657197952, + 0.026010196655988693, + 0.04689928516745567, + 0.03516984358429909, + 0.01592564396560192, + 0.015885211527347565, + 0.03301036357879639, + 0.031817726790905, + 0.00503692589700222, + 0.007746393326669931, + 0.006739351432770491, + 0.007508841343224049, + 0.004453456960618496, + 0.006757726892828941, + 0.0038291134405881166, + 0.0035265847109258175, + 0.005152741447091103, + 0.006043550558388233, + 0.005974294152110815, + 0.006346080917865038, + 0.003136283252388239, + 0.0024586631916463375, + 0.004242611117660999, + 0.0031427538488060236, + 0.008548039942979813, + 0.009898755699396133, + 0.006921366788446903, + 0.0037463733460754156, + 0.003504778491333127, + 0.0050398483872413635, + 0.0036931138020008802, + 0.0059100859798491, + 0.0047506545670330524, + 0.004137578420341015, + 0.0018347289878875017, + 0.0033951089717447758, + 0.002009626477956772, + 0.0021703869570046663, + 0.0026059322990477085, + 0.002489434788003564, + 0.004020655527710915, + 0.0024847278837114573, + 0.0032783998176455498, + 0.00477564474567771, + 0.0021098393481224775, + 0.0034799452405422926, + 0.0020672674290835857, + 0.0025882034096866846, + 0.015162686817348003, + 0.00955461896955967, + 0.010142998769879341, + 0.013788142241537571, + 0.0075011830776929855, + 0.008831790648400784, + 0.007767945993691683, + 0.0103303799405694, + 0.0065976823680102825, + 0.003356285160407424, + 0.005782136227935553, + 0.0049967337399721146, + 0.0049705891869962215, + 0.01146022416651249, + 0.005447860807180405, + 0.00934513658285141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.018209194764494896 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607622765004635 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.01607622765004635, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607622765004635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27979373931884766, + 0.12727724015712738, + 0.08589892089366913, + 0.08451277762651443, + 0.1510169804096222, + 0.07054446637630463, + 0.03232257068157196, + 0.06954023241996765, + 0.05327720567584038, + 0.04581589251756668 + ], + "credit_map": [ + 0.021572023630142212, + 0.022318070754408836, + 0.07051656395196915, + 0.026622619479894638, + 0.02752409316599369, + 0.022790318354964256, + 0.03234757110476494, + 0.05610246956348419, + 0.011228340677917004, + 0.016025451943278313, + 0.019043343141674995, + 0.012612867169082165, + 0.021442266181111336, + 0.01526699960231781, + 0.013554326258599758, + 0.018103627488017082, + 0.010624917224049568, + 0.008007606491446495, + 0.015942860394716263, + 0.012049113400280476, + 0.00723158149048686, + 0.005001908168196678, + 0.01565735600888729, + 0.011383568868041039, + 0.008100751787424088, + 0.01714283972978592, + 0.012757533229887486, + 0.01259648147970438, + 0.00867601577192545, + 0.012224539183080196, + 0.006916018668562174, + 0.006098591256886721, + 0.024377580732107162, + 0.037990767508745193, + 0.025972452014684677, + 0.019021540880203247, + 0.01208785455673933, + 0.011229880154132843, + 0.009893570095300674, + 0.010443336330354214, + 0.01303743477910757, + 0.01621883548796177, + 0.00895578321069479, + 0.007241175509989262, + 0.0061454675160348415, + 0.006567252334207296, + 0.006307398900389671, + 0.006071116309612989, + 0.006243414245545864, + 0.005030976142734289, + 0.003838584292680025, + 0.003799914615228772, + 0.002456306479871273, + 0.003480809275060892, + 0.0026928081642836332, + 0.004779755603522062, + 0.01205410622060299, + 0.005147225689142942, + 0.008523060008883476, + 0.007949386723339558, + 0.0070576295256614685, + 0.013098450377583504, + 0.006605240982025862, + 0.00910513661801815, + 0.011266278102993965, + 0.005350627936422825, + 0.005856256000697613, + 0.008035531267523766, + 0.004149934276938438, + 0.0058335852809250355, + 0.00513966241851449, + 0.007645327132195234, + 0.006787226535379887, + 0.002464447868987918, + 0.005061541683971882, + 0.004202760756015778, + 0.00397392176091671, + 0.008211780339479446, + 0.004893692675977945, + 0.010220520198345184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01607622765004635 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001470029354095 + ], + "opposite_logprob_tokens": [ + -2.695014715194702 + ], + "logprob_reward": -0.07001470029354095, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001470029354095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10463577508926392, + 0.10436057299375534, + 0.11863397806882858, + 0.06276729702949524, + 0.10500211268663406, + 0.1602054089307785, + 0.08971057087182999, + 0.06061718240380287, + 0.10473266243934631, + 0.08933442831039429 + ], + "credit_map": [ + 0.009857804514467716, + 0.0086134672164917, + 0.021954987198114395, + 0.012602072209119797, + 0.012223146855831146, + 0.010102447122335434, + 0.012009129859507084, + 0.017272716388106346, + 0.011497492901980877, + 0.012817637994885445, + 0.014648323878645897, + 0.009147054515779018, + 0.011891677975654602, + 0.013412552885711193, + 0.015111475251615047, + 0.01583435945212841, + 0.01146879605948925, + 0.01591501198709011, + 0.014609512872993946, + 0.014005959965288639, + 0.008627861738204956, + 0.009336907416582108, + 0.02257745899260044, + 0.022092465311288834, + 0.00758240045979619, + 0.008135063573718071, + 0.008166950196027756, + 0.008332157507538795, + 0.00720211211591959, + 0.00903644785284996, + 0.007336732465773821, + 0.006975435186177492, + 0.015866553410887718, + 0.021748410537838936, + 0.016240010038018227, + 0.0202105063945055, + 0.009730551391839981, + 0.006423041690140963, + 0.00807307381182909, + 0.006709971930831671, + 0.025221843272447586, + 0.03030518628656864, + 0.0148922773078084, + 0.023614978417754173, + 0.016391875222325325, + 0.015504450537264347, + 0.019994881004095078, + 0.014279906637966633, + 0.016150429844856262, + 0.011676250025629997, + 0.01193141471594572, + 0.014428027905523777, + 0.008774025365710258, + 0.008306245319545269, + 0.007062133401632309, + 0.011382052674889565, + 0.009639056399464607, + 0.005701185204088688, + 0.0054078674875199795, + 0.0076206279918551445, + 0.006832093466073275, + 0.01098956260830164, + 0.005464519374072552, + 0.008962269872426987, + 0.022068055346608162, + 0.008116014301776886, + 0.01667817309498787, + 0.01156600657850504, + 0.009189310483634472, + 0.010880870744585991, + 0.010481752455234528, + 0.015752486884593964, + 0.008311492390930653, + 0.012218190357089043, + 0.010087997652590275, + 0.0162328090518713, + 0.013898122124373913, + 0.008190342225134373, + 0.011040082201361656, + 0.009355394169688225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.07001470029354095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577717304229736 + ], + "opposite_logprob_tokens": [ + -0.6327717304229736 + ], + "logprob_reward": -0.7577717304229736, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577717304229736, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4930129945278168, + 0.04686737805604935, + 0.12948961555957794, + 0.07270974665880203, + 0.07237603515386581, + 0.04381202161312103, + 0.042066968977451324, + 0.02189714089035988, + 0.05338628590106964, + 0.02438180148601532 + ], + "credit_map": [ + 0.04248816892504692, + 0.04135211557149887, + 0.12973085045814514, + 0.03885217756032944, + 0.04153750464320183, + 0.031707338988780975, + 0.05753714591264725, + 0.10980769991874695, + 0.0051223901100456715, + 0.005830622278153896, + 0.006341838743537664, + 0.004009863827377558, + 0.006446768529713154, + 0.0071725728921592236, + 0.007245112210512161, + 0.004698209464550018, + 0.010702687315642834, + 0.01773259975016117, + 0.021828539669513702, + 0.02528267912566662, + 0.009738866239786148, + 0.00989497173577547, + 0.02126491442322731, + 0.013044355437159538, + 0.0073760743252933025, + 0.012354379519820213, + 0.010465963743627071, + 0.011421308852732182, + 0.008406955748796463, + 0.010165238752961159, + 0.00588636938482523, + 0.00663345493376255, + 0.011443166993558407, + 0.017062019556760788, + 0.011347266845405102, + 0.009909547865390778, + 0.005683535244315863, + 0.005338017828762531, + 0.006622600834816694, + 0.004969888832420111, + 0.006961761973798275, + 0.008549332618713379, + 0.0035688397474586964, + 0.007921766489744186, + 0.0039058662950992584, + 0.004218799527734518, + 0.004414061550050974, + 0.004271592479199171, + 0.00906985905021429, + 0.008183111436665058, + 0.004373278468847275, + 0.00513442512601614, + 0.0031334704253822565, + 0.003153371624648571, + 0.004529711324721575, + 0.004489741288125515, + 0.003185064299032092, + 0.002943464322015643, + 0.0025284842122346163, + 0.0036015158984810114, + 0.002315964549779892, + 0.00255414517596364, + 0.002043029759079218, + 0.0027254733722656965, + 0.008880994282662868, + 0.005849099718034267, + 0.007639627903699875, + 0.00900894496589899, + 0.0050748856738209724, + 0.006596687715500593, + 0.005103473551571369, + 0.005232575349509716, + 0.003203293075785041, + 0.0017911246977746487, + 0.0032850648276507854, + 0.002588051138445735, + 0.003179448889568448, + 0.0038494779728353024, + 0.002655699383467436, + 0.0038296403363347054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.7577717304229736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow-motion arc shot from an upward angle closely follows the movements of the mythical figure Nezha. He appears as the classic demon child from the film \"Nezha: The Devil Child Comes into the World,\" sporting his signature twin buns, heavy smoky makeup, and a defiant expression. Clad in an exquisite red bellyband, he thrusts his flaming spear with all his might amidst a dark void filled with swirling, ink-style smoke. The spear's tip, blazing like molten lava, spins through the air at high speed, its tip precisely grazing a suspended pool of sticky cinnabar and a cloud of shimmering, icy ink, instantly enveloping them in a cacophony of colors. Then, like a giant brush, the flaming spear strikes with thunderous force across a vast, blank scroll unfolding from thin air. With each slashing and stabbing of the gun's tip, it carved the four characters \"我命著我\" (My destiny is my own) with a wild and powerful brushstroke. Its calligraphic style perfectly replicated the raw and commanding power of the title character from \"Ne Zha.\" Each stroke ignited countless golden sparks and embers, leaving trails of burning against the dark background. The entire scene, illuminated by the blazing gunfire and golden sparks, was imbued with the majestic momentum of defiance against fate and a powerful visual impact.", + "seed": 18325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_1/index_4725_seed_18325_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1234972869977355, + "vlm_logprob": -0.1234972869977355, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.1234972869977355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013592193834483624 + ], + "opposite_logprob_tokens": [ + -4.38859224319458 + ], + "logprob_reward": -0.013592193834483624, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013592193834483624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21811820566654205, + 0.1461549699306488, + 0.08135797083377838, + 0.04930863901972771, + 0.05442965775728226, + 0.09531492739915848, + 0.13552239537239075, + 0.03585655987262726, + 0.10952349007129669, + 0.07441315799951553 + ], + "credit_map": [ + 0.02101864665746689, + 0.040940411388874054, + 0.031371936202049255, + 0.020938729867339134, + 0.023639407008886337, + 0.028083516284823418, + 0.01757115125656128, + 0.03455441817641258, + 0.022075248882174492, + 0.024249892681837082, + 0.017659183591604233, + 0.016306694597005844, + 0.01986200176179409, + 0.01265859417617321, + 0.010694015771150589, + 0.02264932356774807, + 0.015909766778349876, + 0.005809460766613483, + 0.025147462263703346, + 0.005370947066694498, + 0.008397823199629784, + 0.00854584388434887, + 0.004892800003290176, + 0.007283864077180624, + 0.007257798686623573, + 0.005859424825757742, + 0.008730730973184109, + 0.005137184634804726, + 0.0061291796155273914, + 0.006085830740630627, + 0.0045715258456766605, + 0.005536963697522879, + 0.00788856577128172, + 0.008216853253543377, + 0.006849158089607954, + 0.00647333636879921, + 0.006934321951121092, + 0.007181338965892792, + 0.004318708088248968, + 0.006567372940480709, + 0.024769986048340797, + 0.012116670608520508, + 0.01767071709036827, + 0.0075423805974423885, + 0.008202009834349155, + 0.010334985330700874, + 0.006465790327638388, + 0.008212385699152946, + 0.017263630405068398, + 0.022785048931837082, + 0.018942423164844513, + 0.012717352248728275, + 0.01831146702170372, + 0.023393139243125916, + 0.011491641402244568, + 0.010617697611451149, + 0.0029163293074816465, + 0.004086825996637344, + 0.0044909375719726086, + 0.004883259534835815, + 0.004279598593711853, + 0.007291927933692932, + 0.003809895133599639, + 0.004097787197679281, + 0.011602727696299553, + 0.010432105511426926, + 0.008735885843634605, + 0.011263587512075901, + 0.01144379936158657, + 0.02041364647448063, + 0.018719438463449478, + 0.016912298277020454, + 0.007401133421808481, + 0.005538091063499451, + 0.004587491508573294, + 0.007723099086433649, + 0.015477987006306648, + 0.009747005999088287, + 0.01388396043330431, + 0.010054389014840126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.013592193834483624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016089601442217827 + ], + "opposite_logprob_tokens": [ + -4.14108943939209 + ], + "logprob_reward": -0.016089601442217827, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016089601442217827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08347408473491669, + 0.22950926423072815, + 0.290147066116333, + 0.07283531129360199, + 0.049517158418893814, + 0.04458128288388252, + 0.0402713343501091, + 0.04185082018375397, + 0.06798956543207169, + 0.07982410490512848 + ], + "credit_map": [ + 0.009699009358882904, + 0.011707095429301262, + 0.01256903912872076, + 0.008429696783423424, + 0.009565258398652077, + 0.007892572320997715, + 0.010276003740727901, + 0.013335405848920345, + 0.027380743995308876, + 0.02781294286251068, + 0.02737557329237461, + 0.01676538959145546, + 0.040297407656908035, + 0.02536686137318611, + 0.027633508667349815, + 0.03687682002782822, + 0.057352498173713684, + 0.02008098177611828, + 0.10026281327009201, + 0.014443943277001381, + 0.015145535580813885, + 0.054527975618839264, + 0.01093252282589674, + 0.017400817945599556, + 0.010912993922829628, + 0.007317916490137577, + 0.012780682183802128, + 0.008015397004783154, + 0.007639546412974596, + 0.01170314196497202, + 0.006547044031322002, + 0.007918591611087322, + 0.007526967208832502, + 0.010137304663658142, + 0.0064396546222269535, + 0.0066319797188043594, + 0.003997156396508217, + 0.006286527495831251, + 0.003998396452516317, + 0.004499174654483795, + 0.007735093589872122, + 0.005793801974505186, + 0.006882598623633385, + 0.004556156694889069, + 0.004238940309733152, + 0.006275673396885395, + 0.003925889730453491, + 0.005173131823539734, + 0.005981429945677519, + 0.009877361357212067, + 0.005951369181275368, + 0.003743143053725362, + 0.0035286317579448223, + 0.005991656333208084, + 0.0021857914980500937, + 0.003011952852830291, + 0.003461203770712018, + 0.0029873556923121214, + 0.004550648853182793, + 0.005209502298384905, + 0.009499839507043362, + 0.0063281613402068615, + 0.005505607929080725, + 0.004308503121137619, + 0.004930268507450819, + 0.008101369254291058, + 0.00477216811850667, + 0.0056059458293020725, + 0.009194320999085903, + 0.01303868554532528, + 0.008236992172896862, + 0.01410981547087431, + 0.007597652729600668, + 0.006518186070024967, + 0.004654658492654562, + 0.008644647896289825, + 0.015312028117477894, + 0.01176775898784399, + 0.014907271601259708, + 0.010421904735267162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.016089601442217827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635827101767063 + ], + "opposite_logprob_tokens": [ + -3.6513583660125732 + ], + "logprob_reward": -0.02635827101767063, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635827101767063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12972278892993927, + 0.12889468669891357, + 0.10520617663860321, + 0.1575544774532318, + 0.06842929869890213, + 0.06699754297733307, + 0.05588215962052345, + 0.056607842445373535, + 0.15511661767959595, + 0.07558839023113251 + ], + "credit_map": [ + 0.016991019248962402, + 0.019841328263282776, + 0.019024791195988655, + 0.01262717042118311, + 0.014516345225274563, + 0.015557381324470043, + 0.012363414280116558, + 0.018801340833306313, + 0.014564170502126217, + 0.01663331314921379, + 0.01585424691438675, + 0.010517526417970657, + 0.020890213549137115, + 0.016236061230301857, + 0.014470473863184452, + 0.01972866989672184, + 0.014480950310826302, + 0.013294470496475697, + 0.014030838385224342, + 0.008241637609899044, + 0.015714891254901886, + 0.015146427787840366, + 0.010305747389793396, + 0.013991216197609901, + 0.02606014348566532, + 0.022425085306167603, + 0.025951022282242775, + 0.013902595266699791, + 0.020841097459197044, + 0.018167516216635704, + 0.014457722194492817, + 0.015749303624033928, + 0.010679543949663639, + 0.011565170250833035, + 0.010931997559964657, + 0.00867778342217207, + 0.007127038668841124, + 0.008012942969799042, + 0.005659456364810467, + 0.005775365512818098, + 0.01578747108578682, + 0.011529333889484406, + 0.010818209499120712, + 0.005946437828242779, + 0.00559631735086441, + 0.007532148156315088, + 0.004099821671843529, + 0.005687807220965624, + 0.00702713942155242, + 0.013460989110171795, + 0.0075170802883803844, + 0.0054170298390090466, + 0.006141471676528454, + 0.00715340580791235, + 0.0034285355359315872, + 0.0057365079410374165, + 0.004261491354554892, + 0.004900702275335789, + 0.005839921068400145, + 0.00811947975307703, + 0.008673097006976604, + 0.00873282365500927, + 0.008492441847920418, + 0.007587882224470377, + 0.010248353704810143, + 0.017130205407738686, + 0.013009079732000828, + 0.014649081975221634, + 0.020282668992877007, + 0.01949601247906685, + 0.03163018077611923, + 0.028671037405729294, + 0.006775012705475092, + 0.006235385779291391, + 0.004735797643661499, + 0.00735247228294611, + 0.016849277541041374, + 0.011184047907590866, + 0.012923011556267738, + 0.00953338760882616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.02635827101767063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810751438140869 + ], + "opposite_logprob_tokens": [ + -3.288107395172119 + ], + "logprob_reward": -0.03810751438140869, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810751438140869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06118951737880707, + 0.06317155063152313, + 0.08627857267856598, + 0.03446440026164055, + 0.03512454032897949, + 0.09975320845842361, + 0.08958690613508224, + 0.02815953828394413, + 0.1930619478225708, + 0.30920976400375366 + ], + "credit_map": [ + 0.0062218159437179565, + 0.007541260682046413, + 0.012457722797989845, + 0.006617439910769463, + 0.005068739410489798, + 0.007758839521557093, + 0.005054200999438763, + 0.010469503700733185, + 0.008295861072838306, + 0.006865403149276972, + 0.010170130990445614, + 0.0050113620236516, + 0.009885349310934544, + 0.005735850427299738, + 0.004793109837919474, + 0.012414486147463322, + 0.017924800515174866, + 0.0051221647299826145, + 0.030631043016910553, + 0.004570380784571171, + 0.004599594511091709, + 0.014338551089167595, + 0.0034165659453719854, + 0.005675479304045439, + 0.007416270207613707, + 0.0034718383103609085, + 0.007085881195962429, + 0.0031788484193384647, + 0.003965097479522228, + 0.0033763861283659935, + 0.0026527231093496084, + 0.0033173603005707264, + 0.0060207657516002655, + 0.006013613194227219, + 0.0060575432144105434, + 0.003916049376130104, + 0.00290628126822412, + 0.004450040869414806, + 0.0022833410184830427, + 0.0034769095946103334, + 0.012741782702505589, + 0.012796671129763126, + 0.014818049035966396, + 0.013915670104324818, + 0.011254485696554184, + 0.0137040289118886, + 0.009203937835991383, + 0.011318592354655266, + 0.010984682478010654, + 0.015208282507956028, + 0.012136764824390411, + 0.008432300761342049, + 0.01323643233627081, + 0.014971481636166573, + 0.007944874465465546, + 0.006672091782093048, + 0.0031451948452740908, + 0.00268875272013247, + 0.0038402865175157785, + 0.003976575564593077, + 0.003634963883087039, + 0.004374420735985041, + 0.0032104281708598137, + 0.003288920735940337, + 0.01112984586507082, + 0.01310334075242281, + 0.014870893210172653, + 0.02134285494685173, + 0.016113361343741417, + 0.032905980944633484, + 0.04368579387664795, + 0.03990990296006203, + 0.03170422092080116, + 0.023819997906684875, + 0.01991935260593891, + 0.027439096942543983, + 0.06511029601097107, + 0.04503853991627693, + 0.05493593588471413, + 0.0412423349916935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.03810751438140869 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha with twin buns, smoky makeup, and a defiant expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Nezha wearing an exquisite red bellyband and holding a flaming spear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spear tip blaze like molten lava and spin through the air at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spear carve the four characters '我命著我' on a vast, blank scroll unfolding from thin air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an upward angle following Nezha in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233388543128967 + ], + "opposite_logprob_tokens": [ + -0.8983388543128967 + ], + "logprob_reward": -0.5233388543128967, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233388543128967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13505010306835175, + 0.06474549323320389, + 0.06219089403748512, + 0.08692967146635056, + 0.09794678539037704, + 0.3900289535522461, + 0.08186376094818115, + 0.020868027582764626, + 0.04319056496024132, + 0.017185840755701065 + ], + "credit_map": [ + 0.022914890199899673, + 0.014754152856767178, + 0.022452693432569504, + 0.009786169044673443, + 0.01123797707259655, + 0.017077160999178886, + 0.011217974126338959, + 0.025609079748392105, + 0.007614729460328817, + 0.006614149548113346, + 0.010230392217636108, + 0.0044249906204640865, + 0.00719126220792532, + 0.009620372205972672, + 0.009136276319622993, + 0.009913312271237373, + 0.009462226182222366, + 0.005218758713454008, + 0.015926310792565346, + 0.004272786900401115, + 0.006970291491597891, + 0.010706763714551926, + 0.004178277682512999, + 0.005455467849969864, + 0.01791767030954361, + 0.010050192475318909, + 0.020246440544724464, + 0.006658971309661865, + 0.009307309985160828, + 0.009228670969605446, + 0.006316889077425003, + 0.0072035156190395355, + 0.014368219301104546, + 0.013200928457081318, + 0.014586959034204483, + 0.013953409157693386, + 0.009320597164332867, + 0.011996290646493435, + 0.009511884301900864, + 0.011008490808308125, + 0.051806677132844925, + 0.04873555526137352, + 0.058128245174884796, + 0.04744245857000351, + 0.044579144567251205, + 0.05846455693244934, + 0.03360193967819214, + 0.047270338982343674, + 0.008870924822986126, + 0.012486548162996769, + 0.010630253702402115, + 0.008652601391077042, + 0.012129219248890877, + 0.013846008107066154, + 0.007762525696307421, + 0.007485680282115936, + 0.0017058240482583642, + 0.0014506641309708357, + 0.0028503460343927145, + 0.002355048432946205, + 0.002949090674519539, + 0.003108345903456211, + 0.00281184702180326, + 0.0036368591245263815, + 0.0044103385880589485, + 0.007471146993339062, + 0.0034592878073453903, + 0.004409538581967354, + 0.006217926274985075, + 0.004703747108578682, + 0.006157705560326576, + 0.006360869389027357, + 0.0010025434894487262, + 0.0016202207189053297, + 0.0011576644610613585, + 0.0023158646654337645, + 0.004226480145007372, + 0.002064059255644679, + 0.002580908825621009, + 0.0022180986125022173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.5233388543128967 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_tensor([9154, 9154])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_tensor([9154, 9154])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8c6dd003738c2546f016fa270da435762b7e778b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_tensor([9154, 9154])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "handheld, slightly shaky, medium-long shot captures a bustling, sunny public square as a man dressed entirely in black approaches two young children. The camera tracks him with subtle, unsteady movements, enhancing the scene's raw immediacy. The man, dressed in a black leather jacket, a black T-shirt, and dark jeans, moves with an unconventional rhythm, twisting his torso while speaking enthusiastically and gesturing. His face is initially obscured by dark sunglasses. Two children, a boy and a girl, wearing brightly colored hoodies, stand close together, their eyes wide and uncertain as they watch his approach. As the man closes in distance, he deliberately slows down, smoothly removing his sunglasses to reveal a gentle smile. He then bends down from the waist, lowering his upper body to level with the children in a non-threatening gesture. In response, the children's previously motionless hands begin to move; one child fidgets with his sleeve, another makes a small, hesitant gesture forward. Throughout this interaction, the background remains vibrant: figures in the distance stroll through the square, and at one point, a passerby briskly walks through the foreground, momentarily interrupting the frame. The entire scene is bathed in bright, cool daylight, creating a stark contrast and a tense, unfolding atmosphere of curiosity, all captured with an erratic camera that suggests a spontaneous, unstaged moment.", + "seed": 8436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_9154_seed_8436_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0740912415087223, + "vlm_logprob": -0.0740912415087223, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.0740912415087223, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02495272271335125 + ], + "opposite_logprob_tokens": [ + -4.0249528884887695 + ], + "logprob_reward": -0.02495272271335125, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02495272271335125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22618649899959564, + 0.07088825106620789, + 0.08097191900014877, + 0.1552966982126236, + 0.05269405245780945, + 0.1689733862876892, + 0.06952621042728424, + 0.04616370424628258, + 0.10265666991472244, + 0.026642609387636185 + ], + "credit_map": [ + 0.016949599608778954, + 0.017501376569271088, + 0.042303360998630524, + 0.023156432434916496, + 0.05135834962129593, + 0.03291960060596466, + 0.025363832712173462, + 0.016633937135338783, + 0.006574738305062056, + 0.008835281245410442, + 0.009900275617837906, + 0.01087965164333582, + 0.007996347732841969, + 0.01107486616820097, + 0.006361243315041065, + 0.009265851229429245, + 0.009199848398566246, + 0.013307983055710793, + 0.009189959615468979, + 0.011684686876833439, + 0.009628684259951115, + 0.015297635458409786, + 0.006136944051831961, + 0.006526176817715168, + 0.019243530929088593, + 0.01880616322159767, + 0.023592213168740273, + 0.02663206309080124, + 0.016365734860301018, + 0.02029062993824482, + 0.012932638637721539, + 0.01743372716009617, + 0.006515762768685818, + 0.006100575439631939, + 0.0053438106551766396, + 0.008453122340142727, + 0.007958290167152882, + 0.00445562181994319, + 0.007054881192743778, + 0.006811989936977625, + 0.009447960183024406, + 0.011687838472425938, + 0.058187540620565414, + 0.013578148558735847, + 0.051239944994449615, + 0.013044532388448715, + 0.005936901550740004, + 0.005850526969879866, + 0.007829621434211731, + 0.009947952814400196, + 0.007912700064480305, + 0.009224249981343746, + 0.00866620521992445, + 0.01057513803243637, + 0.006294796708971262, + 0.009075545705854893, + 0.006863033398985863, + 0.009440377354621887, + 0.004127777647227049, + 0.005173943936824799, + 0.003680393798276782, + 0.006308951415121555, + 0.004765136633068323, + 0.005804089363664389, + 0.011434477753937244, + 0.009017711505293846, + 0.011916949413716793, + 0.014973020181059837, + 0.009651020169258118, + 0.017750483006238937, + 0.013504496775567532, + 0.014408509247004986, + 0.002086745109409094, + 0.0034598486963659525, + 0.0029179842676967382, + 0.003916490823030472, + 0.002504499163478613, + 0.0040330239571630955, + 0.0030821403488516808, + 0.004641876555979252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.02495272271335125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005273716524243355 + ], + "opposite_logprob_tokens": [ + -5.255273818969727 + ], + "logprob_reward": -0.005273716524243355, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005273716524243355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12764155864715576, + 0.1391316056251526, + 0.06089383363723755, + 0.087796151638031, + 0.05167517438530922, + 0.1387491226196289, + 0.046353261917829514, + 0.06615698337554932, + 0.14969313144683838, + 0.13190920650959015 + ], + "credit_map": [ + 0.013241149485111237, + 0.008912631310522556, + 0.0302608422935009, + 0.015640292316675186, + 0.01493641547858715, + 0.00989410001784563, + 0.024529069662094116, + 0.010227064602077007, + 0.015128583647310734, + 0.0195299219340086, + 0.02447604574263096, + 0.019683072343468666, + 0.017276691272854805, + 0.012593861669301987, + 0.015617221593856812, + 0.01482621394097805, + 0.00622972846031189, + 0.007727670483291149, + 0.006134511437267065, + 0.009484394453465939, + 0.007925537414848804, + 0.012284397147595882, + 0.006140503101050854, + 0.004967086482793093, + 0.010740487836301327, + 0.016482818871736526, + 0.012083550915122032, + 0.014012034982442856, + 0.007869593799114227, + 0.011256108991801739, + 0.007285895757377148, + 0.008065661415457726, + 0.006641429383307695, + 0.007360486313700676, + 0.006143832579255104, + 0.00703066261485219, + 0.007858089171350002, + 0.004113944713026285, + 0.005470144562423229, + 0.00705658970400691, + 0.008102146908640862, + 0.008052629418671131, + 0.04749884828925133, + 0.00926753506064415, + 0.04661310091614723, + 0.010247444733977318, + 0.004720243159681559, + 0.0042471690103411674, + 0.005971955601125956, + 0.006345695350319147, + 0.006272152066230774, + 0.006700059864670038, + 0.005256441421806812, + 0.006809721235185862, + 0.003380153328180313, + 0.005617084447294474, + 0.007285071536898613, + 0.009236620739102364, + 0.007444043178111315, + 0.0065524340607225895, + 0.008694261312484741, + 0.010461135767400265, + 0.007119724992662668, + 0.00936368852853775, + 0.012240178883075714, + 0.02035372145473957, + 0.014106033369898796, + 0.029534071683883667, + 0.017303775995969772, + 0.01868702471256256, + 0.020119773223996162, + 0.017348546534776688, + 0.013555832207202911, + 0.01824035309255123, + 0.018447035923600197, + 0.02102062478661537, + 0.017896346747875214, + 0.014048973098397255, + 0.014403086155653, + 0.01429696287959814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.005273716524243355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545006349682808 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545006349682808, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545006349682808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1238841786980629, + 0.11587948352098465, + 0.07519598305225372, + 0.08332452923059464, + 0.030764363706111908, + 0.2751592695713043, + 0.06294458359479904, + 0.07516954839229584, + 0.11208076030015945, + 0.04559728875756264 + ], + "credit_map": [ + 0.012353863567113876, + 0.010387025773525238, + 0.027408679947257042, + 0.009810059331357479, + 0.026861300691962242, + 0.012355154380202293, + 0.014462538994848728, + 0.010245545767247677, + 0.011377557180821896, + 0.014286108314990997, + 0.015079156495630741, + 0.015941623598337173, + 0.01771307736635208, + 0.01309168804436922, + 0.015353989787399769, + 0.013036280870437622, + 0.007327907253056765, + 0.009125086478888988, + 0.008443652652204037, + 0.011203846894204617, + 0.009281368926167488, + 0.015688402578234673, + 0.008198196068406105, + 0.005927523598074913, + 0.008638034574687481, + 0.012801517732441425, + 0.012681113556027412, + 0.016500607132911682, + 0.008602151647210121, + 0.008829831145703793, + 0.007433690596371889, + 0.00783758144825697, + 0.0032651021610945463, + 0.0028197893407195807, + 0.003216999350115657, + 0.004420979414135218, + 0.005476260092109442, + 0.00278915255330503, + 0.00439050467684865, + 0.004385577514767647, + 0.0156790129840374, + 0.01549800019711256, + 0.0919775739312172, + 0.016569161787629128, + 0.09634979814291, + 0.02219032123684883, + 0.009568777866661549, + 0.0073266285471618176, + 0.007668789476156235, + 0.009780014865100384, + 0.0070655918680131435, + 0.009358473122119904, + 0.00774426618590951, + 0.009153851307928562, + 0.004570742603391409, + 0.007602854166179895, + 0.008903959766030312, + 0.017428098246455193, + 0.006117553450167179, + 0.0092027448117733, + 0.007950793951749802, + 0.008088527247309685, + 0.00817850511521101, + 0.009299366734921932, + 0.012179624289274216, + 0.009613445959985256, + 0.013582143001258373, + 0.01771852932870388, + 0.010011866688728333, + 0.020016131922602654, + 0.013863323256373405, + 0.01509569026529789, + 0.004677935969084501, + 0.008370286785066128, + 0.005111542530357838, + 0.007188804447650909, + 0.005616276524960995, + 0.006071927957236767, + 0.003567645326256752, + 0.00499286875128746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.012545006349682808 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201881363987923 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014201881363987923, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201881363987923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12202008068561554, + 0.1078033596277237, + 0.03797701373696327, + 0.05044456571340561, + 0.031369127333164215, + 0.16007843613624573, + 0.027773600071668625, + 0.09696777164936066, + 0.11251708120107651, + 0.25304895639419556 + ], + "credit_map": [ + 0.012057364918291569, + 0.01143554411828518, + 0.017324797809123993, + 0.014168495312333107, + 0.011145314201712608, + 0.01045316644012928, + 0.03452714532613754, + 0.010908253490924835, + 0.014699633233249187, + 0.007749546319246292, + 0.022515343502163887, + 0.01806170493364334, + 0.009286629967391491, + 0.012015312910079956, + 0.011078454554080963, + 0.012396731413900852, + 0.0047058230265975, + 0.0036342120729386806, + 0.007788090966641903, + 0.00435481034219265, + 0.003671351121738553, + 0.005316450726240873, + 0.004834508057683706, + 0.003671768819913268, + 0.006138049531728029, + 0.00895677413791418, + 0.006493870634585619, + 0.007679603528231382, + 0.005358885508030653, + 0.006652678828686476, + 0.004241035785526037, + 0.004923662636429071, + 0.0033891338389366865, + 0.004304456524550915, + 0.0034431060776114464, + 0.00487903319299221, + 0.0045743826776742935, + 0.0028619898948818445, + 0.00429811654612422, + 0.003618910675868392, + 0.009431340731680393, + 0.012694484554231167, + 0.05166380852460861, + 0.010255596600472927, + 0.05195097625255585, + 0.013946044258773327, + 0.005570013076066971, + 0.004566176328808069, + 0.004060768056660891, + 0.004138491116464138, + 0.004393571522086859, + 0.003503187792375684, + 0.002989499131217599, + 0.003420422552153468, + 0.002666071057319641, + 0.0026015890762209892, + 0.012100397609174252, + 0.02103719301521778, + 0.009790215641260147, + 0.010662777349352837, + 0.008254088461399078, + 0.012305534444749355, + 0.009785748086869717, + 0.013031820766627789, + 0.01077519915997982, + 0.012318164110183716, + 0.012799220159649849, + 0.017442502081394196, + 0.011761495843529701, + 0.017419667914509773, + 0.013872455805540085, + 0.016128377988934517, + 0.023028919473290443, + 0.04357919469475746, + 0.030556436628103256, + 0.04679598659276962, + 0.03238535299897194, + 0.02452084980905056, + 0.02270398661494255, + 0.02947823517024517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.014201881363987923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134828805923462 + ], + "opposite_logprob_tokens": [ + -1.3134828805923462 + ], + "logprob_reward": -0.3134828805923462, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134828805923462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1049375981092453, + 0.06172511726617813, + 0.04765930771827698, + 0.06511929631233215, + 0.0809137225151062, + 0.1306915581226349, + 0.2638578414916992, + 0.10114365816116333, + 0.0836453065276146, + 0.06030654534697533 + ], + "credit_map": [ + 0.012367868795990944, + 0.007664734497666359, + 0.019998421892523766, + 0.009899203665554523, + 0.023848019540309906, + 0.010148601606488228, + 0.013656129129230976, + 0.007354625966399908, + 0.007036692462861538, + 0.007968497462570667, + 0.010622342117130756, + 0.011823867447674274, + 0.00441542686894536, + 0.0064495522528886795, + 0.007283320184797049, + 0.006125421263277531, + 0.005743100307881832, + 0.007099266164004803, + 0.006311626639217138, + 0.006754783447831869, + 0.0065927570685744286, + 0.004891492426395416, + 0.006270149257034063, + 0.0039961389265954494, + 0.007949849590659142, + 0.007054038345813751, + 0.012001296505331993, + 0.012747197411954403, + 0.006837541703134775, + 0.007271221373230219, + 0.0048784916289150715, + 0.00637966301292181, + 0.010403943248093128, + 0.015833690762519836, + 0.008820298127830029, + 0.0158119797706604, + 0.008232240565121174, + 0.006382796447724104, + 0.006503642071038485, + 0.008925133384764194, + 0.007187994662672281, + 0.008391250856220722, + 0.041947510093450546, + 0.011859566904604435, + 0.04029608145356178, + 0.009768262505531311, + 0.005917660892009735, + 0.005323235876858234, + 0.03525080159306526, + 0.03085491620004177, + 0.03700016438961029, + 0.02865634672343731, + 0.038725659251213074, + 0.04117114469408989, + 0.01945413276553154, + 0.032744698226451874, + 0.012218022719025612, + 0.0194784477353096, + 0.00975528173148632, + 0.01150444895029068, + 0.013546877540647984, + 0.013094037771224976, + 0.00990801490843296, + 0.01163853332400322, + 0.006717780139297247, + 0.013580653816461563, + 0.007307495456188917, + 0.015213980339467525, + 0.012340810149908066, + 0.008325958624482155, + 0.011426199227571487, + 0.008732430636882782, + 0.006303566507995129, + 0.009218636900186539, + 0.007871595211327076, + 0.008697379380464554, + 0.005263368133455515, + 0.008438956923782825, + 0.007069792132824659, + 0.007443254813551903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.3134828805923462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld, slightly shaky, medium-long shot captures a bustling, sunny public square as a man dressed entirely in black approaches two young children. The camera tracks him with subtle, unsteady movements, enhancing the scene's raw immediacy. The man, dressed in a black leather jacket, a black T-shirt, and dark jeans, moves with an unconventional rhythm, twisting his torso while speaking enthusiastically and gesturing. His face is initially obscured by dark sunglasses. Two children, a boy and a girl, wearing brightly colored hoodies, stand close together, their eyes wide and uncertain as they watch his approach. As the man closes in distance, he deliberately slows down, smoothly removing his sunglasses to reveal a gentle smile. He then bends down from the waist, lowering his upper body to level with the children in a non-threatening gesture. In response, the children's previously motionless hands begin to move; one child fidgets with his sleeve, another makes a small, hesitant gesture forward. Throughout this interaction, the background remains vibrant: figures in the distance stroll through the square, and at one point, a passerby briskly walks through the foreground, momentarily interrupting the frame. The entire scene is bathed in bright, cool daylight, creating a stark contrast and a tense, unfolding atmosphere of curiosity, all captured with an erratic camera that suggests a spontaneous, unstaged moment.", + "seed": 8437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_9154_seed_8437_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07133585698902607, + "vlm_logprob": -0.07133585698902607, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.07133585698902607, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09279204159975052 + ], + "opposite_logprob_tokens": [ + -2.59279203414917 + ], + "logprob_reward": -0.09279204159975052, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09279204159975052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1683221012353897, + 0.287140816450119, + 0.15313100814819336, + 0.07631830126047134, + 0.10834964364767075, + 0.03193352371454239, + 0.06191311031579971, + 0.018597831949591637, + 0.035486411303281784, + 0.05880722403526306 + ], + "credit_map": [ + 0.028434528037905693, + 0.028963202610611916, + 0.030688440427184105, + 0.026974167674779892, + 0.012325825169682503, + 0.01695938967168331, + 0.015038671903312206, + 0.008937880396842957, + 0.019935952499508858, + 0.016708988696336746, + 0.10705871134996414, + 0.11173059046268463, + 0.004693444352596998, + 0.01384287141263485, + 0.008461052551865578, + 0.004709214437752962, + 0.017560850828886032, + 0.013544945046305656, + 0.014068154618144035, + 0.023558711633086205, + 0.0525946244597435, + 0.009832100942730904, + 0.009847970679402351, + 0.012123634107410908, + 0.010405221953988075, + 0.009465821087360382, + 0.013329260982573032, + 0.007884387858211994, + 0.010921117849647999, + 0.007812556810677052, + 0.006907651666551828, + 0.009592287242412567, + 0.008894388563930988, + 0.009369912557303905, + 0.016384655609726906, + 0.0074643949046730995, + 0.026068607345223427, + 0.017191052436828613, + 0.009454591199755669, + 0.013522043824195862, + 0.0030665069352835417, + 0.00382041372358799, + 0.004579377360641956, + 0.0035224435850977898, + 0.004484110977500677, + 0.005219477694481611, + 0.0032411785796284676, + 0.004000011365860701, + 0.008918416686356068, + 0.004394675139337778, + 0.0044129518792033195, + 0.005609351210296154, + 0.023463819175958633, + 0.00921549927443266, + 0.00338451424613595, + 0.002513883402571082, + 0.002407585270702839, + 0.0019931390415877104, + 0.0029851116705685854, + 0.002050303854048252, + 0.0012761924881488085, + 0.0026096038054674864, + 0.0011612845119088888, + 0.004114610143005848, + 0.0033247482497245073, + 0.006984833162277937, + 0.006037611979991198, + 0.005388051271438599, + 0.003998192027211189, + 0.003914853557944298, + 0.0027686776593327522, + 0.003069441532716155, + 0.0065545812249183655, + 0.005753599107265472, + 0.0064673530869185925, + 0.006168739404529333, + 0.008282802999019623, + 0.014409828931093216, + 0.004341077525168657, + 0.006829242687672377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.09279204159975052 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011107752099633217 + ], + "opposite_logprob_tokens": [ + -4.511107921600342 + ], + "logprob_reward": -0.011107752099633217, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011107752099633217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12575304508209229, + 0.11056950688362122, + 0.11461858451366425, + 0.056947074830532074, + 0.054334163665771484, + 0.05767231062054634, + 0.1935291737318039, + 0.03658944368362427, + 0.17615054547786713, + 0.0738360807299614 + ], + "credit_map": [ + 0.012545992620289326, + 0.016476964578032494, + 0.025573739781975746, + 0.021547434851527214, + 0.014671056531369686, + 0.011861794628202915, + 0.01253456063568592, + 0.010541515424847603, + 0.024021221324801445, + 0.008999470621347427, + 0.02576444484293461, + 0.018684979528188705, + 0.010163717903196812, + 0.007398864720016718, + 0.00842820294201374, + 0.007108612451702356, + 0.013009734451770782, + 0.009980772621929646, + 0.019850384443998337, + 0.008729074150323868, + 0.0376751683652401, + 0.008264933712780476, + 0.00798313319683075, + 0.009125389158725739, + 0.007157531101256609, + 0.006552010774612427, + 0.008861223235726357, + 0.007370585110038519, + 0.014861089177429676, + 0.003987394273281097, + 0.004301759414374828, + 0.0038554896600544453, + 0.004444628953933716, + 0.007079966366291046, + 0.00875864364206791, + 0.005630367435514927, + 0.007345074322074652, + 0.009493384510278702, + 0.004486896563321352, + 0.007095202337950468, + 0.005272052716463804, + 0.006014195270836353, + 0.005924101918935776, + 0.009367657825350761, + 0.0068855490535497665, + 0.007843991741538048, + 0.006524670869112015, + 0.009840092621743679, + 0.02259530872106552, + 0.011235402897000313, + 0.012629033997654915, + 0.028843004256486893, + 0.07810074090957642, + 0.02550630457699299, + 0.008729208260774612, + 0.005890186410397291, + 0.0038370639085769653, + 0.00643590884283185, + 0.0050108893774449825, + 0.003970324993133545, + 0.0033072330988943577, + 0.004252666607499123, + 0.0035663836169987917, + 0.006208977662026882, + 0.013649798929691315, + 0.025699492543935776, + 0.019527990370988846, + 0.026561882346868515, + 0.021089743822813034, + 0.033288273960351944, + 0.013321300968527794, + 0.023012088611721992, + 0.006917380727827549, + 0.00690776901319623, + 0.008100012317299843, + 0.010877419263124466, + 0.010746528394520283, + 0.014827225357294083, + 0.006343028508126736, + 0.009116721339523792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.011107752099633217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307002201676369 + ], + "opposite_logprob_tokens": [ + -3.168070077896118 + ], + "logprob_reward": -0.04307002201676369, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307002201676369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1323339194059372, + 0.2015436738729477, + 0.1376824676990509, + 0.04856595769524574, + 0.06834733486175537, + 0.10922107100486755, + 0.09786301106214523, + 0.04614682123064995, + 0.0890318751335144, + 0.06926383823156357 + ], + "credit_map": [ + 0.01813185214996338, + 0.013379099778831005, + 0.02471373975276947, + 0.024501800537109375, + 0.01346462219953537, + 0.017751839011907578, + 0.010470077395439148, + 0.009920884855091572, + 0.01713898964226246, + 0.010528074577450752, + 0.06754650175571442, + 0.07863634824752808, + 0.006517895497381687, + 0.00868033803999424, + 0.007281913422048092, + 0.005213615484535694, + 0.01857549324631691, + 0.010654466226696968, + 0.021649891510605812, + 0.009100948460400105, + 0.04909251257777214, + 0.009668262675404549, + 0.009029596112668514, + 0.009911307133734226, + 0.005030850879848003, + 0.005800074432045221, + 0.006030343472957611, + 0.009063451550900936, + 0.005268353037536144, + 0.0037763165310025215, + 0.00707671046257019, + 0.006519859656691551, + 0.005958542227745056, + 0.006985937245190144, + 0.010613025166094303, + 0.006562403868883848, + 0.012688327580690384, + 0.011237782426178455, + 0.006812839303165674, + 0.007488477975130081, + 0.008733232505619526, + 0.010698276571929455, + 0.012620151974260807, + 0.011129038408398628, + 0.013714535161852837, + 0.013176191598176956, + 0.012970170006155968, + 0.0261794775724411, + 0.010282883420586586, + 0.007384773809462786, + 0.007236797828227282, + 0.01590740494430065, + 0.03523996099829674, + 0.011772653087973595, + 0.004391550552099943, + 0.005646989680826664, + 0.004071136936545372, + 0.00566188246011734, + 0.004664439242333174, + 0.008382721804082394, + 0.0029752589762210846, + 0.00568803446367383, + 0.002479529706761241, + 0.01222381554543972, + 0.007618330884724855, + 0.013044571503996849, + 0.008896466344594955, + 0.01457233726978302, + 0.012395389378070831, + 0.01820984296500683, + 0.006726374849677086, + 0.00756856519728899, + 0.006537921726703644, + 0.007182066794484854, + 0.00757786026224494, + 0.00984872318804264, + 0.010804295539855957, + 0.012104777619242668, + 0.006052578333765268, + 0.009155617095530033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.04307002201676369 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298132486641407 + ], + "opposite_logprob_tokens": [ + -3.529813289642334 + ], + "logprob_reward": -0.0298132486641407, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298132486641407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06678266823291779, + 0.11794379353523254, + 0.059342287480831146, + 0.10878095775842667, + 0.07874495536088943, + 0.0437164232134819, + 0.22557950019836426, + 0.03310387209057808, + 0.16215889155864716, + 0.10384665429592133 + ], + "credit_map": [ + 0.0059477961622178555, + 0.00871964544057846, + 0.011936451308429241, + 0.013876982033252716, + 0.007255059666931629, + 0.006612757686525583, + 0.006342506967484951, + 0.006091469898819923, + 0.030689526349306107, + 0.010102598927915096, + 0.0228738933801651, + 0.029158955439925194, + 0.005837815813720226, + 0.007658135145902634, + 0.006524638272821903, + 0.0050982278771698475, + 0.008351284079253674, + 0.005713437683880329, + 0.00950034148991108, + 0.0037977788597345352, + 0.018785065039992332, + 0.004093660973012447, + 0.00466797174885869, + 0.004432749934494495, + 0.013463547453284264, + 0.012948337942361832, + 0.021028323099017143, + 0.009103632532060146, + 0.028263559564948082, + 0.011276421137154102, + 0.00473615201190114, + 0.00796098168939352, + 0.006480223499238491, + 0.011952690780162811, + 0.011000514961779118, + 0.012672808021306992, + 0.01812099479138851, + 0.006896935403347015, + 0.005703731905668974, + 0.005917055532336235, + 0.007490845397114754, + 0.006154034286737442, + 0.005636891350150108, + 0.006350422278046608, + 0.0045587485656142235, + 0.004222610965371132, + 0.003928082529455423, + 0.005374784115701914, + 0.023409049957990646, + 0.014705647714436054, + 0.01341935247182846, + 0.038954250514507294, + 0.09077779203653336, + 0.029549390077590942, + 0.010090929456055164, + 0.00467306561768055, + 0.0029270448721945286, + 0.006494165398180485, + 0.0036621589679270983, + 0.004023130051791668, + 0.0034718711394816637, + 0.004831835627555847, + 0.003541595069691539, + 0.004152071196585894, + 0.013083787634968758, + 0.02537601813673973, + 0.01851845160126686, + 0.024349631741642952, + 0.018541773781180382, + 0.03220265731215477, + 0.011983649805188179, + 0.018102917820215225, + 0.009744812734425068, + 0.010491638444364071, + 0.011240500956773758, + 0.017671603709459305, + 0.016812706366181374, + 0.017340227961540222, + 0.009820210747420788, + 0.010724948719143867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.0298132486641407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989622056484222 + ], + "opposite_logprob_tokens": [ + -1.8048962354660034 + ], + "logprob_reward": -0.17989622056484222, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989622056484222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11612755060195923, + 0.23581193387508392, + 0.4282910227775574, + 0.01980666071176529, + 0.025270139798521996, + 0.024389320984482765, + 0.06355120241641998, + 0.01778339222073555, + 0.04300517216324806, + 0.02596365660429001 + ], + "credit_map": [ + 0.016834184527397156, + 0.015114705078303814, + 0.023489659652113914, + 0.019410178065299988, + 0.008058805018663406, + 0.014780220575630665, + 0.009995223954319954, + 0.008444569073617458, + 0.025056304410099983, + 0.015845857560634613, + 0.07412304729223251, + 0.09061223268508911, + 0.006960439030081034, + 0.009840314276516438, + 0.007799897342920303, + 0.005573855713009834, + 0.04711415246129036, + 0.03723450377583504, + 0.07693560421466827, + 0.03533633425831795, + 0.13071179389953613, + 0.04566589370369911, + 0.025483189150691032, + 0.02980954200029373, + 0.0022254709620028734, + 0.002369594294577837, + 0.002369390334933996, + 0.0019874069839715958, + 0.0022671737242490053, + 0.002045081229880452, + 0.0028923600912094116, + 0.0036501826252788305, + 0.0025370242074131966, + 0.004129406530410051, + 0.0038215930107980967, + 0.002285822993144393, + 0.004451131913810968, + 0.0034999437630176544, + 0.0016765868058428168, + 0.0028686311561614275, + 0.0037809552159160376, + 0.0027068757917732, + 0.0029492946341633797, + 0.00278662727214396, + 0.0028925039805471897, + 0.003638295456767082, + 0.0023178302217274904, + 0.0033169386442750692, + 0.00627932557836175, + 0.004645646084100008, + 0.0044495342299342155, + 0.01037589181214571, + 0.024718383327126503, + 0.00837628822773695, + 0.002926793647930026, + 0.0017793329898267984, + 0.0018158078892156482, + 0.002475498942658305, + 0.002597286133095622, + 0.002641848521307111, + 0.0013257066020742059, + 0.002111390931531787, + 0.0015516116982325912, + 0.0032642397563904524, + 0.0035769147798419, + 0.006246492266654968, + 0.005286120343953371, + 0.004783082287758589, + 0.0072477953508496284, + 0.007703873794525862, + 0.003477994119748473, + 0.00468289852142334, + 0.0027073267847299576, + 0.003068300196900964, + 0.00320394616574049, + 0.00428986269980669, + 0.0023331772536039352, + 0.0030963504686951637, + 0.0030214949510991573, + 0.004243197850883007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.17989622056484222 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld, slightly shaky, medium-long shot captures a bustling, sunny public square as a man dressed entirely in black approaches two young children. The camera tracks him with subtle, unsteady movements, enhancing the scene's raw immediacy. The man, dressed in a black leather jacket, a black T-shirt, and dark jeans, moves with an unconventional rhythm, twisting his torso while speaking enthusiastically and gesturing. His face is initially obscured by dark sunglasses. Two children, a boy and a girl, wearing brightly colored hoodies, stand close together, their eyes wide and uncertain as they watch his approach. As the man closes in distance, he deliberately slows down, smoothly removing his sunglasses to reveal a gentle smile. He then bends down from the waist, lowering his upper body to level with the children in a non-threatening gesture. In response, the children's previously motionless hands begin to move; one child fidgets with his sleeve, another makes a small, hesitant gesture forward. Throughout this interaction, the background remains vibrant: figures in the distance stroll through the square, and at one point, a passerby briskly walks through the foreground, momentarily interrupting the frame. The entire scene is bathed in bright, cool daylight, creating a stark contrast and a tense, unfolding atmosphere of curiosity, all captured with an erratic camera that suggests a spontaneous, unstaged moment.", + "seed": 8438, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_9154_seed_8438_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10747195780277252, + "vlm_logprob": -0.10747195780277252, + "vlm_yes_no_margin": 3.7499999940395354, + "vlm_token_credit": -0.10747195780277252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010513515211641788 + ], + "opposite_logprob_tokens": [ + -5.2605133056640625 + ], + "logprob_reward": -0.010513515211641788, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.010513515211641788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21336425840854645, + 0.04536355659365654, + 0.03680720925331116, + 0.16440610587596893, + 0.022961117327213287, + 0.04363122582435608, + 0.08935244381427765, + 0.2646893560886383, + 0.06458686292171478, + 0.054837871342897415 + ], + "credit_map": [ + 0.07342630624771118, + 0.02126455120742321, + 0.05674387514591217, + 0.011682676151394844, + 0.021862680092453957, + 0.012303234077990055, + 0.00906398519873619, + 0.0070169661194086075, + 0.004876700229942799, + 0.006236504763364792, + 0.005324086640030146, + 0.006824835203588009, + 0.005050155334174633, + 0.00786603707820177, + 0.005645194090902805, + 0.0035400441847741604, + 0.0032450519502162933, + 0.00440782168880105, + 0.002966562518849969, + 0.00425450736656785, + 0.005038495175540447, + 0.00975129660218954, + 0.0038307311479002237, + 0.0033127462957054377, + 0.05191579461097717, + 0.017396414652466774, + 0.007877605967223644, + 0.009256184101104736, + 0.006231263279914856, + 0.058583591133356094, + 0.006292545236647129, + 0.006852729246020317, + 0.0033244723454117775, + 0.003041231771931052, + 0.003474536817520857, + 0.0030381805263459682, + 0.0024873532820492983, + 0.00317439460195601, + 0.0018079715082421899, + 0.0026129777543246746, + 0.006846757605671883, + 0.004036607686430216, + 0.006296443287283182, + 0.004460819065570831, + 0.006001134403049946, + 0.008578130044043064, + 0.0031649458687752485, + 0.004246390890330076, + 0.025552935898303986, + 0.010391484014689922, + 0.019426271319389343, + 0.005935213528573513, + 0.004874910227954388, + 0.012760196812450886, + 0.00402818713337183, + 0.0063832541927695274, + 0.053882207721471786, + 0.04054199531674385, + 0.03940850496292114, + 0.06137090176343918, + 0.013373836874961853, + 0.016806401312351227, + 0.025539491325616837, + 0.013766021467745304, + 0.007762064225971699, + 0.00937258917838335, + 0.005820738151669502, + 0.005005543120205402, + 0.007121566217392683, + 0.014672467485070229, + 0.007197279948741198, + 0.007634617853909731, + 0.006950239650905132, + 0.004751021042466164, + 0.004397472832351923, + 0.01576094888150692, + 0.007751082070171833, + 0.003242664970457554, + 0.006109541282057762, + 0.005874906666576862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.010513515211641788 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004654881544411182 + ], + "opposite_logprob_tokens": [ + -5.379654884338379 + ], + "logprob_reward": -0.004654881544411182, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004654881544411182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07285651564598083, + 0.04566682130098343, + 0.024203434586524963, + 0.6293450593948364, + 0.017578771337866783, + 0.0348295159637928, + 0.02156144008040428, + 0.0428287610411644, + 0.024227024987339973, + 0.0869026929140091 + ], + "credit_map": [ + 0.027920279651880264, + 0.006495852023363113, + 0.019695859402418137, + 0.0039040865376591682, + 0.003662561532109976, + 0.004467485472559929, + 0.0033623133786022663, + 0.0033480734564363956, + 0.0053200735710561275, + 0.009319944307208061, + 0.006540378555655479, + 0.0035600420087575912, + 0.007847598753869534, + 0.005055586341768503, + 0.004157703369855881, + 0.0038654974196106195, + 0.002525928197428584, + 0.0025854080449789762, + 0.0022020158357918262, + 0.0026891836896538734, + 0.004077606834471226, + 0.004834712017327547, + 0.0023308780509978533, + 0.002957702148705721, + 0.22569121420383453, + 0.041318099945783615, + 0.02681686542928219, + 0.032015927135944366, + 0.020333481952548027, + 0.231119304895401, + 0.02616124041378498, + 0.025888901203870773, + 0.002699041971936822, + 0.005884829442948103, + 0.0012238872004672885, + 0.0015470371581614017, + 0.003033614484593272, + 0.001692108460702002, + 0.0007805066416040063, + 0.0007177466177381575, + 0.0055431826040148735, + 0.0038105230778455734, + 0.0061321984976530075, + 0.0035231609363108873, + 0.006046534515917301, + 0.004033156670629978, + 0.0024677340406924486, + 0.003273021662607789, + 0.005227239336818457, + 0.002332865260541439, + 0.003585230791941285, + 0.0022190806921571493, + 0.0025602050591260195, + 0.0027684152591973543, + 0.0014446736313402653, + 0.0014237299328669906, + 0.008764433674514294, + 0.006438776385039091, + 0.006777491420507431, + 0.007213063538074493, + 0.004726199898868799, + 0.00280241621658206, + 0.0035438574850559235, + 0.0025625198613852262, + 0.0037219661753624678, + 0.0036063948646187782, + 0.002415744587779045, + 0.001986684976145625, + 0.0036767099518328905, + 0.0031458663288503885, + 0.0032889137510210276, + 0.002384743420407176, + 0.011002170853316784, + 0.007173424586653709, + 0.00635989336296916, + 0.017764879390597343, + 0.02234654687345028, + 0.006242211442440748, + 0.008205683901906013, + 0.0078078825026750565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.004654881544411182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198238030076027 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198238030076027, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198238030076027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10688766837120056, + 0.021065276116132736, + 0.053375158458948135, + 0.6532037854194641, + 0.012435946613550186, + 0.01788676157593727, + 0.01759369485080242, + 0.03538772836327553, + 0.021723056212067604, + 0.06044089421629906 + ], + "credit_map": [ + 0.037119071930646896, + 0.009805167093873024, + 0.02958916686475277, + 0.005084579344838858, + 0.011265548877418041, + 0.005933862179517746, + 0.0043624225072562695, + 0.003727857256308198, + 0.002349219284951687, + 0.0031870254315435886, + 0.0021466175094246864, + 0.002650198992341757, + 0.0033148352522403, + 0.002380451885983348, + 0.00336289475671947, + 0.0016740331193432212, + 0.005540486890822649, + 0.004581382032483816, + 0.004345240071415901, + 0.006992970127612352, + 0.006382326129823923, + 0.014070628210902214, + 0.004705274943262339, + 0.006756848655641079, + 0.23265138268470764, + 0.04133832827210426, + 0.028628656640648842, + 0.03372754156589508, + 0.01918458193540573, + 0.2437082678079605, + 0.026495488360524178, + 0.027469513937830925, + 0.0033623073250055313, + 0.0021944621112197638, + 0.0011775396997109056, + 0.0008426322019658983, + 0.0017095118528231978, + 0.0011456450447440147, + 0.0008404070395044982, + 0.0011634419206529856, + 0.003297767136245966, + 0.002067716559395194, + 0.002190764993429184, + 0.0015568311791867018, + 0.002709491876885295, + 0.0028946141246706247, + 0.001231166534125805, + 0.0019384092884138227, + 0.00315964431501925, + 0.0023920400999486446, + 0.0020569043699651957, + 0.001551892957650125, + 0.0017340879421681166, + 0.0030427146703004837, + 0.001035160617902875, + 0.0026212504599243402, + 0.0066061667166650295, + 0.006311211735010147, + 0.0059516229666769505, + 0.003710166085511446, + 0.0043359617702662945, + 0.0031381892040371895, + 0.0027342396788299084, + 0.0026001709047704935, + 0.002732119057327509, + 0.002358473604544997, + 0.0015493046957999468, + 0.0014549329644069076, + 0.0035870405845344067, + 0.003793211653828621, + 0.003100795205682516, + 0.0031471787951886654, + 0.008456923998892307, + 0.005201869644224644, + 0.004581922199577093, + 0.007307725492864847, + 0.019897526130080223, + 0.0044997939839959145, + 0.005491124000400305, + 0.005004005506634712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.014198238030076027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033701442182064056 + ], + "opposite_logprob_tokens": [ + -3.4087014198303223 + ], + "logprob_reward": -0.033701442182064056, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033701442182064056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0508011095225811, + 0.02698896825313568, + 0.030370431020855904, + 0.4744361340999603, + 0.04818529263138771, + 0.0392046794295311, + 0.049062393605709076, + 0.06313973665237427, + 0.03631153330206871, + 0.18149976432323456 + ], + "credit_map": [ + 0.010421845130622387, + 0.007340296637266874, + 0.00689803808927536, + 0.003568108193576336, + 0.011045114137232304, + 0.004071400500833988, + 0.003653813386335969, + 0.0038024941459298134, + 0.0034220588859170675, + 0.00499551510438323, + 0.0030513291712850332, + 0.0025375159457325935, + 0.0033744603861123323, + 0.004955730400979519, + 0.0022779651917517185, + 0.002374395728111267, + 0.0029794948641210794, + 0.0047441646456718445, + 0.0022795831318944693, + 0.004217649344354868, + 0.0032589242327958345, + 0.006608016323298216, + 0.0029992149211466312, + 0.0032833823934197426, + 0.16661854088306427, + 0.031145010143518448, + 0.020669378340244293, + 0.024903761222958565, + 0.016406133770942688, + 0.17511941492557526, + 0.019757987931370735, + 0.019815905019640923, + 0.006746689788997173, + 0.015530531294643879, + 0.003429221222177148, + 0.0029485949780792, + 0.008365079760551453, + 0.005486201494932175, + 0.0030935327522456646, + 0.002585441805422306, + 0.0075559052638709545, + 0.004049776587635279, + 0.005162796936929226, + 0.00360490451566875, + 0.007508696056902409, + 0.004851170815527439, + 0.003205362008884549, + 0.0032660644501447678, + 0.006331843324005604, + 0.007160521578043699, + 0.006534317042678595, + 0.006943223997950554, + 0.008276315405964851, + 0.005868231877684593, + 0.004240219481289387, + 0.0037077206652611494, + 0.013055074959993362, + 0.008122250437736511, + 0.009089374914765358, + 0.01333765871822834, + 0.004580180626362562, + 0.004817166831344366, + 0.005932363215833902, + 0.00420566089451313, + 0.0043423674069345, + 0.004287765361368656, + 0.0020771368872374296, + 0.0031089650001376867, + 0.0036296690814197063, + 0.009096660651266575, + 0.004443804733455181, + 0.00532515998929739, + 0.01928948052227497, + 0.013954310677945614, + 0.014984993264079094, + 0.05561266839504242, + 0.02983853965997696, + 0.01114708837121725, + 0.018855750560760498, + 0.017816931009292603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.033701442182064056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47429171204566956 + ], + "opposite_logprob_tokens": [ + -0.9742916822433472 + ], + "logprob_reward": -0.47429171204566956, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47429171204566956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1502886563539505, + 0.04091297462582588, + 0.032279446721076965, + 0.3699624538421631, + 0.026525743305683136, + 0.04301977530121803, + 0.0313839353621006, + 0.04686078801751137, + 0.032461017370224, + 0.22630523145198822 + ], + "credit_map": [ + 0.04253410920500755, + 0.01610790379345417, + 0.03370579704642296, + 0.007329168729484081, + 0.02450127899646759, + 0.010572747327387333, + 0.008800584822893143, + 0.0067370617762207985, + 0.005177769344300032, + 0.005848847795277834, + 0.004056173842400312, + 0.007550723850727081, + 0.00600370904430747, + 0.004879384767264128, + 0.004641241393983364, + 0.002755122957751155, + 0.007081184536218643, + 0.006972676608711481, + 0.0027901106514036655, + 0.004340479616075754, + 0.0024810663890093565, + 0.00366189400665462, + 0.002774127060547471, + 0.0021779052913188934, + 0.1304813027381897, + 0.0227931160479784, + 0.016949929296970367, + 0.01901533640921116, + 0.012561306357383728, + 0.13771776854991913, + 0.014978447929024696, + 0.01546523068100214, + 0.0049707829020917416, + 0.006261782720685005, + 0.0031649696175009012, + 0.0022896751761436462, + 0.0027112734969705343, + 0.003199838800355792, + 0.0015648688422515988, + 0.0023625483736395836, + 0.0071728164330124855, + 0.004967139568179846, + 0.006462551653385162, + 0.004073035437613726, + 0.008029128424823284, + 0.004896503407508135, + 0.0035930790472775698, + 0.0038255176041275263, + 0.009256103076040745, + 0.0022652153857052326, + 0.007196228019893169, + 0.002134159440174699, + 0.0026197554543614388, + 0.003643856616690755, + 0.0014045854331925511, + 0.0028640313539654016, + 0.00754893571138382, + 0.0074199652299284935, + 0.005945282056927681, + 0.012009437195956707, + 0.0028160244692116976, + 0.0028128603007644415, + 0.005858440883457661, + 0.002449837513267994, + 0.005038760136812925, + 0.0051276907324790955, + 0.002725920407101512, + 0.0026441561058163643, + 0.004940931685268879, + 0.0032886045519262552, + 0.005171842873096466, + 0.003523108549416065, + 0.01305465679615736, + 0.011684727855026722, + 0.01394061092287302, + 0.09525196254253387, + 0.04344599321484566, + 0.008144809864461422, + 0.024626104161143303, + 0.016156358644366264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.47429171204566956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld, slightly shaky, medium-long shot captures a bustling, sunny public square as a man dressed entirely in black approaches two young children. The camera tracks him with subtle, unsteady movements, enhancing the scene's raw immediacy. The man, dressed in a black leather jacket, a black T-shirt, and dark jeans, moves with an unconventional rhythm, twisting his torso while speaking enthusiastically and gesturing. His face is initially obscured by dark sunglasses. Two children, a boy and a girl, wearing brightly colored hoodies, stand close together, their eyes wide and uncertain as they watch his approach. As the man closes in distance, he deliberately slows down, smoothly removing his sunglasses to reveal a gentle smile. He then bends down from the waist, lowering his upper body to level with the children in a non-threatening gesture. In response, the children's previously motionless hands begin to move; one child fidgets with his sleeve, another makes a small, hesitant gesture forward. Throughout this interaction, the background remains vibrant: figures in the distance stroll through the square, and at one point, a passerby briskly walks through the foreground, momentarily interrupting the frame. The entire scene is bathed in bright, cool daylight, creating a stark contrast and a tense, unfolding atmosphere of curiosity, all captured with an erratic camera that suggests a spontaneous, unstaged moment.", + "seed": 8439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_10/index_9154_seed_8439_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07826056927442551, + "vlm_logprob": -0.07826056927442551, + "vlm_yes_no_margin": 3.6249999403953552, + "vlm_token_credit": -0.07826056927442551, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03889428824186325 + ], + "opposite_logprob_tokens": [ + -3.6638941764831543 + ], + "logprob_reward": -0.03889428824186325, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03889428824186325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6890183687210083, + 0.08434487134218216, + 0.02570062316954136, + 0.02164594642817974, + 0.03625660389661789, + 0.04667964205145836, + 0.017259862273931503, + 0.009700722992420197, + 0.03767992556095123, + 0.03171340376138687 + ], + "credit_map": [ + 0.03207479417324066, + 0.02404579147696495, + 0.04243229329586029, + 0.02421344630420208, + 0.040426578372716904, + 0.04885680601000786, + 0.24603241682052612, + 0.2309362292289734, + 0.006571289151906967, + 0.00790134072303772, + 0.005272333975881338, + 0.006128533743321896, + 0.018675023689866066, + 0.01333373412489891, + 0.0131435077637434, + 0.01331910490989685, + 0.0023935064673423767, + 0.003939101938158274, + 0.0032082137186080217, + 0.002964457729831338, + 0.0033860704861581326, + 0.0037999332416802645, + 0.0031962101347744465, + 0.0028131306171417236, + 0.0015520213637501001, + 0.0033251659478992224, + 0.002216920955106616, + 0.0020331693813204765, + 0.0027814528439193964, + 0.005752208177000284, + 0.0020306119695305824, + 0.001954394392669201, + 0.003932625520974398, + 0.005663667805492878, + 0.006215056404471397, + 0.0055365776643157005, + 0.0046862284652888775, + 0.005114419851452112, + 0.002293578116223216, + 0.0028144509997218847, + 0.009409219957888126, + 0.005290321074426174, + 0.005940806120634079, + 0.01578829437494278, + 0.002643236890435219, + 0.003554018447175622, + 0.0017795820022001863, + 0.0022741593420505524, + 0.005459160543978214, + 0.001609730301424861, + 0.004321851301938295, + 0.001414808677509427, + 0.0011384383542463183, + 0.0011628965148702264, + 0.0006717310752719641, + 0.0014812437584623694, + 0.001065815449692309, + 0.0015952906105667353, + 0.0016851102700456977, + 0.001605142722837627, + 0.001134344725869596, + 0.001167341019026935, + 0.0005375958862714469, + 0.0009100820170715451, + 0.003950278274714947, + 0.0043923924677073956, + 0.0049097780138254166, + 0.004347219131886959, + 0.00769371772184968, + 0.003985636401921511, + 0.003600347088649869, + 0.0048005543649196625, + 0.003192592179402709, + 0.00479503907263279, + 0.006012679543346167, + 0.004116875119507313, + 0.006691713817417622, + 0.0018545116763561964, + 0.0024837309028953314, + 0.002566263312473893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03889428824186325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004670187830924988 + ], + "opposite_logprob_tokens": [ + -5.379670143127441 + ], + "logprob_reward": -0.004670187830924988, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004670187830924988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4102195203304291, + 0.08749505877494812, + 0.02478557452559471, + 0.021025732159614563, + 0.03237944841384888, + 0.016697246581315994, + 0.032453130930662155, + 0.10023735463619232, + 0.20109020173549652, + 0.07361672073602676 + ], + "credit_map": [ + 0.018175091594457626, + 0.01531639788299799, + 0.025242162868380547, + 0.014229796826839447, + 0.024109436199069023, + 0.0252431258559227, + 0.1483571082353592, + 0.13954640924930573, + 0.007024225778877735, + 0.011585528030991554, + 0.007955946959555149, + 0.013965405523777008, + 0.011801676824688911, + 0.009543958120048046, + 0.01230003871023655, + 0.013318277895450592, + 0.00252752797678113, + 0.003712894394993782, + 0.0019654056522995234, + 0.003257844364270568, + 0.004605955444276333, + 0.0028486826922744513, + 0.0035923251416534185, + 0.002274939091876149, + 0.001905907061882317, + 0.0033237505704164505, + 0.0018640118651092052, + 0.0035251404624432325, + 0.0034549196716398, + 0.002800047630444169, + 0.00233516376465559, + 0.0018167904345318675, + 0.003098048735409975, + 0.005469264462590218, + 0.0033078687265515327, + 0.004993816837668419, + 0.002531567355617881, + 0.005414912011474371, + 0.003144960617646575, + 0.004419011063873768, + 0.002220502123236656, + 0.0025568497367203236, + 0.0026442252565175295, + 0.002626494737342, + 0.0014811359578743577, + 0.002290166914463043, + 0.001470909919589758, + 0.0014069617027416825, + 0.0060569243505597115, + 0.005873195827007294, + 0.005206997040659189, + 0.005650684237480164, + 0.0029036474879831076, + 0.002457398921251297, + 0.001995926024392247, + 0.0023083589039742947, + 0.007006322033703327, + 0.010699419304728508, + 0.016978750005364418, + 0.014021274633705616, + 0.012358201667666435, + 0.017955170944333076, + 0.007723241113126278, + 0.013494977727532387, + 0.017287231981754303, + 0.025243325158953667, + 0.03203307464718819, + 0.025923486799001694, + 0.04394431412220001, + 0.015638228505849838, + 0.019483428448438644, + 0.021537132561206818, + 0.005201964173465967, + 0.009636328555643559, + 0.013653005473315716, + 0.0070776198990643024, + 0.01747943088412285, + 0.00551846157759428, + 0.00681242486461997, + 0.008237486705183983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.004670187830924988 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076931729912758 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.016076931729912758, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076931729912758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37121349573135376, + 0.03569044545292854, + 0.03308483213186264, + 0.029505331069231033, + 0.05531817302107811, + 0.0337299220263958, + 0.040689900517463684, + 0.13738325238227844, + 0.1009468212723732, + 0.16243784129619598 + ], + "credit_map": [ + 0.01712038554251194, + 0.012343618087470531, + 0.023604407906532288, + 0.015455921180546284, + 0.02183063142001629, + 0.022971374914050102, + 0.133536696434021, + 0.12435045093297958, + 0.0033198341261595488, + 0.0038249974604696035, + 0.0025823055766522884, + 0.0034390706568956375, + 0.005322681274265051, + 0.0051593598909676075, + 0.005536755081266165, + 0.0065054381266236305, + 0.003121111076325178, + 0.0038639188278466463, + 0.00305391987785697, + 0.0034366026520729065, + 0.00557288620620966, + 0.0057187569327652454, + 0.004558957181870937, + 0.003758676117286086, + 0.0031332450453191996, + 0.0060963938012719154, + 0.0027122613973915577, + 0.0038770621176809072, + 0.0034007749054580927, + 0.003934389445930719, + 0.0031596345361322165, + 0.0031915693543851376, + 0.00429744366556406, + 0.0062024942599236965, + 0.0044044447131454945, + 0.009325551800429821, + 0.006791920866817236, + 0.011559178121387959, + 0.004268039483577013, + 0.008469093590974808, + 0.005563771817833185, + 0.0050290776416659355, + 0.0048122587613761425, + 0.00393585953861475, + 0.0034219883382320404, + 0.005975861568003893, + 0.002524918643757701, + 0.0024661829229444265, + 0.008638015016913414, + 0.004578718915581703, + 0.007610435597598553, + 0.0033326344564557076, + 0.003906234400346875, + 0.005044115241616964, + 0.002804841846227646, + 0.004774899687618017, + 0.010583286173641682, + 0.017360202968120575, + 0.01588641107082367, + 0.01953931525349617, + 0.016893764957785606, + 0.028681429103016853, + 0.009414826519787312, + 0.01902400515973568, + 0.006371775642037392, + 0.013557705096900463, + 0.01694183424115181, + 0.011651940643787384, + 0.025617506355047226, + 0.005442896857857704, + 0.011027987115085125, + 0.010335169732570648, + 0.011839057318866253, + 0.019048146903514862, + 0.03034541755914688, + 0.014090784825384617, + 0.041706383228302, + 0.013302984647452831, + 0.014600103721022606, + 0.017504949122667313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.016076931729912758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192922696471214 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.018192922696471214, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018192922696471214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20990365743637085, + 0.015362237580120564, + 0.027584195137023926, + 0.025403572246432304, + 0.03513171523809433, + 0.024142032489180565, + 0.07071511447429657, + 0.15940721333026886, + 0.3410086929798126, + 0.09134155511856079 + ], + "credit_map": [ + 0.01036166399717331, + 0.006879590917378664, + 0.013674682937562466, + 0.007787907961755991, + 0.013776525855064392, + 0.01563417539000511, + 0.07207059860229492, + 0.06971850991249084, + 0.0013368064537644386, + 0.0028448670636862516, + 0.0018197896424680948, + 0.0015719195362180471, + 0.0022048749960958958, + 0.001842236379161477, + 0.001818790566176176, + 0.0019229536410421133, + 0.0030365644488483667, + 0.0051917992532253265, + 0.0028478335589170456, + 0.0038922030944377184, + 0.0034452287945896387, + 0.002996544586494565, + 0.0033532085362821817, + 0.0028208130970597267, + 0.002586611546576023, + 0.0041658817790448666, + 0.0023618056438863277, + 0.0047424244694411755, + 0.0027498421259224415, + 0.003524738596752286, + 0.003188240109011531, + 0.002084028674289584, + 0.00363271776586771, + 0.007481697481125593, + 0.004082540515810251, + 0.004816608037799597, + 0.0039552804082632065, + 0.0049432204104959965, + 0.002744043245911598, + 0.0034756073728203773, + 0.006119787227362394, + 0.0023224754258990288, + 0.0040597254410386086, + 0.0026303729973733425, + 0.0020745147485285997, + 0.0035578981041908264, + 0.001482369378209114, + 0.0018948898650705814, + 0.014311832375824451, + 0.012215224094688892, + 0.012199227698147297, + 0.014846239238977432, + 0.005195497535169125, + 0.004366842098534107, + 0.0028440661262720823, + 0.004736185539513826, + 0.011239973828196526, + 0.030885519459843636, + 0.029717128723859787, + 0.02478494867682457, + 0.026927895843982697, + 0.019640546292066574, + 0.007543694693595171, + 0.008667496964335442, + 0.03316499665379524, + 0.04711294174194336, + 0.05540581792593002, + 0.041836224496364594, + 0.08159048110246658, + 0.023965997621417046, + 0.028955034911632538, + 0.028977204114198685, + 0.006396288052201271, + 0.010816284455358982, + 0.017226098105311394, + 0.008289084769785404, + 0.023744812235236168, + 0.006354598794132471, + 0.008568854071199894, + 0.009945531375706196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.018192922696471214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bustling, sunny public square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man dressed in a black leather jacket, black T-shirt, and dark jeans approaching two children?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the two children wear brightly colored hoodies and display wide, uncertain expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man remove his dark sunglasses to reveal a gentle smile and then bend down to the children's level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld and slightly shaky with erratic movements throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134685158729553 + ], + "opposite_logprob_tokens": [ + -1.3134684562683105 + ], + "logprob_reward": -0.3134685158729553, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134685158729553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17029540240764618, + 0.06476949155330658, + 0.054878104478120804, + 0.03976107016205788, + 0.21745382249355316, + 0.04662267491221428, + 0.16653601825237274, + 0.08602356910705566, + 0.07819787412881851, + 0.07546192407608032 + ], + "credit_map": [ + 0.014957786537706852, + 0.01984189823269844, + 0.01329515315592289, + 0.017085010185837746, + 0.024867167696356773, + 0.03709707409143448, + 0.02013634331524372, + 0.023014992475509644, + 0.006158933043479919, + 0.009849455207586288, + 0.00517726456746459, + 0.005908437073230743, + 0.009477761574089527, + 0.009144353680312634, + 0.010024421848356724, + 0.009028873406350613, + 0.005117742344737053, + 0.008225583471357822, + 0.00762954680249095, + 0.00716437678784132, + 0.007644621655344963, + 0.006586409639567137, + 0.006959439720958471, + 0.00555038545280695, + 0.0035490142181515694, + 0.006494942586869001, + 0.0047951010055840015, + 0.0035286913625895977, + 0.005863546393811703, + 0.008273676037788391, + 0.0041722627356648445, + 0.0030838423408567905, + 0.027220575138926506, + 0.05231315642595291, + 0.03137892484664917, + 0.030779173597693443, + 0.022202907130122185, + 0.02261391095817089, + 0.013252836652100086, + 0.017692355439066887, + 0.00691697932779789, + 0.009726488962769508, + 0.006781549658626318, + 0.008740631863474846, + 0.00287028681486845, + 0.004246583208441734, + 0.003061895724385977, + 0.004278265405446291, + 0.03442196920514107, + 0.02674364671111107, + 0.029223619028925896, + 0.03181926906108856, + 0.01191493310034275, + 0.010804123245179653, + 0.008460264652967453, + 0.013148203492164612, + 0.007168090436607599, + 0.014356298372149467, + 0.015050534158945084, + 0.013305293396115303, + 0.011187158524990082, + 0.007884512655436993, + 0.006491081789135933, + 0.010580606758594513, + 0.008910349570214748, + 0.010386263951659203, + 0.011259302496910095, + 0.0098106749355793, + 0.012812088243663311, + 0.008743374608457088, + 0.0076909298077225685, + 0.00858489703387022, + 0.011835708282887936, + 0.013139646500349045, + 0.013601009733974934, + 0.00879677478224039, + 0.009014302864670753, + 0.004165366757661104, + 0.004857897292822599, + 0.010051226243376732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.3134685158729553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_11/index_tensor([1256, 1256])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_11/index_tensor([1256, 1256])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a1c2074e29aba8a367bf1d6d8af2f93cc535a2c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_11/index_tensor([1256, 1256])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, intimate close-up focuses on the hands of an elderly couple, their skin wrinkled with age and marked by time, as they collaborate with gentle precision. The man's broader, more veined hands firmly grasp the dark mahogany frame of a photograph, leaning against the warm cream-colored living room wall. The woman's fingers, more slender, their skin almost translucent, carefully adjust the alignment. They are straightening a slightly curved, glass-covered photograph that captures younger versions of themselves on a sunny beach. The entire action is bathed in the soft, warm glow of an unseen table lamp, which casts gentle shadows of their hands and creates soft reflections on the glass of the photograph. With a final, deliberate, and almost imperceptible nudge, the woman brings the frame into a perfectly level position. Her touch lingers for a moment as her hands secure it in place, creating a serene and affectionate atmosphere of quiet collaboration and shared history.", + "seed": 11339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_11/index_1256_seed_11339_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038967933412641285, + "vlm_logprob": -0.038967933412641285, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.038967933412641285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1043928936123848 + ], + "opposite_logprob_tokens": [ + -2.3543930053710938 + ], + "logprob_reward": -0.1043928936123848, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1043928936123848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1307283639907837, + 0.05465041473507881, + 0.060965582728385925, + 0.04918397217988968, + 0.39787769317626953, + 0.07276266813278198, + 0.04315800219774246, + 0.1036311537027359, + 0.03863564506173134, + 0.04840647429227829 + ], + "credit_map": [ + 0.014909413643181324, + 0.0110292574390769, + 0.013860804960131645, + 0.009697855450212955, + 0.014239937998354435, + 0.023684076964855194, + 0.019902303814888, + 0.023404715582728386, + 0.0064338380470871925, + 0.007062784396111965, + 0.008624322712421417, + 0.0046767061576247215, + 0.007962962612509727, + 0.00731145404279232, + 0.005657659377902746, + 0.00692069111391902, + 0.006052563898265362, + 0.0068916562013328075, + 0.007520739454776049, + 0.006966621149331331, + 0.010480695404112339, + 0.009549141861498356, + 0.005578971933573484, + 0.007925193756818771, + 0.004403016529977322, + 0.005009989254176617, + 0.005290656350553036, + 0.0037544225342571735, + 0.005926972720772028, + 0.009282532148063183, + 0.005430408753454685, + 0.01008597668260336, + 0.08845596760511398, + 0.019906703382730484, + 0.014378705993294716, + 0.010806016623973846, + 0.020749792456626892, + 0.04337365925312042, + 0.05066625028848648, + 0.14954060316085815, + 0.0069347405806183815, + 0.00701460475102067, + 0.010202021338045597, + 0.009068773128092289, + 0.008443744853138924, + 0.012451711110770702, + 0.008534908294677734, + 0.010112172923982143, + 0.006129419431090355, + 0.005343634635210037, + 0.004085809923708439, + 0.0037519927136600018, + 0.004435767885297537, + 0.006259040907025337, + 0.008241786621510983, + 0.004910555202513933, + 0.01151299849152565, + 0.010861670598387718, + 0.011147803626954556, + 0.016746144741773605, + 0.01050974428653717, + 0.01683626137673855, + 0.014510447159409523, + 0.011506094597280025, + 0.004338839557021856, + 0.0046991631388664246, + 0.005571805872023106, + 0.006109829060733318, + 0.0026209240313619375, + 0.004600004758685827, + 0.005504542496055365, + 0.005190541502088308, + 0.006118930410593748, + 0.005654975306242704, + 0.005179127678275108, + 0.0038932671304792166, + 0.004650756251066923, + 0.006704965606331825, + 0.00664918078109622, + 0.009555274620652199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.1043928936123848 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038177166134119034 + ], + "opposite_logprob_tokens": [ + -3.288177251815796 + ], + "logprob_reward": -0.038177166134119034, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038177166134119034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2265632003545761, + 0.03211091458797455, + 0.08029249310493469, + 0.228941410779953, + 0.09974846243858337, + 0.07602665573358536, + 0.03770798444747925, + 0.05249631777405739, + 0.11327145993709564, + 0.052841175347566605 + ], + "credit_map": [ + 0.019687475636601448, + 0.02005443535745144, + 0.05469139292836189, + 0.021404052153229713, + 0.0224022064357996, + 0.020903734490275383, + 0.037828028202056885, + 0.029591869562864304, + 0.0031352536752820015, + 0.0040999725461006165, + 0.003745499299839139, + 0.004722511861473322, + 0.0054626124911010265, + 0.002568176481872797, + 0.004535735119134188, + 0.0038411503192037344, + 0.007124514784663916, + 0.009404047392308712, + 0.012137066572904587, + 0.006792740896344185, + 0.012160589918494225, + 0.009264239110052586, + 0.007815894670784473, + 0.015593399293720722, + 0.039936039596796036, + 0.009998645633459091, + 0.011348522268235683, + 0.005878021940588951, + 0.007940663024783134, + 0.03611106798052788, + 0.030190736055374146, + 0.08753769844770432, + 0.019495639950037003, + 0.00575163634493947, + 0.003912423737347126, + 0.004069420509040356, + 0.0069914390332996845, + 0.012320959009230137, + 0.012527719140052795, + 0.03467922285199165, + 0.007672050967812538, + 0.007630436215549707, + 0.008744001388549805, + 0.009543493390083313, + 0.005999527871608734, + 0.014833199791610241, + 0.009539131075143814, + 0.012064816430211067, + 0.005064837634563446, + 0.004096170887351036, + 0.004145684186369181, + 0.004402764141559601, + 0.0037207237910479307, + 0.004871696699410677, + 0.0069253402762115, + 0.004480766598135233, + 0.00492878770455718, + 0.005437825340777636, + 0.009800971485674381, + 0.010882122442126274, + 0.004751885309815407, + 0.006208823528140783, + 0.0049569434486329556, + 0.00552895525470376, + 0.013800318352878094, + 0.014622618444263935, + 0.013522259891033173, + 0.014439105987548828, + 0.008157890290021896, + 0.015354059636592865, + 0.017436614260077477, + 0.015938587486743927, + 0.0064877476543188095, + 0.004587643779814243, + 0.003859891090542078, + 0.004772500600665808, + 0.007375090382993221, + 0.008837667293846607, + 0.006016136612743139, + 0.010904494673013687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.038177166134119034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024926564656198025 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024926564656198025, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024926564656198025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22894491255283356, + 0.08593712002038956, + 0.04355509951710701, + 0.2274961620569229, + 0.18751606345176697, + 0.029897185042500496, + 0.0449080765247345, + 0.031010659411549568, + 0.07798294723033905, + 0.04275176301598549 + ], + "credit_map": [ + 0.02276225946843624, + 0.024247080087661743, + 0.034482840448617935, + 0.023969518020749092, + 0.025899259373545647, + 0.03650122880935669, + 0.028035860508680344, + 0.03304685652256012, + 0.010716668330132961, + 0.008213691413402557, + 0.010035383515059948, + 0.005810867995023727, + 0.008143650367856026, + 0.020126022398471832, + 0.008050818927586079, + 0.014840015210211277, + 0.0033131029922515154, + 0.005839847028255463, + 0.00633225729689002, + 0.0031995156314224005, + 0.007704702205955982, + 0.005379936192184687, + 0.004233830142766237, + 0.007551907561719418, + 0.040900375694036484, + 0.009789415635168552, + 0.01036454364657402, + 0.005634712986648083, + 0.008655264973640442, + 0.03513261303305626, + 0.030547861009836197, + 0.08647136390209198, + 0.041080158203840256, + 0.008762912824749947, + 0.006862293463200331, + 0.004881827160716057, + 0.00918081309646368, + 0.021033257246017456, + 0.023443402722477913, + 0.07227139919996262, + 0.003493041731417179, + 0.0029470210429280996, + 0.00282484688796103, + 0.004735628608614206, + 0.002810228383168578, + 0.0046868836507201195, + 0.003911371808499098, + 0.0044881622307002544, + 0.004735118243843317, + 0.006477715913206339, + 0.0041121505200862885, + 0.0050820461474359035, + 0.0042131515219807625, + 0.007567249704152346, + 0.006719234399497509, + 0.006001410540193319, + 0.003441285341978073, + 0.004009009804576635, + 0.005856456235051155, + 0.003648812649771571, + 0.0023261650931090117, + 0.005479834973812103, + 0.0030732229351997375, + 0.003175874473527074, + 0.009288107976317406, + 0.01101735606789589, + 0.009130655787885189, + 0.009587581269443035, + 0.0059868642129004, + 0.011413892731070518, + 0.011032745242118835, + 0.010525736957788467, + 0.0050256578251719475, + 0.004655675962567329, + 0.004973439034074545, + 0.0035591714549809694, + 0.007564869709312916, + 0.005391857586801052, + 0.00545012392103672, + 0.006130967754870653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0024926564656198025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370881825685501 + ], + "opposite_logprob_tokens": [ + -3.4087088108062744 + ], + "logprob_reward": -0.03370881825685501, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370881825685501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20192821323871613, + 0.035060182213783264, + 0.11154182255268097, + 0.09779684990644455, + 0.40638095140457153, + 0.04286763444542885, + 0.01868312433362007, + 0.028223885223269463, + 0.019085338339209557, + 0.03843199089169502 + ], + "credit_map": [ + 0.019807299599051476, + 0.01703362911939621, + 0.0394362136721611, + 0.01710040681064129, + 0.013060502707958221, + 0.031082306057214737, + 0.029560590162873268, + 0.03484727069735527, + 0.0027987987268716097, + 0.005239341873675585, + 0.004246560391038656, + 0.003997701685875654, + 0.005770847201347351, + 0.004380588885396719, + 0.0032170992344617844, + 0.005409245379269123, + 0.008310187608003616, + 0.012946883216500282, + 0.014134332537651062, + 0.009469342418015003, + 0.031383026391267776, + 0.014140731655061245, + 0.008542487397789955, + 0.012614832259714603, + 0.011086438782513142, + 0.010659872554242611, + 0.006832832470536232, + 0.006530412007123232, + 0.011972447857260704, + 0.01980789564549923, + 0.010307316668331623, + 0.02059963531792164, + 0.08758580684661865, + 0.020658399909734726, + 0.015095588751137257, + 0.011319811455905437, + 0.01936223730444908, + 0.046352043747901917, + 0.05040577054023743, + 0.15560133755207062, + 0.004946788307279348, + 0.003906704485416412, + 0.004435046575963497, + 0.005573573522269726, + 0.003254157956689596, + 0.007704856339842081, + 0.006513101048767567, + 0.006533410400152206, + 0.0023370643611997366, + 0.0020306396763771772, + 0.002212181920185685, + 0.0018341043032705784, + 0.0023382033687084913, + 0.0031200568191707134, + 0.003054543863981962, + 0.001756328041665256, + 0.002964835613965988, + 0.002906974172219634, + 0.0040589249692857265, + 0.0042516617104411125, + 0.002766742603853345, + 0.003936939872801304, + 0.004188171587884426, + 0.00314963492564857, + 0.0020524640567600727, + 0.002254771301522851, + 0.003816943848505616, + 0.001533366390503943, + 0.0019506008829921484, + 0.002265491057187319, + 0.0030022552236914635, + 0.002209445694461465, + 0.004671046510338783, + 0.0043125213123857975, + 0.0038692981470376253, + 0.0038988760206848383, + 0.007963369600474834, + 0.0045480974949896336, + 0.004362285602837801, + 0.004806492943316698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.03370881825685501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606813259422779 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.01606813259422779, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606813259422779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3184036314487457, + 0.10317578166723251, + 0.04141893610358238, + 0.042959269136190414, + 0.187160924077034, + 0.051525864750146866, + 0.018310708925127983, + 0.02663029544055462, + 0.043955959379673004, + 0.1664586365222931 + ], + "credit_map": [ + 0.04315975308418274, + 0.020423108711838722, + 0.0373900830745697, + 0.01701938733458519, + 0.02515818551182747, + 0.08344432711601257, + 0.030548660084605217, + 0.06126014143228531, + 0.012787862680852413, + 0.008035300299525261, + 0.01255222037434578, + 0.008220872841775417, + 0.009539220482110977, + 0.02208215370774269, + 0.010048026219010353, + 0.01991012878715992, + 0.0032656295225024223, + 0.005057315807789564, + 0.004735726397484541, + 0.0032381131313741207, + 0.0082725640386343, + 0.007535092532634735, + 0.003896917449310422, + 0.005417575594037771, + 0.004132802132517099, + 0.004780872259289026, + 0.003752373391762376, + 0.0028399338480085135, + 0.0058968993835151196, + 0.006620435975492001, + 0.00456182099878788, + 0.010374132543802261, + 0.04088715463876724, + 0.008692837320268154, + 0.00698815518990159, + 0.005394905339926481, + 0.01079067587852478, + 0.019593514502048492, + 0.02317730523645878, + 0.07163635641336441, + 0.005114153493195772, + 0.0050823381170630455, + 0.0053339083679020405, + 0.006400625221431255, + 0.007930977270007133, + 0.007577475160360336, + 0.007095071021467447, + 0.006991315167397261, + 0.001541787525638938, + 0.0026425973046571016, + 0.0018760348903015256, + 0.0019733780063688755, + 0.0021520827431231737, + 0.0033930635545402765, + 0.0021471267100423574, + 0.002584638074040413, + 0.003255546325817704, + 0.003227141685783863, + 0.0033572197426110506, + 0.004712017253041267, + 0.0020952390041202307, + 0.003380950540304184, + 0.0032555297948420048, + 0.003346650628373027, + 0.005387081298977137, + 0.005753283388912678, + 0.004898203536868095, + 0.004756995476782322, + 0.003829579334706068, + 0.00630978774279356, + 0.006880129221826792, + 0.00614089984446764, + 0.020243031904101372, + 0.02287064865231514, + 0.01707572117447853, + 0.012671498581767082, + 0.029196199029684067, + 0.019647883251309395, + 0.019380392506718636, + 0.02537326142191887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.01606813259422779 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up focuses on the hands of an elderly couple, their skin wrinkled with age and marked by time, as they collaborate with gentle precision. The man's broader, more veined hands firmly grasp the dark mahogany frame of a photograph, leaning against the warm cream-colored living room wall. The woman's fingers, more slender, their skin almost translucent, carefully adjust the alignment. They are straightening a slightly curved, glass-covered photograph that captures younger versions of themselves on a sunny beach. The entire action is bathed in the soft, warm glow of an unseen table lamp, which casts gentle shadows of their hands and creates soft reflections on the glass of the photograph. With a final, deliberate, and almost imperceptible nudge, the woman brings the frame into a perfectly level position. Her touch lingers for a moment as her hands secure it in place, creating a serene and affectionate atmosphere of quiet collaboration and shared history.", + "seed": 11340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_11/index_1256_seed_11340_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08507518768310547, + "vlm_logprob": -0.08507518768310547, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.08507518768310547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18935942649841309 + ], + "opposite_logprob_tokens": [ + -1.814359426498413 + ], + "logprob_reward": -0.18935942649841309, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18935942649841309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12432880699634552, + 0.18588723242282867, + 0.10242965072393417, + 0.03286775201559067, + 0.03833390027284622, + 0.10111555457115173, + 0.07037180662155151, + 0.05899472162127495, + 0.15697503089904785, + 0.12869545817375183 + ], + "credit_map": [ + 0.014244471676647663, + 0.015686675906181335, + 0.017519794404506683, + 0.011763595044612885, + 0.015721095725893974, + 0.01724962890148163, + 0.01800292357802391, + 0.014140643179416656, + 0.01748163066804409, + 0.022586898878216743, + 0.021875552833080292, + 0.019159654155373573, + 0.03030342608690262, + 0.02502564899623394, + 0.024916063994169235, + 0.024538373574614525, + 0.010628977790474892, + 0.008866464719176292, + 0.01787673681974411, + 0.011490777134895325, + 0.020554369315505028, + 0.014199926517903805, + 0.009039417840540409, + 0.009772994555532932, + 0.0037537426687777042, + 0.004062802996486425, + 0.003620855277404189, + 0.004500662907958031, + 0.004355430603027344, + 0.004627773072570562, + 0.003331440966576338, + 0.004615049809217453, + 0.0033754322212189436, + 0.002720460994169116, + 0.0048862178809940815, + 0.005116913001984358, + 0.004375786520540714, + 0.005855596158653498, + 0.007301114499568939, + 0.004702383186668158, + 0.012121574021875858, + 0.008601292967796326, + 0.014093861915171146, + 0.017372405156493187, + 0.011127233505249023, + 0.01298413798213005, + 0.013196627609431744, + 0.01161842979490757, + 0.006765352096408606, + 0.0061698914505541325, + 0.009458974935114384, + 0.011211449280381203, + 0.006486206315457821, + 0.008173945359885693, + 0.014384732581675053, + 0.007721260190010071, + 0.0069500720128417015, + 0.006229491904377937, + 0.007718533743172884, + 0.007302698213607073, + 0.0070999665185809135, + 0.011300369165837765, + 0.005810650996863842, + 0.0065829455852508545, + 0.022226696833968163, + 0.01679675653576851, + 0.027539240196347237, + 0.020154979079961777, + 0.030934179201722145, + 0.016453957185149193, + 0.010727147571742535, + 0.012142086401581764, + 0.017555372789502144, + 0.016019601374864578, + 0.015019964426755905, + 0.017639344558119774, + 0.015498648397624493, + 0.026617106050252914, + 0.011167226359248161, + 0.009178201667964458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.18935942649841309 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07922283560037613 + ], + "opposite_logprob_tokens": [ + -2.5792229175567627 + ], + "logprob_reward": -0.07922283560037613, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07922283560037613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7267153263092041, + 0.0550767183303833, + 0.058352287858724594, + 0.011795984581112862, + 0.02067788876593113, + 0.024289466440677643, + 0.032499559223651886, + 0.010539951734244823, + 0.02326934039592743, + 0.03678346052765846 + ], + "credit_map": [ + 0.26811644434928894, + 0.025311321020126343, + 0.030942708253860474, + 0.019663983955979347, + 0.022598231211304665, + 0.2746773064136505, + 0.01527135819196701, + 0.07013396918773651, + 0.0073134261183440685, + 0.007179344538599253, + 0.007112502586096525, + 0.005797660443931818, + 0.006986586377024651, + 0.0070035639218986034, + 0.005127700045704842, + 0.008555935695767403, + 0.0059180110692977905, + 0.006753948051482439, + 0.01198873482644558, + 0.00857342965900898, + 0.00628229696303606, + 0.008759768679738045, + 0.004774905741214752, + 0.005301191471517086, + 0.0011214609257876873, + 0.0010831396793946624, + 0.0021556762512773275, + 0.0021098635625094175, + 0.0019917679019272327, + 0.0013517893385142088, + 0.0008525810553692281, + 0.0011297045275568962, + 0.002252835314720869, + 0.0012357669183984399, + 0.0021664220839738846, + 0.0029301117174327374, + 0.00243173586204648, + 0.003857219126075506, + 0.002170989289879799, + 0.003632809966802597, + 0.0027434341609477997, + 0.0029667753260582685, + 0.003991697449237108, + 0.003245722968131304, + 0.0031450632959604263, + 0.0026169808115810156, + 0.003523886436596513, + 0.002055905759334564, + 0.0029858730267733335, + 0.00231776456348598, + 0.0037095951847732067, + 0.0036312260199338198, + 0.003607814898714423, + 0.004095698241144419, + 0.0076738763600587845, + 0.004477709531784058, + 0.001479133265092969, + 0.000946586427744478, + 0.0015448685735464096, + 0.0012876571854576468, + 0.0007962586241774261, + 0.0015643022488802671, + 0.0015452966326847672, + 0.001375847845338285, + 0.003155865240842104, + 0.0025832699611783028, + 0.003572516841813922, + 0.002979084849357605, + 0.004304640460759401, + 0.002711045788601041, + 0.0017716475995257497, + 0.0021912686061114073, + 0.005113412160426378, + 0.004855796229094267, + 0.0044056144542992115, + 0.004666659515351057, + 0.004881750326603651, + 0.006810421124100685, + 0.0028720563277602196, + 0.003177751088514924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07922283560037613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004105354659259319 + ], + "opposite_logprob_tokens": [ + -5.504105567932129 + ], + "logprob_reward": -0.004105354659259319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004105354659259319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6446241140365601, + 0.13007888197898865, + 0.033144447952508926, + 0.02419537492096424, + 0.01384752243757248, + 0.04159403219819069, + 0.0239738617092371, + 0.021964242681860924, + 0.03293171152472496, + 0.033645786345005035 + ], + "credit_map": [ + 0.24195335805416107, + 0.022504057735204697, + 0.025681916624307632, + 0.01637912541627884, + 0.021141979843378067, + 0.24220949411392212, + 0.015363648533821106, + 0.05939051881432533, + 0.01036910805851221, + 0.013080371543765068, + 0.015143808908760548, + 0.010631298646330833, + 0.020173942670226097, + 0.020809903740882874, + 0.022797424346208572, + 0.017073029652237892, + 0.0033778047654777765, + 0.003683054819703102, + 0.0047575668431818485, + 0.004354985896497965, + 0.003869309090077877, + 0.005101126153022051, + 0.003822626080363989, + 0.0041779763996601105, + 0.002649832982569933, + 0.0024194675497710705, + 0.0024296054616570473, + 0.003915085457265377, + 0.002994731767103076, + 0.002684440463781357, + 0.0037189077120274305, + 0.0033833065535873175, + 0.0017655333504080772, + 0.0010029890108853579, + 0.0017122523859143257, + 0.0022455358412116766, + 0.0019477506866678596, + 0.0019705910235643387, + 0.0016006738878786564, + 0.0016021968331187963, + 0.0038517299108207226, + 0.003109443699941039, + 0.005663345102220774, + 0.004703983198851347, + 0.0038181678391993046, + 0.009107065387070179, + 0.004705389495939016, + 0.006634911056607962, + 0.00254762708209455, + 0.0021618108730763197, + 0.002081743674352765, + 0.0023927106522023678, + 0.0033949878998100758, + 0.0038532970938831568, + 0.0030578875448554754, + 0.004483798053115606, + 0.0021545158233493567, + 0.0014892381150275469, + 0.0027815457433462143, + 0.0028077729512006044, + 0.0028245276771485806, + 0.0030630812980234623, + 0.0040956586599349976, + 0.002747903810814023, + 0.003960042260587215, + 0.0034031348768621683, + 0.00520247220993042, + 0.0037348330952227116, + 0.005970189347863197, + 0.0037175635807216167, + 0.0033673204015940428, + 0.003576158080250025, + 0.004566625691950321, + 0.00595249654725194, + 0.005083896219730377, + 0.0034806972835212946, + 0.004207632038742304, + 0.002784751821309328, + 0.003728266106918454, + 0.0038414227310568094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004105354659259319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14290183782577515 + ], + "opposite_logprob_tokens": [ + -2.01790189743042 + ], + "logprob_reward": -0.14290183782577515, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14290183782577515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1977631002664566, + 0.051682088524103165, + 0.05525955185294151, + 0.0399753712117672, + 0.051909253001213074, + 0.13209283351898193, + 0.044519003480672836, + 0.04029064625501633, + 0.08675581216812134, + 0.2997523248195648 + ], + "credit_map": [ + 0.05256875976920128, + 0.010237419046461582, + 0.008207574486732483, + 0.008055752143263817, + 0.008330161683261395, + 0.06898520141839981, + 0.01236038375645876, + 0.02901786006987095, + 0.00543264439329505, + 0.005986351054161787, + 0.006379354279488325, + 0.004857453051954508, + 0.0074516646564006805, + 0.007612594403326511, + 0.006740941666066647, + 0.007221090607345104, + 0.005525043234229088, + 0.005628523416817188, + 0.008526653051376343, + 0.0070360517129302025, + 0.007018052972853184, + 0.009418124333024025, + 0.005600128788501024, + 0.006506983190774918, + 0.004478361923247576, + 0.0046112388372421265, + 0.004343897569924593, + 0.006584009621292353, + 0.005637911148369312, + 0.004289884120225906, + 0.004933232441544533, + 0.005096836015582085, + 0.006924394052475691, + 0.004725719336420298, + 0.005879887845367193, + 0.009154132567346096, + 0.006729663349688053, + 0.007257120218127966, + 0.004629135597497225, + 0.006609200965613127, + 0.011999904178082943, + 0.011009115725755692, + 0.015142585150897503, + 0.0126462047919631, + 0.012050249613821507, + 0.030652625486254692, + 0.01602914370596409, + 0.022563010454177856, + 0.004612861666828394, + 0.003707030089572072, + 0.0051653701812028885, + 0.004236176144331694, + 0.007481461390852928, + 0.005898368079215288, + 0.006850474048405886, + 0.006567265372723341, + 0.0044484431855380535, + 0.0030731051228940487, + 0.003314479487016797, + 0.0044620949774980545, + 0.006870603188872337, + 0.006812413223087788, + 0.005659446120262146, + 0.005650063045322895, + 0.009692641906440258, + 0.011688996106386185, + 0.012301839888095856, + 0.011299531906843185, + 0.011519569903612137, + 0.010572019033133984, + 0.009068327955901623, + 0.01061288546770811, + 0.04327033832669258, + 0.0423911027610302, + 0.041605815291404724, + 0.03722461313009262, + 0.03823043778538704, + 0.045161228626966476, + 0.02183510549366474, + 0.03003368154168129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.14290183782577515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786483831703663 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786483831703663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786483831703663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17829552292823792, + 0.06246427819132805, + 0.07358798384666443, + 0.027632761746644974, + 0.04959087073802948, + 0.1459289789199829, + 0.09950269013643265, + 0.046108510345220566, + 0.06981468200683594, + 0.2470737248659134 + ], + "credit_map": [ + 0.05429495871067047, + 0.014673068188130856, + 0.015065737999975681, + 0.012167179025709629, + 0.013580960221588612, + 0.02313966117799282, + 0.013974489644169807, + 0.031399473547935486, + 0.005089539103209972, + 0.007026395294815302, + 0.0070693353191018105, + 0.00610206788405776, + 0.011775253340601921, + 0.008193809539079666, + 0.009173334576189518, + 0.008034544065594673, + 0.00731944153085351, + 0.009387468919157982, + 0.012199245393276215, + 0.007666097953915596, + 0.011128301732242107, + 0.009035950526595116, + 0.00843831431120634, + 0.008413165807723999, + 0.003109274199232459, + 0.003365241689607501, + 0.002949176821857691, + 0.0033334095496684313, + 0.003057665890082717, + 0.003916880115866661, + 0.004480383358895779, + 0.0034207291901111603, + 0.004171054810285568, + 0.003445969894528389, + 0.00550721725448966, + 0.0074736191891133785, + 0.004873018246144056, + 0.00712951784953475, + 0.009581833146512508, + 0.007408640813082457, + 0.016712673008441925, + 0.014963330700993538, + 0.024547772482037544, + 0.023375235497951508, + 0.017285676673054695, + 0.015378563664853573, + 0.017941704019904137, + 0.015724018216133118, + 0.009361126460134983, + 0.005880306474864483, + 0.008636020123958588, + 0.01050889864563942, + 0.0074214148335158825, + 0.011058907024562359, + 0.028443699702620506, + 0.018192321062088013, + 0.005063207354396582, + 0.0037881415337324142, + 0.005636344663798809, + 0.0053108492866158485, + 0.0056087118573486805, + 0.008703887462615967, + 0.005505585577338934, + 0.00649178447201848, + 0.008087953552603722, + 0.008974306285381317, + 0.01054118387401104, + 0.007925073616206646, + 0.012628672644495964, + 0.008262370713055134, + 0.006738133262842894, + 0.006656992714852095, + 0.038647621870040894, + 0.05195622146129608, + 0.040418241173028946, + 0.034560561180114746, + 0.02550816722214222, + 0.019243892282247543, + 0.01923861913383007, + 0.017500383779406548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.009786483831703663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up focuses on the hands of an elderly couple, their skin wrinkled with age and marked by time, as they collaborate with gentle precision. The man's broader, more veined hands firmly grasp the dark mahogany frame of a photograph, leaning against the warm cream-colored living room wall. The woman's fingers, more slender, their skin almost translucent, carefully adjust the alignment. They are straightening a slightly curved, glass-covered photograph that captures younger versions of themselves on a sunny beach. The entire action is bathed in the soft, warm glow of an unseen table lamp, which casts gentle shadows of their hands and creates soft reflections on the glass of the photograph. With a final, deliberate, and almost imperceptible nudge, the woman brings the frame into a perfectly level position. Her touch lingers for a moment as her hands secure it in place, creating a serene and affectionate atmosphere of quiet collaboration and shared history.", + "seed": 11341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_11/index_1256_seed_11341_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025420269183814525, + "vlm_logprob": -0.025420269183814525, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.025420269183814525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03731345757842064 + ], + "opposite_logprob_tokens": [ + -3.412313461303711 + ], + "logprob_reward": -0.03731345757842064, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03731345757842064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05264313146471977, + 0.05684855580329895, + 0.07762245833873749, + 0.1009347140789032, + 0.07658573985099792, + 0.16056987643241882, + 0.4012134373188019, + 0.0421142540872097, + 0.011853090487420559, + 0.01961476169526577 + ], + "credit_map": [ + 0.0046678869985044, + 0.007277204655110836, + 0.00518219405785203, + 0.005873546935617924, + 0.008658473379909992, + 0.007670875638723373, + 0.006528730969876051, + 0.006784211844205856, + 0.011578423902392387, + 0.006394030060619116, + 0.006406112108379602, + 0.005333200097084045, + 0.005573927890509367, + 0.007458127103745937, + 0.005125405266880989, + 0.00897932518273592, + 0.006750564556568861, + 0.0038565131835639477, + 0.008251630701124668, + 0.0063751619309186935, + 0.006393243558704853, + 0.016001451760530472, + 0.02383553795516491, + 0.006158356089144945, + 0.014932415448129177, + 0.007207260001450777, + 0.00570262037217617, + 0.003123563015833497, + 0.00952162779867649, + 0.010562111623585224, + 0.03112834133207798, + 0.018756769597530365, + 0.006804381962865591, + 0.012171974405646324, + 0.013562576845288277, + 0.007519105914980173, + 0.008829145692288876, + 0.00970553234219551, + 0.00850705336779356, + 0.009485972113907337, + 0.01578492671251297, + 0.026568232104182243, + 0.029586730524897575, + 0.018109172582626343, + 0.013232533819973469, + 0.015029879286885262, + 0.01997760683298111, + 0.02228078432381153, + 0.047822605818510056, + 0.059516072273254395, + 0.10808946192264557, + 0.050869185477495193, + 0.02359933964908123, + 0.03096391074359417, + 0.0400502048432827, + 0.040302637964487076, + 0.004860216286033392, + 0.006773621309548616, + 0.008548303507268429, + 0.0048797763884067535, + 0.0034354638773947954, + 0.004830400925129652, + 0.0044607920572161674, + 0.0043256767094135284, + 0.0014154347591102123, + 0.001774073694832623, + 0.001840936136431992, + 0.0014920267276465893, + 0.001779619138687849, + 0.0012583937495946884, + 0.001122067915275693, + 0.0011705366196110845, + 0.0023942876141518354, + 0.003261803649365902, + 0.0023878482170403004, + 0.0025567933917045593, + 0.0033133425749838352, + 0.0018248700071126223, + 0.0019191449973732233, + 0.001956669148057699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.03731345757842064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029846802353858948 + ], + "opposite_logprob_tokens": [ + -3.5298469066619873 + ], + "logprob_reward": -0.029846802353858948, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029846802353858948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08807510882616043, + 0.034693293273448944, + 0.16937875747680664, + 0.13682028651237488, + 0.0774649828672409, + 0.0992424488067627, + 0.2221199870109558, + 0.11476100981235504, + 0.014724288135766983, + 0.04271983355283737 + ], + "credit_map": [ + 0.009510377421975136, + 0.009916617535054684, + 0.008589350618422031, + 0.010628558695316315, + 0.01063302531838417, + 0.010926964692771435, + 0.015009415335953236, + 0.012860801070928574, + 0.0037158236373215914, + 0.0038754167035222054, + 0.004487670958042145, + 0.004457044415175915, + 0.004658996127545834, + 0.004268518183380365, + 0.0035722218453884125, + 0.005657602567225695, + 0.011944996193051338, + 0.008152246475219727, + 0.020211748778820038, + 0.008242574520409107, + 0.022357158362865448, + 0.03307805210351944, + 0.049969494342803955, + 0.01542248111218214, + 0.014924506656825542, + 0.013417699374258518, + 0.013635138981044292, + 0.006272841710597277, + 0.013271442614495754, + 0.01764208823442459, + 0.031020361930131912, + 0.02663620561361313, + 0.007397803943604231, + 0.012688802555203438, + 0.01343430858105421, + 0.008628150448203087, + 0.006141774356365204, + 0.0087808296084404, + 0.008942765183746815, + 0.011450545862317085, + 0.008610915392637253, + 0.017893610522150993, + 0.01808321848511696, + 0.008717094548046589, + 0.011313160881400108, + 0.008189571090042591, + 0.014213568530976772, + 0.01222130935639143, + 0.025562360882759094, + 0.0325118713080883, + 0.05685774236917496, + 0.02846909873187542, + 0.01625857688486576, + 0.0161849744617939, + 0.022372441366314888, + 0.023902904242277145, + 0.013684534467756748, + 0.01738719828426838, + 0.026790067553520203, + 0.014571072533726692, + 0.010217804461717606, + 0.011400294490158558, + 0.01110068429261446, + 0.00960935652256012, + 0.0014257915318012238, + 0.002135068876668811, + 0.0021149534732103348, + 0.0021714854519814253, + 0.0018696256447583437, + 0.0015258912462741137, + 0.0015203654766082764, + 0.001961106900125742, + 0.00540428189560771, + 0.00665824580937624, + 0.0064615728333592415, + 0.007279275916516781, + 0.005419113207608461, + 0.0029337818268686533, + 0.004459391813725233, + 0.004104169551283121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.029846802353858948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763415452092886 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.00763415452092886, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763415452092886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17868953943252563, + 0.052587706595659256, + 0.14504119753837585, + 0.17205829918384552, + 0.08724012970924377, + 0.05940648913383484, + 0.07007177174091339, + 0.15303842723369598, + 0.031639646738767624, + 0.050226785242557526 + ], + "credit_map": [ + 0.015788761898875237, + 0.016707152128219604, + 0.018371719866991043, + 0.022077443078160286, + 0.019569382071495056, + 0.027078190818428993, + 0.030516618862748146, + 0.028580280020833015, + 0.012482579797506332, + 0.004371171817183495, + 0.0066694291308522224, + 0.004323097877204418, + 0.007442596834152937, + 0.008142436854541302, + 0.004286817740648985, + 0.0048695835284888744, + 0.013460424728691578, + 0.007925826124846935, + 0.016121188178658485, + 0.00683137122541666, + 0.020170213654637337, + 0.02341201715171337, + 0.04323498532176018, + 0.013885175809264183, + 0.022602124139666557, + 0.015488555654883385, + 0.01353940088301897, + 0.01306769996881485, + 0.012495107017457485, + 0.026533257216215134, + 0.039251815527677536, + 0.029080351814627647, + 0.00717140082269907, + 0.015256432816386223, + 0.013508128002285957, + 0.00797650683671236, + 0.008832735940814018, + 0.01364020723849535, + 0.009231958538293839, + 0.01162276417016983, + 0.005004494916647673, + 0.007154236547648907, + 0.011112973093986511, + 0.006937370635569096, + 0.009555292315781116, + 0.007217905018478632, + 0.0057350145652890205, + 0.0066892122849822044, + 0.006594019941985607, + 0.009314321912825108, + 0.0112529955804348, + 0.008016773499548435, + 0.01117621548473835, + 0.006594356149435043, + 0.00910553615540266, + 0.008017562329769135, + 0.017695417627692223, + 0.01960141584277153, + 0.04329393059015274, + 0.01769639365375042, + 0.011079780757427216, + 0.013996938243508339, + 0.016769878566265106, + 0.012904676608741283, + 0.003445397363975644, + 0.0041419025510549545, + 0.0045356363989412785, + 0.00409994600340724, + 0.004359870683401823, + 0.0032565142028033733, + 0.004150666296482086, + 0.0036497178953140974, + 0.006081849802285433, + 0.006202531047165394, + 0.006808895617723465, + 0.006701597012579441, + 0.010061541572213173, + 0.0045013693161308765, + 0.0053612361662089825, + 0.004507767502218485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.00763415452092886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811221942305565 + ], + "opposite_logprob_tokens": [ + -3.288112163543701 + ], + "logprob_reward": -0.03811221942305565, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811221942305565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1195947527885437, + 0.03908821567893028, + 0.04283134639263153, + 0.14904938638210297, + 0.054845646023750305, + 0.10265974700450897, + 0.11444319039583206, + 0.24167440831661224, + 0.05086618661880493, + 0.08494716137647629 + ], + "credit_map": [ + 0.012674414552748203, + 0.013429339975118637, + 0.014262445271015167, + 0.013161622919142246, + 0.010762703604996204, + 0.017605755478143692, + 0.01619313657283783, + 0.021505335345864296, + 0.008849372155964375, + 0.0031481655314564705, + 0.005765013862401247, + 0.003059056354686618, + 0.0047716181725263596, + 0.005573604255914688, + 0.0034667651634663343, + 0.00445461506024003, + 0.004233543761074543, + 0.005026704631745815, + 0.0051004826091229916, + 0.0036344206891953945, + 0.005543550942093134, + 0.007086634635925293, + 0.007233098149299622, + 0.00497290538623929, + 0.01351683959364891, + 0.017681101337075233, + 0.020360659807920456, + 0.015819650143384933, + 0.01379356812685728, + 0.022859524935483932, + 0.023374978452920914, + 0.021643051877617836, + 0.004765264689922333, + 0.005867231171578169, + 0.007653178181499243, + 0.0049999481998384, + 0.012432277202606201, + 0.005684474483132362, + 0.006206904072314501, + 0.007236359640955925, + 0.007664799690246582, + 0.013025596737861633, + 0.010797275230288506, + 0.011562095023691654, + 0.01924283802509308, + 0.016038354486227036, + 0.011150619015097618, + 0.01317816786468029, + 0.012868317775428295, + 0.01510263979434967, + 0.024288829416036606, + 0.008528191596269608, + 0.022456280887126923, + 0.011067425832152367, + 0.00929160788655281, + 0.010839898139238358, + 0.025184443220496178, + 0.029924720525741577, + 0.05911310762166977, + 0.025351770222187042, + 0.0270604956895113, + 0.025070223957300186, + 0.024528058245778084, + 0.02544158138334751, + 0.005862042773514986, + 0.005111847072839737, + 0.011135296896100044, + 0.0060609374195337296, + 0.007512805517762899, + 0.006049012765288353, + 0.004122420214116573, + 0.005011819768697023, + 0.010630584321916103, + 0.008043704554438591, + 0.009857891127467155, + 0.011584077030420303, + 0.01699548028409481, + 0.008129158057272434, + 0.008404344320297241, + 0.01130190771073103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.03811221942305565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194712042808533 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194712042808533, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194712042808533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21158652007579803, + 0.03679410740733147, + 0.041864022612571716, + 0.21836945414543152, + 0.23686009645462036, + 0.09842023253440857, + 0.08227868378162384, + 0.025691213086247444, + 0.021696042269468307, + 0.026439674198627472 + ], + "credit_map": [ + 0.018139775842428207, + 0.01967451721429825, + 0.020965907722711563, + 0.02470889501273632, + 0.020217152312397957, + 0.029729951173067093, + 0.04668514430522919, + 0.031465157866477966, + 0.0062532261945307255, + 0.004298689775168896, + 0.005039437208324671, + 0.003438820131123066, + 0.0033086801413446665, + 0.005424375645816326, + 0.003844098187983036, + 0.005186780355870724, + 0.004463223274797201, + 0.003367951838299632, + 0.005838736891746521, + 0.003175384597852826, + 0.007283519022166729, + 0.0065648420713841915, + 0.005282112862914801, + 0.0058882515877485275, + 0.024008674547076225, + 0.018185244873166084, + 0.020928341895341873, + 0.015159565955400467, + 0.025798389688134193, + 0.034741517156362534, + 0.044126272201538086, + 0.0354214645922184, + 0.01984858699142933, + 0.03009006381034851, + 0.031808093190193176, + 0.017241500318050385, + 0.038378190249204636, + 0.0347696878015995, + 0.02703028731048107, + 0.03769366815686226, + 0.00847596861422062, + 0.011142589151859283, + 0.011645017191767693, + 0.007565186358988285, + 0.012490764260292053, + 0.014315061271190643, + 0.013458214700222015, + 0.019327426329255104, + 0.00907735712826252, + 0.011725181713700294, + 0.019757188856601715, + 0.009714351035654545, + 0.008552497252821922, + 0.00640002591535449, + 0.008762660436332226, + 0.008289428427815437, + 0.00259074242785573, + 0.002681503538042307, + 0.004417257383465767, + 0.003416319377720356, + 0.004226074554026127, + 0.0035271099768579006, + 0.002233020728453994, + 0.0025991853326559067, + 0.0022519512567669153, + 0.002788111800327897, + 0.004145617131143808, + 0.0029418610502034426, + 0.003167801769450307, + 0.0020922317635267973, + 0.0020934243220835924, + 0.0022150427103042603, + 0.0035587807651609182, + 0.003788346191868186, + 0.004219140857458115, + 0.004714946262538433, + 0.0032100381795316935, + 0.0015500871231779456, + 0.002657259115949273, + 0.0027410760521888733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.014194712042808533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up focuses on the hands of an elderly couple, their skin wrinkled with age and marked by time, as they collaborate with gentle precision. The man's broader, more veined hands firmly grasp the dark mahogany frame of a photograph, leaning against the warm cream-colored living room wall. The woman's fingers, more slender, their skin almost translucent, carefully adjust the alignment. They are straightening a slightly curved, glass-covered photograph that captures younger versions of themselves on a sunny beach. The entire action is bathed in the soft, warm glow of an unseen table lamp, which casts gentle shadows of their hands and creates soft reflections on the glass of the photograph. With a final, deliberate, and almost imperceptible nudge, the woman brings the frame into a perfectly level position. Her touch lingers for a moment as her hands secure it in place, creating a serene and affectionate atmosphere of quiet collaboration and shared history.", + "seed": 11342, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_11/index_1256_seed_11342_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04956037411466241, + "vlm_logprob": -0.04956037411466241, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04956037411466241, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06466630846261978 + ], + "opposite_logprob_tokens": [ + -2.939666271209717 + ], + "logprob_reward": -0.06466630846261978, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06466630846261978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14508354663848877, + 0.09779657423496246, + 0.06968919187784195, + 0.07158224284648895, + 0.09737297147512436, + 0.20704986155033112, + 0.07033514976501465, + 0.04941052198410034, + 0.13903707265853882, + 0.052642907947301865 + ], + "credit_map": [ + 0.020648540928959846, + 0.023920366540551186, + 0.02301502600312233, + 0.013344712555408478, + 0.015049749985337257, + 0.010379061102867126, + 0.01710498332977295, + 0.021621098741889, + 0.013114448636770248, + 0.01335426140576601, + 0.010808886960148811, + 0.008744565770030022, + 0.013423257507383823, + 0.009489140473306179, + 0.013388166204094887, + 0.015473848208785057, + 0.00827833916991949, + 0.0070785582065582275, + 0.006843043491244316, + 0.004834577441215515, + 0.01010075118392706, + 0.009311959147453308, + 0.01189704891294241, + 0.011344914324581623, + 0.008557380177080631, + 0.007228657603263855, + 0.00828025583177805, + 0.007422452326864004, + 0.009195956401526928, + 0.013500688597559929, + 0.0070147523656487465, + 0.010382098145782948, + 0.01207045279443264, + 0.017150528728961945, + 0.014461984857916832, + 0.008116824552416801, + 0.010982360690832138, + 0.013497334904968739, + 0.00930594652891159, + 0.011787546798586845, + 0.022970004007220268, + 0.02254774421453476, + 0.036205001175403595, + 0.017466679215431213, + 0.022232646122574806, + 0.04000292345881462, + 0.022688064724206924, + 0.022936806082725525, + 0.007348323240876198, + 0.00815992895513773, + 0.013399748131632805, + 0.006751680746674538, + 0.008197764866054058, + 0.010142349638044834, + 0.007934512570500374, + 0.008400838822126389, + 0.00549470866099, + 0.006137330085039139, + 0.00685583520680666, + 0.004339184612035751, + 0.006394097115844488, + 0.007812725380063057, + 0.006265000440180302, + 0.0061116390861570835, + 0.016554372385144234, + 0.013745673932135105, + 0.025590021163225174, + 0.01418826263397932, + 0.01636158674955368, + 0.02102876827120781, + 0.013198381289839745, + 0.0183700080960989, + 0.006073875352740288, + 0.008412465453147888, + 0.006081030238419771, + 0.008885973133146763, + 0.005969217047095299, + 0.0038269164506345987, + 0.0075329882092773914, + 0.005860441830009222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.06466630846261978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381888709962368 + ], + "opposite_logprob_tokens": [ + -3.288188934326172 + ], + "logprob_reward": -0.0381888709962368, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381888709962368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10106981545686722, + 0.04123463109135628, + 0.06530874222517014, + 0.036912113428115845, + 0.24465230107307434, + 0.10151056200265884, + 0.19148071110248566, + 0.1098390445113182, + 0.02138439379632473, + 0.08660771697759628 + ], + "credit_map": [ + 0.014991356059908867, + 0.011634284630417824, + 0.011430896818637848, + 0.012201409786939621, + 0.007720198482275009, + 0.00719434441998601, + 0.017042523249983788, + 0.01885480433702469, + 0.00601580087095499, + 0.010371256619691849, + 0.006245485506951809, + 0.003797802608460188, + 0.003441526321694255, + 0.0031042122282087803, + 0.0038000603672116995, + 0.0044584861025214195, + 0.009210724383592606, + 0.007041746750473976, + 0.007999689318239689, + 0.003382527967914939, + 0.009388094767928123, + 0.013031339272856712, + 0.00573904300108552, + 0.009515575133264065, + 0.004509106744080782, + 0.00438687251880765, + 0.005698759574443102, + 0.003425623755902052, + 0.004195709712803364, + 0.006051010452210903, + 0.0038040250074118376, + 0.004841003566980362, + 0.02982204221189022, + 0.029661450535058975, + 0.04707910865545273, + 0.017804650589823723, + 0.024563465267419815, + 0.03695197030901909, + 0.031997039914131165, + 0.026772577315568924, + 0.011884558014571667, + 0.013268491253256798, + 0.019362816587090492, + 0.008776679635047913, + 0.012376880273222923, + 0.014460385777056217, + 0.009704580530524254, + 0.011676169000566006, + 0.0203980915248394, + 0.020338529720902443, + 0.03747332841157913, + 0.015811339020729065, + 0.021872268989682198, + 0.03482653200626373, + 0.019875463098287582, + 0.020885158330202103, + 0.011842139065265656, + 0.011611075140535831, + 0.024864783510565758, + 0.010100296698510647, + 0.008123457431793213, + 0.013599992729723454, + 0.01603025197982788, + 0.013667041435837746, + 0.002464011311531067, + 0.0029100789688527584, + 0.0038810227997601032, + 0.002270733006298542, + 0.0025039673782885075, + 0.002329906215891242, + 0.002515679458156228, + 0.002508994657546282, + 0.00942972395569086, + 0.011293315328657627, + 0.01627775840461254, + 0.012692420743405819, + 0.008299560286104679, + 0.006883916445076466, + 0.009606418199837208, + 0.012124601751565933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.0381888709962368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744833197444677 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.006744833197444677, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744833197444677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13432380557060242, + 0.09644359350204468, + 0.09482887387275696, + 0.032649364322423935, + 0.09022769331932068, + 0.15389618277549744, + 0.13945990800857544, + 0.11672880500555038, + 0.03825123980641365, + 0.10319055616855621 + ], + "credit_map": [ + 0.018677130341529846, + 0.01848714053630829, + 0.018351811915636063, + 0.01658451184630394, + 0.010705784894526005, + 0.007817471399903297, + 0.014652365818619728, + 0.02904757671058178, + 0.013785158284008503, + 0.015608842484652996, + 0.009429524652659893, + 0.007704778108745813, + 0.015839967876672745, + 0.009900626726448536, + 0.009977483190596104, + 0.014197217300534248, + 0.01429325994104147, + 0.009866557084023952, + 0.009591605514287949, + 0.006400639656931162, + 0.02034548856317997, + 0.01501272153109312, + 0.008427723310887814, + 0.01089087687432766, + 0.003939885646104813, + 0.00480819633230567, + 0.003476879093796015, + 0.0031841641757637262, + 0.006676855031400919, + 0.004172708839178085, + 0.0027350489981472492, + 0.003655626205727458, + 0.010281589813530445, + 0.012547947466373444, + 0.014409392140805721, + 0.0075378501787781715, + 0.010975011624395847, + 0.013690226711332798, + 0.00910460576415062, + 0.011681070551276207, + 0.01818249747157097, + 0.01682337000966072, + 0.01990014687180519, + 0.012595819309353828, + 0.019739868119359016, + 0.03338165953755379, + 0.014615034684538841, + 0.018657784909009933, + 0.015245219692587852, + 0.0136513477191329, + 0.016914337873458862, + 0.01329116616398096, + 0.03183944895863533, + 0.01896924152970314, + 0.014778637327253819, + 0.014770506881177425, + 0.015770018100738525, + 0.015851425006985664, + 0.015131493099033833, + 0.010150204412639141, + 0.01582217961549759, + 0.014850837178528309, + 0.010513583198189735, + 0.01863907463848591, + 0.004413046408444643, + 0.00408231420442462, + 0.005717720370739698, + 0.004000592045485973, + 0.005525908898562193, + 0.006613681558519602, + 0.0032465162221342325, + 0.004651461262255907, + 0.011078407987952232, + 0.009355065412819386, + 0.014249159954488277, + 0.012472708709537983, + 0.010574162006378174, + 0.009949695318937302, + 0.01661222241818905, + 0.018899139016866684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.006744833197444677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12712228298187256 + ], + "opposite_logprob_tokens": [ + -2.127122402191162 + ], + "logprob_reward": -0.12712228298187256, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12712228298187256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19000892341136932, + 0.17349766194820404, + 0.07843916863203049, + 0.0293114110827446, + 0.10462994873523712, + 0.08972299098968506, + 0.11170613020658493, + 0.13293705880641937, + 0.041427478194236755, + 0.04831928759813309 + ], + "credit_map": [ + 0.02519157901406288, + 0.01900487206876278, + 0.01952478103339672, + 0.0198135357350111, + 0.03208067640662193, + 0.016274509951472282, + 0.02394132688641548, + 0.034177616238594055, + 0.022803710773587227, + 0.019365394487977028, + 0.017137503251433372, + 0.013725579716265202, + 0.034910477697849274, + 0.0224749818444252, + 0.017664460465312004, + 0.025415539741516113, + 0.010372026823461056, + 0.009982755407691002, + 0.007037753239274025, + 0.006415797397494316, + 0.011769674718379974, + 0.013560638763010502, + 0.007547254208475351, + 0.011753259226679802, + 0.003809253452345729, + 0.0036364421248435974, + 0.002657730830833316, + 0.0027630538679659367, + 0.003577020950615406, + 0.0050153592601418495, + 0.003384942188858986, + 0.004467607010155916, + 0.01182549074292183, + 0.01437173131853342, + 0.01805420219898224, + 0.008277165703475475, + 0.012325185351073742, + 0.01668969914317131, + 0.011353454552590847, + 0.011733001098036766, + 0.010267840698361397, + 0.011136725544929504, + 0.011183052323758602, + 0.0071312072686851025, + 0.00790625624358654, + 0.018892675638198853, + 0.01038764975965023, + 0.01281757466495037, + 0.012029669247567654, + 0.01232369989156723, + 0.02236996404826641, + 0.008145686239004135, + 0.014857056550681591, + 0.01874975673854351, + 0.011611822992563248, + 0.011618461459875107, + 0.014033442363142967, + 0.012251723557710648, + 0.02690253034234047, + 0.012484405189752579, + 0.010988577269017696, + 0.018520254641771317, + 0.021282613277435303, + 0.016473501920700073, + 0.0049439589492976665, + 0.0045587229542434216, + 0.008863919414579868, + 0.003929168451577425, + 0.00518969539552927, + 0.00593532994389534, + 0.004016768652945757, + 0.003989913035184145, + 0.004940730053931475, + 0.004689901135861874, + 0.006205352023243904, + 0.008097962476313114, + 0.006425455678254366, + 0.006011193618178368, + 0.005810113158077002, + 0.006138576194643974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.12712228298187256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of an elderly couple visible in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hand broader and more veined than the woman's hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark mahogany photograph frame being held against a warm cream-colored wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the photograph inside the frame show a younger couple on a sunny beach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft, warm table lamp casting shadows and reflections on the glass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079574935138226 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079574935138226, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079574935138226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06524892151355743, + 0.032007213681936264, + 0.021188735961914062, + 0.035148173570632935, + 0.07585608959197998, + 0.33684635162353516, + 0.17186206579208374, + 0.17704029381275177, + 0.028361286967992783, + 0.056440841406583786 + ], + "credit_map": [ + 0.008825048804283142, + 0.0073997038416564465, + 0.007005474995821714, + 0.006769777741283178, + 0.009110643528401852, + 0.006593522150069475, + 0.0068349773064255714, + 0.012709776870906353, + 0.004484809003770351, + 0.004383191000670195, + 0.004050679970532656, + 0.002628678921610117, + 0.0034239301458001137, + 0.0034224819391965866, + 0.0033423644490540028, + 0.006271078251302242, + 0.0028477346058934927, + 0.001815388910472393, + 0.0021766540594398975, + 0.001272697001695633, + 0.003851249348372221, + 0.0023197524715214968, + 0.0024328073486685753, + 0.004472453612834215, + 0.004257545340806246, + 0.0037171950098127127, + 0.003952270373702049, + 0.003927029203623533, + 0.004746655933558941, + 0.006307072471827269, + 0.0035705610644072294, + 0.004669849295169115, + 0.008762563578784466, + 0.01130530796945095, + 0.010035562328994274, + 0.004672050941735506, + 0.012553385458886623, + 0.013074294663965702, + 0.006731582805514336, + 0.00872134417295456, + 0.04045978561043739, + 0.038499172776937485, + 0.054252222180366516, + 0.03063814714550972, + 0.03284725546836853, + 0.06432058662176132, + 0.03866918385028839, + 0.037160031497478485, + 0.018640883266925812, + 0.018689515069127083, + 0.03101549856364727, + 0.013614761643111706, + 0.02672232687473297, + 0.02881116233766079, + 0.017210332676768303, + 0.017157597467303276, + 0.02162206918001175, + 0.022817984223365784, + 0.022928187623620033, + 0.017244208604097366, + 0.029056183993816376, + 0.030275244265794754, + 0.017703302204608917, + 0.015393132343888283, + 0.0031770123168826103, + 0.0033716605976223946, + 0.003948701545596123, + 0.002828247146680951, + 0.003569729859009385, + 0.00539221940562129, + 0.002867692383006215, + 0.0032060269732028246, + 0.0058713024482131, + 0.006562972441315651, + 0.010966833680868149, + 0.008194389753043652, + 0.004208455793559551, + 0.004766662605106831, + 0.0060907830484211445, + 0.009779444895684719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.011079574935138226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_12/index_tensor([677, 677])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_12/index_tensor([677, 677])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ae1c5eef01f54081667158cff43caf2e65585077 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_12/index_tensor([677, 677])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle close-up focuses on a heavy, matte black cast iron pot on the grate of a modern gas stove. Inside, a rich, crimson sauce simmers gently, its deep, velvety red surface shimmering in the warm, soft light. Tiny flecks of dark green herbs are suspended in the viscous liquid, while tiny, lazy bubbles steadily rise from the depths, puffing at the surface before bursting with a soft \"plop\" and delicate, concentric ripples, releasing a delicate plume of steam. The pot sits at the center of a sleek black glass cooktop, revealing a low, controlled ring of vibrant blue and orange flames providing steady heat. The entire scene is bathed in a warm, inviting glow from an unseen overhead source, creating soft highlights on the pot's curved sides and the shimmering sauce, establishing a cozy, focused, and deeply comforting atmosphere of patient home cooking.", + "seed": 16515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_12/index_677_seed_16515_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04663454139372334, + "vlm_logprob": -0.04663454139372334, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.04663454139372334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01060174684971571 + ], + "opposite_logprob_tokens": [ + -5.38560152053833 + ], + "logprob_reward": -0.01060174684971571, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01060174684971571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20967721939086914, + 0.07142898440361023, + 0.03759896382689476, + 0.02608110010623932, + 0.09235958755016327, + 0.1026429608464241, + 0.12217473238706589, + 0.21778029203414917, + 0.02593468315899372, + 0.09432143718004227 + ], + "credit_map": [ + 0.008731717243790627, + 0.014087237417697906, + 0.05502423644065857, + 0.02209859900176525, + 0.06224004924297333, + 0.02369651384651661, + 0.012026607990264893, + 0.011772263795137405, + 0.006710379384458065, + 0.010896001011133194, + 0.009600291959941387, + 0.01143614761531353, + 0.008421932347118855, + 0.010324576869606972, + 0.00645295437425375, + 0.00758670549839735, + 0.0026553296484053135, + 0.004084519576281309, + 0.005679577123373747, + 0.006569810211658478, + 0.006792865693569183, + 0.004730867687612772, + 0.0032226641196757555, + 0.003863331861793995, + 0.0019197543151676655, + 0.0031156959012150764, + 0.005026529543101788, + 0.003114410676062107, + 0.005070281680673361, + 0.003089913632720709, + 0.002515241736546159, + 0.0022292733192443848, + 0.00551367225125432, + 0.009848724119365215, + 0.01462581567466259, + 0.024037908762693405, + 0.011086818762123585, + 0.009440956637263298, + 0.009886891581118107, + 0.007918802089989185, + 0.006608931347727776, + 0.008131689392030239, + 0.01659286767244339, + 0.019879909232258797, + 0.01969553902745247, + 0.017979251220822334, + 0.0064454753883183, + 0.007309315726161003, + 0.008854656480252743, + 0.017085092142224312, + 0.014215693809092045, + 0.022009260952472687, + 0.015357195399701595, + 0.01137317530810833, + 0.023562172427773476, + 0.009717494249343872, + 0.013426193036139011, + 0.01082876417785883, + 0.0568745881319046, + 0.015011687763035297, + 0.05683637037873268, + 0.017000898718833923, + 0.0269180741161108, + 0.020883731544017792, + 0.0023088976740837097, + 0.002704022452235222, + 0.004564449656754732, + 0.0033063767477869987, + 0.003304550424218178, + 0.0026146636810153723, + 0.004521006718277931, + 0.0026107158046215773, + 0.00910594966262579, + 0.01298384927213192, + 0.013811400160193443, + 0.01048271544277668, + 0.009710930287837982, + 0.012241055257618427, + 0.017637040466070175, + 0.008348507806658745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01060174684971571 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158319175243378 + ], + "opposite_logprob_tokens": [ + -1.7015831470489502 + ], + "logprob_reward": -0.20158319175243378, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158319175243378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23413614928722382, + 0.1965600848197937, + 0.11756017059087753, + 0.025856073945760727, + 0.047586601227521896, + 0.18052823841571808, + 0.0671369731426239, + 0.03871256858110428, + 0.03702636435627937, + 0.05489683151245117 + ], + "credit_map": [ + 0.010430004447698593, + 0.015789514407515526, + 0.0707661360502243, + 0.01427155639976263, + 0.07722854614257812, + 0.023773539811372757, + 0.01049545593559742, + 0.011381352320313454, + 0.009009811095893383, + 0.011799106374382973, + 0.05770484730601311, + 0.009031815454363823, + 0.07174186408519745, + 0.018668223172426224, + 0.00929839164018631, + 0.009306011721491814, + 0.006751514971256256, + 0.009033061563968658, + 0.02414781041443348, + 0.015051140449941158, + 0.02552998624742031, + 0.01296916976571083, + 0.011736679822206497, + 0.012340802699327469, + 0.0019359783036634326, + 0.00264187715947628, + 0.004692857153713703, + 0.0033045499585568905, + 0.005111741833388805, + 0.0033756685443222523, + 0.0019075404852628708, + 0.0028858580626547337, + 0.0025096347089856863, + 0.004819272551685572, + 0.007957611232995987, + 0.013223947957158089, + 0.005389523692429066, + 0.0045937891118228436, + 0.00553203746676445, + 0.0035607796162366867, + 0.00993639137595892, + 0.016905732452869415, + 0.03441369906067848, + 0.017147302627563477, + 0.04195239022374153, + 0.02142483741044998, + 0.02020883373916149, + 0.018539050593972206, + 0.0053115542978048325, + 0.009653707966208458, + 0.007112666964530945, + 0.014621531590819359, + 0.006076832301914692, + 0.0066198743879795074, + 0.012228221632540226, + 0.005512582138180733, + 0.0030798297375440598, + 0.0033570597879588604, + 0.004844685085117817, + 0.005222449079155922, + 0.009031744673848152, + 0.00471718143671751, + 0.004785060882568359, + 0.0036745513789355755, + 0.0018551074899733067, + 0.0022816394921392202, + 0.005921841133385897, + 0.0050314017571508884, + 0.0062461900524795055, + 0.0034523371141403913, + 0.008605655282735825, + 0.003632188308984041, + 0.006113417912274599, + 0.006170867010951042, + 0.00872485525906086, + 0.005397993139922619, + 0.00870551262050867, + 0.006833144463598728, + 0.007871124893426895, + 0.005079909693449736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.20158319175243378 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008185135084204376 + ], + "opposite_logprob_tokens": [ + -7.125818729400635 + ], + "logprob_reward": -0.0008185135084204376, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008185135084204376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1106235459446907, + 0.2011183649301529, + 0.059286583214998245, + 0.05715884640812874, + 0.047425489872694016, + 0.11727391928434372, + 0.053124912083148956, + 0.033198509365320206, + 0.03630960360169411, + 0.2844802141189575 + ], + "credit_map": [ + 0.005691861268132925, + 0.00852906797081232, + 0.027860300615429878, + 0.007209664676338434, + 0.029435064643621445, + 0.014103948138654232, + 0.008819782175123692, + 0.008973854593932629, + 0.009656377136707306, + 0.014193848706781864, + 0.05453522503376007, + 0.01541898213326931, + 0.06606857478618622, + 0.018329020589590073, + 0.01270644087344408, + 0.010209903120994568, + 0.0032126745209097862, + 0.004464798141270876, + 0.015108340419828892, + 0.0061236838810145855, + 0.016330331563949585, + 0.0055491626262664795, + 0.005189831834286451, + 0.003307763021439314, + 0.003389226272702217, + 0.006059584673494101, + 0.008546053431928158, + 0.008118604309856892, + 0.011276738718152046, + 0.007755652070045471, + 0.005560081452131271, + 0.006452908739447594, + 0.0026285366620868444, + 0.006336937192827463, + 0.005070867482572794, + 0.011006684973835945, + 0.007148814853280783, + 0.00443091569468379, + 0.006646131630986929, + 0.004156602080911398, + 0.009064152836799622, + 0.011601453647017479, + 0.02014981396496296, + 0.015122047625482082, + 0.02177773043513298, + 0.019856425002217293, + 0.008507492952048779, + 0.011194813996553421, + 0.004445409402251244, + 0.008509742096066475, + 0.005695979576557875, + 0.008887737058103085, + 0.005110166501253843, + 0.006598433014005423, + 0.008599820546805859, + 0.005277620628476143, + 0.002054699929431081, + 0.0022730487398803234, + 0.008223420940339565, + 0.002994823968037963, + 0.007420917507261038, + 0.003378442022949457, + 0.003919052891433239, + 0.00293410150334239, + 0.002109065419062972, + 0.005369184538722038, + 0.005344749893993139, + 0.005179916974157095, + 0.005783704109489918, + 0.0034097398165613413, + 0.006205510813742876, + 0.002907732268795371, + 0.023935941979289055, + 0.027050208300352097, + 0.0434034988284111, + 0.04604341462254524, + 0.05675320699810982, + 0.029538458213210106, + 0.02600107528269291, + 0.03175441548228264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.0008185135084204376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630605716258287 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630605716258287, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630605716258287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22860267758369446, + 0.03639262914657593, + 0.18729136884212494, + 0.08019392192363739, + 0.031170539557933807, + 0.2875143587589264, + 0.05496338754892349, + 0.027015186846256256, + 0.01848212443292141, + 0.04837380722165108 + ], + "credit_map": [ + 0.010793626308441162, + 0.015455156564712524, + 0.07170648127794266, + 0.015725424513220787, + 0.07410375773906708, + 0.02092299610376358, + 0.009771820157766342, + 0.010123400948941708, + 0.0023382429499179125, + 0.00399798946455121, + 0.004894357640296221, + 0.008024346083402634, + 0.005553435999900103, + 0.004220356233417988, + 0.004147632978856564, + 0.003216264070942998, + 0.00985872931778431, + 0.011800960637629032, + 0.05108054727315903, + 0.023574072867631912, + 0.05119166150689125, + 0.015652213245630264, + 0.012108588591217995, + 0.012024586088955402, + 0.00429296400398016, + 0.006369340233504772, + 0.021059755235910416, + 0.010031291283667088, + 0.02214478701353073, + 0.00712412316352129, + 0.0042727296240627766, + 0.004898928105831146, + 0.002173469401896, + 0.0029346465598791838, + 0.003871117951348424, + 0.007899203337728977, + 0.004430688917636871, + 0.003515898250043392, + 0.0035291395615786314, + 0.0028163744136691093, + 0.02119489759206772, + 0.02979384735226631, + 0.048485662788152695, + 0.03017016500234604, + 0.05489318072795868, + 0.04125731438398361, + 0.03228029981255531, + 0.029438983649015427, + 0.003332114778459072, + 0.007687314413487911, + 0.009676759131252766, + 0.004862587433308363, + 0.012247959151864052, + 0.005030001513659954, + 0.007962945848703384, + 0.004163696430623531, + 0.002570969983935356, + 0.00176832708530128, + 0.0034522642381489277, + 0.0035318508744239807, + 0.003738495521247387, + 0.003385427175089717, + 0.003893868997693062, + 0.004673979245126247, + 0.0012983462074771523, + 0.0011644810438156128, + 0.004413780756294727, + 0.002995231654495001, + 0.003442991990596056, + 0.0018740890081971884, + 0.0019507241668179631, + 0.0013424769276753068, + 0.004541653674095869, + 0.004570652265101671, + 0.007094613742083311, + 0.0060453834012150764, + 0.010276014916598797, + 0.00538560189306736, + 0.004064956214278936, + 0.006394922733306885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.007630605716258287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538649141788483 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538649141788483, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538649141788483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11432883888483047, + 0.10989540070295334, + 0.035642530769109726, + 0.03568675369024277, + 0.04578235372900963, + 0.4125523865222931, + 0.03850054368376732, + 0.08965235203504562, + 0.03754229098558426, + 0.08041664212942123 + ], + "credit_map": [ + 0.0058715189807116985, + 0.007173615507781506, + 0.030601976439356804, + 0.01205470971763134, + 0.027341946959495544, + 0.011603623628616333, + 0.008900045417249203, + 0.010781394317746162, + 0.006095811724662781, + 0.008168344385921955, + 0.029551608487963676, + 0.009869691915810108, + 0.033994436264038086, + 0.009811507537961006, + 0.006994341965764761, + 0.005409649573266506, + 0.0025764715392142534, + 0.0030976529233157635, + 0.0069570159539580345, + 0.006493085063993931, + 0.004782990552484989, + 0.004456258378922939, + 0.0033500087447464466, + 0.003929044120013714, + 0.0024200573097914457, + 0.0032830066047608852, + 0.007107023149728775, + 0.005185733549296856, + 0.006672241725027561, + 0.00429817708209157, + 0.0027429633773863316, + 0.003977547865360975, + 0.002588629722595215, + 0.00403781421482563, + 0.0066960835829377174, + 0.00925541017204523, + 0.008838677778840065, + 0.005214293021708727, + 0.004583762958645821, + 0.004567678086459637, + 0.022762788459658623, + 0.0329713337123394, + 0.07656296342611313, + 0.06303729861974716, + 0.09361326694488525, + 0.05540730059146881, + 0.03375039994716644, + 0.03444700688123703, + 0.0026963085401803255, + 0.0038434909656643867, + 0.004689395427703857, + 0.009635576978325844, + 0.00383869674988091, + 0.0038328978698700666, + 0.006670435890555382, + 0.003293741960078478, + 0.005988889839500189, + 0.005698224063962698, + 0.0205906480550766, + 0.007679186295717955, + 0.020209243521094322, + 0.007296653464436531, + 0.012412818148732185, + 0.009776681661605835, + 0.0024317791685462, + 0.002354498254135251, + 0.0075098383240401745, + 0.005739740561693907, + 0.006636425852775574, + 0.0036314616445451975, + 0.006227078381925821, + 0.0030114639084786177, + 0.00844623614102602, + 0.010276617482304573, + 0.011636953800916672, + 0.00779933575540781, + 0.011853165924549103, + 0.010503594763576984, + 0.01153129804879427, + 0.008369434624910355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.012538649141788483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up focuses on a heavy, matte black cast iron pot on the grate of a modern gas stove. Inside, a rich, crimson sauce simmers gently, its deep, velvety red surface shimmering in the warm, soft light. Tiny flecks of dark green herbs are suspended in the viscous liquid, while tiny, lazy bubbles steadily rise from the depths, puffing at the surface before bursting with a soft \"plop\" and delicate, concentric ripples, releasing a delicate plume of steam. The pot sits at the center of a sleek black glass cooktop, revealing a low, controlled ring of vibrant blue and orange flames providing steady heat. The entire scene is bathed in a warm, inviting glow from an unseen overhead source, creating soft highlights on the pot's curved sides and the shimmering sauce, establishing a cozy, focused, and deeply comforting atmosphere of patient home cooking.", + "seed": 16516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_12/index_677_seed_16516_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06688442721497267, + "vlm_logprob": -0.06688442721497267, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.06688442721497267, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011475862003862858 + ], + "opposite_logprob_tokens": [ + -5.886476039886475 + ], + "logprob_reward": -0.011475862003862858, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.011475862003862858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2037716656923294, + 0.09775137156248093, + 0.05410495772957802, + 0.08475186675786972, + 0.06570141762495041, + 0.14419206976890564, + 0.12077052146196365, + 0.06084413453936577, + 0.0951380506157875, + 0.07297395914793015 + ], + "credit_map": [ + 0.027991076931357384, + 0.0268396083265543, + 0.027366885915398598, + 0.030642148107290268, + 0.015975384041666985, + 0.027473831549286842, + 0.025199666619300842, + 0.022283054888248444, + 0.010352137498557568, + 0.011350716464221478, + 0.013262913562357426, + 0.01565936580300331, + 0.009475247003138065, + 0.009961470030248165, + 0.018045825883746147, + 0.009643704630434513, + 0.005947769153863192, + 0.00706918491050601, + 0.004926694557070732, + 0.013031616806983948, + 0.004874599166214466, + 0.005801299121230841, + 0.006621581502258778, + 0.00583221297711134, + 0.010713743977248669, + 0.007032637018710375, + 0.006247945129871368, + 0.03123057261109352, + 0.004511568695306778, + 0.006164164282381535, + 0.0109587786719203, + 0.007892453111708164, + 0.00756113464012742, + 0.00629421416670084, + 0.008028772659599781, + 0.009586543776094913, + 0.007113538682460785, + 0.008705166168510914, + 0.011037592776119709, + 0.007374456152319908, + 0.019478317350149155, + 0.009394442662596703, + 0.009621914476156235, + 0.05862319469451904, + 0.00595672894269228, + 0.009899001568555832, + 0.01893749088048935, + 0.012280977331101894, + 0.013957618735730648, + 0.013923399150371552, + 0.014553054235875607, + 0.020067855715751648, + 0.014397626742720604, + 0.01406481396406889, + 0.018930424004793167, + 0.010875729843974113, + 0.006780768278986216, + 0.0074124704115092754, + 0.005975846201181412, + 0.011525757610797882, + 0.006493196822702885, + 0.0068013956770300865, + 0.009972247295081615, + 0.0058824531733989716, + 0.013541590422391891, + 0.006688288412988186, + 0.008460763841867447, + 0.03228456899523735, + 0.006095076911151409, + 0.007891237735748291, + 0.012084858492016792, + 0.008091664873063564, + 0.007964391261339188, + 0.006585725117474794, + 0.006378075573593378, + 0.018247263506054878, + 0.0062949336133897305, + 0.007963541895151138, + 0.010580754838883877, + 0.008959274739027023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.011475862003862858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281338632106781 + ], + "opposite_logprob_tokens": [ + -1.4063386917114258 + ], + "logprob_reward": -0.281338632106781, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281338632106781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2013654112815857, + 0.09903278946876526, + 0.05761886388063431, + 0.11486293375492096, + 0.09053125977516174, + 0.15125398337841034, + 0.15238170325756073, + 0.05184660851955414, + 0.04000505059957504, + 0.041101373732089996 + ], + "credit_map": [ + 0.018185297027230263, + 0.025776084512472153, + 0.028525037690997124, + 0.02668771706521511, + 0.01728532649576664, + 0.032514188438653946, + 0.025547457858920097, + 0.02684430219233036, + 0.012506695464253426, + 0.011949852108955383, + 0.012396696954965591, + 0.014770377427339554, + 0.010000892914831638, + 0.010898256674408913, + 0.014501244761049747, + 0.01200877595692873, + 0.0069448500871658325, + 0.006016414612531662, + 0.006512259598821402, + 0.011592483147978783, + 0.00523140374571085, + 0.007148377597332001, + 0.007290999870747328, + 0.006882083136588335, + 0.014405527152121067, + 0.00937685091048479, + 0.007325090002268553, + 0.043978650122880936, + 0.006235962267965078, + 0.008824252523481846, + 0.013760287314653397, + 0.010956312529742718, + 0.009449797682464123, + 0.011564656160771847, + 0.008488870225846767, + 0.014655319042503834, + 0.00906157772988081, + 0.010778643190860748, + 0.01696549728512764, + 0.009566908702254295, + 0.01708400994539261, + 0.012343705631792545, + 0.01329739112406969, + 0.03950870409607887, + 0.011933655478060246, + 0.016126442700624466, + 0.02554345689713955, + 0.015416618436574936, + 0.017455708235502243, + 0.015265685506165028, + 0.01782337948679924, + 0.026022909209132195, + 0.018422871828079224, + 0.017927682027220726, + 0.02499482035636902, + 0.0144686633720994, + 0.006353034637868404, + 0.005566182546317577, + 0.004880912601947784, + 0.008782495744526386, + 0.0058288308791816235, + 0.006367194466292858, + 0.008524652570486069, + 0.005543307401239872, + 0.004354698117822409, + 0.003905251156538725, + 0.004259059671312571, + 0.008094795048236847, + 0.005132894963026047, + 0.0038045335095375776, + 0.006732336711138487, + 0.00372148584574461, + 0.004381214268505573, + 0.0030260600615292788, + 0.004485693294554949, + 0.013657975010573864, + 0.002755835885182023, + 0.0035815672017633915, + 0.0051879314705729485, + 0.004025097470730543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.281338632106781 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019491974962875247 + ], + "opposite_logprob_tokens": [ + -6.251949310302734 + ], + "logprob_reward": -0.0019491974962875247, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019491974962875247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23286637663841248, + 0.15743233263492584, + 0.07617560774087906, + 0.18216672539710999, + 0.0654001235961914, + 0.06910517811775208, + 0.03547617793083191, + 0.05491982772946358, + 0.0821496993303299, + 0.044308025389909744 + ], + "credit_map": [ + 0.028401264920830727, + 0.02304268442094326, + 0.02746424823999405, + 0.03926748409867287, + 0.019253399223089218, + 0.0225518811494112, + 0.04821912199258804, + 0.024666273966431618, + 0.021402407437562943, + 0.018280329182744026, + 0.014792317524552345, + 0.030629029497504234, + 0.011653770692646503, + 0.014784292317926884, + 0.03216666728258133, + 0.01372350100427866, + 0.007120331283658743, + 0.009173284284770489, + 0.007644152268767357, + 0.012090836651623249, + 0.007415609434247017, + 0.008267977274954319, + 0.01650388166308403, + 0.007959529757499695, + 0.024914104491472244, + 0.01191509235650301, + 0.01175553910434246, + 0.06963279098272324, + 0.008499514311552048, + 0.012076805345714092, + 0.026662029325962067, + 0.01671081967651844, + 0.00827784463763237, + 0.0077093238942325115, + 0.006206415127962828, + 0.010805044323205948, + 0.00716709578409791, + 0.006297351326793432, + 0.012723834253847599, + 0.006213209591805935, + 0.009120705537497997, + 0.006645217537879944, + 0.005374494940042496, + 0.019038360565900803, + 0.004571986850351095, + 0.004159298725426197, + 0.014166888780891895, + 0.006028217729181051, + 0.003919664770364761, + 0.0035466474946588278, + 0.003707863623276353, + 0.006050006952136755, + 0.0033157337456941605, + 0.004534500185400248, + 0.007057054899632931, + 0.003344703698530793, + 0.006622155662626028, + 0.004859641660004854, + 0.004387606400996447, + 0.008438820950686932, + 0.006388130132108927, + 0.006768019404262304, + 0.011723971925675869, + 0.005731477867811918, + 0.009980102069675922, + 0.006617927923798561, + 0.005071348510682583, + 0.030469346791505814, + 0.005196998827159405, + 0.005764343310147524, + 0.010873322375118732, + 0.008176307193934917, + 0.004467083141207695, + 0.003587252227589488, + 0.003415990388020873, + 0.01046033389866352, + 0.003702362533658743, + 0.0042186640202999115, + 0.00995163805782795, + 0.004504696931689978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0019491974962875247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033709511160850525 + ], + "opposite_logprob_tokens": [ + -3.4087095260620117 + ], + "logprob_reward": -0.033709511160850525, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033709511160850525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20787657797336578, + 0.10803874582052231, + 0.0609884113073349, + 0.3322979211807251, + 0.041756898164749146, + 0.051892444491386414, + 0.030474508181214333, + 0.049301568418741226, + 0.05129993334412575, + 0.06607295572757721 + ], + "credit_map": [ + 0.02230416238307953, + 0.03371458500623703, + 0.024324191734194756, + 0.025924239307641983, + 0.022695036605000496, + 0.024334967136383057, + 0.031274013221263885, + 0.023305373266339302, + 0.012885336764156818, + 0.014209510758519173, + 0.011007687076926231, + 0.020028885453939438, + 0.011019142344594002, + 0.009275254793465137, + 0.01962299831211567, + 0.009989931248128414, + 0.00619618222117424, + 0.007617373019456863, + 0.0056616212241351604, + 0.013799484819173813, + 0.004840479232370853, + 0.007469514850527048, + 0.009072041139006615, + 0.00633171945810318, + 0.04638872295618057, + 0.020602835342288017, + 0.021134179085493088, + 0.1405515968799591, + 0.01221479196101427, + 0.023474985733628273, + 0.037130095064640045, + 0.030800718814134598, + 0.004285323899239302, + 0.006478408817201853, + 0.003683029906824231, + 0.007103117182850838, + 0.004270925186574459, + 0.005799369886517525, + 0.0058201272040605545, + 0.004316597245633602, + 0.006167829968035221, + 0.006391030270606279, + 0.005058134440332651, + 0.008611034601926804, + 0.00588031904771924, + 0.006475171074271202, + 0.008726682513952255, + 0.004582245834171772, + 0.003462291555479169, + 0.0038446171674877405, + 0.003242649370804429, + 0.004971614107489586, + 0.003841595258563757, + 0.004125170409679413, + 0.004202080890536308, + 0.002784489654004574, + 0.0059467521496117115, + 0.005844400264322758, + 0.004968436900526285, + 0.007755636237561703, + 0.005535668693482876, + 0.007013568188995123, + 0.007715617306530476, + 0.004521493334323168, + 0.00528760626912117, + 0.005517312791198492, + 0.004903207067400217, + 0.011752812191843987, + 0.004973249044269323, + 0.005347181111574173, + 0.008797960355877876, + 0.004720598924905062, + 0.005665648728609085, + 0.005936273839324713, + 0.006112635601311922, + 0.009357544593513012, + 0.007500541862100363, + 0.006734676659107208, + 0.01764945313334465, + 0.00711618410423398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.033709511160850525 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948933307081461 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005948933307081461, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948933307081461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2277248650789261, + 0.0865025520324707, + 0.03925769776105881, + 0.14318834245204926, + 0.05667915195226669, + 0.12552781403064728, + 0.10688246786594391, + 0.06812220066785812, + 0.06873804330825806, + 0.07737679779529572 + ], + "credit_map": [ + 0.03099297359585762, + 0.022373290732502937, + 0.029153406620025635, + 0.03139800205826759, + 0.018576553091406822, + 0.028134221211075783, + 0.04293717443943024, + 0.024159260094165802, + 0.008583596907556057, + 0.009144624695181847, + 0.010964469984173775, + 0.01453512255102396, + 0.008085171692073345, + 0.01111177820712328, + 0.01519928127527237, + 0.008878505788743496, + 0.003771719289943576, + 0.003816511482000351, + 0.004627020098268986, + 0.005387540441006422, + 0.0045216865837574005, + 0.00457205343991518, + 0.008195753209292889, + 0.004365415778011084, + 0.019905947148799896, + 0.009289361536502838, + 0.010369963943958282, + 0.05967021360993385, + 0.005691954866051674, + 0.010536037385463715, + 0.015140140429139137, + 0.012584719806909561, + 0.0058396002277731895, + 0.00824766606092453, + 0.005142920650541782, + 0.006428639870136976, + 0.006686781998723745, + 0.00995786115527153, + 0.007098454050719738, + 0.00727722467854619, + 0.016091009601950645, + 0.010981746017932892, + 0.007600774988532066, + 0.044677913188934326, + 0.007089939899742603, + 0.00808563269674778, + 0.01924482174217701, + 0.011755988001823425, + 0.011094167828559875, + 0.01329219713807106, + 0.010631847195327282, + 0.01741810515522957, + 0.011922294273972511, + 0.01565563678741455, + 0.017442725598812103, + 0.00942550040781498, + 0.007703412789851427, + 0.005993945989757776, + 0.007438295520842075, + 0.01020059920847416, + 0.006855431944131851, + 0.00865973997861147, + 0.014420764520764351, + 0.006850011646747589, + 0.00949679221957922, + 0.005193996708840132, + 0.004801295232027769, + 0.02332998812198639, + 0.005158423446118832, + 0.005059107206761837, + 0.008777201175689697, + 0.00692124105989933, + 0.007710837759077549, + 0.007520091254264116, + 0.005305982194840908, + 0.018944988027215004, + 0.006674839649349451, + 0.007650849409401417, + 0.013482715003192425, + 0.010086488910019398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.005948933307081461 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up focuses on a heavy, matte black cast iron pot on the grate of a modern gas stove. Inside, a rich, crimson sauce simmers gently, its deep, velvety red surface shimmering in the warm, soft light. Tiny flecks of dark green herbs are suspended in the viscous liquid, while tiny, lazy bubbles steadily rise from the depths, puffing at the surface before bursting with a soft \"plop\" and delicate, concentric ripples, releasing a delicate plume of steam. The pot sits at the center of a sleek black glass cooktop, revealing a low, controlled ring of vibrant blue and orange flames providing steady heat. The entire scene is bathed in a warm, inviting glow from an unseen overhead source, creating soft highlights on the pot's curved sides and the shimmering sauce, establishing a cozy, focused, and deeply comforting atmosphere of patient home cooking.", + "seed": 16517, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_12/index_677_seed_16517_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05820966274477542, + "vlm_logprob": -0.05820966274477542, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.05820966274477542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009568660520017147 + ], + "opposite_logprob_tokens": [ + -5.884568691253662 + ], + "logprob_reward": -0.009568660520017147, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.009568660520017147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07728816568851471, + 0.10844686627388, + 0.1183750182390213, + 0.1967068910598755, + 0.04442514851689339, + 0.04927188903093338, + 0.014789700508117676, + 0.06115789711475372, + 0.3018413484096527, + 0.02769707329571247 + ], + "credit_map": [ + 0.007241522427648306, + 0.00659641157835722, + 0.011177507229149342, + 0.015126807615160942, + 0.00650969659909606, + 0.008450524881482124, + 0.009200617671012878, + 0.012985073029994965, + 0.00916680321097374, + 0.012923575937747955, + 0.017158569768071175, + 0.017576904967427254, + 0.006100766360759735, + 0.011593450792133808, + 0.016966933384537697, + 0.016959859058260918, + 0.011438621208071709, + 0.014325954020023346, + 0.0120850196108222, + 0.02931036986410618, + 0.010113186202943325, + 0.006285021081566811, + 0.020952614024281502, + 0.013864236883819103, + 0.01460674311965704, + 0.045614223927259445, + 0.012606975622475147, + 0.07205715775489807, + 0.00885135680437088, + 0.010657072067260742, + 0.013935528695583344, + 0.01837783493101597, + 0.004961283877491951, + 0.005570673383772373, + 0.004499885719269514, + 0.00731693534180522, + 0.003611260326579213, + 0.004480228293687105, + 0.009103240445256233, + 0.004881641361862421, + 0.003570590866729617, + 0.007939410395920277, + 0.004666813649237156, + 0.01582256518304348, + 0.0024204382207244635, + 0.00333832623437047, + 0.005711765959858894, + 0.005801977124065161, + 0.0019484380027279258, + 0.0017096219817176461, + 0.002112207468599081, + 0.001604021294042468, + 0.0015990021638572216, + 0.0020956932567059994, + 0.0015149287646636367, + 0.002205786295235157, + 0.005818749312311411, + 0.011866653338074684, + 0.00881644245237112, + 0.011445049196481705, + 0.004968960303813219, + 0.004572931677103043, + 0.0063589150086045265, + 0.007310197222977877, + 0.020313795655965805, + 0.08486534655094147, + 0.012056341394782066, + 0.12000486999750137, + 0.011449366807937622, + 0.005686627700924873, + 0.026794450357556343, + 0.020670561119914055, + 0.00332083273679018, + 0.004084675572812557, + 0.005715667735785246, + 0.004969357047230005, + 0.0018958504078909755, + 0.001621516072191298, + 0.003067282261326909, + 0.0030218923930078745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.009568660520017147 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252084881067276 + ], + "opposite_logprob_tokens": [ + -1.5020848512649536 + ], + "logprob_reward": -0.252084881067276, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252084881067276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12262992560863495, + 0.16347651183605194, + 0.2668990194797516, + 0.16321510076522827, + 0.0629330575466156, + 0.07363205403089523, + 0.030965134501457214, + 0.0126801747828722, + 0.06633815914392471, + 0.03723081201314926 + ], + "credit_map": [ + 0.016338061541318893, + 0.017845701426267624, + 0.0146244578063488, + 0.016787895932793617, + 0.010257778689265251, + 0.013768790289759636, + 0.014696558937430382, + 0.018310682848095894, + 0.015894444659352303, + 0.021592196077108383, + 0.01420712098479271, + 0.03538728505373001, + 0.011682955548167229, + 0.015048948116600513, + 0.030963487923145294, + 0.01870007812976837, + 0.027295149862766266, + 0.029423825442790985, + 0.03907318040728569, + 0.05979340523481369, + 0.023887595161795616, + 0.016990266740322113, + 0.03226738050580025, + 0.03816821426153183, + 0.013038866221904755, + 0.03890809416770935, + 0.008644047193229198, + 0.06676694005727768, + 0.006621138192713261, + 0.006097558885812759, + 0.010995287448167801, + 0.012143171392381191, + 0.00875681359320879, + 0.006162844132632017, + 0.00903943832963705, + 0.009999256581068039, + 0.004147178493440151, + 0.005272301379591227, + 0.006733717862516642, + 0.012821502052247524, + 0.010940990410745144, + 0.007495165336877108, + 0.009774945676326752, + 0.009493744000792503, + 0.004074661992490292, + 0.007566530257463455, + 0.006413294933736324, + 0.01787271536886692, + 0.00448549585416913, + 0.0025823735632002354, + 0.003144369460642338, + 0.00459845457226038, + 0.0017634342657402158, + 0.004547334276139736, + 0.003910391591489315, + 0.005933280568569899, + 0.0015993230044841766, + 0.0016290731728076935, + 0.0018614078871905804, + 0.001472091069445014, + 0.000854330719448626, + 0.0013846434885635972, + 0.001797220786102116, + 0.002082084771245718, + 0.0055318535305559635, + 0.01788662001490593, + 0.00405122758820653, + 0.021270010620355606, + 0.0036140140146017075, + 0.002217276953160763, + 0.0051972405053675175, + 0.00656991358846426, + 0.005098001100122929, + 0.0031923705246299505, + 0.008810154162347317, + 0.005584809463471174, + 0.002176763257011771, + 0.0031490109395235777, + 0.0032938853837549686, + 0.005925815086811781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.252084881067276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011854295153170824 + ], + "opposite_logprob_tokens": [ + -6.751185417175293 + ], + "logprob_reward": -0.0011854295153170824, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011854295153170824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13435859978199005, + 0.07745935767889023, + 0.09984885156154633, + 0.2459578812122345, + 0.03449833020567894, + 0.16201278567314148, + 0.036844003945589066, + 0.05571170523762703, + 0.10092020779848099, + 0.05238829553127289 + ], + "credit_map": [ + 0.013400029391050339, + 0.026539001613855362, + 0.008171727880835533, + 0.028757521882653236, + 0.009449887089431286, + 0.012115070596337318, + 0.02105279266834259, + 0.014872567728161812, + 0.008483855053782463, + 0.009544327855110168, + 0.005628630984574556, + 0.015445849858224392, + 0.00563680799677968, + 0.006871523801237345, + 0.01657940447330475, + 0.009268960915505886, + 0.011028863489627838, + 0.014808529056608677, + 0.007538368925452232, + 0.032378822565078735, + 0.006133636925369501, + 0.008274653926491737, + 0.0076448991894721985, + 0.012041078880429268, + 0.017781062051653862, + 0.056719183921813965, + 0.01606016233563423, + 0.08757265657186508, + 0.011352277360856533, + 0.012670464813709259, + 0.020991602912545204, + 0.022810479626059532, + 0.0037392200902104378, + 0.00502891605719924, + 0.0031024045310914516, + 0.005615379195660353, + 0.002903542248532176, + 0.0037382920272648335, + 0.0038266240153461695, + 0.006543950643390417, + 0.014071446843445301, + 0.022800734266638756, + 0.01547485962510109, + 0.05031534284353256, + 0.007417649496346712, + 0.01029224507510662, + 0.0184735506772995, + 0.023166954517364502, + 0.004674552008509636, + 0.0041700247675180435, + 0.004967937711626291, + 0.005330907646566629, + 0.0018477834528312087, + 0.004546693060547113, + 0.005684645846486092, + 0.005621459800750017, + 0.006493884604424238, + 0.004878635052591562, + 0.007274389732629061, + 0.006628586445003748, + 0.005063535645604134, + 0.008391513489186764, + 0.007324578240513802, + 0.009656582027673721, + 0.008535251952707767, + 0.02492089569568634, + 0.0100260553881526, + 0.02482462115585804, + 0.0076883635483682156, + 0.0049470700323581696, + 0.007528356742113829, + 0.012449594214558601, + 0.007459680549800396, + 0.0059092091396451, + 0.009775537997484207, + 0.00849111471325159, + 0.003455097321420908, + 0.003966658841818571, + 0.006935769692063332, + 0.0063952249474823475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.0011854295153170824 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583942532539368 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583942532539368, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583942532539368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07138518989086151, + 0.044238656759262085, + 0.10114966332912445, + 0.2699231803417206, + 0.020442981272935867, + 0.05012340098619461, + 0.02822350151836872, + 0.07549920678138733, + 0.21376605331897736, + 0.12524819374084473 + ], + "credit_map": [ + 0.007655312307178974, + 0.011355159804224968, + 0.005376365967094898, + 0.012193740345537663, + 0.005094096064567566, + 0.010836220346391201, + 0.006929216906428337, + 0.011945078149437904, + 0.005148065276443958, + 0.00471845269203186, + 0.005092184524983168, + 0.0071378047578036785, + 0.00335131143219769, + 0.0033284034579992294, + 0.009533379226922989, + 0.005929053761065006, + 0.009428218938410282, + 0.015490458346903324, + 0.011933660134673119, + 0.015189659781754017, + 0.012648121453821659, + 0.0058870320208370686, + 0.017395276576280594, + 0.01317723747342825, + 0.018484927713871002, + 0.07262910157442093, + 0.011152982711791992, + 0.11409694701433182, + 0.008943094871938229, + 0.007477446459233761, + 0.01826702617108822, + 0.01887165941298008, + 0.0029329005628824234, + 0.0024272638838738203, + 0.0021972714457660913, + 0.0022137537598609924, + 0.0018383236601948738, + 0.0026153468061238527, + 0.0030147419311106205, + 0.003203378990292549, + 0.004781260620802641, + 0.007825735956430435, + 0.003945197910070419, + 0.015459195710718632, + 0.0024645572993904352, + 0.004290848504751921, + 0.005264641251415014, + 0.006091960705816746, + 0.002981391502544284, + 0.005023312754929066, + 0.004931946750730276, + 0.004088867921382189, + 0.0018543555634096265, + 0.0025880250614136457, + 0.002575560938566923, + 0.004180042073130608, + 0.008519547991454601, + 0.006648345850408077, + 0.012626788578927517, + 0.008878630585968494, + 0.005981738213449717, + 0.009996066801249981, + 0.008605416864156723, + 0.014242668636143208, + 0.014605937525629997, + 0.06062416732311249, + 0.010316243395209312, + 0.0797300711274147, + 0.009556442499160767, + 0.005007762461900711, + 0.01780935749411583, + 0.016116062179207802, + 0.014767716638743877, + 0.012791928835213184, + 0.027186408638954163, + 0.019389020279049873, + 0.008421213366091251, + 0.011775548569858074, + 0.013223252259194851, + 0.01769309863448143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.020583942532539368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625400088727474 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625400088727474, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625400088727474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11598016321659088, + 0.08192510157823563, + 0.057455021888017654, + 0.35304054617881775, + 0.03665716201066971, + 0.04086463153362274, + 0.0653173103928566, + 0.07198304682970047, + 0.12126268446445465, + 0.05551436170935631 + ], + "credit_map": [ + 0.009435090236365795, + 0.023967141285538673, + 0.008061027154326439, + 0.01839751936495304, + 0.00910172425210476, + 0.019426623359322548, + 0.009636453352868557, + 0.01795458234846592, + 0.00858944933861494, + 0.009608626365661621, + 0.010758786462247372, + 0.017613139003515244, + 0.007098424714058638, + 0.007569392677396536, + 0.010489455424249172, + 0.010197821073234081, + 0.004464119207113981, + 0.006161720026284456, + 0.008930391632020473, + 0.009099578484892845, + 0.00634378707036376, + 0.003368886886164546, + 0.012593296356499195, + 0.006493240594863892, + 0.024508541449904442, + 0.08963030576705933, + 0.016744567081332207, + 0.14863714575767517, + 0.013475063256919384, + 0.010101302526891232, + 0.024871256202459335, + 0.025072351098060608, + 0.004927386064082384, + 0.004369781352579594, + 0.004473155364394188, + 0.004381284117698669, + 0.003305847989395261, + 0.0037858488503843546, + 0.005948531907051802, + 0.005465322639793158, + 0.005210177507251501, + 0.004412563052028418, + 0.004503057803958654, + 0.006448260974138975, + 0.004069291055202484, + 0.0036958178970962763, + 0.004977117292582989, + 0.0075483424589037895, + 0.007825278677046299, + 0.007362568750977516, + 0.00551540357992053, + 0.007975227199494839, + 0.004015664104372263, + 0.013100537471473217, + 0.006591311655938625, + 0.01293131336569786, + 0.007622461766004562, + 0.007016812451183796, + 0.010917366482317448, + 0.008139965124428272, + 0.005562511272728443, + 0.012185296975076199, + 0.007211420685052872, + 0.013327200897037983, + 0.009535077959299088, + 0.034959156066179276, + 0.007226422429084778, + 0.03866245970129967, + 0.006611633114516735, + 0.0037331285420805216, + 0.008995898067951202, + 0.011538902297616005, + 0.00704271299764514, + 0.005792499054223299, + 0.01576012745499611, + 0.006803187541663647, + 0.002649123314768076, + 0.0038771587423980236, + 0.006455929484218359, + 0.007133621722459793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.007625400088727474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up focuses on a heavy, matte black cast iron pot on the grate of a modern gas stove. Inside, a rich, crimson sauce simmers gently, its deep, velvety red surface shimmering in the warm, soft light. Tiny flecks of dark green herbs are suspended in the viscous liquid, while tiny, lazy bubbles steadily rise from the depths, puffing at the surface before bursting with a soft \"plop\" and delicate, concentric ripples, releasing a delicate plume of steam. The pot sits at the center of a sleek black glass cooktop, revealing a low, controlled ring of vibrant blue and orange flames providing steady heat. The entire scene is bathed in a warm, inviting glow from an unseen overhead source, creating soft highlights on the pot's curved sides and the shimmering sauce, establishing a cozy, focused, and deeply comforting atmosphere of patient home cooking.", + "seed": 16518, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_12/index_677_seed_16518_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.052483067568391564, + "vlm_logprob": -0.052483067568391564, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.052483067568391564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007137988694012165 + ], + "opposite_logprob_tokens": [ + -6.007137775421143 + ], + "logprob_reward": -0.007137988694012165, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007137988694012165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0929064080119133, + 0.2856893539428711, + 0.22627732157707214, + 0.0527605265378952, + 0.08117392659187317, + 0.035165492445230484, + 0.0231113713234663, + 0.0424283966422081, + 0.0642675906419754, + 0.09621959179639816 + ], + "credit_map": [ + 0.012430742383003235, + 0.01480248011648655, + 0.008306695148348808, + 0.009884159080684185, + 0.01543564721941948, + 0.011303137056529522, + 0.012371503747999668, + 0.008372046053409576, + 0.0149020841345191, + 0.03686021640896797, + 0.009702496230602264, + 0.0168316513299942, + 0.10369223356246948, + 0.018445689231157303, + 0.0714893564581871, + 0.013765660114586353, + 0.009008626453578472, + 0.016707943752408028, + 0.005843797232955694, + 0.01024341769516468, + 0.07615378499031067, + 0.01187001820653677, + 0.07631347328424454, + 0.020136268809437752, + 0.0031128607224673033, + 0.0040849135257303715, + 0.0026090273167937994, + 0.010117764584720135, + 0.00857915822416544, + 0.0028977212496101856, + 0.01763257384300232, + 0.0037265114951878786, + 0.006242881994694471, + 0.008089199662208557, + 0.0027357519138604403, + 0.004694413393735886, + 0.02790815569460392, + 0.004659946542233229, + 0.02198464423418045, + 0.004858933389186859, + 0.00296696531586349, + 0.0038512754254043102, + 0.0023044610861688852, + 0.004833696875721216, + 0.005621829070150852, + 0.0031572459265589714, + 0.009314663708209991, + 0.0031153566669672728, + 0.0019330678042024374, + 0.002356067532673478, + 0.0014474192867055535, + 0.006483760662376881, + 0.0029424435924738646, + 0.0017843731911852956, + 0.00422412296757102, + 0.001940119662322104, + 0.0033209563698619604, + 0.0056944433599710464, + 0.00293319602496922, + 0.006658853497356176, + 0.006041825748980045, + 0.0028020525351166725, + 0.011986092664301395, + 0.0029909820295870304, + 0.005849403794854879, + 0.00476060900837183, + 0.0031404532492160797, + 0.008354591205716133, + 0.008513489738106728, + 0.002706060418859124, + 0.025919323787093163, + 0.005023663397878408, + 0.010003048926591873, + 0.006768585182726383, + 0.0047235991805791855, + 0.010920479893684387, + 0.016087524592876434, + 0.004243904259055853, + 0.03657190874218941, + 0.0069005414843559265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.007137988694012165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558759152889252 + ], + "opposite_logprob_tokens": [ + -1.6005876064300537 + ], + "logprob_reward": -0.22558759152889252, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558759152889252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13946792483329773, + 0.29083630442619324, + 0.1679016649723053, + 0.02981700748205185, + 0.17002573609352112, + 0.017790820449590683, + 0.03387446328997612, + 0.029731957241892815, + 0.09687396883964539, + 0.02368013933300972 + ], + "credit_map": [ + 0.011904405429959297, + 0.013003499247133732, + 0.010064775124192238, + 0.02039307728409767, + 0.017092635855078697, + 0.010865649208426476, + 0.04525141790509224, + 0.010892466641962528, + 0.01162361353635788, + 0.034012287855148315, + 0.008435859344899654, + 0.015884634107351303, + 0.10351619124412537, + 0.01651426963508129, + 0.0848357081413269, + 0.016013721004128456, + 0.012075552716851234, + 0.014447015710175037, + 0.009216750971972942, + 0.008242377080023289, + 0.04797036200761795, + 0.017239414155483246, + 0.043301358819007874, + 0.01540883257985115, + 0.0017987516475841403, + 0.002847977913916111, + 0.0017380134668201208, + 0.005220032297074795, + 0.00533386180177331, + 0.0019442765042185783, + 0.008932082913815975, + 0.00200201035477221, + 0.021220862865447998, + 0.023228876292705536, + 0.010319091379642487, + 0.009784375317394733, + 0.05765730142593384, + 0.013312790542840958, + 0.025389643386006355, + 0.009112793952226639, + 0.0018676755717024207, + 0.0028776845429092646, + 0.0014733572024852037, + 0.0023646848276257515, + 0.002723285462707281, + 0.001681046443991363, + 0.0026844381354749203, + 0.002118648961186409, + 0.0021791548933833838, + 0.0032928292639553547, + 0.00219134078361094, + 0.009745774790644646, + 0.004447550978511572, + 0.001990801189094782, + 0.007974996231496334, + 0.002052013762295246, + 0.0030265003442764282, + 0.0043683429248631, + 0.0027428343892097473, + 0.004605029709637165, + 0.0031108565162867308, + 0.002907700603827834, + 0.00599273294210434, + 0.0029779623728245497, + 0.009492943994700909, + 0.006947052199393511, + 0.005177890416234732, + 0.0114162378013134, + 0.013527955859899521, + 0.003906571306288242, + 0.03857046738266945, + 0.007834858261048794, + 0.0027095351833850145, + 0.0019429894164204597, + 0.0013521242653951049, + 0.003663994837552309, + 0.0036575929261744022, + 0.0012605006340891123, + 0.007656486239284277, + 0.0014369144337251782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.22558759152889252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017190692014992237 + ], + "opposite_logprob_tokens": [ + -6.376718997955322 + ], + "logprob_reward": -0.0017190692014992237, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017190692014992237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0560782290995121, + 0.0719204917550087, + 0.5604032874107361, + 0.0169040746986866, + 0.06510819494724274, + 0.019989492371678352, + 0.04298761859536171, + 0.02858833782374859, + 0.08988851308822632, + 0.04813173785805702 + ], + "credit_map": [ + 0.004804751370102167, + 0.010250171646475792, + 0.004261425230652094, + 0.013212976045906544, + 0.00871371105313301, + 0.0034673556219786406, + 0.00814754981547594, + 0.003220286685973406, + 0.008490280248224735, + 0.012853006832301617, + 0.0060155075043439865, + 0.008463704027235508, + 0.009685622528195381, + 0.005619990173727274, + 0.01565529592335224, + 0.005137087311595678, + 0.017914092168211937, + 0.0378851555287838, + 0.012302001938223839, + 0.02970808930695057, + 0.1843978315591812, + 0.024799039587378502, + 0.20329949259757996, + 0.05009761080145836, + 0.0015300641534850001, + 0.002633269177749753, + 0.0018510677618905902, + 0.0032043501269072294, + 0.0020870782900601625, + 0.0015390438493341208, + 0.0025260737165808678, + 0.0015331265749409795, + 0.004786794073879719, + 0.0059229531325399876, + 0.002604347188025713, + 0.0037493116687983274, + 0.02064979262650013, + 0.003850468434393406, + 0.018877601251006126, + 0.004666927270591259, + 0.0013799465959891677, + 0.001922020921483636, + 0.0013670321786776185, + 0.0025510420091450214, + 0.003741051536053419, + 0.00153210514690727, + 0.005611959844827652, + 0.0018843339057639241, + 0.003207830712199211, + 0.004853248130530119, + 0.002300342544913292, + 0.013676279224455357, + 0.004181961063295603, + 0.002770588966086507, + 0.009198394604027271, + 0.0027989745140075684, + 0.0020441971719264984, + 0.003444296307861805, + 0.0021275770850479603, + 0.00573631701990962, + 0.004013014491647482, + 0.0020301456097513437, + 0.007055132184177637, + 0.0021376577205955982, + 0.008739186450839043, + 0.006388003937900066, + 0.004512259736657143, + 0.010892081074416637, + 0.012431207112967968, + 0.0036691937129944563, + 0.03627953305840492, + 0.006977053824812174, + 0.005874534137547016, + 0.0043102530762553215, + 0.0035866263788193464, + 0.014333521015942097, + 0.00323331356048584, + 0.0032609212212264538, + 0.009933420456945896, + 0.003599148476496339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.0017190692014992237 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190933391451836 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018190933391451836, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190933391451836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1580907553434372, + 0.09386976808309555, + 0.09477214515209198, + 0.037778060883283615, + 0.3020850121974945, + 0.044154148548841476, + 0.031675465404987335, + 0.04819437861442566, + 0.14028099179267883, + 0.04909934848546982 + ], + "credit_map": [ + 0.01364316325634718, + 0.015506948344409466, + 0.012367002665996552, + 0.021377913653850555, + 0.020540637895464897, + 0.011022244580090046, + 0.05069650337100029, + 0.012936311773955822, + 0.005255962256342173, + 0.010794753208756447, + 0.006877743639051914, + 0.011410149745643139, + 0.022006206214427948, + 0.007006071042269468, + 0.024226617068052292, + 0.006292248610407114, + 0.006843442562967539, + 0.007309249602258205, + 0.005698880180716515, + 0.009260399267077446, + 0.02013372629880905, + 0.008475054055452347, + 0.029312632977962494, + 0.007738743908703327, + 0.002558659529313445, + 0.0032009556889533997, + 0.002882198430597782, + 0.005822444334626198, + 0.007293610367923975, + 0.0022433861158788204, + 0.011035137809813023, + 0.0027416623197495937, + 0.026766333729028702, + 0.03545641899108887, + 0.011832171119749546, + 0.011258609592914581, + 0.11043889075517654, + 0.016838135197758675, + 0.0736272856593132, + 0.015867125242948532, + 0.0030985011253505945, + 0.005454970523715019, + 0.0028954376466572285, + 0.010178402997553349, + 0.007525560446083546, + 0.0033712894655764103, + 0.008264598436653614, + 0.0033653811551630497, + 0.0024196451995521784, + 0.003942303359508514, + 0.0023256936110556126, + 0.004320032894611359, + 0.005167390685528517, + 0.002189226681366563, + 0.008594288490712643, + 0.002716879593208432, + 0.00406092032790184, + 0.006473745685070753, + 0.0031559716444462538, + 0.00868199486285448, + 0.006184948142617941, + 0.005224836524575949, + 0.010270343162119389, + 0.004141609184443951, + 0.012720631435513496, + 0.009588676504790783, + 0.006999616511166096, + 0.017421800643205643, + 0.020079119130969048, + 0.005676002707332373, + 0.0568692646920681, + 0.01092584803700447, + 0.0030427174642682076, + 0.0036365033593028784, + 0.002768886275589466, + 0.00634720828384161, + 0.010439669713377953, + 0.0028200671076774597, + 0.0157286636531353, + 0.004315622616559267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.018190933391451836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle close-up of a heavy matte black cast iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pot sitting on the grate of a modern gas stove with a sleek black glass cooktop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a rich crimson sauce with visible dark green herb flecks and rising bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there low, controlled rings of vibrant blue and orange flames visible beneath the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft light creating soft highlights on the pot and sauce?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23538470268249512, + 0.279523640871048, + 0.05217611417174339, + 0.02914702706038952, + 0.07964857667684555, + 0.02462516538798809, + 0.08072897046804428, + 0.02846536599099636, + 0.129519522190094, + 0.06078089401125908 + ], + "credit_map": [ + 0.022989990189671516, + 0.024152062833309174, + 0.019656045362353325, + 0.037489987909793854, + 0.023187316954135895, + 0.01900867186486721, + 0.06856195628643036, + 0.02033868245780468, + 0.012547878548502922, + 0.03796025365591049, + 0.008689915761351585, + 0.016342096030712128, + 0.09872514754533768, + 0.017754647880792618, + 0.07380545884370804, + 0.01369826216250658, + 0.0044790045358240604, + 0.005942084360867739, + 0.003952774219214916, + 0.009071159176528454, + 0.005250848829746246, + 0.004877552390098572, + 0.014267480932176113, + 0.004335206001996994, + 0.002070346847176552, + 0.003929058089852333, + 0.0026725928764790297, + 0.00600990979000926, + 0.005467445123940706, + 0.002167884726077318, + 0.004805807955563068, + 0.002023982349783182, + 0.005017190705984831, + 0.009351139888167381, + 0.002994395326822996, + 0.004032269585877657, + 0.027766866609454155, + 0.004367087967693806, + 0.02144613303244114, + 0.004673494026064873, + 0.0021260399371385574, + 0.004175896290689707, + 0.002347863744944334, + 0.002427150262519717, + 0.004058230668306351, + 0.00250297412276268, + 0.004322614520788193, + 0.0026643972378224134, + 0.006172433495521545, + 0.008370601572096348, + 0.004524592310190201, + 0.024033503606915474, + 0.008394264616072178, + 0.005742376204580069, + 0.018272673711180687, + 0.005218520760536194, + 0.0029176196549087763, + 0.004002680536359549, + 0.0023813415318727493, + 0.0044560679234564304, + 0.0032769665122032166, + 0.003087905002757907, + 0.005650950130075216, + 0.002691835630685091, + 0.012892208062112331, + 0.008858850225806236, + 0.006573361810296774, + 0.015567532740533352, + 0.018345359712839127, + 0.005362323950976133, + 0.052066098898649216, + 0.0098537877202034, + 0.004633704666048288, + 0.005301238037645817, + 0.004527190234512091, + 0.016206789761781693, + 0.007123694755136967, + 0.004607126582413912, + 0.012883524410426617, + 0.005497628822922707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.009779755026102066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_tensor([7815, 7815])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_tensor([7815, 7815])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f952f9a215e31213c602831405c074f58d7df868 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_tensor([7815, 7815])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up triggers a smooth, sustained downward tilt, meticulously tracking a delicate square paper tea bag as it is pulled by gravity from the gleaming white ceramic rim of a teacup. The cup, filled to the brim with crystal-clear, still water, rests on a smooth, dark, non-reflective countertop. As the tea bag gently breaks the water's surface tension with an almost imperceptible ripple, its porous paper texture darkens instantly upon contact. The camera follows its graceful and steady descent through the liquid, capturing the moment when wisps of ethereal cloud, rich amber, begin to gently tumble from the paper, unfolding in slow motion like smoke. The scene is illuminated by soft, diffuse sidelighting, creating subtle highlights on the cup's glaze and causing the water to shimmer, enhancing the atmosphere of tranquility and contemplation. As the tea bag softly settles onto the cup's bottom, the camera's movement ends, while a slow, captivating spread of color continues to bloom outward, transforming the clear water into warm, organic hues.", + "seed": 11190, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_7815_seed_11190_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2304783347994089, + "vlm_logprob": -0.2304783347994089, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.2304783347994089, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09127507358789444 + ], + "opposite_logprob_tokens": [ + -2.5912749767303467 + ], + "logprob_reward": -0.09127507358789444, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09127507358789444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2297792285680771, + 0.1650109738111496, + 0.1427488923072815, + 0.08874326199293137, + 0.09859241545200348, + 0.03391082584857941, + 0.08874765038490295, + 0.07217909395694733, + 0.03586385026574135, + 0.04442385584115982 + ], + "credit_map": [ + 0.02507806196808815, + 0.04881225526332855, + 0.01835755817592144, + 0.041085101664066315, + 0.03906068205833435, + 0.010205620899796486, + 0.03633076325058937, + 0.010849183425307274, + 0.06105699762701988, + 0.014346638694405556, + 0.007442514877766371, + 0.00989450141787529, + 0.0169985331594944, + 0.03471875935792923, + 0.012711768038570881, + 0.00784126203507185, + 0.011332808993756771, + 0.017310436815023422, + 0.013641050085425377, + 0.01269734650850296, + 0.04243076220154762, + 0.01321739237755537, + 0.021877432242035866, + 0.01024165004491806, + 0.009389842860400677, + 0.016244225203990936, + 0.006184624508023262, + 0.010136486031115055, + 0.0175928957760334, + 0.010418080724775791, + 0.010174689814448357, + 0.008602415211498737, + 0.004783795680850744, + 0.03142689913511276, + 0.0024673391599208117, + 0.0028373664245009422, + 0.012827317230403423, + 0.03582322597503662, + 0.0036242706701159477, + 0.004802199080586433, + 0.0041258689016103745, + 0.005225601140409708, + 0.0020201997831463814, + 0.002691838424652815, + 0.008530844002962112, + 0.00514226732775569, + 0.0041833422146737576, + 0.0019908633548766375, + 0.008398129604756832, + 0.01685461588203907, + 0.005962195806205273, + 0.008658637292683125, + 0.01909206621348858, + 0.010485832579433918, + 0.013659593649208546, + 0.005636580754071474, + 0.013405346311628819, + 0.010275933891534805, + 0.00376692577265203, + 0.0026578521355986595, + 0.008429326117038727, + 0.02766428515315056, + 0.0020926394499838352, + 0.003886779770255089, + 0.005666040815412998, + 0.005287558306008577, + 0.0031503336504101753, + 0.0017063003033399582, + 0.005981186870485544, + 0.0059209903702139854, + 0.003294615540653467, + 0.004856822546571493, + 0.004062403924763203, + 0.004199369810521603, + 0.0029598206747323275, + 0.01245621033012867, + 0.003927208483219147, + 0.010755096562206745, + 0.0037275967188179493, + 0.002336151897907257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.09127507358789444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578415870666504 + ], + "opposite_logprob_tokens": [ + -0.6328415870666504 + ], + "logprob_reward": -0.7578415870666504, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578415870666504, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.47653117775917053, + 0.14753684401512146, + 0.1010533943772316, + 0.05584365501999855, + 0.12483705580234528, + 0.02125534787774086, + 0.017209790647029877, + 0.012989075854420662, + 0.014538077637553215, + 0.02820562571287155 + ], + "credit_map": [ + 0.06241791322827339, + 0.11159301549196243, + 0.026864750310778618, + 0.04692457243800163, + 0.09908148646354675, + 0.02031904272735119, + 0.08917296677827835, + 0.020157426595687866, + 0.0573335736989975, + 0.014773273840546608, + 0.006912791635841131, + 0.007024935446679592, + 0.013966360129415989, + 0.032727017998695374, + 0.009227343834936619, + 0.005571554880589247, + 0.010193455964326859, + 0.014154266566038132, + 0.006826478056609631, + 0.014289475046098232, + 0.018443165346980095, + 0.013451606035232544, + 0.014743446372449398, + 0.008951500058174133, + 0.005683584604412317, + 0.010874224826693535, + 0.00358111341483891, + 0.006064016837626696, + 0.01167087722569704, + 0.007733828853815794, + 0.005120191723108292, + 0.005115818232297897, + 0.007022586651146412, + 0.035184551030397415, + 0.003612278029322624, + 0.00630818260833621, + 0.016421353444457054, + 0.042789094150066376, + 0.00872036349028349, + 0.004778639413416386, + 0.002925439039245248, + 0.0026732985861599445, + 0.0018907245248556137, + 0.001511974143795669, + 0.004124135244637728, + 0.004011327400803566, + 0.0027159966994076967, + 0.0014024508418515325, + 0.0019607911817729473, + 0.002660860773175955, + 0.001495705102570355, + 0.001989954151213169, + 0.0025975480675697327, + 0.0028810061048716307, + 0.00245563848875463, + 0.0011682860786095262, + 0.0022594439797103405, + 0.0019662438426166773, + 0.0017653991235420108, + 0.000914272153750062, + 0.001611012383364141, + 0.0024713154416531324, + 0.0007470279815606773, + 0.0012543612392619252, + 0.0018959228182211518, + 0.0025012909900397062, + 0.0010455823503434658, + 0.000999297248199582, + 0.0029187120962888002, + 0.0026809582486748695, + 0.0010694308439269662, + 0.0014268823433667421, + 0.0031501224730163813, + 0.003720443695783615, + 0.0015660639619454741, + 0.005297563038766384, + 0.003890946740284562, + 0.0070486292243003845, + 0.001915615051984787, + 0.0016162419924512506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.7578415870666504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16049492359161377 + ], + "opposite_logprob_tokens": [ + -1.9104949235916138 + ], + "logprob_reward": -0.16049492359161377, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16049492359161377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4983319044113159, + 0.02324109897017479, + 0.11512713879346848, + 0.07150077074766159, + 0.12165733426809311, + 0.027868002653121948, + 0.027226213365793228, + 0.04642811417579651, + 0.0598825067281723, + 0.008736873045563698 + ], + "credit_map": [ + 0.06935592740774155, + 0.10683120042085648, + 0.024621514603495598, + 0.049658335745334625, + 0.09495062381029129, + 0.02021237090229988, + 0.11151501536369324, + 0.021186931058764458, + 0.0032457008492201567, + 0.0037552588619291782, + 0.003352127270773053, + 0.0029008963610976934, + 0.003765997476875782, + 0.0016205364372581244, + 0.0026402005460113287, + 0.001960381865501404, + 0.009539574384689331, + 0.015036975964903831, + 0.010563571006059647, + 0.011948890052735806, + 0.03221152350306511, + 0.009744002483785152, + 0.01869058795273304, + 0.007392013445496559, + 0.005896161310374737, + 0.014108228497207165, + 0.0034364028833806515, + 0.0072208428755402565, + 0.018950115889310837, + 0.008146843872964382, + 0.008873570710420609, + 0.004868607502430677, + 0.005932196509093046, + 0.046419400721788406, + 0.0024845036678016186, + 0.003194054588675499, + 0.012315826490521431, + 0.04334136098623276, + 0.004044433124363422, + 0.00392556469887495, + 0.0025272658094763756, + 0.004991220310330391, + 0.0012238961644470692, + 0.0017724605277180672, + 0.008846904151141644, + 0.004024384543299675, + 0.0030301245860755444, + 0.0014517467934638262, + 0.0031815089751034975, + 0.004934642463922501, + 0.0021371166221797466, + 0.0021030500065535307, + 0.006507318466901779, + 0.00382077950052917, + 0.002066606655716896, + 0.00247519020922482, + 0.008079856634140015, + 0.006302936002612114, + 0.005057202652096748, + 0.0026873250026255846, + 0.006893913261592388, + 0.010891451500356197, + 0.0020824093371629715, + 0.004433020483702421, + 0.009149085730314255, + 0.009497473947703838, + 0.004554016515612602, + 0.0030024570878595114, + 0.009687015786767006, + 0.012224993668496609, + 0.005136665888130665, + 0.006630799267441034, + 0.0009268358116969466, + 0.0007837121956981719, + 0.000510802143253386, + 0.0015573586570098996, + 0.0008297741878777742, + 0.0032459732610732317, + 0.00041156308725476265, + 0.0004708535270765424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.16049492359161377 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029835926368832588 + ], + "opposite_logprob_tokens": [ + -3.5298359394073486 + ], + "logprob_reward": -0.029835926368832588, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029835926368832588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37918803095817566, + 0.05441631004214287, + 0.1292702555656433, + 0.08133453130722046, + 0.1403045654296875, + 0.0478849820792675, + 0.046102073043584824, + 0.05275322124361992, + 0.042719919234514236, + 0.026026135310530663 + ], + "credit_map": [ + 0.05080870911478996, + 0.0806874930858612, + 0.024294713512063026, + 0.044793978333473206, + 0.0749603807926178, + 0.016266092658042908, + 0.0673387423157692, + 0.02003791369497776, + 0.007985384203493595, + 0.007413433399051428, + 0.0038055337499827147, + 0.006532169412821531, + 0.010943017899990082, + 0.005516041535884142, + 0.00899559911340475, + 0.0032251274678856134, + 0.012178411707282066, + 0.017188530415296555, + 0.010320756584405899, + 0.011902048252522945, + 0.04128514975309372, + 0.00995240081101656, + 0.01752474717795849, + 0.008918210864067078, + 0.00907047651708126, + 0.0146145885810256, + 0.005442345514893532, + 0.009682951495051384, + 0.013832313939929008, + 0.011728895828127861, + 0.010624926537275314, + 0.0063380226492881775, + 0.010740824043750763, + 0.048480961471796036, + 0.004928635433316231, + 0.009227016009390354, + 0.023140402510762215, + 0.026073776185512543, + 0.011414172127842903, + 0.006298770196735859, + 0.007503986358642578, + 0.006633944809436798, + 0.0030938892159610987, + 0.0033520483411848545, + 0.011851361021399498, + 0.007438906002789736, + 0.004823414143174887, + 0.0031874296255409718, + 0.005007350351661444, + 0.007179453503340483, + 0.0036622444167733192, + 0.004700292367488146, + 0.009608527645468712, + 0.008331298828125, + 0.0035030762664973736, + 0.004109825938940048, + 0.008750253356993198, + 0.00784081220626831, + 0.0035587225575000048, + 0.0026210485957562923, + 0.007683827541768551, + 0.01623455248773098, + 0.0029795020818710327, + 0.003084498457610607, + 0.006617923732846975, + 0.005548752844333649, + 0.003307927632704377, + 0.0019685339648276567, + 0.007240873295813799, + 0.010365081019699574, + 0.0024359601084142923, + 0.005234859883785248, + 0.0033594476990401745, + 0.003463761880993843, + 0.0026936328504234552, + 0.0027628440875560045, + 0.003918258473277092, + 0.005765690002590418, + 0.0019495284650474787, + 0.002112968824803829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.029835926368832588 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294416338205338 + ], + "opposite_logprob_tokens": [ + -2.2379441261291504 + ], + "logprob_reward": -0.11294416338205338, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294416338205338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3704157769680023, + 0.06770628690719604, + 0.06460709124803543, + 0.04740818589925766, + 0.28268975019454956, + 0.043673496693372726, + 0.04663235321640968, + 0.03439546003937721, + 0.03676750510931015, + 0.005704151000827551 + ], + "credit_map": [ + 0.04840059578418732, + 0.07260970771312714, + 0.023369867354631424, + 0.04610876366496086, + 0.06372623145580292, + 0.01644301973283291, + 0.08114012330770493, + 0.01861743815243244, + 0.0213211290538311, + 0.0070470180362463, + 0.004781264811754227, + 0.005039178300648928, + 0.007260625716298819, + 0.012488114647567272, + 0.0059800236485898495, + 0.003788924543187022, + 0.007067039143294096, + 0.008008569478988647, + 0.004733758512884378, + 0.007874584756791592, + 0.01544780284166336, + 0.007407274097204208, + 0.009142501279711723, + 0.004925552289932966, + 0.0032697771675884724, + 0.00783165916800499, + 0.0030885792803019285, + 0.005679446272552013, + 0.012777126394212246, + 0.004628392867743969, + 0.005979390349239111, + 0.004153809044510126, + 0.013592468574643135, + 0.10402350127696991, + 0.005731363780796528, + 0.007796222809702158, + 0.02233242429792881, + 0.10510142147541046, + 0.012387304566800594, + 0.011725042015314102, + 0.00666571781039238, + 0.007258886471390724, + 0.0029521186370402575, + 0.0031145717948675156, + 0.008477701805531979, + 0.0078107016161084175, + 0.004701630678027868, + 0.002692166017368436, + 0.005403186660259962, + 0.008293546736240387, + 0.0026434690225869417, + 0.002397977514192462, + 0.015070033259689808, + 0.006940900813788176, + 0.0024684660602360964, + 0.003414768725633621, + 0.006566287018358707, + 0.004593462683260441, + 0.0016801790334284306, + 0.0011696371948346496, + 0.003654791507869959, + 0.012717245146632195, + 0.0013465051306411624, + 0.002667349064722657, + 0.004610826261341572, + 0.006100175902247429, + 0.0030637322925031185, + 0.002193663502112031, + 0.007306601386517286, + 0.005746698938310146, + 0.003446572693064809, + 0.004299231804907322, + 0.0005472179618664086, + 0.0005380667862482369, + 0.00041337142465636134, + 0.0012484207982197404, + 0.0006635324680246413, + 0.0014176544500514865, + 0.0004836410516873002, + 0.0003922453033737838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.11294416338205338 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up triggers a smooth, sustained downward tilt, meticulously tracking a delicate square paper tea bag as it is pulled by gravity from the gleaming white ceramic rim of a teacup. The cup, filled to the brim with crystal-clear, still water, rests on a smooth, dark, non-reflective countertop. As the tea bag gently breaks the water's surface tension with an almost imperceptible ripple, its porous paper texture darkens instantly upon contact. The camera follows its graceful and steady descent through the liquid, capturing the moment when wisps of ethereal cloud, rich amber, begin to gently tumble from the paper, unfolding in slow motion like smoke. The scene is illuminated by soft, diffuse sidelighting, creating subtle highlights on the cup's glaze and causing the water to shimmer, enhancing the atmosphere of tranquility and contemplation. As the tea bag softly settles onto the cup's bottom, the camera's movement ends, while a slow, captivating spread of color continues to bloom outward, transforming the clear water into warm, organic hues.", + "seed": 11191, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_7815_seed_11191_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2369583085179329, + "vlm_logprob": -0.2369583085179329, + "vlm_yes_no_margin": 1.85, + "vlm_token_credit": -0.2369583085179329, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03742416203022003 + ], + "opposite_logprob_tokens": [ + -3.662424087524414 + ], + "logprob_reward": -0.03742416203022003, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03742416203022003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09359733015298843, + 0.14455458521842957, + 0.03539736196398735, + 0.03936950862407684, + 0.07345130294561386, + 0.045472390949726105, + 0.0860721617937088, + 0.13396196067333221, + 0.28637635707855225, + 0.06174701824784279 + ], + "credit_map": [ + 0.006655002478510141, + 0.012670992873609066, + 0.01036137156188488, + 0.025156566873192787, + 0.015123565681278706, + 0.005090710707008839, + 0.009727745316922665, + 0.008811368606984615, + 0.00861725676804781, + 0.02178376168012619, + 0.009054678492248058, + 0.02381860464811325, + 0.020816680043935776, + 0.00717538595199585, + 0.04836013913154602, + 0.004928069654852152, + 0.003283238038420677, + 0.0031034580897539854, + 0.010262838564813137, + 0.002921334933489561, + 0.00807600375264883, + 0.0015219819033518434, + 0.004366668406873941, + 0.0018618397880345583, + 0.002975590294227004, + 0.005212115123867989, + 0.004591769073158503, + 0.004425603896379471, + 0.008473898284137249, + 0.0020627500489354134, + 0.009268078953027725, + 0.0023597017861902714, + 0.005595098249614239, + 0.014699687249958515, + 0.012365037575364113, + 0.00632764957845211, + 0.018567830324172974, + 0.004475163295865059, + 0.006985425483435392, + 0.004435413051396608, + 0.0033763821702450514, + 0.007207623682916164, + 0.007639355957508087, + 0.0037446485366672277, + 0.010792536661028862, + 0.003177007893100381, + 0.006487300153821707, + 0.0030475349631160498, + 0.009897041134536266, + 0.009244853630661964, + 0.026652559638023376, + 0.00500354403629899, + 0.011227035894989967, + 0.012973237782716751, + 0.0056154499761760235, + 0.005458436440676451, + 0.014330202713608742, + 0.01974506676197052, + 0.04278784990310669, + 0.008230245672166348, + 0.011300893500447273, + 0.008907950483262539, + 0.006208968348801136, + 0.022450776770710945, + 0.018347345292568207, + 0.11465378105640411, + 0.0468527115881443, + 0.013652384281158447, + 0.03375890105962753, + 0.0402020588517189, + 0.007550748065114021, + 0.01135843712836504, + 0.004403778351843357, + 0.00906691886484623, + 0.008002983406186104, + 0.00393622787669301, + 0.011194893158972263, + 0.014195655472576618, + 0.004460124298930168, + 0.006486434955149889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.03742416203022003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761863589286804 + ], + "opposite_logprob_tokens": [ + -0.8261863589286804 + ], + "logprob_reward": -0.5761863589286804, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761863589286804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08281594514846802, + 0.2540946900844574, + 0.1322127878665924, + 0.0461389534175396, + 0.07942793518304825, + 0.03561577573418617, + 0.04954441636800766, + 0.06077249348163605, + 0.17145471274852753, + 0.08792228996753693 + ], + "credit_map": [ + 0.006471829488873482, + 0.014257085509598255, + 0.008624966256320477, + 0.0202806256711483, + 0.010106291621923447, + 0.00968494825065136, + 0.00785135943442583, + 0.00553883844986558, + 0.024355458095669746, + 0.02568000927567482, + 0.018847867846488953, + 0.037195540964603424, + 0.03633978217840195, + 0.013732459396123886, + 0.08839434385299683, + 0.009549233131110668, + 0.011928852647542953, + 0.008708507753908634, + 0.043767303228378296, + 0.008222844451665878, + 0.04356073588132858, + 0.004341796040534973, + 0.008021093904972076, + 0.0036616562865674496, + 0.00309342541731894, + 0.0073018986731767654, + 0.005396049935370684, + 0.005747538525611162, + 0.009756544604897499, + 0.0027234763838350773, + 0.00965262670069933, + 0.0024673917796462774, + 0.009015624411404133, + 0.013830220326781273, + 0.014757603406906128, + 0.008555717766284943, + 0.01376628689467907, + 0.0059967441484332085, + 0.007928846403956413, + 0.0055768839083611965, + 0.0030726115219295025, + 0.006384976673871279, + 0.00429570535197854, + 0.003020354313775897, + 0.009456349536776543, + 0.0033686168026179075, + 0.003394892206415534, + 0.0026222718879580498, + 0.0050409226678311825, + 0.006483149714767933, + 0.015065205283463001, + 0.0028833174146711826, + 0.008069351315498352, + 0.006542196962982416, + 0.0028554005548357964, + 0.002604871056973934, + 0.004144343547523022, + 0.009577534161508083, + 0.01654786430299282, + 0.003523533698171377, + 0.00913956668227911, + 0.010150961577892303, + 0.0044483160600066185, + 0.0032403706572949886, + 0.012072349898517132, + 0.0668754130601883, + 0.03165494650602341, + 0.007972665131092072, + 0.02010723389685154, + 0.01663116179406643, + 0.005240186117589474, + 0.010900753550231457, + 0.0076198577880859375, + 0.013428865000605583, + 0.013794810511171818, + 0.007953856140375137, + 0.01250078808516264, + 0.015743255615234375, + 0.008512976579368114, + 0.008367884904146194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.5761863589286804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487314283847809 + ], + "opposite_logprob_tokens": [ + -1.2237313985824585 + ], + "logprob_reward": -0.3487314283847809, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3487314283847809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0684322640299797, + 0.07322289794683456, + 0.06802235543727875, + 0.11592772603034973, + 0.06366972625255585, + 0.04607192054390907, + 0.030474992468953133, + 0.04616212472319603, + 0.4689703583717346, + 0.01904555968940258 + ], + "credit_map": [ + 0.007956873625516891, + 0.012297914363443851, + 0.006221816409379244, + 0.009171560406684875, + 0.017373504117131233, + 0.005113388877362013, + 0.006971757393330336, + 0.00332545954734087, + 0.004556641448289156, + 0.01233487855643034, + 0.004820479080080986, + 0.00954580307006836, + 0.019077179953455925, + 0.0032125848811119795, + 0.016462888568639755, + 0.0032124475110322237, + 0.006177395582199097, + 0.004765952005982399, + 0.021106746047735214, + 0.005599410273134708, + 0.018223151564598083, + 0.002285938011482358, + 0.007211301010102034, + 0.00265246769413352, + 0.008040016517043114, + 0.015618562698364258, + 0.014121988788247108, + 0.013017496094107628, + 0.022727718576788902, + 0.00735680665820837, + 0.027105074375867844, + 0.007940074428915977, + 0.006030154414474964, + 0.011794830672442913, + 0.013250266201794147, + 0.005219853948801756, + 0.01379694975912571, + 0.0041563804261386395, + 0.005467998329550028, + 0.003953295759856701, + 0.006013134494423866, + 0.006950045935809612, + 0.011853997595608234, + 0.0038389787077903748, + 0.008380613289773464, + 0.002631094306707382, + 0.0036921866703778505, + 0.002711873035877943, + 0.0030376960057765245, + 0.003264406230300665, + 0.007294682320207357, + 0.0018364175921306014, + 0.005250606685876846, + 0.005528988782316446, + 0.002109985798597336, + 0.0021522101014852524, + 0.005457479506731033, + 0.005696420557796955, + 0.00828052219003439, + 0.0032223397865891457, + 0.005278979428112507, + 0.004936108831316233, + 0.003920626826584339, + 0.00936964899301529, + 0.03079341910779476, + 0.18862885236740112, + 0.09145233780145645, + 0.020919298753142357, + 0.05407007783651352, + 0.05338360369205475, + 0.011766908690333366, + 0.01795588620007038, + 0.0019155428744852543, + 0.003248098772019148, + 0.0026241790037602186, + 0.0017546892631798983, + 0.003011539811268449, + 0.003270974848419428, + 0.0014099359977990389, + 0.0018106009811162949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.3487314283847809 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209488958120346 + ], + "opposite_logprob_tokens": [ + -2.8120949268341064 + ], + "logprob_reward": -0.06209488958120346, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209488958120346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057081788778305054, + 0.06704935431480408, + 0.11265730857849121, + 0.062443915754556656, + 0.061991121619939804, + 0.03068837709724903, + 0.040058113634586334, + 0.06094367802143097, + 0.462540864944458, + 0.04454546049237251 + ], + "credit_map": [ + 0.007579838391393423, + 0.008091466501355171, + 0.0037025772035121918, + 0.010586130432784557, + 0.015028972178697586, + 0.0027186076622456312, + 0.006350792944431305, + 0.0030234032310545444, + 0.0046781557612121105, + 0.007962743751704693, + 0.004038064740598202, + 0.008469358086585999, + 0.014018119312822819, + 0.00364458910189569, + 0.021480148658156395, + 0.002758176065981388, + 0.0100557217374444, + 0.00736616924405098, + 0.03820103034377098, + 0.005602553952485323, + 0.036828070878982544, + 0.0032788836397230625, + 0.008021187968552113, + 0.003303696634247899, + 0.004156298004090786, + 0.008528434671461582, + 0.00716613931581378, + 0.006600085645914078, + 0.013234911486506462, + 0.003741378430277109, + 0.015240876004099846, + 0.00377579010091722, + 0.007679910399019718, + 0.009233798831701279, + 0.01657118834555149, + 0.005232317838817835, + 0.010342282243072987, + 0.003992064390331507, + 0.0054208883084356785, + 0.0035186721943318844, + 0.00391559349372983, + 0.004528839141130447, + 0.007354466710239649, + 0.003647042205557227, + 0.004564204718917608, + 0.002211605664342642, + 0.002036159625276923, + 0.002430466702207923, + 0.0036993431858718395, + 0.005762977991253138, + 0.011229204013943672, + 0.00224686274304986, + 0.007939895614981651, + 0.00465672742575407, + 0.0027103882748633623, + 0.0018127172952517867, + 0.0049119796603918076, + 0.020242741331458092, + 0.017869286239147186, + 0.0026808332186192274, + 0.007454911712557077, + 0.0027086278423666954, + 0.0025490254629403353, + 0.0025262683629989624, + 0.028803780674934387, + 0.18526028096675873, + 0.07784070819616318, + 0.019474579021334648, + 0.053760405629873276, + 0.0714576467871666, + 0.010460367426276207, + 0.015483090654015541, + 0.004178707487881184, + 0.007477201987057924, + 0.007166375406086445, + 0.0034259166568517685, + 0.005750194191932678, + 0.011329679749906063, + 0.0022103027440607548, + 0.0030070834327489138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.06209488958120346 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035470366477966 + ], + "opposite_logprob_tokens": [ + -1.910354733467102 + ], + "logprob_reward": -0.16035470366477966, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035470366477966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0533214807510376, + 0.13708506524562836, + 0.03057982586324215, + 0.025734951719641685, + 0.05532205477356911, + 0.040225155651569366, + 0.04689638316631317, + 0.1643868088722229, + 0.44253745675086975, + 0.003910859581083059 + ], + "credit_map": [ + 0.005128642078489065, + 0.011261675506830215, + 0.0037306323647499084, + 0.00848784763365984, + 0.012534034438431263, + 0.003001011675223708, + 0.006693052127957344, + 0.002484588883817196, + 0.012903324328362942, + 0.011667686514556408, + 0.008939377032220364, + 0.014570302329957485, + 0.019675983116030693, + 0.008455431088805199, + 0.05629325658082962, + 0.004579698201268911, + 0.0018988490337505937, + 0.0028254403732717037, + 0.007227766327559948, + 0.002592210890725255, + 0.009626731276512146, + 0.001191752147860825, + 0.0040204222314059734, + 0.00119665393140167, + 0.002816232619807124, + 0.003865557722747326, + 0.003399510635063052, + 0.002988204127177596, + 0.00515935430303216, + 0.001967507181689143, + 0.0032269961666315794, + 0.0023115898948162794, + 0.0052358717657625675, + 0.01120025385171175, + 0.011056669056415558, + 0.0046445457264781, + 0.007746333722025156, + 0.007446896750479937, + 0.003738395171239972, + 0.004253094084560871, + 0.0038985542487353086, + 0.007426354102790356, + 0.00477576395496726, + 0.005612759385257959, + 0.007490117568522692, + 0.003898423397913575, + 0.004639666061848402, + 0.0024835136719048023, + 0.00490908557549119, + 0.008243007585406303, + 0.007598059717565775, + 0.004106130450963974, + 0.007988056167960167, + 0.006824666168540716, + 0.0034393228124827147, + 0.003788057016208768, + 0.018290530890226364, + 0.01759546995162964, + 0.05850448086857796, + 0.010058926418423653, + 0.008569520898163319, + 0.019064251333475113, + 0.007143459282815456, + 0.025160159915685654, + 0.027334552258253098, + 0.17603512108325958, + 0.06984996050596237, + 0.02000199444591999, + 0.0514492429792881, + 0.07122816145420074, + 0.010639200918376446, + 0.01599920727312565, + 0.00032808989635668695, + 0.0005444435519166291, + 0.0006483885226771235, + 0.0002669045061338693, + 0.0006461289594881237, + 0.000626574968919158, + 0.00033794937189668417, + 0.0005123799201101065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.16035470366477966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up triggers a smooth, sustained downward tilt, meticulously tracking a delicate square paper tea bag as it is pulled by gravity from the gleaming white ceramic rim of a teacup. The cup, filled to the brim with crystal-clear, still water, rests on a smooth, dark, non-reflective countertop. As the tea bag gently breaks the water's surface tension with an almost imperceptible ripple, its porous paper texture darkens instantly upon contact. The camera follows its graceful and steady descent through the liquid, capturing the moment when wisps of ethereal cloud, rich amber, begin to gently tumble from the paper, unfolding in slow motion like smoke. The scene is illuminated by soft, diffuse sidelighting, creating subtle highlights on the cup's glaze and causing the water to shimmer, enhancing the atmosphere of tranquility and contemplation. As the tea bag softly settles onto the cup's bottom, the camera's movement ends, while a slow, captivating spread of color continues to bloom outward, transforming the clear water into warm, organic hues.", + "seed": 11192, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_7815_seed_11192_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08532244898378849, + "vlm_logprob": -0.08532244898378849, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.08532244898378849, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055369783192873 + ], + "opposite_logprob_tokens": [ + -3.0553698539733887 + ], + "logprob_reward": -0.055369783192873, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.055369783192873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17736537754535675, + 0.07552355527877808, + 0.1945897489786148, + 0.12762939929962158, + 0.18855153024196625, + 0.05540988966822624, + 0.10111722350120544, + 0.032038334757089615, + 0.023044494912028313, + 0.024730417877435684 + ], + "credit_map": [ + 0.01858336664736271, + 0.0242075826972723, + 0.024417273700237274, + 0.022756528109312057, + 0.017453143373131752, + 0.024487636983394623, + 0.020345577970147133, + 0.025114262476563454, + 0.0076913670636713505, + 0.011749431490898132, + 0.006944199558347464, + 0.011283934116363525, + 0.010899296961724758, + 0.010379262268543243, + 0.010523796081542969, + 0.006052269600331783, + 0.02051834762096405, + 0.02681342139840126, + 0.02178705297410488, + 0.029999149963259697, + 0.023633552715182304, + 0.027823010459542274, + 0.0245676152408123, + 0.019447602331638336, + 0.007280134595930576, + 0.04841168597340584, + 0.012756329961121082, + 0.00954736489802599, + 0.00803615152835846, + 0.024912940338253975, + 0.011605145409703255, + 0.005079643800854683, + 0.009084737859666348, + 0.017820162698626518, + 0.01172754168510437, + 0.01033956091850996, + 0.06194080412387848, + 0.014812562614679337, + 0.013855468481779099, + 0.04897068813443184, + 0.004314095713198185, + 0.005360832903534174, + 0.00798107497394085, + 0.006101152393966913, + 0.006865088827908039, + 0.008476716466248035, + 0.006084022577852011, + 0.010226904414594173, + 0.007570500485599041, + 0.011136136949062347, + 0.01526955422013998, + 0.018184315413236618, + 0.009609650820493698, + 0.014303003437817097, + 0.015384185127913952, + 0.009659882634878159, + 0.0018514186376705766, + 0.0030838539823889732, + 0.005192860495299101, + 0.004165717866271734, + 0.002998372307047248, + 0.005040409043431282, + 0.005262579768896103, + 0.004443120211362839, + 0.001922635012306273, + 0.0034938182216137648, + 0.0029583636205643415, + 0.003846277017146349, + 0.0017754505388438702, + 0.0020944748539477587, + 0.004180251155048609, + 0.0027732241433113813, + 0.001532429363578558, + 0.003848850028589368, + 0.002784163225442171, + 0.003046635305508971, + 0.0023395665921270847, + 0.0024536310229450464, + 0.0050103794783353806, + 0.003714762395247817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.055369783192873 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813466489315033 + ], + "opposite_logprob_tokens": [ + -1.4063466787338257 + ], + "logprob_reward": -0.2813466489315033, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813466489315033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0625588670372963, + 0.18680331110954285, + 0.12921684980392456, + 0.04948537424206734, + 0.34906768798828125, + 0.06591131538152695, + 0.05212700739502907, + 0.035831328481435776, + 0.018743226304650307, + 0.05025501921772957 + ], + "credit_map": [ + 0.0068800258450210094, + 0.00654519023373723, + 0.013004805892705917, + 0.00820609088987112, + 0.005102268885821104, + 0.008312956430017948, + 0.005957018118351698, + 0.008550509810447693, + 0.017283374443650246, + 0.024833591654896736, + 0.021962277591228485, + 0.028316190466284752, + 0.02534802258014679, + 0.024971086531877518, + 0.02623075246810913, + 0.017858000472187996, + 0.01231166161596775, + 0.01489195879548788, + 0.0156620591878891, + 0.021544765681028366, + 0.01744852587580681, + 0.01950168050825596, + 0.016680359840393066, + 0.011175829917192459, + 0.0032046816777437925, + 0.00641234964132309, + 0.008436978794634342, + 0.008179687894880772, + 0.004492639563977718, + 0.004949019756168127, + 0.008661146275699139, + 0.005148869939148426, + 0.016814863309264183, + 0.03403093293309212, + 0.02326720580458641, + 0.010427631437778473, + 0.11752549558877945, + 0.0267802681773901, + 0.015999402850866318, + 0.1042218953371048, + 0.0048714131116867065, + 0.006431648042052984, + 0.012532034888863564, + 0.008047031238675117, + 0.006198596674948931, + 0.01005236990749836, + 0.011986659839749336, + 0.005791559349745512, + 0.004383876919746399, + 0.007946748286485672, + 0.007340754847973585, + 0.008390642702579498, + 0.005720275919884443, + 0.0062734331004321575, + 0.006761580239981413, + 0.005309693049639463, + 0.0023607774637639523, + 0.003209509188309312, + 0.004581781104207039, + 0.0038786809891462326, + 0.004732829984277487, + 0.007280351128429174, + 0.00385654135607183, + 0.005930854007601738, + 0.00103876949287951, + 0.0024132949765771627, + 0.001603605691343546, + 0.0022308623883873224, + 0.0011922995327040553, + 0.00244902609847486, + 0.0032147113233804703, + 0.004600656684488058, + 0.003607620485126972, + 0.008705560117959976, + 0.006224646233022213, + 0.0072238328866660595, + 0.004627284128218889, + 0.004835178144276142, + 0.007727870251983404, + 0.007303023710846901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.2813466489315033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029845761135220528 + ], + "opposite_logprob_tokens": [ + -3.529845714569092 + ], + "logprob_reward": -0.029845761135220528, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029845761135220528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12295544892549515, + 0.12950430810451508, + 0.2381831407546997, + 0.17960618436336517, + 0.10770707577466965, + 0.08402939885854721, + 0.03845193609595299, + 0.04441067948937416, + 0.009762700647115707, + 0.045389167964458466 + ], + "credit_map": [ + 0.011462870053946972, + 0.022141393274068832, + 0.01749933324754238, + 0.016339752823114395, + 0.014936594292521477, + 0.01527833379805088, + 0.013738210313022137, + 0.011558949947357178, + 0.012283770367503166, + 0.01880967803299427, + 0.013314889743924141, + 0.017801031470298767, + 0.019618632271885872, + 0.017637871205806732, + 0.01885865069925785, + 0.011179766617715359, + 0.025132467970252037, + 0.029887672513723373, + 0.03244191035628319, + 0.03897040709853172, + 0.027691243216395378, + 0.03459841012954712, + 0.02729417011141777, + 0.022166859358549118, + 0.011466228403151035, + 0.05380592495203018, + 0.019398821517825127, + 0.018348071724176407, + 0.019926201552152634, + 0.027667125687003136, + 0.019209500402212143, + 0.009784294292330742, + 0.004765471909195185, + 0.00963344145566225, + 0.007905518636107445, + 0.0067837052047252655, + 0.030764291062951088, + 0.009135771542787552, + 0.009745124727487564, + 0.028973745182156563, + 0.00537229236215353, + 0.007970968261361122, + 0.017843306064605713, + 0.012551967985928059, + 0.006707426160573959, + 0.008941124193370342, + 0.016058119013905525, + 0.00858418457210064, + 0.0022750124335289, + 0.005032721906900406, + 0.004738541319966316, + 0.006444770842790604, + 0.004455301444977522, + 0.0056249285116791725, + 0.005229054018855095, + 0.004651600494980812, + 0.00268641603179276, + 0.005187259521335363, + 0.005710483528673649, + 0.005944550037384033, + 0.004525063093751669, + 0.004724373575299978, + 0.00823876354843378, + 0.007393767591565847, + 0.0005673982668668032, + 0.0013308344641700387, + 0.0010427451925352216, + 0.001849954598583281, + 0.0011355720926076174, + 0.0008101824205368757, + 0.00201626680791378, + 0.0010097461054101586, + 0.003377129090949893, + 0.006426100153476, + 0.004630230367183685, + 0.0065068453550338745, + 0.004075835924595594, + 0.004908926319330931, + 0.007260417565703392, + 0.008203678764402866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.029845761135220528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636094018816948 + ], + "opposite_logprob_tokens": [ + -3.6513609886169434 + ], + "logprob_reward": -0.02636094018816948, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636094018816948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11297278106212616, + 0.06766761094331741, + 0.2164720892906189, + 0.13938963413238525, + 0.19136781990528107, + 0.044832371175289154, + 0.057786013931035995, + 0.02588696964085102, + 0.035703737288713455, + 0.10792098194360733 + ], + "credit_map": [ + 0.01348369475454092, + 0.011724523268640041, + 0.020553255453705788, + 0.01997196488082409, + 0.008397716097533703, + 0.014988911338150501, + 0.01079201977699995, + 0.013060701079666615, + 0.006705218926072121, + 0.01027767639607191, + 0.008820528164505959, + 0.010536255314946175, + 0.009144287556409836, + 0.008162742480635643, + 0.007677996996790171, + 0.006342902313917875, + 0.025619512423872948, + 0.02422083355486393, + 0.028679605573415756, + 0.036054112017154694, + 0.02456551603972912, + 0.03287774324417114, + 0.02514527179300785, + 0.019309494644403458, + 0.008493310771882534, + 0.04596090316772461, + 0.013114229775965214, + 0.013655953109264374, + 0.013959111645817757, + 0.02409188635647297, + 0.01201201044023037, + 0.00810223538428545, + 0.015490571968257427, + 0.0195187795907259, + 0.020008916035294533, + 0.024848410859704018, + 0.033996470272541046, + 0.01654292456805706, + 0.03493683040142059, + 0.02602490969002247, + 0.003580391174182296, + 0.004678539931774139, + 0.007209559436887503, + 0.00559379905462265, + 0.004341957159340382, + 0.005925269797444344, + 0.007238894701004028, + 0.006263955496251583, + 0.004520314745604992, + 0.008070019073784351, + 0.006395232398062944, + 0.00785784237086773, + 0.005964886397123337, + 0.012491866946220398, + 0.005277066025882959, + 0.007208784110844135, + 0.0018189033726230264, + 0.003615759080275893, + 0.0039059072732925415, + 0.004096112679690123, + 0.0019698229152709246, + 0.003921644762158394, + 0.004051378928124905, + 0.002507438650354743, + 0.001817538752220571, + 0.005367694888263941, + 0.0037437533028423786, + 0.003273164154961705, + 0.0024435999803245068, + 0.005568561144173145, + 0.004602702800184488, + 0.00888671912252903, + 0.005437343847006559, + 0.017725562676787376, + 0.010510730557143688, + 0.015577108599245548, + 0.008537054993212223, + 0.008328505791723728, + 0.02396542578935623, + 0.01783924549818039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.02636094018816948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368911147117615 + ], + "opposite_logprob_tokens": [ + -3.408689022064209 + ], + "logprob_reward": -0.03368911147117615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368911147117615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12313300371170044, + 0.04280897229909897, + 0.13949206471443176, + 0.05951394513249397, + 0.2946833372116089, + 0.15591269731521606, + 0.15184178948402405, + 0.009012878872454166, + 0.004766551777720451, + 0.018834665417671204 + ], + "credit_map": [ + 0.01288076676428318, + 0.014354463666677475, + 0.02472662925720215, + 0.015923090279102325, + 0.00973832793533802, + 0.016702482476830482, + 0.012967011891305447, + 0.015840228646993637, + 0.004597769118845463, + 0.004537173546850681, + 0.006009979173541069, + 0.007062495220452547, + 0.00466503482311964, + 0.006187742110341787, + 0.005880302283912897, + 0.003868476953357458, + 0.014711171388626099, + 0.017556002363562584, + 0.018439076840877533, + 0.02190839685499668, + 0.016322091221809387, + 0.02003939263522625, + 0.017332905903458595, + 0.013183025643229485, + 0.0035150921903550625, + 0.013840125873684883, + 0.007924114353954792, + 0.0078095123171806335, + 0.006465132813900709, + 0.007619816809892654, + 0.007821294479072094, + 0.0045188600197434425, + 0.014949128031730652, + 0.028998900204896927, + 0.019550511613488197, + 0.010277841240167618, + 0.0979299247264862, + 0.02209261991083622, + 0.017901824787259102, + 0.08298259228467941, + 0.011008298955857754, + 0.016524525359272957, + 0.029705584049224854, + 0.01572311669588089, + 0.014783020131289959, + 0.01795487105846405, + 0.025899630039930344, + 0.024313660338521004, + 0.010480721481144428, + 0.01572573371231556, + 0.023567605763673782, + 0.023246655240654945, + 0.013376239687204361, + 0.03001801297068596, + 0.02050243690609932, + 0.014924395829439163, + 0.0004850878322031349, + 0.0006552473059855402, + 0.0007575438357889652, + 0.0009698532521724701, + 0.0006526021170429885, + 0.0028512850403785706, + 0.0009298439836129546, + 0.0017114155925810337, + 0.0002861888788174838, + 0.0007962770177982748, + 0.0005817254423163831, + 0.00060887640574947, + 0.00048658117884770036, + 0.0004936102195642889, + 0.0007247626199387014, + 0.0007885300437919796, + 0.001111225807107985, + 0.003457496175542474, + 0.0018417686223983765, + 0.0033326074481010437, + 0.0019444876816123724, + 0.0014444144908338785, + 0.004195971880108118, + 0.0015066942432895303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.03368911147117615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up triggers a smooth, sustained downward tilt, meticulously tracking a delicate square paper tea bag as it is pulled by gravity from the gleaming white ceramic rim of a teacup. The cup, filled to the brim with crystal-clear, still water, rests on a smooth, dark, non-reflective countertop. As the tea bag gently breaks the water's surface tension with an almost imperceptible ripple, its porous paper texture darkens instantly upon contact. The camera follows its graceful and steady descent through the liquid, capturing the moment when wisps of ethereal cloud, rich amber, begin to gently tumble from the paper, unfolding in slow motion like smoke. The scene is illuminated by soft, diffuse sidelighting, creating subtle highlights on the cup's glaze and causing the water to shimmer, enhancing the atmosphere of tranquility and contemplation. As the tea bag softly settles onto the cup's bottom, the camera's movement ends, while a slow, captivating spread of color continues to bloom outward, transforming the clear water into warm, organic hues.", + "seed": 11193, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_13/index_7815_seed_11193_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0861473560333252, + "vlm_logprob": -0.0861473560333252, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.0861473560333252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07054818421602249 + ], + "opposite_logprob_tokens": [ + -2.8205482959747314 + ], + "logprob_reward": -0.07054818421602249, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07054818421602249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30734822154045105, + 0.08413428068161011, + 0.09436187893152237, + 0.030627833679318428, + 0.13912811875343323, + 0.13486289978027344, + 0.057185981422662735, + 0.03617953136563301, + 0.08753025531768799, + 0.028640955686569214 + ], + "credit_map": [ + 0.013387508690357208, + 0.039404209703207016, + 0.017096925526857376, + 0.04295695573091507, + 0.07443448901176453, + 0.045995984226465225, + 0.0373539924621582, + 0.036718159914016724, + 0.0043444097973406315, + 0.018335504457354546, + 0.0032201812136918306, + 0.009103949181735516, + 0.011904684826731682, + 0.02111412212252617, + 0.00998683925718069, + 0.006124600768089294, + 0.00699601648375392, + 0.012511477805674076, + 0.00500252190977335, + 0.009692540392279625, + 0.02196487970650196, + 0.01669378951191902, + 0.01041088905185461, + 0.011089772917330265, + 0.003819456323981285, + 0.005073730833828449, + 0.002127900021150708, + 0.003217541379854083, + 0.0056083425879478455, + 0.003758147358894348, + 0.0038026475813239813, + 0.0032200701534748077, + 0.012749159708619118, + 0.01844193786382675, + 0.006018084939569235, + 0.013273661024868488, + 0.02614210918545723, + 0.02589503489434719, + 0.010633973404765129, + 0.025974158197641373, + 0.010075640864670277, + 0.019271276891231537, + 0.006027700379490852, + 0.017749380320310593, + 0.035519909113645554, + 0.021573424339294434, + 0.01686432957649231, + 0.00778124388307333, + 0.010237134993076324, + 0.00717853894457221, + 0.007922809571027756, + 0.005027069244533777, + 0.005999765358865261, + 0.01111410278826952, + 0.0032461562659591436, + 0.006460410077124834, + 0.003490823321044445, + 0.007053293287754059, + 0.003449885407462716, + 0.0044400799088180065, + 0.004765840247273445, + 0.007000673562288284, + 0.0024597428273409605, + 0.0035191953647881746, + 0.02215277962386608, + 0.008706008084118366, + 0.006418110337108374, + 0.0050716884434223175, + 0.022220266982913017, + 0.01195901446044445, + 0.007054908201098442, + 0.0039474861696362495, + 0.003630740335211158, + 0.004134147893637419, + 0.0032013263553380966, + 0.002325783483684063, + 0.0034941283520311117, + 0.004583760164678097, + 0.004672156181186438, + 0.0025989150162786245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.07054818421602249 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604098379611969 + ], + "opposite_logprob_tokens": [ + -1.9104098081588745 + ], + "logprob_reward": -0.1604098379611969, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604098379611969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22224177420139313, + 0.10119472444057465, + 0.041773367673158646, + 0.11783457547426224, + 0.19639050960540771, + 0.11471281945705414, + 0.08421583473682404, + 0.03337367996573448, + 0.04496015980839729, + 0.043302565813064575 + ], + "credit_map": [ + 0.01301712915301323, + 0.03299587592482567, + 0.015557309612631798, + 0.029562076553702354, + 0.05473951995372772, + 0.02800983004271984, + 0.027195168659090996, + 0.021164851263165474, + 0.005069593898952007, + 0.021740002557635307, + 0.005554885137826204, + 0.015274430625140667, + 0.01645762287080288, + 0.013480317778885365, + 0.014367944560945034, + 0.009249921888113022, + 0.0019067913526669145, + 0.008026445284485817, + 0.0030695132445544004, + 0.008357144892215729, + 0.004773351829499006, + 0.007656929548829794, + 0.005303576588630676, + 0.0026796148158609867, + 0.010772602632641792, + 0.01893903687596321, + 0.006561061833053827, + 0.013349547982215881, + 0.02212437428534031, + 0.015133421868085861, + 0.01868821121752262, + 0.012266320176422596, + 0.011317742988467216, + 0.028786147013306618, + 0.007137780077755451, + 0.020220326259732246, + 0.040130265057086945, + 0.04733170196413994, + 0.02491849660873413, + 0.01654803566634655, + 0.01125300768762827, + 0.014161423780024052, + 0.008276966400444508, + 0.0136653957888484, + 0.027400631457567215, + 0.016284920275211334, + 0.014571667648851871, + 0.009098803624510765, + 0.014315665699541569, + 0.010182037949562073, + 0.007320529315620661, + 0.006330825854092836, + 0.012600712478160858, + 0.013655165210366249, + 0.005246650893241167, + 0.014564240351319313, + 0.003600891912356019, + 0.006536419037729502, + 0.002251198049634695, + 0.004062797874212265, + 0.005513311363756657, + 0.005847596097737551, + 0.002568725496530533, + 0.0029927403666079044, + 0.006240927614271641, + 0.004380516242235899, + 0.0029502343386411667, + 0.002315692836418748, + 0.013217288069427013, + 0.005071369931101799, + 0.003979981876909733, + 0.0068041482008993626, + 0.003449575277045369, + 0.0036722731310874224, + 0.0025618011131882668, + 0.002091732807457447, + 0.012081917375326157, + 0.004724573343992233, + 0.01135908905416727, + 0.0033616027794778347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.1604098379611969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299815028905869 + ], + "opposite_logprob_tokens": [ + -2.2379982471466064 + ], + "logprob_reward": -0.11299815028905869, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299815028905869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11919889599084854, + 0.08288534730672836, + 0.11905614286661148, + 0.06325336545705795, + 0.184254452586174, + 0.051304467022418976, + 0.0537734180688858, + 0.03891102969646454, + 0.2595231235027313, + 0.02783975377678871 + ], + "credit_map": [ + 0.007184439804404974, + 0.015372046269476414, + 0.009777538478374481, + 0.01900644227862358, + 0.02452082931995392, + 0.016405055299401283, + 0.01565130427479744, + 0.011281238868832588, + 0.003956095315515995, + 0.01098302286118269, + 0.0037113195285201073, + 0.01067088358104229, + 0.023404724895954132, + 0.01157528255134821, + 0.012095483019948006, + 0.006488528568297625, + 0.009339462034404278, + 0.016610411927103996, + 0.007241010665893555, + 0.016799170523881912, + 0.018101532012224197, + 0.01841931976377964, + 0.018611526116728783, + 0.013933712616562843, + 0.005553568713366985, + 0.008956129662692547, + 0.0039020366966724396, + 0.008397724479436874, + 0.011389960534870625, + 0.006673729047179222, + 0.010561419650912285, + 0.00781879760324955, + 0.020970381796360016, + 0.026741717010736465, + 0.008301254361867905, + 0.01798912324011326, + 0.02287842333316803, + 0.04179420694708824, + 0.009203587658703327, + 0.036375757306814194, + 0.006933895871043205, + 0.008431913331151009, + 0.0024022646248340607, + 0.007212322670966387, + 0.008865635842084885, + 0.008899715729057789, + 0.004274455830454826, + 0.004284262657165527, + 0.008717352524399757, + 0.005865895189344883, + 0.0031137149780988693, + 0.004953842144459486, + 0.009249032475054264, + 0.01097111776471138, + 0.0032184128649532795, + 0.007684055250138044, + 0.006552719045430422, + 0.00511952955275774, + 0.003980638924986124, + 0.004020510241389275, + 0.005660150200128555, + 0.004958861041814089, + 0.003234198549762368, + 0.00538442051038146, + 0.023718826472759247, + 0.01520242914557457, + 0.009795860387384892, + 0.01200451236218214, + 0.08967886120080948, + 0.02263582870364189, + 0.07616184651851654, + 0.01032498013228178, + 0.0035189297050237656, + 0.003799035679548979, + 0.002215897897258401, + 0.0016137990169227123, + 0.007972072809934616, + 0.003990450408309698, + 0.0020852878224104643, + 0.002644282067194581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.11299815028905869 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813275694847107 + ], + "opposite_logprob_tokens": [ + -3.288132667541504 + ], + "logprob_reward": -0.03813275694847107, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813275694847107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06776045262813568, + 0.05366428941488266, + 0.04676825553178787, + 0.024017170071601868, + 0.23137685656547546, + 0.058463919907808304, + 0.07061965018510818, + 0.038068365305662155, + 0.3510964512825012, + 0.05816453695297241 + ], + "credit_map": [ + 0.003903680481016636, + 0.009492884390056133, + 0.004465511068701744, + 0.009675582870841026, + 0.011851278133690357, + 0.009147188626229763, + 0.011610426940023899, + 0.007613897789269686, + 0.0022061083000153303, + 0.007822036743164062, + 0.0025511288549751043, + 0.007759427651762962, + 0.010609584860503674, + 0.009434969164431095, + 0.0077195074409246445, + 0.0055615259334445, + 0.003157456172630191, + 0.006962715648114681, + 0.0022694033104926348, + 0.006984603125602007, + 0.006737655494362116, + 0.007800735533237457, + 0.006916294805705547, + 0.005939388647675514, + 0.0021344025153666735, + 0.004127699416130781, + 0.0016486946260556579, + 0.00368140428327024, + 0.0030035190284252167, + 0.0039472635835409164, + 0.003002431709319353, + 0.0024717545602470636, + 0.02281549572944641, + 0.03472670912742615, + 0.013967601582407951, + 0.020822638645768166, + 0.03170140087604523, + 0.04922477900981903, + 0.01798742450773716, + 0.040130797773599625, + 0.005595462862402201, + 0.00947924517095089, + 0.004775074776262045, + 0.005509089212864637, + 0.008442132733762264, + 0.012087657116353512, + 0.004597689025104046, + 0.007977575063705444, + 0.00919756293296814, + 0.0071245520375669, + 0.00624833395704627, + 0.006930596195161343, + 0.009506694972515106, + 0.015979750081896782, + 0.0037049604579806328, + 0.01192720141261816, + 0.004765631165355444, + 0.00458252290263772, + 0.0024880922865122557, + 0.0042412602342665195, + 0.007450989913195372, + 0.006839926820248365, + 0.003739750012755394, + 0.003960192669183016, + 0.031227802857756615, + 0.026945263147354126, + 0.010287188924849033, + 0.016753729432821274, + 0.10941038280725479, + 0.02620747685432434, + 0.11168338358402252, + 0.018581228330731392, + 0.0063703530468046665, + 0.007164215203374624, + 0.0033683849032968283, + 0.0026895361952483654, + 0.01769484579563141, + 0.0059289890341460705, + 0.011605937965214252, + 0.003342280164361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.03813275694847107 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a square paper tea bag being pulled from a teacup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the teacup made of gleaming white ceramic and filled to the brim with crystal-clear still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tea bag break the water's surface tension creating an almost imperceptible ripple and darken instantly upon contact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of ethereal cloud-like amber color tumble from the paper and bloom outward in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up and perform a smooth, sustained downward tilt tracking the tea bag?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864785075187683 + ], + "opposite_logprob_tokens": [ + -3.048647880554199 + ], + "logprob_reward": -0.04864785075187683, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864785075187683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13514350354671478, + 0.05059316009283066, + 0.047164130955934525, + 0.04534425586462021, + 0.19767551124095917, + 0.3267526626586914, + 0.06575356423854828, + 0.025791915133595467, + 0.06727341562509537, + 0.0385078564286232 + ], + "credit_map": [ + 0.008687933906912804, + 0.019753722473978996, + 0.009963120333850384, + 0.02150711789727211, + 0.02548106759786606, + 0.01785864122211933, + 0.016676804050803185, + 0.015215094201266766, + 0.0025239381939172745, + 0.006747452076524496, + 0.002564572263509035, + 0.005740670021623373, + 0.015330606140196323, + 0.0072145904414355755, + 0.006108902394771576, + 0.004362431354820728, + 0.0025462869089096785, + 0.007970195263624191, + 0.0025349915958940983, + 0.004174469038844109, + 0.01082062628120184, + 0.00993182323873043, + 0.005444663111120462, + 0.003741075284779072, + 0.005945071578025818, + 0.006587647832930088, + 0.003319243900477886, + 0.005435120780020952, + 0.0075761983171105385, + 0.005734685808420181, + 0.006301638204604387, + 0.004444651305675507, + 0.01346502359956503, + 0.024831078946590424, + 0.012579222209751606, + 0.030164320021867752, + 0.03718132525682449, + 0.037204720079898834, + 0.027666782960295677, + 0.014583028852939606, + 0.024663956835865974, + 0.040778953582048416, + 0.015894826501607895, + 0.040755417197942734, + 0.0892980620265007, + 0.04865376651287079, + 0.046390730887651443, + 0.020316945388913155, + 0.008854251354932785, + 0.011113901622593403, + 0.006272632163017988, + 0.00581736396998167, + 0.010468914173543453, + 0.010795962996780872, + 0.004111281130462885, + 0.008319249376654625, + 0.0038029819261282682, + 0.0034123731311410666, + 0.0018434242811053991, + 0.0019163198303431273, + 0.004416880663484335, + 0.003831783775240183, + 0.0022513943258672953, + 0.004316757433116436, + 0.012694770470261574, + 0.011379890143871307, + 0.005394486244767904, + 0.005066931247711182, + 0.014343899674713612, + 0.007257352117449045, + 0.00804090965539217, + 0.003095174441114068, + 0.0033665576484054327, + 0.004157745745033026, + 0.002384521998465061, + 0.0019142659148201346, + 0.01181778684258461, + 0.004500138107687235, + 0.007115741726011038, + 0.0032510969322174788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.04864785075187683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_14/index_tensor([7438, 7438])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_14/index_tensor([7438, 7438])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..64bdfd7111469da5731047a574faca07cea2b6b4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_14/index_tensor([7438, 7438])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium-long shot, shot from a slightly upward angle, slowly pulls out as the subject moves, capturing a fairy in ancient costume on a grand palace terrace. Cool, watery moonlight pours down, casting a holy silver glow on the smooth, carved white marble floor, illuminating a faint silhouette. She stands slender, her long, jet-black hair held together by a simple jade hairpin. Her skin, whiter than snow in the moonlight, reflects a serene expression with a touch of coolness. She wears a pure white, wide-sleeved, flowing fairy dress. The sheer silk flutters gently in the night breeze, and the delicate cloud patterns embroidered with silver thread on the cuffs and hem shimmer with tiny glimmers of light. She slowly raises her slender arm, then twists her wrist gently, steadily and deliberately swinging the long sleeves out gracefully. The wide sleeves, like a flowing white silk ribbon, trace a graceful, lingering arc across the deep night sky, as if silently telling a story a thousand years old. Behind her are misty palaces and a starry night sky, with a bright full moon hanging high in the sky. The whole picture, in cold tones, presents an ultimate tranquility, ethereality and dreamy classical beauty.", + "seed": 6827, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_14/index_7438_seed_6827_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01980085545219481, + "vlm_logprob": -0.01980085545219481, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.01980085545219481, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024905283004045486 + ], + "opposite_logprob_tokens": [ + -4.024905204772949 + ], + "logprob_reward": -0.024905283004045486, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024905283004045486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1031406819820404, + 0.07543648779392242, + 0.08396698534488678, + 0.04507523402571678, + 0.11073478311300278, + 0.2674770653247833, + 0.03409023582935333, + 0.08066552877426147, + 0.043863046914339066, + 0.15554995834827423 + ], + "credit_map": [ + 0.010697020217776299, + 0.00872866902500391, + 0.014299124479293823, + 0.022616975009441376, + 0.008204077370464802, + 0.008531623519957066, + 0.014768671244382858, + 0.015294517390429974, + 0.01526803057640791, + 0.006749001797288656, + 0.00921628437936306, + 0.006567482836544514, + 0.008651801384985447, + 0.0068788183853030205, + 0.007496529724448919, + 0.014608537778258324, + 0.006419048178941011, + 0.008223970420658588, + 0.009566827677190304, + 0.01414488721638918, + 0.006497494410723448, + 0.009619894437491894, + 0.01835177093744278, + 0.011143087409436703, + 0.004116448108106852, + 0.004572518635541201, + 0.008204624988138676, + 0.006110531743615866, + 0.003728005336597562, + 0.005398410838097334, + 0.006775870453566313, + 0.006168820895254612, + 0.008093093521893024, + 0.014459170401096344, + 0.021100185811519623, + 0.019711507484316826, + 0.011209236457943916, + 0.010247213765978813, + 0.013295207172632217, + 0.01261916570365429, + 0.01620921865105629, + 0.017856115475296974, + 0.054740723222494125, + 0.029880858957767487, + 0.01094083022326231, + 0.014275776222348213, + 0.10518988966941833, + 0.018383659422397614, + 0.005605390761047602, + 0.0032884960528463125, + 0.003096485510468483, + 0.005292082671076059, + 0.003338389564305544, + 0.0028357477858662605, + 0.006640914361923933, + 0.00399272795766592, + 0.00891154445707798, + 0.005343545228242874, + 0.008392410352826118, + 0.018889645114541054, + 0.010866833850741386, + 0.010500015690922737, + 0.010012495331466198, + 0.007749043870717287, + 0.006113313138484955, + 0.0038441189099103212, + 0.005737970117479563, + 0.007875286042690277, + 0.004341519437730312, + 0.005911638960242271, + 0.005385396536439657, + 0.00465380446985364, + 0.03731715306639671, + 0.006083420477807522, + 0.008390726521611214, + 0.018969792872667313, + 0.007604272570461035, + 0.05484001338481903, + 0.012808107770979404, + 0.009536470286548138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.024905283004045486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264111328870058 + ], + "opposite_logprob_tokens": [ + -5.2552642822265625 + ], + "logprob_reward": -0.005264111328870058, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264111328870058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11815019696950912, + 0.5727428197860718, + 0.06793423742055893, + 0.021933339536190033, + 0.020496390759944916, + 0.08999043703079224, + 0.032454609870910645, + 0.02772882953286171, + 0.012200501747429371, + 0.03636864945292473 + ], + "credit_map": [ + 0.019212640821933746, + 0.010852952487766743, + 0.012914324179291725, + 0.01834939233958721, + 0.015648657456040382, + 0.012968942523002625, + 0.013949660584330559, + 0.014253640547394753, + 0.09048233181238174, + 0.04895372688770294, + 0.07157864421606064, + 0.06976436823606491, + 0.049251489341259, + 0.05468907952308655, + 0.06904055178165436, + 0.11898262053728104, + 0.00510916393250227, + 0.005970761179924011, + 0.011233292520046234, + 0.011747852899134159, + 0.004576732870191336, + 0.007875449024140835, + 0.010745549574494362, + 0.010675444267690182, + 0.0028205118142068386, + 0.003458307357504964, + 0.0023101367987692356, + 0.0026931024622172117, + 0.0016502379439771175, + 0.003492670366540551, + 0.002944149076938629, + 0.0025642262771725655, + 0.0025606921408325434, + 0.002233930164948106, + 0.001672264188528061, + 0.003393827937543392, + 0.0015749564627185464, + 0.003244226798415184, + 0.0026313101407140493, + 0.003185184672474861, + 0.005223722662776709, + 0.016082903370261192, + 0.005569960922002792, + 0.009305987507104874, + 0.005261596757918596, + 0.010268138721585274, + 0.03133489191532135, + 0.006943248212337494, + 0.0026626233011484146, + 0.0019271564669907093, + 0.0037461116444319487, + 0.006316800136119127, + 0.001969667850062251, + 0.0023917541839182377, + 0.010646452195942402, + 0.0027940489817410707, + 0.0042734285816550255, + 0.0023790227714926004, + 0.002585396636277437, + 0.007026454899460077, + 0.0026323215570300817, + 0.0030606347136199474, + 0.0025451125111430883, + 0.00322646158747375, + 0.0015620142221450806, + 0.0011918442323803902, + 0.0015585038345307112, + 0.00197978806681931, + 0.0016531497240066528, + 0.001542505924589932, + 0.001592283952049911, + 0.0011204129550606012, + 0.0072256713174283504, + 0.001659780158661306, + 0.0031547031830996275, + 0.011827169917523861, + 0.0018449118360877037, + 0.005338781047612429, + 0.0033636586740612984, + 0.0019539762288331985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.005264111328870058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003622637828812003 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.003622637828812003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003622637828812003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09810250997543335, + 0.06697153300046921, + 0.09721849858760834, + 0.10347709059715271, + 0.04710992053151131, + 0.14710751175880432, + 0.2764887809753418, + 0.039960358291864395, + 0.028457757085561752, + 0.09510604292154312 + ], + "credit_map": [ + 0.016082391142845154, + 0.008345468901097775, + 0.012932923622429371, + 0.014926307834684849, + 0.009442958980798721, + 0.010359596461057663, + 0.01134407613426447, + 0.014668789692223072, + 0.011972972191870213, + 0.005342351272702217, + 0.007288523018360138, + 0.0087040476500988, + 0.005629346705973148, + 0.006320226937532425, + 0.011152113787829876, + 0.010561948642134666, + 0.0074273995123803616, + 0.015822380781173706, + 0.010722511447966099, + 0.011968232691287994, + 0.0077911741100251675, + 0.010316097177565098, + 0.018379973247647285, + 0.014790719375014305, + 0.005741751752793789, + 0.005285579711198807, + 0.03543923422694206, + 0.009220529347658157, + 0.002951792674139142, + 0.006842532195150852, + 0.007536812219768763, + 0.030458850786089897, + 0.0034249192103743553, + 0.0056453063152730465, + 0.006304359994828701, + 0.010232483968138695, + 0.004653106443583965, + 0.0047631883062422276, + 0.007523463107645512, + 0.004563089460134506, + 0.009092791937291622, + 0.02845173515379429, + 0.008990869857370853, + 0.01596789062023163, + 0.007695864420384169, + 0.012474514544010162, + 0.04775450751185417, + 0.016679340973496437, + 0.022279947996139526, + 0.010888054035604, + 0.039497677236795425, + 0.04282410815358162, + 0.014569789171218872, + 0.013561169616878033, + 0.11215578019618988, + 0.020712273195385933, + 0.006725249346345663, + 0.004943074658513069, + 0.00471187848597765, + 0.007698227651417255, + 0.0035541681572794914, + 0.003930227365344763, + 0.0040195598267018795, + 0.004377972334623337, + 0.0027425929438322783, + 0.0015791698824614286, + 0.0030286386609077454, + 0.0041303569450974464, + 0.0036621636245399714, + 0.004072732292115688, + 0.006573559250682592, + 0.0026685441844165325, + 0.018710657954216003, + 0.006858421489596367, + 0.008415250107645988, + 0.013000081293284893, + 0.004717128351330757, + 0.028508638963103294, + 0.008869382552802563, + 0.006026474293321371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.003622637828812003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189953975379467 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003189953975379467, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189953975379467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06151015684008598, + 0.06500545889139175, + 0.026457827538251877, + 0.033500734716653824, + 0.026735275983810425, + 0.07883374392986298, + 0.5733822584152222, + 0.04264608398079872, + 0.03080923855304718, + 0.061119191348552704 + ], + "credit_map": [ + 0.008202302269637585, + 0.0061898487620055676, + 0.007379279471933842, + 0.00909732561558485, + 0.006759743671864271, + 0.005867328029125929, + 0.00894990935921669, + 0.009064415469765663, + 0.009293941780924797, + 0.006602734327316284, + 0.008194801397621632, + 0.00950972642749548, + 0.005461962427943945, + 0.005669743288308382, + 0.008073023520410061, + 0.012199530377984047, + 0.002598352963104844, + 0.0023823806550353765, + 0.003279708558693528, + 0.004089914262294769, + 0.002075230237096548, + 0.0034818495623767376, + 0.0043487995862960815, + 0.0042015910148620605, + 0.004515917971730232, + 0.003998450934886932, + 0.006942791398614645, + 0.0033238439355045557, + 0.0017794485902413726, + 0.004590269643813372, + 0.0024410258047282696, + 0.005908987484872341, + 0.002548521850258112, + 0.003626164048910141, + 0.00502674700692296, + 0.004616278689354658, + 0.0020021169912070036, + 0.0023565238807350397, + 0.0038360462058335543, + 0.0027228747494518757, + 0.004614407196640968, + 0.005637489724904299, + 0.012754878960549831, + 0.009359409101307392, + 0.0036200983449816704, + 0.005261737387627363, + 0.030691752210259438, + 0.0068939751945436, + 0.04167388379573822, + 0.020805029198527336, + 0.048133861273527145, + 0.132150337100029, + 0.028479471802711487, + 0.02822010964155197, + 0.23594385385513306, + 0.03797569125890732, + 0.007311077788472176, + 0.004489234648644924, + 0.0028900913894176483, + 0.015995807945728302, + 0.0023007416166365147, + 0.004070463124662638, + 0.002864506794139743, + 0.0027241611387580633, + 0.00302456715144217, + 0.001467867405153811, + 0.003256449243053794, + 0.004018801264464855, + 0.0049618808552622795, + 0.004519824404269457, + 0.005865276791155338, + 0.0036945685278624296, + 0.0051388805732131, + 0.0021561479661613703, + 0.0029802080243825912, + 0.019150588661432266, + 0.0019705910235643387, + 0.023076627403497696, + 0.0030848640017211437, + 0.003561283927410841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.003189953975379467 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062022291123867035 + ], + "opposite_logprob_tokens": [ + -2.8120222091674805 + ], + "logprob_reward": -0.062022291123867035, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062022291123867035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16718269884586334, + 0.11968855559825897, + 0.12565255165100098, + 0.11324474215507507, + 0.07085957378149033, + 0.22698336839675903, + 0.04697071388363838, + 0.03869778290390968, + 0.03719516843557358, + 0.053524792194366455 + ], + "credit_map": [ + 0.027940044179558754, + 0.015296883881092072, + 0.019625434651970863, + 0.0219816192984581, + 0.0156580600887537, + 0.019700879231095314, + 0.01973726600408554, + 0.027242522686719894, + 0.010517771355807781, + 0.010972576215863228, + 0.016318533569574356, + 0.01920769363641739, + 0.011530166491866112, + 0.00891189370304346, + 0.01837052032351494, + 0.023859409615397453, + 0.012308490462601185, + 0.009636940434575081, + 0.02015688084065914, + 0.017554322257637978, + 0.008540097624063492, + 0.018759818747639656, + 0.017471294850111008, + 0.021224720403552055, + 0.005921678151935339, + 0.005070819519460201, + 0.041098494082689285, + 0.010802451521158218, + 0.004177320282906294, + 0.0053835404105484486, + 0.007043470162898302, + 0.033746976405382156, + 0.00929763913154602, + 0.009237355552613735, + 0.007372516207396984, + 0.011098282411694527, + 0.005977770313620567, + 0.008323372341692448, + 0.00850166380405426, + 0.011050975881516933, + 0.013309708796441555, + 0.05360480025410652, + 0.010543988086283207, + 0.022169170901179314, + 0.013072025962173939, + 0.020002279430627823, + 0.07636462152004242, + 0.01791679486632347, + 0.003969124983996153, + 0.0033598411828279495, + 0.007567906752228737, + 0.006047354079782963, + 0.004251522943377495, + 0.0037135062739253044, + 0.013611353002488613, + 0.0044501121155917645, + 0.0052337078377604485, + 0.0038460695650428534, + 0.0029817121103405952, + 0.011670982465147972, + 0.003432017285376787, + 0.00418873643502593, + 0.004256261978298426, + 0.003088301746174693, + 0.0037035620771348476, + 0.002239048480987549, + 0.004997314419597387, + 0.005938255228102207, + 0.004896009340882301, + 0.004871262237429619, + 0.006879978813230991, + 0.0036697390023618937, + 0.008956866338849068, + 0.0026232637465000153, + 0.004065587650984526, + 0.009305414743721485, + 0.0031545860692858696, + 0.018515661358833313, + 0.0032075331546366215, + 0.0036958882119506598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.062022291123867035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a slightly upward angle, slowly pulls out as the subject moves, capturing a fairy in ancient costume on a grand palace terrace. Cool, watery moonlight pours down, casting a holy silver glow on the smooth, carved white marble floor, illuminating a faint silhouette. She stands slender, her long, jet-black hair held together by a simple jade hairpin. Her skin, whiter than snow in the moonlight, reflects a serene expression with a touch of coolness. She wears a pure white, wide-sleeved, flowing fairy dress. The sheer silk flutters gently in the night breeze, and the delicate cloud patterns embroidered with silver thread on the cuffs and hem shimmer with tiny glimmers of light. She slowly raises her slender arm, then twists her wrist gently, steadily and deliberately swinging the long sleeves out gracefully. The wide sleeves, like a flowing white silk ribbon, trace a graceful, lingering arc across the deep night sky, as if silently telling a story a thousand years old. Behind her are misty palaces and a starry night sky, with a bright full moon hanging high in the sky. The whole picture, in cold tones, presents an ultimate tranquility, ethereality and dreamy classical beauty.", + "seed": 6828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_14/index_7438_seed_6828_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07205951381474733, + "vlm_logprob": -0.07205951381474733, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.07205951381474733, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2633143365383148 + ], + "opposite_logprob_tokens": [ + -1.5133143663406372 + ], + "logprob_reward": -0.2633143365383148, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2633143365383148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15023425221443176, + 0.03642367571592331, + 0.11936405301094055, + 0.06576000899076462, + 0.08329285681247711, + 0.08011510968208313, + 0.2822478115558624, + 0.04605573043227196, + 0.05867725610733032, + 0.07782924175262451 + ], + "credit_map": [ + 0.017024924978613853, + 0.015315191820263863, + 0.02417711913585663, + 0.026846056804060936, + 0.01497241947799921, + 0.014825371094048023, + 0.014769451692700386, + 0.022303719073534012, + 0.0036678877659142017, + 0.003473736112937331, + 0.004480151459574699, + 0.007356460206210613, + 0.003485935041680932, + 0.005059155635535717, + 0.005370835307985544, + 0.0035295123234391212, + 0.00660662679001689, + 0.0076857805252075195, + 0.019994152709841728, + 0.015807868912816048, + 0.0061072129756212234, + 0.00917118415236473, + 0.0374528169631958, + 0.016538400202989578, + 0.0045179040171206, + 0.006162604317069054, + 0.009314612485468388, + 0.012873930856585503, + 0.006218332331627607, + 0.006688757799565792, + 0.01383640244603157, + 0.006147466599941254, + 0.005398266017436981, + 0.00597514770925045, + 0.00991873536258936, + 0.013971743173897266, + 0.007041881326586008, + 0.011146171018481255, + 0.014289533719420433, + 0.015551380813121796, + 0.00597588112577796, + 0.0048279049806296825, + 0.018642490729689598, + 0.01516852155327797, + 0.006128034088760614, + 0.009960643947124481, + 0.010110169649124146, + 0.009301470592617989, + 0.026692859828472137, + 0.016150236129760742, + 0.08278995752334595, + 0.07413670420646667, + 0.013445660471916199, + 0.014385998249053955, + 0.03461744636297226, + 0.020028943195939064, + 0.007157237734645605, + 0.003327838843688369, + 0.004915101453661919, + 0.004663820378482342, + 0.0034173724707216024, + 0.00922582671046257, + 0.0066967313177883625, + 0.0066517977975308895, + 0.007542152423411608, + 0.005089397542178631, + 0.009861449711024761, + 0.010003590025007725, + 0.0035092364996671677, + 0.008060912601649761, + 0.00825418159365654, + 0.0063563380390405655, + 0.009787501767277718, + 0.00934743881225586, + 0.01134384237229824, + 0.013016551733016968, + 0.00562250567600131, + 0.006440603639930487, + 0.01446621399372816, + 0.0078045884147286415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.2633143365383148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009803364053368568 + ], + "opposite_logprob_tokens": [ + -4.634803295135498 + ], + "logprob_reward": -0.009803364053368568, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009803364053368568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19801056385040283, + 0.10451792925596237, + 0.06839296221733093, + 0.05953538045287132, + 0.1664244383573532, + 0.04939648509025574, + 0.09265157580375671, + 0.06666585803031921, + 0.08339668810367584, + 0.1110081747174263 + ], + "credit_map": [ + 0.01412068773061037, + 0.015721241012215614, + 0.032732658088207245, + 0.033657852560281754, + 0.014267734251916409, + 0.017073815688490868, + 0.033978112041950226, + 0.03645847737789154, + 0.0054220519959926605, + 0.0062778415158391, + 0.012249690480530262, + 0.016585875302553177, + 0.007991722784936428, + 0.009397800080478191, + 0.023459501564502716, + 0.023133449256420135, + 0.007914039306342602, + 0.006275045685470104, + 0.013430875726044178, + 0.010305802337825298, + 0.005220834631472826, + 0.007910381071269512, + 0.009562408551573753, + 0.007773575372993946, + 0.004093778785318136, + 0.005024883896112442, + 0.006465700455009937, + 0.012442454695701599, + 0.007020516786724329, + 0.004421936813741922, + 0.011224457994103432, + 0.008841650560498238, + 0.009533660486340523, + 0.013849261216819286, + 0.018327534198760986, + 0.02864965982735157, + 0.015822505578398705, + 0.015657484531402588, + 0.02992432750761509, + 0.03466000780463219, + 0.004069722257554531, + 0.0033406447619199753, + 0.011525586247444153, + 0.006835685111582279, + 0.004298877902328968, + 0.004902851767838001, + 0.0068441336043179035, + 0.0075789825059473515, + 0.010093092918395996, + 0.007338649593293667, + 0.01619933731853962, + 0.018535500392317772, + 0.007500962354242802, + 0.008159288205206394, + 0.01248190924525261, + 0.01234283298254013, + 0.009475299157202244, + 0.007059986237436533, + 0.0072268955409526825, + 0.009309832938015461, + 0.0039533874951303005, + 0.010589520446956158, + 0.010875225067138672, + 0.008175707422196865, + 0.0073363143019378185, + 0.011160595342516899, + 0.012782108969986439, + 0.0192762091755867, + 0.004764406476169825, + 0.006461171433329582, + 0.012404637411236763, + 0.009211244992911816, + 0.014081468805670738, + 0.020108703523874283, + 0.018258625641465187, + 0.016076533123850822, + 0.009614052250981331, + 0.009333161637187004, + 0.012795100919902325, + 0.010740518569946289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.009803364053368568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648474045097828 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648474045097828, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648474045097828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13198696076869965, + 0.21929135918617249, + 0.10652688145637512, + 0.04047410935163498, + 0.10485604405403137, + 0.08399569243192673, + 0.12298242747783661, + 0.061354927718639374, + 0.06760795414447784, + 0.06092362105846405 + ], + "credit_map": [ + 0.012134203687310219, + 0.010457352735102177, + 0.02751373127102852, + 0.021351799368858337, + 0.009698878973722458, + 0.012681995518505573, + 0.015291089192032814, + 0.0228579044342041, + 0.008172999136149883, + 0.008610977791249752, + 0.021184351295232773, + 0.025649115443229675, + 0.009474899619817734, + 0.009330234490334988, + 0.07292978465557098, + 0.06393898278474808, + 0.0062211137264966965, + 0.005345549434423447, + 0.018407480791211128, + 0.009866990149021149, + 0.0047599091194570065, + 0.006545462645590305, + 0.03803499415516853, + 0.017345380038022995, + 0.003871290013194084, + 0.0034639344085007906, + 0.008096717298030853, + 0.005477340426295996, + 0.0045874412171542645, + 0.004045797977596521, + 0.004778870847076178, + 0.006152717862278223, + 0.006834486965090036, + 0.006880634464323521, + 0.01319056749343872, + 0.02052287384867668, + 0.007406344171613455, + 0.009847534820437431, + 0.01405339129269123, + 0.026120208203792572, + 0.0075684236362576485, + 0.005501380655914545, + 0.013642566278576851, + 0.010293589904904366, + 0.008304130285978317, + 0.01511458121240139, + 0.008743122220039368, + 0.014827892184257507, + 0.013256769627332687, + 0.011256147176027298, + 0.026093507185578346, + 0.03276349604129791, + 0.007972175255417824, + 0.0059954384341835976, + 0.018378181383013725, + 0.007266714237630367, + 0.006717487704008818, + 0.008036206476390362, + 0.0069787465035915375, + 0.010609086602926254, + 0.0062189241871237755, + 0.005170144606381655, + 0.011727751232683659, + 0.005896579008549452, + 0.007476631551980972, + 0.005875417031347752, + 0.009500779211521149, + 0.007886654697358608, + 0.006389171816408634, + 0.007415753323584795, + 0.008922144770622253, + 0.014141401275992393, + 0.007217697333544493, + 0.008233361877501011, + 0.006611031945794821, + 0.006970924790948629, + 0.0055649979040026665, + 0.0066957720555365086, + 0.007880281656980515, + 0.011749553494155407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.004648474045097828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535588815808296 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535588815808296, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535588815808296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05084701627492905, + 0.03155166283249855, + 0.27953770756721497, + 0.1070193499326706, + 0.037088796496391296, + 0.033139538019895554, + 0.11880480498075485, + 0.14372478425502777, + 0.12235059589147568, + 0.07593569159507751 + ], + "credit_map": [ + 0.005550125613808632, + 0.0035214023664593697, + 0.010196004062891006, + 0.005103410687297583, + 0.00433643301948905, + 0.004896160215139389, + 0.011603886261582375, + 0.005639596376568079, + 0.0026917930226773024, + 0.0038871767465025187, + 0.0038894384633749723, + 0.004721228964626789, + 0.003421517787501216, + 0.003715023398399353, + 0.003940388094633818, + 0.005285096820443869, + 0.013658937998116016, + 0.013523400761187077, + 0.04780149459838867, + 0.029295936226844788, + 0.011781120672821999, + 0.016160111874341965, + 0.1006123349070549, + 0.046704381704330444, + 0.008042213506996632, + 0.007776609156280756, + 0.01219862513244152, + 0.0221034437417984, + 0.009697276167571545, + 0.009682133793830872, + 0.026347700506448746, + 0.01117135863751173, + 0.0033727106638252735, + 0.0029946754220873117, + 0.0069376700557768345, + 0.006700451020151377, + 0.002959164557978511, + 0.0030446317978203297, + 0.004439230542629957, + 0.006640266627073288, + 0.0032904772087931633, + 0.0021580986212939024, + 0.00540499622002244, + 0.004652751609683037, + 0.004031193442642689, + 0.0036261093337088823, + 0.004512322600930929, + 0.00546359084546566, + 0.011142343282699585, + 0.0063960617408156395, + 0.03607574850320816, + 0.033562071621418, + 0.0056633842177689075, + 0.005813537165522575, + 0.01193983480334282, + 0.008211834356188774, + 0.014631151221692562, + 0.022118208929896355, + 0.016283920034766197, + 0.024215133860707283, + 0.013019604608416557, + 0.012738301418721676, + 0.027308503165841103, + 0.013409963808953762, + 0.01405827235430479, + 0.016231445595622063, + 0.019308755174279213, + 0.023196445778012276, + 0.008732471615076065, + 0.01109995786100626, + 0.015178002417087555, + 0.014545264653861523, + 0.008721408434212208, + 0.011974094435572624, + 0.007191134616732597, + 0.012207838706672192, + 0.007565998937934637, + 0.008416160941123962, + 0.008344786241650581, + 0.011514274403452873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.012535588815808296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999580562114716 + ], + "opposite_logprob_tokens": [ + -2.694995880126953 + ], + "logprob_reward": -0.06999580562114716, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999580562114716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0753081664443016, + 0.19119685888290405, + 0.3737059533596039, + 0.0660611093044281, + 0.0481954924762249, + 0.05467648059129715, + 0.06845979392528534, + 0.03635513782501221, + 0.05523234233260155, + 0.030808614566922188 + ], + "credit_map": [ + 0.006065456196665764, + 0.00598911615088582, + 0.012310509569942951, + 0.009362387470901012, + 0.008547164499759674, + 0.011010104790329933, + 0.015214744955301285, + 0.00680868374183774, + 0.00400948990136385, + 0.0032246208284050226, + 0.013219776563346386, + 0.00965841580182314, + 0.007083533797413111, + 0.009444840252399445, + 0.0754065215587616, + 0.0691496729850769, + 0.01860070787370205, + 0.01893037185072899, + 0.06325124949216843, + 0.03891928493976593, + 0.01575351506471634, + 0.02118665724992752, + 0.1350719928741455, + 0.061992187052965164, + 0.004801573231816292, + 0.006218432914465666, + 0.007549460977315903, + 0.013404598459601402, + 0.008729389868676662, + 0.005977102089673281, + 0.0136296721175313, + 0.005750883836299181, + 0.0032304348424077034, + 0.004627179820090532, + 0.005195108707994223, + 0.008049690164625645, + 0.004094739444553852, + 0.005454587284475565, + 0.008772843517363071, + 0.00877091009169817, + 0.005074574612081051, + 0.0031567139085382223, + 0.009194789454340935, + 0.007886482402682304, + 0.0051160333678126335, + 0.008630014955997467, + 0.008472466841340065, + 0.007145404815673828, + 0.004093066789209843, + 0.0033554243855178356, + 0.02081865631043911, + 0.02278982289135456, + 0.00258226552978158, + 0.0028464465867727995, + 0.008051734417676926, + 0.003922381438314915, + 0.0044695367105305195, + 0.0032833972945809364, + 0.004102719947695732, + 0.005893418099731207, + 0.00397584168240428, + 0.005592729430645704, + 0.005080863367766142, + 0.003956635016947985, + 0.005204157438129187, + 0.004891722463071346, + 0.009597431868314743, + 0.008241832256317139, + 0.0029109737370163202, + 0.003801967715844512, + 0.005070835817605257, + 0.015513421967625618, + 0.003532520029693842, + 0.0045238034799695015, + 0.0030721521470695734, + 0.003583302954211831, + 0.003961243201047182, + 0.003176282625645399, + 0.003863054560497403, + 0.005096254404634237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.06999580562114716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a slightly upward angle, slowly pulls out as the subject moves, capturing a fairy in ancient costume on a grand palace terrace. Cool, watery moonlight pours down, casting a holy silver glow on the smooth, carved white marble floor, illuminating a faint silhouette. She stands slender, her long, jet-black hair held together by a simple jade hairpin. Her skin, whiter than snow in the moonlight, reflects a serene expression with a touch of coolness. She wears a pure white, wide-sleeved, flowing fairy dress. The sheer silk flutters gently in the night breeze, and the delicate cloud patterns embroidered with silver thread on the cuffs and hem shimmer with tiny glimmers of light. She slowly raises her slender arm, then twists her wrist gently, steadily and deliberately swinging the long sleeves out gracefully. The wide sleeves, like a flowing white silk ribbon, trace a graceful, lingering arc across the deep night sky, as if silently telling a story a thousand years old. Behind her are misty palaces and a starry night sky, with a bright full moon hanging high in the sky. The whole picture, in cold tones, presents an ultimate tranquility, ethereality and dreamy classical beauty.", + "seed": 6829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_14/index_7438_seed_6829_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02421006690710783, + "vlm_logprob": -0.02421006690710783, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.02421006690710783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07168486714363098 + ], + "opposite_logprob_tokens": [ + -2.8216848373413086 + ], + "logprob_reward": -0.07168486714363098, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07168486714363098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1343061625957489, + 0.11678678542375565, + 0.07259172201156616, + 0.07106224447488785, + 0.07181873172521591, + 0.1210760772228241, + 0.07741574943065643, + 0.19769005477428436, + 0.04987264797091484, + 0.0873798206448555 + ], + "credit_map": [ + 0.011133728548884392, + 0.017229566350579262, + 0.015779459848999977, + 0.014707430265843868, + 0.020383166149258614, + 0.014788256026804447, + 0.018926341086626053, + 0.021358223631978035, + 0.0073367152363061905, + 0.014049920253455639, + 0.015849413350224495, + 0.024122336879372597, + 0.012788938358426094, + 0.011383695527911186, + 0.010206365957856178, + 0.021049410104751587, + 0.005683170165866613, + 0.00674935569986701, + 0.015490904450416565, + 0.014166347682476044, + 0.007785039953887463, + 0.006453210953623056, + 0.0087442547082901, + 0.007519445847719908, + 0.006434462033212185, + 0.008571375161409378, + 0.01021028496325016, + 0.009075860492885113, + 0.008689784444868565, + 0.007271961774677038, + 0.009155258536338806, + 0.011653265915811062, + 0.007108836434781551, + 0.008673173375427723, + 0.01042213849723339, + 0.011796698905527592, + 0.007105788681656122, + 0.008812114596366882, + 0.009453491307795048, + 0.008446495048701763, + 0.010838449001312256, + 0.011761237867176533, + 0.02007882297039032, + 0.01648762635886669, + 0.011905297636985779, + 0.012198139913380146, + 0.01407255046069622, + 0.02373395301401615, + 0.011306650005280972, + 0.013157335110008717, + 0.008534274995326996, + 0.011509343050420284, + 0.006405557505786419, + 0.007940166629850864, + 0.008841182105243206, + 0.009721243754029274, + 0.02559012733399868, + 0.014576883986592293, + 0.0335698202252388, + 0.02785058133304119, + 0.022589460015296936, + 0.01222565583884716, + 0.021809378638863564, + 0.03947816416621208, + 0.00844909343868494, + 0.0055745565332472324, + 0.008301799185574055, + 0.0035240959841758013, + 0.005835434887558222, + 0.010546958073973656, + 0.0031754777301102877, + 0.00446523679420352, + 0.015102825127542019, + 0.0067557902075350285, + 0.01442601066082716, + 0.008797788992524147, + 0.011870802380144596, + 0.01635381393134594, + 0.006105442065745592, + 0.007967359386384487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.07168486714363098 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005965879186987877 + ], + "opposite_logprob_tokens": [ + -5.130965709686279 + ], + "logprob_reward": -0.005965879186987877, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005965879186987877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24589690566062927, + 0.173785001039505, + 0.04628964513540268, + 0.14412298798561096, + 0.126106396317482, + 0.060402244329452515, + 0.062197066843509674, + 0.05924338474869728, + 0.037710659205913544, + 0.044245705008506775 + ], + "credit_map": [ + 0.016238436102867126, + 0.020750945433974266, + 0.06529638171195984, + 0.057109784334897995, + 0.02247190847992897, + 0.017449164763092995, + 0.016236769035458565, + 0.030343525111675262, + 0.008084072731435299, + 0.015865478664636612, + 0.03179445490241051, + 0.044132329523563385, + 0.014059122651815414, + 0.01361183263361454, + 0.01620657555758953, + 0.03003113530576229, + 0.004880605265498161, + 0.006061066873371601, + 0.005496790632605553, + 0.005420566536486149, + 0.006806366145610809, + 0.005675339139997959, + 0.005210996605455875, + 0.006737913936376572, + 0.01262609101831913, + 0.01544034294784069, + 0.022663312032818794, + 0.021043261513113976, + 0.015879474580287933, + 0.01628517545759678, + 0.0156250037252903, + 0.024560311809182167, + 0.008017483167350292, + 0.01048066932708025, + 0.02656424790620804, + 0.02469705417752266, + 0.014087343588471413, + 0.010914207436144352, + 0.017050975933670998, + 0.014294418506324291, + 0.005456661805510521, + 0.008395275101065636, + 0.00862095132470131, + 0.006231776904314756, + 0.007700997870415449, + 0.005671702791005373, + 0.007907684892416, + 0.010417189449071884, + 0.008446456864476204, + 0.011240976862609386, + 0.007053004577755928, + 0.005171041004359722, + 0.006415870040655136, + 0.007050622720271349, + 0.007113525178283453, + 0.009705565869808197, + 0.009785394184291363, + 0.0052846684120595455, + 0.011432785540819168, + 0.006224337965250015, + 0.00630926201120019, + 0.004893254954367876, + 0.007176045328378677, + 0.008137635886669159, + 0.006882266141474247, + 0.003368969773873687, + 0.007313260342925787, + 0.0038487466517835855, + 0.0031491820700466633, + 0.006120813544839621, + 0.0026644221507012844, + 0.004362997133284807, + 0.008371601812541485, + 0.0030722087249159813, + 0.008927798829972744, + 0.004019024781882763, + 0.005759043153375387, + 0.008235183544456959, + 0.002656530821695924, + 0.003204312641173601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.005965879186987877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002498126355931163 + ], + "opposite_logprob_tokens": [ + -6.002498149871826 + ], + "logprob_reward": -0.002498126355931163, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002498126355931163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3248385488986969, + 0.08932098001241684, + 0.04368280619382858, + 0.047361865639686584, + 0.08162780106067657, + 0.07822594791650772, + 0.052228886634111404, + 0.14197511970996857, + 0.04782724753022194, + 0.09291081875562668 + ], + "credit_map": [ + 0.018946439027786255, + 0.022243889048695564, + 0.10757408291101456, + 0.09214438498020172, + 0.016871916130185127, + 0.017792467027902603, + 0.02222757413983345, + 0.02703780308365822, + 0.009735503233969212, + 0.00868160929530859, + 0.010185815393924713, + 0.010397619567811489, + 0.013358231633901596, + 0.013837127014994621, + 0.011275753378868103, + 0.011849320493638515, + 0.005165749695152044, + 0.005322645418345928, + 0.006966423708945513, + 0.00515463063493371, + 0.00744883855804801, + 0.004294347949326038, + 0.004191748797893524, + 0.005138421431183815, + 0.004239270929247141, + 0.006281144451349974, + 0.005971895065158606, + 0.005855557508766651, + 0.005141878966242075, + 0.005745663773268461, + 0.004915548954159021, + 0.009210905060172081, + 0.005473224446177483, + 0.0068294634111225605, + 0.015033137984573841, + 0.014808300882577896, + 0.010855460539460182, + 0.005769209936261177, + 0.01417047530412674, + 0.008688529022037983, + 0.0057081496343016624, + 0.007298008538782597, + 0.009571441449224949, + 0.009129675105214119, + 0.009896582923829556, + 0.007963770069181919, + 0.012585856951773167, + 0.016072465106844902, + 0.00489312969148159, + 0.0046898676082491875, + 0.005829296540468931, + 0.008288293145596981, + 0.01010139286518097, + 0.004059461411088705, + 0.007323545403778553, + 0.007043901365250349, + 0.017675280570983887, + 0.014138886705040932, + 0.020450342446565628, + 0.021539565175771713, + 0.019451288506388664, + 0.009838668629527092, + 0.017195241525769234, + 0.021685844287276268, + 0.006966552697122097, + 0.004696309100836515, + 0.00845397263765335, + 0.004826802760362625, + 0.004952339455485344, + 0.006660158280283213, + 0.0050119454972445965, + 0.006259167566895485, + 0.013362680561840534, + 0.009093226864933968, + 0.010798316448926926, + 0.011043992824852467, + 0.012511592358350754, + 0.015128713101148605, + 0.006846864242106676, + 0.014125430025160313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.002498126355931163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028174726758152246 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028174726758152246, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028174726758152246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30482304096221924, + 0.09620843082666397, + 0.0544048510491848, + 0.06622172892093658, + 0.05958721786737442, + 0.060831137001514435, + 0.06772457808256149, + 0.10880997031927109, + 0.07300405949354172, + 0.10838493704795837 + ], + "credit_map": [ + 0.012908725999295712, + 0.021372675895690918, + 0.09807493537664413, + 0.08681043982505798, + 0.021983318030834198, + 0.017687465995550156, + 0.018032213672995567, + 0.027953267097473145, + 0.0064677102491259575, + 0.011160889640450478, + 0.018137158825993538, + 0.01131978165358305, + 0.012340972200036049, + 0.01009881217032671, + 0.010108214803040028, + 0.01657489314675331, + 0.005430050194263458, + 0.008532116189599037, + 0.009127262979745865, + 0.007582430727779865, + 0.007001405116170645, + 0.005085850600153208, + 0.004780869465321302, + 0.006864867638796568, + 0.007610052824020386, + 0.008911127224564552, + 0.007612911984324455, + 0.00819259975105524, + 0.007918255403637886, + 0.007944222539663315, + 0.006979487370699644, + 0.011053071357309818, + 0.004095279611647129, + 0.005400942172855139, + 0.010920013301074505, + 0.009917535819113255, + 0.007623394951224327, + 0.003797155572101474, + 0.010289832949638367, + 0.007543062325567007, + 0.0065256343223154545, + 0.007615146227180958, + 0.007755560334771872, + 0.005364793818444014, + 0.008389563299715519, + 0.00514843687415123, + 0.009836467914283276, + 0.0101955346763134, + 0.006934357341378927, + 0.009935551322996616, + 0.008917335420846939, + 0.009408754296600819, + 0.005169952753931284, + 0.006449245847761631, + 0.007954015396535397, + 0.012955368496477604, + 0.011532817035913467, + 0.010144317522644997, + 0.013506497256457806, + 0.018361754715442657, + 0.013771338388323784, + 0.011718112975358963, + 0.011953456327319145, + 0.017821677029132843, + 0.01410671230405569, + 0.006170402280986309, + 0.011649358086287975, + 0.00480704614892602, + 0.008099905215203762, + 0.01693008840084076, + 0.0053217425011098385, + 0.00591879989951849, + 0.01814657263457775, + 0.009340880438685417, + 0.015230445191264153, + 0.011608254164457321, + 0.014813139103353024, + 0.019898969680070877, + 0.007526975125074387, + 0.011819692328572273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0028174726758152246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038083989173173904 + ], + "opposite_logprob_tokens": [ + -3.288084030151367 + ], + "logprob_reward": -0.038083989173173904, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038083989173173904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2614550292491913, + 0.24199768900871277, + 0.06315789371728897, + 0.05008232966065407, + 0.06378219276666641, + 0.07806520909070969, + 0.10670232027769089, + 0.06748513132333755, + 0.03700249642133713, + 0.030269654467701912 + ], + "credit_map": [ + 0.011121460236608982, + 0.01778002828359604, + 0.0934716984629631, + 0.08058255910873413, + 0.012905307114124298, + 0.011522474698722363, + 0.011771365068852901, + 0.022300150245428085, + 0.014158939011394978, + 0.025464728474617004, + 0.027726296335458755, + 0.03943568468093872, + 0.03073214553296566, + 0.030231518670916557, + 0.022268788889050484, + 0.05197959765791893, + 0.005579860880970955, + 0.0066153318621218204, + 0.010720402002334595, + 0.009692150168120861, + 0.008828672580420971, + 0.005967430770397186, + 0.005435081664472818, + 0.010318969376385212, + 0.004115675576031208, + 0.004294120240956545, + 0.008149422705173492, + 0.006600921042263508, + 0.004879394546151161, + 0.0051282793283462524, + 0.007494769990444183, + 0.009419751353561878, + 0.006465913262218237, + 0.0073499903082847595, + 0.009562963619828224, + 0.008459731005132198, + 0.006630114279687405, + 0.006806381046772003, + 0.008740246295928955, + 0.009766859002411366, + 0.005747106857597828, + 0.006750165950506926, + 0.010291188023984432, + 0.0091104032471776, + 0.0068223681300878525, + 0.007267054636031389, + 0.01598295010626316, + 0.016093984246253967, + 0.012729734182357788, + 0.007713895756751299, + 0.014003518968820572, + 0.013466058298945427, + 0.009314325638115406, + 0.01221243292093277, + 0.01342856977134943, + 0.02383379079401493, + 0.008106744848191738, + 0.006295805796980858, + 0.010943868197500706, + 0.010904477909207344, + 0.008606167510151863, + 0.0077294353395700455, + 0.0066809989511966705, + 0.00821763463318348, + 0.006085102446377277, + 0.003926693461835384, + 0.0059371828101575375, + 0.003259370569139719, + 0.0035665729083120823, + 0.007867705076932907, + 0.003300143638625741, + 0.003059729468077421, + 0.00535549595952034, + 0.0022703814320266247, + 0.00399532588198781, + 0.002735164947807789, + 0.005169016309082508, + 0.005761871580034494, + 0.0028153718449175358, + 0.0021670281421393156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.038083989173173904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a slightly upward angle, slowly pulls out as the subject moves, capturing a fairy in ancient costume on a grand palace terrace. Cool, watery moonlight pours down, casting a holy silver glow on the smooth, carved white marble floor, illuminating a faint silhouette. She stands slender, her long, jet-black hair held together by a simple jade hairpin. Her skin, whiter than snow in the moonlight, reflects a serene expression with a touch of coolness. She wears a pure white, wide-sleeved, flowing fairy dress. The sheer silk flutters gently in the night breeze, and the delicate cloud patterns embroidered with silver thread on the cuffs and hem shimmer with tiny glimmers of light. She slowly raises her slender arm, then twists her wrist gently, steadily and deliberately swinging the long sleeves out gracefully. The wide sleeves, like a flowing white silk ribbon, trace a graceful, lingering arc across the deep night sky, as if silently telling a story a thousand years old. Behind her are misty palaces and a starry night sky, with a bright full moon hanging high in the sky. The whole picture, in cold tones, presents an ultimate tranquility, ethereality and dreamy classical beauty.", + "seed": 6830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_14/index_7438_seed_6830_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022213070839643478, + "vlm_logprob": -0.022213070839643478, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.022213070839643478, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026515353471040726 + ], + "opposite_logprob_tokens": [ + -3.901515245437622 + ], + "logprob_reward": -0.026515353471040726, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026515353471040726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21054039895534515, + 0.07168800383806229, + 0.09017078578472137, + 0.08439529687166214, + 0.1545666605234146, + 0.0767168253660202, + 0.05352907255291939, + 0.07558813691139221, + 0.1217283084988594, + 0.061076559126377106 + ], + "credit_map": [ + 0.021505853161215782, + 0.027296878397464752, + 0.026105104014277458, + 0.031005581840872765, + 0.03396230190992355, + 0.030656056478619576, + 0.019670536741614342, + 0.020338090136647224, + 0.006093685515224934, + 0.009035689756274223, + 0.007929238490760326, + 0.006633414886891842, + 0.016805633902549744, + 0.009570435620844364, + 0.007059697527438402, + 0.008560210466384888, + 0.007510170340538025, + 0.012701201252639294, + 0.009819942526519299, + 0.01193297654390335, + 0.014404972083866596, + 0.007882921025156975, + 0.014101117849349976, + 0.011817487888038158, + 0.008593590930104256, + 0.008285844698548317, + 0.011222404427826405, + 0.009935837239027023, + 0.011676815338432789, + 0.006709377747029066, + 0.018722444772720337, + 0.009248987771570683, + 0.011714613996446133, + 0.01500024925917387, + 0.01926746405661106, + 0.017063351348042488, + 0.01733432523906231, + 0.012607834301888943, + 0.03748796135187149, + 0.024090854451060295, + 0.006822994910180569, + 0.009436393156647682, + 0.011110717430710793, + 0.01262733619660139, + 0.007144580129534006, + 0.008739287033677101, + 0.008834559470415115, + 0.012000953778624535, + 0.005791948176920414, + 0.00566020468249917, + 0.007965801283717155, + 0.010113043710589409, + 0.005780111998319626, + 0.004205314442515373, + 0.005953859072178602, + 0.008058789186179638, + 0.009172208607196808, + 0.006528116296976805, + 0.013296477496623993, + 0.0110994977876544, + 0.01243864931166172, + 0.004145828075706959, + 0.00974531564861536, + 0.00916204508394003, + 0.01959923654794693, + 0.010598710738122463, + 0.012667452916502953, + 0.007313498295843601, + 0.027287425473332405, + 0.007971123792231083, + 0.014080232940614223, + 0.02221062406897545, + 0.01113097369670868, + 0.0063561322167515755, + 0.010250218212604523, + 0.007504317443817854, + 0.004832296632230282, + 0.007457295432686806, + 0.005167024675756693, + 0.008378300815820694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.026515353471040726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004652864299714565 + ], + "opposite_logprob_tokens": [ + -5.379652976989746 + ], + "logprob_reward": -0.004652864299714565, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004652864299714565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19758901000022888, + 0.22126680612564087, + 0.12612050771713257, + 0.1089952290058136, + 0.14173683524131775, + 0.04426676034927368, + 0.048794250935316086, + 0.037310611456632614, + 0.028708862140774727, + 0.0452110692858696 + ], + "credit_map": [ + 0.029502034187316895, + 0.02001171186566353, + 0.02649029903113842, + 0.03029673360288143, + 0.030148644000291824, + 0.02568327821791172, + 0.016804488375782967, + 0.018651828169822693, + 0.021552395075559616, + 0.03009466826915741, + 0.02561255730688572, + 0.020923327654600143, + 0.054289743304252625, + 0.02263120375573635, + 0.025313787162303925, + 0.02084912732243538, + 0.011061091907322407, + 0.016811968758702278, + 0.015812121331691742, + 0.016487156972289085, + 0.016145769506692886, + 0.012986228801310062, + 0.017308631911873817, + 0.01950754038989544, + 0.008785170502960682, + 0.00932175200432539, + 0.013860565610229969, + 0.010418170131742954, + 0.016730932518839836, + 0.007146294694393873, + 0.0290375929325819, + 0.01369476318359375, + 0.01054870244115591, + 0.008968167006969452, + 0.029920106753706932, + 0.01558438315987587, + 0.009550883434712887, + 0.012402668595314026, + 0.019804859533905983, + 0.03495708107948303, + 0.003828052431344986, + 0.005208541639149189, + 0.007694990839809179, + 0.007284931372851133, + 0.004699044395238161, + 0.004218680784106255, + 0.004159148782491684, + 0.007173373829573393, + 0.005865768529474735, + 0.004157326649874449, + 0.006465492304414511, + 0.007110562641173601, + 0.0046777077950537205, + 0.0036471185740083456, + 0.009109335020184517, + 0.007760938722640276, + 0.003571064444258809, + 0.0038008433766663074, + 0.008294518105685711, + 0.00509675545617938, + 0.004662366583943367, + 0.0025469204410910606, + 0.005257908720523119, + 0.0040802378207445145, + 0.004935610573738813, + 0.0035388863179832697, + 0.004406224004924297, + 0.002640014048665762, + 0.003249320201575756, + 0.002353491960093379, + 0.002970640081912279, + 0.004614678211510181, + 0.007349521853029728, + 0.004272195510566235, + 0.006751445587724447, + 0.004342895932495594, + 0.0037202949170023203, + 0.010767769068479538, + 0.0031802442390471697, + 0.004826705437153578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.004652864299714565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645270295441151 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645270295441151, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645270295441151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23433299362659454, + 0.07010318338871002, + 0.10745711624622345, + 0.1961694359779358, + 0.1184196025133133, + 0.04157860577106476, + 0.06431344151496887, + 0.04133402183651924, + 0.07529319822788239, + 0.05099840462207794 + ], + "credit_map": [ + 0.02786560356616974, + 0.03345785662531853, + 0.03600674122571945, + 0.03384499251842499, + 0.03447647765278816, + 0.03078572079539299, + 0.016023511067032814, + 0.02187209576368332, + 0.007496294565498829, + 0.010220127180218697, + 0.007677963003516197, + 0.006376880221068859, + 0.012753675691783428, + 0.007008847314864397, + 0.008777851238846779, + 0.009791544638574123, + 0.0066434480249881744, + 0.009478745982050896, + 0.014656543731689453, + 0.015013649128377438, + 0.014619218185544014, + 0.008508671075105667, + 0.022954193875193596, + 0.015582648105919361, + 0.017375491559505463, + 0.017602913081645966, + 0.021124832332134247, + 0.024147262796759605, + 0.03539704903960228, + 0.013173791579902172, + 0.044508565217256546, + 0.022839518263936043, + 0.0066593303345143795, + 0.012817115522921085, + 0.011572571471333504, + 0.01186787337064743, + 0.016627077013254166, + 0.007442444562911987, + 0.026210341602563858, + 0.025222845375537872, + 0.004271079320460558, + 0.006422927603125572, + 0.006938200909644365, + 0.005580715369433165, + 0.004384816158562899, + 0.0032177134416997433, + 0.004410639870911837, + 0.00635251309722662, + 0.006250324659049511, + 0.005613709334284067, + 0.010817375034093857, + 0.008886387571692467, + 0.007632233668118715, + 0.0065443492494523525, + 0.00902131199836731, + 0.00954775046557188, + 0.004105567000806332, + 0.004219729918986559, + 0.00806914921849966, + 0.0053114499896764755, + 0.0055650523863732815, + 0.0040940698236227036, + 0.0047105480916798115, + 0.005258454009890556, + 0.007734358310699463, + 0.00820580031722784, + 0.009304055012762547, + 0.007807433605194092, + 0.014231126755475998, + 0.0045859175734221935, + 0.011742751114070415, + 0.011681749485433102, + 0.009342777542769909, + 0.00568259134888649, + 0.007230786606669426, + 0.006214370019733906, + 0.004227367229759693, + 0.007166776340454817, + 0.004598123021423817, + 0.0065356106497347355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.004645270295441151 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250830203294754 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005250830203294754, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250830203294754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16497959196567535, + 0.08680420368909836, + 0.15456967055797577, + 0.1165044903755188, + 0.06360422074794769, + 0.13735263049602509, + 0.12402548640966415, + 0.02902979403734207, + 0.05958816781640053, + 0.06354176253080368 + ], + "credit_map": [ + 0.016518158838152885, + 0.02181551419198513, + 0.023619182407855988, + 0.02019193209707737, + 0.03222618252038956, + 0.02137230522930622, + 0.014083883725106716, + 0.01515242736786604, + 0.006465093232691288, + 0.010772914625704288, + 0.009033692069351673, + 0.007685607299208641, + 0.024687066674232483, + 0.007814399898052216, + 0.010355866514146328, + 0.00998956523835659, + 0.009526241570711136, + 0.01790574938058853, + 0.016026882454752922, + 0.019293591380119324, + 0.03206024318933487, + 0.009759202599525452, + 0.03421546891331673, + 0.01578228548169136, + 0.01495254598557949, + 0.010853859595954418, + 0.017339689657092094, + 0.01953912526369095, + 0.015966933220624924, + 0.00744208786636591, + 0.018316688016057014, + 0.012093568220734596, + 0.006556755397468805, + 0.006551068276166916, + 0.01104151550680399, + 0.008455694653093815, + 0.006556229200214148, + 0.005642069969326258, + 0.008693178184330463, + 0.01010771282017231, + 0.010914966464042664, + 0.02164357528090477, + 0.01660102605819702, + 0.018632572144269943, + 0.027815984562039375, + 0.010037827305495739, + 0.01752178929746151, + 0.01418488472700119, + 0.010161624290049076, + 0.01276693120598793, + 0.014601016417145729, + 0.03052593767642975, + 0.014759995974600315, + 0.013620047830045223, + 0.012707680463790894, + 0.014882249757647514, + 0.002733276691287756, + 0.002676433650776744, + 0.004647449590265751, + 0.003731098957359791, + 0.004158033989369869, + 0.002026168629527092, + 0.003583980491384864, + 0.005473351571708918, + 0.009333196096122265, + 0.0053335935808718204, + 0.008025568909943104, + 0.00426497170701623, + 0.00899912416934967, + 0.0032812729477882385, + 0.006271450314670801, + 0.014078992418944836, + 0.009473301470279694, + 0.007188187912106514, + 0.007776163052767515, + 0.005675786640495062, + 0.005597593728452921, + 0.01346774585545063, + 0.003972827456891537, + 0.010390156880021095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.005250830203294754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a fairy wearing an ancient costume on a grand palace terrace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the floor made of smooth, carved white marble illuminated by cool, watery moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fairy wear a pure white, wide-sleeved flowing dress with silver-thread cloud patterns on the cuffs and hem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fairy slowly raise her arm, twist her wrist, and swing her long sleeves out gracefully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle and slowly pulling out as the subject moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700010359287262 + ], + "opposite_logprob_tokens": [ + -2.6950011253356934 + ], + "logprob_reward": -0.0700010359287262, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700010359287262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12441693991422653, + 0.17117904126644135, + 0.1234729066491127, + 0.1419687420129776, + 0.11553192138671875, + 0.0873812586069107, + 0.13821500539779663, + 0.03275727480649948, + 0.03742756322026253, + 0.027649348601698875 + ], + "credit_map": [ + 0.01330021396279335, + 0.01632085070014, + 0.015700029209256172, + 0.014569601975381374, + 0.021458875387907028, + 0.017791535705327988, + 0.009643460623919964, + 0.01563236676156521, + 0.014620168134570122, + 0.02131938375532627, + 0.0179478507488966, + 0.02126440219581127, + 0.04507071524858475, + 0.012296605855226517, + 0.020318729802966118, + 0.0183411855250597, + 0.00918632373213768, + 0.012767591513693333, + 0.015288245864212513, + 0.01792825385928154, + 0.02030590921640396, + 0.011426622048020363, + 0.024268636479973793, + 0.012301316484808922, + 0.010332807898521423, + 0.013066106475889683, + 0.017719799652695656, + 0.020595986396074295, + 0.023760855197906494, + 0.009248292073607445, + 0.03147595375776291, + 0.01576894335448742, + 0.01094106025993824, + 0.008684314787387848, + 0.016493648290634155, + 0.014962355606257915, + 0.014308521524071693, + 0.008514847606420517, + 0.026700478047132492, + 0.014926685951650143, + 0.007186973467469215, + 0.011175401508808136, + 0.011945218779146671, + 0.011907006613910198, + 0.011988986283540726, + 0.006901949644088745, + 0.013510392978787422, + 0.012765321880578995, + 0.015495644882321358, + 0.010742194950580597, + 0.019069796428084373, + 0.016733938828110695, + 0.011348503641784191, + 0.01109104510396719, + 0.027311690151691437, + 0.026422183960676193, + 0.003566173603758216, + 0.0030802800320088863, + 0.006826512049883604, + 0.0038243138696998358, + 0.00424489239230752, + 0.0018539616139605641, + 0.00461558299139142, + 0.004745555575937033, + 0.005737319588661194, + 0.0033403486013412476, + 0.0066121285781264305, + 0.003359353868290782, + 0.006601287052035332, + 0.002019253559410572, + 0.003552565351128578, + 0.006205303594470024, + 0.004328576847910881, + 0.0039290147833526134, + 0.004025086294859648, + 0.002661519218236208, + 0.004225156735628843, + 0.00230251788161695, + 0.0022750950884073973, + 0.0039023817516863346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.0700010359287262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_15/index_tensor([3491, 3491])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_15/index_tensor([3491, 3491])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..37665e16cff9ea6b3b0988f92963d4c5be9effbc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_15/index_tensor([3491, 3491])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, wide-angle shot captured from a drone's bird's-eye perspective captures the vast expanse of the ocean, stretching to the clean, distant horizon. The deep blue waters swell and subside in a slow, hypnotic rhythm, their surface texture punctuated by countless foamy whitecaps whose crest and dissolve suggest a steady, gentle breeze. Bright, unfiltered sunlight pours down from above, casting a brilliant, shimmering sheen on the water, making it shimmer. Above, the sky is a vast, pale, almost ethereal canvas of blue, dotted with wispy, slow-moving cirrus clouds. A small section of an industrial oil rig can be seen in the upper left corner of the frame; its weathered, rust-red and yellow steel beams and platforms create a stark, geometric contrast to the organic seascape. The composition emphasizes the immense depth and scale of the ocean, creating a tranquil and majestic atmosphere, evoking a profound sense of natural grandeur juxtaposed with the subtle traces of human presence.", + "seed": 8971, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_15/index_3491_seed_8971_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015843475237488747, + "vlm_logprob": -0.015843475237488747, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.015843475237488747, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015047727152705193 + ], + "opposite_logprob_tokens": [ + -4.640047550201416 + ], + "logprob_reward": -0.015047727152705193, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015047727152705193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11255388706922531, + 0.10184835642576218, + 0.04878784716129303, + 0.07354696094989777, + 0.08493920415639877, + 0.034983620047569275, + 0.10492940992116928, + 0.07679387181997299, + 0.33288541436195374, + 0.028731435537338257 + ], + "credit_map": [ + 0.009365084581077099, + 0.017304714769124985, + 0.01457246858626604, + 0.014097564853727818, + 0.01347810123115778, + 0.009599462151527405, + 0.012344248592853546, + 0.02179224044084549, + 0.0063794623129069805, + 0.0248183012008667, + 0.024509305134415627, + 0.011569398455321789, + 0.010775156319141388, + 0.007521810941398144, + 0.005820014514029026, + 0.01045490987598896, + 0.0035664737224578857, + 0.014375578612089157, + 0.005358501803129911, + 0.004099908284842968, + 0.0092926612123847, + 0.0045777917839586735, + 0.0027366522699594498, + 0.0047802794724702835, + 0.005011392291635275, + 0.018485596403479576, + 0.00956609658896923, + 0.015549913048744202, + 0.010016996413469315, + 0.005427361465990543, + 0.0040158433839678764, + 0.005473761353641748, + 0.005792693234980106, + 0.02153737097978592, + 0.011282356455922127, + 0.017178703099489212, + 0.011344164609909058, + 0.0072948746383190155, + 0.0042746602557599545, + 0.006234380882233381, + 0.004215050954371691, + 0.0054619270376861095, + 0.006094763055443764, + 0.004778256174176931, + 0.0038126816507428885, + 0.004137023817747831, + 0.003366241231560707, + 0.0031176761258393526, + 0.007119376212358475, + 0.012781224213540554, + 0.015770483762025833, + 0.011213022284209728, + 0.009591169655323029, + 0.011450070887804031, + 0.016529737040400505, + 0.020474324002861977, + 0.005037209950387478, + 0.008713293820619583, + 0.014696582220494747, + 0.008782639168202877, + 0.007582528050988913, + 0.011306094005703926, + 0.01025642640888691, + 0.010419093072414398, + 0.009815547615289688, + 0.05948220193386078, + 0.04458649829030037, + 0.035010386258363724, + 0.027297569438815117, + 0.05419363081455231, + 0.05499871447682381, + 0.04750087112188339, + 0.0017500885296612978, + 0.005404411815106869, + 0.005418062675744295, + 0.0039717042818665504, + 0.0032183353323489428, + 0.001856214483268559, + 0.00402701273560524, + 0.003085606498643756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.015047727152705193 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004106542095541954 + ], + "opposite_logprob_tokens": [ + -5.504106521606445 + ], + "logprob_reward": -0.004106542095541954, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004106542095541954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12955084443092346, + 0.13432295620441437, + 0.22312824428081512, + 0.1254744976758957, + 0.08092556148767471, + 0.04793871194124222, + 0.05501246824860573, + 0.04712911695241928, + 0.11916843801736832, + 0.03734917566180229 + ], + "credit_map": [ + 0.017842935398221016, + 0.014876737259328365, + 0.018871163949370384, + 0.019840838387608528, + 0.012483692727982998, + 0.011319045908749104, + 0.018938999623060226, + 0.015377422794699669, + 0.008615253493189812, + 0.022030726075172424, + 0.03234086558222771, + 0.017522992566227913, + 0.0131387235596776, + 0.011766055598855019, + 0.01207482349127531, + 0.016833513975143433, + 0.013456027023494244, + 0.02837725542485714, + 0.037529077380895615, + 0.02943328022956848, + 0.021261267364025116, + 0.030028702691197395, + 0.029476899653673172, + 0.03356573358178139, + 0.006141072604805231, + 0.02594424970448017, + 0.0150216743350029, + 0.02441321313381195, + 0.014152844436466694, + 0.009257914498448372, + 0.011490115895867348, + 0.019053397700190544, + 0.005415377207100391, + 0.018971547484397888, + 0.012068432755768299, + 0.01844877377152443, + 0.009608084335923195, + 0.005038068164139986, + 0.005465465597808361, + 0.005909816827625036, + 0.0028505928348749876, + 0.014979694038629532, + 0.004583398345857859, + 0.004289188887923956, + 0.009481973946094513, + 0.005777512677013874, + 0.002730961423367262, + 0.00324539327993989, + 0.003986626397818327, + 0.00971776433289051, + 0.008035175502300262, + 0.0055083041079342365, + 0.005876684095710516, + 0.010883972980082035, + 0.006165352649986744, + 0.004838586784899235, + 0.0025885712821036577, + 0.006975505966693163, + 0.005810153670608997, + 0.005756543483585119, + 0.0058494131080806255, + 0.007146550342440605, + 0.005941471550613642, + 0.007060904521495104, + 0.006784706376492977, + 0.03057420812547207, + 0.014185581356287003, + 0.011454948224127293, + 0.014508738182485104, + 0.019264062866568565, + 0.01224976684898138, + 0.010146436281502247, + 0.0029520371463149786, + 0.007259744219481945, + 0.004659246187657118, + 0.0037747423630207777, + 0.004685177467763424, + 0.005795923061668873, + 0.004283939953893423, + 0.0039383647963404655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.004106542095541954 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195652678608894 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195652678608894, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195652678608894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07543211430311203, + 0.23115628957748413, + 0.023174060508608818, + 0.11113859713077545, + 0.3758790194988251, + 0.01978212036192417, + 0.05664590001106262, + 0.02246182970702648, + 0.046956129372119904, + 0.03737393766641617 + ], + "credit_map": [ + 0.016059881076216698, + 0.008849927224218845, + 0.009681390598416328, + 0.010257097892463207, + 0.008935840800404549, + 0.006763762328773737, + 0.006204165052622557, + 0.00868004746735096, + 0.011345947161316872, + 0.06800249963998795, + 0.05296063423156738, + 0.02322322130203247, + 0.0304282084107399, + 0.018224021419882774, + 0.010680992156267166, + 0.016290739178657532, + 0.002542199334129691, + 0.003380360547453165, + 0.00425973255187273, + 0.003703013528138399, + 0.0021431962959468365, + 0.002660998608916998, + 0.002250523306429386, + 0.0022340358700603247, + 0.00824383832514286, + 0.030234554782509804, + 0.015395967289805412, + 0.02035844326019287, + 0.016332045197486877, + 0.007727907504886389, + 0.005199624691158533, + 0.007646217942237854, + 0.01970664970576763, + 0.10498508810997009, + 0.05191684886813164, + 0.08585181087255478, + 0.05179830268025398, + 0.02526869811117649, + 0.017575383186340332, + 0.018776260316371918, + 0.0027623234782367945, + 0.0027816712390631437, + 0.0039056953974068165, + 0.003594858106225729, + 0.001963150454685092, + 0.002012676326557994, + 0.0014203109312802553, + 0.0013414345448836684, + 0.007233497221022844, + 0.008344292640686035, + 0.009109392762184143, + 0.005526434164494276, + 0.007192968390882015, + 0.006544142495840788, + 0.0062621114775538445, + 0.006433059927076101, + 0.0027308864519000053, + 0.003813046496361494, + 0.003073325613513589, + 0.0019499753834679723, + 0.002814833540469408, + 0.003373132785782218, + 0.002592615084722638, + 0.0021140137687325478, + 0.0025703241117298603, + 0.012866714037954807, + 0.0066919270902872086, + 0.0041885687969625, + 0.005790382158011198, + 0.006671376060694456, + 0.004838821943849325, + 0.0033380144741386175, + 0.003057233989238739, + 0.007158961612731218, + 0.005897060502320528, + 0.003810596652328968, + 0.004702732898294926, + 0.004138264339417219, + 0.004730802960693836, + 0.0038782842457294464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.014195652678608894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073646023869514 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073646023869514, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073646023869514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10691403597593307, + 0.28112295269966125, + 0.1164601668715477, + 0.14324291050434113, + 0.056349631398916245, + 0.0681423544883728, + 0.07645955681800842, + 0.03699192404747009, + 0.07414648681879044, + 0.04016992822289467 + ], + "credit_map": [ + 0.022993987426161766, + 0.013207399286329746, + 0.011792321689426899, + 0.0126199284568429, + 0.010007940232753754, + 0.008856204338371754, + 0.011101641692221165, + 0.016334624961018562, + 0.015780720859766006, + 0.09253744781017303, + 0.0421011745929718, + 0.02041180059313774, + 0.05201846733689308, + 0.025193674489855766, + 0.01186323817819357, + 0.021216467022895813, + 0.012299928814172745, + 0.02856711484491825, + 0.020052555948495865, + 0.012613562867045403, + 0.01256904099136591, + 0.013046986423432827, + 0.0066422875970602036, + 0.010668710805475712, + 0.007788509596139193, + 0.04270116984844208, + 0.017540013417601585, + 0.030466141179203987, + 0.021281516179442406, + 0.010329942218959332, + 0.0056526935659348965, + 0.007482948247343302, + 0.00412860419601202, + 0.014937246218323708, + 0.0068789259530603886, + 0.01253974623978138, + 0.00722174346446991, + 0.004146616905927658, + 0.003272396745160222, + 0.0032243637833744287, + 0.004698028787970543, + 0.02284429594874382, + 0.0062617650255560875, + 0.009102534502744675, + 0.013752499595284462, + 0.004886258393526077, + 0.0033445204608142376, + 0.0032524627167731524, + 0.006769587751477957, + 0.013058646582067013, + 0.009924931451678276, + 0.007999232970178127, + 0.008768687024712563, + 0.006902279332280159, + 0.010845150798559189, + 0.01219105813652277, + 0.0040338607504963875, + 0.0048511577770113945, + 0.008180061355233192, + 0.004867956042289734, + 0.003068712539970875, + 0.002870237920433283, + 0.004908823873847723, + 0.004211120773106813, + 0.003062270814552903, + 0.013486400246620178, + 0.008115556091070175, + 0.007618776988238096, + 0.0069275060668587685, + 0.014560682699084282, + 0.011315836571156979, + 0.00905947107821703, + 0.0041555846109986305, + 0.0074012866243720055, + 0.005545012187212706, + 0.004443448036909103, + 0.004705845844000578, + 0.00653975922614336, + 0.004101070575416088, + 0.0032779276371002197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.016073646023869514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793808236718178 + ], + "opposite_logprob_tokens": [ + -3.5297937393188477 + ], + "logprob_reward": -0.029793808236718178, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793808236718178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14779919385910034, + 0.10763858258724213, + 0.03811929002404213, + 0.04215612635016441, + 0.300150990486145, + 0.03896710276603699, + 0.048155441880226135, + 0.03959282487630844, + 0.1618604212999344, + 0.07556001842021942 + ], + "credit_map": [ + 0.011207900010049343, + 0.0169530026614666, + 0.017061345279216766, + 0.022082895040512085, + 0.014643332920968533, + 0.012395437806844711, + 0.03540483117103577, + 0.018050475046038628, + 0.007170389406383038, + 0.03125143423676491, + 0.020423583686351776, + 0.009890983812510967, + 0.013938186690211296, + 0.011008279398083687, + 0.0061218636110424995, + 0.007833873853087425, + 0.0051881978288292885, + 0.005523476749658585, + 0.005340318661183119, + 0.004532909020781517, + 0.004574424587190151, + 0.004768384620547295, + 0.004115814808756113, + 0.004075766075402498, + 0.0027163883205503225, + 0.008038372732698917, + 0.0053770896047353745, + 0.00980209931731224, + 0.004622739274054766, + 0.0030419223476201296, + 0.003677064087241888, + 0.004880458116531372, + 0.010271599516272545, + 0.08594714105129242, + 0.03667155280709267, + 0.07106655836105347, + 0.03901190310716629, + 0.02701539173722267, + 0.014642930589616299, + 0.015523925423622131, + 0.0024619519244879484, + 0.006575591396540403, + 0.004288144875317812, + 0.0053967139683663845, + 0.006031928583979607, + 0.004012984689325094, + 0.005495966877788305, + 0.0047038234770298, + 0.004355830140411854, + 0.009461339563131332, + 0.005163383204489946, + 0.00476681487634778, + 0.005653703585267067, + 0.007040502969175577, + 0.007026602979749441, + 0.004687273409217596, + 0.00284153176471591, + 0.008158206008374691, + 0.00528654595836997, + 0.00532566336914897, + 0.005444674752652645, + 0.004275654908269644, + 0.003851453075185418, + 0.004409100394695997, + 0.005415038671344519, + 0.02326291613280773, + 0.020011739805340767, + 0.020099524408578873, + 0.014049366116523743, + 0.02252265065908432, + 0.03225904703140259, + 0.024240151047706604, + 0.003556231502443552, + 0.020012304186820984, + 0.007283756509423256, + 0.007392331957817078, + 0.011847686022520065, + 0.011864609085023403, + 0.006659625098109245, + 0.006943475920706987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.029793808236718178 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot captured from a drone's bird's-eye perspective captures the vast expanse of the ocean, stretching to the clean, distant horizon. The deep blue waters swell and subside in a slow, hypnotic rhythm, their surface texture punctuated by countless foamy whitecaps whose crest and dissolve suggest a steady, gentle breeze. Bright, unfiltered sunlight pours down from above, casting a brilliant, shimmering sheen on the water, making it shimmer. Above, the sky is a vast, pale, almost ethereal canvas of blue, dotted with wispy, slow-moving cirrus clouds. A small section of an industrial oil rig can be seen in the upper left corner of the frame; its weathered, rust-red and yellow steel beams and platforms create a stark, geometric contrast to the organic seascape. The composition emphasizes the immense depth and scale of the ocean, creating a tranquil and majestic atmosphere, evoking a profound sense of natural grandeur juxtaposed with the subtle traces of human presence.", + "seed": 8972, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_15/index_3491_seed_8972_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01298833847977221, + "vlm_logprob": -0.01298833847977221, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.01298833847977221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03161539137363434 + ], + "opposite_logprob_tokens": [ + -3.6566154956817627 + ], + "logprob_reward": -0.03161539137363434, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03161539137363434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28499358892440796, + 0.04125451296567917, + 0.05738023295998573, + 0.061247818171978, + 0.01710538938641548, + 0.03403325006365776, + 0.09295540302991867, + 0.09229574352502823, + 0.2815827429294586, + 0.03715129569172859 + ], + "credit_map": [ + 0.015451967716217041, + 0.04397829249501228, + 0.029232561588287354, + 0.03483572229743004, + 0.033495280891656876, + 0.03054642491042614, + 0.03734496235847473, + 0.06010837107896805, + 0.006873446516692638, + 0.00426891865208745, + 0.004746568389236927, + 0.0042899553664028645, + 0.0056373355910182, + 0.005382352042943239, + 0.004309230949729681, + 0.005746702663600445, + 0.006899582222104073, + 0.012106366455554962, + 0.0056641180999577045, + 0.005900386720895767, + 0.007838897407054901, + 0.008852077648043633, + 0.004219030495733023, + 0.005899771116673946, + 0.010606758296489716, + 0.00744810001924634, + 0.010916264727711678, + 0.008610791526734829, + 0.004606850445270538, + 0.0057224552147090435, + 0.006272480357438326, + 0.007064122706651688, + 0.002137066563591361, + 0.0034056585282087326, + 0.002083883620798588, + 0.0015076050767675042, + 0.001812590635381639, + 0.0023796677123755217, + 0.0018994527636095881, + 0.001879464485682547, + 0.0052901073358953, + 0.004055317025631666, + 0.005548980552703142, + 0.003614503890275955, + 0.003651670180261135, + 0.0033613373525440693, + 0.004594974219799042, + 0.003916358109563589, + 0.010072806850075722, + 0.013913450762629509, + 0.013959850184619427, + 0.011033355258405209, + 0.012223043479025364, + 0.01290302723646164, + 0.00995846651494503, + 0.008891407400369644, + 0.009867334738373756, + 0.01655769906938076, + 0.00948962103575468, + 0.005796035286039114, + 0.021562758833169937, + 0.016117282211780548, + 0.007309929467737675, + 0.005595080088824034, + 0.03212449699640274, + 0.04884478822350502, + 0.022304954007267952, + 0.028300389647483826, + 0.04084053263068199, + 0.04027842730283737, + 0.039945513010025024, + 0.028943629935383797, + 0.0031639246735721827, + 0.005724108312278986, + 0.004606086295098066, + 0.00427532522007823, + 0.005932386498898268, + 0.005103664472699165, + 0.003990486264228821, + 0.004355313256382942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.03161539137363434 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00250038574449718 + ], + "opposite_logprob_tokens": [ + -6.002500534057617 + ], + "logprob_reward": -0.00250038574449718, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00250038574449718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15778222680091858, + 0.04030211269855499, + 0.07815243303775787, + 0.07047474384307861, + 0.0360085666179657, + 0.029553376138210297, + 0.07926858961582184, + 0.10585439205169678, + 0.36864376068115234, + 0.033959828317165375 + ], + "credit_map": [ + 0.010589764453470707, + 0.03182188421487808, + 0.02155328169465065, + 0.01363737229257822, + 0.023203084245324135, + 0.019806088879704475, + 0.01705261692404747, + 0.020118137821555138, + 0.003735960926860571, + 0.00804436206817627, + 0.003136918181553483, + 0.00461930176243186, + 0.00445264158770442, + 0.005874405149370432, + 0.005096601787954569, + 0.005341924261301756, + 0.010840069502592087, + 0.01488318108022213, + 0.008098617196083069, + 0.008189049549400806, + 0.008751085959374905, + 0.01069258525967598, + 0.0075125847943127155, + 0.009185264818370342, + 0.011783462949097157, + 0.009448876604437828, + 0.011873627081513405, + 0.00969288032501936, + 0.005753203295171261, + 0.007845005951821804, + 0.005947849713265896, + 0.008129841648042202, + 0.0051009501330554485, + 0.0056549012660980225, + 0.005050186533480883, + 0.004042304120957851, + 0.003342630807310343, + 0.004417036660015583, + 0.004040444269776344, + 0.004360112361609936, + 0.0028992488514631987, + 0.0027924298774451017, + 0.002827535616233945, + 0.0037957485765218735, + 0.0037419109139591455, + 0.0031297639943659306, + 0.005176804959774017, + 0.005189933348447084, + 0.009250431321561337, + 0.011521119624376297, + 0.012141561135649681, + 0.00935825053602457, + 0.00981279369443655, + 0.011194984428584576, + 0.008055487647652626, + 0.007933960296213627, + 0.014468451961874962, + 0.015819286927580833, + 0.01455688662827015, + 0.008335791528224945, + 0.01861215941607952, + 0.014498980715870857, + 0.011256713420152664, + 0.00830612238496542, + 0.03926503658294678, + 0.06881654262542725, + 0.026524024084210396, + 0.037032350897789, + 0.05616578459739685, + 0.055407993495464325, + 0.050977278500795364, + 0.03445476293563843, + 0.0015457364497706294, + 0.007757793180644512, + 0.0016839377349242568, + 0.002472828608006239, + 0.007134386338293552, + 0.006113686133176088, + 0.0028134637977927923, + 0.0044379932805895805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.00250038574449718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093957599252462 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004093957599252462, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093957599252462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19559112191200256, + 0.05695025250315666, + 0.06862831115722656, + 0.07058282196521759, + 0.05143958330154419, + 0.02757604792714119, + 0.06449620425701141, + 0.06771699339151382, + 0.3476637899875641, + 0.04935488477349281 + ], + "credit_map": [ + 0.016838083043694496, + 0.03328432887792587, + 0.03135538473725319, + 0.018458081409335136, + 0.030011314898729324, + 0.02254423126578331, + 0.01586148515343666, + 0.027238210663199425, + 0.009959200397133827, + 0.0070762550458312035, + 0.006991653703153133, + 0.00551023380830884, + 0.005596095696091652, + 0.005915196146816015, + 0.0050637321546673775, + 0.010837883688509464, + 0.008468965999782085, + 0.012020879425108433, + 0.011340025812387466, + 0.005264750216156244, + 0.009741285815834999, + 0.008780552074313164, + 0.005444047972559929, + 0.007567804306745529, + 0.0125376982614398, + 0.009423620998859406, + 0.012395363301038742, + 0.00908293854445219, + 0.006165317725390196, + 0.0075088683515787125, + 0.005760732106864452, + 0.007708280812948942, + 0.008045968599617481, + 0.008441728539764881, + 0.00819624587893486, + 0.004457307513803244, + 0.006752131972461939, + 0.006958552170544863, + 0.004494893364608288, + 0.004092751070857048, + 0.003265032544732094, + 0.003888145089149475, + 0.0038181578274816275, + 0.0024175557773560286, + 0.0038427470717579126, + 0.0033072312362492085, + 0.003463259432464838, + 0.0035739173181355, + 0.006921559572219849, + 0.010558970272541046, + 0.009127842262387276, + 0.006595769431442022, + 0.010203802958130836, + 0.009966609068214893, + 0.005521045532077551, + 0.005600606556981802, + 0.007820752449333668, + 0.008607551455497742, + 0.008200551383197308, + 0.006959779653698206, + 0.010382971726357937, + 0.007824045605957508, + 0.010774483904242516, + 0.007146862801164389, + 0.039519406855106354, + 0.0619933158159256, + 0.027362661436200142, + 0.03502601385116577, + 0.05077982693910599, + 0.05023317411541939, + 0.04865305870771408, + 0.034096330404281616, + 0.007450081408023834, + 0.005723463837057352, + 0.006521198432892561, + 0.006504818331450224, + 0.0056853084824979305, + 0.004971354268491268, + 0.005955950357019901, + 0.006542712450027466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.004093957599252462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542181648314 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542181648314, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542181648314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3075675368309021, + 0.09611062705516815, + 0.08732069283723831, + 0.1448400914669037, + 0.08078180253505707, + 0.03638879209756851, + 0.070340596139431, + 0.033071137964725494, + 0.09908968210220337, + 0.044489141553640366 + ], + "credit_map": [ + 0.02759731002151966, + 0.07663647830486298, + 0.0204591304063797, + 0.019286705181002617, + 0.06076657399535179, + 0.05202816054224968, + 0.0224000196903944, + 0.028393136337399483, + 0.014153017662465572, + 0.013946880586445332, + 0.009808084927499294, + 0.009697461500763893, + 0.013577615842223167, + 0.012231552973389626, + 0.010250965133309364, + 0.012445029802620411, + 0.012956425547599792, + 0.011591321788728237, + 0.014430228620767593, + 0.009651826694607735, + 0.011060190387070179, + 0.00994733814150095, + 0.0070478832349181175, + 0.010635461658239365, + 0.024297939613461494, + 0.019595082849264145, + 0.026256607845425606, + 0.017196405678987503, + 0.01487139705568552, + 0.016392620280385017, + 0.012159839272499084, + 0.014070185832679272, + 0.015581251122057438, + 0.009252704679965973, + 0.013212540186941624, + 0.008555085398256779, + 0.009273446165025234, + 0.00945694837719202, + 0.007500328589230776, + 0.007949492894113064, + 0.008054021745920181, + 0.003880229080095887, + 0.005919605493545532, + 0.0033559980802237988, + 0.0034900358878076077, + 0.003416998777538538, + 0.004281756933778524, + 0.003990144468843937, + 0.007915536873042583, + 0.014411197043955326, + 0.005926718004047871, + 0.004989541135728359, + 0.014444252476096153, + 0.012474888935685158, + 0.005418506450951099, + 0.004759946372359991, + 0.0054073031060397625, + 0.003990081138908863, + 0.005299967247992754, + 0.00364809506572783, + 0.0039780051447451115, + 0.004008112940937281, + 0.003047218080610037, + 0.0036923526786267757, + 0.011971166357398033, + 0.019788680598139763, + 0.007767192088067532, + 0.008625420741736889, + 0.01582874171435833, + 0.015878312289714813, + 0.011404800228774548, + 0.007825363427400589, + 0.005834060721099377, + 0.007184003014117479, + 0.0037479191087186337, + 0.004032057244330645, + 0.0069847372360527515, + 0.006143692880868912, + 0.005614934954792261, + 0.00494772894307971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.012542181648314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418977603316307 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.01418977603316307, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418977603316307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13252224028110504, + 0.027989614754915237, + 0.06191517412662506, + 0.0524299293756485, + 0.01886177621781826, + 0.014035393483936787, + 0.06899947673082352, + 0.061006881296634674, + 0.21456503868103027, + 0.3476744294166565 + ], + "credit_map": [ + 0.00856853649020195, + 0.031147466972470284, + 0.013926669023931026, + 0.008416697382926941, + 0.025563331320881844, + 0.021578330546617508, + 0.009633418172597885, + 0.013687793165445328, + 0.003427772782742977, + 0.0051688579842448235, + 0.0025168354623019695, + 0.003050504717975855, + 0.0025767225306481123, + 0.003770118346437812, + 0.0034272028133273125, + 0.004051600582897663, + 0.007657593581825495, + 0.009140332229435444, + 0.006440052762627602, + 0.006517369300127029, + 0.006377268582582474, + 0.007330401800572872, + 0.00787306297570467, + 0.010579094290733337, + 0.00899802427738905, + 0.006906453985720873, + 0.009289058856666088, + 0.0069785346277058125, + 0.00428320886567235, + 0.0057115256786346436, + 0.004540607798844576, + 0.005722519010305405, + 0.0031108208931982517, + 0.002814290812239051, + 0.0030560388695448637, + 0.002189795719459653, + 0.0016384848859161139, + 0.0022133621387183666, + 0.0019169454462826252, + 0.001922037685289979, + 0.001934210304170847, + 0.0014676412101835012, + 0.001868598279543221, + 0.0018528426298871636, + 0.001342125702649355, + 0.0014391627628356218, + 0.002188456943258643, + 0.0019423561170697212, + 0.008096159435808659, + 0.009359881281852722, + 0.01138075441122055, + 0.007953515276312828, + 0.008273881860077381, + 0.009588056243956089, + 0.006993756629526615, + 0.007353469729423523, + 0.008097981102764606, + 0.007827501744031906, + 0.008841542527079582, + 0.004866538569331169, + 0.010390778072178364, + 0.007843796163797379, + 0.007614844478666782, + 0.005523905158042908, + 0.02440115250647068, + 0.039050448685884476, + 0.016081783920526505, + 0.021976374089717865, + 0.03136768564581871, + 0.031476620584726334, + 0.029776697978377342, + 0.02043428085744381, + 0.015498328022658825, + 0.08593058586120605, + 0.013658368028700352, + 0.018153883516788483, + 0.08351482450962067, + 0.07107609510421753, + 0.024101193994283676, + 0.03574113920331001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.01418977603316307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot captured from a drone's bird's-eye perspective captures the vast expanse of the ocean, stretching to the clean, distant horizon. The deep blue waters swell and subside in a slow, hypnotic rhythm, their surface texture punctuated by countless foamy whitecaps whose crest and dissolve suggest a steady, gentle breeze. Bright, unfiltered sunlight pours down from above, casting a brilliant, shimmering sheen on the water, making it shimmer. Above, the sky is a vast, pale, almost ethereal canvas of blue, dotted with wispy, slow-moving cirrus clouds. A small section of an industrial oil rig can be seen in the upper left corner of the frame; its weathered, rust-red and yellow steel beams and platforms create a stark, geometric contrast to the organic seascape. The composition emphasizes the immense depth and scale of the ocean, creating a tranquil and majestic atmosphere, evoking a profound sense of natural grandeur juxtaposed with the subtle traces of human presence.", + "seed": 8973, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_15/index_3491_seed_8973_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05511645823717117, + "vlm_logprob": -0.05511645823717117, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.05511645823717117, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03575750067830086 + ], + "opposite_logprob_tokens": [ + -3.535757541656494 + ], + "logprob_reward": -0.03575750067830086, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03575750067830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7448316812515259, + 0.03356718644499779, + 0.04189559817314148, + 0.06065633147954941, + 0.012361192144453526, + 0.01750251278281212, + 0.02730061113834381, + 0.038970258086919785, + 0.011636598967015743, + 0.01127801276743412 + ], + "credit_map": [ + 0.009877732954919338, + 0.01928696408867836, + 0.007924237288534641, + 0.008656096644699574, + 0.33947083353996277, + 0.3349069058895111, + 0.009105436503887177, + 0.015603551641106606, + 0.003103302326053381, + 0.0065452526323497295, + 0.0024613363202661276, + 0.002396772149950266, + 0.005218819715082645, + 0.0052175684832036495, + 0.0022496015299111605, + 0.006374537944793701, + 0.002724934834986925, + 0.0024686225224286318, + 0.002190773142501712, + 0.0021865530870854855, + 0.011089813895523548, + 0.011788349598646164, + 0.002785392804071307, + 0.006661165505647659, + 0.002459869720041752, + 0.009619247168302536, + 0.00235652644187212, + 0.0023830942809581757, + 0.017468491569161415, + 0.02000241167843342, + 0.0036159090232104063, + 0.0027507857885211706, + 0.0009984513744711876, + 0.0015201469650492072, + 0.0006644763052463531, + 0.0009589639375917614, + 0.003272992093116045, + 0.0028039091266691685, + 0.0008635396952740848, + 0.0012787136947736144, + 0.001313507091253996, + 0.0012278605718165636, + 0.0015664588427171111, + 0.001301183016039431, + 0.004679807461798191, + 0.004510091617703438, + 0.001253119669854641, + 0.0016504874220117927, + 0.0018892691005021334, + 0.003466306021437049, + 0.0023937684018164873, + 0.00255462434142828, + 0.005640972405672073, + 0.005432195030152798, + 0.00280808936804533, + 0.003115390194579959, + 0.0033845563884824514, + 0.00645110011100769, + 0.004247969947755337, + 0.003256396623328328, + 0.005895445588976145, + 0.0065355789847671986, + 0.0038145650178194046, + 0.005384649615734816, + 0.001907828263938427, + 0.00098597991745919, + 0.0015087733045220375, + 0.0013980470830574632, + 0.002014643745496869, + 0.0016338452696800232, + 0.0009635747992433608, + 0.0012239086208865047, + 0.0006247733836062253, + 0.0016108950367197394, + 0.0007150302408263087, + 0.0011409010039642453, + 0.0023594116792082787, + 0.0026342044584453106, + 0.0009061983437277377, + 0.0012865993194282055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.03575750067830086 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372311219573021 + ], + "opposite_logprob_tokens": [ + -3.4087231159210205 + ], + "logprob_reward": -0.03372311219573021, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372311219573021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7689545750617981, + 0.04383775591850281, + 0.017781708389520645, + 0.024489931762218475, + 0.014943884685635567, + 0.020866934210062027, + 0.017626652494072914, + 0.03281562030315399, + 0.04264488071203232, + 0.016038037836551666 + ], + "credit_map": [ + 0.009714915417134762, + 0.02322569489479065, + 0.00680126529186964, + 0.007520257029682398, + 0.35133054852485657, + 0.3406640589237213, + 0.00764492666348815, + 0.022052889689803123, + 0.0018985857022926211, + 0.0029279000591486692, + 0.0014491005567833781, + 0.0014282938791438937, + 0.014567715115845203, + 0.014410411939024925, + 0.0020075622014701366, + 0.005148180760443211, + 0.001968720695003867, + 0.002331099472939968, + 0.0012509875232353806, + 0.0015654569724574685, + 0.0032108037266880274, + 0.003444288857281208, + 0.0014802730875089765, + 0.0025300767738372087, + 0.002413779031485319, + 0.0023932934273034334, + 0.002219289541244507, + 0.0019768665079027414, + 0.005881913006305695, + 0.006230770144611597, + 0.0013790179509669542, + 0.0019950009882450104, + 0.0008796937181614339, + 0.0013265054440125823, + 0.0011424635304138064, + 0.0008901978144422174, + 0.004103738348931074, + 0.004730418790131807, + 0.0008431794703938067, + 0.0010276858229190111, + 0.0014681541360914707, + 0.0012509437510743737, + 0.0015041239093989134, + 0.0012006413890048862, + 0.006404652260243893, + 0.006370141636580229, + 0.001106227864511311, + 0.0015620486810803413, + 0.0020617367699742317, + 0.0028969740960747004, + 0.0017640708247199655, + 0.0021981175523251295, + 0.002490350278094411, + 0.0023917900398373604, + 0.0015671728178858757, + 0.0022564383689314127, + 0.0020870177540928125, + 0.00391931738704443, + 0.0030296894256025553, + 0.004656336735934019, + 0.005789085756987333, + 0.00496703153476119, + 0.0030597986187785864, + 0.005307340528815985, + 0.001661902293562889, + 0.006087349262088537, + 0.001830143271945417, + 0.003911138977855444, + 0.010587199591100216, + 0.011744758114218712, + 0.002602999098598957, + 0.004219384863972664, + 0.0008300953195430338, + 0.0016706406604498625, + 0.0012067091884091496, + 0.0018066331977024674, + 0.003364351112395525, + 0.003798009594902396, + 0.0012166269589215517, + 0.0021449709311127663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03372311219573021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496174469590187 + ], + "opposite_logprob_tokens": [ + -2.929961681365967 + ], + "logprob_reward": -0.05496174469590187, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496174469590187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.539124071598053, + 0.05132168531417847, + 0.0953482985496521, + 0.11468825489282608, + 0.019797343760728836, + 0.042461469769477844, + 0.03145677223801613, + 0.04544077813625336, + 0.053624458611011505, + 0.00673686433583498 + ], + "credit_map": [ + 0.008713075891137123, + 0.015446105971932411, + 0.006690809968858957, + 0.005111367907375097, + 0.24661794304847717, + 0.24246758222579956, + 0.0061815702356398106, + 0.007895635440945625, + 0.00856795348227024, + 0.008194790221750736, + 0.0048411013558506966, + 0.003087779274210334, + 0.008626666851341724, + 0.01000696700066328, + 0.00335710565559566, + 0.004639322403818369, + 0.00985074695199728, + 0.009187961928546429, + 0.005009908694773912, + 0.0037224218249320984, + 0.028353845700621605, + 0.03094540536403656, + 0.0028802917804569006, + 0.005397719331085682, + 0.008638941682875156, + 0.016556063666939735, + 0.005867446772754192, + 0.004691595211625099, + 0.03319822624325752, + 0.036703288555145264, + 0.003734043799340725, + 0.005298649426549673, + 0.0017498753732070327, + 0.002683387603610754, + 0.0019587271381169558, + 0.0010820194147527218, + 0.005396554246544838, + 0.004622797016054392, + 0.0012735025957226753, + 0.001030481536872685, + 0.002942790277302265, + 0.004842001479119062, + 0.0025379620492458344, + 0.0021484866738319397, + 0.012276659719645977, + 0.013363673351705074, + 0.0017155471723526716, + 0.002634345320984721, + 0.0018493059324100614, + 0.006342766806483269, + 0.0013207232113927603, + 0.0014082653215155005, + 0.008457389660179615, + 0.009363837540149689, + 0.0012230639113113284, + 0.0014914199709892273, + 0.005307056475430727, + 0.0070380945689976215, + 0.0033544383477419615, + 0.0025330015923827887, + 0.010816586203873158, + 0.011232367716729641, + 0.0017756628803908825, + 0.003383568488061428, + 0.004302841145545244, + 0.009808278642594814, + 0.00250904425047338, + 0.0026314747519791126, + 0.014225590042769909, + 0.015648450702428818, + 0.0018185301451012492, + 0.0026802492793649435, + 0.0006574746803380549, + 0.0008997101685963571, + 0.0008867472060956061, + 0.0009118333691731095, + 0.0008241536561399698, + 0.000650562287773937, + 0.0006101438775658607, + 0.0012962395558133721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.05496174469590187 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08910173177719116 + ], + "opposite_logprob_tokens": [ + -2.464101791381836 + ], + "logprob_reward": -0.08910173177719116, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08910173177719116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5954092144966125, + 0.11070460826158524, + 0.04228669032454491, + 0.0615275539457798, + 0.028810562565922737, + 0.028570963069796562, + 0.018593337386846542, + 0.030729226768016815, + 0.07467851042747498, + 0.008689327165484428 + ], + "credit_map": [ + 0.011414797976613045, + 0.017475184053182602, + 0.006888572592288256, + 0.005093079060316086, + 0.2738814353942871, + 0.2659621834754944, + 0.005543394945561886, + 0.009150588884949684, + 0.005776627920567989, + 0.014107150956988335, + 0.00394445052370429, + 0.0031139401253312826, + 0.03507011756300926, + 0.039415083825588226, + 0.0032692430540919304, + 0.006007996387779713, + 0.0019603956025093794, + 0.004053788725286722, + 0.0018024235032498837, + 0.0013436042936518788, + 0.01350495871156454, + 0.014866478741168976, + 0.001361206523142755, + 0.003393835388123989, + 0.003220056416466832, + 0.007394478190690279, + 0.0039011240005493164, + 0.003988363780081272, + 0.016924947500228882, + 0.019074413925409317, + 0.0024616459850221872, + 0.004562521353363991, + 0.0019975306931883097, + 0.0032153651118278503, + 0.001816953532397747, + 0.0018628156976774335, + 0.008784716948866844, + 0.007825683802366257, + 0.0016596035566180944, + 0.0016478963661938906, + 0.001996646635234356, + 0.0028863793704658747, + 0.001644437201321125, + 0.0013401990290731192, + 0.00854477845132351, + 0.009212378412485123, + 0.0012844552984461188, + 0.0016616897191852331, + 0.0013913081493228674, + 0.0031941416673362255, + 0.001192123512737453, + 0.001289745676331222, + 0.004392777569591999, + 0.004432010930031538, + 0.0011604944011196494, + 0.001540736062452197, + 0.002407233929261565, + 0.0050079054199159145, + 0.002187856938689947, + 0.002375710988417268, + 0.006206435151398182, + 0.006710473448038101, + 0.0023653684183955193, + 0.003468243172392249, + 0.002919791266322136, + 0.010721584782004356, + 0.002911393763497472, + 0.006185790058225393, + 0.019310450181365013, + 0.02146751992404461, + 0.0042529418133199215, + 0.006909043528139591, + 0.0009228482376784086, + 0.001183357904665172, + 0.0011704196222126484, + 0.0008938537212088704, + 0.0013580406084656715, + 0.0015050696674734354, + 0.0007119866786524653, + 0.0009437501430511475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.08910173177719116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062038201838731766 + ], + "opposite_logprob_tokens": [ + -2.8120381832122803 + ], + "logprob_reward": -0.062038201838731766, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062038201838731766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36839666962623596, + 0.04886820167303085, + 0.10519915819168091, + 0.08850022405385971, + 0.0454145148396492, + 0.13978812098503113, + 0.0382949598133564, + 0.1263868808746338, + 0.016935862600803375, + 0.022215453907847404 + ], + "credit_map": [ + 0.01946687325835228, + 0.028450921177864075, + 0.008460567332804203, + 0.013072328642010689, + 0.13508816063404083, + 0.12973575294017792, + 0.009701676666736603, + 0.024420347064733505, + 0.00450931629166007, + 0.007178476545959711, + 0.003268203930929303, + 0.0031134109012782574, + 0.007264345418661833, + 0.00857316330075264, + 0.004005917813628912, + 0.010955363512039185, + 0.009121601469814777, + 0.010186666622757912, + 0.006918981671333313, + 0.005692542064934969, + 0.02770293690264225, + 0.030102074146270752, + 0.004985257051885128, + 0.010489092208445072, + 0.008363724686205387, + 0.009253183379769325, + 0.006165350787341595, + 0.004497496876865625, + 0.023865362629294395, + 0.02687147818505764, + 0.004616514779627323, + 0.004867106676101685, + 0.0035796621814370155, + 0.010265613906085491, + 0.002701535588130355, + 0.002880298299714923, + 0.011484523303806782, + 0.009602597914636135, + 0.002511391881853342, + 0.0023888852447271347, + 0.012648165225982666, + 0.026951998472213745, + 0.013877362944185734, + 0.009159823879599571, + 0.02631780132651329, + 0.0268186517059803, + 0.009905052371323109, + 0.014109249226748943, + 0.003352533793076873, + 0.004792660940438509, + 0.004019621293991804, + 0.004015912767499685, + 0.006903809029608965, + 0.00604865001514554, + 0.004127969034016132, + 0.0050337957218289375, + 0.007982023060321808, + 0.023518869653344154, + 0.010203616693615913, + 0.009027275256812572, + 0.026150530204176903, + 0.02790316939353943, + 0.010275638662278652, + 0.011325743980705738, + 0.001000004936940968, + 0.0022379762958735228, + 0.001167822745628655, + 0.0017542054411023855, + 0.0034831888042390347, + 0.0036213623825460672, + 0.0012752765323966742, + 0.002396024763584137, + 0.0023677574936300516, + 0.0021781460382044315, + 0.002399872522801161, + 0.002455528127029538, + 0.0036598106380552053, + 0.003243899904191494, + 0.0021052819211035967, + 0.0038051565643399954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.062038201838731766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot captured from a drone's bird's-eye perspective captures the vast expanse of the ocean, stretching to the clean, distant horizon. The deep blue waters swell and subside in a slow, hypnotic rhythm, their surface texture punctuated by countless foamy whitecaps whose crest and dissolve suggest a steady, gentle breeze. Bright, unfiltered sunlight pours down from above, casting a brilliant, shimmering sheen on the water, making it shimmer. Above, the sky is a vast, pale, almost ethereal canvas of blue, dotted with wispy, slow-moving cirrus clouds. A small section of an industrial oil rig can be seen in the upper left corner of the frame; its weathered, rust-red and yellow steel beams and platforms create a stark, geometric contrast to the organic seascape. The composition emphasizes the immense depth and scale of the ocean, creating a tranquil and majestic atmosphere, evoking a profound sense of natural grandeur juxtaposed with the subtle traces of human presence.", + "seed": 8974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_15/index_3491_seed_8974_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011268593883141876, + "vlm_logprob": -0.011268593883141876, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.011268593883141876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0188587736338377 + ], + "opposite_logprob_tokens": [ + -4.268858909606934 + ], + "logprob_reward": -0.0188587736338377, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0188587736338377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15339425206184387, + 0.06863541901111603, + 0.05680442973971367, + 0.11636713147163391, + 0.12050214409828186, + 0.09236190468072891, + 0.1053449958562851, + 0.12381760776042938, + 0.13498374819755554, + 0.0277884379029274 + ], + "credit_map": [ + 0.015283536165952682, + 0.018941201269626617, + 0.01781560108065605, + 0.01343946810811758, + 0.017625506967306137, + 0.018985863775014877, + 0.01808216981589794, + 0.03322088345885277, + 0.0063094161450862885, + 0.009994586929678917, + 0.006839613430202007, + 0.00713626854121685, + 0.014189728535711765, + 0.010411982424557209, + 0.006328564137220383, + 0.007425243500620127, + 0.007370073813945055, + 0.006598041392862797, + 0.0072989752516150475, + 0.005230250768363476, + 0.007464379072189331, + 0.00680181011557579, + 0.005805456079542637, + 0.01023543905466795, + 0.009113764390349388, + 0.021321238949894905, + 0.011224414221942425, + 0.013235347345471382, + 0.01499253511428833, + 0.017520060762763023, + 0.011537556536495686, + 0.01742219366133213, + 0.019915802404284477, + 0.016000980511307716, + 0.01479518972337246, + 0.0074726734310388565, + 0.01677052676677704, + 0.011717298999428749, + 0.013513673096895218, + 0.0203159861266613, + 0.012257280759513378, + 0.013244513422250748, + 0.009825851768255234, + 0.00822139997035265, + 0.018017765134572983, + 0.011962495744228363, + 0.007391910068690777, + 0.011440678499639034, + 0.010078237392008305, + 0.016785694286227226, + 0.008278506807982922, + 0.014423657208681107, + 0.014557627961039543, + 0.013527711853384972, + 0.013610289432108402, + 0.014083256945014, + 0.01774599961936474, + 0.015321201644837856, + 0.012919426895678043, + 0.013386161997914314, + 0.01948266103863716, + 0.016653889790177345, + 0.016089966520667076, + 0.012218291871249676, + 0.010518093593418598, + 0.012895672582089901, + 0.008000697940587997, + 0.0157142523676157, + 0.02106633596122265, + 0.01513440441340208, + 0.024893183261156082, + 0.026761094108223915, + 0.0026497861836105585, + 0.003335889894515276, + 0.00241872644983232, + 0.0025628504808992147, + 0.0048675090074539185, + 0.0041553545743227005, + 0.0029955802019685507, + 0.004802738316357136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0188587736338377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003627151483669877 + ], + "opposite_logprob_tokens": [ + -5.628627300262451 + ], + "logprob_reward": -0.003627151483669877, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003627151483669877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19917742908000946, + 0.06378694623708725, + 0.04902467504143715, + 0.11955461651086807, + 0.07995299249887466, + 0.12253208458423615, + 0.15337876975536346, + 0.15165072679519653, + 0.025881780311465263, + 0.035059988498687744 + ], + "credit_map": [ + 0.012677768245339394, + 0.02408268116414547, + 0.028367990627884865, + 0.02194376103579998, + 0.021662171930074692, + 0.023643000051379204, + 0.02500816248357296, + 0.04179190471768379, + 0.004835472442209721, + 0.00828569009900093, + 0.005515632219612598, + 0.0077894763089716434, + 0.016131913289427757, + 0.01042277179658413, + 0.005496911238878965, + 0.0053090681321918964, + 0.006815977394580841, + 0.006342849228531122, + 0.0071123866364359856, + 0.004837729502469301, + 0.007170827127993107, + 0.006882192101329565, + 0.003811327973380685, + 0.0060513876378536224, + 0.006178644485771656, + 0.018537869676947594, + 0.01282435841858387, + 0.01435492280870676, + 0.028597045689821243, + 0.015456629917025566, + 0.0069814808666706085, + 0.016623659059405327, + 0.009967826306819916, + 0.010530083440244198, + 0.0069597987458109856, + 0.006014472804963589, + 0.012989651411771774, + 0.008869572542607784, + 0.009822526946663857, + 0.014799060299992561, + 0.008905146270990372, + 0.018534107133746147, + 0.012323742732405663, + 0.017832569777965546, + 0.023007411509752274, + 0.014624757692217827, + 0.014455595053732395, + 0.0128487478941679, + 0.022276096045970917, + 0.020792553201317787, + 0.018103664740920067, + 0.01600874960422516, + 0.021495264023542404, + 0.016251632943749428, + 0.0200300682336092, + 0.018420744687318802, + 0.019599836319684982, + 0.021543186157941818, + 0.014889555051922798, + 0.014227071776986122, + 0.030460277572274208, + 0.023220211267471313, + 0.014084958471357822, + 0.01362562831491232, + 0.0015534720150753856, + 0.00254213809967041, + 0.002294669160619378, + 0.0030186763033270836, + 0.002823441056534648, + 0.0024538864381611347, + 0.007072555832564831, + 0.004122941289097071, + 0.005961218848824501, + 0.004180307034403086, + 0.003618998918682337, + 0.004195087123662233, + 0.004339217208325863, + 0.004229289945214987, + 0.003750517964363098, + 0.004785351920872927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.003627151483669877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1495806723833084, + 0.13088183104991913, + 0.1285790055990219, + 0.05808502063155174, + 0.04072520509362221, + 0.04545849189162254, + 0.22711804509162903, + 0.08669597655534744, + 0.023049280047416687, + 0.1098264679312706 + ], + "credit_map": [ + 0.022989295423030853, + 0.02338404953479767, + 0.019423196092247963, + 0.010260206647217274, + 0.03162554278969765, + 0.01849927194416523, + 0.009661722928285599, + 0.013737394474446774, + 0.017754053696990013, + 0.017414266243577003, + 0.015861809253692627, + 0.01318977028131485, + 0.026691200211644173, + 0.015544841066002846, + 0.00960791576653719, + 0.014817972667515278, + 0.024767594411969185, + 0.015050179325044155, + 0.019174769520759583, + 0.01347094215452671, + 0.018576672300696373, + 0.013965760357677937, + 0.009404626674950123, + 0.014168454334139824, + 0.00581341702491045, + 0.010581722483038902, + 0.005736351013183594, + 0.004148479085415602, + 0.015298277139663696, + 0.007747826166450977, + 0.0037773738149553537, + 0.0049815708771348, + 0.004524191375821829, + 0.00584016228094697, + 0.004382413811981678, + 0.003647126257419586, + 0.00616483623161912, + 0.005021242890506983, + 0.004689593333750963, + 0.006455639842897654, + 0.007368292659521103, + 0.006511647254228592, + 0.005076471250504255, + 0.005326177924871445, + 0.00549723906442523, + 0.005647778976708651, + 0.0037652074825018644, + 0.006265673786401749, + 0.03992493450641632, + 0.030687252059578896, + 0.030045287683606148, + 0.022215060889720917, + 0.035288162529468536, + 0.028440387919545174, + 0.01756344735622406, + 0.022953514009714127, + 0.012686079367995262, + 0.012711925432085991, + 0.009367350488901138, + 0.008919374085962772, + 0.011849376372992992, + 0.009510279633104801, + 0.012578667141497135, + 0.009072929620742798, + 0.003313362831249833, + 0.003168907016515732, + 0.0022919420152902603, + 0.0022121232468634844, + 0.0035870131105184555, + 0.003042639931663871, + 0.002651168964803219, + 0.002782124327495694, + 0.020482782274484634, + 0.016398169100284576, + 0.0104033462703228, + 0.014687363058328629, + 0.01440830435603857, + 0.013558940030634403, + 0.012926324270665646, + 0.0069612376391887665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.005251067224889994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543123215436935 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543123215436935, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543123215436935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18828049302101135, + 0.07352889329195023, + 0.11596725136041641, + 0.029769718647003174, + 0.07321395725011826, + 0.03683412820100784, + 0.08474097400903702, + 0.10827063769102097, + 0.06695300340652466, + 0.22244097292423248 + ], + "credit_map": [ + 0.02127591334283352, + 0.020458152517676353, + 0.02558586746454239, + 0.01940232515335083, + 0.0322808176279068, + 0.02425772324204445, + 0.018196074292063713, + 0.026823602616786957, + 0.0063302041962742805, + 0.008830228820443153, + 0.008263287134468555, + 0.008317063562572002, + 0.016588464379310608, + 0.009064065292477608, + 0.0057723503559827805, + 0.010363235138356686, + 0.021336905658245087, + 0.016928019002079964, + 0.014029020443558693, + 0.00960371270775795, + 0.018575113266706467, + 0.012383818626403809, + 0.009218730032444, + 0.01389192696660757, + 0.00413827458396554, + 0.004479441791772842, + 0.00403875857591629, + 0.002348354086279869, + 0.00450593326240778, + 0.004612864460796118, + 0.002472988562658429, + 0.0031731014605611563, + 0.015449988655745983, + 0.00935401115566492, + 0.010737307369709015, + 0.004655830096453428, + 0.009070435538887978, + 0.007119451649487019, + 0.006524453870952129, + 0.010302472859621048, + 0.003995921462774277, + 0.004941650666296482, + 0.003975844942033291, + 0.003260272555053234, + 0.006894789636135101, + 0.006129406858235598, + 0.003074659500271082, + 0.004561579320579767, + 0.014611262828111649, + 0.01093112863600254, + 0.011384005658328533, + 0.009466055780649185, + 0.013154953718185425, + 0.010048565454781055, + 0.007792133837938309, + 0.0073528592474758625, + 0.017199207097291946, + 0.012294129468500614, + 0.011954942718148232, + 0.011689575389027596, + 0.016012070700526237, + 0.012248028069734573, + 0.016017183661460876, + 0.010855487547814846, + 0.012429909780621529, + 0.007524065673351288, + 0.008591663092374802, + 0.006011857185512781, + 0.00866802129894495, + 0.007135922089219093, + 0.005815786309540272, + 0.010775773786008358, + 0.04059167578816414, + 0.03475487604737282, + 0.01950463093817234, + 0.027798688039183617, + 0.03072541393339634, + 0.029112286865711212, + 0.02600749582052231, + 0.013945892453193665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.012543123215436935 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, wide-angle drone perspective showing a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are deep blue ocean waters visible with slow, rhythmic swells and foamy whitecaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small industrial oil rig visible in the upper left corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil rig feature weathered steel beams and platforms colored rust-red and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky pale blue with wispy, slow-moving cirrus clouds under bright, unfiltered sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606285385787487 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.01606285385787487, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606285385787487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13911812007427216, + 0.21563610434532166, + 0.04614676907658577, + 0.057519301772117615, + 0.09872101992368698, + 0.12159357964992523, + 0.10827303677797318, + 0.16506893932819366, + 0.01867702603340149, + 0.02924611046910286 + ], + "credit_map": [ + 0.008285141550004482, + 0.01662416011095047, + 0.0086121316999197, + 0.012202314101159573, + 0.025794388726353645, + 0.018858348950743675, + 0.01682853326201439, + 0.031913090497255325, + 0.014610511250793934, + 0.029767243191599846, + 0.019645746797323227, + 0.03126826137304306, + 0.03120412863790989, + 0.026373766362667084, + 0.025084365159273148, + 0.037682101130485535, + 0.006067988462746143, + 0.005919171962887049, + 0.0053800796158611774, + 0.0039058183319866657, + 0.00866988766938448, + 0.0063341171480715275, + 0.0039964900352060795, + 0.005873213522136211, + 0.0031084201764315367, + 0.008456197567284107, + 0.007109437603503466, + 0.007133391685783863, + 0.0124433021992445, + 0.00704437168315053, + 0.004603317007422447, + 0.007620863616466522, + 0.009948761202394962, + 0.011256732046604156, + 0.013038844801485538, + 0.01022079586982727, + 0.015042988583445549, + 0.01321858074516058, + 0.011806652881205082, + 0.014187661930918694, + 0.007274791598320007, + 0.017484454438090324, + 0.00920273456722498, + 0.010711912997066975, + 0.027777651324868202, + 0.023939019069075584, + 0.008946279995143414, + 0.016256744042038918, + 0.01634245738387108, + 0.012231392785906792, + 0.013655602931976318, + 0.01297282986342907, + 0.014106550253927708, + 0.010466821491718292, + 0.015610592439770699, + 0.01288678403943777, + 0.02606915310025215, + 0.01863742806017399, + 0.016629869118332863, + 0.015781275928020477, + 0.02718932181596756, + 0.019782232120633125, + 0.024088827893137932, + 0.016890835016965866, + 0.00235371058806777, + 0.0023626494221389294, + 0.0021442584693431854, + 0.002062792656943202, + 0.0024861874990165234, + 0.0024756768252700567, + 0.0022900232579559088, + 0.0025017273146659136, + 0.0028372490778565407, + 0.0028396397829055786, + 0.003448395524173975, + 0.004211430437862873, + 0.0034814116079360247, + 0.0037049120292067528, + 0.0035378839820623398, + 0.005185186862945557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.01606285385787487 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_tensor([2859, 2859])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_tensor([2859, 2859])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4686e7b90d67924522d74b3e71f2ef6589d4715a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_tensor([2859, 2859])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from an extremely low angle, centered on the one-eyed general, Xiahou Dun, clad in an intricate suit of mythical beast armor. The dark black iron is engraved with roaring kirin heads and intricate runes, and gilded patterns, scarred by battle, appear faintly in the dim light beneath a dark, overcast sky. His face contorts with the intensity of his exertion, his single eye blazing with fury. His muscles tense as he raises a heavy, long-handled halberd high above his head, then brings it down with all his might, crashing heavily upon the cracked, scorched earth. Framed by a maximalist composition, the moment of impact seems to drag the world into silky, high-frame slow motion: the ground instantly cracks, and countless rock fragments of varying sizes, interspersed with dazzling arcs of golden energy and lightning, radiate outward in a gesture of exquisite elegance and violence. The impact point becomes the core light source of the picture, illuminating every flying piece of gravel. In the extreme slow motion, you can even see the clear trajectory of each dust particle picked up by the shock wave rolling and spinning in the air. The whole scene is full of epic and destructive power, presenting a violent aesthetic solidified in time.", + "seed": 24062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_2859_seed_24062_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016681201942265033, + "vlm_logprob": -0.016681201942265033, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.016681201942265033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01910405606031418 + ], + "opposite_logprob_tokens": [ + -4.01910400390625 + ], + "logprob_reward": -0.01910405606031418, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01910405606031418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11711687594652176, + 0.0615769661962986, + 0.29446548223495483, + 0.10055059939622879, + 0.0735282450914383, + 0.0636957660317421, + 0.035650499165058136, + 0.042594462633132935, + 0.14451062679290771, + 0.06631051748991013 + ], + "credit_map": [ + 0.015269641764461994, + 0.01754823513329029, + 0.010757343843579292, + 0.016547786071896553, + 0.013676473870873451, + 0.014772030524909496, + 0.014306365512311459, + 0.014239000156521797, + 0.005826189182698727, + 0.005930103827267885, + 0.007058738265186548, + 0.009478936903178692, + 0.006715130526572466, + 0.007727507967501879, + 0.011257903650403023, + 0.0075824568048119545, + 0.022895734757184982, + 0.04307211935520172, + 0.02202758565545082, + 0.06082587316632271, + 0.033676229417324066, + 0.037520162761211395, + 0.04520057514309883, + 0.029247192665934563, + 0.009207692928612232, + 0.0171139407902956, + 0.01635953225195408, + 0.012351141311228275, + 0.013817783445119858, + 0.012574497610330582, + 0.008427386172115803, + 0.010698622092604637, + 0.008926511742174625, + 0.010256262496113777, + 0.010571612045168877, + 0.007390017155557871, + 0.010733542032539845, + 0.009177039377391338, + 0.007985305972397327, + 0.008487961255013943, + 0.00870888028293848, + 0.01026772242039442, + 0.006631242111325264, + 0.008861125446856022, + 0.00703946128487587, + 0.007878655567765236, + 0.006621938198804855, + 0.007686741650104523, + 0.004980122670531273, + 0.003470450872555375, + 0.004266477655619383, + 0.0037298754323273897, + 0.004559936933219433, + 0.005710737779736519, + 0.004144917242228985, + 0.00478797871619463, + 0.008808689191937447, + 0.006295014172792435, + 0.004335420671850443, + 0.004012291319668293, + 0.0062010339461266994, + 0.004661770071834326, + 0.003418444423004985, + 0.0048618009313941, + 0.016096942126750946, + 0.016461588442325592, + 0.011328919790685177, + 0.013589226640760899, + 0.02991541475057602, + 0.01858481951057911, + 0.0152378985658288, + 0.023295825347304344, + 0.01075511984527111, + 0.008709779940545559, + 0.006070779170840979, + 0.008245852775871754, + 0.00981596577912569, + 0.00793734472244978, + 0.005985915660858154, + 0.008789761923253536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.01910405606031418 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811531513929367 + ], + "opposite_logprob_tokens": [ + -3.2881152629852295 + ], + "logprob_reward": -0.03811531513929367, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811531513929367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09334538877010345, + 0.08034802228212357, + 0.13072514533996582, + 0.057730868458747864, + 0.09563735872507095, + 0.04866411164402962, + 0.0441175177693367, + 0.17626714706420898, + 0.23107577860355377, + 0.04208861663937569 + ], + "credit_map": [ + 0.010651673190295696, + 0.012181906029582024, + 0.00968064647167921, + 0.016380297020077705, + 0.0099817905575037, + 0.008715258911252022, + 0.014460062608122826, + 0.011293752118945122, + 0.0052981507033109665, + 0.00783376395702362, + 0.009716394357383251, + 0.017845239490270615, + 0.011576171033084393, + 0.006657957099378109, + 0.009497569873929024, + 0.011922779493033886, + 0.009096606634557247, + 0.014885902404785156, + 0.01404908299446106, + 0.039107147604227066, + 0.012060156092047691, + 0.015135702677071095, + 0.013786019757390022, + 0.012604538351297379, + 0.009436746127903461, + 0.009330413304269314, + 0.00898987427353859, + 0.006440809927880764, + 0.005833875387907028, + 0.006983432453125715, + 0.0045182970352470875, + 0.006197422277182341, + 0.010408978909254074, + 0.012854177504777908, + 0.014570401050150394, + 0.008698827587068081, + 0.01817401684820652, + 0.012828429229557514, + 0.008618392050266266, + 0.009484141133725643, + 0.0059274411760270596, + 0.005666964687407017, + 0.005382114555686712, + 0.004976080264896154, + 0.004638721700757742, + 0.008210561238229275, + 0.004659532103687525, + 0.00920269824564457, + 0.005556368734687567, + 0.0046408819034695625, + 0.00529403006657958, + 0.0037251710891723633, + 0.008110081776976585, + 0.00854252465069294, + 0.0041370689868927, + 0.0041113924235105515, + 0.03028302639722824, + 0.021430401131510735, + 0.016850439831614494, + 0.015736142173409462, + 0.031589310616254807, + 0.024071451276540756, + 0.013423094525933266, + 0.02288329042494297, + 0.02320897951722145, + 0.026235956698656082, + 0.02345399558544159, + 0.031299836933612823, + 0.04095214605331421, + 0.027648495510220528, + 0.022971181198954582, + 0.0353051982820034, + 0.005428485106676817, + 0.004646902903914452, + 0.0055633969604969025, + 0.007049436680972576, + 0.00551167968660593, + 0.007547291461378336, + 0.003522057319059968, + 0.0028193683829158545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.03811531513929367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632380351424217 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632380351424217, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632380351424217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08163478225469589, + 0.030238289386034012, + 0.16004061698913574, + 0.12797534465789795, + 0.09974480420351028, + 0.19854797422885895, + 0.07852219045162201, + 0.03431752324104309, + 0.14055770635604858, + 0.04842079058289528 + ], + "credit_map": [ + 0.011496339924633503, + 0.011195887811481953, + 0.008190257474780083, + 0.00962772872298956, + 0.010716764256358147, + 0.01094995066523552, + 0.011372861452400684, + 0.008084990084171295, + 0.0026831161230802536, + 0.003431031247600913, + 0.002766852965578437, + 0.007686610333621502, + 0.0026849962305277586, + 0.003420526161789894, + 0.004095880314707756, + 0.0034692774061113596, + 0.010944250039756298, + 0.02513701468706131, + 0.012574808672070503, + 0.027586210519075394, + 0.01894822157919407, + 0.021137939766049385, + 0.025396602228283882, + 0.018315566703677177, + 0.0263237114995718, + 0.01867951825261116, + 0.01613595150411129, + 0.013944891281425953, + 0.015597823075950146, + 0.016005750745534897, + 0.011537077836692333, + 0.009750612080097198, + 0.01182788610458374, + 0.009380881674587727, + 0.016960041597485542, + 0.010676977224647999, + 0.01376492902636528, + 0.014032288454473019, + 0.011379350908100605, + 0.011722448281943798, + 0.023756928741931915, + 0.032515183091163635, + 0.02275937981903553, + 0.023256422951817513, + 0.02373938448727131, + 0.027479002252221107, + 0.018992286175489426, + 0.026049384847283363, + 0.011625776067376137, + 0.00817195512354374, + 0.0111923823133111, + 0.008970338851213455, + 0.00905793346464634, + 0.010395945981144905, + 0.008805353194475174, + 0.010302500799298286, + 0.006364823319017887, + 0.00741160474717617, + 0.0034341791179031134, + 0.0031009812373667955, + 0.004110084380954504, + 0.0030043141450732946, + 0.003319100709632039, + 0.0035724379122257233, + 0.016992909833788872, + 0.021651683375239372, + 0.016403960064053535, + 0.01853347010910511, + 0.019990377128124237, + 0.01553184725344181, + 0.012602412141859531, + 0.018851056694984436, + 0.007230669725686312, + 0.005810328293591738, + 0.004840410780161619, + 0.005969636142253876, + 0.0077469227835536, + 0.007541541941463947, + 0.004057398531585932, + 0.0052238814532756805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.007632380351424217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024868296459317207 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024868296459317207, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024868296459317207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06177755445241928, + 0.045847829431295395, + 0.04668958857655525, + 0.051183439791202545, + 0.03905698284506798, + 0.030869074165821075, + 0.057270560413599014, + 0.1888725608587265, + 0.41237834095954895, + 0.06605406105518341 + ], + "credit_map": [ + 0.0076605468057096004, + 0.00727508869022131, + 0.006480400916188955, + 0.009584578685462475, + 0.006966530345380306, + 0.008258244954049587, + 0.009127724915742874, + 0.006424441933631897, + 0.0039925407618284225, + 0.00519853038713336, + 0.003724467707797885, + 0.007477427367120981, + 0.006841904018074274, + 0.004968458786606789, + 0.008718829602003098, + 0.0049256677739322186, + 0.004529151134192944, + 0.006151480600237846, + 0.004377730190753937, + 0.009426511824131012, + 0.004179058596491814, + 0.005939915310591459, + 0.0069259838201105595, + 0.0051597608253359795, + 0.006453890353441238, + 0.009189669042825699, + 0.005349273793399334, + 0.004969523288309574, + 0.008362251333892345, + 0.006497961934655905, + 0.005104216281324625, + 0.005256653763353825, + 0.004422435071319342, + 0.0038937239442020655, + 0.005500018130987883, + 0.00438513932749629, + 0.005328290164470673, + 0.004774341359734535, + 0.00537619786337018, + 0.005376836284995079, + 0.0043341051787137985, + 0.005732424557209015, + 0.003505645552650094, + 0.002887933049350977, + 0.005762359127402306, + 0.0034419912844896317, + 0.0021823616698384285, + 0.003022253280505538, + 0.008770249783992767, + 0.005245685111731291, + 0.006598382256925106, + 0.010434582829475403, + 0.007808383088558912, + 0.006050662137567997, + 0.00595049886032939, + 0.006412115413695574, + 0.033304210752248764, + 0.02463713474571705, + 0.017754418775439262, + 0.014555907808244228, + 0.03567199036478996, + 0.025998612865805626, + 0.015625229105353355, + 0.021325064823031425, + 0.05203309282660484, + 0.04605146497488022, + 0.03502017632126808, + 0.03365004435181618, + 0.07886207103729248, + 0.051717229187488556, + 0.04779369756579399, + 0.0672505795955658, + 0.007419947534799576, + 0.008084705099463463, + 0.007559835910797119, + 0.011568552814424038, + 0.013092368841171265, + 0.007321134675294161, + 0.005772781558334827, + 0.005234736017882824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0024868296459317207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606742851436138 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606742851436138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606742851436138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07238747924566269, + 0.05026231333613396, + 0.07496720552444458, + 0.07862956821918488, + 0.08024798333644867, + 0.07842296361923218, + 0.09316828846931458, + 0.15659533441066742, + 0.2778140604496002, + 0.03750487044453621 + ], + "credit_map": [ + 0.010778190568089485, + 0.009403345175087452, + 0.005540041718631983, + 0.00875614769756794, + 0.008225002326071262, + 0.009834980592131615, + 0.01211621705442667, + 0.007733553647994995, + 0.004695144947618246, + 0.005915121641010046, + 0.005207845941185951, + 0.007550334092229605, + 0.0064314743503928185, + 0.005499846301972866, + 0.009384139440953732, + 0.005578408017754555, + 0.006261149886995554, + 0.009828428737819195, + 0.0061826701276004314, + 0.018720056861639023, + 0.007415765896439552, + 0.009161457419395447, + 0.009322009049355984, + 0.00807566475123167, + 0.008375679142773151, + 0.012830793857574463, + 0.008504369296133518, + 0.007757152896374464, + 0.012853870168328285, + 0.009459752589464188, + 0.008126480504870415, + 0.010721470229327679, + 0.009030836634337902, + 0.0085761034861207, + 0.01326422207057476, + 0.009220224805176258, + 0.010592850856482983, + 0.01027866080403328, + 0.00900489091873169, + 0.010280192829668522, + 0.009025081992149353, + 0.01236716192215681, + 0.009006911888718605, + 0.009034888818860054, + 0.012333232909440994, + 0.009808793663978577, + 0.008531569503247738, + 0.008315318264067173, + 0.009053570218384266, + 0.010814064182341099, + 0.011740378104150295, + 0.011440051719546318, + 0.015033300034701824, + 0.015969526022672653, + 0.009941814467310905, + 0.009175583720207214, + 0.026214156299829483, + 0.023208118975162506, + 0.015229994431138039, + 0.012782673351466656, + 0.026381781324744225, + 0.01846911385655403, + 0.013690539635717869, + 0.020618954673409462, + 0.03976805880665779, + 0.036812543869018555, + 0.036082182079553604, + 0.055882666260004044, + 0.03819584846496582, + 0.024063391610980034, + 0.022688010707497597, + 0.024321338161826134, + 0.005066860467195511, + 0.004692193120718002, + 0.0034542735666036606, + 0.006952858995646238, + 0.0059982487000525, + 0.00409472594037652, + 0.0035849246196448803, + 0.0036607859656214714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.01606742851436138 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, centered on the one-eyed general, Xiahou Dun, clad in an intricate suit of mythical beast armor. The dark black iron is engraved with roaring kirin heads and intricate runes, and gilded patterns, scarred by battle, appear faintly in the dim light beneath a dark, overcast sky. His face contorts with the intensity of his exertion, his single eye blazing with fury. His muscles tense as he raises a heavy, long-handled halberd high above his head, then brings it down with all his might, crashing heavily upon the cracked, scorched earth. Framed by a maximalist composition, the moment of impact seems to drag the world into silky, high-frame slow motion: the ground instantly cracks, and countless rock fragments of varying sizes, interspersed with dazzling arcs of golden energy and lightning, radiate outward in a gesture of exquisite elegance and violence. The impact point becomes the core light source of the picture, illuminating every flying piece of gravel. In the extreme slow motion, you can even see the clear trajectory of each dust particle picked up by the shock wave rolling and spinning in the air. The whole scene is full of epic and destructive power, presenting a violent aesthetic solidified in time.", + "seed": 24063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_2859_seed_24063_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02485650437884033, + "vlm_logprob": -0.02485650437884033, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02485650437884033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04027489200234413 + ], + "opposite_logprob_tokens": [ + -3.2902748584747314 + ], + "logprob_reward": -0.04027489200234413, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04027489200234413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14013829827308655, + 0.15794605016708374, + 0.1290861815214157, + 0.13141417503356934, + 0.05961247533559799, + 0.05561951547861099, + 0.07146307826042175, + 0.027592642232775688, + 0.18604643642902374, + 0.04108119383454323 + ], + "credit_map": [ + 0.019277136772871017, + 0.009086580947041512, + 0.025108229368925095, + 0.01304377056658268, + 0.0201722364872694, + 0.014550924301147461, + 0.026902005076408386, + 0.011997408233582973, + 0.018739020451903343, + 0.010749652050435543, + 0.02662970870733261, + 0.0145133500918746, + 0.012747112661600113, + 0.030626041814684868, + 0.01803961955010891, + 0.025901537388563156, + 0.017389817163348198, + 0.00881990697234869, + 0.013616160489618778, + 0.026407882571220398, + 0.016888689249753952, + 0.0157113429158926, + 0.015094787813723087, + 0.015157588757574558, + 0.017903687432408333, + 0.014106967486441135, + 0.01599985919892788, + 0.011751322075724602, + 0.017535675317049026, + 0.016240887343883514, + 0.024155404418706894, + 0.013720367103815079, + 0.008481081575155258, + 0.00946077797561884, + 0.011701574549078941, + 0.0054255579598248005, + 0.006867624819278717, + 0.006603099871426821, + 0.005594020243734121, + 0.005478741135448217, + 0.008347040973603725, + 0.004123855847865343, + 0.005836570169776678, + 0.006673661060631275, + 0.004882465582340956, + 0.008380569517612457, + 0.006159691140055656, + 0.011215661652386189, + 0.01692325808107853, + 0.005198793951421976, + 0.017200756818056107, + 0.005201609339565039, + 0.006502420641481876, + 0.0063859946094453335, + 0.008014647290110588, + 0.006035591941326857, + 0.004240537527948618, + 0.0029114349745213985, + 0.003137751715257764, + 0.00304625416174531, + 0.002553420141339302, + 0.0029225037433207035, + 0.003889236832037568, + 0.004891503602266312, + 0.022401733323931694, + 0.011979340575635433, + 0.041396625339984894, + 0.013480665162205696, + 0.010644438676536083, + 0.011195720173418522, + 0.06341595202684402, + 0.011531970463693142, + 0.004716844297945499, + 0.004537048749625683, + 0.0060775442980229855, + 0.006276755128055811, + 0.004441292490810156, + 0.004489099606871605, + 0.005849531386047602, + 0.004693078342825174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.04027489200234413 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867305979132652 + ], + "opposite_logprob_tokens": [ + -3.048673152923584 + ], + "logprob_reward": -0.04867305979132652, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867305979132652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23117080330848694, + 0.11496566981077194, + 0.11719047278165817, + 0.10209671407938004, + 0.030180983245372772, + 0.03544127941131592, + 0.25478988885879517, + 0.03170120716094971, + 0.03547416627407074, + 0.04698885977268219 + ], + "credit_map": [ + 0.03405347466468811, + 0.012203493155539036, + 0.02476879209280014, + 0.02530723810195923, + 0.02293691784143448, + 0.020142873749136925, + 0.07126760482788086, + 0.02049041911959648, + 0.012802736833691597, + 0.008814687840640545, + 0.015026973560452461, + 0.020293783396482468, + 0.018444838002324104, + 0.011736449785530567, + 0.015665216371417046, + 0.012180982157588005, + 0.01762455515563488, + 0.007797976955771446, + 0.012834222055971622, + 0.017709065228700638, + 0.01734127849340439, + 0.01716032810509205, + 0.013443272560834885, + 0.013279777951538563, + 0.013838417828083038, + 0.011766293086111546, + 0.012981281615793705, + 0.008935149759054184, + 0.009753616526722908, + 0.015193036757409573, + 0.013440134935081005, + 0.016188783571124077, + 0.003283626865595579, + 0.00369251542724669, + 0.004812398925423622, + 0.004117728676646948, + 0.0048868549056351185, + 0.003470535157248378, + 0.003363951575011015, + 0.0025533714797347784, + 0.005802049767225981, + 0.0028056432493031025, + 0.0029114442877471447, + 0.005241926293820143, + 0.0036840131506323814, + 0.0037394636310636997, + 0.004883801564574242, + 0.006372938863933086, + 0.021657144650816917, + 0.006198879797011614, + 0.10382065922021866, + 0.00875434186309576, + 0.0103526646271348, + 0.006717778742313385, + 0.0882568508386612, + 0.00903157889842987, + 0.005348048638552427, + 0.003683586372062564, + 0.003450030228123069, + 0.0034337081015110016, + 0.0036452971398830414, + 0.0019514302257448435, + 0.007338914554566145, + 0.002850189572200179, + 0.006278198678046465, + 0.0020631703082472086, + 0.006222385913133621, + 0.0018086323980242014, + 0.002844291739165783, + 0.0028279204852879047, + 0.011167781427502632, + 0.0022617869544774294, + 0.004400509409606457, + 0.007376466412097216, + 0.005490201059728861, + 0.007719895802438259, + 0.007073927205055952, + 0.005187658593058586, + 0.004920922685414553, + 0.004819277208298445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.04867305979132652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076697036623955 + ], + "opposite_logprob_tokens": [ + -4.141076564788818 + ], + "logprob_reward": -0.016076697036623955, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076697036623955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06759776175022125, + 0.07305562496185303, + 0.14400774240493774, + 0.16792766749858856, + 0.07803678512573242, + 0.10615932196378708, + 0.2458374947309494, + 0.05157526955008507, + 0.03543201833963394, + 0.030370386317372322 + ], + "credit_map": [ + 0.008844741620123386, + 0.004782524425536394, + 0.009914407506585121, + 0.00850375834852457, + 0.005760125815868378, + 0.006851914804428816, + 0.012182696722447872, + 0.010757585056126118, + 0.010768117383122444, + 0.005378270987421274, + 0.014579126611351967, + 0.008113006129860878, + 0.008465593680739403, + 0.009756644256412983, + 0.009030800312757492, + 0.006964063737541437, + 0.018016494810581207, + 0.011313549242913723, + 0.01606891117990017, + 0.02196125127375126, + 0.019800789654254913, + 0.021882852539420128, + 0.017000989988446236, + 0.01796288974583149, + 0.020379381254315376, + 0.019911715760827065, + 0.020862335339188576, + 0.01729278452694416, + 0.026635272428393364, + 0.018474657088518143, + 0.025915464386343956, + 0.01845603622496128, + 0.0075484164990484715, + 0.008489656262099743, + 0.009539118967950344, + 0.011286863125860691, + 0.010670584626495838, + 0.008166569285094738, + 0.01262877881526947, + 0.009706798940896988, + 0.010149797424674034, + 0.012033908627927303, + 0.01109927985817194, + 0.013694189488887787, + 0.015073755756020546, + 0.010112200863659382, + 0.01698857545852661, + 0.017007606104016304, + 0.018069693818688393, + 0.00566412229090929, + 0.09938527643680573, + 0.009676117449998856, + 0.010868949815630913, + 0.005208189133554697, + 0.08862806856632233, + 0.008337065577507019, + 0.007673849817365408, + 0.004002024885267019, + 0.006616166792809963, + 0.005394259002059698, + 0.005820286925882101, + 0.004324032925069332, + 0.011634554713964462, + 0.006110088899731636, + 0.004557911306619644, + 0.0022126284893602133, + 0.008055293932557106, + 0.002084062434732914, + 0.0022536248434334993, + 0.002093447605147958, + 0.011952978558838367, + 0.002222071634605527, + 0.0029519707895815372, + 0.005110680591315031, + 0.0035362797789275646, + 0.004864188842475414, + 0.003844629740342498, + 0.0034644880797713995, + 0.0033267999533563852, + 0.0032713443506509066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.016076697036623955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06527427583932877, + 0.0956074669957161, + 0.05305936560034752, + 0.03567276895046234, + 0.07296111434698105, + 0.17401273548603058, + 0.33224037289619446, + 0.06002488732337952, + 0.03787932172417641, + 0.07326766103506088 + ], + "credit_map": [ + 0.008936027996242046, + 0.0033460042905062437, + 0.015220080502331257, + 0.006558939814567566, + 0.00651745218783617, + 0.0057585593312978745, + 0.012190657667815685, + 0.006746549159288406, + 0.018077058717608452, + 0.006333072204142809, + 0.01915145292878151, + 0.010471909306943417, + 0.011006324551999569, + 0.013935746625065804, + 0.00909047108143568, + 0.007541430648416281, + 0.007048096042126417, + 0.00533495144918561, + 0.005739330314099789, + 0.006213017273694277, + 0.008123280480504036, + 0.0064745317213237286, + 0.008418216370046139, + 0.0057079424150288105, + 0.0047222631983459, + 0.0040967813692986965, + 0.004377925768494606, + 0.0038385356310755014, + 0.004029793199151754, + 0.0038102269172668457, + 0.00516939302906394, + 0.005627847276628017, + 0.006563051138073206, + 0.005233985837548971, + 0.009080116637051105, + 0.012258419767022133, + 0.007992526516318321, + 0.008297677151858807, + 0.009855312295258045, + 0.013680021278560162, + 0.033424537628889084, + 0.014784224331378937, + 0.015279381535947323, + 0.018703561276197433, + 0.021595211699604988, + 0.03260134905576706, + 0.01871621236205101, + 0.01890825852751732, + 0.02873176336288452, + 0.011148743331432343, + 0.12660393118858337, + 0.0171674694865942, + 0.01737082377076149, + 0.01152186281979084, + 0.10741887986660004, + 0.012276927009224892, + 0.007336340378969908, + 0.006259853485971689, + 0.007748488336801529, + 0.006994411814957857, + 0.0061894552782177925, + 0.0054056704975664616, + 0.012918015010654926, + 0.0071726576425135136, + 0.005297074094414711, + 0.002912907861173153, + 0.007101146504282951, + 0.004843887407332659, + 0.003615398658439517, + 0.003713331650942564, + 0.00740221468731761, + 0.0029933589976280928, + 0.007722586393356323, + 0.012258726172149181, + 0.009522113017737865, + 0.011736118234694004, + 0.00928296335041523, + 0.007334033027291298, + 0.008240645751357079, + 0.007170481141656637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.0031914988067001104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066374257206917 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.016066374257206917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066374257206917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06875961273908615, + 0.075423963367939, + 0.04915044829249382, + 0.12615178525447845, + 0.04587782919406891, + 0.10756535828113556, + 0.08110249787569046, + 0.11031794548034668, + 0.24503573775291443, + 0.09061484783887863 + ], + "credit_map": [ + 0.01169893890619278, + 0.004069439601153135, + 0.01042949315160513, + 0.006673605181276798, + 0.0055227442644536495, + 0.005771458614617586, + 0.017861835658550262, + 0.006732096895575523, + 0.006806693039834499, + 0.005999096669256687, + 0.00999503955245018, + 0.009344893507659435, + 0.008920022286474705, + 0.007496373262256384, + 0.010710153728723526, + 0.016151688992977142, + 0.006128801964223385, + 0.005303202662616968, + 0.0052784704603254795, + 0.006347175687551498, + 0.006058683153241873, + 0.007228209637105465, + 0.006236889399588108, + 0.006569014862179756, + 0.017851661890745163, + 0.01653994619846344, + 0.014879790134727955, + 0.013530385680496693, + 0.01677943393588066, + 0.015320083126425743, + 0.018371528014540672, + 0.012878944166004658, + 0.006372981704771519, + 0.004686466883867979, + 0.006625584792345762, + 0.0051393285393714905, + 0.006031378637999296, + 0.0051889135502278805, + 0.00802887324243784, + 0.003804299980401993, + 0.019553568214178085, + 0.009860830381512642, + 0.009796206839382648, + 0.011245330795645714, + 0.013216857798397541, + 0.019431699067354202, + 0.01065786462277174, + 0.013802990317344666, + 0.012126811780035496, + 0.007445996627211571, + 0.013124365359544754, + 0.006306617986410856, + 0.008817694149911404, + 0.007821979001164436, + 0.018146302551031113, + 0.007312726229429245, + 0.013684907928109169, + 0.009893772192299366, + 0.01560982409864664, + 0.010583641938865185, + 0.010777007788419724, + 0.009895099326968193, + 0.02468486689031124, + 0.015188818797469139, + 0.0334017388522625, + 0.013900045305490494, + 0.052192818373441696, + 0.013815580867230892, + 0.01593577116727829, + 0.017652081325650215, + 0.08231520652770996, + 0.015822485089302063, + 0.01016192976385355, + 0.016843263059854507, + 0.016019416972994804, + 0.013521866872906685, + 0.010402719490230083, + 0.008389703929424286, + 0.008755556307733059, + 0.006520387716591358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.016066374257206917 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, centered on the one-eyed general, Xiahou Dun, clad in an intricate suit of mythical beast armor. The dark black iron is engraved with roaring kirin heads and intricate runes, and gilded patterns, scarred by battle, appear faintly in the dim light beneath a dark, overcast sky. His face contorts with the intensity of his exertion, his single eye blazing with fury. His muscles tense as he raises a heavy, long-handled halberd high above his head, then brings it down with all his might, crashing heavily upon the cracked, scorched earth. Framed by a maximalist composition, the moment of impact seems to drag the world into silky, high-frame slow motion: the ground instantly cracks, and countless rock fragments of varying sizes, interspersed with dazzling arcs of golden energy and lightning, radiate outward in a gesture of exquisite elegance and violence. The impact point becomes the core light source of the picture, illuminating every flying piece of gravel. In the extreme slow motion, you can even see the clear trajectory of each dust particle picked up by the shock wave rolling and spinning in the air. The whole scene is full of epic and destructive power, presenting a violent aesthetic solidified in time.", + "seed": 24064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_2859_seed_24064_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019905232917517425, + "vlm_logprob": -0.019905232917517425, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.019905232917517425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04552606865763664 + ], + "opposite_logprob_tokens": [ + -3.1705260276794434 + ], + "logprob_reward": -0.04552606865763664, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04552606865763664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23417901992797852, + 0.084590844810009, + 0.0682506412267685, + 0.07524047791957855, + 0.07667959481477737, + 0.04378734901547432, + 0.1568322479724884, + 0.10598273575305939, + 0.10680139064788818, + 0.047655750066041946 + ], + "credit_map": [ + 0.031239531934261322, + 0.03528626263141632, + 0.02865249663591385, + 0.02967159077525139, + 0.02304285205900669, + 0.02662288397550583, + 0.0340454988181591, + 0.02561788633465767, + 0.00916645023971796, + 0.009611754678189754, + 0.012131381779909134, + 0.01298960018903017, + 0.007295576389878988, + 0.012911868281662464, + 0.011770045384764671, + 0.008714165538549423, + 0.011776812374591827, + 0.010041489265859127, + 0.009334689937531948, + 0.008715645410120487, + 0.0065031275153160095, + 0.009717495180666447, + 0.006881960202008486, + 0.005279419478029013, + 0.0067856088280677795, + 0.007000313606113195, + 0.01546372752636671, + 0.004767048638314009, + 0.00735650397837162, + 0.014894481748342514, + 0.005884531419724226, + 0.01308825146406889, + 0.007440208923071623, + 0.003834857139736414, + 0.02392392046749592, + 0.003482603933662176, + 0.0040068128146231174, + 0.0041484287939965725, + 0.004543859977275133, + 0.025298895314335823, + 0.007137783337384462, + 0.003965767566114664, + 0.006440042518079281, + 0.004870386328548193, + 0.0055768368765711784, + 0.005297050811350346, + 0.006101734470576048, + 0.004397745709866285, + 0.01604289375245571, + 0.014249100349843502, + 0.04272133484482765, + 0.008852648548781872, + 0.0066591412760317326, + 0.04433126747608185, + 0.011953308247029781, + 0.012022539973258972, + 0.010672091506421566, + 0.006854712497442961, + 0.008726462721824646, + 0.008321375586092472, + 0.010139155201613903, + 0.006390185095369816, + 0.025642137974500656, + 0.02923661284148693, + 0.00849881861358881, + 0.0270423274487257, + 0.00646918173879385, + 0.0053502400405704975, + 0.012781682424247265, + 0.007400136440992355, + 0.02933460846543312, + 0.009924385696649551, + 0.005506379995495081, + 0.005686269607394934, + 0.004808609373867512, + 0.004167366772890091, + 0.007967027835547924, + 0.004960982128977776, + 0.005571609362959862, + 0.008987503126263618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.04552606865763664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113366812467575 + ], + "opposite_logprob_tokens": [ + -3.2881133556365967 + ], + "logprob_reward": -0.038113366812467575, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113366812467575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08709241449832916, + 0.11648707091808319, + 0.04690492898225784, + 0.0920581966638565, + 0.10010445863008499, + 0.1991589367389679, + 0.14898306131362915, + 0.057911161333322525, + 0.09525062888860703, + 0.056049179285764694 + ], + "credit_map": [ + 0.013668608851730824, + 0.012473477981984615, + 0.013336661271750927, + 0.011682075448334217, + 0.006384619511663914, + 0.010329220443964005, + 0.008684899657964706, + 0.010532845743000507, + 0.020048944279551506, + 0.018909335136413574, + 0.020301876589655876, + 0.010618484579026699, + 0.010679363273084164, + 0.019870927557349205, + 0.006775580812245607, + 0.009282561019062996, + 0.007688990328460932, + 0.0062889279797673225, + 0.00606621615588665, + 0.005133836530148983, + 0.005578864365816116, + 0.006919828709214926, + 0.005478932987898588, + 0.003749331459403038, + 0.010562901385128498, + 0.006537330336868763, + 0.014966520480811596, + 0.009713433682918549, + 0.008601252920925617, + 0.018354661762714386, + 0.009274790994822979, + 0.01404730323702097, + 0.010723460465669632, + 0.0060879201628267765, + 0.028441429138183594, + 0.006254204083234072, + 0.006296351086348295, + 0.006547164637595415, + 0.008625011891126633, + 0.027128908783197403, + 0.03372877091169357, + 0.01787244714796543, + 0.03574294596910477, + 0.021437156945466995, + 0.023757118731737137, + 0.02281469851732254, + 0.02422229014337063, + 0.019583512097597122, + 0.016753658652305603, + 0.012339537963271141, + 0.03983648866415024, + 0.007063195575028658, + 0.007113355677574873, + 0.04438494145870209, + 0.009918742813169956, + 0.01157313771545887, + 0.006856740452349186, + 0.005460517480969429, + 0.005772639065980911, + 0.005813669413328171, + 0.007363953627645969, + 0.005529487505555153, + 0.009380428120493889, + 0.011733725666999817, + 0.007559388410300016, + 0.02500752918422222, + 0.005249526351690292, + 0.008635853417217731, + 0.008813617751002312, + 0.007991787046194077, + 0.023572703823447227, + 0.00842022243887186, + 0.0069313873536884785, + 0.006687378976494074, + 0.0053415484726428986, + 0.004076255019754171, + 0.00852466281503439, + 0.004806176759302616, + 0.00690871849656105, + 0.012773050926625729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.038113366812467575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036194308195263147 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036194308195263147, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036194308195263147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07121483236551285, + 0.09923388808965683, + 0.046026989817619324, + 0.09740370512008667, + 0.05077998712658882, + 0.06337681412696838, + 0.0403357595205307, + 0.306656152009964, + 0.18760696053504944, + 0.037364885210990906 + ], + "credit_map": [ + 0.014749675057828426, + 0.00972586777061224, + 0.011676453985273838, + 0.007195676676928997, + 0.005203863140195608, + 0.007107021752744913, + 0.007944845594465733, + 0.007611424662172794, + 0.017116336151957512, + 0.014414787292480469, + 0.016661763191223145, + 0.01106975320726633, + 0.008171920664608479, + 0.013501581735908985, + 0.009381956420838833, + 0.00891579408198595, + 0.007694598287343979, + 0.005938158370554447, + 0.00541582889854908, + 0.004554312210530043, + 0.003624281845986843, + 0.005958812776952982, + 0.004193906206637621, + 0.008647091686725616, + 0.011184467002749443, + 0.010826927609741688, + 0.015368611551821232, + 0.00759779242798686, + 0.010275565087795258, + 0.02197977714240551, + 0.007374495733529329, + 0.012796072289347649, + 0.005176112055778503, + 0.0034614112228155136, + 0.013560212217271328, + 0.00275063025765121, + 0.00420733168721199, + 0.00377061078324914, + 0.004680664744228125, + 0.013173014856874943, + 0.01184271089732647, + 0.00525300856679678, + 0.012163275852799416, + 0.004273995757102966, + 0.00773773156106472, + 0.007572853472083807, + 0.006756913848221302, + 0.007776323240250349, + 0.0045828805305063725, + 0.005835458170622587, + 0.004837371874600649, + 0.003574033034965396, + 0.0044434694573283195, + 0.008895555511116982, + 0.004195766989141703, + 0.003971221391111612, + 0.03699168190360069, + 0.022998595610260963, + 0.02948717027902603, + 0.02901596389710903, + 0.04817325994372368, + 0.027607165277004242, + 0.05847996100783348, + 0.05390235409140587, + 0.01300413254648447, + 0.02517554722726345, + 0.016304386779665947, + 0.01539328321814537, + 0.044965505599975586, + 0.020146602764725685, + 0.023809509351849556, + 0.028808001428842545, + 0.003855828894302249, + 0.004784358199685812, + 0.003279465716332197, + 0.004466784652322531, + 0.007048325147479773, + 0.004633727949112654, + 0.005083718337118626, + 0.004212676547467709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.0036194308195263147 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248754327185452 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.00248754327185452, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248754327185452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09455171972513199, + 0.04480252042412758, + 0.040500178933143616, + 0.04385413974523544, + 0.03214598447084427, + 0.04601123929023743, + 0.06920594722032547, + 0.13665413856506348, + 0.340385764837265, + 0.1518884152173996 + ], + "credit_map": [ + 0.014914237894117832, + 0.010400768369436264, + 0.012743623927235603, + 0.009321111254394054, + 0.009231830947101116, + 0.009155026637017727, + 0.017960647121071815, + 0.010824467986822128, + 0.004818718880414963, + 0.0061183227226138115, + 0.006581013556569815, + 0.00481778709217906, + 0.005962040275335312, + 0.005379531532526016, + 0.005744477733969688, + 0.005380625836551189, + 0.006795488763600588, + 0.007049327250570059, + 0.0057406798005104065, + 0.0051115937530994415, + 0.003547839354723692, + 0.005934882443398237, + 0.0033343627583235502, + 0.0029860029462724924, + 0.005700009874999523, + 0.003170998068526387, + 0.006011112127453089, + 0.004202918149530888, + 0.004844903480261564, + 0.009145792573690414, + 0.004042541608214378, + 0.006735864095389843, + 0.003851187415421009, + 0.002106696367263794, + 0.009044644422829151, + 0.0020840915385633707, + 0.002115514362230897, + 0.0021782228723168373, + 0.0025840348098427057, + 0.008181596174836159, + 0.005671330727636814, + 0.004810138139873743, + 0.005174602847546339, + 0.006094265263527632, + 0.007796471007168293, + 0.005210804753005505, + 0.006945184897631407, + 0.004308442119508982, + 0.0064256452023983, + 0.009063753299415112, + 0.01789463683962822, + 0.004780798684805632, + 0.006662230473011732, + 0.010629381984472275, + 0.006853814236819744, + 0.006895683240145445, + 0.014845350757241249, + 0.017581630498170853, + 0.01429307647049427, + 0.01542609091848135, + 0.013853671960532665, + 0.012115429155528545, + 0.02254817821085453, + 0.025990713387727737, + 0.02249778062105179, + 0.1020660400390625, + 0.018193457275629044, + 0.02423139102756977, + 0.028701677918434143, + 0.0259323101490736, + 0.09129174798727036, + 0.027471337467432022, + 0.01929743029177189, + 0.01724790222942829, + 0.01691538468003273, + 0.025549177080392838, + 0.021678447723388672, + 0.013225053437054157, + 0.015836581587791443, + 0.022138433530926704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.00248754327185452 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05842333287000656, + 0.053114209324121475, + 0.04633038863539696, + 0.06798826158046722, + 0.04609733074903488, + 0.11083652079105377, + 0.2766393721103668, + 0.07549832761287689, + 0.18796201050281525, + 0.07711022347211838 + ], + "credit_map": [ + 0.012423369102180004, + 0.0062177362851798534, + 0.00883557740598917, + 0.006435283925384283, + 0.004764605779200792, + 0.005662194453179836, + 0.008124652318656445, + 0.0059599122032523155, + 0.006806718185544014, + 0.007916438393294811, + 0.009685268625617027, + 0.004492305219173431, + 0.005001517944037914, + 0.009068240411579609, + 0.007004330866038799, + 0.003139388980343938, + 0.007508115842938423, + 0.008774146437644958, + 0.006961863022297621, + 0.004870718810707331, + 0.0035455478355288506, + 0.007307982072234154, + 0.003577043768018484, + 0.0037849736399948597, + 0.009819393046200275, + 0.005133323837071657, + 0.013115132227540016, + 0.004382103215903044, + 0.00659198546782136, + 0.01337352767586708, + 0.004652897361665964, + 0.010919902473688126, + 0.0046087936498224735, + 0.0023182551376521587, + 0.014132475480437279, + 0.0017546077724546194, + 0.002264059614390135, + 0.002663152292370796, + 0.0023867511190474033, + 0.015969233587384224, + 0.021580228582024574, + 0.007952081970870495, + 0.02540246769785881, + 0.011836518533527851, + 0.008776692673563957, + 0.013336352072656155, + 0.011426051147282124, + 0.010526123456656933, + 0.02775624766945839, + 0.02402825653553009, + 0.079068124294281, + 0.0145435631275177, + 0.013819157145917416, + 0.07474666088819504, + 0.019621316343545914, + 0.02305603213608265, + 0.00882104691118002, + 0.007688494399189949, + 0.006752023007720709, + 0.010014551691710949, + 0.012544765137135983, + 0.008100614883005619, + 0.00940354261547327, + 0.012173288501799107, + 0.010957326740026474, + 0.048140883445739746, + 0.00836431048810482, + 0.009066728875041008, + 0.01910434663295746, + 0.012473809532821178, + 0.05718240514397621, + 0.022672200575470924, + 0.00806696992367506, + 0.008252491243183613, + 0.008364731445908546, + 0.005562649108469486, + 0.007745273411273956, + 0.008712532930076122, + 0.01094307191669941, + 0.019462503492832184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.009779755026102066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, centered on the one-eyed general, Xiahou Dun, clad in an intricate suit of mythical beast armor. The dark black iron is engraved with roaring kirin heads and intricate runes, and gilded patterns, scarred by battle, appear faintly in the dim light beneath a dark, overcast sky. His face contorts with the intensity of his exertion, his single eye blazing with fury. His muscles tense as he raises a heavy, long-handled halberd high above his head, then brings it down with all his might, crashing heavily upon the cracked, scorched earth. Framed by a maximalist composition, the moment of impact seems to drag the world into silky, high-frame slow motion: the ground instantly cracks, and countless rock fragments of varying sizes, interspersed with dazzling arcs of golden energy and lightning, radiate outward in a gesture of exquisite elegance and violence. The impact point becomes the core light source of the picture, illuminating every flying piece of gravel. In the extreme slow motion, you can even see the clear trajectory of each dust particle picked up by the shock wave rolling and spinning in the air. The whole scene is full of epic and destructive power, presenting a violent aesthetic solidified in time.", + "seed": 24065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_2/index_2859_seed_24065_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021527032274752854, + "vlm_logprob": -0.021527032274752854, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.021527032274752854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03541260585188866 + ], + "opposite_logprob_tokens": [ + -3.410412549972534 + ], + "logprob_reward": -0.03541260585188866, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03541260585188866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1517166644334793, + 0.046888887882232666, + 0.06847389787435532, + 0.37025585770606995, + 0.05730430781841278, + 0.03967231512069702, + 0.025338411331176758, + 0.06453554332256317, + 0.050280358642339706, + 0.12553372979164124 + ], + "credit_map": [ + 0.009458119980990887, + 0.013568340800702572, + 0.016512548550963402, + 0.012648222036659718, + 0.0351877324283123, + 0.031130071729421616, + 0.00996840838342905, + 0.02324322611093521, + 0.004277616739273071, + 0.006977709010243416, + 0.007069853134453297, + 0.007755655329674482, + 0.00658457912504673, + 0.004378852900117636, + 0.004651317838579416, + 0.0051933047361671925, + 0.005807212088257074, + 0.008107871748507023, + 0.01148996502161026, + 0.011125142686069012, + 0.008273669518530369, + 0.007790280971676111, + 0.008302787318825722, + 0.007576974108815193, + 0.024070080369710922, + 0.02918226644396782, + 0.049778182059526443, + 0.022414252161979675, + 0.06873729079961777, + 0.029936134815216064, + 0.12940646708011627, + 0.01673118770122528, + 0.006644034292548895, + 0.006916233338415623, + 0.008603827096521854, + 0.007318804506212473, + 0.008091245777904987, + 0.008403874933719635, + 0.004944412969052792, + 0.006381879560649395, + 0.0033115707337856293, + 0.005695420783013105, + 0.003530023619532585, + 0.0045980424620211124, + 0.0068601807579398155, + 0.005395588465034962, + 0.0055582113564014435, + 0.004723280668258667, + 0.002808462828397751, + 0.0033353210892528296, + 0.002800766145810485, + 0.0037139244377613068, + 0.0027220253832638264, + 0.003022626042366028, + 0.003385895863175392, + 0.003549390472471714, + 0.00787082314491272, + 0.010971239767968655, + 0.00794320646673441, + 0.011311059817671776, + 0.007032766006886959, + 0.006645433604717255, + 0.0067975400015711784, + 0.005963483359664679, + 0.005749908741563559, + 0.00593860587105155, + 0.006288126111030579, + 0.006853730883449316, + 0.005058536771684885, + 0.005047557409852743, + 0.007297866512089968, + 0.008046030066907406, + 0.007059041876345873, + 0.011563711799681187, + 0.003173051867634058, + 0.0032141096889972687, + 0.04956889525055885, + 0.009553712792694569, + 0.003998666536062956, + 0.03740254044532776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.03541260585188866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812874108552933 + ], + "opposite_logprob_tokens": [ + -3.2881288528442383 + ], + "logprob_reward": -0.03812874108552933, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812874108552933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05803627893328667, + 0.015861455351114273, + 0.02440900169312954, + 0.798367440700531, + 0.013747374527156353, + 0.01843077689409256, + 0.006585333030670881, + 0.011941513046622276, + 0.024886835366487503, + 0.027733949944376945 + ], + "credit_map": [ + 0.003870943794026971, + 0.004035660065710545, + 0.00941970944404602, + 0.004932338371872902, + 0.012788284569978714, + 0.009765874594449997, + 0.003811718663200736, + 0.009411751292645931, + 0.0018000276759266853, + 0.0016746623441576958, + 0.001984060974791646, + 0.00195482699200511, + 0.0019242948619648814, + 0.002564031630754471, + 0.001786125940270722, + 0.0021734251640737057, + 0.0014317012391984463, + 0.0034929285757243633, + 0.003867372404783964, + 0.005055567249655724, + 0.0029647978954017162, + 0.0017982561839744449, + 0.0029327445663511753, + 0.002865634858608246, + 0.05170895531773567, + 0.0689486414194107, + 0.10562016814947128, + 0.04590031877160072, + 0.15430507063865662, + 0.05750530585646629, + 0.2770329713821411, + 0.0373460054397583, + 0.0012754221679642797, + 0.0011533340439200401, + 0.0016472948482260108, + 0.0018627471290528774, + 0.0012355968356132507, + 0.0021145017817616463, + 0.0016145722474902868, + 0.0028439059387892485, + 0.002112681744620204, + 0.002073409268632531, + 0.001720355125144124, + 0.0020780714694410563, + 0.0026176527608186007, + 0.0016194316558539867, + 0.0025190175510942936, + 0.0036901580169796944, + 0.0007601168472319841, + 0.0008603624301031232, + 0.000787809316534549, + 0.00072187784826383, + 0.0006716905045323074, + 0.0009757145307958126, + 0.0008576413383707404, + 0.0009501202148385346, + 0.0010291535872966051, + 0.0019044061191380024, + 0.0018461496802046895, + 0.001067682052962482, + 0.002155480906367302, + 0.0010243430733680725, + 0.0014488081214949489, + 0.001465490204282105, + 0.0018665833631530404, + 0.0034525198861956596, + 0.0020570815540850163, + 0.002397877862676978, + 0.0031118448823690414, + 0.002723730867728591, + 0.0032242373563349247, + 0.006052959710359573, + 0.0014509725151583552, + 0.002415269147604704, + 0.0009315175702795386, + 0.0008063954883255064, + 0.010568453930318356, + 0.0020995894446969032, + 0.0011027121217921376, + 0.008359041064977646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.03812874108552933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259605124592781 + ], + "opposite_logprob_tokens": [ + -5.2552595138549805 + ], + "logprob_reward": -0.005259605124592781, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259605124592781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1392025500535965, + 0.035386569797992706, + 0.027515364810824394, + 0.48402658104896545, + 0.0720524713397026, + 0.07028691470623016, + 0.04660486802458763, + 0.0444365032017231, + 0.025859029963612556, + 0.05462918430566788 + ], + "credit_map": [ + 0.013242285698652267, + 0.010574066080152988, + 0.029446924105286598, + 0.010841233655810356, + 0.02537020668387413, + 0.019320335239171982, + 0.01051345281302929, + 0.019894050434231758, + 0.0018828620668500662, + 0.0053474195301532745, + 0.005210719536989927, + 0.0060793436132371426, + 0.003568714950233698, + 0.0029914947226643562, + 0.004693156573921442, + 0.005612855311483145, + 0.002185489283874631, + 0.0028939321637153625, + 0.003886876394972205, + 0.004919788800179958, + 0.002841337351128459, + 0.003646201454102993, + 0.003812540555372834, + 0.003329197410494089, + 0.03220558911561966, + 0.041888341307640076, + 0.06474652141332626, + 0.028111128136515617, + 0.09247400611639023, + 0.036362964659929276, + 0.16604217886924744, + 0.02219587191939354, + 0.0073369164019823074, + 0.009400418028235435, + 0.007722952868789434, + 0.0078077372163534164, + 0.013773570768535137, + 0.010168559849262238, + 0.0065949708223342896, + 0.009247342124581337, + 0.006792000960558653, + 0.009201610460877419, + 0.006796976551413536, + 0.009617399424314499, + 0.007454797625541687, + 0.009336378425359726, + 0.008919351734220982, + 0.0121683981269598, + 0.004377100616693497, + 0.005687698721885681, + 0.006092175841331482, + 0.004923628643155098, + 0.0055364263243973255, + 0.005924656521528959, + 0.007030007429420948, + 0.007033171132206917, + 0.0028870217502117157, + 0.010967789217829704, + 0.005148734897375107, + 0.002611638978123665, + 0.013558757491409779, + 0.0032379759941250086, + 0.0025950768031179905, + 0.003429506439715624, + 0.002769398968666792, + 0.003520931350067258, + 0.002886381698772311, + 0.0029502296820282936, + 0.003185197478160262, + 0.002985985018312931, + 0.003132956102490425, + 0.0044279503636062145, + 0.0031737766694277525, + 0.0047285244800150394, + 0.0020938466768711805, + 0.0015719188377261162, + 0.020064957439899445, + 0.0040102070197463036, + 0.002736234338954091, + 0.01624971814453602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.005259605124592781 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1054871678352356, + 0.028080083429813385, + 0.021160125732421875, + 0.3507242798805237, + 0.06676420569419861, + 0.19279073178768158, + 0.03492806851863861, + 0.053034912794828415, + 0.06474299728870392, + 0.08228746056556702 + ], + "credit_map": [ + 0.007986965589225292, + 0.0074251568876206875, + 0.011524470522999763, + 0.008178400807082653, + 0.028010742738842964, + 0.016962623223662376, + 0.00954638421535492, + 0.015852419659495354, + 0.002081353450194001, + 0.004382465500384569, + 0.0037960025947541, + 0.004171547945588827, + 0.00361939356662333, + 0.0027938776183873415, + 0.0030494669917970896, + 0.0041859750635921955, + 0.0019604647532105446, + 0.001995051046833396, + 0.0025837095454335213, + 0.002735262271016836, + 0.0033168138470500708, + 0.002383254701271653, + 0.0028662492986768484, + 0.0033193216659128666, + 0.024964714422822, + 0.029661325737833977, + 0.04714730754494667, + 0.020198604092001915, + 0.06691784411668777, + 0.0288112573325634, + 0.11647192388772964, + 0.016551310196518898, + 0.007872284390032291, + 0.005583255551755428, + 0.007395959924906492, + 0.006607590708881617, + 0.006547647528350353, + 0.009930064901709557, + 0.007934688590466976, + 0.014892714098095894, + 0.01931927166879177, + 0.021875061094760895, + 0.01934593729674816, + 0.02520938776433468, + 0.025585683062672615, + 0.021415313705801964, + 0.024385051801800728, + 0.03565504029393196, + 0.003241667291149497, + 0.0039182729087769985, + 0.0059683239087462425, + 0.0050979843363165855, + 0.00359406229108572, + 0.0040332842618227005, + 0.004706055391579866, + 0.0043684206902980804, + 0.006077715661376715, + 0.010653750970959663, + 0.006694643292576075, + 0.005116427782922983, + 0.007942789234220982, + 0.004631482530385256, + 0.004850346595048904, + 0.007067758124321699, + 0.005057559814304113, + 0.006054223515093327, + 0.006666939239948988, + 0.010153939947485924, + 0.00835911650210619, + 0.006628256291151047, + 0.009963332675397396, + 0.011859625577926636, + 0.004908743314445019, + 0.0071761892177164555, + 0.0023465435951948166, + 0.0017673957627266645, + 0.03233729302883148, + 0.006574682891368866, + 0.0029692065436393023, + 0.02420741133391857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.0024877809919416904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from an extremely low angle centered on a one-eyed general?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing dark black iron armor engraved with roaring kirin heads and gilded patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character raise a heavy long-handled halberd and bring it down to crash upon the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the ground cracking with rock fragments, golden energy arcs, and lightning radiating outward upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video presented in silky high-frame slow motion with the impact point serving as the core light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634642831981182 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.02634642831981182, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634642831981182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18453556299209595, + 0.05094241350889206, + 0.03845079243183136, + 0.1587747484445572, + 0.14699864387512207, + 0.061475321650505066, + 0.09997442364692688, + 0.06936772167682648, + 0.10573557019233704, + 0.08374480903148651 + ], + "credit_map": [ + 0.009725203737616539, + 0.009714662097394466, + 0.01953231543302536, + 0.010130706243216991, + 0.05964989587664604, + 0.03215688094496727, + 0.01359257660806179, + 0.030033325776457787, + 0.0066043720580637455, + 0.005419648718088865, + 0.0074416897259652615, + 0.00526708597317338, + 0.006799167022109032, + 0.005752130877226591, + 0.006076311692595482, + 0.007582005579024553, + 0.002872890792787075, + 0.0032949780579656363, + 0.005102760624140501, + 0.007076646666973829, + 0.004440709948539734, + 0.005432679317891598, + 0.00547375762835145, + 0.00475636962801218, + 0.011214196681976318, + 0.013758133165538311, + 0.021475352346897125, + 0.010068395175039768, + 0.029922740533947945, + 0.012038100510835648, + 0.05311160162091255, + 0.007186234928667545, + 0.013494784012436867, + 0.014176450669765472, + 0.015427197329699993, + 0.015270712785422802, + 0.01580488309264183, + 0.023128187283873558, + 0.017845869064331055, + 0.03185056149959564, + 0.005340284202247858, + 0.006414300296455622, + 0.006351854186505079, + 0.006868224591016769, + 0.010901597328484058, + 0.008074904792010784, + 0.008044606074690819, + 0.009479548782110214, + 0.012853116728365421, + 0.013987897895276546, + 0.012732790783047676, + 0.008702494204044342, + 0.011068546213209629, + 0.013295194134116173, + 0.011488578282296658, + 0.015845805406570435, + 0.00770439812913537, + 0.014354304410517216, + 0.01009753905236721, + 0.0050697000697255135, + 0.009565829299390316, + 0.008383156731724739, + 0.005156269297003746, + 0.009036523289978504, + 0.008405838161706924, + 0.011314022354781628, + 0.008800199255347252, + 0.01132749393582344, + 0.015073738060891628, + 0.012398733757436275, + 0.016242695972323418, + 0.022172844037413597, + 0.0057446653954684734, + 0.00853799469769001, + 0.0034096462186425924, + 0.0029716668650507927, + 0.030034372583031654, + 0.007186168804764748, + 0.004038707353174686, + 0.021821584552526474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.02634642831981182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_tensor([575, 575])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_tensor([575, 575])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cb6641858d5e5c558c6bf068cf59d233506c46b7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_tensor([575, 575])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in dramatic slow motion, the film begins with a medium shot tracking a vintage road bike plummeting from the edge of a high, sun-bleached wooden pier. The camera tilts downward, smoothly following the bike's graceful arc through the air. The bike itself, with a weathered burgundy frame and handlebars wrapped in worn brown leather, appears in sharp detail against the vast, deep blue ocean below. Its thin wheels spin rapidly and continuously, the chrome spokes blurring into a hypnotic, shimmering circle that captures the brilliant glow of the low, golden hour sun. As it descends, the moment of impact is both beautiful and violent; the bike pierces the water's surface, creating a silent, explosive plume of crystal-white spray and a surging cloud of jagged bubbles. The entire scene is bathed in the warm, long-lasting light of late afternoon, casting dramatic shadows and illuminating the depths of the water, creating a poignant and visually stunning atmosphere as the bike is rapidly engulfed, its spinning wheels gradually slowing as it sinks into the tranquil, dark abyss.", + "seed": 19062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_575_seed_19062_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018060348159633578, + "vlm_logprob": -0.018060348159633578, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.018060348159633578, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029217522591352463 + ], + "opposite_logprob_tokens": [ + -3.779217481613159 + ], + "logprob_reward": -0.029217522591352463, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.029217522591352463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16591954231262207, + 0.0460343100130558, + 0.04038725048303604, + 0.12622623145580292, + 0.044256750494241714, + 0.18887244164943695, + 0.036703065037727356, + 0.1539425253868103, + 0.14254477620124817, + 0.05511310696601868 + ], + "credit_map": [ + 0.017691336572170258, + 0.01684221439063549, + 0.011249597184360027, + 0.02221895195543766, + 0.024310894310474396, + 0.02001282386481762, + 0.0294862762093544, + 0.024107452481985092, + 0.008494469337165356, + 0.006610038224607706, + 0.003932497464120388, + 0.004568814765661955, + 0.009318121708929539, + 0.00490591861307621, + 0.004632662981748581, + 0.003571788314729929, + 0.004914073273539543, + 0.004685927648097277, + 0.004739725962281227, + 0.00661064125597477, + 0.006744572892785072, + 0.003887137398123741, + 0.00497694406658411, + 0.0038282338064163923, + 0.019773690029978752, + 0.009497405029833317, + 0.038424376398324966, + 0.010530615225434303, + 0.007672250736504793, + 0.006395419128239155, + 0.006991753354668617, + 0.026940731331706047, + 0.008363641798496246, + 0.005674948450177908, + 0.006660946644842625, + 0.006084095686674118, + 0.005378016736358404, + 0.0032199642155319452, + 0.006045965012162924, + 0.002829174976795912, + 0.04851575568318367, + 0.03123423457145691, + 0.030825231224298477, + 0.020361803472042084, + 0.01837806962430477, + 0.008245251141488552, + 0.02196984924376011, + 0.009342270903289318, + 0.0032517113722860813, + 0.008858534507453442, + 0.003992839716374874, + 0.004109859466552734, + 0.006054585333913565, + 0.003558965865522623, + 0.004468365106731653, + 0.0024082057643681765, + 0.017413778230547905, + 0.050592269748449326, + 0.0249143298715353, + 0.006183867808431387, + 0.016824111342430115, + 0.022460807114839554, + 0.007881910540163517, + 0.0076714614406228065, + 0.0294125284999609, + 0.009862535633146763, + 0.03753114491701126, + 0.008887074887752533, + 0.011070895940065384, + 0.023171154782176018, + 0.00862118974328041, + 0.013988271355628967, + 0.0038881548680365086, + 0.003841749858111143, + 0.007225606124848127, + 0.01954643987119198, + 0.003200156381353736, + 0.0033736093901097775, + 0.005094279535114765, + 0.008943120948970318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.029217522591352463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054961517453193665 + ], + "opposite_logprob_tokens": [ + -2.9299614429473877 + ], + "logprob_reward": -0.054961517453193665, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054961517453193665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19810041785240173, + 0.0566384457051754, + 0.047409530729055405, + 0.12176869809627533, + 0.055420346558094025, + 0.058213476091623306, + 0.07805085927248001, + 0.14610543847084045, + 0.19389300048351288, + 0.044399768114089966 + ], + "credit_map": [ + 0.024640528485178947, + 0.020399227738380432, + 0.017812788486480713, + 0.029371367767453194, + 0.028711633756756783, + 0.020073380321264267, + 0.0323253870010376, + 0.02476610615849495, + 0.009669571183621883, + 0.009530787356197834, + 0.007864674553275108, + 0.005230787210166454, + 0.009760454297065735, + 0.004681686405092478, + 0.0056751263327896595, + 0.004225356969982386, + 0.007841265760362148, + 0.005192874930799007, + 0.006191493943333626, + 0.006141087971627712, + 0.006213762331753969, + 0.004608347546309233, + 0.0056972382590174675, + 0.005523459054529667, + 0.01852514035999775, + 0.006809920538216829, + 0.04069935902953148, + 0.008107896894216537, + 0.006224262528121471, + 0.009277616627514362, + 0.006133854389190674, + 0.025990642607212067, + 0.01052572950720787, + 0.0062549179419875145, + 0.008783227764070034, + 0.00653320224955678, + 0.008763778954744339, + 0.004468780476599932, + 0.0074297175742685795, + 0.0026609920896589756, + 0.011198380030691624, + 0.010566513054072857, + 0.007129215169698, + 0.00675552012398839, + 0.007512879557907581, + 0.0036803078837692738, + 0.008468213491141796, + 0.00290244584903121, + 0.007981386967003345, + 0.010681042447686195, + 0.008663155138492584, + 0.009171009995043278, + 0.011748408898711205, + 0.009002081118524075, + 0.014432350173592567, + 0.006371431518346071, + 0.015815244987607002, + 0.05322226881980896, + 0.02545158937573433, + 0.006394055671989918, + 0.015051600523293018, + 0.019416335970163345, + 0.005345626268535852, + 0.005408720578998327, + 0.03720937669277191, + 0.015246567316353321, + 0.0487804114818573, + 0.014337446540594101, + 0.014819726347923279, + 0.03097360208630562, + 0.01263166218996048, + 0.019894208759069443, + 0.006727587431669235, + 0.0029617322143167257, + 0.006744297221302986, + 0.011400705203413963, + 0.003530818270519376, + 0.005136056337505579, + 0.003282097401097417, + 0.0046164728701114655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.054961517453193665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005651307292282581 + ], + "opposite_logprob_tokens": [ + -7.500565052032471 + ], + "logprob_reward": -0.0005651307292282581, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005651307292282581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1590501070022583, + 0.06191612035036087, + 0.03283276781439781, + 0.13245484232902527, + 0.06155665963888168, + 0.04995155707001686, + 0.08982682228088379, + 0.2593362331390381, + 0.10660023987293243, + 0.04647458344697952 + ], + "credit_map": [ + 0.018108194693922997, + 0.017046203836798668, + 0.011568770743906498, + 0.021406976506114006, + 0.024951912462711334, + 0.0165463387966156, + 0.02825346775352955, + 0.021168243139982224, + 0.0098081910982728, + 0.006960932165384293, + 0.005975126754492521, + 0.006826871540397406, + 0.013214209116995335, + 0.006310130935162306, + 0.006679551675915718, + 0.006141103804111481, + 0.0040961094200611115, + 0.003575007664039731, + 0.004719491116702557, + 0.005320469383150339, + 0.004104970954358578, + 0.004232222680002451, + 0.0035595204681158066, + 0.003224979154765606, + 0.020273257046937943, + 0.006088441237807274, + 0.0452704094350338, + 0.00879313051700592, + 0.006171694025397301, + 0.008171373046934605, + 0.005924740340560675, + 0.031761810183525085, + 0.0158613920211792, + 0.007854555733501911, + 0.008894070982933044, + 0.007487152237445116, + 0.0070207552053034306, + 0.0049150967970490456, + 0.00548538938164711, + 0.0040382505394518375, + 0.010239092633128166, + 0.007236072327941656, + 0.00588886346668005, + 0.005968855228275061, + 0.007058588322252035, + 0.0036294322926551104, + 0.006968418601900339, + 0.0029622360598295927, + 0.007884378544986248, + 0.013976933434605598, + 0.010667640715837479, + 0.010882850736379623, + 0.013337157666683197, + 0.00923909991979599, + 0.01641550473868847, + 0.0074232532642781734, + 0.028178377076983452, + 0.09365575760602951, + 0.044180624186992645, + 0.010081845335662365, + 0.026821477338671684, + 0.0357222780585289, + 0.0108500886708498, + 0.009845783933997154, + 0.02283545583486557, + 0.00902100931853056, + 0.02797429822385311, + 0.007668049540370703, + 0.007788899354636669, + 0.015555255115032196, + 0.006255615968257189, + 0.00950165931135416, + 0.004010207951068878, + 0.0031529362313449383, + 0.004380721598863602, + 0.010853178799152374, + 0.0032315331045538187, + 0.00456559332087636, + 0.003246450563892722, + 0.013033963739871979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0005651307292282581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615630092099309 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.003615630092099309, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615630092099309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12079713493585587, + 0.022151634097099304, + 0.04021188244223595, + 0.18060123920440674, + 0.030848396942019463, + 0.09091748297214508, + 0.04852130264043808, + 0.3301118016242981, + 0.09978711605072021, + 0.03605194762349129 + ], + "credit_map": [ + 0.01450478844344616, + 0.013371123000979424, + 0.01014237105846405, + 0.016261503100395203, + 0.015685800462961197, + 0.015530143864452839, + 0.020526278764009476, + 0.014775141142308712, + 0.0035858098417520523, + 0.0032116167712956667, + 0.0031249618623405695, + 0.0021480428986251354, + 0.0034255548380315304, + 0.0027847762685269117, + 0.0020094714127480984, + 0.0018614019500091672, + 0.006755378562957048, + 0.003766902256757021, + 0.004922621883451939, + 0.004356723744422197, + 0.004586866591125727, + 0.005346604157239199, + 0.005293196067214012, + 0.00518359150737524, + 0.027938880026340485, + 0.010207224637269974, + 0.06196108087897301, + 0.012262902222573757, + 0.008157989010214806, + 0.008291767910122871, + 0.008627239614725113, + 0.043154165148735046, + 0.005993088241666555, + 0.0034083682112395763, + 0.0054647428914904594, + 0.0035450714640319347, + 0.003389215562492609, + 0.0031168763525784016, + 0.0029719939921051264, + 0.002959043253213167, + 0.025677630677819252, + 0.019862446933984756, + 0.009811094030737877, + 0.007764432579278946, + 0.009274261072278023, + 0.005948303267359734, + 0.007235910277813673, + 0.00534340925514698, + 0.008284913375973701, + 0.0069616371765732765, + 0.005136548075824976, + 0.005271595437079668, + 0.006433804519474506, + 0.007167582865804434, + 0.005549347493797541, + 0.003715881612151861, + 0.036233123391866684, + 0.1183227002620697, + 0.05854122340679169, + 0.012920673005282879, + 0.03427970036864281, + 0.045758143067359924, + 0.012269578874111176, + 0.011786703951656818, + 0.025058912113308907, + 0.007263505831360817, + 0.025733234360814095, + 0.010449176654219627, + 0.006526305340230465, + 0.009160645306110382, + 0.006517056375741959, + 0.009078290313482285, + 0.005773169454187155, + 0.00406961003318429, + 0.004282193258404732, + 0.0035780465696007013, + 0.0037857007700949907, + 0.005864636041224003, + 0.004672950599342585, + 0.004025645088404417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.003615630092099309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00194193993229419 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.00194193993229419, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00194193993229419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.086444191634655, + 0.045008014887571335, + 0.0308453980833292, + 0.03138508275151253, + 0.026790373027324677, + 0.09832553565502167, + 0.043840449303388596, + 0.2880637049674988, + 0.14775674045085907, + 0.2015405148267746 + ], + "credit_map": [ + 0.009876132942736149, + 0.00879835244268179, + 0.007861452177166939, + 0.01355000026524067, + 0.011678453534841537, + 0.01201718207448721, + 0.011913295835256577, + 0.01074931863695383, + 0.007608611602336168, + 0.006534480955451727, + 0.004567761905491352, + 0.004834617488086224, + 0.008579032495617867, + 0.004291124641895294, + 0.004809826146811247, + 0.003782555926591158, + 0.004663535859435797, + 0.0031566820107400417, + 0.0034254821948707104, + 0.004118155222386122, + 0.0037491319235414267, + 0.004077788908034563, + 0.0039269146509468555, + 0.0037277061492204666, + 0.0027389125898480415, + 0.0037546753883361816, + 0.004713263362646103, + 0.005030790343880653, + 0.004585522226989269, + 0.0029285969212651253, + 0.003359441179782152, + 0.004273878876119852, + 0.005272273905575275, + 0.0029076198115944862, + 0.0026867485139518976, + 0.0036735516041517258, + 0.004785431548953056, + 0.002436308888718486, + 0.003066684352234006, + 0.001961755333468318, + 0.015573997050523758, + 0.01867309957742691, + 0.018738871440291405, + 0.010772339068353176, + 0.010696467943489552, + 0.006571596954017878, + 0.011572621762752533, + 0.0057265362702310085, + 0.00579466437920928, + 0.0071817804127931595, + 0.004914392717182636, + 0.0045476555824279785, + 0.007986541837453842, + 0.004680461250245571, + 0.00565127469599247, + 0.003083675168454647, + 0.035671036690473557, + 0.09505682438611984, + 0.04326808080077171, + 0.01332276314496994, + 0.031002279371023178, + 0.040878213942050934, + 0.015359921380877495, + 0.013504565693438053, + 0.019691694527864456, + 0.016213342547416687, + 0.019184783101081848, + 0.012526915408670902, + 0.015957599505782127, + 0.02236233465373516, + 0.011457729153335094, + 0.030362343415617943, + 0.028735369443893433, + 0.027415873482823372, + 0.02407512441277504, + 0.024429744109511375, + 0.025468062609434128, + 0.02193942666053772, + 0.02804638259112835, + 0.02143050730228424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.00194193993229419 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, the film begins with a medium shot tracking a vintage road bike plummeting from the edge of a high, sun-bleached wooden pier. The camera tilts downward, smoothly following the bike's graceful arc through the air. The bike itself, with a weathered burgundy frame and handlebars wrapped in worn brown leather, appears in sharp detail against the vast, deep blue ocean below. Its thin wheels spin rapidly and continuously, the chrome spokes blurring into a hypnotic, shimmering circle that captures the brilliant glow of the low, golden hour sun. As it descends, the moment of impact is both beautiful and violent; the bike pierces the water's surface, creating a silent, explosive plume of crystal-white spray and a surging cloud of jagged bubbles. The entire scene is bathed in the warm, long-lasting light of late afternoon, casting dramatic shadows and illuminating the depths of the water, creating a poignant and visually stunning atmosphere as the bike is rapidly engulfed, its spinning wheels gradually slowing as it sinks into the tranquil, dark abyss.", + "seed": 19063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_575_seed_19063_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014390303095569833, + "vlm_logprob": -0.014390303095569833, + "vlm_yes_no_margin": 5.6, + "vlm_token_credit": -0.014390303095569833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030344784259796143 + ], + "opposite_logprob_tokens": [ + -3.9053447246551514 + ], + "logprob_reward": -0.030344784259796143, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.030344784259796143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2229463756084442, + 0.09537051618099213, + 0.14298272132873535, + 0.16081193089485168, + 0.07563459128141403, + 0.06031939014792442, + 0.08464417606592178, + 0.051668766885995865, + 0.053889110684394836, + 0.051732394844293594 + ], + "credit_map": [ + 0.02671695500612259, + 0.025575950741767883, + 0.020250050351023674, + 0.029082082211971283, + 0.048768699169158936, + 0.025038879364728928, + 0.019407199695706367, + 0.0281065683811903, + 0.008534437045454979, + 0.009889801032841206, + 0.010465837083756924, + 0.013507881201803684, + 0.016757357865571976, + 0.008462730795145035, + 0.017873838543891907, + 0.009878640994429588, + 0.009724912233650684, + 0.027532080188393593, + 0.011175688356161118, + 0.01623491197824478, + 0.029993383213877678, + 0.01219400018453598, + 0.020861836150288582, + 0.015265928581357002, + 0.018880344927310944, + 0.021918578073382378, + 0.03391287848353386, + 0.01336970180273056, + 0.021510036662220955, + 0.012750196270644665, + 0.02381790615618229, + 0.014652292244136333, + 0.00855215359479189, + 0.010977260768413544, + 0.0050368658266961575, + 0.008107597008347511, + 0.008004278875887394, + 0.010928797535598278, + 0.006346542853862047, + 0.017681103199720383, + 0.008130915462970734, + 0.00668281689286232, + 0.009516254998743534, + 0.009047750383615494, + 0.005972079001367092, + 0.007233140524476767, + 0.00868817139416933, + 0.005048265215009451, + 0.013159317895770073, + 0.011255476623773575, + 0.015040533617138863, + 0.00788920372724533, + 0.008870474994182587, + 0.008627410046756268, + 0.01183828804641962, + 0.007963478565216064, + 0.006918962579220533, + 0.0046837348490953445, + 0.007360256277024746, + 0.004784769378602505, + 0.006161919329315424, + 0.008723458275198936, + 0.006686847656965256, + 0.00634881854057312, + 0.006706769112497568, + 0.004968001041561365, + 0.010809694416821003, + 0.00534010911360383, + 0.003794450080022216, + 0.005708654411137104, + 0.005127368029206991, + 0.011434069834649563, + 0.0050461264327168465, + 0.003467875998467207, + 0.013206073082983494, + 0.006505423691123724, + 0.004063933156430721, + 0.003916820045560598, + 0.004148757550865412, + 0.01137738861143589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.030344784259796143 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809569403529167 + ], + "opposite_logprob_tokens": [ + -3.288095712661743 + ], + "logprob_reward": -0.03809569403529167, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809569403529167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06272683292627335, + 0.1338769793510437, + 0.07759615778923035, + 0.2944624125957489, + 0.10611385852098465, + 0.05476093292236328, + 0.12147517502307892, + 0.07443426549434662, + 0.04153355956077576, + 0.03301982209086418 + ], + "credit_map": [ + 0.007537339348345995, + 0.0077186753042042255, + 0.005401150789111853, + 0.0075893886387348175, + 0.00937473215162754, + 0.008965957909822464, + 0.006722854916006327, + 0.009416735731065273, + 0.016331924125552177, + 0.019919514656066895, + 0.016398748382925987, + 0.01422728318721056, + 0.021597355604171753, + 0.012105211615562439, + 0.020767083391547203, + 0.012529848143458366, + 0.0058286115527153015, + 0.013443892821669579, + 0.006930693052709103, + 0.009687217883765697, + 0.015554266050457954, + 0.009454814717173576, + 0.009584460407495499, + 0.007112203631550074, + 0.03477361425757408, + 0.04526837542653084, + 0.043201711028814316, + 0.028891004621982574, + 0.04787994176149368, + 0.02768367901444435, + 0.03775739669799805, + 0.029006682336330414, + 0.01176205463707447, + 0.01015604566782713, + 0.005655374377965927, + 0.00906738918274641, + 0.012382005341351032, + 0.014317601919174194, + 0.007389253005385399, + 0.03538413345813751, + 0.0059992591850459576, + 0.004637031350284815, + 0.007058434654027224, + 0.004297136794775724, + 0.00839045736938715, + 0.006770544219762087, + 0.011972655542194843, + 0.005635411012917757, + 0.017906494438648224, + 0.012362364679574966, + 0.017508968710899353, + 0.014280034229159355, + 0.012757958844304085, + 0.016762513667345047, + 0.017483780160546303, + 0.012413065880537033, + 0.006279568187892437, + 0.00538074504584074, + 0.011969397775828838, + 0.012085078284144402, + 0.007999861612915993, + 0.0101063447073102, + 0.010744021274149418, + 0.009869251400232315, + 0.0054717231541872025, + 0.004935435950756073, + 0.006828607060015202, + 0.00619665440171957, + 0.004304978996515274, + 0.0049304417334496975, + 0.004727640654891729, + 0.004138079006224871, + 0.0030776069033890963, + 0.0020196144469082355, + 0.006928425282239914, + 0.004924546927213669, + 0.002150813350453973, + 0.003337462665513158, + 0.002791673643514514, + 0.007789678405970335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.03809569403529167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00039057256071828306 + ], + "opposite_logprob_tokens": [ + -7.875390529632568 + ], + "logprob_reward": -0.00039057256071828306, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.00039057256071828306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14966703951358795, + 0.3119315505027771, + 0.09525645524263382, + 0.09813519567251205, + 0.036204010248184204, + 0.0429246686398983, + 0.033634211868047714, + 0.1293535828590393, + 0.06226269155740738, + 0.040630605071783066 + ], + "credit_map": [ + 0.0170355886220932, + 0.016050653532147408, + 0.018055202439427376, + 0.02591680735349655, + 0.023318558931350708, + 0.017414750531315804, + 0.01322749350219965, + 0.018647976219654083, + 0.03309151902794838, + 0.04573126137256622, + 0.05237806960940361, + 0.036320220679044724, + 0.04835484176874161, + 0.03739086911082268, + 0.03214133903384209, + 0.026523403823375702, + 0.008482062257826328, + 0.015327542088925838, + 0.013713980093598366, + 0.015871474519371986, + 0.010265715420246124, + 0.008764682337641716, + 0.012923364527523518, + 0.009907630272209644, + 0.00936135370284319, + 0.01587740331888199, + 0.012970698066055775, + 0.01060723140835762, + 0.014137283898890018, + 0.009386333636939526, + 0.010600857436656952, + 0.015194021165370941, + 0.004520945250988007, + 0.004467567894607782, + 0.002795676700770855, + 0.003333892673254013, + 0.0046644313260912895, + 0.005445415154099464, + 0.003957604523748159, + 0.007018475327640772, + 0.006581468507647514, + 0.00343194673769176, + 0.006737827323377132, + 0.0033574625849723816, + 0.006619164254516363, + 0.006209594197571278, + 0.0059784031473100185, + 0.004008798860013485, + 0.005300421733409166, + 0.0032080342061817646, + 0.005013732239603996, + 0.0032634814269840717, + 0.0038019516505301, + 0.004955249838531017, + 0.0040321811102330685, + 0.004059159196913242, + 0.007679997943341732, + 0.007183239329606295, + 0.015190193429589272, + 0.008212761022150517, + 0.008851702325046062, + 0.01701468788087368, + 0.0331115759909153, + 0.03210941329598427, + 0.007578653749078512, + 0.005864269100129604, + 0.0131286745890975, + 0.00619103480130434, + 0.0038465079851448536, + 0.00635370472446084, + 0.006219400092959404, + 0.01308043859899044, + 0.004123265389353037, + 0.0039659482426941395, + 0.008458927273750305, + 0.005884006153792143, + 0.0032831069547683, + 0.003033997491002083, + 0.0033389313612133265, + 0.008542417548596859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.00039057256071828306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019403931219130754 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.0019403931219130754, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019403931219130754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.078548863530159, + 0.24472765624523163, + 0.08942487835884094, + 0.15291137993335724, + 0.09846830368041992, + 0.0948038101196289, + 0.09310206770896912, + 0.0852494016289711, + 0.03217800334095955, + 0.03058556839823723 + ], + "credit_map": [ + 0.008601696230471134, + 0.008201578631997108, + 0.008770186454057693, + 0.012325436808168888, + 0.00840809103101492, + 0.010466783307492733, + 0.007096040062606335, + 0.014679059386253357, + 0.026709722355008125, + 0.03918788582086563, + 0.0366920530796051, + 0.026938576251268387, + 0.04045792669057846, + 0.03015659935772419, + 0.02377859316766262, + 0.020806310698390007, + 0.009416360408067703, + 0.01732615754008293, + 0.012724165804684162, + 0.011635543778538704, + 0.011010956950485706, + 0.010289306752383709, + 0.010347919538617134, + 0.006674467120319605, + 0.020379822701215744, + 0.023513391613960266, + 0.02595171146094799, + 0.015616148710250854, + 0.015225346200168133, + 0.014983151108026505, + 0.024393809959292412, + 0.012848000042140484, + 0.012175070121884346, + 0.013124632649123669, + 0.0068872361443936825, + 0.010432845912873745, + 0.009774817153811455, + 0.01147060189396143, + 0.007677221205085516, + 0.02692587673664093, + 0.008830990642309189, + 0.007355719339102507, + 0.010402483865618706, + 0.008748918771743774, + 0.012634345330297947, + 0.012917473912239075, + 0.022143881767988205, + 0.011769995093345642, + 0.01434591505676508, + 0.012017919681966305, + 0.014941618777811527, + 0.009883865714073181, + 0.011312953196465969, + 0.009523551911115646, + 0.013883145526051521, + 0.007193101104348898, + 0.007768561597913504, + 0.007088773883879185, + 0.009477553889155388, + 0.007332716137170792, + 0.008729606866836548, + 0.008843427523970604, + 0.018994877114892006, + 0.01701388880610466, + 0.004369176458567381, + 0.002906723413616419, + 0.005089551210403442, + 0.006155978422611952, + 0.0022649322636425495, + 0.0034141866490244865, + 0.004503130447119474, + 0.003474324708804488, + 0.003921329975128174, + 0.003780425526201725, + 0.003684865776449442, + 0.004995285999029875, + 0.0029878285713493824, + 0.003720513079315424, + 0.003848671680316329, + 0.0036466491874307394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.0019403931219130754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011800715001299977 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011800715001299977, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011800715001299977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07078240066766739, + 0.10343721508979797, + 0.04227501153945923, + 0.1294369399547577, + 0.05485266074538231, + 0.06808678060770035, + 0.11689291894435883, + 0.16779734194278717, + 0.07455549389123917, + 0.17188327014446259 + ], + "credit_map": [ + 0.009219151921570301, + 0.007465459406375885, + 0.007530412636697292, + 0.010705463588237762, + 0.009317729622125626, + 0.00797423254698515, + 0.007867412641644478, + 0.010702534578740597, + 0.011018813587725163, + 0.012707388959825039, + 0.013995650224387646, + 0.014069757424294949, + 0.012810549698770046, + 0.013298975303769112, + 0.014966906979680061, + 0.010569158010184765, + 0.003858216106891632, + 0.007091539911925793, + 0.005167132243514061, + 0.006829578895121813, + 0.006631376687437296, + 0.003717204323038459, + 0.005624750163406134, + 0.0033552078530192375, + 0.01664596237242222, + 0.018067743629217148, + 0.01880980283021927, + 0.012253250926733017, + 0.02036202885210514, + 0.013220843859016895, + 0.01474401168525219, + 0.01533327903598547, + 0.0059046619571745396, + 0.007679296191781759, + 0.00541026471182704, + 0.008173316717147827, + 0.006716697011142969, + 0.005273919552564621, + 0.00692956056445837, + 0.008764940313994884, + 0.00860629603266716, + 0.00528419716283679, + 0.006590685807168484, + 0.0065772635862231255, + 0.010849829763174057, + 0.010590740479528904, + 0.011695259250700474, + 0.007892502471804619, + 0.01926727034151554, + 0.011240043677389622, + 0.025019139051437378, + 0.01076252106577158, + 0.010053240694105625, + 0.012935013510286808, + 0.01749533973634243, + 0.010120339691638947, + 0.016755830496549606, + 0.019783496856689453, + 0.024837493896484375, + 0.01442027185112238, + 0.02216489426791668, + 0.01708279550075531, + 0.027651505544781685, + 0.02510104887187481, + 0.011836770921945572, + 0.008711694739758968, + 0.012724199332296848, + 0.008389201946556568, + 0.007832896895706654, + 0.008137183263897896, + 0.009627092629671097, + 0.007296450436115265, + 0.01746346801519394, + 0.007531680166721344, + 0.054595425724983215, + 0.017548782750964165, + 0.005634639877825975, + 0.011775238439440727, + 0.011146130040287971, + 0.046187903732061386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.0011800715001299977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, the film begins with a medium shot tracking a vintage road bike plummeting from the edge of a high, sun-bleached wooden pier. The camera tilts downward, smoothly following the bike's graceful arc through the air. The bike itself, with a weathered burgundy frame and handlebars wrapped in worn brown leather, appears in sharp detail against the vast, deep blue ocean below. Its thin wheels spin rapidly and continuously, the chrome spokes blurring into a hypnotic, shimmering circle that captures the brilliant glow of the low, golden hour sun. As it descends, the moment of impact is both beautiful and violent; the bike pierces the water's surface, creating a silent, explosive plume of crystal-white spray and a surging cloud of jagged bubbles. The entire scene is bathed in the warm, long-lasting light of late afternoon, casting dramatic shadows and illuminating the depths of the water, creating a poignant and visually stunning atmosphere as the bike is rapidly engulfed, its spinning wheels gradually slowing as it sinks into the tranquil, dark abyss.", + "seed": 19064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_575_seed_19064_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02772031454369426, + "vlm_logprob": -0.02772031454369426, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.02772031454369426, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028543706983327866 + ], + "opposite_logprob_tokens": [ + -3.778543710708618 + ], + "logprob_reward": -0.028543706983327866, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028543706983327866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09013649076223373, + 0.043241605162620544, + 0.20944178104400635, + 0.19894646108150482, + 0.15484926104545593, + 0.1179465651512146, + 0.08097172528505325, + 0.039286594837903976, + 0.041778314858675, + 0.023401129990816116 + ], + "credit_map": [ + 0.01565966010093689, + 0.00964425690472126, + 0.010338191874325275, + 0.007327426224946976, + 0.0090020215138793, + 0.009731868281960487, + 0.01866316609084606, + 0.009769896976649761, + 0.003997838590294123, + 0.005061001051217318, + 0.005343726370483637, + 0.00568160368129611, + 0.0034673090558499098, + 0.00763712776824832, + 0.005240083672106266, + 0.006812917534261942, + 0.036617863923311234, + 0.022830946370959282, + 0.030085638165473938, + 0.019264813512563705, + 0.027301372960209846, + 0.03342360258102417, + 0.01772281713783741, + 0.022194745019078255, + 0.024558860808610916, + 0.01954684406518936, + 0.00978584960103035, + 0.02008102647960186, + 0.05290352925658226, + 0.02425755187869072, + 0.02727392502129078, + 0.020538892596960068, + 0.020579583942890167, + 0.02253113128244877, + 0.018091917037963867, + 0.014923886395990849, + 0.030190028250217438, + 0.019956381991505623, + 0.015119603835046291, + 0.013456733897328377, + 0.01865127868950367, + 0.014345256611704826, + 0.01728048361837864, + 0.010501817800104618, + 0.017440099269151688, + 0.011674201115965843, + 0.016135012730956078, + 0.011918424628674984, + 0.015874387696385384, + 0.008111433126032352, + 0.010961854830384254, + 0.0100207244977355, + 0.006428489461541176, + 0.010447917506098747, + 0.007549133617430925, + 0.01157779898494482, + 0.007054888177663088, + 0.006543201860040426, + 0.0025895144790410995, + 0.002984123770147562, + 0.0034726776648312807, + 0.005596814211457968, + 0.004097413271665573, + 0.006947966292500496, + 0.005791032686829567, + 0.0036007652524858713, + 0.005165283102542162, + 0.00414133770391345, + 0.005618698429316282, + 0.007356689311563969, + 0.005098558031022549, + 0.005005954764783382, + 0.003436089726164937, + 0.0027683391235768795, + 0.0026465137489140034, + 0.002334123244509101, + 0.002019783714786172, + 0.0031908659730106592, + 0.003375598229467869, + 0.003629815997555852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.028543706983327866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033027082681656 + ], + "opposite_logprob_tokens": [ + -2.350330352783203 + ], + "logprob_reward": -0.10033027082681656, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033027082681656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11942640691995621, + 0.09043198078870773, + 0.05530519038438797, + 0.09748843312263489, + 0.12523286044597626, + 0.19253098964691162, + 0.14116090536117554, + 0.09170698374509811, + 0.06430523097515106, + 0.022411029785871506 + ], + "credit_map": [ + 0.010010333731770515, + 0.008991912007331848, + 0.012919068336486816, + 0.009124388918280602, + 0.01726653426885605, + 0.016962729394435883, + 0.027592167258262634, + 0.01655927486717701, + 0.009347236715257168, + 0.01094901654869318, + 0.011664436198771, + 0.01270175538957119, + 0.00855670403689146, + 0.014678222127258778, + 0.011231089010834694, + 0.01130351796746254, + 0.007239664904773235, + 0.00762228574603796, + 0.006062473636120558, + 0.008662513457238674, + 0.0053490749560296535, + 0.010090135969221592, + 0.005137321539223194, + 0.005141721107065678, + 0.016089724376797676, + 0.010149597190320492, + 0.01180487871170044, + 0.012013345025479794, + 0.009971067309379578, + 0.012952066026628017, + 0.01112143974751234, + 0.013386308215558529, + 0.017815904691815376, + 0.025955049321055412, + 0.014626097865402699, + 0.010670584626495838, + 0.016607176512479782, + 0.014609056524932384, + 0.009399443864822388, + 0.015549548901617527, + 0.027112280949950218, + 0.020718589425086975, + 0.021511241793632507, + 0.016707949340343475, + 0.03575971722602844, + 0.022245068103075027, + 0.028634315356612206, + 0.019841840490698814, + 0.034049563109874725, + 0.01062738336622715, + 0.016578886657953262, + 0.008523042313754559, + 0.01309499517083168, + 0.03734731674194336, + 0.009193253703415394, + 0.011746455915272236, + 0.01527155376970768, + 0.013821390457451344, + 0.00752511341124773, + 0.00603234488517046, + 0.010428204201161861, + 0.014020235277712345, + 0.008474496193230152, + 0.016133639961481094, + 0.008821581490337849, + 0.0069032409228384495, + 0.006312001496553421, + 0.006123497616499662, + 0.008123192004859447, + 0.012006941251456738, + 0.007513047195971012, + 0.008501722477376461, + 0.0030881259590387344, + 0.0023235809057950974, + 0.0022051511332392693, + 0.002660108730196953, + 0.0023497603833675385, + 0.004272716119885445, + 0.002709997585043311, + 0.002801590831950307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.10033027082681656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036225190851837397 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.0036225190851837397, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036225190851837397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15748818218708038, + 0.10735222697257996, + 0.1275525987148285, + 0.07215788960456848, + 0.07216541469097137, + 0.09103997051715851, + 0.10082933306694031, + 0.15620197355747223, + 0.08045414835214615, + 0.034758251160383224 + ], + "credit_map": [ + 0.011910797096788883, + 0.010897482745349407, + 0.017370430752635002, + 0.012450721114873886, + 0.021970560774207115, + 0.02180853672325611, + 0.03967655077576637, + 0.0214031133800745, + 0.011230451054871082, + 0.011450770311057568, + 0.01253401581197977, + 0.010047238320112228, + 0.00986029114574194, + 0.022470921277999878, + 0.014667801558971405, + 0.01509073842316866, + 0.01904064603149891, + 0.013272098265588284, + 0.019187215715646744, + 0.013501894660294056, + 0.01397489383816719, + 0.022933702915906906, + 0.011808820068836212, + 0.013833323493599892, + 0.010098188184201717, + 0.007520666811615229, + 0.008845693431794643, + 0.010565442033112049, + 0.011544696986675262, + 0.0083211874589324, + 0.00644585071131587, + 0.008816161192953587, + 0.009339706040918827, + 0.014896051958203316, + 0.006543213035911322, + 0.006232951767742634, + 0.01046130433678627, + 0.008860350586473942, + 0.006206035614013672, + 0.009625799022614956, + 0.013780837878584862, + 0.011843719519674778, + 0.012535260058939457, + 0.008082223124802113, + 0.01491392683237791, + 0.010280587710440159, + 0.011269074864685535, + 0.008334340527653694, + 0.014280159026384354, + 0.008761637844145298, + 0.010690083727240562, + 0.010078616440296173, + 0.009708115831017494, + 0.022223196923732758, + 0.008698930032551289, + 0.016388602554798126, + 0.024752367287874222, + 0.027494002133607864, + 0.012666239403188229, + 0.011290145106613636, + 0.01652723178267479, + 0.024521712213754654, + 0.01372961513698101, + 0.025220675393939018, + 0.009475170634686947, + 0.008728626184165478, + 0.007390622515231371, + 0.006259937305003405, + 0.01568445935845375, + 0.016550227999687195, + 0.007299007847905159, + 0.00906609557569027, + 0.006848904304206371, + 0.003145105903968215, + 0.002542279427871108, + 0.003003806108608842, + 0.0029326616786420345, + 0.005689984653145075, + 0.005184899549931288, + 0.005410608369857073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0036225190851837397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12151743471622467, + 0.0930425301194191, + 0.11249089986085892, + 0.09978538006544113, + 0.06185587868094444, + 0.14149658381938934, + 0.14912457764148712, + 0.10291372984647751, + 0.06637350469827652, + 0.051399435847997665 + ], + "credit_map": [ + 0.01172309834510088, + 0.007723322603851557, + 0.013766755349934101, + 0.009563622996211052, + 0.011476930230855942, + 0.015485148876905441, + 0.039426833391189575, + 0.012351722456514835, + 0.009991330094635487, + 0.011182443238794804, + 0.01142409723252058, + 0.011117503978312016, + 0.00877407006919384, + 0.014716424979269505, + 0.012354310601949692, + 0.013482353650033474, + 0.01599302887916565, + 0.01177174225449562, + 0.014606372453272343, + 0.012265621684491634, + 0.01301186066120863, + 0.022757885977625847, + 0.009719387628138065, + 0.01236500684171915, + 0.008654545992612839, + 0.012678318656980991, + 0.007137852255254984, + 0.01990853250026703, + 0.01129985973238945, + 0.012147940695285797, + 0.009578770957887173, + 0.018379559740424156, + 0.009081203490495682, + 0.012689218856394291, + 0.009551341645419598, + 0.004687734413892031, + 0.007572081405669451, + 0.006243109703063965, + 0.005013844929635525, + 0.007017341908067465, + 0.026831459254026413, + 0.014312935061752796, + 0.016786549240350723, + 0.012461744248867035, + 0.022013453766703606, + 0.01588573306798935, + 0.02013414166867733, + 0.013070564717054367, + 0.023558907210826874, + 0.014556391164660454, + 0.013730589300394058, + 0.013202658854424953, + 0.014586552046239376, + 0.03932543098926544, + 0.01101568341255188, + 0.019148364663124084, + 0.01648217998445034, + 0.02102610468864441, + 0.007589494809508324, + 0.0066721695475280285, + 0.010003105737268925, + 0.01869083195924759, + 0.009105676785111427, + 0.013344171456992626, + 0.009181459434330463, + 0.007702005561441183, + 0.004354502074420452, + 0.0056220488622784615, + 0.010361714288592339, + 0.013135463930666447, + 0.006901825312525034, + 0.009114480577409267, + 0.008885142393410206, + 0.0045044613070786, + 0.005024624057114124, + 0.004536361899226904, + 0.0037041266914457083, + 0.010483212769031525, + 0.006483693141490221, + 0.007777816615998745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.003615511115640402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024895647075027227 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024895647075027227, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024895647075027227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0607319250702858, + 0.04481615126132965, + 0.07093554735183716, + 0.07251276820898056, + 0.05814385041594505, + 0.07104364037513733, + 0.08405846357345581, + 0.23593184351921082, + 0.12907737493515015, + 0.17274843156337738 + ], + "credit_map": [ + 0.007389330305159092, + 0.005546064116060734, + 0.00600670650601387, + 0.005140629131346941, + 0.007294953800737858, + 0.007534173782914877, + 0.012664045207202435, + 0.00915602594614029, + 0.004427169915288687, + 0.005187546834349632, + 0.005485572852194309, + 0.004849044140428305, + 0.004615838639438152, + 0.008952632546424866, + 0.00556600745767355, + 0.0057323407381772995, + 0.009034238755702972, + 0.009037197567522526, + 0.009188481606543064, + 0.008432213217020035, + 0.008418991230428219, + 0.01258629560470581, + 0.006633085198700428, + 0.0076050483621656895, + 0.006172665860503912, + 0.007310524582862854, + 0.006264419294893742, + 0.011523076333105564, + 0.009809949435293674, + 0.009276614524424076, + 0.007619349751621485, + 0.014536171220242977, + 0.008499580435454845, + 0.010545465163886547, + 0.005063390824943781, + 0.006655470933765173, + 0.008931837044656277, + 0.007295561488717794, + 0.005094779189676046, + 0.006057766731828451, + 0.012966262176632881, + 0.010608726181089878, + 0.00770281907171011, + 0.008327123709022999, + 0.008861973881721497, + 0.0077033089473843575, + 0.006908664014190435, + 0.007964772172272205, + 0.013298697769641876, + 0.007431555073708296, + 0.010230899788439274, + 0.006339252460747957, + 0.0080288490280509, + 0.01605837419629097, + 0.007929190061986446, + 0.014741647057235241, + 0.03654685989022255, + 0.03676975890994072, + 0.01907898299396038, + 0.01897161640226841, + 0.037471164017915726, + 0.02863459847867489, + 0.02730754390358925, + 0.03115132451057434, + 0.014875764958560467, + 0.018665773794054985, + 0.01039037760347128, + 0.010553309693932533, + 0.028047364205121994, + 0.021488351747393608, + 0.011253537610173225, + 0.0138029083609581, + 0.01891210488975048, + 0.020615966990590096, + 0.017846588045358658, + 0.026594551280140877, + 0.019304443150758743, + 0.022046951577067375, + 0.024273056536912918, + 0.023154769092798233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.0024895647075027227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, the film begins with a medium shot tracking a vintage road bike plummeting from the edge of a high, sun-bleached wooden pier. The camera tilts downward, smoothly following the bike's graceful arc through the air. The bike itself, with a weathered burgundy frame and handlebars wrapped in worn brown leather, appears in sharp detail against the vast, deep blue ocean below. Its thin wheels spin rapidly and continuously, the chrome spokes blurring into a hypnotic, shimmering circle that captures the brilliant glow of the low, golden hour sun. As it descends, the moment of impact is both beautiful and violent; the bike pierces the water's surface, creating a silent, explosive plume of crystal-white spray and a surging cloud of jagged bubbles. The entire scene is bathed in the warm, long-lasting light of late afternoon, casting dramatic shadows and illuminating the depths of the water, creating a poignant and visually stunning atmosphere as the bike is rapidly engulfed, its spinning wheels gradually slowing as it sinks into the tranquil, dark abyss.", + "seed": 19065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_3/index_575_seed_19065_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01533011922147125, + "vlm_logprob": -0.01533011922147125, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.01533011922147125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04566297307610512 + ], + "opposite_logprob_tokens": [ + -3.2956628799438477 + ], + "logprob_reward": -0.04566297307610512, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04566297307610512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4880666136741638, + 0.035245347768068314, + 0.0811222568154335, + 0.07079990953207016, + 0.0854310616850853, + 0.03670874238014221, + 0.09385798126459122, + 0.025841105729341507, + 0.04470733925700188, + 0.03821960836648941 + ], + "credit_map": [ + 0.07662986218929291, + 0.06787654012441635, + 0.06491128355264664, + 0.044281404465436935, + 0.067997045814991, + 0.08349470794200897, + 0.03624963387846947, + 0.04662615433335304, + 0.003895912552252412, + 0.00768781453371048, + 0.004618952516466379, + 0.00296150054782629, + 0.004720095079392195, + 0.0055580963380634785, + 0.0029903496615588665, + 0.002812630496919155, + 0.009528661146759987, + 0.013974713161587715, + 0.004647063557058573, + 0.008377219550311565, + 0.009024056605994701, + 0.011319042183458805, + 0.010991589166224003, + 0.013259915634989738, + 0.006264998111873865, + 0.009589586406946182, + 0.008007577620446682, + 0.011768135242164135, + 0.00837740208953619, + 0.010127732530236244, + 0.007609897293150425, + 0.009054585359990597, + 0.007171685341745615, + 0.010059974156320095, + 0.010274973697960377, + 0.01756567507982254, + 0.012717682868242264, + 0.012713522650301456, + 0.00866173766553402, + 0.006265808362513781, + 0.006159801036119461, + 0.004873317666351795, + 0.004979798104614019, + 0.005662566050887108, + 0.003127859905362129, + 0.0036626006476581097, + 0.005087570287287235, + 0.0031552324071526527, + 0.00972276832908392, + 0.009249567054212093, + 0.013961931690573692, + 0.008267735131084919, + 0.011137062683701515, + 0.013256281614303589, + 0.011913642287254333, + 0.016349000856280327, + 0.0028156989719718695, + 0.002652063500136137, + 0.0030606205109506845, + 0.0043443795293569565, + 0.0032715427223592997, + 0.003423308255150914, + 0.002811765531077981, + 0.00346172833815217, + 0.005983773153275251, + 0.010655193589627743, + 0.0033883326686918736, + 0.004855626728385687, + 0.005424045957624912, + 0.005335613153874874, + 0.004013137426227331, + 0.0050516207702457905, + 0.005913008935749531, + 0.004345918074250221, + 0.00632860092446208, + 0.004155934322625399, + 0.005243402440100908, + 0.0048599522560834885, + 0.00352904899045825, + 0.0038437459152191877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.04566297307610512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350490748882294 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350490748882294, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350490748882294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11042649298906326, + 0.09079976379871368, + 0.0514032281935215, + 0.051255907863378525, + 0.1344706267118454, + 0.09461250901222229, + 0.06022552028298378, + 0.11253048479557037, + 0.1385979801416397, + 0.15567749738693237 + ], + "credit_map": [ + 0.017456520348787308, + 0.01352584920823574, + 0.010847761295735836, + 0.01105014979839325, + 0.015606370754539967, + 0.01775198243558407, + 0.009926666505634785, + 0.014261186122894287, + 0.007174669299274683, + 0.01852664165198803, + 0.0053458246402442455, + 0.013587811961770058, + 0.013367599807679653, + 0.013056420721113682, + 0.007680206559598446, + 0.012060574255883694, + 0.005370401311665773, + 0.00795864686369896, + 0.0047266422770917416, + 0.004559598863124847, + 0.00984759721904993, + 0.0057637388817965984, + 0.0052075134590268135, + 0.00796908512711525, + 0.006573596969246864, + 0.0051964279264211655, + 0.0049318294040858746, + 0.006446404382586479, + 0.007191311102360487, + 0.009455250576138496, + 0.006351805292069912, + 0.005109275691211224, + 0.016375306993722916, + 0.014668548479676247, + 0.020473342388868332, + 0.028435923159122467, + 0.012024478055536747, + 0.017960697412490845, + 0.013393989764153957, + 0.01113833300769329, + 0.011399798095226288, + 0.011259966529905796, + 0.011153350584208965, + 0.018476111814379692, + 0.01230200007557869, + 0.00969892367720604, + 0.012534280307590961, + 0.007788064889609814, + 0.007472648285329342, + 0.00834331102669239, + 0.006035597063601017, + 0.007110367063432932, + 0.008002697490155697, + 0.00820869393646717, + 0.009799983352422714, + 0.005252215079963207, + 0.014585550874471664, + 0.013039122335612774, + 0.019828597083687782, + 0.020135430619120598, + 0.012102204374969006, + 0.00901877786964178, + 0.015809472650289536, + 0.008011321537196636, + 0.02086983248591423, + 0.035086631774902344, + 0.01216387189924717, + 0.018704384565353394, + 0.013851865194737911, + 0.010841126553714275, + 0.015017587691545486, + 0.012062674388289452, + 0.020150063559412956, + 0.017441172152757645, + 0.013906063511967659, + 0.01874859631061554, + 0.02674536593258381, + 0.02037300355732441, + 0.016463175415992737, + 0.02185004949569702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.026350490748882294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006380429258570075 + ], + "opposite_logprob_tokens": [ + -7.375638008117676 + ], + "logprob_reward": -0.0006380429258570075, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006380429258570075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10890992730855942, + 0.06660296022891998, + 0.06100904941558838, + 0.054420653730630875, + 0.10144411027431488, + 0.17956417798995972, + 0.14302869141101837, + 0.03547225892543793, + 0.13283483684062958, + 0.11671333014965057 + ], + "credit_map": [ + 0.013103422708809376, + 0.013773761689662933, + 0.016976645216345787, + 0.011073937639594078, + 0.016896992921829224, + 0.013075766153633595, + 0.009589043445885181, + 0.014420362189412117, + 0.00959761068224907, + 0.0065208422020077705, + 0.009080518037080765, + 0.009566747583448887, + 0.004849927965551615, + 0.010694253258407116, + 0.005420414265245199, + 0.010872646234929562, + 0.009375441819429398, + 0.00779435969889164, + 0.007397226057946682, + 0.006659266073256731, + 0.007632744964212179, + 0.008121900260448456, + 0.007430379278957844, + 0.006597733125090599, + 0.0060607753694057465, + 0.008349133655428886, + 0.005139551125466824, + 0.005691098514944315, + 0.010460576973855495, + 0.006152424495667219, + 0.005761526525020599, + 0.006805569864809513, + 0.010568319819867611, + 0.012518550269305706, + 0.011569556780159473, + 0.02134464681148529, + 0.012649674899876118, + 0.016227714717388153, + 0.009906859137117863, + 0.006658782716840506, + 0.03159989044070244, + 0.020767590031027794, + 0.024027742445468903, + 0.034988660365343094, + 0.016775062307715416, + 0.014849000610411167, + 0.023348502814769745, + 0.01320772897452116, + 0.01958337612450123, + 0.015295388177037239, + 0.025563957169651985, + 0.024113936349749565, + 0.01667032018303871, + 0.01048475131392479, + 0.019913097843527794, + 0.011403870768845081, + 0.004208933562040329, + 0.005361493676900864, + 0.005423864349722862, + 0.005273202434182167, + 0.0047314963303506374, + 0.0030137018766254187, + 0.004248480312526226, + 0.003211087780073285, + 0.020895136520266533, + 0.028663404285907745, + 0.011344082653522491, + 0.01609233021736145, + 0.015047335997223854, + 0.015099894255399704, + 0.013528638519346714, + 0.01216401532292366, + 0.013124063611030579, + 0.012424999848008156, + 0.016841450706124306, + 0.016892869025468826, + 0.016884390264749527, + 0.012584902346134186, + 0.012199947610497475, + 0.01576070673763752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.0006380429258570075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028191369492560625 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028191369492560625, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028191369492560625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20014524459838867, + 0.07807227224111557, + 0.0911964550614357, + 0.03872888907790184, + 0.21903565526008606, + 0.08296208828687668, + 0.08486469089984894, + 0.06880590319633484, + 0.05869974195957184, + 0.07748907059431076 + ], + "credit_map": [ + 0.03251698985695839, + 0.032790061086416245, + 0.012200931087136269, + 0.02024206332862377, + 0.03275231271982193, + 0.038478877395391464, + 0.0133733619004488, + 0.017790645360946655, + 0.007239252328872681, + 0.013106260448694229, + 0.011660408228635788, + 0.014615140855312347, + 0.007923648692667484, + 0.009851641021668911, + 0.006224337965250015, + 0.00745158176869154, + 0.009834627620875835, + 0.01625879295170307, + 0.006218596827238798, + 0.009202071465551853, + 0.012588516809046268, + 0.011182612739503384, + 0.012692569755017757, + 0.013218664564192295, + 0.0059196497313678265, + 0.007267087697982788, + 0.0025380223523825407, + 0.003849772270768881, + 0.004046316724270582, + 0.007893849164247513, + 0.0029645299073308706, + 0.00424965750426054, + 0.024503175169229507, + 0.027597753331065178, + 0.02970898151397705, + 0.04355018213391304, + 0.024390874430537224, + 0.03128493204712868, + 0.021941186860203743, + 0.016058577224612236, + 0.010509459301829338, + 0.009379792958498001, + 0.010719011537730694, + 0.015518369153141975, + 0.0072533912025392056, + 0.007853969000279903, + 0.012880519963800907, + 0.008847564458847046, + 0.010223546996712685, + 0.014297159388661385, + 0.010412108153104782, + 0.009106517769396305, + 0.012384307570755482, + 0.00959062110632658, + 0.009388989768922329, + 0.00946144200861454, + 0.009362081997096539, + 0.00788035523146391, + 0.009755670092999935, + 0.015105407685041428, + 0.007368104998022318, + 0.005030218977481127, + 0.009288574568927288, + 0.005015487316995859, + 0.008256150409579277, + 0.011561509221792221, + 0.0046229236759245396, + 0.004421365447342396, + 0.007642700336873531, + 0.008674201555550098, + 0.006194188725203276, + 0.007326704915612936, + 0.013978729024529457, + 0.009985429234802723, + 0.01400243490934372, + 0.0091032013297081, + 0.006474768277257681, + 0.005962538532912731, + 0.008116563782095909, + 0.009865405969321728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0028191369492560625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage road bike with a weathered burgundy frame and brown leather handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bike start by plummeting from the edge of a high, sun-bleached wooden pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a vast, deep blue ocean below the pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward to follow the bike's arc through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bike pierce the water's surface creating a plume of crystal-white spray and jagged bubbles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011799524072557688 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011799524072557688, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011799524072557688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22206591069698334, + 0.12103091925382614, + 0.06946868449449539, + 0.02898148074746132, + 0.05101504921913147, + 0.07399817556142807, + 0.06446875631809235, + 0.08795478194952011, + 0.15850502252578735, + 0.12251120805740356 + ], + "credit_map": [ + 0.03651939332485199, + 0.035331081598997116, + 0.019057074561715126, + 0.023851051926612854, + 0.029392948374152184, + 0.046118393540382385, + 0.015577818267047405, + 0.016218135133385658, + 0.009557140059769154, + 0.02901405282318592, + 0.00755936186760664, + 0.011282101273536682, + 0.02331848070025444, + 0.02013039030134678, + 0.010903547517955303, + 0.009265843778848648, + 0.006611732300370932, + 0.010212806053459644, + 0.004041300620883703, + 0.008120261132717133, + 0.009941176511347294, + 0.008914933539927006, + 0.00993984006345272, + 0.011686629615724087, + 0.00280768028460443, + 0.003495928831398487, + 0.002726075006648898, + 0.005074846092611551, + 0.0020199131686240435, + 0.00541016086935997, + 0.003041112795472145, + 0.004405762534588575, + 0.0083125876262784, + 0.005138202104717493, + 0.009756654500961304, + 0.007656996138393879, + 0.0048749735578894615, + 0.005120436195284128, + 0.0053361509926617146, + 0.0048190453089773655, + 0.010425573214888573, + 0.009142609313130379, + 0.008915786631405354, + 0.014454187825322151, + 0.00864324439316988, + 0.007192457094788551, + 0.010490444488823414, + 0.004733867011964321, + 0.007339400704950094, + 0.006120653823018074, + 0.00954291969537735, + 0.009496466256678104, + 0.008198393508791924, + 0.006164959631860256, + 0.01211144495755434, + 0.005494516342878342, + 0.011889097280800343, + 0.011133897118270397, + 0.01563519984483719, + 0.01888219080865383, + 0.007590567227452993, + 0.0051049841567873955, + 0.011329551227390766, + 0.006389289163053036, + 0.02697225846350193, + 0.032478757202625275, + 0.01639920100569725, + 0.022285502403974533, + 0.014143552631139755, + 0.01270445715636015, + 0.019698888063430786, + 0.013822388835251331, + 0.02244245819747448, + 0.014261267147958279, + 0.011640421114861965, + 0.010669562965631485, + 0.013973940163850784, + 0.012660035863518715, + 0.009674454107880592, + 0.027189061045646667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0011799524072557688 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_4/index_tensor([3006, 3006])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_4/index_tensor([3006, 3006])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..95d2578b6f3527d23f41e5095300a0ad9380bd9b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_4/index_tensor([3006, 3006])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in dramatic slow motion, this extreme close-up, high-angle shot observes a perfect maraschino cherry falling into a vibrant, luminous cocktail. The camera remains static, fixed on the crystal-clear coupe glass where the glowing electric-cyan liquid awaits. The cherry, a shimmering sphere of deep candied crimson, its slender stem still attached, decisively shatters the tense surface tension. A delicate, crown-shaped spray erupts silently, sending microscopic, jewel-like droplets into the air. As the cherry gracefully sinks into the luminous depths, trailing a stream of tiny bubbles, a series of perfect concentric ripples spreads across the liquid's surface, capturing and distorting sharp, focused rays of light from an unseen source into mesmerizing wavy patterns. The color palette is a study in high-contrast beauty—the intense red of the fruit contrasts with the cool, translucent cyan of the drink, all set against a deeply out-of-focus background of warm golden bokeh lights, creating an exquisite and visually arresting atmosphere of pure sensual indulgence.", + "seed": 25939, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_4/index_3006_seed_25939_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.050326093938201664, + "vlm_logprob": -0.050326093938201664, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.050326093938201664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027960440143942833 + ], + "opposite_logprob_tokens": [ + -3.7779605388641357 + ], + "logprob_reward": -0.027960440143942833, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027960440143942833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10503571480512619, + 0.05233895033597946, + 0.131028950214386, + 0.3894653022289276, + 0.020046202465891838, + 0.04660487174987793, + 0.025508388876914978, + 0.15788348019123077, + 0.01627122424542904, + 0.05581696331501007 + ], + "credit_map": [ + 0.010250250808894634, + 0.007458361331373453, + 0.027515696361660957, + 0.010802207514643669, + 0.009510784409940243, + 0.007116006687283516, + 0.02554875798523426, + 0.00683364924043417, + 0.004694506525993347, + 0.005434897728264332, + 0.007082907482981682, + 0.005646791309118271, + 0.011109509505331516, + 0.0038495282642543316, + 0.007260968908667564, + 0.007259838283061981, + 0.01727326214313507, + 0.012116415426135063, + 0.016571367159485817, + 0.015338001772761345, + 0.030094334855675697, + 0.012994317337870598, + 0.016205046325922012, + 0.010436217300593853, + 0.062443945556879044, + 0.01239918451756239, + 0.039952777326107025, + 0.013691562227904797, + 0.025475991889834404, + 0.1365724503993988, + 0.01965026929974556, + 0.07927911728620529, + 0.002720378339290619, + 0.0019691900815814734, + 0.0018644239753484726, + 0.0030748997814953327, + 0.0035460973158478737, + 0.0017874946352094412, + 0.002241285750642419, + 0.002842432353645563, + 0.00638568727299571, + 0.004570457153022289, + 0.0042108879424631596, + 0.003869266016408801, + 0.01503208838403225, + 0.005505699198693037, + 0.0028807169292122126, + 0.004150065593421459, + 0.002773169195279479, + 0.002684566890820861, + 0.004048556089401245, + 0.0028244764544069767, + 0.003583186771720648, + 0.0036050230264663696, + 0.002910385373979807, + 0.003079024376347661, + 0.014154012314975262, + 0.02368130162358284, + 0.024354571476578712, + 0.006011388264596462, + 0.058971602469682693, + 0.01353688444942236, + 0.009285157546401024, + 0.00788856390863657, + 0.0018168948590755463, + 0.0029759735334664583, + 0.00130041956435889, + 0.0010481496574357152, + 0.004913994576781988, + 0.0014968125615268946, + 0.001621795236133039, + 0.0010971847223117948, + 0.006227281875908375, + 0.006631569936871529, + 0.0029421125072985888, + 0.0021433494985103607, + 0.022579671815037727, + 0.0058693652972579, + 0.005433357786387205, + 0.003990253899246454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.027960440143942833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041116416454315 + ], + "opposite_logprob_tokens": [ + -1.9104111194610596 + ], + "logprob_reward": -0.16041116416454315, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041116416454315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058138590306043625, + 0.03442774713039398, + 0.07557563483715057, + 0.41462230682373047, + 0.022843606770038605, + 0.02755018323659897, + 0.1592065989971161, + 0.050445761531591415, + 0.026431290432810783, + 0.13075828552246094 + ], + "credit_map": [ + 0.00472679641097784, + 0.004041214007884264, + 0.013643283396959305, + 0.005299493204802275, + 0.005245267413556576, + 0.004174801055341959, + 0.017537273466587067, + 0.003470460884273052, + 0.005194240715354681, + 0.0037119020707905293, + 0.004876638762652874, + 0.002952201757580042, + 0.004192929249256849, + 0.003314174013212323, + 0.0042084138840436935, + 0.005977244582027197, + 0.009323368780314922, + 0.006401584483683109, + 0.01353025808930397, + 0.007014941424131393, + 0.01707879826426506, + 0.0059187826700508595, + 0.00977900717407465, + 0.006528889294713736, + 0.06691968441009521, + 0.013761049136519432, + 0.04045160859823227, + 0.017675423994660378, + 0.026009999215602875, + 0.14579090476036072, + 0.02066301740705967, + 0.08335062116384506, + 0.004094236996024847, + 0.003021854907274246, + 0.0027883003931492567, + 0.002266837749630213, + 0.002782277064397931, + 0.0032717071007937193, + 0.0020984485745429993, + 0.002519942820072174, + 0.0029392922297120094, + 0.003216796787455678, + 0.0025901508051902056, + 0.004886668641120195, + 0.005862121935933828, + 0.0029829824343323708, + 0.0028477173764258623, + 0.002224452793598175, + 0.009599466808140278, + 0.027287226170301437, + 0.009650642983615398, + 0.009127664379775524, + 0.05947732552886009, + 0.024479666724801064, + 0.011021283455193043, + 0.008563314564526081, + 0.004696801770478487, + 0.010452181100845337, + 0.003423691261559725, + 0.002479287562891841, + 0.019392505288124084, + 0.004524650052189827, + 0.0025181772653013468, + 0.0029584672302007675, + 0.00323666213080287, + 0.0042018224485218525, + 0.0028496766462922096, + 0.0021603729110211134, + 0.004019209183752537, + 0.002951174508780241, + 0.0029619873967021704, + 0.004050385672599077, + 0.015520778484642506, + 0.017118997871875763, + 0.012301452457904816, + 0.01038477011024952, + 0.03722250461578369, + 0.01422131434082985, + 0.011281046085059643, + 0.012707428075373173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.16041116416454315 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043095141649246216 + ], + "opposite_logprob_tokens": [ + -3.168095111846924 + ], + "logprob_reward": -0.043095141649246216, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043095141649246216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14051386713981628, + 0.031391628086566925, + 0.17906427383422852, + 0.26239848136901855, + 0.020291557535529137, + 0.12125691771507263, + 0.08488798141479492, + 0.07717543095350266, + 0.04185108095407486, + 0.04116876423358917 + ], + "credit_map": [ + 0.018760280683636665, + 0.009833406656980515, + 0.029362071305513382, + 0.016195278614759445, + 0.01244862750172615, + 0.00944752898067236, + 0.03461328148841858, + 0.009853399358689785, + 0.004245865624397993, + 0.0032110954634845257, + 0.004484407603740692, + 0.003153853816911578, + 0.005610222928225994, + 0.0026327853556722403, + 0.004051859490573406, + 0.004001536872237921, + 0.020518267527222633, + 0.014384963549673557, + 0.026738101616501808, + 0.01997186988592148, + 0.043620962649583817, + 0.014046679250895977, + 0.02566460892558098, + 0.014118821360170841, + 0.04249279573559761, + 0.009904639795422554, + 0.025410303846001625, + 0.009470772929489613, + 0.01877882517874241, + 0.09201790392398834, + 0.013153744861483574, + 0.05116947740316391, + 0.0029669925570487976, + 0.0019697940442711115, + 0.0016802003374323249, + 0.0021227013785392046, + 0.0054525951854884624, + 0.002021902007982135, + 0.0019042965723201632, + 0.0021730747539550066, + 0.018048902973532677, + 0.010297824628651142, + 0.008738263510167599, + 0.005737202242016792, + 0.04609547182917595, + 0.012318378314375877, + 0.007377229165285826, + 0.01264363992959261, + 0.005098964553326368, + 0.009537900798022747, + 0.007903574965894222, + 0.004910176154226065, + 0.02982979640364647, + 0.015395741909742355, + 0.007486110553145409, + 0.004725707694888115, + 0.007018047384917736, + 0.012497425079345703, + 0.00883061345666647, + 0.0040497244335711, + 0.02775225043296814, + 0.00694641238078475, + 0.004888502880930901, + 0.005192453507333994, + 0.0050691356882452965, + 0.00746182119473815, + 0.005568610038608313, + 0.004226404242217541, + 0.006386426743119955, + 0.0033654249273240566, + 0.004472441505640745, + 0.00530081195756793, + 0.005372883751988411, + 0.0073872460052371025, + 0.005183972883969545, + 0.003190506249666214, + 0.006320585962384939, + 0.004560534842312336, + 0.004330017603933811, + 0.004823020193725824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.043095141649246216 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09543535858392715, + 0.06699021905660629, + 0.0721394419670105, + 0.40741103887557983, + 0.028293892741203308, + 0.05908957123756409, + 0.04695618525147438, + 0.13207696378231049, + 0.0224146768450737, + 0.06919261813163757 + ], + "credit_map": [ + 0.01316483411937952, + 0.013658522628247738, + 0.012109657749533653, + 0.014058351516723633, + 0.011382699012756348, + 0.008268802426755428, + 0.012557312846183777, + 0.010235179215669632, + 0.006789418403059244, + 0.005819115787744522, + 0.012454644776880741, + 0.008268063887953758, + 0.011552520096302032, + 0.004918314050883055, + 0.008717694319784641, + 0.008470448665320873, + 0.008227787911891937, + 0.005412936210632324, + 0.008918918669223785, + 0.0074956477619707584, + 0.016529420390725136, + 0.007291336543858051, + 0.01044780109077692, + 0.007815593853592873, + 0.06398362666368484, + 0.014200834557414055, + 0.04146481677889824, + 0.015035182237625122, + 0.027373911812901497, + 0.14084334671497345, + 0.022453520447015762, + 0.08205580711364746, + 0.0034642904065549374, + 0.0031948911491781473, + 0.0022373006213456392, + 0.002321123145520687, + 0.00795380026102066, + 0.003781374543905258, + 0.0027037234976887703, + 0.002637390745803714, + 0.007905697450041771, + 0.007092135027050972, + 0.006836066488176584, + 0.006109689828008413, + 0.010394847951829433, + 0.008024240843951702, + 0.005293423309922218, + 0.007433475460857153, + 0.004268258810043335, + 0.00448262132704258, + 0.009172098711133003, + 0.004798066336661577, + 0.006576962303370237, + 0.006795442197471857, + 0.005319633055478334, + 0.005543102975934744, + 0.010940015316009521, + 0.016638193279504776, + 0.027790408581495285, + 0.004413148388266563, + 0.04678869619965553, + 0.011163533665239811, + 0.008984223008155823, + 0.005358744412660599, + 0.003496241755783558, + 0.0028817381244152784, + 0.003577327588573098, + 0.0013277368852868676, + 0.0031623479444533587, + 0.0024819120299071074, + 0.0032366004306823015, + 0.002250773599371314, + 0.007975880056619644, + 0.008242497220635414, + 0.004946645814925432, + 0.004653530661016703, + 0.024942856281995773, + 0.007144388277083635, + 0.006217914633452892, + 0.005068898666650057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.007630960550159216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532763183116913 + ], + "opposite_logprob_tokens": [ + -4.387532711029053 + ], + "logprob_reward": -0.012532763183116913, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532763183116913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0692390501499176, + 0.016419516876339912, + 0.05573000758886337, + 0.059894852340221405, + 0.016538474708795547, + 0.1566518098115921, + 0.1613881140947342, + 0.3205832242965698, + 0.09851284325122833, + 0.04504213482141495 + ], + "credit_map": [ + 0.010373316705226898, + 0.006550484802573919, + 0.01279863715171814, + 0.009159361943602562, + 0.008242898620665073, + 0.006998059339821339, + 0.009501945227384567, + 0.005614343099296093, + 0.0016956611070781946, + 0.0019391673849895597, + 0.003440930275246501, + 0.0014310438418760896, + 0.0017562286229804158, + 0.0013562883250415325, + 0.0022422969341278076, + 0.002557900035753846, + 0.009926819242537022, + 0.005562398582696915, + 0.006924486719071865, + 0.005190886557102203, + 0.011392585933208466, + 0.006057349033653736, + 0.006175340618938208, + 0.004500138573348522, + 0.007376395165920258, + 0.004535453859716654, + 0.006612233817577362, + 0.0027644780930131674, + 0.010888737626373768, + 0.014588315971195698, + 0.0037310300394892693, + 0.009398207068443298, + 0.002467489568516612, + 0.001635093125514686, + 0.0017169294878840446, + 0.0018482502782717347, + 0.0036138067953288555, + 0.0014549767365679145, + 0.0018857895629480481, + 0.0019161393865942955, + 0.024417312815785408, + 0.012423762120306492, + 0.009395873174071312, + 0.009197517298161983, + 0.06069639325141907, + 0.0167580284178257, + 0.009716649539768696, + 0.01404627040028572, + 0.011789793148636818, + 0.027182314544916153, + 0.014225510880351067, + 0.009127342142164707, + 0.05467837303876877, + 0.020243393257260323, + 0.013190152123570442, + 0.010951230302453041, + 0.027632810175418854, + 0.0551149956882, + 0.04215520992875099, + 0.009738875553011894, + 0.12695452570915222, + 0.029625628143548965, + 0.01630222797393799, + 0.013058935292065144, + 0.012078957632184029, + 0.01887151598930359, + 0.0062829842790961266, + 0.006163685582578182, + 0.02948629856109619, + 0.008743439801037312, + 0.009599286131560802, + 0.007286673877388239, + 0.006074980832636356, + 0.007912855595350266, + 0.0049043986946344376, + 0.0037903189659118652, + 0.009631142020225525, + 0.004232962150126696, + 0.004670342896133661, + 0.003825131803750992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.012532763183116913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, this extreme close-up, high-angle shot observes a perfect maraschino cherry falling into a vibrant, luminous cocktail. The camera remains static, fixed on the crystal-clear coupe glass where the glowing electric-cyan liquid awaits. The cherry, a shimmering sphere of deep candied crimson, its slender stem still attached, decisively shatters the tense surface tension. A delicate, crown-shaped spray erupts silently, sending microscopic, jewel-like droplets into the air. As the cherry gracefully sinks into the luminous depths, trailing a stream of tiny bubbles, a series of perfect concentric ripples spreads across the liquid's surface, capturing and distorting sharp, focused rays of light from an unseen source into mesmerizing wavy patterns. The color palette is a study in high-contrast beauty—the intense red of the fruit contrasts with the cool, translucent cyan of the drink, all set against a deeply out-of-focus background of warm golden bokeh lights, creating an exquisite and visually arresting atmosphere of pure sensual indulgence.", + "seed": 25940, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_4/index_3006_seed_25940_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024135226383805275, + "vlm_logprob": -0.024135226383805275, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.024135226383805275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017068514600396156 + ], + "opposite_logprob_tokens": [ + -4.517068386077881 + ], + "logprob_reward": -0.017068514600396156, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017068514600396156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07672480493783951, + 0.08742053806781769, + 0.2883811593055725, + 0.12974001467227936, + 0.07417133450508118, + 0.17606964707374573, + 0.018284142017364502, + 0.0463416688144207, + 0.017591966316103935, + 0.08527468889951706 + ], + "credit_map": [ + 0.011443129740655422, + 0.008747326210141182, + 0.008562913164496422, + 0.010994619689881802, + 0.011734819039702415, + 0.010377527214586735, + 0.007910050451755524, + 0.006954418960958719, + 0.009595667943358421, + 0.01170024648308754, + 0.012049978598952293, + 0.011759471148252487, + 0.014691539108753204, + 0.007210925687104464, + 0.01273947674781084, + 0.007673231419175863, + 0.0370049774646759, + 0.028846653178334236, + 0.025120073929429054, + 0.03000766411423683, + 0.05306774750351906, + 0.035408709198236465, + 0.038723159581422806, + 0.040202174335718155, + 0.022403789684176445, + 0.016027972102165222, + 0.011957231909036636, + 0.016322694718837738, + 0.022198395803570747, + 0.01600274071097374, + 0.013996144756674767, + 0.010831044055521488, + 0.008964740671217442, + 0.007905824109911919, + 0.006729877553880215, + 0.005530145484954119, + 0.018555430695414543, + 0.011791638098657131, + 0.008143722079694271, + 0.006549951620399952, + 0.051820989698171616, + 0.0178934745490551, + 0.011168422177433968, + 0.03826598450541496, + 0.006025086622685194, + 0.019246552139520645, + 0.015498255379498005, + 0.01615089178085327, + 0.0026063460391014814, + 0.0019787221681326628, + 0.0019694955553859472, + 0.0018991118995472789, + 0.0014803077792748809, + 0.0019612377509474754, + 0.004588004667311907, + 0.001800917903892696, + 0.010038323700428009, + 0.0039475844241678715, + 0.004094152245670557, + 0.005545284133404493, + 0.002710994565859437, + 0.0056281075812876225, + 0.011197509244084358, + 0.003179715247824788, + 0.003460478037595749, + 0.0017571819480508566, + 0.0018889224156737328, + 0.0017020200612023473, + 0.0014534547226503491, + 0.001851410954259336, + 0.0037651732563972473, + 0.0017133260844275355, + 0.013937259092926979, + 0.007180479355156422, + 0.004483921453356743, + 0.0048842234537005424, + 0.033833812922239304, + 0.013021687045693398, + 0.005000707693397999, + 0.0029325878713279963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.017068514600396156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374939039349556 + ], + "opposite_logprob_tokens": [ + -3.4087493419647217 + ], + "logprob_reward": -0.03374939039349556, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374939039349556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6215788722038269, + 0.06913496553897858, + 0.049334075301885605, + 0.03619367629289627, + 0.07019493728876114, + 0.03990314528346062, + 0.009136130101978779, + 0.00821277778595686, + 0.022639460861682892, + 0.07367194443941116 + ], + "credit_map": [ + 0.16612643003463745, + 0.02944113500416279, + 0.02918395772576332, + 0.20485857129096985, + 0.013766041025519371, + 0.060082364827394485, + 0.09026513993740082, + 0.027855215594172478, + 0.012221382930874825, + 0.007439841516315937, + 0.009223130531609058, + 0.012754729948937893, + 0.008842313662171364, + 0.005470580421388149, + 0.00684377970173955, + 0.006339206825941801, + 0.009837158024311066, + 0.005566595587879419, + 0.004723719321191311, + 0.006447458174079657, + 0.006929983850568533, + 0.005320459604263306, + 0.005833599250763655, + 0.0046751052141189575, + 0.007925889454782009, + 0.0033516816329210997, + 0.003626518417149782, + 0.004466381389647722, + 0.00582832982763648, + 0.005014983471482992, + 0.003063830081373453, + 0.002916062017902732, + 0.020030025392770767, + 0.0064937276765704155, + 0.005809078458696604, + 0.011372681707143784, + 0.005667614284902811, + 0.009852488525211811, + 0.004703699145466089, + 0.00626562163233757, + 0.008272610604763031, + 0.006652312818914652, + 0.003818012773990631, + 0.005130051169544458, + 0.003012261353433132, + 0.003950697369873524, + 0.006051092408597469, + 0.0030161060858517885, + 0.0013304005842655897, + 0.0009646144462749362, + 0.0013308750931173563, + 0.001507125562056899, + 0.0008950895862653852, + 0.0008607187774032354, + 0.0012141980696469545, + 0.0010331079829484224, + 0.0010988349094986916, + 0.0008838902576826513, + 0.0009329813765361905, + 0.0010273015359416604, + 0.0009918523719534278, + 0.0007035062299109995, + 0.0016184116248041391, + 0.0009559997706674039, + 0.003645215881988406, + 0.002185695106163621, + 0.0021316728089004755, + 0.004446648992598057, + 0.0020596813410520554, + 0.0027470095083117485, + 0.0032710442319512367, + 0.0021524918265640736, + 0.01599145121872425, + 0.0059741889126598835, + 0.00489956745877862, + 0.004605693276971579, + 0.02629205398261547, + 0.007962626405060291, + 0.0045367395505309105, + 0.003409619675949216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.03374939039349556 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337347537279129 + ], + "opposite_logprob_tokens": [ + -3.4087347984313965 + ], + "logprob_reward": -0.0337347537279129, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337347537279129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2991339862346649, + 0.05677824839949608, + 0.1948653608560562, + 0.05901101231575012, + 0.21317537128925323, + 0.056082066148519516, + 0.012755341827869415, + 0.036392904818058014, + 0.01787334680557251, + 0.05393236503005028 + ], + "credit_map": [ + 0.07242283970117569, + 0.017421016469597816, + 0.016950396820902824, + 0.09440132230520248, + 0.012819941155612469, + 0.027280298992991447, + 0.04221648350358009, + 0.015621703118085861, + 0.0084213363006711, + 0.00731666898354888, + 0.008056538179516792, + 0.008685730397701263, + 0.007492607459425926, + 0.004868597257882357, + 0.005401523318141699, + 0.006535247899591923, + 0.026781057938933372, + 0.01983768492937088, + 0.024307776242494583, + 0.032207559794187546, + 0.027828391641378403, + 0.025575757026672363, + 0.0209767147898674, + 0.01735042594373226, + 0.010222481563687325, + 0.008656485006213188, + 0.007676353212445974, + 0.007308477535843849, + 0.007243376225233078, + 0.0075547765009105206, + 0.004888584837317467, + 0.005460479762405157, + 0.04609251022338867, + 0.024004362523555756, + 0.020569907501339912, + 0.03223301097750664, + 0.022650672122836113, + 0.028451578691601753, + 0.015962902456521988, + 0.023210404440760612, + 0.008254996500909328, + 0.007333172019571066, + 0.005082519259303808, + 0.006251222919672728, + 0.005210723727941513, + 0.006743463687598705, + 0.01280013844370842, + 0.004405830521136522, + 0.0017899897648021579, + 0.0016764361644163728, + 0.001641443115659058, + 0.0015066046034917235, + 0.0011764008086174726, + 0.0012825452722609043, + 0.0024690087884664536, + 0.0012129141250625253, + 0.009117563255131245, + 0.004037128295749426, + 0.00331558333709836, + 0.0031294161453843117, + 0.0025138927157968283, + 0.004690675064921379, + 0.006622329819947481, + 0.0029663173481822014, + 0.0028707273304462433, + 0.001926827128045261, + 0.002441219985485077, + 0.001598224975168705, + 0.0016734503442421556, + 0.002078489400446415, + 0.00356440432369709, + 0.0017200034344568849, + 0.009571623988449574, + 0.004550418816506863, + 0.005903509445488453, + 0.004643297288566828, + 0.009050937369465828, + 0.013002816587686539, + 0.004542715381830931, + 0.00266704591922462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0337347537279129 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784476831555367 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784476831555367, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784476831555367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23784661293029785, + 0.14125899970531464, + 0.09821246564388275, + 0.12245170027017593, + 0.08627255260944366, + 0.15641018748283386, + 0.021464617922902107, + 0.06287147104740143, + 0.017909977585077286, + 0.055301446467638016 + ], + "credit_map": [ + 0.05266990512609482, + 0.016596777364611626, + 0.0142460772767663, + 0.07040046900510788, + 0.01819571480154991, + 0.02131001651287079, + 0.03245741128921509, + 0.011970235966145992, + 0.026133162900805473, + 0.012608307413756847, + 0.013382884673774242, + 0.028472980484366417, + 0.02529517561197281, + 0.012330397963523865, + 0.014142069034278393, + 0.008894030936062336, + 0.013780303299427032, + 0.010935477912425995, + 0.009436502121388912, + 0.01074434258043766, + 0.01593765616416931, + 0.01190259587019682, + 0.012974891811609268, + 0.012500688433647156, + 0.015599293634295464, + 0.013977880589663982, + 0.008225538767874241, + 0.01242742594331503, + 0.03153848275542259, + 0.01787506975233555, + 0.01538801845163107, + 0.007419989909976721, + 0.01796928234398365, + 0.008269432000815868, + 0.0055576516315341, + 0.009708345867693424, + 0.01565680094063282, + 0.013516848906874657, + 0.008919944055378437, + 0.006674247328191996, + 0.04486450180411339, + 0.01052093505859375, + 0.010701450519263744, + 0.03180491179227829, + 0.0058988966047763824, + 0.018727343529462814, + 0.021941376850008965, + 0.01195076759904623, + 0.0039007628802210093, + 0.0026064543053507805, + 0.0023170560598373413, + 0.002489494625478983, + 0.0015736399218440056, + 0.002286756644025445, + 0.004509730264544487, + 0.0017807248514145613, + 0.008393403142690659, + 0.0045661055482923985, + 0.005955331958830357, + 0.005834465846419334, + 0.004844712559133768, + 0.006145056802779436, + 0.021046143025159836, + 0.006086253095418215, + 0.003499024547636509, + 0.0018869435880333185, + 0.001689633703790605, + 0.0030555562116205692, + 0.0016445117071270943, + 0.0020500398240983486, + 0.0022330773063004017, + 0.0018511917442083359, + 0.010371691547334194, + 0.004360312595963478, + 0.0037060019094496965, + 0.003449060255661607, + 0.018492352217435837, + 0.006801988929510117, + 0.005064535420387983, + 0.003055507317185402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.009784476831555367 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633899636566639 + ], + "opposite_logprob_tokens": [ + -3.651339054107666 + ], + "logprob_reward": -0.02633899636566639, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633899636566639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5554083585739136, + 0.03833454102277756, + 0.027279719710350037, + 0.04807284474372864, + 0.04417484998703003, + 0.13584385812282562, + 0.011411014012992382, + 0.01959165744483471, + 0.028295734897255898, + 0.09158746153116226 + ], + "credit_map": [ + 0.14622898399829865, + 0.02672911435365677, + 0.026906654238700867, + 0.18132264912128448, + 0.013324555940926075, + 0.05454286187887192, + 0.0810101255774498, + 0.025343406945466995, + 0.0068259998224675655, + 0.00340043636970222, + 0.004540208261460066, + 0.004457299597561359, + 0.007486989721655846, + 0.004109387286007404, + 0.0042204041965305805, + 0.003293816465884447, + 0.004559042397886515, + 0.002467480953782797, + 0.0038564575370401144, + 0.00491691380739212, + 0.003143430920317769, + 0.0030316784977912903, + 0.0029597075190395117, + 0.0023450071457773447, + 0.008016619831323624, + 0.005988057237118483, + 0.004157545976340771, + 0.005357590038329363, + 0.0053352732211351395, + 0.009084563702344894, + 0.004720321390777826, + 0.00541287288069725, + 0.010073118843138218, + 0.005724809132516384, + 0.0034582617226988077, + 0.006664576940238476, + 0.004896693397313356, + 0.005442955531179905, + 0.003858707845211029, + 0.004055728670209646, + 0.022862285375595093, + 0.01982821524143219, + 0.01658068224787712, + 0.015725551173090935, + 0.011515402235090733, + 0.019085777923464775, + 0.01619410514831543, + 0.014051833190023899, + 0.001457778038457036, + 0.0015826047165319324, + 0.0012513796100392938, + 0.0011493166675791144, + 0.0014049677411094308, + 0.001259368029423058, + 0.002173819812014699, + 0.0011317795142531395, + 0.002923305844888091, + 0.002559593878686428, + 0.0031098786275833845, + 0.003333503846079111, + 0.0018074044492095709, + 0.0024230859708040953, + 0.0013568017166107893, + 0.0020780828781425953, + 0.00427269097417593, + 0.003849388100206852, + 0.004549697507172823, + 0.004099616780877113, + 0.0031681526452302933, + 0.002221918199211359, + 0.0035330329556018114, + 0.002601237967610359, + 0.0278908833861351, + 0.006192818749696016, + 0.009477447718381882, + 0.004704242572188377, + 0.026203708723187447, + 0.008433308452367783, + 0.004885547794401646, + 0.003799496917054057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.02633899636566639 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, this extreme close-up, high-angle shot observes a perfect maraschino cherry falling into a vibrant, luminous cocktail. The camera remains static, fixed on the crystal-clear coupe glass where the glowing electric-cyan liquid awaits. The cherry, a shimmering sphere of deep candied crimson, its slender stem still attached, decisively shatters the tense surface tension. A delicate, crown-shaped spray erupts silently, sending microscopic, jewel-like droplets into the air. As the cherry gracefully sinks into the luminous depths, trailing a stream of tiny bubbles, a series of perfect concentric ripples spreads across the liquid's surface, capturing and distorting sharp, focused rays of light from an unseen source into mesmerizing wavy patterns. The color palette is a study in high-contrast beauty—the intense red of the fruit contrasts with the cool, translucent cyan of the drink, all set against a deeply out-of-focus background of warm golden bokeh lights, creating an exquisite and visually arresting atmosphere of pure sensual indulgence.", + "seed": 25941, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_4/index_3006_seed_25941_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01993446600390598, + "vlm_logprob": -0.01993446600390598, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.01993446600390598, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01048083882778883 + ], + "opposite_logprob_tokens": [ + -4.885480880737305 + ], + "logprob_reward": -0.01048083882778883, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01048083882778883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09605224430561066, + 0.10510634630918503, + 0.14883773028850555, + 0.09799320250749588, + 0.04619026184082031, + 0.12723742425441742, + 0.05324744060635567, + 0.15916866064071655, + 0.12854643166065216, + 0.03762021288275719 + ], + "credit_map": [ + 0.014473853632807732, + 0.017363734543323517, + 0.012521497905254364, + 0.010258671827614307, + 0.011352195404469967, + 0.009567908011376858, + 0.011096340604126453, + 0.009418047033250332, + 0.008296721614897251, + 0.020649535581469536, + 0.00802389346063137, + 0.01832360029220581, + 0.014627907425165176, + 0.010262946598231792, + 0.020182205364108086, + 0.004739535041153431, + 0.019390041008591652, + 0.022346703335642815, + 0.015066633000969887, + 0.012213664129376411, + 0.03461705893278122, + 0.022870928049087524, + 0.012785756029188633, + 0.009546958841383457, + 0.014643345959484577, + 0.01563344895839691, + 0.01139877736568451, + 0.014765847474336624, + 0.011152663268148899, + 0.00768913421779871, + 0.01515139825642109, + 0.0075585865415632725, + 0.006196659058332443, + 0.007750530727207661, + 0.004692987073212862, + 0.0051984963938593864, + 0.005556383635848761, + 0.007371691521257162, + 0.00605873903259635, + 0.003364773467183113, + 0.019261043518781662, + 0.01942453719675541, + 0.011398391798138618, + 0.019461441785097122, + 0.01380653865635395, + 0.014530672691762447, + 0.018212443217635155, + 0.011142358183860779, + 0.006205144803971052, + 0.006321751978248358, + 0.006139950826764107, + 0.009663364849984646, + 0.004122552927583456, + 0.006197021808475256, + 0.007514675613492727, + 0.0070829796604812145, + 0.0328347384929657, + 0.01455418486148119, + 0.014730545692145824, + 0.016846751794219017, + 0.010517462156713009, + 0.010798078030347824, + 0.042183827608823776, + 0.01670306921005249, + 0.023650025948882103, + 0.010257238522171974, + 0.016490420326590538, + 0.009495334699749947, + 0.037163250148296356, + 0.011691920459270477, + 0.011575586162507534, + 0.008222651667892933, + 0.004762259777635336, + 0.003123106434941292, + 0.007948190905153751, + 0.003395599313080311, + 0.004987217951565981, + 0.00583394942805171, + 0.0046060094609856606, + 0.002963879844173789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.01048083882778883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007794827222824 + ], + "opposite_logprob_tokens": [ + -2.695077896118164 + ], + "logprob_reward": -0.07007794827222824, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007794827222824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07569745182991028, + 0.17859892547130585, + 0.19579973816871643, + 0.07990221679210663, + 0.07589414715766907, + 0.09257815033197403, + 0.03856488689780235, + 0.05482851713895798, + 0.1743241399526596, + 0.033811841160058975 + ], + "credit_map": [ + 0.007884563878178596, + 0.010228865779936314, + 0.00835419911891222, + 0.014002748765051365, + 0.009341977536678314, + 0.005986412055790424, + 0.014938931912183762, + 0.0049597518518567085, + 0.015480803325772285, + 0.04127107560634613, + 0.01278932485729456, + 0.027368733659386635, + 0.037173643708229065, + 0.011970104649662971, + 0.02539481222629547, + 0.007150416262447834, + 0.024807952344417572, + 0.026781165972352028, + 0.024234814569354057, + 0.01630321517586708, + 0.043623004108667374, + 0.01811831258237362, + 0.016569944098591805, + 0.025361334905028343, + 0.011301238089799881, + 0.01355556957423687, + 0.009243817068636417, + 0.010543794371187687, + 0.00938922818750143, + 0.008956036530435085, + 0.010311600752174854, + 0.006600930821150541, + 0.011173439212143421, + 0.013062434270977974, + 0.006590540055185556, + 0.007132294587790966, + 0.010715014301240444, + 0.013208220712840557, + 0.008399070240557194, + 0.005613135639578104, + 0.015909915789961815, + 0.025768963620066643, + 0.006317812483757734, + 0.008896302431821823, + 0.010126639157533646, + 0.007315644528716803, + 0.011072876863181591, + 0.0071699959225952625, + 0.00391005165874958, + 0.006269962526857853, + 0.004521238151937723, + 0.007263835519552231, + 0.003661828115582466, + 0.0033151160459965467, + 0.004986936692148447, + 0.004635917954146862, + 0.00816871877759695, + 0.005744020454585552, + 0.006243383977562189, + 0.006706929299980402, + 0.006469303276389837, + 0.00546611612662673, + 0.007329145446419716, + 0.0087008997797966, + 0.02447873167693615, + 0.012898641638457775, + 0.030404018238186836, + 0.014427432790398598, + 0.015950514003634453, + 0.007096399087458849, + 0.05084097385406494, + 0.018227415159344673, + 0.003683507675305009, + 0.004399706143885851, + 0.005235759541392326, + 0.0036763562820851803, + 0.003228232264518738, + 0.004070662893354893, + 0.005657482892274857, + 0.0038601309061050415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.07007794827222824 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016098633408546448 + ], + "opposite_logprob_tokens": [ + -4.141098499298096 + ], + "logprob_reward": -0.016098633408546448, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016098633408546448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17966090142726898, + 0.19705256819725037, + 0.09085726737976074, + 0.05298849940299988, + 0.17955107986927032, + 0.08385422825813293, + 0.055692002177238464, + 0.030145810917019844, + 0.09696485847234726, + 0.033232755959033966 + ], + "credit_map": [ + 0.018046250566840172, + 0.029430797323584557, + 0.02676570974290371, + 0.027648955583572388, + 0.018194902688264847, + 0.013935357332229614, + 0.02672204189002514, + 0.018916895613074303, + 0.019220441579818726, + 0.036797404289245605, + 0.020638782531023026, + 0.029482172802090645, + 0.0335933119058609, + 0.02085109055042267, + 0.02401295304298401, + 0.012456436641514301, + 0.012119774706661701, + 0.013111666776239872, + 0.015455846674740314, + 0.008809256367385387, + 0.012907786294817924, + 0.008520874194800854, + 0.008925276808440685, + 0.01100679486989975, + 0.006523977033793926, + 0.008143683895468712, + 0.006884511560201645, + 0.005528182256966829, + 0.009176141582429409, + 0.004322932101786137, + 0.0076860664412379265, + 0.0047230105847120285, + 0.03243515267968178, + 0.046184688806533813, + 0.015066108666360378, + 0.012678723782300949, + 0.030745502561330795, + 0.015541771426796913, + 0.015584458597004414, + 0.011314690113067627, + 0.012182429432868958, + 0.019762661308050156, + 0.007157979998737574, + 0.011378280818462372, + 0.01051661092787981, + 0.007869158871471882, + 0.00899060070514679, + 0.005996518302708864, + 0.008099026046693325, + 0.006839780602604151, + 0.004077439662069082, + 0.004862289410084486, + 0.005335739813745022, + 0.00663454644382, + 0.01294739544391632, + 0.006895791739225388, + 0.003986007999628782, + 0.0037293625064194202, + 0.0048255305737257, + 0.003914015367627144, + 0.0025948553811758757, + 0.002277340507134795, + 0.004836430773139, + 0.003982269670814276, + 0.009611436165869236, + 0.009348636493086815, + 0.019753599539399147, + 0.009670005179941654, + 0.012801977805793285, + 0.005220902152359486, + 0.01820099540054798, + 0.012357317842543125, + 0.0038837529718875885, + 0.004270859993994236, + 0.0038816214073449373, + 0.005469329189509153, + 0.0030898943077772856, + 0.0039864531718194485, + 0.004816215950995684, + 0.003834631061181426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.016098633408546448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022003266494721174 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.0022003266494721174, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022003266494721174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07987070083618164, + 0.10907188057899475, + 0.0827515572309494, + 0.11144261062145233, + 0.2664000391960144, + 0.10841460525989532, + 0.08707992732524872, + 0.09149172902107239, + 0.03979386016726494, + 0.023683112114667892 + ], + "credit_map": [ + 0.011174612678587437, + 0.013111497275531292, + 0.006896603386849165, + 0.00983662623912096, + 0.011591068468987942, + 0.009542465209960938, + 0.01110588014125824, + 0.006611949764192104, + 0.009233838878571987, + 0.01996690407395363, + 0.010986875742673874, + 0.01640419475734234, + 0.02076508104801178, + 0.009879829362034798, + 0.015960348770022392, + 0.005874801892787218, + 0.00990928616374731, + 0.012703824788331985, + 0.01072238851338625, + 0.007093299180269241, + 0.016795163974165916, + 0.0087948739528656, + 0.008033866062760353, + 0.00869885366410017, + 0.014998712576925755, + 0.015658702701330185, + 0.015315902419388294, + 0.012434220872819424, + 0.021609531715512276, + 0.010498798452317715, + 0.011526373215019703, + 0.00940037053078413, + 0.050298016518354416, + 0.05188845843076706, + 0.022719822824001312, + 0.0199282206594944, + 0.04325277358293533, + 0.026095280423760414, + 0.0364367850124836, + 0.01578069105744362, + 0.017642531543970108, + 0.0314953438937664, + 0.007432452868670225, + 0.008014409802854061, + 0.014367015101015568, + 0.008304435759782791, + 0.012182406149804592, + 0.008976011537015438, + 0.013145877048373222, + 0.006793742999434471, + 0.00794473011046648, + 0.011422314681112766, + 0.007422329857945442, + 0.00807082001119852, + 0.019683944061398506, + 0.012596169486641884, + 0.017975131049752235, + 0.007149523589760065, + 0.007528063841164112, + 0.009591435082256794, + 0.007659008260816336, + 0.005447749979794025, + 0.026653286069631577, + 0.009487533941864967, + 0.0043697664514184, + 0.005322068464010954, + 0.006745055317878723, + 0.006428610533475876, + 0.004341651685535908, + 0.0024092223029583693, + 0.005018210504204035, + 0.005159276537597179, + 0.003131814068183303, + 0.0040078130550682545, + 0.0034374885726720095, + 0.0029727844521403313, + 0.0020535436924546957, + 0.0020308734383434057, + 0.003811002243310213, + 0.002237792359665036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.0022003266494721174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000814582861494273 + ], + "opposite_logprob_tokens": [ + -7.125814437866211 + ], + "logprob_reward": -0.000814582861494273, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.000814582861494273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06729110330343246, + 0.05679834261536598, + 0.08418974280357361, + 0.05713658407330513, + 0.10970677435398102, + 0.15726563334465027, + 0.08513760566711426, + 0.047194309532642365, + 0.2584984004497528, + 0.07678146660327911 + ], + "credit_map": [ + 0.0072635747492313385, + 0.007838726043701172, + 0.008024613372981548, + 0.011495817452669144, + 0.007252361159771681, + 0.009058634750545025, + 0.010902857407927513, + 0.005454517900943756, + 0.005250066518783569, + 0.007265507243573666, + 0.005565396044403315, + 0.010740052908658981, + 0.007620719261467457, + 0.003997334744781256, + 0.01251138374209404, + 0.0038478849455714226, + 0.010533861815929413, + 0.012273475527763367, + 0.011493158526718616, + 0.008235438726842403, + 0.019592436030507088, + 0.007597481366246939, + 0.008500930853188038, + 0.005962964612990618, + 0.007336131762713194, + 0.010029051452875137, + 0.008644507266581059, + 0.004651546012610197, + 0.008117788471281528, + 0.006143562030047178, + 0.005801097955554724, + 0.006412902381271124, + 0.014300556853413582, + 0.021841401234269142, + 0.010455516166985035, + 0.0068765245378017426, + 0.019758066162467003, + 0.014009565114974976, + 0.014567030593752861, + 0.00789811834692955, + 0.021574852988123894, + 0.04095244035124779, + 0.009547232650220394, + 0.010879619978368282, + 0.024027559906244278, + 0.023308133706450462, + 0.01600789837539196, + 0.010967913083732128, + 0.012849492020905018, + 0.01324450597167015, + 0.006028635427355766, + 0.005612863693386316, + 0.011579232290387154, + 0.010981949977576733, + 0.0145407784730196, + 0.01030015666037798, + 0.008498264476656914, + 0.0037725206930190325, + 0.004655737429857254, + 0.007561168167740107, + 0.0035272231325507164, + 0.003483827691525221, + 0.011800402775406837, + 0.0038951679598540068, + 0.057381678372621536, + 0.01629175990819931, + 0.025237157940864563, + 0.011484171263873577, + 0.040764641016721725, + 0.02021843008697033, + 0.07148364931344986, + 0.015636930242180824, + 0.00795584637671709, + 0.006990831810981035, + 0.005630871746689081, + 0.004873339086771011, + 0.010751115158200264, + 0.023289816454052925, + 0.010221256874501705, + 0.00706839095801115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.000814582861494273 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, this extreme close-up, high-angle shot observes a perfect maraschino cherry falling into a vibrant, luminous cocktail. The camera remains static, fixed on the crystal-clear coupe glass where the glowing electric-cyan liquid awaits. The cherry, a shimmering sphere of deep candied crimson, its slender stem still attached, decisively shatters the tense surface tension. A delicate, crown-shaped spray erupts silently, sending microscopic, jewel-like droplets into the air. As the cherry gracefully sinks into the luminous depths, trailing a stream of tiny bubbles, a series of perfect concentric ripples spreads across the liquid's surface, capturing and distorting sharp, focused rays of light from an unseen source into mesmerizing wavy patterns. The color palette is a study in high-contrast beauty—the intense red of the fruit contrasts with the cool, translucent cyan of the drink, all set against a deeply out-of-focus background of warm golden bokeh lights, creating an exquisite and visually arresting atmosphere of pure sensual indulgence.", + "seed": 25942, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_4/index_3006_seed_25942_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05658134389668703, + "vlm_logprob": -0.05658134389668703, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.05658134389668703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009166774339973927 + ], + "opposite_logprob_tokens": [ + -5.134166717529297 + ], + "logprob_reward": -0.009166774339973927, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009166774339973927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05473528057336807, + 0.12350667268037796, + 0.06946053355932236, + 0.17883999645709991, + 0.057965733110904694, + 0.028813432902097702, + 0.017277861014008522, + 0.290204793214798, + 0.0829458087682724, + 0.09624991565942764 + ], + "credit_map": [ + 0.007256525103002787, + 0.009702225215733051, + 0.008696544915437698, + 0.0070488546043634415, + 0.006319948472082615, + 0.005118667148053646, + 0.005580739583820105, + 0.005011772271245718, + 0.0372147373855114, + 0.009408729150891304, + 0.04288924112915993, + 0.0066533503122627735, + 0.008707152679562569, + 0.006398123688995838, + 0.00592298386618495, + 0.006312355399131775, + 0.012615535408258438, + 0.009234999306499958, + 0.008655744604766369, + 0.005742340814322233, + 0.005602479446679354, + 0.00778586883097887, + 0.006149924360215664, + 0.01367364265024662, + 0.025183219462633133, + 0.025468584150075912, + 0.029112890362739563, + 0.014876637607812881, + 0.029705479741096497, + 0.01677103526890278, + 0.013542069122195244, + 0.024180078878998756, + 0.014332860708236694, + 0.005949320737272501, + 0.01776362955570221, + 0.004244444891810417, + 0.003933729603886604, + 0.004594586323946714, + 0.0028869747184216976, + 0.004260186571627855, + 0.004198871087282896, + 0.0030712871812283993, + 0.0031300154514610767, + 0.0030259271152317524, + 0.005170927848666906, + 0.002951832255348563, + 0.003113007638603449, + 0.004151566419750452, + 0.002052000258117914, + 0.001959677552804351, + 0.002342632971704006, + 0.0015228615375235677, + 0.001966043608263135, + 0.0024221448693424463, + 0.003138052299618721, + 0.0018744479166343808, + 0.03570154681801796, + 0.02806650660932064, + 0.044468529522418976, + 0.027590196579694748, + 0.03476911410689354, + 0.048450030386447906, + 0.03695308417081833, + 0.034205783158540726, + 0.0183373112231493, + 0.0072370488196611404, + 0.011439906433224678, + 0.007670125924050808, + 0.010247437283396721, + 0.011554354801774025, + 0.006958813406527042, + 0.009500815533101559, + 0.03606831654906273, + 0.008190066553652287, + 0.03414048254489899, + 0.0037915396969765425, + 0.002434720518067479, + 0.003510910551995039, + 0.005002639256417751, + 0.0031112416181713343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.009166774339973927 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994731664657593 + ], + "opposite_logprob_tokens": [ + -1.8049473762512207 + ], + "logprob_reward": -0.17994731664657593, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994731664657593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051745373755693436, + 0.1371963620185852, + 0.08876731991767883, + 0.10426194220781326, + 0.13104265928268433, + 0.16477681696414948, + 0.11981211602687836, + 0.07255567610263824, + 0.03479177877306938, + 0.09505000710487366 + ], + "credit_map": [ + 0.006918171886354685, + 0.009446205571293831, + 0.009213423356413841, + 0.0051703606732189655, + 0.005679852329194546, + 0.0045466055162250996, + 0.005381579976528883, + 0.005389171652495861, + 0.03830324113368988, + 0.013464260846376419, + 0.05423416569828987, + 0.006379889324307442, + 0.005876168608665466, + 0.005674526561051607, + 0.005447196774184704, + 0.007816888391971588, + 0.014986513182520866, + 0.018050989136099815, + 0.011262230575084686, + 0.006834703963249922, + 0.006821481045335531, + 0.009631066583096981, + 0.007481883745640516, + 0.013698452152311802, + 0.022165751084685326, + 0.016020312905311584, + 0.02119404822587967, + 0.008497059345245361, + 0.007830105721950531, + 0.013434330001473427, + 0.006071964744478464, + 0.009048365987837315, + 0.035318974405527115, + 0.014224044047296047, + 0.04193133860826492, + 0.012389825657010078, + 0.005768848117440939, + 0.0049695405177772045, + 0.007483788765966892, + 0.008956292644143105, + 0.02151561714708805, + 0.015576123259961605, + 0.028922419995069504, + 0.01817779615521431, + 0.017420949414372444, + 0.017909999936819077, + 0.01626916415989399, + 0.02898472361266613, + 0.013273284770548344, + 0.017354285344481468, + 0.017398279160261154, + 0.016656197607517242, + 0.012205924838781357, + 0.010688104666769505, + 0.0197345782071352, + 0.012501454912126064, + 0.008269421756267548, + 0.006785286124795675, + 0.009582406841218472, + 0.007284858264029026, + 0.009211715310811996, + 0.012501511722803116, + 0.009640989825129509, + 0.009279475547373295, + 0.004849742166697979, + 0.0033341911621391773, + 0.005400283727794886, + 0.0038867902476340532, + 0.004058114718645811, + 0.004200959112495184, + 0.0033511146903038025, + 0.005710578989237547, + 0.037858519703149796, + 0.005934365559369326, + 0.03635844960808754, + 0.0036666837986558676, + 0.0016651788027957082, + 0.0019227871671319008, + 0.00467258645221591, + 0.0029714247211813927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.17994731664657593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043101079761981964 + ], + "opposite_logprob_tokens": [ + -3.1681010723114014 + ], + "logprob_reward": -0.043101079761981964, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043101079761981964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0523189939558506, + 0.11516048014163971, + 0.2003331184387207, + 0.10832370072603226, + 0.05493276193737984, + 0.061540354043245316, + 0.04953274503350258, + 0.1595057100057602, + 0.0904296338558197, + 0.10792246460914612 + ], + "credit_map": [ + 0.007307171355932951, + 0.015577459707856178, + 0.00642942450940609, + 0.0035104649141430855, + 0.007804298307746649, + 0.004429236054420471, + 0.003305820981040597, + 0.00395512580871582, + 0.02968508191406727, + 0.013712742365896702, + 0.04391750693321228, + 0.0060716792941093445, + 0.005682644434273243, + 0.004508180543780327, + 0.005225406959652901, + 0.0063572595827281475, + 0.031856369227170944, + 0.04421361908316612, + 0.023905543610453606, + 0.017448125407099724, + 0.016695864498615265, + 0.02109850011765957, + 0.01542364526540041, + 0.029691483825445175, + 0.0160800963640213, + 0.01504505705088377, + 0.019683141261339188, + 0.008779870346188545, + 0.01513187400996685, + 0.013535297475755215, + 0.009288376197218895, + 0.010779996402561665, + 0.014574233442544937, + 0.0059648482128977776, + 0.016863690689206123, + 0.0038748413790017366, + 0.0034131971187889576, + 0.0038504204712808132, + 0.003219479927793145, + 0.003172055585309863, + 0.007120325695723295, + 0.005689026787877083, + 0.00927565898746252, + 0.006490675266832113, + 0.008062818087637424, + 0.007069530431181192, + 0.006373277399688959, + 0.011459041386842728, + 0.006414306350052357, + 0.006410334724932909, + 0.006557011976838112, + 0.004958042409271002, + 0.007196501828730106, + 0.006741964723914862, + 0.0064084529876708984, + 0.004846135154366493, + 0.02118484489619732, + 0.015387366525828838, + 0.02873951755464077, + 0.01610349304974079, + 0.013845904730260372, + 0.02350742742419243, + 0.021926939487457275, + 0.01881023310124874, + 0.021176494657993317, + 0.006848027929663658, + 0.012752566486597061, + 0.008134531788527966, + 0.010054226033389568, + 0.012652120552957058, + 0.009193174540996552, + 0.009618506766855717, + 0.04361366853117943, + 0.00638000201433897, + 0.040529269725084305, + 0.004308530129492283, + 0.0018024030141532421, + 0.0024529676884412766, + 0.005276271607726812, + 0.003559355391189456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.043101079761981964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632143795490265 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632143795490265, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632143795490265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020711401477456093, + 0.035150136798620224, + 0.08204269409179688, + 0.07483815401792526, + 0.03440811112523079, + 0.022289197891950607, + 0.009882986545562744, + 0.3516569137573242, + 0.0253978930413723, + 0.3436225354671478 + ], + "credit_map": [ + 0.0033215235453099012, + 0.004279835615307093, + 0.002265667077153921, + 0.0018794599454849958, + 0.0031317020766437054, + 0.0021565842907875776, + 0.0019230230245739222, + 0.0017536074155941606, + 0.008307776413857937, + 0.0037432145327329636, + 0.01054078433662653, + 0.0023864663671702147, + 0.003201135201379657, + 0.0017603576416149735, + 0.003153484547510743, + 0.0020569167099893093, + 0.01276091206818819, + 0.014354932121932507, + 0.010296839289367199, + 0.007954576052725315, + 0.007463898044079542, + 0.008073464035987854, + 0.007388972211629152, + 0.013749097473919392, + 0.012990790419280529, + 0.010668537579476833, + 0.01269400306046009, + 0.005656386259943247, + 0.00855272263288498, + 0.0077063473872840405, + 0.005388932302594185, + 0.01118044089525938, + 0.008779673837125301, + 0.0030229685362428427, + 0.010156983509659767, + 0.0023344880901277065, + 0.0026897923089563847, + 0.003491830313578248, + 0.0019257849780842662, + 0.0020065903663635254, + 0.0024599172174930573, + 0.0020025784615427256, + 0.0034816632978618145, + 0.0025369366630911827, + 0.00284551247023046, + 0.0028628134168684483, + 0.0017831148579716682, + 0.004316661041229963, + 0.0011651358800008893, + 0.001181902247481048, + 0.0017420523799955845, + 0.0008718509343452752, + 0.0014972352655604482, + 0.0011649526422843337, + 0.0013041144702583551, + 0.0009557430166751146, + 0.04118312895298004, + 0.03309160843491554, + 0.051122356206178665, + 0.03317714482545853, + 0.04503626748919487, + 0.0608217716217041, + 0.0461185947060585, + 0.041106030344963074, + 0.0037219764199107885, + 0.0021195742301642895, + 0.003669633297249675, + 0.0025044477079063654, + 0.003958775196224451, + 0.0035836424212902784, + 0.0027679610066115856, + 0.0030718822963535786, + 0.13807189464569092, + 0.020774351432919502, + 0.1323394775390625, + 0.013799222186207771, + 0.004767437465488911, + 0.006265092641115189, + 0.016425566747784615, + 0.011179463006556034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.007632143795490265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a maraschino cherry falling into a cocktail glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle high and does the camera remain static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the liquid in the coupe glass electric-cyan and crystal-clear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cherry shatter the surface tension, creating a crown-shaped spray of droplets and concentric ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deeply out-of-focus scene with warm golden bokeh lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305940493941307 + ], + "opposite_logprob_tokens": [ + -3.1680593490600586 + ], + "logprob_reward": -0.04305940493941307, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305940493941307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08455710113048553, + 0.051736533641815186, + 0.046921081840991974, + 0.06777356564998627, + 0.05109793320298195, + 0.01753481850028038, + 0.05972348526120186, + 0.08280470967292786, + 0.26873308420181274, + 0.2691177725791931 + ], + "credit_map": [ + 0.011544295586645603, + 0.016919830814003944, + 0.009770095348358154, + 0.007853587158024311, + 0.014129683375358582, + 0.009285219013690948, + 0.008146101608872414, + 0.0069082919508218765, + 0.013874667696654797, + 0.004614775534719229, + 0.01875525340437889, + 0.00269132643006742, + 0.0030870921909809113, + 0.0029911119490861893, + 0.002732030348852277, + 0.002990280743688345, + 0.010299653746187687, + 0.010452193208038807, + 0.005862787831574678, + 0.004193577915430069, + 0.0033844790887087584, + 0.0034861210733652115, + 0.003034171648323536, + 0.00620809942483902, + 0.010442203842103481, + 0.009898065589368343, + 0.016452111303806305, + 0.005916188005357981, + 0.00830355379730463, + 0.006040995474904776, + 0.004913401324301958, + 0.005807049572467804, + 0.012994544580578804, + 0.00581404147669673, + 0.015798844397068024, + 0.00427999859675765, + 0.0031442036852240562, + 0.0030514351092278957, + 0.002810192294418812, + 0.0032046714331954718, + 0.0031967705581337214, + 0.0014562905998900533, + 0.0020499241072684526, + 0.001547503750771284, + 0.003011611057445407, + 0.0018898418638855219, + 0.002176902024075389, + 0.0022059727925807238, + 0.00778585858643055, + 0.007912127301096916, + 0.012593449093401432, + 0.004547188524156809, + 0.0066089630126953125, + 0.008801610209047794, + 0.005815316457301378, + 0.005658973008394241, + 0.011974366381764412, + 0.007459667045623064, + 0.009840802289545536, + 0.007403086870908737, + 0.01103474386036396, + 0.014252200722694397, + 0.010489479638636112, + 0.010350363329052925, + 0.0619521364569664, + 0.022681856527924538, + 0.03555510565638542, + 0.019529907032847404, + 0.028635986149311066, + 0.04269612208008766, + 0.02214786596596241, + 0.035534076392650604, + 0.10792964696884155, + 0.01611458510160446, + 0.10329144448041916, + 0.01108417846262455, + 0.004070591647177935, + 0.005097824148833752, + 0.012605862691998482, + 0.008923638612031937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.04305940493941307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_5/index_tensor([8598, 8598])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_5/index_tensor([8598, 8598])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ee401a07bc433bca8460b16d7f5db23d51eedc66 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_5/index_tensor([8598, 8598])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle Steadicam tracking shot follows the heavy, determined steps of a Crusader knight as he journeys through an ancient forest shrouded in thick, gray fog. The knight, clad in a battle-worn suit of chain mail, wears a heavy, once-white tunic, now stained with dirt and dirt. A closed-faced greathelm, with only a narrow eye slit, completely obscures his features, leaving only an air of awe and mystery. One metal-gauntleted hand rests firmly on the hilt of the longsword at his waist, and with each step, the dull thud of metal greaves on slippery, decaying leaves echoes in the silence. All around, twisted, ancient branches loom in the nearly stagnant air, their shadows stretching and swaying silently and ghostly through the thick mist, growing ever more eerie as the knight approaches. As the knight's footsteps slowed in front of a small open space, the camera began to rise smoothly, passing over his shoulder, revealing the astonishing scene ahead with a grand, distant aerial perspective: the outline of a magnificent but dilapidated Gothic church emerged from the mist, its weathered gray stone walls covered with thick ivy, and moist dark green moss grew in the corners and cracks. The broken stained glass windows were like empty eye sockets, silently staring at the uninvited guest. The whole picture presented a desolate, sacred and suspenseful atmosphere under the gloomy, diffuse skylight.", + "seed": 10513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_5/index_8598_seed_10513_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018987872079014777, + "vlm_logprob": -0.018987872079014777, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.018987872079014777, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024800047278404236 + ], + "opposite_logprob_tokens": [ + -3.7748000621795654 + ], + "logprob_reward": -0.024800047278404236, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024800047278404236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11076109111309052, + 0.4880558252334595, + 0.0338536836206913, + 0.04451464116573334, + 0.03502916917204857, + 0.08683203905820847, + 0.04680843651294708, + 0.02672729082405567, + 0.083038330078125, + 0.04437943547964096 + ], + "credit_map": [ + 0.009213902056217194, + 0.01308187935501337, + 0.013898243196308613, + 0.01711985282599926, + 0.012625507079064846, + 0.016637694090604782, + 0.013427660800516605, + 0.014756347984075546, + 0.02398528717458248, + 0.011293538846075535, + 0.12569543719291687, + 0.01343280915170908, + 0.06114550307393074, + 0.19839346408843994, + 0.02181202545762062, + 0.032297804951667786, + 0.004081368446350098, + 0.0024362606927752495, + 0.005429900251328945, + 0.004197431728243828, + 0.004321799147874117, + 0.004053634125739336, + 0.003828679909929633, + 0.005504608154296875, + 0.0036556683480739594, + 0.004684672225266695, + 0.0051770294085145, + 0.006909240502864122, + 0.00491535197943449, + 0.006783594377338886, + 0.0068077510222792625, + 0.005581332370638847, + 0.004389207344502211, + 0.00344837736338377, + 0.004585448186844587, + 0.0034638186916708946, + 0.005090287886559963, + 0.007225578650832176, + 0.003764257999137044, + 0.003062192350625992, + 0.006375377997756004, + 0.008604022674262524, + 0.006264320574700832, + 0.010277092456817627, + 0.01600559987127781, + 0.0087432861328125, + 0.0172727108001709, + 0.013289635069668293, + 0.004115534480661154, + 0.0046501667238771915, + 0.00735563226044178, + 0.004044175148010254, + 0.004800619091838598, + 0.007521286141127348, + 0.008302071131765842, + 0.0060189515352249146, + 0.0038216051179915667, + 0.003880978561937809, + 0.0031487704254686832, + 0.0014802194200456142, + 0.003397679654881358, + 0.004933984484523535, + 0.0027101929299533367, + 0.003353858133777976, + 0.011759714223444462, + 0.012989955022931099, + 0.008651877753436565, + 0.009113138541579247, + 0.011629337444901466, + 0.009860117919743061, + 0.011225249618291855, + 0.007808942813426256, + 0.007149658631533384, + 0.005198481492698193, + 0.004919228609651327, + 0.004188497085124254, + 0.004172155167907476, + 0.0066045369021594524, + 0.005087194032967091, + 0.007059686351567507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.024800047278404236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307059198617935 + ], + "opposite_logprob_tokens": [ + -3.1680705547332764 + ], + "logprob_reward": -0.04307059198617935, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307059198617935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1402299404144287, + 0.10338189452886581, + 0.040549665689468384, + 0.04038146138191223, + 0.09399955719709396, + 0.11565756052732468, + 0.08162811398506165, + 0.1760309934616089, + 0.1573781967163086, + 0.050762616097927094 + ], + "credit_map": [ + 0.011517834849655628, + 0.013015839271247387, + 0.01851811446249485, + 0.016675099730491638, + 0.017276350408792496, + 0.01742314174771309, + 0.020538339391350746, + 0.025265224277973175, + 0.009611889719963074, + 0.005103684961795807, + 0.021130554378032684, + 0.00761146517470479, + 0.014078501611948013, + 0.026607705280184746, + 0.009766235947608948, + 0.009471859782934189, + 0.004439061973243952, + 0.003167329588904977, + 0.006770303472876549, + 0.004189180210232735, + 0.005611886270344257, + 0.00540149537846446, + 0.006113714538514614, + 0.004856690764427185, + 0.007629137020558119, + 0.004674945026636124, + 0.0053612724877893925, + 0.004117892123758793, + 0.0044565098360180855, + 0.00729539105668664, + 0.00337024824693799, + 0.003476069075986743, + 0.015049184672534466, + 0.007981153205037117, + 0.011492508463561535, + 0.008296838961541653, + 0.016500089317560196, + 0.014446382410824299, + 0.011892526410520077, + 0.00834087748080492, + 0.011882293038070202, + 0.016190825030207634, + 0.012130370363593102, + 0.009715049527585506, + 0.01860743761062622, + 0.012110605835914612, + 0.020411068573594093, + 0.014609905891120434, + 0.008644559420645237, + 0.009978152811527252, + 0.013949253596365452, + 0.006594124715775251, + 0.0076105790212750435, + 0.014571753330528736, + 0.011812623590230942, + 0.008467067964375019, + 0.023290742188692093, + 0.02115442603826523, + 0.03414620831608772, + 0.010985823348164558, + 0.01592429354786873, + 0.023736534640192986, + 0.021840715780854225, + 0.024952245876193047, + 0.025411784648895264, + 0.021153556182980537, + 0.01594030112028122, + 0.01315438374876976, + 0.025882098823785782, + 0.01800994575023651, + 0.021292252466082573, + 0.01653388887643814, + 0.007207788061350584, + 0.013238812796771526, + 0.0047539216466248035, + 0.0032849598210304976, + 0.006000209599733353, + 0.006209735292941332, + 0.004387082997709513, + 0.005680109839886427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.04307059198617935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250236950814724 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250236950814724, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250236950814724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15093201398849487, + 0.46030759811401367, + 0.043749790638685226, + 0.06409820169210434, + 0.04530331492424011, + 0.032466720789670944, + 0.03130133077502251, + 0.09286968410015106, + 0.050320301204919815, + 0.02865106239914894 + ], + "credit_map": [ + 0.015147676691412926, + 0.00961498823016882, + 0.019242698326706886, + 0.012316906824707985, + 0.031091490760445595, + 0.01994713954627514, + 0.017701169475913048, + 0.025869928300380707, + 0.025389546528458595, + 0.01190213207155466, + 0.12558548152446747, + 0.013686208054423332, + 0.048733972012996674, + 0.1778269112110138, + 0.024274330586194992, + 0.03290894627571106, + 0.007266295608133078, + 0.002710716100409627, + 0.007509177550673485, + 0.004496933426707983, + 0.0063232011161744595, + 0.005544200073927641, + 0.004679057281464338, + 0.00522020785138011, + 0.005222542677074671, + 0.0052148569375276566, + 0.011960524134337902, + 0.006444698665291071, + 0.005767177790403366, + 0.012501884251832962, + 0.00845478568226099, + 0.008531726896762848, + 0.0062447525560855865, + 0.00438845856115222, + 0.0052845655009150505, + 0.004832522012293339, + 0.00698952516540885, + 0.007852228358387947, + 0.005424800328910351, + 0.004286462906748056, + 0.005014521535485983, + 0.005432560108602047, + 0.00446331687271595, + 0.0030525000765919685, + 0.003639878472313285, + 0.003556295996531844, + 0.003249258501455188, + 0.004058385733515024, + 0.0032568792812526226, + 0.0034796646796166897, + 0.005007875617593527, + 0.002731261309236288, + 0.004774400964379311, + 0.005339834373444319, + 0.0038802637718617916, + 0.002831148449331522, + 0.010902421548962593, + 0.011475720442831516, + 0.015141315758228302, + 0.005489688832312822, + 0.010282404720783234, + 0.016024116426706314, + 0.011672209948301315, + 0.011881791055202484, + 0.009398693218827248, + 0.007358422502875328, + 0.004949078429490328, + 0.005003366153687239, + 0.006565845105797052, + 0.005728024989366531, + 0.006273613777011633, + 0.0050432500429451466, + 0.0055949492380023, + 0.003995981067419052, + 0.0036961499135941267, + 0.002450876636430621, + 0.002432398498058319, + 0.0029993613716214895, + 0.004918983206152916, + 0.0025623603723943233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.005250236950814724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076280031353235245 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.0076280031353235245, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076280031353235245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04818624630570412, + 0.5053231716156006, + 0.02550087869167328, + 0.03061641938984394, + 0.06026064604520798, + 0.04577614367008209, + 0.06227470561861992, + 0.09185183048248291, + 0.06216274946928024, + 0.0680471658706665 + ], + "credit_map": [ + 0.0035848496481776237, + 0.0038472407031804323, + 0.007644387427717447, + 0.006365940906107426, + 0.006121591664850712, + 0.00847491156309843, + 0.006554750259965658, + 0.005592579487711191, + 0.024727152660489082, + 0.011978124268352985, + 0.12957102060317993, + 0.014629299752414227, + 0.06403301656246185, + 0.2052590250968933, + 0.022376786917448044, + 0.0327487513422966, + 0.002937203971669078, + 0.0027436644304543734, + 0.003385493066161871, + 0.0034431542735546827, + 0.0033205619547516108, + 0.002983096055686474, + 0.003792291507124901, + 0.0028954180888831615, + 0.003555845934897661, + 0.003392596961930394, + 0.0034807620104402304, + 0.005891183391213417, + 0.003303915960714221, + 0.0034107498358935118, + 0.004336252808570862, + 0.0032451138831675053, + 0.00891879852861166, + 0.004841435234993696, + 0.007703418843448162, + 0.00560676958411932, + 0.00980339478701353, + 0.011686455458402634, + 0.006564516574144363, + 0.005135861691087484, + 0.00548674026504159, + 0.007352111861109734, + 0.008884047158062458, + 0.004552822094410658, + 0.0048419395461678505, + 0.005408419296145439, + 0.004671155940741301, + 0.004578906577080488, + 0.0078072878532111645, + 0.010048183612525463, + 0.008861037902534008, + 0.00491086021065712, + 0.007094063796103001, + 0.008465022780001163, + 0.008685074746608734, + 0.006403180304914713, + 0.013047067448496819, + 0.010410922579467297, + 0.01343055721372366, + 0.00554789649322629, + 0.009899674914777279, + 0.016269033774733543, + 0.010678193531930447, + 0.012568498030304909, + 0.010852205567061901, + 0.0101412208750844, + 0.00625668466091156, + 0.006204855162650347, + 0.008707066997885704, + 0.009206126444041729, + 0.004550979007035494, + 0.006243613548576832, + 0.013341406360268593, + 0.008428978733718395, + 0.008187659084796906, + 0.005320819094777107, + 0.006773662753403187, + 0.01007370837032795, + 0.007665562909096479, + 0.008255371823906898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.0076280031353235245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190481044352055 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190481044352055, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190481044352055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09531091898679733, + 0.404142290353775, + 0.03885892778635025, + 0.027909161522984505, + 0.06847146898508072, + 0.042039040476083755, + 0.06296256184577942, + 0.07736804336309433, + 0.04425764083862305, + 0.13867993652820587 + ], + "credit_map": [ + 0.008584150113165379, + 0.00927350390702486, + 0.012780719436705112, + 0.010485481470823288, + 0.016367275267839432, + 0.01929624378681183, + 0.008621643297374249, + 0.009901901707053185, + 0.021041493862867355, + 0.009893964044749737, + 0.10616400837898254, + 0.012128057889640331, + 0.04670204967260361, + 0.1600942760705948, + 0.020034436136484146, + 0.02808402292430401, + 0.003880330827087164, + 0.0037623008247464895, + 0.006034898571670055, + 0.004120812751352787, + 0.005299529060721397, + 0.005525700282305479, + 0.005777907092124224, + 0.004457450937479734, + 0.003949707839637995, + 0.0022723625879734755, + 0.005648869089782238, + 0.0025730784982442856, + 0.002693713642656803, + 0.004310771357268095, + 0.003210651921108365, + 0.003250007750466466, + 0.010697301477193832, + 0.005963233299553394, + 0.007726537063717842, + 0.006865584757179022, + 0.012401708401739597, + 0.008375322446227074, + 0.00929772574454546, + 0.007144054397940636, + 0.00470076035708189, + 0.006949292030185461, + 0.006050328258424997, + 0.0034383798483759165, + 0.006163438316434622, + 0.004640017636120319, + 0.005463593173772097, + 0.0046332357451319695, + 0.006085013039410114, + 0.006177637726068497, + 0.011502232402563095, + 0.005780767649412155, + 0.0072775352746248245, + 0.010063236579298973, + 0.008507764898240566, + 0.007568382658064365, + 0.011042707599699497, + 0.009287277236580849, + 0.014076444320380688, + 0.005294901318848133, + 0.006803141441196203, + 0.010595086961984634, + 0.009031355381011963, + 0.011237134225666523, + 0.006744359154254198, + 0.0055039506405591965, + 0.0044149598106741905, + 0.006661093793809414, + 0.005852232687175274, + 0.00640775915235281, + 0.004440865479409695, + 0.004232421051710844, + 0.02143131010234356, + 0.023105844855308533, + 0.015458078123629093, + 0.012173598632216454, + 0.01212441734969616, + 0.014655286446213722, + 0.023750893771648407, + 0.015980524942278862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.014190481044352055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle Steadicam tracking shot follows the heavy, determined steps of a Crusader knight as he journeys through an ancient forest shrouded in thick, gray fog. The knight, clad in a battle-worn suit of chain mail, wears a heavy, once-white tunic, now stained with dirt and dirt. A closed-faced greathelm, with only a narrow eye slit, completely obscures his features, leaving only an air of awe and mystery. One metal-gauntleted hand rests firmly on the hilt of the longsword at his waist, and with each step, the dull thud of metal greaves on slippery, decaying leaves echoes in the silence. All around, twisted, ancient branches loom in the nearly stagnant air, their shadows stretching and swaying silently and ghostly through the thick mist, growing ever more eerie as the knight approaches. As the knight's footsteps slowed in front of a small open space, the camera began to rise smoothly, passing over his shoulder, revealing the astonishing scene ahead with a grand, distant aerial perspective: the outline of a magnificent but dilapidated Gothic church emerged from the mist, its weathered gray stone walls covered with thick ivy, and moist dark green moss grew in the corners and cracks. The broken stained glass windows were like empty eye sockets, silently staring at the uninvited guest. The whole picture presented a desolate, sacred and suspenseful atmosphere under the gloomy, diffuse skylight.", + "seed": 10514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_5/index_8598_seed_10514_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018205211823806168, + "vlm_logprob": -0.018205211823806168, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.018205211823806168, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017602181062102318 + ], + "opposite_logprob_tokens": [ + -4.14260196685791 + ], + "logprob_reward": -0.017602181062102318, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017602181062102318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14756914973258972, + 0.07750669866800308, + 0.032134223729372025, + 0.21402844786643982, + 0.058507710695266724, + 0.032654810696840286, + 0.07988771051168442, + 0.12797924876213074, + 0.06366313248872757, + 0.16606886684894562 + ], + "credit_map": [ + 0.02457786165177822, + 0.011177081614732742, + 0.031825125217437744, + 0.014285469427704811, + 0.015057449229061604, + 0.017671816051006317, + 0.013830858282744884, + 0.0191434808075428, + 0.013726996257901192, + 0.008831145241856575, + 0.014257998205721378, + 0.006869107019156218, + 0.006592349614948034, + 0.005926946178078651, + 0.00727041345089674, + 0.014031745493412018, + 0.005173304583877325, + 0.003366410266608, + 0.003729581832885742, + 0.0028244706336408854, + 0.003239946672692895, + 0.005992348771542311, + 0.0043131085112690926, + 0.0034950529225170612, + 0.011431503109633923, + 0.019805684685707092, + 0.038409918546676636, + 0.01682315580546856, + 0.03154345229268074, + 0.044399410486221313, + 0.0307366531342268, + 0.02087867446243763, + 0.005740471184253693, + 0.009507660754024982, + 0.008609461598098278, + 0.006373583804816008, + 0.005937471520155668, + 0.009275523945689201, + 0.005802026018500328, + 0.007261514198035002, + 0.0054777502082288265, + 0.0052445498295128345, + 0.004684313200414181, + 0.0023628671187907457, + 0.0024143161717802286, + 0.005000393372029066, + 0.003602780168876052, + 0.003867844119668007, + 0.009877964854240417, + 0.008572815917432308, + 0.009727396070957184, + 0.011971363797783852, + 0.008551941253244877, + 0.010691549628973007, + 0.005189911928027868, + 0.015304770320653915, + 0.01711374707520008, + 0.012681880965828896, + 0.016294820234179497, + 0.01553208939731121, + 0.018914302811026573, + 0.016850290820002556, + 0.013955903239548206, + 0.016636205837130547, + 0.01176829170435667, + 0.006623118184506893, + 0.009820613078773022, + 0.011226541362702847, + 0.006937425117939711, + 0.004426428582519293, + 0.008743666112422943, + 0.00411705207079649, + 0.03735056519508362, + 0.010763341560959816, + 0.05053647980093956, + 0.011826066300272942, + 0.01907014288008213, + 0.009671843610703945, + 0.016222355887293816, + 0.010628065094351768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.017602181062102318 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308440908789635 + ], + "opposite_logprob_tokens": [ + -3.1680843830108643 + ], + "logprob_reward": -0.04308440908789635, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308440908789635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13696877658367157, + 0.04608141630887985, + 0.05086192861199379, + 0.32667407393455505, + 0.04505078122019768, + 0.02122669667005539, + 0.09264558553695679, + 0.057565443217754364, + 0.048683393746614456, + 0.17424196004867554 + ], + "credit_map": [ + 0.0308303851634264, + 0.009589064866304398, + 0.027418728917837143, + 0.011541206389665604, + 0.011850239709019661, + 0.018762165680527687, + 0.01018019299954176, + 0.016796762123703957, + 0.0045096189714968204, + 0.005763501860201359, + 0.007290075067430735, + 0.007368511985987425, + 0.003265770385041833, + 0.0049446579068899155, + 0.005877933464944363, + 0.007061340846121311, + 0.010288674384355545, + 0.006540444679558277, + 0.008290693163871765, + 0.0038577180821448565, + 0.003922437317669392, + 0.007179236505180597, + 0.004498189780861139, + 0.006284522823989391, + 0.021484367549419403, + 0.03789837285876274, + 0.048298805952072144, + 0.027543485164642334, + 0.06130926311016083, + 0.053291089832782745, + 0.04137853905558586, + 0.03547012433409691, + 0.003953616600483656, + 0.006775017827749252, + 0.006641841493546963, + 0.00504099577665329, + 0.004137013573199511, + 0.008295955136418343, + 0.004788395948708057, + 0.005417936015874147, + 0.0033101483713835478, + 0.002742633456364274, + 0.003150535747408867, + 0.0020796353928744793, + 0.002195842331275344, + 0.0028258413076400757, + 0.00269385171122849, + 0.002228206256404519, + 0.014392325654625893, + 0.009833765216171741, + 0.011434008367359638, + 0.008966382592916489, + 0.0074919769540429115, + 0.012671416625380516, + 0.006393409334123135, + 0.02146228216588497, + 0.00867453496903181, + 0.006027656607329845, + 0.009419526904821396, + 0.00834285095334053, + 0.0067609986290335655, + 0.007288603577762842, + 0.0053310091607272625, + 0.005720255896449089, + 0.0063193244859576225, + 0.005538205150514841, + 0.008048498071730137, + 0.007137542124837637, + 0.005674934014678001, + 0.004304114263504744, + 0.0066481526009738445, + 0.005012613255530596, + 0.022353220731019974, + 0.020764732733368874, + 0.017634516581892967, + 0.020747212693095207, + 0.03776247426867485, + 0.016526496037840843, + 0.024423088878393173, + 0.014030199497938156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.04308440908789635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031947072129696608 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031947072129696608, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031947072129696608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12332045286893845, + 0.08207256346940994, + 0.03653716295957565, + 0.041078150272369385, + 0.08001381158828735, + 0.07061025500297546, + 0.20590204000473022, + 0.07463817298412323, + 0.07693763077259064, + 0.20888973772525787 + ], + "credit_map": [ + 0.019241539761424065, + 0.007078776601701975, + 0.0343623012304306, + 0.008192733861505985, + 0.008111400529742241, + 0.02007758431136608, + 0.008478151634335518, + 0.01777796447277069, + 0.008488951250910759, + 0.008544557727873325, + 0.01667020283639431, + 0.007848015055060387, + 0.008430283516645432, + 0.008083955384790897, + 0.012946296483278275, + 0.011060303077101707, + 0.00587916886433959, + 0.003641255432739854, + 0.004738639108836651, + 0.0029993609059602022, + 0.002985109109431505, + 0.006997585762292147, + 0.005246305372565985, + 0.004049740731716156, + 0.0030183412600308657, + 0.002908766735345125, + 0.009585182182490826, + 0.002480092691257596, + 0.0034729070030152798, + 0.009695444256067276, + 0.0032385815866291523, + 0.006678836885839701, + 0.0069288210943341255, + 0.007976090535521507, + 0.012321078218519688, + 0.00996343418955803, + 0.007084634620696306, + 0.01658540405333042, + 0.007498233113437891, + 0.011656117625534534, + 0.008387064561247826, + 0.009501119144260883, + 0.009469123557209969, + 0.009157729335129261, + 0.006143850274384022, + 0.0072935838252305984, + 0.014160782098770142, + 0.0064970036037266254, + 0.026193024590611458, + 0.03411486744880676, + 0.02322586625814438, + 0.0418873131275177, + 0.02503451146185398, + 0.019595209509134293, + 0.015242836438119411, + 0.020608410239219666, + 0.009109104052186012, + 0.007895145565271378, + 0.009973599575459957, + 0.010460407473146915, + 0.007703515235334635, + 0.007464572321623564, + 0.00890628807246685, + 0.013125542551279068, + 0.011995524168014526, + 0.00941054429858923, + 0.013817633502185345, + 0.012469030916690826, + 0.007930620573461056, + 0.007873425260186195, + 0.007657150737941265, + 0.005783706903457642, + 0.02417323924601078, + 0.020955590531229973, + 0.020232297480106354, + 0.03694898262619972, + 0.037031855434179306, + 0.021449780091643333, + 0.02685236744582653, + 0.02124563604593277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0031947072129696608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076981201767921 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076981201767921, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076981201767921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045002564787864685, + 0.024007711559534073, + 0.03141886740922928, + 0.02772655338048935, + 0.025549916550517082, + 0.02589818648993969, + 0.07257463783025742, + 0.16238857805728912, + 0.04504626616835594, + 0.5403867959976196 + ], + "credit_map": [ + 0.00603449484333396, + 0.0038130253087729216, + 0.011487887240946293, + 0.004019978456199169, + 0.0037935152649879456, + 0.0056690373457968235, + 0.004142136313021183, + 0.006042484659701586, + 0.0028871572576463223, + 0.0023487829603254795, + 0.005064385011792183, + 0.002089842688292265, + 0.0027260405477136374, + 0.002569462638348341, + 0.0031962450593709946, + 0.0031257933005690575, + 0.004119229502975941, + 0.0033817938528954983, + 0.003963987808674574, + 0.00283749564550817, + 0.0031840773299336433, + 0.005817596800625324, + 0.004555217921733856, + 0.003559466917067766, + 0.0024841607082635164, + 0.0017563003348186612, + 0.007533610332757235, + 0.0023226665798574686, + 0.0017564453883096576, + 0.0071007744409143925, + 0.0025384058244526386, + 0.0022341860458254814, + 0.002516059437766671, + 0.0029306767974048853, + 0.0031372427474707365, + 0.0037573520094156265, + 0.0030876731034368277, + 0.0038330184761434793, + 0.0029040598310530186, + 0.003383832750841975, + 0.0032001498620957136, + 0.0035634334199130535, + 0.0031026501674205065, + 0.004140016157180071, + 0.0035162505228072405, + 0.0026087325531989336, + 0.0034733491484075785, + 0.0022936013992875814, + 0.006493693683296442, + 0.010015360079705715, + 0.008230019360780716, + 0.00899326428771019, + 0.012575594708323479, + 0.010178855620324612, + 0.005938319955021143, + 0.010149521753191948, + 0.017429659143090248, + 0.016734041273593903, + 0.01749391108751297, + 0.02467767894268036, + 0.0320773720741272, + 0.014254213310778141, + 0.024640819057822227, + 0.015080871991813183, + 0.0051905447617173195, + 0.004919133149087429, + 0.005722622387111187, + 0.0055920700542628765, + 0.006121924612671137, + 0.005404396448284388, + 0.007942466996610165, + 0.004153103101998568, + 0.1083369180560112, + 0.04019773378968239, + 0.15816859900951385, + 0.03842220455408096, + 0.06607100367546082, + 0.04013201966881752, + 0.049840182065963745, + 0.03921806812286377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.011076981201767921 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067780554294586 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067780554294586, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067780554294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049732524901628494, + 0.03323351964354515, + 0.01792699098587036, + 0.036100342869758606, + 0.03594779968261719, + 0.02161087840795517, + 0.05424277484416962, + 0.09417504072189331, + 0.05811679735779762, + 0.598913311958313 + ], + "credit_map": [ + 0.007721033878624439, + 0.0028968092519789934, + 0.008644425310194492, + 0.004474838264286518, + 0.005946198012679815, + 0.00827815756201744, + 0.005067525897175074, + 0.006703533697873354, + 0.006998990196734667, + 0.0029752051923424006, + 0.0060309674590826035, + 0.0026040379889309406, + 0.0026358950417488813, + 0.002981043653562665, + 0.0034585753455758095, + 0.0055488040670752525, + 0.0035086111165583134, + 0.0018482469022274017, + 0.002595338737592101, + 0.0013961022486910224, + 0.0015881122089922428, + 0.0035404772497713566, + 0.001977067207917571, + 0.0014730350812897086, + 0.002953367307782173, + 0.0024049158673733473, + 0.00857106875628233, + 0.0023872731253504753, + 0.0029883235692977905, + 0.00913233496248722, + 0.003099019406363368, + 0.004564042668789625, + 0.0030380678363144398, + 0.00429577799513936, + 0.00588727043941617, + 0.005284660030156374, + 0.0036259519401937723, + 0.0058431741781532764, + 0.003838619217276573, + 0.004134276416152716, + 0.0030031928326934576, + 0.0028718437533825636, + 0.002869756892323494, + 0.0019951730500906706, + 0.0019490879494696856, + 0.003211302449926734, + 0.003261844627559185, + 0.0024486787151545286, + 0.004554221872240305, + 0.005544484592974186, + 0.006495253182947636, + 0.004975057207047939, + 0.010637125000357628, + 0.006311499513685703, + 0.004593406803905964, + 0.011131725274026394, + 0.013020923361182213, + 0.010315748862922192, + 0.008762006647884846, + 0.011674691922962666, + 0.015244510024785995, + 0.011586577631533146, + 0.010863320901989937, + 0.012707271613180637, + 0.012480945326387882, + 0.004909350536763668, + 0.007845212705433369, + 0.010091936215758324, + 0.0076908268965780735, + 0.003929222468286753, + 0.007018019910901785, + 0.004151283763349056, + 0.08001553267240524, + 0.06633011251688004, + 0.08536670356988907, + 0.06679688394069672, + 0.09086572378873825, + 0.06964582949876785, + 0.08685465902090073, + 0.05303792282938957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.016067780554294586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle Steadicam tracking shot follows the heavy, determined steps of a Crusader knight as he journeys through an ancient forest shrouded in thick, gray fog. The knight, clad in a battle-worn suit of chain mail, wears a heavy, once-white tunic, now stained with dirt and dirt. A closed-faced greathelm, with only a narrow eye slit, completely obscures his features, leaving only an air of awe and mystery. One metal-gauntleted hand rests firmly on the hilt of the longsword at his waist, and with each step, the dull thud of metal greaves on slippery, decaying leaves echoes in the silence. All around, twisted, ancient branches loom in the nearly stagnant air, their shadows stretching and swaying silently and ghostly through the thick mist, growing ever more eerie as the knight approaches. As the knight's footsteps slowed in front of a small open space, the camera began to rise smoothly, passing over his shoulder, revealing the astonishing scene ahead with a grand, distant aerial perspective: the outline of a magnificent but dilapidated Gothic church emerged from the mist, its weathered gray stone walls covered with thick ivy, and moist dark green moss grew in the corners and cracks. The broken stained glass windows were like empty eye sockets, silently staring at the uninvited guest. The whole picture presented a desolate, sacred and suspenseful atmosphere under the gloomy, diffuse skylight.", + "seed": 10515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_5/index_8598_seed_10515_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023393551353365183, + "vlm_logprob": -0.023393551353365183, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.023393551353365183, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0392831452190876 + ], + "opposite_logprob_tokens": [ + -3.289283037185669 + ], + "logprob_reward": -0.0392831452190876, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0392831452190876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30773526430130005, + 0.3030284345149994, + 0.07692813128232956, + 0.040641654282808304, + 0.03928198292851448, + 0.0241786427795887, + 0.08038868010044098, + 0.03406694531440735, + 0.0633106455206871, + 0.03043961152434349 + ], + "credit_map": [ + 0.03417707234621048, + 0.11215666681528091, + 0.020078804343938828, + 0.01767880655825138, + 0.01863652653992176, + 0.04083940386772156, + 0.014100202359259129, + 0.05006777495145798, + 0.05836008861660957, + 0.01734602451324463, + 0.021028393879532814, + 0.012978085316717625, + 0.017112277448177338, + 0.031508855521678925, + 0.03680020570755005, + 0.10789450258016586, + 0.006057985592633486, + 0.0073329792357981205, + 0.011270644143223763, + 0.0073154703713953495, + 0.010051022283732891, + 0.010943678207695484, + 0.011265342123806477, + 0.012691009789705276, + 0.0040619769133627415, + 0.003995169885456562, + 0.005907057784497738, + 0.003945621661841869, + 0.0034925024956464767, + 0.006880610715597868, + 0.005618201103061438, + 0.006740516517311335, + 0.0037128508556634188, + 0.006596769671887159, + 0.00516597181558609, + 0.005258453078567982, + 0.004905736073851585, + 0.0038130730390548706, + 0.005047548096626997, + 0.004781579133123159, + 0.0014288370730355382, + 0.0032986903097480536, + 0.002599910367280245, + 0.0036302548833191395, + 0.002602531109005213, + 0.005814507603645325, + 0.0019579597283154726, + 0.002845952520146966, + 0.006952318362891674, + 0.013807821087539196, + 0.010784873738884926, + 0.012739837169647217, + 0.006094551645219326, + 0.0076746586710214615, + 0.010046638548374176, + 0.012287979014217854, + 0.003011709311977029, + 0.005330242216587067, + 0.004574292805045843, + 0.00540688494220376, + 0.0042694611474871635, + 0.003610449144616723, + 0.0037908463273197412, + 0.004073061980307102, + 0.006685701664537191, + 0.011117778718471527, + 0.009227202273905277, + 0.009321672841906548, + 0.0047362204641103745, + 0.009403838776051998, + 0.005969557911157608, + 0.006848672404885292, + 0.0023887669667601585, + 0.004103878512978554, + 0.00455032754689455, + 0.004461886361241341, + 0.0014369586715474725, + 0.004874893929809332, + 0.003234798088669777, + 0.005388099234551191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0392831452190876 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866863042116165 + ], + "opposite_logprob_tokens": [ + -3.048668622970581 + ], + "logprob_reward": -0.04866863042116165, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866863042116165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23521286249160767, + 0.29755812883377075, + 0.0629449337720871, + 0.23085224628448486, + 0.029776252806186676, + 0.026104070246219635, + 0.028988556936383247, + 0.04087011516094208, + 0.03347477689385414, + 0.014218060299754143 + ], + "credit_map": [ + 0.022424258291721344, + 0.07666391134262085, + 0.018641695380210876, + 0.014094080775976181, + 0.012475184164941311, + 0.03248375654220581, + 0.011677124537527561, + 0.04675285518169403, + 0.05811626836657524, + 0.016741741448640823, + 0.020449064671993256, + 0.01242871768772602, + 0.015482082962989807, + 0.031009221449494362, + 0.03632164001464844, + 0.1070093885064125, + 0.005832909140735865, + 0.005628520157188177, + 0.011099589988589287, + 0.0060751307755708694, + 0.00665235286578536, + 0.009515216574072838, + 0.008983364328742027, + 0.009157847613096237, + 0.020900677889585495, + 0.023235833272337914, + 0.04402069374918938, + 0.021282536908984184, + 0.033326953649520874, + 0.021816570311784744, + 0.025950219482183456, + 0.040318772196769714, + 0.0025444079656153917, + 0.002970946254208684, + 0.0047664521262049675, + 0.004716312978416681, + 0.0020982439164072275, + 0.004322187975049019, + 0.0037641911767423153, + 0.0045935106463730335, + 0.0017585511086508632, + 0.0024073580279946327, + 0.003540415083989501, + 0.0036743339151144028, + 0.0032531965989619493, + 0.005969040095806122, + 0.002055070362985134, + 0.0034461040049791336, + 0.002403367543593049, + 0.0038431205321103334, + 0.003029269864782691, + 0.0022289431653916836, + 0.004218528512865305, + 0.006143445149064064, + 0.002771456027403474, + 0.004350427538156509, + 0.004085598513484001, + 0.007766058202832937, + 0.005768152419477701, + 0.004676636774092913, + 0.004813197068870068, + 0.004120748955756426, + 0.004170787986367941, + 0.005468934308737516, + 0.002875519683584571, + 0.004201446659862995, + 0.0058974516578018665, + 0.005660353694111109, + 0.0031319442205131054, + 0.004280587192624807, + 0.00323775433935225, + 0.004189721308648586, + 0.0012991202529519796, + 0.0018257354386150837, + 0.002339582657441497, + 0.0021517269778996706, + 0.0009880539728328586, + 0.0019800886511802673, + 0.0017864545807242393, + 0.0018472981173545122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04866863042116165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096213262528181 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096213262528181, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096213262528181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1704261749982834, + 0.31015995144844055, + 0.061376795172691345, + 0.07106103748083115, + 0.03564438596367836, + 0.04795728623867035, + 0.04593227803707123, + 0.03301525115966797, + 0.17557881772518158, + 0.048847999423742294 + ], + "credit_map": [ + 0.03493398055434227, + 0.05116792395710945, + 0.021299155429005623, + 0.007082527037709951, + 0.010164730250835419, + 0.023602915927767754, + 0.009427735581994057, + 0.012747213244438171, + 0.05949477106332779, + 0.01684168167412281, + 0.021406112238764763, + 0.012763596139848232, + 0.01675337366759777, + 0.0324285551905632, + 0.03812190890312195, + 0.11234994232654572, + 0.00563857052475214, + 0.005215900484472513, + 0.007767993025481701, + 0.0038327190559357405, + 0.006180501077324152, + 0.012910811230540276, + 0.010977262631058693, + 0.00885303970426321, + 0.005861884448677301, + 0.0036055089440196753, + 0.012668263167142868, + 0.004101383965462446, + 0.006660697981715202, + 0.008989677764475346, + 0.01314431894570589, + 0.01602930761873722, + 0.003153289668262005, + 0.0029373113065958023, + 0.004420748446136713, + 0.003908176906406879, + 0.0031654764898121357, + 0.007100710179656744, + 0.0061189644038677216, + 0.0048397076316177845, + 0.003326461184769869, + 0.003400983288884163, + 0.006087254732847214, + 0.008350335992872715, + 0.0058325729332864285, + 0.011235148645937443, + 0.003338439157232642, + 0.006386090070009232, + 0.0023714785929769278, + 0.0077845207415521145, + 0.0076047456823289394, + 0.004174346569925547, + 0.005052230786532164, + 0.00690752686932683, + 0.003144310787320137, + 0.008893120102584362, + 0.0022099779453128576, + 0.005999600514769554, + 0.00504743866622448, + 0.004397653974592686, + 0.002758400747552514, + 0.0024659852497279644, + 0.0028498622123152018, + 0.007286332082003355, + 0.020172884687781334, + 0.024351978674530983, + 0.038140106946229935, + 0.0228042583912611, + 0.014630192890763283, + 0.025179563090205193, + 0.013060920871794224, + 0.0172389168292284, + 0.004051961470395327, + 0.004887104965746403, + 0.008989688009023666, + 0.005695491563528776, + 0.0033446780871599913, + 0.00617561861872673, + 0.004996655974537134, + 0.010706800036132336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.004096213262528181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067379651591181755 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.0067379651591181755, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067379651591181755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3031962811946869, + 0.2264663279056549, + 0.04426690936088562, + 0.17166082561016083, + 0.027069464325904846, + 0.024950670078396797, + 0.022696295753121376, + 0.07622030377388, + 0.05434906482696533, + 0.049123868346214294 + ], + "credit_map": [ + 0.04245370626449585, + 0.07065493613481522, + 0.024577472358942032, + 0.019556833431124687, + 0.0163210928440094, + 0.03459757938981056, + 0.01933104172348976, + 0.07570361346006393, + 0.04317948594689369, + 0.012255808338522911, + 0.015426120720803738, + 0.009401594288647175, + 0.012487524189054966, + 0.02449130453169346, + 0.02774832583963871, + 0.08147615194320679, + 0.0037063495256006718, + 0.00504726730287075, + 0.007379853632301092, + 0.0038416872266680002, + 0.004022218286991119, + 0.00901306327432394, + 0.005806753411889076, + 0.005449716467410326, + 0.015235094353556633, + 0.01562356948852539, + 0.03652461618185043, + 0.014564857818186283, + 0.02182168886065483, + 0.01204814575612545, + 0.018374048173427582, + 0.037468794733285904, + 0.0027469934429973364, + 0.0036904041189700365, + 0.0037351050414144993, + 0.0028617403004318476, + 0.002152334665879607, + 0.0045529380440711975, + 0.0031168211717158556, + 0.004213128704577684, + 0.001410592463798821, + 0.003047260921448469, + 0.00283733569085598, + 0.004329025279730558, + 0.0027985922060906887, + 0.005688514094799757, + 0.00208550738170743, + 0.0027538423892110586, + 0.0015841141575947404, + 0.0038950550369918346, + 0.00339168100617826, + 0.0023866568226367235, + 0.0017731345724314451, + 0.003546648658812046, + 0.001862540957517922, + 0.00425646360963583, + 0.008250762708485126, + 0.01202003937214613, + 0.011916027404367924, + 0.008612158708274364, + 0.010723172686994076, + 0.00910539086908102, + 0.007551551330834627, + 0.008041197434067726, + 0.005207669455558062, + 0.005730366334319115, + 0.01023753173649311, + 0.010333406738936901, + 0.003505379892885685, + 0.005980255082249641, + 0.005081078968942165, + 0.008273376151919365, + 0.00475891912356019, + 0.005875552073121071, + 0.00668229628354311, + 0.006567087490111589, + 0.004034587647765875, + 0.009976889006793499, + 0.004229237325489521, + 0.006999295204877853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0067379651591181755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181802704930305 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181802704930305, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181802704930305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35792747139930725, + 0.06155427172780037, + 0.03296773508191109, + 0.07968500256538391, + 0.03348865360021591, + 0.02214912511408329, + 0.049446601420640945, + 0.13698571920394897, + 0.15955181419849396, + 0.06624363362789154 + ], + "credit_map": [ + 0.06414022296667099, + 0.1358484923839569, + 0.020021315664052963, + 0.0168515145778656, + 0.026738453656435013, + 0.03996171057224274, + 0.01903734914958477, + 0.035328444093465805, + 0.005151239689439535, + 0.004976711701601744, + 0.010693655349314213, + 0.00621736329048872, + 0.005980103276669979, + 0.009219471365213394, + 0.00942953210324049, + 0.009886193089187145, + 0.0030073390807956457, + 0.004230722319334745, + 0.005663062911480665, + 0.0034738625399768353, + 0.0033783228136599064, + 0.004901831969618797, + 0.003850130829960108, + 0.004462461918592453, + 0.00869989674538374, + 0.007870089262723923, + 0.012854861095547676, + 0.005966388154774904, + 0.010207908228039742, + 0.007052237633615732, + 0.015625828877091408, + 0.011407792568206787, + 0.005834102164953947, + 0.0042568715289235115, + 0.0044336747378110886, + 0.005032554734498262, + 0.0025498562026768923, + 0.0037912919651716948, + 0.003417826257646084, + 0.0041724746115505695, + 0.001547009451314807, + 0.0024212419521063566, + 0.003019049996510148, + 0.0038733440451323986, + 0.0025510285049676895, + 0.0039625926874578, + 0.0021444223821163177, + 0.0026304370258003473, + 0.003148074494674802, + 0.005495345685631037, + 0.009834512136876583, + 0.004210583865642548, + 0.004577695857733488, + 0.009494831785559654, + 0.0036376346834003925, + 0.009047923609614372, + 0.013953381218016148, + 0.02192341350018978, + 0.021017638966441154, + 0.014922336675226688, + 0.018287092447280884, + 0.014105373993515968, + 0.016435764729976654, + 0.01634073257446289, + 0.01888454332947731, + 0.015120225958526134, + 0.029212795197963715, + 0.022201526910066605, + 0.011481938883662224, + 0.026890629902482033, + 0.01408703438937664, + 0.021673114970326424, + 0.009390516206622124, + 0.008000466972589493, + 0.0069078439846634865, + 0.009262938052415848, + 0.004540748428553343, + 0.012604501098394394, + 0.006220880430191755, + 0.009315737523138523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.018181802704930305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle Steadicam tracking shot follows the heavy, determined steps of a Crusader knight as he journeys through an ancient forest shrouded in thick, gray fog. The knight, clad in a battle-worn suit of chain mail, wears a heavy, once-white tunic, now stained with dirt and dirt. A closed-faced greathelm, with only a narrow eye slit, completely obscures his features, leaving only an air of awe and mystery. One metal-gauntleted hand rests firmly on the hilt of the longsword at his waist, and with each step, the dull thud of metal greaves on slippery, decaying leaves echoes in the silence. All around, twisted, ancient branches loom in the nearly stagnant air, their shadows stretching and swaying silently and ghostly through the thick mist, growing ever more eerie as the knight approaches. As the knight's footsteps slowed in front of a small open space, the camera began to rise smoothly, passing over his shoulder, revealing the astonishing scene ahead with a grand, distant aerial perspective: the outline of a magnificent but dilapidated Gothic church emerged from the mist, its weathered gray stone walls covered with thick ivy, and moist dark green moss grew in the corners and cracks. The broken stained glass windows were like empty eye sockets, silently staring at the uninvited guest. The whole picture presented a desolate, sacred and suspenseful atmosphere under the gloomy, diffuse skylight.", + "seed": 10516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_5/index_8598_seed_10516_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016715689981356262, + "vlm_logprob": -0.016715689981356262, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.016715689981356262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019903697073459625 + ], + "opposite_logprob_tokens": [ + -4.019903659820557 + ], + "logprob_reward": -0.019903697073459625, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019903697073459625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18872271478176117, + 0.12234587967395782, + 0.03974989429116249, + 0.0690351277589798, + 0.05112273246049881, + 0.14065705239772797, + 0.1089337021112442, + 0.10656878352165222, + 0.05904170870780945, + 0.11382238566875458 + ], + "credit_map": [ + 0.027408260852098465, + 0.019282149150967598, + 0.027358882129192352, + 0.028575122356414795, + 0.02007422409951687, + 0.019723093137145042, + 0.02611544169485569, + 0.020185556262731552, + 0.008690391667187214, + 0.00954382587224245, + 0.029744578525424004, + 0.009870658628642559, + 0.008153382688760757, + 0.014147930778563023, + 0.01053735800087452, + 0.03165775164961815, + 0.0039131781086325645, + 0.004826015792787075, + 0.006159692071378231, + 0.005118146538734436, + 0.005307532846927643, + 0.004712113179266453, + 0.00574424397200346, + 0.003968971315771341, + 0.003692208556458354, + 0.00333206239156425, + 0.017427343875169754, + 0.00637979619204998, + 0.006523992400616407, + 0.00867338664829731, + 0.010302762500941753, + 0.012703575193881989, + 0.004605785943567753, + 0.005927026271820068, + 0.009130225516855717, + 0.005401949863880873, + 0.006150910165160894, + 0.006652629002928734, + 0.006420735269784927, + 0.006833471357822418, + 0.013886653818190098, + 0.028310146182775497, + 0.01795199327170849, + 0.014622481539845467, + 0.0205455981194973, + 0.014339447021484375, + 0.021839667111635208, + 0.009161056950688362, + 0.018198100849986076, + 0.028576046228408813, + 0.014022632502019405, + 0.009152710437774658, + 0.01084764301776886, + 0.007566124200820923, + 0.012816331349313259, + 0.007754115387797356, + 0.01483773160725832, + 0.008742406964302063, + 0.019521016627550125, + 0.008589129894971848, + 0.018674038350582123, + 0.015487591736018658, + 0.00850039441138506, + 0.012216467410326004, + 0.007222368847578764, + 0.009485611692070961, + 0.009220041334629059, + 0.004856463987380266, + 0.009434336796402931, + 0.005576267838478088, + 0.005831134971231222, + 0.007415482308715582, + 0.009648078121244907, + 0.010794893838465214, + 0.013279269449412823, + 0.03929513320326805, + 0.009312223643064499, + 0.014239993877708912, + 0.008896496146917343, + 0.008356297388672829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.019903697073459625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307584464550018 + ], + "opposite_logprob_tokens": [ + -3.1680757999420166 + ], + "logprob_reward": -0.04307584464550018, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307584464550018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20789313316345215, + 0.06299276649951935, + 0.03321371600031853, + 0.061886053532361984, + 0.03565511479973793, + 0.22205804288387299, + 0.07757878303527832, + 0.07241801917552948, + 0.07584719359874725, + 0.15045711398124695 + ], + "credit_map": [ + 0.04583536088466644, + 0.031427182257175446, + 0.03267413750290871, + 0.02237902209162712, + 0.011817552149295807, + 0.019034957513213158, + 0.013375483453273773, + 0.031349435448646545, + 0.005266198888421059, + 0.004550180397927761, + 0.013954348862171173, + 0.008675741963088512, + 0.006451485678553581, + 0.0066559636034071445, + 0.00628182478249073, + 0.011157025583088398, + 0.00357682048343122, + 0.0027737203054130077, + 0.005465005058795214, + 0.004358698148280382, + 0.003927411045879126, + 0.0052028680220246315, + 0.004037271253764629, + 0.0038719186559319496, + 0.0039597162976861, + 0.004128853790462017, + 0.011346803978085518, + 0.00786465685814619, + 0.005931079853326082, + 0.007833604700863361, + 0.01220919843763113, + 0.008612140081822872, + 0.0033195018768310547, + 0.004422056954354048, + 0.006305205635726452, + 0.0033573645632714033, + 0.0041443114168941975, + 0.004430565983057022, + 0.004404757171869278, + 0.005271351430565119, + 0.019464628770947456, + 0.02953336574137211, + 0.028273604810237885, + 0.024353351444005966, + 0.030515553429722786, + 0.02214994840323925, + 0.056702855974435806, + 0.011064735241234303, + 0.009617705829441547, + 0.011506280861794949, + 0.008537951856851578, + 0.011007565073668957, + 0.009008325636386871, + 0.00980782974511385, + 0.01180976815521717, + 0.00628335727378726, + 0.009822827763855457, + 0.006550564896315336, + 0.011858785524964333, + 0.008660069666802883, + 0.011634188704192638, + 0.009817156940698624, + 0.006350149866193533, + 0.007724276278167963, + 0.007625684142112732, + 0.009235401637852192, + 0.011764150112867355, + 0.007072398439049721, + 0.01161492895334959, + 0.009287289343774319, + 0.008695326745510101, + 0.010552017018198967, + 0.014608234167098999, + 0.017825592309236526, + 0.014558770693838596, + 0.04379357770085335, + 0.018494172021746635, + 0.01486111804842949, + 0.011356380768120289, + 0.014959275722503662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.04307584464550018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20076531171798706, + 0.12544430792331696, + 0.06439334899187088, + 0.04334542155265808, + 0.029138319194316864, + 0.10844803601503372, + 0.06974391639232635, + 0.06591372936964035, + 0.18429891765117645, + 0.10850866138935089 + ], + "credit_map": [ + 0.040571428835392, + 0.029224997386336327, + 0.03385442495346069, + 0.028835298493504524, + 0.010289865545928478, + 0.018699899315834045, + 0.01309815514832735, + 0.026191245764493942, + 0.014808020554482937, + 0.009957171976566315, + 0.025768892839550972, + 0.012394794262945652, + 0.006583689711987972, + 0.01710081845521927, + 0.014420638792216778, + 0.024410279467701912, + 0.009361272677779198, + 0.005260961130261421, + 0.012315732426941395, + 0.006156786810606718, + 0.007158728316426277, + 0.011525406502187252, + 0.005282403435558081, + 0.007332057226449251, + 0.0032776701264083385, + 0.0026524439454078674, + 0.007014745846390724, + 0.004901806823909283, + 0.004975491669028997, + 0.00819911528378725, + 0.007265052758157253, + 0.0050590974278748035, + 0.0031361898873001337, + 0.0033339462243020535, + 0.006068629678338766, + 0.003103381022810936, + 0.003285687416791916, + 0.003955656196922064, + 0.003867094637826085, + 0.0023877336643636227, + 0.00978055875748396, + 0.018410300835967064, + 0.019114840775728226, + 0.008487917482852936, + 0.01746678166091442, + 0.0141104431822896, + 0.015445507131516933, + 0.005631687585264444, + 0.008949225768446922, + 0.016306046396493912, + 0.008482834324240685, + 0.007425067014992237, + 0.007040341850370169, + 0.006104196421802044, + 0.010673903860151768, + 0.004762296564877033, + 0.00819592084735632, + 0.005878600757569075, + 0.010460296645760536, + 0.00706809526309371, + 0.008599674329161644, + 0.00868977326899767, + 0.008547200821340084, + 0.008474169299006462, + 0.019640706479549408, + 0.038531865924596786, + 0.01689484342932701, + 0.021433066576719284, + 0.028302408754825592, + 0.015561455860733986, + 0.015968549996614456, + 0.027966022491455078, + 0.007011192850768566, + 0.010172955691814423, + 0.011268476024270058, + 0.039876289665699005, + 0.00726998969912529, + 0.0178127009421587, + 0.007795269601047039, + 0.007301796227693558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.00319268717430532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627056445926428 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627056445926428, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627056445926428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08373823016881943, + 0.0506877675652504, + 0.026170767843723297, + 0.08325118571519852, + 0.058819886296987534, + 0.16113616526126862, + 0.09074379503726959, + 0.22512836754322052, + 0.16976487636566162, + 0.05055892467498779 + ], + "credit_map": [ + 0.01570354588329792, + 0.01160292886197567, + 0.011426091194152832, + 0.009754536673426628, + 0.008417616598308086, + 0.008292156271636486, + 0.00833587534725666, + 0.010205487720668316, + 0.005133189260959625, + 0.004877554718405008, + 0.01151319220662117, + 0.00524131441488862, + 0.003668868215754628, + 0.006035447586327791, + 0.005288256332278252, + 0.00892995111644268, + 0.0028463362250477076, + 0.002814625622704625, + 0.0037475863937288523, + 0.004093295894563198, + 0.0031512738205492496, + 0.003402854548767209, + 0.0030648994725197554, + 0.003049897262826562, + 0.004891880322247744, + 0.005018199793994427, + 0.01634748838841915, + 0.008507814258337021, + 0.00950429867953062, + 0.013833250850439072, + 0.013324456289410591, + 0.011823799461126328, + 0.008227922953665257, + 0.006941886153072119, + 0.013678353279829025, + 0.005126162897795439, + 0.006018498446792364, + 0.007664395496249199, + 0.0055808620527386665, + 0.0055818138644099236, + 0.013840108178555965, + 0.029898742213845253, + 0.023387962952256203, + 0.012471027672290802, + 0.02470087818801403, + 0.02548021823167801, + 0.02033768966794014, + 0.011019537225365639, + 0.01656876690685749, + 0.018111510202288628, + 0.01186087355017662, + 0.008138739503920078, + 0.010114237666130066, + 0.008973216637969017, + 0.010112958960235119, + 0.006863495800644159, + 0.022069748491048813, + 0.01650587096810341, + 0.0470394492149353, + 0.01938381791114807, + 0.025222033262252808, + 0.03704128414392471, + 0.024044444784522057, + 0.03382173180580139, + 0.011750375851988792, + 0.027846012264490128, + 0.026967138051986694, + 0.017470046877861023, + 0.03344642370939255, + 0.019041094928979874, + 0.017521461471915245, + 0.01572233811020851, + 0.007565852254629135, + 0.004955723881721497, + 0.004734914284199476, + 0.01272326149046421, + 0.005086024757474661, + 0.00495665380731225, + 0.00489412946626544, + 0.005642369855195284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.007627056445926428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Crusader knight wearing a battle-worn chain mail suit and a closed-faced greathelm with a narrow eye slit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a heavy tunic that is once-white but stained with dirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an ancient forest with twisted branches, slippery decaying leaves, and thick gray fog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle Steadicam tracking shot following the knight's steps and then rise smoothly to reveal a distant Gothic church?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the revealed Gothic church dilapidated, made of weathered gray stone covered in thick ivy and dark green moss, with broken stained glass windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977916456758976 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.00977916456758976, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977916456758976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14939427375793457, + 0.08676847815513611, + 0.018614618107676506, + 0.03919569402933121, + 0.013238322921097279, + 0.15146850049495697, + 0.06328839808702469, + 0.061466068029403687, + 0.2614491283893585, + 0.155116468667984 + ], + "credit_map": [ + 0.03145821765065193, + 0.02450965903699398, + 0.023599276319146156, + 0.013015893287956715, + 0.00734071945771575, + 0.013906262814998627, + 0.009385929442942142, + 0.02617831528186798, + 0.005318342242389917, + 0.006088015157729387, + 0.026697006076574326, + 0.0058168382383883, + 0.00329274102114141, + 0.0075225369073450565, + 0.00487008411437273, + 0.027162928134202957, + 0.002967778593301773, + 0.0017229723744094372, + 0.0028180459048599005, + 0.0017008972354233265, + 0.002126174047589302, + 0.00283027533441782, + 0.0014688513474538922, + 0.0029796261806041002, + 0.0023212383966892958, + 0.0020550901535898447, + 0.010723661631345749, + 0.0034028112422674894, + 0.0034022305626422167, + 0.00363097433000803, + 0.005392412655055523, + 0.008267284370958805, + 0.001288171741180122, + 0.0016426160000264645, + 0.002615631092339754, + 0.0011886231368407607, + 0.0017380004283040762, + 0.0018972645048052073, + 0.0017722051125019789, + 0.0010958131169900298, + 0.012618745677173138, + 0.02132450044155121, + 0.025546563789248466, + 0.014043097384274006, + 0.020682793110609055, + 0.025826470926404, + 0.021185189485549927, + 0.01024114154279232, + 0.013314465060830116, + 0.012784073129296303, + 0.007732018828392029, + 0.004799338523298502, + 0.006082252133637667, + 0.005209260620176792, + 0.00821391586214304, + 0.005153078585863113, + 0.007799752056598663, + 0.005314801819622517, + 0.009395685978233814, + 0.005619702860713005, + 0.0076820990070700645, + 0.009236916899681091, + 0.0071193380281329155, + 0.009297778829932213, + 0.025554412975907326, + 0.02473568357527256, + 0.05065511539578438, + 0.023835649713873863, + 0.036899250000715256, + 0.02504737116396427, + 0.017416659742593765, + 0.057305023074150085, + 0.01567416824400425, + 0.014652120880782604, + 0.016148896887898445, + 0.05466655641794205, + 0.00992928259074688, + 0.021374594420194626, + 0.011018265970051289, + 0.01165260374546051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.00977916456758976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_tensor([4925, 4925])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_tensor([4925, 4925])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..53482671d3e1d32355e125df2012fd78a4dcda91 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_tensor([4925, 4925])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, uninterrupted shot, filmed from a low angle and panning slowly to the right, transports us, in a photorealistic documentary style, deep into the luminescent nighttime jungles of Pandora. In the frame, a massive six-legged beast, head bowed, forages for food. Its six sturdy, arthropod-like legs support its massive frame. Its exoskeleton, a deep blue and purple chitinous exoskeleton with bioluminescent veins, shimmers like obsidian under the dim glow of the surrounding vegetation. Its massive head slowly lowers, its front devoid of visible eyes replaced by a pair of extremely sensitive, feathery tentacles. These tentacles gently sway in the humid air, their tips emitting a soft turquoise glow, as if sensing the chemical signals of their surroundings. The beast carefully uses its intricate mouthparts to probe a patch of bizarre fungi on the ground, emitting a halo of orange light. The background is filled with giant, spiraling ferns and towering trees draped with glowing vines. A thin layer of glowing spores hangs in the air, like floating dust. The entire scene is illuminated entirely by the creatures themselves and the surrounding flora, creating an atmosphere of mystery, tranquility, and primal vitality.", + "seed": 25522, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_4925_seed_25522_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1871155213797465, + "vlm_logprob": -0.1871155213797465, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.1871155213797465, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029978811740875244 + ], + "opposite_logprob_tokens": [ + -3.6549787521362305 + ], + "logprob_reward": -0.029978811740875244, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029978811740875244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12774166464805603, + 0.16528171300888062, + 0.11932246387004852, + 0.13257674872875214, + 0.07652375102043152, + 0.05630956590175629, + 0.03939225897192955, + 0.05919422581791878, + 0.07106394320726395, + 0.1525937020778656 + ], + "credit_map": [ + 0.020836422219872475, + 0.014835161156952381, + 0.023648442700505257, + 0.013582546263933182, + 0.012161063961684704, + 0.014913240447640419, + 0.01296396553516388, + 0.014800827950239182, + 0.013839435763657093, + 0.014915522187948227, + 0.041737113147974014, + 0.01663125678896904, + 0.012023153714835644, + 0.012264636345207691, + 0.01148802600800991, + 0.04238256439566612, + 0.007755191531032324, + 0.00519527355208993, + 0.007759416941553354, + 0.004884697031229734, + 0.005016271024942398, + 0.04134725034236908, + 0.006238728296011686, + 0.04112563654780388, + 0.01768060028553009, + 0.014202289283275604, + 0.016400834545493126, + 0.01178623829036951, + 0.008329596370458603, + 0.02018560841679573, + 0.009391557425260544, + 0.03460002318024635, + 0.008844946511089802, + 0.009528376162052155, + 0.011746219359338284, + 0.009874793700873852, + 0.006605078931897879, + 0.012559090740978718, + 0.008861059322953224, + 0.008504184894263744, + 0.006180528085678816, + 0.007432649377733469, + 0.009324237704277039, + 0.006016015075147152, + 0.004204295575618744, + 0.005601597484201193, + 0.009886643849313259, + 0.007663601543754339, + 0.003877595067024231, + 0.00594360800459981, + 0.005699032451957464, + 0.006359520368278027, + 0.003800122532993555, + 0.0027004475705325603, + 0.004593012388795614, + 0.006418922916054726, + 0.005482148844748735, + 0.010476084426045418, + 0.009270798414945602, + 0.01124071329832077, + 0.005051865708082914, + 0.005188426468521357, + 0.005584220867604017, + 0.006899968720972538, + 0.0061645107343792915, + 0.0077412608079612255, + 0.010780942626297474, + 0.014231191016733646, + 0.005507536232471466, + 0.009550731629133224, + 0.00734336394816637, + 0.009744405746459961, + 0.01339685171842575, + 0.01717514917254448, + 0.02235490083694458, + 0.018925350159406662, + 0.015919579192996025, + 0.015644852072000504, + 0.02729046531021595, + 0.02188655361533165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.029978811740875244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983214497566223 + ], + "opposite_logprob_tokens": [ + -0.5233214497566223 + ], + "logprob_reward": -0.8983214497566223, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983214497566223, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13366402685642242, + 0.24853219091892242, + 0.049359071999788284, + 0.13366547226905823, + 0.04167034476995468, + 0.17568987607955933, + 0.06434760987758636, + 0.04060971364378929, + 0.04934238642454147, + 0.0631192997097969 + ], + "credit_map": [ + 0.015250016935169697, + 0.010569678619503975, + 0.031599260866642, + 0.012638078071177006, + 0.009690962731838226, + 0.03194653242826462, + 0.0074863070622086525, + 0.014483175240457058, + 0.022089583799242973, + 0.014039289206266403, + 0.07560485601425171, + 0.01768157258629799, + 0.01639094389975071, + 0.015073646791279316, + 0.007948320358991623, + 0.07970396429300308, + 0.004676496610045433, + 0.0049643442034721375, + 0.00841012503951788, + 0.005203426815569401, + 0.0038598021492362022, + 0.008263500407338142, + 0.008196042850613594, + 0.005785330198705196, + 0.014422409236431122, + 0.015311701223254204, + 0.01485508680343628, + 0.013865235261619091, + 0.009588816203176975, + 0.021113719791173935, + 0.008241225965321064, + 0.03626726567745209, + 0.003435655729845166, + 0.004967680666595697, + 0.004763557109981775, + 0.006815984379500151, + 0.003527743276208639, + 0.007244272157549858, + 0.005372952204197645, + 0.005542496219277382, + 0.020625216886401176, + 0.02337813377380371, + 0.027779268100857735, + 0.022434284910559654, + 0.014056256040930748, + 0.01899053156375885, + 0.022143930196762085, + 0.026282234117388725, + 0.0057847220450639725, + 0.005635117180645466, + 0.013380232267081738, + 0.0071976035833358765, + 0.00477985292673111, + 0.005831276997923851, + 0.006050666328519583, + 0.015688132494688034, + 0.0032217775005847216, + 0.008224348537623882, + 0.00514276372268796, + 0.007611132692545652, + 0.0036686367820948362, + 0.0028239632956683636, + 0.003958550281822681, + 0.0059585352428257465, + 0.004264183808118105, + 0.005080683622509241, + 0.008715510368347168, + 0.007773247081786394, + 0.0037923739291727543, + 0.008075377903878689, + 0.005104077514261007, + 0.00653692614287138, + 0.006301452871412039, + 0.0057005793787539005, + 0.006274230312556028, + 0.0070619662292301655, + 0.007405590265989304, + 0.013366436585783958, + 0.007806778885424137, + 0.009202254004776478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.8983214497566223 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019451522966846824 + ], + "opposite_logprob_tokens": [ + -6.2519450187683105 + ], + "logprob_reward": -0.0019451522966846824, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019451522966846824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12362014502286911, + 0.4840971827507019, + 0.07140415161848068, + 0.1133715957403183, + 0.04117795452475548, + 0.02288700081408024, + 0.011577277444303036, + 0.019430240616202354, + 0.053776610642671585, + 0.05865788832306862 + ], + "credit_map": [ + 0.013874726369976997, + 0.023922324180603027, + 0.019983425736427307, + 0.01240513939410448, + 0.016737814992666245, + 0.013202137313783169, + 0.011767919175326824, + 0.011726653203368187, + 0.03730097413063049, + 0.02663019485771656, + 0.1534935086965561, + 0.027352329343557358, + 0.028594302013516426, + 0.03156805783510208, + 0.010782725177705288, + 0.16837508976459503, + 0.006227824836969376, + 0.007599653210490942, + 0.009174151346087456, + 0.008203644305467606, + 0.005735580809414387, + 0.009853719733655453, + 0.013732023537158966, + 0.01087755337357521, + 0.013806208036839962, + 0.014886731281876564, + 0.01521997433155775, + 0.018026627600193024, + 0.009910096414387226, + 0.014166232198476791, + 0.006773405708372593, + 0.02058231830596924, + 0.004646311514079571, + 0.007133116014301777, + 0.005855408962816, + 0.003622075542807579, + 0.0050929537974298, + 0.005267637316137552, + 0.003466264111921191, + 0.006094187498092651, + 0.0021525700576603413, + 0.0033824050333350897, + 0.0026679509319365025, + 0.002929768757894635, + 0.0019342057639732957, + 0.0027607157826423645, + 0.0038409142289310694, + 0.0032184701412916183, + 0.0016234772047027946, + 0.0011200675508007407, + 0.0018826064188033342, + 0.001172760152257979, + 0.0012110922252759337, + 0.0014655773993581533, + 0.0013096514157950878, + 0.001792045426554978, + 0.002360070589929819, + 0.0025591065641492605, + 0.0037216837517917156, + 0.003169961040839553, + 0.0016812972025945783, + 0.0015529814409092069, + 0.0019158574286848307, + 0.0024692814331501722, + 0.004580614157021046, + 0.006745719350874424, + 0.009385964833199978, + 0.009932935237884521, + 0.004454897716641426, + 0.006230163387954235, + 0.004803006071597338, + 0.007643305696547031, + 0.005432581529021263, + 0.008421110920608044, + 0.006328044459223747, + 0.009670288302004337, + 0.006954612210392952, + 0.004313322249799967, + 0.009691095910966396, + 0.00784683134406805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0019451522966846824 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036193120758980513 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036193120758980513, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036193120758980513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.147847518324852, + 0.3423164486885071, + 0.05223507434129715, + 0.1321127712726593, + 0.061177704483270645, + 0.030142106115818024, + 0.014516633003950119, + 0.0459931381046772, + 0.052216537296772, + 0.12144213169813156 + ], + "credit_map": [ + 0.018264776095747948, + 0.019354697316884995, + 0.02587774023413658, + 0.020263908430933952, + 0.014568195678293705, + 0.020795419812202454, + 0.014194081537425518, + 0.014528688974678516, + 0.031056849285960197, + 0.015614758245646954, + 0.10741616040468216, + 0.020747045055031776, + 0.02073194459080696, + 0.01892869547009468, + 0.011589877307415009, + 0.11623108386993408, + 0.004800271242856979, + 0.006435805466026068, + 0.007435357663780451, + 0.00727567495778203, + 0.004574785940349102, + 0.006634542718529701, + 0.007440776564180851, + 0.007637855596840382, + 0.014972819946706295, + 0.014615311287343502, + 0.015682704746723175, + 0.01772945560514927, + 0.011258228681981564, + 0.01903526298701763, + 0.009499524720013142, + 0.02931944653391838, + 0.007596198935061693, + 0.009204242378473282, + 0.0089156124740839, + 0.005268024746328592, + 0.006628243252635002, + 0.010428840294480324, + 0.005003603640943766, + 0.0081329345703125, + 0.0033932023216038942, + 0.004107320215553045, + 0.0044743213802576065, + 0.004382372833788395, + 0.0021120826713740826, + 0.00406253058463335, + 0.0036518501583486795, + 0.003958423156291246, + 0.001730813062749803, + 0.0017444281838834286, + 0.002148807281628251, + 0.0019796218257397413, + 0.0013533211313188076, + 0.0013595483032986522, + 0.001973214093595743, + 0.0022268774919211864, + 0.0038918438367545605, + 0.008795074187219143, + 0.007309041917324066, + 0.008449173532426357, + 0.004201680887490511, + 0.0036893070209771395, + 0.003941509872674942, + 0.005715501494705677, + 0.004645445849746466, + 0.0051622334867715836, + 0.008363183587789536, + 0.01124550960958004, + 0.003657339373603463, + 0.006248570512980223, + 0.004880360793322325, + 0.00801389291882515, + 0.01104486733675003, + 0.015229794196784496, + 0.014014620333909988, + 0.019929831847548485, + 0.012804688885807991, + 0.00945516861975193, + 0.02295219525694847, + 0.0160109493881464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0036193120758980513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017128810286521912 + ], + "opposite_logprob_tokens": [ + -6.376712799072266 + ], + "logprob_reward": -0.0017128810286521912, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017128810286521912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18812765181064606, + 0.2507864236831665, + 0.18162521719932556, + 0.07038100063800812, + 0.07408276945352554, + 0.03868279978632927, + 0.03189313784241676, + 0.030523283407092094, + 0.054314401000738144, + 0.07958332449197769 + ], + "credit_map": [ + 0.03472617268562317, + 0.014147951267659664, + 0.02842009626328945, + 0.029516594484448433, + 0.015108196064829826, + 0.024326706305146217, + 0.014570926316082478, + 0.027310987934470177, + 0.026528209447860718, + 0.010640307329595089, + 0.06990434229373932, + 0.015352755784988403, + 0.014056931249797344, + 0.0238559078425169, + 0.013811196200549603, + 0.07663676887750626, + 0.015866395086050034, + 0.011802681721746922, + 0.025562413036823273, + 0.019193023443222046, + 0.012301313690841198, + 0.025872934609651566, + 0.026807226240634918, + 0.044219207018613815, + 0.006501350086182356, + 0.006730918772518635, + 0.010379673913121223, + 0.010217644274234772, + 0.006914244033396244, + 0.010438474826514721, + 0.005912823602557182, + 0.013285861350595951, + 0.006059423089027405, + 0.006988991517573595, + 0.009542164392769337, + 0.009334444999694824, + 0.0065740845166146755, + 0.010858999565243721, + 0.015372647903859615, + 0.009352006018161774, + 0.0026354670990258455, + 0.0039827460423111916, + 0.005003829952329397, + 0.0046341512352228165, + 0.0030748159624636173, + 0.006006783340126276, + 0.00800861231982708, + 0.00533638708293438, + 0.0033824231941252947, + 0.004540898371487856, + 0.0058288127183914185, + 0.0033038000110536814, + 0.0026033497415483, + 0.003966099116951227, + 0.0034221832174807787, + 0.004845569375902414, + 0.0024310376029461622, + 0.0027179429307579994, + 0.005821685306727886, + 0.0024304913822561502, + 0.0027101135347038507, + 0.006766050588339567, + 0.0028594681061804295, + 0.00478648766875267, + 0.005071461666375399, + 0.0054431031458079815, + 0.008349264040589333, + 0.00997914094477892, + 0.003927764482796192, + 0.008070412091910839, + 0.005218473728746176, + 0.008254775777459145, + 0.006133669055998325, + 0.007550165522843599, + 0.010488366708159447, + 0.012977649457752705, + 0.009032121859490871, + 0.009961781091988087, + 0.01286690030246973, + 0.010572661645710468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0017128810286521912 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, uninterrupted shot, filmed from a low angle and panning slowly to the right, transports us, in a photorealistic documentary style, deep into the luminescent nighttime jungles of Pandora. In the frame, a massive six-legged beast, head bowed, forages for food. Its six sturdy, arthropod-like legs support its massive frame. Its exoskeleton, a deep blue and purple chitinous exoskeleton with bioluminescent veins, shimmers like obsidian under the dim glow of the surrounding vegetation. Its massive head slowly lowers, its front devoid of visible eyes replaced by a pair of extremely sensitive, feathery tentacles. These tentacles gently sway in the humid air, their tips emitting a soft turquoise glow, as if sensing the chemical signals of their surroundings. The beast carefully uses its intricate mouthparts to probe a patch of bizarre fungi on the ground, emitting a halo of orange light. The background is filled with giant, spiraling ferns and towering trees draped with glowing vines. A thin layer of glowing spores hangs in the air, like floating dust. The entire scene is illuminated entirely by the creatures themselves and the surrounding flora, creating an atmosphere of mystery, tranquility, and primal vitality.", + "seed": 25523, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_4925_seed_25523_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1036294057033956, + "vlm_logprob": -0.1036294057033956, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.1036294057033956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035745423287153244 + ], + "opposite_logprob_tokens": [ + -3.53574538230896 + ], + "logprob_reward": -0.035745423287153244, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.035745423287153244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3275822699069977, + 0.07973306626081467, + 0.04951820895075798, + 0.06572059541940689, + 0.048727601766586304, + 0.11566847562789917, + 0.06387519836425781, + 0.13937215507030487, + 0.05246712267398834, + 0.05733530968427658 + ], + "credit_map": [ + 0.032559461891651154, + 0.028577422723174095, + 0.07071347534656525, + 0.024391762912273407, + 0.08522890508174896, + 0.03578957915306091, + 0.019417166709899902, + 0.030904483050107956, + 0.009552608244121075, + 0.012287852354347706, + 0.01247228030115366, + 0.008775993250310421, + 0.007846713066101074, + 0.010125597938895226, + 0.006949687376618385, + 0.011722337454557419, + 0.006322101224213839, + 0.0044008465483784676, + 0.007703912910073996, + 0.004965831991285086, + 0.005466705188155174, + 0.008837578818202019, + 0.005320051219314337, + 0.006501181051135063, + 0.006758879870176315, + 0.00704212998971343, + 0.009211902506649494, + 0.008594307117164135, + 0.007158353459089994, + 0.006956014782190323, + 0.012366159819066525, + 0.00763284508138895, + 0.005123303271830082, + 0.006649329327046871, + 0.007175506092607975, + 0.00586689542979002, + 0.006011531222611666, + 0.007442690897732973, + 0.00407958310097456, + 0.006378763355314732, + 0.007871942594647408, + 0.010849918238818645, + 0.024528011679649353, + 0.009682686068117619, + 0.009730729274451733, + 0.01665060967206955, + 0.008781892247498035, + 0.02757267653942108, + 0.004883345682173967, + 0.011320986784994602, + 0.010680309496819973, + 0.009508233517408371, + 0.005954669788479805, + 0.006818793714046478, + 0.004420062992721796, + 0.010288797318935394, + 0.012923184782266617, + 0.01090901717543602, + 0.031930845230817795, + 0.008845140226185322, + 0.010801535099744797, + 0.03554518148303032, + 0.014986085705459118, + 0.013431175611913204, + 0.004538091830909252, + 0.006759576965123415, + 0.008106788620352745, + 0.004958149511367083, + 0.005014342721551657, + 0.008046229369938374, + 0.005142580717802048, + 0.009901364333927631, + 0.0077886502258479595, + 0.008570821024477482, + 0.010683487169444561, + 0.005414423532783985, + 0.007217731326818466, + 0.006152490619570017, + 0.004386137239634991, + 0.007121567148715258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.035745423287153244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742985963821411 + ], + "opposite_logprob_tokens": [ + -0.9742985963821411 + ], + "logprob_reward": -0.4742985963821411, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742985963821411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37062931060791016, + 0.07103259116411209, + 0.031204678118228912, + 0.01939709484577179, + 0.02799287997186184, + 0.10235846042633057, + 0.02622400037944317, + 0.057274237275123596, + 0.09106329083442688, + 0.20282350480556488 + ], + "credit_map": [ + 0.02747240662574768, + 0.019968241453170776, + 0.08857689797878265, + 0.02116226777434349, + 0.13933165371418, + 0.027863329276442528, + 0.014974525198340416, + 0.031279973685741425, + 0.0059075732715427876, + 0.007878519594669342, + 0.012418140657246113, + 0.007846187800168991, + 0.01301676407456398, + 0.013592126779258251, + 0.00393978226929903, + 0.006433497183024883, + 0.002885353984311223, + 0.00365698360837996, + 0.0038705209735780954, + 0.0028920078184455633, + 0.0034225606359541416, + 0.005471317563205957, + 0.00506733451038599, + 0.003938600420951843, + 0.0017450732411816716, + 0.0031550885178148746, + 0.002634400501847267, + 0.0025701113045215607, + 0.0019363644532859325, + 0.0022738478146493435, + 0.002827861811965704, + 0.00225434685125947, + 0.002484367461875081, + 0.0069131203927099705, + 0.0038869197014719248, + 0.003769682487472892, + 0.0036985953338444233, + 0.0028682006523013115, + 0.0021226315293461084, + 0.002249364275485277, + 0.006582936737686396, + 0.019683316349983215, + 0.01750899851322174, + 0.013009757734835148, + 0.015481148846447468, + 0.010625149123370647, + 0.006044761277735233, + 0.013422387652099133, + 0.0020647523924708366, + 0.003925243392586708, + 0.0044169737957417965, + 0.003326400415971875, + 0.002506428398191929, + 0.0036047049798071384, + 0.002663250081241131, + 0.003716247621923685, + 0.005485298112034798, + 0.0065887258388102055, + 0.010338491760194302, + 0.004968666471540928, + 0.004958437755703926, + 0.010035978630185127, + 0.007167733274400234, + 0.007730911020189524, + 0.008689479902386665, + 0.006903441622853279, + 0.026619749143719673, + 0.005438454449176788, + 0.005385000724345446, + 0.024947863072156906, + 0.0040349275805056095, + 0.009044371545314789, + 0.019400835037231445, + 0.032660890370607376, + 0.028961148113012314, + 0.021431032568216324, + 0.030643878504633904, + 0.013942320831120014, + 0.02298804372549057, + 0.03279535844922066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.4742985963821411 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193875541910529 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193875541910529, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193875541910529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39173251390457153, + 0.09180063754320145, + 0.05128273367881775, + 0.040733758360147476, + 0.04536660388112068, + 0.09209390729665756, + 0.16608667373657227, + 0.03504272550344467, + 0.031430959701538086, + 0.054429445415735245 + ], + "credit_map": [ + 0.030145032331347466, + 0.02751517854630947, + 0.10743559896945953, + 0.028827769681811333, + 0.13585175573825836, + 0.024789229035377502, + 0.012589400634169579, + 0.02457854337990284, + 0.007771910168230534, + 0.01027289405465126, + 0.019441740587353706, + 0.010462505742907524, + 0.01855105347931385, + 0.011345332488417625, + 0.00654979981482029, + 0.007405407726764679, + 0.005326955113559961, + 0.008110112510621548, + 0.007044471800327301, + 0.0052770283073186874, + 0.007863737642765045, + 0.00750075513496995, + 0.0038513715844601393, + 0.006308301817625761, + 0.003762746462598443, + 0.0057516805827617645, + 0.005146537441760302, + 0.004754540976136923, + 0.004865194205194712, + 0.005575070157647133, + 0.005012696608901024, + 0.005865293089300394, + 0.0044861966744065285, + 0.00883498415350914, + 0.009003511629998684, + 0.006277097389101982, + 0.00445591239258647, + 0.004785453900694847, + 0.0033703905064612627, + 0.004153054673224688, + 0.00698271207511425, + 0.016758564859628677, + 0.016510287299752235, + 0.01112031377851963, + 0.014804070815443993, + 0.008487042970955372, + 0.0058878022246062756, + 0.01154311466962099, + 0.015539143234491348, + 0.039324965327978134, + 0.025161590427160263, + 0.026917336508631706, + 0.01972808875143528, + 0.014934948645532131, + 0.009661748073995113, + 0.014818862080574036, + 0.0035414830781519413, + 0.005528184585273266, + 0.005987432319670916, + 0.004067228175699711, + 0.004546462092548609, + 0.003882352728396654, + 0.0028647612780332565, + 0.004624823573976755, + 0.0030791147146373987, + 0.0037266649305820465, + 0.005918215494602919, + 0.0027952417731285095, + 0.0025188722647726536, + 0.005848141387104988, + 0.0028902869671583176, + 0.004654422402381897, + 0.006168292835354805, + 0.007877057418227196, + 0.009685840457677841, + 0.0051560793071985245, + 0.00794234499335289, + 0.004684038460254669, + 0.0064393747597932816, + 0.006476419512182474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.003193875541910529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031960143242031336 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.0031960143242031336, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031960143242031336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4208027422428131, + 0.039262812584638596, + 0.05715976282954216, + 0.037797022610902786, + 0.04592268913984299, + 0.11107742786407471, + 0.04864156246185303, + 0.06378991156816483, + 0.10048671811819077, + 0.07505936920642853 + ], + "credit_map": [ + 0.0269465409219265, + 0.019238684326410294, + 0.12161999195814133, + 0.02391243167221546, + 0.1491944044828415, + 0.027647685259580612, + 0.023127257823944092, + 0.029115719720721245, + 0.005037123337388039, + 0.006391175556927919, + 0.006455905269831419, + 0.004676640499383211, + 0.003518088487908244, + 0.006707263179123402, + 0.0026997223030775785, + 0.0037768883630633354, + 0.008407016284763813, + 0.0050445138476789, + 0.007176681887358427, + 0.00557434419170022, + 0.007437500171363354, + 0.007495181169360876, + 0.009214651770889759, + 0.006809864193201065, + 0.0028256974183022976, + 0.005263371393084526, + 0.003988361917436123, + 0.004986677318811417, + 0.005884167272597551, + 0.003974105231463909, + 0.006576081737875938, + 0.0042985547333955765, + 0.004141390323638916, + 0.005242085549980402, + 0.00624687597155571, + 0.006901427172124386, + 0.004474985413253307, + 0.00694736884906888, + 0.005334889516234398, + 0.006633659824728966, + 0.007754461839795113, + 0.013018753379583359, + 0.020817594602704048, + 0.010322128422558308, + 0.014951461926102638, + 0.015424376353621483, + 0.00811669509857893, + 0.020671941339969635, + 0.0040114386938512325, + 0.009659434668719769, + 0.00867124181240797, + 0.006442892365157604, + 0.00478893518447876, + 0.005045351572334766, + 0.0029836189933121204, + 0.00703864311799407, + 0.006187499035149813, + 0.007743093650788069, + 0.011398191563785076, + 0.007228562142699957, + 0.007933800108730793, + 0.011042851954698563, + 0.004914851859211922, + 0.007341058924794197, + 0.009401987306773663, + 0.00506886001676321, + 0.02982478402554989, + 0.005322705954313278, + 0.005329858046025038, + 0.029709555208683014, + 0.005449509713798761, + 0.010379442013800144, + 0.009487527422606945, + 0.010894028469920158, + 0.014292451553046703, + 0.005818283185362816, + 0.011069538071751595, + 0.006886358372867107, + 0.008008783683180809, + 0.008602388203144073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0031960143242031336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017131189815700054 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017131189815700054, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017131189815700054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4080277383327484, + 0.09487441182136536, + 0.08245348185300827, + 0.05208173394203186, + 0.03276539966464043, + 0.07430386543273926, + 0.04460921138525009, + 0.048926591873168945, + 0.09086422622203827, + 0.07109329849481583 + ], + "credit_map": [ + 0.030981995165348053, + 0.0190248042345047, + 0.10724202543497086, + 0.02333850972354412, + 0.13004587590694427, + 0.03425908088684082, + 0.03396183252334595, + 0.029173599556088448, + 0.008893311023712158, + 0.009926088154315948, + 0.014127287082374096, + 0.008137238211929798, + 0.01961158961057663, + 0.016355037689208984, + 0.008171086199581623, + 0.009652773849666119, + 0.006911649368703365, + 0.007250119466334581, + 0.007207308895885944, + 0.007443286012858152, + 0.011970976367592812, + 0.017683345824480057, + 0.012951220385730267, + 0.011035577394068241, + 0.004693229682743549, + 0.004876925144344568, + 0.004803818184882402, + 0.006338765844702721, + 0.004297117702662945, + 0.006099064834415913, + 0.010421833023428917, + 0.010550981387495995, + 0.0030441235285252333, + 0.005575015675276518, + 0.004013934172689915, + 0.005342974793165922, + 0.003013943089172244, + 0.004140421748161316, + 0.0030486800242215395, + 0.004586306866258383, + 0.006308695767074823, + 0.008171187713742256, + 0.013130894862115383, + 0.007917076349258423, + 0.0070558334700763226, + 0.012159491889178753, + 0.005543944891542196, + 0.014016740024089813, + 0.0035081212408840656, + 0.006623316090553999, + 0.007157149724662304, + 0.004502007272094488, + 0.004463460762053728, + 0.0042097195982933044, + 0.0048539238050580025, + 0.009291511960327625, + 0.006561122834682465, + 0.006452777422964573, + 0.007712201680988073, + 0.0041028279811143875, + 0.005463559180498123, + 0.005331727210432291, + 0.005047619342803955, + 0.00825475249439478, + 0.00839413981884718, + 0.005226090084761381, + 0.025268584489822388, + 0.004267904907464981, + 0.005232336465269327, + 0.023073319345712662, + 0.007848948240280151, + 0.011552904732525349, + 0.010137236677110195, + 0.005547540728002787, + 0.01392478309571743, + 0.004453927278518677, + 0.005186567548662424, + 0.01067754440009594, + 0.005779089406132698, + 0.015386614017188549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0017131189815700054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, uninterrupted shot, filmed from a low angle and panning slowly to the right, transports us, in a photorealistic documentary style, deep into the luminescent nighttime jungles of Pandora. In the frame, a massive six-legged beast, head bowed, forages for food. Its six sturdy, arthropod-like legs support its massive frame. Its exoskeleton, a deep blue and purple chitinous exoskeleton with bioluminescent veins, shimmers like obsidian under the dim glow of the surrounding vegetation. Its massive head slowly lowers, its front devoid of visible eyes replaced by a pair of extremely sensitive, feathery tentacles. These tentacles gently sway in the humid air, their tips emitting a soft turquoise glow, as if sensing the chemical signals of their surroundings. The beast carefully uses its intricate mouthparts to probe a patch of bizarre fungi on the ground, emitting a halo of orange light. The background is filled with giant, spiraling ferns and towering trees draped with glowing vines. A thin layer of glowing spores hangs in the air, like floating dust. The entire scene is illuminated entirely by the creatures themselves and the surrounding flora, creating an atmosphere of mystery, tranquility, and primal vitality.", + "seed": 25524, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_4925_seed_25524_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1277808300917968, + "vlm_logprob": -0.1277808300917968, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.1277808300917968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05181527137756348 + ], + "opposite_logprob_tokens": [ + -3.0518152713775635 + ], + "logprob_reward": -0.05181527137756348, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05181527137756348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39205554127693176, + 0.04721803218126297, + 0.052153848111629486, + 0.03185849264264107, + 0.07914015650749207, + 0.03154587373137474, + 0.04227214306592941, + 0.03819533437490463, + 0.17725549638271332, + 0.10830508172512054 + ], + "credit_map": [ + 0.0622282512485981, + 0.043772950768470764, + 0.05797957628965378, + 0.02940002642571926, + 0.04215642064809799, + 0.06413696706295013, + 0.04753453657031059, + 0.04484681785106659, + 0.0055475481785833836, + 0.006372358649969101, + 0.006944019813090563, + 0.004960502497851849, + 0.006238255649805069, + 0.006771651096642017, + 0.004608626011759043, + 0.005775074474513531, + 0.005976829677820206, + 0.0051445974968373775, + 0.006581475026905537, + 0.0071449498645961285, + 0.006785620469599962, + 0.00807251874357462, + 0.006762594450265169, + 0.005685271229594946, + 0.005197472404688597, + 0.002913728356361389, + 0.0042638317681849, + 0.004326522815972567, + 0.0034495610743761063, + 0.00338327931240201, + 0.004428377375006676, + 0.003895722795277834, + 0.008165981620550156, + 0.00934149045497179, + 0.007031412795186043, + 0.011732019484043121, + 0.00777691975235939, + 0.016982728615403175, + 0.009826420806348324, + 0.008283192291855812, + 0.003586891572922468, + 0.0037522779311984777, + 0.004779157228767872, + 0.004062107764184475, + 0.0030697353649884462, + 0.004386800806969404, + 0.004338456317782402, + 0.0035704486072063446, + 0.004815792199224234, + 0.003420488443225622, + 0.00404741708189249, + 0.0047487965784966946, + 0.0038502805400639772, + 0.004973894450813532, + 0.009259548969566822, + 0.00715592922642827, + 0.0023551543708890676, + 0.002377365017309785, + 0.002845533424988389, + 0.0030928112100809813, + 0.003345462027937174, + 0.012072745710611343, + 0.0032907899003475904, + 0.008815478533506393, + 0.012218380346894264, + 0.011181401088833809, + 0.030814848840236664, + 0.006045806687325239, + 0.011867266148328781, + 0.06974898278713226, + 0.00738552724942565, + 0.027993302792310715, + 0.007967733778059483, + 0.008140834048390388, + 0.011114819906651974, + 0.01132418867200613, + 0.008897551335394382, + 0.034450359642505646, + 0.006675161421298981, + 0.019734438508749008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.05181527137756348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761515498161316 + ], + "opposite_logprob_tokens": [ + -0.8261515498161316 + ], + "logprob_reward": -0.5761515498161316, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761515498161316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2187969833612442, + 0.059537094086408615, + 0.07015442848205566, + 0.0474376305937767, + 0.0551605224609375, + 0.030798958614468575, + 0.020480919629335403, + 0.20616982877254486, + 0.16399210691452026, + 0.12747147679328918 + ], + "credit_map": [ + 0.03786003217101097, + 0.037137214094400406, + 0.026451371610164642, + 0.022557059302926064, + 0.026726800948381424, + 0.021859053522348404, + 0.01969364657998085, + 0.02651182748377323, + 0.009052597917616367, + 0.00518518453463912, + 0.011405792087316513, + 0.006113129202276468, + 0.0039005684666335583, + 0.007875708863139153, + 0.006584275979548693, + 0.009419837966561317, + 0.01009052712470293, + 0.0051144883036613464, + 0.007453231140971184, + 0.006477524526417255, + 0.011558313854038715, + 0.00986985769122839, + 0.009560883045196533, + 0.010029601864516735, + 0.005455607082694769, + 0.0057505108416080475, + 0.006142941769212484, + 0.007230737712234259, + 0.004992075730115175, + 0.006553523242473602, + 0.00586513290181756, + 0.005447104573249817, + 0.006346370093524456, + 0.006790972780436277, + 0.007592547684907913, + 0.008837506175041199, + 0.006504605989903212, + 0.007741166744381189, + 0.006139000412076712, + 0.0052083516493439674, + 0.004133729264140129, + 0.00250836880877614, + 0.003623923286795616, + 0.0024603509809821844, + 0.0029689634684473276, + 0.004207132384181023, + 0.006923822686076164, + 0.003972669132053852, + 0.0024587649386376143, + 0.0022374868858605623, + 0.0020344918593764305, + 0.002118668518960476, + 0.0021606916561722755, + 0.002229791134595871, + 0.004421390127390623, + 0.002819634974002838, + 0.007824369706213474, + 0.006238373927772045, + 0.011058788746595383, + 0.00545959547162056, + 0.020438196137547493, + 0.0843583196401596, + 0.005770853720605373, + 0.0650213286280632, + 0.011466378346085548, + 0.010952681303024292, + 0.012900084257125854, + 0.0053677330724895, + 0.011282240971922874, + 0.06325171142816544, + 0.007476136088371277, + 0.04129514843225479, + 0.01143679954111576, + 0.009428792633116245, + 0.013775103725492954, + 0.014273417182266712, + 0.010995009914040565, + 0.03120104782283306, + 0.010207707062363625, + 0.02615359239280224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.5761515498161316 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735360249876976 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735360249876976, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735360249876976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30351850390434265, + 0.07335483282804489, + 0.07518763840198517, + 0.12170534580945969, + 0.1063702180981636, + 0.05167827382683754, + 0.08650847524404526, + 0.08234217762947083, + 0.044899046421051025, + 0.05443551018834114 + ], + "credit_map": [ + 0.041772857308387756, + 0.03521869331598282, + 0.04536430910229683, + 0.027178408578038216, + 0.035197630524635315, + 0.05406315252184868, + 0.033989112824201584, + 0.030734317377209663, + 0.008104062639176846, + 0.008928714320063591, + 0.009081508032977581, + 0.008538915775716305, + 0.009471643716096878, + 0.010197935625910759, + 0.008026331663131714, + 0.011005721054971218, + 0.010529730468988419, + 0.008962366729974747, + 0.007616284769028425, + 0.013839168474078178, + 0.007661830168217421, + 0.009250950068235397, + 0.007611892186105251, + 0.009715411812067032, + 0.011628789827227592, + 0.011288263835012913, + 0.01696672849357128, + 0.01438905019313097, + 0.01645619422197342, + 0.014743056148290634, + 0.021791402250528336, + 0.014441855251789093, + 0.012096290476620197, + 0.00923745334148407, + 0.009411056526005268, + 0.01566382311284542, + 0.009663563221693039, + 0.028805052861571312, + 0.008385978639125824, + 0.013107001781463623, + 0.009068287909030914, + 0.004486855585128069, + 0.005087121855467558, + 0.004661863669753075, + 0.004849691409617662, + 0.006817407440394163, + 0.008499879390001297, + 0.008207166567444801, + 0.010701749473810196, + 0.010599656961858273, + 0.010537464171648026, + 0.011267077177762985, + 0.008571825921535492, + 0.011643524281680584, + 0.01074483897536993, + 0.012442336417734623, + 0.0031215606722980738, + 0.002861408516764641, + 0.0047289286740124226, + 0.003409209894016385, + 0.0076216948218643665, + 0.030750779435038567, + 0.003954724874347448, + 0.02589387074112892, + 0.0034619539510458708, + 0.003508237423375249, + 0.0072555746883153915, + 0.004747371654957533, + 0.005032271612435579, + 0.01131516881287098, + 0.005990354344248772, + 0.00358811323530972, + 0.006271754391491413, + 0.00708326930180192, + 0.007555915974080563, + 0.007091662846505642, + 0.006155471317470074, + 0.009100416675209999, + 0.004832613747566938, + 0.00634440453723073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.006735360249876976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490278100594878 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.002490278100594878, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490278100594878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19559817016124725, + 0.04286125302314758, + 0.03822457790374756, + 0.05796623229980469, + 0.09188789129257202, + 0.06602489203214645, + 0.034585367888212204, + 0.14085781574249268, + 0.1416577398777008, + 0.19033604860305786 + ], + "credit_map": [ + 0.028460193425416946, + 0.02141832560300827, + 0.03222528100013733, + 0.01500269677489996, + 0.019150005653500557, + 0.024099409580230713, + 0.02318909578025341, + 0.0320531502366066, + 0.005143428687006235, + 0.004361679311841726, + 0.006920520681887865, + 0.004924027714878321, + 0.003922458738088608, + 0.006478220224380493, + 0.004565337207168341, + 0.00654557952657342, + 0.004733928479254246, + 0.004118465818464756, + 0.00437673507258296, + 0.005051479209214449, + 0.0053259036503732204, + 0.004219523165374994, + 0.003614715998992324, + 0.006783828604966402, + 0.006687222048640251, + 0.00546296825632453, + 0.006854234728962183, + 0.00729808583855629, + 0.007482936605811119, + 0.006638714112341404, + 0.01152864471077919, + 0.006013425998389721, + 0.008103867061436176, + 0.004808397032320499, + 0.00816405564546585, + 0.005809491034597158, + 0.007853380404412746, + 0.030913474038243294, + 0.0056643239222466946, + 0.020570900291204453, + 0.007900204509496689, + 0.004718326032161713, + 0.008401378057897091, + 0.004889445845037699, + 0.006125749088823795, + 0.008486990816891193, + 0.015745533630251884, + 0.009757264517247677, + 0.004557774867862463, + 0.004136009141802788, + 0.003642287105321884, + 0.004326150752604008, + 0.0032823390793055296, + 0.004247728269547224, + 0.004292081110179424, + 0.0061009968630969524, + 0.005862572696059942, + 0.004279521759599447, + 0.00760014308616519, + 0.0036042630672454834, + 0.01410822756588459, + 0.05717179551720619, + 0.004614522680640221, + 0.043616753071546555, + 0.009410711005330086, + 0.007470786105841398, + 0.03638095036149025, + 0.0056615304201841354, + 0.009634171612560749, + 0.0536070354282856, + 0.007500553969293833, + 0.01199200190603733, + 0.015056527219712734, + 0.010929050855338573, + 0.016051512211561203, + 0.010242346674203873, + 0.01188618689775467, + 0.0723857656121254, + 0.00860512163490057, + 0.045179542154073715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.002490278100594878 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017116909148171544 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017116909148171544, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017116909148171544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20565393567085266, + 0.039216186851263046, + 0.041039370000362396, + 0.022573189809918404, + 0.02019248902797699, + 0.060974299907684326, + 0.027211206033825874, + 0.19806921482086182, + 0.17195507884025574, + 0.21311503648757935 + ], + "credit_map": [ + 0.03375951945781708, + 0.022024042904376984, + 0.026420563459396362, + 0.018933279439806938, + 0.022914277389645576, + 0.02070806361734867, + 0.029907215386629105, + 0.030986974015831947, + 0.004956848919391632, + 0.00413037883117795, + 0.007281474303454161, + 0.004601989407092333, + 0.002883845940232277, + 0.005126515869051218, + 0.0035383908543735743, + 0.006696741562336683, + 0.005535664036870003, + 0.0046443031169474125, + 0.003829161636531353, + 0.0032870075665414333, + 0.004596359096467495, + 0.007309925742447376, + 0.004335174802690744, + 0.007501775398850441, + 0.0020757887978106737, + 0.0028578077908605337, + 0.002588309347629547, + 0.0034371628426015377, + 0.002807599725201726, + 0.002924850443378091, + 0.0032958481460809708, + 0.0025858222506940365, + 0.0024597588926553726, + 0.002164975507184863, + 0.002190086292102933, + 0.002934205112978816, + 0.002036578254774213, + 0.0021709701977670193, + 0.0036167448852211237, + 0.0026191691868007183, + 0.007370194885879755, + 0.005191801581531763, + 0.007248319685459137, + 0.005346362479031086, + 0.005866864696145058, + 0.008483360521495342, + 0.01421362068504095, + 0.007253777235746384, + 0.002620911691337824, + 0.0022027618251740932, + 0.002844625385478139, + 0.0020217946730554104, + 0.002568656112998724, + 0.0037100431509315968, + 0.004056823439896107, + 0.007185590453445911, + 0.007826395332813263, + 0.006040330044925213, + 0.010733081959187984, + 0.00423846859484911, + 0.01987478882074356, + 0.08112706243991852, + 0.005074176471680403, + 0.06315489858388901, + 0.011401079595088959, + 0.009085549972951412, + 0.04104847088456154, + 0.0063626947812736034, + 0.01179796177893877, + 0.06648031622171402, + 0.008191358298063278, + 0.01758764311671257, + 0.018046371638774872, + 0.0103559335693717, + 0.016987113282084465, + 0.012182663194835186, + 0.011645759455859661, + 0.07477637380361557, + 0.010475379414856434, + 0.05864545702934265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0017116909148171544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, uninterrupted shot, filmed from a low angle and panning slowly to the right, transports us, in a photorealistic documentary style, deep into the luminescent nighttime jungles of Pandora. In the frame, a massive six-legged beast, head bowed, forages for food. Its six sturdy, arthropod-like legs support its massive frame. Its exoskeleton, a deep blue and purple chitinous exoskeleton with bioluminescent veins, shimmers like obsidian under the dim glow of the surrounding vegetation. Its massive head slowly lowers, its front devoid of visible eyes replaced by a pair of extremely sensitive, feathery tentacles. These tentacles gently sway in the humid air, their tips emitting a soft turquoise glow, as if sensing the chemical signals of their surroundings. The beast carefully uses its intricate mouthparts to probe a patch of bizarre fungi on the ground, emitting a halo of orange light. The background is filled with giant, spiraling ferns and towering trees draped with glowing vines. A thin layer of glowing spores hangs in the air, like floating dust. The entire scene is illuminated entirely by the creatures themselves and the surrounding flora, creating an atmosphere of mystery, tranquility, and primal vitality.", + "seed": 25525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_6/index_4925_seed_25525_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024707395071163773, + "vlm_logprob": -0.024707395071163773, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.024707395071163773, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027975162491202354 + ], + "opposite_logprob_tokens": [ + -3.777975082397461 + ], + "logprob_reward": -0.027975162491202354, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027975162491202354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.195967897772789, + 0.05190446600317955, + 0.10102029144763947, + 0.0756453424692154, + 0.053659480065107346, + 0.05984734371304512, + 0.1808922290802002, + 0.04318275302648544, + 0.10279861092567444, + 0.13508158922195435 + ], + "credit_map": [ + 0.018835457041859627, + 0.01768786646425724, + 0.016496624797582626, + 0.01945965364575386, + 0.02791506238281727, + 0.02109513431787491, + 0.026067722588777542, + 0.04841037839651108, + 0.005910711362957954, + 0.005128089804202318, + 0.005470710806548595, + 0.004636356607079506, + 0.005583551712334156, + 0.006971618626266718, + 0.010723869316279888, + 0.007479558233171701, + 0.010480889119207859, + 0.012421262450516224, + 0.00776382302865386, + 0.008636817336082458, + 0.011566813103854656, + 0.013687987811863422, + 0.015925714746117592, + 0.020536983385682106, + 0.006659545935690403, + 0.009156553074717522, + 0.007546690292656422, + 0.007915769703686237, + 0.008871030993759632, + 0.011483002454042435, + 0.008574803359806538, + 0.01543794572353363, + 0.004536292981356382, + 0.005864246748387814, + 0.006475426722317934, + 0.005099165253341198, + 0.007046192418783903, + 0.008851965889334679, + 0.0056601632386446, + 0.010126031003892422, + 0.00405651843175292, + 0.005871026776731014, + 0.008232044987380505, + 0.009944862686097622, + 0.0050083426758646965, + 0.010253649204969406, + 0.007245670538395643, + 0.009235233068466187, + 0.013246482238173485, + 0.017397938296198845, + 0.016491470858454704, + 0.018383348360657692, + 0.016313917934894562, + 0.030389908701181412, + 0.050881657749414444, + 0.017787504941225052, + 0.0035365945659577847, + 0.004730896092951298, + 0.006262824870646, + 0.004824880976229906, + 0.005067971535027027, + 0.005703390575945377, + 0.006037039216607809, + 0.0070191556587815285, + 0.0072844806127250195, + 0.018857836723327637, + 0.026863593608140945, + 0.013027357868850231, + 0.010037622414529324, + 0.011493727564811707, + 0.006966743618249893, + 0.008267243392765522, + 0.012269468046724796, + 0.027297360822558403, + 0.02336597442626953, + 0.023302951827645302, + 0.014175823889672756, + 0.014928121119737625, + 0.009589067660272121, + 0.010152824223041534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.027975162491202354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905025571584702 + ], + "opposite_logprob_tokens": [ + -2.46405029296875 + ], + "logprob_reward": -0.08905025571584702, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905025571584702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08503416180610657, + 0.025997407734394073, + 0.05161915346980095, + 0.027424635365605354, + 0.040372271090745926, + 0.0851321592926979, + 0.1820528209209442, + 0.12103883177042007, + 0.3063531816005707, + 0.0749753788113594 + ], + "credit_map": [ + 0.007603724952787161, + 0.010026980191469193, + 0.007137825712561607, + 0.010474396869540215, + 0.01562771573662758, + 0.009338895790278912, + 0.00871444121003151, + 0.01611017994582653, + 0.0020356811583042145, + 0.002851025899872184, + 0.0024205294903367758, + 0.0025865123607218266, + 0.0033464536536484957, + 0.004398945719003677, + 0.004755023866891861, + 0.003603236749768257, + 0.0031597407069057226, + 0.004562458489090204, + 0.0032778300810605288, + 0.003677456174045801, + 0.0050021447241306305, + 0.006590158678591251, + 0.013042168691754341, + 0.012307196855545044, + 0.00320030702278018, + 0.003375339088961482, + 0.0037175205070525408, + 0.0029599072877317667, + 0.003317747265100479, + 0.003990016411989927, + 0.0027663996443152428, + 0.004097397439181805, + 0.003251797752454877, + 0.0037476695142686367, + 0.0033387586008757353, + 0.0032580718398094177, + 0.007676177192479372, + 0.006972695700824261, + 0.005307930521667004, + 0.006819167640060186, + 0.006123737897723913, + 0.011587904766201973, + 0.009500013664364815, + 0.01772492751479149, + 0.007400635164231062, + 0.014892254956066608, + 0.008726964704692364, + 0.00917571410536766, + 0.01218096911907196, + 0.013898923993110657, + 0.017111744731664658, + 0.01707557775080204, + 0.015768811106681824, + 0.032797615975141525, + 0.05213399603962898, + 0.021085185930132866, + 0.009495425969362259, + 0.014193005859851837, + 0.01633322983980179, + 0.022711876779794693, + 0.015616166405379772, + 0.015121224336326122, + 0.014078876934945583, + 0.013489019125699997, + 0.018694715574383736, + 0.04462728276848793, + 0.0480683371424675, + 0.03661840409040451, + 0.04226701334118843, + 0.066556416451931, + 0.01926138624548912, + 0.0302596315741539, + 0.006882194429636002, + 0.008679332211613655, + 0.010208644904196262, + 0.009904741309583187, + 0.008993183262646198, + 0.012250378727912903, + 0.009493265300989151, + 0.008563634008169174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.08905025571584702 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011826909612864256 + ], + "opposite_logprob_tokens": [ + -6.751182556152344 + ], + "logprob_reward": -0.0011826909612864256, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011826909612864256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23224256932735443, + 0.06082926690578461, + 0.11249777674674988, + 0.20485882461071014, + 0.07652875781059265, + 0.0495784655213356, + 0.06200818344950676, + 0.043193183839321136, + 0.0597497895359993, + 0.09851328283548355 + ], + "credit_map": [ + 0.01697777956724167, + 0.012518836185336113, + 0.014307243749499321, + 0.0121848089620471, + 0.05851659178733826, + 0.02771722711622715, + 0.020003072917461395, + 0.07001698762178421, + 0.00601585628464818, + 0.010241490788757801, + 0.00692386319860816, + 0.006868074648082256, + 0.007381032221019268, + 0.006917490158230066, + 0.007495556958019733, + 0.008985898457467556, + 0.016859911382198334, + 0.014598722569644451, + 0.011445295065641403, + 0.008151881396770477, + 0.015440769493579865, + 0.013465783558785915, + 0.016350097954273224, + 0.016185306012630463, + 0.03156178072094917, + 0.02772616781294346, + 0.026531586423516273, + 0.021456191316246986, + 0.025912921875715256, + 0.02884572185575962, + 0.015793418511748314, + 0.02703101374208927, + 0.007003443781286478, + 0.011043976061046124, + 0.008531780913472176, + 0.008745575323700905, + 0.008386447094380856, + 0.01204646285623312, + 0.007588994689285755, + 0.013182067312300205, + 0.004480109084397554, + 0.0060974350199103355, + 0.006403949111700058, + 0.008585216477513313, + 0.004617475438863039, + 0.007326936349272728, + 0.006281048990786076, + 0.005786286201328039, + 0.00413499865680933, + 0.005518027115613222, + 0.008631961420178413, + 0.007331929635256529, + 0.006390559487044811, + 0.010534185916185379, + 0.012297319248318672, + 0.007169198710471392, + 0.003000676864758134, + 0.003489540424197912, + 0.007104304153472185, + 0.007054126355797052, + 0.005226567387580872, + 0.0056452034041285515, + 0.0053562382236123085, + 0.006316521670669317, + 0.0048792921006679535, + 0.008327807299792767, + 0.008805884048342705, + 0.009544589556753635, + 0.0067614600993692875, + 0.011212179437279701, + 0.004165711812674999, + 0.006052861921489239, + 0.007894131354987621, + 0.017261574044823647, + 0.019575199112296104, + 0.01346905343234539, + 0.010835184715688229, + 0.010559515096247196, + 0.009201951324939728, + 0.00971666444092989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0011826909612864256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036165802739560604 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036165802739560604, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036165802739560604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2549482583999634, + 0.04313718527555466, + 0.07451502233743668, + 0.11455769091844559, + 0.04923520237207413, + 0.04234980419278145, + 0.08700396120548248, + 0.06359463930130005, + 0.10148371756076813, + 0.1691744178533554 + ], + "credit_map": [ + 0.015604522079229355, + 0.009636455215513706, + 0.011773509904742241, + 0.01098292414098978, + 0.06651100516319275, + 0.031084217131137848, + 0.02466658689081669, + 0.08468908071517944, + 0.008829314261674881, + 0.004223158583045006, + 0.004747556056827307, + 0.0033006849698722363, + 0.003591360291466117, + 0.004731758497655392, + 0.00751511100679636, + 0.0061982497572898865, + 0.008981654420495033, + 0.006622144021093845, + 0.006393561139702797, + 0.0052826846949756145, + 0.015079685486853123, + 0.008217256516218185, + 0.012670698575675488, + 0.011267347261309624, + 0.016854211688041687, + 0.013981294818222523, + 0.015312870033085346, + 0.011665688827633858, + 0.014144090935587883, + 0.0164266899228096, + 0.01040378212928772, + 0.015769077464938164, + 0.004254672210663557, + 0.007616933900862932, + 0.006111483089625835, + 0.007244215812534094, + 0.006704830564558506, + 0.007382507435977459, + 0.00427005160599947, + 0.005650513805449009, + 0.003350034588947892, + 0.004226204473525286, + 0.00761612830683589, + 0.00558666093274951, + 0.004401107784360647, + 0.006858494598418474, + 0.005226655397564173, + 0.00508452020585537, + 0.006074642296880484, + 0.0069468216970562935, + 0.007151069585233927, + 0.00986095517873764, + 0.008160114288330078, + 0.015281865373253822, + 0.022357195615768433, + 0.011171307414770126, + 0.004930940922349691, + 0.006942871492356062, + 0.009515943005681038, + 0.008301218040287495, + 0.007849090732634068, + 0.010009179823100567, + 0.005899928975850344, + 0.010145474225282669, + 0.009207132272422314, + 0.01516773086041212, + 0.015791524201631546, + 0.01682879403233528, + 0.013903200626373291, + 0.013386592268943787, + 0.006736111361533403, + 0.01046264823526144, + 0.013372303918004036, + 0.02977660298347473, + 0.02812895178794861, + 0.02941916510462761, + 0.014612733386456966, + 0.017608465626835823, + 0.016796555370092392, + 0.019459659233689308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0036165802739560604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a single, uninterrupted shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive six-legged beast with an arthropod-like exoskeleton visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the beast's exoskeleton deep blue and purple with bioluminescent veins, and does it have a head with feathery tentacles instead of eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the beast's turquoise-tipped tentacles and orange-haloed fungi in a nighttime jungle environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017122859135270119 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017122859135270119, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017122859135270119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3782736659049988, + 0.0857718363404274, + 0.12719488143920898, + 0.0661819651722908, + 0.045858707278966904, + 0.03518103063106537, + 0.05231154337525368, + 0.04058919847011566, + 0.07939473539590836, + 0.08924241364002228 + ], + "credit_map": [ + 0.020632313564419746, + 0.022133836522698402, + 0.01690337248146534, + 0.024476123973727226, + 0.08560287952423096, + 0.046313777565956116, + 0.041109777987003326, + 0.12110156565904617, + 0.010304136201739311, + 0.005942474585026503, + 0.010159801691770554, + 0.006336294114589691, + 0.007109495811164379, + 0.013168349862098694, + 0.018882911652326584, + 0.013868369162082672, + 0.009425358846783638, + 0.010708306916058064, + 0.010277792811393738, + 0.00852122437208891, + 0.02771192416548729, + 0.015555613674223423, + 0.020970042794942856, + 0.024024613201618195, + 0.010205845348536968, + 0.008567688055336475, + 0.007653187960386276, + 0.006045101210474968, + 0.009034129790961742, + 0.010038129985332489, + 0.0056436737067997456, + 0.008994204923510551, + 0.0026937949005514383, + 0.0031882114708423615, + 0.003322241362184286, + 0.004144284874200821, + 0.006488797254860401, + 0.008675266988575459, + 0.006181519012898207, + 0.011164587922394276, + 0.0027702772058546543, + 0.0033728787675499916, + 0.0059102498926222324, + 0.005677486304193735, + 0.003000354627147317, + 0.005865122191607952, + 0.004589470569044352, + 0.003995190374553204, + 0.003314673202112317, + 0.004884795285761356, + 0.007662015035748482, + 0.005355654284358025, + 0.008872142992913723, + 0.008348232135176659, + 0.007676995359361172, + 0.006197034846991301, + 0.0038605232257395983, + 0.004027002956718206, + 0.005264720879495144, + 0.0046745790168643, + 0.004949023015797138, + 0.005262838676571846, + 0.0062410179525613785, + 0.006309493910521269, + 0.004834622610360384, + 0.01191948726773262, + 0.009825434535741806, + 0.010179202072322369, + 0.007314363960176706, + 0.020909322425723076, + 0.0039032287895679474, + 0.010509076528251171, + 0.009054633788764477, + 0.01227875892072916, + 0.014718817546963692, + 0.01705826073884964, + 0.006912951357662678, + 0.007545021362602711, + 0.013048063032329082, + 0.008625902235507965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0017122859135270119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_tensor([7531, 7531])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_tensor([7531, 7531])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a6a6ea7993b73959cf614cc2011b0ad2d4589745 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_tensor([7531, 7531])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a tight, close-up, static shot taken from a slight overhead angle, a pair of deft hands in flawless white gloves add the finishing touches. The gloves' texture is delicate and smooth, as if meticulously sculpted from pure white icing, perfectly echoing the fluid lines of a fairytale illustration. They intently handle a Cantonese-style mooncake, fresh from the oven, a captivating golden hue. The intricate auspicious cloud pattern on its surface appears three-dimensional and vivid under the soft sidelight. One hand gently supports the mooncake, while the other, using a pair of delicate tweezers, delicately grasps a thin piece of gold foil and, with steady deliberation, precisely places it on the center pistil of the mooncake. Warm light diffused from one side of the frame casts a honey-hued glow over the scene, creating an atmosphere of tranquility, concentration, and exquisite beauty. The overall color palette is soft and warm, imbued with the dreamy and sweet beauty of a fairy tale.", + "seed": 30708, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_7531_seed_30708_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01636392413638532, + "vlm_logprob": -0.01636392413638532, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.01636392413638532, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023842062801122665 + ], + "opposite_logprob_tokens": [ + -3.8988420963287354 + ], + "logprob_reward": -0.023842062801122665, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023842062801122665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10183148831129074, + 0.09341602772474289, + 0.16637007892131805, + 0.0511801615357399, + 0.4821379482746124, + 0.026433255523443222, + 0.037761490792036057, + 0.008847154676914215, + 0.013458836823701859, + 0.018563589081168175 + ], + "credit_map": [ + 0.011614068411290646, + 0.007358472794294357, + 0.009693542495369911, + 0.009555313736200333, + 0.01365101058036089, + 0.017602482810616493, + 0.010718956589698792, + 0.021637629717588425, + 0.00884255301207304, + 0.006597473751753569, + 0.0066554248332977295, + 0.009546998888254166, + 0.02732851170003414, + 0.011333663016557693, + 0.013861879706382751, + 0.009249517694115639, + 0.0254739448428154, + 0.004431101027876139, + 0.027617469429969788, + 0.004478471353650093, + 0.011340776458382607, + 0.07204995304346085, + 0.0065129604190588, + 0.014465391635894775, + 0.010738101787865162, + 0.005686827003955841, + 0.006971814204007387, + 0.006653092335909605, + 0.003371048951521516, + 0.005353210028260946, + 0.006761288270354271, + 0.0056447735987603664, + 0.06543559581041336, + 0.024745285511016846, + 0.12285968661308289, + 0.029610414057970047, + 0.02561962977051735, + 0.05730954557657242, + 0.034694328904151917, + 0.1218634694814682, + 0.0036652027629315853, + 0.001970625249668956, + 0.00622640922665596, + 0.001933310879394412, + 0.0018432491924613714, + 0.0025824655313044786, + 0.002272677607834339, + 0.005939314141869545, + 0.005105103366076946, + 0.0010495830792933702, + 0.006238273344933987, + 0.0013010491384193301, + 0.0035940702073276043, + 0.01267325971275568, + 0.0024967119097709656, + 0.0053034392185509205, + 0.0010310540674254298, + 0.0005969571648165584, + 0.0007743492024019361, + 0.0007786927744746208, + 0.001100068213418126, + 0.0015214881859719753, + 0.0018130794633179903, + 0.0012314645573496819, + 0.0013479095650836825, + 0.0005504456348717213, + 0.0012039884459227324, + 0.0010177524527534842, + 0.002224132651463151, + 0.0013757870765402913, + 0.003579833311960101, + 0.0021589857060462236, + 0.0020612231455743313, + 0.0017337350873276591, + 0.0017964002909138799, + 0.0014488918241113424, + 0.0028295114170759916, + 0.002277866704389453, + 0.0025623193942010403, + 0.0038536384236067533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.023842062801122665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020631704479455948 + ], + "opposite_logprob_tokens": [ + -3.895631790161133 + ], + "logprob_reward": -0.020631704479455948, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020631704479455948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21618826687335968, + 0.10666993260383606, + 0.16774658858776093, + 0.1131238266825676, + 0.18845346570014954, + 0.027956921607255936, + 0.05683102831244469, + 0.02523951418697834, + 0.033390067517757416, + 0.06440038233995438 + ], + "credit_map": [ + 0.022755173966288567, + 0.015876321122050285, + 0.018992409110069275, + 0.02458297833800316, + 0.032355859875679016, + 0.031322114169597626, + 0.02760700136423111, + 0.04269639775156975, + 0.011352869682013988, + 0.006962037645280361, + 0.008594317361712456, + 0.006755738519132137, + 0.031087489798665047, + 0.017170986160635948, + 0.012994558550417423, + 0.01175193302333355, + 0.021056726574897766, + 0.0042564235627651215, + 0.017141427844762802, + 0.0050595588982105255, + 0.011385691352188587, + 0.05099580064415932, + 0.011461238376796246, + 0.04638972133398056, + 0.0340220183134079, + 0.008708552457392216, + 0.024900471791625023, + 0.006315693724900484, + 0.004915035795420408, + 0.01196699496358633, + 0.007842488586902618, + 0.014452584087848663, + 0.04683482274413109, + 0.011594289913773537, + 0.05165969952940941, + 0.01130196638405323, + 0.009024616330862045, + 0.020387211814522743, + 0.017036043107509613, + 0.020614830777049065, + 0.0038895057514309883, + 0.0020459864754229784, + 0.0035820745397359133, + 0.0020517983939498663, + 0.0018234484596177936, + 0.004524990916252136, + 0.006168011575937271, + 0.0038711028173565865, + 0.005252165254205465, + 0.0021051540970802307, + 0.010494091548025608, + 0.0023357791360467672, + 0.005310161970555782, + 0.02227453887462616, + 0.004475084599107504, + 0.004584053065627813, + 0.00291829532943666, + 0.002625954570248723, + 0.0022587841376662254, + 0.0018991198157891631, + 0.003752937074750662, + 0.0028919053729623556, + 0.0056867278181016445, + 0.0032057880889624357, + 0.0045010242611169815, + 0.0022279464174062014, + 0.0030822993721812963, + 0.0023452506866306067, + 0.004421428311616182, + 0.005357674323022366, + 0.005347093567252159, + 0.0061073508113622665, + 0.010604631155729294, + 0.005025969352573156, + 0.006261968053877354, + 0.004641566891223192, + 0.012069241143763065, + 0.008986772038042545, + 0.009117714129388332, + 0.007692517712712288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.020631704479455948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232852790504694 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.0232852790504694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232852790504694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055469121783971786, + 0.07034683227539062, + 0.20179708302021027, + 0.14710462093353271, + 0.33726030588150024, + 0.019608810544013977, + 0.060559824109077454, + 0.015318199060857296, + 0.025984032079577446, + 0.06655119359493256 + ], + "credit_map": [ + 0.008801430463790894, + 0.0046532610431313515, + 0.006187043152749538, + 0.004491972737014294, + 0.0071957902982831, + 0.008869991637766361, + 0.0065608154982328415, + 0.008708825334906578, + 0.006340933497995138, + 0.005372205283492804, + 0.005566165782511234, + 0.005154743324965239, + 0.020377296954393387, + 0.009232793934643269, + 0.009916783310472965, + 0.008385914377868176, + 0.024548854678869247, + 0.004078539554029703, + 0.021062878891825676, + 0.004096922464668751, + 0.008232569321990013, + 0.08697432279586792, + 0.006831018254160881, + 0.04597198963165283, + 0.036428190767765045, + 0.007510303054004908, + 0.039561569690704346, + 0.008107268251478672, + 0.008474166505038738, + 0.012871685437858105, + 0.011226357892155647, + 0.022925088182091713, + 0.11024303734302521, + 0.01818249747157097, + 0.05771725997328758, + 0.010367301292717457, + 0.014376566745340824, + 0.03510447219014168, + 0.020575013011693954, + 0.07069416344165802, + 0.0025408389046788216, + 0.0012006842298433185, + 0.0026838239282369614, + 0.0014154567616060376, + 0.0015810191398486495, + 0.003339301561936736, + 0.0040212469175457954, + 0.0028264399152249098, + 0.005211223848164082, + 0.0024674220476299524, + 0.0053316764533519745, + 0.0030137519352138042, + 0.005976472981274128, + 0.021318109706044197, + 0.0074211903847754, + 0.009819977916777134, + 0.002864977577701211, + 0.001079607056453824, + 0.0017652057576924562, + 0.0014337020693346858, + 0.0016031281556934118, + 0.002096693031489849, + 0.0023702888283878565, + 0.002104596933349967, + 0.0020368436817079782, + 0.0015243481611832976, + 0.002305848989635706, + 0.0025290152989327908, + 0.005669719073921442, + 0.002597686368972063, + 0.006249743979424238, + 0.003070827340707183, + 0.008774250745773315, + 0.006122628226876259, + 0.008432191796600819, + 0.004200026858597994, + 0.009322967380285263, + 0.009008326567709446, + 0.007895088754594326, + 0.012795718386769295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.0232852790504694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015170981641858816 + ], + "opposite_logprob_tokens": [ + -6.501517295837402 + ], + "logprob_reward": -0.0015170981641858816, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015170981641858816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1184823289513588, + 0.0614008903503418, + 0.11204429715871811, + 0.024153273552656174, + 0.36051124334335327, + 0.15616685152053833, + 0.0515899583697319, + 0.024801360443234444, + 0.027608687058091164, + 0.06324107944965363 + ], + "credit_map": [ + 0.01551011297851801, + 0.009268014691770077, + 0.013656173832714558, + 0.010991949588060379, + 0.015234718099236488, + 0.021261759102344513, + 0.011159725487232208, + 0.021399876102805138, + 0.006595575250685215, + 0.007046711631119251, + 0.006732199341058731, + 0.006862999405711889, + 0.0065724216401577, + 0.009812938049435616, + 0.00839484017342329, + 0.00938321091234684, + 0.013211347162723541, + 0.0030365169513970613, + 0.02155701443552971, + 0.0032765960786491632, + 0.006392777897417545, + 0.051043249666690826, + 0.0033769800793379545, + 0.010149814188480377, + 0.00264953076839447, + 0.001573840738274157, + 0.002528485609218478, + 0.002782328985631466, + 0.0026756753213703632, + 0.004873592872172594, + 0.0033783335238695145, + 0.0036914858501404524, + 0.12140609323978424, + 0.019231781363487244, + 0.09940791875123978, + 0.01467571035027504, + 0.01180570013821125, + 0.0307187307626009, + 0.025351708754897118, + 0.037913598120212555, + 0.030652116984128952, + 0.006881821434944868, + 0.036340657621622086, + 0.008851878345012665, + 0.00840315967798233, + 0.02181011252105236, + 0.015193955972790718, + 0.028033152222633362, + 0.004871181678026915, + 0.0020311735570430756, + 0.007751394994556904, + 0.003181245643645525, + 0.005463423673063517, + 0.018184753134846687, + 0.00551986787468195, + 0.0045869192108511925, + 0.00423954613506794, + 0.0012533855624496937, + 0.0031676776707172394, + 0.0028183443937450647, + 0.002259714063256979, + 0.003035866655409336, + 0.003973169252276421, + 0.004053656477481127, + 0.004841470159590244, + 0.0016531959408894181, + 0.0028546564280986786, + 0.0015174539294093847, + 0.003130959579721093, + 0.0035952210891991854, + 0.0031922163907438517, + 0.006823512259870768, + 0.00760812358930707, + 0.005328675266355276, + 0.006906510796397924, + 0.0040611326694488525, + 0.010472460649907589, + 0.008750895969569683, + 0.007856967858970165, + 0.012256313115358353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0015170981641858816 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09117068350315094, + 0.06626495718955994, + 0.07400872558355331, + 0.04880998283624649, + 0.28520482778549194, + 0.0956706553697586, + 0.15656878054141998, + 0.03238047659397125, + 0.06650017946958542, + 0.0834207609295845 + ], + "credit_map": [ + 0.01652391441166401, + 0.004491471219807863, + 0.009128866717219353, + 0.005757019855082035, + 0.011461653746664524, + 0.012163124978542328, + 0.012634295970201492, + 0.019010333344340324, + 0.006954566575586796, + 0.004518799018114805, + 0.006302765570580959, + 0.00670910207554698, + 0.016040289774537086, + 0.010963710024952888, + 0.006431858986616135, + 0.008343854919075966, + 0.01578318513929844, + 0.0023899574298411608, + 0.008687707595527172, + 0.0027065016329288483, + 0.0036519570276141167, + 0.02306310087442398, + 0.003544396022334695, + 0.014181914739310741, + 0.0067269629798829556, + 0.003015156602486968, + 0.009390058927237988, + 0.0038175536319613457, + 0.002963907551020384, + 0.006076170597225428, + 0.0060019586235284805, + 0.010818211361765862, + 0.05169828608632088, + 0.009825841523706913, + 0.0775832012295723, + 0.013350892812013626, + 0.011363997124135494, + 0.024835234507918358, + 0.02440851368010044, + 0.07213883101940155, + 0.01232805848121643, + 0.005937757436186075, + 0.012652644887566566, + 0.004671250935643911, + 0.006122967693954706, + 0.012238818220794201, + 0.011065761558711529, + 0.030653389170765877, + 0.014837682247161865, + 0.00429011695086956, + 0.031146567314863205, + 0.004348379094153643, + 0.015214676968753338, + 0.066823810338974, + 0.008185873739421368, + 0.011721665039658546, + 0.006180429831147194, + 0.0017383851809427142, + 0.004089508205652237, + 0.0024840643163770437, + 0.0030166308861225843, + 0.00488306162878871, + 0.0043984875082969666, + 0.0055899047292768955, + 0.010291523300111294, + 0.002938613761216402, + 0.006715941242873669, + 0.003644772106781602, + 0.007040789816528559, + 0.009522880427539349, + 0.008589359931647778, + 0.01775628887116909, + 0.018489554524421692, + 0.007684547454118729, + 0.013411825522780418, + 0.004468876402825117, + 0.010946355760097504, + 0.012062896974384785, + 0.008727538399398327, + 0.007629155647009611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.012543476186692715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tight, close-up, static shot taken from a slight overhead angle, a pair of deft hands in flawless white gloves add the finishing touches. The gloves' texture is delicate and smooth, as if meticulously sculpted from pure white icing, perfectly echoing the fluid lines of a fairytale illustration. They intently handle a Cantonese-style mooncake, fresh from the oven, a captivating golden hue. The intricate auspicious cloud pattern on its surface appears three-dimensional and vivid under the soft sidelight. One hand gently supports the mooncake, while the other, using a pair of delicate tweezers, delicately grasps a thin piece of gold foil and, with steady deliberation, precisely places it on the center pistil of the mooncake. Warm light diffused from one side of the frame casts a honey-hued glow over the scene, creating an atmosphere of tranquility, concentration, and exquisite beauty. The overall color palette is soft and warm, imbued with the dreamy and sweet beauty of a fairy tale.", + "seed": 30709, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_7531_seed_30709_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017944401921704413, + "vlm_logprob": -0.017944401921704413, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.017944401921704413, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014729914255440235 + ], + "opposite_logprob_tokens": [ + -4.514729976654053 + ], + "logprob_reward": -0.014729914255440235, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014729914255440235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33283141255378723, + 0.18618975579738617, + 0.09696416556835175, + 0.06441865861415863, + 0.08706080913543701, + 0.02230783738195896, + 0.05238578841090202, + 0.04256082698702812, + 0.06607447564601898, + 0.049206312745809555 + ], + "credit_map": [ + 0.031730685383081436, + 0.03034788742661476, + 0.022823108360171318, + 0.033665433526039124, + 0.0479438342154026, + 0.09309779852628708, + 0.040176451206207275, + 0.0330461710691452, + 0.017501242458820343, + 0.01693466119468212, + 0.010291431099176407, + 0.020862478762865067, + 0.02998458594083786, + 0.04982984438538551, + 0.022033091634511948, + 0.018752403557300568, + 0.008998892270028591, + 0.009059193544089794, + 0.0052310749888420105, + 0.007996579632163048, + 0.015166708268225193, + 0.020309537649154663, + 0.014912785030901432, + 0.015289376489818096, + 0.005501629784703255, + 0.006466485559940338, + 0.004348109010607004, + 0.006813766900449991, + 0.009270421229302883, + 0.014273569919168949, + 0.009255174547433853, + 0.008489500731229782, + 0.008893397636711597, + 0.009199176914989948, + 0.004634961485862732, + 0.01099796500056982, + 0.01492711715400219, + 0.018259450793266296, + 0.010514087043702602, + 0.009634644724428654, + 0.002838647924363613, + 0.0018541847821325064, + 0.0020164288580417633, + 0.0018202076898887753, + 0.003407138865441084, + 0.003585247090086341, + 0.004018164239823818, + 0.0027678152546286583, + 0.007904873229563236, + 0.004923981614410877, + 0.004506134428083897, + 0.0035283167380839586, + 0.006228527519851923, + 0.00950270239263773, + 0.007894721813499928, + 0.007896524854004383, + 0.005486337002366781, + 0.004271790385246277, + 0.0030793517362326384, + 0.003806971712037921, + 0.004674493335187435, + 0.009655969217419624, + 0.00608932925388217, + 0.005496578756719828, + 0.0068396106362342834, + 0.005886449944227934, + 0.005199892912060022, + 0.00842661689966917, + 0.012419640086591244, + 0.010173294693231583, + 0.00899635162204504, + 0.00813260767608881, + 0.005679289810359478, + 0.004252189304679632, + 0.003600014140829444, + 0.00521113583818078, + 0.009444354102015495, + 0.005469166673719883, + 0.009224464185535908, + 0.006325696595013142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014729914255440235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020631004124879837 + ], + "opposite_logprob_tokens": [ + -3.8956310749053955 + ], + "logprob_reward": -0.020631004124879837, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020631004124879837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09617045521736145, + 0.13442811369895935, + 0.09942443668842316, + 0.08457282930612564, + 0.028227636590600014, + 0.04011191427707672, + 0.03580852970480919, + 0.04703699052333832, + 0.07579958438873291, + 0.3584195673465729 + ], + "credit_map": [ + 0.009418602101504803, + 0.008559716865420341, + 0.010312624275684357, + 0.009022542275488377, + 0.011493015103042126, + 0.01370751578360796, + 0.015524635091423988, + 0.01813180185854435, + 0.013196557760238647, + 0.012612774968147278, + 0.0096588721498847, + 0.014104070141911507, + 0.020837830379605293, + 0.0332028828561306, + 0.015338622964918613, + 0.015476482920348644, + 0.009076766669750214, + 0.008357399143278599, + 0.004948499612510204, + 0.009540114551782608, + 0.015779461711645126, + 0.025850679725408554, + 0.012615105137228966, + 0.013256409205496311, + 0.0074295662343502045, + 0.008694687858223915, + 0.005727687384933233, + 0.008724560029804707, + 0.012107860296964645, + 0.017728889361023903, + 0.01197691261768341, + 0.012182665057480335, + 0.0036808063741773367, + 0.002935683587566018, + 0.0015947548672556877, + 0.0028171816375106573, + 0.0046416958793997765, + 0.005510927177965641, + 0.0032582220155745745, + 0.00378836365416646, + 0.003792860545217991, + 0.0035825353115797043, + 0.0025736696552485228, + 0.004547568038105965, + 0.006652230862528086, + 0.0066622113808989525, + 0.007061013951897621, + 0.005239821504801512, + 0.00479560811072588, + 0.0042176139540970325, + 0.002757900394499302, + 0.0036736007314175367, + 0.0035793965216726065, + 0.004215708002448082, + 0.00473977392539382, + 0.007828927598893642, + 0.005166015587747097, + 0.005006605293601751, + 0.004840017296373844, + 0.004815909080207348, + 0.0046556973829865456, + 0.008249253034591675, + 0.008117854595184326, + 0.006185633130371571, + 0.00968222413212061, + 0.007906251586973667, + 0.005911567714065313, + 0.007183522451668978, + 0.01640062779188156, + 0.01085286121815443, + 0.00937987957149744, + 0.008482652716338634, + 0.041145723313093185, + 0.03133745491504669, + 0.020861897617578506, + 0.03424207121133804, + 0.07058437913656235, + 0.058400414884090424, + 0.06304007768630981, + 0.03880750760436058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.020631004124879837 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305654764175415 + ], + "opposite_logprob_tokens": [ + -3.1680564880371094 + ], + "logprob_reward": -0.04305654764175415, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305654764175415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30633965134620667, + 0.12909968197345734, + 0.13709674775600433, + 0.041080061346292496, + 0.0399402417242527, + 0.02376546710729599, + 0.0431368350982666, + 0.059685394167900085, + 0.10470975190401077, + 0.1151462271809578 + ], + "credit_map": [ + 0.028245707973837852, + 0.03600604087114334, + 0.02384578064084053, + 0.035866495221853256, + 0.040083013474941254, + 0.05815687030553818, + 0.041832540184259415, + 0.0423031710088253, + 0.013435956090688705, + 0.012911368161439896, + 0.011214653961360455, + 0.014257115311920643, + 0.017342330887913704, + 0.02641250193119049, + 0.01800636388361454, + 0.015519383363425732, + 0.013571451418101788, + 0.012139678932726383, + 0.012097286060452461, + 0.013591911643743515, + 0.021291213110089302, + 0.029210668057203293, + 0.017279841005802155, + 0.017914677038788795, + 0.004210611805319786, + 0.004819536581635475, + 0.0036004986613988876, + 0.0044002048671245575, + 0.0059526534751057625, + 0.0063712275587022305, + 0.0060117607936263084, + 0.005713558755815029, + 0.005872635170817375, + 0.0033772329334169626, + 0.004287708085030317, + 0.002997024916112423, + 0.0047977217473089695, + 0.006136953830718994, + 0.00658824248239398, + 0.005882717203348875, + 0.0028509811963886023, + 0.002081005834043026, + 0.003002389334142208, + 0.0020531665068119764, + 0.002813283819705248, + 0.003244909690693021, + 0.00446159765124321, + 0.003258130280300975, + 0.006067153066396713, + 0.004263454116880894, + 0.003705789102241397, + 0.003608993487432599, + 0.003812365932390094, + 0.005176078993827105, + 0.00700058089569211, + 0.009502412751317024, + 0.008340277709066868, + 0.006318702828139067, + 0.004963202401995659, + 0.00448241550475359, + 0.006544989999383688, + 0.012959628365933895, + 0.008381971158087254, + 0.0076941996812820435, + 0.009918810799717903, + 0.011673234403133392, + 0.006902158260345459, + 0.011707350611686707, + 0.01779833436012268, + 0.011985188350081444, + 0.020013922825455666, + 0.014710744842886925, + 0.01877065747976303, + 0.010722501203417778, + 0.010750619694590569, + 0.01302134059369564, + 0.018450191244482994, + 0.013157413341104984, + 0.019215526059269905, + 0.011057966388761997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04305654764175415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015192406717687845 + ], + "opposite_logprob_tokens": [ + -6.501519203186035 + ], + "logprob_reward": -0.0015192406717687845, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015192406717687845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28313854336738586, + 0.12480715662240982, + 0.06418240815401077, + 0.05784213915467262, + 0.0585605725646019, + 0.04664502665400505, + 0.1042725071310997, + 0.07633417844772339, + 0.11905843764543533, + 0.06515904515981674 + ], + "credit_map": [ + 0.0373384952545166, + 0.028307508677244186, + 0.026265526190400124, + 0.028382746502757072, + 0.02993336319923401, + 0.04419761896133423, + 0.04930620267987251, + 0.03940706327557564, + 0.015945544466376305, + 0.010409127920866013, + 0.013498006388545036, + 0.01102472934871912, + 0.016319025307893753, + 0.020470380783081055, + 0.01841691881418228, + 0.01872340776026249, + 0.006753185298293829, + 0.005693449173122644, + 0.006595910061150789, + 0.007862146943807602, + 0.0072538056410849094, + 0.009031597524881363, + 0.010826007463037968, + 0.010166299529373646, + 0.0062349229119718075, + 0.004806437063962221, + 0.006038959138095379, + 0.004293283447623253, + 0.00451606186106801, + 0.008034645579755306, + 0.011173826642334461, + 0.012744003906846046, + 0.008686159737408161, + 0.006035250145941973, + 0.0051967184990644455, + 0.0056333886459469795, + 0.006369741167873144, + 0.006433113943785429, + 0.012444601394236088, + 0.007761597167700529, + 0.0072828480042517185, + 0.0044675725512206554, + 0.00541436392813921, + 0.0037041709292680025, + 0.004732489585876465, + 0.006286414805799723, + 0.008524088189005852, + 0.006233080290257931, + 0.013275779783725739, + 0.011769143864512444, + 0.009426874108612537, + 0.009985012002289295, + 0.014665650203824043, + 0.016501765698194504, + 0.013245274312794209, + 0.01540300715714693, + 0.013694771565496922, + 0.008525463752448559, + 0.008631390519440174, + 0.006036319304257631, + 0.005735273938626051, + 0.012042139656841755, + 0.010819490998983383, + 0.010849328711628914, + 0.014800848439335823, + 0.010372896678745747, + 0.009805653244256973, + 0.009160783141851425, + 0.01552473846822977, + 0.027773039415478706, + 0.015367813408374786, + 0.01625267043709755, + 0.008818809874355793, + 0.006975064519792795, + 0.006507135462015867, + 0.005304369609802961, + 0.010010545141994953, + 0.005199429113417864, + 0.012836352922022343, + 0.009507337585091591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0015192406717687845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978530291467905 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.00978530291467905, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978530291467905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2562563419342041, + 0.12305312603712082, + 0.09520624577999115, + 0.0762736052274704, + 0.08515874296426773, + 0.06888870894908905, + 0.036092307418584824, + 0.058667320758104324, + 0.08147835731506348, + 0.1189252957701683 + ], + "credit_map": [ + 0.036187853664159775, + 0.02864638715982437, + 0.039362505078315735, + 0.025924263522028923, + 0.027087124064564705, + 0.023381322622299194, + 0.03514453396201134, + 0.04052232950925827, + 0.016086285933852196, + 0.01134776882827282, + 0.0129570746794343, + 0.009283932857215405, + 0.012677891179919243, + 0.02413637936115265, + 0.01898834854364395, + 0.017575431615114212, + 0.01635976880788803, + 0.009087829850614071, + 0.011533060111105442, + 0.005893519148230553, + 0.006879017688333988, + 0.009560247883200645, + 0.01685761660337448, + 0.019035184755921364, + 0.010622267611324787, + 0.008021513931453228, + 0.00851678941398859, + 0.006276080850511789, + 0.006736825220286846, + 0.0069382269866764545, + 0.013760318979620934, + 0.015401574783027172, + 0.013463390059769154, + 0.008896324783563614, + 0.008925320580601692, + 0.007443526294082403, + 0.008132868446409702, + 0.009799561463296413, + 0.015992291271686554, + 0.012505460530519485, + 0.008329004049301147, + 0.006269918754696846, + 0.006861778907477856, + 0.005325573030859232, + 0.0068031176924705505, + 0.010277916677296162, + 0.01308396179229021, + 0.011937431059777737, + 0.00632478715851903, + 0.0037939336616545916, + 0.00522697763517499, + 0.00323933782055974, + 0.004791250452399254, + 0.00483421515673399, + 0.004253365099430084, + 0.003628439037129283, + 0.008115776814520359, + 0.005314824171364307, + 0.004932193551212549, + 0.005488826427608728, + 0.0062597645446658134, + 0.014028416946530342, + 0.007500702515244484, + 0.007026811596006155, + 0.010563001036643982, + 0.009117509238421917, + 0.008265115320682526, + 0.006885436829179525, + 0.013396576046943665, + 0.007245210465043783, + 0.012734660878777504, + 0.013270840048789978, + 0.01375969685614109, + 0.01191534660756588, + 0.010549389757215977, + 0.009864074178040028, + 0.016774112358689308, + 0.019345439970493317, + 0.02407045289874077, + 0.01264677383005619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.00978530291467905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tight, close-up, static shot taken from a slight overhead angle, a pair of deft hands in flawless white gloves add the finishing touches. The gloves' texture is delicate and smooth, as if meticulously sculpted from pure white icing, perfectly echoing the fluid lines of a fairytale illustration. They intently handle a Cantonese-style mooncake, fresh from the oven, a captivating golden hue. The intricate auspicious cloud pattern on its surface appears three-dimensional and vivid under the soft sidelight. One hand gently supports the mooncake, while the other, using a pair of delicate tweezers, delicately grasps a thin piece of gold foil and, with steady deliberation, precisely places it on the center pistil of the mooncake. Warm light diffused from one side of the frame casts a honey-hued glow over the scene, creating an atmosphere of tranquility, concentration, and exquisite beauty. The overall color palette is soft and warm, imbued with the dreamy and sweet beauty of a fairy tale.", + "seed": 30710, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_7531_seed_30710_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014278673473745585, + "vlm_logprob": -0.014278673473745585, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.014278673473745585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015044908970594406 + ], + "opposite_logprob_tokens": [ + -4.390044689178467 + ], + "logprob_reward": -0.015044908970594406, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015044908970594406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16432732343673706, + 0.05135129392147064, + 0.04386486858129501, + 0.07236702740192413, + 0.0469488725066185, + 0.0838480219244957, + 0.07192808389663696, + 0.28145673871040344, + 0.11624570935964584, + 0.0676620751619339 + ], + "credit_map": [ + 0.016749825328588486, + 0.032369546592235565, + 0.0240317489951849, + 0.018142754212021828, + 0.014792005531489849, + 0.01712951250374317, + 0.01686198078095913, + 0.02424994669854641, + 0.005202310159802437, + 0.00691479304805398, + 0.010489934124052525, + 0.005140888504683971, + 0.0063996706157922745, + 0.004496337380260229, + 0.006769085768610239, + 0.005938275251537561, + 0.009246737696230412, + 0.006105667911469936, + 0.007220656145364046, + 0.0043189991265535355, + 0.005535179283469915, + 0.002590605989098549, + 0.004712929483503103, + 0.004134092014282942, + 0.006941213738173246, + 0.012770944274961948, + 0.01314124371856451, + 0.007913706824183464, + 0.0077971904538571835, + 0.004929163958877325, + 0.009110713377594948, + 0.009762847796082497, + 0.006234742235392332, + 0.00785682164132595, + 0.010447708889842033, + 0.003929728642106056, + 0.005139667075127363, + 0.002601270331069827, + 0.006705520674586296, + 0.00403341231867671, + 0.009277628734707832, + 0.021279804408550262, + 0.01222122460603714, + 0.007433475460857153, + 0.010770809836685658, + 0.006667977198958397, + 0.007349926978349686, + 0.008847178891301155, + 0.00708193751052022, + 0.01827225461602211, + 0.01117030344903469, + 0.00727416854351759, + 0.009514190256595612, + 0.00449772085994482, + 0.0067923404276371, + 0.007325165439397097, + 0.03617018088698387, + 0.0581686906516552, + 0.042852774262428284, + 0.029547803103923798, + 0.04165840893983841, + 0.013341084122657776, + 0.03631768748164177, + 0.023400112986564636, + 0.014969767071306705, + 0.03027784638106823, + 0.025774961337447166, + 0.008657045662403107, + 0.01109064370393753, + 0.005763330962508917, + 0.007845520041882992, + 0.011866595596075058, + 0.01138641219586134, + 0.013666522689163685, + 0.015969151630997658, + 0.005445832386612892, + 0.0056563010439276695, + 0.003572392975911498, + 0.005363215692341328, + 0.0066022430546581745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.015044908970594406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026399720460176468 + ], + "opposite_logprob_tokens": [ + -3.651399612426758 + ], + "logprob_reward": -0.026399720460176468, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026399720460176468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2104090005159378, + 0.1873232126235962, + 0.04503801465034485, + 0.06275159120559692, + 0.19645921885967255, + 0.05591549351811409, + 0.041998568922281265, + 0.02011553943157196, + 0.0684855654835701, + 0.11150380969047546 + ], + "credit_map": [ + 0.025904014706611633, + 0.0411091111600399, + 0.02908712811768055, + 0.02443031594157219, + 0.024445753544569016, + 0.013960206881165504, + 0.0328822061419487, + 0.018590252846479416, + 0.015539012849330902, + 0.020552439615130424, + 0.026592036709189415, + 0.024853598326444626, + 0.027334751561284065, + 0.014148479327559471, + 0.03159685805439949, + 0.0267060324549675, + 0.01079089380800724, + 0.006266345269978046, + 0.0069029261358082294, + 0.004638001322746277, + 0.004275730345398188, + 0.003352481871843338, + 0.0046005211770534515, + 0.004211118444800377, + 0.006097821984440088, + 0.014169719070196152, + 0.009562653489410877, + 0.009107865393161774, + 0.007139064371585846, + 0.005244794767349958, + 0.006681270897388458, + 0.004748399835079908, + 0.017005741596221924, + 0.05419157072901726, + 0.05120764672756195, + 0.014087940566241741, + 0.020798459649086, + 0.010388302616775036, + 0.01646720990538597, + 0.012312357313930988, + 0.005144577007740736, + 0.010541546158492565, + 0.009013977833092213, + 0.00835357140749693, + 0.0075972178019583225, + 0.003707733005285263, + 0.0064752548933029175, + 0.005081613548099995, + 0.0038784898351877928, + 0.005824839696288109, + 0.00725224893540144, + 0.005285673774778843, + 0.006337624508887529, + 0.002683111699298024, + 0.006696769036352634, + 0.004039807710796595, + 0.002430184744298458, + 0.0026546029839664698, + 0.002888905117288232, + 0.0028727471362799406, + 0.002835480496287346, + 0.0017321077175438404, + 0.0028970506973564625, + 0.0018044611206278205, + 0.007917958311736584, + 0.016625918447971344, + 0.014794182032346725, + 0.005754770245403051, + 0.006844932679086924, + 0.0037759074475616217, + 0.004893026780337095, + 0.007878870703279972, + 0.013680621981620789, + 0.010074079036712646, + 0.019193921238183975, + 0.014970541000366211, + 0.01092876773327589, + 0.006927970331162214, + 0.01766352169215679, + 0.01806439273059368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.026399720460176468 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067780554294586 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067780554294586, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067780554294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23842686414718628, + 0.1189277395606041, + 0.08254453539848328, + 0.1095002144575119, + 0.0750693678855896, + 0.14038264751434326, + 0.0566396564245224, + 0.047240935266017914, + 0.07214681059122086, + 0.059121180325746536 + ], + "credit_map": [ + 0.032927002757787704, + 0.03820423036813736, + 0.03670765832066536, + 0.024045430123806, + 0.02690250799059868, + 0.01894170604646206, + 0.029652992263436317, + 0.031045351177453995, + 0.01265153381973505, + 0.015972498804330826, + 0.017482569441199303, + 0.01496668066829443, + 0.011813216842710972, + 0.011873556300997734, + 0.015979811549186707, + 0.018187878653407097, + 0.013094132766127586, + 0.013987475074827671, + 0.011516216211020947, + 0.00833131279796362, + 0.009266038425266743, + 0.005278714466840029, + 0.009628022089600563, + 0.011442629620432854, + 0.010408801026642323, + 0.03226664289832115, + 0.022961698472499847, + 0.009445872157812119, + 0.009367710910737514, + 0.0063595459796488285, + 0.0075693754479289055, + 0.011120571754872799, + 0.007203223183751106, + 0.013310850597918034, + 0.014284620061516762, + 0.007775219157338142, + 0.00862873625010252, + 0.004343507345765829, + 0.012908540666103363, + 0.006614671554416418, + 0.014141304418444633, + 0.03012612834572792, + 0.020586766302585602, + 0.014538463205099106, + 0.018330959603190422, + 0.014038274064660072, + 0.012436943128705025, + 0.016183817759156227, + 0.005400847177952528, + 0.00989053025841713, + 0.008397163823246956, + 0.006510880310088396, + 0.006815429776906967, + 0.0034714548382908106, + 0.0054269591346383095, + 0.010726395063102245, + 0.0054707713425159454, + 0.008887472562491894, + 0.007444570306688547, + 0.004950296133756638, + 0.0076468875631690025, + 0.0025431339163333178, + 0.005294201895594597, + 0.005003604106605053, + 0.007875092327594757, + 0.016851183027029037, + 0.013532462529838085, + 0.0049811480566859245, + 0.00786216277629137, + 0.0036741679068654776, + 0.008354630321264267, + 0.009015965275466442, + 0.005371718667447567, + 0.015833398327231407, + 0.01231915783137083, + 0.0050317877903580666, + 0.0033300661016255617, + 0.003023585071787238, + 0.0055487994104623795, + 0.008662670850753784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.016067780554294586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013386583887040615 + ], + "opposite_logprob_tokens": [ + -6.626338481903076 + ], + "logprob_reward": -0.0013386583887040615, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013386583887040615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1507118046283722, + 0.05064372345805168, + 0.03928862884640694, + 0.0649089589715004, + 0.42043977975845337, + 0.047056883573532104, + 0.04275475814938545, + 0.05850605666637421, + 0.08349563926458359, + 0.04219372197985649 + ], + "credit_map": [ + 0.01948355697095394, + 0.03754165768623352, + 0.024840181693434715, + 0.013353030197322369, + 0.01112386491149664, + 0.010654020123183727, + 0.012744499370455742, + 0.02097100391983986, + 0.005502893589437008, + 0.006392219569534063, + 0.00818535778671503, + 0.005852279253304005, + 0.006084085907787085, + 0.004226373042911291, + 0.0077682617120444775, + 0.006632254458963871, + 0.0044106850400567055, + 0.004977044649422169, + 0.0064361244440078735, + 0.004914938472211361, + 0.0053678350523114204, + 0.0032817781902849674, + 0.0042311870492994785, + 0.005669042002409697, + 0.009142907336354256, + 0.013156556524336338, + 0.009508495219051838, + 0.006929086521267891, + 0.007780652958899736, + 0.004508578218519688, + 0.0063962568528950214, + 0.0074864281341433525, + 0.033447012305259705, + 0.12728482484817505, + 0.12149181216955185, + 0.02914341539144516, + 0.04543014615774155, + 0.018233316019177437, + 0.02378804050385952, + 0.021621262654662132, + 0.00563901336863637, + 0.010517430491745472, + 0.007506876718252897, + 0.005567084066569805, + 0.006466650869697332, + 0.0030657064635306597, + 0.004392377566546202, + 0.003901749150827527, + 0.004529136698693037, + 0.005390595179051161, + 0.006572941318154335, + 0.005192971788346767, + 0.0060157389380037785, + 0.0031838032882660627, + 0.006309892050921917, + 0.0055596837773919106, + 0.007143750321120024, + 0.009759812615811825, + 0.009163688868284225, + 0.0059599378146231174, + 0.009661325253546238, + 0.003382938215509057, + 0.008745629340410233, + 0.0046889763325452805, + 0.008716985583305359, + 0.01901489496231079, + 0.016589000821113586, + 0.007986093871295452, + 0.011347216553986073, + 0.004460206255316734, + 0.006981744430959225, + 0.00839950144290924, + 0.005113402847200632, + 0.008589166216552258, + 0.007434412837028503, + 0.0037989276461303234, + 0.004107939079403877, + 0.0029746973887085915, + 0.003822934115305543, + 0.006352245807647705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0013386583887040615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125422989949584 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.0125422989949584, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125422989949584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13931971788406372, + 0.04952509328722954, + 0.04931060969829559, + 0.15023498237133026, + 0.08662505447864532, + 0.05120127275586128, + 0.025177085772156715, + 0.04461007937788963, + 0.31254369020462036, + 0.09145241975784302 + ], + "credit_map": [ + 0.019210442900657654, + 0.036748819053173065, + 0.0206377524882555, + 0.013969704508781433, + 0.013092977926135063, + 0.009626733139157295, + 0.012936854735016823, + 0.013096441514790058, + 0.005692663602530956, + 0.006693144794553518, + 0.007746552582830191, + 0.006639812141656876, + 0.0052931890822947025, + 0.005420756991952658, + 0.00590559421107173, + 0.006133381277322769, + 0.009391129948198795, + 0.005997912492603064, + 0.009336823597550392, + 0.006193602457642555, + 0.006900364998728037, + 0.0022961189970374107, + 0.00503320200368762, + 0.004161457065492868, + 0.013538400642573833, + 0.028095269575715065, + 0.025038177147507668, + 0.0206287931650877, + 0.018630241975188255, + 0.008653663098812103, + 0.02103814296424389, + 0.014612292870879173, + 0.009579175151884556, + 0.01616031490266323, + 0.016188696026802063, + 0.008939489722251892, + 0.009150856174528599, + 0.006757471710443497, + 0.009172160178422928, + 0.010676880367100239, + 0.004698548000305891, + 0.010432456620037556, + 0.007508069276809692, + 0.005559923592954874, + 0.007708867080509663, + 0.004999430384486914, + 0.004897633567452431, + 0.005396342370659113, + 0.002435089321807027, + 0.002906848443672061, + 0.00416466873139143, + 0.003783627413213253, + 0.003374990075826645, + 0.0017179878195747733, + 0.0038356450386345387, + 0.0029582290444523096, + 0.005972014740109444, + 0.007638570386916399, + 0.0073783183470368385, + 0.004501644056290388, + 0.0065809632651507854, + 0.00250904425047338, + 0.005861358717083931, + 0.004168164450675249, + 0.034013036638498306, + 0.06839539855718613, + 0.05710732936859131, + 0.029877260327339172, + 0.04112575575709343, + 0.017321031540632248, + 0.026287751272320747, + 0.03841613605618477, + 0.014820915646851063, + 0.013044651597738266, + 0.021511690691113472, + 0.00670013390481472, + 0.009318718686699867, + 0.007777618709951639, + 0.006449881009757519, + 0.011828814633190632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.0125422989949584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tight, close-up, static shot taken from a slight overhead angle, a pair of deft hands in flawless white gloves add the finishing touches. The gloves' texture is delicate and smooth, as if meticulously sculpted from pure white icing, perfectly echoing the fluid lines of a fairytale illustration. They intently handle a Cantonese-style mooncake, fresh from the oven, a captivating golden hue. The intricate auspicious cloud pattern on its surface appears three-dimensional and vivid under the soft sidelight. One hand gently supports the mooncake, while the other, using a pair of delicate tweezers, delicately grasps a thin piece of gold foil and, with steady deliberation, precisely places it on the center pistil of the mooncake. Warm light diffused from one side of the frame casts a honey-hued glow over the scene, creating an atmosphere of tranquility, concentration, and exquisite beauty. The overall color palette is soft and warm, imbued with the dreamy and sweet beauty of a fairy tale.", + "seed": 30711, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_7/index_7531_seed_30711_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017180757084861397, + "vlm_logprob": -0.017180757084861397, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.017180757084861397, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012005806900560856 + ], + "opposite_logprob_tokens": [ + -4.637005805969238 + ], + "logprob_reward": -0.012005806900560856, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012005806900560856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.247062087059021, + 0.10440518707036972, + 0.09355245530605316, + 0.07030139118432999, + 0.05111638829112053, + 0.1269095242023468, + 0.0701436921954155, + 0.06243574246764183, + 0.05285760387778282, + 0.12121590226888657 + ], + "credit_map": [ + 0.0452955961227417, + 0.024735599756240845, + 0.051113516092300415, + 0.02059941180050373, + 0.023489076644182205, + 0.03219511732459068, + 0.03111448511481285, + 0.01851927489042282, + 0.014839149080216885, + 0.02076157182455063, + 0.017620844766497612, + 0.004880898632109165, + 0.014262772165238857, + 0.016242247074842453, + 0.009288226254284382, + 0.006509474944323301, + 0.02807515859603882, + 0.010252712294459343, + 0.02831929177045822, + 0.003798112506046891, + 0.0061165811493992805, + 0.007436441723257303, + 0.005011786241084337, + 0.004542367998510599, + 0.014437543228268623, + 0.006999183911830187, + 0.013573949225246906, + 0.004993544891476631, + 0.010362044908106327, + 0.006629126612097025, + 0.007794369477778673, + 0.005511635914444923, + 0.007903494872152805, + 0.0076270680874586105, + 0.005577261559665203, + 0.0050632343627512455, + 0.00918518379330635, + 0.005562506150454283, + 0.006534268613904715, + 0.0036633694544434547, + 0.025127114728093147, + 0.011890867725014687, + 0.024102546274662018, + 0.009099713526666164, + 0.01738555170595646, + 0.014704263769090176, + 0.017298007383942604, + 0.0073014721274375916, + 0.02562575787305832, + 0.0033105064649134874, + 0.02587151899933815, + 0.0030835813377052546, + 0.0026120792608708143, + 0.004289709962904453, + 0.0027434052899479866, + 0.0026071311440318823, + 0.021890878677368164, + 0.003459308762103319, + 0.02218874730169773, + 0.0028600706718862057, + 0.0028521406929939985, + 0.003948697820305824, + 0.0024807737208902836, + 0.0027551264502108097, + 0.007357924710959196, + 0.005819329060614109, + 0.009632389061152935, + 0.0024697210174053907, + 0.008119666948914528, + 0.010506161488592625, + 0.0049966247752308846, + 0.003955783788114786, + 0.042679350823163986, + 0.0059662312269210815, + 0.04333276301622391, + 0.005234626587480307, + 0.004610027652233839, + 0.009213929064571857, + 0.005130257923156023, + 0.005048710852861404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.012005806900560856 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311694949865341 + ], + "opposite_logprob_tokens": [ + -3.168117046356201 + ], + "logprob_reward": -0.04311694949865341, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311694949865341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07638554275035858, + 0.174015611410141, + 0.060604508966207504, + 0.07100023329257965, + 0.023073606193065643, + 0.12934397161006927, + 0.24849775433540344, + 0.10573223978281021, + 0.06349349021911621, + 0.04785303398966789 + ], + "credit_map": [ + 0.013382893055677414, + 0.006749571301043034, + 0.017388930544257164, + 0.005084012635052204, + 0.008305233903229237, + 0.012841006740927696, + 0.006404722109436989, + 0.006229172460734844, + 0.028095755726099014, + 0.04441995918750763, + 0.019239310175180435, + 0.008598718792200089, + 0.02008146233856678, + 0.025077978149056435, + 0.016075430437922478, + 0.012426991015672684, + 0.012350372038781643, + 0.008752943947911263, + 0.013047220185399055, + 0.0030194558203220367, + 0.007118419744074345, + 0.007313088048249483, + 0.004936689510941505, + 0.004066317807883024, + 0.018470389768481255, + 0.005319428164511919, + 0.020729437470436096, + 0.0031456726137548685, + 0.008481919765472412, + 0.007047663908451796, + 0.004687928128987551, + 0.003117790911346674, + 0.004269938915967941, + 0.0026920887175947428, + 0.004081150516867638, + 0.0016821844037622213, + 0.0022076440509408712, + 0.0025892003905028105, + 0.003959296736866236, + 0.0015921032754704356, + 0.034469954669475555, + 0.007696826010942459, + 0.03531217202544212, + 0.006485637277364731, + 0.013355460949242115, + 0.014034559018909931, + 0.011699859984219074, + 0.006289500277489424, + 0.09029461443424225, + 0.009993735700845718, + 0.09350056946277618, + 0.008938205428421497, + 0.009340564720332623, + 0.014956234954297543, + 0.010976563207805157, + 0.010497268289327621, + 0.03736834600567818, + 0.004626829642802477, + 0.03800021857023239, + 0.003991603385657072, + 0.005321634002029896, + 0.0066891140304505825, + 0.004535227548331022, + 0.0051992665976285934, + 0.01197460200637579, + 0.0049929190427064896, + 0.01185938436537981, + 0.0035625998862087727, + 0.009972472675144672, + 0.010629355907440186, + 0.006028833799064159, + 0.004473322536796331, + 0.011401665396988392, + 0.003803438972681761, + 0.011893071234226227, + 0.002869717078283429, + 0.007076329551637173, + 0.005317502189427614, + 0.0031445366330444813, + 0.0023467731662094593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.04311694949865341 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186133354902267 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186133354902267, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186133354902267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09341747313737869, + 0.2905104458332062, + 0.03937859460711479, + 0.028875960037112236, + 0.07174956053495407, + 0.05259287357330322, + 0.15193796157836914, + 0.04093865677714348, + 0.1097707748413086, + 0.12082778662443161 + ], + "credit_map": [ + 0.013533476740121841, + 0.0107234176248312, + 0.014621823094785213, + 0.007885263301432133, + 0.011051113717257977, + 0.018147841095924377, + 0.009435819461941719, + 0.008018706925213337, + 0.031501077115535736, + 0.05246984586119652, + 0.038846246898174286, + 0.021420005708932877, + 0.03872822970151901, + 0.0609828382730484, + 0.0261062141507864, + 0.020455924794077873, + 0.00950558390468359, + 0.005373867694288492, + 0.008922279812395573, + 0.0017841614317148924, + 0.003898477181792259, + 0.005023301113396883, + 0.0025259931571781635, + 0.0023449240252375603, + 0.004916542675346136, + 0.0029505016282200813, + 0.0046356418170034885, + 0.00212698127143085, + 0.004123218357563019, + 0.004632501397281885, + 0.002998159034177661, + 0.0024924108292907476, + 0.013234645128250122, + 0.008469653315842152, + 0.013018418103456497, + 0.006235375069081783, + 0.006979482714086771, + 0.009290311485528946, + 0.009692174382507801, + 0.004829488694667816, + 0.012271531857550144, + 0.003507357556372881, + 0.015060585923492908, + 0.0033791055902838707, + 0.003927149344235659, + 0.004842827096581459, + 0.006320652086287737, + 0.0032836594618856907, + 0.042780231684446335, + 0.008475628681480885, + 0.044721316546201706, + 0.007281535305082798, + 0.010548091493546963, + 0.013390305452048779, + 0.013151682913303375, + 0.011589139699935913, + 0.010239948518574238, + 0.004192889668047428, + 0.00938440766185522, + 0.0033667010720819235, + 0.0038349574897438288, + 0.002846767660230398, + 0.004356169607490301, + 0.0027168048545718193, + 0.0327032171189785, + 0.006984512321650982, + 0.03368446230888367, + 0.0071097989566624165, + 0.006000307854264975, + 0.011998815461993217, + 0.0067804423160851, + 0.004509199410676956, + 0.04114033654332161, + 0.006075574550777674, + 0.04066307842731476, + 0.005489355884492397, + 0.007029145956039429, + 0.010456419549882412, + 0.005304743070155382, + 0.00466912379488349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.018186133354902267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015161458868533373 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015161458868533373, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015161458868533373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07903987914323807, + 0.10029934346675873, + 0.023231791332364082, + 0.14121931791305542, + 0.1058143600821495, + 0.07077918946743011, + 0.19874081015586853, + 0.027683520689606667, + 0.11084237694740295, + 0.14234937727451324 + ], + "credit_map": [ + 0.01643884927034378, + 0.006122131831943989, + 0.01681431755423546, + 0.005782309453934431, + 0.008235933259129524, + 0.011618995107710361, + 0.007423679810017347, + 0.006603662855923176, + 0.012259593233466148, + 0.012106486596167088, + 0.01402838435024023, + 0.007328708190470934, + 0.015771053731441498, + 0.02288074605166912, + 0.008360649459064007, + 0.007563721854239702, + 0.003145386464893818, + 0.00304219126701355, + 0.0032068779692053795, + 0.0016387572977691889, + 0.0040074666030704975, + 0.0033603000920265913, + 0.0028113494627177715, + 0.0020194626413285732, + 0.04815999045968056, + 0.006600684020668268, + 0.05166570842266083, + 0.005924643483012915, + 0.006460609845817089, + 0.009767752140760422, + 0.006271878257393837, + 0.00636805547401309, + 0.023735856637358665, + 0.011209697462618351, + 0.024657927453517914, + 0.006980503909289837, + 0.0116184176877141, + 0.012947316281497478, + 0.008780702948570251, + 0.005883940029889345, + 0.0169138815253973, + 0.00468452600762248, + 0.01686367765069008, + 0.0037333760410547256, + 0.008147465996444225, + 0.009111760184168816, + 0.006312383804470301, + 0.005012122914195061, + 0.07676486670970917, + 0.006662656087428331, + 0.07946836203336716, + 0.007809219881892204, + 0.004609582480043173, + 0.011099456809461117, + 0.006274350453168154, + 0.006052317097783089, + 0.004623434506356716, + 0.0026726587675511837, + 0.004370412789285183, + 0.0017980678239837289, + 0.0028518501203507185, + 0.0046962713822722435, + 0.00343519845046103, + 0.0032356271985918283, + 0.03061208501458168, + 0.006924260873347521, + 0.03241284564137459, + 0.004669092129915953, + 0.01080529298633337, + 0.013050618581473827, + 0.006095176097005606, + 0.006273007486015558, + 0.049960847944021225, + 0.006571211386471987, + 0.050498634576797485, + 0.0056394897401332855, + 0.00943189300596714, + 0.009233205579221249, + 0.006172724533826113, + 0.004841373767703772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.0015161458868533373 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing flawless white gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Cantonese-style mooncake with a golden hue and auspicious cloud pattern visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one hand support the mooncake while the other uses tweezers to place gold foil on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot a tight close-up taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with warm sidelight casting a honey-hued glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078749783337116 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078749783337116, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078749783337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049285829067230225, + 0.128775954246521, + 0.030513405799865723, + 0.045167475938797, + 0.047037046402692795, + 0.04437342658638954, + 0.41199326515197754, + 0.051732562482357025, + 0.10425935685634613, + 0.08686165511608124 + ], + "credit_map": [ + 0.00896403007209301, + 0.004587394185364246, + 0.007420503068715334, + 0.004633757285773754, + 0.005830015055835247, + 0.005896656773984432, + 0.007512175012379885, + 0.004441301804035902, + 0.01545640453696251, + 0.016903197392821312, + 0.0178928654640913, + 0.00999949686229229, + 0.019372817128896713, + 0.026748724281787872, + 0.012503528967499733, + 0.009898914955556393, + 0.00595073914155364, + 0.005600776523351669, + 0.006020428612828255, + 0.002123393816873431, + 0.0021679254714399576, + 0.004090693313628435, + 0.002702869474887848, + 0.001856577699072659, + 0.012162469327449799, + 0.004239157307893038, + 0.01358653325587511, + 0.0023066813591867685, + 0.004306445363909006, + 0.004033043049275875, + 0.0023720574099570513, + 0.0021610872354358435, + 0.008720744401216507, + 0.007912570610642433, + 0.005461288150399923, + 0.0037661464884877205, + 0.004430610686540604, + 0.006176252383738756, + 0.00774356909096241, + 0.0028258662205189466, + 0.01296430453658104, + 0.0028030576650053263, + 0.01375177688896656, + 0.002272064099088311, + 0.00381746725179255, + 0.0031330320052802563, + 0.003371603786945343, + 0.0022601201198995113, + 0.15709282457828522, + 0.01500955130904913, + 0.16228516399860382, + 0.015639880672097206, + 0.011587469838559628, + 0.023882312700152397, + 0.013263166882097721, + 0.013232863508164883, + 0.011736412532627583, + 0.004590757191181183, + 0.008549918420612812, + 0.0029938563238829374, + 0.004739217460155487, + 0.005849181208759546, + 0.006110559217631817, + 0.007162662222981453, + 0.03348413482308388, + 0.0059004779905080795, + 0.03487304598093033, + 0.004409070126712322, + 0.005810686852782965, + 0.010722996667027473, + 0.004587069153785706, + 0.004471879452466965, + 0.029559116810560226, + 0.004484216682612896, + 0.02972318045794964, + 0.003673206316307187, + 0.006481212098151445, + 0.0058647519908845425, + 0.0038011863362044096, + 0.003274784656241536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.011078749783337116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_8/index_tensor([9906, 9906])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_8/index_tensor([9906, 9906])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4060aa4018b5cbf9d69d4663af14d8302119fb12 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_8/index_tensor([9906, 9906])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, angled, Dutch-angle overhead shot closely follows a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels as it deftly navigates a crowded street at a rapid, slightly jerky, game-like pace. The tuk-tuk, with its bright yellow roof, vibrant red body, and sky-blue seats, is covered in a crisp pixel grid. Inside sits a driver, also composed of pixels, whose silhouette is minimalist. The street itself is a dark gray, pixelated asphalt, crowded with equally pixelated motorcycles and blocky cars, while pedestrians, composed of a few pixels, walk along the sidewalk in a fixed, animated loop. The buildings on either side are flat, color-saturated pixel structures dotted with windows and billboards simplified into geometric shapes. The tuk-tuk, making abrupt, near-90-degree turns, finds narrow gaps in the pixelated traffic, cutting in and out of the way with unrealistic precision. The entire world presents a distinct 16-bit retro game aesthetic. Light and shadow are simplified to the superposition of pure color blocks. There are no soft gradients. The highly saturated primary and secondary colors together create a nostalgic atmosphere that is vibrant, slightly chaotic, yet extremely fun to play.", + "seed": 15046, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_8/index_9906_seed_15046_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017983991297660394, + "vlm_logprob": -0.017983991297660394, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.017983991297660394, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006014107260853052 + ], + "opposite_logprob_tokens": [ + -5.881013870239258 + ], + "logprob_reward": -0.006014107260853052, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006014107260853052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14964555203914642, + 0.07980453222990036, + 0.14314334094524384, + 0.06279680877923965, + 0.07499001175165176, + 0.05627119541168213, + 0.04340875893831253, + 0.0487716905772686, + 0.29015570878982544, + 0.05101235955953598 + ], + "credit_map": [ + 0.020752759650349617, + 0.013953843154013157, + 0.018634917214512825, + 0.02646176889538765, + 0.017139233648777008, + 0.008320839144289494, + 0.02953491359949112, + 0.014847291633486748, + 0.009061705321073532, + 0.010001525282859802, + 0.011475123465061188, + 0.010385137982666492, + 0.0062945629470050335, + 0.010720989666879177, + 0.011365026235580444, + 0.010500466451048851, + 0.014180348254740238, + 0.010936351493000984, + 0.014112754724919796, + 0.025182386860251427, + 0.02031201310455799, + 0.01800670102238655, + 0.02592424303293228, + 0.014488550834357738, + 0.004464729689061642, + 0.008401667699217796, + 0.0037413157988339663, + 0.00335995783098042, + 0.019702903926372528, + 0.0050394050776958466, + 0.005396833177655935, + 0.012689992785453796, + 0.012206566520035267, + 0.0049525778740644455, + 0.029308171942830086, + 0.005678587127476931, + 0.0059521994553506374, + 0.004871259443461895, + 0.0039855921640992165, + 0.008035053499042988, + 0.008211962878704071, + 0.00428668363019824, + 0.0074325865134596825, + 0.008724040351808071, + 0.003784036263823509, + 0.00673217186704278, + 0.004405184183269739, + 0.01269452553242445, + 0.0040126582607626915, + 0.004871848970651627, + 0.004642954096198082, + 0.011988035403192043, + 0.003928878344595432, + 0.0033568316139280796, + 0.007514160592108965, + 0.003093391191214323, + 0.006553295534104109, + 0.005221647676080465, + 0.0045790052972733974, + 0.005511596333235502, + 0.006192089524120092, + 0.007896175608038902, + 0.006223300471901894, + 0.006594577804207802, + 0.023536892607808113, + 0.01676122657954693, + 0.10043825954198837, + 0.10299713164567947, + 0.008612449280917645, + 0.01100272685289383, + 0.017373351380228996, + 0.009433673694729805, + 0.010114578530192375, + 0.0073998356238007545, + 0.006954522337764502, + 0.006126872263848782, + 0.005421220324933529, + 0.004125034902244806, + 0.006058705039322376, + 0.0048115914687514305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.006014107260853052 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12580938637256622, + 0.1103908121585846, + 0.09113118052482605, + 0.09468202292919159, + 0.16339731216430664, + 0.04582267627120018, + 0.05314759537577629, + 0.051221903413534164, + 0.23459851741790771, + 0.02979857660830021 + ], + "credit_map": [ + 0.01894209533929825, + 0.008689217269420624, + 0.014308893121778965, + 0.0208012443035841, + 0.01927758939564228, + 0.007515884004533291, + 0.02410772070288658, + 0.01216674130409956, + 0.014565405435860157, + 0.015325958840548992, + 0.011394038796424866, + 0.02045673318207264, + 0.011211677454411983, + 0.01082917395979166, + 0.012531450018286705, + 0.01407637633383274, + 0.008756488561630249, + 0.006782948039472103, + 0.011526975780725479, + 0.011029340326786041, + 0.012341536581516266, + 0.018360145390033722, + 0.010649860836565495, + 0.011683882214128971, + 0.008923722431063652, + 0.013073718175292015, + 0.005342476535588503, + 0.004119109828025103, + 0.032758742570877075, + 0.008416034281253815, + 0.006657986436039209, + 0.015390228480100632, + 0.036271724849939346, + 0.01020508911460638, + 0.06433544307947159, + 0.012677229940891266, + 0.008881349116563797, + 0.013657473027706146, + 0.007265586405992508, + 0.010103427805006504, + 0.005479051731526852, + 0.00456269271671772, + 0.006095604971051216, + 0.006531326565891504, + 0.004204089287668467, + 0.005185104440897703, + 0.003948289435356855, + 0.009816515259444714, + 0.004654301796108484, + 0.004966373555362225, + 0.006214649882167578, + 0.013599109835922718, + 0.004532539285719395, + 0.004025379661470652, + 0.007705535739660263, + 0.007449707482010126, + 0.008285441435873508, + 0.00765093881636858, + 0.007450472563505173, + 0.007014170754700899, + 0.00533360755071044, + 0.005262678489089012, + 0.005929645150899887, + 0.004294945392757654, + 0.01827036216855049, + 0.013499065302312374, + 0.0795404240489006, + 0.08312917500734329, + 0.007323091849684715, + 0.010380329564213753, + 0.014626272022724152, + 0.007829795591533184, + 0.0047293007373809814, + 0.003630387829616666, + 0.004212867934256792, + 0.003504334017634392, + 0.004733726382255554, + 0.00277763232588768, + 0.003360713366419077, + 0.002849615179002285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043041475117206573 + ], + "opposite_logprob_tokens": [ + -3.168041467666626 + ], + "logprob_reward": -0.043041475117206573, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043041475117206573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09568565338850021, + 0.07473418861627579, + 0.10842760652303696, + 0.1487654447555542, + 0.26234522461891174, + 0.057227473706007004, + 0.07697872817516327, + 0.07119517773389816, + 0.08513540774583817, + 0.01950501836836338 + ], + "credit_map": [ + 0.011044556275010109, + 0.007046459708362818, + 0.013039755634963512, + 0.013795068487524986, + 0.01220602449029684, + 0.008774852380156517, + 0.020258385688066483, + 0.00952056422829628, + 0.007040049415081739, + 0.00745929591357708, + 0.009898604825139046, + 0.01052146963775158, + 0.006059059873223305, + 0.009846625849604607, + 0.012147244065999985, + 0.011761842295527458, + 0.007769486866891384, + 0.008331636898219585, + 0.010150388814508915, + 0.0158375333994627, + 0.01872251182794571, + 0.01653994992375374, + 0.016019724309444427, + 0.015056380070745945, + 0.01092439889907837, + 0.01932804100215435, + 0.011340921744704247, + 0.007867768406867981, + 0.04462214559316635, + 0.013605809770524502, + 0.012553030624985695, + 0.028523340821266174, + 0.059520427137613297, + 0.012612585909664631, + 0.1210486888885498, + 0.020253794267773628, + 0.007940679788589478, + 0.012190541252493858, + 0.01048439834266901, + 0.018294133245944977, + 0.0060393670573830605, + 0.008337263017892838, + 0.010647590272128582, + 0.008948407135903835, + 0.006630386225879192, + 0.0062376330606639385, + 0.003395695239305496, + 0.006991141010075808, + 0.00829320214688778, + 0.008027135394513607, + 0.012132116593420506, + 0.01941235177218914, + 0.004863823764026165, + 0.004676415119320154, + 0.012829449959099293, + 0.006744245067238808, + 0.009248652495443821, + 0.00980528537184, + 0.006983147002756596, + 0.006826560012996197, + 0.009260445833206177, + 0.013144141994416714, + 0.008858381770551205, + 0.007068579085171223, + 0.007664377801120281, + 0.006303165107965469, + 0.027001406997442245, + 0.025595078244805336, + 0.00328277749940753, + 0.005124034825712442, + 0.005341262556612492, + 0.004823317751288414, + 0.003729545511305332, + 0.002202293137088418, + 0.0028851069509983063, + 0.0022329373750835657, + 0.002770128659904003, + 0.0014997214311733842, + 0.002575109712779522, + 0.0016101774526759982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.043041475117206573 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368403762578964 + ], + "opposite_logprob_tokens": [ + -3.408684015274048 + ], + "logprob_reward": -0.03368403762578964, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368403762578964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07462374120950699, + 0.09379767626523972, + 0.12227291613817215, + 0.15393765270709991, + 0.13258954882621765, + 0.07838728278875351, + 0.057496532797813416, + 0.13125541806221008, + 0.12618185579776764, + 0.029457399621605873 + ], + "credit_map": [ + 0.009515243582427502, + 0.005735154263675213, + 0.008436515927314758, + 0.009151185862720013, + 0.00757466908544302, + 0.006232800427824259, + 0.017244970425963402, + 0.010733203031122684, + 0.010730168782174587, + 0.007837916724383831, + 0.011247550137341022, + 0.019900869578123093, + 0.007714692037552595, + 0.010769549757242203, + 0.014565402641892433, + 0.011031527072191238, + 0.010362936183810234, + 0.008883194997906685, + 0.01229202002286911, + 0.016078155487775803, + 0.019420314580202103, + 0.023623762652277946, + 0.01681223139166832, + 0.014800299890339375, + 0.014912224374711514, + 0.01662587560713291, + 0.01580888405442238, + 0.010219325311481953, + 0.0424402579665184, + 0.014911839738488197, + 0.019996829330921173, + 0.019022425636649132, + 0.018847275525331497, + 0.008520442992448807, + 0.050046201795339584, + 0.011423655785620213, + 0.010810516774654388, + 0.011903142556548119, + 0.0068460991606116295, + 0.01419219933450222, + 0.009744932875037193, + 0.0169095229357481, + 0.009289510548114777, + 0.009026908315718174, + 0.008688800036907196, + 0.006636560428887606, + 0.00571815250441432, + 0.012372897006571293, + 0.0059519619680941105, + 0.006016391329467297, + 0.007608985994011164, + 0.00793499406427145, + 0.008884506300091743, + 0.006807142868638039, + 0.007049477193504572, + 0.007243072148412466, + 0.01676762104034424, + 0.0160624198615551, + 0.020740479230880737, + 0.019253747537732124, + 0.012734956108033657, + 0.010344913229346275, + 0.02368854731321335, + 0.01166273932904005, + 0.012501423247158527, + 0.007496952544897795, + 0.03680584952235222, + 0.04081117734313011, + 0.005278000142425299, + 0.006723958067595959, + 0.008217230439186096, + 0.008347265422344208, + 0.005122300237417221, + 0.0033315285108983517, + 0.004042064771056175, + 0.004795421380549669, + 0.0028475187718868256, + 0.002713518450036645, + 0.003262101672589779, + 0.0033429458271712065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.03368403762578964 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004366874636616558 + ], + "opposite_logprob_tokens": [ + -7.750436782836914 + ], + "logprob_reward": -0.0004366874636616558, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004366874636616558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2321019470691681, + 0.05958806350827217, + 0.05648750066757202, + 0.09796542674303055, + 0.11760273575782776, + 0.12623541057109833, + 0.03716356307268143, + 0.047855012118816376, + 0.18384627997875214, + 0.041154082864522934 + ], + "credit_map": [ + 0.026481064036488533, + 0.017182299867272377, + 0.0217802282422781, + 0.03993970528244972, + 0.03138548135757446, + 0.014428641647100449, + 0.06018379330635071, + 0.020720748230814934, + 0.004967158194631338, + 0.004435005597770214, + 0.00787637010216713, + 0.010309364646673203, + 0.005278606899082661, + 0.010146082378923893, + 0.008027159608900547, + 0.008548316545784473, + 0.005682243034243584, + 0.005927247926592827, + 0.005737257190048695, + 0.007360280025750399, + 0.010210752487182617, + 0.005610927008092403, + 0.008493965491652489, + 0.0074648247100412846, + 0.006155022885650396, + 0.011273611336946487, + 0.01076322142034769, + 0.00825112871825695, + 0.025980515405535698, + 0.007151771802455187, + 0.011500676162540913, + 0.016889475286006927, + 0.014601051807403564, + 0.007661357056349516, + 0.04060885310173035, + 0.007578170392662287, + 0.00739352498203516, + 0.013295331038534641, + 0.006581765599548817, + 0.019882680848240852, + 0.01430124044418335, + 0.019612543284893036, + 0.02204136550426483, + 0.015154168009757996, + 0.010258627124130726, + 0.020042534917593002, + 0.006555522326380014, + 0.018269412219524384, + 0.0032523872796446085, + 0.004692150745540857, + 0.0058287144638597965, + 0.005317007657140493, + 0.0049455431289970875, + 0.003559017088264227, + 0.004725398030132055, + 0.004843344446271658, + 0.006737329065799713, + 0.006975204683840275, + 0.006671334151178598, + 0.006294067949056625, + 0.0061495318077504635, + 0.0046424404717981815, + 0.005945321638137102, + 0.004439783748239279, + 0.01620532013475895, + 0.010697269812226295, + 0.06259044259786606, + 0.061660949140787125, + 0.006461306009441614, + 0.00787657592445612, + 0.011438294313848019, + 0.006916123908013105, + 0.007243163883686066, + 0.00487598218023777, + 0.005199464038014412, + 0.006703558843582869, + 0.006439966149628162, + 0.003735695034265518, + 0.004329297225922346, + 0.002626952948048711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0004366874636616558 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, angled, Dutch-angle overhead shot closely follows a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels as it deftly navigates a crowded street at a rapid, slightly jerky, game-like pace. The tuk-tuk, with its bright yellow roof, vibrant red body, and sky-blue seats, is covered in a crisp pixel grid. Inside sits a driver, also composed of pixels, whose silhouette is minimalist. The street itself is a dark gray, pixelated asphalt, crowded with equally pixelated motorcycles and blocky cars, while pedestrians, composed of a few pixels, walk along the sidewalk in a fixed, animated loop. The buildings on either side are flat, color-saturated pixel structures dotted with windows and billboards simplified into geometric shapes. The tuk-tuk, making abrupt, near-90-degree turns, finds narrow gaps in the pixelated traffic, cutting in and out of the way with unrealistic precision. The entire world presents a distinct 16-bit retro game aesthetic. Light and shadow are simplified to the superposition of pure color blocks. There are no soft gradients. The highly saturated primary and secondary colors together create a nostalgic atmosphere that is vibrant, slightly chaotic, yet extremely fun to play.", + "seed": 15047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_8/index_9906_seed_15047_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028422770276665687, + "vlm_logprob": -0.028422770276665687, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.028422770276665687, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004003130830824375 + ], + "opposite_logprob_tokens": [ + -6.504003047943115 + ], + "logprob_reward": -0.004003130830824375, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.004003130830824375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.243887260556221, + 0.09443968534469604, + 0.052912186831235886, + 0.21436448395252228, + 0.06401873379945755, + 0.02341701090335846, + 0.04761119931936264, + 0.029844382777810097, + 0.14870446920394897, + 0.08080057054758072 + ], + "credit_map": [ + 0.034889332950115204, + 0.026242462918162346, + 0.02765068970620632, + 0.04192601144313812, + 0.026724256575107574, + 0.027502883225679398, + 0.033656731247901917, + 0.025294888764619827, + 0.010125011205673218, + 0.009145054966211319, + 0.012461818754673004, + 0.01015617698431015, + 0.006676258984953165, + 0.018845215439796448, + 0.011330722831189632, + 0.015699435025453568, + 0.007238635793328285, + 0.0044206599704921246, + 0.012313621118664742, + 0.006431629415601492, + 0.004214323591440916, + 0.00450692605227232, + 0.0075244782492518425, + 0.006261908914893866, + 0.023396877571940422, + 0.030391888692975044, + 0.05281822383403778, + 0.013553164899349213, + 0.013400724157691002, + 0.02394029125571251, + 0.014422380365431309, + 0.04244092106819153, + 0.009500337764620781, + 0.006495194043964148, + 0.007728349883109331, + 0.013047561049461365, + 0.005384692922234535, + 0.008162468671798706, + 0.007501300889998674, + 0.006198828108608723, + 0.0033593063708394766, + 0.0027207687962800264, + 0.0038930014707148075, + 0.0036056488752365112, + 0.0015476290136575699, + 0.0015898074489086866, + 0.0027494134847074747, + 0.0039514354430139065, + 0.004894102457910776, + 0.00427328422665596, + 0.007318075746297836, + 0.007450060918927193, + 0.0039253938011825085, + 0.0061975750140845776, + 0.007040554657578468, + 0.006512152962386608, + 0.0032730938401073217, + 0.00306717143394053, + 0.004494174383580685, + 0.004173269961029291, + 0.0025788056664168835, + 0.0029045974370092154, + 0.0038567655719816685, + 0.00549650564789772, + 0.012274790555238724, + 0.009765819646418095, + 0.032520443201065063, + 0.0083896704018116, + 0.02704843319952488, + 0.012351174838840961, + 0.010952706448733807, + 0.035401441156864166, + 0.007071704603731632, + 0.01017655897885561, + 0.012304434552788734, + 0.01268075592815876, + 0.0060829902067780495, + 0.01147118117660284, + 0.008381768129765987, + 0.01263117603957653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.004003130830824375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544536031782627 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544536031782627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544536031782627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08597324043512344, + 0.12229829281568527, + 0.11987137049436569, + 0.14970044791698456, + 0.038196880370378494, + 0.013545948080718517, + 0.025205053389072418, + 0.015164047479629517, + 0.39447158575057983, + 0.035573046654462814 + ], + "credit_map": [ + 0.01100301742553711, + 0.010662032291293144, + 0.010990026406943798, + 0.015542738139629364, + 0.012404724024236202, + 0.007568009663373232, + 0.006871238816529512, + 0.010931462049484253, + 0.008497127331793308, + 0.010649087838828564, + 0.010696402750909328, + 0.007556728553026915, + 0.006983422674238682, + 0.03510821983218193, + 0.010530393570661545, + 0.03227691724896431, + 0.012551968917250633, + 0.012066267430782318, + 0.01788787730038166, + 0.01707407273352146, + 0.011077680625021458, + 0.011562473140656948, + 0.02183772809803486, + 0.0158133152872324, + 0.01397387869656086, + 0.014427052810788155, + 0.023477183654904366, + 0.01571924239397049, + 0.011081920005381107, + 0.015730977058410645, + 0.016447046771645546, + 0.03884316608309746, + 0.006996799726039171, + 0.0018798938253894448, + 0.004023013636469841, + 0.004321268759667873, + 0.00222757738083601, + 0.005775535013526678, + 0.002973056398332119, + 0.009999735280871391, + 0.002115492708981037, + 0.0011799813946709037, + 0.0033763202372938395, + 0.0011641494929790497, + 0.0006861843867227435, + 0.0009752041078172624, + 0.0015174465952441096, + 0.0025311701465398073, + 0.002879134612157941, + 0.0027033009100705385, + 0.00196095765568316, + 0.005254453979432583, + 0.001874892390333116, + 0.0027392699848860502, + 0.004346160218119621, + 0.003446887945756316, + 0.0016547741834074259, + 0.0016382206231355667, + 0.0022795607801526785, + 0.00163837184663862, + 0.0016621795948594809, + 0.0017293165437877178, + 0.002034104196354747, + 0.0025275209918618202, + 0.043199729174375534, + 0.03456994518637657, + 0.03782081976532936, + 0.019961217418313026, + 0.12725931406021118, + 0.07812050729990005, + 0.03177529200911522, + 0.021764779463410378, + 0.003123253583908081, + 0.003282326739281416, + 0.004947093315422535, + 0.003946957644075155, + 0.0033362761605530977, + 0.004393111914396286, + 0.004604537971317768, + 0.007939490489661694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.012544536031782627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999480724334717 + ], + "opposite_logprob_tokens": [ + -2.6949949264526367 + ], + "logprob_reward": -0.06999480724334717, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999480724334717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09910944104194641, + 0.1827804297208786, + 0.10766983777284622, + 0.13504436612129211, + 0.03159835562109947, + 0.055003147572278976, + 0.06849900633096695, + 0.032121624797582626, + 0.26533955335617065, + 0.02283422090113163 + ], + "credit_map": [ + 0.011696327477693558, + 0.009350034408271313, + 0.01375869382172823, + 0.01856531761586666, + 0.011416148394346237, + 0.010237438604235649, + 0.012924568727612495, + 0.011160913854837418, + 0.017862558364868164, + 0.01720133237540722, + 0.02563137374818325, + 0.015188932418823242, + 0.011334223672747612, + 0.042746663093566895, + 0.015051398426294327, + 0.037763938307762146, + 0.01613914966583252, + 0.00824686884880066, + 0.023008141666650772, + 0.012188265100121498, + 0.0069536552764475346, + 0.012996615841984749, + 0.014476037584245205, + 0.013661107048392296, + 0.016659356653690338, + 0.01404629647731781, + 0.025471022352576256, + 0.01548411138355732, + 0.006071884650737047, + 0.012013372033834457, + 0.012757336720824242, + 0.032540980726480484, + 0.0044299764558672905, + 0.0026304875500500202, + 0.003926405217498541, + 0.006629765033721924, + 0.0028905419167131186, + 0.004117014817893505, + 0.0038520819507539272, + 0.0031220850069075823, + 0.007794417440891266, + 0.004535176791250706, + 0.017072470858693123, + 0.00511115463450551, + 0.0023779100738465786, + 0.002902761334553361, + 0.0031592953018844128, + 0.01204996183514595, + 0.00919685885310173, + 0.007074358407407999, + 0.008730634115636349, + 0.01001487672328949, + 0.005396997090429068, + 0.00978043396025896, + 0.006587184965610504, + 0.011717663146555424, + 0.0038952205795794725, + 0.005009458865970373, + 0.005492973607033491, + 0.003992911893874407, + 0.0030027832835912704, + 0.004094226751476526, + 0.0026005490217357874, + 0.004033500328660011, + 0.034218233078718185, + 0.00993010587990284, + 0.1047816127538681, + 0.028816744685173035, + 0.023959999904036522, + 0.016451753675937653, + 0.015378548763692379, + 0.03180255740880966, + 0.0025210201274603605, + 0.0019498884212225676, + 0.003020164556801319, + 0.003617144888266921, + 0.002354143653064966, + 0.002246433636173606, + 0.003166288137435913, + 0.003959137015044689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.06999480724334717 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493714660406113 + ], + "opposite_logprob_tokens": [ + -2.9299371242523193 + ], + "logprob_reward": -0.05493714660406113, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493714660406113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291268765926361, + 0.21759435534477234, + 0.0782821923494339, + 0.2669781446456909, + 0.034621499478816986, + 0.02774723432958126, + 0.045381393283605576, + 0.027194947004318237, + 0.11531101912260056, + 0.057762302458286285 + ], + "credit_map": [ + 0.01515333354473114, + 0.01790536753833294, + 0.011574978940188885, + 0.024462949484586716, + 0.013363794423639774, + 0.01943717896938324, + 0.0157452505081892, + 0.011484016664326191, + 0.01653008908033371, + 0.019085291773080826, + 0.020431675016880035, + 0.011721993796527386, + 0.010763934813439846, + 0.0652124285697937, + 0.01602778024971485, + 0.057821158319711685, + 0.012629889883100986, + 0.00770525773987174, + 0.014060182496905327, + 0.008139410987496376, + 0.006528505589812994, + 0.009329515509307384, + 0.010752271860837936, + 0.009137162007391453, + 0.039963316172361374, + 0.019960787147283554, + 0.06920317560434341, + 0.028070488944649696, + 0.015294517390429974, + 0.022237973287701607, + 0.01936573162674904, + 0.05288216471672058, + 0.006006770301610231, + 0.0024969251826405525, + 0.003944812808185816, + 0.0050298552960157394, + 0.002622634870931506, + 0.00540243461728096, + 0.0031491885893046856, + 0.005968880839645863, + 0.004674740135669708, + 0.0023107368033379316, + 0.006186953280121088, + 0.003035404486581683, + 0.0017620966536924243, + 0.0023258253931999207, + 0.0029421914368867874, + 0.00450928695499897, + 0.006893141195178032, + 0.004710847046226263, + 0.007765265181660652, + 0.006607254035770893, + 0.0030017008539289236, + 0.004647019784897566, + 0.005759750492870808, + 0.005996412597596645, + 0.0034142902586609125, + 0.0033099991269409657, + 0.006243743933737278, + 0.0025756920222193003, + 0.0020871346350759268, + 0.0036634092684835196, + 0.001837541931308806, + 0.004063135012984276, + 0.0211033895611763, + 0.007944205775856972, + 0.021415499970316887, + 0.011379414238035679, + 0.02170281670987606, + 0.008227388374507427, + 0.012694110162556171, + 0.010844196192920208, + 0.005003087222576141, + 0.005107109900563955, + 0.012014999985694885, + 0.009800666943192482, + 0.004465483594685793, + 0.004504730459302664, + 0.0069405208341777325, + 0.009925699792802334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.05493714660406113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006342306733131409 + ], + "opposite_logprob_tokens": [ + -7.37563419342041 + ], + "logprob_reward": -0.0006342306733131409, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006342306733131409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1709839552640915, + 0.10195278376340866, + 0.11245197057723999, + 0.17590385675430298, + 0.06381915509700775, + 0.07137146592140198, + 0.046277452260255814, + 0.044969476759433746, + 0.1416105329990387, + 0.07065930962562561 + ], + "credit_map": [ + 0.02417694963514805, + 0.015798522159457207, + 0.023168154060840607, + 0.020164625719189644, + 0.01774575747549534, + 0.02814990095794201, + 0.01994393765926361, + 0.021836111322045326, + 0.010491239838302135, + 0.009578936733305454, + 0.012175137177109718, + 0.008113741874694824, + 0.0060323141515254974, + 0.026026517152786255, + 0.0067933849059045315, + 0.02274150773882866, + 0.02096719853579998, + 0.006433239206671715, + 0.02439950220286846, + 0.013111557811498642, + 0.0063444701954722404, + 0.015546726062893867, + 0.012057636864483356, + 0.013591645285487175, + 0.02001653052866459, + 0.02174641564488411, + 0.021781394258141518, + 0.018596256151795387, + 0.011032826267182827, + 0.02809249423444271, + 0.01862099952995777, + 0.03601692616939545, + 0.008579921908676624, + 0.004147662781178951, + 0.006931598763912916, + 0.012729479931294918, + 0.0050542037934064865, + 0.009881489910185337, + 0.00640629418194294, + 0.010088508017361164, + 0.012290025129914284, + 0.007439549546688795, + 0.010676415637135506, + 0.010331768542528152, + 0.004952403716742992, + 0.005352809093892574, + 0.009041731245815754, + 0.011286764405667782, + 0.004147138446569443, + 0.005746626760810614, + 0.005668881814926863, + 0.007183392532169819, + 0.0046945782378315926, + 0.004702532663941383, + 0.007133461069315672, + 0.007000841200351715, + 0.004770892206579447, + 0.004622738808393478, + 0.007132476195693016, + 0.0054528298787772655, + 0.004304840229451656, + 0.005546840373426676, + 0.005019402131438255, + 0.008119456470012665, + 0.022765042260289192, + 0.008761945180594921, + 0.027548156678676605, + 0.015096161514520645, + 0.013560819439589977, + 0.01786559261381626, + 0.013842614367604256, + 0.022170189768075943, + 0.012341327965259552, + 0.007386818062514067, + 0.011843706481158733, + 0.008328421041369438, + 0.005171468481421471, + 0.007342308294028044, + 0.0063042109832167625, + 0.011941046454012394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.0006342306733131409 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, angled, Dutch-angle overhead shot closely follows a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels as it deftly navigates a crowded street at a rapid, slightly jerky, game-like pace. The tuk-tuk, with its bright yellow roof, vibrant red body, and sky-blue seats, is covered in a crisp pixel grid. Inside sits a driver, also composed of pixels, whose silhouette is minimalist. The street itself is a dark gray, pixelated asphalt, crowded with equally pixelated motorcycles and blocky cars, while pedestrians, composed of a few pixels, walk along the sidewalk in a fixed, animated loop. The buildings on either side are flat, color-saturated pixel structures dotted with windows and billboards simplified into geometric shapes. The tuk-tuk, making abrupt, near-90-degree turns, finds narrow gaps in the pixelated traffic, cutting in and out of the way with unrealistic precision. The entire world presents a distinct 16-bit retro game aesthetic. Light and shadow are simplified to the superposition of pure color blocks. There are no soft gradients. The highly saturated primary and secondary colors together create a nostalgic atmosphere that is vibrant, slightly chaotic, yet extremely fun to play.", + "seed": 15048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_8/index_9906_seed_15048_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02139917149906978, + "vlm_logprob": -0.02139917149906978, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.02139917149906978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0034557406324893236 + ], + "opposite_logprob_tokens": [ + -6.253455638885498 + ], + "logprob_reward": -0.0034557406324893236, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0034557406324893236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3079390823841095, + 0.06516151875257492, + 0.047417961061000824, + 0.09707029163837433, + 0.06383118778467178, + 0.07890096306800842, + 0.06628234684467316, + 0.07073897123336792, + 0.14201052486896515, + 0.060647137463092804 + ], + "credit_map": [ + 0.03900887444615364, + 0.02545137144625187, + 0.054807934910058975, + 0.02358526736497879, + 0.10682273656129837, + 0.022229338064789772, + 0.018511082977056503, + 0.01752246916294098, + 0.013333790004253387, + 0.007942387834191322, + 0.009238830767571926, + 0.005762424785643816, + 0.005096692591905594, + 0.007550670299679041, + 0.004562924150377512, + 0.011673799715936184, + 0.005354409571737051, + 0.005311612505465746, + 0.006727491971105337, + 0.004743666388094425, + 0.003770919283851981, + 0.008643700741231441, + 0.005464581307023764, + 0.007401579525321722, + 0.012381667271256447, + 0.009918599389493465, + 0.007790068630129099, + 0.004287933465093374, + 0.03371724486351013, + 0.007239654194563627, + 0.0038857250474393368, + 0.017849398776888847, + 0.01615082658827305, + 0.0046346490271389484, + 0.019523823633790016, + 0.00425161886960268, + 0.00439629377797246, + 0.004386895336210728, + 0.0046009100042283535, + 0.005886172875761986, + 0.011632432229816914, + 0.014302839525043964, + 0.013005774468183517, + 0.008326337672770023, + 0.005567504558712244, + 0.010202499106526375, + 0.008379259146749973, + 0.007484319154173136, + 0.008678095415234566, + 0.01148198638111353, + 0.012010721489787102, + 0.006624769885092974, + 0.004636051598936319, + 0.007262930739670992, + 0.005886855535209179, + 0.009700939059257507, + 0.013271178118884563, + 0.010868015699088573, + 0.008802317082881927, + 0.006711254362016916, + 0.005125753115862608, + 0.008232627995312214, + 0.00704411743208766, + 0.01068370696157217, + 0.037558283656835556, + 0.01465519331395626, + 0.04747365042567253, + 0.009024753235280514, + 0.009427069686353207, + 0.008127284236252308, + 0.007142523303627968, + 0.008601765148341656, + 0.01238005980849266, + 0.01044461503624916, + 0.009107154794037342, + 0.005968034733086824, + 0.005563797894865274, + 0.00569386500865221, + 0.004810159560292959, + 0.006679450627416372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0034557406324893236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585110411047935 + ], + "opposite_logprob_tokens": [ + -3.895585060119629 + ], + "logprob_reward": -0.020585110411047935, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585110411047935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25333699584007263, + 0.04684862866997719, + 0.10525749623775482, + 0.07240292429924011, + 0.025425616651773453, + 0.07249423861503601, + 0.08217933773994446, + 0.07611719518899918, + 0.23864509165287018, + 0.027292493730783463 + ], + "credit_map": [ + 0.03335648775100708, + 0.027222497388720512, + 0.03118347004055977, + 0.029560042545199394, + 0.07465824484825134, + 0.021124539896845818, + 0.022642455995082855, + 0.013589252717792988, + 0.011450450867414474, + 0.00805614609271288, + 0.006451010704040527, + 0.004181080497801304, + 0.0033045755699276924, + 0.0037986617535352707, + 0.002904996508732438, + 0.006701706908643246, + 0.011970298364758492, + 0.01751215010881424, + 0.013334561139345169, + 0.011524478904902935, + 0.009252423420548439, + 0.019990980625152588, + 0.008719435892999172, + 0.012953161261975765, + 0.012843739241361618, + 0.006894410587847233, + 0.00543854059651494, + 0.004579558037221432, + 0.02553854137659073, + 0.005226814653724432, + 0.0035944534465670586, + 0.008286864496767521, + 0.00453502032905817, + 0.002278998726978898, + 0.006319690030068159, + 0.002783306874334812, + 0.001983610913157463, + 0.003080847207456827, + 0.0022554872557520866, + 0.0021886550821363926, + 0.016393255442380905, + 0.013222604990005493, + 0.008784489706158638, + 0.007097011432051659, + 0.00603278074413538, + 0.0075788721442222595, + 0.006596860941499472, + 0.006788359489291906, + 0.017538582906126976, + 0.011359630152583122, + 0.014408850111067295, + 0.007365765981376171, + 0.006126989144831896, + 0.009809351526200771, + 0.005775712430477142, + 0.009794460609555244, + 0.013593940995633602, + 0.010454442352056503, + 0.007675461936742067, + 0.006108606234192848, + 0.007858667522668839, + 0.00835909228771925, + 0.008557960391044617, + 0.013509017415344715, + 0.07174395024776459, + 0.020720548927783966, + 0.07815592736005783, + 0.013009539805352688, + 0.018969381228089333, + 0.013620667159557343, + 0.011604168452322483, + 0.010820914059877396, + 0.008731849491596222, + 0.004040070343762636, + 0.0037588360719382763, + 0.0020114670041948557, + 0.0015995409339666367, + 0.001874917303211987, + 0.00133330631069839, + 0.003942504990845919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.020585110411047935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368449956178665 + ], + "opposite_logprob_tokens": [ + -3.408684492111206 + ], + "logprob_reward": -0.03368449956178665, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368449956178665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02662501484155655, + 0.026981130242347717, + 0.013811659067869186, + 0.08510041236877441, + 0.01512120384722948, + 0.021483173593878746, + 0.014460223726928234, + 0.04328214004635811, + 0.7445030212402344, + 0.00863204151391983 + ], + "credit_map": [ + 0.00343858334235847, + 0.005049899220466614, + 0.003332182765007019, + 0.003301213262602687, + 0.0024576548021286726, + 0.0030302945524454117, + 0.002987485146149993, + 0.0030277008190751076, + 0.006051251664757729, + 0.003777592210099101, + 0.003987059462815523, + 0.0020533299539238214, + 0.002438376424834132, + 0.003008226864039898, + 0.0020837080664932728, + 0.003581584431231022, + 0.001702059875242412, + 0.0018135084537789226, + 0.0028615400660783052, + 0.0012904334580525756, + 0.001164135872386396, + 0.0019602440297603607, + 0.001577202812768519, + 0.001442533335648477, + 0.010536872781813145, + 0.009567148983478546, + 0.0057172467932105064, + 0.004409639164805412, + 0.029009312391281128, + 0.006275984458625317, + 0.00370229035615921, + 0.015881916508078575, + 0.0033762999810278416, + 0.0016343743773177266, + 0.0030856032390147448, + 0.0011056121438741684, + 0.0013679906260222197, + 0.0017253205878660083, + 0.0011856795754283667, + 0.0016403222689405084, + 0.005272325593978167, + 0.004711939953267574, + 0.0025928982067853212, + 0.0018162288470193744, + 0.002023124834522605, + 0.001988861011341214, + 0.001470612594857812, + 0.0016071825521066785, + 0.0023709984961897135, + 0.002177829621359706, + 0.0017668609507381916, + 0.0014953779755160213, + 0.0015231722500175238, + 0.001688015297986567, + 0.0014590314822271466, + 0.0019789363723248243, + 0.010381503030657768, + 0.0076692537404596806, + 0.004379906691610813, + 0.004118529614061117, + 0.0031521101482212543, + 0.0034187398850917816, + 0.004264972638338804, + 0.005897121503949165, + 0.221553236246109, + 0.06636396795511246, + 0.24628475308418274, + 0.04083753377199173, + 0.05829502269625664, + 0.041266750544309616, + 0.034422993659973145, + 0.035478733479976654, + 0.002215155865997076, + 0.0012817634269595146, + 0.0012012316146865487, + 0.000641371647361666, + 0.0005147658521309495, + 0.0006827261531725526, + 0.0004916070029139519, + 0.00160341861192137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.03368449956178665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863615706562996 + ], + "opposite_logprob_tokens": [ + -3.0486361980438232 + ], + "logprob_reward": -0.04863615706562996, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863615706562996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1987968236207962, + 0.07983946800231934, + 0.03835896775126457, + 0.14602988958358765, + 0.03618153929710388, + 0.04815151169896126, + 0.05043950304389, + 0.05569132789969444, + 0.30494996905326843, + 0.04156099632382393 + ], + "credit_map": [ + 0.01679036393761635, + 0.01320371963083744, + 0.052364420145750046, + 0.013325830921530724, + 0.06592203676700592, + 0.014559712260961533, + 0.010094269178807735, + 0.012536484748125076, + 0.01664368063211441, + 0.006595670245587826, + 0.011632739566266537, + 0.0061401017010211945, + 0.0056688920594751835, + 0.010401920415461063, + 0.007601470686495304, + 0.01515499409288168, + 0.005925901234149933, + 0.00608300743624568, + 0.006659593433141708, + 0.003298494964838028, + 0.0033199896570295095, + 0.004196061287075281, + 0.004066566936671734, + 0.004809352569282055, + 0.013761806301772594, + 0.017033517360687256, + 0.010874980129301548, + 0.010142430663108826, + 0.03763941675424576, + 0.013782890513539314, + 0.008211282081902027, + 0.03458356857299805, + 0.010856719687581062, + 0.00315998587757349, + 0.009974358603358269, + 0.0026451286394149065, + 0.001901363953948021, + 0.002517040353268385, + 0.0021623545326292515, + 0.0029645871836692095, + 0.01048494502902031, + 0.009911627508699894, + 0.008308440446853638, + 0.0039012006018310785, + 0.003610602580010891, + 0.004023312125355005, + 0.0029096370562911034, + 0.005001746118068695, + 0.008967427536845207, + 0.008293992839753628, + 0.008054215461015701, + 0.005188097711652517, + 0.00467882351949811, + 0.0058279926888644695, + 0.0037805514875799417, + 0.005648399703204632, + 0.009524213150143623, + 0.00845029205083847, + 0.0061300997622311115, + 0.004837745800614357, + 0.004864866379648447, + 0.006049308925867081, + 0.006295177154242992, + 0.009539628401398659, + 0.08839868009090424, + 0.027392227202653885, + 0.10360737890005112, + 0.0177190862596035, + 0.02102852426469326, + 0.01660737954080105, + 0.014394506812095642, + 0.015802185982465744, + 0.009292655624449253, + 0.005730282515287399, + 0.006234276108443737, + 0.00435832841321826, + 0.003191843628883362, + 0.0038160698022693396, + 0.00409178389236331, + 0.004845758900046349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.04863615706562996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006343498243950307 + ], + "opposite_logprob_tokens": [ + -7.37563419342041 + ], + "logprob_reward": -0.0006343498243950307, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006343498243950307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13360929489135742, + 0.10621917992830276, + 0.07115332782268524, + 0.0465862900018692, + 0.054031964391469955, + 0.08350144326686859, + 0.052752535790205, + 0.046432074159383774, + 0.2309150993824005, + 0.17479872703552246 + ], + "credit_map": [ + 0.017165588214993477, + 0.02190440706908703, + 0.01910797506570816, + 0.012454294599592686, + 0.0303423423320055, + 0.011448540724813938, + 0.008799535222351551, + 0.012386620044708252, + 0.019567880779504776, + 0.009491237811744213, + 0.015046902932226658, + 0.008754466660320759, + 0.009444409050047398, + 0.01342128124088049, + 0.007778817322105169, + 0.022714192047715187, + 0.01097808126360178, + 0.010740003548562527, + 0.014093635603785515, + 0.006639937404543161, + 0.005481111351400614, + 0.008960829116404057, + 0.00660359812900424, + 0.007656133733689785, + 0.010787979699671268, + 0.004114571958780289, + 0.0039015531074255705, + 0.003232147777453065, + 0.013302278704941273, + 0.005061744712293148, + 0.0031151683069765568, + 0.0030708492267876863, + 0.013058795593678951, + 0.005620412994176149, + 0.009631259366869926, + 0.005568130407482386, + 0.004672618117183447, + 0.005324156954884529, + 0.005072203930467367, + 0.005084392614662647, + 0.018744908273220062, + 0.016482453793287277, + 0.011250092647969723, + 0.006751362234354019, + 0.007551851216703653, + 0.008517679758369923, + 0.005860893987119198, + 0.00834219716489315, + 0.008273431099951267, + 0.00850161723792553, + 0.007440631277859211, + 0.004492356441915035, + 0.0040258378721773624, + 0.0054405080154538155, + 0.005685938056558371, + 0.008892214857041836, + 0.01017999742180109, + 0.006884375121444464, + 0.006971342954784632, + 0.003953032195568085, + 0.0030029229819774628, + 0.003917554393410683, + 0.003153514349833131, + 0.008369334042072296, + 0.06838245689868927, + 0.02021288126707077, + 0.07326800376176834, + 0.013483582064509392, + 0.01858087256550789, + 0.014175203628838062, + 0.011015485040843487, + 0.011796609498560429, + 0.05452737584710121, + 0.024103693664073944, + 0.023206451907753944, + 0.014572883024811745, + 0.009164056740701199, + 0.012384019792079926, + 0.011084084399044514, + 0.025756150484085083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0006343498243950307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, angled, Dutch-angle overhead shot closely follows a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels as it deftly navigates a crowded street at a rapid, slightly jerky, game-like pace. The tuk-tuk, with its bright yellow roof, vibrant red body, and sky-blue seats, is covered in a crisp pixel grid. Inside sits a driver, also composed of pixels, whose silhouette is minimalist. The street itself is a dark gray, pixelated asphalt, crowded with equally pixelated motorcycles and blocky cars, while pedestrians, composed of a few pixels, walk along the sidewalk in a fixed, animated loop. The buildings on either side are flat, color-saturated pixel structures dotted with windows and billboards simplified into geometric shapes. The tuk-tuk, making abrupt, near-90-degree turns, finds narrow gaps in the pixelated traffic, cutting in and out of the way with unrealistic precision. The entire world presents a distinct 16-bit retro game aesthetic. Light and shadow are simplified to the superposition of pure color blocks. There are no soft gradients. The highly saturated primary and secondary colors together create a nostalgic atmosphere that is vibrant, slightly chaotic, yet extremely fun to play.", + "seed": 15049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_8/index_9906_seed_15049_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02726041155401617, + "vlm_logprob": -0.02726041155401617, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.02726041155401617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005021341610699892 + ], + "opposite_logprob_tokens": [ + -5.880021572113037 + ], + "logprob_reward": -0.005021341610699892, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.005021341610699892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1093892976641655, + 0.04398145526647568, + 0.07742652297019958, + 0.34098777174949646, + 0.10286670178174973, + 0.06555050611495972, + 0.042484454810619354, + 0.038020212203264236, + 0.07965417206287384, + 0.09963889420032501 + ], + "credit_map": [ + 0.008930710144340992, + 0.008929172530770302, + 0.00876268744468689, + 0.019494760781526566, + 0.011125042103230953, + 0.01114441268146038, + 0.016499685123562813, + 0.024502836167812347, + 0.005107278469949961, + 0.003621994750574231, + 0.004529037978500128, + 0.003895209403708577, + 0.006208632607012987, + 0.0068036154843866825, + 0.007456773892045021, + 0.006358916871249676, + 0.008529211394488811, + 0.006812745239585638, + 0.012422706000506878, + 0.009872647933661938, + 0.009981155395507812, + 0.011529600247740746, + 0.00864399503916502, + 0.009634468704462051, + 0.07753943651914597, + 0.02393156662583351, + 0.022071566432714462, + 0.015298446640372276, + 0.1304735243320465, + 0.0253517497330904, + 0.025898167863488197, + 0.02042333222925663, + 0.014737281948328018, + 0.007952779531478882, + 0.010683729313313961, + 0.003646926488727331, + 0.011086860671639442, + 0.010022955015301704, + 0.010470207780599594, + 0.03426596522331238, + 0.009044029749929905, + 0.011509988456964493, + 0.008631858043372631, + 0.007189585827291012, + 0.00806572288274765, + 0.00822544377297163, + 0.0054245865903794765, + 0.0074592954479157925, + 0.006214607506990433, + 0.006637402344495058, + 0.00585927115753293, + 0.0032087548170238733, + 0.004330283030867577, + 0.009934773668646812, + 0.002607691800221801, + 0.003691678401082754, + 0.004508424550294876, + 0.0043192291632294655, + 0.005849087610840797, + 0.0043754675425589085, + 0.00476152403280139, + 0.007035247515887022, + 0.0025116722099483013, + 0.004659566562622786, + 0.009418432600796223, + 0.007439923472702503, + 0.023411760106682777, + 0.012705398723483086, + 0.007973572239279747, + 0.006710303481668234, + 0.00532485218718648, + 0.006669928319752216, + 0.010321413166821003, + 0.008854511193931103, + 0.012493590824306011, + 0.01716979220509529, + 0.010234220884740353, + 0.00975379440933466, + 0.020458832383155823, + 0.010352746583521366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.005021341610699892 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646904490888119 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008646904490888119, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646904490888119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048358988016843796, + 0.024968452751636505, + 0.06869328022003174, + 0.6271023750305176, + 0.029233714565634727, + 0.06527562439441681, + 0.018044406548142433, + 0.014874688349664211, + 0.03620206564664841, + 0.06724648922681808 + ], + "credit_map": [ + 0.005868894513696432, + 0.004019135609269142, + 0.003644991898909211, + 0.005861278157681227, + 0.005849420092999935, + 0.008468189276754856, + 0.0057961358688771725, + 0.008850940503180027, + 0.002369794063270092, + 0.0022684368304908276, + 0.002719136420637369, + 0.0020454050973057747, + 0.0031510235276073217, + 0.004654017277061939, + 0.004199596121907234, + 0.0035610452760010958, + 0.00894106924533844, + 0.00723332678899169, + 0.009367130696773529, + 0.007499208208173513, + 0.009572743438184261, + 0.008751272223889828, + 0.007918564602732658, + 0.009409966878592968, + 0.127596914768219, + 0.055538080632686615, + 0.03254079818725586, + 0.024091793224215508, + 0.2579290270805359, + 0.04964001476764679, + 0.04494483396410942, + 0.034820880740880966, + 0.0040343152359128, + 0.0018088025972247124, + 0.0031387144699692726, + 0.0013984463876113296, + 0.003139659995213151, + 0.0026767791714519262, + 0.0031801455188542604, + 0.009856849908828735, + 0.007249923422932625, + 0.008390182629227638, + 0.010696733370423317, + 0.0069054290652275085, + 0.0074161384254693985, + 0.00715981237590313, + 0.00750719616189599, + 0.009950211271643639, + 0.0020006962586194277, + 0.001764471991918981, + 0.0026244157925248146, + 0.003185446374118328, + 0.0015459563583135605, + 0.003718125633895397, + 0.0014816292095929384, + 0.0017236651619896293, + 0.0016358481952920556, + 0.002735027577728033, + 0.001808990491554141, + 0.001784747582860291, + 0.002779593225568533, + 0.0021586481016129255, + 0.0009432322694920003, + 0.0010286009637638927, + 0.0036732181906700134, + 0.00579035235568881, + 0.006171812303364277, + 0.004913997370749712, + 0.005775681231170893, + 0.00373635021969676, + 0.003077917732298374, + 0.0030627381056547165, + 0.006226688157767057, + 0.005855077877640724, + 0.009954135864973068, + 0.012925623916089535, + 0.007897242903709412, + 0.007450256496667862, + 0.010660997591912746, + 0.006276469677686691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.008646904490888119 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895220816135406 + ], + "opposite_logprob_tokens": [ + -2.5789523124694824 + ], + "logprob_reward": -0.07895220816135406, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895220816135406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12496522814035416, + 0.04856770485639572, + 0.04272906482219696, + 0.32942983508110046, + 0.18645133078098297, + 0.039684977382421494, + 0.033134568482637405, + 0.04865860939025879, + 0.084163598716259, + 0.062215108424425125 + ], + "credit_map": [ + 0.021296223625540733, + 0.018169092014431953, + 0.00912528857588768, + 0.012244123965501785, + 0.015479238703846931, + 0.014922218397259712, + 0.01552542019635439, + 0.018203619867563248, + 0.004597478546202183, + 0.0049233948811888695, + 0.006246211938560009, + 0.004881125874817371, + 0.006240655668079853, + 0.008137974888086319, + 0.006380061153322458, + 0.007160802371799946, + 0.005324199795722961, + 0.003155134618282318, + 0.006685853935778141, + 0.0069582718424499035, + 0.004972707945853472, + 0.005792100448161364, + 0.00441709766164422, + 0.00542370043694973, + 0.07413573563098907, + 0.024972781538963318, + 0.018571561202406883, + 0.012225557118654251, + 0.13238923251628876, + 0.02516493946313858, + 0.024634674191474915, + 0.01733536832034588, + 0.021147144958376884, + 0.06692326068878174, + 0.012254692614078522, + 0.005129523575305939, + 0.042418237775564194, + 0.015208846889436245, + 0.006678400561213493, + 0.016691219061613083, + 0.004150920081883669, + 0.00524824857711792, + 0.005822907667607069, + 0.004984237253665924, + 0.003892604261636734, + 0.004854531027376652, + 0.0037823973689228296, + 0.006949129048734903, + 0.004878954030573368, + 0.004215717781335115, + 0.004311719443649054, + 0.0037210688460618258, + 0.0035359924659132957, + 0.006585913710296154, + 0.0024031195789575577, + 0.003482084022834897, + 0.004768598824739456, + 0.004656610079109669, + 0.006733157206326723, + 0.005743622779846191, + 0.0073804534040391445, + 0.00940645020455122, + 0.003461767453700304, + 0.00650795130059123, + 0.00964115746319294, + 0.010957969352602959, + 0.025422144681215286, + 0.013513617217540741, + 0.007090666797012091, + 0.004580800421535969, + 0.005347789265215397, + 0.007609446067363024, + 0.005178982391953468, + 0.005068566650152206, + 0.009590179659426212, + 0.011152366176247597, + 0.005904418881982565, + 0.009481065906584263, + 0.009969563223421574, + 0.005869969259947538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.07895220816135406 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043048325926065445 + ], + "opposite_logprob_tokens": [ + -3.16804838180542 + ], + "logprob_reward": -0.043048325926065445, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043048325926065445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09544489532709122, + 0.06016168370842934, + 0.05713210999965668, + 0.09634985774755478, + 0.22175508737564087, + 0.050965480506420135, + 0.056574176996946335, + 0.07335536181926727, + 0.12014000117778778, + 0.1681213229894638 + ], + "credit_map": [ + 0.013062993995845318, + 0.010642478242516518, + 0.006556428037583828, + 0.01079742144793272, + 0.011973537504673004, + 0.011716838926076889, + 0.01592770405113697, + 0.014767497777938843, + 0.005478844512254, + 0.007074199616909027, + 0.004994500428438187, + 0.008137117139995098, + 0.006736076436936855, + 0.008687973953783512, + 0.011176900938153267, + 0.007876069284975529, + 0.01011749729514122, + 0.004872865509241819, + 0.005654884036630392, + 0.005316951312124729, + 0.0077775814570486546, + 0.006722467020153999, + 0.005074517801403999, + 0.01159534603357315, + 0.021397782489657402, + 0.006529012229293585, + 0.008276990614831448, + 0.004363151267170906, + 0.03354329243302345, + 0.008083309046924114, + 0.008091595023870468, + 0.006064726039767265, + 0.022424567490816116, + 0.06596953421831131, + 0.01425006054341793, + 0.006279843859374523, + 0.04736682027578354, + 0.01859235018491745, + 0.010997707955539227, + 0.03587420657277107, + 0.007798674050718546, + 0.006424035411328077, + 0.006669154856353998, + 0.005652044899761677, + 0.007094316650182009, + 0.006930335890501738, + 0.004367015790194273, + 0.006029904820024967, + 0.007755011785775423, + 0.005890942644327879, + 0.009392274543642998, + 0.008187608793377876, + 0.004842592403292656, + 0.00971242319792509, + 0.004002890083938837, + 0.006790430750697851, + 0.00830735731869936, + 0.010823419317603111, + 0.011297805234789848, + 0.009955576620995998, + 0.009055079892277718, + 0.009475747123360634, + 0.006000235676765442, + 0.008440137840807438, + 0.012651213444769382, + 0.012979244813323021, + 0.03159140422940254, + 0.016641857102513313, + 0.012434856034815311, + 0.008523885160684586, + 0.009181857109069824, + 0.0161356832832098, + 0.012246338650584221, + 0.011311459355056286, + 0.026390820741653442, + 0.03468633070588112, + 0.01585967466235161, + 0.02296813391149044, + 0.02876068279147148, + 0.01589788869023323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.043048325926065445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Thai tuk-tuk constructed entirely from chunky, vibrantly colored pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tuk-tuk have a bright yellow roof, a vibrant red body, and sky-blue seats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a Dutch-angle overhead shot that follows the tuk-tuk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tuk-tuk navigate the street with abrupt, near-90-degree turns and unrealistic precision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a distinct 16-bit retro game aesthetic with no soft gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006332775810733438 + ], + "opposite_logprob_tokens": [ + -7.375633239746094 + ], + "logprob_reward": -0.0006332775810733438, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006332775810733438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20366202294826508, + 0.04027972370386124, + 0.10571637004613876, + 0.21871478855609894, + 0.026216426864266396, + 0.0621708482503891, + 0.04056774824857712, + 0.0458710640668869, + 0.08749619126319885, + 0.1693047732114792 + ], + "credit_map": [ + 0.018839426338672638, + 0.020092373713850975, + 0.016966406255960464, + 0.030416255816817284, + 0.0240824893116951, + 0.020502639934420586, + 0.03485345467925072, + 0.03790898621082306, + 0.004664586391299963, + 0.004172659944742918, + 0.003864758415147662, + 0.005144375842064619, + 0.005145059432834387, + 0.0061040474101901054, + 0.005842899903655052, + 0.005341339390724897, + 0.013926148414611816, + 0.008266031742095947, + 0.015577461570501328, + 0.01479410007596016, + 0.012602746486663818, + 0.01234977412968874, + 0.011516128666698933, + 0.016683977097272873, + 0.0472797155380249, + 0.016207939013838768, + 0.013888411223888397, + 0.011396244168281555, + 0.0833345577120781, + 0.016333507373929024, + 0.016395235434174538, + 0.013879172503948212, + 0.0019223153358325362, + 0.0024530922528356314, + 0.0022340419236570597, + 0.0013528052950277925, + 0.003633002983406186, + 0.003212988143786788, + 0.002780115930363536, + 0.008628064766526222, + 0.007373226806521416, + 0.00802698265761137, + 0.011124784126877785, + 0.006991920527070761, + 0.006484380457550287, + 0.006728576961904764, + 0.007852566428482533, + 0.007588408887386322, + 0.005333087872713804, + 0.005479010287672281, + 0.006598516367375851, + 0.006441140081733465, + 0.002971501089632511, + 0.006914766505360603, + 0.002648478141054511, + 0.004181246738880873, + 0.0046741715632379055, + 0.00410182261839509, + 0.009392698295414448, + 0.00552697479724884, + 0.005018784664571285, + 0.008691712282598019, + 0.0031614804174751043, + 0.005303423386067152, + 0.009107068181037903, + 0.013762974180281162, + 0.015938356518745422, + 0.012462574988603592, + 0.007429557852447033, + 0.007289879024028778, + 0.00958013255149126, + 0.01192564982920885, + 0.013754037208855152, + 0.013134638778865337, + 0.02270834520459175, + 0.03314073756337166, + 0.01661059819161892, + 0.022838769480586052, + 0.030376050621271133, + 0.01674160547554493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.0006332775810733438 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_9/index_tensor([1032, 1032])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_9/index_tensor([1032, 1032])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..26757ffd5b8a1e323c7ea5d6d4cd14befc13b00e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_9/index_tensor([1032, 1032])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up macro shot, captured as a high-speed photographic sequence, remains static to record the explosive dispersal of a large seed head. Initially, the frame is filled with the perfectly intricately spherical seed head, its hundreds of delicate, white, feathery papillae interwoven into a fragile sphere, each connected to a dark achene core. The scene is backlit by brilliant afternoon sunlight, causing the silky filaments to glow with a soft, ethereal halo against a very blurred background of vibrant green grass. Suddenly, the sequence captures the violent explosion: the sphere instantly disintegrates, propelling the seeds outward in a chaotic, 360-degree explosion. Each unique, high-resolution photograph in the series freezes a sequential moment of the event—from the initial rupture to the cloud of seeds expanding to fill the frame, their tiny parachutes unfolding and catching the light. The movement is captured so clearly that the spiraling trails of individual seeds are visible, creating a powerful, dramatic, and scientifically accurate visualization of a fleeting natural phenomenon.", + "seed": 7629, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_9/index_1032_seed_7629_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17111086598597466, + "vlm_logprob": -0.17111086598597466, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.17111086598597466, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011734389699995518 + ], + "opposite_logprob_tokens": [ + -5.511734485626221 + ], + "logprob_reward": -0.011734389699995518, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.011734389699995518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19098138809204102, + 0.053891588002443314, + 0.058223702013492584, + 0.03480646014213562, + 0.02701682783663273, + 0.01645324006676674, + 0.14729684591293335, + 0.30668705701828003, + 0.1258058249950409, + 0.038837071508169174 + ], + "credit_map": [ + 0.04796186462044716, + 0.008973287418484688, + 0.025866778567433357, + 0.009589116089046001, + 0.009922081604599953, + 0.044809941202402115, + 0.008442767895758152, + 0.03541555255651474, + 0.01385200023651123, + 0.0027944238390773535, + 0.007086510770022869, + 0.0029810459818691015, + 0.0030065595638006926, + 0.012626848183572292, + 0.002962764585390687, + 0.008581437170505524, + 0.013268636539578438, + 0.002931425580754876, + 0.007377467583864927, + 0.0029237340204417706, + 0.0034625844564288855, + 0.013171925209462643, + 0.002869103802368045, + 0.012218824587762356, + 0.006140470504760742, + 0.0013631603214889765, + 0.004407710395753384, + 0.0022726012393832207, + 0.0015957776922732592, + 0.005908547434955835, + 0.0019812441896647215, + 0.011136947199702263, + 0.004799257963895798, + 0.0011400063522160053, + 0.0039732870645821095, + 0.0013514946913346648, + 0.0012458775890991092, + 0.005940410774201155, + 0.0011802640510722995, + 0.007386229000985622, + 0.0035328639205545187, + 0.0012436945689842105, + 0.0021781495306640863, + 0.0013243495486676693, + 0.001404986483976245, + 0.0023857613559812307, + 0.0012424711603671312, + 0.0031409624498337507, + 0.024052636697888374, + 0.015848902985453606, + 0.020621564239263535, + 0.010504874400794506, + 0.010803879238665104, + 0.015638168901205063, + 0.01381611917167902, + 0.03601069748401642, + 0.1040920540690422, + 0.022029418498277664, + 0.05859251320362091, + 0.015035144984722137, + 0.01568731665611267, + 0.038042739033699036, + 0.011274552904069424, + 0.04193330183625221, + 0.014750906266272068, + 0.006044765468686819, + 0.010936870239675045, + 0.005959334317594767, + 0.008269504643976688, + 0.03746752440929413, + 0.008538030087947845, + 0.03383889049291611, + 0.005977956112474203, + 0.00465532299131155, + 0.006867346353828907, + 0.0035539802629500628, + 0.0036023075226694345, + 0.004803755786269903, + 0.003729612333700061, + 0.005646789912134409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.011734389699995518 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793803095817566 + ], + "opposite_logprob_tokens": [ + -4.634793758392334 + ], + "logprob_reward": -0.009793803095817566, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793803095817566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06651223450899124, + 0.035037219524383545, + 0.025954250246286392, + 0.021795189008116722, + 0.01688213273882866, + 0.008245610632002354, + 0.03584039956331253, + 0.12316696345806122, + 0.6104019284248352, + 0.056164078414440155 + ], + "credit_map": [ + 0.017528926953673363, + 0.0031011595856398344, + 0.009027685038745403, + 0.00396308908239007, + 0.0034522153437137604, + 0.016195980831980705, + 0.0035770011600106955, + 0.00966617837548256, + 0.009755522012710571, + 0.0017984308069571853, + 0.004740027245134115, + 0.0016156063647940755, + 0.0018710409058257937, + 0.008458542637526989, + 0.001768662128597498, + 0.005029385909438133, + 0.005160737317055464, + 0.0011326012900099158, + 0.003440215950831771, + 0.0014070238685235381, + 0.0012150747934356332, + 0.00505853071808815, + 0.0012105271453037858, + 0.007329537067562342, + 0.0035863120574504137, + 0.0009103842312470078, + 0.0025547181721776724, + 0.001583681209012866, + 0.0011176420375704765, + 0.003922785632312298, + 0.0012908568605780602, + 0.006828807760030031, + 0.0028404947370290756, + 0.000839848245959729, + 0.0023706594947725534, + 0.001168385031633079, + 0.000756829627789557, + 0.002690813271328807, + 0.00113222049549222, + 0.005082882009446621, + 0.001449235831387341, + 0.0005787353147752583, + 0.0011811087606474757, + 0.000887888134457171, + 0.0006168887484818697, + 0.001156939659267664, + 0.0007943678647279739, + 0.0015804462600499392, + 0.006875430699437857, + 0.0036259216722100973, + 0.003912785556167364, + 0.003839648561552167, + 0.003158962121233344, + 0.0042463745921850204, + 0.003499282756820321, + 0.006681992672383785, + 0.04461197555065155, + 0.00862666592001915, + 0.022628961130976677, + 0.005990930367261171, + 0.006072122137993574, + 0.015261240303516388, + 0.004626828711479902, + 0.01534823328256607, + 0.06559252738952637, + 0.016354138031601906, + 0.058075569570064545, + 0.014032918959856033, + 0.029207048937678337, + 0.21717429161071777, + 0.019398488104343414, + 0.19056691229343414, + 0.009465133771300316, + 0.005011968780308962, + 0.010334367863833904, + 0.005133986473083496, + 0.004069976042956114, + 0.007429927587509155, + 0.005568372551351786, + 0.009150344878435135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.009793803095817566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673760985955596 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.00673760985955596, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673760985955596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1559935361146927, + 0.07903687655925751, + 0.033615365624427795, + 0.17718340456485748, + 0.02139352820813656, + 0.00970886554569006, + 0.1199643686413765, + 0.1585533767938614, + 0.22199244797229767, + 0.02255830727517605 + ], + "credit_map": [ + 0.01994474232196808, + 0.004988722503185272, + 0.0220644511282444, + 0.006764071993529797, + 0.0068558151833713055, + 0.023270096629858017, + 0.008729802444577217, + 0.06337582319974899, + 0.011300337500870228, + 0.0022387844510376453, + 0.011093533597886562, + 0.0028441979084163904, + 0.0032202403526753187, + 0.012616122141480446, + 0.0033411409240216017, + 0.032382506877183914, + 0.006888418458402157, + 0.0019436171278357506, + 0.004707658663392067, + 0.002001179149374366, + 0.0027573485858738422, + 0.006637189537286758, + 0.0023838263005018234, + 0.006296122912317514, + 0.02871950902044773, + 0.006136036943644285, + 0.023658810183405876, + 0.0086072301492095, + 0.006331711541861296, + 0.030920252203941345, + 0.0075221494771540165, + 0.06528769433498383, + 0.005005724728107452, + 0.0010487863328307867, + 0.0028672374319285154, + 0.0013255189405754209, + 0.00108354224357754, + 0.0053686657920479774, + 0.0011666642967611551, + 0.0035273844841867685, + 0.0018752918113023043, + 0.0006474251276813447, + 0.0012536668218672276, + 0.0007317563286051154, + 0.0007220437983050942, + 0.001488598296418786, + 0.0007074931636452675, + 0.002282590139657259, + 0.01698712445795536, + 0.007845851592719555, + 0.01976686343550682, + 0.006890404503792524, + 0.007725463714450598, + 0.014327675104141235, + 0.00918503012508154, + 0.03723595291376114, + 0.05845773220062256, + 0.009621655568480492, + 0.029294723644852638, + 0.007100858259946108, + 0.00813534576445818, + 0.018962515518069267, + 0.00655206898227334, + 0.020428471267223358, + 0.030168810859322548, + 0.0055345105938613415, + 0.02398873120546341, + 0.005050773732364178, + 0.009714994579553604, + 0.07601170241832733, + 0.007288143038749695, + 0.06423477083444595, + 0.002783974399790168, + 0.0021239432971924543, + 0.002694417256861925, + 0.0021264164242893457, + 0.0023178127594292164, + 0.0030177216976881027, + 0.003214387223124504, + 0.004279632121324539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.00673760985955596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261085748672485 + ], + "opposite_logprob_tokens": [ + -0.5761085748672485 + ], + "logprob_reward": -0.8261085748672485, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261085748672485, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.024751851335167885, + 0.021753080189228058, + 0.03295197710394859, + 0.014459274709224701, + 0.015207392163574696, + 0.006496166344732046, + 0.0366586335003376, + 0.19424717128276825, + 0.6238796710968018, + 0.029594751074910164 + ], + "credit_map": [ + 0.004220194648951292, + 0.0012185125378891826, + 0.0031586219556629658, + 0.0015662397490814328, + 0.0019388364162296057, + 0.004607944283634424, + 0.002181768650189042, + 0.005859734024852514, + 0.0050988043658435345, + 0.000953959533944726, + 0.0029545752331614494, + 0.0015189404366537929, + 0.0012096422724425793, + 0.0054456316865980625, + 0.0015255322214215994, + 0.003045993158593774, + 0.007314146962016821, + 0.00162347755394876, + 0.004189173690974712, + 0.0021402882412075996, + 0.0018881655996665359, + 0.007803978398442268, + 0.0016320173162966967, + 0.0063607278279960155, + 0.002051842864602804, + 0.0005509743932634592, + 0.0023241958115249872, + 0.0005884277052246034, + 0.0005698393797501922, + 0.0025351173244416714, + 0.0006465850747190416, + 0.005192291922867298, + 0.0031274703796952963, + 0.0007159814704209566, + 0.0018033396918326616, + 0.0007219494436867535, + 0.0008289121906273067, + 0.003335084766149521, + 0.0008805471588857472, + 0.0037941073533147573, + 0.0012322042603045702, + 0.0004858499742113054, + 0.0008344457601197064, + 0.0004877372703049332, + 0.00048364404938183725, + 0.0008733160211704671, + 0.0004705941246356815, + 0.0016283747972920537, + 0.006990157533437014, + 0.0024763925466686487, + 0.004578050225973129, + 0.0032027259003371, + 0.002620832296088338, + 0.004701072815805674, + 0.003263267455622554, + 0.008826133795082569, + 0.06958962976932526, + 0.010918842628598213, + 0.037475842982530594, + 0.007046851329505444, + 0.009800381027162075, + 0.022521233186125755, + 0.006422344129532576, + 0.03047204576432705, + 0.06978842616081238, + 0.016787322238087654, + 0.06096600368618965, + 0.015174610540270805, + 0.030869361013174057, + 0.21951055526733398, + 0.02189003862440586, + 0.1888934075832367, + 0.004694693721830845, + 0.0017378802876919508, + 0.00535048171877861, + 0.0024844578001648188, + 0.002662024227902293, + 0.003847292857244611, + 0.003882682416588068, + 0.004935238044708967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.8261085748672485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011799524072557688 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011799524072557688, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011799524072557688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08930961042642593, + 0.17876024544239044, + 0.1109139546751976, + 0.056234125047922134, + 0.05653353035449982, + 0.04985834285616875, + 0.14749489724636078, + 0.0924704521894455, + 0.11983955651521683, + 0.09858530014753342 + ], + "credit_map": [ + 0.016610363498330116, + 0.004615189973264933, + 0.01215363759547472, + 0.007986270822584629, + 0.005593855399638414, + 0.01765481010079384, + 0.00508718891069293, + 0.01960829272866249, + 0.046695560216903687, + 0.008412186056375504, + 0.02478817291557789, + 0.008441896177828312, + 0.009674042463302612, + 0.04741602763533592, + 0.0067845843732357025, + 0.026547765359282494, + 0.02141772396862507, + 0.005281706340610981, + 0.014147637411952019, + 0.004895388148725033, + 0.006620065774768591, + 0.021680515259504318, + 0.004325942602008581, + 0.03254497051239014, + 0.012501118704676628, + 0.003002858255058527, + 0.008072401396930218, + 0.0036955829709768295, + 0.0029516874346882105, + 0.01442575454711914, + 0.0035256892442703247, + 0.00805902760475874, + 0.010429156012833118, + 0.002950526773929596, + 0.0077614993788301945, + 0.0037131637800484896, + 0.0029541444964706898, + 0.011207113973796368, + 0.0032097797375172377, + 0.014308140613138676, + 0.007351805455982685, + 0.0028028676752001047, + 0.008521432057023048, + 0.00410080561414361, + 0.0026666042394936085, + 0.00892558041960001, + 0.002436842070892453, + 0.01305240485817194, + 0.028133153915405273, + 0.00923561304807663, + 0.025784162804484367, + 0.00762238260358572, + 0.008126799948513508, + 0.024400748312473297, + 0.00837967824190855, + 0.035812363028526306, + 0.024224039167165756, + 0.004955760668963194, + 0.01932804100215435, + 0.006009229458868504, + 0.006312698591500521, + 0.007945295423269272, + 0.007120445370674133, + 0.016574935987591743, + 0.02009325847029686, + 0.005861102137714624, + 0.0185116995126009, + 0.006090726237744093, + 0.006324274465441704, + 0.024058973416686058, + 0.008467183448374271, + 0.03043232299387455, + 0.0159041415899992, + 0.012277303263545036, + 0.016309263184666634, + 0.01054780837148428, + 0.007463471964001656, + 0.013804938644170761, + 0.012319879606366158, + 0.009958488866686821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.0011799524072557688 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro shot, captured as a high-speed photographic sequence, remains static to record the explosive dispersal of a large seed head. Initially, the frame is filled with the perfectly intricately spherical seed head, its hundreds of delicate, white, feathery papillae interwoven into a fragile sphere, each connected to a dark achene core. The scene is backlit by brilliant afternoon sunlight, causing the silky filaments to glow with a soft, ethereal halo against a very blurred background of vibrant green grass. Suddenly, the sequence captures the violent explosion: the sphere instantly disintegrates, propelling the seeds outward in a chaotic, 360-degree explosion. Each unique, high-resolution photograph in the series freezes a sequential moment of the event—from the initial rupture to the cloud of seeds expanding to fill the frame, their tiny parachutes unfolding and catching the light. The movement is captured so clearly that the spiraling trails of individual seeds are visible, creating a powerful, dramatic, and scientifically accurate visualization of a fleeting natural phenomenon.", + "seed": 7630, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_9/index_1032_seed_7630_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009567210380919278, + "vlm_logprob": -0.009567210380919278, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.009567210380919278, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004235939122736454 + ], + "opposite_logprob_tokens": [ + -6.879235744476318 + ], + "logprob_reward": -0.004235939122736454, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.004235939122736454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04954519867897034, + 0.08174919337034225, + 0.023372899740934372, + 0.060214921832084656, + 0.33030250668525696, + 0.04352523759007454, + 0.037216916680336, + 0.14989779889583588, + 0.13686127960681915, + 0.08731411397457123 + ], + "credit_map": [ + 0.006822922267019749, + 0.0037546693347394466, + 0.008981048129498959, + 0.004375135060399771, + 0.0051622651517391205, + 0.005970812868326902, + 0.006945926230400801, + 0.007532423362135887, + 0.009618719108402729, + 0.009515619836747646, + 0.011159672401845455, + 0.006559488829225302, + 0.012349632568657398, + 0.010523125529289246, + 0.010487535037100315, + 0.01153540425002575, + 0.005753623321652412, + 0.0018816738156601787, + 0.0045743174850940704, + 0.0015306846471503377, + 0.002413178328424692, + 0.0028040525503456593, + 0.0015570688992738724, + 0.0028582988306879997, + 0.011157356202602386, + 0.005184879992157221, + 0.011973231099545956, + 0.004867864307016134, + 0.007635562680661678, + 0.006406555417925119, + 0.009286011569201946, + 0.003703461028635502, + 0.023061905056238174, + 0.02129204384982586, + 0.13107788562774658, + 0.013311528600752354, + 0.021565597504377365, + 0.09859417378902435, + 0.008059737272560596, + 0.013339603319764137, + 0.004445679951459169, + 0.008238607086241245, + 0.003110570600256324, + 0.003901078598573804, + 0.006834068801254034, + 0.006448866333812475, + 0.005047130398452282, + 0.005499234888702631, + 0.003491475246846676, + 0.006482212338596582, + 0.002649627858772874, + 0.0036138922441750765, + 0.006001215428113937, + 0.006658040452748537, + 0.003924292977899313, + 0.004396157339215279, + 0.0058484976179897785, + 0.048756156116724014, + 0.0038767389487475157, + 0.006161962635815144, + 0.007813282310962677, + 0.05958721786737442, + 0.006451442372053862, + 0.011402497068047523, + 0.006016656290739775, + 0.0252755768597126, + 0.005247098859399557, + 0.05796550214290619, + 0.005815759301185608, + 0.009678061120212078, + 0.019382281228899956, + 0.007480339612811804, + 0.0037034007254987955, + 0.0057113151997327805, + 0.005241325590759516, + 0.021100062876939774, + 0.0058396970853209496, + 0.010822842828929424, + 0.029744548723101616, + 0.005150922108441591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.004235939122736454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100131336599588 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100131336599588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100131336599588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07795441150665283, + 0.047356415539979935, + 0.0489007942378521, + 0.08651458472013474, + 0.27333930134773254, + 0.05252170190215111, + 0.027064595371484756, + 0.23263178765773773, + 0.10719077289104462, + 0.04652564972639084 + ], + "credit_map": [ + 0.0096749197691679, + 0.006275502499192953, + 0.015460641123354435, + 0.006374180316925049, + 0.00909226480871439, + 0.008618767373263836, + 0.011044326238334179, + 0.011413810774683952, + 0.005561911500990391, + 0.004696660675108433, + 0.006141006015241146, + 0.0034544169902801514, + 0.007965375669300556, + 0.007341199088841677, + 0.005314424633979797, + 0.006881421431899071, + 0.011128352954983711, + 0.004361311439424753, + 0.012026755139231682, + 0.0029843980446457863, + 0.003193984506651759, + 0.006604881025850773, + 0.003422665875405073, + 0.005178442224860191, + 0.017551014199852943, + 0.0090590575709939, + 0.019209153950214386, + 0.007326449733227491, + 0.006950709503144026, + 0.010229832492768764, + 0.010396270081400871, + 0.005792091600596905, + 0.01839272677898407, + 0.018618470057845116, + 0.10580568015575409, + 0.012201888486742973, + 0.018378091976046562, + 0.08057650178670883, + 0.008005584590137005, + 0.011360355652868748, + 0.005929980427026749, + 0.011670074425637722, + 0.0037395090330392122, + 0.004258653149008751, + 0.007360948249697685, + 0.007533140480518341, + 0.005021811928600073, + 0.007007586304098368, + 0.002035722602158785, + 0.004177388269454241, + 0.002146750455722213, + 0.002378924749791622, + 0.006049128714948893, + 0.003495403565466404, + 0.0032189430203288794, + 0.003562332596629858, + 0.009104088880121708, + 0.0818747952580452, + 0.005595468450337648, + 0.008939127437770367, + 0.014105593785643578, + 0.0922321230173111, + 0.00918029434978962, + 0.011600293219089508, + 0.0043490054085850716, + 0.012965449132025242, + 0.0048975832760334015, + 0.04231902211904526, + 0.005634128116071224, + 0.014416741207242012, + 0.015234313905239105, + 0.00737452507019043, + 0.0045967139303684235, + 0.0026210960932075977, + 0.008162061683833599, + 0.007012228015810251, + 0.0065102423541247845, + 0.005152287892997265, + 0.008107181638479233, + 0.0043638404458761215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.004100131336599588 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463957479223609 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.00463957479223609, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463957479223609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04541942477226257, + 0.030255986377596855, + 0.03334450721740723, + 0.025506064295768738, + 0.6129500865936279, + 0.07840944826602936, + 0.02059236541390419, + 0.06904160231351852, + 0.04734140262007713, + 0.03713906183838844 + ], + "credit_map": [ + 0.007740156725049019, + 0.002804746385663748, + 0.01016194000840187, + 0.0034360734280198812, + 0.004300383850932121, + 0.004114641807973385, + 0.005279260221868753, + 0.007582227233797312, + 0.004714990966022015, + 0.0028132048901170492, + 0.005402892827987671, + 0.002190748695284128, + 0.0041633490473032, + 0.0029236122500151396, + 0.004274954553693533, + 0.0037722373381257057, + 0.006383426953107119, + 0.002826171927154064, + 0.005619509611278772, + 0.0019269466865807772, + 0.0032508629374206066, + 0.005924687720835209, + 0.002922413870692253, + 0.00449049286544323, + 0.004343186039477587, + 0.00293338461779058, + 0.004253951832652092, + 0.0022342344745993614, + 0.003711260622367263, + 0.0025594786275178194, + 0.003762315958738327, + 0.0017082546837627888, + 0.044326119124889374, + 0.03969591110944748, + 0.2399781197309494, + 0.025104613974690437, + 0.039557162672281265, + 0.1799807995557785, + 0.019290393218398094, + 0.025017065927386284, + 0.007470281794667244, + 0.015523083508014679, + 0.00591342244297266, + 0.007269629277288914, + 0.01422237977385521, + 0.01086663268506527, + 0.008750587701797485, + 0.00839343760162592, + 0.001814369112253189, + 0.003396610263735056, + 0.0016677308594807982, + 0.0018787673907354474, + 0.005030995700508356, + 0.0025845966301858425, + 0.0016862940974533558, + 0.002533003920689225, + 0.0031145953107625246, + 0.013325965963304043, + 0.0032219213899224997, + 0.006483220960944891, + 0.0068472204729914665, + 0.021332740783691406, + 0.005432180594652891, + 0.009283767081797123, + 0.0023268582299351692, + 0.004585492890328169, + 0.002176063135266304, + 0.01944580115377903, + 0.0025329836644232273, + 0.007097449619323015, + 0.00603852653875947, + 0.003138232044875622, + 0.002573759527876973, + 0.002136891707777977, + 0.0040702675469219685, + 0.009436123073101044, + 0.005282175727188587, + 0.003029037266969681, + 0.008156146854162216, + 0.002454663161188364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.00463957479223609 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033680811524391174 + ], + "opposite_logprob_tokens": [ + -3.4086809158325195 + ], + "logprob_reward": -0.033680811524391174, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033680811524391174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012820085510611534, + 0.033563002943992615, + 0.03482736647129059, + 0.09139111638069153, + 0.535620927810669, + 0.07610227167606354, + 0.039976298809051514, + 0.046934857964515686, + 0.10260381549596786, + 0.026160256937146187 + ], + "credit_map": [ + 0.001576912240125239, + 0.0012789919273927808, + 0.002427608473226428, + 0.0014874535845592618, + 0.0013533071614801884, + 0.0013045943342149258, + 0.0016150107840076089, + 0.001776207354851067, + 0.004318065941333771, + 0.0038406625390052795, + 0.0052371518686413765, + 0.0024481695145368576, + 0.004898589104413986, + 0.0038654583040624857, + 0.00382724916562438, + 0.005127656273543835, + 0.008347691036760807, + 0.0025174508336931467, + 0.009618796408176422, + 0.0021078402642160654, + 0.0021967110224068165, + 0.0034861459862440825, + 0.0029982859268784523, + 0.003554444294422865, + 0.01760660670697689, + 0.012062778696417809, + 0.018166575580835342, + 0.009809484705328941, + 0.007225823123008013, + 0.012024122290313244, + 0.007958200760185719, + 0.0065375203266739845, + 0.035583868622779846, + 0.035624291747808456, + 0.20710748434066772, + 0.024162499234080315, + 0.03605801984667778, + 0.15964344143867493, + 0.01432651188224554, + 0.023114817216992378, + 0.008472952991724014, + 0.01233056653290987, + 0.006455415394157171, + 0.007113552652299404, + 0.01005761418491602, + 0.010506127960979939, + 0.009444175288081169, + 0.011721867136657238, + 0.002931882394477725, + 0.006502468138933182, + 0.003312965389341116, + 0.0058157602325081825, + 0.008222965523600578, + 0.005091237369924784, + 0.002237646607682109, + 0.005861373618245125, + 0.0037070410326123238, + 0.008924995549023151, + 0.0041448320262134075, + 0.004269023425877094, + 0.005221263039857149, + 0.008295568637549877, + 0.004891966935247183, + 0.007480165455490351, + 0.005757302511483431, + 0.0213758684694767, + 0.0043278951197862625, + 0.04221617430448532, + 0.004408741369843483, + 0.006998996715992689, + 0.009519577026367188, + 0.007999257184565067, + 0.002264639362692833, + 0.001463517197407782, + 0.003152006771415472, + 0.004853562917560339, + 0.0022529722191393375, + 0.0028175064362585545, + 0.00679708132520318, + 0.0025589719880372286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.033680811524391174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001179595128633082 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.001179595128633082, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001179595128633082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05868219956755638, + 0.058244265615940094, + 0.031084351241588593, + 0.029271969571709633, + 0.3087434470653534, + 0.05985414981842041, + 0.03419291228055954, + 0.15611089766025543, + 0.10354152321815491, + 0.16027428209781647 + ], + "credit_map": [ + 0.015924248844385147, + 0.0034252223558723927, + 0.011734511703252792, + 0.0037986382376402617, + 0.003614900168031454, + 0.005341163370758295, + 0.005254322662949562, + 0.00958919059485197, + 0.013695122674107552, + 0.004990013781934977, + 0.009316076524555683, + 0.0030530262738466263, + 0.0054502845741808414, + 0.005908951163291931, + 0.0049514188431203365, + 0.01087937317788601, + 0.0070380158722400665, + 0.0022000912576913834, + 0.008570066653192043, + 0.0017165385652333498, + 0.0024630597326904535, + 0.0028858440928161144, + 0.0027836966328322887, + 0.003427038434892893, + 0.004963821265846491, + 0.005349739454686642, + 0.003648802638053894, + 0.0036043424624949694, + 0.002499957336112857, + 0.004053054843097925, + 0.002166370628401637, + 0.002985881408676505, + 0.023493396118283272, + 0.02028711885213852, + 0.12432242184877396, + 0.011530574411153793, + 0.018648508936166763, + 0.09227026998996735, + 0.006097956094890833, + 0.012093206867575645, + 0.006530322600156069, + 0.012416196055710316, + 0.004239160567522049, + 0.005329896695911884, + 0.009538951329886913, + 0.009258687496185303, + 0.005212281830608845, + 0.007328652776777744, + 0.003434828482568264, + 0.006048258394002914, + 0.003328740131109953, + 0.003625422017648816, + 0.007156797219067812, + 0.004067960660904646, + 0.002829669276252389, + 0.0037012360990047455, + 0.0062077040784060955, + 0.0412912517786026, + 0.004197654780000448, + 0.0055067152716219425, + 0.009072993882000446, + 0.06148427352309227, + 0.008236156776547432, + 0.020114142447710037, + 0.006429712288081646, + 0.024234363809227943, + 0.007055596448481083, + 0.02249414101243019, + 0.010132480412721634, + 0.016604747623205185, + 0.008943120948970318, + 0.007647362072020769, + 0.0065992772579193115, + 0.008229185827076435, + 0.01199421938508749, + 0.047703228890895844, + 0.009610232897102833, + 0.009452365338802338, + 0.05743589252233505, + 0.009249881841242313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.001179595128633082 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro shot, captured as a high-speed photographic sequence, remains static to record the explosive dispersal of a large seed head. Initially, the frame is filled with the perfectly intricately spherical seed head, its hundreds of delicate, white, feathery papillae interwoven into a fragile sphere, each connected to a dark achene core. The scene is backlit by brilliant afternoon sunlight, causing the silky filaments to glow with a soft, ethereal halo against a very blurred background of vibrant green grass. Suddenly, the sequence captures the violent explosion: the sphere instantly disintegrates, propelling the seeds outward in a chaotic, 360-degree explosion. Each unique, high-resolution photograph in the series freezes a sequential moment of the event—from the initial rupture to the cloud of seeds expanding to fill the frame, their tiny parachutes unfolding and catching the light. The movement is captured so clearly that the spiraling trails of individual seeds are visible, creating a powerful, dramatic, and scientifically accurate visualization of a fleeting natural phenomenon.", + "seed": 7631, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_9/index_1032_seed_7631_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08539423973998055, + "vlm_logprob": -0.08539423973998055, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.08539423973998055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012287630699574947 + ], + "opposite_logprob_tokens": [ + -6.012287616729736 + ], + "logprob_reward": -0.012287630699574947, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.012287630699574947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20722602307796478, + 0.09778351336717606, + 0.05649822577834129, + 0.10653843730688095, + 0.097012460231781, + 0.13918790221214294, + 0.053043466061353683, + 0.1509547233581543, + 0.036482296884059906, + 0.05527297407388687 + ], + "credit_map": [ + 0.058118440210819244, + 0.010073582641780376, + 0.040132082998752594, + 0.013315431773662567, + 0.008286016061902046, + 0.03719677776098251, + 0.013395542278885841, + 0.026708140969276428, + 0.015700435265898705, + 0.004183600656688213, + 0.01945207640528679, + 0.006150207947939634, + 0.005767100490629673, + 0.025938406586647034, + 0.007830118760466576, + 0.012761573307216167, + 0.010790195316076279, + 0.003434219863265753, + 0.011517428793013096, + 0.004039946477860212, + 0.004128764849156141, + 0.00678325816988945, + 0.003495597280561924, + 0.012308817356824875, + 0.00833678338676691, + 0.005760579835623503, + 0.022613447159528732, + 0.005962274502962828, + 0.009101669304072857, + 0.036914121359586716, + 0.005961783230304718, + 0.011887779459357262, + 0.022127192467451096, + 0.008890436962246895, + 0.01614486798644066, + 0.016757139936089516, + 0.007203655783087015, + 0.006927486974745989, + 0.012226139195263386, + 0.00673554465174675, + 0.01734674535691738, + 0.01097020972520113, + 0.016209466382861137, + 0.015947408974170685, + 0.009061351418495178, + 0.03883730247616768, + 0.012447146698832512, + 0.018368272110819817, + 0.0073100971058011055, + 0.004721485078334808, + 0.0060473280027508736, + 0.005561982747167349, + 0.005573252681642771, + 0.010613116435706615, + 0.0045416513457894325, + 0.008674551732838154, + 0.013072838075459003, + 0.015317694284021854, + 0.012440165504813194, + 0.023857543244957924, + 0.013334013521671295, + 0.01431182213127613, + 0.04003444314002991, + 0.01858619414269924, + 0.005865845829248428, + 0.0036274094600230455, + 0.005061924923211336, + 0.0038997000083327293, + 0.003757564816623926, + 0.004296792671084404, + 0.002971370005980134, + 0.0070016877725720406, + 0.007478863932192326, + 0.010654293932020664, + 0.004436085000634193, + 0.006090421695262194, + 0.006854930892586708, + 0.006386908702552319, + 0.005176573526114225, + 0.008194895461201668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.012287630699574947 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550304643809795 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.012550304643809795, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550304643809795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08341420441865921, + 0.15248920023441315, + 0.04022924602031708, + 0.06208622828125954, + 0.10937542468309402, + 0.4260771572589874, + 0.021417012438178062, + 0.06184316799044609, + 0.021333128213882446, + 0.0217352956533432 + ], + "credit_map": [ + 0.021466199308633804, + 0.004375224933028221, + 0.026699021458625793, + 0.004320249892771244, + 0.00427826726809144, + 0.007121719885617495, + 0.00484953448176384, + 0.01030396856367588, + 0.04174211621284485, + 0.009207861497998238, + 0.012841369025409222, + 0.012803759425878525, + 0.011104932054877281, + 0.027116762474179268, + 0.01655377447605133, + 0.02111862041056156, + 0.009867058135569096, + 0.001981810899451375, + 0.007053549401462078, + 0.0018856944516301155, + 0.0023147340398281813, + 0.007102287374436855, + 0.0018998414743691683, + 0.008124266751110554, + 0.0038169906474649906, + 0.0035268678329885006, + 0.009913943707942963, + 0.004010545089840889, + 0.004343979526311159, + 0.021015314385294914, + 0.0036511982325464487, + 0.011807379312813282, + 0.03601463511586189, + 0.006804139353334904, + 0.030169950798153877, + 0.01054276805371046, + 0.004941140301525593, + 0.006386150605976582, + 0.007956170476973057, + 0.006560466252267361, + 0.09920192509889603, + 0.014962781220674515, + 0.058151666074991226, + 0.01322336494922638, + 0.01906142756342888, + 0.16718855500221252, + 0.012915544211864471, + 0.04137185588479042, + 0.002699560485780239, + 0.0017500538378953934, + 0.00262671266682446, + 0.001790623995475471, + 0.0017700358293950558, + 0.004385246429592371, + 0.002707498613744974, + 0.0036872774362564087, + 0.006475868634879589, + 0.005810053087770939, + 0.008449842222034931, + 0.008066249080002308, + 0.007650892715901136, + 0.006601131521165371, + 0.012032310478389263, + 0.006756812799721956, + 0.003384457901120186, + 0.0020193762611597776, + 0.0030167342629283667, + 0.0016245893202722073, + 0.002735647838562727, + 0.0024624764919281006, + 0.0023177829571068287, + 0.0037720606196671724, + 0.002785125747323036, + 0.003948540892452002, + 0.002003471367061138, + 0.002168188104405999, + 0.0029445632826536894, + 0.0026342635974287987, + 0.0023184448946267366, + 0.002932692877948284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.012550304643809795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194829389452934 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014194829389452934, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194829389452934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3374277949333191, + 0.09646281599998474, + 0.017470471560955048, + 0.12272127717733383, + 0.03402763977646828, + 0.17555566132068634, + 0.01803266629576683, + 0.095073401927948, + 0.06181564927101135, + 0.041412610560655594 + ], + "credit_map": [ + 0.05335789918899536, + 0.012105383910238743, + 0.11740203946828842, + 0.01560965832322836, + 0.0116499038413167, + 0.09737242758274078, + 0.011897050775587559, + 0.018033429980278015, + 0.01924019120633602, + 0.0060631814412772655, + 0.02173641137778759, + 0.008612899109721184, + 0.005897145252674818, + 0.015060513280332088, + 0.010147086344659328, + 0.00970539078116417, + 0.0021835253573954105, + 0.0017545036971569061, + 0.0015975326532498002, + 0.0028240883257240057, + 0.0017842691158875823, + 0.0022451227996498346, + 0.0023574335500597954, + 0.0027239969931542873, + 0.015609986148774624, + 0.004612117540091276, + 0.028714541345834732, + 0.004522989038378, + 0.007239441387355328, + 0.04930761456489563, + 0.004348985385149717, + 0.008365601301193237, + 0.003619984956458211, + 0.004203019198030233, + 0.0038239455316215754, + 0.005591386929154396, + 0.0029819810297340155, + 0.0035940902307629585, + 0.00627704244107008, + 0.003936188295483589, + 0.041937943547964096, + 0.00816916674375534, + 0.025402143597602844, + 0.005284540355205536, + 0.009395414032042027, + 0.06693089008331299, + 0.0047139450907707214, + 0.013721618801355362, + 0.002707612933591008, + 0.0013157122302800417, + 0.002366220811381936, + 0.001263509038835764, + 0.0010928220581263304, + 0.005880947690457106, + 0.001389558194205165, + 0.0020162833388894796, + 0.00829332135617733, + 0.008816777728497982, + 0.011701068840920925, + 0.013264253735542297, + 0.011038308963179588, + 0.009678568691015244, + 0.02231682650744915, + 0.009964271448552608, + 0.008358972147107124, + 0.004702080972492695, + 0.008469398133456707, + 0.004800507333129644, + 0.0071380846202373505, + 0.00732693076133728, + 0.006949730683118105, + 0.014069945551455021, + 0.0050014713779091835, + 0.0063910880126059055, + 0.0036462561693042517, + 0.004512708634138107, + 0.00833925697952509, + 0.004363593645393848, + 0.005026761908084154, + 0.0041314722038805485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.014194829389452934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38701868057250977 + ], + "opposite_logprob_tokens": [ + -1.1370186805725098 + ], + "logprob_reward": -0.38701868057250977, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38701868057250977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13288219273090363, + 0.09853781759738922, + 0.03170990198850632, + 0.05355115979909897, + 0.07077442109584808, + 0.41261762380599976, + 0.043725088238716125, + 0.0871472954750061, + 0.03628842532634735, + 0.03276601806282997 + ], + "credit_map": [ + 0.017100492492318153, + 0.00378097640350461, + 0.036097072064876556, + 0.0058831204660236835, + 0.005024208687245846, + 0.048717524856328964, + 0.006197354756295681, + 0.010081450454890728, + 0.014093633741140366, + 0.003061716677621007, + 0.022801415994763374, + 0.004439240321516991, + 0.00447609880939126, + 0.03775487095117569, + 0.005617768038064241, + 0.006293082144111395, + 0.008060039952397346, + 0.0014071695040911436, + 0.006195284426212311, + 0.0019391090609133244, + 0.0018619004404172301, + 0.0045371088199317455, + 0.0018297219648957253, + 0.005879569333046675, + 0.005564800463616848, + 0.004551089834421873, + 0.01224861666560173, + 0.004466583486646414, + 0.005863694939762354, + 0.0076789031736552715, + 0.004627540707588196, + 0.008549927733838558, + 0.014732452109456062, + 0.009076617658138275, + 0.012367465533316135, + 0.010213705711066723, + 0.0070534381084144115, + 0.005227216519415379, + 0.0071358149871230125, + 0.004967722110450268, + 0.09712492674589157, + 0.014561399817466736, + 0.05616036430001259, + 0.011208916082978249, + 0.018701350316405296, + 0.16105102002620697, + 0.01369075570255518, + 0.04011891782283783, + 0.005956294946372509, + 0.003967640921473503, + 0.005123000591993332, + 0.0035340224858373404, + 0.0037001462187618017, + 0.009801565669476986, + 0.00425534974783659, + 0.007387071382254362, + 0.00646279426291585, + 0.008158184587955475, + 0.009688661433756351, + 0.012820243835449219, + 0.00955691933631897, + 0.008456568233668804, + 0.02204744890332222, + 0.009956484660506248, + 0.004531762562692165, + 0.0036307284608483315, + 0.004861060064285994, + 0.003308095969259739, + 0.0036830671597272158, + 0.004575890488922596, + 0.004099169280380011, + 0.007598654367029667, + 0.004312117118388414, + 0.005265591200441122, + 0.0027707910630851984, + 0.0035429152194410563, + 0.0045752739533782005, + 0.004228705540299416, + 0.003756089834496379, + 0.004314534831792116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.38701868057250977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009197533945553005 + ], + "opposite_logprob_tokens": [ + -7.000919818878174 + ], + "logprob_reward": -0.0009197533945553005, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009197533945553005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1846858710050583, + 0.04953545704483986, + 0.058662693947553635, + 0.09674181789159775, + 0.15790407359600067, + 0.14110325276851654, + 0.032628126442432404, + 0.06230006366968155, + 0.029565012082457542, + 0.186873659491539 + ], + "credit_map": [ + 0.061903852969408035, + 0.008427757769823074, + 0.05999002978205681, + 0.00842333398759365, + 0.005975495092570782, + 0.014942706562578678, + 0.008090654388070107, + 0.016932053491473198, + 0.004643920809030533, + 0.002294258214533329, + 0.01492238137871027, + 0.0026515559293329716, + 0.0021691322326660156, + 0.016401134431362152, + 0.0020537818782031536, + 0.004399292636662722, + 0.004561998415738344, + 0.003655108856037259, + 0.006583699956536293, + 0.002314188750460744, + 0.0040184794925153255, + 0.015245206654071808, + 0.002766571007668972, + 0.019517436623573303, + 0.01049212459474802, + 0.004575404338538647, + 0.02330951578915119, + 0.003871041117236018, + 0.005011694505810738, + 0.012407523579895496, + 0.004988855682313442, + 0.032085660845041275, + 0.04474053159356117, + 0.010819276794791222, + 0.038302090018987656, + 0.017782850190997124, + 0.00784611888229847, + 0.011574968695640564, + 0.013510036282241344, + 0.013328203931450844, + 0.03219420090317726, + 0.010938805527985096, + 0.019765807315707207, + 0.008102089166641235, + 0.01045786589384079, + 0.04132119566202164, + 0.006513896398246288, + 0.01180938258767128, + 0.005790547002106905, + 0.002280402695760131, + 0.00408913753926754, + 0.0018278503557667136, + 0.0027495590038597584, + 0.007551113143563271, + 0.002565407892689109, + 0.005774110555648804, + 0.005534762050956488, + 0.005952723789960146, + 0.00766976922750473, + 0.008363800123333931, + 0.008607146330177784, + 0.006441392935812473, + 0.01300655584782362, + 0.006723916158080101, + 0.004606957547366619, + 0.003577158087864518, + 0.0025429765228182077, + 0.003944418393075466, + 0.0030685155652463436, + 0.003126881318166852, + 0.004465598147362471, + 0.004232506267726421, + 0.023041054606437683, + 0.026291975751519203, + 0.015201820991933346, + 0.02008138783276081, + 0.03691183403134346, + 0.020117996260523796, + 0.02308790199458599, + 0.02213970012962818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0009197533945553005 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro shot, captured as a high-speed photographic sequence, remains static to record the explosive dispersal of a large seed head. Initially, the frame is filled with the perfectly intricately spherical seed head, its hundreds of delicate, white, feathery papillae interwoven into a fragile sphere, each connected to a dark achene core. The scene is backlit by brilliant afternoon sunlight, causing the silky filaments to glow with a soft, ethereal halo against a very blurred background of vibrant green grass. Suddenly, the sequence captures the violent explosion: the sphere instantly disintegrates, propelling the seeds outward in a chaotic, 360-degree explosion. Each unique, high-resolution photograph in the series freezes a sequential moment of the event—from the initial rupture to the cloud of seeds expanding to fill the frame, their tiny parachutes unfolding and catching the light. The movement is captured so clearly that the spiraling trails of individual seeds are visible, creating a powerful, dramatic, and scientifically accurate visualization of a fleeting natural phenomenon.", + "seed": 7632, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000124/group_9/index_1032_seed_7632_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015328959631733596, + "vlm_logprob": -0.015328959631733596, + "vlm_yes_no_margin": 5.625, + "vlm_token_credit": -0.015328959631733596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0069961813278496265 + ], + "opposite_logprob_tokens": [ + -6.881996154785156 + ], + "logprob_reward": -0.0069961813278496265, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0069961813278496265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07684928923845291, + 0.059669166803359985, + 0.04591601714491844, + 0.027886871248483658, + 0.267031729221344, + 0.16433005034923553, + 0.16465167701244354, + 0.05132681131362915, + 0.10509761422872543, + 0.03724078834056854 + ], + "credit_map": [ + 0.008560724556446075, + 0.006612684577703476, + 0.011538708582520485, + 0.008681909181177616, + 0.006913107819855213, + 0.015288583002984524, + 0.0066051422618329525, + 0.012648427858948708, + 0.005424770060926676, + 0.009348507970571518, + 0.012110435403883457, + 0.006432170979678631, + 0.007022854872047901, + 0.006871051155030727, + 0.006746291648596525, + 0.0057130781933665276, + 0.005470988806337118, + 0.00358667247928679, + 0.007376363500952721, + 0.003236933844164014, + 0.0035139978863298893, + 0.006330132484436035, + 0.009562348946928978, + 0.006838573608547449, + 0.004138493444770575, + 0.0038546212017536163, + 0.004493747837841511, + 0.002580685308203101, + 0.0026761640328913927, + 0.0025192489847540855, + 0.004523093346506357, + 0.0031008149962872267, + 0.03369008004665375, + 0.024910595268011093, + 0.06608358025550842, + 0.025176117196679115, + 0.03590916842222214, + 0.029487425461411476, + 0.02366412989795208, + 0.02811061218380928, + 0.009597753174602985, + 0.04817032068967819, + 0.009576912969350815, + 0.009416399523615837, + 0.02961702086031437, + 0.017315799370408058, + 0.011475005187094212, + 0.02916082926094532, + 0.011099347844719887, + 0.058389775454998016, + 0.005692232865840197, + 0.008495602756738663, + 0.04828524962067604, + 0.0102580226957798, + 0.015557567588984966, + 0.006873863283544779, + 0.005999299697577953, + 0.005735575687140226, + 0.005902040284126997, + 0.005467833019793034, + 0.009010819718241692, + 0.006445657927542925, + 0.006032298319041729, + 0.00673328572884202, + 0.006873213220387697, + 0.03675740957260132, + 0.01446028333157301, + 0.008446848019957542, + 0.010281410068273544, + 0.011341340839862823, + 0.008002194575965405, + 0.00893490482121706, + 0.0035031097941100597, + 0.005584062542766333, + 0.0038395682349801064, + 0.004836871288716793, + 0.005582194775342941, + 0.00365096190944314, + 0.004809715319424868, + 0.0054343040101230145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.0069961813278496265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036231130361557007 + ], + "opposite_logprob_tokens": [ + -5.628623008728027 + ], + "logprob_reward": -0.0036231130361557007, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036231130361557007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07847161591053009, + 0.05010383203625679, + 0.06817367672920227, + 0.06447747349739075, + 0.2622220814228058, + 0.12399303913116455, + 0.11329711973667145, + 0.08928831666707993, + 0.0896901786327362, + 0.060282666236162186 + ], + "credit_map": [ + 0.00875299982726574, + 0.01002942118793726, + 0.014420654624700546, + 0.009778848849236965, + 0.011730600148439407, + 0.007627568207681179, + 0.008895494043827057, + 0.007236018776893616, + 0.003962878603488207, + 0.0074049630202353, + 0.00801360234618187, + 0.0065183211117982864, + 0.006369197275489569, + 0.005260400008410215, + 0.007496500853449106, + 0.00507796835154295, + 0.011280601844191551, + 0.008314971812069416, + 0.008014166727662086, + 0.008986207656562328, + 0.007249376270920038, + 0.007756239268928766, + 0.008432376198470592, + 0.008139739744365215, + 0.0069180564023554325, + 0.010883639566600323, + 0.008574250154197216, + 0.007256161421537399, + 0.006822542287409306, + 0.006077780853956938, + 0.010466442443430424, + 0.007478605955839157, + 0.030846090987324715, + 0.026248004287481308, + 0.062014345079660416, + 0.027506746351718903, + 0.03586260601878166, + 0.028213554993271828, + 0.025166399776935577, + 0.026364360004663467, + 0.007033222354948521, + 0.03564201667904854, + 0.012521978467702866, + 0.006821147631853819, + 0.024717630818486214, + 0.010203571990132332, + 0.009733295068144798, + 0.017320185899734497, + 0.008673788979649544, + 0.02819669246673584, + 0.007756824139505625, + 0.010758521035313606, + 0.028970425948500633, + 0.008559523150324821, + 0.014237424358725548, + 0.0061439163982868195, + 0.013991856016218662, + 0.007696560118347406, + 0.00915008969604969, + 0.009961291216313839, + 0.013745183125138283, + 0.011381765827536583, + 0.012027974240481853, + 0.011333596892654896, + 0.006468560546636581, + 0.02738259732723236, + 0.01196887157857418, + 0.0074640922248363495, + 0.010135084390640259, + 0.012132649309933186, + 0.00636394415050745, + 0.007774375844746828, + 0.006455677095800638, + 0.007082816679030657, + 0.008224881254136562, + 0.009227617643773556, + 0.00647864630445838, + 0.005050178617238998, + 0.008051110431551933, + 0.00971173495054245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.0036231130361557007 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031945884693413973 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031945884693413973, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031945884693413973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17521150410175323, + 0.09328143298625946, + 0.03244926780462265, + 0.0499841682612896, + 0.08871512115001678, + 0.14928901195526123, + 0.059369202703237534, + 0.09659169614315033, + 0.07033640146255493, + 0.18477219343185425 + ], + "credit_map": [ + 0.017161836847662926, + 0.014826653525233269, + 0.021716101095080376, + 0.01957160048186779, + 0.016680922359228134, + 0.036056213080883026, + 0.014276047237217426, + 0.034922122955322266, + 0.010587488301098347, + 0.011386101134121418, + 0.017728766426444054, + 0.010415940545499325, + 0.010446458123624325, + 0.010574896819889545, + 0.01034656260162592, + 0.0117952236905694, + 0.004710606299340725, + 0.0027927502524107695, + 0.0035627507604658604, + 0.0028607684653252363, + 0.004250733647495508, + 0.003397565335035324, + 0.005010680761188269, + 0.005863409023731947, + 0.005321453791111708, + 0.006954014301300049, + 0.005773148965090513, + 0.005111231002956629, + 0.005415829364210367, + 0.005963140167295933, + 0.008396041579544544, + 0.00704931141808629, + 0.010306389071047306, + 0.01006993092596531, + 0.020352249965071678, + 0.008716467767953873, + 0.00905988272279501, + 0.010664124973118305, + 0.0077813309617340565, + 0.01176474243402481, + 0.01022612676024437, + 0.01577598601579666, + 0.011768229305744171, + 0.013397877104580402, + 0.025885330513119698, + 0.025372657924890518, + 0.02366945520043373, + 0.023193351924419403, + 0.0062194508500397205, + 0.007904148660600185, + 0.006443955935537815, + 0.00721878744661808, + 0.00978853926062584, + 0.005847922060638666, + 0.010096235200762749, + 0.005850163288414478, + 0.010066593997180462, + 0.010167625732719898, + 0.009294746443629265, + 0.011282423511147499, + 0.01539023220539093, + 0.00926546473056078, + 0.012117704376578331, + 0.01900690607726574, + 0.004923461005091667, + 0.007597324904054403, + 0.017780784517526627, + 0.006370996590703726, + 0.008430812507867813, + 0.011781635694205761, + 0.008026083931326866, + 0.005425300449132919, + 0.016868164762854576, + 0.03283531218767166, + 0.01671765372157097, + 0.01642158254981041, + 0.025121554732322693, + 0.02051416225731373, + 0.024015242233872414, + 0.03227853402495384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0031945884693413973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201859563589096 + ], + "opposite_logprob_tokens": [ + -2.812018632888794 + ], + "logprob_reward": -0.06201859563589096, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201859563589096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044156111776828766, + 0.06087295711040497, + 0.019250134006142616, + 0.03314225375652313, + 0.07339043915271759, + 0.05397672951221466, + 0.4352240562438965, + 0.13862501084804535, + 0.10396449267864227, + 0.0373978428542614 + ], + "credit_map": [ + 0.004946877248585224, + 0.005159937310963869, + 0.008466379716992378, + 0.005614167079329491, + 0.004469170700758696, + 0.006525741424411535, + 0.004785574972629547, + 0.004188258666545153, + 0.004771085921674967, + 0.01233038492500782, + 0.010684221982955933, + 0.0063623953610658646, + 0.007235290016978979, + 0.007079802453517914, + 0.006059439852833748, + 0.006350331474095583, + 0.00235649268142879, + 0.0018161989282816648, + 0.0024498947896063328, + 0.0021024986635893583, + 0.0024935887195169926, + 0.002103739185258746, + 0.0024150421377271414, + 0.0035126779694110155, + 0.0028060434851795435, + 0.006930061616003513, + 0.003481072373688221, + 0.003937880974262953, + 0.0030836963560432196, + 0.0028723974246531725, + 0.0056045628152787685, + 0.004426533356308937, + 0.009714002721011639, + 0.008459948934614658, + 0.018775349482893944, + 0.006826834287494421, + 0.008003865368664265, + 0.007673720829188824, + 0.00714730704203248, + 0.006789408624172211, + 0.0043445550836622715, + 0.010019998997449875, + 0.004684246610850096, + 0.004914591554552317, + 0.01317609753459692, + 0.006336046848446131, + 0.006574901286512613, + 0.003926279488950968, + 0.02745915949344635, + 0.1405586302280426, + 0.010942536406219006, + 0.016318803653120995, + 0.10635179281234741, + 0.04198306053876877, + 0.040577854961156845, + 0.051032163202762604, + 0.019915500655770302, + 0.013077173382043839, + 0.014415286481380463, + 0.016470443457365036, + 0.019679779186844826, + 0.01383153721690178, + 0.01575140655040741, + 0.0254838764667511, + 0.009028283879160881, + 0.02745000831782818, + 0.007151245139539242, + 0.007352186366915703, + 0.011879964731633663, + 0.02121220901608467, + 0.007822679355740547, + 0.012067915871739388, + 0.004709488712251186, + 0.005980360321700573, + 0.004361839033663273, + 0.005857840180397034, + 0.003766651963815093, + 0.003508463501930237, + 0.0050522037781775, + 0.004160990938544273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.06201859563589096 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a large seed head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial seed head spherical with white feathery papillae and dark achene cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene backlit by sunlight causing the filaments to glow with a soft halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video capture the seed head exploding and dispersing seeds in a chaotic 360-degree motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a very blurred area of vibrant green grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008123196894302964 + ], + "opposite_logprob_tokens": [ + -7.125812530517578 + ], + "logprob_reward": -0.0008123196894302964, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008123196894302964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671479195356369, + 0.03676402568817139, + 0.045169647783041, + 0.03823000565171242, + 0.10414201766252518, + 0.05588448420166969, + 0.21328718960285187, + 0.07129020988941193, + 0.17538155615329742, + 0.0927029401063919 + ], + "credit_map": [ + 0.020401721820235252, + 0.012575292028486729, + 0.030934270471334457, + 0.017443662509322166, + 0.0169331394135952, + 0.02811184525489807, + 0.01738237403333187, + 0.023365620523691177, + 0.004052917938679457, + 0.003347208257764578, + 0.00835356954485178, + 0.0036383490078151226, + 0.0026905338745564222, + 0.006443839054554701, + 0.0034574742894619703, + 0.004780133254826069, + 0.0050771781243383884, + 0.005703635513782501, + 0.006960389204323292, + 0.00433478644117713, + 0.003553971415385604, + 0.0066058519296348095, + 0.006131965201348066, + 0.006801866460591555, + 0.0037417514249682426, + 0.007703018374741077, + 0.004303427878767252, + 0.003970376681536436, + 0.004194091539829969, + 0.003669418627396226, + 0.006126172840595245, + 0.004521746188402176, + 0.012766167521476746, + 0.011598045006394386, + 0.023920521140098572, + 0.00805287342518568, + 0.012261941097676754, + 0.011968489736318588, + 0.00942436046898365, + 0.014149616472423077, + 0.0036963587626814842, + 0.010687611065804958, + 0.006296925246715546, + 0.00352539517916739, + 0.004912829026579857, + 0.008034306578338146, + 0.006230674218386412, + 0.012500383891165257, + 0.013501632958650589, + 0.07131381332874298, + 0.005855009891092777, + 0.009475789964199066, + 0.05513504892587662, + 0.017806606367230415, + 0.01984158344566822, + 0.02035771682858467, + 0.005108016077429056, + 0.009622922167181969, + 0.0062830522656440735, + 0.00789053924381733, + 0.016269955784082413, + 0.006453365553170443, + 0.006905285641551018, + 0.01275707595050335, + 0.00798182561993599, + 0.06502806395292282, + 0.04769665002822876, + 0.00834873877465725, + 0.011056226678192616, + 0.0063815005123615265, + 0.010131069459021091, + 0.018757479265332222, + 0.011167922057211399, + 0.017134234309196472, + 0.008973506279289722, + 0.011601978912949562, + 0.012713512405753136, + 0.007818018086254597, + 0.008827670477330685, + 0.014466097578406334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.0008123196894302964 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_9/index_tensor([8647, 8647])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_9/index_tensor([8647, 8647])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b75d996c12479c0e523c4797db141fccca2eb1bc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_9/index_tensor([8647, 8647])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot, close to the ground, follows a flame-red sports car as it tears through the urban night at breakneck speed. The car boasts a streamlined body and a low chassis, its wet paint gleaming and shimmering under the passing neon lights. It hurtles across the mirror-like black asphalt, freshly soaked by rain, its wheels spinning at high speed, sending up a shower of fine spray behind it. The car's bright LED taillights cast two elongated, trembling streaks of red light on the slippery road, interweaving with the cold blue and purple hues cast by the surrounding skyscrapers, creating a stark contrast. The entire scene is imbued with a sense of speed and urgency, creating a tense atmosphere of desperate escape through the cold yet glamorous urban jungle.", + "seed": 17420, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_9/index_8647_seed_17420_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04998158058151603, + "vlm_logprob": -0.04998158058151603, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.04998158058151603, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09699927270412445 + ], + "opposite_logprob_tokens": [ + -2.971999168395996 + ], + "logprob_reward": -0.09699927270412445, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.09699927270412445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18520091474056244, + 0.09012430906295776, + 0.124997578561306, + 0.11637451499700546, + 0.0442078672349453, + 0.15369543433189392, + 0.06860927492380142, + 0.07395938783884048, + 0.08006008714437485, + 0.06277059763669968 + ], + "credit_map": [ + 0.02256336249411106, + 0.018771184608340263, + 0.01644202508032322, + 0.04122213274240494, + 0.02017490193247795, + 0.01177703682333231, + 0.02542554959654808, + 0.02882472425699234, + 0.009782458655536175, + 0.011655092239379883, + 0.016274819150567055, + 0.0158088356256485, + 0.0049625602550804615, + 0.006697310600429773, + 0.012499898672103882, + 0.012443337589502335, + 0.012647150084376335, + 0.016427429392933846, + 0.009216503240168095, + 0.01944454200565815, + 0.014996585436165333, + 0.011288229376077652, + 0.027414795011281967, + 0.013562350533902645, + 0.010249807499349117, + 0.006894741207361221, + 0.01603398099541664, + 0.017935335636138916, + 0.005871872883290052, + 0.025016194209456444, + 0.011448295786976814, + 0.022924285382032394, + 0.004497861489653587, + 0.0032145401928573847, + 0.007967772893607616, + 0.004306210670620203, + 0.006722958292812109, + 0.007441062014549971, + 0.003605441190302372, + 0.00645202212035656, + 0.014715264551341534, + 0.008181936107575893, + 0.016906287521123886, + 0.013542245142161846, + 0.01689213514328003, + 0.029680388048291206, + 0.019910115748643875, + 0.03386705741286278, + 0.009879762306809425, + 0.004893598612397909, + 0.008825400844216347, + 0.008954931981861591, + 0.006260171066969633, + 0.009892717935144901, + 0.008812488988041878, + 0.011090205982327461, + 0.010516408830881119, + 0.005224455147981644, + 0.011138111352920532, + 0.008303375914692879, + 0.004806510638445616, + 0.01194776315242052, + 0.0056145950220525265, + 0.016408171504735947, + 0.012261574156582355, + 0.010602464899420738, + 0.015186911448836327, + 0.00677482457831502, + 0.005753712262958288, + 0.008418922312557697, + 0.010013251565396786, + 0.011048431508243084, + 0.00767412269487977, + 0.008976076729595661, + 0.009285731241106987, + 0.0071644713170826435, + 0.006812587846070528, + 0.009248082526028156, + 0.007601413410156965, + 0.006008106283843517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.09699927270412445 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308886453509331 + ], + "opposite_logprob_tokens": [ + -3.168088912963867 + ], + "logprob_reward": -0.04308886453509331, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308886453509331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19372643530368805, + 0.18467672169208527, + 0.18372617661952972, + 0.08330102264881134, + 0.036336224526166916, + 0.0822518840432167, + 0.07740577310323715, + 0.05427905544638634, + 0.055545467883348465, + 0.04875120893120766 + ], + "credit_map": [ + 0.014049030840396881, + 0.017215710133314133, + 0.018147027119994164, + 0.03276408836245537, + 0.024976519867777824, + 0.02756621316075325, + 0.03374581038951874, + 0.02526205964386463, + 0.013546139933168888, + 0.013120763003826141, + 0.020959023386240005, + 0.027164051309227943, + 0.009368915110826492, + 0.02115103229880333, + 0.022466639056801796, + 0.05690016970038414, + 0.01916760765016079, + 0.027373524382710457, + 0.01953047886490822, + 0.025160178542137146, + 0.037548936903476715, + 0.01778474636375904, + 0.021500688046216965, + 0.01566002331674099, + 0.008362552151083946, + 0.006292298436164856, + 0.009269782342016697, + 0.01177194807678461, + 0.0050636716187000275, + 0.01863558031618595, + 0.0069465674459934235, + 0.016958627849817276, + 0.003686376381665468, + 0.0030282922089099884, + 0.003685476491227746, + 0.00493807764723897, + 0.006755463313311338, + 0.005191906355321407, + 0.0037774359807372093, + 0.00527319498360157, + 0.009669231250882149, + 0.006649022921919823, + 0.014182542450726032, + 0.009538811631500721, + 0.006511992309242487, + 0.008998798206448555, + 0.012215515598654747, + 0.014485982246696949, + 0.009784504771232605, + 0.004770986270159483, + 0.008308875374495983, + 0.008134748786687851, + 0.006753615103662014, + 0.011913175694644451, + 0.010772342793643475, + 0.016967523843050003, + 0.009069079533219337, + 0.004890828859061003, + 0.008511370979249477, + 0.004858299158513546, + 0.003236503340303898, + 0.009723549708724022, + 0.003566971281543374, + 0.010422457940876484, + 0.007485288195312023, + 0.009014459326863289, + 0.008606916293501854, + 0.004072846844792366, + 0.0056518679484725, + 0.00606845086440444, + 0.006038876716047525, + 0.008606767281889915, + 0.006667817011475563, + 0.005770546849817038, + 0.0064296964555978775, + 0.0032036793418228626, + 0.004545450676232576, + 0.009986191987991333, + 0.004838662687689066, + 0.007309165317565203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04308886453509331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820537194609642 + ], + "opposite_logprob_tokens": [ + -3.529820442199707 + ], + "logprob_reward": -0.029820537194609642, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820537194609642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14676068723201752, + 0.1481623649597168, + 0.22875678539276123, + 0.09698538482189178, + 0.06495429575443268, + 0.06502311676740646, + 0.06268637627363205, + 0.0515645295381546, + 0.0595954954624176, + 0.07551094144582748 + ], + "credit_map": [ + 0.011646552942693233, + 0.014599221758544445, + 0.01509152539074421, + 0.02333175204694271, + 0.028538430109620094, + 0.014164745807647705, + 0.016288314014673233, + 0.023100141435861588, + 0.018933171406388283, + 0.014073905535042286, + 0.03293662145733833, + 0.023203857243061066, + 0.01026125904172659, + 0.011743016541004181, + 0.020003626123070717, + 0.017006900161504745, + 0.02339780330657959, + 0.03474666550755501, + 0.01719730906188488, + 0.032822366803884506, + 0.04173213243484497, + 0.023341866210103035, + 0.03025451861321926, + 0.025264110416173935, + 0.009936482645571232, + 0.006156006827950478, + 0.015745092183351517, + 0.021827690303325653, + 0.005842526908963919, + 0.01163991168141365, + 0.013157875277101994, + 0.01267980132251978, + 0.006882853340357542, + 0.006248914171010256, + 0.009004401043057442, + 0.005904131103307009, + 0.014928537420928478, + 0.006061747204512358, + 0.005355057306587696, + 0.010568658821284771, + 0.0055991574190557, + 0.004774557892233133, + 0.008167240768671036, + 0.006892748177051544, + 0.008562546223402023, + 0.009821794927120209, + 0.0104274433106184, + 0.010777625255286694, + 0.0075778537429869175, + 0.004844485316425562, + 0.006997960619628429, + 0.010970432311296463, + 0.004939311183989048, + 0.00904364138841629, + 0.007945706136524677, + 0.010366987437009811, + 0.011735851876437664, + 0.005009257700294256, + 0.008035428822040558, + 0.007133187260478735, + 0.0025706433225423098, + 0.0063391379080712795, + 0.0029823221266269684, + 0.0077587030827999115, + 0.00769740529358387, + 0.006748164538294077, + 0.010004570707678795, + 0.006503645330667496, + 0.004803975112736225, + 0.006078179460018873, + 0.009263393469154835, + 0.008496161550283432, + 0.010583159513771534, + 0.008426613174378872, + 0.00969718862324953, + 0.00668142968788743, + 0.00651751970872283, + 0.011415038257837296, + 0.009053267538547516, + 0.013136732392013073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.029820537194609642 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010450384579598904 + ], + "opposite_logprob_tokens": [ + -6.876045227050781 + ], + "logprob_reward": -0.0010450384579598904, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010450384579598904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10916341841220856, + 0.048574917018413544, + 0.09562511742115021, + 0.06498933583498001, + 0.0320548340678215, + 0.035170551389455795, + 0.0833032876253128, + 0.10055956989526749, + 0.3108815848827362, + 0.11967740952968597 + ], + "credit_map": [ + 0.012824117206037045, + 0.009963715448975563, + 0.014283538796007633, + 0.022931547835469246, + 0.01185936201363802, + 0.00740067521110177, + 0.012673872523009777, + 0.01722658798098564, + 0.00610211631283164, + 0.005101046059280634, + 0.009249072521924973, + 0.008228585124015808, + 0.003062584437429905, + 0.004071917850524187, + 0.005431714002043009, + 0.007327883504331112, + 0.010262390598654747, + 0.010395241901278496, + 0.008051189593970776, + 0.014969482086598873, + 0.00915067084133625, + 0.009842266328632832, + 0.01927865482866764, + 0.013675226829946041, + 0.008203218691051006, + 0.009163334034383297, + 0.007084054872393608, + 0.010449294000864029, + 0.006355856545269489, + 0.00868549570441246, + 0.007914441637694836, + 0.007133640814572573, + 0.004884424153715372, + 0.0036168668884783983, + 0.0038701938465237617, + 0.0035792151466012, + 0.004392048344016075, + 0.003338002832606435, + 0.003981500864028931, + 0.004392581060528755, + 0.0033875321969389915, + 0.00250886264257133, + 0.0061534373089671135, + 0.003029340412467718, + 0.0029512441251426935, + 0.005551734007894993, + 0.00660462724044919, + 0.004983772523701191, + 0.00962867122143507, + 0.0068729342892766, + 0.008503035642206669, + 0.012952650897204876, + 0.007530310191214085, + 0.01195301953703165, + 0.010059152729809284, + 0.015803512185811996, + 0.015375332906842232, + 0.007683297153562307, + 0.014803031459450722, + 0.008110666647553444, + 0.012297769077122211, + 0.017818482592701912, + 0.007182032335549593, + 0.01728896051645279, + 0.04087696596980095, + 0.042501747608184814, + 0.053285371512174606, + 0.03758490830659866, + 0.021793941035866737, + 0.022760994732379913, + 0.0567614920437336, + 0.035316161811351776, + 0.018467072397470474, + 0.017046837136149406, + 0.018241574987769127, + 0.013919806107878685, + 0.010242496617138386, + 0.014659848064184189, + 0.014677765779197216, + 0.012422013096511364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0010450384579598904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895419001579285 + ], + "opposite_logprob_tokens": [ + -2.5789542198181152 + ], + "logprob_reward": -0.07895419001579285, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895419001579285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20479150116443634, + 0.3246519863605499, + 0.02911362797021866, + 0.058193057775497437, + 0.1292724609375, + 0.03700133040547371, + 0.030182547867298126, + 0.043350543826818466, + 0.06119505316019058, + 0.08224785327911377 + ], + "credit_map": [ + 0.032615650445222855, + 0.01867716759443283, + 0.02316875196993351, + 0.03834468126296997, + 0.017735512927174568, + 0.01536396611481905, + 0.026713043451309204, + 0.03217273950576782, + 0.027349349111318588, + 0.016900131478905678, + 0.03020976111292839, + 0.06664851307868958, + 0.012593894265592098, + 0.020036345347762108, + 0.038064874708652496, + 0.112849161028862, + 0.004341684281826019, + 0.003375376109033823, + 0.0039005735889077187, + 0.0035170787014067173, + 0.0029192366637289524, + 0.0029855193570256233, + 0.004518152214586735, + 0.0035560105461627245, + 0.00778103806078434, + 0.003680912312120199, + 0.006108296103775501, + 0.005638491827994585, + 0.0034743808209896088, + 0.011815238744020462, + 0.008154922164976597, + 0.01153978705406189, + 0.012336836196482182, + 0.009310866706073284, + 0.018650837242603302, + 0.010622691363096237, + 0.015536725521087646, + 0.023746224120259285, + 0.013806025497615337, + 0.0252622589468956, + 0.003682149341329932, + 0.002185480436310172, + 0.00441896915435791, + 0.0037950549740344286, + 0.0030175603460520506, + 0.007203971967101097, + 0.004950485657900572, + 0.00774766318500042, + 0.0037071704864501953, + 0.0026285857893526554, + 0.0034027297515422106, + 0.004041360225528479, + 0.0031550119165331125, + 0.004127641674131155, + 0.004251935984939337, + 0.004868113901466131, + 0.006477372720837593, + 0.0031524738296866417, + 0.007085631601512432, + 0.0029972533229738474, + 0.0034526525996625423, + 0.008120203390717506, + 0.0021132135298103094, + 0.009951749816536903, + 0.00814284011721611, + 0.00870425347238779, + 0.0111312847584486, + 0.007113803178071976, + 0.006011325400322676, + 0.003710090648382902, + 0.01058190781623125, + 0.005799554288387299, + 0.013236317783594131, + 0.007412075996398926, + 0.011606310494244099, + 0.006799729075282812, + 0.006348213646560907, + 0.01254327967762947, + 0.008421636186540127, + 0.015880299732089043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.07895419001579285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, close to the ground, follows a flame-red sports car as it tears through the urban night at breakneck speed. The car boasts a streamlined body and a low chassis, its wet paint gleaming and shimmering under the passing neon lights. It hurtles across the mirror-like black asphalt, freshly soaked by rain, its wheels spinning at high speed, sending up a shower of fine spray behind it. The car's bright LED taillights cast two elongated, trembling streaks of red light on the slippery road, interweaving with the cold blue and purple hues cast by the surrounding skyscrapers, creating a stark contrast. The entire scene is imbued with a sense of speed and urgency, creating a tense atmosphere of desperate escape through the cold yet glamorous urban jungle.", + "seed": 17421, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_9/index_8647_seed_17421_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0482770788599737, + "vlm_logprob": -0.0482770788599737, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.0482770788599737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04962071776390076 + ], + "opposite_logprob_tokens": [ + -4.174620628356934 + ], + "logprob_reward": -0.04962071776390076, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.04962071776390076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0975603312253952, + 0.06320900470018387, + 0.18426170945167542, + 0.05594109371304512, + 0.04288080334663391, + 0.1344478726387024, + 0.06223927065730095, + 0.14901921153068542, + 0.13138678669929504, + 0.07905395328998566 + ], + "credit_map": [ + 0.010895361192524433, + 0.011271719820797443, + 0.01794450171291828, + 0.014522942714393139, + 0.007490027695894241, + 0.012889632023870945, + 0.010413981042802334, + 0.012132154777646065, + 0.007505742833018303, + 0.006973601412028074, + 0.011908113956451416, + 0.011640147306025028, + 0.005665421951562166, + 0.006716111674904823, + 0.004226970952004194, + 0.008572886697947979, + 0.02095746248960495, + 0.026917925104498863, + 0.022097008302807808, + 0.01969912275671959, + 0.01630621775984764, + 0.02490842342376709, + 0.015198691748082638, + 0.038176845759153366, + 0.006863981485366821, + 0.007893509231507778, + 0.008267895318567753, + 0.006651679519563913, + 0.0065123774111270905, + 0.0067958831787109375, + 0.00571035360917449, + 0.007245407439768314, + 0.004333025310188532, + 0.00503581203520298, + 0.006010625511407852, + 0.0048502120189368725, + 0.0031087726820260286, + 0.006717628799378872, + 0.0063715423457324505, + 0.006453180685639381, + 0.016473209485411644, + 0.009269632399082184, + 0.020266270264983177, + 0.02361387573182583, + 0.017304634675383568, + 0.011733361519873142, + 0.019061975181102753, + 0.01672489009797573, + 0.007780180778354406, + 0.00465437863022089, + 0.0134201031178236, + 0.004749685525894165, + 0.007945035584270954, + 0.009428415447473526, + 0.004766189027577639, + 0.009495270438492298, + 0.013664436526596546, + 0.009009607136249542, + 0.022014308720827103, + 0.013266597874462605, + 0.019493568688631058, + 0.022805752232670784, + 0.02502005733549595, + 0.023744875565171242, + 0.011377663351595402, + 0.013906579464673996, + 0.017394280061125755, + 0.0073315780609846115, + 0.01114200334995985, + 0.029809575527906418, + 0.007809469476342201, + 0.03261563181877136, + 0.010599604807794094, + 0.009296631440520287, + 0.009758904576301575, + 0.007503305561840534, + 0.01447830256074667, + 0.010095510631799698, + 0.005933934822678566, + 0.011387748643755913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.04962071776390076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988358438014984 + ], + "opposite_logprob_tokens": [ + -1.804883599281311 + ], + "logprob_reward": -0.17988358438014984, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988358438014984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06936439126729965, + 0.05859653651714325, + 0.09388525038957596, + 0.0790826827287674, + 0.045188747346401215, + 0.09149907529354095, + 0.03967079892754555, + 0.1298425942659378, + 0.259510338306427, + 0.1333596557378769 + ], + "credit_map": [ + 0.008273330517113209, + 0.008436474949121475, + 0.008323621936142445, + 0.009191593155264854, + 0.007822811603546143, + 0.00883571058511734, + 0.00766021478921175, + 0.010820629075169563, + 0.0065622455440461636, + 0.00907994620501995, + 0.008870387449860573, + 0.010341175831854343, + 0.0061631458811461926, + 0.006143379490822554, + 0.004727931227535009, + 0.006708316970616579, + 0.01303864736109972, + 0.012497247196733952, + 0.010345528833568096, + 0.010846205987036228, + 0.00680791400372982, + 0.013438493944704533, + 0.007918535731732845, + 0.018992673605680466, + 0.007925081066787243, + 0.0073386067524552345, + 0.011458061635494232, + 0.010743007063865662, + 0.007492721546441317, + 0.011827795766294003, + 0.0100063132122159, + 0.012291090562939644, + 0.005579584278166294, + 0.005102836526930332, + 0.010500745847821236, + 0.004948252346366644, + 0.0026829158887267113, + 0.006136658601462841, + 0.0045180064626038074, + 0.005719744134694338, + 0.012439154088497162, + 0.010712621733546257, + 0.009177531115710735, + 0.015482516959309578, + 0.01305519137531519, + 0.010030889883637428, + 0.010435649193823338, + 0.010165508836507797, + 0.004309536423534155, + 0.003695244900882244, + 0.00683712400496006, + 0.004893688019365072, + 0.004163386300206184, + 0.005728131625801325, + 0.004949420690536499, + 0.005094267427921295, + 0.013887994922697544, + 0.019134990870952606, + 0.018222851678729057, + 0.015454884618520737, + 0.014268076047301292, + 0.01458247285336256, + 0.016232816502451897, + 0.018058501183986664, + 0.012033533304929733, + 0.013061434961855412, + 0.016061125323176384, + 0.008435114286839962, + 0.01962473802268505, + 0.08725651353597641, + 0.011136103421449661, + 0.09190177172422409, + 0.012300711125135422, + 0.022392699494957924, + 0.023241471499204636, + 0.014083057641983032, + 0.01644299365580082, + 0.014056083746254444, + 0.01359629351645708, + 0.017246341332793236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.17988358438014984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003623231779783964 + ], + "opposite_logprob_tokens": [ + -5.628623008728027 + ], + "logprob_reward": -0.003623231779783964, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003623231779783964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11743661016225815, + 0.1478501409292221, + 0.27205395698547363, + 0.08419094979763031, + 0.05229217931628227, + 0.09992469102144241, + 0.04023108258843422, + 0.07603441923856735, + 0.03137003630399704, + 0.07861586660146713 + ], + "credit_map": [ + 0.012448742985725403, + 0.021815946325659752, + 0.014436698518693447, + 0.014437416568398476, + 0.018480516970157623, + 0.010074068792164326, + 0.010484309867024422, + 0.015258921310305595, + 0.01467185653746128, + 0.022137489169836044, + 0.01802639849483967, + 0.023243630304932594, + 0.017378902062773705, + 0.013410268351435661, + 0.02456650882959366, + 0.014415095560252666, + 0.029593098908662796, + 0.035623930394649506, + 0.028720121830701828, + 0.029754232615232468, + 0.023518988862633705, + 0.0410444401204586, + 0.024135515093803406, + 0.05966363102197647, + 0.00907499436289072, + 0.007505243178457022, + 0.015379288233816624, + 0.010090744122862816, + 0.00598122738301754, + 0.013300489634275436, + 0.00958289299160242, + 0.01327606663107872, + 0.004548857919871807, + 0.007556600496172905, + 0.006741438526660204, + 0.0067169275134801865, + 0.00489621888846159, + 0.007845532149076462, + 0.007064784877002239, + 0.006921814754605293, + 0.013174855150282383, + 0.008352729491889477, + 0.013375655747950077, + 0.011324665509164333, + 0.01595616154372692, + 0.010546352714300156, + 0.014457852579653263, + 0.012736419215798378, + 0.005014028400182724, + 0.002690065884962678, + 0.007843075320124626, + 0.0035303891636431217, + 0.004725549835711718, + 0.007905448786914349, + 0.0028083526995033026, + 0.005714169703423977, + 0.008411799557507038, + 0.010628657415509224, + 0.007307093124836683, + 0.009465564042329788, + 0.007480928674340248, + 0.010691088624298573, + 0.010748395696282387, + 0.011300885118544102, + 0.0029487875290215015, + 0.003014791291207075, + 0.004752831533551216, + 0.0016189288580790162, + 0.004850616678595543, + 0.007111365906894207, + 0.0025980633217841387, + 0.004474653396755457, + 0.008259457536041737, + 0.012659106403589249, + 0.014018871821463108, + 0.008518158458173275, + 0.008634822443127632, + 0.007285820785909891, + 0.009320729412138462, + 0.009918902069330215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.003623231779783964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006348263123072684 + ], + "opposite_logprob_tokens": [ + -7.375634670257568 + ], + "logprob_reward": -0.0006348263123072684, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006348263123072684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11409484595060349, + 0.02757367305457592, + 0.24102145433425903, + 0.06589951366186142, + 0.03624145686626434, + 0.16727803647518158, + 0.0773061141371727, + 0.06472282111644745, + 0.15973736345767975, + 0.04612476006150246 + ], + "credit_map": [ + 0.010239997878670692, + 0.011970741674304008, + 0.02135413885116577, + 0.016640935093164444, + 0.008584557101130486, + 0.013274246826767921, + 0.013976198621094227, + 0.018054025247693062, + 0.002616982441395521, + 0.00392600754275918, + 0.004097653552889824, + 0.004638671875, + 0.003689167555421591, + 0.002664361149072647, + 0.0028510664124041796, + 0.003089763456955552, + 0.027378030121326447, + 0.03494163230061531, + 0.027222272008657455, + 0.03429792821407318, + 0.022038724273443222, + 0.032172005623579025, + 0.02768339216709137, + 0.03528744354844093, + 0.007144485600292683, + 0.009802635759115219, + 0.007438828237354755, + 0.007171811070293188, + 0.008233319036662579, + 0.010426634922623634, + 0.005383838899433613, + 0.010297958739101887, + 0.0042324247770011425, + 0.0031329826451838017, + 0.006198661867529154, + 0.0054416111670434475, + 0.0023792183492332697, + 0.00489072548225522, + 0.0048957704566419125, + 0.005070061422884464, + 0.021032365038990974, + 0.009713537059724331, + 0.026963163167238235, + 0.0190916508436203, + 0.02660888433456421, + 0.01846347376704216, + 0.022304005920886993, + 0.0231009591370821, + 0.008731558918952942, + 0.005279822740703821, + 0.010663279332220554, + 0.006702343467622995, + 0.012770403176546097, + 0.014359788037836552, + 0.007193868514150381, + 0.011605052277445793, + 0.006352999713271856, + 0.007390407379716635, + 0.008070536889135838, + 0.007769709452986717, + 0.00866105780005455, + 0.009145633317530155, + 0.007859990932047367, + 0.009472480975091457, + 0.02288067154586315, + 0.021251577883958817, + 0.027193792164325714, + 0.011097332462668419, + 0.022959571331739426, + 0.02435901202261448, + 0.013735284097492695, + 0.01626012660562992, + 0.004750146064907312, + 0.006193738896399736, + 0.0073615312576293945, + 0.003916907124221325, + 0.005946196150034666, + 0.006397143471986055, + 0.00625567976385355, + 0.005303418263792992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.0006348263123072684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623034063726664 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623034063726664, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623034063726664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042821004986763, + 0.0294000506401062, + 0.044814832508563995, + 0.08326450735330582, + 0.23482826352119446, + 0.08916629105806351, + 0.02919306606054306, + 0.040131013840436935, + 0.2691922187805176, + 0.13718882203102112 + ], + "credit_map": [ + 0.003057679161429405, + 0.008500311523675919, + 0.004830940626561642, + 0.007050239481031895, + 0.0050581274554133415, + 0.005683807656168938, + 0.00392503384500742, + 0.004714865703135729, + 0.002885873196646571, + 0.004168336279690266, + 0.0036920609418302774, + 0.004179764073342085, + 0.0042174262925982475, + 0.0034781768918037415, + 0.003722720080986619, + 0.0030556945130228996, + 0.006784919183701277, + 0.010902932845056057, + 0.004489494953304529, + 0.005001605022698641, + 0.0037579238414764404, + 0.004902927670627832, + 0.00307268719188869, + 0.00590234249830246, + 0.008674285374581814, + 0.011515079066157341, + 0.013828440569341183, + 0.010584157891571522, + 0.009097968228161335, + 0.011277289129793644, + 0.006519512739032507, + 0.011767774820327759, + 0.027757728472352028, + 0.03155623748898506, + 0.03571855276823044, + 0.025305522605776787, + 0.020816495642066002, + 0.03416913375258446, + 0.028166186064481735, + 0.0313383974134922, + 0.01141919195652008, + 0.009137652814388275, + 0.012865369208157063, + 0.00928738433867693, + 0.01125419419258833, + 0.01214208360761404, + 0.011107049882411957, + 0.011953365057706833, + 0.004120688419789076, + 0.0028674902860075235, + 0.004091466311365366, + 0.0028910976834595203, + 0.005922283045947552, + 0.003842079546302557, + 0.002307833405211568, + 0.003150124801322818, + 0.003361839335411787, + 0.00365965673699975, + 0.005291941575706005, + 0.003109708661213517, + 0.005842109210789204, + 0.008356194943189621, + 0.002817051950842142, + 0.007692511659115553, + 0.02140817791223526, + 0.014814410358667374, + 0.015443922951817513, + 0.007062933873385191, + 0.027593687176704407, + 0.08416048437356949, + 0.01712023839354515, + 0.08158832788467407, + 0.014663920737802982, + 0.01448579877614975, + 0.014921678230166435, + 0.008795054629445076, + 0.024607442319393158, + 0.02428908459842205, + 0.014484785497188568, + 0.020941048860549927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.007623034063726664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, close to the ground, follows a flame-red sports car as it tears through the urban night at breakneck speed. The car boasts a streamlined body and a low chassis, its wet paint gleaming and shimmering under the passing neon lights. It hurtles across the mirror-like black asphalt, freshly soaked by rain, its wheels spinning at high speed, sending up a shower of fine spray behind it. The car's bright LED taillights cast two elongated, trembling streaks of red light on the slippery road, interweaving with the cold blue and purple hues cast by the surrounding skyscrapers, creating a stark contrast. The entire scene is imbued with a sense of speed and urgency, creating a tense atmosphere of desperate escape through the cold yet glamorous urban jungle.", + "seed": 17422, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_9/index_8647_seed_17422_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02906325818039477, + "vlm_logprob": -0.02906325818039477, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.02906325818039477, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03300319239497185 + ], + "opposite_logprob_tokens": [ + -4.658003330230713 + ], + "logprob_reward": -0.03300319239497185, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.03300319239497185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30339086055755615, + 0.11575546860694885, + 0.04010561853647232, + 0.058530692011117935, + 0.030099283903837204, + 0.07638876140117645, + 0.06836802512407303, + 0.04724800959229469, + 0.1670374572277069, + 0.09307584166526794 + ], + "credit_map": [ + 0.008891016244888306, + 0.007633572444319725, + 0.019438093528151512, + 0.012880727648735046, + 0.01232366356998682, + 0.010235405527055264, + 0.11174899339675903, + 0.12023939192295074, + 0.00732586532831192, + 0.007882737554609776, + 0.015108915977180004, + 0.01901397667825222, + 0.024387231096625328, + 0.013277966529130936, + 0.013093403540551662, + 0.015665367245674133, + 0.003749212482944131, + 0.003524211933836341, + 0.0070631252601742744, + 0.008782768622040749, + 0.003063590032979846, + 0.0025152815505862236, + 0.005740143358707428, + 0.00566728413105011, + 0.006136857904493809, + 0.007039086893200874, + 0.006276540458202362, + 0.010794282890856266, + 0.0036667552776634693, + 0.004035200923681259, + 0.01107193436473608, + 0.009510036557912827, + 0.0020488437730818987, + 0.0035950031597167253, + 0.004364634398370981, + 0.004622049629688263, + 0.0025253326166421175, + 0.004413654562085867, + 0.003656795946881175, + 0.004872969351708889, + 0.005933269392699003, + 0.005950635299086571, + 0.013608154840767384, + 0.008030787110328674, + 0.00603835005313158, + 0.010563580319285393, + 0.015150628052651882, + 0.011113361455500126, + 0.00923198089003563, + 0.007206299342215061, + 0.008955205790698528, + 0.00555813405662775, + 0.01027020625770092, + 0.008441824465990067, + 0.007267639972269535, + 0.011436737142503262, + 0.006014327052980661, + 0.0016581341624259949, + 0.006171963177621365, + 0.002490784740075469, + 0.004096765071153641, + 0.004446662496775389, + 0.011976598761975765, + 0.010392771102488041, + 0.022535549476742744, + 0.007748458068817854, + 0.050569090992212296, + 0.0075171818025410175, + 0.007912752218544483, + 0.011511602438986301, + 0.011067801155149937, + 0.04817502573132515, + 0.022013084962964058, + 0.008959239348769188, + 0.021804362535476685, + 0.006455256603658199, + 0.006076526828110218, + 0.00906695332378149, + 0.008487424813210964, + 0.010212995111942291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.03300319239497185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898525893688202 + ], + "opposite_logprob_tokens": [ + -2.5789852142333984 + ], + "logprob_reward": -0.07898525893688202, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898525893688202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.431679904460907, + 0.08217117190361023, + 0.03867996856570244, + 0.043694160878658295, + 0.12930452823638916, + 0.08689138293266296, + 0.04347283020615578, + 0.04982297867536545, + 0.05735669285058975, + 0.036926399916410446 + ], + "credit_map": [ + 0.012736808508634567, + 0.010646738111972809, + 0.016419796273112297, + 0.0167898777872324, + 0.026691405102610588, + 0.026045750826597214, + 0.1479109525680542, + 0.17443858087062836, + 0.006357560865581036, + 0.011410078965127468, + 0.012376423925161362, + 0.013878322206437588, + 0.008072410710155964, + 0.00803665816783905, + 0.009057738818228245, + 0.012981979176402092, + 0.003468084614723921, + 0.003888163948431611, + 0.005724456161260605, + 0.005878138821572065, + 0.00467923516407609, + 0.004228684119880199, + 0.005238086450845003, + 0.005575118120759726, + 0.0053827716037631035, + 0.004866580944508314, + 0.007366281934082508, + 0.00920364074409008, + 0.002927409950643778, + 0.003270501270890236, + 0.006921515800058842, + 0.0037554597947746515, + 0.008159453980624676, + 0.011554133147001266, + 0.011706502176821232, + 0.02041986770927906, + 0.009743918664753437, + 0.023564275354146957, + 0.017199479043483734, + 0.026956908404827118, + 0.005836800672113895, + 0.009953011758625507, + 0.010153968818485737, + 0.006600980646908283, + 0.015330512076616287, + 0.012775937095284462, + 0.009628139436244965, + 0.01661202684044838, + 0.005032161716371775, + 0.00545970955863595, + 0.006319898646324873, + 0.002967291045933962, + 0.006739032454788685, + 0.0061204866506159306, + 0.004358749836683273, + 0.006475504487752914, + 0.0047477902844548225, + 0.0032739632297307253, + 0.005017128307372332, + 0.002688963431864977, + 0.006040461827069521, + 0.005983930081129074, + 0.012867973186075687, + 0.009202769957482815, + 0.00708686513826251, + 0.0033833410125225782, + 0.009639524854719639, + 0.005407768301665783, + 0.005339358001947403, + 0.0060631283558905125, + 0.005482610315084457, + 0.01495409943163395, + 0.00931883230805397, + 0.0028035577852278948, + 0.007573659531772137, + 0.0032235225662589073, + 0.0023197506088763475, + 0.005078991409391165, + 0.003240272169932723, + 0.0033678151667118073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.07898525893688202 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028247239533811808 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028247239533811808, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028247239533811808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18543942272663116, + 0.10781960189342499, + 0.06735292077064514, + 0.09812798351049423, + 0.09091509133577347, + 0.11043605208396912, + 0.048754092305898666, + 0.045427095144987106, + 0.18480488657951355, + 0.06092284619808197 + ], + "credit_map": [ + 0.008503606542944908, + 0.01066140178591013, + 0.011304995976388454, + 0.013873710297048092, + 0.024188198149204254, + 0.020598990842700005, + 0.040874432772397995, + 0.055434081703424454, + 0.0068358308635652065, + 0.007214865647256374, + 0.01819481886923313, + 0.021593129262328148, + 0.012172256596386433, + 0.00938948430120945, + 0.01769556850194931, + 0.014723644591867924, + 0.004390594083815813, + 0.004585182294249535, + 0.014890001155436039, + 0.013875567354261875, + 0.0042840465903282166, + 0.004553360398858786, + 0.011545976623892784, + 0.009228184819221497, + 0.014231698587536812, + 0.010950888507068157, + 0.017201313748955727, + 0.01841464638710022, + 0.00643989909440279, + 0.007667682599276304, + 0.012810799293220043, + 0.010411052964627743, + 0.005971367005258799, + 0.008033000864088535, + 0.007857747375965118, + 0.01143885962665081, + 0.006641594227403402, + 0.020052410662174225, + 0.009922437369823456, + 0.020997673273086548, + 0.004974156618118286, + 0.007687711156904697, + 0.010718648321926594, + 0.010361624881625175, + 0.007525201421231031, + 0.0163055919110775, + 0.03775186464190483, + 0.015111255459487438, + 0.006210619583725929, + 0.004158492665737867, + 0.006736916955560446, + 0.004023590125143528, + 0.005102973897010088, + 0.006841784808784723, + 0.006481517571955919, + 0.009198193438351154, + 0.007411216385662556, + 0.0027594338171184063, + 0.004863171372562647, + 0.0022671285551041365, + 0.0037888798397034407, + 0.005355794448405504, + 0.007951854728162289, + 0.011029615066945553, + 0.02204260043799877, + 0.007270941976457834, + 0.04069884866476059, + 0.007253523450344801, + 0.010390094481408596, + 0.016355065628886223, + 0.014455367811024189, + 0.066338449716568, + 0.016159186139702797, + 0.005725297145545483, + 0.009669548831880093, + 0.004064522683620453, + 0.004674956202507019, + 0.009929100051522255, + 0.004552729427814484, + 0.006147508975118399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0028247239533811808 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007182164117693901 + ], + "opposite_logprob_tokens": [ + -7.250718116760254 + ], + "logprob_reward": -0.0007182164117693901, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007182164117693901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11970485001802444, + 0.10812487453222275, + 0.08339453488588333, + 0.0627286285161972, + 0.04435780271887779, + 0.16176247596740723, + 0.12294961512088776, + 0.07776640355587006, + 0.14532077312469482, + 0.07389001548290253 + ], + "credit_map": [ + 0.006967068184167147, + 0.008807421661913395, + 0.017181890085339546, + 0.012795321643352509, + 0.00794949196279049, + 0.008525101467967033, + 0.032269734889268875, + 0.025208812206983566, + 0.006969425827264786, + 0.005810578353703022, + 0.01952761970460415, + 0.012480436824262142, + 0.01760914735496044, + 0.011759631335735321, + 0.01449479442089796, + 0.0194732416421175, + 0.0076669384725391865, + 0.007160676643252373, + 0.014593460597097874, + 0.01390007883310318, + 0.007598281837999821, + 0.007017809897661209, + 0.012178982608020306, + 0.013278314843773842, + 0.006046663969755173, + 0.010727880522608757, + 0.005865158513188362, + 0.011275716125965118, + 0.00550976162776351, + 0.004657422192394733, + 0.011194519698619843, + 0.007451505865901709, + 0.0037535156589001417, + 0.005070856306701899, + 0.004773716442286968, + 0.007222558371722698, + 0.0037337832618504763, + 0.007575995288789272, + 0.004525340627878904, + 0.007702036760747433, + 0.012411314062774181, + 0.009157662279903889, + 0.01880384236574173, + 0.012234099209308624, + 0.014552971348166466, + 0.024102969095110893, + 0.04119298607110977, + 0.029306629672646523, + 0.014441125094890594, + 0.01308386866003275, + 0.014593016356229782, + 0.010492531582713127, + 0.016725288704037666, + 0.01541147381067276, + 0.014151956886053085, + 0.02405034750699997, + 0.007694416679441929, + 0.002828744240105152, + 0.0076788379810750484, + 0.004019677639007568, + 0.0066017997451126575, + 0.007236151956021786, + 0.020561974495649338, + 0.021144801750779152, + 0.0252281092107296, + 0.009966346435248852, + 0.02487340196967125, + 0.007298195268958807, + 0.013565621338784695, + 0.016149386763572693, + 0.015311210416257381, + 0.03292849659919739, + 0.022952746599912643, + 0.005921120755374432, + 0.013953689485788345, + 0.002537128748372197, + 0.005758877377957106, + 0.012996716424822807, + 0.0032058688811957836, + 0.006563865579664707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.0007182164117693901 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029784899204969406 + ], + "opposite_logprob_tokens": [ + -3.529784917831421 + ], + "logprob_reward": -0.029784899204969406, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029784899204969406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4178144037723541, + 0.06856009364128113, + 0.011471724137663841, + 0.015209924429655075, + 0.03687262535095215, + 0.013913115486502647, + 0.09734845161437988, + 0.12134084850549698, + 0.18861965835094452, + 0.028849173337221146 + ], + "credit_map": [ + 0.018411407247185707, + 0.009708755649626255, + 0.03499706834554672, + 0.019034575670957565, + 0.012781891971826553, + 0.02088688500225544, + 0.13725662231445312, + 0.16473717987537384, + 0.004793035332113504, + 0.005262588616460562, + 0.007884436286985874, + 0.009936830028891563, + 0.012854736298322678, + 0.010311824269592762, + 0.007339040283113718, + 0.010177599266171455, + 0.0008115864475257695, + 0.0007127354620024562, + 0.0017336778109893203, + 0.0021236580796539783, + 0.0010779749136418104, + 0.0012843309668824077, + 0.001746195601299405, + 0.00198156526312232, + 0.002122343983501196, + 0.0015426098834723234, + 0.002358965342864394, + 0.002384674036875367, + 0.0010087917326018214, + 0.0010390946408733726, + 0.0027719505596905947, + 0.0019814937841147184, + 0.0025152741000056267, + 0.0030389016028493643, + 0.003766549751162529, + 0.005289349239319563, + 0.002688017673790455, + 0.007493465673178434, + 0.004317956976592541, + 0.007763109169900417, + 0.0009428809862583876, + 0.001122217276133597, + 0.001562098623253405, + 0.0013655469520017505, + 0.002008747309446335, + 0.0020985081791877747, + 0.00214460794813931, + 0.00266850832849741, + 0.009947910904884338, + 0.010645987465977669, + 0.011079368181526661, + 0.007161150220781565, + 0.018482137471437454, + 0.014073537662625313, + 0.010196761228144169, + 0.015761589631438255, + 0.01193218119442463, + 0.005319644697010517, + 0.010687599889934063, + 0.0046585616655647755, + 0.007993512786924839, + 0.011919315904378891, + 0.032901257276535034, + 0.035928770899772644, + 0.027604399248957634, + 0.010551177896559238, + 0.03631702437996864, + 0.00786218885332346, + 0.01570817083120346, + 0.019308606162667274, + 0.018409380689263344, + 0.05285871401429176, + 0.00493211206048727, + 0.0029114617500454187, + 0.005280446261167526, + 0.0017270364332944155, + 0.002517228713259101, + 0.003838498145341873, + 0.0029796618036925793, + 0.004662728402763605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.029784899204969406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, close to the ground, follows a flame-red sports car as it tears through the urban night at breakneck speed. The car boasts a streamlined body and a low chassis, its wet paint gleaming and shimmering under the passing neon lights. It hurtles across the mirror-like black asphalt, freshly soaked by rain, its wheels spinning at high speed, sending up a shower of fine spray behind it. The car's bright LED taillights cast two elongated, trembling streaks of red light on the slippery road, interweaving with the cold blue and purple hues cast by the surrounding skyscrapers, creating a stark contrast. The entire scene is imbued with a sense of speed and urgency, creating a tense atmosphere of desperate escape through the cold yet glamorous urban jungle.", + "seed": 17423, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_9/index_8647_seed_17423_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.048861346766352655, + "vlm_logprob": -0.048861346766352655, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.048861346766352655, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06312651187181473 + ], + "opposite_logprob_tokens": [ + -3.563126564025879 + ], + "logprob_reward": -0.06312651187181473, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.06312651187181473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1026858240365982, + 0.049730896949768066, + 0.20315341651439667, + 0.0771520584821701, + 0.055397216230630875, + 0.03148157149553299, + 0.051998037844896317, + 0.05221601203083992, + 0.10858031362295151, + 0.26760464906692505 + ], + "credit_map": [ + 0.007511445786803961, + 0.00986199639737606, + 0.010355298407375813, + 0.014689816161990166, + 0.01573963835835457, + 0.01550633180886507, + 0.014827018603682518, + 0.014194282703101635, + 0.0028766056057065725, + 0.0035250410437583923, + 0.005476623773574829, + 0.005091451574116945, + 0.005367141682654619, + 0.00832523312419653, + 0.006454163696616888, + 0.01261463575065136, + 0.019411925226449966, + 0.01488415990024805, + 0.01396456640213728, + 0.021545285359025, + 0.02019309625029564, + 0.061172448098659515, + 0.015459776856005192, + 0.03652215749025345, + 0.008108197711408138, + 0.006915968842804432, + 0.010395645163953304, + 0.011872845701873302, + 0.007917912676930428, + 0.01097810734063387, + 0.010384379886090755, + 0.010578999295830727, + 0.006209095939993858, + 0.005724012386053801, + 0.007843229919672012, + 0.006797649431973696, + 0.005341863259673119, + 0.006040106527507305, + 0.004913879092782736, + 0.012527378275990486, + 0.004252858459949493, + 0.003829855937510729, + 0.004441397730261087, + 0.003308557439595461, + 0.004003193229436874, + 0.004198307637125254, + 0.0036443897988647223, + 0.0038030152209103107, + 0.007395725231617689, + 0.003847965504974127, + 0.007822198793292046, + 0.006192130036652088, + 0.003961305599659681, + 0.006836824119091034, + 0.007449322380125523, + 0.008492564782500267, + 0.0072750523686409, + 0.005814573261886835, + 0.006586072500795126, + 0.007004835642874241, + 0.006662016734480858, + 0.007466891314834356, + 0.00653861416503787, + 0.004867954179644585, + 0.021364258602261543, + 0.008213619701564312, + 0.01884310133755207, + 0.00881878286600113, + 0.010270430706441402, + 0.011460295878350735, + 0.005924014840275049, + 0.023685812950134277, + 0.044349946081638336, + 0.02235008217394352, + 0.01969318278133869, + 0.0366402231156826, + 0.025305824354290962, + 0.05252937600016594, + 0.029016263782978058, + 0.03771976754069328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.06312651187181473 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032757371664047 + ], + "opposite_logprob_tokens": [ + -2.350327491760254 + ], + "logprob_reward": -0.10032757371664047, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032757371664047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07270912081003189, + 0.051603954285383224, + 0.4207853078842163, + 0.045286405831575394, + 0.06779879331588745, + 0.05809686705470085, + 0.052950430661439896, + 0.04689489305019379, + 0.04727613925933838, + 0.13659809529781342 + ], + "credit_map": [ + 0.006030682008713484, + 0.008667496964335442, + 0.004684640560299158, + 0.010740864090621471, + 0.010749424807727337, + 0.011224937625229359, + 0.011265140026807785, + 0.009345936588943005, + 0.004201825242489576, + 0.005394813604652882, + 0.006956552155315876, + 0.009884494356811047, + 0.004500974901020527, + 0.008167612366378307, + 0.005914812441915274, + 0.006582867354154587, + 0.0440724678337574, + 0.029848612844944, + 0.0299689881503582, + 0.043497323989868164, + 0.03987138718366623, + 0.12436848133802414, + 0.035609129816293716, + 0.07354889065027237, + 0.005758336279541254, + 0.0059878891333937645, + 0.00443588150665164, + 0.007183666806668043, + 0.004173965658992529, + 0.006012719590216875, + 0.004910416901111603, + 0.006823529489338398, + 0.00843929685652256, + 0.00831102766096592, + 0.00865783728659153, + 0.009411429055035114, + 0.0058087557554244995, + 0.0064950319938361645, + 0.00841484498232603, + 0.012260573916137218, + 0.005357036832720041, + 0.0068318843841552734, + 0.008120560087263584, + 0.005250635091215372, + 0.009032856673002243, + 0.006878084037452936, + 0.006689466070383787, + 0.00993634108453989, + 0.008621823973953724, + 0.005520794540643692, + 0.008325895294547081, + 0.007494980003684759, + 0.0041860854253172874, + 0.005682310555130243, + 0.006859423127025366, + 0.006259119138121605, + 0.0063694315031170845, + 0.004756436217576265, + 0.007259031292051077, + 0.006982117425650358, + 0.0048430864699184895, + 0.003783301217481494, + 0.007211432326585054, + 0.005690054967999458, + 0.007468219846487045, + 0.002819628454744816, + 0.010230351239442825, + 0.003949997015297413, + 0.004321841057389975, + 0.0022807372733950615, + 0.003643795382231474, + 0.012561570852994919, + 0.022181803360581398, + 0.011187534779310226, + 0.009345722384750843, + 0.01922234706580639, + 0.013450234197080135, + 0.02614591270685196, + 0.016787024214863777, + 0.018277514725923538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.10032757371664047 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197020515799522 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197020515799522, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197020515799522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1544840782880783, + 0.06370139867067337, + 0.15026725828647614, + 0.12425839155912399, + 0.07547309994697571, + 0.0567975714802742, + 0.07215467095375061, + 0.0713941752910614, + 0.06852086633443832, + 0.16294850409030914 + ], + "credit_map": [ + 0.012027801014482975, + 0.01609761454164982, + 0.0132825318723917, + 0.0243210569024086, + 0.023070963099598885, + 0.022210873663425446, + 0.018045933917164803, + 0.025427302345633507, + 0.00377075863070786, + 0.007471483200788498, + 0.0063284761272370815, + 0.007060722913593054, + 0.006158012431114912, + 0.010311443358659744, + 0.008868887089192867, + 0.01373161282390356, + 0.010984046384692192, + 0.015819240361452103, + 0.019410409033298492, + 0.029421953484416008, + 0.012592621147632599, + 0.016162237152457237, + 0.023711072281003, + 0.0221656896173954, + 0.01386652048677206, + 0.01398226898163557, + 0.017321797087788582, + 0.018088025972247124, + 0.015566134825348854, + 0.016530031338334084, + 0.013966373167932034, + 0.014937241561710835, + 0.008124662563204765, + 0.00665291165933013, + 0.010586443357169628, + 0.012581783346831799, + 0.00650833360850811, + 0.007140378467738628, + 0.007811474613845348, + 0.01606711558997631, + 0.010867245495319366, + 0.005213080905377865, + 0.008929992094635963, + 0.004658777266740799, + 0.00699454452842474, + 0.006511698942631483, + 0.0051898714154958725, + 0.008432362228631973, + 0.015081186778843403, + 0.006618499755859375, + 0.012839769944548607, + 0.008746190927922726, + 0.0044503649696707726, + 0.0065962630324065685, + 0.008202796801924706, + 0.009619598276913166, + 0.010562693700194359, + 0.006624937057495117, + 0.01671394146978855, + 0.008626764640212059, + 0.007309656590223312, + 0.005892230197787285, + 0.009719273075461388, + 0.0059446836821734905, + 0.008306194096803665, + 0.007101774215698242, + 0.008027346804738045, + 0.011310248635709286, + 0.007772285491228104, + 0.007316716015338898, + 0.01105163712054491, + 0.007634664885699749, + 0.025671526789665222, + 0.012985060922801495, + 0.011124035343527794, + 0.02193175069987774, + 0.014623244293034077, + 0.03391651436686516, + 0.018212342634797096, + 0.024484021589159966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.018197020515799522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006342306733131409 + ], + "opposite_logprob_tokens": [ + -7.37563419342041 + ], + "logprob_reward": -0.0006342306733131409, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006342306733131409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08024746924638748, + 0.014426357112824917, + 0.03478962555527687, + 0.04382624104619026, + 0.03390471264719963, + 0.026257244870066643, + 0.03718862682580948, + 0.03832748904824257, + 0.0881601870059967, + 0.6028720736503601 + ], + "credit_map": [ + 0.00564540084451437, + 0.007735112681984901, + 0.009540479630231857, + 0.013986137695610523, + 0.011288982816040516, + 0.009806430898606777, + 0.011618112213909626, + 0.010626803152263165, + 0.001364074763841927, + 0.0019405630882829428, + 0.002233348088338971, + 0.0020149971824139357, + 0.0013520031934604049, + 0.0020870377775281668, + 0.0012903785100206733, + 0.00214395415969193, + 0.0019432639237493277, + 0.002484791213646531, + 0.003440473461523652, + 0.004899339750409126, + 0.00340358167886734, + 0.008745459839701653, + 0.0035610157065093517, + 0.006311700213700533, + 0.0042000156827270985, + 0.004832531791180372, + 0.004167275968939066, + 0.005793081596493721, + 0.0069312239065766335, + 0.006022405344992876, + 0.005232269875705242, + 0.006647434085607529, + 0.006010735873132944, + 0.003940338734537363, + 0.005146585404872894, + 0.002984705613926053, + 0.0036487625911831856, + 0.004434196278452873, + 0.002730407053604722, + 0.005008981097489595, + 0.0030765344854444265, + 0.0016854606801643968, + 0.003915093839168549, + 0.002378933597356081, + 0.0033991613890975714, + 0.0033504043240100145, + 0.0031618038192391396, + 0.0052898540161550045, + 0.005284524988383055, + 0.0037622847594320774, + 0.006074842996895313, + 0.006106880027800798, + 0.002902991371229291, + 0.003736959770321846, + 0.0049535674042999744, + 0.004366571083664894, + 0.00551292859017849, + 0.0038510197773575783, + 0.0056360261514782906, + 0.004812663421034813, + 0.004908306989818811, + 0.005099934991449118, + 0.004949761554598808, + 0.003556842217221856, + 0.01452189777046442, + 0.006821103394031525, + 0.019670188426971436, + 0.008233736269176006, + 0.007942860946059227, + 0.007409187499433756, + 0.007192856632173061, + 0.01636834628880024, + 0.11044754087924957, + 0.04256989061832428, + 0.047466013580560684, + 0.0703057274222374, + 0.057684607803821564, + 0.12503212690353394, + 0.06461542844772339, + 0.08475068211555481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0006342306733131409 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking a flame-red sports car moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint wet and gleaming under neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the car's taillights casting elongated red streaks on the wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a background of skyscrapers with cold blue and purple hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062021397054195404 + ], + "opposite_logprob_tokens": [ + -2.812021493911743 + ], + "logprob_reward": -0.062021397054195404, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062021397054195404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11852526664733887, + 0.053096428513526917, + 0.08670884370803833, + 0.0207168348133564, + 0.11952869594097137, + 0.015183473005890846, + 0.06116117164492607, + 0.05189456045627594, + 0.029236480593681335, + 0.44394826889038086 + ], + "credit_map": [ + 0.008143262937664986, + 0.009216287173330784, + 0.01261398196220398, + 0.01544968131929636, + 0.01816827431321144, + 0.019838344305753708, + 0.016503121703863144, + 0.018592309206724167, + 0.003053738735616207, + 0.002952143782749772, + 0.004105265252292156, + 0.004643382504582405, + 0.007331847678869963, + 0.008883127942681313, + 0.008075343444943428, + 0.014051579870283604, + 0.007354813627898693, + 0.005062179174274206, + 0.00827199500054121, + 0.0063414061442017555, + 0.006210274528712034, + 0.03162743151187897, + 0.003950699232518673, + 0.01789003796875477, + 0.0027282906230539083, + 0.0021351969335228205, + 0.002791269449517131, + 0.0032527842558920383, + 0.0021975210402160883, + 0.00287757720798254, + 0.002329128561541438, + 0.002405068138614297, + 0.024539271369576454, + 0.008154704235494137, + 0.014531657099723816, + 0.010564303025603294, + 0.013399246148765087, + 0.019537324085831642, + 0.011652241460978985, + 0.017149947583675385, + 0.0017756834859028459, + 0.001133578596636653, + 0.0021642460487782955, + 0.001524952705949545, + 0.0016822457546368241, + 0.0019421633332967758, + 0.001933713210746646, + 0.003026889869943261, + 0.008696858771145344, + 0.004016113467514515, + 0.008747669868171215, + 0.006199266761541367, + 0.005125008057802916, + 0.009789043106138706, + 0.00894135981798172, + 0.009645851328969002, + 0.007947849109768867, + 0.00466820178553462, + 0.009484768845140934, + 0.004353152122348547, + 0.005894132889807224, + 0.00779081741347909, + 0.006023666355758905, + 0.005731975194066763, + 0.008251974359154701, + 0.0026908693835139275, + 0.0034642580430954695, + 0.0019040838815271854, + 0.002326644491404295, + 0.00345429964363575, + 0.003065272467210889, + 0.004079076461493969, + 0.08437369763851166, + 0.034257881343364716, + 0.03326062485575676, + 0.04966290667653084, + 0.048100315034389496, + 0.08657961338758469, + 0.0464584082365036, + 0.061254821717739105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.062021397054195404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_0/index_tensor([9823, 9823])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_0/index_tensor([9823, 9823])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1f0aa7d869ea46d75e568a0dc3a8ff08b12fe9e0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_0/index_tensor([9823, 9823])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, extremely low-angle tracking shot effortlessly glides just inches above the ground, keeping perfect pace with the sleek, futuristic train as it traverses the vast desert at breakneck speed. The camera leans close to the train's gleaming metal shell, a seamless aerodynamic marvel of polished pearlescent composite material, capturing every distorted reflection of the passing scenery. Thin, pulsating rays of cool white light outline the train, highlighting its silent, overwhelming speed. It hurtles beneath a clear, deep blue sky, past rolling, wind-sculpted dunes of fine, reddish-orange sand. The low, golden sunlight of late afternoon casts long, dramatic shadows and creates a brilliant, fleeting glow across the train's surface. The sheer force of its passage kicks up a massive, turbulent plume of sand in its wake, forming a swirling vortex against the horizon. The entire scene is imbued with an awe-inspiring atmosphere of technological prowess and epic, solitary progress, creating a visceral sense of speed through the juxtaposition of ultra-advanced machinery and the timeless, pristine desert.", + "seed": 31536, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_0/index_9823_seed_31536_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.058431504480540755, + "vlm_logprob": -0.058431504480540755, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.058431504480540755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05950595438480377 + ], + "opposite_logprob_tokens": [ + -2.9345059394836426 + ], + "logprob_reward": -0.05950595438480377, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05950595438480377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24277298152446747, + 0.06498079001903534, + 0.0350060872733593, + 0.050657302141189575, + 0.024793624877929688, + 0.051161058247089386, + 0.13849779963493347, + 0.07329601049423218, + 0.1849290281534195, + 0.1339053362607956 + ], + "credit_map": [ + 0.034665536135435104, + 0.07817701250314713, + 0.03587593138217926, + 0.01153496466577053, + 0.026503393426537514, + 0.027026889845728874, + 0.013385433703660965, + 0.015603837557137012, + 0.012628032825887203, + 0.00878443568944931, + 0.005382320377975702, + 0.003105015028268099, + 0.004920887295156717, + 0.0062403385527431965, + 0.005479377694427967, + 0.018440386280417442, + 0.0057539488188922405, + 0.00611052755266428, + 0.0031806183978915215, + 0.0020368662662804127, + 0.004909968934953213, + 0.002554616890847683, + 0.0022525035310536623, + 0.008207036182284355, + 0.007179549429565668, + 0.009480420500040054, + 0.0051196967251598835, + 0.003843815764412284, + 0.006942756474018097, + 0.006469872780144215, + 0.004108339548110962, + 0.007512854877859354, + 0.003193022683262825, + 0.003180041676387191, + 0.0034247480798512697, + 0.003574347123503685, + 0.0044155847281217575, + 0.002318632323294878, + 0.002180691109970212, + 0.0025065583176910877, + 0.016101043671369553, + 0.006303538102656603, + 0.004004416987299919, + 0.003373869927600026, + 0.004569591488689184, + 0.009704582393169403, + 0.002055616583675146, + 0.005048409104347229, + 0.037546560168266296, + 0.013421128503978252, + 0.012220070697367191, + 0.0098604466766119, + 0.012473448179662228, + 0.030369393527507782, + 0.007190852425992489, + 0.015415910631418228, + 0.010371833108365536, + 0.007393310312181711, + 0.009888186119496822, + 0.013744484633207321, + 0.007992603816092014, + 0.006770925596356392, + 0.00982669461518526, + 0.0073079802095890045, + 0.015333672054111958, + 0.030474038794636726, + 0.02332489751279354, + 0.030682453885674477, + 0.028497973456978798, + 0.016475152224302292, + 0.019042473286390305, + 0.021098388358950615, + 0.01312256883829832, + 0.014961867593228817, + 0.010349305346608162, + 0.0122686717659235, + 0.01984512433409691, + 0.016400765627622604, + 0.02898380346596241, + 0.017973242327570915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.05950595438480377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801830165088177 + ], + "opposite_logprob_tokens": [ + -4.634801864624023 + ], + "logprob_reward": -0.009801830165088177, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801830165088177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15052738785743713, + 0.05301905423402786, + 0.02966165356338024, + 0.02598518691956997, + 0.020970793440937996, + 0.1412920355796814, + 0.10911877453327179, + 0.02257973700761795, + 0.16721589863300323, + 0.2796294689178467 + ], + "credit_map": [ + 0.02047613635659218, + 0.028406403958797455, + 0.0258032139390707, + 0.017299838364124298, + 0.014977178536355495, + 0.012393583543598652, + 0.010211351327598095, + 0.020959677174687386, + 0.010305200703442097, + 0.010427160188555717, + 0.005741422530263662, + 0.0030521706212311983, + 0.006222805939614773, + 0.003997673746198416, + 0.0036913659423589706, + 0.009581257589161396, + 0.004558189772069454, + 0.00400918535888195, + 0.0032442498486489058, + 0.0029553340282291174, + 0.0031887537334114313, + 0.002529137535020709, + 0.0025531467981636524, + 0.006623656954616308, + 0.004107961431145668, + 0.005277162883430719, + 0.002463721204549074, + 0.0020860936492681503, + 0.003496261779218912, + 0.003014064161106944, + 0.0022214881610125303, + 0.0033184350468218327, + 0.003334527602419257, + 0.003092403756454587, + 0.002029679948464036, + 0.0021704561077058315, + 0.00241434620693326, + 0.0023891949094831944, + 0.0014779958873987198, + 0.004062189254909754, + 0.05070171132683754, + 0.01700294017791748, + 0.01062910445034504, + 0.006681713275611401, + 0.010779219679534435, + 0.030040215700864792, + 0.004980384837836027, + 0.010476752184331417, + 0.027846794575452805, + 0.012524597346782684, + 0.006218624766916037, + 0.0059804245829582214, + 0.01045643724501133, + 0.022378526628017426, + 0.006223413627594709, + 0.01748996414244175, + 0.0030429635662585497, + 0.0023123938590288162, + 0.0024144137278199196, + 0.0035271430388092995, + 0.002652009716257453, + 0.0014566670870408416, + 0.003251573769375682, + 0.003922571893781424, + 0.015776459127664566, + 0.016787605360150337, + 0.020226776599884033, + 0.03113655000925064, + 0.0197796318680048, + 0.01162759494036436, + 0.02738877944648266, + 0.024492504075169563, + 0.03266238421201706, + 0.03625046834349632, + 0.02514510229229927, + 0.03421403840184212, + 0.02792353928089142, + 0.04121227562427521, + 0.052020471543073654, + 0.030201198533177376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.009801830165088177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294746398925781 + ], + "opposite_logprob_tokens": [ + -2.237947463989258 + ], + "logprob_reward": -0.11294746398925781, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294746398925781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32178473472595215, + 0.05927005782723427, + 0.03656447306275368, + 0.009783339686691761, + 0.014675034210085869, + 0.19802691042423248, + 0.06628239899873734, + 0.054332830011844635, + 0.14990101754665375, + 0.0893792062997818 + ], + "credit_map": [ + 0.04495084658265114, + 0.09564995020627975, + 0.045479364693164825, + 0.022159352898597717, + 0.027720503509044647, + 0.0369606576859951, + 0.01664465293288231, + 0.03221940994262695, + 0.010916122235357761, + 0.015263439156115055, + 0.003982976078987122, + 0.0028947151731699705, + 0.005485245492309332, + 0.005703803617507219, + 0.0028492470737546682, + 0.012174507603049278, + 0.005929291248321533, + 0.006801576353609562, + 0.00342436577193439, + 0.0021667699329555035, + 0.0036505155730992556, + 0.0054174321703612804, + 0.0029687134083360434, + 0.006205807905644178, + 0.002107726177200675, + 0.0013767959317192435, + 0.0008671183022670448, + 0.0007393773994408548, + 0.0010068711126223207, + 0.0009942864999175072, + 0.0007125271367840469, + 0.0019786376506090164, + 0.0020764365326613188, + 0.001699139829725027, + 0.0014790871646255255, + 0.0019031158881261945, + 0.001983321039006114, + 0.0014503179118037224, + 0.0012708534486591816, + 0.002812762279063463, + 0.06448496133089066, + 0.027659881860017776, + 0.015331807546317577, + 0.010944507084786892, + 0.015650777146220207, + 0.03847946226596832, + 0.007699955720454454, + 0.017775563523173332, + 0.013967398554086685, + 0.007410820107907057, + 0.006560294423252344, + 0.004169434774667025, + 0.006387848872691393, + 0.01046588271856308, + 0.0037661984097212553, + 0.013554519042372704, + 0.00733915064483881, + 0.006033694371581078, + 0.006935071665793657, + 0.01052376814186573, + 0.00597446458414197, + 0.005099229980260134, + 0.006617453880608082, + 0.005809993017464876, + 0.013085328042507172, + 0.022762130945920944, + 0.0164045300334692, + 0.025280212983489037, + 0.0232327152043581, + 0.011382345110177994, + 0.022270217537879944, + 0.015483533032238483, + 0.011521135456860065, + 0.010846396908164024, + 0.00826960802078247, + 0.011432947590947151, + 0.008864665403962135, + 0.009053617715835571, + 0.018140759319067, + 0.011250080540776253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.11294746398925781 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495215207338333 + ], + "opposite_logprob_tokens": [ + -2.9299521446228027 + ], + "logprob_reward": -0.05495215207338333, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495215207338333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13525263965129852, + 0.06813476234674454, + 0.056490540504455566, + 0.06004814803600311, + 0.1076529398560524, + 0.09145860373973846, + 0.22773540019989014, + 0.03164692968130112, + 0.085053451359272, + 0.13652661442756653 + ], + "credit_map": [ + 0.01950308308005333, + 0.040445029735565186, + 0.019341852515935898, + 0.007666845805943012, + 0.015467094257473946, + 0.01497790589928627, + 0.007438847795128822, + 0.010411977767944336, + 0.012466897256672382, + 0.016140973195433617, + 0.006834481842815876, + 0.004168360028415918, + 0.007633198518306017, + 0.005679079797118902, + 0.00345930689945817, + 0.01175246573984623, + 0.009260687045753002, + 0.011485274881124496, + 0.006057004444301128, + 0.004657143261283636, + 0.006155469920486212, + 0.0072872815653681755, + 0.0053047738038003445, + 0.006282908376306295, + 0.008733613416552544, + 0.009518368169665337, + 0.0065357680432498455, + 0.004653973039239645, + 0.0076962546445429325, + 0.006062337663024664, + 0.005270880647003651, + 0.011576956138014793, + 0.01491242740303278, + 0.01637653447687626, + 0.011732556857168674, + 0.015711652114987373, + 0.01749344728887081, + 0.01049583125859499, + 0.010269559919834137, + 0.010660935193300247, + 0.01936868019402027, + 0.02240697667002678, + 0.007940338924527168, + 0.0078094471246004105, + 0.00956606026738882, + 0.008759093470871449, + 0.005407300777733326, + 0.01020070817321539, + 0.062486596405506134, + 0.023788155987858772, + 0.01982625387609005, + 0.014816689305007458, + 0.02130279876291752, + 0.049260757863521576, + 0.012063276022672653, + 0.024190865457057953, + 0.0030061688739806414, + 0.0030397383961826563, + 0.004262164235115051, + 0.006278326269239187, + 0.005784559994935989, + 0.0020753268618136644, + 0.004388473927974701, + 0.002812172519043088, + 0.008040140382945538, + 0.010965252295136452, + 0.010483380407094955, + 0.016133787110447884, + 0.011694732122123241, + 0.0064402478747069836, + 0.00998103991150856, + 0.011314873583614826, + 0.012613989412784576, + 0.011115279980003834, + 0.012922123074531555, + 0.01635863445699215, + 0.012541345320641994, + 0.022900646552443504, + 0.02882707677781582, + 0.01924750581383705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.05495215207338333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495012179017067 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.05495012179017067, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495012179017067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05867020785808563, + 0.044835399836301804, + 0.015342687256634235, + 0.023605309426784515, + 0.0209044162184, + 0.046668753027915955, + 0.12954849004745483, + 0.04136098548769951, + 0.37031280994415283, + 0.2487509399652481 + ], + "credit_map": [ + 0.007861947640776634, + 0.0101527264341712, + 0.009983937256038189, + 0.008470994420349598, + 0.0052723996341228485, + 0.004167880862951279, + 0.0057091605849564075, + 0.0070511591620743275, + 0.009056180715560913, + 0.013817057944834232, + 0.0033429681789129972, + 0.0014193874085322022, + 0.0037847061175853014, + 0.004989649634808302, + 0.002040879800915718, + 0.006384569685906172, + 0.003177524311468005, + 0.002508692443370819, + 0.0018199705518782139, + 0.0010081692598760128, + 0.002272107405588031, + 0.0014356173342093825, + 0.0009101797477342188, + 0.002210426377132535, + 0.003759389976039529, + 0.004688385408371687, + 0.0020949465688318014, + 0.0014961013803258538, + 0.0029147923924028873, + 0.0031577167101204395, + 0.0019903825595974922, + 0.0035035947803407907, + 0.0035541942343115807, + 0.002854445716366172, + 0.0013081480283290148, + 0.001765123917721212, + 0.0024296147748827934, + 0.003006465034559369, + 0.0015864921733736992, + 0.0043999324552714825, + 0.008109228685498238, + 0.012654729187488556, + 0.003714967053383589, + 0.004189994186162949, + 0.0040176245383918285, + 0.002964022569358349, + 0.002852214267477393, + 0.00816597230732441, + 0.035634372383356094, + 0.014263222925364971, + 0.007740450091660023, + 0.00734726432710886, + 0.012247136794030666, + 0.028657950460910797, + 0.006975000258535147, + 0.01668309047818184, + 0.0032797767780721188, + 0.003989272750914097, + 0.005552086979150772, + 0.007210878189653158, + 0.005877206102013588, + 0.002713577589020133, + 0.007955151610076427, + 0.00478303711861372, + 0.03336223587393761, + 0.027895590290427208, + 0.04457254707813263, + 0.06762046366930008, + 0.04659227654337883, + 0.02598625421524048, + 0.06630360335111618, + 0.057979848235845566, + 0.02444474957883358, + 0.01862926222383976, + 0.023906325921416283, + 0.030561834573745728, + 0.021428357809782028, + 0.04487833008170128, + 0.05042704939842224, + 0.03447502851486206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.05495012179017067 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, extremely low-angle tracking shot effortlessly glides just inches above the ground, keeping perfect pace with the sleek, futuristic train as it traverses the vast desert at breakneck speed. The camera leans close to the train's gleaming metal shell, a seamless aerodynamic marvel of polished pearlescent composite material, capturing every distorted reflection of the passing scenery. Thin, pulsating rays of cool white light outline the train, highlighting its silent, overwhelming speed. It hurtles beneath a clear, deep blue sky, past rolling, wind-sculpted dunes of fine, reddish-orange sand. The low, golden sunlight of late afternoon casts long, dramatic shadows and creates a brilliant, fleeting glow across the train's surface. The sheer force of its passage kicks up a massive, turbulent plume of sand in its wake, forming a swirling vortex against the horizon. The entire scene is imbued with an awe-inspiring atmosphere of technological prowess and epic, solitary progress, creating a visceral sense of speed through the juxtaposition of ultra-advanced machinery and the timeless, pristine desert.", + "seed": 31537, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_0/index_9823_seed_31537_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05968505451455712, + "vlm_logprob": -0.05968505451455712, + "vlm_yes_no_margin": 3.350000023841858, + "vlm_token_credit": -0.05968505451455712, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036173105239868164 + ], + "opposite_logprob_tokens": [ + -3.411173105239868 + ], + "logprob_reward": -0.036173105239868164, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036173105239868164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.198788583278656, + 0.0515470989048481, + 0.02893289551138878, + 0.031551819294691086, + 0.055949799716472626, + 0.2606031596660614, + 0.26959386467933655, + 0.026571836322546005, + 0.03275054693222046, + 0.04371044039726257 + ], + "credit_map": [ + 0.03035491146147251, + 0.02955273538827896, + 0.013505753129720688, + 0.01701967790722847, + 0.03758509084582329, + 0.021613281220197678, + 0.028183721005916595, + 0.020973388105630875, + 0.0071832649409770966, + 0.007998649030923843, + 0.00589150981977582, + 0.0036272972356528044, + 0.006993955001235008, + 0.006381264887750149, + 0.005597002804279327, + 0.007874147035181522, + 0.0035352110862731934, + 0.004275532905012369, + 0.0035860338248312473, + 0.004162913653999567, + 0.0027541599702090025, + 0.0035688942298293114, + 0.0029836895409971476, + 0.004066456574946642, + 0.005302550736814737, + 0.004509634803980589, + 0.0013290662318468094, + 0.0033214024733752012, + 0.005530176218599081, + 0.006061030086129904, + 0.002943894360214472, + 0.0025540634524077177, + 0.011830289848148823, + 0.006904541980475187, + 0.0033143542241305113, + 0.004452867899090052, + 0.006382856518030167, + 0.011748875491321087, + 0.006598186679184437, + 0.004717818461358547, + 0.03919851407408714, + 0.044124070554971695, + 0.013437415473163128, + 0.02801806665956974, + 0.040952008217573166, + 0.04597040265798569, + 0.029557308182120323, + 0.019345324486494064, + 0.08156555145978928, + 0.04369262233376503, + 0.007910984568297863, + 0.006553777493536472, + 0.021744227036833763, + 0.08337238430976868, + 0.007370419334620237, + 0.017383884638547897, + 0.0034403284080326557, + 0.0038826586678624153, + 0.00188640842679888, + 0.0028953992296010256, + 0.0037209009751677513, + 0.0037566572427749634, + 0.004095534328371286, + 0.0028939461335539818, + 0.0033356756903231144, + 0.0038107933942228556, + 0.0020618264097720385, + 0.005530704278498888, + 0.007661606650799513, + 0.0033064312301576138, + 0.004664973821491003, + 0.00237853080034256, + 0.004918823484331369, + 0.006798334885388613, + 0.005226897541433573, + 0.0029039797373116016, + 0.00706004723906517, + 0.006992250680923462, + 0.00581897608935833, + 0.003991124220192432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.036173105239868164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410511763766408 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.00410511763766408, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410511763766408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06227751076221466, + 0.044147562235593796, + 0.023077843710780144, + 0.014065162278711796, + 0.018540697172284126, + 0.022809665650129318, + 0.6966351270675659, + 0.023447399958968163, + 0.009403395466506481, + 0.08559564501047134 + ], + "credit_map": [ + 0.006515595596283674, + 0.0069382283836603165, + 0.009297456592321396, + 0.010049735195934772, + 0.010041600093245506, + 0.006847715470939875, + 0.005396476946771145, + 0.007190702483057976, + 0.006562583614140749, + 0.008967610076069832, + 0.003158991690725088, + 0.00370803102850914, + 0.006494320463389158, + 0.005349267274141312, + 0.004946080502122641, + 0.004960677586495876, + 0.0034650017041713, + 0.003650993574410677, + 0.002029651077464223, + 0.002131447661668062, + 0.0027039211709052324, + 0.002770290244370699, + 0.002997035626322031, + 0.003329501487314701, + 0.0021047862246632576, + 0.0019519905326887965, + 0.0010411132825538516, + 0.0013683311408385634, + 0.0019521318608894944, + 0.002411846537142992, + 0.0019508982077240944, + 0.0012840634444728494, + 0.0022735372185707092, + 0.0024907896295189857, + 0.0013194183120504022, + 0.001548752305097878, + 0.0026788688264787197, + 0.0023004186805337667, + 0.0038178213872015476, + 0.002111089648678899, + 0.002639570739120245, + 0.0024255274329334497, + 0.0015294441254809499, + 0.003084588563069701, + 0.0037092987913638353, + 0.003645261051133275, + 0.003848151070997119, + 0.0019278248073533177, + 0.2112729698419571, + 0.11293745785951614, + 0.01995650678873062, + 0.016961587592959404, + 0.055237654596567154, + 0.21614724397659302, + 0.019347205758094788, + 0.04477446898818016, + 0.0023305558133870363, + 0.0020657742861658335, + 0.0030842190608382225, + 0.0027154425624758005, + 0.0025418249424546957, + 0.0025407918728888035, + 0.004146053921431303, + 0.004022737964987755, + 0.001323962351307273, + 0.0013785974588245153, + 0.0005039477255195379, + 0.0007484034285880625, + 0.0029296488501131535, + 0.00117672816850245, + 0.0006832578219473362, + 0.0006588498363271356, + 0.009144692681729794, + 0.007287661079317331, + 0.016343725845217705, + 0.0077986940741539, + 0.013429274782538414, + 0.0122240474447608, + 0.00924601498991251, + 0.010121537372469902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.00410511763766408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427995264530182 + ], + "opposite_logprob_tokens": [ + -2.0177996158599854 + ], + "logprob_reward": -0.1427995264530182, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1427995264530182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09547723084688187, + 0.09703512489795685, + 0.16263291239738464, + 0.02756667509675026, + 0.11587049812078476, + 0.1043500155210495, + 0.21099254488945007, + 0.12492892146110535, + 0.0198870487511158, + 0.04125900939106941 + ], + "credit_map": [ + 0.013342675752937794, + 0.012920246459543705, + 0.00812689121812582, + 0.009083135984838009, + 0.019051186740398407, + 0.014512493275105953, + 0.009458400309085846, + 0.008982211351394653, + 0.012719781138002872, + 0.014872909523546696, + 0.008681179955601692, + 0.009789520874619484, + 0.012071119621396065, + 0.009613036178052425, + 0.01237921230494976, + 0.01690838299691677, + 0.018566755577921867, + 0.016936197876930237, + 0.015072813257575035, + 0.0289171002805233, + 0.01000937819480896, + 0.0140297319740057, + 0.036586061120033264, + 0.022514892742037773, + 0.003391377627849579, + 0.0028038981836289167, + 0.002985714003443718, + 0.003267472144216299, + 0.0034002449829131365, + 0.004252918064594269, + 0.004426062572747469, + 0.003038989147171378, + 0.025068966671824455, + 0.01674766279757023, + 0.007647526450455189, + 0.00843621976673603, + 0.014194672927260399, + 0.02217513881623745, + 0.010144941508769989, + 0.011455374769866467, + 0.010468256659805775, + 0.013090603984892368, + 0.00659398315474391, + 0.015896636992692947, + 0.013706779107451439, + 0.018452053889632225, + 0.01798255927860737, + 0.008159150369465351, + 0.061889320611953735, + 0.03473982959985733, + 0.007676078472286463, + 0.008291599340736866, + 0.016323430463671684, + 0.06149568781256676, + 0.007446293719112873, + 0.013130320236086845, + 0.017280763015151024, + 0.01274056639522314, + 0.008584112860262394, + 0.01875264383852482, + 0.01219622790813446, + 0.018612824380397797, + 0.022720223292708397, + 0.014041568152606487, + 0.0022519631311297417, + 0.0026530411560088396, + 0.0017239877488464117, + 0.0044198972173035145, + 0.002554407576099038, + 0.0022023264318704605, + 0.0023904580157250166, + 0.0016909701516851783, + 0.003278292715549469, + 0.005244643427431583, + 0.003929812461137772, + 0.005824326071888208, + 0.005661713890731335, + 0.005188580136746168, + 0.0070938472636044025, + 0.00503779761493206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.1427995264530182 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633783593773842 + ], + "opposite_logprob_tokens": [ + -3.6513378620147705 + ], + "logprob_reward": -0.02633783593773842, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633783593773842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05380246788263321, + 0.03224090114235878, + 0.04050031676888466, + 0.05416421592235565, + 0.18860971927642822, + 0.18982185423374176, + 0.357173889875412, + 0.022714875638484955, + 0.04425656050443649, + 0.016715234145522118 + ], + "credit_map": [ + 0.006586229894310236, + 0.007346897851675749, + 0.004744849633425474, + 0.00550489779561758, + 0.00925261527299881, + 0.006097183097153902, + 0.00803979393094778, + 0.006229998543858528, + 0.004329855553805828, + 0.004884166177362204, + 0.003364977426826954, + 0.002408834872767329, + 0.004068989772349596, + 0.003969312645494938, + 0.0037095309235155582, + 0.005505230277776718, + 0.005562438629567623, + 0.005236573051661253, + 0.0035569542087614536, + 0.005180403124541044, + 0.002977074822410941, + 0.003979487810283899, + 0.006959923543035984, + 0.007047457154840231, + 0.008567824959754944, + 0.0091866385191679, + 0.0027096057310700417, + 0.005125403869897127, + 0.007699084468185902, + 0.009713025763630867, + 0.005781172309070826, + 0.005381457973271608, + 0.01731465943157673, + 0.01986510120332241, + 0.018037021160125732, + 0.023387666791677475, + 0.024867134168744087, + 0.025027574971318245, + 0.03634142130613327, + 0.023769132792949677, + 0.02313588187098503, + 0.02948666363954544, + 0.011227772571146488, + 0.025491496548056602, + 0.03029470331966877, + 0.02859853021800518, + 0.027338648214936256, + 0.014248144812881947, + 0.10873476415872574, + 0.057638444006443024, + 0.010216915048658848, + 0.008125854656100273, + 0.02825876697897911, + 0.11113197356462479, + 0.009972484782338142, + 0.0230946633964777, + 0.0025351541116833687, + 0.00313961342908442, + 0.0020279919262975454, + 0.002672103699296713, + 0.002292326418682933, + 0.0024054059758782387, + 0.0050281318835914135, + 0.002614147961139679, + 0.006794695742428303, + 0.007446570787578821, + 0.0024686981923878193, + 0.005246828775852919, + 0.008565392345190048, + 0.006159457378089428, + 0.004355891142040491, + 0.003219024045392871, + 0.0013772522797808051, + 0.0015832409262657166, + 0.0026719800662249327, + 0.001602580421604216, + 0.002690614201128483, + 0.0023258961737155914, + 0.002588945673778653, + 0.0018747232388705015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.02633783593773842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900968730449677 + ], + "opposite_logprob_tokens": [ + -2.4640097618103027 + ], + "logprob_reward": -0.08900968730449677, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900968730449677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15848198533058167, + 0.1390729397535324, + 0.09207162261009216, + 0.03343531861901283, + 0.1185079962015152, + 0.10846680402755737, + 0.21290764212608337, + 0.05792662501335144, + 0.015955470502376556, + 0.06317362934350967 + ], + "credit_map": [ + 0.017841622233390808, + 0.014441058970987797, + 0.019567027688026428, + 0.03158833086490631, + 0.020991990342736244, + 0.013825658708810806, + 0.022046683356165886, + 0.018179606646299362, + 0.01916460692882538, + 0.02555559203028679, + 0.011193078942596912, + 0.01261907909065485, + 0.01993958279490471, + 0.01699712686240673, + 0.016738174483180046, + 0.016865691170096397, + 0.011282198131084442, + 0.014965102076530457, + 0.010975529439747334, + 0.0075765736401081085, + 0.009367311373353004, + 0.011151907034218311, + 0.01126028597354889, + 0.015492708422243595, + 0.004279555287212133, + 0.005440585780888796, + 0.0022946076933294535, + 0.003995175007730722, + 0.005356851499527693, + 0.00391857884824276, + 0.0045563820749521255, + 0.0035935835912823677, + 0.01796817034482956, + 0.013582493178546429, + 0.009166753850877285, + 0.011988209560513496, + 0.012834195978939533, + 0.01941497251391411, + 0.02112317830324173, + 0.012430011294782162, + 0.020194459706544876, + 0.01810513250529766, + 0.005859516095370054, + 0.008481902070343494, + 0.015263883396983147, + 0.02407359518110752, + 0.007410392630845308, + 0.009077916853129864, + 0.06302277743816376, + 0.034257613122463226, + 0.006677190773189068, + 0.006078383885324001, + 0.018491892144083977, + 0.06510166823863983, + 0.005940975621342659, + 0.01333712600171566, + 0.006550958380103111, + 0.01023299340158701, + 0.00464972946792841, + 0.0050739385187625885, + 0.00844992883503437, + 0.005968300625681877, + 0.010822602547705173, + 0.0061781746335327625, + 0.0025364502798765898, + 0.0027495103422552347, + 0.0009495148551650345, + 0.001793091301806271, + 0.0031645293347537518, + 0.0017851808806881309, + 0.0018023871816694736, + 0.0011748052202165127, + 0.007214606739580631, + 0.008944420143961906, + 0.005199426785111427, + 0.006732186768203974, + 0.010240801610052586, + 0.007486925460398197, + 0.01121729426085949, + 0.006137961056083441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.08900968730449677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, extremely low-angle tracking shot effortlessly glides just inches above the ground, keeping perfect pace with the sleek, futuristic train as it traverses the vast desert at breakneck speed. The camera leans close to the train's gleaming metal shell, a seamless aerodynamic marvel of polished pearlescent composite material, capturing every distorted reflection of the passing scenery. Thin, pulsating rays of cool white light outline the train, highlighting its silent, overwhelming speed. It hurtles beneath a clear, deep blue sky, past rolling, wind-sculpted dunes of fine, reddish-orange sand. The low, golden sunlight of late afternoon casts long, dramatic shadows and creates a brilliant, fleeting glow across the train's surface. The sheer force of its passage kicks up a massive, turbulent plume of sand in its wake, forming a swirling vortex against the horizon. The entire scene is imbued with an awe-inspiring atmosphere of technological prowess and epic, solitary progress, creating a visceral sense of speed through the juxtaposition of ultra-advanced machinery and the timeless, pristine desert.", + "seed": 31538, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_0/index_9823_seed_31538_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04799846895039082, + "vlm_logprob": -0.04799846895039082, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04799846895039082, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0825439989566803 + ], + "opposite_logprob_tokens": [ + -2.5825440883636475 + ], + "logprob_reward": -0.0825439989566803, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0825439989566803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12050726264715195, + 0.04040834680199623, + 0.06155584007501602, + 0.05563319846987724, + 0.0714559406042099, + 0.16624869406223297, + 0.05131423845887184, + 0.0999981239438057, + 0.06436973065137863, + 0.2685086727142334 + ], + "credit_map": [ + 0.011920483782887459, + 0.020300837233662605, + 0.012425708584487438, + 0.020140552893280983, + 0.02316002920269966, + 0.010043024085462093, + 0.011129772290587425, + 0.01138683594763279, + 0.004150559194386005, + 0.00465650437399745, + 0.004032500553876162, + 0.005497563164681196, + 0.0099344402551651, + 0.004841430112719536, + 0.003626978024840355, + 0.0036683669313788414, + 0.005967587232589722, + 0.0061718011274933815, + 0.011781134642660618, + 0.004391171038150787, + 0.011973632499575615, + 0.014049187302589417, + 0.0036373997572809458, + 0.003583919955417514, + 0.006292732898145914, + 0.005640801973640919, + 0.006740499287843704, + 0.007395151536911726, + 0.011571935378015041, + 0.007169615011662245, + 0.0040304213762283325, + 0.006792031694203615, + 0.007740304339677095, + 0.0071312845684587955, + 0.00974977109581232, + 0.009933780878782272, + 0.014683039858937263, + 0.00583310192450881, + 0.007761050947010517, + 0.008623601868748665, + 0.01908358559012413, + 0.023796910420060158, + 0.0337190218269825, + 0.012664047069847584, + 0.014261314645409584, + 0.03478759527206421, + 0.01364817563444376, + 0.014288021251559258, + 0.005219374317675829, + 0.004560212139040232, + 0.00779731897637248, + 0.006444265134632587, + 0.008738799951970577, + 0.005055150482803583, + 0.008062557317316532, + 0.005436552222818136, + 0.012016881257295609, + 0.012853242456912994, + 0.012127434834837914, + 0.010782112367451191, + 0.020727500319480896, + 0.008321529254317284, + 0.011648495681583881, + 0.011520922183990479, + 0.005318604409694672, + 0.004907854367047548, + 0.007114199921488762, + 0.0059110308066010475, + 0.009821703657507896, + 0.00766765559092164, + 0.014005706645548344, + 0.009622969664633274, + 0.022173777222633362, + 0.03257850185036659, + 0.034222472459077835, + 0.037992365658283234, + 0.06031566113233566, + 0.025259489193558693, + 0.03164223954081535, + 0.02432413212954998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0825439989566803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266720429062843 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.005266720429062843, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266720429062843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10636585205793381, + 0.04077918455004692, + 0.029882226139307022, + 0.09065595269203186, + 0.1256805956363678, + 0.11392302811145782, + 0.04079226776957512, + 0.08183102309703827, + 0.21888208389282227, + 0.15120777487754822 + ], + "credit_map": [ + 0.012543927878141403, + 0.0156815517693758, + 0.012178979814052582, + 0.012681052088737488, + 0.023729141801595688, + 0.010861373506486416, + 0.00956741813570261, + 0.009122411720454693, + 0.003648267360404134, + 0.0040395185351371765, + 0.0044388906098902225, + 0.008425235748291016, + 0.006075192708522081, + 0.003997515421360731, + 0.004974936600774527, + 0.005179630126804113, + 0.0037188557907938957, + 0.004128340166062117, + 0.002561679109930992, + 0.004105620551854372, + 0.008332873694598675, + 0.002258411841467023, + 0.002541061257943511, + 0.0022353874519467354, + 0.01025369856506586, + 0.01247898768633604, + 0.011354017071425915, + 0.012654080055654049, + 0.017169641330838203, + 0.008759569376707077, + 0.006507359445095062, + 0.011478611268103123, + 0.013498567044734955, + 0.014450831338763237, + 0.014745869673788548, + 0.014632713980972767, + 0.023122724145650864, + 0.011736813001334667, + 0.011484366841614246, + 0.022008711472153664, + 0.013649620115756989, + 0.010590380057692528, + 0.022085977718234062, + 0.011721224524080753, + 0.012051586993038654, + 0.022884748876094818, + 0.012657353654503822, + 0.008282145485281944, + 0.006724359933286905, + 0.004218704532831907, + 0.004928471054881811, + 0.003825398627668619, + 0.006877911277115345, + 0.005019757896661758, + 0.005023686680942774, + 0.004173981491476297, + 0.007177084218710661, + 0.007043789606541395, + 0.013812743127346039, + 0.010579580441117287, + 0.010155702009797096, + 0.015155255794525146, + 0.011087881401181221, + 0.00681898882612586, + 0.01692933961749077, + 0.018592240288853645, + 0.02496243827044964, + 0.02897719480097294, + 0.026889633387327194, + 0.018483050167560577, + 0.058155857026576996, + 0.025892343372106552, + 0.012491102330386639, + 0.014962799847126007, + 0.024619024246931076, + 0.021866479888558388, + 0.027537433430552483, + 0.016677597537636757, + 0.01624239794909954, + 0.016810959205031395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.005266720429062843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370397910475731 + ], + "opposite_logprob_tokens": [ + -3.4087040424346924 + ], + "logprob_reward": -0.03370397910475731, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370397910475731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07858704775571823, + 0.027615852653980255, + 0.08067876100540161, + 0.022791346535086632, + 0.04340663179755211, + 0.23704656958580017, + 0.028562109917402267, + 0.06312073767185211, + 0.16964048147201538, + 0.24855050444602966 + ], + "credit_map": [ + 0.008861188776791096, + 0.010693774558603764, + 0.008648605085909367, + 0.010872798971831799, + 0.013217403553426266, + 0.00815365370362997, + 0.009481522254645824, + 0.00865810178220272, + 0.0034657518845051527, + 0.0038269723299890757, + 0.004184102173894644, + 0.004763571545481682, + 0.004275237210094929, + 0.0021617538295686245, + 0.002545802854001522, + 0.002392663387581706, + 0.007294614799320698, + 0.006759010720998049, + 0.02028360590338707, + 0.004768969025462866, + 0.00794050469994545, + 0.02556583285331726, + 0.003885475220158696, + 0.004180749412626028, + 0.0021722507663071156, + 0.0024026078172028065, + 0.002218579640612006, + 0.0030419197864830494, + 0.004703233018517494, + 0.0026449004653841257, + 0.003196039004251361, + 0.0024118151050060987, + 0.004659435246139765, + 0.0046721710823476315, + 0.005302173551172018, + 0.004165547899901867, + 0.008631083182990551, + 0.004172995686531067, + 0.004106104373931885, + 0.007697117980569601, + 0.018102474510669708, + 0.038714922964572906, + 0.06900285929441452, + 0.01201287005096674, + 0.014919321052730083, + 0.060504935681819916, + 0.0086662657558918, + 0.015122911892831326, + 0.00406422046944499, + 0.0036520299036055803, + 0.002834624843671918, + 0.0034580847714096308, + 0.004312486387789249, + 0.002836057683452964, + 0.003417774336412549, + 0.003986830357462168, + 0.005439328029751778, + 0.006938111502677202, + 0.007794322445988655, + 0.008814883418381214, + 0.012433276511728764, + 0.006620653439313173, + 0.009731141850352287, + 0.005349018145352602, + 0.012438408099114895, + 0.01646830514073372, + 0.019786162301898003, + 0.019979052245616913, + 0.02179335430264473, + 0.014895154163241386, + 0.047975778579711914, + 0.016304273158311844, + 0.02253498136997223, + 0.02358577959239483, + 0.024917641654610634, + 0.03629440441727638, + 0.040949877351522446, + 0.025821302086114883, + 0.03861119970679283, + 0.03583529591560364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.03370397910475731 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818695291876793 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818695291876793, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818695291876793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021939324215054512, + 0.02220160886645317, + 0.13662958145141602, + 0.025113480165600777, + 0.016432756558060646, + 0.6996029615402222, + 0.01799723505973816, + 0.027661044150590897, + 0.015282636508345604, + 0.017139391973614693 + ], + "credit_map": [ + 0.0026346337981522083, + 0.0032030900474637747, + 0.0027692122384905815, + 0.003943053539842367, + 0.00338565232232213, + 0.0018503996543586254, + 0.002188208280131221, + 0.0019650741014629602, + 0.0019566041883081198, + 0.0020481818355619907, + 0.0025921615306288004, + 0.005042228382080793, + 0.0028756866231560707, + 0.0023757051676511765, + 0.002740339608862996, + 0.0025707033928483725, + 0.011057665571570396, + 0.010728398337960243, + 0.03400243818759918, + 0.009589314460754395, + 0.014040558598935604, + 0.042621348053216934, + 0.0072137340903282166, + 0.0073761302046477795, + 0.0024493574164807796, + 0.002914703218266368, + 0.0031893285922706127, + 0.0032043226528912783, + 0.004144955426454544, + 0.0026223682798445225, + 0.003307783743366599, + 0.0032806633971631527, + 0.0017719583120197058, + 0.0017254614504054189, + 0.0022095723543316126, + 0.0017179062124341726, + 0.0030404780991375446, + 0.0016368937212973833, + 0.0014510644832625985, + 0.002879421692341566, + 0.05093539506196976, + 0.10654055327177048, + 0.2070675641298294, + 0.032973963767290115, + 0.04672670364379883, + 0.19039219617843628, + 0.018913909792900085, + 0.04605269432067871, + 0.002765125362202525, + 0.0019956971518695354, + 0.002060742350295186, + 0.002036125399172306, + 0.0029321121983230114, + 0.002261265181005001, + 0.001900164526887238, + 0.002046002773568034, + 0.002546439180150628, + 0.0035293076653033495, + 0.004222442861646414, + 0.003453061217442155, + 0.004501029849052429, + 0.004457138013094664, + 0.0025985653046518564, + 0.002353060757741332, + 0.0014699468156322837, + 0.0012754540657624602, + 0.0019642391707748175, + 0.001957897562533617, + 0.0020763438660651445, + 0.0014756133314222097, + 0.002140970202162862, + 0.002922171726822853, + 0.0019435933791100979, + 0.0022788173519074917, + 0.0023169012274593115, + 0.0020681689493358135, + 0.0028836652636528015, + 0.0013436084846034646, + 0.0020348893012851477, + 0.0022697465028613806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.01818695291876793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002906933426857 + ], + "opposite_logprob_tokens": [ + -2.3502907752990723 + ], + "logprob_reward": -0.1002906933426857, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002906933426857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08927157521247864, + 0.021391890943050385, + 0.046032994985580444, + 0.030673615634441376, + 0.027092618867754936, + 0.5454680919647217, + 0.04264134541153908, + 0.057632926851511, + 0.09033835679292679, + 0.049456674605607986 + ], + "credit_map": [ + 0.011988576501607895, + 0.011436440981924534, + 0.011747341603040695, + 0.01114690862596035, + 0.015658076852560043, + 0.008530333638191223, + 0.010176208801567554, + 0.008587684482336044, + 0.0022101716604083776, + 0.002992850262671709, + 0.0021955687552690506, + 0.0034565962851047516, + 0.003508320078253746, + 0.0022982803639024496, + 0.0025707536842674017, + 0.0021593505516648293, + 0.0050780288875103, + 0.0036702256184071302, + 0.008918704465031624, + 0.004653968848288059, + 0.007255918346345425, + 0.011209885589778423, + 0.002879155334085226, + 0.0023671058006584644, + 0.002966376021504402, + 0.003579891985282302, + 0.004864223301410675, + 0.005282046739012003, + 0.005033115856349468, + 0.0026376843452453613, + 0.0031115198507905006, + 0.00319875730201602, + 0.003092237515375018, + 0.003028960432857275, + 0.0027943640016019344, + 0.0026636680122464895, + 0.004445826169103384, + 0.0031808416824787855, + 0.0024607679806649685, + 0.0054259528405964375, + 0.044069018214941025, + 0.07660683244466782, + 0.1566629558801651, + 0.03190561383962631, + 0.03896746784448624, + 0.14542369544506073, + 0.016898568719625473, + 0.03493393957614899, + 0.005579405464231968, + 0.00439144903793931, + 0.006569385062903166, + 0.00533626414835453, + 0.006162925157696009, + 0.003891999600455165, + 0.005584409460425377, + 0.005125507712364197, + 0.004261804278939962, + 0.005411739461123943, + 0.009573336690664291, + 0.008621415123343468, + 0.008790629915893078, + 0.009656490758061409, + 0.0060617937706410885, + 0.00525572057813406, + 0.005465770605951548, + 0.00897037424147129, + 0.012396173551678658, + 0.015453163534402847, + 0.010446165688335896, + 0.007622158620506525, + 0.022907037287950516, + 0.007077512796968222, + 0.005080641712993383, + 0.006680251099169254, + 0.005605199374258518, + 0.00745461555197835, + 0.006701760459691286, + 0.004366954322904348, + 0.006478399969637394, + 0.00708885258063674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.1002906933426857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, extremely low-angle tracking shot effortlessly glides just inches above the ground, keeping perfect pace with the sleek, futuristic train as it traverses the vast desert at breakneck speed. The camera leans close to the train's gleaming metal shell, a seamless aerodynamic marvel of polished pearlescent composite material, capturing every distorted reflection of the passing scenery. Thin, pulsating rays of cool white light outline the train, highlighting its silent, overwhelming speed. It hurtles beneath a clear, deep blue sky, past rolling, wind-sculpted dunes of fine, reddish-orange sand. The low, golden sunlight of late afternoon casts long, dramatic shadows and creates a brilliant, fleeting glow across the train's surface. The sheer force of its passage kicks up a massive, turbulent plume of sand in its wake, forming a swirling vortex against the horizon. The entire scene is imbued with an awe-inspiring atmosphere of technological prowess and epic, solitary progress, creating a visceral sense of speed through the juxtaposition of ultra-advanced machinery and the timeless, pristine desert.", + "seed": 31539, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_0/index_9823_seed_31539_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.038456227816641333, + "vlm_logprob": -0.038456227816641333, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.038456227816641333, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013414955697953701 + ], + "opposite_logprob_tokens": [ + -4.6384148597717285 + ], + "logprob_reward": -0.013414955697953701, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013414955697953701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2958744168281555, + 0.19567692279815674, + 0.0487271212041378, + 0.05401450768113136, + 0.06713838875293732, + 0.21535706520080566, + 0.016996655613183975, + 0.022354165092110634, + 0.03499188274145126, + 0.048868853598833084 + ], + "credit_map": [ + 0.047572098672389984, + 0.025021465495228767, + 0.054856281727552414, + 0.02735745720565319, + 0.027295170351862907, + 0.05643106624484062, + 0.023635633289813995, + 0.03370525315403938, + 0.0465550571680069, + 0.015231519006192684, + 0.0168645977973938, + 0.008980071172118187, + 0.011620152741670609, + 0.0684138759970665, + 0.010077490471303463, + 0.017934167757630348, + 0.00598204554989934, + 0.003886597231030464, + 0.008356420323252678, + 0.004878550302237272, + 0.009731261059641838, + 0.005705528426915407, + 0.00473377387970686, + 0.005452949088066816, + 0.010212979279458523, + 0.0036133145913481712, + 0.006300767883658409, + 0.004847102332860231, + 0.005473277065902948, + 0.014408371411263943, + 0.0035021305084228516, + 0.0056565687991678715, + 0.006729777902364731, + 0.004213693086057901, + 0.011220664717257023, + 0.006912195589393377, + 0.013361421413719654, + 0.009184260852634907, + 0.008372243493795395, + 0.007144132163375616, + 0.047957874834537506, + 0.018819062039256096, + 0.0175168476998806, + 0.012690577656030655, + 0.01812782697379589, + 0.06467727571725845, + 0.012904131785035133, + 0.022663483396172523, + 0.0021252459846436977, + 0.0015580952167510986, + 0.00293113780207932, + 0.0013312594965100288, + 0.0033715569879859686, + 0.0025484522338956594, + 0.0014046182623133063, + 0.0017262906767427921, + 0.002188846468925476, + 0.001500487676821649, + 0.0036245586816221476, + 0.002546395640820265, + 0.0040994370356202126, + 0.003554114606231451, + 0.002476918511092663, + 0.002363410545513034, + 0.003258295124396682, + 0.0026755454018712044, + 0.005599240772426128, + 0.00418311869725585, + 0.006155609618872404, + 0.0043256282806396484, + 0.005222429987043142, + 0.0035720164887607098, + 0.010315687395632267, + 0.00395773071795702, + 0.004371013026684523, + 0.003575593465939164, + 0.0038584175053983927, + 0.014567503705620766, + 0.0032066982239484787, + 0.005016213282942772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.013414955697953701 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029811859130859375 + ], + "opposite_logprob_tokens": [ + -3.5298118591308594 + ], + "logprob_reward": -0.029811859130859375, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029811859130859375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3776717782020569, + 0.10903915762901306, + 0.05139506235718727, + 0.1250293105840683, + 0.03609022498130798, + 0.16604416072368622, + 0.04825408384203911, + 0.037224579602479935, + 0.026232313364744186, + 0.023019280284643173 + ], + "credit_map": [ + 0.0866154208779335, + 0.0345282219350338, + 0.03686864674091339, + 0.02435341291129589, + 0.030094576999545097, + 0.10785266757011414, + 0.021981095895171165, + 0.03537774085998535, + 0.018372224643826485, + 0.00746889179572463, + 0.012661420740187168, + 0.006536392495036125, + 0.014873688109219074, + 0.03051973506808281, + 0.009056363254785538, + 0.009550445713102818, + 0.007505943067371845, + 0.003735828446224332, + 0.007772746495902538, + 0.0048037683591246605, + 0.006746675819158554, + 0.010161515325307846, + 0.005666685756295919, + 0.005001903045922518, + 0.02882513217628002, + 0.011319939978420734, + 0.012255854904651642, + 0.007558730430901051, + 0.008781883865594864, + 0.03967763110995293, + 0.005146847106516361, + 0.011463291943073273, + 0.0036645233631134033, + 0.0025935517624020576, + 0.006830998230725527, + 0.004696966614574194, + 0.005931968800723553, + 0.004729429725557566, + 0.003970016725361347, + 0.0036727699916809797, + 0.03700844198465347, + 0.013780343346297741, + 0.013128083199262619, + 0.009675317443907261, + 0.011492465622723103, + 0.053619347512722015, + 0.008953438140451908, + 0.01838671788573265, + 0.005155546125024557, + 0.002858776366338134, + 0.010217277333140373, + 0.004466052167117596, + 0.007807958405464888, + 0.009071511216461658, + 0.004040492698550224, + 0.004636468831449747, + 0.0029649147763848305, + 0.002529814140871167, + 0.00804049987345934, + 0.0052916621789336205, + 0.004604380112141371, + 0.0040918923914432526, + 0.003961967304348946, + 0.005739447195082903, + 0.002207527169957757, + 0.0016875627916306257, + 0.005676375236362219, + 0.0029541528783738613, + 0.003996643703430891, + 0.0031798265408724546, + 0.003333507338538766, + 0.0031967158429324627, + 0.003931835759431124, + 0.0022162077948451042, + 0.0032956611830741167, + 0.002140999538823962, + 0.002356995129957795, + 0.004994926042854786, + 0.001953018014319241, + 0.0021296378690749407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.029811859130859375 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902692049741745 + ], + "opposite_logprob_tokens": [ + -2.464026927947998 + ], + "logprob_reward": -0.08902692049741745, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902692049741745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3283982276916504, + 0.2198353409767151, + 0.06012200191617012, + 0.03334576636552811, + 0.017324503511190414, + 0.15561352670192719, + 0.018022971227765083, + 0.015349999070167542, + 0.027253273874521255, + 0.12473440170288086 + ], + "credit_map": [ + 0.07543176412582397, + 0.027773316949605942, + 0.0263441763818264, + 0.020243562757968903, + 0.02771502546966076, + 0.10574259608983994, + 0.016124607995152473, + 0.029023146256804466, + 0.053369104862213135, + 0.01683705486357212, + 0.017213229089975357, + 0.010175456292927265, + 0.012595292180776596, + 0.07893843948841095, + 0.00952732190489769, + 0.02117942087352276, + 0.011136187240481377, + 0.005086968187242746, + 0.008459256030619144, + 0.004360739141702652, + 0.007588525768369436, + 0.013392706401646137, + 0.005193950142711401, + 0.004903669003397226, + 0.005805064458400011, + 0.0027441177517175674, + 0.005308401770889759, + 0.0030748548451811075, + 0.0035090013407170773, + 0.007846765220165253, + 0.002274256432428956, + 0.0027833005879074335, + 0.001603471813723445, + 0.001189583563245833, + 0.0033404340501874685, + 0.001800264697521925, + 0.0028098064940422773, + 0.00214062980376184, + 0.0028645507991313934, + 0.0015757613582536578, + 0.033276431262493134, + 0.012127413414418697, + 0.013116497546434402, + 0.009654836729168892, + 0.012062686495482922, + 0.048591263592243195, + 0.009841260500252247, + 0.016943130642175674, + 0.00170708482619375, + 0.0011933846399188042, + 0.003381200134754181, + 0.0027084010653197765, + 0.0028438952285796404, + 0.002286065835505724, + 0.0021405061706900597, + 0.0017624308820813894, + 0.0015656378818675876, + 0.0012685286346822977, + 0.0031479147728532553, + 0.001959983492270112, + 0.0026428597047924995, + 0.0015492703532800078, + 0.0016833526315167546, + 0.0015324504347518086, + 0.0022179957013577223, + 0.001863158424384892, + 0.00620697345584631, + 0.0046160160563886166, + 0.004146606661379337, + 0.002868599258363247, + 0.0024001682177186012, + 0.002933749696239829, + 0.024631338194012642, + 0.013317901641130447, + 0.012388142757117748, + 0.009725628420710564, + 0.01047858688980341, + 0.03747817128896713, + 0.005829473491758108, + 0.010885155759751797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.08902692049741745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368968516588211 + ], + "opposite_logprob_tokens": [ + -3.4086897373199463 + ], + "logprob_reward": -0.03368968516588211, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368968516588211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16322626173496246, + 0.18785598874092102, + 0.1945469230413437, + 0.1384245753288269, + 0.02983253262937069, + 0.15222828090190887, + 0.018359452486038208, + 0.023834142833948135, + 0.015455427579581738, + 0.07623639702796936 + ], + "credit_map": [ + 0.015590852126479149, + 0.011334288865327835, + 0.036129504442214966, + 0.01918940804898739, + 0.020943833515048027, + 0.020839771255850792, + 0.014899970032274723, + 0.024298639968037605, + 0.04364313185214996, + 0.014591016806662083, + 0.015827946364879608, + 0.014089170843362808, + 0.010925610549747944, + 0.062100134789943695, + 0.00951290037482977, + 0.017166072502732277, + 0.04560898616909981, + 0.017462996765971184, + 0.014871541410684586, + 0.011749044060707092, + 0.01620825007557869, + 0.060464002192020416, + 0.01071823388338089, + 0.017463859170675278, + 0.030558912083506584, + 0.010965016670525074, + 0.012170204892754555, + 0.010176127776503563, + 0.010482395999133587, + 0.043466053903102875, + 0.006729471031576395, + 0.01387639157474041, + 0.002645724220201373, + 0.001997439656406641, + 0.0044622099958360195, + 0.004461999516934156, + 0.005486936308443546, + 0.0031334045343101025, + 0.004323549102991819, + 0.003321269294247031, + 0.03233363851904869, + 0.01172090508043766, + 0.01248221192508936, + 0.01128126960247755, + 0.011864489875733852, + 0.04835435375571251, + 0.0073956991545856, + 0.016795717179775238, + 0.003077215049415827, + 0.001460260828025639, + 0.0027769592124968767, + 0.0015241276705637574, + 0.0022801761515438557, + 0.003812578273937106, + 0.0016186233842745423, + 0.0018095109844580293, + 0.0022262069396674633, + 0.00140888593159616, + 0.007346282713115215, + 0.002512057079002261, + 0.002798920264467597, + 0.00279153767041862, + 0.002018334111198783, + 0.0027319181244820356, + 0.0014220940647646785, + 0.0010970537550747395, + 0.002981423167511821, + 0.00243331678211689, + 0.0022270241752266884, + 0.0017561294371262193, + 0.0017433635657653213, + 0.0017950220499187708, + 0.012844057753682137, + 0.006984658073633909, + 0.009126816876232624, + 0.007157447747886181, + 0.007877015508711338, + 0.019314885139465332, + 0.005956921726465225, + 0.006974590942263603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03368968516588211 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, futuristic train with a polished pearlescent composite shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at breakneck speed through a vast desert with rolling, reddish-orange sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at an extremely low angle, gliding just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, pulsating rays of cool white light visible outlining the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the train kick up a massive, turbulent plume of sand forming a swirling vortex in its wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337718591094017 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337718591094017, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337718591094017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3038134276866913, + 0.14472748339176178, + 0.21864570677280426, + 0.039560481905937195, + 0.022705400362610817, + 0.1642114222049713, + 0.0122505072504282, + 0.038986898958683014, + 0.02040308155119419, + 0.034695569425821304 + ], + "credit_map": [ + 0.06984956562519073, + 0.02804649993777275, + 0.035458631813526154, + 0.019805554300546646, + 0.01865129917860031, + 0.08469834923744202, + 0.01671121083199978, + 0.030592313036322594, + 0.028822101652622223, + 0.010586893185973167, + 0.018772166222333908, + 0.012362061999738216, + 0.011164691299200058, + 0.03983883187174797, + 0.00948173925280571, + 0.013699001632630825, + 0.046635642647743225, + 0.019376184791326523, + 0.026084506884217262, + 0.013992383144795895, + 0.019880441948771477, + 0.05888375639915466, + 0.014528069645166397, + 0.01926470547914505, + 0.00637803552672267, + 0.003228588495403528, + 0.009252733550965786, + 0.0026884586550295353, + 0.00408409908413887, + 0.00814269483089447, + 0.0022268593311309814, + 0.0035590108018368483, + 0.0023597117979079485, + 0.0014364166418090463, + 0.002715291688218713, + 0.0022028624080121517, + 0.004080506041646004, + 0.00419293949380517, + 0.0031665910501033068, + 0.002551081357523799, + 0.03683047369122505, + 0.013411028310656548, + 0.010984208434820175, + 0.009893616661429405, + 0.011863903142511845, + 0.053665388375520706, + 0.009908165782690048, + 0.01765464060008526, + 0.0015309317968785763, + 0.0008394381147809327, + 0.002864617621526122, + 0.001472504111006856, + 0.0017166229663416743, + 0.0011749588884413242, + 0.0012485020561143756, + 0.0014029319863766432, + 0.00373002327978611, + 0.002616453915834427, + 0.008414039388298988, + 0.00545957637950778, + 0.005329322535544634, + 0.0041863336227834225, + 0.004535970743745565, + 0.004715180490165949, + 0.0017996923997998238, + 0.0012471004156395793, + 0.003829586086794734, + 0.0030457302927970886, + 0.0032801623456180096, + 0.002772171050310135, + 0.002434616908431053, + 0.001994021702557802, + 0.006571551784873009, + 0.003238377394154668, + 0.0037960377521812916, + 0.003068571677431464, + 0.0026544714346528053, + 0.009226598776876926, + 0.0029726570937782526, + 0.003167305141687393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.026337718591094017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_tensor([1858, 1858])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_tensor([1858, 1858])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..94bd020908b6b0383d78fbf69d8476ffa0d5a590 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_tensor([1858, 1858])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, shot from a slightly overhead, level angle, slowly draws the viewer into a living room reimagined by imagination. A beam of warm afternoon sunlight, carrying with it a distinctly dancing dust mote, slants through an off-screen window, casting a bright, warm patch of light on the thick, soft beige wool carpet—the arena of battle. In the center of this vast, rolling hills, an epic duel unfolds: a hand-sewn gray plush mouse, its fabric-stitched eyes gleaming with determination. It lowers its body slightly, using a shiny bottle cap as a shield, its other paw clutching a broken toothpick as a sword, poised in a defensive stance. Its opponent, a colorful feathered staff, tipped with fluffy, sharp feathers, hovers in mid-air like an elegant knight wielding a lance, its shaft slicing a graceful arc, poised for a swift and deadly dive. In the background, the legs of the wooden coffee table and the corner of the sofa appear blurred and monumental in the shallow depth of field, resembling distant mountains and cliffs, further enhancing the grandeur of this microscopic battle. The entire scene is imbued with a sense of childlike fantasy and epic solemnity, with warm and soft tones. Every frame seems hand-painted, imbued with the texture and warmth of life, perfectly recreating the exquisite and delicate animation style of Hayao Miyazaki.", + "seed": 10772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_1858_seed_10772_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08658236600458621, + "vlm_logprob": -0.08658236600458621, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.08658236600458621, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20861651003360748 + ], + "opposite_logprob_tokens": [ + -1.7086164951324463 + ], + "logprob_reward": -0.20861651003360748, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20861651003360748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5075640678405762, + 0.09339337795972824, + 0.06505381315946579, + 0.04598408192396164, + 0.05631166696548462, + 0.08659783005714417, + 0.03462174907326698, + 0.05850960314273834, + 0.037109289318323135, + 0.014854544773697853 + ], + "credit_map": [ + 0.10062213242053986, + 0.026102563366293907, + 0.024876292794942856, + 0.02380121871829033, + 0.0633532702922821, + 0.15279795229434967, + 0.06485389918088913, + 0.05115674436092377, + 0.013690227642655373, + 0.012100137770175934, + 0.006726250518113375, + 0.012702059000730515, + 0.01807853765785694, + 0.012915256433188915, + 0.007826939225196838, + 0.009353972040116787, + 0.009872153401374817, + 0.008499703370034695, + 0.004428347107023001, + 0.005814547184854746, + 0.008229101076722145, + 0.013092274777591228, + 0.007031247019767761, + 0.008086441084742546, + 0.005421527661383152, + 0.007021127734333277, + 0.00361011759378016, + 0.006749663036316633, + 0.005796217359602451, + 0.004869080148637295, + 0.006104412488639355, + 0.006411935668438673, + 0.00991429015994072, + 0.005926358979195356, + 0.0059378789737820625, + 0.0051209693774580956, + 0.005178509745746851, + 0.012315227650105953, + 0.004056715872138739, + 0.00786171667277813, + 0.01447125431150198, + 0.006421868689358234, + 0.009356794878840446, + 0.006575587205588818, + 0.00828571803867817, + 0.025706114247441292, + 0.005983063951134682, + 0.009797426871955395, + 0.005440596956759691, + 0.003895984962582588, + 0.002845178823918104, + 0.004344013053923845, + 0.003919744398444891, + 0.006911437958478928, + 0.0036074884701520205, + 0.0036573067773133516, + 0.008315782994031906, + 0.008315574377775192, + 0.004395106807351112, + 0.003845350118353963, + 0.008456544950604439, + 0.009382312186062336, + 0.004011339973658323, + 0.011787586845457554, + 0.0075004929676651955, + 0.0029391625430434942, + 0.0032224729657173157, + 0.0027198109310120344, + 0.0033106189221143723, + 0.008493875153362751, + 0.00389173929579556, + 0.005031118169426918, + 0.0020559635013341904, + 0.0018332358449697495, + 0.0016650052275508642, + 0.0020195639226585627, + 0.0013395148562267423, + 0.0018614595755934715, + 0.0019600153900682926, + 0.0021197861060500145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.20861651003360748 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899638265371323 + ], + "opposite_logprob_tokens": [ + -2.578996419906616 + ], + "logprob_reward": -0.07899638265371323, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899638265371323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28893497586250305, + 0.1036081537604332, + 0.11735124886035919, + 0.09241003543138504, + 0.04833769053220749, + 0.060667041689157486, + 0.0559762604534626, + 0.07208404690027237, + 0.10531099140644073, + 0.05531957000494003 + ], + "credit_map": [ + 0.06292759627103806, + 0.02995910868048668, + 0.014176885597407818, + 0.01734703779220581, + 0.034017324447631836, + 0.08095517754554749, + 0.025443485006690025, + 0.024108389392495155, + 0.013114901259541512, + 0.01905892975628376, + 0.008851976133883, + 0.013595170341432095, + 0.017300967127084732, + 0.012533928267657757, + 0.008415556512773037, + 0.010736721567809582, + 0.017518188804388046, + 0.020648417994379997, + 0.006603674963116646, + 0.015946069732308388, + 0.013672326691448689, + 0.015885120257735252, + 0.0159748662263155, + 0.011102578602731228, + 0.011202631518244743, + 0.015522165223956108, + 0.006622158922255039, + 0.023669369518756866, + 0.009903778322041035, + 0.009942494332790375, + 0.008259153924882412, + 0.007288280874490738, + 0.006932172458618879, + 0.005437995307147503, + 0.005426053889095783, + 0.005980690475553274, + 0.006943436339497566, + 0.00497029721736908, + 0.004215112887322903, + 0.008431931026279926, + 0.008855657652020454, + 0.007619394455105066, + 0.005849051754921675, + 0.007132449187338352, + 0.00853003654628992, + 0.00990108773112297, + 0.005695562344044447, + 0.007083798758685589, + 0.008477121591567993, + 0.005769127979874611, + 0.004582011606544256, + 0.011741051450371742, + 0.004948882851749659, + 0.006596028804779053, + 0.005282538942992687, + 0.0085794972255826, + 0.009857689961791039, + 0.006895727477967739, + 0.006655512843281031, + 0.010136140510439873, + 0.006752585992217064, + 0.009319068863987923, + 0.008663159795105457, + 0.013804159127175808, + 0.017575468868017197, + 0.011404938995838165, + 0.010210239328444004, + 0.012149825692176819, + 0.009384398348629475, + 0.01655927672982216, + 0.012423215433955193, + 0.015603626146912575, + 0.008165049366652966, + 0.0075821285136044025, + 0.005503827705979347, + 0.006366552785038948, + 0.005871966481208801, + 0.006567946635186672, + 0.007749465294182301, + 0.007512636948376894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.07899638265371323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897710800170898 + ], + "opposite_logprob_tokens": [ + -2.578977108001709 + ], + "logprob_reward": -0.07897710800170898, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897710800170898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24614228308200836, + 0.08416227996349335, + 0.26933932304382324, + 0.06100809946656227, + 0.061819855123758316, + 0.044816385954618454, + 0.07327054440975189, + 0.046206749975681305, + 0.06667771935462952, + 0.046556755900382996 + ], + "credit_map": [ + 0.06082351133227348, + 0.016450954601168633, + 0.009517945349216461, + 0.011592499911785126, + 0.027128834277391434, + 0.07707955688238144, + 0.02365691028535366, + 0.019892076030373573, + 0.012070810422301292, + 0.013415948487818241, + 0.007977630943059921, + 0.01060689240694046, + 0.012289322912693024, + 0.011470353230834007, + 0.008368081413209438, + 0.00796323362737894, + 0.039944250136613846, + 0.05187629535794258, + 0.013750439509749413, + 0.03230428323149681, + 0.038991935551166534, + 0.04650142788887024, + 0.025395069271326065, + 0.020575597882270813, + 0.007021649740636349, + 0.008483326993882656, + 0.005766688846051693, + 0.017268864437937737, + 0.004027622751891613, + 0.006917678751051426, + 0.005862540565431118, + 0.005659725982695818, + 0.009138738736510277, + 0.00558792520314455, + 0.006567867938429117, + 0.012356474064290524, + 0.008292204700410366, + 0.0055209314450621605, + 0.006554469931870699, + 0.0078012398444116116, + 0.006798775400966406, + 0.005740583408623934, + 0.005119891371577978, + 0.004260353744029999, + 0.007050191052258015, + 0.006502940785139799, + 0.0037055034190416336, + 0.005638146307319403, + 0.0116919269785285, + 0.010481161065399647, + 0.005276077426970005, + 0.013432097621262074, + 0.007806998211890459, + 0.008866866119205952, + 0.005451834760606289, + 0.010263578966259956, + 0.006190224085003138, + 0.005459120962768793, + 0.005137572064995766, + 0.007032702211290598, + 0.00529642915353179, + 0.005521217826753855, + 0.004564952105283737, + 0.0070045338943600655, + 0.011734819039702415, + 0.006732783280313015, + 0.0064969658851623535, + 0.006711713504046202, + 0.005892103537917137, + 0.012916145846247673, + 0.005654545035213232, + 0.010538636706769466, + 0.005990440957248211, + 0.006921816151589155, + 0.005106635857373476, + 0.0073982528410851955, + 0.007732805795967579, + 0.004208249505609274, + 0.004966716282069683, + 0.0042318361811339855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.07897710800170898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284463211894035 + ], + "opposite_logprob_tokens": [ + -3.773284435272217 + ], + "logprob_reward": -0.023284463211894035, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284463211894035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29141390323638916, + 0.0766773372888565, + 0.04908694326877594, + 0.0873367115855217, + 0.26689550280570984, + 0.050835125148296356, + 0.020721323788166046, + 0.09428849071264267, + 0.03529009595513344, + 0.02745460905134678 + ], + "credit_map": [ + 0.05254856497049332, + 0.028157323598861694, + 0.02749284729361534, + 0.01690683700144291, + 0.041615843772888184, + 0.0677507296204567, + 0.031122736632823944, + 0.025819025933742523, + 0.009243322536349297, + 0.008409159258008003, + 0.01151652168482542, + 0.009046928025782108, + 0.013595832511782646, + 0.010069514624774456, + 0.008304083719849586, + 0.00649196607992053, + 0.007049822714179754, + 0.009279473684728146, + 0.0038793436251580715, + 0.004569149110466242, + 0.005925399251282215, + 0.009134071879088879, + 0.0035079997032880783, + 0.005741679109632969, + 0.012354211881756783, + 0.01047192420810461, + 0.008770734071731567, + 0.018924225121736526, + 0.007994979619979858, + 0.011009084060788155, + 0.007503352593630552, + 0.010308203287422657, + 0.04364970326423645, + 0.037906061857938766, + 0.03159284219145775, + 0.028519293293356895, + 0.02961164526641369, + 0.025941502302885056, + 0.024864915758371353, + 0.04480952396988869, + 0.00945613719522953, + 0.0047349948436021805, + 0.005188067443668842, + 0.003215877339243889, + 0.004845099523663521, + 0.013252855278551579, + 0.0034216074272990227, + 0.006720480974763632, + 0.0032580839470028877, + 0.002961053978651762, + 0.002260654466226697, + 0.0018520819721743464, + 0.0018998985178768635, + 0.0038922291714698076, + 0.0013452921994030476, + 0.0032520287204533815, + 0.012665342539548874, + 0.011599504388868809, + 0.010735641233623028, + 0.007838409394025803, + 0.015722015872597694, + 0.01141747273504734, + 0.007861562073230743, + 0.01644854061305523, + 0.005680479109287262, + 0.005214353557676077, + 0.0033337625209242105, + 0.0039185332134366035, + 0.003911216743290424, + 0.005704604554921389, + 0.0024886352475732565, + 0.0050385091453790665, + 0.0036387688014656305, + 0.004316522739827633, + 0.002244866220280528, + 0.003376137698069215, + 0.005974999628961086, + 0.0029313198756426573, + 0.002266588853672147, + 0.002705403370782733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.023284463211894035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303736612200737 + ], + "opposite_logprob_tokens": [ + -3.1680374145507812 + ], + "logprob_reward": -0.04303736612200737, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303736612200737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2126365453004837, + 0.10786242038011551, + 0.12894944846630096, + 0.12033171206712723, + 0.07050380855798721, + 0.0850161761045456, + 0.06209700554609299, + 0.0789448618888855, + 0.05508385971188545, + 0.07857420295476913 + ], + "credit_map": [ + 0.04181262105703354, + 0.023551788181066513, + 0.016435852274298668, + 0.021052593365311623, + 0.02450709417462349, + 0.036209866404533386, + 0.026417972519993782, + 0.022648749873042107, + 0.014730901457369328, + 0.012867975980043411, + 0.014101005159318447, + 0.01004460547119379, + 0.022328685969114304, + 0.011840326711535454, + 0.00981246680021286, + 0.012136449106037617, + 0.017169486731290817, + 0.020087795332074165, + 0.01138476375490427, + 0.017729394137859344, + 0.013379717245697975, + 0.022492120042443275, + 0.011921643279492855, + 0.014784518629312515, + 0.013979904353618622, + 0.010240718722343445, + 0.012426596134901047, + 0.02385559305548668, + 0.011714154854416847, + 0.012157410383224487, + 0.02373906411230564, + 0.012218262068927288, + 0.01097089983522892, + 0.011705833487212658, + 0.00816960446536541, + 0.005662679206579924, + 0.009459520690143108, + 0.009488319978117943, + 0.006537753157317638, + 0.008509190753102303, + 0.013056128285825253, + 0.008733696304261684, + 0.009292815811932087, + 0.008152678608894348, + 0.007855230011045933, + 0.019026603549718857, + 0.00625608628615737, + 0.012642927467823029, + 0.010104895569384098, + 0.007140821777284145, + 0.006004204507917166, + 0.004397801123559475, + 0.006504362914711237, + 0.011598790064454079, + 0.004678466822952032, + 0.011667657643556595, + 0.011604500003159046, + 0.007996753789484501, + 0.006332521326839924, + 0.007958765141665936, + 0.00963960587978363, + 0.01408242154866457, + 0.005408646538853645, + 0.015921643003821373, + 0.008155865594744682, + 0.007236616685986519, + 0.005263053346425295, + 0.004947544075548649, + 0.007712951395660639, + 0.008128659799695015, + 0.006047163158655167, + 0.0075920033268630505, + 0.010295955464243889, + 0.013712937943637371, + 0.005670920945703983, + 0.006421761587262154, + 0.019905485212802887, + 0.010298012755811214, + 0.0057057482190430164, + 0.006563376635313034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.04303736612200737 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a slightly overhead, level angle, slowly draws the viewer into a living room reimagined by imagination. A beam of warm afternoon sunlight, carrying with it a distinctly dancing dust mote, slants through an off-screen window, casting a bright, warm patch of light on the thick, soft beige wool carpet—the arena of battle. In the center of this vast, rolling hills, an epic duel unfolds: a hand-sewn gray plush mouse, its fabric-stitched eyes gleaming with determination. It lowers its body slightly, using a shiny bottle cap as a shield, its other paw clutching a broken toothpick as a sword, poised in a defensive stance. Its opponent, a colorful feathered staff, tipped with fluffy, sharp feathers, hovers in mid-air like an elegant knight wielding a lance, its shaft slicing a graceful arc, poised for a swift and deadly dive. In the background, the legs of the wooden coffee table and the corner of the sofa appear blurred and monumental in the shallow depth of field, resembling distant mountains and cliffs, further enhancing the grandeur of this microscopic battle. The entire scene is imbued with a sense of childlike fantasy and epic solemnity, with warm and soft tones. Every frame seems hand-painted, imbued with the texture and warmth of life, perfectly recreating the exquisite and delicate animation style of Hayao Miyazaki.", + "seed": 10773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_1858_seed_10773_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12933653295040132, + "vlm_logprob": -0.12933653295040132, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.12933653295040132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3988407850265503 + ], + "opposite_logprob_tokens": [ + -1.1488407850265503 + ], + "logprob_reward": -0.3988407850265503, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3988407850265503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3300534188747406, + 0.0974472165107727, + 0.053327277302742004, + 0.09672018140554428, + 0.05178796499967575, + 0.04912707209587097, + 0.15094219148159027, + 0.07429283857345581, + 0.04238751158118248, + 0.053914330899715424 + ], + "credit_map": [ + 0.045820772647857666, + 0.03667106479406357, + 0.03740708902478218, + 0.02231617085635662, + 0.03735816851258278, + 0.08260561525821686, + 0.026689598336815834, + 0.0411849282681942, + 0.01326250284910202, + 0.010270517319440842, + 0.01079039927572012, + 0.011566020548343658, + 0.00976476352661848, + 0.017136000096797943, + 0.00861101783812046, + 0.016045983880758286, + 0.01050726417452097, + 0.010292339138686657, + 0.004995838738977909, + 0.008568141609430313, + 0.004340888932347298, + 0.004978791810572147, + 0.004366742912679911, + 0.005277263466268778, + 0.021049482747912407, + 0.017136700451374054, + 0.011233012191951275, + 0.009882197715342045, + 0.00822755042463541, + 0.010909897275269032, + 0.009378106333315372, + 0.008903236128389835, + 0.00564128952100873, + 0.008055025711655617, + 0.010013160295784473, + 0.004296489059925079, + 0.005124058108776808, + 0.004474773537367582, + 0.003911138046532869, + 0.010272029787302017, + 0.007811177521944046, + 0.007497952785342932, + 0.0055852062068879604, + 0.004782131873071194, + 0.004252136219292879, + 0.004841516260057688, + 0.0062139020301401615, + 0.008143044076859951, + 0.020641909912228584, + 0.017643149942159653, + 0.02227529138326645, + 0.017498884350061417, + 0.023992549628019333, + 0.017665307968854904, + 0.013813421130180359, + 0.01741166226565838, + 0.010119901970028877, + 0.006630757823586464, + 0.012216582894325256, + 0.007667625322937965, + 0.009177013300359249, + 0.012928969226777554, + 0.006838890723884106, + 0.00871308520436287, + 0.006346362177282572, + 0.004106039647012949, + 0.006537623703479767, + 0.004861717578023672, + 0.004003927577286959, + 0.007412869017571211, + 0.004669881425797939, + 0.004449088592082262, + 0.005964959971606731, + 0.006659624166786671, + 0.006506807636469603, + 0.007385746575891972, + 0.00711047975346446, + 0.007528477814048529, + 0.007169300690293312, + 0.005588930565863848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.3988407850265503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981787547469139 + ], + "opposite_logprob_tokens": [ + -3.529817819595337 + ], + "logprob_reward": -0.02981787547469139, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981787547469139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4142577052116394, + 0.06869135797023773, + 0.11590724438428879, + 0.05753341689705849, + 0.0511634387075901, + 0.06491392850875854, + 0.07014670968055725, + 0.04216654226183891, + 0.08706092834472656, + 0.02815878577530384 + ], + "credit_map": [ + 0.0564204677939415, + 0.0396759957075119, + 0.04562859609723091, + 0.03396834060549736, + 0.05058955028653145, + 0.09594903886318207, + 0.0479402132332325, + 0.04408549144864082, + 0.009281089529395103, + 0.009935677982866764, + 0.006181835196912289, + 0.005414033308625221, + 0.012841194868087769, + 0.011002257466316223, + 0.006516446825116873, + 0.007518811151385307, + 0.014903002418577671, + 0.009421100839972496, + 0.009127354249358177, + 0.01547339465469122, + 0.015981541946530342, + 0.022632883861660957, + 0.01351363305002451, + 0.014854323118925095, + 0.008904386311769485, + 0.011084990575909615, + 0.006601236294955015, + 0.005978708155453205, + 0.0064384969882667065, + 0.0061601693741977215, + 0.007794961798936129, + 0.0045704650692641735, + 0.004529602825641632, + 0.004383438266813755, + 0.012605701573193073, + 0.004977673292160034, + 0.003009468549862504, + 0.004969150293618441, + 0.0034897569566965103, + 0.013198638334870338, + 0.00843733549118042, + 0.006552731152623892, + 0.005815379787236452, + 0.007351553998887539, + 0.0090941172093153, + 0.010304450057446957, + 0.010097873397171497, + 0.007260480895638466, + 0.00746174855157733, + 0.00755611015483737, + 0.009433690458536148, + 0.00772517453879118, + 0.013197487220168114, + 0.006270392797887325, + 0.010001451708376408, + 0.008500645868480206, + 0.005167381837964058, + 0.0030932105146348476, + 0.004143150988966227, + 0.003733824472874403, + 0.006941352970898151, + 0.005135438404977322, + 0.009816738776862621, + 0.004135440569370985, + 0.009450960904359818, + 0.007612866349518299, + 0.008265499025583267, + 0.008983532898128033, + 0.01232985407114029, + 0.01851358450949192, + 0.014309548772871494, + 0.007595079950988293, + 0.002539420034736395, + 0.0024097550194710493, + 0.004175439011305571, + 0.004793799947947264, + 0.0040733702480793, + 0.0024163671769201756, + 0.0038860784843564034, + 0.003864555386826396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.02981787547469139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294214427471161 + ], + "opposite_logprob_tokens": [ + -2.2379422187805176 + ], + "logprob_reward": -0.11294214427471161, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294214427471161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3506712317466736, + 0.07293613255023956, + 0.07707463949918747, + 0.09319888055324554, + 0.088503897190094, + 0.09875776618719101, + 0.10334709286689758, + 0.022572260349988937, + 0.07289615273475647, + 0.02004196122288704 + ], + "credit_map": [ + 0.044161561876535416, + 0.03026457689702511, + 0.036418262869119644, + 0.031528279185295105, + 0.04773646965622902, + 0.07851926982402802, + 0.047214969992637634, + 0.0348278246819973, + 0.016209250316023827, + 0.009830707684159279, + 0.007024608552455902, + 0.005650269333273172, + 0.008681530132889748, + 0.01128728874027729, + 0.006373321637511253, + 0.00787916500121355, + 0.011375288479030132, + 0.011082069016993046, + 0.00706066656857729, + 0.007020012475550175, + 0.01455242745578289, + 0.011402292177081108, + 0.00847953837364912, + 0.006102338898926973, + 0.018924478441476822, + 0.01450918335467577, + 0.010002762079238892, + 0.007878528907895088, + 0.012755480594933033, + 0.012385891750454903, + 0.007203629706054926, + 0.00953892432153225, + 0.007250108290463686, + 0.006356155499815941, + 0.01596185192465782, + 0.006866790354251862, + 0.00788886658847332, + 0.013088379055261612, + 0.013537077233195305, + 0.017554663121700287, + 0.018900586292147636, + 0.013285276480019093, + 0.010613122954964638, + 0.009808175265789032, + 0.007587645668536425, + 0.01901930198073387, + 0.00665705930441618, + 0.01288660429418087, + 0.014809119515120983, + 0.012798214331269264, + 0.010368118062615395, + 0.010256868787109852, + 0.020184602588415146, + 0.01023673266172409, + 0.010609637945890427, + 0.014083796180784702, + 0.0032817383762449026, + 0.001658739405684173, + 0.0017766719684004784, + 0.0023504288401454687, + 0.003651463659480214, + 0.002992532914504409, + 0.004439914133399725, + 0.002420771401375532, + 0.009568128734827042, + 0.006187248509377241, + 0.010216223075985909, + 0.01062716357409954, + 0.008234882727265358, + 0.014803416095674038, + 0.0059128026477992535, + 0.007346286438405514, + 0.0020272338297218084, + 0.0022648468147963285, + 0.0034603746607899666, + 0.0035532088950276375, + 0.0026997567620128393, + 0.0013398566516116261, + 0.002662639133632183, + 0.002034044126048684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.11294214427471161 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204346567392349 + ], + "opposite_logprob_tokens": [ + -2.8120434284210205 + ], + "logprob_reward": -0.06204346567392349, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204346567392349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31051626801490784, + 0.09654441475868225, + 0.30153876543045044, + 0.04967862367630005, + 0.02475518174469471, + 0.05418916791677475, + 0.06829801946878433, + 0.04021274670958519, + 0.03235134854912758, + 0.021915439516305923 + ], + "credit_map": [ + 0.0396759919822216, + 0.02791287750005722, + 0.03437520191073418, + 0.03224656358361244, + 0.04831962287425995, + 0.05572591722011566, + 0.042100775986909866, + 0.030159326270222664, + 0.01898197829723358, + 0.011385680176317692, + 0.009843380190432072, + 0.008089499548077583, + 0.010767374187707901, + 0.02189226634800434, + 0.005216553341597319, + 0.010367684997618198, + 0.07401628792285919, + 0.08371607959270477, + 0.02330540120601654, + 0.02669587917625904, + 0.02436905913054943, + 0.03587273508310318, + 0.014756807126104832, + 0.01880650967359543, + 0.012620145455002785, + 0.007328378967940807, + 0.005695863626897335, + 0.004339238163083792, + 0.0052467468194663525, + 0.005269031971693039, + 0.0036284385714679956, + 0.005550778936594725, + 0.0035429242998361588, + 0.00402351189404726, + 0.002524129580706358, + 0.002479634480550885, + 0.0017450673040002584, + 0.003894230118021369, + 0.004251846112310886, + 0.0022938356269150972, + 0.009856905788183212, + 0.0073835295625030994, + 0.005825251340866089, + 0.004986080806702375, + 0.003745059948414564, + 0.004906534217298031, + 0.00865767989307642, + 0.008828123100101948, + 0.009439468383789062, + 0.006665934342890978, + 0.010262587107717991, + 0.007799804210662842, + 0.009003653191030025, + 0.008951522409915924, + 0.008887942880392075, + 0.007287104148417711, + 0.005936398170888424, + 0.0032651007641106844, + 0.005873420275747776, + 0.00431952765211463, + 0.004479425959289074, + 0.0069100637920200825, + 0.0050187669694423676, + 0.004410044755786657, + 0.004537311382591724, + 0.003542043501511216, + 0.00483822263777256, + 0.004823697730898857, + 0.0030256719328463078, + 0.004126038867980242, + 0.0037201608065515757, + 0.003738203551620245, + 0.0027030848432332277, + 0.0026943269185721874, + 0.0022732289507985115, + 0.0027915143873542547, + 0.003402539761736989, + 0.002656643744558096, + 0.0031666357535868883, + 0.0022274667862802744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.06204346567392349 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043038394302129745 + ], + "opposite_logprob_tokens": [ + -3.1680383682250977 + ], + "logprob_reward": -0.043038394302129745, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043038394302129745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5330411791801453, + 0.08259086310863495, + 0.052869297564029694, + 0.06253601610660553, + 0.05269843712449074, + 0.04052220657467842, + 0.05972236022353172, + 0.036486394703388214, + 0.04695742577314377, + 0.03257578983902931 + ], + "credit_map": [ + 0.07156344503164291, + 0.050022225826978683, + 0.059054456651210785, + 0.04233897849917412, + 0.05959990248084068, + 0.13520851731300354, + 0.0604996494948864, + 0.054754018783569336, + 0.01111247856169939, + 0.01218855008482933, + 0.00787108950316906, + 0.007647808641195297, + 0.010082547552883625, + 0.016424277797341347, + 0.006333337165415287, + 0.010930776596069336, + 0.009858843870460987, + 0.008303712122142315, + 0.005062181502580643, + 0.0076536983251571655, + 0.0066758510656654835, + 0.004788314923644066, + 0.005772273056209087, + 0.004754422698169947, + 0.010268590413033962, + 0.00935015082359314, + 0.008364187553524971, + 0.007893896661698818, + 0.006249480415135622, + 0.007518442813307047, + 0.007007431238889694, + 0.005883839912712574, + 0.004752892069518566, + 0.0037103856448084116, + 0.013871821574866772, + 0.0034962077625095844, + 0.0037984533701092005, + 0.004621542524546385, + 0.003480056067928672, + 0.014967075549066067, + 0.00608915975317359, + 0.00572317885234952, + 0.004761474207043648, + 0.0038612468633800745, + 0.004967051558196545, + 0.004178398288786411, + 0.004781882744282484, + 0.006159814074635506, + 0.008742059580981731, + 0.006259469781070948, + 0.005261353682726622, + 0.00457517895847559, + 0.009025697596371174, + 0.011210906319320202, + 0.007839024998247623, + 0.006808667443692684, + 0.004904563073068857, + 0.0031043486669659615, + 0.005479053128510714, + 0.004376330878585577, + 0.004441578406840563, + 0.005186315160244703, + 0.005371300969272852, + 0.0036229060497134924, + 0.006445040460675955, + 0.004227487836033106, + 0.004649867303669453, + 0.005792996846139431, + 0.005310071632266045, + 0.010935899801552296, + 0.00478267390280962, + 0.004813385661691427, + 0.004738382529467344, + 0.005126462318003178, + 0.00454128161072731, + 0.004247245844453573, + 0.0033007082529366016, + 0.0034927567467093468, + 0.004147490952163935, + 0.0029814639128744602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.043038394302129745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a slightly overhead, level angle, slowly draws the viewer into a living room reimagined by imagination. A beam of warm afternoon sunlight, carrying with it a distinctly dancing dust mote, slants through an off-screen window, casting a bright, warm patch of light on the thick, soft beige wool carpet—the arena of battle. In the center of this vast, rolling hills, an epic duel unfolds: a hand-sewn gray plush mouse, its fabric-stitched eyes gleaming with determination. It lowers its body slightly, using a shiny bottle cap as a shield, its other paw clutching a broken toothpick as a sword, poised in a defensive stance. Its opponent, a colorful feathered staff, tipped with fluffy, sharp feathers, hovers in mid-air like an elegant knight wielding a lance, its shaft slicing a graceful arc, poised for a swift and deadly dive. In the background, the legs of the wooden coffee table and the corner of the sofa appear blurred and monumental in the shallow depth of field, resembling distant mountains and cliffs, further enhancing the grandeur of this microscopic battle. The entire scene is imbued with a sense of childlike fantasy and epic solemnity, with warm and soft tones. Every frame seems hand-painted, imbued with the texture and warmth of life, perfectly recreating the exquisite and delicate animation style of Hayao Miyazaki.", + "seed": 10774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_1858_seed_10774_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2146565154194832, + "vlm_logprob": -0.2146565154194832, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.2146565154194832, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8998947143554688 + ], + "opposite_logprob_tokens": [ + -0.5248947143554688 + ], + "logprob_reward": -0.8998947143554688, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8998947143554688, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.38373154401779175, + 0.055377330631017685, + 0.04152189567685127, + 0.050106145441532135, + 0.11216196417808533, + 0.18063771724700928, + 0.1080809235572815, + 0.016644645482301712, + 0.024187181144952774, + 0.027550658211112022 + ], + "credit_map": [ + 0.02977007068693638, + 0.06252485513687134, + 0.03263280913233757, + 0.039710912853479385, + 0.07422970235347748, + 0.05552401393651962, + 0.061017271131277084, + 0.028321893885731697, + 0.004682584200054407, + 0.006544000934809446, + 0.009504837915301323, + 0.004023158457130194, + 0.009292618371546268, + 0.010710102505981922, + 0.004960097372531891, + 0.005659932736307383, + 0.00488433800637722, + 0.007272171322256327, + 0.006012797821313143, + 0.0035565842408686876, + 0.004641459323465824, + 0.005319520831108093, + 0.0038465005345642567, + 0.005988524295389652, + 0.006589053198695183, + 0.008992486633360386, + 0.006590807344764471, + 0.0045067560859024525, + 0.005467110313475132, + 0.00789813045412302, + 0.005700142588466406, + 0.0043616583570837975, + 0.010734191164374352, + 0.016474364325404167, + 0.013881536200642586, + 0.007731225807219744, + 0.017785895615816116, + 0.02338903211057186, + 0.010531243868172169, + 0.011634470894932747, + 0.01806643232703209, + 0.004188432823866606, + 0.04918888583779335, + 0.00854260940104723, + 0.019562041386961937, + 0.0656195729970932, + 0.005633257795125246, + 0.009836491197347641, + 0.010102779604494572, + 0.0030981588643044233, + 0.026796724647283554, + 0.004766987636685371, + 0.013602674007415771, + 0.037085648626089096, + 0.006683343555778265, + 0.005944611504673958, + 0.002181745134294033, + 0.0013942610239610076, + 0.0019781156443059444, + 0.00145390210673213, + 0.0028566012624651194, + 0.0027150933165103197, + 0.0027158695738762617, + 0.0013490577694028616, + 0.0035822484642267227, + 0.002180409850552678, + 0.002598794410005212, + 0.0020114819053560495, + 0.004930926952511072, + 0.0028003614861518145, + 0.0035612951032817364, + 0.0025216632056981325, + 0.0043569207191467285, + 0.0018192637944594026, + 0.002680794335901737, + 0.002247418509796262, + 0.005435233935713768, + 0.003576484741643071, + 0.004207581747323275, + 0.003226959379389882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.8998947143554688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371746465563774 + ], + "opposite_logprob_tokens": [ + -3.408717393875122 + ], + "logprob_reward": -0.03371746465563774, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371746465563774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28912830352783203, + 0.1457352489233017, + 0.0774451270699501, + 0.08856832981109619, + 0.0826815515756607, + 0.22327682375907898, + 0.013900267891585827, + 0.05017832666635513, + 0.018097344785928726, + 0.010988698340952396 + ], + "credit_map": [ + 0.020125217735767365, + 0.03776586800813675, + 0.02687114290893078, + 0.039809875190258026, + 0.043496787548065186, + 0.039444077759981155, + 0.04964730143547058, + 0.03196800872683525, + 0.012362142093479633, + 0.018660973757505417, + 0.023197030648589134, + 0.014484754763543606, + 0.011490562930703163, + 0.026012752205133438, + 0.022803883999586105, + 0.016723131760954857, + 0.008504198864102364, + 0.012595833279192448, + 0.011361971497535706, + 0.006039019208401442, + 0.011000520549714565, + 0.014450130052864552, + 0.007209057454019785, + 0.006284390576183796, + 0.010410605929791927, + 0.012041318230330944, + 0.012704478576779366, + 0.009486407041549683, + 0.010230793617665768, + 0.012581483460962772, + 0.013725160621106625, + 0.0073880781419575214, + 0.010720822028815746, + 0.00652132136747241, + 0.009169071912765503, + 0.007412095088511705, + 0.010639345273375511, + 0.018758714199066162, + 0.011792910285294056, + 0.0076672667637467384, + 0.021251220256090164, + 0.005600159056484699, + 0.06226992607116699, + 0.01142707746475935, + 0.02335895597934723, + 0.07710322737693787, + 0.008580836467444897, + 0.013685386627912521, + 0.0016736978432163596, + 0.0011341019999235868, + 0.0012992126867175102, + 0.0014293183339759707, + 0.0027438460383564234, + 0.002006390830501914, + 0.0015102113829925656, + 0.0021034874953329563, + 0.008154654875397682, + 0.0046070171520113945, + 0.004841554444283247, + 0.0038132835179567337, + 0.011320448480546474, + 0.007273874711245298, + 0.00554991140961647, + 0.004617577418684959, + 0.002164737321436405, + 0.002615013625472784, + 0.0019721342250704765, + 0.0018241307698190212, + 0.0028927235398441553, + 0.002582929562777281, + 0.0026445011608302593, + 0.0014011709718033671, + 0.0013319646241143346, + 0.0013584860134869814, + 0.0011387054109945893, + 0.0010939689818769693, + 0.001990181626752019, + 0.001278161071240902, + 0.0016644964925944805, + 0.0011327322572469711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03371746465563774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536060065031052 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012536060065031052, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536060065031052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08055984228849411, + 0.03599542751908302, + 0.18116246163845062, + 0.09554766863584518, + 0.15681147575378418, + 0.3139042556285858, + 0.047160182148218155, + 0.0481443889439106, + 0.02832558937370777, + 0.012388714589178562 + ], + "credit_map": [ + 0.007363492157310247, + 0.014952261932194233, + 0.0064844004809856415, + 0.0062726824544370174, + 0.021578693762421608, + 0.011948389001190662, + 0.007465892005711794, + 0.004494031425565481, + 0.0026269867084920406, + 0.004325831774622202, + 0.004422337748110294, + 0.0036546685732901096, + 0.004441695753484964, + 0.007169220130890608, + 0.0062056672759354115, + 0.003149025607854128, + 0.01868225261569023, + 0.036703161895275116, + 0.025296470150351524, + 0.013892297632992268, + 0.022804250940680504, + 0.03410036489367485, + 0.014646518044173717, + 0.01503714919090271, + 0.011233304627239704, + 0.012363401241600513, + 0.015983348712325096, + 0.006865256000310183, + 0.01104107964783907, + 0.019836202263832092, + 0.009919993579387665, + 0.008305090479552746, + 0.015611407347023487, + 0.022328048944473267, + 0.019348228350281715, + 0.011347035877406597, + 0.02469228208065033, + 0.033525701612234116, + 0.014158105477690697, + 0.01580067165195942, + 0.031214719638228416, + 0.007481777109205723, + 0.08582557737827301, + 0.014852665364742279, + 0.033895526081323624, + 0.11387645453214645, + 0.009834611788392067, + 0.01692293770611286, + 0.004270846024155617, + 0.0014615529216825962, + 0.01253803912550211, + 0.002037202473729849, + 0.004980502184480429, + 0.017099862918257713, + 0.0024338962975889444, + 0.0023382825311273336, + 0.007308370899409056, + 0.004118958488106728, + 0.004125121980905533, + 0.004231579601764679, + 0.010313920676708221, + 0.00856559444218874, + 0.006241781637072563, + 0.0032390658743679523, + 0.0034149158746004105, + 0.0035235974937677383, + 0.003077722853049636, + 0.0025081862695515156, + 0.003898149821907282, + 0.006777637172490358, + 0.003660708898678422, + 0.0014646750641986728, + 0.0016282731667160988, + 0.001968618482351303, + 0.0014926603762432933, + 0.0015065694460645318, + 0.001560691394843161, + 0.0019451990956440568, + 0.0013980047078803182, + 0.000888699316419661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.012536060065031052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905080705881119 + ], + "opposite_logprob_tokens": [ + -2.464050769805908 + ], + "logprob_reward": -0.08905080705881119, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905080705881119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19491541385650635, + 0.02489968203008175, + 0.17045453190803528, + 0.02401420846581459, + 0.017091231420636177, + 0.42415618896484375, + 0.036019131541252136, + 0.04949258267879486, + 0.04813283681869507, + 0.010824190452694893 + ], + "credit_map": [ + 0.01516732107847929, + 0.02983299270272255, + 0.020622966811060905, + 0.02595346048474312, + 0.038887713104486465, + 0.02526474930346012, + 0.026378832757472992, + 0.012807376682758331, + 0.002167442115023732, + 0.0027656482998281717, + 0.0028618641663342714, + 0.0030597199220210314, + 0.0031814787071198225, + 0.003907893318682909, + 0.004575790837407112, + 0.002379846293479204, + 0.020337071269750595, + 0.03720536455512047, + 0.018069207668304443, + 0.012735982425510883, + 0.024094846099615097, + 0.027987508103251457, + 0.013413132168352604, + 0.01661142334342003, + 0.0028315752279013395, + 0.0036659634206444025, + 0.0032616613898426294, + 0.00251549924723804, + 0.0032528892625123262, + 0.003501820843666792, + 0.0029567049350589514, + 0.0020280955359339714, + 0.0015963696641847491, + 0.0028857102151960135, + 0.002148958155885339, + 0.0017244284972548485, + 0.002153010806068778, + 0.0021515425760298967, + 0.0027121054008603096, + 0.0017191069200634956, + 0.0426233746111393, + 0.009503012523055077, + 0.11445526033639908, + 0.020039966329932213, + 0.046369072049856186, + 0.15584398806095123, + 0.01334167830646038, + 0.021979838609695435, + 0.004137871321290731, + 0.003172065829858184, + 0.0037905999924987555, + 0.0033410745672881603, + 0.00718836160376668, + 0.005649249535053968, + 0.004031640011817217, + 0.004708268214017153, + 0.0048874784260988235, + 0.0036810801830142736, + 0.006759071256965399, + 0.006247557699680328, + 0.007553745526820421, + 0.007924439385533333, + 0.006310681346803904, + 0.006128527224063873, + 0.0060912263579666615, + 0.005995025392621756, + 0.005411253310739994, + 0.003621427109465003, + 0.007391864899545908, + 0.012017526663839817, + 0.004858681000769138, + 0.0027458348777145147, + 0.0012248913990333676, + 0.0010384846245869994, + 0.00124186126049608, + 0.0009685674449428916, + 0.0020027956925332546, + 0.0011939193354919553, + 0.0019802225288003683, + 0.0011734486324712634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.08905080705881119 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038083530962467194 + ], + "opposite_logprob_tokens": [ + -3.288083553314209 + ], + "logprob_reward": -0.038083530962467194, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038083530962467194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1044011190533638, + 0.043817684054374695, + 0.0785594955086708, + 0.19939370453357697, + 0.22375324368476868, + 0.2698362469673157, + 0.01062089391052723, + 0.024087507277727127, + 0.03093673475086689, + 0.014593357220292091 + ], + "credit_map": [ + 0.007139534689486027, + 0.015644805505871773, + 0.013052846305072308, + 0.011322700418531895, + 0.021547283977270126, + 0.015173719264566898, + 0.01347644068300724, + 0.007043775636702776, + 0.00482653733342886, + 0.005977250635623932, + 0.005341280251741409, + 0.0039838640950620174, + 0.005429502110928297, + 0.007853461429476738, + 0.005917514208704233, + 0.004488269332796335, + 0.007310682442039251, + 0.012581613846123219, + 0.009761261753737926, + 0.00770933274179697, + 0.012710027396678925, + 0.015639066696166992, + 0.007686604280024767, + 0.005160897970199585, + 0.02124338410794735, + 0.02617923729121685, + 0.03985041007399559, + 0.01473967730998993, + 0.02284083142876625, + 0.04225337505340576, + 0.015450254082679749, + 0.01683652587234974, + 0.022276876494288445, + 0.03129839524626732, + 0.02797795459628105, + 0.016305962577462196, + 0.035148464143276215, + 0.047585658729076385, + 0.020200015977025032, + 0.022959904745221138, + 0.027421820908784866, + 0.0059427921660244465, + 0.0723782330751419, + 0.012786288745701313, + 0.02950485609471798, + 0.09954482316970825, + 0.008499492891132832, + 0.013757920823991299, + 0.0012427186593413353, + 0.0010692549403756857, + 0.0013354986440390348, + 0.0011445311829447746, + 0.0017738108290359378, + 0.001502691418863833, + 0.0013571541057899594, + 0.001195233198814094, + 0.003591324668377638, + 0.0022488117683678865, + 0.002492789411917329, + 0.0021277356427162886, + 0.005380266811698675, + 0.0036263654474169016, + 0.0025412351824343204, + 0.0020789785776287317, + 0.003595719812437892, + 0.00394238717854023, + 0.003137524938210845, + 0.002641690196469426, + 0.005601324141025543, + 0.006329565774649382, + 0.0036951000802218914, + 0.0019934182055294514, + 0.0017643027240410447, + 0.0020541767589747906, + 0.0016058130422607064, + 0.0013282138388603926, + 0.0019682180136442184, + 0.002794070402160287, + 0.0022010686807334423, + 0.0008774924790486693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.038083530962467194 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a slightly overhead, level angle, slowly draws the viewer into a living room reimagined by imagination. A beam of warm afternoon sunlight, carrying with it a distinctly dancing dust mote, slants through an off-screen window, casting a bright, warm patch of light on the thick, soft beige wool carpet—the arena of battle. In the center of this vast, rolling hills, an epic duel unfolds: a hand-sewn gray plush mouse, its fabric-stitched eyes gleaming with determination. It lowers its body slightly, using a shiny bottle cap as a shield, its other paw clutching a broken toothpick as a sword, poised in a defensive stance. Its opponent, a colorful feathered staff, tipped with fluffy, sharp feathers, hovers in mid-air like an elegant knight wielding a lance, its shaft slicing a graceful arc, poised for a swift and deadly dive. In the background, the legs of the wooden coffee table and the corner of the sofa appear blurred and monumental in the shallow depth of field, resembling distant mountains and cliffs, further enhancing the grandeur of this microscopic battle. The entire scene is imbued with a sense of childlike fantasy and epic solemnity, with warm and soft tones. Every frame seems hand-painted, imbued with the texture and warmth of life, perfectly recreating the exquisite and delicate animation style of Hayao Miyazaki.", + "seed": 10775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_1/index_1858_seed_10775_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3189780872315168, + "vlm_logprob": -0.3189780872315168, + "vlm_yes_no_margin": 2.025, + "vlm_token_credit": -0.3189780872315168, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.314631700515747 + ], + "opposite_logprob_tokens": [ + -0.3146316409111023 + ], + "logprob_reward": -1.314631700515747, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.314631700515747, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.352950781583786, + 0.10015343129634857, + 0.06700384616851807, + 0.031243832781910896, + 0.06143692135810852, + 0.0661989152431488, + 0.14323294162750244, + 0.05216221138834953, + 0.07166194170713425, + 0.05395511910319328 + ], + "credit_map": [ + 0.043549392372369766, + 0.06647682189941406, + 0.05329229310154915, + 0.04584725201129913, + 0.04010681062936783, + 0.027702776715159416, + 0.03920786455273628, + 0.03676756098866463, + 0.0098771583288908, + 0.020449047908186913, + 0.014192693866789341, + 0.01183504331856966, + 0.009628662839531898, + 0.006179118063300848, + 0.016911761835217476, + 0.011079948395490646, + 0.008413898758590221, + 0.010816940106451511, + 0.009046408347785473, + 0.007720645051449537, + 0.012085403315722942, + 0.008815798908472061, + 0.005232656374573708, + 0.004872096702456474, + 0.0034387013874948025, + 0.004414945840835571, + 0.0036723650991916656, + 0.0025747730396687984, + 0.004286485258489847, + 0.006897745653986931, + 0.0029745656065642834, + 0.0029842506628483534, + 0.009940284304320812, + 0.013373149558901787, + 0.007943416945636272, + 0.009409057907760143, + 0.00454515079036355, + 0.004709240980446339, + 0.004027631599456072, + 0.007488987408578396, + 0.008190724067389965, + 0.011459028348326683, + 0.00875459797680378, + 0.009136089123785496, + 0.007810602895915508, + 0.005917015951126814, + 0.0061646923422813416, + 0.00876616220921278, + 0.011675751768052578, + 0.010600434616208076, + 0.038089901208877563, + 0.04434226453304291, + 0.006795587949454784, + 0.008616711013019085, + 0.005503651686012745, + 0.0176086463034153, + 0.007747779134660959, + 0.014240523800253868, + 0.009007318876683712, + 0.0065665473230183125, + 0.003853963688015938, + 0.00392200518399477, + 0.002679193392395973, + 0.004144878126680851, + 0.010818498209118843, + 0.01495440024882555, + 0.00952620804309845, + 0.010782423429191113, + 0.008323108777403831, + 0.007211846765130758, + 0.005376373417675495, + 0.0046690842136740685, + 0.009168841876089573, + 0.01024833507835865, + 0.005601570941507816, + 0.006214048247784376, + 0.006777993403375149, + 0.005657559260725975, + 0.005969719961285591, + 0.004317053128033876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -1.314631700515747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297738552093506 + ], + "opposite_logprob_tokens": [ + -2.2379775047302246 + ], + "logprob_reward": -0.11297738552093506, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297738552093506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30539849400520325, + 0.03328756242990494, + 0.1096184030175209, + 0.02139180898666382, + 0.0947532206773758, + 0.14998076856136322, + 0.06988977640867233, + 0.08451695740222931, + 0.09109437465667725, + 0.040068693459033966 + ], + "credit_map": [ + 0.04139845818281174, + 0.062119510024785995, + 0.04844974726438522, + 0.03867003321647644, + 0.0285196490585804, + 0.01982242241501808, + 0.033620916306972504, + 0.03279774263501167, + 0.0037984035443514585, + 0.006790374405682087, + 0.00418474804610014, + 0.004583865869790316, + 0.0040110438130795956, + 0.002613537013530731, + 0.004282949026674032, + 0.0030226404778659344, + 0.010708559304475784, + 0.0134770218282938, + 0.014337210915982723, + 0.00917139183729887, + 0.019581735134124756, + 0.015023991465568542, + 0.017322342842817307, + 0.009996145032346249, + 0.0028086737729609013, + 0.003887818194925785, + 0.002563049318268895, + 0.002950941678136587, + 0.0022410808596760035, + 0.002511790255084634, + 0.0017472635954618454, + 0.0026811910793185234, + 0.008071760646998882, + 0.008889945223927498, + 0.014826658181846142, + 0.008266497403383255, + 0.007934377528727055, + 0.02988346293568611, + 0.00478679034858942, + 0.012093723751604557, + 0.012461076490581036, + 0.010782705619931221, + 0.02664163149893284, + 0.012587152421474457, + 0.01692250929772854, + 0.04134473204612732, + 0.012068582698702812, + 0.01717236451804638, + 0.004981298930943012, + 0.00723248440772295, + 0.01117160078138113, + 0.013904248364269733, + 0.011114642955362797, + 0.006496655289083719, + 0.0075294263660907745, + 0.007459417451173067, + 0.007029119413346052, + 0.0166470967233181, + 0.00943613052368164, + 0.00895635224878788, + 0.014323449693620205, + 0.008808303624391556, + 0.009914160706102848, + 0.009402344934642315, + 0.014183142222464085, + 0.019536204636096954, + 0.011336316354572773, + 0.012726115062832832, + 0.013007648289203644, + 0.009083583019673824, + 0.0055114044807851315, + 0.005709960125386715, + 0.005888358224183321, + 0.005416767671704292, + 0.004454659763723612, + 0.004188167862594128, + 0.006354250013828278, + 0.0049918233416974545, + 0.0044853040017187595, + 0.004289364442229271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.11297738552093506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905265480279922 + ], + "opposite_logprob_tokens": [ + -2.464052677154541 + ], + "logprob_reward": -0.08905265480279922, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905265480279922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11730095744132996, + 0.04441003128886223, + 0.09925227612257004, + 0.04174170643091202, + 0.16905167698860168, + 0.16447144746780396, + 0.17187263071537018, + 0.1277148574590683, + 0.046679381281137466, + 0.01750504970550537 + ], + "credit_map": [ + 0.01717250607907772, + 0.018912291154265404, + 0.019207796081900597, + 0.01412779651582241, + 0.010107699781656265, + 0.007843698374927044, + 0.018394265323877335, + 0.011534901335835457, + 0.005547274369746447, + 0.008119549602270126, + 0.005355894099920988, + 0.004766369704157114, + 0.004016423597931862, + 0.0031450476963073015, + 0.007631910499185324, + 0.005827560555189848, + 0.012802465818822384, + 0.012126361951231956, + 0.01557197980582714, + 0.009729364886879921, + 0.014038910157978535, + 0.011069334112107754, + 0.015152079053223133, + 0.008761787787079811, + 0.00470119109377265, + 0.006132233887910843, + 0.00537468446418643, + 0.004441923461854458, + 0.003976929467171431, + 0.0066188033670187, + 0.006045432295650244, + 0.004450507927685976, + 0.015748148784041405, + 0.015489282086491585, + 0.026820361614227295, + 0.013029021210968494, + 0.01722385175526142, + 0.05213439092040062, + 0.009946073405444622, + 0.018660547211766243, + 0.015015332028269768, + 0.012224700301885605, + 0.029920700937509537, + 0.013352169655263424, + 0.019404305145144463, + 0.04429928958415985, + 0.011999132111668587, + 0.018255826085805893, + 0.01409587450325489, + 0.009369966574013233, + 0.056635912507772446, + 0.04673337563872337, + 0.01398876029998064, + 0.010715872049331665, + 0.009143790230154991, + 0.011189082637429237, + 0.016208399087190628, + 0.027139004319906235, + 0.022921938449144363, + 0.018258528783917427, + 0.01073510479182005, + 0.010902024805545807, + 0.011532101780176163, + 0.01001774799078703, + 0.00891718827188015, + 0.008997988887131214, + 0.0059350235387682915, + 0.00683069322258234, + 0.004617855418473482, + 0.004306980408728123, + 0.002926930785179138, + 0.00414672028273344, + 0.0024218098260462284, + 0.003050381550565362, + 0.0020136304665356874, + 0.0018924508476629853, + 0.002614263677969575, + 0.0017995741218328476, + 0.002501417649909854, + 0.0012115214485675097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.08905265480279922 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290986195206642 + ], + "opposite_logprob_tokens": [ + -3.7732908725738525 + ], + "logprob_reward": -0.023290986195206642, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290986195206642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16068364679813385, + 0.058209795504808426, + 0.3352431356906891, + 0.215358704328537, + 0.031157108023762703, + 0.02480159141123295, + 0.050135232508182526, + 0.040186744183301926, + 0.06811647117137909, + 0.01610754430294037 + ], + "credit_map": [ + 0.019952351227402687, + 0.029499346390366554, + 0.02337426319718361, + 0.02009899728000164, + 0.016374940052628517, + 0.013482450507581234, + 0.02035593055188656, + 0.01754537597298622, + 0.005855039693415165, + 0.014245201833546162, + 0.007605404127389193, + 0.008440850302577019, + 0.005681511014699936, + 0.0036905405577272177, + 0.008036153391003609, + 0.004655094351619482, + 0.04999921843409538, + 0.07382415235042572, + 0.057704485952854156, + 0.04866206273436546, + 0.0273452028632164, + 0.03103126771748066, + 0.024039801210165024, + 0.02263694256544113, + 0.034050796180963516, + 0.05062687769532204, + 0.03229806572198868, + 0.03178701177239418, + 0.0161746758967638, + 0.016399487853050232, + 0.012423230335116386, + 0.021598560735583305, + 0.003103010356426239, + 0.005282587371766567, + 0.00419623451307416, + 0.003946634940803051, + 0.00335390935651958, + 0.005648711230605841, + 0.0022304446902126074, + 0.0033955767285078764, + 0.0026091067120432854, + 0.003527070162817836, + 0.0038290333468466997, + 0.003070061095058918, + 0.0035662795417010784, + 0.0028603672981262207, + 0.002673760987818241, + 0.002665911801159382, + 0.004002516157925129, + 0.004501320421695709, + 0.018761441111564636, + 0.0076827360317111015, + 0.002508097793906927, + 0.002830653451383114, + 0.0017992286011576653, + 0.008049236610531807, + 0.005089324899017811, + 0.011397999711334705, + 0.006557098589837551, + 0.00483090803027153, + 0.003491391194984317, + 0.0031655593775212765, + 0.002254902385175228, + 0.003399562556296587, + 0.010973633266985416, + 0.01344740018248558, + 0.008624907582998276, + 0.010435683652758598, + 0.007949559949338436, + 0.006515254732221365, + 0.005303209647536278, + 0.004866821225732565, + 0.0016379822045564651, + 0.0024009665939956903, + 0.0021205737721174955, + 0.002586603397503495, + 0.0023664566688239574, + 0.0016171042807400227, + 0.001644441857933998, + 0.0017334152944386005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.023290986195206642 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a living room featuring a thick, soft beige wool carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hand-sewn gray plush mouse visible, using a shiny bottle cap as a shield and a broken toothpick as a sword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a colorful feathered staff with fluffy, sharp feathers hovering in mid-air like a knight wielding a lance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm afternoon sunlight with a dancing dust mote slanting through an off-screen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring blurred legs of a wooden coffee table and a sofa corner resembling distant mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054937709122896194 + ], + "opposite_logprob_tokens": [ + -2.9299376010894775 + ], + "logprob_reward": -0.054937709122896194, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054937709122896194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2849236726760864, + 0.13816703855991364, + 0.03939111530780792, + 0.04230509698390961, + 0.03983946889638901, + 0.06254575401544571, + 0.06905742734670639, + 0.1475471407175064, + 0.09483389556407928, + 0.08138936758041382 + ], + "credit_map": [ + 0.03992787003517151, + 0.05662865191698074, + 0.04153939709067345, + 0.035397861152887344, + 0.029709836468100548, + 0.020882930606603622, + 0.031008610501885414, + 0.029828514903783798, + 0.014516119845211506, + 0.022804250940680504, + 0.0205155648291111, + 0.01942623034119606, + 0.015525822527706623, + 0.010003264993429184, + 0.01885438710451126, + 0.016521399840712547, + 0.004977526143193245, + 0.0069853756576776505, + 0.004834623076021671, + 0.004743464291095734, + 0.005502632819116116, + 0.004480146337300539, + 0.0040152231231331825, + 0.0038521201349794865, + 0.0065527427941560745, + 0.006854530423879623, + 0.0055326735600829124, + 0.006207426078617573, + 0.0037114801816642284, + 0.0056797959841787815, + 0.002767567289993167, + 0.004998878110200167, + 0.004504082724452019, + 0.009170610457658768, + 0.005149045959115028, + 0.005465182941406965, + 0.0030726930126547813, + 0.0049167051911354065, + 0.0031595572363585234, + 0.004401588346809149, + 0.006327226758003235, + 0.006742691621184349, + 0.010950267314910889, + 0.007588254753500223, + 0.007192900404334068, + 0.012964732013642788, + 0.003604222321882844, + 0.007175451144576073, + 0.005019725300371647, + 0.0044891186989843845, + 0.02558846026659012, + 0.013177154585719109, + 0.004674315452575684, + 0.003734805854037404, + 0.004085782449692488, + 0.008288060314953327, + 0.01783665269613266, + 0.04649752750992775, + 0.022963441908359528, + 0.018582846969366074, + 0.011557981371879578, + 0.012004676274955273, + 0.007639188319444656, + 0.010464818216860294, + 0.014522525481879711, + 0.019179318100214005, + 0.011882919818162918, + 0.013840213418006897, + 0.011519440449774265, + 0.00877875741571188, + 0.00828908197581768, + 0.006821631453931332, + 0.012727566063404083, + 0.017014959827065468, + 0.00829347874969244, + 0.0111387362703681, + 0.010666019283235073, + 0.006770221050828695, + 0.00813748687505722, + 0.006640892941504717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.054937709122896194 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_tensor([1238, 1238])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_tensor([1238, 1238])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7b95f540f9487ee165918c2fd55429902cc20eb2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_tensor([1238, 1238])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up shot from a slightly overhead angle captures a pair of steady, deft hands, absorbed in the work of a meticulously detailed artist. One hand firmly grips a classic yellow hexagonal wooden pencil, while the other, with steady and deliberate control, manipulates a finely crafted utility knife with a cold metallic sheen. The sharp tip of the blade meticulously scrapes away layer by layer from the fragile graphite lead, sending fine dust drifting onto the dark wooden work surface with each precise cut. The previously intact lead has been carved into an incredible miniature structure—a chain of interconnected, hair-thin links, their intricate form clearly visible in the lens. A bright desk lamp, shining from above and to the side, casts concentrated light, not only sharp highlights on the metal blade but also intricate shadows on the carved links, enhancing the work's three-dimensional quality. The entire painting is permeated with an atmosphere of extreme concentration and almost breathless tranquility, highlighting the artist's incredible skills and patience.", + "seed": 5423, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_1238_seed_5423_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07686522044241428, + "vlm_logprob": -0.07686522044241428, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.07686522044241428, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02361435629427433 + ], + "opposite_logprob_tokens": [ + -4.398614406585693 + ], + "logprob_reward": -0.02361435629427433, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02361435629427433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2985752522945404, + 0.1333593726158142, + 0.09195573627948761, + 0.05166845768690109, + 0.11393576860427856, + 0.041446443647146225, + 0.12711220979690552, + 0.043839484453201294, + 0.06764101982116699, + 0.03046625852584839 + ], + "credit_map": [ + 0.05142272636294365, + 0.026010997593402863, + 0.0349920317530632, + 0.02706136368215084, + 0.04071526974439621, + 0.046454619616270065, + 0.04935257509350777, + 0.022565675899386406, + 0.0148509182035923, + 0.0164120402187109, + 0.016133006662130356, + 0.015660511329770088, + 0.019164349883794785, + 0.018806342035531998, + 0.021424338221549988, + 0.010907856747508049, + 0.015941275283694267, + 0.0064059095457196236, + 0.01342804916203022, + 0.008313240483403206, + 0.009900003671646118, + 0.016691962257027626, + 0.010602529160678387, + 0.010672769509255886, + 0.008315548300743103, + 0.005033830180764198, + 0.007222163025289774, + 0.005435682367533445, + 0.0062135979533195496, + 0.008231088519096375, + 0.006511390674859285, + 0.004705156199634075, + 0.01727735623717308, + 0.010818392038345337, + 0.017779164016246796, + 0.01325428206473589, + 0.011399597860872746, + 0.01641763001680374, + 0.016427557915449142, + 0.010561788454651833, + 0.0073196133598685265, + 0.004880208522081375, + 0.008801653981208801, + 0.0057691289111971855, + 0.0037674924824386835, + 0.0029137711971998215, + 0.004927115049213171, + 0.003067461308091879, + 0.016784893348813057, + 0.010949413292109966, + 0.01572996936738491, + 0.01712055131793022, + 0.013328821398317814, + 0.010186413303017616, + 0.03033561259508133, + 0.012676531448960304, + 0.004447926301509142, + 0.003853459842503071, + 0.005495268385857344, + 0.005639185197651386, + 0.005288425367325544, + 0.004204799421131611, + 0.010568387806415558, + 0.004342035856097937, + 0.00948183611035347, + 0.006733456160873175, + 0.010840504430234432, + 0.010129204019904137, + 0.006126631982624531, + 0.009594502858817577, + 0.00983731634914875, + 0.004897573031485081, + 0.004575286991894245, + 0.0022154527250677347, + 0.005093964282423258, + 0.0034574863966554403, + 0.0035574063658714294, + 0.004304141737520695, + 0.003822706639766693, + 0.003439816413447261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02361435629427433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223271049559116 + ], + "opposite_logprob_tokens": [ + -4.264223098754883 + ], + "logprob_reward": -0.014223271049559116, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223271049559116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13558778166770935, + 0.23498767614364624, + 0.1264144629240036, + 0.04759275168180466, + 0.14941440522670746, + 0.07937851548194885, + 0.05780923739075661, + 0.060568917542696, + 0.0395718552172184, + 0.06867444515228271 + ], + "credit_map": [ + 0.0200604647397995, + 0.010306202806532383, + 0.01706770434975624, + 0.013203293085098267, + 0.026559889316558838, + 0.019455119967460632, + 0.01871819980442524, + 0.010216906666755676, + 0.036613982170820236, + 0.022224225103855133, + 0.03673597425222397, + 0.0192855354398489, + 0.037070438265800476, + 0.04286949336528778, + 0.019626202061772346, + 0.02056181989610195, + 0.01731519028544426, + 0.009921654127538204, + 0.01634765788912773, + 0.010353240184485912, + 0.01681293547153473, + 0.026350082829594612, + 0.013717216439545155, + 0.015596486628055573, + 0.008331086486577988, + 0.0038345549255609512, + 0.010119350627064705, + 0.005749959498643875, + 0.00403272919356823, + 0.005057210568338633, + 0.0064778937958180904, + 0.003989964723587036, + 0.03863102197647095, + 0.00885085016489029, + 0.043027397245168686, + 0.012567252852022648, + 0.00982845202088356, + 0.010613755322992802, + 0.01564229652285576, + 0.010253393091261387, + 0.02464623749256134, + 0.00407680869102478, + 0.02724592201411724, + 0.0048921238631010056, + 0.004572203848510981, + 0.003479165956377983, + 0.005584991071373224, + 0.004881061147898436, + 0.008366372436285019, + 0.003735278034582734, + 0.007114011328667402, + 0.005488527473062277, + 0.00899579469114542, + 0.008266102522611618, + 0.008209743537008762, + 0.007633404806256294, + 0.008335215039551258, + 0.004034638870507479, + 0.007822503335773945, + 0.005406023468822241, + 0.007980790920555592, + 0.012094796635210514, + 0.005873558577150106, + 0.009021389298141003, + 0.006509431637823582, + 0.003495482262223959, + 0.00783629436045885, + 0.004064880311489105, + 0.0033824697602540255, + 0.005402261856943369, + 0.004763699136674404, + 0.004117332398891449, + 0.012519275769591331, + 0.005431131459772587, + 0.019002696499228477, + 0.007034735754132271, + 0.006476104725152254, + 0.004636230878531933, + 0.009083552286028862, + 0.0044907135888934135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.014223271049559116 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633562870323658 + ], + "opposite_logprob_tokens": [ + -3.6513357162475586 + ], + "logprob_reward": -0.02633562870323658, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633562870323658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18705525994300842, + 0.21273159980773926, + 0.10573038458824158, + 0.060790594667196274, + 0.18407461047172546, + 0.02975274994969368, + 0.12465165555477142, + 0.03709392994642258, + 0.028433669358491898, + 0.029685506597161293 + ], + "credit_map": [ + 0.03085211291909218, + 0.01699858158826828, + 0.03742673620581627, + 0.017693208530545235, + 0.021846242249011993, + 0.027125345543026924, + 0.02139478363096714, + 0.013718256726861, + 0.03097529523074627, + 0.03356141224503517, + 0.02747650071978569, + 0.03197921812534332, + 0.018756313249468803, + 0.02556140162050724, + 0.029650496318936348, + 0.01477098185569048, + 0.018402637913823128, + 0.012406954541802406, + 0.021876169368624687, + 0.01030031405389309, + 0.010358215309679508, + 0.011723835952579975, + 0.01318210456520319, + 0.007480159867554903, + 0.01102506648749113, + 0.005455951206386089, + 0.015722298994660378, + 0.007474270183593035, + 0.0038825967349112034, + 0.005035026930272579, + 0.007760117761790752, + 0.004435273353010416, + 0.05226021632552147, + 0.011583831161260605, + 0.06100088357925415, + 0.014149417169392109, + 0.009101765230298042, + 0.008804806508123875, + 0.018313590437173843, + 0.008860114961862564, + 0.004000716842710972, + 0.0022564386017620564, + 0.004263273905962706, + 0.0026156301610171795, + 0.004206742160022259, + 0.005896218586713076, + 0.003140628570690751, + 0.003373105078935623, + 0.014902553521096706, + 0.01642714999616146, + 0.015530685894191265, + 0.01776212453842163, + 0.013646895065903664, + 0.018164921551942825, + 0.01624775491654873, + 0.01196957379579544, + 0.004782455507665873, + 0.0026420201174914837, + 0.003887129481881857, + 0.003384870709851384, + 0.003959108144044876, + 0.005255307070910931, + 0.008361070416867733, + 0.004821969661861658, + 0.004069499205797911, + 0.003200391074642539, + 0.003823344362899661, + 0.005400404799729586, + 0.002210537437349558, + 0.0020096555817872286, + 0.00535696791484952, + 0.002362870378419757, + 0.005358821712434292, + 0.0022626023273915052, + 0.007760568521916866, + 0.0031141715589910746, + 0.002891400596126914, + 0.0020658764988183975, + 0.0041787587106227875, + 0.0020533085335046053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.02633562870323658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732163019478321 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732163019478321, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732163019478321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13000330328941345, + 0.08480270951986313, + 0.07122327387332916, + 0.0914742648601532, + 0.33248239755630493, + 0.05168117955327034, + 0.12386874109506607, + 0.03202613443136215, + 0.03678407892584801, + 0.04565386101603508 + ], + "credit_map": [ + 0.027346480637788773, + 0.012960896827280521, + 0.02471122145652771, + 0.01502758078277111, + 0.011494944803416729, + 0.013911545276641846, + 0.014912239275872707, + 0.009638400748372078, + 0.01617794670164585, + 0.008114083670079708, + 0.018010852858424187, + 0.009180532768368721, + 0.011401332914829254, + 0.009630953893065453, + 0.007216223515570164, + 0.005070788785815239, + 0.013853976503014565, + 0.007353500928729773, + 0.01673467829823494, + 0.009247572161257267, + 0.005098072811961174, + 0.005071707535535097, + 0.008093535900115967, + 0.00577022647485137, + 0.023902904242277145, + 0.007445212919265032, + 0.026033079251646996, + 0.011664225719869137, + 0.005537156946957111, + 0.007529832422733307, + 0.004903844092041254, + 0.004458016715943813, + 0.101105235517025, + 0.021505672484636307, + 0.11682718247175217, + 0.021868474781513214, + 0.012498906813561916, + 0.01655108854174614, + 0.027923982590436935, + 0.0142018748447299, + 0.007249515503644943, + 0.004456625785678625, + 0.00988801009953022, + 0.00547951040789485, + 0.005803724285215139, + 0.0081306342035532, + 0.004455978516489267, + 0.006217180751264095, + 0.014787624590098858, + 0.012760316953063011, + 0.014437824487686157, + 0.01965794712305069, + 0.012650436721742153, + 0.01574021950364113, + 0.02304137870669365, + 0.010792992077767849, + 0.004118209704756737, + 0.0027513494715094566, + 0.004297837149351835, + 0.004065059125423431, + 0.003177944105118513, + 0.00463096471503377, + 0.0055555310100317, + 0.0034292410127818584, + 0.007290204055607319, + 0.003033473389223218, + 0.009403080679476261, + 0.005067136138677597, + 0.002291041426360607, + 0.0022458822932094336, + 0.005011890083551407, + 0.002441373886540532, + 0.008183459751307964, + 0.003464927664026618, + 0.012512966059148312, + 0.004703683312982321, + 0.0044789700768888, + 0.003863015677779913, + 0.005422396119683981, + 0.0030244430527091026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.006732163019478321 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342068314552307 + ], + "opposite_logprob_tokens": [ + -1.3134206533432007 + ], + "logprob_reward": -0.31342068314552307, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342068314552307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21837759017944336, + 0.10741882771253586, + 0.22216926515102386, + 0.08046770095825195, + 0.08358817547559738, + 0.10772939026355743, + 0.04077945277094841, + 0.06830596923828125, + 0.05566718056797981, + 0.015496510080993176 + ], + "credit_map": [ + 0.03504783660173416, + 0.013911597430706024, + 0.03278995677828789, + 0.014734169468283653, + 0.03559447079896927, + 0.04065574333071709, + 0.026187730953097343, + 0.01945606991648674, + 0.017440907657146454, + 0.008543355390429497, + 0.013376245275139809, + 0.011206855066120625, + 0.015620670281350613, + 0.018940763548016548, + 0.012006213888525963, + 0.010283815674483776, + 0.03583213686943054, + 0.018073875457048416, + 0.04079350084066391, + 0.018788576126098633, + 0.031639646738767624, + 0.02414129115641117, + 0.030715255066752434, + 0.022184977307915688, + 0.02111208066344261, + 0.00506561528891325, + 0.026708852499723434, + 0.004304577596485615, + 0.00568908266723156, + 0.007618449162691832, + 0.004414256662130356, + 0.0055547733791172504, + 0.02559317648410797, + 0.004621192812919617, + 0.02795066125690937, + 0.005164835602045059, + 0.004653090611100197, + 0.005254105664789677, + 0.005893699824810028, + 0.0044574011117219925, + 0.023150650784373283, + 0.007769627030938864, + 0.020700346678495407, + 0.00812102947384119, + 0.010063438676297665, + 0.013407877646386623, + 0.015441307798027992, + 0.009075098671019077, + 0.005135036539286375, + 0.0030137926805764437, + 0.004895650316029787, + 0.003119708737358451, + 0.005840051919221878, + 0.008262697607278824, + 0.005572430323809385, + 0.004940084181725979, + 0.009152304381132126, + 0.005461917724460363, + 0.013459631241858006, + 0.007835855707526207, + 0.005985171068459749, + 0.006828220561146736, + 0.01208295114338398, + 0.00749990763142705, + 0.007980207912623882, + 0.0053013041615486145, + 0.008585152216255665, + 0.008411593735218048, + 0.005706844385713339, + 0.006101657636463642, + 0.009226803667843342, + 0.004353612195700407, + 0.003129248507320881, + 0.0012182360514998436, + 0.0030893811490386724, + 0.0017167377518489957, + 0.0016949954442679882, + 0.0018445103196427226, + 0.001558826770633459, + 0.0012445731554180384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.31342068314552307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle captures a pair of steady, deft hands, absorbed in the work of a meticulously detailed artist. One hand firmly grips a classic yellow hexagonal wooden pencil, while the other, with steady and deliberate control, manipulates a finely crafted utility knife with a cold metallic sheen. The sharp tip of the blade meticulously scrapes away layer by layer from the fragile graphite lead, sending fine dust drifting onto the dark wooden work surface with each precise cut. The previously intact lead has been carved into an incredible miniature structure—a chain of interconnected, hair-thin links, their intricate form clearly visible in the lens. A bright desk lamp, shining from above and to the side, casts concentrated light, not only sharp highlights on the metal blade but also intricate shadows on the carved links, enhancing the work's three-dimensional quality. The entire painting is permeated with an atmosphere of extreme concentration and almost breathless tranquility, highlighting the artist's incredible skills and patience.", + "seed": 5424, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_1238_seed_5424_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05741906063631177, + "vlm_logprob": -0.05741906063631177, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.05741906063631177, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016345882788300514 + ], + "opposite_logprob_tokens": [ + -5.266345977783203 + ], + "logprob_reward": -0.016345882788300514, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.016345882788300514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19341638684272766, + 0.14104510843753815, + 0.11534267663955688, + 0.14393536746501923, + 0.031609855592250824, + 0.030293073505163193, + 0.04116109758615494, + 0.09176253527402878, + 0.17371027171611786, + 0.03772363439202309 + ], + "credit_map": [ + 0.011717810295522213, + 0.02728968858718872, + 0.03226993605494499, + 0.026831382885575294, + 0.02180258370935917, + 0.014681309461593628, + 0.02723041921854019, + 0.03159325569868088, + 0.013346970081329346, + 0.017679832875728607, + 0.034463461488485336, + 0.019427387043833733, + 0.015263795852661133, + 0.009297116659581661, + 0.011943905614316463, + 0.019622642546892166, + 0.006719328463077545, + 0.018058491870760918, + 0.020830081775784492, + 0.01391635276377201, + 0.012400573119521141, + 0.008093702606856823, + 0.011731848120689392, + 0.02359228953719139, + 0.008756628260016441, + 0.029232926666736603, + 0.02743717096745968, + 0.020692234858870506, + 0.015644485130906105, + 0.01030879095196724, + 0.016976777464151382, + 0.014886369928717613, + 0.0022606176789849997, + 0.0046457694843411446, + 0.004012544639408588, + 0.0032282338943332434, + 0.0023194211535155773, + 0.004769267980009317, + 0.003048634622246027, + 0.0073253680020570755, + 0.0020175445824861526, + 0.006026584655046463, + 0.004508905578404665, + 0.003971743397414684, + 0.003058344591408968, + 0.0033542935270816088, + 0.0024132162798196077, + 0.004942439030855894, + 0.0023941081017255783, + 0.005564627703279257, + 0.0061613949947059155, + 0.008311298675835133, + 0.004839894361793995, + 0.0030302314553409815, + 0.005772157106548548, + 0.0050873830914497375, + 0.0069568646140396595, + 0.021570436656475067, + 0.01466214470565319, + 0.01052700262516737, + 0.013107170350849628, + 0.0068861679174005985, + 0.009547206573188305, + 0.008505546487867832, + 0.012402171269059181, + 0.031253959983587265, + 0.031170280650258064, + 0.015206032432615757, + 0.021227043122053146, + 0.01932164467871189, + 0.01812255196273327, + 0.02500658854842186, + 0.003215807257220149, + 0.005415159277617931, + 0.006036826875060797, + 0.006357187405228615, + 0.0030963774770498276, + 0.0038045989349484444, + 0.005650343839079142, + 0.0041473302990198135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.016345882788300514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036360598169267178 + ], + "opposite_logprob_tokens": [ + -5.628635883331299 + ], + "logprob_reward": -0.0036360598169267178, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036360598169267178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12071553617715836, + 0.09510015696287155, + 0.1803124099969864, + 0.15940120816230774, + 0.12200705707073212, + 0.0356692336499691, + 0.07292796671390533, + 0.06044263765215874, + 0.08178559690713882, + 0.07163824141025543 + ], + "credit_map": [ + 0.008929119445383549, + 0.01929323747754097, + 0.022858086973428726, + 0.015415729954838753, + 0.01328812725841999, + 0.008803715929389, + 0.01581687666475773, + 0.016310641542077065, + 0.007300643716007471, + 0.014008358120918274, + 0.01487575564533472, + 0.009844853542745113, + 0.01093421969562769, + 0.009951691143214703, + 0.010479084216058254, + 0.017705552279949188, + 0.010900517925620079, + 0.035810209810733795, + 0.04929332062602043, + 0.017986999824643135, + 0.016199257224798203, + 0.011735565029084682, + 0.012415740638971329, + 0.025970805436372757, + 0.009109694510698318, + 0.03619472682476044, + 0.0397619903087616, + 0.015719586983323097, + 0.01496630348265171, + 0.0115954065695405, + 0.013378504663705826, + 0.018674995750188828, + 0.007891752757132053, + 0.019252723082900047, + 0.024203618988394737, + 0.010852315463125706, + 0.012558904476463795, + 0.015276489779353142, + 0.01293307263404131, + 0.01903817430138588, + 0.002252441830933094, + 0.006574451457709074, + 0.007685611955821514, + 0.0024411361664533615, + 0.00371852470561862, + 0.004501763731241226, + 0.003333935048431158, + 0.005161369685083628, + 0.004606241825968027, + 0.012587004341185093, + 0.01229518186300993, + 0.006014400161802769, + 0.008195440284907818, + 0.009075365029275417, + 0.009558525867760181, + 0.010595807805657387, + 0.0038610410410910845, + 0.010519037954509258, + 0.011014124378561974, + 0.0053654620423913, + 0.008292268961668015, + 0.006784615572541952, + 0.007239651866257191, + 0.0073664309456944466, + 0.004598635248839855, + 0.01790180243551731, + 0.010638212785124779, + 0.005721922963857651, + 0.012096407823264599, + 0.00858774222433567, + 0.00888703390955925, + 0.013353842310607433, + 0.004942237865179777, + 0.012262868694961071, + 0.016533641144633293, + 0.007085913326591253, + 0.006692815106362104, + 0.007964784279465675, + 0.0065408190712332726, + 0.009615160524845123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.0036360598169267178 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247984081506729 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005247984081506729, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247984081506729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34668612480163574, + 0.22358739376068115, + 0.14556364715099335, + 0.05331052094697952, + 0.060284119099378586, + 0.01883883774280548, + 0.04425930231809616, + 0.021918226033449173, + 0.03823990374803543, + 0.04731198400259018 + ], + "credit_map": [ + 0.022504959255456924, + 0.07498226314783096, + 0.10547348111867905, + 0.030710909515619278, + 0.030265720561146736, + 0.0159770455211401, + 0.02909545972943306, + 0.037676285952329636, + 0.018700355663895607, + 0.02754237689077854, + 0.05079777538776398, + 0.03801831230521202, + 0.020045340061187744, + 0.01129164919257164, + 0.01973726414144039, + 0.03745431825518608, + 0.008810203522443771, + 0.023441527038812637, + 0.02933892048895359, + 0.024051237851381302, + 0.013649282045662403, + 0.006898917723447084, + 0.01322162989526987, + 0.026151936501264572, + 0.004581972490996122, + 0.00903369765728712, + 0.007575924042612314, + 0.010277886874973774, + 0.0046584466472268105, + 0.005976030603051186, + 0.004588867537677288, + 0.0066176955588161945, + 0.003622675547376275, + 0.009614274837076664, + 0.009603066369891167, + 0.008653967641294003, + 0.007951172068715096, + 0.007569394540041685, + 0.006430900190025568, + 0.006838666275143623, + 0.0014953033532947302, + 0.0027970410883426666, + 0.0027026417665183544, + 0.0022891417611390352, + 0.0017921526450663805, + 0.0016486399108543992, + 0.002187547041103244, + 0.0039263698272407055, + 0.0030800048261880875, + 0.008079824037849903, + 0.010000831447541714, + 0.004266023635864258, + 0.005671257618814707, + 0.003399054752662778, + 0.004266498144716024, + 0.005495809484273195, + 0.0019855061545968056, + 0.004806737881153822, + 0.0032083417754620314, + 0.002214731415733695, + 0.00242629274725914, + 0.002332843141630292, + 0.0026005259715020657, + 0.0023432469461113214, + 0.0029502392280846834, + 0.005829798057675362, + 0.006762651726603508, + 0.0043144626542925835, + 0.0034923229832202196, + 0.004475418012589216, + 0.004716208670288324, + 0.005698801949620247, + 0.0030273394659161568, + 0.008638832718133926, + 0.007569524925202131, + 0.005293507128953934, + 0.004680593963712454, + 0.006778444163501263, + 0.003764685709029436, + 0.007559057325124741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005247984081506729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777157567441463 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009777157567441463, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777157567441463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2771593928337097, + 0.2426426112651825, + 0.10372608155012131, + 0.10348553210496902, + 0.053024716675281525, + 0.029679207131266594, + 0.019640130922198296, + 0.05820746347308159, + 0.043979041278362274, + 0.06845588237047195 + ], + "credit_map": [ + 0.023656465113162994, + 0.0379737913608551, + 0.04818856716156006, + 0.04338535666465759, + 0.027199292555451393, + 0.014980186708271503, + 0.04809608682990074, + 0.03367964178323746, + 0.012423071078956127, + 0.058739639818668365, + 0.061296187341213226, + 0.03448006510734558, + 0.01760975457727909, + 0.007756843231618404, + 0.029740402474999428, + 0.02059663087129593, + 0.006233366206288338, + 0.01987273059785366, + 0.016512706875801086, + 0.016977224498987198, + 0.009005332365632057, + 0.004844815470278263, + 0.015521220862865448, + 0.01475868746638298, + 0.005843901541084051, + 0.025922080501914024, + 0.026112575083971024, + 0.011170031502842903, + 0.010122058913111687, + 0.005842581856995821, + 0.009290535002946854, + 0.009181767702102661, + 0.0035872701555490494, + 0.007877960801124573, + 0.01148180104792118, + 0.006670328788459301, + 0.004583515226840973, + 0.003858675714582205, + 0.005956142675131559, + 0.009009020403027534, + 0.0021666812244802713, + 0.0058089508675038815, + 0.0044172401539981365, + 0.004310572054237127, + 0.0021236196625977755, + 0.002557467669248581, + 0.0032115052454173565, + 0.005083169788122177, + 0.0018304386176168919, + 0.0030935234390199184, + 0.0030129356309771538, + 0.0026173351798206568, + 0.0015086977509781718, + 0.0014674641424790025, + 0.0023421389050781727, + 0.003767597721889615, + 0.004376229364424944, + 0.0064453561790287495, + 0.00857514888048172, + 0.012699201703071594, + 0.0055558644235134125, + 0.004336799960583448, + 0.009645836427807808, + 0.0065730237402021885, + 0.0030734671745449305, + 0.009655858390033245, + 0.005979187786579132, + 0.003200183855369687, + 0.0055417511612176895, + 0.005536792799830437, + 0.004953973926603794, + 0.0060378266498446465, + 0.004857473075389862, + 0.013661458157002926, + 0.014546033926308155, + 0.006854763720184565, + 0.007809954695403576, + 0.006483651697635651, + 0.00541150476783514, + 0.008831038139760494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009777157567441463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520882189273834 + ], + "opposite_logprob_tokens": [ + -1.502088189125061 + ], + "logprob_reward": -0.2520882189273834, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520882189273834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3433850407600403, + 0.28030791878700256, + 0.03758945316076279, + 0.04386097937822342, + 0.07184147089719772, + 0.03722213953733444, + 0.023118769749999046, + 0.08824057132005692, + 0.05836955085396767, + 0.016064101830124855 + ], + "credit_map": [ + 0.03219703957438469, + 0.056866277009248734, + 0.11392059922218323, + 0.05005629360675812, + 0.025767169892787933, + 0.013344010338187218, + 0.01915736123919487, + 0.03207629546523094, + 0.018096456304192543, + 0.07052064687013626, + 0.08864090591669083, + 0.02467488870024681, + 0.01720254123210907, + 0.016853122040629387, + 0.011813914403319359, + 0.03250548243522644, + 0.00296120997518301, + 0.005845802836120129, + 0.008499089628458023, + 0.0033025986049324274, + 0.002310764277353883, + 0.0042914096266031265, + 0.002706871833652258, + 0.007671711500734091, + 0.0028399541042745113, + 0.008917894214391708, + 0.009121001698076725, + 0.006040486507117748, + 0.004080995451658964, + 0.0029911817982792854, + 0.003922349773347378, + 0.005947118625044823, + 0.0044060456566512585, + 0.010630283504724503, + 0.011964944191277027, + 0.008436532691121101, + 0.008044875226914883, + 0.007494253572076559, + 0.008480940014123917, + 0.012383606284856796, + 0.00269110850058496, + 0.006052282638847828, + 0.008531889878213406, + 0.0029740522149950266, + 0.0027852796483784914, + 0.0031672418117523193, + 0.003544293576851487, + 0.0074759963899850845, + 0.0014088451862335205, + 0.0032057317439466715, + 0.0030005653388798237, + 0.003229452297091484, + 0.002078669611364603, + 0.0029241859447211027, + 0.0024517634883522987, + 0.004819560330361128, + 0.006876688916236162, + 0.018928704783320427, + 0.016162898391485214, + 0.00947577878832817, + 0.010560469701886177, + 0.007996056228876114, + 0.008389340713620186, + 0.009850643575191498, + 0.003654165891930461, + 0.010148362256586552, + 0.010279180482029915, + 0.008901392109692097, + 0.005828782916069031, + 0.005860394798219204, + 0.006412973161786795, + 0.007284303195774555, + 0.001010821433737874, + 0.0025840604212135077, + 0.00207143253646791, + 0.0025552078150212765, + 0.001696108840405941, + 0.0015973852714523673, + 0.002461523050442338, + 0.002087564207613468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.2520882189273834 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle captures a pair of steady, deft hands, absorbed in the work of a meticulously detailed artist. One hand firmly grips a classic yellow hexagonal wooden pencil, while the other, with steady and deliberate control, manipulates a finely crafted utility knife with a cold metallic sheen. The sharp tip of the blade meticulously scrapes away layer by layer from the fragile graphite lead, sending fine dust drifting onto the dark wooden work surface with each precise cut. The previously intact lead has been carved into an incredible miniature structure—a chain of interconnected, hair-thin links, their intricate form clearly visible in the lens. A bright desk lamp, shining from above and to the side, casts concentrated light, not only sharp highlights on the metal blade but also intricate shadows on the carved links, enhancing the work's three-dimensional quality. The entire painting is permeated with an atmosphere of extreme concentration and almost breathless tranquility, highlighting the artist's incredible skills and patience.", + "seed": 5425, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_1238_seed_5425_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.045643884828314184, + "vlm_logprob": -0.045643884828314184, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.045643884828314184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01172508206218481 + ], + "opposite_logprob_tokens": [ + -5.511724948883057 + ], + "logprob_reward": -0.01172508206218481, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.01172508206218481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13470043241977692, + 0.03574914485216141, + 0.10257048159837723, + 0.20547521114349365, + 0.06907684355974197, + 0.07585561275482178, + 0.047781240195035934, + 0.0631583109498024, + 0.09738276153802872, + 0.1682499498128891 + ], + "credit_map": [ + 0.01847914047539234, + 0.02680017240345478, + 0.02058284543454647, + 0.012622070498764515, + 0.013713454827666283, + 0.011420284397900105, + 0.01604411192238331, + 0.015038350597023964, + 0.005723314359784126, + 0.005871125962585211, + 0.005628141108900309, + 0.003702601883560419, + 0.0031010538805276155, + 0.0023303297348320484, + 0.004734869115054607, + 0.004657709505409002, + 0.011407401412725449, + 0.015610598027706146, + 0.014411446638405323, + 0.006860458292067051, + 0.009875022806227207, + 0.00952109694480896, + 0.012621793895959854, + 0.022262658923864365, + 0.02301136404275894, + 0.02439015544950962, + 0.0551937073469162, + 0.010769660584628582, + 0.013767365366220474, + 0.024708783254027367, + 0.017177607864141464, + 0.036456555128097534, + 0.008820808492600918, + 0.010418553836643696, + 0.01043774839490652, + 0.007819894701242447, + 0.007636770140379667, + 0.007003805134445429, + 0.00875157117843628, + 0.008187695406377316, + 0.005938834976404905, + 0.004561309702694416, + 0.022650476545095444, + 0.004724475089460611, + 0.0039992970414459705, + 0.007580440025776625, + 0.005408714059740305, + 0.020992064848542213, + 0.0062785581685602665, + 0.005868755280971527, + 0.010088278912007809, + 0.0037789626512676477, + 0.005469345021992922, + 0.004429234191775322, + 0.0048560891300439835, + 0.007012017536908388, + 0.008261211216449738, + 0.008651074022054672, + 0.011097184382379055, + 0.0045043351128697395, + 0.0038313958793878555, + 0.004855475854128599, + 0.007628523278981447, + 0.014329112134873867, + 0.008978336118161678, + 0.007356578949838877, + 0.030965052545070648, + 0.006730292458087206, + 0.004177688155323267, + 0.0066913096234202385, + 0.009379302151501179, + 0.023104200139641762, + 0.017020991072058678, + 0.007558648940175772, + 0.05559266731142998, + 0.012345943599939346, + 0.010302816517651081, + 0.007719408720731735, + 0.00997866503894329, + 0.04773079976439476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.01172508206218481 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019547895062714815 + ], + "opposite_logprob_tokens": [ + -6.251954555511475 + ], + "logprob_reward": -0.0019547895062714815, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019547895062714815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11570052057504654, + 0.08325393497943878, + 0.10054769366979599, + 0.08339732885360718, + 0.03728341683745384, + 0.13128264248371124, + 0.02501724287867546, + 0.07838742434978485, + 0.034309275448322296, + 0.31082049012184143 + ], + "credit_map": [ + 0.012363641522824764, + 0.012188988737761974, + 0.022097744047641754, + 0.005562401842325926, + 0.008531199768185616, + 0.014234011061489582, + 0.014045827090740204, + 0.026676710695028305, + 0.011476326733827591, + 0.009845576249063015, + 0.013017469085752964, + 0.006544463336467743, + 0.0069078425876796246, + 0.009890157729387283, + 0.008054066449403763, + 0.01751803234219551, + 0.01340431161224842, + 0.015439361333847046, + 0.014022900722920895, + 0.009589118883013725, + 0.010364609770476818, + 0.008918752893805504, + 0.018432514742016792, + 0.010376120917499065, + 0.011414993554353714, + 0.007038402836769819, + 0.014900756999850273, + 0.006263910327106714, + 0.00757068395614624, + 0.012057206593453884, + 0.010906158946454525, + 0.01324521191418171, + 0.004204926546663046, + 0.005061620846390724, + 0.006137416232377291, + 0.004924194887280464, + 0.003407691605389118, + 0.00437501585111022, + 0.004275651182979345, + 0.004896901547908783, + 0.010350377298891544, + 0.004775171168148518, + 0.04746396839618683, + 0.007879580371081829, + 0.0039590708911418915, + 0.006763359531760216, + 0.005546869244426489, + 0.04454423114657402, + 0.0033315683249384165, + 0.004008101765066385, + 0.004765945952385664, + 0.0022604106925427914, + 0.0028318362310528755, + 0.0017941722180694342, + 0.002751324325799942, + 0.003273884765803814, + 0.010650702752172947, + 0.012032938189804554, + 0.011140963062644005, + 0.005997174419462681, + 0.007138135842978954, + 0.007657787762582302, + 0.012330805882811546, + 0.011438917368650436, + 0.004734067711979151, + 0.0031689880415797234, + 0.009094701148569584, + 0.0018207482062280178, + 0.0025641440879553556, + 0.0029730028472840786, + 0.0032234240788966417, + 0.006730199791491032, + 0.024455325677990913, + 0.008932730183005333, + 0.12100064754486084, + 0.012659951113164425, + 0.009637955576181412, + 0.012155036441981792, + 0.01233275979757309, + 0.10964606702327728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0019547895062714815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019378946162760258 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.0019378946162760258, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019378946162760258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14311622083187103, + 0.09739310294389725, + 0.2268400639295578, + 0.09243284910917282, + 0.10727803409099579, + 0.0635652020573616, + 0.05099030211567879, + 0.029108643531799316, + 0.020502733066678047, + 0.16877280175685883 + ], + "credit_map": [ + 0.01743026077747345, + 0.018572475761175156, + 0.025247937068343163, + 0.008670806884765625, + 0.012218012474477291, + 0.015413098968565464, + 0.018431583419442177, + 0.027132054790854454, + 0.0123166898265481, + 0.009129859507083893, + 0.014708123169839382, + 0.007666022516787052, + 0.009168768301606178, + 0.011521500535309315, + 0.011383842676877975, + 0.021498292684555054, + 0.02095300331711769, + 0.0361865870654583, + 0.031975213438272476, + 0.025586452335119247, + 0.019115429371595383, + 0.01782257854938507, + 0.03410748019814491, + 0.04109332337975502, + 0.011160893365740776, + 0.01430566981434822, + 0.020026883110404015, + 0.010714746080338955, + 0.004577482119202614, + 0.009275010786950588, + 0.007637979928404093, + 0.014734197407960892, + 0.012118404731154442, + 0.013908498920500278, + 0.01747809909284115, + 0.01726021058857441, + 0.007459545042365789, + 0.0092848539352417, + 0.012292574159801006, + 0.017475850880146027, + 0.0055857771076262, + 0.0041840230114758015, + 0.016998566687107086, + 0.004595902282744646, + 0.004295343067497015, + 0.006134012248367071, + 0.005053875967860222, + 0.016717705875635147, + 0.0081095639616251, + 0.0065298196859657764, + 0.007529522757977247, + 0.003832787275314331, + 0.008993365801870823, + 0.00456866854801774, + 0.005835752002894878, + 0.005590823944658041, + 0.004379930905997753, + 0.007184602785855532, + 0.004550943151116371, + 0.0027690145652741194, + 0.0024720749352127314, + 0.0017483648844063282, + 0.003282272955402732, + 0.0027214419096708298, + 0.0027620813343673944, + 0.0023715002462267876, + 0.004160380456596613, + 0.0031967118848115206, + 0.0013560722582042217, + 0.001868820283561945, + 0.001717716921120882, + 0.003069451078772545, + 0.013396220281720161, + 0.004625326953828335, + 0.06653346121311188, + 0.006523028016090393, + 0.004934724885970354, + 0.006506786681711674, + 0.006112606730312109, + 0.06014065816998482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0019378946162760258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106884703040123 + ], + "opposite_logprob_tokens": [ + -4.511068820953369 + ], + "logprob_reward": -0.01106884703040123, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106884703040123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19916267693042755, + 0.0695984736084938, + 0.1266850233078003, + 0.07974784076213837, + 0.09443426132202148, + 0.1231611967086792, + 0.07720594108104706, + 0.07903289049863815, + 0.05920005217194557, + 0.09177166223526001 + ], + "credit_map": [ + 0.034262511879205704, + 0.02450631745159626, + 0.040022112429142, + 0.01407440286129713, + 0.01635678857564926, + 0.017577435821294785, + 0.021068338304758072, + 0.03129475563764572, + 0.011706880293786526, + 0.012339772656559944, + 0.011538472026586533, + 0.009850134141743183, + 0.0059364126063883305, + 0.004221239592880011, + 0.006501548923552036, + 0.007504011504352093, + 0.01418328657746315, + 0.011484571732580662, + 0.021248146891593933, + 0.011767849326133728, + 0.008436353877186775, + 0.010580204427242279, + 0.019588109105825424, + 0.029396500438451767, + 0.013101928867399693, + 0.009506608359515667, + 0.013413527049124241, + 0.007207981310784817, + 0.004672147333621979, + 0.007993754930794239, + 0.009338770993053913, + 0.014513126574456692, + 0.010779970325529575, + 0.01070572528988123, + 0.019914137199521065, + 0.00846929382532835, + 0.004562898073345423, + 0.009757531806826591, + 0.008282660506665707, + 0.02196205034852028, + 0.009712282568216324, + 0.004020478576421738, + 0.04590017348527908, + 0.00648135133087635, + 0.003607988590374589, + 0.006298597902059555, + 0.004967351444065571, + 0.04217297583818436, + 0.012467884458601475, + 0.011195670813322067, + 0.013398099690675735, + 0.005740993190556765, + 0.009582537226378918, + 0.006342776585370302, + 0.008637330494821072, + 0.00984064768999815, + 0.01091435831040144, + 0.012369995936751366, + 0.011771241202950478, + 0.007397602312266827, + 0.007563576102256775, + 0.006473703775554895, + 0.010356257669627666, + 0.012186149135231972, + 0.009219223633408546, + 0.00801052525639534, + 0.011247213929891586, + 0.010950776748359203, + 0.0029697269201278687, + 0.003833702066913247, + 0.007408567238599062, + 0.005560316611081362, + 0.008490188978612423, + 0.004762628581374884, + 0.031723909080028534, + 0.00575247360393405, + 0.003944910131394863, + 0.004773781169205904, + 0.004645573440939188, + 0.027678193524479866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01106884703040123 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153281092643738 + ], + "opposite_logprob_tokens": [ + -1.7015328407287598 + ], + "logprob_reward": -0.20153281092643738, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153281092643738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45630747079849243, + 0.0180136077105999, + 0.050957292318344116, + 0.04206426441669464, + 0.05078000947833061, + 0.09968127310276031, + 0.09905432909727097, + 0.03244935721158981, + 0.03245934844017029, + 0.11823296546936035 + ], + "credit_map": [ + 0.055525414645671844, + 0.06855277717113495, + 0.10434003174304962, + 0.020346906036138535, + 0.045714057981967926, + 0.04140394553542137, + 0.04804611951112747, + 0.0723782628774643, + 0.00241257157176733, + 0.0020499478559941053, + 0.002507336437702179, + 0.0015810637269169092, + 0.0012532930122688413, + 0.0019334517419338226, + 0.001847582869231701, + 0.004428362473845482, + 0.004854364320635796, + 0.004302619025111198, + 0.009119014255702496, + 0.0038199513219296932, + 0.003353774780407548, + 0.004383724648505449, + 0.00769968144595623, + 0.013424169272184372, + 0.004704264923930168, + 0.007376482244580984, + 0.01108024176210165, + 0.0038009658455848694, + 0.0022537780459970236, + 0.003578623291105032, + 0.0031779403798282146, + 0.006091973278671503, + 0.004930607974529266, + 0.004062280058860779, + 0.012607870623469353, + 0.00589657062664628, + 0.002953591290861368, + 0.004463221877813339, + 0.004344879649579525, + 0.01152098923921585, + 0.008632802404463291, + 0.004092400427907705, + 0.03195329010486603, + 0.005819070152938366, + 0.003862332319840789, + 0.009368835017085075, + 0.005935085937380791, + 0.030017465353012085, + 0.010104523040354252, + 0.01125397440046072, + 0.019787557423114777, + 0.00789820495992899, + 0.014149581082165241, + 0.008321627974510193, + 0.010721100494265556, + 0.01681777462363243, + 0.0033321157097816467, + 0.0038956492207944393, + 0.005578449461609125, + 0.002443253993988037, + 0.0019207196310162544, + 0.0029335147701203823, + 0.003535251831635833, + 0.008810406550765038, + 0.004292451310902834, + 0.003141582477837801, + 0.007164350710809231, + 0.00238819676451385, + 0.0027936757542192936, + 0.0031761277932673693, + 0.0031339225824922323, + 0.006369046401232481, + 0.009361594915390015, + 0.0030442902352660894, + 0.04687679558992386, + 0.004757952876389027, + 0.003469217335805297, + 0.004344964865595102, + 0.004392753355205059, + 0.041985414922237396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.20153281092643738 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle captures a pair of steady, deft hands, absorbed in the work of a meticulously detailed artist. One hand firmly grips a classic yellow hexagonal wooden pencil, while the other, with steady and deliberate control, manipulates a finely crafted utility knife with a cold metallic sheen. The sharp tip of the blade meticulously scrapes away layer by layer from the fragile graphite lead, sending fine dust drifting onto the dark wooden work surface with each precise cut. The previously intact lead has been carved into an incredible miniature structure—a chain of interconnected, hair-thin links, their intricate form clearly visible in the lens. A bright desk lamp, shining from above and to the side, casts concentrated light, not only sharp highlights on the metal blade but also intricate shadows on the carved links, enhancing the work's three-dimensional quality. The entire painting is permeated with an atmosphere of extreme concentration and almost breathless tranquility, highlighting the artist's incredible skills and patience.", + "seed": 5426, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_10/index_1238_seed_5426_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.049161864817142485, + "vlm_logprob": -0.049161864817142485, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.049161864817142485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011986253783106804 + ], + "opposite_logprob_tokens": [ + -7.011986255645752 + ], + "logprob_reward": -0.011986253783106804, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.011986253783106804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06305111944675446, + 0.04149661585688591, + 0.02505551092326641, + 0.03195521607995033, + 0.07485639303922653, + 0.08977383375167847, + 0.023995639756321907, + 0.2261938452720642, + 0.3700203001499176, + 0.05360151082277298 + ], + "credit_map": [ + 0.00768708623945713, + 0.006430933717638254, + 0.006855325773358345, + 0.010991748422384262, + 0.007633783854544163, + 0.009458135813474655, + 0.007139455061405897, + 0.006854650564491749, + 0.003825627500191331, + 0.002794438973069191, + 0.005204620771110058, + 0.005878187250345945, + 0.006177356466650963, + 0.0037868907675147057, + 0.006159838754683733, + 0.007669661659747362, + 0.002599598141387105, + 0.0024190940894186497, + 0.0026282374747097492, + 0.003506097476929426, + 0.004341882187873125, + 0.0022285834420472383, + 0.003988201264292002, + 0.003343818709254265, + 0.003516833996400237, + 0.0028158333152532578, + 0.004611708223819733, + 0.004454889800399542, + 0.005457086488604546, + 0.002874281955882907, + 0.00416236137971282, + 0.004062220919877291, + 0.00838569737970829, + 0.006236253771930933, + 0.010061136446893215, + 0.006744104903191328, + 0.0057953037321567535, + 0.014621702954173088, + 0.008917689323425293, + 0.014094499871134758, + 0.009131131693720818, + 0.010928773321211338, + 0.013127914629876614, + 0.011957035399973392, + 0.007862887345254421, + 0.012847580946981907, + 0.00715585658326745, + 0.016762658953666687, + 0.003309104358777404, + 0.002877922961488366, + 0.0031382341403514147, + 0.002882743254303932, + 0.0025410130620002747, + 0.002907434944063425, + 0.003995549399405718, + 0.002343639964237809, + 0.01779625564813614, + 0.011213819496333599, + 0.038584161549806595, + 0.020166918635368347, + 0.015537581406533718, + 0.046165332198143005, + 0.01631906069815159, + 0.06041073054075241, + 0.03134801983833313, + 0.02251007780432701, + 0.053638506680727005, + 0.017780670896172523, + 0.018234696239233017, + 0.08404216915369034, + 0.03144475445151329, + 0.11102142184972763, + 0.006264733616262674, + 0.0027385330758988857, + 0.007078724913299084, + 0.004070894327014685, + 0.002849952084943652, + 0.013039923273026943, + 0.004874362610280514, + 0.012684389017522335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.011986253783106804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093251407146454 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093251407146454, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093251407146454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19162596762180328, + 0.1720147728919983, + 0.09811542928218842, + 0.039596494287252426, + 0.049100324511528015, + 0.06638869643211365, + 0.027852630242705345, + 0.2008085697889328, + 0.05085404962301254, + 0.10364308208227158 + ], + "credit_map": [ + 0.018944114446640015, + 0.017726384103298187, + 0.02139378897845745, + 0.03464971110224724, + 0.029606226831674576, + 0.020260941237211227, + 0.028175922110676765, + 0.020868897438049316, + 0.01740230992436409, + 0.01281816978007555, + 0.01767857000231743, + 0.028072210028767586, + 0.02841220237314701, + 0.01451040618121624, + 0.022047080099582672, + 0.031073827296495438, + 0.008808515965938568, + 0.008938015438616276, + 0.014757692813873291, + 0.01959885284304619, + 0.010962282307446003, + 0.009458578191697598, + 0.011662134900689125, + 0.01392934750765562, + 0.004595486912876368, + 0.005149765871465206, + 0.004163885954767466, + 0.005671123042702675, + 0.005456340499222279, + 0.004071567207574844, + 0.005673845764249563, + 0.004814479500055313, + 0.006682435050606728, + 0.007467212621122599, + 0.006853578146547079, + 0.005846200045198202, + 0.004630715120583773, + 0.005404273513704538, + 0.005810541100800037, + 0.006405364256352186, + 0.009499581530690193, + 0.005176038481295109, + 0.010809492319822311, + 0.0060699013993144035, + 0.0037484539207071066, + 0.011825545690953732, + 0.004258027300238609, + 0.015001656487584114, + 0.003305393736809492, + 0.0046905046328902245, + 0.003957388922572136, + 0.004870988428592682, + 0.0019766211044043303, + 0.0024311766028404236, + 0.0035415301099419594, + 0.0030790281016379595, + 0.017151320353150368, + 0.00779558252543211, + 0.03574617579579353, + 0.010420268401503563, + 0.0139427175745368, + 0.04151317477226257, + 0.018835317343473434, + 0.05540401116013527, + 0.009662618860602379, + 0.007820991799235344, + 0.0061415149830281734, + 0.006411930080503225, + 0.004784118849784136, + 0.006303723435848951, + 0.004362943582236767, + 0.005366204772144556, + 0.010400999337434769, + 0.0034470614045858383, + 0.01505451649427414, + 0.005915103945881128, + 0.0042428672313690186, + 0.025832664221525192, + 0.00930425338447094, + 0.02944561094045639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.011093251407146454 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633077144622803 + ], + "opposite_logprob_tokens": [ + -4.758633136749268 + ], + "logprob_reward": -0.008633077144622803, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633077144622803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14271481335163116, + 0.07751049846410751, + 0.06729785352945328, + 0.04680691286921501, + 0.1541702300310135, + 0.1483570635318756, + 0.08694390952587128, + 0.06634268909692764, + 0.12703286111354828, + 0.08282312750816345 + ], + "credit_map": [ + 0.014963526278734207, + 0.01054968498647213, + 0.015244601294398308, + 0.021236902102828026, + 0.014535479247570038, + 0.012650417163968086, + 0.03068215399980545, + 0.02285204641520977, + 0.008774185553193092, + 0.005160795524716377, + 0.009316623210906982, + 0.010759998112916946, + 0.014056356623768806, + 0.006184311583638191, + 0.011405262164771557, + 0.011852966621518135, + 0.006368803326040506, + 0.00443518441170454, + 0.008611656725406647, + 0.012749256566166878, + 0.005942607298493385, + 0.007373336236923933, + 0.012400832027196884, + 0.009416177868843079, + 0.0045096250250935555, + 0.0028568189591169357, + 0.005022812634706497, + 0.006807024125009775, + 0.004109561443328857, + 0.005740307737141848, + 0.010767430067062378, + 0.00699333380907774, + 0.014439206570386887, + 0.026754340156912804, + 0.01726469025015831, + 0.021506473422050476, + 0.01905898191034794, + 0.013522472232580185, + 0.024716196581721306, + 0.01690788008272648, + 0.02116476185619831, + 0.016883546486496925, + 0.02030893787741661, + 0.019555499777197838, + 0.013591514900326729, + 0.01801050640642643, + 0.014213350601494312, + 0.02462894655764103, + 0.011515147052705288, + 0.012448428198695183, + 0.012321402318775654, + 0.013410468585789204, + 0.009166315197944641, + 0.009436258114874363, + 0.009743628092110157, + 0.00890226662158966, + 0.008901305496692657, + 0.005889879539608955, + 0.007984256371855736, + 0.00696056243032217, + 0.009145373478531837, + 0.00972129125148058, + 0.00785505399107933, + 0.009884961880743504, + 0.01624567247927189, + 0.019154556095600128, + 0.014479443430900574, + 0.014110970310866833, + 0.009252571500837803, + 0.019127968698740005, + 0.009613514877855778, + 0.025048168376088142, + 0.010297340340912342, + 0.0024659198243170977, + 0.010835579596459866, + 0.0046766409650444984, + 0.0032358088064938784, + 0.023240286856889725, + 0.0078011672012507915, + 0.02027038484811783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.008633077144622803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532057240605354 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532057240605354, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532057240605354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02797813154757023, + 0.013974422588944435, + 0.02739356830716133, + 0.012084673158824444, + 0.05629490688443184, + 0.23599670827388763, + 0.019686520099639893, + 0.5522048473358154, + 0.045803435146808624, + 0.00858279038220644 + ], + "credit_map": [ + 0.0028545160312205553, + 0.003168376861140132, + 0.0033442063722759485, + 0.006002915557473898, + 0.002507933182641864, + 0.002193981548771262, + 0.003490990726277232, + 0.004415210802108049, + 0.0016399018932133913, + 0.0011210041120648384, + 0.0017288257367908955, + 0.0022238229867070913, + 0.0013989013386890292, + 0.0011405623517930508, + 0.0021076034754514694, + 0.0026138005778193474, + 0.0027430104091763496, + 0.0032078027725219727, + 0.0030061400029808283, + 0.004586225375533104, + 0.004137055948376656, + 0.0028162910602986813, + 0.0034494269639253616, + 0.0034476150758564472, + 0.0015001122374087572, + 0.00132736936211586, + 0.0013452956918627024, + 0.0014752565184608102, + 0.0016612311592325568, + 0.0015226522227749228, + 0.0015484970062971115, + 0.0017042590770870447, + 0.005466721020638943, + 0.005290328990668058, + 0.007135421968996525, + 0.008386574685573578, + 0.006966079119592905, + 0.009378910064697266, + 0.00619520852342248, + 0.007475666701793671, + 0.024447117000818253, + 0.0181431844830513, + 0.04136894643306732, + 0.02645343542098999, + 0.011789110489189625, + 0.041252247989177704, + 0.01469203270971775, + 0.05785064026713371, + 0.0027862475253641605, + 0.001814907300285995, + 0.002424575388431549, + 0.002723569516092539, + 0.002257359679788351, + 0.0034954524599015713, + 0.0018813760252669454, + 0.002303031273186207, + 0.04448724910616875, + 0.02432684600353241, + 0.09833122044801712, + 0.03676290065050125, + 0.035411495715379715, + 0.11915949732065201, + 0.036637503653764725, + 0.15708814561367035, + 0.007865192368626595, + 0.004804505035281181, + 0.005768394563347101, + 0.0053562019020318985, + 0.00581909017637372, + 0.005027218256145716, + 0.00495512317866087, + 0.0062077101320028305, + 0.0008564798627048731, + 0.0006597483297809958, + 0.0008640781161375344, + 0.0013116038171574473, + 0.001396368839778006, + 0.0014002887764945626, + 0.0011756432941183448, + 0.0009185791132040322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.012532057240605354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of steady hands visible working on a miniature structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand gripping a yellow hexagonal wooden pencil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other hand manipulating a utility knife with a metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the carved object a chain of interconnected hair-thin links?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a bright desk lamp casting sharp highlights and shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156468451023102 + ], + "opposite_logprob_tokens": [ + -1.7015646696090698 + ], + "logprob_reward": -0.20156468451023102, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156468451023102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15232223272323608, + 0.06802372634410858, + 0.06926476210355759, + 0.09556274116039276, + 0.02900887094438076, + 0.08524005115032196, + 0.01809694990515709, + 0.20587988197803497, + 0.17752806842327118, + 0.09907266497612 + ], + "credit_map": [ + 0.017761649563908577, + 0.015017579309642315, + 0.017238950356841087, + 0.027874119579792023, + 0.015818564221262932, + 0.019659467041492462, + 0.021823124960064888, + 0.01712879166007042, + 0.006051240488886833, + 0.0070457737892866135, + 0.008152308873832226, + 0.013881534337997437, + 0.008205557242035866, + 0.0070554520934820175, + 0.00699493708088994, + 0.010636923834681511, + 0.007895994931459427, + 0.006753524299710989, + 0.008019168861210346, + 0.011292669922113419, + 0.007495175581425428, + 0.00804037693887949, + 0.0098956860601902, + 0.009872164577245712, + 0.011884159408509731, + 0.011423110030591488, + 0.012225458398461342, + 0.012819943018257618, + 0.011853355914354324, + 0.011547496542334557, + 0.012337160296738148, + 0.011472062207758427, + 0.004291122313588858, + 0.0041417814791202545, + 0.0032831274438649416, + 0.003992661368101835, + 0.002905383938923478, + 0.0031758099794387817, + 0.0039366865530610085, + 0.0032822987996041775, + 0.010429772548377514, + 0.007564985193312168, + 0.013627366162836552, + 0.010831418447196484, + 0.004940045066177845, + 0.012894541956484318, + 0.007228656206279993, + 0.017723267897963524, + 0.002408071653917432, + 0.0020227949135005474, + 0.0024250992573797703, + 0.0035589009057730436, + 0.00161341717466712, + 0.0015048125060275197, + 0.0021635156590491533, + 0.002400337252765894, + 0.015766022726893425, + 0.010492491535842419, + 0.03536394611001015, + 0.015369908884167671, + 0.015047350898385048, + 0.04286803677678108, + 0.013972203247249126, + 0.05699992552399635, + 0.012756680138409138, + 0.01400382537394762, + 0.025974653661251068, + 0.008940725587308407, + 0.011088617146015167, + 0.03694669157266617, + 0.01577744074165821, + 0.05203942954540253, + 0.011174087412655354, + 0.0041180020198225975, + 0.012368859723210335, + 0.009481925517320633, + 0.004863441456109285, + 0.025191346183419228, + 0.010678998194634914, + 0.02119600400328636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.20156468451023102 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_11/index_tensor([177, 177])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_11/index_tensor([177, 177])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e52641f6aeb0136b947e91e23c33884edd93079 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_11/index_tensor([177, 177])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "rapid montage of surreal, high-resolution stills shot from dynamic low angles with a telephoto lens freezes a diverse group of marathon runners in their final, agonizing sprint toward the finish line. Each flashing frame focuses on a different runner, their face a raw canvas of remarkably determined determination: brows furrowed, sweat dripping from temples and chins, jaws clenched, eyes narrowed, unwavering focus on the goal beyond the frame. Their sweat-soaked tracksuits, vibrant in neon yellow, electric blue, and bold red, cling to their lean, muscular bodies, their race bibs pinned to their chests. The camera captures a series of powerful micro-moments: the explosive push of their feet, the splash of water on the wet asphalt; the tense tension of their calf muscles in the harsh light; the decisive pumping of their arms through the air. In the background, the cheering crowd and finish line structure are rendered in a soft, colorful bokeh, emphasizing the athletes' solitary struggle. The scene is lit by bright, direct overhead sunlight, creating high contrast and sharp shadows that make every drop of sweat and strained sinew stand out with crystal clarity. The atmosphere is intensely visceral, filled with the silent yet powerful emotion of human endurance.", + "seed": 15351, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_11/index_177_seed_15351_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.27463455144315957, + "vlm_logprob": -0.27463455144315957, + "vlm_yes_no_margin": 2.999999988079071, + "vlm_token_credit": -0.27463455144315957, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011375096626579762 + ], + "opposite_logprob_tokens": [ + -5.0113749504089355 + ], + "logprob_reward": -0.011375096626579762, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011375096626579762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08838040381669998, + 0.09389135986566544, + 0.1000475138425827, + 0.12701663374900818, + 0.09334631264209747, + 0.1011432632803917, + 0.11928300559520721, + 0.08094186335802078, + 0.09057159721851349, + 0.10537803173065186 + ], + "credit_map": [ + 0.01070481538772583, + 0.01601470820605755, + 0.010303323157131672, + 0.008539590053260326, + 0.012119035236537457, + 0.008325846865773201, + 0.010023183189332485, + 0.012349903583526611, + 0.012753456830978394, + 0.012799389660358429, + 0.014010593295097351, + 0.011045149527490139, + 0.017522817477583885, + 0.011055211536586285, + 0.007176922168582678, + 0.007527818437665701, + 0.023929042741656303, + 0.014397758059203625, + 0.010895408689975739, + 0.007054370827972889, + 0.011197125539183617, + 0.014051774516701698, + 0.007881452329456806, + 0.010640583001077175, + 0.02817182056605816, + 0.015599310398101807, + 0.017589371651411057, + 0.009027360007166862, + 0.014835840091109276, + 0.01581088826060295, + 0.012392262928187847, + 0.013589777983725071, + 0.013753184117376804, + 0.013848005793988705, + 0.008830548264086246, + 0.00920239556580782, + 0.009746614843606949, + 0.015277063474059105, + 0.008007160387933254, + 0.014681338332593441, + 0.017334653064608574, + 0.013707209378480911, + 0.010032172314822674, + 0.010640982538461685, + 0.01352530438452959, + 0.012433579191565514, + 0.010865254327654839, + 0.012604107148945332, + 0.010310342535376549, + 0.014123245142400265, + 0.010451103560626507, + 0.017970064654946327, + 0.013391190208494663, + 0.011993560940027237, + 0.02485530637204647, + 0.01618819311261177, + 0.018700698390603065, + 0.01076426263898611, + 0.010477130301296711, + 0.005916610825806856, + 0.00904460996389389, + 0.012717410922050476, + 0.006728154141455889, + 0.006592989433556795, + 0.011418714188039303, + 0.018212564289569855, + 0.0096439765766263, + 0.009734204970300198, + 0.01563992165029049, + 0.008740240707993507, + 0.008855781517922878, + 0.00832618959248066, + 0.013013459742069244, + 0.020100126042962074, + 0.009857983328402042, + 0.011758143082261086, + 0.019421664997935295, + 0.008813138119876385, + 0.010848426260054111, + 0.01156508270651102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.011375096626579762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129041314125061 + ], + "opposite_logprob_tokens": [ + -2.2379040718078613 + ], + "logprob_reward": -0.1129041314125061, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129041314125061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2303728461265564, + 0.06296820193529129, + 0.0630059763789177, + 0.05031943321228027, + 0.1374054104089737, + 0.05305523797869682, + 0.11085031181573868, + 0.0660611093044281, + 0.041479095816612244, + 0.18448235094547272 + ], + "credit_map": [ + 0.021767985075712204, + 0.03836409002542496, + 0.017348753288388252, + 0.02156963013112545, + 0.038663100451231, + 0.027703948318958282, + 0.038854945451021194, + 0.02610039710998535, + 0.008385246619582176, + 0.00793349090963602, + 0.008423713967204094, + 0.007823636755347252, + 0.01245336513966322, + 0.007343844510614872, + 0.004895440768450499, + 0.0057094707153737545, + 0.012680880725383759, + 0.012081596069037914, + 0.006890933495014906, + 0.005134317092597485, + 0.008324733935296535, + 0.007310176733881235, + 0.005455728154629469, + 0.005127611104398966, + 0.008308829739689827, + 0.006842596922069788, + 0.008490349166095257, + 0.004536291118711233, + 0.005314571782946587, + 0.006255413871258497, + 0.004976052790880203, + 0.005595329683274031, + 0.018738241866230965, + 0.02439119666814804, + 0.018768515437841415, + 0.013718054629862309, + 0.01846935600042343, + 0.01610684208571911, + 0.011884771287441254, + 0.015328447334468365, + 0.0060146707110106945, + 0.009494982659816742, + 0.006438751704990864, + 0.005220706574618816, + 0.009513862431049347, + 0.0070178816094994545, + 0.004219822585582733, + 0.005134562961757183, + 0.009666341356933117, + 0.013727091252803802, + 0.011571498587727547, + 0.018954049795866013, + 0.015635168179869652, + 0.013404401950538158, + 0.013882700353860855, + 0.014009078964591026, + 0.00955269392579794, + 0.009481461718678474, + 0.008619134314358234, + 0.007736051455140114, + 0.010294252075254917, + 0.0072997817769646645, + 0.0054337880574166775, + 0.007643951568752527, + 0.0069512552581727505, + 0.00582811189815402, + 0.004668834153562784, + 0.003044029464945197, + 0.005469887517392635, + 0.004271495155990124, + 0.005981291178613901, + 0.005264194216579199, + 0.017395080998539925, + 0.03950509428977966, + 0.021897858008742332, + 0.02557274140417576, + 0.02190236561000347, + 0.01668079011142254, + 0.021808335557579994, + 0.01972009427845478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.1129041314125061 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199530705809593 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199530705809593, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199530705809593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1002252846956253, + 0.044647786766290665, + 0.05911978706717491, + 0.04007093980908394, + 0.20547877252101898, + 0.02521836943924427, + 0.06028902903199196, + 0.04217715188860893, + 0.026415066793560982, + 0.39635777473449707 + ], + "credit_map": [ + 0.01029307022690773, + 0.019913984462618828, + 0.007945438846945763, + 0.007884415797889233, + 0.018294870853424072, + 0.00870269164443016, + 0.016971200704574585, + 0.010219616815447807, + 0.004974362440407276, + 0.006235227454453707, + 0.005393777042627335, + 0.004360475577414036, + 0.009653403423726559, + 0.004860789515078068, + 0.00402436126023531, + 0.0051453933119773865, + 0.010061660781502724, + 0.009955736808478832, + 0.007696591317653656, + 0.004920146428048611, + 0.007944604381918907, + 0.006059393752366304, + 0.005512928124517202, + 0.006968728732317686, + 0.005966485943645239, + 0.004053621552884579, + 0.0031277011148631573, + 0.004192203748971224, + 0.004528635647147894, + 0.006071372888982296, + 0.004915751051157713, + 0.0072151729837059975, + 0.03621004521846771, + 0.029192274436354637, + 0.02253769524395466, + 0.0389973409473896, + 0.024706700816750526, + 0.021441690623760223, + 0.018298601731657982, + 0.014094448648393154, + 0.003925520461052656, + 0.003456506412476301, + 0.0025827432982623577, + 0.00226271478459239, + 0.0030402911361306906, + 0.004634059965610504, + 0.0022915941663086414, + 0.0030249394476413727, + 0.009157221764326096, + 0.007803065236657858, + 0.004755887668579817, + 0.005402528215199709, + 0.009318875148892403, + 0.006169685162603855, + 0.004913654178380966, + 0.01276811771094799, + 0.0045575592666864395, + 0.0049213808961212635, + 0.004471098072826862, + 0.007709846831858158, + 0.004789277445524931, + 0.0052753412164747715, + 0.006542409770190716, + 0.00391024025157094, + 0.004212692379951477, + 0.0023554651997983456, + 0.003458879655227065, + 0.0025229607708752155, + 0.00227583316154778, + 0.003791108028963208, + 0.0033370591700077057, + 0.004461069591343403, + 0.03994134068489075, + 0.0783499926328659, + 0.04881902411580086, + 0.05353006720542908, + 0.043619219213724136, + 0.037629127502441406, + 0.04577188566327095, + 0.048697151243686676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.014199530705809593 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071676388382912 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071676388382912, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071676388382912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0760212391614914, + 0.03074137307703495, + 0.056965265423059464, + 0.15553535521030426, + 0.09928791224956512, + 0.02453460358083248, + 0.43244481086730957, + 0.03701912239193916, + 0.018556438386440277, + 0.06889389455318451 + ], + "credit_map": [ + 0.007596607785671949, + 0.014166844077408314, + 0.0075407326221466064, + 0.007305611856281757, + 0.013983039185404778, + 0.006955109536647797, + 0.011036770418286324, + 0.007436518091708422, + 0.00395490787923336, + 0.004624254535883665, + 0.004997703712433577, + 0.0029395478777587414, + 0.006409946829080582, + 0.0029744496569037437, + 0.002170041436329484, + 0.002670519519597292, + 0.009038465097546577, + 0.007471852004528046, + 0.009008863940834999, + 0.005994877312332392, + 0.007260676007717848, + 0.006705267820507288, + 0.005551563575863838, + 0.005933696869760752, + 0.029959559440612793, + 0.018931826576590538, + 0.028495538979768753, + 0.011364349164068699, + 0.020287130028009415, + 0.018772901967167854, + 0.01406847033649683, + 0.01365557312965393, + 0.011179113760590553, + 0.012177404947578907, + 0.009299295954406261, + 0.01658075675368309, + 0.011232474818825722, + 0.01321418583393097, + 0.00859513133764267, + 0.017009541392326355, + 0.0034813748206943274, + 0.0027546100318431854, + 0.0035346357617527246, + 0.0029646435286849737, + 0.003156108781695366, + 0.0036644628271460533, + 0.0026860132347792387, + 0.002292755525559187, + 0.035995155572891235, + 0.06468737125396729, + 0.05581516772508621, + 0.052667539566755295, + 0.06077974662184715, + 0.04941368103027344, + 0.06497230380773544, + 0.048113830387592316, + 0.004364040680229664, + 0.004753707908093929, + 0.0038056024350225925, + 0.004979350138455629, + 0.0055351098999381065, + 0.005832287948578596, + 0.00467837555333972, + 0.003070645034313202, + 0.00232627522200346, + 0.0034229399170726538, + 0.001577426097355783, + 0.001541223842650652, + 0.003122482681646943, + 0.0019449719693511724, + 0.0029234651010483503, + 0.0016976522747427225, + 0.006756295450031757, + 0.015697017312049866, + 0.00813679676502943, + 0.00845725741237402, + 0.00864072423428297, + 0.005580179393291473, + 0.008790739811956882, + 0.00683488417416811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.011071676388382912 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236223220825195 + ], + "opposite_logprob_tokens": [ + -0.34862226247787476 + ], + "logprob_reward": -1.2236223220825195, + "yes_no_margin_reward": -0.8750000596046448, + "token_credit_reward": -1.2236223220825195, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20130155980587006, + 0.04134242609143257, + 0.043401170521974564, + 0.042372096329927444, + 0.14354202151298523, + 0.021245598793029785, + 0.2164309173822403, + 0.04199831932783127, + 0.05657867714762688, + 0.19178716838359833 + ], + "credit_map": [ + 0.017098495736718178, + 0.03263682872056961, + 0.012048440985381603, + 0.015761610120534897, + 0.03640565648674965, + 0.023611322045326233, + 0.03585169091820717, + 0.02788752317428589, + 0.004744304809719324, + 0.004120923578739166, + 0.0030679795891046524, + 0.004040321335196495, + 0.005574754439294338, + 0.005464743822813034, + 0.0036409199237823486, + 0.010688484646379948, + 0.01325914729386568, + 0.0044582318514585495, + 0.006209047511219978, + 0.0028980830684304237, + 0.00377824530005455, + 0.005089255981147289, + 0.0036913123913109303, + 0.004017853178083897, + 0.004997281823307276, + 0.004515902604907751, + 0.0039026117883622646, + 0.0037865370977669954, + 0.006357845850288868, + 0.005596988834440708, + 0.00465451879426837, + 0.008560413494706154, + 0.02083028294146061, + 0.01865357905626297, + 0.01460305042564869, + 0.026300175115466118, + 0.017179986461997032, + 0.01546940766274929, + 0.013254386372864246, + 0.017251167446374893, + 0.0030194248538464308, + 0.0027416625525802374, + 0.0029349394608289003, + 0.0022212774492800236, + 0.0024754651822149754, + 0.0033884963486343622, + 0.002193063497543335, + 0.0022712715435773134, + 0.02606743760406971, + 0.02914496511220932, + 0.021743563935160637, + 0.027159718796610832, + 0.025210166350007057, + 0.025101393461227417, + 0.031100749969482422, + 0.03090294823050499, + 0.007584122940897942, + 0.007075250614434481, + 0.004499622154980898, + 0.00408227788284421, + 0.006157183088362217, + 0.00473283464089036, + 0.004623229615390301, + 0.0032438046764582396, + 0.009663764387369156, + 0.005898048635572195, + 0.006893380545079708, + 0.003956833854317665, + 0.006404690444469452, + 0.009649454616010189, + 0.005709686782211065, + 0.008402825333178043, + 0.018170055001974106, + 0.04235048592090607, + 0.021290726959705353, + 0.023875009268522263, + 0.022859035059809685, + 0.01705772615969181, + 0.025477169081568718, + 0.020706983283162117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -1.2236223220825195 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "rapid montage of surreal, high-resolution stills shot from dynamic low angles with a telephoto lens freezes a diverse group of marathon runners in their final, agonizing sprint toward the finish line. Each flashing frame focuses on a different runner, their face a raw canvas of remarkably determined determination: brows furrowed, sweat dripping from temples and chins, jaws clenched, eyes narrowed, unwavering focus on the goal beyond the frame. Their sweat-soaked tracksuits, vibrant in neon yellow, electric blue, and bold red, cling to their lean, muscular bodies, their race bibs pinned to their chests. The camera captures a series of powerful micro-moments: the explosive push of their feet, the splash of water on the wet asphalt; the tense tension of their calf muscles in the harsh light; the decisive pumping of their arms through the air. In the background, the cheering crowd and finish line structure are rendered in a soft, colorful bokeh, emphasizing the athletes' solitary struggle. The scene is lit by bright, direct overhead sunlight, creating high contrast and sharp shadows that make every drop of sweat and strained sinew stand out with crystal clarity. The atmosphere is intensely visceral, filled with the silent yet powerful emotion of human endurance.", + "seed": 15352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_11/index_177_seed_15352_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.25939358798787, + "vlm_logprob": -0.25939358798787, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.25939358798787, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02577570267021656 + ], + "opposite_logprob_tokens": [ + -4.025775909423828 + ], + "logprob_reward": -0.02577570267021656, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02577570267021656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1701706051826477, + 0.12741678953170776, + 0.1731722354888916, + 0.08344221860170364, + 0.048939988017082214, + 0.07820399105548859, + 0.10731098800897598, + 0.03018251061439514, + 0.12851744890213013, + 0.05264323949813843 + ], + "credit_map": [ + 0.035307370126247406, + 0.01838960126042366, + 0.0256468765437603, + 0.01324122492223978, + 0.04192180931568146, + 0.01302577368915081, + 0.010536798276007175, + 0.012101146392524242, + 0.013926895335316658, + 0.01834685169160366, + 0.030959315598011017, + 0.011595536023378372, + 0.011292288079857826, + 0.02221951074898243, + 0.009241878055036068, + 0.009834510274231434, + 0.016384094953536987, + 0.029125872999429703, + 0.02008342742919922, + 0.01566586084663868, + 0.029828853905200958, + 0.020184503868222237, + 0.014718885533511639, + 0.027180735021829605, + 0.007908851839601994, + 0.018067732453346252, + 0.006547772325575352, + 0.011085337027907372, + 0.01719416119158268, + 0.009813684970140457, + 0.004575302824378014, + 0.008249372243881226, + 0.0037356847897171974, + 0.005125863011926413, + 0.0043783001601696014, + 0.005697072949260473, + 0.0086824931204319, + 0.005482946056872606, + 0.009217068552970886, + 0.00662055891007185, + 0.012266782112419605, + 0.008374902419745922, + 0.009132057428359985, + 0.01158886682242155, + 0.008928744122385979, + 0.010861578397452831, + 0.007634636480361223, + 0.009416422806680202, + 0.012690231204032898, + 0.022001491859555244, + 0.009227846749126911, + 0.014923784881830215, + 0.01289317011833191, + 0.009839432314038277, + 0.017678380012512207, + 0.008056653663516045, + 0.0048193130642175674, + 0.0030345129780471325, + 0.0034312440548092127, + 0.003308151848614216, + 0.003217156510800123, + 0.005753261037170887, + 0.0026905741542577744, + 0.003928298130631447, + 0.017240386456251144, + 0.017903460189700127, + 0.01316185761243105, + 0.010522572323679924, + 0.019837772473692894, + 0.019199859350919724, + 0.014337345026433468, + 0.016314193606376648, + 0.00395269226282835, + 0.011762754060328007, + 0.00962579995393753, + 0.004452531225979328, + 0.005904237274080515, + 0.0053777811117470264, + 0.007925842888653278, + 0.003641601651906967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.02577570267021656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17984892427921295 + ], + "opposite_logprob_tokens": [ + -1.8048489093780518 + ], + "logprob_reward": -0.17984892427921295, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17984892427921295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33775240182876587, + 0.038515347987413406, + 0.10217991471290588, + 0.07770814001560211, + 0.04101954773068428, + 0.09951229393482208, + 0.08819154649972916, + 0.05844564363360405, + 0.06962525099515915, + 0.08704987168312073 + ], + "credit_map": [ + 0.035011693835258484, + 0.019374024122953415, + 0.10855183005332947, + 0.01659124530851841, + 0.11389133334159851, + 0.01549429539591074, + 0.01320632454007864, + 0.015631655231118202, + 0.002853079466149211, + 0.005854141432791948, + 0.005126052536070347, + 0.004048412665724754, + 0.004100635182112455, + 0.005594989750534296, + 0.005002748686820269, + 0.005935290828347206, + 0.012790033593773842, + 0.016820501536130905, + 0.012251478619873524, + 0.009807808324694633, + 0.016096411272883415, + 0.011383461765944958, + 0.00813085213303566, + 0.014899371191859245, + 0.009774914011359215, + 0.01649680733680725, + 0.007091933395713568, + 0.008430952206254005, + 0.01713259145617485, + 0.008397379890084267, + 0.005334088113158941, + 0.005049466621130705, + 0.006052134558558464, + 0.008562516421079636, + 0.0029396580066531897, + 0.0038876328617334366, + 0.004952826537191868, + 0.007436541374772787, + 0.003536962205544114, + 0.003651277394965291, + 0.022476352751255035, + 0.009661565534770489, + 0.011581971310079098, + 0.011646570637822151, + 0.008733686059713364, + 0.014535544440150261, + 0.009546572342514992, + 0.011330035515129566, + 0.011679215356707573, + 0.01564212515950203, + 0.00743273738771677, + 0.012372628785669804, + 0.015532825142145157, + 0.007486988790333271, + 0.012004457414150238, + 0.006040561478585005, + 0.01013319194316864, + 0.005951014347374439, + 0.005405241157859564, + 0.007564223371446133, + 0.007584471721202135, + 0.010167334228754044, + 0.00634774612262845, + 0.005292417015880346, + 0.005164149217307568, + 0.02011626958847046, + 0.0037731218617409468, + 0.004554728511720896, + 0.0060104248113930225, + 0.006399883888661861, + 0.005717290099710226, + 0.017889391630887985, + 0.008461947552859783, + 0.00807227287441492, + 0.011001181788742542, + 0.0064779287204146385, + 0.005971889477223158, + 0.018101686611771584, + 0.02335462160408497, + 0.005608341656625271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.17984892427921295 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636047434061766 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636047434061766, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636047434061766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2008894979953766, + 0.08645650744438171, + 0.08664535731077194, + 0.08196673542261124, + 0.04943374544382095, + 0.09182223677635193, + 0.0666898563504219, + 0.0965227410197258, + 0.15521496534347534, + 0.08435838669538498 + ], + "credit_map": [ + 0.022696122527122498, + 0.0158202163875103, + 0.03330770879983902, + 0.01986168511211872, + 0.06078319251537323, + 0.012864152900874615, + 0.011445547454059124, + 0.024110877886414528, + 0.00894817803055048, + 0.014502102509140968, + 0.011572466231882572, + 0.00915626809000969, + 0.007268349640071392, + 0.016469169408082962, + 0.008234363049268723, + 0.010305614210665226, + 0.009286077693104744, + 0.01495201513171196, + 0.012779411859810352, + 0.01083935983479023, + 0.010854721069335938, + 0.0072896662168204784, + 0.01170418132096529, + 0.00893991719931364, + 0.009655441157519817, + 0.01564125530421734, + 0.008814620785415173, + 0.01046303566545248, + 0.01255716010928154, + 0.008749139495193958, + 0.008507836610078812, + 0.007578241638839245, + 0.003993745893239975, + 0.004992001689970493, + 0.0037676841020584106, + 0.005309055559337139, + 0.0073830485343933105, + 0.006584661081433296, + 0.011699059046804905, + 0.005704487673938274, + 0.012157680466771126, + 0.009510207921266556, + 0.007229671347886324, + 0.011134753003716469, + 0.008935710415244102, + 0.013842007145285606, + 0.016477765515446663, + 0.012534437701106071, + 0.011602398008108139, + 0.006839570123702288, + 0.005680756643414497, + 0.006829468067735434, + 0.0147762605920434, + 0.006016015540808439, + 0.009068372659385204, + 0.005877018440514803, + 0.010775650851428509, + 0.01246987096965313, + 0.011210345663130283, + 0.012768467888236046, + 0.01341977808624506, + 0.013855966739356518, + 0.012009900063276291, + 0.010012765415012836, + 0.016791794449090958, + 0.04304591566324234, + 0.009757012128829956, + 0.012307059019804, + 0.015016313642263412, + 0.015127801336348057, + 0.01499477494508028, + 0.02817428857088089, + 0.006382833234965801, + 0.023699361830949783, + 0.006427236832678318, + 0.005724963266402483, + 0.005104193929582834, + 0.005318992305546999, + 0.025087302550673485, + 0.006613495294004679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.007636047434061766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798321425914764 + ], + "opposite_logprob_tokens": [ + -3.5297982692718506 + ], + "logprob_reward": -0.029798321425914764, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798321425914764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32153069972991943, + 0.16480770707130432, + 0.08041438460350037, + 0.030976485460996628, + 0.09457466751337051, + 0.09716596454381943, + 0.07376591861248016, + 0.03832422196865082, + 0.032601241022348404, + 0.06583869457244873 + ], + "credit_map": [ + 0.0344664640724659, + 0.022852517664432526, + 0.08091078698635101, + 0.017627574503421783, + 0.10902102291584015, + 0.020061245188117027, + 0.019039466977119446, + 0.01755163073539734, + 0.01712796650826931, + 0.032284580171108246, + 0.023545876145362854, + 0.01700984127819538, + 0.013417873531579971, + 0.017169497907161713, + 0.013808660209178925, + 0.03044341690838337, + 0.009249988943338394, + 0.011688160710036755, + 0.007710346020758152, + 0.008528508245944977, + 0.012949377298355103, + 0.008606569841504097, + 0.007199697662144899, + 0.014481736347079277, + 0.0044093928299844265, + 0.00529243890196085, + 0.004201733972877264, + 0.003913413267582655, + 0.0037887427024543285, + 0.0037713239435106516, + 0.0024750027805566788, + 0.0031244365964084864, + 0.009309298358857632, + 0.011249948292970657, + 0.007011220324784517, + 0.00831461139023304, + 0.017821412533521652, + 0.022428706288337708, + 0.010639408603310585, + 0.00780006218701601, + 0.013214299455285072, + 0.010826759040355682, + 0.008599812164902687, + 0.014203782193362713, + 0.009153787977993488, + 0.013795074075460434, + 0.014243531972169876, + 0.013128918595612049, + 0.008432813920080662, + 0.010390431620180607, + 0.009498808532953262, + 0.010053737089037895, + 0.01122339628636837, + 0.007100462447851896, + 0.009995464235544205, + 0.007070799823850393, + 0.004644522909075022, + 0.004150997847318649, + 0.0036402468103915453, + 0.003912028390914202, + 0.004066395107656717, + 0.006989858113229275, + 0.005346081685274839, + 0.005574093200266361, + 0.004081794526427984, + 0.004677793476730585, + 0.00357689056545496, + 0.00330947944894433, + 0.004854580387473106, + 0.00499306945130229, + 0.003945880103856325, + 0.0031617525964975357, + 0.004938862286508083, + 0.014328491874039173, + 0.007248377427458763, + 0.004719980526715517, + 0.004593337420374155, + 0.004695952404290438, + 0.02060178481042385, + 0.0047119068913161755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.029798321425914764 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539089441299438 + ], + "opposite_logprob_tokens": [ + -0.42890897393226624 + ], + "logprob_reward": -1.0539089441299438, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539089441299438, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3082340359687805, + 0.07452066242694855, + 0.048295050859451294, + 0.022424621507525444, + 0.046882547438144684, + 0.08165895938873291, + 0.05947006866335869, + 0.16688501834869385, + 0.10828890651464462, + 0.08334014564752579 + ], + "credit_map": [ + 0.03852345049381256, + 0.019880492240190506, + 0.08337581902742386, + 0.016238298267126083, + 0.09134967625141144, + 0.021453281864523888, + 0.016830939799547195, + 0.020582087337970734, + 0.006864782422780991, + 0.008877081796526909, + 0.011430097743868828, + 0.007818682119250298, + 0.011147907003760338, + 0.012320376932621002, + 0.007966599427163601, + 0.008095137774944305, + 0.005280622746795416, + 0.006867874413728714, + 0.00534583767876029, + 0.0045150588266551495, + 0.00841380562633276, + 0.005050698760896921, + 0.005387714598327875, + 0.007433438207954168, + 0.002727668033912778, + 0.003392704762518406, + 0.002339320257306099, + 0.0036322460509836674, + 0.0025858187582343817, + 0.0030080857686698437, + 0.002184544689953327, + 0.0025542331859469414, + 0.004664494190365076, + 0.006426983047276735, + 0.004771547857671976, + 0.0043290224857628345, + 0.0073260096833109856, + 0.007438939996063709, + 0.007657469715923071, + 0.004268080461770296, + 0.009147703647613525, + 0.008083241991698742, + 0.008439118973910809, + 0.01073205005377531, + 0.008853452280163765, + 0.00926533155143261, + 0.014015954919159412, + 0.013122101314365864, + 0.006568874232470989, + 0.009767715819180012, + 0.004425769206136465, + 0.006910770665854216, + 0.011065545491874218, + 0.004954745061695576, + 0.007253481075167656, + 0.008523166179656982, + 0.015221260488033295, + 0.015942255035042763, + 0.015207921154797077, + 0.017466822639107704, + 0.020013470202684402, + 0.04029606282711029, + 0.02261532098054886, + 0.020121905952692032, + 0.013424200937151909, + 0.020324354991316795, + 0.01014446746557951, + 0.010870173573493958, + 0.012687631882727146, + 0.009960467927157879, + 0.009404249489307404, + 0.021473361179232597, + 0.007713783532381058, + 0.011512873694300652, + 0.008292525075376034, + 0.005585346836596727, + 0.006108065135776997, + 0.01905263401567936, + 0.01939522475004196, + 0.005679689347743988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -1.0539089441299438 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "rapid montage of surreal, high-resolution stills shot from dynamic low angles with a telephoto lens freezes a diverse group of marathon runners in their final, agonizing sprint toward the finish line. Each flashing frame focuses on a different runner, their face a raw canvas of remarkably determined determination: brows furrowed, sweat dripping from temples and chins, jaws clenched, eyes narrowed, unwavering focus on the goal beyond the frame. Their sweat-soaked tracksuits, vibrant in neon yellow, electric blue, and bold red, cling to their lean, muscular bodies, their race bibs pinned to their chests. The camera captures a series of powerful micro-moments: the explosive push of their feet, the splash of water on the wet asphalt; the tense tension of their calf muscles in the harsh light; the decisive pumping of their arms through the air. In the background, the cheering crowd and finish line structure are rendered in a soft, colorful bokeh, emphasizing the athletes' solitary struggle. The scene is lit by bright, direct overhead sunlight, creating high contrast and sharp shadows that make every drop of sweat and strained sinew stand out with crystal clarity. The atmosphere is intensely visceral, filled with the silent yet powerful emotion of human endurance.", + "seed": 15353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_11/index_177_seed_15353_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.5200599920935929, + "vlm_logprob": -0.5200599920935929, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.5200599920935929, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015726236626505852 + ], + "opposite_logprob_tokens": [ + -4.390726089477539 + ], + "logprob_reward": -0.015726236626505852, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015726236626505852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1845860630273819, + 0.1146848127245903, + 0.10755090415477753, + 0.09732875972986221, + 0.10555816441774368, + 0.11126143485307693, + 0.13499067723751068, + 0.07936859875917435, + 0.04316719248890877, + 0.02150336280465126 + ], + "credit_map": [ + 0.019239915534853935, + 0.030872000381350517, + 0.02122187614440918, + 0.017632417380809784, + 0.029170121997594833, + 0.02617163397371769, + 0.022410046309232712, + 0.017868049442768097, + 0.010886028409004211, + 0.013110734522342682, + 0.019868629053235054, + 0.012059714645147324, + 0.011141479015350342, + 0.006986088119447231, + 0.014106808230280876, + 0.026525333523750305, + 0.013184754177927971, + 0.014491356909275055, + 0.010983163490891457, + 0.01171304751187563, + 0.017757387831807137, + 0.014175447635352612, + 0.013550580479204655, + 0.01169516146183014, + 0.01775720901787281, + 0.010450350120663643, + 0.010893985629081726, + 0.013148284517228603, + 0.014517311938107014, + 0.010450270026922226, + 0.009504424408078194, + 0.010606921277940273, + 0.011012143455445766, + 0.017222708091139793, + 0.008477708324790001, + 0.012366550043225288, + 0.013269685208797455, + 0.014547662809491158, + 0.011159426532685757, + 0.017502276226878166, + 0.014225220307707787, + 0.010446353815495968, + 0.01157500222325325, + 0.02434384450316429, + 0.014477898366749287, + 0.012122721411287785, + 0.014460849575698376, + 0.009609539061784744, + 0.04833274334669113, + 0.01241127960383892, + 0.004450258798897266, + 0.007292720954865217, + 0.03076614998281002, + 0.01027067843824625, + 0.004847669508308172, + 0.016619181260466576, + 0.00973250437527895, + 0.015400591306388378, + 0.00866713933646679, + 0.006951084826141596, + 0.007497676182538271, + 0.011421614326536655, + 0.004564598202705383, + 0.015133390203118324, + 0.004765212535858154, + 0.005006685387343168, + 0.00419590575620532, + 0.004995344206690788, + 0.00605933228507638, + 0.005781985819339752, + 0.008086124435067177, + 0.00427660159766674, + 0.0033464941661804914, + 0.0023567243479192257, + 0.0024889998603612185, + 0.0022207116708159447, + 0.0033703113440424204, + 0.0022910188417881727, + 0.0032391075510531664, + 0.0021899957209825516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.015726236626505852 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22552527487277985 + ], + "opposite_logprob_tokens": [ + -1.6005252599716187 + ], + "logprob_reward": -0.22552527487277985, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22552527487277985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12927401065826416, + 0.09032102674245834, + 0.058047451078891754, + 0.04983782768249512, + 0.1646372228860855, + 0.1540510654449463, + 0.19937923550605774, + 0.060989752411842346, + 0.05946596339344978, + 0.033996473997831345 + ], + "credit_map": [ + 0.011602089740335941, + 0.01604907028377056, + 0.018685417249798775, + 0.017177367582917213, + 0.018257545307278633, + 0.022692328318953514, + 0.013000523671507835, + 0.011809675954282284, + 0.01256598625332117, + 0.008585900999605656, + 0.013562380336225033, + 0.010684225708246231, + 0.007909727282822132, + 0.006927154492586851, + 0.012324145063757896, + 0.017761508002877235, + 0.010097344405949116, + 0.007612595800310373, + 0.005785745568573475, + 0.005211147014051676, + 0.007814639247953892, + 0.00729529932141304, + 0.009225922636687756, + 0.0050047617405653, + 0.006545762065798044, + 0.006463245023041964, + 0.005865634419023991, + 0.009071535430848598, + 0.006156235001981258, + 0.0051027643494307995, + 0.005586452316492796, + 0.005046200938522816, + 0.017407268285751343, + 0.023576555773615837, + 0.018160639330744743, + 0.01843593455851078, + 0.01815466396510601, + 0.0234215147793293, + 0.016158824786543846, + 0.029321834444999695, + 0.018829040229320526, + 0.01785610243678093, + 0.015310344286262989, + 0.03299478441476822, + 0.019936291500926018, + 0.01661592163145542, + 0.019698554649949074, + 0.012810024432837963, + 0.018886016681790352, + 0.07986442744731903, + 0.005006671417504549, + 0.005532707087695599, + 0.06047588214278221, + 0.011711234226822853, + 0.005950842518359423, + 0.011951453983783722, + 0.008826085366308689, + 0.006482942495495081, + 0.006316867657005787, + 0.006031759548932314, + 0.008212434127926826, + 0.009433272294700146, + 0.005810897331684828, + 0.009875492192804813, + 0.007722223177552223, + 0.009668531827628613, + 0.007539854850620031, + 0.007955140434205532, + 0.007145553827285767, + 0.00743041792884469, + 0.00724544283002615, + 0.004758797585964203, + 0.003878719871863723, + 0.0032797602470964193, + 0.0034525925293564796, + 0.0026017038617283106, + 0.005113436374813318, + 0.005941899958997965, + 0.004355887416750193, + 0.005372475832700729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.22552527487277985 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953436251729727 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953436251729727, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953436251729727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11935155838727951, + 0.04035814478993416, + 0.06946863979101181, + 0.04788503423333168, + 0.11096982657909393, + 0.15917940437793732, + 0.3152669370174408, + 0.05191367492079735, + 0.046705473214387894, + 0.03890126198530197 + ], + "credit_map": [ + 0.010441369377076626, + 0.0193156898021698, + 0.01602785289287567, + 0.012370200827717781, + 0.01577424257993698, + 0.01150420494377613, + 0.021339276805520058, + 0.012578723952174187, + 0.005521477200090885, + 0.005292851477861404, + 0.004051370080560446, + 0.005373876541852951, + 0.006532004568725824, + 0.004138249438256025, + 0.004587180912494659, + 0.004861138295382261, + 0.010157263837754726, + 0.007381681352853775, + 0.008126560598611832, + 0.005488820374011993, + 0.009697040542960167, + 0.007307048887014389, + 0.01332774106413126, + 0.00798249151557684, + 0.013439543545246124, + 0.004833346698433161, + 0.006159576121717691, + 0.003673298517242074, + 0.007475036196410656, + 0.005611997097730637, + 0.002952286973595619, + 0.003739953273907304, + 0.012992323376238346, + 0.01473612803965807, + 0.013647557236254215, + 0.013386432081460953, + 0.013141512870788574, + 0.014713532291352749, + 0.010366244241595268, + 0.017986109480261803, + 0.01854025200009346, + 0.01983361504971981, + 0.015056941658258438, + 0.033919766545295715, + 0.02123517543077469, + 0.01750529743731022, + 0.020554915070533752, + 0.012533444911241531, + 0.038531992584466934, + 0.0865478515625, + 0.007700761780142784, + 0.009159249253571033, + 0.13123346865177155, + 0.016301332041621208, + 0.00837714597582817, + 0.017415190115571022, + 0.008272417820990086, + 0.013847826980054379, + 0.005620595999062061, + 0.0045026917941868305, + 0.004706891253590584, + 0.005838310811668634, + 0.0044388859532773495, + 0.004686058033257723, + 0.005105820018798113, + 0.006476751994341612, + 0.007369453087449074, + 0.006601615343242884, + 0.006164315156638622, + 0.005345616955310106, + 0.0055146911181509495, + 0.0041272141970694065, + 0.005493576172739267, + 0.003649276215583086, + 0.003295895643532276, + 0.00166141742374748, + 0.010053908452391624, + 0.010376648977398872, + 0.0023194877430796623, + 0.002051054500043392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.005953436251729727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818542532622814 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818542532622814, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818542532622814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15151147544384003, + 0.189837247133255, + 0.044521763920784, + 0.1782747209072113, + 0.07774084806442261, + 0.17222096025943756, + 0.10984106361865997, + 0.020663853734731674, + 0.03019486740231514, + 0.025193188339471817 + ], + "credit_map": [ + 0.012662496417760849, + 0.02268410101532936, + 0.017482556402683258, + 0.01336947362869978, + 0.026211630553007126, + 0.01984298974275589, + 0.022811612114310265, + 0.016446616500616074, + 0.018857920542359352, + 0.03615952655673027, + 0.021231181919574738, + 0.020406825467944145, + 0.022786002606153488, + 0.020681584253907204, + 0.02368597686290741, + 0.02602822706103325, + 0.004863370209932327, + 0.007176663260906935, + 0.0060384850949049, + 0.003974894993007183, + 0.0061718495562672615, + 0.007652380038052797, + 0.0047636558301746845, + 0.0038804628420621157, + 0.02427666075527668, + 0.032733332365751266, + 0.01220728363841772, + 0.01380984764546156, + 0.04447463899850845, + 0.02349230833351612, + 0.010430244728922844, + 0.01685039885342121, + 0.007668694015592337, + 0.012659162282943726, + 0.009464006870985031, + 0.007519276812672615, + 0.011549999937415123, + 0.010429037734866142, + 0.007908392697572708, + 0.010542281903326511, + 0.02116447500884533, + 0.019825663417577744, + 0.016553319990634918, + 0.03650566190481186, + 0.02320374920964241, + 0.0188088808208704, + 0.022180603817105293, + 0.013978599570691586, + 0.013770010322332382, + 0.03590068593621254, + 0.003260673489421606, + 0.003955516964197159, + 0.036781881004571915, + 0.005468331277370453, + 0.003618682501837611, + 0.007085284683853388, + 0.0025801199954003096, + 0.0023898265790194273, + 0.002542853122577071, + 0.0018753628246486187, + 0.0029942630790174007, + 0.0032717136200517416, + 0.0021571707911789417, + 0.0028525430243462324, + 0.002792840590700507, + 0.0026943786069750786, + 0.004851556848734617, + 0.004289351403713226, + 0.004028335213661194, + 0.004674389958381653, + 0.0038652289658784866, + 0.0029987848829478025, + 0.004368335008621216, + 0.0030574991833418608, + 0.0025326095055788755, + 0.001752172363922, + 0.004169078543782234, + 0.005071711260825396, + 0.0023719770833849907, + 0.0018698039930313826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.002818542532622814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.350276470184326 + ], + "opposite_logprob_tokens": [ + -0.10027646273374557 + ], + "logprob_reward": -2.350276470184326, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.350276470184326, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.29485583305358887, + 0.07469293475151062, + 0.0987367257475853, + 0.16267475485801697, + 0.04544608294963837, + 0.021465210244059563, + 0.19601884484291077, + 0.03536100685596466, + 0.04404280707240105, + 0.026705821976065636 + ], + "credit_map": [ + 0.026575138792395592, + 0.03564586117863655, + 0.033261820673942566, + 0.043864745646715164, + 0.03292675316333771, + 0.04626259207725525, + 0.04812372475862503, + 0.0281952116638422, + 0.011270360089838505, + 0.008700643666088581, + 0.00965452753007412, + 0.0054693035781383514, + 0.010946180671453476, + 0.0076010823249816895, + 0.009409673511981964, + 0.011641163378953934, + 0.012127604335546494, + 0.010918463580310345, + 0.009476680308580399, + 0.011387349106371403, + 0.015990132465958595, + 0.011737866327166557, + 0.016425980255007744, + 0.010672647505998611, + 0.023756155744194984, + 0.027211325243115425, + 0.01196100190281868, + 0.01906672492623329, + 0.0353255569934845, + 0.018119527027010918, + 0.011919242329895496, + 0.015315218828618526, + 0.006049418356269598, + 0.005365305580198765, + 0.008623339235782623, + 0.0052471826784312725, + 0.0049230922013521194, + 0.005674687214195728, + 0.004424479324370623, + 0.005138580687344074, + 0.002829131204634905, + 0.002500993199646473, + 0.002173847984522581, + 0.003994580823928118, + 0.0025809290818870068, + 0.002399977995082736, + 0.0024500207509845495, + 0.0025357294362038374, + 0.014591298066079617, + 0.07317814230918884, + 0.005822066683322191, + 0.005543428007513285, + 0.07183990627527237, + 0.010221259668469429, + 0.006786272395402193, + 0.00803646445274353, + 0.004788305144757032, + 0.005217238795012236, + 0.006823087111115456, + 0.002945659914985299, + 0.004253590013831854, + 0.004827733151614666, + 0.0030209592077881098, + 0.0034844335168600082, + 0.005194855388253927, + 0.005389938596636057, + 0.006474136840552092, + 0.0056743323802948, + 0.006007872521877289, + 0.0059187025763094425, + 0.005825813859701157, + 0.003557157702744007, + 0.00448610819876194, + 0.002759364666417241, + 0.0022390105295926332, + 0.0021322742104530334, + 0.004799850285053253, + 0.004132501315325499, + 0.0032378926407545805, + 0.0029188194312155247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -2.350276470184326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "rapid montage of surreal, high-resolution stills shot from dynamic low angles with a telephoto lens freezes a diverse group of marathon runners in their final, agonizing sprint toward the finish line. Each flashing frame focuses on a different runner, their face a raw canvas of remarkably determined determination: brows furrowed, sweat dripping from temples and chins, jaws clenched, eyes narrowed, unwavering focus on the goal beyond the frame. Their sweat-soaked tracksuits, vibrant in neon yellow, electric blue, and bold red, cling to their lean, muscular bodies, their race bibs pinned to their chests. The camera captures a series of powerful micro-moments: the explosive push of their feet, the splash of water on the wet asphalt; the tense tension of their calf muscles in the harsh light; the decisive pumping of their arms through the air. In the background, the cheering crowd and finish line structure are rendered in a soft, colorful bokeh, emphasizing the athletes' solitary struggle. The scene is lit by bright, direct overhead sunlight, creating high contrast and sharp shadows that make every drop of sweat and strained sinew stand out with crystal clarity. The atmosphere is intensely visceral, filled with the silent yet powerful emotion of human endurance.", + "seed": 15354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_11/index_177_seed_15354_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.25956870447844266, + "vlm_logprob": -0.25956870447844266, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.25956870447844266, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033386506140232086 + ], + "opposite_logprob_tokens": [ + -3.533386468887329 + ], + "logprob_reward": -0.033386506140232086, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.033386506140232086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04818913713097572, + 0.17293095588684082, + 0.04614345356822014, + 0.05495784059166908, + 0.08863186836242676, + 0.18489380180835724, + 0.09555398672819138, + 0.05923350527882576, + 0.13154293596744537, + 0.11792252212762833 + ], + "credit_map": [ + 0.005979796871542931, + 0.00452616298571229, + 0.007441266439855099, + 0.005560680758208036, + 0.004643662367016077, + 0.011087666265666485, + 0.003688349388539791, + 0.005261552985757589, + 0.025083418935537338, + 0.011142358183860779, + 0.014026746153831482, + 0.06193099915981293, + 0.013011694885790348, + 0.012553507462143898, + 0.013499831780791283, + 0.02168240211904049, + 0.004809889476746321, + 0.004892012570053339, + 0.005080357193946838, + 0.003560922807082534, + 0.006520305760204792, + 0.005433682352304459, + 0.01160344760864973, + 0.004242831841111183, + 0.006104408297687769, + 0.008801568299531937, + 0.009282941929996014, + 0.006400485057383776, + 0.007104974240064621, + 0.004710452165454626, + 0.007065626326948404, + 0.005487383808940649, + 0.013125770725309849, + 0.014432933181524277, + 0.011377432383596897, + 0.013817830011248589, + 0.00926471408456564, + 0.006561760324984789, + 0.008285782299935818, + 0.011765646748244762, + 0.020102621987462044, + 0.01709439419209957, + 0.008536180481314659, + 0.011130048893392086, + 0.007756172679364681, + 0.06477247178554535, + 0.009657852351665497, + 0.04584405943751335, + 0.011786217801272869, + 0.009684315882623196, + 0.011850862763822079, + 0.009154019877314568, + 0.006790155544877052, + 0.012921242043375969, + 0.017244916409254074, + 0.016122259199619293, + 0.006633848417550325, + 0.005695200525224209, + 0.0073065138421952724, + 0.007588796317577362, + 0.007863730192184448, + 0.007109459023922682, + 0.009322999976575375, + 0.00771295977756381, + 0.026587530970573425, + 0.0283319354057312, + 0.015281048603355885, + 0.015601816587150097, + 0.0198166836053133, + 0.008941576816141605, + 0.010684795677661896, + 0.006297546438872814, + 0.011052909307181835, + 0.013298316858708858, + 0.011438759043812752, + 0.01194884441792965, + 0.016018826514482498, + 0.020476412028074265, + 0.0111317727714777, + 0.022556684911251068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.033386506140232086 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3133876621723175 + ], + "opposite_logprob_tokens": [ + -1.3133876323699951 + ], + "logprob_reward": -0.3133876621723175, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3133876621723175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11447548866271973, + 0.12561534345149994, + 0.08604750782251358, + 0.059192534536123276, + 0.0511217899620533, + 0.27367308735847473, + 0.10726255178451538, + 0.04452702775597572, + 0.09189733117818832, + 0.04618729650974274 + ], + "credit_map": [ + 0.015822766348719597, + 0.01203597616404295, + 0.020827265456318855, + 0.011893521063029766, + 0.011524423025548458, + 0.022726479917764664, + 0.008860727772116661, + 0.01078434195369482, + 0.016886962577700615, + 0.008985236287117004, + 0.009821764193475246, + 0.04605808109045029, + 0.008919362910091877, + 0.01020942721515894, + 0.008982149884104729, + 0.015752363950014114, + 0.009001036174595356, + 0.010279078036546707, + 0.0067818909883499146, + 0.008957537822425365, + 0.01187853142619133, + 0.0102354371920228, + 0.019364874809980392, + 0.009549127891659737, + 0.005235051270574331, + 0.006396918091922998, + 0.005878188647329807, + 0.007125255186110735, + 0.010075668804347515, + 0.0065194228664040565, + 0.0055268071591854095, + 0.01243522297590971, + 0.007423229981213808, + 0.005605020560324192, + 0.0062424661591649055, + 0.0034363127779215574, + 0.006186029873788357, + 0.007397902198135853, + 0.0046354979276657104, + 0.010195331647992134, + 0.04419919103384018, + 0.03407644107937813, + 0.019207045435905457, + 0.023992590606212616, + 0.02005285769701004, + 0.044819425791502, + 0.022443752735853195, + 0.0648818165063858, + 0.0168670192360878, + 0.012986813671886921, + 0.013484074734151363, + 0.011614112183451653, + 0.014065484516322613, + 0.011888786219060421, + 0.017455682158470154, + 0.008900577202439308, + 0.007075127679854631, + 0.00707079516723752, + 0.0057238442823290825, + 0.004760940093547106, + 0.005316837225109339, + 0.005095842760056257, + 0.00524379825219512, + 0.004239845089614391, + 0.0144509831443429, + 0.014888924546539783, + 0.014513117261230946, + 0.010175560601055622, + 0.009023809805512428, + 0.011355518363416195, + 0.009340834803879261, + 0.008148590102791786, + 0.005032161250710487, + 0.004858986474573612, + 0.004942379891872406, + 0.004526844248175621, + 0.006743133999407291, + 0.007349689491093159, + 0.0047538746148347855, + 0.007980233058333397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.3133876621723175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865103214979172 + ], + "opposite_logprob_tokens": [ + -3.0486509799957275 + ], + "logprob_reward": -0.04865103214979172, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865103214979172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15017129480838776, + 0.26517805457115173, + 0.10563219338655472, + 0.07656310498714447, + 0.025625500828027725, + 0.12123766541481018, + 0.11662620306015015, + 0.06491582095623016, + 0.03721389174461365, + 0.036836251616477966 + ], + "credit_map": [ + 0.02640710398554802, + 0.009854542091488838, + 0.016542978584766388, + 0.014048810116946697, + 0.01618964783847332, + 0.044823575764894485, + 0.008818352594971657, + 0.01348628755658865, + 0.027431149035692215, + 0.015913719311356544, + 0.01806257851421833, + 0.09947698563337326, + 0.019754374399781227, + 0.01593313179910183, + 0.01988568902015686, + 0.048720452934503555, + 0.01037783082574606, + 0.01380064059048891, + 0.00906249601393938, + 0.011361680924892426, + 0.017074603587388992, + 0.01177303772419691, + 0.022477727383375168, + 0.009704186581075191, + 0.00806692335754633, + 0.009399321861565113, + 0.006782223470509052, + 0.007035501766949892, + 0.015316829085350037, + 0.008352723903954029, + 0.0065964157693088055, + 0.015013170428574085, + 0.0037181545048952103, + 0.0029511861503124237, + 0.002932286821305752, + 0.00258128484711051, + 0.0034485335927456617, + 0.002574499463662505, + 0.0033192913979291916, + 0.004100262653082609, + 0.00977952778339386, + 0.00869643222540617, + 0.007120993919670582, + 0.007384399883449078, + 0.004986077547073364, + 0.024893449619412422, + 0.01068202406167984, + 0.04769476503133774, + 0.01869923807680607, + 0.014666210860013962, + 0.013297059573233128, + 0.012382137589156628, + 0.015495280735194683, + 0.009462352842092514, + 0.023482445627450943, + 0.009141483344137669, + 0.010488087311387062, + 0.008920202031731606, + 0.0064040240831673145, + 0.0062787183560431, + 0.010007064789533615, + 0.008679434657096863, + 0.0067391288466751575, + 0.007399155758321285, + 0.006576356012374163, + 0.0041365670040249825, + 0.006009996403008699, + 0.004353892523795366, + 0.002639373764395714, + 0.004905174020677805, + 0.004606181290000677, + 0.003986351657658815, + 0.004852653946727514, + 0.003833506954833865, + 0.0050131832249462605, + 0.0036643047351390123, + 0.004076393786817789, + 0.004701313562691212, + 0.00452260859310627, + 0.006172284483909607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.04865103214979172 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095382057130337 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095382057130337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095382057130337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16013532876968384, + 0.1797863245010376, + 0.03200225159525871, + 0.03702681139111519, + 0.16556739807128906, + 0.09298790246248245, + 0.06160507723689079, + 0.041937220841646194, + 0.08842238783836365, + 0.14052928984165192 + ], + "credit_map": [ + 0.03862452879548073, + 0.01131489872932434, + 0.012906115502119064, + 0.008479212410748005, + 0.009087427519261837, + 0.06385280936956406, + 0.006868121679872274, + 0.009002229198813438, + 0.025809530168771744, + 0.011522707529366016, + 0.012367956340312958, + 0.06736671179533005, + 0.013102971948683262, + 0.008940916508436203, + 0.013260511681437492, + 0.027414998039603233, + 0.004335799254477024, + 0.004292771220207214, + 0.003181443316861987, + 0.002612500451505184, + 0.005581267643719912, + 0.003927520476281643, + 0.005281961988657713, + 0.0027889867778867483, + 0.003741806373000145, + 0.004245768766850233, + 0.0035691391676664352, + 0.0035061200615018606, + 0.007585977204144001, + 0.0028977668844163418, + 0.0026013569440692663, + 0.008878879249095917, + 0.022262144833803177, + 0.03348679840564728, + 0.03196348994970322, + 0.019936775788664818, + 0.018445715308189392, + 0.009847396053373814, + 0.013979865238070488, + 0.0156452227383852, + 0.006972049362957478, + 0.00865999050438404, + 0.007642722688615322, + 0.007601588498800993, + 0.007167204283177853, + 0.018835457041859627, + 0.006962309125810862, + 0.029146580025553703, + 0.007826452143490314, + 0.00602843938395381, + 0.007421744987368584, + 0.006736858282238245, + 0.005187297705560923, + 0.006442667916417122, + 0.00945263635367155, + 0.012508982792496681, + 0.007842829450964928, + 0.005661588162183762, + 0.005224800668656826, + 0.003481747815385461, + 0.004068916197866201, + 0.004492613021284342, + 0.007036437746137381, + 0.004128291737288237, + 0.011582911014556885, + 0.013766799122095108, + 0.012222074903547764, + 0.013776644133031368, + 0.007466960698366165, + 0.01146533340215683, + 0.008911967277526855, + 0.009229700081050396, + 0.014649748802185059, + 0.016479559242725372, + 0.01521226204931736, + 0.014670085161924362, + 0.01746668666601181, + 0.024485137313604355, + 0.014680900610983372, + 0.022884901612997055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.004095382057130337 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of a rapid montage of still images rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the subjects shown as a diverse group of marathon runners in a sprinting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the runners wear sweat-soaked tracksuits in neon yellow, electric blue, or bold red colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered with a soft, colorful bokeh effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, direct overhead sunlight creating high contrast and sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983229398727417 + ], + "opposite_logprob_tokens": [ + -0.5233229398727417 + ], + "logprob_reward": -0.8983229398727417, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983229398727417, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4566802382469177, + 0.24978400766849518, + 0.025722788646817207, + 0.013906151987612247, + 0.03580131009221077, + 0.018749477341771126, + 0.01592087559401989, + 0.026205962523818016, + 0.05744205415248871, + 0.09978719055652618 + ], + "credit_map": [ + 0.11481539905071259, + 0.031246518716216087, + 0.03255489841103554, + 0.017254674807190895, + 0.020858677104115486, + 0.20307032763957977, + 0.013014547526836395, + 0.023865165188908577, + 0.03152833878993988, + 0.016002509742975235, + 0.017661865800619125, + 0.09403872489929199, + 0.018044792115688324, + 0.017613191157579422, + 0.019536565989255905, + 0.035357989370822906, + 0.002593905432149768, + 0.0030426857993006706, + 0.002718317788094282, + 0.00197398429736495, + 0.004649067763239145, + 0.003087428165599704, + 0.005010966677218676, + 0.002646429231390357, + 0.0015803955029696226, + 0.002483651740476489, + 0.0017531277844682336, + 0.0018879900453612208, + 0.0017884011613205075, + 0.0015108553925529122, + 0.0015342282131314278, + 0.0013675004011020064, + 0.006067704875022173, + 0.0053762285970151424, + 0.0068861632607877254, + 0.0036810538731515408, + 0.0034966806415468454, + 0.003738601692020893, + 0.00315624731592834, + 0.0033986263442784548, + 0.0029624649323523045, + 0.002567300572991371, + 0.001804879866540432, + 0.0020026573911309242, + 0.0016320492140948772, + 0.004160077776759863, + 0.001975438091903925, + 0.0016446076333522797, + 0.0023453908506780863, + 0.002161833457648754, + 0.0028328122571110725, + 0.0014400907093659043, + 0.0013191570760682225, + 0.0014488346641883254, + 0.0026931054890155792, + 0.001679649343714118, + 0.0042134979739785194, + 0.003138722851872444, + 0.002379754325374961, + 0.0018436110112816095, + 0.003412621794268489, + 0.005259419325739145, + 0.0035330296959728003, + 0.0024253029841929674, + 0.009900121949613094, + 0.012591655366122723, + 0.007722224108874798, + 0.007117186672985554, + 0.007389240898191929, + 0.0047422354109585285, + 0.004757132846862078, + 0.0032222566660493612, + 0.009936254471540451, + 0.011821236461400986, + 0.009438278153538704, + 0.010589194484055042, + 0.013871507719159126, + 0.0168034415692091, + 0.009154140949249268, + 0.018173128366470337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.8983229398727417 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_12/index_tensor([7303, 7303])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_12/index_tensor([7303, 7303])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..372345397cabe2f4ce564b7e9c696d07b2756be6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_12/index_tensor([7303, 7303])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, fixed camera holds a steady medium-wide shot across a stark, featureless white void, illuminated by a brilliant, shadowless, diffuse light that fills the infinite, clinical space. A luxury touring motorcycle, reimagined in the surreal, chrome-infused style of Sorayama, moves within the frame. Its body is a single, impossibly smooth form of liquid-polished chrome, devoid of a visible engine, reflecting the distortions of its surroundings with mirror-like perfection in its stark whiteness. Seamlessly blending into the machine's chassis is a gynoid figure, her lithe, metallic form leaning forward with hands integrated into the armrests, her polished face expressionless as her entire body remains completely still. The motorcycle suddenly collides with an invisible wall, and as the camera remains still, the impact triggers a surreal deformation. The front of the chrome machine deflects viscously, its contact points twisting and rippling outward as the solid metal begins to slowly and deliberately ooze from the invisible obstacle, stretching and twisting like liquid mercury. The giantess's fused body twists in unison, her polished form pulling and flowing with the destruction of the bike, creating a captivating atmosphere of cold, silent, bizarre physics and fluid deconstruction.", + "seed": 6555, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_12/index_7303_seed_6555_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03587587359361351, + "vlm_logprob": -0.03587587359361351, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03587587359361351, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11389477550983429 + ], + "opposite_logprob_tokens": [ + -2.3638947010040283 + ], + "logprob_reward": -0.11389477550983429, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11389477550983429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11107185482978821, + 0.1030169427394867, + 0.10538658499717712, + 0.16874337196350098, + 0.1135951429605484, + 0.09881765395402908, + 0.07473108917474747, + 0.034303173422813416, + 0.12526430189609528, + 0.06506988406181335 + ], + "credit_map": [ + 0.014114993624389172, + 0.009569568559527397, + 0.00969325564801693, + 0.012912226840853691, + 0.013425275683403015, + 0.015822554007172585, + 0.02037126012146473, + 0.01516273245215416, + 0.009305162355303764, + 0.010192777030169964, + 0.010366751812398434, + 0.008830998092889786, + 0.013477740809321404, + 0.017533767968416214, + 0.01692948490381241, + 0.016380267217755318, + 0.012220676988363266, + 0.009517652913928032, + 0.011913308873772621, + 0.008289874531328678, + 0.015893395990133286, + 0.01434553787112236, + 0.015717357397079468, + 0.01748877950012684, + 0.013174137100577354, + 0.011029344983398914, + 0.013280197978019714, + 0.024613438174128532, + 0.015656717121601105, + 0.02491016872227192, + 0.03642565384507179, + 0.029653731733560562, + 0.021575985476374626, + 0.00836941972374916, + 0.01141485944390297, + 0.007125427480787039, + 0.008126506581902504, + 0.019241902977228165, + 0.008617691695690155, + 0.029123356565833092, + 0.014618800021708012, + 0.010033431462943554, + 0.009836249984800816, + 0.010141339153051376, + 0.017731014639139175, + 0.011087065562605858, + 0.009874389506876469, + 0.01549536269158125, + 0.013755707070231438, + 0.006083155050873756, + 0.006352163385599852, + 0.005415455438196659, + 0.011656228452920914, + 0.015029222704470158, + 0.008700123988091946, + 0.007739040069282055, + 0.0052639273926615715, + 0.0033275410532951355, + 0.004323979839682579, + 0.0036857344675809145, + 0.0036410633474588394, + 0.004040637984871864, + 0.005076739937067032, + 0.004943552426993847, + 0.02328520640730858, + 0.007852167822420597, + 0.011525132693350315, + 0.010427499189972878, + 0.011978148482739925, + 0.029385633766651154, + 0.01279461570084095, + 0.018015900626778603, + 0.009138817898929119, + 0.004395267926156521, + 0.00649397075176239, + 0.006835888605564833, + 0.007057474926114082, + 0.013627449981868267, + 0.007511913310736418, + 0.010009100660681725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.11389477550983429 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814113512635231 + ], + "opposite_logprob_tokens": [ + -3.2881412506103516 + ], + "logprob_reward": -0.03814113512635231, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814113512635231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12895187735557556, + 0.17553052306175232, + 0.07683470845222473, + 0.050489384680986404, + 0.08821701258420944, + 0.12744340300559998, + 0.07598286867141724, + 0.04901980981230736, + 0.14800512790679932, + 0.07952534407377243 + ], + "credit_map": [ + 0.010714538395404816, + 0.009527082554996014, + 0.007892446592450142, + 0.01129066850990057, + 0.014808944426476955, + 0.026304306462407112, + 0.017365243285894394, + 0.031048623844981194, + 0.014454330317676067, + 0.02217591367661953, + 0.02258361130952835, + 0.020386910066008568, + 0.01854129135608673, + 0.020480340346693993, + 0.03598989173769951, + 0.02091822400689125, + 0.005605135578662157, + 0.0035506715066730976, + 0.0033546732738614082, + 0.003425221424549818, + 0.0075153871439397335, + 0.020557286217808723, + 0.009568183682858944, + 0.023258134722709656, + 0.0057464465498924255, + 0.003944457974284887, + 0.004413095768541098, + 0.006830894388258457, + 0.006422713398933411, + 0.005668568424880505, + 0.010060671716928482, + 0.007402530871331692, + 0.012572496198117733, + 0.00482206791639328, + 0.011005968786776066, + 0.0058397105894982815, + 0.00676564360037446, + 0.01651598885655403, + 0.011504259891808033, + 0.019190866500139236, + 0.013356257230043411, + 0.014878838323056698, + 0.015779512003064156, + 0.015288487076759338, + 0.0253670122474432, + 0.015538353472948074, + 0.0148592134937644, + 0.012375714257359505, + 0.011592370457947254, + 0.007780901156365871, + 0.010745897889137268, + 0.006738458760082722, + 0.010237167589366436, + 0.008361565880477428, + 0.010087202303111553, + 0.010439293459057808, + 0.006554719991981983, + 0.003878614166751504, + 0.0065303985029459, + 0.004819564986974001, + 0.008249716833233833, + 0.0063604507595300674, + 0.005795496981590986, + 0.006830844562500715, + 0.019678374752402306, + 0.009383867494761944, + 0.01658339984714985, + 0.017783528193831444, + 0.015898726880550385, + 0.026253145188093185, + 0.020228775218129158, + 0.022195301949977875, + 0.010664764791727066, + 0.00689293909817934, + 0.010423490777611732, + 0.008337081409990788, + 0.009350431151688099, + 0.011652681045234203, + 0.012880302034318447, + 0.009323645383119583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.03814113512635231 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958294961601496 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958294961601496, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958294961601496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11271516978740692, + 0.15820538997650146, + 0.06968752294778824, + 0.0792742520570755, + 0.07727498561143875, + 0.07506022602319717, + 0.0765594020485878, + 0.058841779828071594, + 0.21811820566654205, + 0.07426311820745468 + ], + "credit_map": [ + 0.012672250159084797, + 0.012157882563769817, + 0.012015418149530888, + 0.01587495394051075, + 0.016260121017694473, + 0.013000819832086563, + 0.0178755484521389, + 0.01285815704613924, + 0.01333913765847683, + 0.017509261146187782, + 0.01728338561952114, + 0.014931563287973404, + 0.016484614461660385, + 0.026187261566519737, + 0.0269259475171566, + 0.025544190779328346, + 0.006661964114755392, + 0.0029816909227520227, + 0.0047960723750293255, + 0.0035914338659495115, + 0.005413551814854145, + 0.015843281522393227, + 0.009891995228827, + 0.020507527515292168, + 0.007579145021736622, + 0.00848184060305357, + 0.0074363709427416325, + 0.005957101471722126, + 0.012090926058590412, + 0.007828641682863235, + 0.019900787621736526, + 0.009999432601034641, + 0.007750203832983971, + 0.00624085683375597, + 0.01297683548182249, + 0.007201330736279488, + 0.008459205739200115, + 0.010063683614134789, + 0.01471229363232851, + 0.009870568290352821, + 0.009817755781114101, + 0.007070960011333227, + 0.009876134805381298, + 0.007081519812345505, + 0.012570311315357685, + 0.010488627478480339, + 0.009866039268672466, + 0.008288866840302944, + 0.00988338328897953, + 0.011000905185937881, + 0.006939257960766554, + 0.00634590582922101, + 0.015072472393512726, + 0.007087627425789833, + 0.009770441800355911, + 0.01045939326286316, + 0.007920913398265839, + 0.00556461326777935, + 0.00660582585260272, + 0.005909549072384834, + 0.011814006604254246, + 0.006648995913565159, + 0.006956845056265593, + 0.0074210274033248425, + 0.03251103684306145, + 0.017370574176311493, + 0.022417519241571426, + 0.027727659791707993, + 0.02255563624203205, + 0.03640821576118469, + 0.02849842794239521, + 0.030629118904471397, + 0.01176774688065052, + 0.00528572266921401, + 0.009462458081543446, + 0.0077652884647250175, + 0.008840802125632763, + 0.014274096116423607, + 0.0088966004550457, + 0.007970396429300308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.005958294961601496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819116808474064 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819116808474064, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819116808474064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1579582244157791, + 0.08349772542715073, + 0.05617928504943848, + 0.10783829540014267, + 0.053690675646066666, + 0.05206460505723953, + 0.039447128772735596, + 0.013198532164096832, + 0.3830983638763428, + 0.053027208894491196 + ], + "credit_map": [ + 0.011850006878376007, + 0.007445775903761387, + 0.008233685046434402, + 0.009450254961848259, + 0.012926203198730946, + 0.046571072190999985, + 0.011228397488594055, + 0.05025281012058258, + 0.00820858497172594, + 0.010422449558973312, + 0.009645704180002213, + 0.00955540593713522, + 0.01033216156065464, + 0.009768945164978504, + 0.01654108427464962, + 0.009023375809192657, + 0.005482891108840704, + 0.0035492635797709227, + 0.0061058346182107925, + 0.003548060078173876, + 0.009277471341192722, + 0.007627528626471758, + 0.012107293121516705, + 0.008480934426188469, + 0.009981335140764713, + 0.008732836693525314, + 0.009491930715739727, + 0.011295937933027744, + 0.01757648028433323, + 0.01164792850613594, + 0.02583049237728119, + 0.013281342573463917, + 0.004549618810415268, + 0.004482886288315058, + 0.005206715781241655, + 0.00547068240121007, + 0.00782626960426569, + 0.006623120047152042, + 0.011979608796536922, + 0.007551767863333225, + 0.008807800710201263, + 0.004767030943185091, + 0.006783084943890572, + 0.0037132464349269867, + 0.00754587585106492, + 0.006958257872611284, + 0.006440973840653896, + 0.007048328872770071, + 0.00781296007335186, + 0.0027142318431288004, + 0.006986771244555712, + 0.003029405605047941, + 0.0026241522282361984, + 0.006795408204197884, + 0.004883582703769207, + 0.004600611515343189, + 0.0021070062648504972, + 0.001455154037103057, + 0.001740556675940752, + 0.0013161863898858428, + 0.0016131968004629016, + 0.0015710274456068873, + 0.001616813475266099, + 0.0017785897944122553, + 0.056161001324653625, + 0.0302825178951025, + 0.04029293730854988, + 0.05099264904856682, + 0.0372924879193306, + 0.05873957648873329, + 0.054528575390577316, + 0.054808590561151505, + 0.007810430135577917, + 0.0036491290666162968, + 0.0065139709040522575, + 0.004643714986741543, + 0.00669542932882905, + 0.00751077476888895, + 0.008905700407922268, + 0.0072980523109436035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.01819116808474064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031939942855387926 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.0031939942855387926, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031939942855387926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14827309548854828, + 0.14311029016971588, + 0.044036637991666794, + 0.05092713236808777, + 0.054782480001449585, + 0.07777035236358643, + 0.07172228395938873, + 0.035528454929590225, + 0.32084208726882935, + 0.05300715193152428 + ], + "credit_map": [ + 0.012183664366602898, + 0.006470397114753723, + 0.006980129983276129, + 0.007696972694247961, + 0.011551355943083763, + 0.04259594529867172, + 0.010722202248871326, + 0.05007244274020195, + 0.01326022855937481, + 0.013889436610043049, + 0.013487250544130802, + 0.013509836979210377, + 0.012826257385313511, + 0.02626177668571472, + 0.02395615726709366, + 0.025919361039996147, + 0.0037252046167850494, + 0.003673730418086052, + 0.005216087680310011, + 0.0029801970813423395, + 0.005907106678932905, + 0.007196326740086079, + 0.007288375869393349, + 0.008049611933529377, + 0.0056863706558942795, + 0.004302567802369595, + 0.005811958573758602, + 0.004494211636483669, + 0.008118429221212864, + 0.005831450689584017, + 0.00908189732581377, + 0.007600254379212856, + 0.008918601088225842, + 0.0035945207346230745, + 0.007707290817052126, + 0.0036070863716304302, + 0.0034283066634088755, + 0.007180272601544857, + 0.007189997471868992, + 0.013156413100659847, + 0.010492014698684216, + 0.007315386086702347, + 0.010648826137185097, + 0.006824877578765154, + 0.01298134308308363, + 0.010735117830336094, + 0.009410392493009567, + 0.009362400509417057, + 0.011377697810530663, + 0.009588891640305519, + 0.005862821824848652, + 0.005180086009204388, + 0.014202754944562912, + 0.01032500620931387, + 0.008124268613755703, + 0.007060764357447624, + 0.0035466053523123264, + 0.00399246159940958, + 0.004890795797109604, + 0.004189500119537115, + 0.005855593830347061, + 0.0037799938581883907, + 0.004437669645994902, + 0.004835839383304119, + 0.04587617516517639, + 0.026534317061305046, + 0.03427377715706825, + 0.0444604717195034, + 0.03185136616230011, + 0.04669126123189926, + 0.047060780227184296, + 0.04409397765994072, + 0.007604938931763172, + 0.004362816922366619, + 0.007744248025119305, + 0.005611186847090721, + 0.007728352211415768, + 0.007753516081720591, + 0.006138353142887354, + 0.006063742097467184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0031939942855387926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, fixed camera holds a steady medium-wide shot across a stark, featureless white void, illuminated by a brilliant, shadowless, diffuse light that fills the infinite, clinical space. A luxury touring motorcycle, reimagined in the surreal, chrome-infused style of Sorayama, moves within the frame. Its body is a single, impossibly smooth form of liquid-polished chrome, devoid of a visible engine, reflecting the distortions of its surroundings with mirror-like perfection in its stark whiteness. Seamlessly blending into the machine's chassis is a gynoid figure, her lithe, metallic form leaning forward with hands integrated into the armrests, her polished face expressionless as her entire body remains completely still. The motorcycle suddenly collides with an invisible wall, and as the camera remains still, the impact triggers a surreal deformation. The front of the chrome machine deflects viscously, its contact points twisting and rippling outward as the solid metal begins to slowly and deliberately ooze from the invisible obstacle, stretching and twisting like liquid mercury. The giantess's fused body twists in unison, her polished form pulling and flowing with the destruction of the bike, creating a captivating atmosphere of cold, silent, bizarre physics and fluid deconstruction.", + "seed": 6556, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_12/index_7303_seed_6556_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029420359525829554, + "vlm_logprob": -0.029420359525829554, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.029420359525829554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04339653626084328 + ], + "opposite_logprob_tokens": [ + -3.418396472930908 + ], + "logprob_reward": -0.04339653626084328, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04339653626084328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10167235881090164, + 0.03904837742447853, + 0.02734491601586342, + 0.46975177526474, + 0.23326589167118073, + 0.02531745471060276, + 0.037216830998659134, + 0.01376757025718689, + 0.036598339676856995, + 0.016016440466046333 + ], + "credit_map": [ + 0.0149392606690526, + 0.018141288310289383, + 0.005774096120148897, + 0.006677873898297548, + 0.009741302579641342, + 0.012792122550308704, + 0.006219742354005575, + 0.027386680245399475, + 0.006107660476118326, + 0.0041675507090985775, + 0.0022802739404141903, + 0.002608007052913308, + 0.0035412448924034834, + 0.004393170587718487, + 0.0023068885784596205, + 0.013643580488860607, + 0.0041427589021623135, + 0.002911669434979558, + 0.0019300204003229737, + 0.00236750696785748, + 0.002622917527332902, + 0.003155626356601715, + 0.001672732993029058, + 0.008541682735085487, + 0.14241209626197815, + 0.027467435225844383, + 0.017435921356081963, + 0.01455222349613905, + 0.019339904189109802, + 0.19398318231105804, + 0.015252245590090752, + 0.03930877521634102, + 0.06550483405590057, + 0.012225337326526642, + 0.008596501313149929, + 0.010983331128954887, + 0.010186485946178436, + 0.08526284992694855, + 0.010633264668285847, + 0.029873274266719818, + 0.00592373963445425, + 0.001693873549811542, + 0.0016755061224102974, + 0.0017135731177404523, + 0.0013628688175231218, + 0.007656564004719257, + 0.0021248518023639917, + 0.0031664767302572727, + 0.005830103997141123, + 0.002196389948949218, + 0.0016524548409506679, + 0.0021311475429683924, + 0.003281222190707922, + 0.01012328453361988, + 0.0033029350452125072, + 0.008699294179677963, + 0.0016938312910497189, + 0.0017972177593037486, + 0.0016877014422789216, + 0.002654434647411108, + 0.0017661717720329762, + 0.0011723644565790892, + 0.001721131382510066, + 0.0012747171567752957, + 0.004952788352966309, + 0.0032773118000477552, + 0.0019930952694267035, + 0.002106277970597148, + 0.006233510095626116, + 0.008531244471669197, + 0.0024907500483095646, + 0.007013362366706133, + 0.00219735293649137, + 0.0017614922253414989, + 0.0013762410962954164, + 0.0011068294988945127, + 0.002153210574761033, + 0.0028182382229715586, + 0.0012007239274680614, + 0.0034023502375930548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.04339653626084328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814411908388138 + ], + "opposite_logprob_tokens": [ + -3.288144111633301 + ], + "logprob_reward": -0.03814411908388138, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814411908388138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2035323679447174, + 0.04864184930920601, + 0.06669352948665619, + 0.2351876199245453, + 0.12329819053411484, + 0.05213538929820061, + 0.12104751169681549, + 0.045878272503614426, + 0.07228270173072815, + 0.0313025563955307 + ], + "credit_map": [ + 0.0308865737169981, + 0.03181014955043793, + 0.0132571617141366, + 0.013373827561736107, + 0.019902996718883514, + 0.025134092196822166, + 0.012263254262506962, + 0.05690430477261543, + 0.0060955495573580265, + 0.01065044105052948, + 0.005867623724043369, + 0.006027961149811745, + 0.004513676278293133, + 0.0038310098461806774, + 0.003539964323863387, + 0.008115624077618122, + 0.010630971752107143, + 0.019416337832808495, + 0.004590026568621397, + 0.004714243113994598, + 0.006775822956115007, + 0.007681793998926878, + 0.0034829438664019108, + 0.009401391260325909, + 0.0592975914478302, + 0.030020080506801605, + 0.0090570580214262, + 0.009278546087443829, + 0.013137534260749817, + 0.07856722921133041, + 0.010235648602247238, + 0.025593936443328857, + 0.02827090583741665, + 0.010779859498143196, + 0.007389534264802933, + 0.009420979768037796, + 0.010224378667771816, + 0.03470981866121292, + 0.006711197551339865, + 0.01579151675105095, + 0.010803323239088058, + 0.006263507530093193, + 0.004135962575674057, + 0.004453721456229687, + 0.004908906761556864, + 0.013043010607361794, + 0.004218022804707289, + 0.00430893711745739, + 0.01659862883388996, + 0.013678126968443394, + 0.007439617067575455, + 0.010823209770023823, + 0.012937751598656178, + 0.028222842141985893, + 0.010732592083513737, + 0.020614737644791603, + 0.005823079496622086, + 0.005461548920720816, + 0.005262898281216621, + 0.007377436384558678, + 0.005779485683888197, + 0.0048964256420731544, + 0.005846897605806589, + 0.005430498160421848, + 0.010350595228374004, + 0.005847084801644087, + 0.005919900722801685, + 0.007342495024204254, + 0.00902361050248146, + 0.012714202515780926, + 0.01002266351133585, + 0.011062147095799446, + 0.004367201589047909, + 0.0036204741336405277, + 0.0034437994472682476, + 0.00306510366499424, + 0.004163173492997885, + 0.004449575208127499, + 0.003355285618454218, + 0.004837943706661463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.03814411908388138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059562805108726025 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.0059562805108726025, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059562805108726025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12292686849832535, + 0.03252329304814339, + 0.0690363347530365, + 0.12870672345161438, + 0.13133785128593445, + 0.051235247403383255, + 0.25313520431518555, + 0.0680650994181633, + 0.0999765545129776, + 0.043056804686784744 + ], + "credit_map": [ + 0.018888233229517937, + 0.015241207554936409, + 0.007715614512562752, + 0.007322569843381643, + 0.011553159914910793, + 0.017042988911271095, + 0.006805541925132275, + 0.038357559591531754, + 0.004067626316100359, + 0.008773650042712688, + 0.0031847949139773846, + 0.0035965824499726295, + 0.0028158396016806364, + 0.0025206943973898888, + 0.002613907679915428, + 0.004950196482241154, + 0.011806672438979149, + 0.01538925152271986, + 0.004857804160565138, + 0.004479983355849981, + 0.00633226428180933, + 0.013998929411172867, + 0.003039437346160412, + 0.00913198757916689, + 0.028399039059877396, + 0.016806140542030334, + 0.007620939519256353, + 0.00569666363298893, + 0.00944039598107338, + 0.04056373983621597, + 0.00664675235748291, + 0.013533055782318115, + 0.031771302223205566, + 0.009720551781356335, + 0.0071899741888046265, + 0.009310293011367321, + 0.008808842860162258, + 0.04083116352558136, + 0.0066956570371985435, + 0.017010066658258438, + 0.00726081570610404, + 0.008815908804535866, + 0.005996036343276501, + 0.005872519686818123, + 0.006598702631890774, + 0.005799213889986277, + 0.006096602883189917, + 0.004795446060597897, + 0.03810855373740196, + 0.021948961541056633, + 0.010829953476786613, + 0.014133254997432232, + 0.026239434257149696, + 0.06813965737819672, + 0.020821664482355118, + 0.05291372537612915, + 0.008358996361494064, + 0.010498128831386566, + 0.006831275764852762, + 0.011117825284600258, + 0.0082395663484931, + 0.008174153044819832, + 0.008044647984206676, + 0.006800503470003605, + 0.014248576946556568, + 0.0075813597068190575, + 0.007305069826543331, + 0.008806650526821613, + 0.012385109439492226, + 0.021383071318268776, + 0.011852103285491467, + 0.016414614394307137, + 0.005091565195471048, + 0.005226242821663618, + 0.005441803019493818, + 0.006218547932803631, + 0.004808538593351841, + 0.004318604711443186, + 0.004796324297785759, + 0.007155179511755705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.0059562805108726025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549643374979496 + ], + "opposite_logprob_tokens": [ + -2.929964303970337 + ], + "logprob_reward": -0.0549643374979496, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549643374979496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05274582281708717, + 0.017322035506367683, + 0.050741903483867645, + 0.4970344305038452, + 0.058710936456918716, + 0.06957171857357025, + 0.0952252745628357, + 0.04201420769095421, + 0.08425845205783844, + 0.032375212758779526 + ], + "credit_map": [ + 0.007398038636893034, + 0.006246731150895357, + 0.004777625668793917, + 0.0034876293502748013, + 0.004704612772911787, + 0.006088766269385815, + 0.0037256376817822456, + 0.01631678082048893, + 0.0021838562097400427, + 0.0045007504522800446, + 0.0017571626231074333, + 0.0018947664648294449, + 0.0018407402094453573, + 0.0011857596691697836, + 0.001473552780225873, + 0.0024854489602148533, + 0.00856736209243536, + 0.011683925054967403, + 0.003796261502429843, + 0.003377792425453663, + 0.0042021917179226875, + 0.010477930307388306, + 0.0027025341987609863, + 0.00593390641734004, + 0.15451766550540924, + 0.020835064351558685, + 0.016519619151949883, + 0.01282294187694788, + 0.020379994064569473, + 0.21604663133621216, + 0.015295922756195068, + 0.040616586804389954, + 0.016985908150672913, + 0.003314689500257373, + 0.0020113384816795588, + 0.002574449870735407, + 0.0028221500106155872, + 0.021833183243870735, + 0.0022114363964647055, + 0.006957779638469219, + 0.011244195513427258, + 0.00961797870695591, + 0.006905281916260719, + 0.006820604670792818, + 0.007568136788904667, + 0.012198058888316154, + 0.008442779071629047, + 0.006774678826332092, + 0.014471580274403095, + 0.007633985951542854, + 0.003969815094023943, + 0.005244972184300423, + 0.009459441527724266, + 0.02584022842347622, + 0.007728539872914553, + 0.020876716822385788, + 0.005273211747407913, + 0.005430905614048243, + 0.004447863437235355, + 0.006805433426052332, + 0.005029934924095869, + 0.004416355863213539, + 0.00549655593931675, + 0.005113945342600346, + 0.012230024673044682, + 0.009812397882342339, + 0.00874553807079792, + 0.009719552472233772, + 0.016415558755397797, + 0.005231794901192188, + 0.013523849658668041, + 0.008579735644161701, + 0.004296948201954365, + 0.00467383349314332, + 0.003662946168333292, + 0.004381480626761913, + 0.005330276675522327, + 0.0029660696163773537, + 0.00398679543286562, + 0.003076861845329404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.0549643374979496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464052427560091 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.00464052427560091, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464052427560091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05886015668511391, + 0.020503070205450058, + 0.05542432516813278, + 0.4673273265361786, + 0.04867813363671303, + 0.08140835911035538, + 0.0979289561510086, + 0.04239589720964432, + 0.09130515158176422, + 0.036168623715639114 + ], + "credit_map": [ + 0.008420465514063835, + 0.00823264755308628, + 0.004235377069562674, + 0.003582883160561323, + 0.005451565142720938, + 0.008447336032986641, + 0.003429303178563714, + 0.017060577869415283, + 0.0024618501774966717, + 0.0053691258653998375, + 0.002005218295380473, + 0.0020994411315768957, + 0.001882003853097558, + 0.0018765917047858238, + 0.002190516795963049, + 0.0026183221489191055, + 0.008835245855152607, + 0.014833882451057434, + 0.004195318557322025, + 0.0040793572552502155, + 0.005142452195286751, + 0.009479127824306488, + 0.00294370180927217, + 0.005915239453315735, + 0.14530684053897858, + 0.02062821574509144, + 0.015524921007454395, + 0.012248590588569641, + 0.01891050674021244, + 0.20223398506641388, + 0.014411186799407005, + 0.038063064217567444, + 0.01117550116032362, + 0.0034928880631923676, + 0.002837277250364423, + 0.003650429891422391, + 0.003252278547734022, + 0.01442481204867363, + 0.003106936113908887, + 0.0067380089312791824, + 0.012876397930085659, + 0.011853502131998539, + 0.008480078540742397, + 0.008170703426003456, + 0.009048338048160076, + 0.01367925200611353, + 0.009826586581766605, + 0.0074734981171786785, + 0.014830591157078743, + 0.008247612975537777, + 0.004141263663768768, + 0.0056337700225412846, + 0.009910061955451965, + 0.026024188846349716, + 0.008098045364022255, + 0.02104342170059681, + 0.005281254183501005, + 0.005400675814598799, + 0.004507163539528847, + 0.00684852059930563, + 0.004854581318795681, + 0.004322741646319628, + 0.0058753835037350655, + 0.005305573809891939, + 0.013312190771102905, + 0.010097741149365902, + 0.009066629223525524, + 0.009770645759999752, + 0.01722221076488495, + 0.008679156191647053, + 0.013500450178980827, + 0.009656128473579884, + 0.004719964228570461, + 0.005444200709462166, + 0.004087310750037432, + 0.004983993712812662, + 0.005566983949393034, + 0.003405848518013954, + 0.004186187870800495, + 0.003774133510887623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.00464052427560091 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, fixed camera holds a steady medium-wide shot across a stark, featureless white void, illuminated by a brilliant, shadowless, diffuse light that fills the infinite, clinical space. A luxury touring motorcycle, reimagined in the surreal, chrome-infused style of Sorayama, moves within the frame. Its body is a single, impossibly smooth form of liquid-polished chrome, devoid of a visible engine, reflecting the distortions of its surroundings with mirror-like perfection in its stark whiteness. Seamlessly blending into the machine's chassis is a gynoid figure, her lithe, metallic form leaning forward with hands integrated into the armrests, her polished face expressionless as her entire body remains completely still. The motorcycle suddenly collides with an invisible wall, and as the camera remains still, the impact triggers a surreal deformation. The front of the chrome machine deflects viscously, its contact points twisting and rippling outward as the solid metal begins to slowly and deliberately ooze from the invisible obstacle, stretching and twisting like liquid mercury. The giantess's fused body twists in unison, her polished form pulling and flowing with the destruction of the bike, creating a captivating atmosphere of cold, silent, bizarre physics and fluid deconstruction.", + "seed": 6557, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_12/index_7303_seed_6557_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.034425050299614665, + "vlm_logprob": -0.034425050299614665, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.034425050299614665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02653009630739689 + ], + "opposite_logprob_tokens": [ + -3.9015300273895264 + ], + "logprob_reward": -0.02653009630739689, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02653009630739689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07689769566059113, + 0.1461879163980484, + 0.044832486659288406, + 0.11386440694332123, + 0.10111088305711746, + 0.3277435004711151, + 0.040433913469314575, + 0.06893432140350342, + 0.04934640973806381, + 0.030648378655314445 + ], + "credit_map": [ + 0.011410481296479702, + 0.012431111186742783, + 0.006554760970175266, + 0.0049814931116998196, + 0.020592575892806053, + 0.008813303895294666, + 0.00498890969902277, + 0.0071250624023377895, + 0.024708572775125504, + 0.019770704209804535, + 0.02008240297436714, + 0.021174730733036995, + 0.009770907461643219, + 0.02148187719285488, + 0.012358935549855232, + 0.016839779913425446, + 0.007565180771052837, + 0.004621791187673807, + 0.006955873686820269, + 0.0035610508639365435, + 0.008607091382145882, + 0.006635185796767473, + 0.002814686158671975, + 0.004071625415235758, + 0.01703261025249958, + 0.012047046795487404, + 0.009582043625414371, + 0.01008603349328041, + 0.015008289366960526, + 0.023336652666330338, + 0.008916470222175121, + 0.017855264246463776, + 0.015103514306247234, + 0.018071595579385757, + 0.010500822216272354, + 0.01471007615327835, + 0.01846063695847988, + 0.008646219968795776, + 0.0063095372170209885, + 0.009308474138379097, + 0.12008113414049149, + 0.024553583934903145, + 0.1153503954410553, + 0.019555123522877693, + 0.014374932274222374, + 0.011391455307602882, + 0.010717909783124924, + 0.011718970723450184, + 0.009091034531593323, + 0.005853644106537104, + 0.007097275462001562, + 0.004086367320269346, + 0.003996938467025757, + 0.002462982665747404, + 0.0029266299679875374, + 0.004919042345136404, + 0.011687767691910267, + 0.009582278318703175, + 0.009863930754363537, + 0.00791908334940672, + 0.00742989731952548, + 0.007786361034959555, + 0.004980164580047131, + 0.009684844873845577, + 0.010360132902860641, + 0.007251444272696972, + 0.0068350741639733315, + 0.005547839216887951, + 0.004709371365606785, + 0.004382767714560032, + 0.0036177965812385082, + 0.006641985848546028, + 0.005617997609078884, + 0.003403703449293971, + 0.003572758985683322, + 0.0028007489163428545, + 0.0037637953646481037, + 0.0027016159147024155, + 0.003657061606645584, + 0.005130698438733816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.02653009630739689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208043918013573 + ], + "opposite_logprob_tokens": [ + -2.8120803833007812 + ], + "logprob_reward": -0.06208043918013573, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208043918013573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20157140493392944, + 0.15280234813690186, + 0.09847227483987808, + 0.2924630045890808, + 0.07126616686582565, + 0.05174979940056801, + 0.05394576117396355, + 0.03591926395893097, + 0.021337175741791725, + 0.020472746342420578 + ], + "credit_map": [ + 0.030300812795758247, + 0.026662254706025124, + 0.016865434125065804, + 0.01664002612233162, + 0.04120694100856781, + 0.023920856416225433, + 0.019328463822603226, + 0.02664663828909397, + 0.025868386030197144, + 0.015417389571666718, + 0.025848835706710815, + 0.014246354810893536, + 0.026113159954547882, + 0.022202085703611374, + 0.011941433884203434, + 0.011164720170199871, + 0.0157626923173666, + 0.016645662486553192, + 0.011308145709335804, + 0.009030986577272415, + 0.012604630552232265, + 0.016073934733867645, + 0.007546323351562023, + 0.009499906562268734, + 0.04282485693693161, + 0.032223157584667206, + 0.03157088905572891, + 0.0355849452316761, + 0.023754101246595383, + 0.05863120034337044, + 0.022078003734350204, + 0.045795854181051254, + 0.010144921019673347, + 0.009463750757277012, + 0.008862343616783619, + 0.0078020826913416386, + 0.01462211087346077, + 0.005525118205696344, + 0.005993090569972992, + 0.008852758444845676, + 0.009637308306992054, + 0.005529328249394894, + 0.008783635683357716, + 0.009000485762953758, + 0.006086386740207672, + 0.004235680680721998, + 0.004202182870358229, + 0.0042747980915009975, + 0.005465730093419552, + 0.006271762307733297, + 0.006417334079742432, + 0.009409560821950436, + 0.008525568060576916, + 0.004480601754039526, + 0.005654466804116964, + 0.007720743305981159, + 0.005510466173291206, + 0.004764068406075239, + 0.004594109952449799, + 0.0043239290826022625, + 0.003922875039279461, + 0.00498235272243619, + 0.0032608709298074245, + 0.004560594912618399, + 0.003169574309140444, + 0.00321592902764678, + 0.0030786260031163692, + 0.0025900164619088173, + 0.002639171201735735, + 0.001595775131136179, + 0.0021785420831292868, + 0.0028695447836071253, + 0.004188894294202328, + 0.002209044061601162, + 0.0028881062753498554, + 0.0013751272344961762, + 0.0019966198597103357, + 0.0016262130811810493, + 0.002388930646702647, + 0.0037998156622052193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.06208043918013573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745306774973869 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745306774973869, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745306774973869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09247753769159317, + 0.20988936722278595, + 0.0902051031589508, + 0.14111149311065674, + 0.14638829231262207, + 0.12845811247825623, + 0.07991202920675278, + 0.04998936876654625, + 0.02918170392513275, + 0.03238697722554207 + ], + "credit_map": [ + 0.013645684346556664, + 0.01275032851845026, + 0.010446437634527683, + 0.007464519701898098, + 0.012839702889323235, + 0.012506166473031044, + 0.00945049524307251, + 0.013374216854572296, + 0.03611195459961891, + 0.0205440241843462, + 0.03628851845860481, + 0.020713888108730316, + 0.028876325115561485, + 0.03099466860294342, + 0.019005708396434784, + 0.017354298382997513, + 0.014079977758228779, + 0.014626394957304, + 0.012637029401957989, + 0.008165158331394196, + 0.010042405687272549, + 0.013557550497353077, + 0.0071531324647367, + 0.009943461045622826, + 0.020538240671157837, + 0.015215215273201466, + 0.016924452036619186, + 0.01596948131918907, + 0.01497683860361576, + 0.02600765787065029, + 0.010643347166478634, + 0.020836278796195984, + 0.022040991112589836, + 0.014608336612582207, + 0.017239276319742203, + 0.01665462739765644, + 0.03421127423644066, + 0.012436977587640285, + 0.010894016362726688, + 0.018302809447050095, + 0.035710643976926804, + 0.009364375844597816, + 0.03578798845410347, + 0.01946614496409893, + 0.008155339397490025, + 0.006364983972162008, + 0.0064221955835819244, + 0.007186457980424166, + 0.007751804310828447, + 0.010332629084587097, + 0.008973118849098682, + 0.01412623655050993, + 0.011884740553796291, + 0.006990250200033188, + 0.008461612276732922, + 0.011391651816666126, + 0.006400194484740496, + 0.006467399187386036, + 0.006757010705769062, + 0.006309479475021362, + 0.005397661589086056, + 0.00667272275313735, + 0.004660432692617178, + 0.0073244753293693066, + 0.006756811402738094, + 0.0034187007695436478, + 0.003931891638785601, + 0.0030393507331609726, + 0.0025122954975813627, + 0.003081751521676779, + 0.0014180431608110666, + 0.005022861063480377, + 0.004982944577932358, + 0.0035928539000451565, + 0.004292556084692478, + 0.0024895924143493176, + 0.00393686443567276, + 0.0029740193858742714, + 0.0038743652403354645, + 0.006243783514946699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.006745306774973869 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002848386764526 + ], + "opposite_logprob_tokens": [ + -2.69502854347229 + ], + "logprob_reward": -0.07002848386764526, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002848386764526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13771221041679382, + 0.1368609070777893, + 0.11658305674791336, + 0.07965251803398132, + 0.0805494412779808, + 0.21152034401893616, + 0.09264230728149414, + 0.053754180669784546, + 0.055395398288965225, + 0.035329669713974 + ], + "credit_map": [ + 0.01867886632680893, + 0.018249493092298508, + 0.013511101715266705, + 0.010091245174407959, + 0.03194139897823334, + 0.018668558448553085, + 0.0118870222941041, + 0.014684520661830902, + 0.026387721300125122, + 0.01775418594479561, + 0.014566181227564812, + 0.012102588079869747, + 0.017282819375395775, + 0.02283129282295704, + 0.01528517622500658, + 0.0106509393081069, + 0.01793677546083927, + 0.018668916076421738, + 0.017709771171212196, + 0.010712973773479462, + 0.013184195384383202, + 0.016858456656336784, + 0.008082055486738682, + 0.013429914601147175, + 0.012193276546895504, + 0.008136840537190437, + 0.009306848049163818, + 0.00758548965677619, + 0.015911618247628212, + 0.007706025615334511, + 0.006382605526596308, + 0.01242980919778347, + 0.013667992316186428, + 0.011774137616157532, + 0.009680662304162979, + 0.012816064059734344, + 0.008906431496143341, + 0.006139308214187622, + 0.005343795288354158, + 0.012221051380038261, + 0.07070932537317276, + 0.01637307181954384, + 0.0672077015042305, + 0.01320503931492567, + 0.00997970812022686, + 0.010270447470247746, + 0.009329299442470074, + 0.014445737935602665, + 0.009496198035776615, + 0.013353287242352962, + 0.010771515779197216, + 0.01801937259733677, + 0.011829749681055546, + 0.00839416217058897, + 0.008651192300021648, + 0.012126829475164413, + 0.006418496370315552, + 0.006598483305424452, + 0.0073413364589214325, + 0.00707633513957262, + 0.006214170716702938, + 0.005838337354362011, + 0.005506713874638081, + 0.008760308846831322, + 0.010719728656113148, + 0.00626026326790452, + 0.005927367135882378, + 0.006658377591520548, + 0.0059496439062058926, + 0.005669441539794207, + 0.0044042328372597694, + 0.009806342422962189, + 0.00436857296153903, + 0.0036366533022373915, + 0.004616590682417154, + 0.003208839800208807, + 0.003838307922706008, + 0.0032331349793821573, + 0.005098969209939241, + 0.007328598760068417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.07002848386764526 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740925367921591 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006740925367921591, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740925367921591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1266590803861618, + 0.11026040464639664, + 0.14879591763019562, + 0.09821023046970367, + 0.09006091952323914, + 0.17556746304035187, + 0.07475554198026657, + 0.08762097358703613, + 0.038736119866371155, + 0.049333371222019196 + ], + "credit_map": [ + 0.02066797949373722, + 0.015203729271888733, + 0.011269933544099331, + 0.01004431676119566, + 0.02260892651975155, + 0.01798674277961254, + 0.01292821392416954, + 0.015949245542287827, + 0.022781740874052048, + 0.013016819953918457, + 0.012033029459416866, + 0.009526640176773071, + 0.01058190781623125, + 0.019610324874520302, + 0.012514067813754082, + 0.010195865295827389, + 0.023623241111636162, + 0.02676515281200409, + 0.020131109282374382, + 0.013548343442380428, + 0.01707819104194641, + 0.021170172840356827, + 0.009512673132121563, + 0.01696702651679516, + 0.014110088348388672, + 0.00949818640947342, + 0.012459568679332733, + 0.011226216331124306, + 0.015912039205431938, + 0.010274243541061878, + 0.007565057370811701, + 0.017164835706353188, + 0.014738802798092365, + 0.009405719116330147, + 0.01117737591266632, + 0.012784157879650593, + 0.013234768994152546, + 0.007924782112240791, + 0.006126410327851772, + 0.014668896794319153, + 0.05804703012108803, + 0.01197428535670042, + 0.05638846755027771, + 0.017850933596491814, + 0.008650360628962517, + 0.006789566949009895, + 0.0077227577567100525, + 0.008144054561853409, + 0.00895333755761385, + 0.009175990708172321, + 0.009189797565340996, + 0.016439523547887802, + 0.00815637968480587, + 0.006385428365319967, + 0.007109708618372679, + 0.009345382452011108, + 0.01000489667057991, + 0.010509265586733818, + 0.01159170176833868, + 0.011400219053030014, + 0.011298475787043571, + 0.009354243054986, + 0.009042788296937943, + 0.01441938430070877, + 0.005965754389762878, + 0.003752250922843814, + 0.005043991841375828, + 0.004582998808473349, + 0.004772937390953302, + 0.0032873349264264107, + 0.0043759956024587154, + 0.006954859010875225, + 0.007160147186368704, + 0.005538160912692547, + 0.006373380310833454, + 0.004896220751106739, + 0.005618426017463207, + 0.004921942483633757, + 0.005272109992802143, + 0.009552986361086369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.006740925367921591 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, fixed camera holds a steady medium-wide shot across a stark, featureless white void, illuminated by a brilliant, shadowless, diffuse light that fills the infinite, clinical space. A luxury touring motorcycle, reimagined in the surreal, chrome-infused style of Sorayama, moves within the frame. Its body is a single, impossibly smooth form of liquid-polished chrome, devoid of a visible engine, reflecting the distortions of its surroundings with mirror-like perfection in its stark whiteness. Seamlessly blending into the machine's chassis is a gynoid figure, her lithe, metallic form leaning forward with hands integrated into the armrests, her polished face expressionless as her entire body remains completely still. The motorcycle suddenly collides with an invisible wall, and as the camera remains still, the impact triggers a surreal deformation. The front of the chrome machine deflects viscously, its contact points twisting and rippling outward as the solid metal begins to slowly and deliberately ooze from the invisible obstacle, stretching and twisting like liquid mercury. The giantess's fused body twists in unison, her polished form pulling and flowing with the destruction of the bike, creating a captivating atmosphere of cold, silent, bizarre physics and fluid deconstruction.", + "seed": 6558, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_12/index_7303_seed_6558_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04350792434997856, + "vlm_logprob": -0.04350792434997856, + "vlm_yes_no_margin": 3.950000023841858, + "vlm_token_credit": -0.04350792434997856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14721430838108063 + ], + "opposite_logprob_tokens": [ + -2.022214412689209 + ], + "logprob_reward": -0.14721430838108063, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14721430838108063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6050271391868591, + 0.11240854859352112, + 0.03825356066226959, + 0.05268813297152519, + 0.04314868897199631, + 0.020247621461749077, + 0.024096302688121796, + 0.03870154544711113, + 0.044123779982328415, + 0.021304616704583168 + ], + "credit_map": [ + 0.14407719671726227, + 0.0713464692234993, + 0.0325700044631958, + 0.044832903891801834, + 0.05637630820274353, + 0.15638990700244904, + 0.057499900460243225, + 0.041934456676244736, + 0.018516069278120995, + 0.014201278798282146, + 0.009004898369312286, + 0.010720748454332352, + 0.015096260234713554, + 0.027434002608060837, + 0.008639898151159286, + 0.008795400150120258, + 0.004858515225350857, + 0.004169690888375044, + 0.004073135554790497, + 0.0055238027125597, + 0.008617914281785488, + 0.0034772444050759077, + 0.0042198109440505505, + 0.0033134466502815485, + 0.009931939654052258, + 0.007376459892839193, + 0.004407529253512621, + 0.004563993774354458, + 0.011742683127522469, + 0.005072750151157379, + 0.004521973431110382, + 0.005070810206234455, + 0.006544946227222681, + 0.004200848750770092, + 0.004516691900789738, + 0.007691849023103714, + 0.0060623143799602985, + 0.004842662252485752, + 0.0037682335823774338, + 0.005521147977560759, + 0.0032974390778690577, + 0.0033726731780916452, + 0.0015784051502123475, + 0.0022009604144841433, + 0.0037494495045393705, + 0.0016585750272497535, + 0.002236388623714447, + 0.0021537335123866796, + 0.0033472799696028233, + 0.002858351217582822, + 0.0020153492223471403, + 0.0027556882705539465, + 0.0037347294855862856, + 0.002182623138651252, + 0.0038283895701169968, + 0.003373893676325679, + 0.005935000255703926, + 0.004832092206925154, + 0.0029150068294256926, + 0.004750152584165335, + 0.006948636844754219, + 0.004472523462027311, + 0.004761172458529472, + 0.004086960107088089, + 0.006800518836826086, + 0.005623211618512869, + 0.0040329634211957455, + 0.004752568434923887, + 0.009507138282060623, + 0.0046083275228738785, + 0.0048052482306957245, + 0.0039938073605299, + 0.0029733446426689625, + 0.002120386343449354, + 0.0021289566066116095, + 0.0024105103220790625, + 0.0041594356298446655, + 0.0022190017625689507, + 0.0025630383752286434, + 0.002729943720623851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.14721430838108063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814595565199852 + ], + "opposite_logprob_tokens": [ + -3.2881460189819336 + ], + "logprob_reward": -0.03814595565199852, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814595565199852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4024636149406433, + 0.1391986906528473, + 0.05923633649945259, + 0.023498618975281715, + 0.01684587076306343, + 0.04202353581786156, + 0.07561183720827103, + 0.036502573639154434, + 0.02048015594482422, + 0.18413878977298737 + ], + "credit_map": [ + 0.09037722647190094, + 0.06461803615093231, + 0.02200973592698574, + 0.029461827129125595, + 0.043375883251428604, + 0.09773571789264679, + 0.03070639818906784, + 0.024178797379136086, + 0.023531358689069748, + 0.017267150804400444, + 0.014227624051272869, + 0.012644588015973568, + 0.01628389209508896, + 0.028570597991347313, + 0.013595777563750744, + 0.013077707029879093, + 0.008459378033876419, + 0.00591115141287446, + 0.004958744160830975, + 0.005497329402714968, + 0.013168791308999062, + 0.006096935831010342, + 0.008122574537992477, + 0.007021430879831314, + 0.003709397278726101, + 0.0036518475972115993, + 0.0024401629343628883, + 0.0030843308195471764, + 0.0038942412938922644, + 0.0020963044371455908, + 0.0018621035851538181, + 0.0027602314949035645, + 0.0030511291697621346, + 0.0025625345297157764, + 0.0014615835389122367, + 0.0018169148825109005, + 0.002560847671702504, + 0.0020737021695822477, + 0.0015347751323133707, + 0.0017843833193182945, + 0.006228238344192505, + 0.008230626583099365, + 0.003384143114089966, + 0.006245642434805632, + 0.006859368644654751, + 0.003705159528180957, + 0.004015695303678513, + 0.0033546602353453636, + 0.009922534227371216, + 0.007845238782465458, + 0.007155757863074541, + 0.010514949448406696, + 0.008811548352241516, + 0.008976446464657784, + 0.009598772041499615, + 0.012786590494215488, + 0.004833056591451168, + 0.004187566228210926, + 0.002491457387804985, + 0.003956465981900692, + 0.008856169879436493, + 0.004518997855484486, + 0.0037187570706009865, + 0.003940105438232422, + 0.0026938081718981266, + 0.0026602738071233034, + 0.0019364535110071301, + 0.0021343340631574392, + 0.003464014735072851, + 0.0020060217939317226, + 0.0028478847816586494, + 0.002737365197390318, + 0.02346823364496231, + 0.01806691102683544, + 0.015896009281277657, + 0.022073082625865936, + 0.03956529125571251, + 0.018787745386362076, + 0.022968409582972527, + 0.023313095793128014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.03814595565199852 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00526067242026329 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.00526067242026329, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00526067242026329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3648664057254791, + 0.07751183211803436, + 0.10540542006492615, + 0.060939542949199677, + 0.06457875669002533, + 0.018651843070983887, + 0.044461458921432495, + 0.04075903445482254, + 0.03563094884157181, + 0.18719471991062164 + ], + "credit_map": [ + 0.06909023970365524, + 0.04228156805038452, + 0.022493325173854828, + 0.03543549031019211, + 0.0369533896446228, + 0.0776078850030899, + 0.049483977258205414, + 0.03152056410908699, + 0.014651552774012089, + 0.009069284424185753, + 0.005420595873147249, + 0.006363241467624903, + 0.011776860803365707, + 0.013833663426339626, + 0.008906200528144836, + 0.007490434683859348, + 0.014847137965261936, + 0.012074809521436691, + 0.008365812711417675, + 0.013165483251214027, + 0.023466963320970535, + 0.011531624011695385, + 0.010585283860564232, + 0.01136830821633339, + 0.0074034384451806545, + 0.012304547242820263, + 0.004065658897161484, + 0.008257512003183365, + 0.014226425439119339, + 0.005205410532653332, + 0.0046334401704370975, + 0.004843117669224739, + 0.010488119907677174, + 0.008259755559265614, + 0.006102786399424076, + 0.010152418166399002, + 0.010369962081313133, + 0.006160504650324583, + 0.005162457935512066, + 0.007882756181061268, + 0.0030061774887144566, + 0.0029715353157371283, + 0.001566505990922451, + 0.0017497382359579206, + 0.004058086778968573, + 0.0020693866536021233, + 0.0016551188891753554, + 0.0015752946492284536, + 0.007260446902364492, + 0.004746629390865564, + 0.004633219912648201, + 0.0045517669059336185, + 0.0071215881034731865, + 0.0051056151278316975, + 0.0048720878548920155, + 0.0061701093800365925, + 0.005803997162729502, + 0.005942446645349264, + 0.002533138496801257, + 0.00537705747410655, + 0.007921330630779266, + 0.004355996381491423, + 0.004279388580471277, + 0.004545681644231081, + 0.0047701504081487656, + 0.004421375226229429, + 0.0037215230986475945, + 0.005813581869006157, + 0.0037350377533584833, + 0.002730933018028736, + 0.006348119582980871, + 0.004090229514986277, + 0.02243855409324169, + 0.018116116523742676, + 0.014694783836603165, + 0.02027265541255474, + 0.04566090181469917, + 0.019250890240073204, + 0.02407378889620304, + 0.02268703281879425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.00526067242026329 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329925447702408 + ], + "opposite_logprob_tokens": [ + -3.773299217224121 + ], + "logprob_reward": -0.02329925447702408, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329925447702408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3044050931930542, + 0.0472668819129467, + 0.08660680800676346, + 0.053222838789224625, + 0.01738051511347294, + 0.02014348842203617, + 0.016510548070073128, + 0.05940406024456024, + 0.04970370605587959, + 0.3453560769557953 + ], + "credit_map": [ + 0.06681182235479355, + 0.030591780319809914, + 0.0164165198802948, + 0.02961108461022377, + 0.02786649391055107, + 0.07259709388017654, + 0.03939316049218178, + 0.02111711911857128, + 0.007485882379114628, + 0.0054878792725503445, + 0.004238109569996595, + 0.005460853688418865, + 0.006994870491325855, + 0.007133605424314737, + 0.005697939544916153, + 0.004767738748341799, + 0.01226690411567688, + 0.008521190844476223, + 0.008283287286758423, + 0.011648190207779408, + 0.01573198102414608, + 0.00954357534646988, + 0.009263896383345127, + 0.011347786523401737, + 0.008304595947265625, + 0.005873169284313917, + 0.004505742806941271, + 0.009284581989049911, + 0.008835110813379288, + 0.005123954266309738, + 0.005579074379056692, + 0.005716610234230757, + 0.00251599308103323, + 0.0028852412942796946, + 0.0016080944333225489, + 0.002232252387329936, + 0.0034958876203745604, + 0.001750473165884614, + 0.0013089515268802643, + 0.0015836211387068033, + 0.0032923314720392227, + 0.003016292117536068, + 0.0018926530610769987, + 0.00292925164103508, + 0.0031783548183739185, + 0.0018316642381250858, + 0.0019142398377880454, + 0.0020887022837996483, + 0.0026652850210666656, + 0.0020064048003405333, + 0.001832114183343947, + 0.0017006496200338006, + 0.0017741973279044032, + 0.0014955067308619618, + 0.0026075346395373344, + 0.002428855514153838, + 0.00851541105657816, + 0.008093290030956268, + 0.004756447859108448, + 0.010100571438670158, + 0.007631758693605661, + 0.004895241465419531, + 0.007426992524415255, + 0.007984348572790623, + 0.006371838506311178, + 0.005516922567039728, + 0.005175302270799875, + 0.007947279140353203, + 0.004847193136811256, + 0.0038697656709700823, + 0.008195757865905762, + 0.0077796452678740025, + 0.04270767793059349, + 0.03274829313158989, + 0.030572501942515373, + 0.04330930858850479, + 0.0763755515217781, + 0.0333070270717144, + 0.04456130042672157, + 0.041774436831474304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.02329925447702408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a stark, featureless white void with a static, fixed camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a luxury touring motorcycle with a liquid-polished chrome body and no visible engine present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gynoid figure seamlessly blended into the motorcycle's chassis, leaning forward with hands integrated into the armrests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the motorcycle collide with an invisible wall, causing the front to deflect viscously and the metal to ooze like liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gynoid figure twist and flow in unison with the motorcycle's deformation during the collision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036194308195263147 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036194308195263147, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036194308195263147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4088013470172882, + 0.0437709279358387, + 0.0967482402920723, + 0.02959100529551506, + 0.034941431134939194, + 0.021001063287258148, + 0.021841945126652718, + 0.04068773612380028, + 0.06118194758892059, + 0.24143435060977936 + ], + "credit_map": [ + 0.08615615218877792, + 0.04493899643421173, + 0.023857522755861282, + 0.038705941289663315, + 0.037982746958732605, + 0.09802719950675964, + 0.04912829026579857, + 0.030004536733031273, + 0.007195841055363417, + 0.005745908711105585, + 0.003483384847640991, + 0.004617395810782909, + 0.007529193069785833, + 0.006025908514857292, + 0.004702344071120024, + 0.00447095325216651, + 0.014917564578354359, + 0.012655004858970642, + 0.0074550616554915905, + 0.010905707255005836, + 0.024109315127134323, + 0.01010353583842516, + 0.008492734283208847, + 0.008109312504529953, + 0.003629222745075822, + 0.0048742457292973995, + 0.0023182027507573366, + 0.005469117779284716, + 0.005182469729334116, + 0.0030063525773584843, + 0.0025278287939727306, + 0.0025835668202489614, + 0.0051553817465901375, + 0.004232447128742933, + 0.0033113090321421623, + 0.006118546240031719, + 0.006013859063386917, + 0.003013598034158349, + 0.0029250886291265488, + 0.004171199165284634, + 0.0032302315812557936, + 0.003497150493785739, + 0.0013673817738890648, + 0.0022315438836812973, + 0.005424942821264267, + 0.0019527508411556482, + 0.0013711553765460849, + 0.0019259076798334718, + 0.003002522513270378, + 0.0022422957699745893, + 0.0017682303441688418, + 0.0022900348994880915, + 0.0037273250054568052, + 0.0024096115957945585, + 0.003092560451477766, + 0.00330936536192894, + 0.005890036933124065, + 0.00432741641998291, + 0.0033373928163200617, + 0.006036561448127031, + 0.005990347359329462, + 0.004111067391932011, + 0.004729435313493013, + 0.006265479605644941, + 0.009572189301252365, + 0.008742362260818481, + 0.005452807527035475, + 0.008456733077764511, + 0.010904883034527302, + 0.005227476824074984, + 0.005791854113340378, + 0.007033639121800661, + 0.028533609583973885, + 0.023290518671274185, + 0.01837630569934845, + 0.02514561451971531, + 0.06256718933582306, + 0.024791603907942772, + 0.030868304893374443, + 0.027861202135682106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0036194308195263147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_tensor([6229, 6229])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_tensor([6229, 6229])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..89083f3f177f2cce878004bc6dadd443989cfe57 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_tensor([6229, 6229])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning slow-motion macro close-up, shot slightly from above, focuses on an incredibly sharp carving knife. The blade, meticulously honed to a gleaming steel blade, is adorned with a dark wooden handle, smooth and mellow from constant use. A pair of calloused yet steady hands manipulate it, the blade cutting into a block of unpolished, light-colored basswood with steady precision. As the blade advances with unwavering determination, a thin layer of wood gracefully curls and peels away before the blade, instantly shattering under the pressure, disintegrating into countless tiny, perfectly formed spirals of wood that fly into the air. In the warm, focused beam of a work lamp, these flying chips seem to come alive, shimmering like golden dust against the blurred, dark backdrop of the workshop, swirling before slowly drifting to the ground. The image is imbued with a sense of quiet concentration and a celebration of exquisite craftsmanship, rendered in warm tones and a tranquil, poetic atmosphere.", + "seed": 18110, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_6229_seed_18110_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03808981832116842, + "vlm_logprob": -0.03808981832116842, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03808981832116842, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025081658735871315 + ], + "opposite_logprob_tokens": [ + -4.400081634521484 + ], + "logprob_reward": -0.025081658735871315, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.025081658735871315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23971018195152283, + 0.08172599971294403, + 0.06151957809925079, + 0.05517754703760147, + 0.09332886338233948, + 0.1778516322374344, + 0.03733383119106293, + 0.054823458194732666, + 0.13886070251464844, + 0.05966819077730179 + ], + "credit_map": [ + 0.028823208063840866, + 0.03362579271197319, + 0.016175182536244392, + 0.013213744387030602, + 0.05695568025112152, + 0.035001371055841446, + 0.013822142966091633, + 0.04209306836128235, + 0.009001252241432667, + 0.009263253770768642, + 0.005796154495328665, + 0.007246996741741896, + 0.016240093857049942, + 0.016121039167046547, + 0.00832963827997446, + 0.009727566502988338, + 0.0077119083143770695, + 0.009317497722804546, + 0.005020717158913612, + 0.006261371076107025, + 0.010126981884241104, + 0.009455268271267414, + 0.004740904550999403, + 0.00888492725789547, + 0.006057919003069401, + 0.00518618943169713, + 0.0039965310133993626, + 0.00618584593757987, + 0.011660386808216572, + 0.007147243712097406, + 0.008165918290615082, + 0.006777510978281498, + 0.01617334596812725, + 0.019035503268241882, + 0.007814228534698486, + 0.0063524493016302586, + 0.01090084295719862, + 0.01269114762544632, + 0.007890620268881321, + 0.012470724061131477, + 0.017258165404200554, + 0.021923940628767014, + 0.011426882818341255, + 0.014128785580396652, + 0.04924169182777405, + 0.025884205475449562, + 0.020002515986561775, + 0.017985453829169273, + 0.00413535488769412, + 0.0037554604932665825, + 0.003983210306614637, + 0.0031845371704548597, + 0.004888992290943861, + 0.007491551339626312, + 0.0037365106400102377, + 0.006158212665468454, + 0.006335972808301449, + 0.00951655674725771, + 0.004255268722772598, + 0.004584146197885275, + 0.012494673021137714, + 0.0064284889958798885, + 0.005454279016703367, + 0.005754074081778526, + 0.015914147719740868, + 0.013913976959884167, + 0.012736862525343895, + 0.017009375616908073, + 0.02791467122733593, + 0.014268044382333755, + 0.021714486181735992, + 0.015389136970043182, + 0.007168179843574762, + 0.014142326079308987, + 0.004815705586224794, + 0.005898305680602789, + 0.006599635351449251, + 0.005233881529420614, + 0.00705736642703414, + 0.008752788417041302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.025081658735871315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055105146020650864 + ], + "opposite_logprob_tokens": [ + -2.930105209350586 + ], + "logprob_reward": -0.055105146020650864, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055105146020650864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27075299620628357, + 0.16880257427692413, + 0.09757458418607712, + 0.03583613038063049, + 0.04812675341963768, + 0.07183177769184113, + 0.05957084894180298, + 0.11919339001178741, + 0.07653038203716278, + 0.05178056284785271 + ], + "credit_map": [ + 0.017246903851628304, + 0.02729363739490509, + 0.017123861238360405, + 0.020550457760691643, + 0.09635049849748611, + 0.03781866282224655, + 0.01907682791352272, + 0.03529214486479759, + 0.017066556960344315, + 0.016131795942783356, + 0.016430458053946495, + 0.019414322450757027, + 0.031967200338840485, + 0.023619530722498894, + 0.01602013409137726, + 0.028152581304311752, + 0.015180196613073349, + 0.013901140540838242, + 0.007346255239099264, + 0.008314594626426697, + 0.01780349388718605, + 0.01484947931021452, + 0.006990663707256317, + 0.013188758864998817, + 0.004477773793041706, + 0.005654398817569017, + 0.0019996019545942545, + 0.0025471386034041643, + 0.0062606981955468655, + 0.005617970135062933, + 0.0027379554230719805, + 0.006540595553815365, + 0.0064118956215679646, + 0.006601476110517979, + 0.0036455453373491764, + 0.004124937113374472, + 0.007835897617042065, + 0.0057862503454089165, + 0.006913378369063139, + 0.006807372905313969, + 0.009842288680374622, + 0.012655135244131088, + 0.005525520071387291, + 0.005874703172594309, + 0.013586622662842274, + 0.008741922676563263, + 0.008400226011872292, + 0.007205356378108263, + 0.0075314161367714405, + 0.009736132808029652, + 0.004078141413629055, + 0.005973464343696833, + 0.0076287006959319115, + 0.005295201670378447, + 0.011656943708658218, + 0.007670850958675146, + 0.01581440307199955, + 0.01711614988744259, + 0.009547300636768341, + 0.01001716684550047, + 0.026597749441862106, + 0.015423145145177841, + 0.01194079965353012, + 0.012736671604216099, + 0.00887915026396513, + 0.005313948728144169, + 0.005111022852361202, + 0.005499331280589104, + 0.024845324456691742, + 0.00942011084407568, + 0.005443945061415434, + 0.012017549946904182, + 0.006878276355564594, + 0.014181037433445454, + 0.00424741068854928, + 0.004601140506565571, + 0.005579172633588314, + 0.004862355068325996, + 0.004606477916240692, + 0.006824691779911518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.055105146020650864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561134994029999 + ], + "opposite_logprob_tokens": [ + -4.387561321258545 + ], + "logprob_reward": -0.012561134994029999, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561134994029999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22711411118507385, + 0.0665963739156723, + 0.08245304226875305, + 0.06743738055229187, + 0.04158172383904457, + 0.22216026484966278, + 0.024199603125452995, + 0.07557787001132965, + 0.12792587280273438, + 0.06495385617017746 + ], + "credit_map": [ + 0.02239816263318062, + 0.03645090013742447, + 0.01371052861213684, + 0.020612744614481926, + 0.051122356206178665, + 0.02970259264111519, + 0.014409571886062622, + 0.03870721533894539, + 0.009916582144796848, + 0.010705308057367802, + 0.005387991666793823, + 0.007090752478688955, + 0.012696580961346626, + 0.008300123736262321, + 0.005343398544937372, + 0.00715562654659152, + 0.009988219477236271, + 0.014693123288452625, + 0.006405957974493504, + 0.009642047807574272, + 0.013165063224732876, + 0.011225458234548569, + 0.0078225526958704, + 0.009510604664683342, + 0.008737348951399326, + 0.0097636878490448, + 0.004083935637027025, + 0.006039452273398638, + 0.014516416937112808, + 0.008180576376616955, + 0.008100058883428574, + 0.008015892468392849, + 0.0063680484890937805, + 0.006822523660957813, + 0.004156978335231543, + 0.005081333220005035, + 0.0068162549287080765, + 0.005218680016696453, + 0.0031732397619634867, + 0.003944660536944866, + 0.02061355859041214, + 0.028399763628840446, + 0.01341131143271923, + 0.01796582341194153, + 0.0624314583837986, + 0.03246573731303215, + 0.024037523195147514, + 0.022835057228803635, + 0.003179877530783415, + 0.0044551873579621315, + 0.0021724896505475044, + 0.0024158405140042305, + 0.003257795237004757, + 0.0025108272675424814, + 0.0036010604817420244, + 0.0026065208949148655, + 0.01155028771609068, + 0.01583750545978546, + 0.007588175591081381, + 0.005991230718791485, + 0.010230816900730133, + 0.0076284161768853664, + 0.005400349851697683, + 0.011351073160767555, + 0.01650950126349926, + 0.013967562466859818, + 0.01197400875389576, + 0.019927524030208588, + 0.013242148794233799, + 0.0158501286059618, + 0.02325211837887764, + 0.013202863745391369, + 0.008945755660533905, + 0.018387386575341225, + 0.005247926339507103, + 0.005523336585611105, + 0.00845453143119812, + 0.006830838043242693, + 0.004410273395478725, + 0.007153801620006561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.012561134994029999 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905058354139328 + ], + "opposite_logprob_tokens": [ + -2.464050531387329 + ], + "logprob_reward": -0.08905058354139328, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905058354139328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17460082471370697, + 0.04411052539944649, + 0.09425099939107895, + 0.04739520326256752, + 0.03328653424978256, + 0.2783682346343994, + 0.0333230085670948, + 0.03387696295976639, + 0.19198566675186157, + 0.06880206614732742 + ], + "credit_map": [ + 0.01723533123731613, + 0.026418454945087433, + 0.010979918763041496, + 0.012032341212034225, + 0.04599258676171303, + 0.022607265040278435, + 0.014620508998632431, + 0.024714400991797447, + 0.003993320744484663, + 0.00440263980999589, + 0.005477639846503735, + 0.006249453406780958, + 0.008161714300513268, + 0.005724290385842323, + 0.004338293336331844, + 0.005763172637671232, + 0.011445192620158195, + 0.0169401615858078, + 0.008710290305316448, + 0.01604217104613781, + 0.012768064625561237, + 0.011772028170526028, + 0.00808654259890318, + 0.008486553095281124, + 0.005595831666141748, + 0.004739928990602493, + 0.003169233677908778, + 0.005409365054219961, + 0.008624848909676075, + 0.006255613639950752, + 0.0071215154603123665, + 0.00647886423394084, + 0.004057398531585932, + 0.004695236217230558, + 0.003295130794867873, + 0.0046948459930717945, + 0.005190148483961821, + 0.004007154144346714, + 0.0035754479467868805, + 0.0037711714394390583, + 0.024919744580984116, + 0.04055419936776161, + 0.017532873898744583, + 0.020710662007331848, + 0.07943962514400482, + 0.04081599414348602, + 0.027329346165060997, + 0.027065802365541458, + 0.005142013542354107, + 0.003966035787016153, + 0.0017835585167631507, + 0.002009117743000388, + 0.003720081178471446, + 0.011435791850090027, + 0.002545275492593646, + 0.00272113224491477, + 0.004544077441096306, + 0.005543985404074192, + 0.0022054489236325026, + 0.002931087277829647, + 0.006210207939147949, + 0.0042620874010026455, + 0.0037629317957907915, + 0.004417137708514929, + 0.024623630568385124, + 0.018264302983880043, + 0.013718987815082073, + 0.018805010244250298, + 0.04496557638049126, + 0.02313940040767193, + 0.02899145893752575, + 0.01947730965912342, + 0.010755741037428379, + 0.019205590710043907, + 0.0048433756455779076, + 0.006604456342756748, + 0.008543726056814194, + 0.006546640768647194, + 0.004729289095848799, + 0.007573240902274847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.08905058354139328 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650568313896656 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.008650568313896656, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650568313896656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24879443645477295, + 0.060531117022037506, + 0.09995746612548828, + 0.09443114697933197, + 0.09980890899896622, + 0.1399773359298706, + 0.03204381465911865, + 0.0575108528137207, + 0.08393282443284988, + 0.08301206678152084 + ], + "credit_map": [ + 0.026809215545654297, + 0.05035020038485527, + 0.019598202779889107, + 0.027994045987725258, + 0.047475866973400116, + 0.03142894059419632, + 0.017008140683174133, + 0.028129829093813896, + 0.005704028066247702, + 0.0064215287566185, + 0.006474756170064211, + 0.008837197907269001, + 0.01237380038946867, + 0.007334779482334852, + 0.006731176283210516, + 0.006653852295130491, + 0.01303800754249096, + 0.01560944877564907, + 0.01091783121228218, + 0.012874450534582138, + 0.017607973888516426, + 0.012438048608601093, + 0.006967403460294008, + 0.01050430815666914, + 0.015373163856565952, + 0.012684185989201069, + 0.00516824284568429, + 0.0046589006669819355, + 0.019635183736681938, + 0.01654016226530075, + 0.005198005121201277, + 0.015173300169408321, + 0.015756288543343544, + 0.019689688459038734, + 0.008254180662333965, + 0.0068904864601790905, + 0.013540999963879585, + 0.014001356437802315, + 0.007213084027171135, + 0.014462820254266262, + 0.01864641346037388, + 0.02222261019051075, + 0.008633104152977467, + 0.011949160136282444, + 0.02900352142751217, + 0.016276730224490166, + 0.017622409388422966, + 0.015623396262526512, + 0.004822987597435713, + 0.0036902939900755882, + 0.0030861636623740196, + 0.003569947089999914, + 0.0035547001753002405, + 0.004107463173568249, + 0.004867421463131905, + 0.004344837740063667, + 0.007460948079824448, + 0.010331819765269756, + 0.004044257570058107, + 0.0040946537628769875, + 0.01181743759661913, + 0.006905188784003258, + 0.004279076587408781, + 0.008577471598982811, + 0.01095820777118206, + 0.007781344931572676, + 0.00764401676133275, + 0.00665202597156167, + 0.028028413653373718, + 0.006197263021022081, + 0.005858311895281076, + 0.010813250206410885, + 0.01181875355541706, + 0.02451694943010807, + 0.007990951649844646, + 0.0066278912127017975, + 0.011762031354010105, + 0.007186118047684431, + 0.00548053253442049, + 0.007628846913576126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.008650568313896656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion macro close-up, shot slightly from above, focuses on an incredibly sharp carving knife. The blade, meticulously honed to a gleaming steel blade, is adorned with a dark wooden handle, smooth and mellow from constant use. A pair of calloused yet steady hands manipulate it, the blade cutting into a block of unpolished, light-colored basswood with steady precision. As the blade advances with unwavering determination, a thin layer of wood gracefully curls and peels away before the blade, instantly shattering under the pressure, disintegrating into countless tiny, perfectly formed spirals of wood that fly into the air. In the warm, focused beam of a work lamp, these flying chips seem to come alive, shimmering like golden dust against the blurred, dark backdrop of the workshop, swirling before slowly drifting to the ground. The image is imbued with a sense of quiet concentration and a celebration of exquisite craftsmanship, rendered in warm tones and a tranquil, poetic atmosphere.", + "seed": 18111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_6229_seed_18111_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025916131725534797, + "vlm_logprob": -0.025916131725534797, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.025916131725534797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018500275909900665 + ], + "opposite_logprob_tokens": [ + -5.143500328063965 + ], + "logprob_reward": -0.018500275909900665, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.018500275909900665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06882323324680328, + 0.535598635673523, + 0.0826513022184372, + 0.05659204348921776, + 0.09882465749979019, + 0.02675679326057434, + 0.025603685528039932, + 0.036056067794561386, + 0.02897237055003643, + 0.040121208876371384 + ], + "credit_map": [ + 0.008005187846720219, + 0.010165748186409473, + 0.008061525411903858, + 0.00879229512065649, + 0.008608306758105755, + 0.008583943359553814, + 0.010073314420878887, + 0.0065329112112522125, + 0.039354484528303146, + 0.17141692340373993, + 0.13912276923656464, + 0.03207249566912651, + 0.0430571474134922, + 0.025358248502016068, + 0.026971593499183655, + 0.058244943618774414, + 0.005881275981664658, + 0.007138526998460293, + 0.014982831664383411, + 0.010317276231944561, + 0.0046906969510018826, + 0.005193451419472694, + 0.012745105661451817, + 0.02170214056968689, + 0.005180342588573694, + 0.008859951049089432, + 0.005681880284100771, + 0.0071649798192083836, + 0.006240444257855415, + 0.007131408434361219, + 0.009980693459510803, + 0.006352344993501902, + 0.013292305171489716, + 0.017871372401714325, + 0.009570462629199028, + 0.014024256728589535, + 0.007625989615917206, + 0.0058690947480499744, + 0.01968401111662388, + 0.01088716834783554, + 0.002322699408978224, + 0.004503406118601561, + 0.0031600904185324907, + 0.0026289590168744326, + 0.003490130417048931, + 0.003084571799263358, + 0.004508783109486103, + 0.0030581525061279535, + 0.0028194000478833914, + 0.003945241216570139, + 0.0029698002617806196, + 0.0024061391595751047, + 0.002548609161749482, + 0.0026797924656420946, + 0.005034975241869688, + 0.0031997289042919874, + 0.004158608615398407, + 0.004670049995183945, + 0.004641406238079071, + 0.0064337062649428844, + 0.004278453066945076, + 0.002312673255801201, + 0.0059716603718698025, + 0.0035895102191716433, + 0.005040512420237064, + 0.003016952658072114, + 0.0049730101600289345, + 0.004117452073842287, + 0.003811337985098362, + 0.0020525266882032156, + 0.003484059125185013, + 0.0024765178095549345, + 0.0043382104486227036, + 0.00504639744758606, + 0.004893567878752947, + 0.006393995601683855, + 0.003930867183953524, + 0.004144282545894384, + 0.006505765952169895, + 0.0048681204207241535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.018500275909900665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038191623985767365 + ], + "opposite_logprob_tokens": [ + -3.288191556930542 + ], + "logprob_reward": -0.038191623985767365, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038191623985767365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15759488940238953, + 0.14675769209861755, + 0.11530378460884094, + 0.1431647092103958, + 0.14089661836624146, + 0.045146241784095764, + 0.046483807265758514, + 0.055214762687683105, + 0.10666152089834213, + 0.042775943875312805 + ], + "credit_map": [ + 0.017476752400398254, + 0.018293406814336777, + 0.014696931466460228, + 0.01854267343878746, + 0.021127821877598763, + 0.026608001440763474, + 0.02244611084461212, + 0.0184031892567873, + 0.014784582890570164, + 0.04026845842599869, + 0.021629128605127335, + 0.014098647981882095, + 0.013401367701590061, + 0.009209895506501198, + 0.009412381798028946, + 0.023953227326273918, + 0.007975987158715725, + 0.006258259993046522, + 0.029659951105713844, + 0.00786569993942976, + 0.006074465811252594, + 0.01010440569370985, + 0.017286011949181557, + 0.030079005286097527, + 0.012604122050106525, + 0.016310179606080055, + 0.012118104845285416, + 0.017838016152381897, + 0.017122667282819748, + 0.02263353206217289, + 0.02804272063076496, + 0.01649537682533264, + 0.01337712537497282, + 0.017377635464072227, + 0.014583362266421318, + 0.01565438322722912, + 0.011875636875629425, + 0.013287527486681938, + 0.03576295077800751, + 0.01897800713777542, + 0.0051470245234668255, + 0.005974356085062027, + 0.005505694542080164, + 0.00486221956089139, + 0.005670396611094475, + 0.003644389333203435, + 0.0092542115598917, + 0.005087945610284805, + 0.004068813286721706, + 0.005811108276247978, + 0.005185625981539488, + 0.004472021479159594, + 0.005314034875482321, + 0.007672515697777271, + 0.008295384235680103, + 0.0056643057614564896, + 0.00545078981667757, + 0.007551141548901796, + 0.007426052819937468, + 0.00991917122155428, + 0.0055207135155797005, + 0.003348412923514843, + 0.009484809823334217, + 0.006513670086860657, + 0.014955416321754456, + 0.011585965752601624, + 0.014503609389066696, + 0.014977415092289448, + 0.013555286452174187, + 0.01079403143376112, + 0.014941922388970852, + 0.011347880586981773, + 0.004516015760600567, + 0.007826421409845352, + 0.0044017573818564415, + 0.004302304703742266, + 0.0044184485450387, + 0.005586035083979368, + 0.005622166208922863, + 0.006102793384343386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.038191623985767365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076424358412623405 + ], + "opposite_logprob_tokens": [ + -4.8826422691345215 + ], + "logprob_reward": -0.0076424358412623405, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076424358412623405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08371486514806747, + 0.31707826256752014, + 0.17875166237354279, + 0.074983611702919, + 0.10036830604076385, + 0.03857535496354103, + 0.03368859738111496, + 0.05874861031770706, + 0.04948962852358818, + 0.0646011009812355 + ], + "credit_map": [ + 0.012878168374300003, + 0.010392109863460064, + 0.00834549218416214, + 0.009393235668540001, + 0.0092856977134943, + 0.01263374462723732, + 0.011336490511894226, + 0.009449923411011696, + 0.021879609674215317, + 0.10157382488250732, + 0.08346658945083618, + 0.019630158320069313, + 0.02533222921192646, + 0.014082367531955242, + 0.01636294461786747, + 0.0347505584359169, + 0.013282951898872852, + 0.01160837709903717, + 0.04375004768371582, + 0.022728487849235535, + 0.008183185011148453, + 0.010566103272140026, + 0.022963978350162506, + 0.04566853120923042, + 0.007236845791339874, + 0.007169341668486595, + 0.008894279599189758, + 0.010233587585389614, + 0.010896584019064903, + 0.01207510195672512, + 0.010762425139546394, + 0.00771544873714447, + 0.009134543128311634, + 0.00826825387775898, + 0.012200222350656986, + 0.012902804650366306, + 0.006632391829043627, + 0.008575844578444958, + 0.027915069833397865, + 0.014739182777702808, + 0.002884829184040427, + 0.005086736753582954, + 0.005231172777712345, + 0.005143155809491873, + 0.004653003066778183, + 0.0030768411234021187, + 0.007948082871735096, + 0.004551535937935114, + 0.0037539734039455652, + 0.004476992879062891, + 0.0036673746071755886, + 0.0034854619298130274, + 0.00521253002807498, + 0.003428703173995018, + 0.005280068144202232, + 0.004383491817861795, + 0.005754297133535147, + 0.006154761649668217, + 0.006464050151407719, + 0.009714343585073948, + 0.00560399703681469, + 0.005538936238735914, + 0.012337865307927132, + 0.007180357351899147, + 0.00533766346052289, + 0.0059094480238854885, + 0.00819140113890171, + 0.008981705643236637, + 0.006226196885108948, + 0.004822199698537588, + 0.0047902981750667095, + 0.0052307117730379105, + 0.005827438551932573, + 0.011749890632927418, + 0.007611453998833895, + 0.008138188160955906, + 0.005741027183830738, + 0.006538713816553354, + 0.009197235107421875, + 0.009797151200473309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.0076424358412623405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062050189822912216 + ], + "opposite_logprob_tokens": [ + -2.8120501041412354 + ], + "logprob_reward": -0.062050189822912216, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062050189822912216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06786664575338364, + 0.7067503333091736, + 0.02475871704518795, + 0.04672079160809517, + 0.03470320627093315, + 0.03730793297290802, + 0.018016289919614792, + 0.015209600329399109, + 0.018943890929222107, + 0.02972269058227539 + ], + "credit_map": [ + 0.008425869978964329, + 0.009252136573195457, + 0.006414349656552076, + 0.007743109483271837, + 0.008610889315605164, + 0.009600892663002014, + 0.010955086909234524, + 0.006864306982606649, + 0.055610816925764084, + 0.22215469181537628, + 0.1555710732936859, + 0.048926301300525665, + 0.06006705388426781, + 0.03724659979343414, + 0.03895176574587822, + 0.0882219597697258, + 0.002614356344565749, + 0.004054306074976921, + 0.0034892065450549126, + 0.002792294602841139, + 0.0016723484732210636, + 0.0031212184112519026, + 0.003965749870985746, + 0.003049233928322792, + 0.0045054699294269085, + 0.005490421317517757, + 0.0048169782385230064, + 0.005974882282316685, + 0.005495261866599321, + 0.007741295732557774, + 0.007615004666149616, + 0.0050814710557460785, + 0.0031732763163745403, + 0.005303641781210899, + 0.003985099960118532, + 0.0032919268123805523, + 0.0036478163674473763, + 0.002253926359117031, + 0.008322249166667461, + 0.004725268110632896, + 0.0034824565518647432, + 0.010172800160944462, + 0.003658056491985917, + 0.003117637475952506, + 0.00580952363088727, + 0.003807506989687681, + 0.00321524147875607, + 0.004044703207910061, + 0.0014136708341538906, + 0.0020330408588051796, + 0.0022255892399698496, + 0.0023936566431075335, + 0.002606434980407357, + 0.00159696105401963, + 0.0035965752322226763, + 0.002150357933714986, + 0.0020775513257831335, + 0.001969309290871024, + 0.002139424206689, + 0.0018400653498247266, + 0.0021281379740685225, + 0.0011403653770685196, + 0.0023582614958286285, + 0.001556483213789761, + 0.003259532153606415, + 0.0022789831273257732, + 0.002807068172842264, + 0.002762576099485159, + 0.0022136124316602945, + 0.0014725282089784741, + 0.0023067600559443235, + 0.0018428274197503924, + 0.0027185571379959583, + 0.005342214833945036, + 0.0034917769953608513, + 0.00309909600764513, + 0.002623498672619462, + 0.004095977637916803, + 0.00415838323533535, + 0.0041931793093681335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.062050189822912216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196133067831397 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.003196133067831397, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196133067831397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07106181979179382, + 0.3105016052722931, + 0.14097391068935394, + 0.04095945507287979, + 0.17349721491336823, + 0.0425834022462368, + 0.10351128876209259, + 0.07114098221063614, + 0.0175420343875885, + 0.028228260576725006 + ], + "credit_map": [ + 0.008070823736488819, + 0.010026318952441216, + 0.008845207281410694, + 0.009041340090334415, + 0.007283157669007778, + 0.01160652469843626, + 0.00895027257502079, + 0.007238172460347414, + 0.01990961842238903, + 0.09601298719644547, + 0.09770829975605011, + 0.017755625769495964, + 0.025287840515375137, + 0.013839859515428543, + 0.015904951840639114, + 0.024082420393824577, + 0.009454249404370785, + 0.005715598352253437, + 0.04475416988134384, + 0.009274464100599289, + 0.004251981154084206, + 0.00715015409514308, + 0.015379433520138264, + 0.044993869960308075, + 0.004203882534056902, + 0.005308871157467365, + 0.004460932686924934, + 0.005364216864109039, + 0.004014147445559502, + 0.00673097325488925, + 0.006395944394171238, + 0.004480486735701561, + 0.01678592711687088, + 0.01751124858856201, + 0.01905660890042782, + 0.021687833592295647, + 0.012500900775194168, + 0.021780962124466896, + 0.04144958406686783, + 0.02272416092455387, + 0.0043393210507929325, + 0.004926958587020636, + 0.006455150432884693, + 0.005953031592071056, + 0.0037992324214428663, + 0.004766878671944141, + 0.007586056832224131, + 0.004756777081638575, + 0.008102374151349068, + 0.013713376596570015, + 0.011149290949106216, + 0.011609543114900589, + 0.010032291524112225, + 0.014255234971642494, + 0.020605431869626045, + 0.014043755829334259, + 0.007804405875504017, + 0.007402376271784306, + 0.009987146593630314, + 0.0116856275126338, + 0.0073731811717152596, + 0.007273703347891569, + 0.011766387149691582, + 0.007848160341382027, + 0.002161924261599779, + 0.0024024415761232376, + 0.002726217033341527, + 0.00306400447152555, + 0.0023752707056701183, + 0.0015553312841802835, + 0.0016488807741552591, + 0.0016079653287306428, + 0.003462339285761118, + 0.0038622915744781494, + 0.002764336299151182, + 0.003087353892624378, + 0.0037171037402004004, + 0.002545173978433013, + 0.005345492623746395, + 0.0034441733732819557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.003196133067831397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion macro close-up, shot slightly from above, focuses on an incredibly sharp carving knife. The blade, meticulously honed to a gleaming steel blade, is adorned with a dark wooden handle, smooth and mellow from constant use. A pair of calloused yet steady hands manipulate it, the blade cutting into a block of unpolished, light-colored basswood with steady precision. As the blade advances with unwavering determination, a thin layer of wood gracefully curls and peels away before the blade, instantly shattering under the pressure, disintegrating into countless tiny, perfectly formed spirals of wood that fly into the air. In the warm, focused beam of a work lamp, these flying chips seem to come alive, shimmering like golden dust against the blurred, dark backdrop of the workshop, swirling before slowly drifting to the ground. The image is imbued with a sense of quiet concentration and a celebration of exquisite craftsmanship, rendered in warm tones and a tranquil, poetic atmosphere.", + "seed": 18112, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_6229_seed_18112_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01881639934144914, + "vlm_logprob": -0.01881639934144914, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.01881639934144914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01703675463795662 + ], + "opposite_logprob_tokens": [ + -5.1420369148254395 + ], + "logprob_reward": -0.01703675463795662, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01703675463795662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15457752346992493, + 0.0492684543132782, + 0.06711893528699875, + 0.05302439630031586, + 0.040070585906505585, + 0.05671928450465202, + 0.06765077263116837, + 0.07037382572889328, + 0.14722280204296112, + 0.293973445892334 + ], + "credit_map": [ + 0.028264271095395088, + 0.013853549025952816, + 0.027697836980223656, + 0.014330469071865082, + 0.011124083772301674, + 0.012576061300933361, + 0.021609678864479065, + 0.025121578946709633, + 0.006483448203653097, + 0.0057614450342953205, + 0.008223612792789936, + 0.004220378585159779, + 0.00458412105217576, + 0.005979586858302355, + 0.008000052534043789, + 0.006015811115503311, + 0.0072996546514332294, + 0.009846542961895466, + 0.012156323529779911, + 0.007046769373118877, + 0.005365750752389431, + 0.004586951807141304, + 0.012357503175735474, + 0.00845943484455347, + 0.00451994314789772, + 0.004537591710686684, + 0.007253325078636408, + 0.007256734184920788, + 0.010190569795668125, + 0.00473628006875515, + 0.00666324608027935, + 0.007866711355745792, + 0.005382806994020939, + 0.0030673700384795666, + 0.00860309973359108, + 0.0032777932938188314, + 0.004014029633253813, + 0.0032478009816259146, + 0.005981018301099539, + 0.0064966692589223385, + 0.004704422317445278, + 0.004840950947254896, + 0.006602042820304632, + 0.006128211971372366, + 0.007258770056068897, + 0.007610407657921314, + 0.011549578048288822, + 0.008024904876947403, + 0.007720413617789745, + 0.010241985321044922, + 0.01031722966581583, + 0.009396782144904137, + 0.008514934219419956, + 0.00449556577950716, + 0.009600961580872536, + 0.007362907286733389, + 0.008477016352117062, + 0.007040601689368486, + 0.011181974783539772, + 0.008128930814564228, + 0.009043664671480656, + 0.005648837424814701, + 0.012411616742610931, + 0.008441184647381306, + 0.03327236324548721, + 0.023440014570951462, + 0.01676836423575878, + 0.014296745881438255, + 0.014277354814112186, + 0.01116862427443266, + 0.009171039797365665, + 0.02482828125357628, + 0.09744884818792343, + 0.01966814696788788, + 0.011115819215774536, + 0.010071616619825363, + 0.020886341109871864, + 0.021111920475959778, + 0.016754740849137306, + 0.09691599756479263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01703675463795662 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01611212268471718 + ], + "opposite_logprob_tokens": [ + -4.141112327575684 + ], + "logprob_reward": -0.01611212268471718, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01611212268471718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12239481508731842, + 0.060936249792575836, + 0.05313704162836075, + 0.033332664519548416, + 0.03444558009505272, + 0.10127036273479462, + 0.06702861934900284, + 0.06126277893781662, + 0.17441421747207642, + 0.2917776107788086 + ], + "credit_map": [ + 0.013428648002445698, + 0.008433116599917412, + 0.01954866386950016, + 0.011881086975336075, + 0.01953287608921528, + 0.012489588931202888, + 0.019387373700737953, + 0.01769346371293068, + 0.007898768410086632, + 0.006346165202558041, + 0.009733887389302254, + 0.006048022303730249, + 0.005500965751707554, + 0.007901281118392944, + 0.00952228344976902, + 0.007984871044754982, + 0.005056412424892187, + 0.00669149961322546, + 0.01017525140196085, + 0.007599427364766598, + 0.004780065733939409, + 0.0036263882648199797, + 0.00548876216635108, + 0.009719236753880978, + 0.004571171477437019, + 0.0029507942963391542, + 0.007303979247808456, + 0.004611791577190161, + 0.003083196235820651, + 0.002222742885351181, + 0.003032927867025137, + 0.0055560641922056675, + 0.005261285696178675, + 0.002546049654483795, + 0.0060388776473701, + 0.0024811786133795977, + 0.003975237254053354, + 0.004009569529443979, + 0.005043803248554468, + 0.005089579615741968, + 0.0056760008446872234, + 0.007220563944429159, + 0.009113647975027561, + 0.01089321356266737, + 0.009706500917673111, + 0.018319960683584213, + 0.025654261931777, + 0.014686215668916702, + 0.007560631260275841, + 0.008610299788415432, + 0.0077811642549932, + 0.009245573543012142, + 0.007509964518249035, + 0.006802031770348549, + 0.00993574783205986, + 0.009583206847310066, + 0.006696649361401796, + 0.0075826747342944145, + 0.011779896914958954, + 0.0068128169514238834, + 0.006675048731267452, + 0.005423157941550016, + 0.008664988912642002, + 0.007627545390278101, + 0.021460706368088722, + 0.01897921971976757, + 0.046507809311151505, + 0.01994333416223526, + 0.01566784456372261, + 0.01614813506603241, + 0.012187836691737175, + 0.023519335314631462, + 0.10227406024932861, + 0.019805727526545525, + 0.00935025978833437, + 0.009389229118824005, + 0.014111055992543697, + 0.02025648206472397, + 0.014932109974324703, + 0.10165869444608688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.01611212268471718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008655531331896782 + ], + "opposite_logprob_tokens": [ + -4.758655548095703 + ], + "logprob_reward": -0.008655531331896782, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008655531331896782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20684179663658142, + 0.0964161679148674, + 0.067506343126297, + 0.1489744931459427, + 0.04118681326508522, + 0.06273771077394485, + 0.07913380861282349, + 0.039808884263038635, + 0.152894988656044, + 0.1044989600777626 + ], + "credit_map": [ + 0.025604313239455223, + 0.019228441640734673, + 0.036516234278678894, + 0.02762858383357525, + 0.015875352546572685, + 0.015031715855002403, + 0.035136688500642776, + 0.03182046860456467, + 0.014156002551317215, + 0.007638213690370321, + 0.013340271078050137, + 0.008532660081982613, + 0.008120033890008926, + 0.011898865923285484, + 0.01801753230392933, + 0.014712586998939514, + 0.004450737964361906, + 0.007862412370741367, + 0.013849123381078243, + 0.010684819892048836, + 0.006136763840913773, + 0.0050199758261442184, + 0.009877368807792664, + 0.009625139646232128, + 0.02080804854631424, + 0.012905743904411793, + 0.0281373243778944, + 0.02177206240594387, + 0.015339421108365059, + 0.011206060647964478, + 0.015902142971754074, + 0.02290368266403675, + 0.006048611830919981, + 0.0036653843708336353, + 0.006517886184155941, + 0.0034709819592535496, + 0.004766637925058603, + 0.004226722754538059, + 0.006769408006221056, + 0.005721178837120533, + 0.004595048725605011, + 0.0059360321611166, + 0.007434626575559378, + 0.006187220569700003, + 0.0073104919865727425, + 0.011322816833853722, + 0.011413419619202614, + 0.00853805523365736, + 0.008603174239397049, + 0.011612926609814167, + 0.014210810884833336, + 0.012548694387078285, + 0.008452781476080418, + 0.00590329896658659, + 0.009014173410832882, + 0.008787942118942738, + 0.003906823694705963, + 0.00436232378706336, + 0.0073844376020133495, + 0.004561126232147217, + 0.0056953830644488335, + 0.00323758483864367, + 0.00581076554954052, + 0.004850438330322504, + 0.02892286516726017, + 0.014938748441636562, + 0.0358714759349823, + 0.013077568262815475, + 0.013915722258388996, + 0.011915836483240128, + 0.010460441000759602, + 0.023792332038283348, + 0.03504612669348717, + 0.006382187828421593, + 0.005742636509239674, + 0.003923706244677305, + 0.005611190106719732, + 0.00752397533506155, + 0.0061577740125358105, + 0.03411136567592621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.008655531331896782 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865875095129013 + ], + "opposite_logprob_tokens": [ + -3.048658847808838 + ], + "logprob_reward": -0.04865875095129013, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865875095129013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14912086725234985, + 0.06561773270368576, + 0.03845454752445221, + 0.06978801637887955, + 0.04709085822105408, + 0.06978179514408112, + 0.15395772457122803, + 0.04782246798276901, + 0.15305586159229279, + 0.2053101658821106 + ], + "credit_map": [ + 0.014226067811250687, + 0.012994245626032352, + 0.02259560488164425, + 0.014807639643549919, + 0.018266770988702774, + 0.020713049918413162, + 0.026954298838973045, + 0.018563171848654747, + 0.010549522936344147, + 0.009491450153291225, + 0.009552420116961002, + 0.004983038175851107, + 0.006656861398369074, + 0.007510218303650618, + 0.006722280755639076, + 0.010151936672627926, + 0.0030551711097359657, + 0.004720420576632023, + 0.0067957318387925625, + 0.004436710383743048, + 0.0037184772081673145, + 0.0031848924700170755, + 0.005906103178858757, + 0.006637035869061947, + 0.0125837167724967, + 0.012299099005758762, + 0.011298292316496372, + 0.006288653705269098, + 0.008654313161969185, + 0.0042467569001019, + 0.005129432771354914, + 0.00928774755448103, + 0.007418766152113676, + 0.003887010971084237, + 0.009777238592505455, + 0.004371549002826214, + 0.004046565853059292, + 0.005393180530518293, + 0.0051069678738713264, + 0.00708957202732563, + 0.005193253979086876, + 0.006299654021859169, + 0.007678873371332884, + 0.0069800508208572865, + 0.00926172360777855, + 0.009149190038442612, + 0.015009982511401176, + 0.010209059342741966, + 0.018694676458835602, + 0.025855092331767082, + 0.026213303208351135, + 0.025021038949489594, + 0.016277074813842773, + 0.009522229433059692, + 0.015549938194453716, + 0.01682436093688011, + 0.005127110052853823, + 0.00649992423132062, + 0.009747251868247986, + 0.005160167347639799, + 0.004688818473368883, + 0.00417796615511179, + 0.00685349153354764, + 0.005567734129726887, + 0.03006983920931816, + 0.01765541173517704, + 0.02648886665701866, + 0.013227504678070545, + 0.016780881211161613, + 0.015596087090671062, + 0.01079406775534153, + 0.022443190217018127, + 0.0630788505077362, + 0.01515002641826868, + 0.008654969744384289, + 0.007352521177381277, + 0.017611859366297722, + 0.01852661743760109, + 0.013998563401401043, + 0.060936734080314636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.04865875095129013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036188371013849974 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036188371013849974, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036188371013849974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19436277449131012, + 0.13116051256656647, + 0.06097787618637085, + 0.057199444621801376, + 0.25343945622444153, + 0.05616002902388573, + 0.0453256294131279, + 0.06804651021957397, + 0.0483933724462986, + 0.08493440598249435 + ], + "credit_map": [ + 0.02423369511961937, + 0.016022518277168274, + 0.031849224120378494, + 0.025354431942105293, + 0.022844569757580757, + 0.014433028176426888, + 0.029690152034163475, + 0.02993514947593212, + 0.023502198979258537, + 0.016947904601693153, + 0.016667652875185013, + 0.008794239722192287, + 0.008912894874811172, + 0.015310205519199371, + 0.022204209119081497, + 0.018821213394403458, + 0.007053365930914879, + 0.008937460370361805, + 0.00977284461259842, + 0.008412598632276058, + 0.006253603380173445, + 0.004427809733897448, + 0.008677884005010128, + 0.007442313712090254, + 0.00498274015262723, + 0.005606715567409992, + 0.011915386654436588, + 0.0062123872339725494, + 0.007711431477218866, + 0.00555817224085331, + 0.005709528923034668, + 0.009503082372248173, + 0.03445465490221977, + 0.0199558287858963, + 0.045135848224163055, + 0.024291517212986946, + 0.02802158147096634, + 0.028904510661959648, + 0.038497403264045715, + 0.03417813405394554, + 0.0041573140770196915, + 0.0048676468431949615, + 0.006458400748670101, + 0.0058243153616786, + 0.007565817330032587, + 0.009404699318110943, + 0.009972470812499523, + 0.00790936779230833, + 0.004490474238991737, + 0.005658337846398354, + 0.008163390681147575, + 0.00894625298678875, + 0.004728246480226517, + 0.0033839987590909004, + 0.005134686827659607, + 0.004820248112082481, + 0.0071572246961295605, + 0.007373715750873089, + 0.012338481843471527, + 0.009481649845838547, + 0.007685793563723564, + 0.0044895657338202, + 0.011525822803378105, + 0.007994257844984531, + 0.008466171100735664, + 0.006831921171396971, + 0.008983651176095009, + 0.005821727216243744, + 0.0043130116537213326, + 0.004342177417129278, + 0.002946123480796814, + 0.006688595749437809, + 0.023237990215420723, + 0.0067888847552239895, + 0.005128426011651754, + 0.003992588724941015, + 0.010454954579472542, + 0.006152082700282335, + 0.0070520248264074326, + 0.022127464413642883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0036188371013849974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning slow-motion macro close-up, shot slightly from above, focuses on an incredibly sharp carving knife. The blade, meticulously honed to a gleaming steel blade, is adorned with a dark wooden handle, smooth and mellow from constant use. A pair of calloused yet steady hands manipulate it, the blade cutting into a block of unpolished, light-colored basswood with steady precision. As the blade advances with unwavering determination, a thin layer of wood gracefully curls and peels away before the blade, instantly shattering under the pressure, disintegrating into countless tiny, perfectly formed spirals of wood that fly into the air. In the warm, focused beam of a work lamp, these flying chips seem to come alive, shimmering like golden dust against the blurred, dark backdrop of the workshop, swirling before slowly drifting to the ground. The image is imbued with a sense of quiet concentration and a celebration of exquisite craftsmanship, rendered in warm tones and a tranquil, poetic atmosphere.", + "seed": 18113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_13/index_6229_seed_18113_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025624483544379472, + "vlm_logprob": -0.025624483544379472, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.025624483544379472, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03935844451189041 + ], + "opposite_logprob_tokens": [ + -3.789358377456665 + ], + "logprob_reward": -0.03935844451189041, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03935844451189041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11277294158935547, + 0.07925532013177872, + 0.08728605508804321, + 0.07518520951271057, + 0.1929357945919037, + 0.10827270895242691, + 0.09086580574512482, + 0.06456570327281952, + 0.12126205116510391, + 0.06759842485189438 + ], + "credit_map": [ + 0.015403075143694878, + 0.011302012018859386, + 0.012226969003677368, + 0.006940592546015978, + 0.01730847731232643, + 0.017571762204170227, + 0.012442263774573803, + 0.019577784463763237, + 0.011517008766531944, + 0.0069349417462944984, + 0.00974340084940195, + 0.00831944216042757, + 0.011383455246686935, + 0.009309432469308376, + 0.010894806124269962, + 0.011152829974889755, + 0.008629891090095043, + 0.007903454825282097, + 0.02143670618534088, + 0.004569351673126221, + 0.02685076929628849, + 0.00862172432243824, + 0.0035674097016453743, + 0.0057067424058914185, + 0.014182855375111103, + 0.014298892579972744, + 0.01201554760336876, + 0.005875151138752699, + 0.007080490235239267, + 0.009969374164938927, + 0.0047868224792182446, + 0.0069760726764798164, + 0.026881858706474304, + 0.03098882921040058, + 0.024157587438821793, + 0.0250615905970335, + 0.02710643969476223, + 0.018472040072083473, + 0.01839160919189453, + 0.021875854581594467, + 0.012495378963649273, + 0.021834248676896095, + 0.01504037156701088, + 0.007818811573088169, + 0.014296069741249084, + 0.01227317750453949, + 0.009870181791484356, + 0.014644467271864414, + 0.009745366871356964, + 0.018509022891521454, + 0.01506301574409008, + 0.010103647597134113, + 0.013263233006000519, + 0.01013229787349701, + 0.005174784921109676, + 0.008874434977769852, + 0.010981839150190353, + 0.008195525035262108, + 0.016013214364647865, + 0.005394430831074715, + 0.009097038768231869, + 0.0049300952814519405, + 0.0034529399126768112, + 0.006500623654574156, + 0.018499011173844337, + 0.03728700801730156, + 0.015154541470110416, + 0.0056045809760689735, + 0.01581536792218685, + 0.011343231424689293, + 0.006464401725679636, + 0.011093905195593834, + 0.017119698226451874, + 0.009528283961117268, + 0.008457613177597523, + 0.005234811455011368, + 0.005157135892659426, + 0.007889860309660435, + 0.004761847201734781, + 0.009449182078242302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.03935844451189041 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023371342569589615 + ], + "opposite_logprob_tokens": [ + -3.773371458053589 + ], + "logprob_reward": -0.023371342569589615, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023371342569589615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09844409674406052, + 0.14239132404327393, + 0.06514862179756165, + 0.04566170647740364, + 0.12223659455776215, + 0.08677755296230316, + 0.0984974056482315, + 0.1929355412721634, + 0.1046583354473114, + 0.043248724192380905 + ], + "credit_map": [ + 0.008512509055435658, + 0.011089148931205273, + 0.008215978741645813, + 0.006424515973776579, + 0.015529163181781769, + 0.013426748104393482, + 0.021610228344798088, + 0.013635803014039993, + 0.019654933363199234, + 0.017746223136782646, + 0.018168393522500992, + 0.015068169683218002, + 0.021024126559495926, + 0.017926378175616264, + 0.01866927742958069, + 0.014133820310235023, + 0.0065127708949148655, + 0.006726582068949938, + 0.01486945990473032, + 0.004282848443835974, + 0.01728801429271698, + 0.0073007321916520596, + 0.0033379471860826015, + 0.004830262623727322, + 0.007638716138899326, + 0.008549836464226246, + 0.005643581971526146, + 0.0035763715859502554, + 0.005932615138590336, + 0.0054831416346132755, + 0.0034054194111377, + 0.005432025995105505, + 0.014030683785676956, + 0.015918325632810593, + 0.013871700502932072, + 0.012945660389959812, + 0.01973557472229004, + 0.014938137494027615, + 0.01509131770581007, + 0.01570519246160984, + 0.00957966223359108, + 0.016881708055734634, + 0.01159743219614029, + 0.007020120043307543, + 0.011606214568018913, + 0.007142132613807917, + 0.011603660881519318, + 0.011346624232828617, + 0.011518470011651516, + 0.018751978874206543, + 0.01751696690917015, + 0.011776977218687534, + 0.015178199857473373, + 0.008859395049512386, + 0.005232643336057663, + 0.009662776254117489, + 0.04270205274224281, + 0.01308658067137003, + 0.08141013979911804, + 0.007743201218545437, + 0.028537767007946968, + 0.006497738882899284, + 0.004763062112033367, + 0.008194994181394577, + 0.023326614871621132, + 0.0190342515707016, + 0.03736293688416481, + 0.004254491999745369, + 0.007802478037774563, + 0.005458333529531956, + 0.0035295747220516205, + 0.0038896501064300537, + 0.008683147840201855, + 0.0049718087539076805, + 0.005559604614973068, + 0.004114109557121992, + 0.004397253040224314, + 0.005811980925500393, + 0.003660796443000436, + 0.006050021853297949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.023371342569589615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0182164516299963 + ], + "opposite_logprob_tokens": [ + -4.018216609954834 + ], + "logprob_reward": -0.0182164516299963, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0182164516299963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08344235271215439, + 0.0876515805721283, + 0.0973704382777214, + 0.08734675496816635, + 0.13101139664649963, + 0.1323988437652588, + 0.10205014050006866, + 0.05680681765079498, + 0.1457185000181198, + 0.07620315998792648 + ], + "credit_map": [ + 0.013336783275008202, + 0.010823379270732403, + 0.008133226074278355, + 0.004820406436920166, + 0.015364248305559158, + 0.007857291959226131, + 0.009498693980276585, + 0.013608334586024284, + 0.01270085759460926, + 0.00935336947441101, + 0.01042860746383667, + 0.007370481733232737, + 0.012791902758181095, + 0.011485934257507324, + 0.011005842126905918, + 0.012514579109847546, + 0.010408615693449974, + 0.00786737259477377, + 0.02445249632000923, + 0.005399956833571196, + 0.02751181274652481, + 0.010411328636109829, + 0.004655763972550631, + 0.006663101725280285, + 0.01731024496257305, + 0.016228603199124336, + 0.012070485390722752, + 0.006902116816490889, + 0.006861551199108362, + 0.015188704244792461, + 0.004466247744858265, + 0.008318805135786533, + 0.01843004673719406, + 0.02223760075867176, + 0.017132321372628212, + 0.01615009643137455, + 0.0122580137103796, + 0.012223456986248493, + 0.013120008632540703, + 0.019459860399365425, + 0.015038438141345978, + 0.01962757669389248, + 0.020445924252271652, + 0.01082651037722826, + 0.016002947464585304, + 0.01890234462916851, + 0.012978638522326946, + 0.01857648231089115, + 0.010984329506754875, + 0.014065347611904144, + 0.016200652346014977, + 0.007926234044134617, + 0.021781498566269875, + 0.011981016024947166, + 0.00824248231947422, + 0.010868582874536514, + 0.007388796657323837, + 0.007217918988317251, + 0.012729212641716003, + 0.006013533566147089, + 0.008309529162943363, + 0.006790193263441324, + 0.0038864915259182453, + 0.004471141844987869, + 0.037932269275188446, + 0.026523694396018982, + 0.04270155355334282, + 0.00626074755564332, + 0.014781301841139793, + 0.006512640044093132, + 0.0056157950311899185, + 0.005390516482293606, + 0.014496984891593456, + 0.010094028897583485, + 0.010146180167794228, + 0.0054977512918412685, + 0.00757928192615509, + 0.00951573345810175, + 0.005620278883725405, + 0.013252925127744675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.0182164516299963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307379201054573 + ], + "opposite_logprob_tokens": [ + -3.168073892593384 + ], + "logprob_reward": -0.04307379201054573, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307379201054573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.079095259308815, + 0.08232388645410538, + 0.04932260513305664, + 0.036642640829086304, + 0.05070517584681511, + 0.051774900406599045, + 0.14400792121887207, + 0.14773467183113098, + 0.33163684606552124, + 0.026756059378385544 + ], + "credit_map": [ + 0.009852256625890732, + 0.009171048179268837, + 0.007756480947136879, + 0.004692307207733393, + 0.014322740957140923, + 0.011270463466644287, + 0.0099659264087677, + 0.012064037844538689, + 0.012857667170464993, + 0.008198625408113003, + 0.01167618203908205, + 0.008391447365283966, + 0.0117959538474679, + 0.0073811509646475315, + 0.010994098149240017, + 0.011028758250176907, + 0.005224864464253187, + 0.0050443727523088455, + 0.008691452443599701, + 0.0044586495496332645, + 0.008077608421444893, + 0.005725166294723749, + 0.005113955121487379, + 0.006986535154283047, + 0.006932635325938463, + 0.006998093333095312, + 0.004550515208393335, + 0.0028736779931932688, + 0.003130856668576598, + 0.00618516281247139, + 0.0023776839952915907, + 0.0035940175876021385, + 0.007349278312176466, + 0.010909893549978733, + 0.0076680476777255535, + 0.005461139138787985, + 0.00587178859859705, + 0.004397465381771326, + 0.0035509485751390457, + 0.005496612284332514, + 0.006535559426993132, + 0.013935421593487263, + 0.007330725435167551, + 0.005002286285161972, + 0.005111096892505884, + 0.0052153426222503185, + 0.0033673467114567757, + 0.005277121905237436, + 0.017972750589251518, + 0.032771218568086624, + 0.020178282633423805, + 0.014779697172343731, + 0.02210734598338604, + 0.014427040703594685, + 0.008112749084830284, + 0.013658841140568256, + 0.02553398162126541, + 0.013458258472383022, + 0.04068561643362045, + 0.014526273123919964, + 0.018680276349186897, + 0.017668284475803375, + 0.008110365830361843, + 0.009071613661944866, + 0.08311135321855545, + 0.06946773082017899, + 0.08486808091402054, + 0.018637171015143394, + 0.033381059765815735, + 0.01421294640749693, + 0.01505627017468214, + 0.012902244925498962, + 0.006970296148210764, + 0.0034276549704372883, + 0.003657455323264003, + 0.001998747931793332, + 0.0022982005029916763, + 0.0032528347801417112, + 0.00188373529817909, + 0.00326713384129107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.04307379201054573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knife blade made of gleaming steel with a dark wooden handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are calloused hands shown manipulating the knife to cut a block of light-colored basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wood chips shatter into tiny spirals that fly into the air as the blade advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot slightly from above under the warm beam of a work lamp against a blurred dark backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102386999875307 + ], + "opposite_logprob_tokens": [ + -5.5041022300720215 + ], + "logprob_reward": -0.004102386999875307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102386999875307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08527019619941711, + 0.0663391500711441, + 0.06996329873800278, + 0.025084238499403, + 0.09874825924634933, + 0.11887525767087936, + 0.230636328458786, + 0.1210508868098259, + 0.1599137932062149, + 0.024118589237332344 + ], + "credit_map": [ + 0.013314592652022839, + 0.009111716412007809, + 0.007748709060251713, + 0.004603450186550617, + 0.013057307340204716, + 0.009383474476635456, + 0.014275672845542431, + 0.013775276020169258, + 0.01134577114135027, + 0.008119593374431133, + 0.008489422500133514, + 0.005742387846112251, + 0.007071842905133963, + 0.004577540792524815, + 0.012072283774614334, + 0.00892029982060194, + 0.007748014759272337, + 0.006088262423872948, + 0.01848694682121277, + 0.004434893373399973, + 0.020868508145213127, + 0.005506827030330896, + 0.003381012938916683, + 0.003448830684646964, + 0.004048669245094061, + 0.005460988264530897, + 0.003372795647010207, + 0.001900366274639964, + 0.0026849580463021994, + 0.0032590357586741447, + 0.0021314704790711403, + 0.002225954318419099, + 0.012978306971490383, + 0.017084553837776184, + 0.014110802672803402, + 0.010018596425652504, + 0.01755988411605358, + 0.010679081082344055, + 0.007550597656518221, + 0.008766437880694866, + 0.015877971425652504, + 0.029963916167616844, + 0.017516346648335457, + 0.019055919721722603, + 0.009600254707038403, + 0.010447612032294273, + 0.0064114839769899845, + 0.010001754388213158, + 0.02692110277712345, + 0.03871278092265129, + 0.0359349325299263, + 0.021814387291669846, + 0.05210880562663078, + 0.026405079290270805, + 0.014372539706528187, + 0.014366681687533855, + 0.022526191547513008, + 0.00989216286689043, + 0.0399695448577404, + 0.006903256755322218, + 0.019542904570698738, + 0.009115679189562798, + 0.004745212849229574, + 0.008355936035513878, + 0.0457209087908268, + 0.018467793241143227, + 0.057468634098768234, + 0.0064423708245158195, + 0.01260402426123619, + 0.007528409827500582, + 0.005040490068495274, + 0.006641163025051355, + 0.004905665758997202, + 0.003069092519581318, + 0.0028141899965703487, + 0.0016230108449235559, + 0.003180254017934203, + 0.0031484635546803474, + 0.001808004337362945, + 0.003569908905774355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.004102386999875307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_14/index_tensor([6128, 6128])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_14/index_tensor([6128, 6128])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6b6697f48144f1eafaeca244732c95e1e5e52ed2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_14/index_tensor([6128, 6128])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In captivating graphic novel style, a medium shot establishes an ancient wisteria vine, its forms rendered in bold, black, ink-like lines and flat, dramatic fields of color. The camera immediately begins a slow, deliberate push-in, creating an increasingly intense sense of focus and tension. The vine's woody, gnarled trunk, a twisted mass of dark brown and black, clings tenaciously to a stark, shadowed wall of muted, cool gray stone. From its dense branches, clusters of dense purple flowers dangle like drooping grapes, their colors a solid, unshaded field of deep purple and lavender. Suddenly, the entire plant shakes with a sharp, unnatural energy, a high-frequency tremor that spreads from the base of the trunk to the tips of the dangling flowers. As the camera continues its relentless advance, it isolates a specific flower that seems to be the center of the disturbance, vibrating with a distinctly unnatural intensity against the still, dark background. Its petals tremble as if filled with a silent, encompassing force, creating a palpable atmosphere of suspense and flourishing magic.", + "seed": 25360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_14/index_6128_seed_25360_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1872960586566478, + "vlm_logprob": -0.1872960586566478, + "vlm_yes_no_margin": 3.1999999940395356, + "vlm_token_credit": -0.1872960586566478, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31538495421409607 + ], + "opposite_logprob_tokens": [ + -1.3153849840164185 + ], + "logprob_reward": -0.31538495421409607, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31538495421409607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1202392578125, + 0.03056659735739231, + 0.07916698604822159, + 0.17426671087741852, + 0.0461745448410511, + 0.16960223019123077, + 0.18546798825263977, + 0.04424314573407173, + 0.07327906042337418, + 0.07699348032474518 + ], + "credit_map": [ + 0.011674176901578903, + 0.01710035465657711, + 0.009757140651345253, + 0.012302424758672714, + 0.017890801653265953, + 0.02375699020922184, + 0.012119591236114502, + 0.015637779608368874, + 0.004138975869864225, + 0.00384526327252388, + 0.003340053139254451, + 0.0035222184378653765, + 0.004082703497260809, + 0.004536739084869623, + 0.003592767985537648, + 0.0035078765358775854, + 0.009735248051583767, + 0.011182497255504131, + 0.008819900453090668, + 0.007806765381246805, + 0.013094396330416203, + 0.011069110594689846, + 0.010119018144905567, + 0.007340054959058762, + 0.015446565113961697, + 0.02402118407189846, + 0.012906871736049652, + 0.02019531838595867, + 0.03725043684244156, + 0.024112308397889137, + 0.022294212132692337, + 0.018039816990494728, + 0.0046910447999835014, + 0.00565633038058877, + 0.00521106505766511, + 0.00919041782617569, + 0.004860145039856434, + 0.004770898260176182, + 0.006757042836397886, + 0.005037597846239805, + 0.013128098100423813, + 0.015617306344211102, + 0.02462957240641117, + 0.032513998448848724, + 0.01938503608107567, + 0.019878489896655083, + 0.022396696731448174, + 0.022053034976124763, + 0.011884144507348537, + 0.00864983256906271, + 0.014513321220874786, + 0.021454613655805588, + 0.00974484346807003, + 0.015659362077713013, + 0.05665290728211403, + 0.04690894857048988, + 0.0036122642923146486, + 0.003803564002737403, + 0.006026134360581636, + 0.008187992498278618, + 0.004473171196877956, + 0.005389287602156401, + 0.005061042495071888, + 0.007689688354730606, + 0.007028079591691494, + 0.010298317298293114, + 0.011133390478789806, + 0.008097304962575436, + 0.00940022338181734, + 0.007969767786562443, + 0.00823152344673872, + 0.011120452545583248, + 0.007151877507567406, + 0.006483811419457197, + 0.009098279289901257, + 0.010318240150809288, + 0.0128556489944458, + 0.01587223820388317, + 0.007897697389125824, + 0.007315690163522959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.31538495421409607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822346519678831 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822346519678831, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822346519678831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1565641313791275, + 0.141116663813591, + 0.12817391753196716, + 0.05279857665300369, + 0.04382359981536865, + 0.1603608876466751, + 0.15319067239761353, + 0.02387138269841671, + 0.07932736724615097, + 0.06077273190021515 + ], + "credit_map": [ + 0.01437399536371231, + 0.02363990619778633, + 0.01215990912169218, + 0.01645255647599697, + 0.02141181379556656, + 0.02437404729425907, + 0.02619921788573265, + 0.017952701076865196, + 0.013787487521767616, + 0.01798182912170887, + 0.012666275724768639, + 0.02786615677177906, + 0.021173488348722458, + 0.013963675126433372, + 0.017972590401768684, + 0.015705173835158348, + 0.010186908766627312, + 0.01725921779870987, + 0.012382104992866516, + 0.02016443945467472, + 0.01856737583875656, + 0.023735668510198593, + 0.011159799061715603, + 0.014718415215611458, + 0.004383069463074207, + 0.00667722849175334, + 0.004354557953774929, + 0.00834002997726202, + 0.008101709187030792, + 0.00894351489841938, + 0.006860565859824419, + 0.005137904081493616, + 0.004253990948200226, + 0.006695195101201534, + 0.006692112423479557, + 0.006186549086123705, + 0.0035361202899366617, + 0.004855733830481768, + 0.007032932713627815, + 0.004570971243083477, + 0.016075531020760536, + 0.013750874437391758, + 0.021114205941557884, + 0.025675654411315918, + 0.016584012657403946, + 0.017573200166225433, + 0.02537883073091507, + 0.02420857921242714, + 0.007562495302408934, + 0.005516310688108206, + 0.006496800575405359, + 0.020090147852897644, + 0.006566846277564764, + 0.0071381679736077785, + 0.05514802411198616, + 0.04467189311981201, + 0.0025286246091127396, + 0.002272383077070117, + 0.0035067847929894924, + 0.0031833085231482983, + 0.002758409595116973, + 0.0033535079564899206, + 0.0028096104506403208, + 0.003458756022155285, + 0.008075183257460594, + 0.00876392051577568, + 0.013345175422728062, + 0.009883291088044643, + 0.007677353918552399, + 0.008915039710700512, + 0.010539875365793705, + 0.012127540074288845, + 0.0064551942050457, + 0.0070366268046200275, + 0.00779752666130662, + 0.008117524906992912, + 0.010711129754781723, + 0.010098353959619999, + 0.005295122042298317, + 0.0052612582221627235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.002822346519678831 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011811431031674147 + ], + "opposite_logprob_tokens": [ + -6.751181125640869 + ], + "logprob_reward": -0.0011811431031674147, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011811431031674147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06305725872516632, + 0.05353735759854317, + 0.04595058038830757, + 0.043886784464120865, + 0.07629106938838959, + 0.1592651754617691, + 0.26858532428741455, + 0.08317717164754868, + 0.0899747759103775, + 0.11627450585365295 + ], + "credit_map": [ + 0.008671099320054054, + 0.007963276468217373, + 0.005306370090693235, + 0.007255935110151768, + 0.007863491773605347, + 0.011210394091904163, + 0.0071288724429905415, + 0.007657818496227264, + 0.0090678995475173, + 0.006689961533993483, + 0.005684227682650089, + 0.0054326411336660385, + 0.007570870220661163, + 0.007138690911233425, + 0.007049061823636293, + 0.004904003813862801, + 0.004731900990009308, + 0.005867998581379652, + 0.005802082363516092, + 0.00523786898702383, + 0.008778492920100689, + 0.006004150491207838, + 0.005066049750894308, + 0.004462035838514566, + 0.003097446635365486, + 0.005505036562681198, + 0.0036791020538657904, + 0.0067734494805336, + 0.007150175515562296, + 0.005467011593282223, + 0.006270770914852619, + 0.005943790543824434, + 0.005885873921215534, + 0.009366185404360294, + 0.01056587416678667, + 0.010839048773050308, + 0.006060469895601273, + 0.007322702556848526, + 0.01620739884674549, + 0.01004351582378149, + 0.015133858658373356, + 0.014209870249032974, + 0.02505023404955864, + 0.021357592195272446, + 0.016602158546447754, + 0.023045731708407402, + 0.020627260208129883, + 0.02323845960199833, + 0.01513004396110773, + 0.0091584213078022, + 0.012303728610277176, + 0.02971442975103855, + 0.011888844892382622, + 0.012272532097995281, + 0.09770806133747101, + 0.08040926605463028, + 0.009058116935193539, + 0.007912707515060902, + 0.010986934415996075, + 0.011672955006361008, + 0.007491773925721645, + 0.011060884222388268, + 0.011339935474097729, + 0.013653869740664959, + 0.008265245705842972, + 0.012972882948815823, + 0.011796561069786549, + 0.012637165375053883, + 0.01093220990151167, + 0.008479481562972069, + 0.01195551734417677, + 0.012935714796185493, + 0.014856424182653427, + 0.007680128328502178, + 0.010684908367693424, + 0.017556337639689445, + 0.010264513082802296, + 0.03759622201323509, + 0.00645694974809885, + 0.011179021559655666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.0011811431031674147 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742838442325592 + ], + "opposite_logprob_tokens": [ + -0.9742838144302368 + ], + "logprob_reward": -0.4742838442325592, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.4742838442325592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08592875301837921, + 0.10835298895835876, + 0.158733531832695, + 0.10182232409715652, + 0.08878293633460999, + 0.08717645704746246, + 0.14755021035671234, + 0.07991992682218552, + 0.04887572303414345, + 0.09285721182823181 + ], + "credit_map": [ + 0.008951823227107525, + 0.013498513028025627, + 0.006136956624686718, + 0.00875863991677761, + 0.01185378897935152, + 0.016960246488451958, + 0.009462163783609867, + 0.010306616313755512, + 0.012748800218105316, + 0.012401419691741467, + 0.013327891007065773, + 0.0148740503937006, + 0.016819532960653305, + 0.012993920594453812, + 0.01318292785435915, + 0.012004437856376171, + 0.012423928827047348, + 0.020723428577184677, + 0.01565127819776535, + 0.017413657158613205, + 0.022694453597068787, + 0.034848589450120926, + 0.014316768385469913, + 0.02066143974661827, + 0.010043648071587086, + 0.011893412098288536, + 0.011655265465378761, + 0.011990377679467201, + 0.015174906700849533, + 0.017990989610552788, + 0.012689046561717987, + 0.010384677909314632, + 0.007162775844335556, + 0.009714413434267044, + 0.014065001159906387, + 0.01582631655037403, + 0.0067622363567352295, + 0.010411671362817287, + 0.01501673087477684, + 0.009823796339333057, + 0.00816241092979908, + 0.012907302007079124, + 0.013981964439153671, + 0.01382994744926691, + 0.008435435593128204, + 0.005298040807247162, + 0.012215749360620975, + 0.012345612049102783, + 0.013468636199831963, + 0.009778725914657116, + 0.01586805284023285, + 0.018928533419966698, + 0.009134364314377308, + 0.008106421679258347, + 0.04350678622722626, + 0.02875869907438755, + 0.008916998282074928, + 0.00939786247909069, + 0.008586614392697811, + 0.010135862976312637, + 0.008897490799427032, + 0.012990565039217472, + 0.00991825945675373, + 0.011076273396611214, + 0.0035642560105770826, + 0.004586223512887955, + 0.00567144388332963, + 0.00587888341397047, + 0.007276829332113266, + 0.005191846750676632, + 0.010112829506397247, + 0.006593410391360521, + 0.014491141773760319, + 0.006998962257057428, + 0.005551925394684076, + 0.012928544543683529, + 0.01346687413752079, + 0.025363510474562645, + 0.006387727800756693, + 0.007668526377528906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.4742838442325592 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428080052137375 + ], + "opposite_logprob_tokens": [ + -2.017807960510254 + ], + "logprob_reward": -0.1428080052137375, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428080052137375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04332459717988968, + 0.04870758578181267, + 0.09523117542266846, + 0.08480587601661682, + 0.0837380588054657, + 0.1432133913040161, + 0.14327837526798248, + 0.04890720173716545, + 0.07193145155906677, + 0.23686225712299347 + ], + "credit_map": [ + 0.004006478004157543, + 0.004732862580567598, + 0.003617540467530489, + 0.004111975431442261, + 0.005552209448069334, + 0.008544297888875008, + 0.0074710245244205, + 0.005288211163133383, + 0.008480802178382874, + 0.005614332854747772, + 0.0054974304512143135, + 0.007272834423929453, + 0.007224951405078173, + 0.005500856786966324, + 0.0048657553270459175, + 0.004250621423125267, + 0.009982715360820293, + 0.012586085125803947, + 0.009380507282912731, + 0.009324925020337105, + 0.019811009988188744, + 0.013048594817519188, + 0.010414822958409786, + 0.010682513937354088, + 0.007751970551908016, + 0.012375274673104286, + 0.005884002428501844, + 0.007972273975610733, + 0.020222749561071396, + 0.01352731790393591, + 0.010033528320491314, + 0.007038759998977184, + 0.0058410316705703735, + 0.009111211635172367, + 0.012437458150088787, + 0.013909108005464077, + 0.006487911567091942, + 0.007348463870584965, + 0.018431279808282852, + 0.01017159316688776, + 0.014488181099295616, + 0.01958027109503746, + 0.019418977200984955, + 0.01909278891980648, + 0.015771830454468727, + 0.01683506928384304, + 0.01743190363049507, + 0.020594365894794464, + 0.008663579821586609, + 0.009677113965153694, + 0.011293774470686913, + 0.015401199460029602, + 0.009757672436535358, + 0.016793152317404747, + 0.039504460990428925, + 0.03218742460012436, + 0.005344349425286055, + 0.005215901881456375, + 0.006375037133693695, + 0.005625901278108358, + 0.004934609401971102, + 0.006917842198163271, + 0.006662692874670029, + 0.007830864749848843, + 0.00641093123704195, + 0.007071422412991524, + 0.010857608169317245, + 0.009078916162252426, + 0.007939293049275875, + 0.009090091101825237, + 0.011524398811161518, + 0.009958790615200996, + 0.03676684573292732, + 0.021067317575216293, + 0.02749461494386196, + 0.02579960599541664, + 0.03758102282881737, + 0.037404485046863556, + 0.023152170702815056, + 0.02759619802236557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.1428080052137375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In captivating graphic novel style, a medium shot establishes an ancient wisteria vine, its forms rendered in bold, black, ink-like lines and flat, dramatic fields of color. The camera immediately begins a slow, deliberate push-in, creating an increasingly intense sense of focus and tension. The vine's woody, gnarled trunk, a twisted mass of dark brown and black, clings tenaciously to a stark, shadowed wall of muted, cool gray stone. From its dense branches, clusters of dense purple flowers dangle like drooping grapes, their colors a solid, unshaded field of deep purple and lavender. Suddenly, the entire plant shakes with a sharp, unnatural energy, a high-frequency tremor that spreads from the base of the trunk to the tips of the dangling flowers. As the camera continues its relentless advance, it isolates a specific flower that seems to be the center of the disturbance, vibrating with a distinctly unnatural intensity against the still, dark background. Its petals tremble as if filled with a silent, encompassing force, creating a palpable atmosphere of suspense and flourishing magic.", + "seed": 25361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_14/index_6128_seed_25361_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14254506112774834, + "vlm_logprob": -0.14254506112774834, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.14254506112774834, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38844606280326843 + ], + "opposite_logprob_tokens": [ + -1.1384460926055908 + ], + "logprob_reward": -0.38844606280326843, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38844606280326843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3578273355960846, + 0.07511091977357864, + 0.06296727061271667, + 0.08457829058170319, + 0.09788459539413452, + 0.06246718019247055, + 0.05507088452577591, + 0.05098187178373337, + 0.04855765774846077, + 0.10455398261547089 + ], + "credit_map": [ + 0.027602527290582657, + 0.03018498234450817, + 0.057276464998722076, + 0.04180894419550896, + 0.047642167657613754, + 0.04609799012541771, + 0.03919754922389984, + 0.06801676750183105, + 0.005912275984883308, + 0.00816840399056673, + 0.010222972370684147, + 0.013218230567872524, + 0.012215234339237213, + 0.007330274675041437, + 0.008159865625202656, + 0.009883667342364788, + 0.005014773458242416, + 0.005316731054335833, + 0.007658489979803562, + 0.011009519919753075, + 0.007218071259558201, + 0.005395959597080946, + 0.011565091088414192, + 0.0097886323928833, + 0.0075193727388978004, + 0.006454881746321917, + 0.008469169028103352, + 0.01859593577682972, + 0.008252584375441074, + 0.007847316563129425, + 0.01310912985354662, + 0.014329914003610611, + 0.0073972661048173904, + 0.007617019582539797, + 0.012845411896705627, + 0.029098525643348694, + 0.006457087118178606, + 0.007192441262304783, + 0.015960941091179848, + 0.011315915733575821, + 0.007972397841513157, + 0.0042075468227267265, + 0.010094721801578999, + 0.007472327910363674, + 0.004418056923896074, + 0.006285801064223051, + 0.013131245970726013, + 0.008885085582733154, + 0.007269004825502634, + 0.004009796306490898, + 0.008290287107229233, + 0.005518460180610418, + 0.003974997438490391, + 0.008200219832360744, + 0.006415559910237789, + 0.01139256451278925, + 0.005780580919235945, + 0.004807140678167343, + 0.007088367827236652, + 0.0055006761103868484, + 0.005549462512135506, + 0.010387610644102097, + 0.006096204277127981, + 0.005771834868937731, + 0.007224995642900467, + 0.0037639872170984745, + 0.007408521603792906, + 0.005111704580485821, + 0.005611357279121876, + 0.005822503473609686, + 0.006962778512388468, + 0.006651814561337233, + 0.010412242263555527, + 0.01170066837221384, + 0.01845617964863777, + 0.013712137006223202, + 0.010292837396264076, + 0.011908584274351597, + 0.014107930473983288, + 0.013963420875370502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.38844606280326843 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821514382958412 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821514382958412, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821514382958412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14738455414772034, + 0.041012704372406006, + 0.317321240901947, + 0.2582976520061493, + 0.043287429958581924, + 0.03754943609237671, + 0.03824945166707039, + 0.03560233861207962, + 0.053424399346113205, + 0.027870729565620422 + ], + "credit_map": [ + 0.0102893877774477, + 0.012937936931848526, + 0.01445736363530159, + 0.0223085917532444, + 0.020172078162431717, + 0.01927306316792965, + 0.020379550755023956, + 0.027566591277718544, + 0.0032472750172019005, + 0.004493295680731535, + 0.0050855414010584354, + 0.008450362831354141, + 0.0058774095959961414, + 0.003826900152489543, + 0.004576764069497585, + 0.005455157719552517, + 0.02075877971947193, + 0.03655070438981056, + 0.04130813106894493, + 0.07203054428100586, + 0.024204440414905548, + 0.020577162504196167, + 0.0482475571334362, + 0.053643934428691864, + 0.017547577619552612, + 0.02281837724149227, + 0.023704886436462402, + 0.06894581764936447, + 0.01997745968401432, + 0.015544097870588303, + 0.04835193604230881, + 0.041407518088817596, + 0.0050548589788377285, + 0.003397446358576417, + 0.007586127612739801, + 0.008739281445741653, + 0.0030426783487200737, + 0.002937723184004426, + 0.006417503580451012, + 0.006111812777817249, + 0.006176726892590523, + 0.002395882271230221, + 0.007690371014177799, + 0.004452356602996588, + 0.002354786964133382, + 0.004926021676510572, + 0.005162111483514309, + 0.004391181282699108, + 0.006195130757987499, + 0.002818852663040161, + 0.005910374689847231, + 0.0037354626692831516, + 0.0028286189772188663, + 0.004019410349428654, + 0.005058624781668186, + 0.007682978641241789, + 0.0036009771283715963, + 0.0034051057882606983, + 0.006058323197066784, + 0.0028530305717140436, + 0.0040700072422623634, + 0.00602642772719264, + 0.003995815292000771, + 0.005592655390501022, + 0.011346742510795593, + 0.0052984319627285, + 0.006344547029584646, + 0.004595459904521704, + 0.008931497111916542, + 0.0045827655121684074, + 0.006695656105875969, + 0.0056293015368282795, + 0.0025511260610073805, + 0.0030794430058449507, + 0.0031559045892208815, + 0.0039039580151438713, + 0.003130323486402631, + 0.0027384934946894646, + 0.005805057939141989, + 0.003506425069645047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.002821514382958412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00038675934774801135 + ], + "opposite_logprob_tokens": [ + -7.875386714935303 + ], + "logprob_reward": -0.00038675934774801135, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.00038675934774801135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18538843095302582, + 0.06815897673368454, + 0.10345245152711868, + 0.08919069916009903, + 0.06181466206908226, + 0.21644620597362518, + 0.08256746083498001, + 0.07838821411132812, + 0.05345078185200691, + 0.06114206090569496 + ], + "credit_map": [ + 0.015536974184215069, + 0.013503740541636944, + 0.03859701007604599, + 0.018830524757504463, + 0.019550681114196777, + 0.022818580269813538, + 0.014774261973798275, + 0.04177666828036308, + 0.00568281440064311, + 0.008366075344383717, + 0.008526379242539406, + 0.006788000930100679, + 0.014170447364449501, + 0.009128487668931484, + 0.006617724895477295, + 0.008879045024514198, + 0.007624902296811342, + 0.011656158603727818, + 0.012867283076047897, + 0.021887684240937233, + 0.008923142217099667, + 0.007840634323656559, + 0.015648240223526955, + 0.01700439676642418, + 0.006462064106017351, + 0.007087707985192537, + 0.007716462016105652, + 0.016959387809038162, + 0.011161289177834988, + 0.013354361057281494, + 0.01291760802268982, + 0.013531820848584175, + 0.0056044054217636585, + 0.005288781598210335, + 0.009236600250005722, + 0.014972269535064697, + 0.0052021220326423645, + 0.004583077505230904, + 0.010332134552299976, + 0.006595268379896879, + 0.03069741651415825, + 0.017837446182966232, + 0.03657865524291992, + 0.027928858995437622, + 0.015228800475597382, + 0.03051113151013851, + 0.03037828765809536, + 0.027285628020763397, + 0.009992473758757114, + 0.010618296451866627, + 0.010560351423919201, + 0.008624237962067127, + 0.007127170450985432, + 0.013264710083603859, + 0.00838138535618782, + 0.013998836278915405, + 0.008987728506326675, + 0.007892726920545101, + 0.014179617166519165, + 0.005661469884216785, + 0.009993989951908588, + 0.012035706080496311, + 0.0077798436395823956, + 0.011857138946652412, + 0.00755489245057106, + 0.006155701354146004, + 0.0067113302648067474, + 0.00627501355484128, + 0.007452827412635088, + 0.006761495023965836, + 0.007297483738511801, + 0.005242040380835533, + 0.005632293410599232, + 0.010338877327740192, + 0.008236163295805454, + 0.006349929608404636, + 0.007008995395153761, + 0.007479706313461065, + 0.006740528158843517, + 0.009355566464364529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.00038675934774801135 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344568729400635 + ], + "opposite_logprob_tokens": [ + -1.3134456872940063 + ], + "logprob_reward": -0.31344568729400635, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344568729400635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14501042664051056, + 0.04725303128361702, + 0.16931147873401642, + 0.0657913014292717, + 0.07864242047071457, + 0.06417834013700485, + 0.06674963980913162, + 0.10866895318031311, + 0.138438880443573, + 0.11595547944307327 + ], + "credit_map": [ + 0.009191149845719337, + 0.01591799035668373, + 0.015414581634104252, + 0.015546879731118679, + 0.022401748225092888, + 0.025830166414380074, + 0.017863871529698372, + 0.022844040766358376, + 0.003978862427175045, + 0.005388262216001749, + 0.006009673234075308, + 0.006248303689062595, + 0.008038165979087353, + 0.005049665924161673, + 0.005775525234639645, + 0.0067645711824297905, + 0.011319107376039028, + 0.01833520457148552, + 0.021011032164096832, + 0.03385166823863983, + 0.0201689675450325, + 0.012379252351820469, + 0.026267385110259056, + 0.025978868827223778, + 0.006717786192893982, + 0.004907155875116587, + 0.006441718433052301, + 0.013244077563285828, + 0.0059712850488722324, + 0.005304381251335144, + 0.011064963415265083, + 0.012139931321144104, + 0.010684874840080738, + 0.004200648982077837, + 0.013221816159784794, + 0.01757776364684105, + 0.0043446943163871765, + 0.007670364808291197, + 0.010534325614571571, + 0.010407931171357632, + 0.009900499135255814, + 0.00978755485266447, + 0.008348628878593445, + 0.006112216506153345, + 0.005684684030711651, + 0.007761368062347174, + 0.009782674722373486, + 0.006800714880228043, + 0.009030995890498161, + 0.005227786488831043, + 0.010996545664966106, + 0.008228851482272148, + 0.005880337208509445, + 0.0062326970510184765, + 0.011191468685865402, + 0.009960962459445, + 0.015021704137325287, + 0.008631523698568344, + 0.024715371429920197, + 0.008243049494922161, + 0.011713453568518162, + 0.0122040631249547, + 0.014164850115776062, + 0.013974940404295921, + 0.025418849661946297, + 0.01392858661711216, + 0.018003974109888077, + 0.01159432902932167, + 0.02454371005296707, + 0.011568191461265087, + 0.015482265502214432, + 0.01789897121489048, + 0.010786985978484154, + 0.017967229709029198, + 0.01935059204697609, + 0.014348994009196758, + 0.0163351371884346, + 0.010369339026510715, + 0.013031275011599064, + 0.013765930198132992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.31344568729400635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625281810760498 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625281810760498, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625281810760498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09524386376142502, + 0.04026056453585625, + 0.06151065602898598, + 0.05862044543027878, + 0.07896517962217331, + 0.16493722796440125, + 0.07871293276548386, + 0.21997690200805664, + 0.15625730156898499, + 0.04551489278674126 + ], + "credit_map": [ + 0.007807365618646145, + 0.007111573126167059, + 0.014290048740804195, + 0.014112956821918488, + 0.008964777924120426, + 0.010506956838071346, + 0.012384059838950634, + 0.02006613276898861, + 0.0028919633477926254, + 0.002942789811640978, + 0.004857080057263374, + 0.008350741118192673, + 0.004711744375526905, + 0.0036309189163148403, + 0.0059819892048835754, + 0.006893342826515436, + 0.005620771087706089, + 0.005389086902141571, + 0.007627661805599928, + 0.009030048735439777, + 0.004936728626489639, + 0.0056286100298166275, + 0.010326114483177662, + 0.012951638549566269, + 0.004755725152790546, + 0.003441874636337161, + 0.006313376594334841, + 0.012368682771921158, + 0.004498043563216925, + 0.006071458104997873, + 0.009147293865680695, + 0.012023995630443096, + 0.007885963656008244, + 0.005862197373062372, + 0.013417793437838554, + 0.0187367033213377, + 0.006065648514777422, + 0.004745176061987877, + 0.012509975582361221, + 0.009741728194057941, + 0.01616610772907734, + 0.013493919745087624, + 0.02561904862523079, + 0.016599727794528008, + 0.01135765016078949, + 0.029682708904147148, + 0.025142570957541466, + 0.026875510811805725, + 0.008975286036729813, + 0.009044290520250797, + 0.010473423637449741, + 0.009223239496350288, + 0.006536366418004036, + 0.012300824746489525, + 0.011744515039026737, + 0.010414996184408665, + 0.020990248769521713, + 0.020060580223798752, + 0.041811536997556686, + 0.019278261810541153, + 0.027593839913606644, + 0.028862789273262024, + 0.02455070987343788, + 0.03682895377278328, + 0.03395933285355568, + 0.016000866889953613, + 0.018874069675803185, + 0.013680507428944111, + 0.024432828649878502, + 0.016556724905967712, + 0.01695328950881958, + 0.015799695625901222, + 0.004579422529786825, + 0.004440885502845049, + 0.007261417806148529, + 0.004241074901074171, + 0.005635401234030724, + 0.007218474056571722, + 0.006055651698261499, + 0.006082568783313036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.007625281810760498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In captivating graphic novel style, a medium shot establishes an ancient wisteria vine, its forms rendered in bold, black, ink-like lines and flat, dramatic fields of color. The camera immediately begins a slow, deliberate push-in, creating an increasingly intense sense of focus and tension. The vine's woody, gnarled trunk, a twisted mass of dark brown and black, clings tenaciously to a stark, shadowed wall of muted, cool gray stone. From its dense branches, clusters of dense purple flowers dangle like drooping grapes, their colors a solid, unshaded field of deep purple and lavender. Suddenly, the entire plant shakes with a sharp, unnatural energy, a high-frequency tremor that spreads from the base of the trunk to the tips of the dangling flowers. As the camera continues its relentless advance, it isolates a specific flower that seems to be the center of the disturbance, vibrating with a distinctly unnatural intensity against the still, dark background. Its petals tremble as if filled with a silent, encompassing force, creating a palpable atmosphere of suspense and flourishing magic.", + "seed": 25362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_14/index_6128_seed_25362_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23351909139892085, + "vlm_logprob": -0.23351909139892085, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.23351909139892085, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18253059685230255 + ], + "opposite_logprob_tokens": [ + -1.8075306415557861 + ], + "logprob_reward": -0.18253059685230255, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18253059685230255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33383625745773315, + 0.09995929896831512, + 0.12737953662872314, + 0.07661748677492142, + 0.10948523879051208, + 0.07300976663827896, + 0.03028745763003826, + 0.045410141348838806, + 0.06845678389072418, + 0.03555799275636673 + ], + "credit_map": [ + 0.023181932047009468, + 0.01820143684744835, + 0.05832495167851448, + 0.030396904796361923, + 0.01734016276896, + 0.024526182562112808, + 0.059579990804195404, + 0.10228472948074341, + 0.015012922696769238, + 0.012342269532382488, + 0.011495026759803295, + 0.013512121513485909, + 0.005916578229516745, + 0.007968473248183727, + 0.016799569129943848, + 0.016912341117858887, + 0.011749610304832458, + 0.015639062970876694, + 0.015709511935710907, + 0.013206658884882927, + 0.011968004517257214, + 0.02088494785130024, + 0.020336102694272995, + 0.017885643988847733, + 0.006636479869484901, + 0.015226762741804123, + 0.010612850077450275, + 0.009819531813263893, + 0.0040873149409890175, + 0.008688915520906448, + 0.010263538919389248, + 0.011282091960310936, + 0.011878889985382557, + 0.01889483816921711, + 0.01236927229911089, + 0.02313687838613987, + 0.006827679928392172, + 0.008934383280575275, + 0.01101744081825018, + 0.016425859183073044, + 0.011547248810529709, + 0.0043740663677453995, + 0.011137471534311771, + 0.017761308699846268, + 0.0032389501575380564, + 0.0064580319449305534, + 0.006141342222690582, + 0.01235134620219469, + 0.0044798580929636955, + 0.001557616633363068, + 0.005289750639349222, + 0.00680319732055068, + 0.0013927489053457975, + 0.0035367722157388926, + 0.0028078665491193533, + 0.00441964715719223, + 0.008297373540699482, + 0.0022349876817315817, + 0.008046280592679977, + 0.009397193789482117, + 0.002243675757199526, + 0.004011004697531462, + 0.00272583682090044, + 0.008453787304461002, + 0.006164307706058025, + 0.0031368390191346407, + 0.01593972183763981, + 0.011506968177855015, + 0.0033092345111072063, + 0.006313275545835495, + 0.006332061253488064, + 0.01575438305735588, + 0.0018005468882620335, + 0.002146065002307296, + 0.010698075406253338, + 0.008554138243198395, + 0.0016389614902436733, + 0.002263231435790658, + 0.0022989376448094845, + 0.006158034782856703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.18253059685230255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011835244949907064 + ], + "opposite_logprob_tokens": [ + -6.75118350982666 + ], + "logprob_reward": -0.0011835244949907064, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011835244949907064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3764856457710266, + 0.10407381504774094, + 0.07115311920642853, + 0.05587361007928848, + 0.10936665534973145, + 0.07130610197782516, + 0.06335826963186264, + 0.04110189154744148, + 0.0519934743642807, + 0.0552874319255352 + ], + "credit_map": [ + 0.027672499418258667, + 0.035209234803915024, + 0.05855340138077736, + 0.045526716858148575, + 0.014567995443940163, + 0.023800794035196304, + 0.06700786203145981, + 0.10414712131023407, + 0.00929808896034956, + 0.017373064532876015, + 0.008948213420808315, + 0.019087007269263268, + 0.008859992027282715, + 0.008523082360625267, + 0.015258802101016045, + 0.016725560650229454, + 0.005011863075196743, + 0.013735650107264519, + 0.008227032609283924, + 0.010876178741455078, + 0.003830881556496024, + 0.005789677146822214, + 0.013464108109474182, + 0.010217717848718166, + 0.006533080246299505, + 0.008711526170372963, + 0.0076418230310082436, + 0.006418044678866863, + 0.0032681026495993137, + 0.00918370671570301, + 0.006918048486113548, + 0.007199273444712162, + 0.010785546153783798, + 0.012729864567518234, + 0.014941778965294361, + 0.021751876920461655, + 0.007115447428077459, + 0.014077730476856232, + 0.011899041943252087, + 0.016065355390310287, + 0.009563925676047802, + 0.006228185724467039, + 0.007734576240181923, + 0.011467216536402702, + 0.004971165210008621, + 0.015109842643141747, + 0.006317516323179007, + 0.009913671761751175, + 0.007008301559835672, + 0.0036205195356160402, + 0.009758101776242256, + 0.013520468026399612, + 0.004041436593979597, + 0.009831107221543789, + 0.005706840194761753, + 0.009871483780443668, + 0.009550508111715317, + 0.0029660528525710106, + 0.008555255830287933, + 0.006162288598716259, + 0.0022462322376668453, + 0.003038778668269515, + 0.002867317060008645, + 0.005715452134609222, + 0.005454517900943756, + 0.0026194772217422724, + 0.011328008957207203, + 0.0114592295140028, + 0.0024295630864799023, + 0.0035843125078827143, + 0.0033424419816583395, + 0.011775916442275047, + 0.0030647930689156055, + 0.003714397083967924, + 0.02200276590883732, + 0.0120988879352808, + 0.002330692019313574, + 0.0026809340342879295, + 0.0033074580132961273, + 0.006087502930313349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0011835244949907064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004959068610332906 + ], + "opposite_logprob_tokens": [ + -7.625495910644531 + ], + "logprob_reward": -0.0004959068610332906, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004959068610332906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21964654326438904, + 0.0679156482219696, + 0.0918518528342247, + 0.11799772828817368, + 0.11086034029722214, + 0.07554928958415985, + 0.06098997965455055, + 0.07585910707712173, + 0.1180839017033577, + 0.061245620250701904 + ], + "credit_map": [ + 0.01484030019491911, + 0.01125131081789732, + 0.04127729684114456, + 0.017989568412303925, + 0.006901323329657316, + 0.015241263434290886, + 0.04034581035375595, + 0.07179965823888779, + 0.005482764914631844, + 0.007898719981312752, + 0.004098749253898859, + 0.011256351135671139, + 0.006072046235203743, + 0.007366973906755447, + 0.013187738135457039, + 0.012552299536764622, + 0.0062273843213915825, + 0.014321479946374893, + 0.009402078576385975, + 0.014119871892035007, + 0.006168860476464033, + 0.009752076119184494, + 0.018186254426836967, + 0.013673842884600163, + 0.010208535008132458, + 0.021096181124448776, + 0.01536490861326456, + 0.013285367749631405, + 0.006580354180186987, + 0.018538106232881546, + 0.013784044422209263, + 0.019140226766467094, + 0.012881823815405369, + 0.015048021450638771, + 0.017512697726488113, + 0.022373246029019356, + 0.005797252524644136, + 0.010456233285367489, + 0.01002831943333149, + 0.016762735322117805, + 0.009993808344006538, + 0.006024766713380814, + 0.007827017456293106, + 0.012569626793265343, + 0.005867235362529755, + 0.016107672825455666, + 0.005961277522146702, + 0.011197881773114204, + 0.006566578987985849, + 0.004270743578672409, + 0.008863117545843124, + 0.012614503502845764, + 0.004681423772126436, + 0.008967025205492973, + 0.005955507978796959, + 0.009071070700883865, + 0.01836651749908924, + 0.0038244491443037987, + 0.012735458090901375, + 0.010443232953548431, + 0.004912867676466703, + 0.006345827132463455, + 0.005949686281383038, + 0.013281062245368958, + 0.015608469024300575, + 0.00914410687983036, + 0.014826287515461445, + 0.023769505321979523, + 0.011594535782933235, + 0.014471723698079586, + 0.01363495085388422, + 0.015034318901598454, + 0.0030612037517130375, + 0.0045340051874518394, + 0.025894878432154655, + 0.0136257354170084, + 0.002051170915365219, + 0.003343815915286541, + 0.0031847856007516384, + 0.005550021771341562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0004959068610332906 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578043937683105 + ], + "opposite_logprob_tokens": [ + -0.6328043937683105 + ], + "logprob_reward": -0.7578043937683105, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578043937683105, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.071627177298069, + 0.16461467742919922, + 0.1220540925860405, + 0.10150159895420074, + 0.13975781202316284, + 0.07217691838741302, + 0.06793246418237686, + 0.0657103881239891, + 0.13162647187709808, + 0.06299839168787003 + ], + "credit_map": [ + 0.006275437772274017, + 0.010578097775578499, + 0.005478198640048504, + 0.010895026847720146, + 0.005470677278935909, + 0.006737660150974989, + 0.012151043862104416, + 0.014041037298738956, + 0.01838110201060772, + 0.01938219927251339, + 0.01438915729522705, + 0.02274491637945175, + 0.0106330756098032, + 0.01725868694484234, + 0.032365839928388596, + 0.02945970743894577, + 0.007227049209177494, + 0.023060057312250137, + 0.01462844479829073, + 0.023410223424434662, + 0.005585414357483387, + 0.006506213452666998, + 0.022276828065514565, + 0.019359849393367767, + 0.009600055404007435, + 0.021503424271941185, + 0.014982224442064762, + 0.013259712606668472, + 0.0049757808446884155, + 0.009082665666937828, + 0.013234288431704044, + 0.014863453805446625, + 0.015528597868978977, + 0.017467336729168892, + 0.02360774204134941, + 0.027941280975937843, + 0.007143368013203144, + 0.014584350399672985, + 0.013275918550789356, + 0.02020922489464283, + 0.007670009974390268, + 0.0052063679322600365, + 0.007311955094337463, + 0.01326174009591341, + 0.004964184015989304, + 0.015252926386892796, + 0.008014312945306301, + 0.010495418682694435, + 0.007774176076054573, + 0.005915147718042135, + 0.008366452530026436, + 0.012366021983325481, + 0.006208919920027256, + 0.011203435249626637, + 0.007075203116983175, + 0.009023109450936317, + 0.01104759331792593, + 0.0062248120084404945, + 0.01135107222944498, + 0.010655222460627556, + 0.004671378992497921, + 0.0054864780977368355, + 0.005991592071950436, + 0.010282243601977825, + 0.017616983503103256, + 0.009171633049845695, + 0.01906495727598667, + 0.019524341449141502, + 0.011234126053750515, + 0.01807592809200287, + 0.0169378612190485, + 0.020000647753477097, + 0.005126518663018942, + 0.007868719287216663, + 0.011999350972473621, + 0.008863492868840694, + 0.005035772919654846, + 0.009387449361383915, + 0.00633662985637784, + 0.008380462415516376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.7578043937683105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558103501796722 + ], + "opposite_logprob_tokens": [ + -1.6005810499191284 + ], + "logprob_reward": -0.22558103501796722, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558103501796722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10914988815784454, + 0.21007975935935974, + 0.06412679702043533, + 0.06323543936014175, + 0.059628214687108994, + 0.09003934264183044, + 0.07705116271972656, + 0.11488290876150131, + 0.12777753174304962, + 0.0840289443731308 + ], + "credit_map": [ + 0.007077307440340519, + 0.0064363800920546055, + 0.017531586810946465, + 0.010037391446530819, + 0.005908985156565905, + 0.009185584262013435, + 0.02008582092821598, + 0.03288684040307999, + 0.023123765364289284, + 0.03284741938114166, + 0.024090034887194633, + 0.031068146228790283, + 0.013180672191083431, + 0.015846244990825653, + 0.03578164428472519, + 0.03414183482527733, + 0.0057066953741014, + 0.011616322211921215, + 0.007352240849286318, + 0.010778396390378475, + 0.004207646008580923, + 0.0057991063222289085, + 0.009448213502764702, + 0.009218176826834679, + 0.00868156086653471, + 0.00777636980637908, + 0.008209590800106525, + 0.006586216390132904, + 0.005140406545251608, + 0.010768604464828968, + 0.007817517966032028, + 0.008255178108811378, + 0.006513696629554033, + 0.008677889592945576, + 0.008709876798093319, + 0.01325181219726801, + 0.0024522768799215555, + 0.005289472173899412, + 0.006018566433340311, + 0.008714620023965836, + 0.013098453171551228, + 0.005723722279071808, + 0.014177781529724598, + 0.019858496263623238, + 0.005103216040879488, + 0.013044323772192001, + 0.006330136675387621, + 0.012703211978077888, + 0.012495940551161766, + 0.0065277209505438805, + 0.009889273904263973, + 0.016055958345532417, + 0.0060721938498318195, + 0.009610298089683056, + 0.0058236075565218925, + 0.010576172731816769, + 0.02774455025792122, + 0.00560784013941884, + 0.02379745990037918, + 0.0207611583173275, + 0.005737545900046825, + 0.006879471242427826, + 0.009583679027855396, + 0.01477120816707611, + 0.014588328078389168, + 0.007846401073038578, + 0.02147929184138775, + 0.019538814201951027, + 0.008785063400864601, + 0.016741709783673286, + 0.014501592144370079, + 0.024296337738633156, + 0.007479146588593721, + 0.01172074768692255, + 0.015836454927921295, + 0.011812139302492142, + 0.006352998781949282, + 0.011857425794005394, + 0.008467941544950008, + 0.010502098128199577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.22558103501796722 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In captivating graphic novel style, a medium shot establishes an ancient wisteria vine, its forms rendered in bold, black, ink-like lines and flat, dramatic fields of color. The camera immediately begins a slow, deliberate push-in, creating an increasingly intense sense of focus and tension. The vine's woody, gnarled trunk, a twisted mass of dark brown and black, clings tenaciously to a stark, shadowed wall of muted, cool gray stone. From its dense branches, clusters of dense purple flowers dangle like drooping grapes, their colors a solid, unshaded field of deep purple and lavender. Suddenly, the entire plant shakes with a sharp, unnatural energy, a high-frequency tremor that spreads from the base of the trunk to the tips of the dangling flowers. As the camera continues its relentless advance, it isolates a specific flower that seems to be the center of the disturbance, vibrating with a distinctly unnatural intensity against the still, dark background. Its petals tremble as if filled with a silent, encompassing force, creating a palpable atmosphere of suspense and flourishing magic.", + "seed": 25363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_14/index_6128_seed_25363_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16812086800346152, + "vlm_logprob": -0.16812086800346152, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.16812086800346152, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18219035863876343 + ], + "opposite_logprob_tokens": [ + -1.8071904182434082 + ], + "logprob_reward": -0.18219035863876343, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18219035863876343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1340045928955078, + 0.09634651988744736, + 0.0609634630382061, + 0.07184592634439468, + 0.08874239772558212, + 0.036043066531419754, + 0.05350630357861519, + 0.18988163769245148, + 0.22728756070137024, + 0.04137846454977989 + ], + "credit_map": [ + 0.016228072345256805, + 0.010641425848007202, + 0.026933327317237854, + 0.011007643304765224, + 0.00986017007380724, + 0.017244825139641762, + 0.013483238406479359, + 0.028605885803699493, + 0.00685998285189271, + 0.008977687917649746, + 0.016972748562693596, + 0.013574793003499508, + 0.010110420174896717, + 0.014361187815666199, + 0.01177800353616476, + 0.013711693696677685, + 0.005883520003408194, + 0.006811616942286491, + 0.011201414279639721, + 0.008082694374024868, + 0.00469396635890007, + 0.004867127165198326, + 0.01092828344553709, + 0.008494838140904903, + 0.005478044040501118, + 0.007230559829622507, + 0.011369355954229832, + 0.008940068073570728, + 0.005471402313560247, + 0.007601872086524963, + 0.01566367596387863, + 0.010090945288538933, + 0.007644516881555319, + 0.012814505025744438, + 0.01064333040267229, + 0.007820305414497852, + 0.010595719330012798, + 0.007523671258240938, + 0.019479401409626007, + 0.012220945209264755, + 0.004819533322006464, + 0.0033617771696299314, + 0.0036414244677871466, + 0.0032062355894595385, + 0.004804510623216629, + 0.0034309825859963894, + 0.006238909903913736, + 0.00653969356790185, + 0.006926394533365965, + 0.006879955064505339, + 0.005984639748930931, + 0.009835383854806423, + 0.006904332432895899, + 0.006182348821312189, + 0.005434304475784302, + 0.005358944181352854, + 0.016002807766199112, + 0.01777508296072483, + 0.022665709257125854, + 0.03497682139277458, + 0.0214192233979702, + 0.028878573328256607, + 0.02298724465072155, + 0.02517617680132389, + 0.021009741351008415, + 0.017749851569533348, + 0.025198280811309814, + 0.02619031071662903, + 0.02151123620569706, + 0.05949006974697113, + 0.0266571007668972, + 0.02948097512125969, + 0.004685367923229933, + 0.00368366832844913, + 0.004665512591600418, + 0.004520963411778212, + 0.005487225949764252, + 0.00636181328445673, + 0.0064375633373856544, + 0.005536350887268782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.18219035863876343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024917051196098328 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024917051196098328, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024917051196098328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15663990378379822, + 0.1820221096277237, + 0.09408380091190338, + 0.05826939269900322, + 0.06541577726602554, + 0.03544772043824196, + 0.04580197110772133, + 0.14309176802635193, + 0.17570577561855316, + 0.043521784245967865 + ], + "credit_map": [ + 0.012480588629841805, + 0.011867995373904705, + 0.039082903414964676, + 0.013516131788492203, + 0.013669384643435478, + 0.01350904069840908, + 0.016731709241867065, + 0.03578215464949608, + 0.012646014802157879, + 0.020213237032294273, + 0.03767295554280281, + 0.029069334268569946, + 0.013474319130182266, + 0.014266914688050747, + 0.027868418022990227, + 0.026810912415385246, + 0.007832464762032032, + 0.009591666050255299, + 0.011856804601848125, + 0.012814339250326157, + 0.00818540994077921, + 0.009371524676680565, + 0.020810004323720932, + 0.013621586374938488, + 0.004869287833571434, + 0.004565181210637093, + 0.010963276028633118, + 0.007140904199331999, + 0.0039863684214651585, + 0.005788540933281183, + 0.012657827697694302, + 0.00829800870269537, + 0.005167362745851278, + 0.006150534842163324, + 0.007461540400981903, + 0.0058114659041166306, + 0.006666953209787607, + 0.006739431992173195, + 0.01842299848794937, + 0.008995492942631245, + 0.005321957636624575, + 0.003698378801345825, + 0.0038426239043474197, + 0.003629237413406372, + 0.004308376926928759, + 0.002972750458866358, + 0.005276545416563749, + 0.006397848948836327, + 0.006111114285886288, + 0.005261983256787062, + 0.005197831429541111, + 0.006913406774401665, + 0.006041559856384993, + 0.005494925193488598, + 0.00504253851249814, + 0.00573861226439476, + 0.012591137550771236, + 0.013272815383970737, + 0.017218902707099915, + 0.02687707357108593, + 0.015332183800637722, + 0.02103087119758129, + 0.016835292801260948, + 0.01993349939584732, + 0.01590336300432682, + 0.013600494712591171, + 0.0196113009005785, + 0.019949741661548615, + 0.017060620710253716, + 0.04550528526306152, + 0.020872287452220917, + 0.0232026856392622, + 0.0074347457848489285, + 0.0035714299883693457, + 0.004928780719637871, + 0.004265518859028816, + 0.005504913162440062, + 0.007689691614359617, + 0.005243007093667984, + 0.004883696790784597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0024917051196098328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008126770262606442 + ], + "opposite_logprob_tokens": [ + -7.125812530517578 + ], + "logprob_reward": -0.0008126770262606442, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008126770262606442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16567684710025787, + 0.09729494154453278, + 0.12958483397960663, + 0.07146621495485306, + 0.054120589047670364, + 0.1131298616528511, + 0.06861605495214462, + 0.08146458119153976, + 0.14644382894039154, + 0.07220222055912018 + ], + "credit_map": [ + 0.023859785869717598, + 0.013667176477611065, + 0.029255429282784462, + 0.016355492174625397, + 0.01766314171254635, + 0.02544831670820713, + 0.017281072214245796, + 0.02214644104242325, + 0.00708381412550807, + 0.011262485757470131, + 0.015407409518957138, + 0.013013127259910107, + 0.012101195752620697, + 0.012017497792840004, + 0.012694411911070347, + 0.01371501013636589, + 0.009086702018976212, + 0.017210718244314194, + 0.015781261026859283, + 0.01735680364072323, + 0.009588165208697319, + 0.010691480711102486, + 0.031267717480659485, + 0.01860199309885502, + 0.006092099007219076, + 0.01027692761272192, + 0.009864968247711658, + 0.011549416929483414, + 0.004805343225598335, + 0.005849255248904228, + 0.013342955149710178, + 0.009685256518423557, + 0.0047243013978004456, + 0.006861092988401651, + 0.006743896286934614, + 0.006424937397241592, + 0.006248319987207651, + 0.006338344421237707, + 0.009573913179337978, + 0.0072057913057506084, + 0.014838661067187786, + 0.01152565237134695, + 0.015485013835132122, + 0.014392376877367496, + 0.010020527988672256, + 0.010293933562934399, + 0.02077682688832283, + 0.015796884894371033, + 0.010483874939382076, + 0.010597477667033672, + 0.008392379619181156, + 0.013126074336469173, + 0.007019569166004658, + 0.006025336217135191, + 0.006245143245905638, + 0.006726207211613655, + 0.00740877166390419, + 0.00775446742773056, + 0.0107375243678689, + 0.016804946586489677, + 0.01064215786755085, + 0.010933391749858856, + 0.006977585144340992, + 0.010205741971731186, + 0.013271463103592396, + 0.014147316105663776, + 0.01664957031607628, + 0.016877131536602974, + 0.014623897150158882, + 0.03599641099572182, + 0.01600981503725052, + 0.01886824704706669, + 0.012334609404206276, + 0.0073954383842647076, + 0.008190174587070942, + 0.006517346017062664, + 0.00885822344571352, + 0.014653329737484455, + 0.006491764448583126, + 0.007761341519653797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0008126770262606442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761370062828064 + ], + "opposite_logprob_tokens": [ + -0.8261370062828064 + ], + "logprob_reward": -0.5761370062828064, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761370062828064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3145385980606079, + 0.1465163677930832, + 0.07533806562423706, + 0.06283578276634216, + 0.08351793885231018, + 0.08962182700634003, + 0.03437194228172302, + 0.06895061582326889, + 0.07390160113573074, + 0.050407279282808304 + ], + "credit_map": [ + 0.04189717769622803, + 0.021477894857525826, + 0.07819990813732147, + 0.019234083592891693, + 0.0246885959059, + 0.03998168185353279, + 0.01974516361951828, + 0.06931409239768982, + 0.012139059603214264, + 0.015121682547032833, + 0.021839557215571404, + 0.026950344443321228, + 0.016157377511262894, + 0.01834931969642639, + 0.016322674229741096, + 0.019636349752545357, + 0.0058128186501562595, + 0.0061165099032223225, + 0.01017849426716566, + 0.008170530200004578, + 0.006309491582214832, + 0.007663982454687357, + 0.019285988062620163, + 0.011800253763794899, + 0.004589705262333155, + 0.009178200736641884, + 0.007586710620671511, + 0.010910702869296074, + 0.004569859243929386, + 0.006355490069836378, + 0.010452130809426308, + 0.009192978963255882, + 0.007068711798638105, + 0.0090033495798707, + 0.009867153130471706, + 0.006894315592944622, + 0.008961549960076809, + 0.007321572862565517, + 0.02277974970638752, + 0.011621540412306786, + 0.013693857938051224, + 0.009664119221270084, + 0.012981871142983437, + 0.011700083501636982, + 0.007328611332923174, + 0.007546895649284124, + 0.014121884480118752, + 0.012584502808749676, + 0.004484823439270258, + 0.0055693951435387135, + 0.004185461439192295, + 0.005556322634220123, + 0.003283084835857153, + 0.00343068758957088, + 0.00360690476372838, + 0.004255262669175863, + 0.006086344365030527, + 0.006381528917700052, + 0.007621776312589645, + 0.009694220498204231, + 0.00849703885614872, + 0.012325402349233627, + 0.008471605367958546, + 0.009872698225080967, + 0.006038398947566748, + 0.005081817973405123, + 0.006801382172852755, + 0.008350543677806854, + 0.009365959092974663, + 0.019563235342502594, + 0.009885470382869244, + 0.008814792148768902, + 0.008827773854136467, + 0.003147275885567069, + 0.008815654553472996, + 0.005225635599344969, + 0.006167342886328697, + 0.007996113039553165, + 0.004789514467120171, + 0.005437971092760563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.5761370062828064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a graphic novel style with bold black ink-like lines and flat fields of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ancient wisteria vine with a woody, gnarled trunk in dark brown and black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are clusters of dense purple flowers visible dangling from the vine's branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant shake with a sharp, unnatural high-frequency tremor spreading from the base to the tips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in that isolates a specific vibrating flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897259294986725 + ], + "opposite_logprob_tokens": [ + -2.578972578048706 + ], + "logprob_reward": -0.07897259294986725, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897259294986725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1547238975763321, + 0.10788246244192123, + 0.11989256739616394, + 0.08216378092765808, + 0.05886302888393402, + 0.14788632094860077, + 0.15744806826114655, + 0.07236077636480331, + 0.05668091028928757, + 0.04209814965724945 + ], + "credit_map": [ + 0.019276868551969528, + 0.016980847343802452, + 0.030451752245426178, + 0.014827253296971321, + 0.012684809975326061, + 0.01438374537974596, + 0.014457661658525467, + 0.03166096284985542, + 0.0070428098551929, + 0.010953612625598907, + 0.022342044860124588, + 0.013597694225609303, + 0.012440619990229607, + 0.009435830637812614, + 0.01702931337058544, + 0.015040541999042034, + 0.01044229045510292, + 0.013094806112349033, + 0.015085157938301563, + 0.018264329060912132, + 0.01135542243719101, + 0.013129927217960358, + 0.021910198032855988, + 0.016610439866781235, + 0.0055415574461221695, + 0.010077808983623981, + 0.009300459176301956, + 0.014148078858852386, + 0.005533295683562756, + 0.008123795501887798, + 0.01748768240213394, + 0.011951095424592495, + 0.004410003311932087, + 0.007973182015120983, + 0.008245081640779972, + 0.0070885890163481236, + 0.004687292501330376, + 0.00430506793782115, + 0.012382997199892998, + 0.009770819917321205, + 0.017500342801213264, + 0.013456403277814388, + 0.01440648827701807, + 0.018693270161747932, + 0.010946198366582394, + 0.018714070320129395, + 0.020725945010781288, + 0.03344361111521721, + 0.016675332561135292, + 0.016588540747761726, + 0.02149067632853985, + 0.033642206341028214, + 0.010821089148521423, + 0.018985150381922722, + 0.015384411439299583, + 0.023860657587647438, + 0.007549348287284374, + 0.005773460958153009, + 0.008952045813202858, + 0.014216513372957706, + 0.008776830509305, + 0.008327009156346321, + 0.011082667857408524, + 0.0076829008758068085, + 0.006007840391248465, + 0.005169217474758625, + 0.005386509466916323, + 0.006512572523206472, + 0.005975006613880396, + 0.013854737393558025, + 0.0058361319825053215, + 0.00793889444321394, + 0.004273354075849056, + 0.0054430230520665646, + 0.005586426239460707, + 0.005052172113209963, + 0.0057848235592246056, + 0.0049862368032336235, + 0.005120741669088602, + 0.005851373076438904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.07897259294986725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_15/index_tensor([9977, 9977])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_15/index_tensor([9977, 9977])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..49a9e9f1571431db01eebf757b0d2ab9bf4623d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_15/index_tensor([9977, 9977])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Set in a bright and cheerful 3D cartoon-style world, a dynamic low-angle tracking shot follows a small wooden rowboat as it plows through the waves. Poised at the prow is a proud black cat, sleek and black with a sleek digitally rendered texture. A bright red triangular scarf flutters in the breeze, and large, luminous emerald green eyes gaze resolutely into the distance. Constructed of warm brown wood planks with distinct grain, the boat sails across a shimmering azure sea, its bow constantly cleaving through crystal-clear white waves, sending stylized cartoon water droplets splattering across the clear sky, while a few fluffy, cotton-candy-like clouds punctuate the clear sky. The cat holds its head high, effortlessly balancing as the boat rhythmically rises and falls, its tail held high in a playful curl, radiating the confidence and authority of a captain. The brilliant sunlight pours down, infusing the entire scene with extremely high color saturation, creating soft highlights on the cat's smooth back and the wet side of the boat, and creating a joyful atmosphere full of vitality and optimistic adventurous spirit.", + "seed": 20495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_15/index_9977_seed_20495_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0796273882035166, + "vlm_logprob": -0.0796273882035166, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.0796273882035166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005692224483937025 + ], + "opposite_logprob_tokens": [ + -7.755692005157471 + ], + "logprob_reward": -0.005692224483937025, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.005692224483937025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1403890997171402, + 0.03423851728439331, + 0.05359824001789093, + 0.025593755766749382, + 0.07108283787965775, + 0.3014873266220093, + 0.0700560137629509, + 0.036906030029058456, + 0.11404627561569214, + 0.15260186791419983 + ], + "credit_map": [ + 0.019696924835443497, + 0.019685158506035805, + 0.02661842294037342, + 0.012142667546868324, + 0.017646631225943565, + 0.0186093058437109, + 0.0085084717720747, + 0.017481524497270584, + 0.00436027254909277, + 0.0044379932805895805, + 0.004630472511053085, + 0.004729012493044138, + 0.005176709499210119, + 0.004277053754776716, + 0.0023932636249810457, + 0.004233743064105511, + 0.009562092833220959, + 0.0044810413382947445, + 0.014482758939266205, + 0.0035155529621988535, + 0.00493574095889926, + 0.007268672343343496, + 0.00370346219278872, + 0.005648919381201267, + 0.005196769721806049, + 0.0023628261405974627, + 0.0035185744054615498, + 0.0022369101643562317, + 0.003065153257921338, + 0.004673016723245382, + 0.00203756638802588, + 0.0025029415264725685, + 0.011757249012589455, + 0.016456400975584984, + 0.008290036581456661, + 0.0054069096222519875, + 0.008745185099542141, + 0.010100330226123333, + 0.005419913679361343, + 0.004906811751425266, + 0.098324254155159, + 0.0261001568287611, + 0.01479793805629015, + 0.007288179825991392, + 0.02599986083805561, + 0.034333236515522, + 0.011437000706791878, + 0.08320672065019608, + 0.010799972340464592, + 0.013298659585416317, + 0.01699724607169628, + 0.0060890777967870235, + 0.006299747619777918, + 0.0069585395976901054, + 0.004212360363453627, + 0.005400415044277906, + 0.004203916061669588, + 0.004473425447940826, + 0.004604277666658163, + 0.0052135903388261795, + 0.005713148973882198, + 0.005742067936807871, + 0.0027270743157714605, + 0.004228533245623112, + 0.0270222220569849, + 0.009879677556455135, + 0.03652867302298546, + 0.00876433216035366, + 0.008838731795549393, + 0.012638314627110958, + 0.005016244016587734, + 0.0053580827079713345, + 0.05430484563112259, + 0.011246105656027794, + 0.016558390110731125, + 0.005108985118567944, + 0.00845028180629015, + 0.045526184141635895, + 0.005301240365952253, + 0.006105845794081688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.005692224483937025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870270252227783 + ], + "opposite_logprob_tokens": [ + -1.1370270252227783 + ], + "logprob_reward": -0.3870270252227783, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870270252227783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12947997450828552, + 0.07350493222475052, + 0.06244448199868202, + 0.13946396112442017, + 0.3092670142650604, + 0.03656657040119171, + 0.0942559465765953, + 0.045317474752664566, + 0.07735161483287811, + 0.03234809264540672 + ], + "credit_map": [ + 0.01990833692252636, + 0.015489256009459496, + 0.018831126391887665, + 0.014781147241592407, + 0.0169631689786911, + 0.01999286375939846, + 0.009674927219748497, + 0.01383914239704609, + 0.010100406594574451, + 0.006794854998588562, + 0.01438085176050663, + 0.007929513230919838, + 0.006689115427434444, + 0.011332242749631405, + 0.00393030047416687, + 0.012347648851573467, + 0.008064781315624714, + 0.010496634989976883, + 0.0069757746532559395, + 0.005374082364141941, + 0.009599943645298481, + 0.008479627780616283, + 0.007010502740740776, + 0.0064431349746882915, + 0.032518498599529266, + 0.012742739170789719, + 0.011724387295544147, + 0.011853232979774475, + 0.0165899358689785, + 0.0356539748609066, + 0.008151184767484665, + 0.010230008512735367, + 0.04566606506705284, + 0.07016027718782425, + 0.05910775437951088, + 0.01981649547815323, + 0.040220316499471664, + 0.036581575870513916, + 0.01857021637260914, + 0.019144317135214806, + 0.01029179897159338, + 0.003767970483750105, + 0.0028677047230303288, + 0.0017874805489555001, + 0.0029412219300866127, + 0.004502648487687111, + 0.0016214197967201471, + 0.00878632441163063, + 0.014447720721364021, + 0.01161572989076376, + 0.015960322692990303, + 0.012227545492351055, + 0.015430178493261337, + 0.008170661516487598, + 0.007671261206269264, + 0.00873252097517252, + 0.00623867753893137, + 0.005950291641056538, + 0.009576882235705853, + 0.004846795927733183, + 0.0067941839806735516, + 0.004241231828927994, + 0.0030676410533487797, + 0.004601770546287298, + 0.011976453475654125, + 0.012577971443533897, + 0.01914627104997635, + 0.009859095327556133, + 0.007410387974232435, + 0.005159314256161451, + 0.005850754678249359, + 0.0053713638335466385, + 0.008074943907558918, + 0.004116648808121681, + 0.006840227171778679, + 0.002774842781946063, + 0.002421978861093521, + 0.004876258783042431, + 0.0015349648892879486, + 0.0017082265112549067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.3870270252227783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021969962399452925 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021969962399452925, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021969962399452925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14918099343776703, + 0.059653814882040024, + 0.3720346689224243, + 0.07990734279155731, + 0.051543790847063065, + 0.12384109199047089, + 0.07032018899917603, + 0.020674534142017365, + 0.037199512124061584, + 0.0356440544128418 + ], + "credit_map": [ + 0.021458016708493233, + 0.020550210028886795, + 0.019468797370791435, + 0.01450409460812807, + 0.026040710508823395, + 0.01998463086783886, + 0.012784161604940891, + 0.014390365220606327, + 0.00735435402020812, + 0.006862388923764229, + 0.006209866143763065, + 0.004474273417145014, + 0.006687267683446407, + 0.008886661380529404, + 0.005360268987715244, + 0.01381873432546854, + 0.0694257840514183, + 0.024513712152838707, + 0.11929421126842499, + 0.03044857829809189, + 0.020878512412309647, + 0.05394520238041878, + 0.020972147583961487, + 0.03255651518702507, + 0.012389853596687317, + 0.006887699011713266, + 0.01491731870919466, + 0.006155497394502163, + 0.007550480775535107, + 0.010434514842927456, + 0.005231350660324097, + 0.016340630128979683, + 0.00745554780587554, + 0.009716130793094635, + 0.007537554483860731, + 0.004000843968242407, + 0.007435562089085579, + 0.00624347198754549, + 0.003148186719045043, + 0.006006494164466858, + 0.03807426616549492, + 0.011240802705287933, + 0.006332113407552242, + 0.003933724947273731, + 0.009840072132647038, + 0.012285545468330383, + 0.004797749686986208, + 0.0373368114233017, + 0.009348387829959393, + 0.01045207493007183, + 0.009787274524569511, + 0.007654385641217232, + 0.010907412506639957, + 0.008706162683665752, + 0.006051366683095694, + 0.007413126528263092, + 0.002754210028797388, + 0.0026445831172168255, + 0.0034398448187857866, + 0.0024870082270354033, + 0.0033956514671444893, + 0.00263034226372838, + 0.001230923691764474, + 0.0020919714588671923, + 0.006055227015167475, + 0.0037810427602380514, + 0.007507376838475466, + 0.0033482590224593878, + 0.00534471170976758, + 0.004821864888072014, + 0.002877109916880727, + 0.0034639225341379642, + 0.009932802990078926, + 0.004692671354860067, + 0.0045752935111522675, + 0.0022635136265307665, + 0.003120526671409607, + 0.006961587350815535, + 0.0019610272720456123, + 0.0021366302389651537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0021969962399452925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015093612018972635 + ], + "opposite_logprob_tokens": [ + -6.501509189605713 + ], + "logprob_reward": -0.0015093612018972635, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015093612018972635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028324708342552185, + 0.02485026605427265, + 0.6239871382713318, + 0.056548357009887695, + 0.07252383977174759, + 0.07707858830690384, + 0.015949198976159096, + 0.012902343645691872, + 0.030066533014178276, + 0.057769063860177994 + ], + "credit_map": [ + 0.004575598519295454, + 0.0039417133666574955, + 0.004342763684689999, + 0.0031048620585352182, + 0.0030532244127243757, + 0.003974460065364838, + 0.0018523781327530742, + 0.003479706821963191, + 0.0031588433776050806, + 0.0019425094360485673, + 0.004632053896784782, + 0.004080664366483688, + 0.0019344573374837637, + 0.0033901913557201624, + 0.0024817741941660643, + 0.0032297708094120026, + 0.11365098506212234, + 0.0406237468123436, + 0.19563500583171844, + 0.05053705349564552, + 0.03549416735768318, + 0.09245879203081131, + 0.03731425106525421, + 0.05827310308814049, + 0.012463759630918503, + 0.008090505376458168, + 0.0037856290582567453, + 0.0028683915734291077, + 0.008273101411759853, + 0.014342769980430603, + 0.003216808894649148, + 0.003507389919832349, + 0.015071890316903591, + 0.007412988692522049, + 0.021690480411052704, + 0.004398878663778305, + 0.004340886604040861, + 0.00939471460878849, + 0.003746055532246828, + 0.006467940751463175, + 0.024696342647075653, + 0.006197869312018156, + 0.0038798884488642216, + 0.0018743379041552544, + 0.005956354551017284, + 0.00770745612680912, + 0.00295708398334682, + 0.023809241130948067, + 0.002488019410520792, + 0.0022878237068653107, + 0.0033134296536445618, + 0.0016529714921489358, + 0.001991144847124815, + 0.001542000682093203, + 0.0011521498672664165, + 0.0015216568717733026, + 0.0016475035808980465, + 0.0018679203931242228, + 0.0012602848000824451, + 0.0012835738016292453, + 0.002572853583842516, + 0.0020322196651250124, + 0.0008883979171514511, + 0.0013495893217623234, + 0.004250552970916033, + 0.0037663616240024567, + 0.003971891012042761, + 0.0033360873349010944, + 0.0030166928190737963, + 0.0057675340212881565, + 0.0023792057763785124, + 0.0035782044287770987, + 0.01669500023126602, + 0.005404322873800993, + 0.012851609848439693, + 0.0032489646691828966, + 0.0024430337361991405, + 0.012769070453941822, + 0.001886405749246478, + 0.0024706521071493626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.0015093612018972635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017113338690251112 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017113338690251112, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017113338690251112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04587920382618904, + 0.019030466675758362, + 0.2454596608877182, + 0.045066237449645996, + 0.4919387400150299, + 0.07360489666461945, + 0.02370494231581688, + 0.008680115453898907, + 0.015211429446935654, + 0.031424254179000854 + ], + "credit_map": [ + 0.006384590175002813, + 0.005043157376348972, + 0.008885529823601246, + 0.005049238912761211, + 0.006161438301205635, + 0.0060388678684830666, + 0.003385628806427121, + 0.0049307565204799175, + 0.002645527245476842, + 0.0019279097905382514, + 0.0034342510625720024, + 0.002356935292482376, + 0.002740553580224514, + 0.0027806954458355904, + 0.0013088902924209833, + 0.0018357032677158713, + 0.04327087849378586, + 0.01724536344408989, + 0.07586619257926941, + 0.020047705620527267, + 0.014571106992661953, + 0.03766120597720146, + 0.014258785173296928, + 0.022538427263498306, + 0.01010183710604906, + 0.0061670332215726376, + 0.002981755416840315, + 0.00235198182053864, + 0.0063218967989087105, + 0.01158927846699953, + 0.002694097813218832, + 0.0028583607636392117, + 0.07159779965877533, + 0.11640609055757523, + 0.08903805911540985, + 0.03328399732708931, + 0.06321515142917633, + 0.0596885122358799, + 0.028617188334465027, + 0.03009195253252983, + 0.022732403129339218, + 0.006177686154842377, + 0.0045881494879722595, + 0.0020767985843122005, + 0.006348935421556234, + 0.007193647790700197, + 0.0029271352104842663, + 0.02156013995409012, + 0.0035260876175016165, + 0.004162159748375416, + 0.004999599419534206, + 0.0019072792492806911, + 0.0034114120062440634, + 0.0022849494125694036, + 0.0016836096765473485, + 0.0017298446036875248, + 0.0011798023479059339, + 0.0011215942213311791, + 0.0010760078439489007, + 0.0008030511671677232, + 0.0017116556409746408, + 0.0012577930465340614, + 0.0006183524383231997, + 0.000911858631297946, + 0.002804454416036606, + 0.0019304647576063871, + 0.0014818515628576279, + 0.0012914090184494853, + 0.0021104621700942516, + 0.003455362282693386, + 0.0010932876029983163, + 0.0010441378690302372, + 0.010325551964342594, + 0.002675060648471117, + 0.0036769555881619453, + 0.0013843385968357325, + 0.0017743089701980352, + 0.009009062312543392, + 0.0012195395538583398, + 0.0013594385236501694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.0017113338690251112 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set in a bright and cheerful 3D cartoon-style world, a dynamic low-angle tracking shot follows a small wooden rowboat as it plows through the waves. Poised at the prow is a proud black cat, sleek and black with a sleek digitally rendered texture. A bright red triangular scarf flutters in the breeze, and large, luminous emerald green eyes gaze resolutely into the distance. Constructed of warm brown wood planks with distinct grain, the boat sails across a shimmering azure sea, its bow constantly cleaving through crystal-clear white waves, sending stylized cartoon water droplets splattering across the clear sky, while a few fluffy, cotton-candy-like clouds punctuate the clear sky. The cat holds its head high, effortlessly balancing as the boat rhythmically rises and falls, its tail held high in a playful curl, radiating the confidence and authority of a captain. The brilliant sunlight pours down, infusing the entire scene with extremely high color saturation, creating soft highlights on the cat's smooth back and the wet side of the boat, and creating a joyful atmosphere full of vitality and optimistic adventurous spirit.", + "seed": 20496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_15/index_9977_seed_20496_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1987110327463597, + "vlm_logprob": -0.1987110327463597, + "vlm_yes_no_margin": 4.675000005960465, + "vlm_token_credit": -0.1987110327463597, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005568942986428738 + ], + "opposite_logprob_tokens": [ + -7.005568981170654 + ], + "logprob_reward": -0.005568942986428738, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.005568942986428738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2256031483411789, + 0.06567920744419098, + 0.04519648477435112, + 0.03170805796980858, + 0.042253825813531876, + 0.061152536422014236, + 0.1488078385591507, + 0.08170820772647858, + 0.1281958371400833, + 0.16969485580921173 + ], + "credit_map": [ + 0.0371725857257843, + 0.01563449762761593, + 0.017856424674391747, + 0.04436719790101051, + 0.012607774697244167, + 0.0728071928024292, + 0.013414773158729076, + 0.011742698960006237, + 0.015353739261627197, + 0.005226003471761942, + 0.00578246358782053, + 0.00580142205581069, + 0.0037500285543501377, + 0.011464318260550499, + 0.007007606327533722, + 0.011293631047010422, + 0.00805551465600729, + 0.005970080383121967, + 0.0031781417783349752, + 0.0071246731095016, + 0.0036781886592507362, + 0.007436005864292383, + 0.005202183499932289, + 0.004551699850708246, + 0.006462572608143091, + 0.003049697959795594, + 0.0022503454238176346, + 0.0032128384336829185, + 0.0026248793583363295, + 0.005712549667805433, + 0.0036853619385510683, + 0.004709810018539429, + 0.0064489278011024, + 0.004213668871670961, + 0.0043324995785951614, + 0.0049512130208313465, + 0.005846734624356031, + 0.00480049941688776, + 0.005612475331872702, + 0.006047810893505812, + 0.0066451202146708965, + 0.005023483652621508, + 0.01205372903496027, + 0.0075997342355549335, + 0.004358287900686264, + 0.008472708985209465, + 0.008627025410532951, + 0.008372447453439236, + 0.007295938208699226, + 0.012639477849006653, + 0.017014682292938232, + 0.023629939183592796, + 0.013336645439267159, + 0.008198576979339123, + 0.05023457109928131, + 0.01645803079009056, + 0.006257359404116869, + 0.00984737928956747, + 0.017027312889695168, + 0.010410738177597523, + 0.008642679080367088, + 0.005559988785535097, + 0.013125302270054817, + 0.010837449692189693, + 0.009143631905317307, + 0.011494378559291363, + 0.010802471078932285, + 0.01336878165602684, + 0.0074600898660719395, + 0.00898228120058775, + 0.04045391455292702, + 0.026490285992622375, + 0.015390989370644093, + 0.037638649344444275, + 0.024108488112688065, + 0.02676873281598091, + 0.013534114696085453, + 0.012331458739936352, + 0.013365843333303928, + 0.026556581258773804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.005568942986428738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742274284362793 + ], + "opposite_logprob_tokens": [ + -0.4742274582386017 + ], + "logprob_reward": -0.9742274284362793, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9742274284362793, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.32671377062797546, + 0.038580719381570816, + 0.028627602383494377, + 0.019721239805221558, + 0.04863646626472473, + 0.05022212862968445, + 0.08092464506626129, + 0.300735741853714, + 0.044466327875852585, + 0.06137137487530708 + ], + "credit_map": [ + 0.05185072869062424, + 0.014978047460317612, + 0.014425469562411308, + 0.08738143742084503, + 0.00783885084092617, + 0.12118253856897354, + 0.01595398038625717, + 0.01310269720852375, + 0.008183151483535767, + 0.004410356283187866, + 0.002454340923577547, + 0.004606207367032766, + 0.003229736816138029, + 0.00661392230540514, + 0.004147161263972521, + 0.0049358392134308815, + 0.0049341218546032906, + 0.003892045933753252, + 0.002330536488443613, + 0.0032581211999058723, + 0.0026833193842321634, + 0.004787989892065525, + 0.0038632419891655445, + 0.0028782214503735304, + 0.0050674802623689175, + 0.002208688063547015, + 0.0014786324463784695, + 0.0019729381892830133, + 0.0021639622282236814, + 0.0031461683101952076, + 0.0018551010871306062, + 0.0018282674718648195, + 0.006324379704892635, + 0.008205550722777843, + 0.003921224270015955, + 0.007523227017372847, + 0.0058280290104448795, + 0.006406608037650585, + 0.006359410937875509, + 0.004068027716130018, + 0.003987076226621866, + 0.007527119014412165, + 0.006726743653416634, + 0.007577575743198395, + 0.00500934012234211, + 0.004476914647966623, + 0.00960380770266056, + 0.00531354732811451, + 0.0064839632250368595, + 0.013590343296527863, + 0.010051492601633072, + 0.010129974223673344, + 0.012308784760534763, + 0.006843033712357283, + 0.01145106554031372, + 0.010065978392958641, + 0.010812525637447834, + 0.013170836493372917, + 0.03212647885084152, + 0.12171536684036255, + 0.010435299947857857, + 0.010598304681479931, + 0.021252045407891273, + 0.08062487095594406, + 0.002692324807867408, + 0.004435025621205568, + 0.004244981333613396, + 0.004213445354253054, + 0.0034781363792717457, + 0.004486120771616697, + 0.011271809227764606, + 0.009644481353461742, + 0.004340529441833496, + 0.008781149052083492, + 0.009384379722177982, + 0.009085000492632389, + 0.005506352987140417, + 0.0067399065010249615, + 0.007805550005286932, + 0.009728499688208103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.9742274284362793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036157488357275724 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.0036157488357275724, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036157488357275724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6992080211639404, + 0.03858286514878273, + 0.04732264578342438, + 0.015576578676700592, + 0.022274240851402283, + 0.01669851876795292, + 0.047312360256910324, + 0.019911736249923706, + 0.05849805474281311, + 0.03461507707834244 + ], + "credit_map": [ + 0.10673333704471588, + 0.02628321945667267, + 0.03251084312796593, + 0.20076365768909454, + 0.012157384306192398, + 0.2638764977455139, + 0.029782071709632874, + 0.027101021260023117, + 0.006994054652750492, + 0.0038375449366867542, + 0.003993211779743433, + 0.00411552470177412, + 0.0032510776072740555, + 0.005351250525563955, + 0.004885127302259207, + 0.006155075505375862, + 0.006736984476447105, + 0.005569730419665575, + 0.00451310770586133, + 0.007434282451868057, + 0.0036059683188796043, + 0.0069865575060248375, + 0.006727252621203661, + 0.0057487646117806435, + 0.003918762784451246, + 0.0019249195465818048, + 0.001063089119270444, + 0.00122500688303262, + 0.0010223696008324623, + 0.0027953318785876036, + 0.0012984809000045061, + 0.002328617963939905, + 0.003783225081861019, + 0.002944381209090352, + 0.002286568284034729, + 0.0029067243449389935, + 0.0017569416668266058, + 0.002827166812494397, + 0.002933668904006481, + 0.0028355643153190613, + 0.001816154457628727, + 0.0030385341960936785, + 0.002291342243552208, + 0.002223602030426264, + 0.0021105383057147264, + 0.0013540128711611032, + 0.002159493276849389, + 0.0017048417357727885, + 0.002497149398550391, + 0.005285594146698713, + 0.005655789282172918, + 0.007942418567836285, + 0.005219167098402977, + 0.0030656405724585056, + 0.0116832684725523, + 0.005963333882391453, + 0.0014927118318155408, + 0.0026407241821289062, + 0.003658501897007227, + 0.003596274182200432, + 0.001323615200817585, + 0.0016150599112734199, + 0.0031527874525636435, + 0.0024320606607943773, + 0.0032894727773964405, + 0.0031219765078276396, + 0.004045721143484116, + 0.007203573826700449, + 0.002589298179373145, + 0.0037385502364486456, + 0.02085297554731369, + 0.013656487688422203, + 0.003362905001267791, + 0.007092183921486139, + 0.004124883562326431, + 0.0051498194225132465, + 0.003360659582540393, + 0.0036271200515329838, + 0.00304213585332036, + 0.004855368286371231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0036157488357275724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015113847330212593 + ], + "opposite_logprob_tokens": [ + -6.501511573791504 + ], + "logprob_reward": -0.0015113847330212593, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015113847330212593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4993753135204315, + 0.03341517969965935, + 0.0642397329211235, + 0.01638699136674404, + 0.042009126394987106, + 0.018289916217327118, + 0.038343098014593124, + 0.1766594499349594, + 0.07357163727283478, + 0.03770953789353371 + ], + "credit_map": [ + 0.07483252137899399, + 0.02005929686129093, + 0.021899696439504623, + 0.14452926814556122, + 0.009088536724448204, + 0.18614733219146729, + 0.02184842713177204, + 0.020970221608877182, + 0.004344628192484379, + 0.0028036555740982294, + 0.003159993328154087, + 0.0037670149467885494, + 0.0023089745081961155, + 0.004481656476855278, + 0.006635854486376047, + 0.005913402419537306, + 0.009926484897732735, + 0.006689544301480055, + 0.006518550682812929, + 0.006974355783313513, + 0.0046569956466555595, + 0.009853538125753403, + 0.01180062722414732, + 0.007819635793566704, + 0.0028414863627403975, + 0.0015070090303197503, + 0.0013480563648045063, + 0.0020055121276527643, + 0.0013869424583390355, + 0.0022796313278377056, + 0.0025829002261161804, + 0.0024354534689337015, + 0.0060514723882079124, + 0.0035081536043435335, + 0.004558443557471037, + 0.007501681335270405, + 0.0023589374031871557, + 0.006008786614984274, + 0.006588030606508255, + 0.0054336246103048325, + 0.0018465658649802208, + 0.0036470124032348394, + 0.002414085902273655, + 0.0019889655523002148, + 0.0024856289383023977, + 0.002051668707281351, + 0.002046535722911358, + 0.001809453358873725, + 0.0026929883752018213, + 0.0053943367674946785, + 0.004461622331291437, + 0.005193221382796764, + 0.005093976389616728, + 0.003668146440759301, + 0.006343410816043615, + 0.005495394580066204, + 0.0059350901283323765, + 0.010679254308342934, + 0.018079247325658798, + 0.06539510190486908, + 0.008245312608778477, + 0.008617721498012543, + 0.017965249717235565, + 0.041742462664842606, + 0.0035966646391898394, + 0.003752390155568719, + 0.005036938935518265, + 0.008303175680339336, + 0.0032613289076834917, + 0.003993087913841009, + 0.027229946106672287, + 0.0183981042355299, + 0.0028567274566739798, + 0.003937904257327318, + 0.004599171224981546, + 0.005824296269565821, + 0.0035323358606547117, + 0.005314861889928579, + 0.004216219298541546, + 0.007428021635860205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0015113847330212593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631658740341663 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.008631658740341663, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631658740341663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7866102457046509, + 0.022469429299235344, + 0.018839526921510696, + 0.011024856008589268, + 0.019365813583135605, + 0.011403582990169525, + 0.020017769187688828, + 0.03189939633011818, + 0.04628793150186539, + 0.03208151459693909 + ], + "credit_map": [ + 0.11910376697778702, + 0.029942449182271957, + 0.03603357821702957, + 0.22781012952327728, + 0.013484875671565533, + 0.29606330394744873, + 0.0324891060590744, + 0.03168294206261635, + 0.005023397970944643, + 0.0024005237501114607, + 0.0018198165344074368, + 0.002314291661605239, + 0.0017642190214246511, + 0.0032475434709340334, + 0.0022581538651138544, + 0.0036414782516658306, + 0.003504181280732155, + 0.001580091891810298, + 0.0019916011951863766, + 0.0016603831900283694, + 0.0012219877680763602, + 0.003878997405990958, + 0.002560477703809738, + 0.0024418027605861425, + 0.0019031264819204807, + 0.001010631094686687, + 0.0008974314550869167, + 0.001409354037605226, + 0.0008820333750918508, + 0.001654524588957429, + 0.0015857818070799112, + 0.0016819714801385999, + 0.0033749628346413374, + 0.0021846990566700697, + 0.0021370183676481247, + 0.002608886919915676, + 0.0014104795409366488, + 0.0024010061752051115, + 0.002636815421283245, + 0.002611942123621702, + 0.001452994649298489, + 0.0014773219591006637, + 0.002150297863408923, + 0.0011140239657834172, + 0.001037881476804614, + 0.0014248231891542673, + 0.001210387097671628, + 0.001535851159133017, + 0.001308181555941701, + 0.0024113566614687443, + 0.002287555020302534, + 0.0031237457878887653, + 0.002274400321766734, + 0.0015308941947296262, + 0.005140294786542654, + 0.0019413372501730919, + 0.0014047871809452772, + 0.002309258095920086, + 0.0050583514384925365, + 0.008791034109890461, + 0.00209820200689137, + 0.0016947713447734714, + 0.0050229402258992195, + 0.00552004761993885, + 0.002546833362430334, + 0.0028634455520659685, + 0.003266765736043453, + 0.005521044135093689, + 0.0025163134559988976, + 0.00289011187851429, + 0.015838081017136574, + 0.01084532868117094, + 0.003082239767536521, + 0.0037113765720278025, + 0.003639905946329236, + 0.005210408940911293, + 0.0027216661255806684, + 0.003935025539249182, + 0.0033308821730315685, + 0.006450008600950241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.008631658740341663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set in a bright and cheerful 3D cartoon-style world, a dynamic low-angle tracking shot follows a small wooden rowboat as it plows through the waves. Poised at the prow is a proud black cat, sleek and black with a sleek digitally rendered texture. A bright red triangular scarf flutters in the breeze, and large, luminous emerald green eyes gaze resolutely into the distance. Constructed of warm brown wood planks with distinct grain, the boat sails across a shimmering azure sea, its bow constantly cleaving through crystal-clear white waves, sending stylized cartoon water droplets splattering across the clear sky, while a few fluffy, cotton-candy-like clouds punctuate the clear sky. The cat holds its head high, effortlessly balancing as the boat rhythmically rises and falls, its tail held high in a playful curl, radiating the confidence and authority of a captain. The brilliant sunlight pours down, infusing the entire scene with extremely high color saturation, creating soft highlights on the cat's smooth back and the wet side of the boat, and creating a joyful atmosphere full of vitality and optimistic adventurous spirit.", + "seed": 20497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_15/index_9977_seed_20497_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.038084139913553375, + "vlm_logprob": -0.038084139913553375, + "vlm_yes_no_margin": 5.725, + "vlm_token_credit": -0.038084139913553375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004120788536965847 + ], + "opposite_logprob_tokens": [ + -7.629120826721191 + ], + "logprob_reward": -0.004120788536965847, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.004120788536965847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15063735842704773, + 0.14024609327316284, + 0.07477828115224838, + 0.048049259930849075, + 0.1639271378517151, + 0.10340029746294022, + 0.10781404376029968, + 0.06160455569624901, + 0.06681233644485474, + 0.08273059874773026 + ], + "credit_map": [ + 0.026696261018514633, + 0.016313163563609123, + 0.02421291172504425, + 0.019302697852253914, + 0.012482673861086369, + 0.02105776220560074, + 0.013810387812554836, + 0.016761505976319313, + 0.021327009424567223, + 0.014036871492862701, + 0.015814904123544693, + 0.019794408231973648, + 0.018051156774163246, + 0.01379432249814272, + 0.01684160903096199, + 0.020585816353559494, + 0.011541814543306828, + 0.011601671576499939, + 0.012451911345124245, + 0.007444798015058041, + 0.01144223939627409, + 0.006535280961543322, + 0.006450767628848553, + 0.007309804670512676, + 0.006392337381839752, + 0.006878068204969168, + 0.005263194907456636, + 0.0052798595279455185, + 0.0063468776643276215, + 0.006443249061703682, + 0.004517653025686741, + 0.00692802295088768, + 0.013921923004090786, + 0.030696118250489235, + 0.012126847170293331, + 0.010075158439576626, + 0.044365864247083664, + 0.02457333914935589, + 0.01468663290143013, + 0.013481281697750092, + 0.014179689809679985, + 0.016684580594301224, + 0.01304006390273571, + 0.012901931069791317, + 0.016815178096294403, + 0.0074805282056331635, + 0.012867562472820282, + 0.009430773556232452, + 0.009608457796275616, + 0.01665216125547886, + 0.012590447440743446, + 0.011539649218320847, + 0.016873952001333237, + 0.014394399709999561, + 0.014271100051701069, + 0.011883877217769623, + 0.008766885846853256, + 0.00949008110910654, + 0.009696941822767258, + 0.008273151703178883, + 0.005479698069393635, + 0.0053366911597549915, + 0.007430283352732658, + 0.007130820769816637, + 0.009845243766903877, + 0.009617852978408337, + 0.013347919099032879, + 0.007272243499755859, + 0.006537438370287418, + 0.0048119318671524525, + 0.009114251472055912, + 0.006265456322580576, + 0.009840739890933037, + 0.009680707938969135, + 0.01457963790744543, + 0.0108461519703269, + 0.013405843637883663, + 0.008368780836462975, + 0.007447025738656521, + 0.008561717346310616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.004120788536965847 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985698580741882 + ], + "opposite_logprob_tokens": [ + -1.8048570156097412 + ], + "logprob_reward": -0.17985698580741882, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985698580741882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09226910769939423, + 0.09085626155138016, + 0.14079135656356812, + 0.05188354477286339, + 0.2366357147693634, + 0.10798760503530502, + 0.10064155608415604, + 0.1277453750371933, + 0.021222412586212158, + 0.02996707707643509 + ], + "credit_map": [ + 0.014628546312451363, + 0.008817492984235287, + 0.01097497995942831, + 0.014556230045855045, + 0.008376235142350197, + 0.018444746732711792, + 0.00899624451994896, + 0.007474619895219803, + 0.01497074868530035, + 0.009048441424965858, + 0.012280350551009178, + 0.009186680428683758, + 0.01692195236682892, + 0.012130563147366047, + 0.008177467621862888, + 0.008140048943459988, + 0.0236881822347641, + 0.020929500460624695, + 0.024075323715806007, + 0.016859661787748337, + 0.012355937622487545, + 0.017741024494171143, + 0.012104197405278683, + 0.01303750928491354, + 0.006604151334613562, + 0.006149597465991974, + 0.007658588234335184, + 0.005279978737235069, + 0.008239383809268475, + 0.005980262067168951, + 0.0071034482680261135, + 0.004868129268288612, + 0.030932508409023285, + 0.04645150899887085, + 0.0217833761125803, + 0.020725533366203308, + 0.03451790288090706, + 0.02491421438753605, + 0.03277026489377022, + 0.024540385231375694, + 0.010463903658092022, + 0.015728872269392014, + 0.01362990029156208, + 0.011747817508876324, + 0.01374727301299572, + 0.009904804639518261, + 0.02022443898022175, + 0.012540583498775959, + 0.010871975682675838, + 0.014196248725056648, + 0.01292776595801115, + 0.011357731185853481, + 0.016445500776171684, + 0.012530882842838764, + 0.013483760878443718, + 0.00882768165320158, + 0.018221449106931686, + 0.012795775197446346, + 0.015619296580553055, + 0.01858975738286972, + 0.009664450772106647, + 0.011896707117557526, + 0.023300902917981148, + 0.017657024785876274, + 0.0045969183556735516, + 0.0020447485148906708, + 0.004076001234352589, + 0.0023701211903244257, + 0.0014053686754778028, + 0.001546772662550211, + 0.002433170098811388, + 0.002749309642240405, + 0.003927858546376228, + 0.00406777486205101, + 0.0031748004257678986, + 0.0034858668223023415, + 0.004981922917068005, + 0.004115292802453041, + 0.0027736618649214506, + 0.0034398972056806087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.17985698580741882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818067092448473 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.002818067092448473, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818067092448473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2829192280769348, + 0.20331229269504547, + 0.1025221198797226, + 0.025396833196282387, + 0.06187213584780693, + 0.030551692470908165, + 0.03622790798544884, + 0.18390198051929474, + 0.030762817710638046, + 0.042532965540885925 + ], + "credit_map": [ + 0.04124230146408081, + 0.027074391022324562, + 0.028657499700784683, + 0.05488196387887001, + 0.02411300502717495, + 0.047402627766132355, + 0.03352291136980057, + 0.026024529710412025, + 0.03330589458346367, + 0.022128095850348473, + 0.0267415139824152, + 0.01962820626795292, + 0.03850475698709488, + 0.027046101167798042, + 0.016698993742465973, + 0.019258731976151466, + 0.01524127833545208, + 0.011789600364863873, + 0.015793638303875923, + 0.012089638039469719, + 0.016729149967432022, + 0.011981314048171043, + 0.008503878489136696, + 0.010393618606030941, + 0.003178076120093465, + 0.0037828770000487566, + 0.002913449890911579, + 0.0022068526595830917, + 0.005632012151181698, + 0.0034320014528930187, + 0.0020852333400398493, + 0.002166330348700285, + 0.007259372156113386, + 0.009393498301506042, + 0.005885659251362085, + 0.004388572182506323, + 0.011413694359362125, + 0.007989531382918358, + 0.009211676195263863, + 0.006330134347081184, + 0.0037538090255111456, + 0.006722563412040472, + 0.003309407038614154, + 0.0034050627145916224, + 0.005441031418740749, + 0.002461573574692011, + 0.0034751032944768667, + 0.0019831403624266386, + 0.0037322628777474165, + 0.005298868287354708, + 0.003912872169166803, + 0.004280345980077982, + 0.0061954716220498085, + 0.004576514009386301, + 0.005103376694023609, + 0.0031281958799809217, + 0.025737570598721504, + 0.01858147606253624, + 0.022329533472657204, + 0.025119362398982048, + 0.016962893307209015, + 0.020305555313825607, + 0.03162459656596184, + 0.023240990936756134, + 0.005293603520840406, + 0.004953268449753523, + 0.004815602209419012, + 0.0034748578909784555, + 0.004548687022179365, + 0.0026434529572725296, + 0.0032593272626399994, + 0.001774017931893468, + 0.005552916321903467, + 0.0059392936527729034, + 0.005453366786241531, + 0.004542860668152571, + 0.009424082934856415, + 0.0033288998529314995, + 0.004318857099860907, + 0.003972687758505344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.002818067092448473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004363300104159862 + ], + "opposite_logprob_tokens": [ + -7.750436305999756 + ], + "logprob_reward": -0.0004363300104159862, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004363300104159862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22749066352844238, + 0.12147875130176544, + 0.12535297870635986, + 0.02635817602276802, + 0.2531859278678894, + 0.12160492688417435, + 0.04394068941473961, + 0.0281063262373209, + 0.029101435095071793, + 0.023380085825920105 + ], + "credit_map": [ + 0.0446748211979866, + 0.022508256137371063, + 0.025725120678544044, + 0.02160150744020939, + 0.02253103069961071, + 0.05498175323009491, + 0.014346606098115444, + 0.02112160623073578, + 0.019331714138388634, + 0.014137348160147667, + 0.020476998761296272, + 0.013294198550283909, + 0.018350401893258095, + 0.013432493433356285, + 0.01172714401036501, + 0.010728471912443638, + 0.018605854362249374, + 0.017010584473609924, + 0.01988607831299305, + 0.016050290316343307, + 0.014523079618811607, + 0.011410838924348354, + 0.013067039661109447, + 0.014799224212765694, + 0.003620072500780225, + 0.004061008337885141, + 0.0032407238613814116, + 0.0024713012389838696, + 0.0047662341967225075, + 0.0032695895060896873, + 0.002836582250893116, + 0.0020926666911691427, + 0.018368341028690338, + 0.04250712692737579, + 0.019909612834453583, + 0.0168985016644001, + 0.06195256859064102, + 0.04239166900515556, + 0.0323517732322216, + 0.01880633272230625, + 0.01130568515509367, + 0.03231019526720047, + 0.00835274439305067, + 0.01006690040230751, + 0.02450980432331562, + 0.010180498473346233, + 0.014688421040773392, + 0.010190689004957676, + 0.004490599501878023, + 0.005761085543781519, + 0.004592591430991888, + 0.004458436742424965, + 0.007374891545623541, + 0.006264748517423868, + 0.006909111514687538, + 0.004089226946234703, + 0.0035815895535051823, + 0.0035846910905092955, + 0.003206237917765975, + 0.0043725487776100636, + 0.0025422426406294107, + 0.002522416412830353, + 0.004498991649597883, + 0.0037976119201630354, + 0.007045516278594732, + 0.003114572959020734, + 0.006735470145940781, + 0.0023841452784836292, + 0.0025949592236429453, + 0.0024895223323255777, + 0.002318888669833541, + 0.00241836323402822, + 0.0034097887109965086, + 0.0023178542032837868, + 0.00421651266515255, + 0.003449386917054653, + 0.0030517473351210356, + 0.002109325025230646, + 0.0023212151136249304, + 0.002504258882254362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0004363300104159862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031885281205177307 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.0031885281205177307, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031885281205177307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23699051141738892, + 0.5099614262580872, + 0.05040052533149719, + 0.011141113005578518, + 0.056705620139837265, + 0.03655751422047615, + 0.030931925401091576, + 0.02812841534614563, + 0.01523423008620739, + 0.023948676884174347 + ], + "credit_map": [ + 0.04267454892396927, + 0.02422764152288437, + 0.027558153495192528, + 0.03639082983136177, + 0.014997106045484543, + 0.04803283512592316, + 0.01851624809205532, + 0.024593155831098557, + 0.08300277590751648, + 0.04990187659859657, + 0.07307688146829605, + 0.05996968224644661, + 0.07723315060138702, + 0.056510355323553085, + 0.055494099855422974, + 0.05477266386151314, + 0.007050869055092335, + 0.006657137535512447, + 0.009989879094064236, + 0.004284181632101536, + 0.0072770919650793076, + 0.006830436643213034, + 0.004298967309296131, + 0.004011964425444603, + 0.001573337591253221, + 0.0019406159408390522, + 0.001247712760232389, + 0.0012486372143030167, + 0.001292238011956215, + 0.0014160377904772758, + 0.0012783632846549153, + 0.0011441701790317893, + 0.0038437156472355127, + 0.010889524593949318, + 0.004700704477727413, + 0.003668113611638546, + 0.015899961814284325, + 0.008069776929914951, + 0.005287878215312958, + 0.0043459441512823105, + 0.0035883246455341578, + 0.007166578434407711, + 0.003863676218315959, + 0.003850120585411787, + 0.0057365368120372295, + 0.003134938422590494, + 0.005527070257812738, + 0.003690272569656372, + 0.0029039965011179447, + 0.004882542882114649, + 0.00486963614821434, + 0.003455026540905237, + 0.00408290745690465, + 0.003935693763196468, + 0.0033440827392041683, + 0.003458041697740555, + 0.0045976038090884686, + 0.004248844459652901, + 0.0029940353706479073, + 0.003986244555562735, + 0.0024801157414913177, + 0.002225482603535056, + 0.004227672703564167, + 0.0033684181980788708, + 0.002437925897538662, + 0.001385461655445397, + 0.00438616331666708, + 0.0016462273197248578, + 0.0012611339334398508, + 0.0010157786309719086, + 0.0016101901419460773, + 0.001491350820288062, + 0.0035316043067723513, + 0.0019055032171308994, + 0.004969044588506222, + 0.0038874594029039145, + 0.002240648726001382, + 0.002877289429306984, + 0.0020114975050091743, + 0.0025256325025111437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0031885281205177307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set in a bright and cheerful 3D cartoon-style world, a dynamic low-angle tracking shot follows a small wooden rowboat as it plows through the waves. Poised at the prow is a proud black cat, sleek and black with a sleek digitally rendered texture. A bright red triangular scarf flutters in the breeze, and large, luminous emerald green eyes gaze resolutely into the distance. Constructed of warm brown wood planks with distinct grain, the boat sails across a shimmering azure sea, its bow constantly cleaving through crystal-clear white waves, sending stylized cartoon water droplets splattering across the clear sky, while a few fluffy, cotton-candy-like clouds punctuate the clear sky. The cat holds its head high, effortlessly balancing as the boat rhythmically rises and falls, its tail held high in a playful curl, radiating the confidence and authority of a captain. The brilliant sunlight pours down, infusing the entire scene with extremely high color saturation, creating soft highlights on the cat's smooth back and the wet side of the boat, and creating a joyful atmosphere full of vitality and optimistic adventurous spirit.", + "seed": 20498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_15/index_9977_seed_20498_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010063104308210313, + "vlm_logprob": -0.010063104308210313, + "vlm_yes_no_margin": 5.75, + "vlm_token_credit": -0.010063104308210313, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0039220331236720085 + ], + "opposite_logprob_tokens": [ + -7.253921985626221 + ], + "logprob_reward": -0.0039220331236720085, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0039220331236720085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15811002254486084, + 0.06973098963499069, + 0.07063542306423187, + 0.09545725584030151, + 0.04115051031112671, + 0.1560794860124588, + 0.051357269287109375, + 0.20475979149341583, + 0.08456005156040192, + 0.06815925240516663 + ], + "credit_map": [ + 0.02147577702999115, + 0.015276908874511719, + 0.022390414029359818, + 0.02063557505607605, + 0.024347301572561264, + 0.015268731862306595, + 0.019275113940238953, + 0.0194401815533638, + 0.011213991791009903, + 0.008153746835887432, + 0.011836945079267025, + 0.008718423545360565, + 0.008507263846695423, + 0.008735557086765766, + 0.005723211448639631, + 0.006841838359832764, + 0.010527699254453182, + 0.007135140243917704, + 0.014303037896752357, + 0.010363406501710415, + 0.007156601175665855, + 0.007770403288304806, + 0.005545339547097683, + 0.007833781652152538, + 0.014020601287484169, + 0.009631145745515823, + 0.024997975677251816, + 0.008940136060118675, + 0.008572831749916077, + 0.0071067968383431435, + 0.011952990666031837, + 0.010234765708446503, + 0.008436544798314571, + 0.004429572727531195, + 0.005958371330052614, + 0.00440990412607789, + 0.004083557520061731, + 0.0035556091461330652, + 0.00458414014428854, + 0.0056928046979010105, + 0.017018577083945274, + 0.026158105581998825, + 0.01582733355462551, + 0.023210272192955017, + 0.025299308821558952, + 0.01728040538728237, + 0.01458785030990839, + 0.01669761724770069, + 0.0053380574099719524, + 0.009023448452353477, + 0.004415723029524088, + 0.00658499775454402, + 0.0074418894946575165, + 0.005961508955806494, + 0.004930868279188871, + 0.007660774048417807, + 0.0212466549128294, + 0.058884572237730026, + 0.0145749906077981, + 0.016611291095614433, + 0.01290859840810299, + 0.013435454107820988, + 0.013838821090757847, + 0.053259383887052536, + 0.009007402695715427, + 0.010420214384794235, + 0.019250862300395966, + 0.007429936900734901, + 0.009866325184702873, + 0.007023765239864588, + 0.007813461124897003, + 0.013748071156442165, + 0.011194631457328796, + 0.008632820099592209, + 0.013134706765413284, + 0.0072647444903850555, + 0.0092413779348135, + 0.005002226680517197, + 0.00823331717401743, + 0.005455420818179846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.0039220331236720085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809729963541031 + ], + "opposite_logprob_tokens": [ + -3.288097381591797 + ], + "logprob_reward": -0.03809729963541031, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809729963541031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18179400265216827, + 0.06565290689468384, + 0.09815003722906113, + 0.2810324430465698, + 0.07987719774246216, + 0.04954781383275986, + 0.031928662210702896, + 0.06453495472669601, + 0.1016571968793869, + 0.04582485929131508 + ], + "credit_map": [ + 0.025716617703437805, + 0.02007601596415043, + 0.02920636348426342, + 0.02467946521937847, + 0.02136395499110222, + 0.015692109242081642, + 0.01762101985514164, + 0.02743842452764511, + 0.009703931398689747, + 0.005933163687586784, + 0.011047734878957272, + 0.007883679121732712, + 0.010372493416070938, + 0.007940396666526794, + 0.005940802861005068, + 0.006830695550888777, + 0.017933517694473267, + 0.008597751148045063, + 0.0260841753333807, + 0.010819165036082268, + 0.006607414688915014, + 0.008553392253816128, + 0.007897417061030865, + 0.01165718026459217, + 0.05278448760509491, + 0.024997472763061523, + 0.07088032364845276, + 0.023677242919802666, + 0.019106963649392128, + 0.01948806643486023, + 0.023960519582033157, + 0.046137336641550064, + 0.014849286526441574, + 0.004557089414447546, + 0.026501376181840897, + 0.008642658591270447, + 0.004278417676687241, + 0.006689560133963823, + 0.007574259769171476, + 0.006784539669752121, + 0.006398489698767662, + 0.007542244624346495, + 0.003839865094050765, + 0.0072877490893006325, + 0.005940495524555445, + 0.004197389353066683, + 0.005003547761589289, + 0.009338025003671646, + 0.004351005889475346, + 0.005954998545348644, + 0.002740101423114538, + 0.004506110679358244, + 0.004427754320204258, + 0.003334719454869628, + 0.003476162441074848, + 0.003137806197628379, + 0.007626399863511324, + 0.01883620396256447, + 0.0043909624218940735, + 0.0050534033216536045, + 0.004798154346644878, + 0.004047170281410217, + 0.004660493694245815, + 0.015122154727578163, + 0.011388411745429039, + 0.0056237285025417805, + 0.03675530478358269, + 0.008160235360264778, + 0.006356953177601099, + 0.004859520122408867, + 0.008918561041355133, + 0.0195944681763649, + 0.0072174775414168835, + 0.005838997196406126, + 0.008029947988688946, + 0.004212557803839445, + 0.006183885503560305, + 0.003365017706528306, + 0.0053222524002194405, + 0.005654716864228249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.03809729963541031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017132379580289125 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017132379580289125, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017132379580289125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1848953664302826, + 0.14318038523197174, + 0.08137058466672897, + 0.18132543563842773, + 0.05297404155135155, + 0.04507169872522354, + 0.0464940182864666, + 0.08860260248184204, + 0.11349736899137497, + 0.06258854269981384 + ], + "credit_map": [ + 0.029525304213166237, + 0.01778104528784752, + 0.025331107899546623, + 0.02839810401201248, + 0.028916848823428154, + 0.015904659405350685, + 0.014579923823475838, + 0.02445836178958416, + 0.021875744685530663, + 0.013109697960317135, + 0.024859260767698288, + 0.015968916937708855, + 0.024048641324043274, + 0.017622575163841248, + 0.01090170256793499, + 0.014793838374316692, + 0.014122920110821724, + 0.009593595750629902, + 0.015597670339047909, + 0.009319682605564594, + 0.00924196932464838, + 0.008318652398884296, + 0.008006945252418518, + 0.007169141434133053, + 0.029271449893712997, + 0.020495928823947906, + 0.037199776619672775, + 0.014726623892784119, + 0.01520078256726265, + 0.01879877969622612, + 0.0192171148955822, + 0.026414968073368073, + 0.007848001085221767, + 0.0075776707381010056, + 0.008171950466930866, + 0.005522843915969133, + 0.007490959018468857, + 0.0047430372796952724, + 0.004421374294906855, + 0.0071981982327997684, + 0.005035880487412214, + 0.005534638650715351, + 0.00512590492144227, + 0.003634081454947591, + 0.00599987618625164, + 0.005088147707283497, + 0.0058470056392252445, + 0.008806159719824791, + 0.004310158547013998, + 0.006669480353593826, + 0.0035725226625800133, + 0.004997951444238424, + 0.008047185838222504, + 0.006214481312781572, + 0.005753589794039726, + 0.006928642280399799, + 0.00867227092385292, + 0.01779942959547043, + 0.007544602267444134, + 0.007673752959817648, + 0.008702769875526428, + 0.006438183598220348, + 0.008147922344505787, + 0.023623665794730186, + 0.016867009922862053, + 0.012654531747102737, + 0.01932886429131031, + 0.009972171857953072, + 0.02144455537199974, + 0.009233998134732246, + 0.012253656052052975, + 0.01174256857484579, + 0.00902633648365736, + 0.005957739893347025, + 0.01596510224044323, + 0.005565652158111334, + 0.006743557285517454, + 0.00433304812759161, + 0.006957654375582933, + 0.008039447478950024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0017132379580289125 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001334729720838368 + ], + "opposite_logprob_tokens": [ + -6.6263346672058105 + ], + "logprob_reward": -0.001334729720838368, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001334729720838368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17286980152130127, + 0.12138725817203522, + 0.10224555432796478, + 0.2968898117542267, + 0.048412028700113297, + 0.03993517905473709, + 0.03792298585176468, + 0.05826370790600777, + 0.09303135424852371, + 0.0290423221886158 + ], + "credit_map": [ + 0.033090513199567795, + 0.012974110431969166, + 0.028720827773213387, + 0.027936536818742752, + 0.017131004482507706, + 0.015326123684644699, + 0.012839707545936108, + 0.02485097572207451, + 0.021391496062278748, + 0.012364460155367851, + 0.02483323961496353, + 0.021847547963261604, + 0.013641824945807457, + 0.009254452772438526, + 0.00958425085991621, + 0.008469986729323864, + 0.016161680221557617, + 0.009166417643427849, + 0.02219611220061779, + 0.01418408751487732, + 0.009109964594244957, + 0.010358836501836777, + 0.010726344771683216, + 0.010342122055590153, + 0.04680623859167099, + 0.03603212535381317, + 0.0575505830347538, + 0.02419750764966011, + 0.025062331929802895, + 0.028473440557718277, + 0.033497463911771774, + 0.04527011513710022, + 0.007140308618545532, + 0.0046515981666743755, + 0.01320276316255331, + 0.005457800347357988, + 0.005952383857220411, + 0.0034720292314887047, + 0.003909359220415354, + 0.004625784233212471, + 0.006248237565159798, + 0.0037160469219088554, + 0.005060693249106407, + 0.004510669969022274, + 0.003384881652891636, + 0.004425713326781988, + 0.005426890216767788, + 0.007162049412727356, + 0.0034201969392597675, + 0.005131247919052839, + 0.0033541086595505476, + 0.0045715197920799255, + 0.005856213625520468, + 0.004231254104524851, + 0.004931761883199215, + 0.0064266822300851345, + 0.006008892320096493, + 0.017749037593603134, + 0.0027702930383384228, + 0.005062351934611797, + 0.0051760245114564896, + 0.0041955383494496346, + 0.004209374543279409, + 0.01309219654649496, + 0.011830635368824005, + 0.00874798558652401, + 0.01259230263531208, + 0.01042342558503151, + 0.01489902101457119, + 0.008655515499413013, + 0.010314490646123886, + 0.015567971393465996, + 0.002299067098647356, + 0.003058650065213442, + 0.0054127536714077, + 0.0017900150269269943, + 0.003993939608335495, + 0.0028576746117323637, + 0.001987694762647152, + 0.007642528507858515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.001334729720838368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a bright and cheerful 3D cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot following a small wooden rowboat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black cat with a red triangular scarf and emerald green eyes visible at the prow of the boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat constructed of warm brown wood planks with distinct grain, sailing through a shimmering azure sea with white waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature brilliant sunlight with extremely high color saturation and fluffy cotton-candy-like clouds in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248221103101969 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005248221103101969, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248221103101969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3127943277359009, + 0.11453091353178024, + 0.0551699697971344, + 0.17700763046741486, + 0.03454428166151047, + 0.032292675226926804, + 0.05202949419617653, + 0.07095333933830261, + 0.0965186357498169, + 0.05415870249271393 + ], + "credit_map": [ + 0.05975758656859398, + 0.02172025665640831, + 0.06035721302032471, + 0.04273786395788193, + 0.0361248143017292, + 0.026234213262796402, + 0.023052407428622246, + 0.04280997812747955, + 0.020139876753091812, + 0.011031572706997395, + 0.02069759927690029, + 0.022472543641924858, + 0.013611708767712116, + 0.007793237920850515, + 0.010863278061151505, + 0.007921099662780762, + 0.008801005780696869, + 0.006151849403977394, + 0.009978893212974072, + 0.0066894469782710075, + 0.00846035499125719, + 0.00477093318477273, + 0.005143269896507263, + 0.005174214486032724, + 0.02960694395005703, + 0.017730887979269028, + 0.04474838823080063, + 0.012378886342048645, + 0.012896421365439892, + 0.015462519600987434, + 0.011157824657857418, + 0.03302575647830963, + 0.005427680443972349, + 0.0030025821179151535, + 0.010016730055212975, + 0.003195520956069231, + 0.0027691086288541555, + 0.0036261952482163906, + 0.0028486342635005713, + 0.003657830646261573, + 0.00509838480502367, + 0.0034498947206884623, + 0.004855906590819359, + 0.004346129484474659, + 0.004792018327862024, + 0.002440063748508692, + 0.00440655741840601, + 0.002903718501329422, + 0.005423218011856079, + 0.008700033649802208, + 0.00491005415096879, + 0.006455196999013424, + 0.00741228275001049, + 0.005854361690580845, + 0.00585519801825285, + 0.007419150788336992, + 0.007269199471920729, + 0.021077970042824745, + 0.004068220965564251, + 0.005456320010125637, + 0.006866303738206625, + 0.005076082423329353, + 0.005369134712964296, + 0.01577010564506054, + 0.010108093731105328, + 0.007765533402562141, + 0.024576740339398384, + 0.009385762736201286, + 0.008262605406343937, + 0.007509224582463503, + 0.007872704416513443, + 0.02103797346353531, + 0.007893750444054604, + 0.007613422814756632, + 0.005033697932958603, + 0.004183315206319094, + 0.007367295678704977, + 0.005240939557552338, + 0.004921401385217905, + 0.011904881335794926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005248221103101969 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_tensor([2147, 2147])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_tensor([2147, 2147])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4f0c603fbd18c34d8512630e39fe881942269cba --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_tensor([2147, 2147])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly tilting, long shot captures large, fluffy flakes of snow, like cotton wool slowed down in time, drifting leisurely down a quiet, night-shrouded city street corner. The warm orange glow of an old cast iron street lamp illuminates the vertical column of air, making each swirling, falling snowflake distinctly visible within the beam, its edges shimmering with a soft halo, its trails trailing like fleeting golden threads. The camera slowly pans down to reveal the ancient cobblestone streets below. The damp stone surfaces reflect the light, creating a mottled contrast with the light, freshly fallen snow. The outlines of brick buildings in the background appear soft and blurred against the snow, with only a few windows filtering through faint interior light, adding a touch of homely life to the chilly evening. The entire image is bathed in a distinct warm yellow filter, with slightly desaturated colors and a subtle film grain, creating a deeply tranquil, nostalgic, and poetic retro atmosphere, as if a cherished winter memory is relived before our eyes.", + "seed": 32760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_2147_seed_32760_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14239130169153214, + "vlm_logprob": -0.14239130169153214, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.14239130169153214, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6342698335647583 + ], + "opposite_logprob_tokens": [ + -0.7592698335647583 + ], + "logprob_reward": -0.6342698335647583, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6342698335647583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037708114832639694, + 0.025045517832040787, + 0.4854549169540405, + 0.019914506003260612, + 0.09720875322818756, + 0.08923839032649994, + 0.063246950507164, + 0.10774579644203186, + 0.054298631846904755, + 0.020138459280133247 + ], + "credit_map": [ + 0.0036518522538244724, + 0.0031604617834091187, + 0.00571815762668848, + 0.011060080491006374, + 0.00232351990416646, + 0.005185054615139961, + 0.0035615283995866776, + 0.0030474578961730003, + 0.002805811120197177, + 0.002316102385520935, + 0.00343327596783638, + 0.004648418165743351, + 0.0024349403101950884, + 0.0026921159587800503, + 0.0036347084678709507, + 0.003080144291743636, + 0.05471532791852951, + 0.028955325484275818, + 0.15060313045978546, + 0.04735808074474335, + 0.016427747905254364, + 0.14106152951717377, + 0.018875852227211, + 0.02745789662003517, + 0.0016888193786144257, + 0.002250485587865114, + 0.002954127499833703, + 0.002849308308213949, + 0.002089484827592969, + 0.0022798622958362103, + 0.003716936567798257, + 0.0020854799076914787, + 0.007844262756407261, + 0.00782974436879158, + 0.0139691811054945, + 0.029994817450642586, + 0.005911221262067556, + 0.015963904559612274, + 0.009905402548611164, + 0.005790208000689745, + 0.007547863759100437, + 0.005401304457336664, + 0.03427215665578842, + 0.01464301347732544, + 0.004276548977941275, + 0.005084426142275333, + 0.004697326570749283, + 0.01331573911011219, + 0.005516085308045149, + 0.009690850041806698, + 0.014896064065396786, + 0.010280335322022438, + 0.005792825948446989, + 0.006134920287877321, + 0.006260610185563564, + 0.004675254691392183, + 0.012372039258480072, + 0.008034934289753437, + 0.02399812825024128, + 0.017461232841014862, + 0.008152139373123646, + 0.01058711763471365, + 0.012819629162549973, + 0.01432056725025177, + 0.00412317831069231, + 0.00308533338829875, + 0.010752109810709953, + 0.008690448477864265, + 0.0040963697247207165, + 0.0038037479389458895, + 0.010619739070534706, + 0.009127701632678509, + 0.0018114774720743299, + 0.0024723869282752275, + 0.00270812283270061, + 0.0028717266395688057, + 0.001982914749532938, + 0.0018437702674418688, + 0.003558647120371461, + 0.0028894112911075354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.6342698335647583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598189905285835 + ], + "opposite_logprob_tokens": [ + -3.8955981731414795 + ], + "logprob_reward": -0.020598189905285835, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598189905285835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0983523577451706, + 0.08901762217283249, + 0.23052939772605896, + 0.06543759256601334, + 0.04905799403786659, + 0.07544662803411484, + 0.09852495044469833, + 0.06540621817111969, + 0.205562025308609, + 0.02266518399119377 + ], + "credit_map": [ + 0.012057237327098846, + 0.009684927761554718, + 0.010884812101721764, + 0.02508956380188465, + 0.006727181375026703, + 0.014741404913365841, + 0.008875838480889797, + 0.010291391052305698, + 0.011975189670920372, + 0.008561105467379093, + 0.014312617480754852, + 0.01175233256071806, + 0.008007630705833435, + 0.013170183636248112, + 0.00791072379797697, + 0.01332783978432417, + 0.021822389215230942, + 0.01313364040106535, + 0.07783867418766022, + 0.02735605090856552, + 0.00820244662463665, + 0.06007658317685127, + 0.011736955493688583, + 0.010362658649682999, + 0.0072183930315077305, + 0.0059494078159332275, + 0.01430866215378046, + 0.009352869354188442, + 0.004823875613510609, + 0.0072340164333581924, + 0.007970200851559639, + 0.008580168709158897, + 0.004184318240731955, + 0.004716638475656509, + 0.005067556165158749, + 0.011647067032754421, + 0.003235951531678438, + 0.012386661022901535, + 0.0040795449167490005, + 0.003740256419405341, + 0.005553996190428734, + 0.00568488659337163, + 0.029231877997517586, + 0.017785482108592987, + 0.004238691180944443, + 0.004469729494303465, + 0.0039886124432086945, + 0.004493352025747299, + 0.009233623743057251, + 0.013241562061011791, + 0.024477502331137657, + 0.01501717884093523, + 0.008952388539910316, + 0.009834472090005875, + 0.009589744731783867, + 0.008178478106856346, + 0.007736104540526867, + 0.005341254640370607, + 0.013675915077328682, + 0.01043118629604578, + 0.005146082490682602, + 0.006457798182964325, + 0.007232164964079857, + 0.00938571710139513, + 0.01087698619812727, + 0.006849922705441713, + 0.05270392820239067, + 0.06869473308324814, + 0.007002347148954868, + 0.0072136372327804565, + 0.036459341645240784, + 0.015761137008666992, + 0.0016167275607585907, + 0.0019060461781919003, + 0.003470000112429261, + 0.005014495924115181, + 0.0022282698191702366, + 0.0025347606278955936, + 0.002501519862562418, + 0.0033933643717318773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.020598189905285835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635362558066845 + ], + "opposite_logprob_tokens": [ + -3.651353597640991 + ], + "logprob_reward": -0.02635362558066845, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635362558066845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03590155765414238, + 0.021740155294537544, + 0.1776445508003235, + 0.08021707832813263, + 0.05978675186634064, + 0.23098932206630707, + 0.07489635795354843, + 0.06662038713693619, + 0.22054073214530945, + 0.03166307508945465 + ], + "credit_map": [ + 0.004718862473964691, + 0.006670902017503977, + 0.004977743607014418, + 0.004743120167404413, + 0.0034087025560438633, + 0.0037512804847210646, + 0.0038134895730763674, + 0.0038174577057361603, + 0.004144264385104179, + 0.0019132227171212435, + 0.002390463137999177, + 0.004021544475108385, + 0.001925127930007875, + 0.0025559377390891314, + 0.0026898644864559174, + 0.0020997305400669575, + 0.016955414786934853, + 0.010454472154378891, + 0.0631326362490654, + 0.03600909188389778, + 0.007416317705065012, + 0.02515687420964241, + 0.01205521821975708, + 0.006464522331953049, + 0.008823366835713387, + 0.008468756452202797, + 0.020753592252731323, + 0.011959300376474857, + 0.005038897041231394, + 0.008938539773225784, + 0.0069754174910485744, + 0.009259210899472237, + 0.004633806645870209, + 0.006001363042742014, + 0.007076963782310486, + 0.009203124791383743, + 0.006410668604075909, + 0.010291559621691704, + 0.0077738589607179165, + 0.008395408280193806, + 0.020263170823454857, + 0.011478707194328308, + 0.09147781133651733, + 0.01753396727144718, + 0.01084538921713829, + 0.016142619773745537, + 0.009422698989510536, + 0.05382494255900383, + 0.006485655903816223, + 0.015645276755094528, + 0.010546102188527584, + 0.012028425931930542, + 0.009295297786593437, + 0.006793709937483072, + 0.007216671947389841, + 0.006885225418955088, + 0.006906851660460234, + 0.007436820305883884, + 0.01133549865335226, + 0.012928428128361702, + 0.007153948303312063, + 0.006575530394911766, + 0.006959076970815659, + 0.007324232719838619, + 0.018522920086979866, + 0.018795540556311607, + 0.05724511295557022, + 0.03847905620932579, + 0.012185316532850266, + 0.012538835406303406, + 0.035648781806230545, + 0.027125172317028046, + 0.003174449549987912, + 0.0025528494734317064, + 0.004852373618632555, + 0.004342109430581331, + 0.004860924556851387, + 0.003153358120471239, + 0.00426449254155159, + 0.004462515469640493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.02635362558066845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540297582745552 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.012540297582745552, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540297582745552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06283954530954361, + 0.07646675407886505, + 0.08268512040376663, + 0.02266090176999569, + 0.05564753711223602, + 0.5589520335197449, + 0.04894373193383217, + 0.028325803577899933, + 0.03894826024770737, + 0.024530239403247833 + ], + "credit_map": [ + 0.008544455282390118, + 0.012458058074116707, + 0.007281665224581957, + 0.007416431792080402, + 0.0063315024599432945, + 0.0075252181850373745, + 0.0076716127805411816, + 0.005610606633126736, + 0.00960355345159769, + 0.007021900732070208, + 0.010723292827606201, + 0.011205249466001987, + 0.006124446168541908, + 0.01610374078154564, + 0.007501737214624882, + 0.008182835765182972, + 0.008808501064777374, + 0.004854858387261629, + 0.027439316734671593, + 0.009050318039953709, + 0.0028711771592497826, + 0.022775724530220032, + 0.002690213732421398, + 0.004195016343146563, + 0.002127578016370535, + 0.0027027057949453592, + 0.0059306020848453045, + 0.0030047816690057516, + 0.002065953565761447, + 0.002785623073577881, + 0.001856315997429192, + 0.002187340520322323, + 0.005239797290414572, + 0.004885904025286436, + 0.00820925273001194, + 0.015895901247859, + 0.004115547984838486, + 0.00624223705381155, + 0.007101332303136587, + 0.0039575607515871525, + 0.045811451971530914, + 0.025938814505934715, + 0.2331816703081131, + 0.04683014377951622, + 0.025861982256174088, + 0.039485953748226166, + 0.019767293706536293, + 0.12207476794719696, + 0.0043842606246471405, + 0.007705411873757839, + 0.01280994713306427, + 0.007335251662880182, + 0.004559166729450226, + 0.0048259710893034935, + 0.004426214378327131, + 0.002897505881264806, + 0.0031218568328768015, + 0.0025225249119102955, + 0.006386935710906982, + 0.004540181718766689, + 0.0024813604541122913, + 0.002845251467078924, + 0.003203189466148615, + 0.0032245046459138393, + 0.0027133007533848286, + 0.002258285414427519, + 0.008480671793222427, + 0.009137507528066635, + 0.001890774117782712, + 0.001979171298444271, + 0.006607855670154095, + 0.005880691111087799, + 0.002076221164315939, + 0.0024640520568937063, + 0.003258843207731843, + 0.0030111197847872972, + 0.0029924344271421432, + 0.003283774247393012, + 0.0038223336450755596, + 0.00362146133556962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.012540297582745552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194561824202538 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.018194561824202538, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194561824202538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06529131531715393, + 0.03744133561849594, + 0.10570821166038513, + 0.047038327902555466, + 0.025978492572903633, + 0.06251563131809235, + 0.11525964736938477, + 0.02762986719608307, + 0.474610835313797, + 0.03852631896734238 + ], + "credit_map": [ + 0.011066927574574947, + 0.009142233990132809, + 0.010067055933177471, + 0.01171292457729578, + 0.004906000103801489, + 0.006585695780813694, + 0.0059490809217095375, + 0.005861392244696617, + 0.004596077837049961, + 0.0033296479377895594, + 0.005791230592876673, + 0.004843517672270536, + 0.005040162242949009, + 0.005144691094756126, + 0.004997151903808117, + 0.0036988549400120974, + 0.011808598414063454, + 0.00724012078717351, + 0.017142366617918015, + 0.01200722623616457, + 0.005148142576217651, + 0.02964864671230316, + 0.011588508263230324, + 0.011124609969556332, + 0.0060231927782297134, + 0.00503843417391181, + 0.006086525972932577, + 0.006966501474380493, + 0.00382151803933084, + 0.0045705800876021385, + 0.008971880190074444, + 0.005559691693633795, + 0.0026661392766982317, + 0.002771178027614951, + 0.004795509856194258, + 0.003134450875222683, + 0.003120468696579337, + 0.003268957370892167, + 0.0027896605897694826, + 0.003432127181440592, + 0.005973434541374445, + 0.004328148439526558, + 0.012177946977317333, + 0.011155885644257069, + 0.003191102761775255, + 0.004551983904093504, + 0.004837313201278448, + 0.016299813985824585, + 0.01010679081082344, + 0.019553830847144127, + 0.024113211780786514, + 0.01860324665904045, + 0.010846494697034359, + 0.012069782242178917, + 0.010932899080216885, + 0.00903338473290205, + 0.002695862203836441, + 0.002276776125654578, + 0.004453623667359352, + 0.005163718480616808, + 0.0023714264389127493, + 0.002156543545424938, + 0.003753868630155921, + 0.004758048802614212, + 0.054080795496702194, + 0.03793023154139519, + 0.14044569432735443, + 0.05325554683804512, + 0.025481486693024635, + 0.03383302316069603, + 0.055320318788290024, + 0.07426375150680542, + 0.0035886699333786964, + 0.005032326094806194, + 0.006449946202337742, + 0.00786565151065588, + 0.0025889019016176462, + 0.0026217414997518063, + 0.00466130068525672, + 0.005717783235013485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.018194561824202538 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly tilting, long shot captures large, fluffy flakes of snow, like cotton wool slowed down in time, drifting leisurely down a quiet, night-shrouded city street corner. The warm orange glow of an old cast iron street lamp illuminates the vertical column of air, making each swirling, falling snowflake distinctly visible within the beam, its edges shimmering with a soft halo, its trails trailing like fleeting golden threads. The camera slowly pans down to reveal the ancient cobblestone streets below. The damp stone surfaces reflect the light, creating a mottled contrast with the light, freshly fallen snow. The outlines of brick buildings in the background appear soft and blurred against the snow, with only a few windows filtering through faint interior light, adding a touch of homely life to the chilly evening. The entire image is bathed in a distinct warm yellow filter, with slightly desaturated colors and a subtle film grain, creating a deeply tranquil, nostalgic, and poetic retro atmosphere, as if a cherished winter memory is relived before our eyes.", + "seed": 32761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_2147_seed_32761_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023690763302147387, + "vlm_logprob": -0.023690763302147387, + "vlm_yes_no_margin": 3.849999952316284, + "vlm_token_credit": -0.023690763302147387, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02716860920190811 + ], + "opposite_logprob_tokens": [ + -3.6521685123443604 + ], + "logprob_reward": -0.02716860920190811, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02716860920190811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1715024709701538, + 0.1485593169927597, + 0.06809109449386597, + 0.05388769134879112, + 0.26571735739707947, + 0.04527156800031662, + 0.02426644042134285, + 0.0964205339550972, + 0.0657413899898529, + 0.06054221838712692 + ], + "credit_map": [ + 0.012340656481683254, + 0.03363719582557678, + 0.015338746830821037, + 0.022539736703038216, + 0.023260360583662987, + 0.011558949016034603, + 0.03717942535877228, + 0.01564738340675831, + 0.012187710031867027, + 0.04227295517921448, + 0.010375122539699078, + 0.017904488369822502, + 0.027690576389431953, + 0.007968743331730366, + 0.020090971142053604, + 0.010068732313811779, + 0.007265465334057808, + 0.012557337991893291, + 0.009599791839718819, + 0.007328456733375788, + 0.008924823254346848, + 0.005455507896840572, + 0.009311627596616745, + 0.00764807453379035, + 0.006354349199682474, + 0.007826322689652443, + 0.005188713315874338, + 0.011365627869963646, + 0.0053045907989144325, + 0.0038473564200103283, + 0.007901571691036224, + 0.006099152844399214, + 0.025347381830215454, + 0.021040290594100952, + 0.03836746886372566, + 0.07839556038379669, + 0.0165113378316164, + 0.021869750693440437, + 0.033586520701646805, + 0.030599024146795273, + 0.00823722779750824, + 0.0057950858026742935, + 0.004582954570651054, + 0.00591430626809597, + 0.0035118372179567814, + 0.004321006126701832, + 0.005381837487220764, + 0.0075273094698786736, + 0.0017557389801368117, + 0.0022664840798825026, + 0.004611189942806959, + 0.004974101670086384, + 0.0031064206268638372, + 0.0030165575444698334, + 0.002754450077190995, + 0.0017814955208450556, + 0.008619267493486404, + 0.011757216416299343, + 0.009939589537680149, + 0.012390533462166786, + 0.016499662771821022, + 0.01017637737095356, + 0.01462413091212511, + 0.0124137494713068, + 0.0038443736266344786, + 0.00532766617834568, + 0.009661120362579823, + 0.015377704054117203, + 0.011798779480159283, + 0.005290019325911999, + 0.008954509161412716, + 0.005487209185957909, + 0.0033053550869226456, + 0.005583086051046848, + 0.002852809615433216, + 0.004231534898281097, + 0.022395862266421318, + 0.004580152686685324, + 0.014583710581064224, + 0.0030096995178610086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.02716860920190811 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819959655404091 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.01819959655404091, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819959655404091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07701127976179123, + 0.031860508024692535, + 0.14686933159828186, + 0.21263132989406586, + 0.0736866369843483, + 0.0782293900847435, + 0.04833764582872391, + 0.0576019324362278, + 0.03377678245306015, + 0.23999515175819397 + ], + "credit_map": [ + 0.00903109647333622, + 0.01051551103591919, + 0.008272056467831135, + 0.013058631680905819, + 0.00948653556406498, + 0.008922653272747993, + 0.010226620361208916, + 0.007498177234083414, + 0.003676316002383828, + 0.005926296580582857, + 0.002104393672198057, + 0.0026804618537425995, + 0.004218952264636755, + 0.0035757028963416815, + 0.004740292672067881, + 0.004938090220093727, + 0.017987128347158432, + 0.013927890919148922, + 0.027035119011998177, + 0.015345873311161995, + 0.01492911484092474, + 0.010479793883860111, + 0.029549280181527138, + 0.01761513575911522, + 0.03596915304660797, + 0.018130691722035408, + 0.01840067468583584, + 0.01796281337738037, + 0.025306887924671173, + 0.01900370977818966, + 0.03498911112546921, + 0.04286829009652138, + 0.008878611028194427, + 0.005416577216237783, + 0.00971333310008049, + 0.016349919140338898, + 0.007335350848734379, + 0.007887208834290504, + 0.008924116380512714, + 0.009181520901620388, + 0.00999472290277481, + 0.007833091542124748, + 0.010597710497677326, + 0.008292948827147484, + 0.007012495305389166, + 0.00877170730382204, + 0.014148615300655365, + 0.011578096076846123, + 0.003960529342293739, + 0.003776364726945758, + 0.005128554068505764, + 0.007937104441225529, + 0.008529563434422016, + 0.009135130792856216, + 0.00587194599211216, + 0.003998456057161093, + 0.0038169301114976406, + 0.016180122271180153, + 0.004167707171291113, + 0.004590059630572796, + 0.004890994634479284, + 0.005351657047867775, + 0.005019791889935732, + 0.01358466874808073, + 0.0025293088983744383, + 0.0030960668809711933, + 0.005701929330825806, + 0.0049850819632411, + 0.004492396023124456, + 0.005005362443625927, + 0.005116766318678856, + 0.0028498717583715916, + 0.012297145090997219, + 0.019352678209543228, + 0.006496810354292393, + 0.018187811598181725, + 0.08327743411064148, + 0.011240635998547077, + 0.07792451977729797, + 0.011218128725886345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.01819959655404091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809512034058571 + ], + "opposite_logprob_tokens": [ + -3.288095235824585 + ], + "logprob_reward": -0.03809512034058571, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809512034058571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11161983758211136, + 0.12687844038009644, + 0.06784222275018692, + 0.1400466114282608, + 0.09912008792161942, + 0.14241161942481995, + 0.04494717717170715, + 0.08007266372442245, + 0.08632353693246841, + 0.10073781758546829 + ], + "credit_map": [ + 0.016056835651397705, + 0.01820380613207817, + 0.015553529374301434, + 0.014690387062728405, + 0.012898971326649189, + 0.006901692599058151, + 0.01850394904613495, + 0.008810663595795631, + 0.010381864383816719, + 0.03494669124484062, + 0.012059113010764122, + 0.015510919503867626, + 0.022919809445738792, + 0.008918440900743008, + 0.014809563755989075, + 0.007332035340368748, + 0.0067850155755877495, + 0.007775678765028715, + 0.00938799511641264, + 0.008857274428009987, + 0.008038015104830265, + 0.0050159720703959465, + 0.014361455105245113, + 0.00762081379070878, + 0.02219124138355255, + 0.011389275081455708, + 0.015345939435064793, + 0.016896359622478485, + 0.015516371466219425, + 0.013391422107815742, + 0.02173355221748352, + 0.0235824566334486, + 0.012039086781442165, + 0.007803135085850954, + 0.015127896331250668, + 0.02235558070242405, + 0.009878002107143402, + 0.010304524563252926, + 0.010841689072549343, + 0.010770176537334919, + 0.013524996116757393, + 0.009945206344127655, + 0.016436589881777763, + 0.015150655061006546, + 0.011014212854206562, + 0.010388691909611225, + 0.03602726012468338, + 0.029924005270004272, + 0.004539616405963898, + 0.0039419024251401424, + 0.005091015249490738, + 0.006435704417526722, + 0.0071548218838870525, + 0.007464277092367411, + 0.005593379959464073, + 0.004726456478238106, + 0.004109656438231468, + 0.0077525293454527855, + 0.006483875215053558, + 0.006530967075377703, + 0.01991480030119419, + 0.011204857379198074, + 0.005824652034789324, + 0.0182513315230608, + 0.004707645624876022, + 0.010514410212635994, + 0.010158209130167961, + 0.014266496524214745, + 0.020753730088472366, + 0.008349377661943436, + 0.009884006343781948, + 0.0076896511018276215, + 0.005050724372267723, + 0.004515177104622126, + 0.004686164669692516, + 0.012846074998378754, + 0.020560862496495247, + 0.005836131051182747, + 0.03867034614086151, + 0.008572334423661232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.03809512034058571 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863721314817667 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.00863721314817667, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863721314817667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059976283460855484, + 0.06922789663076401, + 0.18226438760757446, + 0.0718708336353302, + 0.07034265995025635, + 0.18821203708648682, + 0.03728240728378296, + 0.0669965147972107, + 0.038396794348955154, + 0.2154301404953003 + ], + "credit_map": [ + 0.006128019653260708, + 0.010910214856266975, + 0.007532195188105106, + 0.009268145076930523, + 0.00820342730730772, + 0.004312578123062849, + 0.009829863905906677, + 0.003791843308135867, + 0.0070623611100018024, + 0.01524578221142292, + 0.009545568376779556, + 0.009547760710120201, + 0.009510235860943794, + 0.00697677256539464, + 0.006595962680876255, + 0.004743455909192562, + 0.017177850008010864, + 0.01883184351027012, + 0.014510071836411953, + 0.023795057088136673, + 0.023689253255724907, + 0.013626853935420513, + 0.049287836998701096, + 0.02134564146399498, + 0.009527833200991154, + 0.005597900133579969, + 0.005323359277099371, + 0.008241577073931694, + 0.007412254810333252, + 0.0063445596024394035, + 0.014475403353571892, + 0.014947944320738316, + 0.007166070863604546, + 0.005499597638845444, + 0.010293657891452312, + 0.020060887560248375, + 0.0047197043895721436, + 0.005865473300218582, + 0.009035935625433922, + 0.007701335474848747, + 0.011818538419902325, + 0.016027119010686874, + 0.01935642771422863, + 0.012140733189880848, + 0.03335314989089966, + 0.04084121063351631, + 0.029186464846134186, + 0.025488391518592834, + 0.002665411913767457, + 0.003914396744221449, + 0.004420998506247997, + 0.007096824236214161, + 0.0055664521642029285, + 0.0058830599300563335, + 0.004768875427544117, + 0.0029663920868188143, + 0.0053480928763747215, + 0.013338505290448666, + 0.007003897801041603, + 0.005989364814013243, + 0.009349261410534382, + 0.0069493260234594345, + 0.00463205436244607, + 0.014386013150215149, + 0.0022541803773492575, + 0.003992754500359297, + 0.005510636139661074, + 0.0076025123707950115, + 0.0037435945123434067, + 0.006110304035246372, + 0.005797898396849632, + 0.003384915878996253, + 0.010441487655043602, + 0.015952419489622116, + 0.0057305204682052135, + 0.018638810142874718, + 0.07048427313566208, + 0.010471788235008717, + 0.07291190326213837, + 0.010798967443406582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.00863721314817667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353277266025543 + ], + "opposite_logprob_tokens": [ + -3.651353359222412 + ], + "logprob_reward": -0.026353277266025543, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353277266025543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13964781165122986, + 0.09418021142482758, + 0.14999708533287048, + 0.0749131590127945, + 0.08131832629442215, + 0.08977887034416199, + 0.038157813251018524, + 0.08943166583776474, + 0.11189161241054535, + 0.13068343698978424 + ], + "credit_map": [ + 0.013162416405975819, + 0.01861824095249176, + 0.01929347962141037, + 0.021182367578148842, + 0.01827211119234562, + 0.010207828134298325, + 0.027784371748566628, + 0.011126985773444176, + 0.008168336004018784, + 0.024702340364456177, + 0.007572112139314413, + 0.009871125221252441, + 0.016477854922413826, + 0.007651489693671465, + 0.012780213728547096, + 0.0069567239843308926, + 0.019761094823479652, + 0.013155977241694927, + 0.034775152802467346, + 0.014832214452326298, + 0.013405540958046913, + 0.009974763728678226, + 0.025818806141614914, + 0.018273524940013885, + 0.012909737415611744, + 0.006788135971873999, + 0.00617020670324564, + 0.006368182133883238, + 0.008692161180078983, + 0.009116471745073795, + 0.012568362057209015, + 0.012299896217882633, + 0.008404538035392761, + 0.006648311857134104, + 0.012867380864918232, + 0.01954471692442894, + 0.007271552458405495, + 0.008897287771105766, + 0.009271549992263317, + 0.008412979543209076, + 0.007895134389400482, + 0.004176526330411434, + 0.013713499531149864, + 0.011243521235883236, + 0.005832699593156576, + 0.0073055666871368885, + 0.02192731201648712, + 0.01768460124731064, + 0.002316276077181101, + 0.0032009875867515802, + 0.006159126292914152, + 0.009926416911184788, + 0.004211357794702053, + 0.004481752403080463, + 0.005574295297265053, + 0.0022875992581248283, + 0.0047781337052583694, + 0.009452142752707005, + 0.006081555038690567, + 0.006861412897706032, + 0.027885008603334427, + 0.011857300065457821, + 0.00644312659278512, + 0.01607297919690609, + 0.005860721226781607, + 0.011158688925206661, + 0.010747711174190044, + 0.02312123216688633, + 0.012812512926757336, + 0.01789233833551407, + 0.016406890004873276, + 0.01389150507748127, + 0.006908688694238663, + 0.01035736221820116, + 0.003703406546264887, + 0.009242242202162743, + 0.04513741657137871, + 0.006505019962787628, + 0.04223854839801788, + 0.006590740755200386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.026353277266025543 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly tilting, long shot captures large, fluffy flakes of snow, like cotton wool slowed down in time, drifting leisurely down a quiet, night-shrouded city street corner. The warm orange glow of an old cast iron street lamp illuminates the vertical column of air, making each swirling, falling snowflake distinctly visible within the beam, its edges shimmering with a soft halo, its trails trailing like fleeting golden threads. The camera slowly pans down to reveal the ancient cobblestone streets below. The damp stone surfaces reflect the light, creating a mottled contrast with the light, freshly fallen snow. The outlines of brick buildings in the background appear soft and blurred against the snow, with only a few windows filtering through faint interior light, adding a touch of homely life to the chilly evening. The entire image is bathed in a distinct warm yellow filter, with slightly desaturated colors and a subtle film grain, creating a deeply tranquil, nostalgic, and poetic retro atmosphere, as if a cherished winter memory is relived before our eyes.", + "seed": 32762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_2147_seed_32762_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09410831928253174, + "vlm_logprob": -0.09410831928253174, + "vlm_yes_no_margin": 3.474999988079071, + "vlm_token_credit": -0.09410831928253174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38825923204421997 + ], + "opposite_logprob_tokens": [ + -1.1382591724395752 + ], + "logprob_reward": -0.38825923204421997, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38825923204421997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3640643358230591, + 0.087042897939682, + 0.06155402585864067, + 0.036595024168491364, + 0.15124014019966125, + 0.033999744802713394, + 0.06777600198984146, + 0.13344408571720123, + 0.022538678720593452, + 0.04174507409334183 + ], + "credit_map": [ + 0.03287369757890701, + 0.06490959972143173, + 0.04232246056199074, + 0.061050523072481155, + 0.046382781118154526, + 0.03944574296474457, + 0.038810670375823975, + 0.03826884180307388, + 0.005502719432115555, + 0.013771542347967625, + 0.006410657428205013, + 0.011028100736439228, + 0.014809665270149708, + 0.012634700164198875, + 0.012340147979557514, + 0.010545368306338787, + 0.004131916910409927, + 0.006034414749592543, + 0.008585503324866295, + 0.008181830868124962, + 0.010098972357809544, + 0.0066789486445486546, + 0.0066602821461856365, + 0.011182158254086971, + 0.0021032600197941065, + 0.003211789298802614, + 0.002408870030194521, + 0.0023835853207856417, + 0.00303544569760561, + 0.010512757115066051, + 0.003233768744394183, + 0.00970554631203413, + 0.014536975882947445, + 0.026713619008660316, + 0.01295765396207571, + 0.010103488340973854, + 0.01089237816631794, + 0.049169305711984634, + 0.009802961722016335, + 0.01706375926733017, + 0.002646715147420764, + 0.0037108834367245436, + 0.0023711177054792643, + 0.0028748675249516964, + 0.006708825472742319, + 0.0044693476520478725, + 0.005793366115540266, + 0.005424618721008301, + 0.004403870552778244, + 0.004808090627193451, + 0.009657063521444798, + 0.011234188452363014, + 0.009648787789046764, + 0.009151787497103214, + 0.009067812003195286, + 0.009804404340684414, + 0.011150475591421127, + 0.008393807336688042, + 0.01312252227216959, + 0.012563131749629974, + 0.01096227578818798, + 0.031244590878486633, + 0.011553768999874592, + 0.03445351868867874, + 0.0019353360403329134, + 0.0026502087712287903, + 0.0027561900205910206, + 0.004068668000400066, + 0.00219538901001215, + 0.003327881218865514, + 0.0028669352177530527, + 0.0027380725368857384, + 0.0019374132389202714, + 0.0036727809347212315, + 0.003532210597768426, + 0.009993934072554111, + 0.012481626123189926, + 0.003463233122602105, + 0.003987804055213928, + 0.0026760718319565058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.38825923204421997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029814520850777626 + ], + "opposite_logprob_tokens": [ + -3.5298144817352295 + ], + "logprob_reward": -0.029814520850777626, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029814520850777626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07130218297243118, + 0.15617763996124268, + 0.12325771152973175, + 0.055436477065086365, + 0.1459508091211319, + 0.058075543493032455, + 0.06682992726564407, + 0.1414688527584076, + 0.08077625185251236, + 0.10072457045316696 + ], + "credit_map": [ + 0.007345861755311489, + 0.014234392903745174, + 0.008308026939630508, + 0.012515723705291748, + 0.010713373310863972, + 0.005450312979519367, + 0.004853671882301569, + 0.00788082368671894, + 0.012473654001951218, + 0.027259135618805885, + 0.010166640393435955, + 0.017086798325181007, + 0.019391952082514763, + 0.029231132939457893, + 0.012026829645037651, + 0.028541509062051773, + 0.008158105425536633, + 0.011929732747375965, + 0.018063146620988846, + 0.025304069742560387, + 0.013155022636055946, + 0.013221868313848972, + 0.011674667708575726, + 0.02175111323595047, + 0.005463241599500179, + 0.0071014342829585075, + 0.008570828475058079, + 0.007704006973654032, + 0.0044733439572155476, + 0.0066108861938118935, + 0.007422544993460178, + 0.008090189658105373, + 0.013193521648645401, + 0.023955363780260086, + 0.012537088245153427, + 0.009667102247476578, + 0.010212136432528496, + 0.044615183025598526, + 0.011296538636088371, + 0.020473884418606758, + 0.006612836848944426, + 0.006720747798681259, + 0.003940732683986425, + 0.006130727473646402, + 0.01355011761188507, + 0.007948771119117737, + 0.008141648955643177, + 0.005029963795095682, + 0.004648133646696806, + 0.00573632912710309, + 0.007556237280368805, + 0.009611491113901138, + 0.013527845963835716, + 0.010260012000799179, + 0.006813162937760353, + 0.008676724508404732, + 0.01650449074804783, + 0.011716742999851704, + 0.012600756250321865, + 0.019161079078912735, + 0.017092525959014893, + 0.04208161681890488, + 0.009829140268266201, + 0.012482504360377789, + 0.005845330189913511, + 0.0069150556810200214, + 0.008783254772424698, + 0.009187151677906513, + 0.015688763931393623, + 0.007455458864569664, + 0.009691881015896797, + 0.017209364101290703, + 0.0043347557075321674, + 0.006136646028608084, + 0.01107887364923954, + 0.03303076699376106, + 0.025831567123532295, + 0.0049848491325974464, + 0.00964517705142498, + 0.005681938491761684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.029814520850777626 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810235112905502 + ], + "opposite_logprob_tokens": [ + -3.288102388381958 + ], + "logprob_reward": -0.03810235112905502, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810235112905502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051644809544086456, + 0.01795736514031887, + 0.07369764894247055, + 0.03644389286637306, + 0.09668048471212387, + 0.05071680620312691, + 0.10674573481082916, + 0.29244667291641235, + 0.12672947347164154, + 0.14693713188171387 + ], + "credit_map": [ + 0.0045169624499976635, + 0.010249500162899494, + 0.005164007190614939, + 0.007537560071796179, + 0.0070899599231779575, + 0.005307198502123356, + 0.006612295750528574, + 0.005167326424270868, + 0.0013797247083857656, + 0.003702720860019326, + 0.001280125230550766, + 0.002291747136041522, + 0.0039141494780778885, + 0.0016433726996183395, + 0.0012905916664749384, + 0.002454933011904359, + 0.005532546900212765, + 0.007888630032539368, + 0.01044490560889244, + 0.012357250787317753, + 0.00938795693218708, + 0.00902597513049841, + 0.008636471815407276, + 0.010423910804092884, + 0.003455001162365079, + 0.0048479787074029446, + 0.003341121133416891, + 0.0034548030234873295, + 0.004191372077912092, + 0.00785926915705204, + 0.0034079488832503557, + 0.005886395927518606, + 0.008820543065667152, + 0.01626603864133358, + 0.008076098747551441, + 0.006791313644498587, + 0.006614464335143566, + 0.029479647055268288, + 0.007494581863284111, + 0.013137796893715858, + 0.004720423836261034, + 0.0056519596837460995, + 0.0058006481267511845, + 0.005124865565448999, + 0.007047153078019619, + 0.006995266769081354, + 0.006997527088969946, + 0.008378960192203522, + 0.007793902885168791, + 0.010308008641004562, + 0.01310749351978302, + 0.012990073300898075, + 0.015103669837117195, + 0.014584359712898731, + 0.01620774157345295, + 0.016650492325425148, + 0.03182004392147064, + 0.018274206668138504, + 0.020830320194363594, + 0.026749568060040474, + 0.03503485023975372, + 0.09487734735012054, + 0.016946014016866684, + 0.047914307564496994, + 0.008857287466526031, + 0.011847775429487228, + 0.013800905086100101, + 0.01398527342826128, + 0.022010793909430504, + 0.011745309457182884, + 0.01778627187013626, + 0.02669585682451725, + 0.00598894665017724, + 0.009050455875694752, + 0.013249307870864868, + 0.04998178035020828, + 0.043876804411411285, + 0.007776430808007717, + 0.009829426184296608, + 0.007183977868407965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.03810235112905502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673500495031476 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.00673500495031476, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673500495031476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1764279305934906, + 0.03465479612350464, + 0.10248952358961105, + 0.06348866969347, + 0.07635296881198883, + 0.09011660516262054, + 0.13756506145000458, + 0.18133626878261566, + 0.05646732822060585, + 0.08110089600086212 + ], + "credit_map": [ + 0.01633341982960701, + 0.035110827535390854, + 0.02025221288204193, + 0.02860729582607746, + 0.021020900458097458, + 0.01719309203326702, + 0.022138573229312897, + 0.015771621838212013, + 0.0023394874297082424, + 0.002564701484516263, + 0.0026268207002431154, + 0.003250931389629841, + 0.0042798989452421665, + 0.008340755477547646, + 0.003650244791060686, + 0.007601956836879253, + 0.008015298284590244, + 0.010303734801709652, + 0.013282068073749542, + 0.022862324491143227, + 0.0137924924492836, + 0.009670042432844639, + 0.013134690001606941, + 0.011428874917328358, + 0.004654113668948412, + 0.006101911887526512, + 0.005805895198136568, + 0.007543978746980429, + 0.006711872294545174, + 0.014912126585841179, + 0.006862530950456858, + 0.01089624036103487, + 0.007120517548173666, + 0.013359437696635723, + 0.006277954205870628, + 0.006235177628695965, + 0.005375741049647331, + 0.02299823798239231, + 0.005680613219738007, + 0.00930529274046421, + 0.0083106504753232, + 0.010521871037781239, + 0.007909511215984821, + 0.009523704648017883, + 0.014304419048130512, + 0.014248603954911232, + 0.012194536626338959, + 0.013103307224810123, + 0.01650017313659191, + 0.018937382847070694, + 0.011937438510358334, + 0.008650293573737144, + 0.03475901111960411, + 0.025542160496115685, + 0.011062640696763992, + 0.010175959207117558, + 0.014908366836607456, + 0.013429137878119946, + 0.025462200865149498, + 0.025347644463181496, + 0.016699081286787987, + 0.026771582663059235, + 0.01750941202044487, + 0.041208840906620026, + 0.004266819451004267, + 0.0076265474781394005, + 0.006316428072750568, + 0.008665258064866066, + 0.007809320464730263, + 0.006242562085390091, + 0.006059363018721342, + 0.009481027722358704, + 0.003770862938836217, + 0.0063692862167954445, + 0.009074760600924492, + 0.024662794545292854, + 0.01673217862844467, + 0.005458469968289137, + 0.008737263269722462, + 0.0062952833250164986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.00673500495031476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1711990386247635, + 0.08713672310113907, + 0.06766415387392044, + 0.04738914594054222, + 0.09052299708127975, + 0.057066820561885834, + 0.16158807277679443, + 0.0762089192867279, + 0.034332454204559326, + 0.20689161121845245 + ], + "credit_map": [ + 0.016518117859959602, + 0.0283362977206707, + 0.02229938469827175, + 0.026401447132229805, + 0.02074446715414524, + 0.019102247431874275, + 0.018970144912600517, + 0.01882695034146309, + 0.0077719311229884624, + 0.009828946553170681, + 0.008040721528232098, + 0.011393173597753048, + 0.0089461300522089, + 0.01479247584939003, + 0.011271141469478607, + 0.015092209912836552, + 0.006791464053094387, + 0.006709500681608915, + 0.006395671982318163, + 0.013241825625300407, + 0.010872847400605679, + 0.006694888696074486, + 0.008684164844453335, + 0.008273797109723091, + 0.005268078297376633, + 0.010793991386890411, + 0.003205318236723542, + 0.004943312611430883, + 0.004271793179214001, + 0.007661805488169193, + 0.005296722520142794, + 0.005948129575699568, + 0.008511209860444069, + 0.014899027533829212, + 0.007310335990041494, + 0.00686252536252141, + 0.006737738847732544, + 0.0284867100417614, + 0.006611558608710766, + 0.011103905737400055, + 0.0045166765339672565, + 0.009176111780107021, + 0.0033526686020195484, + 0.005545365158468485, + 0.009292224422097206, + 0.0115060331299901, + 0.0051277694292366505, + 0.008549976162612438, + 0.01226761657744646, + 0.012419315055012703, + 0.024465596303343773, + 0.028122033923864365, + 0.017992651090025902, + 0.022346802055835724, + 0.019303591921925545, + 0.02467048354446888, + 0.0070482236333191395, + 0.008825693279504776, + 0.009591801092028618, + 0.013845537789165974, + 0.010361691005527973, + 0.007783838082104921, + 0.010552311316132545, + 0.008199834264814854, + 0.0040231249295175076, + 0.0033666060771793127, + 0.0035623693838715553, + 0.0036276939790695906, + 0.004895612597465515, + 0.004272203426808119, + 0.004682690370827913, + 0.0059021590277552605, + 0.008336277678608894, + 0.01451239176094532, + 0.018630100414156914, + 0.06863923370838165, + 0.05896584689617157, + 0.011137972585856915, + 0.01610475406050682, + 0.010565049946308136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.007630487438291311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly tilting, long shot captures large, fluffy flakes of snow, like cotton wool slowed down in time, drifting leisurely down a quiet, night-shrouded city street corner. The warm orange glow of an old cast iron street lamp illuminates the vertical column of air, making each swirling, falling snowflake distinctly visible within the beam, its edges shimmering with a soft halo, its trails trailing like fleeting golden threads. The camera slowly pans down to reveal the ancient cobblestone streets below. The damp stone surfaces reflect the light, creating a mottled contrast with the light, freshly fallen snow. The outlines of brick buildings in the background appear soft and blurred against the snow, with only a few windows filtering through faint interior light, adding a touch of homely life to the chilly evening. The entire image is bathed in a distinct warm yellow filter, with slightly desaturated colors and a subtle film grain, creating a deeply tranquil, nostalgic, and poetic retro atmosphere, as if a cherished winter memory is relived before our eyes.", + "seed": 32763, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_2/index_2147_seed_32763_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.044715655222535136, + "vlm_logprob": -0.044715655222535136, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.044715655222535136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12792126834392548 + ], + "opposite_logprob_tokens": [ + -2.1279213428497314 + ], + "logprob_reward": -0.12792126834392548, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12792126834392548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41730162501335144, + 0.03538834676146507, + 0.019586259499192238, + 0.022800691425800323, + 0.05684087797999382, + 0.27747583389282227, + 0.01822900027036667, + 0.02039170451462269, + 0.01609962247312069, + 0.11588599532842636 + ], + "credit_map": [ + 0.051846012473106384, + 0.012184923514723778, + 0.1408136636018753, + 0.04265022650361061, + 0.017218418419361115, + 0.023750251159071922, + 0.029433395713567734, + 0.09940476715564728, + 0.003756217658519745, + 0.0030480176210403442, + 0.009329178370535374, + 0.005098253488540649, + 0.002169159008190036, + 0.0031307516619563103, + 0.004102575592696667, + 0.004754193127155304, + 0.0020225285552442074, + 0.0020992839708924294, + 0.0026593483053147793, + 0.002898637903854251, + 0.0021433879155665636, + 0.002117433585226536, + 0.0026653369423002005, + 0.0029803027864545584, + 0.0017003781395033002, + 0.0037694440688937902, + 0.002026638947427273, + 0.0026063495315611362, + 0.0023210393264889717, + 0.002973450580611825, + 0.0025595324113965034, + 0.00484385946765542, + 0.005236220546066761, + 0.004616450984030962, + 0.0025990731082856655, + 0.018006417900323868, + 0.006529389414936304, + 0.008772308006882668, + 0.006391802802681923, + 0.004689216613769531, + 0.014158758334815502, + 0.009211484342813492, + 0.013103712350130081, + 0.017999855801463127, + 0.02333931438624859, + 0.0843542143702507, + 0.015920130535960197, + 0.09938837587833405, + 0.0019015584839507937, + 0.002079328987747431, + 0.0026082166004925966, + 0.0027845785953104496, + 0.0025192226748913527, + 0.002160230651497841, + 0.002139429096132517, + 0.0020364343654364347, + 0.0029838704504072666, + 0.002528403652831912, + 0.0028032788541167974, + 0.0024728463031351566, + 0.002244719071313739, + 0.002441995544359088, + 0.002067356603220105, + 0.0028492321725934744, + 0.0017746402882039547, + 0.0021246611140668392, + 0.0023574994411319494, + 0.0026730585377663374, + 0.001527857966721058, + 0.001848268904723227, + 0.0021978821605443954, + 0.0015957539435476065, + 0.010017771273851395, + 0.018329571932554245, + 0.01377159170806408, + 0.02263522520661354, + 0.009071331471204758, + 0.008665021508932114, + 0.016408568248152733, + 0.016986913979053497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.12792126834392548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866613447666168 + ], + "opposite_logprob_tokens": [ + -3.04866623878479 + ], + "logprob_reward": -0.04866613447666168, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866613447666168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1701502948999405, + 0.16655880212783813, + 0.08250382542610168, + 0.04282509535551071, + 0.1618451327085495, + 0.15052562952041626, + 0.06406739354133606, + 0.10513056814670563, + 0.021867286413908005, + 0.03452597185969353 + ], + "credit_map": [ + 0.022337842732667923, + 0.008912542834877968, + 0.053655724972486496, + 0.01830928958952427, + 0.0070123509503901005, + 0.007104234769940376, + 0.015262339264154434, + 0.037555962800979614, + 0.019446345046162605, + 0.01139824464917183, + 0.013447239063680172, + 0.022557362914085388, + 0.015161820687353611, + 0.03807632252573967, + 0.02326185815036297, + 0.02320960722863674, + 0.009337354451417923, + 0.006838662549853325, + 0.014662184752523899, + 0.013061740435659885, + 0.006998870521783829, + 0.009388629347085953, + 0.012043598107993603, + 0.010172794573009014, + 0.004209313075989485, + 0.005474167875945568, + 0.006334987003356218, + 0.005369029473513365, + 0.004674836993217468, + 0.005898438394069672, + 0.005016540177166462, + 0.00584778469055891, + 0.013962335884571075, + 0.015640707686543465, + 0.010090501978993416, + 0.04965585470199585, + 0.01657136343419552, + 0.011350558139383793, + 0.015914294868707657, + 0.028659522533416748, + 0.006983340252190828, + 0.005242249928414822, + 0.00674675265327096, + 0.00822898093611002, + 0.01269457209855318, + 0.046638719737529755, + 0.008901936002075672, + 0.0550890788435936, + 0.007954001426696777, + 0.007210195995867252, + 0.008721936494112015, + 0.009727566502988338, + 0.006614758633077145, + 0.00799152534455061, + 0.007917524315416813, + 0.007929885759949684, + 0.010421182960271835, + 0.010922429151833057, + 0.021909790113568306, + 0.015576853416860104, + 0.009083578363060951, + 0.008275904692709446, + 0.019353603944182396, + 0.009587222710251808, + 0.001702839625068009, + 0.002170802094042301, + 0.0035217173863202333, + 0.004292011260986328, + 0.0021431585773825645, + 0.0020223897881805897, + 0.0036236534360796213, + 0.002390712033957243, + 0.004039762541651726, + 0.0038613369688391685, + 0.004851228557527065, + 0.007670741993933916, + 0.0032841493375599384, + 0.0030259890481829643, + 0.003995614126324654, + 0.00379715277813375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04866613447666168 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291101679205894 + ], + "opposite_logprob_tokens": [ + -3.7732911109924316 + ], + "logprob_reward": -0.023291101679205894, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291101679205894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44327741861343384, + 0.051459453999996185, + 0.03227841481566429, + 0.035100001841783524, + 0.05177614837884903, + 0.2652696967124939, + 0.04650210216641426, + 0.020976969972252846, + 0.03777225688099861, + 0.015587525442242622 + ], + "credit_map": [ + 0.055233318358659744, + 0.01048089750111103, + 0.16107679903507233, + 0.047419123351573944, + 0.01587045192718506, + 0.019097935408353806, + 0.022515330463647842, + 0.11158353090286255, + 0.007591793779283762, + 0.004441460594534874, + 0.0127238929271698, + 0.008040403015911579, + 0.001967852935194969, + 0.003263920545578003, + 0.006967158056795597, + 0.00646296888589859, + 0.00336386333219707, + 0.0037206937558948994, + 0.003135364968329668, + 0.005441374611109495, + 0.0034837445709854364, + 0.0034551590215414762, + 0.00429480941966176, + 0.005383404437452555, + 0.00359466508962214, + 0.0035193238873034716, + 0.003505441127344966, + 0.003013217356055975, + 0.0034581637009978294, + 0.004942703526467085, + 0.004475090652704239, + 0.00859139859676361, + 0.004647755064070225, + 0.005212307441979647, + 0.005871057976037264, + 0.009204655885696411, + 0.010510450229048729, + 0.004054136574268341, + 0.0069100577384233475, + 0.005365727469325066, + 0.01115370448678732, + 0.013783179223537445, + 0.010246014222502708, + 0.015517810359597206, + 0.022264378145337105, + 0.08042783290147781, + 0.016240492463111877, + 0.09563630074262619, + 0.005637822672724724, + 0.004949689377099276, + 0.007477034814655781, + 0.006748983170837164, + 0.005550802685320377, + 0.00640265503898263, + 0.005738334264606237, + 0.003996779210865498, + 0.0032647510524839163, + 0.0024015188682824373, + 0.0025504196528345346, + 0.003106101183220744, + 0.0022322561126202345, + 0.0019001411274075508, + 0.002077455399557948, + 0.003444325877353549, + 0.0025160801596939564, + 0.0032823672518134117, + 0.007290796842426062, + 0.008530397899448872, + 0.003292013658210635, + 0.0035598238464444876, + 0.0065612574107944965, + 0.0027395151555538177, + 0.0014637671411037445, + 0.001978191314265132, + 0.0022019934840500355, + 0.003149330383166671, + 0.0013908229302614927, + 0.0016886143712326884, + 0.0022230930626392365, + 0.001491712173447013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.023291101679205894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627766579389572 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627766579389572, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627766579389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4469218850135803, + 0.03363135829567909, + 0.03635909780859947, + 0.03791658207774162, + 0.10881825536489487, + 0.08551809936761856, + 0.07775711268186569, + 0.11232413351535797, + 0.04322214052081108, + 0.017531346529722214 + ], + "credit_map": [ + 0.05594921112060547, + 0.00961381196975708, + 0.16241124272346497, + 0.04751234129071236, + 0.015729282051324844, + 0.0192900188267231, + 0.02483418956398964, + 0.11158183217048645, + 0.003861568868160248, + 0.0028178165666759014, + 0.005859105382114649, + 0.004767259117215872, + 0.002650362206622958, + 0.005269346758723259, + 0.0043313149362802505, + 0.004074582830071449, + 0.004309686366468668, + 0.005576340481638908, + 0.003955901600420475, + 0.006367218680679798, + 0.0026475859340280294, + 0.00353005132637918, + 0.00526612950488925, + 0.004706182982772589, + 0.003375727217644453, + 0.0060798912309110165, + 0.0030556039419025183, + 0.006412004120647907, + 0.0043908823281526566, + 0.003956940490752459, + 0.0049871779046952724, + 0.005658354610204697, + 0.007574848365038633, + 0.011862984858453274, + 0.010000387206673622, + 0.02310735173523426, + 0.012531999498605728, + 0.012075519189238548, + 0.010199509561061859, + 0.02146565355360508, + 0.005801182705909014, + 0.013429033569991589, + 0.005519047845155001, + 0.008186811581254005, + 0.00888823438435793, + 0.015462993644177914, + 0.0075029004365205765, + 0.020727897062897682, + 0.013277960009872913, + 0.007958936505019665, + 0.00871032103896141, + 0.011411028914153576, + 0.008901538327336311, + 0.009815853089094162, + 0.008719383738934994, + 0.00896209105849266, + 0.011411952786147594, + 0.01429824624210596, + 0.016297506168484688, + 0.01507651712745428, + 0.013547741807997227, + 0.01682189106941223, + 0.012880261987447739, + 0.011990023776888847, + 0.005214050877839327, + 0.004417343530803919, + 0.006780230440199375, + 0.005587771534919739, + 0.004290328826755285, + 0.006135648116469383, + 0.0044260285794734955, + 0.006370734889060259, + 0.0018146095098927617, + 0.002626141533255577, + 0.002853285288438201, + 0.003444577334448695, + 0.001200970495119691, + 0.0018428441835567355, + 0.002065181266516447, + 0.0016837368020787835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.007627766579389572 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are large, fluffy snowflakes visible drifting slowly down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a quiet, night-shrouded city street corner illuminated by a warm orange cast iron street lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly pan down to reveal ancient cobblestone streets below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are brick buildings visible in the background with soft, blurred outlines and faint interior light from windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image bathed in a warm yellow filter with slightly desaturated colors and subtle film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072005033493042 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016072005033493042, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072005033493042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2085060328245163, + 0.0534205362200737, + 0.05087503790855408, + 0.05115732550621033, + 0.10905759036540985, + 0.31246450543403625, + 0.07558032125234604, + 0.03160299360752106, + 0.07023000717163086, + 0.03710561990737915 + ], + "credit_map": [ + 0.022895291447639465, + 0.007153602782636881, + 0.0704820528626442, + 0.024736149236559868, + 0.0073786950670182705, + 0.011103220283985138, + 0.015962162986397743, + 0.048794880509376526, + 0.006450220476835966, + 0.004921113606542349, + 0.01142624020576477, + 0.007296320982277393, + 0.004139810800552368, + 0.005528969690203667, + 0.006232624873518944, + 0.007425233256071806, + 0.0057090288028120995, + 0.005767283029854298, + 0.006534767337143421, + 0.010776183567941189, + 0.003891834057867527, + 0.004888209979981184, + 0.007642803248018026, + 0.005664930213242769, + 0.003913465887308121, + 0.007474901154637337, + 0.004829728044569492, + 0.004701512865722179, + 0.0042155045084655285, + 0.00791078433394432, + 0.006049961782991886, + 0.012061464600265026, + 0.011915463954210281, + 0.009607559069991112, + 0.00644126208499074, + 0.03960179165005684, + 0.007324636448174715, + 0.009412561543285847, + 0.01083375047892332, + 0.013920565135776997, + 0.01706032268702984, + 0.01568710058927536, + 0.014445730485022068, + 0.01831267960369587, + 0.026877833530306816, + 0.094502754509449, + 0.01807907596230507, + 0.10749901086091995, + 0.007831239141523838, + 0.005645886994898319, + 0.011932183988392353, + 0.009262294508516788, + 0.008429411798715591, + 0.01430905144661665, + 0.008790544234216213, + 0.00937971007078886, + 0.0033732184674590826, + 0.0033611655235290527, + 0.00441730348393321, + 0.0041334195993840694, + 0.003307766281068325, + 0.0030110774096101522, + 0.006606133189052343, + 0.003392912680283189, + 0.007101220078766346, + 0.007264681626111269, + 0.010602452792227268, + 0.01118775550276041, + 0.0067034219391644, + 0.008873521350324154, + 0.008684548549354076, + 0.00981240812689066, + 0.004096404183655977, + 0.0054528312757611275, + 0.00643386784940958, + 0.006877319887280464, + 0.002960065146908164, + 0.003051596228033304, + 0.0045183878391981125, + 0.0037151493597775698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.016072005033493042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_tensor([1935, 1935])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_tensor([1935, 1935])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0307c55332ba912334e21341e2d411d9ca667258 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_tensor([1935, 1935])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot closely follows a heavy tank as it advances at breakneck speed across a vast, muddy plain, perfectly capturing the realism of a documentary. The tank's dark olive green camouflage paint is speckled with scratches and dried mud. Its thick armor plates reflect a dull metallic sheen under the diffuse light of a gray sky. Its long gun barrel points steadily forward as it moves. Beneath it, its wide metal tracks bite fiercely into the slippery earth with each rotation, sending chunks of dark brown mud and murky water flying high to the rear and sides, creating a dynamic, powerful trail. The background is a vast, empty plain, covered in ruts of varying depths and puddles of water. Only the sparse outlines of a few dead trees can be seen on the distant horizon. The overall cold tones, dominated by the tank's military green and the dark brown of the earth, create an atmosphere of raw power, tension, and undeniable mechanical majesty.", + "seed": 8800, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_1935_seed_8800_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04005538681522012, + "vlm_logprob": -0.04005538681522012, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04005538681522012, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1313485950231552 + ], + "opposite_logprob_tokens": [ + -3.2563486099243164 + ], + "logprob_reward": -0.1313485950231552, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.1313485950231552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12747544050216675, + 0.12895239889621735, + 0.14990787208080292, + 0.05868172273039818, + 0.08216334879398346, + 0.08104854077100754, + 0.04616599529981613, + 0.08854074776172638, + 0.20305778086185455, + 0.03400622308254242 + ], + "credit_map": [ + 0.01899103634059429, + 0.012034077197313309, + 0.016407154500484467, + 0.014671615324914455, + 0.012306456454098225, + 0.011781007051467896, + 0.01912832446396351, + 0.02215576358139515, + 0.022677386179566383, + 0.013936631381511688, + 0.009919548407196999, + 0.012697029858827591, + 0.01444533746689558, + 0.015431545674800873, + 0.016414986923336983, + 0.023429932072758675, + 0.012406406924128532, + 0.008339355699717999, + 0.004496019799262285, + 0.05970790237188339, + 0.03882686793804169, + 0.006690924987196922, + 0.009487269446253777, + 0.009953119792044163, + 0.00529643427580595, + 0.004297257401049137, + 0.006469326093792915, + 0.006058488041162491, + 0.007861264981329441, + 0.008456328883767128, + 0.009448976255953312, + 0.010793647728860378, + 0.008720649406313896, + 0.008772253058850765, + 0.012227353639900684, + 0.008489971049129963, + 0.01339740864932537, + 0.0141132278367877, + 0.008363047614693642, + 0.008079441264271736, + 0.007904700003564358, + 0.010954710654914379, + 0.007285767234861851, + 0.0085813133046031, + 0.012590101920068264, + 0.011966456659138203, + 0.010862423107028008, + 0.010903066024184227, + 0.0036976051051169634, + 0.004897774197161198, + 0.007017381489276886, + 0.007612919434905052, + 0.00628119520843029, + 0.004929862450808287, + 0.005536472890526056, + 0.006192783359438181, + 0.006473436951637268, + 0.005994216073304415, + 0.014220915734767914, + 0.01950337365269661, + 0.012705272063612938, + 0.009837803430855274, + 0.010285872966051102, + 0.009519862942397594, + 0.029887456446886063, + 0.014342274516820908, + 0.012947081588208675, + 0.02098226174712181, + 0.05277719348669052, + 0.03363330289721489, + 0.015206138603389263, + 0.02328205481171608, + 0.004132875707000494, + 0.002383939456194639, + 0.004377978388220072, + 0.0026145344600081444, + 0.006505335681140423, + 0.00557553768157959, + 0.003544152481481433, + 0.004871867597103119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.1313485950231552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029840324074029922 + ], + "opposite_logprob_tokens": [ + -3.5298402309417725 + ], + "logprob_reward": -0.029840324074029922, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029840324074029922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06606853753328323, + 0.09483393281698227, + 0.1263587325811386, + 0.07151807099580765, + 0.039710428565740585, + 0.15779361128807068, + 0.0649072676897049, + 0.05379357933998108, + 0.1293237805366516, + 0.1956920325756073 + ], + "credit_map": [ + 0.007512338925153017, + 0.004559487570077181, + 0.008957366459071636, + 0.008217982016503811, + 0.007396535016596317, + 0.008416075259447098, + 0.009604223072528839, + 0.011404530145227909, + 0.010528653860092163, + 0.007381669711321592, + 0.009571053087711334, + 0.01190237794071436, + 0.012970774434506893, + 0.01147958543151617, + 0.016202669590711594, + 0.014797147363424301, + 0.014673039317131042, + 0.00767784146592021, + 0.006793573498725891, + 0.04816242679953575, + 0.02434321492910385, + 0.005420513451099396, + 0.0089195491746068, + 0.01036857906728983, + 0.005943540018051863, + 0.007807411253452301, + 0.009285178035497665, + 0.00916572380810976, + 0.0128718800842762, + 0.007550985552370548, + 0.008914600126445293, + 0.009978755377233028, + 0.003636156441643834, + 0.0033412084449082613, + 0.0060532353818416595, + 0.004217518027871847, + 0.007420258596539497, + 0.00675704563036561, + 0.004183429758995771, + 0.00410157535225153, + 0.0172265712171793, + 0.020377375185489655, + 0.019794194027781487, + 0.014090865850448608, + 0.03178403899073601, + 0.023951515555381775, + 0.015995241701602936, + 0.014573811553418636, + 0.007486072834581137, + 0.005274253431707621, + 0.005736994091421366, + 0.011605171486735344, + 0.010341845452785492, + 0.0061373235657811165, + 0.007218890357762575, + 0.011106722056865692, + 0.005375231150537729, + 0.0037312896456569433, + 0.0060935793444514275, + 0.01569899171590805, + 0.006078950595110655, + 0.005843697115778923, + 0.004985498730093241, + 0.005986340809613466, + 0.020399728789925575, + 0.009999292902648449, + 0.009840482845902443, + 0.010490276850759983, + 0.036164790391922, + 0.020886186510324478, + 0.009217094630002975, + 0.01232592947781086, + 0.013229692354798317, + 0.010912705212831497, + 0.028118398040533066, + 0.020789891481399536, + 0.05403017997741699, + 0.030791783705353737, + 0.019736697897315025, + 0.018082687631249428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.029840324074029922 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410939147695899 + ], + "opposite_logprob_tokens": [ + -5.5041093826293945 + ], + "logprob_reward": -0.00410939147695899, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410939147695899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1261395663022995, + 0.08013367652893066, + 0.14250338077545166, + 0.04154520854353905, + 0.05391335487365723, + 0.03392665460705757, + 0.0525597408413887, + 0.044287923723459244, + 0.39195603132247925, + 0.033034469932317734 + ], + "credit_map": [ + 0.012525264173746109, + 0.010821252129971981, + 0.016469020396471024, + 0.01854170486330986, + 0.01890898309648037, + 0.012224341742694378, + 0.018659109249711037, + 0.017989899963140488, + 0.010669419541954994, + 0.008417698554694653, + 0.008753744885325432, + 0.007085571531206369, + 0.012862038798630238, + 0.009155727922916412, + 0.013864071108400822, + 0.009325404651463032, + 0.013197177089750767, + 0.010304749943315983, + 0.007009746041148901, + 0.05199417099356651, + 0.030832795426249504, + 0.007638176437467337, + 0.011532565578818321, + 0.009993993677198887, + 0.003065994707867503, + 0.005414797458797693, + 0.005587924737483263, + 0.005775128491222858, + 0.004445235710591078, + 0.006038854829967022, + 0.00616172282025218, + 0.005055552814155817, + 0.005565110128372908, + 0.004334840923547745, + 0.010103142820298672, + 0.00560238491743803, + 0.005914718843996525, + 0.012302854098379612, + 0.0054071578197181225, + 0.004683143924921751, + 0.0038710644003003836, + 0.004833298735320568, + 0.004173622466623783, + 0.0036699341144412756, + 0.005086157936602831, + 0.0040034279227256775, + 0.003846194362267852, + 0.004442955832928419, + 0.004137209616601467, + 0.005496785510331392, + 0.007866316474974155, + 0.006762136705219746, + 0.008269689977169037, + 0.0044614882208406925, + 0.007911927998065948, + 0.007654185872524977, + 0.0035320285242050886, + 0.002740544266998768, + 0.0069884201511740685, + 0.0064780013635754585, + 0.0062324837781488895, + 0.005049636587500572, + 0.005558026488870382, + 0.007708779536187649, + 0.06847663223743439, + 0.03250345587730408, + 0.027615346014499664, + 0.03280407190322876, + 0.12454815208911896, + 0.04984793812036514, + 0.02616373635828495, + 0.02999669313430786, + 0.0025394491385668516, + 0.0022958156187087297, + 0.005680437665432692, + 0.0038730415981262922, + 0.0060914368368685246, + 0.003771335817873478, + 0.004852979443967342, + 0.003929974045604467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.00410939147695899 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642413653433323 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642413653433323, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642413653433323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1143544390797615, + 0.06177115440368652, + 0.09333236515522003, + 0.08176656067371368, + 0.10742805898189545, + 0.10633888840675354, + 0.04598863422870636, + 0.057406291365623474, + 0.28490686416625977, + 0.046706780791282654 + ], + "credit_map": [ + 0.012806148268282413, + 0.010171692818403244, + 0.018315091729164124, + 0.015483248978853226, + 0.009892711415886879, + 0.011586575768887997, + 0.014952494762837887, + 0.02114647440612316, + 0.00904419831931591, + 0.006125280167907476, + 0.00695881899446249, + 0.007412446662783623, + 0.007776000536978245, + 0.007768626790493727, + 0.00861288607120514, + 0.008072898723185062, + 0.0071057165041565895, + 0.006656699348241091, + 0.003366701304912567, + 0.03143785521388054, + 0.021381525322794914, + 0.005447033792734146, + 0.011182727292180061, + 0.0067541045136749744, + 0.007043180987238884, + 0.009873427450656891, + 0.01097725797444582, + 0.010751607827842236, + 0.007916470058262348, + 0.008341257460415363, + 0.015254427678883076, + 0.011608924716711044, + 0.009454102255403996, + 0.009092715568840504, + 0.01633669249713421, + 0.01093900203704834, + 0.012798829935491085, + 0.014668471179902554, + 0.014448653906583786, + 0.019689584150910378, + 0.011126893572509289, + 0.014312732964754105, + 0.011189645156264305, + 0.00931375753134489, + 0.022454943507909775, + 0.016446612775325775, + 0.011192924343049526, + 0.010301378555595875, + 0.004092705901712179, + 0.0047949207946658134, + 0.007457468193024397, + 0.006613994017243385, + 0.006872720550745726, + 0.004566498100757599, + 0.005374191794544458, + 0.006216134410351515, + 0.005213585216552019, + 0.0027201268821954727, + 0.006900334265083075, + 0.017261745408177376, + 0.007382072973996401, + 0.004890024196356535, + 0.0056585888378322124, + 0.007379814982414246, + 0.0483984611928463, + 0.01925925724208355, + 0.01930220238864422, + 0.026062555611133575, + 0.0867842510342598, + 0.036085546016693115, + 0.019209736958146095, + 0.02980484813451767, + 0.002667604712769389, + 0.002155739814043045, + 0.010973244905471802, + 0.005034209229052067, + 0.009187028743326664, + 0.004944982472807169, + 0.006009805016219616, + 0.005734169390052557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.008642413653433323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633620984852314 + ], + "opposite_logprob_tokens": [ + -3.651336193084717 + ], + "logprob_reward": -0.02633620984852314, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633620984852314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15205907821655273, + 0.1849256157875061, + 0.113250233232975, + 0.07229296118021011, + 0.08495254069566727, + 0.05806046724319458, + 0.08349107950925827, + 0.09639932215213776, + 0.08812762796878815, + 0.0664411261677742 + ], + "credit_map": [ + 0.01904456317424774, + 0.012677637860178947, + 0.022915516048669815, + 0.015135065652430058, + 0.018746836110949516, + 0.01873362809419632, + 0.024462642148137093, + 0.02034318447113037, + 0.019076678901910782, + 0.012300550006330013, + 0.01728096976876259, + 0.023874077945947647, + 0.035749226808547974, + 0.025255976244807243, + 0.032100144773721695, + 0.01928798481822014, + 0.01167283020913601, + 0.009040317498147488, + 0.00692368671298027, + 0.03711532801389694, + 0.02168649435043335, + 0.008102393709123135, + 0.010868688113987446, + 0.007840498350560665, + 0.006733466871082783, + 0.006609756499528885, + 0.008560121059417725, + 0.009040364064276218, + 0.008785407058894634, + 0.00991823710501194, + 0.01192531082779169, + 0.010720295831561089, + 0.008122761733829975, + 0.009534371085464954, + 0.010871522128582, + 0.010171576403081417, + 0.014364177361130714, + 0.011801077984273434, + 0.011231296695768833, + 0.008855752646923065, + 0.006780988536775112, + 0.005277832038700581, + 0.0069175781682133675, + 0.0042465743608772755, + 0.011469314806163311, + 0.009345781989395618, + 0.007005689200013876, + 0.007016708608716726, + 0.006027807015925646, + 0.007893433794379234, + 0.00859066192060709, + 0.0114320432767272, + 0.014156538061797619, + 0.006917196325957775, + 0.013902278617024422, + 0.014571116305887699, + 0.00893731601536274, + 0.004397855140268803, + 0.009209036827087402, + 0.038055844604969025, + 0.012081131339073181, + 0.0036792149767279625, + 0.009449399076402187, + 0.010589515790343285, + 0.012281075119972229, + 0.006112813949584961, + 0.007162858732044697, + 0.0062258280813694, + 0.018244924023747444, + 0.020392341539263725, + 0.007006396539509296, + 0.01070138905197382, + 0.003899402916431427, + 0.003874394344165921, + 0.01347351260483265, + 0.010087311267852783, + 0.008788960054516792, + 0.005869169719517231, + 0.01066796574741602, + 0.009780408814549446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02633620984852314 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot closely follows a heavy tank as it advances at breakneck speed across a vast, muddy plain, perfectly capturing the realism of a documentary. The tank's dark olive green camouflage paint is speckled with scratches and dried mud. Its thick armor plates reflect a dull metallic sheen under the diffuse light of a gray sky. Its long gun barrel points steadily forward as it moves. Beneath it, its wide metal tracks bite fiercely into the slippery earth with each rotation, sending chunks of dark brown mud and murky water flying high to the rear and sides, creating a dynamic, powerful trail. The background is a vast, empty plain, covered in ruts of varying depths and puddles of water. Only the sparse outlines of a few dead trees can be seen on the distant horizon. The overall cold tones, dominated by the tank's military green and the dark brown of the earth, create an atmosphere of raw power, tension, and undeniable mechanical majesty.", + "seed": 8801, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_1935_seed_8801_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.046054681623354556, + "vlm_logprob": -0.046054681623354556, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.046054681623354556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1391076147556305 + ], + "opposite_logprob_tokens": [ + -4.514107704162598 + ], + "logprob_reward": -0.1391076147556305, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.1391076147556305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12470650672912598, + 0.07903111726045609, + 0.03168793022632599, + 0.1719897836446762, + 0.1679011881351471, + 0.12152767181396484, + 0.04195536673069, + 0.03843243420124054, + 0.14812849462032318, + 0.07463956624269485 + ], + "credit_map": [ + 0.013648930005729198, + 0.01295360829681158, + 0.025779295712709427, + 0.025748038664460182, + 0.007728546857833862, + 0.009356948547065258, + 0.013587637804448605, + 0.015903495252132416, + 0.013634984381496906, + 0.013197928667068481, + 0.010411322116851807, + 0.0082929115742445, + 0.00932014174759388, + 0.007764340378344059, + 0.007979162968695164, + 0.008430317044258118, + 0.003699270775541663, + 0.003536304458975792, + 0.004704954568296671, + 0.0035254063550382853, + 0.003908324986696243, + 0.0026956575457006693, + 0.00493961526080966, + 0.0046783932484686375, + 0.02763514220714569, + 0.04895317927002907, + 0.009130527265369892, + 0.05124910920858383, + 0.007116403896361589, + 0.0095661710947752, + 0.007179604843258858, + 0.011159644462168217, + 0.018899235874414444, + 0.020715398713946342, + 0.022948751226067543, + 0.024115202948451042, + 0.024843621999025345, + 0.02210346609354019, + 0.019868142902851105, + 0.014407343231141567, + 0.015077789314091206, + 0.01459270529448986, + 0.031209072098135948, + 0.01561509445309639, + 0.009359193965792656, + 0.01160645205527544, + 0.013187906704843044, + 0.010879451408982277, + 0.004720691591501236, + 0.0035439173225313425, + 0.009368524886667728, + 0.007898587733507156, + 0.004490676335990429, + 0.0030448210891336203, + 0.0034374897368252277, + 0.005450656171888113, + 0.004316613078117371, + 0.004894360434263945, + 0.005172785371541977, + 0.007899628020823002, + 0.003768246155232191, + 0.0036289191339164972, + 0.005589565262198448, + 0.003162315348163247, + 0.015259651467204094, + 0.010913880541920662, + 0.027484990656375885, + 0.01828370802104473, + 0.017225785180926323, + 0.0066644358448684216, + 0.037448205053806305, + 0.014847817830741405, + 0.006014596205204725, + 0.00709033478051424, + 0.01021028496325016, + 0.012681445106863976, + 0.009944128803908825, + 0.006214182823896408, + 0.011594104580581188, + 0.010890488512814045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.1391076147556305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498058721423149 + ], + "opposite_logprob_tokens": [ + -2.929980516433716 + ], + "logprob_reward": -0.05498058721423149, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498058721423149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12868060171604156, + 0.04946436733007431, + 0.03625503554940224, + 0.43393197655677795, + 0.16530217230319977, + 0.018483156338334084, + 0.028484726324677467, + 0.040841154754161835, + 0.04973255842924118, + 0.0488242469727993 + ], + "credit_map": [ + 0.013792592100799084, + 0.01191773172467947, + 0.025999153032898903, + 0.02136186696588993, + 0.013123041950166225, + 0.014230594038963318, + 0.012363591231405735, + 0.0158920306712389, + 0.006152203306555748, + 0.006842572242021561, + 0.0076494477689266205, + 0.007887318730354309, + 0.005907565820962191, + 0.004192434251308441, + 0.004776359535753727, + 0.00605646800249815, + 0.006812707055360079, + 0.004918801132589579, + 0.004949189722537994, + 0.003910730592906475, + 0.005624637473374605, + 0.003663809271529317, + 0.0033677818719297647, + 0.0030073807574808598, + 0.0524444505572319, + 0.048733942210674286, + 0.034647632390260696, + 0.14063136279582977, + 0.020317135378718376, + 0.052787620574235916, + 0.02553226239979267, + 0.058837562799453735, + 0.020015347748994827, + 0.01938972808420658, + 0.02548140473663807, + 0.022075502201914787, + 0.025220345705747604, + 0.022789467126131058, + 0.016435623168945312, + 0.013894760049879551, + 0.0022080738563090563, + 0.002180898329243064, + 0.002291949698701501, + 0.003095271997153759, + 0.0025178827345371246, + 0.0018720560474321246, + 0.0025682304985821247, + 0.0017487917793914676, + 0.002605155110359192, + 0.003087723860517144, + 0.0027730101719498634, + 0.003909010905772448, + 0.006865554954856634, + 0.0031817303970456123, + 0.0035375445149838924, + 0.0025249975733458996, + 0.005191470496356487, + 0.005818136502057314, + 0.008118696510791779, + 0.00713912071660161, + 0.0037170362193137407, + 0.0030673188157379627, + 0.004376532044261694, + 0.003412848338484764, + 0.005819414276629686, + 0.004502581898123026, + 0.0077676051296293736, + 0.006679947953671217, + 0.004471611697226763, + 0.0022436759900301695, + 0.013980518095195293, + 0.004267201293259859, + 0.004327958915382624, + 0.0037659972440451384, + 0.007295829709619284, + 0.01023754570633173, + 0.005206763278692961, + 0.003037799149751663, + 0.00662996806204319, + 0.00832238420844078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.05498058721423149 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032052830792963505 + ], + "opposite_logprob_tokens": [ + -5.753205299377441 + ], + "logprob_reward": -0.0032052830792963505, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032052830792963505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09441573172807693, + 0.10635977983474731, + 0.07417702674865723, + 0.15602095425128937, + 0.04455343633890152, + 0.14710339903831482, + 0.08598827570676804, + 0.09312842786312103, + 0.09349337220191956, + 0.10475960373878479 + ], + "credit_map": [ + 0.010828418657183647, + 0.007848828099668026, + 0.016869425773620605, + 0.01688295044004917, + 0.008555775508284569, + 0.009014889597892761, + 0.010736700147390366, + 0.013678744435310364, + 0.01712608151137829, + 0.012539105489850044, + 0.017262538895010948, + 0.013207520358264446, + 0.010464268736541271, + 0.011534548364579678, + 0.010150752030313015, + 0.014074962586164474, + 0.014370408840477467, + 0.009630456566810608, + 0.01014962699264288, + 0.00752796558663249, + 0.0114746717736125, + 0.007774537894874811, + 0.007788797374814749, + 0.005460557993501425, + 0.03921790048480034, + 0.04944947734475136, + 0.008098862133920193, + 0.018726671114563942, + 0.008188443258404732, + 0.019164299592375755, + 0.006935023237019777, + 0.006240278948098421, + 0.005506616085767746, + 0.005237048026174307, + 0.00609883526340127, + 0.005857976619154215, + 0.005633292254060507, + 0.006743381265550852, + 0.005346902180463076, + 0.004129385109990835, + 0.017937330529093742, + 0.01760794408619404, + 0.03150865063071251, + 0.019926514476537704, + 0.013763990253210068, + 0.01588890701532364, + 0.016675880178809166, + 0.01379417721182108, + 0.010600936599075794, + 0.00868242047727108, + 0.015264037996530533, + 0.015101930126547813, + 0.009350680746138096, + 0.006283382885158062, + 0.009743930771946907, + 0.010960955172777176, + 0.010538853704929352, + 0.012143691070377827, + 0.018259935081005096, + 0.02284078113734722, + 0.0069085657596588135, + 0.0056815482676029205, + 0.010267849080264568, + 0.006487202364951372, + 0.011604161001741886, + 0.014484313316643238, + 0.017924930900335312, + 0.010684025473892689, + 0.01057019829750061, + 0.008283540606498718, + 0.011050441302359104, + 0.008891759440302849, + 0.00745342206209898, + 0.008359936065971851, + 0.014917399734258652, + 0.019727027043700218, + 0.013242164626717567, + 0.007753404788672924, + 0.01964285969734192, + 0.013663394376635551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.0032052830792963505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031935188453644514 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031935188453644514, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031935188453644514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1177840456366539, + 0.16926376521587372, + 0.05127793177962303, + 0.10814468562602997, + 0.13817870616912842, + 0.11506294459104538, + 0.025414850562810898, + 0.06329841166734695, + 0.0804210677742958, + 0.13115353882312775 + ], + "credit_map": [ + 0.013576955534517765, + 0.008578881621360779, + 0.02673439122736454, + 0.018899085000157356, + 0.011760085821151733, + 0.011434875428676605, + 0.011368738487362862, + 0.015431038103997707, + 0.027275722473859787, + 0.024756282567977905, + 0.02307218126952648, + 0.015957022085785866, + 0.018197689205408096, + 0.018291188403964043, + 0.013176297768950462, + 0.028537394478917122, + 0.010020429268479347, + 0.004597572144120932, + 0.007762317545711994, + 0.0064034368842840195, + 0.005660850554704666, + 0.005592318717390299, + 0.006036386359483004, + 0.0052046203054487705, + 0.0131111154332757, + 0.02412862703204155, + 0.006495590787380934, + 0.03265482559800148, + 0.005465992260724306, + 0.007831268012523651, + 0.004827000666409731, + 0.013630268163979053, + 0.01660463772714138, + 0.019075077027082443, + 0.018778547644615173, + 0.018498457968235016, + 0.02205681800842285, + 0.018232587724924088, + 0.013492394238710403, + 0.01144019141793251, + 0.013945522718131542, + 0.011750761419534683, + 0.025615308433771133, + 0.01568755879998207, + 0.010823868215084076, + 0.012116238474845886, + 0.013750476762652397, + 0.011373210698366165, + 0.0027537692803889513, + 0.002342278603464365, + 0.00555800087749958, + 0.004736783914268017, + 0.0026237405836582184, + 0.0025292716454714537, + 0.0021624653600156307, + 0.0027085391338914633, + 0.006950980052351952, + 0.007231600116938353, + 0.013300769962370396, + 0.008215134963393211, + 0.004172006156295538, + 0.004902771208435297, + 0.012799413874745369, + 0.005725738126784563, + 0.010529257357120514, + 0.008251163177192211, + 0.01453639380633831, + 0.014014851301908493, + 0.007580587640404701, + 0.004998611751943827, + 0.010013938881456852, + 0.010496264323592186, + 0.01074980292469263, + 0.013845999725162983, + 0.016049783676862717, + 0.022150270640850067, + 0.015543201006948948, + 0.009215356782078743, + 0.026018934324383736, + 0.017580203711986542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.0031935188453644514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029786404222249985 + ], + "opposite_logprob_tokens": [ + -3.5297863483428955 + ], + "logprob_reward": -0.029786404222249985, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029786404222249985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11929295212030411, + 0.16899463534355164, + 0.0742344781756401, + 0.11755052953958511, + 0.09186960756778717, + 0.12456535547971725, + 0.01594855822622776, + 0.07719630748033524, + 0.138233944773674, + 0.07211364060640335 + ], + "credit_map": [ + 0.014352506957948208, + 0.010634238831698895, + 0.0340263694524765, + 0.01892818696796894, + 0.006091775372624397, + 0.01179872639477253, + 0.007521003019064665, + 0.015940146520733833, + 0.01962617039680481, + 0.024006642401218414, + 0.025482336059212685, + 0.02085215225815773, + 0.02328123338520527, + 0.01622343249619007, + 0.015381810255348682, + 0.024140866473317146, + 0.014632217586040497, + 0.007033222820609808, + 0.011678053066134453, + 0.00978451780974865, + 0.008670180104672909, + 0.007838664576411247, + 0.008252755738794804, + 0.006344863213598728, + 0.015888245776295662, + 0.015897247940301895, + 0.009003622457385063, + 0.037904225289821625, + 0.0035265691112726927, + 0.013718648813664913, + 0.009023897349834442, + 0.012588082812726498, + 0.01011002715677023, + 0.011504759080708027, + 0.011441930197179317, + 0.012042808346450329, + 0.014307592995464802, + 0.012709992937743664, + 0.011627637781202793, + 0.008124863728880882, + 0.016757166013121605, + 0.015051910653710365, + 0.02346416376531124, + 0.01669515296816826, + 0.010769077576696873, + 0.013032064773142338, + 0.01572851464152336, + 0.013067302294075489, + 0.0016023953212425113, + 0.0011211881646886468, + 0.0030390045139938593, + 0.0026802043430507183, + 0.0014461243990808725, + 0.0015444669406861067, + 0.002620122628286481, + 0.0018950516823679209, + 0.008215462788939476, + 0.009943430311977863, + 0.015944788232445717, + 0.01750251278281212, + 0.004970467649400234, + 0.004278853069990873, + 0.010477275587618351, + 0.005863513797521591, + 0.016124030575156212, + 0.01668042689561844, + 0.02422988787293434, + 0.014043162576854229, + 0.018056968227028847, + 0.007200844585895538, + 0.02920473739504814, + 0.012693884782493114, + 0.005045617464929819, + 0.005360850133001804, + 0.013584052212536335, + 0.01580173894762993, + 0.007176227867603302, + 0.004537547472864389, + 0.00873114075511694, + 0.011876462027430534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.029786404222249985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot closely follows a heavy tank as it advances at breakneck speed across a vast, muddy plain, perfectly capturing the realism of a documentary. The tank's dark olive green camouflage paint is speckled with scratches and dried mud. Its thick armor plates reflect a dull metallic sheen under the diffuse light of a gray sky. Its long gun barrel points steadily forward as it moves. Beneath it, its wide metal tracks bite fiercely into the slippery earth with each rotation, sending chunks of dark brown mud and murky water flying high to the rear and sides, creating a dynamic, powerful trail. The background is a vast, empty plain, covered in ruts of varying depths and puddles of water. Only the sparse outlines of a few dead trees can be seen on the distant horizon. The overall cold tones, dominated by the tank's military green and the dark brown of the earth, create an atmosphere of raw power, tension, and undeniable mechanical majesty.", + "seed": 8802, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_1935_seed_8802_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023825705354101955, + "vlm_logprob": -0.023825705354101955, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.023825705354101955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08460620790719986 + ], + "opposite_logprob_tokens": [ + -3.584606170654297 + ], + "logprob_reward": -0.08460620790719986, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.08460620790719986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27251049876213074, + 0.035581573843955994, + 0.07765227556228638, + 0.04097110033035278, + 0.055258866399526596, + 0.02856610156595707, + 0.06932442635297775, + 0.042883265763521194, + 0.32822898030281067, + 0.049022939056158066 + ], + "credit_map": [ + 0.026119016110897064, + 0.07784966379404068, + 0.09095869958400726, + 0.02142854779958725, + 0.01573353074491024, + 0.008834208361804485, + 0.01982843689620495, + 0.011758374981582165, + 0.004251453094184399, + 0.0029364274814724922, + 0.00641700578853488, + 0.0038959889207035303, + 0.005102487746626139, + 0.002890424570068717, + 0.005522873252630234, + 0.004564912524074316, + 0.0077570355497300625, + 0.0049367183819413185, + 0.015867292881011963, + 0.010750677436590195, + 0.006213763263076544, + 0.005193616729229689, + 0.011943627148866653, + 0.014989539049565792, + 0.0029922451358288527, + 0.002394063165411353, + 0.005519269499927759, + 0.004901069216430187, + 0.0082932747900486, + 0.004346335772424936, + 0.0060686832293868065, + 0.006456159986555576, + 0.006222858559340239, + 0.0036606674548238516, + 0.004904423374682665, + 0.014234707690775394, + 0.004823051858693361, + 0.004337085876613855, + 0.007898025214672089, + 0.009178047999739647, + 0.002249964978545904, + 0.0016636200016364455, + 0.0030390326865017414, + 0.008527331054210663, + 0.00388343702070415, + 0.0022773139644414186, + 0.003665495663881302, + 0.003259906079620123, + 0.004763314966112375, + 0.0028243863489478827, + 0.006902111228555441, + 0.02671809308230877, + 0.003837076248601079, + 0.0025116950273513794, + 0.015993159264326096, + 0.00577459204941988, + 0.0036906260065734386, + 0.0029752403497695923, + 0.0045043788850307465, + 0.007216752041131258, + 0.005548832472413778, + 0.0039086975157260895, + 0.008408366702497005, + 0.006630370859056711, + 0.04764226824045181, + 0.011801269836723804, + 0.01709553971886635, + 0.1307288557291031, + 0.03006940707564354, + 0.009849307127296925, + 0.043773867189884186, + 0.03726845607161522, + 0.0067969318479299545, + 0.00475619500502944, + 0.011098652146756649, + 0.0057298000901937485, + 0.005207160487771034, + 0.0032620886340737343, + 0.006731983739882708, + 0.005440124776214361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.08460620790719986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821235567331314 + ], + "opposite_logprob_tokens": [ + -4.01821231842041 + ], + "logprob_reward": -0.01821235567331314, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821235567331314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4233326315879822, + 0.11186540871858597, + 0.07006087899208069, + 0.0824340283870697, + 0.09054823219776154, + 0.036623403429985046, + 0.021861892193555832, + 0.025225577875971794, + 0.1130203902721405, + 0.02502751536667347 + ], + "credit_map": [ + 0.05120350420475006, + 0.13440659642219543, + 0.09762398898601532, + 0.030160384252667427, + 0.03594443202018738, + 0.020315950736403465, + 0.02449970319867134, + 0.029178094118833542, + 0.009710259735584259, + 0.006730182096362114, + 0.015091548673808575, + 0.008253252133727074, + 0.025068579241633415, + 0.01665024273097515, + 0.013874826952815056, + 0.016486521810293198, + 0.007764520589262247, + 0.0052406364120543, + 0.007311791647225618, + 0.009479012340307236, + 0.004965886473655701, + 0.0033215489238500595, + 0.00951597560197115, + 0.022461507469415665, + 0.0048838150687515736, + 0.003382785012945533, + 0.008546308614313602, + 0.011463700793683529, + 0.012491933070123196, + 0.014432867057621479, + 0.016815997660160065, + 0.010416623204946518, + 0.004316144622862339, + 0.0047233267687261105, + 0.004781687632203102, + 0.03493799269199371, + 0.00770356273278594, + 0.007507102098315954, + 0.015072125010192394, + 0.011506300419569016, + 0.0035360122565180063, + 0.0016159794759005308, + 0.004092691000550985, + 0.00780832301825285, + 0.004916978534311056, + 0.0070352209731936455, + 0.002672677394002676, + 0.004945523105561733, + 0.001483971020206809, + 0.000955200579483062, + 0.001971825957298279, + 0.008293629623949528, + 0.001827811705879867, + 0.0009226109832525253, + 0.0049899970181286335, + 0.001416848273947835, + 0.0025585629045963287, + 0.0014463030965998769, + 0.0025866213254630566, + 0.005092239007353783, + 0.0022216392681002617, + 0.002369948662817478, + 0.004561282694339752, + 0.0043889801017940044, + 0.012312053702771664, + 0.004137265495955944, + 0.005383489187806845, + 0.04297780245542526, + 0.013595261611044407, + 0.005424791481345892, + 0.014421159401535988, + 0.014768578112125397, + 0.002446985337883234, + 0.0018291883170604706, + 0.005223995074629784, + 0.003382989903911948, + 0.003966866992413998, + 0.001890560844913125, + 0.002988521009683609, + 0.0032984090503305197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.01821235567331314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019497923785820603 + ], + "opposite_logprob_tokens": [ + -6.251949787139893 + ], + "logprob_reward": -0.0019497923785820603, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019497923785820603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37317025661468506, + 0.06663629412651062, + 0.06741773337125778, + 0.04900747910141945, + 0.04066244140267372, + 0.01698835752904415, + 0.04823013022542, + 0.06767883151769638, + 0.07151933014392853, + 0.19868916273117065 + ], + "credit_map": [ + 0.034002501517534256, + 0.10468801856040955, + 0.12634143233299255, + 0.03499622270464897, + 0.020001081749796867, + 0.012636794708669186, + 0.02063535340130329, + 0.01986883208155632, + 0.012316901236772537, + 0.007424455601722002, + 0.012196154333651066, + 0.00631288206204772, + 0.006867389660328627, + 0.005039506126195192, + 0.007894551381468773, + 0.008584453724324703, + 0.008321387693285942, + 0.005399389658123255, + 0.01159022655338049, + 0.008972631767392159, + 0.007571561262011528, + 0.004745388403534889, + 0.013692882843315601, + 0.007124266121536493, + 0.0032073850743472576, + 0.0023511608596891165, + 0.00689556123688817, + 0.005312222987413406, + 0.008752395398914814, + 0.006087856832891703, + 0.008319186046719551, + 0.00808170810341835, + 0.005136523395776749, + 0.0028342369478195906, + 0.0029071024619042873, + 0.006680980324745178, + 0.005482386332005262, + 0.004714382812380791, + 0.004122696351259947, + 0.008784130215644836, + 0.001456319005228579, + 0.0010991727467626333, + 0.0014886785065755248, + 0.0021294911857694387, + 0.003379564732313156, + 0.0027344597037881613, + 0.0022680810652673244, + 0.002432590816169977, + 0.003626440651714802, + 0.0022667015437036753, + 0.005376373417675495, + 0.018302619457244873, + 0.0043914299458265305, + 0.0021306020207703114, + 0.008263473398983479, + 0.0038724890910089016, + 0.007836260832846165, + 0.004875220824033022, + 0.007239185739308596, + 0.009633713401854038, + 0.010252047330141068, + 0.0064926459454, + 0.012348167598247528, + 0.009001593105494976, + 0.006247640121728182, + 0.00296388054266572, + 0.005613795481622219, + 0.013527827337384224, + 0.011929257772862911, + 0.006577059626579285, + 0.011373595334589481, + 0.013286269269883633, + 0.03476160019636154, + 0.01978113502264023, + 0.047499995678663254, + 0.023235715925693512, + 0.017112813889980316, + 0.013014320284128189, + 0.022004596889019012, + 0.021278992295265198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0019497923785820603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673760985955596 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.00673760985955596, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673760985955596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14356118440628052, + 0.05895321071147919, + 0.08150269836187363, + 0.03982441499829292, + 0.20685997605323792, + 0.028105221688747406, + 0.030715761706233025, + 0.04583461582660675, + 0.10575545579195023, + 0.2588874101638794 + ], + "credit_map": [ + 0.019393231719732285, + 0.04419850930571556, + 0.01777924783527851, + 0.01050107553601265, + 0.009429430589079857, + 0.008855331689119339, + 0.012272740714251995, + 0.02113163098692894, + 0.007491336204111576, + 0.005808166228234768, + 0.011259222403168678, + 0.006015291437506676, + 0.006149980705231428, + 0.005411066580563784, + 0.007966973818838596, + 0.008851180784404278, + 0.0046660397201776505, + 0.00408121757209301, + 0.013437759131193161, + 0.00844167172908783, + 0.007270217407494783, + 0.004859365057200193, + 0.023678049445152283, + 0.01506837923079729, + 0.0028740079142153263, + 0.0027869215700775385, + 0.0031407191418111324, + 0.005529030226171017, + 0.009557703509926796, + 0.004429904744029045, + 0.0050840661861002445, + 0.006422064267098904, + 0.010663172230124474, + 0.010684028267860413, + 0.013351053930819035, + 0.08053462207317352, + 0.016323044896125793, + 0.011945471167564392, + 0.04144243523478508, + 0.021916160359978676, + 0.0032741574104875326, + 0.0018190708942711353, + 0.002263690810650587, + 0.00609979871660471, + 0.004267825745046139, + 0.003563659032806754, + 0.003546880092471838, + 0.003270140616223216, + 0.0020211932715028524, + 0.0012632649159058928, + 0.002702758414670825, + 0.012514237314462662, + 0.0020792456343770027, + 0.0012663837987929583, + 0.006028612144291401, + 0.0028400688897818327, + 0.003245230298489332, + 0.0025698370300233364, + 0.004298544488847256, + 0.0069651915691792965, + 0.008403480984270573, + 0.003214358352124691, + 0.009223802015185356, + 0.007914177142083645, + 0.016653243452310562, + 0.004074770025908947, + 0.005823350045830011, + 0.039579685777425766, + 0.008148503489792347, + 0.004413462709635496, + 0.01462897751480341, + 0.01243347767740488, + 0.041507262736558914, + 0.02501632459461689, + 0.053232405334711075, + 0.028828945010900497, + 0.03661995753645897, + 0.01806451752781868, + 0.029892468824982643, + 0.025725573301315308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.00673760985955596 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622560951858759 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622560951858759, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622560951858759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09947598725557327, + 0.05417373403906822, + 0.0495007261633873, + 0.029246140271425247, + 0.10789980739355087, + 0.029768142849206924, + 0.014398745261132717, + 0.05955946445465088, + 0.12219598889350891, + 0.4337812066078186 + ], + "credit_map": [ + 0.005717349238693714, + 0.015183322131633759, + 0.031655170023441315, + 0.01131688803434372, + 0.007024797610938549, + 0.005539454519748688, + 0.011067710816860199, + 0.011971295811235905, + 0.010586919263005257, + 0.004820068366825581, + 0.007255356293171644, + 0.004731213673949242, + 0.0077222189866006374, + 0.005717853549867868, + 0.006437914911657572, + 0.006902188528329134, + 0.0046846335753798485, + 0.003373592859134078, + 0.007850156165659428, + 0.00574855599552393, + 0.004274428356438875, + 0.003209014656022191, + 0.010375398211181164, + 0.009984944947063923, + 0.0023341639898717403, + 0.0015338717494159937, + 0.0032992896158248186, + 0.0043062507174909115, + 0.006594845559448004, + 0.003765944391489029, + 0.0033223002683371305, + 0.004089474678039551, + 0.0050745802000164986, + 0.006487762089818716, + 0.007283070124685764, + 0.04288369044661522, + 0.009636237286031246, + 0.004939232021570206, + 0.018586182966828346, + 0.013009058311581612, + 0.0026455558836460114, + 0.0025416354183107615, + 0.0028467585798352957, + 0.004176416900008917, + 0.00668732263147831, + 0.0026333194691687822, + 0.003812818555161357, + 0.004424314945936203, + 0.0012357155792415142, + 0.0010593240149319172, + 0.0018095361301675439, + 0.0029007564298808575, + 0.0019831685349345207, + 0.0011292421258985996, + 0.0023076594807207584, + 0.001973344013094902, + 0.005070737097412348, + 0.003335541347041726, + 0.006081028841435909, + 0.009957260452210903, + 0.008679203689098358, + 0.004280592314898968, + 0.011272924020886421, + 0.010882175527513027, + 0.01871342770755291, + 0.004534165374934673, + 0.007310948334634304, + 0.04202185943722725, + 0.011293947696685791, + 0.0061757201328873634, + 0.01701168343424797, + 0.0151342349126935, + 0.04648146778345108, + 0.02966124564409256, + 0.06593523919582367, + 0.05044858530163765, + 0.10305096954107285, + 0.03399106115102768, + 0.060038018971681595, + 0.04417463019490242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.007622560951858759 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot closely follows a heavy tank as it advances at breakneck speed across a vast, muddy plain, perfectly capturing the realism of a documentary. The tank's dark olive green camouflage paint is speckled with scratches and dried mud. Its thick armor plates reflect a dull metallic sheen under the diffuse light of a gray sky. Its long gun barrel points steadily forward as it moves. Beneath it, its wide metal tracks bite fiercely into the slippery earth with each rotation, sending chunks of dark brown mud and murky water flying high to the rear and sides, creating a dynamic, powerful trail. The background is a vast, empty plain, covered in ruts of varying depths and puddles of water. Only the sparse outlines of a few dead trees can be seen on the distant horizon. The overall cold tones, dominated by the tank's military green and the dark brown of the earth, create an atmosphere of raw power, tension, and undeniable mechanical majesty.", + "seed": 8803, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_3/index_1935_seed_8803_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027280693314969538, + "vlm_logprob": -0.027280693314969538, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.027280693314969538, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10285374522209167 + ], + "opposite_logprob_tokens": [ + -4.352853775024414 + ], + "logprob_reward": -0.10285374522209167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.10285374522209167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11476177722215652, + 0.014197430573403835, + 0.02278151549398899, + 0.28615957498550415, + 0.06141381338238716, + 0.23554062843322754, + 0.02382553368806839, + 0.13384248316287994, + 0.044893089681863785, + 0.06258412450551987 + ], + "credit_map": [ + 0.012039260007441044, + 0.015370698645710945, + 0.011951536871492863, + 0.01114603690803051, + 0.01935172826051712, + 0.013666950166225433, + 0.013015864416956902, + 0.01821969635784626, + 0.0014897782821208239, + 0.0017551666824147105, + 0.001383533119224012, + 0.0019241338595747948, + 0.0018936789128929377, + 0.001579479780048132, + 0.0019486859673634171, + 0.0022229738533496857, + 0.0022593114990741014, + 0.002207550685852766, + 0.005519099999219179, + 0.003207453293725848, + 0.002820180030539632, + 0.0014830827713012695, + 0.0019831163808703423, + 0.003301720367744565, + 0.03174008056521416, + 0.0077237957157194614, + 0.12347240000963211, + 0.024277718737721443, + 0.008929642848670483, + 0.0055261338129639626, + 0.01430900301784277, + 0.07018078863620758, + 0.0044946991838514805, + 0.014044891111552715, + 0.00673007732257247, + 0.006617773324251175, + 0.011943040415644646, + 0.0061554028652608395, + 0.007109268102794886, + 0.004318667575716972, + 0.014586090110242367, + 0.024081259965896606, + 0.028060687705874443, + 0.021844256669282913, + 0.024758268147706985, + 0.048216596245765686, + 0.02451549470424652, + 0.04947797954082489, + 0.0023001322988420725, + 0.005011747125536203, + 0.0028330220375210047, + 0.0024938518181443214, + 0.002385940169915557, + 0.003086061915382743, + 0.0029058854561299086, + 0.002808893332257867, + 0.007781501393765211, + 0.010210109874606133, + 0.019154047593474388, + 0.007233431562781334, + 0.011001653969287872, + 0.025283226743340492, + 0.04229283332824707, + 0.01088566891849041, + 0.005814117379486561, + 0.00874299742281437, + 0.005255053285509348, + 0.004337443970143795, + 0.00534065393730998, + 0.0037478478625416756, + 0.005778949707746506, + 0.005876023788005114, + 0.00679380539804697, + 0.009241986088454723, + 0.00686156190931797, + 0.008402192033827305, + 0.0059534902684390545, + 0.007601174991577864, + 0.008346260525286198, + 0.00938365701586008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.10285374522209167 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212694019079208 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212694019079208, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212694019079208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11075585335493088, + 0.1621931791305542, + 0.04238176345825195, + 0.3890141248703003, + 0.05220954865217209, + 0.044820357114076614, + 0.05416891723871231, + 0.05100163444876671, + 0.03961540758609772, + 0.05383919179439545 + ], + "credit_map": [ + 0.017103806138038635, + 0.016269858926534653, + 0.011649755761027336, + 0.012403285130858421, + 0.01877419464290142, + 0.00982049759477377, + 0.011525589972734451, + 0.013208869844675064, + 0.029330281540751457, + 0.012929647229611874, + 0.008998673409223557, + 0.021004769951105118, + 0.016584321856498718, + 0.011174778454005718, + 0.013451291248202324, + 0.04871942847967148, + 0.003625518409535289, + 0.0030862209387123585, + 0.007520826067775488, + 0.006976873613893986, + 0.009790432639420033, + 0.0035352755803614855, + 0.003319593844935298, + 0.004527022130787373, + 0.039845436811447144, + 0.012533404864370823, + 0.16646939516067505, + 0.0344109982252121, + 0.014011559076607227, + 0.008037020452320576, + 0.01771494187414646, + 0.0959913432598114, + 0.003846478648483753, + 0.007489550393074751, + 0.00902661681175232, + 0.006220659241080284, + 0.004867021460086107, + 0.0064392490312457085, + 0.008241940289735794, + 0.006078034173697233, + 0.003219087840989232, + 0.009822499006986618, + 0.00602131150662899, + 0.004674395080655813, + 0.005880307871848345, + 0.003509554546326399, + 0.007150717545300722, + 0.0045424834825098515, + 0.006475885398685932, + 0.007059983443468809, + 0.007571674883365631, + 0.005489890929311514, + 0.0054664285853505135, + 0.004436032380908728, + 0.011615635827183723, + 0.006053386256098747, + 0.004341981839388609, + 0.009713837876915932, + 0.005434827413409948, + 0.005227652844041586, + 0.009880749508738518, + 0.004183575045317411, + 0.005390086676925421, + 0.006828922778367996, + 0.004261313937604427, + 0.006020740605890751, + 0.004518870264291763, + 0.003858637996017933, + 0.0049610142596066, + 0.00299597205594182, + 0.006737822666764259, + 0.006261036265641451, + 0.00575388316065073, + 0.00853011105209589, + 0.00443658884614706, + 0.00569309014827013, + 0.009945997968316078, + 0.007511850446462631, + 0.006548048462718725, + 0.005419623106718063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.014212694019079208 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002827458083629608 + ], + "opposite_logprob_tokens": [ + -5.8778276443481445 + ], + "logprob_reward": -0.002827458083629608, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002827458083629608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10714056342840195, + 0.06437884271144867, + 0.06799985468387604, + 0.3170413672924042, + 0.04379017651081085, + 0.17735715210437775, + 0.06520811468362808, + 0.08157768100500107, + 0.04031150043010712, + 0.03519474342465401 + ], + "credit_map": [ + 0.010324869304895401, + 0.016590410843491554, + 0.0100975651293993, + 0.012633467093110085, + 0.009357917122542858, + 0.015533287078142166, + 0.014003725722432137, + 0.018599318340420723, + 0.007702691480517387, + 0.0073410202749073505, + 0.00888911820948124, + 0.008232240565121174, + 0.005789697635918856, + 0.0046630678698420525, + 0.006739581003785133, + 0.015021421015262604, + 0.008238325826823711, + 0.005982609000056982, + 0.015168525278568268, + 0.00921555608510971, + 0.005850550252944231, + 0.003885596990585327, + 0.005066690035164356, + 0.0145920030772686, + 0.033738985657691956, + 0.009087620303034782, + 0.1363075077533722, + 0.02645232528448105, + 0.010253463871777058, + 0.006486320868134499, + 0.01619541086256504, + 0.07851973921060562, + 0.0029679499566555023, + 0.0064119515009224415, + 0.006465918850153685, + 0.005833637434989214, + 0.0047972239553928375, + 0.005150820594280958, + 0.007033554837107658, + 0.005129119381308556, + 0.010060331784188747, + 0.018376989290118217, + 0.01992952823638916, + 0.012765482068061829, + 0.02036062441766262, + 0.03870551288127899, + 0.019240615889430046, + 0.03791806846857071, + 0.007932042703032494, + 0.007723820861428976, + 0.0067494711838662624, + 0.005696437321603298, + 0.004697091411799192, + 0.007420357782393694, + 0.014463512226939201, + 0.010525383055210114, + 0.006521887611597776, + 0.012666256166994572, + 0.008448806591331959, + 0.006379798520356417, + 0.01314238179475069, + 0.010357722640037537, + 0.017555616796016693, + 0.006505216937512159, + 0.00503629120066762, + 0.007596801966428757, + 0.0038290515076369047, + 0.004185730125755072, + 0.0035810789559036493, + 0.004393453244119883, + 0.004917103331536055, + 0.006771990563720465, + 0.00352917960844934, + 0.0038315479177981615, + 0.0034382587764412165, + 0.004735142923891544, + 0.0027394031640142202, + 0.004656338598579168, + 0.005891189910471439, + 0.006373681593686342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.002827458083629608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673630740493536 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.00673630740493536, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673630740493536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022961750626564026, + 0.005715024191886187, + 0.009909908287227154, + 0.8108131289482117, + 0.021986978128552437, + 0.048288315534591675, + 0.032745301723480225, + 0.03079332783818245, + 0.006530501879751682, + 0.010255744680762291 + ], + "credit_map": [ + 0.0037109768018126488, + 0.003736187005415559, + 0.003252241760492325, + 0.0020755482837557793, + 0.0023440991062670946, + 0.0033362829126417637, + 0.0023550065234303474, + 0.0021514082327485085, + 0.0006382848368957639, + 0.0006646258407272398, + 0.0008852147147990763, + 0.000714181165676564, + 0.0005556087708100677, + 0.000506045704241842, + 0.0008131085778586566, + 0.0009379541734233499, + 0.0012410825584083796, + 0.0007955176406539977, + 0.0018923816969618201, + 0.0012501365272328258, + 0.0014996060635894537, + 0.0007007652893662453, + 0.0007161548128351569, + 0.0018142641056329012, + 0.09028192609548569, + 0.020222729071974754, + 0.3539862036705017, + 0.06850408017635345, + 0.024404292926192284, + 0.013479047454893589, + 0.03875479847192764, + 0.20118002593517303, + 0.001477331155911088, + 0.004023066256195307, + 0.0033626402728259563, + 0.001858759205788374, + 0.003634959924966097, + 0.0020532230846583843, + 0.003049046965315938, + 0.002527950331568718, + 0.0027623153291642666, + 0.005120008252561092, + 0.005840431898832321, + 0.005133467726409435, + 0.0050657629035413265, + 0.00992811843752861, + 0.004739874042570591, + 0.009698338806629181, + 0.0035503492690622807, + 0.0037735968362540007, + 0.004473452921956778, + 0.002238882938399911, + 0.0030520439613610506, + 0.003883051685988903, + 0.008305971510708332, + 0.003467951202765107, + 0.00235095270909369, + 0.0035503716208040714, + 0.003598592011258006, + 0.002406396670266986, + 0.00375016569159925, + 0.004146073944866657, + 0.0074785370379686356, + 0.0035122395493090153, + 0.001234024646691978, + 0.0011145650641992688, + 0.0008101394632831216, + 0.0005264354986138642, + 0.0005967889446765184, + 0.0007185047143138945, + 0.000628613110166043, + 0.0009014304378069937, + 0.0009393281070515513, + 0.0011851181043311954, + 0.0010919085470959544, + 0.0018699767533689737, + 0.0012238307390362024, + 0.0012383057037368417, + 0.0013968258863314986, + 0.001310450490564108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.00673630740493536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking the tank as it moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank painted dark olive green with visible scratches and dried mud speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tank's long gun barrel point steadily forward while it advances?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are chunks of dark brown mud and murky water flying from the tracks to the rear and sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, empty plain with ruts, puddles, and sparse dead trees on the distant horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009773261845111847 + ], + "opposite_logprob_tokens": [ + -4.634773254394531 + ], + "logprob_reward": -0.009773261845111847, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009773261845111847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04645492136478424, + 0.03650110214948654, + 0.014114823192358017, + 0.6786704659461975, + 0.030011745169758797, + 0.0246710404753685, + 0.13365581631660461, + 0.015944547951221466, + 0.009805090725421906, + 0.010170473717153072 + ], + "credit_map": [ + 0.006561432499438524, + 0.006274373736232519, + 0.005121333058923483, + 0.004447490442544222, + 0.005586098413914442, + 0.003795418655499816, + 0.011259406805038452, + 0.0034093656577169895, + 0.006911756005138159, + 0.0022810662630945444, + 0.0023800779599696398, + 0.005197446793317795, + 0.0026103376876562834, + 0.0017984587466344237, + 0.0028711992781609297, + 0.012450761161744595, + 0.0016160471132025123, + 0.0011864948319271207, + 0.0036598211154341698, + 0.0018632466671988368, + 0.0012205502716824412, + 0.0008748517720960081, + 0.0011803308734670281, + 0.002513479208573699, + 0.07610003650188446, + 0.016804780811071396, + 0.29597070813179016, + 0.05714590102434158, + 0.02042311243712902, + 0.011360753327608109, + 0.03269819915294647, + 0.16816698014736176, + 0.0016447093803435564, + 0.004575829952955246, + 0.0056327455677092075, + 0.0023024138063192368, + 0.005025152117013931, + 0.0032854443415999413, + 0.004815207794308662, + 0.0027302440721541643, + 0.0015657049370929599, + 0.0041441041976213455, + 0.0029385555535554886, + 0.001939288224093616, + 0.0038104490377008915, + 0.0028280161786824465, + 0.003358887042850256, + 0.004086036700755358, + 0.010150291956961155, + 0.013482225127518177, + 0.012897374108433723, + 0.00936880148947239, + 0.006813965272158384, + 0.020800938829779625, + 0.04622705280780792, + 0.01391516998410225, + 0.0014301170594990253, + 0.0025009382516145706, + 0.0017660715384408832, + 0.0012763526756316423, + 0.0028909880202263594, + 0.0013826257782056928, + 0.0028560757637023926, + 0.0018413792131468654, + 0.0012237406335771084, + 0.0015883378218859434, + 0.0012409205082803965, + 0.0009337136289104819, + 0.0012011812068521976, + 0.0009293826296925545, + 0.0011909033637493849, + 0.0014969103503972292, + 0.0013622369151562452, + 0.0012910026125609875, + 0.0011043043341487646, + 0.0011977937538176775, + 0.0009246697300113738, + 0.0015254960162565112, + 0.001307050115428865, + 0.0014579202979803085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.009773261845111847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_4/index_tensor([5397, 5397])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_4/index_tensor([5397, 5397])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6badb94372564f234f2a7ad5471619b8cf2af890 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_4/index_tensor([5397, 5397])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a world imbued with Eastern Zen, a close-up shot from a slightly downward angle captures a plump, emerald green cabbage plummeting with grace and resolve into a pristine earthenware pot below. The cabbage's leaves, rendered in varying shades of green ink, shift from a lush, verdant exterior to a vibrant, jade-white interior. The veins are outlined with precise, powerful lines of light ink, imbued with the texture of life. The moment the cabbage touches the pot's clear, gently simmering broth, the liquid suddenly explodes in all directions. The splashing water is artistically manipulated into a series of powerful, outward-spreading dots of thick and thin ink and vibrant, white-washed brushstrokes, capturing the peak of movement against a backdrop of rice paper-like white. The rising steam from the pot transforms into wisps of soft, flowing ink, blending seamlessly into the background. The whole picture combines movement and stillness, with the vivid emerald green becoming the only visual focus in the simple black, white and gray tones, creating a sense of purity and tranquility that is both full of instantaneous explosive power and extremely poetic and picturesque.", + "seed": 17917, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_4/index_5397_seed_17917_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15757039161399006, + "vlm_logprob": -0.15757039161399006, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.15757039161399006, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006813744083046913 + ], + "opposite_logprob_tokens": [ + -5.7568135261535645 + ], + "logprob_reward": -0.006813744083046913, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.006813744083046913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06103682890534401, + 0.027791693806648254, + 0.0398104302585125, + 0.0390956848859787, + 0.04751014709472656, + 0.045152146369218826, + 0.5137900114059448, + 0.09234663099050522, + 0.024640550836920738, + 0.10882586985826492 + ], + "credit_map": [ + 0.005255416501313448, + 0.007203457877039909, + 0.005710171535611153, + 0.007050442975014448, + 0.00706250686198473, + 0.01287059672176838, + 0.008888175711035728, + 0.006996063981205225, + 0.004563163500279188, + 0.006137654185295105, + 0.0015933437971398234, + 0.004536808002740145, + 0.002838273299857974, + 0.0020909535232931376, + 0.0035711368545889854, + 0.002460361924022436, + 0.0043648360297083855, + 0.004151333589106798, + 0.0027976003475487232, + 0.009336165152490139, + 0.003961202688515186, + 0.006798208691179752, + 0.0051030609756708145, + 0.003298024879768491, + 0.004552484955638647, + 0.0033898174297064543, + 0.004088896326720715, + 0.0070014772936701775, + 0.00390001037158072, + 0.0048644207417964935, + 0.007716268766671419, + 0.0035823092330247164, + 0.003879684954881668, + 0.004263703711330891, + 0.004351476673036814, + 0.005796987097710371, + 0.004642032086849213, + 0.007602323312312365, + 0.010133085772395134, + 0.006840854417532682, + 0.004264854826033115, + 0.004762489814311266, + 0.003924967721104622, + 0.005949521437287331, + 0.008817484602332115, + 0.00446389801800251, + 0.009729025885462761, + 0.0032399047631770372, + 0.15866436064243317, + 0.031997907906770706, + 0.024944571778178215, + 0.025285029783844948, + 0.025141263380646706, + 0.1770889163017273, + 0.032975681126117706, + 0.037692245095968246, + 0.016397641971707344, + 0.0071096098981797695, + 0.012891823425889015, + 0.009735223837196827, + 0.005650194361805916, + 0.011180147528648376, + 0.009008712135255337, + 0.020373273640871048, + 0.004893328528851271, + 0.002672591246664524, + 0.0030669141560792923, + 0.002559771528467536, + 0.0019098289776593447, + 0.0035549532622098923, + 0.0030602505430579185, + 0.0029229125939309597, + 0.013893438503146172, + 0.012917660176753998, + 0.010662449523806572, + 0.02768421731889248, + 0.007952740415930748, + 0.010061439126729965, + 0.011630400083959103, + 0.014023523777723312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.006813744083046913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577655911445618 + ], + "opposite_logprob_tokens": [ + -0.6327655911445618 + ], + "logprob_reward": -0.7577655911445618, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577655911445618, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08990778028964996, + 0.24873511493206024, + 0.08863501250743866, + 0.1604517251253128, + 0.14220258593559265, + 0.04631485417485237, + 0.12371505796909332, + 0.04882006347179413, + 0.030865950509905815, + 0.020351843908429146 + ], + "credit_map": [ + 0.00939224660396576, + 0.009831797331571579, + 0.006655953358858824, + 0.006973223760724068, + 0.010297130793333054, + 0.020894240587949753, + 0.007607437204569578, + 0.018255749717354774, + 0.06589465588331223, + 0.09782557934522629, + 0.009796500205993652, + 0.024809138849377632, + 0.01785026304423809, + 0.00916079618036747, + 0.01324525848031044, + 0.010152933187782764, + 0.009794955141842365, + 0.008995296433568, + 0.008519533090293407, + 0.01913159154355526, + 0.007950277999043465, + 0.014430978335440159, + 0.010390075854957104, + 0.0094223003834486, + 0.017816970124840736, + 0.02185947820544243, + 0.015332711860537529, + 0.0317072793841362, + 0.01416855864226818, + 0.015410149469971657, + 0.023957962170243263, + 0.02019861899316311, + 0.011056877672672272, + 0.017559172585606575, + 0.0136157451197505, + 0.01891339384019375, + 0.017386702820658684, + 0.018207911401987076, + 0.027498841285705566, + 0.017963934689760208, + 0.004345319699496031, + 0.00473593408241868, + 0.004590757191181183, + 0.008077914826571941, + 0.00628194585442543, + 0.00481807254254818, + 0.00983316358178854, + 0.0036317475605756044, + 0.031774844974279404, + 0.00797984004020691, + 0.008175634779036045, + 0.009958191774785519, + 0.006258603651076555, + 0.035597942769527435, + 0.010800599120557308, + 0.013169403187930584, + 0.008104237727820873, + 0.004179623909294605, + 0.007736196741461754, + 0.005574381910264492, + 0.002558012492954731, + 0.005841556470841169, + 0.004612939432263374, + 0.010213114321231842, + 0.004086341708898544, + 0.004283727612346411, + 0.0030944363679736853, + 0.00502786273136735, + 0.003210856346413493, + 0.0039773439057171345, + 0.0033052973449230194, + 0.0038800842594355345, + 0.002942803082987666, + 0.0017375372117385268, + 0.0020505674183368683, + 0.004314864985644817, + 0.0015324150444939733, + 0.003930871374905109, + 0.0018764599226415157, + 0.001966326031833887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.7577655911445618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016080215573310852 + ], + "opposite_logprob_tokens": [ + -4.141080379486084 + ], + "logprob_reward": -0.016080215573310852, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016080215573310852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18729166686534882, + 0.12596043944358826, + 0.08771713823080063, + 0.0992501899600029, + 0.0911453515291214, + 0.07475942373275757, + 0.13871467113494873, + 0.03339097276329994, + 0.030476784333586693, + 0.1312934011220932 + ], + "credit_map": [ + 0.017831766977906227, + 0.03501974791288376, + 0.012001064606010914, + 0.020093390718102455, + 0.03073478303849697, + 0.023686125874519348, + 0.022558510303497314, + 0.025366248562932014, + 0.031688690185546875, + 0.0450366847217083, + 0.005793687887489796, + 0.01316339336335659, + 0.010382973589003086, + 0.005028968211263418, + 0.009236259385943413, + 0.005629772786051035, + 0.009349184110760689, + 0.007408121135085821, + 0.00624378677457571, + 0.020551757887005806, + 0.009748587384819984, + 0.01414396706968546, + 0.011744234710931778, + 0.008527486585080624, + 0.011894473806023598, + 0.013160253874957561, + 0.011328373104333878, + 0.016888154670596123, + 0.01064934954047203, + 0.012038502842187881, + 0.0110486363992095, + 0.012242442928254604, + 0.0070790937170386314, + 0.009061728604137897, + 0.007190813776105642, + 0.011890225112438202, + 0.009957913309335709, + 0.012551647610962391, + 0.02129998616874218, + 0.012113932520151138, + 0.007518895901739597, + 0.007895078510046005, + 0.007642213720828295, + 0.00818429235368967, + 0.018816864117980003, + 0.008223273791372776, + 0.010043284855782986, + 0.00643550930544734, + 0.04448533058166504, + 0.008581950329244137, + 0.006140996236354113, + 0.0069152843207120895, + 0.0070494101382792, + 0.04872838035225868, + 0.008571933954954147, + 0.008241374045610428, + 0.005469965282827616, + 0.0032659138087183237, + 0.00349110784009099, + 0.00522979348897934, + 0.0021469753701239824, + 0.004468865692615509, + 0.003331944812089205, + 0.005986403673887253, + 0.005231771618127823, + 0.002445171121507883, + 0.0036964656319469213, + 0.004060601815581322, + 0.0017924680141732097, + 0.004574114456772804, + 0.003508747322484851, + 0.005167439114302397, + 0.015817658975720406, + 0.014403235167264938, + 0.014839268289506435, + 0.025798587128520012, + 0.010603910312056541, + 0.015041155740618706, + 0.014065586030483246, + 0.020723989233374596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.016080215573310852 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525260902941227 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.00525260902941227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525260902941227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025508610531687737, + 0.6169338226318359, + 0.020199907943606377, + 0.02401379495859146, + 0.04092208668589592, + 0.022784005850553513, + 0.1896737664937973, + 0.038774266839027405, + 0.0058767045848071575, + 0.01531312521547079 + ], + "credit_map": [ + 0.002435800852254033, + 0.0028686525765806437, + 0.0032333238050341606, + 0.0029140939004719257, + 0.004207998048514128, + 0.0020414027385413647, + 0.004458872601389885, + 0.003348461352288723, + 0.15104787051677704, + 0.22952647507190704, + 0.02627742849290371, + 0.06949512660503387, + 0.049892183393239975, + 0.0230965968221426, + 0.03928065672516823, + 0.028317401185631752, + 0.002262461232021451, + 0.0018766149878501892, + 0.0019795733969658613, + 0.0034574430901557207, + 0.0019288050243631005, + 0.004372545052319765, + 0.0025034525897353888, + 0.0018190082628279924, + 0.0025955894961953163, + 0.003031084081158042, + 0.002073360374197364, + 0.005209733732044697, + 0.0018797465600073338, + 0.0025227079167962074, + 0.002993819070979953, + 0.0037077509332448244, + 0.003211905714124441, + 0.004455145448446274, + 0.003886511316522956, + 0.005438534542918205, + 0.004477005451917648, + 0.00567274447530508, + 0.008493182249367237, + 0.005287050735205412, + 0.0022065099328756332, + 0.0019709232728928328, + 0.0022551603615283966, + 0.0032847682014107704, + 0.003166369628161192, + 0.001906850258819759, + 0.0059783970937132835, + 0.0020150241907685995, + 0.057265233248472214, + 0.01174630131572485, + 0.009729159064590931, + 0.010376493446528912, + 0.009136028587818146, + 0.06408757716417313, + 0.012629964388906956, + 0.014702979475259781, + 0.00700347451493144, + 0.0030421435367316008, + 0.005377201363444328, + 0.0042062783613801, + 0.0021713473834097385, + 0.004625894594937563, + 0.003349332371726632, + 0.008998588658869267, + 0.000830021221190691, + 0.0007552141905762255, + 0.0004894907469861209, + 0.0007099962094798684, + 0.000766221433877945, + 0.0009883916936814785, + 0.0006529431557282805, + 0.0006844252930022776, + 0.001921959570609033, + 0.001812714384868741, + 0.0018371895421296358, + 0.0027066306211054325, + 0.0012177618918940425, + 0.0016994340112432837, + 0.0016204215353354812, + 0.0024970127269625664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.00525260902941227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019397982396185398 + ], + "opposite_logprob_tokens": [ + -6.25193977355957 + ], + "logprob_reward": -0.0019397982396185398, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019397982396185398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1024094745516777, + 0.22830457985401154, + 0.08294858783483505, + 0.04732825979590416, + 0.11152941733598709, + 0.08005188405513763, + 0.21050673723220825, + 0.09165145456790924, + 0.022352999076247215, + 0.02291666716337204 + ], + "credit_map": [ + 0.010587871074676514, + 0.010036049410700798, + 0.010795572772622108, + 0.0124586820602417, + 0.01126446295529604, + 0.014795445837080479, + 0.013703292235732079, + 0.01876809448003769, + 0.04842711612582207, + 0.0765329971909523, + 0.011318633332848549, + 0.02857675589621067, + 0.02190074510872364, + 0.011035746894776821, + 0.017465073615312576, + 0.013047504238784313, + 0.009234786964952946, + 0.008140305988490582, + 0.007996131666004658, + 0.011585979722440243, + 0.008421867154538631, + 0.013842927291989326, + 0.01215948536992073, + 0.011567095294594765, + 0.0051053352653980255, + 0.006609251257032156, + 0.004397374112159014, + 0.009280974976718426, + 0.004122440237551928, + 0.00465237395837903, + 0.006722483783960342, + 0.0064380234107375145, + 0.008776616305112839, + 0.011775045655667782, + 0.010438099503517151, + 0.014573698863387108, + 0.012251808308064938, + 0.015498529188334942, + 0.022835740819573402, + 0.015379868447780609, + 0.012499656528234482, + 0.009388687089085579, + 0.007748411502689123, + 0.011686941608786583, + 0.009689874947071075, + 0.005761750042438507, + 0.016778571531176567, + 0.006497983820736408, + 0.0626378208398819, + 0.013642806559801102, + 0.008753184229135513, + 0.012235001660883427, + 0.01128354575484991, + 0.07340551912784576, + 0.014647840522229671, + 0.013900981284677982, + 0.015228016301989555, + 0.007198641076683998, + 0.013205953873693943, + 0.010598648339509964, + 0.007833733223378658, + 0.01132509857416153, + 0.00782502256333828, + 0.018436340615153313, + 0.003653524676337838, + 0.0021774994675070047, + 0.0026497391518205404, + 0.002914481796324253, + 0.0019665840081870556, + 0.003430279204621911, + 0.0022218136582523584, + 0.0033390745520591736, + 0.004026323091238737, + 0.0021226233802735806, + 0.0018775536445900798, + 0.0034688019659370184, + 0.0016377961728721857, + 0.0051584430038928986, + 0.00221654842607677, + 0.002408575266599655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.0019397982396185398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a world imbued with Eastern Zen, a close-up shot from a slightly downward angle captures a plump, emerald green cabbage plummeting with grace and resolve into a pristine earthenware pot below. The cabbage's leaves, rendered in varying shades of green ink, shift from a lush, verdant exterior to a vibrant, jade-white interior. The veins are outlined with precise, powerful lines of light ink, imbued with the texture of life. The moment the cabbage touches the pot's clear, gently simmering broth, the liquid suddenly explodes in all directions. The splashing water is artistically manipulated into a series of powerful, outward-spreading dots of thick and thin ink and vibrant, white-washed brushstrokes, capturing the peak of movement against a backdrop of rice paper-like white. The rising steam from the pot transforms into wisps of soft, flowing ink, blending seamlessly into the background. The whole picture combines movement and stillness, with the vivid emerald green becoming the only visual focus in the simple black, white and gray tones, creating a sense of purity and tranquility that is both full of instantaneous explosive power and extremely poetic and picturesque.", + "seed": 17918, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_4/index_5397_seed_17918_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2302175737451762, + "vlm_logprob": -0.2302175737451762, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.2302175737451762, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003023697528988123 + ], + "opposite_logprob_tokens": [ + -6.503023624420166 + ], + "logprob_reward": -0.003023697528988123, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.003023697528988123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10885649919509888, + 0.057598479092121124, + 0.07500270754098892, + 0.1800609529018402, + 0.1844947636127472, + 0.1338876187801361, + 0.050743941217660904, + 0.08110532909631729, + 0.054617028683423996, + 0.07363267242908478 + ], + "credit_map": [ + 0.013878505676984787, + 0.01214093528687954, + 0.011165970005095005, + 0.010362876579165459, + 0.021199598908424377, + 0.013704679906368256, + 0.013797040097415447, + 0.01260689739137888, + 0.008829495869576931, + 0.008489939384162426, + 0.00487636448815465, + 0.007115410175174475, + 0.008708776906132698, + 0.007908875122666359, + 0.005906984675675631, + 0.005762634798884392, + 0.008438223972916603, + 0.012273493222892284, + 0.004195357672870159, + 0.008366718888282776, + 0.018362009897828102, + 0.009735137224197388, + 0.009143806993961334, + 0.004487961530685425, + 0.01999642327427864, + 0.02303246408700943, + 0.015108521096408367, + 0.024402642622590065, + 0.023560741916298866, + 0.02188734896481037, + 0.030962662771344185, + 0.02111014723777771, + 0.02931840531527996, + 0.029134877026081085, + 0.011306479573249817, + 0.022072838619351387, + 0.02043234184384346, + 0.02756218984723091, + 0.03203646466135979, + 0.01263117790222168, + 0.01154992263764143, + 0.012762263417243958, + 0.007780841086059809, + 0.013270474970340729, + 0.032551247626543045, + 0.03307168185710907, + 0.014322353526949883, + 0.008578845299780369, + 0.006582767702639103, + 0.008129525929689407, + 0.0044819051399827, + 0.004346060566604137, + 0.007759845349937677, + 0.0093440692871809, + 0.004484628327190876, + 0.005615144036710262, + 0.011898933909833431, + 0.008994529023766518, + 0.00813919398933649, + 0.00741012766957283, + 0.012870410457253456, + 0.01719585806131363, + 0.008333481848239899, + 0.006262797862291336, + 0.008459928445518017, + 0.00530107319355011, + 0.004883529618382454, + 0.004007355775684118, + 0.011502435430884361, + 0.011102981865406036, + 0.005196957383304834, + 0.004162767436355352, + 0.010602705180644989, + 0.0085254842415452, + 0.006549787241965532, + 0.006676997058093548, + 0.014529839158058167, + 0.009743371978402138, + 0.01012397836893797, + 0.006880517117679119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.003023697528988123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1369878053665161 + ], + "opposite_logprob_tokens": [ + -0.3869878351688385 + ], + "logprob_reward": -1.1369878053665161, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1369878053665161, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03565823659300804, + 0.04730113595724106, + 0.051301222294569016, + 0.2758474051952362, + 0.039031948894262314, + 0.045900892466306686, + 0.05846051871776581, + 0.2669447064399719, + 0.12669603526592255, + 0.05285792052745819 + ], + "credit_map": [ + 0.004111620131880045, + 0.007738105021417141, + 0.0023521799594163895, + 0.0028738363180309534, + 0.00837224256247282, + 0.004033118486404419, + 0.0037488576490432024, + 0.0024282755330204964, + 0.008248161524534225, + 0.004239619243890047, + 0.005115419626235962, + 0.0067039537243545055, + 0.00483564380556345, + 0.005679967347532511, + 0.0068462989293038845, + 0.005632069427520037, + 0.005689708981662989, + 0.008133398368954659, + 0.0029256166890263557, + 0.005835062358528376, + 0.010439007543027401, + 0.006070962641388178, + 0.0079337852075696, + 0.004273681435734034, + 0.03625534847378731, + 0.03238226845860481, + 0.02894364856183529, + 0.024405675008893013, + 0.029362071305513382, + 0.04000980779528618, + 0.043234214186668396, + 0.04125437140464783, + 0.00693254079669714, + 0.005593381356447935, + 0.0033046354074031115, + 0.003534255549311638, + 0.005341006442904472, + 0.006012482102960348, + 0.0053468006663024426, + 0.0029668465722352266, + 0.00374611490406096, + 0.0035765909124165773, + 0.002011859556660056, + 0.002334914170205593, + 0.013953193090856075, + 0.014644992537796497, + 0.0030814867932349443, + 0.0025517402682453394, + 0.00942491739988327, + 0.005161042790859938, + 0.003751712152734399, + 0.005527082365006208, + 0.01097206212580204, + 0.010986756533384323, + 0.006308892276138067, + 0.006328055635094643, + 0.0340072400867939, + 0.04460139200091362, + 0.017180345952510834, + 0.02973894402384758, + 0.03847028687596321, + 0.04127734526991844, + 0.03439072519540787, + 0.027278432622551918, + 0.01566656306385994, + 0.013183275237679482, + 0.013191924430429935, + 0.008621860295534134, + 0.02969072386622429, + 0.023533033207058907, + 0.01271325908601284, + 0.010095397010445595, + 0.007159111555665731, + 0.006395640783011913, + 0.006605167407542467, + 0.005333161447197199, + 0.008952057920396328, + 0.0065053910948336124, + 0.007475484162569046, + 0.004431903827935457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -1.1369878053665161 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028236540965735912 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.0028236540965735912, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028236540965735912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025449678301811218, + 0.0464305579662323, + 0.049888137727975845, + 0.26789307594299316, + 0.05602627992630005, + 0.0328599289059639, + 0.06610474735498428, + 0.27095866203308105, + 0.14891715347766876, + 0.03547181934118271 + ], + "credit_map": [ + 0.002770048566162586, + 0.0039275698363780975, + 0.002127429936081171, + 0.002422886900603771, + 0.0055716270580887794, + 0.0034617423079907894, + 0.0030154718551784754, + 0.0021529002115130424, + 0.008305171504616737, + 0.004415993578732014, + 0.0048641012981534, + 0.005843408405780792, + 0.005356013774871826, + 0.005724744871258736, + 0.006034713704138994, + 0.0058864085003733635, + 0.005995470564812422, + 0.0069476161152124405, + 0.0030958272982388735, + 0.005121899303048849, + 0.009685803204774857, + 0.009072320535779, + 0.005917794071137905, + 0.0040514045394957066, + 0.03431633487343788, + 0.03168122470378876, + 0.02743399143218994, + 0.025755174458026886, + 0.02935367450118065, + 0.03759952262043953, + 0.04259561374783516, + 0.039157524704933167, + 0.009594988077878952, + 0.007639063056558371, + 0.003986903466284275, + 0.006123457569628954, + 0.007045572157949209, + 0.008727259002625942, + 0.009000108577311039, + 0.003908928018063307, + 0.003521579084917903, + 0.0031950194388628006, + 0.00253803888335824, + 0.002942198421806097, + 0.007214693818241358, + 0.0074968705885112286, + 0.003245485480874777, + 0.002706041093915701, + 0.012045771814882755, + 0.005999827291816473, + 0.004210143815726042, + 0.0051843151450157166, + 0.011347725987434387, + 0.01280209794640541, + 0.007776735350489616, + 0.006738128140568733, + 0.033669717609882355, + 0.047160759568214417, + 0.017421169206500053, + 0.030329402536153793, + 0.03825823962688446, + 0.0398683063685894, + 0.03660416603088379, + 0.027646886184811592, + 0.01788138411939144, + 0.015481929294764996, + 0.014357384294271469, + 0.009211874566972256, + 0.036502331495285034, + 0.028451116755604744, + 0.01423796359449625, + 0.012793163768947124, + 0.004581991583108902, + 0.005521045997738838, + 0.003966394811868668, + 0.0039235628210008144, + 0.005895830225199461, + 0.004310506395995617, + 0.0050437599420547485, + 0.0022287284955382347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.0028236540965735912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638507030904293 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638507030904293, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638507030904293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026695286855101585, + 0.04351871833205223, + 0.03628550097346306, + 0.26818224787712097, + 0.03314487263560295, + 0.039670202881097794, + 0.058005623519420624, + 0.3285154700279236, + 0.11279450356960297, + 0.05318756401538849 + ], + "credit_map": [ + 0.0034808446653187275, + 0.005044735502451658, + 0.0019463757053017616, + 0.0020863076206296682, + 0.006274435203522444, + 0.003255880903452635, + 0.0027321227826178074, + 0.001874585635960102, + 0.008376069366931915, + 0.004000149201601744, + 0.004652806092053652, + 0.00532419141381979, + 0.004607620183378458, + 0.005636093206703663, + 0.005869878455996513, + 0.005051907151937485, + 0.004596109036356211, + 0.004660106264054775, + 0.0022787866182625294, + 0.0038417770992964506, + 0.006824072916060686, + 0.006746156141161919, + 0.004113819915801287, + 0.0032246701885014772, + 0.03446408361196518, + 0.03138625621795654, + 0.028142018243670464, + 0.024589963257312775, + 0.02905268222093582, + 0.03855197876691818, + 0.04216611385345459, + 0.03982916101813316, + 0.003931197337806225, + 0.006642231252044439, + 0.0025920728221535683, + 0.004443584941327572, + 0.004037180915474892, + 0.004769294057041407, + 0.004693860188126564, + 0.0020354550797492266, + 0.003023032797500491, + 0.002953153569251299, + 0.0015887424815446138, + 0.0026397050824016333, + 0.0122520811855793, + 0.012624947354197502, + 0.0027818060480058193, + 0.0018067341297864914, + 0.008956789039075375, + 0.0062035764567554, + 0.0041405330412089825, + 0.0057688625529408455, + 0.010501870885491371, + 0.011382218450307846, + 0.005580902565270662, + 0.00547086913138628, + 0.039366237819194794, + 0.04549097642302513, + 0.017805416136980057, + 0.03171633183956146, + 0.05758092924952507, + 0.07594579458236694, + 0.031958289444446564, + 0.028651513159275055, + 0.01400700956583023, + 0.01239755004644394, + 0.011829531751573086, + 0.008041630499064922, + 0.025760691612958908, + 0.019691215828061104, + 0.011554006487131119, + 0.009512875229120255, + 0.007598697207868099, + 0.006010811775922775, + 0.006203418131917715, + 0.005505098029971123, + 0.00898187980055809, + 0.0065300846472382545, + 0.00829257071018219, + 0.004065004177391529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.004638507030904293 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.00361420470289886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361420470289886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04002763703465462, + 0.052489135414361954, + 0.03950498625636101, + 0.2143663614988327, + 0.08400247991085052, + 0.112810879945755, + 0.053318120539188385, + 0.23066046833992004, + 0.10723043978214264, + 0.06558945775032043 + ], + "credit_map": [ + 0.005191485397517681, + 0.005411683116108179, + 0.0030913131777197123, + 0.004496302921324968, + 0.008292099460959435, + 0.0052743894048035145, + 0.004767091479152441, + 0.0035032734740525484, + 0.010143735446035862, + 0.004705455154180527, + 0.005713059101253748, + 0.0066231596283614635, + 0.005523218307644129, + 0.0067085521295666695, + 0.007381943985819817, + 0.005690015386790037, + 0.0052030510269105434, + 0.005466894246637821, + 0.002328525297343731, + 0.003746681846678257, + 0.007559710647910833, + 0.005905515048652887, + 0.005356640089303255, + 0.003937963396310806, + 0.029019756242632866, + 0.02431655116379261, + 0.02313920110464096, + 0.018541239202022552, + 0.021994253620505333, + 0.03162249177694321, + 0.03310784325003624, + 0.03262502700090408, + 0.01312309317290783, + 0.013668082654476166, + 0.005739586427807808, + 0.009129481390118599, + 0.010518164373934269, + 0.012327824719250202, + 0.01348017156124115, + 0.006016076542437077, + 0.007839557714760303, + 0.005897802300751209, + 0.00337648275308311, + 0.009706727229058743, + 0.0343887098133564, + 0.03415299206972122, + 0.01206968817859888, + 0.005378910806030035, + 0.008228566497564316, + 0.005481556989252567, + 0.004669291898608208, + 0.0049191126599907875, + 0.008916506543755531, + 0.009592527523636818, + 0.005401840899139643, + 0.0061087217181921005, + 0.029696546494960785, + 0.03463855758309364, + 0.014511914923787117, + 0.02342926152050495, + 0.03698796033859253, + 0.041226163506507874, + 0.02750103734433651, + 0.022669026628136635, + 0.014124873094260693, + 0.011631281115114689, + 0.009625677950680256, + 0.007948305457830429, + 0.024972738698124886, + 0.019477317109704018, + 0.011740029789507389, + 0.007710220292210579, + 0.010118773207068443, + 0.00785649660974741, + 0.007198999170213938, + 0.006115072872489691, + 0.012496557086706161, + 0.008148892782628536, + 0.009092617779970169, + 0.00456205103546381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.00361420470289886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a world imbued with Eastern Zen, a close-up shot from a slightly downward angle captures a plump, emerald green cabbage plummeting with grace and resolve into a pristine earthenware pot below. The cabbage's leaves, rendered in varying shades of green ink, shift from a lush, verdant exterior to a vibrant, jade-white interior. The veins are outlined with precise, powerful lines of light ink, imbued with the texture of life. The moment the cabbage touches the pot's clear, gently simmering broth, the liquid suddenly explodes in all directions. The splashing water is artistically manipulated into a series of powerful, outward-spreading dots of thick and thin ink and vibrant, white-washed brushstrokes, capturing the peak of movement against a backdrop of rice paper-like white. The rising steam from the pot transforms into wisps of soft, flowing ink, blending seamlessly into the background. The whole picture combines movement and stillness, with the vivid emerald green becoming the only visual focus in the simple black, white and gray tones, creating a sense of purity and tranquility that is both full of instantaneous explosive power and extremely poetic and picturesque.", + "seed": 17919, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_4/index_5397_seed_17919_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18419571281410754, + "vlm_logprob": -0.18419571281410754, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.18419571281410754, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010591131635010242 + ], + "opposite_logprob_tokens": [ + -4.76059103012085 + ], + "logprob_reward": -0.010591131635010242, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010591131635010242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17106565833091736, + 0.18131525814533234, + 0.06085800379514694, + 0.06954370439052582, + 0.0588512159883976, + 0.04895481467247009, + 0.15664249658584595, + 0.025481639429926872, + 0.175954207777977, + 0.05133306235074997 + ], + "credit_map": [ + 0.028859533369541168, + 0.017766650766134262, + 0.012205898761749268, + 0.021411750465631485, + 0.01725226826965809, + 0.04964350908994675, + 0.010390833020210266, + 0.01353519968688488, + 0.027403494343161583, + 0.025109698995947838, + 0.008004898205399513, + 0.031001409515738487, + 0.019948607310652733, + 0.05152616649866104, + 0.009999319911003113, + 0.008321658708155155, + 0.008658881299197674, + 0.0069808815605938435, + 0.005132547114044428, + 0.009695958346128464, + 0.007555576507002115, + 0.009185226634144783, + 0.008826504461467266, + 0.0048224241472780704, + 0.009101458825170994, + 0.015144436620175838, + 0.006767146289348602, + 0.007693788968026638, + 0.005795746110379696, + 0.009964176453649998, + 0.009637533687055111, + 0.005439405329525471, + 0.008182846941053867, + 0.007954929955303669, + 0.006323778536170721, + 0.0068604834377765656, + 0.006604905705899, + 0.008768577128648758, + 0.008251895196735859, + 0.005903794430196285, + 0.00850690808147192, + 0.004985401406884193, + 0.005002472549676895, + 0.006926484405994415, + 0.0042655314318835735, + 0.006592320278286934, + 0.008076634258031845, + 0.004599056672304869, + 0.01762845553457737, + 0.01648535579442978, + 0.01713978499174118, + 0.008162532933056355, + 0.010071913711726665, + 0.03116639517247677, + 0.011178343556821346, + 0.04480971395969391, + 0.0026052857283502817, + 0.003602356184273958, + 0.0025572103913873434, + 0.002745667938143015, + 0.002850618213415146, + 0.004767107777297497, + 0.0034830651711672544, + 0.002870326628908515, + 0.010775148868560791, + 0.01399495080113411, + 0.009350511245429516, + 0.010981926694512367, + 0.006874608341604471, + 0.059120070189237595, + 0.009748483076691628, + 0.05510849133133888, + 0.005797130521386862, + 0.006076507735997438, + 0.0035818605683743954, + 0.005077685695141554, + 0.0036710791755467653, + 0.011547594331204891, + 0.009787263348698616, + 0.00579393794760108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.010591131635010242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898283064365387 + ], + "opposite_logprob_tokens": [ + -0.523283064365387 + ], + "logprob_reward": -0.898283064365387, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898283064365387, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22843538224697113, + 0.09374154359102249, + 0.14011308550834656, + 0.08899368345737457, + 0.05720371752977371, + 0.026102622970938683, + 0.16529834270477295, + 0.05368274450302124, + 0.044806815683841705, + 0.1016220822930336 + ], + "credit_map": [ + 0.038599591702222824, + 0.028947722166776657, + 0.023008612915873528, + 0.030689511448144913, + 0.022039759904146194, + 0.039527591317892075, + 0.021626971662044525, + 0.023995619267225266, + 0.013594154268503189, + 0.009184311144053936, + 0.006525206845253706, + 0.016089648008346558, + 0.0071058436296880245, + 0.029134275391697884, + 0.005633012857288122, + 0.00647508492693305, + 0.023002387955784798, + 0.011184047907590866, + 0.019020916894078255, + 0.014432836323976517, + 0.009360509924590588, + 0.027703948318958282, + 0.01308528333902359, + 0.022323142737150192, + 0.015469002537429333, + 0.013709870167076588, + 0.008832397870719433, + 0.007551869843155146, + 0.007057228125631809, + 0.015381420962512493, + 0.01269685011357069, + 0.008295046165585518, + 0.00586027791723609, + 0.0081553990021348, + 0.0037107786629348993, + 0.010985349304974079, + 0.006267149467021227, + 0.006293864920735359, + 0.012535939924418926, + 0.0033949564676731825, + 0.0028912629932165146, + 0.0037560435011982918, + 0.0030650843400508165, + 0.003612637985497713, + 0.00269640376791358, + 0.0025561004877090454, + 0.003994245547801256, + 0.003530845744535327, + 0.01916457526385784, + 0.06268289685249329, + 0.01008833572268486, + 0.010011973790824413, + 0.0138086611405015, + 0.021983228623867035, + 0.009488304145634174, + 0.018070382997393608, + 0.003979641944169998, + 0.009552596136927605, + 0.003945782780647278, + 0.005332520697265863, + 0.0037312202621251345, + 0.0054373182356357574, + 0.006076891906559467, + 0.01562677137553692, + 0.005595953203737736, + 0.003426551353186369, + 0.002693649148568511, + 0.0015472548548132181, + 0.0018882593140006065, + 0.015713002532720566, + 0.002023638691753149, + 0.01191850658506155, + 0.011700187809765339, + 0.015156472101807594, + 0.008339119143784046, + 0.0089124059304595, + 0.008646178990602493, + 0.027614550665020943, + 0.009649776853621006, + 0.01160339079797268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.898283064365387 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262451246380806 + ], + "opposite_logprob_tokens": [ + -5.25526237487793 + ], + "logprob_reward": -0.005262451246380806, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262451246380806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19503524899482727, + 0.08725464344024658, + 0.1932811588048935, + 0.08526867628097534, + 0.08758917450904846, + 0.03541982173919678, + 0.09957001358270645, + 0.026279248297214508, + 0.020553747192025185, + 0.1697482466697693 + ], + "credit_map": [ + 0.03165906295180321, + 0.026029424741864204, + 0.014149948954582214, + 0.02477928251028061, + 0.023048117756843567, + 0.042494095861911774, + 0.01499708741903305, + 0.017878221347928047, + 0.012186896055936813, + 0.01136151421815157, + 0.005384855438023806, + 0.015558682382106781, + 0.007802891079336405, + 0.022706272080540657, + 0.006554455496370792, + 0.005699076224118471, + 0.033170394599437714, + 0.016541220247745514, + 0.03123980388045311, + 0.019858844578266144, + 0.016918964684009552, + 0.025282353162765503, + 0.019653242081403732, + 0.03061634860932827, + 0.01022341102361679, + 0.012553556822240353, + 0.008588682860136032, + 0.011506053619086742, + 0.007845844142138958, + 0.011732361279428005, + 0.016340535134077072, + 0.006478226277977228, + 0.015935005620121956, + 0.01039656437933445, + 0.005340484902262688, + 0.015396647155284882, + 0.005857082083821297, + 0.010701456107199192, + 0.017251065000891685, + 0.006710863672196865, + 0.006852830294519663, + 0.004653786309063435, + 0.0037204513791948557, + 0.003635906148701906, + 0.0033385998103767633, + 0.004449131898581982, + 0.004170473664999008, + 0.004598640836775303, + 0.010188755579292774, + 0.03443634510040283, + 0.005708238109946251, + 0.0051736547611653805, + 0.007026030216366053, + 0.00784797128289938, + 0.0057346271350979805, + 0.023454388603568077, + 0.0029716170392930508, + 0.005001099314540625, + 0.0030939222779124975, + 0.0017899919766932726, + 0.0019170562736690044, + 0.0030054026283323765, + 0.0015841205604374409, + 0.006916038691997528, + 0.001793971168808639, + 0.0019009121460840106, + 0.0013095111353322864, + 0.0011201329762116075, + 0.0012941102031618357, + 0.006295092403888702, + 0.0015802668640390038, + 0.005259750410914421, + 0.018591800704598427, + 0.026352470740675926, + 0.01172420009970665, + 0.014587348327040672, + 0.014531168155372143, + 0.043399419635534286, + 0.025038782507181168, + 0.015523062087595463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.005262451246380806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642422776669264 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.004642422776669264, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642422776669264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2650423049926758, + 0.05716802552342415, + 0.0789630189538002, + 0.06223531812429428, + 0.08220817893743515, + 0.07572578638792038, + 0.08150006085634232, + 0.07712256163358688, + 0.15322977304458618, + 0.06680500507354736 + ], + "credit_map": [ + 0.042265865951776505, + 0.04345092177391052, + 0.021324021741747856, + 0.029482295736670494, + 0.029643956571817398, + 0.05585532262921333, + 0.019319768995046616, + 0.023700151592493057, + 0.009085661731660366, + 0.009012732654809952, + 0.006061940919607878, + 0.007797336205840111, + 0.007115740329027176, + 0.006690873298794031, + 0.00559579860419035, + 0.00580794271081686, + 0.012989681214094162, + 0.008007032796740532, + 0.008306320756673813, + 0.013427868485450745, + 0.0060736085288226604, + 0.013938043266534805, + 0.008182198740541935, + 0.00803827028721571, + 0.008217141032218933, + 0.010956895537674427, + 0.00696819880977273, + 0.005950501188635826, + 0.004834734834730625, + 0.010795648209750652, + 0.00879411306232214, + 0.005718081258237362, + 0.012541765347123146, + 0.011200875043869019, + 0.005350319668650627, + 0.014135041274130344, + 0.007919407449662685, + 0.010858851484954357, + 0.01476383674889803, + 0.00543808052316308, + 0.010585442185401917, + 0.0089806467294693, + 0.007494209334254265, + 0.012703008949756622, + 0.007794609293341637, + 0.0085625434294343, + 0.013576204888522625, + 0.006029119249433279, + 0.008282763883471489, + 0.016690189018845558, + 0.006306135561317205, + 0.004679582547396421, + 0.004697030875831842, + 0.012130375020205975, + 0.005128735210746527, + 0.023585256189107895, + 0.009157863445580006, + 0.009923129342496395, + 0.010336952283978462, + 0.008720899000763893, + 0.008810168132185936, + 0.009831216186285019, + 0.012269719503819942, + 0.008072616532444954, + 0.011938203126192093, + 0.01186192873865366, + 0.0069120293483138084, + 0.00632916996255517, + 0.006979056168347597, + 0.0546853207051754, + 0.008531013503670692, + 0.04599307104945183, + 0.009590593166649342, + 0.010625588707625866, + 0.007496018428355455, + 0.007375768851488829, + 0.007158441934734583, + 0.006491596810519695, + 0.010427163913846016, + 0.007639830932021141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.004642422776669264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199494047090411 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.002199494047090411, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199494047090411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3349069654941559, + 0.10115363448858261, + 0.10696401447057724, + 0.08989156037569046, + 0.03860808536410332, + 0.04500609263777733, + 0.08835870027542114, + 0.04204949364066124, + 0.09626122564077377, + 0.05680025368928909 + ], + "credit_map": [ + 0.05919627100229263, + 0.050016265362501144, + 0.029083721339702606, + 0.042731042951345444, + 0.04027286171913147, + 0.05891416594386101, + 0.02503500133752823, + 0.029657648876309395, + 0.012629833072423935, + 0.012170455418527126, + 0.006372499279677868, + 0.015287558548152447, + 0.010722772218286991, + 0.030115781351923943, + 0.006264090072363615, + 0.007590638007968664, + 0.01750633679330349, + 0.009078871458768845, + 0.011045192368328571, + 0.01518435962498188, + 0.007687620352953672, + 0.019717367365956306, + 0.009943359531462193, + 0.016800902783870697, + 0.01652996987104416, + 0.009392482228577137, + 0.00991026684641838, + 0.00832919217646122, + 0.009073878638446331, + 0.017353080213069916, + 0.008217325434088707, + 0.01108536683022976, + 0.0066437553614377975, + 0.0055042351596057415, + 0.002850799821317196, + 0.004091263748705387, + 0.004008863586932421, + 0.004712335765361786, + 0.006454075686633587, + 0.0043427543714642525, + 0.005188183393329382, + 0.005648351740092039, + 0.0044216071255505085, + 0.00694115599617362, + 0.00465584360063076, + 0.005929280538111925, + 0.006906537804752588, + 0.005315133836120367, + 0.01415424793958664, + 0.01807921938598156, + 0.007331764325499535, + 0.005616508424282074, + 0.006892136298120022, + 0.016887253150343895, + 0.00748926866799593, + 0.011908305808901787, + 0.004213809501379728, + 0.0057142507284879684, + 0.0048551131039857864, + 0.0035232556983828545, + 0.0037751514464616776, + 0.0053241066634655, + 0.004714202135801315, + 0.009929606691002846, + 0.007008176762610674, + 0.007530780043452978, + 0.004060774575918913, + 0.0033739502541720867, + 0.003794695483520627, + 0.03554311394691467, + 0.0045683193020522594, + 0.03038141317665577, + 0.006680341437458992, + 0.008066135458648205, + 0.004788350313901901, + 0.004612363874912262, + 0.004358499310910702, + 0.015376314520835876, + 0.005080940667539835, + 0.007837307639420033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.002199494047090411 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a world imbued with Eastern Zen, a close-up shot from a slightly downward angle captures a plump, emerald green cabbage plummeting with grace and resolve into a pristine earthenware pot below. The cabbage's leaves, rendered in varying shades of green ink, shift from a lush, verdant exterior to a vibrant, jade-white interior. The veins are outlined with precise, powerful lines of light ink, imbued with the texture of life. The moment the cabbage touches the pot's clear, gently simmering broth, the liquid suddenly explodes in all directions. The splashing water is artistically manipulated into a series of powerful, outward-spreading dots of thick and thin ink and vibrant, white-washed brushstrokes, capturing the peak of movement against a backdrop of rice paper-like white. The rising steam from the pot transforms into wisps of soft, flowing ink, blending seamlessly into the background. The whole picture combines movement and stillness, with the vivid emerald green becoming the only visual focus in the simple black, white and gray tones, creating a sense of purity and tranquility that is both full of instantaneous explosive power and extremely poetic and picturesque.", + "seed": 17920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_4/index_5397_seed_17920_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15638453187420964, + "vlm_logprob": -0.15638453187420964, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.15638453187420964, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005143742077052593 + ], + "opposite_logprob_tokens": [ + -5.755143642425537 + ], + "logprob_reward": -0.005143742077052593, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.005143742077052593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30383020639419556, + 0.09622325748205185, + 0.09617789834737778, + 0.11711986362934113, + 0.026487702503800392, + 0.18593692779541016, + 0.055579934269189835, + 0.025207728147506714, + 0.04020155966281891, + 0.05323495343327522 + ], + "credit_map": [ + 0.050147950649261475, + 0.03284931927919388, + 0.05491086468100548, + 0.035185668617486954, + 0.029844628646969795, + 0.03501380607485771, + 0.03694869577884674, + 0.02892925962805748, + 0.012281280010938644, + 0.022106288000941277, + 0.015363059937953949, + 0.008372356183826923, + 0.013621714897453785, + 0.009788542054593563, + 0.008571203798055649, + 0.006118808872997761, + 0.011760225519537926, + 0.015001296065747738, + 0.014632819220423698, + 0.009946825914084911, + 0.011105697602033615, + 0.008494085632264614, + 0.013953392393887043, + 0.011283546686172485, + 0.015708979219198227, + 0.010605418123304844, + 0.008091303519904613, + 0.007676855660974979, + 0.015247401781380177, + 0.03875156491994858, + 0.012881111353635788, + 0.00815722905099392, + 0.0030771829187870026, + 0.003015459282323718, + 0.0035724062472581863, + 0.0020373230800032616, + 0.005040404386818409, + 0.002479239832609892, + 0.00386070366948843, + 0.0034049833193421364, + 0.020356960594654083, + 0.014161018654704094, + 0.02801675535738468, + 0.014117017388343811, + 0.017916865646839142, + 0.05354625731706619, + 0.015679864212870598, + 0.02214217744767666, + 0.013602614402770996, + 0.003637149464339018, + 0.004739745985716581, + 0.004460531752556562, + 0.003925488796085119, + 0.004134561866521835, + 0.005612958688288927, + 0.01546687912195921, + 0.002648387337103486, + 0.001993466168642044, + 0.002782682189717889, + 0.0028858783189207315, + 0.003956438973546028, + 0.0020156067330390215, + 0.0038845606613904238, + 0.005040706600993872, + 0.0050899311900138855, + 0.003562095807865262, + 0.005303184036165476, + 0.0038348992820829153, + 0.004818334244191647, + 0.004606941249221563, + 0.004063824657350779, + 0.00892234593629837, + 0.006262159440666437, + 0.005185682792216539, + 0.009227127768099308, + 0.00459771603345871, + 0.007157493848353624, + 0.0073099019937217236, + 0.004216821398586035, + 0.00927804410457611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.005143742077052593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577705979347229 + ], + "opposite_logprob_tokens": [ + -0.6327705979347229 + ], + "logprob_reward": -0.7577705979347229, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577705979347229, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1070450022816658, + 0.09265447407960892, + 0.14357753098011017, + 0.45060113072395325, + 0.04750697314739227, + 0.03319395333528519, + 0.05396813899278641, + 0.023497112095355988, + 0.0343349426984787, + 0.01362079381942749 + ], + "credit_map": [ + 0.01603889837861061, + 0.011751099489629269, + 0.015441658906638622, + 0.016313092783093452, + 0.009547855705022812, + 0.011732489801943302, + 0.012775767594575882, + 0.013444124720990658, + 0.014445914886891842, + 0.014442394487559795, + 0.02071264386177063, + 0.012139524333178997, + 0.007577738258987665, + 0.005387590732425451, + 0.008757859468460083, + 0.009190799668431282, + 0.02258947491645813, + 0.01508037094026804, + 0.03367265686392784, + 0.012196959927678108, + 0.013274230062961578, + 0.01225125603377819, + 0.013577966950833797, + 0.020934592932462692, + 0.08193054050207138, + 0.03827202692627907, + 0.019963281229138374, + 0.01395540963858366, + 0.04467574506998062, + 0.19075676798820496, + 0.028221987187862396, + 0.032825350761413574, + 0.004690282978117466, + 0.005639287643134594, + 0.005207562819123268, + 0.004260253626853228, + 0.009683976881206036, + 0.00490166712552309, + 0.007702421862632036, + 0.005421516485512257, + 0.0031993859447538853, + 0.0028539872728288174, + 0.003913528751581907, + 0.003327392740175128, + 0.005180833861231804, + 0.007360665127635002, + 0.0036042104475200176, + 0.0037539496552199125, + 0.014940817840397358, + 0.003344962140545249, + 0.009035266004502773, + 0.0035923223476856947, + 0.0033266020473092794, + 0.0033364766277372837, + 0.004545438103377819, + 0.011846249923110008, + 0.0019383099861443043, + 0.0030396406073123217, + 0.0029399357736110687, + 0.0023051307071000338, + 0.002618329366669059, + 0.002838925225660205, + 0.002608447801321745, + 0.005208390299230814, + 0.0039038953837007284, + 0.003254317445680499, + 0.004797737579792738, + 0.0028897649608552456, + 0.006168120075017214, + 0.003564354032278061, + 0.004241362679749727, + 0.005515387747436762, + 0.001805537729524076, + 0.0015015978133305907, + 0.0018928456120193005, + 0.0013228837633505464, + 0.0015951356617733836, + 0.0014709122478961945, + 0.0016877056332305074, + 0.002344174310564995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.7577705979347229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649504743516445 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649504743516445, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649504743516445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09701666980981827, + 0.06844229251146317, + 0.09167119115591049, + 0.5519202351570129, + 0.0318344421684742, + 0.03196081519126892, + 0.05474309250712395, + 0.026350224390625954, + 0.025044649839401245, + 0.021016409620642662 + ], + "credit_map": [ + 0.01667865179479122, + 0.010763328522443771, + 0.0286241564899683, + 0.008607923053205013, + 0.008494023233652115, + 0.00930535327643156, + 0.006637483835220337, + 0.007905752398073673, + 0.010740098543465137, + 0.011128638871014118, + 0.017327826470136642, + 0.006976443342864513, + 0.007049154955893755, + 0.004404766950756311, + 0.005766135640442371, + 0.005049229599535465, + 0.01267683319747448, + 0.01162286102771759, + 0.01996765471994877, + 0.00944554340094328, + 0.009762794710695744, + 0.008492004126310349, + 0.009668024256825447, + 0.010035479441285133, + 0.10223209112882614, + 0.04571075737476349, + 0.023021256551146507, + 0.017231881618499756, + 0.05192985385656357, + 0.23573964834213257, + 0.03549511358141899, + 0.04055961221456528, + 0.003464993555098772, + 0.004026909824460745, + 0.004876525606960058, + 0.002459878334775567, + 0.004667036700993776, + 0.003214007243514061, + 0.004407134372740984, + 0.004717960488051176, + 0.0028047184459865093, + 0.002455084817484021, + 0.0050213332287967205, + 0.002324255183339119, + 0.004422365687787533, + 0.006514050532132387, + 0.0037193470634520054, + 0.0046996623277664185, + 0.013478263281285763, + 0.00271913455799222, + 0.003722449531778693, + 0.0024527304340153933, + 0.0022362566087394953, + 0.0045413970947265625, + 0.004618085455149412, + 0.02097477950155735, + 0.002558767329901457, + 0.0028030816465616226, + 0.003997039515525103, + 0.002819616813212633, + 0.003750747535377741, + 0.002714965958148241, + 0.002945245709270239, + 0.004760758951306343, + 0.003106615971773863, + 0.0022629068698734045, + 0.0036825942806899548, + 0.0025029568932950497, + 0.0030505105387419462, + 0.003545881947502494, + 0.003459933213889599, + 0.003433248959481716, + 0.002553630853071809, + 0.002042301930487156, + 0.002477699425071478, + 0.0017537042731419206, + 0.0031195024494081736, + 0.0030552237294614315, + 0.0020135340746492147, + 0.004000812768936157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.008649504743516445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067402152344584465 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.0067402152344584465, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067402152344584465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18758298456668854, + 0.12709499895572662, + 0.12945863604545593, + 0.08488858491182327, + 0.0986894890666008, + 0.10420408844947815, + 0.13972200453281403, + 0.05421339347958565, + 0.049775682389736176, + 0.024370137602090836 + ], + "credit_map": [ + 0.03020135872066021, + 0.019911017268896103, + 0.034494366496801376, + 0.022765720263123512, + 0.01866471953690052, + 0.024035222828388214, + 0.019862039014697075, + 0.01764853671193123, + 0.019373347982764244, + 0.01593739353120327, + 0.029739273712038994, + 0.01587030664086342, + 0.010915180668234825, + 0.00961750466376543, + 0.011854910291731358, + 0.01378708891570568, + 0.017275281250476837, + 0.02020946331322193, + 0.02316110208630562, + 0.01113896630704403, + 0.019691037014126778, + 0.01230042614042759, + 0.01327733974903822, + 0.012405016459524632, + 0.008837897330522537, + 0.009145832620561123, + 0.008701073005795479, + 0.006253646221011877, + 0.016350435093045235, + 0.011996204033493996, + 0.013151486404240131, + 0.010452011600136757, + 0.009722490794956684, + 0.012740680016577244, + 0.012666798196732998, + 0.007356502581387758, + 0.020503714680671692, + 0.0079800458624959, + 0.013493170030415058, + 0.014226083643734455, + 0.011426282115280628, + 0.008341929875314236, + 0.014879494905471802, + 0.008614149875938892, + 0.010993543080985546, + 0.02664433978497982, + 0.010105454362928867, + 0.013198896311223507, + 0.03170671686530113, + 0.007705504074692726, + 0.013532347045838833, + 0.00943697988986969, + 0.010404077358543873, + 0.009682980366051197, + 0.012338036671280861, + 0.04491537809371948, + 0.005325694568455219, + 0.0052763898856937885, + 0.009042530320584774, + 0.0051462105475366116, + 0.007638394832611084, + 0.005950658116489649, + 0.007128573954105377, + 0.008704942651093006, + 0.005596587434411049, + 0.004782690200954676, + 0.006052230950444937, + 0.004126733634620905, + 0.009389250539243221, + 0.005419247783720493, + 0.006666103377938271, + 0.007742837071418762, + 0.0027426190208643675, + 0.002039593178778887, + 0.0037715148646384478, + 0.0020058329682797194, + 0.0025137129705399275, + 0.0030620284378528595, + 0.00314574153162539, + 0.005089092999696732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.0067402152344584465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, emerald green cabbage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cabbage fall into a pristine earthenware pot containing clear, gently simmering broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the liquid explode outward into dots of thick and thin ink and white-washed brushstrokes upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background and steam rendered in a rice paper-like white style with soft, flowing ink wisps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036185993812978268 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036185993812978268, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036185993812978268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1733030080795288, + 0.05177391692996025, + 0.11652939766645432, + 0.48704057931900024, + 0.018101563677191734, + 0.058252789080142975, + 0.031733833253383636, + 0.01651790924370289, + 0.026319170370697975, + 0.020427845418453217 + ], + "credit_map": [ + 0.02575906366109848, + 0.019340073689818382, + 0.028277143836021423, + 0.027164742350578308, + 0.01609560288488865, + 0.01926722750067711, + 0.02122781053185463, + 0.01617136411368847, + 0.009168455377221107, + 0.008425443433225155, + 0.009140011854469776, + 0.007227200549095869, + 0.004745136480778456, + 0.004055717494338751, + 0.004828756209462881, + 0.004183192737400532, + 0.0180349200963974, + 0.01587492786347866, + 0.02550681121647358, + 0.010331814177334309, + 0.013231990858912468, + 0.00949349720031023, + 0.010271796025335789, + 0.013783641159534454, + 0.0878189280629158, + 0.040553539991378784, + 0.02231142483651638, + 0.014829951338469982, + 0.047335222363471985, + 0.20669768750667572, + 0.03135821968317032, + 0.036135610193014145, + 0.0023443105164915323, + 0.0018437332473695278, + 0.0022851566318422556, + 0.001525455736555159, + 0.003054344328120351, + 0.002361525781452656, + 0.002982444129884243, + 0.0017045929562300444, + 0.00601048581302166, + 0.004164982121437788, + 0.010173306800425053, + 0.003977571148425341, + 0.004967242479324341, + 0.018963685259222984, + 0.004131022375077009, + 0.00586449122056365, + 0.005297800526022911, + 0.0027012263890355825, + 0.004286966752260923, + 0.003552590496838093, + 0.0034690722823143005, + 0.004355713725090027, + 0.004797786008566618, + 0.0032726777717471123, + 0.002004059962928295, + 0.0018978335428982973, + 0.0024974835105240345, + 0.00168418581597507, + 0.0015986502403393388, + 0.001912783132866025, + 0.001686978037469089, + 0.003235934069380164, + 0.0027391284238547087, + 0.0027600203175097704, + 0.003341649193316698, + 0.0023124688304960728, + 0.00501226494088769, + 0.0030667355749756098, + 0.0034669265151023865, + 0.0036199777387082577, + 0.00267033651471138, + 0.0022002351470291615, + 0.003163327230140567, + 0.0018673611339181662, + 0.002341087209060788, + 0.0026412690058350563, + 0.0023477531503885984, + 0.003196478122845292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0036185993812978268 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_5/index_tensor([6069, 6069])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_5/index_tensor([6069, 6069])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5eeff5bc73468f40b82dfdb31d9f1cf5cab03076 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_5/index_tensor([6069, 6069])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, medium-close-up, static shot focuses horizontally on an elderly haenyeo, clad in a tight, traditional black rubber diving suit. The form-fitting material reflects a faint sheen in the light, still stained with a hint of salt. She sits on a patch of black volcanic rock, smoothed by the sea. Behind her is the damp mudflats left by the slowly receding tide and the shimmering, calm sea. The golden evening sun, streaming down at a low angle, casts a soft shimmer on her graying, wet bun and shoulders. Her face, a product of time and the sea, is a map of deep wrinkles across her tanned skin, yet her eyes are remarkably clear and luminous. She speaks to the side, a warm, proud smile on her face. Her calloused hands rest naturally on her knees, occasionally lifting in a small, powerful gesture as she speaks. The entire picture is full of rich film grain and extremely sharp details. From every deep wrinkle on her cheek to the worn texture of her diving suit, they are clearly visible. The perfect light accurately captures a flicker in her deep and intelligent eyes, creating an interview atmosphere that is both real and intimate as well as nostalgic and poetic.", + "seed": 18028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_5/index_6069_seed_18028_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10088155046105385, + "vlm_logprob": -0.10088155046105385, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.10088155046105385, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1054157093167305 + ], + "opposite_logprob_tokens": [ + -2.3554158210754395 + ], + "logprob_reward": -0.1054157093167305, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1054157093167305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21926404535770416, + 0.08105387538671494, + 0.04600760340690613, + 0.016076141968369484, + 0.33202868700027466, + 0.2348470836877823, + 0.01191222108900547, + 0.0288290586322546, + 0.00926902424544096, + 0.02071230299770832 + ], + "credit_map": [ + 0.01737189292907715, + 0.025217564776539803, + 0.021555813029408455, + 0.02520296536386013, + 0.02488170750439167, + 0.021072279661893845, + 0.030048605054616928, + 0.05391320586204529, + 0.010768767446279526, + 0.007116618100553751, + 0.014201966114342213, + 0.006922184955328703, + 0.0071532102301716805, + 0.01252145878970623, + 0.00865041371434927, + 0.01371924951672554, + 0.006656485144048929, + 0.004421286750584841, + 0.006902527529746294, + 0.0070998999290168285, + 0.0055050053633749485, + 0.004330177325755358, + 0.005059402901679277, + 0.006032814271748066, + 0.0019147269194945693, + 0.002118093892931938, + 0.0026601857971400023, + 0.0012796144001185894, + 0.002782546915113926, + 0.0012934627011418343, + 0.0019644242711365223, + 0.002063086023554206, + 0.025111384689807892, + 0.02226010337471962, + 0.02372070588171482, + 0.09886368364095688, + 0.019359489902853966, + 0.09892410784959793, + 0.022038014605641365, + 0.02175116539001465, + 0.018354184925556183, + 0.013075386174023151, + 0.017066149041056633, + 0.07612916082143784, + 0.012602987699210644, + 0.0707269012928009, + 0.01366960909217596, + 0.013222681358456612, + 0.0015348604647442698, + 0.0012736569624394178, + 0.0019128667190670967, + 0.0009668026468716562, + 0.0012313423212617636, + 0.0014153921511024237, + 0.0014272202970460057, + 0.0021500778384506702, + 0.002848066622391343, + 0.004857745487242937, + 0.0035378425382077694, + 0.002974202623590827, + 0.00301532750017941, + 0.004628439899533987, + 0.0032393967267125845, + 0.0037280351389199495, + 0.0019114749738946557, + 0.0007956595509313047, + 0.0011967895552515984, + 0.001141587970778346, + 0.00085665681399405, + 0.0013523853849619627, + 0.0006815681699663401, + 0.0013329010689631104, + 0.0022955879103392363, + 0.0017611391376703978, + 0.0022388051729649305, + 0.002123794984072447, + 0.002006317488849163, + 0.003564313752576709, + 0.002507120603695512, + 0.004215221852064133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.1054157093167305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609652116894722 + ], + "opposite_logprob_tokens": [ + -4.141096591949463 + ], + "logprob_reward": -0.01609652116894722, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609652116894722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21869121491909027, + 0.07139012217521667, + 0.047793760895729065, + 0.05672644451260567, + 0.37607815861701965, + 0.12787310779094696, + 0.022261615842580795, + 0.016917521134018898, + 0.013908405788242817, + 0.04835961014032364 + ], + "credit_map": [ + 0.04169631376862526, + 0.01037480030208826, + 0.0170463677495718, + 0.01271758321672678, + 0.018177015706896782, + 0.07827678322792053, + 0.021728336811065674, + 0.01867402158677578, + 0.009531669318675995, + 0.006362868472933769, + 0.010476967319846153, + 0.007134813349694014, + 0.006001666653901339, + 0.012186609208583832, + 0.007584388833492994, + 0.012111141346395016, + 0.010261204093694687, + 0.0034493915736675262, + 0.004657127894461155, + 0.003279908560216427, + 0.004718164913356304, + 0.008260994218289852, + 0.004698698874562979, + 0.00846826657652855, + 0.00822488684207201, + 0.0029918993823230267, + 0.005239594727754593, + 0.016170451417565346, + 0.003287859261035919, + 0.0131599809974432, + 0.0038844356313347816, + 0.0037673383485525846, + 0.026214255020022392, + 0.0194918904453516, + 0.027035754173994064, + 0.11469507962465286, + 0.019689347594976425, + 0.11455342918634415, + 0.026653943583369255, + 0.027744470164179802, + 0.010193802416324615, + 0.007378395646810532, + 0.00939762033522129, + 0.040720030665397644, + 0.007066068705171347, + 0.03832661360502243, + 0.007278638891875744, + 0.007511948235332966, + 0.004642955493181944, + 0.002285934053361416, + 0.0024941880255937576, + 0.0024612871930003166, + 0.002582149114459753, + 0.0022563051898032427, + 0.0021917149424552917, + 0.0033470825292170048, + 0.0024316341150552034, + 0.002560267923399806, + 0.002039856044575572, + 0.0018796464428305626, + 0.0018951351521536708, + 0.0018827376188710332, + 0.0014078784734010696, + 0.002820364898070693, + 0.002151723485440016, + 0.0017409747233614326, + 0.0015422600554302335, + 0.001962724607437849, + 0.0017257030121982098, + 0.0021363813430070877, + 0.0011315657757222652, + 0.0015170729020610452, + 0.006109369918704033, + 0.0037009811494499445, + 0.005913155153393745, + 0.006199365481734276, + 0.005037197843194008, + 0.008772457018494606, + 0.005391915328800678, + 0.0072351666167378426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.01609652116894722 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016083618625998497 + ], + "opposite_logprob_tokens": [ + -4.141083717346191 + ], + "logprob_reward": -0.016083618625998497, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016083618625998497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40242770314216614, + 0.0476941280066967, + 0.03780016303062439, + 0.017775479704141617, + 0.36361634731292725, + 0.038604311645030975, + 0.02455674298107624, + 0.0299956277012825, + 0.019091447815299034, + 0.018438085913658142 + ], + "credit_map": [ + 0.05219687893986702, + 0.023488778620958328, + 0.05486121028661728, + 0.02570059336721897, + 0.030940625816583633, + 0.1331648826599121, + 0.03358364850282669, + 0.04849106818437576, + 0.008856028318405151, + 0.005587006453424692, + 0.007409712765365839, + 0.005713643040508032, + 0.0039903647266328335, + 0.0057712215930223465, + 0.0049568708054721355, + 0.005409278441220522, + 0.005122513975948095, + 0.0030110380612313747, + 0.00603401567786932, + 0.004146398510783911, + 0.0039035857189446688, + 0.007567308377474546, + 0.003187221474945545, + 0.004828081000596285, + 0.0025470361579209566, + 0.0018618369940668344, + 0.0021563859190791845, + 0.0024808975867927074, + 0.002156479051336646, + 0.002922412706539035, + 0.0020868529099971056, + 0.0015635776799172163, + 0.027911154553294182, + 0.02137123979628086, + 0.0285273976624012, + 0.10984522849321365, + 0.019619813188910484, + 0.10699065029621124, + 0.02536158636212349, + 0.023989267647266388, + 0.005926225800067186, + 0.0022085506934672594, + 0.004056822974234819, + 0.0067410594783723354, + 0.0026410245336592197, + 0.01068746205419302, + 0.002581329084932804, + 0.003761838423088193, + 0.005499863997101784, + 0.001902031246572733, + 0.002534928498789668, + 0.002344651846215129, + 0.0024427687749266624, + 0.0023423866368830204, + 0.002416338538751006, + 0.005073772743344307, + 0.003441512119024992, + 0.004729345440864563, + 0.0036087168846279383, + 0.0036066703032702208, + 0.0046958946622908115, + 0.00326375849545002, + 0.0036202615592628717, + 0.0030294680036604404, + 0.0026972091291099787, + 0.0013848724775016308, + 0.0026001769583672285, + 0.0018966586794704199, + 0.0023430867586284876, + 0.002464513760060072, + 0.001687905052676797, + 0.0040170252323150635, + 0.002170171355828643, + 0.0020903530530631542, + 0.0024704746901988983, + 0.0027807194273918867, + 0.001711257267743349, + 0.0025654397904872894, + 0.001968563534319401, + 0.0026811074931174517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.016083618625998497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198074772953987 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018198074772953987, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198074772953987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20181448757648468, + 0.08954343944787979, + 0.041859857738018036, + 0.07647107541561127, + 0.09382074326276779, + 0.32498809695243835, + 0.026607057079672813, + 0.02414485067129135, + 0.03593527898192406, + 0.08481508493423462 + ], + "credit_map": [ + 0.028105478733778, + 0.01227796170860529, + 0.01492341235280037, + 0.016604524105787277, + 0.017242666333913803, + 0.052745308727025986, + 0.022262349724769592, + 0.03765280172228813, + 0.007169547025114298, + 0.005341535434126854, + 0.015830790624022484, + 0.011235468089580536, + 0.007821289822459221, + 0.017511000856757164, + 0.008953526616096497, + 0.015680288895964622, + 0.005404196213930845, + 0.003336699213832617, + 0.007262264844030142, + 0.005323487799614668, + 0.0034178656060248613, + 0.005335349589586258, + 0.004390595946460962, + 0.007389399688690901, + 0.00694654555991292, + 0.008793634362518787, + 0.014017431996762753, + 0.010563264600932598, + 0.006915706675499678, + 0.010095599107444286, + 0.008688559755682945, + 0.010450341738760471, + 0.00562006002292037, + 0.010130618698894978, + 0.010368100367486477, + 0.019501518458127975, + 0.006297197658568621, + 0.020346185192465782, + 0.008295266889035702, + 0.01326180249452591, + 0.025045905262231827, + 0.01753993146121502, + 0.02337920293211937, + 0.10243486613035202, + 0.018282286822795868, + 0.09920355677604675, + 0.01998482644557953, + 0.019117550924420357, + 0.005316960159689188, + 0.0025824850890785456, + 0.0022672333288937807, + 0.0020390364807099104, + 0.0035250517539680004, + 0.0028563805390149355, + 0.0028178140055388212, + 0.005202096421271563, + 0.003043129574507475, + 0.002803722396492958, + 0.0026756778825074434, + 0.0021370770409703255, + 0.0030102406162768602, + 0.0034554668236523867, + 0.0021200766786932945, + 0.004899461753666401, + 0.007772677578032017, + 0.0032120002433657646, + 0.003708508098497987, + 0.0042799473740160465, + 0.00362088973633945, + 0.004809472244232893, + 0.0044721998274326324, + 0.004059586673974991, + 0.008580800145864487, + 0.008620955049991608, + 0.010796861723065376, + 0.010396566241979599, + 0.007373277097940445, + 0.011355319991707802, + 0.010777458548545837, + 0.016913849860429764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.018198074772953987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34861382842063904 + ], + "opposite_logprob_tokens": [ + -1.2236138582229614 + ], + "logprob_reward": -0.34861382842063904, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34861382842063904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1196928545832634, + 0.04407334327697754, + 0.014246241189539433, + 0.044641636312007904, + 0.521181046962738, + 0.14649656414985657, + 0.04116522893309593, + 0.04096774384379387, + 0.013756908476352692, + 0.013778389431536198 + ], + "credit_map": [ + 0.011835007928311825, + 0.01251711044460535, + 0.013340839184820652, + 0.014453181996941566, + 0.011096138507127762, + 0.009834337048232555, + 0.016645682975649834, + 0.029970571398735046, + 0.008655539713799953, + 0.00554400309920311, + 0.004381328355520964, + 0.00608366122469306, + 0.005514195654541254, + 0.004683102015405893, + 0.004560597240924835, + 0.004650919232517481, + 0.0015313650947064161, + 0.0013635955983772874, + 0.0023128336761146784, + 0.002132260240614414, + 0.0014499039389193058, + 0.0013759833527728915, + 0.002121944911777973, + 0.001958355074748397, + 0.00508395629003644, + 0.003091888502240181, + 0.0043180654756724834, + 0.012695347890257835, + 0.0024434456136077642, + 0.008734606206417084, + 0.004581076093018055, + 0.003693254664540291, + 0.037399087101221085, + 0.02726481296122074, + 0.03696074336767197, + 0.16357257962226868, + 0.02572844922542572, + 0.16193325817584991, + 0.03499634191393852, + 0.033325787633657455, + 0.011819876730442047, + 0.008649628609418869, + 0.010841649025678635, + 0.047591496258974075, + 0.008259228430688381, + 0.042692963033914566, + 0.008549054153263569, + 0.008092685602605343, + 0.0050138989463448524, + 0.004856693092733622, + 0.005095981992781162, + 0.003495346289128065, + 0.004696309566497803, + 0.004669006448239088, + 0.005002886056900024, + 0.008335111662745476, + 0.006822316441684961, + 0.005849455017596483, + 0.004367035813629627, + 0.00424509821459651, + 0.0045068878680467606, + 0.00427604652941227, + 0.003470304422080517, + 0.007430601865053177, + 0.0017737161833792925, + 0.0013722238363698125, + 0.001838432508520782, + 0.0016651269979774952, + 0.0016899271868169308, + 0.0016008664388209581, + 0.0013206342700868845, + 0.0024959822185337543, + 0.0017720693722367287, + 0.0014593580272048712, + 0.0017668476793915033, + 0.0016327211633324623, + 0.000999363255687058, + 0.001724875532090664, + 0.0017771711573004723, + 0.0026459842920303345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.34861382842063904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, medium-close-up, static shot focuses horizontally on an elderly haenyeo, clad in a tight, traditional black rubber diving suit. The form-fitting material reflects a faint sheen in the light, still stained with a hint of salt. She sits on a patch of black volcanic rock, smoothed by the sea. Behind her is the damp mudflats left by the slowly receding tide and the shimmering, calm sea. The golden evening sun, streaming down at a low angle, casts a soft shimmer on her graying, wet bun and shoulders. Her face, a product of time and the sea, is a map of deep wrinkles across her tanned skin, yet her eyes are remarkably clear and luminous. She speaks to the side, a warm, proud smile on her face. Her calloused hands rest naturally on her knees, occasionally lifting in a small, powerful gesture as she speaks. The entire picture is full of rich film grain and extremely sharp details. From every deep wrinkle on her cheek to the worn texture of her diving suit, they are clearly visible. The perfect light accurately captures a flicker in her deep and intelligent eyes, creating an interview atmosphere that is both real and intimate as well as nostalgic and poetic.", + "seed": 18029, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_5/index_6069_seed_18029_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12556634657084942, + "vlm_logprob": -0.12556634657084942, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.12556634657084942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08298010379076004 + ], + "opposite_logprob_tokens": [ + -2.582980155944824 + ], + "logprob_reward": -0.08298010379076004, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08298010379076004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1459515392780304, + 0.07662206888198853, + 0.08071480691432953, + 0.05324038863182068, + 0.06853830814361572, + 0.0717877671122551, + 0.15840932726860046, + 0.10185075551271439, + 0.07314107567071915, + 0.16974394023418427 + ], + "credit_map": [ + 0.015401198528707027, + 0.015806330367922783, + 0.018666153773665428, + 0.013170791789889336, + 0.028596017509698868, + 0.012040668167173862, + 0.019091645255684853, + 0.02317873015999794, + 0.010789458639919758, + 0.010802678763866425, + 0.006880050525069237, + 0.007134731393307447, + 0.017305854707956314, + 0.007022171281278133, + 0.008035881444811821, + 0.0086512491106987, + 0.015029697678983212, + 0.01569078490138054, + 0.014226018451154232, + 0.008210713043808937, + 0.0081167072057724, + 0.005432254634797573, + 0.005412545055150986, + 0.008596086874604225, + 0.004786977078765631, + 0.00482823746278882, + 0.006015668623149395, + 0.006799129769206047, + 0.010930164717137814, + 0.004673733841627836, + 0.008546271361410618, + 0.006660205312073231, + 0.009510282427072525, + 0.010218966752290726, + 0.009608487598598003, + 0.005645911209285259, + 0.007472067605704069, + 0.007302329409867525, + 0.0077305505983531475, + 0.01104971393942833, + 0.007413751911371946, + 0.012096203863620758, + 0.009385790675878525, + 0.00612685177475214, + 0.009874416515231133, + 0.009180291555821896, + 0.00907198153436184, + 0.008638482540845871, + 0.018655406311154366, + 0.01879422552883625, + 0.015740135684609413, + 0.02195977047085762, + 0.03287920728325844, + 0.015049386769533157, + 0.018634241074323654, + 0.01669694483280182, + 0.0126936761662364, + 0.017374271526932716, + 0.012739906087517738, + 0.008561828173696995, + 0.015160603448748589, + 0.009308844804763794, + 0.0138790188357234, + 0.012132602743804455, + 0.011765602976083755, + 0.012702791951596737, + 0.007316296920180321, + 0.005220281891524792, + 0.013895364478230476, + 0.005947519559413195, + 0.00729124341160059, + 0.009001978673040867, + 0.054125797003507614, + 0.07169303297996521, + 0.006430481560528278, + 0.005852216389030218, + 0.009201484732329845, + 0.007538549602031708, + 0.005316959228366613, + 0.009585421532392502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.08298010379076004 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005269092041999102 + ], + "opposite_logprob_tokens": [ + -5.2552690505981445 + ], + "logprob_reward": -0.005269092041999102, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005269092041999102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05379811301827431, + 0.029095813632011414, + 0.08339846134185791, + 0.01215332094579935, + 0.08227240294218063, + 0.07836456596851349, + 0.021288104355335236, + 0.02349570393562317, + 0.02101871185004711, + 0.5951148271560669 + ], + "credit_map": [ + 0.009734426625072956, + 0.011125193908810616, + 0.006091144401580095, + 0.003944788593798876, + 0.005261018872261047, + 0.0071850912645459175, + 0.004015197977423668, + 0.006441249046474695, + 0.004551010672003031, + 0.0050077592022717, + 0.0036176368594169617, + 0.0019365904154255986, + 0.004379199352115393, + 0.0027461566496640444, + 0.0034848032519221306, + 0.0033726568799465895, + 0.012219064868986607, + 0.011359651573002338, + 0.008058473467826843, + 0.0049620079807937145, + 0.018319882452487946, + 0.005252755247056484, + 0.0043657878413796425, + 0.0188608355820179, + 0.0011951288906857371, + 0.0014366705436259508, + 0.0020314299035817385, + 0.0013053208822384477, + 0.0018386113224551082, + 0.0015527780633419752, + 0.001198955811560154, + 0.0015944251790642738, + 0.02881615050137043, + 0.03118601255118847, + 0.004403617233037949, + 0.002438519150018692, + 0.003095106454566121, + 0.0030247285030782223, + 0.0018882561707869172, + 0.007420011796057224, + 0.01600872352719307, + 0.01636953465640545, + 0.015064801089465618, + 0.004659607075154781, + 0.006059146020561457, + 0.006394416093826294, + 0.004710814915597439, + 0.009097526781260967, + 0.0021124742925167084, + 0.002770338673144579, + 0.0027031295467168093, + 0.002362918108701706, + 0.004349834285676479, + 0.002104715211316943, + 0.002586115151643753, + 0.00229857861995697, + 0.002091961679980159, + 0.0024622913915663958, + 0.002849506912752986, + 0.0026137467939406633, + 0.00455447006970644, + 0.0031885269563645124, + 0.002620843704789877, + 0.0031143566593527794, + 0.0028090043924748898, + 0.0032287321519106627, + 0.0035072334576398134, + 0.0017564492300152779, + 0.003773293225094676, + 0.002059316961094737, + 0.0018892516382038593, + 0.0019954312592744827, + 0.1990153044462204, + 0.26318061351776123, + 0.016703417524695396, + 0.010445692576467991, + 0.029708394780755043, + 0.02779621258378029, + 0.016280056908726692, + 0.03198510780930519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.005269092041999102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644304238259792 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644304238259792, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644304238259792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03524142876267433, + 0.027581673115491867, + 0.1303938925266266, + 0.011890112422406673, + 0.08418894559144974, + 0.014990076422691345, + 0.01551886647939682, + 0.019543636590242386, + 0.03218964487314224, + 0.6284616589546204 + ], + "credit_map": [ + 0.004569604992866516, + 0.004706370644271374, + 0.005315784830600023, + 0.004044480621814728, + 0.004308303352445364, + 0.004590155556797981, + 0.004051716532558203, + 0.00365501269698143, + 0.003582623088732362, + 0.003463772824034095, + 0.0022642246913164854, + 0.002635802375152707, + 0.00448803324252367, + 0.0024317018687725067, + 0.0044941529631614685, + 0.004221362993121147, + 0.008645854890346527, + 0.010733542963862419, + 0.007814173586666584, + 0.008281642571091652, + 0.0413973368704319, + 0.00993398018181324, + 0.009706079959869385, + 0.0338812954723835, + 0.001151676056906581, + 0.0012721503153443336, + 0.0014450181042775512, + 0.001388015691190958, + 0.0020269686356186867, + 0.0015667738625779748, + 0.0016052997671067715, + 0.0014342092908918858, + 0.032562099397182465, + 0.03179709240794182, + 0.003444535192102194, + 0.0023094359785318375, + 0.0034711286425590515, + 0.0032830059062689543, + 0.0021620020270347595, + 0.005159646272659302, + 0.0023415936157107353, + 0.0033728540875017643, + 0.0019614261109381914, + 0.0007587624131701887, + 0.0019372812239453197, + 0.002246381249278784, + 0.0009723922703415155, + 0.001399385160766542, + 0.0022132068406790495, + 0.0020762737840414047, + 0.0018071415834128857, + 0.001569311018101871, + 0.002972224960103631, + 0.0015184673247858882, + 0.001515421667136252, + 0.0018468204652890563, + 0.0018938444554805756, + 0.002513672923669219, + 0.002843961352482438, + 0.0019236539956182241, + 0.0035864447709172964, + 0.002369905589148402, + 0.0023979204706847668, + 0.00201423279941082, + 0.007446177303791046, + 0.008839697577059269, + 0.002425784943625331, + 0.0016127544222399592, + 0.00518867326900363, + 0.0020621202420443296, + 0.0022863075137138367, + 0.0023281285539269447, + 0.2106924057006836, + 0.27762192487716675, + 0.017434459179639816, + 0.01094137504696846, + 0.031442396342754364, + 0.029392385855317116, + 0.017258398234844208, + 0.03367830812931061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.008644304238259792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076298960484564304 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.0076298960484564304, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076298960484564304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03468945249915123, + 0.032956480979919434, + 0.03168684244155884, + 0.014348690398037434, + 0.03147140145301819, + 0.014994736760854721, + 0.010077799670398235, + 0.029697051271796227, + 0.026272611692547798, + 0.7738049626350403 + ], + "credit_map": [ + 0.0053390213288366795, + 0.003388790413737297, + 0.004862549714744091, + 0.0027356697246432304, + 0.0054558380506932735, + 0.0038236321415752172, + 0.004776595626026392, + 0.0043073552660644054, + 0.004203116986900568, + 0.005705325398594141, + 0.002801887458190322, + 0.003145081689581275, + 0.006206774152815342, + 0.003367741359397769, + 0.0034913381095975637, + 0.004035213962197304, + 0.002207815879955888, + 0.0019804774783551693, + 0.002409605775028467, + 0.0018194337608292699, + 0.00951414555311203, + 0.0023836640175431967, + 0.0019171064486727118, + 0.009454592131078243, + 0.0013527798000723124, + 0.0010958814527839422, + 0.0020918312948197126, + 0.0019042821368202567, + 0.002316659316420555, + 0.0016675549559295177, + 0.001924842712469399, + 0.0019948570989072323, + 0.009247167967259884, + 0.009205205366015434, + 0.0025348595809191465, + 0.001298869145102799, + 0.0020332415588200092, + 0.0025066924281418324, + 0.0017262797337025404, + 0.0029190820641815662, + 0.0026376694440841675, + 0.003741932101547718, + 0.001888992264866829, + 0.0008886329596862197, + 0.0018937025452032685, + 0.001138573745265603, + 0.0010300303110852838, + 0.0017752011772245169, + 0.0011610977817326784, + 0.0011437609791755676, + 0.00126769975759089, + 0.000847638351842761, + 0.0019190331222489476, + 0.0014705293579027057, + 0.0011297577293589711, + 0.0011382820084691048, + 0.004587088245898485, + 0.006948932074010372, + 0.002221103757619858, + 0.0027290484867990017, + 0.004734965972602367, + 0.0034557953476905823, + 0.0026135214138776064, + 0.0024065945763140917, + 0.004773441702127457, + 0.004823619034141302, + 0.0036138815339654684, + 0.003309039631858468, + 0.0037431835662573576, + 0.0017221798188984394, + 0.0020025211852043867, + 0.002284742658957839, + 0.25968435406684875, + 0.3421030640602112, + 0.021552277728915215, + 0.013389399275183678, + 0.038498517125844955, + 0.03604539856314659, + 0.02135242521762848, + 0.04117947071790695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0076298960484564304 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233083367347717 + ], + "opposite_logprob_tokens": [ + -0.8983083367347717 + ], + "logprob_reward": -0.5233083367347717, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233083367347717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050833091139793396, + 0.026253266260027885, + 0.07289128750562668, + 0.00851520337164402, + 0.1417563110589981, + 0.007919009774923325, + 0.016173871234059334, + 0.014875051565468311, + 0.026792602613568306, + 0.6339903473854065 + ], + "credit_map": [ + 0.00764934066683054, + 0.00948068406432867, + 0.005247196648269892, + 0.00459040142595768, + 0.006477872841060162, + 0.007866494357585907, + 0.004157797433435917, + 0.005363308824598789, + 0.0035512205213308334, + 0.0030997965950518847, + 0.0035899998620152473, + 0.002467792946845293, + 0.0039063189178705215, + 0.0024275979958474636, + 0.0031157145276665688, + 0.004094826523214579, + 0.004886312410235405, + 0.005896548740565777, + 0.005421760492026806, + 0.002296188147738576, + 0.023439329117536545, + 0.0057078879326581955, + 0.004679554607719183, + 0.02056371420621872, + 0.0010023596696555614, + 0.0008440047386102378, + 0.0010177764343097806, + 0.0011721192859113216, + 0.0013019859325140715, + 0.0009168382966890931, + 0.0012032940285280347, + 0.0010568250436335802, + 0.05107796564698219, + 0.054149191826581955, + 0.005751750431954861, + 0.0036058654077351093, + 0.0056609176099300385, + 0.0044874222949147224, + 0.0035905512049794197, + 0.013432648964226246, + 0.0013039724435657263, + 0.0016610630555078387, + 0.001168272690847516, + 0.0006870509823784232, + 0.0007086509722284973, + 0.0006970374379307032, + 0.0005642119795084, + 0.0011287510860711336, + 0.00256897509098053, + 0.0021883791778236628, + 0.0016786815831437707, + 0.002658648183569312, + 0.002153603592887521, + 0.0012787163723260164, + 0.0019160208757966757, + 0.0017308464739471674, + 0.0023004356771707535, + 0.0024292548187077045, + 0.0021003051660954952, + 0.002279258333146572, + 0.0015026985201984644, + 0.0011438921792432666, + 0.0013511196011677384, + 0.0017680884338915348, + 0.0068011595867574215, + 0.007453322410583496, + 0.0018562105251476169, + 0.001707893330603838, + 0.003563212463632226, + 0.0017554035875946283, + 0.001822528662160039, + 0.0018328748410567641, + 0.21215909719467163, + 0.2802623212337494, + 0.01776544377207756, + 0.010465356521308422, + 0.03174015134572983, + 0.029775306582450867, + 0.017462221905589104, + 0.03436049818992615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.5233083367347717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, medium-close-up, static shot focuses horizontally on an elderly haenyeo, clad in a tight, traditional black rubber diving suit. The form-fitting material reflects a faint sheen in the light, still stained with a hint of salt. She sits on a patch of black volcanic rock, smoothed by the sea. Behind her is the damp mudflats left by the slowly receding tide and the shimmering, calm sea. The golden evening sun, streaming down at a low angle, casts a soft shimmer on her graying, wet bun and shoulders. Her face, a product of time and the sea, is a map of deep wrinkles across her tanned skin, yet her eyes are remarkably clear and luminous. She speaks to the side, a warm, proud smile on her face. Her calloused hands rest naturally on her knees, occasionally lifting in a small, powerful gesture as she speaks. The entire picture is full of rich film grain and extremely sharp details. From every deep wrinkle on her cheek to the worn texture of her diving suit, they are clearly visible. The perfect light accurately captures a flicker in her deep and intelligent eyes, creating an interview atmosphere that is both real and intimate as well as nostalgic and poetic.", + "seed": 18030, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_5/index_6069_seed_18030_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13196192793548106, + "vlm_logprob": -0.13196192793548106, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.13196192793548106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09740554541349411 + ], + "opposite_logprob_tokens": [ + -2.472405433654785 + ], + "logprob_reward": -0.09740554541349411, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09740554541349411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11687781661748886, + 0.058932576328516006, + 0.0674508810043335, + 0.09728012979030609, + 0.0648261308670044, + 0.045571208000183105, + 0.14868170022964478, + 0.07576002925634384, + 0.027190856635570526, + 0.2974286675453186 + ], + "credit_map": [ + 0.008773457258939743, + 0.005723142065107822, + 0.011468170210719109, + 0.009185288101434708, + 0.03670516982674599, + 0.011628331616520882, + 0.013964195735752583, + 0.01943005435168743, + 0.007897911593317986, + 0.004572599660605192, + 0.003132039215415716, + 0.002942665247246623, + 0.009124891832470894, + 0.004160306416451931, + 0.012310698628425598, + 0.014791463501751423, + 0.0050097559578716755, + 0.00384279596619308, + 0.003729018848389387, + 0.004624552559107542, + 0.011729472316801548, + 0.00713289575651288, + 0.017752952873706818, + 0.013629438355565071, + 0.013889964669942856, + 0.010163286700844765, + 0.010840751230716705, + 0.00912720337510109, + 0.01836635358631611, + 0.009560640901327133, + 0.012542495504021645, + 0.012789440341293812, + 0.0037430929951369762, + 0.0031257453374564648, + 0.010059176944196224, + 0.00523093156516552, + 0.01030020322650671, + 0.011704254895448685, + 0.010872107930481434, + 0.009790614247322083, + 0.004571679513901472, + 0.0023325763177126646, + 0.00805277843028307, + 0.004209408536553383, + 0.0077604674734175205, + 0.004179348237812519, + 0.008126802742481232, + 0.006338144652545452, + 0.006259606685489416, + 0.009466415271162987, + 0.004769702907651663, + 0.00638612313196063, + 0.014866719953715801, + 0.019350893795490265, + 0.05354905128479004, + 0.03403318673372269, + 0.004073970019817352, + 0.003257239470258355, + 0.006577757652848959, + 0.00475238636136055, + 0.006673585623502731, + 0.011141381226480007, + 0.01064015831798315, + 0.028643548488616943, + 0.0034137226175516844, + 0.0012470560614019632, + 0.002090955851599574, + 0.0019148248247802258, + 0.0030859485268592834, + 0.003280077828094363, + 0.006078311707824469, + 0.006079956889152527, + 0.017130261287093163, + 0.007275511510670185, + 0.022423263639211655, + 0.03164113685488701, + 0.0416160449385643, + 0.017354443669319153, + 0.10381001234054565, + 0.05617799609899521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.09740554541349411 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009806433692574501 + ], + "opposite_logprob_tokens": [ + -4.6348066329956055 + ], + "logprob_reward": -0.009806433692574501, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009806433692574501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14078684151172638, + 0.16821064054965973, + 0.09935130923986435, + 0.22101210057735443, + 0.06431816518306732, + 0.06872935593128204, + 0.08450888842344284, + 0.041193582117557526, + 0.06128714233636856, + 0.05060199275612831 + ], + "credit_map": [ + 0.01855221949517727, + 0.014084363356232643, + 0.031692635267972946, + 0.012500452809035778, + 0.019921137019991875, + 0.00912526622414589, + 0.019057000055909157, + 0.015853771939873695, + 0.035203445702791214, + 0.019089657813310623, + 0.015119423158466816, + 0.008686328306794167, + 0.019183596596121788, + 0.007578947115689516, + 0.018478821963071823, + 0.04487042501568794, + 0.020771365612745285, + 0.01059942040592432, + 0.023471249267458916, + 0.005540380720049143, + 0.013560664840042591, + 0.004694545175880194, + 0.009679179638624191, + 0.01103450357913971, + 0.040308207273483276, + 0.01776237227022648, + 0.03634992986917496, + 0.013317291624844074, + 0.028021663427352905, + 0.011175684630870819, + 0.028475675731897354, + 0.045601293444633484, + 0.005911264568567276, + 0.009293098002672195, + 0.008355283178389072, + 0.002749646082520485, + 0.005877257790416479, + 0.012170098721981049, + 0.004173068329691887, + 0.015788452699780464, + 0.003934422507882118, + 0.002345689805224538, + 0.003309286665171385, + 0.0023217336274683475, + 0.0063009257428348064, + 0.007012610789388418, + 0.018359052017331123, + 0.0251456331461668, + 0.005769859068095684, + 0.007334491237998009, + 0.007701240945607424, + 0.005101323593407869, + 0.010573829524219036, + 0.011730311438441277, + 0.023304635658860207, + 0.012993194162845612, + 0.003677049884572625, + 0.00237983837723732, + 0.009413925930857658, + 0.0036937131080776453, + 0.004017191007733345, + 0.009627064689993858, + 0.00368102896027267, + 0.004703771788626909, + 0.005518454592674971, + 0.003190193558111787, + 0.005247240420430899, + 0.0042130667716264725, + 0.007302211597561836, + 0.01695152372121811, + 0.011358553543686867, + 0.007505899295210838, + 0.003048646030947566, + 0.0013440719339996576, + 0.003857160685583949, + 0.006023205816745758, + 0.007942798547446728, + 0.0032891505397856236, + 0.016619950532913208, + 0.008477007038891315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.009806433692574501 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199244514107704 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.018199244514107704, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199244514107704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14913608133792877, + 0.19650904834270477, + 0.08737464994192123, + 0.06040916591882706, + 0.06552784144878387, + 0.13257303833961487, + 0.08926799893379211, + 0.10935037583112717, + 0.04630785435438156, + 0.06354393064975739 + ], + "credit_map": [ + 0.02135905623435974, + 0.013559037819504738, + 0.015909014269709587, + 0.011435498483479023, + 0.021813485771417618, + 0.020389322191476822, + 0.023817826062440872, + 0.020852843299508095, + 0.016238924115896225, + 0.008012850768864155, + 0.011667124927043915, + 0.00935649685561657, + 0.015317454002797604, + 0.007656234782189131, + 0.06695406883955002, + 0.06130589544773102, + 0.012237229384481907, + 0.009604097343981266, + 0.007510886527597904, + 0.007307330146431923, + 0.01431773230433464, + 0.004578068386763334, + 0.018310926854610443, + 0.013508379459381104, + 0.011282349936664104, + 0.0059745595790445805, + 0.009952000342309475, + 0.003939024172723293, + 0.009723088704049587, + 0.0049639916978776455, + 0.0076590306125581264, + 0.006915120407938957, + 0.0075979409739375114, + 0.007661046925932169, + 0.010785120539367199, + 0.004344005137681961, + 0.009387626312673092, + 0.004081366583704948, + 0.00950133427977562, + 0.012169402092695236, + 0.011079536750912666, + 0.011913006193935871, + 0.006589103955775499, + 0.006387393921613693, + 0.019378580152988434, + 0.014910602942109108, + 0.018304968252778053, + 0.04400984197854996, + 0.005012043286114931, + 0.004175442270934582, + 0.008349692448973656, + 0.004978738259524107, + 0.012581122107803822, + 0.012862484902143478, + 0.019632969051599503, + 0.021675508469343185, + 0.007160299923270941, + 0.00392949627712369, + 0.017280466854572296, + 0.007794137112796307, + 0.008601019158959389, + 0.006679176352918148, + 0.017035698518157005, + 0.04087008535861969, + 0.0042882924899458885, + 0.0023705032654106617, + 0.0039574624970555305, + 0.003972603473812342, + 0.007348217535763979, + 0.007844824343919754, + 0.007588024251163006, + 0.008937924169003963, + 0.0038927877321839333, + 0.0021066591143608093, + 0.004776368848979473, + 0.007837536744773388, + 0.009509450756013393, + 0.005612643901258707, + 0.018815696239471436, + 0.010992784984409809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.018199244514107704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083701625466347 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083701625466347, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083701625466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08322899788618088, + 0.10607762634754181, + 0.08523285388946533, + 0.05855756625533104, + 0.09125051647424698, + 0.07414061576128006, + 0.04597293213009834, + 0.09009309858083725, + 0.22972092032432556, + 0.13572487235069275 + ], + "credit_map": [ + 0.007736158091574907, + 0.00491517735645175, + 0.008602690882980824, + 0.00880508590489626, + 0.02136625349521637, + 0.008828038349747658, + 0.008903010748326778, + 0.014072583056986332, + 0.020414356142282486, + 0.011386268772184849, + 0.01598675549030304, + 0.006346428766846657, + 0.015199001878499985, + 0.004306551069021225, + 0.017675982788205147, + 0.014762280508875847, + 0.010996104218065739, + 0.006470289081335068, + 0.013707500882446766, + 0.004128258675336838, + 0.01050618477165699, + 0.005491981282830238, + 0.011724922806024551, + 0.022207606583833694, + 0.011476156301796436, + 0.005580665078014135, + 0.009471961297094822, + 0.003876518225297332, + 0.009319020435214043, + 0.00432741641998291, + 0.006602718494832516, + 0.00790311023592949, + 0.009701570495963097, + 0.009043535217642784, + 0.018518278375267982, + 0.004861258901655674, + 0.011023029685020447, + 0.007091626990586519, + 0.01126314140856266, + 0.01974807120859623, + 0.004793256055563688, + 0.007447330746799707, + 0.010525992140173912, + 0.005805978551506996, + 0.010239381343126297, + 0.010773470625281334, + 0.008911505341529846, + 0.01564370095729828, + 0.005187136121094227, + 0.00427494291216135, + 0.006813921499997377, + 0.0030536712147295475, + 0.006908608600497246, + 0.0069237882271409035, + 0.007127473130822182, + 0.00568339042365551, + 0.004267397336661816, + 0.003968993667513132, + 0.007533444557338953, + 0.006238359957933426, + 0.011676665395498276, + 0.0201053936034441, + 0.0105287479236722, + 0.025774097070097923, + 0.031210746616125107, + 0.013571557588875294, + 0.019444458186626434, + 0.016978541389107704, + 0.03187922388315201, + 0.054356202483177185, + 0.035266246646642685, + 0.027013933286070824, + 0.007924619130790234, + 0.002578800544142723, + 0.008060178719460964, + 0.014295890927314758, + 0.019235022366046906, + 0.005387033335864544, + 0.05228973925113678, + 0.025953590869903564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.011083701625466347 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233147144317627 + ], + "opposite_logprob_tokens": [ + -0.8983147144317627 + ], + "logprob_reward": -0.5233147144317627, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233147144317627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07816623151302338, + 0.14569203555583954, + 0.12249968945980072, + 0.0676821768283844, + 0.08309929072856903, + 0.2889770567417145, + 0.07798472791910172, + 0.059205666184425354, + 0.05373897776007652, + 0.022954154759645462 + ], + "credit_map": [ + 0.007103912066668272, + 0.006661888211965561, + 0.012673556804656982, + 0.005314396694302559, + 0.017865819856524467, + 0.010598636232316494, + 0.007791426498442888, + 0.010156596079468727, + 0.009560972452163696, + 0.004556142725050449, + 0.006745166145265102, + 0.006799452006816864, + 0.013729642145335674, + 0.006439161952584982, + 0.051673367619514465, + 0.04618813842535019, + 0.011935718357563019, + 0.005413740407675505, + 0.025063714012503624, + 0.005844707600772381, + 0.011815347708761692, + 0.006186836864799261, + 0.015945013612508774, + 0.04029460996389389, + 0.015295888297259808, + 0.005748718976974487, + 0.010629437863826752, + 0.003491502720862627, + 0.00885451678186655, + 0.003286835039034486, + 0.007399227935820818, + 0.012976044788956642, + 0.004176854621618986, + 0.010170314460992813, + 0.004955512471497059, + 0.0030023399740457535, + 0.008989381603896618, + 0.023665370419621468, + 0.007130875717848539, + 0.02100863866508007, + 0.020883802324533463, + 0.013040628284215927, + 0.035978175699710846, + 0.01703125610947609, + 0.0323888324201107, + 0.01308322325348854, + 0.07600127905607224, + 0.08056984096765518, + 0.004181345459073782, + 0.003416260937228799, + 0.0035934792831540108, + 0.003741720924153924, + 0.011312486603856087, + 0.011630810797214508, + 0.02552678994834423, + 0.01458183117210865, + 0.0036682693753391504, + 0.0027806146536022425, + 0.010570729151368141, + 0.0071372962556779385, + 0.007946280762553215, + 0.010481503792107105, + 0.007011164911091328, + 0.00960980262607336, + 0.006659837905317545, + 0.0025377245619893074, + 0.003828437067568302, + 0.00415791617706418, + 0.0070757209323346615, + 0.01240137405693531, + 0.007302805781364441, + 0.009775160811841488, + 0.0023748718667775393, + 0.0011057066731154919, + 0.0029174452647566795, + 0.0031067486852407455, + 0.005130511708557606, + 0.002029673894867301, + 0.0029853014275431633, + 0.003303894307464361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.5233147144317627 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, medium-close-up, static shot focuses horizontally on an elderly haenyeo, clad in a tight, traditional black rubber diving suit. The form-fitting material reflects a faint sheen in the light, still stained with a hint of salt. She sits on a patch of black volcanic rock, smoothed by the sea. Behind her is the damp mudflats left by the slowly receding tide and the shimmering, calm sea. The golden evening sun, streaming down at a low angle, casts a soft shimmer on her graying, wet bun and shoulders. Her face, a product of time and the sea, is a map of deep wrinkles across her tanned skin, yet her eyes are remarkably clear and luminous. She speaks to the side, a warm, proud smile on her face. Her calloused hands rest naturally on her knees, occasionally lifting in a small, powerful gesture as she speaks. The entire picture is full of rich film grain and extremely sharp details. From every deep wrinkle on her cheek to the worn texture of her diving suit, they are clearly visible. The perfect light accurately captures a flicker in her deep and intelligent eyes, creating an interview atmosphere that is both real and intimate as well as nostalgic and poetic.", + "seed": 18031, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_5/index_6069_seed_18031_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11045024152845144, + "vlm_logprob": -0.11045024152845144, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.11045024152845144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1320897936820984 + ], + "opposite_logprob_tokens": [ + -2.132089853286743 + ], + "logprob_reward": -0.1320897936820984, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1320897936820984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08104727417230606, + 0.06089768558740616, + 0.43666529655456543, + 0.05646257475018501, + 0.03875410556793213, + 0.03721173480153084, + 0.040422748774290085, + 0.12886793911457062, + 0.0864759311079979, + 0.03319466859102249 + ], + "credit_map": [ + 0.009080672636628151, + 0.010155224241316319, + 0.010849973186850548, + 0.00894505437463522, + 0.014437262900173664, + 0.007866086438298225, + 0.011053565889596939, + 0.008659430779516697, + 0.01094422210007906, + 0.005523018538951874, + 0.006220691371709108, + 0.004649898037314415, + 0.006517388392239809, + 0.006207659374922514, + 0.004963381681591272, + 0.015871426090598106, + 0.05460774526000023, + 0.062169548124074936, + 0.06298124045133591, + 0.05156697332859039, + 0.059035319834947586, + 0.05565657094120979, + 0.05111080780625343, + 0.039537087082862854, + 0.00820580217987299, + 0.0045083132572472095, + 0.008445960469543934, + 0.006022301036864519, + 0.005258030723780394, + 0.005442039575427771, + 0.007030199281871319, + 0.011549929156899452, + 0.005053549073636532, + 0.0039939843118190765, + 0.009159935638308525, + 0.0041086263954639435, + 0.004282495938241482, + 0.004193259868770838, + 0.0038643418811261654, + 0.004097912460565567, + 0.003779847640544176, + 0.0038768756203353405, + 0.009006076492369175, + 0.0053453086875379086, + 0.003529758658260107, + 0.004133070819079876, + 0.003241518745198846, + 0.004299275577068329, + 0.0064735449850559235, + 0.0020751317497342825, + 0.005858500022441149, + 0.0038561581168323755, + 0.002710319124162197, + 0.0025818634312599897, + 0.004763387143611908, + 0.012103844434022903, + 0.014970656484365463, + 0.006043977104127407, + 0.01036074385046959, + 0.011902698315680027, + 0.006100354716181755, + 0.006665400695055723, + 0.02235441468656063, + 0.05046970397233963, + 0.016742292791604996, + 0.010032682679593563, + 0.00849448423832655, + 0.00761011615395546, + 0.004222281277179718, + 0.008158477023243904, + 0.00685151806101203, + 0.02436407469213009, + 0.004208933562040329, + 0.003516600700095296, + 0.0052366480231285095, + 0.002943842438980937, + 0.004325909540057182, + 0.0056766788475215435, + 0.0030595287680625916, + 0.00422652717679739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.1320897936820984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01109513733536005 + ], + "opposite_logprob_tokens": [ + -4.51109504699707 + ], + "logprob_reward": -0.01109513733536005, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01109513733536005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14255213737487793, + 0.15935096144676208, + 0.13487158715724945, + 0.03515639901161194, + 0.03677397593855858, + 0.03656455874443054, + 0.07129229605197906, + 0.29309728741645813, + 0.060740355402231216, + 0.02960042841732502 + ], + "credit_map": [ + 0.01673014834523201, + 0.01851828396320343, + 0.03365708887577057, + 0.014827229082584381, + 0.009552866220474243, + 0.011108510196208954, + 0.012710610404610634, + 0.025447409600019455, + 0.022332919761538506, + 0.01843998394906521, + 0.033244404941797256, + 0.015240037813782692, + 0.008664148859679699, + 0.01227812934666872, + 0.009142672643065453, + 0.040008679032325745, + 0.015217363834381104, + 0.014267470687627792, + 0.021294396370649338, + 0.012531773187220097, + 0.014113790355622768, + 0.027750475332140923, + 0.012512780725955963, + 0.01718355156481266, + 0.005100327543914318, + 0.0031910003162920475, + 0.009442182257771492, + 0.0037950887344777584, + 0.002786003053188324, + 0.004013656172901392, + 0.0032872436568140984, + 0.003540900768712163, + 0.003988605458289385, + 0.005136593710631132, + 0.007778521161526442, + 0.0037196376360952854, + 0.0049484469927847385, + 0.0034131307620555162, + 0.004301926586776972, + 0.0034871185198426247, + 0.005208449903875589, + 0.0034604521933943033, + 0.008064723573625088, + 0.004871428944170475, + 0.0035300152376294136, + 0.004568865988403559, + 0.0030666044913232327, + 0.003794018877670169, + 0.013891905546188354, + 0.003509238362312317, + 0.006744743324816227, + 0.004609262105077505, + 0.003485536901280284, + 0.004733051173388958, + 0.008963764645159245, + 0.02535480074584484, + 0.035680949687957764, + 0.009669989347457886, + 0.012526450678706169, + 0.027668092399835587, + 0.009078522212803364, + 0.015468497760593891, + 0.054551150649785995, + 0.12845362722873688, + 0.013149896636605263, + 0.004495234694331884, + 0.0035792775452136993, + 0.006057913415133953, + 0.0026017636992037296, + 0.004217279143631458, + 0.005425913259387016, + 0.021213078871369362, + 0.0045738834887743, + 0.002414248650893569, + 0.008419121615588665, + 0.003171399235725403, + 0.0025286104064434767, + 0.0025128896813839674, + 0.002396379364654422, + 0.0035838973708450794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.01109513733536005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059540290385484695 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.0059540290385484695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059540290385484695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15588687360286713, + 0.11299574375152588, + 0.4315304160118103, + 0.032494157552719116, + 0.03350898250937462, + 0.041733767837285995, + 0.048102859407663345, + 0.09535006433725357, + 0.0369914285838604, + 0.011405764147639275 + ], + "credit_map": [ + 0.020451711490750313, + 0.02585359290242195, + 0.018178541213274002, + 0.01456527691334486, + 0.01607201248407364, + 0.014395258389413357, + 0.018129920586943626, + 0.028240563347935677, + 0.014002935029566288, + 0.012440843507647514, + 0.02087162248790264, + 0.01176142692565918, + 0.010374631732702255, + 0.0114166010171175, + 0.008661056868731976, + 0.023466631770133972, + 0.050421092659235, + 0.05698515847325325, + 0.06630606949329376, + 0.041171152144670486, + 0.06029944494366646, + 0.06478400528430939, + 0.03986571729183197, + 0.05169777199625969, + 0.005224175285547972, + 0.0032540385145694017, + 0.005944824777543545, + 0.003310099244117737, + 0.00223552156239748, + 0.005398108623921871, + 0.003372505772858858, + 0.003754883538931608, + 0.004905307199805975, + 0.004303491208702326, + 0.005211607553064823, + 0.0031805119942873716, + 0.005958458408713341, + 0.0028213285841047764, + 0.0034126536920666695, + 0.0037156243342906237, + 0.005733877886086702, + 0.004585443064570427, + 0.009733596816658974, + 0.00423791166394949, + 0.002539679640904069, + 0.007112307474017143, + 0.003206829307600856, + 0.004584121517837048, + 0.008198010735213757, + 0.002781184855848551, + 0.005142856389284134, + 0.003923104144632816, + 0.0027996820863336325, + 0.003358860732987523, + 0.005880643613636494, + 0.016018519178032875, + 0.010406045243144035, + 0.007957653142511845, + 0.004289240576326847, + 0.008741334080696106, + 0.0037807661574333906, + 0.005146914627403021, + 0.016514163464307785, + 0.03851393982768059, + 0.006515996530652046, + 0.0051557221449911594, + 0.0026528227608650923, + 0.003692689584568143, + 0.003636236535385251, + 0.0031000312883406878, + 0.0025954931043088436, + 0.00964243896305561, + 0.0015777485677972436, + 0.0012442130828276277, + 0.0019060259219259024, + 0.0008807111298665404, + 0.0015939611475914717, + 0.0015119636664167047, + 0.0011174221290275455, + 0.0015737188514322042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0059540290385484695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607106626033783 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.01607106626033783, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607106626033783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15468540787696838, + 0.06986312568187714, + 0.21064744889736176, + 0.09738107770681381, + 0.041696690022945404, + 0.07477615028619766, + 0.08870509266853333, + 0.06336834281682968, + 0.16166767477989197, + 0.03720898553729057 + ], + "credit_map": [ + 0.017671555280685425, + 0.019253158941864967, + 0.027851469814777374, + 0.01534920372068882, + 0.014953789301216602, + 0.016117878258228302, + 0.0163873303681612, + 0.027101019397377968, + 0.008354558609426022, + 0.010844414122402668, + 0.007425555028021336, + 0.007330988068133593, + 0.0077893054112792015, + 0.006015911232680082, + 0.005380784627050161, + 0.016721611842513084, + 0.021918242797255516, + 0.022946758195757866, + 0.03333898261189461, + 0.017206253483891487, + 0.022309165447950363, + 0.04531396925449371, + 0.018850594758987427, + 0.028763489797711372, + 0.014912666752934456, + 0.008890228345990181, + 0.021873749792575836, + 0.009485144168138504, + 0.008632526732981205, + 0.015630247071385384, + 0.008225287310779095, + 0.009731227532029152, + 0.00449003279209137, + 0.005200878717005253, + 0.006557953078299761, + 0.005256164353340864, + 0.005373145453631878, + 0.005265363957732916, + 0.0050666132010519505, + 0.00448653707280755, + 0.01052751112729311, + 0.006741246208548546, + 0.014470791444182396, + 0.008291705511510372, + 0.006949923001229763, + 0.011958622373640537, + 0.007058200892060995, + 0.00877815019339323, + 0.01697648875415325, + 0.004178790375590324, + 0.007606188766658306, + 0.0061009228229522705, + 0.003976527135819197, + 0.00571738975122571, + 0.01149796787649393, + 0.03265082463622093, + 0.0075505697168409824, + 0.0036216925363987684, + 0.006139910314232111, + 0.005900486372411251, + 0.004179094452410936, + 0.004884344059973955, + 0.009399032220244408, + 0.021693209186196327, + 0.022929176688194275, + 0.037419456988573074, + 0.014379416592419147, + 0.012571519240736961, + 0.014438079670071602, + 0.018565595149993896, + 0.009646745398640633, + 0.03171766921877861, + 0.005046140868216753, + 0.0036545852199196815, + 0.008440559729933739, + 0.004228741861879826, + 0.0028978255577385426, + 0.004905571695417166, + 0.003679634304717183, + 0.004355926997959614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.01607106626033783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman wearing a tight, traditional black rubber diving suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman sitting on a patch of black volcanic rock with damp mudflats and a calm sea in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature golden evening sunlight at a low angle illuminating the subject's wet hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman shown speaking with a warm smile, her calloused hands resting on her knees and making small gestures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a static medium-close-up shot with rich film grain and extremely sharp details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870411813259125 + ], + "opposite_logprob_tokens": [ + -1.1370412111282349 + ], + "logprob_reward": -0.3870411813259125, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870411813259125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09461762756109238, + 0.05816172808408737, + 0.21471932530403137, + 0.11973965913057327, + 0.19692231714725494, + 0.034446317702531815, + 0.05846988409757614, + 0.09156523644924164, + 0.10835292190313339, + 0.02300490438938141 + ], + "credit_map": [ + 0.01151889655739069, + 0.008413352072238922, + 0.01662246324121952, + 0.009592588990926743, + 0.015096291899681091, + 0.008385327644646168, + 0.012480726465582848, + 0.01250799372792244, + 0.008075372315943241, + 0.005993610713630915, + 0.009368857368826866, + 0.006576788146048784, + 0.00702913710847497, + 0.005675133317708969, + 0.004225446842610836, + 0.011217384599149227, + 0.026379460468888283, + 0.03299926966428757, + 0.03125828504562378, + 0.023065565153956413, + 0.027776967734098434, + 0.028760235756635666, + 0.022537631914019585, + 0.02194194123148918, + 0.01325814425945282, + 0.010919232852756977, + 0.03194378316402435, + 0.017342615872621536, + 0.007928197272121906, + 0.014164585620164871, + 0.011308470740914345, + 0.012874661013484001, + 0.033070821315050125, + 0.02877851575613022, + 0.016434045508503914, + 0.016516054049134254, + 0.044211823493242264, + 0.020395822823047638, + 0.02074699103832245, + 0.016768261790275574, + 0.00433321250602603, + 0.002512306673452258, + 0.008682145737111568, + 0.0035772223491221666, + 0.0032022418454289436, + 0.00434127589687705, + 0.002948241773992777, + 0.004849673248827457, + 0.007046667858958244, + 0.004016755614429712, + 0.011434298940002918, + 0.0059980591759085655, + 0.005143245216459036, + 0.00833814311772585, + 0.004995075985789299, + 0.011497643776237965, + 0.010256936773657799, + 0.004528975114226341, + 0.003575604408979416, + 0.008617710322141647, + 0.004325174260884523, + 0.0048712738789618015, + 0.01678086258471012, + 0.03860871121287346, + 0.020474808290600777, + 0.01973281428217888, + 0.014118947088718414, + 0.009695925749838352, + 0.007439610548317432, + 0.008381254971027374, + 0.00675679836422205, + 0.021752778440713882, + 0.003247759537771344, + 0.0018628999823704362, + 0.0036706109531223774, + 0.001666493364609778, + 0.0026467221323400736, + 0.004490801598876715, + 0.0021773420739918947, + 0.0032422784715890884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.3870411813259125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_tensor([5840, 5840])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_tensor([5840, 5840])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..919ad10e0615ec6d002897c0a61158c8e23c0247 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_tensor([5840, 5840])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, shot from a level angle and perfectly centered, focuses squarely on a young monk. His shaved head reflects a soft sheen in the warm candlelight. He sits cross-legged on a round, dark brown cushion, clad in a loose, gray cotton and linen robe, its folds flowing naturally along his straight back. His eyes are closed, his long eyelashes casting a subtle shadow on his cheeks. His features are delicate, his lips slightly pursed, his expression serene and focused. His hands are folded before him in the meditation mudra, like a statue, lost in deep meditation. The sanctuary he inhabits is a simple, rustic setting, framed by dark wooden walls and a slightly out-of-focus shoji paper door. Not far from him, a white candle burns quietly on a rustic ceramic candlestick. The flickering orange flame is the only dynamic element and the only light source in the picture. It casts soft and ever-changing light and shadow on the monk's focused face and body outline, separating him from the dark background and creating an atmosphere of extreme tranquility, solemnity and inner strength, as if time has frozen at this moment.", + "seed": 17476, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_5840_seed_17476_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04622963014990091, + "vlm_logprob": -0.04622963014990091, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04622963014990091, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01654617115855217 + ], + "opposite_logprob_tokens": [ + -4.641546249389648 + ], + "logprob_reward": -0.01654617115855217, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01654617115855217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38277286291122437, + 0.08249431848526001, + 0.06193032115697861, + 0.04785262048244476, + 0.06512083113193512, + 0.06741063296794891, + 0.05337480828166008, + 0.06419572979211807, + 0.0994764044880867, + 0.07537142187356949 + ], + "credit_map": [ + 0.06286422908306122, + 0.043645165860652924, + 0.046367719769477844, + 0.03892457112669945, + 0.056713346391916275, + 0.03830660879611969, + 0.05158896744251251, + 0.044362276792526245, + 0.014400370419025421, + 0.00906606949865818, + 0.011041334830224514, + 0.009217408485710621, + 0.01107447873800993, + 0.007739477790892124, + 0.008969990536570549, + 0.010985197499394417, + 0.00826801173388958, + 0.007755336817353964, + 0.007723053451627493, + 0.007141515146940947, + 0.00618848716840148, + 0.0075339823961257935, + 0.006773275788873434, + 0.010546660050749779, + 0.006969677284359932, + 0.0062843444757163525, + 0.00662597082555294, + 0.003989978693425655, + 0.0058943284675478935, + 0.006416962947696447, + 0.005056134890764952, + 0.006615223363041878, + 0.011005771346390247, + 0.004981353413313627, + 0.009906914085149765, + 0.005547371227294207, + 0.006063874810934067, + 0.007663864642381668, + 0.006557992193847895, + 0.013393701985478401, + 0.010978992097079754, + 0.00617353105917573, + 0.011835357174277306, + 0.00598690053448081, + 0.007113640196621418, + 0.006909837946295738, + 0.006127531174570322, + 0.012284846045076847, + 0.007620168849825859, + 0.0062919119372963905, + 0.007551577873528004, + 0.004132542293518782, + 0.005212412681430578, + 0.010149984620511532, + 0.005819452460855246, + 0.006596763618290424, + 0.00825485959649086, + 0.005894858855754137, + 0.007924998179078102, + 0.00580174382776022, + 0.005774632096290588, + 0.00692375935614109, + 0.010714389383792877, + 0.012906492687761784, + 0.016366899013519287, + 0.010910539887845516, + 0.014372380450367928, + 0.008964793756604195, + 0.011541567742824554, + 0.010986712761223316, + 0.012525932863354683, + 0.013807586394250393, + 0.01339268684387207, + 0.008196256123483181, + 0.011099373921751976, + 0.006344955414533615, + 0.007687200326472521, + 0.010343320667743683, + 0.007426508702337742, + 0.010881123133003712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.01654617115855217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357805356383324 + ], + "opposite_logprob_tokens": [ + -3.651357889175415 + ], + "logprob_reward": -0.026357805356383324, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357805356383324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40320253372192383, + 0.13939185440540314, + 0.08683400601148605, + 0.08582249283790588, + 0.04242894425988197, + 0.05540543794631958, + 0.04929827153682709, + 0.042449701577425, + 0.05401081219315529, + 0.04115591198205948 + ], + "credit_map": [ + 0.04058777913451195, + 0.045106191188097, + 0.03997324779629707, + 0.05414082854986191, + 0.04740730673074722, + 0.05035390704870224, + 0.06708481162786484, + 0.058548454195261, + 0.015097592025995255, + 0.013862280175089836, + 0.014127316884696484, + 0.014339527115225792, + 0.018102047964930534, + 0.017657291144132614, + 0.030191341415047646, + 0.016014464199543, + 0.008557327091693878, + 0.012634523212909698, + 0.008587097749114037, + 0.009857740253210068, + 0.010941670276224613, + 0.008550532162189484, + 0.017112672328948975, + 0.01059244479984045, + 0.01205222588032484, + 0.009070200845599174, + 0.013705971650779247, + 0.007525073830038309, + 0.010834277607500553, + 0.010625116527080536, + 0.011082551442086697, + 0.010927075520157814, + 0.006137773394584656, + 0.0037962296046316624, + 0.005758394952863455, + 0.004738685209304094, + 0.006566981319338083, + 0.005165785551071167, + 0.0048734587617218494, + 0.005391636397689581, + 0.009472139179706573, + 0.006676954682916403, + 0.0072893970645964146, + 0.0048337699845433235, + 0.009488115087151527, + 0.004179228097200394, + 0.007073734421283007, + 0.00639209384098649, + 0.008247936144471169, + 0.006264474708586931, + 0.006184189114719629, + 0.0037014831323176622, + 0.005313021596521139, + 0.007487668190151453, + 0.005146571435034275, + 0.006952928379178047, + 0.004153157584369183, + 0.006293566431850195, + 0.00447538448497653, + 0.004819017369300127, + 0.004987139720469713, + 0.00390065205283463, + 0.008787419646978378, + 0.005033363122493029, + 0.007214206736534834, + 0.005522884428501129, + 0.005927735473960638, + 0.004624766297638416, + 0.005732642952352762, + 0.0048576844856143, + 0.01302191149443388, + 0.007108980789780617, + 0.0053907339461147785, + 0.006121598184108734, + 0.004974121693521738, + 0.004211912862956524, + 0.005900226533412933, + 0.0036957489792257547, + 0.005773442331701517, + 0.005088127218186855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026357805356383324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048657048493623734 + ], + "opposite_logprob_tokens": [ + -3.048656940460205 + ], + "logprob_reward": -0.048657048493623734, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048657048493623734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3327682614326477, + 0.1304791420698166, + 0.0855124369263649, + 0.09789833426475525, + 0.0644952729344368, + 0.055552344769239426, + 0.10234029591083527, + 0.05296207219362259, + 0.04639533907175064, + 0.03159649297595024 + ], + "credit_map": [ + 0.042792338877916336, + 0.03734012320637703, + 0.04199189320206642, + 0.02976325899362564, + 0.05767429247498512, + 0.03672640398144722, + 0.03675422444939613, + 0.04972575604915619, + 0.022783061489462852, + 0.011412914842367172, + 0.016918139532208443, + 0.009003357961773872, + 0.018765706568956375, + 0.019829904660582542, + 0.015153489075601101, + 0.016612572595477104, + 0.01753324456512928, + 0.01007169671356678, + 0.012664233334362507, + 0.0071354578249156475, + 0.010509141720831394, + 0.006070598028600216, + 0.010442795231938362, + 0.01108527835458517, + 0.01171413529664278, + 0.011476392857730389, + 0.014216486364603043, + 0.012138858437538147, + 0.012491649016737938, + 0.009910128079354763, + 0.013141512870788574, + 0.01280918437987566, + 0.012919571250677109, + 0.006061915773898363, + 0.009536846540868282, + 0.0047907414846122265, + 0.008450489491224289, + 0.004244720563292503, + 0.007508991751819849, + 0.010981998406350613, + 0.007283566053956747, + 0.005817940458655357, + 0.00705233309417963, + 0.004526602569967508, + 0.009485186077654362, + 0.007263761945068836, + 0.007081945892423391, + 0.00704101100564003, + 0.015904149040579796, + 0.011340196244418621, + 0.017046969383955002, + 0.007513754535466433, + 0.009725104086101055, + 0.02025892771780491, + 0.007499364670366049, + 0.013051837682723999, + 0.00761124724522233, + 0.007445922587066889, + 0.007264292798936367, + 0.00452475156635046, + 0.00601575942710042, + 0.008024973794817924, + 0.004861559718847275, + 0.0072135659866034985, + 0.0067059691064059734, + 0.007937433198094368, + 0.005735809449106455, + 0.004283640533685684, + 0.005039385985583067, + 0.007011754438281059, + 0.004574011079967022, + 0.005107342731207609, + 0.0060336412861943245, + 0.003353593870997429, + 0.004266343545168638, + 0.003463967237621546, + 0.0032727159559726715, + 0.003197635989636183, + 0.003709646174684167, + 0.0042989542707800865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.048657048493623734 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532881461083889 + ], + "opposite_logprob_tokens": [ + -4.387532711029053 + ], + "logprob_reward": -0.012532881461083889, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532881461083889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.278964102268219, + 0.09827563166618347, + 0.10633555054664612, + 0.16990256309509277, + 0.06492672860622406, + 0.06167306378483772, + 0.06954256445169449, + 0.04746459797024727, + 0.05405648052692413, + 0.04885873571038246 + ], + "credit_map": [ + 0.0389329195022583, + 0.03837190195918083, + 0.03038228489458561, + 0.031487077474594116, + 0.0391666553914547, + 0.03636699542403221, + 0.03058037720620632, + 0.03367588296532631, + 0.01524612121284008, + 0.008467422798275948, + 0.013413635082542896, + 0.009541367180645466, + 0.01396301481872797, + 0.011467888019979, + 0.01076831016689539, + 0.015407875180244446, + 0.018026361241936684, + 0.01624196767807007, + 0.013038561679422855, + 0.010219798423349857, + 0.010318334214389324, + 0.018400905653834343, + 0.0073074945248663425, + 0.01278211921453476, + 0.02567330189049244, + 0.019352078437805176, + 0.02430412918329239, + 0.013993576169013977, + 0.01860586740076542, + 0.029696214944124222, + 0.014570244587957859, + 0.023707151412963867, + 0.009177884086966515, + 0.007298004347831011, + 0.010484295897185802, + 0.0049952557310462, + 0.006088170688599348, + 0.007695710752159357, + 0.006226394325494766, + 0.012961010448634624, + 0.008624146692454815, + 0.008675952441990376, + 0.007733819540590048, + 0.0060127596370875835, + 0.006716221570968628, + 0.010183461010456085, + 0.005774896126240492, + 0.007951810024678707, + 0.010537014342844486, + 0.01037745550274849, + 0.008428039960563183, + 0.005959790199995041, + 0.009404826909303665, + 0.008032934740185738, + 0.007628407794982195, + 0.009174098260700703, + 0.006749468855559826, + 0.006312135141342878, + 0.006370441056787968, + 0.004213166423141956, + 0.0053331078961491585, + 0.006458774674683809, + 0.005208577029407024, + 0.006818926427513361, + 0.007198713719844818, + 0.005876094102859497, + 0.006803280673921108, + 0.004574974067509174, + 0.007467824500054121, + 0.005194336175918579, + 0.009501819498836994, + 0.007439437322318554, + 0.00840828288346529, + 0.005970320664346218, + 0.007604314014315605, + 0.003957107197493315, + 0.005654278211295605, + 0.004328783601522446, + 0.00407553231343627, + 0.008860117755830288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012532881461083889 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705424427986145 + ], + "opposite_logprob_tokens": [ + -2.127054214477539 + ], + "logprob_reward": -0.12705424427986145, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705424427986145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36368173360824585, + 0.1304311752319336, + 0.0711265355348587, + 0.12329421937465668, + 0.038223013281822205, + 0.05745507404208183, + 0.0910317599773407, + 0.050634026527404785, + 0.036345988512039185, + 0.03777656331658363 + ], + "credit_map": [ + 0.047287944704294205, + 0.05615532025694847, + 0.046788088977336884, + 0.036195866763591766, + 0.03975163772702217, + 0.05435456708073616, + 0.036079373210668564, + 0.047068897634744644, + 0.0259222611784935, + 0.010597757063806057, + 0.018400775268673897, + 0.008902872912585735, + 0.012223606929183006, + 0.01686062477529049, + 0.01155498530715704, + 0.025968287140130997, + 0.011721775867044926, + 0.010148992761969566, + 0.009020546451210976, + 0.005335430148988962, + 0.006237627472728491, + 0.008870492689311504, + 0.006656702142208815, + 0.013134961016476154, + 0.017282867804169655, + 0.017166748642921448, + 0.017232097685337067, + 0.014189854264259338, + 0.013829745352268219, + 0.014805658720433712, + 0.01153796911239624, + 0.017249267548322678, + 0.005955999251455069, + 0.004593226592987776, + 0.005689305253326893, + 0.003420258639380336, + 0.0035187420435249805, + 0.005280894227325916, + 0.004096886608749628, + 0.005667696241289377, + 0.008436169475317001, + 0.008602999150753021, + 0.007607914041727781, + 0.004031595308333635, + 0.00453310227021575, + 0.008695423603057861, + 0.006847876124083996, + 0.008699988014996052, + 0.01264782715588808, + 0.012241010554134846, + 0.01018210593611002, + 0.007727222517132759, + 0.010626417584717274, + 0.010493948124349117, + 0.013626303523778915, + 0.013486907817423344, + 0.006963811814785004, + 0.006690343376249075, + 0.0062357960268855095, + 0.0042997850105166435, + 0.005419136956334114, + 0.006778395269066095, + 0.00746883312240243, + 0.006777918431907892, + 0.004941828548908234, + 0.004091838840395212, + 0.005148657131940126, + 0.0033078703563660383, + 0.003915018867701292, + 0.004744897596538067, + 0.005050440318882465, + 0.0051454342901706696, + 0.0060848998837172985, + 0.004602842964231968, + 0.005380177870392799, + 0.004468540195375681, + 0.004162194672971964, + 0.00326604419387877, + 0.002920924685895443, + 0.0068909358233213425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.12705424427986145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a level angle and perfectly centered, focuses squarely on a young monk. His shaved head reflects a soft sheen in the warm candlelight. He sits cross-legged on a round, dark brown cushion, clad in a loose, gray cotton and linen robe, its folds flowing naturally along his straight back. His eyes are closed, his long eyelashes casting a subtle shadow on his cheeks. His features are delicate, his lips slightly pursed, his expression serene and focused. His hands are folded before him in the meditation mudra, like a statue, lost in deep meditation. The sanctuary he inhabits is a simple, rustic setting, framed by dark wooden walls and a slightly out-of-focus shoji paper door. Not far from him, a white candle burns quietly on a rustic ceramic candlestick. The flickering orange flame is the only dynamic element and the only light source in the picture. It casts soft and ever-changing light and shadow on the monk's focused face and body outline, separating him from the dark background and creating an atmosphere of extreme tranquility, solemnity and inner strength, as if time has frozen at this moment.", + "seed": 17477, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_5840_seed_17477_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03825773522257805, + "vlm_logprob": -0.03825773522257805, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.03825773522257805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013308502733707428 + ], + "opposite_logprob_tokens": [ + -4.763308525085449 + ], + "logprob_reward": -0.013308502733707428, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013308502733707428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37121009826660156, + 0.0805000588297844, + 0.0639483779668808, + 0.08886110037565231, + 0.06073826551437378, + 0.06199454516172409, + 0.04192954674363136, + 0.05287505313754082, + 0.09183823317289352, + 0.08610472828149796 + ], + "credit_map": [ + 0.05040580779314041, + 0.0360351987183094, + 0.04745083302259445, + 0.0553436316549778, + 0.05897236242890358, + 0.032333988696336746, + 0.03966321796178818, + 0.051005057990550995, + 0.010735228657722473, + 0.005599030293524265, + 0.01055537536740303, + 0.011768334545195103, + 0.011722603812813759, + 0.009956235066056252, + 0.008995293639600277, + 0.011167949065566063, + 0.00911638978868723, + 0.005773322656750679, + 0.008665758185088634, + 0.006852458696812391, + 0.00601270142942667, + 0.01038600504398346, + 0.006698074284940958, + 0.010443655773997307, + 0.011735350824892521, + 0.006483811419457197, + 0.009225653484463692, + 0.014195338822901249, + 0.01298702135682106, + 0.009770373813807964, + 0.00941522978246212, + 0.015048312954604626, + 0.007829980924725533, + 0.005292796064168215, + 0.009559767320752144, + 0.007884765975177288, + 0.008889591321349144, + 0.005561232101172209, + 0.0058434647507965565, + 0.009876660071313381, + 0.006413325201719999, + 0.004973193630576134, + 0.00975108053535223, + 0.007329491898417473, + 0.008304468356072903, + 0.007491250056773424, + 0.007580982521176338, + 0.010150742717087269, + 0.005429702345281839, + 0.0032914909534156322, + 0.007514778524637222, + 0.004785529803484678, + 0.004249878693372011, + 0.0032821781933307648, + 0.00666528707370162, + 0.006710695568472147, + 0.009920076467096806, + 0.005433500744402409, + 0.00628464762121439, + 0.005512807052582502, + 0.005864507984369993, + 0.0071167126297950745, + 0.004880927503108978, + 0.007861871272325516, + 0.015899546444416046, + 0.008036048151552677, + 0.009390674531459808, + 0.008712402544915676, + 0.013035100884735584, + 0.012876508757472038, + 0.008607589639723301, + 0.015280351042747498, + 0.013442432507872581, + 0.008238950744271278, + 0.012274878099560738, + 0.008484277874231339, + 0.012280821800231934, + 0.00962921418249607, + 0.006797152571380138, + 0.01495699305087328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.013308502733707428 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812900349497795 + ], + "opposite_logprob_tokens": [ + -3.529812812805176 + ], + "logprob_reward": -0.029812900349497795, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812900349497795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40717002749443054, + 0.14697466790676117, + 0.07370688021183014, + 0.08806240558624268, + 0.04629398509860039, + 0.03218219056725502, + 0.06685082614421844, + 0.039999429136514664, + 0.03821365162730217, + 0.06054593622684479 + ], + "credit_map": [ + 0.041543588042259216, + 0.030490748584270477, + 0.038575511425733566, + 0.07579874247312546, + 0.05007653683423996, + 0.038374461233615875, + 0.07691999524831772, + 0.05539043992757797, + 0.015282860025763512, + 0.007931455038487911, + 0.01405257172882557, + 0.029250888153910637, + 0.01837042346596718, + 0.0100473053753376, + 0.03137587010860443, + 0.020663300529122353, + 0.00872404407709837, + 0.0046614850871264935, + 0.009038493037223816, + 0.013593471609055996, + 0.008057385683059692, + 0.006834637373685837, + 0.012657706625759602, + 0.01013965718448162, + 0.00996664259582758, + 0.007010399829596281, + 0.0132517721503973, + 0.012119698338210583, + 0.01392392534762621, + 0.007390862796455622, + 0.010003365576267242, + 0.014395750127732754, + 0.006417650263756514, + 0.002906487323343754, + 0.004230850841850042, + 0.00688589084893465, + 0.005920180585235357, + 0.007071900647133589, + 0.006041537504643202, + 0.006819489877671003, + 0.004408266860991716, + 0.0028141632210463285, + 0.004466712474822998, + 0.0040273284539580345, + 0.0042722951620817184, + 0.0036840392276644707, + 0.0035041128285229206, + 0.005005276761949062, + 0.009127129800617695, + 0.005267452914267778, + 0.008468416519463062, + 0.006449975538998842, + 0.011097574606537819, + 0.009188422933220863, + 0.005691452417522669, + 0.011560400947928429, + 0.004223318304866552, + 0.0030735894106328487, + 0.0050425962544977665, + 0.005572315771132708, + 0.00650037219747901, + 0.00444819638505578, + 0.00460337707772851, + 0.006535667926073074, + 0.0066414144821465015, + 0.0027373479679226875, + 0.0059470827691257, + 0.0034559573978185654, + 0.003605090780183673, + 0.00652599660679698, + 0.002991709392517805, + 0.006309055723249912, + 0.007424251176416874, + 0.00408818619325757, + 0.009339380078017712, + 0.007705542724579573, + 0.010947105474770069, + 0.00487399660050869, + 0.0062367962673306465, + 0.009930682368576527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.029812900349497795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430646575987339 + ], + "opposite_logprob_tokens": [ + -3.168064594268799 + ], + "logprob_reward": -0.0430646575987339, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430646575987339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24927420914173126, + 0.12440723925828934, + 0.07150685042142868, + 0.0624195821583271, + 0.05566636472940445, + 0.0853671282529831, + 0.1266012191772461, + 0.08584969490766525, + 0.10884956270456314, + 0.03005814366042614 + ], + "credit_map": [ + 0.03774381801486015, + 0.02040698379278183, + 0.03965108096599579, + 0.03030817210674286, + 0.026433059945702553, + 0.023140957579016685, + 0.032501399517059326, + 0.039088740944862366, + 0.016514407470822334, + 0.009516068734228611, + 0.01896793767809868, + 0.014545460231602192, + 0.014358904212713242, + 0.013106375001370907, + 0.018973492085933685, + 0.01842459663748741, + 0.010810930281877518, + 0.0055207195691764355, + 0.008535942994058132, + 0.007801061496138573, + 0.00904160737991333, + 0.009617575444281101, + 0.009033379144966602, + 0.011145637370646, + 0.009767825715243816, + 0.005360990762710571, + 0.007563886698335409, + 0.00732820201665163, + 0.006839307956397533, + 0.007974947802722454, + 0.007940523326396942, + 0.009643899276852608, + 0.00898529589176178, + 0.0056382897309958935, + 0.006465592421591282, + 0.005346968304365873, + 0.007020002696663141, + 0.009403000585734844, + 0.004817853681743145, + 0.007989360950887203, + 0.012929534539580345, + 0.007400285918265581, + 0.01045441348105669, + 0.008890395052731037, + 0.013343937695026398, + 0.010629834607243538, + 0.008606999181210995, + 0.013111729174852371, + 0.019087616354227066, + 0.010908473283052444, + 0.015690777450799942, + 0.012667401693761349, + 0.02110842429101467, + 0.01531427912414074, + 0.011098595336079597, + 0.02072564885020256, + 0.015291725285351276, + 0.00917164608836174, + 0.009678049944341183, + 0.008540824986994267, + 0.010815382935106754, + 0.013543880544602871, + 0.006827329285442829, + 0.011980855837464333, + 0.01896701566874981, + 0.009578844532370567, + 0.013603853061795235, + 0.010313102044165134, + 0.011217879131436348, + 0.018809055909514427, + 0.009196829050779343, + 0.017162984237074852, + 0.004183121956884861, + 0.0023423880338668823, + 0.0055986265651881695, + 0.0027597814332693815, + 0.0049926904030144215, + 0.0023661977611482143, + 0.0022210427559912205, + 0.0055942959152162075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0430646575987339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060860827565193 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.016060860827565193, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060860827565193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3346729576587677, + 0.13450078666210175, + 0.07804208248853683, + 0.10804656893014908, + 0.05368737131357193, + 0.0682678073644638, + 0.04662155732512474, + 0.04738892242312431, + 0.08531684428453445, + 0.043455127626657486 + ], + "credit_map": [ + 0.04623877629637718, + 0.029032696038484573, + 0.03937586396932602, + 0.03263367712497711, + 0.05297263339161873, + 0.0399087518453598, + 0.03999687358736992, + 0.05451370030641556, + 0.023336263373494148, + 0.010763639584183693, + 0.016635801643133163, + 0.011948350816965103, + 0.01788424327969551, + 0.020384062081575394, + 0.014413084834814072, + 0.01913534477353096, + 0.012933007441461086, + 0.0060922689735889435, + 0.009295249357819557, + 0.0065399957820773125, + 0.009393621236085892, + 0.013736516237258911, + 0.008017227984964848, + 0.012034191749989986, + 0.012726725079119205, + 0.007402531802654266, + 0.018525881692767143, + 0.009273143485188484, + 0.01389772817492485, + 0.015528842806816101, + 0.011765961535274982, + 0.018925746902823448, + 0.00958846416324377, + 0.0045167128555476665, + 0.007730721030384302, + 0.005480033345520496, + 0.007135162595659494, + 0.005552942864596844, + 0.0060407062992453575, + 0.0076426295563578606, + 0.010812281630933285, + 0.005044452380388975, + 0.0069410973228514194, + 0.00775599479675293, + 0.010149109177291393, + 0.008147910237312317, + 0.008478541858494282, + 0.010938417166471481, + 0.005392484832555056, + 0.0035200542770326138, + 0.004234391264617443, + 0.0069153811782598495, + 0.00642551202327013, + 0.005585885606706142, + 0.008147439919412136, + 0.006400408688932657, + 0.00619638804346323, + 0.0037407586351037025, + 0.004206992220133543, + 0.005694428458809853, + 0.006296159699559212, + 0.008069120347499847, + 0.0060492646880447865, + 0.007135808002203703, + 0.014080360531806946, + 0.007630546577274799, + 0.015172414481639862, + 0.00792065542191267, + 0.00802595354616642, + 0.01297029945999384, + 0.005573596339672804, + 0.013943021185696125, + 0.0068752639926970005, + 0.00365628395229578, + 0.006009258795529604, + 0.004584021866321564, + 0.005160849075764418, + 0.0073143732734024525, + 0.0038968774024397135, + 0.005958197172731161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.016060860827565193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904175460338593 + ], + "opposite_logprob_tokens": [ + -2.4640417098999023 + ], + "logprob_reward": -0.08904175460338593, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904175460338593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29710400104522705, + 0.09650097787380219, + 0.09225452691316605, + 0.045205257833004, + 0.11503074318170547, + 0.07722753286361694, + 0.04805893450975418, + 0.0921936184167862, + 0.10792258381843567, + 0.028501877561211586 + ], + "credit_map": [ + 0.04477404057979584, + 0.02384716458618641, + 0.032205864787101746, + 0.03853524848818779, + 0.03745127096772194, + 0.04852157086133957, + 0.029518067836761475, + 0.04225073754787445, + 0.01251049991697073, + 0.007891453802585602, + 0.013854160904884338, + 0.010587800294160843, + 0.010512229055166245, + 0.017011644318699837, + 0.010037394240498543, + 0.014095784164965153, + 0.011524713598191738, + 0.007690862752497196, + 0.014955541118979454, + 0.010996917262673378, + 0.011976098641753197, + 0.009698467329144478, + 0.01122726034373045, + 0.014184665866196156, + 0.006290877237915993, + 0.00316686206497252, + 0.007389914710074663, + 0.004940614569932222, + 0.005706542171537876, + 0.004827931057661772, + 0.005449521355330944, + 0.007432993967086077, + 0.012473720125854015, + 0.009635847993195057, + 0.018920499831438065, + 0.015232175588607788, + 0.01645852066576481, + 0.009321656078100204, + 0.014974910765886307, + 0.018013404682278633, + 0.008781366981565952, + 0.005265446845442057, + 0.01614456996321678, + 0.00820118561387062, + 0.009836520068347454, + 0.00692148320376873, + 0.008770318701863289, + 0.01330664474517107, + 0.0051744976080954075, + 0.003606278682127595, + 0.009727832861244678, + 0.004717645701020956, + 0.005416944622993469, + 0.004037013743072748, + 0.006018541287630796, + 0.009360178373754025, + 0.012544075958430767, + 0.007179239299148321, + 0.011953512206673622, + 0.011492358520627022, + 0.013669192790985107, + 0.010638521052896976, + 0.011086858808994293, + 0.013629850000143051, + 0.018089469522237778, + 0.010247545316815376, + 0.016791056841611862, + 0.009979835711419582, + 0.012592135928571224, + 0.011203723959624767, + 0.011440873146057129, + 0.01757793501019478, + 0.0042921495623886585, + 0.0024356164503842592, + 0.002836971078068018, + 0.00358630926348269, + 0.003135116770863533, + 0.005635544192045927, + 0.0024571747053414583, + 0.004122994840145111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.08904175460338593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a level angle and perfectly centered, focuses squarely on a young monk. His shaved head reflects a soft sheen in the warm candlelight. He sits cross-legged on a round, dark brown cushion, clad in a loose, gray cotton and linen robe, its folds flowing naturally along his straight back. His eyes are closed, his long eyelashes casting a subtle shadow on his cheeks. His features are delicate, his lips slightly pursed, his expression serene and focused. His hands are folded before him in the meditation mudra, like a statue, lost in deep meditation. The sanctuary he inhabits is a simple, rustic setting, framed by dark wooden walls and a slightly out-of-focus shoji paper door. Not far from him, a white candle burns quietly on a rustic ceramic candlestick. The flickering orange flame is the only dynamic element and the only light source in the picture. It casts soft and ever-changing light and shadow on the monk's focused face and body outline, separating him from the dark background and creating an atmosphere of extreme tranquility, solemnity and inner strength, as if time has frozen at this moment.", + "seed": 17478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_5840_seed_17478_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04706752505153418, + "vlm_logprob": -0.04706752505153418, + "vlm_yes_no_margin": 3.450000023841858, + "vlm_token_credit": -0.04706752505153418, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027942240238189697 + ], + "opposite_logprob_tokens": [ + -3.777942180633545 + ], + "logprob_reward": -0.027942240238189697, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027942240238189697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.399229496717453, + 0.09396969527006149, + 0.08445703983306885, + 0.09844601154327393, + 0.0497976690530777, + 0.046877000480890274, + 0.039603594690561295, + 0.03673738241195679, + 0.09304980933666229, + 0.05783235654234886 + ], + "credit_map": [ + 0.05304165929555893, + 0.03762393072247505, + 0.08187244087457657, + 0.06386323273181915, + 0.03423549234867096, + 0.05044355243444443, + 0.03844094276428223, + 0.039708226919174194, + 0.009198818355798721, + 0.012719987891614437, + 0.017833037301898003, + 0.01645706407725811, + 0.0076780472882092, + 0.012019587680697441, + 0.008978401310741901, + 0.009084749966859818, + 0.011249842122197151, + 0.008396311663091183, + 0.009751257486641407, + 0.015067847445607185, + 0.008081476204097271, + 0.014138062484562397, + 0.008688529022037983, + 0.009083718992769718, + 0.011988773941993713, + 0.01245386153459549, + 0.01922675408422947, + 0.01696665957570076, + 0.008344973437488079, + 0.011858618818223476, + 0.008527657017111778, + 0.009078710339963436, + 0.004557808395475149, + 0.005298506934195757, + 0.0072259013541042805, + 0.008278241381049156, + 0.0038400255143642426, + 0.009601356461644173, + 0.0049140118062496185, + 0.006081817205995321, + 0.004693199414759874, + 0.0037137854378670454, + 0.007515543606132269, + 0.007970225997269154, + 0.004873403813689947, + 0.008206600323319435, + 0.003999353852123022, + 0.005904886871576309, + 0.004539818968623877, + 0.005825878120958805, + 0.005544520448893309, + 0.00655833724886179, + 0.004124930594116449, + 0.005971238017082214, + 0.0034023800399154425, + 0.0036364945117384195, + 0.0038509462028741837, + 0.0035334196873009205, + 0.0045159016735851765, + 0.005576451774686575, + 0.0033203132916241884, + 0.005016681272536516, + 0.0037448396906256676, + 0.007178831845521927, + 0.01804642379283905, + 0.015501519665122032, + 0.015328149311244488, + 0.015574995428323746, + 0.0073823644779622555, + 0.007112847175449133, + 0.006244769785553217, + 0.007858742028474808, + 0.009281019680202007, + 0.0064450339414179325, + 0.009768488816916943, + 0.009957339614629745, + 0.0060568745248019695, + 0.005777179263532162, + 0.005249318666756153, + 0.005297102499753237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.027942240238189697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590949803590775 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020590949803590775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590949803590775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31937849521636963, + 0.15418380498886108, + 0.10320326685905457, + 0.0523773692548275, + 0.04510118439793587, + 0.06279433518648148, + 0.05312914773821831, + 0.05516133829951286, + 0.09789133816957474, + 0.05677971988916397 + ], + "credit_map": [ + 0.04818664491176605, + 0.04221852868795395, + 0.05079756677150726, + 0.0547093041241169, + 0.02814513072371483, + 0.03800434246659279, + 0.028258059173822403, + 0.029058940708637238, + 0.03037281148135662, + 0.021311406046152115, + 0.01985919661819935, + 0.02370702475309372, + 0.01409414503723383, + 0.015262611210346222, + 0.01554682943969965, + 0.014029774814844131, + 0.021136905997991562, + 0.014804325997829437, + 0.014377979561686516, + 0.016433216631412506, + 0.007447879761457443, + 0.010253931395709515, + 0.00730566680431366, + 0.011443356052041054, + 0.006544996052980423, + 0.006612719036638737, + 0.008462348952889442, + 0.0102715278044343, + 0.005058938171714544, + 0.006642341148108244, + 0.003887476399540901, + 0.00489701796323061, + 0.007311950903385878, + 0.005110027734190226, + 0.005860080476850271, + 0.00712004816159606, + 0.004948124755173922, + 0.006121425423771143, + 0.003818661440163851, + 0.004810862243175507, + 0.014623957686126232, + 0.00858717318624258, + 0.011666061356663704, + 0.010065166279673576, + 0.004814780782908201, + 0.005505134351551533, + 0.0033477533143013716, + 0.004184302408248186, + 0.008745688013732433, + 0.008166908286511898, + 0.007098445203155279, + 0.011373362503945827, + 0.00421611312776804, + 0.005911870859563351, + 0.004028812050819397, + 0.0035879460629075766, + 0.012140767648816109, + 0.006415525916963816, + 0.009642861783504486, + 0.008816640824079514, + 0.004318543244153261, + 0.0050524440594017506, + 0.0045654745772480965, + 0.004209076054394245, + 0.01987408474087715, + 0.017161371186375618, + 0.016164470463991165, + 0.01887466013431549, + 0.006968518253415823, + 0.007496332749724388, + 0.005883705802261829, + 0.005468183662742376, + 0.00954744964838028, + 0.0046564312651753426, + 0.008677837438881397, + 0.011250907555222511, + 0.005257004406303167, + 0.008306579664349556, + 0.0049669998697936535, + 0.004116505850106478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020590949803590775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980584278702736 + ], + "opposite_logprob_tokens": [ + -3.529805898666382 + ], + "logprob_reward": -0.02980584278702736, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980584278702736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3039665222167969, + 0.15643441677093506, + 0.09900186955928802, + 0.049191415309906006, + 0.08083949983119965, + 0.06336011737585068, + 0.07860743254423141, + 0.054418571293354034, + 0.048554677516222, + 0.06562551856040955 + ], + "credit_map": [ + 0.03387834131717682, + 0.03221217170357704, + 0.047214947640895844, + 0.051192935556173325, + 0.029796848073601723, + 0.06132311746478081, + 0.022206943482160568, + 0.026141226291656494, + 0.03388430178165436, + 0.022781308740377426, + 0.020617114380002022, + 0.025588853284716606, + 0.014298928901553154, + 0.01561226136982441, + 0.013335722498595715, + 0.010315931402146816, + 0.0204167403280735, + 0.014685112982988358, + 0.014986325986683369, + 0.015746768563985825, + 0.0076927486807107925, + 0.011440396308898926, + 0.00731084868311882, + 0.006722929421812296, + 0.00919129978865385, + 0.00493644829839468, + 0.009381459094583988, + 0.009535718709230423, + 0.003504863940179348, + 0.0046445890329778194, + 0.004126515239477158, + 0.003870520042255521, + 0.016694264486432076, + 0.012484915554523468, + 0.014405584894120693, + 0.013290099799633026, + 0.0061470260843634605, + 0.008512232452630997, + 0.00469718873500824, + 0.004608187358826399, + 0.01168854534626007, + 0.008351045660674572, + 0.010520652867853642, + 0.00991410668939352, + 0.004493645392358303, + 0.006763814482837915, + 0.005221887491643429, + 0.006406419910490513, + 0.013198348693549633, + 0.01196971070021391, + 0.013193747028708458, + 0.016840193420648575, + 0.005615354515612125, + 0.006839614361524582, + 0.0053718858398497105, + 0.005578584969043732, + 0.011179494671523571, + 0.006385641638189554, + 0.009034547954797745, + 0.008253342472016811, + 0.0037155221216380596, + 0.005861440673470497, + 0.005075366701930761, + 0.0049132127314805984, + 0.007410666439682245, + 0.007295717019587755, + 0.007701357826590538, + 0.00866580381989479, + 0.004292281344532967, + 0.00437527010217309, + 0.0034376911353319883, + 0.005375887732952833, + 0.012697975151240826, + 0.006814229302108288, + 0.011727957986295223, + 0.012354475446045399, + 0.005450280383229256, + 0.006867721676826477, + 0.005337001755833626, + 0.004375876393169165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02980584278702736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191421680152416 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191421680152416, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191421680152416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2941899001598358, + 0.13831062614917755, + 0.15836235880851746, + 0.13901185989379883, + 0.05781610682606697, + 0.04614494740962982, + 0.040091633796691895, + 0.032808657735586166, + 0.04382536560297012, + 0.049438588321208954 + ], + "credit_map": [ + 0.03632507100701332, + 0.03293872997164726, + 0.07934388518333435, + 0.053107164800167084, + 0.025371020659804344, + 0.02562263049185276, + 0.020534604787826538, + 0.020946750417351723, + 0.03314206749200821, + 0.023395538330078125, + 0.01728248968720436, + 0.024532822892069817, + 0.008919527754187584, + 0.011916790157556534, + 0.010262977331876755, + 0.008858409710228443, + 0.03944312781095505, + 0.027638455852866173, + 0.022498667240142822, + 0.02999856136739254, + 0.010195357725024223, + 0.010572313331067562, + 0.00971658993512392, + 0.008299283683300018, + 0.03309064731001854, + 0.019173866137862206, + 0.029209664091467857, + 0.024124955758452415, + 0.0091929966583848, + 0.009510776959359646, + 0.007796677760779858, + 0.006912254728376865, + 0.007574566174298525, + 0.005317707546055317, + 0.01224431674927473, + 0.009921842254698277, + 0.003932558465749025, + 0.005666854791343212, + 0.005490977317094803, + 0.00766728026792407, + 0.011561606079339981, + 0.006929771509021521, + 0.008211766369640827, + 0.008014938794076443, + 0.0031765648163855076, + 0.0029948772862553596, + 0.002946676919236779, + 0.0023087426088750362, + 0.006357261445373297, + 0.003977679181843996, + 0.008950027637183666, + 0.007232200354337692, + 0.00403995206579566, + 0.0051846872083842754, + 0.0021589475218206644, + 0.0021908741910010576, + 0.0034500251058489084, + 0.004060621839016676, + 0.006017846055328846, + 0.004768824204802513, + 0.002855787519365549, + 0.004830586723983288, + 0.003292055567726493, + 0.0035329065285623074, + 0.0061256433837115765, + 0.007562790531665087, + 0.0072828633710742, + 0.008237304165959358, + 0.0034506770316511393, + 0.004481977317482233, + 0.0029412417206913233, + 0.0037428634241223335, + 0.009430352598428726, + 0.007208980619907379, + 0.008835593238472939, + 0.008473188616335392, + 0.00418146513402462, + 0.0052070836536586285, + 0.0031418069265782833, + 0.002960115671157837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.014191421680152416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280717074871063 + ], + "opposite_logprob_tokens": [ + -2.0178072452545166 + ], + "logprob_reward": -0.14280717074871063, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280717074871063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40723875164985657, + 0.11416055262088776, + 0.04527539014816284, + 0.062144726514816284, + 0.03961077705025673, + 0.07609845697879791, + 0.06994018703699112, + 0.08502195030450821, + 0.053815074265003204, + 0.046694036573171616 + ], + "credit_map": [ + 0.07689692825078964, + 0.053377121686935425, + 0.09254490584135056, + 0.07008819282054901, + 0.0315701849758625, + 0.0356462262570858, + 0.020628320053219795, + 0.02648692950606346, + 0.022094326093792915, + 0.019104285165667534, + 0.020156504586338997, + 0.01886754482984543, + 0.00943977665156126, + 0.009036167524755001, + 0.0076254005543887615, + 0.007836557924747467, + 0.005312200170010328, + 0.00796855241060257, + 0.005310931243002415, + 0.008029598742723465, + 0.004873795434832573, + 0.005011289846152067, + 0.004381075501441956, + 0.004387952852994204, + 0.007890460081398487, + 0.0094189727678895, + 0.011913535185158253, + 0.011762428097426891, + 0.005333489738404751, + 0.0060650939121842384, + 0.0050389403477311134, + 0.004721817094832659, + 0.005783787928521633, + 0.00603690929710865, + 0.007882354781031609, + 0.006690099369734526, + 0.0037868665531277657, + 0.0031328652985394, + 0.0028660534881055355, + 0.0034318489488214254, + 0.01402024831622839, + 0.010908074676990509, + 0.015942875295877457, + 0.012822234071791172, + 0.005283622536808252, + 0.007823022082448006, + 0.004149483051151037, + 0.005148908589035273, + 0.00723018916323781, + 0.007719274144619703, + 0.01110172551125288, + 0.012105448171496391, + 0.006637703161686659, + 0.010106266476213932, + 0.007051707245409489, + 0.007987882941961288, + 0.012239595875144005, + 0.01006005983799696, + 0.017707735300064087, + 0.01351789478212595, + 0.007182961795479059, + 0.0106012262403965, + 0.0070114596746861935, + 0.006701029371470213, + 0.010617994703352451, + 0.00902940146625042, + 0.009678679518401623, + 0.009334912523627281, + 0.003959649708122015, + 0.004745065234601498, + 0.0032876376062631607, + 0.003161737462505698, + 0.008180630393326283, + 0.006332728546112776, + 0.009047840721905231, + 0.007189363706856966, + 0.003417286789044738, + 0.0052079916931688786, + 0.003307414473965764, + 0.004010786302387714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.14280717074871063 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a level angle and perfectly centered, focuses squarely on a young monk. His shaved head reflects a soft sheen in the warm candlelight. He sits cross-legged on a round, dark brown cushion, clad in a loose, gray cotton and linen robe, its folds flowing naturally along his straight back. His eyes are closed, his long eyelashes casting a subtle shadow on his cheeks. His features are delicate, his lips slightly pursed, his expression serene and focused. His hands are folded before him in the meditation mudra, like a statue, lost in deep meditation. The sanctuary he inhabits is a simple, rustic setting, framed by dark wooden walls and a slightly out-of-focus shoji paper door. Not far from him, a white candle burns quietly on a rustic ceramic candlestick. The flickering orange flame is the only dynamic element and the only light source in the picture. It casts soft and ever-changing light and shadow on the monk's focused face and body outline, separating him from the dark background and creating an atmosphere of extreme tranquility, solemnity and inner strength, as if time has frozen at this moment.", + "seed": 17479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_6/index_5840_seed_17479_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026535192504525185, + "vlm_logprob": -0.026535192504525185, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.026535192504525185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01139525044709444 + ], + "opposite_logprob_tokens": [ + -5.011395454406738 + ], + "logprob_reward": -0.01139525044709444, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01139525044709444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33865582942962646, + 0.11286357790231705, + 0.08146088570356369, + 0.05239897966384888, + 0.06772243976593018, + 0.06433653831481934, + 0.08388327062129974, + 0.05721860006451607, + 0.07594507932662964, + 0.06551478803157806 + ], + "credit_map": [ + 0.03787962347269058, + 0.037237800657749176, + 0.03857129067182541, + 0.034776121377944946, + 0.05342031270265579, + 0.05307980626821518, + 0.04110650345683098, + 0.042584363371133804, + 0.013433161191642284, + 0.009556016884744167, + 0.009733851067721844, + 0.014047274366021156, + 0.01526904571801424, + 0.016552483662962914, + 0.017125003039836884, + 0.01714674010872841, + 0.012058482505381107, + 0.009278682991862297, + 0.011081008240580559, + 0.008041446097195148, + 0.008399007841944695, + 0.011204608716070652, + 0.009832275100052357, + 0.011565368622541428, + 0.007323985453695059, + 0.004238776862621307, + 0.00784015841782093, + 0.006116815842688084, + 0.006996666081249714, + 0.006961322855204344, + 0.0057511720806360245, + 0.007170084863901138, + 0.01057472638785839, + 0.007360114250332117, + 0.009583248756825924, + 0.007368030492216349, + 0.007046042010188103, + 0.00968377385288477, + 0.009186845272779465, + 0.006919654551893473, + 0.011418142355978489, + 0.005729238502681255, + 0.005872730631381273, + 0.004271896556019783, + 0.004402657505124807, + 0.01671001687645912, + 0.006040861364454031, + 0.009890989400446415, + 0.01764928735792637, + 0.00509117916226387, + 0.01086210273206234, + 0.005593190900981426, + 0.009295476600527763, + 0.02039494179189205, + 0.005746385548263788, + 0.009250706061720848, + 0.008973248302936554, + 0.005548327229917049, + 0.012793940491974354, + 0.004340075422078371, + 0.005633961874991655, + 0.0073713213205337524, + 0.005261264741420746, + 0.007296463008970022, + 0.011894055642187595, + 0.005481612868607044, + 0.015929771587252617, + 0.008800022304058075, + 0.008348174393177032, + 0.00954227801412344, + 0.007883562706410885, + 0.008065602742135525, + 0.008602667599916458, + 0.008473007939755917, + 0.010824134573340416, + 0.006889048498123884, + 0.007442675065249205, + 0.00786123052239418, + 0.008374637924134731, + 0.007047387771308422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01139525044709444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232975073158741 + ], + "opposite_logprob_tokens": [ + -3.7732975482940674 + ], + "logprob_reward": -0.0232975073158741, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232975073158741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35479897260665894, + 0.13229160010814667, + 0.10818737745285034, + 0.05347216874361038, + 0.07872244715690613, + 0.06197208911180496, + 0.04952263459563255, + 0.04508446902036667, + 0.0646011158823967, + 0.05134708061814308 + ], + "credit_map": [ + 0.04138809069991112, + 0.032268378883600235, + 0.04875379428267479, + 0.04598349705338478, + 0.05027715489268303, + 0.04288739711046219, + 0.05349303036928177, + 0.03974763676524162, + 0.021325502544641495, + 0.009305155836045742, + 0.012855680659413338, + 0.0136096877977252, + 0.014365613460540771, + 0.030001364648342133, + 0.016301987692713737, + 0.014526600949466228, + 0.02054544910788536, + 0.006284789647907019, + 0.007863194681704044, + 0.008773589506745338, + 0.011728279292583466, + 0.030574850738048553, + 0.010281631723046303, + 0.012135593220591545, + 0.007686890196055174, + 0.008530154824256897, + 0.007549962494522333, + 0.0060115051455795765, + 0.005726359318941832, + 0.006704216357320547, + 0.004773883149027824, + 0.006489194463938475, + 0.015631699934601784, + 0.004265053663402796, + 0.009653361514210701, + 0.004942282568663359, + 0.006632463540881872, + 0.02149549312889576, + 0.007340333424508572, + 0.008761756122112274, + 0.008556520566344261, + 0.0048419879749417305, + 0.005806291475892067, + 0.0046865795738995075, + 0.006058924365788698, + 0.014209031127393246, + 0.00853659026324749, + 0.009276163764297962, + 0.00841229222714901, + 0.003920637536793947, + 0.005630311090499163, + 0.0059457262977957726, + 0.0042315698228776455, + 0.007199388463050127, + 0.00625558290630579, + 0.007927128113806248, + 0.006176653318107128, + 0.004022551700472832, + 0.004902535118162632, + 0.004886704497039318, + 0.006649605929851532, + 0.007908351719379425, + 0.005786874331533909, + 0.004751190077513456, + 0.01061580516397953, + 0.00865919515490532, + 0.009446172043681145, + 0.006670041475445032, + 0.004760598298162222, + 0.008757036179304123, + 0.005334734916687012, + 0.010357532650232315, + 0.006267547607421875, + 0.004649290814995766, + 0.005506427958607674, + 0.004895286168903112, + 0.007981915958225727, + 0.008421164005994797, + 0.006822738330811262, + 0.006802706979215145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0232975073158741 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818707026541233 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818707026541233, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818707026541233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36735063791275024, + 0.1463707983493805, + 0.10429387539625168, + 0.05370355397462845, + 0.05799786373972893, + 0.05637527257204056, + 0.06015557050704956, + 0.04076652228832245, + 0.05496211722493172, + 0.05802379548549652 + ], + "credit_map": [ + 0.04031022638082504, + 0.03121713176369667, + 0.033378954976797104, + 0.05256865546107292, + 0.06023317947983742, + 0.04610629379749298, + 0.05229520425200462, + 0.05124100670218468, + 0.02322169952094555, + 0.011227001436054707, + 0.011216928251087666, + 0.01468526478856802, + 0.019843170419335365, + 0.031218884512782097, + 0.014350886456668377, + 0.02060696668922901, + 0.015575933270156384, + 0.010164469480514526, + 0.006944150198251009, + 0.013559000566601753, + 0.01693107932806015, + 0.01895241066813469, + 0.011319915764033794, + 0.010846912860870361, + 0.008752032183110714, + 0.006201160605996847, + 0.004412159789353609, + 0.005342530552297831, + 0.00820434745401144, + 0.009045870043337345, + 0.006314131431281567, + 0.00543132284656167, + 0.008013828657567501, + 0.004642527084797621, + 0.006042906083166599, + 0.007323137018829584, + 0.008537190966308117, + 0.008472508750855923, + 0.007636586204171181, + 0.0073291766457259655, + 0.009106694720685482, + 0.0043428135104477406, + 0.003928461577743292, + 0.0046761068515479565, + 0.0062014819122850895, + 0.013090617023408413, + 0.006742095109075308, + 0.008287002332508564, + 0.013023539446294308, + 0.005248498637229204, + 0.007160899229347706, + 0.004008934833109379, + 0.004623372107744217, + 0.015069028362631798, + 0.0033602111507207155, + 0.007661082316190004, + 0.0048876493237912655, + 0.004462579730898142, + 0.0038039940409362316, + 0.004623426124453545, + 0.005409521982073784, + 0.006525774952024221, + 0.0059394920244812965, + 0.005114084575325251, + 0.008739709854125977, + 0.006745744496583939, + 0.010246196761727333, + 0.0053506214171648026, + 0.004665329121053219, + 0.006682702340185642, + 0.004773510154336691, + 0.007758300751447678, + 0.008638972416520119, + 0.0055192140862345695, + 0.00597943551838398, + 0.005895503796637058, + 0.006149210035800934, + 0.010976627469062805, + 0.00856527965515852, + 0.006299552507698536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01818707026541233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777984581887722 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.009777984581887722, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777984581887722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2677152752876282, + 0.16910408437252045, + 0.10453607887029648, + 0.13292467594146729, + 0.06854484230279922, + 0.04684580862522125, + 0.03958661109209061, + 0.03944309428334236, + 0.09801338613033295, + 0.033286187797784805 + ], + "credit_map": [ + 0.03982189670205116, + 0.018921498209238052, + 0.05494871735572815, + 0.033538199961185455, + 0.030025994405150414, + 0.03201950341463089, + 0.03321271017193794, + 0.025226712226867676, + 0.02935010939836502, + 0.023150647059082985, + 0.028363879770040512, + 0.0129084512591362, + 0.013065573759377003, + 0.027634892612695694, + 0.011715836822986603, + 0.022914685308933258, + 0.017694395035505295, + 0.01569291576743126, + 0.015257548540830612, + 0.007328305393457413, + 0.008775227703154087, + 0.016764607280492783, + 0.0076838331297039986, + 0.015339239500463009, + 0.022544482722878456, + 0.018116122111678123, + 0.022443510591983795, + 0.014160899445414543, + 0.010402301326394081, + 0.018939336761832237, + 0.011042213998734951, + 0.015275790356099606, + 0.0126167768612504, + 0.007971311919391155, + 0.008459898643195629, + 0.004524955525994301, + 0.00501603726297617, + 0.015550157986581326, + 0.005512563977390528, + 0.008893132209777832, + 0.006993675138801336, + 0.0061211357824504375, + 0.00815875269472599, + 0.0035883712116628885, + 0.004205521661788225, + 0.006530972197651863, + 0.004015190061181784, + 0.007232186384499073, + 0.005182713735848665, + 0.006086450070142746, + 0.005974941421300173, + 0.00420638732612133, + 0.003832119982689619, + 0.005104162264615297, + 0.003805700922384858, + 0.005394130479544401, + 0.006683128885924816, + 0.0038809804245829582, + 0.00421190308406949, + 0.003116948064416647, + 0.0035191660281270742, + 0.009591283276677132, + 0.002801312832161784, + 0.005638368893414736, + 0.014947050251066685, + 0.015285028144717216, + 0.017766602337360382, + 0.010124417021870613, + 0.008054829202592373, + 0.011442221701145172, + 0.007638473063707352, + 0.012754757888615131, + 0.005345128942281008, + 0.0037698952946811914, + 0.004937719088047743, + 0.004069083370268345, + 0.00338751170784235, + 0.004425495862960815, + 0.0038813361898064613, + 0.003470012452453375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009777984581887722 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed from a level angle and perfectly centered on a young monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monk wearing a loose gray cotton and linen robe and sitting cross-legged on a round dark brown cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the monk's hands folded in a meditation mudra with his eyes closed and eyelashes casting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a rustic sanctuary with dark wooden walls, a slightly out-of-focus shoji paper door, and a white candle burning on a ceramic candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source a flickering orange candle flame that casts soft, changing light and shadows on the monk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001814991235733 + ], + "opposite_logprob_tokens": [ + -2.6950180530548096 + ], + "logprob_reward": -0.07001814991235733, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001814991235733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3411582112312317, + 0.1021035760641098, + 0.05261228233575821, + 0.08485770225524902, + 0.17079779505729675, + 0.049267858266830444, + 0.04807537794113159, + 0.056008245795965195, + 0.06739452481269836, + 0.027724429965019226 + ], + "credit_map": [ + 0.06358879804611206, + 0.032705940306186676, + 0.04972492903470993, + 0.028957292437553406, + 0.03276108577847481, + 0.059790968894958496, + 0.028653452172875404, + 0.044975731521844864, + 0.018062174320220947, + 0.011567777022719383, + 0.012644532136619091, + 0.008493087254464626, + 0.011305214837193489, + 0.019713539630174637, + 0.008966529741883278, + 0.011350715532898903, + 0.007452985271811485, + 0.0036359629593789577, + 0.009412059560418129, + 0.004694437142461538, + 0.006746280938386917, + 0.008136101998388767, + 0.005074179731309414, + 0.007460273336619139, + 0.011206460185348988, + 0.009186554700136185, + 0.007890298962593079, + 0.008268777281045914, + 0.009027496911585331, + 0.017090264707803726, + 0.008367293514311314, + 0.013820555992424488, + 0.036545827984809875, + 0.010883934795856476, + 0.02145378850400448, + 0.015168524347245693, + 0.01139887236058712, + 0.04144083335995674, + 0.014229290187358856, + 0.019676722586154938, + 0.007086388301104307, + 0.004638545215129852, + 0.012401765212416649, + 0.0045777796767652035, + 0.004559044260531664, + 0.006056740880012512, + 0.003347506746649742, + 0.006600087508559227, + 0.006541706621646881, + 0.007759732659906149, + 0.009055044502019882, + 0.004297195468097925, + 0.004809771664440632, + 0.005996244959533215, + 0.0037218451034277678, + 0.005893837194889784, + 0.011581427417695522, + 0.0031530950218439102, + 0.008842121809720993, + 0.004086717497557402, + 0.0033647327218204737, + 0.014427216723561287, + 0.0037040472961962223, + 0.006848882883787155, + 0.009680735878646374, + 0.010763166472315788, + 0.012808326631784439, + 0.006356930360198021, + 0.0061847371980547905, + 0.008201082237064838, + 0.004783320240676403, + 0.008616224862635136, + 0.005522320047020912, + 0.0014048856683075428, + 0.003291047178208828, + 0.0022135479375720024, + 0.002388706663623452, + 0.006467907689511776, + 0.0027597276493906975, + 0.003676285035908222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07001814991235733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_tensor([3682, 3682])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_tensor([3682, 3682])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9ab99b9e0fdc87d3b17f79631e3d7d97462c8e21 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_tensor([3682, 3682])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle close-up captures a large, speckled seashell, its form defined by the elegant, flowing lines of traditional Oriental calligraphy. Its surface, resembling the intricate pattern of a partridge's feathers, is a tapestry of tan, sienna, and cream, rendered in soft, bleeding gradations of delicate ink. Half-buried in dark, shimmering wet sand, the shell's smooth, curved surface catches the diffused light. The scene unfolds with the rich texture, subtle grain, and deep color saturation characteristic of 35mm film. A thin, translucent sheet of water slowly and rhythmically creeps up the shore, its edge swirling as it gently caresses the shell's base before retreating with a quiet sigh, leaving the sand and shell's underside shimmering. The entire scene is bathed in a soft, overcast light, eliminating harsh shadows and enhancing the tranquil, contemplative atmosphere, creating an enchanting fusion of natural beauty and timeless artistic expression.", + "seed": 32177, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_3682_seed_32177_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02177913784980774, + "vlm_logprob": -0.02177913784980774, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.02177913784980774, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013903550803661346 + ], + "opposite_logprob_tokens": [ + -4.513903617858887 + ], + "logprob_reward": -0.013903550803661346, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013903550803661346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10186392813920975, + 0.047346021980047226, + 0.04814831167459488, + 0.04194329306483269, + 0.17480061948299408, + 0.18115845322608948, + 0.12387287616729736, + 0.09385722875595093, + 0.05488189682364464, + 0.1321273297071457 + ], + "credit_map": [ + 0.016452865675091743, + 0.014007458463311195, + 0.016913658007979393, + 0.009907030500471592, + 0.009478980675339699, + 0.015190361998975277, + 0.008102407678961754, + 0.011811167933046818, + 0.007160682696849108, + 0.00449305959045887, + 0.005714400205761194, + 0.0036055927630513906, + 0.005204301327466965, + 0.00931543204933405, + 0.004647387657314539, + 0.007205168250948191, + 0.007880622521042824, + 0.00561915896832943, + 0.008033928461372852, + 0.004162325989454985, + 0.0032313065603375435, + 0.0069038583897054195, + 0.004709105938673019, + 0.00760800763964653, + 0.0077141947112977505, + 0.004566744901239872, + 0.004600413143634796, + 0.0035873211454600096, + 0.004170582164078951, + 0.010024195536971092, + 0.0031997051555663347, + 0.004080142360180616, + 0.042446963489055634, + 0.015399551950395107, + 0.021187372505664825, + 0.016518738120794296, + 0.009759996086359024, + 0.039726272225379944, + 0.017017235979437828, + 0.012744498439133167, + 0.05916569009423256, + 0.0102993194013834, + 0.011380404233932495, + 0.008043724112212658, + 0.007559951860457659, + 0.06450393050909042, + 0.00785819161683321, + 0.012347251176834106, + 0.03728139027953148, + 0.012606139294803143, + 0.012817658483982086, + 0.006516825407743454, + 0.006436267402023077, + 0.03654424101114273, + 0.005561340134590864, + 0.006109023466706276, + 0.021144794300198555, + 0.011271484196186066, + 0.01176538597792387, + 0.0072981249541044235, + 0.007197391241788864, + 0.020632335916161537, + 0.0065886895172297955, + 0.007959025911986828, + 0.012369650416076183, + 0.00650970870628953, + 0.00820468831807375, + 0.004906489979475737, + 0.004299790132790804, + 0.008634086698293686, + 0.0040081217885017395, + 0.005949369631707668, + 0.03749176859855652, + 0.012422031722962856, + 0.01056104525923729, + 0.006154023576527834, + 0.009635650552809238, + 0.04264426976442337, + 0.00493836821988225, + 0.008280175738036633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.013903550803661346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052612656727433205 + ], + "opposite_logprob_tokens": [ + -5.255261421203613 + ], + "logprob_reward": -0.0052612656727433205, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052612656727433205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08777406066656113, + 0.04415583238005638, + 0.056307654827833176, + 0.10140908509492874, + 0.35408109426498413, + 0.05952664837241173, + 0.12187240272760391, + 0.030629707500338554, + 0.025783883407711983, + 0.11845966428518295 + ], + "credit_map": [ + 0.013830861076712608, + 0.009085682220757008, + 0.02115556225180626, + 0.007568709552288055, + 0.006721647921949625, + 0.011439546942710876, + 0.006990708876401186, + 0.01098133996129036, + 0.008465452119708061, + 0.0048674349673092365, + 0.006814613938331604, + 0.0038795936852693558, + 0.004272771067917347, + 0.0061691696755588055, + 0.004107497166842222, + 0.00557930301874876, + 0.009416847489774227, + 0.006124354433268309, + 0.007278112694621086, + 0.004641471430659294, + 0.004929018206894398, + 0.013627076521515846, + 0.004272114485502243, + 0.006018661428242922, + 0.02205709181725979, + 0.009345541708171368, + 0.013223551213741302, + 0.005981642287224531, + 0.007045451086014509, + 0.024801576510071754, + 0.006259236019104719, + 0.012694990262389183, + 0.10528817027807236, + 0.026586314663290977, + 0.03762408345937729, + 0.018532950431108475, + 0.016932470723986626, + 0.10324310511350632, + 0.02323680929839611, + 0.022637203335762024, + 0.012696334160864353, + 0.004974099341779947, + 0.006127168424427509, + 0.0045237354934215546, + 0.0037795898970216513, + 0.017770567908883095, + 0.003929567523300648, + 0.005725589115172625, + 0.04187764227390289, + 0.006904253736138344, + 0.0076688616536557674, + 0.004203096963465214, + 0.0047604674473404884, + 0.04653908312320709, + 0.004192627500742674, + 0.0057263728231191635, + 0.007677961140871048, + 0.0032890888396650553, + 0.0025795865803956985, + 0.00280921533703804, + 0.0024553893599659204, + 0.006945592351257801, + 0.0024474498350173235, + 0.0024254275485873222, + 0.004113683942705393, + 0.004574836231768131, + 0.0037941080518066883, + 0.0028184212278574705, + 0.0022988838609308004, + 0.004018273204565048, + 0.0020781017374247313, + 0.00208757515065372, + 0.03962209075689316, + 0.00896492786705494, + 0.0061231981962919235, + 0.0031000201124697924, + 0.007296890020370483, + 0.044414222240448, + 0.0030309115536510944, + 0.005907406564801931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.0052612656727433205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004226744174957 + ], + "opposite_logprob_tokens": [ + -2.695042371749878 + ], + "logprob_reward": -0.07004226744174957, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004226744174957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16106298565864563, + 0.0280443225055933, + 0.04418540000915527, + 0.06427913159132004, + 0.2967422902584076, + 0.07999194413423538, + 0.1796463578939438, + 0.0812569260597229, + 0.03689034283161163, + 0.027900278568267822 + ], + "credit_map": [ + 0.028679246082901955, + 0.016469094902276993, + 0.025270676240324974, + 0.010678381659090519, + 0.017300333827733994, + 0.028913671150803566, + 0.010202210396528244, + 0.02354937046766281, + 0.002925284206867218, + 0.0023973349016159773, + 0.0051924483850598335, + 0.0028981512878090143, + 0.00282350555062294, + 0.003723053727298975, + 0.0027904307935386896, + 0.005294113885611296, + 0.005847534630447626, + 0.005699176341295242, + 0.007237742654979229, + 0.004084027372300625, + 0.0037017727736383677, + 0.008088359609246254, + 0.004136507865041494, + 0.005390278995037079, + 0.012441122904419899, + 0.006253765895962715, + 0.00968911312520504, + 0.003730789292603731, + 0.0053468612022697926, + 0.015506059862673283, + 0.004307386931031942, + 0.0070040263235569, + 0.08676034212112427, + 0.020755816251039505, + 0.03616955503821373, + 0.013077383860945702, + 0.014517175033688545, + 0.08801712095737457, + 0.018607083708047867, + 0.018837815150618553, + 0.02083083987236023, + 0.006532502360641956, + 0.007298353593796492, + 0.004213277716189623, + 0.004388771485537291, + 0.027196645736694336, + 0.004388951696455479, + 0.00514261145144701, + 0.06157302111387253, + 0.011432995088398457, + 0.01013245526701212, + 0.006000393070280552, + 0.00747301010414958, + 0.0687749907374382, + 0.006032976321876049, + 0.00822652131319046, + 0.015754161402583122, + 0.00916855689138174, + 0.01061584148555994, + 0.007107931189239025, + 0.0066190920770168304, + 0.01685444824397564, + 0.006942760664969683, + 0.008194136433303356, + 0.006795873865485191, + 0.004884436726570129, + 0.00404769042506814, + 0.0046319239772856236, + 0.0029641122091561556, + 0.005801643710583448, + 0.004147415980696678, + 0.0036172468680888414, + 0.009488970972597599, + 0.0020398444030433893, + 0.001509171910583973, + 0.0007037503528408706, + 0.0017147044418379664, + 0.010199868120253086, + 0.0009224496898241341, + 0.001321519841440022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.07004226744174957 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036182431504130363 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036182431504130363, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036182431504130363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12404707074165344, + 0.03751545399427414, + 0.046829964965581894, + 0.044021815061569214, + 0.04641517251729965, + 0.03265528008341789, + 0.4345974326133728, + 0.10570888221263885, + 0.06288374960422516, + 0.06532520800828934 + ], + "credit_map": [ + 0.024831101298332214, + 0.015292550437152386, + 0.01684214361011982, + 0.009915019385516644, + 0.01241226401180029, + 0.01958613470196724, + 0.008959286846220493, + 0.016208557412028313, + 0.005444066599011421, + 0.004249424673616886, + 0.006190076004713774, + 0.0030342426616698503, + 0.003822609316557646, + 0.005829963833093643, + 0.0037701865658164024, + 0.005174876190721989, + 0.00798837747424841, + 0.004965478088706732, + 0.00702175498008728, + 0.0035069254226982594, + 0.004461180418729782, + 0.007974017411470413, + 0.0036938798148185015, + 0.007218342274427414, + 0.007478214334696531, + 0.003922741394490004, + 0.006285299081355333, + 0.0026500492822378874, + 0.0030617460142821074, + 0.01179022341966629, + 0.003095924388617277, + 0.005737613886594772, + 0.008502769283950329, + 0.004951161332428455, + 0.01044651959091425, + 0.003720158012583852, + 0.004336331970989704, + 0.006955049466341734, + 0.0033709220588207245, + 0.004132254980504513, + 0.006857356987893581, + 0.003176376223564148, + 0.0029068347066640854, + 0.002732743276283145, + 0.001890111598186195, + 0.010030404664576054, + 0.0028582459781318903, + 0.00220320257358253, + 0.15437354147434235, + 0.02575555630028248, + 0.023118965327739716, + 0.013631057925522327, + 0.016883019357919693, + 0.1685396134853363, + 0.013484359718859196, + 0.01881127804517746, + 0.020628049969673157, + 0.012865817174315453, + 0.011354711838066578, + 0.007710248231887817, + 0.009472174569964409, + 0.022690027952194214, + 0.008975749835371971, + 0.01201208122074604, + 0.009779398329555988, + 0.009300820529460907, + 0.010086962953209877, + 0.006604312919080257, + 0.005009302869439125, + 0.011522947810590267, + 0.004997094161808491, + 0.005582901649177074, + 0.020048072561621666, + 0.005027645267546177, + 0.00414699874818325, + 0.0024141983594745398, + 0.0040788808837533, + 0.02304491586983204, + 0.0022462892811745405, + 0.004318207502365112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0036182431504130363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607036218047142 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.01607036218047142, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607036218047142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07879725098609924, + 0.01792966201901436, + 0.02835112437605858, + 0.08077803254127502, + 0.07599090784788132, + 0.028557270765304565, + 0.4530710279941559, + 0.13050605356693268, + 0.08277441561222076, + 0.023244280368089676 + ], + "credit_map": [ + 0.01342814788222313, + 0.0105616245418787, + 0.014242570847272873, + 0.006468542385846376, + 0.006135869305580854, + 0.010884253308176994, + 0.005209943279623985, + 0.011866303160786629, + 0.0026908402796834707, + 0.0024112965911626816, + 0.0031248191371560097, + 0.0014372913865372539, + 0.0016746746841818094, + 0.0029679653234779835, + 0.001552117639221251, + 0.0020706569775938988, + 0.005304394755512476, + 0.003289927961304784, + 0.003689539385959506, + 0.0023062853142619133, + 0.0028296378441154957, + 0.005043895449489355, + 0.0022514446172863245, + 0.0036359981168061495, + 0.011253583244979382, + 0.008854484185576439, + 0.012778916396200657, + 0.007279166020452976, + 0.008170676417648792, + 0.014149750582873821, + 0.008390303701162338, + 0.009901149198412895, + 0.013716387562453747, + 0.007874101400375366, + 0.011590766720473766, + 0.008379330858588219, + 0.007341673597693443, + 0.013271533884108067, + 0.006588748190551996, + 0.00722836796194315, + 0.008685282431542873, + 0.0016658251406624913, + 0.0017439438961446285, + 0.001069081132300198, + 0.001442736480385065, + 0.01116841472685337, + 0.0012792282504960895, + 0.0015027581248432398, + 0.1663120836019516, + 0.024634944275021553, + 0.021487785503268242, + 0.012430360540747643, + 0.016257906332612038, + 0.18105360865592957, + 0.012549555860459805, + 0.01834479346871376, + 0.02564697526395321, + 0.01679973490536213, + 0.014631900005042553, + 0.009661090560257435, + 0.01169486902654171, + 0.02762512117624283, + 0.01069148164242506, + 0.013754884712398052, + 0.012646599672734737, + 0.012986547313630581, + 0.01298452727496624, + 0.008331451565027237, + 0.006871596910059452, + 0.014894140884280205, + 0.006561184301972389, + 0.00749836815521121, + 0.006317323539406061, + 0.0021414824295789003, + 0.001892706728540361, + 0.001404346665367484, + 0.0016375855775550008, + 0.006869779899716377, + 0.0010669255862012506, + 0.0019141307566314936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.01607036218047142 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a large, speckled seashell, its form defined by the elegant, flowing lines of traditional Oriental calligraphy. Its surface, resembling the intricate pattern of a partridge's feathers, is a tapestry of tan, sienna, and cream, rendered in soft, bleeding gradations of delicate ink. Half-buried in dark, shimmering wet sand, the shell's smooth, curved surface catches the diffused light. The scene unfolds with the rich texture, subtle grain, and deep color saturation characteristic of 35mm film. A thin, translucent sheet of water slowly and rhythmically creeps up the shore, its edge swirling as it gently caresses the shell's base before retreating with a quiet sigh, leaving the sand and shell's underside shimmering. The entire scene is bathed in a soft, overcast light, eliminating harsh shadows and enhancing the tranquil, contemplative atmosphere, creating an enchanting fusion of natural beauty and timeless artistic expression.", + "seed": 32178, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_3682_seed_32178_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018948657717555762, + "vlm_logprob": -0.018948657717555762, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.018948657717555762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00982968881726265 + ], + "opposite_logprob_tokens": [ + -4.884829521179199 + ], + "logprob_reward": -0.00982968881726265, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00982968881726265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04165056347846985, + 0.020054513588547707, + 0.03525828942656517, + 0.009103851392865181, + 0.00992119126021862, + 0.6229367256164551, + 0.041000157594680786, + 0.08057670295238495, + 0.07978298515081406, + 0.05971507355570793 + ], + "credit_map": [ + 0.005033115856349468, + 0.0046839225105941296, + 0.004349374212324619, + 0.007841741666197777, + 0.002912005875259638, + 0.00733090890571475, + 0.0037144129164516926, + 0.005785079672932625, + 0.002047382527962327, + 0.0020433319732546806, + 0.003743506968021393, + 0.00501156784594059, + 0.0017819993663579226, + 0.0019488552352413535, + 0.0016749242786318064, + 0.00180294387973845, + 0.0044874041341245174, + 0.0032406311947852373, + 0.005757034290581942, + 0.005548421293497086, + 0.004045345354825258, + 0.00314071262255311, + 0.004902439657598734, + 0.004136299714446068, + 0.0010342937894165516, + 0.00122916791588068, + 0.0011276748264208436, + 0.0013799479929730296, + 0.0010740433353930712, + 0.0012811162741854787, + 0.0010336252162232995, + 0.0009439816349186003, + 0.001275937189348042, + 0.0009791668271645904, + 0.0015232909936457872, + 0.0013069768901914358, + 0.0008480834658257663, + 0.001300010597333312, + 0.0009458626736886799, + 0.0017418633215129375, + 0.10406047850847244, + 0.03707243502140045, + 0.0735408216714859, + 0.025459768250584602, + 0.04251521825790405, + 0.17701895534992218, + 0.04803366959095001, + 0.1152353584766388, + 0.004225218202918768, + 0.002171529922634363, + 0.013075658120214939, + 0.0024162086192518473, + 0.0016928151017054915, + 0.004448140971362591, + 0.0017164861783385277, + 0.011254098266363144, + 0.008043217472732067, + 0.006744612008333206, + 0.02097838930785656, + 0.008224345743656158, + 0.004703720565885305, + 0.009082460775971413, + 0.0037623639218509197, + 0.01903758943080902, + 0.012547256425023079, + 0.008271734230220318, + 0.020730428397655487, + 0.007849812507629395, + 0.005239983554929495, + 0.009445064701139927, + 0.00474415672942996, + 0.010954555124044418, + 0.0069616627879440784, + 0.003970280755311251, + 0.016955357044935226, + 0.00439751660451293, + 0.0034801673609763384, + 0.014163278043270111, + 0.004819048568606377, + 0.004967760294675827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.00982968881726265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259131081402302 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005259131081402302, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259131081402302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0774620845913887, + 0.12005352228879929, + 0.08919212967157364, + 0.02842669002711773, + 0.07481909543275833, + 0.34279605746269226, + 0.13604237139225006, + 0.0692158043384552, + 0.02472512423992157, + 0.03726714849472046 + ], + "credit_map": [ + 0.009742701426148415, + 0.012226521968841553, + 0.00726662902161479, + 0.008682181127369404, + 0.008589781820774078, + 0.011864476837217808, + 0.008086500689387321, + 0.011003293097019196, + 0.018239136785268784, + 0.011425305157899857, + 0.012602928094565868, + 0.01479242742061615, + 0.011038461700081825, + 0.024546852335333824, + 0.008187786675989628, + 0.019220618531107903, + 0.010141042992472649, + 0.005231248214840889, + 0.028196845203638077, + 0.007681237068027258, + 0.005101817660033703, + 0.01422326173633337, + 0.005789926741272211, + 0.012826737947762012, + 0.005563286133110523, + 0.002042611362412572, + 0.007820652797818184, + 0.0030190080869942904, + 0.001483370317146182, + 0.0032344628125429153, + 0.002030320232734084, + 0.0032329752575606108, + 0.013633714988827705, + 0.0041162194684147835, + 0.007704330608248711, + 0.004399357363581657, + 0.0059737893752753735, + 0.0181575920432806, + 0.008949204348027706, + 0.01188488956540823, + 0.05753403902053833, + 0.020635858178138733, + 0.03945435211062431, + 0.014379484578967094, + 0.0236067995429039, + 0.09664198756217957, + 0.027002694085240364, + 0.06354080885648727, + 0.01480900775641203, + 0.007161185145378113, + 0.04496866837143898, + 0.00821116752922535, + 0.005136930849403143, + 0.01384736318141222, + 0.005550283472985029, + 0.036357756704092026, + 0.008032519370317459, + 0.00666464539244771, + 0.012249059975147247, + 0.008002984337508678, + 0.00524090277031064, + 0.012266712263226509, + 0.0040575796738266945, + 0.012701403349637985, + 0.0032287093345075846, + 0.00201615272089839, + 0.004696237854659557, + 0.002026749076321721, + 0.002414627932012081, + 0.0038699526339769363, + 0.001838131807744503, + 0.0046345628798007965, + 0.004004020243883133, + 0.003276654053479433, + 0.007063332013785839, + 0.0037236101925373077, + 0.002657987643033266, + 0.008557264693081379, + 0.00227985717356205, + 0.005704422015696764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.005259131081402302 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054974380880594254 + ], + "opposite_logprob_tokens": [ + -2.929974317550659 + ], + "logprob_reward": -0.054974380880594254, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054974380880594254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11969899386167526, + 0.11098695546388626, + 0.06450792402029037, + 0.031200772151350975, + 0.03737533837556839, + 0.2913931608200073, + 0.16699926555156708, + 0.07325675338506699, + 0.054242510348558426, + 0.050338368862867355 + ], + "credit_map": [ + 0.013631458394229412, + 0.01359296590089798, + 0.016410088166594505, + 0.018976200371980667, + 0.009524867869913578, + 0.02029842883348465, + 0.011181158944964409, + 0.016083817929029465, + 0.009319779463112354, + 0.011476345360279083, + 0.02025521919131279, + 0.031536709517240524, + 0.009228046052157879, + 0.008454132825136185, + 0.012211248278617859, + 0.008505466394126415, + 0.009655237197875977, + 0.004778160247951746, + 0.011417551897466183, + 0.004599816165864468, + 0.00473216874524951, + 0.017449142411351204, + 0.0042852843180298805, + 0.007590560708194971, + 0.004508843179792166, + 0.002416839823126793, + 0.005178048741072416, + 0.004286950919777155, + 0.003314546076580882, + 0.003807642962783575, + 0.0043746125884354115, + 0.0033132857643067837, + 0.006416781805455685, + 0.0033873801585286856, + 0.003251544898375869, + 0.003912828862667084, + 0.0036188741214573383, + 0.007843172177672386, + 0.0043427892960608006, + 0.004601959139108658, + 0.04804541543126106, + 0.018556436523795128, + 0.03509991243481636, + 0.01285302359610796, + 0.021133538335561752, + 0.08264614641666412, + 0.020549530163407326, + 0.0525091178715229, + 0.017644304782152176, + 0.008723071776330471, + 0.05568657070398331, + 0.009594883769750595, + 0.0060083563439548016, + 0.017082862555980682, + 0.006977731361985207, + 0.045281462371349335, + 0.010772992856800556, + 0.009295188821852207, + 0.012441757135093212, + 0.007759603671729565, + 0.005604511126875877, + 0.01316813938319683, + 0.0049857404083013535, + 0.009228807874023914, + 0.015368265099823475, + 0.0032070574816316366, + 0.0054007298313081264, + 0.002606682712212205, + 0.002902002539485693, + 0.008351301774382591, + 0.0030233568977564573, + 0.013383112847805023, + 0.007225629407912493, + 0.004213639535009861, + 0.011915524490177631, + 0.0036237030290067196, + 0.003207732457667589, + 0.0081934854388237, + 0.002811847720295191, + 0.00914679653942585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.054974380880594254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099181387573481 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099181387573481, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099181387573481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07928192615509033, + 0.0769190862774849, + 0.0937383696436882, + 0.054649461060762405, + 0.06291575729846954, + 0.45110756158828735, + 0.02638431265950203, + 0.0377182774245739, + 0.061774466186761856, + 0.05551072582602501 + ], + "credit_map": [ + 0.010147009044885635, + 0.00891161523759365, + 0.007764142472296953, + 0.00920657068490982, + 0.007886901497840881, + 0.014989418908953667, + 0.008857281878590584, + 0.011518999934196472, + 0.010059638880193233, + 0.007323386147618294, + 0.013639802113175392, + 0.0163797065615654, + 0.0067255315370857716, + 0.007046473678201437, + 0.007617348805069923, + 0.008127202279865742, + 0.011128894984722137, + 0.005664070602506399, + 0.029559554532170296, + 0.006112368777394295, + 0.005433063022792339, + 0.018811888992786407, + 0.007786293048411608, + 0.00924224779009819, + 0.008629552088677883, + 0.004661974031478167, + 0.00841930229216814, + 0.007990152575075626, + 0.005189594347029924, + 0.006743154022842646, + 0.007023104466497898, + 0.005992630030959845, + 0.010925635695457458, + 0.005083858035504818, + 0.006390857044607401, + 0.004874852951616049, + 0.005845033098012209, + 0.015040949918329716, + 0.006367063149809837, + 0.00838751532137394, + 0.06916620582342148, + 0.0355822779238224, + 0.04727447032928467, + 0.023679295554757118, + 0.043765608221292496, + 0.1265442967414856, + 0.031551212072372437, + 0.07354424893856049, + 0.002866984810680151, + 0.002120313700288534, + 0.006362975109368563, + 0.002340937964618206, + 0.0012775094946846366, + 0.0029769816901534796, + 0.0021078179124742746, + 0.0063307941891252995, + 0.004929068963974714, + 0.004202346317470074, + 0.00713752768933773, + 0.0046995109878480434, + 0.0028290932532399893, + 0.00643472746014595, + 0.003341841045767069, + 0.004144167527556419, + 0.017378389835357666, + 0.006391952745616436, + 0.014388340525329113, + 0.005237109959125519, + 0.003808305598795414, + 0.005968802608549595, + 0.0032585931476205587, + 0.005342978518456221, + 0.007677962072193623, + 0.003563772654160857, + 0.00454220175743103, + 0.0032170552294701338, + 0.003526560962200165, + 0.014020068570971489, + 0.003980839159339666, + 0.014982271008193493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.004099181387573481 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058090642094612 + ], + "opposite_logprob_tokens": [ + -3.895581007003784 + ], + "logprob_reward": -0.02058090642094612, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058090642094612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053754717111587524, + 0.02738489955663681, + 0.04038785398006439, + 0.0172495748847723, + 0.0232416819781065, + 0.5068385004997253, + 0.09966359287500381, + 0.1228516474366188, + 0.06887617707252502, + 0.039751362055540085 + ], + "credit_map": [ + 0.007329053245484829, + 0.0067726788111031055, + 0.007576627191156149, + 0.0063292887061834335, + 0.004276877269148827, + 0.008258816786110401, + 0.004870573524385691, + 0.008340799249708652, + 0.0036126135382801294, + 0.002235219581052661, + 0.004270932171493769, + 0.00402858667075634, + 0.0021270618308335543, + 0.0048041315749287605, + 0.002509834012016654, + 0.0037965215742588043, + 0.006223715841770172, + 0.004291976802051067, + 0.005402323789894581, + 0.00529842684045434, + 0.004882663954049349, + 0.005750573240220547, + 0.005122116766870022, + 0.0034160560462623835, + 0.0028446139767766, + 0.0014289143728092313, + 0.0023684122134000063, + 0.0025259838439524174, + 0.0016913089202716947, + 0.001954501960426569, + 0.002341019921004772, + 0.0020948206074535847, + 0.003697717096656561, + 0.0031888883095234632, + 0.0030845648143440485, + 0.0023298682644963264, + 0.002227513352409005, + 0.003613836131989956, + 0.001845493447035551, + 0.0032537998631596565, + 0.08492149412631989, + 0.03197436034679413, + 0.0582880936563015, + 0.020746540278196335, + 0.035625774413347244, + 0.14313843846321106, + 0.03819504380226135, + 0.09394871443510056, + 0.010963804088532925, + 0.005601127631962299, + 0.03288837522268295, + 0.005599966738373041, + 0.003747090697288513, + 0.01039203256368637, + 0.0040616849437355995, + 0.026409508660435677, + 0.0153560321778059, + 0.011438756249845028, + 0.027601467445492744, + 0.009932702407240868, + 0.008854305371642113, + 0.017043398693203926, + 0.006297646556049585, + 0.0263273436576128, + 0.013864357024431229, + 0.004693914204835892, + 0.0085015082731843, + 0.0031648012809455395, + 0.005570013541728258, + 0.015553667210042477, + 0.004095652140676975, + 0.013432259671390057, + 0.005879499483853579, + 0.0037157931365072727, + 0.0061520473100245, + 0.00494270958006382, + 0.0036333086900413036, + 0.00552193820476532, + 0.005181214772164822, + 0.004724851343780756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.02058090642094612 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a large, speckled seashell, its form defined by the elegant, flowing lines of traditional Oriental calligraphy. Its surface, resembling the intricate pattern of a partridge's feathers, is a tapestry of tan, sienna, and cream, rendered in soft, bleeding gradations of delicate ink. Half-buried in dark, shimmering wet sand, the shell's smooth, curved surface catches the diffused light. The scene unfolds with the rich texture, subtle grain, and deep color saturation characteristic of 35mm film. A thin, translucent sheet of water slowly and rhythmically creeps up the shore, its edge swirling as it gently caresses the shell's base before retreating with a quiet sigh, leaving the sand and shell's underside shimmering. The entire scene is bathed in a soft, overcast light, eliminating harsh shadows and enhancing the tranquil, contemplative atmosphere, creating an enchanting fusion of natural beauty and timeless artistic expression.", + "seed": 32179, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_3682_seed_32179_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022146473918110133, + "vlm_logprob": -0.022146473918110133, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.022146473918110133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008151240646839142 + ], + "opposite_logprob_tokens": [ + -5.133151054382324 + ], + "logprob_reward": -0.008151240646839142, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008151240646839142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05527287721633911, + 0.06805117428302765, + 0.11147122830152512, + 0.385524183511734, + 0.11327806860208511, + 0.13583143055438995, + 0.014481033198535442, + 0.025319084525108337, + 0.07345395535230637, + 0.01731685735285282 + ], + "credit_map": [ + 0.006410534959286451, + 0.006789398845285177, + 0.00683700293302536, + 0.0069920942187309265, + 0.006108646281063557, + 0.006697932258248329, + 0.006386342458426952, + 0.009050929918885231, + 0.018847377970814705, + 0.003714113961905241, + 0.004962023813277483, + 0.004589562304317951, + 0.004944386426359415, + 0.007763784844428301, + 0.0049462467432022095, + 0.018283681944012642, + 0.025157179683446884, + 0.007231981493532658, + 0.026042014360427856, + 0.005814410746097565, + 0.00926207471638918, + 0.014515435323119164, + 0.010541122406721115, + 0.012907023541629314, + 0.13721168041229248, + 0.022258702665567398, + 0.02224319986999035, + 0.020554710179567337, + 0.013712418265640736, + 0.024313192814588547, + 0.01956060901284218, + 0.1256697177886963, + 0.02963312529027462, + 0.006734293885529041, + 0.010967734269797802, + 0.00745844841003418, + 0.012540573254227638, + 0.012520335614681244, + 0.00728839123621583, + 0.026135172694921494, + 0.020908471196889877, + 0.009599861688911915, + 0.024501552805304527, + 0.01313578151166439, + 0.01655493676662445, + 0.018526559695601463, + 0.011003181338310242, + 0.021601106971502304, + 0.0027267059776932, + 0.0010046737734228373, + 0.002405772916972637, + 0.00152740185149014, + 0.001205145614221692, + 0.001736034406349063, + 0.001549626118503511, + 0.0023256719578057528, + 0.0031712730415165424, + 0.0028259486425668, + 0.003909596242010593, + 0.002238757209852338, + 0.003598578041419387, + 0.004232796374708414, + 0.002480175346136093, + 0.002861961955204606, + 0.007437432184815407, + 0.0025609158910810947, + 0.0265265591442585, + 0.004104967229068279, + 0.0037224662955850363, + 0.005940046161413193, + 0.003245726926252246, + 0.01991584710776806, + 0.002537535270676017, + 0.00254236813634634, + 0.001797099830582738, + 0.0010142255341634154, + 0.0013132032472640276, + 0.0051706875674426556, + 0.0010872847633436322, + 0.001854454749263823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.008151240646839142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032011240255087614 + ], + "opposite_logprob_tokens": [ + -5.753201007843018 + ], + "logprob_reward": -0.0032011240255087614, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032011240255087614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05082559958100319, + 0.02919626422226429, + 0.12612071633338928, + 0.3978322744369507, + 0.08931519836187363, + 0.025954488664865494, + 0.019923238083720207, + 0.0331612192094326, + 0.2159074991941452, + 0.011763499118387699 + ], + "credit_map": [ + 0.00630180723965168, + 0.0048424010165035725, + 0.005929083563387394, + 0.0044000414200127125, + 0.0047571477480232716, + 0.011272643692791462, + 0.006158150732517242, + 0.007164325099438429, + 0.0023646256886422634, + 0.0013367388164624572, + 0.008661577478051186, + 0.002424167701974511, + 0.0019989314023405313, + 0.003297068178653717, + 0.00211321166716516, + 0.00699994433671236, + 0.02765600197017193, + 0.0071665882132947445, + 0.04344595968723297, + 0.008506519719958305, + 0.007003144361078739, + 0.012720699422061443, + 0.006751263048499823, + 0.012870538979768753, + 0.14662684500217438, + 0.019109835848212242, + 0.023037618026137352, + 0.014041492715477943, + 0.012957089580595493, + 0.022765027359128, + 0.01917383074760437, + 0.1401205211877823, + 0.033333081752061844, + 0.0038436756003648043, + 0.004896074999123812, + 0.0036863889545202255, + 0.0033888500183820724, + 0.006535560358315706, + 0.004812866449356079, + 0.028818704187870026, + 0.004301439039409161, + 0.00211937609128654, + 0.0037616149056702852, + 0.002978281583636999, + 0.0036468796897679567, + 0.0034827415365725756, + 0.0026841324288398027, + 0.002980025252327323, + 0.003532369388267398, + 0.001543347374536097, + 0.0021460610441863537, + 0.0013168735895305872, + 0.0022707588504999876, + 0.00521484250202775, + 0.0016989039722830057, + 0.002200080081820488, + 0.0056271785870194435, + 0.0022090633865445852, + 0.00423953402787447, + 0.0019207180012017488, + 0.0032876976765692234, + 0.009240458719432354, + 0.0027396262157708406, + 0.0038969404995441437, + 0.019282830879092216, + 0.010283468291163445, + 0.08355390280485153, + 0.012257489375770092, + 0.010377098806202412, + 0.016814716160297394, + 0.008451135829091072, + 0.05488685891032219, + 0.002197353169322014, + 0.001371648977510631, + 0.001245012041181326, + 0.000868482980877161, + 0.0013058275217190385, + 0.0022158490028232336, + 0.0009311321773566306, + 0.00162819295655936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.0032011240255087614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899671047925949 + ], + "opposite_logprob_tokens": [ + -2.5789966583251953 + ], + "logprob_reward": -0.07899671047925949, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899671047925949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05390476435422897, + 0.1582239866256714, + 0.10003091394901276, + 0.178378164768219, + 0.18581999838352203, + 0.09330495446920395, + 0.026446251198649406, + 0.0704483613371849, + 0.11921986192464828, + 0.014222724363207817 + ], + "credit_map": [ + 0.0068666948936879635, + 0.004205400124192238, + 0.0067954896949231625, + 0.005597714334726334, + 0.004354578442871571, + 0.008147652260959148, + 0.005786230321973562, + 0.012151006609201431, + 0.012110460549592972, + 0.007667050696909428, + 0.046693652868270874, + 0.012807698920369148, + 0.0067790960893034935, + 0.012715606018900871, + 0.009633773006498814, + 0.04981665313243866, + 0.021096229553222656, + 0.006798576097935438, + 0.030368156731128693, + 0.007728675380349159, + 0.005945798475295305, + 0.009910792112350464, + 0.008513928391039371, + 0.009668758139014244, + 0.022121254354715347, + 0.025973351672291756, + 0.017349619418382645, + 0.023699646815657616, + 0.017321022227406502, + 0.028758373111486435, + 0.026938727125525475, + 0.016216160729527473, + 0.06212364509701729, + 0.010500149801373482, + 0.0100317457690835, + 0.007576796691864729, + 0.013069252483546734, + 0.01543614361435175, + 0.01015987154096365, + 0.056922394782304764, + 0.025667622685432434, + 0.006712875794619322, + 0.022310351952910423, + 0.004373755306005478, + 0.00766910333186388, + 0.01032621692866087, + 0.006252454128116369, + 0.009992570616304874, + 0.003982533700764179, + 0.0021932825911790133, + 0.002442412544041872, + 0.002953590825200081, + 0.0032151476480066776, + 0.004976590629667044, + 0.0033583438489586115, + 0.0033243512734770775, + 0.011935578659176826, + 0.004655109718441963, + 0.00791916809976101, + 0.004502159543335438, + 0.0061408281326293945, + 0.018404358997941017, + 0.0067090727388858795, + 0.0101820919662714, + 0.011023690924048424, + 0.0042077586986124516, + 0.044607844203710556, + 0.006933858152478933, + 0.006056254729628563, + 0.009512115269899368, + 0.0050621721893548965, + 0.03181616589426994, + 0.002033042488619685, + 0.0015954716363921762, + 0.0014643139438703656, + 0.0013125586556270719, + 0.0017126832390204072, + 0.002190619707107544, + 0.0018934519030153751, + 0.002020582091063261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.07899671047925949 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021999699529260397 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.0021999699529260397, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021999699529260397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06386242061853409, + 0.037766676396131516, + 0.2139313966035843, + 0.0636158287525177, + 0.20295682549476624, + 0.059208944439888, + 0.028279030695557594, + 0.033880338072776794, + 0.27535945177078247, + 0.021139008924365044 + ], + "credit_map": [ + 0.01024586707353592, + 0.0048850406892597675, + 0.012158749625086784, + 0.005141148343682289, + 0.005246578715741634, + 0.006369518116116524, + 0.006945360917598009, + 0.01287015900015831, + 0.008457496762275696, + 0.003606664016842842, + 0.0028033670969307423, + 0.003898314433172345, + 0.0031530994456261396, + 0.005497616715729237, + 0.0031459846068173647, + 0.007204135414212942, + 0.051203884184360504, + 0.010895110666751862, + 0.0758812353014946, + 0.01348487101495266, + 0.010355761274695396, + 0.019256573170423508, + 0.010077444836497307, + 0.022776536643505096, + 0.013188779354095459, + 0.0068604908883571625, + 0.006375002209097147, + 0.00606335885822773, + 0.0046047791838645935, + 0.003749866271391511, + 0.008293971419334412, + 0.014479582197964191, + 0.06600393354892731, + 0.00914450641721487, + 0.016247332096099854, + 0.009549709036946297, + 0.011126372963190079, + 0.02420053817331791, + 0.010110417380928993, + 0.056574027985334396, + 0.010685211047530174, + 0.003177529200911522, + 0.015425155870616436, + 0.0029975033830851316, + 0.006323506589978933, + 0.00726387370377779, + 0.004409446846693754, + 0.008926722221076488, + 0.003921916708350182, + 0.002401899080723524, + 0.0029503940604627132, + 0.0023482481483370066, + 0.004412857815623283, + 0.006392314098775387, + 0.002904090564697981, + 0.0029473132453858852, + 0.0054964483715593815, + 0.003620023839175701, + 0.0032866087276488543, + 0.0032021179795265198, + 0.005099077243357897, + 0.00608529057353735, + 0.003251333488151431, + 0.0038394404109567404, + 0.02400572970509529, + 0.013918421231210232, + 0.10696391761302948, + 0.01581285335123539, + 0.013565212488174438, + 0.02151499129831791, + 0.010672512464225292, + 0.06890585273504257, + 0.003932328894734383, + 0.0023434346076101065, + 0.003355144988745451, + 0.0017994442023336887, + 0.002132449299097061, + 0.004016731400042772, + 0.0015909395879134536, + 0.001968535827472806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.0021999699529260397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183324486017227 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183324486017227, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183324486017227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0967567190527916, + 0.10927581787109375, + 0.06917347759008408, + 0.12553727626800537, + 0.20743757486343384, + 0.04672490060329437, + 0.02794225886464119, + 0.07274016737937927, + 0.2148178368806839, + 0.02959398180246353 + ], + "credit_map": [ + 0.009579924866557121, + 0.008644767105579376, + 0.013893096707761288, + 0.0063849021680653095, + 0.010372895747423172, + 0.01880662888288498, + 0.009638247080147266, + 0.01943625882267952, + 0.007098664995282888, + 0.0058190105482935905, + 0.0311269648373127, + 0.01103118248283863, + 0.007138576824218035, + 0.012252960354089737, + 0.0064242067746818066, + 0.028384245932102203, + 0.01016569696366787, + 0.005394835956394672, + 0.015069776214659214, + 0.005053823813796043, + 0.006892604753375053, + 0.007726458832621574, + 0.009626843966543674, + 0.009243440814316273, + 0.01565888151526451, + 0.017485838383436203, + 0.017451690509915352, + 0.019020330160856247, + 0.010105852037668228, + 0.016948362812399864, + 0.016334086656570435, + 0.012532235123217106, + 0.0661805272102356, + 0.00959769356995821, + 0.015382316894829273, + 0.009245052002370358, + 0.010916956700384617, + 0.019036434590816498, + 0.011338138952851295, + 0.06574045866727829, + 0.006890407297760248, + 0.003544554812833667, + 0.010094653815031052, + 0.005109101999551058, + 0.006085075903683901, + 0.005469094030559063, + 0.004448288585990667, + 0.005083724856376648, + 0.004243789706379175, + 0.002032609423622489, + 0.003010665997862816, + 0.0018164666835218668, + 0.003435592632740736, + 0.00786872860044241, + 0.00215539475902915, + 0.0033790101297199726, + 0.014433072879910469, + 0.005521951708942652, + 0.00806223601102829, + 0.005989734083414078, + 0.00996371079236269, + 0.01507889200001955, + 0.006193571258336306, + 0.007496993523091078, + 0.020370133221149445, + 0.008801130577921867, + 0.08053857088088989, + 0.011998526751995087, + 0.010475315153598785, + 0.01664668507874012, + 0.009028054773807526, + 0.05695942044258118, + 0.0055876970291137695, + 0.002553217113018036, + 0.004327943082898855, + 0.002016729209572077, + 0.0028179739601910114, + 0.005830571986734867, + 0.003040245035663247, + 0.003419603919610381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.018183324486017227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a large, speckled seashell, its form defined by the elegant, flowing lines of traditional Oriental calligraphy. Its surface, resembling the intricate pattern of a partridge's feathers, is a tapestry of tan, sienna, and cream, rendered in soft, bleeding gradations of delicate ink. Half-buried in dark, shimmering wet sand, the shell's smooth, curved surface catches the diffused light. The scene unfolds with the rich texture, subtle grain, and deep color saturation characteristic of 35mm film. A thin, translucent sheet of water slowly and rhythmically creeps up the shore, its edge swirling as it gently caresses the shell's base before retreating with a quiet sigh, leaving the sand and shell's underside shimmering. The entire scene is bathed in a soft, overcast light, eliminating harsh shadows and enhancing the tranquil, contemplative atmosphere, creating an enchanting fusion of natural beauty and timeless artistic expression.", + "seed": 32180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_7/index_3682_seed_32180_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027491439366713166, + "vlm_logprob": -0.027491439366713166, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.027491439366713166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027395740151405334 + ], + "opposite_logprob_tokens": [ + -3.777395725250244 + ], + "logprob_reward": -0.027395740151405334, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027395740151405334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15844790637493134, + 0.07858975976705551, + 0.10564758628606796, + 0.0634295865893364, + 0.2066715657711029, + 0.0712786540389061, + 0.12850569188594818, + 0.037922002375125885, + 0.1270684152841568, + 0.02243882417678833 + ], + "credit_map": [ + 0.03289252147078514, + 0.013675017282366753, + 0.03424772620201111, + 0.01579776406288147, + 0.013124320656061172, + 0.0153597891330719, + 0.01691504567861557, + 0.016435716301202774, + 0.011054358445107937, + 0.009125054813921452, + 0.011037778109312057, + 0.009180146269500256, + 0.009628505446016788, + 0.007127543445676565, + 0.010130021721124649, + 0.01130635105073452, + 0.015618267469108105, + 0.0063979350961744785, + 0.022123508155345917, + 0.007595069706439972, + 0.007297069765627384, + 0.010977520607411861, + 0.012741093523800373, + 0.022897126153111458, + 0.010859285481274128, + 0.008302946574985981, + 0.007663415279239416, + 0.006597395520657301, + 0.007871533744037151, + 0.008033688180148602, + 0.007758572697639465, + 0.006342747714370489, + 0.020142311230301857, + 0.014357238076627254, + 0.026210833340883255, + 0.038169991225004196, + 0.020590009167790413, + 0.01862614043056965, + 0.04889105632901192, + 0.01968398503959179, + 0.008906525559723377, + 0.005932743661105633, + 0.011717605404555798, + 0.008951999247074127, + 0.007239928934723139, + 0.011042368598282337, + 0.010949508287012577, + 0.006537974812090397, + 0.012216417118906975, + 0.010223042219877243, + 0.012096783146262169, + 0.02281353622674942, + 0.014889799058437347, + 0.013450787402689457, + 0.029858605936169624, + 0.012956718914210796, + 0.005079005379229784, + 0.002614544937387109, + 0.004205940756946802, + 0.0031038993038237095, + 0.004464718513190746, + 0.006159100215882063, + 0.004752829670906067, + 0.00754196010529995, + 0.018791500478982925, + 0.01107261423021555, + 0.014011891558766365, + 0.008901629596948624, + 0.01272583194077015, + 0.02568119950592518, + 0.01749463565647602, + 0.018389102071523666, + 0.0032019196078181267, + 0.0021183162461966276, + 0.0026161547284573317, + 0.0020234589464962482, + 0.00206377194263041, + 0.0032496817875653505, + 0.0031136288307607174, + 0.004051891155540943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.027395740151405334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264941602945328 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264941602945328, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264941602945328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06220236420631409, + 0.05127406865358353, + 0.19216834008693695, + 0.10845094919204712, + 0.03745502606034279, + 0.07837269455194473, + 0.2754696309566498, + 0.06161273643374443, + 0.03934700787067413, + 0.09364720433950424 + ], + "credit_map": [ + 0.012074366211891174, + 0.005820485297590494, + 0.011505911126732826, + 0.005553187802433968, + 0.004924731329083443, + 0.006556709297001362, + 0.0059421490877866745, + 0.009824826382100582, + 0.008635113015770912, + 0.004831492435187101, + 0.006188762374222279, + 0.005246842745691538, + 0.004889193922281265, + 0.006508014164865017, + 0.0068841008469462395, + 0.008090551011264324, + 0.044708698987960815, + 0.011561047285795212, + 0.03315577656030655, + 0.007028630468994379, + 0.00837576575577259, + 0.01933753490447998, + 0.028787696734070778, + 0.03921319171786308, + 0.025034114718437195, + 0.00824750866740942, + 0.013394569046795368, + 0.006972219329327345, + 0.007848454639315605, + 0.012364613823592663, + 0.013998618349432945, + 0.02059084363281727, + 0.005223393905907869, + 0.003555542789399624, + 0.0035675629042088985, + 0.004343247041106224, + 0.004665174055844545, + 0.004185561556369066, + 0.008250619284808636, + 0.0036639259196817875, + 0.012049749493598938, + 0.006623733788728714, + 0.015419318340718746, + 0.005806711036711931, + 0.007276545744389296, + 0.012491753324866295, + 0.007978874258697033, + 0.01072600856423378, + 0.02493387833237648, + 0.02311372198164463, + 0.028698358684778214, + 0.049985334277153015, + 0.03312307596206665, + 0.028103314340114594, + 0.057829685509204865, + 0.02968224696815014, + 0.007704149931669235, + 0.004314335063099861, + 0.00885082595050335, + 0.004644503351300955, + 0.004838145337998867, + 0.010907553136348724, + 0.008056702092289925, + 0.012296517379581928, + 0.006477812770754099, + 0.0030641183257102966, + 0.005135564133524895, + 0.002351491479203105, + 0.0031313758809119463, + 0.008730040863156319, + 0.0045922682620584965, + 0.005864337552338839, + 0.011909881606698036, + 0.008897947147488594, + 0.009494641795754433, + 0.008211486972868443, + 0.008126790635287762, + 0.016683151945471764, + 0.013901819474995136, + 0.01642148196697235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.005264941602945328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901378720998764 + ], + "opposite_logprob_tokens": [ + -2.5790138244628906 + ], + "logprob_reward": -0.07901378720998764, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901378720998764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08030635118484497, + 0.15098440647125244, + 0.21727631986141205, + 0.08849550038576126, + 0.0785975456237793, + 0.06363925337791443, + 0.19931410253047943, + 0.05821128934621811, + 0.038570836186409, + 0.02460438758134842 + ], + "credit_map": [ + 0.01393716037273407, + 0.006772192660719156, + 0.014743844047188759, + 0.008496823720633984, + 0.009635071270167828, + 0.007168679032474756, + 0.01053554005920887, + 0.009017035365104675, + 0.030789809301495552, + 0.008507435210049152, + 0.05762604996562004, + 0.012436061166226864, + 0.00628718500956893, + 0.009381270036101341, + 0.00889753270894289, + 0.017059067264199257, + 0.035666629672050476, + 0.014431223273277283, + 0.05302249267697334, + 0.01184651255607605, + 0.015215628780424595, + 0.013335379771888256, + 0.04566938057541847, + 0.028089070692658424, + 0.012079744599759579, + 0.008702126331627369, + 0.011646026745438576, + 0.007553144358098507, + 0.01044602319598198, + 0.009480418637394905, + 0.013473869301378727, + 0.015114150010049343, + 0.011157780885696411, + 0.009967798367142677, + 0.01347251608967781, + 0.006871549878269434, + 0.010615000501275063, + 0.008610446937382221, + 0.010369249619543552, + 0.00753320287913084, + 0.009843657724559307, + 0.004073323681950569, + 0.013175077736377716, + 0.008172937668859959, + 0.005012407898902893, + 0.007456880994141102, + 0.007901041768491268, + 0.00800392311066389, + 0.018808307126164436, + 0.012976625002920628, + 0.01635502092540264, + 0.043820761144161224, + 0.016962502151727676, + 0.022546619176864624, + 0.046435412019491196, + 0.02140885777771473, + 0.008164357393980026, + 0.0046281758695840836, + 0.006578135769814253, + 0.004120419733226299, + 0.00648524658754468, + 0.013058900833129883, + 0.007127044722437859, + 0.008049010299146175, + 0.005634907633066177, + 0.0032574620563536882, + 0.004617749713361263, + 0.0028725028969347477, + 0.0036340809892863035, + 0.00783508364111185, + 0.004689996130764484, + 0.006029053125530481, + 0.003542362479493022, + 0.002279821317642927, + 0.0027261176146566868, + 0.0018091276288032532, + 0.001920804032124579, + 0.0044233775697648525, + 0.0028269619215279818, + 0.005075816065073013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.07901378720998764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493488835170865 + ], + "opposite_logprob_tokens": [ + -6.002493381500244 + ], + "logprob_reward": -0.002493488835170865, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493488835170865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11282183229923248, + 0.03790156543254852, + 0.16044634580612183, + 0.05916011333465576, + 0.07255163788795471, + 0.02282613515853882, + 0.38797828555107117, + 0.09149575233459473, + 0.017901888117194176, + 0.03691638633608818 + ], + "credit_map": [ + 0.0238645076751709, + 0.009326017461717129, + 0.026713797822594643, + 0.009395838715136051, + 0.007228341419249773, + 0.017327792942523956, + 0.008291476406157017, + 0.010674072429537773, + 0.007097114343196154, + 0.0048247999511659145, + 0.006714144721627235, + 0.0031212179455906153, + 0.004139377735555172, + 0.0026985188014805317, + 0.0041300649754703045, + 0.005176330450922251, + 0.029447486624121666, + 0.01097998395562172, + 0.039994578808546066, + 0.008908108808100224, + 0.00829034112393856, + 0.014046861790120602, + 0.026240777224302292, + 0.02253822609782219, + 0.007373180240392685, + 0.005644298158586025, + 0.006043316796422005, + 0.0061968485824763775, + 0.006297046784311533, + 0.0062576583586633205, + 0.010362928733229637, + 0.01098483894020319, + 0.008843282237648964, + 0.007953488267958164, + 0.008044387213885784, + 0.008878299966454506, + 0.009433811530470848, + 0.00746498815715313, + 0.014050069265067577, + 0.007883316837251186, + 0.00399815384298563, + 0.0022903922945261, + 0.003937578294426203, + 0.0017753725405782461, + 0.0023701719474047422, + 0.0040104868821799755, + 0.0020749373361468315, + 0.0023690443485975266, + 0.034671612083911896, + 0.030540065839886665, + 0.03873062878847122, + 0.07770081609487534, + 0.04487094283103943, + 0.037718020379543304, + 0.08304392546415329, + 0.040702320635318756, + 0.012585372664034367, + 0.006608942057937384, + 0.01186625100672245, + 0.006212144624441862, + 0.007006136234849691, + 0.018453987315297127, + 0.010312196798622608, + 0.018450727686285973, + 0.002833422040566802, + 0.00185789552051574, + 0.0022936349269002676, + 0.001588733633980155, + 0.0016351501690223813, + 0.0029407048132270575, + 0.002079735277220607, + 0.0026726138312369585, + 0.004909133072942495, + 0.0037464897613972425, + 0.005024046171456575, + 0.003347852034494281, + 0.004179391544312239, + 0.0040372442454099655, + 0.005719766486436129, + 0.00595246534794569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.002493488835170865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large, speckled seashell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seashell half-buried in dark, shimmering wet sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin, translucent sheet of water slowly creep up the shore and retreat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static, low-angle close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and overcast, eliminating harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289239034056664 + ], + "opposite_logprob_tokens": [ + -3.773289203643799 + ], + "logprob_reward": -0.023289239034056664, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289239034056664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08586855232715607, + 0.0884542167186737, + 0.2454567700624466, + 0.10101142525672913, + 0.055132292211055756, + 0.040586791932582855, + 0.32342055439949036, + 0.024970663711428642, + 0.021595003083348274, + 0.013503653928637505 + ], + "credit_map": [ + 0.014872748404741287, + 0.006538386456668377, + 0.015530728735029697, + 0.008222563192248344, + 0.0074151791632175446, + 0.011979111470282078, + 0.01044086366891861, + 0.010868983343243599, + 0.024374324828386307, + 0.004859274718910456, + 0.03153051435947418, + 0.00692692631855607, + 0.004855047445744276, + 0.004294583108276129, + 0.006756508257240057, + 0.004857048857957125, + 0.047887299209833145, + 0.015961112454533577, + 0.03262439742684364, + 0.011458619497716427, + 0.01240642461925745, + 0.02278464287519455, + 0.04783828556537628, + 0.05449599772691727, + 0.02128377929329872, + 0.008373958058655262, + 0.014673029072582722, + 0.00975398812443018, + 0.006648252252489328, + 0.00932473037391901, + 0.016143765300512314, + 0.014809923246502876, + 0.005256511736661196, + 0.0034678163938224316, + 0.005319814197719097, + 0.011206164956092834, + 0.005553559400141239, + 0.003616516711190343, + 0.01461015921086073, + 0.006101757287979126, + 0.006491384468972683, + 0.003715626196935773, + 0.007766868453472853, + 0.004767924081534147, + 0.004439496900886297, + 0.006019493564963341, + 0.0038098974619060755, + 0.0035761038307100534, + 0.03180792182683945, + 0.023454656824469566, + 0.030368193984031677, + 0.06971580535173416, + 0.03308528661727905, + 0.028483975678682327, + 0.06972875446081161, + 0.03677598759531975, + 0.004215918481349945, + 0.0023884575348347425, + 0.002296237740665674, + 0.002117742784321308, + 0.0027785999700427055, + 0.005443149711936712, + 0.0024902925360947847, + 0.00324026751331985, + 0.0030991798266768456, + 0.0018044115277007222, + 0.00292061734944582, + 0.0014071163022890687, + 0.0021839747205376625, + 0.005094255320727825, + 0.0024904056917876005, + 0.0025950444396585226, + 0.0021206869278103113, + 0.000998709467239678, + 0.002001982880756259, + 0.0008269495447166264, + 0.0010649611940607429, + 0.002351218368858099, + 0.0014035251224413514, + 0.002735620830208063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.023289239034056664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_8/index_tensor([5466, 5466])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_8/index_tensor([5466, 5466])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..77674a6d25510261729ddcc428c23229c27e21e2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_8/index_tensor([5466, 5466])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The drone camera quickly pulls back to an extreme wide shot, capturing a breathtaking panorama of a powerful snowmobile launching from a jagged, snow-capped peak. The scene is set against a vast, cloudless azure sky. The intense, bright, high-altitude sun casts sharp, dramatic shadows across the pristine, untouched powder, causing millions of ice crystals to glitter. The snowmobile itself is a high-performance vehicle, equipped with a glossy electric blue-black chassis, capable of explosive acceleration to the edge. Its rider, clad in a full-body black snowsuit and a matching helmet with a tinted visor, pulls forward tightly on the handlebars. As the vehicle catapults over the sharp cornice into the vast emptiness, its rear track violently ejects a massive, billowing cloud of fine white powder that erupts behind it, creating a breathtaking spectacle. For a moment, the snowmobile hangs suspended in mid-air, creating a dynamic silhouette of the endless mountain range, evoking a powerful sense of adrenaline and a breathtaking sense of freedom.", + "seed": 33032, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_8/index_5466_seed_33032_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3206802872940898, + "vlm_logprob": -0.3206802872940898, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.3206802872940898, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5091217756271362 + ], + "opposite_logprob_tokens": [ + -0.2591218054294586 + ], + "logprob_reward": -1.5091217756271362, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5091217756271362, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1496472805738449, + 0.0966271236538887, + 0.07907487452030182, + 0.13717317581176758, + 0.049932606518268585, + 0.08151658624410629, + 0.11345142871141434, + 0.08863094449043274, + 0.14165395498275757, + 0.06229199469089508 + ], + "credit_map": [ + 0.02307830937206745, + 0.023581670597195625, + 0.031049130484461784, + 0.014957345090806484, + 0.01587667316198349, + 0.011861100792884827, + 0.015098380856215954, + 0.014144673943519592, + 0.012040391564369202, + 0.01447345782071352, + 0.016897113993763924, + 0.010616596788167953, + 0.011699358932673931, + 0.0074149202555418015, + 0.010782848112285137, + 0.012702438980340958, + 0.009801993146538734, + 0.010644379071891308, + 0.009437188506126404, + 0.012082892470061779, + 0.00928541086614132, + 0.008357125334441662, + 0.007733339909464121, + 0.011732546612620354, + 0.023465322330594063, + 0.01795349270105362, + 0.02512860856950283, + 0.018353266641497612, + 0.008444881066679955, + 0.01188657246530056, + 0.012523341923952103, + 0.019417695701122284, + 0.008816261775791645, + 0.007836636155843735, + 0.0062309070490300655, + 0.0036938406992703676, + 0.0053632319904863834, + 0.00640909094363451, + 0.004412218928337097, + 0.0071704150177538395, + 0.01040591113269329, + 0.010517921298742294, + 0.011194608174264431, + 0.007811522111296654, + 0.007134812884032726, + 0.006958429701626301, + 0.009258351288735867, + 0.01823502592742443, + 0.012235400266945362, + 0.01220718864351511, + 0.022328341379761696, + 0.017565952613949776, + 0.008703785017132759, + 0.010777799412608147, + 0.014027558267116547, + 0.015605402179062366, + 0.010807798244059086, + 0.012572212144732475, + 0.013531007803976536, + 0.012412754818797112, + 0.008739817887544632, + 0.007942743599414825, + 0.011851096525788307, + 0.010773510672152042, + 0.020523829385638237, + 0.015758102759718895, + 0.019602570682764053, + 0.021518653258681297, + 0.013101635500788689, + 0.01465352438390255, + 0.014797035604715347, + 0.0216985996812582, + 0.009839670732617378, + 0.009846985340118408, + 0.006557806394994259, + 0.00809525977820158, + 0.0068321432918310165, + 0.007753589190542698, + 0.006500777788460255, + 0.006865758914500475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -1.5091217756271362 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011117536574602127 + ], + "opposite_logprob_tokens": [ + -4.511117458343506 + ], + "logprob_reward": -0.011117536574602127, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011117536574602127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21443931758403778, + 0.04295109957456589, + 0.03665327653288841, + 0.062049947679042816, + 0.04759356752038002, + 0.06283419579267502, + 0.06621092557907104, + 0.05824785307049751, + 0.09785962104797363, + 0.3111601769924164 + ], + "credit_map": [ + 0.028345132246613503, + 0.03192848339676857, + 0.048937369138002396, + 0.02484888583421707, + 0.024273375049233437, + 0.018012341111898422, + 0.016100702807307243, + 0.021993037313222885, + 0.005378181580454111, + 0.0056292274966835976, + 0.006703103892505169, + 0.004990163259208202, + 0.007814576849341393, + 0.004750669002532959, + 0.004063209053128958, + 0.0036219661124050617, + 0.005498427897691727, + 0.005461422260850668, + 0.004581015557050705, + 0.006700228434056044, + 0.00554462056607008, + 0.002644779859110713, + 0.003337249858304858, + 0.002885532798245549, + 0.010660575702786446, + 0.009207407012581825, + 0.008530660532414913, + 0.010809381492435932, + 0.00493696378543973, + 0.0036107744090259075, + 0.007379258517175913, + 0.006914925295859575, + 0.00940334890037775, + 0.007862811908125877, + 0.006296319421380758, + 0.005768381524831057, + 0.004283001646399498, + 0.003054571570828557, + 0.004724844824522734, + 0.006200290285050869, + 0.011119992472231388, + 0.008619943633675575, + 0.009974452666938305, + 0.005932622589170933, + 0.004719028249382973, + 0.006631932221353054, + 0.006963647902011871, + 0.008872577920556068, + 0.007628805004060268, + 0.007917163893580437, + 0.010800446383655071, + 0.00817389041185379, + 0.005977810360491276, + 0.006902832072228193, + 0.00784446019679308, + 0.01096551213413477, + 0.007339191623032093, + 0.007815108634531498, + 0.007338344119489193, + 0.00804103072732687, + 0.006377765443176031, + 0.006791454274207354, + 0.007472853641957045, + 0.007072098553180695, + 0.015450785867869854, + 0.012315933592617512, + 0.015916554257273674, + 0.011979402974247932, + 0.008676371537148952, + 0.008830172941088676, + 0.012980748899281025, + 0.011709650047123432, + 0.0581057146191597, + 0.04573483765125275, + 0.03509391471743584, + 0.04134344682097435, + 0.02725978195667267, + 0.032950229942798615, + 0.042040545493364334, + 0.02863171137869358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.011117536574602127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870280995965004 + ], + "opposite_logprob_tokens": [ + -3.0487027168273926 + ], + "logprob_reward": -0.04870280995965004, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870280995965004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1562308371067047, + 0.31871968507766724, + 0.17842070758342743, + 0.1573910117149353, + 0.02383454702794552, + 0.030854981392621994, + 0.038660336285829544, + 0.031189214438199997, + 0.022584401071071625, + 0.04211430251598358 + ], + "credit_map": [ + 0.020978672429919243, + 0.01991717517375946, + 0.024616064503788948, + 0.01186390407383442, + 0.02211560122668743, + 0.014857624657452106, + 0.014901885762810707, + 0.026979917660355568, + 0.04674498364329338, + 0.04145169258117676, + 0.03409423679113388, + 0.028645655140280724, + 0.07610767334699631, + 0.025498701259493828, + 0.031747836619615555, + 0.034428920596838, + 0.021079223603010178, + 0.023626044392585754, + 0.01448080874979496, + 0.019104499369859695, + 0.03955890238285065, + 0.01663450337946415, + 0.024791637435555458, + 0.019145093858242035, + 0.022650128230452538, + 0.02009626105427742, + 0.02404141053557396, + 0.021135108545422554, + 0.012333234772086143, + 0.01206999458372593, + 0.01736300438642502, + 0.02770187519490719, + 0.003109462559223175, + 0.002891791518777609, + 0.0022009427193552256, + 0.002617479767650366, + 0.002604482229799032, + 0.0025149176362901926, + 0.0036693899892270565, + 0.004226080607622862, + 0.004650355316698551, + 0.0043419464491307735, + 0.004534832201898098, + 0.0028396283742040396, + 0.0025053478311747313, + 0.003965245559811592, + 0.002375925425440073, + 0.005641700234264135, + 0.004990747664123774, + 0.004221728071570396, + 0.006872331723570824, + 0.005277625285089016, + 0.0036868988536298275, + 0.003964509814977646, + 0.004592996556311846, + 0.005053497385233641, + 0.003907087258994579, + 0.003996996209025383, + 0.003634162712842226, + 0.0036826669238507748, + 0.0040382579900324345, + 0.005150757264345884, + 0.0033959192223846912, + 0.003383366856724024, + 0.00263267126865685, + 0.0026520213577896357, + 0.002723919926211238, + 0.002869846299290657, + 0.0023568738251924515, + 0.003004950936883688, + 0.0026929015293717384, + 0.0036512145306915045, + 0.0038085496053099632, + 0.005000689532607794, + 0.006428068969398737, + 0.00545507250353694, + 0.0058136447332799435, + 0.006027144379913807, + 0.005185278598219156, + 0.004395854193717241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.04870280995965004 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819265007972717 + ], + "opposite_logprob_tokens": [ + -3.5298192501068115 + ], + "logprob_reward": -0.029819265007972717, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819265007972717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11836294084787369, + 0.09095551073551178, + 0.12791837751865387, + 0.16609066724777222, + 0.1583552360534668, + 0.08372844010591507, + 0.0736645832657814, + 0.08562518656253815, + 0.05039818584918976, + 0.044900938868522644 + ], + "credit_map": [ + 0.013166193850338459, + 0.0174709502607584, + 0.023290198296308517, + 0.011831462383270264, + 0.014657367952167988, + 0.01147080771625042, + 0.01078113541007042, + 0.015694821253418922, + 0.011764369904994965, + 0.013982854783535004, + 0.02304702438414097, + 0.00853728223592043, + 0.008195582777261734, + 0.006445115432143211, + 0.010338419117033482, + 0.00864485464990139, + 0.015248454175889492, + 0.01663411222398281, + 0.025705089792609215, + 0.024254288524389267, + 0.00971924141049385, + 0.010141021572053432, + 0.010277505032718182, + 0.015938667580485344, + 0.0309480968862772, + 0.02482897974550724, + 0.025468986481428146, + 0.02532474882900715, + 0.01065058633685112, + 0.009550838731229305, + 0.01812991127371788, + 0.02118850313127041, + 0.028207719326019287, + 0.024506788700819016, + 0.01158551499247551, + 0.011701276525855064, + 0.01769110932946205, + 0.01899653673171997, + 0.021901654079556465, + 0.023764610290527344, + 0.008726214990019798, + 0.009482089430093765, + 0.013851219788193703, + 0.00885874405503273, + 0.006937025580555201, + 0.008541041985154152, + 0.008083881810307503, + 0.019248221069574356, + 0.010483789257705212, + 0.008328442461788654, + 0.010879828594624996, + 0.007574769202619791, + 0.0056187366135418415, + 0.006668465677648783, + 0.00874342117458582, + 0.01536712795495987, + 0.009901182726025581, + 0.013298096135258675, + 0.0090647516772151, + 0.01026246976107359, + 0.009854436852037907, + 0.010237807407975197, + 0.012107870541512966, + 0.010898569598793983, + 0.006264092400670052, + 0.00601563835516572, + 0.005953266751021147, + 0.007414912339299917, + 0.0050738356076180935, + 0.00675544049590826, + 0.006708095781505108, + 0.006212898530066013, + 0.006076187361031771, + 0.006895188707858324, + 0.0051392060704529285, + 0.0060534668155014515, + 0.004540187772363424, + 0.005368831567466259, + 0.005405096802860498, + 0.005422770045697689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.029819265007972717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640049301087856 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640049301087856, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640049301087856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08648497611284256, + 0.05124548822641373, + 0.04448585584759712, + 0.10249727219343185, + 0.04162396118044853, + 0.09663069993257523, + 0.1246090680360794, + 0.2028122842311859, + 0.07011867314577103, + 0.17949168384075165 + ], + "credit_map": [ + 0.015744918957352638, + 0.013484954833984375, + 0.017069309949874878, + 0.006688718218356371, + 0.008425715379416943, + 0.00773251848295331, + 0.009280447848141193, + 0.008058394305408001, + 0.0055297804065048695, + 0.006687780376523733, + 0.007960332557559013, + 0.006340997293591499, + 0.006594428326934576, + 0.0050217993557453156, + 0.006277346983551979, + 0.00683302478864789, + 0.0063180094584822655, + 0.005634939763695002, + 0.007447251584380865, + 0.006940973922610283, + 0.006025935988873243, + 0.004004323855042458, + 0.00447589298710227, + 0.003638532478362322, + 0.01713000237941742, + 0.015212876722216606, + 0.01479114219546318, + 0.014712770469486713, + 0.008294963277876377, + 0.0071082953363657, + 0.010785426944494247, + 0.014461799524724483, + 0.006279892288148403, + 0.006613230798393488, + 0.0034309972543269396, + 0.003171349409967661, + 0.005454611498862505, + 0.005050033330917358, + 0.003979003522545099, + 0.00764484703540802, + 0.013112685643136501, + 0.013075720518827438, + 0.010605640709400177, + 0.008667333982884884, + 0.015770111232995987, + 0.012955134734511375, + 0.009681476280093193, + 0.012762599624693394, + 0.012652111239731312, + 0.014169326983392239, + 0.0188377033919096, + 0.017896179109811783, + 0.011134746484458447, + 0.015108163468539715, + 0.015299402177333832, + 0.01951144076883793, + 0.026328913867473602, + 0.031187627464532852, + 0.013797199353575706, + 0.01619459129869938, + 0.035382796078920364, + 0.029357077553868294, + 0.022559434175491333, + 0.028004666790366173, + 0.0113744605332613, + 0.009242571890354156, + 0.01067384798079729, + 0.008575962856411934, + 0.006111851893365383, + 0.006212460808455944, + 0.007682191673666239, + 0.01024533249437809, + 0.020468616858124733, + 0.023765070363879204, + 0.019158517941832542, + 0.021143250167369843, + 0.020139826461672783, + 0.026984382420778275, + 0.02133476920425892, + 0.026497256010770798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.004640049301087856 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The drone camera quickly pulls back to an extreme wide shot, capturing a breathtaking panorama of a powerful snowmobile launching from a jagged, snow-capped peak. The scene is set against a vast, cloudless azure sky. The intense, bright, high-altitude sun casts sharp, dramatic shadows across the pristine, untouched powder, causing millions of ice crystals to glitter. The snowmobile itself is a high-performance vehicle, equipped with a glossy electric blue-black chassis, capable of explosive acceleration to the edge. Its rider, clad in a full-body black snowsuit and a matching helmet with a tinted visor, pulls forward tightly on the handlebars. As the vehicle catapults over the sharp cornice into the vast emptiness, its rear track violently ejects a massive, billowing cloud of fine white powder that erupts behind it, creating a breathtaking spectacle. For a moment, the snowmobile hangs suspended in mid-air, creating a dynamic silhouette of the endless mountain range, evoking a powerful sense of adrenaline and a breathtaking sense of freedom.", + "seed": 33033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_8/index_5466_seed_33033_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16226077489554883, + "vlm_logprob": -0.16226077489554883, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.16226077489554883, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6994143724441528 + ], + "opposite_logprob_tokens": [ + -0.6994143724441528 + ], + "logprob_reward": -0.6994143724441528, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6994143724441528, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05477268993854523, + 0.162337988615036, + 0.020790299400687218, + 0.028306737542152405, + 0.1762312352657318, + 0.0787695124745369, + 0.231474369764328, + 0.09729349613189697, + 0.12646479904651642, + 0.023558929562568665 + ], + "credit_map": [ + 0.0065027461387217045, + 0.004775260109454393, + 0.011835819110274315, + 0.01102676521986723, + 0.0036573978140950203, + 0.0035213343799114227, + 0.004820498172193766, + 0.008632870391011238, + 0.015621036291122437, + 0.01726478338241577, + 0.019479667767882347, + 0.015025495551526546, + 0.03246377781033516, + 0.02188313566148281, + 0.021424392238259315, + 0.0191756933927536, + 0.003064208198338747, + 0.002930904272943735, + 0.003548359964042902, + 0.0032449590507894754, + 0.0015469935024157166, + 0.001366482232697308, + 0.0018289170693606138, + 0.0032594737131148577, + 0.006905783899128437, + 0.005742755252867937, + 0.0027827522717416286, + 0.0035908131394535303, + 0.002821976086124778, + 0.0018941758899018168, + 0.0018992749974131584, + 0.002669207053259015, + 0.045261964201927185, + 0.04475877061486244, + 0.01534611452370882, + 0.019784269854426384, + 0.012695650570094585, + 0.011136890389025211, + 0.010769497603178024, + 0.01647806540131569, + 0.0174811203032732, + 0.012578763999044895, + 0.010070063173770905, + 0.008851625956594944, + 0.005895710550248623, + 0.005895871669054031, + 0.010078204795718193, + 0.007918152026832104, + 0.03283194452524185, + 0.08489257842302322, + 0.008776681497693062, + 0.009558877907693386, + 0.019871728494763374, + 0.013787771575152874, + 0.01469418965280056, + 0.04706060513854027, + 0.00994928739964962, + 0.01314078364521265, + 0.010033135302364826, + 0.00850431714206934, + 0.01449357159435749, + 0.013330971822142601, + 0.016370153054594994, + 0.011471273377537727, + 0.010482165031135082, + 0.035405147820711136, + 0.010682396590709686, + 0.010523590259253979, + 0.03245263174176216, + 0.009863262996077538, + 0.00972082931548357, + 0.007334776222705841, + 0.004164416808634996, + 0.004712915513664484, + 0.002624206943437457, + 0.0020758602768182755, + 0.002284479094669223, + 0.0037526781670749187, + 0.0019902146887034178, + 0.001954158768057823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.6994143724441528 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008674204349517822 + ], + "opposite_logprob_tokens": [ + -4.758674144744873 + ], + "logprob_reward": -0.008674204349517822, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008674204349517822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04151264950633049, + 0.2669847309589386, + 0.01707218773663044, + 0.026445358991622925, + 0.1604849249124527, + 0.012726548127830029, + 0.13564632833003998, + 0.18390654027462006, + 0.1162991151213646, + 0.03892163559794426 + ], + "credit_map": [ + 0.004772478248924017, + 0.004387447610497475, + 0.005993938073515892, + 0.006727238651365042, + 0.002959293546155095, + 0.004816730506718159, + 0.004925098270177841, + 0.006930425297468901, + 0.03221699222922325, + 0.028227191418409348, + 0.03713103011250496, + 0.023759564384818077, + 0.047939177602529526, + 0.0324694849550724, + 0.03392650559544563, + 0.03131479024887085, + 0.0032839998602867126, + 0.002395175863057375, + 0.0027009714394807816, + 0.0018373328493908048, + 0.001901533454656601, + 0.0011110160266980529, + 0.0016806351486593485, + 0.0021615219302475452, + 0.0034909627865999937, + 0.0033086109906435013, + 0.0031345728784799576, + 0.004751753527671099, + 0.0035120940301567316, + 0.0023236337583512068, + 0.0024908226914703846, + 0.003432909492403269, + 0.03572455793619156, + 0.03321406990289688, + 0.019209306687116623, + 0.02079816535115242, + 0.012420518323779106, + 0.014081519097089767, + 0.011050020344555378, + 0.013986765407025814, + 0.002082534832879901, + 0.0019180496456101537, + 0.0016171411843970418, + 0.001156502403318882, + 0.0016455521108582616, + 0.0009891537483781576, + 0.0019489553524181247, + 0.0013686586171388626, + 0.020127372816205025, + 0.053019266575574875, + 0.006219504401087761, + 0.00456894189119339, + 0.010245663113892078, + 0.006977717392146587, + 0.005328922066837549, + 0.0291589368134737, + 0.029331255704164505, + 0.02559654414653778, + 0.032649312168359756, + 0.016967441886663437, + 0.026507677510380745, + 0.030608413740992546, + 0.008728629909455776, + 0.013517267070710659, + 0.004954831674695015, + 0.046068813651800156, + 0.005513663869351149, + 0.0035552328918129206, + 0.04584638029336929, + 0.004738633055239916, + 0.0027613581623882055, + 0.0028602061793208122, + 0.00892720278352499, + 0.011980503797531128, + 0.0042445543222129345, + 0.002020403044298291, + 0.003622716758400202, + 0.0033352081663906574, + 0.0024987163487821817, + 0.0022923313081264496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.008674204349517822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007361203432083 + ], + "opposite_logprob_tokens": [ + -2.6950736045837402 + ], + "logprob_reward": -0.07007361203432083, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007361203432083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2115931361913681, + 0.047584403306245804, + 0.15916238725185394, + 0.1616801619529724, + 0.1668129712343216, + 0.03556538000702858, + 0.07462377846240997, + 0.038076359778642654, + 0.059833720326423645, + 0.045067694038152695 + ], + "credit_map": [ + 0.016527125611901283, + 0.014283505268394947, + 0.05381327122449875, + 0.04049212858080864, + 0.01405303180217743, + 0.013798742555081844, + 0.02448916621506214, + 0.03413617983460426, + 0.005420244764536619, + 0.004943886771798134, + 0.00901847518980503, + 0.005757134407758713, + 0.005685491021722555, + 0.006118465680629015, + 0.0051427120342850685, + 0.0054979934357106686, + 0.018825476989150047, + 0.018653670325875282, + 0.027296504005789757, + 0.017609601840376854, + 0.022626984864473343, + 0.014360521920025349, + 0.016306940466165543, + 0.02348269335925579, + 0.027661070227622986, + 0.023456061258912086, + 0.013205799274146557, + 0.021182779222726822, + 0.030590958893299103, + 0.010982167907059193, + 0.01593220792710781, + 0.018669119104743004, + 0.02498547174036503, + 0.02921873703598976, + 0.023252392187714577, + 0.0248810276389122, + 0.015760507434606552, + 0.01512130256742239, + 0.01397483516484499, + 0.019618691876530647, + 0.003677929053083062, + 0.004003829322755337, + 0.005274557508528233, + 0.005052080377936363, + 0.004300910979509354, + 0.003805470187216997, + 0.004840505309402943, + 0.004610097035765648, + 0.010649019852280617, + 0.028661148622632027, + 0.0030066994950175285, + 0.0027886023744940758, + 0.0059123532846570015, + 0.0042768437415361404, + 0.0039480444975197315, + 0.015381061471998692, + 0.004882956389337778, + 0.006057558581233025, + 0.0042600748129189014, + 0.0030169528909027576, + 0.006834316998720169, + 0.0048871273174881935, + 0.004649853799492121, + 0.003487520618364215, + 0.004831623286008835, + 0.01695438288152218, + 0.004234475549310446, + 0.003735271282494068, + 0.017464227974414825, + 0.005688972771167755, + 0.0038867478724569082, + 0.0030380208045244217, + 0.010573389939963818, + 0.01355622336268425, + 0.003517479868605733, + 0.003080270253121853, + 0.004736736882477999, + 0.004465090576559305, + 0.002297267783433199, + 0.0028412353713065386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.07007361203432083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595736801624298 + ], + "opposite_logprob_tokens": [ + -3.8955957889556885 + ], + "logprob_reward": -0.020595736801624298, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595736801624298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14575345814228058, + 0.07531765103340149, + 0.09526314586400986, + 0.04177457466721535, + 0.17689603567123413, + 0.20185969769954681, + 0.127226322889328, + 0.031200792640447617, + 0.0850161612033844, + 0.01969221793115139 + ], + "credit_map": [ + 0.01067831926047802, + 0.01139094214886427, + 0.0368693508207798, + 0.023930994793772697, + 0.009720821864902973, + 0.016699593514204025, + 0.017671674489974976, + 0.018791763111948967, + 0.007838720455765724, + 0.007939464412629604, + 0.012407984584569931, + 0.009614570066332817, + 0.00940338522195816, + 0.0079936059191823, + 0.009366470389068127, + 0.010753451846539974, + 0.017739545553922653, + 0.014193758368492126, + 0.016728930175304413, + 0.013035774230957031, + 0.007352734915912151, + 0.007484743371605873, + 0.007082862313836813, + 0.011644791811704636, + 0.004665129352360964, + 0.00583418644964695, + 0.00543750636279583, + 0.0054887449368834496, + 0.007747694384306669, + 0.003993877675384283, + 0.0036880222614854574, + 0.004919414408504963, + 0.026833385229110718, + 0.029089339077472687, + 0.02496020682156086, + 0.034357961267232895, + 0.009573102928698063, + 0.016105394810438156, + 0.01266384031623602, + 0.02331281267106533, + 0.055342718958854675, + 0.04146619513630867, + 0.03199661895632744, + 0.020315952599048615, + 0.011492240242660046, + 0.0112077621743083, + 0.012543274089694023, + 0.0174949299544096, + 0.01166705321520567, + 0.013813387602567673, + 0.013031170703470707, + 0.010848085395991802, + 0.01009551715105772, + 0.03163871914148331, + 0.017217060551047325, + 0.01891533099114895, + 0.0032993771601468325, + 0.00381087907589972, + 0.003291747532784939, + 0.00397480046376586, + 0.0036349317524582148, + 0.004068227484822273, + 0.004993973299860954, + 0.004126856103539467, + 0.005380726885050535, + 0.02420067973434925, + 0.004858251195400953, + 0.005318627692759037, + 0.025394100695848465, + 0.01109679788351059, + 0.0038508488796651363, + 0.004916127305477858, + 0.002818051492795348, + 0.0028237386140972376, + 0.002579864114522934, + 0.002442691707983613, + 0.0021967622451484203, + 0.002857232466340065, + 0.001822206424549222, + 0.002151669003069401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.020595736801624298 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545948848128319 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545948848128319, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545948848128319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11048590391874313, + 0.06528227776288986, + 0.06722401827573776, + 0.08213374018669128, + 0.12197776883840561, + 0.11887066066265106, + 0.1851930320262909, + 0.12011781334877014, + 0.08939209580421448, + 0.03932269290089607 + ], + "credit_map": [ + 0.010766718536615372, + 0.009280839934945107, + 0.022995134815573692, + 0.020106740295886993, + 0.008720560930669308, + 0.01011358480900526, + 0.01109314151108265, + 0.0174091849476099, + 0.0061667682603001595, + 0.0064599462784826756, + 0.008862572722136974, + 0.00834257248789072, + 0.011228013783693314, + 0.007378787733614445, + 0.008585541509091854, + 0.008258072659373283, + 0.016033515334129333, + 0.01263743732124567, + 0.009968841448426247, + 0.006979001220315695, + 0.00527543481439352, + 0.004022795241326094, + 0.004949015565216541, + 0.007357981987297535, + 0.017391109839081764, + 0.013719937764108181, + 0.009808224625885487, + 0.01204769778996706, + 0.006529749371111393, + 0.008454736322164536, + 0.005975548177957535, + 0.008206736296415329, + 0.01918250136077404, + 0.020182302221655846, + 0.014538297429680824, + 0.016707489266991615, + 0.014681237749755383, + 0.010577736422419548, + 0.012916098348796368, + 0.013192103244364262, + 0.017541054636240005, + 0.014973962679505348, + 0.018221134319901466, + 0.015652917325496674, + 0.013417116366326809, + 0.010655260644853115, + 0.015290568582713604, + 0.013118645176291466, + 0.022950373589992523, + 0.05571027845144272, + 0.012525714002549648, + 0.010113544762134552, + 0.012647558003664017, + 0.02421502023935318, + 0.014231524430215359, + 0.032799020409584045, + 0.01648024469614029, + 0.016670726239681244, + 0.014547068625688553, + 0.011752036400139332, + 0.019957298412919044, + 0.011214339174330235, + 0.019817864522337914, + 0.009678234346210957, + 0.01032298430800438, + 0.019610192626714706, + 0.01055221352726221, + 0.009308438748121262, + 0.01851228065788746, + 0.008737915195524693, + 0.006068767514079809, + 0.006279305554926395, + 0.0055047739297151566, + 0.00599335040897131, + 0.006326434202492237, + 0.002663255436345935, + 0.004087110981345177, + 0.007299173157662153, + 0.003365470329299569, + 0.004083123058080673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.012545948848128319 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The drone camera quickly pulls back to an extreme wide shot, capturing a breathtaking panorama of a powerful snowmobile launching from a jagged, snow-capped peak. The scene is set against a vast, cloudless azure sky. The intense, bright, high-altitude sun casts sharp, dramatic shadows across the pristine, untouched powder, causing millions of ice crystals to glitter. The snowmobile itself is a high-performance vehicle, equipped with a glossy electric blue-black chassis, capable of explosive acceleration to the edge. Its rider, clad in a full-body black snowsuit and a matching helmet with a tinted visor, pulls forward tightly on the handlebars. As the vehicle catapults over the sharp cornice into the vast emptiness, its rear track violently ejects a massive, billowing cloud of fine white powder that erupts behind it, creating a breathtaking spectacle. For a moment, the snowmobile hangs suspended in mid-air, creating a dynamic silhouette of the endless mountain range, evoking a powerful sense of adrenaline and a breathtaking sense of freedom.", + "seed": 33034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_8/index_5466_seed_33034_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08022204404696823, + "vlm_logprob": -0.08022204404696823, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.08022204404696823, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2926121652126312 + ], + "opposite_logprob_tokens": [ + -1.4176121950149536 + ], + "logprob_reward": -0.2926121652126312, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2926121652126312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0836290717124939, + 0.2216409593820572, + 0.09465701133012772, + 0.08814801275730133, + 0.16723422706127167, + 0.1501888632774353, + 0.035624999552965164, + 0.06201159954071045, + 0.062180593609809875, + 0.03468463197350502 + ], + "credit_map": [ + 0.013341471552848816, + 0.014658213593065739, + 0.014192070811986923, + 0.01170800719410181, + 0.007828759029507637, + 0.011141505092382431, + 0.00543218944221735, + 0.005326859187334776, + 0.00950370542705059, + 0.07481534779071808, + 0.009928819723427296, + 0.01348374504595995, + 0.0173215139657259, + 0.07195072621107101, + 0.012225618585944176, + 0.012411508709192276, + 0.008929000236093998, + 0.015238992869853973, + 0.016344904899597168, + 0.009543228894472122, + 0.012920710258185863, + 0.009955969639122486, + 0.014025948010385036, + 0.0076982625760138035, + 0.008787634782493114, + 0.006661528721451759, + 0.011034494265913963, + 0.01600673981010914, + 0.008022602647542953, + 0.00948214903473854, + 0.015964554622769356, + 0.0121883200481534, + 0.015829529613256454, + 0.013244344852864742, + 0.025538397952914238, + 0.03702476620674133, + 0.02056325413286686, + 0.016375703737139702, + 0.019451145082712173, + 0.01920711062848568, + 0.012474916875362396, + 0.012524822726845741, + 0.024497775360941887, + 0.016510266810655594, + 0.013916541822254658, + 0.016882486641407013, + 0.03297838196158409, + 0.02040369063615799, + 0.002624306594952941, + 0.004197017755359411, + 0.002586079528555274, + 0.00415049958974123, + 0.003900475800037384, + 0.0032776896841824055, + 0.00891120359301567, + 0.005977730732411146, + 0.005029768217355013, + 0.00993939395993948, + 0.0051852320320904255, + 0.0064111193642020226, + 0.012580608017742634, + 0.0064990161918103695, + 0.009690236300230026, + 0.006676229182630777, + 0.005914553068578243, + 0.0097411023452878, + 0.00484906230121851, + 0.006716094445437193, + 0.012142480351030827, + 0.006078937090933323, + 0.010171896778047085, + 0.0065664672292768955, + 0.002592257224023342, + 0.004461545497179031, + 0.0036982432939112186, + 0.004177866969257593, + 0.0049269553273916245, + 0.006168360821902752, + 0.004962915554642677, + 0.0036964905448257923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.2926121652126312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00364841241389513 + ], + "opposite_logprob_tokens": [ + -5.628648281097412 + ], + "logprob_reward": -0.00364841241389513, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00364841241389513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06359491497278214, + 0.41119900345802307, + 0.07186270505189896, + 0.03319777175784111, + 0.15006893873214722, + 0.06633022427558899, + 0.07354513555765152, + 0.07960726320743561, + 0.01959528774023056, + 0.030998723581433296 + ], + "credit_map": [ + 0.011851920746266842, + 0.011717239394783974, + 0.008862940594553947, + 0.007446408271789551, + 0.0072089070454239845, + 0.007745508570224047, + 0.004546387121081352, + 0.004215606953948736, + 0.020932307466864586, + 0.17184816300868988, + 0.007500645704567432, + 0.014280869625508785, + 0.022899847477674484, + 0.1556854546070099, + 0.007384852506220341, + 0.010666870512068272, + 0.007714165840297937, + 0.008607800118625164, + 0.016542838886380196, + 0.009903833270072937, + 0.007565631065517664, + 0.006632593460381031, + 0.0067488690838217735, + 0.008146975189447403, + 0.00306926341727376, + 0.004288328345865011, + 0.003276781179010868, + 0.004659492988139391, + 0.004078653175383806, + 0.005495523568242788, + 0.004494425840675831, + 0.003835307667031884, + 0.013745499774813652, + 0.017703544348478317, + 0.014249828644096851, + 0.025226058438420296, + 0.01464823167771101, + 0.017151180654764175, + 0.02764572575688362, + 0.019698888063430786, + 0.0057382709346711636, + 0.007035389542579651, + 0.01184867974370718, + 0.008583573624491692, + 0.00462615629658103, + 0.006994255352765322, + 0.012457400560379028, + 0.009046505205333233, + 0.006617357023060322, + 0.015479586087167263, + 0.004911675583571196, + 0.007089801598340273, + 0.01082729920744896, + 0.014264285564422607, + 0.007733286824077368, + 0.006621848791837692, + 0.007055453024804592, + 0.008801002986729145, + 0.008770962245762348, + 0.01234411634504795, + 0.0069641307927668095, + 0.014833301305770874, + 0.010535486042499542, + 0.01030281838029623, + 0.0016669840551912785, + 0.0024815667420625687, + 0.0028202037792652845, + 0.0022024649661034346, + 0.002715574111789465, + 0.002331525320187211, + 0.0033441646955907345, + 0.0020328061655163765, + 0.0033018470276147127, + 0.004463126417249441, + 0.004377090372145176, + 0.002956005046144128, + 0.004342763219028711, + 0.0045786830596625805, + 0.004220737610012293, + 0.0027584738563746214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.00364841241389513 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790226012468338 + ], + "opposite_logprob_tokens": [ + -2.5790226459503174 + ], + "logprob_reward": -0.0790226012468338, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790226012468338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027318589389324188, + 0.8019888997077942, + 0.047664742916822433, + 0.04469544067978859, + 0.012251368723809719, + 0.017439018934965134, + 0.019453097134828568, + 0.005630097817629576, + 0.01157469768077135, + 0.011984135024249554 + ], + "credit_map": [ + 0.003108425298705697, + 0.0045942203141748905, + 0.003093010513111949, + 0.003100045956671238, + 0.0034352883230894804, + 0.003708264324814081, + 0.0033878295216709375, + 0.00289150164462626, + 0.04123459756374359, + 0.338845431804657, + 0.014050275087356567, + 0.026461003348231316, + 0.04290727525949478, + 0.3058342933654785, + 0.01368306390941143, + 0.018972894176840782, + 0.005486777052283287, + 0.010160669684410095, + 0.008645813912153244, + 0.005430047400295734, + 0.004301294684410095, + 0.006332250777631998, + 0.003218963975086808, + 0.004088918212801218, + 0.0031346131581813097, + 0.006696072872728109, + 0.005063146352767944, + 0.0063311937265098095, + 0.005361196584999561, + 0.005378290079534054, + 0.006786269601434469, + 0.005944657139480114, + 0.0012442045845091343, + 0.0016290725907310843, + 0.0015834555961191654, + 0.0017569523770362139, + 0.0012108475202694535, + 0.001497106160968542, + 0.002028974937275052, + 0.0013007554225623608, + 0.0013799937441945076, + 0.002078417921438813, + 0.002326438669115305, + 0.0021220191847532988, + 0.0012761278776451945, + 0.0023834926541894674, + 0.0034800933208316565, + 0.002392434049397707, + 0.001695109298452735, + 0.003266021143645048, + 0.001101226662285626, + 0.001766785979270935, + 0.003416728461161256, + 0.002931412309408188, + 0.003096926724538207, + 0.00217888574115932, + 0.0004874127625953406, + 0.0005842210375703871, + 0.0004321879241615534, + 0.0006748383166268468, + 0.0007000341429375112, + 0.000598020211327821, + 0.0013021198101341724, + 0.0008512631175108254, + 0.0009943435434252024, + 0.0013564859982579947, + 0.0009781519183889031, + 0.001452013966627419, + 0.001801128382794559, + 0.0013931875582784414, + 0.0021922262385487556, + 0.0014071583282202482, + 0.0011481578694656491, + 0.0017508928431198, + 0.001017685397528112, + 0.0011250016978010535, + 0.0017448817379772663, + 0.001709655043669045, + 0.0021425134036689997, + 0.0013453460996970534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.0790226012468338 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196435645222664 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196435645222664, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196435645222664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042266689240932465, + 0.5265074968338013, + 0.09698501229286194, + 0.03242351487278938, + 0.05026058852672577, + 0.039182472974061966, + 0.0553397461771965, + 0.037529874593019485, + 0.08335476368665695, + 0.036149829626083374 + ], + "credit_map": [ + 0.004824352916330099, + 0.0043104165233671665, + 0.007756995502859354, + 0.005944595206528902, + 0.0038816388696432114, + 0.005876871757209301, + 0.004756187088787556, + 0.004915632773190737, + 0.03017103485763073, + 0.22481350600719452, + 0.009354020468890667, + 0.019436363130807877, + 0.02474602684378624, + 0.19829566776752472, + 0.008919961750507355, + 0.010770972818136215, + 0.009142395108938217, + 0.010907137766480446, + 0.012848217971622944, + 0.02074437215924263, + 0.008867403492331505, + 0.008330782875418663, + 0.008967341855168343, + 0.017177363857626915, + 0.0026904877740889788, + 0.004524783231317997, + 0.004601025953888893, + 0.004704689607024193, + 0.003589354222640395, + 0.003953937906771898, + 0.0048478953540325165, + 0.0035113380290567875, + 0.003372614039108157, + 0.008733319118618965, + 0.004313901532441378, + 0.0073656439781188965, + 0.005040748976171017, + 0.006883204914629459, + 0.008147377520799637, + 0.006403779610991478, + 0.003152377437800169, + 0.0033093304373323917, + 0.005454579368233681, + 0.006512449588626623, + 0.003342940704897046, + 0.004294256214052439, + 0.007639130111783743, + 0.005477409344166517, + 0.004204591736197472, + 0.00658299820497632, + 0.0033709080889821053, + 0.006380555219948292, + 0.007453653495758772, + 0.005088615696877241, + 0.013508306816220284, + 0.008750115521252155, + 0.003542088670656085, + 0.004975840914994478, + 0.0035602678544819355, + 0.004156833980232477, + 0.005107907112687826, + 0.005504639819264412, + 0.006153714377433062, + 0.004528581164777279, + 0.00723387161269784, + 0.009185806848108768, + 0.009684029966592789, + 0.011233103461563587, + 0.01021234504878521, + 0.00978788547217846, + 0.015289529226720333, + 0.010728199034929276, + 0.003451296128332615, + 0.004980390891432762, + 0.0031174542382359505, + 0.0035525227431207895, + 0.004990615416318178, + 0.005533219780772924, + 0.006579962093383074, + 0.003944369498640299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.018196435645222664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630605716258287 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630605716258287, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630605716258287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03588965907692909, + 0.46575847268104553, + 0.052505046129226685, + 0.05088607966899872, + 0.02822323888540268, + 0.022365691140294075, + 0.06285376846790314, + 0.06693696230649948, + 0.1060812696814537, + 0.10849969834089279 + ], + "credit_map": [ + 0.0037308731116354465, + 0.003707811702042818, + 0.0053616915829479694, + 0.004980115685611963, + 0.004433704074472189, + 0.004670883063226938, + 0.0047455099411308765, + 0.004259076435118914, + 0.02479281835258007, + 0.19888699054718018, + 0.007921052165329456, + 0.015961699187755585, + 0.023959515616297722, + 0.17732922732830048, + 0.007316504139453173, + 0.009590717032551765, + 0.004556827247142792, + 0.005186726804822683, + 0.009812834672629833, + 0.00852454174309969, + 0.006412542890757322, + 0.004991766531020403, + 0.006709663197398186, + 0.006310149561613798, + 0.0039946590550243855, + 0.00442842161282897, + 0.005874786525964737, + 0.0076437825337052345, + 0.005387989804148674, + 0.004621949512511492, + 0.011693630367517471, + 0.007240865379571915, + 0.0025868734810501337, + 0.002686522202566266, + 0.0031830950174480677, + 0.00503367418423295, + 0.0026244008913636208, + 0.00250536878593266, + 0.005714861676096916, + 0.003888444509357214, + 0.0022276027593761683, + 0.0024990381207317114, + 0.003175985300913453, + 0.003167671849951148, + 0.001744022243656218, + 0.003665702184662223, + 0.003280896693468094, + 0.0026047732681035995, + 0.004968992900103331, + 0.009653977118432522, + 0.0034553660079836845, + 0.005927427671849728, + 0.010097463615238667, + 0.008873763494193554, + 0.011884103529155254, + 0.007992682047188282, + 0.00552937388420105, + 0.010840031318366528, + 0.00437907874584198, + 0.0068545215763151646, + 0.013701049610972404, + 0.006270143203437328, + 0.011771152727305889, + 0.007591616827994585, + 0.008342554792761803, + 0.01059287041425705, + 0.01017404068261385, + 0.014791407622396946, + 0.015027587302029133, + 0.012758246622979641, + 0.020634843036532402, + 0.013759732246398926, + 0.010112294927239418, + 0.01775641366839409, + 0.008334281854331493, + 0.014731098897755146, + 0.016740141436457634, + 0.011053980328142643, + 0.017347967252135277, + 0.012423532083630562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.007630605716258287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The drone camera quickly pulls back to an extreme wide shot, capturing a breathtaking panorama of a powerful snowmobile launching from a jagged, snow-capped peak. The scene is set against a vast, cloudless azure sky. The intense, bright, high-altitude sun casts sharp, dramatic shadows across the pristine, untouched powder, causing millions of ice crystals to glitter. The snowmobile itself is a high-performance vehicle, equipped with a glossy electric blue-black chassis, capable of explosive acceleration to the edge. Its rider, clad in a full-body black snowsuit and a matching helmet with a tinted visor, pulls forward tightly on the handlebars. As the vehicle catapults over the sharp cornice into the vast emptiness, its rear track violently ejects a massive, billowing cloud of fine white powder that erupts behind it, creating a breathtaking spectacle. For a moment, the snowmobile hangs suspended in mid-air, creating a dynamic silhouette of the endless mountain range, evoking a powerful sense of adrenaline and a breathtaking sense of freedom.", + "seed": 33035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_8/index_5466_seed_33035_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2994329683482647, + "vlm_logprob": -0.2994329683482647, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.2994329683482647, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1480391025543213 + ], + "opposite_logprob_tokens": [ + -0.3980390727519989 + ], + "logprob_reward": -1.1480391025543213, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1480391025543213, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12664245069026947, + 0.0784192904829979, + 0.13342612981796265, + 0.054613593965768814, + 0.08332440257072449, + 0.08868876099586487, + 0.16007839143276215, + 0.11542879790067673, + 0.10115150362253189, + 0.05822661519050598 + ], + "credit_map": [ + 0.01461621094495058, + 0.015484516508877277, + 0.01778561994433403, + 0.01561559084802866, + 0.01127894688397646, + 0.01763862743973732, + 0.01562669314444065, + 0.01859625242650509, + 0.012256021611392498, + 0.009175533428788185, + 0.011556694284081459, + 0.008920220658183098, + 0.009056320413947105, + 0.009787367656826973, + 0.008523436263203621, + 0.009143699891865253, + 0.013998998329043388, + 0.02054438367486, + 0.016311461105942726, + 0.019924303516745567, + 0.014805188402533531, + 0.01690763421356678, + 0.014745143242180347, + 0.01618901640176773, + 0.004866399336606264, + 0.008514679968357086, + 0.006638987921178341, + 0.008127426728606224, + 0.005502948071807623, + 0.007597329095005989, + 0.006233178079128265, + 0.007132646162062883, + 0.00913554709404707, + 0.009322580881416798, + 0.008928188122808933, + 0.011066000908613205, + 0.0070810746401548386, + 0.015892820432782173, + 0.007602060679346323, + 0.014296126551926136, + 0.011833126656711102, + 0.009960399009287357, + 0.011302702128887177, + 0.012034490704536438, + 0.008510230109095573, + 0.010854310356080532, + 0.010043480433523655, + 0.014150021597743034, + 0.01427442952990532, + 0.01597459241747856, + 0.03534282371401787, + 0.028991607949137688, + 0.012304090894758701, + 0.011358041316270828, + 0.01908843033015728, + 0.02274438552558422, + 0.016172118484973907, + 0.01592032052576542, + 0.00961228646337986, + 0.014134803786873817, + 0.019937261939048767, + 0.010579206980764866, + 0.016880936920642853, + 0.012191860936582088, + 0.014818268828094006, + 0.020704617723822594, + 0.009128831326961517, + 0.00751880370080471, + 0.017490357160568237, + 0.008396509103477001, + 0.01311721745878458, + 0.009976896457374096, + 0.006403310224413872, + 0.008047854527831078, + 0.005186442285776138, + 0.004480029456317425, + 0.006262746173888445, + 0.007064017467200756, + 0.013027382083237171, + 0.007754831574857235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -1.1480391025543213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012587739154696465 + ], + "opposite_logprob_tokens": [ + -4.387587547302246 + ], + "logprob_reward": -0.012587739154696465, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012587739154696465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0748642310500145, + 0.05398545414209366, + 0.04720788076519966, + 0.027511101216077805, + 0.028877992182970047, + 0.4208863377571106, + 0.21980713307857513, + 0.022725600749254227, + 0.045421816408634186, + 0.058712512254714966 + ], + "credit_map": [ + 0.010658349841833115, + 0.010170980356633663, + 0.011226858012378216, + 0.00903068296611309, + 0.005300196819007397, + 0.010038827545940876, + 0.007345818914473057, + 0.011092511005699635, + 0.011219817213714123, + 0.009377343580126762, + 0.008230580016970634, + 0.0035446977708488703, + 0.004308572504669428, + 0.00758080231025815, + 0.003919062204658985, + 0.005804573185741901, + 0.006309939082711935, + 0.007567721884697676, + 0.00611446239054203, + 0.0055687157437205315, + 0.003061505500227213, + 0.00887995958328247, + 0.004501670598983765, + 0.0052038999274373055, + 0.0021244280505925417, + 0.0028717140667140484, + 0.0046887267380952835, + 0.00427602231502533, + 0.00282963365316391, + 0.0050027742981910706, + 0.003265207866206765, + 0.002452591434121132, + 0.005180325824767351, + 0.00396774522960186, + 0.003644360462203622, + 0.005673022475093603, + 0.001816792762838304, + 0.0032679077703505754, + 0.0022138438653200865, + 0.003113994374871254, + 0.04783432185649872, + 0.07877647876739502, + 0.11597798019647598, + 0.07218464463949203, + 0.02447165548801422, + 0.023699287325143814, + 0.026590496301651, + 0.03135140985250473, + 0.05579197779297829, + 0.06829118728637695, + 0.023955289274454117, + 0.030853187665343285, + 0.008886389434337616, + 0.009344426915049553, + 0.007799023296684027, + 0.014885633252561092, + 0.002484773052856326, + 0.00242014997638762, + 0.0032050481531769037, + 0.0031485948711633682, + 0.00314485770650208, + 0.0021503104362636805, + 0.0027283516246825457, + 0.0034435135312378407, + 0.00681992806494236, + 0.008697560988366604, + 0.004494363907724619, + 0.005341109819710255, + 0.006092481780797243, + 0.0029850671999156475, + 0.0055546932853758335, + 0.0054366025142371655, + 0.006704952102154493, + 0.007395448163151741, + 0.007138609886169434, + 0.005762991029769182, + 0.005682863295078278, + 0.00874342955648899, + 0.006191173102706671, + 0.011093039996922016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.012587739154696465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2522491216659546 + ], + "opposite_logprob_tokens": [ + -1.5022491216659546 + ], + "logprob_reward": -0.2522491216659546, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2522491216659546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10499708354473114, + 0.10603734850883484, + 0.12437734752893448, + 0.05939008668065071, + 0.10532800108194351, + 0.1311211735010147, + 0.2122626006603241, + 0.06771428883075714, + 0.03366664797067642, + 0.055105481296777725 + ], + "credit_map": [ + 0.013442326337099075, + 0.015286006033420563, + 0.01289413496851921, + 0.013492930680513382, + 0.009983117692172527, + 0.013508296571671963, + 0.010821547359228134, + 0.015568719245493412, + 0.005991505924612284, + 0.00596996396780014, + 0.022643500939011574, + 0.009466802701354027, + 0.012888450175523758, + 0.01636495441198349, + 0.025073466822504997, + 0.0076387058943510056, + 0.016363738104701042, + 0.018154658377170563, + 0.015803707763552666, + 0.012703212909400463, + 0.014983517117798328, + 0.023238655179739, + 0.013001896440982819, + 0.010127966292202473, + 0.004557938314974308, + 0.00621890788897872, + 0.010333898477256298, + 0.007851187139749527, + 0.004899478051811457, + 0.010448917746543884, + 0.006743474397808313, + 0.008336282335221767, + 0.009149515070021152, + 0.010408889502286911, + 0.012251039035618305, + 0.019570710137486458, + 0.013873469084501266, + 0.016583578661084175, + 0.011349571868777275, + 0.012141217477619648, + 0.010232522152364254, + 0.019919296726584435, + 0.032324690371751785, + 0.021126162260770798, + 0.012691711075603962, + 0.01036548800766468, + 0.012513157911598682, + 0.011948147788643837, + 0.015063833445310593, + 0.026424750685691833, + 0.027469610795378685, + 0.05259018763899803, + 0.01587582379579544, + 0.026189498603343964, + 0.025758616626262665, + 0.02289028838276863, + 0.005623691249638796, + 0.006143651437014341, + 0.008247937075793743, + 0.009963798336684704, + 0.009333045221865177, + 0.006763083394616842, + 0.012097695842385292, + 0.009541384875774384, + 0.004455224610865116, + 0.005332617554813623, + 0.004807509481906891, + 0.00462046405300498, + 0.003149256110191345, + 0.0026282682083547115, + 0.003974518273025751, + 0.0046987892128527164, + 0.0075908382423222065, + 0.009849096648395061, + 0.0063483500853180885, + 0.005603297613561153, + 0.005430395249277353, + 0.004402401391416788, + 0.00898763071745634, + 0.006893472280353308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.2522491216659546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007994502782822 + ], + "opposite_logprob_tokens": [ + -2.695080041885376 + ], + "logprob_reward": -0.07007994502782822, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007994502782822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08723683655261993, + 0.05463268607854843, + 0.04841214790940285, + 0.09769503027200699, + 0.1919427514076233, + 0.07973191887140274, + 0.07071753591299057, + 0.03441070392727852, + 0.0416923388838768, + 0.2935280501842499 + ], + "credit_map": [ + 0.01031939685344696, + 0.014442069455981255, + 0.01252214889973402, + 0.010008460842072964, + 0.006356782745569944, + 0.013358530588448048, + 0.009244468063116074, + 0.010984990745782852, + 0.005386780481785536, + 0.00559125654399395, + 0.009395699948072433, + 0.007550251204520464, + 0.00639888783916831, + 0.006532193627208471, + 0.009969537146389484, + 0.003808081615716219, + 0.005330556072294712, + 0.005838362965732813, + 0.00754665769636631, + 0.006229414604604244, + 0.007569376844912767, + 0.005713379010558128, + 0.00571758346632123, + 0.0044668205082416534, + 0.007540245074778795, + 0.013318938203155994, + 0.014214527793228626, + 0.014148717746138573, + 0.007861725986003876, + 0.016664888709783554, + 0.009139468893408775, + 0.014806528575718403, + 0.01837180182337761, + 0.022288337349891663, + 0.025790007784962654, + 0.023821042850613594, + 0.016744103282690048, + 0.042884666472673416, + 0.015582771971821785, + 0.02646002732217312, + 0.007219481747597456, + 0.011530984193086624, + 0.018816322088241577, + 0.015412824228405952, + 0.004961127415299416, + 0.008380871266126633, + 0.005891263950616121, + 0.007519054226577282, + 0.005824113730341196, + 0.00796300545334816, + 0.008607515133917332, + 0.00973608996719122, + 0.005957954563200474, + 0.014688625931739807, + 0.007054655812680721, + 0.010885583236813545, + 0.004616617225110531, + 0.004491467960178852, + 0.004330748692154884, + 0.004840652458369732, + 0.004171930253505707, + 0.00342745054513216, + 0.00445495406165719, + 0.004076888784766197, + 0.004340798128396273, + 0.005760216154158115, + 0.0039910608902573586, + 0.0046363468281924725, + 0.0061426106840372086, + 0.0035720011219382286, + 0.007742648012936115, + 0.005506661720573902, + 0.03451867774128914, + 0.04683293402194977, + 0.027204399928450584, + 0.025383086875081062, + 0.026882948353886604, + 0.036411840468645096, + 0.057948414236307144, + 0.03834579512476921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.07007994502782822 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera pull back quickly to an extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set against a vast, cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snowmobile a high-performance vehicle with a glossy electric blue-black chassis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snowmobile launch from a jagged, snow-capped peak and eject a massive cloud of white powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the snowmobile hang suspended in mid-air after launching over a sharp cornice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208933338522911 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208933338522911, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208933338522911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13859347999095917, + 0.05537122115492821, + 0.15756618976593018, + 0.11374368518590927, + 0.06347259134054184, + 0.12961837649345398, + 0.11277831345796585, + 0.05811364948749542, + 0.05359242856502533, + 0.11715003848075867 + ], + "credit_map": [ + 0.0142868896946311, + 0.021621162071824074, + 0.01752493530511856, + 0.020223204046487808, + 0.009927617385983467, + 0.020731961354613304, + 0.013094078749418259, + 0.02118363417685032, + 0.007652195170521736, + 0.008008414879441261, + 0.007676847744733095, + 0.005053307395428419, + 0.007845066487789154, + 0.0067668212577700615, + 0.007275481708347797, + 0.0050930785946547985, + 0.017710169777274132, + 0.02853299491107464, + 0.020644070580601692, + 0.015294828452169895, + 0.01499711163341999, + 0.030575938522815704, + 0.014386872760951519, + 0.015424193814396858, + 0.0117849325761199, + 0.015600377693772316, + 0.0164241511374712, + 0.013824108056724072, + 0.007429138291627169, + 0.021477805450558662, + 0.009556431323289871, + 0.017646731808781624, + 0.007780367508530617, + 0.007101942785084248, + 0.007857885211706161, + 0.008630956523120403, + 0.00546900974586606, + 0.00858275219798088, + 0.006773359142243862, + 0.011276304721832275, + 0.01208228524774313, + 0.02526894398033619, + 0.038045696914196014, + 0.021764222532510757, + 0.006899077445268631, + 0.006117870099842548, + 0.011051143519580364, + 0.008389121852815151, + 0.011907588690519333, + 0.018094249069690704, + 0.01782357133924961, + 0.025077784433960915, + 0.010328599251806736, + 0.00563874701038003, + 0.011369667947292328, + 0.012538093142211437, + 0.01308264210820198, + 0.01180100254714489, + 0.006934634875506163, + 0.006404507905244827, + 0.0053800251334905624, + 0.0032959310337901115, + 0.005518803372979164, + 0.005696095060557127, + 0.0055980384349823, + 0.008398162201046944, + 0.005446107126772404, + 0.007230233401060104, + 0.006830820348113775, + 0.004038471728563309, + 0.008262582123279572, + 0.0077880085445940495, + 0.01724395900964737, + 0.023871876299381256, + 0.009824661538004875, + 0.008745945058763027, + 0.010372361168265343, + 0.012672362849116325, + 0.019099460914731026, + 0.015319399535655975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.014208933338522911 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_9/index_tensor([8235, 8235])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_9/index_tensor([8235, 8235])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..746cabdc7a96ea2f9461415195fee3d86b1388a7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_9/index_tensor([8235, 8235])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, low-angle, static, full-body shot captures a god-warrior's unwavering stand against catastrophic events, created in the epic, high-stakes style of an action film scene. He is clad head to toe in gleaming celestial armor, its polished, iridescent white gold surface intricately etched with cosmic patterns, reflecting the chaotic ambient light. His feet planted broadly on a jagged, damp rock shelf, his posture resolute, his hands poised, firmly gripping a mighty three-pointed spear. In the center of his forehead, a third eye blazes with intense, pulsating azure power, illuminating his grim resolve. Behind him, a massive river flood rages, a roaring, tumbling wall of mud, water, and debris violently tearing through a deep gorge, stirring up a thick spray. The atmosphere is tense and dramatic, dominated by a cool, desaturated palette of dark grays and browns, contrasting sharply with the brilliant radiance of the armor and the dazzling blue glow of his power. The camera remains locked, emphasizing his steadfastness as he stands as the sole point of light and order against the overwhelming and destructive flood.", + "seed": 13311, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_9/index_8235_seed_13311_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05472815111279487, + "vlm_logprob": -0.05472815111279487, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05472815111279487, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004712190479040146 + ], + "opposite_logprob_tokens": [ + -6.504712104797363 + ], + "logprob_reward": -0.004712190479040146, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.004712190479040146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26886695623397827, + 0.07701410353183746, + 0.07547826319932938, + 0.0704793930053711, + 0.04470699280500412, + 0.13137371838092804, + 0.030252045020461082, + 0.06885559856891632, + 0.1843959093093872, + 0.048577018082141876 + ], + "credit_map": [ + 0.03113717958331108, + 0.02914193458855152, + 0.04552555829286575, + 0.0529998317360878, + 0.025098320096731186, + 0.028625277802348137, + 0.02699577808380127, + 0.029343094676733017, + 0.00722986925393343, + 0.008692190051078796, + 0.012412209995090961, + 0.01586863584816456, + 0.006359405349940062, + 0.009542697109282017, + 0.010023829527199268, + 0.006885258946567774, + 0.007951711304485798, + 0.008769027888774872, + 0.0239250548183918, + 0.0060620843432843685, + 0.00468507269397378, + 0.01248735561966896, + 0.006190267391502857, + 0.0054076798260211945, + 0.014740929938852787, + 0.010279699228703976, + 0.009043379686772823, + 0.011462101712822914, + 0.004369992762804031, + 0.00671745277941227, + 0.005990331526845694, + 0.007875503972172737, + 0.009056932292878628, + 0.008859727531671524, + 0.005190890282392502, + 0.005260748788714409, + 0.0032799229957163334, + 0.005843975581228733, + 0.0031504707876592875, + 0.004064321983605623, + 0.018643107265233994, + 0.017546728253364563, + 0.0170845165848732, + 0.028490355238318443, + 0.008874570950865746, + 0.01658306084573269, + 0.008440898731350899, + 0.0157104954123497, + 0.003612063592299819, + 0.004691213835030794, + 0.004426578991115093, + 0.004207602236419916, + 0.002933749696239829, + 0.003324999939650297, + 0.004524004179984331, + 0.0025318320840597153, + 0.007119425572454929, + 0.00793163850903511, + 0.013460169546306133, + 0.0058521246537566185, + 0.007147991098463535, + 0.009275173768401146, + 0.010266107507050037, + 0.007802970707416534, + 0.028099531307816505, + 0.026911692693829536, + 0.027087122201919556, + 0.024823175743222237, + 0.023334085941314697, + 0.02239127643406391, + 0.016409842297434807, + 0.015339184552431107, + 0.006543717347085476, + 0.006520599592477083, + 0.009450851939618587, + 0.005040246527642012, + 0.004595160484313965, + 0.007440107874572277, + 0.0049793655052781105, + 0.00400696974247694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.004712190479040146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549461729824543 + ], + "opposite_logprob_tokens": [ + -2.929946184158325 + ], + "logprob_reward": -0.0549461729824543, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549461729824543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39847853779792786, + 0.13793182373046875, + 0.10244610160589218, + 0.07426884025335312, + 0.06841545552015305, + 0.02780388854444027, + 0.028480494394898415, + 0.06444892287254333, + 0.0470237210392952, + 0.05070215091109276 + ], + "credit_map": [ + 0.04184352606534958, + 0.0414968878030777, + 0.07044150680303574, + 0.059532564133405685, + 0.03902332857251167, + 0.05336068570613861, + 0.052030351012945175, + 0.040749691426754, + 0.012317854911088943, + 0.015809297561645508, + 0.018231401219964027, + 0.028563283383846283, + 0.01328404899686575, + 0.016795353963971138, + 0.019171027466654778, + 0.013759558089077473, + 0.010505552403628826, + 0.010343991219997406, + 0.0316869281232357, + 0.010806641541421413, + 0.007541830185800791, + 0.016528546810150146, + 0.007117013446986675, + 0.007915602065622807, + 0.016567731276154518, + 0.01550996396690607, + 0.00826987624168396, + 0.007611056789755821, + 0.005086667835712433, + 0.0086112841963768, + 0.006316624581813812, + 0.006295630242675543, + 0.009180241264402866, + 0.012632234953343868, + 0.00710729043930769, + 0.009496876038610935, + 0.007374654058367014, + 0.008774194866418839, + 0.00488350261002779, + 0.008966457098722458, + 0.0023761168122291565, + 0.0026381919160485268, + 0.0046801092103123665, + 0.0037720331456512213, + 0.0036905177403241396, + 0.003988522104918957, + 0.0036327203270047903, + 0.0030256793834269047, + 0.003393623512238264, + 0.0037270919419825077, + 0.0042993249371647835, + 0.0036730424035340548, + 0.0036124244797974825, + 0.002978348173201084, + 0.0026427595876157284, + 0.004153879825025797, + 0.009695482440292835, + 0.011360013857483864, + 0.012427065521478653, + 0.006300542503595352, + 0.004245452117174864, + 0.00884830392897129, + 0.005452023819088936, + 0.006120034959167242, + 0.007263946812599897, + 0.006627616472542286, + 0.007928612641990185, + 0.0058272406458854675, + 0.0040070232935249805, + 0.006229400634765625, + 0.004489616025239229, + 0.00465026730671525, + 0.00861847773194313, + 0.010377813130617142, + 0.006703529506921768, + 0.005197395570576191, + 0.003731878474354744, + 0.00675547868013382, + 0.004046767950057983, + 0.00527080986648798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0549461729824543 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493917688727379 + ], + "opposite_logprob_tokens": [ + -2.9299392700195312 + ], + "logprob_reward": -0.05493917688727379, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493917688727379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.268103688955307, + 0.13958999514579773, + 0.1606082171201706, + 0.06241898611187935, + 0.06067106872797012, + 0.10307323187589645, + 0.023111121729016304, + 0.060001567006111145, + 0.06887505203485489, + 0.053547076880931854 + ], + "credit_map": [ + 0.037463460117578506, + 0.03261040896177292, + 0.055579010397195816, + 0.03642309829592705, + 0.020787514746189117, + 0.036157798022031784, + 0.022338440641760826, + 0.026743950322270393, + 0.016741232946515083, + 0.01713140867650509, + 0.024243533611297607, + 0.0148787135258317, + 0.015310395509004593, + 0.017576584592461586, + 0.015787148848176003, + 0.017920974642038345, + 0.016018429771065712, + 0.018607499077916145, + 0.04038960486650467, + 0.01686386950314045, + 0.012689323164522648, + 0.023389004170894623, + 0.01875261589884758, + 0.013897870667278767, + 0.010826432146131992, + 0.008561075665056705, + 0.007166814524680376, + 0.009792491793632507, + 0.00508614769205451, + 0.00643700547516346, + 0.00741798710078001, + 0.007131032645702362, + 0.006781273987144232, + 0.008109263144433498, + 0.006426856853067875, + 0.011164876632392406, + 0.0053739226423203945, + 0.00733213871717453, + 0.005760703235864639, + 0.009722036309540272, + 0.015957869589328766, + 0.015260965563356876, + 0.010792907327413559, + 0.021045545116066933, + 0.010326574556529522, + 0.012555547058582306, + 0.008050736971199512, + 0.009083091281354427, + 0.003568526590242982, + 0.004578154068440199, + 0.0026288575027137995, + 0.0018635503947734833, + 0.002635046374052763, + 0.0027243203949183226, + 0.0020022084936499596, + 0.0031104579102247953, + 0.008488403633236885, + 0.009266159497201443, + 0.012653831392526627, + 0.00668262830004096, + 0.0032341580372303724, + 0.008015492931008339, + 0.005058228503912687, + 0.006602661684155464, + 0.011833494529128075, + 0.010333613492548466, + 0.011803745292127132, + 0.0074848392978310585, + 0.0046360669657588005, + 0.01023679506033659, + 0.005471325013786554, + 0.007075164467096329, + 0.006997371558099985, + 0.009227939881384373, + 0.007634076289832592, + 0.005350705701857805, + 0.004617971368134022, + 0.007092674262821674, + 0.005472492426633835, + 0.007153840269893408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.05493917688727379 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904818445444107 + ], + "opposite_logprob_tokens": [ + -2.464048147201538 + ], + "logprob_reward": -0.08904818445444107, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904818445444107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2684915363788605, + 0.08806920051574707, + 0.10727711766958237, + 0.14844468235969543, + 0.08161713927984238, + 0.03966313228011131, + 0.042010243982076645, + 0.07259547710418701, + 0.0813930332660675, + 0.07043853402137756 + ], + "credit_map": [ + 0.025824490934610367, + 0.027785304933786392, + 0.040873266756534576, + 0.042858924716711044, + 0.02766767516732216, + 0.03257223963737488, + 0.03615916147828102, + 0.03475044667720795, + 0.006939893122762442, + 0.010559910908341408, + 0.01372869685292244, + 0.017489811405539513, + 0.008451067842543125, + 0.0106246592476964, + 0.010718501172959805, + 0.00955665111541748, + 0.01199062168598175, + 0.010589193552732468, + 0.034347791224718094, + 0.01210696529597044, + 0.006592930294573307, + 0.017265593633055687, + 0.006657406222075224, + 0.007726609706878662, + 0.031624529510736465, + 0.031242944300174713, + 0.0172664076089859, + 0.019844355061650276, + 0.009558469988405704, + 0.017888229340314865, + 0.009759933687746525, + 0.011259796097874641, + 0.011097804643213749, + 0.01269923709332943, + 0.010956809855997562, + 0.013612487353384495, + 0.004250055178999901, + 0.010120480321347713, + 0.006176503375172615, + 0.012703748419880867, + 0.0033120447769761086, + 0.004473111592233181, + 0.0042259530164301395, + 0.007042238488793373, + 0.004498000256717205, + 0.005272134207189083, + 0.0037943972274661064, + 0.007045248989015818, + 0.006577267777174711, + 0.008042129687964916, + 0.004175129812210798, + 0.003908198326826096, + 0.004575361032038927, + 0.00471806013956666, + 0.0033561186864972115, + 0.0066579729318618774, + 0.009820598177611828, + 0.011979700066149235, + 0.015573540702462196, + 0.008120648562908173, + 0.0030734085012227297, + 0.009813937358558178, + 0.005958949215710163, + 0.008254692889750004, + 0.011188727803528309, + 0.013303986750543118, + 0.016676289960741997, + 0.008143407292664051, + 0.005970973987132311, + 0.01289631612598896, + 0.006043897941708565, + 0.007169423159211874, + 0.009397937916219234, + 0.012931399047374725, + 0.009044748730957508, + 0.0071349903009831905, + 0.007920143194496632, + 0.009430225007236004, + 0.00595156429335475, + 0.008627521805465221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.08904818445444107 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999503076076508 + ], + "opposite_logprob_tokens": [ + -2.6949949264526367 + ], + "logprob_reward": -0.06999503076076508, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999503076076508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10387421399354935, + 0.025287702679634094, + 0.04533735662698746, + 0.06916916370391846, + 0.14600782096385956, + 0.06564119458198547, + 0.05824742093682289, + 0.17922015488147736, + 0.2018030732870102, + 0.10541189461946487 + ], + "credit_map": [ + 0.01426614262163639, + 0.023620307445526123, + 0.01726408861577511, + 0.011845704168081284, + 0.009616775438189507, + 0.011486477218568325, + 0.007948863320052624, + 0.007825854234397411, + 0.003106427611783147, + 0.0032813982106745243, + 0.004946296103298664, + 0.0032505167182534933, + 0.002127228304743767, + 0.0032662220764905214, + 0.00286160409450531, + 0.0024480095598846674, + 0.005478978157043457, + 0.005284302402287722, + 0.011590461246669292, + 0.0044128866866230965, + 0.0024003360886126757, + 0.007246153894811869, + 0.0026884544640779495, + 0.006235782522708178, + 0.0071980273351073265, + 0.011102838441729546, + 0.007752956822514534, + 0.015675699338316917, + 0.0046874103136360645, + 0.007924596779048443, + 0.00723818875849247, + 0.007589450106024742, + 0.020766237750649452, + 0.02210928127169609, + 0.017384853214025497, + 0.02249717153608799, + 0.006487224716693163, + 0.019610870629549026, + 0.010363913141191006, + 0.02678826078772545, + 0.008903330191969872, + 0.010790578089654446, + 0.008520486764609814, + 0.013795781880617142, + 0.0034164516255259514, + 0.008612174540758133, + 0.0048789167776703835, + 0.00672348216176033, + 0.010720619931817055, + 0.01084815338253975, + 0.0070173791609704494, + 0.004665612243115902, + 0.004336884245276451, + 0.00688373064622283, + 0.0034285681322216988, + 0.010346470400691032, + 0.02717684581875801, + 0.03255973011255264, + 0.03799496591091156, + 0.01864575780928135, + 0.0070620640181005, + 0.02452809177339077, + 0.012655667960643768, + 0.018597040325403214, + 0.024229643866419792, + 0.028004011139273643, + 0.04436223581433296, + 0.0301552452147007, + 0.01265039388090372, + 0.03189200907945633, + 0.01274962443858385, + 0.017759915441274643, + 0.019814617931842804, + 0.019118640571832657, + 0.01744510419666767, + 0.009274995885789394, + 0.006028593983501196, + 0.015391355380415916, + 0.006957487668842077, + 0.011381099931895733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.06999503076076508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle, static, full-body shot captures a god-warrior's unwavering stand against catastrophic events, created in the epic, high-stakes style of an action film scene. He is clad head to toe in gleaming celestial armor, its polished, iridescent white gold surface intricately etched with cosmic patterns, reflecting the chaotic ambient light. His feet planted broadly on a jagged, damp rock shelf, his posture resolute, his hands poised, firmly gripping a mighty three-pointed spear. In the center of his forehead, a third eye blazes with intense, pulsating azure power, illuminating his grim resolve. Behind him, a massive river flood rages, a roaring, tumbling wall of mud, water, and debris violently tearing through a deep gorge, stirring up a thick spray. The atmosphere is tense and dramatic, dominated by a cool, desaturated palette of dark grays and browns, contrasting sharply with the brilliant radiance of the armor and the dazzling blue glow of his power. The camera remains locked, emphasizing his steadfastness as he stands as the sole point of light and order against the overwhelming and destructive flood.", + "seed": 13312, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_9/index_8235_seed_13312_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05203275950625539, + "vlm_logprob": -0.05203275950625539, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.05203275950625539, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005407231394201517 + ], + "opposite_logprob_tokens": [ + -5.755407333374023 + ], + "logprob_reward": -0.005407231394201517, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.005407231394201517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2118169665336609, + 0.05655572935938835, + 0.05033449828624725, + 0.08289852738380432, + 0.03394758328795433, + 0.07709458470344543, + 0.05751879885792732, + 0.18885590136051178, + 0.1755238026380539, + 0.06545364111661911 + ], + "credit_map": [ + 0.01838553138077259, + 0.02059910073876381, + 0.02806365117430687, + 0.02997530810534954, + 0.015820151194930077, + 0.02936566434800625, + 0.036482956260442734, + 0.03312460333108902, + 0.008589649572968483, + 0.004905213136225939, + 0.006739710457623005, + 0.006557193119078875, + 0.005456474609673023, + 0.009522327221930027, + 0.00732779735699296, + 0.007457363419234753, + 0.0055513228289783, + 0.005065515171736479, + 0.010671277530491352, + 0.004048788454383612, + 0.005046010948717594, + 0.006102926563471556, + 0.00675352243706584, + 0.00709513807669282, + 0.009557475335896015, + 0.006119483150541782, + 0.020067350938916206, + 0.00959050003439188, + 0.006314325612038374, + 0.014228956773877144, + 0.007849718444049358, + 0.009170716628432274, + 0.005237083416432142, + 0.0027368664741516113, + 0.006355696823447943, + 0.004077497869729996, + 0.0036454801447689533, + 0.0041930582374334335, + 0.004114973824471235, + 0.0035869237035512924, + 0.008982706815004349, + 0.008155674673616886, + 0.014339519664645195, + 0.0071661267429590225, + 0.006734379567205906, + 0.01246607955545187, + 0.008854628540575504, + 0.010395471006631851, + 0.009637049399316311, + 0.00561338709667325, + 0.008083268068730831, + 0.003761144122108817, + 0.0038251145742833614, + 0.011558979749679565, + 0.005062552634626627, + 0.009977306239306927, + 0.02479645051062107, + 0.008666693232953548, + 0.012146919965744019, + 0.009204994887113571, + 0.007594433147460222, + 0.046601731330156326, + 0.009167051874101162, + 0.07067762315273285, + 0.03972189128398895, + 0.015038082376122475, + 0.017186634242534637, + 0.021837836131453514, + 0.011165613308548927, + 0.04034164175391197, + 0.015536296181380749, + 0.014695816673338413, + 0.012719645164906979, + 0.0077037448063492775, + 0.007209997624158859, + 0.005353943444788456, + 0.003576310584321618, + 0.00802282802760601, + 0.006004120223224163, + 0.014863042160868645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.005407231394201517 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810303658246994 + ], + "opposite_logprob_tokens": [ + -3.2881031036376953 + ], + "logprob_reward": -0.03810303658246994, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810303658246994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20731723308563232, + 0.049805980175733566, + 0.03966410085558891, + 0.06006829813122749, + 0.032098446041345596, + 0.05704258009791374, + 0.11342789232730865, + 0.3261081576347351, + 0.06159954518079758, + 0.05286787077784538 + ], + "credit_map": [ + 0.029379401355981827, + 0.017955070361495018, + 0.026384053751826286, + 0.03206128254532814, + 0.016475196927785873, + 0.027970246970653534, + 0.03181695193052292, + 0.02527502365410328, + 0.005416041240096092, + 0.005466986447572708, + 0.006680361460894346, + 0.005068742670118809, + 0.0065341973677277565, + 0.007479500956833363, + 0.007659163326025009, + 0.005500987637788057, + 0.004305324517190456, + 0.003864116733893752, + 0.007914133369922638, + 0.004353392869234085, + 0.004136997740715742, + 0.005027238745242357, + 0.005072004161775112, + 0.004990892484784126, + 0.006477410905063152, + 0.004001947585493326, + 0.015306385233998299, + 0.007616280112415552, + 0.004804654978215694, + 0.009967484511435032, + 0.0057539003901183605, + 0.006140236277133226, + 0.002964144106954336, + 0.004135625436902046, + 0.0037494793068617582, + 0.0037283424753695726, + 0.004779196344316006, + 0.0046386877074837685, + 0.0059266905300319195, + 0.0021762780379503965, + 0.007452663034200668, + 0.00479815574362874, + 0.010866455733776093, + 0.007029885891824961, + 0.004495781846344471, + 0.009863143786787987, + 0.005810015369206667, + 0.006726480089128017, + 0.030750185251235962, + 0.008975398726761341, + 0.013429208658635616, + 0.005836001131683588, + 0.005803354550153017, + 0.02753848023712635, + 0.006637847516685724, + 0.014457416720688343, + 0.10847704857587814, + 0.02567271888256073, + 0.017160488292574883, + 0.014356324449181557, + 0.010936410166323185, + 0.02876555174589157, + 0.013099819421768188, + 0.10763978958129883, + 0.012000343762338161, + 0.008820297196507454, + 0.006276938132941723, + 0.0051088533364236355, + 0.004981463309377432, + 0.014888705685734749, + 0.004197924397885799, + 0.0053250170312821865, + 0.005096284206956625, + 0.0067795924842357635, + 0.005404000170528889, + 0.004439906682819128, + 0.0038720096927136183, + 0.013215320184826851, + 0.00527387298643589, + 0.008786882273852825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.03810303658246994 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903750032186508 + ], + "opposite_logprob_tokens": [ + -2.4640374183654785 + ], + "logprob_reward": -0.08903750032186508, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903750032186508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23253776133060455, + 0.11980052292346954, + 0.05030323192477226, + 0.06415262073278427, + 0.04983366280794144, + 0.06145830079913139, + 0.07074137032032013, + 0.1324247121810913, + 0.11902676522731781, + 0.09972111135721207 + ], + "credit_map": [ + 0.0266113243997097, + 0.020239025354385376, + 0.02872786670923233, + 0.03973013162612915, + 0.01758946105837822, + 0.030696174129843712, + 0.037327028810977936, + 0.031616754829883575, + 0.0119064562022686, + 0.012080793268978596, + 0.015456123277544975, + 0.02178283780813217, + 0.012541918084025383, + 0.015909837558865547, + 0.01568029820919037, + 0.014442252926528454, + 0.004231932107359171, + 0.004688219632953405, + 0.007220890838652849, + 0.007614404894411564, + 0.006250368896871805, + 0.006491774693131447, + 0.008859347552061081, + 0.004946291912347078, + 0.007357950787991285, + 0.005024048499763012, + 0.01279368158429861, + 0.008473028428852558, + 0.00687723234295845, + 0.01105236355215311, + 0.006272132508456707, + 0.006302179303020239, + 0.007196170277893543, + 0.004933403339236975, + 0.006100187543779612, + 0.00502568855881691, + 0.007012475747615099, + 0.007172707002609968, + 0.007971457205712795, + 0.004421574994921684, + 0.007023954764008522, + 0.007185665890574455, + 0.008732800371944904, + 0.005236586555838585, + 0.006110301706939936, + 0.00975764635950327, + 0.007883070036768913, + 0.009528275579214096, + 0.016520913690328598, + 0.005671782419085503, + 0.010102641768753529, + 0.0049497089348733425, + 0.004748912062495947, + 0.018061231821775436, + 0.0035498093347996473, + 0.007136375177651644, + 0.020671602338552475, + 0.01260144542902708, + 0.013858023099601269, + 0.009423698298633099, + 0.007716997526586056, + 0.037203527987003326, + 0.009911151602864265, + 0.021038256585597992, + 0.01862768828868866, + 0.010345480404794216, + 0.01618207059800625, + 0.018725423142313957, + 0.008942663669586182, + 0.018683934584259987, + 0.014775545336306095, + 0.012743957340717316, + 0.02782936580479145, + 0.008462951518595219, + 0.008328882977366447, + 0.0070145949721336365, + 0.0054353876039385796, + 0.012665752321481705, + 0.010591383092105389, + 0.019392792135477066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.08903750032186508 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865511879324913 + ], + "opposite_logprob_tokens": [ + -3.0486550331115723 + ], + "logprob_reward": -0.04865511879324913, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865511879324913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2149943858385086, + 0.07414847612380981, + 0.04336589202284813, + 0.08107569068670273, + 0.04051557555794716, + 0.09472823143005371, + 0.05120260640978813, + 0.17581965029239655, + 0.03320411592721939, + 0.19094538688659668 + ], + "credit_map": [ + 0.02973952703177929, + 0.018787387758493423, + 0.024950530380010605, + 0.032946664839982986, + 0.015751896426081657, + 0.02888759970664978, + 0.0334039032459259, + 0.03052687458693981, + 0.01012425310909748, + 0.007967188954353333, + 0.006654354277998209, + 0.008555236272513866, + 0.009464152157306671, + 0.012222816236317158, + 0.011282567866146564, + 0.00787790585309267, + 0.005670595448464155, + 0.003944090101867914, + 0.008142547681927681, + 0.004652055911719799, + 0.004508737940341234, + 0.005065527744591236, + 0.0059009152464568615, + 0.005481420550495386, + 0.010128745809197426, + 0.0055899666622281075, + 0.019237464293837547, + 0.008159941993653774, + 0.0057043335400521755, + 0.015382591634988785, + 0.006929324008524418, + 0.009943323209881783, + 0.005227834917604923, + 0.005103271454572678, + 0.006344619672745466, + 0.0035445995163172483, + 0.005705488845705986, + 0.004679588135331869, + 0.00620421115309, + 0.0037059616297483444, + 0.011499343439936638, + 0.007741070818156004, + 0.016240576282143593, + 0.01249113492667675, + 0.006299018394201994, + 0.018531298264861107, + 0.008481016382575035, + 0.01344477292150259, + 0.009050566703081131, + 0.004517590161412954, + 0.0076703960075974464, + 0.004647755529731512, + 0.004006834235042334, + 0.007882448844611645, + 0.005172549746930599, + 0.008254464715719223, + 0.020396221429109573, + 0.01471663173288107, + 0.014020505361258984, + 0.009872183203697205, + 0.009347868151962757, + 0.05373939126729965, + 0.01059788279235363, + 0.04312896728515625, + 0.0073189386166632175, + 0.003595777554437518, + 0.002337676240131259, + 0.003283990314230323, + 0.0025200422387570143, + 0.009730477817356586, + 0.0022563827224075794, + 0.002160828560590744, + 0.04599222168326378, + 0.01888274773955345, + 0.014580733142793179, + 0.015362115576863289, + 0.016456743702292442, + 0.05242585018277168, + 0.01272889319807291, + 0.014516082592308521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.04865511879324913 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896091043949127 + ], + "opposite_logprob_tokens": [ + -2.57896089553833 + ], + "logprob_reward": -0.07896091043949127, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896091043949127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2186097502708435, + 0.01974327117204666, + 0.06488709896802902, + 0.030453892424702644, + 0.04920118302106857, + 0.1901475191116333, + 0.028101108968257904, + 0.33412784337997437, + 0.037743620574474335, + 0.026984745636582375 + ], + "credit_map": [ + 0.03463650494813919, + 0.02313999831676483, + 0.03301539644598961, + 0.02917330339550972, + 0.016321947798132896, + 0.029678257182240486, + 0.02511131763458252, + 0.02753301151096821, + 0.0033466960303485394, + 0.0021575645077973604, + 0.002295037265866995, + 0.0023180292919278145, + 0.001416900777257979, + 0.0033136773854494095, + 0.0021489381324499846, + 0.0027464283630251884, + 0.009657083079218864, + 0.006232777610421181, + 0.013941013254225254, + 0.005784850567579269, + 0.005363414529711008, + 0.008340389467775822, + 0.006783867720514536, + 0.008783705532550812, + 0.0030942545272409916, + 0.0017619343707337976, + 0.007862369529902935, + 0.003443711204454303, + 0.0022278029937297106, + 0.00630878284573555, + 0.0026809475384652615, + 0.0030740899965167046, + 0.007971201092004776, + 0.003944439347833395, + 0.011870361864566803, + 0.004151104483753443, + 0.004653849173337221, + 0.0070096212439239025, + 0.004516230430454016, + 0.005084373988211155, + 0.029325993731617928, + 0.018564503639936447, + 0.0340939424932003, + 0.026211965829133987, + 0.013430752791464329, + 0.03116920217871666, + 0.01752021349966526, + 0.019830936565995216, + 0.005938356276601553, + 0.0035184125881642103, + 0.00436687096953392, + 0.002038130769506097, + 0.002126899315044284, + 0.004806696902960539, + 0.0020397277548909187, + 0.0032660160213708878, + 0.05944036319851875, + 0.01445862092077732, + 0.02046986110508442, + 0.01711537130177021, + 0.01056843250989914, + 0.06639041751623154, + 0.013888098299503326, + 0.13179665803909302, + 0.009823241271078587, + 0.004335883539170027, + 0.0027690858114510775, + 0.002083705272525549, + 0.002004762180149555, + 0.011749709025025368, + 0.002180845243856311, + 0.00279638753272593, + 0.007896567694842815, + 0.0023615818936377764, + 0.002149753738194704, + 0.002004950540140271, + 0.0014715885045006871, + 0.003269059583544731, + 0.0023854454047977924, + 0.005445798393338919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.07896091043949127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle, static, full-body shot captures a god-warrior's unwavering stand against catastrophic events, created in the epic, high-stakes style of an action film scene. He is clad head to toe in gleaming celestial armor, its polished, iridescent white gold surface intricately etched with cosmic patterns, reflecting the chaotic ambient light. His feet planted broadly on a jagged, damp rock shelf, his posture resolute, his hands poised, firmly gripping a mighty three-pointed spear. In the center of his forehead, a third eye blazes with intense, pulsating azure power, illuminating his grim resolve. Behind him, a massive river flood rages, a roaring, tumbling wall of mud, water, and debris violently tearing through a deep gorge, stirring up a thick spray. The atmosphere is tense and dramatic, dominated by a cool, desaturated palette of dark grays and browns, contrasting sharply with the brilliant radiance of the armor and the dazzling blue glow of his power. The camera remains locked, emphasizing his steadfastness as he stands as the sole point of light and order against the overwhelming and destructive flood.", + "seed": 13313, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_9/index_8235_seed_13313_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05387312900274992, + "vlm_logprob": -0.05387312900274992, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.05387312900274992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005143742077052593 + ], + "opposite_logprob_tokens": [ + -6.505143642425537 + ], + "logprob_reward": -0.005143742077052593, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.005143742077052593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23110836744308472, + 0.11647678166627884, + 0.061745963990688324, + 0.10506213456392288, + 0.1101304143667221, + 0.06640860438346863, + 0.06209912896156311, + 0.07124942541122437, + 0.0921887755393982, + 0.08353034406900406 + ], + "credit_map": [ + 0.029976116493344307, + 0.02808234840631485, + 0.02792036533355713, + 0.022194085642695427, + 0.02893768437206745, + 0.029243722558021545, + 0.026863140985369682, + 0.03789090737700462, + 0.02282647229731083, + 0.02144562266767025, + 0.012952509336173534, + 0.013459485024213791, + 0.01046005915850401, + 0.012763923034071922, + 0.007680265698581934, + 0.014888457953929901, + 0.008339430205523968, + 0.009295662865042686, + 0.006987274158746004, + 0.011437277309596539, + 0.004838364198803902, + 0.007277187891304493, + 0.005377826280891895, + 0.008192951790988445, + 0.009077801369130611, + 0.031255919486284256, + 0.006795776542276144, + 0.007284726947546005, + 0.011424745433032513, + 0.020221250131726265, + 0.007562220562249422, + 0.011439698748290539, + 0.010787381790578365, + 0.0166506078094244, + 0.007677892688661814, + 0.009605140425264835, + 0.01139208022505045, + 0.01923532597720623, + 0.007039372809231281, + 0.027742622420191765, + 0.007726224139332771, + 0.010922417975962162, + 0.008291853591799736, + 0.01171280350536108, + 0.004172664601355791, + 0.009015091694891453, + 0.005284439772367477, + 0.009283112362027168, + 0.006016491446644068, + 0.009329909458756447, + 0.005791873205453157, + 0.008543577045202255, + 0.008035143837332726, + 0.006973592098802328, + 0.01028119120746851, + 0.007127352990210056, + 0.00801767036318779, + 0.014186855405569077, + 0.008081996813416481, + 0.007635926362127066, + 0.005551027599722147, + 0.009749522432684898, + 0.00699276477098465, + 0.011033671908080578, + 0.013641221448779106, + 0.02128673531115055, + 0.010494615882635117, + 0.009142580442130566, + 0.007706620264798403, + 0.012664435431361198, + 0.007172395475208759, + 0.010080172680318356, + 0.014140226878225803, + 0.012823673896491528, + 0.010642045177519321, + 0.007869255729019642, + 0.011203385889530182, + 0.010112492367625237, + 0.009537569247186184, + 0.007201703265309334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005143742077052593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634909749031067 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.02634909749031067, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634909749031067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1761876344680786, + 0.0707942321896553, + 0.09441612660884857, + 0.32874229550361633, + 0.08200512826442719, + 0.04757559671998024, + 0.02861637994647026, + 0.055266477167606354, + 0.05896226316690445, + 0.057433851063251495 + ], + "credit_map": [ + 0.020628036931157112, + 0.018798070028424263, + 0.020990464836359024, + 0.017517464235424995, + 0.017514405772089958, + 0.031041357666254044, + 0.021463200449943542, + 0.02823461964726448, + 0.01361983921378851, + 0.014143139123916626, + 0.007609526626765728, + 0.008176214061677456, + 0.005244900938123465, + 0.008127703331410885, + 0.006089088506996632, + 0.007783815730363131, + 0.011345693841576576, + 0.013058103621006012, + 0.007703364361077547, + 0.010657128877937794, + 0.015011539682745934, + 0.010087106376886368, + 0.013177305459976196, + 0.013375888578593731, + 0.02687341719865799, + 0.14115087687969208, + 0.010696658864617348, + 0.007839197292923927, + 0.013000153005123138, + 0.10912502557039261, + 0.007729387376457453, + 0.012327572330832481, + 0.00856867153197527, + 0.010993318632245064, + 0.0053705088794231415, + 0.005975015927106142, + 0.010232037864625454, + 0.014141632243990898, + 0.007007637992501259, + 0.019716300070285797, + 0.005006744526326656, + 0.006178391631692648, + 0.004896911792457104, + 0.006351094227284193, + 0.0054478719830513, + 0.006449698470532894, + 0.003732238430529833, + 0.009512643329799175, + 0.0022857077419757843, + 0.0038463063538074493, + 0.002460283925756812, + 0.002576284809038043, + 0.005346206948161125, + 0.003081094240769744, + 0.004099583253264427, + 0.0049209133721888065, + 0.006285517942160368, + 0.009264170192182064, + 0.006939612329006195, + 0.007469231262803078, + 0.0063289739191532135, + 0.0072685242630541325, + 0.004900799598544836, + 0.006809648592025042, + 0.006323511246591806, + 0.009328346699476242, + 0.0048051076009869576, + 0.0063641793094575405, + 0.010221104137599468, + 0.007678441237658262, + 0.00746399350464344, + 0.006777577567845583, + 0.0057563805021345615, + 0.011633709073066711, + 0.009379447437822819, + 0.00745417270809412, + 0.006007858086377382, + 0.006941401399672031, + 0.005930193699896336, + 0.004330689087510109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.02634909749031067 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054949332028627396 + ], + "opposite_logprob_tokens": [ + -2.9299492835998535 + ], + "logprob_reward": -0.054949332028627396, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054949332028627396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19775980710983276, + 0.0788278877735138, + 0.05978803709149361, + 0.22130145132541656, + 0.06602147221565247, + 0.029977019876241684, + 0.1197364553809166, + 0.04211781173944473, + 0.04577292874455452, + 0.1386970579624176 + ], + "credit_map": [ + 0.023833895102143288, + 0.02121386118233204, + 0.018195360898971558, + 0.018517522141337395, + 0.018190080299973488, + 0.03141121566295624, + 0.02914375066757202, + 0.03725413978099823, + 0.008294183760881424, + 0.008798245340585709, + 0.008070485666394234, + 0.010660328902304173, + 0.009564845822751522, + 0.009334344416856766, + 0.015105944126844406, + 0.008999517187476158, + 0.0068295677192509174, + 0.008978855796158314, + 0.004864195361733437, + 0.008222525008022785, + 0.0077880495227873325, + 0.007289727218449116, + 0.00836680456995964, + 0.007448319811373949, + 0.018736889585852623, + 0.09458458423614502, + 0.006774744018912315, + 0.0058769346214830875, + 0.009065059013664722, + 0.06880621612071991, + 0.006024068221449852, + 0.011432984843850136, + 0.006139521487057209, + 0.008272480219602585, + 0.003066598204895854, + 0.003965901210904121, + 0.00659882090985775, + 0.0112462742254138, + 0.006348621565848589, + 0.020383264869451523, + 0.0034343695733696222, + 0.004231433849781752, + 0.0033410685136914253, + 0.004526154603809118, + 0.0034971514251083136, + 0.0037079723551869392, + 0.002671451773494482, + 0.0045674205757677555, + 0.009658689610660076, + 0.014277143403887749, + 0.009521194733679295, + 0.011443673633038998, + 0.01649230159819126, + 0.015828369185328484, + 0.0229075625538826, + 0.019607527181506157, + 0.004433027468621731, + 0.007897438481450081, + 0.006199055816978216, + 0.00516839325428009, + 0.0036932227667421103, + 0.0054311733692884445, + 0.0032584485597908497, + 0.0060370564460754395, + 0.0066904835402965546, + 0.009071260690689087, + 0.004163270816206932, + 0.0050088814459741116, + 0.005527157336473465, + 0.006545950658619404, + 0.003757745958864689, + 0.00500818295404315, + 0.013368776999413967, + 0.0240866057574749, + 0.015850896015763283, + 0.018852297216653824, + 0.018158558756113052, + 0.016055051237344742, + 0.019502047449350357, + 0.012822829186916351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.054949332028627396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700187087059021 + ], + "opposite_logprob_tokens": [ + -2.695018768310547 + ], + "logprob_reward": -0.0700187087059021, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700187087059021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21795988082885742, + 0.07371038943529129, + 0.05168360844254494, + 0.2719045877456665, + 0.12771610915660858, + 0.03534674644470215, + 0.033972933888435364, + 0.06268279999494553, + 0.05131981521844864, + 0.07370319217443466 + ], + "credit_map": [ + 0.024284910410642624, + 0.025730499997735023, + 0.028264595195651054, + 0.02357712760567665, + 0.022094637155532837, + 0.03789130970835686, + 0.029081592336297035, + 0.027035176753997803, + 0.010636103339493275, + 0.01245210599154234, + 0.009425665251910686, + 0.009280548430979252, + 0.007708510849624872, + 0.00782178994268179, + 0.0055811526253819466, + 0.010804500430822372, + 0.00677570141851902, + 0.0068227797746658325, + 0.005215432029217482, + 0.004850526340305805, + 0.006236435379832983, + 0.0066614458337426186, + 0.006795220077037811, + 0.00832605641335249, + 0.022569572553038597, + 0.11463607102632523, + 0.012086943723261356, + 0.009026684798300266, + 0.011001335456967354, + 0.08537543565034866, + 0.0070626335218548775, + 0.010145855136215687, + 0.012387700378894806, + 0.015918046236038208, + 0.006338529288768768, + 0.008295300416648388, + 0.01088015642017126, + 0.02753691002726555, + 0.008873507380485535, + 0.03748594596982002, + 0.004272035323083401, + 0.0041392189450562, + 0.005379387177526951, + 0.004444621503353119, + 0.002015025820583105, + 0.0049057831056416035, + 0.0027258419431746006, + 0.00746482377871871, + 0.0033943301532417536, + 0.004965786822140217, + 0.0038726460188627243, + 0.0032466494012624025, + 0.004768752958625555, + 0.004294696729630232, + 0.0035100048407912254, + 0.005920059978961945, + 0.0075098806992173195, + 0.011583891697227955, + 0.009651190601289272, + 0.008373892866075039, + 0.003849571803584695, + 0.008152625523507595, + 0.003983984235674143, + 0.009577745571732521, + 0.007943088188767433, + 0.011187577620148659, + 0.005932033993303776, + 0.006612966302782297, + 0.003801266197115183, + 0.006349056493490934, + 0.0037808020133525133, + 0.005713012535125017, + 0.007678565103560686, + 0.015371942892670631, + 0.013920890167355537, + 0.01005607470870018, + 0.0060536363162100315, + 0.009029578417539597, + 0.005767760332673788, + 0.0058247302658855915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0700187087059021 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290476471185684 + ], + "opposite_logprob_tokens": [ + -2.2379047870635986 + ], + "logprob_reward": -0.11290476471185684, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290476471185684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1680157333612442, + 0.10609373450279236, + 0.10584837943315506, + 0.23431652784347534, + 0.03546242415904999, + 0.0478014275431633, + 0.01799013651907444, + 0.08850886672735214, + 0.11900889128446579, + 0.07695396989583969 + ], + "credit_map": [ + 0.01865708827972412, + 0.02105296403169632, + 0.016683319583535194, + 0.022214751690626144, + 0.010264333337545395, + 0.03488989919424057, + 0.012748343870043755, + 0.031505037099123, + 0.015005076304078102, + 0.019703976809978485, + 0.014992092736065388, + 0.01533510908484459, + 0.007600977551192045, + 0.012164649553596973, + 0.007010847795754671, + 0.014281000010669231, + 0.017231004312634468, + 0.020043320953845978, + 0.010821052826941013, + 0.01330097857862711, + 0.00659907516092062, + 0.014215100556612015, + 0.007986580021679401, + 0.01565125770866871, + 0.0207502543926239, + 0.09991961717605591, + 0.009063095785677433, + 0.008122261613607407, + 0.008786571212112904, + 0.07074321061372757, + 0.0056249964982271194, + 0.011306502856314182, + 0.00602473272010684, + 0.0064131817780435085, + 0.004313950426876545, + 0.003830548608675599, + 0.003145494032651186, + 0.006575188133865595, + 0.002106946427375078, + 0.0030523783061653376, + 0.006324786227196455, + 0.006609127391129732, + 0.006020703818649054, + 0.005525608547031879, + 0.0029032803140580654, + 0.008095345459878445, + 0.0035859846975654364, + 0.00873658712953329, + 0.0021778158843517303, + 0.0026214802637696266, + 0.0016766375629231334, + 0.0017904328415170312, + 0.002128247171640396, + 0.0022983539383858442, + 0.00322527764365077, + 0.0020718907471746206, + 0.010625242255628109, + 0.01735854335129261, + 0.015753936022520065, + 0.010545726865530014, + 0.004358820151537657, + 0.01219947636127472, + 0.005535636562854052, + 0.012131478637456894, + 0.0175032801926136, + 0.03050384111702442, + 0.012313089333474636, + 0.013574487529695034, + 0.007317608222365379, + 0.01746969297528267, + 0.008081980980932713, + 0.012244908139109612, + 0.008679021149873734, + 0.017746010795235634, + 0.011998835951089859, + 0.009133952669799328, + 0.006369962356984615, + 0.00910277757793665, + 0.006898978725075722, + 0.007024427875876427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.11290476471185684 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle, static, full-body shot captures a god-warrior's unwavering stand against catastrophic events, created in the epic, high-stakes style of an action film scene. He is clad head to toe in gleaming celestial armor, its polished, iridescent white gold surface intricately etched with cosmic patterns, reflecting the chaotic ambient light. His feet planted broadly on a jagged, damp rock shelf, his posture resolute, his hands poised, firmly gripping a mighty three-pointed spear. In the center of his forehead, a third eye blazes with intense, pulsating azure power, illuminating his grim resolve. Behind him, a massive river flood rages, a roaring, tumbling wall of mud, water, and debris violently tearing through a deep gorge, stirring up a thick spray. The atmosphere is tense and dramatic, dominated by a cool, desaturated palette of dark grays and browns, contrasting sharply with the brilliant radiance of the armor and the dazzling blue glow of his power. The camera remains locked, emphasizing his steadfastness as he stands as the sole point of light and order against the overwhelming and destructive flood.", + "seed": 13314, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000153/group_9/index_8235_seed_13314_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.25243052542209626, + "vlm_logprob": -0.25243052542209626, + "vlm_yes_no_margin": 1.475, + "vlm_token_credit": -0.25243052542209626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5810459852218628 + ], + "opposite_logprob_tokens": [ + -0.8310459852218628 + ], + "logprob_reward": -0.5810459852218628, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5810459852218628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40409085154533386, + 0.047754544764757156, + 0.05153106153011322, + 0.0529104508459568, + 0.039014022797346115, + 0.03202942758798599, + 0.06288556754589081, + 0.16924241185188293, + 0.08418440818786621, + 0.05635729059576988 + ], + "credit_map": [ + 0.10075920075178146, + 0.05051634460687637, + 0.04287717491388321, + 0.03875742480158806, + 0.03888814151287079, + 0.06649082899093628, + 0.02966596931219101, + 0.036135755479335785, + 0.005863596685230732, + 0.006024099886417389, + 0.004415244795382023, + 0.006079853978008032, + 0.007959777489304543, + 0.0060673984698951244, + 0.007111607119441032, + 0.004232965409755707, + 0.007768925745040178, + 0.005533869843930006, + 0.010112080723047256, + 0.005305939354002476, + 0.004751450382173061, + 0.00739081809297204, + 0.004035524558275938, + 0.006632453296333551, + 0.005655257496982813, + 0.005987329874187708, + 0.003858142765238881, + 0.0062585268169641495, + 0.007205142639577389, + 0.005933875218033791, + 0.011157081462442875, + 0.00685509480535984, + 0.005153616424649954, + 0.004409712739288807, + 0.004650089889764786, + 0.005486750975251198, + 0.004207318183034658, + 0.0045288801193237305, + 0.0044405171647667885, + 0.006137141026556492, + 0.004137229640036821, + 0.0030504276510328054, + 0.004380491562187672, + 0.003570265369489789, + 0.003199125872924924, + 0.006535602267831564, + 0.00372061925008893, + 0.0034356683026999235, + 0.009547989815473557, + 0.00382061256095767, + 0.0073829395696520805, + 0.004993999842554331, + 0.004592362325638533, + 0.00962196197360754, + 0.006748277693986893, + 0.016177421435713768, + 0.01947830617427826, + 0.013382893055677414, + 0.022707751020789146, + 0.017194371670484543, + 0.019952844828367233, + 0.019723238423466682, + 0.02395588345825672, + 0.032847121357917786, + 0.011343677528202534, + 0.008545641787350178, + 0.010633918456733227, + 0.00754516152665019, + 0.010377384722232819, + 0.016002118587493896, + 0.009930809959769249, + 0.009805692359805107, + 0.007465657312422991, + 0.004778269678354263, + 0.006529208738356829, + 0.005365757271647453, + 0.00761466845870018, + 0.008282508701086044, + 0.00822632946074009, + 0.008094885386526585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.5810459852218628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281355082988739 + ], + "opposite_logprob_tokens": [ + -1.4063551425933838 + ], + "logprob_reward": -0.281355082988739, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281355082988739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2813037931919098, + 0.052241332828998566, + 0.05758170410990715, + 0.06287208199501038, + 0.08828083425760269, + 0.04091393202543259, + 0.04674321413040161, + 0.08405200392007828, + 0.08990655839443207, + 0.19610455632209778 + ], + "credit_map": [ + 0.05313192680478096, + 0.03780098259449005, + 0.03296699747443199, + 0.031092718243598938, + 0.032045017927885056, + 0.04064140096306801, + 0.026754125952720642, + 0.0268706064671278, + 0.011753086932003498, + 0.006760822143405676, + 0.0062407744117081165, + 0.0048372684977948666, + 0.005671558901667595, + 0.007330575026571751, + 0.005133608356118202, + 0.004513640888035297, + 0.009793587028980255, + 0.007173758931457996, + 0.00795403216034174, + 0.0051153721287846565, + 0.004940255545079708, + 0.008172063156962395, + 0.004996915347874165, + 0.00943571887910366, + 0.009948004968464375, + 0.0059318868443369865, + 0.011716339737176895, + 0.006049098446965218, + 0.004433302208781242, + 0.008200984448194504, + 0.0058292206376791, + 0.010763240978121758, + 0.010569103062152863, + 0.010840632021427155, + 0.014798409305512905, + 0.01096264086663723, + 0.007886427454650402, + 0.011246820911765099, + 0.008347692899405956, + 0.013629104010760784, + 0.005657619796693325, + 0.004414757248014212, + 0.004721538629382849, + 0.004828670062124729, + 0.004923035856336355, + 0.007262029219418764, + 0.005066807381808758, + 0.004039473831653595, + 0.0066616348922252655, + 0.003508520545437932, + 0.00395849347114563, + 0.00587060023099184, + 0.003968977835029364, + 0.006753290072083473, + 0.008102824911475182, + 0.007918871939182281, + 0.009578072465956211, + 0.008452733978629112, + 0.013717703521251678, + 0.00890835840255022, + 0.00873568281531334, + 0.012557885609567165, + 0.011745434254407883, + 0.010356131941080093, + 0.012047762982547283, + 0.008661570027470589, + 0.012432635761797428, + 0.008027601055800915, + 0.011264251545071602, + 0.016649635508656502, + 0.009469250217080116, + 0.011353849433362484, + 0.025424795225262642, + 0.01644117757678032, + 0.024132026359438896, + 0.018071193248033524, + 0.0267984289675951, + 0.0286876130849123, + 0.026192257180809975, + 0.030357064679265022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.281355082988739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038191318511963 + ], + "opposite_logprob_tokens": [ + -1.9103819131851196 + ], + "logprob_reward": -0.16038191318511963, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038191318511963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37904343008995056, + 0.1075674444437027, + 0.07237988710403442, + 0.0987967848777771, + 0.06310489773750305, + 0.033897608518600464, + 0.044391125440597534, + 0.11325900256633759, + 0.042249176651239395, + 0.04531065374612808 + ], + "credit_map": [ + 0.09221533685922623, + 0.04865872114896774, + 0.03798387572169304, + 0.03082900121808052, + 0.04574817791581154, + 0.060564056038856506, + 0.027657747268676758, + 0.035386521369218826, + 0.015209339559078217, + 0.012935969047248363, + 0.010202625766396523, + 0.011347715742886066, + 0.01942395232617855, + 0.014903527684509754, + 0.013676808215677738, + 0.009867507964372635, + 0.01246038731187582, + 0.008464922197163105, + 0.011479495093226433, + 0.008400315418839455, + 0.005909482948482037, + 0.009973639622330666, + 0.005561214406043291, + 0.010130428709089756, + 0.012566476128995419, + 0.010497485287487507, + 0.008381793275475502, + 0.015004454180598259, + 0.011801586486399174, + 0.012773524969816208, + 0.009954183362424374, + 0.01781727932393551, + 0.007313174195587635, + 0.006749384105205536, + 0.010120922699570656, + 0.01075364463031292, + 0.006676504388451576, + 0.00678232591599226, + 0.006493901368230581, + 0.00821504183113575, + 0.004512222018092871, + 0.0032011703588068485, + 0.004897187929600477, + 0.0035058180801570415, + 0.0028892094269394875, + 0.005974466912448406, + 0.0036726088728755713, + 0.0052449265494942665, + 0.005444170441478491, + 0.003303845878690481, + 0.005640074145048857, + 0.004361176863312721, + 0.003910883329808712, + 0.006769205909222364, + 0.004849114455282688, + 0.010112655349075794, + 0.010930302552878857, + 0.011028640903532505, + 0.01581370271742344, + 0.016698045656085014, + 0.014835390262305737, + 0.011461536400020123, + 0.017584308981895447, + 0.014907077886164188, + 0.005173007026314735, + 0.00635214289650321, + 0.004354325123131275, + 0.004771434236317873, + 0.00472696777433157, + 0.0062246485613286495, + 0.005049761384725571, + 0.00559689337387681, + 0.005580482538789511, + 0.005121887661516666, + 0.004101716447621584, + 0.004073768854141235, + 0.008031255565583706, + 0.006624849978834391, + 0.005457374732941389, + 0.006319318898022175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.16038191318511963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603955179452896 + ], + "opposite_logprob_tokens": [ + -1.9103955030441284 + ], + "logprob_reward": -0.1603955179452896, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603955179452896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.331264466047287, + 0.05403103306889534, + 0.060321468859910965, + 0.07339022308588028, + 0.08419637382030487, + 0.03907791152596474, + 0.07967431098222733, + 0.06176239997148514, + 0.04002843052148819, + 0.17625340819358826 + ], + "credit_map": [ + 0.07833068072795868, + 0.04093322902917862, + 0.035326819866895676, + 0.04144875332713127, + 0.027698379009962082, + 0.051261838525533676, + 0.026384221389889717, + 0.029880497604608536, + 0.012829363346099854, + 0.006071880459785461, + 0.006792339961975813, + 0.004882392473518848, + 0.004781982861459255, + 0.008008741773664951, + 0.005667291581630707, + 0.004997034557163715, + 0.006730871275067329, + 0.0077945454977452755, + 0.006664881017059088, + 0.005485705099999905, + 0.006902618799358606, + 0.008560271002352238, + 0.006553796585649252, + 0.011628776788711548, + 0.00646359845995903, + 0.007029273081570864, + 0.011373041197657585, + 0.012603864073753357, + 0.008518151007592678, + 0.008162221871316433, + 0.01042198110371828, + 0.008818085305392742, + 0.010359466075897217, + 0.011258097365498543, + 0.012691791169345379, + 0.011803146451711655, + 0.007808265741914511, + 0.009542123414576054, + 0.0086636608466506, + 0.012069822289049625, + 0.005163120571523905, + 0.0037882879842072725, + 0.005636332556605339, + 0.00333958538249135, + 0.005022732075303793, + 0.00804927758872509, + 0.003687845775857568, + 0.00439072772860527, + 0.00978601910173893, + 0.004339682403951883, + 0.008923212066292763, + 0.006874053739011288, + 0.006715032272040844, + 0.011215838603675365, + 0.009727375581860542, + 0.022093089297413826, + 0.007337889168411493, + 0.007786917500197887, + 0.008177646435797215, + 0.00900415051728487, + 0.006871976889669895, + 0.006143460050225258, + 0.007061055861413479, + 0.009379297494888306, + 0.004014116711914539, + 0.004086112137883902, + 0.004699533339589834, + 0.006356419995427132, + 0.005277243442833424, + 0.004509415477514267, + 0.005989611614495516, + 0.005095976870507002, + 0.022070489823818207, + 0.01438435073941946, + 0.022797340527176857, + 0.015494407154619694, + 0.026442106813192368, + 0.021848438307642937, + 0.026179790496826172, + 0.02703648805618286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.1603955179452896 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a god-warrior standing in a resolute, full-body pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior clad in gleaming celestial armor with a polished, iridescent white gold surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior gripping a mighty three-pointed spear with both hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a third eye on the warrior's forehead blazing with intense, pulsating azure power?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a massive river flood raging through a deep gorge with mud, water, and debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897412776947021 + ], + "opposite_logprob_tokens": [ + -2.5789742469787598 + ], + "logprob_reward": -0.07897412776947021, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897412776947021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24939775466918945, + 0.054506540298461914, + 0.08890677243471146, + 0.1166333556175232, + 0.05651048198342323, + 0.03650283440947533, + 0.18959656357765198, + 0.16611401736736298, + 0.0137641541659832, + 0.028067566454410553 + ], + "credit_map": [ + 0.06638035923242569, + 0.040635790675878525, + 0.02320042997598648, + 0.019728627055883408, + 0.016455290839076042, + 0.04117564857006073, + 0.016270650550723076, + 0.025550946593284607, + 0.008554819971323013, + 0.010340589098632336, + 0.006253517232835293, + 0.005423619877547026, + 0.0036893996875733137, + 0.007883789949119091, + 0.004622644279152155, + 0.007738159969449043, + 0.013622961938381195, + 0.011749135330319405, + 0.012726559303700924, + 0.01370919868350029, + 0.00437407661229372, + 0.011942082084715366, + 0.006791764870285988, + 0.013990993611514568, + 0.023959210142493248, + 0.012002440169453621, + 0.013682588934898376, + 0.014193934388458729, + 0.007262705359607935, + 0.018500033766031265, + 0.009812897071242332, + 0.01721954345703125, + 0.008369360119104385, + 0.0067831105552613735, + 0.0064402371644973755, + 0.008227662183344364, + 0.005457836203277111, + 0.006669736001640558, + 0.0064207580871880054, + 0.0081417765468359, + 0.00588823389261961, + 0.0036432535853236914, + 0.003976309206336737, + 0.0053589968010783195, + 0.0026902626268565655, + 0.007642865180969238, + 0.003550053108483553, + 0.0037528565153479576, + 0.022426309064030647, + 0.015158738940954208, + 0.029982231557369232, + 0.020854176953434944, + 0.0115583511069417, + 0.0303481575101614, + 0.017055653035640717, + 0.04221294820308685, + 0.020856795832514763, + 0.017813079059123993, + 0.032734762877225876, + 0.02201639860868454, + 0.013674231246113777, + 0.024600625038146973, + 0.015817157924175262, + 0.01860097423195839, + 0.0016838805750012398, + 0.0011463552946224809, + 0.0023825198877602816, + 0.0013372498797252774, + 0.0012379817198961973, + 0.0016321223229169846, + 0.0023568354081362486, + 0.001987209776416421, + 0.003141660476103425, + 0.002719016745686531, + 0.003774527693167329, + 0.003124222392216325, + 0.004075403790920973, + 0.0029683257453143597, + 0.003960359841585159, + 0.004304050002247095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.07897412776947021 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_0/index_tensor([118, 118])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_0/index_tensor([118, 118])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..861eec6b163df0d915ab3153659ce3b4f8f0114e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_0/index_tensor([118, 118])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a highly detailed, cinematic 3D animation style, a medium shot, shot from a slightly elevated angle, slowly pans rightward around the center of the scene, focusing the viewer's attention on a young practitioner seated cross-legged at the center of a vast, ancient stone Bagua formation. His long, jet-black hair is tied high with a simple wooden hairpin, his face delicate and focused, his eyes closed, as if lost in deep meditation. He wears a dark blue silk robe, the hem and cuffs meticulously embroidered with intricate cloud and star patterns in silver thread, shimmering with a mysterious luster in the dim light. Each deeply engraved rune in the Bagua formation emanates a soft, luminous cyan light, casting its primary light upward, dramatically outlining his resolute jawline and the deep folds of his robe. His hands, resting on his chest, rapidly form complex hand seals at a speed barely perceptible to the naked eye. Each interlacing, spreading, and closing of his fingers is precise and powerful, imbued with a sense of ritual. As his hand gestures danced, the surrounding air visibly twisted and rippled. Countless luminous energy particles, as if drawn by an invisible force, converged from all directions, forming spirals of translucent air that coiled around his arms and body, accelerating in time with the rhythm of his hand seals. The backdrop was a deep, boundless night sky, dotted with stars. Distant nebulae shone in dreamy purples and blues, creating a mysterious atmosphere of intense focus, tranquility, and profound inner strength.", + "seed": 28482, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_0/index_118_seed_28482_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13138183895498515, + "vlm_logprob": -0.13138183895498515, + "vlm_yes_no_margin": 3.175000023841858, + "vlm_token_credit": -0.13138183895498515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009574091993272305 + ], + "opposite_logprob_tokens": [ + -5.134573936462402 + ], + "logprob_reward": -0.009574091993272305, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009574091993272305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14058704674243927, + 0.1101883128285408, + 0.05077934265136719, + 0.08443164080381393, + 0.05348046496510506, + 0.03569038212299347, + 0.262081116437912, + 0.12989772856235504, + 0.05959436669945717, + 0.07326965034008026 + ], + "credit_map": [ + 0.017918046563863754, + 0.017352188006043434, + 0.019518470391631126, + 0.0207369327545166, + 0.013868510723114014, + 0.020675504580140114, + 0.018097806721925735, + 0.012419587932527065, + 0.014145412482321262, + 0.014593949541449547, + 0.015122567303478718, + 0.010824747383594513, + 0.013017837889492512, + 0.020976250991225243, + 0.012261220254004002, + 0.009246326982975006, + 0.0064096031710505486, + 0.006546165328472853, + 0.007627819664776325, + 0.005608776584267616, + 0.006483035162091255, + 0.0069219619035720825, + 0.005104501266032457, + 0.006077478639781475, + 0.010230647400021553, + 0.0077651990577578545, + 0.027537701651453972, + 0.00832053367048502, + 0.005256566219031811, + 0.011761864647269249, + 0.005284179467707872, + 0.008274948224425316, + 0.006012183614075184, + 0.007964950986206532, + 0.010627924464643002, + 0.004090548027306795, + 0.005422795657068491, + 0.006587690208107233, + 0.00572583032771945, + 0.007048542611300945, + 0.006006981246173382, + 0.0038430490531027317, + 0.004799749236553907, + 0.0033238499891012907, + 0.00506812147796154, + 0.0038548125885427, + 0.004556693602353334, + 0.0042371246963739395, + 0.034541040658950806, + 0.04734528064727783, + 0.04257363826036453, + 0.0183476060628891, + 0.015294683165848255, + 0.07450154423713684, + 0.01267554983496666, + 0.01680176891386509, + 0.03374383598566055, + 0.013459066860377789, + 0.009783066809177399, + 0.0067370799370110035, + 0.007218137849122286, + 0.04295020177960396, + 0.005530310794711113, + 0.010476023890078068, + 0.008388402871787548, + 0.006191902793943882, + 0.009588430635631084, + 0.005890251137316227, + 0.008920402266085148, + 0.007644945289939642, + 0.006156417075544596, + 0.006813612766563892, + 0.010430173948407173, + 0.008428565226495266, + 0.00973574910312891, + 0.007302423007786274, + 0.011379310861229897, + 0.009509432129561901, + 0.009663419798016548, + 0.00682057486847043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.009574091993272305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016088545322418213 + ], + "opposite_logprob_tokens": [ + -4.141088485717773 + ], + "logprob_reward": -0.016088545322418213, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016088545322418213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26275691390037537, + 0.07547223567962646, + 0.08550120890140533, + 0.05385222285985947, + 0.08332368731498718, + 0.0522337481379509, + 0.10850872099399567, + 0.06416025012731552, + 0.08989376574754715, + 0.12429720908403397 + ], + "credit_map": [ + 0.02919865772128105, + 0.03503548726439476, + 0.04176449403166771, + 0.05259658768773079, + 0.028633963316679, + 0.022291794419288635, + 0.03262314572930336, + 0.020612770691514015, + 0.006560805253684521, + 0.008896134793758392, + 0.011189223267138004, + 0.009284745901823044, + 0.010828234255313873, + 0.010881448164582253, + 0.010815206915140152, + 0.007016434334218502, + 0.014271567575633526, + 0.011253135278820992, + 0.015291326679289341, + 0.008063607849180698, + 0.00781332515180111, + 0.010683473199605942, + 0.009264734573662281, + 0.008860042318701744, + 0.008481352590024471, + 0.00466273445636034, + 0.006593219470232725, + 0.006395787000656128, + 0.004698734264820814, + 0.007015401031821966, + 0.006553169805556536, + 0.00945182517170906, + 0.010364743880927563, + 0.008230449631810188, + 0.010940764099359512, + 0.006428075488656759, + 0.011461161077022552, + 0.011848700232803822, + 0.012671312317252159, + 0.011378482915461063, + 0.008162975311279297, + 0.00779906427487731, + 0.004449853673577309, + 0.004409736022353172, + 0.008111213333904743, + 0.005980906542390585, + 0.007042822428047657, + 0.006277179345488548, + 0.014351298101246357, + 0.017558898776769638, + 0.0162753164768219, + 0.00830310583114624, + 0.00856809038668871, + 0.027856722474098206, + 0.007187573239207268, + 0.008407720364630222, + 0.008983055129647255, + 0.007977561093866825, + 0.008406857959926128, + 0.007984340190887451, + 0.007261291611939669, + 0.0070015788078308105, + 0.006185735110193491, + 0.01035983394831419, + 0.017079606652259827, + 0.0071831364184618, + 0.011494380421936512, + 0.008034403435885906, + 0.010494418442249298, + 0.019252734258770943, + 0.008180467411875725, + 0.00817461870610714, + 0.01490541361272335, + 0.013320681639015675, + 0.011368363164365292, + 0.013728874735534191, + 0.022600550204515457, + 0.016906578093767166, + 0.020346371456980705, + 0.011120366863906384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.016088545322418213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193302020430565 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193302020430565, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193302020430565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16635717451572418, + 0.10331520438194275, + 0.06646460294723511, + 0.07222837954759598, + 0.05753367021679878, + 0.09462445229291916, + 0.18687857687473297, + 0.10784070193767548, + 0.08237932622432709, + 0.06237798556685448 + ], + "credit_map": [ + 0.01947544328868389, + 0.022520683705806732, + 0.028347300365567207, + 0.029888469725847244, + 0.018708521500229836, + 0.017395617440342903, + 0.01652638427913189, + 0.013494743034243584, + 0.01048087328672409, + 0.014960791915655136, + 0.016225406900048256, + 0.011279520578682423, + 0.0077870371751487255, + 0.01758825033903122, + 0.013022364117205143, + 0.0119709437713027, + 0.008551239967346191, + 0.008203228004276752, + 0.009941576980054379, + 0.007567543536424637, + 0.010790679603815079, + 0.008263401687145233, + 0.005893754307180643, + 0.007253166753798723, + 0.008595830760896206, + 0.005828189663589001, + 0.022242218255996704, + 0.0062512620352208614, + 0.008108984678983688, + 0.009429983794689178, + 0.005193463526666164, + 0.006578441709280014, + 0.007905143313109875, + 0.00833260826766491, + 0.007807561662048101, + 0.004699067212641239, + 0.006729246117174625, + 0.008147574961185455, + 0.008209224790334702, + 0.005703239236027002, + 0.018553543835878372, + 0.012125804089009762, + 0.020417988300323486, + 0.007467169314622879, + 0.005659312941133976, + 0.011332952417433262, + 0.006977984681725502, + 0.01208969485014677, + 0.02558409608900547, + 0.03196964040398598, + 0.02998708002269268, + 0.01328620407730341, + 0.012848755344748497, + 0.05186130851507187, + 0.00871522817760706, + 0.012626240961253643, + 0.02699456736445427, + 0.012194757349789143, + 0.008013474754989147, + 0.005401679314672947, + 0.007084723096340895, + 0.0339730829000473, + 0.005083287134766579, + 0.009095128625631332, + 0.012702537700533867, + 0.007087890524417162, + 0.012003721669316292, + 0.00847352109849453, + 0.010350892320275307, + 0.014179005287587643, + 0.009294417686760426, + 0.008287337608635426, + 0.00867834035307169, + 0.006746836472302675, + 0.006822758354246616, + 0.006823996547609568, + 0.010658030398190022, + 0.008669141680002213, + 0.008725387044250965, + 0.005253489129245281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.014193302020430565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742823541164398 + ], + "opposite_logprob_tokens": [ + -0.9742823839187622 + ], + "logprob_reward": -0.4742823541164398, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742823541164398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2981383502483368, + 0.06358067691326141, + 0.1013527438044548, + 0.04569020867347717, + 0.0936228558421135, + 0.05026711896061897, + 0.08780639618635178, + 0.07684684544801712, + 0.0559256337583065, + 0.12676918506622314 + ], + "credit_map": [ + 0.0284709595143795, + 0.050412051379680634, + 0.03439788520336151, + 0.05134578421711922, + 0.036480724811553955, + 0.029405737295746803, + 0.04214679077267647, + 0.0254784245043993, + 0.007193189579993486, + 0.009624611586332321, + 0.006760532967746258, + 0.00606872234493494, + 0.009609824046492577, + 0.008379492908716202, + 0.009690839797258377, + 0.0062534622848033905, + 0.016433149576187134, + 0.013172101229429245, + 0.014749334193766117, + 0.008074202574789524, + 0.01396525464951992, + 0.012304892763495445, + 0.011209936812520027, + 0.01144387200474739, + 0.006698822136968374, + 0.004876752849668264, + 0.007363057229667902, + 0.00372370146214962, + 0.006706593092530966, + 0.005838139448314905, + 0.005874252878129482, + 0.00460888771340251, + 0.011893432587385178, + 0.012855653651058674, + 0.0113875363022089, + 0.008183849975466728, + 0.013187858276069164, + 0.012411574833095074, + 0.014877655543386936, + 0.008825298398733139, + 0.007999666966497898, + 0.00644994480535388, + 0.006310633383691311, + 0.004938967991620302, + 0.00636252760887146, + 0.005920282565057278, + 0.0066632055677473545, + 0.005621890537440777, + 0.010229668579995632, + 0.01489471085369587, + 0.008247419260442257, + 0.009230349212884903, + 0.010701377876102924, + 0.01821189932525158, + 0.007709520868957043, + 0.008581453002989292, + 0.013491025194525719, + 0.00806012749671936, + 0.009355593472719193, + 0.00635119341313839, + 0.006536251399666071, + 0.018277939409017563, + 0.005109015852212906, + 0.009665696881711483, + 0.009877290576696396, + 0.006476040929555893, + 0.005221921484917402, + 0.003951306454837322, + 0.009441121481359005, + 0.011873832903802395, + 0.0047917356714606285, + 0.004292382393032312, + 0.018218349665403366, + 0.013366300612688065, + 0.009585612453520298, + 0.015642164275050163, + 0.01952992007136345, + 0.020428787916898727, + 0.019680408760905266, + 0.010317639447748661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.4742823541164398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427709013223648 + ], + "opposite_logprob_tokens": [ + -2.017771005630493 + ], + "logprob_reward": -0.1427709013223648, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1427709013223648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1723446100950241, + 0.1172209158539772, + 0.05370936915278435, + 0.07487811893224716, + 0.08524139970541, + 0.04093065857887268, + 0.1642037183046341, + 0.10328640788793564, + 0.056272588670253754, + 0.1319122165441513 + ], + "credit_map": [ + 0.02174379676580429, + 0.01930728368461132, + 0.02956032007932663, + 0.022851163521409035, + 0.015085174702107906, + 0.033583689481019974, + 0.015965409576892853, + 0.014247764833271503, + 0.012381268665194511, + 0.01277119480073452, + 0.014820908196270466, + 0.012126035057008266, + 0.017437860369682312, + 0.019856547936797142, + 0.01635769195854664, + 0.011469417251646519, + 0.006582530215382576, + 0.006599212996661663, + 0.0066307480446994305, + 0.005196953192353249, + 0.009219854138791561, + 0.006610182113945484, + 0.005989591125398874, + 0.006880294065922499, + 0.009591333568096161, + 0.0053807636722922325, + 0.015054616145789623, + 0.008674594573676586, + 0.006636430975049734, + 0.009346249513328075, + 0.011363225057721138, + 0.008830904960632324, + 0.008012047968804836, + 0.009949898347258568, + 0.010886603966355324, + 0.010274488478899002, + 0.00992557406425476, + 0.009665396064519882, + 0.018747227266430855, + 0.007780164014548063, + 0.005994059611111879, + 0.0039877379313111305, + 0.006202611606568098, + 0.004587148316204548, + 0.006369218695908785, + 0.004340158775448799, + 0.004710017237812281, + 0.004739704076200724, + 0.02477852813899517, + 0.030734915286302567, + 0.014821630902588367, + 0.008603719063103199, + 0.012872199527919292, + 0.05235879495739937, + 0.010935829021036625, + 0.009098108857870102, + 0.020129546523094177, + 0.012127055786550045, + 0.010639602318406105, + 0.008096198551356792, + 0.012697341851890087, + 0.02205408737063408, + 0.008695702068507671, + 0.008846878074109554, + 0.01096184179186821, + 0.004746923688799143, + 0.006403276696801186, + 0.004823372233659029, + 0.006140175741165876, + 0.012448753230273724, + 0.006376138422638178, + 0.004372106399387121, + 0.015308745205402374, + 0.015374147333204746, + 0.0187511146068573, + 0.012403092347085476, + 0.019177505746483803, + 0.019756659865379333, + 0.015335246920585632, + 0.01580570638179779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.1427709013223648 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a highly detailed, cinematic 3D animation style, a medium shot, shot from a slightly elevated angle, slowly pans rightward around the center of the scene, focusing the viewer's attention on a young practitioner seated cross-legged at the center of a vast, ancient stone Bagua formation. His long, jet-black hair is tied high with a simple wooden hairpin, his face delicate and focused, his eyes closed, as if lost in deep meditation. He wears a dark blue silk robe, the hem and cuffs meticulously embroidered with intricate cloud and star patterns in silver thread, shimmering with a mysterious luster in the dim light. Each deeply engraved rune in the Bagua formation emanates a soft, luminous cyan light, casting its primary light upward, dramatically outlining his resolute jawline and the deep folds of his robe. His hands, resting on his chest, rapidly form complex hand seals at a speed barely perceptible to the naked eye. Each interlacing, spreading, and closing of his fingers is precise and powerful, imbued with a sense of ritual. As his hand gestures danced, the surrounding air visibly twisted and rippled. Countless luminous energy particles, as if drawn by an invisible force, converged from all directions, forming spirals of translucent air that coiled around his arms and body, accelerating in time with the rhythm of his hand seals. The backdrop was a deep, boundless night sky, dotted with stars. Distant nebulae shone in dreamy purples and blues, creating a mysterious atmosphere of intense focus, tranquility, and profound inner strength.", + "seed": 28483, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_0/index_118_seed_28483_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09331110082566738, + "vlm_logprob": -0.09331110082566738, + "vlm_yes_no_margin": 3.049999976158142, + "vlm_token_credit": -0.09331110082566738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018836896866559982 + ], + "opposite_logprob_tokens": [ + -4.268836975097656 + ], + "logprob_reward": -0.018836896866559982, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018836896866559982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3164047598838806, + 0.09350588172674179, + 0.048240140080451965, + 0.08192931860685349, + 0.028171060606837273, + 0.10111550986766815, + 0.03477386385202408, + 0.05335696414113045, + 0.08639063686132431, + 0.15611188113689423 + ], + "credit_map": [ + 0.04974057897925377, + 0.047446344047784805, + 0.05922077223658562, + 0.058664895594120026, + 0.028331058099865913, + 0.018509671092033386, + 0.023879803717136383, + 0.03061162307858467, + 0.011659672483801842, + 0.010856164619326591, + 0.018947098404169083, + 0.013622426427900791, + 0.012473022565245628, + 0.009371357969939709, + 0.007996561005711555, + 0.008579575456678867, + 0.006678236182779074, + 0.0063007245771586895, + 0.007395128719508648, + 0.006713701877743006, + 0.007622271776199341, + 0.003380633657798171, + 0.00526445871219039, + 0.00488498667255044, + 0.009288010187447071, + 0.008743218146264553, + 0.008045559749007225, + 0.009289097040891647, + 0.01054770965129137, + 0.01541566289961338, + 0.011095631867647171, + 0.00950443372130394, + 0.0034255185164511204, + 0.0027566899079829454, + 0.0034731982741504908, + 0.002258078195154667, + 0.004335971549153328, + 0.004404514562338591, + 0.0033480864949524403, + 0.004169002640992403, + 0.012719705700874329, + 0.011929814703762531, + 0.01583106815814972, + 0.011089274659752846, + 0.01477281004190445, + 0.01138176117092371, + 0.012148654088377953, + 0.01124242227524519, + 0.004475160501897335, + 0.005131976678967476, + 0.004370362963527441, + 0.0030727144330739975, + 0.006937618367373943, + 0.003736022859811783, + 0.002781278220936656, + 0.004268730524927378, + 0.007278157398104668, + 0.005642857402563095, + 0.008284698240458965, + 0.005187496542930603, + 0.006083398591727018, + 0.005811462178826332, + 0.004329661373049021, + 0.010739229619503021, + 0.010489458218216896, + 0.015499911271035671, + 0.012607228942215443, + 0.005305161699652672, + 0.005924294237047434, + 0.015594647265970707, + 0.005661042407155037, + 0.01530888956040144, + 0.019741231575608253, + 0.026852255687117577, + 0.02365470677614212, + 0.012565163895487785, + 0.01620654948055744, + 0.018559999763965607, + 0.012548243626952171, + 0.025983726605772972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.018836896866559982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331218123435974 + ], + "opposite_logprob_tokens": [ + -3.7733120918273926 + ], + "logprob_reward": -0.02331218123435974, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331218123435974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32220733165740967, + 0.060400791466236115, + 0.053244203329086304, + 0.0658426582813263, + 0.07241584360599518, + 0.13882960379123688, + 0.0370306670665741, + 0.08143302798271179, + 0.04071585088968277, + 0.12787999212741852 + ], + "credit_map": [ + 0.06008755788207054, + 0.0413888655602932, + 0.07086087018251419, + 0.04853889346122742, + 0.027781564742326736, + 0.02051951363682747, + 0.01980324648320675, + 0.03322684392333031, + 0.006395657081156969, + 0.0067786891013383865, + 0.009173098020255566, + 0.006601749919354916, + 0.011610275134444237, + 0.00722010713070631, + 0.0060304030776023865, + 0.006590810138732195, + 0.006455703172832727, + 0.00714797992259264, + 0.008446072228252888, + 0.00420784205198288, + 0.010436639189720154, + 0.004446119535714388, + 0.006695471238344908, + 0.00540837412700057, + 0.00798057671636343, + 0.0054226284846663475, + 0.007086926139891148, + 0.008403749205172062, + 0.009670509025454521, + 0.008354535326361656, + 0.010067371651530266, + 0.008856362663209438, + 0.0062433755956590176, + 0.005751064512878656, + 0.007527414709329605, + 0.01063067652285099, + 0.010351673699915409, + 0.009611651301383972, + 0.010465272702276707, + 0.011834714561700821, + 0.014718078076839447, + 0.015502489171922207, + 0.018861781805753708, + 0.01514438632875681, + 0.021664921194314957, + 0.015421183779835701, + 0.015141110867261887, + 0.02237565815448761, + 0.004480880219489336, + 0.004975178744643927, + 0.00509144039824605, + 0.004469428211450577, + 0.006226277444511652, + 0.0037773442454636097, + 0.004084767308086157, + 0.003925347700715065, + 0.010563744232058525, + 0.01013027410954237, + 0.013344078324735165, + 0.008880090899765491, + 0.007982386276125908, + 0.008901698514819145, + 0.0062884013168513775, + 0.015342354774475098, + 0.004009260796010494, + 0.005109000485390425, + 0.007018683943897486, + 0.0028460135217756033, + 0.0034512935671955347, + 0.005246782209724188, + 0.002887249691411853, + 0.01014756504446268, + 0.015190157108008862, + 0.020908266305923462, + 0.015041333623230457, + 0.010297971777617931, + 0.01257694698870182, + 0.01992165297269821, + 0.011983470991253853, + 0.0219601858407259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.02331218123435974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070010140538216 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016070010140538216, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070010140538216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2689325213432312, + 0.14692482352256775, + 0.04658707603812218, + 0.09396252036094666, + 0.050277743488550186, + 0.09599519520998001, + 0.04627564549446106, + 0.07341818511486053, + 0.04231668636202812, + 0.13530957698822021 + ], + "credit_map": [ + 0.0315367616713047, + 0.04527757689356804, + 0.03533829376101494, + 0.03740602731704712, + 0.040275853127241135, + 0.02549724653363228, + 0.022078903391957283, + 0.031521834433078766, + 0.016072455793619156, + 0.020018283277750015, + 0.025007212534546852, + 0.02670835517346859, + 0.01854025572538376, + 0.014094645157456398, + 0.0120923463255167, + 0.014391264878213406, + 0.006454457063227892, + 0.0065491884015500546, + 0.0092721963301301, + 0.006253962870687246, + 0.006249045953154564, + 0.0038535804487764835, + 0.003978090826421976, + 0.003976551350206137, + 0.0109103349968791, + 0.010904389433562756, + 0.010776681825518608, + 0.011231091804802418, + 0.011318575590848923, + 0.017078381031751633, + 0.010236199013888836, + 0.011506853625178337, + 0.006189406383782625, + 0.005274051800370216, + 0.00557391531765461, + 0.004660633392632008, + 0.008896660059690475, + 0.008609757758677006, + 0.00567862531170249, + 0.005394688341766596, + 0.012876778841018677, + 0.010761011391878128, + 0.014961070381104946, + 0.012028202414512634, + 0.012369614094495773, + 0.010680611245334148, + 0.010600173845887184, + 0.011717729270458221, + 0.004878438543528318, + 0.00604264484718442, + 0.008094385266304016, + 0.004673726391047239, + 0.007781429216265678, + 0.005538711324334145, + 0.0036784119438380003, + 0.005587893072515726, + 0.010073677636682987, + 0.007531345821917057, + 0.012893086299300194, + 0.007866816595196724, + 0.00755720492452383, + 0.008844001218676567, + 0.005352512467652559, + 0.013299527578055859, + 0.006544549949467182, + 0.005899867974221706, + 0.005583152174949646, + 0.0031673340126872063, + 0.0044207279570400715, + 0.006600221153348684, + 0.003992692567408085, + 0.006108138244599104, + 0.01719404011964798, + 0.027367377653717995, + 0.01679462194442749, + 0.012004245072603226, + 0.01609942503273487, + 0.016712835058569908, + 0.01032788585871458, + 0.018809139728546143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.016070010140538216 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28131765127182007 + ], + "opposite_logprob_tokens": [ + -1.4063177108764648 + ], + "logprob_reward": -0.28131765127182007, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28131765127182007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16893579065799713, + 0.06747101247310638, + 0.04446738213300705, + 0.09967854619026184, + 0.05254316329956055, + 0.20477840304374695, + 0.07290465384721756, + 0.1777290254831314, + 0.02849166840314865, + 0.08300039917230606 + ], + "credit_map": [ + 0.03684602677822113, + 0.01708276942372322, + 0.05219157412648201, + 0.024497326463460922, + 0.011157023720443249, + 0.007854915224015713, + 0.009844617918133736, + 0.009461529552936554, + 0.006890549324452877, + 0.00698454026132822, + 0.01355371531099081, + 0.007976283319294453, + 0.0099115539342165, + 0.006470106076449156, + 0.008773085661232471, + 0.006911174394190311, + 0.005398621316999197, + 0.005977611057460308, + 0.006315873935818672, + 0.0035513746552169323, + 0.0096546346321702, + 0.0034032489638775587, + 0.005962991155683994, + 0.004203023388981819, + 0.012794895097613335, + 0.011208182200789452, + 0.008738246746361256, + 0.005888050422072411, + 0.01291106641292572, + 0.031606245785951614, + 0.009056986309587955, + 0.007474876940250397, + 0.006896007806062698, + 0.004719736985862255, + 0.005887059960514307, + 0.00700152013450861, + 0.00633652601391077, + 0.008776823058724403, + 0.005942859221249819, + 0.006982630118727684, + 0.015420366078615189, + 0.02573183923959732, + 0.020555293187499046, + 0.022511456161737442, + 0.0404309518635273, + 0.024870647117495537, + 0.03420782461762428, + 0.02105000428855419, + 0.009593851864337921, + 0.010465976782143116, + 0.008395864628255367, + 0.011128648184239864, + 0.010957680642604828, + 0.0072519113309681416, + 0.007078238762915134, + 0.00803247932344675, + 0.02376558445394039, + 0.02035556361079216, + 0.026129072532057762, + 0.019038746133446693, + 0.018267134204506874, + 0.02163832262158394, + 0.0161142535507679, + 0.03242034092545509, + 0.0037532446440309286, + 0.005670115817338228, + 0.0032366039231419563, + 0.0022935918532311916, + 0.003055310808122158, + 0.003950605168938637, + 0.0032378712203353643, + 0.003294323803856969, + 0.009831054136157036, + 0.014694025740027428, + 0.009246536530554295, + 0.00890738982707262, + 0.008727352134883404, + 0.008247646503150463, + 0.009620913304388523, + 0.013725482858717442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.28131765127182007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270187646150589 + ], + "opposite_logprob_tokens": [ + -2.127018690109253 + ], + "logprob_reward": -0.1270187646150589, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270187646150589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18317627906799316, + 0.03474994748830795, + 0.03664961829781532, + 0.045477453619241714, + 0.050195083022117615, + 0.10398915410041809, + 0.035567380487918854, + 0.165523499250412, + 0.052810799330472946, + 0.29186075925827026 + ], + "credit_map": [ + 0.01846635527908802, + 0.024454837664961815, + 0.04752233251929283, + 0.02373659797012806, + 0.02056763507425785, + 0.01754521019756794, + 0.01382053829729557, + 0.017062773928046227, + 0.003848829073831439, + 0.003956690896302462, + 0.006008490454405546, + 0.0029570134356617928, + 0.006095176097005606, + 0.0031394655816257, + 0.004179712850600481, + 0.004564575385302305, + 0.004283259157091379, + 0.0030502837616950274, + 0.005059100221842527, + 0.004487163387238979, + 0.006337299942970276, + 0.0025047920644283295, + 0.005765020381659269, + 0.005162704735994339, + 0.005308110732585192, + 0.005810702685266733, + 0.004471971187740564, + 0.004680075217038393, + 0.0051744673401117325, + 0.010084224864840508, + 0.005000084638595581, + 0.004947819747030735, + 0.006230256054550409, + 0.004738058894872665, + 0.00633515790104866, + 0.006125743966549635, + 0.009667927399277687, + 0.007924881763756275, + 0.0044237710535526276, + 0.00474928691983223, + 0.00953938439488411, + 0.008509218692779541, + 0.01123556774109602, + 0.012158192694187164, + 0.019458744674921036, + 0.013680205680429935, + 0.016168219968676567, + 0.013239631429314613, + 0.0035983084235340357, + 0.005507842171937227, + 0.005097711458802223, + 0.003305217018350959, + 0.006423210259526968, + 0.004403877537697554, + 0.003629941027611494, + 0.0036012742202728987, + 0.022664984688162804, + 0.012221047654747963, + 0.035283785313367844, + 0.012674665078520775, + 0.010341883637011051, + 0.021600190550088882, + 0.009364423342049122, + 0.04137253016233444, + 0.00631039310246706, + 0.010775650851428509, + 0.005317333620041609, + 0.004456279333680868, + 0.004998324904590845, + 0.005505033303052187, + 0.0051212129183113575, + 0.010326574556529522, + 0.036564670503139496, + 0.0653049647808075, + 0.035375796258449554, + 0.02494710311293602, + 0.03742433339357376, + 0.034696564078330994, + 0.021550308912992477, + 0.03599703311920166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.1270187646150589 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a highly detailed, cinematic 3D animation style, a medium shot, shot from a slightly elevated angle, slowly pans rightward around the center of the scene, focusing the viewer's attention on a young practitioner seated cross-legged at the center of a vast, ancient stone Bagua formation. His long, jet-black hair is tied high with a simple wooden hairpin, his face delicate and focused, his eyes closed, as if lost in deep meditation. He wears a dark blue silk robe, the hem and cuffs meticulously embroidered with intricate cloud and star patterns in silver thread, shimmering with a mysterious luster in the dim light. Each deeply engraved rune in the Bagua formation emanates a soft, luminous cyan light, casting its primary light upward, dramatically outlining his resolute jawline and the deep folds of his robe. His hands, resting on his chest, rapidly form complex hand seals at a speed barely perceptible to the naked eye. Each interlacing, spreading, and closing of his fingers is precise and powerful, imbued with a sense of ritual. As his hand gestures danced, the surrounding air visibly twisted and rippled. Countless luminous energy particles, as if drawn by an invisible force, converged from all directions, forming spirals of translucent air that coiled around his arms and body, accelerating in time with the rhythm of his hand seals. The backdrop was a deep, boundless night sky, dotted with stars. Distant nebulae shone in dreamy purples and blues, creating a mysterious atmosphere of intense focus, tranquility, and profound inner strength.", + "seed": 28484, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_0/index_118_seed_28484_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07618272816762328, + "vlm_logprob": -0.07618272816762328, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.07618272816762328, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007237288635224104 + ], + "opposite_logprob_tokens": [ + -5.632237434387207 + ], + "logprob_reward": -0.007237288635224104, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007237288635224104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1977793574333191, + 0.07740657776594162, + 0.0490972064435482, + 0.12524841725826263, + 0.1049245074391365, + 0.06412708014249802, + 0.09785973280668259, + 0.08719774335622787, + 0.09908106178045273, + 0.09727828204631805 + ], + "credit_map": [ + 0.025831397622823715, + 0.024575624614953995, + 0.029606152325868607, + 0.01796349138021469, + 0.023450369015336037, + 0.02493162639439106, + 0.019296593964099884, + 0.0321241058409214, + 0.012384878471493721, + 0.006787578575313091, + 0.008798636496067047, + 0.007788477931171656, + 0.009758509695529938, + 0.01176354568451643, + 0.010257115587592125, + 0.009867841377854347, + 0.006336730904877186, + 0.007067695260047913, + 0.006432756781578064, + 0.0067086792550981045, + 0.007334416266530752, + 0.003764393739402294, + 0.006247785408049822, + 0.0052047474309802055, + 0.032995276153087616, + 0.008605489507317543, + 0.012386428192257881, + 0.008113233372569084, + 0.010392063297331333, + 0.011569565162062645, + 0.006513493601232767, + 0.03467286005616188, + 0.01749405264854431, + 0.010243861936032772, + 0.0106094591319561, + 0.011552729643881321, + 0.014177327044308186, + 0.01821657456457615, + 0.010879293084144592, + 0.0117512047290802, + 0.007872973568737507, + 0.006272436119616032, + 0.006574857048690319, + 0.006545379292219877, + 0.008927048183977604, + 0.009702809154987335, + 0.008607584051787853, + 0.009623990394175053, + 0.013311104848980904, + 0.009450851008296013, + 0.013984189368784428, + 0.008195490576326847, + 0.017929604277014732, + 0.011841999366879463, + 0.010274713858962059, + 0.012871772982180119, + 0.009319428354501724, + 0.011919361539185047, + 0.010812461376190186, + 0.007912622764706612, + 0.012519953772425652, + 0.012020518071949482, + 0.010376756079494953, + 0.012316640466451645, + 0.014331741258502007, + 0.008472612127661705, + 0.014473751187324524, + 0.009888486936688423, + 0.011629901826381683, + 0.013522435910999775, + 0.012188870459794998, + 0.014573264867067337, + 0.017681041732430458, + 0.01254270039498806, + 0.01277029700577259, + 0.007150883320719004, + 0.010670030489563942, + 0.018055018037557602, + 0.00719441520050168, + 0.01121390238404274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.007237288635224104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091718450188637 + ], + "opposite_logprob_tokens": [ + -4.511091709136963 + ], + "logprob_reward": -0.011091718450188637, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091718450188637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20970864593982697, + 0.030139872804284096, + 0.03611883521080017, + 0.24455125629901886, + 0.23052562773227692, + 0.06636083871126175, + 0.04183458909392357, + 0.04750390723347664, + 0.04360327124595642, + 0.04965308681130409 + ], + "credit_map": [ + 0.02796078659594059, + 0.027674267068505287, + 0.03348856419324875, + 0.02399395778775215, + 0.02879909984767437, + 0.021610064432024956, + 0.023405861109495163, + 0.02277606911957264, + 0.004833335988223553, + 0.0027295330073684454, + 0.004743983969092369, + 0.0027327907737344503, + 0.005503321532160044, + 0.0028586252592504025, + 0.003631745930761099, + 0.0031065403018146753, + 0.005346471909433603, + 0.0030526777263730764, + 0.003619762137532234, + 0.003596831578761339, + 0.00978168286383152, + 0.002139579737558961, + 0.004134470131248236, + 0.004447362385690212, + 0.07795056700706482, + 0.015967831015586853, + 0.02021143212914467, + 0.0072324504144489765, + 0.005632119253277779, + 0.019685061648488045, + 0.006131747271865606, + 0.09174007177352905, + 0.05309143662452698, + 0.009775625541806221, + 0.020446302369236946, + 0.01731983572244644, + 0.01040363498032093, + 0.021566836163401604, + 0.014644645154476166, + 0.08327732980251312, + 0.00985323078930378, + 0.007422479800879955, + 0.0040736994706094265, + 0.004593633580952883, + 0.01581784151494503, + 0.008169658482074738, + 0.008931205607950687, + 0.0074990964494645596, + 0.006008198019117117, + 0.0050073303282260895, + 0.004484331235289574, + 0.003211172530427575, + 0.008313251659274101, + 0.0049685221165418625, + 0.0049070739187300205, + 0.004934710916131735, + 0.004122213460505009, + 0.005927405785769224, + 0.006218533962965012, + 0.004383621737360954, + 0.007965631783008575, + 0.006129485554993153, + 0.004340603947639465, + 0.008416416123509407, + 0.0070402598939836025, + 0.0043247458525002, + 0.005883281584829092, + 0.0033413530327379704, + 0.00690928939729929, + 0.005609091371297836, + 0.005148045253008604, + 0.005347206722944975, + 0.007596959359943867, + 0.005530037917196751, + 0.004890602547675371, + 0.0035892403684556484, + 0.008423221297562122, + 0.007706633303314447, + 0.00678494805470109, + 0.0051314448937773705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.011091718450188637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107061468064785 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.01107061468064785, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107061468064785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28302472829818726, + 0.08933821320533752, + 0.02270250767469406, + 0.2602003812789917, + 0.1112164631485939, + 0.043097611516714096, + 0.06044518202543259, + 0.05409327149391174, + 0.03623177856206894, + 0.03964981809258461 + ], + "credit_map": [ + 0.03875993192195892, + 0.027845559641718864, + 0.051999013870954514, + 0.025517389178276062, + 0.03572540730237961, + 0.0327603742480278, + 0.028482310473918915, + 0.04193473234772682, + 0.011976045556366444, + 0.009543942287564278, + 0.016944365575909615, + 0.007874459959566593, + 0.009943577460944653, + 0.016243748366832733, + 0.005592659581452608, + 0.011219417676329613, + 0.0026843200903385878, + 0.0020472905598580837, + 0.003972338512539864, + 0.0029964593704789877, + 0.004104332532733679, + 0.002028633141890168, + 0.0023408157285302877, + 0.0025283172726631165, + 0.08534583449363708, + 0.017195256426930428, + 0.02161601185798645, + 0.007331326138228178, + 0.006840650923550129, + 0.019227758049964905, + 0.005975265987217426, + 0.09666828066110611, + 0.012351682409644127, + 0.007377553731203079, + 0.010735186748206615, + 0.011304191313683987, + 0.012340371496975422, + 0.0198726374655962, + 0.01429018285125494, + 0.02294466271996498, + 0.005296789575368166, + 0.003936578519642353, + 0.003576667048037052, + 0.004954086616635323, + 0.0058976211585104465, + 0.006265668664127588, + 0.007497110869735479, + 0.005673088598996401, + 0.00836910493671894, + 0.004429529421031475, + 0.011581694707274437, + 0.0048749567940831184, + 0.006211426574736834, + 0.009053832851350307, + 0.004142921883612871, + 0.011781718581914902, + 0.005300234071910381, + 0.007032504305243492, + 0.008182003162801266, + 0.004347207956016064, + 0.008274240419268608, + 0.007086536381393671, + 0.0046170917339622974, + 0.009253458119928837, + 0.004020352382212877, + 0.004408795852214098, + 0.0033719539642333984, + 0.0039518470875918865, + 0.007243107073009014, + 0.0045687672682106495, + 0.004612554330378771, + 0.0040544006042182446, + 0.00532177509739995, + 0.004752561915665865, + 0.004868756048381329, + 0.0030720909126102924, + 0.005588841158896685, + 0.006021745502948761, + 0.004181989934295416, + 0.005842056591063738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.01107061468064785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134327828884125 + ], + "opposite_logprob_tokens": [ + -1.3134328126907349 + ], + "logprob_reward": -0.3134327828884125, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134327828884125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13936001062393188, + 0.03471466898918152, + 0.02591124176979065, + 0.2814607322216034, + 0.09291284531354904, + 0.1707179993391037, + 0.048914793878793716, + 0.0804925262928009, + 0.06441555172204971, + 0.06109967455267906 + ], + "credit_map": [ + 0.019844161346554756, + 0.014556746929883957, + 0.026964284479618073, + 0.01584807224571705, + 0.01640293188393116, + 0.012899904511868954, + 0.014738449826836586, + 0.018105443567037582, + 0.004744625184684992, + 0.003593328408896923, + 0.004715045914053917, + 0.003377920016646385, + 0.006282447837293148, + 0.003451775060966611, + 0.003908704500645399, + 0.0046408218331635, + 0.0034596947953104973, + 0.0024720728397369385, + 0.002465910278260708, + 0.0030789747834205627, + 0.005668980069458485, + 0.0029427483677864075, + 0.0030748294666409492, + 0.0027480321004986763, + 0.0909663513302803, + 0.018441317602992058, + 0.02152540534734726, + 0.009092324413359165, + 0.008947198279201984, + 0.020915305241942406, + 0.008200242184102535, + 0.10337258130311966, + 0.012150405906140804, + 0.007370217237621546, + 0.006937413010746241, + 0.010632573626935482, + 0.012416142039000988, + 0.010942920111119747, + 0.010840498842298985, + 0.0216226726770401, + 0.022165890783071518, + 0.0177194494754076, + 0.011920610442757607, + 0.016122527420520782, + 0.031501058489084244, + 0.021382413804531097, + 0.02964908629655838, + 0.020256957039237022, + 0.004548821598291397, + 0.005085047334432602, + 0.007281663361936808, + 0.0051474375650286674, + 0.008473705500364304, + 0.006024191156029701, + 0.0049952976405620575, + 0.007358628790825605, + 0.008876856416463852, + 0.010103259235620499, + 0.009083599783480167, + 0.008944179862737656, + 0.015049237757921219, + 0.010039282031357288, + 0.007500177249312401, + 0.010895935818552971, + 0.008030829019844532, + 0.007187282666563988, + 0.005797164048999548, + 0.0052422829903662205, + 0.01212527696043253, + 0.008053378202021122, + 0.011331685818731785, + 0.006647647824138403, + 0.007897008210420609, + 0.007177392952144146, + 0.00527936452999711, + 0.004413881804794073, + 0.01067451760172844, + 0.006791332270950079, + 0.010853666812181473, + 0.008012503385543823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.3134327828884125 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808123618364334 + ], + "opposite_logprob_tokens": [ + -3.288081169128418 + ], + "logprob_reward": -0.03808123618364334, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808123618364334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23174269497394562, + 0.037999436259269714, + 0.026262067258358, + 0.0746452659368515, + 0.27709805965423584, + 0.0521991066634655, + 0.037801455706357956, + 0.11926847696304321, + 0.07661530375480652, + 0.06636811047792435 + ], + "credit_map": [ + 0.032993730157613754, + 0.02873177081346512, + 0.03829767554998398, + 0.025289110839366913, + 0.029814889654517174, + 0.02718430571258068, + 0.024182913824915886, + 0.025248311460018158, + 0.005277928896248341, + 0.003890927881002426, + 0.005369007587432861, + 0.003901261370629072, + 0.007445759139955044, + 0.0035998427774757147, + 0.004468129947781563, + 0.004046577960252762, + 0.0032464847899973392, + 0.001978002954274416, + 0.00381285697221756, + 0.002957864897325635, + 0.004644350148737431, + 0.0025897969026118517, + 0.0031774593517184258, + 0.0038552512414753437, + 0.017636172473430634, + 0.0046731336042284966, + 0.006077499128878117, + 0.003986387979239225, + 0.0076807960867881775, + 0.007027412764728069, + 0.004446517676115036, + 0.023117348551750183, + 0.07286597788333893, + 0.009913991205394268, + 0.022627990692853928, + 0.011158310808241367, + 0.009780017659068108, + 0.026173992082476616, + 0.010880665853619576, + 0.1136971265077591, + 0.005808164365589619, + 0.006336137652397156, + 0.004419980105012655, + 0.003693632548674941, + 0.011984656564891338, + 0.007439234293997288, + 0.0068518915213644505, + 0.005665409378707409, + 0.003997348248958588, + 0.0036290022544562817, + 0.007082358002662659, + 0.0033877980895340443, + 0.005555220413953066, + 0.004572986625134945, + 0.004054048098623753, + 0.005522694438695908, + 0.01086140051484108, + 0.013241078704595566, + 0.016635479405522346, + 0.010163619183003902, + 0.02076147496700287, + 0.015858184546232224, + 0.011858740821480751, + 0.0198884978890419, + 0.009013433009386063, + 0.009688091464340687, + 0.005532979965209961, + 0.006086449604481459, + 0.017747923731803894, + 0.009220722131431103, + 0.009753993712365627, + 0.009571710601449013, + 0.009547648020088673, + 0.008416510187089443, + 0.005642218049615622, + 0.004542532842606306, + 0.012777384370565414, + 0.010637502186000347, + 0.008733098395168781, + 0.006071220617741346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.03808123618364334 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a highly detailed, cinematic 3D animation style, a medium shot, shot from a slightly elevated angle, slowly pans rightward around the center of the scene, focusing the viewer's attention on a young practitioner seated cross-legged at the center of a vast, ancient stone Bagua formation. His long, jet-black hair is tied high with a simple wooden hairpin, his face delicate and focused, his eyes closed, as if lost in deep meditation. He wears a dark blue silk robe, the hem and cuffs meticulously embroidered with intricate cloud and star patterns in silver thread, shimmering with a mysterious luster in the dim light. Each deeply engraved rune in the Bagua formation emanates a soft, luminous cyan light, casting its primary light upward, dramatically outlining his resolute jawline and the deep folds of his robe. His hands, resting on his chest, rapidly form complex hand seals at a speed barely perceptible to the naked eye. Each interlacing, spreading, and closing of his fingers is precise and powerful, imbued with a sense of ritual. As his hand gestures danced, the surrounding air visibly twisted and rippled. Countless luminous energy particles, as if drawn by an invisible force, converged from all directions, forming spirals of translucent air that coiled around his arms and body, accelerating in time with the rhythm of his hand seals. The backdrop was a deep, boundless night sky, dotted with stars. Distant nebulae shone in dreamy purples and blues, creating a mysterious atmosphere of intense focus, tranquility, and profound inner strength.", + "seed": 28485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_0/index_118_seed_28485_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10959338173270225, + "vlm_logprob": -0.10959338173270225, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.10959338173270225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008737546391785145 + ], + "opposite_logprob_tokens": [ + -5.508737564086914 + ], + "logprob_reward": -0.008737546391785145, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008737546391785145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25206899642944336, + 0.059838831424713135, + 0.05208781734108925, + 0.02275252901017666, + 0.024564532563090324, + 0.1014890968799591, + 0.24040724337100983, + 0.1563151627779007, + 0.05745839700102806, + 0.03301737830042839 + ], + "credit_map": [ + 0.0375312864780426, + 0.024431699886918068, + 0.02586400881409645, + 0.020906805992126465, + 0.038344018161296844, + 0.052122462540864944, + 0.026814572513103485, + 0.02605416625738144, + 0.010984877124428749, + 0.0064848363399505615, + 0.009540537372231483, + 0.005387422163039446, + 0.005715036764740944, + 0.010518930852413177, + 0.0050939228385686874, + 0.00611327076330781, + 0.00936105940490961, + 0.005572055000811815, + 0.010068327188491821, + 0.004803680349141359, + 0.0043863472528755665, + 0.007597778923809528, + 0.0036253458820283413, + 0.00667322613298893, + 0.004058142192661762, + 0.0028860564343631268, + 0.0028938069008290768, + 0.0026993483770638704, + 0.0021148959640413523, + 0.0035752393305301666, + 0.0019305429887026548, + 0.0025944989174604416, + 0.0034587651025503874, + 0.003104538656771183, + 0.0039673177525401115, + 0.0031180791556835175, + 0.002320217899978161, + 0.003684430615976453, + 0.0024370336905121803, + 0.002474150387570262, + 0.014757533557713032, + 0.010085763409733772, + 0.01610584184527397, + 0.010411716066300869, + 0.006459298077970743, + 0.010884491726756096, + 0.00979958102107048, + 0.022984877228736877, + 0.05938690900802612, + 0.02060479298233986, + 0.03196224197745323, + 0.019221285358071327, + 0.016648972406983376, + 0.05869685858488083, + 0.0162823349237442, + 0.01760387234389782, + 0.025825580582022667, + 0.014236394315958023, + 0.03312710300087929, + 0.012681112624704838, + 0.013505849987268448, + 0.031103815883398056, + 0.011609682813286781, + 0.014225632883608341, + 0.008776627480983734, + 0.005378319416195154, + 0.007389454636722803, + 0.007328948006033897, + 0.006666701752692461, + 0.00794392079114914, + 0.007258588448166847, + 0.006715843919664621, + 0.005403642076998949, + 0.0035705487243831158, + 0.004207868594676256, + 0.003526462474837899, + 0.005550498142838478, + 0.003929862752556801, + 0.003580627962946892, + 0.003247870597988367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.008737546391785145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085729002952576 + ], + "opposite_logprob_tokens": [ + -4.141085624694824 + ], + "logprob_reward": -0.016085729002952576, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085729002952576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18411053717136383, + 0.05958009138703346, + 0.029446743428707123, + 0.015872906893491745, + 0.03183624893426895, + 0.020325638353824615, + 0.49132412672042847, + 0.09412441402673721, + 0.04515090212225914, + 0.02822834998369217 + ], + "credit_map": [ + 0.02114923670887947, + 0.017995432019233704, + 0.022677874192595482, + 0.01731782592833042, + 0.026088068261742592, + 0.03639118745923042, + 0.022314907982945442, + 0.020176032558083534, + 0.008897383697330952, + 0.006086837500333786, + 0.007063074968755245, + 0.006450789514929056, + 0.006187932565808296, + 0.00794608611613512, + 0.009776501916348934, + 0.00717149069532752, + 0.005769038572907448, + 0.0028265141882002354, + 0.007084980141371489, + 0.002611463190987706, + 0.002121060388162732, + 0.0033323869574815035, + 0.0017625540494918823, + 0.003938749898225069, + 0.0026527794543653727, + 0.0018451546784490347, + 0.00127206661272794, + 0.002653024857863784, + 0.001768867252394557, + 0.002449752064421773, + 0.0019354047253727913, + 0.0012958594597876072, + 0.00531869288533926, + 0.002940966747701168, + 0.005292417947202921, + 0.00439299363642931, + 0.0025916409213095903, + 0.004149972926825285, + 0.003526938846334815, + 0.003622626420110464, + 0.0027945430483669043, + 0.002654863055795431, + 0.00316763948649168, + 0.0027859313413500786, + 0.002507551107555628, + 0.0019178144866600633, + 0.002186867641285062, + 0.002310429932549596, + 0.10851818323135376, + 0.04742765054106712, + 0.07392265647649765, + 0.0409010611474514, + 0.03540987893939018, + 0.11488074064254761, + 0.03582817316055298, + 0.03443581238389015, + 0.01601560227572918, + 0.0072911824099719524, + 0.016477536410093307, + 0.007200534921139479, + 0.010106596164405346, + 0.01669738069176674, + 0.010405410081148148, + 0.00993018876761198, + 0.006440954748541117, + 0.004442297387868166, + 0.005269370973110199, + 0.006491718348115683, + 0.00548735773190856, + 0.006730143912136555, + 0.0063407025299966335, + 0.003948358818888664, + 0.004516679793596268, + 0.004110051319003105, + 0.003599827876314521, + 0.0024445555172860622, + 0.00469930749386549, + 0.00425729202106595, + 0.0023095528595149517, + 0.002291086595505476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.016085729002952576 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192714355885983 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192714355885983, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192714355885983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16418735682964325, + 0.12168736010789871, + 0.06667839735746384, + 0.03910019248723984, + 0.0738503485918045, + 0.06257924437522888, + 0.2057468742132187, + 0.16881389915943146, + 0.06280820816755295, + 0.03454810008406639 + ], + "credit_map": [ + 0.025381561368703842, + 0.0152704156935215, + 0.024198222905397415, + 0.016135061159729958, + 0.020495226606726646, + 0.02911585569381714, + 0.013548932038247585, + 0.020042087882757187, + 0.026797538623213768, + 0.010161267593502998, + 0.01920655369758606, + 0.008901900611817837, + 0.008579074405133724, + 0.027534356340765953, + 0.007911612279713154, + 0.012595057487487793, + 0.011872011236846447, + 0.006901968270540237, + 0.01729395054280758, + 0.005916770547628403, + 0.00511851767078042, + 0.006834280211478472, + 0.0038916508201509714, + 0.008849251084029675, + 0.005893043242394924, + 0.004094150848686695, + 0.004238169174641371, + 0.00440393527969718, + 0.00431570690125227, + 0.006743995472788811, + 0.003887394443154335, + 0.005523796658962965, + 0.010667447000741959, + 0.008996064774692059, + 0.011701826006174088, + 0.008594789542257786, + 0.007290416397154331, + 0.009996111504733562, + 0.007821776904165745, + 0.00878191739320755, + 0.009039178490638733, + 0.005523269530385733, + 0.012878148816525936, + 0.007428842131048441, + 0.004638939164578915, + 0.008183179423213005, + 0.006246351171284914, + 0.008641333319246769, + 0.049565549939870834, + 0.017828717827796936, + 0.027613839134573936, + 0.01555558666586876, + 0.013354380615055561, + 0.05455484986305237, + 0.012534049339592457, + 0.01473989337682724, + 0.02898448146879673, + 0.013461182825267315, + 0.029492301866412163, + 0.013042953796684742, + 0.0187554731965065, + 0.029978374019265175, + 0.01937960460782051, + 0.015719516202807426, + 0.009053298272192478, + 0.006065091118216515, + 0.0074830432422459126, + 0.009330742061138153, + 0.007728977128863335, + 0.009221900254487991, + 0.008668346330523491, + 0.005256804171949625, + 0.004341490566730499, + 0.004847932606935501, + 0.005121381022036076, + 0.003568728920072317, + 0.006162323988974094, + 0.004019274841994047, + 0.003314564237371087, + 0.0031724043656140566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.014192714355885983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486327826976776 + ], + "opposite_logprob_tokens": [ + -1.2236328125 + ], + "logprob_reward": -0.3486327826976776, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486327826976776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2940230667591095, + 0.10614077001810074, + 0.03157356381416321, + 0.025170717388391495, + 0.10809957981109619, + 0.057239409536123276, + 0.12070594727993011, + 0.18170469999313354, + 0.04190277308225632, + 0.03343946859240532 + ], + "credit_map": [ + 0.02851407416164875, + 0.02416851557791233, + 0.020608829334378242, + 0.03242380917072296, + 0.03372977301478386, + 0.05369966849684715, + 0.06654336303472519, + 0.034335024654865265, + 0.01519029587507248, + 0.011600660160183907, + 0.013703524135053158, + 0.013130954466760159, + 0.010682575404644012, + 0.010971235111355782, + 0.01798725314438343, + 0.012874272651970387, + 0.0038792656268924475, + 0.004686927422881126, + 0.004122742917388678, + 0.004191989079117775, + 0.0041801719926297665, + 0.0039060746785253286, + 0.003688100026920438, + 0.0029182902071624994, + 0.0035922834649682045, + 0.0028716858942061663, + 0.003074235050007701, + 0.004106074571609497, + 0.0031264268327504396, + 0.0029338437598198652, + 0.002878655446693301, + 0.0025875107385218143, + 0.01562386192381382, + 0.012410464696586132, + 0.018188076093792915, + 0.013606403023004532, + 0.010923974215984344, + 0.01372483465820551, + 0.01166580431163311, + 0.01195615530014038, + 0.009215944446623325, + 0.005679996684193611, + 0.008024532347917557, + 0.0075467354618012905, + 0.006772938650101423, + 0.004823992494493723, + 0.008492590859532356, + 0.006682679522782564, + 0.02604096382856369, + 0.010972169227898121, + 0.020101744681596756, + 0.009793746285140514, + 0.009225844405591488, + 0.026970256119966507, + 0.008824014104902744, + 0.008777215145528316, + 0.028838850557804108, + 0.01613549515604973, + 0.043714720755815506, + 0.015009680762887001, + 0.013922282494604588, + 0.03338668867945671, + 0.017222167924046516, + 0.01347482018172741, + 0.0058266217820346355, + 0.0038723722100257874, + 0.005579949356615543, + 0.005560311488807201, + 0.005134955048561096, + 0.005852833390235901, + 0.0069324905052781105, + 0.003143239300698042, + 0.006517203990370035, + 0.004957115277647972, + 0.0036356861237436533, + 0.003918264526873827, + 0.004258633125573397, + 0.004683953244239092, + 0.0029883128590881824, + 0.0024803001433610916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.3486327826976776 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of a young practitioner seated cross-legged in the center of a vast ancient stone Bagua formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner wearing a dark blue silk robe with silver cloud and star embroidery, and does his jet-black hair feature a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the engraved runes in the Bagua formation emit a soft luminous cyan light that casts upward illumination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the practitioner perform rapid, complex hand seals on his chest while translucent energy particles spiral around his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly elevated angle and slowly panning rightward around the center of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16031813621520996 + ], + "opposite_logprob_tokens": [ + -1.91031813621521 + ], + "logprob_reward": -0.16031813621520996, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16031813621520996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15248771011829376, + 0.09236393123865128, + 0.04078724607825279, + 0.03514818474650383, + 0.05025612190365791, + 0.1371953785419464, + 0.08352609723806381, + 0.2735553979873657, + 0.09572313725948334, + 0.038956791162490845 + ], + "credit_map": [ + 0.020368333905935287, + 0.016218308359384537, + 0.022344214841723442, + 0.016155695542693138, + 0.025783585384488106, + 0.02270868606865406, + 0.012515136040747166, + 0.01639375276863575, + 0.017511652782559395, + 0.012590072117745876, + 0.008847041986882687, + 0.009058230556547642, + 0.011294414289295673, + 0.015294915996491909, + 0.009086835198104382, + 0.00868077203631401, + 0.008295251987874508, + 0.004797500558197498, + 0.008292571641504765, + 0.00302260834723711, + 0.004012954421341419, + 0.005156326107680798, + 0.003041736548766494, + 0.004168294835835695, + 0.006672235205769539, + 0.003946244250983, + 0.0045138513669371605, + 0.0037272535264492035, + 0.003850361332297325, + 0.0055862655863165855, + 0.0032520140521228313, + 0.0035999591927975416, + 0.007465834263712168, + 0.007375424262136221, + 0.006747799925506115, + 0.005876199342310429, + 0.006008201744407415, + 0.0061225551180541515, + 0.0055863093584775925, + 0.005073797423392534, + 0.02133900858461857, + 0.011387541890144348, + 0.02014409936964512, + 0.01119353249669075, + 0.00992704089730978, + 0.014667704701423645, + 0.01770607940852642, + 0.030830373987555504, + 0.012810821644961834, + 0.00711140688508749, + 0.021631471812725067, + 0.006593623664230108, + 0.006212986074388027, + 0.011618351563811302, + 0.00942021980881691, + 0.008127219043672085, + 0.0458923801779747, + 0.02233375422656536, + 0.0507463738322258, + 0.022780045866966248, + 0.028904348611831665, + 0.0493769645690918, + 0.029939424246549606, + 0.0235820971429348, + 0.014266270212829113, + 0.010179609060287476, + 0.011652029119431973, + 0.011198940686881542, + 0.012474575079977512, + 0.014940408989787102, + 0.012344676069915295, + 0.008666622452437878, + 0.004152361303567886, + 0.006209525279700756, + 0.005289728753268719, + 0.003656302811577916, + 0.007620217744261026, + 0.005051410291343927, + 0.003614790504798293, + 0.003362454939633608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.16031813621520996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_tensor([6972, 6972])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_tensor([6972, 6972])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..480df1dbf068a18f92e4971ace0e7c1caeaf1cd8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_tensor([6972, 6972])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows two enormous butterflies as they soar at breakneck speed across a lush meadow bathed in intense, high-contrast light and shadow. In this stark black and white world, the leading butterfly's wings, nearly snow-white, punctuated only by delicate black veins, appear fragile and translucent in the sunlight. The pursuer, in stark contrast, possesses velvety black wings emblazoned with intricate white geometric patterns, each flap imbued with power and determination. The camera tracks their flight paths, capturing the former's desperate evasion with a swift and erratic path, while the latter pursues with a more direct and aggressive approach. Each powerful downward thrust and graceful upward stroke of their massive wings is captured with breathtaking clarity, even the subtle vibrations at the wing edges and the minute air currents they create, violently swaying the knee-high grass and unidentified wildflowers below. The strong light casts a huge, fast-moving, ever-changing shadow of a butterfly on the grass, which, combined with the sharp shadows of the grass leaves themselves, creates a dynamic, dreamy, and tense surreal picture.", + "seed": 16231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_6972_seed_16231_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3451941512525082, + "vlm_logprob": -0.3451941512525082, + "vlm_yes_no_margin": 1.775, + "vlm_token_credit": -0.3451941512525082, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6356915831565857 + ], + "opposite_logprob_tokens": [ + -0.7606915831565857 + ], + "logprob_reward": -0.6356915831565857, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6356915831565857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3155903220176697, + 0.1798247992992401, + 0.03595645725727081, + 0.04536525160074234, + 0.13317368924617767, + 0.01250661164522171, + 0.013060903176665306, + 0.07821281999349594, + 0.046734899282455444, + 0.13957421481609344 + ], + "credit_map": [ + 0.030598213896155357, + 0.04336174204945564, + 0.0820264145731926, + 0.03451357036828995, + 0.025277649983763695, + 0.029953733086586, + 0.0462711863219738, + 0.023587841540575027, + 0.014157051220536232, + 0.06245879456400871, + 0.028886154294013977, + 0.03848806396126747, + 0.008865314535796642, + 0.00780969625338912, + 0.009475820697844028, + 0.009683901444077492, + 0.004147155210375786, + 0.003996505867689848, + 0.0038305302150547504, + 0.007297448348253965, + 0.0068007539957761765, + 0.0027004736475646496, + 0.0037830020301043987, + 0.0034005881752818823, + 0.004793648608028889, + 0.008974282070994377, + 0.006549415178596973, + 0.008418551646173, + 0.004422134254127741, + 0.005495226476341486, + 0.0038928715512156487, + 0.0028191213496029377, + 0.008316299878060818, + 0.009774668142199516, + 0.03741303086280823, + 0.007010180037468672, + 0.041410110890865326, + 0.014267150312662125, + 0.007420634850859642, + 0.00756161380559206, + 0.0021484135650098324, + 0.0017540673725306988, + 0.0013837015721946955, + 0.001122111571021378, + 0.0016809910302981734, + 0.0018388996832072735, + 0.0014299229951575398, + 0.001148503739386797, + 0.0016698455438017845, + 0.0019843389745801687, + 0.0017129960469901562, + 0.0016600501257926226, + 0.0021470757201313972, + 0.0016667525051161647, + 0.0011701059993356466, + 0.001049738610163331, + 0.008303260430693626, + 0.005553732626140118, + 0.019487302750349045, + 0.013051237910985947, + 0.005247795023024082, + 0.0072194053791463375, + 0.009432714432477951, + 0.00991736724972725, + 0.016944320872426033, + 0.011656208895146847, + 0.0026454583276063204, + 0.002548767952248454, + 0.0026635974645614624, + 0.002259398577734828, + 0.0026367814280092716, + 0.005380365997552872, + 0.008164703845977783, + 0.05085734650492668, + 0.0052162958309054375, + 0.006793309934437275, + 0.047425463795661926, + 0.007582162506878376, + 0.007491567637771368, + 0.006043350324034691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.6356915831565857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983269929885864 + ], + "opposite_logprob_tokens": [ + -0.5233269929885864 + ], + "logprob_reward": -0.8983269929885864, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983269929885864, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.42276832461357117, + 0.2897094190120697, + 0.03641781955957413, + 0.04995284602046013, + 0.02299484610557556, + 0.010151523165404797, + 0.01425950974225998, + 0.05192948877811432, + 0.041630908846855164, + 0.06018523499369621 + ], + "credit_map": [ + 0.039972610771656036, + 0.04413511976599693, + 0.1303287297487259, + 0.02943611703813076, + 0.025384962558746338, + 0.03851667046546936, + 0.09081097692251205, + 0.024183139204978943, + 0.020640011876821518, + 0.10431776195764542, + 0.10937582701444626, + 0.01452269684523344, + 0.012519368901848793, + 0.010468922555446625, + 0.011146845296025276, + 0.0067180064506828785, + 0.003552256152033806, + 0.0037097411695867777, + 0.0038330245297402143, + 0.005300304852426052, + 0.009563867934048176, + 0.003850281471386552, + 0.0036214550491422415, + 0.002986888401210308, + 0.00518215075135231, + 0.010554678738117218, + 0.006961603183299303, + 0.009508280083537102, + 0.004736431874334812, + 0.005691213998943567, + 0.00438139820471406, + 0.0029370898846536875, + 0.0021931896917521954, + 0.0029569377657026052, + 0.003260769648477435, + 0.003443089546635747, + 0.004732986446470022, + 0.0022111546713858843, + 0.002414627466350794, + 0.001782091916538775, + 0.0018056557746604085, + 0.0012924503535032272, + 0.0012843137374147773, + 0.0011465460993349552, + 0.0013220352120697498, + 0.0010968727292492986, + 0.0013773838290944695, + 0.0008262647315859795, + 0.0018013054504990578, + 0.0018844070145860314, + 0.001886621699668467, + 0.0015937138814479113, + 0.0023927725851535797, + 0.002343533094972372, + 0.0013425707584246993, + 0.0010145845590159297, + 0.006272762082517147, + 0.00467846542596817, + 0.011560337617993355, + 0.00795553345233202, + 0.004293219652026892, + 0.006245214957743883, + 0.005155725404620171, + 0.00576823391020298, + 0.012700224295258522, + 0.010871564038097858, + 0.003942260053008795, + 0.0023050736635923386, + 0.0023457480128854513, + 0.0036920453421771526, + 0.00207813223823905, + 0.003695863066241145, + 0.0036671003326773643, + 0.015070809051394463, + 0.005266801919788122, + 0.007220026571303606, + 0.01555565930902958, + 0.0042924475856125355, + 0.0036616409197449684, + 0.005450748372823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.8983269929885864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07907712459564209 + ], + "opposite_logprob_tokens": [ + -2.5790772438049316 + ], + "logprob_reward": -0.07907712459564209, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07907712459564209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3565008044242859, + 0.37173137068748474, + 0.024861672893166542, + 0.03981717303395271, + 0.04265287145972252, + 0.013286914676427841, + 0.011366532184183598, + 0.03892887011170387, + 0.05786893516778946, + 0.042984865605831146 + ], + "credit_map": [ + 0.0333138071000576, + 0.050329189747571945, + 0.0960216298699379, + 0.03608952835202217, + 0.02853507176041603, + 0.03457026556134224, + 0.05251568183302879, + 0.025125637650489807, + 0.026850398629903793, + 0.1373540163040161, + 0.1396520584821701, + 0.018870461732149124, + 0.014881918206810951, + 0.012133440002799034, + 0.013547743670642376, + 0.008441325277090073, + 0.0021954355761408806, + 0.0030504027381539345, + 0.0025036188308149576, + 0.0031785934697836637, + 0.006706938147544861, + 0.002091585658490658, + 0.002769782207906246, + 0.0023653169628232718, + 0.00433188583701849, + 0.006169552449136972, + 0.005338000133633614, + 0.008483903482556343, + 0.0039035617373883724, + 0.005042966455221176, + 0.003848076332360506, + 0.002699228236451745, + 0.0035727820359170437, + 0.004529484082013369, + 0.007109787315130234, + 0.004415408242493868, + 0.013249429874122143, + 0.005173200741410255, + 0.002504921518266201, + 0.0020978585816919804, + 0.0027613588608801365, + 0.0018379100365564227, + 0.0017590352799743414, + 0.0014324570074677467, + 0.0015288692666217685, + 0.0017799048218876123, + 0.0012304564006626606, + 0.0009569234098307788, + 0.0015041754813864827, + 0.0018688640557229519, + 0.001594795729033649, + 0.0013920983765274286, + 0.0017256123246625066, + 0.001544569036923349, + 0.0009622281650081277, + 0.0007741886074654758, + 0.004718709737062454, + 0.002142679411917925, + 0.011025427840650082, + 0.006265643052756786, + 0.0026146303862333298, + 0.0031647130381315947, + 0.0044747525826096535, + 0.004522311966866255, + 0.021406816318631172, + 0.012893641367554665, + 0.0038631863426417112, + 0.0022618777584284544, + 0.0038450707215815783, + 0.0028748083859682083, + 0.0033507468178868294, + 0.007372787222266197, + 0.003243581857532263, + 0.011530401185154915, + 0.003745543770492077, + 0.004670264665037394, + 0.008900577202439308, + 0.0040177274495363235, + 0.0029883491806685925, + 0.003888418199494481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.07907712459564209 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551363557577133 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551363557577133, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551363557577133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32503747940063477, + 0.35648205876350403, + 0.040617138147354126, + 0.03990340605378151, + 0.025324208661913872, + 0.011596345342695713, + 0.014910894446074963, + 0.04969032108783722, + 0.05729154869914055, + 0.07914664596319199 + ], + "credit_map": [ + 0.03175782784819603, + 0.033046919852495193, + 0.09909851849079132, + 0.021851038560271263, + 0.02030632644891739, + 0.029713064432144165, + 0.07005622237920761, + 0.019207503646612167, + 0.02453547716140747, + 0.1254143863916397, + 0.14084656536579132, + 0.02061435393989086, + 0.014101524837315083, + 0.010880472138524055, + 0.012477019801735878, + 0.007612239569425583, + 0.004684211686253548, + 0.0044614402577281, + 0.0036350656300783157, + 0.004678430501371622, + 0.01088083628565073, + 0.00397859700024128, + 0.004692792892456055, + 0.003605758072808385, + 0.004482583608478308, + 0.007725091185420752, + 0.005527929402887821, + 0.006945021916180849, + 0.003606752958148718, + 0.004538875073194504, + 0.004375115502625704, + 0.0027020308189094067, + 0.0026435309555381536, + 0.004988825414329767, + 0.0033105171751230955, + 0.005042777396738529, + 0.0031065926887094975, + 0.0019953318405896425, + 0.002357647754251957, + 0.0018789832247421145, + 0.0017169173806905746, + 0.001298805931583047, + 0.001505572465248406, + 0.0013036875752732158, + 0.0018921777373179793, + 0.0014918857486918569, + 0.0014669244410470128, + 0.0009203730733133852, + 0.001848906627856195, + 0.0019111536676064134, + 0.0023660294245928526, + 0.0019906118977814913, + 0.0021014909725636244, + 0.0020561902783811092, + 0.0015310589224100113, + 0.0011054511414840817, + 0.00622278917580843, + 0.003700185799971223, + 0.012425391003489494, + 0.00796602200716734, + 0.004405450075864792, + 0.004175299312919378, + 0.005656299646943808, + 0.005138877779245377, + 0.01980707049369812, + 0.013237904757261276, + 0.003966306336224079, + 0.0033882963471114635, + 0.003535635070875287, + 0.0029735600110143423, + 0.0038148851599544287, + 0.006567883770912886, + 0.00476181972771883, + 0.022699635475873947, + 0.0060496581718325615, + 0.008398009464144707, + 0.021466780453920364, + 0.004764897748827934, + 0.00456438260152936, + 0.0064414492808282375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.012551363557577133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032369196414948 + ], + "opposite_logprob_tokens": [ + -2.3503236770629883 + ], + "logprob_reward": -0.10032369196414948, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032369196414948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46509668231010437, + 0.24503035843372345, + 0.021855346858501434, + 0.03846615180373192, + 0.03266677260398865, + 0.009267613291740417, + 0.018944190815091133, + 0.05453420802950859, + 0.016273928806185722, + 0.09786473959684372 + ], + "credit_map": [ + 0.04393842816352844, + 0.0509059801697731, + 0.14207884669303894, + 0.03420320153236389, + 0.029067272320389748, + 0.042260121554136276, + 0.09620227664709091, + 0.026440542191267014, + 0.017407018691301346, + 0.08905931562185287, + 0.09036222845315933, + 0.01264452189207077, + 0.010682513937354088, + 0.008838583715260029, + 0.009553489275276661, + 0.006482680793851614, + 0.00262093055061996, + 0.002180067589506507, + 0.002139393938705325, + 0.0029136326629668474, + 0.005124651361256838, + 0.001713054021820426, + 0.002779342932626605, + 0.002384273800998926, + 0.003983007278293371, + 0.008000126108527184, + 0.005253077019006014, + 0.0069716149009764194, + 0.003954456653445959, + 0.004363974556326866, + 0.0035049139987677336, + 0.002434981521219015, + 0.0021952760871499777, + 0.0029153863433748484, + 0.006889212876558304, + 0.0030253666918724775, + 0.009069650433957577, + 0.0034430879168212414, + 0.0032720589078962803, + 0.0018567362567409873, + 0.0008430646266788244, + 0.0011112511856481433, + 0.0010047147516161203, + 0.0007836812874302268, + 0.0021590166725218296, + 0.0009625927195884287, + 0.00147747911978513, + 0.000925813103094697, + 0.0024045556783676147, + 0.0022607212886214256, + 0.002557443454861641, + 0.002125850412994623, + 0.003090617246925831, + 0.0031424397602677345, + 0.0018785606371238828, + 0.0014840016374364495, + 0.006012235768139362, + 0.004614005796611309, + 0.01269804872572422, + 0.008643366396427155, + 0.0039918613620102406, + 0.006147853564471006, + 0.00596701679751277, + 0.0064598252065479755, + 0.003609581384807825, + 0.0022066370584070683, + 0.0020891763269901276, + 0.0015482730232179165, + 0.0013430803082883358, + 0.0022785982582718134, + 0.0015263429377228022, + 0.0016722393920645118, + 0.005268894135951996, + 0.03528900817036629, + 0.0036915000528097153, + 0.004963353276252747, + 0.03537958115339279, + 0.004283606074750423, + 0.0038614734075963497, + 0.005127323791384697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.10032369196414948 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows two enormous butterflies as they soar at breakneck speed across a lush meadow bathed in intense, high-contrast light and shadow. In this stark black and white world, the leading butterfly's wings, nearly snow-white, punctuated only by delicate black veins, appear fragile and translucent in the sunlight. The pursuer, in stark contrast, possesses velvety black wings emblazoned with intricate white geometric patterns, each flap imbued with power and determination. The camera tracks their flight paths, capturing the former's desperate evasion with a swift and erratic path, while the latter pursues with a more direct and aggressive approach. Each powerful downward thrust and graceful upward stroke of their massive wings is captured with breathtaking clarity, even the subtle vibrations at the wing edges and the minute air currents they create, violently swaying the knee-high grass and unidentified wildflowers below. The strong light casts a huge, fast-moving, ever-changing shadow of a butterfly on the grass, which, combined with the sharp shadows of the grass leaves themselves, creates a dynamic, dreamy, and tense surreal picture.", + "seed": 16232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_6972_seed_16232_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09558988101780415, + "vlm_logprob": -0.09558988101780415, + "vlm_yes_no_margin": 3.275000011920929, + "vlm_token_credit": -0.09558988101780415, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0278935506939888 + ], + "opposite_logprob_tokens": [ + -3.777893543243408 + ], + "logprob_reward": -0.0278935506939888, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0278935506939888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2815660238265991, + 0.08756093680858612, + 0.017974913120269775, + 0.01899668201804161, + 0.4013001024723053, + 0.016274958848953247, + 0.08237773925065994, + 0.027687499299645424, + 0.04068432003259659, + 0.025576738640666008 + ], + "credit_map": [ + 0.0178860891610384, + 0.015290064737200737, + 0.02826603874564171, + 0.01724386215209961, + 0.01972617208957672, + 0.09196288883686066, + 0.028661496937274933, + 0.06252940744161606, + 0.008105730637907982, + 0.021024290472269058, + 0.009191736578941345, + 0.01609932817518711, + 0.009878179989755154, + 0.007737102918326855, + 0.00855829194188118, + 0.006966277025640011, + 0.0016222442500293255, + 0.002224183874204755, + 0.0024323859252035618, + 0.0018687315750867128, + 0.0024992499966174364, + 0.002287463750690222, + 0.002306881360709667, + 0.002733772387728095, + 0.0030347832944244146, + 0.001818066113628447, + 0.0022732981014996767, + 0.0019585108384490013, + 0.0032892345916479826, + 0.002914935816079378, + 0.0018623743671923876, + 0.0018454798264428973, + 0.1280428022146225, + 0.017541874200105667, + 0.07380255311727524, + 0.01586664468050003, + 0.015886403620243073, + 0.09248185157775879, + 0.014843137934803963, + 0.04283482953906059, + 0.003145191352814436, + 0.0013074732851237059, + 0.0014756455784663558, + 0.0013642216799780726, + 0.0015512736281380057, + 0.00469482596963644, + 0.0017081870464608073, + 0.0010281401919201016, + 0.009380319155752659, + 0.015496722422540188, + 0.012640822678804398, + 0.013000245206058025, + 0.007548497058451176, + 0.0083584263920784, + 0.009000714868307114, + 0.006951994728296995, + 0.003858306212350726, + 0.002938787452876568, + 0.003772031981498003, + 0.0035615211818367243, + 0.002926052315160632, + 0.003827973734587431, + 0.0036501765716820955, + 0.003152650548145175, + 0.0045134215615689754, + 0.006031329277902842, + 0.002774167340248823, + 0.006465120241045952, + 0.005216163117438555, + 0.006205184850841761, + 0.0034221354871988297, + 0.0060567958280444145, + 0.003222488099709153, + 0.003210692200809717, + 0.005357176065444946, + 0.0021149679087102413, + 0.0033151968382298946, + 0.0028615056071430445, + 0.0015431586652994156, + 0.003951555117964745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0278935506939888 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38708847761154175 + ], + "opposite_logprob_tokens": [ + -1.1370885372161865 + ], + "logprob_reward": -0.38708847761154175, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38708847761154175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5368731617927551, + 0.053988534957170486, + 0.012446138076484203, + 0.015616551041603088, + 0.2526718080043793, + 0.04079942777752876, + 0.021909674629569054, + 0.01055259071290493, + 0.04025500640273094, + 0.014887074008584023 + ], + "credit_map": [ + 0.041195809841156006, + 0.023606328293681145, + 0.034485794603824615, + 0.017127711325883865, + 0.030526546761393547, + 0.18911121785640717, + 0.05492899566888809, + 0.14589077234268188, + 0.004302592948079109, + 0.007792171090841293, + 0.006027906201779842, + 0.010520891286432743, + 0.005640207324177027, + 0.005666011944413185, + 0.007568431552499533, + 0.006470328662544489, + 0.0010344141628593206, + 0.001904757460579276, + 0.0018613151041790843, + 0.001272212015464902, + 0.0018575509311631322, + 0.0015131570398807526, + 0.0013761272421106696, + 0.0016266056336462498, + 0.001997095998376608, + 0.0019378734286874533, + 0.0023921329993754625, + 0.0019702804274857044, + 0.0016133198514580727, + 0.0028050129767507315, + 0.001334687927737832, + 0.0015661503421142697, + 0.0843484178185463, + 0.010365953668951988, + 0.03500884026288986, + 0.010545219294726849, + 0.008198860101401806, + 0.06895844638347626, + 0.00892594177275896, + 0.026320137083530426, + 0.006166055332869291, + 0.004756396170705557, + 0.005151471123099327, + 0.004830868449062109, + 0.005302703473716974, + 0.005295857321470976, + 0.005577465053647757, + 0.003718616673722863, + 0.0033972538076341152, + 0.003181006759405136, + 0.0037382205482572317, + 0.002685394138097763, + 0.002101725433021784, + 0.0017092974158003926, + 0.0017494783969596028, + 0.003347300225868821, + 0.0020552093628793955, + 0.0013044566148892045, + 0.0009496324346400797, + 0.0009023499442264438, + 0.0013772208476439118, + 0.0016400731401517987, + 0.0008647064678370953, + 0.0014589432394132018, + 0.005238512996584177, + 0.005223880056291819, + 0.004477465525269508, + 0.0037519014440476894, + 0.005770925432443619, + 0.007780243176966906, + 0.002924238331615925, + 0.00508784269914031, + 0.002420994220301509, + 0.0022247033193707466, + 0.0028236559592187405, + 0.0014949581818655133, + 0.0017763131763786077, + 0.0014936532825231552, + 0.0011259594466537237, + 0.0015268366551026702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.38708847761154175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820182055234909 + ], + "opposite_logprob_tokens": [ + -4.01820182800293 + ], + "logprob_reward": -0.01820182055234909, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820182055234909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7583655118942261, + 0.022287171334028244, + 0.009473655372858047, + 0.01152059156447649, + 0.11644081771373749, + 0.015721507370471954, + 0.025113604962825775, + 0.015118672512471676, + 0.0106442179530859, + 0.01531428936868906 + ], + "credit_map": [ + 0.052724987268447876, + 0.02841799706220627, + 0.05051388591527939, + 0.02652222104370594, + 0.04344945773482323, + 0.26978087425231934, + 0.0774942934513092, + 0.20946180820465088, + 0.002146890154108405, + 0.0026562444400042295, + 0.002014327561482787, + 0.0033082591835409403, + 0.003274201648309827, + 0.002222029259428382, + 0.004287612624466419, + 0.0023776073940098286, + 0.0007962355157360435, + 0.0012597910827025771, + 0.0013853871496394277, + 0.0009328363812528551, + 0.0013588166330009699, + 0.0012656582985073328, + 0.001258760690689087, + 0.0012161695631220937, + 0.0012420759303495288, + 0.001288775121793151, + 0.0019269586773589253, + 0.001387517317198217, + 0.0012539782328531146, + 0.0021484685130417347, + 0.0013321838341653347, + 0.000940633995924145, + 0.03765539824962616, + 0.005121197085827589, + 0.022623341530561447, + 0.004207413177937269, + 0.003850799286738038, + 0.02669358253479004, + 0.004156252834945917, + 0.012132829055190086, + 0.003028788836672902, + 0.0017803239170461893, + 0.0016302423318848014, + 0.0016248043393716216, + 0.001950389938428998, + 0.0030512420926243067, + 0.0014870185405015945, + 0.0011686979560181499, + 0.005723211448639631, + 0.003464349079877138, + 0.003427047049626708, + 0.0025009203236550093, + 0.002709680702537298, + 0.0030795682687312365, + 0.0017150096828117967, + 0.002493818523362279, + 0.002000900451093912, + 0.0017488800222054124, + 0.0019032900454476476, + 0.0015612334245815873, + 0.0015422241995111108, + 0.002371822949498892, + 0.0019000069005414844, + 0.0020903139375150204, + 0.0010744171449914575, + 0.0011893842602148652, + 0.0008959705592133105, + 0.0015929914079606533, + 0.0018329588929191232, + 0.001734058721922338, + 0.0013394986744970083, + 0.0009849383495748043, + 0.002458827570080757, + 0.001811012509278953, + 0.0023737051524221897, + 0.0012213967274874449, + 0.0021720011718571186, + 0.002327438909560442, + 0.0009959540329873562, + 0.001953952945768833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.01820182055234909 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073680594563484 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073680594563484, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073680594563484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25743722915649414, + 0.11579327285289764, + 0.0775463655591011, + 0.03295378386974335, + 0.24480551481246948, + 0.03079456277191639, + 0.11480950564146042, + 0.0486447811126709, + 0.04344967007637024, + 0.033765338361263275 + ], + "credit_map": [ + 0.023248059675097466, + 0.02635282278060913, + 0.03333955630660057, + 0.026339136064052582, + 0.02933591790497303, + 0.045229747891426086, + 0.023901358246803284, + 0.049690645188093185, + 0.011353340931236744, + 0.026429465040564537, + 0.012652630917727947, + 0.01888537034392357, + 0.01427399180829525, + 0.009966271929442883, + 0.01173314917832613, + 0.010499058291316032, + 0.006012905389070511, + 0.009008707478642464, + 0.010471826419234276, + 0.010083182714879513, + 0.007810044568032026, + 0.013889770023524761, + 0.010404560714960098, + 0.00986536592245102, + 0.0045105707831680775, + 0.004156647250056267, + 0.003930315840989351, + 0.0029942456167191267, + 0.00431472435593605, + 0.006914379540830851, + 0.0031148442067205906, + 0.003018059302121401, + 0.08101893216371536, + 0.011904877610504627, + 0.02833734080195427, + 0.008944693952798843, + 0.012082175351679325, + 0.07008536905050278, + 0.008295385167002678, + 0.024136744439601898, + 0.003521617501974106, + 0.004393054638057947, + 0.003031224710866809, + 0.003131123259663582, + 0.004534573759883642, + 0.004899523686617613, + 0.003316451096907258, + 0.003966995514929295, + 0.013067568652331829, + 0.020084436982870102, + 0.020866265520453453, + 0.015899192541837692, + 0.01036914438009262, + 0.012778930366039276, + 0.011899890378117561, + 0.00984407588839531, + 0.007341634016484022, + 0.0052283755503594875, + 0.00570383807644248, + 0.004399911500513554, + 0.005798004567623138, + 0.00721405353397131, + 0.005821904167532921, + 0.0071370601654052734, + 0.005632758606225252, + 0.005461364984512329, + 0.00408818619325757, + 0.008015281520783901, + 0.0048575433902442455, + 0.006860231980681419, + 0.004844099748879671, + 0.003690205980092287, + 0.004799084272235632, + 0.004304039292037487, + 0.004929894581437111, + 0.0023298331070691347, + 0.004495036788284779, + 0.005159464664757252, + 0.002272470621392131, + 0.0054755136370658875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.011073680594563484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033691875636577606 + ], + "opposite_logprob_tokens": [ + -3.408691883087158 + ], + "logprob_reward": -0.033691875636577606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033691875636577606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47740447521209717, + 0.09772517532110214, + 0.01247025839984417, + 0.05698294937610626, + 0.19843506813049316, + 0.02828137017786503, + 0.08346045017242432, + 0.012510478496551514, + 0.02597905322909355, + 0.006750765722244978 + ], + "credit_map": [ + 0.034222859889268875, + 0.01991041749715805, + 0.033983394503593445, + 0.017317412421107292, + 0.02907566912472248, + 0.16749218106269836, + 0.048874929547309875, + 0.1265275925397873, + 0.01004592515528202, + 0.02407022938132286, + 0.010137763805687428, + 0.015253808349370956, + 0.011894519440829754, + 0.008651632815599442, + 0.011066407896578312, + 0.006604887079447508, + 0.0010245818411931396, + 0.0015040506841614842, + 0.0013260595733299851, + 0.001998336287215352, + 0.0015712050953879952, + 0.0016203271225094795, + 0.0017037264769896865, + 0.001721972250379622, + 0.006087800022214651, + 0.007429283577948809, + 0.00850465428084135, + 0.0062787714414298534, + 0.008082074113190174, + 0.009833046235144138, + 0.005859988741576672, + 0.004907331895083189, + 0.06898537278175354, + 0.007632014807313681, + 0.023434258997440338, + 0.005954301450401545, + 0.0065278853289783, + 0.05987315997481346, + 0.006413670722395182, + 0.019614417105913162, + 0.00576407928019762, + 0.0024583078920841217, + 0.0030350012239068747, + 0.0025301268324255943, + 0.0037283513229340315, + 0.006530446000397205, + 0.0025297871325165033, + 0.0017052695620805025, + 0.012458325363695621, + 0.01571156457066536, + 0.010999547317624092, + 0.011228151619434357, + 0.008363577537238598, + 0.00728142075240612, + 0.0075296685099601746, + 0.009888200089335442, + 0.0021954297553747892, + 0.0013807093491777778, + 0.0014434285694733262, + 0.001438987092114985, + 0.001445416477508843, + 0.0013954068999737501, + 0.0019714266527444124, + 0.0012396742822602391, + 0.0023512972984462976, + 0.0037672631442546844, + 0.0015524661866948009, + 0.003466307884082198, + 0.005238335579633713, + 0.004385712090879679, + 0.0031369293574243784, + 0.0020807397086173296, + 0.0007179370149970055, + 0.0009981811745092273, + 0.0010592469479888678, + 0.0008661242900416255, + 0.0010030710836872458, + 0.0009186584502458572, + 0.0005725151277147233, + 0.0006150314002297819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.033691875636577606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows two enormous butterflies as they soar at breakneck speed across a lush meadow bathed in intense, high-contrast light and shadow. In this stark black and white world, the leading butterfly's wings, nearly snow-white, punctuated only by delicate black veins, appear fragile and translucent in the sunlight. The pursuer, in stark contrast, possesses velvety black wings emblazoned with intricate white geometric patterns, each flap imbued with power and determination. The camera tracks their flight paths, capturing the former's desperate evasion with a swift and erratic path, while the latter pursues with a more direct and aggressive approach. Each powerful downward thrust and graceful upward stroke of their massive wings is captured with breathtaking clarity, even the subtle vibrations at the wing edges and the minute air currents they create, violently swaying the knee-high grass and unidentified wildflowers below. The strong light casts a huge, fast-moving, ever-changing shadow of a butterfly on the grass, which, combined with the sharp shadows of the grass leaves themselves, creates a dynamic, dreamy, and tense surreal picture.", + "seed": 16233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_6972_seed_16233_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09077653810381889, + "vlm_logprob": -0.09077653810381889, + "vlm_yes_no_margin": 3.6000000119209288, + "vlm_token_credit": -0.09077653810381889, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3949412703514099 + ], + "opposite_logprob_tokens": [ + -1.1449413299560547 + ], + "logprob_reward": -0.3949412703514099, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3949412703514099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2003227323293686, + 0.0448952354490757, + 0.07199756801128387, + 0.036451444029808044, + 0.09868216514587402, + 0.06721722334623337, + 0.03662789985537529, + 0.045966845005750656, + 0.07445592433214188, + 0.3233829438686371 + ], + "credit_map": [ + 0.02862493507564068, + 0.023098960518836975, + 0.03168538212776184, + 0.03150654956698418, + 0.0200993362814188, + 0.021241849288344383, + 0.016839127987623215, + 0.02722659334540367, + 0.00534489331766963, + 0.005712677724659443, + 0.004535456653684378, + 0.004783149808645248, + 0.00476719019934535, + 0.0038019968196749687, + 0.006462075747549534, + 0.009487795643508434, + 0.006808197125792503, + 0.0075450618751347065, + 0.007443652022629976, + 0.007360283751040697, + 0.007996872998774052, + 0.008323835209012032, + 0.010504371486604214, + 0.016015296801924706, + 0.003553513903170824, + 0.004856190178543329, + 0.0031372234225273132, + 0.005895825568586588, + 0.004956947639584541, + 0.005181682761758566, + 0.0036812883336097, + 0.005188771057873964, + 0.014756720513105392, + 0.009523658081889153, + 0.015197644010186195, + 0.016260966658592224, + 0.007943098433315754, + 0.011368165723979473, + 0.009931820444762707, + 0.013700088486075401, + 0.01251948345452547, + 0.010430174879729748, + 0.007861806079745293, + 0.0071604591794312, + 0.010375689715147018, + 0.007661108858883381, + 0.0046229795552790165, + 0.006585521157830954, + 0.004654211923480034, + 0.0037957928143441677, + 0.0046915290877223015, + 0.0045204972848296165, + 0.004839676432311535, + 0.005258519668132067, + 0.005252593196928501, + 0.0036150815431028605, + 0.005745301488786936, + 0.004443636629730463, + 0.00391496904194355, + 0.005005276761949062, + 0.0067072524689137936, + 0.006291024852544069, + 0.009061927907168865, + 0.004797455854713917, + 0.016387488692998886, + 0.007913813926279545, + 0.008382887579500675, + 0.005387717392295599, + 0.009547951631247997, + 0.013633142225444317, + 0.006604332942515612, + 0.0065985871478915215, + 0.0563434436917305, + 0.03432062640786171, + 0.04125136509537697, + 0.054519083350896835, + 0.04390450194478035, + 0.02953205816447735, + 0.031949594616889954, + 0.031562283635139465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.3949412703514099 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0206155888736248 + ], + "opposite_logprob_tokens": [ + -3.895615577697754 + ], + "logprob_reward": -0.0206155888736248, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0206155888736248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23350872099399567, + 0.1680091917514801, + 0.06850478798151016, + 0.09056378901004791, + 0.05333366245031357, + 0.10707005113363266, + 0.07717433571815491, + 0.0329304039478302, + 0.05393632873892784, + 0.11496875435113907 + ], + "credit_map": [ + 0.04132632538676262, + 0.02402377687394619, + 0.04303082823753357, + 0.037945639342069626, + 0.015541144646704197, + 0.024393949657678604, + 0.019501550123095512, + 0.027745485305786133, + 0.018283216282725334, + 0.012385251000523567, + 0.014802802354097366, + 0.023021893575787544, + 0.01508633978664875, + 0.018326889723539352, + 0.028330238536000252, + 0.03777255862951279, + 0.009458259679377079, + 0.007264501880854368, + 0.008831572718918324, + 0.006227811798453331, + 0.008375579491257668, + 0.0073709129355847836, + 0.00835664663463831, + 0.012619506567716599, + 0.01088747475296259, + 0.009303374215960503, + 0.008491280488669872, + 0.01763816550374031, + 0.0072167785838246346, + 0.011464061215519905, + 0.007282562553882599, + 0.018280087038874626, + 0.006991076283156872, + 0.01147205475717783, + 0.006785650737583637, + 0.003857321571558714, + 0.005933311302214861, + 0.005707873497158289, + 0.005890750791877508, + 0.0066956221126019955, + 0.010832441039383411, + 0.012300378642976284, + 0.009498805738985538, + 0.010463587939739227, + 0.014383249916136265, + 0.011610065586864948, + 0.017534801736474037, + 0.020446715876460075, + 0.011476879008114338, + 0.006573287304490805, + 0.013705350458621979, + 0.009046711027622223, + 0.009006872773170471, + 0.011944073252379894, + 0.00822366587817669, + 0.007197492755949497, + 0.004121008329093456, + 0.0043430146761238575, + 0.003825652413070202, + 0.003969961311668158, + 0.004101327154785395, + 0.004554951563477516, + 0.005453759804368019, + 0.0025607324205338955, + 0.009483471512794495, + 0.006327019073069096, + 0.007058208342641592, + 0.005911016836762428, + 0.007469742558896542, + 0.009687493555247784, + 0.00451281713321805, + 0.0034865594934672117, + 0.01387070119380951, + 0.011027254164218903, + 0.01315024122595787, + 0.024794377386569977, + 0.01643271930515766, + 0.013840372674167156, + 0.010508286766707897, + 0.011344803497195244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0206155888736248 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819397695362568 + ], + "opposite_logprob_tokens": [ + -4.018194198608398 + ], + "logprob_reward": -0.01819397695362568, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819397695362568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32063400745391846, + 0.07561890035867691, + 0.11191093921661377, + 0.03809099271893501, + 0.06729258596897125, + 0.08756941556930542, + 0.09296749532222748, + 0.040902432054281235, + 0.06234855204820633, + 0.10266466438770294 + ], + "credit_map": [ + 0.05223443731665611, + 0.047429706901311874, + 0.03883380815386772, + 0.04905594140291214, + 0.030471164733171463, + 0.031931400299072266, + 0.030804824084043503, + 0.03987272456288338, + 0.012311326339840889, + 0.008145288564264774, + 0.005609460175037384, + 0.008661746978759766, + 0.008351747877895832, + 0.00824100710451603, + 0.01044214703142643, + 0.013856173492968082, + 0.01179514080286026, + 0.025743169710040092, + 0.012724229134619236, + 0.011572261340916157, + 0.010590463876724243, + 0.013549764640629292, + 0.010417022742331028, + 0.01551888883113861, + 0.005124012008309364, + 0.006293428596109152, + 0.003258030628785491, + 0.00503080477938056, + 0.004187805112451315, + 0.005922760348767042, + 0.00352336373180151, + 0.004750786814838648, + 0.007640278432518244, + 0.008673349395394325, + 0.011432413943111897, + 0.006137669086456299, + 0.005955606233328581, + 0.00862823985517025, + 0.008458862081170082, + 0.010366168804466724, + 0.008636933751404285, + 0.018140114843845367, + 0.009441065602004528, + 0.00885408092290163, + 0.009449568577110767, + 0.009104415774345398, + 0.011901097372174263, + 0.012042135000228882, + 0.011100602336227894, + 0.013610010035336018, + 0.011812943033874035, + 0.014890584163367748, + 0.015424913726747036, + 0.009564254432916641, + 0.010656395927071571, + 0.005907788872718811, + 0.003924404736608267, + 0.005179979372769594, + 0.004193217493593693, + 0.005977118853479624, + 0.005864505656063557, + 0.0051824175752699375, + 0.005501384846866131, + 0.005079400725662708, + 0.011764219962060452, + 0.007716507650911808, + 0.005044229328632355, + 0.004513624124228954, + 0.005246927961707115, + 0.018351677805185318, + 0.005053899250924587, + 0.004657466430217028, + 0.015542435459792614, + 0.012415328994393349, + 0.013345228508114815, + 0.021648166701197624, + 0.012781444936990738, + 0.008479599840939045, + 0.009395158849656582, + 0.00905731413513422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.01819397695362568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059441933408379555 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.0059441933408379555, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059441933408379555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22443899512290955, + 0.09072870761156082, + 0.15990108251571655, + 0.03846633434295654, + 0.05505039915442467, + 0.08351410925388336, + 0.084404855966568, + 0.0460045225918293, + 0.0470648892223835, + 0.17042608559131622 + ], + "credit_map": [ + 0.04184194654226303, + 0.02843661606311798, + 0.03464101627469063, + 0.023951826617121696, + 0.025038572028279305, + 0.017562443390488625, + 0.021485526114702225, + 0.031481049954891205, + 0.008098341524600983, + 0.010451206006109715, + 0.009016125462949276, + 0.013159167021512985, + 0.010985415428876877, + 0.009080925025045872, + 0.013277974911034107, + 0.016659552231431007, + 0.019050773233175278, + 0.03591505065560341, + 0.017629047855734825, + 0.02210506983101368, + 0.0127693060785532, + 0.018397176638245583, + 0.01943238079547882, + 0.014602284878492355, + 0.0045545948669314384, + 0.004975474905222654, + 0.003814778756350279, + 0.00464269332587719, + 0.004852484446018934, + 0.00519872410222888, + 0.004399747587740421, + 0.006027834489941597, + 0.006836039014160633, + 0.007259584963321686, + 0.00822952575981617, + 0.00616335729137063, + 0.003911741077899933, + 0.005076040979474783, + 0.008517575450241566, + 0.009056536480784416, + 0.010711842216551304, + 0.010039793327450752, + 0.006136823445558548, + 0.005631649401038885, + 0.0209350548684597, + 0.010671539232134819, + 0.005725922994315624, + 0.013661479577422142, + 0.009697690606117249, + 0.011133179999887943, + 0.009869517758488655, + 0.012912928126752377, + 0.01298905722796917, + 0.00838239211589098, + 0.013090082444250584, + 0.0063300104811787605, + 0.005030039232224226, + 0.007719101849943399, + 0.004947942215949297, + 0.004454139154404402, + 0.0063631488010287285, + 0.00719864945858717, + 0.005552931688725948, + 0.004738569259643555, + 0.010044116526842117, + 0.004765904042869806, + 0.003740189829841256, + 0.0031241762917488813, + 0.004293741192668676, + 0.013435511849820614, + 0.004153536166995764, + 0.0035077109932899475, + 0.017910586670041084, + 0.01835758611559868, + 0.014359205029904842, + 0.024523021653294563, + 0.026954280212521553, + 0.02754499390721321, + 0.018925828859210014, + 0.021850580349564552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0059441933408379555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187660999596119 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.014187660999596119, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187660999596119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19919390976428986, + 0.182423934340477, + 0.08776507526636124, + 0.043978720903396606, + 0.07125348597764969, + 0.04694410786032677, + 0.11609066277742386, + 0.0420413613319397, + 0.12582361698150635, + 0.08448512107133865 + ], + "credit_map": [ + 0.023535750806331635, + 0.027442816644906998, + 0.03142736852169037, + 0.041702061891555786, + 0.016209203749895096, + 0.018956322222948074, + 0.01633531227707863, + 0.02358507178723812, + 0.017730161547660828, + 0.015923408791422844, + 0.019930025562644005, + 0.02434396743774414, + 0.015860682353377342, + 0.018644072115421295, + 0.030450552701950073, + 0.03954106941819191, + 0.008663225919008255, + 0.012382568791508675, + 0.008204706013202667, + 0.008819973096251488, + 0.011436307802796364, + 0.009701509028673172, + 0.010679682716727257, + 0.01787710376083851, + 0.004121577367186546, + 0.006503187585622072, + 0.0045458609238266945, + 0.004683602135628462, + 0.0064459568820893764, + 0.006536219269037247, + 0.00489262817427516, + 0.006249688565731049, + 0.012322849594056606, + 0.0068359035067260265, + 0.010615442879498005, + 0.008099565282464027, + 0.00595035869628191, + 0.008334974758327007, + 0.010892299935221672, + 0.008202086202800274, + 0.005538099445402622, + 0.006792815867811441, + 0.005120933521538973, + 0.005791010335087776, + 0.0048680840991437435, + 0.00571292731910944, + 0.006677729543298483, + 0.006442506797611713, + 0.014260097406804562, + 0.014779736287891865, + 0.020987754687666893, + 0.017242183908820152, + 0.015587800182402134, + 0.011597078293561935, + 0.012007782235741615, + 0.009628244675695896, + 0.004057119134813547, + 0.004296140745282173, + 0.004433775320649147, + 0.0043916176073253155, + 0.004785583354532719, + 0.004290726967155933, + 0.009958312846720219, + 0.005828089080750942, + 0.0190512016415596, + 0.02132287807762623, + 0.017244497314095497, + 0.01738692633807659, + 0.013505164533853531, + 0.015683608129620552, + 0.011830965057015419, + 0.009798373095691204, + 0.012470054440200329, + 0.008527915924787521, + 0.011224254034459591, + 0.01700688526034355, + 0.011951377615332603, + 0.007787204347550869, + 0.009538521990180016, + 0.005978906061500311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014187660999596119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows two enormous butterflies as they soar at breakneck speed across a lush meadow bathed in intense, high-contrast light and shadow. In this stark black and white world, the leading butterfly's wings, nearly snow-white, punctuated only by delicate black veins, appear fragile and translucent in the sunlight. The pursuer, in stark contrast, possesses velvety black wings emblazoned with intricate white geometric patterns, each flap imbued with power and determination. The camera tracks their flight paths, capturing the former's desperate evasion with a swift and erratic path, while the latter pursues with a more direct and aggressive approach. Each powerful downward thrust and graceful upward stroke of their massive wings is captured with breathtaking clarity, even the subtle vibrations at the wing edges and the minute air currents they create, violently swaying the knee-high grass and unidentified wildflowers below. The strong light casts a huge, fast-moving, ever-changing shadow of a butterfly on the grass, which, combined with the sharp shadows of the grass leaves themselves, creates a dynamic, dreamy, and tense surreal picture.", + "seed": 16234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_1/index_6972_seed_16234_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.042366760317236185, + "vlm_logprob": -0.042366760317236185, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.042366760317236185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14701108634471893 + ], + "opposite_logprob_tokens": [ + -2.0220110416412354 + ], + "logprob_reward": -0.14701108634471893, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14701108634471893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04300334304571152, + 0.14923790097236633, + 0.054774001240730286, + 0.04839443042874336, + 0.02945452556014061, + 0.026368238031864166, + 0.03399040922522545, + 0.5336571931838989, + 0.011585905216634274, + 0.06953404098749161 + ], + "credit_map": [ + 0.007116579916328192, + 0.00698621291667223, + 0.002986261388286948, + 0.003937837202101946, + 0.006390823051333427, + 0.004414523486047983, + 0.004020018968731165, + 0.007151092868298292, + 0.05528676137328148, + 0.010323120281100273, + 0.009341980330646038, + 0.008733736351132393, + 0.006928135175257921, + 0.042833130806684494, + 0.007207391317933798, + 0.00858366023749113, + 0.006134280003607273, + 0.007504868786782026, + 0.007919251918792725, + 0.009766686707735062, + 0.003404204035177827, + 0.007529452443122864, + 0.004552430007606745, + 0.007962832227349281, + 0.006872878409922123, + 0.004875066224485636, + 0.007429579738527536, + 0.006584611721336842, + 0.004031065851449966, + 0.005706069990992546, + 0.006181959994137287, + 0.006713201757520437, + 0.003432526485994458, + 0.002738155424594879, + 0.004906002432107925, + 0.00517251156270504, + 0.002296222373843193, + 0.0020558827091008425, + 0.003746137721464038, + 0.0051070875488221645, + 0.0032369340769946575, + 0.0019668801687657833, + 0.0033075115643441677, + 0.003978558350354433, + 0.002678146818652749, + 0.003008082043379545, + 0.004105288535356522, + 0.004086839035153389, + 0.002709205960854888, + 0.0026987395249307156, + 0.008250851184129715, + 0.004103014245629311, + 0.0020250887610018253, + 0.0032093797344714403, + 0.005779021419584751, + 0.005215111654251814, + 0.1323840171098709, + 0.015045396983623505, + 0.028900310397148132, + 0.01417637150734663, + 0.022925661876797676, + 0.0700182244181633, + 0.0870320126414299, + 0.16317521035671234, + 0.0014807891566306353, + 0.001372471684589982, + 0.0018523582257330418, + 0.0016124625690281391, + 0.0010881902417168021, + 0.001173335243947804, + 0.0013579204678535461, + 0.0016483782092109323, + 0.012044193223118782, + 0.003457192797213793, + 0.006959197577089071, + 0.006228872574865818, + 0.014526731334626675, + 0.013629808090627193, + 0.005347217433154583, + 0.007340837270021439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.14701108634471893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314977064728737 + ], + "opposite_logprob_tokens": [ + -3.773314952850342 + ], + "logprob_reward": -0.023314977064728737, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314977064728737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025430791079998016, + 0.011702721007168293, + 0.021394819021224976, + 0.015978693962097168, + 0.029732920229434967, + 0.05752217397093773, + 0.052237797528505325, + 0.7285694479942322, + 0.017253471538424492, + 0.04017722234129906 + ], + "credit_map": [ + 0.0057588014751672745, + 0.006197736598551273, + 0.0015395812224596739, + 0.0013207003939896822, + 0.004809177480638027, + 0.0018055748660117388, + 0.00185589084867388, + 0.002143328543752432, + 0.0020974755752831697, + 0.0020610010251402855, + 0.0012163846986368299, + 0.0013903100043535233, + 0.0014633291866630316, + 0.0012928195064887404, + 0.0013221201952546835, + 0.000859281572047621, + 0.002091160509735346, + 0.0030997195281088352, + 0.0027228076942265034, + 0.00406095152720809, + 0.002013612538576126, + 0.003474217839539051, + 0.0016459714388474822, + 0.0022863775957375765, + 0.0014377804473042488, + 0.0015269746072590351, + 0.002323533408343792, + 0.002289550844579935, + 0.001148216426372528, + 0.0014193356037139893, + 0.0022261447738856077, + 0.003607158549129963, + 0.002471392974257469, + 0.0026089695747941732, + 0.004793199710547924, + 0.005235959775745869, + 0.002994404872879386, + 0.002528700977563858, + 0.004483796656131744, + 0.0046164956875145435, + 0.0037640982773154974, + 0.0031282424461096525, + 0.008050834760069847, + 0.009599950164556503, + 0.005315420683473349, + 0.004189812578260899, + 0.013322645798325539, + 0.010151167400181293, + 0.004195849411189556, + 0.003065672004595399, + 0.012090661562979221, + 0.005764237139374018, + 0.0023002594243735075, + 0.003161385655403137, + 0.012117619626224041, + 0.009542112238705158, + 0.17846673727035522, + 0.02099587954580784, + 0.04144011065363884, + 0.019626513123512268, + 0.03133413568139076, + 0.0956946313381195, + 0.11749143153429031, + 0.22352001070976257, + 0.0027828747406601906, + 0.0010165172861889005, + 0.0027816323563456535, + 0.0021205341909080744, + 0.0011286661028862, + 0.0018027801997959614, + 0.002748943865299225, + 0.0028715224470943213, + 0.007837099954485893, + 0.003252865746617317, + 0.003080968977883458, + 0.002275944920256734, + 0.009466979652643204, + 0.007241609040647745, + 0.0031296901870518923, + 0.003892061533406377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.023314977064728737 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202234335243702 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014202234335243702, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202234335243702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06764665991067886, + 0.19818957149982452, + 0.07339255511760712, + 0.01932618021965027, + 0.035409729927778244, + 0.01240913849323988, + 0.016613587737083435, + 0.5308240056037903, + 0.011036963202059269, + 0.03515166416764259 + ], + "credit_map": [ + 0.010418794117867947, + 0.007717584725469351, + 0.00581642659381032, + 0.006624102126806974, + 0.009424926713109016, + 0.006186602171510458, + 0.009549801237881184, + 0.011908418498933315, + 0.07123925536870956, + 0.013818822801113129, + 0.015448464080691338, + 0.010103157721459866, + 0.009760645218193531, + 0.05681521072983742, + 0.008725272491574287, + 0.012278721667826176, + 0.006720896810293198, + 0.010335764847695827, + 0.010245520621538162, + 0.014436192810535431, + 0.005158469080924988, + 0.009574513882398605, + 0.0052933297120034695, + 0.011627866886556149, + 0.002733551198616624, + 0.0017859890358522534, + 0.00289303925819695, + 0.0024742719251662493, + 0.0014891508035361767, + 0.0025861719623208046, + 0.0022522208746522665, + 0.0031117836479097605, + 0.0036302278749644756, + 0.0036297773476690054, + 0.00832605455070734, + 0.005904602352529764, + 0.002569165313616395, + 0.0022365206386893988, + 0.004313284996896982, + 0.00480009475722909, + 0.0014788394328206778, + 0.0009405520395375788, + 0.0011031387839466333, + 0.0016359465662389994, + 0.0014479095116257668, + 0.001349018421024084, + 0.0023416541516780853, + 0.002112078480422497, + 0.0015439667040482163, + 0.002843765541911125, + 0.0020156383980065584, + 0.001910935970954597, + 0.0018936032429337502, + 0.0017611677758395672, + 0.0024286978878080845, + 0.0022158115170896053, + 0.11651123315095901, + 0.015734177082777023, + 0.05035526305437088, + 0.0187135748565197, + 0.023746438324451447, + 0.06428491324186325, + 0.09086904674768448, + 0.1506093293428421, + 0.0016885515069589019, + 0.0018751220777630806, + 0.0011998642003163695, + 0.001351154875010252, + 0.001434584497474134, + 0.0012851273640990257, + 0.0009200676577165723, + 0.0012824906734749675, + 0.006225917488336563, + 0.0017651199596002698, + 0.004086391534656286, + 0.0030615548603236675, + 0.007648952305316925, + 0.006769648753106594, + 0.002572040306404233, + 0.0030220355838537216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.014202234335243702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02108849212527275, + 0.0436062291264534, + 0.0671798437833786, + 0.009187575429677963, + 0.020804625004529953, + 0.01388874277472496, + 0.010687021538615227, + 0.8034754395484924, + 0.0040877885185182095, + 0.005994314327836037 + ], + "credit_map": [ + 0.003341173054650426, + 0.0035004965029656887, + 0.002193127293139696, + 0.0018215199233964086, + 0.002621793420985341, + 0.001951972022652626, + 0.0025772100780159235, + 0.0030811999458819628, + 0.015380553901195526, + 0.0028679282404482365, + 0.0028835146222263575, + 0.0022864555940032005, + 0.002062266692519188, + 0.012795898132026196, + 0.0024739643558859825, + 0.0028556480538100004, + 0.007713377941399813, + 0.007928858511149883, + 0.00915443617850542, + 0.01144920103251934, + 0.0055829985067248344, + 0.010892114602029324, + 0.006070865783840418, + 0.00838799774646759, + 0.0007838299497961998, + 0.0007310103974305093, + 0.0011630587978288531, + 0.0016059490153566003, + 0.0011510937474668026, + 0.0008472013869322836, + 0.0018714964389801025, + 0.0010339353466406465, + 0.001959045883268118, + 0.0015485157491639256, + 0.005239236168563366, + 0.0033276770263910294, + 0.0017124175792559981, + 0.0011304147774353623, + 0.00296322931535542, + 0.0029240886215120554, + 0.0011766530806198716, + 0.0014971696073189378, + 0.0019704890437424183, + 0.0016163837863132358, + 0.0024198738392442465, + 0.001424525398761034, + 0.0016012379201129079, + 0.002182410331442952, + 0.0007334368419833481, + 0.0014980128034949303, + 0.0014186338521540165, + 0.0015037264674901962, + 0.0010357825085520744, + 0.0010953355813398957, + 0.0019778688438236713, + 0.0014242250472307205, + 0.19069623947143555, + 0.023372676223516464, + 0.055455923080444336, + 0.024029308930039406, + 0.03472530469298363, + 0.10317826271057129, + 0.13089562952518463, + 0.2411220371723175, + 0.0005161594599485397, + 0.00041501125087961555, + 0.0004321097512729466, + 0.00048295536544173956, + 0.0005566972540691495, + 0.00044089899165555835, + 0.0005695972940884531, + 0.0006743590929545462, + 0.0006788548780605197, + 0.0006348609458655119, + 0.0009818609105423093, + 0.000705538026522845, + 0.00127163075376302, + 0.000916992430575192, + 0.00039007054874673486, + 0.0004145058337599039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.006734057795256376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in stark black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two enormous butterflies visible soaring at high speed across a lush meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading butterfly's wing color snow-white with delicate black veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing butterfly's wing color velvety black with intricate white geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the butterflies' flight paths with a low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571446046233177 + ], + "opposite_logprob_tokens": [ + -3.89557147026062 + ], + "logprob_reward": -0.020571446046233177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571446046233177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0339340940117836, + 0.016909688711166382, + 0.026304803788661957, + 0.011172733269631863, + 0.0359782911837101, + 0.06357020884752274, + 0.056235749274492264, + 0.6885208487510681, + 0.04379008337855339, + 0.023583507165312767 + ], + "credit_map": [ + 0.005980540998280048, + 0.006427721120417118, + 0.0022489142138510942, + 0.002672268310561776, + 0.0052588991820812225, + 0.002967601642012596, + 0.004256946500390768, + 0.004121202044188976, + 0.004964741412550211, + 0.0013363017933443189, + 0.0017118859104812145, + 0.0012909984216094017, + 0.0011639122385531664, + 0.0036868026945739985, + 0.0013008936075493693, + 0.001454151002690196, + 0.004927304573357105, + 0.0022663804702460766, + 0.0032577048987150192, + 0.00285895774140954, + 0.0021042274311184883, + 0.005859444849193096, + 0.001971814315766096, + 0.0030589685775339603, + 0.000889008748345077, + 0.0012115351855754852, + 0.0013239823747426271, + 0.0016910050762817264, + 0.000989214633591473, + 0.0012546106008812785, + 0.0016958002233877778, + 0.0021175749134272337, + 0.004000476561486721, + 0.0035408865660429, + 0.007224066182971001, + 0.0046927896328270435, + 0.0036424673162400723, + 0.002503157826140523, + 0.005016528535634279, + 0.005357918329536915, + 0.00451141269877553, + 0.0045918794348835945, + 0.009448308497667313, + 0.00961958896368742, + 0.006995965261012316, + 0.004711898509413004, + 0.013990616425871849, + 0.009700533002614975, + 0.004893389530479908, + 0.00456579914316535, + 0.01146059762686491, + 0.005351137835532427, + 0.0034090958070009947, + 0.005021531600505114, + 0.014664486050605774, + 0.006869709584861994, + 0.16273558139801025, + 0.020153190940618515, + 0.04946095123887062, + 0.019964585080742836, + 0.03004554472863674, + 0.08837507665157318, + 0.11175573617219925, + 0.20603017508983612, + 0.007325282320380211, + 0.0026477572973817587, + 0.0070761567912995815, + 0.005151339340955019, + 0.002970780711621046, + 0.004714746959507465, + 0.0071376352570950985, + 0.006766383070498705, + 0.004565846174955368, + 0.0021467208862304688, + 0.00366015313193202, + 0.002236796310171485, + 0.004230131395161152, + 0.0028140340000391006, + 0.00196346384473145, + 0.001966359792277217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.020571446046233177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_tensor([3718, 3718])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_tensor([3718, 3718])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8246ff8989715ac65ee40759103f4bc9b8553d23 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_tensor([3718, 3718])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Rendered in a vibrant Pixar style, the scene begins with a high-angle shot, the camera slowly and gracefully descending toward a stylized map, magically realized on a textured, ancient scroll. The scroll, with its warm, sheepskin-yellow hue and delicately worn edges, unfolds within a soft, dark void, its surface shimmering with an inner light. Above, a winding tour route weaves through the terracotta excavation pits, rendered with bold, animated lines, connecting the charming pit icons against a landscape of saturated greens and earthy browns. As the camera smoothly advances, it focuses on a whimsically animated cargo ship, constructed from soft, rounded planks and a cheerful design, gliding steadily along a painted, bright blue river on the map. The ship's decks are meticulously arranged with rows of adorable miniature terracotta warriors, each with an oversized head and simplified, friendly features. The ship is deliberately moored in a clearly marked port; a small, stylized anchor hangs from the bow on a thin rope, and when it hits the painted river, it doesn't splash but instead sends up a delightful, vibrant puff of golden-brown dust that gently dissipates, perfectly capturing an atmosphere of fun and magical adventure.", + "seed": 15325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_3718_seed_15325_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018967143213376402, + "vlm_logprob": -0.018967143213376402, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.018967143213376402, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03924291208386421 + ], + "opposite_logprob_tokens": [ + -3.78924298286438 + ], + "logprob_reward": -0.03924291208386421, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03924291208386421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24932965636253357, + 0.07226095348596573, + 0.12093885242938995, + 0.13049070537090302, + 0.07896354049444199, + 0.05504544451832771, + 0.09296706318855286, + 0.11357943713665009, + 0.02917349711060524, + 0.05725080892443657 + ], + "credit_map": [ + 0.018582213670015335, + 0.02156927064061165, + 0.06752073019742966, + 0.029556332156062126, + 0.051178980618715286, + 0.020903252065181732, + 0.019315162673592567, + 0.02070373296737671, + 0.005281132645905018, + 0.008846067823469639, + 0.013163694180548191, + 0.012973594479262829, + 0.0044257752597332, + 0.005658620037138462, + 0.008735602721571922, + 0.013176471926271915, + 0.007631855085492134, + 0.014406289905309677, + 0.010824370197951794, + 0.023327818140387535, + 0.017502466216683388, + 0.009397154673933983, + 0.020039435476064682, + 0.01780947856605053, + 0.008145337924361229, + 0.010525897145271301, + 0.04530496895313263, + 0.009986422024667263, + 0.007787665817886591, + 0.009569535963237286, + 0.01870458945631981, + 0.02046629600226879, + 0.00903720036149025, + 0.007799975108355284, + 0.014163359068334103, + 0.00957569945603609, + 0.006767058279365301, + 0.009066742844879627, + 0.010429812595248222, + 0.012123702093958855, + 0.005567715037614107, + 0.007368483114987612, + 0.007692505139857531, + 0.004527289420366287, + 0.00832188967615366, + 0.008041753433644772, + 0.006618210580199957, + 0.006907606963068247, + 0.011153447441756725, + 0.013816040940582752, + 0.00842877384275198, + 0.009768943302333355, + 0.00882816780358553, + 0.01903548277914524, + 0.010898216627538204, + 0.011037995107471943, + 0.02988205850124359, + 0.009961489588022232, + 0.009639039635658264, + 0.00909925065934658, + 0.006625190377235413, + 0.01840571127831936, + 0.008886540308594704, + 0.021080171689391136, + 0.00478111207485199, + 0.0047622742131352425, + 0.0030390669126063585, + 0.0036851228214800358, + 0.0030748690478503704, + 0.003148658201098442, + 0.003522302024066448, + 0.0031600939109921455, + 0.014047116972506046, + 0.006402983795851469, + 0.013951036147773266, + 0.005106258671730757, + 0.004258005879819393, + 0.006127381231635809, + 0.003647745354101062, + 0.003710287855938077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.03924291208386421 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643240667879581 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643240667879581, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643240667879581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1136467382311821, + 0.13588201999664307, + 0.21722619235515594, + 0.09080661088228226, + 0.07715307921171188, + 0.10705641657114029, + 0.062840037047863, + 0.1429038792848587, + 0.03184651583433151, + 0.020638512447476387 + ], + "credit_map": [ + 0.010939741507172585, + 0.012157357297837734, + 0.019272442907094955, + 0.01668424904346466, + 0.011034040711820126, + 0.00964641198515892, + 0.01703871786594391, + 0.01687377505004406, + 0.009458665736019611, + 0.01229275856167078, + 0.024847399443387985, + 0.019674085080623627, + 0.014129847288131714, + 0.012267379090189934, + 0.015349358320236206, + 0.02786252088844776, + 0.012275765649974346, + 0.03546563908457756, + 0.01973322406411171, + 0.05126847326755524, + 0.025361502543091774, + 0.017492396757006645, + 0.03029770590364933, + 0.02533147856593132, + 0.010241830721497536, + 0.011856332421302795, + 0.012321127578616142, + 0.007939930073916912, + 0.010523338802158833, + 0.010973934084177017, + 0.012869595550000668, + 0.014080519787967205, + 0.009215157479047775, + 0.010046778246760368, + 0.010662865824997425, + 0.010704395361244678, + 0.009502257220447063, + 0.008154262788593769, + 0.008207929320633411, + 0.010659436695277691, + 0.012975437566637993, + 0.011514915153384209, + 0.012604889459908009, + 0.012857433408498764, + 0.010354405269026756, + 0.01528205070644617, + 0.016425244510173798, + 0.015042039565742016, + 0.006557159125804901, + 0.007900784723460674, + 0.005874148104339838, + 0.008806193247437477, + 0.007317026145756245, + 0.008573736995458603, + 0.008762387558817863, + 0.00904859509319067, + 0.034886956214904785, + 0.008784683421254158, + 0.010966847650706768, + 0.008952854201197624, + 0.0081967543810606, + 0.046402864158153534, + 0.008626875467598438, + 0.016086040064692497, + 0.00388676137663424, + 0.005013213027268648, + 0.0033990254160016775, + 0.003369859652593732, + 0.004516672808676958, + 0.004932870157063007, + 0.004601495340466499, + 0.00212661549448967, + 0.0031808859203010798, + 0.002634174656122923, + 0.002380050951614976, + 0.002158908871933818, + 0.0026069283485412598, + 0.0025114535819739103, + 0.002291678683832288, + 0.0028744321316480637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.008643240667879581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021977098658680916 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021977098658680916, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021977098658680916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15762706100940704, + 0.09174185246229172, + 0.12067603319883347, + 0.11337584257125854, + 0.15144211053848267, + 0.04371798038482666, + 0.14109668135643005, + 0.09959796071052551, + 0.03615494444966316, + 0.044569529592990875 + ], + "credit_map": [ + 0.018392542377114296, + 0.018309930339455605, + 0.01753712072968483, + 0.022384250536561012, + 0.0313449427485466, + 0.015501371584832668, + 0.017334630712866783, + 0.016822269186377525, + 0.007315763738006353, + 0.013227148912847042, + 0.011231094598770142, + 0.01604175940155983, + 0.009384271688759327, + 0.007258950732648373, + 0.013783151283860207, + 0.013499717228114605, + 0.0089637516066432, + 0.02112615294754505, + 0.015747828409075737, + 0.015542282722890377, + 0.013318676501512527, + 0.010259726084768772, + 0.017173532396554947, + 0.018544083461165428, + 0.006423674989491701, + 0.009050006978213787, + 0.03966677188873291, + 0.007125374861061573, + 0.006992726121097803, + 0.006838925648480654, + 0.012464636005461216, + 0.02481371909379959, + 0.019193682819604874, + 0.017939865589141846, + 0.02076691761612892, + 0.01399679109454155, + 0.0160662941634655, + 0.018927808851003647, + 0.01669880747795105, + 0.027851952239871025, + 0.006022622808814049, + 0.005826329812407494, + 0.004747751168906689, + 0.0038026601541787386, + 0.004709172062575817, + 0.0050559570081532, + 0.0066937413066625595, + 0.006859745364636183, + 0.013272816315293312, + 0.015540257096290588, + 0.01709611341357231, + 0.020856274291872978, + 0.012278309091925621, + 0.020706025883555412, + 0.020996669307351112, + 0.020350204780697823, + 0.014117177575826645, + 0.008900119923055172, + 0.01213159691542387, + 0.010341540910303593, + 0.008209167048335075, + 0.027476567775011063, + 0.010516053065657616, + 0.007905744016170502, + 0.0038182849530130625, + 0.004552725702524185, + 0.004553147125989199, + 0.0046624355018138885, + 0.005015718284994364, + 0.00430441414937377, + 0.005958057940006256, + 0.003290159162133932, + 0.008448300883173943, + 0.005123711656779051, + 0.008520358242094517, + 0.005384772550314665, + 0.004707390442490578, + 0.004140864126384258, + 0.005037800874561071, + 0.0032063338439911604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0021977098658680916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017102628480643034 + ], + "opposite_logprob_tokens": [ + -6.376710414886475 + ], + "logprob_reward": -0.0017102628480643034, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017102628480643034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07172247767448425, + 0.10893158614635468, + 0.09099162369966507, + 0.0422087125480175, + 0.10513556003570557, + 0.04725439473986626, + 0.09771490842103958, + 0.26032599806785583, + 0.07765840739011765, + 0.09805633872747421 + ], + "credit_map": [ + 0.008007212541997433, + 0.00774816470220685, + 0.01334606483578682, + 0.009262420237064362, + 0.008463090285658836, + 0.007071236148476601, + 0.01072268933057785, + 0.007101598661392927, + 0.008088694885373116, + 0.009591135196387768, + 0.01424562931060791, + 0.022489342838525772, + 0.010412049479782581, + 0.00862561073154211, + 0.018781285732984543, + 0.016697829589247704, + 0.006101280450820923, + 0.013782141730189323, + 0.010586671531200409, + 0.020986409857869148, + 0.00860762782394886, + 0.007026202976703644, + 0.011579161509871483, + 0.01232212409377098, + 0.003757700091227889, + 0.0070913126692175865, + 0.005848992615938187, + 0.005102692637592554, + 0.0037444490008056164, + 0.00310120009817183, + 0.005821407772600651, + 0.007740958593785763, + 0.01187763549387455, + 0.012330042198300362, + 0.01489248313009739, + 0.00950363278388977, + 0.019981103017926216, + 0.010391714051365852, + 0.009897980839014053, + 0.016260972246527672, + 0.004462238866835833, + 0.005085176322609186, + 0.006740535609424114, + 0.006037392653524876, + 0.00548483757302165, + 0.00486673042178154, + 0.007643512450158596, + 0.0069339694455266, + 0.011959194205701351, + 0.016328632831573486, + 0.013902431353926659, + 0.009261790663003922, + 0.010276151821017265, + 0.015038293786346912, + 0.010626290924847126, + 0.010322117246687412, + 0.050784140825271606, + 0.01897117868065834, + 0.023771993815898895, + 0.02391110174357891, + 0.018327686935663223, + 0.08819181472063065, + 0.0207742340862751, + 0.015593858435750008, + 0.009986001066863537, + 0.013205694034695625, + 0.010556812398135662, + 0.008149306289851665, + 0.00995382945984602, + 0.009190171957015991, + 0.010171613655984402, + 0.006444977130740881, + 0.02190023846924305, + 0.011144844815135002, + 0.02217617817223072, + 0.009603466838598251, + 0.009168845601379871, + 0.00810403935611248, + 0.009505636058747768, + 0.006453095935285091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.0017102628480643034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043041590601205826 + ], + "opposite_logprob_tokens": [ + -3.168041706085205 + ], + "logprob_reward": -0.043041590601205826, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043041590601205826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18028733134269714, + 0.06720077991485596, + 0.07334162294864655, + 0.09948054701089859, + 0.07002858817577362, + 0.049454465508461, + 0.06682023406028748, + 0.09428777545690536, + 0.045691732317209244, + 0.25340694189071655 + ], + "credit_map": [ + 0.011917431838810444, + 0.014838563278317451, + 0.05043261870741844, + 0.022392714396119118, + 0.03914313018321991, + 0.01112349797040224, + 0.016814792528748512, + 0.013624574989080429, + 0.005652758292853832, + 0.007991825230419636, + 0.008693734183907509, + 0.013230224139988422, + 0.0065767802298069, + 0.005013135261833668, + 0.010737888514995575, + 0.00930443499237299, + 0.005158461630344391, + 0.010684042237699032, + 0.008864509873092175, + 0.012570072896778584, + 0.0067225913517177105, + 0.00708788400515914, + 0.013837830163538456, + 0.00841622892767191, + 0.006826276425272226, + 0.011943605728447437, + 0.03197881206870079, + 0.006201482377946377, + 0.0055498965084552765, + 0.004974669776856899, + 0.012394645251333714, + 0.019611170515418053, + 0.006393061485141516, + 0.007574916817247868, + 0.010425602085888386, + 0.006323945242911577, + 0.008439597673714161, + 0.009950184263288975, + 0.007221718318760395, + 0.01369955949485302, + 0.006010279059410095, + 0.005158834625035524, + 0.008182178251445293, + 0.0058989510871469975, + 0.00450772512704134, + 0.007617258466780186, + 0.006379764527082443, + 0.005699478089809418, + 0.008799586445093155, + 0.008522874675691128, + 0.009055586531758308, + 0.006907733157277107, + 0.0069259339943528175, + 0.009583212435245514, + 0.008472888730466366, + 0.00855241809040308, + 0.01434190385043621, + 0.008019382134079933, + 0.006840905174612999, + 0.01001462060958147, + 0.007644664961844683, + 0.024101518094539642, + 0.00903510395437479, + 0.014289671555161476, + 0.006353343836963177, + 0.006925829220563173, + 0.005924257915467024, + 0.004798182286322117, + 0.005801650695502758, + 0.005664363503456116, + 0.006312765646725893, + 0.003911342937499285, + 0.06731805950403214, + 0.02718920260667801, + 0.06195506826043129, + 0.022176144644618034, + 0.020382242277264595, + 0.02121674083173275, + 0.01917460933327675, + 0.013994863256812096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.043041590601205826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a vibrant Pixar style, the scene begins with a high-angle shot, the camera slowly and gracefully descending toward a stylized map, magically realized on a textured, ancient scroll. The scroll, with its warm, sheepskin-yellow hue and delicately worn edges, unfolds within a soft, dark void, its surface shimmering with an inner light. Above, a winding tour route weaves through the terracotta excavation pits, rendered with bold, animated lines, connecting the charming pit icons against a landscape of saturated greens and earthy browns. As the camera smoothly advances, it focuses on a whimsically animated cargo ship, constructed from soft, rounded planks and a cheerful design, gliding steadily along a painted, bright blue river on the map. The ship's decks are meticulously arranged with rows of adorable miniature terracotta warriors, each with an oversized head and simplified, friendly features. The ship is deliberately moored in a clearly marked port; a small, stylized anchor hangs from the bow on a thin rope, and when it hits the painted river, it doesn't splash but instead sends up a delightful, vibrant puff of golden-brown dust that gently dissipates, perfectly capturing an atmosphere of fun and magical adventure.", + "seed": 15326, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_3718_seed_15326_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020527484267950057, + "vlm_logprob": -0.020527484267950057, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.020527484267950057, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02281453087925911 + ], + "opposite_logprob_tokens": [ + -4.772814750671387 + ], + "logprob_reward": -0.02281453087925911, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.02281453087925911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3603861331939697, + 0.07392029464244843, + 0.11252336204051971, + 0.038994476199150085, + 0.09883087873458862, + 0.08396735787391663, + 0.06259258091449738, + 0.03765090927481651, + 0.03698223829269409, + 0.09415175765752792 + ], + "credit_map": [ + 0.029206102713942528, + 0.025930892676115036, + 0.07235312461853027, + 0.022838076576590538, + 0.02459953911602497, + 0.08142302930355072, + 0.026695359498262405, + 0.0773400142788887, + 0.00853080302476883, + 0.008408849127590656, + 0.011373474262654781, + 0.007419733330607414, + 0.007392756640911102, + 0.014351335354149342, + 0.006976054981350899, + 0.009467294439673424, + 0.007044119294732809, + 0.012591562233865261, + 0.01660187914967537, + 0.017674071714282036, + 0.01155117992311716, + 0.019312875345349312, + 0.016274411231279373, + 0.011473267339169979, + 0.0037753190845251083, + 0.004348393529653549, + 0.006676708348095417, + 0.006759400479495525, + 0.004899266641587019, + 0.003931843210011721, + 0.0060103596188127995, + 0.002593187615275383, + 0.008231271989643574, + 0.014001142233610153, + 0.009911125525832176, + 0.009724183939397335, + 0.01919754222035408, + 0.014890078455209732, + 0.011885269545018673, + 0.010990278795361519, + 0.016585737466812134, + 0.010000423528254032, + 0.007896735332906246, + 0.011858740821480751, + 0.009170359000563622, + 0.00861172005534172, + 0.014294704422354698, + 0.0055489493533968925, + 0.008175777271389961, + 0.008328120224177837, + 0.005095798522233963, + 0.01050793007016182, + 0.007325653452426195, + 0.008090957067906857, + 0.009840321727097034, + 0.005228029098361731, + 0.007675571832805872, + 0.003603816730901599, + 0.003317160066217184, + 0.0034610475413501263, + 0.004746994003653526, + 0.005863370839506388, + 0.004228695761412382, + 0.004754253197461367, + 0.005230322480201721, + 0.005621292628347874, + 0.0026211480144411325, + 0.004527518525719643, + 0.003976271953433752, + 0.006117756478488445, + 0.005402870010584593, + 0.0034850575029850006, + 0.013549027033150196, + 0.008049515075981617, + 0.006547265220433474, + 0.009838217869400978, + 0.007142312359064817, + 0.026893267408013344, + 0.008297447115182877, + 0.013834710232913494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.02281453087925911 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329622581601143 + ], + "opposite_logprob_tokens": [ + -3.7732961177825928 + ], + "logprob_reward": -0.02329622581601143, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329622581601143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29023778438568115, + 0.05532879754900932, + 0.12908349931240082, + 0.13197244703769684, + 0.11657357215881348, + 0.037508901208639145, + 0.04405321925878525, + 0.09979745745658875, + 0.041196271777153015, + 0.05424799397587776 + ], + "credit_map": [ + 0.027652477845549583, + 0.025272231549024582, + 0.04499795660376549, + 0.025587299838662148, + 0.029111674055457115, + 0.06728927791118622, + 0.028601111844182014, + 0.0417257621884346, + 0.005946467164903879, + 0.008086717687547207, + 0.007941457442939281, + 0.007068075705319643, + 0.006724955514073372, + 0.008397514000535011, + 0.004371857736259699, + 0.006791750900447369, + 0.009591101668775082, + 0.01721298135817051, + 0.016223467886447906, + 0.016049783676862717, + 0.015212072990834713, + 0.027219058945775032, + 0.015169409103691578, + 0.012405632063746452, + 0.01216018758714199, + 0.012355305254459381, + 0.03381180763244629, + 0.018483556807041168, + 0.012706649489700794, + 0.013073672540485859, + 0.01298807468265295, + 0.016393182799220085, + 0.0100747374817729, + 0.016215506941080093, + 0.013791082426905632, + 0.015081321820616722, + 0.02034231275320053, + 0.015641218051314354, + 0.014763202518224716, + 0.010664187371730804, + 0.004890059120953083, + 0.005419893190264702, + 0.004012048244476318, + 0.005894556175917387, + 0.004133997950702906, + 0.00432162219658494, + 0.005613603629171848, + 0.003223121166229248, + 0.00517906853929162, + 0.005413285456597805, + 0.003624934470281005, + 0.006927993148565292, + 0.004961691331118345, + 0.004959046375006437, + 0.008785123005509377, + 0.004202080890536308, + 0.020632395520806313, + 0.010363459587097168, + 0.010862201452255249, + 0.01138642244040966, + 0.012431799434125423, + 0.010256132110953331, + 0.012213556095957756, + 0.011651492677628994, + 0.006458580493927002, + 0.007677784655243158, + 0.0021853779908269644, + 0.005307986866682768, + 0.005362227559089661, + 0.005334772169589996, + 0.006137263495475054, + 0.0027322801761329174, + 0.006154398899525404, + 0.009513230063021183, + 0.004063844680786133, + 0.008505456149578094, + 0.005851338151842356, + 0.007099214941263199, + 0.007430465891957283, + 0.00563004519790411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.02329622581601143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952014587819576 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005952014587819576, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952014587819576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19748491048812866, + 0.05267652869224548, + 0.16004525125026703, + 0.06185159459710121, + 0.13259464502334595, + 0.073941171169281, + 0.14798365533351898, + 0.06192751228809357, + 0.03606820106506348, + 0.07542657852172852 + ], + "credit_map": [ + 0.01629575714468956, + 0.015514892525970936, + 0.033577706664800644, + 0.029430748894810677, + 0.015507453121244907, + 0.036242105066776276, + 0.01951899379491806, + 0.0313972644507885, + 0.006375828292220831, + 0.007336833514273167, + 0.0066770827397704124, + 0.005702613387256861, + 0.006043161265552044, + 0.007888751104474068, + 0.004837949760258198, + 0.007814310491085052, + 0.0104284156113863, + 0.018534716218709946, + 0.02297164686024189, + 0.024413663893938065, + 0.014810333959758282, + 0.02873297594487667, + 0.021699875593185425, + 0.01845361851155758, + 0.005483257118612528, + 0.005797578487545252, + 0.009495006874203682, + 0.01066959835588932, + 0.007893508300185204, + 0.007271209266036749, + 0.010358565486967564, + 0.00488286791369319, + 0.010931141674518585, + 0.019107310101389885, + 0.01774163730442524, + 0.014993797056376934, + 0.021803949028253555, + 0.018833773210644722, + 0.017603104934096336, + 0.011579928919672966, + 0.014915674924850464, + 0.00912488717585802, + 0.008660814724862576, + 0.011520933359861374, + 0.007587422616779804, + 0.00779387541115284, + 0.008910796605050564, + 0.00542676355689764, + 0.018874842673540115, + 0.019047453999519348, + 0.01273495052009821, + 0.02507295459508896, + 0.01887868531048298, + 0.018089819699525833, + 0.02475881576538086, + 0.010526137426495552, + 0.012559094466269016, + 0.004980872385203838, + 0.006450919434428215, + 0.005851311609148979, + 0.0056160870008170605, + 0.011824366636574268, + 0.0059916796162724495, + 0.008653179742395878, + 0.005499956663697958, + 0.0049871718510985374, + 0.0029572336934506893, + 0.004341511055827141, + 0.003991363570094109, + 0.005882115103304386, + 0.004427431151270866, + 0.003981419838964939, + 0.010055575519800186, + 0.007872730493545532, + 0.004640025086700916, + 0.004021766595542431, + 0.006088360212743282, + 0.024051418527960777, + 0.006318374071270227, + 0.012378332205116749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.005952014587819576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019386084750294685 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019386084750294685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019386084750294685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21235397458076477, + 0.09544041007757187, + 0.0789952352643013, + 0.10663054883480072, + 0.06232760474085808, + 0.07673090696334839, + 0.15776053071022034, + 0.13612815737724304, + 0.04075184836983681, + 0.03288085386157036 + ], + "credit_map": [ + 0.015511464327573776, + 0.011310852132737637, + 0.06345086544752121, + 0.01468116044998169, + 0.011603823862969875, + 0.022929063066840172, + 0.012811212800443172, + 0.06005552411079407, + 0.007168784271925688, + 0.011630506254732609, + 0.0117550790309906, + 0.015182367525994778, + 0.008988851681351662, + 0.01159296091645956, + 0.015697509050369263, + 0.0134243443608284, + 0.005765006877481937, + 0.008634260855615139, + 0.013579810038208961, + 0.009458149783313274, + 0.006705283187329769, + 0.012772077694535255, + 0.009305051527917385, + 0.012775595299899578, + 0.008558571338653564, + 0.01132242288440466, + 0.01803458295762539, + 0.021299652755260468, + 0.012142322957515717, + 0.010722974315285683, + 0.015110718086361885, + 0.009439302608370781, + 0.006138993427157402, + 0.00932273082435131, + 0.00856691598892212, + 0.007705660071223974, + 0.0105040417984128, + 0.008377156220376492, + 0.0074172839522361755, + 0.004294825252145529, + 0.011370843276381493, + 0.01053010392934084, + 0.007103279698640108, + 0.013943243771791458, + 0.009538411162793636, + 0.00877826102077961, + 0.009998977184295654, + 0.005467787850648165, + 0.015208336524665356, + 0.0200492013245821, + 0.012919312343001366, + 0.028120581060647964, + 0.020953135564923286, + 0.020074132829904556, + 0.028598995879292488, + 0.011836840771138668, + 0.022410394623875618, + 0.01851525902748108, + 0.012098491191864014, + 0.018071942031383514, + 0.014979283325374126, + 0.017234085127711296, + 0.01898198574781418, + 0.013836711645126343, + 0.005482966545969248, + 0.006200819276273251, + 0.0034681984689086676, + 0.005579181481152773, + 0.004726613871753216, + 0.005821088794618845, + 0.00568807777017355, + 0.003784900065511465, + 0.005822184961289167, + 0.004324328154325485, + 0.002915632212534547, + 0.003968399949371815, + 0.0035178016405552626, + 0.004940430633723736, + 0.0038830717094242573, + 0.0035090043675154448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0019386084750294685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863604158163071 + ], + "opposite_logprob_tokens": [ + -3.048635959625244 + ], + "logprob_reward": -0.04863604158163071, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863604158163071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3670707941055298, + 0.05618688836693764, + 0.053319334983825684, + 0.08813648670911789, + 0.11260030418634415, + 0.07238162308931351, + 0.06564809381961823, + 0.10968413203954697, + 0.02722311206161976, + 0.04774923250079155 + ], + "credit_map": [ + 0.020599860697984695, + 0.0203452967107296, + 0.09617020189762115, + 0.041477881371974945, + 0.018058519810438156, + 0.05645056068897247, + 0.025075163692235947, + 0.08889330178499222, + 0.00484407227486372, + 0.0075890119187533855, + 0.006709185428917408, + 0.007735962048172951, + 0.00654960423707962, + 0.007655174471437931, + 0.005540027748793364, + 0.009563852101564407, + 0.003809436457231641, + 0.005561350379139185, + 0.00864648912101984, + 0.008227543905377388, + 0.0054769027046859264, + 0.008638259954750538, + 0.0075194635428488255, + 0.0054398830980062485, + 0.008279253728687763, + 0.009038212709128857, + 0.010425778105854988, + 0.01666557788848877, + 0.01395760104060173, + 0.009495879523456097, + 0.013204901479184628, + 0.0070692794397473335, + 0.011648941785097122, + 0.014222322963178158, + 0.011231554672122002, + 0.014290384948253632, + 0.017495926469564438, + 0.01696009375154972, + 0.015329227782785892, + 0.01142184529453516, + 0.010699553415179253, + 0.007805350236594677, + 0.0067738937214016914, + 0.01219931524246931, + 0.007831566035747528, + 0.007811251096427441, + 0.012062319554388523, + 0.0071983784437179565, + 0.008562371134757996, + 0.006218736059963703, + 0.006768218241631985, + 0.008024834096431732, + 0.0081507358700037, + 0.008968820795416832, + 0.011953442357480526, + 0.0070009310729801655, + 0.01611936464905739, + 0.012329263612627983, + 0.010172899812459946, + 0.016397256404161453, + 0.014675455167889595, + 0.013383681885898113, + 0.016410868614912033, + 0.010195351205766201, + 0.0047266860492527485, + 0.00515044666826725, + 0.0015211738646030426, + 0.0036984861362725496, + 0.003082717303186655, + 0.0040721832774579525, + 0.00274060876108706, + 0.002230808138847351, + 0.005275822710245848, + 0.008527599275112152, + 0.003289446933194995, + 0.004692575894296169, + 0.005140969064086676, + 0.009435705840587616, + 0.005987520795315504, + 0.005399592686444521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.04863604158163071 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a vibrant Pixar style, the scene begins with a high-angle shot, the camera slowly and gracefully descending toward a stylized map, magically realized on a textured, ancient scroll. The scroll, with its warm, sheepskin-yellow hue and delicately worn edges, unfolds within a soft, dark void, its surface shimmering with an inner light. Above, a winding tour route weaves through the terracotta excavation pits, rendered with bold, animated lines, connecting the charming pit icons against a landscape of saturated greens and earthy browns. As the camera smoothly advances, it focuses on a whimsically animated cargo ship, constructed from soft, rounded planks and a cheerful design, gliding steadily along a painted, bright blue river on the map. The ship's decks are meticulously arranged with rows of adorable miniature terracotta warriors, each with an oversized head and simplified, friendly features. The ship is deliberately moored in a clearly marked port; a small, stylized anchor hangs from the bow on a thin rope, and when it hits the painted river, it doesn't splash but instead sends up a delightful, vibrant puff of golden-brown dust that gently dissipates, perfectly capturing an atmosphere of fun and magical adventure.", + "seed": 15327, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_3718_seed_15327_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015168705605901778, + "vlm_logprob": -0.015168705605901778, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.015168705605901778, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01198295596987009 + ], + "opposite_logprob_tokens": [ + -6.1369829177856445 + ], + "logprob_reward": -0.01198295596987009, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.01198295596987009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0580291710793972, + 0.05524046719074249, + 0.02816343680024147, + 0.030969083309173584, + 0.1152300164103508, + 0.07915432006120682, + 0.20905162394046783, + 0.09248226135969162, + 0.2934458255767822, + 0.03823379427194595 + ], + "credit_map": [ + 0.00607292540371418, + 0.0080193392932415, + 0.0062387543730437756, + 0.006819719914346933, + 0.005462931469082832, + 0.006280702073127031, + 0.011150273494422436, + 0.00798452366143465, + 0.005649752449244261, + 0.005380809307098389, + 0.004752518609166145, + 0.008602077141404152, + 0.005098444409668446, + 0.0051788813434541225, + 0.013407614082098007, + 0.007170369848608971, + 0.0025772186927497387, + 0.004723520018160343, + 0.0030094340909272432, + 0.002984336344525218, + 0.0054551297798752785, + 0.002879049628973007, + 0.003366652177646756, + 0.003168096300214529, + 0.00318435812368989, + 0.0056215738877654076, + 0.0027765852864831686, + 0.0027244261000305414, + 0.005424289032816887, + 0.002971313428133726, + 0.004786000121384859, + 0.003480535699054599, + 0.012653356418013573, + 0.01567898690700531, + 0.01103572640568018, + 0.020479785278439522, + 0.011893403716385365, + 0.013493314385414124, + 0.021057888865470886, + 0.008937562815845013, + 0.004859793931245804, + 0.012720917351543903, + 0.011048679240047932, + 0.008361601270735264, + 0.009748145006597042, + 0.015623576939105988, + 0.011257988400757313, + 0.005533617455512285, + 0.022679045796394348, + 0.050700265914201736, + 0.0265685822814703, + 0.013620415702462196, + 0.038511011749506, + 0.026177454739809036, + 0.01261025108397007, + 0.018184594810009003, + 0.013336790725588799, + 0.024763189256191254, + 0.006362303160130978, + 0.008719075471162796, + 0.011621169745922089, + 0.016796426847577095, + 0.00639306241646409, + 0.004490246530622244, + 0.021154701709747314, + 0.014656135812401772, + 0.011999145150184631, + 0.07601971179246902, + 0.020749332383275032, + 0.056739386171102524, + 0.0807199478149414, + 0.011407461948692799, + 0.004581467714160681, + 0.00442973617464304, + 0.005023514851927757, + 0.003976421896368265, + 0.004045943729579449, + 0.0036118431016802788, + 0.004106804728507996, + 0.008458063006401062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.01198295596987009 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540768831968307 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540768831968307, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540768831968307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13254928588867188, + 0.14020048081874847, + 0.08012289553880692, + 0.04372056573629379, + 0.1689821183681488, + 0.03849581629037857, + 0.15923956036567688, + 0.0672670379281044, + 0.12083086371421814, + 0.04859137907624245 + ], + "credit_map": [ + 0.012451909482479095, + 0.012035001069307327, + 0.017294354736804962, + 0.020598577335476875, + 0.012905868701636791, + 0.014710387215018272, + 0.016317036002874374, + 0.02623615600168705, + 0.021636225283145905, + 0.019041111692786217, + 0.010123109444975853, + 0.017133817076683044, + 0.012106646783649921, + 0.017877697944641113, + 0.02807099185883999, + 0.014210877008736134, + 0.008503067307174206, + 0.009645297192037106, + 0.007538840174674988, + 0.00998293049633503, + 0.013648362830281258, + 0.00937565229833126, + 0.009353031404316425, + 0.012075717560946941, + 0.0031600375659763813, + 0.0055980621837079525, + 0.0031047319062054157, + 0.003637752030044794, + 0.012435934506356716, + 0.005337664391845465, + 0.005597492679953575, + 0.004848890472203493, + 0.014598513953387737, + 0.020677082240581512, + 0.014605946838855743, + 0.03418201580643654, + 0.015224194154143333, + 0.021489761769771576, + 0.03414658457040787, + 0.014058012515306473, + 0.0024086833000183105, + 0.0063340612687170506, + 0.0037490627728402615, + 0.00539338169619441, + 0.006112359464168549, + 0.005819764919579029, + 0.00608140230178833, + 0.0025971003342419863, + 0.01770167611539364, + 0.040751341730356216, + 0.016690826043486595, + 0.01179327443242073, + 0.030194954946637154, + 0.021479174494743347, + 0.012996825389564037, + 0.007631487213075161, + 0.008936932310461998, + 0.017629630863666534, + 0.005496606696397066, + 0.007786202244460583, + 0.009285597130656242, + 0.008318457752466202, + 0.0054083350114524364, + 0.004405276384204626, + 0.01197361946105957, + 0.006955469958484173, + 0.0051694512367248535, + 0.02433028258383274, + 0.009698556736111641, + 0.02972988598048687, + 0.025598108768463135, + 0.007375495508313179, + 0.00448254682123661, + 0.0037725649308413267, + 0.005891705863177776, + 0.004260791465640068, + 0.005087998230010271, + 0.010919079184532166, + 0.004048455040901899, + 0.010128235444426537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.012540768831968307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015124559868127108 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015124559868127108, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015124559868127108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12006150186061859, + 0.0654430240392685, + 0.06466095894575119, + 0.11662570387125015, + 0.07133328914642334, + 0.12919050455093384, + 0.15609271824359894, + 0.06760773807764053, + 0.13793042302131653, + 0.0710541158914566 + ], + "credit_map": [ + 0.013149730861186981, + 0.016375698149204254, + 0.013948618434369564, + 0.013680311851203442, + 0.015132238157093525, + 0.012437578290700912, + 0.018252138048410416, + 0.01708519086241722, + 0.006677062716335058, + 0.009003277868032455, + 0.005032728426158428, + 0.012185855768620968, + 0.005881255026906729, + 0.006425752304494381, + 0.013183094561100006, + 0.007053998298943043, + 0.005243618506938219, + 0.0073796906508505344, + 0.005217329133301973, + 0.0093416478484869, + 0.007534469477832317, + 0.0062245456501841545, + 0.011209164746105671, + 0.01251048780977726, + 0.010177863761782646, + 0.015708671882748604, + 0.008958774618804455, + 0.013319306075572968, + 0.018270017579197884, + 0.012460905127227306, + 0.027588825672864914, + 0.010141342878341675, + 0.009585882537066936, + 0.01094034407287836, + 0.009103999473154545, + 0.008117740042507648, + 0.008540799841284752, + 0.007956481538712978, + 0.009512235410511494, + 0.007575803902000189, + 0.010751093737781048, + 0.02041022665798664, + 0.019215388223528862, + 0.019094059243798256, + 0.014462582767009735, + 0.014801555313169956, + 0.01902282051742077, + 0.011432778090238571, + 0.013966711238026619, + 0.028583163395524025, + 0.03138559311628342, + 0.010978324338793755, + 0.022854220122098923, + 0.02535170316696167, + 0.016124555841088295, + 0.006848454009741545, + 0.010604562237858772, + 0.008629604242742062, + 0.005872350186109543, + 0.009403944946825504, + 0.010540545918047428, + 0.007756482809782028, + 0.008483247831463814, + 0.006316997576504946, + 0.00975481141358614, + 0.009144292213022709, + 0.006520441733300686, + 0.027856381610035896, + 0.017347684130072594, + 0.036523718386888504, + 0.024748671799898148, + 0.006034412421286106, + 0.006698478478938341, + 0.007769554387778044, + 0.008619215339422226, + 0.005871099885553122, + 0.007007213309407234, + 0.017981750890612602, + 0.005985135678201914, + 0.011121670715510845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.0015124559868127108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177094760350883 + ], + "opposite_logprob_tokens": [ + -6.7511773109436035 + ], + "logprob_reward": -0.001177094760350883, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177094760350883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0534895583987236, + 0.02560189552605152, + 0.0664280578494072, + 0.034297823905944824, + 0.04500744491815567, + 0.057541634887456894, + 0.04786592721939087, + 0.03942522406578064, + 0.5896431803703308, + 0.04069928079843521 + ], + "credit_map": [ + 0.0056149461306631565, + 0.006608475930988789, + 0.006470377556979656, + 0.006959724240005016, + 0.004983585327863693, + 0.005981730297207832, + 0.010092961601912975, + 0.006777752190828323, + 0.003182365093380213, + 0.002761953044682741, + 0.0020791911520063877, + 0.004284789320081472, + 0.0027005125302821398, + 0.0025288916658610106, + 0.004609689116477966, + 0.003454504068940878, + 0.0045096976682543755, + 0.00839305017143488, + 0.00648913299664855, + 0.008855738677084446, + 0.010227998718619347, + 0.00765142822638154, + 0.009176312945783138, + 0.011124690994620323, + 0.003873850218951702, + 0.004194905981421471, + 0.0032049475703388453, + 0.002805965719744563, + 0.004998394288122654, + 0.003708440577611327, + 0.005563360173255205, + 0.005947954952716827, + 0.00412128446623683, + 0.006692762020975351, + 0.005377370864152908, + 0.007952100597321987, + 0.0040108272805809975, + 0.006034851539880037, + 0.006892650853842497, + 0.0039255935698747635, + 0.003723699366673827, + 0.010945267044007778, + 0.00692392373457551, + 0.007019364275038242, + 0.007257772609591484, + 0.01198533084243536, + 0.006256094668060541, + 0.0034301821142435074, + 0.004307789262384176, + 0.0028743110597133636, + 0.019191600382328033, + 0.00319014023989439, + 0.0023076306097209454, + 0.008968319743871689, + 0.002100822515785694, + 0.004925309214740992, + 0.004807681310921907, + 0.00479364488273859, + 0.0038782437331974506, + 0.007169523276388645, + 0.005757544655352831, + 0.004117279779165983, + 0.005555144045501947, + 0.003346159588545561, + 0.04533010721206665, + 0.028479186818003654, + 0.02308439463376999, + 0.1412835270166397, + 0.04305697977542877, + 0.1318380981683731, + 0.15245284140110016, + 0.02411801367998123, + 0.005042182747274637, + 0.005588695406913757, + 0.0034297448582947254, + 0.002687255386263132, + 0.004966894164681435, + 0.0112839974462986, + 0.0038917020428925753, + 0.003808803390711546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.001177094760350883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486302524805069 + ], + "opposite_logprob_tokens": [ + -3.0486302375793457 + ], + "logprob_reward": -0.0486302524805069, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486302524805069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08581022173166275, + 0.026926636695861816, + 0.035142190754413605, + 0.07539688795804977, + 0.09861655533313751, + 0.10641112178564072, + 0.20389120280742645, + 0.09891551733016968, + 0.17898039519786835, + 0.08990925550460815 + ], + "credit_map": [ + 0.01096265111118555, + 0.01025627925992012, + 0.008348396047949791, + 0.00881156325340271, + 0.008997704833745956, + 0.018649253994226456, + 0.010719106532633305, + 0.00906527228653431, + 0.004107500892132521, + 0.0037881038151681423, + 0.002349733142182231, + 0.003534784074872732, + 0.002362349070608616, + 0.0037015737034380436, + 0.004472475498914719, + 0.002610116731375456, + 0.0029934081248939037, + 0.00535622239112854, + 0.0032604404259473085, + 0.004850790835916996, + 0.006087937392294407, + 0.0036945699248462915, + 0.004259373992681503, + 0.004639447666704655, + 0.006050858646631241, + 0.008097285404801369, + 0.004354776348918676, + 0.00855652242898941, + 0.015168110840022564, + 0.009238884784281254, + 0.01721416972577572, + 0.006716277450323105, + 0.011322264559566975, + 0.01378433033823967, + 0.009523040615022182, + 0.017038075253367424, + 0.011042637750506401, + 0.012191230431199074, + 0.016558952629566193, + 0.007156027015298605, + 0.007669604849070311, + 0.016735613346099854, + 0.0138741759583354, + 0.01148358453065157, + 0.012413877062499523, + 0.022497229278087616, + 0.013778502121567726, + 0.007958539761602879, + 0.019811447709798813, + 0.05092230439186096, + 0.022707747295498848, + 0.014735008589923382, + 0.04157936945557594, + 0.028163889423012733, + 0.015684280544519424, + 0.010287152603268623, + 0.014004895463585854, + 0.024917710572481155, + 0.006908860523253679, + 0.010434400290250778, + 0.011944286525249481, + 0.009333676658570766, + 0.013312412425875664, + 0.008059275336563587, + 0.011386923491954803, + 0.00892557017505169, + 0.00756104988977313, + 0.05058877915143967, + 0.012401951476931572, + 0.027243562042713165, + 0.05300017073750496, + 0.007872387766838074, + 0.009619275107979774, + 0.008937743492424488, + 0.007997835986316204, + 0.007586461491882801, + 0.01050218939781189, + 0.025149056687951088, + 0.009832379408180714, + 0.010284313932061195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.0486302524805069 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a vibrant Pixar style, the scene begins with a high-angle shot, the camera slowly and gracefully descending toward a stylized map, magically realized on a textured, ancient scroll. The scroll, with its warm, sheepskin-yellow hue and delicately worn edges, unfolds within a soft, dark void, its surface shimmering with an inner light. Above, a winding tour route weaves through the terracotta excavation pits, rendered with bold, animated lines, connecting the charming pit icons against a landscape of saturated greens and earthy browns. As the camera smoothly advances, it focuses on a whimsically animated cargo ship, constructed from soft, rounded planks and a cheerful design, gliding steadily along a painted, bright blue river on the map. The ship's decks are meticulously arranged with rows of adorable miniature terracotta warriors, each with an oversized head and simplified, friendly features. The ship is deliberately moored in a clearly marked port; a small, stylized anchor hangs from the bow on a thin rope, and when it hits the painted river, it doesn't splash but instead sends up a delightful, vibrant puff of golden-brown dust that gently dissipates, perfectly capturing an atmosphere of fun and magical adventure.", + "seed": 15328, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_10/index_3718_seed_15328_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020588148664683104, + "vlm_logprob": -0.020588148664683104, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.020588148664683104, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015246162191033363 + ], + "opposite_logprob_tokens": [ + -4.890246391296387 + ], + "logprob_reward": -0.015246162191033363, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.015246162191033363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07616226375102997, + 0.23775392770767212, + 0.12704189121723175, + 0.10395629703998566, + 0.018880097195506096, + 0.12299694120883942, + 0.04342694953083992, + 0.13137342035770416, + 0.07735872268676758, + 0.061049509793519974 + ], + "credit_map": [ + 0.009054997935891151, + 0.008182434365153313, + 0.008770924992859364, + 0.009477498009800911, + 0.006404411979019642, + 0.00802281592041254, + 0.0112734604626894, + 0.01497572846710682, + 0.020855341106653214, + 0.018104540184140205, + 0.06187662482261658, + 0.021455006673932076, + 0.028372228145599365, + 0.0177492406219244, + 0.02114463597536087, + 0.048196304589509964, + 0.025549380108714104, + 0.029317909851670265, + 0.020891869440674782, + 0.008370435796678066, + 0.013208230026066303, + 0.01088396180421114, + 0.009681317023932934, + 0.009138782508671284, + 0.029993977397680283, + 0.023221276700496674, + 0.006790084298700094, + 0.006854229606688023, + 0.0067211939021945, + 0.010756347328424454, + 0.004346862435340881, + 0.015272319316864014, + 0.002295455429702997, + 0.001906682038679719, + 0.002145129255950451, + 0.002461383817717433, + 0.0022161679808050394, + 0.0023229310754686594, + 0.003262785729020834, + 0.002269561169669032, + 0.0317019447684288, + 0.007927633821964264, + 0.006647328846156597, + 0.007157139014452696, + 0.008315077051520348, + 0.01654791831970215, + 0.007464074995368719, + 0.03723582997918129, + 0.005732349120080471, + 0.0049570887349545956, + 0.005543964449316263, + 0.005499804392457008, + 0.007084719371050596, + 0.006343530025333166, + 0.003870381973683834, + 0.004395115189254284, + 0.01622549630701542, + 0.014077140018343925, + 0.01854931004345417, + 0.02036851830780506, + 0.012071422301232815, + 0.022585226222872734, + 0.014735550619661808, + 0.012760755605995655, + 0.008677417412400246, + 0.00775974290445447, + 0.010342564433813095, + 0.012703333050012589, + 0.008337649516761303, + 0.008370202966034412, + 0.010382276028394699, + 0.010785532183945179, + 0.00612758519127965, + 0.008006751537322998, + 0.00812053307890892, + 0.009632358327507973, + 0.010490613989531994, + 0.007048054598271847, + 0.006398070137947798, + 0.0052255382761359215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.015246162191033363 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029791494831442833 + ], + "opposite_logprob_tokens": [ + -3.5297915935516357 + ], + "logprob_reward": -0.029791494831442833, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029791494831442833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13921619951725006, + 0.3225914537906647, + 0.06226535141468048, + 0.13263286650180817, + 0.024488775059580803, + 0.14467386901378632, + 0.045179471373558044, + 0.057906411588191986, + 0.02781638503074646, + 0.04322919622063637 + ], + "credit_map": [ + 0.014907157979905605, + 0.014441106468439102, + 0.02528294362127781, + 0.017825458198785782, + 0.013087913393974304, + 0.013594057410955429, + 0.02823171392083168, + 0.011845837347209454, + 0.024775715544819832, + 0.028727460652589798, + 0.08039926737546921, + 0.03657438978552818, + 0.03596310317516327, + 0.021179286763072014, + 0.03332463651895523, + 0.061647579073905945, + 0.008479726500809193, + 0.015260783024132252, + 0.004842516966164112, + 0.005119826178997755, + 0.004973460920155048, + 0.005265102256089449, + 0.004879160784184933, + 0.013444776646792889, + 0.013931402005255222, + 0.02072211541235447, + 0.015989772975444794, + 0.012096100486814976, + 0.013275468721985817, + 0.02255389280617237, + 0.00993907731026411, + 0.024125030264258385, + 0.0027681973297148943, + 0.002038778970018029, + 0.0028646665159612894, + 0.004289377015084028, + 0.003407243872061372, + 0.003163616405799985, + 0.002955291187390685, + 0.0030016042292118073, + 0.03692816197872162, + 0.009877807460725307, + 0.016696935519576073, + 0.016101723536849022, + 0.007606277242302895, + 0.011815211735665798, + 0.012464047409594059, + 0.033183708786964417, + 0.0066347611136734486, + 0.005635720677673817, + 0.004210252780467272, + 0.0037125099916011095, + 0.005419413559138775, + 0.007001325022429228, + 0.004555422812700272, + 0.008010064251720905, + 0.009315199218690395, + 0.005414440296590328, + 0.007485909853130579, + 0.008601474575698376, + 0.005794061813503504, + 0.007845297455787659, + 0.006007092539221048, + 0.007442936301231384, + 0.003198325866833329, + 0.004764570854604244, + 0.004235782660543919, + 0.003482369240373373, + 0.004294644109904766, + 0.0027203389909118414, + 0.002604815876111388, + 0.0025155378971248865, + 0.00433691218495369, + 0.0044588553719222546, + 0.006633092649281025, + 0.007407477591186762, + 0.007606219500303268, + 0.0042779529467225075, + 0.004775685258209705, + 0.0037329993210732937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.029791494831442833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019386084750294685 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019386084750294685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019386084750294685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08513545989990234, + 0.4814926087856293, + 0.10362511873245239, + 0.057392653077840805, + 0.032216932624578476, + 0.051104944199323654, + 0.03189205378293991, + 0.05754590779542923, + 0.01812436804175377, + 0.08146996051073074 + ], + "credit_map": [ + 0.008411559276282787, + 0.007392671424895525, + 0.0071416995488107204, + 0.010197879746556282, + 0.005903339479118586, + 0.009796745143830776, + 0.012649602256715298, + 0.023641962558031082, + 0.03481977432966232, + 0.03349831700325012, + 0.08778849244117737, + 0.04405250772833824, + 0.047183044254779816, + 0.0818093791604042, + 0.07181297242641449, + 0.08052811026573181, + 0.02198922447860241, + 0.017552141100168228, + 0.01494551356881857, + 0.009165257215499878, + 0.010249579325318336, + 0.011986665427684784, + 0.007237354293465614, + 0.010499371215701103, + 0.007010788191109896, + 0.008376522921025753, + 0.005598265677690506, + 0.0039781229570508, + 0.004039796069264412, + 0.010001055896282196, + 0.002792340936139226, + 0.015595756471157074, + 0.0037182241212576628, + 0.002935531083494425, + 0.0037338261026889086, + 0.003519823309034109, + 0.00493142893537879, + 0.0039054080843925476, + 0.0033811472821980715, + 0.006091543473303318, + 0.012200593948364258, + 0.003099178895354271, + 0.0027004394214600325, + 0.0027199320029467344, + 0.0031478696037083864, + 0.006552883889526129, + 0.0037903538905084133, + 0.016893692314624786, + 0.004239862784743309, + 0.0033793405164033175, + 0.0027043207082897425, + 0.0029899724759161472, + 0.004059475846588612, + 0.00569012900814414, + 0.004009346477687359, + 0.004819605499505997, + 0.009998833760619164, + 0.005716586019843817, + 0.006269881967455149, + 0.0067958165891468525, + 0.005627351813018322, + 0.011517911218106747, + 0.005417786072939634, + 0.0062017361633479595, + 0.00258920481428504, + 0.0020736120641231537, + 0.0025456654839217663, + 0.002720460994169116, + 0.002425603801384568, + 0.001659590401686728, + 0.0017457532230764627, + 0.0023644776083528996, + 0.009973232634365559, + 0.010943513363599777, + 0.009092215448617935, + 0.013148922473192215, + 0.013766517862677574, + 0.007912777364253998, + 0.009352498687803745, + 0.007280274759978056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0019386084750294685 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010387268848717213 + ], + "opposite_logprob_tokens": [ + -6.876038551330566 + ], + "logprob_reward": -0.0010387268848717213, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010387268848717213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1620468646287918, + 0.16608364880084991, + 0.18387649953365326, + 0.08066301047801971, + 0.03424793481826782, + 0.143403097987175, + 0.04511090740561485, + 0.057401590049266815, + 0.0513329803943634, + 0.07583343982696533 + ], + "credit_map": [ + 0.01540431473404169, + 0.014830539934337139, + 0.02691388688981533, + 0.020358415320515633, + 0.010213902220129967, + 0.019973568618297577, + 0.03591476008296013, + 0.018437469378113747, + 0.012718813493847847, + 0.012994700111448765, + 0.0393366664648056, + 0.013748315162956715, + 0.01833309419453144, + 0.027127087116241455, + 0.014935842715203762, + 0.0268891341984272, + 0.03138600289821625, + 0.047582078725099564, + 0.024920262396335602, + 0.014619190245866776, + 0.0161207914352417, + 0.018438829109072685, + 0.015339729376137257, + 0.015469598583877087, + 0.015327392145991325, + 0.010739371180534363, + 0.010654889047145844, + 0.007619483862072229, + 0.006355999503284693, + 0.014211962930858135, + 0.004681412596255541, + 0.011072498746216297, + 0.0055753751657903194, + 0.003174465149641037, + 0.0033781894017010927, + 0.004341952968388796, + 0.0030355683993548155, + 0.003686149138957262, + 0.003677075495943427, + 0.007379160262644291, + 0.038202304393053055, + 0.008377108722925186, + 0.009213977493345737, + 0.010663261637091637, + 0.006945135071873665, + 0.014893503859639168, + 0.008594833314418793, + 0.04651297628879547, + 0.006423467304557562, + 0.004479894880205393, + 0.0058458345010876656, + 0.0048472401686012745, + 0.00690061878412962, + 0.007012390065938234, + 0.003938067704439163, + 0.005663389805704355, + 0.008236591704189777, + 0.00790177471935749, + 0.006263645365834236, + 0.008743244223296642, + 0.00724034896120429, + 0.008896290324628353, + 0.004233119543641806, + 0.005886574275791645, + 0.007616916671395302, + 0.005756731145083904, + 0.009137575514614582, + 0.006919702980667353, + 0.005918309558182955, + 0.006525079254060984, + 0.0056831189431250095, + 0.003775546792894602, + 0.007601489778608084, + 0.010731120593845844, + 0.008974358439445496, + 0.011314693838357925, + 0.014383852481842041, + 0.007712106686085463, + 0.009036805480718613, + 0.0060790181159973145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0010387268848717213 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and consistent with a Pixar style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene begin with a high-angle shot that slowly and gracefully descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a stylized map rendered on a textured, ancient scroll with a warm, sheepskin-yellow hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a whimsically animated cargo ship constructed from soft, rounded planks gliding along a painted, bright blue river on the map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship's anchor send up a vibrant puff of golden-brown dust instead of a splash when it hits the river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492575094103813 + ], + "opposite_logprob_tokens": [ + -2.9299256801605225 + ], + "logprob_reward": -0.05492575094103813, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492575094103813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056911490857601166, + 0.21063290536403656, + 0.15334583818912506, + 0.16898469626903534, + 0.030052974820137024, + 0.13472586870193481, + 0.045808080583810806, + 0.04669448733329773, + 0.03773541375994682, + 0.11510828882455826 + ], + "credit_map": [ + 0.006659484468400478, + 0.0060704960487782955, + 0.006117729935795069, + 0.007137765642255545, + 0.004513543099164963, + 0.0071508814580738544, + 0.010790490545332432, + 0.008471105247735977, + 0.013084695674479008, + 0.011330519802868366, + 0.060265522450208664, + 0.018952421844005585, + 0.026345256716012955, + 0.012991282157599926, + 0.021655330434441566, + 0.046007897704839706, + 0.02684125490486622, + 0.04591480642557144, + 0.019715571776032448, + 0.008994163945317268, + 0.014704728499054909, + 0.011426459066569805, + 0.009299610741436481, + 0.016449254006147385, + 0.024580063298344612, + 0.0394534207880497, + 0.016908880323171616, + 0.010384749621152878, + 0.012232039123773575, + 0.022861823439598083, + 0.010671897791326046, + 0.03189181908965111, + 0.003470181953161955, + 0.0028672944754362106, + 0.0036704193335026503, + 0.0036677857860922813, + 0.0040449947118759155, + 0.0045728557743132114, + 0.0038564323913306, + 0.0039030108600854874, + 0.02874538116157055, + 0.007270310074090958, + 0.009553901851177216, + 0.008695347234606743, + 0.007846925407648087, + 0.01767347753047943, + 0.00970454327762127, + 0.045235976576805115, + 0.006203839555382729, + 0.0056654284708201885, + 0.004474300891160965, + 0.004097788129001856, + 0.006576398387551308, + 0.007087919861078262, + 0.004768769256770611, + 0.006933634635061026, + 0.007016798947006464, + 0.004080385435372591, + 0.004471984226256609, + 0.0049969591200351715, + 0.004163803998380899, + 0.013885660097002983, + 0.0033755411859601736, + 0.004703357350081205, + 0.005562259815633297, + 0.005120911169797182, + 0.004135308787226677, + 0.005682751536369324, + 0.005602086894214153, + 0.003147557144984603, + 0.0038273108657449484, + 0.004657226148992777, + 0.015354071743786335, + 0.014749659225344658, + 0.01212703250348568, + 0.01842746138572693, + 0.017413903027772903, + 0.010354680009186268, + 0.013382740318775177, + 0.013298746198415756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.05492575094103813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_11/index_tensor([7978, 7978])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_11/index_tensor([7978, 7978])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4f1a600b2606a29c25ec12a568ce9fc6b5efa4d7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_11/index_tensor([7978, 7978])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle shot, like viewing through the powerful lens of a deep-space telescope, captures the majestic and sweeping celestial panorama of Saturn, the sixth planet from the Sun. This massive gas giant glows a soft, yellowish hue, its thin, hazy atmosphere subtly zoned, allowing distant sunlight to penetrate its upper layers. Its defining feature is a vast, intricate ring system that prominently encircles the planet's equator, tilted slightly toward the viewer. Composed of countless shimmering ice and rock particles, these rings are rendered in surreal detail, revealing unique concentric bands and gaps that create mesmerizing patterns of light and shadow. Sunlight catches the ice particles, giving the rings a silvery-white glow, while the planet casts a long, curved shadow across its vast plane. The luminous planet and its rings stand in stark contrast to the stark, absolute blackness of outer space, their light piercing the void. This tilted, slightly overhead perspective offers a complete, unobstructed view, emphasizing the celestial body's silent, awe-inspiring grandeur and creating a tranquil yet powerful atmosphere of cosmic discovery.", + "seed": 14042, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_11/index_7978_seed_14042_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015882928436622024, + "vlm_logprob": -0.015882928436622024, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.015882928436622024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005138286389410496 + ], + "opposite_logprob_tokens": [ + -5.630138397216797 + ], + "logprob_reward": -0.005138286389410496, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.005138286389410496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15055257081985474, + 0.0589943528175354, + 0.05030741170048714, + 0.061125703155994415, + 0.0758926048874855, + 0.05950070917606354, + 0.2822829484939575, + 0.0916510745882988, + 0.03946249186992645, + 0.13023017346858978 + ], + "credit_map": [ + 0.011680688709020615, + 0.018087521195411682, + 0.013792638666927814, + 0.022487839683890343, + 0.023162437602877617, + 0.01679084077477455, + 0.022020677104592323, + 0.022529929876327515, + 0.004821192938834429, + 0.005949561949819326, + 0.010889967903494835, + 0.0092460373416543, + 0.007252623327076435, + 0.003988869022578001, + 0.008233548142015934, + 0.008612553589046001, + 0.003556684358045459, + 0.00428934209048748, + 0.0074945357628166676, + 0.00990646705031395, + 0.005182970315217972, + 0.004235879052430391, + 0.007661107927560806, + 0.007980423048138618, + 0.0034714231733232737, + 0.006601247936487198, + 0.006973083596676588, + 0.010944684967398643, + 0.008037948980927467, + 0.005646152421832085, + 0.009093624539673328, + 0.010357535444200039, + 0.0023130225017666817, + 0.006847524084150791, + 0.0072807795368134975, + 0.022145209833979607, + 0.006910429336130619, + 0.003909962717443705, + 0.017769061028957367, + 0.008716616779565811, + 0.005318025127053261, + 0.008642837405204773, + 0.0053227925673127174, + 0.006063760258257389, + 0.010936981067061424, + 0.006846682634204626, + 0.008242336101830006, + 0.008127296343445778, + 0.012905295938253403, + 0.0312684066593647, + 0.017476331442594528, + 0.02099636197090149, + 0.027203097939491272, + 0.10510995984077454, + 0.018501954153180122, + 0.04882154241204262, + 0.008693713694810867, + 0.012274465523660183, + 0.007206877693533897, + 0.015512816607952118, + 0.012664152309298515, + 0.010514848865568638, + 0.0174301378428936, + 0.007354065775871277, + 0.005199060309678316, + 0.005269698332995176, + 0.004327352624386549, + 0.003929965663701296, + 0.005783302243798971, + 0.0038736453279852867, + 0.00632483372464776, + 0.004754630383104086, + 0.009755476377904415, + 0.010203218087553978, + 0.007043428719043732, + 0.005054082255810499, + 0.04613522067666054, + 0.007744456175714731, + 0.008349497802555561, + 0.035944800823926926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.005138286389410496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00040082051418721676 + ], + "opposite_logprob_tokens": [ + -7.875401020050049 + ], + "logprob_reward": -0.00040082051418721676, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.00040082051418721676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10907438397407532, + 0.019388442859053612, + 0.04009781405329704, + 0.050003763288259506, + 0.05086366832256317, + 0.07638484984636307, + 0.12024138867855072, + 0.40285632014274597, + 0.030915517359972, + 0.10017379373311996 + ], + "credit_map": [ + 0.0077535230666399, + 0.012267806567251682, + 0.012665060348808765, + 0.018217051401734352, + 0.015921276062726974, + 0.010387236252427101, + 0.015465697273612022, + 0.01639673300087452, + 0.0016674918588250875, + 0.001957858446985483, + 0.002587382448837161, + 0.003391475649550557, + 0.0022603680845350027, + 0.001676854444667697, + 0.00295240618288517, + 0.002894607139751315, + 0.0026904018595814705, + 0.0037285888101905584, + 0.005548755172640085, + 0.007970265112817287, + 0.004762320313602686, + 0.0030359153170138597, + 0.0056626019068062305, + 0.006698965094983578, + 0.003112259553745389, + 0.006638348568230867, + 0.005650151055306196, + 0.007448198739439249, + 0.007777937687933445, + 0.004031770396977663, + 0.0065717631950974464, + 0.00877333153039217, + 0.0030795326456427574, + 0.004753970541059971, + 0.003942341543734074, + 0.006990443915128708, + 0.007710980251431465, + 0.004026815760880709, + 0.01203132327646017, + 0.008328264579176903, + 0.007915469817817211, + 0.010032524354755878, + 0.005811381619423628, + 0.008504443801939487, + 0.01602041907608509, + 0.009928543120622635, + 0.009209704585373402, + 0.008962368592619896, + 0.00721722049638629, + 0.015620443038642406, + 0.009422038681805134, + 0.010366911068558693, + 0.018073629587888718, + 0.03536534309387207, + 0.009093720465898514, + 0.015082092024385929, + 0.016256969422101974, + 0.016271376982331276, + 0.018210487440228462, + 0.11232342571020126, + 0.032316576689481735, + 0.03669976815581322, + 0.1103818342089653, + 0.06039590761065483, + 0.0027286666445434093, + 0.0050118472427129745, + 0.003511655842885375, + 0.004848827142268419, + 0.004173275548964739, + 0.002516265958547592, + 0.004890819080173969, + 0.00323416106402874, + 0.005993745289742947, + 0.014648683369159698, + 0.008061448112130165, + 0.0046723647974431515, + 0.030732451006770134, + 0.009695727378129959, + 0.005644160322844982, + 0.02072521485388279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.00040082051418721676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647525027394295 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647525027394295, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647525027394295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12467360496520996, + 0.04274881258606911, + 0.11018284410238266, + 0.09649838507175446, + 0.09971650689840317, + 0.09777523577213287, + 0.1479984074831009, + 0.0704101175069809, + 0.09354287385940552, + 0.11645318567752838 + ], + "credit_map": [ + 0.007779579143971205, + 0.014486822299659252, + 0.015023204497992992, + 0.02107490971684456, + 0.01707671955227852, + 0.013230955228209496, + 0.017694097012281418, + 0.01830732636153698, + 0.0038094634655863047, + 0.00539777847006917, + 0.005782960914075375, + 0.007302148267626762, + 0.006063420791178942, + 0.0033977862913161516, + 0.005226305685937405, + 0.005768947768956423, + 0.007466975133866072, + 0.007813036441802979, + 0.016839154064655304, + 0.023046577349305153, + 0.011397196911275387, + 0.00842037983238697, + 0.01852147839963436, + 0.016678044572472572, + 0.00607592798769474, + 0.012095671147108078, + 0.009639922529459, + 0.01345139928162098, + 0.0166445542126894, + 0.0080207334831357, + 0.012447443790733814, + 0.018122734501957893, + 0.004716300871223211, + 0.01006527990102768, + 0.0067055439576506615, + 0.022368744015693665, + 0.011308363638818264, + 0.005258820950984955, + 0.025994131341576576, + 0.013299321755766869, + 0.010384510271251202, + 0.015752365812659264, + 0.007147401105612516, + 0.010504701174795628, + 0.017781708389520645, + 0.009852292947471142, + 0.01347903162240982, + 0.012873224914073944, + 0.010449320077896118, + 0.019133374094963074, + 0.01090959832072258, + 0.010229135863482952, + 0.01603706367313862, + 0.04573979601264, + 0.010789474472403526, + 0.02471064031124115, + 0.004811975639313459, + 0.008971835486590862, + 0.005533894523978233, + 0.011824316345155239, + 0.009606384672224522, + 0.00942855142056942, + 0.014510206878185272, + 0.0057229530066251755, + 0.008368181064724922, + 0.016753418371081352, + 0.010991360992193222, + 0.009159545414149761, + 0.013420973904430866, + 0.009465849958360195, + 0.015520477667450905, + 0.009863067418336868, + 0.008042139001190662, + 0.011276975274085999, + 0.008385485969483852, + 0.0051375022158026695, + 0.03688008338212967, + 0.006433149799704552, + 0.008649752475321293, + 0.03164809197187424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.004647525027394295 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864955693483353 + ], + "opposite_logprob_tokens": [ + -3.048649549484253 + ], + "logprob_reward": -0.04864955693483353, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864955693483353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12195704877376556, + 0.025540189817547798, + 0.039924297481775284, + 0.06401212513446808, + 0.10690343379974365, + 0.02896432764828205, + 0.031769782304763794, + 0.30721068382263184, + 0.06404825299978256, + 0.2096698135137558 + ], + "credit_map": [ + 0.009854929521679878, + 0.013180678710341454, + 0.014080826193094254, + 0.01955544948577881, + 0.0174162108451128, + 0.013515672646462917, + 0.017288707196712494, + 0.017064576968550682, + 0.0021536468993872404, + 0.0029761490877717733, + 0.0030916116666048765, + 0.004167763516306877, + 0.0038116255309432745, + 0.0022293247748166323, + 0.0034863303881138563, + 0.003623738419264555, + 0.0026715679559856653, + 0.0028663750272244215, + 0.00701261917129159, + 0.008717344142496586, + 0.0033255419693887234, + 0.0028120637871325016, + 0.006517868489027023, + 0.006000918336212635, + 0.003607293125241995, + 0.00793907605111599, + 0.005300361197441816, + 0.009248371236026287, + 0.010061739943921566, + 0.005991452839225531, + 0.00884569063782692, + 0.013018136844038963, + 0.0057162148877978325, + 0.014333131723105907, + 0.006537908222526312, + 0.010325233452022076, + 0.024377262219786644, + 0.012418280355632305, + 0.018482152372598648, + 0.014713255688548088, + 0.0039724078960716724, + 0.0033581696916371584, + 0.0017802630318328738, + 0.0028004280757158995, + 0.003948418889194727, + 0.004164821933954954, + 0.005044123157858849, + 0.00389569578692317, + 0.004414776340126991, + 0.004248770885169506, + 0.002732539549469948, + 0.004065872635692358, + 0.004916232544928789, + 0.002352067967876792, + 0.004902836866676807, + 0.004136682488024235, + 0.012174122966825962, + 0.012513762339949608, + 0.014124898239970207, + 0.08559402823448181, + 0.02649122290313244, + 0.025972312316298485, + 0.08322715759277344, + 0.04711318761110306, + 0.005117540247738361, + 0.011254687793552876, + 0.006607862189412117, + 0.010066901333630085, + 0.011621545068919659, + 0.00606598611921072, + 0.00640149088576436, + 0.006912245415151119, + 0.017802631482481956, + 0.029328063130378723, + 0.01382728572934866, + 0.01388471107929945, + 0.054566118866205215, + 0.01753995195031166, + 0.014745991677045822, + 0.04797505959868431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.04864955693483353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578453317284584 + ], + "opposite_logprob_tokens": [ + -3.895578384399414 + ], + "logprob_reward": -0.020578453317284584, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578453317284584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08346934616565704, + 0.02368481084704399, + 0.08472079783678055, + 0.12604199349880219, + 0.09798523038625717, + 0.2925463318824768, + 0.07625942677259445, + 0.1581035852432251, + 0.02268604002892971, + 0.034502360969781876 + ], + "credit_map": [ + 0.005689677316695452, + 0.008749767206609249, + 0.014513799920678139, + 0.016231995075941086, + 0.00905918050557375, + 0.005664662458002567, + 0.010833372361958027, + 0.012726910412311554, + 0.002711554756388068, + 0.002952004550024867, + 0.0026921562384814024, + 0.004640651401132345, + 0.003009817795827985, + 0.0014073364436626434, + 0.002827025717124343, + 0.003444268135353923, + 0.005997856147587299, + 0.0063955639488995075, + 0.013005216605961323, + 0.017307093366980553, + 0.009266780689358711, + 0.005957378074526787, + 0.013515466824173927, + 0.013275462202727795, + 0.009364934638142586, + 0.01550219114869833, + 0.013801053166389465, + 0.021489311009645462, + 0.017312897369265556, + 0.010147852823138237, + 0.01650569587945938, + 0.021918073296546936, + 0.00789317861199379, + 0.012711926363408566, + 0.008328024297952652, + 0.02016407810151577, + 0.012964616529643536, + 0.007811201270669699, + 0.01523128617554903, + 0.012880939058959484, + 0.025779936462640762, + 0.05805029720067978, + 0.0224442221224308, + 0.03465644642710686, + 0.054812438786029816, + 0.03392857313156128, + 0.031015822663903236, + 0.031858645379543304, + 0.006675330922007561, + 0.011233152821660042, + 0.006635231897234917, + 0.005555096548050642, + 0.011516598984599113, + 0.016848638653755188, + 0.006847342010587454, + 0.01094804983586073, + 0.009100320748984814, + 0.014168997295200825, + 0.009724137373268604, + 0.03870118781924248, + 0.01262516900897026, + 0.01825360208749771, + 0.04057101905345917, + 0.014959188178181648, + 0.003017744515091181, + 0.0026515356730669737, + 0.0027844354044646025, + 0.002137406263500452, + 0.0030483820009976625, + 0.002519825007766485, + 0.003142330562695861, + 0.003384384559467435, + 0.0031612261664122343, + 0.006773436442017555, + 0.003562475088983774, + 0.002264839131385088, + 0.007831262424588203, + 0.0032123292330652475, + 0.001921263406984508, + 0.005775537807494402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.020578453317284584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle shot, like viewing through the powerful lens of a deep-space telescope, captures the majestic and sweeping celestial panorama of Saturn, the sixth planet from the Sun. This massive gas giant glows a soft, yellowish hue, its thin, hazy atmosphere subtly zoned, allowing distant sunlight to penetrate its upper layers. Its defining feature is a vast, intricate ring system that prominently encircles the planet's equator, tilted slightly toward the viewer. Composed of countless shimmering ice and rock particles, these rings are rendered in surreal detail, revealing unique concentric bands and gaps that create mesmerizing patterns of light and shadow. Sunlight catches the ice particles, giving the rings a silvery-white glow, while the planet casts a long, curved shadow across its vast plane. The luminous planet and its rings stand in stark contrast to the stark, absolute blackness of outer space, their light piercing the void. This tilted, slightly overhead perspective offers a complete, unobstructed view, emphasizing the celestial body's silent, awe-inspiring grandeur and creating a tranquil yet powerful atmosphere of cosmic discovery.", + "seed": 14043, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_11/index_7978_seed_14043_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013464257831219583, + "vlm_logprob": -0.013464257831219583, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.013464257831219583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00746520422399044 + ], + "opposite_logprob_tokens": [ + -5.382465362548828 + ], + "logprob_reward": -0.00746520422399044, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00746520422399044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0377933606505394, + 0.05576495826244354, + 0.061735786497592926, + 0.0555625855922699, + 0.06745138764381409, + 0.37866726517677307, + 0.04630166292190552, + 0.034316301345825195, + 0.020594703033566475, + 0.241812065243721 + ], + "credit_map": [ + 0.004381800536066294, + 0.002623838372528553, + 0.006613139994442463, + 0.004238375928252935, + 0.004967617802321911, + 0.005478119011968374, + 0.003762129694223404, + 0.005728339776396751, + 0.006368221249431372, + 0.003446859773248434, + 0.006190161686390638, + 0.00828003603965044, + 0.00711585721001029, + 0.008231600746512413, + 0.008205891586840153, + 0.007926322519779205, + 0.008390184491872787, + 0.0022672193590551615, + 0.01052539236843586, + 0.004283045884221792, + 0.002224347786977887, + 0.025537149980664253, + 0.004723627120256424, + 0.0037848157808184624, + 0.005096411798149347, + 0.0017526878509670496, + 0.009443563409149647, + 0.008893763646483421, + 0.0025574348401278257, + 0.021345430985093117, + 0.0041636135429143906, + 0.0023096709046512842, + 0.005041962489485741, + 0.001511337119154632, + 0.02693404257297516, + 0.0057448334991931915, + 0.0016809599474072456, + 0.02035447023808956, + 0.0026190150529146194, + 0.0035647691693156958, + 0.0262385755777359, + 0.015531273558735847, + 0.045969415456056595, + 0.13014358282089233, + 0.012327236123383045, + 0.02077682875096798, + 0.013976583257317543, + 0.11370375007390976, + 0.0053826309740543365, + 0.0031457734294235706, + 0.008601726032793522, + 0.005383749958127737, + 0.0052902307361364365, + 0.005742970388382673, + 0.004412398207932711, + 0.008342176675796509, + 0.004298658575862646, + 0.0015850160270929337, + 0.011369832791388035, + 0.0026974990032613277, + 0.002265122253447771, + 0.0027820244431495667, + 0.0016466184752061963, + 0.007671525701880455, + 0.0028335892129689455, + 0.001908796257339418, + 0.004250639118254185, + 0.001876152353361249, + 0.001938923029229045, + 0.0025311268400400877, + 0.0022730922792106867, + 0.0029823833610862494, + 0.03732921555638313, + 0.014169995673000813, + 0.04885900765657425, + 0.018043892458081245, + 0.017297590151429176, + 0.03289202228188515, + 0.01669939234852791, + 0.05652092769742012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.00746520422399044 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003995097358711064 + ], + "opposite_logprob_tokens": [ + -7.875399589538574 + ], + "logprob_reward": -0.0003995097358711064, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003995097358711064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03812667354941368, + 0.03744484856724739, + 0.08294644951820374, + 0.3993358314037323, + 0.1360965222120285, + 0.07267879694700241, + 0.031406134366989136, + 0.02797558531165123, + 0.0676024779677391, + 0.10638667643070221 + ], + "credit_map": [ + 0.004835434723645449, + 0.0024613591376692057, + 0.005847517400979996, + 0.004306368995457888, + 0.005577110685408115, + 0.005204787477850914, + 0.004220753908157349, + 0.005673338659107685, + 0.004006899427622557, + 0.0022177535574883223, + 0.007021631114184856, + 0.0056823184713721275, + 0.0032598571851849556, + 0.004806388169527054, + 0.004094365052878857, + 0.006355635821819305, + 0.011944612488150597, + 0.0029384640511125326, + 0.013077685609459877, + 0.004567986354231834, + 0.0029841652140021324, + 0.0349985808134079, + 0.006776530761271715, + 0.005658427253365517, + 0.03936992958188057, + 0.009573877789080143, + 0.06430844217538834, + 0.06788889318704605, + 0.01371141616255045, + 0.16970331966876984, + 0.02059865929186344, + 0.014181300066411495, + 0.010052381083369255, + 0.0031103738583624363, + 0.05618854984641075, + 0.014439886435866356, + 0.0033439593389630318, + 0.03751392662525177, + 0.005447528325021267, + 0.0059999325312674046, + 0.00600493885576725, + 0.003102669259533286, + 0.011767588555812836, + 0.022265387699007988, + 0.002574074547737837, + 0.00707095256075263, + 0.0030010929331183434, + 0.016892094165086746, + 0.003471417585387826, + 0.001861486816778779, + 0.008002766408026218, + 0.0033718859776854515, + 0.003984819166362286, + 0.003200333099812269, + 0.0024701447691768408, + 0.005043281242251396, + 0.0029091176111251116, + 0.0013698800466954708, + 0.008163162507116795, + 0.0021643650252372026, + 0.0016349432989954948, + 0.0029149376787245274, + 0.001523156650364399, + 0.007296022027730942, + 0.0036549395881593227, + 0.0019290720811113715, + 0.02364417165517807, + 0.0031548512633889914, + 0.0035057077184319496, + 0.003069594269618392, + 0.0031385463662445545, + 0.02550560049712658, + 0.017369842156767845, + 0.00547017389908433, + 0.022694407030940056, + 0.007024438586086035, + 0.006861243862658739, + 0.014034077525138855, + 0.006715128198266029, + 0.026217365637421608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.0003995097358711064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002206393051892519 + ], + "opposite_logprob_tokens": [ + -6.127206325531006 + ], + "logprob_reward": -0.002206393051892519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002206393051892519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04674634709954262, + 0.044968992471694946, + 0.20440202951431274, + 0.038208067417144775, + 0.4004206955432892, + 0.09306511282920837, + 0.039457906037569046, + 0.015538345091044903, + 0.0467708483338356, + 0.07042156904935837 + ], + "credit_map": [ + 0.0058526708744466305, + 0.003203602507710457, + 0.009162722155451775, + 0.006076327990740538, + 0.0043850252404809, + 0.007554712705314159, + 0.004983477294445038, + 0.005527813453227282, + 0.004993564914911985, + 0.00224889675155282, + 0.008230269886553288, + 0.007188995368778706, + 0.003736368613317609, + 0.005991511978209019, + 0.004668583627790213, + 0.007910806685686111, + 0.030220426619052887, + 0.006769504398107529, + 0.021326161921024323, + 0.01288862619549036, + 0.006682827603071928, + 0.08389580994844437, + 0.03073563612997532, + 0.011883051134645939, + 0.0032602036371827126, + 0.0018850087653845549, + 0.007934465073049068, + 0.0067199463956058025, + 0.0016919991467148066, + 0.011397033929824829, + 0.0030252698343247175, + 0.002294145757332444, + 0.029681269079446793, + 0.006312752142548561, + 0.16615630686283112, + 0.030879227444529533, + 0.008772448636591434, + 0.12718118727207184, + 0.01453760452568531, + 0.016899989917874336, + 0.0070625245571136475, + 0.0039011382032185793, + 0.008735236711800098, + 0.03345879912376404, + 0.002131671877577901, + 0.004123662132769823, + 0.0035246522165834904, + 0.030127447098493576, + 0.0048217251896858215, + 0.0029014679603278637, + 0.00677333353087306, + 0.0041830893605947495, + 0.007809916045516729, + 0.003617816837504506, + 0.004612729884684086, + 0.0047378321178257465, + 0.0015494574327021837, + 0.0012151378905400634, + 0.0028509944677352905, + 0.0009940431918948889, + 0.0024533739779144526, + 0.0019829459488391876, + 0.001434376579709351, + 0.0030580174643546343, + 0.002797153312712908, + 0.001349021797068417, + 0.016368312761187553, + 0.002192756626754999, + 0.0024044939782470465, + 0.0019496214808896184, + 0.002073951531201601, + 0.017635537311434746, + 0.010069700889289379, + 0.004430626519024372, + 0.013632059097290039, + 0.005507475696504116, + 0.005566390231251717, + 0.010526767931878567, + 0.005141320172697306, + 0.015547233633697033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.002206393051892519 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04571636766195297, + 0.03244038298726082, + 0.05075233057141304, + 0.26026755571365356, + 0.3174576461315155, + 0.03672269731760025, + 0.03802383691072464, + 0.021649466827511787, + 0.07229984551668167, + 0.12466984987258911 + ], + "credit_map": [ + 0.005378248170018196, + 0.0023096553049981594, + 0.007963454350829124, + 0.005490040872246027, + 0.006460185628384352, + 0.006212848704308271, + 0.004828983452171087, + 0.0070729488506913185, + 0.0033037152606993914, + 0.0015677681658416986, + 0.007752780336886644, + 0.0031746679451316595, + 0.0026775780133903027, + 0.003882741089910269, + 0.0021949869114905596, + 0.007886144332587719, + 0.007417742162942886, + 0.0015441299183294177, + 0.006484080106019974, + 0.002934178104624152, + 0.001535960240289569, + 0.020940903574228287, + 0.006933336611837149, + 0.002961995778605342, + 0.02479040063917637, + 0.00670271459966898, + 0.04263294115662575, + 0.04567408561706543, + 0.008635265752673149, + 0.1095350980758667, + 0.012592561542987823, + 0.009704504162073135, + 0.02356060966849327, + 0.005077379290014505, + 0.13484157621860504, + 0.024910693988204002, + 0.006958669051527977, + 0.09692078828811646, + 0.011703165248036385, + 0.013484777882695198, + 0.0039129056967794895, + 0.0020110115874558687, + 0.01092219166457653, + 0.005113522056490183, + 0.00348185864277184, + 0.004413051996380091, + 0.002458583563566208, + 0.004409572575241327, + 0.0027918859850615263, + 0.0023031209129840136, + 0.007919459603726864, + 0.004518851637840271, + 0.0055959029123187065, + 0.0040686288848519325, + 0.00444850092753768, + 0.006377488374710083, + 0.002229807199910283, + 0.0015472081722691655, + 0.0034776008687913418, + 0.002041458385065198, + 0.0019306527683511376, + 0.0030512786470353603, + 0.0031644750852137804, + 0.004206985700875521, + 0.00408259266987443, + 0.002655569463968277, + 0.02397461235523224, + 0.0036291913129389286, + 0.004181776661425829, + 0.0033984777983278036, + 0.00399154145270586, + 0.02638608030974865, + 0.01764579862356186, + 0.007489289157092571, + 0.025052588433027267, + 0.01031134556978941, + 0.009753720834851265, + 0.01749015785753727, + 0.00916250143200159, + 0.02776445262134075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.014192949049174786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305723309516907 + ], + "opposite_logprob_tokens": [ + -3.1680572032928467 + ], + "logprob_reward": -0.04305723309516907, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305723309516907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08365844190120697, + 0.05308876931667328, + 0.058504167944192886, + 0.3305427134037018, + 0.1820538491010666, + 0.20485155284404755, + 0.02203863486647606, + 0.01878098025918007, + 0.017946776002645493, + 0.0285341739654541 + ], + "credit_map": [ + 0.010167195461690426, + 0.006418419070541859, + 0.014765267260372639, + 0.01158213336020708, + 0.0077364505268633366, + 0.01236469391733408, + 0.007527103181928396, + 0.013097167946398258, + 0.007796734105795622, + 0.0034281595144420862, + 0.007502300199121237, + 0.008904163725674152, + 0.00464995251968503, + 0.009118864312767982, + 0.005555546842515469, + 0.00613304553553462, + 0.00892571359872818, + 0.002194014610722661, + 0.005415392108261585, + 0.003057594411075115, + 0.00252136099152267, + 0.024299798533320427, + 0.0086855823174119, + 0.0034047053195536137, + 0.03198162093758583, + 0.008044872432947159, + 0.05293513089418411, + 0.05720888450741768, + 0.011288737878203392, + 0.14005504548549652, + 0.01711367629468441, + 0.011914744973182678, + 0.016025006771087646, + 0.003889102954417467, + 0.07954327762126923, + 0.014450045302510262, + 0.00409949803724885, + 0.043192073702812195, + 0.007373076863586903, + 0.013481751084327698, + 0.014413012191653252, + 0.00837497878819704, + 0.01686592772603035, + 0.07489453256130219, + 0.004318523220717907, + 0.008571104146540165, + 0.007367210928350687, + 0.07004624605178833, + 0.0027158991433680058, + 0.001862810691818595, + 0.0027191617991775274, + 0.0021149669773876667, + 0.0030941821169108152, + 0.0033260437194257975, + 0.002346630208194256, + 0.0038589383475482464, + 0.001694505219347775, + 0.001449280884116888, + 0.004412748850882053, + 0.001797826960682869, + 0.0022484303917735815, + 0.002141684992238879, + 0.0017250385135412216, + 0.0033114636316895485, + 0.0011710679391399026, + 0.0005695128347724676, + 0.005653531290590763, + 0.000905507302377373, + 0.0009993466082960367, + 0.001310603809542954, + 0.0007843304192647338, + 0.006552874110639095, + 0.0043762559071183205, + 0.0021698293276131153, + 0.004446179606020451, + 0.0024522156454622746, + 0.00251821824349463, + 0.003963482100516558, + 0.0027182975318282843, + 0.005889694206416607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.04305723309516907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle shot, like viewing through the powerful lens of a deep-space telescope, captures the majestic and sweeping celestial panorama of Saturn, the sixth planet from the Sun. This massive gas giant glows a soft, yellowish hue, its thin, hazy atmosphere subtly zoned, allowing distant sunlight to penetrate its upper layers. Its defining feature is a vast, intricate ring system that prominently encircles the planet's equator, tilted slightly toward the viewer. Composed of countless shimmering ice and rock particles, these rings are rendered in surreal detail, revealing unique concentric bands and gaps that create mesmerizing patterns of light and shadow. Sunlight catches the ice particles, giving the rings a silvery-white glow, while the planet casts a long, curved shadow across its vast plane. The luminous planet and its rings stand in stark contrast to the stark, absolute blackness of outer space, their light piercing the void. This tilted, slightly overhead perspective offers a complete, unobstructed view, emphasizing the celestial body's silent, awe-inspiring grandeur and creating a tranquil yet powerful atmosphere of cosmic discovery.", + "seed": 14044, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_11/index_7978_seed_14044_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.009960811934433877, + "vlm_logprob": -0.009960811934433877, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.009960811934433877, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005617901682853699 + ], + "opposite_logprob_tokens": [ + -5.505618095397949 + ], + "logprob_reward": -0.005617901682853699, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005617901682853699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15162868797779083, + 0.08149363845586777, + 0.030010761693120003, + 0.03780191391706467, + 0.046165093779563904, + 0.06855157017707825, + 0.06560245901346207, + 0.0713452473282814, + 0.04213568940758705, + 0.4052649438381195 + ], + "credit_map": [ + 0.028986167162656784, + 0.01827111840248108, + 0.0237138569355011, + 0.011123002506792545, + 0.01723938249051571, + 0.0207829512655735, + 0.013149809092283249, + 0.018362397328019142, + 0.00794606190174818, + 0.008124509826302528, + 0.012750448659062386, + 0.0035738651640713215, + 0.025273319333791733, + 0.014812725596129894, + 0.005546029191464186, + 0.00346668204292655, + 0.002141551347449422, + 0.0038880116771906614, + 0.001934013213030994, + 0.0019125263206660748, + 0.004396284930408001, + 0.005277400370687246, + 0.005587486084550619, + 0.004873486235737801, + 0.006787888705730438, + 0.003481548512354493, + 0.005167465656995773, + 0.002512201201170683, + 0.006051278673112392, + 0.004758819472044706, + 0.002956132171675563, + 0.006086576730012894, + 0.0024497462436556816, + 0.004667527973651886, + 0.0027747030835598707, + 0.004173508379608393, + 0.007311011664569378, + 0.006844516843557358, + 0.010587523691356182, + 0.007356556132435799, + 0.0038468462880700827, + 0.006238899659365416, + 0.0030260817147791386, + 0.005378457717597485, + 0.010020503774285316, + 0.011197471059858799, + 0.012575849890708923, + 0.016267459839582443, + 0.005147404968738556, + 0.00744688231498003, + 0.004290791228413582, + 0.004988986533135176, + 0.00990425143390894, + 0.011584770865738392, + 0.010381899774074554, + 0.01185747142881155, + 0.006202513352036476, + 0.006183807272464037, + 0.004534823354333639, + 0.007477219216525555, + 0.01132467482239008, + 0.015819067135453224, + 0.009721623733639717, + 0.010081518441438675, + 0.002886400558054447, + 0.004219178110361099, + 0.003860911587253213, + 0.003537854878231883, + 0.006365750916302204, + 0.008838631212711334, + 0.006541151087731123, + 0.00588581059128046, + 0.024903738871216774, + 0.028407646343111992, + 0.027388202026486397, + 0.04836934432387352, + 0.09562857449054718, + 0.08040831983089447, + 0.05490751191973686, + 0.04525158554315567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.005617901682853699 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005062728887423873 + ], + "opposite_logprob_tokens": [ + -7.625506401062012 + ], + "logprob_reward": -0.0005062728887423873, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0005062728887423873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1188628226518631, + 0.45984795689582825, + 0.020685410127043724, + 0.018716098740696907, + 0.04291478917002678, + 0.03535940870642662, + 0.049588512629270554, + 0.08549020439386368, + 0.036841824650764465, + 0.1316930055618286 + ], + "credit_map": [ + 0.02167820930480957, + 0.016064152121543884, + 0.017838113009929657, + 0.008102215826511383, + 0.014569020830094814, + 0.01744622364640236, + 0.009613740257918835, + 0.013551143929362297, + 0.04058866575360298, + 0.03793925791978836, + 0.07295314967632294, + 0.01715671643614769, + 0.15444760024547577, + 0.08615221828222275, + 0.0276603065431118, + 0.02295004017651081, + 0.0017345233354717493, + 0.0030723209492862225, + 0.0019526607356965542, + 0.0014072083868086338, + 0.0033021126873791218, + 0.0033952516969293356, + 0.002064514672383666, + 0.003756818827241659, + 0.0014856699854135513, + 0.0029532723128795624, + 0.0013136108173057437, + 0.0013591527240350842, + 0.0025319871492683887, + 0.0028170845471322536, + 0.00346379354596138, + 0.0027915292885154486, + 0.0024326404090970755, + 0.004460674244910479, + 0.0021167355589568615, + 0.003880121512338519, + 0.007922406308352947, + 0.007156492676585913, + 0.008238157257437706, + 0.006707559339702129, + 0.002189385937526822, + 0.003179355524480343, + 0.0016705987509340048, + 0.0025505046360194683, + 0.006033151876181364, + 0.005484988447278738, + 0.004685304127633572, + 0.009566119872033596, + 0.003873309586197138, + 0.005176139529794455, + 0.002880031242966652, + 0.00415654806420207, + 0.010119060985744, + 0.008633577264845371, + 0.008329939097166061, + 0.006419908720999956, + 0.006198620889335871, + 0.008225659839808941, + 0.005007365718483925, + 0.010463490150868893, + 0.013784787617623806, + 0.019137993454933167, + 0.011987414211034775, + 0.010684872046113014, + 0.002389164874330163, + 0.004990545101463795, + 0.002356418641284108, + 0.003377733752131462, + 0.0059660691767930984, + 0.007609924301505089, + 0.005039494019001722, + 0.005112473852932453, + 0.008503809571266174, + 0.010598615743219852, + 0.011072774417698383, + 0.014294051565229893, + 0.026135431602597237, + 0.020741356536746025, + 0.015824319794774055, + 0.024522647261619568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0005062728887423873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101199563592672 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101199563592672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101199563592672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10420452058315277, + 0.18487776815891266, + 0.07632333040237427, + 0.05783018469810486, + 0.11951521039009094, + 0.04893187806010246, + 0.05065668746829033, + 0.09801089018583298, + 0.04771542176604271, + 0.2119341343641281 + ], + "credit_map": [ + 0.019297195598483086, + 0.01179547794163227, + 0.016906768083572388, + 0.006302670110017061, + 0.013207758776843548, + 0.016116660088300705, + 0.009409665130078793, + 0.011168312281370163, + 0.015654688701033592, + 0.016782237216830254, + 0.025499533861875534, + 0.009692547842860222, + 0.043323080986738205, + 0.03128217160701752, + 0.024852123111486435, + 0.0177913885563612, + 0.008176109753549099, + 0.009134186431765556, + 0.00883050262928009, + 0.004722900222986937, + 0.01007898524403572, + 0.010267015546560287, + 0.010536949150264263, + 0.014576681889593601, + 0.006711641326546669, + 0.006811119150370359, + 0.006852555088698864, + 0.005905685480684042, + 0.0065715438686311245, + 0.007111157290637493, + 0.00864182785153389, + 0.009224656969308853, + 0.007779938634485006, + 0.013199240900576115, + 0.006897107232362032, + 0.010732807219028473, + 0.01705605909228325, + 0.01665830984711647, + 0.02571023628115654, + 0.02148151397705078, + 0.0028955077286809683, + 0.0037628996651619673, + 0.0028686788864433765, + 0.00498613016679883, + 0.0074000232852995396, + 0.0079270601272583, + 0.007478639949113131, + 0.01161293126642704, + 0.004967264365404844, + 0.006552801467478275, + 0.0035259185824543238, + 0.00471679400652647, + 0.008106134831905365, + 0.009130582213401794, + 0.007530989125370979, + 0.0061262003146111965, + 0.007298447657376528, + 0.008287658914923668, + 0.006558466702699661, + 0.010663440451025963, + 0.023311063647270203, + 0.020444747060537338, + 0.012902776710689068, + 0.008544293232262135, + 0.0032293270342051983, + 0.005740239284932613, + 0.003706280142068863, + 0.004419333767145872, + 0.008032429963350296, + 0.010659796185791492, + 0.006426049862056971, + 0.005501968786120415, + 0.01015779934823513, + 0.015938645228743553, + 0.01269572414457798, + 0.02968706749379635, + 0.043496377766132355, + 0.039927493780851364, + 0.024294061586260796, + 0.03573696315288544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.004101199563592672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797395691275597 + ], + "opposite_logprob_tokens": [ + -3.529797315597534 + ], + "logprob_reward": -0.029797395691275597, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797395691275597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0834309458732605, + 0.2630749046802521, + 0.05626695603132248, + 0.029087040573358536, + 0.0842873677611351, + 0.10061203688383102, + 0.04803425818681717, + 0.12156885117292404, + 0.03560158237814903, + 0.17803604900836945 + ], + "credit_map": [ + 0.014997587539255619, + 0.0111344363540411, + 0.01252604741603136, + 0.00615558261051774, + 0.009804618544876575, + 0.010793766006827354, + 0.007867938838899136, + 0.01015097089111805, + 0.022661510854959488, + 0.02152644842863083, + 0.041763704270124435, + 0.009474826976656914, + 0.08995422720909119, + 0.04977306351065636, + 0.015539164654910564, + 0.012381934560835361, + 0.0091320239007473, + 0.00605362793430686, + 0.008246015757322311, + 0.003187011694535613, + 0.006231538951396942, + 0.007728755474090576, + 0.00621830066666007, + 0.009469677694141865, + 0.002441511955112219, + 0.003161662956699729, + 0.002028709976002574, + 0.0031429484952241182, + 0.004216072615236044, + 0.0052977693267166615, + 0.005183068104088306, + 0.0036152976099401712, + 0.005051043350249529, + 0.010297753848135471, + 0.003908889833837748, + 0.004690433386713266, + 0.02156364731490612, + 0.01889786496758461, + 0.01018854696303606, + 0.009689186699688435, + 0.0073085967451334, + 0.008812202140688896, + 0.006051986012607813, + 0.007899784483015537, + 0.016562091186642647, + 0.013296399265527725, + 0.012138660065829754, + 0.02854231745004654, + 0.0027159221936017275, + 0.006308338604867458, + 0.0019662498962134123, + 0.003578658914193511, + 0.009546631015837193, + 0.0075506665743887424, + 0.009680848568677902, + 0.006686941720545292, + 0.010334419086575508, + 0.009512418881058693, + 0.007870529778301716, + 0.011434901505708694, + 0.02845611795783043, + 0.025066165253520012, + 0.013031336478888988, + 0.015862958505749702, + 0.0027471142821013927, + 0.003889017505571246, + 0.0018552165711298585, + 0.0036464419681578875, + 0.006589766126126051, + 0.007437217049300671, + 0.004906156565994024, + 0.004530654288828373, + 0.008680427446961403, + 0.009755855426192284, + 0.013953584246337414, + 0.028808314353227615, + 0.03950336202979088, + 0.029429227113723755, + 0.022702498361468315, + 0.025202780961990356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.029797395691275597 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781289845705032 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781289845705032, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781289845705032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04026883840560913, + 0.7219691872596741, + 0.015096028335392475, + 0.052661407738924026, + 0.010579640045762062, + 0.006994232069700956, + 0.01292408723384142, + 0.049794185906648636, + 0.010791171342134476, + 0.07892119884490967 + ], + "credit_map": [ + 0.00815073773264885, + 0.0049149878323078156, + 0.006103688385337591, + 0.0029923105612397194, + 0.005573580507189035, + 0.0059787817299366, + 0.003598059993237257, + 0.0029566900338977575, + 0.06572968512773514, + 0.05979229137301445, + 0.11791567504405975, + 0.025594908744096756, + 0.24544689059257507, + 0.13600797951221466, + 0.03982992842793465, + 0.0316518135368824, + 0.0016297833062708378, + 0.0018312567844986916, + 0.0015949229709804058, + 0.001103628077544272, + 0.001957178581506014, + 0.002244061091914773, + 0.0025857530999928713, + 0.00214944314211607, + 0.00844544917345047, + 0.004695393610745668, + 0.0073067303746938705, + 0.0037927990779280663, + 0.007240453734993935, + 0.007320570293813944, + 0.006240579299628735, + 0.007619433104991913, + 0.001023165532387793, + 0.0011502041015774012, + 0.0007422651397064328, + 0.0008090329938568175, + 0.00186538218986243, + 0.0017764002550393343, + 0.0014107435708865523, + 0.00180244620423764, + 0.0006243382813408971, + 0.0008091703057289124, + 0.00038414844311773777, + 0.0006012996309436858, + 0.0009788468014448881, + 0.001419873209670186, + 0.0011711184633895755, + 0.0010054368758574128, + 0.0009913081303238869, + 0.0015038447454571724, + 0.0007085343240760267, + 0.0014866968849673867, + 0.0017552360659465194, + 0.0021649582777172327, + 0.0025360139552503824, + 0.0017774958396330476, + 0.004692235495895147, + 0.005440722219645977, + 0.0030851627234369516, + 0.004423615988343954, + 0.0087726516649127, + 0.009174293838441372, + 0.008184511214494705, + 0.00602099159732461, + 0.0008000584784895182, + 0.0011382607044652104, + 0.0010613679187372327, + 0.0009729950106702745, + 0.0017710758838802576, + 0.0021180110052227974, + 0.001308673177845776, + 0.001620728988200426, + 0.0038566719740629196, + 0.004100081976503134, + 0.005585054866969585, + 0.0110150296241045, + 0.019452739506959915, + 0.014849701896309853, + 0.009339794516563416, + 0.010722119361162186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.009781289845705032 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle shot, like viewing through the powerful lens of a deep-space telescope, captures the majestic and sweeping celestial panorama of Saturn, the sixth planet from the Sun. This massive gas giant glows a soft, yellowish hue, its thin, hazy atmosphere subtly zoned, allowing distant sunlight to penetrate its upper layers. Its defining feature is a vast, intricate ring system that prominently encircles the planet's equator, tilted slightly toward the viewer. Composed of countless shimmering ice and rock particles, these rings are rendered in surreal detail, revealing unique concentric bands and gaps that create mesmerizing patterns of light and shadow. Sunlight catches the ice particles, giving the rings a silvery-white glow, while the planet casts a long, curved shadow across its vast plane. The luminous planet and its rings stand in stark contrast to the stark, absolute blackness of outer space, their light piercing the void. This tilted, slightly overhead perspective offers a complete, unobstructed view, emphasizing the celestial body's silent, awe-inspiring grandeur and creating a tranquil yet powerful atmosphere of cosmic discovery.", + "seed": 14045, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_11/index_7978_seed_14045_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013364114600699395, + "vlm_logprob": -0.013364114600699395, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.013364114600699395, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015706051141023636 + ], + "opposite_logprob_tokens": [ + -4.2657060623168945 + ], + "logprob_reward": -0.015706051141023636, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015706051141023636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24004842340946198, + 0.07617641240358353, + 0.027347449213266373, + 0.03240957483649254, + 0.019056428223848343, + 0.2408626675605774, + 0.046192578971385956, + 0.10821564495563507, + 0.12854500114917755, + 0.08114579319953918 + ], + "credit_map": [ + 0.025947963818907738, + 0.026321055367588997, + 0.019940655678510666, + 0.02553439326584339, + 0.03775729238986969, + 0.039237771183252335, + 0.021747402846813202, + 0.0435619130730629, + 0.00943470187485218, + 0.009040012024343014, + 0.005604119505733252, + 0.009263717569410801, + 0.016689231619238853, + 0.011279997415840626, + 0.0063291313126683235, + 0.008535514585673809, + 0.004203474149107933, + 0.003375933039933443, + 0.0023293965496122837, + 0.0034101945348083973, + 0.0038607297465205193, + 0.003886478953063488, + 0.0022333378437906504, + 0.004047907888889313, + 0.004163462668657303, + 0.005148422904312611, + 0.0020441259257495403, + 0.003016295377165079, + 0.005792520008981228, + 0.0045368061400949955, + 0.0029846366960555315, + 0.004723309073597193, + 0.0040809521451592445, + 0.0028033307753503323, + 0.0013992742169648409, + 0.002347278641536832, + 0.0023511373437941074, + 0.002213394967839122, + 0.0020984180737286806, + 0.0017626462504267693, + 0.054250434041023254, + 0.02402360551059246, + 0.023556649684906006, + 0.031744346022605896, + 0.024832159280776978, + 0.027712877839803696, + 0.021551726385951042, + 0.03319087252020836, + 0.007720503490418196, + 0.005882191937416792, + 0.005735903047025204, + 0.004031110089272261, + 0.009912865236401558, + 0.0058466652408242226, + 0.003242114093154669, + 0.003821230959147215, + 0.01569373719394207, + 0.013105103746056557, + 0.008775541558861732, + 0.012747872620821, + 0.019955389201641083, + 0.012912014499306679, + 0.01560161355882883, + 0.009424387477338314, + 0.015115011483430862, + 0.012042096816003323, + 0.014686324633657932, + 0.01951034553349018, + 0.013186252675950527, + 0.01959950476884842, + 0.017201488837599754, + 0.01720399409532547, + 0.008424127474427223, + 0.00617244141176343, + 0.008438761346042156, + 0.006873367819935083, + 0.01691437140107155, + 0.009618903510272503, + 0.013647085055708885, + 0.011056735180318356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.015706051141023636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00031585473334416747 + ], + "opposite_logprob_tokens": [ + -8.12531566619873 + ], + "logprob_reward": -0.00031585473334416747, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.00031585473334416747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06600328534841537, + 0.0748467668890953, + 0.037060610949993134, + 0.025448450818657875, + 0.013321145437657833, + 0.29616811871528625, + 0.07185995578765869, + 0.08666935563087463, + 0.13461141288280487, + 0.1940108984708786 + ], + "credit_map": [ + 0.007305876351892948, + 0.008977522142231464, + 0.005428715143352747, + 0.008237718604505062, + 0.011032812297344208, + 0.00928825419396162, + 0.006176369730383158, + 0.009556015022099018, + 0.011487197130918503, + 0.005613845307379961, + 0.013164395466446877, + 0.009845170192420483, + 0.008826103061437607, + 0.011591955088078976, + 0.00597093440592289, + 0.00834716483950615, + 0.004613071214407682, + 0.004675067029893398, + 0.0037690200842916965, + 0.003139905398711562, + 0.00531369261443615, + 0.006010891869664192, + 0.002576360944658518, + 0.00696260342374444, + 0.003944136202335358, + 0.0023842966184020042, + 0.003172248834744096, + 0.0021155213471502066, + 0.003299025585874915, + 0.003423842368647456, + 0.002534192055463791, + 0.0045751892030239105, + 0.002893551718443632, + 0.0015201829373836517, + 0.000953051436226815, + 0.0014262532349675894, + 0.0015564564382657409, + 0.00152560172136873, + 0.001819936907850206, + 0.0016261112177744508, + 0.06756375730037689, + 0.0285818949341774, + 0.030048929154872894, + 0.04001574590802193, + 0.029160043224692345, + 0.033865392208099365, + 0.026381367817521095, + 0.040550995618104935, + 0.012783169746398926, + 0.007918150164186954, + 0.009137251414358616, + 0.006184759549796581, + 0.013604206033051014, + 0.009161741472780704, + 0.006315929349511862, + 0.006754745729267597, + 0.016614224761724472, + 0.010195063427090645, + 0.007238028571009636, + 0.010557619854807854, + 0.015906674787402153, + 0.009525285102427006, + 0.010439913719892502, + 0.0061925435438752174, + 0.0121768144890666, + 0.011806736700236797, + 0.012738102115690708, + 0.02262149751186371, + 0.013521561399102211, + 0.020493026822805405, + 0.02247377298772335, + 0.018779905512928963, + 0.01705332100391388, + 0.01689973473548889, + 0.015782814472913742, + 0.011019046418368816, + 0.049109525978565216, + 0.023716477677226067, + 0.03606802225112915, + 0.024361953139305115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.00031585473334416747 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822346519678831 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822346519678831, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822346519678831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09698548167943954, + 0.08390642702579498, + 0.04076746106147766, + 0.02166677638888359, + 0.019166745245456696, + 0.16556963324546814, + 0.07774659246206284, + 0.109585240483284, + 0.16909031569957733, + 0.2155153453350067 + ], + "credit_map": [ + 0.011224158108234406, + 0.011930522508919239, + 0.009808430448174477, + 0.012196161784231663, + 0.020782899111509323, + 0.014284639619290829, + 0.0059013525024056435, + 0.010857322253286839, + 0.013311746530234814, + 0.00658452557399869, + 0.014944132417440414, + 0.00932205282151699, + 0.00846041552722454, + 0.012683182954788208, + 0.006356102880090475, + 0.012244264595210552, + 0.004247778560966253, + 0.004306800197809935, + 0.004157437011599541, + 0.00338209536857903, + 0.004818152170628309, + 0.0071429903618991375, + 0.0032998372334986925, + 0.0094123724848032, + 0.00252993474714458, + 0.0030395069625228643, + 0.0014907537261024117, + 0.0019338926067575812, + 0.003942757844924927, + 0.002857450395822525, + 0.002262981841340661, + 0.0036094989627599716, + 0.0050704386085271835, + 0.0024306639097630978, + 0.0016931082354858518, + 0.002208102261647582, + 0.0020529672037810087, + 0.002241426380351186, + 0.0015525557100772858, + 0.0019174822373315692, + 0.03474404662847519, + 0.015444699674844742, + 0.017723554745316505, + 0.01948666200041771, + 0.017162887379527092, + 0.018498841673135757, + 0.018696201965212822, + 0.02381274662911892, + 0.017334960401058197, + 0.0073662130162119865, + 0.013695629313588142, + 0.006358408369123936, + 0.011423791758716106, + 0.00851292535662651, + 0.005403025541454554, + 0.007651646155864, + 0.02216240018606186, + 0.01260827761143446, + 0.00989130511879921, + 0.01443481259047985, + 0.018360139802098274, + 0.011870482936501503, + 0.012616155669093132, + 0.007641668897122145, + 0.014570503495633602, + 0.013309595175087452, + 0.015244923532009125, + 0.027872925624251366, + 0.016416963189840317, + 0.02537079155445099, + 0.03175075724720955, + 0.024553848430514336, + 0.020866381004452705, + 0.014339182525873184, + 0.02096686325967312, + 0.017163164913654327, + 0.04754287749528885, + 0.025848539546132088, + 0.041185345500707626, + 0.02760297805070877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.002822346519678831 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793230816721916 + ], + "opposite_logprob_tokens": [ + -3.5297932624816895 + ], + "logprob_reward": -0.029793230816721916, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793230816721916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08876009285449982, + 0.08149366080760956, + 0.053118009120225906, + 0.03871103748679161, + 0.018303273245692253, + 0.36815762519836426, + 0.05555636063218117, + 0.04989602044224739, + 0.07451547682285309, + 0.1714884489774704 + ], + "credit_map": [ + 0.0156939048320055, + 0.00944785587489605, + 0.00766251003369689, + 0.00912870280444622, + 0.014511557295918465, + 0.014464116655290127, + 0.005771930795162916, + 0.012079508043825626, + 0.014644124545156956, + 0.006070662755519152, + 0.012810582295060158, + 0.0112488754093647, + 0.005647916812449694, + 0.011849706061184406, + 0.007071000058203936, + 0.012150779366493225, + 0.005420024041086435, + 0.005115619394928217, + 0.005883289035409689, + 0.004303325433284044, + 0.005394912790507078, + 0.009383855387568474, + 0.004157430492341518, + 0.01345954928547144, + 0.005694743245840073, + 0.00625150790438056, + 0.0023619455751031637, + 0.00354396621696651, + 0.008525491692125797, + 0.005036612972617149, + 0.003574743866920471, + 0.003722019959241152, + 0.0032919698860496283, + 0.0025058265309780836, + 0.0017350484849885106, + 0.0018992644036188722, + 0.0029063208494335413, + 0.0023184933234006166, + 0.001205438282340765, + 0.0024409089237451553, + 0.08609198033809662, + 0.03537317365407944, + 0.03443307429552078, + 0.05396376922726631, + 0.036351464688777924, + 0.04326680675148964, + 0.028521370142698288, + 0.050155963748693466, + 0.01065674889832735, + 0.006873868405818939, + 0.0075304703786969185, + 0.005454809404909611, + 0.010989585891366005, + 0.0057884263806045055, + 0.0037528418470174074, + 0.004509604535996914, + 0.009376929141581059, + 0.005915454588830471, + 0.004065772984176874, + 0.005597230512648821, + 0.009052806533873081, + 0.0054552326910197735, + 0.005205739289522171, + 0.005226851440966129, + 0.007752709556370974, + 0.01130366325378418, + 0.005180951673537493, + 0.00831630639731884, + 0.017488764598965645, + 0.01056938897818327, + 0.007790499832481146, + 0.006113189738243818, + 0.01967732235789299, + 0.01657385751605034, + 0.01973194070160389, + 0.016532234847545624, + 0.015972428023815155, + 0.015170833095908165, + 0.042137157171964645, + 0.02569265477359295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.029793230816721916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gas giant planet resembling Saturn with a yellowish hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, intricate ring system visible encircling the planet's equator?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rings rendered with silvery-white ice particles showing concentric bands and gaps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static, high-angle shot tilted slightly toward the viewer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background depicted as stark, absolute blackness of outer space?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183089792728424 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183089792728424, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183089792728424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06375399976968765, + 0.18435552716255188, + 0.02267724834382534, + 0.035826604813337326, + 0.010979656130075455, + 0.4672844409942627, + 0.01632889173924923, + 0.029491625726222992, + 0.03451898694038391, + 0.13478298485279083 + ], + "credit_map": [ + 0.006913753692060709, + 0.006411650683730841, + 0.005542056169360876, + 0.008296829648315907, + 0.00828388798981905, + 0.010124183259904385, + 0.007182321511209011, + 0.010999326594173908, + 0.032258085906505585, + 0.01368710957467556, + 0.032192472368478775, + 0.019133999943733215, + 0.019655801355838776, + 0.027844995260238647, + 0.013074665330350399, + 0.02650843746960163, + 0.0028192216996103525, + 0.002410735934972763, + 0.0015716843772679567, + 0.002779188798740506, + 0.0027222686912864447, + 0.0039023980498313904, + 0.002087212633341551, + 0.004384542349725962, + 0.005815836135298014, + 0.0029925887938588858, + 0.003999552223831415, + 0.0028901260811835527, + 0.003156007034704089, + 0.005303563550114632, + 0.0036420186515897512, + 0.008026923052966595, + 0.0019273451762273908, + 0.0015390746993944049, + 0.0008376744226552546, + 0.0016551014268770814, + 0.0013017297023907304, + 0.0011960604460909963, + 0.0013599785743281245, + 0.001162694301456213, + 0.10607974231243134, + 0.044462889432907104, + 0.04641859978437424, + 0.06798708438873291, + 0.04559839516878128, + 0.05318804085254669, + 0.03920289874076843, + 0.06434685736894608, + 0.003425213973969221, + 0.0015646459069103003, + 0.0018718797946348786, + 0.0022367662750184536, + 0.002146664308384061, + 0.0021144647616893053, + 0.001438132836483419, + 0.0015311271417886019, + 0.006245858501642942, + 0.003418431617319584, + 0.0027077673003077507, + 0.003727007657289505, + 0.0057033272460103035, + 0.0030509955249726772, + 0.002682432299479842, + 0.0019558127969503403, + 0.004555893130600452, + 0.0030110992956906557, + 0.003523251973092556, + 0.004004190675914288, + 0.004214199725538492, + 0.004671995062381029, + 0.005755293648689985, + 0.004783074837177992, + 0.014227933250367641, + 0.008738112635910511, + 0.014336811378598213, + 0.008846459910273552, + 0.026412690058350563, + 0.014900116249918938, + 0.02827855758368969, + 0.019042326137423515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.018183089792728424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_12/index_tensor([1078, 1078])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_12/index_tensor([1078, 1078])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..767baa6272e84cadd92c6f53f0e4a61e43b5d473 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_12/index_tensor([1078, 1078])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "violently shaky first-person perspective camera places the viewer in a pitching ship cabin. Every violent pitch and jolt of the ship shakes the frame, simulating the dizzying sensation of losing balance amidst surging waves. A single, dim light, cast from a bare bulb hanging from the ceiling, carves a sharp path through the murky air filled with suspended dust. The beam shifts with the swaying of the ship, causing shadows to dance wildly within the room. This light illuminates the bleakness of the room: dark, water-stained wood panels on the walls, a simple fixed bunk with a thin, wrinkled mattress, and no one else. An unsteady gaze shifts to the small, salt-stained round porthole, revealing only the endless, roaring, rolling gray waves outside. These waves merge with the gloomy sky, creating a monotonous, oppressive gray world. The entire scene is extremely dark in tone, dominated by cold gray and dirty brown. The only warm color comes from the faint and flickering light, but this light not only does not bring warmth, but instead intensifies the deep, suffocating loneliness and desolation in the empty room.", + "seed": 21830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_12/index_1078_seed_21830_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011902691307477653, + "vlm_logprob": -0.011902691307477653, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.011902691307477653, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03952197730541229 + ], + "opposite_logprob_tokens": [ + -3.5395219326019287 + ], + "logprob_reward": -0.03952197730541229, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03952197730541229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2118210643529892, + 0.1014101505279541, + 0.11288241297006607, + 0.06366400420665741, + 0.14772331714630127, + 0.15905746817588806, + 0.04444247856736183, + 0.07115079462528229, + 0.030048709362745285, + 0.0577995702624321 + ], + "credit_map": [ + 0.014902236871421337, + 0.02006560005247593, + 0.027281828224658966, + 0.016625935211777687, + 0.007468282710760832, + 0.06520962715148926, + 0.016699299216270447, + 0.04356823489069939, + 0.008836152032017708, + 0.014040282927453518, + 0.013892162591218948, + 0.017551014199852943, + 0.007089253515005112, + 0.011752811260521412, + 0.012683063745498657, + 0.015565415844321251, + 0.004836482927203178, + 0.007517661899328232, + 0.02734176628291607, + 0.03785792738199234, + 0.005028070881962776, + 0.0050322674214839935, + 0.016179939731955528, + 0.0090883057564497, + 0.006785234902054071, + 0.008729593828320503, + 0.008147991262376308, + 0.010190311819314957, + 0.004398545250296593, + 0.008182722143828869, + 0.010299497283995152, + 0.006930111907422543, + 0.01425770204514265, + 0.01696608029305935, + 0.019394706934690475, + 0.023757699877023697, + 0.01658383011817932, + 0.013149970211088657, + 0.026231378316879272, + 0.017381958663463593, + 0.012083948589861393, + 0.022750910371541977, + 0.02205854468047619, + 0.027632256969809532, + 0.014566529542207718, + 0.01307002641260624, + 0.023351499810814857, + 0.023543743416666985, + 0.00461047887802124, + 0.003611390246078372, + 0.0045380485244095325, + 0.006087046582251787, + 0.006726977881044149, + 0.008384494110941887, + 0.00528038339689374, + 0.005203659180551767, + 0.010293976403772831, + 0.009874763898551464, + 0.015307389199733734, + 0.011620930396020412, + 0.007491203024983406, + 0.005459913052618504, + 0.005248494911938906, + 0.0058541190810501575, + 0.003295931965112686, + 0.003916894551366568, + 0.0054051922634243965, + 0.006624382920563221, + 0.0033737351186573505, + 0.0020618156995624304, + 0.002463272539898753, + 0.002907484071329236, + 0.010599162429571152, + 0.0053125787526369095, + 0.007539444603025913, + 0.009270930662751198, + 0.0028801667504012585, + 0.007748047821223736, + 0.009414974600076675, + 0.005034261383116245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.03952197730541229 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795927442610264 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.009795927442610264, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795927442610264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12396008521318436, + 0.12461484968662262, + 0.25222188234329224, + 0.0659613087773323, + 0.14423692226409912, + 0.04263561964035034, + 0.08907929807901382, + 0.038283951580524445, + 0.051616568118333817, + 0.06738948822021484 + ], + "credit_map": [ + 0.010572045110166073, + 0.01972775347530842, + 0.013044108636677265, + 0.020909124985337257, + 0.010750286281108856, + 0.017146142199635506, + 0.01619144342839718, + 0.015619187615811825, + 0.012466352432966232, + 0.015868810936808586, + 0.014606179669499397, + 0.01772434078156948, + 0.008473016321659088, + 0.013371217995882034, + 0.017673268914222717, + 0.02443167008459568, + 0.014242448844015598, + 0.03068576380610466, + 0.04823422431945801, + 0.06377114355564117, + 0.019774025306105614, + 0.014135959558188915, + 0.03500402718782425, + 0.02637430466711521, + 0.005976989399641752, + 0.007861027494072914, + 0.008610150776803493, + 0.014667217619717121, + 0.004682176746428013, + 0.00829065591096878, + 0.007233994547277689, + 0.008639102801680565, + 0.012764365412294865, + 0.012989077717065811, + 0.01973130740225315, + 0.02272305265069008, + 0.011248412542045116, + 0.010165182873606682, + 0.025972142815589905, + 0.028643393889069557, + 0.0048624733462929726, + 0.006534312851727009, + 0.005489580798894167, + 0.008367753587663174, + 0.0033770364243537188, + 0.003206048160791397, + 0.005829618312418461, + 0.004968797322362661, + 0.014089133590459824, + 0.010212361812591553, + 0.009780501015484333, + 0.011510959826409817, + 0.011554368771612644, + 0.013266762718558311, + 0.009166743606328964, + 0.009498474188148975, + 0.004046970512717962, + 0.004609036259353161, + 0.00806057546287775, + 0.0051241712644696236, + 0.0029531281907111406, + 0.00318987132050097, + 0.004439177457243204, + 0.005861022975295782, + 0.005173801444470882, + 0.005612464156001806, + 0.010886549949645996, + 0.011016079224646091, + 0.0047727832570672035, + 0.0033199130557477474, + 0.005364483222365379, + 0.0054704975336790085, + 0.005945321638137102, + 0.005535296630114317, + 0.013089502230286598, + 0.01416291669011116, + 0.003144885180518031, + 0.007357412949204445, + 0.012412904761731625, + 0.00574125349521637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.009795927442610264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956043489277363 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956043489277363, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956043489277363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35803642868995667, + 0.05743515491485596, + 0.03760989382863045, + 0.07298281043767929, + 0.18982553482055664, + 0.06206759437918663, + 0.08188236504793167, + 0.046056147664785385, + 0.020230073481798172, + 0.07387401163578033 + ], + "credit_map": [ + 0.027573034167289734, + 0.02160249464213848, + 0.03249077498912811, + 0.029726622626185417, + 0.018398582935333252, + 0.11312759667634964, + 0.0327729769051075, + 0.08234431594610214, + 0.003811539150774479, + 0.006281787063926458, + 0.005976927001029253, + 0.008046852424740791, + 0.007630150765180588, + 0.005203375592827797, + 0.010125817731022835, + 0.010358704254031181, + 0.001959144603461027, + 0.003244071500375867, + 0.007848558947443962, + 0.010217237286269665, + 0.002266073366627097, + 0.00255896570160985, + 0.005873449146747589, + 0.003642395604401827, + 0.0075622242875397205, + 0.0064743394032120705, + 0.009391160681843758, + 0.012248186394572258, + 0.00587299233302474, + 0.005058936309069395, + 0.013834941200911999, + 0.012540027499198914, + 0.014119809493422508, + 0.01578797772526741, + 0.02502947486937046, + 0.029210813343524933, + 0.021231504157185555, + 0.017050107941031456, + 0.03224334120750427, + 0.03515251353383064, + 0.00430525466799736, + 0.00774986669421196, + 0.010840622708201408, + 0.00965167861431837, + 0.005519069265574217, + 0.006241837982088327, + 0.00784373003989458, + 0.009915532544255257, + 0.011163847520947456, + 0.008215140551328659, + 0.012415137141942978, + 0.00981778372079134, + 0.014725527726113796, + 0.010121189057826996, + 0.007628873456269503, + 0.0077948663383722305, + 0.004725632257759571, + 0.004909072536975145, + 0.0061221616342663765, + 0.005503145977854729, + 0.004883705172687769, + 0.005626876838505268, + 0.004668248351663351, + 0.009617307223379612, + 0.002166442573070526, + 0.003086454700678587, + 0.0030678848270326853, + 0.0035216170363128185, + 0.0018471436342224479, + 0.001427645212970674, + 0.0023871574085205793, + 0.0027257276233285666, + 0.010875464417040348, + 0.006937055382877588, + 0.008452219888567924, + 0.012018068693578243, + 0.005182584282010794, + 0.014352274127304554, + 0.01018490269780159, + 0.00587143748998642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.005956043489277363 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031943507492542267 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.0031943507492542267, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031943507492542267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17002294957637787, + 0.07245855778455734, + 0.11751458793878555, + 0.06455878913402557, + 0.12990042567253113, + 0.06021130830049515, + 0.17801757156848907, + 0.09438043087720871, + 0.05238154157996178, + 0.06055394187569618 + ], + "credit_map": [ + 0.011355810798704624, + 0.015259569510817528, + 0.01805766485631466, + 0.014667931012809277, + 0.006763834971934557, + 0.04994603991508484, + 0.020607907325029373, + 0.03336416557431221, + 0.006191429682075977, + 0.011730089783668518, + 0.008764655329287052, + 0.012697715312242508, + 0.006050023250281811, + 0.007529614493250847, + 0.011928231455385685, + 0.007566788699477911, + 0.0042843082919716835, + 0.005587573163211346, + 0.031857170164585114, + 0.043287791311740875, + 0.006466391962021589, + 0.0028730721678584814, + 0.015018961392343044, + 0.00813930481672287, + 0.005944765172898769, + 0.006724261213093996, + 0.00855004321783781, + 0.012327829375863075, + 0.004486634396016598, + 0.006811507977545261, + 0.010159949772059917, + 0.00955379568040371, + 0.011287963017821312, + 0.013253789395093918, + 0.020452838391065598, + 0.024423636496067047, + 0.007614970672875643, + 0.007329031825065613, + 0.023868892341852188, + 0.021669287234544754, + 0.004340255632996559, + 0.008647925220429897, + 0.01269873883575201, + 0.009028158150613308, + 0.0059076156467199326, + 0.005042850971221924, + 0.008732179179787636, + 0.0058135828003287315, + 0.018737878650426865, + 0.01933753304183483, + 0.013557136058807373, + 0.01703638769686222, + 0.029169457033276558, + 0.03442063555121422, + 0.018570052459836006, + 0.02718847058713436, + 0.012721368111670017, + 0.0117191756144166, + 0.012368227355182171, + 0.018619781360030174, + 0.008030092343688011, + 0.008916206657886505, + 0.009831077419221401, + 0.012174495495855808, + 0.005607650615274906, + 0.0065752314403653145, + 0.013206087984144688, + 0.010166673921048641, + 0.004602770321071148, + 0.0021321801468729973, + 0.005321124102920294, + 0.004769820719957352, + 0.006974714808166027, + 0.005677293054759502, + 0.012056011706590652, + 0.013753396458923817, + 0.002881015418097377, + 0.00607289420440793, + 0.0077171712182462215, + 0.005421436857432127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0031943507492542267 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010451575508341193 + ], + "opposite_logprob_tokens": [ + -6.876045227050781 + ], + "logprob_reward": -0.0010451575508341193, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010451575508341193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19743238389492035, + 0.12608276307582855, + 0.1090909019112587, + 0.11069934070110321, + 0.07081589102745056, + 0.07600802928209305, + 0.12859563529491425, + 0.06959383189678192, + 0.04212278500199318, + 0.06955844163894653 + ], + "credit_map": [ + 0.017992498353123665, + 0.013588234782218933, + 0.022813094779849052, + 0.017335686832666397, + 0.009141975082457066, + 0.054664116352796555, + 0.01444092858582735, + 0.047455862164497375, + 0.007221473380923271, + 0.014847511425614357, + 0.015198009088635445, + 0.022935306653380394, + 0.012045030482113361, + 0.014228051528334618, + 0.019375480711460114, + 0.020231902599334717, + 0.004764573182910681, + 0.006364540196955204, + 0.02465645782649517, + 0.03523801639676094, + 0.005378613714128733, + 0.0036750284489244223, + 0.019055426120758057, + 0.009958242997527122, + 0.010802011005580425, + 0.01193828135728836, + 0.013416909612715244, + 0.018081897869706154, + 0.009285874664783478, + 0.008086551912128925, + 0.021730447188019753, + 0.017357364296913147, + 0.009168791584670544, + 0.007321087643504143, + 0.009590326808393002, + 0.012132378295063972, + 0.005637663416564465, + 0.005125392694026232, + 0.013090896420180798, + 0.008749355562031269, + 0.006127106957137585, + 0.010617103427648544, + 0.0184159092605114, + 0.011418220587074757, + 0.007210021372884512, + 0.005909630097448826, + 0.009766618721187115, + 0.006543422117829323, + 0.02145339362323284, + 0.014927408657968044, + 0.01966203562915325, + 0.014799349941313267, + 0.014379345811903477, + 0.017217833548784256, + 0.013574711047112942, + 0.01258154958486557, + 0.010149645619094372, + 0.009186120703816414, + 0.010943516157567501, + 0.015364550985395908, + 0.004830007441341877, + 0.004199483897536993, + 0.008762779645621777, + 0.0061577195301651955, + 0.00424959184601903, + 0.0044220020063221455, + 0.005868822336196899, + 0.0058810594491660595, + 0.004574073012918234, + 0.003361382754519582, + 0.005197461228817701, + 0.00856839306652546, + 0.009403744712471962, + 0.006434926297515631, + 0.009176229126751423, + 0.012132092379033566, + 0.004401718266308308, + 0.011977195739746094, + 0.011181369423866272, + 0.004851158242672682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0010451575508341193 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "violently shaky first-person perspective camera places the viewer in a pitching ship cabin. Every violent pitch and jolt of the ship shakes the frame, simulating the dizzying sensation of losing balance amidst surging waves. A single, dim light, cast from a bare bulb hanging from the ceiling, carves a sharp path through the murky air filled with suspended dust. The beam shifts with the swaying of the ship, causing shadows to dance wildly within the room. This light illuminates the bleakness of the room: dark, water-stained wood panels on the walls, a simple fixed bunk with a thin, wrinkled mattress, and no one else. An unsteady gaze shifts to the small, salt-stained round porthole, revealing only the endless, roaring, rolling gray waves outside. These waves merge with the gloomy sky, creating a monotonous, oppressive gray world. The entire scene is extremely dark in tone, dominated by cold gray and dirty brown. The only warm color comes from the faint and flickering light, but this light not only does not bring warmth, but instead intensifies the deep, suffocating loneliness and desolation in the empty room.", + "seed": 21831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_12/index_1078_seed_21831_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014198473165743053, + "vlm_logprob": -0.014198473165743053, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.014198473165743053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04218493402004242 + ], + "opposite_logprob_tokens": [ + -3.292184829711914 + ], + "logprob_reward": -0.04218493402004242, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04218493402004242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06492041051387787, + 0.07671705633401871, + 0.28061211109161377, + 0.15463535487651825, + 0.14913003146648407, + 0.04570314288139343, + 0.08216505497694016, + 0.03460369631648064, + 0.06411177664995193, + 0.04740145802497864 + ], + "credit_map": [ + 0.009589825756847858, + 0.0070579396560788155, + 0.006979592144489288, + 0.009445593692362309, + 0.008800010196864605, + 0.008361040614545345, + 0.0071245841681957245, + 0.007561817299574614, + 0.012062516063451767, + 0.011262258514761925, + 0.010474584996700287, + 0.012807760387659073, + 0.00909885298460722, + 0.0062961275689303875, + 0.008541394025087357, + 0.006173558067530394, + 0.02064858004450798, + 0.07791256159543991, + 0.11622697114944458, + 0.01896890439093113, + 0.012745015323162079, + 0.008627533912658691, + 0.012863376177847385, + 0.012619144283235073, + 0.010874815285205841, + 0.010150646790862083, + 0.011409304104745388, + 0.0098272068426013, + 0.026422804221510887, + 0.04816574975848198, + 0.01208922453224659, + 0.02569557912647724, + 0.016413642093539238, + 0.03944768011569977, + 0.019041774794459343, + 0.019651085138320923, + 0.01712198182940483, + 0.011343121528625488, + 0.011987349949777126, + 0.014123381115496159, + 0.003981934394687414, + 0.006488441023975611, + 0.006234932690858841, + 0.006747759412974119, + 0.004724715370684862, + 0.007244172506034374, + 0.004981287755072117, + 0.00529989181086421, + 0.005231925752013922, + 0.009587662294507027, + 0.011212265118956566, + 0.014022179879248142, + 0.00694594020023942, + 0.008353140205144882, + 0.016997046768665314, + 0.009814884513616562, + 0.0027600349858403206, + 0.008439010940492153, + 0.004031179938465357, + 0.005218234844505787, + 0.004742254037410021, + 0.0022643005941063166, + 0.003324666526168585, + 0.00382400956004858, + 0.007703304756432772, + 0.00522217433899641, + 0.014255906455218792, + 0.00898768100887537, + 0.009139524772763252, + 0.006571329198777676, + 0.004158273339271545, + 0.008073573000729084, + 0.006972632370889187, + 0.008773311041295528, + 0.005355069879442453, + 0.006056501064449549, + 0.0061773317866027355, + 0.00440216576680541, + 0.005257840268313885, + 0.004406603053212166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.04218493402004242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260791163891554 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005260791163891554, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260791163891554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14736832678318024, + 0.14742539823055267, + 0.10156675428152084, + 0.03191134333610535, + 0.07852879166603088, + 0.060607459396123886, + 0.15574729442596436, + 0.10438309609889984, + 0.09837256371974945, + 0.07408899068832397 + ], + "credit_map": [ + 0.02223190851509571, + 0.014552141539752483, + 0.024280786514282227, + 0.0185692235827446, + 0.014472356997430325, + 0.01821443997323513, + 0.015756241977214813, + 0.019291231408715248, + 0.023660406470298767, + 0.01592019945383072, + 0.02129579521715641, + 0.02617163397371769, + 0.016545727849006653, + 0.01201984379440546, + 0.01785469241440296, + 0.013957091607153416, + 0.006703119724988937, + 0.02246677502989769, + 0.030705144628882408, + 0.00931155402213335, + 0.007238872814923525, + 0.009869069792330265, + 0.007658766582608223, + 0.0076134540140628815, + 0.0028658693190664053, + 0.004387958440929651, + 0.004127041436731815, + 0.003350013168528676, + 0.006249690428376198, + 0.004852075129747391, + 0.002975133480504155, + 0.003103559371083975, + 0.0075522917322814465, + 0.01905236765742302, + 0.009969256818294525, + 0.010245244950056076, + 0.009523560293018818, + 0.006693949922919273, + 0.00957367941737175, + 0.005918443202972412, + 0.004965826403349638, + 0.008618761785328388, + 0.007856548763811588, + 0.010358570143580437, + 0.006181060336530209, + 0.00719120679423213, + 0.008267118595540524, + 0.007168364245444536, + 0.011104515753686428, + 0.018506944179534912, + 0.021405242383480072, + 0.026557883247733116, + 0.017122061923146248, + 0.01904921792447567, + 0.029112011194229126, + 0.01288942527025938, + 0.009734985418617725, + 0.020982038229703903, + 0.01246619038283825, + 0.013759511522948742, + 0.015704326331615448, + 0.013914812356233597, + 0.010391793213784695, + 0.007429442368447781, + 0.005163888912647963, + 0.007680147886276245, + 0.016051921993494034, + 0.007936431095004082, + 0.03286445140838623, + 0.01668284647166729, + 0.004841578658670187, + 0.00715130427852273, + 0.0068762535229325294, + 0.00873407069593668, + 0.008746220730245113, + 0.01137510221451521, + 0.011425967328250408, + 0.01037691067904234, + 0.006609369534999132, + 0.009945101104676723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.005260791163891554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641231805086136 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.008641231805086136, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641231805086136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10061636567115784, + 0.09678107500076294, + 0.0965782180428505, + 0.03561139106750488, + 0.0765869989991188, + 0.048950113356113434, + 0.06785975396633148, + 0.14501096308231354, + 0.09426957368850708, + 0.23773562908172607 + ], + "credit_map": [ + 0.009503796696662903, + 0.007861290127038956, + 0.018439898267388344, + 0.009876811876893044, + 0.009889721870422363, + 0.02414696291089058, + 0.00940176472067833, + 0.011496104300022125, + 0.01855672337114811, + 0.012706141918897629, + 0.014212574809789658, + 0.017381437122821808, + 0.009034747257828712, + 0.006200763862580061, + 0.010874091647565365, + 0.007814571261405945, + 0.009909647516906261, + 0.01022693794220686, + 0.020684828981757164, + 0.014692756347358227, + 0.008588846772909164, + 0.013681705109775066, + 0.010043546557426453, + 0.008749923668801785, + 0.0024871272034943104, + 0.0028997401241213083, + 0.005848418455570936, + 0.0031943791545927525, + 0.005024919752031565, + 0.00915477517992258, + 0.003492537187412381, + 0.003509486559778452, + 0.007798252161592245, + 0.02483482100069523, + 0.0063843028619885445, + 0.010049781762063503, + 0.009269009344279766, + 0.005254183430224657, + 0.007462651934474707, + 0.005533984396606684, + 0.0035742369946092367, + 0.005507126450538635, + 0.007048055529594421, + 0.009613268077373505, + 0.004913991317152977, + 0.007307213731110096, + 0.006753102410584688, + 0.004233106505125761, + 0.005000383127480745, + 0.007800636347383261, + 0.008713941089808941, + 0.009949274361133575, + 0.010128305293619633, + 0.008265278302133083, + 0.008895043283700943, + 0.009106878191232681, + 0.015294359065592289, + 0.024625029414892197, + 0.019176362082362175, + 0.024063223972916603, + 0.018324822187423706, + 0.015835078433156013, + 0.013723506592214108, + 0.013968554325401783, + 0.013764685019850731, + 0.010965988971292973, + 0.013156355358660221, + 0.013194441795349121, + 0.013002966530621052, + 0.010485878214240074, + 0.007531374227255583, + 0.012167864479124546, + 0.023612113669514656, + 0.048998162150382996, + 0.02587869204580784, + 0.028439905494451523, + 0.03739950805902481, + 0.030657486990094185, + 0.020453063771128654, + 0.022296655923128128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.008641231805086136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187073335051537 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014187073335051537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187073335051537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0830869972705841, + 0.05475205183029175, + 0.18009741604328156, + 0.018302233889698982, + 0.04773025959730148, + 0.07904768735170364, + 0.27906712889671326, + 0.11781935393810272, + 0.07337935268878937, + 0.06671755015850067 + ], + "credit_map": [ + 0.008487953804433346, + 0.00597388856112957, + 0.01789749786257744, + 0.008473559282720089, + 0.00711468793451786, + 0.021061528474092484, + 0.0061269900761544704, + 0.007950888015329838, + 0.006855477578938007, + 0.007293677888810635, + 0.008534248918294907, + 0.011226129718124866, + 0.0053570163436234, + 0.005295167211443186, + 0.00643742922693491, + 0.0037529044784605503, + 0.012241407297551632, + 0.05358736962080002, + 0.0718667209148407, + 0.011809838935732841, + 0.008631120435893536, + 0.007733854465186596, + 0.007785675581544638, + 0.006441416218876839, + 0.0018151155672967434, + 0.001903618685901165, + 0.0037419935688376427, + 0.0021413657814264297, + 0.0021560334134846926, + 0.0023008729331195354, + 0.0016495379386469722, + 0.0025936937890946865, + 0.0031296436209231615, + 0.009239737875759602, + 0.005554842762649059, + 0.00446530245244503, + 0.00821309071034193, + 0.0042917439714074135, + 0.007907280698418617, + 0.004928612150251865, + 0.006394358351826668, + 0.011536434292793274, + 0.012230386026203632, + 0.00964565109461546, + 0.007997228763997555, + 0.006575261242687702, + 0.012917214073240757, + 0.011751139536499977, + 0.019353102892637253, + 0.029696151614189148, + 0.03940393775701523, + 0.04507051408290863, + 0.02549036219716072, + 0.03552553057670593, + 0.05602886900305748, + 0.02849864773452282, + 0.009424243122339249, + 0.01631275564432144, + 0.01436124462634325, + 0.01897498220205307, + 0.02086595818400383, + 0.012378587387502193, + 0.01232946291565895, + 0.013172106817364693, + 0.007185015827417374, + 0.010299276560544968, + 0.013081034645438194, + 0.007424092385917902, + 0.01760788820683956, + 0.00978869292885065, + 0.0035124013666063547, + 0.004480945877730846, + 0.008764020167291164, + 0.006806729827076197, + 0.007578812073916197, + 0.012692132033407688, + 0.0072710709646344185, + 0.008421606384217739, + 0.008204988203942776, + 0.006978186313062906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.014187073335051537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007183355046436191 + ], + "opposite_logprob_tokens": [ + -7.250718116760254 + ], + "logprob_reward": -0.0007183355046436191, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007183355046436191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17783568799495697, + 0.13207943737506866, + 0.11048856377601624, + 0.04711679741740227, + 0.05238855630159378, + 0.0736362487077713, + 0.12228497862815857, + 0.0909653902053833, + 0.09245068579912186, + 0.10075362026691437 + ], + "credit_map": [ + 0.021996762603521347, + 0.01530417613685131, + 0.03160635009407997, + 0.019369177520275116, + 0.015297611244022846, + 0.04044094309210777, + 0.01697707176208496, + 0.016843609511852264, + 0.024329151958227158, + 0.01729533076286316, + 0.0185539647936821, + 0.02386699989438057, + 0.015213960781693459, + 0.009102649055421352, + 0.013588561676442623, + 0.010128838010132313, + 0.013830992393195629, + 0.01064395997673273, + 0.03221200034022331, + 0.019614337012171745, + 0.008054191246628761, + 0.0066194310784339905, + 0.011880911886692047, + 0.007632745895534754, + 0.005174309480935335, + 0.004162435419857502, + 0.007437374442815781, + 0.005568949040025473, + 0.005525661166757345, + 0.005860215052962303, + 0.004477052483707666, + 0.008910801261663437, + 0.005692644976079464, + 0.00895509123802185, + 0.007095810491591692, + 0.005458359140902758, + 0.006790128070861101, + 0.005044714547693729, + 0.007152338977903128, + 0.00619947025552392, + 0.005488326773047447, + 0.010098349303007126, + 0.009814168326556683, + 0.018092267215251923, + 0.006339966785162687, + 0.007163102272897959, + 0.009718326851725578, + 0.006921745371073484, + 0.008343525230884552, + 0.014888772740960121, + 0.017760876566171646, + 0.021581562235951424, + 0.016352811828255653, + 0.013982482254505157, + 0.018932664766907692, + 0.010442293249070644, + 0.007312685716897249, + 0.012982753105461597, + 0.016024351119995117, + 0.01590270735323429, + 0.009875532239675522, + 0.011536350473761559, + 0.0086861876770854, + 0.008644823916256428, + 0.008182941935956478, + 0.006238000467419624, + 0.02213641069829464, + 0.008595980703830719, + 0.02186400070786476, + 0.013621699996292591, + 0.004666030872613192, + 0.007145621348172426, + 0.008609755896031857, + 0.01501246728003025, + 0.013096311129629612, + 0.01964901201426983, + 0.011367101222276688, + 0.01171672809869051, + 0.011133121326565742, + 0.010169129818677902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0007183355046436191 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "violently shaky first-person perspective camera places the viewer in a pitching ship cabin. Every violent pitch and jolt of the ship shakes the frame, simulating the dizzying sensation of losing balance amidst surging waves. A single, dim light, cast from a bare bulb hanging from the ceiling, carves a sharp path through the murky air filled with suspended dust. The beam shifts with the swaying of the ship, causing shadows to dance wildly within the room. This light illuminates the bleakness of the room: dark, water-stained wood panels on the walls, a simple fixed bunk with a thin, wrinkled mattress, and no one else. An unsteady gaze shifts to the small, salt-stained round porthole, revealing only the endless, roaring, rolling gray waves outside. These waves merge with the gloomy sky, creating a monotonous, oppressive gray world. The entire scene is extremely dark in tone, dominated by cold gray and dirty brown. The only warm color comes from the faint and flickering light, but this light not only does not bring warmth, but instead intensifies the deep, suffocating loneliness and desolation in the empty room.", + "seed": 21832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_12/index_1078_seed_21832_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.044463551789522174, + "vlm_logprob": -0.044463551789522174, + "vlm_yes_no_margin": 4.549999952316284, + "vlm_token_credit": -0.044463551789522174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18480026721954346 + ], + "opposite_logprob_tokens": [ + -1.8098002672195435 + ], + "logprob_reward": -0.18480026721954346, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18480026721954346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2656673192977905, + 0.03979097306728363, + 0.1048651710152626, + 0.036090556532144547, + 0.10071872919797897, + 0.030959047377109528, + 0.04055320844054222, + 0.033867765218019485, + 0.20343804359436035, + 0.14404919743537903 + ], + "credit_map": [ + 0.028181806206703186, + 0.037579212337732315, + 0.07567394524812698, + 0.02478509210050106, + 0.014446394518017769, + 0.02259449101984501, + 0.027212757617235184, + 0.03519359603524208, + 0.007042091805487871, + 0.003982279449701309, + 0.004061076790094376, + 0.0032728286460042, + 0.004699096083641052, + 0.00454009510576725, + 0.006431267596781254, + 0.005762238521128893, + 0.009870410896837711, + 0.032057031989097595, + 0.02817004732787609, + 0.008071226067841053, + 0.007686627097427845, + 0.006019536405801773, + 0.007195205893367529, + 0.00579508813098073, + 0.003784879809245467, + 0.004685861524194479, + 0.0032462948001921177, + 0.005422381218522787, + 0.004754757042974234, + 0.003447160590440035, + 0.006651108153164387, + 0.004098116420209408, + 0.006069159600883722, + 0.008069111034274101, + 0.0066296313889324665, + 0.01353122852742672, + 0.020611463114619255, + 0.008922875858843327, + 0.029501503333449364, + 0.007383759133517742, + 0.002875568578019738, + 0.008730283938348293, + 0.007983334362506866, + 0.003561155404895544, + 0.002198847709223628, + 0.0015904998872429132, + 0.002608116716146469, + 0.0014112420612946153, + 0.003003813326358795, + 0.007458345964550972, + 0.0047292220406234264, + 0.007675830274820328, + 0.0026488706935197115, + 0.0057941218838095665, + 0.004066547844558954, + 0.00517645850777626, + 0.0034381987061351538, + 0.006955883465707302, + 0.003783631371334195, + 0.0036655962467193604, + 0.0025430924724787474, + 0.004281748551875353, + 0.0027638161554932594, + 0.006435796152800322, + 0.034361809492111206, + 0.038302887231111526, + 0.041749488562345505, + 0.02419601008296013, + 0.01674775406718254, + 0.011697224341332912, + 0.016094885766506195, + 0.020287979394197464, + 0.008257721550762653, + 0.006734584458172321, + 0.005664939060807228, + 0.005737577099353075, + 0.005288476124405861, + 0.007009165361523628, + 0.05506136640906334, + 0.05029537156224251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.18480026721954346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084203496575356 + ], + "opposite_logprob_tokens": [ + -4.14108419418335 + ], + "logprob_reward": -0.016084203496575356, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084203496575356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2610468864440918, + 0.06731728464365005, + 0.07545408606529236, + 0.05060529708862305, + 0.040871668606996536, + 0.11187054216861725, + 0.024746205657720566, + 0.022298691794276237, + 0.18547993898391724, + 0.16030944883823395 + ], + "credit_map": [ + 0.026534169912338257, + 0.041031669825315475, + 0.07398007065057755, + 0.019765885546803474, + 0.014208806678652763, + 0.022295793518424034, + 0.020004596561193466, + 0.04322587326169014, + 0.011434609070420265, + 0.007278690580278635, + 0.008990496397018433, + 0.009370608255267143, + 0.006176685448735952, + 0.006918652448803186, + 0.00805920921266079, + 0.009088335558772087, + 0.009777144528925419, + 0.015071824193000793, + 0.009844431653618813, + 0.012322318740189075, + 0.007151507772505283, + 0.005759653635323048, + 0.007678985595703125, + 0.007848221808671951, + 0.004613956902176142, + 0.0053340415470302105, + 0.003968944773077965, + 0.007234722375869751, + 0.0054794494062662125, + 0.004044327884912491, + 0.014299344271421432, + 0.0056305075995624065, + 0.0023270135279744864, + 0.004001971799880266, + 0.0034544190857559443, + 0.00651213526725769, + 0.007190763484686613, + 0.0030847410671412945, + 0.010806615464389324, + 0.0034940054174512625, + 0.005914322566241026, + 0.03990502655506134, + 0.04057138040661812, + 0.008249311707913876, + 0.005736472550779581, + 0.0027037099935114384, + 0.003913779277354479, + 0.004876540973782539, + 0.002234439132735133, + 0.0040555414743721485, + 0.0034416562411934137, + 0.005117450840771198, + 0.001175362034700811, + 0.003544767387211323, + 0.0029778487514704466, + 0.0021991401445120573, + 0.002079279161989689, + 0.003503696992993355, + 0.002759945811703801, + 0.0030141854658722878, + 0.0021784661803394556, + 0.002747858641669154, + 0.002425853395834565, + 0.003589405445381999, + 0.03178340569138527, + 0.03288315236568451, + 0.04087655618786812, + 0.022383809089660645, + 0.014088183641433716, + 0.011961518786847591, + 0.013920709490776062, + 0.017582600936293602, + 0.009726742282509804, + 0.0093531534075737, + 0.005808803718537092, + 0.006276022642850876, + 0.010190747678279877, + 0.005481138825416565, + 0.059134893119335175, + 0.054337937384843826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.016084203496575356 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022034193389117718 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.0022034193389117718, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022034193389117718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18519747257232666, + 0.04854002594947815, + 0.1372900754213333, + 0.05613788589835167, + 0.054797086864709854, + 0.15611180663108826, + 0.05179031565785408, + 0.03433872386813164, + 0.049747299402952194, + 0.22604936361312866 + ], + "credit_map": [ + 0.009796136058866978, + 0.053596433252096176, + 0.012744549661874771, + 0.007568525616079569, + 0.026685643941164017, + 0.014459465630352497, + 0.00914816278964281, + 0.051198557019233704, + 0.006057691760361195, + 0.004743698053061962, + 0.004801182076334953, + 0.006087537854909897, + 0.006768724415451288, + 0.00445725629106164, + 0.009234458208084106, + 0.006389479152858257, + 0.013775496743619442, + 0.03521436080336571, + 0.032126497477293015, + 0.010293912142515182, + 0.013759682886302471, + 0.013056060299277306, + 0.008848550729453564, + 0.010215511545538902, + 0.007205847185105085, + 0.010558979585766792, + 0.0052400012500584126, + 0.008449098095297813, + 0.006491397973150015, + 0.006798283662647009, + 0.006674750242382288, + 0.004719529300928116, + 0.0038191648200154305, + 0.005094403866678476, + 0.0029371916316449642, + 0.007327136117964983, + 0.010944286361336708, + 0.0034648042637854815, + 0.016995886340737343, + 0.0042142122983932495, + 0.008673067204654217, + 0.05627076327800751, + 0.052712585777044296, + 0.014532449655234814, + 0.00856320932507515, + 0.004764636512845755, + 0.0055099730379879475, + 0.005085135344415903, + 0.0032933824695646763, + 0.006902613677084446, + 0.005992189049720764, + 0.009769678115844727, + 0.003030601888895035, + 0.010560944676399231, + 0.0053748153150081635, + 0.006866095587611198, + 0.0035381081979721785, + 0.007241642102599144, + 0.0037107905372977257, + 0.004525642842054367, + 0.003135946812108159, + 0.0033858444076031446, + 0.003717010607942939, + 0.005083738826215267, + 0.008527122437953949, + 0.009468227624893188, + 0.008452208712697029, + 0.005899527110159397, + 0.00402097450569272, + 0.0040011280216276646, + 0.004242461174726486, + 0.005135650280863047, + 0.014210875146090984, + 0.007238570135086775, + 0.007217319682240486, + 0.007873183116316795, + 0.005839094985276461, + 0.007195145823061466, + 0.09430092573165894, + 0.08217425644397736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0022034193389117718 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185783177614212 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.018185783177614212, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018185783177614212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13850530982017517, + 0.20542018115520477, + 0.15904204547405243, + 0.0833173468708992, + 0.10694895684719086, + 0.08287429064512253, + 0.06405073404312134, + 0.03814307972788811, + 0.031621161848306656, + 0.09007684141397476 + ], + "credit_map": [ + 0.013472733087837696, + 0.019834721460938454, + 0.031464286148548126, + 0.012744099833071232, + 0.012044013477861881, + 0.012112538330256939, + 0.017313135787844658, + 0.019519811496138573, + 0.019126519560813904, + 0.018963875249028206, + 0.02288726158440113, + 0.02324661612510681, + 0.02164880558848381, + 0.020137842744588852, + 0.04500774294137955, + 0.0344015434384346, + 0.02991820126771927, + 0.037094682455062866, + 0.035681113600730896, + 0.014074889943003654, + 0.011197803542017937, + 0.009190360084176064, + 0.009640405885875225, + 0.012244598008692265, + 0.007131560239940882, + 0.008050630800426006, + 0.005399566609412432, + 0.012121246196329594, + 0.00956050120294094, + 0.008160061202943325, + 0.018795661628246307, + 0.014098127372562885, + 0.0062008509412407875, + 0.009478643536567688, + 0.007010417524725199, + 0.018569618463516235, + 0.02141493745148182, + 0.004650746006518602, + 0.032126251608133316, + 0.007497505750507116, + 0.006681807804852724, + 0.019505685195326805, + 0.021282663568854332, + 0.010122070088982582, + 0.006951275747269392, + 0.005179916974157095, + 0.007214549463242292, + 0.0059363278560340405, + 0.004926659632474184, + 0.012101688422262669, + 0.008815096691250801, + 0.011394641362130642, + 0.0035548871383070946, + 0.009308434091508389, + 0.007436413783580065, + 0.006512925960123539, + 0.003620769362896681, + 0.006771591026335955, + 0.004316670820116997, + 0.004824673291295767, + 0.003133336780592799, + 0.004425442777574062, + 0.00320895342156291, + 0.00784164760261774, + 0.0046926094219088554, + 0.006985452026128769, + 0.004265647381544113, + 0.004173415247350931, + 0.0032744116615504026, + 0.0023923914413899183, + 0.0028190873563289642, + 0.0030181482434272766, + 0.01069450844079256, + 0.012250750325620174, + 0.009549072943627834, + 0.007969966158270836, + 0.013129200786352158, + 0.0065185134299099445, + 0.016262786462903023, + 0.01370205171406269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.018185783177614212 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010440857149660587 + ], + "opposite_logprob_tokens": [ + -6.876044273376465 + ], + "logprob_reward": -0.0010440857149660587, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010440857149660587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16176185011863708, + 0.051677923649549484, + 0.0691019669175148, + 0.05222295969724655, + 0.03484183922410011, + 0.0939209908246994, + 0.051515061408281326, + 0.01866171322762966, + 0.03937568515539169, + 0.42691999673843384 + ], + "credit_map": [ + 0.021568574011325836, + 0.024922626093029976, + 0.02528679557144642, + 0.015612480230629444, + 0.016329942271113396, + 0.012182910926640034, + 0.016442660242319107, + 0.029415855184197426, + 0.006325442809611559, + 0.005471725482493639, + 0.005926225334405899, + 0.005992598365992308, + 0.006543513387441635, + 0.00556562002748251, + 0.007146425545215607, + 0.008706372231245041, + 0.014924467541277409, + 0.01474067848175764, + 0.00906462874263525, + 0.006736577954143286, + 0.0064835622906684875, + 0.006156625226140022, + 0.005069564562290907, + 0.00592586537823081, + 0.006689423229545355, + 0.006127024069428444, + 0.005066701211035252, + 0.009261793456971645, + 0.0056570968590676785, + 0.005038572940975428, + 0.006280092988163233, + 0.008102251216769218, + 0.0026464746333658695, + 0.002869061892852187, + 0.002878698520362377, + 0.004055961966514587, + 0.0054770163260400295, + 0.0020713796839118004, + 0.009645496495068073, + 0.00519774854183197, + 0.005780620034784079, + 0.03385613486170769, + 0.03199945017695427, + 0.007695337291806936, + 0.004917114973068237, + 0.0027671868447214365, + 0.003576403483748436, + 0.0033287405967712402, + 0.004174528643488884, + 0.008393460884690285, + 0.006109439302235842, + 0.01000343356281519, + 0.002471343846991658, + 0.008951355703175068, + 0.004682514816522598, + 0.006728983484208584, + 0.0022136522457003593, + 0.0035454800818115473, + 0.002555601531639695, + 0.0020851739682257175, + 0.0015383227728307247, + 0.0016906539676710963, + 0.0016853574197739363, + 0.0033474715892225504, + 0.006956625729799271, + 0.00784421619027853, + 0.008806510828435421, + 0.005017441231757402, + 0.002616102108731866, + 0.002381537575274706, + 0.002346297027543187, + 0.0034069535322487354, + 0.02586059458553791, + 0.016610493883490562, + 0.01296214945614338, + 0.013862947933375835, + 0.010092022828757763, + 0.013143723830580711, + 0.17853310704231262, + 0.1558549702167511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0010440857149660587 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "violently shaky first-person perspective camera places the viewer in a pitching ship cabin. Every violent pitch and jolt of the ship shakes the frame, simulating the dizzying sensation of losing balance amidst surging waves. A single, dim light, cast from a bare bulb hanging from the ceiling, carves a sharp path through the murky air filled with suspended dust. The beam shifts with the swaying of the ship, causing shadows to dance wildly within the room. This light illuminates the bleakness of the room: dark, water-stained wood panels on the walls, a simple fixed bunk with a thin, wrinkled mattress, and no one else. An unsteady gaze shifts to the small, salt-stained round porthole, revealing only the endless, roaring, rolling gray waves outside. These waves merge with the gloomy sky, creating a monotonous, oppressive gray world. The entire scene is extremely dark in tone, dominated by cold gray and dirty brown. The only warm color comes from the faint and flickering light, but this light not only does not bring warmth, but instead intensifies the deep, suffocating loneliness and desolation in the empty room.", + "seed": 21833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_12/index_1078_seed_21833_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01854883509222418, + "vlm_logprob": -0.01854883509222418, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.01854883509222418, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07059350609779358 + ], + "opposite_logprob_tokens": [ + -2.8205935955047607 + ], + "logprob_reward": -0.07059350609779358, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07059350609779358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4204876720905304, + 0.04923807457089424, + 0.0688294768333435, + 0.06368157267570496, + 0.06281419843435287, + 0.03404079005122185, + 0.025206172838807106, + 0.18237686157226562, + 0.018440091982483864, + 0.07488510757684708 + ], + "credit_map": [ + 0.021337036043405533, + 0.039310477674007416, + 0.17060403525829315, + 0.02245500683784485, + 0.015646157786250114, + 0.12577952444553375, + 0.014883521944284439, + 0.010471900925040245, + 0.0038523701950907707, + 0.0055210473947227, + 0.005714703351259232, + 0.003911223262548447, + 0.006914123427122831, + 0.008497379720211029, + 0.008065661415457726, + 0.006761565804481506, + 0.006460759788751602, + 0.007796719670295715, + 0.0082730483263731, + 0.00869883131235838, + 0.00933823175728321, + 0.007502347230911255, + 0.010329566895961761, + 0.010429967194795609, + 0.006048255600035191, + 0.00879448838531971, + 0.006881541106849909, + 0.011069929227232933, + 0.005763276945799589, + 0.008771832101047039, + 0.006862256675958633, + 0.009489988908171654, + 0.008770476095378399, + 0.005312252324074507, + 0.005293380469083786, + 0.003954812418669462, + 0.006598277948796749, + 0.00603475421667099, + 0.009851261973381042, + 0.016998974606394768, + 0.0017781250644475222, + 0.0020422660745680332, + 0.010035518556833267, + 0.0055345455184578896, + 0.0040746089071035385, + 0.002303385641425848, + 0.005771247204393148, + 0.002501091454178095, + 0.001226256601512432, + 0.00801645964384079, + 0.0037833789829164743, + 0.0049484227783977985, + 0.0037178874481469393, + 0.0011051589390262961, + 0.0015889422502368689, + 0.0008196649141609669, + 0.009143073111772537, + 0.009802945889532566, + 0.0716857984662056, + 0.006350083276629448, + 0.00822720117866993, + 0.06709445267915726, + 0.0042794132605195045, + 0.005793880671262741, + 0.002654094249010086, + 0.0028147133998572826, + 0.0025554851163178682, + 0.0019563203677535057, + 0.001204677508212626, + 0.0034987982362508774, + 0.0014264762867242098, + 0.0023295246064662933, + 0.005871891509741545, + 0.022519465535879135, + 0.011102618649601936, + 0.01387932151556015, + 0.007958447560667992, + 0.0041956109926104546, + 0.004601706750690937, + 0.004756041336804628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.07059350609779358 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553953565657139 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012553953565657139, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553953565657139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3641110956668854, + 0.08872607350349426, + 0.11226314306259155, + 0.06913124024868011, + 0.029388779774308205, + 0.02565009705722332, + 0.03302772343158722, + 0.1659565418958664, + 0.025682879611849785, + 0.08606241643428802 + ], + "credit_map": [ + 0.02048049494624138, + 0.02441323734819889, + 0.1337103694677353, + 0.0200932789593935, + 0.017512353137135506, + 0.1197986751794815, + 0.014600302092730999, + 0.01350240595638752, + 0.008101297542452812, + 0.011967359110713005, + 0.013341284357011318, + 0.013787206262350082, + 0.010135720483958721, + 0.007819046266376972, + 0.012554388493299484, + 0.01101977564394474, + 0.016897981986403465, + 0.011235734447836876, + 0.013629615306854248, + 0.009243829175829887, + 0.013199916109442711, + 0.015262441709637642, + 0.017211178317666054, + 0.015582461841404438, + 0.007316346745938063, + 0.01096058078110218, + 0.007017910480499268, + 0.00938121136277914, + 0.00698461988940835, + 0.009561274200677872, + 0.008481970056891441, + 0.009427336044609547, + 0.0025687271263450384, + 0.0033952819649130106, + 0.0025405711494386196, + 0.002344131702557206, + 0.0029879147186875343, + 0.0066365888342261314, + 0.004296781960874796, + 0.004618787672370672, + 0.0014135389355942607, + 0.0018215798772871494, + 0.006189522333443165, + 0.00425793556496501, + 0.003080671187490225, + 0.0018865952733904123, + 0.0048805284313857555, + 0.002119729295372963, + 0.002531000878661871, + 0.005965894553810358, + 0.005198600236326456, + 0.006331420969218016, + 0.005451313219964504, + 0.00245546642690897, + 0.0029636514373123646, + 0.0021303805988281965, + 0.008107336238026619, + 0.008684678003191948, + 0.06513284146785736, + 0.005497577600181103, + 0.006630934774875641, + 0.0620901994407177, + 0.003990250173956156, + 0.005822746083140373, + 0.002462925622239709, + 0.00267585925757885, + 0.0035631582140922546, + 0.003452793462201953, + 0.0023257059510797262, + 0.005516625940799713, + 0.002746441401541233, + 0.0029393737204372883, + 0.006190038286149502, + 0.02541143260896206, + 0.012329373508691788, + 0.01776083931326866, + 0.007242269814014435, + 0.005048719700425863, + 0.005724490620195866, + 0.006355263292789459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.012553953565657139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036222815979272127 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036222815979272127, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036222815979272127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7459891438484192, + 0.023873232305049896, + 0.012468568049371243, + 0.05513277277350426, + 0.029884226620197296, + 0.0679253339767456, + 0.020004745572805405, + 0.019983116537332535, + 0.012674404308199883, + 0.012064504437148571 + ], + "credit_map": [ + 0.037480130791664124, + 0.07403268665075302, + 0.29729053378105164, + 0.03576849400997162, + 0.047971177846193314, + 0.20565113425254822, + 0.025230105966329575, + 0.022564880549907684, + 0.0021178103052079678, + 0.0025626234710216522, + 0.003298236755654216, + 0.003519900608807802, + 0.0023408986162394285, + 0.002730608219280839, + 0.003597097937017679, + 0.003706055460497737, + 0.00215904857032001, + 0.0013212125049903989, + 0.001845663064159453, + 0.0012539521558210254, + 0.0014327015960589051, + 0.0010972078889608383, + 0.0015136395813897252, + 0.0018451425712555647, + 0.005578909069299698, + 0.007230291608721018, + 0.004976903088390827, + 0.007848130539059639, + 0.006183375138789415, + 0.009344239719212055, + 0.006742974277585745, + 0.007227947004139423, + 0.003175504971295595, + 0.0030019041150808334, + 0.002436757553368807, + 0.002745044883340597, + 0.003060797927901149, + 0.0036518541164696217, + 0.004800058901309967, + 0.007012304849922657, + 0.0034597176127135754, + 0.004111127927899361, + 0.019882438704371452, + 0.013919061981141567, + 0.009739683009684086, + 0.003086329437792301, + 0.00955639686435461, + 0.004170581232756376, + 0.0014849741710349917, + 0.0035869311541318893, + 0.003798912977799773, + 0.003166197333484888, + 0.0026611664798110723, + 0.0017026836285367608, + 0.0022049471735954285, + 0.0013989332364872098, + 0.0009556452860124409, + 0.0013615747448056936, + 0.007512677926570177, + 0.0009228640701621771, + 0.000901638122741133, + 0.006550962571054697, + 0.0006641546497121453, + 0.001113599049858749, + 0.0017490909667685628, + 0.0023209780920296907, + 0.0017032326431944966, + 0.0012193679576739669, + 0.0007900622440502048, + 0.0018878127448260784, + 0.0011293473653495312, + 0.0018745128763839602, + 0.0011407228885218501, + 0.0035870419815182686, + 0.0017050029709935188, + 0.0017185938777402043, + 0.0012393329525366426, + 0.000673630042001605, + 0.001060227514244616, + 0.0009399521513842046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0036222815979272127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046363710425794125 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.0046363710425794125, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046363710425794125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3875894248485565, + 0.1338285356760025, + 0.1262492537498474, + 0.0499173142015934, + 0.03999839723110199, + 0.06927749514579773, + 0.07302756607532501, + 0.040034160017967224, + 0.032370008528232574, + 0.047707896679639816 + ], + "credit_map": [ + 0.018899643793702126, + 0.0204453244805336, + 0.13259375095367432, + 0.01913553476333618, + 0.028372932225465775, + 0.12881368398666382, + 0.023471955209970474, + 0.01585657149553299, + 0.01077287271618843, + 0.011189410462975502, + 0.015955839306116104, + 0.014084439724683762, + 0.019454220309853554, + 0.02173369750380516, + 0.021416597068309784, + 0.019221443682909012, + 0.016400810331106186, + 0.012018800713121891, + 0.014841247349977493, + 0.011806117370724678, + 0.016340728849172592, + 0.014989308081567287, + 0.0198071151971817, + 0.020045116543769836, + 0.0037440794985741377, + 0.006407239940017462, + 0.0036354365292936563, + 0.00463928421959281, + 0.0039364248514175415, + 0.013280857354402542, + 0.004155040252953768, + 0.010118949227035046, + 0.003177376464009285, + 0.0034177883062511683, + 0.003148892894387245, + 0.0025246390141546726, + 0.004139786586165428, + 0.009812887758016586, + 0.005834067240357399, + 0.007942954078316689, + 0.0036289156414568424, + 0.002912002382799983, + 0.021036401391029358, + 0.012884963303804398, + 0.008933718316257, + 0.0036929540801793337, + 0.012046369723975658, + 0.004142154008150101, + 0.004250745289027691, + 0.02185387723147869, + 0.012519522570073605, + 0.014314485713839531, + 0.006118424702435732, + 0.004333818797022104, + 0.005975870881229639, + 0.0036608108785003424, + 0.0042477804236114025, + 0.004200010094791651, + 0.0054683075286448, + 0.004803433082997799, + 0.004543073009699583, + 0.005415191408246756, + 0.0065964097157120705, + 0.004759947303682566, + 0.003692727070301771, + 0.00552357966080308, + 0.004573131445795298, + 0.003773616161197424, + 0.002913696924224496, + 0.004819606430828571, + 0.0035427033435553312, + 0.0035309467930346727, + 0.005594845861196518, + 0.009427480399608612, + 0.007559226360172033, + 0.00527386786416173, + 0.005794617347419262, + 0.0037453132681548595, + 0.004953945986926556, + 0.005358594935387373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0046363710425794125 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person and extremely shaky with violent pitching and jolting motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set inside a ship cabin featuring dark water-stained wood panels and a fixed bunk with a thin wrinkled mattress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single dim light source from a bare bulb hanging from the ceiling that casts a shifting beam through murky air filled with suspended dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small salt-stained round porthole visible showing endless roaring rolling gray waves merging with a gloomy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone extremely dark dominated by cold gray and dirty brown with only a faint flickering warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013380631571635604 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013380631571635604, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013380631571635604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37099677324295044, + 0.058882128447294235, + 0.06947942823171616, + 0.052513428032398224, + 0.024403197690844536, + 0.07120420783758163, + 0.01952773705124855, + 0.27623650431632996, + 0.017307063564658165, + 0.039449505507946014 + ], + "credit_map": [ + 0.020781513303518295, + 0.043275393545627594, + 0.14709848165512085, + 0.01940755546092987, + 0.022601541131734848, + 0.0894242599606514, + 0.01381751149892807, + 0.014590539038181305, + 0.006150033324956894, + 0.010246838442981243, + 0.008544580079615116, + 0.009298145771026611, + 0.005041331518441439, + 0.004257626365870237, + 0.007602154742926359, + 0.007741427514702082, + 0.01042299997061491, + 0.009260051883757114, + 0.007863209582865238, + 0.0074732499197125435, + 0.008112781681120396, + 0.007868294604122639, + 0.010129163973033428, + 0.008349686861038208, + 0.0038208290934562683, + 0.004992588888853788, + 0.0033235899172723293, + 0.004550762940198183, + 0.0023370312992483377, + 0.016938405111432076, + 0.004465047270059586, + 0.01208517886698246, + 0.0020033579785376787, + 0.001916923327371478, + 0.0035358325112611055, + 0.0022092873696237803, + 0.002476136200129986, + 0.0038824905641376972, + 0.003589141881093383, + 0.0047900318168103695, + 0.002886171219870448, + 0.0028131655417382717, + 0.0252293162047863, + 0.012011831626296043, + 0.01053664367645979, + 0.002421083627268672, + 0.009778151288628578, + 0.005527857691049576, + 0.0011940959375351667, + 0.004509979858994484, + 0.002778887515887618, + 0.003964294213801622, + 0.003049717517569661, + 0.0011714738793671131, + 0.0018972153775393963, + 0.0009620750206522644, + 0.012990009039640427, + 0.013137519359588623, + 0.10803117603063583, + 0.00859217718243599, + 0.010354076512157917, + 0.10831424593925476, + 0.005129815544933081, + 0.009687514044344425, + 0.0018000707495957613, + 0.0016173013718798757, + 0.0023853841703385115, + 0.0021052267402410507, + 0.0013271194184198976, + 0.002832191763445735, + 0.001613881322555244, + 0.003625890240073204, + 0.003367363242432475, + 0.010726779699325562, + 0.006217829417437315, + 0.006212776526808739, + 0.004462227690964937, + 0.002692208159714937, + 0.0033000484108924866, + 0.00247027981095016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0013380631571635604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_tensor([2210, 2210])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_tensor([2210, 2210])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1e06940b53a9a4c113d87cd357150c5895d0b978 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_tensor([2210, 2210])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled, extreme close-up captures a hand holding a rectangular power bank on a rich, dark wood-grained wooden table. The power bank features a modern design, with a matte black, subtly textured plastic shell and smooth, rounded edges. The hand, belonging to a young adult with fair skin and closely manicured nails, curiously holds the device. The thumb first traces the cool, smooth contours of the power bank's side, before the fingers carefully rotate the object. Then, with deliberate precision, the thumb moves to a small, circular button flush with the shell. The thumb pad presses firmly, and in response, a series of four tiny, pinprick-sized blue LEDs next to the button (previously dark) illuminate one by one, casting a faint, cool glow on the skin of the thumb. The entire scene is bathed in a soft, warm light from an unseen light source on the right, creating gentle shadows and highlighting the subtle textures of the skin and device, evoking an atmosphere of quiet, focused preparation.", + "seed": 29660, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_2210_seed_29660_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02422829531133175, + "vlm_logprob": -0.02422829531133175, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02422829531133175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0161399245262146 + ], + "opposite_logprob_tokens": [ + -6.266139984130859 + ], + "logprob_reward": -0.0161399245262146, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0161399245262146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14569757878780365, + 0.037220489233732224, + 0.04466306045651436, + 0.15681512653827667, + 0.05001801624894142, + 0.05973818153142929, + 0.07818087190389633, + 0.09751928597688675, + 0.12353106588125229, + 0.20661640167236328 + ], + "credit_map": [ + 0.03021389991044998, + 0.009257852099835873, + 0.020639881491661072, + 0.018684545531868935, + 0.010834514163434505, + 0.025494325906038284, + 0.011972369626164436, + 0.018600184470415115, + 0.008950717747211456, + 0.00191782065667212, + 0.00633143400773406, + 0.006104439962655306, + 0.0032887631095945835, + 0.005892355926334858, + 0.0020950196776539087, + 0.0026399362832307816, + 0.006317690946161747, + 0.0029797793831676245, + 0.006537796929478645, + 0.004210830200463533, + 0.00496651092544198, + 0.01202075183391571, + 0.002679570345208049, + 0.004950129892677069, + 0.020615145564079285, + 0.015809470787644386, + 0.024938732385635376, + 0.012151234783232212, + 0.03216058760881424, + 0.031228240579366684, + 0.00781771820038557, + 0.01209399662911892, + 0.006312476936727762, + 0.005669746547937393, + 0.007584684528410435, + 0.00433144997805357, + 0.009402320720255375, + 0.006953851785510778, + 0.004337724763900042, + 0.005425761919468641, + 0.007208588533103466, + 0.006139752920717001, + 0.008867722935974598, + 0.003951254766434431, + 0.012245766818523407, + 0.011959065683186054, + 0.003654654137790203, + 0.005711376201361418, + 0.010160580277442932, + 0.0056503391824662685, + 0.008674606680870056, + 0.010238339193165302, + 0.0063027082942426205, + 0.006238309666514397, + 0.018187478184700012, + 0.012728515081107616, + 0.013281896710395813, + 0.010290992446243763, + 0.01418332476168871, + 0.010022202506661415, + 0.012776842340826988, + 0.0076888916082680225, + 0.014757019467651844, + 0.01451811846345663, + 0.020358087494969368, + 0.013970574364066124, + 0.019536154344677925, + 0.016813715919852257, + 0.011643833480775356, + 0.017249446362257004, + 0.009012075141072273, + 0.014947174116969109, + 0.02590741030871868, + 0.018089698627591133, + 0.03538430482149124, + 0.011755443178117275, + 0.01976539008319378, + 0.048962175846099854, + 0.02861904725432396, + 0.018132928758859634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0161399245262146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305892944335938 + ], + "opposite_logprob_tokens": [ + -3.773305892944336 + ], + "logprob_reward": -0.023305892944335938, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305892944335938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34359925985336304, + 0.08997021615505219, + 0.05522293224930763, + 0.029530204832553864, + 0.03708796948194504, + 0.022276345640420914, + 0.05651745945215225, + 0.13124993443489075, + 0.05789991095662117, + 0.17664572596549988 + ], + "credit_map": [ + 0.040214769542217255, + 0.02458420768380165, + 0.04611176997423172, + 0.030000736936926842, + 0.02946573868393898, + 0.08206111192703247, + 0.027435947209596634, + 0.06372499465942383, + 0.011529923416674137, + 0.0066137006506323814, + 0.011425911448895931, + 0.010060001164674759, + 0.010492310859262943, + 0.014314022846519947, + 0.00811687856912613, + 0.01741746813058853, + 0.006990840658545494, + 0.0034890337847173214, + 0.008410302922129631, + 0.005634933244436979, + 0.006959952414035797, + 0.013608021661639214, + 0.0034761156421154737, + 0.006653734482824802, + 0.004302596673369408, + 0.0021761732641607523, + 0.004333709366619587, + 0.0032815493177622557, + 0.003763209329918027, + 0.005290641449391842, + 0.0018585574580356479, + 0.004523768089711666, + 0.005510405171662569, + 0.003339557908475399, + 0.005063636694103479, + 0.004759969189763069, + 0.0053100124932825565, + 0.0060201119631528854, + 0.0033657436724752188, + 0.0037185309920459986, + 0.002662699669599533, + 0.002034147270023823, + 0.0031808267813175917, + 0.0023957721423357725, + 0.00326109048910439, + 0.004024926107376814, + 0.0018538880394771695, + 0.002862995257601142, + 0.004819326568394899, + 0.005708025768399239, + 0.005675445310771465, + 0.005826155189424753, + 0.00559563934803009, + 0.006278323940932751, + 0.014324985444545746, + 0.008289556950330734, + 0.0153771648183465, + 0.018177464604377747, + 0.018682660534977913, + 0.010301027446985245, + 0.028276709839701653, + 0.013850737363100052, + 0.013747801072895527, + 0.012836363166570663, + 0.01114720106124878, + 0.00704804016277194, + 0.008398850448429585, + 0.006380760110914707, + 0.0072571211494505405, + 0.008230793289840221, + 0.004077015910297632, + 0.005360127426683903, + 0.019514959305524826, + 0.014496092684566975, + 0.02258840762078762, + 0.01710374467074871, + 0.018855225294828415, + 0.019293688237667084, + 0.04186978563666344, + 0.022923817858099937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.023305892944335938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062133315950632095 + ], + "opposite_logprob_tokens": [ + -2.812133312225342 + ], + "logprob_reward": -0.062133315950632095, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062133315950632095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12755022943019867, + 0.06128689646720886, + 0.02498609386384487, + 0.03603944554924965, + 0.14789821207523346, + 0.12424838542938232, + 0.128927543759346, + 0.18584784865379333, + 0.0702778697013855, + 0.09293738752603531 + ], + "credit_map": [ + 0.027815569192171097, + 0.00676727294921875, + 0.019425077363848686, + 0.016238218173384666, + 0.008940452709794044, + 0.02332438714802265, + 0.007838418707251549, + 0.01720084808766842, + 0.013658405281603336, + 0.003840943332761526, + 0.009661759249866009, + 0.0054802363738417625, + 0.005667254328727722, + 0.012582032941281796, + 0.003247810062021017, + 0.007148456759750843, + 0.005443199537694454, + 0.0015162840718403459, + 0.0038798607420176268, + 0.002244720933958888, + 0.002584073692560196, + 0.0052320100367069244, + 0.0012992427218705416, + 0.002786705270409584, + 0.005638906732201576, + 0.002205251483246684, + 0.005412442144006491, + 0.004326886031776667, + 0.0038211082573980093, + 0.010609086602926254, + 0.0015714075416326523, + 0.0024543567560613155, + 0.01768684573471546, + 0.014146022498607635, + 0.023295562714338303, + 0.011918880976736546, + 0.029987452551722527, + 0.0247013159096241, + 0.011877129785716534, + 0.014285024255514145, + 0.01747020334005356, + 0.013180158101022243, + 0.01767672970890999, + 0.008957846090197563, + 0.02350594662129879, + 0.023876404389739037, + 0.008311742916703224, + 0.011269363574683666, + 0.01653232052922249, + 0.01158587820827961, + 0.018690306693315506, + 0.010877110995352268, + 0.013292213901877403, + 0.021475329995155334, + 0.021661505103111267, + 0.014812881126999855, + 0.017822561785578728, + 0.016818201169371605, + 0.023384639993309975, + 0.013047964312136173, + 0.0333377830684185, + 0.0174070093780756, + 0.04321468248963356, + 0.020815033465623856, + 0.00859824102371931, + 0.007038971874862909, + 0.009013306349515915, + 0.010101373307406902, + 0.007126124110072851, + 0.009340579621493816, + 0.007687246892601252, + 0.011372033506631851, + 0.010839618742465973, + 0.005986139178276062, + 0.013135876506567001, + 0.008932341821491718, + 0.007742635905742645, + 0.025038478896021843, + 0.013410652056336403, + 0.007851654663681984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.062133315950632095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783178567886353 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783178567886353, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783178567886353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0450298897922039, + 0.023144399747252464, + 0.012936346232891083, + 0.027546020224690437, + 0.05029412358999252, + 0.1034267246723175, + 0.22752122581005096, + 0.07107912749052048, + 0.0896276906132698, + 0.34939447045326233 + ], + "credit_map": [ + 0.009778384119272232, + 0.00306537002325058, + 0.00712242815643549, + 0.005177655722945929, + 0.002755741821601987, + 0.004236446227878332, + 0.004440685734152794, + 0.008453180082142353, + 0.00282532861456275, + 0.002220352180302143, + 0.003218193305656314, + 0.0021331612952053547, + 0.0023362156935036182, + 0.004677537363022566, + 0.0021713085006922483, + 0.0035623020958155394, + 0.0017929720925167203, + 0.0008897876832634211, + 0.002099249977618456, + 0.0012365917209535837, + 0.002343772677704692, + 0.0022183824330568314, + 0.0008505342993885279, + 0.0015050559304654598, + 0.0034277166705578566, + 0.0024617330636829138, + 0.004255370702594519, + 0.002409636974334717, + 0.005980524700134993, + 0.00397847592830658, + 0.0019291207427158952, + 0.0031034427229315042, + 0.007181847933679819, + 0.004753900691866875, + 0.007645664736628532, + 0.004764306824654341, + 0.008500555530190468, + 0.00700451759621501, + 0.0032832385040819645, + 0.0071600922383368015, + 0.012128627859055996, + 0.010984603315591812, + 0.01517149806022644, + 0.008106219582259655, + 0.021568777039647102, + 0.017940888181328773, + 0.00741945393383503, + 0.010106655769050121, + 0.022952767089009285, + 0.022403404116630554, + 0.030263938009738922, + 0.017781775444746017, + 0.01988386921584606, + 0.0326174795627594, + 0.047038767486810684, + 0.03457922860980034, + 0.00678285351023078, + 0.008175496943295002, + 0.009447585791349411, + 0.006355766206979752, + 0.015052453614771366, + 0.008415615186095238, + 0.008830140344798565, + 0.008019212633371353, + 0.01243390329182148, + 0.008990099653601646, + 0.014139038510620594, + 0.013412215746939182, + 0.007805232424288988, + 0.009003844112157822, + 0.008940130472183228, + 0.014903229661285877, + 0.036897651851177216, + 0.02769026905298233, + 0.046880342066287994, + 0.036963775753974915, + 0.035494882613420486, + 0.0454961434006691, + 0.07846927642822266, + 0.04150211066007614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.009783178567886353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977916456758976 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.00977916456758976, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977916456758976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.183371901512146, + 0.050396502017974854, + 0.04737149551510811, + 0.034129198640584946, + 0.10377583652734756, + 0.16694720089435577, + 0.11552593111991882, + 0.01890210062265396, + 0.036723822355270386, + 0.24285604059696198 + ], + "credit_map": [ + 0.04228111729025841, + 0.008523612283170223, + 0.029043857008218765, + 0.02972426451742649, + 0.010755505412817001, + 0.02034521847963333, + 0.014963106252253056, + 0.027735209092497826, + 0.009273079223930836, + 0.003269071225076914, + 0.007548397872596979, + 0.005761838052421808, + 0.004845087882131338, + 0.0071981861256062984, + 0.004701189696788788, + 0.007799651939421892, + 0.008972742594778538, + 0.003332681255415082, + 0.007421549875289202, + 0.0068067763932049274, + 0.006341366097331047, + 0.007626118138432503, + 0.003276415169239044, + 0.003593847621232271, + 0.004935785662382841, + 0.00320756365545094, + 0.005167276132851839, + 0.003815349191427231, + 0.005907353479415178, + 0.006228425540030003, + 0.001755419303663075, + 0.0031120264902710915, + 0.011567922309041023, + 0.009970935061573982, + 0.01467859372496605, + 0.011806339025497437, + 0.017234256491065025, + 0.016312602907419205, + 0.009167690761387348, + 0.013037499971687794, + 0.019159091636538506, + 0.01739683747291565, + 0.02386617846786976, + 0.012970381416380405, + 0.03499329835176468, + 0.03268331661820412, + 0.011079943738877773, + 0.014798140153288841, + 0.012020072899758816, + 0.012635455466806889, + 0.016701964661478996, + 0.011088663712143898, + 0.008574090898036957, + 0.011793036945164204, + 0.02740299329161644, + 0.015309646725654602, + 0.001908308593556285, + 0.0018512139795348048, + 0.00216269469819963, + 0.0014286880614235997, + 0.002082979306578636, + 0.0022281671408563852, + 0.004927656147629023, + 0.002312391297891736, + 0.0045676082372665405, + 0.003811480710282922, + 0.0046709198504686356, + 0.004689685069024563, + 0.00407298281788826, + 0.004012924153357744, + 0.003982021473348141, + 0.006916197948157787, + 0.029340438544750214, + 0.01955229789018631, + 0.03889721632003784, + 0.018632855266332626, + 0.023738425225019455, + 0.04439641907811165, + 0.04723450168967247, + 0.021063867956399918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.00977916456758976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, extreme close-up captures a hand holding a rectangular power bank on a rich, dark wood-grained wooden table. The power bank features a modern design, with a matte black, subtly textured plastic shell and smooth, rounded edges. The hand, belonging to a young adult with fair skin and closely manicured nails, curiously holds the device. The thumb first traces the cool, smooth contours of the power bank's side, before the fingers carefully rotate the object. Then, with deliberate precision, the thumb moves to a small, circular button flush with the shell. The thumb pad presses firmly, and in response, a series of four tiny, pinprick-sized blue LEDs next to the button (previously dark) illuminate one by one, casting a faint, cool glow on the skin of the thumb. The entire scene is bathed in a soft, warm light from an unseen light source on the right, creating gentle shadows and highlighting the subtle textures of the skin and device, evoking an atmosphere of quiet, focused preparation.", + "seed": 29661, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_2210_seed_29661_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024051358550786973, + "vlm_logprob": -0.024051358550786973, + "vlm_yes_no_margin": 4.399999952316284, + "vlm_token_credit": -0.024051358550786973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02150048315525055 + ], + "opposite_logprob_tokens": [ + -7.021500587463379 + ], + "logprob_reward": -0.02150048315525055, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.02150048315525055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10767965018749237, + 0.07858934253454208, + 0.17101137340068817, + 0.13648760318756104, + 0.07254769653081894, + 0.046770840883255005, + 0.06341772526502609, + 0.13821332156658173, + 0.127745121717453, + 0.057537294924259186 + ], + "credit_map": [ + 0.014763559214770794, + 0.007473978213965893, + 0.016641533002257347, + 0.02171996794641018, + 0.01323843002319336, + 0.007202665321528912, + 0.013902708888053894, + 0.012736808508634567, + 0.007331259548664093, + 0.008956795558333397, + 0.0089117968454957, + 0.009848264046013355, + 0.008235029876232147, + 0.009100413881242275, + 0.0152305131778121, + 0.010975276120007038, + 0.017973514273762703, + 0.01551468763500452, + 0.020220914855599403, + 0.029465163126587868, + 0.014371892437338829, + 0.014971127733588219, + 0.03871659189462662, + 0.01977747492492199, + 0.01623331755399704, + 0.008740254677832127, + 0.016951318830251694, + 0.017783693969249725, + 0.009008427150547504, + 0.017756657674908638, + 0.03238636255264282, + 0.017627572640776634, + 0.007472177967429161, + 0.005952074192464352, + 0.008224875666201115, + 0.011543627828359604, + 0.005228519905358553, + 0.007978436537086964, + 0.017003867775201797, + 0.009144115261733532, + 0.00511948624625802, + 0.003522330429404974, + 0.005751327145844698, + 0.01020206417888403, + 0.0034169554710388184, + 0.0037788974586874247, + 0.009771060198545456, + 0.0052087195217609406, + 0.005934423301368952, + 0.004161237273365259, + 0.009450949728488922, + 0.015672512352466583, + 0.0036189816892147064, + 0.005107537377625704, + 0.012931924313306808, + 0.0065401531755924225, + 0.018087267875671387, + 0.011081025004386902, + 0.018493646755814552, + 0.033514201641082764, + 0.00828703586012125, + 0.008809368126094341, + 0.02383032813668251, + 0.016110466793179512, + 0.015249466523528099, + 0.010092880576848984, + 0.018884586170315742, + 0.03421054407954216, + 0.0067489659413695335, + 0.006853914353996515, + 0.024182025343179703, + 0.01152273640036583, + 0.008325589820742607, + 0.005254710558801889, + 0.0071275304071605206, + 0.012902507558465004, + 0.0035129135940223932, + 0.004508838057518005, + 0.0076701464131474495, + 0.008235056884586811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.02150048315525055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635873481631279 + ], + "opposite_logprob_tokens": [ + -3.6513588428497314 + ], + "logprob_reward": -0.02635873481631279, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635873481631279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2971065044403076, + 0.12298133224248886, + 0.2299356460571289, + 0.06030670925974846, + 0.01194249652326107, + 0.030547479167580605, + 0.04033491760492325, + 0.08218560367822647, + 0.04970389977097511, + 0.07495538145303726 + ], + "credit_map": [ + 0.04313717037439346, + 0.023224974051117897, + 0.047687649726867676, + 0.06038054823875427, + 0.052723150700330734, + 0.020013226196169853, + 0.030883820727467537, + 0.019055983051657677, + 0.014653586782515049, + 0.009630437009036541, + 0.014992681331932545, + 0.021629516035318375, + 0.012008560821413994, + 0.0124166589230299, + 0.018479526042938232, + 0.019170379266142845, + 0.02366066165268421, + 0.020608894526958466, + 0.026800543069839478, + 0.039300452917814255, + 0.01931016705930233, + 0.021124793216586113, + 0.05102017521858215, + 0.028109971433877945, + 0.0055336058139801025, + 0.004424682818353176, + 0.0070879035629332066, + 0.009484713897109032, + 0.004801157861948013, + 0.007544462103396654, + 0.014476444572210312, + 0.006953742820769548, + 0.001312813488766551, + 0.0009257663041353226, + 0.0019509777193889022, + 0.002096226904541254, + 0.0008602694724686444, + 0.0010674754157662392, + 0.0021611065603792667, + 0.0015678612980991602, + 0.0029191719368100166, + 0.00160156877245754, + 0.0042681521736085415, + 0.008233398199081421, + 0.002335380995646119, + 0.002630967181175947, + 0.005148420576006174, + 0.0034104210790246725, + 0.005045450758188963, + 0.0032732055988162756, + 0.00506735872477293, + 0.007337340619415045, + 0.0027065319009125233, + 0.003486211411654949, + 0.00776422256603837, + 0.005654600448906422, + 0.010477735660970211, + 0.005744506139308214, + 0.010670982301235199, + 0.017274172976613045, + 0.005272320471704006, + 0.006934701930731535, + 0.014213438145816326, + 0.011597748845815659, + 0.005749623756855726, + 0.003829775843769312, + 0.006750273052603006, + 0.01105537824332714, + 0.003676628228276968, + 0.0036205612123012543, + 0.009530847892165184, + 0.0054908134043216705, + 0.0070022395811975, + 0.007419530302286148, + 0.009258435107767582, + 0.009023510850965977, + 0.008474646136164665, + 0.009836206212639809, + 0.01461592223495245, + 0.009324898011982441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.02635873481631279 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814366087317467 + ], + "opposite_logprob_tokens": [ + -3.2881436347961426 + ], + "logprob_reward": -0.03814366087317467, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814366087317467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18628282845020294, + 0.3140825033187866, + 0.05374223738908768, + 0.07922909408807755, + 0.15148505568504333, + 0.03555592522025108, + 0.027742234990000725, + 0.043718017637729645, + 0.07736190408468246, + 0.0308002270758152 + ], + "credit_map": [ + 0.016397016122937202, + 0.018762309104204178, + 0.02210521139204502, + 0.029721349477767944, + 0.03881526365876198, + 0.014745495282113552, + 0.02092408575117588, + 0.024812089279294014, + 0.03597822040319443, + 0.022057443857192993, + 0.04321131482720375, + 0.05067269131541252, + 0.03343517705798149, + 0.03343870863318443, + 0.05164942517876625, + 0.04363951086997986, + 0.005522690247744322, + 0.0036704076919704676, + 0.00684580160304904, + 0.008565104566514492, + 0.004780117888003588, + 0.008448010310530663, + 0.008314475417137146, + 0.007595631293952465, + 0.008620609529316425, + 0.005730201490223408, + 0.010184093378484249, + 0.010479298420250416, + 0.005479452200233936, + 0.010431412607431412, + 0.018564686179161072, + 0.009739345870912075, + 0.016881879419088364, + 0.011800129897892475, + 0.017534496262669563, + 0.023564718663692474, + 0.011828791350126266, + 0.016399288550019264, + 0.033650290220975876, + 0.019825471565127373, + 0.005891953594982624, + 0.0016863339114934206, + 0.004901271313428879, + 0.008852624334394932, + 0.002271045232191682, + 0.002658446552231908, + 0.005668293219059706, + 0.0036259572952985764, + 0.002931323368102312, + 0.0017225960036739707, + 0.003918981179594994, + 0.006460198666900396, + 0.0019697912503033876, + 0.002676332136616111, + 0.004305321257561445, + 0.0037576924078166485, + 0.005044238641858101, + 0.0032961152028292418, + 0.006089215166866779, + 0.01090753823518753, + 0.0026788502000272274, + 0.0031314215157181025, + 0.007347633596509695, + 0.0052230097353458405, + 0.010745125822722912, + 0.007938715629279613, + 0.009877553209662437, + 0.018385451287031174, + 0.004398352932184935, + 0.004026097245514393, + 0.014248332008719444, + 0.007742284797132015, + 0.0031755217351019382, + 0.0026424089446663857, + 0.004038309678435326, + 0.004576895385980606, + 0.002827141899615526, + 0.0027936010155826807, + 0.006393520161509514, + 0.004352830350399017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.03814366087317467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187187612056732 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018187187612056732, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018187187612056732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16261982917785645, + 0.10911277681589127, + 0.04376104101538658, + 0.04258151724934578, + 0.05473240837454796, + 0.02112617902457714, + 0.10826078057289124, + 0.0910707637667656, + 0.34562137722969055, + 0.021113330498337746 + ], + "credit_map": [ + 0.020091716200113297, + 0.016115184873342514, + 0.02017633430659771, + 0.026771198958158493, + 0.03420863300561905, + 0.012837007641792297, + 0.018398767337203026, + 0.014020979404449463, + 0.012092744000256062, + 0.008395025506615639, + 0.016034603118896484, + 0.015410312451422215, + 0.012200508266687393, + 0.010445190593600273, + 0.020636118948459625, + 0.013898271135985851, + 0.0054816086776554585, + 0.003040377050638199, + 0.004803900606930256, + 0.007722066715359688, + 0.0027882957365363836, + 0.003991669975221157, + 0.010304819792509079, + 0.0056283012963831425, + 0.004098943900316954, + 0.0036405993159860373, + 0.007221568375825882, + 0.005322800949215889, + 0.002669903915375471, + 0.0071058813482522964, + 0.007999851368367672, + 0.004521967843174934, + 0.005354811903089285, + 0.004899939987808466, + 0.006601207423955202, + 0.007284060120582581, + 0.003923366777598858, + 0.0062993476167321205, + 0.011996595188975334, + 0.008373080752789974, + 0.0019215461798012257, + 0.0016088738339021802, + 0.002649135421961546, + 0.004362665116786957, + 0.0017053784104064107, + 0.001904598786495626, + 0.0036852078046649694, + 0.003288772189989686, + 0.01250334084033966, + 0.008022272028028965, + 0.014936289750039577, + 0.026255235075950623, + 0.005852511618286371, + 0.01040400005877018, + 0.018391592428088188, + 0.011895541101694107, + 0.010010289028286934, + 0.007395740132778883, + 0.01265972014516592, + 0.021852949634194374, + 0.0050634522922337055, + 0.006316755432635546, + 0.016722165048122406, + 0.011049691587686539, + 0.03666694834828377, + 0.032555948942899704, + 0.04743875935673714, + 0.0848398506641388, + 0.021447788923978806, + 0.018515694886446, + 0.06676816195249557, + 0.03738822042942047, + 0.002361044753342867, + 0.00204350333660841, + 0.0025611850433051586, + 0.0030646738596260548, + 0.0018137593287974596, + 0.0021281002555042505, + 0.003858318319544196, + 0.003282745135948062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.018187187612056732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606672629714012 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.01606672629714012, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606672629714012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31588807702064514, + 0.04510911926627159, + 0.03101368248462677, + 0.04544650763273239, + 0.16227968037128448, + 0.09733813256025314, + 0.053344856947660446, + 0.12139708548784256, + 0.08205194771289825, + 0.046130940318107605 + ], + "credit_map": [ + 0.04559557884931564, + 0.021671151742339134, + 0.05108972638845444, + 0.054390113800764084, + 0.06895836442708969, + 0.022869782522320747, + 0.03274649754166603, + 0.01856684498488903, + 0.0039119054563343525, + 0.003524789121001959, + 0.005718803033232689, + 0.006638077087700367, + 0.0040705446153879166, + 0.005909599829465151, + 0.007148385047912598, + 0.008187011815607548, + 0.0032504177652299404, + 0.0025029524695128202, + 0.0044218627735972404, + 0.004036694299429655, + 0.0023622598964720964, + 0.003745834808796644, + 0.005496069323271513, + 0.005197588820010424, + 0.008810008876025677, + 0.002415207913145423, + 0.006989825516939163, + 0.006359711289405823, + 0.00295455614104867, + 0.005878046620637178, + 0.006897623650729656, + 0.005141523666679859, + 0.017667267471551895, + 0.012673648074269295, + 0.018900852650403976, + 0.02751643769443035, + 0.012532215565443039, + 0.01800849288702011, + 0.03427552431821823, + 0.02070523239672184, + 0.01562187634408474, + 0.004179688170552254, + 0.012959195300936699, + 0.02531352825462818, + 0.005302417557686567, + 0.007924534380435944, + 0.01385441329330206, + 0.012182467617094517, + 0.004538467153906822, + 0.004195471294224262, + 0.007122463081032038, + 0.013207096606492996, + 0.002767699072137475, + 0.004232030361890793, + 0.00990369450300932, + 0.007377936504781246, + 0.01566232554614544, + 0.009234226308763027, + 0.016635287553071976, + 0.02979181334376335, + 0.007106636185199022, + 0.00815650075674057, + 0.020777009427547455, + 0.014033272862434387, + 0.00798057671636343, + 0.007497054059058428, + 0.010899968445301056, + 0.018143855035305023, + 0.005714554339647293, + 0.005279002245515585, + 0.01578529179096222, + 0.010751633904874325, + 0.0049045574851334095, + 0.004376502241939306, + 0.005964953918009996, + 0.005774086806923151, + 0.004133279900997877, + 0.004361649509519339, + 0.009359360672533512, + 0.007256547454744577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.01606672629714012 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, extreme close-up captures a hand holding a rectangular power bank on a rich, dark wood-grained wooden table. The power bank features a modern design, with a matte black, subtly textured plastic shell and smooth, rounded edges. The hand, belonging to a young adult with fair skin and closely manicured nails, curiously holds the device. The thumb first traces the cool, smooth contours of the power bank's side, before the fingers carefully rotate the object. Then, with deliberate precision, the thumb moves to a small, circular button flush with the shell. The thumb pad presses firmly, and in response, a series of four tiny, pinprick-sized blue LEDs next to the button (previously dark) illuminate one by one, casting a faint, cool glow on the skin of the thumb. The entire scene is bathed in a soft, warm light from an unseen light source on the right, creating gentle shadows and highlighting the subtle textures of the skin and device, evoking an atmosphere of quiet, focused preparation.", + "seed": 29662, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_2210_seed_29662_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027070524357259273, + "vlm_logprob": -0.027070524357259273, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.027070524357259273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018902407959103584 + ], + "opposite_logprob_tokens": [ + -5.89390230178833 + ], + "logprob_reward": -0.018902407959103584, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.018902407959103584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09191086143255234, + 0.10673587024211884, + 0.12387817353010178, + 0.08544701337814331, + 0.07855655252933502, + 0.04735160991549492, + 0.06198945641517639, + 0.3347392976284027, + 0.034079860895872116, + 0.03531132638454437 + ], + "credit_map": [ + 0.011332154273986816, + 0.00899520143866539, + 0.02128615230321884, + 0.007303639315068722, + 0.01170328725129366, + 0.01675439067184925, + 0.006513449363410473, + 0.008022572845220566, + 0.013423547148704529, + 0.010913996957242489, + 0.016295205801725388, + 0.007389663718640804, + 0.013459093868732452, + 0.019044985994696617, + 0.011509833857417107, + 0.014699531719088554, + 0.010408527217805386, + 0.00858384370803833, + 0.01864440366625786, + 0.0051035694777965546, + 0.02472532168030739, + 0.037685051560401917, + 0.007275036536157131, + 0.011452405713498592, + 0.014398211613297462, + 0.008610651828348637, + 0.014778036624193192, + 0.005590911954641342, + 0.007736385799944401, + 0.015876146033406258, + 0.0074555822648108006, + 0.01100108027458191, + 0.011529028415679932, + 0.010200566612184048, + 0.0112505117431283, + 0.008290974423289299, + 0.007987749762833118, + 0.01085889246314764, + 0.01029561273753643, + 0.008143207058310509, + 0.006305964197963476, + 0.0062692975625395775, + 0.0069525097496807575, + 0.004820375237613916, + 0.0045426469296216965, + 0.005303678568452597, + 0.005715657956898212, + 0.007441474590450525, + 0.012687922455370426, + 0.014953989535570145, + 0.007216956000775099, + 0.0030913865193724632, + 0.004924052394926548, + 0.009368772618472576, + 0.004462086595594883, + 0.005284285172820091, + 0.030341017991304398, + 0.052450262010097504, + 0.05174277350306511, + 0.03601646050810814, + 0.05433795601129532, + 0.03210495039820671, + 0.05006270110607147, + 0.027683136984705925, + 0.0039320215582847595, + 0.004317025654017925, + 0.005140744615346193, + 0.00615620706230402, + 0.0033235924784094095, + 0.003189573995769024, + 0.00319623458199203, + 0.004824455827474594, + 0.0036014143843203783, + 0.0034566111862659454, + 0.006842348724603653, + 0.002633495721966028, + 0.004571187775582075, + 0.006293697282671928, + 0.003463849425315857, + 0.004448719788342714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.018902407959103584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498984083533287 + ], + "opposite_logprob_tokens": [ + -2.929989814758301 + ], + "logprob_reward": -0.05498984083533287, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498984083533287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12700434029102325, + 0.07386746257543564, + 0.1348782479763031, + 0.11091465502977371, + 0.22144676744937897, + 0.11322706937789917, + 0.1339341104030609, + 0.029863376170396805, + 0.028393452987074852, + 0.026470473036170006 + ], + "credit_map": [ + 0.014632461592555046, + 0.016894327476620674, + 0.023651907220482826, + 0.014127135276794434, + 0.01756427250802517, + 0.018591128289699554, + 0.010172604583203793, + 0.011370517313480377, + 0.008161221630871296, + 0.011392464861273766, + 0.00951338466256857, + 0.005682809744030237, + 0.012374158948659897, + 0.008954215794801712, + 0.009710402227938175, + 0.00807880237698555, + 0.015375904738903046, + 0.021442348137497902, + 0.021144045516848564, + 0.01615242287516594, + 0.015027666464447975, + 0.01733512617647648, + 0.014593109488487244, + 0.013807638548314571, + 0.014883408322930336, + 0.011705942451953888, + 0.016949592158198357, + 0.010301589034497738, + 0.010612539947032928, + 0.022662201896309853, + 0.011327369138598442, + 0.012472020462155342, + 0.026533707976341248, + 0.0315835066139698, + 0.02575678378343582, + 0.021859874948859215, + 0.022404896095395088, + 0.03480418398976326, + 0.033880725502967834, + 0.024623101577162743, + 0.014631202444434166, + 0.016906145960092545, + 0.01834356226027012, + 0.013188991695642471, + 0.010799378156661987, + 0.014200652949512005, + 0.013993220403790474, + 0.011163927614688873, + 0.028334544971585274, + 0.05289166793227196, + 0.011302705854177475, + 0.0048634884878993034, + 0.017178360372781754, + 0.005954388063400984, + 0.0064874025993049145, + 0.006921561434864998, + 0.0030787496361881495, + 0.004706585314124823, + 0.004855161998420954, + 0.0026263047475367785, + 0.0038869448471814394, + 0.003090877551585436, + 0.0036114281974732876, + 0.004007325973361731, + 0.0027807701844722033, + 0.005310100503265858, + 0.0040993173606693745, + 0.003885401878505945, + 0.004393286537379026, + 0.0026181095745414495, + 0.0031223930418491364, + 0.002184076700359583, + 0.002749537117779255, + 0.0035702341701835394, + 0.00427937600761652, + 0.0035591903142631054, + 0.0029262725729495287, + 0.003939785994589329, + 0.002465807134285569, + 0.002980269258841872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.05498984083533287 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029850274324417114 + ], + "opposite_logprob_tokens": [ + -3.5298502445220947 + ], + "logprob_reward": -0.029850274324417114, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029850274324417114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23213563859462738, + 0.1193387359380722, + 0.0739673376083374, + 0.08909652382135391, + 0.05570882558822632, + 0.07194303721189499, + 0.07755774259567261, + 0.21262389421463013, + 0.04248364269733429, + 0.02514464221894741 + ], + "credit_map": [ + 0.024447545409202576, + 0.029059462249279022, + 0.040398940443992615, + 0.03160418942570686, + 0.02892802655696869, + 0.0258405152708292, + 0.02260645106434822, + 0.0292504895478487, + 0.013583788648247719, + 0.012271680869162083, + 0.018295783549547195, + 0.009260272607207298, + 0.01495357695966959, + 0.01482094544917345, + 0.011393005028367043, + 0.02475968934595585, + 0.00641372986137867, + 0.0070090568624436855, + 0.009488440118730068, + 0.005213071592152119, + 0.012364521622657776, + 0.01642422378063202, + 0.007668626494705677, + 0.0093856705352664, + 0.01725141704082489, + 0.0067618233151733875, + 0.009835263714194298, + 0.004053060431033373, + 0.010683316737413406, + 0.021881647408008575, + 0.006677146069705486, + 0.011952849105000496, + 0.007830125279724598, + 0.007839573547244072, + 0.007145978044718504, + 0.004710127599537373, + 0.0060168784111738205, + 0.008204973302781582, + 0.007883810438215733, + 0.006077357567846775, + 0.009050541557371616, + 0.0072456952184438705, + 0.008054854348301888, + 0.00593770993873477, + 0.006894640624523163, + 0.011191986501216888, + 0.009724941104650497, + 0.013842671178281307, + 0.02163175866007805, + 0.030713820829987526, + 0.007329812739044428, + 0.002452921587973833, + 0.0073739611543715, + 0.0032006921246647835, + 0.0023424550890922546, + 0.002512322971597314, + 0.017643829807639122, + 0.0302627794444561, + 0.04452664405107498, + 0.020914627239108086, + 0.037750985473394394, + 0.01588503271341324, + 0.028454985469579697, + 0.01718500629067421, + 0.004323057364672422, + 0.005665332078933716, + 0.0069610257633030415, + 0.004131009336560965, + 0.0051054940558969975, + 0.005847327411174774, + 0.003677983535453677, + 0.006772415246814489, + 0.0022006924264132977, + 0.003349656006321311, + 0.003409882541745901, + 0.003698852611705661, + 0.003509389702230692, + 0.003627859754487872, + 0.0022861503530293703, + 0.0030621597543358803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.029850274324417114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635594829916954 + ], + "opposite_logprob_tokens": [ + -3.6513559818267822 + ], + "logprob_reward": -0.02635594829916954, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635594829916954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039157699793577194, + 0.04208854213356972, + 0.015437777154147625, + 0.018295500427484512, + 0.022455457597970963, + 0.008586131036281586, + 0.021701665595173836, + 0.7972190976142883, + 0.011703560128808022, + 0.023354604840278625 + ], + "credit_map": [ + 0.004554808139801025, + 0.004851643927395344, + 0.007337644230574369, + 0.004853885155171156, + 0.0049284049309790134, + 0.0038821727503091097, + 0.0034732986241579056, + 0.0052758436650037766, + 0.004568113014101982, + 0.0033954973332583904, + 0.00496151577681303, + 0.0028042984195053577, + 0.005388469900935888, + 0.006956000812351704, + 0.004880133084952831, + 0.009134511463344097, + 0.0017052139155566692, + 0.0019975933246314526, + 0.0023532493505626917, + 0.0011784178204834461, + 0.0019446194637566805, + 0.0016382940812036395, + 0.0020227301865816116, + 0.00259765749797225, + 0.0036575382109731436, + 0.001429519266821444, + 0.002293130150064826, + 0.0009854973759502172, + 0.0020258540753275156, + 0.004495023749768734, + 0.0009745575953274965, + 0.0024343784898519516, + 0.002982216188684106, + 0.0029915706254541874, + 0.00433606281876564, + 0.0025187195278704166, + 0.0020967796444892883, + 0.003016273258253932, + 0.0021733592730015516, + 0.002340473933145404, + 0.0011213081888854504, + 0.0011426045093685389, + 0.0013667611638084054, + 0.000760789611376822, + 0.0008582684095017612, + 0.0014256794238463044, + 0.0008791549480520189, + 0.0010315643157809973, + 0.004176759626716375, + 0.007934917695820332, + 0.002176654525101185, + 0.001092919148504734, + 0.002547957468777895, + 0.001444642897695303, + 0.0011164292227476835, + 0.0012113840784877539, + 0.06784163415431976, + 0.11090100556612015, + 0.17252486944198608, + 0.07652902603149414, + 0.14089147746562958, + 0.05271371081471443, + 0.11404063552618027, + 0.06177671253681183, + 0.001082013244740665, + 0.00134174560662359, + 0.00210781954228878, + 0.0008730508852750063, + 0.0013543282402679324, + 0.0016030642436817288, + 0.0011301198974251747, + 0.002211417071521282, + 0.0019880139734596014, + 0.0025913435965776443, + 0.0036161281168460846, + 0.002456406829878688, + 0.0030377409420907497, + 0.0032702304888516665, + 0.0023470071610063314, + 0.00404773373156786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.02635594829916954 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254150368273258 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254150368273258, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254150368273258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15943576395511627, + 0.07785354554653168, + 0.018361307680606842, + 0.01582682505249977, + 0.02373417839407921, + 0.018342429772019386, + 0.022770889103412628, + 0.6066026091575623, + 0.016498761251568794, + 0.04057370126247406 + ], + "credit_map": [ + 0.026575971394777298, + 0.013234865851700306, + 0.04984678328037262, + 0.012457842007279396, + 0.012103012762963772, + 0.025648346170783043, + 0.009124532341957092, + 0.010444401763379574, + 0.01111546065658331, + 0.0042613293044269085, + 0.02229275554418564, + 0.0041951825842261314, + 0.006913546938449144, + 0.01290774904191494, + 0.005038051400333643, + 0.011129467748105526, + 0.002066300017759204, + 0.002286241389811039, + 0.004345680121332407, + 0.0018789960304275155, + 0.0014105562586337328, + 0.002023519715294242, + 0.0017720058094710112, + 0.0025780065916478634, + 0.0032321519684046507, + 0.0011589645873755217, + 0.001985336886718869, + 0.0007642850978299975, + 0.0018005635356530547, + 0.003915165085345507, + 0.0012804226716980338, + 0.0016899340553209186, + 0.0034818798303604126, + 0.003834808710962534, + 0.004151617176830769, + 0.0022161840461194515, + 0.002533943625167012, + 0.002047184621915221, + 0.0032732828985899687, + 0.0021952746901661158, + 0.0020237197168171406, + 0.0019089942798018456, + 0.003504203399643302, + 0.001681431313045323, + 0.0016219235258176923, + 0.0029992738272994757, + 0.0017683246405795217, + 0.002834557555615902, + 0.005256928037852049, + 0.0075782062485814095, + 0.002007473958656192, + 0.0009356279042549431, + 0.0022584598045796156, + 0.0014717541635036469, + 0.0013632354093715549, + 0.0018992008408531547, + 0.051568105816841125, + 0.08470369875431061, + 0.1304004043340683, + 0.058263469487428665, + 0.1065070629119873, + 0.04152314364910126, + 0.08711676299571991, + 0.046519916504621506, + 0.0013447351520881057, + 0.002816945780068636, + 0.0022466504015028477, + 0.0017499448731541634, + 0.002454740460962057, + 0.00187988905236125, + 0.001964263152331114, + 0.0020415904000401497, + 0.0038616841193288565, + 0.004071542993187904, + 0.007098041474819183, + 0.0039413487538695335, + 0.004527335986495018, + 0.006382420193403959, + 0.004665939137339592, + 0.006025385577231646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.005254150368273258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, extreme close-up captures a hand holding a rectangular power bank on a rich, dark wood-grained wooden table. The power bank features a modern design, with a matte black, subtly textured plastic shell and smooth, rounded edges. The hand, belonging to a young adult with fair skin and closely manicured nails, curiously holds the device. The thumb first traces the cool, smooth contours of the power bank's side, before the fingers carefully rotate the object. Then, with deliberate precision, the thumb moves to a small, circular button flush with the shell. The thumb pad presses firmly, and in response, a series of four tiny, pinprick-sized blue LEDs next to the button (previously dark) illuminate one by one, casting a faint, cool glow on the skin of the thumb. The entire scene is bathed in a soft, warm light from an unseen light source on the right, creating gentle shadows and highlighting the subtle textures of the skin and device, evoking an atmosphere of quiet, focused preparation.", + "seed": 29663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_13/index_2210_seed_29663_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02513892836868763, + "vlm_logprob": -0.02513892836868763, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02513892836868763, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02177264541387558 + ], + "opposite_logprob_tokens": [ + -6.771772861480713 + ], + "logprob_reward": -0.02177264541387558, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.02177264541387558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07659176737070084, + 0.04344124719500542, + 0.03430648520588875, + 0.04608841985464096, + 0.0516124963760376, + 0.05327964201569557, + 0.16131462156772614, + 0.3478136360645294, + 0.12418301403522491, + 0.06136868894100189 + ], + "credit_map": [ + 0.011465847492218018, + 0.006232334300875664, + 0.010139135643839836, + 0.01371234841644764, + 0.007908293046057224, + 0.005579509772360325, + 0.008396634832024574, + 0.013157668523490429, + 0.006934347562491894, + 0.004288747441023588, + 0.008495420217514038, + 0.008403871208429337, + 0.00431987177580595, + 0.0027866901364177465, + 0.003951030317693949, + 0.004261269234120846, + 0.00512806186452508, + 0.0027732327580451965, + 0.005396610591560602, + 0.004115259740501642, + 0.002695453353226185, + 0.003161001019179821, + 0.006855279207229614, + 0.0041815899312496185, + 0.007388511672616005, + 0.0060348594561219215, + 0.006833244115114212, + 0.008392967283725739, + 0.003389610443264246, + 0.0025025925133377314, + 0.007400991395115852, + 0.004145645536482334, + 0.006016691215336323, + 0.007699638605117798, + 0.005204938352108002, + 0.0052483100444078445, + 0.005493731703609228, + 0.0058705443516373634, + 0.00733935134485364, + 0.00873928889632225, + 0.008413017727434635, + 0.006159152835607529, + 0.00997338816523552, + 0.005186902824789286, + 0.0039018234238028526, + 0.00431476766243577, + 0.007431265898048878, + 0.007899324409663677, + 0.0301757100969553, + 0.01532525010406971, + 0.034633174538612366, + 0.015754228457808495, + 0.01108111534267664, + 0.009494442492723465, + 0.021590212360024452, + 0.023260489106178284, + 0.06879599392414093, + 0.030848689377307892, + 0.07623416930437088, + 0.03563349321484566, + 0.020017027854919434, + 0.01587984710931778, + 0.05366503447294235, + 0.0467393584549427, + 0.021298611536622047, + 0.011460590176284313, + 0.023223116993904114, + 0.012862189672887325, + 0.006711508147418499, + 0.006171791814267635, + 0.022599350661039352, + 0.019855856895446777, + 0.011154993437230587, + 0.0062939999625086784, + 0.012150031514465809, + 0.007626018021255732, + 0.004091699607670307, + 0.002960794372484088, + 0.009504665620625019, + 0.007586484309285879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.02177264541387558 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331392839550972 + ], + "opposite_logprob_tokens": [ + -3.7733139991760254 + ], + "logprob_reward": -0.02331392839550972, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331392839550972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06470127403736115, + 0.12476285547018051, + 0.0456060953438282, + 0.07760333269834518, + 0.06032363325357437, + 0.07926958054304123, + 0.18879537284374237, + 0.14065741002559662, + 0.11919372528791428, + 0.0990867018699646 + ], + "credit_map": [ + 0.010270957835018635, + 0.005364524200558662, + 0.00793819222599268, + 0.011966842226684093, + 0.0059591541066765785, + 0.006074088159948587, + 0.009185023605823517, + 0.00794249214231968, + 0.018350308761000633, + 0.023321956396102905, + 0.01601823791861534, + 0.022941410541534424, + 0.01439037173986435, + 0.01221687812358141, + 0.009185749106109142, + 0.008337941020727158, + 0.00703035481274128, + 0.005329624749720097, + 0.007916225120425224, + 0.005338692106306553, + 0.00490092858672142, + 0.004095360171049833, + 0.006662476807832718, + 0.0043324357829988, + 0.009105835109949112, + 0.004890713840723038, + 0.009308390319347382, + 0.00676123658195138, + 0.007478201761841774, + 0.005892825312912464, + 0.016287297010421753, + 0.017878836020827293, + 0.007311414927244186, + 0.006371099501848221, + 0.009223312139511108, + 0.009107770398259163, + 0.004260190296918154, + 0.004495549947023392, + 0.011918620206415653, + 0.007635675836354494, + 0.013551481999456882, + 0.011068089865148067, + 0.01443847082555294, + 0.007383494172245264, + 0.004950518719851971, + 0.00771971745416522, + 0.012317890301346779, + 0.007839908823370934, + 0.04084666818380356, + 0.02156219445168972, + 0.04629519581794739, + 0.016117388382554054, + 0.010343947447836399, + 0.01189862098544836, + 0.021096842363476753, + 0.020634520798921585, + 0.027776816859841347, + 0.013251226395368576, + 0.030459651723504066, + 0.013986147940158844, + 0.008318261243402958, + 0.006851886864751577, + 0.0215484369546175, + 0.01846497505903244, + 0.023964373394846916, + 0.012313791550695896, + 0.025309931486845016, + 0.012943176552653313, + 0.00625925837084651, + 0.00502769323065877, + 0.01954851858317852, + 0.013826973736286163, + 0.014166098088026047, + 0.012648135423660278, + 0.01556001603603363, + 0.010875004343688488, + 0.007927462458610535, + 0.006101000588387251, + 0.014075039885938168, + 0.017733946442604065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.02331392839550972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311478137969971 + ], + "opposite_logprob_tokens": [ + -3.16811466217041 + ], + "logprob_reward": -0.04311478137969971, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311478137969971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20172269642353058, + 0.10083363950252533, + 0.03216533362865448, + 0.0687967836856842, + 0.03716909885406494, + 0.02375410497188568, + 0.11151430755853653, + 0.05788814648985863, + 0.323865681886673, + 0.042290229350328445 + ], + "credit_map": [ + 0.03620132431387901, + 0.017915336415171623, + 0.032772455364465714, + 0.030758952721953392, + 0.0175912044942379, + 0.01640799082815647, + 0.014777995645999908, + 0.0352974534034729, + 0.014966084621846676, + 0.010179723612964153, + 0.013838845305144787, + 0.01524051558226347, + 0.012394189834594727, + 0.007342853583395481, + 0.010695830918848515, + 0.01617560163140297, + 0.004940008744597435, + 0.003967201337218285, + 0.005222446285188198, + 0.003428827039897442, + 0.0035423063673079014, + 0.0031802940648049116, + 0.0036364514380693436, + 0.004247796721756458, + 0.01002315990626812, + 0.004507414065301418, + 0.01080497819930315, + 0.006499103736132383, + 0.005391236394643784, + 0.006505440454930067, + 0.010063553228974342, + 0.015001892112195492, + 0.004972679540514946, + 0.003921562805771828, + 0.005146990064531565, + 0.004207128193229437, + 0.004155091941356659, + 0.004051638767123222, + 0.005323190242052078, + 0.005390814505517483, + 0.0036053943913429976, + 0.0028979992493987083, + 0.003804992651566863, + 0.00211155554279685, + 0.0020014147739857435, + 0.002458366798236966, + 0.0032458044588565826, + 0.003628577571362257, + 0.02321377955377102, + 0.009835433214902878, + 0.02731357514858246, + 0.010322807356715202, + 0.006806406192481518, + 0.005839868448674679, + 0.013704349286854267, + 0.014478086493909359, + 0.010995801538228989, + 0.005267550703138113, + 0.011251961812376976, + 0.00650560250505805, + 0.00395794864743948, + 0.003076049732044339, + 0.009600464254617691, + 0.007232768926769495, + 0.06339015811681747, + 0.032200537621974945, + 0.0694689005613327, + 0.03375774249434471, + 0.01704237423837185, + 0.013982648961246014, + 0.05289760231971741, + 0.04112571105360985, + 0.008477865718305111, + 0.0036685424856841564, + 0.009686664678156376, + 0.004401924088597298, + 0.0029548348393291235, + 0.002104732906445861, + 0.005399088840931654, + 0.005596574861556292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.04311478137969971 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296575993299484 + ], + "opposite_logprob_tokens": [ + -3.773296594619751 + ], + "logprob_reward": -0.023296575993299484, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296575993299484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09805350005626678, + 0.031515851616859436, + 0.03100104257464409, + 0.0894281268119812, + 0.03758253529667854, + 0.04695838317275047, + 0.0912347286939621, + 0.33312085270881653, + 0.2015858292579651, + 0.03951914608478546 + ], + "credit_map": [ + 0.015940295532345772, + 0.007250943221151829, + 0.01677493005990982, + 0.013269123621284962, + 0.007686128374189138, + 0.009497475810348988, + 0.009102101437747478, + 0.018532507121562958, + 0.004641751293092966, + 0.002552666934207082, + 0.004298549611121416, + 0.004402306396514177, + 0.0038109333254396915, + 0.0027855474036186934, + 0.0030012440402060747, + 0.006022856570780277, + 0.00482914038002491, + 0.0019110525026917458, + 0.004652808420360088, + 0.00338495965115726, + 0.0031462565530091524, + 0.0035565278958529234, + 0.0047930460423231125, + 0.004727253690361977, + 0.008843721821904182, + 0.005211338866502047, + 0.008991381153464317, + 0.010090169496834278, + 0.009899728931486607, + 0.006529474165290594, + 0.027899913489818573, + 0.011962412856519222, + 0.004821919370442629, + 0.004744634963572025, + 0.005353660322725773, + 0.003470291383564472, + 0.0039021065458655357, + 0.003979725297540426, + 0.0060652149841189384, + 0.005244987551122904, + 0.008220137096941471, + 0.0053827883675694466, + 0.009042220190167427, + 0.005461763124912977, + 0.0032422763761132956, + 0.004117320291697979, + 0.005165474023669958, + 0.0063264076597988605, + 0.017604680731892586, + 0.008629932068288326, + 0.020109403878450394, + 0.008772876113653183, + 0.006252625957131386, + 0.0054413252510130405, + 0.011883525177836418, + 0.012540362775325775, + 0.06830334663391113, + 0.030283059924840927, + 0.07503236830234528, + 0.033564649522304535, + 0.01794305443763733, + 0.015560982748866081, + 0.05126994103193283, + 0.041163478046655655, + 0.03966987878084183, + 0.01904100365936756, + 0.04558701440691948, + 0.02022973634302616, + 0.01027367077767849, + 0.008517539128661156, + 0.033424027264118195, + 0.024842973798513412, + 0.006507258862257004, + 0.0035819525364786386, + 0.005888418760150671, + 0.0045295599848032, + 0.005839457269757986, + 0.0037095840089023113, + 0.003869106061756611, + 0.005593806505203247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.023296575993299484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a hand holding a rectangular power bank on a dark wood-grained table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the power bank matte black with a subtly textured plastic shell and smooth rounded edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand belong to a young adult with fair skin and closely manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the thumb press a small circular button causing four tiny blue LEDs to illuminate one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a soft warm light from the right creating gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196710661053658 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014196710661053658, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196710661053658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08825422078371048, + 0.03679945692420006, + 0.03710891306400299, + 0.05099589005112648, + 0.03826962783932686, + 0.029765427112579346, + 0.02347811870276928, + 0.41246718168258667, + 0.16736425459384918, + 0.11549685150384903 + ], + "credit_map": [ + 0.01554715633392334, + 0.007356499787420034, + 0.014783049933612347, + 0.010260247625410557, + 0.008046826347708702, + 0.007349652703851461, + 0.006715383380651474, + 0.018195409327745438, + 0.0054482934065163136, + 0.0036993755493313074, + 0.004664779640734196, + 0.005581346806138754, + 0.0033401830587536097, + 0.003532690228894353, + 0.003895833622664213, + 0.006636958569288254, + 0.005969444289803505, + 0.0026047565042972565, + 0.0048727961257100105, + 0.003540872596204281, + 0.003503483720123768, + 0.004301038105040789, + 0.006345354951918125, + 0.00597116956487298, + 0.006097929552197456, + 0.004603360779583454, + 0.00625340361148119, + 0.005934175569564104, + 0.0038948263972997665, + 0.0033069667406380177, + 0.01298272330313921, + 0.007922508753836155, + 0.005494479089975357, + 0.003447113325819373, + 0.005380828399211168, + 0.003359101479873061, + 0.00353672681376338, + 0.0033550348598510027, + 0.006497430615127087, + 0.007198919076472521, + 0.0044558607041835785, + 0.003961687441915274, + 0.0043987673707306385, + 0.003573466558009386, + 0.00252516008913517, + 0.0029943673871457577, + 0.003852671477943659, + 0.004003448411822319, + 0.0035981214605271816, + 0.0027785468846559525, + 0.003495045006275177, + 0.0029273517429828644, + 0.0020853704772889614, + 0.001796238822862506, + 0.0033423553686589003, + 0.0034550921991467476, + 0.08373697102069855, + 0.03646000102162361, + 0.09158875793218613, + 0.04175250604748726, + 0.023426536470651627, + 0.018235141411423683, + 0.06458839774131775, + 0.05267892777919769, + 0.03280336782336235, + 0.017141681164503098, + 0.0362057201564312, + 0.016939477995038033, + 0.008525293320417404, + 0.006807873956859112, + 0.027646295726299286, + 0.021294552832841873, + 0.017774295061826706, + 0.01298371609300375, + 0.02076439931988716, + 0.013262706808745861, + 0.008259887807071209, + 0.006862220820039511, + 0.016812726855278015, + 0.0187769066542387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.014196710661053658 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_14/index_tensor([399, 399])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_14/index_tensor([399, 399])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..462a5b59ded3b1ff71bc51bcaceaf83012fbab11 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_14/index_tensor([399, 399])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot shot from an extremely low angle, the camera slowly tilting upward as the helicopter ascends, captures the moment a modern multi-purpose helicopter, painted jet black, slowly ascends vertically. Its massive main rotor spins at an astonishing speed, forming a hazy, translucent disk. The resulting powerful downdraft violently impacts the open, dry sandy ground below. Instantly, a dense cloud of yellow-brown dust, mixed with fine gravel and hay, tumbles and spreads in all directions, forming a miniature sandstorm that nearly engulfs the entire ground. At the center of this chaotic vortex of dust, the helicopter's sleek metal fuselage reflects a cool luster in the glaring afternoon sun. With a steady and composed demeanor, it resolutely defies gravity, its sturdy landing gear barely lifted off the ground. Sunlight pierces the dense dust, casting dramatic beams of light. The stark contrast between yellow sand and black iron creates a grand atmosphere imbued with power, tension, and oppression.", + "seed": 26823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_14/index_399_seed_26823_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14688798785209656, + "vlm_logprob": -0.14688798785209656, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.14688798785209656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43544915318489075 + ], + "opposite_logprob_tokens": [ + -1.0604491233825684 + ], + "logprob_reward": -0.43544915318489075, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43544915318489075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06370802223682404, + 0.02427458018064499, + 0.09894458949565887, + 0.21419954299926758, + 0.0729738399386406, + 0.17750699818134308, + 0.22830674052238464, + 0.07118117064237595, + 0.027038199827075005, + 0.021866286173462868 + ], + "credit_map": [ + 0.005807056557387114, + 0.004980725236237049, + 0.009521669708192348, + 0.008597245439887047, + 0.008802621625363827, + 0.008832396008074284, + 0.008689267560839653, + 0.008477043360471725, + 0.0032693534158170223, + 0.0026595164090394974, + 0.0029953266493976116, + 0.002993459813296795, + 0.0034510460682213306, + 0.002805321244522929, + 0.00277188653126359, + 0.0033286719117313623, + 0.008207350969314575, + 0.008599384687840939, + 0.029342472553253174, + 0.006705851759761572, + 0.006061451975256205, + 0.0061838035471737385, + 0.00889815017580986, + 0.024946123361587524, + 0.025705616921186447, + 0.020709093660116196, + 0.04473906010389328, + 0.027302080765366554, + 0.03258636221289635, + 0.0193655826151371, + 0.014661880210042, + 0.029129881411790848, + 0.013498163782060146, + 0.0048629469238221645, + 0.011001174338161945, + 0.0076551372185349464, + 0.014464015141129494, + 0.006873962469398975, + 0.006807985249906778, + 0.007810457609593868, + 0.03315829858183861, + 0.021028604358434677, + 0.04650072753429413, + 0.013476463034749031, + 0.014151699841022491, + 0.01483658142387867, + 0.013005607761442661, + 0.021349024027585983, + 0.055367711931467056, + 0.03489146754145622, + 0.042803965508937836, + 0.01472797803580761, + 0.021348673850297928, + 0.021403852850198746, + 0.019239261746406555, + 0.018523838371038437, + 0.01850222423672676, + 0.005358023103326559, + 0.02004186250269413, + 0.0037798916455358267, + 0.006676225457340479, + 0.0042113251984119415, + 0.007047075778245926, + 0.005564546212553978, + 0.007637639995664358, + 0.0020650119986385107, + 0.00864173099398613, + 0.0016492338618263602, + 0.00184285594150424, + 0.0016433647833764553, + 0.0017930378671735525, + 0.0017653265967965126, + 0.0037632896564900875, + 0.003496026387438178, + 0.003063495736569166, + 0.002009286778047681, + 0.0025239167734980583, + 0.0017151663778349757, + 0.0023268440272659063, + 0.002968262182548642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.43544915318489075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133025765419006 + ], + "opposite_logprob_tokens": [ + -1.4063302278518677 + ], + "logprob_reward": -0.28133025765419006, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133025765419006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1391160637140274, + 0.052033569663763046, + 0.11233541369438171, + 0.1270751655101776, + 0.08363638818264008, + 0.32605159282684326, + 0.05886431038379669, + 0.04279811307787895, + 0.03447255864739418, + 0.023616837337613106 + ], + "credit_map": [ + 0.012531162239611149, + 0.016764620319008827, + 0.024457043036818504, + 0.014528345316648483, + 0.017934134230017662, + 0.016344906762242317, + 0.017943046987056732, + 0.018612800166010857, + 0.005967744626104832, + 0.006404457613825798, + 0.007987365126609802, + 0.0072713736444711685, + 0.00663722725585103, + 0.006140704266726971, + 0.006434560753405094, + 0.005190138705074787, + 0.014011451043188572, + 0.00849058199673891, + 0.017337292432785034, + 0.009405852295458317, + 0.01550347451120615, + 0.01708388701081276, + 0.013435262255370617, + 0.01706760935485363, + 0.020775076001882553, + 0.013273460790514946, + 0.017475249245762825, + 0.015314017422497272, + 0.012737484648823738, + 0.0133747523650527, + 0.01292063482105732, + 0.021204493939876556, + 0.017658527940511703, + 0.008673523552715778, + 0.013389560393989086, + 0.007242320571094751, + 0.0080006318166852, + 0.009353441186249256, + 0.00715107936412096, + 0.012167304754257202, + 0.10325133055448532, + 0.024992264807224274, + 0.12127040326595306, + 0.01906369812786579, + 0.014002284966409206, + 0.01427915133535862, + 0.013787294737994671, + 0.015405172482132912, + 0.015091908164322376, + 0.009817195124924183, + 0.0117751844227314, + 0.0037203608080744743, + 0.005262852646410465, + 0.004886912181973457, + 0.0040213060565292835, + 0.0042885891161859035, + 0.0045449635945260525, + 0.004786114674061537, + 0.004897190723568201, + 0.004868670832365751, + 0.0057568480260670185, + 0.005020543001592159, + 0.0069107296876609325, + 0.006013053003698587, + 0.010354624129831791, + 0.0035169944167137146, + 0.010681705549359322, + 0.0021236438769847155, + 0.0020409924909472466, + 0.0015396618982777, + 0.002285000868141651, + 0.0019299344858154655, + 0.0035909905564039946, + 0.00238794251345098, + 0.0027204807847738266, + 0.0017766552045941353, + 0.0021544555202126503, + 0.002321748062968254, + 0.002504427218809724, + 0.006160137243568897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.28133025765419006 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100131336599588 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100131336599588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100131336599588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19483479857444763, + 0.08228380978107452, + 0.14395475387573242, + 0.20136244595050812, + 0.09495825320482254, + 0.08733384311199188, + 0.054785922169685364, + 0.04638343304395676, + 0.04630130156874657, + 0.0478014312684536 + ], + "credit_map": [ + 0.020699145272374153, + 0.019027018919587135, + 0.025261014699935913, + 0.016900310292840004, + 0.02573217824101448, + 0.031627874821424484, + 0.02006748504936695, + 0.03551975637674332, + 0.010223684832453728, + 0.005991050973534584, + 0.016918223351240158, + 0.008509410545229912, + 0.007855376228690147, + 0.012940127402544022, + 0.008453503251075745, + 0.011392430402338505, + 0.015895109623670578, + 0.008246787823736668, + 0.037939801812171936, + 0.009016876108944416, + 0.00995801854878664, + 0.015642523765563965, + 0.012807291001081467, + 0.034448325634002686, + 0.028254764154553413, + 0.032126035541296005, + 0.032102201133966446, + 0.022436700761318207, + 0.03461180254817009, + 0.017238706350326538, + 0.01462752465158701, + 0.019964711740612984, + 0.01809864304959774, + 0.008915810845792294, + 0.009216777049005032, + 0.00961465947329998, + 0.01587553881108761, + 0.008590891025960445, + 0.012051976285874844, + 0.012593956664204597, + 0.016761887818574905, + 0.012530148029327393, + 0.011579613201320171, + 0.00866270437836647, + 0.008589730598032475, + 0.009308955632150173, + 0.0073185511864721775, + 0.012582248076796532, + 0.014310477301478386, + 0.008205881342291832, + 0.01166609674692154, + 0.0035128064919263124, + 0.005624319426715374, + 0.003729622345417738, + 0.004714759532362223, + 0.003021953394636512, + 0.0062532685697078705, + 0.005316078197211027, + 0.0076556578278541565, + 0.005059132352471352, + 0.005607956554740667, + 0.00495065376162529, + 0.006559980101883411, + 0.004980701953172684, + 0.010124754160642624, + 0.0053171939216554165, + 0.013151304796338081, + 0.002871762728318572, + 0.004300833214074373, + 0.003108396427705884, + 0.004128466360270977, + 0.0032985894940793514, + 0.009816384874284267, + 0.005674473475664854, + 0.007975611835718155, + 0.003993591293692589, + 0.005765725392848253, + 0.004570499528199434, + 0.004691469483077526, + 0.005313673987984657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.004100131336599588 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024874242953956127 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024874242953956127, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024874242953956127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14777401089668274, + 0.0807369127869606, + 0.09186869114637375, + 0.137719988822937, + 0.04802966117858887, + 0.11630355566740036, + 0.06065693497657776, + 0.12447076290845871, + 0.1523023098707199, + 0.04013720527291298 + ], + "credit_map": [ + 0.01842137612402439, + 0.013022826053202152, + 0.019222434610128403, + 0.024111295118927956, + 0.013044409453868866, + 0.02513858489692211, + 0.017515316605567932, + 0.01729775220155716, + 0.01037411205470562, + 0.008996632881462574, + 0.012541912496089935, + 0.01439711358398199, + 0.007195658050477505, + 0.008771933615207672, + 0.00987966638058424, + 0.008579891175031662, + 0.010259067639708519, + 0.007960399612784386, + 0.022041302174329758, + 0.0069121853448450565, + 0.00664425129070878, + 0.007232669275254011, + 0.010564551688730717, + 0.02025427110493183, + 0.017930099740624428, + 0.013785641640424728, + 0.023838337510824203, + 0.014446045272052288, + 0.01739044301211834, + 0.012888321653008461, + 0.009576053358614445, + 0.027865048497915268, + 0.008433857932686806, + 0.003858751617372036, + 0.00648519629612565, + 0.00516421627253294, + 0.007618269417434931, + 0.0057459622621536255, + 0.004678718280047178, + 0.006044688168913126, + 0.028026379644870758, + 0.012838800437748432, + 0.03175393491983414, + 0.008852154947817326, + 0.009050100110471249, + 0.009798134677112103, + 0.008748315274715424, + 0.007235731463879347, + 0.01191470492631197, + 0.009927192702889442, + 0.010737198404967785, + 0.004052756354212761, + 0.007049610838294029, + 0.00599542073905468, + 0.005323633085936308, + 0.0056564235128462315, + 0.03906743973493576, + 0.007857928052544594, + 0.04202284291386604, + 0.007760188542306423, + 0.005727923940867186, + 0.007789126597344875, + 0.0069040535017848015, + 0.00734125729650259, + 0.04021995887160301, + 0.013264232315123081, + 0.05899825692176819, + 0.008437743410468102, + 0.00838837493211031, + 0.007030744571238756, + 0.006474302615970373, + 0.009488698095083237, + 0.006331644486635923, + 0.006023049354553223, + 0.004198725800961256, + 0.003939896821975708, + 0.006511606741696596, + 0.0037879280280321836, + 0.003995809238404036, + 0.005348543170839548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.0024874242953956127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072972789406776 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011072972789406776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072972789406776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12351572513580322, + 0.0392233245074749, + 0.07974354922771454, + 0.09415197372436523, + 0.11022290587425232, + 0.12908683717250824, + 0.07598497718572617, + 0.17651039361953735, + 0.05619495362043381, + 0.1153654232621193 + ], + "credit_map": [ + 0.01654929853975773, + 0.012374518439173698, + 0.02294047176837921, + 0.010507837869226933, + 0.01597987301647663, + 0.013978565111756325, + 0.012120537459850311, + 0.019064612686634064, + 0.004237786866724491, + 0.00419738981872797, + 0.0043383389711380005, + 0.004667936824262142, + 0.0036253805737942457, + 0.005487161688506603, + 0.006923494394868612, + 0.005745835602283478, + 0.01052585244178772, + 0.008178528398275375, + 0.011170392856001854, + 0.007775372359901667, + 0.008090871386229992, + 0.010276717133820057, + 0.011514346115291119, + 0.01221147459000349, + 0.010605367831885815, + 0.010870098136365414, + 0.016148129478096962, + 0.011014062911272049, + 0.012960202060639858, + 0.012232575565576553, + 0.007110885810106993, + 0.01321064680814743, + 0.020690402016043663, + 0.010305576957762241, + 0.014880155213177204, + 0.014024198055267334, + 0.012480358593165874, + 0.014342119917273521, + 0.01155551802366972, + 0.011944570578634739, + 0.032831933349370956, + 0.010606820695102215, + 0.043786413967609406, + 0.008607756346464157, + 0.0076810684986412525, + 0.008587130345404148, + 0.007368412334471941, + 0.00961730070412159, + 0.014280618168413639, + 0.010200992226600647, + 0.014782867394387722, + 0.005980752874165773, + 0.007703578565269709, + 0.007851185277104378, + 0.00811972189694643, + 0.007065259851515293, + 0.05216414853930473, + 0.015694893896579742, + 0.05614764988422394, + 0.011414917185902596, + 0.00810401700437069, + 0.0103243887424469, + 0.012254021130502224, + 0.010406354442238808, + 0.010049065575003624, + 0.007001069374382496, + 0.01723434403538704, + 0.0037615764886140823, + 0.004183688201010227, + 0.004720922559499741, + 0.004524808377027512, + 0.00471947668120265, + 0.021810252219438553, + 0.01613641530275345, + 0.012791427783668041, + 0.013803429901599884, + 0.013112067244946957, + 0.011591237969696522, + 0.010323735885322094, + 0.015796847641468048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.011072972789406776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly tilting upward as the helicopter ascends, captures the moment a modern multi-purpose helicopter, painted jet black, slowly ascends vertically. Its massive main rotor spins at an astonishing speed, forming a hazy, translucent disk. The resulting powerful downdraft violently impacts the open, dry sandy ground below. Instantly, a dense cloud of yellow-brown dust, mixed with fine gravel and hay, tumbles and spreads in all directions, forming a miniature sandstorm that nearly engulfs the entire ground. At the center of this chaotic vortex of dust, the helicopter's sleek metal fuselage reflects a cool luster in the glaring afternoon sun. With a steady and composed demeanor, it resolutely defies gravity, its sturdy landing gear barely lifted off the ground. Sunlight pierces the dense dust, casting dramatic beams of light. The stark contrast between yellow sand and black iron creates a grand atmosphere imbued with power, tension, and oppression.", + "seed": 26824, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_14/index_399_seed_26824_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09298008908517659, + "vlm_logprob": -0.09298008908517659, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.09298008908517659, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21064144372940063 + ], + "opposite_logprob_tokens": [ + -1.7106413841247559 + ], + "logprob_reward": -0.21064144372940063, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21064144372940063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05904928222298622, + 0.045515742152929306, + 0.05083782225847244, + 0.12428268045186996, + 0.05668899044394493, + 0.07586048543453217, + 0.06685908138751984, + 0.17881035804748535, + 0.09389085322618484, + 0.24820473790168762 + ], + "credit_map": [ + 0.005645513068884611, + 0.006050106603652239, + 0.008976561017334461, + 0.007760010194033384, + 0.00811931025236845, + 0.007038736250251532, + 0.008961041457951069, + 0.006498001515865326, + 0.005149534437805414, + 0.005169328302145004, + 0.004129549954086542, + 0.003927934914827347, + 0.009048535488545895, + 0.006297013256698847, + 0.006776483729481697, + 0.005017364863306284, + 0.005517924670130014, + 0.006484136916697025, + 0.007312519010156393, + 0.00555418711155653, + 0.006832915358245373, + 0.006694586481899023, + 0.007244007661938667, + 0.005197545047849417, + 0.025251422077417374, + 0.02688571624457836, + 0.010873638093471527, + 0.009938661940395832, + 0.01358023565262556, + 0.013962332159280777, + 0.01103695947676897, + 0.012753724120557308, + 0.006696233991533518, + 0.009731621481478214, + 0.01295302715152502, + 0.004978494718670845, + 0.00475457776337862, + 0.0059418142773211, + 0.0038507890421897173, + 0.007782434578984976, + 0.011755268089473248, + 0.006212793756276369, + 0.017631247639656067, + 0.0060572451911866665, + 0.005732985213398933, + 0.008312429301440716, + 0.009309504181146622, + 0.010849007405340672, + 0.011199177242815495, + 0.006154739297926426, + 0.00937578920274973, + 0.007903136312961578, + 0.006449772976338863, + 0.009429442696273327, + 0.008366097696125507, + 0.007980928756296635, + 0.031736619770526886, + 0.011032485403120518, + 0.021062759682536125, + 0.019016357138752937, + 0.03607241064310074, + 0.016786830499768257, + 0.02037777192890644, + 0.022725116461515427, + 0.012972493655979633, + 0.007183327339589596, + 0.014023292809724808, + 0.011400901712477207, + 0.015090442262589931, + 0.009161616675555706, + 0.012818889692425728, + 0.011239881627261639, + 0.04311774671077728, + 0.03048068843781948, + 0.026872755959630013, + 0.023643486201763153, + 0.03368422016501427, + 0.03242916613817215, + 0.018242817372083664, + 0.03973384201526642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.21064144372940063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558008134365082 + ], + "opposite_logprob_tokens": [ + -1.600580096244812 + ], + "logprob_reward": -0.22558008134365082, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558008134365082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05809466913342476, + 0.10176995396614075, + 0.05307232216000557, + 0.09618619084358215, + 0.14349974691867828, + 0.12001313269138336, + 0.17011137306690216, + 0.1386955827474594, + 0.07289617508649826, + 0.045660946518182755 + ], + "credit_map": [ + 0.006018989719450474, + 0.00628367206081748, + 0.008403330110013485, + 0.006465309765189886, + 0.006467542611062527, + 0.00766369653865695, + 0.009180945344269276, + 0.007611178793013096, + 0.011651392094790936, + 0.012562518008053303, + 0.01127548236399889, + 0.009162098169326782, + 0.018521957099437714, + 0.014249174855649471, + 0.012539369985461235, + 0.011807961389422417, + 0.008376003243029118, + 0.004442362580448389, + 0.008500708267092705, + 0.006316645070910454, + 0.004715056624263525, + 0.00667976401746273, + 0.008166060782968998, + 0.005875722039490938, + 0.011901102028787136, + 0.011522582732141018, + 0.02010328136384487, + 0.008520342409610748, + 0.011775915510952473, + 0.01128168124705553, + 0.011251519434154034, + 0.009829763323068619, + 0.020418286323547363, + 0.029100915417075157, + 0.021809494122862816, + 0.012509332969784737, + 0.015773149207234383, + 0.01513007003813982, + 0.014158611185848713, + 0.014599897898733616, + 0.022653721272945404, + 0.009968127124011517, + 0.026532476767897606, + 0.00994114950299263, + 0.009837389923632145, + 0.014283054508268833, + 0.008575187996029854, + 0.018222011625766754, + 0.040353380143642426, + 0.01673831231892109, + 0.029434988275170326, + 0.015479166060686111, + 0.010487585328519344, + 0.021208129823207855, + 0.011530374176800251, + 0.024879444390535355, + 0.029751388356089592, + 0.01047550980001688, + 0.02656226046383381, + 0.01411015447229147, + 0.009143723174929619, + 0.01877499930560589, + 0.009678380563855171, + 0.020199155434966087, + 0.011569165624678135, + 0.0045036431401968, + 0.014349592849612236, + 0.00885852426290512, + 0.006063537206500769, + 0.007999546825885773, + 0.00876566395163536, + 0.010786504484713078, + 0.007125242613255978, + 0.0032406363170593977, + 0.006256044376641512, + 0.005744283087551594, + 0.006575794890522957, + 0.0053915344178676605, + 0.004835042171180248, + 0.006492367014288902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.22558008134365082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022033003624528646 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.0022033003624528646, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022033003624528646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11754455417394638, + 0.13176485896110535, + 0.08294694125652313, + 0.13916224241256714, + 0.06463167816400528, + 0.082940012216568, + 0.11569514125585556, + 0.06908537447452545, + 0.051760297268629074, + 0.1444689780473709 + ], + "credit_map": [ + 0.012028629891574383, + 0.015098060481250286, + 0.014685939066112041, + 0.012341509573161602, + 0.018899042159318924, + 0.016077794134616852, + 0.016956457868218422, + 0.011457103304564953, + 0.015410023741424084, + 0.014346405863761902, + 0.016242455691099167, + 0.015587001107633114, + 0.015566889196634293, + 0.017690351232886314, + 0.019585780799388885, + 0.017335930839180946, + 0.009896914474666119, + 0.010217213071882725, + 0.010919749736785889, + 0.009060736745595932, + 0.01098500844091177, + 0.012003390118479729, + 0.012693147175014019, + 0.00717077637091279, + 0.021641133353114128, + 0.021820716559886932, + 0.01820632629096508, + 0.013917160220444202, + 0.01148662157356739, + 0.020654840394854546, + 0.016316566616296768, + 0.015118864364922047, + 0.009433316998183727, + 0.006363289896398783, + 0.010004095733165741, + 0.006469359155744314, + 0.0061139799654483795, + 0.00793843250721693, + 0.006800621747970581, + 0.011508574709296227, + 0.011088959872722626, + 0.006626345682889223, + 0.013097288087010384, + 0.008256038650870323, + 0.008812458254396915, + 0.007704526651650667, + 0.016168899834156036, + 0.011185486800968647, + 0.024138590320944786, + 0.0058989389799535275, + 0.022429537028074265, + 0.011651729233562946, + 0.008842350915074348, + 0.01306742150336504, + 0.013865563087165356, + 0.015801001340150833, + 0.013984479010105133, + 0.006290931720286608, + 0.007473730947822332, + 0.006369404960423708, + 0.009891459718346596, + 0.006926311645656824, + 0.0073399473913013935, + 0.010809104889631271, + 0.007876000367105007, + 0.004160317126661539, + 0.007275517098605633, + 0.005618968512862921, + 0.007129631470888853, + 0.005522867664694786, + 0.00857143010944128, + 0.0056055570021271706, + 0.02239895798265934, + 0.01497735921293497, + 0.020097240805625916, + 0.014361660927534103, + 0.020704949274659157, + 0.01605820469558239, + 0.01319176610559225, + 0.022678814828395844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.0022033003624528646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031918552704155445 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031918552704155445, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031918552704155445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1344638168811798, + 0.0798586830496788, + 0.07999932765960693, + 0.1509721279144287, + 0.11147256940603256, + 0.07267443090677261, + 0.05332494154572487, + 0.16240324079990387, + 0.05240776762366295, + 0.10242313891649246 + ], + "credit_map": [ + 0.012418786995112896, + 0.019109148532152176, + 0.01394124235957861, + 0.017065219581127167, + 0.016744019463658333, + 0.015148394741117954, + 0.024815553799271584, + 0.015221428126096725, + 0.008960445411503315, + 0.010405194014310837, + 0.01185198687016964, + 0.011151975020766258, + 0.009022435173392296, + 0.010716402903199196, + 0.010796260088682175, + 0.006953976582735777, + 0.011590246111154556, + 0.009103652089834213, + 0.013154796324670315, + 0.009351514279842377, + 0.009087874554097652, + 0.011393697001039982, + 0.010288834571838379, + 0.006028702016919851, + 0.03343004360795021, + 0.038572944700717926, + 0.014165098778903484, + 0.012311616912484169, + 0.01443680003285408, + 0.01640537567436695, + 0.010425555519759655, + 0.011224672198295593, + 0.019289229065179825, + 0.009433584287762642, + 0.017545880749821663, + 0.0104571133852005, + 0.012842774391174316, + 0.015034404583275318, + 0.008526400662958622, + 0.01834317483007908, + 0.010913411155343056, + 0.005874947644770145, + 0.016724709421396255, + 0.006730237975716591, + 0.006893464829772711, + 0.008766179904341698, + 0.006989683490246534, + 0.009781784377992153, + 0.013304511085152626, + 0.004077466670423746, + 0.009773224592208862, + 0.0046430546790361404, + 0.004707805812358856, + 0.007620427757501602, + 0.0037463114131242037, + 0.005452134180814028, + 0.03188757225871086, + 0.013084033504128456, + 0.026814522221684456, + 0.01649576798081398, + 0.0207965224981308, + 0.01932195946574211, + 0.014018474146723747, + 0.01998439058661461, + 0.009883755818009377, + 0.003811754984781146, + 0.01068824715912342, + 0.005461473483592272, + 0.0035928613506257534, + 0.006266214419156313, + 0.004172514658421278, + 0.008530941791832447, + 0.019685637205839157, + 0.007965891622006893, + 0.011885784566402435, + 0.010686066001653671, + 0.0169520303606987, + 0.013587276451289654, + 0.011550662107765675, + 0.010109786875545979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.0031918552704155445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283764719963074 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283764719963074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283764719963074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08502569049596786, + 0.06113451346755028, + 0.12668482959270477, + 0.19229824841022491, + 0.05689983814954758, + 0.12217764556407928, + 0.09328968077898026, + 0.14594867825508118, + 0.05142812430858612, + 0.06511284410953522 + ], + "credit_map": [ + 0.009354215115308762, + 0.008704776875674725, + 0.010665987618267536, + 0.012144454754889011, + 0.009573979303240776, + 0.010925821028649807, + 0.013644163496792316, + 0.010012289509177208, + 0.008395642042160034, + 0.0065854196436703205, + 0.009407030418515205, + 0.007634846959263086, + 0.006626706104725599, + 0.007380209863185883, + 0.008476636372506618, + 0.00662802392616868, + 0.01634948141872883, + 0.016621358692646027, + 0.0192397590726614, + 0.013526009395718575, + 0.01541050523519516, + 0.018176591023802757, + 0.016357306391000748, + 0.01100381463766098, + 0.036476586014032364, + 0.030292432755231857, + 0.03572715446352959, + 0.016086652874946594, + 0.012948431074619293, + 0.02500448003411293, + 0.012842546217143536, + 0.02291995845735073, + 0.008005127310752869, + 0.009359373711049557, + 0.010411165654659271, + 0.0055629494599998, + 0.005213907919824123, + 0.007641772739589214, + 0.004126503132283688, + 0.006579037755727768, + 0.021451108157634735, + 0.011171029880642891, + 0.022144490852952003, + 0.011809197254478931, + 0.011608229950070381, + 0.013686943799257278, + 0.01607641391456127, + 0.014230222441256046, + 0.020967744290828705, + 0.010573478415608406, + 0.012765486724674702, + 0.006767119746655226, + 0.00817328505218029, + 0.011876608245074749, + 0.009307564236223698, + 0.012858393602073193, + 0.03297792375087738, + 0.01690642163157463, + 0.021548999473452568, + 0.013288507238030434, + 0.009921404533088207, + 0.016827845945954323, + 0.013231412507593632, + 0.021246159449219704, + 0.009649342857301235, + 0.004211810417473316, + 0.008078665472567081, + 0.004717606585472822, + 0.006869727745652199, + 0.004977172706276178, + 0.006500684190541506, + 0.00642311479896307, + 0.010777907446026802, + 0.006715134251862764, + 0.007999001070857048, + 0.007911891676485538, + 0.009236250072717667, + 0.007747416850179434, + 0.007337862625718117, + 0.007387382909655571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.023283764719963074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly tilting upward as the helicopter ascends, captures the moment a modern multi-purpose helicopter, painted jet black, slowly ascends vertically. Its massive main rotor spins at an astonishing speed, forming a hazy, translucent disk. The resulting powerful downdraft violently impacts the open, dry sandy ground below. Instantly, a dense cloud of yellow-brown dust, mixed with fine gravel and hay, tumbles and spreads in all directions, forming a miniature sandstorm that nearly engulfs the entire ground. At the center of this chaotic vortex of dust, the helicopter's sleek metal fuselage reflects a cool luster in the glaring afternoon sun. With a steady and composed demeanor, it resolutely defies gravity, its sturdy landing gear barely lifted off the ground. Sunlight pierces the dense dust, casting dramatic beams of light. The stark contrast between yellow sand and black iron creates a grand atmosphere imbued with power, tension, and oppression.", + "seed": 26825, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_14/index_399_seed_26825_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13878966183401645, + "vlm_logprob": -0.13878966183401645, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.13878966183401645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09992679953575134 + ], + "opposite_logprob_tokens": [ + -2.474926710128784 + ], + "logprob_reward": -0.09992679953575134, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09992679953575134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22510795295238495, + 0.1237235888838768, + 0.22554843127727509, + 0.06685380637645721, + 0.05069105699658394, + 0.055965326726436615, + 0.023760423064231873, + 0.14732305705547333, + 0.02386586368083954, + 0.057160504162311554 + ], + "credit_map": [ + 0.06893754750490189, + 0.01725834049284458, + 0.00961772445589304, + 0.010058719664812088, + 0.021457789465785027, + 0.07670941948890686, + 0.008637513034045696, + 0.012430891394615173, + 0.019006401300430298, + 0.01980910822749138, + 0.011054527945816517, + 0.007519491016864777, + 0.017539294436573982, + 0.02471744269132614, + 0.015306712128221989, + 0.008770602755248547, + 0.023727722465991974, + 0.02927990071475506, + 0.023638196289539337, + 0.02964824251830578, + 0.02605469338595867, + 0.025177566334605217, + 0.03779172524809837, + 0.03023037500679493, + 0.011303394101560116, + 0.008296395651996136, + 0.0067023844458162785, + 0.007894247770309448, + 0.006957824807614088, + 0.007201232016086578, + 0.008008028380572796, + 0.010490295477211475, + 0.0047221058048307896, + 0.003526385873556137, + 0.010479680262506008, + 0.0036760170478373766, + 0.004419082310050726, + 0.007836658507585526, + 0.00421030493453145, + 0.011820820160210133, + 0.007274264469742775, + 0.0031719859689474106, + 0.01570499688386917, + 0.005345967598259449, + 0.004486863035708666, + 0.007813949137926102, + 0.005041829310357571, + 0.007125467527657747, + 0.002682152669876814, + 0.0013042171485722065, + 0.00596985686570406, + 0.002474916400387883, + 0.0018940300215035677, + 0.004464040044695139, + 0.0016155685298144817, + 0.0033556411508470774, + 0.020599771291017532, + 0.011226577684283257, + 0.02051059529185295, + 0.02056523971259594, + 0.010096954181790352, + 0.02609037607908249, + 0.014646475203335285, + 0.02358705922961235, + 0.0022671029437333345, + 0.002156376140192151, + 0.002280140994116664, + 0.003971992526203394, + 0.0021686360705643892, + 0.002368176355957985, + 0.003272262867540121, + 0.005381173454225063, + 0.006537465378642082, + 0.005458259955048561, + 0.004391966387629509, + 0.00510668708011508, + 0.010680723004043102, + 0.0053688217885792255, + 0.009066235274076462, + 0.010550339706242085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.09992679953575134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761268138885498 + ], + "opposite_logprob_tokens": [ + -0.8261268138885498 + ], + "logprob_reward": -0.5761268138885498, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761268138885498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2253045290708542, + 0.07806133478879929, + 0.09366410225629807, + 0.07980126142501831, + 0.08888024091720581, + 0.07361264526844025, + 0.0336562916636467, + 0.25338858366012573, + 0.017328396439552307, + 0.05630258098244667 + ], + "credit_map": [ + 0.06126874312758446, + 0.015293790958821774, + 0.011266500689089298, + 0.012085333466529846, + 0.02267702855169773, + 0.07331313192844391, + 0.015994763001799583, + 0.013405242003500462, + 0.013115509413182735, + 0.008126756176352501, + 0.007727842777967453, + 0.006242021452635527, + 0.010635101236402988, + 0.010188639163970947, + 0.015164545737206936, + 0.0068609160371124744, + 0.014581874012947083, + 0.011073412373661995, + 0.015048202127218246, + 0.010268832556903362, + 0.010404734872281551, + 0.008887628093361855, + 0.013981195166707039, + 0.009418230503797531, + 0.013971623964607716, + 0.00912743154913187, + 0.007821626029908657, + 0.009003983810544014, + 0.010106176137924194, + 0.00883505679666996, + 0.008935683406889439, + 0.01199967972934246, + 0.015617798082530499, + 0.009029421955347061, + 0.015327583067119122, + 0.007268856279551983, + 0.008300894871354103, + 0.013001509010791779, + 0.007480863481760025, + 0.012853316031396389, + 0.010614035651087761, + 0.006269351579248905, + 0.016737909987568855, + 0.006782068870961666, + 0.007379966787993908, + 0.011082937009632587, + 0.007007320411503315, + 0.00773905823007226, + 0.004174056462943554, + 0.0018030568026006222, + 0.00984966941177845, + 0.003835956798866391, + 0.0026481756940484047, + 0.004953918978571892, + 0.002750915940850973, + 0.00364054087549448, + 0.07924117147922516, + 0.008550475351512432, + 0.013996267691254616, + 0.009151369333267212, + 0.006547922268509865, + 0.10335797816514969, + 0.010086088441312313, + 0.022457310929894447, + 0.0017923846608027816, + 0.002195275155827403, + 0.0021353098563849926, + 0.0027178607415407896, + 0.001102197216823697, + 0.001754092052578926, + 0.002227843040600419, + 0.003403434529900551, + 0.006376324221491814, + 0.005388421472162008, + 0.006262972019612789, + 0.0053511918522417545, + 0.0068961684592068195, + 0.005046313628554344, + 0.009352843277156353, + 0.011628342792391777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.5761268138885498 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003625251119956374 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.003625251119956374, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003625251119956374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34335383772850037, + 0.07616626471281052, + 0.08699221163988113, + 0.06073228642344475, + 0.09082664549350739, + 0.1609622985124588, + 0.049947816878557205, + 0.04410107433795929, + 0.034616366028785706, + 0.052301254123449326 + ], + "credit_map": [ + 0.11129130423069, + 0.023426441475749016, + 0.011209501884877682, + 0.009601764380931854, + 0.03172789514064789, + 0.12287799268960953, + 0.013885979540646076, + 0.019332952797412872, + 0.01091004442423582, + 0.013715443201363087, + 0.0068950229324400425, + 0.0073121716268360615, + 0.009978814981877804, + 0.00917203351855278, + 0.00966261513531208, + 0.00852012075483799, + 0.008113927207887173, + 0.010679922997951508, + 0.009275578893721104, + 0.01036726962774992, + 0.010624995455145836, + 0.00993864145129919, + 0.016081178560853004, + 0.011910699307918549, + 0.010188538581132889, + 0.006158909294754267, + 0.008197679184377193, + 0.006726165767759085, + 0.006635667756199837, + 0.00737330736592412, + 0.0076371366158127785, + 0.007814883254468441, + 0.009918062016367912, + 0.007337288931012154, + 0.016294363886117935, + 0.008312910795211792, + 0.009411800652742386, + 0.010975665412843227, + 0.015456418506801128, + 0.013120139017701149, + 0.016555842012166977, + 0.010731639340519905, + 0.04733450710773468, + 0.014021515846252441, + 0.010045262984931469, + 0.026288308203220367, + 0.01414578128606081, + 0.021839439868927002, + 0.010429462417960167, + 0.0027344804257154465, + 0.012432183139026165, + 0.0042306226678192616, + 0.0028299717232584953, + 0.0072548105381429195, + 0.0033947003539651632, + 0.006641587242484093, + 0.005158757790923119, + 0.004334132187068462, + 0.005529897287487984, + 0.00486194621771574, + 0.004531593061983585, + 0.006890768185257912, + 0.005449796561151743, + 0.007344180718064308, + 0.004485410638153553, + 0.0036886867601424456, + 0.003951006568968296, + 0.005143254064023495, + 0.0029542255215346813, + 0.0034331385977566242, + 0.003242603735998273, + 0.0077180382795631886, + 0.004972252529114485, + 0.004454366862773895, + 0.006398279685527086, + 0.005883235018700361, + 0.004974398296326399, + 0.004200538154691458, + 0.012302872724831104, + 0.009115305729210377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.003625251119956374 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003194232005625963 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.003194232005625963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003194232005625963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14513669908046722, + 0.15520845353603363, + 0.13844069838523865, + 0.05334082618355751, + 0.10799986869096756, + 0.08640408515930176, + 0.03238806128501892, + 0.19853447377681732, + 0.022453362122178078, + 0.06009349972009659 + ], + "credit_map": [ + 0.0329032726585865, + 0.01965649425983429, + 0.01287860982120037, + 0.010887611657381058, + 0.015633506700396538, + 0.03291819989681244, + 0.008699123747646809, + 0.011559873819351196, + 0.021946219727396965, + 0.021793052554130554, + 0.015025891363620758, + 0.017342200502753258, + 0.02751319110393524, + 0.0212971530854702, + 0.016837934032082558, + 0.013452806510031223, + 0.015838393941521645, + 0.019886814057826996, + 0.01424368005245924, + 0.021176885813474655, + 0.02001434937119484, + 0.014174483716487885, + 0.02067900076508522, + 0.01242709532380104, + 0.01093343272805214, + 0.005248312838375568, + 0.007825062610208988, + 0.006252843886613846, + 0.004022605251520872, + 0.005751076154410839, + 0.005362765863537788, + 0.007944729179143906, + 0.014859221875667572, + 0.01100984402000904, + 0.014804944396018982, + 0.010891802608966827, + 0.011334461160004139, + 0.015668045729398727, + 0.015240984968841076, + 0.014190556481480598, + 0.010045343078672886, + 0.005754782352596521, + 0.02149183675646782, + 0.008370095863938332, + 0.006104482337832451, + 0.012616287916898727, + 0.012608109973371029, + 0.009413146413862705, + 0.004140513017773628, + 0.001878289389424026, + 0.006359550170600414, + 0.0038458837661892176, + 0.0025435644201934338, + 0.005735096521675587, + 0.0029963261913508177, + 0.004888838157057762, + 0.04050304740667343, + 0.010112497955560684, + 0.032484471797943115, + 0.014101671054959297, + 0.01172538660466671, + 0.04802869260311127, + 0.01349386852234602, + 0.028084835037589073, + 0.003023460740223527, + 0.0023691621609032154, + 0.0023074564523994923, + 0.0034517126623541117, + 0.0019017101731151342, + 0.002111271023750305, + 0.0028780153952538967, + 0.004410572815686464, + 0.005215395707637072, + 0.005785858724266291, + 0.004176728427410126, + 0.007035484537482262, + 0.00610607024282217, + 0.005177158862352371, + 0.011825923807919025, + 0.014770873822271824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.003194232005625963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075212620198727 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075212620198727, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075212620198727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16269996762275696, + 0.05245599523186684, + 0.10895513743162155, + 0.05274191498756409, + 0.07403428107500076, + 0.058076243847608566, + 0.01964688114821911, + 0.3723183572292328, + 0.025392545387148857, + 0.07367873191833496 + ], + "credit_map": [ + 0.04715590178966522, + 0.013975612819194794, + 0.009637957438826561, + 0.010361963883042336, + 0.014937024563550949, + 0.04581465572118759, + 0.010552356019616127, + 0.010264506563544273, + 0.008058210834860802, + 0.005052029155194759, + 0.005559088662266731, + 0.006002178881317377, + 0.00592201529070735, + 0.008567780256271362, + 0.007083151023834944, + 0.006211541593074799, + 0.01446352619677782, + 0.01703098975121975, + 0.011275112628936768, + 0.012644699774682522, + 0.015167046338319778, + 0.011451725848019123, + 0.016514571383595467, + 0.010407467372715473, + 0.00979520846158266, + 0.00571462232619524, + 0.006908144801855087, + 0.00400597183033824, + 0.005251931492239237, + 0.005309666972607374, + 0.0043462845496833324, + 0.01141008548438549, + 0.007777463179081678, + 0.006504366640001535, + 0.015131226740777493, + 0.005890657193958759, + 0.007833289913833141, + 0.011084752157330513, + 0.006640894804149866, + 0.013171630911529064, + 0.007377789821475744, + 0.0038764493074268103, + 0.017925646156072617, + 0.0051756794564425945, + 0.003193673212081194, + 0.00970880314707756, + 0.004421365447342396, + 0.006396835204213858, + 0.0025936707388609648, + 0.0009939884766936302, + 0.00505016790702939, + 0.002144001657143235, + 0.0011031655594706535, + 0.0031637651845812798, + 0.0013162515824660659, + 0.0032818708568811417, + 0.1235930323600769, + 0.01088697463274002, + 0.018410852178931236, + 0.011449696496129036, + 0.009752717800438404, + 0.1603938788175583, + 0.01000785082578659, + 0.027823330834507942, + 0.002294880570843816, + 0.0016774390824139118, + 0.003481535939499736, + 0.0020829022396355867, + 0.0021779013331979513, + 0.002375443233177066, + 0.004196371417492628, + 0.007106072269380093, + 0.006626727059483528, + 0.006078257691115141, + 0.00834711454808712, + 0.007320323958992958, + 0.011735685169696808, + 0.0061700972728431225, + 0.012749267742037773, + 0.014651255682110786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.011075212620198727 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly tilting upward as the helicopter ascends, captures the moment a modern multi-purpose helicopter, painted jet black, slowly ascends vertically. Its massive main rotor spins at an astonishing speed, forming a hazy, translucent disk. The resulting powerful downdraft violently impacts the open, dry sandy ground below. Instantly, a dense cloud of yellow-brown dust, mixed with fine gravel and hay, tumbles and spreads in all directions, forming a miniature sandstorm that nearly engulfs the entire ground. At the center of this chaotic vortex of dust, the helicopter's sleek metal fuselage reflects a cool luster in the glaring afternoon sun. With a steady and composed demeanor, it resolutely defies gravity, its sturdy landing gear barely lifted off the ground. Sunlight pierces the dense dust, casting dramatic beams of light. The stark contrast between yellow sand and black iron creates a grand atmosphere imbued with power, tension, and oppression.", + "seed": 26826, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_14/index_399_seed_26826_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06603531055152416, + "vlm_logprob": -0.06603531055152416, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06603531055152416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1677396148443222 + ], + "opposite_logprob_tokens": [ + -1.9177396297454834 + ], + "logprob_reward": -0.1677396148443222, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1677396148443222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2263696938753128, + 0.05144229531288147, + 0.037096913903951645, + 0.057138364762067795, + 0.4488481879234314, + 0.015436342917382717, + 0.047730688005685806, + 0.07849658280611038, + 0.02528093382716179, + 0.012160006910562515 + ], + "credit_map": [ + 0.012384453788399696, + 0.07225462049245834, + 0.010232679545879364, + 0.011839481070637703, + 0.07878260314464569, + 0.01467651966959238, + 0.015341897495090961, + 0.010857447981834412, + 0.007004509214311838, + 0.006999093107879162, + 0.010591842234134674, + 0.004137615207582712, + 0.0063913483172655106, + 0.005378360860049725, + 0.006473944988101721, + 0.004465580452233553, + 0.010796521790325642, + 0.0028024560306221247, + 0.007513388525694609, + 0.0028127033729106188, + 0.00236991117708385, + 0.004930815193802118, + 0.0029785255901515484, + 0.0028925908263772726, + 0.010105971246957779, + 0.005996365565806627, + 0.006908325478434563, + 0.0061373161152005196, + 0.009019915014505386, + 0.007856900803744793, + 0.005120940040796995, + 0.005992627702653408, + 0.07734248787164688, + 0.07379046082496643, + 0.03518921136856079, + 0.043530963361263275, + 0.040362995117902756, + 0.10858482122421265, + 0.02529444359242916, + 0.04475279897451401, + 0.002152262721210718, + 0.0030237664468586445, + 0.002259442349895835, + 0.001767400885000825, + 0.001521094236522913, + 0.001692183781415224, + 0.001268825144506991, + 0.0017513673519715667, + 0.007631480693817139, + 0.008696701377630234, + 0.01016611885279417, + 0.003869280219078064, + 0.0032879160717129707, + 0.0025844399351626635, + 0.0032854340970516205, + 0.008209314197301865, + 0.010438330471515656, + 0.007046116515994072, + 0.019053606316447258, + 0.01034976914525032, + 0.00962830800563097, + 0.005643763113766909, + 0.005137261003255844, + 0.011199423111975193, + 0.005788219161331654, + 0.0014254478737711906, + 0.00900303479284048, + 0.00207145931199193, + 0.0009129924001172185, + 0.0007950767176225781, + 0.001687826355919242, + 0.0035968783777207136, + 0.0013045077212154865, + 0.0008433019393123686, + 0.0018571149557828903, + 0.001103412825614214, + 0.0010691668139770627, + 0.0011441154638305306, + 0.0016502155922353268, + 0.0031881711911410093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.1677396148443222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281316101551056 + ], + "opposite_logprob_tokens": [ + -2.017813205718994 + ], + "logprob_reward": -0.14281316101551056, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281316101551056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2352793961763382, + 0.12748810648918152, + 0.09151215851306915, + 0.034190788865089417, + 0.1710512638092041, + 0.02181314118206501, + 0.08004643768072128, + 0.08676210790872574, + 0.0655505433678627, + 0.08630602806806564 + ], + "credit_map": [ + 0.017083968967199326, + 0.06348574161529541, + 0.010713384486734867, + 0.015400406904518604, + 0.0843079686164856, + 0.015075402334332466, + 0.01505389716476202, + 0.014158629812300205, + 0.01590087078511715, + 0.020363187417387962, + 0.022636760026216507, + 0.0119757866486907, + 0.015831008553504944, + 0.015767429023981094, + 0.015546967275440693, + 0.009466093964874744, + 0.020440425723791122, + 0.012200449593365192, + 0.017452538013458252, + 0.007225531153380871, + 0.008226821199059486, + 0.010200947523117065, + 0.007280218414962292, + 0.008485226891934872, + 0.006277534645050764, + 0.0049202642403542995, + 0.004263301845639944, + 0.003321093739941716, + 0.004014464095234871, + 0.004709508735686541, + 0.0031921707559376955, + 0.003492452437058091, + 0.030774489045143127, + 0.025903403759002686, + 0.020569248124957085, + 0.013958794064819813, + 0.012815059162676334, + 0.04216369614005089, + 0.006776015739887953, + 0.018090562894940376, + 0.0027688932605087757, + 0.003237469820305705, + 0.004241493996232748, + 0.0032806929666548967, + 0.001835441216826439, + 0.0017423884710296988, + 0.0018333855550736189, + 0.002873375080525875, + 0.012069826014339924, + 0.009481808170676231, + 0.020194081589579582, + 0.006173931062221527, + 0.005796149373054504, + 0.004618364851921797, + 0.005270788911730051, + 0.01644148863852024, + 0.01270214095711708, + 0.007344601210206747, + 0.030253693461418152, + 0.00907177571207285, + 0.004968834575265646, + 0.004774772096425295, + 0.005605919286608696, + 0.012040364556014538, + 0.013061645440757275, + 0.003811826463788748, + 0.020338261500000954, + 0.006369485519826412, + 0.0036729583516716957, + 0.0033658549655228853, + 0.005867154337465763, + 0.009063351899385452, + 0.011171345598995686, + 0.00831763818860054, + 0.012609658762812614, + 0.010767925530672073, + 0.0067575303837656975, + 0.006924743764102459, + 0.009935388341546059, + 0.01982179842889309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.14281316101551056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636875845491886 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007636875845491886, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636875845491886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12985195219516754, + 0.08417169749736786, + 0.2825254499912262, + 0.1178131178021431, + 0.12551987171173096, + 0.02134334295988083, + 0.034137364476919174, + 0.042071156203746796, + 0.12812261283397675, + 0.034443438053131104 + ], + "credit_map": [ + 0.009612394496798515, + 0.03915507718920708, + 0.00564807141199708, + 0.011683488264679909, + 0.03214993700385094, + 0.00885012000799179, + 0.008439025841653347, + 0.014313838444650173, + 0.008811729960143566, + 0.016239436343312263, + 0.010093350894749165, + 0.005428141914308071, + 0.016130181029438972, + 0.012837793678045273, + 0.010006499476730824, + 0.004624573048204184, + 0.051599640399217606, + 0.03754126653075218, + 0.046453654766082764, + 0.022026274353265762, + 0.03981998190283775, + 0.032702427357435226, + 0.023568427190184593, + 0.028813792392611504, + 0.02315778099000454, + 0.013204590417444706, + 0.021635811775922775, + 0.008842932060360909, + 0.009519794024527073, + 0.023339372128248215, + 0.00864843837916851, + 0.009464397095143795, + 0.02132066711783409, + 0.019040176644921303, + 0.01253997627645731, + 0.011330421082675457, + 0.011667014099657536, + 0.029439810663461685, + 0.007617571856826544, + 0.01256424281746149, + 0.002833710517734289, + 0.0031892138067632914, + 0.002992437221109867, + 0.002321690320968628, + 0.003033222397789359, + 0.002191416220739484, + 0.0015578764723613858, + 0.0032237758859992027, + 0.00460607698187232, + 0.004579298663884401, + 0.0073399730026721954, + 0.002639046171680093, + 0.0029532096814364195, + 0.002686556661501527, + 0.002859810832887888, + 0.006473393179476261, + 0.005761693697422743, + 0.002805310068652034, + 0.014271185733377934, + 0.004550243727862835, + 0.0027102320455014706, + 0.0028422900941222906, + 0.0028922997880727053, + 0.0062379008159041405, + 0.034807682037353516, + 0.005082498304545879, + 0.05332934856414795, + 0.008360733278095722, + 0.004347741138190031, + 0.0032553982455283403, + 0.005308430641889572, + 0.01363078597933054, + 0.004241586662828922, + 0.003744577057659626, + 0.004625000990927219, + 0.004762018099427223, + 0.002420883858576417, + 0.0044004833325743675, + 0.004604917950928211, + 0.005643971264362335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.007636875845491886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736070849001408 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736070849001408, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736070849001408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10053873062133789, + 0.2272140234708786, + 0.13775181770324707, + 0.12094201147556305, + 0.16938471794128418, + 0.032138511538505554, + 0.05642063170671463, + 0.06289635598659515, + 0.0648408830165863, + 0.027872327715158463 + ], + "credit_map": [ + 0.009875727817416191, + 0.015528068877756596, + 0.011663788929581642, + 0.014489441178739071, + 0.013067648746073246, + 0.010284827090799809, + 0.011876058764755726, + 0.013753172941505909, + 0.02543146163225174, + 0.034483473747968674, + 0.04680483415722847, + 0.019239384680986404, + 0.0298775602132082, + 0.02602524869143963, + 0.025263695046305656, + 0.02008836902678013, + 0.037145163863897324, + 0.010651669465005398, + 0.024365387856960297, + 0.00953422300517559, + 0.012618964537978172, + 0.016524208709597588, + 0.011523709632456303, + 0.015388498082756996, + 0.02583738975226879, + 0.015860002487897873, + 0.02453925646841526, + 0.011825747787952423, + 0.008070164360105991, + 0.01645338535308838, + 0.009032715111970901, + 0.009323349222540855, + 0.029604502022266388, + 0.026572134345769882, + 0.013730576261878014, + 0.015882711857557297, + 0.01487720850855112, + 0.04163575917482376, + 0.009390321560204029, + 0.017691506072878838, + 0.004279000218957663, + 0.0047777630388736725, + 0.007798109203577042, + 0.0039123487658798695, + 0.0026506914291530848, + 0.002541015390306711, + 0.002547771902754903, + 0.00363180972635746, + 0.008052930235862732, + 0.008243042044341564, + 0.01255122758448124, + 0.00490484619513154, + 0.004221491981297731, + 0.003505526576191187, + 0.004709579050540924, + 0.010231989435851574, + 0.008156795985996723, + 0.004574055317789316, + 0.020183563232421875, + 0.006874468177556992, + 0.0047376141883432865, + 0.0040746466256678104, + 0.004308714065700769, + 0.009986497461795807, + 0.0071646347641944885, + 0.0059800054877996445, + 0.012085850350558758, + 0.010764400474727154, + 0.004747586324810982, + 0.006110782735049725, + 0.005698781926184893, + 0.012288843281567097, + 0.00421934062615037, + 0.0028248634189367294, + 0.003836089977994561, + 0.0034904752392321825, + 0.002794603118672967, + 0.0027039397973567247, + 0.0035693917889147997, + 0.004433623980730772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.006736070849001408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the helicopter painted jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from an extremely low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter ascend vertically while its main rotor spins rapidly forming a hazy disk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense cloud of yellow-brown dust mixed with gravel and hay spreading on the dry sandy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do beams of sunlight pierce the dust to create a stark contrast between the yellow sand and black helicopter?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250830203294754 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005250830203294754, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250830203294754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2444090098142624, + 0.029331395402550697, + 0.11845728754997253, + 0.045474182814359665, + 0.3991445004940033, + 0.05326753482222557, + 0.026639586314558983, + 0.02840844914317131, + 0.014080071821808815, + 0.0407879613339901 + ], + "credit_map": [ + 0.016589179635047913, + 0.06098673865199089, + 0.013700311072170734, + 0.016902783885598183, + 0.08181340247392654, + 0.015916945412755013, + 0.020503299310803413, + 0.01799635775387287, + 0.0036793933250010014, + 0.0038666059263050556, + 0.005620297510176897, + 0.0036892222706228495, + 0.0027769734151661396, + 0.002739734947681427, + 0.00412273732945323, + 0.0028364311438053846, + 0.02027874067425728, + 0.015430818311870098, + 0.01937643624842167, + 0.009833530522882938, + 0.01573135517537594, + 0.012445515021681786, + 0.014442593790590763, + 0.010918299667537212, + 0.007903678342700005, + 0.004401651211082935, + 0.008517840877175331, + 0.0038214693777263165, + 0.004855947103351355, + 0.008398119360208511, + 0.0035471427254378796, + 0.004028336610645056, + 0.07009627670049667, + 0.06956061720848083, + 0.0289236418902874, + 0.037825267761945724, + 0.03469850867986679, + 0.0962991863489151, + 0.021047918125987053, + 0.040693122893571854, + 0.006497214548289776, + 0.006457215175032616, + 0.012980814091861248, + 0.007601842749863863, + 0.0036210305988788605, + 0.004074270371347666, + 0.0034519352484494448, + 0.00858321599662304, + 0.0031460903119295835, + 0.0026502441614866257, + 0.005755397025495768, + 0.002775338478386402, + 0.002747707534581423, + 0.0020263169426470995, + 0.0024787927977740765, + 0.0050596995279192924, + 0.0037664964329451323, + 0.002728155581280589, + 0.008407768793404102, + 0.0025051170960068703, + 0.002873799065127969, + 0.0020128267351537943, + 0.0016788053326308727, + 0.0044354829005897045, + 0.0019130130531266332, + 0.0013429487589746714, + 0.0029981539119035006, + 0.001894359476864338, + 0.0014346351381391287, + 0.0008950440678745508, + 0.001459617749787867, + 0.0021423001307994127, + 0.005577233154326677, + 0.004066978581249714, + 0.005407317075878382, + 0.004188200458884239, + 0.003407126758247614, + 0.0038890251889824867, + 0.00588710093870759, + 0.008364983834326267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.005250830203294754 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_15/index_tensor([6547, 6547])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_15/index_tensor([6547, 6547])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7b66b62294344e34dd805fed997dcb4248d938af --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_15/index_tensor([6547, 6547])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, static, wide-angle shot immerses the viewer in a lush bamboo forest, rendered in the classic painterly animation style inspired by Studio Ghibli. A sudden thunderstorm erupts, and the towering jade-green stems, their smooth, segmented surfaces already saturated with moisture, begin to sway violently in a chaotic yet graceful dance, propelled by a powerful, unseen wind. The entire forest bends and creaks, its dense canopy trembling violently. Simultaneously, a torrential downpour begins, plump, almost jewel-like drops cascading down in thick, visible flakes, splashing across the vibrant green foliage and dark, mossy earth below. The light of the dim, overcast sky is momentarily interrupted by a soft, distant flash of lightning, which blooms silently, bathing the scene in a pale, ethereal glow. This makes the damp green pop with incredible vitality, throws the forest's complex textures into sharp, beautiful relief, and then casts them in atmospheric shadows, creating a mood both dramatic and profoundly tranquil.", + "seed": 11047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_15/index_6547_seed_11047_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3578782869502902, + "vlm_logprob": -0.3578782869502902, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.3578782869502902, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008458266966044903 + ], + "opposite_logprob_tokens": [ + -5.008458137512207 + ], + "logprob_reward": -0.008458266966044903, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008458266966044903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12805022299289703, + 0.10843399912118912, + 0.03622909635305405, + 0.41179659962654114, + 0.13402844965457916, + 0.023331966251134872, + 0.05371372029185295, + 0.03151383623480797, + 0.039479468017816544, + 0.03342267498373985 + ], + "credit_map": [ + 0.01881711371243, + 0.014102471061050892, + 0.016885604709386826, + 0.017207032069563866, + 0.016262134537100792, + 0.013593160547316074, + 0.017111761495471, + 0.014070953242480755, + 0.01511778961867094, + 0.008461392484605312, + 0.005435251630842686, + 0.006841999012976885, + 0.03700628504157066, + 0.00889240950345993, + 0.010951310396194458, + 0.015727566555142403, + 0.0034419691655784845, + 0.003991289529949427, + 0.0052091325633227825, + 0.004688022658228874, + 0.007420598529279232, + 0.0032536922954022884, + 0.003682957962155342, + 0.004541432019323111, + 0.05258266627788544, + 0.02650841511785984, + 0.04654831066727638, + 0.028717080131173134, + 0.021546993404626846, + 0.0219196155667305, + 0.16389434039592743, + 0.05007917061448097, + 0.01098702009767294, + 0.02800874225795269, + 0.00987959187477827, + 0.020487867295742035, + 0.01730409450829029, + 0.022156227380037308, + 0.013135308399796486, + 0.012069595977663994, + 0.0015569796087220311, + 0.0028192708268761635, + 0.0028457562439143658, + 0.00347984884865582, + 0.003018044400960207, + 0.0032699035946279764, + 0.0028201164677739143, + 0.0035220447462052107, + 0.00474495580419898, + 0.007986742071807384, + 0.004861661698669195, + 0.009196219965815544, + 0.009092841297388077, + 0.004489785525947809, + 0.007931197062134743, + 0.005410322453826666, + 0.002669047797098756, + 0.0041675204411149025, + 0.0024843651335686445, + 0.0029559091199189425, + 0.006802072748541832, + 0.0052118366584181786, + 0.0028681617695838213, + 0.004354923497885466, + 0.0033095853868871927, + 0.004484518896788359, + 0.005391908343881369, + 0.0046782298013567924, + 0.006623659748584032, + 0.005330145359039307, + 0.004348109010607004, + 0.005313310772180557, + 0.0020568594336509705, + 0.0047805216163396835, + 0.004402465187013149, + 0.004601533990353346, + 0.0054947203025221825, + 0.004487140569835901, + 0.004078453406691551, + 0.0035209807101637125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.008458266966044903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4062645435333252 + ], + "opposite_logprob_tokens": [ + -0.2812644839286804 + ], + "logprob_reward": -1.4062645435333252, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4062645435333252, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3987221419811249, + 0.08900050073862076, + 0.04219422861933708, + 0.07953299582004547, + 0.13224124908447266, + 0.05748048797249794, + 0.06058147922158241, + 0.053470104932785034, + 0.0442306287586689, + 0.042546287178993225 + ], + "credit_map": [ + 0.045046087354421616, + 0.04450720548629761, + 0.041961122304201126, + 0.05016122758388519, + 0.032640304416418076, + 0.04537041112780571, + 0.05964651703834534, + 0.07938921451568604, + 0.006602671928703785, + 0.008168463595211506, + 0.007027830928564072, + 0.00848577544093132, + 0.006509129889309406, + 0.0064773051999509335, + 0.023411616683006287, + 0.02231769636273384, + 0.004285155329853296, + 0.005458878818899393, + 0.00786335114389658, + 0.00597635842859745, + 0.0030925008468329906, + 0.004320042673498392, + 0.004582993220537901, + 0.006614943966269493, + 0.008886845782399178, + 0.006945554632693529, + 0.011398687027394772, + 0.007783220615237951, + 0.008654025383293629, + 0.007153322920203209, + 0.01857626438140869, + 0.010135067626833916, + 0.011673050932586193, + 0.028819410130381584, + 0.017425961792469025, + 0.020087091252207756, + 0.016869837418198586, + 0.012842796742916107, + 0.014337356202304363, + 0.010185723192989826, + 0.0043146382085978985, + 0.0071074627339839935, + 0.006851609796285629, + 0.008928664959967136, + 0.006063329055905342, + 0.005479814950376749, + 0.010634027421474457, + 0.008100934326648712, + 0.006280793808400631, + 0.009212502278387547, + 0.004932265728712082, + 0.007405356038361788, + 0.01037845853716135, + 0.0064772809855639935, + 0.009598548524081707, + 0.006296262610703707, + 0.0042754970490932465, + 0.006230955012142658, + 0.004070737399160862, + 0.00506027415394783, + 0.011525693349540234, + 0.010224001482129097, + 0.005683600436896086, + 0.006399338599294424, + 0.003161825705319643, + 0.004147062543779612, + 0.00715935043990612, + 0.005588842555880547, + 0.005503928288817406, + 0.005638145841658115, + 0.005591543857008219, + 0.007439924404025078, + 0.003669935278594494, + 0.0065809073857963085, + 0.0051967790350317955, + 0.005689449142664671, + 0.005550947971642017, + 0.0052735330536961555, + 0.006459602154791355, + 0.004125125706195831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -1.4062645435333252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561927139759064 + ], + "opposite_logprob_tokens": [ + -1.6006193161010742 + ], + "logprob_reward": -0.22561927139759064, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561927139759064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12385871261358261, + 0.05419750511646271, + 0.023443564772605896, + 0.4756655991077423, + 0.12645618617534637, + 0.10424288362264633, + 0.02319369465112686, + 0.02030252479016781, + 0.021754443645477295, + 0.026884829625487328 + ], + "credit_map": [ + 0.00976338516920805, + 0.009616359136998653, + 0.012635709717869759, + 0.013568147085607052, + 0.009720075875520706, + 0.016632089391350746, + 0.020290926098823547, + 0.03163202479481697, + 0.010201649740338326, + 0.0030172481201589108, + 0.004718000069260597, + 0.003803862724453211, + 0.008716371841728687, + 0.0025097860489040613, + 0.009157217107713223, + 0.012073369696736336, + 0.002626431407406926, + 0.002096733544021845, + 0.004313422832638025, + 0.003839522134512663, + 0.0022771370131522417, + 0.0018506437772884965, + 0.004148590844124556, + 0.0022910828702151775, + 0.05699525773525238, + 0.03350606560707092, + 0.05260414630174637, + 0.033393580466508865, + 0.0285330917686224, + 0.02535848505795002, + 0.18405814468860626, + 0.06121683865785599, + 0.013227016665041447, + 0.017894679680466652, + 0.011585814878344536, + 0.016603510826826096, + 0.02340552769601345, + 0.020230989903211594, + 0.014500109478831291, + 0.00900853332132101, + 0.008193228393793106, + 0.016300614923238754, + 0.011502404697239399, + 0.020231764763593674, + 0.009252110496163368, + 0.01710299216210842, + 0.009457007981836796, + 0.012202766723930836, + 0.0015760499518364668, + 0.003828266868367791, + 0.002338072983548045, + 0.004917670972645283, + 0.0025606609415262938, + 0.0020566496532410383, + 0.004088062793016434, + 0.0018282600212842226, + 0.0013335958356037736, + 0.00217517651617527, + 0.002936777425929904, + 0.002392557682469487, + 0.0031930303666740656, + 0.002373205963522196, + 0.002908125286921859, + 0.002990058157593012, + 0.0015088203363120556, + 0.0025505332741886377, + 0.004152390640228987, + 0.003121674060821533, + 0.0022666756995022297, + 0.0018029228085651994, + 0.002596418373286724, + 0.0037550085689872503, + 0.0017744764918461442, + 0.004826794844120741, + 0.003563235979527235, + 0.0035362616181373596, + 0.003360432805493474, + 0.0032344392966479063, + 0.0036513060331344604, + 0.002937883138656616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.22561927139759064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003616452217102 + ], + "opposite_logprob_tokens": [ + -2.3503615856170654 + ], + "logprob_reward": -0.1003616452217102, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003616452217102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1107269898056984, + 0.07307063788175583, + 0.10641350597143173, + 0.3001008927822113, + 0.2344328910112381, + 0.04999274015426636, + 0.028142694383859634, + 0.021196026355028152, + 0.020775917917490005, + 0.05514770746231079 + ], + "credit_map": [ + 0.01329068373888731, + 0.0070072682574391365, + 0.011067875660955906, + 0.012041223235428333, + 0.006345799658447504, + 0.012369642034173012, + 0.019992733374238014, + 0.028611764311790466, + 0.0077553801238536835, + 0.007959526963531971, + 0.006192976608872414, + 0.007551613729447126, + 0.016289152204990387, + 0.006110303569585085, + 0.013595002703368664, + 0.0076166847720742226, + 0.009475341066718102, + 0.009555568918585777, + 0.023881379514932632, + 0.012038894928991795, + 0.015809347853064537, + 0.008476270362734795, + 0.012502980418503284, + 0.014673725701868534, + 0.03930065408349037, + 0.017545344308018684, + 0.03415432944893837, + 0.02253059856593609, + 0.02175237238407135, + 0.014233577996492386, + 0.11473902314901352, + 0.035844989120960236, + 0.024506958201527596, + 0.020099176093935966, + 0.025533625856041908, + 0.024660976603627205, + 0.04630440101027489, + 0.04918617755174637, + 0.02471972443163395, + 0.019421858713030815, + 0.0038303520996123552, + 0.008271932601928711, + 0.005118298809975386, + 0.008991463109850883, + 0.005536552518606186, + 0.00872744433581829, + 0.004320158623158932, + 0.005196537356823683, + 0.002357218647375703, + 0.0038243690505623817, + 0.002581258537247777, + 0.0038867804687470198, + 0.004278591368347406, + 0.0031558277551084757, + 0.004876305349171162, + 0.003182344138622284, + 0.0014969423646107316, + 0.002744109369814396, + 0.0023404851090162992, + 0.0024780654348433018, + 0.0031582056544721127, + 0.003739936975762248, + 0.002465498633682728, + 0.0027727812994271517, + 0.00144993606954813, + 0.0028880340978503227, + 0.0029210939537733793, + 0.0023322291672229767, + 0.0022522658109664917, + 0.0030470273923128843, + 0.0027149859815835953, + 0.0031703461427241564, + 0.003894947236403823, + 0.009344113059341908, + 0.006476214155554771, + 0.006755651906132698, + 0.009100264869630337, + 0.007561465259641409, + 0.007109911646693945, + 0.004905141424387693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.1003616452217102 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048687707632780075 + ], + "opposite_logprob_tokens": [ + -3.048687696456909 + ], + "logprob_reward": -0.048687707632780075, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048687707632780075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14955957233905792, + 0.10017647594213486, + 0.0603298619389534, + 0.1806190460920334, + 0.22733117640018463, + 0.06664412468671799, + 0.05623084679245949, + 0.02037898823618889, + 0.09861953556537628, + 0.040110379457473755 + ], + "credit_map": [ + 0.0158651415258646, + 0.013973225839436054, + 0.013640735298395157, + 0.02453814260661602, + 0.020489176735281944, + 0.032680246978998184, + 0.012545805424451828, + 0.015827100723981857, + 0.013253352604806423, + 0.0077852183021605015, + 0.005769364535808563, + 0.007589629385620356, + 0.028124384582042694, + 0.007908739149570465, + 0.012853695079684258, + 0.016892092302441597, + 0.006657087709754705, + 0.00616462342441082, + 0.007371000479906797, + 0.010832172818481922, + 0.006463260389864445, + 0.0050036427564918995, + 0.008946954272687435, + 0.008891122415661812, + 0.019405227154493332, + 0.013014956377446651, + 0.01962454989552498, + 0.016720961779356003, + 0.012339010834693909, + 0.011444020085036755, + 0.06241185590624809, + 0.025658471509814262, + 0.019111061468720436, + 0.034774985164403915, + 0.023435015231370926, + 0.025881638750433922, + 0.030712364241480827, + 0.04526254162192345, + 0.028587130829691887, + 0.019566427916288376, + 0.004906611982733011, + 0.010886452160775661, + 0.006824263837188482, + 0.009003255516290665, + 0.00817696563899517, + 0.009560134261846542, + 0.010084995068609715, + 0.007201450411230326, + 0.006445319391787052, + 0.007334977388381958, + 0.006551109254360199, + 0.010274780914187431, + 0.006750081200152636, + 0.005857401993125677, + 0.0070908269844949245, + 0.0059263501316308975, + 0.001798149780370295, + 0.0028090542182326317, + 0.0026438389904797077, + 0.002880666870623827, + 0.0030942142475396395, + 0.0025238278321921825, + 0.002492277417331934, + 0.0021369587630033493, + 0.007984460331499577, + 0.009487641975283623, + 0.014469338580965996, + 0.012998071499168873, + 0.009686030447483063, + 0.012189148925244808, + 0.014951412566006184, + 0.01685342937707901, + 0.0025727490428835154, + 0.006141562480479479, + 0.003488647285848856, + 0.005335761699825525, + 0.007810849696397781, + 0.005529319401830435, + 0.005056297406554222, + 0.004175193142145872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.048687707632780075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, wide-angle shot immerses the viewer in a lush bamboo forest, rendered in the classic painterly animation style inspired by Studio Ghibli. A sudden thunderstorm erupts, and the towering jade-green stems, their smooth, segmented surfaces already saturated with moisture, begin to sway violently in a chaotic yet graceful dance, propelled by a powerful, unseen wind. The entire forest bends and creaks, its dense canopy trembling violently. Simultaneously, a torrential downpour begins, plump, almost jewel-like drops cascading down in thick, visible flakes, splashing across the vibrant green foliage and dark, mossy earth below. The light of the dim, overcast sky is momentarily interrupted by a soft, distant flash of lightning, which blooms silently, bathing the scene in a pale, ethereal glow. This makes the damp green pop with incredible vitality, throws the forest's complex textures into sharp, beautiful relief, and then casts them in atmospheric shadows, creating a mood both dramatic and profoundly tranquil.", + "seed": 11048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_15/index_6547_seed_11048_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3652496924623847, + "vlm_logprob": -0.3652496924623847, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.3652496924623847, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011841834522783756 + ], + "opposite_logprob_tokens": [ + -4.761841773986816 + ], + "logprob_reward": -0.011841834522783756, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011841834522783756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06877988576889038, + 0.0739041194319725, + 0.04868730902671814, + 0.2274160385131836, + 0.20955215394496918, + 0.0458499938249588, + 0.13075146079063416, + 0.07879281789064407, + 0.08882243186235428, + 0.027443761005997658 + ], + "credit_map": [ + 0.004293494857847691, + 0.010228808969259262, + 0.008005709387362003, + 0.008976086042821407, + 0.012354876846075058, + 0.01142644602805376, + 0.007346785627305508, + 0.006147684063762426, + 0.0050894636660814285, + 0.024616651237010956, + 0.0037810439243912697, + 0.0041836341843008995, + 0.009924166835844517, + 0.006084274500608444, + 0.015194653533399105, + 0.0050302306190133095, + 0.0019395409617573023, + 0.0029021771624684334, + 0.001695839106105268, + 0.0017867032438516617, + 0.016536766663193703, + 0.018308265134692192, + 0.0028848012443631887, + 0.0026332170236855745, + 0.008630586788058281, + 0.012133069336414337, + 0.005643852055072784, + 0.004918494261801243, + 0.07862778753042221, + 0.09422034025192261, + 0.012661498971283436, + 0.010580407455563545, + 0.02252809889614582, + 0.03076651133596897, + 0.016401316970586777, + 0.011591189540922642, + 0.03254014998674393, + 0.05479389801621437, + 0.017718706279993057, + 0.023212267085909843, + 0.004427646286785603, + 0.004845038056373596, + 0.007580169010907412, + 0.006548201199620962, + 0.005354895256459713, + 0.004630837123841047, + 0.006853665225207806, + 0.005609543062746525, + 0.011007792316377163, + 0.02224014513194561, + 0.009952912107110023, + 0.014029676094651222, + 0.017833048477768898, + 0.022104820236563683, + 0.014732252806425095, + 0.018850823864340782, + 0.005470743868499994, + 0.018037553876638412, + 0.005295780021697283, + 0.004537367261946201, + 0.013202289119362831, + 0.01708444207906723, + 0.008961147628724575, + 0.006203494966030121, + 0.004156009294092655, + 0.009497343562543392, + 0.003372922306880355, + 0.0046469648368656635, + 0.03096677176654339, + 0.02813074178993702, + 0.003944133874028921, + 0.004107543267309666, + 0.002376010175794363, + 0.005165846087038517, + 0.002621634630486369, + 0.0023693349212408066, + 0.003941935487091541, + 0.006190439686179161, + 0.0026146939489990473, + 0.0021638667676597834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.011841834522783756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4062875509262085 + ], + "opposite_logprob_tokens": [ + -0.2812875211238861 + ], + "logprob_reward": -1.4062875509262085, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4062875509262085, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09574279934167862, + 0.04880187660455704, + 0.1798843890428543, + 0.05053245648741722, + 0.06634628027677536, + 0.08251038938760757, + 0.04052949324250221, + 0.13202327489852905, + 0.06863758713006973, + 0.2349914163351059 + ], + "credit_map": [ + 0.007411077618598938, + 0.01152783166617155, + 0.01270362176001072, + 0.010205759666860104, + 0.012894634157419205, + 0.016762858256697655, + 0.014421364292502403, + 0.009815650060772896, + 0.0035536191426217556, + 0.008139535784721375, + 0.0067487298510968685, + 0.005581429228186607, + 0.0063567194156348705, + 0.008653474040329456, + 0.005547297652810812, + 0.004221073817461729, + 0.006513767410069704, + 0.010007714852690697, + 0.005311180837452412, + 0.005930183921009302, + 0.06213809549808502, + 0.07141699641942978, + 0.009913405403494835, + 0.008653036318719387, + 0.0034305613953620195, + 0.004419112578034401, + 0.0021156203001737595, + 0.0031228228472173214, + 0.013048062101006508, + 0.015048093162477016, + 0.0046964241191744804, + 0.004651761148124933, + 0.007327870465815067, + 0.009301125071942806, + 0.005999194458127022, + 0.005504352040588856, + 0.010269896127283573, + 0.009209390729665756, + 0.009244561195373535, + 0.009489892981946468, + 0.006528340745717287, + 0.008280892856419086, + 0.009907189756631851, + 0.015508166514337063, + 0.00891188532114029, + 0.01189875602722168, + 0.010771596804261208, + 0.010703564621508121, + 0.00310041313059628, + 0.005807833280414343, + 0.002902862848713994, + 0.004705987870693207, + 0.005741180386394262, + 0.011452308855950832, + 0.0036948812194168568, + 0.0031240293756127357, + 0.01241814624518156, + 0.02056313492357731, + 0.010884755291044712, + 0.007055504247546196, + 0.019622480496764183, + 0.038023531436920166, + 0.007440671324729919, + 0.016015052795410156, + 0.004159742034971714, + 0.01770995929837227, + 0.00436088303104043, + 0.0026202728040516376, + 0.009790650568902493, + 0.024071842432022095, + 0.0030752161983400583, + 0.0028490156400948763, + 0.01624191738665104, + 0.044436026364564896, + 0.018764568492770195, + 0.016375824809074402, + 0.037102654576301575, + 0.06134586036205292, + 0.02351052686572075, + 0.017214056104421616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -1.4062875509262085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813935875892639 + ], + "opposite_logprob_tokens": [ + -1.4063935279846191 + ], + "logprob_reward": -0.2813935875892639, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813935875892639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09556025266647339, + 0.06915649771690369, + 0.22924241423606873, + 0.08600831031799316, + 0.17847400903701782, + 0.06724774092435837, + 0.03878641501069069, + 0.0280250683426857, + 0.109864242374897, + 0.09763503074645996 + ], + "credit_map": [ + 0.0058832778595387936, + 0.017824312672019005, + 0.010002484545111656, + 0.015038227662444115, + 0.009415669366717339, + 0.01597902923822403, + 0.009753193706274033, + 0.011664054356515408, + 0.004664569161832333, + 0.015802521258592606, + 0.007132720667868853, + 0.006097991485148668, + 0.009684785269200802, + 0.010569637641310692, + 0.009265310131013393, + 0.005938962567597628, + 0.008524912409484386, + 0.012535924091935158, + 0.005764171946793795, + 0.006192934233695269, + 0.07880400866270065, + 0.09391316771507263, + 0.012508939020335674, + 0.010998353362083435, + 0.004083019215613604, + 0.005463720299303532, + 0.003164874855428934, + 0.0029434305615723133, + 0.02767137624323368, + 0.032605599611997604, + 0.0057745748199522495, + 0.0043017202988266945, + 0.019461235031485558, + 0.024783646687865257, + 0.012688239105045795, + 0.0095493383705616, + 0.029082927852869034, + 0.05018234625458717, + 0.014778646640479565, + 0.01794763281941414, + 0.004403972532600164, + 0.007548488676548004, + 0.00669664004817605, + 0.012917652726173401, + 0.007125590927898884, + 0.00995655357837677, + 0.009357229806482792, + 0.009241611696779728, + 0.004302590619772673, + 0.006370986346155405, + 0.0028864697087556124, + 0.0034628703724592924, + 0.006916749756783247, + 0.008404497988522053, + 0.003621556330472231, + 0.0028206927236169577, + 0.0024027936160564423, + 0.003173939883708954, + 0.003579177428036928, + 0.0030147165525704622, + 0.0036640337202697992, + 0.005337063688784838, + 0.0031170097645372152, + 0.0037363332230597734, + 0.006870905868709087, + 0.038644056767225266, + 0.002220665803179145, + 0.002821955131366849, + 0.010809578001499176, + 0.04313146695494652, + 0.0019267899915575981, + 0.0034388313069939613, + 0.009013214148581028, + 0.020751919597387314, + 0.009504744783043861, + 0.007950684987008572, + 0.013652313500642776, + 0.01846526749432087, + 0.010224293917417526, + 0.008072593249380589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.2813935875892639 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035409778356552 + ], + "opposite_logprob_tokens": [ + -2.3503541946411133 + ], + "logprob_reward": -0.10035409778356552, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035409778356552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1317507028579712, + 0.07958224415779114, + 0.10909955948591232, + 0.05544305592775345, + 0.14945493638515472, + 0.051682643592357635, + 0.0568392351269722, + 0.06383993476629257, + 0.21838325262069702, + 0.08392440527677536 + ], + "credit_map": [ + 0.010288044810295105, + 0.017365073785185814, + 0.01916799321770668, + 0.019862910732626915, + 0.015443304553627968, + 0.02133471705019474, + 0.018913570791482925, + 0.00937508326023817, + 0.005960950627923012, + 0.009025408886373043, + 0.00775474077090621, + 0.007051568012684584, + 0.012976380996406078, + 0.013141273520886898, + 0.016678035259246826, + 0.0069938888773322105, + 0.0042009782046079636, + 0.006149002350866795, + 0.0026712920516729355, + 0.002894382458180189, + 0.03797708824276924, + 0.04496755078434944, + 0.0053359828889369965, + 0.004903280176222324, + 0.0028274524956941605, + 0.004629253875464201, + 0.0028558142948895693, + 0.00346477422863245, + 0.013488435186445713, + 0.017614347860217094, + 0.006186055485159159, + 0.004376920871436596, + 0.010471602901816368, + 0.014567176811397076, + 0.010442397557199001, + 0.01251540333032608, + 0.03609094023704529, + 0.026942653581500053, + 0.01506626047194004, + 0.023358507081866264, + 0.0034440981689840555, + 0.004943639971315861, + 0.005967704579234123, + 0.009862039238214493, + 0.0067535205744206905, + 0.007473974954336882, + 0.006788104772567749, + 0.006449562497437, + 0.0041085826233029366, + 0.008925543166697025, + 0.004093407653272152, + 0.004898713901638985, + 0.01057529915124178, + 0.010131597518920898, + 0.006994139868766069, + 0.007111951243132353, + 0.0052659702487289906, + 0.00977002177387476, + 0.006337180733680725, + 0.004362776409834623, + 0.010020734742283821, + 0.013379156589508057, + 0.008337241597473621, + 0.006366851273924112, + 0.01302824541926384, + 0.050028230994939804, + 0.004595128819346428, + 0.005035726353526115, + 0.043019410222768784, + 0.08923644572496414, + 0.005542524624615908, + 0.00789753906428814, + 0.006472854875028133, + 0.010858808644115925, + 0.007486557587981224, + 0.007940308190882206, + 0.018517615273594856, + 0.015280533581972122, + 0.009229293093085289, + 0.008138432167470455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.10035409778356552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026371391490101814 + ], + "opposite_logprob_tokens": [ + -3.651371479034424 + ], + "logprob_reward": -0.026371391490101814, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371391490101814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17569617927074432, + 0.035466134548187256, + 0.12089727818965912, + 0.09253890067338943, + 0.10748712718486786, + 0.10162908583879471, + 0.07729192823171616, + 0.09590049088001251, + 0.1325531303882599, + 0.060539670288562775 + ], + "credit_map": [ + 0.012543057091534138, + 0.022369790822267532, + 0.020144585520029068, + 0.021213194355368614, + 0.028765179216861725, + 0.03193143010139465, + 0.021239573135972023, + 0.017489362508058548, + 0.0028581961523741484, + 0.004522822797298431, + 0.004002734553068876, + 0.004642415791749954, + 0.004069543909281492, + 0.006841449998319149, + 0.0053586214780807495, + 0.003170349635183811, + 0.004535376559942961, + 0.006621015723794699, + 0.0030518975108861923, + 0.0028965186793357134, + 0.04237304627895355, + 0.04985888674855232, + 0.00624366058036685, + 0.005316872149705887, + 0.004948956426233053, + 0.0068702176213264465, + 0.004702338017523289, + 0.004925375804305077, + 0.025212537497282028, + 0.030850796028971672, + 0.008004933595657349, + 0.00702374754473567, + 0.01420296635478735, + 0.01726553589105606, + 0.008441625162959099, + 0.0070184669457376, + 0.015911731868982315, + 0.026930924504995346, + 0.007915136404335499, + 0.009800750762224197, + 0.006409875117242336, + 0.01105379220098257, + 0.011564524844288826, + 0.01820099540054798, + 0.011855699121952057, + 0.015979550778865814, + 0.012564229778945446, + 0.014000412076711655, + 0.008229542523622513, + 0.011561590246856213, + 0.005797824822366238, + 0.0063866921700537205, + 0.016116347163915634, + 0.01635393127799034, + 0.006681974045932293, + 0.006164027843624353, + 0.008277179673314095, + 0.013227821327745914, + 0.010503523051738739, + 0.006770521402359009, + 0.014140481129288673, + 0.01942293532192707, + 0.012033337727189064, + 0.011524693109095097, + 0.008280104957520962, + 0.037872619926929474, + 0.0034513550344854593, + 0.004674593918025494, + 0.01756955124437809, + 0.05061058700084686, + 0.004765360616147518, + 0.005328943487256765, + 0.004539099056273699, + 0.01483945269137621, + 0.004165570251643658, + 0.004906168673187494, + 0.00996409635990858, + 0.015170864760875702, + 0.0033608409576117992, + 0.003593576606363058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.026371391490101814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, wide-angle shot immerses the viewer in a lush bamboo forest, rendered in the classic painterly animation style inspired by Studio Ghibli. A sudden thunderstorm erupts, and the towering jade-green stems, their smooth, segmented surfaces already saturated with moisture, begin to sway violently in a chaotic yet graceful dance, propelled by a powerful, unseen wind. The entire forest bends and creaks, its dense canopy trembling violently. Simultaneously, a torrential downpour begins, plump, almost jewel-like drops cascading down in thick, visible flakes, splashing across the vibrant green foliage and dark, mossy earth below. The light of the dim, overcast sky is momentarily interrupted by a soft, distant flash of lightning, which blooms silently, bathing the scene in a pale, ethereal glow. This makes the damp green pop with incredible vitality, throws the forest's complex textures into sharp, beautiful relief, and then casts them in atmospheric shadows, creating a mood both dramatic and profoundly tranquil.", + "seed": 11049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_15/index_6547_seed_11049_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.37256654975935816, + "vlm_logprob": -0.37256654975935816, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.37256654975935816, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006045981775969267 + ], + "opposite_logprob_tokens": [ + -5.506045818328857 + ], + "logprob_reward": -0.006045981775969267, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006045981775969267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24375122785568237, + 0.17652468383312225, + 0.10055580735206604, + 0.04407463222742081, + 0.039529815316200256, + 0.08531422168016434, + 0.10474098473787308, + 0.0456843376159668, + 0.09706053137779236, + 0.0627637505531311 + ], + "credit_map": [ + 0.01976696588099003, + 0.03505018725991249, + 0.03628313168883324, + 0.04484211280941963, + 0.02245841547846794, + 0.028974728658795357, + 0.03106178157031536, + 0.02531391941010952, + 0.014973676763474941, + 0.03498581051826477, + 0.015630943700671196, + 0.015655895695090294, + 0.014459885656833649, + 0.008714534342288971, + 0.056113351136446, + 0.01599060371518135, + 0.01197755429893732, + 0.015066488645970821, + 0.010562218725681305, + 0.014819640666246414, + 0.009155672043561935, + 0.010153388604521751, + 0.016128983348608017, + 0.012691870331764221, + 0.0030578398145735264, + 0.005274402443319559, + 0.006109274458140135, + 0.006521676201373339, + 0.005393597763031721, + 0.003401842899620533, + 0.007392406929284334, + 0.006923592183738947, + 0.003471646923571825, + 0.0036691096611320972, + 0.004675081465393305, + 0.004794673062860966, + 0.00504763238132, + 0.003620613133534789, + 0.007779729086905718, + 0.006471330299973488, + 0.009446029551327229, + 0.013416259549558163, + 0.007506689988076687, + 0.0070501165464520454, + 0.015887316316366196, + 0.013676599599421024, + 0.009436865337193012, + 0.008894357830286026, + 0.018925011157989502, + 0.009747279807925224, + 0.016079368069767952, + 0.00895904190838337, + 0.009851424023509026, + 0.012312428094446659, + 0.009781725704669952, + 0.019084712490439415, + 0.0036547407507896423, + 0.0067960359156131744, + 0.005351439584046602, + 0.005286332219839096, + 0.008136418648064137, + 0.00560757564380765, + 0.005703927017748356, + 0.005147870630025864, + 0.009537361562252045, + 0.009710598737001419, + 0.01503253635019064, + 0.010669413022696972, + 0.014764618128538132, + 0.010596233420073986, + 0.01427424605935812, + 0.012475531548261642, + 0.006924155633896589, + 0.00490693049505353, + 0.015319421887397766, + 0.008496868424117565, + 0.008628441952168941, + 0.007117490749806166, + 0.007040745113044977, + 0.004329698625952005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.006045981775969267 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.406261920928955 + ], + "opposite_logprob_tokens": [ + -0.28126198053359985 + ], + "logprob_reward": -1.406261920928955, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.406261920928955, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10805156081914902, + 0.12965086102485657, + 0.055316295474767685, + 0.03443024307489395, + 0.038716576993465424, + 0.19156917929649353, + 0.15028904378414154, + 0.06551100313663483, + 0.0685022696852684, + 0.15796296298503876 + ], + "credit_map": [ + 0.010179135948419571, + 0.01433565840125084, + 0.01640496775507927, + 0.016733651980757713, + 0.007588917855173349, + 0.012750223278999329, + 0.01826474815607071, + 0.0117942551150918, + 0.011610249988734722, + 0.038922782987356186, + 0.01772688888013363, + 0.011906412430107594, + 0.010478807613253593, + 0.01972668617963791, + 0.010310688987374306, + 0.008968345820903778, + 0.0069908094592392445, + 0.004680247511714697, + 0.006009652744978666, + 0.00958924274891615, + 0.009432081133127213, + 0.006517338566482067, + 0.0055565317161381245, + 0.006540396250784397, + 0.002664131810888648, + 0.0030604321509599686, + 0.005181091837584972, + 0.00523374555632472, + 0.0031087426468729973, + 0.003419579705223441, + 0.0052723209373652935, + 0.0064901988953351974, + 0.0025699161924421787, + 0.0030276128090918064, + 0.006627670489251614, + 0.008021441288292408, + 0.004330715164542198, + 0.0034375181421637535, + 0.005688369274139404, + 0.005013333633542061, + 0.016498452052474022, + 0.02745615877211094, + 0.016346564516425133, + 0.01879495568573475, + 0.03964194282889366, + 0.027419673278927803, + 0.0223341453820467, + 0.02307727001607418, + 0.013216162100434303, + 0.023580437526106834, + 0.016936803236603737, + 0.010474076494574547, + 0.033213019371032715, + 0.023371972143650055, + 0.012574122287333012, + 0.01692245341837406, + 0.005497333128005266, + 0.012541992589831352, + 0.005048559512943029, + 0.006954217329621315, + 0.01107054390013218, + 0.009090221486985683, + 0.006842486560344696, + 0.008465650491416454, + 0.005984670482575893, + 0.009280909784138203, + 0.010389993898570538, + 0.010639307089149952, + 0.008932692930102348, + 0.006676127202808857, + 0.007970844395458698, + 0.008627726696431637, + 0.013945242390036583, + 0.008624518290162086, + 0.04203344136476517, + 0.014393801800906658, + 0.014528332278132439, + 0.03962040692567825, + 0.013590003363788128, + 0.011227217502892017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -1.406261920928955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559529542922974 + ], + "opposite_logprob_tokens": [ + -1.600595235824585 + ], + "logprob_reward": -0.22559529542922974, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559529542922974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21147912740707397, + 0.21391865611076355, + 0.044721443206071854, + 0.059907324612140656, + 0.04182248190045357, + 0.16664282977581024, + 0.05501769855618477, + 0.06265733391046524, + 0.07942242920398712, + 0.06441070884466171 + ], + "credit_map": [ + 0.01899346336722374, + 0.03639977052807808, + 0.02544933557510376, + 0.029162632301449776, + 0.016478927806019783, + 0.03665127977728844, + 0.028002562001347542, + 0.020341133698821068, + 0.01780378632247448, + 0.017720792442560196, + 0.041660524904727936, + 0.018576491624116898, + 0.02111843228340149, + 0.04680285230278969, + 0.03263501077890396, + 0.017600756138563156, + 0.00524546904489398, + 0.004842092283070087, + 0.004387179855257273, + 0.010041404515504837, + 0.004803990479558706, + 0.003651713253930211, + 0.006791795138269663, + 0.004957797005772591, + 0.00444398308172822, + 0.0061430856585502625, + 0.009307986125349998, + 0.008756364695727825, + 0.0050063044764101505, + 0.004966662731021643, + 0.010991033166646957, + 0.010291899554431438, + 0.003123137867078185, + 0.004429264459758997, + 0.00601676432415843, + 0.010492623783648014, + 0.004192742984741926, + 0.0038448753766715527, + 0.005050244275480509, + 0.0046728262677788734, + 0.01485586166381836, + 0.024661151692271233, + 0.016717512160539627, + 0.015816999599337578, + 0.03328919783234596, + 0.023538826033473015, + 0.01603497564792633, + 0.021728286519646645, + 0.01635025441646576, + 0.004160642623901367, + 0.006643661297857761, + 0.004375622607767582, + 0.00969694834202528, + 0.004566226154565811, + 0.005624228622764349, + 0.0036001100670546293, + 0.005448315758258104, + 0.007455128710716963, + 0.010882603004574776, + 0.00789764802902937, + 0.007076905574649572, + 0.008665043860673904, + 0.008934094570577145, + 0.00629758695140481, + 0.007536053657531738, + 0.008441844955086708, + 0.012022456154227257, + 0.008799411356449127, + 0.014449302107095718, + 0.007640959694981575, + 0.008816140703856945, + 0.011716260574758053, + 0.005665204022079706, + 0.004218636080622673, + 0.01663302443921566, + 0.006422665901482105, + 0.005431273020803928, + 0.015160555951297283, + 0.005621444899588823, + 0.005257900804281235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.22559529542922974 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20162470638751984 + ], + "opposite_logprob_tokens": [ + -1.7016247510910034 + ], + "logprob_reward": -0.20162470638751984, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20162470638751984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1879335641860962, + 0.15440593659877777, + 0.05454064905643463, + 0.027351778000593185, + 0.08577100187540054, + 0.13120807707309723, + 0.04293796792626381, + 0.09427962452173233, + 0.10691943019628525, + 0.11465201526880264 + ], + "credit_map": [ + 0.014879077672958374, + 0.024961698800325394, + 0.024572670459747314, + 0.027760792523622513, + 0.01892218180000782, + 0.023279886692762375, + 0.029212553054094315, + 0.024344701319932938, + 0.012530461885035038, + 0.046208858489990234, + 0.01352692861109972, + 0.011277889832854271, + 0.009448220953345299, + 0.028663823381066322, + 0.022032704204320908, + 0.010717052966356277, + 0.004821405746042728, + 0.004882356151938438, + 0.005388146732002497, + 0.013787846080958843, + 0.005506190471351147, + 0.005392168648540974, + 0.008426024578511715, + 0.0063365125097334385, + 0.0022554942406713963, + 0.002390239154919982, + 0.004265320021659136, + 0.003652801038697362, + 0.002656318247318268, + 0.0023619034327566624, + 0.0054487609304487705, + 0.0043209390714764595, + 0.006204800680279732, + 0.007184962276369333, + 0.009558354504406452, + 0.01448417454957962, + 0.016939247027039528, + 0.007228202652186155, + 0.010642261244356632, + 0.013529003597795963, + 0.013487141579389572, + 0.02926127053797245, + 0.010591216385364532, + 0.010122224688529968, + 0.022801076993346214, + 0.0152689004316926, + 0.010649959556758404, + 0.01902627944946289, + 0.005193295888602734, + 0.00583950849249959, + 0.00396790960803628, + 0.003895640606060624, + 0.008522991091012955, + 0.0069533465430140495, + 0.005050417501479387, + 0.0035148565657436848, + 0.006074697710573673, + 0.015614000149071217, + 0.008198648691177368, + 0.009933157823979855, + 0.020293250679969788, + 0.012928382493555546, + 0.009442168287932873, + 0.011795311234891415, + 0.010003604926168919, + 0.014186380431056023, + 0.012901317328214645, + 0.011629637330770493, + 0.01724385656416416, + 0.012568502686917782, + 0.010610515251755714, + 0.01777561567723751, + 0.010785359889268875, + 0.0075459168292582035, + 0.030219562351703644, + 0.010780608281493187, + 0.009520906023681164, + 0.02839677594602108, + 0.008094796910881996, + 0.009308087639510632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.20162470638751984 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330484427511692 + ], + "opposite_logprob_tokens": [ + -3.7733049392700195 + ], + "logprob_reward": -0.02330484427511692, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330484427511692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15618759393692017, + 0.10695606470108032, + 0.046849604696035385, + 0.01890108734369278, + 0.05579258129000664, + 0.05713316798210144, + 0.0980866476893425, + 0.07727137207984924, + 0.3464723527431488, + 0.036349520087242126 + ], + "credit_map": [ + 0.012071636505424976, + 0.02461034059524536, + 0.017741894349455833, + 0.02188996784389019, + 0.015731053426861763, + 0.023753978312015533, + 0.022203512489795685, + 0.018185224384069443, + 0.009122334420681, + 0.010924643836915493, + 0.009068969637155533, + 0.008820679970085621, + 0.00902070477604866, + 0.01980859972536564, + 0.031240055337548256, + 0.008950072340667248, + 0.006168828811496496, + 0.007179051637649536, + 0.005295213777571917, + 0.0072154877707362175, + 0.0040306346490979195, + 0.005486736074090004, + 0.006874742452055216, + 0.004598909988999367, + 0.0017772988649085164, + 0.0016382618341594934, + 0.002705885097384453, + 0.0027955726254731417, + 0.001869523199275136, + 0.001811710768379271, + 0.0033537279814481735, + 0.0029491076711565256, + 0.005086823832243681, + 0.005181765183806419, + 0.0096653513610363, + 0.007272524293512106, + 0.007983730174601078, + 0.004446911625564098, + 0.007849827408790588, + 0.008305642753839493, + 0.004978360142558813, + 0.008479543961584568, + 0.008926517330110073, + 0.005783754866570234, + 0.008778291754424572, + 0.0063223340548574924, + 0.007417944725602865, + 0.00644641974940896, + 0.026075640693306923, + 0.01380062010139227, + 0.009355967864394188, + 0.006906136404722929, + 0.008661511354148388, + 0.007802742067724466, + 0.007627802435308695, + 0.017856229096651077, + 0.005839130375534296, + 0.011704766191542149, + 0.007887283340096474, + 0.0075886729173362255, + 0.01349648367613554, + 0.011102735996246338, + 0.00823427177965641, + 0.011418030597269535, + 0.0337635762989521, + 0.03707941994071007, + 0.04292049631476402, + 0.04142308607697487, + 0.04817581921815872, + 0.039568960666656494, + 0.039369914680719376, + 0.06417109072208405, + 0.0033148296643048525, + 0.0032361310441046953, + 0.006211999338120222, + 0.004314860794693232, + 0.005329491570591927, + 0.007043967954814434, + 0.003438734682276845, + 0.003459505969658494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.02330484427511692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, wide-angle shot immerses the viewer in a lush bamboo forest, rendered in the classic painterly animation style inspired by Studio Ghibli. A sudden thunderstorm erupts, and the towering jade-green stems, their smooth, segmented surfaces already saturated with moisture, begin to sway violently in a chaotic yet graceful dance, propelled by a powerful, unseen wind. The entire forest bends and creaks, its dense canopy trembling violently. Simultaneously, a torrential downpour begins, plump, almost jewel-like drops cascading down in thick, visible flakes, splashing across the vibrant green foliage and dark, mossy earth below. The light of the dim, overcast sky is momentarily interrupted by a soft, distant flash of lightning, which blooms silently, bathing the scene in a pale, ethereal glow. This makes the damp green pop with incredible vitality, throws the forest's complex textures into sharp, beautiful relief, and then casts them in atmospheric shadows, creating a mood both dramatic and profoundly tranquil.", + "seed": 11050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_15/index_6547_seed_11050_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3456131400540471, + "vlm_logprob": -0.3456131400540471, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.3456131400540471, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005567283369600773 + ], + "opposite_logprob_tokens": [ + -5.6305670738220215 + ], + "logprob_reward": -0.005567283369600773, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.005567283369600773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1325153112411499, + 0.21254555881023407, + 0.1097690686583519, + 0.1800827831029892, + 0.06264187395572662, + 0.06965501606464386, + 0.06239280477166176, + 0.05151628330349922, + 0.04324319213628769, + 0.07563813030719757 + ], + "credit_map": [ + 0.011223521083593369, + 0.013083859346807003, + 0.02611367218196392, + 0.016856973990797997, + 0.013121988624334335, + 0.018267270177602768, + 0.01568465679883957, + 0.018163375556468964, + 0.010665975511074066, + 0.050991516560316086, + 0.024872567504644394, + 0.014874228276312351, + 0.05569470301270485, + 0.016179034486413002, + 0.017537398263812065, + 0.021730145439505577, + 0.01023117545992136, + 0.00980261992663145, + 0.036229055374860764, + 0.013614132069051266, + 0.008335614576935768, + 0.008660191670060158, + 0.014465326443314552, + 0.008430948480963707, + 0.0091421352699399, + 0.009701324626803398, + 0.07316667586565018, + 0.05364377796649933, + 0.006286194548010826, + 0.006376547273248434, + 0.013344612903892994, + 0.00842151790857315, + 0.003746939357370138, + 0.0119658587500453, + 0.00969994068145752, + 0.006221930496394634, + 0.004833501297980547, + 0.0049371058121323586, + 0.008312254212796688, + 0.01292433775961399, + 0.004931891802698374, + 0.003488270565867424, + 0.015388598665595055, + 0.010439892299473286, + 0.004001614172011614, + 0.00470888614654541, + 0.011735928244888783, + 0.014959936030209064, + 0.003834665985777974, + 0.0035559353418648243, + 0.013723710551857948, + 0.011192448437213898, + 0.006850605830550194, + 0.005018732976168394, + 0.01203832682222128, + 0.006178379524499178, + 0.003441763808950782, + 0.009810070507228374, + 0.004842765163630247, + 0.007136252708733082, + 0.00776727544143796, + 0.006006601266562939, + 0.007450097240507603, + 0.005061456933617592, + 0.0029659236315637827, + 0.004286420997232199, + 0.009564642794430256, + 0.004201836884021759, + 0.004264524206519127, + 0.0037992645520716906, + 0.005810718983411789, + 0.008349859155714512, + 0.005980613641440868, + 0.0031253232154995203, + 0.016696084290742874, + 0.009195628575980663, + 0.00576899666339159, + 0.0036614008713513613, + 0.013381509110331535, + 0.017828576266765594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.005567283369600773 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.406264066696167 + ], + "opposite_logprob_tokens": [ + -0.2812640368938446 + ], + "logprob_reward": -1.406264066696167, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.406264066696167, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04586513340473175, + 0.6723379492759705, + 0.022599542513489723, + 0.018324745818972588, + 0.009458385407924652, + 0.05106145143508911, + 0.05793607980012894, + 0.03388552367687225, + 0.05916546657681465, + 0.029365673661231995 + ], + "credit_map": [ + 0.004652883391827345, + 0.005594570655375719, + 0.00790336262434721, + 0.004231737926602364, + 0.004576482810080051, + 0.00648789806291461, + 0.005729735363274813, + 0.006688459776341915, + 0.03312474861741066, + 0.1881847083568573, + 0.06395547837018967, + 0.03999137505888939, + 0.1780855357646942, + 0.04823332652449608, + 0.04758521541953087, + 0.07317757606506348, + 0.0012533155968412757, + 0.0018278755014762282, + 0.00512506952509284, + 0.003926290664821863, + 0.0015677455812692642, + 0.0013076248578727245, + 0.003922598902136087, + 0.003669022349640727, + 0.0014493124326691031, + 0.0015453975647687912, + 0.0024158372543752193, + 0.0039049116894602776, + 0.00130884547252208, + 0.0013572388561442494, + 0.002284042304381728, + 0.004059161059558392, + 0.0007373588741756976, + 0.0008320000488311052, + 0.001636708271689713, + 0.0013579631922766566, + 0.001013021799735725, + 0.0009719842346385121, + 0.0015638814074918628, + 0.0013454679865390062, + 0.003462951397523284, + 0.002182268537580967, + 0.017011158168315887, + 0.004306113813072443, + 0.002107761101797223, + 0.002744348719716072, + 0.003458541352301836, + 0.015788303688168526, + 0.003710981458425522, + 0.0038054699543863535, + 0.012399333529174328, + 0.009816570207476616, + 0.009268772788345814, + 0.005246518645435572, + 0.007684438489377499, + 0.006003992632031441, + 0.0026163579896092415, + 0.006621940061450005, + 0.0037818471901118755, + 0.003899396862834692, + 0.00553990388289094, + 0.003979107830673456, + 0.0037716326769441366, + 0.0036753355525434017, + 0.004631330259144306, + 0.008062050677835941, + 0.005213144700974226, + 0.006345044821500778, + 0.012423080392181873, + 0.007563658058643341, + 0.009539872407913208, + 0.005387281998991966, + 0.0018911869265139103, + 0.0018783813575282693, + 0.006306818686425686, + 0.002661429112777114, + 0.0032139476388692856, + 0.001266995444893837, + 0.003339682938531041, + 0.008807232603430748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -1.406264066696167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994312942028046 + ], + "opposite_logprob_tokens": [ + -1.8049430847167969 + ], + "logprob_reward": -0.17994312942028046, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994312942028046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08274303376674652, + 0.3888147473335266, + 0.031308043748140335, + 0.06772662699222565, + 0.02693689987063408, + 0.13326682150363922, + 0.048544853925704956, + 0.08354241400957108, + 0.03958175331354141, + 0.09753487259149551 + ], + "credit_map": [ + 0.006403666455298662, + 0.007375605404376984, + 0.013316228054463863, + 0.012675746344029903, + 0.005795720964670181, + 0.01619045063853264, + 0.010208324529230595, + 0.010777290910482407, + 0.01842878945171833, + 0.03770517557859421, + 0.07109399884939194, + 0.031128330156207085, + 0.06000760570168495, + 0.0461634024977684, + 0.039900194853544235, + 0.08438724279403687, + 0.0028077252209186554, + 0.002701477613300085, + 0.004551379010081291, + 0.006998723838478327, + 0.0035537888761609793, + 0.0023864763788878918, + 0.005214730743318796, + 0.0030937392730265856, + 0.003938707988709211, + 0.004012601915746927, + 0.02324291504919529, + 0.016282550990581512, + 0.004085192922502756, + 0.003275846829637885, + 0.004826490301638842, + 0.00806231889873743, + 0.002032204531133175, + 0.004663487896323204, + 0.0042128306813538074, + 0.003779360791668296, + 0.0015574542339891195, + 0.0019434578716754913, + 0.002355673350393772, + 0.006392429117113352, + 0.008266379125416279, + 0.0040244776755571365, + 0.045800432562828064, + 0.012729352340102196, + 0.0048579806461930275, + 0.006964649073779583, + 0.007489127106964588, + 0.043134428560733795, + 0.0033641967456787825, + 0.003649801714345813, + 0.007584395818412304, + 0.012500190176069736, + 0.00697538023814559, + 0.004209293983876705, + 0.005248554516583681, + 0.005013037472963333, + 0.005531354807317257, + 0.009556183591485023, + 0.012926309369504452, + 0.008420804515480995, + 0.014276303350925446, + 0.010617118328809738, + 0.012384779751300812, + 0.00982955563813448, + 0.002659552963450551, + 0.004541618283838034, + 0.007489078678190708, + 0.00443684309720993, + 0.0034599630162119865, + 0.004903826396912336, + 0.005593202542513609, + 0.006497666239738464, + 0.00796799547970295, + 0.005435421597212553, + 0.008585850708186626, + 0.006755280774086714, + 0.02665146254003048, + 0.0050887237302958965, + 0.009096802212297916, + 0.027953334152698517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.17994312942028046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297962069511414 + ], + "opposite_logprob_tokens": [ + -2.2379796504974365 + ], + "logprob_reward": -0.11297962069511414, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297962069511414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060213956981897354, + 0.13550695776939392, + 0.2708877623081207, + 0.020763101056218147, + 0.036026421934366226, + 0.07351105660200119, + 0.0639488473534584, + 0.18844722211360931, + 0.02763928286731243, + 0.12305540591478348 + ], + "credit_map": [ + 0.0046319784596562386, + 0.006004763767123222, + 0.00946679525077343, + 0.00603524362668395, + 0.006365107838064432, + 0.014594770036637783, + 0.00576813705265522, + 0.0073471576906740665, + 0.0062775928527116776, + 0.030919134616851807, + 0.010072733275592327, + 0.0077130114659667015, + 0.04368717595934868, + 0.013049589470028877, + 0.01364814955741167, + 0.010139573365449905, + 0.018632030114531517, + 0.024727553129196167, + 0.09009214490652084, + 0.03687087818980217, + 0.027547113597393036, + 0.016262473538517952, + 0.030794069170951843, + 0.025961481034755707, + 0.0015364981954917312, + 0.001449868781492114, + 0.004742120858281851, + 0.00339585286565125, + 0.0012110865209251642, + 0.001525397994555533, + 0.0023458998184651136, + 0.004556375555694103, + 0.0026158310938626528, + 0.0033842907287180424, + 0.006579394452273846, + 0.005915321875363588, + 0.0036016518715769053, + 0.0036504967138171196, + 0.004545785486698151, + 0.005733649246394634, + 0.0050500864163041115, + 0.003015976632013917, + 0.023187704384326935, + 0.007790931034833193, + 0.003284422680735588, + 0.004124010913074017, + 0.006097706966102123, + 0.020960219204425812, + 0.0036652155686169863, + 0.004391023889183998, + 0.010308702476322651, + 0.01919180527329445, + 0.007108825258910656, + 0.004967337474226952, + 0.007070157676935196, + 0.007245780900120735, + 0.012118488550186157, + 0.026821065694093704, + 0.019637931138277054, + 0.032369546592235565, + 0.022015094757080078, + 0.029941102489829063, + 0.02488604746758938, + 0.020657937973737717, + 0.002094326540827751, + 0.003079009708017111, + 0.003994875121861696, + 0.0029498727526515722, + 0.00441287737339735, + 0.0027635833248496056, + 0.00490662781521678, + 0.0034381109289824963, + 0.008361994288861752, + 0.007938958704471588, + 0.016988709568977356, + 0.011995241977274418, + 0.025560906156897545, + 0.007048553321510553, + 0.029353730380535126, + 0.015807317569851875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.11297962069511414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering jade-green stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems swaying violently due to a powerful wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrential downpour with thick, visible raindrops falling on the foliage and earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a soft, distant flash of lightning illuminate the scene with a pale glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a classic painterly animation style inspired by Studio Ghibli?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331160008907318 + ], + "opposite_logprob_tokens": [ + -3.7733116149902344 + ], + "logprob_reward": -0.02331160008907318, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331160008907318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15791401267051697, + 0.2213028371334076, + 0.05733109265565872, + 0.0889207050204277, + 0.06518084555864334, + 0.11294691264629364, + 0.025245286524295807, + 0.0890742689371109, + 0.07726441323757172, + 0.10481961071491241 + ], + "credit_map": [ + 0.01340810302644968, + 0.012077506631612778, + 0.02801242098212242, + 0.023482181131839752, + 0.014516415074467659, + 0.01968199759721756, + 0.02296353690326214, + 0.023771842941641808, + 0.009239796549081802, + 0.03678879141807556, + 0.033460237085819244, + 0.01331528089940548, + 0.04690485820174217, + 0.02055864781141281, + 0.01866731606423855, + 0.04236790910363197, + 0.004997731186449528, + 0.008091533556580544, + 0.01030973345041275, + 0.01073534321039915, + 0.0054612308740615845, + 0.0052396575920283794, + 0.007387091871351004, + 0.005108772311359644, + 0.004501674324274063, + 0.00479683605954051, + 0.03181768208742142, + 0.02854784019291401, + 0.0038769827224314213, + 0.003440587315708399, + 0.0073351250030100346, + 0.004603978246450424, + 0.00556802237406373, + 0.007816526107490063, + 0.011681480333209038, + 0.008746621198952198, + 0.007352098356932402, + 0.006108455825597048, + 0.008330403827130795, + 0.009577239863574505, + 0.007368027698248625, + 0.004080083686858416, + 0.038136739283800125, + 0.010324341244995594, + 0.004503028467297554, + 0.006464323960244656, + 0.008898369036614895, + 0.03317200019955635, + 0.0020553695503622293, + 0.002150767482817173, + 0.003847433952614665, + 0.005453547462821007, + 0.0031691777985543013, + 0.002497419249266386, + 0.0034165207762271166, + 0.00265505095012486, + 0.005832748021930456, + 0.014658503234386444, + 0.010965882800519466, + 0.012003470212221146, + 0.011171833612024784, + 0.008371259085834026, + 0.017213404178619385, + 0.008857171051204205, + 0.0049243951216340065, + 0.009234407916665077, + 0.013679065741598606, + 0.008675848133862019, + 0.00804173108190298, + 0.00866050273180008, + 0.010878977365791798, + 0.013169488869607449, + 0.008293953724205494, + 0.008413846604526043, + 0.011499281041324139, + 0.008686584420502186, + 0.025712791830301285, + 0.005299592390656471, + 0.006586097180843353, + 0.030327456071972847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.02331160008907318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_tensor([4177, 4177])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_tensor([4177, 4177])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8b91fad588c029f490bf85512bc1bfe85b6018d0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_tensor([4177, 4177])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, deliberate horizontal tracking shot slides from left to right, revealing a profoundly still scene within a dim, ancient stone chamber. A sharp beam of sunlight, piercing the oppressive darkness from an unseen high window, slants across the frame, illuminating the central scene. An ancient, weathered crown rests solemnly upon a plush crimson velvet cushion. The crown itself is crafted from heavy, tarnished gold, its luster long faded to a dull glow, its surface marred by tiny scratches and the occasional deep dent from passing years. Its simple yet powerful design is interrupted by the empty surroundings where the jewel once rested. The velvet cushion, rich in color but slightly frayed at the edges, shows signs of age, sinking deeply beneath the weight of the crown. Within the brilliant light, countless motes of dust dance and twirl in a slow, captivating ballet, catching the light like tiny, glittering embers. The camera's steady sideways drift causes light to slide across the crown's textured surface, creating a subtle play of highlights and shadows that emphasize its age and history, evoking an atmosphere of powerful, silent, and contemplative forgotten majesty.", + "seed": 18631, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_4177_seed_18631_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018177160108461977, + "vlm_logprob": -0.018177160108461977, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.018177160108461977, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016388336196541786 + ], + "opposite_logprob_tokens": [ + -4.266388416290283 + ], + "logprob_reward": -0.016388336196541786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016388336196541786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08721946179866791, + 0.6069703102111816, + 0.027093028649687767, + 0.03242188319563866, + 0.11249465495347977, + 0.015709005296230316, + 0.0098035903647542, + 0.013885358348488808, + 0.062010347843170166, + 0.032392386347055435 + ], + "credit_map": [ + 0.004625106230378151, + 0.005471846554428339, + 0.027464354410767555, + 0.011834864504635334, + 0.008267724886536598, + 0.01308499276638031, + 0.0070479982532560825, + 0.009422577917575836, + 0.024641867727041245, + 0.010306705720722675, + 0.25108224153518677, + 0.03960428386926651, + 0.015451261773705482, + 0.19261065125465393, + 0.03873208165168762, + 0.03454121947288513, + 0.0020707419607788324, + 0.00233958987519145, + 0.0050591351464390755, + 0.003298809751868248, + 0.002009372226893902, + 0.004276755265891552, + 0.0028167827986180782, + 0.005221843253821135, + 0.003684480907395482, + 0.0032912802416831255, + 0.0050216419622302055, + 0.004961246624588966, + 0.003046000376343727, + 0.004630305338650942, + 0.0039900075644254684, + 0.003796921344473958, + 0.008171440102159977, + 0.005871150642633438, + 0.017412930727005005, + 0.015005681663751602, + 0.007654285989701748, + 0.03295629471540451, + 0.015417799353599548, + 0.010005071759223938, + 0.0010989948641508818, + 0.0007492754375562072, + 0.0016772415256127715, + 0.0017491469625383615, + 0.0011365060927346349, + 0.0025354267563670874, + 0.002361021935939789, + 0.004401391372084618, + 0.0010935977334156632, + 0.001174028031527996, + 0.0013154339976608753, + 0.001323773874901235, + 0.000990477972663939, + 0.0015470900107175112, + 0.001379598630592227, + 0.000979590811766684, + 0.001043537980876863, + 0.0012591215781867504, + 0.0021202308125793934, + 0.0020399729255586863, + 0.0011310882400721312, + 0.001649973215535283, + 0.0026517545338720083, + 0.001989679178223014, + 0.004404712002724409, + 0.004634042736142874, + 0.011581296101212502, + 0.006587743293493986, + 0.006279938854277134, + 0.008403928950428963, + 0.011645525693893433, + 0.008473155088722706, + 0.0028885584324598312, + 0.003190228482708335, + 0.003540212055668235, + 0.004519423004239798, + 0.0035854538436979055, + 0.005046150181442499, + 0.004582665860652924, + 0.005039694719016552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.016388336196541786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016096873208880424 + ], + "opposite_logprob_tokens": [ + -4.141097068786621 + ], + "logprob_reward": -0.016096873208880424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096873208880424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38067588210105896, + 0.14059317111968994, + 0.04167383909225464, + 0.034756049513816833, + 0.0915345847606659, + 0.02740277536213398, + 0.013421749696135521, + 0.010554991662502289, + 0.1420702487230301, + 0.11731670051813126 + ], + "credit_map": [ + 0.021156538277864456, + 0.009897027164697647, + 0.15632763504981995, + 0.03284916281700134, + 0.019452422857284546, + 0.0669526681303978, + 0.031194161623716354, + 0.04284626618027687, + 0.009875723160803318, + 0.008796703070402145, + 0.028439948335289955, + 0.012494786642491817, + 0.009023432619869709, + 0.042012400925159454, + 0.014824026264250278, + 0.015126151032745838, + 0.003081388771533966, + 0.004693808499723673, + 0.007875721901655197, + 0.0049101607874035835, + 0.0031546407844871283, + 0.007580684497952461, + 0.0040087467059493065, + 0.006368685979396105, + 0.0037160604260861874, + 0.003832477144896984, + 0.005997464060783386, + 0.005319925490766764, + 0.0022421940229833126, + 0.005478013306856155, + 0.003878060495480895, + 0.004291854798793793, + 0.007647993508726358, + 0.0037058733869343996, + 0.014322358183562756, + 0.0074193524196743965, + 0.005114040337502956, + 0.032449278980493546, + 0.011326992884278297, + 0.009548692032694817, + 0.0020461627282202244, + 0.001219533383846283, + 0.003183429129421711, + 0.003280382137745619, + 0.002415822818875313, + 0.0040305922739207745, + 0.004351028241217136, + 0.006875824648886919, + 0.0015722179086878896, + 0.0012550611281767488, + 0.0018955041887238622, + 0.0017093806527554989, + 0.001120421220548451, + 0.002075108466669917, + 0.0018745154375210404, + 0.0019195419736206532, + 0.0007525653345510364, + 0.0007545198313891888, + 0.0016424991190433502, + 0.0016292702639475465, + 0.0007694429950788617, + 0.0011597886914387345, + 0.0018866988830268383, + 0.001960206311196089, + 0.01050797663629055, + 0.009614703245460987, + 0.028845906257629395, + 0.01627700962126255, + 0.011753260158002377, + 0.01922043226659298, + 0.02715645544230938, + 0.018694495782256126, + 0.01140573713928461, + 0.009531841613352299, + 0.013278682716190815, + 0.012931694276630878, + 0.010121519677340984, + 0.027163144201040268, + 0.017051884904503822, + 0.015832193195819855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.016096873208880424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822702983394265 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.002822702983394265, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822702983394265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09710072726011276, + 0.3123241662979126, + 0.030575793236494064, + 0.09570702910423279, + 0.14850465953350067, + 0.1109367087483406, + 0.02913161739706993, + 0.029431717470288277, + 0.12064554542303085, + 0.025642015039920807 + ], + "credit_map": [ + 0.0072782705537974834, + 0.007052079774439335, + 0.027407733723521233, + 0.014908606186509132, + 0.010895281098783016, + 0.012420181185007095, + 0.009198523126542568, + 0.007940055802464485, + 0.015021217055618763, + 0.006720177363604307, + 0.12880593538284302, + 0.021093063056468964, + 0.008793157525360584, + 0.08988023549318314, + 0.02149455063045025, + 0.020515860989689827, + 0.00203117192722857, + 0.001957736909389496, + 0.005198047496378422, + 0.0032404393423348665, + 0.0032220149878412485, + 0.007386503275483847, + 0.004080390091985464, + 0.003459490370005369, + 0.008360197767615318, + 0.004275218117982149, + 0.015930918976664543, + 0.00863152090460062, + 0.0068369596265256405, + 0.017531689256429672, + 0.01438287366181612, + 0.01975766383111477, + 0.009085330180823803, + 0.00976130273193121, + 0.025196794420480728, + 0.02814357355237007, + 0.008194218389689922, + 0.025193363428115845, + 0.02480599470436573, + 0.018124084919691086, + 0.007009454071521759, + 0.006755542010068893, + 0.018335195258259773, + 0.01684749312698841, + 0.01160729955881834, + 0.017696106806397438, + 0.0175761841237545, + 0.015109443105757236, + 0.003006586106494069, + 0.0024833260104060173, + 0.00435213465243578, + 0.004557698033750057, + 0.0028151869773864746, + 0.004855850245803595, + 0.0039061536081135273, + 0.003154682694002986, + 0.002901927102357149, + 0.00215195887722075, + 0.005312040913850069, + 0.00316889607347548, + 0.0023580254055559635, + 0.0036447534803301096, + 0.005132031626999378, + 0.004762089811265469, + 0.007882945239543915, + 0.008130761794745922, + 0.01896786130964756, + 0.013569947332143784, + 0.009039031341671944, + 0.020375650376081467, + 0.021121051162481308, + 0.021558305248618126, + 0.002750373911112547, + 0.0024051766376942396, + 0.0037197605706751347, + 0.003411844838410616, + 0.002101558493450284, + 0.0036036698147654533, + 0.003965088166296482, + 0.003684545401483774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.002822702983394265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18342088162899017, + 0.4471514821052551, + 0.04652351140975952, + 0.04966512322425842, + 0.07135587930679321, + 0.045834291726350784, + 0.016297997906804085, + 0.014997916296124458, + 0.05384449660778046, + 0.07090838253498077 + ], + "credit_map": [ + 0.011741640977561474, + 0.008909683674573898, + 0.05680153891444206, + 0.025823388248682022, + 0.01888514682650566, + 0.02664879709482193, + 0.01836164854466915, + 0.016249055042862892, + 0.019497228786349297, + 0.009579338133335114, + 0.18404558300971985, + 0.03023773804306984, + 0.010850337333977222, + 0.13605079054832458, + 0.029731648042798042, + 0.027158819139003754, + 0.0034908917732536793, + 0.004090727772563696, + 0.00674248905852437, + 0.006037602666765451, + 0.00486226100474596, + 0.011069935746490955, + 0.005433972459286451, + 0.004795635584741831, + 0.004863021429628134, + 0.004075449425727129, + 0.006678152829408646, + 0.005637900438159704, + 0.0046128761023283005, + 0.007326468359678984, + 0.007043017074465752, + 0.009428238496184349, + 0.005495033692568541, + 0.004962752107530832, + 0.013834003359079361, + 0.007330768276005983, + 0.0044176699593663216, + 0.021507585421204567, + 0.00720368092879653, + 0.00660438509657979, + 0.00304677477106452, + 0.0035379757173359394, + 0.004977080039680004, + 0.005270256660878658, + 0.0035733694676309824, + 0.007730359677225351, + 0.006002670619636774, + 0.011695809662342072, + 0.0015314000193029642, + 0.0017249668017029762, + 0.0023853760212659836, + 0.001902924501337111, + 0.0013903817161917686, + 0.0031408194918185472, + 0.0018685092218220234, + 0.0023536207154393196, + 0.0013649279717355967, + 0.0010773205431178212, + 0.002416446339339018, + 0.0018652660073712468, + 0.0012605456868186593, + 0.002868897747248411, + 0.001873746165074408, + 0.0022707669995725155, + 0.0035722425673156977, + 0.00381947448477149, + 0.010400005616247654, + 0.005107722710818052, + 0.004790508653968573, + 0.007973629981279373, + 0.0109191183000803, + 0.007261800114065409, + 0.006728222128003836, + 0.005392991006374359, + 0.008282259106636047, + 0.00817074254155159, + 0.007299627177417278, + 0.015688812360167503, + 0.010119724087417126, + 0.00922601018100977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.012536413036286831 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043041475117206573 + ], + "opposite_logprob_tokens": [ + -3.168041467666626 + ], + "logprob_reward": -0.043041475117206573, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043041475117206573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2564970552921295, + 0.1544695347547531, + 0.03652871400117874, + 0.06413678824901581, + 0.1817212849855423, + 0.08129201829433441, + 0.03219906985759735, + 0.020156538113951683, + 0.07299506664276123, + 0.10000388324260712 + ], + "credit_map": [ + 0.01635226607322693, + 0.009586655534803867, + 0.09919123351573944, + 0.025458918884396553, + 0.015114971436560154, + 0.043629925698041916, + 0.021725933998823166, + 0.025437159463763237, + 0.0057401107624173164, + 0.006343870889395475, + 0.05700397491455078, + 0.010927875526249409, + 0.009505354799330235, + 0.04682885482907295, + 0.010080751962959766, + 0.008038750849664211, + 0.0023795361630618572, + 0.0027478630654513836, + 0.006886375602334738, + 0.0042517418041825294, + 0.005867500323802233, + 0.004987942520529032, + 0.004381053615361452, + 0.005026699043810368, + 0.006425034254789352, + 0.009778638370335102, + 0.00804657582193613, + 0.007088026497513056, + 0.00464484840631485, + 0.011640762910246849, + 0.008232789114117622, + 0.008280114270746708, + 0.014960750006139278, + 0.007910340093076229, + 0.02833242155611515, + 0.015090631321072578, + 0.010569987818598747, + 0.06521997600793839, + 0.022029919549822807, + 0.017607269808650017, + 0.006638761609792709, + 0.0038595604710280895, + 0.009362603537738323, + 0.01322314701974392, + 0.005372087005525827, + 0.013626291416585445, + 0.011445257812738419, + 0.017764313146471977, + 0.003358320100232959, + 0.0026950305327773094, + 0.0045674326829612255, + 0.0037615897599607706, + 0.0032352448906749487, + 0.005917937029153109, + 0.004026852548122406, + 0.00463666208088398, + 0.002050631446763873, + 0.0009759669774211943, + 0.0021507591009140015, + 0.0020943244453519583, + 0.001267551793716848, + 0.004603283479809761, + 0.0033937422558665276, + 0.003620279487222433, + 0.005229829344898462, + 0.0046139070764184, + 0.015199893154203892, + 0.008971202187240124, + 0.0048990980722010136, + 0.010391980409622192, + 0.013988396152853966, + 0.009700764901936054, + 0.010145788080990314, + 0.008321264758706093, + 0.010647810995578766, + 0.008366775698959827, + 0.011722448281943798, + 0.019689735025167465, + 0.014226983301341534, + 0.01688307523727417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.043041475117206573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate horizontal tracking shot slides from left to right, revealing a profoundly still scene within a dim, ancient stone chamber. A sharp beam of sunlight, piercing the oppressive darkness from an unseen high window, slants across the frame, illuminating the central scene. An ancient, weathered crown rests solemnly upon a plush crimson velvet cushion. The crown itself is crafted from heavy, tarnished gold, its luster long faded to a dull glow, its surface marred by tiny scratches and the occasional deep dent from passing years. Its simple yet powerful design is interrupted by the empty surroundings where the jewel once rested. The velvet cushion, rich in color but slightly frayed at the edges, shows signs of age, sinking deeply beneath the weight of the crown. Within the brilliant light, countless motes of dust dance and twirl in a slow, captivating ballet, catching the light like tiny, glittering embers. The camera's steady sideways drift causes light to slide across the crown's textured surface, creating a subtle play of highlights and shadows that emphasize its age and history, evoking an atmosphere of powerful, silent, and contemplative forgotten majesty.", + "seed": 18632, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_4177_seed_18632_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02925687753595412, + "vlm_logprob": -0.02925687753595412, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.02925687753595412, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01968923583626747 + ], + "opposite_logprob_tokens": [ + -4.519689083099365 + ], + "logprob_reward": -0.01968923583626747, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01968923583626747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3776734471321106, + 0.174033984541893, + 0.09059473872184753, + 0.05121897533535957, + 0.03936215490102768, + 0.020742714405059814, + 0.07826998084783554, + 0.05567588657140732, + 0.0793619155883789, + 0.03306626155972481 + ], + "credit_map": [ + 0.09673316031694412, + 0.03305304795503616, + 0.014046312309801579, + 0.01930966228246689, + 0.033676985651254654, + 0.028421154245734215, + 0.12995649874210358, + 0.02247658744454384, + 0.02687470056116581, + 0.021129416301846504, + 0.006386358756572008, + 0.00800591055303812, + 0.020824581384658813, + 0.010213589295744896, + 0.0279676653444767, + 0.05263174697756767, + 0.01715526357293129, + 0.009910843335092068, + 0.005558742210268974, + 0.005654480773955584, + 0.013489572331309319, + 0.018390031531453133, + 0.007247955538332462, + 0.013187847100198269, + 0.009899321012198925, + 0.010402203537523746, + 0.002818530658259988, + 0.004046199377626181, + 0.005597489885985851, + 0.00796094536781311, + 0.0036943063605576754, + 0.00679997680708766, + 0.005669440142810345, + 0.005707564298063517, + 0.002476943889632821, + 0.0031069559045135975, + 0.006105305626988411, + 0.008570239879190922, + 0.0028989005368202925, + 0.004826809279620647, + 0.0030877289827913046, + 0.0030429745092988014, + 0.0032127429731190205, + 0.0014908185694366693, + 0.0022941159550100565, + 0.0030362545512616634, + 0.0018042128067463636, + 0.002773865358904004, + 0.010107850655913353, + 0.01287491712719202, + 0.006412480492144823, + 0.004896392580121756, + 0.01955440267920494, + 0.012851725332438946, + 0.0055989609099924564, + 0.00597325200214982, + 0.014310397207736969, + 0.008624888956546783, + 0.005150300916284323, + 0.004227601923048496, + 0.004476623144000769, + 0.011296344920992851, + 0.003634141990914941, + 0.003955586813390255, + 0.022452695295214653, + 0.005744535475969315, + 0.027574703097343445, + 0.005860438104718924, + 0.00435938686132431, + 0.005794062279164791, + 0.0037863412871956825, + 0.0037897578440606594, + 0.004804808646440506, + 0.004233610816299915, + 0.004432071465998888, + 0.0036841847468167543, + 0.006982491817325354, + 0.0029105269350111485, + 0.003270544344559312, + 0.002748024882748723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01968923583626747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029829099774360657 + ], + "opposite_logprob_tokens": [ + -3.5298290252685547 + ], + "logprob_reward": -0.029829099774360657, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029829099774360657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10634525120258331, + 0.2940048575401306, + 0.11074798554182053, + 0.059884727001190186, + 0.06743643432855606, + 0.09481959044933319, + 0.08533217012882233, + 0.036680493503808975, + 0.09014169871807098, + 0.05460679158568382 + ], + "credit_map": [ + 0.014530327171087265, + 0.011472629383206367, + 0.008651557378470898, + 0.010025283321738243, + 0.012653698213398457, + 0.010405403561890125, + 0.02219576947391033, + 0.01641058176755905, + 0.04868638142943382, + 0.03153074160218239, + 0.010739284567534924, + 0.012197836302220821, + 0.032287366688251495, + 0.018494782969355583, + 0.048911355435848236, + 0.09115710109472275, + 0.02394525147974491, + 0.01329080481082201, + 0.009195670485496521, + 0.008569418452680111, + 0.018110452219843864, + 0.0156752597540617, + 0.007746219635009766, + 0.014214912429451942, + 0.011782417073845863, + 0.007678251713514328, + 0.003667324548587203, + 0.005618085619062185, + 0.007753225043416023, + 0.009847009554505348, + 0.004619312472641468, + 0.008919098414480686, + 0.016916342079639435, + 0.0063955276273190975, + 0.003677734872326255, + 0.006550889927893877, + 0.007534154690802097, + 0.007623359560966492, + 0.004595463164150715, + 0.014142964966595173, + 0.013501905836164951, + 0.02467595972120762, + 0.01046980544924736, + 0.005741043481975794, + 0.011586219072341919, + 0.014193880371749401, + 0.00721063744276762, + 0.007440134882926941, + 0.01068828348070383, + 0.015250744298100471, + 0.0057436940260231495, + 0.005075356457382441, + 0.021798081696033478, + 0.010677173733711243, + 0.007588702253997326, + 0.008510138839483261, + 0.0077145639806985855, + 0.00616768142208457, + 0.0038589921314269304, + 0.002789183519780636, + 0.003875449998304248, + 0.007212326396256685, + 0.0024697016924619675, + 0.0025925959926098585, + 0.02395821176469326, + 0.007705854717642069, + 0.028067881241440773, + 0.006919315550476313, + 0.0060572028160095215, + 0.007359278853982687, + 0.005370552651584148, + 0.004703395999968052, + 0.008859160356223583, + 0.007283835206180811, + 0.007321246899664402, + 0.005506450776010752, + 0.011512595228850842, + 0.00538357300683856, + 0.004579541739076376, + 0.004160386510193348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.029829099774360657 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624063218012452 + ], + "opposite_logprob_tokens": [ + -5.628623962402344 + ], + "logprob_reward": -0.003624063218012452, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624063218012452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12474503368139267, + 0.034908782690763474, + 0.040836889296770096, + 0.16915813088417053, + 0.4084014594554901, + 0.026291707530617714, + 0.04916485399007797, + 0.04990246519446373, + 0.0697610080242157, + 0.026829678565263748 + ], + "credit_map": [ + 0.03336448222398758, + 0.011666381731629372, + 0.004437008406966925, + 0.005875641014426947, + 0.009320464916527271, + 0.00916093960404396, + 0.043276142328977585, + 0.007643967401236296, + 0.004782782401889563, + 0.00546247698366642, + 0.0032168247271329165, + 0.0027642017230391502, + 0.005486947018653154, + 0.004589741583913565, + 0.0033725968096405268, + 0.0052332100458443165, + 0.0068691568449139595, + 0.005147177260369062, + 0.0032911666203290224, + 0.0036758107598870993, + 0.005032022949308157, + 0.00589989498257637, + 0.005309928208589554, + 0.005611728876829147, + 0.024731408804655075, + 0.04061473160982132, + 0.007838897407054901, + 0.013686192221939564, + 0.02976466342806816, + 0.022438667714595795, + 0.008035057224333286, + 0.022048506885766983, + 0.1143202930688858, + 0.05340535566210747, + 0.013661539182066917, + 0.01701963134109974, + 0.045109935104846954, + 0.045398835092782974, + 0.018005680292844772, + 0.1014801561832428, + 0.003972779959440231, + 0.0032194301020354033, + 0.0038429228588938713, + 0.0024783918634057045, + 0.0032196210231631994, + 0.0033425737638026476, + 0.0029472080059349537, + 0.0032687794882804155, + 0.0061200205236673355, + 0.009328965097665787, + 0.004262933507561684, + 0.004774728789925575, + 0.0091088293120265, + 0.008596634492278099, + 0.003209981368854642, + 0.003762762527912855, + 0.007420072797685862, + 0.005975712090730667, + 0.004645598586648703, + 0.004256921354681253, + 0.005140372086316347, + 0.00750538008287549, + 0.005325449630618095, + 0.009632959961891174, + 0.013502830639481544, + 0.009415588341653347, + 0.017064278945326805, + 0.007969260215759277, + 0.0050887162797153, + 0.004052557051181793, + 0.005733751691877842, + 0.0069340188056230545, + 0.0034333907533437014, + 0.0032532960176467896, + 0.0027107750065624714, + 0.0036979459691792727, + 0.0044411844573915005, + 0.002572936238721013, + 0.0035898268688470125, + 0.0031303237192332745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.003624063218012452 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191539026796818 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191539026796818, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191539026796818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29319608211517334, + 0.20085789263248444, + 0.06465685367584229, + 0.0883873850107193, + 0.060736458748579025, + 0.07195006310939789, + 0.034305673092603683, + 0.05410844832658768, + 0.10563547909259796, + 0.026165654882788658 + ], + "credit_map": [ + 0.08038535714149475, + 0.019950004294514656, + 0.009926551952958107, + 0.014570259489119053, + 0.016912860795855522, + 0.023441901430487633, + 0.1069893017411232, + 0.021019848063588142, + 0.027762504294514656, + 0.027767697349190712, + 0.008788464590907097, + 0.009272743947803974, + 0.025683866813778877, + 0.014221325516700745, + 0.030816854909062386, + 0.05654444918036461, + 0.015004097484052181, + 0.007811263203620911, + 0.004774205386638641, + 0.0052817161194980145, + 0.009596139192581177, + 0.0077711003832519054, + 0.0057664853520691395, + 0.00865184422582388, + 0.01773279719054699, + 0.01281916443258524, + 0.005221950821578503, + 0.007286428939551115, + 0.00981492642313242, + 0.014172261580824852, + 0.008350544609129429, + 0.012989312410354614, + 0.008591994643211365, + 0.01058514229953289, + 0.003547298489138484, + 0.004764719400554895, + 0.010964364744722843, + 0.00902594719082117, + 0.004636459052562714, + 0.008620532229542732, + 0.01021815650165081, + 0.01652289554476738, + 0.008142671547830105, + 0.004401951562613249, + 0.008841939270496368, + 0.010931426659226418, + 0.00619654031470418, + 0.006694484036415815, + 0.004285457544028759, + 0.0058077177964150906, + 0.0039637829177081585, + 0.002598532708361745, + 0.0068196142092347145, + 0.004355122800916433, + 0.0039090923964977264, + 0.0025663510896265507, + 0.011473790742456913, + 0.00858686026185751, + 0.004885913338512182, + 0.003358814399689436, + 0.005392147693783045, + 0.010415870696306229, + 0.004743007943034172, + 0.005252042319625616, + 0.02181040309369564, + 0.00883612409234047, + 0.025119008496403694, + 0.010159575380384922, + 0.012797204777598381, + 0.00911355298012495, + 0.010916495695710182, + 0.0068831127136945724, + 0.0038304277695715427, + 0.0043411036022007465, + 0.0026340538170188665, + 0.003430987475439906, + 0.004590983968228102, + 0.0027738362550735474, + 0.0021960679441690445, + 0.0023681942839175463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014191539026796818 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895044982433319 + ], + "opposite_logprob_tokens": [ + -2.5789504051208496 + ], + "logprob_reward": -0.07895044982433319, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895044982433319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2072528600692749, + 0.12016284465789795, + 0.05438665300607681, + 0.05918951332569122, + 0.10658413171768188, + 0.10628477483987808, + 0.05872931331396103, + 0.10480912029743195, + 0.15380163490772247, + 0.028799118474125862 + ], + "credit_map": [ + 0.05691593140363693, + 0.01528141088783741, + 0.008570062927901745, + 0.009044402278959751, + 0.014407962560653687, + 0.0162434633821249, + 0.07241339981555939, + 0.014376216568052769, + 0.017342280596494675, + 0.011665884405374527, + 0.004690633155405521, + 0.004894652869552374, + 0.01348848082125187, + 0.007804354652762413, + 0.020638104528188705, + 0.039638448506593704, + 0.007959874346852303, + 0.0068704853765666485, + 0.004426372237503529, + 0.008421579375863075, + 0.006725119426846504, + 0.007051228079944849, + 0.006544508505612612, + 0.006387482397258282, + 0.01056710071861744, + 0.010153581388294697, + 0.0030823119450360537, + 0.004578292369842529, + 0.009700159542262554, + 0.009039881639182568, + 0.0033306835684925318, + 0.008737497963011265, + 0.017628895118832588, + 0.02191859483718872, + 0.005362934432923794, + 0.00685101980343461, + 0.017298730090260506, + 0.015891144052147865, + 0.006146708969026804, + 0.015486103482544422, + 0.014226350001990795, + 0.024742117151618004, + 0.010407430119812489, + 0.00750645250082016, + 0.01449335552752018, + 0.014219957403838634, + 0.009363433346152306, + 0.011325688101351261, + 0.007313570939004421, + 0.008564063347876072, + 0.0056965709663927555, + 0.006063472479581833, + 0.008480001240968704, + 0.009639891795814037, + 0.007012970745563507, + 0.005958767142146826, + 0.0157946664839983, + 0.017254026606678963, + 0.008385173976421356, + 0.006949564907699823, + 0.018255116418004036, + 0.015592483803629875, + 0.009214294143021107, + 0.013363787904381752, + 0.018464557826519012, + 0.023958545178174973, + 0.01568225957453251, + 0.02613053098320961, + 0.016644898802042007, + 0.011800401844084263, + 0.019786493852734566, + 0.021333957090973854, + 0.0037456322461366653, + 0.00476994551718235, + 0.003566476283594966, + 0.0035769084934145212, + 0.004522612784057856, + 0.0024315996561199427, + 0.003109576180577278, + 0.003076364751905203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.07895044982433319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate horizontal tracking shot slides from left to right, revealing a profoundly still scene within a dim, ancient stone chamber. A sharp beam of sunlight, piercing the oppressive darkness from an unseen high window, slants across the frame, illuminating the central scene. An ancient, weathered crown rests solemnly upon a plush crimson velvet cushion. The crown itself is crafted from heavy, tarnished gold, its luster long faded to a dull glow, its surface marred by tiny scratches and the occasional deep dent from passing years. Its simple yet powerful design is interrupted by the empty surroundings where the jewel once rested. The velvet cushion, rich in color but slightly frayed at the edges, shows signs of age, sinking deeply beneath the weight of the crown. Within the brilliant light, countless motes of dust dance and twirl in a slow, captivating ballet, catching the light like tiny, glittering embers. The camera's steady sideways drift causes light to slide across the crown's textured surface, creating a subtle play of highlights and shadows that emphasize its age and history, evoking an atmosphere of powerful, silent, and contemplative forgotten majesty.", + "seed": 18633, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_4177_seed_18633_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03801699606701732, + "vlm_logprob": -0.03801699606701732, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03801699606701732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026880968362092972 + ], + "opposite_logprob_tokens": [ + -3.776880979537964 + ], + "logprob_reward": -0.026880968362092972, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026880968362092972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09687338024377823, + 0.03545766696333885, + 0.054717496037483215, + 0.020277906209230423, + 0.048207491636276245, + 0.09185972809791565, + 0.0379432737827301, + 0.13776107132434845, + 0.028680654242634773, + 0.4482212960720062 + ], + "credit_map": [ + 0.008687475696206093, + 0.005944217089563608, + 0.00606285035610199, + 0.014863327145576477, + 0.00962925236672163, + 0.018822835758328438, + 0.02146945334970951, + 0.011393978260457516, + 0.004491451196372509, + 0.0029527032747864723, + 0.002468075370416045, + 0.003279665019363165, + 0.0036495784297585487, + 0.007379464339464903, + 0.005573093891143799, + 0.00566363800317049, + 0.00971038918942213, + 0.005119701381772757, + 0.004808208905160427, + 0.0060190255753695965, + 0.003751120064407587, + 0.00891958549618721, + 0.007811345625668764, + 0.008578128181397915, + 0.0030566162895411253, + 0.002073611132800579, + 0.0012759565142914653, + 0.0022604891564697027, + 0.0031461629550904036, + 0.0030621879268437624, + 0.0031579704955220222, + 0.002244915347546339, + 0.0099556939676404, + 0.005335867404937744, + 0.0032765185460448265, + 0.00458860956132412, + 0.005869282875210047, + 0.007533104624599218, + 0.004609014373272657, + 0.0070394049398601055, + 0.021332930773496628, + 0.014709802344441414, + 0.003749263472855091, + 0.0054909116588532925, + 0.018968287855386734, + 0.014931517653167248, + 0.007573842536658049, + 0.0051031894981861115, + 0.008775015361607075, + 0.005302625708281994, + 0.0030118003487586975, + 0.0031523515935987234, + 0.0057439543306827545, + 0.0060935718938708305, + 0.002994778100401163, + 0.0028691799379885197, + 0.02352437749505043, + 0.024709496647119522, + 0.01605377346277237, + 0.009371503256261349, + 0.027771122753620148, + 0.017038481310009956, + 0.007970855571329594, + 0.011321467347443104, + 0.004414395894855261, + 0.00517135439440608, + 0.0025644695851951838, + 0.002878746250644326, + 0.005488916765898466, + 0.003779036458581686, + 0.002937932498753071, + 0.0014458036748692393, + 0.15722766518592834, + 0.02465108223259449, + 0.009351598098874092, + 0.01179086696356535, + 0.026858972385525703, + 0.18932199478149414, + 0.014578690752387047, + 0.014440462924540043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.026880968362092972 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00411330908536911 + ], + "opposite_logprob_tokens": [ + -5.50411319732666 + ], + "logprob_reward": -0.00411330908536911, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00411330908536911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05102846026420593, + 0.22712191939353943, + 0.08649899065494537, + 0.028949221596121788, + 0.24359869956970215, + 0.018997523933649063, + 0.03658599779009819, + 0.06745380163192749, + 0.029928749427199364, + 0.2098366618156433 + ], + "credit_map": [ + 0.005758268758654594, + 0.0035038446076214314, + 0.0052888658829033375, + 0.005544818006455898, + 0.004448514897376299, + 0.008864597417414188, + 0.00816851295530796, + 0.00945103820413351, + 0.018458213657140732, + 0.020309045910835266, + 0.015093059279024601, + 0.01654226705431938, + 0.038626037538051605, + 0.043522875756025314, + 0.04737985134124756, + 0.027190567925572395, + 0.00951378047466278, + 0.0064294361509382725, + 0.00752060255035758, + 0.008874915540218353, + 0.0066300262697041035, + 0.015455619432032108, + 0.014664974063634872, + 0.01740962825715542, + 0.0028633642941713333, + 0.0038938040379434824, + 0.00149018713273108, + 0.0030776350758969784, + 0.004486537538468838, + 0.004429982975125313, + 0.004615169949829578, + 0.004092540126293898, + 0.04293259605765343, + 0.024370677769184113, + 0.022083507850766182, + 0.026198506355285645, + 0.03009776957333088, + 0.03720435872673988, + 0.025514932349324226, + 0.035196345299482346, + 0.002531945239752531, + 0.0030364645645022392, + 0.0017210290534421802, + 0.0017087701708078384, + 0.0027034454979002476, + 0.0026807994581758976, + 0.0023837415501475334, + 0.002231327584013343, + 0.00703999912366271, + 0.005177337676286697, + 0.00460425578057766, + 0.002500997856259346, + 0.004626404028385878, + 0.006083102896809578, + 0.0032224294263869524, + 0.003331469837576151, + 0.011458970606327057, + 0.011773078702390194, + 0.00828277226537466, + 0.005097343120723963, + 0.010064622387290001, + 0.008319206535816193, + 0.006062522530555725, + 0.006395286414772272, + 0.004815630614757538, + 0.006976693868637085, + 0.002679933328181505, + 0.0023027032148092985, + 0.004799391608685255, + 0.004504004027694464, + 0.0024258026387542486, + 0.0014245909405872226, + 0.0746321827173233, + 0.011531943455338478, + 0.004827772732824087, + 0.005832628346979618, + 0.009924798272550106, + 0.08921618014574051, + 0.006970532238483429, + 0.006900615990161896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.00411330908536911 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255454685539007 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255454685539007, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255454685539007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17762404680252075, + 0.11040907353162766, + 0.14734897017478943, + 0.10254998505115509, + 0.08371574431657791, + 0.0459505096077919, + 0.06335952877998352, + 0.03579701855778694, + 0.06428947299718857, + 0.16895560920238495 + ], + "credit_map": [ + 0.01751735433936119, + 0.018143096938729286, + 0.013828251510858536, + 0.0243925079703331, + 0.023581750690937042, + 0.03031863085925579, + 0.028246428817510605, + 0.021596036851406097, + 0.009338886477053165, + 0.011695900931954384, + 0.005933037959039211, + 0.0110301673412323, + 0.018554769456386566, + 0.018632663413882256, + 0.02020714245736599, + 0.015016512013971806, + 0.010979239828884602, + 0.009642435237765312, + 0.018542837351560593, + 0.022793468087911606, + 0.010329010896384716, + 0.024409951642155647, + 0.023110169917345047, + 0.027541859075427055, + 0.010536646470427513, + 0.01192175131291151, + 0.0070187714882195, + 0.00819764006882906, + 0.012109273113310337, + 0.016763607040047646, + 0.02117331512272358, + 0.014828980900347233, + 0.014353028498589993, + 0.008744846098124981, + 0.008994955569505692, + 0.009380501694977283, + 0.00890122726559639, + 0.012095845304429531, + 0.008936433121562004, + 0.012308907695114613, + 0.00881064124405384, + 0.007002722006291151, + 0.003644652431830764, + 0.004098236095160246, + 0.0058671883307397366, + 0.00651873555034399, + 0.0050058080814778805, + 0.005002520978450775, + 0.009660864248871803, + 0.008733662776648998, + 0.00835611391812563, + 0.004447001498192549, + 0.010772529058158398, + 0.00955524668097496, + 0.005890995729714632, + 0.005943109281361103, + 0.005701020359992981, + 0.005771999713033438, + 0.003785340115427971, + 0.0030993542168289423, + 0.004419177770614624, + 0.00425583915784955, + 0.005343637894839048, + 0.0034206497948616743, + 0.009511824697256088, + 0.014148533344268799, + 0.005175935570150614, + 0.005883164703845978, + 0.007636884227395058, + 0.012451332993805408, + 0.005751945078372955, + 0.003729855641722679, + 0.057147230952978134, + 0.010622264817357063, + 0.004093926399946213, + 0.004779656417667866, + 0.013064253143966198, + 0.06689396500587463, + 0.005855885334312916, + 0.006498442962765694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.005255454685539007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070143431425095 + ], + "opposite_logprob_tokens": [ + -4.511070251464844 + ], + "logprob_reward": -0.011070143431425095, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070143431425095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12316365540027618, + 0.12831394374370575, + 0.10656267404556274, + 0.031674161553382874, + 0.22086714208126068, + 0.02977791801095009, + 0.046561844646930695, + 0.05031134933233261, + 0.04363887384533882, + 0.21912841498851776 + ], + "credit_map": [ + 0.01212294027209282, + 0.010224276222288609, + 0.00927144568413496, + 0.01893298327922821, + 0.016557732596993446, + 0.01855745166540146, + 0.022711239755153656, + 0.014785594306886196, + 0.010495306923985481, + 0.011388145387172699, + 0.00886897649616003, + 0.009810454212129116, + 0.023031214252114296, + 0.02356707490980625, + 0.027528686448931694, + 0.013624075800180435, + 0.013609256595373154, + 0.009795984253287315, + 0.010119667276740074, + 0.015530645847320557, + 0.010699138045310974, + 0.016377024352550507, + 0.015814155340194702, + 0.014616794884204865, + 0.0035249129869043827, + 0.0037332293577492237, + 0.00134957954287529, + 0.0023661574814468622, + 0.0055695706978440285, + 0.0045452299527823925, + 0.006257226690649986, + 0.004328254144638777, + 0.03769654035568237, + 0.02099435403943062, + 0.020640017464756966, + 0.02629741095006466, + 0.026012325659394264, + 0.03180328384041786, + 0.024637162685394287, + 0.032786041498184204, + 0.006795091088861227, + 0.0038108909502625465, + 0.0019849957898259163, + 0.001710446085780859, + 0.004763057921081781, + 0.0048232125118374825, + 0.0025677229277789593, + 0.0033224988728761673, + 0.008140096440911293, + 0.00796123780310154, + 0.003887234954163432, + 0.004298646468669176, + 0.006644625216722488, + 0.0075456234626472, + 0.004312482196837664, + 0.0037718990352004766, + 0.008333695121109486, + 0.007401284761726856, + 0.00536586157977581, + 0.0038258638232946396, + 0.008592997677624226, + 0.006336309481412172, + 0.0059771169908344746, + 0.004478216636925936, + 0.006443398538976908, + 0.009320996701717377, + 0.0036457255482673645, + 0.0033460790291428566, + 0.007975231856107712, + 0.0061098006553947926, + 0.004705111496150494, + 0.002092530718073249, + 0.07743664085865021, + 0.012159507721662521, + 0.005012116394937038, + 0.006568310782313347, + 0.01087642926722765, + 0.09230928122997284, + 0.007685160730034113, + 0.007080951239913702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.011070143431425095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427651047706604 + ], + "opposite_logprob_tokens": [ + -2.0177650451660156 + ], + "logprob_reward": -0.1427651047706604, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427651047706604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09988754242658615, + 0.1802915632724762, + 0.07025185227394104, + 0.0813218355178833, + 0.11665116250514984, + 0.030789690092206, + 0.061925485730171204, + 0.26855310797691345, + 0.045370426028966904, + 0.04495738074183464 + ], + "credit_map": [ + 0.008184676058590412, + 0.007973501458764076, + 0.006610826123505831, + 0.01384605560451746, + 0.01046523917466402, + 0.017550358548760414, + 0.021355561912059784, + 0.01390132401138544, + 0.013808250427246094, + 0.01624700427055359, + 0.011461972258985043, + 0.014503550715744495, + 0.03034009411931038, + 0.033746521919965744, + 0.03632735460996628, + 0.023856813088059425, + 0.007007165811955929, + 0.0050630345940589905, + 0.009259840473532677, + 0.010651946067810059, + 0.004872783087193966, + 0.011578449048101902, + 0.008873309008777142, + 0.012945326045155525, + 0.009452464058995247, + 0.007830715738236904, + 0.005433422047644854, + 0.009831358678638935, + 0.0119247455149889, + 0.012411534786224365, + 0.013636065647006035, + 0.010801534168422222, + 0.019623905420303345, + 0.010548342950642109, + 0.00824363436549902, + 0.010363834910094738, + 0.010913461446762085, + 0.02102762460708618, + 0.00962891522794962, + 0.026301439851522446, + 0.005671476013958454, + 0.0034974615555256605, + 0.0026046037673950195, + 0.0018990851240232587, + 0.005279011558741331, + 0.004941396880894899, + 0.002526749623939395, + 0.004369908012449741, + 0.011445616371929646, + 0.009093757718801498, + 0.0064953044056892395, + 0.0044226255267858505, + 0.007900484837591648, + 0.010702804662287235, + 0.0054860711097717285, + 0.0063788252882659435, + 0.04443078488111496, + 0.04491347074508667, + 0.030873754993081093, + 0.017907485365867615, + 0.05624766647815704, + 0.032424814999103546, + 0.01915217563509941, + 0.022602958604693413, + 0.00777227571234107, + 0.008693121373653412, + 0.004161003977060318, + 0.0031965700909495354, + 0.005063826218247414, + 0.009945877827703953, + 0.0038892398588359356, + 0.0026485100388526917, + 0.007517832797020674, + 0.005622813943773508, + 0.0030319381039589643, + 0.0032400465570390224, + 0.009500752203166485, + 0.00915408693253994, + 0.004141896963119507, + 0.002748009981587529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.1427651047706604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate horizontal tracking shot slides from left to right, revealing a profoundly still scene within a dim, ancient stone chamber. A sharp beam of sunlight, piercing the oppressive darkness from an unseen high window, slants across the frame, illuminating the central scene. An ancient, weathered crown rests solemnly upon a plush crimson velvet cushion. The crown itself is crafted from heavy, tarnished gold, its luster long faded to a dull glow, its surface marred by tiny scratches and the occasional deep dent from passing years. Its simple yet powerful design is interrupted by the empty surroundings where the jewel once rested. The velvet cushion, rich in color but slightly frayed at the edges, shows signs of age, sinking deeply beneath the weight of the crown. Within the brilliant light, countless motes of dust dance and twirl in a slow, captivating ballet, catching the light like tiny, glittering embers. The camera's steady sideways drift causes light to slide across the crown's textured surface, creating a subtle play of highlights and shadows that emphasize its age and history, evoking an atmosphere of powerful, silent, and contemplative forgotten majesty.", + "seed": 18634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_2/index_4177_seed_18634_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020423682499676944, + "vlm_logprob": -0.020423682499676944, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.020423682499676944, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037752751260995865 + ], + "opposite_logprob_tokens": [ + -3.412752866744995 + ], + "logprob_reward": -0.037752751260995865, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.037752751260995865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19187189638614655, + 0.2081228643655777, + 0.0911230593919754, + 0.16855034232139587, + 0.05703199282288551, + 0.034812357276678085, + 0.046408481895923615, + 0.05867216736078262, + 0.053671158850193024, + 0.08973564207553864 + ], + "credit_map": [ + 0.023582877591252327, + 0.027175860479474068, + 0.021025754511356354, + 0.022243758663535118, + 0.03947187215089798, + 0.024969035759568214, + 0.017480267211794853, + 0.015922492370009422, + 0.019528472796082497, + 0.041312325745821, + 0.04374704509973526, + 0.024934619665145874, + 0.028214219957590103, + 0.02127673476934433, + 0.01778561808168888, + 0.011323845945298672, + 0.010866230353713036, + 0.018557025119662285, + 0.009634867310523987, + 0.011358683928847313, + 0.013330807909369469, + 0.011492155492305756, + 0.011233517900109291, + 0.0046497732400894165, + 0.01712832599878311, + 0.023091675713658333, + 0.008139573968946934, + 0.013463740237057209, + 0.03574797511100769, + 0.02620891109108925, + 0.024718860164284706, + 0.020051300525665283, + 0.006952365860342979, + 0.010525135323405266, + 0.004587762523442507, + 0.004280915949493647, + 0.007304798811674118, + 0.006799390073865652, + 0.006821609102189541, + 0.009760018438100815, + 0.0070584239438176155, + 0.005313510540872812, + 0.0029804816003888845, + 0.002822315786033869, + 0.004131990484893322, + 0.0045762392692267895, + 0.0035888892598450184, + 0.0043405103497207165, + 0.013088220730423927, + 0.004244369454681873, + 0.002860529348254204, + 0.003026794409379363, + 0.0045386385172605515, + 0.012806509621441364, + 0.0031549730338156223, + 0.002688452135771513, + 0.009943947196006775, + 0.0065104723908007145, + 0.004912098404020071, + 0.004431212320923805, + 0.0112477270886302, + 0.008524746634066105, + 0.00954237300902605, + 0.0035595984663814306, + 0.0083997193723917, + 0.009857356548309326, + 0.009542439132928848, + 0.004444984253495932, + 0.005674650426954031, + 0.007104940712451935, + 0.004110993351787329, + 0.004536082968115807, + 0.012599446810781956, + 0.011662730947136879, + 0.009898546151816845, + 0.011721971444785595, + 0.009571152739226818, + 0.015668489038944244, + 0.012836639769375324, + 0.005776677280664444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.037752751260995865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002839939668774605 + ], + "opposite_logprob_tokens": [ + -5.877840042114258 + ], + "logprob_reward": -0.002839939668774605, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002839939668774605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08098871260881424, + 0.6161908507347107, + 0.08340273052453995, + 0.03801363706588745, + 0.017777349799871445, + 0.04477952793240547, + 0.029459118843078613, + 0.012422258965671062, + 0.02238299511373043, + 0.05458280071616173 + ], + "credit_map": [ + 0.00960992556065321, + 0.012414413504302502, + 0.007804314140230417, + 0.007954066619277, + 0.016352031379938126, + 0.011086716316640377, + 0.007022069301456213, + 0.008745172061026096, + 0.06797131150960922, + 0.08298647403717041, + 0.09003549814224243, + 0.0774540826678276, + 0.09220720827579498, + 0.09733710438013077, + 0.0622018538415432, + 0.04599730670452118, + 0.008048314601182938, + 0.010413147509098053, + 0.00847090594470501, + 0.005179909989237785, + 0.02186906710267067, + 0.011013021692633629, + 0.013969238847494125, + 0.004439132753759623, + 0.002922568703070283, + 0.004603877663612366, + 0.003731663804501295, + 0.004577097948640585, + 0.00652739591896534, + 0.005607375409454107, + 0.004325026646256447, + 0.005718631204217672, + 0.0021803593263030052, + 0.00324204470962286, + 0.0014337123138830066, + 0.0017429441213607788, + 0.0021072784438729286, + 0.0022146510891616344, + 0.002288993215188384, + 0.0025673657655715942, + 0.007891193963587284, + 0.005762229207903147, + 0.003571450477465987, + 0.003975707106292248, + 0.005678183399140835, + 0.005722468253225088, + 0.0048290458507835865, + 0.007349257357418537, + 0.007589888758957386, + 0.0036524091847240925, + 0.001591466017998755, + 0.0012486049672588706, + 0.0030461722053587437, + 0.008762054145336151, + 0.0018561050528660417, + 0.0017124192090705037, + 0.0019823373295366764, + 0.001348277204670012, + 0.0014391362201422453, + 0.0009431844810023904, + 0.0021580823231488466, + 0.0017709991661831737, + 0.002019293140619993, + 0.0007609504973515868, + 0.0033541882876306772, + 0.0035401880741119385, + 0.002238510176539421, + 0.00266578933224082, + 0.003243943676352501, + 0.002095386851578951, + 0.003240308491513133, + 0.0020046818535774946, + 0.007757655344903469, + 0.007085305172950029, + 0.005961055867373943, + 0.006323030684143305, + 0.00937620084732771, + 0.009729484096169472, + 0.005873559508472681, + 0.002476512687280774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.002839939668774605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059529622085392475 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.0059529622085392475, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059529622085392475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15172424912452698, + 0.09603136777877808, + 0.10017935186624527, + 0.06305459886789322, + 0.04007283225655556, + 0.07058199495077133, + 0.2529933750629425, + 0.11403366923332214, + 0.05070451274514198, + 0.06062411144375801 + ], + "credit_map": [ + 0.018981782719492912, + 0.02670617215335369, + 0.022878030315041542, + 0.015862053260207176, + 0.02952180616557598, + 0.01884590834379196, + 0.007789976894855499, + 0.011138512752950191, + 0.007408383302390575, + 0.01891295239329338, + 0.017752282321453094, + 0.01553293690085411, + 0.010676279664039612, + 0.011892670765519142, + 0.007853852584958076, + 0.0060020070523023605, + 0.010374451987445354, + 0.012916143983602524, + 0.012007211335003376, + 0.011829757131636143, + 0.016251109540462494, + 0.01795606128871441, + 0.011830453760921955, + 0.007014159113168716, + 0.005517344456166029, + 0.007491603028029203, + 0.003418031381443143, + 0.00764993904158473, + 0.012823156081140041, + 0.01083065290004015, + 0.00870059709995985, + 0.006623279768973589, + 0.0069745066575706005, + 0.0057971361093223095, + 0.0034046205691993237, + 0.0028286639135330915, + 0.006733606103807688, + 0.005311114247888327, + 0.003966793417930603, + 0.005056392401456833, + 0.012717987410724163, + 0.010175212286412716, + 0.009542467072606087, + 0.0044809323735535145, + 0.007897271774709225, + 0.008895352482795715, + 0.004525404889136553, + 0.012347370386123657, + 0.05063612386584282, + 0.03498796001076698, + 0.013566938228905201, + 0.02277720347046852, + 0.04551376402378082, + 0.05267075076699257, + 0.019292764365673065, + 0.013547864742577076, + 0.018582820892333984, + 0.013462304137647152, + 0.009832801297307014, + 0.007801415864378214, + 0.02085208334028721, + 0.016743851825594902, + 0.019338056445121765, + 0.007420334033668041, + 0.008075540885329247, + 0.011185776442289352, + 0.006399050820618868, + 0.0038889278657734394, + 0.007277593482285738, + 0.0059813703410327435, + 0.004324791021645069, + 0.003571461420506239, + 0.007148525211960077, + 0.009906010702252388, + 0.007332244887948036, + 0.006832127925008535, + 0.010035655461251736, + 0.0075792125426232815, + 0.008426095359027386, + 0.0033642363268882036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0059529622085392475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536648660898209 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536648660898209, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536648660898209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0996614322066307, + 0.14350326359272003, + 0.19371381402015686, + 0.08795223385095596, + 0.16779501736164093, + 0.048873044550418854, + 0.07266999036073685, + 0.018231146037578583, + 0.12848198413848877, + 0.039117973297834396 + ], + "credit_map": [ + 0.01651313714683056, + 0.014844991266727448, + 0.011500096879899502, + 0.010358289815485477, + 0.015818996354937553, + 0.013411659747362137, + 0.007350939325988293, + 0.009863331913948059, + 0.016137434169650078, + 0.019072433933615685, + 0.01960570365190506, + 0.023104606196284294, + 0.02305789105594158, + 0.0209035687148571, + 0.013672512955963612, + 0.007949115708470345, + 0.02143390290439129, + 0.018333645537495613, + 0.018301749601960182, + 0.01175777893513441, + 0.04082441329956055, + 0.04187869280576706, + 0.028958644717931747, + 0.012225007638335228, + 0.006918672472238541, + 0.010886849835515022, + 0.007830813527107239, + 0.01059228926897049, + 0.0176243856549263, + 0.012179144658148289, + 0.009875289164483547, + 0.012044795788824558, + 0.021669620648026466, + 0.030738787725567818, + 0.014431633986532688, + 0.014538914896547794, + 0.021666353568434715, + 0.020427322015166283, + 0.020531415939331055, + 0.023790989071130753, + 0.006956867873668671, + 0.006723743863403797, + 0.0035914841573685408, + 0.0044496674090623856, + 0.006765430793166161, + 0.006823592819273472, + 0.0042260559275746346, + 0.009336204268038273, + 0.017791064456105232, + 0.007222609128803015, + 0.004455545451492071, + 0.0050556715577840805, + 0.009197359904646873, + 0.01913006603717804, + 0.005502606276422739, + 0.004315071739256382, + 0.0025482354685664177, + 0.0031497557647526264, + 0.0019666976295411587, + 0.0014278749004006386, + 0.002329018898308277, + 0.0031804742757230997, + 0.002467556158080697, + 0.0011615344556048512, + 0.01853570155799389, + 0.020186273381114006, + 0.019619852304458618, + 0.011726107448339462, + 0.016712091863155365, + 0.018309177830815315, + 0.010847222059965134, + 0.012545580975711346, + 0.006226054858416319, + 0.00554091390222311, + 0.004337099380791187, + 0.0036735280882567167, + 0.008329457603394985, + 0.0045915222726762295, + 0.0043055154383182526, + 0.0021138833835721016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.012536648660898209 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dim, ancient stone chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sharp beam of sunlight slanting across the frame from an unseen high window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ancient, weathered crown made of tarnished gold resting on a plush crimson velvet cushion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless motes of dust visible dancing in the light like tiny embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, deliberate horizontal tracking shot sliding from left to right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303611069917679 + ], + "opposite_logprob_tokens": [ + -3.1680362224578857 + ], + "logprob_reward": -0.04303611069917679, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303611069917679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0812326967716217, + 0.19873479008674622, + 0.13982176780700684, + 0.05698937922716141, + 0.04156804084777832, + 0.15609440207481384, + 0.051733408123254776, + 0.033698271960020065, + 0.11732592433691025, + 0.12280136346817017 + ], + "credit_map": [ + 0.010303905233740807, + 0.010906633920967579, + 0.010458146221935749, + 0.009088600054383278, + 0.013980547897517681, + 0.00943213701248169, + 0.00944868940860033, + 0.0076140305027365685, + 0.01828603260219097, + 0.04414065554738045, + 0.04686848446726799, + 0.033911559730768204, + 0.01800500974059105, + 0.016320478171110153, + 0.010502503253519535, + 0.010700042359530926, + 0.009536745958030224, + 0.020460408180952072, + 0.011686151847243309, + 0.015119197778403759, + 0.041818924248218536, + 0.010322831571102142, + 0.026415735483169556, + 0.004461756907403469, + 0.004814356565475464, + 0.005555751733481884, + 0.005595138296484947, + 0.007017378695309162, + 0.012097508646547794, + 0.008628947660326958, + 0.007275195326656103, + 0.006005098577588797, + 0.006462355609983206, + 0.0039541940204799175, + 0.002777369925752282, + 0.0035044392570853233, + 0.006636822130531073, + 0.005542827304452658, + 0.004976562689989805, + 0.007713465951383114, + 0.02385297417640686, + 0.02010815031826496, + 0.013844497501850128, + 0.014552685432136059, + 0.021912425756454468, + 0.018920550122857094, + 0.017876053228974342, + 0.02502705156803131, + 0.009388118050992489, + 0.008039465174078941, + 0.0038916636258363724, + 0.0033105164766311646, + 0.007787264417856932, + 0.010291003622114658, + 0.0048726084642112255, + 0.004152762703597546, + 0.005506841000169516, + 0.00461951270699501, + 0.003978276159614325, + 0.002431467641144991, + 0.005932836327701807, + 0.0036862725391983986, + 0.005570188630372286, + 0.0019728727638721466, + 0.018706755712628365, + 0.021800175309181213, + 0.019146619364619255, + 0.009968765079975128, + 0.012716029770672321, + 0.01628313958644867, + 0.007096836809068918, + 0.011607591062784195, + 0.014150240458548069, + 0.014623486436903477, + 0.014398230239748955, + 0.011239221319556236, + 0.019353684037923813, + 0.030047696083784103, + 0.011049474589526653, + 0.007939323782920837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.04303611069917679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_tensor([6395, 6395])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_tensor([6395, 6395])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dab836a53f556bf6aff2b24c43690d1b09fe3ec6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_tensor([6395, 6395])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle shot captures a scene of pure terror within a dark, narrow corridor, rendered in a stark, horror-cartoon style with bold outlines and high-contrast shadows. At the center of the frame, the body of a monstrous crimson symbiote violently contorts. Its form, a viscous, shimmering crimson substance, constantly twists and melts into nightmarish shapes, its surface bubbling like a thick, boiling liquid. A single long fluorescent light in the low ceiling flickers erratically, flickering between darkness and harsh, cold illumination. In these stroboscopic flashes, one glimpses the creature's grotesque transformations: a gaping, enormous maw filled with jagged, needle-like teeth, gapes intermittently, spurting out sleek, powerful tendrils that whip across the grimy tile floor. Each pulse of light reveals a new, more grotesque outline—one moment a jumble of amorphous flesh, the next a spiky, monstrous form—as the creature's body twitches and reshapes itself with unnatural speed, intensifying the raw, graphic terror and creating an atmosphere of pure, inescapable dread.", + "seed": 12888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_6395_seed_12888_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021796598844230176, + "vlm_logprob": -0.021796598844230176, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.021796598844230176, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03162786364555359 + ], + "opposite_logprob_tokens": [ + -3.906627893447876 + ], + "logprob_reward": -0.03162786364555359, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03162786364555359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5665264129638672, + 0.06008727103471756, + 0.05209798738360405, + 0.020930131897330284, + 0.025548569858074188, + 0.1098279356956482, + 0.04051259160041809, + 0.02719285525381565, + 0.029711972922086716, + 0.06756428629159927 + ], + "credit_map": [ + 0.04428694769740105, + 0.013530203141272068, + 0.2238466739654541, + 0.03331559896469116, + 0.020998692139983177, + 0.18039491772651672, + 0.019073136150836945, + 0.031080210581421852, + 0.006081485189497471, + 0.010282782837748528, + 0.01341787539422512, + 0.005266720429062843, + 0.006111137568950653, + 0.007868478074669838, + 0.005882482975721359, + 0.005176307167857885, + 0.005607753060758114, + 0.007204486057162285, + 0.006254201289266348, + 0.007668042089790106, + 0.004272973630577326, + 0.005913186352699995, + 0.007473888341337442, + 0.007703453302383423, + 0.002101789927110076, + 0.0013600600650534034, + 0.00467830803245306, + 0.0023201850708574057, + 0.0018678437918424606, + 0.002878485480323434, + 0.0026310423854738474, + 0.0030924163293093443, + 0.0030108210630714893, + 0.003431599121540785, + 0.0064597525633871555, + 0.0016572967870160937, + 0.001971589867025614, + 0.0045236204750835896, + 0.0015223637456074357, + 0.0029715250711888075, + 0.013004244305193424, + 0.025988759472966194, + 0.015781641006469727, + 0.007192142773419619, + 0.01424986869096756, + 0.015692995861172676, + 0.008486922830343246, + 0.009431359358131886, + 0.004397066775709391, + 0.0036222271155565977, + 0.004396183881908655, + 0.005343258380889893, + 0.004110081121325493, + 0.006207501981407404, + 0.006525979842990637, + 0.00591028667986393, + 0.003952493891119957, + 0.002773925429210067, + 0.004534659907221794, + 0.0031346012838184834, + 0.002562427194789052, + 0.004698015749454498, + 0.002310981275513768, + 0.0032257498241961002, + 0.0062340134754776955, + 0.004168030805885792, + 0.005360077600926161, + 0.002297627739608288, + 0.0021430342458188534, + 0.004821856506168842, + 0.0025778617709875107, + 0.0021094700787216425, + 0.012602860108017921, + 0.00873291864991188, + 0.012290840968489647, + 0.007077062968164682, + 0.003474801778793335, + 0.011968886479735374, + 0.005314422771334648, + 0.00610248651355505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.03162786364555359 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001955979270860553 + ], + "opposite_logprob_tokens": [ + -6.251955986022949 + ], + "logprob_reward": -0.001955979270860553, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001955979270860553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6779095530509949, + 0.03411637246608734, + 0.06727821379899979, + 0.040304914116859436, + 0.02161039412021637, + 0.028963755816221237, + 0.021854665130376816, + 0.028552770614624023, + 0.03482823446393013, + 0.04458111152052879 + ], + "credit_map": [ + 0.051409680396318436, + 0.01362773310393095, + 0.2742632031440735, + 0.03035818412899971, + 0.023005938157439232, + 0.23640765249729156, + 0.01598239503800869, + 0.03285478428006172, + 0.003531766589730978, + 0.002233760431408882, + 0.008712831884622574, + 0.004205566365271807, + 0.002374876756221056, + 0.005549056455492973, + 0.0032485013362020254, + 0.004260013811290264, + 0.00535459304228425, + 0.005227889399975538, + 0.013125641271471977, + 0.008757157251238823, + 0.003906517755240202, + 0.011445541866123676, + 0.00935311708599329, + 0.010107754729688168, + 0.0047407615929841995, + 0.0018187332898378372, + 0.008931304328143597, + 0.0032449865248054266, + 0.0017763637006282806, + 0.0113284382969141, + 0.0026639695279300213, + 0.005800357088446617, + 0.0031560827046632767, + 0.0023299625609070063, + 0.005440013017505407, + 0.002016142476350069, + 0.0014931847108528018, + 0.003168778959661722, + 0.0016716360114514828, + 0.0023345942609012127, + 0.003401783760637045, + 0.004855233710259199, + 0.004252889659255743, + 0.0017675362760201097, + 0.004014349542558193, + 0.004562957677990198, + 0.002868437906727195, + 0.0032405657693743706, + 0.0034343122970312834, + 0.0015005605528131127, + 0.0024952366948127747, + 0.0027923977468162775, + 0.002133352914825082, + 0.00286145880818367, + 0.0037673229817301035, + 0.0028700237162411213, + 0.004415564704686403, + 0.0021290022414177656, + 0.00410187803208828, + 0.004303034860640764, + 0.002668109955266118, + 0.005241209641098976, + 0.0027374131605029106, + 0.0029565587174147367, + 0.005740373395383358, + 0.0062363361939787865, + 0.004888454452157021, + 0.003232682356610894, + 0.0034163091331720352, + 0.004801747854799032, + 0.002460917690768838, + 0.004051414784044027, + 0.005836359225213528, + 0.004093492403626442, + 0.007358396891504526, + 0.005450512282550335, + 0.0026189463678747416, + 0.006728514097630978, + 0.004282784182578325, + 0.008212106302380562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.001955979270860553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866204410791397 + ], + "opposite_logprob_tokens": [ + -3.048661947250366 + ], + "logprob_reward": -0.04866204410791397, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866204410791397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36201196908950806, + 0.08074460923671722, + 0.26037463545799255, + 0.06403835117816925, + 0.019275160506367683, + 0.05039758235216141, + 0.03420848771929741, + 0.03448738530278206, + 0.03933237865567207, + 0.055129438638687134 + ], + "credit_map": [ + 0.028394922614097595, + 0.01217623334378004, + 0.13444599509239197, + 0.025440603494644165, + 0.015580098144710064, + 0.11118672788143158, + 0.015987368300557137, + 0.018800007179379463, + 0.008099263533949852, + 0.01285285223275423, + 0.017119931057095528, + 0.007735744584351778, + 0.007135087624192238, + 0.010103192180395126, + 0.007245549466460943, + 0.010452987626194954, + 0.02568059042096138, + 0.011955908499658108, + 0.041227906942367554, + 0.03202018886804581, + 0.01949135959148407, + 0.048683129251003265, + 0.05827077478170395, + 0.023044776171445847, + 0.007236669771373272, + 0.004997088573873043, + 0.008011353202164173, + 0.006134531926363707, + 0.004054693505167961, + 0.010904891416430473, + 0.008489985018968582, + 0.014209137298166752, + 0.002968844724819064, + 0.0014096254017204046, + 0.002627280307933688, + 0.0028499008622020483, + 0.0015558538725599647, + 0.0021383105777204037, + 0.002331128343939781, + 0.0033942176960408688, + 0.005491213407367468, + 0.009888255037367344, + 0.007648210506886244, + 0.004370483569800854, + 0.005339660216122866, + 0.006664687301963568, + 0.005654903594404459, + 0.005340165458619595, + 0.003323410404846072, + 0.003588891588151455, + 0.0038295157719403505, + 0.003957364242523909, + 0.004074722062796354, + 0.004647025838494301, + 0.006467326078563929, + 0.004320232197642326, + 0.005718845408409834, + 0.002188110025599599, + 0.006858654320240021, + 0.004078906495124102, + 0.0033009243197739124, + 0.005143479909747839, + 0.0033956451807171106, + 0.003802819177508354, + 0.007330439984798431, + 0.0032959957607090473, + 0.0067186132073402405, + 0.004391422029584646, + 0.0037195852492004633, + 0.005137328058481216, + 0.004744274076074362, + 0.003994717728346586, + 0.00881358701735735, + 0.004225620999932289, + 0.009201317094266415, + 0.0047346181236207485, + 0.005907958839088678, + 0.00778973288834095, + 0.007157229818403721, + 0.007299376651644707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04866204410791397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634689211845398 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.02634689211845398, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634689211845398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2375335544347763, + 0.12471063435077667, + 0.14133243262767792, + 0.16287311911582947, + 0.0519380196928978, + 0.10629289597272873, + 0.03650674596428871, + 0.03147851675748825, + 0.05576445534825325, + 0.051569558680057526 + ], + "credit_map": [ + 0.02650575526058674, + 0.020587550476193428, + 0.04280005767941475, + 0.03909693658351898, + 0.024845609441399574, + 0.029803842306137085, + 0.02779124677181244, + 0.02610255777835846, + 0.014128862880170345, + 0.01886497251689434, + 0.02306896448135376, + 0.009769133292138577, + 0.011035828851163387, + 0.021439088508486748, + 0.012422342784702778, + 0.01398144569247961, + 0.012349310331046581, + 0.01986142434179783, + 0.029884932562708855, + 0.014595715329051018, + 0.009794758632779121, + 0.02704007364809513, + 0.012152446433901787, + 0.015653779730200768, + 0.017698166891932487, + 0.008386137895286083, + 0.028289420530200005, + 0.020220410078763962, + 0.010613560676574707, + 0.024721655994653702, + 0.018954919651150703, + 0.03398886322975159, + 0.0066218506544828415, + 0.004620632156729698, + 0.007438573986291885, + 0.004481734707951546, + 0.0041922470554709435, + 0.006809753831475973, + 0.005294145550578833, + 0.012479082681238651, + 0.018302302807569504, + 0.00884915515780449, + 0.020268261432647705, + 0.007889566011726856, + 0.014540835283696651, + 0.013919898308813572, + 0.009957268834114075, + 0.012565606273710728, + 0.00503547303378582, + 0.003504726802930236, + 0.0048259529285132885, + 0.004115004558116198, + 0.003328045830130577, + 0.005021289456635714, + 0.006547820754349232, + 0.0041284323669970036, + 0.005731784272938967, + 0.003108508186414838, + 0.004730472341179848, + 0.002821882488206029, + 0.0029848101548850536, + 0.005921488627791405, + 0.0027407314628362656, + 0.0034388387575745583, + 0.009550967253744602, + 0.005764387547969818, + 0.0074206143617630005, + 0.006987659726291895, + 0.006946909707039595, + 0.0066955918446183205, + 0.0067842514254152775, + 0.005614074878394604, + 0.005373325198888779, + 0.004571025725454092, + 0.013787351548671722, + 0.003298260271549225, + 0.0027984813787043095, + 0.013620720244944096, + 0.0030974149703979492, + 0.005022982135415077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.02634689211845398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000390215078368783 + ], + "opposite_logprob_tokens": [ + -7.87539005279541 + ], + "logprob_reward": -0.000390215078368783, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.000390215078368783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42662981152534485, + 0.09823399782180786, + 0.16569629311561584, + 0.06307228654623032, + 0.03953336179256439, + 0.0463155098259449, + 0.028721492737531662, + 0.06253817677497864, + 0.04474072903394699, + 0.024518365040421486 + ], + "credit_map": [ + 0.03968445584177971, + 0.01982932910323143, + 0.1400262862443924, + 0.039365120232105255, + 0.025267869234085083, + 0.10118581354618073, + 0.02804291434586048, + 0.033228009939193726, + 0.011104440316557884, + 0.013889587484300137, + 0.028197795152664185, + 0.0058008176274597645, + 0.008453873917460442, + 0.014570525847375393, + 0.00850688386708498, + 0.007710072677582502, + 0.01427621953189373, + 0.012909228913486004, + 0.030620655044913292, + 0.02008133940398693, + 0.009510097093880177, + 0.025838341563940048, + 0.01728602685034275, + 0.03517437353730202, + 0.007085573393851519, + 0.0026042263489216566, + 0.011380059644579887, + 0.004832917358726263, + 0.0027498682029545307, + 0.009753936901688576, + 0.007898111827671528, + 0.0167675893753767, + 0.007182059343904257, + 0.003342870157212019, + 0.0073080300353467464, + 0.003833731636404991, + 0.002546757459640503, + 0.004642334301024675, + 0.0036162605974823236, + 0.007061315700411797, + 0.006499190349131823, + 0.003866512794047594, + 0.008781428448855877, + 0.003503864398226142, + 0.004612749442458153, + 0.006692791823297739, + 0.005796486511826515, + 0.006562486756592989, + 0.0046772160567343235, + 0.0021226468961685896, + 0.005030793137848377, + 0.0029690403025597334, + 0.002077721757814288, + 0.004500361625105143, + 0.003995517734438181, + 0.003348194994032383, + 0.00865026842802763, + 0.0045950827188789845, + 0.006467293947935104, + 0.009677167050540447, + 0.006282588467001915, + 0.011289611458778381, + 0.0072520216926932335, + 0.008324140682816505, + 0.0062416428700089455, + 0.0030967642087489367, + 0.00743704242631793, + 0.005751516669988632, + 0.00516019482165575, + 0.004991771187633276, + 0.005261995363980532, + 0.006799801252782345, + 0.003187673632055521, + 0.0020042143296450377, + 0.003425591392442584, + 0.002546678762882948, + 0.003213543677702546, + 0.003075594548135996, + 0.0028649207670241594, + 0.004200148396193981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.000390215078368783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures a scene of pure terror within a dark, narrow corridor, rendered in a stark, horror-cartoon style with bold outlines and high-contrast shadows. At the center of the frame, the body of a monstrous crimson symbiote violently contorts. Its form, a viscous, shimmering crimson substance, constantly twists and melts into nightmarish shapes, its surface bubbling like a thick, boiling liquid. A single long fluorescent light in the low ceiling flickers erratically, flickering between darkness and harsh, cold illumination. In these stroboscopic flashes, one glimpses the creature's grotesque transformations: a gaping, enormous maw filled with jagged, needle-like teeth, gapes intermittently, spurting out sleek, powerful tendrils that whip across the grimy tile floor. Each pulse of light reveals a new, more grotesque outline—one moment a jumble of amorphous flesh, the next a spiky, monstrous form—as the creature's body twitches and reshapes itself with unnatural speed, intensifying the raw, graphic terror and creating an atmosphere of pure, inescapable dread.", + "seed": 12889, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_6395_seed_12889_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1753489183145575, + "vlm_logprob": -0.1753489183145575, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.1753489183145575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5323230028152466 + ], + "opposite_logprob_tokens": [ + -0.9073230028152466 + ], + "logprob_reward": -0.5323230028152466, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5323230028152466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5353144407272339, + 0.043807290494441986, + 0.026135727763175964, + 0.02746802382171154, + 0.026773296296596527, + 0.03830178454518318, + 0.02848970703780651, + 0.15942862629890442, + 0.07774627208709717, + 0.03653484955430031 + ], + "credit_map": [ + 0.06565916538238525, + 0.056244079023599625, + 0.09339440613985062, + 0.05694134905934334, + 0.04470432549715042, + 0.04889893904328346, + 0.06814131140708923, + 0.10133081674575806, + 0.006378178019076586, + 0.0036708954721689224, + 0.004970231559127569, + 0.004846873227506876, + 0.004368784371763468, + 0.006187791004776955, + 0.0036697792820632458, + 0.00971475150436163, + 0.002979682292789221, + 0.0038063088431954384, + 0.002537989057600498, + 0.0028568352572619915, + 0.0029811717104166746, + 0.003992107696831226, + 0.0022539980709552765, + 0.004727632738649845, + 0.004908445756882429, + 0.002993798814713955, + 0.0025107460096478462, + 0.0024172295816242695, + 0.005268097389489412, + 0.004074655473232269, + 0.002490642247721553, + 0.002804405987262726, + 0.004772598389536142, + 0.0021264844108372927, + 0.0026386675890535116, + 0.004468883387744427, + 0.0020331747364252806, + 0.003663207171484828, + 0.003921558149158955, + 0.0031487212982028723, + 0.007970492355525494, + 0.007155153434723616, + 0.0033439388498663902, + 0.0032905547413975, + 0.004492462147027254, + 0.005184298846870661, + 0.0035896161571145058, + 0.003275262890383601, + 0.003783388063311577, + 0.003426211653277278, + 0.0024806195870041847, + 0.0025439506862312555, + 0.0043381513096392155, + 0.003981300164014101, + 0.003909436520189047, + 0.004026647191494703, + 0.025185061618685722, + 0.02768256701529026, + 0.022597702220082283, + 0.019301503896713257, + 0.011629276908934116, + 0.02238517813384533, + 0.016659120097756386, + 0.013988194055855274, + 0.01160106435418129, + 0.014516966417431831, + 0.008035829290747643, + 0.0077805593609809875, + 0.011171949096024036, + 0.009457207284867764, + 0.0067451936192810535, + 0.008437497541308403, + 0.005013869609683752, + 0.00624102121219039, + 0.0031137599144130945, + 0.004229388199746609, + 0.006141668185591698, + 0.004495029337704182, + 0.0038418895564973354, + 0.003458223771303892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.5323230028152466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032130067702382803 + ], + "opposite_logprob_tokens": [ + -5.753212928771973 + ], + "logprob_reward": -0.0032130067702382803, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032130067702382803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5483959317207336, + 0.06796468049287796, + 0.045992184430360794, + 0.031541354954242706, + 0.045239366590976715, + 0.051346555352211, + 0.020463889464735985, + 0.12394407391548157, + 0.039610590785741806, + 0.02550135925412178 + ], + "credit_map": [ + 0.06779897958040237, + 0.059564389288425446, + 0.09744550287723541, + 0.0563092902302742, + 0.04668319597840309, + 0.049842484295368195, + 0.06816596537828445, + 0.10258609801530838, + 0.010364986956119537, + 0.00513940816745162, + 0.008695748634636402, + 0.007619666401296854, + 0.006230705883353949, + 0.008992007933557034, + 0.00650275032967329, + 0.014419407583773136, + 0.005400687921792269, + 0.005176014732569456, + 0.003961249720305204, + 0.004484339151531458, + 0.00462502334266901, + 0.00731883617118001, + 0.0042020161636173725, + 0.010824019089341164, + 0.006061050575226545, + 0.0034505671355873346, + 0.0031106227543205023, + 0.0025977410841733217, + 0.004286315757781267, + 0.004632939584553242, + 0.0030571024399250746, + 0.004345014691352844, + 0.009991784580051899, + 0.005268004257231951, + 0.0036171076353639364, + 0.005322718061506748, + 0.00399312237277627, + 0.006640213076025248, + 0.004550370387732983, + 0.005856044124811888, + 0.012075512669980526, + 0.010422506369650364, + 0.005108962766826153, + 0.004196427762508392, + 0.004538962617516518, + 0.00684646051377058, + 0.004049812443554401, + 0.004107911139726639, + 0.0039939614944159985, + 0.0031622869428247213, + 0.0019637413788586855, + 0.0017907802248373628, + 0.0025960272178053856, + 0.0031074383296072483, + 0.0019131280714645982, + 0.0019365273183211684, + 0.01591121032834053, + 0.019314460456371307, + 0.016173094511032104, + 0.01875990442931652, + 0.0097767049446702, + 0.01580868288874626, + 0.017477963119745255, + 0.01072204951196909, + 0.0058157360181212425, + 0.007946941070258617, + 0.003427308052778244, + 0.0040819402784109116, + 0.0061882915906608105, + 0.00493445573374629, + 0.0034742425195872784, + 0.0037416766863316298, + 0.003623099299147725, + 0.003474426455795765, + 0.002403718652203679, + 0.0036895747762173414, + 0.004197472240775824, + 0.003117513842880726, + 0.0028594930190593004, + 0.00213606096804142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0032130067702382803 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039176285266876 + ], + "opposite_logprob_tokens": [ + -1.9103918075561523 + ], + "logprob_reward": -0.16039176285266876, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039176285266876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35566267371177673, + 0.06333889812231064, + 0.2189687043428421, + 0.05545983836054802, + 0.09924357384443283, + 0.07155594229698181, + 0.045440491288900375, + 0.029702916741371155, + 0.023945748805999756, + 0.03668120130896568 + ], + "credit_map": [ + 0.039964012801647186, + 0.0379423052072525, + 0.06599747389554977, + 0.039629820734262466, + 0.03128823637962341, + 0.031108619645237923, + 0.049363620579242706, + 0.060368604958057404, + 0.009220097213983536, + 0.009709887206554413, + 0.009600910358130932, + 0.007192769553512335, + 0.007197629660367966, + 0.007432896178215742, + 0.006745276506990194, + 0.006239434238523245, + 0.03136008605360985, + 0.020276349037885666, + 0.022884685546159744, + 0.01978076621890068, + 0.03267253562808037, + 0.02701163850724697, + 0.020818224176764488, + 0.04416445270180702, + 0.011067686602473259, + 0.005410478916019201, + 0.005034455098211765, + 0.003958592656999826, + 0.010778479278087616, + 0.007803641259670258, + 0.004803215619176626, + 0.006603289395570755, + 0.016275834292173386, + 0.012903101742267609, + 0.0066858879290521145, + 0.014847773127257824, + 0.010254859924316406, + 0.012921527028083801, + 0.012028073891997337, + 0.013326515443623066, + 0.009573487564921379, + 0.009949917905032635, + 0.007550176698714495, + 0.007402217015624046, + 0.007218255661427975, + 0.010150107555091381, + 0.01066555269062519, + 0.009046235121786594, + 0.005258695688098669, + 0.007051215972751379, + 0.004880554508417845, + 0.005142595618963242, + 0.00539819709956646, + 0.005012832581996918, + 0.0064841583371162415, + 0.006212243810296059, + 0.003731394885107875, + 0.003592799184843898, + 0.0031050050165504217, + 0.004045093897730112, + 0.0035780516918748617, + 0.0037834399845451117, + 0.0042291060090065, + 0.003638028632849455, + 0.0039008718449622393, + 0.0030798150692135096, + 0.003860330209136009, + 0.0028022690676152706, + 0.002447684993967414, + 0.002762943971902132, + 0.0021472135558724403, + 0.00294462195597589, + 0.00687616178765893, + 0.006284961476922035, + 0.002910342998802662, + 0.003928008023649454, + 0.004891871474683285, + 0.00489046610891819, + 0.00392264686524868, + 0.0029767449013888836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.16039176285266876 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989253997802734 + ], + "opposite_logprob_tokens": [ + -1.8048925399780273 + ], + "logprob_reward": -0.17989253997802734, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989253997802734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12363258749246597, + 0.15187954902648926, + 0.08518703281879425, + 0.10677814483642578, + 0.0721488744020462, + 0.0641133189201355, + 0.06790712475776672, + 0.2026607245206833, + 0.06388290226459503, + 0.06180978938937187 + ], + "credit_map": [ + 0.016583042219281197, + 0.01373112853616476, + 0.021250758320093155, + 0.014828776009380817, + 0.012655550613999367, + 0.012928742915391922, + 0.0140733877196908, + 0.017581194639205933, + 0.019813628867268562, + 0.011531193740665913, + 0.01960953325033188, + 0.015542144887149334, + 0.012704561464488506, + 0.022796357050538063, + 0.011411178857088089, + 0.038470957428216934, + 0.018207384273409843, + 0.010167237371206284, + 0.00836376566439867, + 0.005444124341011047, + 0.009888933971524239, + 0.013060448691248894, + 0.005556392949074507, + 0.014498751610517502, + 0.0155147360637784, + 0.014893699437379837, + 0.010250658728182316, + 0.01159837655723095, + 0.011449545621871948, + 0.015852201730012894, + 0.00992751307785511, + 0.017291409894824028, + 0.013171963393688202, + 0.006464072503149509, + 0.00609356677159667, + 0.011550422757863998, + 0.006581959780305624, + 0.009434064850211143, + 0.01088901050388813, + 0.007963815703988075, + 0.012082177214324474, + 0.01155188400298357, + 0.00723628792911768, + 0.005140145309269428, + 0.006872751750051975, + 0.008462261408567429, + 0.005846929736435413, + 0.006920878775417805, + 0.008308636024594307, + 0.0069800736382603645, + 0.0060314275324344635, + 0.0060982853174209595, + 0.010776971466839314, + 0.00920562632381916, + 0.009447833523154259, + 0.011058269999921322, + 0.032196518033742905, + 0.03653346747159958, + 0.025069965049624443, + 0.024030610918998718, + 0.017449907958507538, + 0.0277777798473835, + 0.02190341427922249, + 0.01769905909895897, + 0.008349711075425148, + 0.011943759396672249, + 0.007533412426710129, + 0.006585391703993082, + 0.00868114735931158, + 0.0076369186863303185, + 0.00579513143748045, + 0.007357432506978512, + 0.006951592396944761, + 0.01425959262996912, + 0.004767530597746372, + 0.006823527626693249, + 0.0095231207087636, + 0.007941415533423424, + 0.004834756720811129, + 0.006708251778036356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.17989253997802734 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009242791566066444 + ], + "opposite_logprob_tokens": [ + -7.000924110412598 + ], + "logprob_reward": -0.0009242791566066444, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009242791566066444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5970053672790527, + 0.0565548799932003, + 0.06028394401073456, + 0.03422879800200462, + 0.05267148092389107, + 0.023790989071130753, + 0.04094037413597107, + 0.07976251095533371, + 0.022822484374046326, + 0.03193926066160202 + ], + "credit_map": [ + 0.07602250576019287, + 0.06195595860481262, + 0.10008370131254196, + 0.06108315289020538, + 0.048802681267261505, + 0.05639330670237541, + 0.07316190749406815, + 0.11950211971998215, + 0.006956745870411396, + 0.003907690290361643, + 0.0055096871219575405, + 0.009067436680197716, + 0.006894449703395367, + 0.006614001467823982, + 0.007351889740675688, + 0.010252974927425385, + 0.010645749047398567, + 0.005290617235004902, + 0.006749688182026148, + 0.00541231082752347, + 0.005113754887133837, + 0.009656588546931744, + 0.004816313739866018, + 0.012598917819559574, + 0.0062418291345238686, + 0.004806300159543753, + 0.002705691382288933, + 0.0035212726797908545, + 0.005525586195290089, + 0.004304945003241301, + 0.0030010282061994076, + 0.0041221450082957745, + 0.009378763847053051, + 0.005183031782507896, + 0.004539380315691233, + 0.0074347639456391335, + 0.005870024207979441, + 0.007001778576523066, + 0.006877741776406765, + 0.0063859885558485985, + 0.003296569688245654, + 0.0028423888143152, + 0.002360563725233078, + 0.0032089988235384226, + 0.0024760395754128695, + 0.0027522221207618713, + 0.0039012401830404997, + 0.0029529633466154337, + 0.005170539487153292, + 0.0042273299768567085, + 0.005361846182495356, + 0.004540619906038046, + 0.005646337289363146, + 0.004826318472623825, + 0.005866092164069414, + 0.00530128413811326, + 0.013911937363445759, + 0.017245585098862648, + 0.009503495879471302, + 0.008101650513708591, + 0.006419931538403034, + 0.011259996332228184, + 0.007010897155851126, + 0.006309007294476032, + 0.00338081456720829, + 0.0034336945973336697, + 0.002839820459485054, + 0.002535350387915969, + 0.002953633200377226, + 0.00270740850828588, + 0.00199539284221828, + 0.0029763670172542334, + 0.004199070390313864, + 0.0048968056216835976, + 0.0028581968508660793, + 0.0044438824988901615, + 0.005125479307025671, + 0.003718966618180275, + 0.003845724742859602, + 0.002851132769137621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0009242791566066444 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures a scene of pure terror within a dark, narrow corridor, rendered in a stark, horror-cartoon style with bold outlines and high-contrast shadows. At the center of the frame, the body of a monstrous crimson symbiote violently contorts. Its form, a viscous, shimmering crimson substance, constantly twists and melts into nightmarish shapes, its surface bubbling like a thick, boiling liquid. A single long fluorescent light in the low ceiling flickers erratically, flickering between darkness and harsh, cold illumination. In these stroboscopic flashes, one glimpses the creature's grotesque transformations: a gaping, enormous maw filled with jagged, needle-like teeth, gapes intermittently, spurting out sleek, powerful tendrils that whip across the grimy tile floor. Each pulse of light reveals a new, more grotesque outline—one moment a jumble of amorphous flesh, the next a spiky, monstrous form—as the creature's body twitches and reshapes itself with unnatural speed, intensifying the raw, graphic terror and creating an atmosphere of pure, inescapable dread.", + "seed": 12890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_6395_seed_12890_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0426623938488774, + "vlm_logprob": -0.0426623938488774, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.0426623938488774, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0609457902610302 + ], + "opposite_logprob_tokens": [ + -2.935945749282837 + ], + "logprob_reward": -0.0609457902610302, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0609457902610302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16445785760879517, + 0.07346072047948837, + 0.07063129544258118, + 0.06183486431837082, + 0.049178339540958405, + 0.10415124148130417, + 0.05096890404820442, + 0.044779080897569656, + 0.22658616304397583, + 0.153951495885849 + ], + "credit_map": [ + 0.033899936825037, + 0.020327091217041016, + 0.015571067109704018, + 0.01343106385320425, + 0.018352393060922623, + 0.023666417226195335, + 0.019617749378085136, + 0.01959214359521866, + 0.008341168984770775, + 0.00826150830835104, + 0.012169144116342068, + 0.007886175997555256, + 0.01055434811860323, + 0.008699186146259308, + 0.00819054339081049, + 0.009358654730021954, + 0.009300864301621914, + 0.01091746799647808, + 0.007639054208993912, + 0.00706909317523241, + 0.009026481769979, + 0.01045713946223259, + 0.0070048244670033455, + 0.009216377511620522, + 0.008527572266757488, + 0.013789670541882515, + 0.006619805004447699, + 0.0073194801807403564, + 0.005936100613325834, + 0.0053995936177670956, + 0.009540063329041004, + 0.004702581558376551, + 0.005996732972562313, + 0.007140560541301966, + 0.006324941758066416, + 0.00706197414547205, + 0.004465581849217415, + 0.0039935242384672165, + 0.005965756718069315, + 0.008229272440075874, + 0.01106816902756691, + 0.01085667870938778, + 0.008974607102572918, + 0.015981215983629227, + 0.020598571747541428, + 0.011781043373048306, + 0.014713146723806858, + 0.010177824646234512, + 0.005366258788853884, + 0.0038554631173610687, + 0.003275770926848054, + 0.007824237458407879, + 0.006329795345664024, + 0.006236711051315069, + 0.012046710588037968, + 0.006033963058143854, + 0.004672112409025431, + 0.004749268759042025, + 0.004811463411897421, + 0.005445130635052919, + 0.009426291100680828, + 0.005149493459612131, + 0.004278670996427536, + 0.006246653851121664, + 0.03847524896264076, + 0.022294573485851288, + 0.03290964663028717, + 0.03734668344259262, + 0.021973133087158203, + 0.024884885177016258, + 0.024818697944283485, + 0.02388332225382328, + 0.030896857380867004, + 0.01405349001288414, + 0.01993165723979473, + 0.013746863231062889, + 0.016564054414629936, + 0.018539993092417717, + 0.025600912049412727, + 0.014617674052715302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0609457902610302 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002832331694662571 + ], + "opposite_logprob_tokens": [ + -5.877832412719727 + ], + "logprob_reward": -0.002832331694662571, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002832331694662571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14529633522033691, + 0.1047978475689888, + 0.07194314152002335, + 0.05377858877182007, + 0.05969934165477753, + 0.11308418959379196, + 0.06878901273012161, + 0.09691541641950607, + 0.04133358225226402, + 0.2443625032901764 + ], + "credit_map": [ + 0.012611759826540947, + 0.010143773630261421, + 0.016249872744083405, + 0.016660474240779877, + 0.02447340078651905, + 0.016691405326128006, + 0.02293766476213932, + 0.02552798017859459, + 0.009488201700150967, + 0.011273473501205444, + 0.012344335205852985, + 0.013513127341866493, + 0.021303700283169746, + 0.014836116693913937, + 0.010425386019051075, + 0.01161351427435875, + 0.008756318129599094, + 0.007936506532132626, + 0.009575968608260155, + 0.007012896705418825, + 0.010594429448246956, + 0.009060886688530445, + 0.009266207925975323, + 0.009739929810166359, + 0.0059807379730045795, + 0.0042318678461015224, + 0.006581265013664961, + 0.010443130508065224, + 0.004450355190783739, + 0.0033046307507902384, + 0.010427228175103664, + 0.008359372615814209, + 0.007163291797041893, + 0.00940136518329382, + 0.00591125525534153, + 0.010846924968063831, + 0.006099711637943983, + 0.004765789955854416, + 0.007008467800915241, + 0.008502539247274399, + 0.010266847908496857, + 0.00856273528188467, + 0.010551555082201958, + 0.017124515026807785, + 0.02651345729827881, + 0.011642436496913433, + 0.014887875877320766, + 0.013534776866436005, + 0.006066363770514727, + 0.004816829226911068, + 0.005067773163318634, + 0.008176873438060284, + 0.009267115965485573, + 0.007608846295624971, + 0.019200369715690613, + 0.00858483836054802, + 0.01242627389729023, + 0.011427165940403938, + 0.01621037907898426, + 0.014746533706784248, + 0.00968234334141016, + 0.010951895266771317, + 0.008917802013456821, + 0.012553020380437374, + 0.005918808747082949, + 0.003874968970194459, + 0.0060793934389948845, + 0.005810869857668877, + 0.006161066237837076, + 0.0041387006640434265, + 0.004500293638557196, + 0.004849477671086788, + 0.05494456738233566, + 0.019499661400914192, + 0.03336768597364426, + 0.0209820494055748, + 0.021770598366856575, + 0.029742904007434845, + 0.04239498823881149, + 0.02166004851460457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.002832331694662571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900486141443253 + ], + "opposite_logprob_tokens": [ + -2.5790047645568848 + ], + "logprob_reward": -0.07900486141443253, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900486141443253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15267032384872437, + 0.08909408748149872, + 0.07229582220315933, + 0.09107650816440582, + 0.033626988530159, + 0.07636579126119614, + 0.07857886701822281, + 0.13270701467990875, + 0.04840267822146416, + 0.2251819372177124 + ], + "credit_map": [ + 0.021129243075847626, + 0.013637312687933445, + 0.015972483903169632, + 0.0168431606143713, + 0.021937664598226547, + 0.01917259208858013, + 0.020799407735466957, + 0.02317844331264496, + 0.01053275354206562, + 0.011272993870079517, + 0.012731376104056835, + 0.009074882604181767, + 0.015195687301456928, + 0.013143894262611866, + 0.0088880006223917, + 0.00825449451804161, + 0.007929366081953049, + 0.009877325035631657, + 0.009726695716381073, + 0.0066238935105502605, + 0.009640766307711601, + 0.00982301402837038, + 0.0057472786866128445, + 0.012927471660077572, + 0.01454604510217905, + 0.019543997943401337, + 0.010633069090545177, + 0.01032605767250061, + 0.007506514433771372, + 0.006729997228831053, + 0.013536941260099411, + 0.008253874257206917, + 0.003242292907088995, + 0.003753108438104391, + 0.003919564187526703, + 0.004456742666661739, + 0.0036027338355779648, + 0.0033683020155876875, + 0.00481633422896266, + 0.006467905361205339, + 0.012232281267642975, + 0.014399060979485512, + 0.008515727706253529, + 0.008127045817673206, + 0.00989792961627245, + 0.008855585940182209, + 0.00677095353603363, + 0.007567200809717178, + 0.008525599725544453, + 0.005175527650862932, + 0.004699135664850473, + 0.011185248382389545, + 0.009577067568898201, + 0.009270037524402142, + 0.02087019383907318, + 0.009276052005589008, + 0.013707095757126808, + 0.01194978877902031, + 0.014039193280041218, + 0.010572611354291439, + 0.03082272782921791, + 0.013337026350200176, + 0.017458129674196243, + 0.02082042582333088, + 0.007666959892958403, + 0.00542093813419342, + 0.005968524143099785, + 0.004546045791357756, + 0.008866498246788979, + 0.0054713706485927105, + 0.0062549784779548645, + 0.004207358695566654, + 0.04881894960999489, + 0.01972883939743042, + 0.02891198918223381, + 0.018960535526275635, + 0.0259951651096344, + 0.026382479816675186, + 0.03767118230462074, + 0.01871277391910553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.07900486141443253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002926617860794 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.07002926617860794, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002926617860794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19444073736667633, + 0.07821396738290787, + 0.10811254382133484, + 0.09106700867414474, + 0.21934860944747925, + 0.16981764137744904, + 0.034189365804195404, + 0.04607873409986496, + 0.03532823175191879, + 0.02340320311486721 + ], + "credit_map": [ + 0.027735481038689613, + 0.03694624826312065, + 0.019223198294639587, + 0.017447801306843758, + 0.023403363302350044, + 0.02339959889650345, + 0.01790275052189827, + 0.028382284566760063, + 0.008838736452162266, + 0.0110210832208395, + 0.011142797768115997, + 0.009922510012984276, + 0.010836266912519932, + 0.011768242344260216, + 0.0064150444231927395, + 0.00826928112655878, + 0.011283677071332932, + 0.010977820493280888, + 0.009677558206021786, + 0.012903383001685143, + 0.016235988587141037, + 0.01735837757587433, + 0.017225082963705063, + 0.012450654059648514, + 0.012111702002584934, + 0.013025227002799511, + 0.014867585152387619, + 0.012002844363451004, + 0.008215177804231644, + 0.007922365330159664, + 0.012740978971123695, + 0.010181121528148651, + 0.039663538336753845, + 0.035938046872615814, + 0.023525027558207512, + 0.025082195177674294, + 0.022317474707961082, + 0.017840659245848656, + 0.02453657239675522, + 0.03044508583843708, + 0.014996306970715523, + 0.01250938419252634, + 0.02166184037923813, + 0.021035797894001007, + 0.037970539182424545, + 0.02162482775747776, + 0.023907244205474854, + 0.01611168310046196, + 0.004290113225579262, + 0.002567974152043462, + 0.0029708317015320063, + 0.004973548464477062, + 0.0048474399372935295, + 0.0036901128478348255, + 0.006744169630110264, + 0.00410517817363143, + 0.006363983266055584, + 0.005252229515463114, + 0.006356573663651943, + 0.006215435918420553, + 0.0059003387577831745, + 0.0052114129066467285, + 0.0035320832394063473, + 0.007246678229421377, + 0.004015256185084581, + 0.0028767527546733618, + 0.0039046125020831823, + 0.006393779069185257, + 0.005712791811674833, + 0.004144495818763971, + 0.00452668871730566, + 0.0037538541946560144, + 0.00448444951325655, + 0.0021225239615887403, + 0.003256401978433132, + 0.002024977235123515, + 0.002175366273149848, + 0.003049769438803196, + 0.0040587130934000015, + 0.0022310034837573767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.07002926617860794 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004997196956537664 + ], + "opposite_logprob_tokens": [ + -7.625499725341797 + ], + "logprob_reward": -0.0004997196956537664, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004997196956537664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12079716473817825, + 0.11752396076917648, + 0.12011440098285675, + 0.07883352041244507, + 0.040152981877326965, + 0.17785131931304932, + 0.04943006485700607, + 0.19713155925273895, + 0.03590773046016693, + 0.06225735694169998 + ], + "credit_map": [ + 0.012046129442751408, + 0.013037845492362976, + 0.009856708347797394, + 0.015658706426620483, + 0.019938983023166656, + 0.015370993874967098, + 0.014756506308913231, + 0.020131288096308708, + 0.010960236191749573, + 0.011877443641424179, + 0.012842762283980846, + 0.01369287259876728, + 0.023210305720567703, + 0.015666134655475616, + 0.014139613136649132, + 0.01513458788394928, + 0.014262893237173557, + 0.014233493246138096, + 0.015876274555921555, + 0.011724687181413174, + 0.01987769827246666, + 0.015100563876330853, + 0.016172830015420914, + 0.012865962460637093, + 0.011522512882947922, + 0.01041479129344225, + 0.012215481139719486, + 0.011456104926764965, + 0.007873115129768848, + 0.006608182098716497, + 0.008151155896484852, + 0.010592170059680939, + 0.004133650567382574, + 0.002616596408188343, + 0.007620758377015591, + 0.006981855724006891, + 0.002727503189817071, + 0.003433719277381897, + 0.004750683903694153, + 0.00788821093738079, + 0.013731141574680805, + 0.013572829775512218, + 0.020708855241537094, + 0.02038205787539482, + 0.04763050749897957, + 0.01962122693657875, + 0.02284901961684227, + 0.019355669617652893, + 0.007821034640073776, + 0.00416887691244483, + 0.004616976249963045, + 0.005045728292316198, + 0.005441483575850725, + 0.0052701132372021675, + 0.011368327774107456, + 0.00569752836599946, + 0.01729602739214897, + 0.01790502667427063, + 0.019710639491677284, + 0.014601239934563637, + 0.051518116146326065, + 0.02113438956439495, + 0.024195589125156403, + 0.030770519748330116, + 0.0052709258161485195, + 0.0034679370000958443, + 0.0036206261720508337, + 0.003895964240655303, + 0.0069220419973134995, + 0.00460326811298728, + 0.004449451807886362, + 0.00367751345038414, + 0.008399518206715584, + 0.006242783740162849, + 0.007294676266610622, + 0.005383369978517294, + 0.00746954046189785, + 0.007699714042246342, + 0.00991102959960699, + 0.009856721386313438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.0004997196956537664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures a scene of pure terror within a dark, narrow corridor, rendered in a stark, horror-cartoon style with bold outlines and high-contrast shadows. At the center of the frame, the body of a monstrous crimson symbiote violently contorts. Its form, a viscous, shimmering crimson substance, constantly twists and melts into nightmarish shapes, its surface bubbling like a thick, boiling liquid. A single long fluorescent light in the low ceiling flickers erratically, flickering between darkness and harsh, cold illumination. In these stroboscopic flashes, one glimpses the creature's grotesque transformations: a gaping, enormous maw filled with jagged, needle-like teeth, gapes intermittently, spurting out sleek, powerful tendrils that whip across the grimy tile floor. Each pulse of light reveals a new, more grotesque outline—one moment a jumble of amorphous flesh, the next a spiky, monstrous form—as the creature's body twitches and reshapes itself with unnatural speed, intensifying the raw, graphic terror and creating an atmosphere of pure, inescapable dread.", + "seed": 12891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_3/index_6395_seed_12891_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06280181243782863, + "vlm_logprob": -0.06280181243782863, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.06280181243782863, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14659790694713593 + ], + "opposite_logprob_tokens": [ + -2.1465978622436523 + ], + "logprob_reward": -0.14659790694713593, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14659790694713593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18731778860092163, + 0.10852333903312683, + 0.04965890944004059, + 0.028451308608055115, + 0.09178335219621658, + 0.05282208323478699, + 0.12940333783626556, + 0.10583174228668213, + 0.1657649427652359, + 0.0804431289434433 + ], + "credit_map": [ + 0.014579139649868011, + 0.008215103298425674, + 0.021388448774814606, + 0.03576608747243881, + 0.021028263494372368, + 0.04020380228757858, + 0.01827307417988777, + 0.027863873168826103, + 0.011488095857203007, + 0.006479782052338123, + 0.013992046937346458, + 0.01776272803544998, + 0.011649036779999733, + 0.014988218434154987, + 0.0155588798224926, + 0.016604559496045113, + 0.007911182008683681, + 0.003918493166565895, + 0.005789716728031635, + 0.008969402872025967, + 0.0046872966922819614, + 0.004726845305413008, + 0.005832365248352289, + 0.007823606953024864, + 0.005056923720985651, + 0.0035692108795046806, + 0.004342310596257448, + 0.00229385937564075, + 0.0026746131479740143, + 0.003758820705115795, + 0.0026916563510894775, + 0.004063913598656654, + 0.011990474537014961, + 0.005204771179705858, + 0.01164971198886633, + 0.007539159618318081, + 0.011214660480618477, + 0.022715147584676743, + 0.011074810288846493, + 0.010394618846476078, + 0.009762786328792572, + 0.005353144370019436, + 0.008115020580589771, + 0.0061742342077195644, + 0.005068570375442505, + 0.0070540085434913635, + 0.006186749320477247, + 0.005107571836560965, + 0.017965611070394516, + 0.010662454180419445, + 0.02225041761994362, + 0.023665722459554672, + 0.009577195160090923, + 0.013661188073456287, + 0.015055832453072071, + 0.016564922407269478, + 0.012830170802772045, + 0.012016850523650646, + 0.015763936564326286, + 0.015205206349492073, + 0.010015678592026234, + 0.01729341596364975, + 0.008674460463225842, + 0.014032023958861828, + 0.015893975272774696, + 0.02846323698759079, + 0.023742949590086937, + 0.025225473567843437, + 0.014507682994008064, + 0.021964335814118385, + 0.014251147396862507, + 0.02171614207327366, + 0.010537759400904179, + 0.006400330923497677, + 0.01393855083733797, + 0.010524162091314793, + 0.005940992850810289, + 0.007288661785423756, + 0.012010120786726475, + 0.013802556321024895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.14659790694713593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001959072658792138 + ], + "opposite_logprob_tokens": [ + -6.251958847045898 + ], + "logprob_reward": -0.001959072658792138, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001959072658792138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22019599378108978, + 0.08414947986602783, + 0.02650628052651882, + 0.062026068568229675, + 0.0849979817867279, + 0.02770787477493286, + 0.084952712059021, + 0.14044289290905, + 0.18189331889152527, + 0.0871274471282959 + ], + "credit_map": [ + 0.015658803284168243, + 0.00671743880957365, + 0.02199835330247879, + 0.0329696387052536, + 0.019725583493709564, + 0.06026468053460121, + 0.02193065918982029, + 0.04093082249164581, + 0.007256653625518084, + 0.0038483852986246347, + 0.011620048433542252, + 0.01640554890036583, + 0.008538994938135147, + 0.010979914106428623, + 0.008503650315105915, + 0.01699628122150898, + 0.003180371131747961, + 0.0015345386927947402, + 0.0022434729617089033, + 0.0037979662884026766, + 0.003859306452795863, + 0.005059436429291964, + 0.0030796562787145376, + 0.0037515326403081417, + 0.011037060059607029, + 0.003738541854545474, + 0.009689203463494778, + 0.005935595836490393, + 0.006252201274037361, + 0.007297946140170097, + 0.006384793668985367, + 0.011690725572407246, + 0.010979242622852325, + 0.004287797957658768, + 0.011020327918231487, + 0.008474307134747505, + 0.009877542965114117, + 0.01985025219619274, + 0.010455982759594917, + 0.010052526369690895, + 0.005577133968472481, + 0.0025423704646527767, + 0.004346154164522886, + 0.003894180292263627, + 0.0025912337005138397, + 0.002201669616624713, + 0.003759087296202779, + 0.00279604597017169, + 0.011861581355333328, + 0.0067585487850010395, + 0.014385330490767956, + 0.014476722106337547, + 0.006291836965829134, + 0.008995014242827892, + 0.00987477507442236, + 0.012308903969824314, + 0.01731945015490055, + 0.01670127548277378, + 0.02152644470334053, + 0.021096331998705864, + 0.013810778968036175, + 0.01918385550379753, + 0.013191940262913704, + 0.017612800002098083, + 0.01838824339210987, + 0.02687060460448265, + 0.023716019466519356, + 0.03035997413098812, + 0.01654861681163311, + 0.026566021144390106, + 0.017739666625857353, + 0.021704182028770447, + 0.012995823286473751, + 0.007366887293756008, + 0.012556340545415878, + 0.009080291725695133, + 0.00521291047334671, + 0.010346509516239166, + 0.01252342201769352, + 0.017045266926288605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.001959072658792138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708091735839844 + ], + "opposite_logprob_tokens": [ + -2.1270809173583984 + ], + "logprob_reward": -0.12708091735839844, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708091735839844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26248690485954285, + 0.19241201877593994, + 0.0848151445388794, + 0.06699664145708084, + 0.027652394026517868, + 0.02996586635708809, + 0.09630591422319412, + 0.09057075530290604, + 0.08423607796430588, + 0.0645582526922226 + ], + "credit_map": [ + 0.023702047765254974, + 0.01786108687520027, + 0.040100179612636566, + 0.05288929119706154, + 0.031134119257330894, + 0.03335336595773697, + 0.03835432976484299, + 0.025092504918575287, + 0.017837608233094215, + 0.007971084676682949, + 0.02230875939130783, + 0.04999871924519539, + 0.01778881624341011, + 0.02585745044052601, + 0.020738432183861732, + 0.029911139979958534, + 0.01216051448136568, + 0.005835722200572491, + 0.011113942600786686, + 0.015376058407127857, + 0.00712611572816968, + 0.009424869902431965, + 0.007792806252837181, + 0.01598510891199112, + 0.009713792242109776, + 0.004360660910606384, + 0.006437244359403849, + 0.005467453505843878, + 0.01042689848691225, + 0.011711831204593182, + 0.0108102485537529, + 0.008068510331213474, + 0.0063405828550457954, + 0.0034885869827121496, + 0.004039700608700514, + 0.0025403390172868967, + 0.0023359814658761024, + 0.003875406226143241, + 0.002318981336429715, + 0.0027128153014928102, + 0.0062905666418373585, + 0.002034958451986313, + 0.004779466427862644, + 0.005893170367926359, + 0.0028200983069837093, + 0.002415748778730631, + 0.002525377320125699, + 0.003206480760127306, + 0.013173836283385754, + 0.00796710979193449, + 0.01652531884610653, + 0.017704283818602562, + 0.006619772873818874, + 0.010184179991483688, + 0.01140770222991705, + 0.012723713181912899, + 0.012153243646025658, + 0.007416355889290571, + 0.011090777814388275, + 0.009862754493951797, + 0.00882982648909092, + 0.01146306935697794, + 0.01914878562092781, + 0.010605943389236927, + 0.007976607419550419, + 0.01642461307346821, + 0.011015839874744415, + 0.012087234295904636, + 0.0078286062926054, + 0.012746665626764297, + 0.005636367481201887, + 0.010520141571760178, + 0.00988633744418621, + 0.005583335645496845, + 0.006157119292765856, + 0.0063269552774727345, + 0.006434776820242405, + 0.010811151005327702, + 0.009734228253364563, + 0.009624357335269451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.12708091735839844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381014309823513 + ], + "opposite_logprob_tokens": [ + -3.2881014347076416 + ], + "logprob_reward": -0.0381014309823513, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381014309823513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17910587787628174, + 0.09821534156799316, + 0.16322633624076843, + 0.024635959416627884, + 0.026882879436016083, + 0.10442489385604858, + 0.1168551817536354, + 0.07349593192338943, + 0.18749959766864777, + 0.025658000260591507 + ], + "credit_map": [ + 0.019355978816747665, + 0.015038677491247654, + 0.023783830925822258, + 0.03282344713807106, + 0.014959868043661118, + 0.025041613727808, + 0.02747223898768425, + 0.02063022181391716, + 0.014146271161735058, + 0.009705689735710621, + 0.014537212438881397, + 0.022161660715937614, + 0.007354250643402338, + 0.009883378632366657, + 0.008592015132308006, + 0.01183486171066761, + 0.025762714445590973, + 0.015382477082312107, + 0.0193291325122118, + 0.03387732431292534, + 0.010858425870537758, + 0.018268484622240067, + 0.01187742780894041, + 0.02787034772336483, + 0.0038181289564818144, + 0.002923141699284315, + 0.002732587978243828, + 0.0018445325549691916, + 0.0026927494909614325, + 0.003560332814231515, + 0.002858058549463749, + 0.004206429235637188, + 0.005322945304214954, + 0.003155916929244995, + 0.004083509091287851, + 0.003615607973188162, + 0.0024929335340857506, + 0.0027983852196484804, + 0.00274611241184175, + 0.002667470369488001, + 0.012479132041335106, + 0.012474904768168926, + 0.012671153992414474, + 0.015822185203433037, + 0.009879771620035172, + 0.014986610040068626, + 0.015471776016056538, + 0.010639366693794727, + 0.01564926467835903, + 0.010051432996988297, + 0.01938643679022789, + 0.021162275224924088, + 0.009324474260210991, + 0.012663445435464382, + 0.013294242322444916, + 0.015323609113693237, + 0.009848017245531082, + 0.008061970584094524, + 0.008767464198172092, + 0.0075675412081182, + 0.008119909092783928, + 0.01019066758453846, + 0.012035621330142021, + 0.008904741145670414, + 0.01902826316654682, + 0.03534918650984764, + 0.025702621787786484, + 0.03010997176170349, + 0.016234656795859337, + 0.023273948580026627, + 0.014554859139025211, + 0.023246092721819878, + 0.003485022811219096, + 0.0022607219871133566, + 0.0025360630825161934, + 0.003069095080718398, + 0.002642212901264429, + 0.003784997621551156, + 0.0045473696663975716, + 0.0033325161784887314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0381014309823513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, narrow corridor with a grimy tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a horror-cartoon style with bold outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a single long fluorescent light in the low ceiling flickering erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a monstrous crimson symbiote visible that contorts, twists, and melts into nightmarish shapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00026973424246534705 + ], + "opposite_logprob_tokens": [ + -8.250269889831543 + ], + "logprob_reward": -0.00026973424246534705, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.00026973424246534705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1557462066411972, + 0.11163516342639923, + 0.060664817690849304, + 0.0468251034617424, + 0.03662426024675369, + 0.03763248398900032, + 0.11580399423837662, + 0.11015108972787857, + 0.2520340085029602, + 0.07288289815187454 + ], + "credit_map": [ + 0.008364738896489143, + 0.004990636836737394, + 0.017260154709219933, + 0.022981910035014153, + 0.016321493312716484, + 0.0441545732319355, + 0.012090133503079414, + 0.02958255633711815, + 0.011490019038319588, + 0.004264087416231632, + 0.011576063930988312, + 0.024821782484650612, + 0.014664201065897942, + 0.015210690908133984, + 0.014414489269256592, + 0.015193821862339973, + 0.006952058058232069, + 0.0029461777303367853, + 0.007341375574469566, + 0.007602708414196968, + 0.008221590891480446, + 0.006625676527619362, + 0.011673797853291035, + 0.009301434271037579, + 0.008444256149232388, + 0.0029934344347566366, + 0.006739369593560696, + 0.004333030432462692, + 0.005018402822315693, + 0.005965132731944323, + 0.005250727292150259, + 0.008080749772489071, + 0.007738115265965462, + 0.0028291079215705395, + 0.004910353105515242, + 0.0047392332926392555, + 0.00282361195422709, + 0.006157104857265949, + 0.0039445324800908566, + 0.0034821999724954367, + 0.006190498359501362, + 0.004945190157741308, + 0.005078674294054508, + 0.004669977817684412, + 0.004698759410530329, + 0.004213240928947926, + 0.004360119812190533, + 0.0034760264679789543, + 0.015680810436606407, + 0.009842649102210999, + 0.019540105015039444, + 0.021556787192821503, + 0.00839533656835556, + 0.012182075530290604, + 0.013156670145690441, + 0.01544955838471651, + 0.014428436756134033, + 0.012765776365995407, + 0.013698680326342583, + 0.013216798193752766, + 0.012663176283240318, + 0.015650246292352676, + 0.014214574359357357, + 0.013513405807316303, + 0.024158285930752754, + 0.042910944670438766, + 0.03839997202157974, + 0.045898426324129105, + 0.020389322191476822, + 0.02879883348941803, + 0.01687142252922058, + 0.03460680693387985, + 0.009671139530837536, + 0.007684577256441116, + 0.006959849968552589, + 0.007505813613533974, + 0.00650571845471859, + 0.008458811789751053, + 0.013007246889173985, + 0.013089747168123722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.00026973424246534705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_4/index_tensor([7289, 7289])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_4/index_tensor([7289, 7289])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e212cc026c9ee51fb740b2d09f8ce6622a111ba7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_4/index_tensor([7289, 7289])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a level angle, then slowly pans to the right, captures an elderly couple lovingly supporting each other. They stand side by side on a tranquil hillside, a vast expanse of sky stained a golden purple by the setting sun behind them. The man's hair is meticulously combed, his wrinkles smoothed by a genuine smile. He wears a soft, light gray V-neck wool sweater, one arm firmly and protectively wrapped around his wife's shoulders. The woman, with her hair in an elegant silver bun and a similarly gentle smile on her face, nestles gently against her husband, a warm ochre cashmere shawl draped over her shoulders. As the camera pans out, a huge, blazing sunset enters the edge of the frame, and a warm and dazzling anamorphic lens flare sweeps horizontally across the entire frame, instantly enveloping them in a dreamy golden light. The light penetrates their silver hair and outlines their silhouettes with a soft light edge. The whole scene is filled with an atmosphere of nostalgia, tranquility and affection.", + "seed": 11966, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_4/index_7289_seed_11966_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030006344616413116, + "vlm_logprob": -0.030006344616413116, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.030006344616413116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08483490347862244 + ], + "opposite_logprob_tokens": [ + -2.5848348140716553 + ], + "logprob_reward": -0.08483490347862244, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08483490347862244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10941371321678162, + 0.06729502230882645, + 0.04084327444434166, + 0.03155244141817093, + 0.047127410769462585, + 0.15573160350322723, + 0.33819323778152466, + 0.05606618523597717, + 0.06057003140449524, + 0.09320708364248276 + ], + "credit_map": [ + 0.027906572446227074, + 0.025580767542123795, + 0.011339976452291012, + 0.009421263821423054, + 0.010764800012111664, + 0.005718850996345282, + 0.010972505435347557, + 0.007708976976573467, + 0.008663970977067947, + 0.014442341402173042, + 0.006768566090613604, + 0.00811609998345375, + 0.00736412825062871, + 0.007721928879618645, + 0.008647601120173931, + 0.005570387002080679, + 0.006328786723315716, + 0.004144861362874508, + 0.00525259505957365, + 0.004876041319221258, + 0.003767810994759202, + 0.008310343138873577, + 0.005528674460947514, + 0.0026341599877923727, + 0.0034484134521335363, + 0.003850894747301936, + 0.00545889139175415, + 0.003879828378558159, + 0.00436074985191226, + 0.003193035488948226, + 0.004292255267500877, + 0.0030683691147714853, + 0.00630906643345952, + 0.005300067365169525, + 0.005292790941894054, + 0.005548219196498394, + 0.005533477291464806, + 0.00987321324646473, + 0.005346414167433977, + 0.003924163058400154, + 0.03630613535642624, + 0.014201107434928417, + 0.0386396124958992, + 0.020671118050813675, + 0.01493077352643013, + 0.008148173801600933, + 0.01328535471111536, + 0.009549320675432682, + 0.08723962306976318, + 0.03283705562353134, + 0.11507704108953476, + 0.04242696613073349, + 0.018610278144478798, + 0.015761595219373703, + 0.017677564173936844, + 0.008563097566366196, + 0.007040069438517094, + 0.004568974953144789, + 0.011671585030853748, + 0.0064609358087182045, + 0.005560779012739658, + 0.0047901994548738, + 0.01123122964054346, + 0.004742414224892855, + 0.009748244658112526, + 0.008521592244505882, + 0.009945190511643887, + 0.00647668307647109, + 0.0061467308551073074, + 0.004543982911854982, + 0.010116075165569782, + 0.005071531515568495, + 0.007010150235146284, + 0.008066367357969284, + 0.010904114693403244, + 0.010377347469329834, + 0.015131914988160133, + 0.0125167490914464, + 0.018079137429594994, + 0.011121304705739021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.08483490347862244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223036356270313 + ], + "opposite_logprob_tokens": [ + -4.264223098754883 + ], + "logprob_reward": -0.014223036356270313, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223036356270313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08493969589471817, + 0.05828296020627022, + 0.03752085193991661, + 0.01429932750761509, + 0.027743089944124222, + 0.0792183130979538, + 0.6038036942481995, + 0.020112179219722748, + 0.036493729799985886, + 0.03758616000413895 + ], + "credit_map": [ + 0.02203037030994892, + 0.020251795649528503, + 0.006701332051306963, + 0.007616529241204262, + 0.006889779586344957, + 0.004851067438721657, + 0.009277147240936756, + 0.007321672979742289, + 0.008779752999544144, + 0.013933444395661354, + 0.0036153430119156837, + 0.006328413728624582, + 0.005263478960841894, + 0.010076184757053852, + 0.006502147298306227, + 0.0037841876037418842, + 0.006224124226719141, + 0.0057253241539001465, + 0.003542740363627672, + 0.002129372674971819, + 0.002927467692643404, + 0.009790435433387756, + 0.004011206794530153, + 0.0031701785046607256, + 0.001617046887986362, + 0.001271994085982442, + 0.0024271304719150066, + 0.0016637115040794015, + 0.0018055873224511743, + 0.0020130190532654524, + 0.0022829067893326283, + 0.001217929646372795, + 0.0034865965135395527, + 0.0028391776140779257, + 0.0035912194289267063, + 0.0026139786932617426, + 0.0039812419563531876, + 0.005758570972830057, + 0.003093952778726816, + 0.00237835175357759, + 0.013829010538756847, + 0.007479035761207342, + 0.014559503644704819, + 0.015675729140639305, + 0.009121359325945377, + 0.004749328829348087, + 0.008153902366757393, + 0.005650436971336603, + 0.15624696016311646, + 0.058413099497556686, + 0.20781834423542023, + 0.07649296522140503, + 0.03193989768624306, + 0.028121214359998703, + 0.02988598309457302, + 0.014885268174111843, + 0.003713026409968734, + 0.0023631653748452663, + 0.004232855048030615, + 0.0018202437786385417, + 0.0023737542796880007, + 0.0012409384362399578, + 0.0025754296220839024, + 0.0017927648732438684, + 0.00461591174826026, + 0.0037099644541740417, + 0.007443545386195183, + 0.004641642794013023, + 0.005225933622568846, + 0.0029873312450945377, + 0.005023117642849684, + 0.0028462845366448164, + 0.003526997519657016, + 0.003074257867410779, + 0.005207047797739506, + 0.004410240333527327, + 0.006027502939105034, + 0.004298512823879719, + 0.007106971926987171, + 0.003934626467525959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.014223036356270313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652104064822197 + ], + "opposite_logprob_tokens": [ + -4.758652210235596 + ], + "logprob_reward": -0.008652104064822197, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652104064822197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2217017263174057, + 0.028963211923837662, + 0.03946959599852562, + 0.013399563729763031, + 0.02338530495762825, + 0.12096710503101349, + 0.4341158866882324, + 0.023917553946375847, + 0.03845999017357826, + 0.05562001094222069 + ], + "credit_map": [ + 0.08201596140861511, + 0.0771401971578598, + 0.009159067645668983, + 0.011347617022693157, + 0.011758057400584221, + 0.008110185153782368, + 0.011886876076459885, + 0.010283773764967918, + 0.0030227508395910263, + 0.00349254859611392, + 0.003279470605775714, + 0.0036437525413930416, + 0.003973593469709158, + 0.0041075763292610645, + 0.004551183432340622, + 0.0028923407662659883, + 0.006343494169414043, + 0.00525214709341526, + 0.004436278250068426, + 0.0025474070571362972, + 0.0033084829337894917, + 0.010139917023479939, + 0.004180720541626215, + 0.0032611479982733727, + 0.0015102463075891137, + 0.0020185408648103476, + 0.0021046570036560297, + 0.0012381024425849319, + 0.0018316612113267183, + 0.0014755778247490525, + 0.0020736870355904102, + 0.0011470930185168982, + 0.0028429508674889803, + 0.003095218213275075, + 0.003060587216168642, + 0.0018975932616740465, + 0.0031957386527210474, + 0.005112625192850828, + 0.002492345869541168, + 0.00168824871070683, + 0.028821002691984177, + 0.010797045193612576, + 0.031708039343357086, + 0.017048915848135948, + 0.012059271335601807, + 0.004768301732838154, + 0.009256573393940926, + 0.006507967133074999, + 0.11252786219120026, + 0.04202790930867195, + 0.14874103665351868, + 0.055051639676094055, + 0.023003634065389633, + 0.019886041060090065, + 0.02159259095788002, + 0.011285185813903809, + 0.005054070148617029, + 0.0028714018408209085, + 0.005749039817601442, + 0.002549596596509218, + 0.0018380862893536687, + 0.0014855522895231843, + 0.002510342514142394, + 0.0018594684079289436, + 0.0068397908471524715, + 0.004317580256611109, + 0.007125689648091793, + 0.004743121098726988, + 0.0038706494960933924, + 0.002282479079440236, + 0.006108570843935013, + 0.003172110766172409, + 0.005460278131067753, + 0.00479832012206316, + 0.007171269040554762, + 0.008456872776150703, + 0.008666766807436943, + 0.005633241962641478, + 0.010088056325912476, + 0.005345209967344999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.008652104064822197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368576616048813 + ], + "opposite_logprob_tokens": [ + -3.4086856842041016 + ], + "logprob_reward": -0.03368576616048813, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368576616048813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09551016241312027, + 0.04788947477936745, + 0.08159564435482025, + 0.037017662078142166, + 0.026124143972992897, + 0.12327436357736588, + 0.13229434192180634, + 0.07382748275995255, + 0.24315662682056427, + 0.13931000232696533 + ], + "credit_map": [ + 0.028773188591003418, + 0.028166502714157104, + 0.008091500028967857, + 0.007203387562185526, + 0.005149901378899813, + 0.005351513624191284, + 0.007348125334829092, + 0.005426051095128059, + 0.0071809361688792706, + 0.010008540004491806, + 0.004132013767957687, + 0.0037975935265421867, + 0.0031474276911467314, + 0.011000349186360836, + 0.005356086418032646, + 0.0032665312755852938, + 0.010339331813156605, + 0.00938349962234497, + 0.010257448069751263, + 0.007331811357289553, + 0.00744706392288208, + 0.01641940511763096, + 0.013040446676313877, + 0.0073766522109508514, + 0.0034168772399425507, + 0.004865237046033144, + 0.006348281167447567, + 0.0045135533437132835, + 0.004471153486520052, + 0.00498689291998744, + 0.005468236282467842, + 0.0029474354814738035, + 0.0037562784273177385, + 0.0028248766902834177, + 0.0026563662104308605, + 0.0035558599047362804, + 0.0028802286833524704, + 0.00409209867939353, + 0.003752143820747733, + 0.0026062969118356705, + 0.028258657082915306, + 0.01030617393553257, + 0.03290501981973648, + 0.01617744006216526, + 0.010723034851253033, + 0.006543408147990704, + 0.009434020146727562, + 0.008926616981625557, + 0.03041728399693966, + 0.020062226802110672, + 0.0325968861579895, + 0.015125835314393044, + 0.009677796624600887, + 0.009040603414177895, + 0.008687575347721577, + 0.006686147768050432, + 0.007600882556289434, + 0.011038864962756634, + 0.01069073099642992, + 0.00967755913734436, + 0.009436126798391342, + 0.005906401202082634, + 0.01414746418595314, + 0.005329464096575975, + 0.039427462965250015, + 0.03687329217791557, + 0.056712713092565536, + 0.03164169192314148, + 0.022503582760691643, + 0.013642650097608566, + 0.024712087586522102, + 0.01764318346977234, + 0.015332797542214394, + 0.010608727112412453, + 0.02634870447218418, + 0.033600665628910065, + 0.014430836774408817, + 0.012202614918351173, + 0.017391063272953033, + 0.009394600987434387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.03368576616048813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635913021862507 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.008635913021862507, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635913021862507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08695821464061737, + 0.03808308392763138, + 0.03208037465810776, + 0.014530680142343044, + 0.01126093976199627, + 0.029923440888524055, + 0.4303697347640991, + 0.011983294039964676, + 0.07740936428308487, + 0.267400860786438 + ], + "credit_map": [ + 0.026966024190187454, + 0.025520194321870804, + 0.006319459527730942, + 0.0059501114301383495, + 0.0046780649572610855, + 0.007001436781138182, + 0.005504791159182787, + 0.005018135998398066, + 0.0067098322324454784, + 0.011566485278308392, + 0.0027038000989705324, + 0.004360112827271223, + 0.002810850739479065, + 0.005250654648989439, + 0.002793879248201847, + 0.0018874678062275052, + 0.005135396961122751, + 0.004451177082955837, + 0.003677132772281766, + 0.001610688166692853, + 0.002879951149225235, + 0.008174922317266464, + 0.003429003991186619, + 0.0027221045456826687, + 0.0014599284622818232, + 0.0015953521942719817, + 0.0027639467734843493, + 0.0016432394040748477, + 0.0018840334378182888, + 0.0014250273816287518, + 0.0026168676558882, + 0.001142285531386733, + 0.0015711571322754025, + 0.001413789694197476, + 0.0015729781007394195, + 0.00165832182392478, + 0.001452121650800109, + 0.0015507249627262354, + 0.0012614901643246412, + 0.0007803560001775622, + 0.0027847359888255596, + 0.0025150084402412176, + 0.003112290520220995, + 0.004211657680571079, + 0.003820814425125718, + 0.00523508433252573, + 0.00439273938536644, + 0.0038511117454618216, + 0.11197058111429214, + 0.042596425861120224, + 0.14795587956905365, + 0.055280569940805435, + 0.022442877292633057, + 0.019805477932095528, + 0.020884646102786064, + 0.00943327322602272, + 0.0019855061545968056, + 0.0017054729396477342, + 0.0026134212967008352, + 0.0013164097908884287, + 0.0010415413416922092, + 0.000704365607816726, + 0.0016363593749701977, + 0.000980217126198113, + 0.011694065295159817, + 0.010764981620013714, + 0.022219810634851456, + 0.011792329140007496, + 0.005880323704332113, + 0.004127958789467812, + 0.006657586898654699, + 0.004272304475307465, + 0.03406330943107605, + 0.019050784409046173, + 0.058104608207941055, + 0.08894718438386917, + 0.01809275522828102, + 0.013117827475070953, + 0.02482721209526062, + 0.011197181418538094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.008635913021862507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, then slowly pans to the right, captures an elderly couple lovingly supporting each other. They stand side by side on a tranquil hillside, a vast expanse of sky stained a golden purple by the setting sun behind them. The man's hair is meticulously combed, his wrinkles smoothed by a genuine smile. He wears a soft, light gray V-neck wool sweater, one arm firmly and protectively wrapped around his wife's shoulders. The woman, with her hair in an elegant silver bun and a similarly gentle smile on her face, nestles gently against her husband, a warm ochre cashmere shawl draped over her shoulders. As the camera pans out, a huge, blazing sunset enters the edge of the frame, and a warm and dazzling anamorphic lens flare sweeps horizontally across the entire frame, instantly enveloping them in a dreamy golden light. The light penetrates their silver hair and outlines their silhouettes with a soft light edge. The whole scene is filled with an atmosphere of nostalgia, tranquility and affection.", + "seed": 11967, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_4/index_7289_seed_11967_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04789770357310772, + "vlm_logprob": -0.04789770357310772, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04789770357310772, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06442099809646606 + ], + "opposite_logprob_tokens": [ + -3.0644209384918213 + ], + "logprob_reward": -0.06442099809646606, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06442099809646606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1254328489303589, + 0.063299261033535, + 0.323881059885025, + 0.07041120529174805, + 0.05595570057630539, + 0.04398398473858833, + 0.03580057993531227, + 0.05455048382282257, + 0.05824314057826996, + 0.16844168305397034 + ], + "credit_map": [ + 0.014073953032493591, + 0.01881827414035797, + 0.015115172602236271, + 0.01269825454801321, + 0.01759505085647106, + 0.01839391142129898, + 0.014015731401741505, + 0.014722497202455997, + 0.0076090553775429726, + 0.01235842052847147, + 0.007637560833245516, + 0.007872024551033974, + 0.007477275095880032, + 0.005816959775984287, + 0.007604117039591074, + 0.0069238473661243916, + 0.04981395602226257, + 0.0827033668756485, + 0.026971811428666115, + 0.030163686722517014, + 0.032265763729810715, + 0.03779066354036331, + 0.031040847301483154, + 0.03313096985220909, + 0.008025817573070526, + 0.012472706846892834, + 0.0059432825073599815, + 0.010059203952550888, + 0.010282501578330994, + 0.004823561757802963, + 0.011072118766605854, + 0.007732014171779156, + 0.006025389302521944, + 0.005567345302551985, + 0.006232299841940403, + 0.00923947710543871, + 0.008813300170004368, + 0.006163176614791155, + 0.007583822123706341, + 0.006330890115350485, + 0.005426366347819567, + 0.004617816302925348, + 0.00461009843274951, + 0.0063055939972400665, + 0.005789961200207472, + 0.0059380484744906425, + 0.00603331858292222, + 0.005262778140604496, + 0.004782583564519882, + 0.0028026499785482883, + 0.0029186801984906197, + 0.008223576471209526, + 0.004438772331923246, + 0.0036192971747368574, + 0.0050424979999661446, + 0.003972521051764488, + 0.006518126931041479, + 0.00707708764821291, + 0.004143835976719856, + 0.009636756964027882, + 0.010143371298909187, + 0.00546066090464592, + 0.005380365066230297, + 0.006190279498696327, + 0.007397327106446028, + 0.00834856927394867, + 0.008115893229842186, + 0.007112983148545027, + 0.0069770184345543385, + 0.005395510699599981, + 0.005336822476238012, + 0.009559017606079578, + 0.02072196640074253, + 0.020004451274871826, + 0.020307466387748718, + 0.025301888585090637, + 0.020426103845238686, + 0.02549157850444317, + 0.017919540405273438, + 0.018268687650561333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.06442099809646606 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610555499792099 + ], + "opposite_logprob_tokens": [ + -4.141105651855469 + ], + "logprob_reward": -0.01610555499792099, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610555499792099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13382215797901154, + 0.09933502227067947, + 0.12112561613321304, + 0.08644016832113266, + 0.03800459951162338, + 0.02559482865035534, + 0.017487356439232826, + 0.12630176544189453, + 0.08605387806892395, + 0.26583465933799744 + ], + "credit_map": [ + 0.015012829564511776, + 0.02057868242263794, + 0.013476449064910412, + 0.010379752144217491, + 0.01907210424542427, + 0.02680385671555996, + 0.01212418545037508, + 0.016374295577406883, + 0.017603902146220207, + 0.024280939251184464, + 0.006776556838303804, + 0.008349395357072353, + 0.010463641956448555, + 0.010087188333272934, + 0.009282001294195652, + 0.01249138917773962, + 0.012490534223616123, + 0.014166965149343014, + 0.01524488627910614, + 0.012679039500653744, + 0.02632228285074234, + 0.01450903806835413, + 0.01456028874963522, + 0.011152582243084908, + 0.011798294261097908, + 0.019220460206270218, + 0.008977794088423252, + 0.009455354884266853, + 0.011388774961233139, + 0.0063063171692192554, + 0.009868944995105267, + 0.009424231015145779, + 0.0038673721719533205, + 0.006164549849927425, + 0.0029768741223961115, + 0.005854995921254158, + 0.0056587643921375275, + 0.0033042796421796083, + 0.004590279888361692, + 0.00558748422190547, + 0.003901829244568944, + 0.0020583004225045443, + 0.0035962481051683426, + 0.0031936676241457462, + 0.002684196922928095, + 0.004783097188919783, + 0.0027458842378109694, + 0.0026316025760024786, + 0.001938663306646049, + 0.0018916567787528038, + 0.001433806261047721, + 0.002876137848943472, + 0.0028516841121017933, + 0.0014614462852478027, + 0.0030325709376484156, + 0.002001391025260091, + 0.012604424729943275, + 0.013668935745954514, + 0.007263342849910259, + 0.017522944137454033, + 0.030982039868831635, + 0.01095704734325409, + 0.016606325283646584, + 0.016696712002158165, + 0.010887578129768372, + 0.012653855606913567, + 0.008548504672944546, + 0.009660162031650543, + 0.015109431929886341, + 0.009201993234455585, + 0.009613373316824436, + 0.010378973558545113, + 0.0755828320980072, + 0.016679270192980766, + 0.09726209193468094, + 0.022819167003035545, + 0.011634891852736473, + 0.019317584112286568, + 0.010867860168218613, + 0.01167095173150301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.01610555499792099 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652577176690102 + ], + "opposite_logprob_tokens": [ + -4.758652687072754 + ], + "logprob_reward": -0.008652577176690102, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652577176690102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1331014633178711, + 0.0636167973279953, + 0.2800401747226715, + 0.07058296352624893, + 0.0586324967443943, + 0.029503678902983665, + 0.024236423894762993, + 0.0693730041384697, + 0.10175783187150955, + 0.16915512084960938 + ], + "credit_map": [ + 0.015554649755358696, + 0.016979075968265533, + 0.011117598973214626, + 0.02030283398926258, + 0.016609396785497665, + 0.021933482959866524, + 0.015496297739446163, + 0.015108149498701096, + 0.01042536273598671, + 0.012560039758682251, + 0.005313056521117687, + 0.007024198304861784, + 0.006658500526100397, + 0.0064668855629861355, + 0.008259780704975128, + 0.006908982526510954, + 0.037510693073272705, + 0.05534457787871361, + 0.028665073215961456, + 0.0205815639346838, + 0.059960197657346725, + 0.023471826687455177, + 0.027306562289595604, + 0.027199698612093925, + 0.00958202127367258, + 0.013465408235788345, + 0.007313155103474855, + 0.010059972293674946, + 0.00882640853524208, + 0.006577487103641033, + 0.008991030976176262, + 0.005767477676272392, + 0.006089990492910147, + 0.008061542175710201, + 0.004523545037955046, + 0.01189199648797512, + 0.009773229248821735, + 0.003527672728523612, + 0.008114630356431007, + 0.006649894639849663, + 0.003799769328907132, + 0.0038267665077000856, + 0.0024272657465189695, + 0.005063743330538273, + 0.003227005945518613, + 0.003646582830697298, + 0.003671343205496669, + 0.003841203870251775, + 0.002890774980187416, + 0.0021903703454881907, + 0.0018672950100153685, + 0.003531961003318429, + 0.00454778503626585, + 0.0024300459772348404, + 0.0036967915948480368, + 0.0030813992489129305, + 0.007545924745500088, + 0.009245455265045166, + 0.0037607986014336348, + 0.013840570114552975, + 0.010984111577272415, + 0.005470590200275183, + 0.009009324945509434, + 0.009516225196421146, + 0.012308666482567787, + 0.015906324610114098, + 0.008506661280989647, + 0.014298192225396633, + 0.018213296309113503, + 0.00971789937466383, + 0.012352865189313889, + 0.010453932918608189, + 0.02821490727365017, + 0.017153076827526093, + 0.03699568286538124, + 0.02230289950966835, + 0.0170210599899292, + 0.020704174414277077, + 0.00929272174835205, + 0.017470603808760643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.008652577176690102 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702201306819916 + ], + "opposite_logprob_tokens": [ + -2.1270220279693604 + ], + "logprob_reward": -0.12702201306819916, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702201306819916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0884147509932518, + 0.0865994468331337, + 0.10834398865699768, + 0.10163968056440353, + 0.06348113715648651, + 0.038531046360731125, + 0.025036556646227837, + 0.044865842908620834, + 0.0933748111128807, + 0.34971269965171814 + ], + "credit_map": [ + 0.0100489417091012, + 0.010865354910492897, + 0.00879710540175438, + 0.010556467808783054, + 0.014268692582845688, + 0.015159272588789463, + 0.007617896888405085, + 0.011101015843451023, + 0.017555303871631622, + 0.026317304000258446, + 0.006578405853360891, + 0.008017283864319324, + 0.00867157056927681, + 0.0066863480024039745, + 0.006188123486936092, + 0.006585105787962675, + 0.017570121213793755, + 0.017068421468138695, + 0.01434249710291624, + 0.007656603120267391, + 0.015372267924249172, + 0.016937021166086197, + 0.009668920189142227, + 0.009728135541081429, + 0.01357525959610939, + 0.020102722570300102, + 0.010247595608234406, + 0.011979146860539913, + 0.014727904461324215, + 0.007587380241602659, + 0.013215071521699429, + 0.010204599238932133, + 0.005924235563725233, + 0.009708346799015999, + 0.005343024153262377, + 0.012346084229648113, + 0.009763664565980434, + 0.00501085352152586, + 0.007887410931289196, + 0.007497516926378012, + 0.004639550112187862, + 0.004092616029083729, + 0.002656408352777362, + 0.007535623852163553, + 0.005902181379497051, + 0.003705537412315607, + 0.0053230463527143, + 0.004676083102822304, + 0.0025949948467314243, + 0.002495663706213236, + 0.001310503575950861, + 0.0036975450348109007, + 0.006064220331609249, + 0.0022739102132618427, + 0.003791878465563059, + 0.002807838609442115, + 0.0039604962803423405, + 0.005127605050802231, + 0.002243880182504654, + 0.007076816633343697, + 0.008821641094982624, + 0.003584539284929633, + 0.006835818290710449, + 0.007215047255158424, + 0.011420364491641521, + 0.01439774502068758, + 0.014036718755960464, + 0.013127398677170277, + 0.013044222258031368, + 0.01058462355285883, + 0.0062208017334342, + 0.010542931035161018, + 0.058371737599372864, + 0.039210326969623566, + 0.06709413975477219, + 0.04170447215437889, + 0.039171479642391205, + 0.04403705522418022, + 0.02780556119978428, + 0.03231795132160187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.12702201306819916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287374526262283 + ], + "opposite_logprob_tokens": [ + -3.773287296295166 + ], + "logprob_reward": -0.023287374526262283, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287374526262283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08705542236566544, + 0.24507473409175873, + 0.2322171926498413, + 0.10042350739240646, + 0.06085324287414551, + 0.019329488277435303, + 0.01947362720966339, + 0.038327693939208984, + 0.07662086933851242, + 0.12062419205904007 + ], + "credit_map": [ + 0.010493187233805656, + 0.009614323265850544, + 0.015062729828059673, + 0.010469459928572178, + 0.009833155199885368, + 0.009956881403923035, + 0.01281002163887024, + 0.008815660141408443, + 0.04440223425626755, + 0.0642436221241951, + 0.027361251413822174, + 0.022414907813072205, + 0.023354841396212578, + 0.022983558475971222, + 0.01821638084948063, + 0.022097967565059662, + 0.043286874890327454, + 0.0666404590010643, + 0.01036108285188675, + 0.018885377794504166, + 0.04219678416848183, + 0.018945790827274323, + 0.016784368082880974, + 0.015116484835743904, + 0.015407825820147991, + 0.021216262131929398, + 0.013457722030580044, + 0.01252251211553812, + 0.011029593646526337, + 0.006627219263464212, + 0.01107410341501236, + 0.009088268503546715, + 0.00600971607491374, + 0.010198721662163734, + 0.0038589362520724535, + 0.011483770795166492, + 0.008705414831638336, + 0.004886569920927286, + 0.00784830842167139, + 0.007861805148422718, + 0.002300142077729106, + 0.002158072078600526, + 0.0020218880381435156, + 0.0025376698467880487, + 0.002530633006244898, + 0.002587445080280304, + 0.0023405104875564575, + 0.0028531276620924473, + 0.0020138616673648357, + 0.001680994057096541, + 0.0017483215779066086, + 0.004580385517328978, + 0.0024995834100991488, + 0.0019661965779960155, + 0.002487066201865673, + 0.002497217385098338, + 0.00448880298063159, + 0.005214067175984383, + 0.0024011023342609406, + 0.00806761346757412, + 0.0062386468052864075, + 0.0030071446672081947, + 0.004809393547475338, + 0.004100921098142862, + 0.009643294848501682, + 0.01081157848238945, + 0.01001587975770235, + 0.01259723212569952, + 0.008298086933791637, + 0.007377935107797384, + 0.007082417607307434, + 0.010794448666274548, + 0.019511602818965912, + 0.015738902613520622, + 0.02054302580654621, + 0.012234947644174099, + 0.018292240798473358, + 0.015020808205008507, + 0.009222152642905712, + 0.010060516186058521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.023287374526262283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, then slowly pans to the right, captures an elderly couple lovingly supporting each other. They stand side by side on a tranquil hillside, a vast expanse of sky stained a golden purple by the setting sun behind them. The man's hair is meticulously combed, his wrinkles smoothed by a genuine smile. He wears a soft, light gray V-neck wool sweater, one arm firmly and protectively wrapped around his wife's shoulders. The woman, with her hair in an elegant silver bun and a similarly gentle smile on her face, nestles gently against her husband, a warm ochre cashmere shawl draped over her shoulders. As the camera pans out, a huge, blazing sunset enters the edge of the frame, and a warm and dazzling anamorphic lens flare sweeps horizontally across the entire frame, instantly enveloping them in a dreamy golden light. The light penetrates their silver hair and outlines their silhouettes with a soft light edge. The whole scene is filled with an atmosphere of nostalgia, tranquility and affection.", + "seed": 11968, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_4/index_7289_seed_11968_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031094479840248823, + "vlm_logprob": -0.031094479840248823, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.031094479840248823, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07056429237127304 + ], + "opposite_logprob_tokens": [ + -2.8205642700195312 + ], + "logprob_reward": -0.07056429237127304, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07056429237127304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08469989895820618, + 0.060018181800842285, + 0.019398093223571777, + 0.021420689299702644, + 0.543019711971283, + 0.14827477931976318, + 0.013776861131191254, + 0.020137500017881393, + 0.022258805111050606, + 0.06699545681476593 + ], + "credit_map": [ + 0.01095869392156601, + 0.010281933471560478, + 0.009213604032993317, + 0.010091368108987808, + 0.011353257112205029, + 0.01700465939939022, + 0.008302762173116207, + 0.007493625394999981, + 0.006709076929837465, + 0.006237384397536516, + 0.00541697395965457, + 0.00775486696511507, + 0.008528107777237892, + 0.008084002882242203, + 0.008899664506316185, + 0.008388105779886246, + 0.0025100966449826956, + 0.0021572846453636885, + 0.0017144472803920507, + 0.002132169436663389, + 0.0025639154482632875, + 0.002985146827995777, + 0.002795125124976039, + 0.0025399099104106426, + 0.0028895086143165827, + 0.002168599283322692, + 0.00212524994276464, + 0.0035454474855214357, + 0.003422702429816127, + 0.0029400165658444166, + 0.002275988459587097, + 0.00205317628569901, + 0.22786207497119904, + 0.01560682337731123, + 0.01283679436892271, + 0.01087221410125494, + 0.009475858882069588, + 0.24455207586288452, + 0.01010627206414938, + 0.011707616038620472, + 0.01839827001094818, + 0.022915242239832878, + 0.019618378952145576, + 0.02826991118490696, + 0.014359630644321442, + 0.02053109183907509, + 0.010761674493551254, + 0.01342057716101408, + 0.0017584313172847033, + 0.0012902474263682961, + 0.0020077817607671022, + 0.0019723898731172085, + 0.0017660645535215735, + 0.002164604142308235, + 0.0015613005962222815, + 0.0012560425093397498, + 0.001882322714664042, + 0.0018318763468414545, + 0.001794368145056069, + 0.0031986315734684467, + 0.0034208358265459538, + 0.0012996488949283957, + 0.004404446110129356, + 0.0023053695913404226, + 0.00251406361348927, + 0.0025573140010237694, + 0.0022603359539061785, + 0.003852179506793618, + 0.0034865422639995813, + 0.0015266352565959096, + 0.0037316076923161745, + 0.0023301278706640005, + 0.006822440307587385, + 0.004861894529312849, + 0.005707297939807177, + 0.008960624225437641, + 0.010261940769851208, + 0.006334628444164991, + 0.017747193574905396, + 0.006299430504441261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.07056429237127304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061290293931961 + ], + "opposite_logprob_tokens": [ + -3.895612955093384 + ], + "logprob_reward": -0.02061290293931961, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061290293931961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09098733961582184, + 0.04679694399237633, + 0.043950360268354416, + 0.061851661652326584, + 0.38920557498931885, + 0.05340912565588951, + 0.049083586782217026, + 0.09026464074850082, + 0.035189367830753326, + 0.1392614096403122 + ], + "credit_map": [ + 0.011686769314110279, + 0.010104705579578876, + 0.007440788205713034, + 0.015489698387682438, + 0.012337514199316502, + 0.009777416475117207, + 0.014520077034831047, + 0.009630377404391766, + 0.006100847851485014, + 0.002916192403063178, + 0.0030002016574144363, + 0.01033608615398407, + 0.006131813395768404, + 0.006294773425906897, + 0.006296541076153517, + 0.00572048919275403, + 0.005205658730119467, + 0.0045433733612298965, + 0.0031294093932956457, + 0.00967549066990614, + 0.006477548740804195, + 0.005469799507409334, + 0.0052023958414793015, + 0.004246687050908804, + 0.009820275008678436, + 0.005595885682851076, + 0.004387719091027975, + 0.009688596241176128, + 0.007816318422555923, + 0.009799443185329437, + 0.007157092448323965, + 0.007586331572383642, + 0.15073297917842865, + 0.01696951314806938, + 0.009419208392500877, + 0.013120656833052635, + 0.014211880974471569, + 0.16183559596538544, + 0.011129510588943958, + 0.011786211282014847, + 0.007103919982910156, + 0.0059533133171498775, + 0.004082123748958111, + 0.009167209267616272, + 0.008019601926207542, + 0.007718756794929504, + 0.005672790110111237, + 0.005691410042345524, + 0.007063747849315405, + 0.0034247732255607843, + 0.006645975634455681, + 0.005394871812313795, + 0.004426743369549513, + 0.011408544145524502, + 0.0060897222720086575, + 0.004629204981029034, + 0.007969624362885952, + 0.0068915230222046375, + 0.008640329353511333, + 0.013604245148599148, + 0.016504328697919846, + 0.0057968501932919025, + 0.019479958340525627, + 0.011377780698239803, + 0.0035957470536231995, + 0.002912361640483141, + 0.0037645772099494934, + 0.004941364284604788, + 0.005350447725504637, + 0.002382345264777541, + 0.008410186506807804, + 0.0038323390763252974, + 0.015545650385320187, + 0.014822056517004967, + 0.011623942293226719, + 0.017032278701663017, + 0.0198662206530571, + 0.012569384649395943, + 0.03416850417852402, + 0.013633368536829948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.02061290293931961 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262095481157303 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005262095481157303, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262095481157303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09020641446113586, + 0.05348115786910057, + 0.0366818867623806, + 0.03214845433831215, + 0.49800562858581543, + 0.04105928912758827, + 0.0747155025601387, + 0.06416654586791992, + 0.032709088176488876, + 0.0768260657787323 + ], + "credit_map": [ + 0.011234554462134838, + 0.012108390219509602, + 0.008384949527680874, + 0.012526237405836582, + 0.013602517545223236, + 0.008356108330190182, + 0.012348410673439503, + 0.011645253747701645, + 0.009414146654307842, + 0.005154834594577551, + 0.004123102407902479, + 0.00938239973038435, + 0.0061422246508300304, + 0.008094383403658867, + 0.0052439370192587376, + 0.005926125682890415, + 0.00583430053666234, + 0.003341262461617589, + 0.002904019085690379, + 0.00812823697924614, + 0.004808882251381874, + 0.004961954429745674, + 0.0037130529526621103, + 0.0029901782982051373, + 0.005691425874829292, + 0.0033694293815642595, + 0.00256862654350698, + 0.005929466336965561, + 0.003690518671646714, + 0.005252270493656397, + 0.002939946949481964, + 0.0027067698538303375, + 0.2071242481470108, + 0.01499013788998127, + 0.011377652175724506, + 0.01092000026255846, + 0.009641427546739578, + 0.22208788990974426, + 0.010134579613804817, + 0.01172971073538065, + 0.0048264265060424805, + 0.00455454969778657, + 0.005911281798034906, + 0.007444028276950121, + 0.004841613583266735, + 0.004557380918413401, + 0.004041214473545551, + 0.004882795736193657, + 0.010105178691446781, + 0.005684342700988054, + 0.010912868194282055, + 0.009031231515109539, + 0.007444634567946196, + 0.015089530497789383, + 0.009394604712724686, + 0.007053114473819733, + 0.0060377540066838264, + 0.005852478090673685, + 0.007248704321682453, + 0.010075590573251247, + 0.011027202941477299, + 0.003856844734400511, + 0.012307812459766865, + 0.007760158274322748, + 0.003969704732298851, + 0.004891104996204376, + 0.005607674829661846, + 0.0057129873894155025, + 0.0038700897712260485, + 0.0018154539866372943, + 0.003474361961707473, + 0.0033677101600915194, + 0.007991877384483814, + 0.005787759553641081, + 0.006360090337693691, + 0.011010462418198586, + 0.011650319211184978, + 0.0069466326385736465, + 0.019846120849251747, + 0.007232797797769308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.005262095481157303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054937370121479034 + ], + "opposite_logprob_tokens": [ + -2.9299373626708984 + ], + "logprob_reward": -0.054937370121479034, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054937370121479034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062018465250730515, + 0.052719682455062866, + 0.03920433670282364, + 0.016719311475753784, + 0.3600603938102722, + 0.0463719516992569, + 0.06588897854089737, + 0.029467657208442688, + 0.12542356550693512, + 0.2021256983280182 + ], + "credit_map": [ + 0.007852171547710896, + 0.007770512718707323, + 0.006414144299924374, + 0.007486896123737097, + 0.007216302212327719, + 0.009114790707826614, + 0.00909908302128315, + 0.007064559496939182, + 0.007771476171910763, + 0.008651429787278175, + 0.004591348581016064, + 0.008265923708677292, + 0.006051226053386927, + 0.005966243799775839, + 0.0059547582641243935, + 0.005467271897941828, + 0.006861243862658739, + 0.006088694091886282, + 0.003510131733492017, + 0.006429973524063826, + 0.0037659937515854836, + 0.004768724553287029, + 0.004054847173392773, + 0.0037247256841510534, + 0.002644368913024664, + 0.003026246326044202, + 0.001617992646060884, + 0.0023507499136030674, + 0.001774635398760438, + 0.0020301819313317537, + 0.0016309311613440514, + 0.0016442027408629656, + 0.1472366452217102, + 0.011678742244839668, + 0.00882622878998518, + 0.010047050192952156, + 0.008139204233884811, + 0.15832437574863434, + 0.007632584776729345, + 0.008175482973456383, + 0.004929691087454557, + 0.0073664248920977116, + 0.006073970813304186, + 0.008113312534987926, + 0.005845606792718172, + 0.004708446562290192, + 0.004653554409742355, + 0.0046809399500489235, + 0.008701290935277939, + 0.005087843630462885, + 0.01011477317661047, + 0.009264918975532055, + 0.0056647141464054585, + 0.014258275739848614, + 0.00641991151496768, + 0.006377245765179396, + 0.0030603462364524603, + 0.002760815666988492, + 0.003218343947082758, + 0.003778548212721944, + 0.004514877684414387, + 0.004118674900382757, + 0.00376824289560318, + 0.004247803706675768, + 0.01216108724474907, + 0.011638066731393337, + 0.01407579705119133, + 0.023951103910803795, + 0.016689151525497437, + 0.008022598922252655, + 0.024625511839985847, + 0.014260233379900455, + 0.020272308960556984, + 0.021181175485253334, + 0.0174521766602993, + 0.031002415344119072, + 0.03474696725606918, + 0.01723080314695835, + 0.04128098860383034, + 0.018958842381834984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.054937370121479034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095738288015127 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095738288015127, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095738288015127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046504851430654526, + 0.0385621152818203, + 0.013871007598936558, + 0.016209691762924194, + 0.33617302775382996, + 0.034424494951963425, + 0.06642117351293564, + 0.039046160876750946, + 0.18497885763645172, + 0.2238086760044098 + ], + "credit_map": [ + 0.005671605933457613, + 0.007169544231146574, + 0.0036376523785293102, + 0.006040418054908514, + 0.0055208150297403336, + 0.0063507636077702045, + 0.007020189426839352, + 0.005093857180327177, + 0.006572546903043985, + 0.00407101446762681, + 0.002187378006055951, + 0.005842830520123243, + 0.005711591336876154, + 0.006660056300461292, + 0.0043345969170331955, + 0.003182098036631942, + 0.002205061260610819, + 0.0015477814013138413, + 0.0010146701242774725, + 0.0023379286285489798, + 0.0016022383933886886, + 0.0019343869062140584, + 0.0017562290886417031, + 0.0014727114466950297, + 0.002702294383198023, + 0.0015245871618390083, + 0.0015343263512477279, + 0.0023926906287670135, + 0.002032896503806114, + 0.0023446998093277216, + 0.0020521061960607767, + 0.0016260907286778092, + 0.1402839571237564, + 0.009008638560771942, + 0.00903416983783245, + 0.008048959076404572, + 0.006173280067741871, + 0.1496010273694992, + 0.006634348537772894, + 0.007388629950582981, + 0.004721296485513449, + 0.003326686332002282, + 0.00417325971648097, + 0.005353027489036322, + 0.004843524657189846, + 0.006047359202057123, + 0.002894286997616291, + 0.0030650494154542685, + 0.010096178390085697, + 0.005139579530805349, + 0.011559991165995598, + 0.007082567550241947, + 0.005419009365141392, + 0.015970377251505852, + 0.005090788472443819, + 0.006062672473490238, + 0.004707625135779381, + 0.0034861350432038307, + 0.006945920642465353, + 0.0047929310239851475, + 0.004956728778779507, + 0.005217142403125763, + 0.003914327826350927, + 0.005025344435125589, + 0.02048526145517826, + 0.018302060663700104, + 0.033760517835617065, + 0.029977712780237198, + 0.02454436384141445, + 0.012202423065900803, + 0.02463683672249317, + 0.021069655194878578, + 0.03467773646116257, + 0.040278077125549316, + 0.03379269689321518, + 0.03533139452338219, + 0.02253950759768486, + 0.01697518490254879, + 0.022708723321557045, + 0.017505329102277756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.004095738288015127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, then slowly pans to the right, captures an elderly couple lovingly supporting each other. They stand side by side on a tranquil hillside, a vast expanse of sky stained a golden purple by the setting sun behind them. The man's hair is meticulously combed, his wrinkles smoothed by a genuine smile. He wears a soft, light gray V-neck wool sweater, one arm firmly and protectively wrapped around his wife's shoulders. The woman, with her hair in an elegant silver bun and a similarly gentle smile on her face, nestles gently against her husband, a warm ochre cashmere shawl draped over her shoulders. As the camera pans out, a huge, blazing sunset enters the edge of the frame, and a warm and dazzling anamorphic lens flare sweeps horizontally across the entire frame, instantly enveloping them in a dreamy golden light. The light penetrates their silver hair and outlines their silhouettes with a soft light edge. The whole scene is filled with an atmosphere of nostalgia, tranquility and affection.", + "seed": 11969, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_4/index_7289_seed_11969_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03105220217257738, + "vlm_logprob": -0.03105220217257738, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03105220217257738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07434587925672531 + ], + "opposite_logprob_tokens": [ + -2.824345827102661 + ], + "logprob_reward": -0.07434587925672531, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07434587925672531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1562785804271698, + 0.05845874175429344, + 0.06757790595293045, + 0.106063611805439, + 0.03876492753624916, + 0.08273637294769287, + 0.040283203125, + 0.05716880410909653, + 0.24819380044937134, + 0.1444740891456604 + ], + "credit_map": [ + 0.020614104345440865, + 0.01856835186481476, + 0.01483109500259161, + 0.02016010694205761, + 0.02392483502626419, + 0.01645772159099579, + 0.020943356677889824, + 0.020779000595211983, + 0.006718606222420931, + 0.006200352683663368, + 0.008603280410170555, + 0.008286246098577976, + 0.007803698070347309, + 0.003599055577069521, + 0.010868288576602936, + 0.006379213649779558, + 0.007913944311439991, + 0.008146470412611961, + 0.009415851905941963, + 0.009510968811810017, + 0.007443315349519253, + 0.013071534223854542, + 0.005295346956700087, + 0.006780473981052637, + 0.01799686625599861, + 0.009431911632418633, + 0.010192040354013443, + 0.008970091119408607, + 0.010276035405695438, + 0.028559710830450058, + 0.010981007479131222, + 0.00965595431625843, + 0.006431820336729288, + 0.0032067412976175547, + 0.00406732177361846, + 0.005571880843490362, + 0.005703385919332504, + 0.006509995087981224, + 0.0040159448981285095, + 0.0032578371465206146, + 0.02278258092701435, + 0.003871738212183118, + 0.003655987558886409, + 0.007493957877159119, + 0.006525383796542883, + 0.021510163322091103, + 0.009572549723088741, + 0.007324006408452988, + 0.004639990162104368, + 0.003618069691583514, + 0.005383669398725033, + 0.006642053369432688, + 0.006508293095976114, + 0.005397304892539978, + 0.0047367410734295845, + 0.0033570807427167892, + 0.007984091527760029, + 0.004939333535730839, + 0.00645374646410346, + 0.009216886013746262, + 0.009096173569560051, + 0.004730405751615763, + 0.00848367903381586, + 0.006264486815780401, + 0.025138940662145615, + 0.024963125586509705, + 0.028509754687547684, + 0.027362395077943802, + 0.04504462704062462, + 0.024328885599970818, + 0.05293876305222511, + 0.019907301291823387, + 0.013081059791147709, + 0.01892935484647751, + 0.008876985870301723, + 0.018551655113697052, + 0.025119954720139503, + 0.017794599756598473, + 0.026864858344197273, + 0.015255618840456009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.07434587925672531 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014234436675906181 + ], + "opposite_logprob_tokens": [ + -4.26423454284668 + ], + "logprob_reward": -0.014234436675906181, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014234436675906181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18168963491916656, + 0.1326608508825302, + 0.09549244493246078, + 0.038166869431734085, + 0.05020221322774887, + 0.08710109442472458, + 0.03964797407388687, + 0.09322168678045273, + 0.14540031552314758, + 0.136416956782341 + ], + "credit_map": [ + 0.025793135166168213, + 0.02460310235619545, + 0.02160770818591118, + 0.020648736506700516, + 0.02280273661017418, + 0.018690286204218864, + 0.02958856336772442, + 0.017955346032977104, + 0.016436681151390076, + 0.015269824303686619, + 0.009247426874935627, + 0.03608370199799538, + 0.022833706811070442, + 0.0061044711619615555, + 0.01807457022368908, + 0.008610449731349945, + 0.010888935998082161, + 0.010204882360994816, + 0.005121602676808834, + 0.01710907183587551, + 0.019983820617198944, + 0.007928241044282913, + 0.013315488584339619, + 0.010940393432974815, + 0.0048700859770178795, + 0.002879068488255143, + 0.004916988778859377, + 0.005654261913150549, + 0.005869690794497728, + 0.006215841043740511, + 0.004143385216593742, + 0.0036175413988530636, + 0.0076796310022473335, + 0.0047056362964212894, + 0.0025030127726495266, + 0.011111886240541935, + 0.007604543585330248, + 0.006644018460065126, + 0.005699828267097473, + 0.00425365474075079, + 0.016537265852093697, + 0.006122961174696684, + 0.006290947552770376, + 0.010568263940513134, + 0.011296689510345459, + 0.014345990493893623, + 0.013822042383253574, + 0.008116932585835457, + 0.004864310380071402, + 0.00401734234765172, + 0.005642713978886604, + 0.004605265334248543, + 0.006375058088451624, + 0.0047326418571174145, + 0.005655925255268812, + 0.0037547170650213957, + 0.01056422758847475, + 0.0071713970974087715, + 0.010012740269303322, + 0.015113214030861855, + 0.014100192114710808, + 0.007862590253353119, + 0.01408142875880003, + 0.014315895736217499, + 0.015720456838607788, + 0.011475345119833946, + 0.019316520541906357, + 0.015653744339942932, + 0.022693350911140442, + 0.016258087009191513, + 0.026435235515236855, + 0.017847564071416855, + 0.014081957750022411, + 0.015218407846987247, + 0.012201556004583836, + 0.019045155495405197, + 0.01862187497317791, + 0.01746252551674843, + 0.020341593772172928, + 0.019443880766630173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.014234436675906181 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642317563295364 + ], + "opposite_logprob_tokens": [ + -4.8826422691345215 + ], + "logprob_reward": -0.007642317563295364, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642317563295364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14737437665462494, + 0.09627474844455719, + 0.16083021461963654, + 0.05216275155544281, + 0.0356275737285614, + 0.16303576529026031, + 0.03193995729088783, + 0.07118096947669983, + 0.1218922883272171, + 0.11968143284320831 + ], + "credit_map": [ + 0.017988713458180428, + 0.02007977105677128, + 0.01933431811630726, + 0.020237406715750694, + 0.02165754698216915, + 0.013715865090489388, + 0.019718268886208534, + 0.01464247889816761, + 0.010890820994973183, + 0.015013390220701694, + 0.007925266399979591, + 0.021512284874916077, + 0.013656861148774624, + 0.00493780430406332, + 0.014890913851559162, + 0.007447399664670229, + 0.019795022904872894, + 0.0160282664000988, + 0.013423267751932144, + 0.04067634046077728, + 0.021379519253969193, + 0.019827770069241524, + 0.01838166080415249, + 0.011318343691527843, + 0.0060770828276872635, + 0.0047326297499239445, + 0.003091315273195505, + 0.014869640581309795, + 0.007388238795101643, + 0.004239085130393505, + 0.006635745987296104, + 0.005129009485244751, + 0.0057284110225737095, + 0.003509742207825184, + 0.0020326708909124136, + 0.00769271282479167, + 0.0053835767321288586, + 0.0052835578098893166, + 0.003522471757605672, + 0.002474429551512003, + 0.05159410461783409, + 0.007599540054798126, + 0.00874180719256401, + 0.011869062669575214, + 0.011084474623203278, + 0.04997679218649864, + 0.013829912059009075, + 0.00834005419164896, + 0.004054851830005646, + 0.0027265422977507114, + 0.004115923773497343, + 0.004885026719421148, + 0.005208210088312626, + 0.004921139217913151, + 0.003426266135647893, + 0.0026019925717264414, + 0.007915996015071869, + 0.005457149352878332, + 0.0076639545150101185, + 0.012615214101970196, + 0.009718180634081364, + 0.0053403340280056, + 0.010999557562172413, + 0.01147057581692934, + 0.011482015252113342, + 0.011669585481286049, + 0.01067543588578701, + 0.014663529582321644, + 0.024914046749472618, + 0.012178893201053143, + 0.024791186675429344, + 0.011517594568431377, + 0.011761005036532879, + 0.014498193748295307, + 0.01123020425438881, + 0.015274032019078732, + 0.01946987956762314, + 0.014981279149651527, + 0.022048380225896835, + 0.010418446734547615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.007642317563295364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941095411777496 + ], + "opposite_logprob_tokens": [ + -2.929941177368164 + ], + "logprob_reward": -0.054941095411777496, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941095411777496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056953780353069305, + 0.036670759320259094, + 0.057926151901483536, + 0.03658188134431839, + 0.025920843705534935, + 0.17332184314727783, + 0.056019410490989685, + 0.11650227755308151, + 0.12924893200397491, + 0.3108541667461395 + ], + "credit_map": [ + 0.007681576535105705, + 0.005497304257005453, + 0.012291298247873783, + 0.005648213438689709, + 0.008048204705119133, + 0.004794998560100794, + 0.007124844938516617, + 0.005867335479706526, + 0.004376165568828583, + 0.00435675447806716, + 0.005819335114210844, + 0.005834824405610561, + 0.0053116753697395325, + 0.003946349956095219, + 0.004119182005524635, + 0.0029064707923680544, + 0.008753487840294838, + 0.005447066854685545, + 0.005199078470468521, + 0.009603899903595448, + 0.007574727293103933, + 0.011101111769676208, + 0.005057111382484436, + 0.00518966605886817, + 0.005875007249414921, + 0.0031986278481781483, + 0.002994430484250188, + 0.004077311605215073, + 0.004166224505752325, + 0.008595290593802929, + 0.0036819856613874435, + 0.003993000369518995, + 0.004367115441709757, + 0.001993173733353615, + 0.0019375476986169815, + 0.004259854089468718, + 0.003652827348560095, + 0.0044525437988340855, + 0.0030603015329688787, + 0.0021974772680550814, + 0.0667768344283104, + 0.0062492964789271355, + 0.00614533293992281, + 0.005290650762617588, + 0.005874917842447758, + 0.0670168325304985, + 0.008399607613682747, + 0.007568359375, + 0.006517856847494841, + 0.005607596132904291, + 0.005720514804124832, + 0.009885980747640133, + 0.009713501669466496, + 0.006651837844401598, + 0.006770994048565626, + 0.0051511237397789955, + 0.016701428219676018, + 0.007554811425507069, + 0.010503559373319149, + 0.01653284952044487, + 0.017769478261470795, + 0.014486170373857021, + 0.016161059960722923, + 0.016792912036180496, + 0.01855958066880703, + 0.008947907947003841, + 0.021889768540859222, + 0.01207734178751707, + 0.013708537444472313, + 0.013830848038196564, + 0.0227824579924345, + 0.017452476546168327, + 0.034922175109386444, + 0.019159259274601936, + 0.028270794078707695, + 0.03440704569220543, + 0.022393062710762024, + 0.04427941516041756, + 0.04305119067430496, + 0.08437120914459229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.054941095411777496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a medium shot of an elderly couple standing side by side on a hillside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a light gray V-neck wool sweater with meticulously combed hair and a genuine smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a warm ochre cashmere shawl with her hair in an elegant silver bun and a gentle smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a level angle and slowly pan to the right, eventually panning out to reveal a huge blazing sunset entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm and dazzling anamorphic lens flare sweeping horizontally across the frame, creating a dreamy golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097281955182552 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097281955182552, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097281955182552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0742124393582344, + 0.08479168266057968, + 0.051811426877975464, + 0.02167479135096073, + 0.053589288145303726, + 0.06130412966012955, + 0.05135491117835045, + 0.1775234043598175, + 0.10684795677661896, + 0.31689000129699707 + ], + "credit_map": [ + 0.008867011405527592, + 0.007871229201555252, + 0.011163702234625816, + 0.014288796111941338, + 0.011064521037042141, + 0.006146908272057772, + 0.007002119440585375, + 0.007808153983205557, + 0.012019726447761059, + 0.008646137081086636, + 0.009361949749290943, + 0.0178902018815279, + 0.011720201000571251, + 0.007101308088749647, + 0.011937346309423447, + 0.006114809308201075, + 0.008141146041452885, + 0.005970408674329519, + 0.007331541273742914, + 0.006308066658675671, + 0.005292828194797039, + 0.011680785566568375, + 0.0031333542428910732, + 0.0039532966911792755, + 0.0024629300460219383, + 0.002857386600226164, + 0.002959997858852148, + 0.003505710046738386, + 0.002477724803611636, + 0.002903946442529559, + 0.0025638709776103497, + 0.001943225972354412, + 0.016039295122027397, + 0.002283018548041582, + 0.005471319891512394, + 0.00356479175388813, + 0.004241602495312691, + 0.01500090304762125, + 0.0033343732357025146, + 0.0036539859138429165, + 0.02032996155321598, + 0.002529395977035165, + 0.003112025326117873, + 0.003685673698782921, + 0.003293208312243223, + 0.020561091601848602, + 0.004062631167471409, + 0.0037301438860595226, + 0.006026809569448233, + 0.005617218092083931, + 0.005460785701870918, + 0.011932178400456905, + 0.006545472890138626, + 0.005298662465065718, + 0.006039457395672798, + 0.004434328526258469, + 0.021338757127523422, + 0.012691383250057697, + 0.015721790492534637, + 0.024731643497943878, + 0.027870452031493187, + 0.019434407353401184, + 0.026463337242603302, + 0.029271624982357025, + 0.013508418574929237, + 0.008753364905714989, + 0.010570745915174484, + 0.008970587514340878, + 0.011831088922917843, + 0.012464191764593124, + 0.011942772194743156, + 0.028806783258914948, + 0.0314369760453701, + 0.017448274418711662, + 0.03356299549341202, + 0.027751358225941658, + 0.029774682596325874, + 0.04149135574698448, + 0.04080374166369438, + 0.09462062269449234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.004097281955182552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_5/index_tensor([410, 410])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_5/index_tensor([410, 410])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7c5a13d17d13998554fdf63bb1f4bdc740c2b9b9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_5/index_tensor([410, 410])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, extreme wide shot captures a stunning timelapse of the iconic New York City skyline, seamlessly transitioning from late afternoon to late night. The scene begins with brilliant afternoon sunlight casting long, sharp shadows across the glass and steel skyscrapers that dot the city's grid. As the day progresses, the sun rapidly descends, bathing the entire cityscape in a warm golden hour glow; the buildings' western facades blaze in intense orange and yellow light, while their eastern shadows dramatically lengthen, cascading across canyons of concrete and glass. The sky transforms into a vibrant palette of deep orange, pink, and purple before fading into a cool, deep indigo dusk. It is in this moment that the city begins to awaken with renewed energy, and thousands of office and apartment windows begin to flicker, first as sparse points of light, then expanding into a magnificent, shimmering tapestry that redefines architectural form. The scene climaxes in the pitch black of night, the skyline a dazzling spectacle of white, yellow, and blue lights against a dark sky, the steady red pulse of aircraft warning beacons looming overhead on the tallest buildings, creating an atmosphere of majestic, never-ending urban energy.", + "seed": 28828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_5/index_410_seed_28828_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022688629943877457, + "vlm_logprob": -0.022688629943877457, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.022688629943877457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07284442335367203 + ], + "opposite_logprob_tokens": [ + -2.8228445053100586 + ], + "logprob_reward": -0.07284442335367203, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07284442335367203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10007699579000473, + 0.11680710315704346, + 0.05109133943915367, + 0.1693810671567917, + 0.07195967435836792, + 0.042996227741241455, + 0.061745841056108475, + 0.08536814898252487, + 0.1217198595404625, + 0.17885369062423706 + ], + "credit_map": [ + 0.011642195284366608, + 0.01088790874928236, + 0.014485557563602924, + 0.01183386892080307, + 0.012627637945115566, + 0.015623698942363262, + 0.01026044599711895, + 0.01271568238735199, + 0.01140736322849989, + 0.017160583287477493, + 0.009909539483487606, + 0.016738547012209892, + 0.014829371124505997, + 0.015851715579628944, + 0.01040079165250063, + 0.02050919458270073, + 0.005405886098742485, + 0.005809386260807514, + 0.008411242626607418, + 0.006289833690971136, + 0.006219730246812105, + 0.0061082132160663605, + 0.0060417973436415195, + 0.006805249955505133, + 0.02819911576807499, + 0.015819761902093887, + 0.03783236816525459, + 0.02484939619898796, + 0.01467852108180523, + 0.012114054523408413, + 0.019138071686029434, + 0.01674976758658886, + 0.008070128969848156, + 0.0052140555344522, + 0.01917574182152748, + 0.02144583687186241, + 0.004265869501978159, + 0.004736389499157667, + 0.0045690834522247314, + 0.0044825756922364235, + 0.006065695080906153, + 0.004526536911725998, + 0.0067216381430625916, + 0.007945962250232697, + 0.004278393927961588, + 0.004244249779731035, + 0.004530887119472027, + 0.0046828631311655045, + 0.007242608815431595, + 0.008175349794328213, + 0.00816248171031475, + 0.010361781343817711, + 0.0071877408772706985, + 0.006104893516749144, + 0.006600471679121256, + 0.007910509593784809, + 0.014865430071949959, + 0.012643746100366116, + 0.014920175075531006, + 0.011574131436645985, + 0.006016467232257128, + 0.0062454091385006905, + 0.009032076224684715, + 0.010070710442960262, + 0.01906132511794567, + 0.021225489675998688, + 0.018103374168276787, + 0.00891872402280569, + 0.013759060762822628, + 0.011614592745900154, + 0.01446472853422165, + 0.0145725654438138, + 0.02505386807024479, + 0.023412903770804405, + 0.02657194994390011, + 0.02427922748029232, + 0.022174693644046783, + 0.017842121422290802, + 0.01576407253742218, + 0.02375485561788082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.07284442335367203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033728066831827164 + ], + "opposite_logprob_tokens": [ + -3.4087281227111816 + ], + "logprob_reward": -0.033728066831827164, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033728066831827164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11824801564216614, + 0.11461687833070755, + 0.038013357669115067, + 0.09295260161161423, + 0.3417987525463104, + 0.03811218962073326, + 0.04692302271723747, + 0.03714965283870697, + 0.06911993026733398, + 0.10306555777788162 + ], + "credit_map": [ + 0.01256804820150137, + 0.013503454625606537, + 0.0229208804666996, + 0.020198892802000046, + 0.014108915813267231, + 0.01412472315132618, + 0.011364726349711418, + 0.009458369575440884, + 0.010723596438765526, + 0.017221368849277496, + 0.011371996253728867, + 0.013380019925534725, + 0.01808490790426731, + 0.020151207223534584, + 0.010214364156126976, + 0.013469411060214043, + 0.003396207233890891, + 0.007005962543189526, + 0.0037446499336510897, + 0.004431365989148617, + 0.0044287038035690784, + 0.004678578115999699, + 0.0034501187037676573, + 0.006877770181745291, + 0.017787497490644455, + 0.009223739616572857, + 0.018905872479081154, + 0.017173536121845245, + 0.00994046963751316, + 0.004892217461019754, + 0.007812264375388622, + 0.007217005826532841, + 0.027003690600395203, + 0.011553365737199783, + 0.11645340174436569, + 0.1295326203107834, + 0.010232041589915752, + 0.012029195204377174, + 0.0178869366645813, + 0.01710747368633747, + 0.005957809276878834, + 0.0035378995817154646, + 0.0058577596209943295, + 0.005884863436222076, + 0.0030339763034135103, + 0.004614683799445629, + 0.004620240069925785, + 0.00460495799779892, + 0.004316297825425863, + 0.00522250821813941, + 0.007285102270543575, + 0.011226939037442207, + 0.0036770417355000973, + 0.0027619486209005117, + 0.0054075163789093494, + 0.0070256697945296764, + 0.005189519841223955, + 0.00543089909479022, + 0.006281082984060049, + 0.0033968999050557613, + 0.003404504619538784, + 0.00456638541072607, + 0.003215707838535309, + 0.005664654076099396, + 0.011250937357544899, + 0.00782725214958191, + 0.011631611734628677, + 0.004762259311974049, + 0.005830351263284683, + 0.009352019056677818, + 0.006586345378309488, + 0.011879155412316322, + 0.018983427435159683, + 0.009676114656031132, + 0.01711171865463257, + 0.006650973577052355, + 0.013318772427737713, + 0.016553539782762527, + 0.006399642210453749, + 0.014371359720826149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.033728066831827164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010524216340854764 + ], + "opposite_logprob_tokens": [ + -6.876052379608154 + ], + "logprob_reward": -0.0010524216340854764, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010524216340854764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08334814757108688, + 0.07217821478843689, + 0.024478670209646225, + 0.15694767236709595, + 0.2937093675136566, + 0.03960255905985832, + 0.034144025295972824, + 0.07980157434940338, + 0.05820855125784874, + 0.15758119523525238 + ], + "credit_map": [ + 0.009194833226501942, + 0.008752635680139065, + 0.010736502707004547, + 0.007131443824619055, + 0.011135484091937542, + 0.014900384470820427, + 0.006298370193690062, + 0.015198497101664543, + 0.0069129010662436485, + 0.012374545447528362, + 0.004728053696453571, + 0.00873787235468626, + 0.011629428714513779, + 0.01132564339786768, + 0.007299295626580715, + 0.009170474484562874, + 0.0018838789546862245, + 0.0041244011372327805, + 0.0030950901564210653, + 0.0033476995304226875, + 0.0035685538314282894, + 0.0026510399766266346, + 0.002844055648893118, + 0.002963949227705598, + 0.042057640850543976, + 0.010279358364641666, + 0.053859174251556396, + 0.011230968870222569, + 0.00819876417517662, + 0.009881320409476757, + 0.011349649168550968, + 0.010090788826346397, + 0.02031179703772068, + 0.011947769671678543, + 0.10319128632545471, + 0.11380339413881302, + 0.009407146833837032, + 0.011643171310424805, + 0.013348128646612167, + 0.01005665585398674, + 0.004746466875076294, + 0.005028323270380497, + 0.0056080580689013, + 0.007827376015484333, + 0.0035933726467192173, + 0.003663318697363138, + 0.004996048752218485, + 0.0041395933367311954, + 0.00503199826925993, + 0.004426199477165937, + 0.004440004006028175, + 0.004818576853722334, + 0.0035004985984414816, + 0.0036812215112149715, + 0.004490813706070185, + 0.00375471543520689, + 0.01807638816535473, + 0.010176390409469604, + 0.01998666487634182, + 0.007322854828089476, + 0.005430659744888544, + 0.0060894060879945755, + 0.00676902336999774, + 0.005950184538960457, + 0.008760339580476284, + 0.007300100289285183, + 0.006465396378189325, + 0.004367652349174023, + 0.007091132923960686, + 0.00946142990142107, + 0.007340034004300833, + 0.007422465831041336, + 0.028286974877119064, + 0.014390625059604645, + 0.018009673804044724, + 0.010516174137592316, + 0.018209610134363174, + 0.03952327370643616, + 0.007853249087929726, + 0.020791610702872276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.0010524216340854764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104523919522762 + ], + "opposite_logprob_tokens": [ + -5.5041046142578125 + ], + "logprob_reward": -0.004104523919522762, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104523919522762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09045159816741943, + 0.1126231849193573, + 0.07754068821668625, + 0.21624425053596497, + 0.1930564045906067, + 0.05655945464968681, + 0.03765426576137543, + 0.06054314598441124, + 0.05881962925195694, + 0.09650736302137375 + ], + "credit_map": [ + 0.012706662528216839, + 0.008691485971212387, + 0.015592820942401886, + 0.00844552181661129, + 0.009058481082320213, + 0.015939444303512573, + 0.007201731204986572, + 0.012815458700060844, + 0.008738420903682709, + 0.01329098455607891, + 0.008054216392338276, + 0.010965926572680473, + 0.029568804427981377, + 0.016754761338233948, + 0.009228860028088093, + 0.01602121628820896, + 0.007176260929554701, + 0.010897495783865452, + 0.00840536504983902, + 0.009325466118752956, + 0.01103854551911354, + 0.012264484539628029, + 0.008593430742621422, + 0.009839647449553013, + 0.02836243435740471, + 0.0237155519425869, + 0.027722693979740143, + 0.0396433025598526, + 0.021387625485658646, + 0.020741373300552368, + 0.026704832911491394, + 0.0279664508998394, + 0.02412215806543827, + 0.013291141018271446, + 0.05014629662036896, + 0.04831171780824661, + 0.013059000484645367, + 0.017152972519397736, + 0.013248047791421413, + 0.013725093565881252, + 0.008802173659205437, + 0.006224981043487787, + 0.009101596660912037, + 0.009651850908994675, + 0.005434689112007618, + 0.005858818534761667, + 0.006406206171959639, + 0.005079141352325678, + 0.004789679776877165, + 0.005126649048179388, + 0.006497885100543499, + 0.005687180440872908, + 0.003288965206593275, + 0.0028210096061229706, + 0.004858384840190411, + 0.0045845163986086845, + 0.011899877339601517, + 0.008587912656366825, + 0.013454108498990536, + 0.006057895720005035, + 0.004482869058847427, + 0.004768432583659887, + 0.005119991023093462, + 0.00617206422612071, + 0.009512406773865223, + 0.007731002289801836, + 0.008773078210651875, + 0.004645956680178642, + 0.005959207192063332, + 0.007813943549990654, + 0.006274744868278503, + 0.008109288290143013, + 0.013504840433597565, + 0.011298324912786484, + 0.01915442757308483, + 0.00845468882471323, + 0.01217876747250557, + 0.011227075941860676, + 0.006515796761959791, + 0.014173444360494614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.004104523919522762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017137139802798629 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017137139802798629, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017137139802798629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1259596049785614, + 0.13204863667488098, + 0.0383230559527874, + 0.10285937041044235, + 0.12636026740074158, + 0.05752937123179436, + 0.052618350833654404, + 0.08217243105173111, + 0.10935122519731522, + 0.1727776676416397 + ], + "credit_map": [ + 0.014975412748754025, + 0.012689250521361828, + 0.022081373259425163, + 0.015640240162611008, + 0.015490471385419369, + 0.018960263580083847, + 0.009992449544370174, + 0.01613014191389084, + 0.01100347563624382, + 0.02124778740108013, + 0.0075321197509765625, + 0.01379142701625824, + 0.02734980173408985, + 0.021047841757535934, + 0.014341771602630615, + 0.015734415501356125, + 0.0035645836032927036, + 0.004476974252611399, + 0.0036004565190523863, + 0.005236511584371328, + 0.006899264641106129, + 0.004579784814268351, + 0.0034227599389851093, + 0.0065427194349467754, + 0.027253296226263046, + 0.0061594112776219845, + 0.037596870213747025, + 0.007121425587683916, + 0.00646997569128871, + 0.006239297334104776, + 0.0077359420247375965, + 0.004283146467059851, + 0.012987504713237286, + 0.005384383257478476, + 0.042704276740550995, + 0.04230516031384468, + 0.004672059323638678, + 0.006936318706721067, + 0.006235840730369091, + 0.005134733393788338, + 0.012156124226748943, + 0.007000873796641827, + 0.007726339157670736, + 0.008380172774195671, + 0.005811687093228102, + 0.006025864277034998, + 0.005905099678784609, + 0.004523209761828184, + 0.007362836040556431, + 0.006960350088775158, + 0.007376173045486212, + 0.007092600222676992, + 0.004955689888447523, + 0.004370071925222874, + 0.007138697896152735, + 0.007361929398030043, + 0.01902555488049984, + 0.011432578787207603, + 0.019390098750591278, + 0.007160672452300787, + 0.0058951182290911674, + 0.006456349510699511, + 0.005696943961083889, + 0.007115119602531195, + 0.017267512157559395, + 0.012186447158455849, + 0.01371924951672554, + 0.005259265657514334, + 0.011254347860813141, + 0.016210293397307396, + 0.012125001288950443, + 0.021329116076231003, + 0.03210671991109848, + 0.015967080369591713, + 0.019677067175507545, + 0.011275248602032661, + 0.02048666588962078, + 0.039918649941682816, + 0.009271964430809021, + 0.024074256420135498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.0017137139802798629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme wide shot captures a stunning timelapse of the iconic New York City skyline, seamlessly transitioning from late afternoon to late night. The scene begins with brilliant afternoon sunlight casting long, sharp shadows across the glass and steel skyscrapers that dot the city's grid. As the day progresses, the sun rapidly descends, bathing the entire cityscape in a warm golden hour glow; the buildings' western facades blaze in intense orange and yellow light, while their eastern shadows dramatically lengthen, cascading across canyons of concrete and glass. The sky transforms into a vibrant palette of deep orange, pink, and purple before fading into a cool, deep indigo dusk. It is in this moment that the city begins to awaken with renewed energy, and thousands of office and apartment windows begin to flicker, first as sparse points of light, then expanding into a magnificent, shimmering tapestry that redefines architectural form. The scene climaxes in the pitch black of night, the skyline a dazzling spectacle of white, yellow, and blue lights against a dark sky, the steady red pulse of aircraft warning beacons looming overhead on the tallest buildings, creating an atmosphere of majestic, never-ending urban energy.", + "seed": 28829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_5/index_410_seed_28829_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020807960187084973, + "vlm_logprob": -0.020807960187084973, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.020807960187084973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03065582364797592 + ], + "opposite_logprob_tokens": [ + -4.030655860900879 + ], + "logprob_reward": -0.03065582364797592, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03065582364797592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10832152515649796, + 0.08878424763679504, + 0.18967504799365997, + 0.12432164698839188, + 0.049565430730581284, + 0.04601512849330902, + 0.05555534362792969, + 0.038691189140081406, + 0.22102074325084686, + 0.07804974168539047 + ], + "credit_map": [ + 0.00845959223806858, + 0.010703048668801785, + 0.014530007727444172, + 0.012194199487566948, + 0.02386258728802204, + 0.009069094434380531, + 0.0185199324041605, + 0.010983053594827652, + 0.006147075444459915, + 0.006636789999902248, + 0.010418626479804516, + 0.007304486818611622, + 0.014928219839930534, + 0.011117245070636272, + 0.022918259724974632, + 0.009313542395830154, + 0.016018755733966827, + 0.0313510037958622, + 0.026074694469571114, + 0.025430670008063316, + 0.029998818412423134, + 0.021094994619488716, + 0.01938529498875141, + 0.020320812240242958, + 0.00805389229208231, + 0.009374966844916344, + 0.018121566623449326, + 0.028404133394360542, + 0.01511971466243267, + 0.013900882564485073, + 0.01663786545395851, + 0.01470861490815878, + 0.004733905661851168, + 0.004046356305480003, + 0.0052968962118029594, + 0.005266683641821146, + 0.00805622898042202, + 0.009077692404389381, + 0.006564839277416468, + 0.006522823590785265, + 0.004057421814650297, + 0.0050066988915205, + 0.006459021009504795, + 0.005900091957300901, + 0.00546141155064106, + 0.00512157566845417, + 0.00707232067361474, + 0.006936587858945131, + 0.006717142183333635, + 0.005977050866931677, + 0.008254723623394966, + 0.009173993021249771, + 0.006406542845070362, + 0.005127667915076017, + 0.008776272647082806, + 0.00512195099145174, + 0.0065922182984650135, + 0.0030589464586228132, + 0.004534071311354637, + 0.005679893307387829, + 0.003345837350934744, + 0.004126305226236582, + 0.007686297409236431, + 0.0036676148883998394, + 0.023822566494345665, + 0.02225104160606861, + 0.036599740386009216, + 0.03055151365697384, + 0.024957681074738503, + 0.019164958968758583, + 0.0257109384983778, + 0.03796227648854256, + 0.00856807641685009, + 0.0086224265396595, + 0.0143532520160079, + 0.008942091837525368, + 0.00812853779643774, + 0.008437390439212322, + 0.009565119631588459, + 0.011432845145463943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.03065582364797592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208525598049164 + ], + "opposite_logprob_tokens": [ + -2.8120851516723633 + ], + "logprob_reward": -0.06208525598049164, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208525598049164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10733268409967422, + 0.11633012443780899, + 0.05253863334655762, + 0.15217085182666779, + 0.10812588781118393, + 0.13234448432922363, + 0.05050507187843323, + 0.08070135116577148, + 0.15025237202644348, + 0.04969857633113861 + ], + "credit_map": [ + 0.011656016111373901, + 0.00908348336815834, + 0.014592813327908516, + 0.009829114191234112, + 0.020121706649661064, + 0.009070262312889099, + 0.021111752837896347, + 0.011867544613778591, + 0.006943506188690662, + 0.004399921745061874, + 0.025039343163371086, + 0.008822349831461906, + 0.011843008920550346, + 0.009033729322254658, + 0.04100854694843292, + 0.00923971552401781, + 0.005730015225708485, + 0.0052766744047403336, + 0.007578645832836628, + 0.0067088729701936245, + 0.0053827459923923016, + 0.00683357659727335, + 0.007687466684728861, + 0.007340637501329184, + 0.013591510243713856, + 0.010064863599836826, + 0.019765036180615425, + 0.02416377142071724, + 0.018095893785357475, + 0.029095645993947983, + 0.01881590113043785, + 0.01857822760939598, + 0.010877585038542747, + 0.007052265573292971, + 0.015769904479384422, + 0.009201698005199432, + 0.012370621785521507, + 0.028601640835404396, + 0.009601141326129436, + 0.014651022851467133, + 0.0171792171895504, + 0.008682969026267529, + 0.019370906054973602, + 0.011907704174518585, + 0.014381407760083675, + 0.02728123404085636, + 0.014275423251092434, + 0.019265620037913322, + 0.006691471673548222, + 0.004885106813162565, + 0.0067540062591433525, + 0.006071694660931826, + 0.005247997585684061, + 0.00771370017901063, + 0.0076697939075529575, + 0.0054713040590286255, + 0.007471649441868067, + 0.011425412259995937, + 0.010651154443621635, + 0.015470023266971111, + 0.008149919100105762, + 0.005790675524622202, + 0.012204037979245186, + 0.009538478218019009, + 0.012888665311038494, + 0.016707690432667732, + 0.02119608409702778, + 0.019171176478266716, + 0.017318768426775932, + 0.01130178477615118, + 0.02729509398341179, + 0.024373114109039307, + 0.007135599385946989, + 0.006007877644151449, + 0.007003411650657654, + 0.004726429935544729, + 0.007023812271654606, + 0.006721664220094681, + 0.005514045245945454, + 0.005565738305449486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.06208525598049164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017233534017577767 + ], + "opposite_logprob_tokens": [ + -6.376723289489746 + ], + "logprob_reward": -0.0017233534017577767, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017233534017577767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27812957763671875, + 0.06345682591199875, + 0.07971738278865814, + 0.1465177834033966, + 0.08472362905740738, + 0.07268864661455154, + 0.05016307160258293, + 0.033255137503147125, + 0.14970551431179047, + 0.041642505675554276 + ], + "credit_map": [ + 0.020023798570036888, + 0.02797059528529644, + 0.037931423634290695, + 0.028273506090044975, + 0.06592918932437897, + 0.02542913891375065, + 0.042936019599437714, + 0.029635878279805183, + 0.004879583138972521, + 0.005488886497914791, + 0.00964601431041956, + 0.005890154745429754, + 0.007717952132225037, + 0.0064690071158111095, + 0.01720382086932659, + 0.006161401513963938, + 0.0072369505651295185, + 0.010573362931609154, + 0.01157588604837656, + 0.011027969419956207, + 0.008318133652210236, + 0.006627078168094158, + 0.012755022384226322, + 0.011602972634136677, + 0.011636278592050076, + 0.014992599375545979, + 0.01881856471300125, + 0.025233427062630653, + 0.025749322026968002, + 0.013324388302862644, + 0.020543891936540604, + 0.016219306737184525, + 0.008058037608861923, + 0.006852157413959503, + 0.013169393874704838, + 0.00979684665799141, + 0.009137618355453014, + 0.008496103808283806, + 0.01925772987306118, + 0.009955726563930511, + 0.009227775037288666, + 0.006908841896802187, + 0.010604476556181908, + 0.008976059965789318, + 0.007040468044579029, + 0.009582817554473877, + 0.011064975522458553, + 0.0092832176014781, + 0.006043822504580021, + 0.0036970828659832478, + 0.00831163302063942, + 0.006888281553983688, + 0.0043248943984508514, + 0.009075377136468887, + 0.0061364369466900826, + 0.00568554038181901, + 0.003187986323609948, + 0.004246243741363287, + 0.004185330588370562, + 0.005266045220196247, + 0.003575744340196252, + 0.0033466401509940624, + 0.0056154741905629635, + 0.003831669455394149, + 0.014891326427459717, + 0.017514487728476524, + 0.02411247231066227, + 0.016669945791363716, + 0.0218028761446476, + 0.01541963778436184, + 0.01746276393532753, + 0.021831989288330078, + 0.006018692161887884, + 0.0044629620388150215, + 0.0065243360586464405, + 0.0033086640760302544, + 0.004760327748954296, + 0.0035397454630583525, + 0.0033438752871006727, + 0.009683899581432343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0017233534017577767 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036245384253561497 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.0036245384253561497, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036245384253561497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13959179818630219, + 0.044445667415857315, + 0.0892055407166481, + 0.1737760603427887, + 0.1392352432012558, + 0.09247394651174545, + 0.06783300638198853, + 0.06585249304771423, + 0.10341846197843552, + 0.08416781574487686 + ], + "credit_map": [ + 0.01573110930621624, + 0.017436549067497253, + 0.018720317631959915, + 0.01807519420981407, + 0.02289879508316517, + 0.015067900530993938, + 0.0156999584287405, + 0.015961961820721626, + 0.0035915328189730644, + 0.003975132014602423, + 0.009391131810843945, + 0.0035639472771435976, + 0.005090204533189535, + 0.003228109097108245, + 0.012003457173705101, + 0.003602151293307543, + 0.009676143527030945, + 0.009520166553556919, + 0.010537837632000446, + 0.01244660560041666, + 0.013711328618228436, + 0.012067543342709541, + 0.00873878039419651, + 0.012507122941315174, + 0.012927812524139881, + 0.010583566501736641, + 0.022471731528639793, + 0.031028367578983307, + 0.020518241450190544, + 0.031467173248529434, + 0.02544429339468479, + 0.01933487132191658, + 0.009990389458835125, + 0.009434403851628304, + 0.02002379298210144, + 0.014303408563137054, + 0.0156893040984869, + 0.017585378140211105, + 0.0362139493227005, + 0.015994610264897346, + 0.011462219059467316, + 0.00510271405801177, + 0.012675333768129349, + 0.007327963598072529, + 0.01250387728214264, + 0.021551206707954407, + 0.009492757730185986, + 0.01235786359757185, + 0.006376476958394051, + 0.008345822803676128, + 0.009420538321137428, + 0.008666343055665493, + 0.007606337778270245, + 0.006556002888828516, + 0.010859238915145397, + 0.01000223122537136, + 0.005930656101554632, + 0.00568949431180954, + 0.008400698192417622, + 0.012112594209611416, + 0.007721306756138802, + 0.0061795334331691265, + 0.010235782712697983, + 0.009582419879734516, + 0.008886011317372322, + 0.012965398840606213, + 0.014018652029335499, + 0.012269464321434498, + 0.015043983235955238, + 0.00923224724829197, + 0.016419697552919388, + 0.014582997187972069, + 0.007895438000559807, + 0.008052422665059566, + 0.015263085253536701, + 0.012297533452510834, + 0.010600448586046696, + 0.0066324444487690926, + 0.011211439967155457, + 0.01221500150859356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0036245384253561497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17124724388122559, + 0.047032590955495834, + 0.08835722506046295, + 0.06111836060881615, + 0.17469678819179535, + 0.08818159252405167, + 0.05286790430545807, + 0.051139313727617264, + 0.2071799486875534, + 0.05817903205752373 + ], + "credit_map": [ + 0.012947519309818745, + 0.02008916065096855, + 0.0252669807523489, + 0.018719203770160675, + 0.0341002531349659, + 0.016048632562160492, + 0.025681277737021446, + 0.018394213169813156, + 0.006573129445314407, + 0.005066659767180681, + 0.005438348278403282, + 0.004370211623609066, + 0.004999689757823944, + 0.008068486116826534, + 0.007466798648238182, + 0.005049268249422312, + 0.009856915101408958, + 0.0109781539067626, + 0.00973697192966938, + 0.011420861817896366, + 0.011459166184067726, + 0.0154897291213274, + 0.01027616299688816, + 0.009139270521700382, + 0.005729310214519501, + 0.004593024030327797, + 0.008970042690634727, + 0.009027790278196335, + 0.008957401849329472, + 0.00840598251670599, + 0.008567658253014088, + 0.006867154501378536, + 0.024954942986369133, + 0.010318011976778507, + 0.022130386903882027, + 0.01638125255703926, + 0.01939016953110695, + 0.04116370901465416, + 0.01676301471889019, + 0.023595314472913742, + 0.016907116398215294, + 0.005385327152907848, + 0.01284777745604515, + 0.008067627437412739, + 0.010409422218799591, + 0.014147105626761913, + 0.01082230918109417, + 0.009594914503395557, + 0.005444367881864309, + 0.004849082324653864, + 0.007690764032304287, + 0.006441309582442045, + 0.004788902588188648, + 0.007622388191521168, + 0.0076743923127651215, + 0.008356700651347637, + 0.004141951911151409, + 0.005019082222133875, + 0.00618445361033082, + 0.009784539230167866, + 0.007115284912288189, + 0.004718306474387646, + 0.007319859694689512, + 0.006855840794742107, + 0.019860051572322845, + 0.02816755324602127, + 0.028846843168139458, + 0.021391233429312706, + 0.031700409948825836, + 0.02184552326798439, + 0.026057546958327293, + 0.029310815036296844, + 0.006732146255671978, + 0.005784842651337385, + 0.01130552776157856, + 0.007381217088550329, + 0.007089028134942055, + 0.004686967935413122, + 0.0069602238945662975, + 0.008239082992076874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.005950829479843378 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme wide shot captures a stunning timelapse of the iconic New York City skyline, seamlessly transitioning from late afternoon to late night. The scene begins with brilliant afternoon sunlight casting long, sharp shadows across the glass and steel skyscrapers that dot the city's grid. As the day progresses, the sun rapidly descends, bathing the entire cityscape in a warm golden hour glow; the buildings' western facades blaze in intense orange and yellow light, while their eastern shadows dramatically lengthen, cascading across canyons of concrete and glass. The sky transforms into a vibrant palette of deep orange, pink, and purple before fading into a cool, deep indigo dusk. It is in this moment that the city begins to awaken with renewed energy, and thousands of office and apartment windows begin to flicker, first as sparse points of light, then expanding into a magnificent, shimmering tapestry that redefines architectural form. The scene climaxes in the pitch black of night, the skyline a dazzling spectacle of white, yellow, and blue lights against a dark sky, the steady red pulse of aircraft warning beacons looming overhead on the tallest buildings, creating an atmosphere of majestic, never-ending urban energy.", + "seed": 28830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_5/index_410_seed_28830_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025910275138448922, + "vlm_logprob": -0.025910275138448922, + "vlm_yes_no_margin": 4.849999952316284, + "vlm_token_credit": -0.025910275138448922, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030676979571580887 + ], + "opposite_logprob_tokens": [ + -4.03067684173584 + ], + "logprob_reward": -0.030676979571580887, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.030676979571580887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06000986322760582, + 0.053087268024683, + 0.3481403887271881, + 0.04400075599551201, + 0.041654206812381744, + 0.04610580950975418, + 0.03330056741833687, + 0.03420465812087059, + 0.060166098177433014, + 0.27933037281036377 + ], + "credit_map": [ + 0.00441714096814394, + 0.005204035434871912, + 0.006532040424644947, + 0.006935256067663431, + 0.005118325352668762, + 0.005359855480492115, + 0.013367880135774612, + 0.013075334951281548, + 0.005268768407404423, + 0.005427592899650335, + 0.010993575677275658, + 0.004151980858296156, + 0.00358207686804235, + 0.007860779762268066, + 0.007218798156827688, + 0.00858369655907154, + 0.023022092878818512, + 0.009828250855207443, + 0.12770965695381165, + 0.014211502857506275, + 0.008739486336708069, + 0.017574749886989594, + 0.01165839098393917, + 0.13539625704288483, + 0.002295452868565917, + 0.0018832137575373054, + 0.0064457315020263195, + 0.005921412259340286, + 0.002080080332234502, + 0.0031441834289580584, + 0.010576792992651463, + 0.011653895489871502, + 0.0040031857788562775, + 0.004074060823768377, + 0.009333228692412376, + 0.004368038382381201, + 0.0029508299194276333, + 0.0037277666851878166, + 0.004658113699406385, + 0.008538988418877125, + 0.005944191012531519, + 0.0034332734066993, + 0.005710945464670658, + 0.0063295867294073105, + 0.006554568186402321, + 0.0059952507726848125, + 0.0064260573126375675, + 0.0057119387201964855, + 0.004031633026897907, + 0.0037673888728022575, + 0.0039689322002232075, + 0.0057978215627372265, + 0.005628741811960936, + 0.002819043816998601, + 0.0037137228064239025, + 0.0035732861142605543, + 0.009220564737915993, + 0.0023999959230422974, + 0.009415138512849808, + 0.004528315272182226, + 0.0022645751014351845, + 0.0015445470344275236, + 0.003386747557669878, + 0.0014447715366259217, + 0.006349167786538601, + 0.00690448097884655, + 0.007951364852488041, + 0.006858563981950283, + 0.008033200167119503, + 0.007397881709039211, + 0.010826212354004383, + 0.005845227744430304, + 0.024083686992526054, + 0.11746624857187271, + 0.08930590748786926, + 0.013509029522538185, + 0.012239065021276474, + 0.01207165140658617, + 0.005730258300900459, + 0.004924548789858818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.030676979571580887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906476199626923 + ], + "opposite_logprob_tokens": [ + -2.464064836502075 + ], + "logprob_reward": -0.08906476199626923, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906476199626923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20655229687690735, + 0.03025975450873375, + 0.25177282094955444, + 0.03929360210895538, + 0.07000049948692322, + 0.05959360674023628, + 0.02741321176290512, + 0.09376182407140732, + 0.0396893247961998, + 0.1816631257534027 + ], + "credit_map": [ + 0.01491350308060646, + 0.017289070412516594, + 0.026336422190070152, + 0.014983681961894035, + 0.026597393676638603, + 0.048055440187454224, + 0.025397030636668205, + 0.03297971561551094, + 0.0022513887379318476, + 0.00384469423443079, + 0.005747050978243351, + 0.002551750745624304, + 0.003002147888764739, + 0.004205917939543724, + 0.003485976019874215, + 0.005170824937522411, + 0.015831124037504196, + 0.007978171110153198, + 0.09018788486719131, + 0.011244896799325943, + 0.006894320715218782, + 0.015135355293750763, + 0.009792889468371868, + 0.09470813721418381, + 0.003099267603829503, + 0.002886308589950204, + 0.006470595020800829, + 0.007006406784057617, + 0.0027949188370257616, + 0.0050336155109107494, + 0.007218468468636274, + 0.0047840154729783535, + 0.011569674126803875, + 0.00962680485099554, + 0.011758650653064251, + 0.010642778128385544, + 0.005180862732231617, + 0.0054217916913330555, + 0.005102328024804592, + 0.010697595775127411, + 0.008835026994347572, + 0.005945220123976469, + 0.007145623676478863, + 0.007151128724217415, + 0.011831657961010933, + 0.009614767506718636, + 0.004200192168354988, + 0.004869979340583086, + 0.0034609022550284863, + 0.002956413896754384, + 0.0030826351139694452, + 0.0034337148535996675, + 0.007199340499937534, + 0.002944265492260456, + 0.0023993130307644606, + 0.0019366221968084574, + 0.03173990547657013, + 0.004985319450497627, + 0.033910367637872696, + 0.011110401712357998, + 0.002847240073606372, + 0.0031009307131171227, + 0.0032795837614685297, + 0.002788063371554017, + 0.0036522045265883207, + 0.004189891275018454, + 0.005195193458348513, + 0.0048476215451955795, + 0.004514241125434637, + 0.0050826771184802055, + 0.007441523484885693, + 0.004765963181853294, + 0.014142734929919243, + 0.06895442306995392, + 0.059664759784936905, + 0.011024268344044685, + 0.006819578353315592, + 0.00849500484764576, + 0.007175033446401358, + 0.005387298762798309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.08906476199626923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009303532424382865 + ], + "opposite_logprob_tokens": [ + -7.000930309295654 + ], + "logprob_reward": -0.0009303532424382865, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009303532424382865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13989757001399994, + 0.054003093391656876, + 0.2422323226928711, + 0.034595370292663574, + 0.03310544788837433, + 0.0485483817756176, + 0.028293929994106293, + 0.08302228897809982, + 0.08106520771980286, + 0.25523635745048523 + ], + "credit_map": [ + 0.00909279752522707, + 0.01127808727324009, + 0.01665012538433075, + 0.013487218879163265, + 0.01868552342057228, + 0.031781964004039764, + 0.016658281907439232, + 0.02226356789469719, + 0.00450783409178257, + 0.004568183328956366, + 0.009331172332167625, + 0.004764389246702194, + 0.0033897156827151775, + 0.012784143909811974, + 0.005913722328841686, + 0.008743931539356709, + 0.014261560514569283, + 0.007265086751431227, + 0.08581940829753876, + 0.011483166366815567, + 0.006990884430706501, + 0.019167207181453705, + 0.010063208639621735, + 0.08718179911375046, + 0.0022025825455784798, + 0.0020377140026539564, + 0.004840765614062548, + 0.00304672890342772, + 0.002135959453880787, + 0.0022593883331865072, + 0.009881483390927315, + 0.008190746419131756, + 0.003980773966759443, + 0.00400684354826808, + 0.006738731171935797, + 0.00311226025223732, + 0.0026055118069052696, + 0.0031451524700969458, + 0.003591070184484124, + 0.005925102159380913, + 0.005057518370449543, + 0.003713140031322837, + 0.00630456255748868, + 0.00516982888802886, + 0.006903518456965685, + 0.006293232087045908, + 0.009255991317331791, + 0.005850587040185928, + 0.0036348288413137197, + 0.0027489534113556147, + 0.0029453476890921593, + 0.0036468571051955223, + 0.007428497076034546, + 0.0029426850378513336, + 0.0026094329077750444, + 0.0023373286239802837, + 0.024102432653307915, + 0.006273011676967144, + 0.027649428695440292, + 0.009960951283574104, + 0.0029172718059271574, + 0.003666162956506014, + 0.0049310107715427876, + 0.0035220191348344088, + 0.008700167760252953, + 0.010763256810605526, + 0.011766567826271057, + 0.008846085518598557, + 0.010050461627542973, + 0.012072354555130005, + 0.012748923152685165, + 0.00611739419400692, + 0.02092048153281212, + 0.10619284957647324, + 0.08396518975496292, + 0.014046802185475826, + 0.00965573638677597, + 0.010498644784092903, + 0.004495507571846247, + 0.005461160559207201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.0009303532424382865 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624063218012452 + ], + "opposite_logprob_tokens": [ + -5.628623962402344 + ], + "logprob_reward": -0.003624063218012452, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624063218012452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11248498409986496, + 0.04547824710607529, + 0.2788807451725006, + 0.031154152005910873, + 0.04204864427447319, + 0.024004371836781502, + 0.032997455447912216, + 0.09728261083364487, + 0.03085898980498314, + 0.3048098087310791 + ], + "credit_map": [ + 0.008264278061687946, + 0.0077170091681182384, + 0.012767119333148003, + 0.00932373944669962, + 0.01312139630317688, + 0.024508286267518997, + 0.01509913057088852, + 0.02168401889503002, + 0.0030471214558929205, + 0.005052171181887388, + 0.007681845221668482, + 0.0025503283832222223, + 0.003306521801277995, + 0.00979750044643879, + 0.0052186730317771435, + 0.008824086748063564, + 0.01973845064640045, + 0.008250099606812, + 0.09327178448438644, + 0.01598353311419487, + 0.009608457796275616, + 0.023718446493148804, + 0.014084064401686192, + 0.09422589838504791, + 0.0021909582428634167, + 0.00184789695776999, + 0.004275130573660135, + 0.0023066357243806124, + 0.0019518571207299829, + 0.002981238765642047, + 0.009293489158153534, + 0.0063069467432796955, + 0.005509707611054182, + 0.005219032522290945, + 0.007930981926620007, + 0.005313938949257135, + 0.0037707947194576263, + 0.0035806866362690926, + 0.003170605516061187, + 0.007552896626293659, + 0.003899397561326623, + 0.0020200160797685385, + 0.0032261619344353676, + 0.002609189134091139, + 0.002891794079914689, + 0.003147510578855872, + 0.0034019139129668474, + 0.002808388788253069, + 0.0042397077195346355, + 0.004373098723590374, + 0.0038626338355243206, + 0.005606155842542648, + 0.00402951380237937, + 0.0027700704522430897, + 0.004417293705046177, + 0.003698980202898383, + 0.031410861760377884, + 0.005270898807793856, + 0.03550035506486893, + 0.012007676996290684, + 0.002652812283486128, + 0.0030168474186211824, + 0.004395709373056889, + 0.0030274451710283756, + 0.0045553469099104404, + 0.003890726249665022, + 0.004896968603134155, + 0.0030788183212280273, + 0.0038545681163668633, + 0.004925260785967112, + 0.003320762189105153, + 0.0023365390952676535, + 0.023475676774978638, + 0.12204627692699432, + 0.1056024432182312, + 0.017916783690452576, + 0.010660401545464993, + 0.013395344838500023, + 0.0055595142766833305, + 0.006153378169983625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.003624063218012452 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052552176639437675 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.0052552176639437675, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052552176639437675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17684327065944672, + 0.04725023731589317, + 0.2594420909881592, + 0.055769823491573334, + 0.06270820647478104, + 0.06448668986558914, + 0.03576533496379852, + 0.11180426180362701, + 0.0644688829779625, + 0.1214611828327179 + ], + "credit_map": [ + 0.01561366580426693, + 0.011098664253950119, + 0.019599374383687973, + 0.012480341829359531, + 0.02161192148923874, + 0.03968162462115288, + 0.026478642597794533, + 0.03027902916073799, + 0.0041930763982236385, + 0.006069732829928398, + 0.009954195469617844, + 0.002543199574574828, + 0.0035948222503066063, + 0.007087741978466511, + 0.006031256169080734, + 0.007776213344186544, + 0.017197299748659134, + 0.009364735335111618, + 0.09337479621171951, + 0.009182617999613285, + 0.006323952227830887, + 0.01755440980195999, + 0.009038557298481464, + 0.09740572422742844, + 0.0031181012745946646, + 0.0036831528414040804, + 0.007483653724193573, + 0.007133528124541044, + 0.0028979405760765076, + 0.0038395316805690527, + 0.014737922698259354, + 0.012875991873443127, + 0.010543475858867168, + 0.007201059255748987, + 0.010802141390740871, + 0.0066144573502242565, + 0.0045122792944312096, + 0.005817195400595665, + 0.005781736224889755, + 0.011435865424573421, + 0.012386866845190525, + 0.005603795405477285, + 0.00879712961614132, + 0.010139781050384045, + 0.0067248777486383915, + 0.0068642315454781055, + 0.005245800130069256, + 0.00872421357780695, + 0.00533710652962327, + 0.0037804199382662773, + 0.004156071692705154, + 0.004257993306964636, + 0.007619106210768223, + 0.003703133435919881, + 0.0039000927936285734, + 0.0030114115215837955, + 0.03332970663905144, + 0.008494925685226917, + 0.03726913034915924, + 0.012903144583106041, + 0.0042357612401247025, + 0.0054717762395739555, + 0.005637087859213352, + 0.0044627259485423565, + 0.0075281355530023575, + 0.007510406896471977, + 0.01005464792251587, + 0.006884935777634382, + 0.008426841348409653, + 0.009739805944263935, + 0.00927993468940258, + 0.0050441729836165905, + 0.010657536797225475, + 0.03591329604387283, + 0.02703743427991867, + 0.007587532512843609, + 0.010105263441801071, + 0.0074388873763382435, + 0.01318181399255991, + 0.009539422579109669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0052552176639437675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme wide shot captures a stunning timelapse of the iconic New York City skyline, seamlessly transitioning from late afternoon to late night. The scene begins with brilliant afternoon sunlight casting long, sharp shadows across the glass and steel skyscrapers that dot the city's grid. As the day progresses, the sun rapidly descends, bathing the entire cityscape in a warm golden hour glow; the buildings' western facades blaze in intense orange and yellow light, while their eastern shadows dramatically lengthen, cascading across canyons of concrete and glass. The sky transforms into a vibrant palette of deep orange, pink, and purple before fading into a cool, deep indigo dusk. It is in this moment that the city begins to awaken with renewed energy, and thousands of office and apartment windows begin to flicker, first as sparse points of light, then expanding into a magnificent, shimmering tapestry that redefines architectural form. The scene climaxes in the pitch black of night, the skyline a dazzling spectacle of white, yellow, and blue lights against a dark sky, the steady red pulse of aircraft warning beacons looming overhead on the tallest buildings, creating an atmosphere of majestic, never-ending urban energy.", + "seed": 28831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_5/index_410_seed_28831_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018978411960415543, + "vlm_logprob": -0.018978411960415543, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.018978411960415543, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03861837461590767 + ], + "opposite_logprob_tokens": [ + -3.913618326187134 + ], + "logprob_reward": -0.03861837461590767, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03861837461590767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10130525380373001, + 0.04134251922369003, + 0.0712195485830307, + 0.04389190673828125, + 0.21608027815818787, + 0.09950099885463715, + 0.05760759860277176, + 0.13287098705768585, + 0.10059972107410431, + 0.13558118045330048 + ], + "credit_map": [ + 0.009555652737617493, + 0.02243136428296566, + 0.013171752914786339, + 0.01631530560553074, + 0.010616069659590721, + 0.00733200553804636, + 0.010282104834914207, + 0.011600998230278492, + 0.004539537243545055, + 0.005794383119791746, + 0.005628332030028105, + 0.005692249163985252, + 0.004890382289886475, + 0.005071902647614479, + 0.004753598012030125, + 0.004972133785486221, + 0.004441208206117153, + 0.005853119771927595, + 0.018402349203824997, + 0.007593450136482716, + 0.004827718250453472, + 0.005771313793957233, + 0.009439128451049328, + 0.014891262166202068, + 0.005696413107216358, + 0.004347032867372036, + 0.0051591601222753525, + 0.005223310086876154, + 0.006237316410988569, + 0.00598165625706315, + 0.00486913463100791, + 0.006377886049449444, + 0.023243308067321777, + 0.026988035067915916, + 0.02818971313536167, + 0.02319052256643772, + 0.04527023434638977, + 0.025719130411744118, + 0.024289723485708237, + 0.019189592450857162, + 0.016943570226430893, + 0.010069835931062698, + 0.007437606807798147, + 0.006284915842115879, + 0.010873794555664062, + 0.023106401786208153, + 0.007915380410850048, + 0.01686949096620083, + 0.0071625132113695145, + 0.013852174393832684, + 0.005770121701061726, + 0.006462068762630224, + 0.003837991738691926, + 0.006173260044306517, + 0.005210577510297298, + 0.009138890542089939, + 0.02970234863460064, + 0.028535403311252594, + 0.011412073858082294, + 0.01383077260106802, + 0.011468786746263504, + 0.01950892060995102, + 0.009411877021193504, + 0.009000795893371105, + 0.01449211873114109, + 0.012889156118035316, + 0.00899025984108448, + 0.009728585369884968, + 0.015432773157954216, + 0.016532866284251213, + 0.00991728249937296, + 0.01261668000370264, + 0.016732115298509598, + 0.01716541312634945, + 0.008469662629067898, + 0.01533563993871212, + 0.011571984738111496, + 0.015209976583719254, + 0.03433500602841377, + 0.016761381179094315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.03861837461590767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043092746287584305 + ], + "opposite_logprob_tokens": [ + -3.168092727661133 + ], + "logprob_reward": -0.043092746287584305, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043092746287584305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10553059726953506, + 0.05695899948477745, + 0.20851676166057587, + 0.08883906900882721, + 0.03657921403646469, + 0.05076740309596062, + 0.0495014563202858, + 0.1969863921403885, + 0.05855800583958626, + 0.14776213467121124 + ], + "credit_map": [ + 0.014099645428359509, + 0.01327452715486288, + 0.013602710328996181, + 0.021530363708734512, + 0.009938017465174198, + 0.009820960462093353, + 0.012596198357641697, + 0.010668179020285606, + 0.008026385679841042, + 0.009688615798950195, + 0.00894633773714304, + 0.005733662284910679, + 0.005141050089150667, + 0.0063047767616808414, + 0.0055855391547083855, + 0.007532630115747452, + 0.01970537006855011, + 0.016029387712478638, + 0.07248847186565399, + 0.02336876466870308, + 0.01329442486166954, + 0.011463575065135956, + 0.01408684253692627, + 0.038079939782619476, + 0.011135398410260677, + 0.009910810738801956, + 0.011517692357301712, + 0.008946954272687435, + 0.011405416764318943, + 0.007904581725597382, + 0.010506740771234035, + 0.017511479556560516, + 0.00489529175683856, + 0.0056632934138178825, + 0.006019692402333021, + 0.0037962589412927628, + 0.003319612005725503, + 0.0049722217954695225, + 0.0038760665338486433, + 0.004036777652800083, + 0.0074064298532903194, + 0.005994992330670357, + 0.0034079866018146276, + 0.003066723933443427, + 0.004488363396376371, + 0.016229012981057167, + 0.003322135889902711, + 0.0068517569452524185, + 0.00612186873331666, + 0.01480075716972351, + 0.0030680436175316572, + 0.003242173232138157, + 0.003942484501749277, + 0.006040172651410103, + 0.0037810413632541895, + 0.008504917845129967, + 0.049641720950603485, + 0.07488206028938293, + 0.009525405243039131, + 0.00919750239700079, + 0.011412283405661583, + 0.02436143346130848, + 0.00850218627601862, + 0.009463800117373466, + 0.006727042142301798, + 0.008359458297491074, + 0.006060526706278324, + 0.007474101148545742, + 0.008757755160331726, + 0.008104508742690086, + 0.006819009780883789, + 0.006255602464079857, + 0.021224284544587135, + 0.021461308002471924, + 0.014459831640124321, + 0.013479460962116718, + 0.020090345293283463, + 0.0191370639950037, + 0.021373512223362923, + 0.01653631217777729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.043092746287584305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001189477858133614 + ], + "opposite_logprob_tokens": [ + -6.751189708709717 + ], + "logprob_reward": -0.001189477858133614, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001189477858133614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1039641797542572, + 0.05604816600680351, + 0.08138532936573029, + 0.03656456992030144, + 0.06756304949522018, + 0.037582412362098694, + 0.20174255967140198, + 0.1521562933921814, + 0.06402716040611267, + 0.19896630942821503 + ], + "credit_map": [ + 0.010812067426741123, + 0.014372806064784527, + 0.013818441890180111, + 0.01776134967803955, + 0.008664081804454327, + 0.012610373087227345, + 0.012106808833777905, + 0.01381825190037489, + 0.007365140598267317, + 0.006941045634448528, + 0.0076858894899487495, + 0.0063360002823174, + 0.008420458994805813, + 0.007279905490577221, + 0.005509795155376196, + 0.006509934086352587, + 0.009746484458446503, + 0.0049204146489501, + 0.02713998593389988, + 0.010133140720427036, + 0.00412409845739603, + 0.004817246925085783, + 0.006320635322481394, + 0.01418332289904356, + 0.005106573924422264, + 0.0028443261981010437, + 0.006001269910484552, + 0.004491667728871107, + 0.0037379551213234663, + 0.0027223187498748302, + 0.003456196514889598, + 0.008204261772334576, + 0.006830874364823103, + 0.00840696506202221, + 0.009915324859321117, + 0.013604125939309597, + 0.008959971368312836, + 0.005573343485593796, + 0.007639921270310879, + 0.00663252966478467, + 0.005199790000915527, + 0.004242122173309326, + 0.0031805660109966993, + 0.002446707570925355, + 0.0034449787344783545, + 0.010340468026697636, + 0.002827952615916729, + 0.005899826530367136, + 0.024517396464943886, + 0.06719862669706345, + 0.010727287270128727, + 0.010398787446320057, + 0.01353677362203598, + 0.017916616052389145, + 0.015003517270088196, + 0.042443547397851944, + 0.04071272909641266, + 0.03431714326143265, + 0.00700382050126791, + 0.01659976691007614, + 0.01817820779979229, + 0.017226124182343483, + 0.011246521957218647, + 0.006871968973428011, + 0.009142004884779453, + 0.01126460637897253, + 0.0052725267596542835, + 0.005730029661208391, + 0.011103994213044643, + 0.010038587264716625, + 0.0055656759068369865, + 0.005909737199544907, + 0.022891491651535034, + 0.028739945963025093, + 0.01164291426539421, + 0.02182627096772194, + 0.028575187548995018, + 0.020685950294137, + 0.04412226751446724, + 0.02048228494822979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.001189477858133614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002206512028351426 + ], + "opposite_logprob_tokens": [ + -6.127206325531006 + ], + "logprob_reward": -0.002206512028351426, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002206512028351426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08652645349502563, + 0.08020485937595367, + 0.10624181479215622, + 0.06994067877531052, + 0.029584424570202827, + 0.08919201046228409, + 0.2101507931947708, + 0.25002142786979675, + 0.04891595616936684, + 0.02922155149281025 + ], + "credit_map": [ + 0.009001186117529869, + 0.012525783851742744, + 0.010767999105155468, + 0.012333586812019348, + 0.010144717991352081, + 0.009803857654333115, + 0.012017626315355301, + 0.009931692853569984, + 0.0070837559178471565, + 0.011875089257955551, + 0.010256302542984486, + 0.010649425908923149, + 0.010645531117916107, + 0.010087750852108002, + 0.010380387306213379, + 0.009226619265973568, + 0.013815863989293575, + 0.011445223353803158, + 0.017122292891144753, + 0.013358366675674915, + 0.010871632024645805, + 0.009680296294391155, + 0.015102497301995754, + 0.014845643192529678, + 0.007716568186879158, + 0.0064554596319794655, + 0.010888219811022282, + 0.00951350573450327, + 0.005154546350240707, + 0.0063401395455002785, + 0.009581945836544037, + 0.014290292747318745, + 0.0025875356514006853, + 0.003546718508005142, + 0.003839714452624321, + 0.00695286737754941, + 0.001922326278872788, + 0.002811134560033679, + 0.0038435664027929306, + 0.004080562852323055, + 0.014705478213727474, + 0.01176957692950964, + 0.007397992070764303, + 0.0059959059581160545, + 0.006316212937235832, + 0.0293195191770792, + 0.006231037434190512, + 0.007456287741661072, + 0.026283377781510353, + 0.07494793832302094, + 0.008456778712570667, + 0.007928116247057915, + 0.012526166625320911, + 0.01601572521030903, + 0.014777339063584805, + 0.04921532794833183, + 0.04250349849462509, + 0.02934580296278, + 0.01897958293557167, + 0.03124680370092392, + 0.040014784783124924, + 0.044756561517715454, + 0.025440899655222893, + 0.017733503133058548, + 0.00785472709685564, + 0.007999860681593418, + 0.0037699760869145393, + 0.004720482509583235, + 0.008143696002662182, + 0.006888199597597122, + 0.0047297789715230465, + 0.004809235688298941, + 0.0050314124673604965, + 0.004188061226159334, + 0.002780167618766427, + 0.0031590787693858147, + 0.0034651141613721848, + 0.003592147957533598, + 0.003834305563941598, + 0.003171265125274658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.002206512028351426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, extreme wide shot of the New York City skyline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene transition from late afternoon sunlight with long shadows to a golden hour glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sky transform from deep orange, pink, and purple into a cool, deep indigo dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thousands of office and apartment windows flicker on, expanding from sparse points to a shimmering tapestry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene climax in pitch black night with white, yellow, and blue lights and red aircraft warning beacons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784949012100697 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009784949012100697, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784949012100697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06273191422224045, + 0.03876141458749771, + 0.060362085700035095, + 0.0465974397957325, + 0.03146986663341522, + 0.04821588471531868, + 0.1362747848033905, + 0.41554614901542664, + 0.08595121651887894, + 0.07408922165632248 + ], + "credit_map": [ + 0.006537561304867268, + 0.01093320269137621, + 0.007718886248767376, + 0.012343845330178738, + 0.005293876398354769, + 0.004449149593710899, + 0.009051780216395855, + 0.0064036184921860695, + 0.004206697456538677, + 0.004660237114876509, + 0.004896259400993586, + 0.005154309328645468, + 0.004675201140344143, + 0.004315773956477642, + 0.005742926616221666, + 0.005110008176416159, + 0.005249775946140289, + 0.0064301518723368645, + 0.0127851078286767, + 0.00796149019151926, + 0.005247948225587606, + 0.0074088010005652905, + 0.004309201613068581, + 0.01096961461007595, + 0.004266710951924324, + 0.004506031982600689, + 0.0066818115301430225, + 0.006547164171934128, + 0.005453834310173988, + 0.006292346399277449, + 0.005966939963400364, + 0.006882607005536556, + 0.0037713428027927876, + 0.004761833697557449, + 0.005003232043236494, + 0.0034258293453603983, + 0.00251369783654809, + 0.004387303255498409, + 0.00455985963344574, + 0.003046770114451647, + 0.010703641921281815, + 0.005263577215373516, + 0.004828969947993755, + 0.003013388253748417, + 0.005540249403566122, + 0.012023760937154293, + 0.0036022167187184095, + 0.0032400840427726507, + 0.01693914271891117, + 0.04537331685423851, + 0.005549769848585129, + 0.005690339021384716, + 0.007121644914150238, + 0.011812116019427776, + 0.010119350627064705, + 0.033669110387563705, + 0.08403846621513367, + 0.08340489864349365, + 0.024676207453012466, + 0.05104571580886841, + 0.06476094573736191, + 0.05352352187037468, + 0.03374423831701279, + 0.020352160558104515, + 0.009529423899948597, + 0.01026625744998455, + 0.00701798964291811, + 0.010520453564822674, + 0.01693711057305336, + 0.010342593304812908, + 0.010802676901221275, + 0.010534712113440037, + 0.010511912405490875, + 0.012246821075677872, + 0.0062346993945539, + 0.007267081178724766, + 0.010278930887579918, + 0.009306524880230427, + 0.010164430364966393, + 0.008078834973275661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.009784949012100697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_tensor([8853, 8853])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_tensor([8853, 8853])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b5b5ee458c84a712550409ee8d8fd0308309b817 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_tensor([8853, 8853])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, close-up shot, capturing this crucial moment with an almost handheld quality, focuses on a female general within a magnificent Tang Dynasty-style palace. Her towering, jet-black bun is slightly disheveled, a few sweat-soaked strands plastering her dusty face. Yet, her phoenix eyes blaze with an unyielding, resolute light. Her brow furrowed, she exerts all her might to bring an out-of-control black warhorse to a halt. She wears a meticulously crafted \"Fairy Knight Breastplate,\" its silver-white metal engraved with intricate vine patterns. As she leans back and tenses her muscles, the ancient magical rune inlaid in the center of the breastplate glows with a brilliant azure light, its halo gleaming brightly against the rising dust. She gripped the rough leather reins tightly, her knuckles white from the strain, as she fiercely battled the steed's ferocious power. Its eyes, wide with fear, and its nostrils, foaming with froth, blurred into the background. Surrounding her were the hazy outlines of the palace's massive vermilion columns and carved, painted rafters. Golden beams from the high windows pierced the perspiration and dust, forming beams of light that precisely illuminated the beads of sweat on her cheeks and the magical glow flowing across her breastplate. The entire scene was filled with tension, power, and a sacred, epic atmosphere.", + "seed": 32703, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_8853_seed_32703_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0906017180532217, + "vlm_logprob": -0.0906017180532217, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.0906017180532217, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0246278066188097 + ], + "opposite_logprob_tokens": [ + -3.774627923965454 + ], + "logprob_reward": -0.0246278066188097, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0246278066188097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16278982162475586, + 0.2272637039422989, + 0.062075112015008926, + 0.06183147430419922, + 0.06427895277738571, + 0.036206476390361786, + 0.16752225160598755, + 0.05702630430459976, + 0.08569194376468658, + 0.07531395554542542 + ], + "credit_map": [ + 0.027549391612410545, + 0.02735212817788124, + 0.018195340409874916, + 0.021371377632021904, + 0.01842152699828148, + 0.02047242969274521, + 0.01649102196097374, + 0.012936610728502274, + 0.023824863135814667, + 0.0196924339979887, + 0.05658653751015663, + 0.015918180346488953, + 0.026398317888379097, + 0.0594758577644825, + 0.014067677780985832, + 0.011299828067421913, + 0.006909976247698069, + 0.010025634430348873, + 0.00914805382490158, + 0.0048863813281059265, + 0.006265449803322554, + 0.008777376264333725, + 0.007074953522533178, + 0.008987288922071457, + 0.01606723479926586, + 0.0071197147481143475, + 0.005926923826336861, + 0.005908752791583538, + 0.004803750663995743, + 0.01088157668709755, + 0.00552715826779604, + 0.005596360191702843, + 0.007758180610835552, + 0.00827210582792759, + 0.007407253608107567, + 0.004825000651180744, + 0.010721760801970959, + 0.010732456110417843, + 0.008428406901657581, + 0.006133792921900749, + 0.0038429710548371077, + 0.005544010084122419, + 0.002667232183739543, + 0.003242329927161336, + 0.005338383372873068, + 0.005234714597463608, + 0.003878924995660782, + 0.006457910872995853, + 0.015013055875897408, + 0.013201583176851273, + 0.004266762174665928, + 0.0034411814995110035, + 0.06482955813407898, + 0.050705939531326294, + 0.007948060519993305, + 0.008116126991808414, + 0.005326510407030582, + 0.011853125877678394, + 0.005200177896767855, + 0.006264352239668369, + 0.006371452007442713, + 0.006173941306769848, + 0.006969064008444548, + 0.00886768102645874, + 0.009438352659344673, + 0.013277893885970116, + 0.00859865266829729, + 0.00948602706193924, + 0.01572476141154766, + 0.009565841406583786, + 0.011755544692277908, + 0.007844868116080761, + 0.01295175775885582, + 0.008862770162522793, + 0.008039642125368118, + 0.00690176198258996, + 0.008244002237915993, + 0.010775212198495865, + 0.00898236595094204, + 0.010556451044976711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.0246278066188097 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486550748348236 + ], + "opposite_logprob_tokens": [ + -1.223655104637146 + ], + "logprob_reward": -0.3486550748348236, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486550748348236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07107546180486679, + 0.17160829901695251, + 0.17719922959804535, + 0.1144646555185318, + 0.05287814885377884, + 0.15886808931827545, + 0.0540412962436676, + 0.04393133521080017, + 0.03397393599152565, + 0.12195951491594315 + ], + "credit_map": [ + 0.00744312908500433, + 0.010543965734541416, + 0.006311885546892881, + 0.00955104362219572, + 0.011506487615406513, + 0.009324870072305202, + 0.007750607095658779, + 0.008643466979265213, + 0.018802443519234657, + 0.014139307662844658, + 0.04733666777610779, + 0.011591409333050251, + 0.012138226069509983, + 0.049478404223918915, + 0.00815413985401392, + 0.009967693127691746, + 0.026265302672982216, + 0.02992541342973709, + 0.03251713514328003, + 0.015286571346223354, + 0.01935199648141861, + 0.027667757123708725, + 0.012266646139323711, + 0.013918410986661911, + 0.02618112601339817, + 0.013569314032793045, + 0.013123215176165104, + 0.009113358333706856, + 0.01372474990785122, + 0.015925662592053413, + 0.008547267876565456, + 0.014279961585998535, + 0.007444231305271387, + 0.0055615645833313465, + 0.0078099677339196205, + 0.004336746875196695, + 0.007079530041664839, + 0.00869833305478096, + 0.006563565693795681, + 0.005384213291108608, + 0.018741464242339134, + 0.013810182921588421, + 0.008626064285635948, + 0.01144898310303688, + 0.020161138847470284, + 0.028115302324295044, + 0.015517037361860275, + 0.04244791343808174, + 0.007168794982135296, + 0.0070090265944600105, + 0.0033663229551166296, + 0.0039376141503453255, + 0.018410850316286087, + 0.006182733457535505, + 0.004456896800547838, + 0.003509056754410267, + 0.0034947311505675316, + 0.00817931443452835, + 0.00404257932677865, + 0.00381674082018435, + 0.006074950564652681, + 0.004940221551805735, + 0.007650584448128939, + 0.005732210353016853, + 0.003825317369773984, + 0.004662961699068546, + 0.00295437965542078, + 0.0038277776911854744, + 0.005789626855403185, + 0.0029009487479925156, + 0.006788185331970453, + 0.0032247405033558607, + 0.017246222123503685, + 0.013978355564177036, + 0.012705960310995579, + 0.00865593645721674, + 0.017589980736374855, + 0.016643265262246132, + 0.0162615105509758, + 0.018878286704421043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.3486550748348236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305985942482948 + ], + "opposite_logprob_tokens": [ + -3.168059825897217 + ], + "logprob_reward": -0.04305985942482948, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305985942482948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06365688145160675, + 0.09402065724134445, + 0.0523427277803421, + 0.17686796188354492, + 0.2783621847629547, + 0.1319376528263092, + 0.07501707226037979, + 0.03232638165354729, + 0.04616159573197365, + 0.04930690675973892 + ], + "credit_map": [ + 0.009358773939311504, + 0.008357171900570393, + 0.005800880491733551, + 0.00941047165542841, + 0.008023026399314404, + 0.0072059850208461285, + 0.009298932738602161, + 0.006201630458235741, + 0.010648765601217747, + 0.011060771532356739, + 0.024946415796875954, + 0.005617812275886536, + 0.0050628227181732655, + 0.027744688093662262, + 0.005217744503170252, + 0.0037216204218566418, + 0.00821677315980196, + 0.00997238140553236, + 0.00596687151119113, + 0.0040639485232532024, + 0.006759469863027334, + 0.005179587285965681, + 0.0053751361556351185, + 0.006808553822338581, + 0.032206956297159195, + 0.018308768048882484, + 0.02348257414996624, + 0.017007136717438698, + 0.019709289073944092, + 0.022976597771048546, + 0.015449711121618748, + 0.027726899832487106, + 0.030192632228136063, + 0.032778386026620865, + 0.04307742416858673, + 0.022539230063557625, + 0.03637262061238289, + 0.05356457829475403, + 0.03514279052615166, + 0.024694502353668213, + 0.01682361401617527, + 0.009145529009401798, + 0.008664115332067013, + 0.009369770996272564, + 0.020112641155719757, + 0.02140323631465435, + 0.012360487133264542, + 0.03405824303627014, + 0.008507851511240005, + 0.005568374879658222, + 0.0025247959420084953, + 0.001970661571249366, + 0.0296222735196352, + 0.017771149054169655, + 0.005003804806619883, + 0.004048150032758713, + 0.0029014325700700283, + 0.006063732784241438, + 0.002862341236323118, + 0.0027649737894535065, + 0.004237389657646418, + 0.004017213359475136, + 0.006398703437298536, + 0.0030805962160229683, + 0.004747429862618446, + 0.00637240381911397, + 0.004987742751836777, + 0.005778492894023657, + 0.0060262251645326614, + 0.00704405689612031, + 0.006362883374094963, + 0.004842355381697416, + 0.008311711251735687, + 0.005529732909053564, + 0.0049546854570508, + 0.0040335506200790405, + 0.00739710871130228, + 0.006249219179153442, + 0.006187202408909798, + 0.006643690634518862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.04305985942482948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598655566573143 + ], + "opposite_logprob_tokens": [ + -3.8955986499786377 + ], + "logprob_reward": -0.020598655566573143, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598655566573143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0843549370765686, + 0.0821114331483841, + 0.10091620683670044, + 0.05530279874801636, + 0.1071883961558342, + 0.0991649478673935, + 0.07467308640480042, + 0.04513956606388092, + 0.23250465095043182, + 0.11864396184682846 + ], + "credit_map": [ + 0.01685355044901371, + 0.014338307082653046, + 0.010070095770061016, + 0.01173272356390953, + 0.009263585321605206, + 0.009665220975875854, + 0.006921602878719568, + 0.0055098505690693855, + 0.008747829124331474, + 0.010447992943227291, + 0.023392118513584137, + 0.004712789785116911, + 0.0041281371377408504, + 0.023301977664232254, + 0.002697570947930217, + 0.004683017730712891, + 0.012772793881595135, + 0.01978687010705471, + 0.0180814191699028, + 0.009141831658780575, + 0.010434594005346298, + 0.015019794926047325, + 0.008019796572625637, + 0.007659108377993107, + 0.011626188643276691, + 0.0056352135725319386, + 0.00713666807860136, + 0.0037060321774333715, + 0.00590808130800724, + 0.009332280606031418, + 0.0055643050000071526, + 0.00639403285458684, + 0.012438112869858742, + 0.013353235088288784, + 0.014198283664882183, + 0.00876652542501688, + 0.015870526432991028, + 0.019374433904886246, + 0.01477096788585186, + 0.008416309021413326, + 0.013146484270691872, + 0.010918471030890942, + 0.008027209900319576, + 0.008360862731933594, + 0.0132267065346241, + 0.016743551939725876, + 0.008508546277880669, + 0.020233118906617165, + 0.010392322205007076, + 0.007825094275176525, + 0.00283717829734087, + 0.0025039040483534336, + 0.031447555869817734, + 0.011473233811557293, + 0.00474974513053894, + 0.0034440516028553247, + 0.00486729945987463, + 0.009282959625124931, + 0.003981566056609154, + 0.004128533881157637, + 0.005344143602997065, + 0.006781940348446369, + 0.005249624140560627, + 0.005503499414771795, + 0.035082701593637466, + 0.024677708745002747, + 0.01977471448481083, + 0.017666848376393318, + 0.03469626605510712, + 0.0391322560608387, + 0.033172857016325, + 0.0283013004809618, + 0.013578562997281551, + 0.013651115819811821, + 0.009765978902578354, + 0.008572805672883987, + 0.011495723389089108, + 0.02540571242570877, + 0.018055692315101624, + 0.0181183610111475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.020598655566573143 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606719382107258 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.01606719382107258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606719382107258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10121352970600128, + 0.0548674575984478, + 0.07302034646272659, + 0.0906255915760994, + 0.2304696887731552, + 0.09316101670265198, + 0.1954220086336136, + 0.03675705939531326, + 0.0748722106218338, + 0.04959101602435112 + ], + "credit_map": [ + 0.01809508167207241, + 0.019336747005581856, + 0.012225966900587082, + 0.013169807381927967, + 0.012957388535141945, + 0.01279088668525219, + 0.006612756755203009, + 0.006024904083460569, + 0.005609362851828337, + 0.009324992075562477, + 0.009790744632482529, + 0.005495923571288586, + 0.0052983989007771015, + 0.010775353759527206, + 0.0036882394924759865, + 0.004884443711489439, + 0.009181930683553219, + 0.014318706467747688, + 0.010546309873461723, + 0.00592774897813797, + 0.005968149285763502, + 0.010814973153173923, + 0.00785887148231268, + 0.008403650484979153, + 0.013613463379442692, + 0.01296699233353138, + 0.010670225135982037, + 0.008583629503846169, + 0.009934389032423496, + 0.01027293037623167, + 0.012589151971042156, + 0.011994806118309498, + 0.02946692518889904, + 0.028076140210032463, + 0.029226982966065407, + 0.020203368738293648, + 0.036241598427295685, + 0.04079585149884224, + 0.02841121330857277, + 0.018047619611024857, + 0.010983875021338463, + 0.013466392643749714, + 0.009421506896615028, + 0.010888741351664066, + 0.016465799883008003, + 0.010311884805560112, + 0.00827052816748619, + 0.013352287001907825, + 0.02342318370938301, + 0.016887202858924866, + 0.005476145073771477, + 0.004865297116339207, + 0.08423121273517609, + 0.042406387627124786, + 0.010511151514947414, + 0.007621424738317728, + 0.0032990279141813517, + 0.008022639900445938, + 0.003149787662550807, + 0.0035064523108303547, + 0.003661727998405695, + 0.0038962194230407476, + 0.004244922194629908, + 0.006976285018026829, + 0.011239555664360523, + 0.008650800213217735, + 0.004976229276508093, + 0.005798706319183111, + 0.011172406375408173, + 0.012748690322041512, + 0.011749752797186375, + 0.008536062203347683, + 0.005260203965008259, + 0.006791998166590929, + 0.003555924165993929, + 0.004258523229509592, + 0.00707221357151866, + 0.008205440826714039, + 0.008352844975888729, + 0.0060938685201108456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.01606719382107258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up shot, capturing this crucial moment with an almost handheld quality, focuses on a female general within a magnificent Tang Dynasty-style palace. Her towering, jet-black bun is slightly disheveled, a few sweat-soaked strands plastering her dusty face. Yet, her phoenix eyes blaze with an unyielding, resolute light. Her brow furrowed, she exerts all her might to bring an out-of-control black warhorse to a halt. She wears a meticulously crafted \"Fairy Knight Breastplate,\" its silver-white metal engraved with intricate vine patterns. As she leans back and tenses her muscles, the ancient magical rune inlaid in the center of the breastplate glows with a brilliant azure light, its halo gleaming brightly against the rising dust. She gripped the rough leather reins tightly, her knuckles white from the strain, as she fiercely battled the steed's ferocious power. Its eyes, wide with fear, and its nostrils, foaming with froth, blurred into the background. Surrounding her were the hazy outlines of the palace's massive vermilion columns and carved, painted rafters. Golden beams from the high windows pierced the perspiration and dust, forming beams of light that precisely illuminated the beads of sweat on her cheeks and the magical glow flowing across her breastplate. The entire scene was filled with tension, power, and a sacred, epic atmosphere.", + "seed": 32704, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_8853_seed_32704_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21845658570528032, + "vlm_logprob": -0.21845658570528032, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.21845658570528032, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3498671352863312 + ], + "opposite_logprob_tokens": [ + -1.2248671054840088 + ], + "logprob_reward": -0.3498671352863312, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3498671352863312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036569807678461075, + 0.027164842933416367, + 0.17333656549453735, + 0.06418614089488983, + 0.10361217707395554, + 0.06225290894508362, + 0.3188861608505249, + 0.04581381380558014, + 0.1315334141254425, + 0.03664423152804375 + ], + "credit_map": [ + 0.004582176450639963, + 0.004305433016270399, + 0.0060059186071157455, + 0.003752188291400671, + 0.005568531807512045, + 0.003924483899027109, + 0.00510701909661293, + 0.003324055578559637, + 0.002320677973330021, + 0.0026397323235869408, + 0.003389789490029216, + 0.002269672928377986, + 0.0030529939103871584, + 0.0028781844303011894, + 0.006419818848371506, + 0.004193970933556557, + 0.014837573282420635, + 0.016398049890995026, + 0.05107518658041954, + 0.00845321360975504, + 0.007814944721758366, + 0.05906049534678459, + 0.007931229658424854, + 0.007765865419059992, + 0.008558926172554493, + 0.006099663209170103, + 0.00863543339073658, + 0.006911611650139093, + 0.0055893040262162685, + 0.0099793067201972, + 0.006230922415852547, + 0.012180972844362259, + 0.012514365836977959, + 0.014631683006882668, + 0.015569181181490421, + 0.018816765397787094, + 0.011387279257178307, + 0.012142239138484001, + 0.011904379352927208, + 0.006646276451647282, + 0.008858413435518742, + 0.011650842614471912, + 0.006624175235629082, + 0.009481018409132957, + 0.00659150630235672, + 0.0060999346897006035, + 0.00794496014714241, + 0.005002052057534456, + 0.04882613196969032, + 0.03390764817595482, + 0.04092444106936455, + 0.08672423660755157, + 0.019503751769661903, + 0.025014331564307213, + 0.0369214303791523, + 0.02706417255103588, + 0.005767288152128458, + 0.007269583642482758, + 0.006667601875960827, + 0.00594636844471097, + 0.004941096995025873, + 0.0062811062671244144, + 0.004413289949297905, + 0.004527472425252199, + 0.022737829014658928, + 0.02010600082576275, + 0.035344868898391724, + 0.018616583198308945, + 0.011561316438019276, + 0.007532544434070587, + 0.009272782132029533, + 0.00636147940531373, + 0.0034727740567177534, + 0.005888040643185377, + 0.00378397130407393, + 0.005282159429043531, + 0.004602065775543451, + 0.004698059055954218, + 0.004775245673954487, + 0.004141911398619413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.3498671352863312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289233386516571 + ], + "opposite_logprob_tokens": [ + -1.0539233684539795 + ], + "logprob_reward": -0.4289233386516571, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289233386516571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048498157411813736, + 0.025202423334121704, + 0.3745330274105072, + 0.04790796712040901, + 0.04929506778717041, + 0.038235004991292953, + 0.12559765577316284, + 0.06886614859104156, + 0.15419243276119232, + 0.06767214834690094 + ], + "credit_map": [ + 0.005188749637454748, + 0.00672910874709487, + 0.008152459748089314, + 0.00796640757471323, + 0.005411313381046057, + 0.004414182156324387, + 0.0061633954755961895, + 0.004472540225833654, + 0.002434142166748643, + 0.003000829601660371, + 0.003278691554442048, + 0.0024389822501689196, + 0.003290147753432393, + 0.0027490570209920406, + 0.005129919853061438, + 0.002880654064938426, + 0.03287481889128685, + 0.0368359200656414, + 0.11280796676874161, + 0.014485429972410202, + 0.016289066523313522, + 0.13606765866279602, + 0.007962354458868504, + 0.01720983162522316, + 0.006005759350955486, + 0.006532030180096626, + 0.005990916863083839, + 0.008473874069750309, + 0.0052767531014978886, + 0.0054240599274635315, + 0.005973586812615395, + 0.004230988211929798, + 0.006281026639044285, + 0.004998133517801762, + 0.004705742001533508, + 0.004510524217039347, + 0.0115470290184021, + 0.006575245875865221, + 0.006150825414806604, + 0.00452653830870986, + 0.005535668693482876, + 0.004479660652577877, + 0.00492420420050621, + 0.004175449255853891, + 0.00375384115613997, + 0.004916064906865358, + 0.004439609125256538, + 0.006010507233440876, + 0.014774356968700886, + 0.010389268398284912, + 0.015168110840022564, + 0.02383185550570488, + 0.010235733352601528, + 0.015330838970839977, + 0.01792430318892002, + 0.017943184822797775, + 0.006838509812951088, + 0.009212478995323181, + 0.008153709582984447, + 0.008393806405365467, + 0.007140825968235731, + 0.012598836794495583, + 0.00816638395190239, + 0.00836160033941269, + 0.031069422140717506, + 0.012225125916302204, + 0.04686374217271805, + 0.010565497912466526, + 0.011418181471526623, + 0.01631125807762146, + 0.011592622846364975, + 0.014146575704216957, + 0.007548850029706955, + 0.0077790552750229836, + 0.006908235140144825, + 0.011855732649564743, + 0.011345269158482552, + 0.00972966942936182, + 0.006093751173466444, + 0.006411583162844181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.4289233386516571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306979477405548 + ], + "opposite_logprob_tokens": [ + -3.168069839477539 + ], + "logprob_reward": -0.04306979477405548, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306979477405548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04933542013168335, + 0.031847983598709106, + 0.23837946355342865, + 0.07007928937673569, + 0.04672960191965103, + 0.09607025980949402, + 0.06066546216607094, + 0.2758021950721741, + 0.052343178540468216, + 0.07874718308448792 + ], + "credit_map": [ + 0.006472857668995857, + 0.005706530064344406, + 0.008356133475899696, + 0.00425319466739893, + 0.00805660430341959, + 0.005472684279084206, + 0.005982310976833105, + 0.005035107489675283, + 0.0032532704062759876, + 0.0039020259864628315, + 0.005202104803174734, + 0.00311471545137465, + 0.0033872090280056, + 0.0037660840898752213, + 0.005490810144692659, + 0.003731759963557124, + 0.020381072536110878, + 0.025106269866228104, + 0.06582161039113998, + 0.010388503782451153, + 0.013636729680001736, + 0.07744120061397552, + 0.009363650344312191, + 0.016240417957305908, + 0.01103577297180891, + 0.009783080779016018, + 0.007517910096794367, + 0.006362357642501593, + 0.006702658254653215, + 0.011884542182087898, + 0.005674700252711773, + 0.011118271388113499, + 0.005306378472596407, + 0.007280660793185234, + 0.006625953130424023, + 0.006316254381090403, + 0.004947168752551079, + 0.005634394474327564, + 0.0048948354087769985, + 0.005723961628973484, + 0.011164365336298943, + 0.020503317937254906, + 0.011293327435851097, + 0.012679003179073334, + 0.011005544103682041, + 0.01058292668312788, + 0.011940370313823223, + 0.006901406217366457, + 0.010510469786822796, + 0.0062911356799304485, + 0.007657259237021208, + 0.00594333978369832, + 0.006558718625456095, + 0.00840444304049015, + 0.0056490590795874596, + 0.00965103879570961, + 0.02927672676742077, + 0.0570148341357708, + 0.04040998965501785, + 0.05265846475958824, + 0.030317334458231926, + 0.01870989240705967, + 0.02375497668981552, + 0.02365995943546295, + 0.006021210458129644, + 0.007838306948542595, + 0.010257946327328682, + 0.007625564467161894, + 0.005342455580830574, + 0.003947456832975149, + 0.005555293522775173, + 0.0057549444027245045, + 0.007912830449640751, + 0.008979875594377518, + 0.007317736744880676, + 0.014134163968265057, + 0.012994441203773022, + 0.011494462378323078, + 0.007335705682635307, + 0.008577970787882805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.04306979477405548 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2522290349006653 + ], + "opposite_logprob_tokens": [ + -1.5022289752960205 + ], + "logprob_reward": -0.2522290349006653, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2522290349006653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058164820075035095, + 0.052081745117902756, + 0.39136144518852234, + 0.13770022988319397, + 0.059946946799755096, + 0.028903305530548096, + 0.11499373614788055, + 0.051504019647836685, + 0.036125969141721725, + 0.06921779364347458 + ], + "credit_map": [ + 0.007466258481144905, + 0.006936971098184586, + 0.011822570115327835, + 0.008096639066934586, + 0.006461332552134991, + 0.00610841391608119, + 0.00570652773603797, + 0.00556610943749547, + 0.007318382151424885, + 0.007962609641253948, + 0.006012891884893179, + 0.005827882792800665, + 0.005161982495337725, + 0.005245694424957037, + 0.009236648678779602, + 0.005315648391842842, + 0.03378854691982269, + 0.03707114979624748, + 0.10892942547798157, + 0.016937900334596634, + 0.02220471203327179, + 0.13238951563835144, + 0.014573019929230213, + 0.025467168539762497, + 0.01856953836977482, + 0.019751882180571556, + 0.02355142869055271, + 0.013083386234939098, + 0.016904639080166817, + 0.018824977800250053, + 0.01345913391560316, + 0.013555237092077732, + 0.005685667973011732, + 0.008420838974416256, + 0.0072740064933896065, + 0.010964643210172653, + 0.007755757309496403, + 0.007108726073056459, + 0.006975655909627676, + 0.005761650390923023, + 0.003987768664956093, + 0.004353238735347986, + 0.004097797907888889, + 0.003359180875122547, + 0.003043590812012553, + 0.003713932354003191, + 0.0035143529530614614, + 0.0028334420640021563, + 0.016487659886479378, + 0.008814540691673756, + 0.017612889409065247, + 0.01736970618367195, + 0.008717834018170834, + 0.014427817426621914, + 0.015198047272861004, + 0.016365239396691322, + 0.006776776164770126, + 0.008744556456804276, + 0.00786807481199503, + 0.008437130600214005, + 0.005945116281509399, + 0.0052153002470731735, + 0.004523030947893858, + 0.0039940341375768185, + 0.0041682361625134945, + 0.00411187345162034, + 0.0058273570612072945, + 0.0034153189044445753, + 0.006886057090014219, + 0.0047716922126710415, + 0.004045853391289711, + 0.0028995804022997618, + 0.010715547949075699, + 0.00829396490007639, + 0.005480943713337183, + 0.02018015645444393, + 0.006346413865685463, + 0.006386778317391872, + 0.006594300270080566, + 0.005219695158302784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.2522290349006653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193624913692474 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193624913692474, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193624913692474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06696511059999466, + 0.040790338069200516, + 0.1377318799495697, + 0.08721268177032471, + 0.0897994339466095, + 0.07311677187681198, + 0.0666198804974556, + 0.1351672261953354, + 0.04941505938768387, + 0.2531816065311432 + ], + "credit_map": [ + 0.009008139371871948, + 0.007988451048731804, + 0.011558986268937588, + 0.006903289817273617, + 0.006956190802156925, + 0.008230963721871376, + 0.00892773736268282, + 0.007391352206468582, + 0.004042491316795349, + 0.005344764329493046, + 0.005728449672460556, + 0.0040672835893929005, + 0.007373495027422905, + 0.005443180445581675, + 0.005156603641808033, + 0.0036340728402137756, + 0.011850956827402115, + 0.013337709940969944, + 0.03971094265580177, + 0.006515809800475836, + 0.006447095889598131, + 0.04811104014515877, + 0.004576495382934809, + 0.007181833498179913, + 0.01668018475174904, + 0.011182278394699097, + 0.007936466485261917, + 0.009184611961245537, + 0.009646603837609291, + 0.019119439646601677, + 0.0077791898511350155, + 0.00568391103297472, + 0.008966807276010513, + 0.01025979034602642, + 0.00982499960809946, + 0.009079364128410816, + 0.02108626998960972, + 0.014848602935671806, + 0.008647019043564796, + 0.007086578756570816, + 0.009346230886876583, + 0.012143549509346485, + 0.007245638407766819, + 0.007987690158188343, + 0.009748716838657856, + 0.008595599792897701, + 0.011452668346464634, + 0.006596681196242571, + 0.009886027313768864, + 0.00849286001175642, + 0.006865356117486954, + 0.007410390768200159, + 0.005383780226111412, + 0.011133466847240925, + 0.008068528957664967, + 0.009379473514854908, + 0.021957429125905037, + 0.01750245876610279, + 0.01915460266172886, + 0.015830188989639282, + 0.015110921114683151, + 0.01649593375623226, + 0.014233355410397053, + 0.01488233171403408, + 0.006116263568401337, + 0.006978968158364296, + 0.009688902646303177, + 0.005623064935207367, + 0.008730736561119556, + 0.004497398156672716, + 0.004283551592379808, + 0.0034961714409291744, + 0.024320827797055244, + 0.035498566925525665, + 0.024373646825551987, + 0.0416167676448822, + 0.032462868839502335, + 0.03609374910593033, + 0.0341036282479763, + 0.02471153438091278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.018193624913692474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up shot, capturing this crucial moment with an almost handheld quality, focuses on a female general within a magnificent Tang Dynasty-style palace. Her towering, jet-black bun is slightly disheveled, a few sweat-soaked strands plastering her dusty face. Yet, her phoenix eyes blaze with an unyielding, resolute light. Her brow furrowed, she exerts all her might to bring an out-of-control black warhorse to a halt. She wears a meticulously crafted \"Fairy Knight Breastplate,\" its silver-white metal engraved with intricate vine patterns. As she leans back and tenses her muscles, the ancient magical rune inlaid in the center of the breastplate glows with a brilliant azure light, its halo gleaming brightly against the rising dust. She gripped the rough leather reins tightly, her knuckles white from the strain, as she fiercely battled the steed's ferocious power. Its eyes, wide with fear, and its nostrils, foaming with froth, blurred into the background. Surrounding her were the hazy outlines of the palace's massive vermilion columns and carved, painted rafters. Golden beams from the high windows pierced the perspiration and dust, forming beams of light that precisely illuminated the beads of sweat on her cheeks and the magical glow flowing across her breastplate. The entire scene was filled with tension, power, and a sacred, epic atmosphere.", + "seed": 32705, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_8853_seed_32705_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2748145774006844, + "vlm_logprob": -0.2748145774006844, + "vlm_yes_no_margin": 1.7499999761581422, + "vlm_token_credit": -0.2748145774006844, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7597399950027466 + ], + "opposite_logprob_tokens": [ + -0.6347399950027466 + ], + "logprob_reward": -0.7597399950027466, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7597399950027466, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20229563117027283, + 0.04328017681837082, + 0.059981606900691986, + 0.0718560740351677, + 0.1639091521501541, + 0.08907586336135864, + 0.10455476492643356, + 0.08873124420642853, + 0.14167137444019318, + 0.03464413061738014 + ], + "credit_map": [ + 0.019086139276623726, + 0.02392721176147461, + 0.030548669397830963, + 0.020108802244067192, + 0.02940431423485279, + 0.018293937668204308, + 0.034538641571998596, + 0.026387900114059448, + 0.005186204332858324, + 0.006121088284999132, + 0.005083384923636913, + 0.0032874189782887697, + 0.007904529571533203, + 0.003938041627407074, + 0.007097368594259024, + 0.004662143066525459, + 0.005839180201292038, + 0.0065272157080471516, + 0.003810866503044963, + 0.005407530814409256, + 0.01396146509796381, + 0.008002362214028835, + 0.007659102790057659, + 0.008773887529969215, + 0.00603583175688982, + 0.007284276653081179, + 0.012641442939639091, + 0.012544789351522923, + 0.006195897702127695, + 0.009799827821552753, + 0.008983084000647068, + 0.008370921947062016, + 0.01902846433222294, + 0.02321304939687252, + 0.013323422521352768, + 0.02259134314954281, + 0.012691614218056202, + 0.01820496656000614, + 0.029494691640138626, + 0.025361597537994385, + 0.007434609811753035, + 0.020393306389451027, + 0.014805911108851433, + 0.004263450391590595, + 0.00675560487434268, + 0.006770539563149214, + 0.023442119359970093, + 0.0052103204652667046, + 0.009789232164621353, + 0.018473217263817787, + 0.009941342286765575, + 0.009135976433753967, + 0.017019063234329224, + 0.00973240565508604, + 0.02157036028802395, + 0.008893169462680817, + 0.007968392223119736, + 0.015754500404000282, + 0.017749235033988953, + 0.009910211898386478, + 0.009972529485821724, + 0.005438576452434063, + 0.014017194509506226, + 0.007920605130493641, + 0.010667262598872185, + 0.012592460960149765, + 0.0060477969236671925, + 0.014896689914166927, + 0.04907285422086716, + 0.022672930732369423, + 0.00899818167090416, + 0.01672319695353508, + 0.0038384634535759687, + 0.0034854323603212833, + 0.0035014767199754715, + 0.004797155968844891, + 0.003855721326544881, + 0.005343221127986908, + 0.006362884771078825, + 0.0034597732592374086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.7597399950027466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708909809589386 + ], + "opposite_logprob_tokens": [ + -2.127089023590088 + ], + "logprob_reward": -0.12708909809589386, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12708909809589386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07523223012685776, + 0.03338782861828804, + 0.03026166930794716, + 0.06258472055196762, + 0.15181979537010193, + 0.11907217651605606, + 0.19569094479084015, + 0.16388854384422302, + 0.08583728224039078, + 0.08222480118274689 + ], + "credit_map": [ + 0.006707489490509033, + 0.00977312307804823, + 0.00867929495871067, + 0.010711428709328175, + 0.012084763497114182, + 0.007663012482225895, + 0.0090479776263237, + 0.01056514773517847, + 0.004473975393921137, + 0.004127179272472858, + 0.005705706309527159, + 0.0026698687579482794, + 0.003459321102127433, + 0.004930410534143448, + 0.004457530565559864, + 0.003563839243724942, + 0.003597624134272337, + 0.0062147509306669235, + 0.004643561784178019, + 0.0031769226770848036, + 0.0033657008316367865, + 0.002740277675911784, + 0.0034735503140836954, + 0.003049283754080534, + 0.007160560693591833, + 0.006626706104725599, + 0.010061956942081451, + 0.010454193688929081, + 0.004374978598207235, + 0.007467181887477636, + 0.007581764366477728, + 0.008857380598783493, + 0.033085890114307404, + 0.025697382166981697, + 0.01613624393939972, + 0.022409966215491295, + 0.014586519449949265, + 0.012391883879899979, + 0.017224857583642006, + 0.010287064127624035, + 0.007375767920166254, + 0.03438844531774521, + 0.011625770479440689, + 0.005873485468327999, + 0.008526983670890331, + 0.011678339913487434, + 0.034894268959760666, + 0.004709125962108374, + 0.021772334352135658, + 0.03235943987965584, + 0.019800642505288124, + 0.014972465112805367, + 0.031026329845190048, + 0.018918149173259735, + 0.038210295140743256, + 0.01863129623234272, + 0.017884474247694016, + 0.03149402141571045, + 0.04129538685083389, + 0.0179457925260067, + 0.017790811136364937, + 0.012649931013584137, + 0.015053994953632355, + 0.009774143807590008, + 0.011447240598499775, + 0.010672143660485744, + 0.004814656917005777, + 0.010730276815593243, + 0.0212544035166502, + 0.01018885150551796, + 0.007587676867842674, + 0.009142039343714714, + 0.006149287335574627, + 0.00709817698225379, + 0.007359358482062817, + 0.011262806132435799, + 0.006307685747742653, + 0.010614735074341297, + 0.02232225239276886, + 0.01111050508916378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.12708909809589386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700417086482048 + ], + "opposite_logprob_tokens": [ + -2.6950416564941406 + ], + "logprob_reward": -0.0700417086482048, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700417086482048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07468011975288391, + 0.02872374653816223, + 0.03584735468029976, + 0.027357585728168488, + 0.2293224036693573, + 0.0714588612318039, + 0.05000689625740051, + 0.18841025233268738, + 0.1553623080253601, + 0.13883045315742493 + ], + "credit_map": [ + 0.008118918165564537, + 0.01213119924068451, + 0.007672452367842197, + 0.010199369862675667, + 0.007759519852697849, + 0.007308797910809517, + 0.011483688838779926, + 0.010006179101765156, + 0.003582988167181611, + 0.005563816521316767, + 0.0023779154289513826, + 0.0028873903211206198, + 0.0035253874957561493, + 0.0030262598302215338, + 0.005693324841558933, + 0.002066666027531028, + 0.0038547939620912075, + 0.0072720665484666824, + 0.0028843514155596495, + 0.0037908097729086876, + 0.007109455298632383, + 0.003535652533173561, + 0.0038208630867302418, + 0.0035793648567050695, + 0.0035541774705052376, + 0.0018951944075524807, + 0.004607670474797487, + 0.004257852211594582, + 0.002461884170770645, + 0.0036339065991342068, + 0.0030540635343641043, + 0.0038928373251110315, + 0.024847794324159622, + 0.035979464650154114, + 0.014839465729892254, + 0.03161393851041794, + 0.01542078610509634, + 0.021392228081822395, + 0.04509003460407257, + 0.040138717740774155, + 0.005498516373336315, + 0.01276057492941618, + 0.013183428905904293, + 0.003118879860267043, + 0.0072791059501469135, + 0.009651701897382736, + 0.01373792253434658, + 0.006228725891560316, + 0.006069941911846399, + 0.008282899856567383, + 0.005942035932093859, + 0.00565812923014164, + 0.006835154257714748, + 0.004851897247135639, + 0.008322499692440033, + 0.004044342320412397, + 0.025963881984353065, + 0.030683608725667, + 0.028986450284719467, + 0.01771734282374382, + 0.030626462772488594, + 0.011164129711687565, + 0.02871100604534149, + 0.01455738302320242, + 0.012006989680230618, + 0.014793166890740395, + 0.006306151393800974, + 0.016680344939231873, + 0.05418839678168297, + 0.024164339527487755, + 0.008808569982647896, + 0.01841435581445694, + 0.010778062045574188, + 0.010924265719950199, + 0.011449324898421764, + 0.01777190901339054, + 0.010428767651319504, + 0.018295081332325935, + 0.04002614691853523, + 0.019156893715262413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.0700417086482048 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3873882591724396 + ], + "opposite_logprob_tokens": [ + -1.1373882293701172 + ], + "logprob_reward": -0.3873882591724396, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3873882591724396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08040308952331543, + 0.055767722427845, + 0.039522603154182434, + 0.059000786393880844, + 0.10797564685344696, + 0.09302260726690292, + 0.412381112575531, + 0.08198259770870209, + 0.0395609587430954, + 0.030382875353097916 + ], + "credit_map": [ + 0.012921715155243874, + 0.01442719716578722, + 0.00980190560221672, + 0.009081795811653137, + 0.009552246890962124, + 0.007144480478018522, + 0.009906013496220112, + 0.007567740511149168, + 0.006510529667139053, + 0.010608812794089317, + 0.005793062504380941, + 0.006511805113404989, + 0.005513887386769056, + 0.005342064891010523, + 0.011351331137120724, + 0.004136228002607822, + 0.0050088888965547085, + 0.008577937260270119, + 0.00378977763466537, + 0.003910683561116457, + 0.005982444621622562, + 0.003222462022677064, + 0.0043118055909872055, + 0.004718605894595385, + 0.005823169369250536, + 0.00827033631503582, + 0.010199230164289474, + 0.010074437595903873, + 0.005015850532799959, + 0.00784114096313715, + 0.006493549793958664, + 0.005283070728182793, + 0.017830416560173035, + 0.015493308193981647, + 0.007425366900861263, + 0.014897777698934078, + 0.008566894568502903, + 0.009606692008674145, + 0.0179083701223135, + 0.01624683290719986, + 0.006452600937336683, + 0.026138247922062874, + 0.0072749825194478035, + 0.0039020271506160498, + 0.005674305837601423, + 0.01129466574639082, + 0.0277407206594944, + 0.004545058589428663, + 0.07785414904356003, + 0.055783361196517944, + 0.05619140714406967, + 0.04260741174221039, + 0.05488457530736923, + 0.03288736194372177, + 0.05438268929719925, + 0.03779013082385063, + 0.0086002666503191, + 0.014098871499300003, + 0.01527581736445427, + 0.009236309677362442, + 0.010153766721487045, + 0.006068156100809574, + 0.012356655672192574, + 0.006192747037857771, + 0.005189064424484968, + 0.00575445918366313, + 0.002463748212903738, + 0.00644941721111536, + 0.007597468327730894, + 0.004878650885075331, + 0.003360178554430604, + 0.0038679717108607292, + 0.0029068489093333483, + 0.002674506977200508, + 0.0026635618414729834, + 0.003564986400306225, + 0.002117771888151765, + 0.003370740218088031, + 0.008714611642062664, + 0.004369845148175955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.3873882591724396 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813826084136963 + ], + "opposite_logprob_tokens": [ + -3.529813766479492 + ], + "logprob_reward": -0.029813826084136963, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813826084136963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08925912529230118, + 0.07755675166845322, + 0.1671326607465744, + 0.077865369617939, + 0.06885972619056702, + 0.0783575028181076, + 0.13642434775829315, + 0.129563108086586, + 0.09892725944519043, + 0.0760541558265686 + ], + "credit_map": [ + 0.008445222862064838, + 0.015690810978412628, + 0.014436773955821991, + 0.01165810413658619, + 0.013206559233367443, + 0.008265077136456966, + 0.008893968537449837, + 0.00866260752081871, + 0.012950686737895012, + 0.012609300203621387, + 0.010807400569319725, + 0.007865658961236477, + 0.00770049262791872, + 0.008385954424738884, + 0.01099590864032507, + 0.006241351366043091, + 0.019667282700538635, + 0.04318639636039734, + 0.019333332777023315, + 0.0219388697296381, + 0.023411447182297707, + 0.01438229251652956, + 0.013467594049870968, + 0.011745449155569077, + 0.007003611419349909, + 0.009632416069507599, + 0.011470040306448936, + 0.012531273998320103, + 0.006253749132156372, + 0.010467018000781536, + 0.01079159788787365, + 0.009715660475194454, + 0.013737850822508335, + 0.01011024322360754, + 0.007831089198589325, + 0.009481724351644516, + 0.006186278071254492, + 0.006043375935405493, + 0.00796117540448904, + 0.007507992442697287, + 0.006143794860690832, + 0.016547681763768196, + 0.0106209022924304, + 0.004397047683596611, + 0.010027441196143627, + 0.009177104569971561, + 0.01642380654811859, + 0.005019721575081348, + 0.021319856867194176, + 0.01962757483124733, + 0.018550939857959747, + 0.01900092512369156, + 0.016256097704172134, + 0.010535057634115219, + 0.020889172330498695, + 0.010244731791317463, + 0.014091654680669308, + 0.018046187236905098, + 0.016106482595205307, + 0.012254470959305763, + 0.017668751999735832, + 0.008638283237814903, + 0.023726308718323708, + 0.019030973315238953, + 0.009519600309431553, + 0.009696850553154945, + 0.004452378489077091, + 0.008346321061253548, + 0.029816128313541412, + 0.01584399864077568, + 0.006540045607835054, + 0.014711938798427582, + 0.006741469260305166, + 0.005776438862085342, + 0.008123010396957397, + 0.011555429548025131, + 0.0060084545984864235, + 0.006046482361853123, + 0.019876724109053612, + 0.011926149018108845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.029813826084136963 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up shot, capturing this crucial moment with an almost handheld quality, focuses on a female general within a magnificent Tang Dynasty-style palace. Her towering, jet-black bun is slightly disheveled, a few sweat-soaked strands plastering her dusty face. Yet, her phoenix eyes blaze with an unyielding, resolute light. Her brow furrowed, she exerts all her might to bring an out-of-control black warhorse to a halt. She wears a meticulously crafted \"Fairy Knight Breastplate,\" its silver-white metal engraved with intricate vine patterns. As she leans back and tenses her muscles, the ancient magical rune inlaid in the center of the breastplate glows with a brilliant azure light, its halo gleaming brightly against the rising dust. She gripped the rough leather reins tightly, her knuckles white from the strain, as she fiercely battled the steed's ferocious power. Its eyes, wide with fear, and its nostrils, foaming with froth, blurred into the background. Surrounding her were the hazy outlines of the palace's massive vermilion columns and carved, painted rafters. Golden beams from the high windows pierced the perspiration and dust, forming beams of light that precisely illuminated the beads of sweat on her cheeks and the magical glow flowing across her breastplate. The entire scene was filled with tension, power, and a sacred, epic atmosphere.", + "seed": 32706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_6/index_8853_seed_32706_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0666356610134244, + "vlm_logprob": -0.0666356610134244, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.0666356610134244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03880072012543678 + ], + "opposite_logprob_tokens": [ + -3.2888007164001465 + ], + "logprob_reward": -0.03880072012543678, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03880072012543678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20106567442417145, + 0.043318331241607666, + 0.0644928365945816, + 0.22364594042301178, + 0.07734247297048569, + 0.13918285071849823, + 0.0664258748292923, + 0.046449195593595505, + 0.04036175087094307, + 0.09771512448787689 + ], + "credit_map": [ + 0.028756605461239815, + 0.03255818784236908, + 0.026469841599464417, + 0.022829286754131317, + 0.022993840277194977, + 0.026027070358395576, + 0.018712321296334267, + 0.02271851897239685, + 0.00869248528033495, + 0.006704676430672407, + 0.006422083824872971, + 0.00591700803488493, + 0.004515336826443672, + 0.003248696681112051, + 0.004763563629239798, + 0.003054477507248521, + 0.007302782032638788, + 0.008939198218286037, + 0.008326128125190735, + 0.0045426688157022, + 0.014673493802547455, + 0.00707830348983407, + 0.008640949614346027, + 0.004989318549633026, + 0.023206425830721855, + 0.025413362309336662, + 0.05670569837093353, + 0.01764206402003765, + 0.019259704276919365, + 0.01376158557832241, + 0.04681906849145889, + 0.020838038995862007, + 0.01236067246645689, + 0.008958871476352215, + 0.009790806099772453, + 0.008931235410273075, + 0.010929351672530174, + 0.009856888093054295, + 0.0074398331344127655, + 0.009074813686311245, + 0.03228580579161644, + 0.021500028669834137, + 0.018766392022371292, + 0.009554761461913586, + 0.015661293640732765, + 0.011693326756358147, + 0.018696142360568047, + 0.011025107465684414, + 0.008723322302103043, + 0.009870843961834908, + 0.007600376382470131, + 0.007696488872170448, + 0.01128376368433237, + 0.010620119981467724, + 0.00658100750297308, + 0.00404996145516634, + 0.00449824845418334, + 0.004955493379384279, + 0.005814829375594854, + 0.007877861149609089, + 0.0053140223026275635, + 0.005951093975454569, + 0.0056689707562327385, + 0.006368673872202635, + 0.006506726611405611, + 0.004716081079095602, + 0.003925126977264881, + 0.006187997758388519, + 0.00567334471270442, + 0.005722520407289267, + 0.003996109124273062, + 0.003633842570707202, + 0.01613577827811241, + 0.012380225583910942, + 0.009507126174867153, + 0.013529079966247082, + 0.015929363667964935, + 0.010589294135570526, + 0.01020300667732954, + 0.009441250003874302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.03880072012543678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203349679708481 + ], + "opposite_logprob_tokens": [ + -2.8120334148406982 + ], + "logprob_reward": -0.06203349679708481, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203349679708481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1457744985818863, + 0.09192068874835968, + 0.035996560007333755, + 0.12266455590724945, + 0.057372916489839554, + 0.13832256197929382, + 0.07373887300491333, + 0.12364445626735687, + 0.13432413339614868, + 0.07624072581529617 + ], + "credit_map": [ + 0.029456226155161858, + 0.021457966417074203, + 0.021540187299251556, + 0.015826361253857613, + 0.013109678402543068, + 0.013346080668270588, + 0.01531593594700098, + 0.015722056850790977, + 0.010715454816818237, + 0.010035689920186996, + 0.011918265372514725, + 0.014632985927164555, + 0.014229354448616505, + 0.008694479241967201, + 0.013972319662570953, + 0.00772214075550437, + 0.00481141172349453, + 0.005536102689802647, + 0.00397228728979826, + 0.00458289822563529, + 0.00494525907561183, + 0.00479378504678607, + 0.0033484897576272488, + 0.0040063271299004555, + 0.012943632900714874, + 0.01529819518327713, + 0.02941752038896084, + 0.009646071121096611, + 0.011931142769753933, + 0.007931205444037914, + 0.023522721603512764, + 0.011974063701927662, + 0.00987668801099062, + 0.0075772968120872974, + 0.005816006567329168, + 0.005065020639449358, + 0.009629128500819206, + 0.0077151525765657425, + 0.0051531414501369, + 0.006540476344525814, + 0.01701718382537365, + 0.008964527398347855, + 0.038173895329236984, + 0.009615406394004822, + 0.014578836038708687, + 0.00915539264678955, + 0.013984397053718567, + 0.026832913979887962, + 0.008693504147231579, + 0.011521980166435242, + 0.01050384808331728, + 0.008387362584471703, + 0.013335767202079296, + 0.010287298820912838, + 0.006360066123306751, + 0.004649043083190918, + 0.016312310472130775, + 0.01863778755068779, + 0.014635485596954823, + 0.024353042244911194, + 0.010828224010765553, + 0.01270910445600748, + 0.012819831259548664, + 0.013348669745028019, + 0.02570388652384281, + 0.016731958836317062, + 0.01114218682050705, + 0.024955587461590767, + 0.01556346844881773, + 0.011717129498720169, + 0.015825415030121803, + 0.012684488669037819, + 0.012155483476817608, + 0.009751064702868462, + 0.007555563468486071, + 0.01244096364825964, + 0.009964926168322563, + 0.007073622662574053, + 0.009486431255936623, + 0.0078126797452569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.06203349679708481 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818074844777584 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.01818074844777584, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818074844777584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07566727697849274, + 0.11012770980596542, + 0.0337054505944252, + 0.5364644527435303, + 0.049943093210458755, + 0.05213570222258568, + 0.0426425002515316, + 0.02444944530725479, + 0.022071752697229385, + 0.05279257148504257 + ], + "credit_map": [ + 0.010904224589467049, + 0.007758717052638531, + 0.009784352965652943, + 0.008556087501347065, + 0.009490695782005787, + 0.00871170312166214, + 0.009634481742978096, + 0.010827024467289448, + 0.016393279656767845, + 0.014603173360228539, + 0.013313516043126583, + 0.014921192079782486, + 0.012833048589527607, + 0.010316219180822372, + 0.019831515848636627, + 0.007915781810879707, + 0.005068890284746885, + 0.005282352678477764, + 0.00441625015810132, + 0.0036943659652024508, + 0.005164641886949539, + 0.004304409027099609, + 0.0033225512597709894, + 0.0024519937578588724, + 0.04921351373195648, + 0.06259959191083908, + 0.12605077028274536, + 0.03760335594415665, + 0.06855306029319763, + 0.04726514220237732, + 0.09485919773578644, + 0.05031987652182579, + 0.008446942083537579, + 0.006032485049217939, + 0.007108233403414488, + 0.0050672185607254505, + 0.00677713006734848, + 0.006097030360251665, + 0.00453529879450798, + 0.005878755822777748, + 0.009272337891161442, + 0.006197306793183088, + 0.0058837211690843105, + 0.0042668720707297325, + 0.007618339732289314, + 0.0048667630180716515, + 0.007742427289485931, + 0.006287937518209219, + 0.005501609295606613, + 0.004812677390873432, + 0.012129351496696472, + 0.0033060992136597633, + 0.006017028819769621, + 0.003909555729478598, + 0.00420622481033206, + 0.0027599569875746965, + 0.002840906148776412, + 0.003910774365067482, + 0.0029763083439320326, + 0.0028777429834008217, + 0.0029057839419692755, + 0.0038669751957058907, + 0.002555098384618759, + 0.002515855710953474, + 0.0037115630693733692, + 0.002928362926468253, + 0.0019004767527803779, + 0.003192620351910591, + 0.002541556488722563, + 0.002745730336755514, + 0.002676223171874881, + 0.0023752194829285145, + 0.006895372644066811, + 0.006298943422734737, + 0.005378572270274162, + 0.0063475435599684715, + 0.009694045409560204, + 0.006308263633400202, + 0.006409010849893093, + 0.00546081829816103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.01818074844777584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20162928104400635 + ], + "opposite_logprob_tokens": [ + -1.7016292810440063 + ], + "logprob_reward": -0.20162928104400635, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20162928104400635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0748516246676445, + 0.053953781723976135, + 0.042875126004219055, + 0.2128043919801712, + 0.03795544430613518, + 0.11373760551214218, + 0.12624678015708923, + 0.04913434386253357, + 0.228354811668396, + 0.06008606031537056 + ], + "credit_map": [ + 0.016619907692074776, + 0.010281609371304512, + 0.013122675940394402, + 0.005636495538055897, + 0.0072394683957099915, + 0.0068614548072218895, + 0.006389889866113663, + 0.008700123988091946, + 0.007009866181761026, + 0.006007847376167774, + 0.0075171105563640594, + 0.00587856350466609, + 0.0072095924988389015, + 0.006751410663127899, + 0.008365904912352562, + 0.005213486030697823, + 0.007015673443675041, + 0.006860881112515926, + 0.0064142313785851, + 0.00437581492587924, + 0.004739358555525541, + 0.004629762843251228, + 0.005368900950998068, + 0.0034705032594501972, + 0.04753023758530617, + 0.043916720896959305, + 0.04691280052065849, + 0.013887129724025726, + 0.015306644141674042, + 0.015077795833349228, + 0.016678135842084885, + 0.013494919054210186, + 0.00583501486107707, + 0.002843618392944336, + 0.004680671729147434, + 0.0052792648784816265, + 0.004921139683574438, + 0.0042600687593221664, + 0.006065049208700657, + 0.004070619121193886, + 0.01700703427195549, + 0.005833040457218885, + 0.0364219956099987, + 0.007192608900368214, + 0.006464532110840082, + 0.007185948546975851, + 0.008430526591837406, + 0.025201918557286263, + 0.016373708844184875, + 0.019052032381296158, + 0.047837384045124054, + 0.008665424771606922, + 0.00948711484670639, + 0.008885959163308144, + 0.009279321879148483, + 0.006665821652859449, + 0.0066207279451191425, + 0.006686091888695955, + 0.005179471801966429, + 0.004994094371795654, + 0.008306136354804039, + 0.007090077269822359, + 0.004685716703534126, + 0.005572030320763588, + 0.048344895243644714, + 0.025739183649420738, + 0.01756628416478634, + 0.03234698250889778, + 0.023014454171061516, + 0.02987583354115486, + 0.026073981076478958, + 0.02539318986237049, + 0.009477864019572735, + 0.007961256429553032, + 0.007136047817766666, + 0.007845921441912651, + 0.009397263638675213, + 0.006979513913393021, + 0.0059419674798846245, + 0.005346226971596479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.20162928104400635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female general wearing a silver-white breastplate with vine patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general attempting to stop a black warhorse that is foaming at the mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the breastplate feature a central magical rune glowing with brilliant azure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a Tang Dynasty-style palace with vermilion columns and carved rafters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden beams of light piercing through dust and sweat on the general's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534058652818203 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012534058652818203, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534058652818203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25049301981925964, + 0.049859464168548584, + 0.04501881077885628, + 0.21163831651210785, + 0.09344808757305145, + 0.0900590792298317, + 0.09422294795513153, + 0.04992327466607094, + 0.03544061258435249, + 0.07989636808633804 + ], + "credit_map": [ + 0.04837316647171974, + 0.038752686232328415, + 0.035855747759342194, + 0.019022392109036446, + 0.03141062706708908, + 0.023315630853176117, + 0.028000323101878166, + 0.02576242759823799, + 0.006921006366610527, + 0.006103620398789644, + 0.006703347899019718, + 0.005916398484259844, + 0.005488847382366657, + 0.004529687575995922, + 0.009546107612550259, + 0.004650451708585024, + 0.008030793629586697, + 0.006653605028986931, + 0.005776960402727127, + 0.004855785518884659, + 0.005299638491123915, + 0.004739843308925629, + 0.006182181648910046, + 0.0034800032153725624, + 0.02307085320353508, + 0.024644264951348305, + 0.05273279920220375, + 0.0186577457934618, + 0.019272038713097572, + 0.01686224713921547, + 0.038059718906879425, + 0.018338637426495552, + 0.014551198109984398, + 0.01159600354731083, + 0.013537305407226086, + 0.009347861632704735, + 0.012950223870575428, + 0.01224423386156559, + 0.008622226305305958, + 0.010599028319120407, + 0.021258380264043808, + 0.009553425945341587, + 0.010267614386975765, + 0.007368563208729029, + 0.0097856679931283, + 0.01020395290106535, + 0.012422272004187107, + 0.009199205785989761, + 0.012014736421406269, + 0.011579266749322414, + 0.03245474025607109, + 0.006989332847297192, + 0.008333277888596058, + 0.007867569103837013, + 0.007549089379608631, + 0.007434941362589598, + 0.005187754984945059, + 0.007149708457291126, + 0.0060952999629080296, + 0.00595902930945158, + 0.008207542821764946, + 0.00832161121070385, + 0.0050541008822619915, + 0.003948228899389505, + 0.010676882229745388, + 0.003049224615097046, + 0.002822933718562126, + 0.003632191801443696, + 0.003414758248254657, + 0.002673561219125986, + 0.005584024824202061, + 0.0035870345309376717, + 0.00993301346898079, + 0.009130991995334625, + 0.007536477874964476, + 0.010053295642137527, + 0.013734672218561172, + 0.011435789987444878, + 0.009203259833157063, + 0.008868871256709099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.012534058652818203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_tensor([977, 977])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_tensor([977, 977])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..16a251a23b454d4e9594aa2b6c28a74b5209b703 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_tensor([977, 977])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, still shot captures a scene imbued with the aesthetics of Ghibli: the sky a clear azure, dotted with fluffy, white clouds like cotton candy. In the warm afternoon sun, a vast expanse of rolling green hills is tinged with a soft golden glow. Amidst this tranquil, vibrant meadow, an antique electric sewing machine, painted black with faded gold filigree, seems to come to life as it tumbles down the hillside with a clumsy yet persistent tread. The machine's cast iron body gleams as it rolls, its wooden base scraping against the soft grass, dragging an old, braided electrical cord behind it, tangling and dragging. Its movement is not a smooth roll, but a bumpy process of swaying, sometimes rolling sideways, and sometimes rushing forward on the wheels, as if a creature that has just learned to walk is exploring the world. The whole scene is extremely quiet with a hint of strange fairy tale color, creating a warm, nostalgic and imaginative fantasy atmosphere.", + "seed": 186, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_977_seed_186_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009753062576055526, + "vlm_logprob": -0.009753062576055526, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.009753062576055526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022115076426416636 + ], + "opposite_logprob_tokens": [ + -7.627211570739746 + ], + "logprob_reward": -0.0022115076426416636, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0022115076426416636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09302183240652084, + 0.07418440282344818, + 0.04558888450264931, + 0.033219143748283386, + 0.03168635815382004, + 0.03803060203790665, + 0.5628382563591003, + 0.03219149261713028, + 0.03398561105132103, + 0.0552532933652401 + ], + "credit_map": [ + 0.011935168877243996, + 0.018315669149160385, + 0.012202664278447628, + 0.013739146292209625, + 0.004521347116678953, + 0.006703615188598633, + 0.014821137301623821, + 0.010783095844089985, + 0.007856730371713638, + 0.006602284964174032, + 0.010103218257427216, + 0.009144958108663559, + 0.006110963877290487, + 0.010272767394781113, + 0.012168340384960175, + 0.011925146915018559, + 0.003677234286442399, + 0.004912762437015772, + 0.010997883044183254, + 0.009371892549097538, + 0.0029090445023030043, + 0.004012962803244591, + 0.004802926909178495, + 0.004904181696474552, + 0.004658868070691824, + 0.005439447239041328, + 0.005825499538332224, + 0.005142223089933395, + 0.0036590825766324997, + 0.002918808488175273, + 0.003222452476620674, + 0.0023527657613158226, + 0.00431152293458581, + 0.005196704994887114, + 0.0036219204775989056, + 0.003378362162038684, + 0.0049171228893101215, + 0.003395954379811883, + 0.003584451973438263, + 0.0032803225331008434, + 0.005338883027434349, + 0.004520738031715155, + 0.004617027007043362, + 0.0033724235836416483, + 0.005096631124615669, + 0.006112373899668455, + 0.004815871827304363, + 0.004156657960265875, + 0.02976846694946289, + 0.011513937264680862, + 0.17783373594284058, + 0.017240500077605247, + 0.015063732862472534, + 0.24442538619041443, + 0.027604132890701294, + 0.03938845545053482, + 0.003485416527837515, + 0.0036746128462255, + 0.004755137953907251, + 0.003659221576526761, + 0.0033949692733585835, + 0.005227250978350639, + 0.002957309829071164, + 0.005037576891481876, + 0.004509533755481243, + 0.0028183558024466038, + 0.005851455964148045, + 0.002795223845168948, + 0.002346823224797845, + 0.007844219915568829, + 0.002604204462841153, + 0.005215797573328018, + 0.005791259463876486, + 0.006472197361290455, + 0.00894211232662201, + 0.006624419242143631, + 0.008062222972512245, + 0.009765737690031528, + 0.004876685794442892, + 0.0047186631709337234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0022115076426416636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953673273324966 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.005953673273324966, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953673273324966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08923271298408508, + 0.027113014832139015, + 0.05709531158208847, + 0.030596289783716202, + 0.05790642276406288, + 0.07375717163085938, + 0.5705885291099548, + 0.027220066636800766, + 0.04123266041278839, + 0.025257781147956848 + ], + "credit_map": [ + 0.011845104396343231, + 0.01415394525974989, + 0.013404437340795994, + 0.019443832337856293, + 0.0047320351004600525, + 0.0037698072846978903, + 0.01762474700808525, + 0.004258804954588413, + 0.0031116174068301916, + 0.003742688102647662, + 0.00452940259128809, + 0.0036475968081504107, + 0.002415519440546632, + 0.0030363888945430517, + 0.0033086766488850117, + 0.003321122145280242, + 0.004869627300649881, + 0.0067426301538944244, + 0.012651398777961731, + 0.010316106490790844, + 0.0041708373464643955, + 0.005923938471823931, + 0.006086518056690693, + 0.006334255449473858, + 0.002646957291290164, + 0.0042689028196036816, + 0.004291558638215065, + 0.0039682574570178986, + 0.003429246135056019, + 0.0037630684673786163, + 0.004294875543564558, + 0.003933421336114407, + 0.004976813681423664, + 0.005964323412626982, + 0.0075722793117165565, + 0.010763108730316162, + 0.010542629286646843, + 0.004954397678375244, + 0.008626759052276611, + 0.004506111145019531, + 0.006871815305203199, + 0.008176777511835098, + 0.00849588867276907, + 0.006795891560614109, + 0.011197809129953384, + 0.011126949451863766, + 0.013777286745607853, + 0.007314751856029034, + 0.028623344376683235, + 0.01977957785129547, + 0.17689715325832367, + 0.01831042766571045, + 0.016417555510997772, + 0.24354717135429382, + 0.027901532128453255, + 0.039111752063035965, + 0.0033093569800257683, + 0.0032245495822280645, + 0.004144094884395599, + 0.003246094798669219, + 0.0025064717046916485, + 0.004270494915544987, + 0.0030480357818305492, + 0.0034709691535681486, + 0.00434528524056077, + 0.004945592954754829, + 0.0071594445034861565, + 0.004456353839486837, + 0.003693198785185814, + 0.008600587025284767, + 0.003278204007074237, + 0.004753992427140474, + 0.0026032179594039917, + 0.0038730925880372524, + 0.003529404290020466, + 0.003747074631974101, + 0.0026108052115887403, + 0.004157347604632378, + 0.0025622607208788395, + 0.0021745783742517233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.005953673273324966 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095738288015127 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095738288015127, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095738288015127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08036381751298904, + 0.0889817401766777, + 0.040809161961078644, + 0.03713202103972435, + 0.02488350123167038, + 0.03559314087033272, + 0.5774900317192078, + 0.03646054118871689, + 0.061132170259952545, + 0.017153820022940636 + ], + "credit_map": [ + 0.012049146927893162, + 0.01889302395284176, + 0.010424351319670677, + 0.014596113935112953, + 0.004969545640051365, + 0.0036593603435903788, + 0.012161105871200562, + 0.00361118302680552, + 0.00964417401701212, + 0.008787588216364384, + 0.01066345814615488, + 0.010766982100903988, + 0.006394009571522474, + 0.013548611663281918, + 0.014420337975025177, + 0.014756584540009499, + 0.0037319690454751253, + 0.0038138919044286013, + 0.012361305765807629, + 0.010229594074189663, + 0.002020214218646288, + 0.0026960603427141905, + 0.0031688143499195576, + 0.0027873138897120953, + 0.002530075376853347, + 0.003926042001694441, + 0.009810532443225384, + 0.007476904895156622, + 0.0030983590986579657, + 0.002354516414925456, + 0.003519111080095172, + 0.004416482523083687, + 0.0022256558295339346, + 0.002043549669906497, + 0.00342054245993495, + 0.004639626480638981, + 0.002866248833015561, + 0.0029387641698122025, + 0.003485052613541484, + 0.0032640632707625628, + 0.0030549983493983746, + 0.005310015752911568, + 0.0038431454449892044, + 0.004465697333216667, + 0.00533817894756794, + 0.004211746621876955, + 0.004990211687982082, + 0.004379148129373789, + 0.029674503952264786, + 0.011517010629177094, + 0.18338559567928314, + 0.018061332404613495, + 0.015146022662520409, + 0.251636803150177, + 0.027643384411931038, + 0.04042544215917587, + 0.002702170517295599, + 0.0051374537870287895, + 0.005288220010697842, + 0.0053374674171209335, + 0.004964540712535381, + 0.005714136175811291, + 0.0037578933406621218, + 0.003558659926056862, + 0.008767778985202312, + 0.005234880838543177, + 0.013815653510391712, + 0.0033952670637518167, + 0.004413665272295475, + 0.01476849615573883, + 0.003507358254864812, + 0.007229073904454708, + 0.002200373448431492, + 0.002661733189597726, + 0.002866779686883092, + 0.0020588082261383533, + 0.0017887181602418423, + 0.0025545668322592974, + 0.001511131296865642, + 0.0015117096481844783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.004095738288015127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368692100048065 + ], + "opposite_logprob_tokens": [ + -3.408686876296997 + ], + "logprob_reward": -0.03368692100048065, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368692100048065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07118731737136841, + 0.05877460539340973, + 0.049133431166410446, + 0.07060129195451736, + 0.03876778110861778, + 0.08637844771146774, + 0.43170875310897827, + 0.10424120724201202, + 0.04674664884805679, + 0.04246043413877487 + ], + "credit_map": [ + 0.00774025171995163, + 0.015111775137484074, + 0.015288883820176125, + 0.012123985216021538, + 0.00307856104336679, + 0.0036704728845506907, + 0.007374541833996773, + 0.006798849441111088, + 0.006474512629210949, + 0.005253134295344353, + 0.007164972834289074, + 0.006832003127783537, + 0.004449839238077402, + 0.009281670674681664, + 0.009784760884940624, + 0.009533717297017574, + 0.0053701032884418964, + 0.005481223110109568, + 0.010769866406917572, + 0.0094752237200737, + 0.0036410470493137836, + 0.004443848505616188, + 0.005361432209610939, + 0.00459069199860096, + 0.008368046954274178, + 0.010450424626469612, + 0.009155701845884323, + 0.010934876278042793, + 0.008062543347477913, + 0.005807538982480764, + 0.0072846985422074795, + 0.010537475347518921, + 0.004127888008952141, + 0.005675272550433874, + 0.004644979257136583, + 0.006474432535469532, + 0.004817445296794176, + 0.003465901594609022, + 0.005434574093669653, + 0.0041272942908108234, + 0.007844435051083565, + 0.009801057167351246, + 0.011768518015742302, + 0.007784727960824966, + 0.01044685859233141, + 0.01804540678858757, + 0.010888597927987576, + 0.009798852726817131, + 0.02286417968571186, + 0.0169405285269022, + 0.12789972126483917, + 0.013526707887649536, + 0.018672967329621315, + 0.1776479035615921, + 0.02401321940124035, + 0.030143586918711662, + 0.013902616687119007, + 0.008473481051623821, + 0.018368525430560112, + 0.010674047283828259, + 0.007870478555560112, + 0.01476393360644579, + 0.011832193471491337, + 0.01835593767464161, + 0.00497034564614296, + 0.004591497592628002, + 0.009711580350995064, + 0.002847270341590047, + 0.003516754601150751, + 0.01197932381182909, + 0.0027230384293943644, + 0.00640684412792325, + 0.004616837948560715, + 0.0059165190905332565, + 0.006578494794666767, + 0.005867781583219767, + 0.004028457682579756, + 0.007977932691574097, + 0.003432122291997075, + 0.004042292479425669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.03368692100048065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028174726758152246 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028174726758152246, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028174726758152246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20199690759181976, + 0.09142131358385086, + 0.05463035777211189, + 0.1828872263431549, + 0.07018855959177017, + 0.0861365795135498, + 0.08552441745996475, + 0.06256755441427231, + 0.10163231194019318, + 0.0630146861076355 + ], + "credit_map": [ + 0.029022710397839546, + 0.05006631463766098, + 0.035594791173934937, + 0.02534695342183113, + 0.012526029720902443, + 0.008956697769463062, + 0.03230856731534004, + 0.008174853399395943, + 0.010052979923784733, + 0.011937948875129223, + 0.01436594221740961, + 0.011019757017493248, + 0.010810495354235172, + 0.011664088815450668, + 0.012528687715530396, + 0.009041420184075832, + 0.006174994166940451, + 0.007395581807941198, + 0.007719419896602631, + 0.006211569998413324, + 0.005210449919104576, + 0.007034964393824339, + 0.007944509387016296, + 0.006938872393220663, + 0.02922005020081997, + 0.033201344311237335, + 0.028647666797041893, + 0.025983398780226707, + 0.014449035748839378, + 0.015989432111382484, + 0.015037977136671543, + 0.020358338952064514, + 0.011630827561020851, + 0.008470515720546246, + 0.011256684549152851, + 0.008916071616113186, + 0.00683011207729578, + 0.007516905199736357, + 0.0069610741920769215, + 0.008606375195086002, + 0.009203232824802399, + 0.01075028721243143, + 0.011165370233356953, + 0.008549976162612438, + 0.011508621275424957, + 0.015478960238397121, + 0.009729736484587193, + 0.00975039228796959, + 0.004468309693038464, + 0.010503624565899372, + 0.020149800926446915, + 0.004088719841092825, + 0.0051184180192649364, + 0.02903079055249691, + 0.006413409952074289, + 0.005751344375312328, + 0.008472120389342308, + 0.00607823021709919, + 0.010325700044631958, + 0.0056986636482179165, + 0.004489803221076727, + 0.00982513464987278, + 0.006466892082244158, + 0.011211013421416283, + 0.011893399059772491, + 0.008342940360307693, + 0.022226529195904732, + 0.006480025593191385, + 0.005909294821321964, + 0.02699826844036579, + 0.006113414652645588, + 0.013668453320860863, + 0.007512388750910759, + 0.00804795604199171, + 0.008851468563079834, + 0.00804715696722269, + 0.008473247289657593, + 0.008685266599059105, + 0.0053537506610155106, + 0.008043460547924042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0028174726758152246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, still shot captures a scene imbued with the aesthetics of Ghibli: the sky a clear azure, dotted with fluffy, white clouds like cotton candy. In the warm afternoon sun, a vast expanse of rolling green hills is tinged with a soft golden glow. Amidst this tranquil, vibrant meadow, an antique electric sewing machine, painted black with faded gold filigree, seems to come to life as it tumbles down the hillside with a clumsy yet persistent tread. The machine's cast iron body gleams as it rolls, its wooden base scraping against the soft grass, dragging an old, braided electrical cord behind it, tangling and dragging. Its movement is not a smooth roll, but a bumpy process of swaying, sometimes rolling sideways, and sometimes rushing forward on the wheels, as if a creature that has just learned to walk is exploring the world. The whole scene is extremely quiet with a hint of strange fairy tale color, creating a warm, nostalgic and imaginative fantasy atmosphere.", + "seed": 187, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_977_seed_187_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0064617620781064035, + "vlm_logprob": -0.0064617620781064035, + "vlm_yes_no_margin": 5.974999952316284, + "vlm_token_credit": -0.0064617620781064035, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006191005813889205 + ], + "opposite_logprob_tokens": [ + -8.375618934631348 + ], + "logprob_reward": -0.0006191005813889205, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.0006191005813889205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13537240028381348, + 0.08559499680995941, + 0.22944633662700653, + 0.030952349305152893, + 0.0639507994055748, + 0.18157975375652313, + 0.12566575407981873, + 0.04714219272136688, + 0.04233648627996445, + 0.05795886740088463 + ], + "credit_map": [ + 0.01345447264611721, + 0.016909394413232803, + 0.019725516438484192, + 0.01707347296178341, + 0.029243825003504753, + 0.014826592057943344, + 0.014981266111135483, + 0.009157862514257431, + 0.01157338172197342, + 0.015211964026093483, + 0.011136236600577831, + 0.011203965172171593, + 0.008564738556742668, + 0.00754786329343915, + 0.010899321176111698, + 0.009457537904381752, + 0.010095752775669098, + 0.009986357763409615, + 0.09028896689414978, + 0.007119198329746723, + 0.0038566526491194963, + 0.008971580304205418, + 0.006162055302411318, + 0.09296580404043198, + 0.002821073867380619, + 0.004025242291390896, + 0.004404108971357346, + 0.003953629173338413, + 0.00414741225540638, + 0.0032383305951952934, + 0.004522847011685371, + 0.003839707002043724, + 0.005083393305540085, + 0.00794935505837202, + 0.008826172910630703, + 0.008753592148423195, + 0.012090620584785938, + 0.007256197743117809, + 0.006932319141924381, + 0.007059162948280573, + 0.013637962751090527, + 0.012142453342676163, + 0.031086577102541924, + 0.04281797632575035, + 0.021144766360521317, + 0.01898619905114174, + 0.018771717324852943, + 0.022992119193077087, + 0.0072896163910627365, + 0.009547662921249866, + 0.022551704198122025, + 0.02064387872815132, + 0.01565425656735897, + 0.011778025887906551, + 0.019678674638271332, + 0.018521953374147415, + 0.003587069222703576, + 0.005973338149487972, + 0.006095658987760544, + 0.004645814187824726, + 0.010477697476744652, + 0.00791841372847557, + 0.003221872029826045, + 0.005222335923463106, + 0.003117865649983287, + 0.0037965045776218176, + 0.004771416075527668, + 0.006932911928743124, + 0.004148178733885288, + 0.006325745023787022, + 0.005737419240176678, + 0.007506449241191149, + 0.003236197866499424, + 0.003658483037725091, + 0.01964598335325718, + 0.0036445162259042263, + 0.002601457992568612, + 0.004259697161614895, + 0.0033211102709174156, + 0.017591426149010658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.0006191005813889205 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978105328977108 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.00978105328977108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978105328977108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21626470983028412, + 0.050753846764564514, + 0.17944906651973724, + 0.049610435962677, + 0.062273621559143066, + 0.10985855013132095, + 0.14332692325115204, + 0.04741517826914787, + 0.026729540899395943, + 0.11431814730167389 + ], + "credit_map": [ + 0.014810546301305294, + 0.02681049145758152, + 0.024585526436567307, + 0.019878698512911797, + 0.06266555190086365, + 0.029764462262392044, + 0.02430468425154686, + 0.013444744050502777, + 0.003885317128151655, + 0.0051116058602929115, + 0.007541414350271225, + 0.009164362214505672, + 0.006982394959777594, + 0.0039309849962592125, + 0.007584089413285255, + 0.006553671322762966, + 0.008038843050599098, + 0.008075481280684471, + 0.060522351413965225, + 0.008676066994667053, + 0.006126062478870153, + 0.011203055270016193, + 0.0121176578104496, + 0.06468953937292099, + 0.003572201356291771, + 0.005881914868950844, + 0.0063945455476641655, + 0.005777981132268906, + 0.007444138638675213, + 0.0054631358943879604, + 0.006732176523655653, + 0.00834434013813734, + 0.005109535064548254, + 0.008974460884928703, + 0.008310635574162006, + 0.011460477486252785, + 0.0068990387953817844, + 0.006752634886652231, + 0.0077843982726335526, + 0.006982432212680578, + 0.00763180386275053, + 0.009508050046861172, + 0.01862088032066822, + 0.01832050271332264, + 0.01959049515426159, + 0.010785250924527645, + 0.01084231212735176, + 0.014559243805706501, + 0.007606762461364269, + 0.010857852175831795, + 0.026810944080352783, + 0.022487681359052658, + 0.01805700547993183, + 0.013173622079193592, + 0.024317936971783638, + 0.020015111193060875, + 0.004032035358250141, + 0.004310912918299437, + 0.006118196994066238, + 0.005563654471188784, + 0.011361897923052311, + 0.005778300575911999, + 0.005511260591447353, + 0.004738913383334875, + 0.0018624399090185761, + 0.0031370953656733036, + 0.0030272079166024923, + 0.0035343249328434467, + 0.0031133403535932302, + 0.0036775385960936546, + 0.004068667069077492, + 0.004308923147618771, + 0.005666343495249748, + 0.006586237344890833, + 0.04061288386583328, + 0.005845090374350548, + 0.0035909810103476048, + 0.007361048832535744, + 0.006248078308999538, + 0.038407474756240845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.00978105328977108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028152139857411385 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.0028152139857411385, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028152139857411385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12856723368167877, + 0.06331299245357513, + 0.1335987001657486, + 0.03428248316049576, + 0.08840794116258621, + 0.08983206748962402, + 0.21981734037399292, + 0.12287013977766037, + 0.0714782252907753, + 0.04783286154270172 + ], + "credit_map": [ + 0.009967101737856865, + 0.01513441652059555, + 0.02029341273009777, + 0.013609442859888077, + 0.029383936896920204, + 0.016430947929620743, + 0.013495448976755142, + 0.010252531617879868, + 0.008448511362075806, + 0.01003092247992754, + 0.0065713380463421345, + 0.008108329027891159, + 0.007481797132641077, + 0.005035483743995428, + 0.009884477593004704, + 0.007752130273729563, + 0.005635278765112162, + 0.005686487071216106, + 0.046950988471508026, + 0.008603670634329319, + 0.004523261915892363, + 0.006536182481795549, + 0.007298130542039871, + 0.04836468771100044, + 0.0030882975552231073, + 0.003965023905038834, + 0.004305616952478886, + 0.003990518860518932, + 0.004378878511488438, + 0.004149203654378653, + 0.005936563014984131, + 0.004468381870537996, + 0.004435181617736816, + 0.007276403252035379, + 0.012389685027301311, + 0.016768015921115875, + 0.013247794471681118, + 0.0071701728738844395, + 0.014140626415610313, + 0.012980062514543533, + 0.005637799389660358, + 0.00678011542186141, + 0.016088686883449554, + 0.015897396951913834, + 0.014897066168487072, + 0.00944917369633913, + 0.00933092925697565, + 0.011750901117920876, + 0.01652209274470806, + 0.017653057351708412, + 0.04149344563484192, + 0.03223975747823715, + 0.024169811978936195, + 0.02411133237183094, + 0.034729938954114914, + 0.028897911310195923, + 0.013141692616045475, + 0.0076876492239534855, + 0.015149833634495735, + 0.01869180053472519, + 0.01575927995145321, + 0.01660739816725254, + 0.017588162794709206, + 0.01824432611465454, + 0.0055701048113405704, + 0.010968207381665707, + 0.00997503288090229, + 0.010220468044281006, + 0.010027708485722542, + 0.00566442683339119, + 0.010021088644862175, + 0.009031188674271107, + 0.004140865057706833, + 0.0051921638660132885, + 0.007490973919630051, + 0.0046821171417832375, + 0.005686511751264334, + 0.004881450440734625, + 0.009807727299630642, + 0.005951054394245148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.0028152139857411385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817583292722702 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.01817583292722702, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01817583292722702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24183358252048492, + 0.0512939915060997, + 0.09741964936256409, + 0.08021539449691772, + 0.09584622085094452, + 0.0520361103117466, + 0.10244353860616684, + 0.10070377588272095, + 0.044300276786088943, + 0.13390743732452393 + ], + "credit_map": [ + 0.015128168277442455, + 0.027902472764253616, + 0.030232449993491173, + 0.023566341027617455, + 0.06353182345628738, + 0.03607536852359772, + 0.028495322912931442, + 0.016901642084121704, + 0.0032867593690752983, + 0.005092481151223183, + 0.007758560124784708, + 0.006136177107691765, + 0.010106666013598442, + 0.006455549504607916, + 0.007161930203437805, + 0.005295868963003159, + 0.004573311656713486, + 0.004605413880199194, + 0.0326552540063858, + 0.006078250706195831, + 0.0031455534044653177, + 0.005191933363676071, + 0.007486947812139988, + 0.033682990819215775, + 0.006350727751851082, + 0.009165958501398563, + 0.0104798199608922, + 0.010111276060342789, + 0.01345470268279314, + 0.00914663728326559, + 0.011918995529413223, + 0.009587287902832031, + 0.006763993296772242, + 0.012726442888379097, + 0.012656350620090961, + 0.01361188106238842, + 0.01608986221253872, + 0.0110201146453619, + 0.010717184282839298, + 0.012260389514267445, + 0.004627155605703592, + 0.0057844785042107105, + 0.007600302342325449, + 0.00685510179027915, + 0.00876524392515421, + 0.004620603285729885, + 0.006772853899747133, + 0.007010375149548054, + 0.0063727907836437225, + 0.008258894085884094, + 0.017641866579651833, + 0.015594637021422386, + 0.014684519730508327, + 0.009666066616773605, + 0.01605973020195961, + 0.014165041968226433, + 0.007663576398044825, + 0.009781047701835632, + 0.015489993616938591, + 0.011421136558055878, + 0.021565228700637817, + 0.011595268733799458, + 0.013181068003177643, + 0.010006469674408436, + 0.003998969681560993, + 0.004919897299259901, + 0.005913770757615566, + 0.005815077573060989, + 0.005518412217497826, + 0.00561748119071126, + 0.005982513539493084, + 0.006534156855195761, + 0.007050710264593363, + 0.0223796758800745, + 0.02463463880121708, + 0.01909293793141842, + 0.010793342255055904, + 0.011976322159171104, + 0.01418230403214693, + 0.023797525092959404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.01817583292722702 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009176096064038575 + ], + "opposite_logprob_tokens": [ + -7.000917434692383 + ], + "logprob_reward": -0.0009176096064038575, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009176096064038575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23048558831214905, + 0.08702508360147476, + 0.28300604224205017, + 0.0668129101395607, + 0.07967562228441238, + 0.08118536323308945, + 0.057249847799539566, + 0.021927764639258385, + 0.025808658450841904, + 0.06682316213846207 + ], + "credit_map": [ + 0.015796199440956116, + 0.028295772150158882, + 0.02656080387532711, + 0.024048812687397003, + 0.06127363443374634, + 0.03258312866091728, + 0.028248118236660957, + 0.01367910299450159, + 0.011776542291045189, + 0.017127690836787224, + 0.009913926012814045, + 0.010818103328347206, + 0.008029836229979992, + 0.009618035517632961, + 0.009417083114385605, + 0.01032386813312769, + 0.012826121412217617, + 0.012271896004676819, + 0.10718636214733124, + 0.008650342002511024, + 0.006514487322419882, + 0.015915391966700554, + 0.009158525615930557, + 0.11048292368650436, + 0.004047025926411152, + 0.0070371669717133045, + 0.009623819962143898, + 0.010290125384926796, + 0.01060548797249794, + 0.007171145640313625, + 0.010516287758946419, + 0.007521838415414095, + 0.009791563265025616, + 0.008317137137055397, + 0.010812804102897644, + 0.011826282367110252, + 0.010044043883681297, + 0.01077818963676691, + 0.008364386856555939, + 0.009741206653416157, + 0.005914540030062199, + 0.004881220869719982, + 0.014291642233729362, + 0.019308045506477356, + 0.010163243860006332, + 0.007017865777015686, + 0.007870701141655445, + 0.011738099157810211, + 0.004964738618582487, + 0.005703844130039215, + 0.008574912324547768, + 0.009703271090984344, + 0.006271189544349909, + 0.006375108379870653, + 0.007992124184966087, + 0.007664655800908804, + 0.0015179344918578863, + 0.0021348530426621437, + 0.002773614600300789, + 0.0032764195930212736, + 0.0034162693191319704, + 0.0026471451856195927, + 0.0034874530974775553, + 0.0026740743778645992, + 0.0018781990511342883, + 0.003328985068947077, + 0.0027034892700612545, + 0.0032589342445135117, + 0.003087804652750492, + 0.0033434522338211536, + 0.003965416457504034, + 0.0042423768900334835, + 0.00338574917986989, + 0.00493566133081913, + 0.023378631100058556, + 0.003471892327070236, + 0.0019406869541853666, + 0.004140638746321201, + 0.0030040706042200327, + 0.02256583422422409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0009176096064038575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, still shot captures a scene imbued with the aesthetics of Ghibli: the sky a clear azure, dotted with fluffy, white clouds like cotton candy. In the warm afternoon sun, a vast expanse of rolling green hills is tinged with a soft golden glow. Amidst this tranquil, vibrant meadow, an antique electric sewing machine, painted black with faded gold filigree, seems to come to life as it tumbles down the hillside with a clumsy yet persistent tread. The machine's cast iron body gleams as it rolls, its wooden base scraping against the soft grass, dragging an old, braided electrical cord behind it, tangling and dragging. Its movement is not a smooth roll, but a bumpy process of swaying, sometimes rolling sideways, and sometimes rushing forward on the wheels, as if a creature that has just learned to walk is exploring the world. The whole scene is extremely quiet with a hint of strange fairy tale color, creating a warm, nostalgic and imaginative fantasy atmosphere.", + "seed": 188, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_977_seed_188_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.009924912988208235, + "vlm_logprob": -0.009924912988208235, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.009924912988208235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0027585094794631004 + ], + "opposite_logprob_tokens": [ + -6.877758502960205 + ], + "logprob_reward": -0.0027585094794631004, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0027585094794631004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03728483244776726, + 0.09127239137887955, + 0.020832909271121025, + 0.036080360412597656, + 0.010138248093426228, + 0.029323652386665344, + 0.017975714057683945, + 0.013733944855630398, + 0.045501720160245895, + 0.6978563070297241 + ], + "credit_map": [ + 0.006302462425082922, + 0.004404543433338404, + 0.004137296695262194, + 0.0038820875342935324, + 0.0038019870407879353, + 0.00431273365393281, + 0.004097835160791874, + 0.006345887202769518, + 0.013992253690958023, + 0.011102579534053802, + 0.00991086196154356, + 0.010053422302007675, + 0.009719212539494038, + 0.014255435205996037, + 0.010983163490891457, + 0.011255455203354359, + 0.0037219536025077105, + 0.0040137083269655704, + 0.0022976850159466267, + 0.0025686563458293676, + 0.0017834495520219207, + 0.002935354597866535, + 0.0018551774555817246, + 0.0016569227445870638, + 0.0033477088436484337, + 0.005317748058587313, + 0.003299343166872859, + 0.0034841701854020357, + 0.004339597653597593, + 0.008977165445685387, + 0.002609473653137684, + 0.0047051492147147655, + 0.001070577185600996, + 0.001359898946247995, + 0.0015900024445727468, + 0.0011284291977062821, + 0.001379349734634161, + 0.001047919038683176, + 0.0014827473787590861, + 0.0010793233523145318, + 0.009315039031207561, + 0.010182629339396954, + 0.0019283290021121502, + 0.0015294429613277316, + 0.0013335312251001596, + 0.0021911023650318384, + 0.001404586830176413, + 0.0014389887219294906, + 0.0021624763030558825, + 0.0015558430459350348, + 0.003228291403502226, + 0.002010124968364835, + 0.0025015664286911488, + 0.0021192608401179314, + 0.0020432258024811745, + 0.002354924101382494, + 0.0014756544260308146, + 0.0026132583152502775, + 0.0030202874913811684, + 0.0009450822253711522, + 0.001871631946414709, + 0.001309491810388863, + 0.0008768297266215086, + 0.0016217080410569906, + 0.005469698458909988, + 0.006625491194427013, + 0.0038348245434463024, + 0.003913416992872953, + 0.00793615821748972, + 0.0076164244674146175, + 0.006057508755475283, + 0.004048193339258432, + 0.21753183007240295, + 0.09507540613412857, + 0.22508086264133453, + 0.032757725566625595, + 0.013493540696799755, + 0.013054786249995232, + 0.017057936638593674, + 0.08380420506000519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.0027585094794631004 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08322808146476746, + 0.0333508662879467, + 0.045687947422266006, + 0.03095615841448307, + 0.24203431606292725, + 0.04557652026414871, + 0.020843936130404472, + 0.10489653050899506, + 0.1533142328262329, + 0.24011139571666718 + ], + "credit_map": [ + 0.014599047601222992, + 0.01289940346032381, + 0.008142532780766487, + 0.010207530111074448, + 0.008748858235776424, + 0.008117282763123512, + 0.009362936019897461, + 0.011150490492582321, + 0.004436900373548269, + 0.0033376303035765886, + 0.004097715485841036, + 0.0048238355666399, + 0.004016022197902203, + 0.003274698741734028, + 0.004222100600600243, + 0.0051419599913060665, + 0.005195660050958395, + 0.006084744352847338, + 0.008086834102869034, + 0.006927241571247578, + 0.005473332479596138, + 0.004689518827944994, + 0.004804781638085842, + 0.0044258334673941135, + 0.00409074267372489, + 0.004342568572610617, + 0.0035961493849754333, + 0.0038631458301097155, + 0.00440026493743062, + 0.004510920029133558, + 0.0030074066016823053, + 0.003144962014630437, + 0.08153925836086273, + 0.025351407006382942, + 0.016148461028933525, + 0.008155541494488716, + 0.005446553695946932, + 0.02099943906068802, + 0.014381765387952328, + 0.07001189142465591, + 0.005305407103151083, + 0.0055506848730146885, + 0.00439205439761281, + 0.00646272674202919, + 0.009785129688680172, + 0.00403484096750617, + 0.005110188852995634, + 0.004935489967465401, + 0.0027260850183665752, + 0.0017438095528632402, + 0.0022800720762461424, + 0.0048292516730725765, + 0.0025332877412438393, + 0.0017574573867022991, + 0.0027405517175793648, + 0.0022334211971610785, + 0.015754973515868187, + 0.017096320167183876, + 0.0075520738027989864, + 0.0076454454101622105, + 0.01861884631216526, + 0.015937410295009613, + 0.009085657075047493, + 0.01320580579340458, + 0.017797397449612617, + 0.022834371775388718, + 0.013300304301083088, + 0.012427224777638912, + 0.025492755696177483, + 0.02578483149409294, + 0.020138708874583244, + 0.015538637526333332, + 0.06489001214504242, + 0.030488312244415283, + 0.0676623061299324, + 0.021213293075561523, + 0.008067281916737556, + 0.007630498614162207, + 0.011274224147200584, + 0.028885461390018463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.01108264084905386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863721314817667 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.00863721314817667, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863721314817667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02553500421345234, + 0.027732253074645996, + 0.01743648760020733, + 0.07599776238203049, + 0.46125686168670654, + 0.05570700392127037, + 0.020925045013427734, + 0.04166439548134804, + 0.03522965684533119, + 0.23851551115512848 + ], + "credit_map": [ + 0.003132707206532359, + 0.003439328633248806, + 0.003210413735359907, + 0.0036932507064193487, + 0.002525156596675515, + 0.0033722075168043375, + 0.0026864083483815193, + 0.0034755340311676264, + 0.004135903436690569, + 0.003708538133651018, + 0.002373764291405678, + 0.0021501320879906416, + 0.0030259054619818926, + 0.0063829259015619755, + 0.002808990888297558, + 0.003146094735711813, + 0.0019819780718535185, + 0.00204907963052392, + 0.0021830445621162653, + 0.00197677593678236, + 0.0023143617436289787, + 0.0022793691605329514, + 0.002516845939680934, + 0.0021350327879190445, + 0.007118742913007736, + 0.008789296261966228, + 0.008467047475278378, + 0.010767900384962559, + 0.006876810919493437, + 0.011999832466244698, + 0.00983724370598793, + 0.01214088499546051, + 0.15543194115161896, + 0.04847559705376625, + 0.030961086973547935, + 0.015219035558402538, + 0.009477023035287857, + 0.03965366631746292, + 0.027871323749423027, + 0.13416719436645508, + 0.006865217816084623, + 0.005987261887639761, + 0.010117078199982643, + 0.006269257050007582, + 0.004685701336711645, + 0.00948207639157772, + 0.005512392148375511, + 0.006788024213165045, + 0.00265250401571393, + 0.0020547888707369566, + 0.0022413304541260004, + 0.002691489877179265, + 0.003964726347476244, + 0.002278211759403348, + 0.002828498836606741, + 0.0022134948521852493, + 0.009002170525491238, + 0.004110772628337145, + 0.011801207438111305, + 0.0028073291759938, + 0.0030916715040802956, + 0.0028544573578983545, + 0.0030604421626776457, + 0.004936342593282461, + 0.0040693399496376514, + 0.004909633658826351, + 0.0027876130770891905, + 0.003195831784978509, + 0.006457806099206209, + 0.0059632123447954655, + 0.004601346328854561, + 0.00324487267062068, + 0.071770578622818, + 0.031947437673807144, + 0.07422974705696106, + 0.013614850118756294, + 0.005450722761452198, + 0.005471407901495695, + 0.007153918035328388, + 0.028876835480332375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.00863721314817667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334119960665703 + ], + "opposite_logprob_tokens": [ + -3.651334047317505 + ], + "logprob_reward": -0.026334119960665703, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334119960665703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05502329021692276, + 0.09001260250806808, + 0.04672551900148392, + 0.057296574115753174, + 0.3481399714946747, + 0.05008995532989502, + 0.02193404734134674, + 0.03166944533586502, + 0.052186354994773865, + 0.24692228436470032 + ], + "credit_map": [ + 0.00788235291838646, + 0.0068418653681874275, + 0.008472898975014687, + 0.008888598531484604, + 0.005031491629779339, + 0.004763210657984018, + 0.006501486524939537, + 0.006641382351517677, + 0.014527317136526108, + 0.010716388933360577, + 0.01038520410656929, + 0.009948110207915306, + 0.008785954676568508, + 0.014559994451701641, + 0.009882166981697083, + 0.011207465082406998, + 0.007625569589436054, + 0.008939052931964397, + 0.006533490493893623, + 0.004881268832832575, + 0.005264714825898409, + 0.004632271360605955, + 0.00403206842020154, + 0.004817083477973938, + 0.0066762445494532585, + 0.007872073911130428, + 0.004990106914192438, + 0.00699644023552537, + 0.007885927334427834, + 0.012701816856861115, + 0.00422121025621891, + 0.005952751729637384, + 0.1168665662407875, + 0.03643755614757538, + 0.023384487256407738, + 0.011475630104541779, + 0.007850744761526585, + 0.030002983286976814, + 0.02107427455484867, + 0.10104773193597794, + 0.006088268477469683, + 0.006077818106859922, + 0.007690261583775282, + 0.00653472077101469, + 0.00501621700823307, + 0.007820745930075645, + 0.004721854347735643, + 0.006140068173408508, + 0.002708769403398037, + 0.001850651460699737, + 0.0021103834733366966, + 0.003703280119225383, + 0.0034813592210412025, + 0.0022439402528107166, + 0.003311282256618142, + 0.0025243815034627914, + 0.008761769160628319, + 0.007775530684739351, + 0.0023752585984766483, + 0.0016515306197106838, + 0.004418639000505209, + 0.002843031194061041, + 0.00167753454297781, + 0.0021661529317498207, + 0.005809050519019365, + 0.007047771476209164, + 0.004684483632445335, + 0.0047820936888456345, + 0.008762575685977936, + 0.007816944271326065, + 0.008057286031544209, + 0.005226146895438433, + 0.07659073173999786, + 0.03349446505308151, + 0.0792408287525177, + 0.011990581639111042, + 0.005316057242453098, + 0.004917077720165253, + 0.00610709423199296, + 0.029265452176332474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.026334119960665703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008120815036818385 + ], + "opposite_logprob_tokens": [ + -7.12581205368042 + ], + "logprob_reward": -0.0008120815036818385, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008120815036818385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046809226274490356, + 0.026389574632048607, + 0.01716374047100544, + 0.02366616576910019, + 0.2089671939611435, + 0.015752289444208145, + 0.01130136102437973, + 0.12000402063131332, + 0.021236136555671692, + 0.5087102055549622 + ], + "credit_map": [ + 0.006212469656020403, + 0.005704415030777454, + 0.0032982646953314543, + 0.004103946965187788, + 0.005575781222432852, + 0.009250369854271412, + 0.006816294509917498, + 0.005847682245075703, + 0.004243205767124891, + 0.003926759585738182, + 0.00329165649600327, + 0.0025612348690629005, + 0.00224221870303154, + 0.004501361399888992, + 0.002241293666884303, + 0.003381843212991953, + 0.002564529189839959, + 0.0035333719570189714, + 0.0018178676255047321, + 0.0021345734130591154, + 0.002274435246363282, + 0.001954671461135149, + 0.0014959345571696758, + 0.0013883581850677729, + 0.0023873751051723957, + 0.0029809153638780117, + 0.002516700653359294, + 0.0030619350727647543, + 0.002998037962242961, + 0.0035537981893867254, + 0.00286684138700366, + 0.003300562035292387, + 0.07073075324296951, + 0.021949749439954758, + 0.013874400407075882, + 0.006721979007124901, + 0.004118410404771566, + 0.018080348148941994, + 0.01260419748723507, + 0.060887373983860016, + 0.0033006365410983562, + 0.003547581611201167, + 0.0017055298667401075, + 0.0014404680114239454, + 0.0013435326982289553, + 0.0016688850009813905, + 0.0013333427486941218, + 0.0014123116852715611, + 0.0016454266151413321, + 0.0009434266248717904, + 0.002177517395466566, + 0.0016642927657812834, + 0.0012805666774511337, + 0.00112459110096097, + 0.0011371326399967074, + 0.0013284068554639816, + 0.04120735451579094, + 0.026298604905605316, + 0.02582288719713688, + 0.004245790187269449, + 0.004089146386831999, + 0.0033247610554099083, + 0.004509291145950556, + 0.010506192222237587, + 0.002540934132412076, + 0.0031612440943717957, + 0.0018497866112738848, + 0.001807538908906281, + 0.0033485267776995897, + 0.0037699993699789047, + 0.0027317428030073643, + 0.0020263628102838993, + 0.15841154754161835, + 0.0693611353635788, + 0.16382886469364166, + 0.023922845721244812, + 0.009954032488167286, + 0.009459746070206165, + 0.012602767907083035, + 0.06116924062371254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.0008120815036818385 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, still shot captures a scene imbued with the aesthetics of Ghibli: the sky a clear azure, dotted with fluffy, white clouds like cotton candy. In the warm afternoon sun, a vast expanse of rolling green hills is tinged with a soft golden glow. Amidst this tranquil, vibrant meadow, an antique electric sewing machine, painted black with faded gold filigree, seems to come to life as it tumbles down the hillside with a clumsy yet persistent tread. The machine's cast iron body gleams as it rolls, its wooden base scraping against the soft grass, dragging an old, braided electrical cord behind it, tangling and dragging. Its movement is not a smooth roll, but a bumpy process of swaying, sometimes rolling sideways, and sometimes rushing forward on the wheels, as if a creature that has just learned to walk is exploring the world. The whole scene is extremely quiet with a hint of strange fairy tale color, creating a warm, nostalgic and imaginative fantasy atmosphere.", + "seed": 189, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_7/index_977_seed_189_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010990399692673236, + "vlm_logprob": -0.010990399692673236, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.010990399692673236, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00289794709533453 + ], + "opposite_logprob_tokens": [ + -6.3778977394104 + ], + "logprob_reward": -0.00289794709533453, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.00289794709533453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1362231969833374, + 0.07350369542837143, + 0.025736575946211815, + 0.03528919816017151, + 0.0390775129199028, + 0.2417229264974594, + 0.1994934380054474, + 0.07534433156251907, + 0.10340656340122223, + 0.0702025517821312 + ], + "credit_map": [ + 0.01681661233305931, + 0.016093097627162933, + 0.020680103451013565, + 0.02021924965083599, + 0.014526447281241417, + 0.01427356619387865, + 0.012789350003004074, + 0.020824767649173737, + 0.008012516424059868, + 0.008402219042181969, + 0.010168328881263733, + 0.009662728756666183, + 0.01169082522392273, + 0.009709248319268227, + 0.008109766989946365, + 0.007748068310320377, + 0.0033994088880717754, + 0.003163144690915942, + 0.0039036592934280634, + 0.0030850195325911045, + 0.00327578978613019, + 0.003920697141438723, + 0.002515008207410574, + 0.0024738512001931667, + 0.004171577747911215, + 0.004758971277624369, + 0.008219621144235134, + 0.0038380564656108618, + 0.0036439623218029737, + 0.003907785750925541, + 0.0039727333933115005, + 0.002776490291580558, + 0.0030918142292648554, + 0.003675054991617799, + 0.005380216054618359, + 0.004327437374740839, + 0.004364120773971081, + 0.0031044166535139084, + 0.008123558945953846, + 0.007010894361883402, + 0.04263288900256157, + 0.07474593818187714, + 0.0157487615942955, + 0.015691939741373062, + 0.014139406383037567, + 0.020894553512334824, + 0.018705448135733604, + 0.039163995534181595, + 0.022327736020088196, + 0.012262617237865925, + 0.028875427320599556, + 0.015449896454811096, + 0.02580222487449646, + 0.044896434992551804, + 0.021083133295178413, + 0.02879597432911396, + 0.009602271020412445, + 0.006162008736282587, + 0.012457684613764286, + 0.00785804446786642, + 0.007330876775085926, + 0.011470593512058258, + 0.007754367310553789, + 0.012708486057817936, + 0.010948389768600464, + 0.008490871638059616, + 0.019766977056860924, + 0.01363122183829546, + 0.00833371002227068, + 0.009676116518676281, + 0.014929231256246567, + 0.017630044370889664, + 0.005725349765270948, + 0.004804606083780527, + 0.008783464320003986, + 0.014349186792969704, + 0.00462164543569088, + 0.003213057992979884, + 0.013597049750387669, + 0.01510818861424923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.00289794709533453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740333512425423 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740333512425423, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740333512425423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24153809249401093, + 0.08756029605865479, + 0.03939468041062355, + 0.04111606255173683, + 0.039053287357091904, + 0.09439945966005325, + 0.03629203885793686, + 0.052538711577653885, + 0.1769779920578003, + 0.19112935662269592 + ], + "credit_map": [ + 0.027648238465189934, + 0.029405655339360237, + 0.025724118575453758, + 0.03837704285979271, + 0.030163640156388283, + 0.027104035019874573, + 0.027290059253573418, + 0.03582529351115227, + 0.007784890942275524, + 0.009448475204408169, + 0.013575018383562565, + 0.014321860857307911, + 0.011657069437205791, + 0.012657553888857365, + 0.009094354696571827, + 0.009021070785820484, + 0.004679849836975336, + 0.004825667943805456, + 0.004865542054176331, + 0.0052574193105101585, + 0.0063748229295015335, + 0.004481488838791847, + 0.005043250974267721, + 0.0038666375912725925, + 0.004582539200782776, + 0.005271292291581631, + 0.006925900466740131, + 0.0036889908369630575, + 0.007346884347498417, + 0.005254616495221853, + 0.004829831887036562, + 0.003216005628928542, + 0.0030513072852045298, + 0.003973241429775953, + 0.0036384323611855507, + 0.004748473409563303, + 0.006781823001801968, + 0.004481673706322908, + 0.007437652908265591, + 0.004940683022141457, + 0.015586485154926777, + 0.017617328092455864, + 0.008909783326089382, + 0.012014870531857014, + 0.008501246571540833, + 0.009293904528021812, + 0.0132945841178298, + 0.009181258268654346, + 0.004977976903319359, + 0.002240478992462158, + 0.0070632509887218475, + 0.0038324156776070595, + 0.0030390725005418062, + 0.005104586482048035, + 0.0030340016819536686, + 0.007000258192420006, + 0.00575537933036685, + 0.006792531348764896, + 0.007130580022931099, + 0.006226000376045704, + 0.00676051015034318, + 0.007188400253653526, + 0.006429454777389765, + 0.006255859974771738, + 0.02110985666513443, + 0.01959165930747986, + 0.023921873420476913, + 0.016920704394578934, + 0.010884388349950314, + 0.016883378848433495, + 0.025683868676424026, + 0.0419822633266449, + 0.017399227246642113, + 0.015498507767915726, + 0.015237134881317616, + 0.01855781115591526, + 0.03202107921242714, + 0.0221840962767601, + 0.042695045471191406, + 0.027536453679203987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.006740333512425423 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067327553406357765 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.0067327553406357765, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067327553406357765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15631064772605896, + 0.08397465199232101, + 0.18375566601753235, + 0.04618168622255325, + 0.1304740607738495, + 0.11064193397760391, + 0.16445773839950562, + 0.040721289813518524, + 0.05409570783376694, + 0.029386622831225395 + ], + "credit_map": [ + 0.02678205631673336, + 0.031085964292287827, + 0.016277115792036057, + 0.017272835597395897, + 0.016214124858379364, + 0.018635887652635574, + 0.013797306455671787, + 0.016245361417531967, + 0.007376824505627155, + 0.009157204069197178, + 0.015482652001082897, + 0.014606015756726265, + 0.010241792537271976, + 0.009826812893152237, + 0.009258045814931393, + 0.008025304414331913, + 0.029254645109176636, + 0.0236574225127697, + 0.03318387269973755, + 0.021244170144200325, + 0.019741393625736237, + 0.02150803804397583, + 0.01978333480656147, + 0.015382783487439156, + 0.00746322562918067, + 0.005356904119253159, + 0.0072478330694139, + 0.004589882213622332, + 0.005548226181417704, + 0.005327697843313217, + 0.006524323485791683, + 0.004123595543205738, + 0.009849674999713898, + 0.01297407504171133, + 0.011530144140124321, + 0.012442049570381641, + 0.02461494691669941, + 0.00926757138222456, + 0.03434532880783081, + 0.015450268983840942, + 0.016560640186071396, + 0.03359562158584595, + 0.0075677246786653996, + 0.007253759540617466, + 0.006417499389499426, + 0.007621658034622669, + 0.007875132374465466, + 0.023749900981783867, + 0.030367346480488777, + 0.013446765020489693, + 0.017181379720568657, + 0.019447561353445053, + 0.012955006211996078, + 0.013241822831332684, + 0.011357572861015797, + 0.04646027833223343, + 0.0041254619136452675, + 0.0038669987116008997, + 0.008005904965102673, + 0.005001813173294067, + 0.003583061508834362, + 0.005902567412704229, + 0.004284370690584183, + 0.00595111446455121, + 0.005442901514470577, + 0.005314445123076439, + 0.00853647943586111, + 0.010155700147151947, + 0.0045467340387403965, + 0.004784842953085899, + 0.00847623310983181, + 0.006838371977210045, + 0.0023304629139602184, + 0.0024481427390128374, + 0.003239973681047559, + 0.004684465005993843, + 0.004716168157756329, + 0.002442006254568696, + 0.005750413052737713, + 0.0037749900948256254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.0067327553406357765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808720037341118 + ], + "opposite_logprob_tokens": [ + -3.2880871295928955 + ], + "logprob_reward": -0.03808720037341118, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808720037341118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18133264780044556, + 0.12062960118055344, + 0.09550157189369202, + 0.12137605994939804, + 0.11913236230611801, + 0.07471059262752533, + 0.03475029394030571, + 0.03776983544230461, + 0.1612292230129242, + 0.0535677894949913 + ], + "credit_map": [ + 0.025915643200278282, + 0.02956444025039673, + 0.018095754086971283, + 0.025045420974493027, + 0.021873243153095245, + 0.019510632380843163, + 0.016846967861056328, + 0.024480560794472694, + 0.010520962998270988, + 0.013744940049946308, + 0.017324190586805344, + 0.017958803102374077, + 0.016408251598477364, + 0.016832705587148666, + 0.013021235354244709, + 0.014818520285189152, + 0.008641828782856464, + 0.010661911219358444, + 0.014527869410812855, + 0.009617116302251816, + 0.013775710947811604, + 0.015442496165633202, + 0.012503130361437798, + 0.010331512428820133, + 0.015254102647304535, + 0.014704114757478237, + 0.022369224578142166, + 0.0166444294154644, + 0.011888120323419571, + 0.0156110730022192, + 0.013961507007479668, + 0.010943488217890263, + 0.010314451530575752, + 0.014937176369130611, + 0.01237387116998434, + 0.013477231375873089, + 0.016091644763946533, + 0.019894249737262726, + 0.016981957480311394, + 0.015061774291098118, + 0.010697416961193085, + 0.007370797451585531, + 0.011494934558868408, + 0.0083082877099514, + 0.0037202469538897276, + 0.008636650629341602, + 0.00841835979372263, + 0.01606389693915844, + 0.0034464909695088863, + 0.0035843944642692804, + 0.004989502020180225, + 0.004434711299836636, + 0.0038681530859321356, + 0.006186393555253744, + 0.004388358909636736, + 0.003852289402857423, + 0.005042129661887884, + 0.005044696386903524, + 0.0036582606844604015, + 0.005581775214523077, + 0.004283731337636709, + 0.00494834128767252, + 0.004387742839753628, + 0.004823159892112017, + 0.022380076348781586, + 0.01594839245080948, + 0.026496855542063713, + 0.014412744902074337, + 0.008896016515791416, + 0.014909686520695686, + 0.017883889377117157, + 0.040301546454429626, + 0.005177266430109739, + 0.00467738788574934, + 0.004703597631305456, + 0.007181162014603615, + 0.009453200735151768, + 0.006472742184996605, + 0.008087125606834888, + 0.007815306074917316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.03808720037341118 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast expanse of rolling green hills under a clear azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an antique electric sewing machine painted black with faded gold filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sewing machine moving down the hillside in a clumsy, bumpy motion involving swaying and rolling sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an old braided electrical cord visible dragging behind the machine and tangling in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a warm, nostalgic, and imaginative fantasy atmosphere featuring soft golden lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000493762141559273 + ], + "opposite_logprob_tokens": [ + -7.62549352645874 + ], + "logprob_reward": -0.000493762141559273, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.000493762141559273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22942307591438293, + 0.07998256385326385, + 0.0478367917239666, + 0.04059645161032677, + 0.029183831065893173, + 0.05114838108420372, + 0.152495339512825, + 0.03620129078626633, + 0.25692448019981384, + 0.07620780169963837 + ], + "credit_map": [ + 0.04022479057312012, + 0.045146457850933075, + 0.020940139889717102, + 0.02728303149342537, + 0.0247739739716053, + 0.025141559541225433, + 0.017938850447535515, + 0.02797427959740162, + 0.008101807907223701, + 0.008390982635319233, + 0.017771458253264427, + 0.013345160521566868, + 0.008632288314402103, + 0.009963366203010082, + 0.007568559609353542, + 0.00620893994346261, + 0.005111432168632746, + 0.005238458048552275, + 0.00592022854834795, + 0.006027892231941223, + 0.0068304771557450294, + 0.007683427073061466, + 0.005867257248610258, + 0.005157620180398226, + 0.005227529909461737, + 0.004617186263203621, + 0.007648560684174299, + 0.005940583534538746, + 0.003347043413668871, + 0.005102963652461767, + 0.005368668120354414, + 0.003343915333971381, + 0.004608168266713619, + 0.0039154766127467155, + 0.003555122297257185, + 0.003327180165797472, + 0.0028341261204332113, + 0.0037374256644397974, + 0.004100335296243429, + 0.0031059973407536745, + 0.009266684763133526, + 0.010312899015843868, + 0.006375226192176342, + 0.004699489567428827, + 0.0035275942645967007, + 0.005718495696783066, + 0.005293227266520262, + 0.005954765249043703, + 0.025570454075932503, + 0.01385927852243185, + 0.015549949370324612, + 0.017522819340229034, + 0.019315047189593315, + 0.024159390479326248, + 0.014156381599605083, + 0.02236201800405979, + 0.003723854897543788, + 0.0027434120420366526, + 0.007053811103105545, + 0.0029553743079304695, + 0.004092111252248287, + 0.00795944593846798, + 0.0033910057973116636, + 0.00428227661177516, + 0.03999052196741104, + 0.027742428705096245, + 0.03565458953380585, + 0.018674494698643684, + 0.011131182312965393, + 0.02469341643154621, + 0.027634987607598305, + 0.07140286266803741, + 0.007848820649087429, + 0.008644389919936657, + 0.009028921835124493, + 0.011343305930495262, + 0.007123077753931284, + 0.005954258609563112, + 0.011226399801671505, + 0.0150386281311512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.000493762141559273 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_8/index_tensor([2080, 2080])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_8/index_tensor([2080, 2080])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6a0504fb409f80c7187d3f93993d7d11be206d18 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_8/index_tensor([2080, 2080])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "first-person perspective depicts a slow and steady upward tracking shot, as if we were aboard a spaceship with a vast, seamless glass dome, steadily and deliberately ascending upstream. The dominant image is a magnificent galaxy, not static, but a majestic cosmic waterfall cascading from the very top of the frame. Countless luminous stars, nebulae of fiery magenta and deep sapphire blue, and golden cosmic dust converge into a dynamic torrent of light. As we slowly ascend, nearby star clusters glide gracefully past the glass dome like luminous bubbles, shattering and dissipating, while distant galaxies maintain their majestic, unchanging presence. The entire scene is illuminated by this radiant stream of cosmic light, whose ever-changing light casts soft yet dynamic shadows on the (imaginary) cockpit. The overall color tone dynamically alternates between deep cool colors and fiery warm colors, creating an awe-inspiring, tranquil, epic and otherworldly atmosphere, allowing people to fully immerse themselves in this magnificent journey against the flow of light and time.", + "seed": 10871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_8/index_2080_seed_10871_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04783326806500554, + "vlm_logprob": -0.04783326806500554, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04783326806500554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059684306383132935 + ], + "opposite_logprob_tokens": [ + -3.0596842765808105 + ], + "logprob_reward": -0.059684306383132935, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.059684306383132935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20822648704051971, + 0.14970995485782623, + 0.04668833315372467, + 0.04734073579311371, + 0.05815066024661064, + 0.15041588246822357, + 0.039053551852703094, + 0.06189049780368805, + 0.19835875928401947, + 0.04016507789492607 + ], + "credit_map": [ + 0.015397455543279648, + 0.031198883429169655, + 0.03278578445315361, + 0.020327210426330566, + 0.02457159012556076, + 0.028798962011933327, + 0.027588749304413795, + 0.02755788154900074, + 0.01730523444712162, + 0.013943351805210114, + 0.022514846175909042, + 0.024337930604815483, + 0.016004711389541626, + 0.019071288406848907, + 0.019076339900493622, + 0.017456278204917908, + 0.006113103125244379, + 0.008745988830924034, + 0.005396823398768902, + 0.005710701923817396, + 0.005279984325170517, + 0.00440607825294137, + 0.006157236639410257, + 0.004878421314060688, + 0.0053145200945436954, + 0.004071665462106466, + 0.008399425074458122, + 0.00598950358107686, + 0.005651024170219898, + 0.005179289728403091, + 0.007850871421396732, + 0.004884445574134588, + 0.0058726356364786625, + 0.006213142070919275, + 0.008766830898821354, + 0.008628823794424534, + 0.007157126907259226, + 0.00588678615167737, + 0.01018377486616373, + 0.005441547371447086, + 0.008008643053472042, + 0.004336888436228037, + 0.056910641491413116, + 0.006039667408913374, + 0.004337869584560394, + 0.05561528727412224, + 0.0038842300418764353, + 0.011282671242952347, + 0.0043993256986141205, + 0.0038371833506971598, + 0.009066098369657993, + 0.003973727580159903, + 0.0038843953516334295, + 0.006452173460274935, + 0.003887626575306058, + 0.0035530258901417255, + 0.005659631919115782, + 0.004523961804807186, + 0.01571301929652691, + 0.005516077391803265, + 0.0031405941117554903, + 0.015592226758599281, + 0.00431312108412385, + 0.007431871723383665, + 0.010850242339074612, + 0.004461870063096285, + 0.05665965750813484, + 0.007643251679837704, + 0.007089618593454361, + 0.0793318897485733, + 0.006907701026648283, + 0.025414545089006424, + 0.003530448768287897, + 0.004236568231135607, + 0.00435774028301239, + 0.007547738030552864, + 0.0037488299421966076, + 0.007079745642840862, + 0.004774284549057484, + 0.004889731295406818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.059684306383132935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789995789527893 + ], + "opposite_logprob_tokens": [ + -2.5789995193481445 + ], + "logprob_reward": -0.0789995789527893, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789995789527893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02090260572731495, + 0.016835743561387062, + 0.022017458453774452, + 0.009464223869144917, + 0.022620871663093567, + 0.3099123239517212, + 0.016488978639245033, + 0.135978102684021, + 0.42214518785476685, + 0.02363450825214386 + ], + "credit_map": [ + 0.0027043104637414217, + 0.0027954289689660072, + 0.002924407832324505, + 0.0020133822690695524, + 0.002377085154876113, + 0.0027344301342964172, + 0.0026883776299655437, + 0.00266518397256732, + 0.0017257619183510542, + 0.0020026848651468754, + 0.0016547604463994503, + 0.0029464366380125284, + 0.0019027611706405878, + 0.0019998454954475164, + 0.002281506545841694, + 0.0023219871800392866, + 0.0023389928974211216, + 0.0036303901579231024, + 0.0027006673626601696, + 0.002556926803663373, + 0.0026904100086539984, + 0.003446802031248808, + 0.0021607077214866877, + 0.0024925617035478354, + 0.001151941018179059, + 0.0008871296304278076, + 0.0013085288228467107, + 0.0009928378276526928, + 0.0012224613456055522, + 0.0011202206369489431, + 0.0011277804151177406, + 0.0016533239977434278, + 0.002642698585987091, + 0.0031526831444352865, + 0.003328697057440877, + 0.0032977662049233913, + 0.0026754497084766626, + 0.0027113286778330803, + 0.00289357197470963, + 0.0019186766585335135, + 0.01987762562930584, + 0.007666845340281725, + 0.12331279367208481, + 0.013896603137254715, + 0.007812041323632002, + 0.1088777557015419, + 0.008223870769143105, + 0.020244820043444633, + 0.0015751359751448035, + 0.0021933955140411854, + 0.0020040932577103376, + 0.001686115749180317, + 0.003005378879606724, + 0.0019698927644640207, + 0.0020731149706989527, + 0.001981851179152727, + 0.007868874818086624, + 0.006925726775079966, + 0.011781997978687286, + 0.0046490333043038845, + 0.004855954088270664, + 0.045160070061683655, + 0.005156691186130047, + 0.04957975819706917, + 0.0264575257897377, + 0.01103247981518507, + 0.1314769834280014, + 0.01767539419233799, + 0.01495411992073059, + 0.16533851623535156, + 0.013597935438156128, + 0.041612204164266586, + 0.002968200482428074, + 0.002403959399089217, + 0.0031276578083634377, + 0.00287221884354949, + 0.0029329294338822365, + 0.0029358600731939077, + 0.003606014419347048, + 0.002787666628137231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.0789995789527893 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101437050849199 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101437050849199, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101437050849199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04298132285475731, + 0.05493912100791931, + 0.0422932468354702, + 0.05270832031965256, + 0.04936700686812401, + 0.6584476232528687, + 0.023018544539809227, + 0.020957134664058685, + 0.04170836880803108, + 0.013579299673438072 + ], + "credit_map": [ + 0.004253027029335499, + 0.008281445130705833, + 0.00461628008633852, + 0.004016994498670101, + 0.008347253315150738, + 0.005272973328828812, + 0.0047014751471579075, + 0.0034918738529086113, + 0.005546059459447861, + 0.006579467095434666, + 0.005278446711599827, + 0.007218922022730112, + 0.006095574703067541, + 0.007919462397694588, + 0.008253288455307484, + 0.008047900162637234, + 0.005110476166009903, + 0.005954979453235865, + 0.005133694037795067, + 0.004595151636749506, + 0.004496391396969557, + 0.004307206254452467, + 0.006344916764646769, + 0.006350433453917503, + 0.004643503110855818, + 0.004890918266028166, + 0.009852049872279167, + 0.006716337986290455, + 0.006781664676964283, + 0.006005451083183289, + 0.008166369050741196, + 0.005652029067277908, + 0.005793449003249407, + 0.005269940476864576, + 0.006264118477702141, + 0.006235998589545488, + 0.0065468065440654755, + 0.005855064373463392, + 0.009480956941843033, + 0.003920671995729208, + 0.04096144810318947, + 0.013395171612501144, + 0.26627397537231445, + 0.027485717087984085, + 0.015751779079437256, + 0.23440784215927124, + 0.015263435430824757, + 0.04490826278924942, + 0.0019092120928689837, + 0.002301724860444665, + 0.003911038860678673, + 0.0016066835960373282, + 0.0024183548521250486, + 0.006010372191667557, + 0.0019542318768799305, + 0.0029069269075989723, + 0.0018753116019070148, + 0.0018283870304003358, + 0.002550031989812851, + 0.002117362106218934, + 0.001574867987073958, + 0.004628116264939308, + 0.0022241640836000443, + 0.004158894531428814, + 0.0035597688984125853, + 0.002844679169356823, + 0.008536718785762787, + 0.0037324714940041304, + 0.004025822039693594, + 0.005449915304780006, + 0.002827799180522561, + 0.010731194168329239, + 0.001323564792983234, + 0.0013825611677020788, + 0.0019054223084822297, + 0.0020274915732443333, + 0.0012260277289897203, + 0.002155100693926215, + 0.0018706574337556958, + 0.0016884732758626342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.004101437050849199 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361869648098946 + ], + "opposite_logprob_tokens": [ + -3.6513619422912598 + ], + "logprob_reward": -0.026361869648098946, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361869648098946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03726748377084732, + 0.029319528490304947, + 0.02683124504983425, + 0.01994437724351883, + 0.07408220320940018, + 0.47218087315559387, + 0.06261065602302551, + 0.17172323167324066, + 0.09424103796482086, + 0.011799346655607224 + ], + "credit_map": [ + 0.004289109259843826, + 0.006021473091095686, + 0.006132934708148241, + 0.003287757048383355, + 0.004847064148634672, + 0.0045365565456449986, + 0.004825954791158438, + 0.0033266341779381037, + 0.0032119492534548044, + 0.0032663748133927584, + 0.004275879822671413, + 0.004406493157148361, + 0.0033049241174012423, + 0.0036485602613538504, + 0.0029387527611106634, + 0.0042665936052799225, + 0.003175160149112344, + 0.0030217920430004597, + 0.003620374947786331, + 0.003257198492065072, + 0.004810277838259935, + 0.0025699245743453503, + 0.0035171860363334417, + 0.0028593300376087427, + 0.003011611057445407, + 0.0016783784376457334, + 0.0025187984574586153, + 0.002769995015114546, + 0.001506482600234449, + 0.0023418671917170286, + 0.0031231443863362074, + 0.0029940991662442684, + 0.006777138914912939, + 0.008922100998461246, + 0.008786959573626518, + 0.010279728099703789, + 0.009583651088178158, + 0.007953069172799587, + 0.015579938888549805, + 0.006199620198458433, + 0.02822042815387249, + 0.009610519744455814, + 0.18879900872707367, + 0.019557006657123566, + 0.011258644051849842, + 0.1707644760608673, + 0.011316805146634579, + 0.03265397250652313, + 0.006100663915276527, + 0.008600159548223019, + 0.005415335297584534, + 0.006449342705309391, + 0.011764181777834892, + 0.0066331978887319565, + 0.009285638108849525, + 0.008362137712538242, + 0.007329972460865974, + 0.007705770432949066, + 0.007316115777939558, + 0.004700219724327326, + 0.0037816513795405626, + 0.06703249365091324, + 0.006667543668299913, + 0.06718945503234863, + 0.006980672013014555, + 0.0035034394823014736, + 0.031140940263867378, + 0.005022142082452774, + 0.0036862390115857124, + 0.03202179819345474, + 0.00326113891787827, + 0.008624670095741749, + 0.0012231703149154782, + 0.0011781712528318167, + 0.0013868066016584635, + 0.0018024644814431667, + 0.001079575507901609, + 0.002447147388011217, + 0.00131600396707654, + 0.0013660074910148978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.026361869648098946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001914829015732 + ], + "opposite_logprob_tokens": [ + -2.695019245147705 + ], + "logprob_reward": -0.07001914829015732, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001914829015732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10140322148799896, + 0.0712117925286293, + 0.0876537561416626, + 0.0415850467979908, + 0.04133634641766548, + 0.3268929123878479, + 0.06221967563033104, + 0.14479520916938782, + 0.07787492126226425, + 0.045027170330286026 + ], + "credit_map": [ + 0.012953490018844604, + 0.013103483244776726, + 0.015734143555164337, + 0.0094532435759902, + 0.012383295223116875, + 0.012484114617109299, + 0.01146332174539566, + 0.013828124850988388, + 0.00839877501130104, + 0.008376352488994598, + 0.009555432945489883, + 0.009177714586257935, + 0.008001917973160744, + 0.00953791942447424, + 0.007239051163196564, + 0.0109246289357543, + 0.010576467961072922, + 0.01896059699356556, + 0.007682467345148325, + 0.01125040277838707, + 0.010676427744328976, + 0.009294562973082066, + 0.009644604288041592, + 0.009568221867084503, + 0.004250544589012861, + 0.003245377680286765, + 0.008060943335294724, + 0.004655034281313419, + 0.004477268550544977, + 0.005882186349481344, + 0.005585602484643459, + 0.005428089760243893, + 0.0037258442025631666, + 0.006304795388132334, + 0.00522643281146884, + 0.004936103709042072, + 0.005081063602119684, + 0.004570527467876673, + 0.007524851243942976, + 0.003966726828366518, + 0.025757119059562683, + 0.02231430821120739, + 0.11203131079673767, + 0.022544126957654953, + 0.016437295824289322, + 0.08810245990753174, + 0.01790648140013218, + 0.02179979719221592, + 0.00557943619787693, + 0.008940835483372211, + 0.006727272644639015, + 0.0056461067870259285, + 0.00852410402148962, + 0.009239091537892818, + 0.009682883508503437, + 0.007879937998950481, + 0.00992182269692421, + 0.006791784893721342, + 0.02387508936226368, + 0.007931013591587543, + 0.006839679088443518, + 0.04825128987431526, + 0.00826029758900404, + 0.032924238592386246, + 0.006215238478034735, + 0.003747022245079279, + 0.024655990302562714, + 0.005094380583614111, + 0.0038057039491832256, + 0.02039705030620098, + 0.0038950068410485983, + 0.010064523667097092, + 0.006108236499130726, + 0.006269820965826511, + 0.0065536219626665115, + 0.005710292141884565, + 0.004554059822112322, + 0.006034962832927704, + 0.004632676485925913, + 0.005163499154150486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.07001914829015732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person perspective depicts a slow and steady upward tracking shot, as if we were aboard a spaceship with a vast, seamless glass dome, steadily and deliberately ascending upstream. The dominant image is a magnificent galaxy, not static, but a majestic cosmic waterfall cascading from the very top of the frame. Countless luminous stars, nebulae of fiery magenta and deep sapphire blue, and golden cosmic dust converge into a dynamic torrent of light. As we slowly ascend, nearby star clusters glide gracefully past the glass dome like luminous bubbles, shattering and dissipating, while distant galaxies maintain their majestic, unchanging presence. The entire scene is illuminated by this radiant stream of cosmic light, whose ever-changing light casts soft yet dynamic shadows on the (imaginary) cockpit. The overall color tone dynamically alternates between deep cool colors and fiery warm colors, creating an awe-inspiring, tranquil, epic and otherworldly atmosphere, allowing people to fully immerse themselves in this magnificent journey against the flow of light and time.", + "seed": 10872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_8/index_2080_seed_10872_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1632862638682127, + "vlm_logprob": -0.1632862638682127, + "vlm_yes_no_margin": 2.2249999761581423, + "vlm_token_credit": -0.1632862638682127, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.48263031244277954 + ], + "opposite_logprob_tokens": [ + -0.9826303124427795 + ], + "logprob_reward": -0.48263031244277954, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.48263031244277954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4067699909210205, + 0.04322817549109459, + 0.029450012370944023, + 0.08447029441595078, + 0.0891343280673027, + 0.10350955277681351, + 0.09117615222930908, + 0.026790650561451912, + 0.04934316873550415, + 0.07612772285938263 + ], + "credit_map": [ + 0.039243023842573166, + 0.052842944860458374, + 0.03659951686859131, + 0.026054203510284424, + 0.04930458590388298, + 0.06640756875276566, + 0.04638025909662247, + 0.08993784338235855, + 0.0047383238561451435, + 0.0034935050643980503, + 0.005785168614238501, + 0.0033863240387290716, + 0.005294452421367168, + 0.007199262734502554, + 0.0058883000165224075, + 0.007442834321409464, + 0.0028905114158988, + 0.002402391517534852, + 0.0031615060288459063, + 0.0030724857933819294, + 0.004810469690710306, + 0.004819427616894245, + 0.0037748864851891994, + 0.004518332425504923, + 0.005670508835464716, + 0.006181888747960329, + 0.005940072238445282, + 0.007208275143057108, + 0.006578402128070593, + 0.023172002285718918, + 0.0056116217747330666, + 0.024107515811920166, + 0.007558078970760107, + 0.013001540675759315, + 0.007946961559355259, + 0.006431483663618565, + 0.010941245593130589, + 0.01898343861103058, + 0.0072088586166501045, + 0.01706271432340145, + 0.008455755189061165, + 0.01052150223404169, + 0.016251258552074432, + 0.01629963144659996, + 0.012461015954613686, + 0.01421502884477377, + 0.013014967553317547, + 0.012290396727621555, + 0.007810949347913265, + 0.016468685120344162, + 0.012037252075970173, + 0.009895800612866879, + 0.010887746699154377, + 0.014988322742283344, + 0.009498417377471924, + 0.009588978253304958, + 0.002715465845540166, + 0.002462174044921994, + 0.002954456489533186, + 0.002093961462378502, + 0.002223532646894455, + 0.005020578857511282, + 0.0036162491887807846, + 0.005704231094568968, + 0.006130688823759556, + 0.004063523840159178, + 0.008887042291462421, + 0.006025731563568115, + 0.006794845685362816, + 0.006080606020987034, + 0.004639833699911833, + 0.006720895413309336, + 0.00831360649317503, + 0.007350376807153225, + 0.01673128642141819, + 0.008572365157306194, + 0.010618355125188828, + 0.00999178271740675, + 0.005949430167675018, + 0.008600511588156223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.48263031244277954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283187687397003 + ], + "opposite_logprob_tokens": [ + -2.017831802368164 + ], + "logprob_reward": -0.14283187687397003, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283187687397003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1824321299791336, + 0.049897655844688416, + 0.12736162543296814, + 0.07785096019506454, + 0.051657889038324356, + 0.09277171641588211, + 0.12946905195713043, + 0.041693028062582016, + 0.17149905860424042, + 0.07536688446998596 + ], + "credit_map": [ + 0.022027604281902313, + 0.04529484361410141, + 0.01114269532263279, + 0.010023576207458973, + 0.01764019764959812, + 0.024638982489705086, + 0.0121105732396245, + 0.039553649723529816, + 0.006504860706627369, + 0.006227585952728987, + 0.007193289697170258, + 0.004806687589734793, + 0.004974403418600559, + 0.008031859993934631, + 0.004994033370167017, + 0.007164931856095791, + 0.012293953448534012, + 0.008128016255795956, + 0.022122425958514214, + 0.010699350386857986, + 0.014949574135243893, + 0.019538667052984238, + 0.012067669071257114, + 0.027561968192458153, + 0.005954507272690535, + 0.007378405891358852, + 0.005599077790975571, + 0.009415295906364918, + 0.007503808010369539, + 0.01838233880698681, + 0.004584718029946089, + 0.019032813608646393, + 0.004198719747364521, + 0.007378472946584225, + 0.006508296821266413, + 0.006549432408064604, + 0.0063646952621638775, + 0.006347252521663904, + 0.006192192900925875, + 0.0081188278272748, + 0.007922936230897903, + 0.0077415648847818375, + 0.016684208065271378, + 0.013364465907216072, + 0.008835241198539734, + 0.012794464826583862, + 0.008541095070540905, + 0.016887741163372993, + 0.011182849295437336, + 0.02126975543797016, + 0.01554605271667242, + 0.013101283460855484, + 0.013927306048572063, + 0.02053431235253811, + 0.013786349445581436, + 0.020121140405535698, + 0.00502370297908783, + 0.00515530351549387, + 0.005347326397895813, + 0.005811572540551424, + 0.004652045201510191, + 0.004562302492558956, + 0.004918586928397417, + 0.0062221866101026535, + 0.019372887909412384, + 0.022741103544831276, + 0.02788182534277439, + 0.021458061411976814, + 0.01996280811727047, + 0.022161051630973816, + 0.014450956135988235, + 0.02347036823630333, + 0.00996385607868433, + 0.008751796558499336, + 0.01612253114581108, + 0.006800659466534853, + 0.0127716651186347, + 0.009666148573160172, + 0.00531706353649497, + 0.005973162595182657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.14283187687397003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598655566573143 + ], + "opposite_logprob_tokens": [ + -3.8955986499786377 + ], + "logprob_reward": -0.020598655566573143, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598655566573143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18860368430614471, + 0.07063502818346024, + 0.13469535112380981, + 0.16079048812389374, + 0.10731986910104752, + 0.07332389801740646, + 0.04961042106151581, + 0.10445056110620499, + 0.0756397470831871, + 0.034930940717458725 + ], + "credit_map": [ + 0.018273577094078064, + 0.021688083186745644, + 0.02081756852567196, + 0.01851797103881836, + 0.02627943456172943, + 0.030202047899365425, + 0.02582063525915146, + 0.02700437419116497, + 0.009841226041316986, + 0.006723913364112377, + 0.010160891339182854, + 0.008044756017625332, + 0.008288447745144367, + 0.008745220489799976, + 0.007373489905148745, + 0.01145708467811346, + 0.015081975609064102, + 0.008650663308799267, + 0.021751506254076958, + 0.00989605113863945, + 0.017850730568170547, + 0.022511696442961693, + 0.011810819618403912, + 0.027141908183693886, + 0.013716802932322025, + 0.015230627730488777, + 0.012754817493259907, + 0.020688118413090706, + 0.014486782252788544, + 0.0359225794672966, + 0.011757602915167809, + 0.036233142018318176, + 0.009147774428129196, + 0.01176067627966404, + 0.01378202810883522, + 0.01110828761011362, + 0.012352239340543747, + 0.01759435422718525, + 0.014927631244063377, + 0.01664688065648079, + 0.007650340441614389, + 0.007208541966974735, + 0.01107489038258791, + 0.009235778823494911, + 0.008505810983479023, + 0.011119265109300613, + 0.006099769379943609, + 0.012429501861333847, + 0.004901411011815071, + 0.0071014720015227795, + 0.006988754030317068, + 0.005800748709589243, + 0.005534913390874863, + 0.007570961955934763, + 0.005845740903168917, + 0.005866422317922115, + 0.011265669949352741, + 0.011527201160788536, + 0.015093164518475533, + 0.009929734282195568, + 0.00857425294816494, + 0.016940433531999588, + 0.0078001380898058414, + 0.023319974541664124, + 0.008629688061773777, + 0.00785724725574255, + 0.012937994673848152, + 0.008825968950986862, + 0.010148690082132816, + 0.009778653271496296, + 0.007669514045119286, + 0.009791988879442215, + 0.003975271247327328, + 0.0034816039260476828, + 0.007740079425275326, + 0.003144263755530119, + 0.005602325778454542, + 0.00525527261197567, + 0.0027630296535789967, + 0.00296909362077713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.020598655566573143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005249708890915 + ], + "opposite_logprob_tokens": [ + -2.6950523853302 + ], + "logprob_reward": -0.07005249708890915, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005249708890915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.436388224363327, + 0.047235842794179916, + 0.07250049710273743, + 0.08369507640600204, + 0.05608496814966202, + 0.09633909165859222, + 0.05899687856435776, + 0.04163162782788277, + 0.025533758103847504, + 0.08159400522708893 + ], + "credit_map": [ + 0.04001273214817047, + 0.05584486573934555, + 0.038661204278469086, + 0.02700347825884819, + 0.04960227012634277, + 0.07532389461994171, + 0.050641193985939026, + 0.09929861128330231, + 0.006925491616129875, + 0.004592018201947212, + 0.004722995683550835, + 0.004385017789900303, + 0.005791569594293833, + 0.0076015121303498745, + 0.004983182530850172, + 0.008234054781496525, + 0.00659830030053854, + 0.0039036464877426624, + 0.010705467313528061, + 0.005276061128824949, + 0.011124192737042904, + 0.0129080293700099, + 0.006768592167645693, + 0.015216199681162834, + 0.008016488514840603, + 0.009674911387264729, + 0.008188960142433643, + 0.009617640636861324, + 0.011038384400308132, + 0.0137188620865345, + 0.00935650710016489, + 0.014083320274949074, + 0.004308422561734915, + 0.007404748350381851, + 0.005444612354040146, + 0.005732523277401924, + 0.00664372555911541, + 0.011004207655787468, + 0.006350044626742601, + 0.009196684695780277, + 0.008875913918018341, + 0.010183979757130146, + 0.014586523175239563, + 0.014603981748223305, + 0.0105965631082654, + 0.013286695815622807, + 0.011402773670852184, + 0.012802657671272755, + 0.005213536787778139, + 0.010377965867519379, + 0.008326767943799496, + 0.006856800988316536, + 0.006213048007339239, + 0.009506206959486008, + 0.0058502615429461, + 0.006652293261140585, + 0.005107637029141188, + 0.004481206648051739, + 0.006620450410991907, + 0.0036668581888079643, + 0.003603211371228099, + 0.007004172541201115, + 0.0038879564963281155, + 0.007260138168931007, + 0.003058383706957102, + 0.002817565342411399, + 0.003856508992612362, + 0.002296313177794218, + 0.004191339947283268, + 0.003256245981901884, + 0.003028975334018469, + 0.0030284251552075148, + 0.010883129201829433, + 0.0072959722019732, + 0.019320467486977577, + 0.00921441987156868, + 0.010436668992042542, + 0.01053529605269432, + 0.006160201039165258, + 0.007747847121208906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.07005249708890915 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031797736883163 + ], + "opposite_logprob_tokens": [ + -2.35031795501709 + ], + "logprob_reward": -0.10031797736883163, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031797736883163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3542589843273163, + 0.049389444291591644, + 0.06535475701093674, + 0.07023090869188309, + 0.07755327969789505, + 0.05431881174445152, + 0.051609888672828674, + 0.02463642880320549, + 0.06607121229171753, + 0.18657630681991577 + ], + "credit_map": [ + 0.03341731056571007, + 0.047047004103660583, + 0.02871949039399624, + 0.021853765472769737, + 0.04038665071129799, + 0.06040697917342186, + 0.04020889848470688, + 0.08221889287233353, + 0.007663818541914225, + 0.006192886270582676, + 0.0047348798252642155, + 0.005865941289812326, + 0.0049112774431705475, + 0.007435919251292944, + 0.005680892150849104, + 0.006903828587383032, + 0.0064201438799500465, + 0.00468951603397727, + 0.008082165382802486, + 0.0061516668647527695, + 0.007913853973150253, + 0.012507561594247818, + 0.00606202706694603, + 0.013527821749448776, + 0.005358515307307243, + 0.0042383745312690735, + 0.006119955331087112, + 0.011041439138352871, + 0.006625457666814327, + 0.016364308074116707, + 0.0048636263236403465, + 0.015619229525327682, + 0.006020172964781523, + 0.009560158476233482, + 0.006549451500177383, + 0.007525170221924782, + 0.0078092035837471485, + 0.017060864716768265, + 0.007978564128279686, + 0.015049690380692482, + 0.006609899923205376, + 0.004100905265659094, + 0.010985426604747772, + 0.008105174638330936, + 0.0063176522962749004, + 0.007387724705040455, + 0.005009690299630165, + 0.005802334286272526, + 0.004577807150781155, + 0.009394333697855473, + 0.006323471665382385, + 0.006392703391611576, + 0.005505557172000408, + 0.008644514717161655, + 0.005114285741001368, + 0.005657215602695942, + 0.0026909904554486275, + 0.0023927059955894947, + 0.002893403172492981, + 0.002197111025452614, + 0.0023305481299757957, + 0.004964700434356928, + 0.0023603348527103662, + 0.004806634038686752, + 0.006853519473224878, + 0.005725131835788488, + 0.008735054172575474, + 0.01128319464623928, + 0.008224811404943466, + 0.007780063897371292, + 0.008469009771943092, + 0.009000430814921856, + 0.02507590875029564, + 0.017388861626386642, + 0.03550378978252411, + 0.02243540808558464, + 0.027739739045500755, + 0.023745840415358543, + 0.0167697761207819, + 0.017916997894644737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.10031797736883163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person perspective depicts a slow and steady upward tracking shot, as if we were aboard a spaceship with a vast, seamless glass dome, steadily and deliberately ascending upstream. The dominant image is a magnificent galaxy, not static, but a majestic cosmic waterfall cascading from the very top of the frame. Countless luminous stars, nebulae of fiery magenta and deep sapphire blue, and golden cosmic dust converge into a dynamic torrent of light. As we slowly ascend, nearby star clusters glide gracefully past the glass dome like luminous bubbles, shattering and dissipating, while distant galaxies maintain their majestic, unchanging presence. The entire scene is illuminated by this radiant stream of cosmic light, whose ever-changing light casts soft yet dynamic shadows on the (imaginary) cockpit. The overall color tone dynamically alternates between deep cool colors and fiery warm colors, creating an awe-inspiring, tranquil, epic and otherworldly atmosphere, allowing people to fully immerse themselves in this magnificent journey against the flow of light and time.", + "seed": 10873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_8/index_2080_seed_10873_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04250485124066472, + "vlm_logprob": -0.04250485124066472, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04250485124066472, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08795051276683807 + ], + "opposite_logprob_tokens": [ + -2.7129504680633545 + ], + "logprob_reward": -0.08795051276683807, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08795051276683807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4034481644630432, + 0.0372123084962368, + 0.04677474871277809, + 0.07329043000936508, + 0.0523090735077858, + 0.048315875232219696, + 0.11407184600830078, + 0.07082898169755936, + 0.06595316529273987, + 0.08779547363519669 + ], + "credit_map": [ + 0.051962800323963165, + 0.07174850255250931, + 0.0430828332901001, + 0.07809168845415115, + 0.03092307038605213, + 0.04313035309314728, + 0.05982330068945885, + 0.0246855691075325, + 0.005124720744788647, + 0.004811118822544813, + 0.004213958978652954, + 0.004780749324709177, + 0.0030350061133503914, + 0.006786029785871506, + 0.0036584138870239258, + 0.004802309907972813, + 0.004844407085329294, + 0.009122415445744991, + 0.004207972437143326, + 0.004302095156162977, + 0.0049962191842496395, + 0.005149281118065119, + 0.007786042522639036, + 0.006366312503814697, + 0.006875477731227875, + 0.004826589021831751, + 0.015743793919682503, + 0.0051805623807013035, + 0.0067141554318368435, + 0.023404542356729507, + 0.004202971700578928, + 0.006342328153550625, + 0.006305949296802282, + 0.005261278245598078, + 0.006124527659267187, + 0.005831776186823845, + 0.0078580928966403, + 0.00947926752269268, + 0.004489172250032425, + 0.006959004793316126, + 0.004134566523134708, + 0.0046336837112903595, + 0.006060936022549868, + 0.005313723813742399, + 0.007323634345084429, + 0.00877823680639267, + 0.004129811190068722, + 0.007941276766359806, + 0.010043411515653133, + 0.010479324497282505, + 0.010497656650841236, + 0.010778500698506832, + 0.02793753333389759, + 0.019521024078130722, + 0.009910578839480877, + 0.014903807081282139, + 0.004803821444511414, + 0.008324136026203632, + 0.006077212747186422, + 0.00852493941783905, + 0.012675396166741848, + 0.010021536611020565, + 0.007614647503942251, + 0.012787283398211002, + 0.0073143234476447105, + 0.009499620646238327, + 0.006834946572780609, + 0.009534262120723724, + 0.00916336104273796, + 0.006775372661650181, + 0.007364320103079081, + 0.009466948918998241, + 0.008087947964668274, + 0.01113148033618927, + 0.009220319800078869, + 0.016879763454198837, + 0.00726450327783823, + 0.01029279362410307, + 0.01102739479392767, + 0.013891266658902168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.08795051276683807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496829003095627 + ], + "opposite_logprob_tokens": [ + -2.9299683570861816 + ], + "logprob_reward": -0.05496829003095627, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496829003095627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13868622481822968, + 0.15077047049999237, + 0.040633972734212875, + 0.0670381635427475, + 0.0774105042219162, + 0.09915965795516968, + 0.1095823422074318, + 0.1334201991558075, + 0.09355293214321136, + 0.08974558860063553 + ], + "credit_map": [ + 0.01883004419505596, + 0.018965156748890877, + 0.013547252863645554, + 0.018260883167386055, + 0.01786879077553749, + 0.01280861534178257, + 0.02757788449525833, + 0.010827586054801941, + 0.016546031460165977, + 0.0140233701094985, + 0.030411235988140106, + 0.03461070731282234, + 0.013948584906756878, + 0.015490525402128696, + 0.014402884989976883, + 0.01133711263537407, + 0.005807384382933378, + 0.005245860666036606, + 0.005433422513306141, + 0.00598784489557147, + 0.0031649633310735226, + 0.004306256305426359, + 0.00495428591966629, + 0.005733950529247522, + 0.009163123555481434, + 0.0041542197577655315, + 0.010667809285223484, + 0.006498717237263918, + 0.006581544876098633, + 0.017676454037427902, + 0.005143478512763977, + 0.007152811624109745, + 0.008171342313289642, + 0.006702051497995853, + 0.014101214706897736, + 0.009977119043469429, + 0.012780971825122833, + 0.009076453745365143, + 0.008668581023812294, + 0.007932763546705246, + 0.012889700010418892, + 0.008904782123863697, + 0.017966661602258682, + 0.011071648448705673, + 0.010132316499948502, + 0.011137038469314575, + 0.007531751878559589, + 0.019525747746229172, + 0.012125128880143166, + 0.011457891203463078, + 0.013972923159599304, + 0.011241835542023182, + 0.020538894459605217, + 0.015708157792687416, + 0.01022416353225708, + 0.01431333925575018, + 0.012019414454698563, + 0.027567412704229355, + 0.01718071661889553, + 0.01709471270442009, + 0.016278184950351715, + 0.014739388599991798, + 0.014035839587450027, + 0.014504518359899521, + 0.011366973631083965, + 0.013803058303892612, + 0.008568892255425453, + 0.01730640046298504, + 0.010333750396966934, + 0.0110758813098073, + 0.01127694547176361, + 0.009821015410125256, + 0.010799591429531574, + 0.011260408908128738, + 0.00987514853477478, + 0.01696673035621643, + 0.009216371923685074, + 0.010212212800979614, + 0.0108411880210042, + 0.010573930107057095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.05496829003095627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955214146524668 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955214146524668, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955214146524668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14045602083206177, + 0.0914444550871849, + 0.27343785762786865, + 0.1306135058403015, + 0.2102009654045105, + 0.02031148038804531, + 0.023421723395586014, + 0.0547783300280571, + 0.018729375675320625, + 0.03660639375448227 + ], + "credit_map": [ + 0.015505611896514893, + 0.01497819647192955, + 0.01367457490414381, + 0.01905766688287258, + 0.017985256388783455, + 0.020435426384210587, + 0.022319307550787926, + 0.016499970108270645, + 0.01392503920942545, + 0.010254043154418468, + 0.013050828129053116, + 0.014606580138206482, + 0.00827078428119421, + 0.012810100801289082, + 0.008116382174193859, + 0.010410682298243046, + 0.035854414105415344, + 0.07152985036373138, + 0.027145639061927795, + 0.023447975516319275, + 0.02489391900599003, + 0.028984753414988518, + 0.023215005174279213, + 0.038366276770830154, + 0.01058214996010065, + 0.00799238309264183, + 0.028934327885508537, + 0.011705762706696987, + 0.010715451091527939, + 0.04075608775019646, + 0.006672647316008806, + 0.013254676014184952, + 0.01828153245151043, + 0.02110164612531662, + 0.035061679780483246, + 0.023115500807762146, + 0.041621576994657516, + 0.029210515320301056, + 0.02122904546558857, + 0.020579438656568527, + 0.002590559422969818, + 0.0017103217542171478, + 0.002846052637323737, + 0.0016723882872611284, + 0.002693049143999815, + 0.003388776211068034, + 0.0017657559365034103, + 0.003644574899226427, + 0.0022194397170096636, + 0.0025619175285100937, + 0.0029811174608767033, + 0.0021008667536079884, + 0.0038956741336733103, + 0.004395132884383202, + 0.002677339129149914, + 0.0025902320630848408, + 0.0046865083277225494, + 0.008879697881639004, + 0.007707011420279741, + 0.008410383947193623, + 0.005879036616533995, + 0.005877121817320585, + 0.006038210354745388, + 0.007300354074686766, + 0.0022850031964480877, + 0.002722526201978326, + 0.00185370200779289, + 0.002047500805929303, + 0.0030660689808428288, + 0.0025988249108195305, + 0.0018774140626192093, + 0.002278332831338048, + 0.0038897101767361164, + 0.004361427389085293, + 0.003762629348784685, + 0.005169473588466644, + 0.004902046639472246, + 0.004743675701320171, + 0.004240130074322224, + 0.005537297111004591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.005955214146524668 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591065287590027 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020591065287590027, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591065287590027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09562808275222778, + 0.058057110756635666, + 0.2136768102645874, + 0.046438027173280716, + 0.20101754367351532, + 0.1361221820116043, + 0.06137571856379509, + 0.07963885366916656, + 0.05132291838526726, + 0.05672271549701691 + ], + "credit_map": [ + 0.013313954696059227, + 0.013664117082953453, + 0.009738250635564327, + 0.013444209471344948, + 0.009453468024730682, + 0.014730959199368954, + 0.012447272427380085, + 0.00883584562689066, + 0.007527076173573732, + 0.011151999235153198, + 0.009039154276251793, + 0.007495193276554346, + 0.004472556058317423, + 0.006608123425394297, + 0.006989594548940659, + 0.0047734142281115055, + 0.025267507880926132, + 0.04905044287443161, + 0.020554915070533752, + 0.018121622502803802, + 0.024256346747279167, + 0.022854899987578392, + 0.018834013491868973, + 0.03473706543445587, + 0.005913682747632265, + 0.005032390356063843, + 0.005918254144489765, + 0.005726091098040342, + 0.005466096103191376, + 0.006998323369771242, + 0.004795530345290899, + 0.00658766133710742, + 0.020301634445786476, + 0.024193530902266502, + 0.03451821580529213, + 0.021439339965581894, + 0.027326535433530807, + 0.028849810361862183, + 0.022390786558389664, + 0.021997690200805664, + 0.01406670268625021, + 0.011821837164461613, + 0.014407647773623466, + 0.012915235012769699, + 0.019426001235842705, + 0.028300315141677856, + 0.015723945572972298, + 0.019460508599877357, + 0.004764168057590723, + 0.004728261847048998, + 0.0070387329906225204, + 0.0059761106967926025, + 0.014430388808250427, + 0.010538480244576931, + 0.006298073101788759, + 0.007601500954478979, + 0.007325123529881239, + 0.0130159268155694, + 0.009286943823099136, + 0.009381859563291073, + 0.01125166192650795, + 0.008568902499973774, + 0.009179678745567799, + 0.011628757230937481, + 0.0066938078962266445, + 0.0065060872584581375, + 0.004788237623870373, + 0.009451007470488548, + 0.007307399529963732, + 0.006545253098011017, + 0.00564885139465332, + 0.004382276441901922, + 0.0068540628999471664, + 0.01035146415233612, + 0.005568105261772871, + 0.010211414657533169, + 0.004627458285540342, + 0.005487031769007444, + 0.007625882048159838, + 0.005997295491397381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.020591065287590027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043059173971414566 + ], + "opposite_logprob_tokens": [ + -3.1680591106414795 + ], + "logprob_reward": -0.043059173971414566, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043059173971414566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13176608085632324, + 0.08015105873346329, + 0.07498668879270554, + 0.04396236687898636, + 0.1139744445681572, + 0.07370717823505402, + 0.11117298156023026, + 0.24052296578884125, + 0.06985056400299072, + 0.05990564450621605 + ], + "credit_map": [ + 0.01527006272226572, + 0.01773669384419918, + 0.013924858532845974, + 0.02281864732503891, + 0.013703471049666405, + 0.02008698508143425, + 0.016397669911384583, + 0.011827697046101093, + 0.009535218589007854, + 0.007747394032776356, + 0.012654952704906464, + 0.012106635607779026, + 0.010183502919971943, + 0.01167425699532032, + 0.008136006072163582, + 0.008113096468150616, + 0.00881778821349144, + 0.017268536612391472, + 0.007170509081333876, + 0.0068541462533175945, + 0.007711504120379686, + 0.008842675015330315, + 0.007162414491176605, + 0.011159118264913559, + 0.006213670130819082, + 0.004426354076713324, + 0.005514835007488728, + 0.004227459896355867, + 0.004865105729550123, + 0.008444948121905327, + 0.004619785118848085, + 0.00565020926296711, + 0.012066428549587727, + 0.012416950426995754, + 0.015387061983346939, + 0.014452105388045311, + 0.018662801012396812, + 0.014615556225180626, + 0.012549093924462795, + 0.013824447058141232, + 0.008428123779594898, + 0.006944950670003891, + 0.013738895766437054, + 0.007813338190317154, + 0.01050221361219883, + 0.010409246198832989, + 0.005858052521944046, + 0.010012352839112282, + 0.00961507298052311, + 0.01647772826254368, + 0.021955246105790138, + 0.01283078920096159, + 0.012633923441171646, + 0.011300729587674141, + 0.011539053171873093, + 0.014820443466305733, + 0.024147192016243935, + 0.03668256103992462, + 0.02870781533420086, + 0.025966264307498932, + 0.04256436973810196, + 0.029695307835936546, + 0.0252042505890131, + 0.027555197477340698, + 0.008977114222943783, + 0.009192004799842834, + 0.008196627721190453, + 0.011641668155789375, + 0.007313575595617294, + 0.007567627355456352, + 0.007636321242898703, + 0.009325627237558365, + 0.008627297356724739, + 0.00809512473642826, + 0.0049207573756575584, + 0.007637751754373312, + 0.007714622188359499, + 0.006900718901306391, + 0.009251653216779232, + 0.006757717113941908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.043059173971414566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person perspective depicts a slow and steady upward tracking shot, as if we were aboard a spaceship with a vast, seamless glass dome, steadily and deliberately ascending upstream. The dominant image is a magnificent galaxy, not static, but a majestic cosmic waterfall cascading from the very top of the frame. Countless luminous stars, nebulae of fiery magenta and deep sapphire blue, and golden cosmic dust converge into a dynamic torrent of light. As we slowly ascend, nearby star clusters glide gracefully past the glass dome like luminous bubbles, shattering and dissipating, while distant galaxies maintain their majestic, unchanging presence. The entire scene is illuminated by this radiant stream of cosmic light, whose ever-changing light casts soft yet dynamic shadows on the (imaginary) cockpit. The overall color tone dynamically alternates between deep cool colors and fiery warm colors, creating an awe-inspiring, tranquil, epic and otherworldly atmosphere, allowing people to fully immerse themselves in this magnificent journey against the flow of light and time.", + "seed": 10874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_8/index_2080_seed_10874_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06429953649640083, + "vlm_logprob": -0.06429953649640083, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.06429953649640083, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12048093974590302 + ], + "opposite_logprob_tokens": [ + -2.245481014251709 + ], + "logprob_reward": -0.12048093974590302, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12048093974590302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11389399319887161, + 0.2046358585357666, + 0.06917967647314072, + 0.0714184045791626, + 0.06776244938373566, + 0.09215924888849258, + 0.09068144857883453, + 0.1741224080324173, + 0.0461726188659668, + 0.069973886013031 + ], + "credit_map": [ + 0.016707584261894226, + 0.01705917902290821, + 0.017261117696762085, + 0.021189076825976372, + 0.011592426337301731, + 0.011866014450788498, + 0.007868790067732334, + 0.010349802672863007, + 0.025401325896382332, + 0.02929546870291233, + 0.030221497640013695, + 0.026003623381257057, + 0.031208299100399017, + 0.023574158549308777, + 0.015917908400297165, + 0.023013588041067123, + 0.00773966358974576, + 0.012545821256935596, + 0.009461402893066406, + 0.0069359662011265755, + 0.00936302449554205, + 0.009309681132435799, + 0.006119330879300833, + 0.007704786024987698, + 0.00828343816101551, + 0.015544158406555653, + 0.007237676065415144, + 0.007590946275740862, + 0.010624013841152191, + 0.008029932156205177, + 0.007424537092447281, + 0.006683703046292067, + 0.007420769426971674, + 0.011893189512193203, + 0.00821749772876501, + 0.007783970329910517, + 0.006678143050521612, + 0.007641087286174297, + 0.009379900991916656, + 0.008747885003685951, + 0.00492705637589097, + 0.02850434184074402, + 0.010298146866261959, + 0.007231005467474461, + 0.013235250487923622, + 0.008948572911322117, + 0.007040814496576786, + 0.011974062770605087, + 0.01058783009648323, + 0.007498716935515404, + 0.014939081855118275, + 0.010157880373299122, + 0.007895784452557564, + 0.012865150347352028, + 0.013466225005686283, + 0.013270778581500053, + 0.010930516757071018, + 0.013855739496648312, + 0.01813308335840702, + 0.01677439548075199, + 0.014214739203453064, + 0.04101099446415901, + 0.03139444440603256, + 0.027808506041765213, + 0.005146846175193787, + 0.005600936245173216, + 0.005852698814123869, + 0.005266136955469847, + 0.00642595998942852, + 0.005761364474892616, + 0.006134343799203634, + 0.005984330549836159, + 0.006502426229417324, + 0.007266463711857796, + 0.00831483956426382, + 0.009553886950016022, + 0.008392896503210068, + 0.008753781206905842, + 0.011617028154432774, + 0.009572571143507957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.12048093974590302 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706968188285828 + ], + "opposite_logprob_tokens": [ + -2.1270697116851807 + ], + "logprob_reward": -0.12706968188285828, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706968188285828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043420061469078064, + 0.09466197341680527, + 0.08957332372665405, + 0.0847063809633255, + 0.06553611904382706, + 0.14461639523506165, + 0.09733282774686813, + 0.24363090097904205, + 0.08402305841445923, + 0.05249897018074989 + ], + "credit_map": [ + 0.008222604170441628, + 0.004256543703377247, + 0.0055612181313335896, + 0.005623931530863047, + 0.007693672087043524, + 0.005285273306071758, + 0.0027807482983917, + 0.0039960709400475025, + 0.012181230820715427, + 0.005855238530784845, + 0.026546979323029518, + 0.00893304031342268, + 0.007444954942911863, + 0.008377796038985252, + 0.004124632570892572, + 0.021198108792304993, + 0.007824288681149483, + 0.011412291787564754, + 0.00835429411381483, + 0.0068751233629882336, + 0.02173100784420967, + 0.011571084149181843, + 0.011156411841511726, + 0.010648827068507671, + 0.009455445222556591, + 0.015111323446035385, + 0.007767933886498213, + 0.007781127467751503, + 0.012545241042971611, + 0.011168528348207474, + 0.012343344278633595, + 0.008533436805009842, + 0.005845344625413418, + 0.008132831193506718, + 0.006562090944498777, + 0.010062630288302898, + 0.008050546050071716, + 0.007788433227688074, + 0.012159532867372036, + 0.006934707518666983, + 0.014442064799368382, + 0.016038062050938606, + 0.025867236778140068, + 0.02693559043109417, + 0.018658950924873352, + 0.016613425686955452, + 0.012191186659038067, + 0.013869866728782654, + 0.0083767743781209, + 0.008185936138033867, + 0.011304885149002075, + 0.009452523663640022, + 0.007822236977517605, + 0.017519962042570114, + 0.020148925483226776, + 0.014521585777401924, + 0.014496395364403725, + 0.013093644753098488, + 0.020049655809998512, + 0.014419982209801674, + 0.01490685623139143, + 0.08137030154466629, + 0.032453760504722595, + 0.05284030735492706, + 0.0110179977491498, + 0.008270482532680035, + 0.00946778617799282, + 0.010046486742794514, + 0.011078578419983387, + 0.012188724242150784, + 0.010500557720661163, + 0.011452441103756428, + 0.004524239804595709, + 0.0028819753788411617, + 0.01790771633386612, + 0.004015775863081217, + 0.00500538619235158, + 0.0072399633936584, + 0.004744890611618757, + 0.006179023068398237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.12706968188285828 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085116304457188 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011085116304457188, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085116304457188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03152671456336975, + 0.13820235431194305, + 0.06948424130678177, + 0.09216507524251938, + 0.07360473275184631, + 0.15288926661014557, + 0.07992822676897049, + 0.25433531403541565, + 0.08522012829780579, + 0.02264392003417015 + ], + "credit_map": [ + 0.0044772750698029995, + 0.0025591517332941294, + 0.005794643424451351, + 0.003969790879637003, + 0.005004602950066328, + 0.004232549574226141, + 0.00241365609690547, + 0.0030750457663089037, + 0.018488524481654167, + 0.006740669719874859, + 0.03753545507788658, + 0.010217934846878052, + 0.010664599016308784, + 0.013862216845154762, + 0.005705535411834717, + 0.034987423568964005, + 0.0066407350823283195, + 0.008046261966228485, + 0.0073436046950519085, + 0.00467669079080224, + 0.017128149047493935, + 0.010075364261865616, + 0.008854851126670837, + 0.006718586198985577, + 0.010476458817720413, + 0.01677420176565647, + 0.008599108085036278, + 0.008900280110538006, + 0.014749151654541492, + 0.011702707037329674, + 0.011740094982087612, + 0.009223072789609432, + 0.008296232670545578, + 0.009787609800696373, + 0.00713984714820981, + 0.009353948757052422, + 0.010583581402897835, + 0.009107380174100399, + 0.012530525214970112, + 0.006805606186389923, + 0.014085646718740463, + 0.01909256912767887, + 0.022707022726535797, + 0.022708812728524208, + 0.026190591976046562, + 0.023111753165721893, + 0.012256513349711895, + 0.012736363336443901, + 0.00693722628057003, + 0.006691483780741692, + 0.009397074580192566, + 0.007834183983504772, + 0.005821810103952885, + 0.014110309071838856, + 0.016348928213119507, + 0.012787215411663055, + 0.015463996678590775, + 0.013744453899562359, + 0.017445659264922142, + 0.013682807795703411, + 0.015423727221786976, + 0.09020613133907318, + 0.03176827356219292, + 0.05660027265548706, + 0.011145546101033688, + 0.008853957988321781, + 0.010179074481129646, + 0.009505948051810265, + 0.011646329425275326, + 0.012845837511122227, + 0.010073443874716759, + 0.01096998993307352, + 0.00196303753182292, + 0.0016901263734325767, + 0.005724360700696707, + 0.0024807758163660765, + 0.002472790889441967, + 0.0027454053051769733, + 0.0036082188598811626, + 0.001959205372259021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.011085116304457188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420364435762167 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.01420364435762167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420364435762167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09378788620233536, + 0.11783178150653839, + 0.05153684690594673, + 0.05144524574279785, + 0.06322020292282104, + 0.05784422531723976, + 0.07566610723733902, + 0.13573552668094635, + 0.06098116189241409, + 0.29195091128349304 + ], + "credit_map": [ + 0.011360698379576206, + 0.010538166388869286, + 0.011603753082454205, + 0.016623571515083313, + 0.012668321840465069, + 0.01262928731739521, + 0.007099434267729521, + 0.011264661327004433, + 0.015504341572523117, + 0.012545978650450706, + 0.0270723644644022, + 0.011410138569772243, + 0.012768948450684547, + 0.01150442287325859, + 0.00699002156034112, + 0.020035572350025177, + 0.004854448139667511, + 0.009418703615665436, + 0.008641303516924381, + 0.004480813164263964, + 0.005569763015955687, + 0.008289831690490246, + 0.003971564583480358, + 0.0063104210421442986, + 0.00525993388146162, + 0.007333078887313604, + 0.007115400396287441, + 0.006554854102432728, + 0.00860008504241705, + 0.006173143163323402, + 0.005955666769295931, + 0.004453087225556374, + 0.006918596103787422, + 0.00784500315785408, + 0.007581303361803293, + 0.007501389365643263, + 0.007671640254557133, + 0.0076028346084058285, + 0.008883464150130749, + 0.009215977042913437, + 0.005542191211134195, + 0.01087912730872631, + 0.006194776389747858, + 0.003860295517370105, + 0.008047547191381454, + 0.009449048899114132, + 0.004598286002874374, + 0.009272956289350986, + 0.009632697328925133, + 0.0054692612029612064, + 0.014742112718522549, + 0.0062141879461705685, + 0.006595247890800238, + 0.011896642856299877, + 0.010143525898456573, + 0.010972436517477036, + 0.00821546372026205, + 0.007722238078713417, + 0.01204773411154747, + 0.008222670294344425, + 0.008548073470592499, + 0.046385910362005234, + 0.016701368615031242, + 0.027892081066966057, + 0.008844972588121891, + 0.007175571750849485, + 0.009436842054128647, + 0.0048236967995762825, + 0.009057741612195969, + 0.007876289077103138, + 0.0064750793389976025, + 0.007290973793715239, + 0.02764083631336689, + 0.019017890095710754, + 0.06534332036972046, + 0.026526127010583878, + 0.039325226098299026, + 0.03681661933660507, + 0.052704401314258575, + 0.024576522409915924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.01420364435762167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, simulating a view from inside a spaceship with a glass dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow and steady upward tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dominant visual a cosmic waterfall of light cascading from the top of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do nearby star clusters glide past the glass dome like luminous bubbles that shatter and dissipate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft yet dynamic shadows cast on the imaginary cockpit by the radiant stream of cosmic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865830019116402 + ], + "opposite_logprob_tokens": [ + -3.0486583709716797 + ], + "logprob_reward": -0.04865830019116402, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865830019116402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13031359016895294, + 0.08111967891454697, + 0.10483237355947495, + 0.10191434621810913, + 0.054212529212236404, + 0.09295479208230972, + 0.1466539055109024, + 0.1261584609746933, + 0.07432461529970169, + 0.0875156968832016 + ], + "credit_map": [ + 0.02779371663928032, + 0.014708555303514004, + 0.01648596301674843, + 0.019061259925365448, + 0.015325621701776981, + 0.01408880390226841, + 0.011466064490377903, + 0.011383607052266598, + 0.009932439774274826, + 0.00855233520269394, + 0.016450045630335808, + 0.011526969261467457, + 0.00789548084139824, + 0.011738770641386509, + 0.005170508753508329, + 0.009853136725723743, + 0.015572688542306423, + 0.019247479736804962, + 0.014423731714487076, + 0.00948599074035883, + 0.011958453804254532, + 0.014860247261822224, + 0.009353907778859138, + 0.009929874911904335, + 0.010046187788248062, + 0.013355394825339317, + 0.009067190811038017, + 0.009345158003270626, + 0.017919445410370827, + 0.014191575348377228, + 0.0170524250715971, + 0.010936971753835678, + 0.007522277534008026, + 0.007420793641358614, + 0.005887545645236969, + 0.006295972503721714, + 0.006923000793904066, + 0.006139830220490694, + 0.0077865635976195335, + 0.006236549001187086, + 0.009964056313037872, + 0.009747564792633057, + 0.012938708066940308, + 0.009708528406918049, + 0.011520802043378353, + 0.01746295765042305, + 0.012130241841077805, + 0.009481930173933506, + 0.012885416857898235, + 0.010621560737490654, + 0.017831573262810707, + 0.01597057655453682, + 0.011469680815935135, + 0.025652555748820305, + 0.030448317527770996, + 0.02177422307431698, + 0.009595812298357487, + 0.007502158172428608, + 0.01042535062879324, + 0.008993166498839855, + 0.010064440779387951, + 0.0331948921084404, + 0.025662384927272797, + 0.020720254629850388, + 0.00622481619939208, + 0.006469844840466976, + 0.007632893044501543, + 0.010836523026227951, + 0.011389357037842274, + 0.009162924252450466, + 0.011605366133153439, + 0.011002886109054089, + 0.00899734441190958, + 0.006625547539442778, + 0.01060893852263689, + 0.008527985773980618, + 0.014009584672749043, + 0.011260898783802986, + 0.018904151394963264, + 0.008581249974668026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.04865830019116402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_9/index_tensor([4600, 4600])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_9/index_tensor([4600, 4600])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2242f75802171b796df9f74f34b9a4dd053102db --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_9/index_tensor([4600, 4600])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely close, static shot is locked on the car's analog speedometer, the only sharp point of focus in the frame. The speedometer's face is a deep matte black, its numerals and hash marks glowing a vibrant, fiery orange. The polished red needle sweeps unwaveringly and rapidly upward, rapidly climbing past the 100, 120, and 140 mph markers in smooth, relentless motion. In the background, the tunnel environment is rendered as a swirl of intense, natural motion blur; overhead sodium lights and the concrete walls carve a chaotic jumble of warm yellow and dark gray horizontal lines, creating a monumental sense of speed. The flickering, distorted reflections of these light streaks slide across the instrument panel's clear glass cover, the only sign of the world beyond. The stark contrast between the bright, glowing dial and the completely blurred surroundings, combined with the dark, high-contrast color palette, creates an intense and thrilling atmosphere of pure, unadulterated speed.", + "seed": 25305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_9/index_4600_seed_25305_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05611462723463774, + "vlm_logprob": -0.05611462723463774, + "vlm_yes_no_margin": 3.424999976158142, + "vlm_token_credit": -0.05611462723463774, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15217028558254242 + ], + "opposite_logprob_tokens": [ + -2.152170181274414 + ], + "logprob_reward": -0.15217028558254242, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.15217028558254242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26932093501091003, + 0.14088405668735504, + 0.11449073255062103, + 0.045328184962272644, + 0.062450043857097626, + 0.08932656049728394, + 0.1259828805923462, + 0.07299772650003433, + 0.036329835653305054, + 0.042888984084129333 + ], + "credit_map": [ + 0.026640594005584717, + 0.02805340848863125, + 0.032022636383771896, + 0.03524940088391304, + 0.04414835944771767, + 0.028789488598704338, + 0.03924816474318504, + 0.03516888618469238, + 0.015425360761582851, + 0.027021730318665504, + 0.01791587844491005, + 0.01799517497420311, + 0.0166013240814209, + 0.013945178128778934, + 0.013336940668523312, + 0.018642472103238106, + 0.022252844646573067, + 0.008581845089793205, + 0.010471261106431484, + 0.014465251937508583, + 0.00963730737566948, + 0.007394672837108374, + 0.034609388560056686, + 0.007078155409544706, + 0.005321504082530737, + 0.005563249345868826, + 0.0069296713918447495, + 0.005617871880531311, + 0.006984150968492031, + 0.004152082372456789, + 0.005265369080007076, + 0.005494289565831423, + 0.010187757201492786, + 0.009956862777471542, + 0.01066550426185131, + 0.005990035831928253, + 0.008807496167719364, + 0.006270881742238998, + 0.005825144238770008, + 0.004746366757899523, + 0.024436166509985924, + 0.007104651536792517, + 0.03053194098174572, + 0.007456157356500626, + 0.002874848898500204, + 0.002965579740703106, + 0.008018159307539463, + 0.00593905383720994, + 0.03200459107756615, + 0.014399685896933079, + 0.025573624297976494, + 0.041680168360471725, + 0.0038411845453083515, + 0.0024130085948854685, + 0.004242083057761192, + 0.0018285327823832631, + 0.012944119051098824, + 0.01684414967894554, + 0.007938300259411335, + 0.007921455428004265, + 0.00992176029831171, + 0.005398436915129423, + 0.006479500327259302, + 0.005550006404519081, + 0.00700759282335639, + 0.006616939324885607, + 0.004449696745723486, + 0.004253684543073177, + 0.003696004394441843, + 0.002710184082388878, + 0.0030016114469617605, + 0.004594122525304556, + 0.004632364492863417, + 0.007909140549600124, + 0.00694693811237812, + 0.009060616604983807, + 0.003147493815049529, + 0.0028821188025176525, + 0.004145858343690634, + 0.004164454992860556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.15217028558254242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023321032524108887 + ], + "opposite_logprob_tokens": [ + -3.7733211517333984 + ], + "logprob_reward": -0.023321032524108887, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023321032524108887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22184886038303375, + 0.1415257751941681, + 0.16353264451026917, + 0.049546338617801666, + 0.06430542469024658, + 0.14630864560604095, + 0.06977114081382751, + 0.08071984350681305, + 0.0375605933368206, + 0.02488068677484989 + ], + "credit_map": [ + 0.030955132097005844, + 0.0336117297410965, + 0.02125626988708973, + 0.02206747978925705, + 0.03166918456554413, + 0.023821251466870308, + 0.032301370054483414, + 0.026166444644331932, + 0.015777120366692543, + 0.02413949742913246, + 0.015738988295197487, + 0.014784068800508976, + 0.016404487192630768, + 0.014687618240714073, + 0.01955394074320793, + 0.02044004760682583, + 0.03470320627093315, + 0.012636172585189342, + 0.012170428410172462, + 0.018115384504199028, + 0.012754367664456367, + 0.011237523518502712, + 0.05183342099189758, + 0.010082138702273369, + 0.00542135676369071, + 0.006882389076054096, + 0.005525624845176935, + 0.004983560647815466, + 0.007748427800834179, + 0.004210478160530329, + 0.008170479908585548, + 0.006604023277759552, + 0.010394822806119919, + 0.009850455448031425, + 0.011453074403107166, + 0.006058624479919672, + 0.00889156386256218, + 0.006711745169013739, + 0.0054627046920359135, + 0.005482435692101717, + 0.038294222205877304, + 0.008727777749300003, + 0.054863620549440384, + 0.014131643809378147, + 0.004353980999439955, + 0.003994705155491829, + 0.0120106041431427, + 0.009932093322277069, + 0.013876831158995628, + 0.005273742601275444, + 0.01565634086728096, + 0.024512413889169693, + 0.0023712385445833206, + 0.00267258333042264, + 0.0032070865854620934, + 0.002200900577008724, + 0.016308618709445, + 0.021564224734902382, + 0.007614310830831528, + 0.006830115802586079, + 0.009424932301044464, + 0.006363075226545334, + 0.006690323352813721, + 0.005924243479967117, + 0.005686088465154171, + 0.0059835040010511875, + 0.004016475286334753, + 0.0043190522119402885, + 0.005635657347738743, + 0.0027469282504171133, + 0.004495179746299982, + 0.004677707329392433, + 0.004044353496283293, + 0.0037658067885786295, + 0.0036900152917951345, + 0.003222925355657935, + 0.0024019968695938587, + 0.0024190908297896385, + 0.0026820721104741096, + 0.002654426731169224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.023321032524108887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002904266119003 + ], + "opposite_logprob_tokens": [ + -2.6950290203094482 + ], + "logprob_reward": -0.07002904266119003, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002904266119003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1220027282834053, + 0.11808694154024124, + 0.07689616829156876, + 0.04885938763618469, + 0.03841323405504227, + 0.2819575071334839, + 0.10355137288570404, + 0.05399039015173912, + 0.10643768310546875, + 0.04980456456542015 + ], + "credit_map": [ + 0.020895501598715782, + 0.021326467394828796, + 0.012054916471242905, + 0.010432939045131207, + 0.018986910581588745, + 0.014144569635391235, + 0.010009948164224625, + 0.01415148377418518, + 0.015517155639827251, + 0.027072947472333908, + 0.013501102104783058, + 0.011320515535771847, + 0.014113017357885838, + 0.008901852183043957, + 0.01083382498472929, + 0.01682651787996292, + 0.011862759478390217, + 0.008161922916769981, + 0.009181587025523186, + 0.008554099127650261, + 0.009389295242726803, + 0.006279249209910631, + 0.017680473625659943, + 0.005786790046840906, + 0.0041734036058187485, + 0.0062927426770329475, + 0.0056809294037520885, + 0.005858692806214094, + 0.007279293145984411, + 0.006105063483119011, + 0.007774224039167166, + 0.0056950394064188, + 0.005807952955365181, + 0.004687042441219091, + 0.005688962526619434, + 0.003958297427743673, + 0.005996078718453646, + 0.004721228498965502, + 0.003383887931704521, + 0.004169786348938942, + 0.0759153664112091, + 0.020832180976867676, + 0.09987320750951767, + 0.027327505871653557, + 0.008821690455079079, + 0.006405988708138466, + 0.022385798394680023, + 0.02039577253162861, + 0.034254513680934906, + 0.02741609513759613, + 0.008049195632338524, + 0.015927381813526154, + 0.0058695655316114426, + 0.004004285670816898, + 0.00445815222337842, + 0.003572190413251519, + 0.007273494731634855, + 0.006879317574203014, + 0.009972797706723213, + 0.008147960528731346, + 0.005645781755447388, + 0.003761259838938713, + 0.004794820211827755, + 0.007514956407248974, + 0.013076881878077984, + 0.012641370296478271, + 0.011817138642072678, + 0.01503976434469223, + 0.009015158750116825, + 0.01457222644239664, + 0.011496446095407009, + 0.018778694793581963, + 0.004409367218613625, + 0.008334465324878693, + 0.007018929347395897, + 0.01092612650245428, + 0.0059126694686710835, + 0.005196144338697195, + 0.004334740806370974, + 0.0036721222568303347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.07002904266119003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252371542155743 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252371542155743, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252371542155743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25057676434516907, + 0.11509652435779572, + 0.32059231400489807, + 0.028393728658556938, + 0.02822517789900303, + 0.10704076290130615, + 0.04710595682263374, + 0.047587472945451736, + 0.035345934331417084, + 0.02003541961312294 + ], + "credit_map": [ + 0.025607295334339142, + 0.024997688829898834, + 0.03324981406331062, + 0.03984754532575607, + 0.04053911939263344, + 0.02402220480144024, + 0.031288690865039825, + 0.031024370342493057, + 0.015266877599060535, + 0.01990090310573578, + 0.00986490212380886, + 0.008606359362602234, + 0.016056304797530174, + 0.01089490856975317, + 0.016980091109871864, + 0.0175261739641428, + 0.07849615812301636, + 0.01975264772772789, + 0.02256208285689354, + 0.035531818866729736, + 0.01958916336297989, + 0.015089621767401695, + 0.11406837403774261, + 0.01550243142992258, + 0.0032422603107988834, + 0.003904670709744096, + 0.004080663900822401, + 0.003330330830067396, + 0.004472473636269569, + 0.0024730898439884186, + 0.003141380613669753, + 0.003748856484889984, + 0.0031333104707300663, + 0.0030523338355123997, + 0.0040038153529167175, + 0.0026249729562550783, + 0.004916512407362461, + 0.004970364272594452, + 0.003147410461679101, + 0.002376455580815673, + 0.013578512705862522, + 0.019053803756833076, + 0.01343795470893383, + 0.00972028449177742, + 0.020930135622620583, + 0.006899828091263771, + 0.013209765776991844, + 0.01021046843379736, + 0.00974978506565094, + 0.004274427890777588, + 0.010973148047924042, + 0.016355181112885475, + 0.0014850201550871134, + 0.0012659529456868768, + 0.0018578548915684223, + 0.0011445839190855622, + 0.0071676867082715034, + 0.009211495518684387, + 0.006336481310427189, + 0.006069459952414036, + 0.005922751501202583, + 0.003584682708606124, + 0.004352353513240814, + 0.004942559637129307, + 0.005301448050886393, + 0.005134252831339836, + 0.00462579820305109, + 0.004850068595260382, + 0.003333328291773796, + 0.004030765034258366, + 0.0031867974903434515, + 0.0048834760673344135, + 0.0025472925044596195, + 0.003991545643657446, + 0.0032364926300942898, + 0.0038254123646765947, + 0.0015993486158549786, + 0.0015306008281186223, + 0.0017760783666744828, + 0.001528646913357079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.005252371542155743 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800403863191605 + ], + "opposite_logprob_tokens": [ + -3.5298004150390625 + ], + "logprob_reward": -0.029800403863191605, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800403863191605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08959472924470901, + 0.06444048881530762, + 0.1658092439174652, + 0.041626155376434326, + 0.03519989922642708, + 0.08454225957393646, + 0.26854151487350464, + 0.16161519289016724, + 0.033703990280628204, + 0.054926495999097824 + ], + "credit_map": [ + 0.00849879439920187, + 0.010477742180228233, + 0.008350186981260777, + 0.006999680306762457, + 0.017972972244024277, + 0.013780583627521992, + 0.011000930331647396, + 0.012513847090303898, + 0.008623536676168442, + 0.014211256988346577, + 0.006985444109886885, + 0.007529032416641712, + 0.006504353601485491, + 0.006063330452889204, + 0.0073445201851427555, + 0.0071790143847465515, + 0.03859914094209671, + 0.013233166188001633, + 0.011592536233365536, + 0.018224047496914864, + 0.011026078835129738, + 0.008545906282961369, + 0.056028205901384354, + 0.008560175076127052, + 0.005263310391455889, + 0.005819549318403006, + 0.004498286638408899, + 0.005627689883112907, + 0.007276903837919235, + 0.0034558612387627363, + 0.005388237070292234, + 0.004296321887522936, + 0.005130224861204624, + 0.0045013511553406715, + 0.004954364616423845, + 0.00449194572865963, + 0.005935631226748228, + 0.0036104926839470863, + 0.003064447082579136, + 0.003511442570015788, + 0.011587864719331264, + 0.012830984778702259, + 0.015323545783758163, + 0.0069392104633152485, + 0.012090027332305908, + 0.006343238055706024, + 0.009354542940855026, + 0.010072853416204453, + 0.01604505255818367, + 0.0309322327375412, + 0.077817402780056, + 0.10648319125175476, + 0.012136900797486305, + 0.008420052006840706, + 0.010636414401233196, + 0.006070284638553858, + 0.02600906789302826, + 0.030980851501226425, + 0.0227106474339962, + 0.02462194301187992, + 0.017087800428271294, + 0.011378437280654907, + 0.011884721927344799, + 0.016941752284765244, + 0.003959042020142078, + 0.005462178960442543, + 0.004197911825031042, + 0.004573184531182051, + 0.003618387971073389, + 0.004004726652055979, + 0.003267919411882758, + 0.004620639141649008, + 0.005565997213125229, + 0.007272835820913315, + 0.008197400718927383, + 0.010178392753005028, + 0.005640001036226749, + 0.0061610485427081585, + 0.008431636728346348, + 0.003479190170764923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.029800403863191605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, static shot is locked on the car's analog speedometer, the only sharp point of focus in the frame. The speedometer's face is a deep matte black, its numerals and hash marks glowing a vibrant, fiery orange. The polished red needle sweeps unwaveringly and rapidly upward, rapidly climbing past the 100, 120, and 140 mph markers in smooth, relentless motion. In the background, the tunnel environment is rendered as a swirl of intense, natural motion blur; overhead sodium lights and the concrete walls carve a chaotic jumble of warm yellow and dark gray horizontal lines, creating a monumental sense of speed. The flickering, distorted reflections of these light streaks slide across the instrument panel's clear glass cover, the only sign of the world beyond. The stark contrast between the bright, glowing dial and the completely blurred surroundings, combined with the dark, high-contrast color palette, creates an intense and thrilling atmosphere of pure, unadulterated speed.", + "seed": 25306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_9/index_4600_seed_25306_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02690376788377762, + "vlm_logprob": -0.02690376788377762, + "vlm_yes_no_margin": 3.849999952316284, + "vlm_token_credit": -0.02690376788377762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02018318697810173 + ], + "opposite_logprob_tokens": [ + -4.895183086395264 + ], + "logprob_reward": -0.02018318697810173, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.02018318697810173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15737701952457428, + 0.06597457081079483, + 0.06166863813996315, + 0.11973188817501068, + 0.18623225390911102, + 0.06861443817615509, + 0.16674059629440308, + 0.03183232620358467, + 0.1134311631321907, + 0.02839706838130951 + ], + "credit_map": [ + 0.03294489160180092, + 0.03837503120303154, + 0.021919121965765953, + 0.017877839505672455, + 0.010143522173166275, + 0.008673625998198986, + 0.012352893128991127, + 0.015090109780430794, + 0.006435819435864687, + 0.00821601040661335, + 0.007573925890028477, + 0.014599560759961605, + 0.007328879088163376, + 0.005430823192000389, + 0.0066231791861355305, + 0.00976637750864029, + 0.010353383608162403, + 0.014522915706038475, + 0.0068428595550358295, + 0.008527088910341263, + 0.0035862962249666452, + 0.005605859216302633, + 0.00467227678745985, + 0.007557960227131844, + 0.01684408448636532, + 0.017979569733142853, + 0.013062595389783382, + 0.014055808074772358, + 0.006712879985570908, + 0.012966309674084187, + 0.010820182971656322, + 0.0272904671728611, + 0.05533768981695175, + 0.05513278394937515, + 0.012643049471080303, + 0.016560513526201248, + 0.010308263823390007, + 0.010965388268232346, + 0.012089983560144901, + 0.013194580562412739, + 0.012629445642232895, + 0.01619778946042061, + 0.005622232798486948, + 0.009910105727612972, + 0.004417366813868284, + 0.005786275025457144, + 0.005605689249932766, + 0.00844554416835308, + 0.030155982822179794, + 0.017726268619298935, + 0.016729403287172318, + 0.02178887464106083, + 0.013385545462369919, + 0.016391484066843987, + 0.017603985965251923, + 0.03295905143022537, + 0.005835161544382572, + 0.0048417337238788605, + 0.0031479503959417343, + 0.0032321333419531584, + 0.0022344375029206276, + 0.004696073941886425, + 0.004587836097925901, + 0.0032570003531873226, + 0.01760932430624962, + 0.0169925969094038, + 0.018164001405239105, + 0.014704316854476929, + 0.010470901615917683, + 0.00875515304505825, + 0.012276552617549896, + 0.014458322897553444, + 0.003490413073450327, + 0.004339206963777542, + 0.0036982735618948936, + 0.003769195405766368, + 0.002323936205357313, + 0.003335728542879224, + 0.004151917994022369, + 0.003288396866992116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.02018318697810173 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306808114051819 + ], + "opposite_logprob_tokens": [ + -3.1680681705474854 + ], + "logprob_reward": -0.04306808114051819, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306808114051819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0779116153717041, + 0.08076111227273941, + 0.12429695576429367, + 0.1255980283021927, + 0.10951121151447296, + 0.08593548834323883, + 0.060736577957868576, + 0.050462864339351654, + 0.07813030481338501, + 0.20665580034255981 + ], + "credit_map": [ + 0.011494792066514492, + 0.016777772456407547, + 0.007991195656359196, + 0.010547597892582417, + 0.008633648976683617, + 0.005528497043997049, + 0.0070726522244513035, + 0.009865461848676205, + 0.008177532814443111, + 0.014785366132855415, + 0.007924981415271759, + 0.01896779052913189, + 0.00993923470377922, + 0.0049814507365226746, + 0.008369560353457928, + 0.007615204434841871, + 0.018974820151925087, + 0.023976607248187065, + 0.011305847205221653, + 0.013998406007885933, + 0.009462960064411163, + 0.011060979217290878, + 0.014185143634676933, + 0.02133220247924328, + 0.01763305813074112, + 0.02221369929611683, + 0.014930679462850094, + 0.018348706886172295, + 0.00740954652428627, + 0.012820141389966011, + 0.011304021812975407, + 0.020938176661729813, + 0.0366620309650898, + 0.018383685499429703, + 0.010580837726593018, + 0.009446052834391594, + 0.00546859297901392, + 0.007258927915245295, + 0.006837711203843355, + 0.01487337239086628, + 0.021323734894394875, + 0.02309591881930828, + 0.006674509029835463, + 0.01283805351704359, + 0.004432243295013905, + 0.006979663390666246, + 0.004921805579215288, + 0.005669572856277227, + 0.010498651303350925, + 0.008113555610179901, + 0.007337470073252916, + 0.00616793055087328, + 0.006389258895069361, + 0.007466794457286596, + 0.005571882240474224, + 0.009191039949655533, + 0.007906190119683743, + 0.00933982152491808, + 0.0054276264272630215, + 0.010102078318595886, + 0.003551417263224721, + 0.004431428853422403, + 0.0051423329859972, + 0.004561971873044968, + 0.01001021359115839, + 0.007111733313649893, + 0.00790998712182045, + 0.006791634019464254, + 0.008721974678337574, + 0.012201891280710697, + 0.01071474514901638, + 0.014668131247162819, + 0.045701608061790466, + 0.047596100717782974, + 0.019400913268327713, + 0.023890970274806023, + 0.02027658373117447, + 0.014871390536427498, + 0.01623430848121643, + 0.01868392713367939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.04306808114051819 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818765513598919 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.01818765513598919, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818765513598919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11040836572647095, + 0.11247549206018448, + 0.13956035673618317, + 0.14390291273593903, + 0.19945265352725983, + 0.05940506234765053, + 0.0638694316148758, + 0.07483453303575516, + 0.05100957676768303, + 0.04508161544799805 + ], + "credit_map": [ + 0.009936382994055748, + 0.008351743221282959, + 0.013563127256929874, + 0.02918756753206253, + 0.008189868181943893, + 0.011196035891771317, + 0.00898281391710043, + 0.021000834181904793, + 0.01413168478757143, + 0.017815956845879555, + 0.01220716256648302, + 0.029739823192358017, + 0.009388183243572712, + 0.007032071705907583, + 0.011035887524485588, + 0.011124723590910435, + 0.020928608253598213, + 0.027976861223578453, + 0.014249448664486408, + 0.01972479373216629, + 0.010951435193419456, + 0.01100502535700798, + 0.011102922260761261, + 0.023621253669261932, + 0.019381143152713776, + 0.021336590871214867, + 0.015839284285902977, + 0.019039884209632874, + 0.00834103487432003, + 0.015350633300840855, + 0.016108538955450058, + 0.028505800291895866, + 0.07073327898979187, + 0.058198608458042145, + 0.014158429577946663, + 0.01750064454972744, + 0.009229389019310474, + 0.009471799246966839, + 0.008807506412267685, + 0.011352994479238987, + 0.01303081214427948, + 0.01232577208429575, + 0.005279427394270897, + 0.006735648959875107, + 0.0038385679945349693, + 0.005687332712113857, + 0.006741166114807129, + 0.005766336340457201, + 0.011246783658862114, + 0.00726571399718523, + 0.007384114433079958, + 0.007870377041399479, + 0.005723468493670225, + 0.0064551508985459805, + 0.006780215539038181, + 0.011143605224788189, + 0.012500726617872715, + 0.008453764952719212, + 0.010235672816634178, + 0.008353027515113354, + 0.00492339301854372, + 0.007077967748045921, + 0.01634344644844532, + 0.006946539506316185, + 0.006288870237767696, + 0.0047842166386544704, + 0.004633526783436537, + 0.00526662589982152, + 0.0072898948565125465, + 0.0052323429845273495, + 0.008785796351730824, + 0.008728301152586937, + 0.006512457970529795, + 0.0092738326638937, + 0.00505347503349185, + 0.00574859231710434, + 0.004634436219930649, + 0.00337835680693388, + 0.00479522068053484, + 0.005685244686901569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.01818765513598919 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029791725799441338 + ], + "opposite_logprob_tokens": [ + -3.529791831970215 + ], + "logprob_reward": -0.029791725799441338, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029791725799441338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2368154525756836, + 0.1621994525194168, + 0.14296041429042816, + 0.07657627016305923, + 0.21383079886436462, + 0.05505327880382538, + 0.03410254046320915, + 0.013287832029163837, + 0.029960542917251587, + 0.035213373601436615 + ], + "credit_map": [ + 0.04274348169565201, + 0.05955904349684715, + 0.020221682265400887, + 0.026254290714859962, + 0.017304638400673866, + 0.00866722408682108, + 0.018060941249132156, + 0.0440041609108448, + 0.021098557859659195, + 0.026688406243920326, + 0.01544706616550684, + 0.045781124383211136, + 0.01645711064338684, + 0.008316651917994022, + 0.014796656556427479, + 0.013613879680633545, + 0.02607126720249653, + 0.033455830067396164, + 0.014040699228644371, + 0.016830943524837494, + 0.010311363264918327, + 0.010482419282197952, + 0.010989736765623093, + 0.02077815681695938, + 0.011744005605578423, + 0.013606520369648933, + 0.006891853176057339, + 0.009627076797187328, + 0.004216695204377174, + 0.007930121384561062, + 0.005631121341139078, + 0.01692887209355831, + 0.08018989861011505, + 0.06872501969337463, + 0.01312966737896204, + 0.014819982461631298, + 0.00859080534428358, + 0.010127288289368153, + 0.008198619820177555, + 0.010049537755548954, + 0.013168488629162312, + 0.015394811518490314, + 0.004325416870415211, + 0.008979741483926773, + 0.002851681085303426, + 0.0032932055182754993, + 0.003022299613803625, + 0.004017638508230448, + 0.006571839563548565, + 0.004349995870143175, + 0.0038946098648011684, + 0.005478542763739824, + 0.0032169169280678034, + 0.002557623665779829, + 0.003974255174398422, + 0.004058758728206158, + 0.0022080966737121344, + 0.0019496707245707512, + 0.0017966924933716655, + 0.0021078153513371944, + 0.0008659708546474576, + 0.0010443547507748008, + 0.002325968351215124, + 0.0009892627131193876, + 0.003617443609982729, + 0.0031625120900571346, + 0.0027284296229481697, + 0.004257527645677328, + 0.00421005068346858, + 0.003081150818616152, + 0.004741276614367962, + 0.004162152297794819, + 0.007018084637820721, + 0.0056135281920433044, + 0.003637992311269045, + 0.003450487507507205, + 0.003133933525532484, + 0.0036962281446903944, + 0.0027878882829099894, + 0.0058752321638166904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.029791725799441338 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288190364837646 + ], + "opposite_logprob_tokens": [ + -3.7732882499694824 + ], + "logprob_reward": -0.023288190364837646, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288190364837646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12272030860185623, + 0.11028943210840225, + 0.12538856267929077, + 0.2221400886774063, + 0.2261005938053131, + 0.032801516354084015, + 0.0635511726140976, + 0.019389837980270386, + 0.047393765300512314, + 0.030224720016121864 + ], + "credit_map": [ + 0.010570001788437366, + 0.008498267270624638, + 0.014740176498889923, + 0.027339523658156395, + 0.006520660128444433, + 0.009511814452707767, + 0.012850362807512283, + 0.032689500600099564, + 0.014788827858865261, + 0.018446333706378937, + 0.009292630478739738, + 0.03064371645450592, + 0.010372125543653965, + 0.0060496157966554165, + 0.009630057960748672, + 0.011066118255257607, + 0.01971515826880932, + 0.02417444996535778, + 0.012788696214556694, + 0.014919396489858627, + 0.011917443946003914, + 0.011956821195781231, + 0.014315957203507423, + 0.015600642189383507, + 0.02976870909333229, + 0.03651832789182663, + 0.024181317538022995, + 0.0320899672806263, + 0.012802207842469215, + 0.023267751559615135, + 0.023470014333724976, + 0.04004179686307907, + 0.07495081424713135, + 0.06429300457239151, + 0.017061764374375343, + 0.02047213725745678, + 0.011177895590662956, + 0.011923995800316334, + 0.011771128512918949, + 0.014449858106672764, + 0.006111287046223879, + 0.00800881627947092, + 0.0030942370649427176, + 0.004973268136382103, + 0.0019294875673949718, + 0.002605003071948886, + 0.0025777667760849, + 0.0035016504116356373, + 0.012409215793013573, + 0.0069410428404808044, + 0.006853481288999319, + 0.008997802622616291, + 0.005883875768631697, + 0.005268139764666557, + 0.007150758523494005, + 0.010046852752566338, + 0.003490561619400978, + 0.0030049069318920374, + 0.0023500530514866114, + 0.002346395282074809, + 0.001372762257233262, + 0.0022088850382715464, + 0.00304632680490613, + 0.001569947344250977, + 0.0054431795142591, + 0.005437277257442474, + 0.004544478375464678, + 0.006608244962990284, + 0.006179075688123703, + 0.006336571183055639, + 0.007668215315788984, + 0.005176723934710026, + 0.004638197366148233, + 0.004246294964104891, + 0.002888550516217947, + 0.003397481981664896, + 0.0023530072066932917, + 0.002741229021921754, + 0.0028492670971900225, + 0.007110692095011473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.023288190364837646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, static shot is locked on the car's analog speedometer, the only sharp point of focus in the frame. The speedometer's face is a deep matte black, its numerals and hash marks glowing a vibrant, fiery orange. The polished red needle sweeps unwaveringly and rapidly upward, rapidly climbing past the 100, 120, and 140 mph markers in smooth, relentless motion. In the background, the tunnel environment is rendered as a swirl of intense, natural motion blur; overhead sodium lights and the concrete walls carve a chaotic jumble of warm yellow and dark gray horizontal lines, creating a monumental sense of speed. The flickering, distorted reflections of these light streaks slide across the instrument panel's clear glass cover, the only sign of the world beyond. The stark contrast between the bright, glowing dial and the completely blurred surroundings, combined with the dark, high-contrast color palette, creates an intense and thrilling atmosphere of pure, unadulterated speed.", + "seed": 25307, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_9/index_4600_seed_25307_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023258828138932586, + "vlm_logprob": -0.023258828138932586, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.023258828138932586, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07062739133834839 + ], + "opposite_logprob_tokens": [ + -2.820627450942993 + ], + "logprob_reward": -0.07062739133834839, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07062739133834839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31875622272491455, + 0.08710384368896484, + 0.18248499929904938, + 0.03453664854168892, + 0.12629340589046478, + 0.0804307833313942, + 0.04413265734910965, + 0.051238447427749634, + 0.047711167484521866, + 0.02731184847652912 + ], + "credit_map": [ + 0.04119221866130829, + 0.013963144272565842, + 0.06888088583946228, + 0.1291678249835968, + 0.020577067509293556, + 0.014852851629257202, + 0.014835692010819912, + 0.01528654620051384, + 0.010318126529455185, + 0.009991956874728203, + 0.010674572549760342, + 0.010760081000626087, + 0.008517315611243248, + 0.013723211362957954, + 0.011883729137480259, + 0.011234845966100693, + 0.03633178398013115, + 0.03317445144057274, + 0.014151209965348244, + 0.019080307334661484, + 0.014860523864626884, + 0.024466298520565033, + 0.016773685812950134, + 0.023646734654903412, + 0.0033087010961025953, + 0.003561546327546239, + 0.004737003706395626, + 0.004050873685628176, + 0.004648583009839058, + 0.00491297198459506, + 0.004897128324955702, + 0.0044198413379490376, + 0.013170537538826466, + 0.014512193389236927, + 0.017982495948672295, + 0.013468894176185131, + 0.019139911979436874, + 0.012631445191800594, + 0.013061566278338432, + 0.02232636883854866, + 0.011317756026983261, + 0.010355101898312569, + 0.008210795000195503, + 0.012689722701907158, + 0.007138644810765982, + 0.010448206216096878, + 0.008881251327693462, + 0.011389301158487797, + 0.005008738022297621, + 0.005530356429517269, + 0.004602271597832441, + 0.005173502489924431, + 0.005954263731837273, + 0.005986506585031748, + 0.005674290936440229, + 0.006202725227922201, + 0.007325988728553057, + 0.007359712850302458, + 0.007021103519946337, + 0.009621006436645985, + 0.005517608020454645, + 0.0044049848802387714, + 0.004947197623550892, + 0.005040847230702639, + 0.005120537709444761, + 0.00580204650759697, + 0.004218729678541422, + 0.00601247837767005, + 0.005440616048872471, + 0.006734134163707495, + 0.005995674058794975, + 0.008386951871216297, + 0.0035203269217163324, + 0.0034552293363958597, + 0.0036797053180634975, + 0.004639092832803726, + 0.001743043540045619, + 0.003899781499058008, + 0.0033427644520998, + 0.0030319057404994965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.07062739133834839 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647759281098843 + ], + "opposite_logprob_tokens": [ + -4.88264799118042 + ], + "logprob_reward": -0.007647759281098843, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647759281098843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17492496967315674, + 0.11686265468597412, + 0.1214039996266365, + 0.0670241266489029, + 0.049771882593631744, + 0.12261532247066498, + 0.05965755134820938, + 0.14009852707386017, + 0.06425679475069046, + 0.08338408172130585 + ], + "credit_map": [ + 0.029100999236106873, + 0.029646700248122215, + 0.012697761878371239, + 0.028929999098181725, + 0.019600560888648033, + 0.01984522119164467, + 0.0169928390532732, + 0.018110908567905426, + 0.012480907142162323, + 0.01330147311091423, + 0.01991262473165989, + 0.014825498685240746, + 0.010598554275929928, + 0.014839429408311844, + 0.014836257323622704, + 0.016067923977971077, + 0.01437422912567854, + 0.014106448739767075, + 0.009142939001321793, + 0.018300190567970276, + 0.01468183659017086, + 0.020044272765517235, + 0.014032941311597824, + 0.016721157357096672, + 0.006366552785038948, + 0.007975105196237564, + 0.007500964216887951, + 0.006911175791174173, + 0.013110076077282429, + 0.009471812285482883, + 0.008882925845682621, + 0.006805513519793749, + 0.0061045256443321705, + 0.005541049409657717, + 0.005279048811644316, + 0.006538748741149902, + 0.006066326983273029, + 0.005757695529609919, + 0.005962552037090063, + 0.00852193869650364, + 0.013667639344930649, + 0.016422249376773834, + 0.012402845546603203, + 0.017299504950642586, + 0.014440248720347881, + 0.016539154574275017, + 0.01585392653942108, + 0.015989769250154495, + 0.006205454468727112, + 0.007028873544186354, + 0.005790675524622202, + 0.00918006431311369, + 0.008875141851603985, + 0.007715070154517889, + 0.007488238625228405, + 0.0073740375228226185, + 0.02188195288181305, + 0.028893068432807922, + 0.010853047482669353, + 0.01611538603901863, + 0.018491800874471664, + 0.020804770290851593, + 0.012519676238298416, + 0.010538836009800434, + 0.012244584038853645, + 0.010507252998650074, + 0.005776309873908758, + 0.006897471379488707, + 0.0059099020436406136, + 0.008779151365160942, + 0.0055680908262729645, + 0.008574040606617928, + 0.011030778288841248, + 0.011569909751415253, + 0.009030161425471306, + 0.014490518718957901, + 0.005904335994273424, + 0.013297412544488907, + 0.00839453935623169, + 0.009666435420513153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.007647759281098843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287957534193993 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023287957534193993, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287957534193993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47861140966415405, + 0.04626386985182762, + 0.05986995995044708, + 0.034359898418188095, + 0.03822899982333183, + 0.06266017258167267, + 0.0688643604516983, + 0.09274018555879593, + 0.059407830238342285, + 0.05899329483509064 + ], + "credit_map": [ + 0.06064704805612564, + 0.018841486424207687, + 0.10828270763158798, + 0.19566801190376282, + 0.030141564086079597, + 0.019753389060497284, + 0.021205134689807892, + 0.024072103202342987, + 0.004577834624797106, + 0.004435120150446892, + 0.008986528031527996, + 0.007058470509946346, + 0.0038447959814220667, + 0.00674627348780632, + 0.004628241062164307, + 0.005986608564853668, + 0.005130130797624588, + 0.005236126948148012, + 0.0037637243513017893, + 0.00951701495796442, + 0.00824436079710722, + 0.008608072064816952, + 0.010272147133946419, + 0.009098390117287636, + 0.0031561548821628094, + 0.00391410244628787, + 0.004569407552480698, + 0.004633817821741104, + 0.005933356937021017, + 0.004656498786062002, + 0.004192158579826355, + 0.0033044065348803997, + 0.0035067463759332895, + 0.003599590854719281, + 0.0041569909080863, + 0.005867957137525082, + 0.004893419798463583, + 0.005948722828179598, + 0.003997278865426779, + 0.006258298177272081, + 0.01017756201326847, + 0.007034642156213522, + 0.00824909657239914, + 0.008303848095238209, + 0.0057794381864368916, + 0.008648526854813099, + 0.006981933489441872, + 0.007485130801796913, + 0.007307757157832384, + 0.007947209291160107, + 0.008108479902148247, + 0.009006889536976814, + 0.008142205886542797, + 0.008824219927191734, + 0.010032190009951591, + 0.009495417587459087, + 0.024713872000575066, + 0.008013236336410046, + 0.005836694035679102, + 0.026422247290611267, + 0.007901039905846119, + 0.005636337213218212, + 0.007831932976841927, + 0.006384831853210926, + 0.005036693532019854, + 0.0075456853955984116, + 0.004785072989761829, + 0.012739265337586403, + 0.0067048040218651295, + 0.0075546493753790855, + 0.0062754335813224316, + 0.0087662348523736, + 0.0061522480100393295, + 0.007081274408847094, + 0.006602696608752012, + 0.010588765144348145, + 0.007981214672327042, + 0.006856692489236593, + 0.009389885701239109, + 0.004340524785220623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.023287957534193993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021948551293462515 + ], + "opposite_logprob_tokens": [ + -6.127194881439209 + ], + "logprob_reward": -0.0021948551293462515, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021948551293462515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1828451156616211, + 0.10602767765522003, + 0.04969240352511406, + 0.059828344732522964, + 0.14302091300487518, + 0.13418810069561005, + 0.1094340905547142, + 0.123796246945858, + 0.0238750372081995, + 0.06729206442832947 + ], + "credit_map": [ + 0.02498585358262062, + 0.021741991862654686, + 0.037360429763793945, + 0.03601757064461708, + 0.014619899913668633, + 0.017823562026023865, + 0.012929016724228859, + 0.0173668060451746, + 0.012641256675124168, + 0.01236374769359827, + 0.02030321955680847, + 0.012907098047435284, + 0.008758998475968838, + 0.0135046923533082, + 0.013076542876660824, + 0.012472128495573997, + 0.006531071849167347, + 0.005517399869859219, + 0.006461452692747116, + 0.005328064784407616, + 0.006194766145199537, + 0.005359855946153402, + 0.00815832894295454, + 0.006141463294625282, + 0.007284254301339388, + 0.008583741262555122, + 0.009075148031115532, + 0.0065641882829368114, + 0.007803342305123806, + 0.007344500627368689, + 0.0077205197885632515, + 0.005452650599181652, + 0.024787981063127518, + 0.015571553260087967, + 0.017496252432465553, + 0.013507240451872349, + 0.016705680638551712, + 0.021082131192088127, + 0.010866427794098854, + 0.02300364151597023, + 0.017911400645971298, + 0.018292512744665146, + 0.012691721320152283, + 0.0196539293974638, + 0.013466205447912216, + 0.017857208847999573, + 0.01601654291152954, + 0.01829857937991619, + 0.016334902495145798, + 0.015384471043944359, + 0.01320244837552309, + 0.016270315274596214, + 0.011979646980762482, + 0.014771465212106705, + 0.010192550718784332, + 0.011298286728560925, + 0.0158365648239851, + 0.025783546268939972, + 0.015307758934795856, + 0.01928631216287613, + 0.011675187386572361, + 0.016044436022639275, + 0.008681103587150574, + 0.011181333102285862, + 0.002350381575524807, + 0.003154339734464884, + 0.002245465060696006, + 0.003386939410120249, + 0.003252754220739007, + 0.003316159127280116, + 0.0024773168843239546, + 0.0036916823592036963, + 0.010564615949988365, + 0.009584550745785236, + 0.009056518785655499, + 0.010582881048321724, + 0.0059814453125, + 0.009826741181313992, + 0.0073615750297904015, + 0.0043337405659258366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0021948551293462515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536177411675453 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012536177411675453, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536177411675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2660043239593506, + 0.06834429502487183, + 0.1228727474808693, + 0.03960855305194855, + 0.06177308037877083, + 0.06180274859070778, + 0.11086496710777283, + 0.098055399954319, + 0.06366405636072159, + 0.1070098727941513 + ], + "credit_map": [ + 0.08994556963443756, + 0.028455322608351707, + 0.03618606925010681, + 0.03232051059603691, + 0.02253957837820053, + 0.0231368076056242, + 0.012928185984492302, + 0.02049226500093937, + 0.0079164057970047, + 0.00837934110313654, + 0.014356415718793869, + 0.00899034645408392, + 0.005064589902758598, + 0.007378663867712021, + 0.007475367281585932, + 0.008783163502812386, + 0.02175213024020195, + 0.021455664187669754, + 0.009650488384068012, + 0.01742190308868885, + 0.010373112745583057, + 0.01674400269985199, + 0.00925503671169281, + 0.016220401972532272, + 0.0033690212294459343, + 0.003944971598684788, + 0.008083305321633816, + 0.004988929722458124, + 0.004876891616731882, + 0.005457431077957153, + 0.004455992486327887, + 0.004432006739079952, + 0.007162958849221468, + 0.005739493761211634, + 0.006478370167315006, + 0.00718978326767683, + 0.010095808655023575, + 0.008710741065442562, + 0.007359660230576992, + 0.009036260657012463, + 0.010506304912269115, + 0.004147024359554052, + 0.012991699390113354, + 0.007844991981983185, + 0.004650673363357782, + 0.00839927140623331, + 0.00580280227586627, + 0.007459978573024273, + 0.01390200387686491, + 0.013340468518435955, + 0.01215454563498497, + 0.011542510241270065, + 0.012524012476205826, + 0.017029879614710808, + 0.011528599075973034, + 0.018842946738004684, + 0.027426909655332565, + 0.009818200953304768, + 0.007348906248807907, + 0.024339288473129272, + 0.007921913638710976, + 0.008634517900645733, + 0.006757375784218311, + 0.005808272864669561, + 0.004958512727171183, + 0.00788982305675745, + 0.005233416333794594, + 0.009773443453013897, + 0.007544321473687887, + 0.009547930210828781, + 0.0077722081914544106, + 0.010944397188723087, + 0.014379757456481457, + 0.014913790859282017, + 0.010761012323200703, + 0.018923401832580566, + 0.007719754707068205, + 0.018693851307034492, + 0.012627745047211647, + 0.008990554139018059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.012536177411675453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, static shot is locked on the car's analog speedometer, the only sharp point of focus in the frame. The speedometer's face is a deep matte black, its numerals and hash marks glowing a vibrant, fiery orange. The polished red needle sweeps unwaveringly and rapidly upward, rapidly climbing past the 100, 120, and 140 mph markers in smooth, relentless motion. In the background, the tunnel environment is rendered as a swirl of intense, natural motion blur; overhead sodium lights and the concrete walls carve a chaotic jumble of warm yellow and dark gray horizontal lines, creating a monumental sense of speed. The flickering, distorted reflections of these light streaks slide across the instrument panel's clear glass cover, the only sign of the world beyond. The stark contrast between the bright, glowing dial and the completely blurred surroundings, combined with the dark, high-contrast color palette, creates an intense and thrilling atmosphere of pure, unadulterated speed.", + "seed": 25308, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000159/group_9/index_4600_seed_25308_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012333238171413541, + "vlm_logprob": -0.012333238171413541, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.012333238171413541, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009229730814695358 + ], + "opposite_logprob_tokens": [ + -6.00922966003418 + ], + "logprob_reward": -0.009229730814695358, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.009229730814695358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11186335235834122, + 0.18449772894382477, + 0.18239717185497284, + 0.06404919922351837, + 0.030830178409814835, + 0.05919986590743065, + 0.07674572616815567, + 0.09757585823535919, + 0.1051592007279396, + 0.08768171817064285 + ], + "credit_map": [ + 0.009758444502949715, + 0.011726628057658672, + 0.01646869257092476, + 0.016424132511019707, + 0.006830335129052401, + 0.016749566420912743, + 0.009116205386817455, + 0.02478935196995735, + 0.020481839776039124, + 0.01794554851949215, + 0.025149794295430183, + 0.024443743750452995, + 0.02907758764922619, + 0.018904361873865128, + 0.023169245570898056, + 0.025325607508420944, + 0.021282074972987175, + 0.01987130381166935, + 0.01736566051840782, + 0.024030033499002457, + 0.023397020995616913, + 0.024147072806954384, + 0.025042861700057983, + 0.027261139824986458, + 0.006557702552527189, + 0.009405072778463364, + 0.009649790823459625, + 0.006002872250974178, + 0.007913005538284779, + 0.00808962807059288, + 0.007520901970565319, + 0.0089102266356349, + 0.003133058547973633, + 0.0036303335800766945, + 0.002882137196138501, + 0.004375360906124115, + 0.004840988665819168, + 0.004935289733111858, + 0.0030874384101480246, + 0.00394557137042284, + 0.007276794407516718, + 0.008965082466602325, + 0.0048618982546031475, + 0.008507315069437027, + 0.005012607201933861, + 0.007803220301866531, + 0.009196294471621513, + 0.0075766546651721, + 0.006141205318272114, + 0.008306026458740234, + 0.01040731929242611, + 0.010728314518928528, + 0.004948337562382221, + 0.01256844773888588, + 0.010951656848192215, + 0.012694417499005795, + 0.015219773165881634, + 0.01176732312887907, + 0.009122950956225395, + 0.009302826598286629, + 0.014417894184589386, + 0.018879428505897522, + 0.007334242574870586, + 0.01153141912072897, + 0.01668219454586506, + 0.007773090619593859, + 0.017390044406056404, + 0.007528208661824465, + 0.007523110136389732, + 0.011327121406793594, + 0.006951615680009127, + 0.029983816668391228, + 0.00825888104736805, + 0.024593042209744453, + 0.010897508822381496, + 0.009351164102554321, + 0.004019262734800577, + 0.010276766493916512, + 0.008947783149778843, + 0.011337308213114738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.009229730814695358 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032048076391220093 + ], + "opposite_logprob_tokens": [ + -5.753204822540283 + ], + "logprob_reward": -0.0032048076391220093, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032048076391220093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2533408999443054, + 0.22766755521297455, + 0.15123985707759857, + 0.06849081069231033, + 0.06904640793800354, + 0.03142533078789711, + 0.0703958049416542, + 0.03805769234895706, + 0.038444239646196365, + 0.05189137160778046 + ], + "credit_map": [ + 0.03046957217156887, + 0.019718613475561142, + 0.04587630182504654, + 0.02082519605755806, + 0.012564231641590595, + 0.025590630248188972, + 0.016681626439094543, + 0.08161472529172897, + 0.014291709288954735, + 0.021057672798633575, + 0.028925945982336998, + 0.021745407953858376, + 0.03402966260910034, + 0.039143990725278854, + 0.021961510181427002, + 0.04651165381073952, + 0.020243851467967033, + 0.013003910891711712, + 0.01796433888375759, + 0.024693310260772705, + 0.016381075605750084, + 0.010739997029304504, + 0.028336670249700546, + 0.01987670734524727, + 0.00621304614469409, + 0.009267657063901424, + 0.011763895861804485, + 0.008679719641804695, + 0.008335079997777939, + 0.004412087611854076, + 0.009313016198575497, + 0.010506308637559414, + 0.00885225459933281, + 0.009009789675474167, + 0.006357480771839619, + 0.007654250133782625, + 0.010807046666741371, + 0.012486309744417667, + 0.005255395546555519, + 0.008623872883617878, + 0.005038742441684008, + 0.004350758623331785, + 0.003397696651518345, + 0.004860485903918743, + 0.0021269607823342085, + 0.003577273339033127, + 0.003896067850291729, + 0.004177342634648085, + 0.007552423980087042, + 0.007101789582520723, + 0.0109557481482625, + 0.01035543717443943, + 0.005047131329774857, + 0.0103486692532897, + 0.00863019097596407, + 0.010404417291283607, + 0.003794895950704813, + 0.0030312661547213793, + 0.005298221949487925, + 0.0055556041188538074, + 0.004576064646244049, + 0.006604870315641165, + 0.004281960427761078, + 0.004914808552712202, + 0.005351159255951643, + 0.0031460740137845278, + 0.005416274536401033, + 0.004369054920971394, + 0.005238803569227457, + 0.005791031289845705, + 0.004658046178519726, + 0.0044737933203577995, + 0.0054036034271121025, + 0.0035287358332425356, + 0.013002205640077591, + 0.007883723825216293, + 0.0026607403997331858, + 0.004583402071148157, + 0.004881245084106922, + 0.009947714395821095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0032048076391220093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033704210072755814 + ], + "opposite_logprob_tokens": [ + -3.4087042808532715 + ], + "logprob_reward": -0.033704210072755814, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033704210072755814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09128151088953018, + 0.1851312816143036, + 0.13459889590740204, + 0.0792994350194931, + 0.018104512244462967, + 0.19562941789627075, + 0.041523631662130356, + 0.03220778703689575, + 0.048458635807037354, + 0.1737649291753769 + ], + "credit_map": [ + 0.007560059428215027, + 0.010182841680943966, + 0.013352062553167343, + 0.013395524583756924, + 0.007181308697909117, + 0.01033759955316782, + 0.006699005141854286, + 0.022573117166757584, + 0.011579298414289951, + 0.021367410197854042, + 0.020011164247989655, + 0.010947329923510551, + 0.0379662811756134, + 0.03033628687262535, + 0.01701790653169155, + 0.03590560704469681, + 0.018634000793099403, + 0.009897630661725998, + 0.016907162964344025, + 0.019739564508199692, + 0.015127844177186489, + 0.012784807942807674, + 0.023880071938037872, + 0.017627807334065437, + 0.008899489417672157, + 0.009495564736425877, + 0.009806127287447453, + 0.00919350516051054, + 0.009462318383157253, + 0.009443040937185287, + 0.011581159196794033, + 0.011418232694268227, + 0.001953048282302916, + 0.0019347012275829911, + 0.0022383229807019234, + 0.0020043025724589825, + 0.002049925271421671, + 0.003291246946901083, + 0.002041479339823127, + 0.0025914853904396296, + 0.021619725972414017, + 0.03287011757493019, + 0.018571773543953896, + 0.03183263912796974, + 0.014170614071190357, + 0.021076012402772903, + 0.029937125742435455, + 0.025551417842507362, + 0.0035269884392619133, + 0.005709745455533266, + 0.004520387388765812, + 0.006463596597313881, + 0.004023411311209202, + 0.006543127819895744, + 0.0049024117179214954, + 0.005833962466567755, + 0.003345056204125285, + 0.0029327701777219772, + 0.0040844520553946495, + 0.0052309646271169186, + 0.002737888367846608, + 0.00561683950945735, + 0.004322415683418512, + 0.003937401343137026, + 0.005302308592945337, + 0.0036680144257843494, + 0.009925207123160362, + 0.006405309308320284, + 0.0020657102577388287, + 0.004884416703134775, + 0.008353488519787788, + 0.007854180410504341, + 0.017582055181264877, + 0.05261804908514023, + 0.041183244436979294, + 0.009820987470448017, + 0.00875814538449049, + 0.018990609794855118, + 0.007693047635257244, + 0.01711878925561905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.033704210072755814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001334610627964139 + ], + "opposite_logprob_tokens": [ + -6.6263346672058105 + ], + "logprob_reward": -0.001334610627964139, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001334610627964139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14606115221977234, + 0.2548551857471466, + 0.23341605067253113, + 0.06466832756996155, + 0.05641026049852371, + 0.05716316029429436, + 0.0739051103591919, + 0.036150719970464706, + 0.03806071728467941, + 0.0393092967569828 + ], + "credit_map": [ + 0.009151733480393887, + 0.011449512094259262, + 0.0271354541182518, + 0.010414172895252705, + 0.007397233974188566, + 0.018305376172065735, + 0.009709088131785393, + 0.05249858647584915, + 0.029800523072481155, + 0.01851346530020237, + 0.031530216336250305, + 0.012444858439266682, + 0.03638158738613129, + 0.04971631243824959, + 0.018389981240034103, + 0.05807822197675705, + 0.0336458683013916, + 0.02552099898457527, + 0.02118827775120735, + 0.02606526017189026, + 0.03981681913137436, + 0.026108313351869583, + 0.02887273021042347, + 0.03219779208302498, + 0.006199706811457872, + 0.006533026695251465, + 0.011885161511600018, + 0.008845284581184387, + 0.004498979076743126, + 0.007828877307474613, + 0.009169424884021282, + 0.00970787089318037, + 0.007861598394811153, + 0.006743425969034433, + 0.003673126455396414, + 0.003935675136744976, + 0.010302335023880005, + 0.012966067530214787, + 0.004249445162713528, + 0.006678585894405842, + 0.006139929406344891, + 0.008992367424070835, + 0.004266457632184029, + 0.007606666535139084, + 0.004639104939997196, + 0.00886874832212925, + 0.008562056347727776, + 0.008087826892733574, + 0.007985131815075874, + 0.007196287624537945, + 0.011205315589904785, + 0.007725792936980724, + 0.006219025701284409, + 0.012439490295946598, + 0.008062552660703659, + 0.013071514666080475, + 0.00475185364484787, + 0.0030588179361075163, + 0.005125550553202629, + 0.004831835627555847, + 0.003930084407329559, + 0.006360682658851147, + 0.003829872002825141, + 0.004262025933712721, + 0.0053656515665352345, + 0.004865933209657669, + 0.005119119770824909, + 0.0030845364090055227, + 0.004130999557673931, + 0.006166855338960886, + 0.002758883638307452, + 0.006568734999746084, + 0.003638853784650564, + 0.0026511780451983213, + 0.012141423299908638, + 0.003659092588350177, + 0.0021201581694185734, + 0.0028342760633677244, + 0.003762325970456004, + 0.008501987904310226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.001334610627964139 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extremely close, static view focused on a car's analog speedometer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the speedometer face deep matte black with numerals and hash marks glowing vibrant fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a polished red needle sweep rapidly upward past the 100, 120, and 140 mph markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a tunnel environment rendered as a swirl of intense motion blur with warm yellow and dark gray horizontal lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are flickering, distorted reflections of light streaks visible sliding across the instrument panel's clear glass cover?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192831702530384 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192831702530384, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192831702530384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12202399969100952, + 0.14419184625148773, + 0.321848601102829, + 0.07041099667549133, + 0.06068621575832367, + 0.027595896273851395, + 0.08537928760051727, + 0.08101938664913177, + 0.03962869569659233, + 0.047215115278959274 + ], + "credit_map": [ + 0.009758242405951023, + 0.008781746961176395, + 0.024745026603341103, + 0.008378639817237854, + 0.00909313652664423, + 0.012312005273997784, + 0.006822820752859116, + 0.04213237017393112, + 0.025900721549987793, + 0.013214030303061008, + 0.0160886999219656, + 0.010982107371091843, + 0.014730295166373253, + 0.023608943447470665, + 0.011462015099823475, + 0.028205033391714096, + 0.04017642140388489, + 0.026881391182541847, + 0.03757813572883606, + 0.04850231483578682, + 0.033987969160079956, + 0.04133803769946098, + 0.04500454291701317, + 0.048379816114902496, + 0.007512228563427925, + 0.01070558000355959, + 0.00723282340914011, + 0.006258358247578144, + 0.013281695544719696, + 0.007704521995037794, + 0.008148696273565292, + 0.00956709310412407, + 0.006628165487200022, + 0.008121712133288383, + 0.005313476547598839, + 0.004583735018968582, + 0.011341439560055733, + 0.011712457053363323, + 0.004058710765093565, + 0.008926521986722946, + 0.0031678760424256325, + 0.004301639273762703, + 0.0027637090533971786, + 0.003769386326894164, + 0.0029562702402472496, + 0.003775930730625987, + 0.003067239886149764, + 0.003793844487518072, + 0.010576385073363781, + 0.013724706135690212, + 0.01180645078420639, + 0.008465991355478764, + 0.008141515776515007, + 0.013173874467611313, + 0.009771778248250484, + 0.00971858948469162, + 0.01179770939052105, + 0.008383483625948429, + 0.009952806867659092, + 0.007644607685506344, + 0.012389441952109337, + 0.011900330893695354, + 0.008805172517895699, + 0.010145829990506172, + 0.007040579803287983, + 0.004574361257255077, + 0.004967115353792906, + 0.0027858749963343143, + 0.004108058754354715, + 0.005982663482427597, + 0.003689937060698867, + 0.006480103824287653, + 0.0046524303033947945, + 0.005535883828997612, + 0.011417782865464687, + 0.005942400544881821, + 0.0033165474887937307, + 0.0050339363515377045, + 0.00424655107781291, + 0.007069586310535669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014192831702530384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_0/index_tensor([8624, 8624])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_0/index_tensor([8624, 8624])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e2c30a58961f2e6d12cb2ab0a5e0d6a0fb53dd4b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_0/index_tensor([8624, 8624])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a powerful, low-angle, medium shot, the entire scene is rendered in the style of a painterly work, with the paint applied thickly (impasto) to create distinct textures and strokes. At the center of the frame, a muscular woodcutter, chopping with all his might, sports a thick brown beard, a weathered face, and a focused expression strained by the intensity of the effort. He wears a frayed, dark red flannel shirt, sleeves rolled up to the elbows, revealing forearms muscular from constant labor. He grips a heavy chopping axe, its handle smooth from years of use, raised above his head. The blade gleams a cold metallic sheen in the dappled sunlight filtering through the forest floor, poised to bring down with a thunderous force. Before him lies a massive tree trunk, its bark rough and cracked, a deep cut revealing the yellowed, fresh wood. At the peak of his power accumulation, countless wood chips of various sizes were scattered by the previous attack, solidifying into pieces of golden fragments illuminated by the light in the air, forming a sharp contrast with the deep forest green and earthy brown in the background. The whole scene, with the strong interweaving of light and shadow, perfectly captured this moment full of primitive power and dynamic beauty.", + "seed": 17155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_0/index_8624_seed_17155_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03454670556820929, + "vlm_logprob": -0.03454670556820929, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.03454670556820929, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14447304606437683 + ], + "opposite_logprob_tokens": [ + -2.019473075866699 + ], + "logprob_reward": -0.14447304606437683, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14447304606437683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18794474005699158, + 0.08006161451339722, + 0.11058869957923889, + 0.05879886448383331, + 0.11087364703416824, + 0.10363883525133133, + 0.09692732244729996, + 0.04315653070807457, + 0.10088939219713211, + 0.1071203276515007 + ], + "credit_map": [ + 0.026612762361764908, + 0.03171464055776596, + 0.019606485962867737, + 0.018159043043851852, + 0.022613121196627617, + 0.027652455493807793, + 0.016437385231256485, + 0.02514885924756527, + 0.018932482227683067, + 0.0051329247653484344, + 0.0055576809681952, + 0.0035071708261966705, + 0.004862257279455662, + 0.011357584968209267, + 0.0053467415273189545, + 0.025364773347973824, + 0.01581847108900547, + 0.010401329025626183, + 0.006080854218453169, + 0.005881306249648333, + 0.030214475467801094, + 0.02052323706448078, + 0.008314431644976139, + 0.013354606926441193, + 0.007811196148395538, + 0.0023083093110471964, + 0.0033625157084316015, + 0.0037566826213151217, + 0.018209591507911682, + 0.013409762643277645, + 0.005235754419118166, + 0.0047050477005541325, + 0.030558310449123383, + 0.009377961046993732, + 0.008004146628081799, + 0.006326624192297459, + 0.006572804879397154, + 0.014966452494263649, + 0.007861568592488766, + 0.027205782011151314, + 0.013172156177461147, + 0.015828009694814682, + 0.007240547798573971, + 0.011418343521654606, + 0.01090172864496708, + 0.014483184553682804, + 0.014566127210855484, + 0.016028743237257004, + 0.013174965977668762, + 0.021055154502391815, + 0.00907730683684349, + 0.007433667778968811, + 0.007295665331184864, + 0.017448296770453453, + 0.009055376052856445, + 0.012386880815029144, + 0.007057413924485445, + 0.0059456913731992245, + 0.005204793531447649, + 0.004320517648011446, + 0.005882286466658115, + 0.004645786713808775, + 0.00413493299856782, + 0.00596510898321867, + 0.01302288193255663, + 0.00993595365434885, + 0.011146094650030136, + 0.005312163848429918, + 0.009326159954071045, + 0.03429277986288071, + 0.005953556392341852, + 0.011899799108505249, + 0.011796317994594574, + 0.010597940534353256, + 0.00860445387661457, + 0.0067440601997077465, + 0.02277602069079876, + 0.02039511688053608, + 0.01228253822773695, + 0.0139238853007555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.14447304606437683 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036207374650985003 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036207374650985003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036207374650985003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11306630820035934, + 0.17545472085475922, + 0.10371477901935577, + 0.125505730509758, + 0.08170188963413239, + 0.10656257718801498, + 0.08720111101865768, + 0.09853752702474594, + 0.040625181049108505, + 0.06763017177581787 + ], + "credit_map": [ + 0.010105897672474384, + 0.009769562631845474, + 0.008424962870776653, + 0.009624461643397808, + 0.01617475040256977, + 0.023882396519184113, + 0.01872643642127514, + 0.016357839107513428, + 0.029335301369428635, + 0.051679469645023346, + 0.013953056186437607, + 0.011624989099800587, + 0.012166924774646759, + 0.020709434524178505, + 0.01117035560309887, + 0.02481517195701599, + 0.009373413398861885, + 0.006226067431271076, + 0.004625899251550436, + 0.004018011502921581, + 0.03851606696844101, + 0.017220905050635338, + 0.007042905781418085, + 0.016691509634256363, + 0.016305522993206978, + 0.004046596121042967, + 0.004615848883986473, + 0.005896006710827351, + 0.0357641875743866, + 0.04600435495376587, + 0.005734697915613651, + 0.007138515822589397, + 0.01619741879403591, + 0.006685835774987936, + 0.004812981002032757, + 0.004913866985589266, + 0.005462419707328081, + 0.025210758671164513, + 0.006758107338100672, + 0.011660503223538399, + 0.011862949468195438, + 0.018906671553850174, + 0.007542339153587818, + 0.010286970995366573, + 0.015957478433847427, + 0.01241408009082079, + 0.016769088804721832, + 0.012822989374399185, + 0.008070179261267185, + 0.005245279520750046, + 0.006018500775098801, + 0.00833776593208313, + 0.00972640048712492, + 0.024370284751057625, + 0.009306800551712513, + 0.01612590253353119, + 0.01208045519888401, + 0.010044612921774387, + 0.008782869204878807, + 0.0087362639605999, + 0.01623779907822609, + 0.012520436197519302, + 0.012002384290099144, + 0.018132705241441727, + 0.0051973192021250725, + 0.002791693899780512, + 0.004519133362919092, + 0.0020026888232678175, + 0.004785205237567425, + 0.01426564622670412, + 0.0023082843981683254, + 0.0047552092000842094, + 0.014797172509133816, + 0.006950248498469591, + 0.009310402907431126, + 0.005105292424559593, + 0.0063629150390625, + 0.008029920049011707, + 0.006868799217045307, + 0.010205422528088093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0036207374650985003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625281810760498 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625281810760498, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625281810760498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15712560713291168, + 0.05078842490911484, + 0.15567457675933838, + 0.0900079533457756, + 0.10392957925796509, + 0.19782471656799316, + 0.09666690230369568, + 0.047074247151613235, + 0.06360059976577759, + 0.03730744123458862 + ], + "credit_map": [ + 0.016696445643901825, + 0.01656992919743061, + 0.012333840131759644, + 0.010379252955317497, + 0.020326264202594757, + 0.045992445200681686, + 0.015048062428832054, + 0.019779354333877563, + 0.011063871905207634, + 0.007977982982993126, + 0.004456365015357733, + 0.0035177506506443024, + 0.003755364567041397, + 0.007309455890208483, + 0.004314333666115999, + 0.008393296040594578, + 0.019111791625618935, + 0.006727531086653471, + 0.007101504597812891, + 0.0055618891492486, + 0.035167939960956573, + 0.026544591411948204, + 0.027629464864730835, + 0.02782985381782055, + 0.012087665498256683, + 0.003958386369049549, + 0.0037039946764707565, + 0.0050498116761446, + 0.023981161415576935, + 0.031034689396619797, + 0.004548976663500071, + 0.0056432574056088924, + 0.01483202911913395, + 0.00696076825261116, + 0.00812135636806488, + 0.00852065160870552, + 0.013105537742376328, + 0.021023597568273544, + 0.011574815027415752, + 0.019790813326835632, + 0.02251720055937767, + 0.01457893941551447, + 0.008548267185688019, + 0.009877071715891361, + 0.024099169299006462, + 0.06228480488061905, + 0.014732830226421356, + 0.041186414659023285, + 0.009695245884358883, + 0.0068861921317875385, + 0.008105721324682236, + 0.01903527043759823, + 0.00683205621317029, + 0.014429707080125809, + 0.018126534298062325, + 0.0135561628267169, + 0.005267415661364794, + 0.0047135124914348125, + 0.005298363044857979, + 0.004211296793073416, + 0.00701006269082427, + 0.00608291057869792, + 0.004652518779039383, + 0.009838168509304523, + 0.008899911306798458, + 0.0050302837044000626, + 0.006918513216078281, + 0.0033005510922521353, + 0.006131516769528389, + 0.023908201605081558, + 0.002995360642671585, + 0.006416259799152613, + 0.007261918392032385, + 0.004284217022359371, + 0.003918126691132784, + 0.0025949261616915464, + 0.0034212993923574686, + 0.00875918846577406, + 0.002549625001847744, + 0.004518134519457817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.007625281810760498 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059457337483763695 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.0059457337483763695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059457337483763695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10823051631450653, + 0.10950964689254761, + 0.09071878343820572, + 0.12022698670625687, + 0.0548781119287014, + 0.13326704502105713, + 0.13801990449428558, + 0.10307487100362778, + 0.035116467624902725, + 0.10695762187242508 + ], + "credit_map": [ + 0.011821913532912731, + 0.011704778298735619, + 0.010161242447793484, + 0.009909210726618767, + 0.012943439185619354, + 0.018535854294896126, + 0.015701990574598312, + 0.017452090978622437, + 0.022789087146520615, + 0.03696359321475029, + 0.006084863096475601, + 0.005535971373319626, + 0.006158494856208563, + 0.01372271403670311, + 0.006462082266807556, + 0.011792839504778385, + 0.012149094603955746, + 0.011406134814023972, + 0.00556849455460906, + 0.007024559658020735, + 0.01109415851533413, + 0.01993929222226143, + 0.01311218086630106, + 0.010424874722957611, + 0.016952034085989, + 0.004011018201708794, + 0.004542191047221422, + 0.00548521988093853, + 0.036497294902801514, + 0.038723092526197433, + 0.007502986118197441, + 0.0065131462179124355, + 0.011761796660721302, + 0.005786017049103975, + 0.005985605996102095, + 0.005674275103956461, + 0.0041181412525475025, + 0.007967428304255009, + 0.0040539889596402645, + 0.00953085720539093, + 0.01589912921190262, + 0.015784094110131264, + 0.009083027951419353, + 0.012463027611374855, + 0.017371591180562973, + 0.025533020496368408, + 0.019693685695528984, + 0.017439469695091248, + 0.018490610644221306, + 0.0413844920694828, + 0.012099380604922771, + 0.011748136021196842, + 0.011127891018986702, + 0.017006641253829002, + 0.012768374755978584, + 0.013394372537732124, + 0.012016880325973034, + 0.01416455116122961, + 0.009169274941086769, + 0.009847206063568592, + 0.01427542231976986, + 0.016093993559479713, + 0.011142553761601448, + 0.016364997252821922, + 0.005302153527736664, + 0.0028091648127883673, + 0.0030957891140133142, + 0.0023835853207856417, + 0.00432974100112915, + 0.008957553654909134, + 0.002371383598074317, + 0.005867100320756435, + 0.01336007658392191, + 0.009940669871866703, + 0.008472991175949574, + 0.005768773145973682, + 0.018808618187904358, + 0.027593960985541344, + 0.009328925982117653, + 0.013683597557246685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.0059457337483763695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068728752434254 + ], + "opposite_logprob_tokens": [ + -4.511068820953369 + ], + "logprob_reward": -0.011068728752434254, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068728752434254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17545506358146667, + 0.11875274032354355, + 0.15171314775943756, + 0.07006925344467163, + 0.05897703394293785, + 0.09196698665618896, + 0.07543904334306717, + 0.03882637992501259, + 0.17360171675682068, + 0.045198600739240646 + ], + "credit_map": [ + 0.021431269124150276, + 0.016023140400648117, + 0.013997240923345089, + 0.016703473404049873, + 0.03204537555575371, + 0.03121037594974041, + 0.017639124765992165, + 0.026405051350593567, + 0.02677087113261223, + 0.03801797330379486, + 0.006879698950797319, + 0.006294489838182926, + 0.006606195587664843, + 0.0133773572742939, + 0.010343515314161777, + 0.010462632402777672, + 0.030037466436624527, + 0.018131261691451073, + 0.01033846940845251, + 0.010903929360210896, + 0.016934270039200783, + 0.02415366843342781, + 0.028073713183403015, + 0.013140350580215454, + 0.009586870670318604, + 0.002779041649773717, + 0.0034641597885638475, + 0.0038915411569178104, + 0.019668221473693848, + 0.021878192201256752, + 0.0035842135548591614, + 0.0052170101553201675, + 0.011489042080938816, + 0.007718996144831181, + 0.006249716039747, + 0.008422160521149635, + 0.003318938659504056, + 0.010277003049850464, + 0.0045029460452497005, + 0.0069982316344976425, + 0.01351092103868723, + 0.007264213636517525, + 0.005494126118719578, + 0.006959906313568354, + 0.010075315833091736, + 0.025595124810934067, + 0.008871187455952168, + 0.014196187257766724, + 0.010407799854874611, + 0.020869536325335503, + 0.0068378387950360775, + 0.006113860290497541, + 0.008149207569658756, + 0.0073760515078902245, + 0.007007301319390535, + 0.008677440695464611, + 0.00500426534563303, + 0.003896709531545639, + 0.00493089621886611, + 0.0037145179230719805, + 0.0047028507106006145, + 0.004546343348920345, + 0.006097983103245497, + 0.005932812578976154, + 0.024249862879514694, + 0.011345372535288334, + 0.014278775081038475, + 0.008080166764557362, + 0.016290396451950073, + 0.07212618738412857, + 0.007094423286616802, + 0.020136527717113495, + 0.00913824699819088, + 0.003930795472115278, + 0.004152744077146053, + 0.0033719942439347506, + 0.007117980159819126, + 0.005943126045167446, + 0.0056650093756616116, + 0.005878702737390995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.011068728752434254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a powerful, low-angle, medium shot, the entire scene is rendered in the style of a painterly work, with the paint applied thickly (impasto) to create distinct textures and strokes. At the center of the frame, a muscular woodcutter, chopping with all his might, sports a thick brown beard, a weathered face, and a focused expression strained by the intensity of the effort. He wears a frayed, dark red flannel shirt, sleeves rolled up to the elbows, revealing forearms muscular from constant labor. He grips a heavy chopping axe, its handle smooth from years of use, raised above his head. The blade gleams a cold metallic sheen in the dappled sunlight filtering through the forest floor, poised to bring down with a thunderous force. Before him lies a massive tree trunk, its bark rough and cracked, a deep cut revealing the yellowed, fresh wood. At the peak of his power accumulation, countless wood chips of various sizes were scattered by the previous attack, solidifying into pieces of golden fragments illuminated by the light in the air, forming a sharp contrast with the deep forest green and earthy brown in the background. The whole scene, with the strong interweaving of light and shadow, perfectly captured this moment full of primitive power and dynamic beauty.", + "seed": 17156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_0/index_8624_seed_17156_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02581646330654621, + "vlm_logprob": -0.02581646330654621, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02581646330654621, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1019911915063858 + ], + "opposite_logprob_tokens": [ + -2.3519911766052246 + ], + "logprob_reward": -0.1019911915063858, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1019911915063858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21836836636066437, + 0.07656198740005493, + 0.08948975801467896, + 0.12324921041727066, + 0.05198913812637329, + 0.09749799966812134, + 0.12520597875118256, + 0.13337945938110352, + 0.04955941066145897, + 0.03469865024089813 + ], + "credit_map": [ + 0.024128137156367302, + 0.025172417983412743, + 0.034762170165777206, + 0.019428091123700142, + 0.02106807753443718, + 0.037305232137441635, + 0.020913541316986084, + 0.03559069707989693, + 0.011805856600403786, + 0.01063995435833931, + 0.01246942300349474, + 0.008978093042969704, + 0.007297556847333908, + 0.009148103184998035, + 0.00834497157484293, + 0.007878025993704796, + 0.009090044535696507, + 0.008471885696053505, + 0.012214668095111847, + 0.016288306564092636, + 0.00923082884401083, + 0.009523427113890648, + 0.015542581677436829, + 0.009128010831773281, + 0.016732219606637955, + 0.0094775864854455, + 0.01143396832048893, + 0.008828219957649708, + 0.01519258227199316, + 0.021831149235367775, + 0.015187731944024563, + 0.024565747007727623, + 0.0073655713349580765, + 0.005147886462509632, + 0.006603946443647146, + 0.006407969631254673, + 0.0037870800588279963, + 0.008388929069042206, + 0.005190365016460419, + 0.009097394533455372, + 0.013178079389035702, + 0.009505272842943668, + 0.01059053186327219, + 0.010612900368869305, + 0.013670380227267742, + 0.013027068227529526, + 0.01361885853111744, + 0.013294906355440617, + 0.02198265679180622, + 0.009047087281942368, + 0.01871788688004017, + 0.010695002973079681, + 0.01298935804516077, + 0.020478207617998123, + 0.013257695361971855, + 0.018038088455796242, + 0.015785424038767815, + 0.012779020704329014, + 0.013342216610908508, + 0.011329149827361107, + 0.024091143161058426, + 0.016399815678596497, + 0.016476495191454887, + 0.02317618951201439, + 0.004750765394419432, + 0.0046469648368656635, + 0.006084822583943605, + 0.004067998379468918, + 0.006075487472116947, + 0.006967627443373203, + 0.007240996230393648, + 0.009724748320877552, + 0.004997628275305033, + 0.005652783904224634, + 0.003548927139490843, + 0.003054215805605054, + 0.004468950908631086, + 0.003919190727174282, + 0.0035423613153398037, + 0.0055145942606031895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.1019911915063858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100012592971325 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004100012592971325, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100012592971325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.154489666223526, + 0.16919390857219696, + 0.04839487001299858, + 0.1001623272895813, + 0.13002537190914154, + 0.06054264307022095, + 0.08958697319030762, + 0.10088670998811722, + 0.08116819709539413, + 0.065549336373806 + ], + "credit_map": [ + 0.02335362322628498, + 0.02295610122382641, + 0.011580952443182468, + 0.011017882265150547, + 0.022121436893939972, + 0.021829994395375252, + 0.020018571987748146, + 0.021611103788018227, + 0.020340776070952415, + 0.017857076600193977, + 0.024217231199145317, + 0.0199842918664217, + 0.02350238338112831, + 0.021459445357322693, + 0.01838643290102482, + 0.023446260020136833, + 0.006192388478666544, + 0.005530251655727625, + 0.00823427364230156, + 0.003600558266043663, + 0.005469297990202904, + 0.006535458844155073, + 0.0069481381215155125, + 0.0058845048770308495, + 0.010436944663524628, + 0.010274938307702541, + 0.008626527152955532, + 0.007813329808413982, + 0.014328188262879848, + 0.014930473640561104, + 0.010898829437792301, + 0.022853096947073936, + 0.04040466621518135, + 0.008238539099693298, + 0.006878465414047241, + 0.003714188002049923, + 0.0049925632774829865, + 0.047270722687244415, + 0.004753184970468283, + 0.013773047365248203, + 0.007854168303310871, + 0.005822350271046162, + 0.004949419759213924, + 0.004070551134645939, + 0.010204661637544632, + 0.00873857457190752, + 0.007530673872679472, + 0.011372245848178864, + 0.01380185317248106, + 0.009596528485417366, + 0.013445014134049416, + 0.009392790496349335, + 0.0059116599150002, + 0.019607961177825928, + 0.010425876826047897, + 0.007405294105410576, + 0.010923637077212334, + 0.010831760242581367, + 0.013047318905591965, + 0.008991007693111897, + 0.01661805994808674, + 0.01202685758471489, + 0.012211148627102375, + 0.016236918047070503, + 0.009616723284125328, + 0.007230227347463369, + 0.0077649434097111225, + 0.007314447313547134, + 0.011423634365200996, + 0.012990961782634258, + 0.010942007414996624, + 0.013885250315070152, + 0.01099152397364378, + 0.006877661682665348, + 0.005776120815426111, + 0.005641339812427759, + 0.008127851411700249, + 0.007745199836790562, + 0.008630778640508652, + 0.011758861131966114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.004100012592971325 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067326370626688 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.0067326370626688, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067326370626688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15268824994564056, + 0.14003132283687592, + 0.09074612706899643, + 0.2616426646709442, + 0.08430897444486618, + 0.04960745573043823, + 0.03174673020839691, + 0.047774869948625565, + 0.09185712039470673, + 0.04959647357463837 + ], + "credit_map": [ + 0.024005446583032608, + 0.013095049187541008, + 0.017326595261693, + 0.014613686129450798, + 0.013727780431509018, + 0.025059899315238, + 0.020251626148819923, + 0.024608170613646507, + 0.016485748812556267, + 0.01825047843158245, + 0.018698634579777718, + 0.013282397761940956, + 0.01999904029071331, + 0.0178259015083313, + 0.014501594938337803, + 0.020987529307603836, + 0.010386815294623375, + 0.008355457335710526, + 0.01065954938530922, + 0.014730149880051613, + 0.009813941083848476, + 0.011903195641934872, + 0.014305946417152882, + 0.010591063648462296, + 0.03402038291096687, + 0.020327864214777946, + 0.023161930963397026, + 0.016739822924137115, + 0.03175213187932968, + 0.04486691579222679, + 0.0319206528365612, + 0.05885295942425728, + 0.021550148725509644, + 0.007565703708678484, + 0.006759559269994497, + 0.005360196344554424, + 0.006278265733271837, + 0.021207811310887337, + 0.005159364081919193, + 0.010427924804389477, + 0.007565568201243877, + 0.004473450127989054, + 0.005207537207752466, + 0.003038619412109256, + 0.008618478663265705, + 0.00696924701333046, + 0.004647215828299522, + 0.009087340906262398, + 0.0038782774936407804, + 0.0030004126019775867, + 0.0032700547017157078, + 0.0025607473216950893, + 0.0037033394910395145, + 0.004169333726167679, + 0.004385665524750948, + 0.006778901442885399, + 0.007072193082422018, + 0.004791701678186655, + 0.00516930827870965, + 0.0030598826706409454, + 0.008231990039348602, + 0.006099453661590815, + 0.003950436133891344, + 0.009399903006851673, + 0.008180858567357063, + 0.0057999566197395325, + 0.005683532916009426, + 0.005804682616144419, + 0.014534101821482182, + 0.015240013599395752, + 0.010263069532811642, + 0.026350902393460274, + 0.008187214843928814, + 0.006701156962662935, + 0.004676221404224634, + 0.0050895665772259235, + 0.005855188705027103, + 0.005178410094231367, + 0.0040496960282325745, + 0.009859014302492142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0067326370626688 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762398075312376 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.00762398075312376, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762398075312376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1447267085313797, + 0.11617633700370789, + 0.07217347621917725, + 0.15086127817630768, + 0.15714915096759796, + 0.12267599254846573, + 0.04577087610960007, + 0.051203761249780655, + 0.08415830135345459, + 0.0551040880382061 + ], + "credit_map": [ + 0.022279908880591393, + 0.015441935509443283, + 0.019998937845230103, + 0.012513551861047745, + 0.01868329383432865, + 0.021728478372097015, + 0.01650405488908291, + 0.017576556652784348, + 0.015272748656570911, + 0.01509807538241148, + 0.016123760491609573, + 0.013256300240755081, + 0.01683535799384117, + 0.014671359211206436, + 0.011119241826236248, + 0.013799495063722134, + 0.00947936624288559, + 0.00859126728028059, + 0.011144934222102165, + 0.004490224178880453, + 0.008396410383284092, + 0.009768590331077576, + 0.009104575030505657, + 0.011198112741112709, + 0.01475514005869627, + 0.027116673067212105, + 0.012922326102852821, + 0.019600635394454002, + 0.019587751477956772, + 0.023373136296868324, + 0.014350194483995438, + 0.019155427813529968, + 0.045473162084817886, + 0.008406207896769047, + 0.0092543950304389, + 0.005653160158544779, + 0.007123905699700117, + 0.05860530957579613, + 0.006707889027893543, + 0.015925120562314987, + 0.013362741097807884, + 0.015691464766860008, + 0.00809593964368105, + 0.010279983282089233, + 0.020115848630666733, + 0.01671738736331463, + 0.016811804845929146, + 0.021600820124149323, + 0.006109262816607952, + 0.005857965908944607, + 0.006221783347427845, + 0.004318473394960165, + 0.004542321898043156, + 0.008143401704728603, + 0.005060745403170586, + 0.005516922101378441, + 0.0060635400004684925, + 0.005283957812935114, + 0.006221224553883076, + 0.0038777184672653675, + 0.009370685555040836, + 0.006041001062840223, + 0.005722207948565483, + 0.008623424917459488, + 0.00860567670315504, + 0.0068405368365347385, + 0.007727406919002533, + 0.010612319223582745, + 0.009220650419592857, + 0.014095338992774487, + 0.01582718826830387, + 0.011229189112782478, + 0.009268159046769142, + 0.007751005236059427, + 0.006320823449641466, + 0.008220498450100422, + 0.0056954617612063885, + 0.004954915959388018, + 0.007849828340113163, + 0.005043394863605499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.00762398075312376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634494617581367 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634494617581367, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634494617581367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23530346155166626, + 0.17058469355106354, + 0.09701226651668549, + 0.1110399067401886, + 0.0839260146021843, + 0.0837147980928421, + 0.05706709250807762, + 0.06674397736787796, + 0.038593217730522156, + 0.05601451173424721 + ], + "credit_map": [ + 0.029390661045908928, + 0.026347991079092026, + 0.04762353003025055, + 0.03343745321035385, + 0.027282891795039177, + 0.02134658768773079, + 0.023685695603489876, + 0.02618865855038166, + 0.026107022538781166, + 0.018292581662535667, + 0.027558378875255585, + 0.020525794476270676, + 0.019843949005007744, + 0.022450894117355347, + 0.017833132296800613, + 0.01797294057905674, + 0.013447187840938568, + 0.01019345037639141, + 0.012737937271595001, + 0.014932909980416298, + 0.009523400105535984, + 0.014267059043049812, + 0.009765098802745342, + 0.012145224027335644, + 0.010628610849380493, + 0.020109673961997032, + 0.012486112304031849, + 0.012785768136382103, + 0.011111106723546982, + 0.015513228252530098, + 0.01245134137570858, + 0.015954066067934036, + 0.01799697056412697, + 0.007040385622531176, + 0.007424229755997658, + 0.006173632573336363, + 0.005068100523203611, + 0.022482283413410187, + 0.005223770160228014, + 0.012516639195382595, + 0.012153887189924717, + 0.012951967306435108, + 0.010477161966264248, + 0.00957581028342247, + 0.009955408982932568, + 0.010470093227922916, + 0.009392119944095612, + 0.008738353848457336, + 0.007188877090811729, + 0.008057436905801296, + 0.007807085756212473, + 0.00667492114007473, + 0.005470811855047941, + 0.008086959831416607, + 0.00684300996363163, + 0.006937989965081215, + 0.009987804107367992, + 0.006523354910314083, + 0.012300794012844563, + 0.006552266422659159, + 0.008483185432851315, + 0.007586957421153784, + 0.008166560903191566, + 0.007143056485801935, + 0.0035652522929012775, + 0.0042219869792461395, + 0.003214300610125065, + 0.00413079047575593, + 0.007262419443577528, + 0.00654561584815383, + 0.004874874372035265, + 0.0047779749147593975, + 0.00971094612032175, + 0.007406154647469521, + 0.008351883850991726, + 0.00829511322081089, + 0.0044297934509813786, + 0.005140564404428005, + 0.007845927961170673, + 0.0048341285437345505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008634494617581367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a powerful, low-angle, medium shot, the entire scene is rendered in the style of a painterly work, with the paint applied thickly (impasto) to create distinct textures and strokes. At the center of the frame, a muscular woodcutter, chopping with all his might, sports a thick brown beard, a weathered face, and a focused expression strained by the intensity of the effort. He wears a frayed, dark red flannel shirt, sleeves rolled up to the elbows, revealing forearms muscular from constant labor. He grips a heavy chopping axe, its handle smooth from years of use, raised above his head. The blade gleams a cold metallic sheen in the dappled sunlight filtering through the forest floor, poised to bring down with a thunderous force. Before him lies a massive tree trunk, its bark rough and cracked, a deep cut revealing the yellowed, fresh wood. At the peak of his power accumulation, countless wood chips of various sizes were scattered by the previous attack, solidifying into pieces of golden fragments illuminated by the light in the air, forming a sharp contrast with the deep forest green and earthy brown in the background. The whole scene, with the strong interweaving of light and shadow, perfectly captured this moment full of primitive power and dynamic beauty.", + "seed": 17157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_0/index_8624_seed_17157_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03609768003225326, + "vlm_logprob": -0.03609768003225326, + "vlm_yes_no_margin": 4.0499999761581424, + "vlm_token_credit": -0.03609768003225326, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12874780595302582 + ], + "opposite_logprob_tokens": [ + -2.1287477016448975 + ], + "logprob_reward": -0.12874780595302582, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12874780595302582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17709706723690033, + 0.14930960536003113, + 0.1670677214860916, + 0.10852757096290588, + 0.08680138736963272, + 0.04545239731669426, + 0.04633739963173866, + 0.0739198699593544, + 0.032379187643527985, + 0.11310775578022003 + ], + "credit_map": [ + 0.011084544472396374, + 0.02362179569900036, + 0.024890659376978874, + 0.02526901289820671, + 0.009842639788985252, + 0.02032565511763096, + 0.03080442175269127, + 0.03125835582613945, + 0.008690286427736282, + 0.010325146839022636, + 0.015020783990621567, + 0.01810823567211628, + 0.008244910277426243, + 0.03289913386106491, + 0.031483832746744156, + 0.024537282064557076, + 0.01001389417797327, + 0.028230512514710426, + 0.012666663154959679, + 0.03247317299246788, + 0.011059165000915527, + 0.027938319370150566, + 0.020488541573286057, + 0.024197464808821678, + 0.008522611111402512, + 0.010226558893918991, + 0.009952021762728691, + 0.017818883061408997, + 0.011930832639336586, + 0.016439422965049744, + 0.01653376966714859, + 0.017103468999266624, + 0.005573302041739225, + 0.011111602187156677, + 0.008914642035961151, + 0.016476718708872795, + 0.00732216564938426, + 0.017000209540128708, + 0.009668561629951, + 0.0107341930270195, + 0.0032084297854453325, + 0.005118665751069784, + 0.00390915060415864, + 0.007578491233289242, + 0.002349614165723324, + 0.01032606977969408, + 0.005503959953784943, + 0.00745801767334342, + 0.006406050641089678, + 0.006335424724966288, + 0.005420947913080454, + 0.00780620938166976, + 0.0031432397663593292, + 0.006115196272730827, + 0.005134508479386568, + 0.005975823383778334, + 0.003380921669304371, + 0.007453518453985453, + 0.0031975761521607637, + 0.013061248697340488, + 0.005969812627881765, + 0.008016693405807018, + 0.02007455565035343, + 0.012765549123287201, + 0.002455135341733694, + 0.004828648641705513, + 0.00389246572740376, + 0.005697740241885185, + 0.0032208915799856186, + 0.0037708673626184464, + 0.004630570765584707, + 0.0038828719407320023, + 0.008560840040445328, + 0.015417475253343582, + 0.010804246179759502, + 0.019117960706353188, + 0.006426941603422165, + 0.015253867022693157, + 0.016313236206769943, + 0.02121320366859436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.12874780595302582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646931774914265 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646931774914265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646931774914265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.159354567527771, + 0.08271259814500809, + 0.11379776149988174, + 0.15018410980701447, + 0.14348261058330536, + 0.023353153839707375, + 0.04346345737576485, + 0.11277932673692703, + 0.04064145311713219, + 0.13023105263710022 + ], + "credit_map": [ + 0.010958376340568066, + 0.014279009774327278, + 0.024888603016734123, + 0.02231697551906109, + 0.00804122444242239, + 0.023078573867678642, + 0.028524920344352722, + 0.027266860008239746, + 0.010758329182863235, + 0.013664653524756432, + 0.009941726922988892, + 0.011554551310837269, + 0.006095122545957565, + 0.007377432659268379, + 0.011582338251173496, + 0.011738430708646774, + 0.005540794692933559, + 0.015270905569195747, + 0.01309318095445633, + 0.021124133840203285, + 0.007086706813424826, + 0.02254427596926689, + 0.012075046077370644, + 0.017062710598111153, + 0.014405401423573494, + 0.030954817309975624, + 0.010469181463122368, + 0.026832502335309982, + 0.01193194929510355, + 0.010393825359642506, + 0.019570933654904366, + 0.025625472888350487, + 0.008381732739508152, + 0.019838007166981697, + 0.013690359890460968, + 0.03757448121905327, + 0.01126936823129654, + 0.01968957670032978, + 0.01300584152340889, + 0.020033234730362892, + 0.002619983395561576, + 0.0026213822420686483, + 0.0027203024365007877, + 0.002860249252989888, + 0.00216684490442276, + 0.0031114742159843445, + 0.0037964906077831984, + 0.0034564246889203787, + 0.0035351915284991264, + 0.008731045760214329, + 0.003977915272116661, + 0.008703180588781834, + 0.0033841757103800774, + 0.004863658454269171, + 0.00485979113727808, + 0.005408493336290121, + 0.003962748683989048, + 0.01253124512732029, + 0.005780417937785387, + 0.020999453961849213, + 0.008588384836912155, + 0.011836799792945385, + 0.03200041875243187, + 0.017079848796129227, + 0.0025448082014918327, + 0.007564056199043989, + 0.004556337371468544, + 0.008001187816262245, + 0.003526353742927313, + 0.0037380678113549948, + 0.005302577279508114, + 0.0054080612026154995, + 0.007269627880305052, + 0.024831540882587433, + 0.008898802101612091, + 0.020404912531375885, + 0.013714027591049671, + 0.012705167755484581, + 0.02172316424548626, + 0.02068379335105419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.004646931774914265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573781803250313 + ], + "opposite_logprob_tokens": [ + -3.895573854446411 + ], + "logprob_reward": -0.020573781803250313, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573781803250313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11133388429880142, + 0.05945581942796707, + 0.04570349305868149, + 0.048030171543359756, + 0.09700866043567657, + 0.030605653300881386, + 0.0560460202395916, + 0.4863382875919342, + 0.018614819273352623, + 0.046863213181495667 + ], + "credit_map": [ + 0.007714743725955486, + 0.018658027052879333, + 0.011190859600901604, + 0.019853558391332626, + 0.0060370671562850475, + 0.014471429400146008, + 0.017530985176563263, + 0.01587720774114132, + 0.005045957863330841, + 0.005242041777819395, + 0.005114672239869833, + 0.006643522996455431, + 0.004319461528211832, + 0.011603092774748802, + 0.011385992169380188, + 0.010101079940795898, + 0.0024748072028160095, + 0.005271182395517826, + 0.005014664959162474, + 0.005854352377355099, + 0.0049489219672977924, + 0.010613803751766682, + 0.005354475695639849, + 0.006171284709125757, + 0.0036043107975274324, + 0.008283168077468872, + 0.0037308388855308294, + 0.009821885265409946, + 0.00414402736350894, + 0.004455782473087311, + 0.005526094697415829, + 0.008464068174362183, + 0.005096010863780975, + 0.00854509323835373, + 0.0071577955968678, + 0.022044016048312187, + 0.009990188293159008, + 0.011947267688810825, + 0.019407525658607483, + 0.01282076258212328, + 0.002880794694647193, + 0.003641905263066292, + 0.0033960039727389812, + 0.004949889145791531, + 0.0024327512364834547, + 0.003789992071688175, + 0.0051528471522033215, + 0.004361472092568874, + 0.004260587971657515, + 0.007935933768749237, + 0.006808706559240818, + 0.010463879443705082, + 0.005281833466142416, + 0.0068037668243050575, + 0.007445496041327715, + 0.007045821752399206, + 0.02278776839375496, + 0.06983008980751038, + 0.023244770243763924, + 0.14558182656764984, + 0.04572714492678642, + 0.050148069858551025, + 0.046446479856967926, + 0.08257213979959488, + 0.0011910017346963286, + 0.0023329623509198427, + 0.0017165610333904624, + 0.004072018899023533, + 0.0015232706209644675, + 0.002493709558621049, + 0.0023058922961354256, + 0.0029794033616781235, + 0.0017949853790923953, + 0.007957610301673412, + 0.0023687586653977633, + 0.0078013925813138485, + 0.004554521292448044, + 0.005216061603277922, + 0.009812118485569954, + 0.0073577663861215115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.020573781803250313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946681834757328 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946681834757328, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946681834757328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10861075669527054, + 0.1504688560962677, + 0.091551773250103, + 0.14050090312957764, + 0.08129923790693283, + 0.14496566355228424, + 0.04470154270529747, + 0.07680179178714752, + 0.09936769306659698, + 0.0617317296564579 + ], + "credit_map": [ + 0.007343048229813576, + 0.017367610707879066, + 0.012316922657191753, + 0.014994463883340359, + 0.005507089197635651, + 0.010820559225976467, + 0.02041841484606266, + 0.019842645153403282, + 0.018868213519454002, + 0.0236797071993351, + 0.02007932774722576, + 0.023694835603237152, + 0.009894615970551968, + 0.015035660937428474, + 0.01778457872569561, + 0.02143191359937191, + 0.005767178721725941, + 0.006693465169519186, + 0.006910229567438364, + 0.010968667455017567, + 0.007010651286691427, + 0.021655825898051262, + 0.016211718320846558, + 0.016334034502506256, + 0.010655472055077553, + 0.025482170283794403, + 0.00944052916020155, + 0.022149641066789627, + 0.01246840413659811, + 0.020523980259895325, + 0.01909620873630047, + 0.02068449556827545, + 0.004316315054893494, + 0.00940406322479248, + 0.007038325071334839, + 0.01246473379433155, + 0.004459760617464781, + 0.021343346685171127, + 0.00867049302905798, + 0.013602204620838165, + 0.007972422987222672, + 0.021226394921541214, + 0.01206879410892725, + 0.031600866466760635, + 0.009455688297748566, + 0.0176743995398283, + 0.020938463509082794, + 0.024028632789850235, + 0.005360784009099007, + 0.006847289856523275, + 0.004766523372381926, + 0.006459575146436691, + 0.004699899349361658, + 0.0056433179415762424, + 0.005209372844547033, + 0.005714782979339361, + 0.004098779987543821, + 0.009813422337174416, + 0.004295137710869312, + 0.01639922335743904, + 0.006014072801917791, + 0.009089496918022633, + 0.014027245342731476, + 0.013064417988061905, + 0.006778412964195013, + 0.01549814734607935, + 0.009771851822733879, + 0.021928850561380386, + 0.008367208763957024, + 0.008633766323328018, + 0.014707360416650772, + 0.013682090677320957, + 0.0037560933269560337, + 0.01186341978609562, + 0.003670401405543089, + 0.009520666673779488, + 0.005959601607173681, + 0.006199888419359922, + 0.011436102911829948, + 0.009325557388365269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.005946681834757328 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573198795318604 + ], + "opposite_logprob_tokens": [ + -3.895573139190674 + ], + "logprob_reward": -0.020573198795318604, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573198795318604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27571406960487366, + 0.0634034126996994, + 0.08771546930074692, + 0.06724517047405243, + 0.1508864462375641, + 0.039259664714336395, + 0.04305840656161308, + 0.04713556542992592, + 0.10198717564344406, + 0.12359470874071121 + ], + "credit_map": [ + 0.019587436690926552, + 0.03874091058969498, + 0.0283082015812397, + 0.06409838795661926, + 0.014039691537618637, + 0.02923521026968956, + 0.0363457016646862, + 0.04535849019885063, + 0.0037986079696565866, + 0.0083800433203578, + 0.0060309539549052715, + 0.016495617106556892, + 0.004329327959567308, + 0.0063984510488808155, + 0.008176179602742195, + 0.009794226847589016, + 0.00512397987768054, + 0.008381943218410015, + 0.008335490711033344, + 0.00995428953319788, + 0.006667901761829853, + 0.019203532487154007, + 0.015936342999339104, + 0.014111975207924843, + 0.004597880877554417, + 0.00878383219242096, + 0.007056593894958496, + 0.01497511938214302, + 0.004803359508514404, + 0.006711401976644993, + 0.009733057580888271, + 0.010583911091089249, + 0.008111709728837013, + 0.015804242342710495, + 0.014507998712360859, + 0.03369154408574104, + 0.012159180827438831, + 0.029018815606832504, + 0.016101382672786713, + 0.021491559222340584, + 0.0029552276246249676, + 0.005168589763343334, + 0.003679771674796939, + 0.005147148855030537, + 0.004203328397125006, + 0.006583214737474918, + 0.005695495288819075, + 0.005826882552355528, + 0.0027663197834044695, + 0.008342458866536617, + 0.004380308091640472, + 0.008299197070300579, + 0.003472822718322277, + 0.006156133953481913, + 0.004409869201481342, + 0.005231292452663183, + 0.0027156441938132048, + 0.008628053590655327, + 0.0034498258028179407, + 0.010844703763723373, + 0.0035588087048381567, + 0.004202261101454496, + 0.006789233069866896, + 0.006947025191038847, + 0.00705983629450202, + 0.013814164325594902, + 0.010231364518404007, + 0.025083331391215324, + 0.008058495819568634, + 0.007093710359185934, + 0.015135303139686584, + 0.015510953962802887, + 0.006762627046555281, + 0.018726075068116188, + 0.008300410583615303, + 0.023355582728981972, + 0.010898097418248653, + 0.012839201837778091, + 0.019387491047382355, + 0.023325197398662567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.020573198795318604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a powerful, low-angle, medium shot, the entire scene is rendered in the style of a painterly work, with the paint applied thickly (impasto) to create distinct textures and strokes. At the center of the frame, a muscular woodcutter, chopping with all his might, sports a thick brown beard, a weathered face, and a focused expression strained by the intensity of the effort. He wears a frayed, dark red flannel shirt, sleeves rolled up to the elbows, revealing forearms muscular from constant labor. He grips a heavy chopping axe, its handle smooth from years of use, raised above his head. The blade gleams a cold metallic sheen in the dappled sunlight filtering through the forest floor, poised to bring down with a thunderous force. Before him lies a massive tree trunk, its bark rough and cracked, a deep cut revealing the yellowed, fresh wood. At the peak of his power accumulation, countless wood chips of various sizes were scattered by the previous attack, solidifying into pieces of golden fragments illuminated by the light in the air, forming a sharp contrast with the deep forest green and earthy brown in the background. The whole scene, with the strong interweaving of light and shadow, perfectly captured this moment full of primitive power and dynamic beauty.", + "seed": 17158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_0/index_8624_seed_17158_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020744804572314025, + "vlm_logprob": -0.020744804572314025, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.020744804572314025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07221047580242157 + ], + "opposite_logprob_tokens": [ + -2.6972105503082275 + ], + "logprob_reward": -0.07221047580242157, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07221047580242157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24824166297912598, + 0.25845927000045776, + 0.03853125125169754, + 0.05314108356833458, + 0.08733981102705002, + 0.05806738883256912, + 0.03181327506899834, + 0.04382660984992981, + 0.12969979643821716, + 0.05087987706065178 + ], + "credit_map": [ + 0.035881973803043365, + 0.022779416292905807, + 0.016210362315177917, + 0.022196125239133835, + 0.024040067568421364, + 0.04369264841079712, + 0.051775917410850525, + 0.0316651351749897, + 0.027899377048015594, + 0.023161882534623146, + 0.017841728404164314, + 0.01895902305841446, + 0.02801417000591755, + 0.016020016744732857, + 0.07290910929441452, + 0.053653933107852936, + 0.0062826815992593765, + 0.004136419855058193, + 0.002689570654183626, + 0.0044381506741046906, + 0.004831152036786079, + 0.005042937118560076, + 0.006422203965485096, + 0.0046881320886313915, + 0.005991778336465359, + 0.005358164664357901, + 0.0031559185590595007, + 0.00427008094266057, + 0.005389797501266003, + 0.013091868720948696, + 0.00811971165239811, + 0.0077637555077672005, + 0.01296226680278778, + 0.009225095622241497, + 0.005130743142217398, + 0.014697885140776634, + 0.007867594249546528, + 0.010765542276203632, + 0.014470390044152737, + 0.012220285832881927, + 0.012937042862176895, + 0.004649939015507698, + 0.0026968757156282663, + 0.00549476221203804, + 0.004690380766987801, + 0.01785697601735592, + 0.00592816574499011, + 0.0038132392801344395, + 0.0039847674779593945, + 0.0050633070059120655, + 0.0018151313997805119, + 0.0029199046548455954, + 0.005997896660119295, + 0.004939934238791466, + 0.0038439540658146143, + 0.003248379100114107, + 0.0074935778975486755, + 0.003259864170104265, + 0.0027349311858415604, + 0.003913891036063433, + 0.0038429475389420986, + 0.005696213338524103, + 0.007455042097717524, + 0.009430141188204288, + 0.020583977922797203, + 0.008045583963394165, + 0.0073374491184949875, + 0.012593292631208897, + 0.01581120863556862, + 0.020886920392513275, + 0.023587603121995926, + 0.02085374854505062, + 0.007674998138099909, + 0.009016396477818489, + 0.0026514993514865637, + 0.0056648291647434235, + 0.004202071111649275, + 0.007107938174158335, + 0.007226832676678896, + 0.007335307076573372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.07221047580242157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645982291549444 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.004645982291549444, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645982291549444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08720532059669495, + 0.14259815216064453, + 0.0646899938583374, + 0.08470486104488373, + 0.05690620467066765, + 0.03439054265618324, + 0.0661405399441719, + 0.08751837909221649, + 0.07808643579483032, + 0.29775962233543396 + ], + "credit_map": [ + 0.011847791261970997, + 0.007679898291826248, + 0.008235529996454716, + 0.010279337875545025, + 0.013411644846200943, + 0.01188667118549347, + 0.014293094165623188, + 0.009571356698870659, + 0.024233773350715637, + 0.016251586377620697, + 0.007233524229377508, + 0.009050250984728336, + 0.02602110616862774, + 0.016546839848160744, + 0.028298310935497284, + 0.01496275793761015, + 0.007349902763962746, + 0.005679439753293991, + 0.003474628319963813, + 0.004838207736611366, + 0.011120698414742947, + 0.010227041319012642, + 0.01456101518124342, + 0.0074390568770468235, + 0.01689162291586399, + 0.009559253230690956, + 0.004401419777423143, + 0.004967780318111181, + 0.01399956177920103, + 0.009592185728251934, + 0.017609739676117897, + 0.007683298550546169, + 0.0074217417277395725, + 0.004709939938038588, + 0.0025708735920488834, + 0.008130019530653954, + 0.005294196307659149, + 0.008362474851310253, + 0.012695031240582466, + 0.007721931207925081, + 0.008957936428487301, + 0.002349493559449911, + 0.0023015395272523165, + 0.0024368043523281813, + 0.0014306429075077176, + 0.012325872667133808, + 0.0024655372835695744, + 0.002122717211022973, + 0.008667808026075363, + 0.010705532506108284, + 0.0036066207103431225, + 0.006083832122385502, + 0.011362062767148018, + 0.010751824826002121, + 0.007410626858472824, + 0.007552230730652809, + 0.01437646895647049, + 0.01006812322884798, + 0.00561324181035161, + 0.007394464686512947, + 0.00844605267047882, + 0.013891873881220818, + 0.013081871904432774, + 0.014646278694272041, + 0.011304385028779507, + 0.01134475041180849, + 0.00415736623108387, + 0.012062056921422482, + 0.007486422546207905, + 0.009153997525572777, + 0.01230709720402956, + 0.010270361788570881, + 0.04301463067531586, + 0.041139356791973114, + 0.017112089321017265, + 0.021229881793260574, + 0.03632296621799469, + 0.04939073696732521, + 0.043717529624700546, + 0.04583242908120155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.004645982291549444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059465630911290646 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.0059465630911290646, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059465630911290646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16339418292045593, + 0.34333929419517517, + 0.0760289803147316, + 0.08685504645109177, + 0.06444749236106873, + 0.04258149489760399, + 0.07585824280977249, + 0.061469901353120804, + 0.05282546207308769, + 0.03319988772273064 + ], + "credit_map": [ + 0.019862914457917213, + 0.01532787550240755, + 0.016250232234597206, + 0.019256718456745148, + 0.016876088455319405, + 0.02424117550253868, + 0.020033646374940872, + 0.03154553845524788, + 0.06501288712024689, + 0.014330267906188965, + 0.009837983176112175, + 0.012560073286294937, + 0.016850681975483894, + 0.02118946984410286, + 0.1302219182252884, + 0.07333602756261826, + 0.009112105704843998, + 0.004491084720939398, + 0.003514296840876341, + 0.0046170903369784355, + 0.013012118637561798, + 0.017192909494042397, + 0.01641269586980343, + 0.007676674984395504, + 0.008616396225988865, + 0.010979924350976944, + 0.004965273197740316, + 0.006150614470243454, + 0.011026524007320404, + 0.014876004308462143, + 0.016557764261960983, + 0.013682539574801922, + 0.008547053672373295, + 0.0055455490946769714, + 0.003125209826976061, + 0.008229526691138744, + 0.005776476114988327, + 0.00955103524029255, + 0.01500402856618166, + 0.008668620139360428, + 0.007881969213485718, + 0.0030535138212144375, + 0.0025667096488177776, + 0.0038591595366597176, + 0.003014982445165515, + 0.010287631303071976, + 0.005985305644571781, + 0.005932220723479986, + 0.011619003489613533, + 0.00887480191886425, + 0.0038668017368763685, + 0.005389966070652008, + 0.012631429359316826, + 0.014919075183570385, + 0.010399773716926575, + 0.008157391101121902, + 0.009212682023644447, + 0.00530980434268713, + 0.005081626120954752, + 0.005998257547616959, + 0.006028736010193825, + 0.009827530942857265, + 0.011678478680551052, + 0.008332786150276661, + 0.008381201885640621, + 0.00566850183531642, + 0.002850125078111887, + 0.004413902759552002, + 0.006395064759999514, + 0.007374824956059456, + 0.009289759211242199, + 0.008452081121504307, + 0.0044672670774161816, + 0.00457192724570632, + 0.001976022031158209, + 0.004356974270194769, + 0.003367031691595912, + 0.005082826130092144, + 0.004493824206292629, + 0.004884013440459967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0059465630911290646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732518319040537 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732518319040537, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732518319040537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12916293740272522, + 0.147578164935112, + 0.07725762575864792, + 0.07787203043699265, + 0.12474438548088074, + 0.07054296135902405, + 0.04221375659108162, + 0.1235332190990448, + 0.1707620471715927, + 0.03633289039134979 + ], + "credit_map": [ + 0.013483481481671333, + 0.012556234374642372, + 0.009874810464680195, + 0.029966631904244423, + 0.015984445810317993, + 0.014931698329746723, + 0.01690680906176567, + 0.015458819456398487, + 0.030133305117487907, + 0.0074759856797754765, + 0.008745060302317142, + 0.012284031137824059, + 0.01341305673122406, + 0.012691738083958626, + 0.021463574841618538, + 0.04137139767408371, + 0.011460859328508377, + 0.00554191879928112, + 0.003706887364387512, + 0.005707204807549715, + 0.011185343377292156, + 0.014791957102715969, + 0.01452602818608284, + 0.010337422601878643, + 0.010867052711546421, + 0.006383855827152729, + 0.004440652672201395, + 0.004686159547418356, + 0.012218225747346878, + 0.01813388615846634, + 0.010817029513418674, + 0.010325168259441853, + 0.018923945724964142, + 0.008561204187572002, + 0.006985879503190517, + 0.016728544607758522, + 0.011632684618234634, + 0.016799839213490486, + 0.025765899568796158, + 0.019346386194229126, + 0.01770145632326603, + 0.004098088014870882, + 0.004624958615750074, + 0.0061350432224571705, + 0.005300314165651798, + 0.02272239699959755, + 0.005944274365901947, + 0.004016426857560873, + 0.007325167767703533, + 0.0056210788898169994, + 0.002353017684072256, + 0.005450127646327019, + 0.006464512553066015, + 0.006041256245225668, + 0.004483252763748169, + 0.004475336987525225, + 0.01966664381325245, + 0.014147808775305748, + 0.01010158658027649, + 0.01498524472117424, + 0.010483764111995697, + 0.016099514439702034, + 0.019201070070266724, + 0.01884758286178112, + 0.02524646557867527, + 0.025244057178497314, + 0.01423296332359314, + 0.026492761448025703, + 0.01414198987185955, + 0.026156505569815636, + 0.020795609802007675, + 0.018451692536473274, + 0.005649368278682232, + 0.005145384930074215, + 0.0020972846541553736, + 0.004162836354225874, + 0.004452436696738005, + 0.004865425173193216, + 0.005231892690062523, + 0.004728260450065136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.006732518319040537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a painterly style with thick impasto textures and visible brush strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot framed as a low-angle medium shot of a muscular woodcutter with a thick brown beard and focused expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter wearing a frayed dark red flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woodcutter holding a heavy axe raised above his head, poised to strike a massive tree trunk with a visible deep cut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are golden wood chips scattered in the air around the woodcutter, contrasting with the deep forest green and earthy brown background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188483357429504 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188483357429504, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188483357429504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1485087275505066, + 0.1770980954170227, + 0.02731395699083805, + 0.07695113122463226, + 0.053809694945812225, + 0.03659289330244064, + 0.02399558573961258, + 0.0466860756278038, + 0.3665323257446289, + 0.042511552572250366 + ], + "credit_map": [ + 0.021096043288707733, + 0.017943209037184715, + 0.009905544109642506, + 0.018235445022583008, + 0.020653078332543373, + 0.02216433919966221, + 0.018645266070961952, + 0.01986580528318882, + 0.033541202545166016, + 0.017378002405166626, + 0.016420932486653328, + 0.0231460053473711, + 0.021285058930516243, + 0.015740899369120598, + 0.03125197812914848, + 0.01833401992917061, + 0.0031073768623173237, + 0.002098609460517764, + 0.001828362001106143, + 0.0030703216325491667, + 0.00429500499740243, + 0.00618725223466754, + 0.0037917285226285458, + 0.0029353033751249313, + 0.017446262761950493, + 0.006524856202304363, + 0.0034148921258747578, + 0.005258675664663315, + 0.011210234835743904, + 0.008047414943575859, + 0.017316607758402824, + 0.007732197176665068, + 0.007601195015013218, + 0.005737705621868372, + 0.00316267809830606, + 0.010070470161736012, + 0.003701453097164631, + 0.0067762406542897224, + 0.011665322817862034, + 0.005094633437693119, + 0.007020802702754736, + 0.0044328453950583935, + 0.0022442429326474667, + 0.0023620021529495716, + 0.003422854235395789, + 0.008741081692278385, + 0.005095554515719414, + 0.0032735138665884733, + 0.003722884226590395, + 0.0035662720911204815, + 0.001595943234860897, + 0.004332622047513723, + 0.002101518679410219, + 0.003528670407831669, + 0.0018900107825174928, + 0.003257666015997529, + 0.007069546263664961, + 0.005428078584372997, + 0.003431921359151602, + 0.005944248288869858, + 0.004978301003575325, + 0.0076885647140443325, + 0.007120124995708466, + 0.00502529414370656, + 0.051843706518411636, + 0.07247129827737808, + 0.023057108744978905, + 0.03702329099178314, + 0.031483519822359085, + 0.029509814456105232, + 0.06273769587278366, + 0.05840590223670006, + 0.006674573756754398, + 0.008051726035773754, + 0.003267226740717888, + 0.004117385949939489, + 0.0044172401539981365, + 0.004692510236054659, + 0.006825369317084551, + 0.004465519450604916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.014188483357429504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_tensor([1983, 1983])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_tensor([1983, 1983])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d8c11ca02238ff5f1416ba769cd0a5aaf7f04543 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_tensor([1983, 1983])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from an extreme overhead angle, the camera tilts downward at breakneck speed, closely following a sleek, silver, metallic fighter jet as it violently tears through a vast, thick sea of white clouds. Almost instantly, it enters a perfect 90-degree vertical dive, plummeting downward at breathtaking speed. Its polished fuselage reflects a sharp gleam in the glaring sunlight at high altitude, its swept wings and single-seat bubble-shaped cockpit clearly visible. Due to the immense air resistance and hypersonic speed, the entire fuselage vibrates visibly, even subtle ripples and distortions appearing in the metal skin. A translucent sonic boom cloud forms around the wing edges and nose. The fighter jet leaves a swirling, chaotic vortex trail through the clouds, filling the entire scene with a sense of pure power, extreme speed, and awe-inspiring tension, creating an adrenaline-pumping visual impact.", + "seed": 31321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_1983_seed_31321_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020908139739185573, + "vlm_logprob": -0.020908139739185573, + "vlm_yes_no_margin": 4.1999999523162845, + "vlm_token_credit": -0.020908139739185573, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035747725516557693 + ], + "opposite_logprob_tokens": [ + -4.410747528076172 + ], + "logprob_reward": -0.035747725516557693, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.035747725516557693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12497318536043167, + 0.11554587632417679, + 0.09460519999265671, + 0.22707529366016388, + 0.10034724324941635, + 0.12398677319288254, + 0.029166042804718018, + 0.09585505723953247, + 0.040197473019361496, + 0.04824787750840187 + ], + "credit_map": [ + 0.011015244759619236, + 0.011219797655940056, + 0.01292050164192915, + 0.00972762145102024, + 0.018192579969763756, + 0.018600087612867355, + 0.009904890321195126, + 0.03339245915412903, + 0.008917854167521, + 0.01638207584619522, + 0.0120224105194211, + 0.01977779157459736, + 0.016572846099734306, + 0.015208820812404156, + 0.010064060799777508, + 0.016600025817751884, + 0.011891983449459076, + 0.011226519010961056, + 0.01868239790201187, + 0.01654883660376072, + 0.007387088146060705, + 0.01305821817368269, + 0.009565641172230244, + 0.006244521122425795, + 0.023948198184370995, + 0.03286953270435333, + 0.018968138843774796, + 0.024927537888288498, + 0.02590344101190567, + 0.03791004419326782, + 0.038577694445848465, + 0.023970700800418854, + 0.009557262994349003, + 0.0140869515016675, + 0.01393099594861269, + 0.015395646914839745, + 0.01129855401813984, + 0.01158488541841507, + 0.01494103018194437, + 0.00955191534012556, + 0.020369218662381172, + 0.008617149665951729, + 0.03082105703651905, + 0.013027696870267391, + 0.008932339958846569, + 0.022293856367468834, + 0.008465014398097992, + 0.0114604402333498, + 0.005962012801319361, + 0.0030564707703888416, + 0.0035922150127589703, + 0.0030235922895371914, + 0.0017961985431611538, + 0.0030678596813231707, + 0.0025230487808585167, + 0.006144641432911158, + 0.014503263868391514, + 0.012689859606325626, + 0.014887718483805656, + 0.010854335501790047, + 0.008787604048848152, + 0.011237788945436478, + 0.011847518384456635, + 0.011046971194446087, + 0.004749591927975416, + 0.0039594462141394615, + 0.003867357037961483, + 0.005335030145943165, + 0.005507130641490221, + 0.005777337122708559, + 0.006660740822553635, + 0.004340836778283119, + 0.009201794862747192, + 0.0064779105596244335, + 0.006350277457386255, + 0.004668368957936764, + 0.0038035905454307795, + 0.004647598601877689, + 0.0061806063167750835, + 0.006917724385857582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.035747725516557693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059386856853962 + ], + "opposite_logprob_tokens": [ + -3.8955938816070557 + ], + "logprob_reward": -0.02059386856853962, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059386856853962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17781852185726166, + 0.22047995030879974, + 0.056615207344293594, + 0.10430959612131119, + 0.0770188570022583, + 0.1062777191400528, + 0.0638766959309578, + 0.09853087365627289, + 0.059441257268190384, + 0.03563133254647255 + ], + "credit_map": [ + 0.0072320387698709965, + 0.018925953656435013, + 0.02822164259850979, + 0.015752043575048447, + 0.029505504295229912, + 0.023878248408436775, + 0.022537440061569214, + 0.0317656509578228, + 0.013563493266701698, + 0.035595644265413284, + 0.026622382923960686, + 0.024641651660203934, + 0.04171339422464371, + 0.032217513769865036, + 0.0220514927059412, + 0.024074377492070198, + 0.008044900372624397, + 0.008666999638080597, + 0.009501435793936253, + 0.0073252152651548386, + 0.004827235825359821, + 0.007907983846962452, + 0.00479523092508316, + 0.005546205677092075, + 0.009333211928606033, + 0.022855881601572037, + 0.01076065469533205, + 0.01332351379096508, + 0.015024606138467789, + 0.012903384864330292, + 0.010170749388635159, + 0.00993758998811245, + 0.0062817903235554695, + 0.009223268367350101, + 0.012129471637308598, + 0.01892509125173092, + 0.004991413094103336, + 0.007138922810554504, + 0.011293633840978146, + 0.007035264279693365, + 0.02130413055419922, + 0.008009802550077438, + 0.023958440870046616, + 0.01184757985174656, + 0.009462734684348106, + 0.012113283388316631, + 0.007131762802600861, + 0.012449988164007664, + 0.013091149739921093, + 0.005294665228575468, + 0.006779939401894808, + 0.0074543943628668785, + 0.005236650351434946, + 0.008635533042252064, + 0.007533268071711063, + 0.009851096197962761, + 0.01838364265859127, + 0.008022135123610497, + 0.01898021437227726, + 0.008973205462098122, + 0.009254739619791508, + 0.017168402671813965, + 0.009387508034706116, + 0.008361025713384151, + 0.005134297534823418, + 0.006137517746537924, + 0.0039023475255817175, + 0.008080060593783855, + 0.00997853185981512, + 0.007387847639620304, + 0.012924709357321262, + 0.005895944777876139, + 0.007615411188453436, + 0.004877128638327122, + 0.0041443416848778725, + 0.0033138063736259937, + 0.003639254719018936, + 0.0031578405760228634, + 0.004328402224928141, + 0.004555149003863335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.02059386856853962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182972446084023 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.018182972446084023, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018182972446084023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10898851603269577, + 0.0922548845410347, + 0.12393262982368469, + 0.15899647772312164, + 0.11245787143707275, + 0.14183849096298218, + 0.08726193010807037, + 0.0942627564072609, + 0.044063277542591095, + 0.03594319522380829 + ], + "credit_map": [ + 0.006198135670274496, + 0.010026308707892895, + 0.01844831369817257, + 0.009140840731561184, + 0.02033711038529873, + 0.016462378203868866, + 0.010410193353891373, + 0.01796523854136467, + 0.005139381159096956, + 0.011150398291647434, + 0.011345872655510902, + 0.011583497747778893, + 0.013413512147963047, + 0.012734489515423775, + 0.007555067539215088, + 0.01933266408741474, + 0.016333535313606262, + 0.013356911949813366, + 0.027147779241204262, + 0.023285778239369392, + 0.009186150506138802, + 0.01720270700752735, + 0.00911938026547432, + 0.00830038357526064, + 0.018015725538134575, + 0.02749302238225937, + 0.01697382889688015, + 0.021033048629760742, + 0.021851226687431335, + 0.017697324976325035, + 0.015986096113920212, + 0.019946206361055374, + 0.009712358005344868, + 0.017450451850891113, + 0.015498434193432331, + 0.022307366132736206, + 0.009139691479504108, + 0.013131864368915558, + 0.015971729531884193, + 0.009245981462299824, + 0.022734858095645905, + 0.013366062194108963, + 0.02581890858709812, + 0.01730332523584366, + 0.01515559945255518, + 0.019481906667351723, + 0.013054254464805126, + 0.014923572540283203, + 0.013807851821184158, + 0.008984400890767574, + 0.01208909135311842, + 0.010470383800566196, + 0.009004488587379456, + 0.009496426209807396, + 0.011051093228161335, + 0.012358194217085838, + 0.01249765232205391, + 0.01092590019106865, + 0.012293321080505848, + 0.013455946929752827, + 0.009470929391682148, + 0.013201053254306316, + 0.012948595918715, + 0.009469352662563324, + 0.002769323531538248, + 0.005469676107168198, + 0.0037019550800323486, + 0.005779342260211706, + 0.006742202676832676, + 0.005720505025237799, + 0.009496811777353287, + 0.0043834601528942585, + 0.006242379546165466, + 0.005669750738888979, + 0.003278225427493453, + 0.004307878669351339, + 0.003111023223027587, + 0.003976449836045504, + 0.0045639341697096825, + 0.0047935545444488525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.018182972446084023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067374915815889835 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.0067374915815889835, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067374915815889835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20539823174476624, + 0.0600440539419651, + 0.08502035588026047, + 0.12329418957233429, + 0.06891587376594543, + 0.10138501226902008, + 0.03589267283678055, + 0.26171305775642395, + 0.020250443369150162, + 0.038086120039224625 + ], + "credit_map": [ + 0.012033980339765549, + 0.01923142559826374, + 0.03689894825220108, + 0.013933675363659859, + 0.027991624549031258, + 0.031860847026109695, + 0.01917312480509281, + 0.04427460953593254, + 0.006183892022818327, + 0.005003457423299551, + 0.007195233833044767, + 0.008231920190155506, + 0.006019067484885454, + 0.012568366713821888, + 0.003925493452697992, + 0.010916621424257755, + 0.012171163223683834, + 0.010140115395188332, + 0.014210063964128494, + 0.014209468849003315, + 0.00587331410497427, + 0.013490267097949982, + 0.007640470750629902, + 0.007285493426024914, + 0.014265673235058784, + 0.01682216301560402, + 0.010526477359235287, + 0.014136062003672123, + 0.014148438349366188, + 0.01835153065621853, + 0.020548490807414055, + 0.014495357871055603, + 0.00710873119533062, + 0.0076584722846746445, + 0.009607960470020771, + 0.012575339525938034, + 0.007024512160569429, + 0.007473678793758154, + 0.011107905767858028, + 0.006359271705150604, + 0.0170120969414711, + 0.00809186976402998, + 0.014896643348038197, + 0.01414551679044962, + 0.011246034875512123, + 0.014335602521896362, + 0.008176780305802822, + 0.013480464927852154, + 0.004363146610558033, + 0.004337780177593231, + 0.00520749157294631, + 0.004656539764255285, + 0.004740905947983265, + 0.0038798688910901546, + 0.004260173067450523, + 0.004446765407919884, + 0.03543136641383171, + 0.03141332417726517, + 0.046223804354667664, + 0.03481372445821762, + 0.020092420279979706, + 0.04332728311419487, + 0.027546998113393784, + 0.022864099591970444, + 0.0016465156804770231, + 0.0034791240468621254, + 0.0019310235511511564, + 0.002865781309083104, + 0.0029031240846961737, + 0.0022664228454232216, + 0.003264753380790353, + 0.0018936996348202229, + 0.005181230138987303, + 0.005386752542108297, + 0.0057122791185975075, + 0.004937415011227131, + 0.003983883187174797, + 0.0035131715703755617, + 0.004425032064318657, + 0.004946356639266014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0067374915815889835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327864058315754 + ], + "opposite_logprob_tokens": [ + -3.7732787132263184 + ], + "logprob_reward": -0.02327864058315754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327864058315754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1128755584359169, + 0.06650176644325256, + 0.06658351421356201, + 0.12751887738704681, + 0.07257730513811111, + 0.24366995692253113, + 0.06399589776992798, + 0.1613035947084427, + 0.05168576166033745, + 0.0332876518368721 + ], + "credit_map": [ + 0.00770303700119257, + 0.009887457825243473, + 0.009358929470181465, + 0.008795016445219517, + 0.017166513949632645, + 0.018574345856904984, + 0.01206064224243164, + 0.02932961843907833, + 0.0056173973716795444, + 0.00983426719903946, + 0.008051902055740356, + 0.007822445593774319, + 0.010164232924580574, + 0.007618142291903496, + 0.006761073134839535, + 0.010632309131324291, + 0.009148708544671535, + 0.009299653582274914, + 0.010643135756254196, + 0.013339987024664879, + 0.00509438244625926, + 0.010121235623955727, + 0.004736039321869612, + 0.004200372379273176, + 0.014141788706183434, + 0.016061460599303246, + 0.011488939635455608, + 0.01330032479017973, + 0.018657702952623367, + 0.0193423330783844, + 0.02075777016580105, + 0.013768572360277176, + 0.00746974116191268, + 0.007842758670449257, + 0.01031765341758728, + 0.012972514145076275, + 0.007374748587608337, + 0.008583227172493935, + 0.009069281630218029, + 0.008947380818426609, + 0.04163653403520584, + 0.01710873283445835, + 0.06212968751788139, + 0.027083057910203934, + 0.0170734953135252, + 0.043057218194007874, + 0.015246101655066013, + 0.02033512480556965, + 0.011238502338528633, + 0.004330468364059925, + 0.009320429526269436, + 0.006277941167354584, + 0.005498987156897783, + 0.009251954033970833, + 0.008454826660454273, + 0.009622788056731224, + 0.026025447994470596, + 0.009051312692463398, + 0.03173127397894859, + 0.021055594086647034, + 0.012595092877745628, + 0.030359728261828423, + 0.0167593564838171, + 0.013725779950618744, + 0.003920356277376413, + 0.005714998114854097, + 0.004733392968773842, + 0.007909473031759262, + 0.005936936475336552, + 0.006754021160304546, + 0.010573805309832096, + 0.006142780650407076, + 0.007015563547611237, + 0.00445591751486063, + 0.003963342867791653, + 0.0029539705719798803, + 0.003416049759835005, + 0.0031252740882337093, + 0.003735723439604044, + 0.004621810745447874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.02327864058315754 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an extreme overhead angle, the camera tilts downward at breakneck speed, closely following a sleek, silver, metallic fighter jet as it violently tears through a vast, thick sea of white clouds. Almost instantly, it enters a perfect 90-degree vertical dive, plummeting downward at breathtaking speed. Its polished fuselage reflects a sharp gleam in the glaring sunlight at high altitude, its swept wings and single-seat bubble-shaped cockpit clearly visible. Due to the immense air resistance and hypersonic speed, the entire fuselage vibrates visibly, even subtle ripples and distortions appearing in the metal skin. A translucent sonic boom cloud forms around the wing edges and nose. The fighter jet leaves a swirling, chaotic vortex trail through the clouds, filling the entire scene with a sense of pure power, extreme speed, and awe-inspiring tension, creating an adrenaline-pumping visual impact.", + "seed": 31322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_1983_seed_31322_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03921443969011307, + "vlm_logprob": -0.03921443969011307, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03921443969011307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1307043433189392 + ], + "opposite_logprob_tokens": [ + -2.505704402923584 + ], + "logprob_reward": -0.1307043433189392, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1307043433189392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09032382816076279, + 0.045037891715765, + 0.11357545852661133, + 0.038134120404720306, + 0.1528598666191101, + 0.1778331845998764, + 0.08109898120164871, + 0.037557605654001236, + 0.10303491353988647, + 0.16054412722587585 + ], + "credit_map": [ + 0.008605037815868855, + 0.01740192621946335, + 0.007821282371878624, + 0.0067833890207111835, + 0.008901795372366905, + 0.012707045301795006, + 0.007215803489089012, + 0.02088753879070282, + 0.005557352676987648, + 0.00694349966943264, + 0.0057217711582779884, + 0.005312664434313774, + 0.005954788066446781, + 0.006221321411430836, + 0.004168224520981312, + 0.005158271640539169, + 0.01605950854718685, + 0.02766195870935917, + 0.014981514774262905, + 0.014954209327697754, + 0.009691061452031136, + 0.01460560504347086, + 0.007241337094455957, + 0.00838025938719511, + 0.003813351271674037, + 0.006324134301394224, + 0.004978514276444912, + 0.00506823742762208, + 0.005121639929711819, + 0.005760465282946825, + 0.0027241043280810118, + 0.004343670327216387, + 0.020847629755735397, + 0.017153119668364525, + 0.019676264375448227, + 0.025605803355574608, + 0.020610956475138664, + 0.016897980123758316, + 0.015794074162840843, + 0.01627403125166893, + 0.0260526891797781, + 0.00997103750705719, + 0.0463082529604435, + 0.013496103696525097, + 0.009719825349748135, + 0.02622479386627674, + 0.007843823172152042, + 0.03821667656302452, + 0.007954820059239864, + 0.008333747275173664, + 0.008703907020390034, + 0.008141727186739445, + 0.007783627603203058, + 0.01462493184953928, + 0.010413721203804016, + 0.015142494812607765, + 0.003716445993632078, + 0.0022545307874679565, + 0.011326813139021397, + 0.004603439476341009, + 0.0025749006308615208, + 0.003824458923190832, + 0.0022919217590242624, + 0.006965096108615398, + 0.01263169851154089, + 0.007481856271624565, + 0.011316381394863129, + 0.007121387403458357, + 0.011897259391844273, + 0.025720050558447838, + 0.0073377275839447975, + 0.019528549164533615, + 0.015129433944821358, + 0.0072118607349693775, + 0.04908525571227074, + 0.00967034325003624, + 0.009787195362150669, + 0.012986348941922188, + 0.010255644097924232, + 0.046418048441410065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.1307043433189392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751463748514652 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751463748514652, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751463748514652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16029547154903412, + 0.06141803413629532, + 0.06842445582151413, + 0.03829764947295189, + 0.09625333547592163, + 0.2856515645980835, + 0.08913937211036682, + 0.033972516655921936, + 0.05745786800980568, + 0.10908976197242737 + ], + "credit_map": [ + 0.01217559352517128, + 0.025103027001023293, + 0.017464689910411835, + 0.020190907642245293, + 0.01787438429892063, + 0.023250000551342964, + 0.01616336964070797, + 0.02807351015508175, + 0.006728240754455328, + 0.008691403083503246, + 0.006667664274573326, + 0.0069534145295619965, + 0.008487764745950699, + 0.008806605823338032, + 0.010485905222594738, + 0.004597039427608252, + 0.008439926430583, + 0.012150888331234455, + 0.008990383706986904, + 0.00849346537142992, + 0.0072300853207707405, + 0.01002742163836956, + 0.005633192602545023, + 0.007459098938852549, + 0.004064877983182669, + 0.00668667908757925, + 0.00499038677662611, + 0.0049337162636220455, + 0.005071451887488365, + 0.005570399574935436, + 0.0025971708819270134, + 0.0043829698115587234, + 0.012495754286646843, + 0.01068209484219551, + 0.01452720444649458, + 0.01484939269721508, + 0.01342866849154234, + 0.009554957039654255, + 0.01049328688532114, + 0.010221980512142181, + 0.029819747433066368, + 0.0221982691437006, + 0.06589928269386292, + 0.023735538125038147, + 0.024244070053100586, + 0.04076298698782921, + 0.016287153586745262, + 0.06270453333854675, + 0.009967464953660965, + 0.008166292682290077, + 0.012689516879618168, + 0.007711815182119608, + 0.009515520185232162, + 0.015354708768427372, + 0.010666714049875736, + 0.015067354775965214, + 0.0033535135444253683, + 0.0024030786007642746, + 0.010789984837174416, + 0.004201302770525217, + 0.003026082646101713, + 0.003977281507104635, + 0.0019985055550932884, + 0.004222770221531391, + 0.007068900857120752, + 0.0070938956923782825, + 0.0056845019571483135, + 0.005816061981022358, + 0.006604440044611692, + 0.011205911636352539, + 0.005557477008551359, + 0.008426686748862267, + 0.00983272772282362, + 0.004747854545712471, + 0.03505943343043327, + 0.006790122017264366, + 0.0063946121372282505, + 0.007114052306860685, + 0.006580182816833258, + 0.03257077932357788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.006751463748514652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695563673973083 + ], + "opposite_logprob_tokens": [ + -3.4086954593658447 + ], + "logprob_reward": -0.033695563673973083, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695563673973083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08721280843019485, + 0.07978850603103638, + 0.051351796835660934, + 0.07753921300172806, + 0.1214914545416832, + 0.09727609902620316, + 0.24342891573905945, + 0.06963981688022614, + 0.13009372353553772, + 0.042177699506282806 + ], + "credit_map": [ + 0.008770549669861794, + 0.010831532068550587, + 0.010301641188561916, + 0.010766509920358658, + 0.011835301294922829, + 0.011974604800343513, + 0.012995746918022633, + 0.00973691139370203, + 0.010638466104865074, + 0.009900625795125961, + 0.010299109853804111, + 0.007751953322440386, + 0.01240498200058937, + 0.009889673441648483, + 0.009257322177290916, + 0.009646361693739891, + 0.008558537811040878, + 0.0066530234180390835, + 0.006596122402697802, + 0.006564082112163305, + 0.004171864595264196, + 0.007838171906769276, + 0.003957412205636501, + 0.007012576330453157, + 0.01191652286797762, + 0.009795566089451313, + 0.008764367550611496, + 0.013342717662453651, + 0.0068602897226810455, + 0.010667179711163044, + 0.0074660200625658035, + 0.008726537227630615, + 0.017033345997333527, + 0.013822266831994057, + 0.01767498441040516, + 0.01271043624728918, + 0.012301741167902946, + 0.018555181100964546, + 0.011347309686243534, + 0.018046177923679352, + 0.011664124205708504, + 0.007230327930301428, + 0.01651610992848873, + 0.009720065630972385, + 0.00860822293907404, + 0.01810508407652378, + 0.007360469084233046, + 0.01807168871164322, + 0.022592084482312202, + 0.024289119988679886, + 0.024953335523605347, + 0.022273993119597435, + 0.02198987826704979, + 0.04767683520913124, + 0.02786463126540184, + 0.05178900808095932, + 0.0076342313550412655, + 0.0041121686808764935, + 0.022256778553128242, + 0.009027707390487194, + 0.004498591646552086, + 0.007640883792191744, + 0.003270423272624612, + 0.011199026368558407, + 0.015175866894423962, + 0.010644130408763885, + 0.013761560432612896, + 0.009307345375418663, + 0.014506730251014233, + 0.03242601454257965, + 0.011131925508379936, + 0.023140141740441322, + 0.004035421181470156, + 0.00443364679813385, + 0.010729138739407063, + 0.006295376922935247, + 0.0033920113928616047, + 0.00320493639446795, + 0.0034346473403275013, + 0.006652518175542355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.033695563673973083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738557480275631 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738557480275631, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738557480275631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16219820082187653, + 0.11361047625541687, + 0.05922996997833252, + 0.08814775943756104, + 0.09683138132095337, + 0.0595909021794796, + 0.08683507889509201, + 0.0441146120429039, + 0.13412339985370636, + 0.1553182154893875 + ], + "credit_map": [ + 0.017408018931746483, + 0.02091079391539097, + 0.016590282320976257, + 0.022332819178700447, + 0.02018507942557335, + 0.028698528185486794, + 0.018824322149157524, + 0.017248347401618958, + 0.011968516744673252, + 0.01390717551112175, + 0.016101043671369553, + 0.01204257644712925, + 0.015109879896044731, + 0.01713646575808525, + 0.01329929567873478, + 0.014045522548258305, + 0.009833278134465218, + 0.011191600933670998, + 0.009569169022142887, + 0.006057257764041424, + 0.004620708059519529, + 0.007943541742861271, + 0.0042069884948432446, + 0.0058074225671589375, + 0.01115037314593792, + 0.01733963005244732, + 0.009966396726667881, + 0.009051043540239334, + 0.009241042658686638, + 0.015592548996210098, + 0.005580936558544636, + 0.01022578589618206, + 0.010925118811428547, + 0.008336643688380718, + 0.011680589988827705, + 0.014631905592978, + 0.012193490751087666, + 0.014238235540688038, + 0.009534487500786781, + 0.015290912240743637, + 0.005610536318272352, + 0.004861959256231785, + 0.011370450258255005, + 0.0053690699860453606, + 0.00606162752956152, + 0.007301646750420332, + 0.004359383601695299, + 0.014656230807304382, + 0.008435790427029133, + 0.007351115345954895, + 0.010289783589541912, + 0.008187355473637581, + 0.007591051980853081, + 0.01661139726638794, + 0.009506670758128166, + 0.018861912190914154, + 0.00491366907954216, + 0.005210631061345339, + 0.00813719630241394, + 0.004202723503112793, + 0.004069013055413961, + 0.008496344089508057, + 0.0033525999169796705, + 0.00573243573307991, + 0.015779918059706688, + 0.010637739673256874, + 0.013506666757166386, + 0.010887835174798965, + 0.016109205782413483, + 0.0328911654651165, + 0.009541054256260395, + 0.02476980909705162, + 0.01503026857972145, + 0.005730268079787493, + 0.05187235400080681, + 0.009263441897928715, + 0.008112161420285702, + 0.010030528530478477, + 0.009020875208079815, + 0.04625830054283142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.006738557480275631 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182270228862762 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182270228862762, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182270228862762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08078383654356003, + 0.056759607046842575, + 0.05159740895032883, + 0.07336200028657913, + 0.07067963480949402, + 0.07401556521654129, + 0.31578248739242554, + 0.07475500553846359, + 0.1376413255929947, + 0.06462313234806061 + ], + "credit_map": [ + 0.005166210699826479, + 0.01361322682350874, + 0.009375778958201408, + 0.007796251680701971, + 0.009960884228348732, + 0.013024093583226204, + 0.007937042973935604, + 0.013910355977714062, + 0.008327999152243137, + 0.008117228746414185, + 0.00802676472812891, + 0.005901803728193045, + 0.006312891375273466, + 0.007952703163027763, + 0.005231849383562803, + 0.006888371426612139, + 0.008413602598011494, + 0.005514475516974926, + 0.006634873803704977, + 0.007846945896744728, + 0.004943203646689653, + 0.007418951019644737, + 0.004952462390065193, + 0.005872894544154406, + 0.009180091321468353, + 0.011836581863462925, + 0.008363169617950916, + 0.012361407279968262, + 0.006862502079457045, + 0.00928556453436613, + 0.006769359577447176, + 0.008703324943780899, + 0.010073614306747913, + 0.007711231242865324, + 0.009323881939053535, + 0.01040447223931551, + 0.009249399416148663, + 0.007000860758125782, + 0.007584601640701294, + 0.00933157280087471, + 0.008136868476867676, + 0.0058430517092347145, + 0.009980671107769012, + 0.011411485262215137, + 0.009086978621780872, + 0.009290067479014397, + 0.008311902172863483, + 0.011954544112086296, + 0.035040441900491714, + 0.026744037866592407, + 0.04761020839214325, + 0.02481258101761341, + 0.03315896913409233, + 0.05809431150555611, + 0.03379840776324272, + 0.056523509323596954, + 0.006836874410510063, + 0.006872006691992283, + 0.01197822391986847, + 0.006963903550058603, + 0.007196887396275997, + 0.012966422364115715, + 0.0061682648956775665, + 0.015772422775626183, + 0.015876593068242073, + 0.007799428887665272, + 0.01467349287122488, + 0.009629989042878151, + 0.017007142305374146, + 0.03545675799250603, + 0.009880460798740387, + 0.027317462489008904, + 0.006455547641962767, + 0.002129080006852746, + 0.022701090201735497, + 0.0039043775759637356, + 0.0030799044761806726, + 0.0039827111177146435, + 0.004546564072370529, + 0.01782386004924774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.018182270228862762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an extreme overhead angle, the camera tilts downward at breakneck speed, closely following a sleek, silver, metallic fighter jet as it violently tears through a vast, thick sea of white clouds. Almost instantly, it enters a perfect 90-degree vertical dive, plummeting downward at breathtaking speed. Its polished fuselage reflects a sharp gleam in the glaring sunlight at high altitude, its swept wings and single-seat bubble-shaped cockpit clearly visible. Due to the immense air resistance and hypersonic speed, the entire fuselage vibrates visibly, even subtle ripples and distortions appearing in the metal skin. A translucent sonic boom cloud forms around the wing edges and nose. The fighter jet leaves a swirling, chaotic vortex trail through the clouds, filling the entire scene with a sense of pure power, extreme speed, and awe-inspiring tension, creating an adrenaline-pumping visual impact.", + "seed": 31323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_1983_seed_31323_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.040092957019805905, + "vlm_logprob": -0.040092957019805905, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.040092957019805905, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09483129531145096 + ], + "opposite_logprob_tokens": [ + -2.8448312282562256 + ], + "logprob_reward": -0.09483129531145096, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09483129531145096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07662276178598404, + 0.043781448155641556, + 0.03290679305791855, + 0.037784144282341, + 0.02611784264445305, + 0.12339027971029282, + 0.14431312680244446, + 0.0802222415804863, + 0.1436084806919098, + 0.29125291109085083 + ], + "credit_map": [ + 0.010247990489006042, + 0.011415026150643826, + 0.006176698952913284, + 0.006318970583379269, + 0.011582435108721256, + 0.009829447604715824, + 0.0067266179248690605, + 0.014325572177767754, + 0.0031605770345777273, + 0.005251677706837654, + 0.003176524769514799, + 0.003946396987885237, + 0.008151416666805744, + 0.0064475503750145435, + 0.006054354831576347, + 0.0075929490849375725, + 0.0024586531799286604, + 0.0041000861674547195, + 0.005108227021992207, + 0.005510720890015364, + 0.0038444986566901207, + 0.004757077433168888, + 0.0026378119364380836, + 0.004489716142416, + 0.004433154594153166, + 0.005549085326492786, + 0.005059095099568367, + 0.0068423086777329445, + 0.0027750374283641577, + 0.006025209091603756, + 0.0026370473206043243, + 0.004463204648345709, + 0.002437620423734188, + 0.005565692204982042, + 0.005669443868100643, + 0.0020902841351926327, + 0.00182492402382195, + 0.002815172541886568, + 0.0021221130155026913, + 0.003592593129724264, + 0.011128879152238369, + 0.03612472862005234, + 0.0317469947040081, + 0.006888661999255419, + 0.012510167434811592, + 0.00896905642002821, + 0.007483516354113817, + 0.008538275957107544, + 0.008349796757102013, + 0.05213972553610802, + 0.04707794263958931, + 0.010499564930796623, + 0.008296669460833073, + 0.006205445621162653, + 0.005798302125185728, + 0.0059456718154251575, + 0.00864106509834528, + 0.011807695031166077, + 0.011579838581383228, + 0.010077995248138905, + 0.00893891230225563, + 0.010484134778380394, + 0.008204096928238869, + 0.010488500818610191, + 0.013268557377159595, + 0.026856228709220886, + 0.0365908220410347, + 0.014388756826519966, + 0.014214593917131424, + 0.012230029329657555, + 0.01151354145258665, + 0.014545944519340992, + 0.019879959523677826, + 0.09754816442728043, + 0.10261502116918564, + 0.021402521058917046, + 0.01397449616342783, + 0.009678817354142666, + 0.014361992478370667, + 0.011791964992880821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.09483129531145096 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097969900816679 + ], + "opposite_logprob_tokens": [ + -4.634797096252441 + ], + "logprob_reward": -0.0097969900816679, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097969900816679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11761774122714996, + 0.06613139063119888, + 0.05793341249227524, + 0.03601483628153801, + 0.05397219583392143, + 0.262069433927536, + 0.1955471783876419, + 0.06748130172491074, + 0.09518221765756607, + 0.048050276935100555 + ], + "credit_map": [ + 0.012601613998413086, + 0.014635666273534298, + 0.010607310570776463, + 0.01430692058056593, + 0.016281360760331154, + 0.019189056009054184, + 0.01430210005491972, + 0.015693724155426025, + 0.005999833345413208, + 0.007187327835708857, + 0.006968969013541937, + 0.005800826475024223, + 0.01014106348156929, + 0.009409444406628609, + 0.008826744742691517, + 0.011797185055911541, + 0.0037936880253255367, + 0.007786437403410673, + 0.006145182531327009, + 0.011496925726532936, + 0.008223007433116436, + 0.0075605628080666065, + 0.008043406531214714, + 0.004884204361587763, + 0.0025580311194062233, + 0.004275779705494642, + 0.005238113924860954, + 0.008165865205228329, + 0.003880774136632681, + 0.005403126124292612, + 0.00272247358225286, + 0.003770675975829363, + 0.0032605126034468412, + 0.013509837910532951, + 0.016789842396974564, + 0.004560872446745634, + 0.0037776241078972816, + 0.0037918423768132925, + 0.003834693692624569, + 0.004446974024176598, + 0.012621897272765636, + 0.10722088068723679, + 0.10593638569116592, + 0.00654755812138319, + 0.009376089088618755, + 0.00558547442778945, + 0.004103800747543573, + 0.010677330195903778, + 0.012220111675560474, + 0.0738743394613266, + 0.06469077616930008, + 0.009890401735901833, + 0.01175801269710064, + 0.007235126569867134, + 0.007410761900246143, + 0.008467649109661579, + 0.007302316837012768, + 0.011069310829043388, + 0.009991824626922607, + 0.008431512862443924, + 0.007646232843399048, + 0.008182989433407784, + 0.00629551662132144, + 0.008561599999666214, + 0.009631133638322353, + 0.016884075477719307, + 0.021147694438695908, + 0.008933177217841148, + 0.010123634710907936, + 0.008803194388747215, + 0.009402640163898468, + 0.01025667879730463, + 0.004556039813905954, + 0.010967091657221317, + 0.010307186283171177, + 0.004721914883702993, + 0.004760752897709608, + 0.0045127542689442635, + 0.0038591932971030474, + 0.004365349654108286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0097969900816679 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000859081745148 + ], + "opposite_logprob_tokens": [ + -2.6950085163116455 + ], + "logprob_reward": -0.07000859081745148, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000859081745148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025770893320441246, + 0.016605298966169357, + 0.05459482595324516, + 0.038069773465394974, + 0.018801923841238022, + 0.6438854336738586, + 0.07921848446130753, + 0.042066000401973724, + 0.03874116763472557, + 0.0422462522983551 + ], + "credit_map": [ + 0.0044049834832549095, + 0.0035154898650944233, + 0.001917930319905281, + 0.0026790043339133263, + 0.00329251098446548, + 0.0029234089888632298, + 0.002287872601300478, + 0.004749692510813475, + 0.001038901973515749, + 0.0020705759525299072, + 0.0012748310109600425, + 0.0014893168117851019, + 0.003245933912694454, + 0.0023325730580836535, + 0.0026419663336127996, + 0.0025112004950642586, + 0.005979993846267462, + 0.005705340765416622, + 0.008229495026171207, + 0.007667204365134239, + 0.005685893818736076, + 0.00827764067798853, + 0.00565821910277009, + 0.007391040679067373, + 0.0032551137264817953, + 0.0044861226342618465, + 0.004895307123661041, + 0.008926821872591972, + 0.0030699914786964655, + 0.005830319132655859, + 0.003305745543912053, + 0.004300350323319435, + 0.001256241463124752, + 0.0027914962265640497, + 0.0025504655204713345, + 0.002090287860482931, + 0.0022375676780939102, + 0.0027199150063097477, + 0.002253228100016713, + 0.0029027231503278017, + 0.029398620128631592, + 0.2680150866508484, + 0.2622911036014557, + 0.01714063063263893, + 0.020939379930496216, + 0.01306837983429432, + 0.0086408955976367, + 0.024391286075115204, + 0.006419893819838762, + 0.019951535388827324, + 0.01644952967762947, + 0.009595663286745548, + 0.00798787921667099, + 0.006272830534726381, + 0.006926504895091057, + 0.005614649038761854, + 0.004823506344109774, + 0.006537589710205793, + 0.00619736360386014, + 0.004177120979875326, + 0.004145459271967411, + 0.005697278305888176, + 0.0043467129580676556, + 0.006140969693660736, + 0.003374150488525629, + 0.007815971970558167, + 0.00978629570454359, + 0.003979241009801626, + 0.0027570039965212345, + 0.003922224976122379, + 0.003685906296595931, + 0.0034203743562102318, + 0.005101535934954882, + 0.006568625569343567, + 0.004530397709459066, + 0.006766836624592543, + 0.005880155134946108, + 0.0034551438875496387, + 0.005829975474625826, + 0.004113584291189909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.07000859081745148 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525260902941227 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.00525260902941227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525260902941227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1254294514656067, + 0.047439008951187134, + 0.03473491221666336, + 0.017853466793894768, + 0.025439875200390816, + 0.4625939428806305, + 0.11498437821865082, + 0.061008404940366745, + 0.07024337351322174, + 0.04027315601706505 + ], + "credit_map": [ + 0.01218333002179861, + 0.017772044986486435, + 0.011996417306363583, + 0.019190484657883644, + 0.013963953591883183, + 0.02003789134323597, + 0.015427525155246258, + 0.01485779881477356, + 0.004659849219024181, + 0.007307653781026602, + 0.004699806682765484, + 0.004658248741179705, + 0.007528284098953009, + 0.006913168355822563, + 0.0046266475692391396, + 0.007045350037515163, + 0.0036900935228914022, + 0.004231923259794712, + 0.004571591969579458, + 0.00610017403960228, + 0.0030248775146901608, + 0.0046914867125451565, + 0.004898895509541035, + 0.003525868058204651, + 0.0011295101139694452, + 0.0032250855583697557, + 0.002401123521849513, + 0.0030890910420566797, + 0.0019555161707103252, + 0.002580226631835103, + 0.0015568457311019301, + 0.0019160698866471648, + 0.0024571623653173447, + 0.003992094192653894, + 0.004019018728286028, + 0.0027284612879157066, + 0.0028267614543437958, + 0.0034176751505583525, + 0.0025822818279266357, + 0.0034164199605584145, + 0.02119404822587967, + 0.1930992603302002, + 0.18958680331707, + 0.012207156047224998, + 0.015020793303847313, + 0.008316120132803917, + 0.0056905001401901245, + 0.017479250207543373, + 0.0062375315465033054, + 0.04370225593447685, + 0.03977425768971443, + 0.006383555941283703, + 0.007292375899851322, + 0.004025984089821577, + 0.0035762169864028692, + 0.003992198966443539, + 0.005957324523478746, + 0.009606068022549152, + 0.011630520224571228, + 0.005790211725980043, + 0.005187577567994595, + 0.009681757539510727, + 0.005199280101805925, + 0.007955665700137615, + 0.005418009124696255, + 0.015675650909543037, + 0.01945725828409195, + 0.004621215630322695, + 0.007208637893199921, + 0.007208091206848621, + 0.0065845525823533535, + 0.004069961607456207, + 0.0032735562417656183, + 0.010059482418000698, + 0.011328848078846931, + 0.0036212019622325897, + 0.0028419955633580685, + 0.0028384036850184202, + 0.004310627467930317, + 0.0019990422297269106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.00525260902941227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575299859046936 + ], + "opposite_logprob_tokens": [ + -3.8955752849578857 + ], + "logprob_reward": -0.020575299859046936, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575299859046936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1825241595506668, + 0.04219392314553261, + 0.09419966489076614, + 0.08390495926141739, + 0.053327322006225586, + 0.050364088267087936, + 0.10679714381694794, + 0.0563771091401577, + 0.2659054696559906, + 0.06440618634223938 + ], + "credit_map": [ + 0.020961636677384377, + 0.02264329046010971, + 0.015916433185338974, + 0.024727676063776016, + 0.023843351751565933, + 0.026484686881303787, + 0.02055802196264267, + 0.027389056980609894, + 0.004122431855648756, + 0.006493816152215004, + 0.005002087447792292, + 0.004565312527120113, + 0.005474462639540434, + 0.005143838468939066, + 0.005897491239011288, + 0.005494485143572092, + 0.007429326418787241, + 0.012186340987682343, + 0.008594367653131485, + 0.016637122258543968, + 0.009294142946600914, + 0.013367857784032822, + 0.011972879059612751, + 0.014717625454068184, + 0.007984096184372902, + 0.015848301351070404, + 0.011478377506136894, + 0.012360622175037861, + 0.009937928058207035, + 0.010839474387466908, + 0.007925434969365597, + 0.007530715316534042, + 0.002650744281709194, + 0.013017005287110806, + 0.016043156385421753, + 0.004394423216581345, + 0.003443139372393489, + 0.004650895483791828, + 0.0035810882691293955, + 0.005546863656491041, + 0.004414082039147615, + 0.01165279746055603, + 0.011798780411481857, + 0.004109831992536783, + 0.004624807275831699, + 0.005282602272927761, + 0.004571373108774424, + 0.003909810446202755, + 0.011062520556151867, + 0.02734849601984024, + 0.027642281726002693, + 0.00976742897182703, + 0.011874453164637089, + 0.007078032940626144, + 0.004833596292883158, + 0.007190330885350704, + 0.006947049405425787, + 0.009573277086019516, + 0.009309323504567146, + 0.0057427166029810905, + 0.00570535147562623, + 0.005932909436523914, + 0.0050232368521392345, + 0.008143235929310322, + 0.02979518286883831, + 0.05417222902178764, + 0.07175259292125702, + 0.022336529567837715, + 0.025877071544528008, + 0.01947632245719433, + 0.018080664798617363, + 0.024414844810962677, + 0.007569876499474049, + 0.01523592695593834, + 0.018829144537448883, + 0.004777329042553902, + 0.0049710641615092754, + 0.004114522133022547, + 0.004854258615523577, + 0.0040540569461882114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.020575299859046936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from an extreme overhead angle, the camera tilts downward at breakneck speed, closely following a sleek, silver, metallic fighter jet as it violently tears through a vast, thick sea of white clouds. Almost instantly, it enters a perfect 90-degree vertical dive, plummeting downward at breathtaking speed. Its polished fuselage reflects a sharp gleam in the glaring sunlight at high altitude, its swept wings and single-seat bubble-shaped cockpit clearly visible. Due to the immense air resistance and hypersonic speed, the entire fuselage vibrates visibly, even subtle ripples and distortions appearing in the metal skin. A translucent sonic boom cloud forms around the wing edges and nose. The fighter jet leaves a swirling, chaotic vortex trail through the clouds, filling the entire scene with a sense of pure power, extreme speed, and awe-inspiring tension, creating an adrenaline-pumping visual impact.", + "seed": 31324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_1/index_1983_seed_31324_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03494314029812813, + "vlm_logprob": -0.03494314029812813, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03494314029812813, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08462461084127426 + ], + "opposite_logprob_tokens": [ + -2.8346245288848877 + ], + "logprob_reward": -0.08462461084127426, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08462461084127426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06947888433933258, + 0.03933059796690941, + 0.08380666375160217, + 0.2637028098106384, + 0.03771352395415306, + 0.03164876624941826, + 0.16943307220935822, + 0.05013040825724602, + 0.19287677109241486, + 0.06187855079770088 + ], + "credit_map": [ + 0.006441763136535883, + 0.010172451846301556, + 0.013530416414141655, + 0.007920745760202408, + 0.00548098050057888, + 0.013939624652266502, + 0.004661970771849155, + 0.007330931723117828, + 0.006858406588435173, + 0.006076025776565075, + 0.00511179631575942, + 0.00554330088198185, + 0.0030739442445337772, + 0.005102070979773998, + 0.0035291474778205156, + 0.004035907331854105, + 0.019085323438048363, + 0.005789916031062603, + 0.011610139161348343, + 0.012631379999220371, + 0.004978731274604797, + 0.011889971792697906, + 0.003999778069555759, + 0.013821427710354328, + 0.06017211079597473, + 0.016123492270708084, + 0.019880324602127075, + 0.03261328116059303, + 0.013848108239471912, + 0.028012223541736603, + 0.01223410852253437, + 0.08081915229558945, + 0.007717486470937729, + 0.0034255622886121273, + 0.0068688164465129375, + 0.005864537321031094, + 0.003158330684527755, + 0.0025759206619113684, + 0.0031080578919500113, + 0.0049948119558393955, + 0.0039080604910850525, + 0.005976132117211819, + 0.006184455007314682, + 0.004377004690468311, + 0.0018984857015311718, + 0.004578007850795984, + 0.001959545537829399, + 0.002767074154689908, + 0.008129692636430264, + 0.008912746794521809, + 0.06973765045404434, + 0.06421981751918793, + 0.002456045476719737, + 0.005572234746068716, + 0.004837889224290848, + 0.005567015614360571, + 0.0046316394582390785, + 0.008282506838440895, + 0.005813748110085726, + 0.008495033718645573, + 0.005341450683772564, + 0.005741783883422613, + 0.006184445694088936, + 0.005639798939228058, + 0.024531904608011246, + 0.029227517545223236, + 0.031246740370988846, + 0.028862226754426956, + 0.018695134669542313, + 0.023465631529688835, + 0.016571156680583954, + 0.020276449620723724, + 0.01594104804098606, + 0.005614256951957941, + 0.011206776835024357, + 0.007287175394594669, + 0.003746954258531332, + 0.005503037478774786, + 0.0042670420370996, + 0.008312256075441837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.08462461084127426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795809164643288 + ], + "opposite_logprob_tokens": [ + -4.634795665740967 + ], + "logprob_reward": -0.009795809164643288, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795809164643288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3537119925022125, + 0.09919371455907822, + 0.11084702610969543, + 0.13192680478096008, + 0.06380221992731094, + 0.03754296898841858, + 0.05293884128332138, + 0.03388364985585213, + 0.05225468799471855, + 0.06389807164669037 + ], + "credit_map": [ + 0.026413511484861374, + 0.023959817364811897, + 0.10607247799634933, + 0.030378255993127823, + 0.020692003890872, + 0.08880984783172607, + 0.020312616601586342, + 0.03707348182797432, + 0.010641497559845448, + 0.011266319081187248, + 0.014218774624168873, + 0.012101185508072376, + 0.014589733444154263, + 0.01155884750187397, + 0.01308436132967472, + 0.01173299178481102, + 0.017322644591331482, + 0.011363133788108826, + 0.01209795568138361, + 0.012159446254372597, + 0.01486374344676733, + 0.015406752936542034, + 0.014560872688889503, + 0.01307247206568718, + 0.01818438246846199, + 0.01834176480770111, + 0.01145130768418312, + 0.01640474423766136, + 0.012520198710262775, + 0.015092856250703335, + 0.01178810652345419, + 0.028143446892499924, + 0.008680070750415325, + 0.008028433658182621, + 0.00982384942471981, + 0.007083750795572996, + 0.008167210035026073, + 0.004207031801342964, + 0.007590562105178833, + 0.010221314616501331, + 0.00487329438328743, + 0.00537631381303072, + 0.00748197827488184, + 0.004739078693091869, + 0.0032205001916736364, + 0.005241934210062027, + 0.002933345502242446, + 0.0036765264812856913, + 0.0037068109959363937, + 0.004304348491132259, + 0.015038946643471718, + 0.016064390540122986, + 0.0033907818142324686, + 0.0022797302808612585, + 0.0038362976629287004, + 0.004317536484450102, + 0.0032701981253921986, + 0.0054251630790531635, + 0.004502007272094488, + 0.0046849665232002735, + 0.004253136459738016, + 0.002954408759251237, + 0.004510254133492708, + 0.004283517599105835, + 0.005446176510304213, + 0.0061223916709423065, + 0.008275923319160938, + 0.008347676135599613, + 0.00623372383415699, + 0.006386869121342897, + 0.006905683781951666, + 0.00453624501824379, + 0.01221957802772522, + 0.0065009333193302155, + 0.011329581029713154, + 0.009668266400694847, + 0.005192435812205076, + 0.007006841711699963, + 0.005453343968838453, + 0.006527092773467302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.009795809164643288 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864535480737686 + ], + "opposite_logprob_tokens": [ + -3.048645257949829 + ], + "logprob_reward": -0.04864535480737686, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864535480737686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1362563669681549, + 0.07825114578008652, + 0.04572812840342522, + 0.29221397638320923, + 0.01678905263543129, + 0.04898189380764961, + 0.21627143025398254, + 0.02314927615225315, + 0.047343116253614426, + 0.09501565992832184 + ], + "credit_map": [ + 0.011409481056034565, + 0.018052680417895317, + 0.03502608835697174, + 0.013364346697926521, + 0.007616369053721428, + 0.030669866129755974, + 0.007385424803942442, + 0.012732107192277908, + 0.010735026560723782, + 0.009136063978075981, + 0.007936222478747368, + 0.007857885211706161, + 0.013371473178267479, + 0.012787805870175362, + 0.010378913953900337, + 0.00604775408282876, + 0.007742313202470541, + 0.00305359810590744, + 0.005978339817374945, + 0.006216688081622124, + 0.003159203566610813, + 0.006969040725380182, + 0.003233399009332061, + 0.009375546127557755, + 0.06796927750110626, + 0.014419433660805225, + 0.022488344460725784, + 0.038907717913389206, + 0.017009558156132698, + 0.02770582027733326, + 0.011562694795429707, + 0.09215112775564194, + 0.0022523808293044567, + 0.0016293812077492476, + 0.0029049839358776808, + 0.002521179849281907, + 0.0021014532539993525, + 0.001201925566419959, + 0.0020538612734526396, + 0.00212388695217669, + 0.009002858772873878, + 0.010272583924233913, + 0.00555029883980751, + 0.00876681599766016, + 0.003954359795898199, + 0.0035059002693742514, + 0.0034659570083022118, + 0.004463118501007557, + 0.009255526587367058, + 0.012100915424525738, + 0.0875871554017067, + 0.07555042207241058, + 0.006760264281183481, + 0.006528991274535656, + 0.009619168937206268, + 0.00886897649616003, + 0.0022609480656683445, + 0.003202251624315977, + 0.002482163952663541, + 0.0026762273628264666, + 0.003450589720159769, + 0.0025092284195125103, + 0.0033980323933064938, + 0.003169833915308118, + 0.006382293533533812, + 0.005385842639952898, + 0.0068377056159079075, + 0.008205501362681389, + 0.006008628290146589, + 0.005501060280948877, + 0.005249366629868746, + 0.0037727162707597017, + 0.023817118257284164, + 0.010968951508402824, + 0.0167729202657938, + 0.010302060283720493, + 0.005919578950852156, + 0.009258881211280823, + 0.006225690245628357, + 0.011750464327633381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.04864535480737686 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074859648942947 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011074859648942947, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074859648942947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20415571331977844, + 0.16306038200855255, + 0.022423330694437027, + 0.08113579452037811, + 0.04466051235795021, + 0.07001674175262451, + 0.11726094037294388, + 0.019503153860569, + 0.04081706702709198, + 0.2369663566350937 + ], + "credit_map": [ + 0.015519384294748306, + 0.03216248005628586, + 0.05585285276174545, + 0.018228299915790558, + 0.007249502930790186, + 0.04699461907148361, + 0.008855154737830162, + 0.019293390214443207, + 0.017740540206432343, + 0.020438354462385178, + 0.014205769635736942, + 0.01663360372185707, + 0.038741953670978546, + 0.023034946992993355, + 0.02089383266866207, + 0.011371377855539322, + 0.0039806426502764225, + 0.0023148132022470236, + 0.002886519068852067, + 0.003966874442994595, + 0.0018639895133674145, + 0.0026784008368849754, + 0.0018394422950223088, + 0.0028926506638526917, + 0.013368334621191025, + 0.009755348786711693, + 0.007859084755182266, + 0.012422965839505196, + 0.006587210576981306, + 0.01299328077584505, + 0.004469616338610649, + 0.01367995422333479, + 0.00820431113243103, + 0.00448448583483696, + 0.009008863009512424, + 0.00723516708239913, + 0.003548126434907317, + 0.0027584691997617483, + 0.003468841779977083, + 0.005952251609414816, + 0.011908515356481075, + 0.0094922985881567, + 0.016119346022605896, + 0.01774056814610958, + 0.002764729782938957, + 0.004258856177330017, + 0.003298745024949312, + 0.004433685448020697, + 0.005185722839087248, + 0.006316730752587318, + 0.04877622425556183, + 0.045371416956186295, + 0.001973774516955018, + 0.003207172267138958, + 0.0035715969279408455, + 0.0028582937084138393, + 0.001823166967369616, + 0.002435508882626891, + 0.002024125773459673, + 0.0018954587867483497, + 0.004178440198302269, + 0.001658281427808106, + 0.003825279651209712, + 0.0016628917073830962, + 0.005794110242277384, + 0.005424031987786293, + 0.005308571271598339, + 0.005557973403483629, + 0.005307161249220371, + 0.004976578056812286, + 0.004372594878077507, + 0.004076046869158745, + 0.062217507511377335, + 0.0319230780005455, + 0.043635837733745575, + 0.028918001800775528, + 0.014455684460699558, + 0.019271666184067726, + 0.015717264264822006, + 0.020827319473028183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.011074859648942947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme overhead angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a sleek, silver, metallic fighter jet with swept wings and a bubble-shaped cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the jet performing a violent 90-degree vertical dive through a thick sea of white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects present, including a vibrating fuselage, a translucent sonic boom cloud, and a swirling vortex trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show the jet's polished fuselage reflecting a sharp gleam in glaring sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575067028403282 + ], + "opposite_logprob_tokens": [ + -3.8955750465393066 + ], + "logprob_reward": -0.020575067028403282, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575067028403282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16711214184761047, + 0.11878903955221176, + 0.05382974073290825, + 0.2501729428768158, + 0.031233955174684525, + 0.07590242475271225, + 0.08652950078248978, + 0.13866692781448364, + 0.028394494205713272, + 0.04936881735920906 + ], + "credit_map": [ + 0.015853160992264748, + 0.023295700550079346, + 0.03820488229393959, + 0.015286731533706188, + 0.00988146010786295, + 0.031915921717882156, + 0.011199276894330978, + 0.021475015208125114, + 0.013425633311271667, + 0.013118328526616096, + 0.014437207020819187, + 0.014280746690928936, + 0.02049209177494049, + 0.017538420855998993, + 0.013370969332754612, + 0.012125648558139801, + 0.0068228160962462425, + 0.005216862540692091, + 0.006729095708578825, + 0.008069688454270363, + 0.00411939388141036, + 0.008044958114624023, + 0.005260716192424297, + 0.009566210210323334, + 0.05261601135134697, + 0.028997138142585754, + 0.021139152348041534, + 0.034674156457185745, + 0.01358148641884327, + 0.021406197920441628, + 0.011949447914958, + 0.06580936163663864, + 0.006460684817284346, + 0.002961899619549513, + 0.005682827904820442, + 0.003857281757518649, + 0.0029126734007149935, + 0.0027356413193047047, + 0.002749518258497119, + 0.0038734287954866886, + 0.005992121063172817, + 0.007280085235834122, + 0.02871904894709587, + 0.019102830439805984, + 0.002237771637737751, + 0.006091195624321699, + 0.002543036127462983, + 0.003936336375772953, + 0.004297137260437012, + 0.004990920890122652, + 0.03490319103002548, + 0.031028321012854576, + 0.0023715775460004807, + 0.0027832151390612125, + 0.0030805557034909725, + 0.0030745859257876873, + 0.010358900763094425, + 0.021776540204882622, + 0.0276225246489048, + 0.03617841377854347, + 0.009594934992492199, + 0.013375054113566875, + 0.008810024708509445, + 0.010950556956231594, + 0.003986443858593702, + 0.0030649674590677023, + 0.004320203326642513, + 0.004383451770991087, + 0.004071402363479137, + 0.0027235401794314384, + 0.003686706768348813, + 0.002157780574634671, + 0.014837890863418579, + 0.005435111001133919, + 0.009217079728841782, + 0.005427035968750715, + 0.0026508416049182415, + 0.003521279664710164, + 0.003060225397348404, + 0.005219355691224337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.020575067028403282 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_tensor([6331, 6331])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_tensor([6331, 6331])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..69b58fd74cfcb4080c84908fb15ca0f6a02232df --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_tensor([6331, 6331])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shaky, handheld, medium shot at eye level immerses the viewer in an intimate and emotionally charged conversation. The camera's subtle, sustained movement focuses on a man in his 30s, his hair disheveled and shadowed at five o'clock, who speaks with passion. He wears a worn, dark blue denim jacket over a simple gray T-shirt, and sits leaning forward on a plush sofa. He gestures vigorously with his right hand, fingers spread and then clenched to punctuate his words. Across from him, a woman in her early 30s, her long, wavy brown hair partially obscuring her face, listens raptly. She wears a soft, cream-colored sweater. The scene is set in a cozy, dimly lit living room at night. A warm floor lamp casts a golden glow, creating soft shadows and highlighting the dancing dust particles in the air. When the man pauses, the woman slowly turns to meet his gaze, her expression softening with empathy, and she nods gently in deep agreement, a silent affirmation that resonates within the raw, authentic atmosphere of the moment.", + "seed": 7683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_6331_seed_7683_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5902728980407119, + "vlm_logprob": -0.5902728980407119, + "vlm_yes_no_margin": 1.7, + "vlm_token_credit": -0.5902728980407119, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.245088577270508 + ], + "opposite_logprob_tokens": [ + -0.12008848041296005 + ], + "logprob_reward": -2.245088577270508, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.245088577270508, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2708437144756317, + 0.18999165296554565, + 0.06566981971263885, + 0.048430878669023514, + 0.19118255376815796, + 0.07439550757408142, + 0.043168507516384125, + 0.047055501490831375, + 0.02546738274395466, + 0.04379444941878319 + ], + "credit_map": [ + 0.020517630502581596, + 0.05451894551515579, + 0.037265803664922714, + 0.04291141405701637, + 0.025698872283101082, + 0.03154268488287926, + 0.03591962531208992, + 0.022468743845820427, + 0.04166065901517868, + 0.020471153780817986, + 0.0265178345143795, + 0.027298446744680405, + 0.015065200626850128, + 0.02458581142127514, + 0.017556659877300262, + 0.016835875809192657, + 0.008188368752598763, + 0.0074390992522239685, + 0.011044745333492756, + 0.010112144984304905, + 0.005455882754176855, + 0.009539729915559292, + 0.008393965661525726, + 0.005495881661772728, + 0.00458164606243372, + 0.004536185879260302, + 0.009330298751592636, + 0.008277171291410923, + 0.004668541252613068, + 0.007295156363397837, + 0.005287643522024155, + 0.004454237408936024, + 0.015103884041309357, + 0.03088812530040741, + 0.02236109972000122, + 0.03389163315296173, + 0.017738910391926765, + 0.032077252864837646, + 0.02113904245197773, + 0.017982598394155502, + 0.013955275528132915, + 0.009719685651361942, + 0.01107845176011324, + 0.007673249579966068, + 0.006206293124705553, + 0.01142122596502304, + 0.007683384232223034, + 0.006657945457845926, + 0.0033976298291236162, + 0.007996316999197006, + 0.004408313427120447, + 0.007819210179150105, + 0.0033913464285433292, + 0.005601137410849333, + 0.004573826212435961, + 0.005980730056762695, + 0.006819466128945351, + 0.006785218603909016, + 0.00670348946005106, + 0.007585130166262388, + 0.004938195459544659, + 0.005331787746399641, + 0.003982242662459612, + 0.004909968469291925, + 0.0027850279584527016, + 0.004557686857879162, + 0.004119685851037502, + 0.004005816299468279, + 0.002682364545762539, + 0.002931042341515422, + 0.0021305924747139215, + 0.002255166880786419, + 0.003684606868773699, + 0.005049805156886578, + 0.0046104504726827145, + 0.005965941585600376, + 0.0031116323079913855, + 0.007617667317390442, + 0.010084593668580055, + 0.003669752273708582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -2.245088577270508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029866356402635574 + ], + "opposite_logprob_tokens": [ + -3.5298664569854736 + ], + "logprob_reward": -0.029866356402635574, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029866356402635574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1844923347234726, + 0.1908850222826004, + 0.06663960963487625, + 0.04881260171532631, + 0.16562004387378693, + 0.08433952182531357, + 0.07380315661430359, + 0.06359022110700607, + 0.03149326145648956, + 0.09032423049211502 + ], + "credit_map": [ + 0.015356979332864285, + 0.027553444728255272, + 0.020404592156410217, + 0.03652595356106758, + 0.019593916833400726, + 0.02059119939804077, + 0.023226594552397728, + 0.02123964950442314, + 0.04292448237538338, + 0.021676775068044662, + 0.02467990852892399, + 0.031600553542375565, + 0.012955586425960064, + 0.023410165682435036, + 0.017309196293354034, + 0.016328347846865654, + 0.00766502320766449, + 0.011109731160104275, + 0.008379523642361164, + 0.01077787671238184, + 0.006130215711891651, + 0.008980262093245983, + 0.007242038380354643, + 0.006354943383485079, + 0.006953911390155554, + 0.006782888900488615, + 0.0060434346087276936, + 0.006024765316396952, + 0.00622025178745389, + 0.0068639046512544155, + 0.004907199647277594, + 0.0050162458792328835, + 0.016857823356986046, + 0.027321314439177513, + 0.019407765939831734, + 0.025915466248989105, + 0.016685716807842255, + 0.023684844374656677, + 0.019147995859384537, + 0.016599124297499657, + 0.015555012039840221, + 0.010876476764678955, + 0.008693440817296505, + 0.010389679111540318, + 0.008660217747092247, + 0.012716257944703102, + 0.0075445170514285564, + 0.009903921745717525, + 0.00920709315687418, + 0.012058382853865623, + 0.008873729035258293, + 0.009622550569474697, + 0.007191285490989685, + 0.0100289611145854, + 0.005972640123218298, + 0.010848515667021275, + 0.008208146318793297, + 0.009836218319833279, + 0.010229040868580341, + 0.01085282489657402, + 0.005539822392165661, + 0.00695411441847682, + 0.007307115942239761, + 0.004662935622036457, + 0.0038563741836696863, + 0.005576661322265863, + 0.003588925115764141, + 0.005558561999350786, + 0.0022768552880734205, + 0.004320092964917421, + 0.003581276861950755, + 0.0027345139533281326, + 0.008527054451406002, + 0.013081755489110947, + 0.009061172604560852, + 0.010909450240433216, + 0.00720909982919693, + 0.021786268800497055, + 0.01207540649920702, + 0.007674016524106264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.029866356402635574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08910118788480759 + ], + "opposite_logprob_tokens": [ + -2.4641010761260986 + ], + "logprob_reward": -0.08910118788480759, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08910118788480759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20466816425323486, + 0.24553941190242767, + 0.10853365808725357, + 0.05405517295002937, + 0.09530157595872879, + 0.06350712478160858, + 0.09629344940185547, + 0.07266093045473099, + 0.031264349818229675, + 0.028176166117191315 + ], + "credit_map": [ + 0.029046716168522835, + 0.03608481585979462, + 0.021722570061683655, + 0.03317520394921303, + 0.0188639834523201, + 0.022936269640922546, + 0.022439686581492424, + 0.020398933440446854, + 0.05203551799058914, + 0.02695327438414097, + 0.04018017649650574, + 0.038496192544698715, + 0.01578763872385025, + 0.031236780807375908, + 0.021722257137298584, + 0.019127564504742622, + 0.010038474574685097, + 0.010745099745690823, + 0.01673317328095436, + 0.027111288160085678, + 0.007440178655087948, + 0.015369839034974575, + 0.012282009236514568, + 0.008813601918518543, + 0.005774770863354206, + 0.007065195124596357, + 0.005524150561541319, + 0.011719424277544022, + 0.005524256266653538, + 0.008138969540596008, + 0.0051596579141914845, + 0.005148753989487886, + 0.016336137428879738, + 0.013432246632874012, + 0.009100942872464657, + 0.014368993230164051, + 0.0097879059612751, + 0.014870219863951206, + 0.008719085715711117, + 0.008686047978699207, + 0.009905964136123657, + 0.012311440892517567, + 0.006595052313059568, + 0.0075754765421152115, + 0.006003332324326038, + 0.009232324548065662, + 0.00499678635969758, + 0.006886749528348446, + 0.008541884832084179, + 0.01695612445473671, + 0.010883131995797157, + 0.020726175978779793, + 0.008142962120473385, + 0.012880814261734486, + 0.010458074510097504, + 0.0077042910270392895, + 0.01591821201145649, + 0.010165772400796413, + 0.008163000456988811, + 0.010079238563776016, + 0.0049051702953875065, + 0.01123115699738264, + 0.006217732559889555, + 0.005980649031698704, + 0.004133225884288549, + 0.004583385307341814, + 0.004150298424065113, + 0.005286411382257938, + 0.0019771328661590815, + 0.005931849591434002, + 0.0027709216810762882, + 0.002431125147268176, + 0.002675327006727457, + 0.0035706060007214546, + 0.0037108859978616238, + 0.004335874691605568, + 0.0026839885395020247, + 0.00420647207647562, + 0.0038049144204705954, + 0.003188095521181822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.08910118788480759 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107992883771658 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.01107992883771658, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107992883771658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24647021293640137, + 0.18963639438152313, + 0.04072779044508934, + 0.03910980746150017, + 0.13571052253246307, + 0.09815884381532669, + 0.14610768854618073, + 0.042770467698574066, + 0.026734862476587296, + 0.03457341715693474 + ], + "credit_map": [ + 0.0405474528670311, + 0.034275565296411514, + 0.03125501051545143, + 0.036454856395721436, + 0.02266274392604828, + 0.026548318564891815, + 0.02684662491083145, + 0.02787965163588524, + 0.04018428549170494, + 0.020948005840182304, + 0.028419148176908493, + 0.029481304809451103, + 0.013732134364545345, + 0.02329033613204956, + 0.017528630793094635, + 0.016052549704909325, + 0.004692130256444216, + 0.004771948792040348, + 0.006693327333778143, + 0.005488808266818523, + 0.0039761182852089405, + 0.006105469074100256, + 0.0053833844140172005, + 0.0036166023928672075, + 0.0050187972374260426, + 0.0049348813481628895, + 0.004716817289590836, + 0.005806939210742712, + 0.003292533103376627, + 0.005680134519934654, + 0.004172638989984989, + 0.005487062968313694, + 0.027601929381489754, + 0.016968542709946632, + 0.015782231464982033, + 0.021391551941633224, + 0.010944141075015068, + 0.0199391171336174, + 0.011670052073895931, + 0.011412948369979858, + 0.024534011259675026, + 0.011707739904522896, + 0.013493629172444344, + 0.00981936790049076, + 0.008557992987334728, + 0.01390813011676073, + 0.007463445886969566, + 0.008674530312418938, + 0.03237343207001686, + 0.017844438552856445, + 0.01816457323729992, + 0.019407719373703003, + 0.014522907324135303, + 0.020252631977200508, + 0.012443854473531246, + 0.011098123155534267, + 0.010323719121515751, + 0.005581118632107973, + 0.0056983535178005695, + 0.005630158819258213, + 0.0035950627643615007, + 0.005245383828878403, + 0.0038985467981547117, + 0.002798123052343726, + 0.0051916795782744884, + 0.0040970537811517715, + 0.003980581182986498, + 0.0040122345089912415, + 0.002623053966090083, + 0.002235428197309375, + 0.002565550385043025, + 0.0020292820408940315, + 0.004532849881798029, + 0.003659385722130537, + 0.0037062254268676043, + 0.002816068008542061, + 0.0024777366779744625, + 0.006907796021550894, + 0.007784651126712561, + 0.0026887021958827972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01107992883771658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762284398078918 + ], + "opposite_logprob_tokens": [ + -0.8262284398078918 + ], + "logprob_reward": -0.5762284398078918, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762284398078918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35121217370033264, + 0.09814664721488953, + 0.07006178796291351, + 0.05261201784014702, + 0.1728912740945816, + 0.05373946577310562, + 0.0500338077545166, + 0.06523259729146957, + 0.05339387431740761, + 0.03267642855644226 + ], + "credit_map": [ + 0.04196164384484291, + 0.0482817180454731, + 0.040064483880996704, + 0.07194554060697556, + 0.03589978069067001, + 0.039959438145160675, + 0.04049815237522125, + 0.03260140120983124, + 0.014265822246670723, + 0.01060387585312128, + 0.01975916139781475, + 0.020278552547097206, + 0.0067201354540884495, + 0.012400444597005844, + 0.005658913403749466, + 0.008459738455712795, + 0.008992929942905903, + 0.008051643148064613, + 0.010122097097337246, + 0.01024599839001894, + 0.006514356471598148, + 0.008616638369858265, + 0.008107740432024002, + 0.009410381317138672, + 0.004896208178251982, + 0.006918127648532391, + 0.008053130470216274, + 0.009320062585175037, + 0.005402577109634876, + 0.0067448243498802185, + 0.005924579221755266, + 0.005352506414055824, + 0.013500653207302094, + 0.021194852888584137, + 0.031209222972393036, + 0.029319925233721733, + 0.016611170023679733, + 0.02955489233136177, + 0.017886033281683922, + 0.01361451018601656, + 0.006816198118031025, + 0.00783323124051094, + 0.006504482124000788, + 0.0088502811267972, + 0.004967347718775272, + 0.007791541516780853, + 0.005877643823623657, + 0.005098738707602024, + 0.005310689564794302, + 0.008591602556407452, + 0.00491652637720108, + 0.008236554451286793, + 0.004177585709840059, + 0.006687619257718325, + 0.004638382233679295, + 0.007474846672266722, + 0.007057255133986473, + 0.010603785514831543, + 0.009278585202991962, + 0.010034448467195034, + 0.006837464869022369, + 0.006632786709815264, + 0.005657879635691643, + 0.009130386635661125, + 0.0070418547838926315, + 0.008434513583779335, + 0.00920670386403799, + 0.00843726098537445, + 0.005367586389183998, + 0.005775214172899723, + 0.004793285857886076, + 0.0043374584056437016, + 0.0019200267270207405, + 0.00402927678078413, + 0.003753998316824436, + 0.004128123167902231, + 0.0027694955933839083, + 0.0060661365278065205, + 0.007384196389466524, + 0.0026251724921166897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.5762284398078918 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shaky, handheld, medium shot at eye level immerses the viewer in an intimate and emotionally charged conversation. The camera's subtle, sustained movement focuses on a man in his 30s, his hair disheveled and shadowed at five o'clock, who speaks with passion. He wears a worn, dark blue denim jacket over a simple gray T-shirt, and sits leaning forward on a plush sofa. He gestures vigorously with his right hand, fingers spread and then clenched to punctuate his words. Across from him, a woman in her early 30s, her long, wavy brown hair partially obscuring her face, listens raptly. She wears a soft, cream-colored sweater. The scene is set in a cozy, dimly lit living room at night. A warm floor lamp casts a golden glow, creating soft shadows and highlighting the dancing dust particles in the air. When the man pauses, the woman slowly turns to meet his gaze, her expression softening with empathy, and she nods gently in deep agreement, a silent affirmation that resonates within the raw, authentic atmosphere of the moment.", + "seed": 7684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_6331_seed_7684_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5195743177086115, + "vlm_logprob": -0.5195743177086115, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.5195743177086115, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.2442500591278076 + ], + "opposite_logprob_tokens": [ + -0.11925008893013 + ], + "logprob_reward": -2.2442500591278076, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.2442500591278076, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16184598207473755, + 0.06801071763038635, + 0.04846959188580513, + 0.05739445611834526, + 0.06352339684963226, + 0.043193817138671875, + 0.30403658747673035, + 0.16234903037548065, + 0.047613129019737244, + 0.04356333985924721 + ], + "credit_map": [ + 0.019366994500160217, + 0.022933028638362885, + 0.028361117467284203, + 0.022461790591478348, + 0.016139622777700424, + 0.013672946020960808, + 0.026068761944770813, + 0.012841718271374702, + 0.01079154945909977, + 0.010920216329395771, + 0.005337527021765709, + 0.010285964235663414, + 0.00555159617215395, + 0.008677664212882519, + 0.01079370267689228, + 0.0056524937972426414, + 0.004163539502769709, + 0.0072073559276759624, + 0.003338045906275511, + 0.009414971806108952, + 0.006044401787221432, + 0.004749609623104334, + 0.008243171498179436, + 0.005308493506163359, + 0.007462324108928442, + 0.01079633180052042, + 0.004865400027483702, + 0.011917110532522202, + 0.00559550104662776, + 0.0036778163630515337, + 0.009860716760158539, + 0.003219256177544594, + 0.006718604825437069, + 0.007760710548609495, + 0.007741154171526432, + 0.004860507324337959, + 0.00580050703138113, + 0.009184683673083782, + 0.01214655488729477, + 0.009310671128332615, + 0.005438964813947678, + 0.004323462024331093, + 0.0035688679199665785, + 0.0039349570870399475, + 0.0031470824033021927, + 0.008439328521490097, + 0.009611997753381729, + 0.004729155451059341, + 0.016862861812114716, + 0.014077290892601013, + 0.007188630755990744, + 0.01838875003159046, + 0.015773558989167213, + 0.11608296632766724, + 0.0974445790052414, + 0.01821795664727688, + 0.01169312559068203, + 0.01676146686077118, + 0.011908425949513912, + 0.030759664252400398, + 0.010968255810439587, + 0.021097565069794655, + 0.034428082406520844, + 0.024732446298003197, + 0.0044391038827598095, + 0.008525675162672997, + 0.0046859984286129475, + 0.005835899151861668, + 0.003686152631416917, + 0.006403734441846609, + 0.006964582018554211, + 0.007071982603520155, + 0.005682921502739191, + 0.0075666760094463825, + 0.004274402279406786, + 0.005222978536039591, + 0.0041841911152005196, + 0.004657920449972153, + 0.006099773105233908, + 0.00587447639554739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -2.2442500591278076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016127724200487137 + ], + "opposite_logprob_tokens": [ + -4.141127586364746 + ], + "logprob_reward": -0.016127724200487137, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016127724200487137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04439413547515869, + 0.025049366056919098, + 0.020175913348793983, + 0.018587134778499603, + 0.022096648812294006, + 0.022326888516545296, + 0.16170738637447357, + 0.636278510093689, + 0.021635184064507484, + 0.027748916298151016 + ], + "credit_map": [ + 0.0071440190076828, + 0.005370988976210356, + 0.00706267636269331, + 0.0067529440857470036, + 0.00444497587159276, + 0.0036770363803952932, + 0.006474101450294256, + 0.0034673907794058323, + 0.0038530132733285427, + 0.004074323922395706, + 0.0018069669604301453, + 0.003770582377910614, + 0.002244045725092292, + 0.0024597099982202053, + 0.005088147707283497, + 0.0017525751609355211, + 0.0032287121284753084, + 0.0026025688275694847, + 0.0017642953898757696, + 0.003834929782897234, + 0.0013611268950626254, + 0.0022656002547591925, + 0.0030622619669884443, + 0.0020564154256135225, + 0.0026048847939819098, + 0.0027197888121008873, + 0.0018690655706450343, + 0.004014894366264343, + 0.0014102010754868388, + 0.0012717500794678926, + 0.0035066897980868816, + 0.00118985865265131, + 0.003993042279034853, + 0.0027722278609871864, + 0.0026612032670527697, + 0.0015139610040932894, + 0.002137420466169715, + 0.002740107476711273, + 0.004206533543765545, + 0.00207214942201972, + 0.0026961402036249638, + 0.0021702104713767767, + 0.0015347417211160064, + 0.0029952742625027895, + 0.001491585629992187, + 0.003513864241540432, + 0.004900479689240456, + 0.0030245890375226736, + 0.008983497507870197, + 0.006911946460604668, + 0.003249804489314556, + 0.010507753118872643, + 0.008461592718958855, + 0.059418488293886185, + 0.05432279035449028, + 0.0098514873534441, + 0.03861137107014656, + 0.048548780381679535, + 0.054778456687927246, + 0.09276801347732544, + 0.04458397254347801, + 0.10049454122781754, + 0.12578527629375458, + 0.13070803880691528, + 0.00327692786231637, + 0.003380158916115761, + 0.0014769628178328276, + 0.0015176462475210428, + 0.001733402255922556, + 0.003528893692418933, + 0.003872427623718977, + 0.002848762320354581, + 0.005312944762408733, + 0.004194847773760557, + 0.0028019777964800596, + 0.0024705848190933466, + 0.002930133603513241, + 0.003023824654519558, + 0.003937592729926109, + 0.003077007597312331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.016127724200487137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194829389452934 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014194829389452934, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194829389452934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1178542748093605, + 0.053724419325590134, + 0.04971940070390701, + 0.03183496370911598, + 0.03824169561266899, + 0.04654765874147415, + 0.13745762407779694, + 0.33293142914772034, + 0.08107815682888031, + 0.11061037331819534 + ], + "credit_map": [ + 0.017784055322408676, + 0.013024498708546162, + 0.015915386378765106, + 0.016268564388155937, + 0.016039418056607246, + 0.012913385406136513, + 0.013623627834022045, + 0.012285348027944565, + 0.008323940448462963, + 0.006673203781247139, + 0.005835123360157013, + 0.0067220120690763, + 0.004624004941433668, + 0.007708907593041658, + 0.008073712699115276, + 0.005763517692685127, + 0.007770514581352472, + 0.008785872720181942, + 0.006555802188813686, + 0.005672353319823742, + 0.00410030921921134, + 0.00631494726985693, + 0.007035443559288979, + 0.003484159242361784, + 0.004753744229674339, + 0.0028943666256964207, + 0.0049260761588811874, + 0.005022105760872364, + 0.002975360956043005, + 0.0031340476125478745, + 0.005365082994103432, + 0.002764184260740876, + 0.004310102667659521, + 0.003437594510614872, + 0.0035403738729655743, + 0.004621031694114208, + 0.002534926403313875, + 0.006047866772860289, + 0.005880200769752264, + 0.007869603112339973, + 0.005270345136523247, + 0.0059321122244000435, + 0.002672500442713499, + 0.004908400122076273, + 0.0030697679612785578, + 0.008353120647370815, + 0.008113681338727474, + 0.00822773389518261, + 0.008383773267269135, + 0.0072798775508999825, + 0.0036621796898543835, + 0.009255915880203247, + 0.0071069467812776566, + 0.047810692340135574, + 0.045404840260744095, + 0.008553408086299896, + 0.018606163561344147, + 0.025500331073999405, + 0.03749172016978264, + 0.04448626562952995, + 0.02393144927918911, + 0.053731050342321396, + 0.06040246784687042, + 0.0687820166349411, + 0.011266806162893772, + 0.01035046111792326, + 0.004336918704211712, + 0.006625500041991472, + 0.008847778663039207, + 0.01773921214044094, + 0.011094397865235806, + 0.010817090049386024, + 0.009775321930646896, + 0.013320699334144592, + 0.008477794006466866, + 0.009905662387609482, + 0.010885725729167461, + 0.02606763131916523, + 0.01386782992631197, + 0.018309716135263443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.014194829389452934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097876638174057 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097876638174057, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097876638174057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14428870379924774, + 0.06319672614336014, + 0.03793179243803024, + 0.05506739020347595, + 0.09644164890050888, + 0.044439949095249176, + 0.271237850189209, + 0.19914625585079193, + 0.0486733540892601, + 0.03957640379667282 + ], + "credit_map": [ + 0.018489938229322433, + 0.02082560770213604, + 0.020261237397789955, + 0.020463382825255394, + 0.013312257826328278, + 0.014939732849597931, + 0.023225300014019012, + 0.012771235778927803, + 0.008398220874369144, + 0.007629385218024254, + 0.005418381653726101, + 0.007774569559842348, + 0.006467863917350769, + 0.008626524358987808, + 0.012553432025015354, + 0.006328346207737923, + 0.0047148666344583035, + 0.003356666537001729, + 0.003360976930707693, + 0.005822425242513418, + 0.0029517386574298143, + 0.006716284900903702, + 0.006166200153529644, + 0.004842632450163364, + 0.00742637412622571, + 0.005947915371507406, + 0.008877938613295555, + 0.007422396447509527, + 0.004600579384714365, + 0.007471691817045212, + 0.008364567533135414, + 0.004955926910042763, + 0.013716758228838444, + 0.013817104510962963, + 0.014200303703546524, + 0.007825281471014023, + 0.00790635496377945, + 0.009994138032197952, + 0.0160814356058836, + 0.012900283560156822, + 0.007091531530022621, + 0.005741233471781015, + 0.005770925432443619, + 0.005573991686105728, + 0.0047853547148406506, + 0.003089696168899536, + 0.006354179698973894, + 0.006033037323504686, + 0.0232902392745018, + 0.016635797917842865, + 0.010966252535581589, + 0.017492033541202545, + 0.014548012986779213, + 0.09702996909618378, + 0.07188857346773148, + 0.019386980682611465, + 0.009384705685079098, + 0.017927294597029686, + 0.012060013599693775, + 0.033400602638721466, + 0.01474131178110838, + 0.031381428241729736, + 0.04278196021914482, + 0.03746892884373665, + 0.00476553151383996, + 0.008591602556407452, + 0.0039201583713293076, + 0.005394397769123316, + 0.003629256971180439, + 0.007317785639315844, + 0.007662136573344469, + 0.007392482832074165, + 0.005038582254201174, + 0.006675928831100464, + 0.00252901972271502, + 0.005220232531428337, + 0.004227287136018276, + 0.005666176788508892, + 0.005834860261529684, + 0.00438431603834033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0097876638174057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31351131200790405 + ], + "opposite_logprob_tokens": [ + -1.3135113716125488 + ], + "logprob_reward": -0.31351131200790405, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31351131200790405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06952980160713196, + 0.07869114726781845, + 0.055569782853126526, + 0.10608121752738953, + 0.23685875535011292, + 0.03421928733587265, + 0.18130770325660706, + 0.044661425054073334, + 0.09323979914188385, + 0.09984111785888672 + ], + "credit_map": [ + 0.009500584565103054, + 0.011868036352097988, + 0.009203641675412655, + 0.011208157055079937, + 0.006503711920231581, + 0.004765983205288649, + 0.012329012155532837, + 0.004150671884417534, + 0.013772934675216675, + 0.010004367679357529, + 0.009890448302030563, + 0.013842005282640457, + 0.006544017232954502, + 0.006021852605044842, + 0.01374003104865551, + 0.004875499755144119, + 0.0080413818359375, + 0.007673133164644241, + 0.0043052067048847675, + 0.008968798443675041, + 0.005615744739770889, + 0.005133586935698986, + 0.008558970876038074, + 0.007272960152477026, + 0.011149764992296696, + 0.014975854195654392, + 0.016865715384483337, + 0.019015096127986908, + 0.008314681239426136, + 0.009771564975380898, + 0.0186014287173748, + 0.007387113757431507, + 0.03316885232925415, + 0.030416833236813545, + 0.02774227038025856, + 0.018719518557190895, + 0.02306046150624752, + 0.03484703600406647, + 0.0405268594622612, + 0.02837691642343998, + 0.004505346994847059, + 0.004119609482586384, + 0.0027501117438077927, + 0.005107697565108538, + 0.0029166974127292633, + 0.004751750733703375, + 0.006118467543274164, + 0.003949607256799936, + 0.010412303730845451, + 0.010017840191721916, + 0.006805967073887587, + 0.012721152044832706, + 0.0105778519064188, + 0.0654246136546135, + 0.05508916452527046, + 0.010258805006742477, + 0.003931849263608456, + 0.0054944963194429874, + 0.002678387565538287, + 0.006646616850048304, + 0.004139865282922983, + 0.008898102678358555, + 0.006774149835109711, + 0.006097958888858557, + 0.012327157892286777, + 0.010698482394218445, + 0.007054443471133709, + 0.012682177126407623, + 0.0061088004149496555, + 0.014531189575791359, + 0.019173381850123405, + 0.010664168745279312, + 0.0157626960426569, + 0.015089688822627068, + 0.010399083606898785, + 0.014375630766153336, + 0.0059995949268341064, + 0.010813500732183456, + 0.016859373077750206, + 0.010541560128331184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.31351131200790405 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shaky, handheld, medium shot at eye level immerses the viewer in an intimate and emotionally charged conversation. The camera's subtle, sustained movement focuses on a man in his 30s, his hair disheveled and shadowed at five o'clock, who speaks with passion. He wears a worn, dark blue denim jacket over a simple gray T-shirt, and sits leaning forward on a plush sofa. He gestures vigorously with his right hand, fingers spread and then clenched to punctuate his words. Across from him, a woman in her early 30s, her long, wavy brown hair partially obscuring her face, listens raptly. She wears a soft, cream-colored sweater. The scene is set in a cozy, dimly lit living room at night. A warm floor lamp casts a golden glow, creating soft shadows and highlighting the dancing dust particles in the air. When the man pauses, the woman slowly turns to meet his gaze, her expression softening with empathy, and she nods gently in deep agreement, a silent affirmation that resonates within the raw, authentic atmosphere of the moment.", + "seed": 7685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_6331_seed_7685_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.5736078944057226, + "vlm_logprob": -0.5736078944057226, + "vlm_yes_no_margin": 1.825000023841858, + "vlm_token_credit": -0.5736078944057226, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.03139066696167 + ], + "opposite_logprob_tokens": [ + -0.15639077126979828 + ], + "logprob_reward": -2.03139066696167, + "yes_no_margin_reward": -1.8749998807907104, + "token_credit_reward": -2.03139066696167, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2778438627719879, + 0.08442864567041397, + 0.055532585829496384, + 0.12991511821746826, + 0.09162271022796631, + 0.08925209194421768, + 0.08023946732282639, + 0.08630891889333725, + 0.06509817391633987, + 0.03975841403007507 + ], + "credit_map": [ + 0.042967602610588074, + 0.0405854657292366, + 0.0367516428232193, + 0.060428790748119354, + 0.025480836629867554, + 0.019675301387906075, + 0.03254053369164467, + 0.01941370777785778, + 0.007846405729651451, + 0.015083088539540768, + 0.013847755268216133, + 0.012766409665346146, + 0.00944250263273716, + 0.006629366893321276, + 0.010462281294167042, + 0.008350839838385582, + 0.006921092513948679, + 0.006960953585803509, + 0.007078786380589008, + 0.009216869249939919, + 0.0056008328683674335, + 0.006160452496260405, + 0.0060601262375712395, + 0.007533471565693617, + 0.019770348444581032, + 0.016140414401888847, + 0.032298170030117035, + 0.02320769801735878, + 0.007912689819931984, + 0.006955564953386784, + 0.013192970305681229, + 0.010437259450554848, + 0.010392237454652786, + 0.01563171111047268, + 0.012195113115012646, + 0.016004450619220734, + 0.009241621010005474, + 0.010275883600115776, + 0.00913002248853445, + 0.008751672692596912, + 0.011478863656520844, + 0.015020913444459438, + 0.011977652087807655, + 0.016202449798583984, + 0.008951324969530106, + 0.007291843648999929, + 0.0100266607478261, + 0.008302384056150913, + 0.008874599821865559, + 0.01575659215450287, + 0.01121712289750576, + 0.011948158964514732, + 0.008911576122045517, + 0.006168772000819445, + 0.008132731541991234, + 0.009229909628629684, + 0.010896866209805012, + 0.014386323280632496, + 0.014050985686480999, + 0.015444064512848854, + 0.0074898372404277325, + 0.006202727556228638, + 0.011022872291505337, + 0.006815234664827585, + 0.008542552590370178, + 0.008050104603171349, + 0.009249601513147354, + 0.009858282282948494, + 0.0049495818093419075, + 0.005894227419048548, + 0.010580958798527718, + 0.007972867228090763, + 0.006341961678117514, + 0.0054759434424340725, + 0.007224365137517452, + 0.006322755943983793, + 0.003415795974433422, + 0.0026980990078300238, + 0.004752048756927252, + 0.00352744129486382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -2.03139066696167 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026410866528749466 + ], + "opposite_logprob_tokens": [ + -3.6514108180999756 + ], + "logprob_reward": -0.026410866528749466, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026410866528749466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3718566298484802, + 0.05053156986832619, + 0.07662606984376907, + 0.21166104078292847, + 0.020245306193828583, + 0.019817687571048737, + 0.035959117114543915, + 0.06699685007333755, + 0.06190945953130722, + 0.08439631015062332 + ], + "credit_map": [ + 0.03572024032473564, + 0.04538371413946152, + 0.050477996468544006, + 0.08927115797996521, + 0.037709008902311325, + 0.02654828317463398, + 0.06082012131810188, + 0.025926072150468826, + 0.005084055941551924, + 0.00831518042832613, + 0.006467550527304411, + 0.007298022508621216, + 0.00481486925855279, + 0.007192202378064394, + 0.0035759643651545048, + 0.007783722132444382, + 0.010188274085521698, + 0.011130006983876228, + 0.01082698069512844, + 0.013016252778470516, + 0.006851756013929844, + 0.0064545744098722935, + 0.008224040269851685, + 0.00993418600410223, + 0.03243919834494591, + 0.021274475380778313, + 0.053243473172187805, + 0.04363967478275299, + 0.012027720920741558, + 0.013696269132196903, + 0.023465313017368317, + 0.011874917894601822, + 0.0019833194091916084, + 0.003171650692820549, + 0.0023420220240950584, + 0.00489922147244215, + 0.0016332867089658976, + 0.002396733732894063, + 0.0017382815713062882, + 0.0020807902328670025, + 0.00257346429862082, + 0.0030188884120434523, + 0.0026385770179331303, + 0.003774861339479685, + 0.0018952079117298126, + 0.0016786142950877547, + 0.0024353014305233955, + 0.0018027739133685827, + 0.0039039880502969027, + 0.004237804561853409, + 0.005882518831640482, + 0.007122809998691082, + 0.003704462666064501, + 0.003899306058883667, + 0.0038951439782977104, + 0.003313083201646805, + 0.00790412724018097, + 0.007942176423966885, + 0.009405006654560566, + 0.015442702919244766, + 0.005712603684514761, + 0.003976785112172365, + 0.011210455559194088, + 0.005402987822890282, + 0.007729170378297567, + 0.009837299585342407, + 0.008362148888409138, + 0.011135869659483433, + 0.005190793890506029, + 0.004509261343628168, + 0.008524352684617043, + 0.00662056403234601, + 0.010578740388154984, + 0.014538067393004894, + 0.01116766594350338, + 0.016296282410621643, + 0.005753233563154936, + 0.004941157530993223, + 0.01244062464684248, + 0.008680528029799461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.026410866528749466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813619911670685 + ], + "opposite_logprob_tokens": [ + -3.2881362438201904 + ], + "logprob_reward": -0.03813619911670685, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813619911670685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26344171166419983, + 0.15734031796455383, + 0.11291196942329407, + 0.13690322637557983, + 0.037121210247278214, + 0.05599932372570038, + 0.05509883910417557, + 0.049175798892974854, + 0.05107197165489197, + 0.08093568682670593 + ], + "credit_map": [ + 0.02896709367632866, + 0.04381296783685684, + 0.027303732931613922, + 0.034991294145584106, + 0.03715311363339424, + 0.026569070294499397, + 0.03547341376543045, + 0.029171006754040718, + 0.015970412641763687, + 0.02105606347322464, + 0.01993720419704914, + 0.03193163499236107, + 0.01382033433765173, + 0.019429121166467667, + 0.015119149349629879, + 0.020076395943760872, + 0.013519826345145702, + 0.01429179310798645, + 0.017456986010074615, + 0.012199290096759796, + 0.009368791244924068, + 0.01509177964180708, + 0.01289517804980278, + 0.018088320270180702, + 0.01834050752222538, + 0.01897692121565342, + 0.027414558455348015, + 0.017216261476278305, + 0.011423664167523384, + 0.012941287830471992, + 0.01047755777835846, + 0.02011246420443058, + 0.0038713172543793917, + 0.0046737841330468655, + 0.004993444308638573, + 0.00797748751938343, + 0.004395980853587389, + 0.00384335545822978, + 0.003967173397541046, + 0.0033986703492701054, + 0.006550296675413847, + 0.008729115128517151, + 0.008670943789184093, + 0.007908722385764122, + 0.003522133221849799, + 0.005807876121252775, + 0.006093094125390053, + 0.00871714111417532, + 0.0061637540347874165, + 0.006566514261066914, + 0.01034537609666586, + 0.006524976808577776, + 0.003970772959291935, + 0.005857203621417284, + 0.008210756815969944, + 0.007459488697350025, + 0.005066614598035812, + 0.00701918825507164, + 0.00755507405847311, + 0.008011291734874249, + 0.004585307091474533, + 0.003947093617171049, + 0.0074759917333722115, + 0.0055152359418570995, + 0.008199414238333702, + 0.0065971254371106625, + 0.009763081558048725, + 0.006763774435967207, + 0.004589139483869076, + 0.004030143842101097, + 0.005859410390257835, + 0.00526988273486495, + 0.013324864208698273, + 0.007660993374884129, + 0.0176104586571455, + 0.009704713709652424, + 0.007334422320127487, + 0.007549649570137262, + 0.009771679528057575, + 0.007978902198374271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03813619911670685 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207640662789345 + ], + "opposite_logprob_tokens": [ + -4.26420783996582 + ], + "logprob_reward": -0.014207640662789345, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207640662789345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3055538237094879, + 0.10221497714519501, + 0.06438926607370377, + 0.14811134338378906, + 0.051149673759937286, + 0.10267997533082962, + 0.08221136778593063, + 0.06757166236639023, + 0.05159834399819374, + 0.02451956644654274 + ], + "credit_map": [ + 0.04185836762189865, + 0.03270278498530388, + 0.04902058467268944, + 0.06982272863388062, + 0.02701663225889206, + 0.019847774878144264, + 0.04303635656833649, + 0.022248608991503716, + 0.014419618062675, + 0.01683400757610798, + 0.024486999958753586, + 0.014895318076014519, + 0.006803120486438274, + 0.008764016442000866, + 0.009786225855350494, + 0.006225679535418749, + 0.009127968922257423, + 0.007953718304634094, + 0.0103536956012249, + 0.011528057977557182, + 0.0054410300217568874, + 0.005339043214917183, + 0.007048746570944786, + 0.007597003132104874, + 0.016160113736987114, + 0.01765940524637699, + 0.026436682790517807, + 0.03873132914304733, + 0.011498961597681046, + 0.00779664097353816, + 0.01855478249490261, + 0.011273438110947609, + 0.0062920209020376205, + 0.007230937015265226, + 0.009353880770504475, + 0.007392751052975655, + 0.0052817463874816895, + 0.005911480635404587, + 0.004846783820539713, + 0.004840068519115448, + 0.015272929333150387, + 0.01578996516764164, + 0.01827242225408554, + 0.01771010272204876, + 0.009804647415876389, + 0.007071282714605331, + 0.010376662015914917, + 0.008381961844861507, + 0.010461840778589249, + 0.010593261569738388, + 0.014128852635622025, + 0.012042015790939331, + 0.007023199461400509, + 0.0061924695037305355, + 0.014076827093958855, + 0.007692908868193626, + 0.008398127742111683, + 0.009220666252076626, + 0.009308267384767532, + 0.014444302767515182, + 0.0054732621647417545, + 0.00419563939794898, + 0.01123698428273201, + 0.005294415634125471, + 0.006873921025544405, + 0.005838737823069096, + 0.00809320155531168, + 0.006817760411649942, + 0.003946778364479542, + 0.005362949334084988, + 0.007611690554767847, + 0.007053303997963667, + 0.003419168759137392, + 0.002917887642979622, + 0.004392930306494236, + 0.00425700331106782, + 0.0015206628013402224, + 0.0017072474583983421, + 0.003354688175022602, + 0.0029499779921025038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.014207640662789345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578940987586975 + ], + "opposite_logprob_tokens": [ + -0.6328940987586975 + ], + "logprob_reward": -0.7578940987586975, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578940987586975, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18315309286117554, + 0.11898421496152878, + 0.08061379194259644, + 0.08451857417821884, + 0.06577576696872711, + 0.11366954445838928, + 0.09494231641292572, + 0.1746073216199875, + 0.04548940062522888, + 0.03824598714709282 + ], + "credit_map": [ + 0.025607196614146233, + 0.02492121048271656, + 0.02277405560016632, + 0.04113234579563141, + 0.016447121277451515, + 0.01373367290943861, + 0.02330349013209343, + 0.015234000980854034, + 0.01475165132433176, + 0.018627522513270378, + 0.020797202363610268, + 0.02100328914821148, + 0.00777492206543684, + 0.011963178403675556, + 0.014358889311552048, + 0.009707549586892128, + 0.009933837689459324, + 0.012351932004094124, + 0.014491484500467777, + 0.010053367353975773, + 0.005946825724095106, + 0.008200653828680515, + 0.008396645076572895, + 0.011239043436944485, + 0.012970782816410065, + 0.008279918693006039, + 0.01900867372751236, + 0.011536601930856705, + 0.0061216456815600395, + 0.0073405043222010136, + 0.009901890531182289, + 0.009358547627925873, + 0.007316713221371174, + 0.008059591054916382, + 0.008307894691824913, + 0.012688362039625645, + 0.004683017265051603, + 0.007345038000494242, + 0.009824964217841625, + 0.007550177164375782, + 0.014180484227836132, + 0.013427330181002617, + 0.01905496045947075, + 0.020173221826553345, + 0.012010087259113789, + 0.011756292544305325, + 0.01388557255268097, + 0.009181588888168335, + 0.009381973184645176, + 0.01434791274368763, + 0.01285086665302515, + 0.012885781936347485, + 0.009922755882143974, + 0.009149783290922642, + 0.015044603496789932, + 0.011358637362718582, + 0.019002946093678474, + 0.030130187049508095, + 0.022071506828069687, + 0.029176047071814537, + 0.014604284428060055, + 0.013400441966950893, + 0.028995931148529053, + 0.017225967720150948, + 0.00600096769630909, + 0.006110003218054771, + 0.006369690876454115, + 0.005496268626302481, + 0.003814691910520196, + 0.005173909943550825, + 0.006375271361321211, + 0.006148595362901688, + 0.005798070691525936, + 0.004472295753657818, + 0.006649412680417299, + 0.005968693643808365, + 0.0029378384351730347, + 0.003743023844435811, + 0.0036980933509767056, + 0.0049785589799284935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.7578940987586975 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "shaky, handheld, medium shot at eye level immerses the viewer in an intimate and emotionally charged conversation. The camera's subtle, sustained movement focuses on a man in his 30s, his hair disheveled and shadowed at five o'clock, who speaks with passion. He wears a worn, dark blue denim jacket over a simple gray T-shirt, and sits leaning forward on a plush sofa. He gestures vigorously with his right hand, fingers spread and then clenched to punctuate his words. Across from him, a woman in her early 30s, her long, wavy brown hair partially obscuring her face, listens raptly. She wears a soft, cream-colored sweater. The scene is set in a cozy, dimly lit living room at night. A warm floor lamp casts a golden glow, creating soft shadows and highlighting the dancing dust particles in the air. When the man pauses, the woman slowly turns to meet his gaze, her expression softening with empathy, and she nods gently in deep agreement, a silent affirmation that resonates within the raw, authentic atmosphere of the moment.", + "seed": 7686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_10/index_6331_seed_7686_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.5187456786632538, + "vlm_logprob": -0.5187456786632538, + "vlm_yes_no_margin": 2.2249999761581423, + "vlm_token_credit": -0.5187456786632538, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.0249173641204834 + ], + "opposite_logprob_tokens": [ + -0.14991728961467743 + ], + "logprob_reward": -2.0249173641204834, + "yes_no_margin_reward": -1.8750001192092896, + "token_credit_reward": -2.0249173641204834, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20335756242275238, + 0.09557028114795685, + 0.08225800842046738, + 0.1870851367712021, + 0.03826314955949783, + 0.10900986194610596, + 0.05600246042013168, + 0.10102546215057373, + 0.07882270961999893, + 0.04860536381602287 + ], + "credit_map": [ + 0.033096689730882645, + 0.020554838702082634, + 0.017639918252825737, + 0.0475311353802681, + 0.016130603849887848, + 0.020839393138885498, + 0.027332961559295654, + 0.020232023671269417, + 0.014633714221417904, + 0.015531989745795727, + 0.014931022189557552, + 0.013569998554885387, + 0.007440845016390085, + 0.009524685330688953, + 0.01047398429363966, + 0.00946403481066227, + 0.011470342986285686, + 0.015827059745788574, + 0.007949652150273323, + 0.017507212236523628, + 0.0062208110466599464, + 0.006028973497450352, + 0.010394497774541378, + 0.006859456654638052, + 0.045921072363853455, + 0.017552297562360764, + 0.01124098151922226, + 0.01675787754356861, + 0.015856880694627762, + 0.04215941205620766, + 0.022608835250139236, + 0.014987793751060963, + 0.004700921475887299, + 0.007932322099804878, + 0.004028999246656895, + 0.006102070678025484, + 0.004513551481068134, + 0.003996029496192932, + 0.004065894056111574, + 0.002923359628766775, + 0.01853836514055729, + 0.013306330889463425, + 0.007772673387080431, + 0.024135330691933632, + 0.008025196380913258, + 0.013300662860274315, + 0.012378792278468609, + 0.011552523821592331, + 0.0069306339137256145, + 0.008661582134664059, + 0.0065643941052258015, + 0.00918181985616684, + 0.004923020489513874, + 0.005740853492170572, + 0.0075239683501422405, + 0.006476182024925947, + 0.012691144831478596, + 0.014641716144979, + 0.011164812371134758, + 0.012574802152812481, + 0.01038017775863409, + 0.01102573610842228, + 0.01596047729253769, + 0.012586594559252262, + 0.009162824600934982, + 0.014104274101555347, + 0.010318818502128124, + 0.008443485014140606, + 0.008405739441514015, + 0.00856314692646265, + 0.011999635957181454, + 0.007824783213436604, + 0.006261851638555527, + 0.009608924388885498, + 0.0057997615076601505, + 0.005851862486451864, + 0.0036553204990923405, + 0.004423823673278093, + 0.005786104127764702, + 0.0072177196852862835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -2.0249173641204834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018245482817292213 + ], + "opposite_logprob_tokens": [ + -4.018245697021484 + ], + "logprob_reward": -0.018245482817292213, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018245482817292213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5259951949119568, + 0.03949674963951111, + 0.02584107592701912, + 0.1181115210056305, + 0.072395920753479, + 0.08592427521944046, + 0.026428304612636566, + 0.050135329365730286, + 0.029814647510647774, + 0.02585708536207676 + ], + "credit_map": [ + 0.1024550050497055, + 0.03408971428871155, + 0.03294602781534195, + 0.17415860295295715, + 0.029777877032756805, + 0.04186473786830902, + 0.08087961375713348, + 0.029823627322912216, + 0.005781786050647497, + 0.006441887468099594, + 0.00569162517786026, + 0.005274659022688866, + 0.00382938957773149, + 0.004033332224935293, + 0.004237856715917587, + 0.004206210374832153, + 0.004403128754347563, + 0.005261609796434641, + 0.0023847739212214947, + 0.004266886506229639, + 0.001832089968957007, + 0.0023162495344877243, + 0.003564328420907259, + 0.0018120105378329754, + 0.03292492404580116, + 0.01201019249856472, + 0.005120954941958189, + 0.009401911869645119, + 0.008985021151602268, + 0.029374513775110245, + 0.012169094756245613, + 0.00812490750104189, + 0.018678713589906693, + 0.007720279041677713, + 0.005237974226474762, + 0.011747716926038265, + 0.004948055371642113, + 0.01270937081426382, + 0.007373225409537554, + 0.003980579320341349, + 0.011973828077316284, + 0.009310571476817131, + 0.005295742303133011, + 0.018630001693964005, + 0.008255871012806892, + 0.015105935744941235, + 0.008833478204905987, + 0.00851884763687849, + 0.002624892396852374, + 0.0047737229615449905, + 0.0026117339730262756, + 0.003481069812551141, + 0.0031821723096072674, + 0.003817964345216751, + 0.0025888257659971714, + 0.00334792397916317, + 0.007789129391312599, + 0.006518818903714418, + 0.006390417460352182, + 0.007790416479110718, + 0.00501265237107873, + 0.006865677889436483, + 0.00440297182649374, + 0.005365245044231415, + 0.00336852646432817, + 0.005659166723489761, + 0.003347853198647499, + 0.0025452927220612764, + 0.002862724242731929, + 0.0048184567131102085, + 0.0036810075398534536, + 0.0035316201392561197, + 0.003014977788552642, + 0.003350625978782773, + 0.0027985398191958666, + 0.0030867604073137045, + 0.0015397791285067797, + 0.0037301841657608747, + 0.0027302741073071957, + 0.005605941638350487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.018245482817292213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4796554446220398, + 0.0313614122569561, + 0.041326478123664856, + 0.03153248503804207, + 0.022151753306388855, + 0.0729026272892952, + 0.03811297565698624, + 0.06455565243959427, + 0.05495794489979744, + 0.16344326734542847 + ], + "credit_map": [ + 0.09313791990280151, + 0.028983771800994873, + 0.030311547219753265, + 0.16032378375530243, + 0.026970703154802322, + 0.03724941611289978, + 0.07494555413722992, + 0.027732735499739647, + 0.005306728184223175, + 0.004422246012836695, + 0.003917531110346317, + 0.004353979602456093, + 0.0026550916954874992, + 0.003993114922195673, + 0.004041111562401056, + 0.0026716089341789484, + 0.010580926202237606, + 0.005501373205333948, + 0.0031096527818590403, + 0.003804897889494896, + 0.002766546094790101, + 0.008320852182805538, + 0.004259692970663309, + 0.0029825353994965553, + 0.00376835186034441, + 0.0032356104347854853, + 0.0031147205736488104, + 0.005198857747018337, + 0.0027084313333034515, + 0.004417348653078079, + 0.005625125020742416, + 0.003464038949459791, + 0.005112059880048037, + 0.0027265490498393774, + 0.002417602576315403, + 0.0039450558833777905, + 0.002162483986467123, + 0.0028424998745322227, + 0.0018648469122126698, + 0.0010806554928421974, + 0.009355136193335056, + 0.006234586238861084, + 0.004145508166402578, + 0.016435736790299416, + 0.007965737022459507, + 0.014154315926134586, + 0.005917010363191366, + 0.008694601245224476, + 0.004461868666112423, + 0.005230724811553955, + 0.0021045347675681114, + 0.006959115620702505, + 0.0045121521688997746, + 0.006840211804956198, + 0.00275193783454597, + 0.005252431612461805, + 0.0082955751568079, + 0.008220192976295948, + 0.0043431902304291725, + 0.008139709942042828, + 0.0059331557713449, + 0.011384161189198494, + 0.006047475151717663, + 0.012192195281386375, + 0.0071854363195598125, + 0.008896825835108757, + 0.00470002181828022, + 0.004909987561404705, + 0.005414922256022692, + 0.010266822762787342, + 0.006127558182924986, + 0.007456369698047638, + 0.017369922250509262, + 0.02083965577185154, + 0.018072212114930153, + 0.021606553345918655, + 0.011224983260035515, + 0.019216356799006462, + 0.016926398500800133, + 0.038187168538570404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081389039754868 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.016081389039754868, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081389039754868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4932168126106262, + 0.05601461976766586, + 0.039829764515161514, + 0.041543424129486084, + 0.07976140826940536, + 0.0726536363363266, + 0.03658950701355934, + 0.097588911652565, + 0.041579972952604294, + 0.04122192785143852 + ], + "credit_map": [ + 0.09290970116853714, + 0.030291687697172165, + 0.031275395303964615, + 0.1641024947166443, + 0.02800249680876732, + 0.03969532996416092, + 0.07783694565296173, + 0.029102755710482597, + 0.008196846581995487, + 0.007819676771759987, + 0.0075743431225419044, + 0.006461401004344225, + 0.006808135658502579, + 0.005775536876171827, + 0.0083535835146904, + 0.005025099031627178, + 0.0062610087916255, + 0.006792418658733368, + 0.005005408078432083, + 0.005623506847769022, + 0.0034569725394248962, + 0.004200473893433809, + 0.0050183855928480625, + 0.0034715901128947735, + 0.006741723511368036, + 0.0068603442050516605, + 0.004218813963234425, + 0.004568599630147219, + 0.003980990499258041, + 0.005952367093414068, + 0.0053551411256194115, + 0.00386544456705451, + 0.012283101677894592, + 0.01070556789636612, + 0.006031838711351156, + 0.017383674159646034, + 0.008713613264262676, + 0.01014384999871254, + 0.00846086535602808, + 0.006038897205144167, + 0.014991830103099346, + 0.007398481015115976, + 0.005968654528260231, + 0.020699195563793182, + 0.0036501954309642315, + 0.005654304753988981, + 0.007268608082085848, + 0.007022371515631676, + 0.0045113395899534225, + 0.00430842861533165, + 0.0030716920737177134, + 0.006791993975639343, + 0.00368690537288785, + 0.005047188606113195, + 0.004944032058119774, + 0.00422792648896575, + 0.01440069917589426, + 0.012373829260468483, + 0.00953706819564104, + 0.018909214064478874, + 0.007882468402385712, + 0.011058025062084198, + 0.011745452880859375, + 0.01168215461075306, + 0.005959311034530401, + 0.008611151948571205, + 0.004194915760308504, + 0.004868547432124615, + 0.0036911428906023502, + 0.005696909502148628, + 0.004730020184069872, + 0.0038279728032648563, + 0.00503895478323102, + 0.006226250436156988, + 0.0046273646876215935, + 0.005354751832783222, + 0.00353326927870512, + 0.005181537941098213, + 0.005229529924690723, + 0.006030268035829067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.016081389039754868 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man in his 30s wearing a dark blue denim jacket and a gray T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman in her early 30s wearing a cream-colored sweater visible across from the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a cozy, dimly lit living room at night with a warm floor lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man gesture vigorously with his right hand while speaking?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234031677246094 + ], + "opposite_logprob_tokens": [ + -0.8984031677246094 + ], + "logprob_reward": -0.5234031677246094, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234031677246094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11369890719652176, + 0.07076148688793182, + 0.08902795612812042, + 0.46235746145248413, + 0.027763037011027336, + 0.08318700641393661, + 0.0471910797059536, + 0.04497753083705902, + 0.032172609120607376, + 0.028862955048680305 + ], + "credit_map": [ + 0.017876561731100082, + 0.011168867349624634, + 0.008078652434051037, + 0.030226271599531174, + 0.011453292332589626, + 0.010300238616764545, + 0.016790367662906647, + 0.007804646622389555, + 0.01253353338688612, + 0.008579674176871777, + 0.007110323291271925, + 0.015462074428796768, + 0.004726796876639128, + 0.008036073297262192, + 0.008664141409099102, + 0.005648864433169365, + 0.02143346518278122, + 0.014506850391626358, + 0.007652027532458305, + 0.008209624327719212, + 0.005910724867135286, + 0.01766638644039631, + 0.008049341849982738, + 0.005599530413746834, + 0.10031761229038239, + 0.053808942437171936, + 0.02579454332590103, + 0.04273074492812157, + 0.04174533858895302, + 0.09999751299619675, + 0.06207302585244179, + 0.03588969632983208, + 0.006557394750416279, + 0.00421353243291378, + 0.0023362599313259125, + 0.0030548647046089172, + 0.0022907499223947525, + 0.00512666767463088, + 0.002650800161063671, + 0.0015327655710279942, + 0.010192704387009144, + 0.009249328635632992, + 0.004665805958211422, + 0.018762048333883286, + 0.0089949369430542, + 0.015362328849732876, + 0.007159263361245394, + 0.008800586685538292, + 0.005242777988314629, + 0.0071984934620559216, + 0.005672489292919636, + 0.007131517864763737, + 0.004753059707581997, + 0.005054680164903402, + 0.006411672569811344, + 0.005726383998990059, + 0.006208482198417187, + 0.00708047766238451, + 0.004810503218322992, + 0.006467271130532026, + 0.00443144841119647, + 0.005784983281046152, + 0.005521285347640514, + 0.004673079587519169, + 0.004575997591018677, + 0.0052695972844958305, + 0.0034296386875212193, + 0.004996173083782196, + 0.0034974971786141396, + 0.003456756705418229, + 0.0035911400336772203, + 0.0033558078575879335, + 0.004570864140987396, + 0.005675614811480045, + 0.0027237532194703817, + 0.003439240623265505, + 0.002466699341312051, + 0.003558059688657522, + 0.0036160233430564404, + 0.0028126989491283894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.5234031677246094 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_11/index_tensor([9180, 9180])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_11/index_tensor([9180, 9180])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2a4794e93885355b7abe684c476dacf8ac566c1d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_11/index_tensor([9180, 9180])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, captured from a static, slightly high-angle perspective, meticulously frames a quiet, physical moment. The polished silver tip of a teaspoon hovers above a perfectly dark, polished countertop, reflecting like black water. A pool of thick, golden syrup, translucent and luminous, rests on the spoon, capturing the soft, warm glow of an invisible overhead light. This light creates a brilliant highlight along the spoon's curve and a distorted, warm reflection on the countertop below. From the spoon's edge, a viscous droplet slowly and deliberately elongates, stretched by gravity into a shimmering amber thread. It finally separates with a silent thud and collapses onto the surface into a perfect little pool of syrup. The impact creates subtle, silent ripples that expand outward with a mesmerizing slow-motion effect. The atmosphere is one of tranquil focus and sensual detail, emphasizing the stark contrast between the warm, glowing syrup and the cool, dark surface.", + "seed": 15717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_11/index_9180_seed_15717_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009014501585625112, + "vlm_logprob": -0.009014501585625112, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.009014501585625112, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023973703384399414 + ], + "opposite_logprob_tokens": [ + -4.64897346496582 + ], + "logprob_reward": -0.023973703384399414, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.023973703384399414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4040001928806305, + 0.03997655212879181, + 0.0775565505027771, + 0.07609324157238007, + 0.059161439538002014, + 0.16579987108707428, + 0.06866490840911865, + 0.028047848492860794, + 0.019369211047887802, + 0.06133018806576729 + ], + "credit_map": [ + 0.04602914676070213, + 0.03891773521900177, + 0.05345289036631584, + 0.04587177559733391, + 0.03776627033948898, + 0.0744466558098793, + 0.050399672240018845, + 0.05711602047085762, + 0.007815509103238583, + 0.00254397583194077, + 0.006038418505340815, + 0.008210227824747562, + 0.001572233741171658, + 0.0033751721493899822, + 0.0034136460162699223, + 0.00700736977159977, + 0.01174716092646122, + 0.004392070695757866, + 0.011597763746976852, + 0.011136554181575775, + 0.0030841976404190063, + 0.00840489286929369, + 0.005661884322762489, + 0.02153201773762703, + 0.010054566897451878, + 0.006073175463825464, + 0.01133379340171814, + 0.009089765138924122, + 0.005322149954736233, + 0.013901556842029095, + 0.004731598775833845, + 0.01558663696050644, + 0.0069270748645067215, + 0.0055231270380318165, + 0.008141404017806053, + 0.00692121684551239, + 0.005908293183892965, + 0.010223854333162308, + 0.00474826293066144, + 0.010768203996121883, + 0.024846673011779785, + 0.0255782101303339, + 0.023594487458467484, + 0.01577869802713394, + 0.013227144256234169, + 0.02789672277867794, + 0.012784876860678196, + 0.022093050181865692, + 0.00917923916131258, + 0.007735965773463249, + 0.010520938783884048, + 0.006159182172268629, + 0.007460341788828373, + 0.013447615318000317, + 0.006283252499997616, + 0.007878363132476807, + 0.0036826147697865963, + 0.0031755631789565086, + 0.003916522022336721, + 0.004134021699428558, + 0.003143857466056943, + 0.003688342636451125, + 0.002695009345188737, + 0.003611915512010455, + 0.002974419156089425, + 0.002464230405166745, + 0.0033492206130176783, + 0.0022134024184197187, + 0.001897878828458488, + 0.0026846984401345253, + 0.0018801646074280143, + 0.0019051963463425636, + 0.008748474530875683, + 0.006084339693188667, + 0.009996599517762661, + 0.008580251596868038, + 0.005446060094982386, + 0.008966765366494656, + 0.006480058655142784, + 0.0070276339538395405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.023973703384399414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221626333892345 + ], + "opposite_logprob_tokens": [ + -4.264221668243408 + ], + "logprob_reward": -0.014221626333892345, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221626333892345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2382201999425888, + 0.14057455956935883, + 0.08892584592103958, + 0.0945579782128334, + 0.08444851636886597, + 0.12948575615882874, + 0.11780405044555664, + 0.04304034262895584, + 0.02343481406569481, + 0.03950796648859978 + ], + "credit_map": [ + 0.028506234288215637, + 0.021902961656451225, + 0.030170239508152008, + 0.03125961497426033, + 0.024355284869670868, + 0.03464341163635254, + 0.028937257826328278, + 0.03844518959522247, + 0.015049176290631294, + 0.01307909656316042, + 0.021442662924528122, + 0.02647496573626995, + 0.010757450014352798, + 0.024795835837721825, + 0.011634276248514652, + 0.017341090366244316, + 0.01000456977635622, + 0.007880202494561672, + 0.012654499150812626, + 0.011456500738859177, + 0.007474942598491907, + 0.017099060118198395, + 0.010507255792617798, + 0.011848815716803074, + 0.014847137033939362, + 0.012118847109377384, + 0.014126977883279324, + 0.011108988896012306, + 0.004119537770748138, + 0.014449045062065125, + 0.006805767770856619, + 0.01698167808353901, + 0.014226026833057404, + 0.007453808560967445, + 0.013122551143169403, + 0.007362424861639738, + 0.006009382661432028, + 0.014787049032747746, + 0.005846284795552492, + 0.015640990808606148, + 0.01958708092570305, + 0.009616687893867493, + 0.016970163211226463, + 0.015780597925186157, + 0.009484752081334591, + 0.024199452251195908, + 0.008639817126095295, + 0.02520720101892948, + 0.01698601245880127, + 0.012948913499712944, + 0.017468847334384918, + 0.009681520983576775, + 0.012654130347073078, + 0.02623523212969303, + 0.008865712210536003, + 0.012963689863681793, + 0.004996506031602621, + 0.004960778169333935, + 0.0064822472631931305, + 0.004750146064907312, + 0.0035482433158904314, + 0.00894003827124834, + 0.003908959683030844, + 0.005453421734273434, + 0.004206856247037649, + 0.0031206973362714052, + 0.004217532463371754, + 0.002439431846141815, + 0.0020476547069847584, + 0.0030268062837421894, + 0.002189568942412734, + 0.0021862683352082968, + 0.0060380264185369015, + 0.004863268230110407, + 0.006715377327054739, + 0.004493405111134052, + 0.0033699236810207367, + 0.006616732105612755, + 0.0027219243347644806, + 0.00468931021168828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.014221626333892345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010552796302363276 + ], + "opposite_logprob_tokens": [ + -6.8760552406311035 + ], + "logprob_reward": -0.0010552796302363276, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010552796302363276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2584901750087738, + 0.21091870963573456, + 0.059419505298137665, + 0.05416583642363548, + 0.08736288547515869, + 0.0530669242143631, + 0.1282883882522583, + 0.03706492856144905, + 0.04740985855460167, + 0.06381279975175858 + ], + "credit_map": [ + 0.02916700206696987, + 0.0176406167447567, + 0.03444919362664223, + 0.03275119513273239, + 0.014916110783815384, + 0.04082728922367096, + 0.03371719270944595, + 0.05502155423164368, + 0.023163706064224243, + 0.018315084278583527, + 0.027568161487579346, + 0.04855852946639061, + 0.011090262793004513, + 0.03628898039460182, + 0.021813133731484413, + 0.02412085235118866, + 0.007394279818981886, + 0.0043735196813941, + 0.008005444891750813, + 0.011018083430826664, + 0.005208301357924938, + 0.007160497363656759, + 0.006698256824165583, + 0.009561121463775635, + 0.008544853888452053, + 0.004380601458251476, + 0.00871209241449833, + 0.006338174920529127, + 0.0026203724555671215, + 0.008112652227282524, + 0.003963307943195105, + 0.011493784375488758, + 0.010244966484606266, + 0.009279251098632812, + 0.013348518870770931, + 0.011442664079368114, + 0.006016389932483435, + 0.01720571704208851, + 0.005929828155785799, + 0.013895551674067974, + 0.010068868286907673, + 0.006174882873892784, + 0.00790657103061676, + 0.004394827876240015, + 0.00403651362285018, + 0.011325336992740631, + 0.0036233479622751474, + 0.0055365730077028275, + 0.019106600433588028, + 0.013043848797678947, + 0.020134098827838898, + 0.012620313093066216, + 0.014344598166644573, + 0.0237474013119936, + 0.008765248581767082, + 0.016526274383068085, + 0.0050496975891292095, + 0.0052215722389519215, + 0.005482434295117855, + 0.003134539583697915, + 0.0035928452853113413, + 0.005976791959255934, + 0.003604046069085598, + 0.005003001075237989, + 0.007718150038272142, + 0.004953729920089245, + 0.008124003186821938, + 0.0045758821070194244, + 0.005816753022372723, + 0.006548211444169283, + 0.0033041576389223337, + 0.006368972826749086, + 0.008590428158640862, + 0.008855360560119152, + 0.009312606416642666, + 0.00940316915512085, + 0.00727556087076664, + 0.01031564548611641, + 0.004305926617234945, + 0.005754103418439627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0010552796302363276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022023487836122513 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022023487836122513, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022023487836122513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2032199501991272, + 0.0905928760766983, + 0.057369060814380646, + 0.05500921979546547, + 0.16266372799873352, + 0.10378392785787582, + 0.14696171879768372, + 0.03134414181113243, + 0.06284207850694656, + 0.08621327579021454 + ], + "credit_map": [ + 0.026730570942163467, + 0.01764264889061451, + 0.02484002150595188, + 0.030610980466008186, + 0.01582939177751541, + 0.03311813995242119, + 0.02392513118684292, + 0.030523065477609634, + 0.013770206831395626, + 0.00959406141191721, + 0.012728692032396793, + 0.012154549360275269, + 0.007905077189207077, + 0.00908929854631424, + 0.00974347535520792, + 0.015607518143951893, + 0.00931121688336134, + 0.004789719358086586, + 0.00881354883313179, + 0.006409208755940199, + 0.003150623058900237, + 0.006340986583381891, + 0.0040705581195652485, + 0.0144831957295537, + 0.008612655103206635, + 0.0038705579936504364, + 0.0084544001147151, + 0.005847119260579348, + 0.002630428643897176, + 0.009922743774950504, + 0.004142825026065111, + 0.011528491973876953, + 0.017575515434145927, + 0.017247306182980537, + 0.024826692417263985, + 0.022205129265785217, + 0.013741955161094666, + 0.0306455846875906, + 0.010941559448838234, + 0.02547997608780861, + 0.01712246611714363, + 0.01126819010823965, + 0.014095312915742397, + 0.008165137842297554, + 0.007430884521454573, + 0.022002581506967545, + 0.0063737197779119015, + 0.01732562854886055, + 0.02039230242371559, + 0.01366768404841423, + 0.02340053766965866, + 0.010834719985723495, + 0.017085298895835876, + 0.03196163848042488, + 0.012538901530206203, + 0.017080647870898247, + 0.0047835963778197765, + 0.0027150390669703484, + 0.004271150566637516, + 0.003167351707816124, + 0.002420770702883601, + 0.004736581817269325, + 0.0031300485134124756, + 0.0061196014285087585, + 0.009828615002334118, + 0.00857491698116064, + 0.0115024009719491, + 0.005833279807120562, + 0.006780959665775299, + 0.010896552354097366, + 0.004959516692906618, + 0.004465837962925434, + 0.010344992391765118, + 0.013173738494515419, + 0.013669360429048538, + 0.011649553664028645, + 0.010510516352951527, + 0.014885807409882545, + 0.005504287779331207, + 0.006475022062659264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0022023487836122513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619549795985222 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619549795985222, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619549795985222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22422167658805847, + 0.24184675514698029, + 0.027580689638853073, + 0.0423140823841095, + 0.1963804066181183, + 0.07793083786964417, + 0.0713997632265091, + 0.01725342497229576, + 0.05520296096801758, + 0.04586949199438095 + ], + "credit_map": [ + 0.027752362191677094, + 0.023960238322615623, + 0.02744685672223568, + 0.026873812079429626, + 0.024948282167315483, + 0.034044623374938965, + 0.029465802013874054, + 0.02972966991364956, + 0.02457141876220703, + 0.02451765350997448, + 0.03209066763520241, + 0.05677200108766556, + 0.01506018452346325, + 0.0407092422246933, + 0.021300243213772774, + 0.026825299486517906, + 0.0035495106130838394, + 0.0032344567589461803, + 0.004150687716901302, + 0.0030949630308896303, + 0.0023702369071543217, + 0.003294342663139105, + 0.0025520510971546173, + 0.005334437824785709, + 0.005620172247290611, + 0.00467489892616868, + 0.007386550307273865, + 0.005160834640264511, + 0.004589519929140806, + 0.006404613610357046, + 0.002987530315294862, + 0.005489957053214312, + 0.021363725885748863, + 0.0239840317517519, + 0.030392184853553772, + 0.025995805859565735, + 0.018092738464474678, + 0.03565983101725578, + 0.015504717826843262, + 0.025387343019247055, + 0.01088602002710104, + 0.0100955069065094, + 0.011653388850390911, + 0.0069717601872980595, + 0.00491370027884841, + 0.012059800326824188, + 0.005981305614113808, + 0.015369337983429432, + 0.009369690902531147, + 0.007692508399486542, + 0.009707434102892876, + 0.006878537591546774, + 0.0060594091191887856, + 0.015107497572898865, + 0.004859121050685644, + 0.011725553311407566, + 0.0029518732335418463, + 0.002072450937703252, + 0.0025344437453895807, + 0.0017874559853225946, + 0.0017779931658878922, + 0.0023118932731449604, + 0.0014299956383183599, + 0.0023873159661889076, + 0.008614682592451572, + 0.006029232870787382, + 0.00858934037387371, + 0.005649865604937077, + 0.005382603965699673, + 0.00792614184319973, + 0.00468165148049593, + 0.008329433389008045, + 0.005634893663227558, + 0.006483994424343109, + 0.006886224262416363, + 0.007081363350152969, + 0.003928437829017639, + 0.007360554300248623, + 0.0031733654905110598, + 0.005320655182003975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.003619549795985222 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, captured from a static, slightly high-angle perspective, meticulously frames a quiet, physical moment. The polished silver tip of a teaspoon hovers above a perfectly dark, polished countertop, reflecting like black water. A pool of thick, golden syrup, translucent and luminous, rests on the spoon, capturing the soft, warm glow of an invisible overhead light. This light creates a brilliant highlight along the spoon's curve and a distorted, warm reflection on the countertop below. From the spoon's edge, a viscous droplet slowly and deliberately elongates, stretched by gravity into a shimmering amber thread. It finally separates with a silent thud and collapses onto the surface into a perfect little pool of syrup. The impact creates subtle, silent ripples that expand outward with a mesmerizing slow-motion effect. The atmosphere is one of tranquil focus and sensual detail, emphasizing the stark contrast between the warm, glowing syrup and the cool, dark surface.", + "seed": 15718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_11/index_9180_seed_15718_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013369890546891838, + "vlm_logprob": -0.013369890546891838, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.013369890546891838, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030720794573426247 + ], + "opposite_logprob_tokens": [ + -4.0307207107543945 + ], + "logprob_reward": -0.030720794573426247, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.030720794573426247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14150723814964294, + 0.11984176933765411, + 0.10369507968425751, + 0.16171175241470337, + 0.07467575371265411, + 0.07101072371006012, + 0.05682047829031944, + 0.07088817656040192, + 0.06576400995254517, + 0.1340850442647934 + ], + "credit_map": [ + 0.01843227446079254, + 0.013496136292815208, + 0.015077497810125351, + 0.025789272040128708, + 0.011794810183346272, + 0.011919887736439705, + 0.02394045516848564, + 0.021056899800896645, + 0.013772563077509403, + 0.018168529495596886, + 0.014687098562717438, + 0.01842569373548031, + 0.013924065977334976, + 0.012845247983932495, + 0.014114972203969955, + 0.013903598301112652, + 0.016675928607583046, + 0.011374549940228462, + 0.011089485138654709, + 0.021574633195996284, + 0.013212204910814762, + 0.00975379254668951, + 0.012135934084653854, + 0.007878552190959454, + 0.03989216685295105, + 0.020044568926095963, + 0.02059362456202507, + 0.024552850052714348, + 0.015492242760956287, + 0.015722671523690224, + 0.012402796186506748, + 0.013010840862989426, + 0.011850896291434765, + 0.01667042449116707, + 0.007553542032837868, + 0.006275621242821217, + 0.01046132855117321, + 0.008756381459534168, + 0.005210597533732653, + 0.007896961644291878, + 0.00728940824046731, + 0.010064171627163887, + 0.005028484854847193, + 0.006882206071168184, + 0.007566446904093027, + 0.012751941569149494, + 0.0063013979233801365, + 0.015126671642065048, + 0.007105731405317783, + 0.009157843887805939, + 0.00489173736423254, + 0.006437504664063454, + 0.00938886497169733, + 0.00616068160161376, + 0.006090973969548941, + 0.007587139960378408, + 0.008607191033661366, + 0.013135053217411041, + 0.007200001273304224, + 0.008384165354073048, + 0.011227237991988659, + 0.008264894597232342, + 0.006930598523467779, + 0.007139034103602171, + 0.009800156578421593, + 0.009296974167227745, + 0.005723560228943825, + 0.006003630813211203, + 0.01237296499311924, + 0.010062583722174168, + 0.0037272193003445864, + 0.008776921778917313, + 0.01798669621348381, + 0.016208158805966377, + 0.012695425190031528, + 0.01164157409220934, + 0.024065988138318062, + 0.0173246618360281, + 0.01580556109547615, + 0.01835697516798973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.030720794573426247 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029829679057002068 + ], + "opposite_logprob_tokens": [ + -3.529829740524292 + ], + "logprob_reward": -0.029829679057002068, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029829679057002068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12627550959587097, + 0.06400306522846222, + 0.07191488146781921, + 0.16752944886684418, + 0.22017066180706024, + 0.08985928446054459, + 0.06440754979848862, + 0.05546681210398674, + 0.05245264992117882, + 0.0879201591014862 + ], + "credit_map": [ + 0.012029738165438175, + 0.02282930165529251, + 0.01211097463965416, + 0.019686341285705566, + 0.01517022680491209, + 0.014178961515426636, + 0.01694982685148716, + 0.013320132158696651, + 0.007698180619627237, + 0.005186796188354492, + 0.008479300886392593, + 0.01238408125936985, + 0.007007324602454901, + 0.005174174904823303, + 0.011141328141093254, + 0.006931882817298174, + 0.008082524873316288, + 0.008061009459197521, + 0.007095316890627146, + 0.011848349124193192, + 0.010348290205001831, + 0.007662518415600061, + 0.01064079999923706, + 0.008176077157258987, + 0.024425772950053215, + 0.021158047020435333, + 0.017195308580994606, + 0.028056006878614426, + 0.020348137244582176, + 0.02115650102496147, + 0.0189665537327528, + 0.016223113983869553, + 0.03937649726867676, + 0.046379365026950836, + 0.022818930447101593, + 0.020384130999445915, + 0.0244562067091465, + 0.01868836022913456, + 0.023597456514835358, + 0.024469705298542976, + 0.010663291439414024, + 0.012661031447350979, + 0.006249499972909689, + 0.006259394809603691, + 0.009832891635596752, + 0.016001597046852112, + 0.00848050694912672, + 0.019711069762706757, + 0.005723549053072929, + 0.009621476754546165, + 0.004916981793940067, + 0.004887180402874947, + 0.010705232620239258, + 0.010224960744380951, + 0.006809724494814873, + 0.01151844672858715, + 0.007231126539409161, + 0.009557100012898445, + 0.004501111805438995, + 0.0048340060748159885, + 0.008474317379295826, + 0.007122259587049484, + 0.005744566209614277, + 0.008002325892448425, + 0.005708654876798391, + 0.0054182629100978374, + 0.0041353777050971985, + 0.003815264208242297, + 0.008526695892214775, + 0.00927380844950676, + 0.004472754895687103, + 0.011101827956736088, + 0.00887656956911087, + 0.011490409262478352, + 0.007503143046051264, + 0.005087363068014383, + 0.0236101932823658, + 0.013228895142674446, + 0.007490427698940039, + 0.01063315849751234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.029829679057002068 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009342834819108248 + ], + "opposite_logprob_tokens": [ + -7.00093412399292 + ], + "logprob_reward": -0.0009342834819108248, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009342834819108248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10840310901403427, + 0.2253643125295639, + 0.0844416618347168, + 0.07053397595882416, + 0.05945160984992981, + 0.04025951400399208, + 0.12614023685455322, + 0.03406135365366936, + 0.03484407812356949, + 0.21650013327598572 + ], + "credit_map": [ + 0.009960287250578403, + 0.013203115202486515, + 0.009234342724084854, + 0.020797714591026306, + 0.007537891622632742, + 0.008969572372734547, + 0.0235451627522707, + 0.015155017375946045, + 0.01630043052136898, + 0.02323196455836296, + 0.024961372837424278, + 0.04502415657043457, + 0.028589503839612007, + 0.02386665716767311, + 0.03818647935986519, + 0.02520374022424221, + 0.010035213083028793, + 0.012257585301995277, + 0.007797862868756056, + 0.012176325544714928, + 0.014259915798902512, + 0.009508018381893635, + 0.010205000638961792, + 0.008201736956834793, + 0.01042974554002285, + 0.010793640278279781, + 0.0063439663499593735, + 0.010256200097501278, + 0.008553304709494114, + 0.009765500202775002, + 0.007655967492610216, + 0.0067356545478105545, + 0.006931119132786989, + 0.01142297312617302, + 0.00512040127068758, + 0.004687851294875145, + 0.008564653806388378, + 0.008053337223827839, + 0.005860540550202131, + 0.008810732513666153, + 0.004574750084429979, + 0.00767754390835762, + 0.004395652562379837, + 0.003910854924470186, + 0.004667344968765974, + 0.005953200161457062, + 0.00348612479865551, + 0.005594043992459774, + 0.023272505030035973, + 0.02432546764612198, + 0.013831638731062412, + 0.01342199556529522, + 0.01815151423215866, + 0.013502886518836021, + 0.008985115215182304, + 0.010649123229086399, + 0.006055885460227728, + 0.004952323157340288, + 0.0034398692660033703, + 0.004305513110011816, + 0.003572803921997547, + 0.0035994872450828552, + 0.00396348861977458, + 0.0041719842702150345, + 0.004469228442758322, + 0.004594051744788885, + 0.002738886047154665, + 0.002891619224101305, + 0.005109965801239014, + 0.004774073138833046, + 0.0040409076027572155, + 0.006225341930985451, + 0.021283943206071854, + 0.02546263299882412, + 0.013356341049075127, + 0.013938471674919128, + 0.04078705981373787, + 0.0364752933382988, + 0.021579494699835777, + 0.043616898357868195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.0009342834819108248 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007239343249239028 + ], + "opposite_logprob_tokens": [ + -7.250723838806152 + ], + "logprob_reward": -0.0007239343249239028, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007239343249239028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06101015582680702, + 0.04816658794879913, + 0.08357863873243332, + 0.07807734608650208, + 0.05761195719242096, + 0.05785827711224556, + 0.2599995732307434, + 0.06654627621173859, + 0.04422563314437866, + 0.24292558431625366 + ], + "credit_map": [ + 0.008238966576755047, + 0.008979599922895432, + 0.0067642610520124435, + 0.010572409257292747, + 0.004821536596864462, + 0.004140536766499281, + 0.010450678877532482, + 0.007042165845632553, + 0.004136651288717985, + 0.009556498378515244, + 0.005009185057133436, + 0.0075864726677536964, + 0.005962387192994356, + 0.00452876603230834, + 0.006325914058834314, + 0.005060713738203049, + 0.011423246003687382, + 0.012473907321691513, + 0.007705638650804758, + 0.0157463476061821, + 0.013195058330893517, + 0.008896537125110626, + 0.007670722436159849, + 0.006467186380177736, + 0.012983402237296104, + 0.011881493963301182, + 0.008745420724153519, + 0.01170535758137703, + 0.008217314258217812, + 0.008515316061675549, + 0.00869392417371273, + 0.007335116155445576, + 0.009706951677799225, + 0.010311723686754704, + 0.006306239869445562, + 0.0050206901505589485, + 0.006753964815288782, + 0.006367230787873268, + 0.006690016482025385, + 0.0064551387913525105, + 0.008664263412356377, + 0.005183475557714701, + 0.007090157363563776, + 0.0054134768433868885, + 0.0074175582267344, + 0.008261178620159626, + 0.00543171726167202, + 0.010396448895335197, + 0.034554373472929, + 0.03865313157439232, + 0.022899260744452477, + 0.01718665473163128, + 0.04280221462249756, + 0.03724818676710129, + 0.024997485801577568, + 0.041658252477645874, + 0.007413901854306459, + 0.008672554977238178, + 0.005407208576798439, + 0.005872494075447321, + 0.010047177784144878, + 0.009429620578885078, + 0.006822285708039999, + 0.012881027534604073, + 0.005266574677079916, + 0.004664284642785788, + 0.003766830312088132, + 0.003448367351666093, + 0.0061486284248530865, + 0.007006238214671612, + 0.0036311622243374586, + 0.010293549858033657, + 0.0214114710688591, + 0.024043116718530655, + 0.015604178421199322, + 0.01488504558801651, + 0.04492081701755524, + 0.04325848072767258, + 0.025417782366275787, + 0.0533846952021122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.0007239343249239028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464076129719615 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.00464076129719615, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464076129719615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11612705141305923, + 0.10689300298690796, + 0.0632258877158165, + 0.05351850390434265, + 0.05112067237496376, + 0.06658238172531128, + 0.11728286743164062, + 0.04376085102558136, + 0.06146007403731346, + 0.32002872228622437 + ], + "credit_map": [ + 0.013111084699630737, + 0.014454707503318787, + 0.013350321911275387, + 0.024602094665169716, + 0.009974022395908833, + 0.011340411379933357, + 0.016680194064974785, + 0.012614213861525059, + 0.008940892294049263, + 0.016900764778256416, + 0.01225802581757307, + 0.019148539751768112, + 0.012596546672284603, + 0.011252127587795258, + 0.014854459092020988, + 0.010941646061837673, + 0.006621545646339655, + 0.005483720917254686, + 0.006848548073321581, + 0.010398305021226406, + 0.008996649645268917, + 0.007702862843871117, + 0.009831869974732399, + 0.007342378608882427, + 0.010185707360506058, + 0.0055288588628172874, + 0.006395604461431503, + 0.007789097726345062, + 0.004771439358592033, + 0.0073937913402915, + 0.00574793666601181, + 0.005706068128347397, + 0.011975187808275223, + 0.008792044594883919, + 0.006616172380745411, + 0.004606625065207481, + 0.0053488388657569885, + 0.0037416736595332623, + 0.005281282588839531, + 0.004758847411721945, + 0.006428121123462915, + 0.006944434251636267, + 0.005460215732455254, + 0.007528951391577721, + 0.010132323019206524, + 0.0107886902987957, + 0.007971815764904022, + 0.011327829211950302, + 0.02086907997727394, + 0.022158175706863403, + 0.012008273974061012, + 0.008500087074935436, + 0.017105573788285255, + 0.01283347699791193, + 0.010025102645158768, + 0.013783099129796028, + 0.005684640724211931, + 0.006200718227773905, + 0.004299356136471033, + 0.0048431819304823875, + 0.006197554990649223, + 0.005503494292497635, + 0.003986718133091927, + 0.007045184727758169, + 0.008720764890313148, + 0.008379451930522919, + 0.0049264575354754925, + 0.004440652206540108, + 0.009515016339719296, + 0.009905774146318436, + 0.004476514179259539, + 0.011095446534454823, + 0.037363432347774506, + 0.04098602011799812, + 0.02210838533937931, + 0.022621963173151016, + 0.053269002586603165, + 0.05024132877588272, + 0.03318702429533005, + 0.060251567512750626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.00464076129719615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, captured from a static, slightly high-angle perspective, meticulously frames a quiet, physical moment. The polished silver tip of a teaspoon hovers above a perfectly dark, polished countertop, reflecting like black water. A pool of thick, golden syrup, translucent and luminous, rests on the spoon, capturing the soft, warm glow of an invisible overhead light. This light creates a brilliant highlight along the spoon's curve and a distorted, warm reflection on the countertop below. From the spoon's edge, a viscous droplet slowly and deliberately elongates, stretched by gravity into a shimmering amber thread. It finally separates with a silent thud and collapses onto the surface into a perfect little pool of syrup. The impact creates subtle, silent ripples that expand outward with a mesmerizing slow-motion effect. The atmosphere is one of tranquil focus and sensual detail, emphasizing the stark contrast between the warm, glowing syrup and the cool, dark surface.", + "seed": 15719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_11/index_9180_seed_15719_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014009703398915007, + "vlm_logprob": -0.014009703398915007, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.014009703398915007, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03036779724061489 + ], + "opposite_logprob_tokens": [ + -3.905367851257324 + ], + "logprob_reward": -0.03036779724061489, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03036779724061489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07496516406536102, + 0.05758616328239441, + 0.07785272598266602, + 0.05596282333135605, + 0.24644877016544342, + 0.21557573974132538, + 0.06920073926448822, + 0.07755564153194427, + 0.06644307076931, + 0.058409184217453 + ], + "credit_map": [ + 0.009861032478511333, + 0.01013992540538311, + 0.010167640633881092, + 0.009454740211367607, + 0.006827821955084801, + 0.005140115972608328, + 0.011858640238642693, + 0.011515244841575623, + 0.0043387338519096375, + 0.005853066686540842, + 0.008390096947550774, + 0.00753400195389986, + 0.0065855602733790874, + 0.006535266991704702, + 0.008735612966120243, + 0.009613825008273125, + 0.008668415248394012, + 0.008104339241981506, + 0.012464835308492184, + 0.01007048785686493, + 0.006633318029344082, + 0.007656777277588844, + 0.01381604466587305, + 0.01043850276619196, + 0.009004288353025913, + 0.004666795954108238, + 0.013805192895233631, + 0.006514469161629677, + 0.0033292018342763186, + 0.0034931227564811707, + 0.005718158092349768, + 0.009431594982743263, + 0.025304444134235382, + 0.02433549240231514, + 0.04665970429778099, + 0.027202554047107697, + 0.014754325151443481, + 0.024889476597309113, + 0.030260024592280388, + 0.053042735904455185, + 0.019083403050899506, + 0.026645425707101822, + 0.039548926055431366, + 0.030029335990548134, + 0.01310608722269535, + 0.02476165071129799, + 0.02501686103641987, + 0.03738406300544739, + 0.006562311202287674, + 0.010158810764551163, + 0.010372540913522243, + 0.008667461574077606, + 0.008553032763302326, + 0.007422164548188448, + 0.007663260214030743, + 0.009801155887544155, + 0.008787603117525578, + 0.002845889888703823, + 0.029620133340358734, + 0.004778796341270208, + 0.0032210578210651875, + 0.003687762888148427, + 0.003332330146804452, + 0.02128206565976143, + 0.0037133977748453617, + 0.004960273392498493, + 0.015154640190303326, + 0.004740564618259668, + 0.006443437188863754, + 0.007558491080999374, + 0.007056775037199259, + 0.016815492883324623, + 0.005119689740240574, + 0.008158627897500992, + 0.014024103991687298, + 0.00547214737161994, + 0.0052951485849916935, + 0.007443524431437254, + 0.005550658795982599, + 0.007345281075686216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.03036779724061489 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372069075703621 + ], + "opposite_logprob_tokens": [ + -3.4087207317352295 + ], + "logprob_reward": -0.03372069075703621, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372069075703621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13014474511146545, + 0.025737926363945007, + 0.02524801529943943, + 0.02254006639122963, + 0.041354574263095856, + 0.04878748953342438, + 0.05045059695839882, + 0.4981093406677246, + 0.08680872619152069, + 0.0708184763789177 + ], + "credit_map": [ + 0.013654380105435848, + 0.013986950740218163, + 0.01816486567258835, + 0.016705652698874474, + 0.011978419497609138, + 0.009906182996928692, + 0.023667216300964355, + 0.022081071510910988, + 0.0028975741006433964, + 0.002567667281255126, + 0.004242677241563797, + 0.002599984174594283, + 0.002720421180129051, + 0.002941938117146492, + 0.003661103779450059, + 0.004106561187654734, + 0.004354394041001797, + 0.003411746583878994, + 0.0037566013634204865, + 0.002529613906517625, + 0.0024258822668343782, + 0.0023866610135883093, + 0.003362573217600584, + 0.003020545234903693, + 0.002495501423254609, + 0.002313476987183094, + 0.003457811428233981, + 0.0037023380864411592, + 0.0025280900299549103, + 0.002487374469637871, + 0.002750033512711525, + 0.0028054402209818363, + 0.0039375717751681805, + 0.004341901745647192, + 0.0064872694201767445, + 0.004930319730192423, + 0.004477014299482107, + 0.004543715622276068, + 0.004887031856924295, + 0.007749751675873995, + 0.003794633550569415, + 0.006250360980629921, + 0.006863794755190611, + 0.00746785756200552, + 0.0044346582144498825, + 0.006294607650488615, + 0.006225503981113434, + 0.007456073071807623, + 0.006695343181490898, + 0.005862508900463581, + 0.009633819572627544, + 0.006167876999825239, + 0.0036178312730044127, + 0.004866098053753376, + 0.005205953028053045, + 0.008401162922382355, + 0.05674520879983902, + 0.015406924299895763, + 0.20097289979457855, + 0.03135000541806221, + 0.010106844827532768, + 0.01655939780175686, + 0.021764323115348816, + 0.14520376920700073, + 0.0051816413179039955, + 0.007556265685707331, + 0.014578122645616531, + 0.008682155050337315, + 0.010174335911870003, + 0.012149697169661522, + 0.012625169940292835, + 0.01586134172976017, + 0.005373821128159761, + 0.009295486845076084, + 0.015992866829037666, + 0.007215452846139669, + 0.006377002224326134, + 0.009593143127858639, + 0.008470194414258003, + 0.008500506170094013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.03372069075703621 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009304723353125155 + ], + "opposite_logprob_tokens": [ + -7.000930309295654 + ], + "logprob_reward": -0.0009304723353125155, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009304723353125155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3143951892852783, + 0.050394702702760696, + 0.03261100873351097, + 0.08532262593507767, + 0.10694729536771774, + 0.04005827754735947, + 0.06353582441806793, + 0.1972273737192154, + 0.042995698750019073, + 0.06651195138692856 + ], + "credit_map": [ + 0.03740708902478218, + 0.033632952719926834, + 0.04131189361214638, + 0.03797461837530136, + 0.02759120985865593, + 0.02706899121403694, + 0.05517468973994255, + 0.05423373728990555, + 0.004657939542084932, + 0.005117850843816996, + 0.007796076592057943, + 0.00559188611805439, + 0.005276272539049387, + 0.0052429488860070705, + 0.008450159803032875, + 0.00826156884431839, + 0.002834062557667494, + 0.0034228165168315172, + 0.005331319756805897, + 0.005093718878924847, + 0.002528218785300851, + 0.0035288475919514894, + 0.005162801127880812, + 0.004709222819656134, + 0.009613598696887493, + 0.006793088745325804, + 0.016549060121178627, + 0.010509556159377098, + 0.007726677227765322, + 0.009335202164947987, + 0.012530235573649406, + 0.012265201658010483, + 0.01565631851553917, + 0.01431251224130392, + 0.0184034276753664, + 0.010054870508611202, + 0.009778181090950966, + 0.011221332475543022, + 0.011483140289783478, + 0.016037514433264732, + 0.0037986801471561193, + 0.004531307611614466, + 0.007902550511062145, + 0.00494250375777483, + 0.0035823481157422066, + 0.005090297199785709, + 0.004500231705605984, + 0.005710359662771225, + 0.006003463640809059, + 0.010047889314591885, + 0.009548220783472061, + 0.006966401357203722, + 0.00716016162186861, + 0.008311931043863297, + 0.007549190893769264, + 0.007948564365506172, + 0.02206900529563427, + 0.0076555474661290646, + 0.07599324733018875, + 0.013940381817519665, + 0.006587691139429808, + 0.008258095942437649, + 0.008014312013983727, + 0.05470911040902138, + 0.0034398145508021116, + 0.0037453556433320045, + 0.0075499447993934155, + 0.005921314004808664, + 0.004513384308665991, + 0.006018185522407293, + 0.005570798180997372, + 0.006236903369426727, + 0.004449724219739437, + 0.00828422512859106, + 0.014956560917198658, + 0.005633222404867411, + 0.0073789372108876705, + 0.009794463403522968, + 0.006447295192629099, + 0.00956752523779869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0009304723353125155 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003911683743353933 + ], + "opposite_logprob_tokens": [ + -7.875391006469727 + ], + "logprob_reward": -0.0003911683743353933, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003911683743353933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04481074586510658, + 0.025711320340633392, + 0.01605958677828312, + 0.020549369975924492, + 0.03703692927956581, + 0.01726689375936985, + 0.07245483994483948, + 0.5438784956932068, + 0.05854073166847229, + 0.16369113326072693 + ], + "credit_map": [ + 0.00426128227263689, + 0.004701938945800066, + 0.006851700600236654, + 0.00476911710575223, + 0.0038141882978379726, + 0.006160635966807604, + 0.005840195342898369, + 0.008411684073507786, + 0.002664742758497596, + 0.002964526182040572, + 0.004841881338506937, + 0.002618251834064722, + 0.0019803238101303577, + 0.0028418160509318113, + 0.003199917497113347, + 0.0045998599380254745, + 0.0017136912792921066, + 0.0015089469961822033, + 0.002872877288609743, + 0.0023017844650894403, + 0.001216283650137484, + 0.001707887277007103, + 0.0023815184831619263, + 0.0023565958254039288, + 0.0022988535929471254, + 0.0023333290591835976, + 0.003983475733548403, + 0.0028008134104311466, + 0.001596171990968287, + 0.001485130749642849, + 0.0030655355658382177, + 0.002986057661473751, + 0.003913671709597111, + 0.004063593689352274, + 0.005803296808153391, + 0.003975134342908859, + 0.0038232936058193445, + 0.004762740805745125, + 0.004198242910206318, + 0.006496951915323734, + 0.0017813495360314846, + 0.002327847993001342, + 0.0022723625879734755, + 0.0017780986381694674, + 0.002032012213021517, + 0.002333722310140729, + 0.0021456191316246986, + 0.0025958791375160217, + 0.005780271720141172, + 0.00973081961274147, + 0.010880878195166588, + 0.011293794959783554, + 0.004926699213683605, + 0.007775365374982357, + 0.011441857554018497, + 0.010625148192048073, + 0.05752035602927208, + 0.023241909220814705, + 0.21246425807476044, + 0.03435441479086876, + 0.014717694371938705, + 0.023724261671304703, + 0.020625894889235497, + 0.15722966194152832, + 0.004559303633868694, + 0.0057442947290837765, + 0.009926949627697468, + 0.004925471264868975, + 0.007007057312875986, + 0.008248959667980671, + 0.006183311343193054, + 0.011945378966629505, + 0.014671276323497295, + 0.02136574499309063, + 0.04525311291217804, + 0.01496356911957264, + 0.016928676515817642, + 0.021701516583561897, + 0.013588491827249527, + 0.015218734741210938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0003911683743353933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463838828727603 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.00463838828727603, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463838828727603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14918164908885956, + 0.033544860780239105, + 0.02907201275229454, + 0.016612259671092033, + 0.11934441328048706, + 0.11174559593200684, + 0.08853738009929657, + 0.20204277336597443, + 0.11556128412485123, + 0.13435783982276917 + ], + "credit_map": [ + 0.013710224069654942, + 0.014122461900115013, + 0.02083607390522957, + 0.019543327391147614, + 0.017597749829292297, + 0.014083236455917358, + 0.023489713668823242, + 0.025798853486776352, + 0.004936646204441786, + 0.0037015597335994244, + 0.006202089134603739, + 0.00411888025701046, + 0.0023120117839425802, + 0.0034096951130777597, + 0.004076550714671612, + 0.0047874306328594685, + 0.0022877694573253393, + 0.0028943479992449284, + 0.004308613948523998, + 0.004338533617556095, + 0.0026382561773061752, + 0.0026048491708934307, + 0.005755277816206217, + 0.004244363866746426, + 0.0018448509508743882, + 0.0021678577177226543, + 0.0025400600861757994, + 0.002009643241763115, + 0.0023736595176160336, + 0.0016166686546057463, + 0.0020403347443789244, + 0.002019185107201338, + 0.01025962270796299, + 0.01246415264904499, + 0.021359600126743317, + 0.012532052583992481, + 0.007979520596563816, + 0.01605978049337864, + 0.012934911996126175, + 0.025754766538739204, + 0.009742301888763905, + 0.01455269381403923, + 0.018562914803624153, + 0.015740055590867996, + 0.0067360312677919865, + 0.013813129626214504, + 0.014700798317790031, + 0.017897669225931168, + 0.00814694445580244, + 0.013147254474461079, + 0.01247025653719902, + 0.012414672411978245, + 0.010966856963932514, + 0.009560027159750462, + 0.01084925141185522, + 0.01098211482167244, + 0.020902059972286224, + 0.01009863056242466, + 0.07583573460578918, + 0.01227480173110962, + 0.008438647724688053, + 0.008998394943773746, + 0.009128978475928307, + 0.05636551231145859, + 0.006956218741834164, + 0.011360062286257744, + 0.018684787675738335, + 0.011417768895626068, + 0.015426691621541977, + 0.018554896116256714, + 0.016881104558706284, + 0.016279757022857666, + 0.008903012610971928, + 0.014602642506361008, + 0.029552694410085678, + 0.012760726734995842, + 0.010562262497842312, + 0.02002430334687233, + 0.017488108947873116, + 0.020464081317186356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.00463838828727603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, captured from a static, slightly high-angle perspective, meticulously frames a quiet, physical moment. The polished silver tip of a teaspoon hovers above a perfectly dark, polished countertop, reflecting like black water. A pool of thick, golden syrup, translucent and luminous, rests on the spoon, capturing the soft, warm glow of an invisible overhead light. This light creates a brilliant highlight along the spoon's curve and a distorted, warm reflection on the countertop below. From the spoon's edge, a viscous droplet slowly and deliberately elongates, stretched by gravity into a shimmering amber thread. It finally separates with a silent thud and collapses onto the surface into a perfect little pool of syrup. The impact creates subtle, silent ripples that expand outward with a mesmerizing slow-motion effect. The atmosphere is one of tranquil focus and sensual detail, emphasizing the stark contrast between the warm, glowing syrup and the cool, dark surface.", + "seed": 15720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_11/index_9180_seed_15720_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05604595830664039, + "vlm_logprob": -0.05604595830664039, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.05604595830664039, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24012809991836548 + ], + "opposite_logprob_tokens": [ + -1.6151280403137207 + ], + "logprob_reward": -0.24012809991836548, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24012809991836548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16880352795124054, + 0.04385992884635925, + 0.03670555353164673, + 0.10018576681613922, + 0.15080557763576508, + 0.24161547422409058, + 0.10796590149402618, + 0.03752252832055092, + 0.07697062194347382, + 0.0355650931596756 + ], + "credit_map": [ + 0.01842491887509823, + 0.016283538192510605, + 0.01517204288393259, + 0.00896503683179617, + 0.01762866973876953, + 0.01432801317423582, + 0.05397426709532738, + 0.02402704581618309, + 0.005579801741987467, + 0.005554910749197006, + 0.008542006835341454, + 0.0056999241933226585, + 0.00338441482745111, + 0.003253810340538621, + 0.0046414476819336414, + 0.0072036138735711575, + 0.0037390775978565216, + 0.002720345975831151, + 0.00901117641478777, + 0.0026481824461370707, + 0.0026619501877576113, + 0.0020086453296244144, + 0.005000412929803133, + 0.008915760554373264, + 0.011208615265786648, + 0.007676814217120409, + 0.02416173182427883, + 0.007862763479351997, + 0.00996385421603918, + 0.007583937142044306, + 0.00904882326722145, + 0.022679228335618973, + 0.030018344521522522, + 0.013405515812337399, + 0.029933447018265724, + 0.0176246240735054, + 0.009681560099124908, + 0.009096060879528522, + 0.013799303211271763, + 0.027246735990047455, + 0.04294155165553093, + 0.02326090633869171, + 0.04298057407140732, + 0.030009379610419273, + 0.019009307026863098, + 0.017809266224503517, + 0.02674250490963459, + 0.03886200860142708, + 0.022093066945672035, + 0.01058945246040821, + 0.018637770786881447, + 0.013092145323753357, + 0.005343725439161062, + 0.008116203360259533, + 0.011142326518893242, + 0.01895122230052948, + 0.005712208803743124, + 0.0028778172563761473, + 0.010543523356318474, + 0.004048228729516268, + 0.0021075699478387833, + 0.002067518886178732, + 0.004335432313382626, + 0.005830234382301569, + 0.0093918452039361, + 0.007141862064599991, + 0.016198012977838516, + 0.006904561538249254, + 0.008094784803688526, + 0.0061851562932133675, + 0.011347353458404541, + 0.011707046069204807, + 0.005079248920083046, + 0.0046332343481481075, + 0.0057684616185724735, + 0.0038507746066898108, + 0.004135288763791323, + 0.0031496488954871893, + 0.00428230594843626, + 0.004666132386773825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.24012809991836548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020617106929421425 + ], + "opposite_logprob_tokens": [ + -3.8956170082092285 + ], + "logprob_reward": -0.020617106929421425, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020617106929421425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3187125325202942, + 0.04105726629495621, + 0.041233181953430176, + 0.15732787549495697, + 0.04935979098081589, + 0.11787792295217514, + 0.09168749302625656, + 0.07917693257331848, + 0.05859090015292168, + 0.04497610777616501 + ], + "credit_map": [ + 0.02558373287320137, + 0.02842363342642784, + 0.02520175091922283, + 0.0186720322817564, + 0.03316774219274521, + 0.028645208105444908, + 0.10973222553730011, + 0.049286194145679474, + 0.0046433499082922935, + 0.0052395230159163475, + 0.006445647217333317, + 0.004430736880749464, + 0.003751527052372694, + 0.0032036416232585907, + 0.005312668159604073, + 0.008030171506106853, + 0.003413429716601968, + 0.002895333105698228, + 0.009492618031799793, + 0.0036545912735164165, + 0.0028088500257581472, + 0.0029644311871379614, + 0.00482224440202117, + 0.011181685142219067, + 0.020017893984913826, + 0.010005056858062744, + 0.03937666118144989, + 0.015610461123287678, + 0.01058040652424097, + 0.007522443309426308, + 0.01576995849609375, + 0.03844500333070755, + 0.009214108809828758, + 0.0046119699254632, + 0.005825686268508434, + 0.004130178596824408, + 0.004924167413264513, + 0.004652431700378656, + 0.007957139052450657, + 0.008044109679758549, + 0.02546241134405136, + 0.008124200627207756, + 0.024761341512203217, + 0.014741477556526661, + 0.006652051582932472, + 0.00790418591350317, + 0.010164974257349968, + 0.020067280158400536, + 0.016545521095395088, + 0.007856039330363274, + 0.017369603738188744, + 0.01224353164434433, + 0.005635808687657118, + 0.006023907568305731, + 0.010345513932406902, + 0.015667561441659927, + 0.012788909487426281, + 0.006299757864326239, + 0.020287372171878815, + 0.008328575640916824, + 0.005348402075469494, + 0.0033190378453582525, + 0.009691355749964714, + 0.013113531284034252, + 0.01016474049538374, + 0.0044824774377048016, + 0.013165587559342384, + 0.010227025486528873, + 0.003232759889215231, + 0.0028188033029437065, + 0.005765222012996674, + 0.008734283968806267, + 0.007145461160689592, + 0.004061334766447544, + 0.010959585197269917, + 0.004759558010846376, + 0.003958596847951412, + 0.0023784576915204525, + 0.004490844905376434, + 0.00722226919606328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.020617106929421425 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032086102291941643 + ], + "opposite_logprob_tokens": [ + -5.753208637237549 + ], + "logprob_reward": -0.0032086102291941643, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032086102291941643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11688444018363953, + 0.0390816256403923, + 0.041241440922021866, + 0.17434586584568024, + 0.30698704719543457, + 0.11841434240341187, + 0.06031317636370659, + 0.032983019948005676, + 0.047565195709466934, + 0.06218386068940163 + ], + "credit_map": [ + 0.010539157316088676, + 0.01094898022711277, + 0.010943619534373283, + 0.008061815053224564, + 0.01274689007550478, + 0.010314730927348137, + 0.03632960841059685, + 0.016999639570713043, + 0.004103171173483133, + 0.004536100663244724, + 0.0069873398169875145, + 0.003610756015405059, + 0.0038337160367518663, + 0.0034114159643650055, + 0.005890797823667526, + 0.006708324421197176, + 0.002985558472573757, + 0.004785554483532906, + 0.006510091945528984, + 0.003607213729992509, + 0.003569406922906637, + 0.0035435864701867104, + 0.008234352804720402, + 0.008005674928426743, + 0.01698216237127781, + 0.015381291508674622, + 0.030076680704951286, + 0.01690554991364479, + 0.022618034854531288, + 0.013748723082244396, + 0.02764657326042652, + 0.0309868473559618, + 0.04512466862797737, + 0.038209468126297, + 0.051566701382398605, + 0.03613228723406792, + 0.022299818694591522, + 0.027384711429476738, + 0.04158946871757507, + 0.0446799211204052, + 0.024387728422880173, + 0.00842601340264082, + 0.022370245307683945, + 0.01481233537197113, + 0.007324506528675556, + 0.008562975563108921, + 0.012812190689146519, + 0.019718345254659653, + 0.013665781356394291, + 0.005515738390386105, + 0.011836164630949497, + 0.006623624823987484, + 0.0029829551931470633, + 0.0041216579265892506, + 0.004687701351940632, + 0.01087955292314291, + 0.004628817550837994, + 0.0035338883753865957, + 0.005991870071738958, + 0.003704406786710024, + 0.0031287746969610453, + 0.003021629760041833, + 0.004186377394944429, + 0.0047872550785541534, + 0.005797678604722023, + 0.004964232910424471, + 0.007632109802216291, + 0.0055539789609611034, + 0.005461364518851042, + 0.004063218366354704, + 0.007348231505602598, + 0.006744378246366978, + 0.010594386607408524, + 0.005837015341967344, + 0.013930443674325943, + 0.0074209789745509624, + 0.005591094493865967, + 0.003330135950818658, + 0.006035937462002039, + 0.00944386888295412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.0032086102291941643 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864406768232584 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.00864406768232584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864406768232584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12888282537460327, + 0.06336479634046555, + 0.07122094184160233, + 0.30290985107421875, + 0.16860823333263397, + 0.06359248608350754, + 0.036501821130514145, + 0.03946823626756668, + 0.04961945116519928, + 0.07583139836788177 + ], + "credit_map": [ + 0.01163935475051403, + 0.013167510740458965, + 0.015092428773641586, + 0.009610245935618877, + 0.014617225155234337, + 0.01231207326054573, + 0.03484000265598297, + 0.01760396920144558, + 0.0074126445688307285, + 0.0076432297937572, + 0.00942624919116497, + 0.004773132503032684, + 0.007171702571213245, + 0.006566164083778858, + 0.010518287308514118, + 0.009853393770754337, + 0.005222688894718885, + 0.004962072242051363, + 0.01589951291680336, + 0.006260506343096495, + 0.005066681653261185, + 0.005543211475014687, + 0.00952000543475151, + 0.018746260553598404, + 0.03157126531004906, + 0.019090639427304268, + 0.0789954736828804, + 0.025925399735569954, + 0.018680782988667488, + 0.02073393203318119, + 0.030038295313715935, + 0.07787404954433441, + 0.0347583070397377, + 0.01721910759806633, + 0.03297186270356178, + 0.02041085995733738, + 0.009851089678704739, + 0.009175101295113564, + 0.01604704186320305, + 0.028174862265586853, + 0.013256540521979332, + 0.004519407171756029, + 0.01311609335243702, + 0.008101564832031727, + 0.004658226855099201, + 0.004042033106088638, + 0.005052844062447548, + 0.010845773853361607, + 0.0033993511460721493, + 0.0038109158631414175, + 0.004360171500593424, + 0.004536949098110199, + 0.004866224713623524, + 0.003161903005093336, + 0.007059548050165176, + 0.00530675845220685, + 0.0065966383554041386, + 0.0030910188797861338, + 0.010124921798706055, + 0.004334162455052137, + 0.002768472768366337, + 0.002204209566116333, + 0.004242718685418367, + 0.006106096785515547, + 0.005426717922091484, + 0.004269057884812355, + 0.01042287889868021, + 0.006410770583897829, + 0.0048110270872712135, + 0.004409213550388813, + 0.006664082407951355, + 0.007205704692751169, + 0.012564055621623993, + 0.006916983984410763, + 0.018423963338136673, + 0.009238634258508682, + 0.0053983163088560104, + 0.003991378005594015, + 0.007366551551967859, + 0.011931517161428928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.00864406768232584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a teaspoon tip hovering over a dark, polished countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective static and slightly high-angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pool of thick, golden, translucent syrup visible on the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a viscous droplet elongate into a shimmering amber thread and separate from the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the droplet's fall and the resulting ripples shown in a slow-motion effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631906773895025 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631906773895025, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631906773895025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1488771140575409, + 0.05811573565006256, + 0.07002905756235123, + 0.1111542209982872, + 0.1375839114189148, + 0.08692616969347, + 0.058051157742738724, + 0.06524374336004257, + 0.17854280769824982, + 0.08547604084014893 + ], + "credit_map": [ + 0.017147649079561234, + 0.0176620464771986, + 0.028042929247021675, + 0.01630576327443123, + 0.01563510298728943, + 0.012369387783110142, + 0.021783089265227318, + 0.01993115060031414, + 0.007519411388784647, + 0.006051895674318075, + 0.009833204559981823, + 0.005422420799732208, + 0.007305094040930271, + 0.005007781554013491, + 0.007864069193601608, + 0.009111864492297173, + 0.005652585066854954, + 0.004721378907561302, + 0.015898549929261208, + 0.006018075626343489, + 0.004772015381604433, + 0.005213605705648661, + 0.008841782808303833, + 0.01891106367111206, + 0.013040493242442608, + 0.01030923891812563, + 0.02355801686644554, + 0.011326470412313938, + 0.008498228155076504, + 0.008372561074793339, + 0.012890612706542015, + 0.023158608004450798, + 0.01751135103404522, + 0.017740461975336075, + 0.014093217439949512, + 0.014929590746760368, + 0.018240170553326607, + 0.01714242622256279, + 0.020051762461662292, + 0.01787494495511055, + 0.017916042357683182, + 0.007290775887668133, + 0.015364494174718857, + 0.00980892963707447, + 0.00740213505923748, + 0.006907052360475063, + 0.00848388485610485, + 0.013752865605056286, + 0.009402314200997353, + 0.005985253024846315, + 0.008093434385955334, + 0.004988952074199915, + 0.006310469936579466, + 0.0056343236938118935, + 0.008827290497720242, + 0.008809122256934643, + 0.009145854972302914, + 0.005400217603892088, + 0.017104675993323326, + 0.006533400155603886, + 0.006433659233152866, + 0.0035304210614413023, + 0.007193023804575205, + 0.00990249402821064, + 0.022037511691451073, + 0.013638269156217575, + 0.03870124742388725, + 0.018636995926499367, + 0.015517499297857285, + 0.01452934555709362, + 0.025465676560997963, + 0.030016275122761726, + 0.010987240821123123, + 0.009195918217301369, + 0.015874849632382393, + 0.010073970071971416, + 0.010144831612706184, + 0.006919957231730223, + 0.010188452899456024, + 0.012090831995010376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.007631906773895025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_12/index_tensor([2138, 2138])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_12/index_tensor([2138, 2138])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d100f96c46b8390ec009c0f384ef5d8f434f9692 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_12/index_tensor([2138, 2138])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angled medium shot captures a scene rendered in a lush, painterly style reminiscent of Studio Ghibli, focusing on an elderly, anthropomorphic white duck standing defiantly within a vast, ancient vault. Wearing an elaborate black silk top hat and a pair of small, gold-rimmed glasses perched on his orange beak, his expression is both resolute and fiercely protective. One of his wings is rigidly extended, forming a protective shield over his three furry yellow nephews, who huddle in fear beneath it, their wide, innocent eyes fixed on an unseen threat. The vault itself is a cavernous space constructed of massive, rough-hewn stone blocks. Raw light, emanating from the ethereal, luminous moss clinging to the walls and floor, casts a soft, magical glow in hues of emerald and sapphire. Piles of treasure—glittering gold coins, ancient leather-bound books, and mysterious, glowing artifacts—scatter the background, their details softened by a hand-painted aesthetic. The atmosphere was thick with tension as a huge, menacing shadow, indistinct in form, crept monstrously and slowly but relentlessly over the distant stone wall, its deliberate advance disturbing millions of motes of dust that had previously danced gently in the soft light and now causing them to swirl in agitated, chaotic patterns.", + "seed": 6826, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_12/index_2138_seed_6826_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08379150219261647, + "vlm_logprob": -0.08379150219261647, + "vlm_yes_no_margin": 3.524999964237213, + "vlm_token_credit": -0.08379150219261647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023634381592273712 + ], + "opposite_logprob_tokens": [ + -3.773634433746338 + ], + "logprob_reward": -0.023634381592273712, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023634381592273712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2035035640001297, + 0.05407659709453583, + 0.02242239937186241, + 0.04611900821328163, + 0.0819174200296402, + 0.1847778707742691, + 0.07778789848089218, + 0.11170501261949539, + 0.11992772668600082, + 0.09776244312524796 + ], + "credit_map": [ + 0.02024073712527752, + 0.015730753540992737, + 0.029484884813427925, + 0.024526188150048256, + 0.014418408274650574, + 0.0178048238158226, + 0.055295608937740326, + 0.026002168655395508, + 0.006359592080116272, + 0.005238641053438187, + 0.0068571376614272594, + 0.006485306657850742, + 0.005433248821645975, + 0.007224315777420998, + 0.011436586268246174, + 0.005041772034019232, + 0.0019675546791404486, + 0.0018940229201689363, + 0.002885535592213273, + 0.002448471263051033, + 0.001826822874136269, + 0.0016812026733532548, + 0.006809436250478029, + 0.0029093557968735695, + 0.004286872688680887, + 0.003949705511331558, + 0.005636048968881369, + 0.005614597350358963, + 0.003598384326323867, + 0.0038038643542677164, + 0.013845560140907764, + 0.005383977200835943, + 0.007218345999717712, + 0.007965008728206158, + 0.010769374668598175, + 0.010791024193167686, + 0.006370452232658863, + 0.0063111623749136925, + 0.023144161328673363, + 0.009347907267510891, + 0.029288101941347122, + 0.016476944088935852, + 0.032830022275447845, + 0.01432703249156475, + 0.025690466165542603, + 0.030102264136075974, + 0.01882358267903328, + 0.017239470034837723, + 0.013409880921244621, + 0.006733936723321676, + 0.00782707054167986, + 0.005508787930011749, + 0.009388431906700134, + 0.010712671093642712, + 0.011519049294292927, + 0.012688076123595238, + 0.019027065485715866, + 0.01185967680066824, + 0.010576693341135979, + 0.010432065464556217, + 0.02384890429675579, + 0.009126023389399052, + 0.017691591754555702, + 0.009142995811998844, + 0.031223544850945473, + 0.008925151079893112, + 0.015653299167752266, + 0.008937111124396324, + 0.011293877847492695, + 0.01613505929708481, + 0.011666936799883842, + 0.016092753037810326, + 0.022547980770468712, + 0.00829133577644825, + 0.007494459860026836, + 0.00678235711529851, + 0.012536274269223213, + 0.02245454490184784, + 0.010475491173565388, + 0.007180015090852976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.023634381592273712 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290403187274933 + ], + "opposite_logprob_tokens": [ + -3.7732903957366943 + ], + "logprob_reward": -0.023290403187274933, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290403187274933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1412142664194107, + 0.31670865416526794, + 0.10999976098537445, + 0.0433233268558979, + 0.02583947777748108, + 0.07229677587747574, + 0.10560709983110428, + 0.06894059479236603, + 0.06362903118133545, + 0.05244094133377075 + ], + "credit_map": [ + 0.016718657687306404, + 0.01526147872209549, + 0.017127826809883118, + 0.01242036372423172, + 0.013202172704041004, + 0.01929386705160141, + 0.027681797742843628, + 0.019508108496665955, + 0.034668829292058945, + 0.026804767549037933, + 0.04259186238050461, + 0.03602779284119606, + 0.02674277499318123, + 0.03976154699921608, + 0.07368545234203339, + 0.036425668746232986, + 0.012737731449306011, + 0.011609386652708054, + 0.012879817746579647, + 0.012361896224319935, + 0.009219507686793804, + 0.014156123623251915, + 0.023656725883483887, + 0.013378577306866646, + 0.005585657898336649, + 0.006163678131997585, + 0.004696553107351065, + 0.0043169232085347176, + 0.004075982607901096, + 0.0055983103811740875, + 0.007651680614799261, + 0.005234546028077602, + 0.003969920799136162, + 0.0021759786177426577, + 0.003334412816911936, + 0.0022323504090309143, + 0.002172005595639348, + 0.004317287355661392, + 0.004736904054880142, + 0.0029006227850914, + 0.012023876421153545, + 0.008077907375991344, + 0.01121276244521141, + 0.004620963707566261, + 0.009746266528964043, + 0.012917278334498405, + 0.006878009997308254, + 0.006819718051701784, + 0.014834294095635414, + 0.009259350597858429, + 0.011565362103283405, + 0.008017919957637787, + 0.010947559028863907, + 0.0178435817360878, + 0.013188156299293041, + 0.01995088905096054, + 0.00784343108534813, + 0.007087577134370804, + 0.006425399333238602, + 0.007848984561860561, + 0.010129213333129883, + 0.007538878824561834, + 0.01011693850159645, + 0.011950171552598476, + 0.005843364167958498, + 0.005667390301823616, + 0.014051253907382488, + 0.0058333310298621655, + 0.00809077825397253, + 0.008437405340373516, + 0.010332082398235798, + 0.005373433232307434, + 0.010468163527548313, + 0.006414161529392004, + 0.0038972217589616776, + 0.004656359553337097, + 0.007879543118178844, + 0.006813359912484884, + 0.006737015210092068, + 0.005575121380388737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.023290403187274933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052458494901657104 + ], + "opposite_logprob_tokens": [ + -5.255245685577393 + ], + "logprob_reward": -0.0052458494901657104, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052458494901657104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11901729553937912, + 0.2296014279127121, + 0.04927636310458183, + 0.02738489769399166, + 0.15020573139190674, + 0.08136992156505585, + 0.0935429260134697, + 0.11191816627979279, + 0.05675626918673515, + 0.08092706650495529 + ], + "credit_map": [ + 0.013538826256990433, + 0.01039879396557808, + 0.01740766502916813, + 0.012636568397283554, + 0.014664447866380215, + 0.014456802047789097, + 0.021100612357258797, + 0.014813586138188839, + 0.023778697475790977, + 0.020831579342484474, + 0.032032933086156845, + 0.02758658491075039, + 0.020748872309923172, + 0.02321062795817852, + 0.05661537125706673, + 0.024796750396490097, + 0.005071691703051329, + 0.005660189315676689, + 0.0062584117986261845, + 0.006095347926020622, + 0.0045887380838394165, + 0.004320988431572914, + 0.012117389589548111, + 0.005163604859262705, + 0.0038817196618765593, + 0.002272403333336115, + 0.003546491963788867, + 0.002130665350705385, + 0.002602849155664444, + 0.005497890990227461, + 0.003839844139292836, + 0.0036130340304225683, + 0.01908312737941742, + 0.01046060025691986, + 0.02008763886988163, + 0.01970851793885231, + 0.009693306870758533, + 0.01943974941968918, + 0.03492703661322594, + 0.016805747523903847, + 0.013617006130516529, + 0.008536248467862606, + 0.012526145204901695, + 0.005244338419288397, + 0.010503046214580536, + 0.015443314798176289, + 0.007324995473027229, + 0.008174828253686428, + 0.01926223561167717, + 0.007258285768330097, + 0.012216969393193722, + 0.006482032127678394, + 0.005107227712869644, + 0.01656065136194229, + 0.006126494146883488, + 0.02052902616560459, + 0.011095005087554455, + 0.010657746344804764, + 0.01596936397254467, + 0.009534083306789398, + 0.015546142123639584, + 0.010647005401551723, + 0.01905045099556446, + 0.019418368116021156, + 0.012805919162929058, + 0.004279095213860273, + 0.015493299812078476, + 0.004398984368890524, + 0.0033746492117643356, + 0.004661818500608206, + 0.005945312790572643, + 0.005797193851321936, + 0.012191671878099442, + 0.011571826413273811, + 0.010344757698476315, + 0.005933328531682491, + 0.007814835757017136, + 0.018554026260972023, + 0.007478667888790369, + 0.007037957664579153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0052458494901657104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182972446084023 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.018182972446084023, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018182972446084023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08330859988927841, + 0.15567930042743683, + 0.04025346413254738, + 0.060078348964452744, + 0.06030796840786934, + 0.04257769137620926, + 0.04535339027643204, + 0.09516821056604385, + 0.3230801522731781, + 0.09419289976358414 + ], + "credit_map": [ + 0.008646473288536072, + 0.00814356654882431, + 0.010393570177257061, + 0.009771977551281452, + 0.009754743427038193, + 0.00909983180463314, + 0.018339697271585464, + 0.009158737026154995, + 0.01480835396796465, + 0.013634889386594296, + 0.021352820098400116, + 0.01935645379126072, + 0.013714928179979324, + 0.014894377440214157, + 0.04096967726945877, + 0.016947798430919647, + 0.003107671858742833, + 0.00417473865672946, + 0.005865808110684156, + 0.0051975613459944725, + 0.003749644849449396, + 0.0027332811150699854, + 0.01086973026394844, + 0.004555028397589922, + 0.006754874251782894, + 0.005508036818355322, + 0.007094307336956263, + 0.007048006169497967, + 0.004479808267205954, + 0.006459727883338928, + 0.016370922327041626, + 0.006362663581967354, + 0.007149299141019583, + 0.004153414163738489, + 0.007510415278375149, + 0.008649853058159351, + 0.005282595753669739, + 0.00905383937060833, + 0.010393418371677399, + 0.008115136064589024, + 0.006188569590449333, + 0.005481662694364786, + 0.006152489222586155, + 0.003939785063266754, + 0.004778191912919283, + 0.006090343464165926, + 0.005755046848207712, + 0.004191603511571884, + 0.005948174744844437, + 0.0035273281391710043, + 0.007968416437506676, + 0.003632650477811694, + 0.0034339313860982656, + 0.007252865005284548, + 0.007973022758960724, + 0.005617000628262758, + 0.01081888284534216, + 0.01066077034920454, + 0.011240352876484394, + 0.008865995332598686, + 0.014978195540606976, + 0.007774726487696171, + 0.019214080646634102, + 0.011615206487476826, + 0.08829325437545776, + 0.02192547917366028, + 0.11164405196905136, + 0.022191008552908897, + 0.01345273107290268, + 0.0220942385494709, + 0.015567012131214142, + 0.02791239321231842, + 0.012003631331026554, + 0.013147151097655296, + 0.010710942558944225, + 0.007642124313861132, + 0.00916154496371746, + 0.020238451659679413, + 0.012229012325406075, + 0.009060041047632694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.018182972446084023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34860390424728394 + ], + "opposite_logprob_tokens": [ + -1.2236039638519287 + ], + "logprob_reward": -0.34860390424728394, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34860390424728394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14546863734722137, + 0.13592751324176788, + 0.0811944231390953, + 0.031322456896305084, + 0.043171025812625885, + 0.12734106183052063, + 0.08895797282457352, + 0.15741707384586334, + 0.0729563906788826, + 0.11624342948198318 + ], + "credit_map": [ + 0.01670175790786743, + 0.011404335498809814, + 0.02076563611626625, + 0.017248429358005524, + 0.015397890470921993, + 0.015307714231312275, + 0.03164130821824074, + 0.017001571133732796, + 0.014067488722503185, + 0.012365144677460194, + 0.01843748241662979, + 0.015665743499994278, + 0.012557236477732658, + 0.014511224813759327, + 0.03239298611879349, + 0.015930207446217537, + 0.008030050434172153, + 0.0072036744095385075, + 0.010012284852564335, + 0.009812328033149242, + 0.006732587236911058, + 0.007821413688361645, + 0.021652914583683014, + 0.009929169900715351, + 0.0030978836584836245, + 0.004006898496299982, + 0.004780532326549292, + 0.0034394427202641964, + 0.0030826714355498552, + 0.0023071744944900274, + 0.007365977857261896, + 0.0032418733462691307, + 0.008009552955627441, + 0.0037368983030319214, + 0.008795913308858871, + 0.003637909423559904, + 0.0035658469423651695, + 0.006815518718212843, + 0.00396564370021224, + 0.004643742460757494, + 0.02016221545636654, + 0.0110758226364851, + 0.02227724902331829, + 0.009805665351450443, + 0.015775736421346664, + 0.021778017282485962, + 0.014104155823588371, + 0.012362204492092133, + 0.01324895117431879, + 0.007528092246502638, + 0.015116717666387558, + 0.010516058653593063, + 0.00907589215785265, + 0.013419142924249172, + 0.010417254641652107, + 0.009635860100388527, + 0.01681329868733883, + 0.015983449295163155, + 0.019765272736549377, + 0.012511915527284145, + 0.02461789734661579, + 0.013096504844725132, + 0.0301081333309412, + 0.024520600214600563, + 0.018540851771831512, + 0.00627516582608223, + 0.005975443869829178, + 0.005166357848793268, + 0.006747369654476643, + 0.012682556174695492, + 0.006471199449151754, + 0.01109744980931282, + 0.020706895738840103, + 0.009244868531823158, + 0.007349120453000069, + 0.00863712653517723, + 0.008490714244544506, + 0.03387802839279175, + 0.015795258805155754, + 0.012141418643295765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.34860390424728394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled medium shot captures a scene rendered in a lush, painterly style reminiscent of Studio Ghibli, focusing on an elderly, anthropomorphic white duck standing defiantly within a vast, ancient vault. Wearing an elaborate black silk top hat and a pair of small, gold-rimmed glasses perched on his orange beak, his expression is both resolute and fiercely protective. One of his wings is rigidly extended, forming a protective shield over his three furry yellow nephews, who huddle in fear beneath it, their wide, innocent eyes fixed on an unseen threat. The vault itself is a cavernous space constructed of massive, rough-hewn stone blocks. Raw light, emanating from the ethereal, luminous moss clinging to the walls and floor, casts a soft, magical glow in hues of emerald and sapphire. Piles of treasure—glittering gold coins, ancient leather-bound books, and mysterious, glowing artifacts—scatter the background, their details softened by a hand-painted aesthetic. The atmosphere was thick with tension as a huge, menacing shadow, indistinct in form, crept monstrously and slowly but relentlessly over the distant stone wall, its deliberate advance disturbing millions of motes of dust that had previously danced gently in the soft light and now causing them to swirl in agitated, chaotic patterns.", + "seed": 6827, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_12/index_2138_seed_6827_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06657178308814764, + "vlm_logprob": -0.06657178308814764, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.06657178308814764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02368246391415596 + ], + "opposite_logprob_tokens": [ + -3.7736823558807373 + ], + "logprob_reward": -0.02368246391415596, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02368246391415596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44281497597694397, + 0.15062670409679413, + 0.06239709258079529, + 0.04524943232536316, + 0.05169174447655678, + 0.05400969088077545, + 0.045210741460323334, + 0.03193198889493942, + 0.07283578813076019, + 0.04323180019855499 + ], + "credit_map": [ + 0.0837700366973877, + 0.060583896934986115, + 0.0353972353041172, + 0.026329193264245987, + 0.03578956052660942, + 0.06977865099906921, + 0.03442296385765076, + 0.09674343466758728, + 0.026918450370430946, + 0.02288690395653248, + 0.011231652460992336, + 0.009284449741244316, + 0.02643231861293316, + 0.01961258426308632, + 0.010486096143722534, + 0.023774269968271255, + 0.011489398777484894, + 0.009011712856590748, + 0.005434234626591206, + 0.00448672054335475, + 0.008262524381279945, + 0.00869041122496128, + 0.005770792253315449, + 0.009251297451555729, + 0.006799506954848766, + 0.005891667678952217, + 0.0037164806853979826, + 0.003461234737187624, + 0.009712258353829384, + 0.005764673929661512, + 0.004763063509017229, + 0.005140543915331364, + 0.00894936453551054, + 0.006243362091481686, + 0.003948214929550886, + 0.003470623167231679, + 0.006337370723485947, + 0.0074916728772223, + 0.0038803003262728453, + 0.01137083675712347, + 0.007162075024098158, + 0.005197932478040457, + 0.005097699351608753, + 0.005088114645332098, + 0.010344967246055603, + 0.005668625235557556, + 0.00821360107511282, + 0.0072366767562925816, + 0.008604876697063446, + 0.0037451067473739386, + 0.00503714382648468, + 0.004179707728326321, + 0.005389511119574308, + 0.005366804543882608, + 0.005698388442397118, + 0.007189204916357994, + 0.007955464534461498, + 0.0036055257078260183, + 0.0025290457997471094, + 0.002476656809449196, + 0.0039932504296302795, + 0.0036709357518702745, + 0.003162788227200508, + 0.004538321867585182, + 0.016415691003203392, + 0.010459105484187603, + 0.006968104280531406, + 0.006058134604245424, + 0.005361135117709637, + 0.01021640282124281, + 0.005398968700319529, + 0.011958256363868713, + 0.0062406244687736034, + 0.004877751227468252, + 0.0042176926508545876, + 0.0048045688308775425, + 0.007420791313052177, + 0.005555715411901474, + 0.004500932991504669, + 0.005613722838461399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02368246391415596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865318909287453 + ], + "opposite_logprob_tokens": [ + -3.0486531257629395 + ], + "logprob_reward": -0.04865318909287453, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865318909287453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41251933574676514, + 0.15959109365940094, + 0.11697039008140564, + 0.051616132259368896, + 0.055548783391714096, + 0.0412001796066761, + 0.06374141573905945, + 0.028116511180996895, + 0.03350243344902992, + 0.037193700671195984 + ], + "credit_map": [ + 0.08365117758512497, + 0.05796177685260773, + 0.040845759212970734, + 0.027738064527511597, + 0.02049393579363823, + 0.0665677934885025, + 0.02893829718232155, + 0.08632255345582962, + 0.030513327568769455, + 0.025349210947752, + 0.016029126942157745, + 0.01040528155863285, + 0.013111833482980728, + 0.02355971559882164, + 0.010892444290220737, + 0.029730157926678658, + 0.026839474216103554, + 0.017685940489172935, + 0.010127360001206398, + 0.0067649721167981625, + 0.005998781882226467, + 0.017531754449009895, + 0.006410655565559864, + 0.025611456483602524, + 0.011086185462772846, + 0.007615539710968733, + 0.004773626569658518, + 0.003081447444856167, + 0.003365825628861785, + 0.007826808840036392, + 0.0038153447676450014, + 0.010051354765892029, + 0.00917085725814104, + 0.007551873102784157, + 0.005081662908196449, + 0.003629310056567192, + 0.006911966484040022, + 0.007676403038203716, + 0.004230744671076536, + 0.011295965872704983, + 0.006204314064234495, + 0.0054876115173101425, + 0.0045888228341937065, + 0.0033847526647150517, + 0.00575067987665534, + 0.00550455879420042, + 0.004409139044582844, + 0.005870301276445389, + 0.01173883955925703, + 0.00878144335001707, + 0.00485535804182291, + 0.0034552644938230515, + 0.006066764704883099, + 0.009657865390181541, + 0.004007674288004637, + 0.01517820730805397, + 0.003621050389483571, + 0.002977035939693451, + 0.003114905906841159, + 0.002755583729594946, + 0.004822473041713238, + 0.0035053216852247715, + 0.0030563385225832462, + 0.0042638024315238, + 0.00681358901783824, + 0.004630000796169043, + 0.003623884404078126, + 0.0029280693270266056, + 0.0030899769626557827, + 0.004412263166159391, + 0.0031372939702123404, + 0.004867355804890394, + 0.006017441861331463, + 0.0046512107364833355, + 0.0039424896240234375, + 0.002745207631960511, + 0.005232033785432577, + 0.0047874306328594685, + 0.003232653019949794, + 0.006585231516510248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04865318909287453 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026332611218094826 + ], + "opposite_logprob_tokens": [ + -3.6513326168060303 + ], + "logprob_reward": -0.026332611218094826, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026332611218094826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42084282636642456, + 0.1862063705921173, + 0.11081893742084503, + 0.0668688714504242, + 0.05055630952119827, + 0.036677539348602295, + 0.04973948746919632, + 0.02389645390212536, + 0.029325436800718307, + 0.025067834183573723 + ], + "credit_map": [ + 0.08849658817052841, + 0.05936172604560852, + 0.04221276193857193, + 0.028820479288697243, + 0.024894610047340393, + 0.0627060979604721, + 0.032358963042497635, + 0.0819915235042572, + 0.034782152622938156, + 0.03528037667274475, + 0.017010914161801338, + 0.01433559413999319, + 0.01391652598977089, + 0.028199294582009315, + 0.014607599005103111, + 0.028073905035853386, + 0.02137005142867565, + 0.018918892368674278, + 0.010107181966304779, + 0.008586765266954899, + 0.010420954786241055, + 0.016076374799013138, + 0.008522083051502705, + 0.01681661605834961, + 0.015382570214569569, + 0.009644321165978909, + 0.005593337118625641, + 0.005852969363331795, + 0.005457284394651651, + 0.008742911741137505, + 0.006713107228279114, + 0.009482366964221, + 0.008146204985678196, + 0.005503253545612097, + 0.004339995328336954, + 0.004331482574343681, + 0.008790283463895321, + 0.0063356561586260796, + 0.004709410015493631, + 0.008400016464293003, + 0.006467258092015982, + 0.004808346275240183, + 0.004374995827674866, + 0.004824929870665073, + 0.0032538319937884808, + 0.004615082871168852, + 0.003910685423761606, + 0.004422403872013092, + 0.008554527536034584, + 0.008678039535880089, + 0.005838382989168167, + 0.00467330776154995, + 0.0052674030885100365, + 0.006684319116175175, + 0.00398468691855669, + 0.006058814004063606, + 0.006532970350235701, + 0.0029905366245657206, + 0.0025043326895684004, + 0.001909898011945188, + 0.002318373415619135, + 0.0031087996903806925, + 0.0018400289118289948, + 0.0026915099006146193, + 0.004335932433605194, + 0.004510197322815657, + 0.0032795665320008993, + 0.002814704552292824, + 0.0040877764113247395, + 0.0034037402365356684, + 0.002965017920359969, + 0.003928500227630138, + 0.004865147639065981, + 0.0029408440459519625, + 0.0028041317127645016, + 0.0025574921164661646, + 0.0027072641532868147, + 0.0028762670699507, + 0.0027197825256735086, + 0.003596901660785079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026332611218094826 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633785881102085 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633785881102085, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633785881102085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.395995557308197, + 0.14340730011463165, + 0.09026145935058594, + 0.044621482491493225, + 0.06318355351686478, + 0.05474051460623741, + 0.052574075758457184, + 0.06020992249250412, + 0.04305548965930939, + 0.05195062980055809 + ], + "credit_map": [ + 0.09694427996873856, + 0.041474368423223495, + 0.04577074199914932, + 0.03134783357381821, + 0.034020572900772095, + 0.05626664310693741, + 0.02872764877974987, + 0.06144346669316292, + 0.03352085128426552, + 0.0171146709471941, + 0.016506323590874672, + 0.0116331297904253, + 0.010977218858897686, + 0.01887192390859127, + 0.011235477402806282, + 0.0235477052628994, + 0.019605480134487152, + 0.01053701899945736, + 0.010707102715969086, + 0.008422951214015484, + 0.009178122505545616, + 0.010794874280691147, + 0.006948923692107201, + 0.014066983945667744, + 0.006399678997695446, + 0.004196677822619677, + 0.00532873347401619, + 0.005071577150374651, + 0.005309344734996557, + 0.005379255395382643, + 0.004704768769443035, + 0.008231446146965027, + 0.011846140027046204, + 0.00712041649967432, + 0.007102973759174347, + 0.005732748657464981, + 0.007672181818634272, + 0.007485730107873678, + 0.005470539443194866, + 0.010752820409834385, + 0.008140443824231625, + 0.009002801962196827, + 0.00747864181175828, + 0.006240545306354761, + 0.006246115081012249, + 0.006291098892688751, + 0.005158658139407635, + 0.006182207725942135, + 0.010296131484210491, + 0.005580609198659658, + 0.005214138422161341, + 0.00412584887817502, + 0.005106874275952578, + 0.007185227703303099, + 0.00393472658470273, + 0.011130520142614841, + 0.010470381937921047, + 0.007520982529968023, + 0.005097610410302877, + 0.005249721463769674, + 0.009000447578728199, + 0.007290728855878115, + 0.005709717981517315, + 0.009870334528386593, + 0.006662019528448582, + 0.0062073045410215855, + 0.004803212825208902, + 0.004063508938997984, + 0.005493605975061655, + 0.0054170782677829266, + 0.003951152320951223, + 0.006457606330513954, + 0.00944810826331377, + 0.007970797829329967, + 0.005165745969861746, + 0.004462163895368576, + 0.007523639127612114, + 0.006359583232551813, + 0.004598828963935375, + 0.006421761121600866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008633785881102085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255568653345108 + ], + "opposite_logprob_tokens": [ + -1.6005568504333496 + ], + "logprob_reward": -0.2255568653345108, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255568653345108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4936874210834503, + 0.1245659813284874, + 0.07019075751304626, + 0.0522555410861969, + 0.052346762269735336, + 0.027412617579102516, + 0.04244856908917427, + 0.058065932244062424, + 0.03567381203174591, + 0.04335258901119232 + ], + "credit_map": [ + 0.1285477876663208, + 0.07265651971101761, + 0.04928883537650108, + 0.03586290404200554, + 0.03347182646393776, + 0.07529956102371216, + 0.03452187776565552, + 0.06403811275959015, + 0.028232552111148834, + 0.021315844729542732, + 0.013494816608726978, + 0.011160720139741898, + 0.009121822193264961, + 0.01718873903155327, + 0.008591665886342525, + 0.015459822490811348, + 0.015125262551009655, + 0.010166890919208527, + 0.00691619049757719, + 0.006659294944256544, + 0.010030350647866726, + 0.008584470488131046, + 0.0050067906267941, + 0.007701511029154062, + 0.011943468824028969, + 0.00936884805560112, + 0.0046417261473834515, + 0.004868621472269297, + 0.0036079720593988895, + 0.006555257365107536, + 0.0047349631786346436, + 0.006534678861498833, + 0.007683135569095612, + 0.005966509692370892, + 0.005062730051577091, + 0.004761914722621441, + 0.009788043797016144, + 0.006217189133167267, + 0.005110123660415411, + 0.007757116574794054, + 0.006096600089222193, + 0.00434210617095232, + 0.003093328792601824, + 0.0025930723641067743, + 0.0018208639230579138, + 0.0037964140065014362, + 0.0025814492255449295, + 0.003088784171268344, + 0.005157004110515118, + 0.004324223380535841, + 0.003285839920863509, + 0.003978777211159468, + 0.007076159585267305, + 0.0057444446720182896, + 0.004952483810484409, + 0.007929635234177113, + 0.013155707158148289, + 0.00640606414526701, + 0.0042650094255805016, + 0.004381966311484575, + 0.007085473742336035, + 0.007526105269789696, + 0.00457413075491786, + 0.010671478696167469, + 0.006957166362553835, + 0.004679781850427389, + 0.0034101081546396017, + 0.0031395559199154377, + 0.005417338572442532, + 0.00469277473166585, + 0.0031707901507616043, + 0.004206293728202581, + 0.009077387861907482, + 0.005943664349615574, + 0.0031066827941685915, + 0.0034998625051230192, + 0.00463670352473855, + 0.005931760184466839, + 0.0030642636120319366, + 0.008092261850833893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2255568653345108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled medium shot captures a scene rendered in a lush, painterly style reminiscent of Studio Ghibli, focusing on an elderly, anthropomorphic white duck standing defiantly within a vast, ancient vault. Wearing an elaborate black silk top hat and a pair of small, gold-rimmed glasses perched on his orange beak, his expression is both resolute and fiercely protective. One of his wings is rigidly extended, forming a protective shield over his three furry yellow nephews, who huddle in fear beneath it, their wide, innocent eyes fixed on an unseen threat. The vault itself is a cavernous space constructed of massive, rough-hewn stone blocks. Raw light, emanating from the ethereal, luminous moss clinging to the walls and floor, casts a soft, magical glow in hues of emerald and sapphire. Piles of treasure—glittering gold coins, ancient leather-bound books, and mysterious, glowing artifacts—scatter the background, their details softened by a hand-painted aesthetic. The atmosphere was thick with tension as a huge, menacing shadow, indistinct in form, crept monstrously and slowly but relentlessly over the distant stone wall, its deliberate advance disturbing millions of motes of dust that had previously danced gently in the soft light and now causing them to swirl in agitated, chaotic patterns.", + "seed": 6828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_12/index_2138_seed_6828_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06828118152916432, + "vlm_logprob": -0.06828118152916432, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.06828118152916432, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008130665868520737 + ], + "opposite_logprob_tokens": [ + -4.8831305503845215 + ], + "logprob_reward": -0.008130665868520737, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008130665868520737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27095040678977966, + 0.08586727827787399, + 0.12870727479457855, + 0.09173066914081573, + 0.04591996222734451, + 0.08126013725996017, + 0.04501602053642273, + 0.13559848070144653, + 0.055403321981430054, + 0.05954640358686447 + ], + "credit_map": [ + 0.0496165007352829, + 0.025388294830918312, + 0.03487750515341759, + 0.02187826856970787, + 0.025308648124337196, + 0.059718403965234756, + 0.01890358328819275, + 0.03525920957326889, + 0.01029952708631754, + 0.012458388693630695, + 0.012836787849664688, + 0.010506890714168549, + 0.010402140207588673, + 0.010096529498696327, + 0.007991163991391659, + 0.011275848373770714, + 0.013561424799263477, + 0.014383248053491116, + 0.01285589300096035, + 0.01512062456458807, + 0.02259792760014534, + 0.015496501699090004, + 0.017488479614257812, + 0.01720316894352436, + 0.007684745825827122, + 0.008413841016590595, + 0.006586332805454731, + 0.009740573354065418, + 0.01611977629363537, + 0.012934979051351547, + 0.021610364317893982, + 0.00864005833864212, + 0.005821618251502514, + 0.004436382558196783, + 0.004034820944070816, + 0.005733753088861704, + 0.007749234326183796, + 0.006087701767683029, + 0.006685999687761068, + 0.005370448809117079, + 0.01159703079611063, + 0.010840578004717827, + 0.007803433109074831, + 0.01155822817236185, + 0.010643010027706623, + 0.009566254913806915, + 0.011971480213105679, + 0.007280129939317703, + 0.004987416788935661, + 0.003983682487159967, + 0.005271766800433397, + 0.007706694770604372, + 0.007589343935251236, + 0.005155070219188929, + 0.006283929105848074, + 0.00403811689466238, + 0.016831709071993828, + 0.014282438904047012, + 0.011961877346038818, + 0.016323212534189224, + 0.022692279890179634, + 0.017565645277500153, + 0.022345656529068947, + 0.013595668599009514, + 0.008445118553936481, + 0.011364508420228958, + 0.006670853588730097, + 0.007605779450386763, + 0.0053695873357355595, + 0.0063958740793168545, + 0.003958515357226133, + 0.0055930837988853455, + 0.008646456524729729, + 0.008652807213366032, + 0.006338183768093586, + 0.007397536188364029, + 0.005536059383302927, + 0.011682908982038498, + 0.004946914501488209, + 0.00634553749114275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.008130665868520737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810739889740944 + ], + "opposite_logprob_tokens": [ + -3.288107395172119 + ], + "logprob_reward": -0.03810739889740944, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810739889740944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12405847758054733, + 0.22563235461711884, + 0.09555426239967346, + 0.10984358936548233, + 0.04412379860877991, + 0.050953350961208344, + 0.0748671367764473, + 0.0606551468372345, + 0.19391988217830658, + 0.020392024889588356 + ], + "credit_map": [ + 0.020126821473240852, + 0.012544619850814342, + 0.016287636011838913, + 0.016339730471372604, + 0.012858706526458263, + 0.013978427276015282, + 0.017298277467489243, + 0.014624256640672684, + 0.026027247309684753, + 0.033983014523983, + 0.03151857480406761, + 0.034287337213754654, + 0.024329498410224915, + 0.02515443228185177, + 0.026759246364235878, + 0.02357299253344536, + 0.010956097394227982, + 0.011673199944198132, + 0.01040672417730093, + 0.011407601647078991, + 0.01782834716141224, + 0.011839578859508038, + 0.014599211513996124, + 0.006843498907983303, + 0.02044695056974888, + 0.028690896928310394, + 0.008387818932533264, + 0.010331395082175732, + 0.011034476570785046, + 0.015864264219999313, + 0.007906690239906311, + 0.007181107997894287, + 0.005583794321864843, + 0.007537422701716423, + 0.0034608221612870693, + 0.005453139077872038, + 0.007000025827437639, + 0.006097550969570875, + 0.005324259866029024, + 0.0036667832173407078, + 0.005514237098395824, + 0.005032292567193508, + 0.005174958147108555, + 0.007004037965089083, + 0.008756743744015694, + 0.005974843166768551, + 0.00856392364948988, + 0.004932315554469824, + 0.007886835373938084, + 0.005956686567515135, + 0.006142758764326572, + 0.011931312270462513, + 0.015273281373083591, + 0.011134910397231579, + 0.01145029067993164, + 0.005091062281280756, + 0.006514211185276508, + 0.006111078429967165, + 0.004786374978721142, + 0.007017194759100676, + 0.010125834494829178, + 0.007142802234739065, + 0.010141946375370026, + 0.0088157057762146, + 0.026412421837449074, + 0.03752772510051727, + 0.02246803790330887, + 0.028522346168756485, + 0.015501036308705807, + 0.021866325289011, + 0.015416448935866356, + 0.026205526664853096, + 0.0027246878016740084, + 0.004499687347561121, + 0.0022762224543839693, + 0.002530896570533514, + 0.0017821865621954203, + 0.002736822236329317, + 0.0019442347111180425, + 0.0018972853431478143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.03810739889740944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774561040103436 + ], + "opposite_logprob_tokens": [ + -4.634774684906006 + ], + "logprob_reward": -0.009774561040103436, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774561040103436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17052914202213287, + 0.07859481871128082, + 0.10706634074449539, + 0.10720834881067276, + 0.12338525056838989, + 0.07943298667669296, + 0.06385969370603561, + 0.0587601400911808, + 0.13242469727993011, + 0.07873858511447906 + ], + "credit_map": [ + 0.0274137482047081, + 0.01770763099193573, + 0.020690619945526123, + 0.01678631454706192, + 0.02703460492193699, + 0.02828150987625122, + 0.017908353358507156, + 0.014706368558108807, + 0.012776223011314869, + 0.008672280237078667, + 0.013052928261458874, + 0.007910351268947124, + 0.008647012524306774, + 0.009842762723565102, + 0.007994541898369789, + 0.009698718786239624, + 0.02007468417286873, + 0.01397701445966959, + 0.016749953851103783, + 0.0104092201218009, + 0.0074054221622645855, + 0.014691008254885674, + 0.010689502581954002, + 0.013069543056190014, + 0.016788417473435402, + 0.019624320790171623, + 0.01278542447835207, + 0.01053111907094717, + 0.00988368783146143, + 0.015454885549843311, + 0.012469515204429626, + 0.009670983068645, + 0.024580376222729683, + 0.03091614693403244, + 0.010850655846297741, + 0.009723830968141556, + 0.011452156119048595, + 0.018477872014045715, + 0.008844076655805111, + 0.008540141396224499, + 0.009856794960796833, + 0.014126414433121681, + 0.010840554721653461, + 0.00959846656769514, + 0.010753531940281391, + 0.010489045642316341, + 0.0071669272147119045, + 0.006601260043680668, + 0.006941254250705242, + 0.008257635869085789, + 0.007890382781624794, + 0.009191921912133694, + 0.009318863973021507, + 0.008187140338122845, + 0.007607039529830217, + 0.006465461105108261, + 0.008304965682327747, + 0.007571256719529629, + 0.004669214133173227, + 0.007573285605758429, + 0.008586217649281025, + 0.007888454012572765, + 0.008333130739629269, + 0.005833622068166733, + 0.014638291671872139, + 0.023376960307359695, + 0.016378480941057205, + 0.02176136150956154, + 0.01100952085107565, + 0.014512313529849052, + 0.010890113189816475, + 0.019857650622725487, + 0.012142403982579708, + 0.016481516882777214, + 0.008178102783858776, + 0.009796814061701298, + 0.007945834659039974, + 0.011616649106144905, + 0.006514601409435272, + 0.006062663625925779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.009774561040103436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093601368367672 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.004093601368367672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093601368367672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0997941717505455, + 0.0675235241651535, + 0.09967432916164398, + 0.15184076130390167, + 0.03990257903933525, + 0.0666586384177208, + 0.054188236594200134, + 0.03472442924976349, + 0.21913711726665497, + 0.16655626893043518 + ], + "credit_map": [ + 0.012816726230084896, + 0.011878935620188713, + 0.012194177135825157, + 0.010072767734527588, + 0.017864534631371498, + 0.01087117101997137, + 0.013488227501511574, + 0.010607616975903511, + 0.010968174785375595, + 0.008158189244568348, + 0.009493494406342506, + 0.007461077068001032, + 0.00952047761529684, + 0.007749693933874369, + 0.006924882531166077, + 0.00724753038957715, + 0.014828524552285671, + 0.010965712368488312, + 0.016077710315585136, + 0.012823752127587795, + 0.01055584754794836, + 0.013961936347186565, + 0.011887672357261181, + 0.008573166094720364, + 0.021345777437090874, + 0.044805582612752914, + 0.011611164547502995, + 0.00997976679354906, + 0.016631193459033966, + 0.022148847579956055, + 0.014777230098843575, + 0.010541180148720741, + 0.005588667467236519, + 0.006346878595650196, + 0.004024540539830923, + 0.005164330825209618, + 0.004936792887747288, + 0.005059472285211086, + 0.005318112671375275, + 0.0034637791104614735, + 0.013657226227223873, + 0.012655926868319511, + 0.007944075390696526, + 0.006592828314751387, + 0.007289845030754805, + 0.008250641636550426, + 0.0052628666162490845, + 0.005005223210901022, + 0.00658686738461256, + 0.008038253523409367, + 0.0072648716159164906, + 0.00798258651047945, + 0.0056227464228868484, + 0.0065389759838581085, + 0.006328623741865158, + 0.005825307685881853, + 0.005802817177027464, + 0.006570159923285246, + 0.0036096363328397274, + 0.004527626559138298, + 0.0027092958334833384, + 0.004234510939568281, + 0.002907948102802038, + 0.004362429957836866, + 0.024295903742313385, + 0.04231027513742447, + 0.02652398869395256, + 0.03417452797293663, + 0.018199199810624123, + 0.02405884489417076, + 0.017578646540641785, + 0.031995709985494614, + 0.025141717866063118, + 0.03804370388388634, + 0.02074972353875637, + 0.02712203562259674, + 0.010151033289730549, + 0.023075252771377563, + 0.008959544822573662, + 0.013313247822225094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.004093601368367672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812996804714203 + ], + "opposite_logprob_tokens": [ + -1.4062997102737427 + ], + "logprob_reward": -0.2812996804714203, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812996804714203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17757529020309448, + 0.17973937094211578, + 0.07385415583848953, + 0.10634413361549377, + 0.04359697550535202, + 0.059574734419584274, + 0.1015784963965416, + 0.0466967411339283, + 0.138569638133049, + 0.07247039675712585 + ], + "credit_map": [ + 0.028575850650668144, + 0.016356227919459343, + 0.013826199807226658, + 0.01829538121819496, + 0.028148019686341286, + 0.03923040255904198, + 0.016198350116610527, + 0.0169448833912611, + 0.02873656526207924, + 0.014615981839597225, + 0.029230069369077682, + 0.024474969133734703, + 0.014983788132667542, + 0.021238114684820175, + 0.020627709105610847, + 0.025832192972302437, + 0.011917823925614357, + 0.00767875649034977, + 0.00751536525785923, + 0.008238976821303368, + 0.012288045138120651, + 0.008999599143862724, + 0.008107848465442657, + 0.009107746183872223, + 0.016499090939760208, + 0.029533494263887405, + 0.008344507776200771, + 0.007246675435453653, + 0.01191532239317894, + 0.015309344045817852, + 0.010770196095108986, + 0.00672551617026329, + 0.00595724955201149, + 0.006812470033764839, + 0.004339336883276701, + 0.007731209974735975, + 0.00485243322327733, + 0.005679643712937832, + 0.004861601162701845, + 0.003363036084920168, + 0.011475504375994205, + 0.008728292770683765, + 0.006531055551022291, + 0.0065116859041154385, + 0.007152948062866926, + 0.007145938463509083, + 0.007591495290398598, + 0.004437822382897139, + 0.011932753957808018, + 0.0168536938726902, + 0.007805118337273598, + 0.011640961281955242, + 0.019110599532723427, + 0.015299811959266663, + 0.012149580754339695, + 0.006785995792597532, + 0.007486596703529358, + 0.009393233805894852, + 0.005639237351715565, + 0.0057125152088701725, + 0.004566017538309097, + 0.005746945738792419, + 0.0045004189014434814, + 0.0036517849657684565, + 0.01503795012831688, + 0.030857421457767487, + 0.016683129593729973, + 0.020403482019901276, + 0.010893445461988449, + 0.015149828977882862, + 0.010381985455751419, + 0.019162410870194435, + 0.012850034981966019, + 0.023022731766104698, + 0.009253572672605515, + 0.006627393886446953, + 0.004209398292005062, + 0.008550652302801609, + 0.0036420668475329876, + 0.0043145534582436085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.2812996804714203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled medium shot captures a scene rendered in a lush, painterly style reminiscent of Studio Ghibli, focusing on an elderly, anthropomorphic white duck standing defiantly within a vast, ancient vault. Wearing an elaborate black silk top hat and a pair of small, gold-rimmed glasses perched on his orange beak, his expression is both resolute and fiercely protective. One of his wings is rigidly extended, forming a protective shield over his three furry yellow nephews, who huddle in fear beneath it, their wide, innocent eyes fixed on an unseen threat. The vault itself is a cavernous space constructed of massive, rough-hewn stone blocks. Raw light, emanating from the ethereal, luminous moss clinging to the walls and floor, casts a soft, magical glow in hues of emerald and sapphire. Piles of treasure—glittering gold coins, ancient leather-bound books, and mysterious, glowing artifacts—scatter the background, their details softened by a hand-painted aesthetic. The atmosphere was thick with tension as a huge, menacing shadow, indistinct in form, crept monstrously and slowly but relentlessly over the distant stone wall, its deliberate advance disturbing millions of motes of dust that had previously danced gently in the soft light and now causing them to swirl in agitated, chaotic patterns.", + "seed": 6829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_12/index_2138_seed_6829_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04420219506137073, + "vlm_logprob": -0.04420219506137073, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.04420219506137073, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010282163508236408 + ], + "opposite_logprob_tokens": [ + -4.635282039642334 + ], + "logprob_reward": -0.010282163508236408, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010282163508236408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16616514325141907, + 0.08958424627780914, + 0.0579213872551918, + 0.04722874239087105, + 0.04959221929311752, + 0.06482259184122086, + 0.06400216370820999, + 0.09058871120214462, + 0.19963328540325165, + 0.170461505651474 + ], + "credit_map": [ + 0.01668412797152996, + 0.024057170376181602, + 0.014004048891365528, + 0.01201025489717722, + 0.03813948109745979, + 0.026753423735499382, + 0.019663384184241295, + 0.014853272587060928, + 0.015812892466783524, + 0.012571350671350956, + 0.007386316545307636, + 0.007592304609715939, + 0.016478294506669044, + 0.012702244333922863, + 0.009381784126162529, + 0.007659068331122398, + 0.009733065962791443, + 0.008659291081130505, + 0.0038611339405179024, + 0.004353750962764025, + 0.012265872210264206, + 0.007724076043814421, + 0.005549061577767134, + 0.005775140132755041, + 0.0036267233081161976, + 0.007508565206080675, + 0.0029235361143946648, + 0.004786960780620575, + 0.012944317422807217, + 0.004522554110735655, + 0.0065465327352285385, + 0.004369554575532675, + 0.007196506485342979, + 0.007743910886347294, + 0.0034715605434030294, + 0.00467326957732439, + 0.010534008964896202, + 0.005502269137650728, + 0.005892077926546335, + 0.004578622058033943, + 0.006605412811040878, + 0.008314560167491436, + 0.0105255963280797, + 0.007727175019681454, + 0.007712730672210455, + 0.007405435666441917, + 0.009610173292458057, + 0.006921517197042704, + 0.006541471462696791, + 0.007968258112668991, + 0.006920597981661558, + 0.009540464729070663, + 0.008354004472494125, + 0.009123953059315681, + 0.00915690790861845, + 0.006396513897925615, + 0.006641048472374678, + 0.009580976329743862, + 0.014074782840907574, + 0.01573256216943264, + 0.011526713147759438, + 0.008906003087759018, + 0.013941818848252296, + 0.0101848179474473, + 0.013091111555695534, + 0.013875548727810383, + 0.021802445873618126, + 0.023657483980059624, + 0.02911650575697422, + 0.012819001451134682, + 0.05574117600917816, + 0.029530027881264687, + 0.02019343338906765, + 0.021076899021863937, + 0.02173003740608692, + 0.014443306252360344, + 0.018409593030810356, + 0.014497875235974789, + 0.026774132624268532, + 0.033336251974105835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.010282163508236408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496445298194885 + ], + "opposite_logprob_tokens": [ + -2.929964542388916 + ], + "logprob_reward": -0.05496445298194885, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496445298194885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08231443166732788, + 0.20549306273460388, + 0.02064829133450985, + 0.024435831233859062, + 0.03021024912595749, + 0.11714349687099457, + 0.05158131197094917, + 0.09819234162569046, + 0.21620982885360718, + 0.1537710577249527 + ], + "credit_map": [ + 0.007356185931712389, + 0.012586222030222416, + 0.007596758659929037, + 0.005767423659563065, + 0.02043677307665348, + 0.010132700204849243, + 0.011486810632050037, + 0.006951566785573959, + 0.046022504568099976, + 0.02938719466328621, + 0.016149308532476425, + 0.016922706738114357, + 0.023882411420345306, + 0.034321658313274384, + 0.016243238002061844, + 0.022564051672816277, + 0.00257692183367908, + 0.002907648915424943, + 0.0016424678033217788, + 0.002181828487664461, + 0.004349505063146353, + 0.0019099899800494313, + 0.0029820462223142385, + 0.0020978841930627823, + 0.0031572417356073856, + 0.003528803354129195, + 0.0018331792671233416, + 0.002145314123481512, + 0.00578472763299942, + 0.002489445498213172, + 0.0033136983402073383, + 0.00218342337757349, + 0.005893011577427387, + 0.005209661088883877, + 0.0026363732758909464, + 0.0020908236037939787, + 0.003473894437775016, + 0.004628937225788832, + 0.002219399670138955, + 0.004058149177581072, + 0.016139809042215347, + 0.01527458056807518, + 0.022323889657855034, + 0.01542679127305746, + 0.011121285147964954, + 0.011718553490936756, + 0.012984872795641422, + 0.012153721414506435, + 0.006826842203736305, + 0.006739506497979164, + 0.0051723625510931015, + 0.006598097272217274, + 0.006048752926290035, + 0.006663245148956776, + 0.008255505934357643, + 0.005277003161609173, + 0.010154983028769493, + 0.009439067915081978, + 0.012713275849819183, + 0.013993525877594948, + 0.017687592655420303, + 0.009322323836386204, + 0.013018334284424782, + 0.01186324842274189, + 0.01625605672597885, + 0.02397477999329567, + 0.024122387170791626, + 0.029369991272687912, + 0.02960778959095478, + 0.011371356435120106, + 0.04556957632303238, + 0.03593791648745537, + 0.015905706211924553, + 0.02049119956791401, + 0.018009424209594727, + 0.019026456400752068, + 0.014732720330357552, + 0.019838249310851097, + 0.020837651565670967, + 0.024929670616984367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.05496445298194885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775268845260143 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775268845260143, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775268845260143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12321296334266663, + 0.17313194274902344, + 0.05171319097280502, + 0.027645377442240715, + 0.021005984395742416, + 0.026961784809827805, + 0.10397765785455704, + 0.19329100847244263, + 0.15302817523479462, + 0.12603183090686798 + ], + "credit_map": [ + 0.017028260976076126, + 0.01897810772061348, + 0.015840670093894005, + 0.010032753460109234, + 0.022385627031326294, + 0.0159295741468668, + 0.012723199091851711, + 0.010294773615896702, + 0.03457758203148842, + 0.023876477032899857, + 0.013843190856277943, + 0.013971898704767227, + 0.025700004771351814, + 0.025608012452721596, + 0.018339907750487328, + 0.017214873805642128, + 0.011063640005886555, + 0.007855714298784733, + 0.005110408179461956, + 0.00400983402505517, + 0.006539347115904093, + 0.007481443230062723, + 0.00445112818852067, + 0.005201681517064571, + 0.0041518500074744225, + 0.004344230517745018, + 0.0026385467499494553, + 0.0024943386670202017, + 0.005448715295642614, + 0.0026676582638174295, + 0.003486912464722991, + 0.0024131275713443756, + 0.0030327122658491135, + 0.0030986096244305372, + 0.001994494814425707, + 0.0018300641095265746, + 0.003729147370904684, + 0.0025945152156054974, + 0.0021322877146303654, + 0.002594157587736845, + 0.002196711953729391, + 0.0031773410737514496, + 0.002837816718965769, + 0.0038359183818101883, + 0.004684519488364458, + 0.002952027367427945, + 0.005003672093153, + 0.0022737798281013966, + 0.011231755837798119, + 0.014018497429788113, + 0.01277597900480032, + 0.010786544531583786, + 0.016455085948109627, + 0.01760438084602356, + 0.011953567154705524, + 0.009151855483651161, + 0.015015418641269207, + 0.018214315176010132, + 0.036546219140291214, + 0.020639056339859962, + 0.030829841271042824, + 0.02107418328523636, + 0.02464689128100872, + 0.026325104758143425, + 0.011185831390321255, + 0.015277326107025146, + 0.016845906153321266, + 0.015258570201694965, + 0.021808648481965065, + 0.011765476316213608, + 0.03951653465628624, + 0.021369893103837967, + 0.01971529982984066, + 0.01759815402328968, + 0.01269945502281189, + 0.012541216798126698, + 0.011931388638913631, + 0.011849171482026577, + 0.01547172013670206, + 0.02422543615102768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.009775268845260143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031909046228975058 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031909046228975058, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031909046228975058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1327548325061798, + 0.2283582240343094, + 0.02306871861219406, + 0.017807811498641968, + 0.012031174264848232, + 0.03660663217306137, + 0.05041626840829849, + 0.05441751703619957, + 0.21390962600708008, + 0.23062922060489655 + ], + "credit_map": [ + 0.029056496918201447, + 0.019274629652500153, + 0.012881421484053135, + 0.010169175453484058, + 0.015048602595925331, + 0.02056831680238247, + 0.013072892092168331, + 0.012683290056884289, + 0.04400709643959999, + 0.03080126829445362, + 0.018288997933268547, + 0.02001967839896679, + 0.03606231138110161, + 0.031141556799411774, + 0.02745825983583927, + 0.020579015836119652, + 0.0033549449872225523, + 0.0036200289614498615, + 0.002185138640925288, + 0.0022230336908251047, + 0.0028496342711150646, + 0.0029954377096146345, + 0.00328762992285192, + 0.0025528669357299805, + 0.0032892641611397266, + 0.0023673742543905973, + 0.0015383482677862048, + 0.0014854874461889267, + 0.0026578232645988464, + 0.003008834086358547, + 0.0016926294192671776, + 0.0017680485034361482, + 0.0023417221382260323, + 0.0017473890911787748, + 0.0011936075752601027, + 0.0011646857019513845, + 0.0013636834919452667, + 0.0019251176854595542, + 0.001019344199448824, + 0.0012756233336403966, + 0.0034359158016741276, + 0.004804783500730991, + 0.004205462988466024, + 0.003953548613935709, + 0.007223283406347036, + 0.003888482926413417, + 0.004387696739286184, + 0.0047074537724256516, + 0.006292755715548992, + 0.008274327032268047, + 0.006267671473324299, + 0.006851304788142443, + 0.005042970180511475, + 0.008068704046308994, + 0.0045257569290697575, + 0.00509277218952775, + 0.008134381845593452, + 0.007429962977766991, + 0.008351892232894897, + 0.006930486299097538, + 0.004396057687699795, + 0.004531317390501499, + 0.00690364558249712, + 0.007739766035228968, + 0.01609322428703308, + 0.018209662288427353, + 0.022227471694350243, + 0.021699951961636543, + 0.030643627047538757, + 0.015412796288728714, + 0.05974818021059036, + 0.029874693602323532, + 0.024758068844676018, + 0.030436517670750618, + 0.02455735020339489, + 0.023414313793182373, + 0.022307178005576134, + 0.027028758078813553, + 0.0325298048555851, + 0.04559721425175667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.0031909046228975058 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly, anthropomorphic white duck wearing a black silk top hat and gold-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are three furry yellow nephews visible huddled beneath the duck's extended wing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, ancient vault constructed of massive, rough-hewn stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by ethereal, luminous moss on the walls and floor casting emerald and sapphire hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a huge, menacing shadow creeping over a distant stone wall, causing dust motes to swirl chaotically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279818534851074 + ], + "opposite_logprob_tokens": [ + -2.0177981853485107 + ], + "logprob_reward": -0.14279818534851074, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279818534851074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06708082556724548, + 0.30479666590690613, + 0.03869606927037239, + 0.014917274937033653, + 0.012107876129448414, + 0.024790603667497635, + 0.027579130604863167, + 0.08241542428731918, + 0.25478893518447876, + 0.17282713949680328 + ], + "credit_map": [ + 0.010397782549262047, + 0.00952513050287962, + 0.006867877673357725, + 0.005647051148116589, + 0.010910219512879848, + 0.00929108913987875, + 0.008053054101765156, + 0.0063886190764606, + 0.05963226407766342, + 0.04124800115823746, + 0.02343938499689102, + 0.02662801556289196, + 0.04704321175813675, + 0.04224998131394386, + 0.03642667084932327, + 0.02812914177775383, + 0.006826070602983236, + 0.006116282194852829, + 0.0038916710764169693, + 0.003282790305092931, + 0.004062937572598457, + 0.005557429976761341, + 0.004662275779992342, + 0.004296612460166216, + 0.0027490619104355574, + 0.002168152714148164, + 0.0012663766974583268, + 0.0013252292992547154, + 0.0023224600590765476, + 0.002106033032760024, + 0.0017604598542675376, + 0.0012195025337859988, + 0.0017258110456168652, + 0.0018256817711517215, + 0.0011822010856121778, + 0.0010377081343904138, + 0.0023618703708052635, + 0.0013730149948969483, + 0.001560914097353816, + 0.0010406740475445986, + 0.0026467652060091496, + 0.003281728830188513, + 0.0023773577995598316, + 0.0028814254328608513, + 0.004884039983153343, + 0.002660743659362197, + 0.0030064634047448635, + 0.003052080050110817, + 0.0028092090506106615, + 0.0033628486562520266, + 0.002687599742785096, + 0.0049493443220853806, + 0.00306873070076108, + 0.0035016390029340982, + 0.004766707308590412, + 0.0024330527521669865, + 0.006514082197099924, + 0.007849055342376232, + 0.01610385626554489, + 0.009741800837218761, + 0.011091639287769794, + 0.0076153031550347805, + 0.01366274431347847, + 0.009836938232183456, + 0.018406741321086884, + 0.01983129419386387, + 0.034676797688007355, + 0.0246559027582407, + 0.03619469702243805, + 0.014560948126018047, + 0.06684781610965729, + 0.039614733308553696, + 0.0200112983584404, + 0.02108454704284668, + 0.016469713300466537, + 0.01935192570090294, + 0.014532527886331081, + 0.021029995754361153, + 0.02619094029068947, + 0.0341561958193779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.14279818534851074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_tensor([5232, 5232])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_tensor([5232, 5232])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0edeeea82f31973d8533d4c8e890b71da96253cd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_tensor([5232, 5232])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a slightly upward angle and handheld, imparts a subtle yet authentic tremor to the image, as if the photographer is holding his breath, overwhelmed by the overwhelming presence of the figure before him. At the center of the frame stands a white-haired, muscular martial arts master, standing like a rock in the center of an ancient Japanese dojo steeped in history. His graying hair is meticulously tied back in a bun, his weathered face etched with deep wrinkles, yet his eyes, sharp as a hawk, seem to pierce the surrounding gloom. He wears a traditional cotton kikofu (deep indigo), slightly faded from years of wear and washing, its edges showing signs of wear and tear, further highlighting his profound experience. His bare feet, firmly planted on the dark wooden floor, polished to a gleaming shine by time, his arms, veins visible, hanging naturally at his sides, his entire posture exuding an unwavering determination and composure. The dojo's backdrop is a hazy shoji paper door and a wall adorned with calligraphy depicting the character \"武.\" Wisps of dust drift slowly in the tangible, Tyndall-like beams of light slanting in from the high windows. The lighting throughout the scene is highly dramatic, with much of the space shrouded in deep shadow. Only the light precisely outlines the grandmaster's silver hair, the muscular lines of his shoulders, and the resolute expression on his face, creating an atmosphere of solemnity, power, and tranquility.", + "seed": 17894, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_5232_seed_17894_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024746794695965947, + "vlm_logprob": -0.024746794695965947, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.024746794695965947, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008151713758707047 + ], + "opposite_logprob_tokens": [ + -5.133151531219482 + ], + "logprob_reward": -0.008151713758707047, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008151713758707047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.335528701543808, + 0.04539908841252327, + 0.051553092896938324, + 0.03841523081064224, + 0.03983418643474579, + 0.11179967224597931, + 0.13579794764518738, + 0.06264587491750717, + 0.11732078343629837, + 0.061705440282821655 + ], + "credit_map": [ + 0.03201412782073021, + 0.04447565972805023, + 0.04153972491621971, + 0.04763535037636757, + 0.03841610625386238, + 0.05713547766208649, + 0.03181060776114464, + 0.04250166192650795, + 0.003334096632897854, + 0.006986795458942652, + 0.004342914558947086, + 0.005654470529407263, + 0.005384253337979317, + 0.005828696768730879, + 0.0071821054443717, + 0.0066857547499239445, + 0.0043007927015423775, + 0.00590873509645462, + 0.005501128267496824, + 0.006313310004770756, + 0.005596133880317211, + 0.009355162270367146, + 0.005366108380258083, + 0.009211724624037743, + 0.0037468676455318928, + 0.0035614462103694677, + 0.004104326479136944, + 0.0051767826080322266, + 0.005342694465070963, + 0.006217700429260731, + 0.0053182863630354404, + 0.0049471245147287846, + 0.0020046827849000692, + 0.0018254652386531234, + 0.0030083029996603727, + 0.00217928783968091, + 0.00350744160823524, + 0.012811252847313881, + 0.0033529975917190313, + 0.011144755408167839, + 0.004936043173074722, + 0.004073816351592541, + 0.006329833064228296, + 0.004873807542026043, + 0.005427369847893715, + 0.040889889001846313, + 0.007048084866255522, + 0.03822081908583641, + 0.0043887849897146225, + 0.0031398688443005085, + 0.006426326930522919, + 0.005008133593946695, + 0.006646808702498674, + 0.05283721163868904, + 0.0059302812442183495, + 0.05142053961753845, + 0.005562006961554289, + 0.007849406450986862, + 0.008046158589422703, + 0.007176111917942762, + 0.008724477142095566, + 0.00791817344725132, + 0.009019284509122372, + 0.008350254967808723, + 0.0048580290749669075, + 0.009075822308659554, + 0.015329534187912941, + 0.01768081821501255, + 0.011068791151046753, + 0.027920955792069435, + 0.010537640191614628, + 0.02084919437766075, + 0.0044433120638132095, + 0.005157762207090855, + 0.006426842883229256, + 0.0071981907822191715, + 0.007836205884814262, + 0.009275498799979687, + 0.010171495378017426, + 0.011196132749319077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.008151713758707047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004182040691376 + ], + "opposite_logprob_tokens": [ + -2.6950418949127197 + ], + "logprob_reward": -0.07004182040691376, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004182040691376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2803102135658264, + 0.07084649056196213, + 0.06561382859945297, + 0.029615893959999084, + 0.037972670048475266, + 0.07969249039888382, + 0.248029425740242, + 0.0718546062707901, + 0.057517025619745255, + 0.05854739993810654 + ], + "credit_map": [ + 0.025146296247839928, + 0.04379231482744217, + 0.029907802119851112, + 0.035800736397504807, + 0.027821410447359085, + 0.03681008145213127, + 0.045068733394145966, + 0.035962849855422974, + 0.007474208250641823, + 0.00794399157166481, + 0.008142507635056973, + 0.011240633204579353, + 0.007590101566165686, + 0.009682182222604752, + 0.007897898554801941, + 0.010874966159462929, + 0.005771254654973745, + 0.010653543286025524, + 0.007146905642002821, + 0.006052064243704081, + 0.009886442683637142, + 0.009006801061332226, + 0.010088750161230564, + 0.007008064072579145, + 0.002015823731198907, + 0.002065651584416628, + 0.003021092852577567, + 0.002568908268585801, + 0.0029924255795776844, + 0.00717660877853632, + 0.004207679536193609, + 0.0055677033960819244, + 0.002353298244997859, + 0.003045870689675212, + 0.0028690334875136614, + 0.002200536895543337, + 0.0037410585209727287, + 0.011472673155367374, + 0.003013118403032422, + 0.009277081117033958, + 0.0038435168098658323, + 0.004494885914027691, + 0.006013397127389908, + 0.004846010357141495, + 0.00517653115093708, + 0.02427562139928341, + 0.008814558386802673, + 0.022227970883250237, + 0.0071844682097435, + 0.0039755660109221935, + 0.008809304796159267, + 0.0043088700622320175, + 0.006065956316888332, + 0.10816174000501633, + 0.00805723201483488, + 0.10146628320217133, + 0.007366199046373367, + 0.005239950027316809, + 0.010171836242079735, + 0.008464161306619644, + 0.009152722544968128, + 0.008290499448776245, + 0.01184722688049078, + 0.011322005651891232, + 0.003330785082653165, + 0.005586260464042425, + 0.007215783931314945, + 0.005385816562920809, + 0.005947291385382414, + 0.01397726684808731, + 0.005083915777504444, + 0.010989903472363949, + 0.0038070192094892263, + 0.006092130672186613, + 0.009396549314260483, + 0.007732628844678402, + 0.008947120048105717, + 0.008202449418604374, + 0.009118620306253433, + 0.005250883288681507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.07004182040691376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946681834757328 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946681834757328, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946681834757328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.257036417722702, + 0.08979768306016922, + 0.061710063368082047, + 0.03199503943324089, + 0.03200118988752365, + 0.09743176400661469, + 0.11854492127895355, + 0.09546766430139542, + 0.05386188626289368, + 0.16215331852436066 + ], + "credit_map": [ + 0.021782945841550827, + 0.055711448192596436, + 0.03147174417972565, + 0.034745506942272186, + 0.030280157923698425, + 0.02980504184961319, + 0.03083464317023754, + 0.02240494079887867, + 0.010676431469619274, + 0.009134967811405659, + 0.01214561890810728, + 0.0141533138230443, + 0.006042689550668001, + 0.013056253083050251, + 0.008053651079535484, + 0.016534751281142235, + 0.005251310765743256, + 0.007426051422953606, + 0.006550303194671869, + 0.008433926850557327, + 0.007521777879446745, + 0.009916537441313267, + 0.0082039600238204, + 0.008406194858253002, + 0.002737682079896331, + 0.003643078962340951, + 0.003289253683760762, + 0.003973264712840319, + 0.004170505329966545, + 0.00547763891518116, + 0.004821465816348791, + 0.0038821531925350428, + 0.002515186322852969, + 0.0030353444162756205, + 0.0023284645285457373, + 0.0021161928307265043, + 0.002486829413101077, + 0.009308087639510632, + 0.0022622349206358194, + 0.007948849350214005, + 0.007569980341941118, + 0.006862272508442402, + 0.0074592349119484425, + 0.005375801119953394, + 0.004575988743454218, + 0.029881779104471207, + 0.007516640704125166, + 0.028190072625875473, + 0.004011484328657389, + 0.005257941782474518, + 0.006434111390262842, + 0.005375096574425697, + 0.0038104411214590073, + 0.046093184500932693, + 0.004137896467000246, + 0.04342476278543472, + 0.00955603551119566, + 0.01349158026278019, + 0.013398980721831322, + 0.012368796393275261, + 0.008326765149831772, + 0.012071236036717892, + 0.009851221926510334, + 0.016403043642640114, + 0.0039171637035906315, + 0.00483864126726985, + 0.009904757142066956, + 0.010247355327010155, + 0.004590560216456652, + 0.009120997972786427, + 0.005362695548683405, + 0.005879716947674751, + 0.01468636840581894, + 0.027270156890153885, + 0.028678864240646362, + 0.029404789209365845, + 0.013409841805696487, + 0.025333072990179062, + 0.012559216469526291, + 0.010811003856360912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.005946681834757328 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015111466636881232 + ], + "opposite_logprob_tokens": [ + -6.501511096954346 + ], + "logprob_reward": -0.0015111466636881232, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015111466636881232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2494194358587265, + 0.1163272112607956, + 0.08368169516324997, + 0.041722238063812256, + 0.045074909925460815, + 0.0406293086707592, + 0.3331141471862793, + 0.03306983411312103, + 0.02962885983288288, + 0.027332322672009468 + ], + "credit_map": [ + 0.03014496900141239, + 0.042828332632780075, + 0.027155248448252678, + 0.03958025202155113, + 0.021667461842298508, + 0.0331137552857399, + 0.022286783903837204, + 0.032642632722854614, + 0.020338404923677444, + 0.011856885626912117, + 0.018549565225839615, + 0.012244731187820435, + 0.007121784146875143, + 0.016829904168844223, + 0.008311430923640728, + 0.02107449434697628, + 0.00852818600833416, + 0.010372519493103027, + 0.009565703570842743, + 0.008781475014984608, + 0.011633032001554966, + 0.01120187621563673, + 0.011756784282624722, + 0.01184211764484644, + 0.00383747392334044, + 0.0037485852371901274, + 0.0036680654156953096, + 0.004220064263790846, + 0.005190540105104446, + 0.00908586010336876, + 0.004696402698755264, + 0.007275245618075132, + 0.004703950602561235, + 0.006405696272850037, + 0.006207641679793596, + 0.004752750974148512, + 0.0054733892902731895, + 0.006665255408734083, + 0.005177957005798817, + 0.005688265897333622, + 0.002826048992574215, + 0.006516473833471537, + 0.004113346338272095, + 0.004386409651488066, + 0.005924363620579243, + 0.005829215981066227, + 0.006714392453432083, + 0.004319055937230587, + 0.012923875823616982, + 0.012118560262024403, + 0.014467473141849041, + 0.012384670786559582, + 0.008836047723889351, + 0.13498681783676147, + 0.011788963340222836, + 0.12560774385929108, + 0.0024217518512159586, + 0.003948650322854519, + 0.005239550489932299, + 0.003917188849300146, + 0.003973762970417738, + 0.004620298743247986, + 0.004559567663818598, + 0.004389064386487007, + 0.0022303611040115356, + 0.003923844080418348, + 0.003886923659592867, + 0.003465871326625347, + 0.003258563345298171, + 0.004693347495049238, + 0.0035081401001662016, + 0.004661808721721172, + 0.0021733257453888655, + 0.0028831413947045803, + 0.003983070142567158, + 0.0044150883331894875, + 0.003571294015273452, + 0.0036515011452138424, + 0.003979971166700125, + 0.0026749286334961653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0015111466636881232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038082610815763474 + ], + "opposite_logprob_tokens": [ + -3.2880825996398926 + ], + "logprob_reward": -0.038082610815763474, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038082610815763474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3136081397533417, + 0.05260134115815163, + 0.06379452347755432, + 0.024185169488191605, + 0.05415649339556694, + 0.10778552293777466, + 0.12011350691318512, + 0.07419237494468689, + 0.09909731894731522, + 0.09046559780836105 + ], + "credit_map": [ + 0.03743620961904526, + 0.06894752383232117, + 0.03320404514670372, + 0.030116895213723183, + 0.03291667625308037, + 0.03701666742563248, + 0.035759229212999344, + 0.0382109060883522, + 0.004575361497700214, + 0.008053608238697052, + 0.005170729011297226, + 0.006818481720983982, + 0.006691979244351387, + 0.00695843854919076, + 0.00708822812885046, + 0.00724452082067728, + 0.005791432224214077, + 0.007877864874899387, + 0.006972801871597767, + 0.006258214823901653, + 0.009992290288209915, + 0.008067095652222633, + 0.011027669534087181, + 0.007807160262018442, + 0.002799025271087885, + 0.0033290633000433445, + 0.0026548600289970636, + 0.0021826045121997595, + 0.0028256988152861595, + 0.004056146834045649, + 0.0032497402280569077, + 0.0030880304984748363, + 0.0026692203246057034, + 0.0032792186830192804, + 0.004506407305598259, + 0.0028228312730789185, + 0.002889785682782531, + 0.018459729850292206, + 0.004150161985307932, + 0.015379142016172409, + 0.006705694831907749, + 0.013243207708001137, + 0.008982104249298573, + 0.008354812860488892, + 0.008428552187979221, + 0.027899309992790222, + 0.009817104786634445, + 0.024354752153158188, + 0.006363498046994209, + 0.0062743364833295345, + 0.008434038609266281, + 0.00691180769354105, + 0.00957557000219822, + 0.03881247341632843, + 0.010276301763951778, + 0.033465489745140076, + 0.006033497862517834, + 0.005263747647404671, + 0.007877799682319164, + 0.009500229731202126, + 0.007337045855820179, + 0.012261267751455307, + 0.011214489117264748, + 0.014704300090670586, + 0.0035708800423890352, + 0.004918307065963745, + 0.008774355985224247, + 0.010799702256917953, + 0.006342034786939621, + 0.032160814851522446, + 0.006730557419359684, + 0.025800667703151703, + 0.008031198754906654, + 0.010207009501755238, + 0.016197960823774338, + 0.017536558210849762, + 0.007651555351912975, + 0.014838749542832375, + 0.008056528866291046, + 0.007946045137941837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.038082610815763474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle and handheld, imparts a subtle yet authentic tremor to the image, as if the photographer is holding his breath, overwhelmed by the overwhelming presence of the figure before him. At the center of the frame stands a white-haired, muscular martial arts master, standing like a rock in the center of an ancient Japanese dojo steeped in history. His graying hair is meticulously tied back in a bun, his weathered face etched with deep wrinkles, yet his eyes, sharp as a hawk, seem to pierce the surrounding gloom. He wears a traditional cotton kikofu (deep indigo), slightly faded from years of wear and washing, its edges showing signs of wear and tear, further highlighting his profound experience. His bare feet, firmly planted on the dark wooden floor, polished to a gleaming shine by time, his arms, veins visible, hanging naturally at his sides, his entire posture exuding an unwavering determination and composure. The dojo's backdrop is a hazy shoji paper door and a wall adorned with calligraphy depicting the character \"武.\" Wisps of dust drift slowly in the tangible, Tyndall-like beams of light slanting in from the high windows. The lighting throughout the scene is highly dramatic, with much of the space shrouded in deep shadow. Only the light precisely outlines the grandmaster's silver hair, the muscular lines of his shoulders, and the resolute expression on his face, creating an atmosphere of solemnity, power, and tranquility.", + "seed": 17895, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_5232_seed_17895_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019460530113428832, + "vlm_logprob": -0.019460530113428832, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.019460530113428832, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011477983556687832 + ], + "opposite_logprob_tokens": [ + -4.761477947235107 + ], + "logprob_reward": -0.011477983556687832, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011477983556687832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43113845586776733, + 0.06188808009028435, + 0.05486300215125084, + 0.06498181074857712, + 0.06224456802010536, + 0.0688452497124672, + 0.08305846154689789, + 0.04759836196899414, + 0.05141913518309593, + 0.07396281510591507 + ], + "credit_map": [ + 0.04249842092394829, + 0.05153856426477432, + 0.10543418675661087, + 0.06375506520271301, + 0.03052869625389576, + 0.0508962906897068, + 0.025073900818824768, + 0.06141337752342224, + 0.0063013071194291115, + 0.012759287841618061, + 0.008410818874835968, + 0.009007194079458714, + 0.007638422306627035, + 0.0062505449168384075, + 0.00526939844712615, + 0.00625110836699605, + 0.00312553602270782, + 0.00303017464466393, + 0.009016147814691067, + 0.012464257888495922, + 0.00411007646471262, + 0.004331210162490606, + 0.01275478396564722, + 0.0060308147221803665, + 0.01184167806059122, + 0.008227262645959854, + 0.009381994605064392, + 0.008838017471134663, + 0.0062157888896763325, + 0.005439456552267075, + 0.007817257195711136, + 0.007220362313091755, + 0.009342968463897705, + 0.00873792078346014, + 0.009159090928733349, + 0.009402702562510967, + 0.008925741538405418, + 0.006109775975346565, + 0.006394081749022007, + 0.004172287881374359, + 0.01224726252257824, + 0.006395082920789719, + 0.010470704175531864, + 0.011342580430209637, + 0.005674436688423157, + 0.006902005989104509, + 0.00955692958086729, + 0.006256250198930502, + 0.015889625996351242, + 0.0080686891451478, + 0.013039419427514076, + 0.009210885502398014, + 0.006298036314547062, + 0.01554397027939558, + 0.008129382506012917, + 0.006878453306853771, + 0.010817544534802437, + 0.0038032864686101675, + 0.007230166811496019, + 0.007999943569302559, + 0.0035927798599004745, + 0.006288452073931694, + 0.0043564848601818085, + 0.0035097042564302683, + 0.006190395914018154, + 0.007296910975128412, + 0.007783825509250164, + 0.007684517186135054, + 0.006164552178233862, + 0.006756749004125595, + 0.004050880670547485, + 0.005491309799253941, + 0.008382104337215424, + 0.007589057553559542, + 0.008532906882464886, + 0.013612178154289722, + 0.010868688113987446, + 0.011689592152833939, + 0.007304070983082056, + 0.005984222516417503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.011477983556687832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309182986617088 + ], + "opposite_logprob_tokens": [ + -3.1680917739868164 + ], + "logprob_reward": -0.04309182986617088, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309182986617088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20802626013755798, + 0.040114350616931915, + 0.04832569137215614, + 0.04587792977690697, + 0.036549314856529236, + 0.06637286394834518, + 0.054395515471696854, + 0.060775596648454666, + 0.09998275339603424, + 0.3395797312259674 + ], + "credit_map": [ + 0.023335687816143036, + 0.03239273279905319, + 0.05019247904419899, + 0.031369976699352264, + 0.016236165538430214, + 0.023586580529808998, + 0.011938950978219509, + 0.018973680213093758, + 0.005896872375160456, + 0.00555045809596777, + 0.005366850178688765, + 0.005497169215232134, + 0.0050579579547047615, + 0.004792854189872742, + 0.003738964907824993, + 0.00421322463080287, + 0.005604861304163933, + 0.005221810191869736, + 0.008787604980170727, + 0.009505132213234901, + 0.003177208825945854, + 0.005843526218086481, + 0.006099465768784285, + 0.004086084198206663, + 0.005132489372044802, + 0.008988561108708382, + 0.006611373275518417, + 0.006593789905309677, + 0.003427838906645775, + 0.0064440080896019936, + 0.004182181786745787, + 0.00449768640100956, + 0.008744604885578156, + 0.005737491883337498, + 0.004463565535843372, + 0.004481187090277672, + 0.0036331370938569307, + 0.003638559952378273, + 0.0029503991827368736, + 0.0029003683011978865, + 0.008974245749413967, + 0.008636618964374065, + 0.008450744673609734, + 0.009812952019274235, + 0.00862265843898058, + 0.008285005576908588, + 0.00799510721117258, + 0.005595530848950148, + 0.012610457837581635, + 0.0047254497185349464, + 0.007840506732463837, + 0.008436108008027077, + 0.005928082391619682, + 0.006603367161005735, + 0.004253088030964136, + 0.003998455125838518, + 0.008831707760691643, + 0.003658188506960869, + 0.008627363480627537, + 0.011359471827745438, + 0.0049074268899858, + 0.005451532546430826, + 0.011814306490123272, + 0.006125595420598984, + 0.009213389828801155, + 0.007632885593920946, + 0.014537275768816471, + 0.017949391156435013, + 0.015616308897733688, + 0.010248856619000435, + 0.012124654836952686, + 0.012659991160035133, + 0.03341502323746681, + 0.02719363197684288, + 0.04142366349697113, + 0.04763013869524002, + 0.04898925498127937, + 0.053519777953624725, + 0.043468303978443146, + 0.04393995553255081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.04309182986617088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198003336787224 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014198003336787224, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198003336787224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16636332869529724, + 0.1061098650097847, + 0.037647794932127, + 0.2121504545211792, + 0.18975605070590973, + 0.07945344597101212, + 0.028771379962563515, + 0.030069055035710335, + 0.06973683834075928, + 0.0799417570233345 + ], + "credit_map": [ + 0.021253598853945732, + 0.0241555068641901, + 0.03936507925391197, + 0.023650649935007095, + 0.011953229084610939, + 0.014973322860896587, + 0.008374463766813278, + 0.022637486457824707, + 0.01928790844976902, + 0.018550053238868713, + 0.01782323606312275, + 0.018538763746619225, + 0.0070519959554076195, + 0.009784026071429253, + 0.005896169692277908, + 0.009177721105515957, + 0.0038722550962120295, + 0.0035658779088407755, + 0.007010484114289284, + 0.008549987338483334, + 0.0026697868015617132, + 0.004160918295383453, + 0.004103619605302811, + 0.0037148662377148867, + 0.02628246322274208, + 0.04075685888528824, + 0.03148597478866577, + 0.030908280983567238, + 0.014325310476124287, + 0.029666386544704437, + 0.017782945185899734, + 0.020942235365509987, + 0.03282172605395317, + 0.03667699173092842, + 0.028351467102766037, + 0.022534172981977463, + 0.016390696167945862, + 0.02689979411661625, + 0.009008456021547318, + 0.017072753980755806, + 0.0235019251704216, + 0.010649128817021847, + 0.011229333467781544, + 0.008316672407090664, + 0.005847386550158262, + 0.009656456299126148, + 0.004203006625175476, + 0.0060495417565107346, + 0.00593524007126689, + 0.002252966398373246, + 0.00392252579331398, + 0.003842257661744952, + 0.003489586990326643, + 0.004298113752156496, + 0.002890370786190033, + 0.002140318276360631, + 0.003675278276205063, + 0.0022644102573394775, + 0.003834432689473033, + 0.0056423828937113285, + 0.0030204348731786013, + 0.004122361075133085, + 0.004216846078634262, + 0.0032929100561887026, + 0.012795615009963512, + 0.013979601673781872, + 0.01179384347051382, + 0.007710771635174751, + 0.004507372155785561, + 0.01054959837347269, + 0.0039458260871469975, + 0.004454206209629774, + 0.01125457976013422, + 0.009107550606131554, + 0.018096525222063065, + 0.009840193204581738, + 0.007305747363716364, + 0.014781217090785503, + 0.005272479262202978, + 0.004283461719751358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.014198003336787224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021971152164041996 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021971152164041996, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021971152164041996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17229537665843964, + 0.14949721097946167, + 0.32098209857940674, + 0.05435415729880333, + 0.030519098043441772, + 0.1454267054796219, + 0.023113535717129707, + 0.03654294088482857, + 0.030763784423470497, + 0.036505091935396194 + ], + "credit_map": [ + 0.017244797199964523, + 0.029859719797968864, + 0.029455453157424927, + 0.029091402888298035, + 0.02167828194797039, + 0.020916497334837914, + 0.01173136942088604, + 0.012317853048443794, + 0.0155708659440279, + 0.027791794389486313, + 0.023623153567314148, + 0.019476447254419327, + 0.016579946503043175, + 0.02309137023985386, + 0.008720019832253456, + 0.014643608592450619, + 0.024332614615559578, + 0.04476086050271988, + 0.05123629793524742, + 0.05504072457551956, + 0.02754884958267212, + 0.06726426631212234, + 0.020341504365205765, + 0.030456990003585815, + 0.006537062581628561, + 0.009251239709556103, + 0.008162717334926128, + 0.006970100104808807, + 0.006658566650003195, + 0.0073988656513392925, + 0.0046263206750154495, + 0.004749289248138666, + 0.006722296122461557, + 0.004182992037385702, + 0.004141040146350861, + 0.0033295874018222094, + 0.0031113354489207268, + 0.004303376656025648, + 0.0022556683979928493, + 0.002472802298143506, + 0.042281705886125565, + 0.01848413795232773, + 0.0215595792979002, + 0.015610934235155582, + 0.010540266521275043, + 0.01735937036573887, + 0.008994361385703087, + 0.010596338659524918, + 0.004351428709924221, + 0.001706573530100286, + 0.0037578928750008345, + 0.003566868370398879, + 0.0025375839322805405, + 0.0037512474227696657, + 0.0015245156828314066, + 0.0019174268236383796, + 0.004769126418977976, + 0.0030769635923206806, + 0.004451355896890163, + 0.007156881038099527, + 0.004200615454465151, + 0.005103331990540028, + 0.0038988087326288223, + 0.003885857993736863, + 0.004116587806493044, + 0.004622783046215773, + 0.004784484393894672, + 0.0036338118370622396, + 0.00325772725045681, + 0.004449845757335424, + 0.003000757424160838, + 0.002897786907851696, + 0.006467633415013552, + 0.004615153651684523, + 0.005175197497010231, + 0.005244385451078415, + 0.0033018654212355614, + 0.005020155571401119, + 0.003309966064989567, + 0.0033707348629832268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0021971152164041996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337718591094017 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337718591094017, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337718591094017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16573183238506317, + 0.14392739534378052, + 0.12998734414577484, + 0.10815177112817764, + 0.07608149945735931, + 0.08313542604446411, + 0.1315808743238449, + 0.08102073520421982, + 0.03557666018605232, + 0.044806525111198425 + ], + "credit_map": [ + 0.022027458995580673, + 0.029113808646798134, + 0.02710583433508873, + 0.028241079300642014, + 0.01450573280453682, + 0.017399800941348076, + 0.013273796066641808, + 0.014064322225749493, + 0.01786022074520588, + 0.022558871656656265, + 0.02250129170715809, + 0.020183907821774483, + 0.022810252383351326, + 0.01673680916428566, + 0.010580096393823624, + 0.01069596130400896, + 0.014482214115560055, + 0.019255200400948524, + 0.019759368151426315, + 0.017646316438913345, + 0.011639817617833614, + 0.02215869352221489, + 0.010266302153468132, + 0.01477943453937769, + 0.010142493061721325, + 0.011786800809204578, + 0.01758401095867157, + 0.016936007887125015, + 0.012161025777459145, + 0.011521476320922375, + 0.016004061326384544, + 0.012015885673463345, + 0.01293532270938158, + 0.014128797687590122, + 0.009956580586731434, + 0.009009871631860733, + 0.009357910603284836, + 0.008532218635082245, + 0.006350534502416849, + 0.005810256581753492, + 0.01921878010034561, + 0.01074677798897028, + 0.010149125941097736, + 0.010623665526509285, + 0.01067847665399313, + 0.008362241089344025, + 0.007641454227268696, + 0.005714906845241785, + 0.04380184784531593, + 0.009657667018473148, + 0.021433383226394653, + 0.014247252605855465, + 0.011399691924452782, + 0.01601681485772133, + 0.0069897109642624855, + 0.00803450308740139, + 0.016899317502975464, + 0.006386562250554562, + 0.00975837279111147, + 0.011629551649093628, + 0.007437786553055048, + 0.011917329393327236, + 0.008043013513088226, + 0.008948801085352898, + 0.004137648735195398, + 0.004846869967877865, + 0.0051696267910301685, + 0.004888112191110849, + 0.0038488805294036865, + 0.005242443177849054, + 0.0040983837097883224, + 0.0033446967136114836, + 0.004811031743884087, + 0.003929539117962122, + 0.005116743966937065, + 0.0069819544441998005, + 0.008732103742659092, + 0.0059971590526402, + 0.005084999371320009, + 0.004152990411967039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.026337718591094017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle and handheld, imparts a subtle yet authentic tremor to the image, as if the photographer is holding his breath, overwhelmed by the overwhelming presence of the figure before him. At the center of the frame stands a white-haired, muscular martial arts master, standing like a rock in the center of an ancient Japanese dojo steeped in history. His graying hair is meticulously tied back in a bun, his weathered face etched with deep wrinkles, yet his eyes, sharp as a hawk, seem to pierce the surrounding gloom. He wears a traditional cotton kikofu (deep indigo), slightly faded from years of wear and washing, its edges showing signs of wear and tear, further highlighting his profound experience. His bare feet, firmly planted on the dark wooden floor, polished to a gleaming shine by time, his arms, veins visible, hanging naturally at his sides, his entire posture exuding an unwavering determination and composure. The dojo's backdrop is a hazy shoji paper door and a wall adorned with calligraphy depicting the character \"武.\" Wisps of dust drift slowly in the tangible, Tyndall-like beams of light slanting in from the high windows. The lighting throughout the scene is highly dramatic, with much of the space shrouded in deep shadow. Only the light precisely outlines the grandmaster's silver hair, the muscular lines of his shoulders, and the resolute expression on his face, creating an atmosphere of solemnity, power, and tranquility.", + "seed": 17896, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_5232_seed_17896_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021742145530879498, + "vlm_logprob": -0.021742145530879498, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.021742145530879498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009162875823676586 + ], + "opposite_logprob_tokens": [ + -5.134162902832031 + ], + "logprob_reward": -0.009162875823676586, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009162875823676586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22533439099788666, + 0.0760122537612915, + 0.12933982908725739, + 0.0935763344168663, + 0.15441295504570007, + 0.08194521814584732, + 0.08047179132699966, + 0.07421082258224487, + 0.03136639669537544, + 0.05333001911640167 + ], + "credit_map": [ + 0.02678089030086994, + 0.01779332011938095, + 0.05542255938053131, + 0.015688180923461914, + 0.029876377433538437, + 0.02147931046783924, + 0.0363444946706295, + 0.021949253976345062, + 0.011538399383425713, + 0.007929080165922642, + 0.00951530784368515, + 0.009620710276067257, + 0.007149702403694391, + 0.009757842868566513, + 0.012422612868249416, + 0.00807859841734171, + 0.019835306331515312, + 0.022738059982657433, + 0.014068600721657276, + 0.013462243601679802, + 0.017088234424591064, + 0.0165741927921772, + 0.014357054606080055, + 0.011216139420866966, + 0.015798214823007584, + 0.012097966857254505, + 0.012611765414476395, + 0.008801684714853764, + 0.010054681450128555, + 0.013344032689929008, + 0.009804371744394302, + 0.011063617654144764, + 0.03207099065184593, + 0.02307889051735401, + 0.024960746988654137, + 0.011489300057291985, + 0.015214486978948116, + 0.011463337577879429, + 0.01138025987893343, + 0.024754952639341354, + 0.008113622665405273, + 0.007259988225996494, + 0.018998853862285614, + 0.007541290018707514, + 0.010836422443389893, + 0.007912250235676765, + 0.006192667409777641, + 0.015090130269527435, + 0.005237623117864132, + 0.00598145741969347, + 0.020514003932476044, + 0.009852495975792408, + 0.0055046118795871735, + 0.005902748089283705, + 0.008214653469622135, + 0.01926420070230961, + 0.007046491373330355, + 0.009173261933028698, + 0.01552790030837059, + 0.007986054755747318, + 0.006611655466258526, + 0.004575984086841345, + 0.006279533728957176, + 0.017009932547807693, + 0.0024800309911370277, + 0.003209529211744666, + 0.004362433683127165, + 0.005347424186766148, + 0.004739977419376373, + 0.003202857216820121, + 0.0036321666557341814, + 0.004391977563500404, + 0.005273413844406605, + 0.004813073202967644, + 0.0114155113697052, + 0.006316035054624081, + 0.004747034516185522, + 0.004288897383958101, + 0.005077847745269537, + 0.01139820646494627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.009162875823676586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867612570524216 + ], + "opposite_logprob_tokens": [ + -3.048676013946533 + ], + "logprob_reward": -0.04867612570524216, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867612570524216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25980305671691895, + 0.06654100120067596, + 0.087928406894207, + 0.08519140630960464, + 0.10642004013061523, + 0.05710044130682945, + 0.09908326715230942, + 0.13138335943222046, + 0.056019704788923264, + 0.05052934214472771 + ], + "credit_map": [ + 0.026524240151047707, + 0.029610242694616318, + 0.05365234613418579, + 0.0259819645434618, + 0.027951505035161972, + 0.02401292510330677, + 0.04960806295275688, + 0.02246173843741417, + 0.006250976584851742, + 0.006528302561491728, + 0.008494632318615913, + 0.006586587522178888, + 0.010328508913516998, + 0.007872167974710464, + 0.013301760889589787, + 0.007178052794188261, + 0.015127424150705338, + 0.011753227561712265, + 0.012324046343564987, + 0.006510153412818909, + 0.007754626218229532, + 0.008917194791138172, + 0.011351891793310642, + 0.014189829118549824, + 0.01960940845310688, + 0.0150672672316432, + 0.012396702542901039, + 0.006841370370239019, + 0.00893188826739788, + 0.007768869400024414, + 0.006965570151805878, + 0.0076103173196315765, + 0.014744484797120094, + 0.01345882285386324, + 0.016069572418928146, + 0.009445790201425552, + 0.018995065242052078, + 0.007139261346310377, + 0.010708028450608253, + 0.01585899293422699, + 0.006779532879590988, + 0.006967846769839525, + 0.0103216040879488, + 0.007534442935138941, + 0.005965897347778082, + 0.005022467114031315, + 0.006247076205909252, + 0.008261567912995815, + 0.014781315810978413, + 0.021829379722476006, + 0.012120128609240055, + 0.010069957934319973, + 0.008849202655255795, + 0.010018641129136086, + 0.008976553566753864, + 0.012438076548278332, + 0.011228351853787899, + 0.016560597345232964, + 0.017712427303195, + 0.018121235072612762, + 0.025165783241391182, + 0.013596705161035061, + 0.013931346125900745, + 0.01506690215319395, + 0.006345947273075581, + 0.00794560369104147, + 0.009833119809627533, + 0.0061354562640190125, + 0.006776768714189529, + 0.004640556406229734, + 0.005734698846936226, + 0.008607552386820316, + 0.0038176169618964195, + 0.005042692646384239, + 0.00854627974331379, + 0.007227323949337006, + 0.007267347536981106, + 0.0037333921063691378, + 0.00839232001453638, + 0.006502366159111261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.04867612570524216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107556652277708 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.01107556652277708, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107556652277708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13780036568641663, + 0.0493677593767643, + 0.1688743382692337, + 0.10009941458702087, + 0.17982538044452667, + 0.10906993597745895, + 0.10542706400156021, + 0.048560403287410736, + 0.04436485469341278, + 0.05661051720380783 + ], + "credit_map": [ + 0.021799100562930107, + 0.011278344318270683, + 0.031153151765465736, + 0.009683813899755478, + 0.013515518046915531, + 0.01566108874976635, + 0.021531499922275543, + 0.013177850283682346, + 0.005535593722015619, + 0.004846593365073204, + 0.006955706514418125, + 0.005676487926393747, + 0.00598540436476469, + 0.007734015118330717, + 0.007748998701572418, + 0.004884956870228052, + 0.03159007057547569, + 0.029693901538848877, + 0.02010994218289852, + 0.017438359558582306, + 0.016624951735138893, + 0.024844741448760033, + 0.015438851900398731, + 0.013133522123098373, + 0.015973519533872604, + 0.018459267914295197, + 0.01359948143362999, + 0.008263780735433102, + 0.009680223651230335, + 0.015490058809518814, + 0.01059508416801691, + 0.008038001134991646, + 0.03403748571872711, + 0.027850044891238213, + 0.03282243013381958, + 0.010658119805157185, + 0.019127313047647476, + 0.014524889178574085, + 0.011600418947637081, + 0.029204683378338814, + 0.012213550508022308, + 0.008549530059099197, + 0.02718566730618477, + 0.0068441652692854404, + 0.012124044820666313, + 0.012041749432682991, + 0.006752146407961845, + 0.023359078913927078, + 0.010092861950397491, + 0.016293222084641457, + 0.025070443749427795, + 0.008474686183035374, + 0.008491133339703083, + 0.007380194962024689, + 0.007143879309296608, + 0.022480638697743416, + 0.007311161607503891, + 0.006667483597993851, + 0.005850725807249546, + 0.005122849252074957, + 0.0045942747965455055, + 0.008418504148721695, + 0.004440224729478359, + 0.0061551788821816444, + 0.005771300755441189, + 0.010668488219380379, + 0.0062567624263465405, + 0.004295307211577892, + 0.0038146318402141333, + 0.005111489444971085, + 0.0034532134886831045, + 0.00499366270378232, + 0.0071398732252418995, + 0.011256737634539604, + 0.01110842451453209, + 0.004595350939780474, + 0.002987401792779565, + 0.007487768307328224, + 0.003694825805723667, + 0.008340136148035526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.01107556652277708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017122859135270119 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017122859135270119, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017122859135270119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08588167279958725, + 0.12906020879745483, + 0.1097550317645073, + 0.03767479211091995, + 0.06617176532745361, + 0.0398808978497982, + 0.1853344738483429, + 0.1326352059841156, + 0.11259718984365463, + 0.10100872069597244 + ], + "credit_map": [ + 0.0158732570707798, + 0.007312098052352667, + 0.01294440682977438, + 0.006134460214525461, + 0.011658812873065472, + 0.0115512665361166, + 0.012115474790334702, + 0.008291896432638168, + 0.010142575018107891, + 0.009918542578816414, + 0.018858490511775017, + 0.014981738291680813, + 0.01811838708817959, + 0.015028322115540504, + 0.029974432662129402, + 0.01203772984445095, + 0.02590281143784523, + 0.02200320176780224, + 0.013870482333004475, + 0.007797892205417156, + 0.008657479658722878, + 0.014539469964802265, + 0.007905572652816772, + 0.009078134782612324, + 0.007515952456742525, + 0.003446036484092474, + 0.0051689185202121735, + 0.003188766073435545, + 0.0058571454137563705, + 0.003705837531015277, + 0.0048818811774253845, + 0.003910258878022432, + 0.012498406693339348, + 0.010599284432828426, + 0.010800602845847607, + 0.00418436573818326, + 0.0079546719789505, + 0.006199983414262533, + 0.004213872831314802, + 0.009720583446323872, + 0.006471517961472273, + 0.005841321777552366, + 0.006501397117972374, + 0.003837150987237692, + 0.003899125149473548, + 0.005432059522718191, + 0.003133572405204177, + 0.004764758516103029, + 0.028643611818552017, + 0.0468713752925396, + 0.02624192275106907, + 0.018311189487576485, + 0.014296529814600945, + 0.018962720409035683, + 0.013409395702183247, + 0.01859775185585022, + 0.014276609756052494, + 0.01856166496872902, + 0.03356147184967995, + 0.011073141358792782, + 0.011954456567764282, + 0.007979277521371841, + 0.008694685995578766, + 0.026533903554081917, + 0.017245154827833176, + 0.02760959044098854, + 0.019486360251903534, + 0.010312731377780437, + 0.007774603087455034, + 0.01006353460252285, + 0.007284504361450672, + 0.012820720672607422, + 0.013548307120800018, + 0.02140359953045845, + 0.022635016590356827, + 0.008107486180961132, + 0.005722078960388899, + 0.008323382586240768, + 0.005035221576690674, + 0.016233637928962708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.0017122859135270119 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808387368917465 + ], + "opposite_logprob_tokens": [ + -3.288083791732788 + ], + "logprob_reward": -0.03808387368917465, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808387368917465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11869944632053375, + 0.15707097947597504, + 0.03910909965634346, + 0.12475497275590897, + 0.09684965759515762, + 0.08800990879535675, + 0.08858394622802734, + 0.09654900431632996, + 0.1554301381111145, + 0.03494282439351082 + ], + "credit_map": [ + 0.020841920748353004, + 0.011893619783222675, + 0.01653650775551796, + 0.008213584311306477, + 0.013818072155117989, + 0.023051545023918152, + 0.014240490272641182, + 0.01010370533913374, + 0.01581697165966034, + 0.012349801138043404, + 0.022589394822716713, + 0.019892916083335876, + 0.02069290168583393, + 0.017411446198821068, + 0.03419975936412811, + 0.014117790386080742, + 0.005519164260476828, + 0.005628787446767092, + 0.004939565900713205, + 0.004536270163953304, + 0.005362675990909338, + 0.005528540350496769, + 0.0046653891913592815, + 0.002928704721853137, + 0.027763552963733673, + 0.017501337453722954, + 0.01741153560578823, + 0.0095509709790349, + 0.012203419581055641, + 0.01772543229162693, + 0.010349816642701626, + 0.012248901650309563, + 0.01928548514842987, + 0.017758077010512352, + 0.014195706695318222, + 0.006971878930926323, + 0.011250708252191544, + 0.007316333241760731, + 0.007837723009288311, + 0.012233740650117397, + 0.01138297002762556, + 0.007872241549193859, + 0.022643305361270905, + 0.007284374441951513, + 0.009379887953400612, + 0.005261074751615524, + 0.005947590339928865, + 0.01823846437036991, + 0.010201912373304367, + 0.016442852094769478, + 0.014979644678533077, + 0.010254468768835068, + 0.010228893719613552, + 0.006938702426850796, + 0.006866934709250927, + 0.012670536525547504, + 0.009464249014854431, + 0.007355593144893646, + 0.024368582293391228, + 0.008704185485839844, + 0.00894914660602808, + 0.006707437336444855, + 0.008030089549720287, + 0.022969724610447884, + 0.01904917322099209, + 0.027836723253130913, + 0.040665995329618454, + 0.010531684383749962, + 0.007584968116134405, + 0.013561418280005455, + 0.0077135940082371235, + 0.028486579656600952, + 0.004252798855304718, + 0.004412451758980751, + 0.0050952318124473095, + 0.004595703911036253, + 0.0037814092356711626, + 0.0037593042943626642, + 0.00434453971683979, + 0.004701385274529457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.03808387368917465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle and handheld, imparts a subtle yet authentic tremor to the image, as if the photographer is holding his breath, overwhelmed by the overwhelming presence of the figure before him. At the center of the frame stands a white-haired, muscular martial arts master, standing like a rock in the center of an ancient Japanese dojo steeped in history. His graying hair is meticulously tied back in a bun, his weathered face etched with deep wrinkles, yet his eyes, sharp as a hawk, seem to pierce the surrounding gloom. He wears a traditional cotton kikofu (deep indigo), slightly faded from years of wear and washing, its edges showing signs of wear and tear, further highlighting his profound experience. His bare feet, firmly planted on the dark wooden floor, polished to a gleaming shine by time, his arms, veins visible, hanging naturally at his sides, his entire posture exuding an unwavering determination and composure. The dojo's backdrop is a hazy shoji paper door and a wall adorned with calligraphy depicting the character \"武.\" Wisps of dust drift slowly in the tangible, Tyndall-like beams of light slanting in from the high windows. The lighting throughout the scene is highly dramatic, with much of the space shrouded in deep shadow. Only the light precisely outlines the grandmaster's silver hair, the muscular lines of his shoulders, and the resolute expression on his face, creating an atmosphere of solemnity, power, and tranquility.", + "seed": 17897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_13/index_5232_seed_17897_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029178553679957985, + "vlm_logprob": -0.029178553679957985, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.029178553679957985, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016405105590820312 + ], + "opposite_logprob_tokens": [ + -4.26640510559082 + ], + "logprob_reward": -0.016405105590820312, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016405105590820312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07196811586618423, + 0.023039547726511955, + 0.015515816397964954, + 0.014363660477101803, + 0.01776125840842724, + 0.01915322057902813, + 0.7959915399551392, + 0.016211699694395065, + 0.013986960053443909, + 0.012008128687739372 + ], + "credit_map": [ + 0.009069123305380344, + 0.012948637828230858, + 0.009711642749607563, + 0.006645039655268192, + 0.008666004985570908, + 0.009113937616348267, + 0.0034067907836288214, + 0.012406942434608936, + 0.002383527113124728, + 0.0027629174292087555, + 0.003167416201904416, + 0.0030888887122273445, + 0.003085246542468667, + 0.0035503124818205833, + 0.0019568558782339096, + 0.003044383367523551, + 0.001402122899889946, + 0.0013698397669941187, + 0.002418042626231909, + 0.0015345256542786956, + 0.002878546016290784, + 0.0019525240641087294, + 0.0015692985616624355, + 0.002390916459262371, + 0.001958961132913828, + 0.0013692069333046675, + 0.0024175604339689016, + 0.001942618633620441, + 0.001722149783745408, + 0.0017467689467594028, + 0.001374624203890562, + 0.0018317715730518103, + 0.0014502763515338302, + 0.000950882094912231, + 0.00449684401974082, + 0.001726083573885262, + 0.001685722847469151, + 0.0014745777007192373, + 0.0023410720750689507, + 0.0036358004435896873, + 0.0017075750511139631, + 0.001362882903777063, + 0.0035025998950004578, + 0.0017447532154619694, + 0.0014198222197592258, + 0.0039763906970620155, + 0.0018735132180154324, + 0.003565683262422681, + 0.03947955742478371, + 0.021844660863280296, + 0.3167169690132141, + 0.044689375907182693, + 0.01797621324658394, + 0.0261966772377491, + 0.033378951251506805, + 0.29570916295051575, + 0.0010889723198488355, + 0.0019666869193315506, + 0.0027822910342365503, + 0.001933556399308145, + 0.0016398814041167498, + 0.0024126507341861725, + 0.00202240445651114, + 0.002365255495533347, + 0.0008657044381834567, + 0.0012166410451754928, + 0.001746947062201798, + 0.0019702366553246975, + 0.0013166384305804968, + 0.0022415779531002045, + 0.0027453338261693716, + 0.0018838803516700864, + 0.001072088023647666, + 0.0015878435224294662, + 0.0015913737006485462, + 0.0012619191547855735, + 0.0014049072051420808, + 0.0017744854558259249, + 0.0019094878807663918, + 0.0014060235116630793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.016405105590820312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004938274621964 + ], + "opposite_logprob_tokens": [ + -2.695049285888672 + ], + "logprob_reward": -0.07004938274621964, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004938274621964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.168767049908638, + 0.06950576603412628, + 0.0680367648601532, + 0.04098246991634369, + 0.037583619356155396, + 0.0739579051733017, + 0.38843294978141785, + 0.06619633734226227, + 0.037534527480602264, + 0.04900258034467697 + ], + "credit_map": [ + 0.024969127029180527, + 0.02044643647968769, + 0.03463858366012573, + 0.01416258979588747, + 0.016294535249471664, + 0.027467619627714157, + 0.0070894630625844, + 0.023698709905147552, + 0.0076293484307825565, + 0.010148540139198303, + 0.008655949495732784, + 0.008689358830451965, + 0.010912662371993065, + 0.01060534454882145, + 0.006050887517631054, + 0.006813672371208668, + 0.006171027198433876, + 0.0066276779398322105, + 0.014922006987035275, + 0.008485954254865646, + 0.007416635286062956, + 0.007155804429203272, + 0.004305908922106028, + 0.012951754033565521, + 0.002490957733243704, + 0.004378408659249544, + 0.007241417653858662, + 0.005457731895148754, + 0.005780150648206472, + 0.005478215403854847, + 0.004077736288309097, + 0.006077858619391918, + 0.004067382775247097, + 0.0023859310895204544, + 0.006843391805887222, + 0.004664130043238401, + 0.003227472770959139, + 0.004915301688015461, + 0.006018959451466799, + 0.005461050663143396, + 0.005513667594641447, + 0.004555040504783392, + 0.02102752774953842, + 0.005094275809824467, + 0.005258089862763882, + 0.004721608944237232, + 0.006626889109611511, + 0.021160801872611046, + 0.021175513043999672, + 0.01141295861452818, + 0.14821355044841766, + 0.024139640852808952, + 0.00981095153838396, + 0.012650957331061363, + 0.021750887855887413, + 0.13927850127220154, + 0.006422285456210375, + 0.010043456219136715, + 0.00916483998298645, + 0.006836516782641411, + 0.005065568722784519, + 0.014177163131535053, + 0.006144267972558737, + 0.008342244662344456, + 0.002447504783049226, + 0.0037804930470883846, + 0.004809531848877668, + 0.005426579620689154, + 0.0037447097711265087, + 0.0058232564479112625, + 0.0067968107759952545, + 0.004705640021711588, + 0.00632227910682559, + 0.006932341493666172, + 0.007480138447135687, + 0.0038455729372799397, + 0.004229388665407896, + 0.007523688022047281, + 0.005847914610058069, + 0.006821256596595049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.07004938274621964 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192831702530384 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192831702530384, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192831702530384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06671634316444397, + 0.07046458125114441, + 0.027595771476626396, + 0.028360027819871902, + 0.07254672795534134, + 0.13564297556877136, + 0.23359991610050201, + 0.19036918878555298, + 0.08832017332315445, + 0.08638434112071991 + ], + "credit_map": [ + 0.01256686169654131, + 0.008205877617001534, + 0.010652526281774044, + 0.005945709999650717, + 0.00830249022692442, + 0.007132996805012226, + 0.003736533224582672, + 0.01017333846539259, + 0.008633571676909924, + 0.004357106052339077, + 0.011377574875950813, + 0.006545167416334152, + 0.004723187070339918, + 0.018557464703917503, + 0.0026852230075746775, + 0.013585273176431656, + 0.0035597472451627254, + 0.002299042185768485, + 0.005632941145449877, + 0.0028636918868869543, + 0.002482364885509014, + 0.003802657127380371, + 0.0016276264796033502, + 0.0053276983089745045, + 0.00396800646558404, + 0.003111628582701087, + 0.004662086255848408, + 0.0028301964048296213, + 0.0027489131316542625, + 0.0032814359292387962, + 0.002376371994614601, + 0.005381384864449501, + 0.008526462130248547, + 0.006330209784209728, + 0.01267924252897501, + 0.003918945323675871, + 0.007185288239270449, + 0.015017535537481308, + 0.007879314012825489, + 0.01100972294807434, + 0.01821010932326317, + 0.01322630513459444, + 0.029004309326410294, + 0.010591177269816399, + 0.009017248637974262, + 0.016915300861001015, + 0.010712106712162495, + 0.027966389432549477, + 0.012934419326484203, + 0.00749983498826623, + 0.0908854678273201, + 0.01336639653891325, + 0.006328243762254715, + 0.007973901927471161, + 0.011160223744809628, + 0.08345142751932144, + 0.01735767163336277, + 0.030527712777256966, + 0.03527887910604477, + 0.013374022208154202, + 0.014385557733476162, + 0.038827210664749146, + 0.01215716265141964, + 0.028460973873734474, + 0.007161485496908426, + 0.019668029621243477, + 0.010476761497557163, + 0.00581238092854619, + 0.006806211546063423, + 0.02106177620589733, + 0.007275035139173269, + 0.010058483108878136, + 0.012432449497282505, + 0.011715872213244438, + 0.015209588222205639, + 0.00553635461255908, + 0.004784376826137304, + 0.014653687365353107, + 0.00919746421277523, + 0.012854538857936859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.014192831702530384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021972341928631067 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021972341928631067, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021972341928631067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16264848411083221, + 0.06528409570455551, + 0.05077460780739784, + 0.027814514935016632, + 0.024378132075071335, + 0.03712458536028862, + 0.4560697674751282, + 0.08688231557607651, + 0.026649344712495804, + 0.06237413361668587 + ], + "credit_map": [ + 0.024650778621435165, + 0.025468794628977776, + 0.02586030773818493, + 0.013496306724846363, + 0.018577223643660545, + 0.01517329178750515, + 0.009397825226187706, + 0.03002394363284111, + 0.007530851289629936, + 0.0060386559925973415, + 0.009652510285377502, + 0.006063157692551613, + 0.005790261086076498, + 0.016318779438734055, + 0.003111903788521886, + 0.010777974501252174, + 0.005730289500206709, + 0.004705773200839758, + 0.009193775244057178, + 0.005013452377170324, + 0.004337015096098185, + 0.009382016025483608, + 0.003298662370070815, + 0.009113624691963196, + 0.001854453468695283, + 0.003133087884634733, + 0.004876235965639353, + 0.004235848318785429, + 0.0036381122190505266, + 0.0034498972818255424, + 0.00290711410343647, + 0.003719767089933157, + 0.0016982476226985455, + 0.002096205949783325, + 0.004395525436848402, + 0.0027319008950144053, + 0.003632382955402136, + 0.0025062093045562506, + 0.003712882986292243, + 0.0036047750618308783, + 0.0062491209246218204, + 0.0038954445626586676, + 0.006492092274129391, + 0.003575734794139862, + 0.002615524223074317, + 0.00399762112647295, + 0.0047723292373120785, + 0.005526717286556959, + 0.023488830775022507, + 0.012765935622155666, + 0.1804918795824051, + 0.025551309809088707, + 0.010847801342606544, + 0.014893273822963238, + 0.019487565383315086, + 0.16854320466518402, + 0.00833878293633461, + 0.01326051726937294, + 0.0159333273768425, + 0.006494716741144657, + 0.00676118116825819, + 0.017363764345645905, + 0.005919551011174917, + 0.012810478918254375, + 0.0020664134062826633, + 0.003989999182522297, + 0.0036943010054528713, + 0.003302352037280798, + 0.002276053884997964, + 0.004461509175598621, + 0.0036224080249667168, + 0.003236309392377734, + 0.004895369987934828, + 0.007383432704955339, + 0.013515917584300041, + 0.006916326470673084, + 0.004456930328160524, + 0.006344818510115147, + 0.007087501231580973, + 0.011773839592933655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0021972341928631067 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot of a white-haired, muscular martial arts master standing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a traditional cotton kikofu that is deep indigo in color and shows signs of wear and fading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting an ancient Japanese dojo featuring a hazy shoji paper door and a wall with calligraphy of the character '武'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting highly dramatic with deep shadows and visible beams of light creating a Tyndall effect with drifting dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle with a handheld tremor effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304821416735649 + ], + "opposite_logprob_tokens": [ + -3.168048143386841 + ], + "logprob_reward": -0.04304821416735649, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304821416735649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17385464906692505, + 0.055292900651693344, + 0.08677522093057632, + 0.0690331757068634, + 0.0941774770617485, + 0.08277558535337448, + 0.2990604043006897, + 0.04751892015337944, + 0.038062721490859985, + 0.053448956459760666 + ], + "credit_map": [ + 0.02641286328434944, + 0.030355650931596756, + 0.02116777002811432, + 0.01628858782351017, + 0.01848907209932804, + 0.023722995072603226, + 0.011228780262172222, + 0.026188937947154045, + 0.006469607353210449, + 0.007889153435826302, + 0.008156155236065388, + 0.0058264038525521755, + 0.0053991698659956455, + 0.010021403431892395, + 0.0030623399652540684, + 0.00846866611391306, + 0.006464843172580004, + 0.00912714097648859, + 0.014048280194401741, + 0.00924694910645485, + 0.014357809908688068, + 0.011914335191249847, + 0.008101747371256351, + 0.013514118269085884, + 0.006405987311154604, + 0.006198003888130188, + 0.011935907416045666, + 0.010295778512954712, + 0.007183298468589783, + 0.009379608556628227, + 0.006711374036967754, + 0.010923221707344055, + 0.006777728907763958, + 0.006029197014868259, + 0.023684091866016388, + 0.008865914307534695, + 0.010445816442370415, + 0.006722822319716215, + 0.010414482094347477, + 0.021237418055534363, + 0.008431881666183472, + 0.00814588088542223, + 0.010886849835515022, + 0.012234830297529697, + 0.007967473939061165, + 0.006527750752866268, + 0.015954775735735893, + 0.012626145035028458, + 0.012918379157781601, + 0.009212934412062168, + 0.11154099553823471, + 0.022120090201497078, + 0.0090663256123662, + 0.012310334481298923, + 0.016925310716032982, + 0.10496604442596436, + 0.0029659883584827185, + 0.004644600208848715, + 0.006040818057954311, + 0.007899179123342037, + 0.004712602589279413, + 0.006579183042049408, + 0.00902443379163742, + 0.005652112420648336, + 0.003078232519328594, + 0.004036667291074991, + 0.004487626254558563, + 0.006327764131128788, + 0.0034677220974117517, + 0.0040488350205123425, + 0.007296532392501831, + 0.005319342948496342, + 0.0038122734986245632, + 0.002807947341352701, + 0.016361944377422333, + 0.00405234145000577, + 0.003221326507627964, + 0.0030309048015624285, + 0.00502663804218173, + 0.015135579742491245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.04304821416735649 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_14/index_tensor([7243, 7243])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_14/index_tensor([7243, 7243])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3b9b40b17823646a4df7d202d42b862f4f2db513 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_14/index_tensor([7243, 7243])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, handheld shot, simulating the effects of night vision goggles, bathed in a grainy, monochromatic phosphorescent green shimmering with digital noise, slowly and steadily follows a towering skeletal figure. The figure is draped in a tattered, dark cloak that sways subtly with its movements, its weathered bones gleaming palely in the eerie green light. One of its gnarled hands grips a massive sickle. The wooden handle is ancient and curved, and the long, curved blade, rusted and chipped, reflects a harsh, bright green streak in the night vision lens's enhanced effect. The figure moves with a ghostly, leisurely pace through a long-abandoned cemetery, its skull turning slightly from side to side, its hollow eye sockets seeming to survey its surroundings. In the background, overgrown and crooked, ancient tombstones appear blurred and distorted in the green glow, their outlines draped in the dark silhouettes of ivy. The ground is covered with thick dead leaves, and each step of the skeleton is silent, without disturbing the dead silence of the land, which together creates an eerie sense of surveillance and supernatural tension, as if the audience is peeking at this ominous exploration process through the first-person perspective of some inhuman device.", + "seed": 9339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_14/index_7243_seed_9339_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.067164873983711, + "vlm_logprob": -0.067164873983711, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.067164873983711, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2845962941646576 + ], + "opposite_logprob_tokens": [ + -1.90959632396698 + ], + "logprob_reward": -0.2845962941646576, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.2845962941646576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19783642888069153, + 0.11127548664808273, + 0.10932237654924393, + 0.06814920157194138, + 0.09545093774795532, + 0.09020478278398514, + 0.07038960605859756, + 0.07844138890504837, + 0.09964435547590256, + 0.07928550243377686 + ], + "credit_map": [ + 0.023636316880583763, + 0.03438163176178932, + 0.025678791105747223, + 0.02446332387626171, + 0.0249639842659235, + 0.028081607073545456, + 0.018914982676506042, + 0.017715774476528168, + 0.009501277469098568, + 0.007778294384479523, + 0.021916847676038742, + 0.009260677732527256, + 0.008491219952702522, + 0.01572287641465664, + 0.0110661955550313, + 0.027538083493709564, + 0.007304492872208357, + 0.011443058960139751, + 0.022090911865234375, + 0.014586737379431725, + 0.007819854654371738, + 0.013022677972912788, + 0.011506704613566399, + 0.02154793217778206, + 0.006781783886253834, + 0.007194324396550655, + 0.009381098672747612, + 0.008552850224077702, + 0.007730100769549608, + 0.011898528784513474, + 0.007776729296892881, + 0.008833780884742737, + 0.006257058586925268, + 0.007109609432518482, + 0.02435324713587761, + 0.008970871567726135, + 0.010025574825704098, + 0.006856170017272234, + 0.006615017540752888, + 0.025263376533985138, + 0.010401755571365356, + 0.012605482712388039, + 0.014727224595844746, + 0.016795504838228226, + 0.008240168914198875, + 0.010522661730647087, + 0.007239434868097305, + 0.009672545827925205, + 0.007882335223257542, + 0.010110558941960335, + 0.008849815465509892, + 0.010811780579388142, + 0.007904655300080776, + 0.008354292251169682, + 0.00915633887052536, + 0.007319826632738113, + 0.00663434574380517, + 0.008963675238192081, + 0.012135952711105347, + 0.01140580978244543, + 0.011548968032002449, + 0.010128495283424854, + 0.008786872029304504, + 0.008837263099849224, + 0.012432374060153961, + 0.01769234612584114, + 0.013140792027115822, + 0.014535010792315006, + 0.01455437857657671, + 0.01017074380069971, + 0.008669240400195122, + 0.008449457585811615, + 0.006609137170016766, + 0.00691014900803566, + 0.01203117985278368, + 0.008959523402154446, + 0.010608270764350891, + 0.015361989848315716, + 0.010849707759916782, + 0.007955539971590042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2845962941646576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004156403709203005 + ], + "opposite_logprob_tokens": [ + -5.504156589508057 + ], + "logprob_reward": -0.004156403709203005, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004156403709203005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1335299164056778, + 0.06133818253874779, + 0.3224538564682007, + 0.09646780788898468, + 0.12344013154506683, + 0.03587181493639946, + 0.0715700313448906, + 0.08465298265218735, + 0.03118745982646942, + 0.03948773443698883 + ], + "credit_map": [ + 0.012742793187499046, + 0.025041691958904266, + 0.01469850167632103, + 0.030463196337223053, + 0.01258523017168045, + 0.01687907986342907, + 0.011545897461473942, + 0.009573533199727535, + 0.005438284017145634, + 0.004214733373373747, + 0.014927994459867477, + 0.005124875344336033, + 0.0044193752110004425, + 0.008761918172240257, + 0.004495648667216301, + 0.013955355621874332, + 0.021097732707858086, + 0.013259549625217915, + 0.0992916077375412, + 0.01721161976456642, + 0.014058017171919346, + 0.03833809867501259, + 0.02045205980539322, + 0.09874517470598221, + 0.005191560368984938, + 0.005175401922315359, + 0.005665413569658995, + 0.006391027942299843, + 0.006680975668132305, + 0.03307220712304115, + 0.005887691862881184, + 0.02840353362262249, + 0.006309507414698601, + 0.0074772946536540985, + 0.03740277886390686, + 0.006839913781732321, + 0.00958284456282854, + 0.007478978950530291, + 0.008851343765854836, + 0.039497483521699905, + 0.003800379578024149, + 0.005305042490363121, + 0.004523148760199547, + 0.005191565956920385, + 0.005144720897078514, + 0.005186181981116533, + 0.0029400780331343412, + 0.0037806991022080183, + 0.007512510288506746, + 0.007176309358328581, + 0.009400124661624432, + 0.010278236120939255, + 0.008282732218503952, + 0.008378489874303341, + 0.010581126436591148, + 0.009960504248738289, + 0.006703423801809549, + 0.011045411229133606, + 0.01110822707414627, + 0.010973246768116951, + 0.014474551193416119, + 0.013743313029408455, + 0.008676326833665371, + 0.007928483188152313, + 0.0039118616841733456, + 0.005121120251715183, + 0.004458462819457054, + 0.0038992001209408045, + 0.004937720950692892, + 0.002913587260991335, + 0.002888341899961233, + 0.00305716460570693, + 0.0037304721772670746, + 0.005072378553450108, + 0.008366711437702179, + 0.005641304887831211, + 0.0027766707353293896, + 0.0043423958122730255, + 0.005196979735046625, + 0.004360826686024666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.004156403709203005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003203381784260273 + ], + "opposite_logprob_tokens": [ + -5.753203392028809 + ], + "logprob_reward": -0.003203381784260273, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003203381784260273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061084602028131485, + 0.07185866683721542, + 0.13679374754428864, + 0.19143465161323547, + 0.354709655046463, + 0.028093378990888596, + 0.08472783863544464, + 0.012589775025844574, + 0.020837508141994476, + 0.03787025064229965 + ], + "credit_map": [ + 0.005761486478149891, + 0.008429394103586674, + 0.006286053452640772, + 0.01383612584322691, + 0.008282596245408058, + 0.006830618251115084, + 0.005512771662324667, + 0.006145549938082695, + 0.006188543979078531, + 0.005433480255305767, + 0.018717817962169647, + 0.004979352932423353, + 0.005970187950879335, + 0.007682199589908123, + 0.009182109497487545, + 0.013704962097108364, + 0.008013058453798294, + 0.013431066647171974, + 0.027454828843474388, + 0.02131655625998974, + 0.013154292479157448, + 0.01180514506995678, + 0.01200978271663189, + 0.029609009623527527, + 0.015375819988548756, + 0.023127034306526184, + 0.020361844450235367, + 0.023016775026917458, + 0.038654036819934845, + 0.0286749005317688, + 0.02341422252357006, + 0.018809998407959938, + 0.02000819705426693, + 0.02083386853337288, + 0.1014145165681839, + 0.02206285297870636, + 0.0350012369453907, + 0.018551884219050407, + 0.02610836923122406, + 0.1107286885380745, + 0.004084300249814987, + 0.003123914822936058, + 0.004195134621113539, + 0.00395590765401721, + 0.004537854343652725, + 0.002210397506132722, + 0.002949932124465704, + 0.0030359348747879267, + 0.007261381018906832, + 0.013982838951051235, + 0.012108031660318375, + 0.015410611405968666, + 0.007417867425829172, + 0.007978403940796852, + 0.01250320766121149, + 0.008065485395491123, + 0.0011657527647912502, + 0.002006387570872903, + 0.001505271065980196, + 0.001389811746776104, + 0.0018923503812402487, + 0.0020063247065991163, + 0.0011582423467189074, + 0.0014656332787126303, + 0.0024681109935045242, + 0.002902975073084235, + 0.003025222336873412, + 0.002723953453823924, + 0.0026042871177196503, + 0.003026396268978715, + 0.002082166029140353, + 0.0020043945405632257, + 0.002766265068203211, + 0.004204985685646534, + 0.00414087250828743, + 0.004707659129053354, + 0.00521453283727169, + 0.007702033966779709, + 0.004360932856798172, + 0.004772965330630541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.003203381784260273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058464288711548 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.02058464288711548, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058464288711548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15933683514595032, + 0.17257124185562134, + 0.10181529074907303, + 0.08631279319524765, + 0.11052978038787842, + 0.08696994185447693, + 0.0741145983338356, + 0.13600899279117584, + 0.04207891225814819, + 0.030261604115366936 + ], + "credit_map": [ + 0.016061631962656975, + 0.024214059114456177, + 0.023875990882515907, + 0.03193201869726181, + 0.022848350927233696, + 0.0151978163048625, + 0.013572223484516144, + 0.01163473166525364, + 0.016764463856816292, + 0.0094094043597579, + 0.05397439002990723, + 0.01430332101881504, + 0.008519096300005913, + 0.016865065321326256, + 0.010165567509829998, + 0.04256991669535637, + 0.006419924553483725, + 0.008313785307109356, + 0.025258619338274002, + 0.009224402718245983, + 0.005865097977221012, + 0.01577110029757023, + 0.00775790773332119, + 0.02320445142686367, + 0.005304784048348665, + 0.004945769440382719, + 0.005380503833293915, + 0.004483018536120653, + 0.005072297994047403, + 0.030443482100963593, + 0.004833408631384373, + 0.02584952302277088, + 0.005843601655215025, + 0.007832132279872894, + 0.028369799256324768, + 0.010327000170946121, + 0.013190147466957569, + 0.0072625926695764065, + 0.007285315077751875, + 0.030419182032346725, + 0.008735316805541515, + 0.01064308825880289, + 0.0099814273416996, + 0.012107880786061287, + 0.015234505757689476, + 0.01403977070003748, + 0.007106288336217403, + 0.009121661074459553, + 0.007325080223381519, + 0.01023950893431902, + 0.010491946712136269, + 0.012676376849412918, + 0.0076560103334486485, + 0.007887139916419983, + 0.008975078351795673, + 0.008863450959324837, + 0.010297425091266632, + 0.018797563388943672, + 0.016569992527365685, + 0.018708527088165283, + 0.022229401394724846, + 0.02188146859407425, + 0.014565525576472282, + 0.01295907236635685, + 0.005713361315429211, + 0.007155953906476498, + 0.005155016668140888, + 0.006544353440403938, + 0.004772081505507231, + 0.0040495553985238075, + 0.004738490097224712, + 0.003950098529458046, + 0.0032951505854725838, + 0.0025804981123656034, + 0.0064370823092758656, + 0.004400548059493303, + 0.0026694508269429207, + 0.002873034914955497, + 0.003838498378172517, + 0.004167337901890278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.02058464288711548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283647373318672 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283647373318672, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283647373318672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1275976449251175, + 0.06226122006773949, + 0.1169203370809555, + 0.1689729541540146, + 0.12428134679794312, + 0.04594706743955612, + 0.052344419062137604, + 0.08900809288024902, + 0.05957695469260216, + 0.15308986604213715 + ], + "credit_map": [ + 0.012745688669383526, + 0.019365301355719566, + 0.013757424429059029, + 0.02452262118458748, + 0.020482666790485382, + 0.012084534391760826, + 0.013544315472245216, + 0.011095110327005386, + 0.0056432136334478855, + 0.005098995752632618, + 0.01436053216457367, + 0.005155567079782486, + 0.0037804157473146915, + 0.005956993903964758, + 0.005128030199557543, + 0.017137477174401283, + 0.008944835513830185, + 0.013652043417096138, + 0.015305140055716038, + 0.024746453389525414, + 0.011469165794551373, + 0.012730702757835388, + 0.01443767175078392, + 0.015634331852197647, + 0.009708365425467491, + 0.01296299695968628, + 0.014664684422314167, + 0.012788334861397743, + 0.018409190699458122, + 0.03862188756465912, + 0.015358651988208294, + 0.04645886272192001, + 0.007062940858304501, + 0.009998301975429058, + 0.029543379321694374, + 0.01054220087826252, + 0.014842516742646694, + 0.011589105241000652, + 0.013419298455119133, + 0.027283618226647377, + 0.004901109728962183, + 0.005808359477669001, + 0.0061803217977285385, + 0.004507009405642748, + 0.00893452763557434, + 0.0062801470048725605, + 0.0040234727784991264, + 0.005312121473252773, + 0.0041644386947155, + 0.007408550940454006, + 0.006845678668469191, + 0.007727486547082663, + 0.0059349010698497295, + 0.006832952611148357, + 0.006179509684443474, + 0.0072509064339101315, + 0.006623832508921623, + 0.011557728052139282, + 0.011938653886318207, + 0.011952579021453857, + 0.013272219337522984, + 0.014563171193003654, + 0.009589462541043758, + 0.009510450065135956, + 0.009670366533100605, + 0.008743857964873314, + 0.008787663653492928, + 0.007042530458420515, + 0.006521867122501135, + 0.006129681598395109, + 0.004832615610212088, + 0.007848379202187061, + 0.012834756635129452, + 0.01139004249125719, + 0.02922622673213482, + 0.014845471829175949, + 0.02037779614329338, + 0.026789702475070953, + 0.01732667163014412, + 0.020299211144447327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.023283647373318672 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld shot, simulating the effects of night vision goggles, bathed in a grainy, monochromatic phosphorescent green shimmering with digital noise, slowly and steadily follows a towering skeletal figure. The figure is draped in a tattered, dark cloak that sways subtly with its movements, its weathered bones gleaming palely in the eerie green light. One of its gnarled hands grips a massive sickle. The wooden handle is ancient and curved, and the long, curved blade, rusted and chipped, reflects a harsh, bright green streak in the night vision lens's enhanced effect. The figure moves with a ghostly, leisurely pace through a long-abandoned cemetery, its skull turning slightly from side to side, its hollow eye sockets seeming to survey its surroundings. In the background, overgrown and crooked, ancient tombstones appear blurred and distorted in the green glow, their outlines draped in the dark silhouettes of ivy. The ground is covered with thick dead leaves, and each step of the skeleton is silent, without disturbing the dead silence of the land, which together creates an eerie sense of surveillance and supernatural tension, as if the audience is peeking at this ominous exploration process through the first-person perspective of some inhuman device.", + "seed": 9340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_14/index_7243_seed_9340_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04941267229150981, + "vlm_logprob": -0.04941267229150981, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.04941267229150981, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19495129585266113 + ], + "opposite_logprob_tokens": [ + -2.069951295852661 + ], + "logprob_reward": -0.19495129585266113, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.19495129585266113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1820783019065857, + 0.11381222307682037, + 0.021245194599032402, + 0.039150528609752655, + 0.019389890134334564, + 0.01031278632581234, + 0.5572859644889832, + 0.020767254754900932, + 0.01758251152932644, + 0.018375318497419357 + ], + "credit_map": [ + 0.008962824940681458, + 0.010851376689970493, + 0.0076003302820026875, + 0.011350090615451336, + 0.015087262727320194, + 0.06604432314634323, + 0.013565294444561005, + 0.048616793006658554, + 0.00570495193824172, + 0.007517393212765455, + 0.005361088085919619, + 0.004815851338207722, + 0.012225506827235222, + 0.035936594009399414, + 0.019858960062265396, + 0.022391866892576218, + 0.002017471706494689, + 0.0026517282240092754, + 0.00327718211337924, + 0.0025417893193662167, + 0.0026175384409725666, + 0.0030196376610547304, + 0.001901494339108467, + 0.003218352794647217, + 0.004636922385543585, + 0.0044350652024149895, + 0.00434326846152544, + 0.004559371154755354, + 0.00445141177624464, + 0.005708339158445597, + 0.005421702284365892, + 0.0055944472551345825, + 0.001979754539206624, + 0.002494766376912594, + 0.002431699773296714, + 0.002993152476847172, + 0.0020650674123317003, + 0.002330343471840024, + 0.002003704197704792, + 0.003091400722041726, + 0.0017139361007139087, + 0.0010577719658613205, + 0.0016588879516348243, + 0.0009799828985705972, + 0.0011358177289366722, + 0.0014154219534248114, + 0.0010555380722507834, + 0.0012954294215887785, + 0.021629296243190765, + 0.02409769780933857, + 0.20837317407131195, + 0.03265886381268501, + 0.00709231523796916, + 0.0343504324555397, + 0.008356492035090923, + 0.22072766721248627, + 0.0020612350199371576, + 0.0024511206429451704, + 0.0028731387574225664, + 0.0021259270142763853, + 0.00264186249114573, + 0.0034468097146600485, + 0.0022331292275339365, + 0.0029340332839637995, + 0.002009589457884431, + 0.0019243821734562516, + 0.002883108099922538, + 0.0018314101034775376, + 0.0014190587680786848, + 0.002069777576252818, + 0.0027561397291719913, + 0.0026890463195741177, + 0.002620571991428733, + 0.0015546249924227595, + 0.0026875862386077642, + 0.0025180019438266754, + 0.0019676601514220238, + 0.0016159324441105127, + 0.00271158036775887, + 0.002699360717087984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.19495129585266113 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032414060551673174 + ], + "opposite_logprob_tokens": [ + -5.753241539001465 + ], + "logprob_reward": -0.0032414060551673174, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032414060551673174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27924585342407227, + 0.0904349833726883, + 0.02012363076210022, + 0.019887743517756462, + 0.044675204902887344, + 0.020483680069446564, + 0.465805321931839, + 0.01908269338309765, + 0.03001958690583706, + 0.010241328738629818 + ], + "credit_map": [ + 0.012339669279754162, + 0.015488711185753345, + 0.011848962865769863, + 0.017421873286366463, + 0.021251434460282326, + 0.10456328839063644, + 0.020790593698620796, + 0.07554131746292114, + 0.005178040359169245, + 0.007387398276478052, + 0.004611507523804903, + 0.0049735368229448795, + 0.009401318617165089, + 0.027267878875136375, + 0.015168284997344017, + 0.01644701324403286, + 0.0018611452542245388, + 0.002217225031927228, + 0.003934919368475676, + 0.001611979678273201, + 0.0019606738351285458, + 0.004244702402502298, + 0.0017676101997494698, + 0.002525374060496688, + 0.001976581523194909, + 0.002440772019326687, + 0.00240513333119452, + 0.002407010179013014, + 0.002144880359992385, + 0.002391750691458583, + 0.0027469538617879152, + 0.003374662483111024, + 0.00539542967453599, + 0.006279872264713049, + 0.006596784573048353, + 0.006124306470155716, + 0.0039311484433710575, + 0.005638636648654938, + 0.003726094728335738, + 0.00698293000459671, + 0.0033513465896248817, + 0.0026009895373135805, + 0.004798875190317631, + 0.002043697517365217, + 0.0010070274583995342, + 0.0028503085486590862, + 0.0009658019989728928, + 0.0028656332287937403, + 0.017963502556085587, + 0.01982610486447811, + 0.17310476303100586, + 0.027971770614385605, + 0.006001739297062159, + 0.02893931046128273, + 0.008114890195429325, + 0.18388323485851288, + 0.0016318081179633737, + 0.0025172214955091476, + 0.001929036807268858, + 0.002015958074480295, + 0.003991951700299978, + 0.0023797848261892796, + 0.002448262646794319, + 0.0021686700638383627, + 0.0033322186209261417, + 0.003084823489189148, + 0.0050521185621619225, + 0.003574018832296133, + 0.0038603704888373613, + 0.0028512398712337017, + 0.004171060398221016, + 0.004093737807124853, + 0.0013626512372866273, + 0.0012717150384560227, + 0.0019896053709089756, + 0.0011581882135942578, + 0.0010057144099846482, + 0.0009679857757873833, + 0.001236424781382084, + 0.0012490429216995835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0032414060551673174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019471748964861035 + ], + "opposite_logprob_tokens": [ + -6.251947402954102 + ], + "logprob_reward": -0.0019471748964861035, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019471748964861035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05341828614473343, + 0.01204126887023449, + 0.03271840140223503, + 0.023748166859149933, + 0.022289082407951355, + 0.021200457587838173, + 0.7877741456031799, + 0.015036286786198616, + 0.0221586711704731, + 0.009615197777748108 + ], + "credit_map": [ + 0.002851477824151516, + 0.004182241391390562, + 0.0036187421064823866, + 0.0045416709035634995, + 0.006173123139888048, + 0.014145425520837307, + 0.006680344231426716, + 0.011225264519453049, + 0.0008661500760354102, + 0.0015921557787805796, + 0.0008480972610414028, + 0.001313744462095201, + 0.001249833614565432, + 0.0024754581972956657, + 0.0018830315675586462, + 0.0018127979710698128, + 0.003673811908811331, + 0.0039037347305566072, + 0.005510455463081598, + 0.003606901504099369, + 0.003975615371018648, + 0.006141749210655689, + 0.0024480377323925495, + 0.0034580982755869627, + 0.002381888683885336, + 0.002340673701837659, + 0.00266264658421278, + 0.003431017743423581, + 0.0034355977550148964, + 0.0033352384343743324, + 0.0026995083317160606, + 0.003461596090346575, + 0.0022414280101656914, + 0.002488135127350688, + 0.0024141394533216953, + 0.0033682044595479965, + 0.004222480114549398, + 0.0023948948364704847, + 0.0031229406595230103, + 0.0020368618424981833, + 0.0018584185745567083, + 0.002417027484625578, + 0.0024925018660724163, + 0.00365957198664546, + 0.0037309555336833, + 0.0022613494656980038, + 0.0025489351246505976, + 0.0022317010443657637, + 0.03079320304095745, + 0.0338386632502079, + 0.2945343255996704, + 0.04649612307548523, + 0.009983133524656296, + 0.04838570952415466, + 0.01247167307883501, + 0.31127139925956726, + 0.0014122987631708384, + 0.0021147732622921467, + 0.0019759461283683777, + 0.0021677305921912193, + 0.0013442994095385075, + 0.0019476445158943534, + 0.002235834486782551, + 0.001837759860791266, + 0.0024163592606782913, + 0.0027025879826396704, + 0.0032224971801042557, + 0.002385259373113513, + 0.0025425637140870094, + 0.003251743270084262, + 0.002732981462031603, + 0.002904680324718356, + 0.0011396839981898665, + 0.0010769644286483526, + 0.00140021787956357, + 0.0012079575099050999, + 0.0013258809922263026, + 0.0007860889891162515, + 0.001314195222221315, + 0.0013642085250467062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.0019471748964861035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634573169052601 + ], + "opposite_logprob_tokens": [ + -3.651345729827881 + ], + "logprob_reward": -0.02634573169052601, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634573169052601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0483354777097702, + 0.0448998399078846, + 0.008690381422638893, + 0.05854355916380882, + 0.010830895975232124, + 0.01584727130830288, + 0.7273315191268921, + 0.034193672239780426, + 0.03186821565032005, + 0.01945924200117588 + ], + "credit_map": [ + 0.004049066919833422, + 0.004240135662257671, + 0.003374434309080243, + 0.005433883983641863, + 0.0075958725064992905, + 0.008740189485251904, + 0.009807082824409008, + 0.0050948127172887325, + 0.0022567950654774904, + 0.0032284108456224203, + 0.0024691219441592693, + 0.0024806896690279245, + 0.004132024012506008, + 0.013271630741655827, + 0.007544254418462515, + 0.009516916237771511, + 0.0010037751635536551, + 0.0006037996499799192, + 0.0016487700631842017, + 0.0006064733024686575, + 0.0008925888687372208, + 0.0020308862440288067, + 0.0006427766056731343, + 0.001261310651898384, + 0.0069971028715372086, + 0.004856535233557224, + 0.006393870338797569, + 0.007752897683531046, + 0.007658944930881262, + 0.010197569616138935, + 0.006656689569354057, + 0.008029947988688946, + 0.0013528118142858148, + 0.0012176460586488247, + 0.00153376255184412, + 0.002220561495050788, + 0.0008314085425809026, + 0.0009621392237022519, + 0.0009711496531963348, + 0.0017414161702618003, + 0.0031578391790390015, + 0.0017982844728976488, + 0.003731207689270377, + 0.0017942184349521995, + 0.0008041528053581715, + 0.0016934274462983012, + 0.0009125307551585138, + 0.0019556116312742233, + 0.02838907763361931, + 0.0315980426967144, + 0.27069970965385437, + 0.04267444089055061, + 0.009342276491224766, + 0.04511333256959915, + 0.011761166155338287, + 0.2877534329891205, + 0.0030934293754398823, + 0.004670134745538235, + 0.0035002727527171373, + 0.0036299594212323427, + 0.006075460463762283, + 0.0041554998606443405, + 0.003536772448569536, + 0.0055321441031992435, + 0.0035214864183217287, + 0.004286143928766251, + 0.006323722656816244, + 0.003308686427772045, + 0.0026615881361067295, + 0.005444227252155542, + 0.0031551136635243893, + 0.0031672464683651924, + 0.002503866795450449, + 0.002844147616997361, + 0.0027921211440116167, + 0.002673412673175335, + 0.0013880252372473478, + 0.002153427340090275, + 0.002382357604801655, + 0.00272188289090991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.02634573169052601 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577752962708473 + ], + "opposite_logprob_tokens": [ + -3.8955776691436768 + ], + "logprob_reward": -0.020577752962708473, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577752962708473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2029884308576584, + 0.20989638566970825, + 0.062486205250024796, + 0.06487154960632324, + 0.06271772086620331, + 0.03796011582016945, + 0.19288335740566254, + 0.06344697624444962, + 0.07499729841947556, + 0.027752015739679337 + ], + "credit_map": [ + 0.012025902047753334, + 0.0155887920409441, + 0.02465270832180977, + 0.024648914113640785, + 0.02107383869588375, + 0.04261363297700882, + 0.027977097779512405, + 0.03440754488110542, + 0.009816927835345268, + 0.01085853110998869, + 0.009718306362628937, + 0.00911899283528328, + 0.01887686364352703, + 0.07093575596809387, + 0.038894642144441605, + 0.041676364839076996, + 0.006551663391292095, + 0.006394597701728344, + 0.012614156119525433, + 0.005248804576694965, + 0.004814358428120613, + 0.012882096692919731, + 0.006182260345667601, + 0.007798265665769577, + 0.007125847041606903, + 0.009086178615689278, + 0.006981126964092255, + 0.008559592999517918, + 0.009901726618409157, + 0.007651132065802813, + 0.00771708320826292, + 0.007848857901990414, + 0.008452202193439007, + 0.007596345618367195, + 0.007873238064348698, + 0.00979340635240078, + 0.007667940575629473, + 0.006262172944843769, + 0.006724299397319555, + 0.008348113857209682, + 0.003363833064213395, + 0.004295633174479008, + 0.004217072390019894, + 0.006378457881510258, + 0.004852543585002422, + 0.004512463696300983, + 0.005216880235821009, + 0.005123233422636986, + 0.008384245447814465, + 0.009472493082284927, + 0.06574881821870804, + 0.01080939918756485, + 0.005573756992816925, + 0.011937608942389488, + 0.010007955133914948, + 0.07094908505678177, + 0.005678246263414621, + 0.010566877201199532, + 0.008158767595887184, + 0.006385007407516241, + 0.008967151865363121, + 0.007302304729819298, + 0.007916162721812725, + 0.00847245566546917, + 0.00899253785610199, + 0.006698181387037039, + 0.010398512706160545, + 0.013476196676492691, + 0.006302418652921915, + 0.006130825262516737, + 0.011606753803789616, + 0.011391873471438885, + 0.0031717794481664896, + 0.0026156450621783733, + 0.005105879157781601, + 0.003311050357297063, + 0.003132494166493416, + 0.001924527226947248, + 0.0036372188478708267, + 0.004853421356528997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.020577752962708473 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld shot, simulating the effects of night vision goggles, bathed in a grainy, monochromatic phosphorescent green shimmering with digital noise, slowly and steadily follows a towering skeletal figure. The figure is draped in a tattered, dark cloak that sways subtly with its movements, its weathered bones gleaming palely in the eerie green light. One of its gnarled hands grips a massive sickle. The wooden handle is ancient and curved, and the long, curved blade, rusted and chipped, reflects a harsh, bright green streak in the night vision lens's enhanced effect. The figure moves with a ghostly, leisurely pace through a long-abandoned cemetery, its skull turning slightly from side to side, its hollow eye sockets seeming to survey its surroundings. In the background, overgrown and crooked, ancient tombstones appear blurred and distorted in the green glow, their outlines draped in the dark silhouettes of ivy. The ground is covered with thick dead leaves, and each step of the skeleton is silent, without disturbing the dead silence of the land, which together creates an eerie sense of surveillance and supernatural tension, as if the audience is peeking at this ominous exploration process through the first-person perspective of some inhuman device.", + "seed": 9341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_14/index_7243_seed_9341_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06361335581168533, + "vlm_logprob": -0.06361335581168533, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.06361335581168533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.242726132273674 + ], + "opposite_logprob_tokens": [ + -1.9927260875701904 + ], + "logprob_reward": -0.242726132273674, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.242726132273674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14573310315608978, + 0.022296268492937088, + 0.3287602663040161, + 0.01614263653755188, + 0.2111920565366745, + 0.04167838767170906, + 0.028010915964841843, + 0.03854513540863991, + 0.11751991510391235, + 0.0501212477684021 + ], + "credit_map": [ + 0.01866786740720272, + 0.014821503311395645, + 0.016975441947579384, + 0.020199133083224297, + 0.038747549057006836, + 0.008631709031760693, + 0.011898135766386986, + 0.01579177752137184, + 0.0026851689908653498, + 0.0029811738058924675, + 0.0031837730202823877, + 0.0023028377909213305, + 0.0033019986003637314, + 0.003044100943952799, + 0.0022978363558650017, + 0.0024993803817778826, + 0.04971297085285187, + 0.033085793256759644, + 0.04358210787177086, + 0.07537448406219482, + 0.028470536693930626, + 0.04910973459482193, + 0.035754356533288956, + 0.013670315966010094, + 0.001454383134841919, + 0.001521793776191771, + 0.002741782693192363, + 0.0013766996562480927, + 0.002819992834702134, + 0.0021047471091151237, + 0.0019394038245081902, + 0.002183835254982114, + 0.00994846224784851, + 0.015924764797091484, + 0.012416339479386806, + 0.01078488864004612, + 0.01832255721092224, + 0.06534405797719955, + 0.06262388825416565, + 0.0158271212130785, + 0.0043268692679703236, + 0.003510169219225645, + 0.005191346164792776, + 0.00562063604593277, + 0.005509598180651665, + 0.006513786967843771, + 0.006706304848194122, + 0.004299679771065712, + 0.0018833392532542348, + 0.0027649805415421724, + 0.0021346905268728733, + 0.005173664540052414, + 0.00256688822992146, + 0.004049989860504866, + 0.00585805531591177, + 0.0035793096758425236, + 0.002814902225509286, + 0.004148263484239578, + 0.005384331569075584, + 0.005068343132734299, + 0.004531352315098047, + 0.005789568647742271, + 0.004324192181229591, + 0.0064841825515031815, + 0.0060454546473920345, + 0.004174126777797937, + 0.003987761214375496, + 0.018307987600564957, + 0.006476414855569601, + 0.005482853855937719, + 0.04328728839755058, + 0.0297580286860466, + 0.0061238850466907024, + 0.00501287542283535, + 0.006855660583823919, + 0.008079013787209988, + 0.005455764010548592, + 0.005955209955573082, + 0.006923303473740816, + 0.005715539213269949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.242726132273674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028745303861796856 + ], + "opposite_logprob_tokens": [ + -5.877874374389648 + ], + "logprob_reward": -0.0028745303861796856, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028745303861796856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04558150842785835, + 0.01453405525535345, + 0.16965390741825104, + 0.014497787691652775, + 0.08833476901054382, + 0.027877327054739, + 0.028356362134218216, + 0.13950443267822266, + 0.2513372302055359, + 0.22032256424427032 + ], + "credit_map": [ + 0.0059081995859742165, + 0.0031238473020493984, + 0.008017289452254772, + 0.006322761531919241, + 0.006510412786155939, + 0.00348399649374187, + 0.004387980792671442, + 0.007827023044228554, + 0.0016906161326915026, + 0.0015387084567919374, + 0.0023440043441951275, + 0.001424739370122552, + 0.0018542025936767459, + 0.0021298364736139774, + 0.00139051943551749, + 0.0021614294964820147, + 0.02523934282362461, + 0.017822109162807465, + 0.022228680551052094, + 0.0391341894865036, + 0.0159585140645504, + 0.024393634870648384, + 0.017839225009083748, + 0.007038217969238758, + 0.0013449061661958694, + 0.0015204120427370071, + 0.0024473550729453564, + 0.0012441725702956319, + 0.0029970163013786077, + 0.0019330689683556557, + 0.0013702186988666654, + 0.0016406389186158776, + 0.006035748869180679, + 0.0037021469324827194, + 0.003955716732889414, + 0.0024526191409677267, + 0.007286107167601585, + 0.029639258980751038, + 0.027847256511449814, + 0.007415912579745054, + 0.0020715089049190283, + 0.002300688996911049, + 0.002484718570485711, + 0.005367457401007414, + 0.003705498529598117, + 0.004945593420416117, + 0.0033953278325498104, + 0.0036065324675291777, + 0.003187281545251608, + 0.0024578082375228405, + 0.0037046915385872126, + 0.0030046976171433926, + 0.0025694728828966618, + 0.00477389944717288, + 0.004826504737138748, + 0.003832004964351654, + 0.008793952874839306, + 0.017122561112046242, + 0.014966328628361225, + 0.019274532794952393, + 0.016745027154684067, + 0.019594911485910416, + 0.01666579581797123, + 0.026341315358877182, + 0.018534332513809204, + 0.011017058975994587, + 0.02029496245086193, + 0.041889261454343796, + 0.020366011187434196, + 0.020460542291402817, + 0.0725308209657669, + 0.046244241297245026, + 0.022403845563530922, + 0.022637708112597466, + 0.02897164598107338, + 0.030583428218960762, + 0.02330506406724453, + 0.03112153336405754, + 0.03171192854642868, + 0.02958739921450615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.0028745303861796856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079339310526848 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079339310526848, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079339310526848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09267944097518921, + 0.02213709056377411, + 0.3077171742916107, + 0.033531345427036285, + 0.15549904108047485, + 0.02114596590399742, + 0.03688346967101097, + 0.15370634198188782, + 0.05318756774067879, + 0.12351256608963013 + ], + "credit_map": [ + 0.01053043082356453, + 0.008468921296298504, + 0.014123384840786457, + 0.012542924843728542, + 0.02195677161216736, + 0.008541764691472054, + 0.007429336663335562, + 0.009085905738174915, + 0.0026467449497431517, + 0.002262408146634698, + 0.0030977288261055946, + 0.002670958638191223, + 0.0017400308279320598, + 0.003167814575135708, + 0.0019918358884751797, + 0.00455956906080246, + 0.04898969829082489, + 0.031575337052345276, + 0.04057709500193596, + 0.06887621432542801, + 0.02620355598628521, + 0.04695015773177147, + 0.032144129276275635, + 0.012400982901453972, + 0.0029089110903441906, + 0.0037201866507530212, + 0.005538029130548239, + 0.0038283299654722214, + 0.003066537668928504, + 0.004237485118210316, + 0.006141098216176033, + 0.004090767353773117, + 0.005901898257434368, + 0.011980232782661915, + 0.00886872410774231, + 0.008502140641212463, + 0.012048117816448212, + 0.05084656924009323, + 0.048328712582588196, + 0.00902263168245554, + 0.0016930220881476998, + 0.0019091595895588398, + 0.001969019416719675, + 0.003450453747063875, + 0.0034406257327646017, + 0.002967566018924117, + 0.003011871362105012, + 0.002704247599467635, + 0.003718713065609336, + 0.0035561274271458387, + 0.004268615040928125, + 0.004825799725949764, + 0.0040176138281822205, + 0.005956536158919334, + 0.006496412213891745, + 0.004043653141707182, + 0.009578646160662174, + 0.015529236756265163, + 0.011548174545168877, + 0.022240595892071724, + 0.017284968867897987, + 0.01904457062482834, + 0.022918961942195892, + 0.035561177879571915, + 0.003684216644614935, + 0.0045000403188169, + 0.0036713825538754463, + 0.005782124120742083, + 0.006506115663796663, + 0.00447521498426795, + 0.013437784276902676, + 0.01113069150596857, + 0.01311508659273386, + 0.014403876848518848, + 0.015335680916905403, + 0.018898312002420425, + 0.012881320901215076, + 0.013592392206192017, + 0.01783335953950882, + 0.017452536150813103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.011079339310526848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232976246625185 + ], + "opposite_logprob_tokens": [ + -3.7732975482940674 + ], + "logprob_reward": -0.0232976246625185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232976246625185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08525974303483963, + 0.05477779358625412, + 0.35509565472602844, + 0.02073882706463337, + 0.18674899637699127, + 0.11980734765529633, + 0.05981704220175743, + 0.025609983131289482, + 0.05735103413462639, + 0.034793589264154434 + ], + "credit_map": [ + 0.01855049468576908, + 0.006583696696907282, + 0.018491247668862343, + 0.00933890137821436, + 0.009090214967727661, + 0.004953053314238787, + 0.008864122442901134, + 0.009388012811541557, + 0.006574130151420832, + 0.008308894000947475, + 0.007985271513462067, + 0.007636595517396927, + 0.005211923271417618, + 0.005416868720203638, + 0.005880663171410561, + 0.007763448636978865, + 0.05890865623950958, + 0.02953392080962658, + 0.05212767794728279, + 0.052004192024469376, + 0.030241921544075012, + 0.07185930013656616, + 0.044591572135686874, + 0.015828410163521767, + 0.0020531846676021814, + 0.001638630055822432, + 0.005111711099743843, + 0.0021286311093717813, + 0.002208033809438348, + 0.00270491442643106, + 0.00268450565636158, + 0.002209216821938753, + 0.006730217952281237, + 0.007052319124341011, + 0.0069196755066514015, + 0.003788499627262354, + 0.010181103833019733, + 0.07255534827709198, + 0.06706768274307251, + 0.012454149313271046, + 0.011676306836307049, + 0.014098543673753738, + 0.00843049492686987, + 0.023989975452423096, + 0.01722666248679161, + 0.015369251370429993, + 0.01824292540550232, + 0.0107731893658638, + 0.006683753337711096, + 0.005604247562587261, + 0.010883891023695469, + 0.008157995529472828, + 0.004427701234817505, + 0.007863499224185944, + 0.007040930446237326, + 0.009155027568340302, + 0.0020434611942619085, + 0.002761820564046502, + 0.0026445258408784866, + 0.0034028859809041023, + 0.0032207064796239138, + 0.0029818539042025805, + 0.003950589802116156, + 0.004604138433933258, + 0.004773695021867752, + 0.0032506620045751333, + 0.00444795424118638, + 0.009814867749810219, + 0.004408566746860743, + 0.0073167867958545685, + 0.014546609483659267, + 0.008791890926659107, + 0.00345870410092175, + 0.00386106432415545, + 0.004984614439308643, + 0.00308642303571105, + 0.003825635649263859, + 0.005636073183268309, + 0.005456430371850729, + 0.004484645556658506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.0232976246625185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808915242552757 + ], + "opposite_logprob_tokens": [ + -3.2880890369415283 + ], + "logprob_reward": -0.03808915242552757, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808915242552757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19441986083984375, + 0.028388068079948425, + 0.047930873930454254, + 0.04445233941078186, + 0.30161169171333313, + 0.07594450563192368, + 0.07161549478769302, + 0.10849583894014359, + 0.037338607013225555, + 0.08980277180671692 + ], + "credit_map": [ + 0.022086679935455322, + 0.017822157591581345, + 0.025174135342240334, + 0.027880921959877014, + 0.04800688847899437, + 0.016700197011232376, + 0.018316496163606644, + 0.018432360142469406, + 0.0034800847060978413, + 0.0036919943522661924, + 0.004717150237411261, + 0.003869022708386183, + 0.0033219829201698303, + 0.0028021270409226418, + 0.0035401377826929092, + 0.0029655660036951303, + 0.008711909875273705, + 0.005250267684459686, + 0.007341216783970594, + 0.006238593254238367, + 0.003801661776378751, + 0.011698436923325062, + 0.0027116602286696434, + 0.0021771204192191362, + 0.004151714500039816, + 0.004340658895671368, + 0.0081280916929245, + 0.004926515277475119, + 0.005079771392047405, + 0.0054938471876084805, + 0.006473305635154247, + 0.005858427379280329, + 0.013546271249651909, + 0.01731737144291401, + 0.02000567503273487, + 0.010421307757496834, + 0.02099197544157505, + 0.10082360357046127, + 0.09630413353443146, + 0.022201351821422577, + 0.00593923032283783, + 0.010675855912268162, + 0.008892493322491646, + 0.006935940124094486, + 0.00619250675663352, + 0.016438117250800133, + 0.008585839532315731, + 0.012284510768949986, + 0.005406732205301523, + 0.007799726910889149, + 0.006150519009679556, + 0.009072449058294296, + 0.007785648573189974, + 0.014503571204841137, + 0.010300981812179089, + 0.01059584878385067, + 0.007212928030639887, + 0.010807253420352936, + 0.010868841782212257, + 0.01868526265025139, + 0.013208691962063313, + 0.01295977272093296, + 0.017406072467565536, + 0.017347006127238274, + 0.003571671899408102, + 0.002558788750320673, + 0.0037247587461024523, + 0.005492075812071562, + 0.0035334357526153326, + 0.005303053185343742, + 0.0055106389336287975, + 0.007644183933734894, + 0.01014750823378563, + 0.013241582550108433, + 0.012142670340836048, + 0.012834427878260612, + 0.012661975808441639, + 0.008170128799974918, + 0.011438962072134018, + 0.00916550774127245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.03808915242552757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld shot, simulating the effects of night vision goggles, bathed in a grainy, monochromatic phosphorescent green shimmering with digital noise, slowly and steadily follows a towering skeletal figure. The figure is draped in a tattered, dark cloak that sways subtly with its movements, its weathered bones gleaming palely in the eerie green light. One of its gnarled hands grips a massive sickle. The wooden handle is ancient and curved, and the long, curved blade, rusted and chipped, reflects a harsh, bright green streak in the night vision lens's enhanced effect. The figure moves with a ghostly, leisurely pace through a long-abandoned cemetery, its skull turning slightly from side to side, its hollow eye sockets seeming to survey its surroundings. In the background, overgrown and crooked, ancient tombstones appear blurred and distorted in the green glow, their outlines draped in the dark silhouettes of ivy. The ground is covered with thick dead leaves, and each step of the skeleton is silent, without disturbing the dead silence of the land, which together creates an eerie sense of surveillance and supernatural tension, as if the audience is peeking at this ominous exploration process through the first-person perspective of some inhuman device.", + "seed": 9342, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_14/index_7243_seed_9342_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07639585379511118, + "vlm_logprob": -0.07639585379511118, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.07639585379511118, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2977961003780365 + ], + "opposite_logprob_tokens": [ + -1.9227961301803589 + ], + "logprob_reward": -0.2977961003780365, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.2977961003780365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23977822065353394, + 0.18486014008522034, + 0.08052268624305725, + 0.03253810852766037, + 0.045644596219062805, + 0.02322099544107914, + 0.07108800858259201, + 0.04269414022564888, + 0.034184593707323074, + 0.24546854197978973 + ], + "credit_map": [ + 0.02501717209815979, + 0.021923094987869263, + 0.029450522735714912, + 0.02991868369281292, + 0.03605874627828598, + 0.044614169746637344, + 0.021194905042648315, + 0.031600914895534515, + 0.011129051446914673, + 0.034960027784109116, + 0.0171921756118536, + 0.022978821769356728, + 0.018488124012947083, + 0.025060541927814484, + 0.023442154750227928, + 0.031609226018190384, + 0.005274227354675531, + 0.012967095710337162, + 0.006392854731529951, + 0.010279195383191109, + 0.013309817761182785, + 0.009979067370295525, + 0.013584332540631294, + 0.00873609259724617, + 0.0028960530180484056, + 0.003640941344201565, + 0.004313007462769747, + 0.003899442031979561, + 0.00401294557377696, + 0.005252161528915167, + 0.004325609188526869, + 0.004197946283966303, + 0.005525345914065838, + 0.004974304232746363, + 0.006397169083356857, + 0.006640678737312555, + 0.00406021811068058, + 0.005138280335813761, + 0.006273963022977114, + 0.006634634453803301, + 0.003489779308438301, + 0.002263634465634823, + 0.003080110065639019, + 0.0038684438914060593, + 0.0029576225206255913, + 0.0023941295221447945, + 0.0029360754415392876, + 0.0022311971988528967, + 0.004393781069666147, + 0.007104525342583656, + 0.007181041873991489, + 0.010084453038871288, + 0.004229514393955469, + 0.013327782042324543, + 0.008403023704886436, + 0.01636388525366783, + 0.007661600597202778, + 0.00433924188837409, + 0.005403675604611635, + 0.006341438740491867, + 0.003519364632666111, + 0.00472219567745924, + 0.004920479375869036, + 0.005786140915006399, + 0.002725068014115095, + 0.004694482311606407, + 0.0032678847201168537, + 0.004354449920356274, + 0.0034293450880795717, + 0.00456570740789175, + 0.00679330388084054, + 0.0043543498031795025, + 0.02850523591041565, + 0.014497324824333191, + 0.010960495099425316, + 0.019301950931549072, + 0.01046695001423359, + 0.07911615073680878, + 0.008340214379131794, + 0.07428020983934402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.2977961003780365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002889032242819667 + ], + "opposite_logprob_tokens": [ + -5.877889156341553 + ], + "logprob_reward": -0.002889032242819667, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002889032242819667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2643948197364807, + 0.04770813509821892, + 0.07582852244377136, + 0.03242088109254837, + 0.13364069163799286, + 0.029912428930401802, + 0.06936689466238022, + 0.1653883457183838, + 0.04984357953071594, + 0.13149571418762207 + ], + "credit_map": [ + 0.024020761251449585, + 0.02610708214342594, + 0.03392919898033142, + 0.03178056329488754, + 0.03970549628138542, + 0.04696618393063545, + 0.026664022356271744, + 0.035221513360738754, + 0.0029353778809309006, + 0.006106344051659107, + 0.004856986925005913, + 0.005391203798353672, + 0.006615378428250551, + 0.0072457389906048775, + 0.006319274660199881, + 0.008237832225859165, + 0.00496879406273365, + 0.007273143623024225, + 0.0071965972892940044, + 0.013509265147149563, + 0.007145033683627844, + 0.009012900292873383, + 0.010017693042755127, + 0.016705099493265152, + 0.0024937475100159645, + 0.0037524159997701645, + 0.005083905067294836, + 0.003488207468762994, + 0.0032019931823015213, + 0.005449493415653706, + 0.004396209958940744, + 0.004554906394332647, + 0.01692970097064972, + 0.012652824632823467, + 0.01593295857310295, + 0.015549723990261555, + 0.02006465382874012, + 0.018893223255872726, + 0.016961712390184402, + 0.01665588654577732, + 0.0024428789038211107, + 0.003165606176480651, + 0.004075467586517334, + 0.004932017996907234, + 0.003645760240033269, + 0.004980995785444975, + 0.0035586394369602203, + 0.0031110618729144335, + 0.00439416104927659, + 0.00744674215093255, + 0.010115143842995167, + 0.011571589857339859, + 0.0063191428780555725, + 0.012174656614661217, + 0.009065424092113972, + 0.008280034177005291, + 0.03187771514058113, + 0.012620167806744576, + 0.022290481254458427, + 0.024534717202186584, + 0.009322610683739185, + 0.022556288167834282, + 0.01813664101064205, + 0.024049725383520126, + 0.0042883241549134254, + 0.005391799844801426, + 0.005403085146099329, + 0.005959098227322102, + 0.004495819564908743, + 0.007454480975866318, + 0.010671395808458328, + 0.006179573480039835, + 0.013310744427144527, + 0.011397900059819221, + 0.01027466356754303, + 0.015568659640848637, + 0.011726072058081627, + 0.03151930123567581, + 0.007738075684756041, + 0.02996028959751129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.002889032242819667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028325694147497416 + ], + "opposite_logprob_tokens": [ + -5.877832412719727 + ], + "logprob_reward": -0.0028325694147497416, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028325694147497416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12216033041477203, + 0.13350588083267212, + 0.18831069767475128, + 0.036839891225099564, + 0.08032596111297607, + 0.01711898297071457, + 0.06566506624221802, + 0.07256033271551132, + 0.054420437663793564, + 0.22909235954284668 + ], + "credit_map": [ + 0.011305629275739193, + 0.010347675532102585, + 0.0159157607704401, + 0.011357882060110569, + 0.015960698947310448, + 0.025226658210158348, + 0.014729896560311317, + 0.01731613092124462, + 0.009202859364449978, + 0.020401325076818466, + 0.019337039440870285, + 0.015781886875629425, + 0.01730269193649292, + 0.016300169751048088, + 0.014906547963619232, + 0.020273355767130852, + 0.010218367911875248, + 0.025541365146636963, + 0.01211625523865223, + 0.028379499912261963, + 0.03373545780777931, + 0.018548300489783287, + 0.030254341661930084, + 0.02951711229979992, + 0.0035284627228975296, + 0.0039029184263199568, + 0.004640912637114525, + 0.003594891866669059, + 0.003948119934648275, + 0.007376838009804487, + 0.005048193968832493, + 0.004799554590135813, + 0.010967566631734371, + 0.009139934554696083, + 0.011445268988609314, + 0.014195944182574749, + 0.0072729382663965225, + 0.008116285316646099, + 0.009300696663558483, + 0.009887320920825005, + 0.0017516672378405929, + 0.0017829617718234658, + 0.0021682961378246546, + 0.002946613123640418, + 0.0017922059632837772, + 0.0028728835750371218, + 0.002159346593543887, + 0.0016450078692287207, + 0.004471665248274803, + 0.006291079800575972, + 0.006343676242977381, + 0.009234601631760597, + 0.0038345481734722853, + 0.013723411597311497, + 0.006449927110224962, + 0.015316161327064037, + 0.010585642419755459, + 0.007206988986581564, + 0.010049251839518547, + 0.011996809393167496, + 0.0060023898258805275, + 0.010257020592689514, + 0.009644580073654652, + 0.006817654240876436, + 0.004406766500324011, + 0.006372555159032345, + 0.007512811571359634, + 0.006416115909814835, + 0.006056470330804586, + 0.006959541700780392, + 0.007947240956127644, + 0.008748934604227543, + 0.02637990564107895, + 0.015958309173583984, + 0.015322350896894932, + 0.02063334360718727, + 0.009319509379565716, + 0.06484759598970413, + 0.009455611929297447, + 0.06717574596405029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.0028325694147497416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048665907233953476 + ], + "opposite_logprob_tokens": [ + -3.048666000366211 + ], + "logprob_reward": -0.048665907233953476, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048665907233953476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1742352694272995, + 0.28937992453575134, + 0.09398052096366882, + 0.06298695504665375, + 0.03750348836183548, + 0.02060505375266075, + 0.06934869289398193, + 0.02302977629005909, + 0.02547570690512657, + 0.20345459878444672 + ], + "credit_map": [ + 0.01417829655110836, + 0.013160981237888336, + 0.022196391597390175, + 0.015129218809306622, + 0.023681173101067543, + 0.04212377592921257, + 0.016365643590688705, + 0.02739979512989521, + 0.018050402402877808, + 0.023498112335801125, + 0.043813224881887436, + 0.03123852051794529, + 0.02046912908554077, + 0.03876062482595444, + 0.027074409648776054, + 0.08647550642490387, + 0.006360418163239956, + 0.012286503799259663, + 0.008145886473357677, + 0.015141054056584835, + 0.013231066055595875, + 0.009858733974397182, + 0.011224852874875069, + 0.017732005566358566, + 0.007008465938270092, + 0.00651529198512435, + 0.00790776964277029, + 0.006590493489056826, + 0.008203658275306225, + 0.011593637056648731, + 0.008687003515660763, + 0.006480633281171322, + 0.003980990499258041, + 0.0033770331647247076, + 0.004182971548289061, + 0.005311459302902222, + 0.004830796737223864, + 0.005636095535010099, + 0.004086657892912626, + 0.0060974834486842155, + 0.0020699703600257635, + 0.0029119602404534817, + 0.0026496865320950747, + 0.003630273509770632, + 0.00235320464707911, + 0.001960130175575614, + 0.002773409243673086, + 0.0022564197424799204, + 0.003994396422058344, + 0.007572019472718239, + 0.008944690227508545, + 0.011709445156157017, + 0.005042781122028828, + 0.009994723834097385, + 0.009479764848947525, + 0.012610876001417637, + 0.0021530857775360346, + 0.003138653701171279, + 0.0028547020629048347, + 0.004751731175929308, + 0.0017766812816262245, + 0.0032694414258003235, + 0.0027400844264775515, + 0.0023453948087990284, + 0.0023740981705486774, + 0.0031247949227690697, + 0.002432058332487941, + 0.0032642704900354147, + 0.003867155872285366, + 0.003390694037079811, + 0.0034196386113762856, + 0.0036029964685440063, + 0.02133258432149887, + 0.013686146587133408, + 0.013721871189773083, + 0.016427386552095413, + 0.011264710687100887, + 0.06041089445352554, + 0.008579056710004807, + 0.05803194269537926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.048665907233953476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video style simulated as night vision goggles with a grainy, monochromatic phosphorescent green look?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering skeletal figure draped in a tattered dark cloak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a long-abandoned cemetery with overgrown, crooked ancient tombstones in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily follow the skeletal figure as it moves through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795659705996513 + ], + "opposite_logprob_tokens": [ + -3.5297956466674805 + ], + "logprob_reward": -0.029795659705996513, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795659705996513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27663713693618774, + 0.10355006903409958, + 0.06325407326221466, + 0.043448373675346375, + 0.0608421266078949, + 0.05336370691657066, + 0.15017645061016083, + 0.03543480858206749, + 0.034633222967386246, + 0.17866000533103943 + ], + "credit_map": [ + 0.02689940109848976, + 0.020398087799549103, + 0.035091087222099304, + 0.027160542085766792, + 0.03947579115629196, + 0.05569205433130264, + 0.03379497304558754, + 0.03812519833445549, + 0.00754530867561698, + 0.016138026490807533, + 0.017425160855054855, + 0.011031512171030045, + 0.014369943179190159, + 0.015402814373373985, + 0.009857065044343472, + 0.011780232191085815, + 0.003931018989533186, + 0.008057717233896255, + 0.0057519166730344296, + 0.00879475474357605, + 0.010121691040694714, + 0.009284991770982742, + 0.010397354140877724, + 0.0069146305322647095, + 0.004096303600817919, + 0.004390000831335783, + 0.006310004275292158, + 0.005315955262631178, + 0.0038022322114557028, + 0.006505269557237625, + 0.007340846583247185, + 0.005687758792191744, + 0.007614362984895706, + 0.007046673446893692, + 0.006882059387862682, + 0.0071636768989264965, + 0.006860310211777687, + 0.007700417190790176, + 0.009777369908988476, + 0.00779725331813097, + 0.006034270394593477, + 0.004850677214562893, + 0.007301750592887402, + 0.00944574549794197, + 0.006132174748927355, + 0.006640193983912468, + 0.007969100028276443, + 0.004989793058484793, + 0.009515529498457909, + 0.014619979076087475, + 0.01588950864970684, + 0.02008056826889515, + 0.009791875258088112, + 0.03133172169327736, + 0.015652284026145935, + 0.03329499438405037, + 0.004580759443342686, + 0.00426317285746336, + 0.004507483448833227, + 0.0066591533832252026, + 0.003511207178235054, + 0.0048922146670520306, + 0.0034195110201835632, + 0.0036013065837323666, + 0.003080896334722638, + 0.004217085428535938, + 0.005542477127164602, + 0.004350648261606693, + 0.004071160219609737, + 0.004536299034953117, + 0.003823915496468544, + 0.005010740831494331, + 0.021652625873684883, + 0.014566447585821152, + 0.010296435095369816, + 0.017979303374886513, + 0.00852205790579319, + 0.047335173934698105, + 0.008939199149608612, + 0.049368780106306076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.029795659705996513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_15/index_tensor([9303, 9303])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_15/index_tensor([9303, 9303])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2f0d1370f9765b3d385ebd82257d3539a3d6d1a4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_15/index_tensor([9303, 9303])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in a vibrant, illustrative style reminiscent of classic animated films, a low-angle, wide-angle shot begins in a slow, majestic rotation, capturing an impossible spectacle. Across a vast, sun-drenched valley, a massive, crumbling castle begins its mighty takeoff. The structure is a chaotic marvel of mismatched stone towers, bent timber extensions, and bolted iron plates, all perched precariously on four massive, piston-driven mechanical legs. As the legs begin to tumble, their heavy, rhythmic steps kicking up clods of earth and grass, the entire fortress pitches forward, gaining momentum with a startling shudder. Thick black smoke and billowing white steam billow violently from a dozen chimneys of various shapes and sizes, staining the bright blue sky red. The camera slowly rotates, following the castle's accelerating path, emphasizing its immense scale against the painterly landscape, culminating in the moment when orange and yellow flames burst from hidden vents in the castle's underbelly, beginning to lift this impossible machine into the air in a breathtaking, dreamlike ascent.", + "seed": 16773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_15/index_9303_seed_16773_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0414887766353786, + "vlm_logprob": -0.0414887766353786, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0414887766353786, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005561118945479393 + ], + "opposite_logprob_tokens": [ + -5.380561351776123 + ], + "logprob_reward": -0.005561118945479393, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005561118945479393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3068864345550537, + 0.03837861493229866, + 0.0762149915099144, + 0.21918077766895294, + 0.18111176788806915, + 0.03640483692288399, + 0.022737160325050354, + 0.03309072554111481, + 0.07017838954925537, + 0.015816330909729004 + ], + "credit_map": [ + 0.10408338904380798, + 0.01191117987036705, + 0.006190429907292128, + 0.00791249144822359, + 0.015889331698417664, + 0.12821975350379944, + 0.015001809224486351, + 0.01767805777490139, + 0.0044309706427156925, + 0.006818091496825218, + 0.004816668573766947, + 0.004339489154517651, + 0.004824915900826454, + 0.00364616047590971, + 0.004417129792273045, + 0.005085188429802656, + 0.013623709790408611, + 0.00816015899181366, + 0.0096835782751441, + 0.007329067215323448, + 0.00786788109689951, + 0.008161920122802258, + 0.00541276391595602, + 0.015975911170244217, + 0.052857447415590286, + 0.018404651433229446, + 0.046644337475299835, + 0.009769503027200699, + 0.008382937870919704, + 0.017250142991542816, + 0.008959293365478516, + 0.05691245198249817, + 0.016481371596455574, + 0.016337405890226364, + 0.02479819767177105, + 0.011694534681737423, + 0.020232221111655235, + 0.03984222561120987, + 0.01986725814640522, + 0.03185855224728584, + 0.004367338493466377, + 0.0032610169146209955, + 0.006039200816303492, + 0.005374642554670572, + 0.0030887178145349026, + 0.005822301376610994, + 0.003140568733215332, + 0.005311049986630678, + 0.0029645583126693964, + 0.002846697112545371, + 0.0025979983620345592, + 0.0032036942429840565, + 0.002482673153281212, + 0.00328294700011611, + 0.00253013102337718, + 0.002828462515026331, + 0.004187571816146374, + 0.003687481163069606, + 0.01178785227239132, + 0.0029049981385469437, + 0.001880490337498486, + 0.0027821611147373915, + 0.0017073818016797304, + 0.0041527897119522095, + 0.010812226682901382, + 0.01895911432802677, + 0.007577324751764536, + 0.006133034359663725, + 0.008519149385392666, + 0.0043709417805075645, + 0.006386853288859129, + 0.00741974450647831, + 0.0020509823225438595, + 0.0022147586569190025, + 0.0022077339235693216, + 0.0018772073090076447, + 0.002040248829871416, + 0.0019572710152715445, + 0.0018410332268103957, + 0.0016270955093204975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.005561118945479393 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595386624336243 + ], + "opposite_logprob_tokens": [ + -3.8955953121185303 + ], + "logprob_reward": -0.020595386624336243, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595386624336243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23732541501522064, + 0.1018531322479248, + 0.034162454307079315, + 0.2664669454097748, + 0.13378994166851044, + 0.06515222787857056, + 0.02470513805747032, + 0.03552308678627014, + 0.05157747119665146, + 0.049444250762462616 + ], + "credit_map": [ + 0.07786233723163605, + 0.00838178489357233, + 0.005625725258141756, + 0.00844309851527214, + 0.013278253376483917, + 0.09700166434049606, + 0.011888496577739716, + 0.014844023622572422, + 0.01195360068231821, + 0.012382863089442253, + 0.010985576547682285, + 0.009764885529875755, + 0.014240500517189503, + 0.012149805203080177, + 0.017259642481803894, + 0.013116239570081234, + 0.005998485255986452, + 0.0025027887895703316, + 0.004823361523449421, + 0.0031521532218903303, + 0.0030229121912270784, + 0.0034743864089250565, + 0.0029418091289699078, + 0.00824655406177044, + 0.06181718409061432, + 0.02201450988650322, + 0.05779295414686203, + 0.011358696036040783, + 0.010191318579018116, + 0.02141992747783661, + 0.010703351348638535, + 0.07116895914077759, + 0.010873115621507168, + 0.011965451762080193, + 0.018748212605714798, + 0.008384141139686108, + 0.014809574000537395, + 0.02983148582279682, + 0.01343788392841816, + 0.025740066543221474, + 0.004967667628079653, + 0.002426838269457221, + 0.02438235469162464, + 0.003703189315274358, + 0.0026626617182046175, + 0.0036299354396760464, + 0.0033972759265452623, + 0.019982295110821724, + 0.0033802613615989685, + 0.002642794279381633, + 0.003532374044880271, + 0.003934576641768217, + 0.0028031894471496344, + 0.002482695272192359, + 0.0028503495268523693, + 0.0030788923613727093, + 0.005531616043299437, + 0.005709538236260414, + 0.006949699018150568, + 0.004282054957002401, + 0.0033089264761656523, + 0.004642005078494549, + 0.0026700557209551334, + 0.002429185202345252, + 0.007574633229523897, + 0.010440962389111519, + 0.008850912563502789, + 0.005007033236324787, + 0.0060086920857429504, + 0.004499868955463171, + 0.004038053099066019, + 0.005157307256013155, + 0.007212169002741575, + 0.007158538326621056, + 0.007602539844810963, + 0.006046314258128405, + 0.0060132406651973724, + 0.005597067065536976, + 0.005995185114443302, + 0.0038191908970475197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.020595386624336243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737255025655031 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737255025655031, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737255025655031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07119502872228622, + 0.036621399223804474, + 0.07088746130466461, + 0.10365612059831619, + 0.36244088411331177, + 0.13639825582504272, + 0.04226095229387283, + 0.04875588044524193, + 0.09820207953453064, + 0.02958195097744465 + ], + "credit_map": [ + 0.012501423247158527, + 0.007205386646091938, + 0.00446641119197011, + 0.006877280306071043, + 0.010163783095777035, + 0.015138628892600536, + 0.010346712544560432, + 0.004495406057685614, + 0.0047038947232067585, + 0.005500713363289833, + 0.00461411802098155, + 0.005357656162232161, + 0.003448318922892213, + 0.004224539268761873, + 0.005025362595915794, + 0.0037467952352017164, + 0.013614085502922535, + 0.007273179944604635, + 0.009458672255277634, + 0.006154583767056465, + 0.006112146191298962, + 0.007673666812479496, + 0.005955861881375313, + 0.014645268209278584, + 0.01896088197827339, + 0.011369680054485798, + 0.020255768671631813, + 0.00534950103610754, + 0.006002511829137802, + 0.0069503504782915115, + 0.0085745332762599, + 0.026192892342805862, + 0.030541032552719116, + 0.03382906690239906, + 0.05043428763747215, + 0.0221552774310112, + 0.040686238557100296, + 0.083140067756176, + 0.036806341260671616, + 0.06484857201576233, + 0.011604001745581627, + 0.003593783127143979, + 0.053504034876823425, + 0.0066354600712656975, + 0.0029182436410337687, + 0.007785467896610498, + 0.006801184732466936, + 0.043556082993745804, + 0.00495493970811367, + 0.004901495296508074, + 0.006711256690323353, + 0.005989734083414078, + 0.005166513845324516, + 0.00429111672565341, + 0.005509437993168831, + 0.004736457020044327, + 0.006787044461816549, + 0.007871003821492195, + 0.008889964781701565, + 0.007440435234457254, + 0.004887916147708893, + 0.005405815783888102, + 0.0038698974531143904, + 0.003603802528232336, + 0.016314806416630745, + 0.02969645895063877, + 0.011416321620345116, + 0.008445695042610168, + 0.010498344898223877, + 0.006422387436032295, + 0.006923085078597069, + 0.008484983816742897, + 0.005591028835624456, + 0.004383424296975136, + 0.004881720524281263, + 0.003530902089551091, + 0.0029897582717239857, + 0.002499023685231805, + 0.003603767603635788, + 0.002102324040606618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.006737255025655031 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035164892673492 + ], + "opposite_logprob_tokens": [ + -1.9103516340255737 + ], + "logprob_reward": -0.16035164892673492, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035164892673492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15446124970912933, + 0.018341543152928352, + 0.03511255979537964, + 0.3645372986793518, + 0.25577202439308167, + 0.07932574301958084, + 0.012062394991517067, + 0.04033033549785614, + 0.027643783017992973, + 0.012413029558956623 + ], + "credit_map": [ + 0.05161181464791298, + 0.005937208421528339, + 0.003661801340058446, + 0.0035935293417423964, + 0.008977903984487057, + 0.06289597600698471, + 0.008704033680260181, + 0.009078997187316418, + 0.0027281553484499454, + 0.002113868948072195, + 0.0026141779962927103, + 0.001858889008872211, + 0.0015051927184686065, + 0.0030777831561863422, + 0.00270254397764802, + 0.0017409335123375058, + 0.006020719651132822, + 0.004074574448168278, + 0.004612020682543516, + 0.002645273692905903, + 0.0035997198428958654, + 0.003779701190069318, + 0.0032646646723151207, + 0.0071158865466713905, + 0.08769471198320389, + 0.03251507878303528, + 0.07999564707279205, + 0.014925297349691391, + 0.009752804413437843, + 0.026597794145345688, + 0.013267144560813904, + 0.09978881478309631, + 0.020765867084264755, + 0.02143058367073536, + 0.036489088088274, + 0.018987998366355896, + 0.025315186008810997, + 0.04908021166920662, + 0.031004605814814568, + 0.05269847437739372, + 0.007095497567206621, + 0.0034969844855368137, + 0.028133844956755638, + 0.004843715112656355, + 0.0032281551975756884, + 0.005581689532846212, + 0.004114989191293716, + 0.022830866277217865, + 0.0011193351820111275, + 0.0013965658145025373, + 0.001424345769919455, + 0.0015944488113746047, + 0.001378720044158399, + 0.0023054287303239107, + 0.0013863140484318137, + 0.0014572370564565063, + 0.0055959345772862434, + 0.005884943529963493, + 0.010258663445711136, + 0.004066755995154381, + 0.0032639717683196068, + 0.004149795044213533, + 0.003331315703690052, + 0.0037789572961628437, + 0.004853154998272657, + 0.005862025078386068, + 0.004603949375450611, + 0.0029731784015893936, + 0.0025666775181889534, + 0.002030354691669345, + 0.0029978889506310225, + 0.0017565531888976693, + 0.0017842313973233104, + 0.0016196263022720814, + 0.0020650289952754974, + 0.0018110317178070545, + 0.001440295483916998, + 0.0012859406415373087, + 0.001502330880612135, + 0.0009045437909662724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.16035164892673492 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198473654687405 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198473654687405, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198473654687405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06009393185377121, + 0.017194224521517754, + 0.034600015729665756, + 0.4349548816680908, + 0.13739581406116486, + 0.06618474423885345, + 0.06921274960041046, + 0.06567445397377014, + 0.07256793230772018, + 0.04212125763297081 + ], + "credit_map": [ + 0.0125607093796134, + 0.004568751435726881, + 0.0026579760015010834, + 0.00433883024379611, + 0.006842200644314289, + 0.016456183046102524, + 0.008219154551625252, + 0.004450129345059395, + 0.0024055116809904575, + 0.0018783231498673558, + 0.002543570240959525, + 0.0018955738050863147, + 0.0014135888777673244, + 0.0027961255982518196, + 0.002159039955586195, + 0.002102492842823267, + 0.006638416089117527, + 0.003317156806588173, + 0.004823828116059303, + 0.002628349233418703, + 0.003824752988293767, + 0.0036769930738955736, + 0.002173809800297022, + 0.007516707759350538, + 0.10516159981489182, + 0.03947371989488602, + 0.09514473378658295, + 0.01777835562825203, + 0.011451025493443012, + 0.031583450734615326, + 0.016129598021507263, + 0.11823242157697678, + 0.012642196379601955, + 0.012698108330368996, + 0.019327402114868164, + 0.009871350601315498, + 0.013410381972789764, + 0.029308022931218147, + 0.015575754456222057, + 0.024562586098909378, + 0.005798600614070892, + 0.0031535690650343895, + 0.022198859602212906, + 0.005425900220870972, + 0.0033667166717350483, + 0.004022062756121159, + 0.004316241014748812, + 0.017902789637446404, + 0.008287225849926472, + 0.008838098496198654, + 0.010103803128004074, + 0.010835264809429646, + 0.008414983749389648, + 0.007917565293610096, + 0.007267679553478956, + 0.007548123132437468, + 0.00846025813370943, + 0.01021366287022829, + 0.008492941036820412, + 0.010304628871381283, + 0.007836868055164814, + 0.010154709219932556, + 0.006327906157821417, + 0.003883481491357088, + 0.010576645843684673, + 0.01682179421186447, + 0.011316710151731968, + 0.007501829881221056, + 0.0075866724364459515, + 0.0065101212821900845, + 0.006182752083986998, + 0.006071402691304684, + 0.006236627697944641, + 0.00726467277854681, + 0.006628323812037706, + 0.005479531828314066, + 0.004640810191631317, + 0.003662344766780734, + 0.004703599028289318, + 0.003505346132442355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.014198473654687405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a vibrant, illustrative style reminiscent of classic animated films, a low-angle, wide-angle shot begins in a slow, majestic rotation, capturing an impossible spectacle. Across a vast, sun-drenched valley, a massive, crumbling castle begins its mighty takeoff. The structure is a chaotic marvel of mismatched stone towers, bent timber extensions, and bolted iron plates, all perched precariously on four massive, piston-driven mechanical legs. As the legs begin to tumble, their heavy, rhythmic steps kicking up clods of earth and grass, the entire fortress pitches forward, gaining momentum with a startling shudder. Thick black smoke and billowing white steam billow violently from a dozen chimneys of various shapes and sizes, staining the bright blue sky red. The camera slowly rotates, following the castle's accelerating path, emphasizing its immense scale against the painterly landscape, culminating in the moment when orange and yellow flames burst from hidden vents in the castle's underbelly, beginning to lift this impossible machine into the air in a breathtaking, dreamlike ascent.", + "seed": 16774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_15/index_9303_seed_16774_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02352063972502947, + "vlm_logprob": -0.02352063972502947, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.02352063972502947, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007134201005101204 + ], + "opposite_logprob_tokens": [ + -5.132134437561035 + ], + "logprob_reward": -0.007134201005101204, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007134201005101204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12589311599731445, + 0.08211137354373932, + 0.19457943737506866, + 0.09254889190196991, + 0.1028246358036995, + 0.13516736030578613, + 0.10254132002592087, + 0.05884614959359169, + 0.0792142003774643, + 0.02627347595989704 + ], + "credit_map": [ + 0.012144461274147034, + 0.011597339063882828, + 0.016936609521508217, + 0.01516315620392561, + 0.015499651432037354, + 0.018649699166417122, + 0.01705654338002205, + 0.01884564757347107, + 0.009033228270709515, + 0.005616813898086548, + 0.007873333059251308, + 0.007467771414667368, + 0.009486737661063671, + 0.021349478513002396, + 0.011429469101130962, + 0.009854541160166264, + 0.011906263418495655, + 0.01833915337920189, + 0.01459950115531683, + 0.01909721828997135, + 0.02024238184094429, + 0.04631457477807999, + 0.01715848222374916, + 0.04692186042666435, + 0.006930596195161343, + 0.011906501837074757, + 0.012943235225975513, + 0.009986511431634426, + 0.011650619097054005, + 0.008964268490672112, + 0.012044562958180904, + 0.01812259666621685, + 0.009539206512272358, + 0.013035405427217484, + 0.012073526158928871, + 0.0253764558583498, + 0.009403828531503677, + 0.007141876500099897, + 0.01606648415327072, + 0.010187852196395397, + 0.021380141377449036, + 0.015444920398294926, + 0.012921019457280636, + 0.019408835098147392, + 0.012302846647799015, + 0.021153824403882027, + 0.01542426273226738, + 0.017131509259343147, + 0.014892245642840862, + 0.01336879376322031, + 0.019969861954450607, + 0.009115876629948616, + 0.00799735076725483, + 0.013023465871810913, + 0.008090184070169926, + 0.01608354225754738, + 0.009391659870743752, + 0.006181240081787109, + 0.008072287775576115, + 0.006884405855089426, + 0.005461953580379486, + 0.008669551461935043, + 0.008190467022359371, + 0.005994588602334261, + 0.009572042152285576, + 0.016589447855949402, + 0.008853155188262463, + 0.009153101593255997, + 0.007074321154505014, + 0.008790937252342701, + 0.01008082740008831, + 0.009100363589823246, + 0.0036581968888640404, + 0.004430369473993778, + 0.0032434826716780663, + 0.003657373134046793, + 0.002798144007101655, + 0.002458715345710516, + 0.003475705161690712, + 0.0025514899753034115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.007134201005101204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054966483265161514 + ], + "opposite_logprob_tokens": [ + -2.929966449737549 + ], + "logprob_reward": -0.054966483265161514, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054966483265161514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08021585643291473, + 0.05386938899755478, + 0.2125079482793808, + 0.06741566956043243, + 0.18030184507369995, + 0.07391553372144699, + 0.03543359413743019, + 0.12538142502307892, + 0.1000867486000061, + 0.0708719864487648 + ], + "credit_map": [ + 0.008812000043690205, + 0.007684601936489344, + 0.01222517341375351, + 0.011643100529909134, + 0.0068790786899626255, + 0.012164894491434097, + 0.010325446724891663, + 0.010481569916009903, + 0.005820844322443008, + 0.0035371517296880484, + 0.004698195960372686, + 0.004742425866425037, + 0.006624071858823299, + 0.015362165868282318, + 0.0060930876061320305, + 0.006991446018218994, + 0.017391882836818695, + 0.023973209783434868, + 0.018302299082279205, + 0.02203923836350441, + 0.020418059080839157, + 0.043983351439237595, + 0.020895984023809433, + 0.04550394415855408, + 0.005014299415051937, + 0.009633992798626423, + 0.008151385933160782, + 0.008533450774848461, + 0.00887013878673315, + 0.004719612188637257, + 0.008907548151910305, + 0.013585246168076992, + 0.01869826763868332, + 0.025274096056818962, + 0.01791035756468773, + 0.038823969662189484, + 0.023868871852755547, + 0.012121258303523064, + 0.025475652888417244, + 0.018129386007785797, + 0.010289745405316353, + 0.009584654122591019, + 0.007853364571928978, + 0.01154082827270031, + 0.008820694871246815, + 0.008213571272790432, + 0.009123081341385841, + 0.008489594794809818, + 0.005444511771202087, + 0.004935176111757755, + 0.005593802314251661, + 0.003715796861797571, + 0.002912112744525075, + 0.0038523224648088217, + 0.003466840600594878, + 0.005513033829629421, + 0.020593497902154922, + 0.020110242068767548, + 0.021562915295362473, + 0.0139844361692667, + 0.009498847648501396, + 0.011816082522273064, + 0.015739137306809425, + 0.01207626610994339, + 0.010498862713575363, + 0.02337748184800148, + 0.010910176672041416, + 0.0112558389082551, + 0.010759064927697182, + 0.00920944008976221, + 0.013658740557730198, + 0.010417154990136623, + 0.009283001534640789, + 0.009847589768469334, + 0.008384456858038902, + 0.008428793400526047, + 0.009253207594156265, + 0.009929596446454525, + 0.00875747762620449, + 0.006987868342548609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.054966483265161514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778338484466076 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778338484466076, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778338484466076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12511098384857178, + 0.11879819631576538, + 0.1109296977519989, + 0.050447892397642136, + 0.10889140516519547, + 0.075734943151474, + 0.2588668465614319, + 0.055804021656513214, + 0.06760852783918381, + 0.02780749276280403 + ], + "credit_map": [ + 0.012453235685825348, + 0.012866283766925335, + 0.015816327184438705, + 0.016075938940048218, + 0.012240112759172916, + 0.020979249849915504, + 0.01662628725171089, + 0.018053550273180008, + 0.013004502281546593, + 0.0092098293825984, + 0.013865521177649498, + 0.01288151927292347, + 0.015263578854501247, + 0.024022208526730537, + 0.01681477017700672, + 0.01373626571148634, + 0.0062238057143986225, + 0.00920129381120205, + 0.007003100123256445, + 0.0075618247501552105, + 0.008262388408184052, + 0.02876303903758526, + 0.012469376437366009, + 0.03144487738609314, + 0.005189945921301842, + 0.0058147055096924305, + 0.006057649850845337, + 0.007425141986459494, + 0.007585533894598484, + 0.004307153634727001, + 0.005554608069360256, + 0.008513147942721844, + 0.012534239329397678, + 0.013051746413111687, + 0.013207045383751392, + 0.02391110546886921, + 0.010868124663829803, + 0.010483851656317711, + 0.013209578581154346, + 0.011625716462731361, + 0.012147733010351658, + 0.007698446977883577, + 0.007756441831588745, + 0.011754931882023811, + 0.009000349789857864, + 0.008276465348899364, + 0.009025515988469124, + 0.010075061582028866, + 0.0443718321621418, + 0.030555497854948044, + 0.04419415444135666, + 0.022041838616132736, + 0.02254491113126278, + 0.027153022587299347, + 0.022862344980239868, + 0.0451432503759861, + 0.009334711357951164, + 0.009173691272735596, + 0.008903556503355503, + 0.007351670879870653, + 0.004282080568373203, + 0.00507926382124424, + 0.00668357266113162, + 0.004995475523173809, + 0.007547877728939056, + 0.013997344300150871, + 0.007793081924319267, + 0.007744540460407734, + 0.005810144357383251, + 0.0072655524127185345, + 0.009927711449563503, + 0.007522272411733866, + 0.0041570295579731464, + 0.0044392128475010395, + 0.003115682629868388, + 0.0025239530950784683, + 0.002716611372306943, + 0.0033959588035941124, + 0.0034999498166143894, + 0.003959096036851406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.009778338484466076 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809546306729317 + ], + "opposite_logprob_tokens": [ + -3.288095474243164 + ], + "logprob_reward": -0.03809546306729317, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809546306729317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05760527402162552, + 0.042313579469919205, + 0.15972772240638733, + 0.06841534376144409, + 0.19614994525909424, + 0.1491445004940033, + 0.04099313169717789, + 0.06196016073226929, + 0.10872175544500351, + 0.11496848613023758 + ], + "credit_map": [ + 0.00940107274800539, + 0.007217749021947384, + 0.007749478332698345, + 0.0060089752078056335, + 0.0065247961319983006, + 0.007594431284815073, + 0.00681519228965044, + 0.006293579004704952, + 0.004731674212962389, + 0.003130301833152771, + 0.005381963215768337, + 0.004479415249079466, + 0.005414724815636873, + 0.007477531675249338, + 0.006139073520898819, + 0.005558894947171211, + 0.015745118260383606, + 0.01249755173921585, + 0.014009439386427402, + 0.01322348415851593, + 0.0180466677993536, + 0.038142766803503036, + 0.010506942868232727, + 0.0375557579100132, + 0.006511129438877106, + 0.007269139401614666, + 0.013721020892262459, + 0.010669197887182236, + 0.005962872877717018, + 0.00514590460807085, + 0.008138914592564106, + 0.010997166857123375, + 0.015259800478816032, + 0.02666585147380829, + 0.023679586127400398, + 0.04528612270951271, + 0.02131483145058155, + 0.01462561171501875, + 0.029063226655125618, + 0.020254921168088913, + 0.01966921053826809, + 0.02092471905052662, + 0.017865430563688278, + 0.017627617344260216, + 0.016166532412171364, + 0.02477305382490158, + 0.014337723143398762, + 0.01778022199869156, + 0.007069143000990152, + 0.003969565033912659, + 0.005684320814907551, + 0.004047422669827938, + 0.004272314719855785, + 0.005099200177937746, + 0.0034525073133409023, + 0.007398656569421291, + 0.010664108209311962, + 0.0077467202208936214, + 0.007972341030836105, + 0.006470206659287214, + 0.006554622668772936, + 0.007189290598034859, + 0.008388020098209381, + 0.006974854506552219, + 0.010661864653229713, + 0.022888582199811935, + 0.011642541736364365, + 0.012342287227511406, + 0.011457710526883602, + 0.011220633052289486, + 0.017168695107102394, + 0.011339443735778332, + 0.014610379934310913, + 0.013991446234285831, + 0.014467920176684856, + 0.013853993266820908, + 0.016857312992215157, + 0.013289147056639194, + 0.016454225406050682, + 0.011444056406617165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.03809546306729317 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076287128031253815 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.0076287128031253815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076287128031253815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053852200508117676, + 0.04528597742319107, + 0.058705829083919525, + 0.04324355721473694, + 0.052877649664878845, + 0.0999472513794899, + 0.3712998032569885, + 0.0653306171298027, + 0.07037222385406494, + 0.13908490538597107 + ], + "credit_map": [ + 0.0057954187504947186, + 0.006874537095427513, + 0.006518985610455275, + 0.008174419403076172, + 0.006012457422912121, + 0.007782375440001488, + 0.006327982526272535, + 0.006366020534187555, + 0.003936985041946173, + 0.004053867422044277, + 0.006585694383829832, + 0.004912570584565401, + 0.004824806936085224, + 0.0073560019955039024, + 0.00830293633043766, + 0.005313111934810877, + 0.0041564470157027245, + 0.005381173919886351, + 0.005685776472091675, + 0.0054578776471316814, + 0.007168398704379797, + 0.012079457752406597, + 0.006696527823805809, + 0.012080161832273006, + 0.005903391633182764, + 0.004680643789470196, + 0.006388481240719557, + 0.006896199192851782, + 0.0038299430161714554, + 0.0032104472629725933, + 0.0038870112039148808, + 0.0084474366158247, + 0.006071504671126604, + 0.007870666682720184, + 0.006057028658688068, + 0.004723961465060711, + 0.005961737595498562, + 0.010261202231049538, + 0.006210573948919773, + 0.005720970220863819, + 0.017987282946705818, + 0.014365346170961857, + 0.01110433042049408, + 0.009513293392956257, + 0.00888617243617773, + 0.012857972644269466, + 0.009998301975429058, + 0.015234542079269886, + 0.06325215846300125, + 0.04320898652076721, + 0.06183070316910744, + 0.033542778342962265, + 0.031565725803375244, + 0.037877682596445084, + 0.03257030248641968, + 0.06745144724845886, + 0.013133812695741653, + 0.005874576047062874, + 0.012840601615607738, + 0.007505982182919979, + 0.005057119764387608, + 0.005745796486735344, + 0.00971253588795662, + 0.005460189189761877, + 0.008382382802665234, + 0.013603868894279003, + 0.007872959598898888, + 0.008406705223023891, + 0.006719163153320551, + 0.007143836468458176, + 0.011176619678735733, + 0.007066687569022179, + 0.01759069785475731, + 0.021189792081713676, + 0.016443541273474693, + 0.013564632274210453, + 0.016271810978651047, + 0.01854669861495495, + 0.018884120509028435, + 0.01659359782934189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.0076287128031253815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a vibrant, illustrative style reminiscent of classic animated films, a low-angle, wide-angle shot begins in a slow, majestic rotation, capturing an impossible spectacle. Across a vast, sun-drenched valley, a massive, crumbling castle begins its mighty takeoff. The structure is a chaotic marvel of mismatched stone towers, bent timber extensions, and bolted iron plates, all perched precariously on four massive, piston-driven mechanical legs. As the legs begin to tumble, their heavy, rhythmic steps kicking up clods of earth and grass, the entire fortress pitches forward, gaining momentum with a startling shudder. Thick black smoke and billowing white steam billow violently from a dozen chimneys of various shapes and sizes, staining the bright blue sky red. The camera slowly rotates, following the castle's accelerating path, emphasizing its immense scale against the painterly landscape, culminating in the moment when orange and yellow flames burst from hidden vents in the castle's underbelly, beginning to lift this impossible machine into the air in a breathtaking, dreamlike ascent.", + "seed": 16775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_15/index_9303_seed_16775_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02194903646595776, + "vlm_logprob": -0.02194903646595776, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.02194903646595776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0035371140111237764 + ], + "opposite_logprob_tokens": [ + -6.003537178039551 + ], + "logprob_reward": -0.0035371140111237764, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0035371140111237764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09138398617506027, + 0.0576372891664505, + 0.08319301903247833, + 0.020683232694864273, + 0.047660186886787415, + 0.42761823534965515, + 0.12657800316810608, + 0.05489553138613701, + 0.05207666754722595, + 0.03827379271388054 + ], + "credit_map": [ + 0.008148210123181343, + 0.011428190395236015, + 0.012043122202157974, + 0.014480485580861568, + 0.014457512646913528, + 0.008767027407884598, + 0.0134578300639987, + 0.008601613342761993, + 0.003595337038859725, + 0.004035078454762697, + 0.01608237996697426, + 0.003622208721935749, + 0.005436733830720186, + 0.014876333065330982, + 0.0069407252594828606, + 0.003048493294045329, + 0.010592895559966564, + 0.016541311517357826, + 0.013823084533214569, + 0.010042986832559109, + 0.008139107376337051, + 0.007778132800012827, + 0.00829805713146925, + 0.00797743909060955, + 0.0023329388350248337, + 0.002596010686829686, + 0.0026106617879122496, + 0.0019247176824137568, + 0.00408894894644618, + 0.003906252793967724, + 0.0017632062081247568, + 0.0014604958705604076, + 0.0062253521755337715, + 0.006233289837837219, + 0.00490777101367712, + 0.0074872891418635845, + 0.005078351590782404, + 0.006577319465577602, + 0.00614546425640583, + 0.0050053526647388935, + 0.08840060234069824, + 0.04251968115568161, + 0.12444182485342026, + 0.040469907224178314, + 0.030756153166294098, + 0.04298253357410431, + 0.026295669376850128, + 0.031751848757267, + 0.02535898983478546, + 0.016932308673858643, + 0.015025303699076176, + 0.0157068632543087, + 0.013925198465585709, + 0.01437979843467474, + 0.012337205931544304, + 0.012912338599562645, + 0.011852776631712914, + 0.003995648119598627, + 0.009188879281282425, + 0.0037552935536950827, + 0.005419151391834021, + 0.005483366083353758, + 0.004334799014031887, + 0.010865616612136364, + 0.00584569713100791, + 0.006721649318933487, + 0.005810542963445187, + 0.010494898073375225, + 0.005328064784407616, + 0.004087642300873995, + 0.007754551246762276, + 0.006033620331436396, + 0.005945054814219475, + 0.0023010512813925743, + 0.006563413888216019, + 0.003097794484347105, + 0.00289746792986989, + 0.0029785623773932457, + 0.0036802461836487055, + 0.010810199193656445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0035371140111237764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811049833893776 + ], + "opposite_logprob_tokens": [ + -3.2881104946136475 + ], + "logprob_reward": -0.03811049833893776, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811049833893776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1472887545824051, + 0.1001778170466423, + 0.04449660703539848, + 0.07697837799787521, + 0.03250117227435112, + 0.19835932552814484, + 0.052863504737615585, + 0.15317632257938385, + 0.06956838816404343, + 0.12458962947130203 + ], + "credit_map": [ + 0.015992559492588043, + 0.01906614936888218, + 0.019399765878915787, + 0.020288564264774323, + 0.02320263721048832, + 0.0161772221326828, + 0.018275940790772438, + 0.014885938726365566, + 0.005819095764309168, + 0.006271694786846638, + 0.02755027823150158, + 0.005532704759389162, + 0.008253330364823341, + 0.02997411973774433, + 0.011370529420673847, + 0.005406080279499292, + 0.00542503222823143, + 0.008276925422251225, + 0.005139131564646959, + 0.004054217599332333, + 0.007110211066901684, + 0.006410237401723862, + 0.004083958454430103, + 0.003996907267719507, + 0.008518324233591557, + 0.009147693403065205, + 0.009252066724002361, + 0.007273953873664141, + 0.0165752824395895, + 0.010939220897853374, + 0.007841777987778187, + 0.0074300761334598064, + 0.003708923002704978, + 0.0039436714723706245, + 0.0034180209040641785, + 0.005202780477702618, + 0.003917416092008352, + 0.004722205922007561, + 0.003978377673774958, + 0.003609780687838793, + 0.043171051889657974, + 0.020741386339068413, + 0.05606677755713463, + 0.01829107105731964, + 0.013197899796068668, + 0.019781189039349556, + 0.012062563560903072, + 0.01504743006080389, + 0.011722467839717865, + 0.005205833353102207, + 0.007433868013322353, + 0.0067701456137001514, + 0.005077152047306299, + 0.00490296958014369, + 0.006176923867315054, + 0.005574157927185297, + 0.01996018923819065, + 0.021385619416832924, + 0.017810150980949402, + 0.015684744343161583, + 0.023240700364112854, + 0.014135698787868023, + 0.018858280032873154, + 0.022100962698459625, + 0.008429902605712414, + 0.01290244609117508, + 0.005902767647057772, + 0.011186596006155014, + 0.007302670273929834, + 0.006256807595491409, + 0.010268200188875198, + 0.007319006137549877, + 0.01880994625389576, + 0.008181605488061905, + 0.022036198526620865, + 0.008608872070908546, + 0.009558121673762798, + 0.009902755729854107, + 0.009955553337931633, + 0.0375366136431694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.03811049833893776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11500591039657593, + 0.06117987260222435, + 0.09130097180604935, + 0.034880079329013824, + 0.05618179216980934, + 0.29677045345306396, + 0.12856554985046387, + 0.06485826522111893, + 0.07984957098960876, + 0.0714075118303299 + ], + "credit_map": [ + 0.008851820603013039, + 0.013210710138082504, + 0.013803595677018166, + 0.022612378001213074, + 0.016365954652428627, + 0.01131858304142952, + 0.01981271617114544, + 0.009030157700181007, + 0.007074122782796621, + 0.006703592836856842, + 0.012287966907024384, + 0.005907461047172546, + 0.007240796461701393, + 0.006935336161404848, + 0.006278178188949823, + 0.008752423338592052, + 0.018227774649858475, + 0.00920496229082346, + 0.01466396078467369, + 0.005961467977613211, + 0.011700225993990898, + 0.012949646450579166, + 0.007110086269676685, + 0.011482851579785347, + 0.0032755143474787474, + 0.003112782957032323, + 0.004746460355818272, + 0.0043364763259887695, + 0.0057297139428555965, + 0.006023951340466738, + 0.003863021032884717, + 0.003792164148762822, + 0.007687605917453766, + 0.010334017686545849, + 0.005951374769210815, + 0.006494545843452215, + 0.007217017933726311, + 0.0068026394583284855, + 0.004868995398283005, + 0.006825605407357216, + 0.06138647347688675, + 0.031183021143078804, + 0.08510453253984451, + 0.029374655336141586, + 0.021420395001769066, + 0.029693633317947388, + 0.016598505899310112, + 0.022009262815117836, + 0.0290650837123394, + 0.01608433946967125, + 0.015681590884923935, + 0.015242482535541058, + 0.012242837809026241, + 0.01384694967418909, + 0.013267504051327705, + 0.013134768232703209, + 0.006329133175313473, + 0.008661976084113121, + 0.007249262183904648, + 0.009228064678609371, + 0.008579489775002003, + 0.006087213754653931, + 0.009999132715165615, + 0.008723998442292213, + 0.008747528307139874, + 0.010362289845943451, + 0.010156968608498573, + 0.011226125992834568, + 0.012624314986169338, + 0.006844612769782543, + 0.009277245961129665, + 0.0106104901060462, + 0.009813399985432625, + 0.009616338647902012, + 0.01046157255768776, + 0.005581477656960487, + 0.008426492102444172, + 0.008201615884900093, + 0.005820402875542641, + 0.013486226089298725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0052520157769322395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864978417754173 + ], + "opposite_logprob_tokens": [ + -3.048649787902832 + ], + "logprob_reward": -0.04864978417754173, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864978417754173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07417425513267517, + 0.048759181052446365, + 0.057496774941682816, + 0.11296554654836655, + 0.10547178238630295, + 0.2896588444709778, + 0.10819061845541, + 0.122678741812706, + 0.05127165466547012, + 0.0293325986713171 + ], + "credit_map": [ + 0.006157823838293552, + 0.008065867237746716, + 0.011230867356061935, + 0.01016257330775261, + 0.01154596358537674, + 0.008780472911894321, + 0.00801771692931652, + 0.010212969034910202, + 0.00541260838508606, + 0.005206937436014414, + 0.007386248093098402, + 0.004229176789522171, + 0.006399529054760933, + 0.006499499082565308, + 0.008534700609743595, + 0.0050904820673167706, + 0.00828018132597208, + 0.00737882312387228, + 0.01229240745306015, + 0.007236910052597523, + 0.0050487034022808075, + 0.0050018285401165485, + 0.005640259012579918, + 0.006617667619138956, + 0.014242802746593952, + 0.012176243588328362, + 0.00987880490720272, + 0.011256938800215721, + 0.020101167261600494, + 0.02203384041786194, + 0.011523926630616188, + 0.011751820333302021, + 0.014675804413855076, + 0.01638275384902954, + 0.012110070325434208, + 0.011044513434171677, + 0.0108222970739007, + 0.016703279688954353, + 0.011486023664474487, + 0.01224703062325716, + 0.05474746599793434, + 0.026361070573329926, + 0.054892707616090775, + 0.029979513958096504, + 0.04136180505156517, + 0.029255792498588562, + 0.02647949941456318, + 0.02658097818493843, + 0.022661123424768448, + 0.01246870867908001, + 0.015641270205378532, + 0.012933296151459217, + 0.01110640075057745, + 0.010805201716721058, + 0.011190449818968773, + 0.011384171433746815, + 0.026256684213876724, + 0.01102419849485159, + 0.017579732462763786, + 0.008203225210309029, + 0.010209706611931324, + 0.011796613223850727, + 0.011740542948246002, + 0.025868039578199387, + 0.005853945389389992, + 0.007452119141817093, + 0.005870105233043432, + 0.007112319581210613, + 0.006952045951038599, + 0.004888257943093777, + 0.006064776796847582, + 0.00707808556035161, + 0.003733690595254302, + 0.0031517446041107178, + 0.0038923800457268953, + 0.005043117329478264, + 0.0027575260028243065, + 0.0035176018718630075, + 0.0029053192120045424, + 0.004331219010055065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.04864978417754173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195770025253296 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195770025253296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195770025253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09304700046777725, + 0.04396342113614082, + 0.07093862444162369, + 0.05695942044258118, + 0.09703179448843002, + 0.13839438557624817, + 0.10338925570249557, + 0.0775640606880188, + 0.14965812861919403, + 0.16905394196510315 + ], + "credit_map": [ + 0.011703670024871826, + 0.00904104020446539, + 0.016118474304676056, + 0.0114449979737401, + 0.015676049515604973, + 0.008820218034088612, + 0.01253274455666542, + 0.00770979980006814, + 0.004570581018924713, + 0.004245339892804623, + 0.00956178829073906, + 0.0034461074974387884, + 0.0054079629480838776, + 0.006147382315248251, + 0.005635574925690889, + 0.004948681220412254, + 0.00880458578467369, + 0.017499256879091263, + 0.006790943443775177, + 0.008795213885605335, + 0.006192808970808983, + 0.008399276994168758, + 0.008054647594690323, + 0.006401885766535997, + 0.006380924955010414, + 0.005479883402585983, + 0.005536087788641453, + 0.005126607604324818, + 0.013281408697366714, + 0.010494541376829147, + 0.0060450234450399876, + 0.004614935256540775, + 0.01441653911024332, + 0.016078172251582146, + 0.01177862100303173, + 0.011933240108191967, + 0.0094896936789155, + 0.012666436843574047, + 0.00911783054471016, + 0.011551255360245705, + 0.0206950381398201, + 0.012967166490852833, + 0.030884306877851486, + 0.017443588003516197, + 0.019028740003705025, + 0.014688631519675255, + 0.01261194422841072, + 0.010074961930513382, + 0.021777072921395302, + 0.011781364679336548, + 0.014445376582443714, + 0.011935725808143616, + 0.011683046817779541, + 0.01188972033560276, + 0.007274881470948458, + 0.012602059170603752, + 0.01369986217468977, + 0.0077156201004981995, + 0.011679599992930889, + 0.0071150329895317554, + 0.0053026010282337666, + 0.008065947331488132, + 0.007581827696412802, + 0.0164035651832819, + 0.020727673545479774, + 0.018852952867746353, + 0.023576896637678146, + 0.016058728098869324, + 0.024098820984363556, + 0.012954342179000378, + 0.012852244079113007, + 0.020536450669169426, + 0.020532505586743355, + 0.028768304735422134, + 0.025032581761479378, + 0.014383905567228794, + 0.02468438632786274, + 0.019051071256399155, + 0.012280957773327827, + 0.024320220574736595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.014195770025253296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a vibrant, illustrative style reminiscent of classic animated films, a low-angle, wide-angle shot begins in a slow, majestic rotation, capturing an impossible spectacle. Across a vast, sun-drenched valley, a massive, crumbling castle begins its mighty takeoff. The structure is a chaotic marvel of mismatched stone towers, bent timber extensions, and bolted iron plates, all perched precariously on four massive, piston-driven mechanical legs. As the legs begin to tumble, their heavy, rhythmic steps kicking up clods of earth and grass, the entire fortress pitches forward, gaining momentum with a startling shudder. Thick black smoke and billowing white steam billow violently from a dozen chimneys of various shapes and sizes, staining the bright blue sky red. The camera slowly rotates, following the castle's accelerating path, emphasizing its immense scale against the painterly landscape, culminating in the moment when orange and yellow flames burst from hidden vents in the castle's underbelly, beginning to lift this impossible machine into the air in a breathtaking, dreamlike ascent.", + "seed": 16776, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_15/index_9303_seed_16776_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.040463393460959195, + "vlm_logprob": -0.040463393460959195, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.040463393460959195, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0039177583530545235 + ], + "opposite_logprob_tokens": [ + -6.128917694091797 + ], + "logprob_reward": -0.0039177583530545235, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0039177583530545235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13014230132102966, + 0.0686991885304451, + 0.07040369510650635, + 0.255056232213974, + 0.03703952580690384, + 0.15909329056739807, + 0.06055118516087532, + 0.04433317482471466, + 0.09559769928455353, + 0.07908369600772858 + ], + "credit_map": [ + 0.015598587691783905, + 0.011738057248294353, + 0.02177688665688038, + 0.016788963228464127, + 0.011149357073009014, + 0.022470206022262573, + 0.015916211530566216, + 0.01470403466373682, + 0.00563786318525672, + 0.004165034741163254, + 0.020874658599495888, + 0.007304809056222439, + 0.004383031744509935, + 0.013680718839168549, + 0.006471964530646801, + 0.006181109230965376, + 0.01106088887900114, + 0.007494226098060608, + 0.017048196867108345, + 0.005043569952249527, + 0.006025063339620829, + 0.010876249521970749, + 0.007826996967196465, + 0.005028506740927696, + 0.009142947383224964, + 0.011276031844317913, + 0.043252717703580856, + 0.008498436771333218, + 0.011332794092595577, + 0.09898132085800171, + 0.014174114912748337, + 0.05839785933494568, + 0.004430157598108053, + 0.0038846274837851524, + 0.0026358747854828835, + 0.004943919833749533, + 0.004682513885200024, + 0.004660234320908785, + 0.004320630803704262, + 0.007481567561626434, + 0.013917740434408188, + 0.01923336088657379, + 0.014302916824817657, + 0.01937771402299404, + 0.015412494540214539, + 0.03346769139170647, + 0.020409153774380684, + 0.022972218692302704, + 0.005735686048865318, + 0.007171210367232561, + 0.005489175673574209, + 0.006354374345391989, + 0.005461344961076975, + 0.010381307452917099, + 0.010826312005519867, + 0.009131774306297302, + 0.0054646446369588375, + 0.004229533951729536, + 0.003181644482538104, + 0.0051129888743162155, + 0.005105643067508936, + 0.004460216499865055, + 0.008652674034237862, + 0.008125828579068184, + 0.011821823194622993, + 0.010400693863630295, + 0.008283916860818863, + 0.010847557336091995, + 0.014490440487861633, + 0.010392749682068825, + 0.014721023850142956, + 0.014639497734606266, + 0.007374906912446022, + 0.009023526683449745, + 0.006984469946473837, + 0.010420178063213825, + 0.013789433054625988, + 0.00620117736980319, + 0.014153859578073025, + 0.011136139743030071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.0039177583530545235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797343984246254 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797343984246254, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797343984246254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13378049433231354, + 0.028585806488990784, + 0.0587814562022686, + 0.17616894841194153, + 0.056485891342163086, + 0.08604603260755539, + 0.06268493086099625, + 0.04165254533290863, + 0.1982010304927826, + 0.15761283040046692 + ], + "credit_map": [ + 0.017225481569767, + 0.01429374422878027, + 0.02381991595029831, + 0.018842801451683044, + 0.012132294476032257, + 0.018188999965786934, + 0.016049502417445183, + 0.013227751478552818, + 0.002697251969948411, + 0.0021352742332965136, + 0.007247716188430786, + 0.003469988238066435, + 0.0024466041941195726, + 0.004941610153764486, + 0.00265441439114511, + 0.0029929487500339746, + 0.0049420674331486225, + 0.003602842567488551, + 0.02134687267243862, + 0.004140094853937626, + 0.003320574527606368, + 0.012492765672504902, + 0.004874615930020809, + 0.004061623010784388, + 0.017313916236162186, + 0.02277647890150547, + 0.02256530523300171, + 0.010112947784364223, + 0.02023933082818985, + 0.04044992849230766, + 0.017277639359235764, + 0.02543339878320694, + 0.005808258429169655, + 0.00631945114582777, + 0.005060551222413778, + 0.011578700505197048, + 0.005035930313169956, + 0.007956276647746563, + 0.006959086749702692, + 0.007767629809677601, + 0.008957883343100548, + 0.011441228911280632, + 0.007793392986059189, + 0.009976603090763092, + 0.008046738803386688, + 0.015251420438289642, + 0.012595847249031067, + 0.011982914060354233, + 0.006260031834244728, + 0.007916387170553207, + 0.006069456692785025, + 0.007457118947058916, + 0.005521587561815977, + 0.009491872042417526, + 0.012022481299936771, + 0.007945995777845383, + 0.006935860496014357, + 0.00443790340796113, + 0.003667013021185994, + 0.004651503171771765, + 0.0046900976449251175, + 0.0035726341884583235, + 0.007767663337290287, + 0.005929868668317795, + 0.025884581729769707, + 0.02810966968536377, + 0.014202097430825233, + 0.01608509197831154, + 0.037235382944345474, + 0.021984564140439034, + 0.021771695464849472, + 0.03292796015739441, + 0.020313076674938202, + 0.02105400338768959, + 0.011197668500244617, + 0.019019166007637978, + 0.020617956295609474, + 0.013374735601246357, + 0.03099244274199009, + 0.02104377932846546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.009797343984246254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637083038687706 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637083038687706, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637083038687706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20347829163074493, + 0.13664619624614716, + 0.0772191509604454, + 0.08083989471197128, + 0.07630673050880432, + 0.1634165346622467, + 0.03228102996945381, + 0.04235823452472687, + 0.080356664955616, + 0.1070973351597786 + ], + "credit_map": [ + 0.03341735899448395, + 0.018207615241408348, + 0.03627902269363403, + 0.025423688814044, + 0.02046157978475094, + 0.028074217960238457, + 0.020326407626271248, + 0.02128838188946247, + 0.011322692967951298, + 0.011412071995437145, + 0.02855181135237217, + 0.019266746938228607, + 0.01221973355859518, + 0.022422177717089653, + 0.016744084656238556, + 0.014706868678331375, + 0.009529449976980686, + 0.010433276183903217, + 0.010548936203122139, + 0.007793016731739044, + 0.006896778475493193, + 0.009406974539160728, + 0.012490755878388882, + 0.01011995505541563, + 0.009699376299977303, + 0.013733644969761372, + 0.008261729031801224, + 0.0066659399308264256, + 0.011112737469375134, + 0.010478612966835499, + 0.00996004231274128, + 0.01092780102044344, + 0.008234509266912937, + 0.009483280591666698, + 0.008514483459293842, + 0.00935245119035244, + 0.005851666908711195, + 0.009100754745304585, + 0.00794629380106926, + 0.01782328262925148, + 0.014638690277934074, + 0.015553024597465992, + 0.0163401011377573, + 0.015375472605228424, + 0.010199662297964096, + 0.036205973476171494, + 0.03129859268665314, + 0.023805003613233566, + 0.003685916541144252, + 0.004302147310227156, + 0.0026478287763893604, + 0.00317886914126575, + 0.002570357406511903, + 0.005434970371425152, + 0.006261372938752174, + 0.004199562128633261, + 0.004431567620486021, + 0.006004094611853361, + 0.0032928762957453728, + 0.0052191209979355335, + 0.005573446862399578, + 0.0036645899526774883, + 0.007762685418128967, + 0.006409849040210247, + 0.01054190844297409, + 0.0075959982350468636, + 0.006633814424276352, + 0.00984839629381895, + 0.009950577281415462, + 0.008001641370356083, + 0.013073454611003399, + 0.0147108668461442, + 0.008799515664577484, + 0.011531269177794456, + 0.006923564709722996, + 0.013901744969189167, + 0.019415192306041718, + 0.008925051428377628, + 0.02553350105881691, + 0.012067487463355064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004637083038687706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798696368932724 + ], + "opposite_logprob_tokens": [ + -1.8048696517944336 + ], + "logprob_reward": -0.1798696368932724, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798696368932724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09413875639438629, + 0.03389182686805725, + 0.11911305785179138, + 0.19636686146259308, + 0.031375836580991745, + 0.07111290097236633, + 0.06424451619386673, + 0.0986819863319397, + 0.15783128142356873, + 0.13324306905269623 + ], + "credit_map": [ + 0.012289254926145077, + 0.011023825965821743, + 0.012489776127040386, + 0.008683435618877411, + 0.008013516664505005, + 0.011399188078939915, + 0.010925421491265297, + 0.019314322620630264, + 0.002715915907174349, + 0.002580255037173629, + 0.005321968346834183, + 0.0033916207030415535, + 0.00277371215634048, + 0.005051746964454651, + 0.0065217334777116776, + 0.0055348677560687065, + 0.008209883235394955, + 0.00687509635463357, + 0.03935565426945686, + 0.006526960991322994, + 0.00621374137699604, + 0.029512429609894753, + 0.006731520406901836, + 0.01568775251507759, + 0.007212796248495579, + 0.006552739534527063, + 0.0336417518556118, + 0.007196388673037291, + 0.008484096266329288, + 0.07527986168861389, + 0.009838391095399857, + 0.04816080629825592, + 0.003339035203680396, + 0.0035918704234063625, + 0.0037080333568155766, + 0.003768104361370206, + 0.00459652952849865, + 0.0034179603680968285, + 0.003442463232204318, + 0.00551183894276619, + 0.008421080186963081, + 0.009604879654943943, + 0.005911427084356546, + 0.006672380957752466, + 0.004681748803704977, + 0.014353008009493351, + 0.01154262013733387, + 0.009925749152898788, + 0.006743140518665314, + 0.005297951865941286, + 0.005693670362234116, + 0.006524515803903341, + 0.005630873143672943, + 0.012117921374738216, + 0.014756296761333942, + 0.007480139844119549, + 0.012583806179463863, + 0.012037991546094418, + 0.0056229932233691216, + 0.01029624417424202, + 0.011892376467585564, + 0.00866373348981142, + 0.02385294996201992, + 0.013731882907450199, + 0.01608269102871418, + 0.014100419357419014, + 0.01338176429271698, + 0.023597098886966705, + 0.015525508671998978, + 0.015827516093850136, + 0.02769635058939457, + 0.03161992132663727, + 0.012843974865972996, + 0.01155682560056448, + 0.01054373662918806, + 0.01996953785419464, + 0.011855983175337315, + 0.013005383312702179, + 0.03563747555017471, + 0.017830125987529755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.1798696368932724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style vibrant and illustrative, reminiscent of classic animated films?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the castle massive, crumbling, and constructed of mismatched stone towers, bent timber extensions, and bolted iron plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the castle stand on four massive, piston-driven mechanical legs that are tumbling and kicking up clods of earth and grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick black smoke and billowing white steam rise from a dozen chimneys, staining the bright blue sky red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly rotate to follow the castle's accelerating path, and do orange and yellow flames burst from hidden vents in the underbelly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095145035535097 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095145035535097, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095145035535097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0988694503903389, + 0.047402262687683105, + 0.12585556507110596, + 0.18159006536006927, + 0.02038975991308689, + 0.07187845557928085, + 0.04538135603070259, + 0.09088524430990219, + 0.2152099311351776, + 0.10253793746232986 + ], + "credit_map": [ + 0.012413401156663895, + 0.009518766775727272, + 0.015829553827643394, + 0.0144611531868577, + 0.009316543117165565, + 0.011640197597444057, + 0.013801575638353825, + 0.01188825536519289, + 0.004365827888250351, + 0.004006871487945318, + 0.009949704632163048, + 0.004696797579526901, + 0.004169085994362831, + 0.007302458398044109, + 0.006822612602263689, + 0.006088904570788145, + 0.009206011891365051, + 0.006738695781677961, + 0.05293372645974159, + 0.007723325863480568, + 0.0062806266359984875, + 0.029126444831490517, + 0.006423160433769226, + 0.007423573173582554, + 0.007250177673995495, + 0.008278075605630875, + 0.02985321171581745, + 0.006822296418249607, + 0.00793355144560337, + 0.0676807388663292, + 0.009603910148143768, + 0.044168103486299515, + 0.002518946537747979, + 0.002858103020116687, + 0.0023129701148718596, + 0.002726481296122074, + 0.0025417839642614126, + 0.0023567453026771545, + 0.0022913236171007156, + 0.0027834060601890087, + 0.006004905793815851, + 0.009813402779400349, + 0.00623853225260973, + 0.00911214854568243, + 0.006473028101027012, + 0.015572124160826206, + 0.0083014452829957, + 0.010362866334617138, + 0.005032646469771862, + 0.0053648632019758224, + 0.003026860998943448, + 0.003902025520801544, + 0.0046045370399951935, + 0.009060712531208992, + 0.008537817746400833, + 0.0058518946170806885, + 0.011242032051086426, + 0.010304925963282585, + 0.005752083845436573, + 0.011795361526310444, + 0.00983983464539051, + 0.009198512881994247, + 0.013438370078802109, + 0.019314119592308998, + 0.02358468435704708, + 0.029677772894501686, + 0.013556656427681446, + 0.023961050435900688, + 0.03159020468592644, + 0.020631812512874603, + 0.03379620984196663, + 0.03841153904795647, + 0.00956386886537075, + 0.013712121173739433, + 0.009308780543506145, + 0.012910286895930767, + 0.01569751463830471, + 0.008954430930316448, + 0.018060483038425446, + 0.014330452308058739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.004095145035535097 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_tensor([8899, 8899])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_tensor([8899, 8899])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..85777e078c527585a3b72b50c302dbea377b1593 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_tensor([8899, 8899])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium-long shot, shot from a low, upward angle, slowly and steadily zooms in on a white-haired elderly man standing alone on the precipice of a precipitous mountain. His long, silvery hair and a simple gray robe made of rough linen flutter violently in the howling wind, yet his upright figure remains motionless, like a mountain, his feet firmly rooted to the rock beneath him, creating a striking contrast between movement and stillness. As the camera gradually moves in, his deeply wrinkled face and deep, ancient eyes, like ancient wells, gaze resolutely at the rolling, mist-shrouded mountains in the distance. Against the overcast sky, the cold, diffuse light accentuates his resolute silhouette, highlighting the folds of his robe and the creases on his face. The entire scene, shrouded in a cold, blue-gray hue, framed by a vast sky and blurred mountain ranges, creates a sense of epic loneliness and a timeless, indestructible, and quiet power.", + "seed": 591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_8899_seed_591_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03529881115537137, + "vlm_logprob": -0.03529881115537137, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.03529881115537137, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008469495922327042 + ], + "opposite_logprob_tokens": [ + -5.258469581604004 + ], + "logprob_reward": -0.008469495922327042, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008469495922327042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128579244017601, + 0.13926704227924347, + 0.03300460800528526, + 0.028117354959249496, + 0.07638662308454514, + 0.09002777189016342, + 0.14069828391075134, + 0.08937429636716843, + 0.15874987840652466, + 0.11579485982656479 + ], + "credit_map": [ + 0.018691811710596085, + 0.01750858686864376, + 0.01388532668352127, + 0.015285198576748371, + 0.013203934766352177, + 0.022097505629062653, + 0.015786856412887573, + 0.012120014056563377, + 0.012444774620234966, + 0.010341388173401356, + 0.038959480822086334, + 0.016902709379792213, + 0.008199124597012997, + 0.031666193157434464, + 0.01080353558063507, + 0.009949829429388046, + 0.0028188081923872232, + 0.004221384413540363, + 0.004143605940043926, + 0.003829605644568801, + 0.0035220119170844555, + 0.003840397344902158, + 0.004889274016022682, + 0.00573952030390501, + 0.002114992355927825, + 0.002981402212753892, + 0.0034466306678950787, + 0.002911079442128539, + 0.005355877801775932, + 0.003847418585792184, + 0.003809300484135747, + 0.0036506536416709423, + 0.009778766892850399, + 0.00751446234062314, + 0.010480557568371296, + 0.008280089125037193, + 0.012390262447297573, + 0.009822271764278412, + 0.010005465708673, + 0.00811474397778511, + 0.013472522608935833, + 0.009659293107688427, + 0.014080426655709743, + 0.00846031866967678, + 0.00863331276923418, + 0.011121787130832672, + 0.013469981960952282, + 0.011130130849778652, + 0.016443854197859764, + 0.026772677898406982, + 0.014603734947741032, + 0.016719646751880646, + 0.021135086193680763, + 0.013478793203830719, + 0.015996897593140602, + 0.015547601506114006, + 0.007355656009167433, + 0.01155246514827013, + 0.011043988168239594, + 0.009733896702528, + 0.009041591547429562, + 0.01009640097618103, + 0.01015017181634903, + 0.02040012553334236, + 0.024406515061855316, + 0.01247369684278965, + 0.03982926160097122, + 0.012819440104067326, + 0.012328972108662128, + 0.01957935094833374, + 0.015500452369451523, + 0.021812185645103455, + 0.017480701208114624, + 0.020471064373850822, + 0.013649802654981613, + 0.01405902486294508, + 0.011843797750771046, + 0.011308313347399235, + 0.015079920180141926, + 0.01190223079174757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.008469495922327042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05504918470978737 + ], + "opposite_logprob_tokens": [ + -2.930049180984497 + ], + "logprob_reward": -0.05504918470978737, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05504918470978737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2954710125923157, + 0.2632867693901062, + 0.05943787470459938, + 0.030244862660765648, + 0.04054545238614082, + 0.05663728713989258, + 0.05454547330737114, + 0.05474962666630745, + 0.06079505756497383, + 0.08428657799959183 + ], + "credit_map": [ + 0.02591986209154129, + 0.03933291882276535, + 0.04310129955410957, + 0.03208334743976593, + 0.029199063777923584, + 0.05126509815454483, + 0.037756502628326416, + 0.036812923848629, + 0.015070131048560143, + 0.056071072816848755, + 0.03847169131040573, + 0.022696981206536293, + 0.01800866797566414, + 0.07574846595525742, + 0.02213227190077305, + 0.015087493695318699, + 0.0050380052998661995, + 0.005965812131762505, + 0.007441581692546606, + 0.00811535120010376, + 0.006278143264353275, + 0.0071174101904034615, + 0.01095703523606062, + 0.008524534292519093, + 0.003167220391333103, + 0.0032901803497225046, + 0.0035843357909470797, + 0.002890537027269602, + 0.004142658319324255, + 0.004383706953376532, + 0.00501311244443059, + 0.003773111617192626, + 0.006643366999924183, + 0.004092245828360319, + 0.005035931710153818, + 0.004989139270037413, + 0.004735706839710474, + 0.0049586850218474865, + 0.004427148494869471, + 0.0056632282212376595, + 0.005093097686767578, + 0.007028039079159498, + 0.006903169211000204, + 0.00771459611132741, + 0.006740787997841835, + 0.0054602595046162605, + 0.011578947305679321, + 0.006118393037468195, + 0.006078425794839859, + 0.00995416659861803, + 0.004967476241290569, + 0.0065514142625033855, + 0.006954357959330082, + 0.005722945090383291, + 0.006424781866371632, + 0.007891904562711716, + 0.005791946779936552, + 0.008802701719105244, + 0.005572723224759102, + 0.0059378184378147125, + 0.0070022945292294025, + 0.005966185126453638, + 0.007858458906412125, + 0.007817498408257961, + 0.0053640250116586685, + 0.005808298010379076, + 0.009642837569117546, + 0.010277501307427883, + 0.006567343138158321, + 0.006532446946948767, + 0.010114272125065327, + 0.006488333456218243, + 0.012784522958099842, + 0.015975620597600937, + 0.01061891671270132, + 0.012149319052696228, + 0.009184892289340496, + 0.009045179933309555, + 0.008012746460735798, + 0.006515377666801214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.05504918470978737 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10037954896688461 + ], + "opposite_logprob_tokens": [ + -2.350379467010498 + ], + "logprob_reward": -0.10037954896688461, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10037954896688461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21364611387252808, + 0.12574990093708038, + 0.05787960812449455, + 0.06160591170191765, + 0.10282842069864273, + 0.08748792111873627, + 0.07211608439683914, + 0.10944969207048416, + 0.10702214390039444, + 0.062214162200689316 + ], + "credit_map": [ + 0.019903825595974922, + 0.023806806653738022, + 0.029432788491249084, + 0.022333376109600067, + 0.02355779893696308, + 0.03004172444343567, + 0.03550264984369278, + 0.0290671419352293, + 0.010355803184211254, + 0.012200540862977505, + 0.026988007128238678, + 0.013093625195324421, + 0.010594978928565979, + 0.02751127816736698, + 0.013491325080394745, + 0.011514334008097649, + 0.004514890257269144, + 0.00488766236230731, + 0.008696161210536957, + 0.006781898904591799, + 0.009658644907176495, + 0.0073068514466285706, + 0.010614641010761261, + 0.005418861750513315, + 0.006391121540218592, + 0.005526106804609299, + 0.006574693135917187, + 0.006454857997596264, + 0.012660154141485691, + 0.00642271526157856, + 0.011794313788414001, + 0.00578194810077548, + 0.010099939070641994, + 0.009168275631964207, + 0.012110612355172634, + 0.012724115513265133, + 0.018729427829384804, + 0.010272258892655373, + 0.019680960103869438, + 0.010042834095656872, + 0.0098345335572958, + 0.01262434758245945, + 0.012581152841448784, + 0.011848017573356628, + 0.01104152575135231, + 0.009062463417649269, + 0.012499763630330563, + 0.007996118627488613, + 0.008885595016181469, + 0.01334658544510603, + 0.007594495080411434, + 0.007268399931490421, + 0.009193789213895798, + 0.006900150328874588, + 0.009424520656466484, + 0.009502552449703217, + 0.01267985813319683, + 0.019181538373231888, + 0.012797731906175613, + 0.01702308841049671, + 0.013782803900539875, + 0.010824485681951046, + 0.012656918726861477, + 0.01050326693803072, + 0.01291735377162695, + 0.014074819162487984, + 0.01405649445950985, + 0.01224254909902811, + 0.011781739071011543, + 0.011723310686647892, + 0.017718259245157242, + 0.012507610954344273, + 0.00846460834145546, + 0.008639252744615078, + 0.008452670648694038, + 0.008373941294848919, + 0.006903990637511015, + 0.0068670352920889854, + 0.0063344263471663, + 0.008178235962986946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.10037954896688461 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18743930757045746, + 0.10027284175157547, + 0.052078571170568466, + 0.03693459928035736, + 0.07122039049863815, + 0.07101281732320786, + 0.08512859791517258, + 0.07696153968572617, + 0.16388560831546783, + 0.15506574511528015 + ], + "credit_map": [ + 0.01597416214644909, + 0.02398192696273327, + 0.040515463799238205, + 0.022613679990172386, + 0.01285272091627121, + 0.038344379514455795, + 0.020470188930630684, + 0.0126867750659585, + 0.006388555746525526, + 0.029781026765704155, + 0.006964069325476885, + 0.008407069370150566, + 0.004674561787396669, + 0.03016405552625656, + 0.006241649389266968, + 0.007651848718523979, + 0.004993775859475136, + 0.006440056953579187, + 0.00639958493411541, + 0.0067040142603218555, + 0.005593679379671812, + 0.007119230926036835, + 0.008257861249148846, + 0.0065703620202839375, + 0.003966663032770157, + 0.004456568509340286, + 0.005040299613028765, + 0.0044167484156787395, + 0.003606262383982539, + 0.0035015239845961332, + 0.004159123636782169, + 0.007787409238517284, + 0.010078518651425838, + 0.007301792036741972, + 0.008786484599113464, + 0.008012755773961544, + 0.011532281525433064, + 0.008581001311540604, + 0.008616008795797825, + 0.008311543613672256, + 0.008259955793619156, + 0.008044342510402203, + 0.009440443478524685, + 0.009247747249901295, + 0.007998750545084476, + 0.010689212009310722, + 0.009267522022128105, + 0.008064847439527512, + 0.010108714923262596, + 0.012972619384527206, + 0.010199064388871193, + 0.01067456603050232, + 0.01034766435623169, + 0.008728224784135818, + 0.010220066644251347, + 0.011877668090164661, + 0.00926507730036974, + 0.012220115400850773, + 0.009497340768575668, + 0.007354337722063065, + 0.008416119031608105, + 0.008288085460662842, + 0.010402039624750614, + 0.011518415994942188, + 0.0188848078250885, + 0.012770264409482479, + 0.027537329122424126, + 0.017251290380954742, + 0.012143894098699093, + 0.01929970271885395, + 0.021471727639436722, + 0.03452659398317337, + 0.03178798407316208, + 0.02326706051826477, + 0.02107226848602295, + 0.018181212246418, + 0.013979107141494751, + 0.020198825746774673, + 0.0148572763428092, + 0.011721995659172535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015148365637287498 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015148365637287498, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015148365637287498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16381295025348663, + 0.18624548614025116, + 0.031404025852680206, + 0.0314641073346138, + 0.04521823301911354, + 0.18798360228538513, + 0.04386601597070694, + 0.08553607761859894, + 0.14542926847934723, + 0.0790402963757515 + ], + "credit_map": [ + 0.014445547945797443, + 0.022224588319659233, + 0.030743328854441643, + 0.021382302045822144, + 0.01836768165230751, + 0.022671077400445938, + 0.01869388110935688, + 0.015284538269042969, + 0.014977119863033295, + 0.0212908573448658, + 0.04346504062414169, + 0.02039623260498047, + 0.009790743701159954, + 0.04715089127421379, + 0.014075024984776974, + 0.015099577605724335, + 0.0031686180736869574, + 0.0032502266112715006, + 0.004221070557832718, + 0.0031320543494075537, + 0.004341823980212212, + 0.0047603496350348, + 0.003516347147524357, + 0.005013536661863327, + 0.003360721981152892, + 0.003686637384817004, + 0.004167672712355852, + 0.0020870438311249018, + 0.005598811432719231, + 0.004916199017316103, + 0.003021285869181156, + 0.004625733010470867, + 0.004946068860590458, + 0.003421367146074772, + 0.008991206996142864, + 0.006527203135192394, + 0.0051362658850848675, + 0.005097668617963791, + 0.0052961367182433605, + 0.005802318453788757, + 0.032114721834659576, + 0.018941141664981842, + 0.033731915056705475, + 0.016941701993346214, + 0.01596718840301037, + 0.030647728592157364, + 0.01571391336619854, + 0.023925282061100006, + 0.005174823570996523, + 0.009190787561237812, + 0.0043821982108056545, + 0.004470974672585726, + 0.005464917980134487, + 0.004743676166981459, + 0.005622423253953457, + 0.004816216416656971, + 0.010266301222145557, + 0.01665029115974903, + 0.007606506813317537, + 0.009990153834223747, + 0.011418784037232399, + 0.009809920564293861, + 0.008817379362881184, + 0.010976739227771759, + 0.015328887850046158, + 0.015285175293684006, + 0.02341124415397644, + 0.015354989096522331, + 0.014404231682419777, + 0.016650259494781494, + 0.025936761870980263, + 0.019057709723711014, + 0.014982241205871105, + 0.012485423125326633, + 0.009010660462081432, + 0.008487086743116379, + 0.00746172247454524, + 0.010701528750360012, + 0.007821696810424328, + 0.008089936338365078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.0015148365637287498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a low, upward angle, slowly and steadily zooms in on a white-haired elderly man standing alone on the precipice of a precipitous mountain. His long, silvery hair and a simple gray robe made of rough linen flutter violently in the howling wind, yet his upright figure remains motionless, like a mountain, his feet firmly rooted to the rock beneath him, creating a striking contrast between movement and stillness. As the camera gradually moves in, his deeply wrinkled face and deep, ancient eyes, like ancient wells, gaze resolutely at the rolling, mist-shrouded mountains in the distance. Against the overcast sky, the cold, diffuse light accentuates his resolute silhouette, highlighting the folds of his robe and the creases on his face. The entire scene, shrouded in a cold, blue-gray hue, framed by a vast sky and blurred mountain ranges, creates a sense of epic loneliness and a timeless, indestructible, and quiet power.", + "seed": 592, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_8899_seed_592_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11860288181342185, + "vlm_logprob": -0.11860288181342185, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.11860288181342185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006936872377991676 + ], + "opposite_logprob_tokens": [ + -5.506937026977539 + ], + "logprob_reward": -0.006936872377991676, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006936872377991676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24522052705287933, + 0.053443096578121185, + 0.0839802622795105, + 0.09038077294826508, + 0.049612388014793396, + 0.05810672789812088, + 0.05108578875660896, + 0.03658100590109825, + 0.09596490114927292, + 0.23562446236610413 + ], + "credit_map": [ + 0.027694586664438248, + 0.03332560136914253, + 0.048175036907196045, + 0.038897059857845306, + 0.019196363165974617, + 0.024215228855609894, + 0.026557279750704765, + 0.02715940587222576, + 0.007080123294144869, + 0.01118744071573019, + 0.006854274310171604, + 0.004636120982468128, + 0.005172996316105127, + 0.004761840216815472, + 0.006561916787177324, + 0.007188388612121344, + 0.013125704601407051, + 0.010570998303592205, + 0.015454888343811035, + 0.011439983732998371, + 0.009765123017132282, + 0.007389953825622797, + 0.008937950246036053, + 0.007295671384781599, + 0.019251562654972076, + 0.008012959733605385, + 0.02230345644056797, + 0.009094745852053165, + 0.010424353182315826, + 0.007118385285139084, + 0.007458376698195934, + 0.006716935429722071, + 0.006593538913875818, + 0.0045999097637832165, + 0.005465286318212748, + 0.0064019314013421535, + 0.007462012115865946, + 0.007359884213656187, + 0.006098550278693438, + 0.005631280597299337, + 0.012634984217584133, + 0.005605571437627077, + 0.007019607815891504, + 0.007085761521011591, + 0.005055755376815796, + 0.0101791862398386, + 0.004391505382955074, + 0.006134362425655127, + 0.006772483699023724, + 0.004552079364657402, + 0.0071356515400111675, + 0.005593345500528812, + 0.007318022195249796, + 0.008821605704724789, + 0.0047205532900989056, + 0.006172054912894964, + 0.00648379884660244, + 0.003317714435979724, + 0.004494651220738888, + 0.0054818266071379185, + 0.004765818826854229, + 0.0049843089655041695, + 0.0039024786092340946, + 0.003150414675474167, + 0.01395502034574747, + 0.008216873742640018, + 0.01316752191632986, + 0.013579217717051506, + 0.013881977647542953, + 0.008243268355727196, + 0.012834400869905949, + 0.012086628936231136, + 0.06344195455312729, + 0.032184939831495285, + 0.05629131570458412, + 0.03403499349951744, + 0.010847457684576511, + 0.02182680368423462, + 0.00728064589202404, + 0.009716369211673737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.006936872377991676 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4746796488761902 + ], + "opposite_logprob_tokens": [ + -0.9746796488761902 + ], + "logprob_reward": -0.4746796488761902, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4746796488761902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29846489429473877, + 0.12819018959999084, + 0.08792311698198318, + 0.0672093853354454, + 0.05404051020741463, + 0.04486503824591637, + 0.029225070029497147, + 0.035647425800561905, + 0.05278775468468666, + 0.20164664089679718 + ], + "credit_map": [ + 0.03737134858965874, + 0.03832218423485756, + 0.05053592473268509, + 0.046772319823503494, + 0.026995670050382614, + 0.040799353271722794, + 0.028119029477238655, + 0.02954906038939953, + 0.027842892333865166, + 0.03570648655295372, + 0.009959686547517776, + 0.01090762298554182, + 0.01357915811240673, + 0.010123242624104023, + 0.010214818641543388, + 0.009856275282800198, + 0.010927954688668251, + 0.015971165150403976, + 0.010445165447890759, + 0.01535973884165287, + 0.008403308689594269, + 0.008068298920989037, + 0.009946187026798725, + 0.008801290765404701, + 0.008263488300144672, + 0.008846084587275982, + 0.009284082800149918, + 0.011328832246363163, + 0.008787436410784721, + 0.006322473753243685, + 0.0061514307744801044, + 0.008225561119616032, + 0.008470998145639896, + 0.005635371897369623, + 0.006149836350232363, + 0.0056027742102742195, + 0.009079749695956707, + 0.008636850863695145, + 0.004816044121980667, + 0.005648886784911156, + 0.009406495839357376, + 0.004000762943178415, + 0.006251677870750427, + 0.007333447691053152, + 0.0037689877208322287, + 0.00712943309918046, + 0.00336627708747983, + 0.003607950871810317, + 0.003853710601106286, + 0.002158022951334715, + 0.003975143190473318, + 0.0039316024631261826, + 0.00606609508395195, + 0.0023584789596498013, + 0.003819952718913555, + 0.003062063129618764, + 0.006606920622289181, + 0.0037696470972150564, + 0.004985147155821323, + 0.005485811270773411, + 0.004246321972459555, + 0.0037916116416454315, + 0.0034881052561104298, + 0.0032738593872636557, + 0.006728138774633408, + 0.00488147372379899, + 0.006341399624943733, + 0.007804065477102995, + 0.005629441700875759, + 0.010404856875538826, + 0.004819180350750685, + 0.006179204210639, + 0.047139398753643036, + 0.02524845488369465, + 0.04759211465716362, + 0.02858288586139679, + 0.009644425474107265, + 0.024235373362898827, + 0.00795792043209076, + 0.011246071197092533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.4746796488761902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909671753644943 + ], + "opposite_logprob_tokens": [ + -2.464096784591675 + ], + "logprob_reward": -0.08909671753644943, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909671753644943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06698112189769745, + 0.03847160562872887, + 0.07349500805139542, + 0.08491424471139908, + 0.051844432950019836, + 0.04063529521226883, + 0.023123564198613167, + 0.029601382091641426, + 0.05404496192932129, + 0.5368883609771729 + ], + "credit_map": [ + 0.00788506492972374, + 0.00882969330996275, + 0.01176032330840826, + 0.007623712997883558, + 0.006823649164289236, + 0.00936051644384861, + 0.009600595571100712, + 0.005097564775496721, + 0.009205475449562073, + 0.010084890760481358, + 0.004364629741758108, + 0.0025976288598030806, + 0.0024614690337330103, + 0.0030894035007804632, + 0.002956278156489134, + 0.0037118273321539164, + 0.007850351743400097, + 0.00948178768157959, + 0.011314231902360916, + 0.008265841752290726, + 0.00924638006836176, + 0.004940787795931101, + 0.013233453035354614, + 0.00916217640042305, + 0.019479183480143547, + 0.006327489856630564, + 0.023084072396159172, + 0.008138538338243961, + 0.006925185211002827, + 0.006816735025495291, + 0.007251086179167032, + 0.006891951430588961, + 0.0061595081351697445, + 0.005945228971540928, + 0.005530799273401499, + 0.0059905825182795525, + 0.006024056114256382, + 0.00948241725564003, + 0.006709841545671225, + 0.006001998204737902, + 0.009786231443285942, + 0.0037741484120488167, + 0.005959496833384037, + 0.004777370952069759, + 0.0033647737000137568, + 0.007749265059828758, + 0.002552113262936473, + 0.002671897877007723, + 0.0028066367376595736, + 0.001884614466689527, + 0.003931637387722731, + 0.003396013518795371, + 0.005003714002668858, + 0.002215804997831583, + 0.0018933911342173815, + 0.001991751603782177, + 0.006960027851164341, + 0.002954876748844981, + 0.003414343111217022, + 0.003821809310466051, + 0.0027450791094452143, + 0.005092501640319824, + 0.001970083685591817, + 0.0026426624972373247, + 0.007018320728093386, + 0.0044280835427343845, + 0.0071625541895627975, + 0.010380787774920464, + 0.005448420997709036, + 0.008197244256734848, + 0.004778886213898659, + 0.006630659103393555, + 0.15424852073192596, + 0.08018653839826584, + 0.14814814925193787, + 0.08446523547172546, + 0.01242360845208168, + 0.029728500172495842, + 0.012651695869863033, + 0.015036126598715782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.08909671753644943 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586861297488213 + ], + "opposite_logprob_tokens": [ + -3.8955869674682617 + ], + "logprob_reward": -0.020586861297488213, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586861297488213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16722099483013153, + 0.033559005707502365, + 0.06804437190294266, + 0.07711029052734375, + 0.05972829461097717, + 0.05316067487001419, + 0.03961711376905441, + 0.03937908634543419, + 0.18567371368408203, + 0.27650654315948486 + ], + "credit_map": [ + 0.022960493341088295, + 0.02549048699438572, + 0.03880295529961586, + 0.016860827803611755, + 0.008791535161435604, + 0.023472653701901436, + 0.014233434572815895, + 0.01660858653485775, + 0.003455939469859004, + 0.004946381784975529, + 0.005777629557996988, + 0.003664272604510188, + 0.004599167034029961, + 0.0032210606150329113, + 0.004614195786416531, + 0.003280355827882886, + 0.00818689540028572, + 0.02026614360511303, + 0.015315869823098183, + 0.004747622646391392, + 0.005063578020781279, + 0.004482581280171871, + 0.005158757325261831, + 0.004822916351258755, + 0.010976580902934074, + 0.008297101594507694, + 0.011615433730185032, + 0.0113326171413064, + 0.012092393822968006, + 0.008181605488061905, + 0.007883535698056221, + 0.006731011904776096, + 0.010990587994456291, + 0.005745801609009504, + 0.006921096704900265, + 0.005240709986537695, + 0.00616602785885334, + 0.01036710012704134, + 0.007461828645318747, + 0.006835137959569693, + 0.012735337018966675, + 0.004961259663105011, + 0.007897872477769852, + 0.0064752427861094475, + 0.004120992962270975, + 0.01012494694441557, + 0.003566889325156808, + 0.0032781288027763367, + 0.005081117153167725, + 0.0036308420822024345, + 0.0065049841068685055, + 0.006375824101269245, + 0.008086121641099453, + 0.004310182295739651, + 0.0025837935972958803, + 0.0030442485585808754, + 0.009805427864193916, + 0.0038342808838933706, + 0.004393081180751324, + 0.005362557712942362, + 0.004088687244802713, + 0.006077767349779606, + 0.00293621770106256, + 0.002881064312532544, + 0.027268756181001663, + 0.016768377274274826, + 0.025502508506178856, + 0.030643874779343605, + 0.01823205128312111, + 0.03087356500327587, + 0.01503511518239975, + 0.02134946547448635, + 0.08386024832725525, + 0.0433460995554924, + 0.07477725297212601, + 0.04224146902561188, + 0.006034661550074816, + 0.012415739707648754, + 0.005873177200555801, + 0.007957855239510536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.020586861297488213 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017143089789897203 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.0017143089789897203, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017143089789897203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13616348803043365, + 0.06581691652536392, + 0.0829075500369072, + 0.025858769193291664, + 0.020515963435173035, + 0.06068605184555054, + 0.019096294417977333, + 0.040996991097927094, + 0.12979379296302795, + 0.4181641638278961 + ], + "credit_map": [ + 0.015072797425091267, + 0.018763933330774307, + 0.02635846845805645, + 0.02178478240966797, + 0.01075727492570877, + 0.014033682644367218, + 0.0123610133305192, + 0.017031539231538773, + 0.016906220465898514, + 0.0225972943007946, + 0.005588860251009464, + 0.005067095160484314, + 0.00320120295509696, + 0.0051500010304152966, + 0.003939618822187185, + 0.003366625402122736, + 0.0062043028883636, + 0.02079911157488823, + 0.028012678027153015, + 0.007012400776147842, + 0.0038614748045802116, + 0.004678739700466394, + 0.006110813468694687, + 0.006228033918887377, + 0.003145703347399831, + 0.0023468760773539543, + 0.004338719882071018, + 0.003492045681923628, + 0.0032430796418339014, + 0.002926531946286559, + 0.003090063575655222, + 0.003275747876614332, + 0.0023133177310228348, + 0.001989668933674693, + 0.002500563859939575, + 0.0025140487123280764, + 0.0021332174073904753, + 0.0034905842039734125, + 0.0030254004523158073, + 0.002549161436036229, + 0.015798531472682953, + 0.0066309780813753605, + 0.00793689489364624, + 0.00451882416382432, + 0.00451512448489666, + 0.013758947141468525, + 0.0033324274700134993, + 0.004194323439151049, + 0.00311519717797637, + 0.001854072674177587, + 0.003349941922351718, + 0.00488825561478734, + 0.0016872241394594312, + 0.0018413492944091558, + 0.0011457150103524327, + 0.00121453742031008, + 0.007467137649655342, + 0.0029729255475103855, + 0.0071376352570950985, + 0.008718978613615036, + 0.0028927160892635584, + 0.004687600303441286, + 0.004002386704087257, + 0.0031176090706139803, + 0.02457866445183754, + 0.011889300309121609, + 0.028567073866724968, + 0.020917052403092384, + 0.009296615608036518, + 0.014388068579137325, + 0.00880480371415615, + 0.011352217756211758, + 0.12866562604904175, + 0.06713281571865082, + 0.11335644870996475, + 0.06299344450235367, + 0.00889196339994669, + 0.01732933521270752, + 0.009327908046543598, + 0.010466620326042175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0017143089789897203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a low, upward angle, slowly and steadily zooms in on a white-haired elderly man standing alone on the precipice of a precipitous mountain. His long, silvery hair and a simple gray robe made of rough linen flutter violently in the howling wind, yet his upright figure remains motionless, like a mountain, his feet firmly rooted to the rock beneath him, creating a striking contrast between movement and stillness. As the camera gradually moves in, his deeply wrinkled face and deep, ancient eyes, like ancient wells, gaze resolutely at the rolling, mist-shrouded mountains in the distance. Against the overcast sky, the cold, diffuse light accentuates his resolute silhouette, highlighting the folds of his robe and the creases on his face. The entire scene, shrouded in a cold, blue-gray hue, framed by a vast sky and blurred mountain ranges, creates a sense of epic loneliness and a timeless, indestructible, and quiet power.", + "seed": 593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_8899_seed_593_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10747425486333669, + "vlm_logprob": -0.10747425486333669, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.10747425486333669, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011522294022142887 + ], + "opposite_logprob_tokens": [ + -4.63652229309082 + ], + "logprob_reward": -0.011522294022142887, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011522294022142887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30069535970687866, + 0.05880846455693245, + 0.039189521223306656, + 0.03160905838012695, + 0.026502136141061783, + 0.040213968604803085, + 0.11691135168075562, + 0.10213494300842285, + 0.12922881543636322, + 0.15470634400844574 + ], + "credit_map": [ + 0.06413988023996353, + 0.07139550149440765, + 0.03437464311718941, + 0.037899430841207504, + 0.023678841069340706, + 0.013826276175677776, + 0.029844889417290688, + 0.025535909458994865, + 0.008019866421818733, + 0.007638968527317047, + 0.008868701756000519, + 0.007657253183424473, + 0.00737769203260541, + 0.005014473106712103, + 0.007665357086807489, + 0.006566154304891825, + 0.006285550072789192, + 0.005295634735375643, + 0.004486486781388521, + 0.004464513156563044, + 0.006830763071775436, + 0.003530119778588414, + 0.0045177447609603405, + 0.003778710961341858, + 0.003311829175800085, + 0.004295229446142912, + 0.004539413843303919, + 0.0038530295714735985, + 0.005100843030959368, + 0.0037665993440896273, + 0.004019148647785187, + 0.002722962060943246, + 0.0028763683512806892, + 0.0033472476061433554, + 0.0027739163488149643, + 0.0027155058924108744, + 0.004455518443137407, + 0.003803672967478633, + 0.0032783353235572577, + 0.0032515707425773144, + 0.003843330079689622, + 0.0042618680745363235, + 0.007284854538738728, + 0.006139950826764107, + 0.007229641545563936, + 0.0033156767021864653, + 0.004026707727462053, + 0.004111938178539276, + 0.01144537702202797, + 0.015294112265110016, + 0.01392209529876709, + 0.016753172501921654, + 0.027630746364593506, + 0.01147523708641529, + 0.01081258524209261, + 0.009578022174537182, + 0.009201621636748314, + 0.012175826355814934, + 0.013154109939932823, + 0.011753383092582226, + 0.025496402755379677, + 0.01011218037456274, + 0.01137912180274725, + 0.00886230356991291, + 0.014529637061059475, + 0.020692354068160057, + 0.01683187298476696, + 0.009924878366291523, + 0.02357439137995243, + 0.017563030123710632, + 0.01344545092433691, + 0.012667197734117508, + 0.012182064354419708, + 0.019776202738285065, + 0.015642113983631134, + 0.017640002071857452, + 0.026907579973340034, + 0.02387682907283306, + 0.01902683638036251, + 0.019654707983136177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.011522294022142887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28162121772766113 + ], + "opposite_logprob_tokens": [ + -1.4066212177276611 + ], + "logprob_reward": -0.28162121772766113, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28162121772766113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4443538188934326, + 0.1727648824453354, + 0.05388631671667099, + 0.033187054097652435, + 0.03841688856482506, + 0.04901425540447235, + 0.056259553879499435, + 0.030871011316776276, + 0.04979892447590828, + 0.07144729048013687 + ], + "credit_map": [ + 0.09049353003501892, + 0.10449094325304031, + 0.04758289456367493, + 0.0713607594370842, + 0.02870854362845421, + 0.02555784024298191, + 0.04287250339984894, + 0.033286795020103455, + 0.0260199923068285, + 0.02151613123714924, + 0.03921566903591156, + 0.029276693239808083, + 0.015360710211098194, + 0.009256463497877121, + 0.017941495403647423, + 0.014177718199789524, + 0.007723710034042597, + 0.006954064592719078, + 0.008745483122766018, + 0.009801756590604782, + 0.00685430783778429, + 0.003778506303206086, + 0.004986661486327648, + 0.005041822791099548, + 0.00499705458059907, + 0.004038343206048012, + 0.00513511011376977, + 0.005213823169469833, + 0.003342280862852931, + 0.0029162883292883635, + 0.0036262895446270704, + 0.003917861729860306, + 0.004875533282756805, + 0.004739413037896156, + 0.00712784705683589, + 0.006883327849209309, + 0.003234531031921506, + 0.003411374054849148, + 0.004124649800360203, + 0.004020211286842823, + 0.006627886090427637, + 0.006378269288688898, + 0.008550700731575489, + 0.008105716668069363, + 0.005667335353791714, + 0.00351913389749825, + 0.005011206958442926, + 0.00515400106087327, + 0.008216599933803082, + 0.0075948163866996765, + 0.011570143513381481, + 0.011441236361861229, + 0.005275186616927385, + 0.003913505934178829, + 0.0041169291362166405, + 0.004131130874156952, + 0.003975849132984877, + 0.003833176800981164, + 0.004493044689297676, + 0.005112437065690756, + 0.0043712458573281765, + 0.0015223403461277485, + 0.003713053883984685, + 0.003849860979244113, + 0.00563578587025404, + 0.006534997373819351, + 0.007938624359667301, + 0.005230220966041088, + 0.006442897487431765, + 0.007723163813352585, + 0.003936468623578548, + 0.0063567631877958775, + 0.008322050794959068, + 0.010646620765328407, + 0.006459991447627544, + 0.005909483879804611, + 0.016348781064152718, + 0.009501769207417965, + 0.006632742006331682, + 0.00762584712356329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.28162121772766113 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17996862530708313 + ], + "opposite_logprob_tokens": [ + -1.8049685955047607 + ], + "logprob_reward": -0.17996862530708313, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17996862530708313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35505467653274536, + 0.1035194993019104, + 0.04047273471951485, + 0.1221526637673378, + 0.05365902557969093, + 0.04343350976705551, + 0.0620366595685482, + 0.07909300923347473, + 0.06795471161603928, + 0.07262350618839264 + ], + "credit_map": [ + 0.0938062071800232, + 0.09587259590625763, + 0.033673517405986786, + 0.0351942703127861, + 0.029429800808429718, + 0.02064194343984127, + 0.024040885269641876, + 0.02239542454481125, + 0.013700502924621105, + 0.016529597342014313, + 0.01735394448041916, + 0.015841634944081306, + 0.011429604142904282, + 0.01039888709783554, + 0.00978992972522974, + 0.008475403301417828, + 0.003956162370741367, + 0.005007681902498007, + 0.004331517964601517, + 0.005235506687313318, + 0.007594551891088486, + 0.005330963525921106, + 0.004772172775119543, + 0.004244174342602491, + 0.018729645758867264, + 0.01692981831729412, + 0.014079547487199306, + 0.01356315053999424, + 0.011739225126802921, + 0.011776253581047058, + 0.017266470938920975, + 0.018068553879857063, + 0.006153489928692579, + 0.006747608073055744, + 0.006996907759457827, + 0.005967167671769857, + 0.008027677424252033, + 0.006479143630713224, + 0.005885064601898193, + 0.007401965092867613, + 0.00682640727609396, + 0.006480954121798277, + 0.005847033578902483, + 0.005641236901283264, + 0.00627036951482296, + 0.003579086624085903, + 0.004245958756655455, + 0.004542460199445486, + 0.007550566457211971, + 0.008076097816228867, + 0.011688335798680782, + 0.009754527360200882, + 0.009937076829373837, + 0.004867146722972393, + 0.005543709732592106, + 0.004619194194674492, + 0.00891290046274662, + 0.008744279853999615, + 0.012817205861210823, + 0.010876809246838093, + 0.011574617587029934, + 0.007491616066545248, + 0.01022920198738575, + 0.008446376770734787, + 0.007429535035043955, + 0.009727472439408302, + 0.009627110324800014, + 0.0065985689871013165, + 0.01590723730623722, + 0.006182018667459488, + 0.006203846540302038, + 0.006278919987380505, + 0.00673848669975996, + 0.009667089208960533, + 0.009532569907605648, + 0.010522802360355854, + 0.012016156688332558, + 0.008253704756498337, + 0.0060383351519703865, + 0.00985436700284481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.17996862530708313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205892935395241 + ], + "opposite_logprob_tokens": [ + -2.812058925628662 + ], + "logprob_reward": -0.06205892935395241, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205892935395241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20252688229084015, + 0.07601714134216309, + 0.0499361976981163, + 0.14771606028079987, + 0.02787589281797409, + 0.0696147233247757, + 0.07150035351514816, + 0.07696356624364853, + 0.031453609466552734, + 0.24639557301998138 + ], + "credit_map": [ + 0.03401154279708862, + 0.029227502644062042, + 0.03082956187427044, + 0.03898808732628822, + 0.01753082312643528, + 0.016957107931375504, + 0.02008327841758728, + 0.014898974448442459, + 0.007819735445082188, + 0.008966512978076935, + 0.011737700551748276, + 0.009601341560482979, + 0.010435327887535095, + 0.009109387174248695, + 0.009747808799147606, + 0.008599333465099335, + 0.005254149436950684, + 0.006512653082609177, + 0.005712037906050682, + 0.0064523546025156975, + 0.008430263958871365, + 0.006652681156992912, + 0.005347426515072584, + 0.005574631504714489, + 0.0228581465780735, + 0.02310449816286564, + 0.014917964115738869, + 0.01847575418651104, + 0.01645522378385067, + 0.022670304402709007, + 0.01562725380063057, + 0.013606918044388294, + 0.0024717424530535936, + 0.0036235135048627853, + 0.0036360565572977066, + 0.003268828149884939, + 0.004881874658167362, + 0.0031484293285757303, + 0.0034992636647075415, + 0.0033461821731179953, + 0.008284268900752068, + 0.00923750177025795, + 0.01183506939560175, + 0.010186261497437954, + 0.01056253258138895, + 0.0075501021929085255, + 0.005545991938561201, + 0.006412993650883436, + 0.010363548994064331, + 0.00975541491061449, + 0.009859327226877213, + 0.006919822655618191, + 0.012517803348600864, + 0.008692662231624126, + 0.007615647744387388, + 0.005776126869022846, + 0.008457579649984837, + 0.009669139981269836, + 0.008443286642432213, + 0.009198477491736412, + 0.020574351772665977, + 0.007276271469891071, + 0.006342214532196522, + 0.00700224656611681, + 0.0040602898225188255, + 0.0045788162387907505, + 0.0047720144502818584, + 0.003107826691120863, + 0.003962745424360037, + 0.0027736974880099297, + 0.0033022593706846237, + 0.004895963240414858, + 0.023560838773846626, + 0.03296293690800667, + 0.02741478942334652, + 0.031873784959316254, + 0.04708768427371979, + 0.024063002318143845, + 0.026152299717068672, + 0.03328024595975876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.06205892935395241 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200207905843854 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.002200207905843854, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200207905843854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5247846841812134, + 0.10863552242517471, + 0.04892579838633537, + 0.06685857474803925, + 0.04711015149950981, + 0.02979009971022606, + 0.06319555640220642, + 0.03197088465094566, + 0.04057617858052254, + 0.038152579218149185 + ], + "credit_map": [ + 0.09111427515745163, + 0.08528317511081696, + 0.08226432651281357, + 0.11534854024648666, + 0.03431741148233414, + 0.03789873421192169, + 0.04367649182677269, + 0.034881699830293655, + 0.014123715460300446, + 0.014025510288774967, + 0.01918911002576351, + 0.01780509762465954, + 0.009600279852747917, + 0.007891061715781689, + 0.013015738688409328, + 0.012985012494027615, + 0.005529382731765509, + 0.006218678317964077, + 0.007701787166297436, + 0.009553268551826477, + 0.006982352584600449, + 0.003785526379942894, + 0.004687368869781494, + 0.004467434249818325, + 0.00921721663326025, + 0.007232335861772299, + 0.012164664454758167, + 0.014946960844099522, + 0.004998304881155491, + 0.004974196664988995, + 0.006712545175105333, + 0.006612348835915327, + 0.007126461248844862, + 0.006312962621450424, + 0.009040679782629013, + 0.01133681833744049, + 0.0032733629923313856, + 0.002608669688925147, + 0.004013469442725182, + 0.003397726221010089, + 0.003865678794682026, + 0.004168021026998758, + 0.005835683085024357, + 0.004959754645824432, + 0.003501882776618004, + 0.0021202657371759415, + 0.0024115247651934624, + 0.002927288645878434, + 0.009860486723482609, + 0.008365055546164513, + 0.013316374272108078, + 0.01341202761977911, + 0.00520612858235836, + 0.003328900318592787, + 0.0047098067589104176, + 0.004996780771762133, + 0.00319670420140028, + 0.0038837091997265816, + 0.0052277697250247, + 0.004738085903227329, + 0.005593718960881233, + 0.002887418493628502, + 0.0031655661296099424, + 0.003277912735939026, + 0.0045233601704239845, + 0.006177967414259911, + 0.0063675823621451855, + 0.004363032523542643, + 0.007425616029649973, + 0.003860395634546876, + 0.0035539702512323856, + 0.004304253030568361, + 0.004601095803081989, + 0.004836530890315771, + 0.003349756821990013, + 0.003072144463658333, + 0.010642608627676964, + 0.003681628266349435, + 0.003715038299560547, + 0.004253771621733904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.002200207905843854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a low, upward angle, slowly and steadily zooms in on a white-haired elderly man standing alone on the precipice of a precipitous mountain. His long, silvery hair and a simple gray robe made of rough linen flutter violently in the howling wind, yet his upright figure remains motionless, like a mountain, his feet firmly rooted to the rock beneath him, creating a striking contrast between movement and stillness. As the camera gradually moves in, his deeply wrinkled face and deep, ancient eyes, like ancient wells, gaze resolutely at the rolling, mist-shrouded mountains in the distance. Against the overcast sky, the cold, diffuse light accentuates his resolute silhouette, highlighting the folds of his robe and the creases on his face. The entire scene, shrouded in a cold, blue-gray hue, framed by a vast sky and blurred mountain ranges, creates a sense of epic loneliness and a timeless, indestructible, and quiet power.", + "seed": 594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_2/index_8899_seed_594_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1761352798435837, + "vlm_logprob": -0.1761352798435837, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.1761352798435837, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014765623025596142 + ], + "opposite_logprob_tokens": [ + -4.389765739440918 + ], + "logprob_reward": -0.014765623025596142, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014765623025596142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3017471730709076, + 0.10742262750864029, + 0.04657822847366333, + 0.06275555491447449, + 0.05233808979392052, + 0.04599824920296669, + 0.039450131356716156, + 0.048509687185287476, + 0.17266306281089783, + 0.12253718823194504 + ], + "credit_map": [ + 0.022964075207710266, + 0.038387637585401535, + 0.04660407453775406, + 0.03558550029993057, + 0.03261524811387062, + 0.046030886471271515, + 0.041003622114658356, + 0.03855612874031067, + 0.008640396408736706, + 0.013773816637694836, + 0.012036727741360664, + 0.01836460642516613, + 0.009104731492698193, + 0.010457186959683895, + 0.01954716071486473, + 0.015497993677854538, + 0.004089553374797106, + 0.006193316541612148, + 0.004267026204615831, + 0.006290807388722897, + 0.00489629665389657, + 0.006995536386966705, + 0.006712930276989937, + 0.007132764905691147, + 0.006534854415804148, + 0.005303061567246914, + 0.005349673796445131, + 0.004755244590342045, + 0.011873705312609673, + 0.014148973859846592, + 0.0074293771758675575, + 0.007360664661973715, + 0.005879157688468695, + 0.008600646629929543, + 0.005352281499654055, + 0.008162494748830795, + 0.005888472311198711, + 0.006377688609063625, + 0.006585371680557728, + 0.005491974297910929, + 0.004394446033984423, + 0.003638547845184803, + 0.00651169940829277, + 0.003972548525780439, + 0.007321520242840052, + 0.008234415203332901, + 0.004879848565906286, + 0.007045222446322441, + 0.004578816704452038, + 0.005692182574421167, + 0.005274017807096243, + 0.004843361210078001, + 0.005116459913551807, + 0.004288483876734972, + 0.004312836565077305, + 0.005343970842659473, + 0.004127774853259325, + 0.009115597233176231, + 0.005601166747510433, + 0.005773843266069889, + 0.006827151868492365, + 0.005637543275952339, + 0.004979775752872229, + 0.00644683325663209, + 0.013111500069499016, + 0.0218967255204916, + 0.025119950994849205, + 0.023313051089644432, + 0.021518375724554062, + 0.020996617153286934, + 0.025183316320180893, + 0.021523531526327133, + 0.00981876626610756, + 0.01772451587021351, + 0.017436103895306587, + 0.011892281472682953, + 0.016304776072502136, + 0.015016558580100536, + 0.016337715089321136, + 0.018006473779678345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.014765623025596142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38749393820762634 + ], + "opposite_logprob_tokens": [ + -1.1374939680099487 + ], + "logprob_reward": -0.38749393820762634, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38749393820762634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24901363253593445, + 0.10120989382266998, + 0.061508115381002426, + 0.26866334676742554, + 0.06583462655544281, + 0.06023675203323364, + 0.04645343869924545, + 0.02936767227947712, + 0.04989474266767502, + 0.06781784445047379 + ], + "credit_map": [ + 0.022431334480643272, + 0.04188872128725052, + 0.028758330270648003, + 0.03213241696357727, + 0.030883874744176865, + 0.04594799131155014, + 0.02508844994008541, + 0.021882515400648117, + 0.007992644794285297, + 0.0203142911195755, + 0.01103302650153637, + 0.014604946598410606, + 0.008416940458118916, + 0.009084356017410755, + 0.019589649513363838, + 0.01017404068261385, + 0.0064267548732459545, + 0.006142050493508577, + 0.0070207854732871056, + 0.010733446106314659, + 0.007144654635339975, + 0.009466111660003662, + 0.008427371270954609, + 0.006146944593638182, + 0.0106620192527771, + 0.008656487800180912, + 0.010596686974167824, + 0.010358806699514389, + 0.0959085151553154, + 0.09625670313835144, + 0.021252445876598358, + 0.014971701428294182, + 0.008671541698276997, + 0.010655397549271584, + 0.0074673006311059, + 0.00896404031664133, + 0.008539067581295967, + 0.007322358898818493, + 0.008176556788384914, + 0.0060383640229702, + 0.005691670812666416, + 0.00881546176970005, + 0.007345355581492186, + 0.007823733612895012, + 0.007256757467985153, + 0.007187179755419493, + 0.00968871358782053, + 0.006427880842238665, + 0.004969916306436062, + 0.00545836566016078, + 0.006470237392932177, + 0.00603328226134181, + 0.005231658462435007, + 0.005079312715679407, + 0.0052907755598425865, + 0.00791989266872406, + 0.0024353405460715294, + 0.003775263438001275, + 0.004084437154233456, + 0.0034175438340753317, + 0.003735105972737074, + 0.003516430500894785, + 0.00315113621763885, + 0.005252413917332888, + 0.00422601168975234, + 0.006823413074016571, + 0.0061136651784181595, + 0.0055171167477965355, + 0.005850988440215588, + 0.00654936395585537, + 0.005668435245752335, + 0.009145751595497131, + 0.007744042202830315, + 0.010999531485140324, + 0.009448601864278316, + 0.005991602782160044, + 0.008167499676346779, + 0.008609624579548836, + 0.006322056520730257, + 0.010534888133406639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.38749393820762634 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871708810329437 + ], + "opposite_logprob_tokens": [ + -1.1371709108352661 + ], + "logprob_reward": -0.3871708810329437, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871708810329437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23909536004066467, + 0.14079800248146057, + 0.05923353508114815, + 0.11329338699579239, + 0.07480542361736298, + 0.05030146241188049, + 0.06288956105709076, + 0.09507673233747482, + 0.08801223337650299, + 0.07649438083171844 + ], + "credit_map": [ + 0.02171844244003296, + 0.03975741192698479, + 0.03562365844845772, + 0.03008759208023548, + 0.02497064135968685, + 0.030978349968791008, + 0.02704886719584465, + 0.028910411521792412, + 0.012898308224976063, + 0.025063112378120422, + 0.01584169641137123, + 0.02497176080942154, + 0.016441840678453445, + 0.01855642907321453, + 0.014859896153211594, + 0.01216494757682085, + 0.005657714791595936, + 0.005163461435586214, + 0.004681549966335297, + 0.006131790112704039, + 0.008623831905424595, + 0.009837966412305832, + 0.00972476415336132, + 0.00941245723515749, + 0.005629521794617176, + 0.007401214446872473, + 0.006091057322919369, + 0.007895725779235363, + 0.03353557735681534, + 0.03346302732825279, + 0.011317805387079716, + 0.007959447801113129, + 0.007013570982962847, + 0.0069122472777962685, + 0.006938157603144646, + 0.008696356788277626, + 0.009850560687482357, + 0.0095223318785429, + 0.013488594442605972, + 0.01238360907882452, + 0.0049439528957009315, + 0.007233091630041599, + 0.007732316851615906, + 0.008503086864948273, + 0.005409733392298222, + 0.004471681546419859, + 0.006506966892629862, + 0.0055006323382258415, + 0.006191850174218416, + 0.009129710495471954, + 0.009925290942192078, + 0.008474885486066341, + 0.006229445803910494, + 0.005904561374336481, + 0.00712163932621479, + 0.009912182576954365, + 0.009633715264499187, + 0.012619453482329845, + 0.01284314226359129, + 0.011321726255118847, + 0.012406058609485626, + 0.011586870066821575, + 0.012343768030405045, + 0.012322001159191132, + 0.006676636636257172, + 0.010417419485747814, + 0.011241908185184002, + 0.010717191733419895, + 0.011258366517722607, + 0.01083559449762106, + 0.009982814081013203, + 0.01688229665160179, + 0.005715426057577133, + 0.00995616801083088, + 0.01157449185848236, + 0.008756824769079685, + 0.009464836679399014, + 0.009405467659235, + 0.009172649122774601, + 0.012448515743017197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.3871708810329437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904622495174408 + ], + "opposite_logprob_tokens": [ + -2.4640462398529053 + ], + "logprob_reward": -0.08904622495174408, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904622495174408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11180289834737778, + 0.07501618564128876, + 0.06509599834680557, + 0.11941847205162048, + 0.028310755267739296, + 0.06595107167959213, + 0.028762733563780785, + 0.023622136563062668, + 0.30900654196739197, + 0.17301322519779205 + ], + "credit_map": [ + 0.009251261129975319, + 0.020462704822421074, + 0.014288094826042652, + 0.015959860756993294, + 0.010992796160280704, + 0.0158709567040205, + 0.011667060665786266, + 0.013310154899954796, + 0.006129085551947355, + 0.011419709771871567, + 0.006467399187386036, + 0.009151939302682877, + 0.008119313977658749, + 0.010424427688121796, + 0.013292876072227955, + 0.010011435486376286, + 0.004885214846581221, + 0.008531177416443825, + 0.006566679570823908, + 0.012439134530723095, + 0.006122209131717682, + 0.008324982598423958, + 0.010802497155964375, + 0.007424101699143648, + 0.005719630979001522, + 0.009575595147907734, + 0.004994361661374569, + 0.005398373119533062, + 0.04093385860323906, + 0.040744151920080185, + 0.006124061997979879, + 0.005928437691181898, + 0.0024654115550220013, + 0.0035084206610918045, + 0.003158766310662031, + 0.0037583084776997566, + 0.00364281446672976, + 0.003599028568714857, + 0.004245903808623552, + 0.003932103049010038, + 0.007207558024674654, + 0.007689106278121471, + 0.008399414829909801, + 0.007986399345099926, + 0.008646970614790916, + 0.007728257682174444, + 0.011582876555621624, + 0.006710493937134743, + 0.003520277561619878, + 0.0033750757575035095, + 0.0035448644775897264, + 0.003073298605158925, + 0.0028223933186382055, + 0.002923749154433608, + 0.003417429281398654, + 0.006085643079131842, + 0.0025896276347339153, + 0.002978831296786666, + 0.0022623203694820404, + 0.0023257494904100895, + 0.003329312661662698, + 0.0029408494010567665, + 0.002876673359423876, + 0.004318772349506617, + 0.024680854752659798, + 0.03710734844207764, + 0.04419621452689171, + 0.041222378611564636, + 0.04637788236141205, + 0.03841811791062355, + 0.04254469275474548, + 0.03445905074477196, + 0.010643589310348034, + 0.024577831849455833, + 0.026971107348799706, + 0.018282877281308174, + 0.026301641017198563, + 0.022875068709254265, + 0.01920708641409874, + 0.02415401116013527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.08904622495174408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium-long shot featuring an elderly man with long, silvery hair wearing a simple gray robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man standing on the precipice of a steep mountain with his feet firmly rooted to the rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's robe flutter violently in the wind while his upright figure remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low, upward angle and slowly zooming in on the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a vast sky, blurred mountain ranges, and an overcast sky with a cold, blue-gray hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021997320000082254 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021997320000082254, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021997320000082254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1495494395494461, + 0.2600754201412201, + 0.08237587660551071, + 0.28295350074768066, + 0.0175323449075222, + 0.027112679556012154, + 0.028616564348340034, + 0.022475946694612503, + 0.09920302033424377, + 0.030105210840702057 + ], + "credit_map": [ + 0.010137676261365414, + 0.016244368627667427, + 0.02493087574839592, + 0.015609846450388432, + 0.01855483092367649, + 0.025603102520108223, + 0.017246035858988762, + 0.02122270129621029, + 0.024020295590162277, + 0.013937964104115963, + 0.01532186008989811, + 0.014097857289016247, + 0.04087204858660698, + 0.07357490062713623, + 0.03303896263241768, + 0.0452115535736084, + 0.008427892811596394, + 0.005522699560970068, + 0.005211982410401106, + 0.006854288280010223, + 0.014306250959634781, + 0.02359023690223694, + 0.009258813224732876, + 0.00920371524989605, + 0.008732838556170464, + 0.004242043010890484, + 0.00580979697406292, + 0.004225715063512325, + 0.12398238480091095, + 0.12012864649295807, + 0.007250318769365549, + 0.008581746369600296, + 0.0017756817396730185, + 0.002004543785005808, + 0.002044437685981393, + 0.002478082198649645, + 0.0021679382771253586, + 0.0015921375015750527, + 0.0029926947318017483, + 0.002476828871294856, + 0.003340840106830001, + 0.003827357431873679, + 0.003192390315234661, + 0.004152873996645212, + 0.0032936017960309982, + 0.002887948416173458, + 0.0037571904249489307, + 0.0026604763697832823, + 0.002972890855744481, + 0.0024038844276219606, + 0.0031382464803755283, + 0.002725812839344144, + 0.002397888107225299, + 0.0030514425598084927, + 0.003620260627940297, + 0.008306137286126614, + 0.0023193175438791513, + 0.003057244699448347, + 0.0030749139841645956, + 0.0028848128858953714, + 0.002449565101414919, + 0.0025878259912133217, + 0.0028898552991449833, + 0.0032124114222824574, + 0.007942491210997105, + 0.009226429276168346, + 0.01118962187319994, + 0.009624599479138851, + 0.010288388468325138, + 0.013352170586585999, + 0.011554665863513947, + 0.026024654507637024, + 0.002596947830170393, + 0.004411091096699238, + 0.0048317257314920425, + 0.0034160283394157887, + 0.002798213856294751, + 0.003358623245730996, + 0.0031918236054480076, + 0.005500758532434702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0021997320000082254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_tensor([7012, 7012])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_tensor([7012, 7012])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9ed5d97354a3a517319b52d41b2ef4a097d8dd4c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_tensor([7012, 7012])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cyberpunk-inspired extreme macro shot plunges the viewer directly into a dark, viscous synthetic fluid, the camera seemingly swept by a powerful electric current. Within this inky black liquid, countless glowing turquoise filaments, resembling synthetic algae, serve as the sole source of light. Each filament possesses a translucent, fibrous texture, its internal structure clearly visible as a complex double-helix spiral of bioluminescent data. Microscopic particles flow rapidly along these pathways like digital information. These threads are caught in a violent and turbulent flow, whipping and twisting erratically, colliding and tangling in a chaotic dance. They momentarily tense, then snap back, forming fleeting, intricate braids before being instantly ripped apart by a powerful, unseen force. Intense green light casts sharp, volumetric rays into the murky liquid, illuminating suspended microscopic particles and creating a dramatic interplay of light and deep shadows. The scene, dominated by cool electric green tones and abyssal blacks, creates a mesmerizing yet unsettling atmosphere of high-tech, microscopic chaos—a beautiful yet violent spectacle of life and data intertwined.", + "seed": 24122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_7012_seed_24122_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05372512557078153, + "vlm_logprob": -0.05372512557078153, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.05372512557078153, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013570320792496204 + ], + "opposite_logprob_tokens": [ + -4.513570308685303 + ], + "logprob_reward": -0.013570320792496204, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013570320792496204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13274694979190826, + 0.05878794938325882, + 0.06523362547159195, + 0.10206537693738937, + 0.17407263815402985, + 0.03416614234447479, + 0.13049793243408203, + 0.059261009097099304, + 0.14685188233852386, + 0.09631654620170593 + ], + "credit_map": [ + 0.013918278738856316, + 0.02207302674651146, + 0.010837179608643055, + 0.011396176181733608, + 0.014807086437940598, + 0.018783362582325935, + 0.021642161533236504, + 0.019289661198854446, + 0.006188059225678444, + 0.009465279057621956, + 0.007952735759317875, + 0.008354765363037586, + 0.00529783871024847, + 0.005443565081804991, + 0.007613249588757753, + 0.008472456596791744, + 0.0068592047318816185, + 0.008958559483289719, + 0.009959856979548931, + 0.007602646015584469, + 0.009257270954549313, + 0.01004001870751381, + 0.00650441600009799, + 0.006051639560610056, + 0.012860559858381748, + 0.02089853771030903, + 0.005358404014259577, + 0.004796088673174381, + 0.010886583477258682, + 0.009434984996914864, + 0.013150008395314217, + 0.02468019910156727, + 0.019701950252056122, + 0.022388527169823647, + 0.020636560395359993, + 0.013157868757843971, + 0.021352488547563553, + 0.03149819374084473, + 0.02245311811566353, + 0.02288392186164856, + 0.003721376648172736, + 0.004007361829280853, + 0.0038577262312173843, + 0.0023736213333904743, + 0.004652415867894888, + 0.0058236378245055676, + 0.004143238998949528, + 0.0055867587216198444, + 0.010756926611065865, + 0.012543083168566227, + 0.02007407695055008, + 0.013197376392781734, + 0.01569538563489914, + 0.019758742302656174, + 0.019584445282816887, + 0.018887870013713837, + 0.004998711869120598, + 0.007172188255935907, + 0.0087380176410079, + 0.006633201614022255, + 0.01062967348843813, + 0.007222185377031565, + 0.0060303108766674995, + 0.007836710661649704, + 0.011438778601586819, + 0.015413457527756691, + 0.014820294454693794, + 0.011333859525620937, + 0.015991071239113808, + 0.019465481862425804, + 0.027375513687729836, + 0.03101341798901558, + 0.005856985226273537, + 0.006457862444221973, + 0.008749001659452915, + 0.005712649319320917, + 0.008210568688809872, + 0.02930685319006443, + 0.011918243020772934, + 0.02010437473654747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.013570320792496204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013517538318410516 + ], + "opposite_logprob_tokens": [ + -6.626351833343506 + ], + "logprob_reward": -0.0013517538318410516, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013517538318410516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11066780239343643, + 0.11878174543380737, + 0.06270752847194672, + 0.10245184600353241, + 0.07825782895088196, + 0.1035047173500061, + 0.11513765901327133, + 0.11402028799057007, + 0.11897087097167969, + 0.0754997581243515 + ], + "credit_map": [ + 0.011882971040904522, + 0.014489805325865746, + 0.010327592492103577, + 0.013109092600643635, + 0.014409626834094524, + 0.014527032151818275, + 0.015753112733364105, + 0.01616855524480343, + 0.013414609245955944, + 0.015173682011663914, + 0.01692822016775608, + 0.01149456761777401, + 0.01453846413642168, + 0.013684016652405262, + 0.018383445218205452, + 0.015164738520979881, + 0.007185700349509716, + 0.008025012910366058, + 0.006555892992764711, + 0.0051750605925917625, + 0.00944148376584053, + 0.008432176895439625, + 0.00824312400072813, + 0.009649072773754597, + 0.008750553242862225, + 0.017743529751896858, + 0.0070118107832968235, + 0.009778755716979504, + 0.016479989513754845, + 0.013630234636366367, + 0.014472053386271, + 0.014584909193217754, + 0.010262430645525455, + 0.007933489046990871, + 0.008642393164336681, + 0.005119931884109974, + 0.00926815252751112, + 0.01229957863688469, + 0.012373512610793114, + 0.012358329258859158, + 0.011915863491594791, + 0.012398355640470982, + 0.010169866494834423, + 0.009831428527832031, + 0.012946597300469875, + 0.010307816788554192, + 0.01534867100417614, + 0.020586110651493073, + 0.012230487540364265, + 0.009585928171873093, + 0.015407530590891838, + 0.00891867931932211, + 0.015336505137383938, + 0.01748979091644287, + 0.012543135322630405, + 0.02362559176981449, + 0.009420261718332767, + 0.015673406422138214, + 0.01751871593296528, + 0.01401270367205143, + 0.018463298678398132, + 0.01315805222839117, + 0.011386840604245663, + 0.014386999420821667, + 0.008184212259948254, + 0.008066206239163876, + 0.016979429870843887, + 0.008829168975353241, + 0.013242166489362717, + 0.019253674894571304, + 0.01822151429951191, + 0.026194484904408455, + 0.004807173740118742, + 0.005969139747321606, + 0.008813208900392056, + 0.005759358871728182, + 0.008733715862035751, + 0.01780993863940239, + 0.01186397671699524, + 0.011743233539164066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.0013517538318410516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604018211364746 + ], + "opposite_logprob_tokens": [ + -1.9104018211364746 + ], + "logprob_reward": -0.1604018211364746, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604018211364746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09841200709342957, + 0.13442882895469666, + 0.0438387393951416, + 0.12939485907554626, + 0.0970325917005539, + 0.09395626932382584, + 0.044866014271974564, + 0.13800185918807983, + 0.16409747302532196, + 0.055971380323171616 + ], + "credit_map": [ + 0.009468638338148594, + 0.015117788687348366, + 0.014743052423000336, + 0.00971392821520567, + 0.010960293933749199, + 0.010746289044618607, + 0.012710082344710827, + 0.014951927587389946, + 0.01195009145885706, + 0.020948175340890884, + 0.015928233042359352, + 0.014012740924954414, + 0.018881745636463165, + 0.016100706532597542, + 0.017975376918911934, + 0.01863175816833973, + 0.007338438183069229, + 0.00541627686470747, + 0.006884382572025061, + 0.00409708172082901, + 0.004769870080053806, + 0.005091862753033638, + 0.0052368734031915665, + 0.005003951955586672, + 0.013219776563346386, + 0.028560610488057137, + 0.007500725332647562, + 0.007689724210649729, + 0.01022476889193058, + 0.012555827386677265, + 0.015378917567431927, + 0.03426452353596687, + 0.010577253066003323, + 0.010944612324237823, + 0.010363348759710789, + 0.006075461395084858, + 0.012578250840306282, + 0.018773358315229416, + 0.013866826891899109, + 0.013853471726179123, + 0.01102439220994711, + 0.011889666318893433, + 0.01058912929147482, + 0.009283732622861862, + 0.009924105368554592, + 0.009025481529533863, + 0.014401499181985855, + 0.0178182665258646, + 0.0034806085750460625, + 0.003984179347753525, + 0.004546192940324545, + 0.00404230784624815, + 0.005444194655865431, + 0.01001797802746296, + 0.006371734198182821, + 0.006978820078074932, + 0.011166973039507866, + 0.02068319171667099, + 0.014638721011579037, + 0.011859362944960594, + 0.02078239619731903, + 0.025103172287344933, + 0.013989382423460484, + 0.019778667017817497, + 0.012307214550673962, + 0.018870137631893158, + 0.01657397486269474, + 0.013775506988167763, + 0.01858610287308693, + 0.02130804769694805, + 0.028243444859981537, + 0.034433040767908096, + 0.003261495614424348, + 0.0034633369650691748, + 0.005957139190286398, + 0.004096639342606068, + 0.0042559257708489895, + 0.014381934888660908, + 0.009595192037522793, + 0.010959720239043236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.1604018211364746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284580558538437 + ], + "opposite_logprob_tokens": [ + -3.773284673690796 + ], + "logprob_reward": -0.023284580558538437, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284580558538437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1793193519115448, + 0.1265677958726883, + 0.04268142580986023, + 0.04986957088112831, + 0.1277269423007965, + 0.08627372235059738, + 0.08522927016019821, + 0.061161402612924576, + 0.19181962311267853, + 0.049350887537002563 + ], + "credit_map": [ + 0.016876062378287315, + 0.030361158773303032, + 0.020148873329162598, + 0.018380938097834587, + 0.02771630510687828, + 0.022181393578648567, + 0.021055538207292557, + 0.022599080577492714, + 0.012937377206981182, + 0.019339127466082573, + 0.018131297081708908, + 0.013238121755421162, + 0.0156778022646904, + 0.01296786405146122, + 0.018718566745519638, + 0.015557640232145786, + 0.005775973666459322, + 0.006201117765158415, + 0.004963396117091179, + 0.005021392367780209, + 0.005508052185177803, + 0.004594072699546814, + 0.004046441987156868, + 0.0065709808841347694, + 0.004367698449641466, + 0.006977061741054058, + 0.004544153343886137, + 0.004995717667043209, + 0.007215721067041159, + 0.005528704263269901, + 0.006799767725169659, + 0.00944074708968401, + 0.013443954288959503, + 0.012442778795957565, + 0.017777366563677788, + 0.013932825066149235, + 0.01409419160336256, + 0.016509518027305603, + 0.021396208554506302, + 0.01813010685145855, + 0.010523497126996517, + 0.01040404662489891, + 0.009691163897514343, + 0.006321452092379332, + 0.00978102907538414, + 0.012955470941960812, + 0.009187689051032066, + 0.017409373074769974, + 0.008262869901955128, + 0.006320387125015259, + 0.008756506256759167, + 0.005790878087282181, + 0.010998464189469814, + 0.018392043188214302, + 0.008449571207165718, + 0.018258554860949516, + 0.004957054276019335, + 0.008015526458621025, + 0.006446166429668665, + 0.004623739514499903, + 0.009077295660972595, + 0.012487255968153477, + 0.006090565584599972, + 0.009463795460760593, + 0.012915556319057941, + 0.019393300637602806, + 0.02280871942639351, + 0.015628160908818245, + 0.020100077614188194, + 0.0346902459859848, + 0.03989911079406738, + 0.02638445980846882, + 0.0031158779747784138, + 0.0034746117889881134, + 0.0038801629561930895, + 0.003893036162480712, + 0.0049046715721488, + 0.013236204162240028, + 0.009822203777730465, + 0.00702411774545908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.023284580558538437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001715153455734 + ], + "opposite_logprob_tokens": [ + -2.695017099380493 + ], + "logprob_reward": -0.07001715153455734, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001715153455734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09432822465896606, + 0.17948468029499054, + 0.06661176681518555, + 0.28544652462005615, + 0.04949699714779854, + 0.08388882130384445, + 0.06927581131458282, + 0.04051296412944794, + 0.08505374193191528, + 0.0459003821015358 + ], + "credit_map": [ + 0.008424486964941025, + 0.015002348460257053, + 0.009042634628713131, + 0.01324333343654871, + 0.015550190582871437, + 0.011346817947924137, + 0.008755500428378582, + 0.01296291034668684, + 0.02330206148326397, + 0.020437175408005714, + 0.02641007862985134, + 0.022845741361379623, + 0.019402585923671722, + 0.022014189511537552, + 0.021937746554613113, + 0.023135105147957802, + 0.011130962520837784, + 0.007888402789831161, + 0.006089893635362387, + 0.0041565303690731525, + 0.011832009069621563, + 0.00845171045511961, + 0.008345946669578552, + 0.008716311305761337, + 0.028860026970505714, + 0.06999615579843521, + 0.014698888175189495, + 0.017564035952091217, + 0.022621920332312584, + 0.028879491612315178, + 0.029680941253900528, + 0.0731450691819191, + 0.005048599559813738, + 0.005796445067971945, + 0.005086344201117754, + 0.005112857092171907, + 0.006204855162650347, + 0.007487575523555279, + 0.008771509863436222, + 0.005988813936710358, + 0.009600906632840633, + 0.011698112823069096, + 0.0100939329713583, + 0.00811330508440733, + 0.011695552617311478, + 0.00898120366036892, + 0.009794743731617928, + 0.013911069370806217, + 0.006166268140077591, + 0.004216516390442848, + 0.007614403963088989, + 0.004985586740076542, + 0.007365156430751085, + 0.018285540863871574, + 0.010897868312895298, + 0.009744471870362759, + 0.0035753254778683186, + 0.006415119394659996, + 0.0034820614382624626, + 0.005042232573032379, + 0.003996903542429209, + 0.007668020203709602, + 0.003883439116179943, + 0.006449863780289888, + 0.005597978830337524, + 0.005385658238083124, + 0.0127464160323143, + 0.00552726723253727, + 0.009546449407935143, + 0.01394492108374834, + 0.01447796355932951, + 0.01782708615064621, + 0.003578929929062724, + 0.003255855990573764, + 0.005229364149272442, + 0.0040817223489284515, + 0.0059203170239925385, + 0.007846708409488201, + 0.00941993948072195, + 0.006567548960447311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.07001715153455734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cyberpunk-inspired extreme macro shot plunges the viewer directly into a dark, viscous synthetic fluid, the camera seemingly swept by a powerful electric current. Within this inky black liquid, countless glowing turquoise filaments, resembling synthetic algae, serve as the sole source of light. Each filament possesses a translucent, fibrous texture, its internal structure clearly visible as a complex double-helix spiral of bioluminescent data. Microscopic particles flow rapidly along these pathways like digital information. These threads are caught in a violent and turbulent flow, whipping and twisting erratically, colliding and tangling in a chaotic dance. They momentarily tense, then snap back, forming fleeting, intricate braids before being instantly ripped apart by a powerful, unseen force. Intense green light casts sharp, volumetric rays into the murky liquid, illuminating suspended microscopic particles and creating a dramatic interplay of light and deep shadows. The scene, dominated by cool electric green tones and abyssal blacks, creates a mesmerizing yet unsettling atmosphere of high-tech, microscopic chaos—a beautiful yet violent spectacle of life and data intertwined.", + "seed": 24123, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_7012_seed_24123_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03369004593696445, + "vlm_logprob": -0.03369004593696445, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03369004593696445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02583436854183674 + ], + "opposite_logprob_tokens": [ + -3.900834321975708 + ], + "logprob_reward": -0.02583436854183674, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02583436854183674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3202502429485321, + 0.1808115541934967, + 0.03681644797325134, + 0.03842330724000931, + 0.0822555348277092, + 0.042080074548721313, + 0.06691466271877289, + 0.05486879125237465, + 0.09240853041410446, + 0.08517076820135117 + ], + "credit_map": [ + 0.021967018023133278, + 0.02696879580616951, + 0.05040939152240753, + 0.04611837863922119, + 0.027649199590086937, + 0.04112714156508446, + 0.039071034640073776, + 0.0669393315911293, + 0.006468813866376877, + 0.009145786985754967, + 0.01404094509780407, + 0.03040497377514839, + 0.007576618809252977, + 0.02706005796790123, + 0.03626283258199692, + 0.04985153675079346, + 0.003322674660012126, + 0.0035171655472368, + 0.004648889414966106, + 0.007018963806331158, + 0.002907742979004979, + 0.004736544564366341, + 0.004178609699010849, + 0.006485858466476202, + 0.0032149055041372776, + 0.0027179738972336054, + 0.0041741253808140755, + 0.0046974713914096355, + 0.00627607386559248, + 0.006942862644791603, + 0.004832483362406492, + 0.005567416548728943, + 0.008632983081042767, + 0.009801681153476238, + 0.012700505554676056, + 0.010034804232418537, + 0.012835688889026642, + 0.009181932546198368, + 0.00834386795759201, + 0.010724078863859177, + 0.005174055229872465, + 0.007014617323875427, + 0.006695875432342291, + 0.003952302038669586, + 0.006327385548502207, + 0.00459969462826848, + 0.003468721406534314, + 0.004847426898777485, + 0.010914077050983906, + 0.007968951016664505, + 0.011075212620198727, + 0.006124273408204317, + 0.007540872786194086, + 0.006451285444200039, + 0.006704216822981834, + 0.010135777294635773, + 0.0038949837908148766, + 0.006598415318876505, + 0.006986715365201235, + 0.006919503211975098, + 0.008765879087150097, + 0.01004028506577015, + 0.00528954342007637, + 0.00637346925213933, + 0.008116821758449078, + 0.010954450815916061, + 0.02415616251528263, + 0.0124319177120924, + 0.006750096101313829, + 0.006074115633964539, + 0.009269583970308304, + 0.014655393548309803, + 0.004080713726580143, + 0.008177137933671474, + 0.02056676335632801, + 0.03365487977862358, + 0.0030865909066051245, + 0.004267432726919651, + 0.0071132187731564045, + 0.004224038682878017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02583436854183674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015298341168090701 + ], + "opposite_logprob_tokens": [ + -6.501529693603516 + ], + "logprob_reward": -0.0015298341168090701, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015298341168090701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1634998917579651, + 0.20250514149665833, + 0.0472121499478817, + 0.2527267336845398, + 0.05560826510190964, + 0.04455859586596489, + 0.0463290773332119, + 0.07443520426750183, + 0.044326744973659515, + 0.06879816949367523 + ], + "credit_map": [ + 0.009769157506525517, + 0.015021935105323792, + 0.025804288685321808, + 0.023791953921318054, + 0.013038711622357368, + 0.019329844042658806, + 0.021372398361563683, + 0.035371605306863785, + 0.008157234638929367, + 0.012065703049302101, + 0.01842985488474369, + 0.02163957618176937, + 0.008182534016668797, + 0.016821084544062614, + 0.07251495122909546, + 0.0446942076086998, + 0.007821341045200825, + 0.006000785157084465, + 0.005373350810259581, + 0.005300746764987707, + 0.004658855497837067, + 0.007175840903073549, + 0.003988319076597691, + 0.006892911158502102, + 0.009229482151567936, + 0.012392242439091206, + 0.007058864459395409, + 0.007815894670784473, + 0.10280783474445343, + 0.020255422219634056, + 0.009672630578279495, + 0.08349438011646271, + 0.007604900747537613, + 0.008839535526931286, + 0.00848817266523838, + 0.005675897933542728, + 0.007801279425621033, + 0.005657640751451254, + 0.004512458574026823, + 0.007028378080576658, + 0.006089902948588133, + 0.006525189615786076, + 0.008047099225223064, + 0.0035578415263444185, + 0.005515689961612225, + 0.00487921666353941, + 0.00468604126945138, + 0.005257614888250828, + 0.009383956901729107, + 0.005199587903916836, + 0.008896167390048504, + 0.0045656622387468815, + 0.005507538095116615, + 0.00492009986191988, + 0.002901407890021801, + 0.004954654257744551, + 0.008313675411045551, + 0.009420200251042843, + 0.00938386283814907, + 0.008343846537172794, + 0.01138579286634922, + 0.013438207097351551, + 0.006653936579823494, + 0.0074956840835511684, + 0.004458438605070114, + 0.007041791453957558, + 0.008861931972205639, + 0.0054122768342494965, + 0.004099145065993071, + 0.003703309455886483, + 0.005010625813156366, + 0.005739226937294006, + 0.004601635038852692, + 0.007079653907567263, + 0.01825726218521595, + 0.023905044421553612, + 0.003117993241176009, + 0.0026732124388217926, + 0.006246115546673536, + 0.002917249221354723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0015298341168090701 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497156083583832 + ], + "opposite_logprob_tokens": [ + -2.92997145652771 + ], + "logprob_reward": -0.05497156083583832, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497156083583832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23238423466682434, + 0.23011189699172974, + 0.027890902012586594, + 0.03197145462036133, + 0.047080669552087784, + 0.056994080543518066, + 0.09390120953321457, + 0.02732810750603676, + 0.036462780088186264, + 0.21587465703487396 + ], + "credit_map": [ + 0.013384662568569183, + 0.016349459066987038, + 0.03156799077987671, + 0.024743929505348206, + 0.027059298008680344, + 0.038095347583293915, + 0.03566425293684006, + 0.045519303530454636, + 0.004630719777196646, + 0.007985780946910381, + 0.01645798794925213, + 0.02534620091319084, + 0.007347020320594311, + 0.015878282487392426, + 0.09372510761022568, + 0.05874079465866089, + 0.003122287103906274, + 0.0022303080186247826, + 0.003955567255616188, + 0.004189305007457733, + 0.0029912437312304974, + 0.0045063067227602005, + 0.002673151670023799, + 0.004222731105983257, + 0.0044480436481535435, + 0.0022876302246004343, + 0.004253869876265526, + 0.0026638591662049294, + 0.004271777346730232, + 0.00624870415776968, + 0.0037876178976148367, + 0.0040099541656672955, + 0.006036645267158747, + 0.007690367754548788, + 0.006739167962223291, + 0.005038592964410782, + 0.006267883349210024, + 0.003803577506914735, + 0.00611503841355443, + 0.005389399826526642, + 0.007374432869255543, + 0.005699021741747856, + 0.008930284529924393, + 0.006002852693200111, + 0.008566232398152351, + 0.00840741116553545, + 0.0058733439072966576, + 0.006140498910099268, + 0.009573427960276604, + 0.011747041717171669, + 0.014570583589375019, + 0.009906502440571785, + 0.01235975231975317, + 0.0100638372823596, + 0.008337273262441158, + 0.01734279654920101, + 0.002467662561684847, + 0.003101949580013752, + 0.004162436816841364, + 0.004603512119501829, + 0.0030665898229926825, + 0.00370198511518538, + 0.003109489567577839, + 0.003114482620730996, + 0.004608502145856619, + 0.005606592632830143, + 0.00602142745628953, + 0.004810040351003408, + 0.003782909596338868, + 0.003746282309293747, + 0.0042990208603441715, + 0.0035880054347217083, + 0.009461745619773865, + 0.018076423555612564, + 0.05985107272863388, + 0.08943771570920944, + 0.007191333919763565, + 0.008321430534124374, + 0.016182944178581238, + 0.00735199311748147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.05497156083583832 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305323585867882 + ], + "opposite_logprob_tokens": [ + -3.168053150177002 + ], + "logprob_reward": -0.04305323585867882, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305323585867882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11842260509729385, + 0.2687349319458008, + 0.04226197674870491, + 0.13398391008377075, + 0.022957075387239456, + 0.07082075625658035, + 0.02637665346264839, + 0.05652344971895218, + 0.16454023122787476, + 0.09537846595048904 + ], + "credit_map": [ + 0.00619706604629755, + 0.00821251142770052, + 0.017181823030114174, + 0.011974979192018509, + 0.010908899828791618, + 0.01749109849333763, + 0.018578985705971718, + 0.02787722647190094, + 0.0062254369258880615, + 0.006230635568499565, + 0.02037304826080799, + 0.03305002301931381, + 0.00879704114049673, + 0.018286824226379395, + 0.11070622503757477, + 0.06506568938493729, + 0.003737781662493944, + 0.003992636688053608, + 0.00804483238607645, + 0.0068160369992256165, + 0.0031306352466344833, + 0.005836925003677607, + 0.004333827178925276, + 0.006369297858327627, + 0.004992412868887186, + 0.007109371479600668, + 0.0038583234418183565, + 0.004644297994673252, + 0.05362424626946449, + 0.010555416345596313, + 0.00595892034471035, + 0.043240923434495926, + 0.0029731718823313713, + 0.00290542165748775, + 0.003966005519032478, + 0.0026619937270879745, + 0.0029316223226487637, + 0.0021613689605146646, + 0.002291852841153741, + 0.0030656366143375635, + 0.009504277259111404, + 0.00946045108139515, + 0.01199738122522831, + 0.005448796320706606, + 0.009297200478613377, + 0.0069991303607821465, + 0.006637435406446457, + 0.011476079002022743, + 0.004850849974900484, + 0.0035797148011624813, + 0.005250940099358559, + 0.0028647223953157663, + 0.0033110210206359625, + 0.0022203160915523767, + 0.0014999286504462361, + 0.0027991575188934803, + 0.005825383588671684, + 0.007110975217074156, + 0.00766461156308651, + 0.008810805156826973, + 0.007505133748054504, + 0.008422919549047947, + 0.005121392197906971, + 0.006062224041670561, + 0.016582755371928215, + 0.01549347024410963, + 0.046260345727205276, + 0.01820438727736473, + 0.014773269183933735, + 0.011816211976110935, + 0.015987837687134743, + 0.025421949103474617, + 0.004766062833368778, + 0.007663904223591089, + 0.026128103956580162, + 0.03646685555577278, + 0.00404377281665802, + 0.0038979847449809313, + 0.008126737549901009, + 0.004285031463950872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.04305323585867882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306123033165932 + ], + "opposite_logprob_tokens": [ + -3.1680612564086914 + ], + "logprob_reward": -0.04306123033165932, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306123033165932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0967886745929718, + 0.4264337420463562, + 0.027523234486579895, + 0.025293176993727684, + 0.03766484931111336, + 0.011280655860900879, + 0.041884079575538635, + 0.03538927063345909, + 0.21397465467453003, + 0.08376768231391907 + ], + "credit_map": [ + 0.006589159369468689, + 0.013367028906941414, + 0.012302805669605732, + 0.014629284851253033, + 0.009726953692734241, + 0.01524266880005598, + 0.012599585577845573, + 0.012331187725067139, + 0.009804563596844673, + 0.014831185340881348, + 0.031321194022893906, + 0.039300862699747086, + 0.013888092711567879, + 0.028495093807578087, + 0.17336031794548035, + 0.11543244123458862, + 0.0020932843908667564, + 0.0037019725423306227, + 0.0036152235697954893, + 0.003966909367591143, + 0.0023452225141227245, + 0.0038995679933577776, + 0.0033381592947989702, + 0.004562896676361561, + 0.001553481793962419, + 0.0018942052265629172, + 0.0018782232655212283, + 0.0016188310692086816, + 0.007592928595840931, + 0.002888613846153021, + 0.0020472374744713306, + 0.0058196550235152245, + 0.003864741651341319, + 0.004800624214112759, + 0.005186034366488457, + 0.0031294391956180334, + 0.004725106526166201, + 0.00392507016658783, + 0.0049689156003296375, + 0.007064920384436846, + 0.0013463489012792706, + 0.0014504014980047941, + 0.0016405185451731086, + 0.001008915714919567, + 0.001462012529373169, + 0.0018826060695573688, + 0.0009624538361094892, + 0.0015273989411070943, + 0.004313798155635595, + 0.0052320705726742744, + 0.007162621710449457, + 0.004159762989729643, + 0.005312777124345303, + 0.004519103094935417, + 0.004197750240564346, + 0.0069861942902207375, + 0.0032821930944919586, + 0.004141580779105425, + 0.004573801066726446, + 0.003781231353059411, + 0.005783537868410349, + 0.0063183242455124855, + 0.003697434440255165, + 0.003811170579865575, + 0.020984487608075142, + 0.019621876999735832, + 0.05853680893778801, + 0.023842858150601387, + 0.01862376183271408, + 0.014955895021557808, + 0.021025346592068672, + 0.03638362139463425, + 0.004185566212981939, + 0.005679169669747353, + 0.025195734575390816, + 0.034726664423942566, + 0.0021543619222939014, + 0.0023144313599914312, + 0.0062217083759605885, + 0.003290042281150818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.04306123033165932 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cyberpunk-inspired extreme macro shot plunges the viewer directly into a dark, viscous synthetic fluid, the camera seemingly swept by a powerful electric current. Within this inky black liquid, countless glowing turquoise filaments, resembling synthetic algae, serve as the sole source of light. Each filament possesses a translucent, fibrous texture, its internal structure clearly visible as a complex double-helix spiral of bioluminescent data. Microscopic particles flow rapidly along these pathways like digital information. These threads are caught in a violent and turbulent flow, whipping and twisting erratically, colliding and tangling in a chaotic dance. They momentarily tense, then snap back, forming fleeting, intricate braids before being instantly ripped apart by a powerful, unseen force. Intense green light casts sharp, volumetric rays into the murky liquid, illuminating suspended microscopic particles and creating a dramatic interplay of light and deep shadows. The scene, dominated by cool electric green tones and abyssal blacks, creates a mesmerizing yet unsettling atmosphere of high-tech, microscopic chaos—a beautiful yet violent spectacle of life and data intertwined.", + "seed": 24124, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_7012_seed_24124_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.033164084795862435, + "vlm_logprob": -0.033164084795862435, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.033164084795862435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052262723445892334 + ], + "opposite_logprob_tokens": [ + -3.052262783050537 + ], + "logprob_reward": -0.052262723445892334, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.052262723445892334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1727219671010971, + 0.09304669499397278, + 0.106251560151577, + 0.07509290426969528, + 0.04892844334244728, + 0.11997047066688538, + 0.1501217931509018, + 0.06777424365282059, + 0.04470203071832657, + 0.12138998508453369 + ], + "credit_map": [ + 0.022428208962082863, + 0.020567141473293304, + 0.03367340564727783, + 0.017480015754699707, + 0.016773227602243423, + 0.026699630543589592, + 0.016466038301587105, + 0.01863427273929119, + 0.013798863627016544, + 0.010015247389674187, + 0.012800916098058224, + 0.010817375034093857, + 0.009571504779160023, + 0.009087080135941505, + 0.012731114402413368, + 0.014224577695131302, + 0.012838169932365417, + 0.013269596733152866, + 0.009538494981825352, + 0.01639288291335106, + 0.012295475229620934, + 0.014747634530067444, + 0.011275703087449074, + 0.015893589705228806, + 0.00926457904279232, + 0.0085088349878788, + 0.007149368524551392, + 0.00649336026981473, + 0.009153601713478565, + 0.013471766375005245, + 0.008857770822942257, + 0.012193607166409492, + 0.004760107956826687, + 0.003901783609762788, + 0.005391302052885294, + 0.006907758302986622, + 0.006365404464304447, + 0.00892218854278326, + 0.004124685190618038, + 0.008555209264159203, + 0.017576515674591064, + 0.01505130436271429, + 0.016282612457871437, + 0.013156355358660221, + 0.01213113497942686, + 0.01892722211778164, + 0.014963489957153797, + 0.011881811544299126, + 0.016032638028264046, + 0.012400149367749691, + 0.027216870337724686, + 0.017132645472884178, + 0.017103102058172226, + 0.02673628181219101, + 0.019529862329363823, + 0.01397022232413292, + 0.005246628541499376, + 0.0076149748638272285, + 0.013470843434333801, + 0.00708351144567132, + 0.005519665777683258, + 0.014383644796907902, + 0.00619335426017642, + 0.008261606097221375, + 0.0033920982386916876, + 0.006090641487389803, + 0.007052109111100435, + 0.00329511403106153, + 0.0043584671802818775, + 0.008311721496284008, + 0.0034507960081100464, + 0.008751076646149158, + 0.010701035149395466, + 0.00968290027230978, + 0.016016263514757156, + 0.00911642424762249, + 0.018256237730383873, + 0.014787787571549416, + 0.02620750479400158, + 0.01662181317806244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.052262723445892334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00119304982945323 + ], + "opposite_logprob_tokens": [ + -6.751193046569824 + ], + "logprob_reward": -0.00119304982945323, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.00119304982945323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291893720626831, + 0.12647868692874908, + 0.13665853440761566, + 0.05696909874677658, + 0.0337914302945137, + 0.1743161529302597, + 0.07444804161787033, + 0.11368787288665771, + 0.10403364896774292, + 0.05042709782719612 + ], + "credit_map": [ + 0.01422877423465252, + 0.014888730831444263, + 0.016834422945976257, + 0.011129410937428474, + 0.017997542396187782, + 0.0208930391818285, + 0.014519313350319862, + 0.018698163330554962, + 0.014873453415930271, + 0.012807130813598633, + 0.025892063975334167, + 0.014583353884518147, + 0.012924924492835999, + 0.013717147521674633, + 0.013827163726091385, + 0.01785345934331417, + 0.02697594277560711, + 0.013779601082205772, + 0.018203817307949066, + 0.013124104589223862, + 0.012601678259670734, + 0.017034027725458145, + 0.013908215798437595, + 0.021031176671385765, + 0.008636395446956158, + 0.006190167274326086, + 0.005048638209700584, + 0.005970584694296122, + 0.004917181562632322, + 0.00986284390091896, + 0.005049681290984154, + 0.011293615214526653, + 0.003609119215980172, + 0.0021806166041642427, + 0.004477465059608221, + 0.0030935914255678654, + 0.004775065463036299, + 0.006222649477422237, + 0.003183660563081503, + 0.006249265745282173, + 0.017252802848815918, + 0.019679401069879532, + 0.02336604706943035, + 0.02287040464580059, + 0.015899796038866043, + 0.02738628163933754, + 0.02920578047633171, + 0.018655672669410706, + 0.011459682136774063, + 0.01517370343208313, + 0.01198970153927803, + 0.005280677229166031, + 0.007036618888378143, + 0.013259817846119404, + 0.004929917398840189, + 0.005317927338182926, + 0.013140193186700344, + 0.012821841984987259, + 0.020995361730456352, + 0.010538700968027115, + 0.008796733804047108, + 0.02483091689646244, + 0.007558056153357029, + 0.015006082132458687, + 0.008436255156993866, + 0.01142827421426773, + 0.017151182517409325, + 0.00863902922719717, + 0.007895998656749725, + 0.02133771777153015, + 0.008847198449075222, + 0.020298004150390625, + 0.00498444028198719, + 0.006208130158483982, + 0.007382371928542852, + 0.004821404814720154, + 0.005766571499407291, + 0.005523554049432278, + 0.009448397904634476, + 0.006292233243584633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.00119304982945323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003326714038849 + ], + "opposite_logprob_tokens": [ + -2.695033311843872 + ], + "logprob_reward": -0.07003326714038849, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003326714038849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10980850458145142, + 0.1783328801393509, + 0.12530694901943207, + 0.08079216629266739, + 0.05753578990697861, + 0.09868349879980087, + 0.12167449295520782, + 0.07016566395759583, + 0.022364120930433273, + 0.13533592224121094 + ], + "credit_map": [ + 0.011560648679733276, + 0.0147045673802495, + 0.01293976604938507, + 0.008762163110077381, + 0.015289045870304108, + 0.01630314067006111, + 0.015501122921705246, + 0.014748047105967999, + 0.014788958244025707, + 0.013878965750336647, + 0.01791660115122795, + 0.023373903706669807, + 0.02301451936364174, + 0.019555561244487762, + 0.03810233250260353, + 0.027702035382390022, + 0.010542384348809719, + 0.013119944371283054, + 0.008618426509201527, + 0.011332271620631218, + 0.028777915984392166, + 0.022074267268180847, + 0.012106020003557205, + 0.01873571053147316, + 0.008835739456117153, + 0.008280299603939056, + 0.006151375360786915, + 0.005921078380197287, + 0.012673525139689445, + 0.019511960446834564, + 0.007073603570461273, + 0.012344586662948132, + 0.0045112925581634045, + 0.004208473488688469, + 0.007050524931401014, + 0.006717386189848185, + 0.008487958461046219, + 0.009531562216579914, + 0.004790973849594593, + 0.012237618677318096, + 0.01730194129049778, + 0.01180949155241251, + 0.011828586459159851, + 0.009772660210728645, + 0.009449472650885582, + 0.01533462293446064, + 0.012655981816351414, + 0.010530742816627026, + 0.018340347334742546, + 0.02763344906270504, + 0.015658138319849968, + 0.010611223988234997, + 0.009811032563447952, + 0.019591018557548523, + 0.011182615533471107, + 0.00884667132049799, + 0.0067786406725645065, + 0.009224748238921165, + 0.014726434834301472, + 0.0059790341183543205, + 0.004381971899420023, + 0.016742611303925514, + 0.005108412355184555, + 0.007223811466246843, + 0.0021757681388407946, + 0.0024622289929538965, + 0.005279733333736658, + 0.0018366617150604725, + 0.00196640333160758, + 0.003642765572294593, + 0.0017427769489586353, + 0.0032577812671661377, + 0.012294751591980457, + 0.011289750225841999, + 0.01387794315814972, + 0.010634945705533028, + 0.022609608247876167, + 0.014537022449076176, + 0.031088178977370262, + 0.019003726541996002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.07003326714038849 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09717495739459991, + 0.08022917807102203, + 0.08592738956212997, + 0.22119730710983276, + 0.05837450176477432, + 0.09391378611326218, + 0.16718760132789612, + 0.030122864991426468, + 0.04886237531900406, + 0.11701002717018127 + ], + "credit_map": [ + 0.009337185882031918, + 0.008649910800158978, + 0.010844827629625797, + 0.007343781180679798, + 0.01885463297367096, + 0.017261913046240807, + 0.013705767691135406, + 0.011176937259733677, + 0.008043790236115456, + 0.0094028664752841, + 0.011036518029868603, + 0.010905696079134941, + 0.009429785422980785, + 0.007828065194189548, + 0.010471236892044544, + 0.013111226260662079, + 0.011573471128940582, + 0.010217645205557346, + 0.0077980030328035355, + 0.008053733967244625, + 0.009610423818230629, + 0.015168585814535618, + 0.01069131214171648, + 0.012814213521778584, + 0.01651054434478283, + 0.02118455246090889, + 0.01430729404091835, + 0.01499538216739893, + 0.030770614743232727, + 0.059675004333257675, + 0.020324286073446274, + 0.04342963546514511, + 0.005204167682677507, + 0.006018640473484993, + 0.0039687189273536205, + 0.008365385234355927, + 0.007397809531539679, + 0.008801107294857502, + 0.006056537851691246, + 0.012562132440507412, + 0.012758105993270874, + 0.009141645394265652, + 0.013144392520189285, + 0.013549333438277245, + 0.0071347118355333805, + 0.01368795894086361, + 0.014377795159816742, + 0.010119838640093803, + 0.015191566199064255, + 0.01852467469871044, + 0.03740276023745537, + 0.01743301749229431, + 0.014951251447200775, + 0.03212583810091019, + 0.017583737149834633, + 0.013974768109619617, + 0.00298322387970984, + 0.004125085659325123, + 0.005424836650490761, + 0.0032004257664084435, + 0.0024953342508524656, + 0.006058509461581707, + 0.0019077311735600233, + 0.003927717916667461, + 0.004369490779936314, + 0.006591179873794317, + 0.009375507943332195, + 0.004198946990072727, + 0.00471304589882493, + 0.008397508412599564, + 0.0038856416940689087, + 0.007331050932407379, + 0.009800433181226254, + 0.007132094353437424, + 0.00804397463798523, + 0.008209188468754292, + 0.021269667893648148, + 0.013748014345765114, + 0.03044245019555092, + 0.018364200368523598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.008636740036308765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369464352726936 + ], + "opposite_logprob_tokens": [ + -3.4086947441101074 + ], + "logprob_reward": -0.03369464352726936, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369464352726936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12476890534162521, + 0.14168429374694824, + 0.14351357519626617, + 0.06966513395309448, + 0.04939776659011841, + 0.0625925213098526, + 0.0639377236366272, + 0.056672144681215286, + 0.15223437547683716, + 0.13553351163864136 + ], + "credit_map": [ + 0.013528663665056229, + 0.011394296772778034, + 0.016800621524453163, + 0.011033738031983376, + 0.018119195476174355, + 0.019531436264514923, + 0.013768698088824749, + 0.020592262968420982, + 0.01802118867635727, + 0.016273148357868195, + 0.020893648266792297, + 0.017170272767543793, + 0.014264772646129131, + 0.012310154736042023, + 0.021526692435145378, + 0.021224407479166985, + 0.013234474696218967, + 0.02004101127386093, + 0.01488705538213253, + 0.017197666689753532, + 0.018712051212787628, + 0.022565064951777458, + 0.015209213830530643, + 0.021667037159204483, + 0.007833659648895264, + 0.0072973440401256084, + 0.00586550310254097, + 0.008531667292118073, + 0.008591454476118088, + 0.013485867530107498, + 0.0070457966066896915, + 0.011013842187821865, + 0.0038245604373514652, + 0.004273157101124525, + 0.0043968576937913895, + 0.005536840762943029, + 0.007575099356472492, + 0.007385882083326578, + 0.00494074821472168, + 0.011464620009064674, + 0.010689365677535534, + 0.007932220585644245, + 0.006760685704648495, + 0.007768937386572361, + 0.0065329233184456825, + 0.009457360953092575, + 0.005632082931697369, + 0.007818947546184063, + 0.005718735978007317, + 0.00877174362540245, + 0.00996832549571991, + 0.007348187733441591, + 0.006686561740934849, + 0.010745596140623093, + 0.007522960659116507, + 0.007175617851316929, + 0.005250925198197365, + 0.0036215863656252623, + 0.013849679380655289, + 0.004351679235696793, + 0.0037340382114052773, + 0.012831548228859901, + 0.003970091696828604, + 0.009062603116035461, + 0.013056322932243347, + 0.019321711733937263, + 0.016711123287677765, + 0.015075369738042355, + 0.014451252296566963, + 0.027151430025696754, + 0.014698204584419727, + 0.03176896646618843, + 0.010402250103652477, + 0.011245857924222946, + 0.009225313551723957, + 0.009922629222273827, + 0.02444513514637947, + 0.015443889424204826, + 0.03641922026872635, + 0.018429217860102654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.03369464352726936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cyberpunk-inspired extreme macro shot plunges the viewer directly into a dark, viscous synthetic fluid, the camera seemingly swept by a powerful electric current. Within this inky black liquid, countless glowing turquoise filaments, resembling synthetic algae, serve as the sole source of light. Each filament possesses a translucent, fibrous texture, its internal structure clearly visible as a complex double-helix spiral of bioluminescent data. Microscopic particles flow rapidly along these pathways like digital information. These threads are caught in a violent and turbulent flow, whipping and twisting erratically, colliding and tangling in a chaotic dance. They momentarily tense, then snap back, forming fleeting, intricate braids before being instantly ripped apart by a powerful, unseen force. Intense green light casts sharp, volumetric rays into the murky liquid, illuminating suspended microscopic particles and creating a dramatic interplay of light and deep shadows. The scene, dominated by cool electric green tones and abyssal blacks, creates a mesmerizing yet unsettling atmosphere of high-tech, microscopic chaos—a beautiful yet violent spectacle of life and data intertwined.", + "seed": 24125, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_3/index_7012_seed_24125_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029487190605141222, + "vlm_logprob": -0.029487190605141222, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.029487190605141222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0177932009100914 + ], + "opposite_logprob_tokens": [ + -4.39279317855835 + ], + "logprob_reward": -0.0177932009100914, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0177932009100914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0566333644092083, + 0.12488555163145065, + 0.01921064965426922, + 0.01610041968524456, + 0.2089061737060547, + 0.03938600793480873, + 0.059235237538814545, + 0.36752238869667053, + 0.08281432092189789, + 0.02530580386519432 + ], + "credit_map": [ + 0.006134701427072287, + 0.007576281670480967, + 0.005920058116316795, + 0.006272362545132637, + 0.010786443017423153, + 0.0069343880750238895, + 0.007762555032968521, + 0.0052465759217739105, + 0.010893840342760086, + 0.01473067793995142, + 0.014361964538693428, + 0.008892545476555824, + 0.034792155027389526, + 0.01265445351600647, + 0.008225517347455025, + 0.020334409549832344, + 0.0023111216723918915, + 0.0017418008064851165, + 0.0022834406699985266, + 0.0016897078603506088, + 0.001988971373066306, + 0.0032535663340240717, + 0.0034211899619549513, + 0.0025208513252437115, + 0.0013615210773423314, + 0.001700670225545764, + 0.0023084247950464487, + 0.0018234822200611234, + 0.00197155075147748, + 0.0032162850257009268, + 0.0019532074220478535, + 0.001765277935191989, + 0.018649553880095482, + 0.014371405355632305, + 0.05667182803153992, + 0.013298887759447098, + 0.014869718812406063, + 0.06111837923526764, + 0.017900871112942696, + 0.012025526724755764, + 0.003444628557190299, + 0.005624429322779179, + 0.00621526874601841, + 0.0034922685008496046, + 0.002819608198478818, + 0.010147680528461933, + 0.004453863948583603, + 0.0031882598996162415, + 0.005849056411534548, + 0.003993521444499493, + 0.00722423754632473, + 0.0073105162009596825, + 0.005376975052058697, + 0.009486903436481953, + 0.0089699262753129, + 0.01102409977465868, + 0.030233873054385185, + 0.02563672885298729, + 0.06536994874477386, + 0.018261851742863655, + 0.03386266902089119, + 0.041269320994615555, + 0.0390436016023159, + 0.11384441703557968, + 0.0056969644501805305, + 0.009138095192611217, + 0.014376622624695301, + 0.00688091991469264, + 0.007331981789320707, + 0.011482611298561096, + 0.014482349157333374, + 0.013424772769212723, + 0.0017874828772619367, + 0.003315976355224848, + 0.002095969393849373, + 0.0026699185837060213, + 0.003639722242951393, + 0.00416911905631423, + 0.004539038985967636, + 0.0030885746236890554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.0177932009100914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001529715140350163 + ], + "opposite_logprob_tokens": [ + -6.501529693603516 + ], + "logprob_reward": -0.001529715140350163, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001529715140350163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07054629921913147, + 0.09059351682662964, + 0.05446793884038925, + 0.03456753492355347, + 0.1290803700685501, + 0.03633671626448631, + 0.0779191330075264, + 0.38719093799591064, + 0.05923264101147652, + 0.060064926743507385 + ], + "credit_map": [ + 0.0076521458104252815, + 0.010046044364571571, + 0.00949547067284584, + 0.007412638049572706, + 0.011218870058655739, + 0.007957487367093563, + 0.008834661915898323, + 0.007928982377052307, + 0.007160069886595011, + 0.012336251325905323, + 0.010605658404529095, + 0.006642550695687532, + 0.01815423183143139, + 0.013920331373810768, + 0.0063652885146439075, + 0.015409127809107304, + 0.005262276623398066, + 0.004280234221369028, + 0.008102450519800186, + 0.0032711916137486696, + 0.005900755990296602, + 0.010286183096468449, + 0.008145727217197418, + 0.009219114668667316, + 0.003123109694570303, + 0.004080233629792929, + 0.004529863130301237, + 0.003469933522865176, + 0.003751772455871105, + 0.006036160048097372, + 0.005488474387675524, + 0.004087982699275017, + 0.009981614537537098, + 0.0063597350381314754, + 0.0365515798330307, + 0.0070100389420986176, + 0.007388297468423843, + 0.04365238919854164, + 0.00877444352954626, + 0.009362267330288887, + 0.00391193013638258, + 0.0044520883820950985, + 0.005906601902097464, + 0.003637926187366247, + 0.003664112649857998, + 0.006754336412996054, + 0.004850182682275772, + 0.003159535815939307, + 0.00699519133195281, + 0.004049939569085836, + 0.014268189668655396, + 0.006899763830006123, + 0.005195776466280222, + 0.01080064196139574, + 0.010095834732055664, + 0.01961379311978817, + 0.03166445717215538, + 0.028779588639736176, + 0.06607487797737122, + 0.019823018461465836, + 0.03628867119550705, + 0.044069793075323105, + 0.041775114834308624, + 0.11871539801359177, + 0.00422806479036808, + 0.00474940799176693, + 0.011279013007879257, + 0.004146910272538662, + 0.00458236038684845, + 0.011354212649166584, + 0.008881519548594952, + 0.01001114770770073, + 0.004231499042361975, + 0.007198629435151815, + 0.009447365999221802, + 0.004479657392948866, + 0.007771021220833063, + 0.008322661742568016, + 0.00998593494296074, + 0.00862815510481596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.001529715140350163 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497652664780617 + ], + "opposite_logprob_tokens": [ + -2.929976463317871 + ], + "logprob_reward": -0.05497652664780617, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497652664780617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08971337974071503, + 0.04527847468852997, + 0.20001138746738434, + 0.03636999800801277, + 0.19617782533168793, + 0.028074482455849648, + 0.05893433094024658, + 0.2525889277458191, + 0.044126056134700775, + 0.04872517287731171 + ], + "credit_map": [ + 0.008497516624629498, + 0.012040983885526657, + 0.010508527047932148, + 0.009382582269608974, + 0.01503458060324192, + 0.011714150197803974, + 0.012397749349474907, + 0.0101372841745615, + 0.004767920356243849, + 0.007166214287281036, + 0.005138089414685965, + 0.005094029940664768, + 0.005916791036725044, + 0.007447505835443735, + 0.0046982308849692345, + 0.005049684084951878, + 0.020329726859927177, + 0.014709713868796825, + 0.02870951034128666, + 0.01526162400841713, + 0.021978141739964485, + 0.033159494400024414, + 0.03122207708656788, + 0.034641072154045105, + 0.0037651758175343275, + 0.004200094845145941, + 0.004591669421643019, + 0.003884869161993265, + 0.0039086067117750645, + 0.006436968222260475, + 0.005579059477895498, + 0.004003551322966814, + 0.016798464581370354, + 0.009962895885109901, + 0.05782272666692734, + 0.011935323476791382, + 0.01028444617986679, + 0.06648559123277664, + 0.012378588318824768, + 0.010509778745472431, + 0.0021186931990087032, + 0.003496363293379545, + 0.004968520253896713, + 0.0028641934040933847, + 0.0020008524879813194, + 0.006785599980503321, + 0.003115295898169279, + 0.0027249599806964397, + 0.002816570922732353, + 0.003570537082850933, + 0.01572541892528534, + 0.003179293591529131, + 0.004813774488866329, + 0.017728714272379875, + 0.005353296175599098, + 0.005746716633439064, + 0.021666783839464188, + 0.018162867054343224, + 0.043659135699272156, + 0.013278660364449024, + 0.023045511916279793, + 0.028673548251390457, + 0.027942145243287086, + 0.0761602371931076, + 0.0029473628383129835, + 0.00479994947090745, + 0.0053851790726184845, + 0.004506858065724373, + 0.004195761401206255, + 0.008706407621502876, + 0.007913904264569283, + 0.005670628044754267, + 0.0039948513731360435, + 0.007277037017047405, + 0.005359368864446878, + 0.004263936076313257, + 0.006699236575514078, + 0.007474921643733978, + 0.00771584315225482, + 0.005939972586929798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.05497652664780617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183792009949684 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183792009949684, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183792009949684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05819099768996239, + 0.07840350270271301, + 0.03751783445477486, + 0.03678322583436966, + 0.12235494703054428, + 0.08875736594200134, + 0.09536714851856232, + 0.3505084216594696, + 0.07258012145757675, + 0.05953644961118698 + ], + "credit_map": [ + 0.00785871222615242, + 0.006744281854480505, + 0.007462085224688053, + 0.008431067690253258, + 0.008009557612240314, + 0.007046185899525881, + 0.007312441244721413, + 0.005326662678271532, + 0.010082020424306393, + 0.00904659554362297, + 0.009567509405314922, + 0.010415252298116684, + 0.0149435019120574, + 0.010540823452174664, + 0.006065078545361757, + 0.007742720190435648, + 0.0034704238642007113, + 0.00417372677475214, + 0.0038628429174423218, + 0.0032144628930836916, + 0.0032553719356656075, + 0.007821107283234596, + 0.005932330619543791, + 0.005787568632513285, + 0.0033571659587323666, + 0.004657665733247995, + 0.004702742677181959, + 0.003596068127080798, + 0.004180889576673508, + 0.007669391110539436, + 0.00410706689581275, + 0.004512236453592777, + 0.01102143619209528, + 0.008931359276175499, + 0.0313396193087101, + 0.007619958836585283, + 0.010663764551281929, + 0.033661406487226486, + 0.011541634798049927, + 0.007575764320790768, + 0.008164826780557632, + 0.007870031520724297, + 0.019758855924010277, + 0.005847613792866468, + 0.006316171959042549, + 0.025613142177462578, + 0.009412282146513462, + 0.005774437449872494, + 0.007176434621214867, + 0.004236194770783186, + 0.027245040982961655, + 0.0050206491723656654, + 0.00585330231115222, + 0.02304304949939251, + 0.007689712569117546, + 0.015102768316864967, + 0.030527476221323013, + 0.030959520488977432, + 0.0593988411128521, + 0.020014652982354164, + 0.03132898360490799, + 0.038880202919244766, + 0.03936178237199783, + 0.10003694146871567, + 0.004681488033384085, + 0.006369437091052532, + 0.00846826285123825, + 0.007325350772589445, + 0.008230678737163544, + 0.013548696413636208, + 0.011923020705580711, + 0.012033181264996529, + 0.005655207671225071, + 0.005037497729063034, + 0.007983850315213203, + 0.004959844518452883, + 0.006895184982568026, + 0.010156710632145405, + 0.009645064361393452, + 0.009203087538480759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.018183792009949684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro shot of a dark, viscous synthetic fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless glowing turquoise filaments visible within the fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the filaments display a visible double-helix spiral structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fluid in a violent, turbulent flow with whipping and twisting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does intense green light cast sharp volumetric rays into the murky liquid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549527183175087 + ], + "opposite_logprob_tokens": [ + -2.929952621459961 + ], + "logprob_reward": -0.0549527183175087, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549527183175087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0676482692360878, + 0.08771759271621704, + 0.028582915663719177, + 0.013690924271941185, + 0.17634929716587067, + 0.050009798258543015, + 0.2822638154029846, + 0.21582293510437012, + 0.041190873831510544, + 0.03672361373901367 + ], + "credit_map": [ + 0.007057636976242065, + 0.010457201860845089, + 0.00680609792470932, + 0.00777284475043416, + 0.013544555753469467, + 0.0062859379686415195, + 0.010776770301163197, + 0.004947222303599119, + 0.006627922412008047, + 0.011038198135793209, + 0.01098839845508337, + 0.007507906295359135, + 0.014024699106812477, + 0.016324305906891823, + 0.005469455849379301, + 0.015736691653728485, + 0.004266312345862389, + 0.003114397171884775, + 0.003611733438447118, + 0.001973868114873767, + 0.002841545967385173, + 0.0042751384899020195, + 0.003587901359423995, + 0.004912018310278654, + 0.0014669016236439347, + 0.0016405531205236912, + 0.001959512708708644, + 0.001183762215077877, + 0.001319503178820014, + 0.0025174319744110107, + 0.0018002663273364305, + 0.0018029926577582955, + 0.018026039004325867, + 0.013070940040051937, + 0.045505259186029434, + 0.01309169176965952, + 0.014014613814651966, + 0.04770185053348541, + 0.01484581921249628, + 0.010093086399137974, + 0.004484205506742001, + 0.0054893833585083485, + 0.008461116813123226, + 0.004461144097149372, + 0.0036768547724932432, + 0.014698744751513004, + 0.004657064564526081, + 0.004081282764673233, + 0.014422372914850712, + 0.00900214072316885, + 0.08713965862989426, + 0.012876789085566998, + 0.017660250887274742, + 0.09576286375522614, + 0.018926970660686493, + 0.026472749188542366, + 0.016006695106625557, + 0.016029058024287224, + 0.038419146090745926, + 0.010094472207129002, + 0.019132692366838455, + 0.024642441421747208, + 0.022527268156409264, + 0.06897114217281342, + 0.002991019282490015, + 0.004633396398276091, + 0.0050766500644385815, + 0.003248515073210001, + 0.004303662106394768, + 0.009450766257941723, + 0.006222131196409464, + 0.005264728330075741, + 0.002338304417207837, + 0.005591860972344875, + 0.003144626971334219, + 0.003528740257024765, + 0.004484944976866245, + 0.00564555823802948, + 0.006535320542752743, + 0.005454257130622864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.0549527183175087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_4/index_tensor([7821, 7821])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_4/index_tensor([7821, 7821])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..73918bf4caeb82fbb631de89ec355bebfecef8ff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_4/index_tensor([7821, 7821])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "quick zoom from a slightly lower angle reveals a super-slow-motion, catastrophic collision between two suits of ornate metal armor. At the precise point of impact, the air erupts in a violent, dazzling burst of orange and white sparks, momentarily illuminating the scene with a piercing, flashing light. To the left, a suit crafted from polished gunmetal steel, its front crest engraved with intricate swirling filigree, collides with the breastplate of a rival suit crafted from blackened, angular iron with a dull bronze edging. The immense force is rendered with visceral clarity as the polished gunmetal visibly dents and dents inward, the black iron of the breastplate groaning and twisting, spiderweb cracks splitting its surface. Sharp metal fragments break off from both suits of armor, swirling into the dim stone arena backdrop. The entire moment is captured in brutal detail, brilliant, fleeting rays of light filtering through the dents and scraping the metal's grain, creating a raw and intense atmosphere of pure, destructive power.", + "seed": 17571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_4/index_7821_seed_17571_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09877778217196465, + "vlm_logprob": -0.09877778217196465, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.09877778217196465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08986945450305939 + ], + "opposite_logprob_tokens": [ + -2.589869499206543 + ], + "logprob_reward": -0.08986945450305939, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08986945450305939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29055535793304443, + 0.03776177763938904, + 0.1844990849494934, + 0.05176844447851181, + 0.11730010062456131, + 0.07856498658657074, + 0.05837129056453705, + 0.07599280029535294, + 0.0369650237262249, + 0.0682210922241211 + ], + "credit_map": [ + 0.03235108032822609, + 0.04749338701367378, + 0.03957187011837959, + 0.05459360405802727, + 0.03079056926071644, + 0.020909316837787628, + 0.03536113351583481, + 0.029484404250979424, + 0.006624505389481783, + 0.006542333867400885, + 0.005385502241551876, + 0.004783352836966515, + 0.0038511953316628933, + 0.003766752080991864, + 0.0044607059098780155, + 0.002347428584471345, + 0.031161028891801834, + 0.016823826357722282, + 0.04431088641285896, + 0.021385153755545616, + 0.020873863250017166, + 0.013083441182971, + 0.021169979125261307, + 0.015690907835960388, + 0.00565505912527442, + 0.006135090254247189, + 0.00687535060569644, + 0.006274348124861717, + 0.00805559940636158, + 0.006862367503345013, + 0.00562094384804368, + 0.006289682351052761, + 0.009619386866688728, + 0.014294161461293697, + 0.012796490453183651, + 0.01502961665391922, + 0.014405608177185059, + 0.016901640221476555, + 0.018343575298786163, + 0.015909617766737938, + 0.005153344478458166, + 0.009176951833069324, + 0.010083473287522793, + 0.008809958584606647, + 0.01016572117805481, + 0.011336746625602245, + 0.015399692580103874, + 0.008439097553491592, + 0.006359403487294912, + 0.004233422223478556, + 0.006932937074452639, + 0.010296196676790714, + 0.0061156246811151505, + 0.007434334605932236, + 0.011409683153033257, + 0.005589691456407309, + 0.005768263712525368, + 0.013035423122346401, + 0.00820785854011774, + 0.011003616265952587, + 0.008825951255857944, + 0.00861128605902195, + 0.012282244861125946, + 0.008258157409727573, + 0.0038692348171025515, + 0.0023684976622462273, + 0.00455692782998085, + 0.002918282523751259, + 0.004596218001097441, + 0.0043502552434802055, + 0.007357192225754261, + 0.006948414724320173, + 0.014609030447900295, + 0.004588845185935497, + 0.003798837075009942, + 0.004185494501143694, + 0.008196837268769741, + 0.007038082927465439, + 0.006787620484828949, + 0.01901635155081749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.08986945450305939 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520831227302551 + ], + "opposite_logprob_tokens": [ + -1.5020830631256104 + ], + "logprob_reward": -0.2520831227302551, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520831227302551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1785288006067276, + 0.057850804179906845, + 0.08739902824163437, + 0.05835675820708275, + 0.1486378163099289, + 0.12828628718852997, + 0.07572556287050247, + 0.07990532368421555, + 0.08682233840227127, + 0.09848726540803909 + ], + "credit_map": [ + 0.02293228730559349, + 0.024439428001642227, + 0.02222895622253418, + 0.031156225129961967, + 0.01845429465174675, + 0.012676620855927467, + 0.026061926037073135, + 0.02057906612753868, + 0.007811412680894136, + 0.007459402084350586, + 0.008059273473918438, + 0.01195030938833952, + 0.004713450092822313, + 0.005369008518755436, + 0.009025001898407936, + 0.0034629476722329855, + 0.012393206357955933, + 0.008950475603342056, + 0.012946183793246746, + 0.009649665094912052, + 0.007685599382966757, + 0.009996941313147545, + 0.009169507771730423, + 0.01660745032131672, + 0.004619520623236895, + 0.006019551772624254, + 0.008058789186179638, + 0.006529469974339008, + 0.010273552499711514, + 0.009582730010151863, + 0.006277778185904026, + 0.0069953701458871365, + 0.012873586267232895, + 0.017233831807971, + 0.019278546795248985, + 0.014768417924642563, + 0.017742570489645004, + 0.027684086933732033, + 0.02346980571746826, + 0.015586979687213898, + 0.011567792855203152, + 0.012039249762892723, + 0.01179217267781496, + 0.01415787823498249, + 0.016760874539613724, + 0.021728122606873512, + 0.028994882479310036, + 0.011245320551097393, + 0.008600468747317791, + 0.007312902249395847, + 0.008068492636084557, + 0.01339147798717022, + 0.007268107030540705, + 0.010939561761915684, + 0.01094273291528225, + 0.009201821871101856, + 0.006637099198997021, + 0.006024346221238375, + 0.008263460360467434, + 0.008860786445438862, + 0.011766365729272366, + 0.01579129509627819, + 0.014943461865186691, + 0.007618507836014032, + 0.009977753274142742, + 0.0077382540330290794, + 0.01173218060284853, + 0.01118922233581543, + 0.00902298279106617, + 0.011921793222427368, + 0.011175036430358887, + 0.014065117575228214, + 0.015093770809471607, + 0.005894877947866917, + 0.008032388985157013, + 0.006301588844507933, + 0.01499558612704277, + 0.010869101621210575, + 0.010182196274399757, + 0.027117755264043808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.2520831227302551 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907882869243622 + ], + "opposite_logprob_tokens": [ + -2.464078903198242 + ], + "logprob_reward": -0.08907882869243622, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907882869243622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13899827003479004, + 0.03742202743887901, + 0.08577868342399597, + 0.03170708566904068, + 0.05643397942185402, + 0.10291009396314621, + 0.08148542046546936, + 0.20550400018692017, + 0.08709157258272171, + 0.17266882956027985 + ], + "credit_map": [ + 0.015093553811311722, + 0.024479348212480545, + 0.01881737820804119, + 0.02539343386888504, + 0.015777459368109703, + 0.009091331623494625, + 0.017322678118944168, + 0.013023088686168194, + 0.005701958667486906, + 0.00492927897721529, + 0.004812626168131828, + 0.00619496637955308, + 0.00412385631352663, + 0.004689912311732769, + 0.004132869187742472, + 0.0028365598991513252, + 0.01108712237328291, + 0.007919538766145706, + 0.016367312520742416, + 0.010943095199763775, + 0.011076651513576508, + 0.007499718572944403, + 0.009329738095402718, + 0.0115555040538311, + 0.0043417466804385185, + 0.004128157161176205, + 0.005234573036432266, + 0.0045216865837574005, + 0.003264911938458681, + 0.0038130800239741802, + 0.003294602734968066, + 0.003108327742666006, + 0.0038996010553091764, + 0.005375394131988287, + 0.008512760512530804, + 0.00570228137075901, + 0.005248507484793663, + 0.012750017456710339, + 0.008618121035397053, + 0.006327295210212469, + 0.006541138049215078, + 0.013399502262473106, + 0.009452477097511292, + 0.013090691529214382, + 0.014914795756340027, + 0.015118255279958248, + 0.019781038165092468, + 0.010612192563712597, + 0.007205972447991371, + 0.006055683363229036, + 0.009946702048182487, + 0.012958591803908348, + 0.008523707278072834, + 0.01176803931593895, + 0.015300764702260494, + 0.009725963696837425, + 0.016409125179052353, + 0.043707653880119324, + 0.015479130670428276, + 0.025314858183264732, + 0.023959731683135033, + 0.023746676743030548, + 0.0330842025578022, + 0.023802610114216805, + 0.010721318423748016, + 0.005564847029745579, + 0.011912429705262184, + 0.007777080871164799, + 0.01086192112416029, + 0.010080056264996529, + 0.015447688288986683, + 0.01472623459994793, + 0.04286818951368332, + 0.009784409776329994, + 0.007392221596091986, + 0.007298422511667013, + 0.01787232793867588, + 0.011934866197407246, + 0.011629513464868069, + 0.06388886272907257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.08907882869243622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866749420762062 + ], + "opposite_logprob_tokens": [ + -3.0486674308776855 + ], + "logprob_reward": -0.04866749420762062, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866749420762062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14743319153785706, + 0.044726379215717316, + 0.08052907884120941, + 0.027009397745132446, + 0.15950335562229156, + 0.08060085773468018, + 0.11177197098731995, + 0.07299996167421341, + 0.11730178445577621, + 0.15812398493289948 + ], + "credit_map": [ + 0.016311218962073326, + 0.023245450109243393, + 0.019896941259503365, + 0.03050612099468708, + 0.011565173044800758, + 0.01261116098612547, + 0.018695607781410217, + 0.014601526781916618, + 0.006585604511201382, + 0.0071511524729430676, + 0.006313407327979803, + 0.006814528722316027, + 0.004122252110391855, + 0.004845609422773123, + 0.005040402058511972, + 0.0038534251507371664, + 0.01208107452839613, + 0.006227779667824507, + 0.019635135307908058, + 0.008435158990323544, + 0.00765030225738883, + 0.007226224523037672, + 0.006935296580195427, + 0.012338116765022278, + 0.003426122944802046, + 0.0028913130518049, + 0.003848192049190402, + 0.0034312193747609854, + 0.0036258785985410213, + 0.003957540728151798, + 0.0028645601123571396, + 0.0029645722825080156, + 0.011430328711867332, + 0.013032334856688976, + 0.029002219438552856, + 0.017842140048742294, + 0.013005532324314117, + 0.03182661160826683, + 0.02487187832593918, + 0.0184923242777586, + 0.005496715661138296, + 0.010851121507585049, + 0.00902180653065443, + 0.009416639804840088, + 0.012492175213992596, + 0.011095135472714901, + 0.013874786905944347, + 0.008352485485374928, + 0.009777843952178955, + 0.009285171516239643, + 0.014061137102544308, + 0.015460988506674767, + 0.012800343334674835, + 0.017083559185266495, + 0.018683571368455887, + 0.014619369059801102, + 0.004501700401306152, + 0.010401347652077675, + 0.006082624197006226, + 0.010180577635765076, + 0.009889771230518818, + 0.010070309042930603, + 0.012344376184046268, + 0.00952925905585289, + 0.01434754952788353, + 0.007971891202032566, + 0.01514721941202879, + 0.010535801760852337, + 0.014789425767958164, + 0.013651913963258266, + 0.02097238041460514, + 0.019885612651705742, + 0.023181453347206116, + 0.013287345878779888, + 0.011062223464250565, + 0.010248706676065922, + 0.022890936583280563, + 0.014519676566123962, + 0.01574351079761982, + 0.04719014838337898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.04866749420762062 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190010726451874 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190010726451874, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190010726451874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2692716121673584, + 0.02583763375878334, + 0.5652331113815308, + 0.017466450110077858, + 0.028682736679911613, + 0.039760831743478775, + 0.011077961884438992, + 0.011097830720245838, + 0.011347899213433266, + 0.02022395469248295 + ], + "credit_map": [ + 0.031193194910883904, + 0.04445375129580498, + 0.036740925163030624, + 0.05375427007675171, + 0.021553246304392815, + 0.016985422000288963, + 0.035017941147089005, + 0.029572859406471252, + 0.004062673542648554, + 0.004760296083986759, + 0.004251439124345779, + 0.003222005907446146, + 0.0028169730212539434, + 0.002462643664330244, + 0.0023337018210440874, + 0.0019279028056189418, + 0.08138483017683029, + 0.05009111389517784, + 0.12528914213180542, + 0.07397620379924774, + 0.06855332851409912, + 0.04028711095452309, + 0.06037772819399834, + 0.06527368724346161, + 0.0017584209563210607, + 0.0021121236495673656, + 0.0023545583244413137, + 0.0025772699154913425, + 0.0026332763954997063, + 0.0020925754215568304, + 0.0019441378535702825, + 0.0019940880592912436, + 0.002347709145396948, + 0.005096019245684147, + 0.0032855088356882334, + 0.004657551646232605, + 0.0029774014838039875, + 0.0037114513106644154, + 0.0034964752849191427, + 0.003110618097707629, + 0.0030991442035883665, + 0.003183621447533369, + 0.0036233337596058846, + 0.0051603903993964195, + 0.004714231938123703, + 0.008137810975313187, + 0.008574426174163818, + 0.00326787238009274, + 0.001018456183373928, + 0.0007816465804353356, + 0.0019976398907601833, + 0.0015323797706514597, + 0.001563327619805932, + 0.0014070463366806507, + 0.0015083964681252837, + 0.0012690689181908965, + 0.0007619025418534875, + 0.0013738382840529084, + 0.0008243995835073292, + 0.001489527989178896, + 0.0011467158328741789, + 0.0019677758682519197, + 0.002179199829697609, + 0.0013544701505452394, + 0.0011520714033395052, + 0.0008820000803098083, + 0.001524216728284955, + 0.0012114403070881963, + 0.001174969132989645, + 0.0017345126252621412, + 0.001968747703358531, + 0.0016999408835545182, + 0.0032536012586206198, + 0.0016461103223264217, + 0.0012761123944073915, + 0.0015559470048174262, + 0.002923941472545266, + 0.002498169895261526, + 0.002674099523574114, + 0.004395973403006792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.014190010726451874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "quick zoom from a slightly lower angle reveals a super-slow-motion, catastrophic collision between two suits of ornate metal armor. At the precise point of impact, the air erupts in a violent, dazzling burst of orange and white sparks, momentarily illuminating the scene with a piercing, flashing light. To the left, a suit crafted from polished gunmetal steel, its front crest engraved with intricate swirling filigree, collides with the breastplate of a rival suit crafted from blackened, angular iron with a dull bronze edging. The immense force is rendered with visceral clarity as the polished gunmetal visibly dents and dents inward, the black iron of the breastplate groaning and twisting, spiderweb cracks splitting its surface. Sharp metal fragments break off from both suits of armor, swirling into the dim stone arena backdrop. The entire moment is captured in brutal detail, brilliant, fleeting rays of light filtering through the dents and scraping the metal's grain, creating a raw and intense atmosphere of pure, destructive power.", + "seed": 17572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_4/index_7821_seed_17572_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12437104787677526, + "vlm_logprob": -0.12437104787677526, + "vlm_yes_no_margin": 2.999999988079071, + "vlm_token_credit": -0.12437104787677526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11754173040390015 + ], + "opposite_logprob_tokens": [ + -2.242541790008545 + ], + "logprob_reward": -0.11754173040390015, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11754173040390015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23633749783039093, + 0.14592400193214417, + 0.1524965614080429, + 0.049324799329042435, + 0.05410699173808098, + 0.09936410933732986, + 0.05606577545404434, + 0.05689369887113571, + 0.10476193577051163, + 0.044724635779857635 + ], + "credit_map": [ + 0.024914052337408066, + 0.035945143550634384, + 0.01844378374516964, + 0.0216339323669672, + 0.03432425111532211, + 0.021439651027321815, + 0.05150618404150009, + 0.028130508959293365, + 0.017598459497094154, + 0.023554453626275063, + 0.011658569797873497, + 0.014612637460231781, + 0.029036235064268112, + 0.014026333577930927, + 0.018472012132406235, + 0.016965290531516075, + 0.01793578267097473, + 0.02380445972084999, + 0.017256781458854675, + 0.020757518708705902, + 0.015236840583384037, + 0.012906267307698727, + 0.026073714718222618, + 0.01852518878877163, + 0.009080935269594193, + 0.006656670477241278, + 0.004320480395108461, + 0.005325703416019678, + 0.006959946360439062, + 0.003945919219404459, + 0.007309315726161003, + 0.00572583032771945, + 0.008184471167623997, + 0.008292841725051403, + 0.0058067538775503635, + 0.007776128593832254, + 0.006262593902647495, + 0.003925629425793886, + 0.008740454912185669, + 0.005118115805089474, + 0.005749976262450218, + 0.016817957162857056, + 0.005105240270495415, + 0.006094763521105051, + 0.008934184908866882, + 0.0036085762549191713, + 0.03795754536986351, + 0.015095861628651619, + 0.004887241870164871, + 0.008478461764752865, + 0.005846040789037943, + 0.007920981384813786, + 0.00783186499029398, + 0.005369447637349367, + 0.008641641587018967, + 0.007090089842677116, + 0.014050942845642567, + 0.004467824008315802, + 0.003517070785164833, + 0.006798867601901293, + 0.005817382130771875, + 0.006764880381524563, + 0.007446428295224905, + 0.008030304685235023, + 0.01272664312273264, + 0.010941539891064167, + 0.011142424307763577, + 0.017912838608026505, + 0.011110726743936539, + 0.01181952003389597, + 0.013198357075452805, + 0.015909889712929726, + 0.0050650727935135365, + 0.005271967966109514, + 0.00533750606700778, + 0.008265497162938118, + 0.006495615467429161, + 0.004646684974431992, + 0.004754890687763691, + 0.004887397866696119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.11754173040390015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42887192964553833 + ], + "opposite_logprob_tokens": [ + -1.0538718700408936 + ], + "logprob_reward": -0.42887192964553833, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42887192964553833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23635321855545044, + 0.1386105865240097, + 0.11384519934654236, + 0.05812278762459755, + 0.07420238852500916, + 0.14442409574985504, + 0.08988051116466522, + 0.04020700603723526, + 0.0792098343372345, + 0.025144314393401146 + ], + "credit_map": [ + 0.022472742944955826, + 0.0317390114068985, + 0.023953042924404144, + 0.04233326390385628, + 0.03423748537898064, + 0.02082665264606476, + 0.03932665288448334, + 0.021464388817548752, + 0.016625650227069855, + 0.013540205545723438, + 0.010755736380815506, + 0.011976629495620728, + 0.025541478767991066, + 0.013220001012086868, + 0.03230065852403641, + 0.014650250785052776, + 0.01256493665277958, + 0.016414806246757507, + 0.010835826396942139, + 0.014029438607394695, + 0.01591808721423149, + 0.010264046490192413, + 0.019572163000702858, + 0.014245901256799698, + 0.007027515210211277, + 0.010734609328210354, + 0.0055306521244347095, + 0.005887240637093782, + 0.0056718685664236546, + 0.004389037843793631, + 0.007687927223742008, + 0.011193936690688133, + 0.00837983749806881, + 0.011095530353486538, + 0.0065226005390286446, + 0.00946708396077156, + 0.00934662390500307, + 0.007848056964576244, + 0.010463805869221687, + 0.011078856885433197, + 0.012401980347931385, + 0.02541879378259182, + 0.015925679355859756, + 0.019459929317235947, + 0.01464843563735485, + 0.01727348007261753, + 0.018796661868691444, + 0.02049914374947548, + 0.006316090933978558, + 0.011439952999353409, + 0.009107228368520737, + 0.014621486887335777, + 0.01112737413495779, + 0.013590008020401001, + 0.013807839713990688, + 0.00987053569406271, + 0.008564413525164127, + 0.0037594519089907408, + 0.004010345786809921, + 0.004447678104043007, + 0.006181421224027872, + 0.00390188698656857, + 0.005136953666806221, + 0.004204859957098961, + 0.00879520270973444, + 0.00869734212756157, + 0.008622484281659126, + 0.013161280192434788, + 0.004718210082501173, + 0.011240020394325256, + 0.009911367669701576, + 0.014063932001590729, + 0.0031999072525650263, + 0.002725475700572133, + 0.002957381075248122, + 0.004363310988992453, + 0.0019349594367668033, + 0.0031394967809319496, + 0.0026522246189415455, + 0.004171561449766159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.42887192964553833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812449797987938 + ], + "opposite_logprob_tokens": [ + -3.2881245613098145 + ], + "logprob_reward": -0.03812449797987938, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812449797987938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17755688726902008, + 0.06887142360210419, + 0.1240115538239479, + 0.08176936209201813, + 0.037854455411434174, + 0.23106975853443146, + 0.056567076593637466, + 0.053333912044763565, + 0.08529088646173477, + 0.08367469906806946 + ], + "credit_map": [ + 0.018357930704951286, + 0.0260357353836298, + 0.010454830713570118, + 0.01815386675298214, + 0.02574126049876213, + 0.016431838274002075, + 0.04347411170601845, + 0.018907323479652405, + 0.008689346723258495, + 0.010031614452600479, + 0.0048293909057974815, + 0.006338011007755995, + 0.0120977982878685, + 0.006048247218132019, + 0.013639889657497406, + 0.007197128143161535, + 0.02314574085175991, + 0.017876697704195976, + 0.012435340322554111, + 0.014681600034236908, + 0.017364833503961563, + 0.008140194229781628, + 0.016669290140271187, + 0.013697855174541473, + 0.016550473868846893, + 0.008898088708519936, + 0.007736484985798597, + 0.00740560470148921, + 0.007561779581010342, + 0.00923963449895382, + 0.010053589940071106, + 0.014323703944683075, + 0.004085124004632235, + 0.005597137846052647, + 0.003236382268369198, + 0.004429405555129051, + 0.005517994519323111, + 0.004006142728030682, + 0.00638215709477663, + 0.004600109998136759, + 0.01302433107048273, + 0.04616592824459076, + 0.023094670847058296, + 0.02296992391347885, + 0.01720185950398445, + 0.01201989222317934, + 0.06971205025911331, + 0.026881108060479164, + 0.005476526916027069, + 0.008139895275235176, + 0.006019388325512409, + 0.005942081566900015, + 0.006418286357074976, + 0.007161841727793217, + 0.010021034628152847, + 0.007388018537312746, + 0.011416837573051453, + 0.005064758937805891, + 0.005924184340983629, + 0.004703220911324024, + 0.008230906911194324, + 0.004910239018499851, + 0.007144298404455185, + 0.00593946548178792, + 0.008076361380517483, + 0.010039383545517921, + 0.011472787708044052, + 0.01759147085249424, + 0.008303605020046234, + 0.011195385828614235, + 0.011210720986127853, + 0.007401172537356615, + 0.007734159007668495, + 0.011589009314775467, + 0.010849513113498688, + 0.01010159496217966, + 0.015160871669650078, + 0.008926652371883392, + 0.011313430033624172, + 0.007999470457434654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.03812449797987938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370501473546028 + ], + "opposite_logprob_tokens": [ + -3.408704996109009 + ], + "logprob_reward": -0.03370501473546028, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370501473546028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09677549451589584, + 0.07043060660362244, + 0.07259532064199448, + 0.13370946049690247, + 0.05666080117225647, + 0.22664989531040192, + 0.05102540925145149, + 0.06806415319442749, + 0.08813033998012543, + 0.13595855236053467 + ], + "credit_map": [ + 0.011042523197829723, + 0.011570025235414505, + 0.010764386504888535, + 0.010038837790489197, + 0.014423731714487076, + 0.012250964529812336, + 0.013647009618580341, + 0.013038012199103832, + 0.005085551645606756, + 0.007173809222877026, + 0.005708770826458931, + 0.006388815585523844, + 0.01578778587281704, + 0.007145669311285019, + 0.014910852536559105, + 0.00822934228926897, + 0.008950500749051571, + 0.010175548493862152, + 0.006704192608594894, + 0.008232327178120613, + 0.009852816350758076, + 0.005188310518860817, + 0.01405126042664051, + 0.009440364316105843, + 0.03175937011837959, + 0.015478159300982952, + 0.011055550538003445, + 0.010414142161607742, + 0.01215796172618866, + 0.013012166135013103, + 0.01865343004465103, + 0.02117866463959217, + 0.006218012887984514, + 0.008481768891215324, + 0.004507477395236492, + 0.00627923896536231, + 0.00894349254667759, + 0.005526287946850061, + 0.008589318953454494, + 0.008115199394524097, + 0.013247250579297543, + 0.0466921329498291, + 0.02065644972026348, + 0.022038493305444717, + 0.016049375757575035, + 0.01320978906005621, + 0.06846069544553757, + 0.02629568614065647, + 0.004482673015445471, + 0.008433421142399311, + 0.005859783384948969, + 0.005695203319191933, + 0.005798580590635538, + 0.005656890571117401, + 0.007707773707807064, + 0.00739107932895422, + 0.019555039703845978, + 0.0056874980218708515, + 0.006585179828107357, + 0.006809895858168602, + 0.007898419164121151, + 0.006671207956969738, + 0.009790209122002125, + 0.005066700745373964, + 0.008501960895955563, + 0.008240622468292713, + 0.007154802326112986, + 0.00988763477653265, + 0.006869191769510508, + 0.01977001316845417, + 0.011489847674965858, + 0.016216272488236427, + 0.019555838778614998, + 0.017507730051875114, + 0.012697843834757805, + 0.03147216513752937, + 0.01938113011419773, + 0.012022227980196476, + 0.01343279518187046, + 0.009888810105621815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.03370501473546028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036120666190981865 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.0036120666190981865, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036120666190981865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.157527357339859, + 0.11252818256616592, + 0.06004047021269798, + 0.0705789178609848, + 0.02572389505803585, + 0.15675599873065948, + 0.09286195039749146, + 0.167842835187912, + 0.11592866480350494, + 0.04021167755126953 + ], + "credit_map": [ + 0.016680195927619934, + 0.022698935121297836, + 0.008551283739507198, + 0.015722189098596573, + 0.0207811426371336, + 0.017502224072813988, + 0.037978511303663254, + 0.0176128800958395, + 0.01550943125039339, + 0.014470289461314678, + 0.009978086687624454, + 0.010415714234113693, + 0.01535863894969225, + 0.011026163585484028, + 0.023952821269631386, + 0.01181704644113779, + 0.0071340566501021385, + 0.006642531603574753, + 0.004968191497027874, + 0.007591756526380777, + 0.008362257853150368, + 0.005207651760429144, + 0.01252200547605753, + 0.007612025365233421, + 0.012501765973865986, + 0.00846807099878788, + 0.005992881953716278, + 0.008770526386797428, + 0.00909536611288786, + 0.006562524940818548, + 0.00779836717993021, + 0.011389420367777348, + 0.003067535813897848, + 0.0033169928938150406, + 0.0030654200818389654, + 0.003662155708298087, + 0.0027251210995018482, + 0.0027763003017753363, + 0.002963124541565776, + 0.00414724787697196, + 0.009004484862089157, + 0.026456577703356743, + 0.007352015003561974, + 0.007513769436627626, + 0.014466119930148125, + 0.00430329842492938, + 0.06277292221784592, + 0.024886827915906906, + 0.0052027627825737, + 0.010958115570247173, + 0.00942574068903923, + 0.016879428178071976, + 0.01309898216277361, + 0.013417895883321762, + 0.013160865753889084, + 0.010718164965510368, + 0.055184368044137955, + 0.012641492299735546, + 0.01446748711168766, + 0.011179529130458832, + 0.022246113047003746, + 0.00875174906104803, + 0.025749750435352325, + 0.017622359097003937, + 0.01132629718631506, + 0.010834514163434505, + 0.017686069011688232, + 0.021134864538908005, + 0.013020474463701248, + 0.013093251734972, + 0.015087914653122425, + 0.013745280914008617, + 0.00517698610201478, + 0.005184862297028303, + 0.004414722789078951, + 0.00964917242527008, + 0.005194159224629402, + 0.003957807086408138, + 0.0038751710671931505, + 0.002758800983428955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.0036120666190981865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "quick zoom from a slightly lower angle reveals a super-slow-motion, catastrophic collision between two suits of ornate metal armor. At the precise point of impact, the air erupts in a violent, dazzling burst of orange and white sparks, momentarily illuminating the scene with a piercing, flashing light. To the left, a suit crafted from polished gunmetal steel, its front crest engraved with intricate swirling filigree, collides with the breastplate of a rival suit crafted from blackened, angular iron with a dull bronze edging. The immense force is rendered with visceral clarity as the polished gunmetal visibly dents and dents inward, the black iron of the breastplate groaning and twisting, spiderweb cracks splitting its surface. Sharp metal fragments break off from both suits of armor, swirling into the dim stone arena backdrop. The entire moment is captured in brutal detail, brilliant, fleeting rays of light filtering through the dents and scraping the metal's grain, creating a raw and intense atmosphere of pure, destructive power.", + "seed": 17573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_4/index_7821_seed_17573_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0917875625193119, + "vlm_logprob": -0.0917875625193119, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.0917875625193119, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053278084844350815 + ], + "opposite_logprob_tokens": [ + -3.0532779693603516 + ], + "logprob_reward": -0.053278084844350815, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.053278084844350815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13679762184619904, + 0.0800723135471344, + 0.20211143791675568, + 0.05101621523499489, + 0.054263751953840256, + 0.15296798944473267, + 0.05758194997906685, + 0.041642703115940094, + 0.16420012712478638, + 0.05934591591358185 + ], + "credit_map": [ + 0.018344778567552567, + 0.015879439190030098, + 0.025934942066669464, + 0.022318774834275246, + 0.015104575082659721, + 0.016166597604751587, + 0.011335904709994793, + 0.011712607927620411, + 0.014828318729996681, + 0.008213412016630173, + 0.006475538946688175, + 0.00606826227158308, + 0.009457776322960854, + 0.006904061883687973, + 0.010664447210729122, + 0.017460493370890617, + 0.0109794232994318, + 0.014028120785951614, + 0.015844671055674553, + 0.025235343724489212, + 0.010803867131471634, + 0.015231926925480366, + 0.06527998298406601, + 0.044708117842674255, + 0.003997046034783125, + 0.005824269261211157, + 0.0074524725787341595, + 0.007944419048726559, + 0.006777246482670307, + 0.005704882089048624, + 0.006990106776356697, + 0.006325775291770697, + 0.003867688123136759, + 0.006207256577908993, + 0.007269409019500017, + 0.006971841212362051, + 0.005454908590763807, + 0.008200575597584248, + 0.008513779379427433, + 0.00777829485014081, + 0.011477991007268429, + 0.01579541526734829, + 0.019102921709418297, + 0.021953430026769638, + 0.015865448862314224, + 0.02304011583328247, + 0.024319514632225037, + 0.021413158625364304, + 0.0050941226072609425, + 0.004146980121731758, + 0.007765190675854683, + 0.007129179313778877, + 0.00710756191983819, + 0.008455825038254261, + 0.010194542817771435, + 0.007688545156270266, + 0.004785545170307159, + 0.004900162108242512, + 0.011611346155405045, + 0.0031128134578466415, + 0.0040176198817789555, + 0.004596403334289789, + 0.004240307491272688, + 0.0043785059824585915, + 0.059554800391197205, + 0.015113409608602524, + 0.03408336266875267, + 0.006467518862336874, + 0.004668402951210737, + 0.012751567177474499, + 0.006888639181852341, + 0.024672426283359528, + 0.006935277488082647, + 0.008176127448678017, + 0.00917831901460886, + 0.007661791983991861, + 0.005876983515918255, + 0.005507708061486483, + 0.00891296099871397, + 0.00709674833342433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.053278084844350815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134091794490814 + ], + "opposite_logprob_tokens": [ + -1.3134092092514038 + ], + "logprob_reward": -0.3134091794490814, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134091794490814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16193391382694244, + 0.09682504087686539, + 0.17781555652618408, + 0.04464303329586983, + 0.12923675775527954, + 0.15409870445728302, + 0.052088622003793716, + 0.038801874965429306, + 0.07421187311410904, + 0.07034461945295334 + ], + "credit_map": [ + 0.024135524407029152, + 0.018471315503120422, + 0.02548672817647457, + 0.025997987017035484, + 0.019259251654148102, + 0.01913028582930565, + 0.014175706543028355, + 0.015277110040187836, + 0.01322612538933754, + 0.009300675243139267, + 0.00844462402164936, + 0.010411432012915611, + 0.011001617647707462, + 0.01050917524844408, + 0.012178031727671623, + 0.02175336517393589, + 0.020382380113005638, + 0.021901246160268784, + 0.014058692380785942, + 0.027521900832653046, + 0.0142874326556921, + 0.013738843612372875, + 0.026589276269078255, + 0.03933577612042427, + 0.005249412730336189, + 0.007135499734431505, + 0.0055728512816131115, + 0.005944696255028248, + 0.0053970711305737495, + 0.004532490856945515, + 0.005542878061532974, + 0.005268130451440811, + 0.007696396205574274, + 0.008919241838157177, + 0.03307554870843887, + 0.009953428991138935, + 0.012259052135050297, + 0.01801631599664688, + 0.017638079822063446, + 0.02167869545519352, + 0.024452950805425644, + 0.01411551982164383, + 0.027265965938568115, + 0.01215844415128231, + 0.015566357411444187, + 0.021047690883278847, + 0.020991630852222443, + 0.018500154837965965, + 0.006125430576503277, + 0.004249366465955973, + 0.006845724303275347, + 0.007066677790135145, + 0.004141899757087231, + 0.009337221272289753, + 0.006693302188068628, + 0.0076289973221719265, + 0.003926720004528761, + 0.003552135080099106, + 0.008062116801738739, + 0.0023132916539907455, + 0.0060284752398729324, + 0.0064749810844659805, + 0.003988682758063078, + 0.004455472342669964, + 0.02405853196978569, + 0.007240061182528734, + 0.014985690824687481, + 0.004583308007568121, + 0.003656720509752631, + 0.005727672483772039, + 0.00441380450502038, + 0.00954608153551817, + 0.005868411622941494, + 0.006959946826100349, + 0.010700458660721779, + 0.011657190509140491, + 0.01018572598695755, + 0.006697576493024826, + 0.010776747018098831, + 0.007498562335968018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.3134091794490814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430784709751606 + ], + "opposite_logprob_tokens": [ + -3.1680784225463867 + ], + "logprob_reward": -0.0430784709751606, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430784709751606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0892709344625473, + 0.08490590751171112, + 0.1504972130060196, + 0.04979902133345604, + 0.09525413811206818, + 0.21821117401123047, + 0.07826746255159378, + 0.10670806467533112, + 0.06442572921514511, + 0.0626603439450264 + ], + "credit_map": [ + 0.01108433399349451, + 0.01102907583117485, + 0.011063482612371445, + 0.012184013612568378, + 0.01501902099698782, + 0.009139709174633026, + 0.011609800159931183, + 0.008141495287418365, + 0.011586353182792664, + 0.00896701030433178, + 0.007711492478847504, + 0.013156379573047161, + 0.011248432099819183, + 0.007673198357224464, + 0.012943149544298649, + 0.011619886383414268, + 0.010663117282092571, + 0.009497479535639286, + 0.01051816251128912, + 0.02410324290394783, + 0.009441664442420006, + 0.008383986540138721, + 0.049313683062791824, + 0.028575873002409935, + 0.004429555963724852, + 0.005862397141754627, + 0.007712179329246283, + 0.007827693596482277, + 0.004620827734470367, + 0.004966315347701311, + 0.006226441822946072, + 0.008153609000146389, + 0.0053928145207464695, + 0.01130429282784462, + 0.011994387954473495, + 0.010814509354531765, + 0.011026624590158463, + 0.012630775570869446, + 0.01871141605079174, + 0.013379313983023167, + 0.042249929159879684, + 0.01855149306356907, + 0.03649912402033806, + 0.013816321268677711, + 0.022753676399588585, + 0.02487957291305065, + 0.030943160876631737, + 0.028517886996269226, + 0.006557676941156387, + 0.006583008449524641, + 0.01313023455440998, + 0.009541243314743042, + 0.007919094525277615, + 0.011777903884649277, + 0.010819760151207447, + 0.011938540264964104, + 0.008951621130108833, + 0.006608169991523027, + 0.011397424153983593, + 0.006747513543814421, + 0.008846430107951164, + 0.03625677525997162, + 0.009197547100484371, + 0.018702581524848938, + 0.021494075655937195, + 0.006177867762744427, + 0.013792410492897034, + 0.0030445146840065718, + 0.0020948152523487806, + 0.005348206963390112, + 0.0029617920517921448, + 0.009512043558061123, + 0.007770904339849949, + 0.007907544262707233, + 0.012631852179765701, + 0.008079525083303452, + 0.005320485215634108, + 0.005205678753554821, + 0.008396406657993793, + 0.007347944192588329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.0430784709751606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038104645907878876 + ], + "opposite_logprob_tokens": [ + -3.28810453414917 + ], + "logprob_reward": -0.038104645907878876, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038104645907878876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11997657269239426, + 0.08523447811603546, + 0.11538438498973846, + 0.09291720390319824, + 0.130235493183136, + 0.1185242086648941, + 0.058679014444351196, + 0.14314259588718414, + 0.019789021462202072, + 0.11611701548099518 + ], + "credit_map": [ + 0.013030133210122585, + 0.01687009446322918, + 0.023704223334789276, + 0.02143239602446556, + 0.015773631632328033, + 0.010695800185203552, + 0.011597545817494392, + 0.006872743368148804, + 0.013360587880015373, + 0.00869712419807911, + 0.008078531362116337, + 0.009409702382981777, + 0.009117625653743744, + 0.006781183648854494, + 0.012703169137239456, + 0.01708654873073101, + 0.01320943608880043, + 0.013426927849650383, + 0.010322924703359604, + 0.019314521923661232, + 0.010845799930393696, + 0.010835208930075169, + 0.015880271792411804, + 0.021549295634031296, + 0.010859081521630287, + 0.011001784354448318, + 0.011979077011346817, + 0.016263097524642944, + 0.009671536274254322, + 0.012422025203704834, + 0.010922403074800968, + 0.00979820266366005, + 0.00889234896749258, + 0.01704043708741665, + 0.017205972224473953, + 0.014901862479746342, + 0.012881665490567684, + 0.02319207787513733, + 0.02057969942688942, + 0.015541436150670052, + 0.01390787586569786, + 0.009632130153477192, + 0.021988509222865105, + 0.008977461606264114, + 0.010171287693083286, + 0.016606321558356285, + 0.020481333136558533, + 0.016759289428591728, + 0.006107024848461151, + 0.006127971224486828, + 0.007967918179929256, + 0.009053980000317097, + 0.005333244800567627, + 0.007916335016489029, + 0.007994601503014565, + 0.00817793607711792, + 0.011511427350342274, + 0.009185148403048515, + 0.026387803256511688, + 0.009835264645516872, + 0.011996750719845295, + 0.04108341410756111, + 0.01136019453406334, + 0.021782588213682175, + 0.0027482248842716217, + 0.0026648917701095343, + 0.004884040914475918, + 0.0019073362927883863, + 0.0017515579238533974, + 0.0016257554525509477, + 0.002049854723736644, + 0.0021573598496615887, + 0.013049660250544548, + 0.01623501442372799, + 0.009267142973840237, + 0.01643475517630577, + 0.01390930823981762, + 0.01095209177583456, + 0.019066771492362022, + 0.017202280461788177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.038104645907878876 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067431420087814 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067431420087814, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067431420087814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4919828772544861, + 0.06943920254707336, + 0.1058911457657814, + 0.03542211651802063, + 0.10300414264202118, + 0.04118399694561958, + 0.038056518882513046, + 0.04220957309007645, + 0.018715456128120422, + 0.05409497395157814 + ], + "credit_map": [ + 0.08168008178472519, + 0.06970938295125961, + 0.08318881690502167, + 0.09483855217695236, + 0.04620658606290817, + 0.033409517258405685, + 0.049423810094594955, + 0.033526092767715454, + 0.010394887998700142, + 0.006876809522509575, + 0.006349393166601658, + 0.005634518805891275, + 0.008830524049699306, + 0.00826520100235939, + 0.01128829549998045, + 0.011799572966992855, + 0.0065184044651687145, + 0.007121536880731583, + 0.006777735892683268, + 0.016017887741327286, + 0.006118848919868469, + 0.005814936012029648, + 0.03512582927942276, + 0.022395970299839973, + 0.003827156964689493, + 0.004411390516906977, + 0.0054479618556797504, + 0.005252702161669731, + 0.003289290936663747, + 0.005120436195284128, + 0.004159469623118639, + 0.003913705702871084, + 0.005637559574097395, + 0.008531897328794003, + 0.015981638804078102, + 0.010731428861618042, + 0.012404747307300568, + 0.01701168529689312, + 0.01633583940565586, + 0.016369344666600227, + 0.0070186154916882515, + 0.0038907171692699194, + 0.0049666231498122215, + 0.004255133215337992, + 0.005612243432551622, + 0.006039447616785765, + 0.004690178204327822, + 0.004711037036031485, + 0.004851591773331165, + 0.0037567478138953447, + 0.005113841034471989, + 0.005234249401837587, + 0.003826049854978919, + 0.005539874080568552, + 0.004219175316393375, + 0.005514988675713539, + 0.003015567548573017, + 0.0028221188113093376, + 0.004114838317036629, + 0.0028749199118465185, + 0.004128045868128538, + 0.013221239671111107, + 0.00463339826092124, + 0.007399444468319416, + 0.002556973835453391, + 0.002063584513962269, + 0.003619356080889702, + 0.001960246590897441, + 0.002849338576197624, + 0.001888983533717692, + 0.002209741622209549, + 0.0015672313747927547, + 0.005289874039590359, + 0.008409642614424229, + 0.007262790109962225, + 0.006050588097423315, + 0.0069708917289972305, + 0.006089646369218826, + 0.007091166451573372, + 0.006930372212082148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.011067431420087814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "quick zoom from a slightly lower angle reveals a super-slow-motion, catastrophic collision between two suits of ornate metal armor. At the precise point of impact, the air erupts in a violent, dazzling burst of orange and white sparks, momentarily illuminating the scene with a piercing, flashing light. To the left, a suit crafted from polished gunmetal steel, its front crest engraved with intricate swirling filigree, collides with the breastplate of a rival suit crafted from blackened, angular iron with a dull bronze edging. The immense force is rendered with visceral clarity as the polished gunmetal visibly dents and dents inward, the black iron of the breastplate groaning and twisting, spiderweb cracks splitting its surface. Sharp metal fragments break off from both suits of armor, swirling into the dim stone arena backdrop. The entire moment is captured in brutal detail, brilliant, fleeting rays of light filtering through the dents and scraping the metal's grain, creating a raw and intense atmosphere of pure, destructive power.", + "seed": 17574, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_4/index_7821_seed_17574_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09908266514539718, + "vlm_logprob": -0.09908266514539718, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.09908266514539718, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06609293073415756 + ], + "opposite_logprob_tokens": [ + -2.8160929679870605 + ], + "logprob_reward": -0.06609293073415756, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06609293073415756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22776366770267487, + 0.11426258087158203, + 0.06069960817694664, + 0.06879069656133652, + 0.06985054910182953, + 0.06957314163446426, + 0.08478403836488724, + 0.09982883930206299, + 0.07405014336109161, + 0.13039670884609222 + ], + "credit_map": [ + 0.023224756121635437, + 0.029147155582904816, + 0.04713631421327591, + 0.030590161681175232, + 0.03540319576859474, + 0.024709703400731087, + 0.023949556052684784, + 0.013602808117866516, + 0.020441699773073196, + 0.014398365281522274, + 0.012558835558593273, + 0.008318091742694378, + 0.013123990036547184, + 0.01696445234119892, + 0.01571103185415268, + 0.012746105901896954, + 0.005561589729040861, + 0.006097530014812946, + 0.008876210078597069, + 0.009793742559850216, + 0.01112319901585579, + 0.005471661686897278, + 0.007522444240748882, + 0.006253234576433897, + 0.006517260335385799, + 0.01017746701836586, + 0.00788898952305317, + 0.009577141143381596, + 0.00859500840306282, + 0.009226451627910137, + 0.007738699670881033, + 0.009069681167602539, + 0.0069474088959395885, + 0.009418769739568233, + 0.006520950701087713, + 0.010859180241823196, + 0.010249855928122997, + 0.005778124555945396, + 0.010964849032461643, + 0.00911141000688076, + 0.007530460599809885, + 0.01581389456987381, + 0.005196939688175917, + 0.0057806889526546, + 0.012595093809068203, + 0.0055635711178183556, + 0.00759064219892025, + 0.009501850232481956, + 0.006056673359125853, + 0.006340364925563335, + 0.008365709334611893, + 0.006793677806854248, + 0.006737605668604374, + 0.006398754194378853, + 0.024042388424277306, + 0.020048867911100388, + 0.00822005420923233, + 0.02476891316473484, + 0.008832411840558052, + 0.01544882357120514, + 0.008721232414245605, + 0.006449814420193434, + 0.014668452553451061, + 0.012719137594103813, + 0.006585756316781044, + 0.008838490583002567, + 0.009864550083875656, + 0.011725625954568386, + 0.005171917378902435, + 0.006800051312893629, + 0.012633314356207848, + 0.012430439703166485, + 0.023991862311959267, + 0.008936799131333828, + 0.03714967891573906, + 0.012509101070463657, + 0.008125354535877705, + 0.005801775027066469, + 0.012136262841522694, + 0.021745888516306877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.06609293073415756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813093960285187 + ], + "opposite_logprob_tokens": [ + -1.4063093662261963 + ], + "logprob_reward": -0.2813093960285187, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813093960285187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5164233446121216, + 0.033459048718214035, + 0.06402453780174255, + 0.04714484512805939, + 0.043880730867385864, + 0.08001358062028885, + 0.08059263974428177, + 0.0200465340167284, + 0.053313206881284714, + 0.06110154837369919 + ], + "credit_map": [ + 0.04815531149506569, + 0.07889959961175919, + 0.20506399869918823, + 0.06358018517494202, + 0.030300002545118332, + 0.027040865272283554, + 0.043930914252996445, + 0.01945246011018753, + 0.005027003586292267, + 0.0037999842315912247, + 0.0044654663652181625, + 0.0027664536610245705, + 0.0051823565736413, + 0.004945965018123388, + 0.004151853732764721, + 0.0031199627555906773, + 0.007649785839021206, + 0.006573578342795372, + 0.008992312476038933, + 0.006408359855413437, + 0.010890613310039043, + 0.009534010663628578, + 0.007727183401584625, + 0.006248691584914923, + 0.0053363158367574215, + 0.006301254034042358, + 0.0031651079189032316, + 0.00952815916389227, + 0.006001933943480253, + 0.005471348762512207, + 0.0039001277182251215, + 0.007440602872520685, + 0.0055056242272257805, + 0.0060812924057245255, + 0.003425748087465763, + 0.006104272790253162, + 0.005428981967270374, + 0.004413133952766657, + 0.006477816496044397, + 0.00644386000931263, + 0.007224501110613346, + 0.013432075269520283, + 0.006691544782370329, + 0.008510597050189972, + 0.015713036060333252, + 0.007732961792498827, + 0.011297833174467087, + 0.009411036036908627, + 0.005718293599784374, + 0.005909562576562166, + 0.0070614805445075035, + 0.005692475941032171, + 0.006360281258821487, + 0.008109469898045063, + 0.021483877673745155, + 0.02025720290839672, + 0.0024016937240958214, + 0.001951705664396286, + 0.0023343570064753294, + 0.003183271735906601, + 0.0013497540494427085, + 0.0018822230631485581, + 0.004248400684446096, + 0.0026951292529702187, + 0.008813407272100449, + 0.0042749326676130295, + 0.00826280191540718, + 0.006883229129016399, + 0.0040739853866398335, + 0.007428851909935474, + 0.0054024942219257355, + 0.008173505775630474, + 0.013186916708946228, + 0.0044601084664464, + 0.01433706097304821, + 0.0070052663795650005, + 0.00488109840080142, + 0.003546981606632471, + 0.004489669110625982, + 0.009194446727633476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.2813093960285187 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811256214976311 + ], + "opposite_logprob_tokens": [ + -3.2881126403808594 + ], + "logprob_reward": -0.03811256214976311, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811256214976311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21180923283100128, + 0.04445033520460129, + 0.05198080465197563, + 0.12181561440229416, + 0.05394664779305458, + 0.06375797837972641, + 0.05609072744846344, + 0.21029621362686157, + 0.06758186966180801, + 0.11827056854963303 + ], + "credit_map": [ + 0.023067446425557137, + 0.030271140858530998, + 0.06657949090003967, + 0.026407450437545776, + 0.015859585255384445, + 0.013473646715283394, + 0.021582771092653275, + 0.014567703939974308, + 0.0077100154012441635, + 0.006890079006552696, + 0.004934001713991165, + 0.003519823309034109, + 0.004750388674438, + 0.005985895171761513, + 0.0053230347111821175, + 0.005337096750736237, + 0.006128608714789152, + 0.005052848719060421, + 0.00798521563410759, + 0.006197870709002018, + 0.00677458755671978, + 0.007958213798701763, + 0.006280551664531231, + 0.005602911580353975, + 0.01210663840174675, + 0.017457276582717896, + 0.006778737064450979, + 0.011838103644549847, + 0.013814607635140419, + 0.017936566844582558, + 0.011431592516601086, + 0.030452096834778786, + 0.0035696295090019703, + 0.007102217059582472, + 0.004026079550385475, + 0.00660292012616992, + 0.007925054989755154, + 0.00415011215955019, + 0.012928820215165615, + 0.007641818840056658, + 0.006297644693404436, + 0.009380651637911797, + 0.005310903303325176, + 0.006332000717520714, + 0.014029993675649166, + 0.005923728458583355, + 0.008721978403627872, + 0.007761085871607065, + 0.004902883432805538, + 0.004483744967728853, + 0.005911972839385271, + 0.005176763515919447, + 0.005498742219060659, + 0.006466159597039223, + 0.013775710016489029, + 0.009874755516648293, + 0.01893606223165989, + 0.05884790048003197, + 0.02068493887782097, + 0.02899295836687088, + 0.02027052454650402, + 0.011156429536640644, + 0.02623235434293747, + 0.025175049901008606, + 0.00920640304684639, + 0.008200710639357567, + 0.011295769363641739, + 0.008916650898754597, + 0.0061520058661699295, + 0.005928690545260906, + 0.00893077440559864, + 0.008950868621468544, + 0.015792474150657654, + 0.011986781843006611, + 0.019734622910618782, + 0.01628040336072445, + 0.019547084346413612, + 0.007460348773747683, + 0.013564917258918285, + 0.01390394289046526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.03811256214976311 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495835840702057 + ], + "opposite_logprob_tokens": [ + -2.9299583435058594 + ], + "logprob_reward": -0.05495835840702057, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495835840702057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18272969126701355, + 0.03915366530418396, + 0.05001608654856682, + 0.08306852728128433, + 0.13485650718212128, + 0.08111529797315598, + 0.13395775854587555, + 0.045545779168605804, + 0.10711798071861267, + 0.14243876934051514 + ], + "credit_map": [ + 0.021967174485325813, + 0.02428518608212471, + 0.037764646112918854, + 0.02622939832508564, + 0.022405100986361504, + 0.017977675423026085, + 0.017734896391630173, + 0.014365600422024727, + 0.0067917038686573505, + 0.00490735424682498, + 0.004365899134427309, + 0.003865601494908333, + 0.004563306458294392, + 0.006359117105603218, + 0.004740025848150253, + 0.0035606520250439644, + 0.006215888541191816, + 0.0059892116114497185, + 0.007511853240430355, + 0.0063819256611168385, + 0.006402455270290375, + 0.006733735557645559, + 0.005345226731151342, + 0.005435783416032791, + 0.010113243944942951, + 0.012920602224767208, + 0.009300965815782547, + 0.011120365932583809, + 0.009226839058101177, + 0.008817145600914955, + 0.008560867048799992, + 0.013008489273488522, + 0.007821924053132534, + 0.018856210634112358, + 0.008054624311625957, + 0.013296186923980713, + 0.02716524712741375, + 0.009928349405527115, + 0.02898486889898777, + 0.020749075338244438, + 0.006481807678937912, + 0.013475907035171986, + 0.0067061930894851685, + 0.008016236126422882, + 0.01438831351697445, + 0.008984691463410854, + 0.010330503806471825, + 0.012731639668345451, + 0.00923234410583973, + 0.017108289524912834, + 0.01267437357455492, + 0.013925393112003803, + 0.010169713757932186, + 0.010313073173165321, + 0.03187016397714615, + 0.02866438776254654, + 0.005743993911892176, + 0.006163871381431818, + 0.0053494637832045555, + 0.007746351417154074, + 0.0030830795876681805, + 0.004665698856115341, + 0.007246802095323801, + 0.005546513479202986, + 0.012752781622111797, + 0.009421388618648052, + 0.011123398318886757, + 0.01712716743350029, + 0.007069485727697611, + 0.01213268656283617, + 0.015498383902013302, + 0.021992679685354233, + 0.01648864522576332, + 0.010263778269290924, + 0.026388036087155342, + 0.017423070967197418, + 0.014972302131354809, + 0.010309082455933094, + 0.019473081454634666, + 0.027120772749185562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.05495835840702057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a quick zoom from a slightly lower angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot in super-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one suit made of polished gunmetal steel with a front crest engraved with intricate swirling filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other suit made of blackened angular iron with a dull bronze edging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact produce a violent burst of orange and white sparks that momentarily illuminates the scene with flashing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054940078407526016 + ], + "opposite_logprob_tokens": [ + -2.9299399852752686 + ], + "logprob_reward": -0.054940078407526016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054940078407526016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34621909260749817, + 0.11979445815086365, + 0.04948343336582184, + 0.09573853015899658, + 0.07668133080005646, + 0.03304208815097809, + 0.09410177171230316, + 0.06839408725500107, + 0.06781148910522461, + 0.04873374104499817 + ], + "credit_map": [ + 0.03391752019524574, + 0.048257309943437576, + 0.053148455917835236, + 0.056793250143527985, + 0.04311059042811394, + 0.0375833697617054, + 0.048565927892923355, + 0.024842659011483192, + 0.016332121565937996, + 0.014571327716112137, + 0.010677333921194077, + 0.010425537824630737, + 0.026344187557697296, + 0.012777874246239662, + 0.018869711086153984, + 0.00979635864496231, + 0.00506560318171978, + 0.006609233096241951, + 0.006862988229840994, + 0.005589645821601152, + 0.008847322314977646, + 0.006061881314963102, + 0.0046548303216695786, + 0.005791925825178623, + 0.009642321616411209, + 0.012203206308186054, + 0.00528945354744792, + 0.007746435236185789, + 0.012027185410261154, + 0.014078318141400814, + 0.009418056346476078, + 0.025333549827337265, + 0.0051728649996221066, + 0.010698880068957806, + 0.004887203685939312, + 0.0076791588217020035, + 0.014538299292325974, + 0.006469404324889183, + 0.014913283288478851, + 0.012322238646447659, + 0.00280546466819942, + 0.006573570426553488, + 0.0025313037913292646, + 0.0039087687619030476, + 0.006181496195495129, + 0.0029598528053611517, + 0.00431798817589879, + 0.003763644490391016, + 0.006417619530111551, + 0.010338464751839638, + 0.009283054620027542, + 0.00922403670847416, + 0.006273436360061169, + 0.006606690585613251, + 0.024701157584786415, + 0.0212573129683733, + 0.006285371258854866, + 0.020310765132308006, + 0.006062525324523449, + 0.008162908256053925, + 0.007348916959017515, + 0.0042707244865596294, + 0.008098012767732143, + 0.007854867726564407, + 0.010385099798440933, + 0.007227297872304916, + 0.009260200895369053, + 0.008999284356832504, + 0.0060885073617100716, + 0.009435441344976425, + 0.007745175156742334, + 0.00867048092186451, + 0.0070119998417794704, + 0.0031275341752916574, + 0.012274907901883125, + 0.005721923895180225, + 0.003684455528855324, + 0.002997493138536811, + 0.004861791152507067, + 0.009053637273609638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.054940078407526016 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_5/index_tensor([4960, 4960])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_5/index_tensor([4960, 4960])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3293765680842a416d4d7c3ee41d254a67e617c4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_5/index_tensor([4960, 4960])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, smooth tracking shot, viewed from a slightly upward angle, follows a giant octopus tentacle undergoing a breathtaking process of petrification. Against the deep, almost pitch-black seafloor, a precise, cool white beam, reminiscent of a deep-sea probe, illuminates this surreal transformation. The shot begins at the still-vibrant end of the tentacle. Its deep purple skin, coated in a slippery, iridescent sheen, reflects in the beam, while its neatly arranged, massive suction cups undergo their final, delicate contractions. As the camera pans steadily to the right, we witness the core of the transformation: the slippery sheen fades at a visible rate, the elastic texture hardens inch by inch, and the deep purple skin rapidly pales, transforming into a lifeless, rocky gray. Countless hairline cracks, like spiderwebs, spread from the hardened surface, rapidly deepening and widening, forming deep grooves, as if ancient rock has been fractured by weathering. The once-soft suction cups have solidified into rigid, stone disks, devoid of all signs of life. The entire process is presented in extreme details in the CGI realistic style. The suspended particles in the water are clearly visible in the beam of light, and the light and shadow form a sharp contrast between the rough rock cracks and the smooth wet skin, creating a magnificent, strange and breathtakingly quiet atmosphere.", + "seed": 25290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_5/index_4960_seed_25290_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16716393940150737, + "vlm_logprob": -0.16716393940150737, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.16716393940150737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08252446353435516 + ], + "opposite_logprob_tokens": [ + -2.582524538040161 + ], + "logprob_reward": -0.08252446353435516, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08252446353435516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07288996130228043, + 0.10556496679782867, + 0.07987690716981888, + 0.17288877069950104, + 0.225539892911911, + 0.14132635295391083, + 0.07705560326576233, + 0.03392329066991806, + 0.04260452091693878, + 0.048329759389162064 + ], + "credit_map": [ + 0.004401760641485453, + 0.005259506404399872, + 0.009154506027698517, + 0.012906712479889393, + 0.007166157476603985, + 0.005264129489660263, + 0.011334001086652279, + 0.017403187230229378, + 0.008309459313750267, + 0.006991323549300432, + 0.015755459666252136, + 0.01136790681630373, + 0.01403573714196682, + 0.011366790160536766, + 0.014269252307713032, + 0.02346903271973133, + 0.005714491009712219, + 0.007058221381157637, + 0.006728935521095991, + 0.010009288787841797, + 0.0069373431615531445, + 0.013049660250544548, + 0.018115244805812836, + 0.012263725511729717, + 0.018170995637774467, + 0.020954065024852753, + 0.018315155059099197, + 0.022927260026335716, + 0.015901794657111168, + 0.02273460477590561, + 0.03186009079217911, + 0.022024812176823616, + 0.025665225461125374, + 0.020499279722571373, + 0.04645003750920296, + 0.028706019744277, + 0.0206147488206625, + 0.026764869689941406, + 0.029831528663635254, + 0.027008183300495148, + 0.025545746088027954, + 0.016152361407876015, + 0.01880151964724064, + 0.013523432426154613, + 0.011112329550087452, + 0.01660095527768135, + 0.012157207354903221, + 0.027432793751358986, + 0.005274539347738028, + 0.0072659640572965145, + 0.007815179415047169, + 0.017207428812980652, + 0.009179523214697838, + 0.008637274615466595, + 0.009271612390875816, + 0.012404080480337143, + 0.003626007353886962, + 0.004022057633846998, + 0.005258292891085148, + 0.004593391437083483, + 0.003038187976926565, + 0.005234001204371452, + 0.003999399021267891, + 0.004151955246925354, + 0.0035590692423284054, + 0.006036467384546995, + 0.005141103640198708, + 0.0060010915622115135, + 0.004865835886448622, + 0.003911741077899933, + 0.006651896983385086, + 0.006437312345951796, + 0.003553834045305848, + 0.004302529618144035, + 0.00437505217269063, + 0.005260789301246405, + 0.0037321792915463448, + 0.008449996821582317, + 0.007862837053835392, + 0.010792539454996586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.08252446353435516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255148183554411 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.01255148183554411, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255148183554411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16701413691043854, + 0.08675161004066467, + 0.09759722650051117, + 0.1361268162727356, + 0.12979602813720703, + 0.08118356019258499, + 0.10823440551757812, + 0.027985481545329094, + 0.046974699944257736, + 0.1183360293507576 + ], + "credit_map": [ + 0.011641168966889381, + 0.014943997375667095, + 0.01927858591079712, + 0.02545628510415554, + 0.021749109029769897, + 0.015899505466222763, + 0.03222372382879257, + 0.025821762159466743, + 0.00795737188309431, + 0.008652840740978718, + 0.01396563183516264, + 0.008747132495045662, + 0.014165788888931274, + 0.008130664937198162, + 0.008874049410223961, + 0.016258135437965393, + 0.009716848842799664, + 0.013023482635617256, + 0.0071146064437925816, + 0.010411207564175129, + 0.009741159155964851, + 0.013193930499255657, + 0.022285904735326767, + 0.012110087089240551, + 0.016750816255807877, + 0.015490439720451832, + 0.015511135570704937, + 0.020433926954865456, + 0.012367470189929008, + 0.01660160906612873, + 0.018896538764238358, + 0.02007487416267395, + 0.016696535050868988, + 0.017272379249334335, + 0.018405728042125702, + 0.019110405817627907, + 0.010711702518165112, + 0.0146439578384161, + 0.016826706007122993, + 0.01612861454486847, + 0.013207294046878815, + 0.013388220220804214, + 0.0069822887890040874, + 0.007967314682900906, + 0.00846825260668993, + 0.008627014234662056, + 0.008918153122067451, + 0.013625026680529118, + 0.0068665905855596066, + 0.012135163880884647, + 0.014642258174717426, + 0.018345018848776817, + 0.012166796252131462, + 0.01388312503695488, + 0.014939519576728344, + 0.015255930833518505, + 0.0019973875023424625, + 0.004789656028151512, + 0.004897046834230423, + 0.003745150752365589, + 0.0021083466708660126, + 0.0033311122097074986, + 0.0033840169198811054, + 0.0037327639292925596, + 0.003828776767477393, + 0.007499219384044409, + 0.0059537459164857864, + 0.00585416704416275, + 0.0058523365296423435, + 0.005057619884610176, + 0.005717274267226458, + 0.0072115580551326275, + 0.012963732704520226, + 0.013136591762304306, + 0.019086569547653198, + 0.012860887683928013, + 0.009429770521819592, + 0.01440555602312088, + 0.018570316955447197, + 0.01788260228931904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.01255148183554411 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636639289557934 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.007636639289557934, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636639289557934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06364165991544724, + 0.12726420164108276, + 0.05695013329386711, + 0.06365624070167542, + 0.07070492953062057, + 0.13070286810398102, + 0.08377886563539505, + 0.035167496651411057, + 0.1898793876171112, + 0.17825423181056976 + ], + "credit_map": [ + 0.004494704771786928, + 0.0052362424321472645, + 0.0085455272346735, + 0.010726885870099068, + 0.006549330428242683, + 0.005224567372351885, + 0.007885648868978024, + 0.014978748746216297, + 0.014257497154176235, + 0.01506443414837122, + 0.028608957305550575, + 0.013928030617535114, + 0.020523501560091972, + 0.012121831998229027, + 0.013122967444360256, + 0.009636985138058662, + 0.005056505557149649, + 0.005227141547948122, + 0.00434999680146575, + 0.006231508683413267, + 0.005505306646227837, + 0.010337449610233307, + 0.012652531266212463, + 0.007589694112539291, + 0.006810534279793501, + 0.006340331397950649, + 0.007117444649338722, + 0.01086968183517456, + 0.006396050564944744, + 0.007315731607377529, + 0.011892426759004593, + 0.0069140419363975525, + 0.008850576356053352, + 0.009238094091415405, + 0.008504132740199566, + 0.011027801781892776, + 0.005602497141808271, + 0.007752492092549801, + 0.010172227397561073, + 0.009557115845382214, + 0.01892993599176407, + 0.021651780232787132, + 0.018383555114269257, + 0.012366166338324547, + 0.010298005305230618, + 0.015774719417095184, + 0.012455870397388935, + 0.020842835307121277, + 0.0043592085130512714, + 0.009253968484699726, + 0.009870585054159164, + 0.014458398334681988, + 0.008421248756349087, + 0.00805679801851511, + 0.016937727108597755, + 0.01242093462496996, + 0.002227669581770897, + 0.005209049209952354, + 0.005200893618166447, + 0.005571693181991577, + 0.003428781870752573, + 0.0031734800431877375, + 0.0055095236748456955, + 0.004846405703574419, + 0.015689173713326454, + 0.02569168619811535, + 0.03025442734360695, + 0.025402376428246498, + 0.022510766983032227, + 0.020816447213292122, + 0.023326385766267776, + 0.02618812397122383, + 0.015642093494534492, + 0.022778352722525597, + 0.019485151395201683, + 0.024416886270046234, + 0.01838969811797142, + 0.020905353128910065, + 0.031912874430418015, + 0.02472381852567196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.007636639289557934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328083276748657 + ], + "opposite_logprob_tokens": [ + -0.7578083276748657 + ], + "logprob_reward": -0.6328083276748657, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328083276748657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14233815670013428, + 0.12115354090929031, + 0.10219543427228928, + 0.06396472454071045, + 0.11878447234630585, + 0.04744522273540497, + 0.12814000248908997, + 0.08165578544139862, + 0.07631111145019531, + 0.11801154911518097 + ], + "credit_map": [ + 0.012934735044836998, + 0.015847977250814438, + 0.013615581206977367, + 0.020817266777157784, + 0.02518336847424507, + 0.010995774529874325, + 0.02610766887664795, + 0.01683579385280609, + 0.015438666567206383, + 0.01779116503894329, + 0.020495224744081497, + 0.018722275272011757, + 0.012808896601200104, + 0.011624794453382492, + 0.012617713771760464, + 0.011654798872768879, + 0.012363027781248093, + 0.012273434549570084, + 0.009660452604293823, + 0.01169603131711483, + 0.009131542406976223, + 0.012003254145383835, + 0.02572859264910221, + 0.009339108131825924, + 0.009555958211421967, + 0.008799922652542591, + 0.007383319549262524, + 0.00660270219668746, + 0.004192882217466831, + 0.005703771952539682, + 0.009686821140348911, + 0.012039346620440483, + 0.012701430357992649, + 0.014423007145524025, + 0.01587248034775257, + 0.012124573811888695, + 0.009756217710673809, + 0.013523895293474197, + 0.022783804684877396, + 0.017599059268832207, + 0.009065516293048859, + 0.007637247443199158, + 0.005695147439837456, + 0.005693592131137848, + 0.004511068109422922, + 0.0037205505650490522, + 0.0038854072336107492, + 0.007236693520098925, + 0.010840242728590965, + 0.012007387354969978, + 0.013036999851465225, + 0.022171474993228912, + 0.009913153015077114, + 0.020031901076436043, + 0.019127611070871353, + 0.021011225879192352, + 0.0062811789102852345, + 0.008326439186930656, + 0.012638213112950325, + 0.008201313205063343, + 0.008226875215768814, + 0.0121535649523139, + 0.012966331094503403, + 0.012861868366599083, + 0.006177590228617191, + 0.007577965501695871, + 0.009267779998481274, + 0.010151649825274944, + 0.008497876115143299, + 0.01000986434519291, + 0.011695021763443947, + 0.012933366000652313, + 0.01041710376739502, + 0.012174602597951889, + 0.014901572838425636, + 0.013806046918034554, + 0.011242067441344261, + 0.01906430721282959, + 0.013747730292379856, + 0.02265811525285244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.6328083276748657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029878467321396 + ], + "opposite_logprob_tokens": [ + -2.3502988815307617 + ], + "logprob_reward": -0.10029878467321396, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029878467321396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11788678914308548, + 0.10329359024763107, + 0.08171115815639496, + 0.06324389576911926, + 0.03873709961771965, + 0.0926297977566719, + 0.24276995658874512, + 0.142926424741745, + 0.07765990495681763, + 0.03914140909910202 + ], + "credit_map": [ + 0.01342023629695177, + 0.013589669018983841, + 0.011278966441750526, + 0.015125380828976631, + 0.018042655661702156, + 0.013526835478842258, + 0.010360008105635643, + 0.022543029859662056, + 0.006539380177855492, + 0.007172468584030867, + 0.014066088944673538, + 0.0133750569075346, + 0.013647914864122868, + 0.01663777604699135, + 0.02072881907224655, + 0.011126085184514523, + 0.007679057773202658, + 0.0094635970890522, + 0.006312082055956125, + 0.009186617098748684, + 0.00998011790215969, + 0.011123844422399998, + 0.014813159592449665, + 0.013152684085071087, + 0.0073346905410289764, + 0.00668244156986475, + 0.006629582028836012, + 0.008938148617744446, + 0.005837710574269295, + 0.010626805946230888, + 0.0091093759983778, + 0.008085140958428383, + 0.0063180322758853436, + 0.0037313555367290974, + 0.004247015342116356, + 0.0041661858558654785, + 0.0030094909016042948, + 0.005428058560937643, + 0.003823368577286601, + 0.0080135902389884, + 0.01752154529094696, + 0.01224987581372261, + 0.009985188022255898, + 0.009488831274211407, + 0.008148408494889736, + 0.009330837987363338, + 0.00785803608596325, + 0.018047070130705833, + 0.020481975749135017, + 0.024835173040628433, + 0.02708720974624157, + 0.043017540127038956, + 0.017822861671447754, + 0.037417151033878326, + 0.03288201615214348, + 0.03922603651881218, + 0.010697648860514164, + 0.020388614386320114, + 0.020603174343705177, + 0.021304601803421974, + 0.014499911107122898, + 0.014841470867395401, + 0.021505335345864296, + 0.019085660576820374, + 0.005747689865529537, + 0.008506127633154392, + 0.011664576828479767, + 0.010929127223789692, + 0.008076891303062439, + 0.00814114511013031, + 0.011986645869910717, + 0.012607702985405922, + 0.004079180769622326, + 0.004558296408504248, + 0.0058036199770867825, + 0.0038940017111599445, + 0.003243349026888609, + 0.006794218439608812, + 0.004776639398187399, + 0.005992102436721325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.10029878467321396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth tracking shot, viewed from a slightly upward angle, follows a giant octopus tentacle undergoing a breathtaking process of petrification. Against the deep, almost pitch-black seafloor, a precise, cool white beam, reminiscent of a deep-sea probe, illuminates this surreal transformation. The shot begins at the still-vibrant end of the tentacle. Its deep purple skin, coated in a slippery, iridescent sheen, reflects in the beam, while its neatly arranged, massive suction cups undergo their final, delicate contractions. As the camera pans steadily to the right, we witness the core of the transformation: the slippery sheen fades at a visible rate, the elastic texture hardens inch by inch, and the deep purple skin rapidly pales, transforming into a lifeless, rocky gray. Countless hairline cracks, like spiderwebs, spread from the hardened surface, rapidly deepening and widening, forming deep grooves, as if ancient rock has been fractured by weathering. The once-soft suction cups have solidified into rigid, stone disks, devoid of all signs of life. The entire process is presented in extreme details in the CGI realistic style. The suspended particles in the water are clearly visible in the beam of light, and the light and shadow form a sharp contrast between the rough rock cracks and the smooth wet skin, creating a magnificent, strange and breathtakingly quiet atmosphere.", + "seed": 25291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_5/index_4960_seed_25291_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.29426494278013704, + "vlm_logprob": -0.29426494278013704, + "vlm_yes_no_margin": 2.2750000238418577, + "vlm_token_credit": -0.29426494278013704, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14548435807228088 + ], + "opposite_logprob_tokens": [ + -2.020484447479248 + ], + "logprob_reward": -0.14548435807228088, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14548435807228088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08077474683523178, + 0.06139243394136429, + 0.0818033441901207, + 0.035166651010513306, + 0.07326240837574005, + 0.35592883825302124, + 0.028137993067502975, + 0.06648043543100357, + 0.1863929033279419, + 0.030660199001431465 + ], + "credit_map": [ + 0.009526398964226246, + 0.012068088166415691, + 0.015465764328837395, + 0.008735611103475094, + 0.00971835944801569, + 0.006265436764806509, + 0.008765069767832756, + 0.010230029001832008, + 0.006595495622605085, + 0.007823809050023556, + 0.008426661603152752, + 0.010128544643521309, + 0.0047077275812625885, + 0.006867336109280586, + 0.00712777953594923, + 0.009715087711811066, + 0.008507312275469303, + 0.008313252590596676, + 0.007902022451162338, + 0.00766275729984045, + 0.010823950171470642, + 0.012674358673393726, + 0.010990414768457413, + 0.014929277822375298, + 0.0030385153368115425, + 0.0033773123286664486, + 0.0038336627185344696, + 0.005181347019970417, + 0.00461954390630126, + 0.003757755970582366, + 0.007633664179593325, + 0.003724853741005063, + 0.012193008325994015, + 0.009126298129558563, + 0.006627770606428385, + 0.007808992173522711, + 0.006580764893442392, + 0.007011443842202425, + 0.007725228555500507, + 0.016188906505703926, + 0.051557812839746475, + 0.03294770047068596, + 0.018595127388834953, + 0.023874888196587563, + 0.1335405707359314, + 0.0340256504714489, + 0.030293412506580353, + 0.03109372779726982, + 0.003556972136721015, + 0.0037957539316266775, + 0.004253862425684929, + 0.0026205924805253744, + 0.003219045465812087, + 0.0037210434675216675, + 0.00297077395953238, + 0.003999951761215925, + 0.007301022298634052, + 0.010176128707826138, + 0.006350097246468067, + 0.006624716334044933, + 0.009745520539581776, + 0.006801579147577286, + 0.008047854527831078, + 0.011433526873588562, + 0.021335452795028687, + 0.022140949964523315, + 0.02118329145014286, + 0.023036696016788483, + 0.01641780138015747, + 0.019868534058332443, + 0.03244467452168465, + 0.029965519905090332, + 0.0028155564796179533, + 0.0031241094693541527, + 0.0019350291695445776, + 0.002797300461679697, + 0.0023530314210802317, + 0.003211989998817444, + 0.0068078539334237576, + 0.007615331560373306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.14548435807228088 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211754314601421 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211754314601421, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211754314601421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09775029122829437, + 0.08554621785879135, + 0.10615731030702591, + 0.022077767178416252, + 0.05672153830528259, + 0.058521170169115067, + 0.021087106317281723, + 0.048266008496284485, + 0.3812008500099182, + 0.1226717159152031 + ], + "credit_map": [ + 0.011086453683674335, + 0.019168881699442863, + 0.013124089688062668, + 0.010399295948445797, + 0.009167906828224659, + 0.010998963378369808, + 0.010380267165601254, + 0.013424435630440712, + 0.010411862283945084, + 0.009039572440087795, + 0.01542326994240284, + 0.014724311418831348, + 0.007687695324420929, + 0.009223518893122673, + 0.009154808707535267, + 0.00988117977976799, + 0.012885934673249722, + 0.010923504829406738, + 0.01101136114448309, + 0.011017762124538422, + 0.013679171912372112, + 0.015368318185210228, + 0.014004058204591274, + 0.01726720482110977, + 0.0028848950751125813, + 0.0023670161608606577, + 0.003264301223680377, + 0.0031218293588608503, + 0.0024022753350436687, + 0.0021303778048604727, + 0.0028806559275835752, + 0.003026416990906, + 0.011142712086439133, + 0.009548872709274292, + 0.004335592966526747, + 0.004943992011249065, + 0.0060033854097127914, + 0.0064454772509634495, + 0.006905459798872471, + 0.00739605026319623, + 0.0075575509108603, + 0.0059091439470648766, + 0.006511632818728685, + 0.0059559401124715805, + 0.009252151474356651, + 0.009367568418383598, + 0.006310232449322939, + 0.0076569486409425735, + 0.0035658576525747776, + 0.002661691978573799, + 0.0028299004770815372, + 0.0019396074349060655, + 0.002966691507026553, + 0.003134249709546566, + 0.001755082281306386, + 0.0022340258583426476, + 0.005430977791547775, + 0.006487981881946325, + 0.0060594696551561356, + 0.005942882038652897, + 0.0048263901844620705, + 0.00422407453879714, + 0.006932220887392759, + 0.00836200825870037, + 0.04983469471335411, + 0.05094014108181, + 0.06614869832992554, + 0.0430692620575428, + 0.02453945018351078, + 0.035627033561468124, + 0.0433754026889801, + 0.06766615062952042, + 0.01725868321955204, + 0.020666681230068207, + 0.007594867143779993, + 0.009935255162417889, + 0.007434002123773098, + 0.013426466844975948, + 0.03446108475327492, + 0.01189467404037714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.014211754314601421 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202586375176907 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202586375176907, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202586375176907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042863402515649796, + 0.027490945532917976, + 0.05434773489832878, + 0.029443027451634407, + 0.13671515882015228, + 0.05628109723329544, + 0.025545768439769745, + 0.07849191874265671, + 0.5092280507087708, + 0.03959288075566292 + ], + "credit_map": [ + 0.005038409493863583, + 0.006790724117308855, + 0.006718134507536888, + 0.004745276179164648, + 0.003797470824792981, + 0.003993426449596882, + 0.005482157692313194, + 0.006297809071838856, + 0.003358864923939109, + 0.003787585301324725, + 0.004747987724840641, + 0.00474645895883441, + 0.0028312592767179012, + 0.002728847088292241, + 0.002452726708725095, + 0.0028372181113809347, + 0.006120580248534679, + 0.006678507663309574, + 0.005295792128890753, + 0.004714352078735828, + 0.005832303781062365, + 0.006889584008604288, + 0.006438749376684427, + 0.012377866543829441, + 0.0029037813656032085, + 0.003568904008716345, + 0.0026065425481647253, + 0.003318773116916418, + 0.003030257299542427, + 0.0038696362171322107, + 0.007254514843225479, + 0.002890619682148099, + 0.023895995691418648, + 0.020490625873208046, + 0.010494271293282509, + 0.011923544108867645, + 0.010683787986636162, + 0.01257940661162138, + 0.013861501589417458, + 0.032786037772893906, + 0.008028827607631683, + 0.008353271521627903, + 0.004274859558790922, + 0.0067364792339503765, + 0.005110111087560654, + 0.0066675362177193165, + 0.006907944101840258, + 0.01020207442343235, + 0.00319485180079937, + 0.0030012251809239388, + 0.0035822864156216383, + 0.002693305956199765, + 0.003433858510106802, + 0.003213301533833146, + 0.00278608244843781, + 0.0036408579908311367, + 0.008018743246793747, + 0.008566804230213165, + 0.00822211243212223, + 0.009701552800834179, + 0.010083560831844807, + 0.008147391490638256, + 0.012788159772753716, + 0.012963601388037205, + 0.06964408606290817, + 0.06344088912010193, + 0.09364854544401169, + 0.05535094439983368, + 0.030899805948138237, + 0.04658818244934082, + 0.05671346187591553, + 0.09294220805168152, + 0.0038046855479478836, + 0.004248659126460552, + 0.0024659063201397657, + 0.003795769065618515, + 0.003548206528648734, + 0.003844971302896738, + 0.009570407681167126, + 0.008314280770719051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.014202586375176907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370443105697632 + ], + "opposite_logprob_tokens": [ + -0.3870442807674408 + ], + "logprob_reward": -1.1370443105697632, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370443105697632, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.052404046058654785, + 0.0801401361823082, + 0.1172565147280693, + 0.03113742545247078, + 0.08653507381677628, + 0.0535343736410141, + 0.028610769659280777, + 0.021968692541122437, + 0.22702009975910187, + 0.30139291286468506 + ], + "credit_map": [ + 0.009092613123357296, + 0.004921833518892527, + 0.006031626835465431, + 0.003480802057310939, + 0.0038978999946266413, + 0.007054519839584827, + 0.0068991039879620075, + 0.011025646701455116, + 0.007914009504020214, + 0.009250261820852757, + 0.008005616255104542, + 0.011831972748041153, + 0.0066084000281989574, + 0.011514278128743172, + 0.010468665510416031, + 0.014546926133334637, + 0.019604891538619995, + 0.023584479466080666, + 0.007945294491946697, + 0.01127791590988636, + 0.00964437797665596, + 0.015028915368020535, + 0.012478339485824108, + 0.01769229955971241, + 0.0034175137989223003, + 0.0033670240081846714, + 0.0028330693021416664, + 0.004164376296103001, + 0.004520817659795284, + 0.003204232780262828, + 0.006040997803211212, + 0.003589395899325609, + 0.014636987820267677, + 0.012715177610516548, + 0.006249893922358751, + 0.008293069899082184, + 0.008515891619026661, + 0.0073723699897527695, + 0.011342165991663933, + 0.017409516498446465, + 0.0068766530603170395, + 0.004286850802600384, + 0.0034014869015663862, + 0.0045613860711455345, + 0.018243838101625443, + 0.004600421525537968, + 0.005356764420866966, + 0.006206975784152746, + 0.004147766157984734, + 0.00427566934376955, + 0.004410943482071161, + 0.002740244846791029, + 0.003435042453929782, + 0.0038098045624792576, + 0.0030059791170060635, + 0.002785319462418556, + 0.0035798391327261925, + 0.0019205108983442187, + 0.003426535753533244, + 0.0025003543123602867, + 0.002032943069934845, + 0.0024570957757532597, + 0.003003796562552452, + 0.0030476178508251905, + 0.03296591714024544, + 0.027647338807582855, + 0.047922294586896896, + 0.026837347075343132, + 0.013668889179825783, + 0.019135545939207077, + 0.018350543454289436, + 0.040492210537195206, + 0.02962627448141575, + 0.032929375767707825, + 0.011612948030233383, + 0.014128466136753559, + 0.015737483277916908, + 0.02411457523703575, + 0.1026381328701973, + 0.07060565054416656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -1.1370443105697632 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038170456886292 + ], + "opposite_logprob_tokens": [ + -1.9103816747665405 + ], + "logprob_reward": -0.16038170456886292, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038170456886292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09446284919977188, + 0.04387569800019264, + 0.08372657746076584, + 0.09901118278503418, + 0.18281881511211395, + 0.20295190811157227, + 0.04719962552189827, + 0.012422286905348301, + 0.17312265932559967, + 0.06040842458605766 + ], + "credit_map": [ + 0.011814018711447716, + 0.011969943530857563, + 0.01651259884238243, + 0.015538564883172512, + 0.011248805560171604, + 0.006809740327298641, + 0.010486912913620472, + 0.010082263499498367, + 0.006249533034861088, + 0.005154559388756752, + 0.0063138618133962154, + 0.006580425426363945, + 0.003650262486189604, + 0.005266937892884016, + 0.00519372196868062, + 0.0054663997143507, + 0.008887180127203465, + 0.02449152059853077, + 0.004328430164605379, + 0.005849150940775871, + 0.0049800933338701725, + 0.007921342737972736, + 0.00576004246249795, + 0.02150881476700306, + 0.01026599109172821, + 0.00972829107195139, + 0.010319370776414871, + 0.011529485695064068, + 0.007895409129559994, + 0.011274420656263828, + 0.028848804533481598, + 0.00914941169321537, + 0.03568368777632713, + 0.026398582383990288, + 0.013373380526900291, + 0.012789668515324593, + 0.012639408931136131, + 0.016678428277373314, + 0.01774357073009014, + 0.04751208797097206, + 0.032595690339803696, + 0.015003182925283909, + 0.01117466576397419, + 0.014343546703457832, + 0.06928792595863342, + 0.022592002525925636, + 0.01891198940575123, + 0.019042914733290672, + 0.009044747799634933, + 0.005647844634950161, + 0.009410036727786064, + 0.004080739803612232, + 0.00443589361384511, + 0.005575188901275396, + 0.004439385142177343, + 0.004565788432955742, + 0.001683316077105701, + 0.0012096200371161103, + 0.0016335349064320326, + 0.0014141610590741038, + 0.0016163189429789782, + 0.0013889619149267673, + 0.0014475586358457804, + 0.0020288159139454365, + 0.023658381775021553, + 0.02092602103948593, + 0.031684935092926025, + 0.018905485048890114, + 0.010486857034265995, + 0.01574743539094925, + 0.020385637879371643, + 0.031327906996011734, + 0.004902197513729334, + 0.0056499880738556385, + 0.0032903451938182116, + 0.004589622840285301, + 0.004164497368037701, + 0.006385996006429195, + 0.015251991339027882, + 0.01617378741502762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.16038170456886292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "slow, smooth tracking shot, viewed from a slightly upward angle, follows a giant octopus tentacle undergoing a breathtaking process of petrification. Against the deep, almost pitch-black seafloor, a precise, cool white beam, reminiscent of a deep-sea probe, illuminates this surreal transformation. The shot begins at the still-vibrant end of the tentacle. Its deep purple skin, coated in a slippery, iridescent sheen, reflects in the beam, while its neatly arranged, massive suction cups undergo their final, delicate contractions. As the camera pans steadily to the right, we witness the core of the transformation: the slippery sheen fades at a visible rate, the elastic texture hardens inch by inch, and the deep purple skin rapidly pales, transforming into a lifeless, rocky gray. Countless hairline cracks, like spiderwebs, spread from the hardened surface, rapidly deepening and widening, forming deep grooves, as if ancient rock has been fractured by weathering. The once-soft suction cups have solidified into rigid, stone disks, devoid of all signs of life. The entire process is presented in extreme details in the CGI realistic style. The suspended particles in the water are clearly visible in the beam of light, and the light and shadow form a sharp contrast between the rough rock cracks and the smooth wet skin, creating a magnificent, strange and breathtakingly quiet atmosphere.", + "seed": 25292, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_5/index_4960_seed_25292_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0961501158773899, + "vlm_logprob": -0.0961501158773899, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.0961501158773899, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11600104719400406 + ], + "opposite_logprob_tokens": [ + -2.2410011291503906 + ], + "logprob_reward": -0.11600104719400406, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11600104719400406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05253956466913223, + 0.4389040470123291, + 0.03364289551973343, + 0.05561113357543945, + 0.051641255617141724, + 0.03650175407528877, + 0.042287979274988174, + 0.1738017052412033, + 0.025625552982091904, + 0.08944416046142578 + ], + "credit_map": [ + 0.005306857638061047, + 0.005438405554741621, + 0.006117676384747028, + 0.007493882440030575, + 0.005390535108745098, + 0.007043667137622833, + 0.007720818743109703, + 0.00802772119641304, + 0.17201314866542816, + 0.015193536877632141, + 0.17200030386447906, + 0.013210069388151169, + 0.017466414719820023, + 0.005162106826901436, + 0.036994513124227524, + 0.006863959599286318, + 0.0066720424219965935, + 0.002905597910284996, + 0.005061679519712925, + 0.003459497820585966, + 0.0028913295827805996, + 0.006509432103484869, + 0.003163715125992894, + 0.0029796005692332983, + 0.00680163037031889, + 0.005442322231829166, + 0.010430525988340378, + 0.005144729278981686, + 0.005128440912812948, + 0.006327579263597727, + 0.007853442803025246, + 0.008482463657855988, + 0.0054506040178239346, + 0.0037734133657068014, + 0.0041200825944542885, + 0.003637738060206175, + 0.005023903213441372, + 0.008431880734860897, + 0.005891437642276287, + 0.0153121929615736, + 0.003909537568688393, + 0.0028482920024544, + 0.004640503320842981, + 0.004782768432050943, + 0.006368258967995644, + 0.00557069294154644, + 0.004225128795951605, + 0.0041565727442502975, + 0.005334288813173771, + 0.009258880279958248, + 0.005822021048516035, + 0.0052318815141916275, + 0.0049511827528476715, + 0.0028005384374409914, + 0.004282484762370586, + 0.004606703761965036, + 0.014397972263395786, + 0.0066589172929525375, + 0.06257478147745132, + 0.019738275557756424, + 0.011916769668459892, + 0.0066442773677408695, + 0.013491155579686165, + 0.038379549980163574, + 0.0034660017117857933, + 0.0021933221723884344, + 0.003381793387234211, + 0.004379891790449619, + 0.0019384160405024886, + 0.0017919979291036725, + 0.00499240355566144, + 0.00348172546364367, + 0.006866809446364641, + 0.007317568175494671, + 0.006635277532041073, + 0.011009070090949535, + 0.023441696539521217, + 0.005345731042325497, + 0.007072353269904852, + 0.021755659952759743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.11600104719400406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793920442461967 + ], + "opposite_logprob_tokens": [ + -4.634793758392334 + ], + "logprob_reward": -0.009793920442461967, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793920442461967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14074629545211792, + 0.3611663579940796, + 0.03004693239927292, + 0.04965304955840111, + 0.0452924519777298, + 0.15107806026935577, + 0.042776286602020264, + 0.05533966049551964, + 0.02012566290795803, + 0.10377519577741623 + ], + "credit_map": [ + 0.01733936183154583, + 0.01550134178251028, + 0.020706111565232277, + 0.024522805586457253, + 0.010479608550667763, + 0.016622819006443024, + 0.017678774893283844, + 0.017895476892590523, + 0.13847702741622925, + 0.012934943661093712, + 0.14325205981731415, + 0.011520974338054657, + 0.014037060551345348, + 0.004062490072101355, + 0.031208399683237076, + 0.005673411302268505, + 0.0050008464604616165, + 0.002443708712235093, + 0.003530407091602683, + 0.0024561297614127398, + 0.0037828381173312664, + 0.005478792358189821, + 0.0033267089165747166, + 0.004027501679956913, + 0.005865833722054958, + 0.0035783315543085337, + 0.017459461465477943, + 0.003155503189191222, + 0.005308627616614103, + 0.005540498532354832, + 0.002825801493600011, + 0.005918992683291435, + 0.005123653914779425, + 0.0029601913411170244, + 0.005058939568698406, + 0.0036783399991691113, + 0.006108352914452553, + 0.006902033928781748, + 0.005183663684874773, + 0.010277279652655125, + 0.029723789542913437, + 0.011863894760608673, + 0.018832866102457047, + 0.020393837243318558, + 0.023323707282543182, + 0.019637910649180412, + 0.01527880597859621, + 0.012023258022964, + 0.0034757000394165516, + 0.002869382966309786, + 0.007046958431601524, + 0.004241274204105139, + 0.009060116484761238, + 0.005692016799002886, + 0.006070289760828018, + 0.004320548847317696, + 0.005249211098998785, + 0.0023961623664945364, + 0.021487116813659668, + 0.009224151261150837, + 0.003488427260890603, + 0.0019461518386378884, + 0.0028688611928373575, + 0.008679580874741077, + 0.0030801394023001194, + 0.002424597041681409, + 0.0023816495668143034, + 0.002820861293002963, + 0.0017369159031659365, + 0.0017454291228204966, + 0.0026531333569437265, + 0.00328293745405972, + 0.007882293313741684, + 0.005353378131985664, + 0.009379786439239979, + 0.005605081096291542, + 0.026872146874666214, + 0.005929829552769661, + 0.009421490132808685, + 0.033331193029880524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.009793920442461967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635420672595501 + ], + "opposite_logprob_tokens": [ + -3.6513543128967285 + ], + "logprob_reward": -0.02635420672595501, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635420672595501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09212248772382736, + 0.24457813799381256, + 0.04108380153775215, + 0.0758267194032669, + 0.07104460150003433, + 0.20612429082393646, + 0.047119442373514175, + 0.13525845110416412, + 0.026071032509207726, + 0.060771021991968155 + ], + "credit_map": [ + 0.01138576865196228, + 0.00943820271641016, + 0.018890807405114174, + 0.014460654929280281, + 0.006586186587810516, + 0.007531306240707636, + 0.013817139901220798, + 0.010012416169047356, + 0.08984962850809097, + 0.010059509426355362, + 0.09380527585744858, + 0.009268631227314472, + 0.00929501373320818, + 0.005813681520521641, + 0.021579550579190254, + 0.004906846676021814, + 0.006671890150755644, + 0.004016718827188015, + 0.005095344502478838, + 0.0036137334536761045, + 0.005609897430986166, + 0.0052054645493626595, + 0.005081559531390667, + 0.005789192859083414, + 0.010116768069565296, + 0.00522337993606925, + 0.026866955682635307, + 0.004878268577158451, + 0.007442265283316374, + 0.0084611801430583, + 0.004851181525737047, + 0.007986721582710743, + 0.007793318945914507, + 0.00504810968413949, + 0.007661298383027315, + 0.00578330596908927, + 0.00841518584638834, + 0.01078806258738041, + 0.009386797435581684, + 0.01616852916777134, + 0.03840741887688637, + 0.015665234997868538, + 0.026973357424139977, + 0.025936689227819443, + 0.031549300998449326, + 0.026532268151640892, + 0.02409331128001213, + 0.01696673408150673, + 0.00713677192106843, + 0.004753755871206522, + 0.006779559887945652, + 0.0059470548294484615, + 0.005492322146892548, + 0.005851864814758301, + 0.0050002820789813995, + 0.006157838273793459, + 0.010452628135681152, + 0.0073011331260204315, + 0.01768890954554081, + 0.030330782756209373, + 0.008593079634010792, + 0.006412082817405462, + 0.008623114787042141, + 0.04585673660039902, + 0.0037504362408071756, + 0.003198224352672696, + 0.002943694358691573, + 0.0030586891807615757, + 0.003705981420353055, + 0.0027710748836398125, + 0.0037904512137174606, + 0.002852482721209526, + 0.005167720839381218, + 0.005472984630614519, + 0.008199475705623627, + 0.006480759475380182, + 0.01630725897848606, + 0.00445995107293129, + 0.00789195392280817, + 0.006790924817323685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.02635420672595501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156458020210266 + ], + "opposite_logprob_tokens": [ + -1.7015645503997803 + ], + "logprob_reward": -0.20156458020210266, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156458020210266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12572331726551056, + 0.14091961085796356, + 0.03367239981889725, + 0.06296446174383163, + 0.07763268798589706, + 0.20348972082138062, + 0.05681784451007843, + 0.12733523547649384, + 0.11633172631263733, + 0.055112969130277634 + ], + "credit_map": [ + 0.014830527827143669, + 0.013930954039096832, + 0.02294471487402916, + 0.019686227664351463, + 0.00925584975630045, + 0.015605490654706955, + 0.016347186639904976, + 0.013122373260557652, + 0.05153517425060272, + 0.007125639356672764, + 0.04878964647650719, + 0.007426506374031305, + 0.005889585707336664, + 0.004704511258751154, + 0.011242243461310863, + 0.004206301644444466, + 0.005900053773075342, + 0.0036996640264987946, + 0.0037483975756913424, + 0.0027645875234156847, + 0.004712975118309259, + 0.004469548352062702, + 0.0038441664073616266, + 0.0045330035500228405, + 0.008539034985005856, + 0.004464426543563604, + 0.02049621008336544, + 0.004088659305125475, + 0.006660356651991606, + 0.006811659783124924, + 0.0039509679190814495, + 0.007953147403895855, + 0.007997809909284115, + 0.006961611099541187, + 0.006233775056898594, + 0.006062911823391914, + 0.008760252967476845, + 0.01264834962785244, + 0.009359610266983509, + 0.01960836537182331, + 0.038382671773433685, + 0.015991196036338806, + 0.025904573500156403, + 0.028258154168725014, + 0.03306952491402626, + 0.025298940017819405, + 0.019397934898734093, + 0.01718672178685665, + 0.005779691506177187, + 0.00665433332324028, + 0.008431578055024147, + 0.00708612659946084, + 0.008860853500664234, + 0.007140083238482475, + 0.00666389474645257, + 0.006201283074915409, + 0.01032107975333929, + 0.003091595135629177, + 0.04553312063217163, + 0.0077932411804795265, + 0.0062241991981863976, + 0.004956733901053667, + 0.007212204392999411, + 0.04220305755734444, + 0.009734549559652805, + 0.0050950320437550545, + 0.006112006492912769, + 0.007464290596544743, + 0.0108342245221138, + 0.006109476089477539, + 0.042448267340660095, + 0.028533879667520523, + 0.004759807139635086, + 0.0038272596430033445, + 0.008227263577282429, + 0.0035125846043229103, + 0.007689837832003832, + 0.0027662517968565226, + 0.006372497882694006, + 0.017957467585802078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.20156458020210266 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703682482242584 + ], + "opposite_logprob_tokens": [ + -2.1270368099212646 + ], + "logprob_reward": -0.12703682482242584, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703682482242584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04526439309120178, + 0.5513795018196106, + 0.06171627715229988, + 0.03377220779657364, + 0.061314262449741364, + 0.026196753606200218, + 0.07040144503116608, + 0.01959267258644104, + 0.09021849930286407, + 0.040143974125385284 + ], + "credit_map": [ + 0.006519740447402, + 0.004402903839945793, + 0.006946423090994358, + 0.006018040701746941, + 0.0038765673525631428, + 0.005536450073122978, + 0.004766437690705061, + 0.007197830360382795, + 0.2149600237607956, + 0.01897016167640686, + 0.21578098833560944, + 0.017186274752020836, + 0.02292633056640625, + 0.006516436580568552, + 0.046785783022642136, + 0.008253449574112892, + 0.01139928214251995, + 0.005695655010640621, + 0.00547699024900794, + 0.005723176058381796, + 0.007464682217687368, + 0.013859146274626255, + 0.005411878693848848, + 0.006685466505587101, + 0.005881432443857193, + 0.002750706858932972, + 0.007520908955484629, + 0.0028094712179154158, + 0.0028078679461032152, + 0.003958380315452814, + 0.0037614754401147366, + 0.004281963221728802, + 0.007958120666444302, + 0.003739147214218974, + 0.004208723083138466, + 0.00484274560585618, + 0.0059611499309539795, + 0.009145321324467659, + 0.007659106981009245, + 0.017799949273467064, + 0.005161885637789965, + 0.0022370994556695223, + 0.0035967431031167507, + 0.0035800421610474586, + 0.003687547752633691, + 0.003350684652104974, + 0.0024049312341958284, + 0.0021778177469968796, + 0.009536052122712135, + 0.007570917718112469, + 0.011538378894329071, + 0.009621295146644115, + 0.00842132605612278, + 0.008796293288469315, + 0.00808869767934084, + 0.006828484125435352, + 0.0035636869724839926, + 0.0022817363496869802, + 0.00368986907415092, + 0.0023470318410545588, + 0.0020334767177700996, + 0.0011969348415732384, + 0.0025996530894190073, + 0.0018802837003022432, + 0.007672187406569719, + 0.003598939860239625, + 0.004807999823242426, + 0.006806802004575729, + 0.006671720650047064, + 0.0042983791790902615, + 0.03367428481578827, + 0.022688183933496475, + 0.003901619231328368, + 0.003931049723178148, + 0.003882140852510929, + 0.0025096344761550426, + 0.012898964807391167, + 0.0034269862808287144, + 0.002919475082308054, + 0.006674101576209068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.12703682482242584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth tracking shot, viewed from a slightly upward angle, follows a giant octopus tentacle undergoing a breathtaking process of petrification. Against the deep, almost pitch-black seafloor, a precise, cool white beam, reminiscent of a deep-sea probe, illuminates this surreal transformation. The shot begins at the still-vibrant end of the tentacle. Its deep purple skin, coated in a slippery, iridescent sheen, reflects in the beam, while its neatly arranged, massive suction cups undergo their final, delicate contractions. As the camera pans steadily to the right, we witness the core of the transformation: the slippery sheen fades at a visible rate, the elastic texture hardens inch by inch, and the deep purple skin rapidly pales, transforming into a lifeless, rocky gray. Countless hairline cracks, like spiderwebs, spread from the hardened surface, rapidly deepening and widening, forming deep grooves, as if ancient rock has been fractured by weathering. The once-soft suction cups have solidified into rigid, stone disks, devoid of all signs of life. The entire process is presented in extreme details in the CGI realistic style. The suspended particles in the water are clearly visible in the beam of light, and the light and shadow form a sharp contrast between the rough rock cracks and the smooth wet skin, creating a magnificent, strange and breathtakingly quiet atmosphere.", + "seed": 25293, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_5/index_4960_seed_25293_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19353778464719654, + "vlm_logprob": -0.19353778464719654, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.19353778464719654, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059539537876844406 + ], + "opposite_logprob_tokens": [ + -2.934539556503296 + ], + "logprob_reward": -0.059539537876844406, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.059539537876844406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25419124960899353, + 0.05942164734005928, + 0.025797415524721146, + 0.1551058292388916, + 0.020870590582489967, + 0.018555982038378716, + 0.03648846223950386, + 0.10324770212173462, + 0.2833459675312042, + 0.04297509789466858 + ], + "credit_map": [ + 0.004841027315706015, + 0.004702379927039146, + 0.007335772272199392, + 0.02702382020652294, + 0.009355037473142147, + 0.010464276187121868, + 0.10574217140674591, + 0.08472677320241928, + 0.004982288461178541, + 0.006165842991322279, + 0.008210769854485989, + 0.008851135149598122, + 0.005651799961924553, + 0.008558680303394794, + 0.008425370790064335, + 0.008575765416026115, + 0.0014216743875294924, + 0.0021540431771427393, + 0.002045594621449709, + 0.003878998337313533, + 0.0014475365169346333, + 0.0019982291851192713, + 0.00600328529253602, + 0.006848055403679609, + 0.019305631518363953, + 0.012644424103200436, + 0.020585963502526283, + 0.02140999399125576, + 0.014710146002471447, + 0.01932520978152752, + 0.024897079914808273, + 0.022227376699447632, + 0.0019572137389332056, + 0.002382043981924653, + 0.0033080538269132376, + 0.002306149573996663, + 0.0035605772864073515, + 0.002255512634292245, + 0.002360390964895487, + 0.0027406502049416304, + 0.0016746819019317627, + 0.001742318389005959, + 0.0030151247046887875, + 0.002064000815153122, + 0.0033711164724081755, + 0.0020377205219119787, + 0.0022131118457764387, + 0.0024379068054258823, + 0.00428791856393218, + 0.0042380946688354015, + 0.004423229955136776, + 0.005327445454895496, + 0.003962634596973658, + 0.006050009746104479, + 0.00450589694082737, + 0.003693235106766224, + 0.008299562148749828, + 0.0060605332255363464, + 0.012426408939063549, + 0.009840177372097969, + 0.007227628491818905, + 0.024923304095864296, + 0.007386102806776762, + 0.027084000408649445, + 0.015214727260172367, + 0.011112802661955357, + 0.014595163986086845, + 0.015543609857559204, + 0.02603832818567753, + 0.09840472042560577, + 0.08170352876186371, + 0.020733118057250977, + 0.006241547409445047, + 0.005878611467778683, + 0.004890122916549444, + 0.005465518217533827, + 0.004639769438654184, + 0.006191886495798826, + 0.004508107900619507, + 0.0051595368422567844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.059539537876844406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052668387070298195 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.0052668387070298195, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052668387070298195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34091392159461975, + 0.08643655478954315, + 0.04218491539359093, + 0.11944884061813354, + 0.03203388303518295, + 0.04166343808174133, + 0.08101105690002441, + 0.06921638548374176, + 0.12721635401248932, + 0.059874724596738815 + ], + "credit_map": [ + 0.009419282898306847, + 0.011379019357264042, + 0.008228464983403683, + 0.03900429233908653, + 0.013416565023362637, + 0.011504503898322582, + 0.13992998003959656, + 0.10803180187940598, + 0.006723991595208645, + 0.007417266257107258, + 0.011605463922023773, + 0.012867455370724201, + 0.007362856063991785, + 0.010926512070000172, + 0.017464647069573402, + 0.012068357318639755, + 0.0024795231875032187, + 0.002824472263455391, + 0.005885469727218151, + 0.005307015031576157, + 0.003573330817744136, + 0.0026325732469558716, + 0.008792072534561157, + 0.010690450668334961, + 0.014230797067284584, + 0.011640071868896484, + 0.015330368652939796, + 0.013727402314543724, + 0.01245676726102829, + 0.01438202802091837, + 0.019923776388168335, + 0.017757626250386238, + 0.0029575794469565153, + 0.003943216521292925, + 0.0052744378335773945, + 0.003315614303573966, + 0.004128254950046539, + 0.0026803279761224985, + 0.004136344883590937, + 0.005598105024546385, + 0.004121210426092148, + 0.003933020867407322, + 0.009009559638798237, + 0.003924950025975704, + 0.005529073998332024, + 0.00550412479788065, + 0.004080663900822401, + 0.0055608307011425495, + 0.007084345445036888, + 0.006113744340837002, + 0.01111892145127058, + 0.01090511679649353, + 0.012761889025568962, + 0.01421389076858759, + 0.010333448648452759, + 0.008479690179228783, + 0.00621142378076911, + 0.008012141101062298, + 0.0066605727188289165, + 0.013381261378526688, + 0.009104624390602112, + 0.008922798559069633, + 0.010000619105994701, + 0.006922942120581865, + 0.006161910481750965, + 0.005403675138950348, + 0.007137850858271122, + 0.010174715891480446, + 0.011513812467455864, + 0.042462654411792755, + 0.03434187173843384, + 0.010019846260547638, + 0.009497527964413166, + 0.00793539173901081, + 0.008315101265907288, + 0.008592669852077961, + 0.007161334622651339, + 0.006644467823207378, + 0.006399840582162142, + 0.005328381434082985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0052668387070298195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741635967046022 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741635967046022, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741635967046022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17312948405742645, + 0.026565641164779663, + 0.03780495002865791, + 0.043560098856687546, + 0.021989481523633003, + 0.017667172476649284, + 0.029985904693603516, + 0.11521479487419128, + 0.43562036752700806, + 0.09846211224794388 + ], + "credit_map": [ + 0.003659241134300828, + 0.0035891830921173096, + 0.0043006050400435925, + 0.01744944043457508, + 0.007174579426646233, + 0.006516391411423683, + 0.07102293521165848, + 0.059417109936475754, + 0.002265370450913906, + 0.0025614730548113585, + 0.003381990594789386, + 0.0037814609240740538, + 0.0028633163310587406, + 0.004107796121388674, + 0.004090042784810066, + 0.0035141902044415474, + 0.003952189814299345, + 0.0033346081618219614, + 0.006927244830876589, + 0.005585250444710255, + 0.0033934256061911583, + 0.0036993378307670355, + 0.005280560813844204, + 0.005632330663502216, + 0.006512681487947702, + 0.003467695787549019, + 0.006598757579922676, + 0.00543398130685091, + 0.004125060513615608, + 0.005138366483151913, + 0.006819451227784157, + 0.005464100744575262, + 0.001535162329673767, + 0.001954660750925541, + 0.004504597280174494, + 0.0025067729875445366, + 0.0019521972863003612, + 0.002425903221592307, + 0.00306239677593112, + 0.004047791473567486, + 0.001633145147934556, + 0.001263944897800684, + 0.0030363465193659067, + 0.0014009480364620686, + 0.002808536170050502, + 0.002211576094850898, + 0.0026120985858142376, + 0.002700577490031719, + 0.0030887078028172255, + 0.0020900347735732794, + 0.00432764645665884, + 0.0037758213002234697, + 0.005105690564960241, + 0.004678292665630579, + 0.00376684358343482, + 0.003152866382151842, + 0.009099394083023071, + 0.004540040623396635, + 0.015268629416823387, + 0.008743141777813435, + 0.006629630923271179, + 0.02989726886153221, + 0.006792951375246048, + 0.03424374386668205, + 0.023598123341798782, + 0.016757434234023094, + 0.022188765928149223, + 0.02332346700131893, + 0.041415728628635406, + 0.15190468728542328, + 0.1253553032875061, + 0.031076868996024132, + 0.017561201006174088, + 0.010272336192429066, + 0.011131655424833298, + 0.010583102703094482, + 0.01377095840871334, + 0.010913319885730743, + 0.009797846898436546, + 0.014431687071919441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.006741635967046022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261317610740662 + ], + "opposite_logprob_tokens": [ + -0.5761317610740662 + ], + "logprob_reward": -0.8261317610740662, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261317610740662, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.551265299320221, + 0.15968888998031616, + 0.08491356670856476, + 0.031335506588220596, + 0.04085833579301834, + 0.016372473910450935, + 0.02757914550602436, + 0.04109061136841774, + 0.016483791172504425, + 0.0304123405367136 + ], + "credit_map": [ + 0.007862294092774391, + 0.007469374220818281, + 0.011332708410918713, + 0.052819885313510895, + 0.018164416775107384, + 0.019815830513834953, + 0.23557746410369873, + 0.1982233226299286, + 0.014106867834925652, + 0.013649646192789078, + 0.025229450315237045, + 0.027478862553834915, + 0.012874564155936241, + 0.0209957305341959, + 0.023644665256142616, + 0.021709104999899864, + 0.004209875129163265, + 0.007621314376592636, + 0.0065973359160125256, + 0.010381522588431835, + 0.004303414840251207, + 0.004215202759951353, + 0.0221217293292284, + 0.025463169440627098, + 0.002938838442787528, + 0.004465777426958084, + 0.003941318951547146, + 0.0025236394722014666, + 0.0032479609362781048, + 0.0035670395009219646, + 0.005035729147493839, + 0.00561520317569375, + 0.0036188550293445587, + 0.0052168648689985275, + 0.005481834057718515, + 0.004844094626605511, + 0.005978570785373449, + 0.0036207842640578747, + 0.006535087712109089, + 0.005562246311455965, + 0.0017712204717099667, + 0.0013845866778865457, + 0.0027149072848260403, + 0.0017121820710599422, + 0.002218768699094653, + 0.0019897338934242725, + 0.0021484517492353916, + 0.0024326234124600887, + 0.0018068490317091346, + 0.0024858026299625635, + 0.004251326899975538, + 0.0029513221234083176, + 0.006269195582717657, + 0.0036724251694977283, + 0.003520366968587041, + 0.002621858147904277, + 0.0034388641361147165, + 0.00494669983163476, + 0.003550562309101224, + 0.007122276816517115, + 0.004705681465566158, + 0.006539629772305489, + 0.0065901498310267925, + 0.004196750000119209, + 0.0014432258903980255, + 0.0010715447133406997, + 0.0014897871296852827, + 0.00335404253564775, + 0.0022740066051483154, + 0.003005491802468896, + 0.002693634247407317, + 0.001152058015577495, + 0.003672743448987603, + 0.004216258879750967, + 0.0041825477965176105, + 0.005255617201328278, + 0.00367121584713459, + 0.0034665518905967474, + 0.00321992882527411, + 0.0027274759486317635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.8261317610740662 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant octopus tentacle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tentacle initially deep purple with a slippery, iridescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tentacle undergo a transformation where it hardens into rocky gray stone with hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a precise, cool white beam against a deep, pitch-black seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right following the tentacle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000914961099625 + ], + "opposite_logprob_tokens": [ + -2.695009231567383 + ], + "logprob_reward": -0.07000914961099625, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000914961099625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27673110365867615, + 0.07085069268941879, + 0.10550142824649811, + 0.05518592894077301, + 0.12595541775226593, + 0.14388562738895416, + 0.07921702414751053, + 0.04040305316448212, + 0.0772227868437767, + 0.025046968832612038 + ], + "credit_map": [ + 0.004764696583151817, + 0.005454505328088999, + 0.004751722794026136, + 0.021798638626933098, + 0.010331766679883003, + 0.011552023701369762, + 0.11403129249811172, + 0.10404647141695023, + 0.005603970028460026, + 0.006236713379621506, + 0.010694500990211964, + 0.01201244443655014, + 0.006138950120657682, + 0.009120351634919643, + 0.011257657781243324, + 0.009786110371351242, + 0.008125636726617813, + 0.011352860368788242, + 0.010430723428726196, + 0.014417999424040318, + 0.008535105735063553, + 0.0071276412345469, + 0.019334886223077774, + 0.02617657370865345, + 0.005189937073737383, + 0.004367562010884285, + 0.006944004446268082, + 0.006976277567446232, + 0.005302773788571358, + 0.008481903001666069, + 0.008472939021885395, + 0.009450532495975494, + 0.011434951797127724, + 0.011574978940188885, + 0.02110706828534603, + 0.01129062008112669, + 0.02599526382982731, + 0.01232712808996439, + 0.014612379483878613, + 0.01761302351951599, + 0.016189107671380043, + 0.01433450635522604, + 0.020384298637509346, + 0.015310987830162048, + 0.018894290551543236, + 0.017960425466299057, + 0.01885668933391571, + 0.02195531688630581, + 0.007370307110249996, + 0.0069113667123019695, + 0.011163679882884026, + 0.011128324083983898, + 0.011245752684772015, + 0.015683872625231743, + 0.00797699298709631, + 0.007736730854958296, + 0.006108725443482399, + 0.003015906782820821, + 0.0034027318470180035, + 0.005173457320779562, + 0.003154333448037505, + 0.00756039097905159, + 0.006125739775598049, + 0.005861767567694187, + 0.004820663947612047, + 0.004469342064112425, + 0.0059874411672353745, + 0.010117755271494389, + 0.006088972091674805, + 0.020293492823839188, + 0.019521623849868774, + 0.0059234932996332645, + 0.00291458354331553, + 0.003072503488510847, + 0.003496393095701933, + 0.0037459132727235556, + 0.003625974291935563, + 0.0027802239637821913, + 0.0023536300286650658, + 0.0030577480792999268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.07000914961099625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_tensor([5837, 5837])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_tensor([5837, 5837])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5d7da1402db6fb7c67986b7db3d7982c71c04faf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_tensor([5837, 5837])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, wide-angle, static shot captures a magnificent spectacle beneath a deep, dark night sky. Hundreds of small, sleek drones, their gleaming metallic black shells looming in the darkness, converge at breakneck speed from all directions, even from the depths of the background, towards an invisible focal point in the very center of the frame. Each drone trails a long, luminous digital light trail of electric blue and fiery magenta, carving elegant, intricate arcs through the sky. Their flight paths intertwine, instantly weaving a vast, dynamic web of light across the clear night sky, completely ablaze. The brightly lit futuristic city skyline below provides subtle ambient light, but the primary illumination for the entire scene comes from the drones themselves, whose light reflects off their sleek shells in sharp highlights. The entire image is imbued with a sense of speed, energy, and precise control, creating an atmosphere of grand, high-tech, and dizzying futuristic visual spectacle.", + "seed": 12353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_5837_seed_12353_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2557055061450228, + "vlm_logprob": -0.2557055061450228, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.2557055061450228, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9744238257408142 + ], + "opposite_logprob_tokens": [ + -0.4744238257408142 + ], + "logprob_reward": -0.9744238257408142, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9744238257408142, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09076593071222305, + 0.14104479551315308, + 0.04716639220714569, + 0.09728758037090302, + 0.1707342565059662, + 0.08812572807073593, + 0.04878128692507744, + 0.054320327937603, + 0.08956905454397202, + 0.1722046136856079 + ], + "credit_map": [ + 0.008755474351346493, + 0.013592101633548737, + 0.011601760052144527, + 0.009349166415631771, + 0.019042717292904854, + 0.011433040723204613, + 0.007612358778715134, + 0.009379317052662373, + 0.020973969250917435, + 0.053791314363479614, + 0.039004240185022354, + 0.004844920709729195, + 0.010404807515442371, + 0.004539219196885824, + 0.003873412264510989, + 0.0036129083018749952, + 0.007236437872052193, + 0.0057945274747908115, + 0.0062952651642262936, + 0.004861011169850826, + 0.005097944755107164, + 0.00916494894772768, + 0.003543311497196555, + 0.005172942765057087, + 0.025251353159546852, + 0.007903163321316242, + 0.025126663967967033, + 0.008508353494107723, + 0.007131075952202082, + 0.011767988093197346, + 0.005076013971120119, + 0.006522963754832745, + 0.016564294695854187, + 0.018392272293567657, + 0.03023679554462433, + 0.014878039248287678, + 0.025440659373998642, + 0.029773622751235962, + 0.020299864932894707, + 0.015148709528148174, + 0.013575104996562004, + 0.01139096450060606, + 0.014037709683179855, + 0.008292695507407188, + 0.009215505793690681, + 0.016440339386463165, + 0.0077148014679551125, + 0.007458600681275129, + 0.005696983076632023, + 0.006434859242290258, + 0.006615350488573313, + 0.00575618352741003, + 0.005533146671950817, + 0.007427438162267208, + 0.005906735081225634, + 0.005410592537373304, + 0.005006593186408281, + 0.019146203994750977, + 0.013763596303761005, + 0.002482454292476177, + 0.0038296410348266363, + 0.004299574997276068, + 0.0021986595820635557, + 0.003593608969822526, + 0.014120501466095448, + 0.008342811837792397, + 0.015299797058105469, + 0.008019697852432728, + 0.00626866752281785, + 0.011428424157202244, + 0.012022648006677628, + 0.014066510833799839, + 0.0632154643535614, + 0.06624968349933624, + 0.004765755031257868, + 0.0038444208912551403, + 0.00836141686886549, + 0.017151223495602608, + 0.004216360859572887, + 0.004400276578962803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.9744238257408142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521384358406067 + ], + "opposite_logprob_tokens": [ + -1.502138376235962 + ], + "logprob_reward": -0.2521384358406067, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521384358406067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08063505589962006, + 0.04085296019911766, + 0.0752810686826706, + 0.13268333673477173, + 0.06008141115307808, + 0.07153985649347305, + 0.04325208067893982, + 0.07100632786750793, + 0.09222854673862457, + 0.3324394226074219 + ], + "credit_map": [ + 0.007917249575257301, + 0.010649860836565495, + 0.00996980257332325, + 0.011548969894647598, + 0.012830040417611599, + 0.008839743211865425, + 0.008186903782188892, + 0.010692491196095943, + 0.0044878083281219006, + 0.005402742885053158, + 0.007181004621088505, + 0.003105688141658902, + 0.006784257013350725, + 0.0072811441496014595, + 0.0026804658118635416, + 0.003929848782718182, + 0.011382133699953556, + 0.006945436354726553, + 0.010273667052388191, + 0.007936782203614712, + 0.007113330531865358, + 0.016946474090218544, + 0.006124939303845167, + 0.008558310568332672, + 0.03579173609614372, + 0.009912002831697464, + 0.035760704427957535, + 0.008516897447407246, + 0.008427076041698456, + 0.020383629947900772, + 0.0052613019943237305, + 0.008629987016320229, + 0.006208387203514576, + 0.006613735109567642, + 0.012410322204232216, + 0.004696174059063196, + 0.007831571623682976, + 0.009821442887187004, + 0.007204461842775345, + 0.005295317154377699, + 0.013738713227212429, + 0.007601912133395672, + 0.014790507033467293, + 0.006516412831842899, + 0.007121772039681673, + 0.00991892535239458, + 0.006064487621188164, + 0.005787125322967768, + 0.004697628319263458, + 0.006138496100902557, + 0.005401765927672386, + 0.005959490314126015, + 0.005912241525948048, + 0.0068053253926336765, + 0.004311115015298128, + 0.004026020411401987, + 0.0074882893823087215, + 0.016652030870318413, + 0.012046711519360542, + 0.004726869985461235, + 0.008537646383047104, + 0.012540730647742748, + 0.003997937776148319, + 0.005016113165766001, + 0.012049569748342037, + 0.00783509947359562, + 0.02401670813560486, + 0.010852070525288582, + 0.0049139466136693954, + 0.010816829279065132, + 0.007555755320936441, + 0.014188561588525772, + 0.12033948302268982, + 0.13270144164562225, + 0.007923384197056293, + 0.006964995991438627, + 0.01481303758919239, + 0.0341225191950798, + 0.007459430489689112, + 0.008115105330944061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.2521384358406067 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254983339458704 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.01254983339458704, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254983339458704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12287408113479614, + 0.11712078750133514, + 0.11202844232320786, + 0.17201541364192963, + 0.18080700933933258, + 0.10932707786560059, + 0.057905301451683044, + 0.0466659739613533, + 0.04931269958615303, + 0.03194313496351242 + ], + "credit_map": [ + 0.010883803479373455, + 0.0187514778226614, + 0.013774177059531212, + 0.013975929468870163, + 0.025424351915717125, + 0.016624001786112785, + 0.01025440078228712, + 0.013185945339500904, + 0.013030314818024635, + 0.02941247820854187, + 0.028219416737556458, + 0.006548322271555662, + 0.017992565408349037, + 0.01039549708366394, + 0.005042453296482563, + 0.006479751784354448, + 0.027690136805176735, + 0.009888028725981712, + 0.03112347051501274, + 0.005691343918442726, + 0.005360668990761042, + 0.015437213703989983, + 0.004933349322527647, + 0.011904244311153889, + 0.05473913624882698, + 0.008000107482075691, + 0.0562935508787632, + 0.00973525270819664, + 0.006039885338395834, + 0.02296910062432289, + 0.005655412096530199, + 0.008582992479205132, + 0.017206506803631783, + 0.021597253158688545, + 0.02976987324655056, + 0.01640789769589901, + 0.029042458161711693, + 0.029097000136971474, + 0.022667618468403816, + 0.015018421225249767, + 0.017289966344833374, + 0.01397150382399559, + 0.01936505362391472, + 0.009531061165034771, + 0.011830558069050312, + 0.01945989392697811, + 0.008774478919804096, + 0.009104574099183083, + 0.006952030584216118, + 0.007767398376017809, + 0.007573199458420277, + 0.0076684788800776005, + 0.0060555096715688705, + 0.008030463010072708, + 0.007478951942175627, + 0.006379278376698494, + 0.003771466203033924, + 0.011713200248777866, + 0.011371193453669548, + 0.00296780769713223, + 0.004722329322248697, + 0.006419913377612829, + 0.002772899344563484, + 0.002927170367911458, + 0.008112210780382156, + 0.0063184406608343124, + 0.00693432055413723, + 0.0045280493795871735, + 0.003886040300130844, + 0.008599871769547462, + 0.005362770985811949, + 0.005570999346673489, + 0.007683111820369959, + 0.006534139160066843, + 0.0038022894877940416, + 0.002288133604452014, + 0.0028739951085299253, + 0.003810008056461811, + 0.002090290654450655, + 0.0028611698653548956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.01254983339458704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013328249333426356 + ], + "opposite_logprob_tokens": [ + -6.626332759857178 + ], + "logprob_reward": -0.0013328249333426356, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013328249333426356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09676173329353333, + 0.12501950562000275, + 0.11778751760721207, + 0.11742033809423447, + 0.17411376535892487, + 0.11398839950561523, + 0.05275305360555649, + 0.05706549063324928, + 0.07918085902929306, + 0.06590936332941055 + ], + "credit_map": [ + 0.010009845718741417, + 0.012942152097821236, + 0.012912723235785961, + 0.011921750381588936, + 0.019056245684623718, + 0.011086784303188324, + 0.008350271731615067, + 0.010481958277523518, + 0.015275713056325912, + 0.04524542763829231, + 0.0353718027472496, + 0.00521820830181241, + 0.009367203339934349, + 0.005817628465592861, + 0.003958888817578554, + 0.004764646291732788, + 0.03589945286512375, + 0.008490636944770813, + 0.04184418171644211, + 0.00700335344299674, + 0.004402084741741419, + 0.009462629444897175, + 0.004571175668388605, + 0.006114002782851458, + 0.022689880803227425, + 0.015094971284270287, + 0.020596660673618317, + 0.013131110928952694, + 0.013189720921218395, + 0.015630563721060753, + 0.007524920627474785, + 0.00956250261515379, + 0.017084114253520966, + 0.01969737745821476, + 0.030236123129725456, + 0.015315841883420944, + 0.02760131284594536, + 0.030004706233739853, + 0.019963158294558525, + 0.014211121015250683, + 0.01624641753733158, + 0.014642233029007912, + 0.016361329704523087, + 0.010942110791802406, + 0.011962125077843666, + 0.021893683820962906, + 0.011361355893313885, + 0.010579134337604046, + 0.005959043744951487, + 0.006613610778003931, + 0.00712805800139904, + 0.006220313720405102, + 0.0058015682734549046, + 0.00787684228271246, + 0.006624657195061445, + 0.006528956815600395, + 0.005552643444389105, + 0.020289411768317223, + 0.013871436007320881, + 0.0025954849552363157, + 0.004072962794452906, + 0.00536589790135622, + 0.0021174035500735044, + 0.00320024648681283, + 0.012971661984920502, + 0.007715615443885326, + 0.012443795800209045, + 0.008290502242743969, + 0.005767052993178368, + 0.010920298285782337, + 0.008836386725306511, + 0.012235547415912151, + 0.022483233362436295, + 0.02160070836544037, + 0.0030845454894006252, + 0.0026186916511505842, + 0.003988533280789852, + 0.006715224590152502, + 0.0023716401774436235, + 0.003046783385798335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.0013328249333426356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038082610815763474 + ], + "opposite_logprob_tokens": [ + -3.2880825996398926 + ], + "logprob_reward": -0.038082610815763474, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038082610815763474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06570809334516525, + 0.14168468117713928, + 0.06356638669967651, + 0.21660879254341125, + 0.0908784419298172, + 0.0335329994559288, + 0.08379259705543518, + 0.11228128522634506, + 0.14285799860954285, + 0.049088750034570694 + ], + "credit_map": [ + 0.0074583860114216805, + 0.008637403137981892, + 0.006920477841049433, + 0.0080553675070405, + 0.014766231179237366, + 0.006387609522789717, + 0.006846535950899124, + 0.006636087317019701, + 0.013549856841564178, + 0.046332020312547684, + 0.04368547722697258, + 0.009069293737411499, + 0.011327437125146389, + 0.00701564084738493, + 0.00497575756162405, + 0.00572919799014926, + 0.010142558254301548, + 0.005794729106128216, + 0.008793167769908905, + 0.006079608108848333, + 0.005057456437498331, + 0.01541825756430626, + 0.005188784562051296, + 0.007091826759278774, + 0.071967713534832, + 0.009785308502614498, + 0.07804477959871292, + 0.01030641421675682, + 0.006051252130419016, + 0.02551640011370182, + 0.006742488592863083, + 0.008194433525204659, + 0.009788937866687775, + 0.00856948271393776, + 0.0178239606320858, + 0.008316664025187492, + 0.01237432099878788, + 0.014830614440143108, + 0.010632438585162163, + 0.008542022667825222, + 0.004514828324317932, + 0.0038970080204308033, + 0.005432007368654013, + 0.003684472292661667, + 0.003538738936185837, + 0.0053281099535524845, + 0.003938750829547644, + 0.00319908419623971, + 0.0089528514072299, + 0.009616514667868614, + 0.012757707387208939, + 0.011766822077333927, + 0.008201247081160545, + 0.011072415858507156, + 0.010739544406533241, + 0.010685497894883156, + 0.013718575239181519, + 0.031837962567806244, + 0.01719239540398121, + 0.005960647948086262, + 0.0105394022539258, + 0.017435230314731598, + 0.005921850446611643, + 0.009675220586359501, + 0.026138918474316597, + 0.011970479972660542, + 0.040987528860569, + 0.014328451827168465, + 0.00750594912096858, + 0.015653036534786224, + 0.011732598766684532, + 0.014541025273501873, + 0.013398423790931702, + 0.011183041147887707, + 0.004226679913699627, + 0.00413832813501358, + 0.0031338734552264214, + 0.006248651072382927, + 0.003612430766224861, + 0.0031473187264055014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.038082610815763474 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, wide-angle, static shot captures a magnificent spectacle beneath a deep, dark night sky. Hundreds of small, sleek drones, their gleaming metallic black shells looming in the darkness, converge at breakneck speed from all directions, even from the depths of the background, towards an invisible focal point in the very center of the frame. Each drone trails a long, luminous digital light trail of electric blue and fiery magenta, carving elegant, intricate arcs through the sky. Their flight paths intertwine, instantly weaving a vast, dynamic web of light across the clear night sky, completely ablaze. The brightly lit futuristic city skyline below provides subtle ambient light, but the primary illumination for the entire scene comes from the drones themselves, whose light reflects off their sleek shells in sharp highlights. The entire image is imbued with a sense of speed, energy, and precise control, creating an atmosphere of grand, high-tech, and dizzying futuristic visual spectacle.", + "seed": 12354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_5837_seed_12354_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0441036228556186, + "vlm_logprob": -0.0441036228556186, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.0441036228556186, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02706453949213028 + ], + "opposite_logprob_tokens": [ + -3.652064561843872 + ], + "logprob_reward": -0.02706453949213028, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02706453949213028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22398501634597778, + 0.19927334785461426, + 0.19024966657161713, + 0.020833328366279602, + 0.056491151452064514, + 0.03202340006828308, + 0.13138478994369507, + 0.0581372044980526, + 0.06788963824510574, + 0.019732490181922913 + ], + "credit_map": [ + 0.01938077248632908, + 0.01868145912885666, + 0.022595077753067017, + 0.028055507689714432, + 0.027411652728915215, + 0.03156035766005516, + 0.053380172699689865, + 0.022920021787285805, + 0.025524679571390152, + 0.02564801089465618, + 0.025192074477672577, + 0.018490534275770187, + 0.022014765068888664, + 0.02574532851576805, + 0.023959599435329437, + 0.03269834816455841, + 0.03930079936981201, + 0.014493520371615887, + 0.016702083870768547, + 0.014212720096111298, + 0.02468418888747692, + 0.03394484892487526, + 0.017837075516581535, + 0.029074417427182198, + 0.0046495115384459496, + 0.0019447077065706253, + 0.001730433781631291, + 0.0014725058572366834, + 0.002610963536426425, + 0.0025371978990733624, + 0.0029423926025629044, + 0.0029456152115017176, + 0.008253318257629871, + 0.004377217497676611, + 0.00598622253164649, + 0.005077728535979986, + 0.0047914101742208, + 0.006410705391317606, + 0.008698977530002594, + 0.012895572930574417, + 0.00402056984603405, + 0.0038160341791808605, + 0.0028054772410541773, + 0.003943000454455614, + 0.005424088798463345, + 0.004058549180626869, + 0.004873206373304129, + 0.0030824707355350256, + 0.01559373177587986, + 0.013092751614749432, + 0.015300215221941471, + 0.011095421388745308, + 0.0160943903028965, + 0.026058996096253395, + 0.02193613536655903, + 0.01221314538270235, + 0.012803190387785435, + 0.004318573512136936, + 0.0028376562986522913, + 0.002983451122418046, + 0.0196219589561224, + 0.006512300111353397, + 0.006021617446094751, + 0.00303845526650548, + 0.008168908767402172, + 0.007475095801055431, + 0.006611221469938755, + 0.0071384296752512455, + 0.007009671535342932, + 0.008458212018013, + 0.01322262641042471, + 0.009805476292967796, + 0.0023466183338314295, + 0.0038320845924317837, + 0.0021370439790189266, + 0.0018100050510838628, + 0.0038755349814891815, + 0.002656973898410797, + 0.0016667416784912348, + 0.0014074865030124784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02706453949213028 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042236328125 + ], + "opposite_logprob_tokens": [ + -2.812042236328125 + ], + "logprob_reward": -0.062042236328125, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042236328125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2057761549949646, + 0.07157356292009354, + 0.18538174033164978, + 0.04171700403094292, + 0.06727403402328491, + 0.04383589327335358, + 0.05369371548295021, + 0.2825169265270233, + 0.03128973767161369, + 0.016941208392381668 + ], + "credit_map": [ + 0.01831834763288498, + 0.018109366297721863, + 0.022191246971488, + 0.029055017977952957, + 0.02257741615176201, + 0.027567129582166672, + 0.043826330453157425, + 0.024131307378411293, + 0.008692224510014057, + 0.007837806828320026, + 0.007083466742187738, + 0.009454098530113697, + 0.009267596527934074, + 0.007448080461472273, + 0.011274060234427452, + 0.010516224429011345, + 0.04729631170630455, + 0.013628099113702774, + 0.018472300842404366, + 0.013480686582624912, + 0.02172859199345112, + 0.030597509816288948, + 0.01469884067773819, + 0.025479404255747795, + 0.00393953500315547, + 0.003672074293717742, + 0.0035204461310058832, + 0.0032837740145623684, + 0.0053330836817622185, + 0.006267134100198746, + 0.006187351420521736, + 0.009513608179986477, + 0.00926526915282011, + 0.004721269477158785, + 0.006498546339571476, + 0.005881964694708586, + 0.0053519606590271, + 0.007901700213551521, + 0.012872908264398575, + 0.01478041335940361, + 0.005937068723142147, + 0.004689028952270746, + 0.005326387006789446, + 0.0030588365625590086, + 0.004301299806684256, + 0.006414946634322405, + 0.0035838524345308542, + 0.010524468496441841, + 0.005782236810773611, + 0.004851694218814373, + 0.004723334684967995, + 0.004784677643328905, + 0.006095367018133402, + 0.01082010567188263, + 0.010299896821379662, + 0.006336402613669634, + 0.08057656139135361, + 0.01743314042687416, + 0.01699097827076912, + 0.011583001352846622, + 0.096594899892807, + 0.027104558423161507, + 0.01821073144674301, + 0.014023066498339176, + 0.0038084685802459717, + 0.0034485578071326017, + 0.0031524719670414925, + 0.004036925733089447, + 0.003174724755808711, + 0.005021314136683941, + 0.004239279311150312, + 0.004407995846122503, + 0.0022811265662312508, + 0.0017135449452325702, + 0.002430766122415662, + 0.001490123337134719, + 0.002721420954912901, + 0.0021462358999997377, + 0.0018917464185506105, + 0.002266243100166321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.062042236328125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022005646023899317 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.0022005646023899317, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022005646023899317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3431234359741211, + 0.1710348129272461, + 0.15786224603652954, + 0.051643308252096176, + 0.04150691628456116, + 0.03248131275177002, + 0.0815332755446434, + 0.03171035274863243, + 0.06663882732391357, + 0.022465404123067856 + ], + "credit_map": [ + 0.027711834758520126, + 0.0294240303337574, + 0.04039884731173515, + 0.04383457824587822, + 0.03699922189116478, + 0.044654205441474915, + 0.07837985455989838, + 0.041720911860466, + 0.016544023528695107, + 0.020754247903823853, + 0.012122420594096184, + 0.018815767019987106, + 0.01774856075644493, + 0.020891176536679268, + 0.026514949277043343, + 0.037643685936927795, + 0.03136788681149483, + 0.012392081320285797, + 0.013855842873454094, + 0.010202000848948956, + 0.02026788704097271, + 0.025891823694109917, + 0.015909958630800247, + 0.02797478251159191, + 0.005345542915165424, + 0.004892549943178892, + 0.005139960441738367, + 0.0048955464735627174, + 0.010490624234080315, + 0.006316570099443197, + 0.008836522698402405, + 0.0057259961031377316, + 0.008388237096369267, + 0.003597668372094631, + 0.003591818269342184, + 0.0031215075869113207, + 0.0036846750881522894, + 0.005193251185119152, + 0.006684249732643366, + 0.007245513144880533, + 0.005283386446535587, + 0.003912778105586767, + 0.0032333803828805685, + 0.0028424779884517193, + 0.0042974380776286125, + 0.004660428036004305, + 0.0038453489542007446, + 0.004406079649925232, + 0.007561922539025545, + 0.0052458844147622585, + 0.005921265110373497, + 0.0088184904307127, + 0.007853579707443714, + 0.018400540575385094, + 0.018976813182234764, + 0.00875478982925415, + 0.005282304715365171, + 0.002845835406333208, + 0.00260400609113276, + 0.0024159655440598726, + 0.00729834008961916, + 0.003552731592208147, + 0.005418451968580484, + 0.0022927175741642714, + 0.007784611079841852, + 0.007332205306738615, + 0.006921492051333189, + 0.0071611409075558186, + 0.00613561924546957, + 0.008001869544386864, + 0.015637224540114403, + 0.007664674427360296, + 0.0026700887829065323, + 0.004378779791295528, + 0.0023516297806054354, + 0.0024747427087277174, + 0.004096061922609806, + 0.002367825713008642, + 0.0021580816246569157, + 0.0019681951962411404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0022005646023899317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021947361528873444 + ], + "opposite_logprob_tokens": [ + -6.127194881439209 + ], + "logprob_reward": -0.0021947361528873444, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021947361528873444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24174180626869202, + 0.14813952147960663, + 0.1592913269996643, + 0.08926822245121002, + 0.1332438439130783, + 0.10274206846952438, + 0.02109806053340435, + 0.03471067175269127, + 0.047891728579998016, + 0.02187274396419525 + ], + "credit_map": [ + 0.019021427258849144, + 0.022626955062150955, + 0.02832498587667942, + 0.051000967621803284, + 0.02547542192041874, + 0.032709751278162, + 0.03189895302057266, + 0.03068333864212036, + 0.015192341059446335, + 0.01123952865600586, + 0.0129871079698205, + 0.015178866684436798, + 0.014046451076865196, + 0.01895354874432087, + 0.02144452929496765, + 0.03909715265035629, + 0.03722825646400452, + 0.011967038735747337, + 0.014881225302815437, + 0.00989984069019556, + 0.020668957382440567, + 0.02599339187145233, + 0.014727594330906868, + 0.023925023153424263, + 0.007610006257891655, + 0.00826263427734375, + 0.00899778213351965, + 0.0074859135784208775, + 0.02107773907482624, + 0.009708045050501823, + 0.01592826470732689, + 0.01019783690571785, + 0.018717819824814796, + 0.01266531553119421, + 0.01237560249865055, + 0.013505082577466965, + 0.010982240550220013, + 0.014825177378952503, + 0.021648874506354332, + 0.02852373942732811, + 0.010734638199210167, + 0.006355504505336285, + 0.006814213469624519, + 0.007527551148086786, + 0.009371801279485226, + 0.01898462325334549, + 0.00958542712032795, + 0.0333683080971241, + 0.0025746948085725307, + 0.0016927714459598064, + 0.001929510966874659, + 0.0021518298890441656, + 0.0021238261833786964, + 0.003886042395606637, + 0.0037386934272944927, + 0.0030006926972419024, + 0.004115165676921606, + 0.0022316062822937965, + 0.003789318958297372, + 0.0024931316729635, + 0.009628569707274437, + 0.0042038955725729465, + 0.0054747131653130054, + 0.0027742674574255943, + 0.006808908190578222, + 0.005669266916811466, + 0.0046730344183743, + 0.004308054223656654, + 0.005128440912812948, + 0.005607011262327433, + 0.010689124464988708, + 0.005007890053093433, + 0.0031890051905065775, + 0.0037160716019570827, + 0.0021913754753768444, + 0.002108391374349594, + 0.003657794324681163, + 0.002969389082863927, + 0.0020361756905913353, + 0.0020045428536832333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0021947361528873444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701603770256042 + ], + "opposite_logprob_tokens": [ + -2.127016067504883 + ], + "logprob_reward": -0.12701603770256042, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12701603770256042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19797958433628082, + 0.13341538608074188, + 0.06657867878675461, + 0.11512259393930435, + 0.07226375490427017, + 0.05894012004137039, + 0.03694208338856697, + 0.17005516588687897, + 0.11776184290647507, + 0.030940817669034004 + ], + "credit_map": [ + 0.020884975790977478, + 0.018991881981492043, + 0.022805001586675644, + 0.028777476400136948, + 0.024639906361699104, + 0.027425521984696388, + 0.036228932440280914, + 0.018225857987999916, + 0.016530079767107964, + 0.014797571115195751, + 0.018447989597916603, + 0.013550953939557076, + 0.015749815851449966, + 0.017117327079176903, + 0.015332740731537342, + 0.021888891234993935, + 0.011094038374722004, + 0.005360859911888838, + 0.005163871683180332, + 0.005437801126390696, + 0.01080499216914177, + 0.011026617139577866, + 0.008241304196417332, + 0.009449180215597153, + 0.011029833927750587, + 0.011264164932072163, + 0.010910775512456894, + 0.0109622897580266, + 0.02464519441127777, + 0.01159832626581192, + 0.02122083120048046, + 0.013491171412169933, + 0.012344978749752045, + 0.008281432092189789, + 0.006356813479214907, + 0.00579379266127944, + 0.00720227649435401, + 0.008523856289684772, + 0.014487738721072674, + 0.009272860363125801, + 0.0073097823187708855, + 0.005629498045891523, + 0.006137780845165253, + 0.005255958065390587, + 0.0074614849872887135, + 0.010045034810900688, + 0.006712145172059536, + 0.010388429276645184, + 0.004081449471414089, + 0.003425241680815816, + 0.00464651407673955, + 0.0031910822726786137, + 0.004542361944913864, + 0.005966637283563614, + 0.00422624871134758, + 0.006862543523311615, + 0.03857167437672615, + 0.011605801992118359, + 0.020012065768241882, + 0.012082226574420929, + 0.03535201773047447, + 0.017166977748274803, + 0.02098936215043068, + 0.014275029301643372, + 0.01638314500451088, + 0.0134707847610116, + 0.012810922227799892, + 0.01261481549590826, + 0.011679114773869514, + 0.01585552841424942, + 0.022930912673473358, + 0.012016610242426395, + 0.004569044336676598, + 0.004130440764129162, + 0.0037958950269967318, + 0.0034725493751466274, + 0.0035006676334887743, + 0.004299412481486797, + 0.0033585787750780582, + 0.0038142288103699684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.12701603770256042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, wide-angle, static shot captures a magnificent spectacle beneath a deep, dark night sky. Hundreds of small, sleek drones, their gleaming metallic black shells looming in the darkness, converge at breakneck speed from all directions, even from the depths of the background, towards an invisible focal point in the very center of the frame. Each drone trails a long, luminous digital light trail of electric blue and fiery magenta, carving elegant, intricate arcs through the sky. Their flight paths intertwine, instantly weaving a vast, dynamic web of light across the clear night sky, completely ablaze. The brightly lit futuristic city skyline below provides subtle ambient light, but the primary illumination for the entire scene comes from the drones themselves, whose light reflects off their sleek shells in sharp highlights. The entire image is imbued with a sense of speed, energy, and precise control, creating an atmosphere of grand, high-tech, and dizzying futuristic visual spectacle.", + "seed": 12355, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_5837_seed_12355_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08023610620293767, + "vlm_logprob": -0.08023610620293767, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.08023610620293767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09090972691774368 + ], + "opposite_logprob_tokens": [ + -2.465909719467163 + ], + "logprob_reward": -0.09090972691774368, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09090972691774368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5810160040855408, + 0.02168859727680683, + 0.04054988548159599, + 0.05260399729013443, + 0.030799061059951782, + 0.05470546707510948, + 0.1161348894238472, + 0.05937321484088898, + 0.02760954387485981, + 0.015519284643232822 + ], + "credit_map": [ + 0.04121781885623932, + 0.06737257540225983, + 0.11299855262041092, + 0.040925901383161545, + 0.10543196648359299, + 0.10930663347244263, + 0.05797401815652847, + 0.04578854888677597, + 0.0021275165490806103, + 0.002080370206385851, + 0.0030287380795925856, + 0.0026214304380118847, + 0.002336060395464301, + 0.0029443101957440376, + 0.00411427766084671, + 0.002435892354696989, + 0.00490536680445075, + 0.0059577953070402145, + 0.006313241086900234, + 0.003007803112268448, + 0.00739396084100008, + 0.005291013978421688, + 0.004007020965218544, + 0.0036736864130944014, + 0.006487500853836536, + 0.0057555329985916615, + 0.011950144544243813, + 0.0042136781848967075, + 0.006520766764879227, + 0.004916814621537924, + 0.0051348526030778885, + 0.007624703459441662, + 0.0034036017023026943, + 0.0038782386109232903, + 0.004632530268281698, + 0.0030152867548167706, + 0.00393524207174778, + 0.004323313944041729, + 0.0038968545850366354, + 0.0037139940541237593, + 0.007320726290345192, + 0.005928556900471449, + 0.009702441282570362, + 0.004276730585843325, + 0.004983951337635517, + 0.0066782403737306595, + 0.006236696150153875, + 0.009578121826052666, + 0.01286864373832941, + 0.012870550155639648, + 0.016771623864769936, + 0.013035024516284466, + 0.011856884695589542, + 0.015174392610788345, + 0.013194900937378407, + 0.020362865179777145, + 0.00760205602273345, + 0.009088590741157532, + 0.012012144550681114, + 0.0033060647547245026, + 0.005477381870150566, + 0.015419544652104378, + 0.0030557187274098396, + 0.003411712357774377, + 0.0036516678519546986, + 0.00350612448528409, + 0.0031504982616752386, + 0.0022350167855620384, + 0.0033654358703643084, + 0.004003673326224089, + 0.003992208279669285, + 0.0037049201782792807, + 0.0013604628620669246, + 0.001862799283117056, + 0.0030339702498167753, + 0.0010542443487793207, + 0.002191302366554737, + 0.0018618693575263023, + 0.0016641905531287193, + 0.0024904459714889526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.09090972691774368 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015610784292221 + ], + "opposite_logprob_tokens": [ + -1.7015610933303833 + ], + "logprob_reward": -0.2015610784292221, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015610784292221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5511258244514465, + 0.019205953925848007, + 0.11015117168426514, + 0.06227956339716911, + 0.022360263392329216, + 0.04849918559193611, + 0.0481363907456398, + 0.09021376073360443, + 0.02652762085199356, + 0.0215002428740263 + ], + "credit_map": [ + 0.04048248752951622, + 0.06837016344070435, + 0.09703449159860611, + 0.0396270789206028, + 0.11094525456428528, + 0.09108250588178635, + 0.057334065437316895, + 0.04624981805682182, + 0.002158993622288108, + 0.001808681758120656, + 0.0025113101582974195, + 0.0023788907565176487, + 0.0024254596792161465, + 0.0025925247464329004, + 0.0030995728448033333, + 0.0022305201273411512, + 0.006830024532973766, + 0.008520303294062614, + 0.03210160508751869, + 0.004898794926702976, + 0.034610815346241, + 0.010460068471729755, + 0.005946328863501549, + 0.006783234886825085, + 0.00772908516228199, + 0.00802446249872446, + 0.01223096065223217, + 0.005446251947432756, + 0.007438800297677517, + 0.006545940414071083, + 0.0063192956149578094, + 0.008544761687517166, + 0.0036942821461707354, + 0.0026937369257211685, + 0.003660226473584771, + 0.0016881288029253483, + 0.0027714360039681196, + 0.003043063683435321, + 0.002359024714678526, + 0.002450364176183939, + 0.00597530510276556, + 0.0050832852721214294, + 0.00853546429425478, + 0.004012536257505417, + 0.004765836521983147, + 0.005784696899354458, + 0.005418950226157904, + 0.008923115208745003, + 0.005269840359687805, + 0.005047404672950506, + 0.006842999719083309, + 0.005678239744156599, + 0.004762654658406973, + 0.006041607353836298, + 0.00528001831844449, + 0.009213623590767384, + 0.012024630792438984, + 0.015419883653521538, + 0.016078738495707512, + 0.005413488484919071, + 0.01008228212594986, + 0.02000143937766552, + 0.005150180775672197, + 0.006043112371116877, + 0.002046202542260289, + 0.0026656854897737503, + 0.005009555257856846, + 0.0016875399742275476, + 0.002401103964075446, + 0.00682076858356595, + 0.002934758784249425, + 0.00296200648881495, + 0.0015937364660203457, + 0.0022448329254984856, + 0.004520220682024956, + 0.001732304459437728, + 0.002018924104049802, + 0.0024746819399297237, + 0.002377118682488799, + 0.004538426175713539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.2015610784292221 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189411610364914 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189411610364914, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189411610364914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5673402547836304, + 0.03452038764953613, + 0.09384719282388687, + 0.05222723260521889, + 0.021144839003682137, + 0.040095746517181396, + 0.0675337165594101, + 0.06439122557640076, + 0.019795214757323265, + 0.03910421207547188 + ], + "credit_map": [ + 0.04096855968236923, + 0.06847099959850311, + 0.10225111991167068, + 0.04104042425751686, + 0.10837477445602417, + 0.09895794838666916, + 0.05998149886727333, + 0.047294918447732925, + 0.00422526802867651, + 0.00515240291133523, + 0.006904685404151678, + 0.0021743469405919313, + 0.005422193091362715, + 0.0033188764937222004, + 0.0036479937843978405, + 0.0036746191326528788, + 0.006997175049036741, + 0.008510051295161247, + 0.02463158406317234, + 0.004674885887652636, + 0.027052059769630432, + 0.009664668701589108, + 0.006086057983338833, + 0.0062306951731443405, + 0.006305904593318701, + 0.006551041733473539, + 0.010423773899674416, + 0.004689793102443218, + 0.005987063050270081, + 0.005567002110183239, + 0.005469948519021273, + 0.007232701871544123, + 0.0028384674806147814, + 0.0026212886441498995, + 0.003426490118727088, + 0.001744585926644504, + 0.0024477727711200714, + 0.0027044247835874557, + 0.0026230579242110252, + 0.0027387505397200584, + 0.004982174374163151, + 0.004252583719789982, + 0.006618547718971968, + 0.0030011525377631187, + 0.0033482287544757128, + 0.005036239977926016, + 0.004623821936547756, + 0.008232994936406612, + 0.00730945123359561, + 0.007235654629766941, + 0.009745929390192032, + 0.00768688041716814, + 0.006890296936035156, + 0.008783047087490559, + 0.007639702875167131, + 0.012242748402059078, + 0.0087037468329072, + 0.010811765678226948, + 0.011293008923530579, + 0.003804926760494709, + 0.007010635454207659, + 0.015174745582044125, + 0.0036914728116244078, + 0.0039009195752441883, + 0.0022675141226500273, + 0.001898799673654139, + 0.003399978158995509, + 0.001238013501279056, + 0.0015846907626837492, + 0.004836427979171276, + 0.002376856515184045, + 0.0021929326467216015, + 0.005051856394857168, + 0.005593887530267239, + 0.006250547245144844, + 0.0028819777071475983, + 0.005483191926032305, + 0.004666911903768778, + 0.003983925562351942, + 0.0051919082179665565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.018189411610364914 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015100754098966718 + ], + "opposite_logprob_tokens": [ + -6.501510143280029 + ], + "logprob_reward": -0.0015100754098966718, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015100754098966718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5576702356338501, + 0.03139028698205948, + 0.09808866679668427, + 0.051181990653276443, + 0.022331103682518005, + 0.03320654481649399, + 0.08611363917589188, + 0.05246232450008392, + 0.02850596234202385, + 0.039049290120601654 + ], + "credit_map": [ + 0.03945164754986763, + 0.06508240103721619, + 0.10567063093185425, + 0.038828663527965546, + 0.10518249869346619, + 0.10142142325639725, + 0.05782041698694229, + 0.044212568551301956, + 0.003588366089388728, + 0.0035459285136312246, + 0.005979395937174559, + 0.0028195527847856283, + 0.0035413745790719986, + 0.003953081090003252, + 0.003652592422440648, + 0.004309998359531164, + 0.006170629523694515, + 0.008068334311246872, + 0.027123037725687027, + 0.004578037187457085, + 0.030316464602947235, + 0.009758762083947659, + 0.005438542924821377, + 0.006634864490479231, + 0.006181485019624233, + 0.005347758065909147, + 0.011627844534814358, + 0.0038152839988470078, + 0.006330079864710569, + 0.005229988135397434, + 0.005498447455465794, + 0.007151102181524038, + 0.0028991831932216883, + 0.002268790500238538, + 0.003097803797572851, + 0.0021488196216523647, + 0.0025868380907922983, + 0.0025059953331947327, + 0.0034484202042222023, + 0.0033752534072846174, + 0.004383768420666456, + 0.0039037796668708324, + 0.005450846627354622, + 0.00268531939946115, + 0.00542198633775115, + 0.003485685447230935, + 0.003871918423101306, + 0.0040032402612268925, + 0.009823762811720371, + 0.009934223257005215, + 0.012956365942955017, + 0.00946071371436119, + 0.008342426270246506, + 0.011091092601418495, + 0.00907466933131218, + 0.015430384315550327, + 0.007110223639756441, + 0.008169705048203468, + 0.009998355060815811, + 0.0029856765177100897, + 0.005082813091576099, + 0.012926621362566948, + 0.002897711005061865, + 0.0032912178430706263, + 0.002684173872694373, + 0.002610687632113695, + 0.0054590473882853985, + 0.0016911905258893967, + 0.002493376610800624, + 0.007471922785043716, + 0.0029866767581552267, + 0.0031088877003639936, + 0.004738151095807552, + 0.004549704026430845, + 0.006181048229336739, + 0.0029670177027583122, + 0.0040838452987372875, + 0.003935165703296661, + 0.006109815556555986, + 0.006484543904662132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0015100754098966718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901023864746094 + ], + "opposite_logprob_tokens": [ + -2.464010238647461 + ], + "logprob_reward": -0.08901023864746094, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901023864746094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.191888228058815, + 0.12642677128314972, + 0.25024569034576416, + 0.04336591809988022, + 0.028388988226652145, + 0.03613605722784996, + 0.06567319482564926, + 0.05158456414937973, + 0.040802597999572754, + 0.16548800468444824 + ], + "credit_map": [ + 0.013921062462031841, + 0.012103443033993244, + 0.05473561957478523, + 0.010744290426373482, + 0.018265388906002045, + 0.05240904539823532, + 0.01204104907810688, + 0.017668342217803, + 0.01369569730013609, + 0.016790341585874557, + 0.021564336493611336, + 0.011533553712069988, + 0.021876178681850433, + 0.012219443917274475, + 0.01165072526782751, + 0.017096497118473053, + 0.016023816540837288, + 0.01408662274479866, + 0.07739761471748352, + 0.013047104701399803, + 0.07613416016101837, + 0.02259160578250885, + 0.014294097200036049, + 0.016670672222971916, + 0.004502517636865377, + 0.004450572654604912, + 0.011612881906330585, + 0.0023387710098177195, + 0.008445769548416138, + 0.0034541075583547354, + 0.00409088097512722, + 0.004470417741686106, + 0.0029819670598953962, + 0.0037183696404099464, + 0.0032768850214779377, + 0.003054768778383732, + 0.0028001389000564814, + 0.004454738460481167, + 0.004185745492577553, + 0.003916377667337656, + 0.003420158987864852, + 0.004189806059002876, + 0.00437770364806056, + 0.0034064813517034054, + 0.005664655473083258, + 0.0042434982024133205, + 0.004324018955230713, + 0.006509740371257067, + 0.009369160048663616, + 0.0077017624862492085, + 0.008890616707503796, + 0.006794666405767202, + 0.007235471159219742, + 0.007632351480424404, + 0.005878526717424393, + 0.012170640751719475, + 0.007943454198539257, + 0.009226012043654919, + 0.006193121895194054, + 0.0037097574677318335, + 0.008722454309463501, + 0.00586785189807415, + 0.005524906329810619, + 0.0043970090337097645, + 0.004565964452922344, + 0.004337622318416834, + 0.00612979382276535, + 0.0028062029741704464, + 0.004007814452052116, + 0.008977394551038742, + 0.005286551546305418, + 0.004691253416240215, + 0.026376908645033836, + 0.024908775463700294, + 0.02141694352030754, + 0.014231543987989426, + 0.019196167588233948, + 0.018768003210425377, + 0.02331353910267353, + 0.017276136204600334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.08901023864746094 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, wide-angle, static shot captures a magnificent spectacle beneath a deep, dark night sky. Hundreds of small, sleek drones, their gleaming metallic black shells looming in the darkness, converge at breakneck speed from all directions, even from the depths of the background, towards an invisible focal point in the very center of the frame. Each drone trails a long, luminous digital light trail of electric blue and fiery magenta, carving elegant, intricate arcs through the sky. Their flight paths intertwine, instantly weaving a vast, dynamic web of light across the clear night sky, completely ablaze. The brightly lit futuristic city skyline below provides subtle ambient light, but the primary illumination for the entire scene comes from the drones themselves, whose light reflects off their sleek shells in sharp highlights. The entire image is imbued with a sense of speed, energy, and precise control, creating an atmosphere of grand, high-tech, and dizzying futuristic visual spectacle.", + "seed": 12356, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_6/index_5837_seed_12356_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1795281393453479, + "vlm_logprob": -0.1795281393453479, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.1795281393453479, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329797506332397 + ], + "opposite_logprob_tokens": [ + -0.7579797506332397 + ], + "logprob_reward": -0.6329797506332397, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329797506332397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08551083505153656, + 0.19041770696640015, + 0.1882883757352829, + 0.11600897461175919, + 0.13493876159191132, + 0.07049360871315002, + 0.05216088891029358, + 0.055145058780908585, + 0.05414293706417084, + 0.05289284139871597 + ], + "credit_map": [ + 0.010326770134270191, + 0.015592139214277267, + 0.014504111371934414, + 0.014764314517378807, + 0.01081614289432764, + 0.005621642805635929, + 0.006333657540380955, + 0.007552053313702345, + 0.017282702028751373, + 0.027441391721367836, + 0.02355102263391018, + 0.03139027580618858, + 0.02746342122554779, + 0.018293943256139755, + 0.021532628685235977, + 0.023462312296032906, + 0.018390990793704987, + 0.034266311675310135, + 0.019991206005215645, + 0.027319900691509247, + 0.04123777151107788, + 0.02421649731695652, + 0.0099674416705966, + 0.012898268178105354, + 0.010125681757926941, + 0.02119666337966919, + 0.015549195930361748, + 0.010057661682367325, + 0.0186889860779047, + 0.02237730845808983, + 0.007452643476426601, + 0.010560829192399979, + 0.016750406473875046, + 0.03565986827015877, + 0.014160329475998878, + 0.012449443340301514, + 0.02777213789522648, + 0.007807842921465635, + 0.00911377277225256, + 0.011224962770938873, + 0.008366514928638935, + 0.014523877762258053, + 0.014424584805965424, + 0.006455410737544298, + 0.01109034288674593, + 0.006857225671410561, + 0.004743052646517754, + 0.00403260113671422, + 0.004796508699655533, + 0.007247441913932562, + 0.007232738193124533, + 0.0076731001026928425, + 0.0076695517636835575, + 0.007516460958868265, + 0.005835255142301321, + 0.004189830739051104, + 0.0064385742880403996, + 0.007353184744715691, + 0.008541068993508816, + 0.0077034831047058105, + 0.0071802688762545586, + 0.005050082225352526, + 0.0050229886546730995, + 0.007855402305722237, + 0.004515401553362608, + 0.008502742275595665, + 0.006285306066274643, + 0.006269833073019981, + 0.00743918726220727, + 0.006805713754147291, + 0.006526056677103043, + 0.007798692677170038, + 0.003268721280619502, + 0.009192729368805885, + 0.004481965210288763, + 0.006567910313606262, + 0.011013252660632133, + 0.006084078457206488, + 0.004844226874411106, + 0.007439953275024891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.6329797506332397 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705351412296295 + ], + "opposite_logprob_tokens": [ + -2.1270534992218018 + ], + "logprob_reward": -0.12705351412296295, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705351412296295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1914382129907608, + 0.11461134254932404, + 0.27238309383392334, + 0.0404278039932251, + 0.056220345199108124, + 0.03869706019759178, + 0.06280426681041718, + 0.03410402312874794, + 0.049259286373853683, + 0.1400545984506607 + ], + "credit_map": [ + 0.024638518691062927, + 0.02905116230249405, + 0.03419177606701851, + 0.03103286772966385, + 0.023684443905949593, + 0.01298697292804718, + 0.012087623588740826, + 0.0237648393958807, + 0.008559087291359901, + 0.02289509028196335, + 0.01519088540226221, + 0.015921764075756073, + 0.021673837676644325, + 0.01146125327795744, + 0.009600963443517685, + 0.009308458305895329, + 0.02701975218951702, + 0.04852982982993126, + 0.028111953288316727, + 0.03751129284501076, + 0.05873839557170868, + 0.038800083100795746, + 0.014585389755666256, + 0.019086407497525215, + 0.004515268374234438, + 0.006959357764571905, + 0.004526271019130945, + 0.00440240791067481, + 0.007524092681705952, + 0.005009365268051624, + 0.0037520627956837416, + 0.003738976549357176, + 0.006204781588166952, + 0.012480220757424831, + 0.006996700540184975, + 0.005629484076052904, + 0.01023886725306511, + 0.003546317107975483, + 0.004930812865495682, + 0.006193157285451889, + 0.0036400991957634687, + 0.006302691996097565, + 0.004962228238582611, + 0.003711806144565344, + 0.00856254156678915, + 0.005087145138531923, + 0.004121028818190098, + 0.0023095181677490473, + 0.004704820923507214, + 0.007916825823485851, + 0.009401281364262104, + 0.011577191762626171, + 0.00832426268607378, + 0.010573230683803558, + 0.005270447116345167, + 0.0050361971370875835, + 0.0038204616867005825, + 0.004421009216457605, + 0.005730955861508846, + 0.005042146425694227, + 0.00433944771066308, + 0.002809182507917285, + 0.002938127378001809, + 0.005002687685191631, + 0.00411098962649703, + 0.010433743707835674, + 0.005533052142709494, + 0.005507979542016983, + 0.009273581206798553, + 0.0064614275470376015, + 0.003911120817065239, + 0.0040273903869092464, + 0.015079983510077, + 0.016363203525543213, + 0.012348453514277935, + 0.03234366327524185, + 0.01789780519902706, + 0.015241463668644428, + 0.020734088495373726, + 0.010045927949249744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.12705351412296295 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07142875343561172, + 0.12255550175905228, + 0.2349681556224823, + 0.07673949748277664, + 0.05443299561738968, + 0.028257114812731743, + 0.04454962536692619, + 0.02619888447225094, + 0.02579924650490284, + 0.3150702118873596 + ], + "credit_map": [ + 0.008049027062952518, + 0.012510920874774456, + 0.013070855289697647, + 0.010351987555623055, + 0.010100849904119968, + 0.00564856082201004, + 0.004342930857092142, + 0.00735362246632576, + 0.01157357357442379, + 0.018466763198375702, + 0.01329951174557209, + 0.019115637987852097, + 0.019743865355849266, + 0.012084247544407845, + 0.013016733340919018, + 0.015255174599587917, + 0.023992113769054413, + 0.04350019991397858, + 0.023447658866643906, + 0.03274846076965332, + 0.049269866198301315, + 0.03387428820133209, + 0.011750574223697186, + 0.016385013237595558, + 0.006174150388687849, + 0.015261777676641941, + 0.013359148986637592, + 0.006604463793337345, + 0.013308864086866379, + 0.011032483540475368, + 0.0044006602838635445, + 0.006597955245524645, + 0.005557727999985218, + 0.011357897892594337, + 0.006806273944675922, + 0.006119303405284882, + 0.009743544273078442, + 0.003856831230223179, + 0.004299905616790056, + 0.006691517308354378, + 0.0031524663791060448, + 0.005103187169879675, + 0.002983467187732458, + 0.00291920593008399, + 0.005763873923569918, + 0.0034420189913362265, + 0.0029011908918619156, + 0.0019917055033147335, + 0.0040626670233905315, + 0.0061575318686664104, + 0.006255226209759712, + 0.006433450151234865, + 0.005924764089286327, + 0.0070236665196716785, + 0.00487670348957181, + 0.003815615549683571, + 0.002841982524842024, + 0.003609553212299943, + 0.004744437988847494, + 0.003529069246724248, + 0.003449472598731518, + 0.0023198090493679047, + 0.0017949686152860522, + 0.003909593913704157, + 0.001930678030475974, + 0.004613462835550308, + 0.0026974426582455635, + 0.003525979584082961, + 0.0035540941171348095, + 0.00385731621645391, + 0.0031337167602032423, + 0.002486554905772209, + 0.03563868626952171, + 0.041643209755420685, + 0.026397904381155968, + 0.060330212116241455, + 0.04663624241948128, + 0.037517789751291275, + 0.03858914598822594, + 0.028317030519247055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.005950829479843378 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004631743300706148 + ], + "opposite_logprob_tokens": [ + -5.379631519317627 + ], + "logprob_reward": -0.004631743300706148, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004631743300706148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0994519516825676, + 0.28250354528427124, + 0.07268550992012024, + 0.057296574115753174, + 0.07384871691465378, + 0.12660668790340424, + 0.046796657145023346, + 0.08999346941709518, + 0.045410748571157455, + 0.10540608316659927 + ], + "credit_map": [ + 0.008591830730438232, + 0.017678551375865936, + 0.012243879027664661, + 0.014999602921307087, + 0.0144790168851614, + 0.008413141593337059, + 0.008268147706985474, + 0.014777780510485172, + 0.026063935831189156, + 0.039464373141527176, + 0.034140922129154205, + 0.04560096934437752, + 0.04203878343105316, + 0.028449760749936104, + 0.03152333199977875, + 0.03522148355841637, + 0.007898261770606041, + 0.012697805650532246, + 0.007993398234248161, + 0.0078849196434021, + 0.013436384499073029, + 0.012253787368535995, + 0.005009869579225779, + 0.005511078052222729, + 0.00597715936601162, + 0.009762580506503582, + 0.008150276727974415, + 0.0057049961760640144, + 0.00899710413068533, + 0.010098426602780819, + 0.003797074081376195, + 0.004808955360203981, + 0.008041799999773502, + 0.012585691176354885, + 0.010593963786959648, + 0.009151313453912735, + 0.01313068438321352, + 0.006249835714697838, + 0.004790233913809061, + 0.009305193088948727, + 0.015046543441712856, + 0.02809811942279339, + 0.027902454137802124, + 0.011612757109105587, + 0.0197781790047884, + 0.01105467975139618, + 0.007197218015789986, + 0.0059167370200157166, + 0.004286321345716715, + 0.007338391616940498, + 0.009698458947241306, + 0.005896829068660736, + 0.005324011202901602, + 0.005327016580849886, + 0.004905315116047859, + 0.004020312335342169, + 0.009093012660741806, + 0.011846835725009441, + 0.018443813547492027, + 0.012164567597210407, + 0.012063347734510899, + 0.00812957901507616, + 0.006351211573928595, + 0.011901096440851688, + 0.004343233071267605, + 0.007044861558824778, + 0.0074054282158613205, + 0.004825021605938673, + 0.006727408152073622, + 0.004864986054599285, + 0.00382155436091125, + 0.006378255784511566, + 0.009416100569069386, + 0.016934433951973915, + 0.008372687734663486, + 0.014307329431176186, + 0.021558014675974846, + 0.013808340765535831, + 0.011218343861401081, + 0.009790838696062565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.004631743300706148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are hundreds of small, sleek drones visible converging from all directions towards the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the drones have gleaming metallic black shells with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the drones trail long, luminous digital light trails of electric blue and fiery magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at night with a deep, dark sky and a futuristic city skyline below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle and static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702485918998718 + ], + "opposite_logprob_tokens": [ + -2.1270248889923096 + ], + "logprob_reward": -0.12702485918998718, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702485918998718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1548977792263031, + 0.05518396198749542, + 0.09998907893896103, + 0.06614335626363754, + 0.1449636071920395, + 0.04669703543186188, + 0.056772585958242416, + 0.15199950337409973, + 0.05162765830755234, + 0.17172551155090332 + ], + "credit_map": [ + 0.014509135857224464, + 0.02193889580667019, + 0.02961442805826664, + 0.02788652665913105, + 0.020092930644750595, + 0.012078004889190197, + 0.00984943751245737, + 0.018928421661257744, + 0.005584032274782658, + 0.012427433393895626, + 0.005721337627619505, + 0.005341211333870888, + 0.01204913854598999, + 0.005345212295651436, + 0.004471981897950172, + 0.004243607632815838, + 0.0068546379916369915, + 0.017674246802926064, + 0.017955347895622253, + 0.008899466134607792, + 0.017064260318875313, + 0.013537432998418808, + 0.008684558793902397, + 0.009319114498794079, + 0.0058014062233269215, + 0.014065355993807316, + 0.01123530138283968, + 0.005202872212976217, + 0.012074009515345097, + 0.007012617774307728, + 0.004470498766750097, + 0.006281282752752304, + 0.0181355532258749, + 0.03162644803524017, + 0.015311384573578835, + 0.017996372655034065, + 0.02886063978075981, + 0.01113863568753004, + 0.01165679283440113, + 0.010237774811685085, + 0.005292645655572414, + 0.00827885139733553, + 0.008868356235325336, + 0.004713727161288261, + 0.007571544498205185, + 0.004912593401968479, + 0.004355611279606819, + 0.0027037004474550486, + 0.005944778211414814, + 0.008879308588802814, + 0.013503532856702805, + 0.006592338904738426, + 0.006905372254550457, + 0.00600845692679286, + 0.003994794562458992, + 0.004944001790136099, + 0.01737929880619049, + 0.025012603029608727, + 0.026308711618185043, + 0.014697481878101826, + 0.024359818547964096, + 0.018162770196795464, + 0.011856388300657272, + 0.014222415164113045, + 0.005494832526892424, + 0.007519757840782404, + 0.007294460665434599, + 0.003950296435505152, + 0.008571930229663849, + 0.005482956301420927, + 0.004039029590785503, + 0.00927438773214817, + 0.014238238334655762, + 0.02102765068411827, + 0.015120351687073708, + 0.0424405038356781, + 0.02104927971959114, + 0.018997544422745705, + 0.026067612692713737, + 0.012784316204488277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.12702485918998718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_tensor([3558, 3558])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_tensor([3558, 3558])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..45664ac3f99a3d37cc1c68441694a029b9750130 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_tensor([3558, 3558])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, dynamic tracking shot follows a man sprinting along a rainy city street. The camera stays in sync with him, emphasizing the speed and force of his run. The man, in his thirties, wears a well-tailored, dark charcoal gray suit, now completely soaked by the downpour, the expensive fabric clinging to his contours. His once-pristine white shirt is ripped at the collar, a crimson silk tie hanging loosely, swinging with his movements. The rain dampens his short, dark hair, strands clinging to his forehead and cheeks. His expression is tense, his jaw clenched, and his eyes reveal a mixture of despair and unwavering focus. Directly ahead of him lies a wide asphalt road, its surface a black mirror in the relentless rain, greedily reflecting the shimmering, iridescent neon lights of the tall buildings on either side—psychedelic pinks, cool blues, and glaring reds. With each strenuous step forward, his polished black Oxfords slammed heavily into the puddle, sending up circles of shimmering, neon-lit water, forming brief, dazzling arcs in the camera. The entire scene is imbued with a tense, urgent atmosphere, like a climactic chase scene in a neo-noir film. The stark contrast between the cool tones of the frame and the saturated neon highlights creates a dramatic tension that is both flamboyant and desperate.", + "seed": 20761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_3558_seed_20761_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04445593021810055, + "vlm_logprob": -0.04445593021810055, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04445593021810055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08782561123371124 + ], + "opposite_logprob_tokens": [ + -2.7128255367279053 + ], + "logprob_reward": -0.08782561123371124, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08782561123371124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13380104303359985, + 0.07314911484718323, + 0.03571326285600662, + 0.5043660402297974, + 0.05101536959409714, + 0.029768845066428185, + 0.04809771105647087, + 0.021114682778716087, + 0.06847792863845825, + 0.03449593484401703 + ], + "credit_map": [ + 0.008894401602447033, + 0.015727682039141655, + 0.014978255145251751, + 0.02182534709572792, + 0.011882591061294079, + 0.022587493062019348, + 0.01709105260670185, + 0.02081422321498394, + 0.005545794032514095, + 0.007334608118981123, + 0.009495222941040993, + 0.00987267680466175, + 0.010219915769994259, + 0.008141577243804932, + 0.013956225477159023, + 0.008583097718656063, + 0.003587219165638089, + 0.0032262837048619986, + 0.004952801391482353, + 0.0067682755179703236, + 0.005759711377322674, + 0.004548291675746441, + 0.0028295903466641903, + 0.004041088279336691, + 0.026663783937692642, + 0.031646303832530975, + 0.1829807013273239, + 0.18216370046138763, + 0.01660047098994255, + 0.015948280692100525, + 0.03474677726626396, + 0.013616041280329227, + 0.004266304895281792, + 0.009668615646660328, + 0.0043236794881522655, + 0.014746961183845997, + 0.003310754895210266, + 0.007531961891800165, + 0.0038948270957916975, + 0.0032722624018788338, + 0.0029430752620100975, + 0.004064888693392277, + 0.0026391902938485146, + 0.002894877688959241, + 0.005674282554537058, + 0.002375561511144042, + 0.007220855914056301, + 0.0019561138469725847, + 0.008343800902366638, + 0.007088208571076393, + 0.01021475251764059, + 0.004622782580554485, + 0.004689191933721304, + 0.004195630084723234, + 0.0049965521320700645, + 0.003946790937334299, + 0.0024743911344558, + 0.0026493521872907877, + 0.0032043098472058773, + 0.001914381398819387, + 0.0027495429385453463, + 0.0031591288279742002, + 0.002810589037835598, + 0.0021529882214963436, + 0.007151599042117596, + 0.00418463908135891, + 0.012503131292760372, + 0.005503740627318621, + 0.008935499005019665, + 0.005961897782981396, + 0.013519439846277237, + 0.010717982426285744, + 0.0030147270299494267, + 0.0020831262227147818, + 0.008744404651224613, + 0.0028388118371367455, + 0.0029295857530087233, + 0.002868346869945526, + 0.004616666119545698, + 0.007400267291814089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.08782561123371124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206833943724632 + ], + "opposite_logprob_tokens": [ + -2.812068223953247 + ], + "logprob_reward": -0.06206833943724632, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206833943724632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09577985107898712, + 0.047904182225465775, + 0.024574771523475647, + 0.7181723713874817, + 0.023983364924788475, + 0.007383781019598246, + 0.023280039429664612, + 0.026117227971553802, + 0.014059353619813919, + 0.01874503307044506 + ], + "credit_map": [ + 0.005752410273998976, + 0.006238615605980158, + 0.009128986857831478, + 0.02598959021270275, + 0.004068527836352587, + 0.03065955452620983, + 0.007159890606999397, + 0.0067822798155248165, + 0.003999172709882259, + 0.004563645459711552, + 0.009075969457626343, + 0.007080136798322201, + 0.004616925492882729, + 0.0036625396460294724, + 0.00996849499642849, + 0.004937298595905304, + 0.0019017882877960801, + 0.0019764676690101624, + 0.002958836732432246, + 0.0040741064585745335, + 0.002484823577105999, + 0.0024970367085188627, + 0.0032715376000851393, + 0.005410173907876015, + 0.03860868886113167, + 0.04562418907880783, + 0.26664042472839355, + 0.2697913944721222, + 0.01352212205529213, + 0.01970505155622959, + 0.04848240688443184, + 0.01579812355339527, + 0.0016611401224508882, + 0.006178951822221279, + 0.0022928540129214525, + 0.009463708847761154, + 0.0009940068703144789, + 0.0009694861364550889, + 0.0015788455493748188, + 0.0008443744154646993, + 0.0009178423206321895, + 0.0008197824354283512, + 0.0009429905330762267, + 0.0008951975614763796, + 0.0010482659563422203, + 0.0009372486383654177, + 0.0011807563714683056, + 0.0006416971446014941, + 0.0032985599245876074, + 0.002292696852236986, + 0.0051967548206448555, + 0.002236314583569765, + 0.00289032026194036, + 0.0017096559749916196, + 0.0035297188442200422, + 0.0021260182838886976, + 0.004294978454709053, + 0.00293337763287127, + 0.004437996074557304, + 0.0013921712525188923, + 0.003671608166769147, + 0.00476191658526659, + 0.0024690856225788593, + 0.00215609441511333, + 0.0014468258013948798, + 0.000916152261197567, + 0.0020890147425234318, + 0.0013604514533653855, + 0.0018199998885393143, + 0.0013646217994391918, + 0.0031187941785901785, + 0.0019434933783486485, + 0.0019635334610939026, + 0.0015542642213404179, + 0.003365016309544444, + 0.0031741708517074585, + 0.0018364401767030358, + 0.0014625213807448745, + 0.002329460345208645, + 0.003059625858440995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.06206833943724632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549791194498539 + ], + "opposite_logprob_tokens": [ + -2.929979085922241 + ], + "logprob_reward": -0.0549791194498539, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549791194498539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18765336275100708, + 0.08001720160245895, + 0.016592374071478844, + 0.5373066663742065, + 0.019400494173169136, + 0.027842259034514427, + 0.053843919187784195, + 0.02976326085627079, + 0.029290443286299706, + 0.018290024250745773 + ], + "credit_map": [ + 0.009400710463523865, + 0.013115908019244671, + 0.009354790672659874, + 0.050321903079748154, + 0.009115270338952541, + 0.068959079682827, + 0.011546705849468708, + 0.0158389899879694, + 0.006530072074383497, + 0.006443032529205084, + 0.010617616586387157, + 0.009653469547629356, + 0.010357912629842758, + 0.00937428418546915, + 0.017380472272634506, + 0.009660347364842892, + 0.0018265871331095695, + 0.0017006697598844767, + 0.0019292199285700917, + 0.0021680444478988647, + 0.0022972216829657555, + 0.0019107350381091237, + 0.0015653667505830526, + 0.0031945286318659782, + 0.029381293803453445, + 0.03398236259818077, + 0.1994086652994156, + 0.20254027843475342, + 0.009947236627340317, + 0.01460262294858694, + 0.035919029265642166, + 0.011525227688252926, + 0.0020534759387373924, + 0.003090623067691922, + 0.0013501237845048308, + 0.002980070887133479, + 0.0019913429860025644, + 0.004275496117770672, + 0.0017882988322526217, + 0.0018710604635998607, + 0.003702452639117837, + 0.004955093376338482, + 0.0035764367785304785, + 0.004809390287846327, + 0.0024015363305807114, + 0.0030169389210641384, + 0.003384467214345932, + 0.001995942322537303, + 0.009632598608732224, + 0.0057096416130661964, + 0.010807684622704983, + 0.004154462832957506, + 0.006988088134676218, + 0.0046293423511087894, + 0.006432347930967808, + 0.005489753559231758, + 0.004778666887432337, + 0.003854253562167287, + 0.0048044705763459206, + 0.0022167693823575974, + 0.0031243287958204746, + 0.0033379429951310158, + 0.004986514803022146, + 0.002660316415131092, + 0.002783081727102399, + 0.0023255159612745047, + 0.0054131909273564816, + 0.003654221538454294, + 0.0028132351581007242, + 0.002904698019847274, + 0.0055239698849618435, + 0.0038725300692021847, + 0.0016606678254902363, + 0.001720037660561502, + 0.0027536479756236076, + 0.0023973812349140644, + 0.0021266601979732513, + 0.0020426653791218996, + 0.0025892050471156836, + 0.0029997597448527813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0549791194498539 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629659026861191 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629659026861191, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629659026861191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2939296066761017, + 0.038923513144254684, + 0.08814540505409241, + 0.23516800999641418, + 0.11632086336612701, + 0.05212787166237831, + 0.043497245758771896, + 0.023318661376833916, + 0.06648164242506027, + 0.042087193578481674 + ], + "credit_map": [ + 0.02281000465154648, + 0.020717835053801537, + 0.04912785813212395, + 0.058545418083667755, + 0.013037919998168945, + 0.08679128438234329, + 0.0215905848890543, + 0.02130870148539543, + 0.004224142525345087, + 0.004975306801497936, + 0.0061925258487463, + 0.004833358805626631, + 0.004249067045748234, + 0.004052366595715284, + 0.004592887591570616, + 0.00580385560169816, + 0.010583418421447277, + 0.012538379058241844, + 0.011709867045283318, + 0.01489180512726307, + 0.012408885173499584, + 0.0062171658501029015, + 0.010974867269396782, + 0.008821005932986736, + 0.01362198032438755, + 0.017498673871159554, + 0.08218812942504883, + 0.08706871420145035, + 0.006103955674916506, + 0.007635420188307762, + 0.015886375680565834, + 0.00516474200412631, + 0.013443917036056519, + 0.02839667536318302, + 0.01848759688436985, + 0.03052249364554882, + 0.00647253030911088, + 0.006060710176825523, + 0.007420218084007502, + 0.005516714882105589, + 0.007398848421871662, + 0.01532799657434225, + 0.006492236163467169, + 0.005823843646794558, + 0.003685497911646962, + 0.0035010974388569593, + 0.006319741252809763, + 0.0035786021035164595, + 0.006082586478441954, + 0.00502490159124136, + 0.008603062480688095, + 0.006646753754466772, + 0.0037760951090604067, + 0.0032393792644143105, + 0.00549119058996439, + 0.004633272998034954, + 0.0033016190864145756, + 0.002335798926651478, + 0.003067447803914547, + 0.0017170527717098594, + 0.003071932354941964, + 0.003397570224478841, + 0.003971435595303774, + 0.002455802634358406, + 0.006759974639862776, + 0.005908002145588398, + 0.013061932288110256, + 0.006382979452610016, + 0.0059315795078873634, + 0.0074231005273759365, + 0.008692958392202854, + 0.012321114540100098, + 0.004701714962720871, + 0.004235691856592894, + 0.00983394030481577, + 0.0035774516873061657, + 0.004256756044924259, + 0.004420324694365263, + 0.004353643860667944, + 0.006707668304443359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.007629659026861191 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776921942830086 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009776921942830086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776921942830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1975635290145874, + 0.036720335483551025, + 0.03297039866447449, + 0.5990533828735352, + 0.025759171694517136, + 0.015502061694860458, + 0.023528072983026505, + 0.0158967524766922, + 0.02724534273147583, + 0.025761058554053307 + ], + "credit_map": [ + 0.013939260505139828, + 0.01515195146203041, + 0.02029068022966385, + 0.03906126692891121, + 0.014391499571502209, + 0.051492661237716675, + 0.01826045848429203, + 0.024975720793008804, + 0.0037410175427794456, + 0.0031938247848302126, + 0.007175071630626917, + 0.006438204552978277, + 0.004673941992223263, + 0.0034701297990977764, + 0.004018947947770357, + 0.004009195603430271, + 0.004059709142893553, + 0.004770361818373203, + 0.004745514132082462, + 0.006058523431420326, + 0.004382514860481024, + 0.0023061588872224092, + 0.0034754108637571335, + 0.003172203665599227, + 0.0323915034532547, + 0.03727004677057266, + 0.22290202975273132, + 0.22648179531097412, + 0.010292322374880314, + 0.016232222318649292, + 0.0405154675245285, + 0.012967943213880062, + 0.0023462779354304075, + 0.005423250608146191, + 0.0020022536627948284, + 0.01007267739623785, + 0.0010198777308687568, + 0.0015791791956871748, + 0.0020245229825377464, + 0.0012911299709230661, + 0.0020328231621533632, + 0.0018888708436861634, + 0.0017319495091214776, + 0.0022136366460472345, + 0.0022069173865020275, + 0.0012983804335817695, + 0.0024676022585481405, + 0.0016618804074823856, + 0.0029689460061490536, + 0.0032249141950160265, + 0.004423969425261021, + 0.003224958200007677, + 0.0021523500327020884, + 0.0019126234110444784, + 0.0036471534986048937, + 0.0019731568172574043, + 0.0020648480858653784, + 0.002251147758215666, + 0.0023692199029028416, + 0.0012542061740532517, + 0.002237798646092415, + 0.002020308980718255, + 0.0021596241276711226, + 0.0015395971713587642, + 0.0028062222991138697, + 0.001461152220144868, + 0.005371584556996822, + 0.001794559066183865, + 0.0033584556076675653, + 0.0024401461705565453, + 0.005544687621295452, + 0.004468534141778946, + 0.002339943777769804, + 0.001790556707419455, + 0.00526969600468874, + 0.004673202987760305, + 0.0025391236413270235, + 0.0015304500702768564, + 0.0035298822913318872, + 0.004088200628757477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.009776921942830086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a man sprinting along a rainy city street. The camera stays in sync with him, emphasizing the speed and force of his run. The man, in his thirties, wears a well-tailored, dark charcoal gray suit, now completely soaked by the downpour, the expensive fabric clinging to his contours. His once-pristine white shirt is ripped at the collar, a crimson silk tie hanging loosely, swinging with his movements. The rain dampens his short, dark hair, strands clinging to his forehead and cheeks. His expression is tense, his jaw clenched, and his eyes reveal a mixture of despair and unwavering focus. Directly ahead of him lies a wide asphalt road, its surface a black mirror in the relentless rain, greedily reflecting the shimmering, iridescent neon lights of the tall buildings on either side—psychedelic pinks, cool blues, and glaring reds. With each strenuous step forward, his polished black Oxfords slammed heavily into the puddle, sending up circles of shimmering, neon-lit water, forming brief, dazzling arcs in the camera. The entire scene is imbued with a tense, urgent atmosphere, like a climactic chase scene in a neo-noir film. The stark contrast between the cool tones of the frame and the saturated neon highlights creates a dramatic tension that is both flamboyant and desperate.", + "seed": 20762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_3558_seed_20762_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05656685270369053, + "vlm_logprob": -0.05656685270369053, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.05656685270369053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12363184243440628 + ], + "opposite_logprob_tokens": [ + -2.2486319541931152 + ], + "logprob_reward": -0.12363184243440628, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12363184243440628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1468103677034378, + 0.02471773512661457, + 0.3940541744232178, + 0.02489103004336357, + 0.03978193178772926, + 0.04984098672866821, + 0.08462098240852356, + 0.022998737171292305, + 0.17215482890605927, + 0.04012926295399666 + ], + "credit_map": [ + 0.010322390124201775, + 0.009178265929222107, + 0.030393725261092186, + 0.023884519934654236, + 0.0326739102602005, + 0.006337669678032398, + 0.02220527082681656, + 0.011814617551863194, + 0.0020614294335246086, + 0.0025164843536913395, + 0.00405800249427557, + 0.00399050023406744, + 0.0033894521184265614, + 0.0017374128801748157, + 0.0045957849361002445, + 0.0023686669301241636, + 0.03845374658703804, + 0.023621559143066406, + 0.1777513027191162, + 0.09694382548332214, + 0.01621282286942005, + 0.012434884905815125, + 0.01856747828423977, + 0.010068527422845364, + 0.0028872000984847546, + 0.0017140604322776198, + 0.003539720783010125, + 0.0030697558540850878, + 0.003324915189296007, + 0.0021171567495912313, + 0.003008051309734583, + 0.005230167880654335, + 0.005326425656676292, + 0.007419065572321415, + 0.004775534849613905, + 0.0035585355944931507, + 0.003840691177174449, + 0.005088405217975378, + 0.0048095365054905415, + 0.004963733721524477, + 0.0043085236102342606, + 0.0038090995512902737, + 0.0072279698215425014, + 0.005118480417877436, + 0.008444856852293015, + 0.008764860220253468, + 0.005285909399390221, + 0.006881279405206442, + 0.007870259694755077, + 0.007648355793207884, + 0.011498554609715939, + 0.011037061922252178, + 0.012359632179141045, + 0.008449061773717403, + 0.014768105000257492, + 0.01098993793129921, + 0.0029710843227803707, + 0.001997716259211302, + 0.0028368046041578054, + 0.0019183624535799026, + 0.0018388812895864248, + 0.005825706757605076, + 0.002923531923443079, + 0.00268664862960577, + 0.014369520358741283, + 0.028617987409234047, + 0.022122016176581383, + 0.026831578463315964, + 0.015322452411055565, + 0.017462730407714844, + 0.021869922056794167, + 0.02555861696600914, + 0.003648789832368493, + 0.0032675773836672306, + 0.004132934380322695, + 0.004605296067893505, + 0.008694727905094624, + 0.0037823249585926533, + 0.007307682652026415, + 0.0046899267472326756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.12363184243440628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900453358888626 + ], + "opposite_logprob_tokens": [ + -2.5790045261383057 + ], + "logprob_reward": -0.07900453358888626, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900453358888626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09760333597660065, + 0.041335657238960266, + 0.40577179193496704, + 0.028498930856585503, + 0.10544877499341965, + 0.05697739124298096, + 0.1244746595621109, + 0.040162261575460434, + 0.0699140653014183, + 0.02981310710310936 + ], + "credit_map": [ + 0.009051625616848469, + 0.00776073569431901, + 0.018025873228907585, + 0.01640218310058117, + 0.018418269231915474, + 0.007892198860645294, + 0.0120431212708354, + 0.008009334094822407, + 0.0035008161794394255, + 0.004662984982132912, + 0.005375029519200325, + 0.006595434620976448, + 0.006706177722662687, + 0.0038895043544471264, + 0.0056707412004470825, + 0.004934967961162329, + 0.03678765147924423, + 0.0254686176776886, + 0.17722058296203613, + 0.09776785224676132, + 0.018939349800348282, + 0.014813845045864582, + 0.02260640077292919, + 0.01216751430183649, + 0.0035502712707966566, + 0.0018678922206163406, + 0.003671059152111411, + 0.002703202422708273, + 0.003943313844501972, + 0.0030681556090712547, + 0.003755244193598628, + 0.005939794238656759, + 0.010224498808383942, + 0.01544615812599659, + 0.01095578819513321, + 0.015862176194787025, + 0.011156418360769749, + 0.009229916147887707, + 0.017539823427796364, + 0.015034004114568233, + 0.0043936544097959995, + 0.0058351121842861176, + 0.004738849122077227, + 0.006827985867857933, + 0.010851841419935226, + 0.008633910678327084, + 0.010931426659226418, + 0.0047646164894104, + 0.011169780977070332, + 0.01561419665813446, + 0.010746399872004986, + 0.015778252854943275, + 0.014786452986299992, + 0.0124839898198843, + 0.02262117527425289, + 0.021274428814649582, + 0.0037165675312280655, + 0.004093076102435589, + 0.006789970211684704, + 0.003344322321936488, + 0.0041980748064816, + 0.008836986497044563, + 0.004932597279548645, + 0.004250670783221722, + 0.010428708977997303, + 0.007023658603429794, + 0.013410255312919617, + 0.010074509307742119, + 0.005406277719885111, + 0.0067164096981287, + 0.007577074691653252, + 0.00927717424929142, + 0.002482029376551509, + 0.002613453194499016, + 0.0048967162147164345, + 0.003880580421537161, + 0.004763798322528601, + 0.0032337873708456755, + 0.003994808066636324, + 0.003947936464101076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.07900453358888626 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982192672789097 + ], + "opposite_logprob_tokens": [ + -3.5298218727111816 + ], + "logprob_reward": -0.02982192672789097, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982192672789097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07386064529418945, + 0.028931500390172005, + 0.04607246443629265, + 0.03157666698098183, + 0.020523201674222946, + 0.10052836686372757, + 0.5803108215332031, + 0.027468135580420494, + 0.05875751003623009, + 0.03197067603468895 + ], + "credit_map": [ + 0.00842270813882351, + 0.010485509410500526, + 0.013759434223175049, + 0.009839636273682117, + 0.011861434206366539, + 0.004299769643694162, + 0.008562961593270302, + 0.00662919320166111, + 0.0021663829684257507, + 0.0023163731675595045, + 0.0038029311690479517, + 0.00441505154594779, + 0.0038604617584496737, + 0.002363009611144662, + 0.0070938714779913425, + 0.0029134193900972605, + 0.00945836491882801, + 0.004522016271948814, + 0.008937646634876728, + 0.0068922340869903564, + 0.006039052736014128, + 0.002575682708993554, + 0.004544693976640701, + 0.0031027686782181263, + 0.002977625699713826, + 0.003441431326791644, + 0.0027237609028816223, + 0.005198175087571144, + 0.004730054643005133, + 0.0030978217255324125, + 0.0057687475346028805, + 0.0036390505265444517, + 0.0027853501960635185, + 0.002466728677973151, + 0.0029857761692255735, + 0.0022063192445784807, + 0.002494045067578554, + 0.002084889682009816, + 0.0029910625889897346, + 0.002509030280634761, + 0.005393523257225752, + 0.005027813371270895, + 0.0053919521160423756, + 0.00801959726959467, + 0.01045521255582571, + 0.02791965939104557, + 0.012487009167671204, + 0.025833595544099808, + 0.05633225291967392, + 0.07268810272216797, + 0.05320630222558975, + 0.0829986035823822, + 0.06312074512243271, + 0.05025463178753853, + 0.11318212002515793, + 0.08852803707122803, + 0.0031978925690054893, + 0.0022802765015512705, + 0.004232132341712713, + 0.0027669367846101522, + 0.002315191086381674, + 0.006943108979612589, + 0.0028886166401207447, + 0.0028439811430871487, + 0.005779646802693605, + 0.008027751930058002, + 0.008161946199834347, + 0.0074064359068870544, + 0.005417760461568832, + 0.006518912501633167, + 0.008842583745718002, + 0.008602474816143513, + 0.0031051384285092354, + 0.0031382963061332703, + 0.004162035882472992, + 0.003750600153580308, + 0.006052691489458084, + 0.0030871375929564238, + 0.005077862646430731, + 0.0035969135351479053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.02982192672789097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979809045791626 + ], + "opposite_logprob_tokens": [ + -3.5297980308532715 + ], + "logprob_reward": -0.02979809045791626, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979809045791626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03127236291766167, + 0.030501119792461395, + 0.8557884097099304, + 0.008856242522597313, + 0.00569615513086319, + 0.013475224375724792, + 0.009045006707310677, + 0.032142117619514465, + 0.008250216953456402, + 0.004973097704350948 + ], + "credit_map": [ + 0.0025326984468847513, + 0.0025408954825252295, + 0.006759130861610174, + 0.004938442260026932, + 0.0060790046118199825, + 0.0015808154130354524, + 0.00410498259589076, + 0.0027363940607756376, + 0.004284634254872799, + 0.0036770941223949194, + 0.003974749241024256, + 0.004374244250357151, + 0.003543479135259986, + 0.0023504497949033976, + 0.002320856088772416, + 0.005975613370537758, + 0.07400201261043549, + 0.04978901892900467, + 0.38463857769966125, + 0.2202572077512741, + 0.034310903400182724, + 0.029133591800928116, + 0.04145420715212822, + 0.022202907130122185, + 0.0011136927641928196, + 0.0009562199120409787, + 0.0010153050534427166, + 0.0011463130358606577, + 0.0005535667878575623, + 0.0011419771471992135, + 0.0005195452831685543, + 0.0024096225388348103, + 0.0006928631337359548, + 0.0010265216697007418, + 0.000667648040689528, + 0.0005191809614188969, + 0.0007036937167868018, + 0.0005425306735560298, + 0.0007244490552693605, + 0.0008192674722522497, + 0.0013027633540332317, + 0.0013472236460074782, + 0.0021600802429020405, + 0.0014005252160131931, + 0.0017255350248888135, + 0.0023072627373039722, + 0.0015004429733380675, + 0.001731390948407352, + 0.0007240092963911593, + 0.0011745607480406761, + 0.0009949079249054193, + 0.0013905108207836747, + 0.0013081257930025458, + 0.0010109739378094673, + 0.0014459829544648528, + 0.0009959351737052202, + 0.0036785926204174757, + 0.003211511066183448, + 0.005247552413493395, + 0.004176333546638489, + 0.003202105639502406, + 0.005920798517763615, + 0.0032751874532550573, + 0.003430034965276718, + 0.0010422335471957922, + 0.0010305834002792835, + 0.0009863142622634768, + 0.0011190446093678474, + 0.0007898643962107599, + 0.0010467618703842163, + 0.0011719194008037448, + 0.0010634958744049072, + 0.0004010756965726614, + 0.0005192196695134044, + 0.0007528795977123082, + 0.0006375760422088206, + 0.0009252431918866932, + 0.0004945998662151396, + 0.0005529556656256318, + 0.0006895482656545937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.02979809045791626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.158018559217453, + 0.03252119570970535, + 0.30761152505874634, + 0.03306020796298981, + 0.08040367811918259, + 0.08989998698234558, + 0.058313265442848206, + 0.09278645366430283, + 0.0789656862616539, + 0.06841941177845001 + ], + "credit_map": [ + 0.022207025438547134, + 0.02064679190516472, + 0.02164573408663273, + 0.02022743783891201, + 0.031232217326760292, + 0.009320227429270744, + 0.02081214264035225, + 0.011926968581974506, + 0.0039263260550796986, + 0.003934969659894705, + 0.0038087251596152782, + 0.004337027203291655, + 0.0040525346994400024, + 0.002533274469897151, + 0.004693468101322651, + 0.005234866868704557, + 0.02708839438855648, + 0.018590793013572693, + 0.13641074299812317, + 0.07822757214307785, + 0.013452219776809216, + 0.009943266399204731, + 0.01601456291973591, + 0.007883977144956589, + 0.003218726720660925, + 0.0042191920801997185, + 0.0047151739709079266, + 0.004559357650578022, + 0.005429616663604975, + 0.003660485614091158, + 0.003410633420571685, + 0.0038470157887786627, + 0.009462536312639713, + 0.016544830054044724, + 0.008172411471605301, + 0.008825779892504215, + 0.008513555862009525, + 0.009519606828689575, + 0.008078777231276035, + 0.01128617487847805, + 0.005420289002358913, + 0.0052008130587637424, + 0.007067276630550623, + 0.008433379232883453, + 0.011203542351722717, + 0.021126052364706993, + 0.01755240559577942, + 0.01389621663838625, + 0.004745247773826122, + 0.006993413902819157, + 0.005799286067485809, + 0.010068120434880257, + 0.006240835878998041, + 0.006825525779277086, + 0.012277424335479736, + 0.005363405216485262, + 0.007787203416228294, + 0.006831815931946039, + 0.01835925132036209, + 0.008966751396656036, + 0.007385054137557745, + 0.026518508791923523, + 0.009832504205405712, + 0.0071053653955459595, + 0.007773945573717356, + 0.010375156998634338, + 0.010901108384132385, + 0.010023967362940311, + 0.008659043349325657, + 0.008471938781440258, + 0.011814695782959461, + 0.010945833288133144, + 0.006491734180599451, + 0.008563621900975704, + 0.007193694822490215, + 0.009679297916591167, + 0.010517382994294167, + 0.007946635596454144, + 0.007750832010060549, + 0.010276209563016891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.020577870309352875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a man sprinting along a rainy city street. The camera stays in sync with him, emphasizing the speed and force of his run. The man, in his thirties, wears a well-tailored, dark charcoal gray suit, now completely soaked by the downpour, the expensive fabric clinging to his contours. His once-pristine white shirt is ripped at the collar, a crimson silk tie hanging loosely, swinging with his movements. The rain dampens his short, dark hair, strands clinging to his forehead and cheeks. His expression is tense, his jaw clenched, and his eyes reveal a mixture of despair and unwavering focus. Directly ahead of him lies a wide asphalt road, its surface a black mirror in the relentless rain, greedily reflecting the shimmering, iridescent neon lights of the tall buildings on either side—psychedelic pinks, cool blues, and glaring reds. With each strenuous step forward, his polished black Oxfords slammed heavily into the puddle, sending up circles of shimmering, neon-lit water, forming brief, dazzling arcs in the camera. The entire scene is imbued with a tense, urgent atmosphere, like a climactic chase scene in a neo-noir film. The stark contrast between the cool tones of the frame and the saturated neon highlights creates a dramatic tension that is both flamboyant and desperate.", + "seed": 20763, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_3558_seed_20763_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.049901886656880376, + "vlm_logprob": -0.049901886656880376, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.049901886656880376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08568403124809265 + ], + "opposite_logprob_tokens": [ + -2.710684061050415 + ], + "logprob_reward": -0.08568403124809265, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08568403124809265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08728642761707306, + 0.08773317188024521, + 0.10689488053321838, + 0.15624217689037323, + 0.22796913981437683, + 0.05788309872150421, + 0.060338716953992844, + 0.02200208604335785, + 0.02682970091700554, + 0.16682060062885284 + ], + "credit_map": [ + 0.006539713591337204, + 0.009245282039046288, + 0.01036650501191616, + 0.01879841648042202, + 0.007777738850563765, + 0.01026588398963213, + 0.01205225009471178, + 0.012240641750395298, + 0.007673812564462423, + 0.00704943435266614, + 0.010760506615042686, + 0.02385706827044487, + 0.009910286404192448, + 0.004814246203750372, + 0.008836043067276478, + 0.014831778593361378, + 0.008818576112389565, + 0.012819623574614525, + 0.008632601238787174, + 0.015562593005597591, + 0.018767205998301506, + 0.011553141288459301, + 0.010803921148180962, + 0.019937213510274887, + 0.015876930207014084, + 0.01793271116912365, + 0.039223771542310715, + 0.016461411491036415, + 0.03282676637172699, + 0.010435680858790874, + 0.011805619113147259, + 0.011679278686642647, + 0.0221072006970644, + 0.020279670134186745, + 0.04420873522758484, + 0.024814827367663383, + 0.0281012412160635, + 0.018779169768095016, + 0.04256700351834297, + 0.027111297473311424, + 0.007321596145629883, + 0.006911901757121086, + 0.008026695810258389, + 0.00872403010725975, + 0.006371357012540102, + 0.007774703670293093, + 0.0064402068965137005, + 0.006312611512839794, + 0.0064166816882789135, + 0.006881055887788534, + 0.010236167348921299, + 0.005479548126459122, + 0.006145050283521414, + 0.008198787458240986, + 0.009609137661755085, + 0.00737229036167264, + 0.0020745445508509874, + 0.00307062198407948, + 0.0026203501038253307, + 0.0035800167825073004, + 0.002337925136089325, + 0.0034896796569228172, + 0.0020883879624307156, + 0.0027405600994825363, + 0.0017185764154419303, + 0.003349542850628495, + 0.00249024061486125, + 0.00369193940423429, + 0.004471893422305584, + 0.004528720397502184, + 0.0028304371517151594, + 0.0037483517080545425, + 0.010165928862988949, + 0.018453657627105713, + 0.015467174351215363, + 0.025311045348644257, + 0.02540716901421547, + 0.0241668950766325, + 0.029348360374569893, + 0.018500380218029022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.08568403124809265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905854821205139 + ], + "opposite_logprob_tokens": [ + -2.4640586376190186 + ], + "logprob_reward": -0.08905854821205139, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905854821205139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676068753004074, + 0.061544258147478104, + 0.09830950945615768, + 0.0650842934846878, + 0.08078484237194061, + 0.018133359029889107, + 0.04179558902978897, + 0.01703660748898983, + 0.09751518815755844, + 0.3521895110607147 + ], + "credit_map": [ + 0.012508070096373558, + 0.016271429136395454, + 0.019275695085525513, + 0.024730954319238663, + 0.01627994515001774, + 0.030934860929846764, + 0.023779619485139847, + 0.023826299235224724, + 0.005671083461493254, + 0.0077421036548912525, + 0.007544379215687513, + 0.015145166777074337, + 0.005113504361361265, + 0.0039314888417720795, + 0.005917176604270935, + 0.010479351505637169, + 0.007321049924939871, + 0.00614283699542284, + 0.005963243078440428, + 0.030181152746081352, + 0.006749155465513468, + 0.005864348728209734, + 0.006475523114204407, + 0.02961220033466816, + 0.008352472446858883, + 0.007878932170569897, + 0.0075724199414253235, + 0.012400456704199314, + 0.010984666645526886, + 0.004492027219384909, + 0.006697875447571278, + 0.006705437321215868, + 0.0088598458096385, + 0.0073239244520664215, + 0.011608668603003025, + 0.008849930949509144, + 0.009629298932850361, + 0.012082891538739204, + 0.012241818010807037, + 0.010188463144004345, + 0.0022976896725594997, + 0.002242601942270994, + 0.0018607412930577993, + 0.0019481530180200934, + 0.002041424857452512, + 0.0032491423189640045, + 0.002407606691122055, + 0.002085999818518758, + 0.005358547903597355, + 0.004055882338434458, + 0.005684785544872284, + 0.004523205570876598, + 0.0039789327420294285, + 0.006110085174441338, + 0.006566622760146856, + 0.005517531652003527, + 0.0018640008056536317, + 0.0019327666377648711, + 0.0021090195514261723, + 0.0028972248546779156, + 0.0016740813152864575, + 0.0021829712204635143, + 0.0015124629717320204, + 0.0028640786185860634, + 0.007746716495603323, + 0.006472117733210325, + 0.010535386390984058, + 0.006866947747766972, + 0.033057376742362976, + 0.008858157321810722, + 0.01327374018728733, + 0.010704746469855309, + 0.02519470453262329, + 0.03453075513243675, + 0.0336526595056057, + 0.04971561208367348, + 0.053191155195236206, + 0.04094577953219414, + 0.07238250970840454, + 0.04257633537054062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.08905854821205139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038119908422231674 + ], + "opposite_logprob_tokens": [ + -3.2881197929382324 + ], + "logprob_reward": -0.038119908422231674, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038119908422231674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14669360220432281, + 0.11430293321609497, + 0.22555561363697052, + 0.08111986517906189, + 0.08371804654598236, + 0.03246350213885307, + 0.030791165307164192, + 0.01419256255030632, + 0.06119691580533981, + 0.20996583998203278 + ], + "credit_map": [ + 0.012442315928637981, + 0.022052491083741188, + 0.017674269154667854, + 0.025898508727550507, + 0.012297909706830978, + 0.023237688466906548, + 0.014455646276473999, + 0.018634775653481483, + 0.009029225446283817, + 0.010209771804511547, + 0.006327979266643524, + 0.035969242453575134, + 0.011554882861673832, + 0.0057434444315731525, + 0.009809296578168869, + 0.025659089908003807, + 0.015631165355443954, + 0.01476196013391018, + 0.013262799941003323, + 0.07090238481760025, + 0.018065880984067917, + 0.012136109173297882, + 0.013289476744830608, + 0.067505843937397, + 0.010286156088113785, + 0.01007546205073595, + 0.007308816071599722, + 0.01405983418226242, + 0.0172842089086771, + 0.008521109819412231, + 0.006850464269518852, + 0.006733810994774103, + 0.012572952546179295, + 0.008322331123054028, + 0.016844531521201134, + 0.008655471727252007, + 0.00826620776206255, + 0.011545645073056221, + 0.008888499811291695, + 0.008622403256595135, + 0.003224041312932968, + 0.003947952762246132, + 0.005343837197870016, + 0.0043335529044270515, + 0.003524658503010869, + 0.004614728968590498, + 0.003337668487802148, + 0.0041370633989572525, + 0.002699590753763914, + 0.002985627157613635, + 0.0030302375089377165, + 0.0032403445802628994, + 0.004629930481314659, + 0.005030677188187838, + 0.004263359121978283, + 0.004911397583782673, + 0.0012715726625174284, + 0.002001984277740121, + 0.002060660859569907, + 0.0014882185496389866, + 0.0020986171439290047, + 0.0017627701163291931, + 0.0014508290914818645, + 0.0020579099655151367, + 0.005887407809495926, + 0.003946976270526648, + 0.004515381529927254, + 0.005842446815222502, + 0.020404405891895294, + 0.00807658489793539, + 0.006889317184686661, + 0.005634396802634001, + 0.015010787174105644, + 0.0203610397875309, + 0.019915003329515457, + 0.03186194226145744, + 0.03090132400393486, + 0.026614466682076454, + 0.041370633989572525, + 0.0239306278526783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.038119908422231674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070127487182617 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070127487182617, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070127487182617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07244165241718292, + 0.07508198916912079, + 0.21125468611717224, + 0.028933826833963394, + 0.36631613969802856, + 0.05464382842183113, + 0.05789574980735779, + 0.030456630513072014, + 0.07536891847848892, + 0.027606569230556488 + ], + "credit_map": [ + 0.006037491373717785, + 0.007356839254498482, + 0.010653422214090824, + 0.009244886226952076, + 0.006363505031913519, + 0.01067526824772358, + 0.010015362873673439, + 0.01209487859159708, + 0.009924151003360748, + 0.0064011067152023315, + 0.006860432680696249, + 0.009921046905219555, + 0.007256897632032633, + 0.00919866468757391, + 0.007537492085248232, + 0.01798219606280327, + 0.015402175486087799, + 0.012698769569396973, + 0.01480778306722641, + 0.0588265098631382, + 0.013903754763305187, + 0.017954163253307343, + 0.015973126515746117, + 0.06168839707970619, + 0.003157815197482705, + 0.0037649066653102636, + 0.0038324096240103245, + 0.004693391267210245, + 0.004040600266307592, + 0.003503270447254181, + 0.003058526199311018, + 0.0028829064685851336, + 0.057217907160520554, + 0.02955324575304985, + 0.09449166059494019, + 0.04161006584763527, + 0.02644439972937107, + 0.04373772069811821, + 0.03678519278764725, + 0.03647593781352043, + 0.005439519416540861, + 0.004940278362482786, + 0.013228686526417732, + 0.006490018684417009, + 0.004799431189894676, + 0.008644471876323223, + 0.004335045348852873, + 0.006766374222934246, + 0.008018613792955875, + 0.011292736977338791, + 0.007232732605189085, + 0.00491173192858696, + 0.006860125809907913, + 0.006845495663583279, + 0.006809882819652557, + 0.005924432538449764, + 0.002458530478179455, + 0.003093597711995244, + 0.00402548024430871, + 0.005632190965116024, + 0.003497765166684985, + 0.006068161688745022, + 0.0025423697661608458, + 0.003138534026220441, + 0.00732044130563736, + 0.0036884115543216467, + 0.008068867959082127, + 0.004635918419808149, + 0.026763584464788437, + 0.007320955395698547, + 0.007537327706813812, + 0.010033412836492062, + 0.002104058861732483, + 0.0031466882210224867, + 0.002806030912324786, + 0.00488869147375226, + 0.0038078746292740107, + 0.003074170555919409, + 0.005167752504348755, + 0.0026113013736903667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.016070127487182617 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057681791484356 + ], + "opposite_logprob_tokens": [ + -3.8955767154693604 + ], + "logprob_reward": -0.02057681791484356, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057681791484356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1303722858428955, + 0.1153578832745552, + 0.12396026402711868, + 0.06477941572666168, + 0.2031547576189041, + 0.05303603410720825, + 0.04617231339216232, + 0.0669066533446312, + 0.057113952934741974, + 0.13914650678634644 + ], + "credit_map": [ + 0.010137463919818401, + 0.012913446873426437, + 0.014056690037250519, + 0.017277073115110397, + 0.0126798739656806, + 0.02605806104838848, + 0.016431929543614388, + 0.020817745476961136, + 0.011372760869562626, + 0.010289611294865608, + 0.007990860380232334, + 0.01774122565984726, + 0.015385693870484829, + 0.013318773359060287, + 0.009574824944138527, + 0.029684120789170265, + 0.008591968566179276, + 0.00777587341144681, + 0.006969674490392208, + 0.043808430433273315, + 0.005193447228521109, + 0.006493214517831802, + 0.006902167573571205, + 0.03822547569870949, + 0.008097205311059952, + 0.008831769227981567, + 0.010296570137143135, + 0.007784275338053703, + 0.011352035216987133, + 0.005960788577795029, + 0.0061839609406888485, + 0.006272809114307165, + 0.022568903863430023, + 0.014064866118133068, + 0.04266439378261566, + 0.022445037961006165, + 0.020579688251018524, + 0.02562316507101059, + 0.030665207654237747, + 0.024543464183807373, + 0.004903135355561972, + 0.00510087376460433, + 0.009389390237629414, + 0.007511907257139683, + 0.005040594842284918, + 0.008847575634717941, + 0.005202555097639561, + 0.007039997726678848, + 0.008307074196636677, + 0.006204680539667606, + 0.00556202745065093, + 0.004507966339588165, + 0.005195743404328823, + 0.005398052278906107, + 0.0060289776884019375, + 0.004967784974724054, + 0.009001115337014198, + 0.011918067000806332, + 0.005560506135225296, + 0.007094032131135464, + 0.009040932171046734, + 0.010793926194310188, + 0.004996020346879959, + 0.008502049371600151, + 0.00590558024123311, + 0.006411266978830099, + 0.00659620575606823, + 0.005909110885113478, + 0.013698214665055275, + 0.005972520913928747, + 0.008115091361105442, + 0.004505959339439869, + 0.009341642260551453, + 0.014260291121900082, + 0.012779452838003635, + 0.02122729830443859, + 0.018517062067985535, + 0.020503442734479904, + 0.028185542672872543, + 0.014331763610243797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.02057681791484356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a man sprinting along a rainy city street. The camera stays in sync with him, emphasizing the speed and force of his run. The man, in his thirties, wears a well-tailored, dark charcoal gray suit, now completely soaked by the downpour, the expensive fabric clinging to his contours. His once-pristine white shirt is ripped at the collar, a crimson silk tie hanging loosely, swinging with his movements. The rain dampens his short, dark hair, strands clinging to his forehead and cheeks. His expression is tense, his jaw clenched, and his eyes reveal a mixture of despair and unwavering focus. Directly ahead of him lies a wide asphalt road, its surface a black mirror in the relentless rain, greedily reflecting the shimmering, iridescent neon lights of the tall buildings on either side—psychedelic pinks, cool blues, and glaring reds. With each strenuous step forward, his polished black Oxfords slammed heavily into the puddle, sending up circles of shimmering, neon-lit water, forming brief, dazzling arcs in the camera. The entire scene is imbued with a tense, urgent atmosphere, like a climactic chase scene in a neo-noir film. The stark contrast between the cool tones of the frame and the saturated neon highlights creates a dramatic tension that is both flamboyant and desperate.", + "seed": 20764, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_7/index_3558_seed_20764_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05123941488564014, + "vlm_logprob": -0.05123941488564014, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.05123941488564014, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11097472906112671 + ], + "opposite_logprob_tokens": [ + -2.3609747886657715 + ], + "logprob_reward": -0.11097472906112671, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11097472906112671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17067018151283264, + 0.13804024457931519, + 0.11055471003055573, + 0.054844483733177185, + 0.05907279998064041, + 0.06800118833780289, + 0.09933705627918243, + 0.07330574095249176, + 0.14125090837478638, + 0.084922656416893 + ], + "credit_map": [ + 0.01634899340569973, + 0.02527068741619587, + 0.02319525182247162, + 0.024588899686932564, + 0.020207656547427177, + 0.021179359406232834, + 0.02349206805229187, + 0.01638726145029068, + 0.014966769143939018, + 0.0225662998855114, + 0.013382426463067532, + 0.014658940024673939, + 0.013798409141600132, + 0.022460483014583588, + 0.02094678208231926, + 0.015260137617588043, + 0.012912235222756863, + 0.010977665893733501, + 0.016739867627620697, + 0.009945190511643887, + 0.019434180110692978, + 0.016802199184894562, + 0.012425899505615234, + 0.011317471973598003, + 0.006608779076486826, + 0.006430402398109436, + 0.008052465505897999, + 0.0073794168420135975, + 0.005535920150578022, + 0.006734873633831739, + 0.003961259499192238, + 0.010141368955373764, + 0.006421664264053106, + 0.006021345034241676, + 0.010734952986240387, + 0.00837779138237238, + 0.003690782468765974, + 0.010044503957033157, + 0.007278600241988897, + 0.006503157317638397, + 0.006048977840691805, + 0.009376638568937778, + 0.010535134002566338, + 0.00809702929109335, + 0.01199113205075264, + 0.0065181925892829895, + 0.009890326298773289, + 0.005543762352317572, + 0.011559909209609032, + 0.012454329058527946, + 0.011980322189629078, + 0.01305396854877472, + 0.00882644858211279, + 0.013467025011777878, + 0.01125666406005621, + 0.01673838682472706, + 0.008574838750064373, + 0.007627181243151426, + 0.008623098954558372, + 0.009342319332063198, + 0.008906587027013302, + 0.010605111718177795, + 0.008607497438788414, + 0.011019106023013592, + 0.011302415281534195, + 0.03209944814443588, + 0.008443666622042656, + 0.007919250056147575, + 0.025992805138230324, + 0.01780114136636257, + 0.02567451447248459, + 0.012017670087516308, + 0.015129775740206242, + 0.0068209366872906685, + 0.02349845878779888, + 0.009219931438565254, + 0.00955596286803484, + 0.004296184051781893, + 0.007017060182988644, + 0.009384340606629848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.11097472906112671 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003537565469742 + ], + "opposite_logprob_tokens": [ + -2.695035457611084 + ], + "logprob_reward": -0.07003537565469742, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003537565469742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09424331039190292, + 0.07340749353170395, + 0.24388927221298218, + 0.07801471650600433, + 0.026317840442061424, + 0.04003911092877388, + 0.06175261363387108, + 0.0740031749010086, + 0.2512149512767792, + 0.057117439806461334 + ], + "credit_map": [ + 0.008211118169128895, + 0.010584641247987747, + 0.0094340480864048, + 0.0126118753105402, + 0.01351204700767994, + 0.014927621930837631, + 0.013978821225464344, + 0.010983148589730263, + 0.007699852809309959, + 0.008798157796263695, + 0.009625115431845188, + 0.010134984739124775, + 0.008624464273452759, + 0.0132528031244874, + 0.00740235298871994, + 0.00786976981908083, + 0.01823608949780464, + 0.017466794699430466, + 0.04414169490337372, + 0.02159079909324646, + 0.048869986087083817, + 0.04032203555107117, + 0.033753253519535065, + 0.019508643075823784, + 0.0065412502735853195, + 0.004066627938300371, + 0.02028052508831024, + 0.007059965282678604, + 0.004975527990609407, + 0.008526775985956192, + 0.006167232990264893, + 0.020396815612912178, + 0.0033362226095050573, + 0.002748535480350256, + 0.0034452001564204693, + 0.003452678443863988, + 0.002136768540367484, + 0.0024311954621225595, + 0.005181436892598867, + 0.003585806582123041, + 0.004792059771716595, + 0.0071006352081894875, + 0.005501972045749426, + 0.003256266238167882, + 0.004367959685623646, + 0.0061410353519022465, + 0.003863109741359949, + 0.005016078241169453, + 0.006110745947808027, + 0.007224482949823141, + 0.009512466378509998, + 0.007367286365479231, + 0.006800463888794184, + 0.00905632134526968, + 0.006043038330972195, + 0.00963781401515007, + 0.009954039938747883, + 0.008558128960430622, + 0.007152594160288572, + 0.00902901217341423, + 0.009584023617208004, + 0.009054555557668209, + 0.009494468569755554, + 0.01117635890841484, + 0.01886054128408432, + 0.05935158580541611, + 0.010945932008326054, + 0.01592867076396942, + 0.03466407209634781, + 0.017352836206555367, + 0.07045374065637589, + 0.023657599464058876, + 0.006114591844379902, + 0.006280316039919853, + 0.00521102687343955, + 0.006289600394666195, + 0.009267269633710384, + 0.004997981246560812, + 0.006730147171765566, + 0.012226511724293232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.07003537565469742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033721499145030975 + ], + "opposite_logprob_tokens": [ + -3.408721446990967 + ], + "logprob_reward": -0.033721499145030975, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033721499145030975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09998514503240585, + 0.10220499336719513, + 0.11843950301408768, + 0.03821195662021637, + 0.03655998408794403, + 0.03554920479655266, + 0.07721785455942154, + 0.052986204624176025, + 0.1576378345489502, + 0.28120723366737366 + ], + "credit_map": [ + 0.00878539215773344, + 0.014856071211397648, + 0.010797246359288692, + 0.012437615543603897, + 0.013676092959940434, + 0.01745169796049595, + 0.012344746850430965, + 0.009636280126869678, + 0.013051596470177174, + 0.02010485716164112, + 0.008361813612282276, + 0.012846753001213074, + 0.013754412531852722, + 0.014500795863568783, + 0.010902611538767815, + 0.008682156912982464, + 0.010908659547567368, + 0.014059028588235378, + 0.020223280414938927, + 0.012656421400606632, + 0.014585213735699654, + 0.014064389280974865, + 0.0141915800049901, + 0.0177509356290102, + 0.005569871515035629, + 0.004112488590180874, + 0.005911414045840502, + 0.005242527928203344, + 0.0037555827293545008, + 0.005401572212576866, + 0.0036745360121130943, + 0.004543965216726065, + 0.0035928625147789717, + 0.002922892803326249, + 0.005503689404577017, + 0.005256446078419685, + 0.002680092118680477, + 0.005471085663884878, + 0.005393358878791332, + 0.005739555228501558, + 0.003309903433546424, + 0.0041985162533819675, + 0.004876696038991213, + 0.0047500235959887505, + 0.005276757292449474, + 0.004474676679819822, + 0.004910414107143879, + 0.0037522190250456333, + 0.007310805376619101, + 0.010710961185395718, + 0.01364816166460514, + 0.01075107604265213, + 0.008333987556397915, + 0.007144898176193237, + 0.009878779761493206, + 0.009439188055694103, + 0.007361050695180893, + 0.006312267854809761, + 0.005435905419290066, + 0.004871903918683529, + 0.00528495479375124, + 0.009282390587031841, + 0.00648464635014534, + 0.007953078486025333, + 0.012575095519423485, + 0.03336658701300621, + 0.011696778237819672, + 0.012803339399397373, + 0.018961068242788315, + 0.011049398221075535, + 0.042240243405103683, + 0.014945310540497303, + 0.03184782341122627, + 0.00879654474556446, + 0.10383429378271103, + 0.023230932652950287, + 0.013781983405351639, + 0.012012509629130363, + 0.012534796260297298, + 0.07516835629940033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.033721499145030975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818566583096981 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818566583096981, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818566583096981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06328509002923965, + 0.10659465193748474, + 0.049173831939697266, + 0.06903574615716934, + 0.045133717358112335, + 0.09823508560657501, + 0.07116616517305374, + 0.03370177745819092, + 0.257946252822876, + 0.20572759211063385 + ], + "credit_map": [ + 0.005758875980973244, + 0.006605023052543402, + 0.008914255537092686, + 0.009931636974215508, + 0.010074554942548275, + 0.008378708735108376, + 0.00800869707018137, + 0.005613339599221945, + 0.01002607960253954, + 0.014886293560266495, + 0.011937553994357586, + 0.012694642879068851, + 0.0139317000284791, + 0.014855348505079746, + 0.015730001032352448, + 0.012533050030469894, + 0.006078704725950956, + 0.006644195411354303, + 0.0071006412617862225, + 0.0063575576059520245, + 0.0058167134411633015, + 0.006537550128996372, + 0.005652553867548704, + 0.004985919687896967, + 0.010120225138962269, + 0.010163775645196438, + 0.01046838704496622, + 0.009712887927889824, + 0.006727836560457945, + 0.007999214343726635, + 0.004788210149854422, + 0.00905520748347044, + 0.010791710577905178, + 0.0052977055311203, + 0.005921869073063135, + 0.004956578835844994, + 0.002894681878387928, + 0.005054523702710867, + 0.004819371737539768, + 0.005397276021540165, + 0.013188258744776249, + 0.022212890908122063, + 0.01425371877849102, + 0.01289130188524723, + 0.00889644492417574, + 0.008450905792415142, + 0.010285272262990475, + 0.008056297898292542, + 0.009578327648341656, + 0.008279064670205116, + 0.008460545912384987, + 0.010213649831712246, + 0.006905626505613327, + 0.009354370646178722, + 0.007496185600757599, + 0.010878403671085835, + 0.004028485156595707, + 0.006295827683061361, + 0.005289973225444555, + 0.003657008521258831, + 0.00438090693205595, + 0.003696298226714134, + 0.003247839631512761, + 0.0031054404098540545, + 0.02269596792757511, + 0.075034499168396, + 0.012615259736776352, + 0.011101244017481804, + 0.018205467611551285, + 0.019537383690476418, + 0.07547963410615921, + 0.02327682077884674, + 0.016560465097427368, + 0.009095103479921818, + 0.06340523809194565, + 0.017967773601412773, + 0.00880217831581831, + 0.010326437652111053, + 0.010961391031742096, + 0.06860902905464172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.01818566583096981 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and executing a dynamic tracking shot following a man sprinting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark charcoal gray suit that is completely soaked and clinging to his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man's white shirt ripped at the collar with a crimson silk tie hanging loosely and swinging?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a wide asphalt road reflecting iridescent neon lights in shades of pink, blue, and red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the man's black Oxfords shown slamming into puddles and creating shimmering, neon-lit water arcs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327980473637581 + ], + "opposite_logprob_tokens": [ + -3.773279905319214 + ], + "logprob_reward": -0.02327980473637581, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327980473637581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11506771296262741, + 0.10464507341384888, + 0.10423588752746582, + 0.05034005269408226, + 0.025403674691915512, + 0.06020783632993698, + 0.08750211447477341, + 0.06040133908390999, + 0.06350136548280716, + 0.3286949396133423 + ], + "credit_map": [ + 0.010090458206832409, + 0.0131323067471385, + 0.011378131806850433, + 0.014995523728430271, + 0.015220217406749725, + 0.01711190678179264, + 0.017485562711954117, + 0.01565360836684704, + 0.01112357433885336, + 0.01837957836687565, + 0.01056336984038353, + 0.011822318658232689, + 0.014855094254016876, + 0.016589166596531868, + 0.011854438111186028, + 0.00945754162967205, + 0.01149689219892025, + 0.015627048909664154, + 0.015581074170768261, + 0.009604987688362598, + 0.01649455912411213, + 0.01339492853730917, + 0.011822187341749668, + 0.010214208625257015, + 0.0055718631483614445, + 0.006833153311163187, + 0.0067399246618151665, + 0.006879257038235664, + 0.005420805886387825, + 0.006715320982038975, + 0.0035991244949400425, + 0.008580602705478668, + 0.002612598007544875, + 0.0024066385813057423, + 0.004148504696786404, + 0.004208879079669714, + 0.0018493920797482133, + 0.002893528202548623, + 0.0031631675083190203, + 0.004120967350900173, + 0.006514364387840033, + 0.01228085346519947, + 0.008564000017940998, + 0.007781550753861666, + 0.005973560269922018, + 0.006328609772026539, + 0.007711574900895357, + 0.005053326021879911, + 0.010921204462647438, + 0.011005599051713943, + 0.010328270494937897, + 0.013336015865206718, + 0.007288181688636541, + 0.014858442358672619, + 0.010982112027704716, + 0.008782287128269672, + 0.006718612741678953, + 0.007952318526804447, + 0.006610000040382147, + 0.006754731293767691, + 0.006519103888422251, + 0.008882641792297363, + 0.007516875397413969, + 0.009447054006159306, + 0.005245424807071686, + 0.00943976454436779, + 0.005230700131505728, + 0.0046296617947518826, + 0.016752280294895172, + 0.009832162410020828, + 0.007102204952389002, + 0.005269163753837347, + 0.036993127316236496, + 0.010006705299019814, + 0.12336359918117523, + 0.027335243299603462, + 0.014434557408094406, + 0.014221676625311375, + 0.013460451737046242, + 0.08887957036495209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.02327980473637581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_8/index_tensor([6926, 6926])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_8/index_tensor([6926, 6926])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..69360eed54274c7c80c9c1d7c20238be27e99f77 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_8/index_tensor([6926, 6926])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, top-down macro shot captures a pristine, off-white sheet of densely textured art paper, its fibrous surface illuminated by soft, directed studio light, casting subtle shadows across its topography. At the center of the frame rests a perfectly placed drop of viscous, inky black ink, its surface shimmering with a wet, smooth sheen. In mesmerizing super-slow motion, an unseen force begins to pull the drop horizontally, causing it to gracefully stretch and elongate across the paper. The body of ink thins, its cohesiveness beginning to break down, splintering into a network of delicate, branching threads. These delicate, syrupy threads hypnotically extend outward, branching into ever more intricate, fragmented patterns resembling tree roots, creating a stark and beautiful contrast against the warm texture of the paper. The deliberate, silent unfolding of the ink creates an atmosphere of concentration, contemplation, and intense artistic intensity, capturing the moment of abstract creation.", + "seed": 9445, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_8/index_6926_seed_9445_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03389081070199609, + "vlm_logprob": -0.03389081070199609, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03389081070199609, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018447145819664 + ], + "opposite_logprob_tokens": [ + -4.768446922302246 + ], + "logprob_reward": -0.018447145819664, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.018447145819664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2414717823266983, + 0.09725137054920197, + 0.039287298917770386, + 0.10613157600164413, + 0.04960020259022713, + 0.06472987681627274, + 0.06889704614877701, + 0.055717259645462036, + 0.1437220275402069, + 0.1331915259361267 + ], + "credit_map": [ + 0.03636917099356651, + 0.04025657847523689, + 0.0288220401853323, + 0.02696140669286251, + 0.01466837152838707, + 0.022676648572087288, + 0.022583063691854477, + 0.04913453385233879, + 0.02089586667716503, + 0.020491063594818115, + 0.014661060646176338, + 0.00839920248836279, + 0.00518686743453145, + 0.006105673499405384, + 0.006485098972916603, + 0.015026540495455265, + 0.007481234148144722, + 0.00797476340085268, + 0.00703023374080658, + 0.0031018508598208427, + 0.002748554339632392, + 0.002591686090454459, + 0.002541643800213933, + 0.005817336030304432, + 0.020384790375828743, + 0.021084297448396683, + 0.015283898450434208, + 0.007642838172614574, + 0.007832113653421402, + 0.007993699982762337, + 0.007091888692229986, + 0.018818054348230362, + 0.008450205437839031, + 0.0071692196652293205, + 0.006160400342196226, + 0.004843696486204863, + 0.004409974906593561, + 0.0061063580214977264, + 0.005105380434542894, + 0.007354965433478355, + 0.011262059211730957, + 0.008096051402390003, + 0.00947833713144064, + 0.006446496117860079, + 0.007210581563413143, + 0.007776199374347925, + 0.005828045774251223, + 0.008632106706500053, + 0.011364845559000969, + 0.009083911776542664, + 0.009841137565672398, + 0.008158810436725616, + 0.004978714510798454, + 0.007990529760718346, + 0.004943724721670151, + 0.01253537181764841, + 0.009270131587982178, + 0.009839828126132488, + 0.008515292778611183, + 0.0062040649354457855, + 0.004940113052725792, + 0.005306922364979982, + 0.005436575971543789, + 0.006204329896718264, + 0.029577387496829033, + 0.027094192802906036, + 0.021871209144592285, + 0.019354749470949173, + 0.008181044831871986, + 0.011013919487595558, + 0.011302374303340912, + 0.01532715279608965, + 0.02234465442597866, + 0.014635566622018814, + 0.017891382798552513, + 0.013772428035736084, + 0.011253776960074902, + 0.019677618518471718, + 0.012751910835504532, + 0.020864197984337807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.018447145819664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308703541755676 + ], + "opposite_logprob_tokens": [ + -3.1680870056152344 + ], + "logprob_reward": -0.04308703541755676, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308703541755676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.303891122341156, + 0.12996216118335724, + 0.09959263354539871, + 0.07488364726305008, + 0.11202506721019745, + 0.06975117325782776, + 0.0564604252576828, + 0.06773271411657333, + 0.050892218947410583, + 0.03480884060263634 + ], + "credit_map": [ + 0.053214944899082184, + 0.058919571340084076, + 0.04014522209763527, + 0.02999037131667137, + 0.01583402417600155, + 0.0193660669028759, + 0.024991437792778015, + 0.06142948940396309, + 0.022989077493548393, + 0.025417445227503777, + 0.01631231978535652, + 0.015256374143064022, + 0.008499277755618095, + 0.00783897191286087, + 0.010285312309861183, + 0.023363390937447548, + 0.015576813369989395, + 0.026553187519311905, + 0.01191262062638998, + 0.009016018360853195, + 0.006130765657871962, + 0.004603417124599218, + 0.0060153622180223465, + 0.019784454256296158, + 0.016748478636145592, + 0.01814737357199192, + 0.009954242967069149, + 0.0053750392980873585, + 0.004996409174054861, + 0.005334808491170406, + 0.0038071086164563894, + 0.010520191863179207, + 0.023330572992563248, + 0.016479041427373886, + 0.015613517723977566, + 0.010324171744287014, + 0.00868489220738411, + 0.010779420845210552, + 0.008355476893484592, + 0.01845797523856163, + 0.011483446694910526, + 0.007158992812037468, + 0.00764413271099329, + 0.010102313943207264, + 0.005852696485817432, + 0.011345475912094116, + 0.006285143084824085, + 0.009878971613943577, + 0.012865060940384865, + 0.009892689995467663, + 0.008646568283438683, + 0.005663246847689152, + 0.003409851808100939, + 0.004181868862360716, + 0.003581162542104721, + 0.008219978772103786, + 0.015132882632315159, + 0.012549263425171375, + 0.011673525907099247, + 0.006432234309613705, + 0.00406201696023345, + 0.00550987059250474, + 0.004645658656954765, + 0.007727261632680893, + 0.009213960729539394, + 0.007517477963119745, + 0.006824063137173653, + 0.0070994240231812, + 0.002963507315143943, + 0.00505084078758955, + 0.00433909660205245, + 0.007883851416409016, + 0.003955724649131298, + 0.0040285661816596985, + 0.004823376890271902, + 0.003594372421503067, + 0.003457198617979884, + 0.00639085378497839, + 0.003066036384552717, + 0.0054927123710513115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04308703541755676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907261490821838 + ], + "opposite_logprob_tokens": [ + -2.4640727043151855 + ], + "logprob_reward": -0.08907261490821838, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907261490821838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14577406644821167, + 0.048479363322257996, + 0.10740530490875244, + 0.08871982246637344, + 0.136260986328125, + 0.05207507312297821, + 0.09563656151294708, + 0.1336897760629654, + 0.12629356980323792, + 0.06566543877124786 + ], + "credit_map": [ + 0.018076561391353607, + 0.022875219583511353, + 0.02078595943748951, + 0.01976037211716175, + 0.011773517355322838, + 0.014083348214626312, + 0.01340491883456707, + 0.025014203041791916, + 0.00818124320358038, + 0.00895867682993412, + 0.005484626162797213, + 0.00513543002307415, + 0.004020709544420242, + 0.004521016031503677, + 0.0038994760252535343, + 0.008278191089630127, + 0.014454560354351997, + 0.018829496577382088, + 0.016805164515972137, + 0.01132114790380001, + 0.008268032222986221, + 0.00969206728041172, + 0.007547946646809578, + 0.020486902445554733, + 0.014638576656579971, + 0.015128979459404945, + 0.013389439322054386, + 0.0071298531256616116, + 0.007480158004909754, + 0.0075907898135483265, + 0.005120644811540842, + 0.01824139431118965, + 0.021705076098442078, + 0.01794508844614029, + 0.015598234720528126, + 0.015813538804650307, + 0.016170762479305267, + 0.013680079020559788, + 0.014892601408064365, + 0.020455610007047653, + 0.0077753146179020405, + 0.00627897260710597, + 0.006019892171025276, + 0.00759479496628046, + 0.0053609213791787624, + 0.005944916512817144, + 0.006933154072612524, + 0.006167114246636629, + 0.016277914866805077, + 0.012442291714251041, + 0.013131207786500454, + 0.010614543221890926, + 0.00844758190214634, + 0.014074159786105156, + 0.009294976480305195, + 0.011353894136846066, + 0.020652730017900467, + 0.016795746982097626, + 0.01645982451736927, + 0.016919514164328575, + 0.012589134275913239, + 0.017469516023993492, + 0.016495302319526672, + 0.016308028250932693, + 0.016501573845744133, + 0.013938124291598797, + 0.015783866867423058, + 0.015425186604261398, + 0.012778835371136665, + 0.019686678424477577, + 0.013447796925902367, + 0.018731527030467987, + 0.010229925625026226, + 0.010559901595115662, + 0.00909237377345562, + 0.00684916228055954, + 0.004267998039722443, + 0.009717708453536034, + 0.005769160110503435, + 0.009179213084280491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.08907261490821838 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643965046852827 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.004643965046852827, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643965046852827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10808194428682327, + 0.029003160074353218, + 0.02627996727824211, + 0.0278993621468544, + 0.03604227676987648, + 0.043242283165454865, + 0.02576814778149128, + 0.07445114105939865, + 0.055950772017240524, + 0.5732809901237488 + ], + "credit_map": [ + 0.015047061257064342, + 0.016504937782883644, + 0.015609845519065857, + 0.011912432499229908, + 0.007081090472638607, + 0.011517300270497799, + 0.010691125877201557, + 0.019718149676918983, + 0.0034894533455371857, + 0.004997011739760637, + 0.0034003430046141148, + 0.002968364395201206, + 0.0026058207731693983, + 0.0035558564122766256, + 0.0029901331290602684, + 0.004996177740395069, + 0.0043515232391655445, + 0.0043404363095760345, + 0.0035618250258266926, + 0.0023838419001549482, + 0.0023517035879194736, + 0.0025539519265294075, + 0.0024828952737152576, + 0.004253789782524109, + 0.0037727481685578823, + 0.0037781810387969017, + 0.004122606478631496, + 0.0029616812244057655, + 0.0026168604381382465, + 0.0029615662060678005, + 0.0028746232856065035, + 0.004811094608157873, + 0.005301665514707565, + 0.004414179362356663, + 0.004050174728035927, + 0.003883578348904848, + 0.003764407942071557, + 0.004519949667155743, + 0.0032552131451666355, + 0.006853108759969473, + 0.00614240812137723, + 0.00540396012365818, + 0.005821951199322939, + 0.007167333271354437, + 0.004942779894918203, + 0.004062137566506863, + 0.003929260186851025, + 0.005772450938820839, + 0.004159816075116396, + 0.003094704821705818, + 0.003480448853224516, + 0.0022726356983184814, + 0.0015076217241585255, + 0.004524091258645058, + 0.0020516400691121817, + 0.004677188582718372, + 0.0105296541005373, + 0.009305416606366634, + 0.010371481068432331, + 0.010065149515867233, + 0.006835657171905041, + 0.010360697284340858, + 0.007572732400149107, + 0.009410353377461433, + 0.008529936894774437, + 0.00782528892159462, + 0.007859003730118275, + 0.006776094902306795, + 0.005441788583993912, + 0.0064270831644535065, + 0.005705741234123707, + 0.007385834585875273, + 0.08604229986667633, + 0.07407514750957489, + 0.07637657225131989, + 0.06767209619283676, + 0.050263095647096634, + 0.06820416450500488, + 0.05525568500161171, + 0.0953918993473053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.004643965046852827 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203292317688465 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203292317688465, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203292317688465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19853658974170685, + 0.056121718138456345, + 0.038744889199733734, + 0.11338293552398682, + 0.11336049437522888, + 0.08866249769926071, + 0.09831877052783966, + 0.048264291137456894, + 0.0783059373497963, + 0.166301891207695 + ], + "credit_map": [ + 0.038230545818805695, + 0.03355026990175247, + 0.02683764509856701, + 0.016097528859972954, + 0.013421903364360332, + 0.017640559002757072, + 0.015110122039914131, + 0.03764801472425461, + 0.009582174010574818, + 0.007880402728915215, + 0.007296425756067038, + 0.007895545102655888, + 0.005364674609154463, + 0.004717051982879639, + 0.0068283225409686565, + 0.006557117216289043, + 0.005879421252757311, + 0.007258806377649307, + 0.005106843542307615, + 0.0036893279757350683, + 0.0028493127319961786, + 0.0031811343505978584, + 0.003443600842729211, + 0.007336442358791828, + 0.016679221764206886, + 0.015200520865619183, + 0.013625407591462135, + 0.011957795359194279, + 0.013116428628563881, + 0.013515105471014977, + 0.013886328786611557, + 0.015402128919959068, + 0.016049014404416084, + 0.01495568547397852, + 0.014200331643223763, + 0.014028605073690414, + 0.011682533659040928, + 0.014512186869978905, + 0.013321234844624996, + 0.01461090799421072, + 0.017086001113057137, + 0.013251802884042263, + 0.010283288545906544, + 0.008759737014770508, + 0.010987975634634495, + 0.01102499570697546, + 0.0070347473956644535, + 0.010233938694000244, + 0.015933597460389137, + 0.0116584412753582, + 0.014311485923826694, + 0.011083326302468777, + 0.00907355546951294, + 0.01188773475587368, + 0.01150511484593153, + 0.012865505181252956, + 0.007089532446116209, + 0.005922005511820316, + 0.005810549482703209, + 0.005740462802350521, + 0.005125959403812885, + 0.006335089448839426, + 0.0046931300312280655, + 0.007547562941908836, + 0.011489521712064743, + 0.01200574729591608, + 0.009312565438449383, + 0.009187002666294575, + 0.007556804455816746, + 0.009451869875192642, + 0.008458467200398445, + 0.010843959636986256, + 0.02025659754872322, + 0.017679180949926376, + 0.022485466673970222, + 0.019754720851778984, + 0.01578342355787754, + 0.02586689405143261, + 0.01694328524172306, + 0.027532311156392097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.014203292317688465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down macro shot captures a pristine, off-white sheet of densely textured art paper, its fibrous surface illuminated by soft, directed studio light, casting subtle shadows across its topography. At the center of the frame rests a perfectly placed drop of viscous, inky black ink, its surface shimmering with a wet, smooth sheen. In mesmerizing super-slow motion, an unseen force begins to pull the drop horizontally, causing it to gracefully stretch and elongate across the paper. The body of ink thins, its cohesiveness beginning to break down, splintering into a network of delicate, branching threads. These delicate, syrupy threads hypnotically extend outward, branching into ever more intricate, fragmented patterns resembling tree roots, creating a stark and beautiful contrast against the warm texture of the paper. The deliberate, silent unfolding of the ink creates an atmosphere of concentration, contemplation, and intense artistic intensity, capturing the moment of abstract creation.", + "seed": 9446, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_8/index_6926_seed_9446_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024460167321376503, + "vlm_logprob": -0.024460167321376503, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.024460167321376503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025282297283411026 + ], + "opposite_logprob_tokens": [ + -4.275282382965088 + ], + "logprob_reward": -0.025282297283411026, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.025282297283411026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5607901215553284, + 0.11443533003330231, + 0.04234699159860611, + 0.03206420689821243, + 0.04169603809714317, + 0.030034303665161133, + 0.049183931201696396, + 0.03876911848783493, + 0.052546046674251556, + 0.038133908063173294 + ], + "credit_map": [ + 0.10284406691789627, + 0.03404379263520241, + 0.0768919438123703, + 0.0359872542321682, + 0.06555843353271484, + 0.10953224450349808, + 0.0463675931096077, + 0.08956478536128998, + 0.025472836568951607, + 0.006755894515663385, + 0.018500154837965965, + 0.007154756225645542, + 0.008808858692646027, + 0.014759586192667484, + 0.009164636954665184, + 0.023818600922822952, + 0.008903385140001774, + 0.0026082205586135387, + 0.008006376214325428, + 0.003303634002804756, + 0.004597189836204052, + 0.004670289810746908, + 0.0031745333690196276, + 0.0070833610370755196, + 0.0066156527027487755, + 0.0018380810506641865, + 0.0060912566259503365, + 0.0030970454681664705, + 0.0025253496132791042, + 0.0027158174198120832, + 0.0031665016431361437, + 0.0060144998133182526, + 0.00814511813223362, + 0.0024883593432605267, + 0.006679947022348642, + 0.0036373550537973642, + 0.0031860526651144028, + 0.006661799270659685, + 0.0029285894706845284, + 0.007968815974891186, + 0.004908799659460783, + 0.0018977290019392967, + 0.004226997960358858, + 0.002688596723601222, + 0.002857911167666316, + 0.004729860927909613, + 0.002653009258210659, + 0.0060713994316756725, + 0.009369495324790478, + 0.0024384623393416405, + 0.008119218051433563, + 0.003745190566405654, + 0.0037176525220274925, + 0.005646750330924988, + 0.004428591579198837, + 0.0117185702547431, + 0.00789463147521019, + 0.0026245703920722008, + 0.006195882335305214, + 0.0028386758640408516, + 0.004659188445657492, + 0.003183437278494239, + 0.005054440815001726, + 0.006318293511867523, + 0.011210297234356403, + 0.003349058562889695, + 0.009909418411552906, + 0.003061816329136491, + 0.004197548143565655, + 0.0047380863688886166, + 0.004928511567413807, + 0.011151308193802834, + 0.005599298980087042, + 0.0021682633087038994, + 0.007080669514834881, + 0.0026677497662603855, + 0.003565288847312331, + 0.004580334760248661, + 0.0038608477916568518, + 0.00861145555973053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.025282297283411026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208839267492294 + ], + "opposite_logprob_tokens": [ + -2.8120884895324707 + ], + "logprob_reward": -0.06208839267492294, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208839267492294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2498907446861267, + 0.13192300498485565, + 0.1329503208398819, + 0.10779808461666107, + 0.1976855993270874, + 0.03483370319008827, + 0.06335610896348953, + 0.01509570050984621, + 0.021088825538754463, + 0.0453779511153698 + ], + "credit_map": [ + 0.042691897600889206, + 0.015048423781991005, + 0.04315352812409401, + 0.02690676599740982, + 0.020626163110136986, + 0.03523124009370804, + 0.02004285529255867, + 0.04618984833359718, + 0.034424684941768646, + 0.007112881634384394, + 0.026389773935079575, + 0.008056449703872204, + 0.0063475631177425385, + 0.012526591308414936, + 0.00855126604437828, + 0.028513776138424873, + 0.031012116000056267, + 0.006756402086466551, + 0.030217407271265984, + 0.009648848325014114, + 0.010051635093986988, + 0.01074643898755312, + 0.00897189136594534, + 0.025545572862029076, + 0.03065686672925949, + 0.0034693309571594, + 0.027329472824931145, + 0.006393544375896454, + 0.007092994172126055, + 0.009416352026164532, + 0.005901531782001257, + 0.017537984997034073, + 0.06084507331252098, + 0.00914874579757452, + 0.04353969916701317, + 0.007918143644928932, + 0.008544878102838993, + 0.013142571784555912, + 0.008933124132454395, + 0.04561334475874901, + 0.009514618664979935, + 0.0018933034734800458, + 0.005394200794398785, + 0.001985869836062193, + 0.002439401112496853, + 0.004085575696080923, + 0.0019314043456688523, + 0.0075893281027674675, + 0.013816695660352707, + 0.002396524650976062, + 0.012411767616868019, + 0.003144944319501519, + 0.0029513640329241753, + 0.005814712028950453, + 0.005235325079411268, + 0.017584770917892456, + 0.002789028687402606, + 0.0008501382544636726, + 0.002712869318202138, + 0.0009622423094697297, + 0.0017054863274097443, + 0.002027388196438551, + 0.001329866237938404, + 0.002718680538237095, + 0.004168690647929907, + 0.0009087466169148684, + 0.0037455251440405846, + 0.0015253485180437565, + 0.0018225369276478887, + 0.00204002158716321, + 0.0022589447908103466, + 0.004619009792804718, + 0.007305300794541836, + 0.003748230868950486, + 0.007774266880005598, + 0.004138756077736616, + 0.006271675694733858, + 0.0034619667567312717, + 0.00400393083691597, + 0.008673817850649357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.06208839267492294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018211184069514275 + ], + "opposite_logprob_tokens": [ + -4.018211364746094 + ], + "logprob_reward": -0.018211184069514275, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018211184069514275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2926930785179138, + 0.09382893890142441, + 0.035745494067668915, + 0.056845199316740036, + 0.08078885078430176, + 0.06068888679146767, + 0.090715691447258, + 0.1077248677611351, + 0.16253963112831116, + 0.01842927560210228 + ], + "credit_map": [ + 0.04942655935883522, + 0.01607280783355236, + 0.040693674236536026, + 0.02199617028236389, + 0.03440495580434799, + 0.0566251315176487, + 0.026877738535404205, + 0.04659607261419296, + 0.0141779575496912, + 0.007718943990767002, + 0.014790781773626804, + 0.009804791770875454, + 0.00843430869281292, + 0.011482605710625648, + 0.008513758890330791, + 0.018905799835920334, + 0.00679789949208498, + 0.00277433218434453, + 0.006421283818781376, + 0.00277100526727736, + 0.004290905315428972, + 0.003406075993552804, + 0.002906791865825653, + 0.0063772052526474, + 0.00835488922894001, + 0.003941644914448261, + 0.008355959318578243, + 0.005473716650158167, + 0.0063432324677705765, + 0.008166099898517132, + 0.004940050188452005, + 0.011269613169133663, + 0.01518680527806282, + 0.004543709568679333, + 0.011397035792469978, + 0.006975519470870495, + 0.007698632776737213, + 0.012593535706400871, + 0.006837661378085613, + 0.015555962920188904, + 0.012244127690792084, + 0.004121275153011084, + 0.009479423984885216, + 0.003671061946079135, + 0.0051961359567940235, + 0.007202496752142906, + 0.005497244652360678, + 0.013277128338813782, + 0.015341228805482388, + 0.003546137595549226, + 0.017236705869436264, + 0.004617827944457531, + 0.005610018502920866, + 0.009694201871752739, + 0.007476325612515211, + 0.027193253859877586, + 0.0184022244066, + 0.006632951553910971, + 0.021090058609843254, + 0.006559069734066725, + 0.011639266274869442, + 0.00976531021296978, + 0.011128447018563747, + 0.022507546469569206, + 0.027834534645080566, + 0.008453909307718277, + 0.03608913719654083, + 0.009489726275205612, + 0.014780435711145401, + 0.014645595103502274, + 0.016819698736071587, + 0.03442661464214325, + 0.0031863118056207895, + 0.0012295551132410765, + 0.003363292198628187, + 0.0011416800552979112, + 0.0022014013957232237, + 0.0018646091921254992, + 0.0022513498552143574, + 0.003191078547388315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.018211184069514275 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006380429258570075 + ], + "opposite_logprob_tokens": [ + -7.375638008117676 + ], + "logprob_reward": -0.0006380429258570075, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006380429258570075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14632976055145264, + 0.08296483755111694, + 0.05607079714536667, + 0.047212276607751846, + 0.09171418845653534, + 0.06234356388449669, + 0.10906577110290527, + 0.15795549750328064, + 0.18338753283023834, + 0.06295574456453323 + ], + "credit_map": [ + 0.024198533967137337, + 0.009949627332389355, + 0.023000653833150864, + 0.01217845268547535, + 0.011682551354169846, + 0.018652578815817833, + 0.012332388199865818, + 0.03433498740196228, + 0.014617441222071648, + 0.00555441016331315, + 0.011974761262536049, + 0.0062212515622377396, + 0.009361320175230503, + 0.011724244803190231, + 0.009143288247287273, + 0.014368118718266487, + 0.011580304242670536, + 0.004159285221248865, + 0.00989698525518179, + 0.004599810112267733, + 0.007525771856307983, + 0.005239668767899275, + 0.004570699762552977, + 0.008498270995914936, + 0.011721865274012089, + 0.002748926170170307, + 0.008542622439563274, + 0.0034365570172667503, + 0.00352757191285491, + 0.0038541413377970457, + 0.004824311938136816, + 0.008556279353797436, + 0.01783650740981102, + 0.005842640995979309, + 0.011896947398781776, + 0.00766358058899641, + 0.007611318025738001, + 0.014594684354960918, + 0.00888196099549532, + 0.017386551946401596, + 0.01025801058858633, + 0.003722844645380974, + 0.00925802905112505, + 0.005032367072999477, + 0.006275249645113945, + 0.008190135471522808, + 0.007207932416349649, + 0.012398995459079742, + 0.02490517683327198, + 0.005634988658130169, + 0.018234549090266228, + 0.006983248516917229, + 0.007620649412274361, + 0.013734892942011356, + 0.012738141231238842, + 0.019214129075407982, + 0.02950098179280758, + 0.012629486620426178, + 0.025197019800543785, + 0.011217674240469933, + 0.016607830300927162, + 0.016359204426407814, + 0.018467828631401062, + 0.027975475415587425, + 0.03603305295109749, + 0.011141197755932808, + 0.036766279488801956, + 0.011434941552579403, + 0.015161635354161263, + 0.016673749312758446, + 0.018207760527729988, + 0.03796892240643501, + 0.00998481921851635, + 0.0038660643622279167, + 0.012023275718092918, + 0.005823907908052206, + 0.006585896480828524, + 0.005211823154240847, + 0.0052837408147752285, + 0.01417621597647667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0006380429258570075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608091965317726 + ], + "opposite_logprob_tokens": [ + -4.141080856323242 + ], + "logprob_reward": -0.01608091965317726, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608091965317726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25428885221481323, + 0.06112855300307274, + 0.05642220750451088, + 0.0574629008769989, + 0.06938718259334564, + 0.04528605565428734, + 0.08339297771453857, + 0.1116417646408081, + 0.21249927580356598, + 0.048490237444639206 + ], + "credit_map": [ + 0.0454266183078289, + 0.018805470317602158, + 0.03513248637318611, + 0.017936812713742256, + 0.028832269832491875, + 0.04387626424431801, + 0.021902790293097496, + 0.04237612336874008, + 0.00753347622230649, + 0.005241428967565298, + 0.009584485553205013, + 0.005242618266493082, + 0.008011309430003166, + 0.008164503611624241, + 0.006710008718073368, + 0.010640723630785942, + 0.011971059255301952, + 0.003212305484339595, + 0.010200726799666882, + 0.003922637552022934, + 0.0064215464517474174, + 0.005387252662330866, + 0.004472276195883751, + 0.010834403336048126, + 0.01288682222366333, + 0.0030267671681940556, + 0.011621289886534214, + 0.004981139209121466, + 0.00566890137270093, + 0.004236963111907244, + 0.005055947694927454, + 0.009985068812966347, + 0.01570913754403591, + 0.003544247942045331, + 0.010900771245360374, + 0.005327718798071146, + 0.006493579596281052, + 0.008167771622538567, + 0.006603581365197897, + 0.012640368193387985, + 0.00831230916082859, + 0.002697102027013898, + 0.006901010405272245, + 0.0032999145332723856, + 0.0034637085627764463, + 0.006147491745650768, + 0.0036930583883076906, + 0.010771460831165314, + 0.02182222157716751, + 0.003971438389271498, + 0.015061507932841778, + 0.005712042097002268, + 0.005258397664874792, + 0.007050731685012579, + 0.007605867460370064, + 0.016910778358578682, + 0.018993081524968147, + 0.0058952821418643, + 0.021892724558711052, + 0.008367459289729595, + 0.009325622580945492, + 0.008912206627428532, + 0.013980266638100147, + 0.024275122210383415, + 0.042772166430950165, + 0.011298811063170433, + 0.046983443200588226, + 0.012926321476697922, + 0.015998361632227898, + 0.01840256340801716, + 0.021212302148342133, + 0.0429052971303463, + 0.006842872593551874, + 0.0033313424792140722, + 0.008745442144572735, + 0.003207435132935643, + 0.0076850298792123795, + 0.005983710754662752, + 0.006192209664732218, + 0.006502190139144659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.01608091965317726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down macro shot captures a pristine, off-white sheet of densely textured art paper, its fibrous surface illuminated by soft, directed studio light, casting subtle shadows across its topography. At the center of the frame rests a perfectly placed drop of viscous, inky black ink, its surface shimmering with a wet, smooth sheen. In mesmerizing super-slow motion, an unseen force begins to pull the drop horizontally, causing it to gracefully stretch and elongate across the paper. The body of ink thins, its cohesiveness beginning to break down, splintering into a network of delicate, branching threads. These delicate, syrupy threads hypnotically extend outward, branching into ever more intricate, fragmented patterns resembling tree roots, creating a stark and beautiful contrast against the warm texture of the paper. The deliberate, silent unfolding of the ink creates an atmosphere of concentration, contemplation, and intense artistic intensity, capturing the moment of abstract creation.", + "seed": 9447, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_8/index_6926_seed_9447_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027407490322366355, + "vlm_logprob": -0.027407490322366355, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.027407490322366355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025287644937634468 + ], + "opposite_logprob_tokens": [ + -4.275287628173828 + ], + "logprob_reward": -0.025287644937634468, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.025287644937634468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17789986729621887, + 0.09910356998443604, + 0.05090296268463135, + 0.11759865283966064, + 0.15895415842533112, + 0.028159311041235924, + 0.11063586175441742, + 0.132171168923378, + 0.08296209573745728, + 0.04161239415407181 + ], + "credit_map": [ + 0.038040388375520706, + 0.017053358256816864, + 0.026894481852650642, + 0.023210646584630013, + 0.012434850446879864, + 0.008556315675377846, + 0.013025866821408272, + 0.0386839359998703, + 0.02073032222688198, + 0.009691295213997364, + 0.015503929927945137, + 0.008746414445340633, + 0.007639431860297918, + 0.006934745702892542, + 0.00959266722202301, + 0.020264752209186554, + 0.007560542784631252, + 0.004654654301702976, + 0.00775158079341054, + 0.00355925434269011, + 0.005769350565969944, + 0.010223527438938618, + 0.004388662986457348, + 0.006995385512709618, + 0.020437927916646004, + 0.014113476499915123, + 0.018482904881238937, + 0.007951796054840088, + 0.01041450072079897, + 0.012616544030606747, + 0.01408709492534399, + 0.019494397565722466, + 0.023712441325187683, + 0.011205592192709446, + 0.036906033754348755, + 0.01185877900570631, + 0.014111632481217384, + 0.013448959216475487, + 0.022523555904626846, + 0.025187144055962563, + 0.004631320480257273, + 0.0018159386236220598, + 0.003289843210950494, + 0.0017378153279423714, + 0.0021171460393816233, + 0.0054639228619635105, + 0.004244163632392883, + 0.004859160166233778, + 0.012586597353219986, + 0.009223797358572483, + 0.014869189821183681, + 0.012480078265070915, + 0.012583338655531406, + 0.015161524526774883, + 0.01843302883207798, + 0.015298299491405487, + 0.018848620355129242, + 0.011153237894177437, + 0.016254981979727745, + 0.014285133220255375, + 0.013833544217050076, + 0.022705508396029472, + 0.015476410277187824, + 0.019613711163401604, + 0.010381829924881458, + 0.005883387289941311, + 0.007816242054104805, + 0.008206471800804138, + 0.00753286387771368, + 0.019217580556869507, + 0.011791779659688473, + 0.012131926603615284, + 0.0064394609071314335, + 0.002730751410126686, + 0.003546816064044833, + 0.004386925604194403, + 0.0032082421239465475, + 0.009432774968445301, + 0.005289597436785698, + 0.006577819585800171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.025287644937634468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033730026334524155 + ], + "opposite_logprob_tokens": [ + -3.4087300300598145 + ], + "logprob_reward": -0.033730026334524155, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033730026334524155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.160347118973732, + 0.3033648133277893, + 0.34661194682121277, + 0.04103878140449524, + 0.0400492362678051, + 0.018039923161268234, + 0.02775273285806179, + 0.017366906628012657, + 0.015511779114603996, + 0.02991679310798645 + ], + "credit_map": [ + 0.024828338995575905, + 0.01924293115735054, + 0.025570830330252647, + 0.024990146979689598, + 0.015231654047966003, + 0.010081615298986435, + 0.014323730953037739, + 0.0260778721421957, + 0.07370022684335709, + 0.022400211542844772, + 0.04815293103456497, + 0.023770764470100403, + 0.032598379999399185, + 0.016940949484705925, + 0.033918850123882294, + 0.05188247188925743, + 0.07505131512880325, + 0.02669234201312065, + 0.056445762515068054, + 0.01854695752263069, + 0.03593241050839424, + 0.05205092579126358, + 0.020483706146478653, + 0.061408501118421555, + 0.012901995331048965, + 0.002396178897470236, + 0.011417281813919544, + 0.00182113959454, + 0.002773342886939645, + 0.0023114338982850313, + 0.00278963684104383, + 0.004627772141247988, + 0.007031742017716169, + 0.002352166920900345, + 0.00595858134329319, + 0.0029344556387513876, + 0.0030860663391649723, + 0.0035834130831062794, + 0.008085855282843113, + 0.007016955874860287, + 0.002829571720212698, + 0.0012657103361561894, + 0.0030632021371275187, + 0.0012422640575096011, + 0.0014686578651890159, + 0.002450584666803479, + 0.001963697373867035, + 0.0037562360521405935, + 0.0035531807225197554, + 0.0021090563386678696, + 0.002873367164283991, + 0.003356313332915306, + 0.0032660409342497587, + 0.0041482215747237206, + 0.004557983949780464, + 0.0038885686080902815, + 0.0026965662837028503, + 0.0018989980453625321, + 0.002675281837582588, + 0.0015334570780396461, + 0.001782886916771531, + 0.002025492722168565, + 0.0021261624060571194, + 0.0026280612219125032, + 0.0019184128614142537, + 0.00134666683152318, + 0.0020383568480610847, + 0.0017437860369682312, + 0.0013818265870213509, + 0.002529155695810914, + 0.0022478033788502216, + 0.0023057698272168636, + 0.004650408402085304, + 0.0022459588944911957, + 0.0035349668469280005, + 0.0029275044798851013, + 0.002516219625249505, + 0.006287925876677036, + 0.0036869735922664404, + 0.0040668342262506485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.033730026334524155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209825351834297 + ], + "opposite_logprob_tokens": [ + -2.812098264694214 + ], + "logprob_reward": -0.06209825351834297, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209825351834297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10527417808771133, + 0.07878673821687698, + 0.0967281386256218, + 0.20446118712425232, + 0.33499446511268616, + 0.025820253416895866, + 0.030387429520487785, + 0.06972270458936691, + 0.03484835475683212, + 0.018976591527462006 + ], + "credit_map": [ + 0.01326280739158392, + 0.013376706279814243, + 0.014421396888792515, + 0.013440949842333794, + 0.01563446596264839, + 0.009511005133390427, + 0.008544825948774815, + 0.01708202064037323, + 0.013685712590813637, + 0.009552784264087677, + 0.011744453571736813, + 0.007472887635231018, + 0.007101979106664658, + 0.0062617468647658825, + 0.008537949994206429, + 0.014429215341806412, + 0.01617485098540783, + 0.010981909930706024, + 0.015611335635185242, + 0.008200288750231266, + 0.013505951501429081, + 0.010962958447635174, + 0.007437524851411581, + 0.013853313401341438, + 0.04362787678837776, + 0.020365988835692406, + 0.03765067458152771, + 0.01281648501753807, + 0.014981860294938087, + 0.021364601328969002, + 0.02389310859143734, + 0.029760591685771942, + 0.05023615434765816, + 0.022417785599827766, + 0.06256994605064392, + 0.02495475858449936, + 0.0272903423756361, + 0.030634108930826187, + 0.06142144277691841, + 0.05546988174319267, + 0.0031557234469801188, + 0.00202987901866436, + 0.0040826499462127686, + 0.0018293093889951706, + 0.0029309066012501717, + 0.005267677828669548, + 0.0033388810697942972, + 0.0031852228567004204, + 0.003953362815082073, + 0.0031846475321799517, + 0.0032291782554239035, + 0.00365639291703701, + 0.0028552834410220385, + 0.0032293901313096285, + 0.005652123596519232, + 0.004627048037946224, + 0.00947503186762333, + 0.0060711451806128025, + 0.009125377051532269, + 0.007135467138141394, + 0.007075668312609196, + 0.012135609053075314, + 0.008724384009838104, + 0.009980015456676483, + 0.0038934971671551466, + 0.0031465294305235147, + 0.004239875823259354, + 0.003815441858023405, + 0.003156152321025729, + 0.007231017109006643, + 0.004126660991460085, + 0.005239177495241165, + 0.002937325509265065, + 0.0016144801629707217, + 0.001714821788482368, + 0.0017471257597208023, + 0.001673018909059465, + 0.0034053705167025328, + 0.0022768876515328884, + 0.0036075583193451166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.06209825351834297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001718117156997323 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001718117156997323, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001718117156997323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20525728166103363, + 0.07101995497941971, + 0.07974673062562943, + 0.16042600572109222, + 0.15305118262767792, + 0.07619775831699371, + 0.04050159454345703, + 0.0770215168595314, + 0.07279017567634583, + 0.06398782134056091 + ], + "credit_map": [ + 0.02651870623230934, + 0.02860592119395733, + 0.029878683388233185, + 0.033879633992910385, + 0.02563895285129547, + 0.01639592833817005, + 0.014742749743163586, + 0.02959670126438141, + 0.013586996123194695, + 0.006833458784967661, + 0.011210337281227112, + 0.005811944603919983, + 0.00744843203574419, + 0.005578426644206047, + 0.006101141218096018, + 0.014449216425418854, + 0.010363741777837276, + 0.006609018426388502, + 0.010488534346222878, + 0.005480493884533644, + 0.009974590502679348, + 0.017807159572839737, + 0.005768828559666872, + 0.013254370540380478, + 0.03366165980696678, + 0.0168040432035923, + 0.03210250288248062, + 0.011319490149617195, + 0.013167769648134708, + 0.0159505195915699, + 0.013932431116700172, + 0.02348758652806282, + 0.02188769169151783, + 0.01287881750613451, + 0.03441257402300835, + 0.01111091673374176, + 0.01673014461994171, + 0.015311386436223984, + 0.018323926255106926, + 0.02239571511745453, + 0.010893957689404488, + 0.0056924172677099705, + 0.00785859115421772, + 0.005150886252522469, + 0.008633201010525227, + 0.018860721960663795, + 0.008687719702720642, + 0.010420260019600391, + 0.005358788184821606, + 0.0043591707944869995, + 0.007756930775940418, + 0.004500541836023331, + 0.0036078221164643764, + 0.005025396589189768, + 0.004471078049391508, + 0.005421863403171301, + 0.009923531673848629, + 0.010191257111728191, + 0.010553871281445026, + 0.008428831584751606, + 0.00861755758523941, + 0.01090309675782919, + 0.008805313147604465, + 0.009598060511052608, + 0.009405596181750298, + 0.005628025624901056, + 0.009359209798276424, + 0.007560430094599724, + 0.006512637250125408, + 0.013192575424909592, + 0.009434787556529045, + 0.011696914210915565, + 0.008386724628508091, + 0.004535929299890995, + 0.0054413750767707825, + 0.008142896927893162, + 0.004529039841145277, + 0.011260808445513248, + 0.009498845785856247, + 0.012192203663289547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.001718117156997323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203409664332867 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203409664332867, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203409664332867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18388719856739044, + 0.10923121869564056, + 0.15999683737754822, + 0.10103373229503632, + 0.2122376561164856, + 0.05446697026491165, + 0.06618682295084, + 0.034752849489450455, + 0.030493419617414474, + 0.04771329462528229 + ], + "credit_map": [ + 0.026385145261883736, + 0.02328408695757389, + 0.024667756631970406, + 0.026798171922564507, + 0.021316904574632645, + 0.012503743171691895, + 0.013472678139805794, + 0.03545869514346123, + 0.01892787590622902, + 0.013174450024962425, + 0.017830992117524147, + 0.009700484573841095, + 0.012693440541625023, + 0.008808096870779991, + 0.007747987750917673, + 0.020347891375422478, + 0.03607015311717987, + 0.015327086672186852, + 0.025777874514460564, + 0.010012039914727211, + 0.01365603692829609, + 0.024301599711179733, + 0.009952076710760593, + 0.02489997260272503, + 0.025011800229549408, + 0.007606439758092165, + 0.02210669405758381, + 0.00617260904982686, + 0.007225214969366789, + 0.009084654971957207, + 0.00646459124982357, + 0.01736173778772354, + 0.03074033372104168, + 0.015783250331878662, + 0.03299293667078018, + 0.01709706149995327, + 0.01626027747988701, + 0.017486942932009697, + 0.04519105330109596, + 0.03668580204248428, + 0.007018440403044224, + 0.00439757714048028, + 0.0055741542018949986, + 0.003121462883427739, + 0.006595099810510874, + 0.014083554036915302, + 0.006346416659653187, + 0.007330263964831829, + 0.007537268567830324, + 0.007149240933358669, + 0.004345493391156197, + 0.007432945072650909, + 0.006842874456197023, + 0.010163028724491596, + 0.012305798940360546, + 0.010410173796117306, + 0.0046012126840651035, + 0.0032657035626471043, + 0.004427915904670954, + 0.003925422206521034, + 0.004185033030807972, + 0.004635165445506573, + 0.004174066241830587, + 0.005538330879062414, + 0.0048325639218091965, + 0.002684167120605707, + 0.004194121342152357, + 0.0034709249157458544, + 0.0026418569032102823, + 0.004252402577549219, + 0.00389626226387918, + 0.004521118942648172, + 0.01026817224919796, + 0.0038065817207098007, + 0.005739371292293072, + 0.005061624571681023, + 0.0039026413578540087, + 0.007148006930947304, + 0.004670160356909037, + 0.007116739638149738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.014203409664332867 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down macro shot captures a pristine, off-white sheet of densely textured art paper, its fibrous surface illuminated by soft, directed studio light, casting subtle shadows across its topography. At the center of the frame rests a perfectly placed drop of viscous, inky black ink, its surface shimmering with a wet, smooth sheen. In mesmerizing super-slow motion, an unseen force begins to pull the drop horizontally, causing it to gracefully stretch and elongate across the paper. The body of ink thins, its cohesiveness beginning to break down, splintering into a network of delicate, branching threads. These delicate, syrupy threads hypnotically extend outward, branching into ever more intricate, fragmented patterns resembling tree roots, creating a stark and beautiful contrast against the warm texture of the paper. The deliberate, silent unfolding of the ink creates an atmosphere of concentration, contemplation, and intense artistic intensity, capturing the moment of abstract creation.", + "seed": 9448, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_8/index_6926_seed_9448_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028827265510335563, + "vlm_logprob": -0.028827265510335563, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.028827265510335563, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0361245833337307 + ], + "opposite_logprob_tokens": [ + -3.6611244678497314 + ], + "logprob_reward": -0.0361245833337307, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0361245833337307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08945218473672867, + 0.16337712109088898, + 0.02618802897632122, + 0.06497792154550552, + 0.33086392283439636, + 0.07629773765802383, + 0.047822944819927216, + 0.05389390513300896, + 0.042658738791942596, + 0.10446754097938538 + ], + "credit_map": [ + 0.00872727669775486, + 0.014081232249736786, + 0.014701549895107746, + 0.014668862335383892, + 0.009765460155904293, + 0.012911414727568626, + 0.008002103306353092, + 0.006594281177967787, + 0.016302773728966713, + 0.042306672781705856, + 0.04029630124568939, + 0.01615237630903721, + 0.009841023944318295, + 0.014043723233044147, + 0.0075238789431750774, + 0.016910361126065254, + 0.003906628582626581, + 0.004597414284944534, + 0.002854591468349099, + 0.0035300941672176123, + 0.0022609340958297253, + 0.002365212421864271, + 0.0026058400981128216, + 0.0040673124603927135, + 0.02072778344154358, + 0.016971327364444733, + 0.01102634146809578, + 0.0035951361060142517, + 0.0019359547877684236, + 0.0019236073130741715, + 0.0014971151249483228, + 0.007300654426217079, + 0.12201903015375137, + 0.036528684198856354, + 0.023090390488505363, + 0.01248341053724289, + 0.009381165727972984, + 0.010490696877241135, + 0.010702257044613361, + 0.10616826266050339, + 0.009722412563860416, + 0.010688303969800472, + 0.014092775993049145, + 0.010466430336236954, + 0.006603587418794632, + 0.007014826405793428, + 0.005837905220687389, + 0.011871498078107834, + 0.009047474712133408, + 0.00729767233133316, + 0.009327400475740433, + 0.0031668664887547493, + 0.002471982967108488, + 0.004698537290096283, + 0.004265626892447472, + 0.007547379937022924, + 0.010077894665300846, + 0.005650163162499666, + 0.013310198672115803, + 0.004413996823132038, + 0.0032717178110033274, + 0.0032424076925963163, + 0.003717184066772461, + 0.010210342705249786, + 0.009304980747401714, + 0.005025628954172134, + 0.006662962958216667, + 0.003186512738466263, + 0.0030570458620786667, + 0.002649968024343252, + 0.0031410853844136, + 0.009630554355680943, + 0.018933214247226715, + 0.01913812942802906, + 0.018099382519721985, + 0.006555096711963415, + 0.008395091630518436, + 0.007490025833249092, + 0.0093529699370265, + 0.01650363579392433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.0361245833337307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005149126052856 + ], + "opposite_logprob_tokens": [ + -2.695051431655884 + ], + "logprob_reward": -0.07005149126052856, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005149126052856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1404298096895218, + 0.18348442018032074, + 0.09417664259672165, + 0.23469537496566772, + 0.16196805238723755, + 0.02347099967300892, + 0.01645697094500065, + 0.020531687885522842, + 0.026529841125011444, + 0.0982561782002449 + ], + "credit_map": [ + 0.01163624506443739, + 0.02181035280227661, + 0.02530568651854992, + 0.03376416116952896, + 0.010854496620595455, + 0.016812818124890327, + 0.01031577866524458, + 0.009930277243256569, + 0.016759149730205536, + 0.03969573229551315, + 0.03464971482753754, + 0.027317458763718605, + 0.01024661399424076, + 0.018603399395942688, + 0.010876086540520191, + 0.025336269289255142, + 0.020433872938156128, + 0.008835705928504467, + 0.02857401967048645, + 0.009724573232233524, + 0.006494483910501003, + 0.005400748457759619, + 0.004673992283642292, + 0.010039249435067177, + 0.039120256900787354, + 0.04816442355513573, + 0.019685598090291023, + 0.006716458592563868, + 0.0066952211782336235, + 0.01012248545885086, + 0.00580948730930686, + 0.09838145226240158, + 0.05921490862965584, + 0.018082259222865105, + 0.013601494021713734, + 0.006358626764267683, + 0.005187315866351128, + 0.005134339910000563, + 0.0055431826040148735, + 0.04884592443704605, + 0.0039767054840922356, + 0.002722988137975335, + 0.004804702941328287, + 0.0015821356792002916, + 0.0022921946365386248, + 0.0018225177191197872, + 0.0022269668988883495, + 0.004042787943035364, + 0.0027302615344524384, + 0.002708983374759555, + 0.004359290469437838, + 0.0011095718946307898, + 0.0008034765487536788, + 0.0011616945266723633, + 0.0007461069035343826, + 0.002837587147951126, + 0.0035535902716219425, + 0.0030925320461392403, + 0.005191349890083075, + 0.0013212653575465083, + 0.0011973080690950155, + 0.0013825963251292706, + 0.0011330173583701253, + 0.0036600299645215273, + 0.005898452363908291, + 0.003239700570702553, + 0.005232962314039469, + 0.0017186087789013982, + 0.001591313979588449, + 0.0021109466906636953, + 0.0013426759978756309, + 0.005395182408392429, + 0.03849790245294571, + 0.009178404696285725, + 0.006497809663414955, + 0.0019051661947742105, + 0.002902188804000616, + 0.0032782109919935465, + 0.0032627652399241924, + 0.0327337346971035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.07005149126052856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023318471387028694 + ], + "opposite_logprob_tokens": [ + -3.7733185291290283 + ], + "logprob_reward": -0.023318471387028694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023318471387028694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12082920968532562, + 0.30194613337516785, + 0.05765113607048988, + 0.13214804232120514, + 0.24010299146175385, + 0.03877398744225502, + 0.036041051149368286, + 0.008073863573372364, + 0.030158143490552902, + 0.03427543491125107 + ], + "credit_map": [ + 0.011283114552497864, + 0.017466098070144653, + 0.015517043881118298, + 0.01883535087108612, + 0.016088172793388367, + 0.018050944432616234, + 0.011567077599465847, + 0.012021411210298538, + 0.029061632230877876, + 0.08134404569864273, + 0.07635097205638885, + 0.036090027540922165, + 0.015254936181008816, + 0.022970756515860558, + 0.01322126854211092, + 0.02765248529613018, + 0.015577936545014381, + 0.00841426569968462, + 0.01267213560640812, + 0.005061486270278692, + 0.003007633378729224, + 0.00302139762789011, + 0.003035779343917966, + 0.006860503926873207, + 0.04075144603848457, + 0.044544052332639694, + 0.01159520260989666, + 0.004754181951284409, + 0.004666982684284449, + 0.003156835911795497, + 0.0026431060396134853, + 0.020036237314343452, + 0.09163486957550049, + 0.027070626616477966, + 0.015377688221633434, + 0.006021896842867136, + 0.0059996843338012695, + 0.006681362632662058, + 0.006622710730880499, + 0.08069416135549545, + 0.007160046603530645, + 0.004440364893525839, + 0.00551687041297555, + 0.0026935478672385216, + 0.0033564893528819084, + 0.0030394375789910555, + 0.004110160283744335, + 0.008457069285213947, + 0.005575297400355339, + 0.004957794677466154, + 0.0068823471665382385, + 0.002802270231768489, + 0.0035291865933686495, + 0.004063937347382307, + 0.0029375033918768167, + 0.005292714573442936, + 0.0013643641723319888, + 0.0009006029576994479, + 0.001011443673633039, + 0.0005819296929985285, + 0.0009737330838106573, + 0.0014126170426607132, + 0.0006376141682267189, + 0.0011915592476725578, + 0.00618688715621829, + 0.0031069645192474127, + 0.003582828212529421, + 0.002469051396474242, + 0.003193808486685157, + 0.0025419758167117834, + 0.0036603943444788456, + 0.005416232161223888, + 0.007078190334141254, + 0.007989143952727318, + 0.005608644802123308, + 0.002055137185379863, + 0.0018150421092286706, + 0.0020327141974121332, + 0.0020557723473757505, + 0.005640791729092598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.023318471387028694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004395472351461649 + ], + "opposite_logprob_tokens": [ + -7.750439643859863 + ], + "logprob_reward": -0.0004395472351461649, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004395472351461649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11507698148488998, + 0.09025798738002777, + 0.06142599508166313, + 0.057090118527412415, + 0.3500750958919525, + 0.023450572043657303, + 0.06668618321418762, + 0.05635176599025726, + 0.04221095144748688, + 0.13737432658672333 + ], + "credit_map": [ + 0.01237540040165186, + 0.013003832660615444, + 0.014438675716519356, + 0.030458062887191772, + 0.010976588353514671, + 0.0162704735994339, + 0.008660669438540936, + 0.008893275633454323, + 0.00678214430809021, + 0.019166870042681694, + 0.024603260681033134, + 0.011984080076217651, + 0.005026856437325478, + 0.007804100401699543, + 0.005041333381086588, + 0.00984934438019991, + 0.013775435276329517, + 0.00557383568957448, + 0.01701461896300316, + 0.00802368763834238, + 0.0033517165575176477, + 0.0035694881808012724, + 0.0032867032568901777, + 0.006830510683357716, + 0.015341612510383129, + 0.017868079245090485, + 0.0052704536356031895, + 0.0024335202760994434, + 0.002667362103238702, + 0.002183476695790887, + 0.001892801490612328, + 0.009432812221348286, + 0.13416725397109985, + 0.040082380175590515, + 0.021278588101267815, + 0.008295061998069286, + 0.008601659908890724, + 0.009961931966245174, + 0.009986571036279202, + 0.11770162731409073, + 0.0032348979730159044, + 0.0029093394987285137, + 0.004288332536816597, + 0.0020204256288707256, + 0.002646863926202059, + 0.002556117018684745, + 0.0017655763076618314, + 0.0040290215983986855, + 0.010523835197091103, + 0.009247913025319576, + 0.009081673808395863, + 0.007858929224312305, + 0.004787822254002094, + 0.005809694528579712, + 0.009076287969946861, + 0.010300029069185257, + 0.010258043184876442, + 0.008858702145516872, + 0.011017129756510258, + 0.0038097219076007605, + 0.003593801287934184, + 0.005757244769483805, + 0.003716630395501852, + 0.009340491145849228, + 0.009048301726579666, + 0.0037362920120358467, + 0.006425183266401291, + 0.0032445418182760477, + 0.003496896242722869, + 0.003169870004057884, + 0.003579096170142293, + 0.009510774165391922, + 0.05073368549346924, + 0.016777221113443375, + 0.009989466518163681, + 0.004145278595387936, + 0.003806410124525428, + 0.0053498405031859875, + 0.0051223584450781345, + 0.04145006462931633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.0004395472351461649 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pristine, off-white sheet of densely textured art paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a perfectly placed drop of viscous, inky black ink at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ink drop stretch, thin, and splinter into a network of branching threads resembling tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the ink unfolding in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202234335243702 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014202234335243702, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202234335243702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06563425064086914, + 0.041423458606004715, + 0.08295492082834244, + 0.07014131546020508, + 0.4798866808414459, + 0.07854347676038742, + 0.027129605412483215, + 0.03234735503792763, + 0.028604645282030106, + 0.09333425015211105 + ], + "credit_map": [ + 0.0062512001022696495, + 0.0069574713706970215, + 0.00989504437893629, + 0.016039006412029266, + 0.005758614744991064, + 0.009962723590433598, + 0.005851928144693375, + 0.004918264225125313, + 0.003230591304600239, + 0.008735291659832, + 0.008419550023972988, + 0.006346914917230606, + 0.002787545556202531, + 0.0044999998062849045, + 0.002610398456454277, + 0.004793168511241674, + 0.01782773993909359, + 0.009271802380681038, + 0.03023739717900753, + 0.005379761103540659, + 0.00420807022601366, + 0.0036080689169466496, + 0.0036560003645718098, + 0.008766084909439087, + 0.02715623565018177, + 0.011219322681427002, + 0.009024734608829021, + 0.003649854101240635, + 0.002813132246956229, + 0.00264928862452507, + 0.0026759766042232513, + 0.010952773503959179, + 0.18855874240398407, + 0.05579938367009163, + 0.026346124708652496, + 0.010977224446833134, + 0.010160328820347786, + 0.01290358416736126, + 0.010935829021036625, + 0.16420546174049377, + 0.015953384339809418, + 0.008218919858336449, + 0.011869940906763077, + 0.0064729866571724415, + 0.005199525970965624, + 0.005045688711106777, + 0.006840319838374853, + 0.018942708149552345, + 0.004339405335485935, + 0.0031964259687811136, + 0.004410693887621164, + 0.0020645521581172943, + 0.002960714278742671, + 0.0031469198875129223, + 0.002477535977959633, + 0.004533358383923769, + 0.0056830202229321, + 0.004456956870853901, + 0.006171568762511015, + 0.0022300169803202152, + 0.0020099380053579807, + 0.0035738551523536444, + 0.0021723241079598665, + 0.006049676798284054, + 0.004566529765725136, + 0.0030386031139642, + 0.004749333020299673, + 0.0023535334039479494, + 0.0029148238245397806, + 0.0025357878766953945, + 0.003765056375414133, + 0.004680975805968046, + 0.037665970623493195, + 0.011087420396506786, + 0.005020131357014179, + 0.0015053687384352088, + 0.002705541206523776, + 0.003632134525105357, + 0.0027892172802239656, + 0.028928464278578758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.014202234335243702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_9/index_tensor([290, 290])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_9/index_tensor([290, 290])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1173d207e8e09dd82f192d037b4f6cd9542b1752 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_9/index_tensor([290, 290])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Set against the chaotic backdrop of a fierce thunderstorm, this eye-level close-up focuses on a black ant desperately clinging to the smooth, trembling surface of a broad green leaf. The camera initially remains static, capturing the grimness of the scene. The ant's sleek black exoskeleton is speckled with tiny water droplets, its body twitching in jerky, unnatural motions, its legs twitching spasmodically against the leaf. A sinister, gray-white fungal stalk emerges eerily from the junction between its head and thorax, its alien form gleaming in the gloomy, stormy light. The camera then begins a slow, deliberate, and tight zoom, pushing past the ant's feeble flailing antennae. As the camera moves closer, the dark, rain-swept background of leaves fades into a complete blur, and the parasitic plant fills the frame. In this extreme close-up, stark, high-contrast shadows emphasize every gruesome detail: the fibrous texture of the stalk, the subtle pulsation at its base, and the eerie way it blends into the ant's body, creating a chilling atmosphere of biological terror.", + "seed": 15023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_9/index_290_seed_15023_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17942294841632248, + "vlm_logprob": -0.17942294841632248, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.17942294841632248, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010064803063869476 + ], + "opposite_logprob_tokens": [ + -5.510064601898193 + ], + "logprob_reward": -0.010064803063869476, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010064803063869476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676090508699417, + 0.06392171233892441, + 0.15950419008731842, + 0.03385423868894577, + 0.026974881067872047, + 0.058795053511857986, + 0.22894054651260376, + 0.02941151335835457, + 0.12901709973812103, + 0.10197166353464127 + ], + "credit_map": [ + 0.029407871887087822, + 0.027550382539629936, + 0.022054504603147507, + 0.014722988940775394, + 0.02491406723856926, + 0.014603479765355587, + 0.01895824261009693, + 0.015397517941892147, + 0.014234145171940327, + 0.014118767343461514, + 0.00899907760322094, + 0.00571871642023325, + 0.00494594918563962, + 0.006943604908883572, + 0.0043718465603888035, + 0.004589602816849947, + 0.03252530097961426, + 0.0330541767179966, + 0.015379785560071468, + 0.012500716373324394, + 0.020397180691361427, + 0.02318449504673481, + 0.00961589440703392, + 0.012846650555729866, + 0.004809271078556776, + 0.008060543797910213, + 0.005381283350288868, + 0.002814621664583683, + 0.0032241519074887037, + 0.003545504529029131, + 0.0025408267974853516, + 0.003478038590401411, + 0.0029891785234212875, + 0.0028609975706785917, + 0.0028861169703304768, + 0.0026434222236275673, + 0.0037810425274074078, + 0.00269638211466372, + 0.006192994769662619, + 0.002924749394878745, + 0.007388833910226822, + 0.011749513447284698, + 0.008838197216391563, + 0.006446231622248888, + 0.007588127627968788, + 0.007442518137395382, + 0.004997981712222099, + 0.0043436517007648945, + 0.08946791291236877, + 0.00737948901951313, + 0.09776165336370468, + 0.011678297072649002, + 0.005389672238379717, + 0.006474868394434452, + 0.005750355776399374, + 0.005038336850702763, + 0.004873118828982115, + 0.002413011621683836, + 0.004164834972470999, + 0.004440411925315857, + 0.0035921414382755756, + 0.0029699336737394333, + 0.0036819446831941605, + 0.003276118775829673, + 0.02465377375483513, + 0.01893368922173977, + 0.020284580066800117, + 0.02052472159266472, + 0.014783717691898346, + 0.010015370324254036, + 0.008893406018614769, + 0.010927844792604446, + 0.020593175664544106, + 0.014629040844738483, + 0.02690376713871956, + 0.010342231951653957, + 0.010366274043917656, + 0.007681324612349272, + 0.00547747639939189, + 0.0059783789329230785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.010064803063869476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261913061141968 + ], + "opposite_logprob_tokens": [ + -0.5761913061141968 + ], + "logprob_reward": -0.8261913061141968, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261913061141968, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1157817617058754, + 0.16427481174468994, + 0.1680251955986023, + 0.014870334416627884, + 0.04070669412612915, + 0.04938884824514389, + 0.12631313502788544, + 0.07986919581890106, + 0.14471007883548737, + 0.09606000036001205 + ], + "credit_map": [ + 0.028004126623272896, + 0.01769966259598732, + 0.011068781837821007, + 0.008167044259607792, + 0.015230940654873848, + 0.01192109752446413, + 0.009029265493154526, + 0.014660831540822983, + 0.03646386042237282, + 0.03561447560787201, + 0.031381528824567795, + 0.013436591252684593, + 0.014241836965084076, + 0.012304705567657948, + 0.009308869019150734, + 0.01152294036000967, + 0.03322957083582878, + 0.029506206512451172, + 0.015712890774011612, + 0.013362732715904713, + 0.022237682715058327, + 0.030351050198078156, + 0.009569958783686161, + 0.014055076986551285, + 0.0023833943996578455, + 0.002438685856759548, + 0.0023676969576627016, + 0.0018581824842840433, + 0.0014378490159288049, + 0.0017070716712623835, + 0.0013070919085294008, + 0.0013703616568818688, + 0.006870104465633631, + 0.005772202741354704, + 0.005801074206829071, + 0.00413325522094965, + 0.005143358837813139, + 0.004759264178574085, + 0.0037484383210539818, + 0.00447899429127574, + 0.007995628751814365, + 0.01453481800854206, + 0.005554073490202427, + 0.004040610510855913, + 0.0038479582872241735, + 0.006358887534588575, + 0.004003584384918213, + 0.0030532819218933582, + 0.0491267554461956, + 0.0033387811854481697, + 0.05421110987663269, + 0.007152270991355181, + 0.0029174135997891426, + 0.00333365797996521, + 0.0031583139207214117, + 0.003074823645874858, + 0.013759033754467964, + 0.010041764937341213, + 0.013684258796274662, + 0.008474287576973438, + 0.00790355633944273, + 0.011451659724116325, + 0.007159145083278418, + 0.007395479828119278, + 0.026798782870173454, + 0.01363812293857336, + 0.02699490822851658, + 0.014715800061821938, + 0.015882456675171852, + 0.015346197411417961, + 0.01616593450307846, + 0.01516787614673376, + 0.013026399537920952, + 0.020548516884446144, + 0.02106260135769844, + 0.010317441076040268, + 0.01114082895219326, + 0.009662666358053684, + 0.004284788854420185, + 0.0060167512856423855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.8261913061141968 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538649141788483 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538649141788483, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538649141788483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13831155002117157, + 0.03429461643099785, + 0.08038262277841568, + 0.04228191450238228, + 0.029093196615576744, + 0.023398609831929207, + 0.45184165239334106, + 0.057540152221918106, + 0.061840612441301346, + 0.08101506531238556 + ], + "credit_map": [ + 0.03276367112994194, + 0.0230852197855711, + 0.015724534168839455, + 0.009916183538734913, + 0.01899123191833496, + 0.013104883022606373, + 0.012868923135101795, + 0.011856905184686184, + 0.00622343085706234, + 0.008885146118700504, + 0.004037008620798588, + 0.002545078517869115, + 0.0031684923451393843, + 0.0036313675809651613, + 0.002067448338493705, + 0.0037366419564932585, + 0.014654443599283695, + 0.014340938068926334, + 0.007692885119467974, + 0.007027047220617533, + 0.011507725343108177, + 0.01292382925748825, + 0.005494554992765188, + 0.0067411987110972404, + 0.004965776111930609, + 0.0054228659719228745, + 0.007285378873348236, + 0.003534110030159354, + 0.0061144339852035046, + 0.00547559792175889, + 0.005768900271505117, + 0.003714853199198842, + 0.004065100569278002, + 0.004135794937610626, + 0.003040836425498128, + 0.002904320601373911, + 0.0034815010149031878, + 0.004013526253402233, + 0.0037184145767241716, + 0.0037337024696171284, + 0.0030341686215251684, + 0.00425401097163558, + 0.0027328412979841232, + 0.002513833111152053, + 0.003943354357033968, + 0.0024855718947947025, + 0.0024440872948616743, + 0.001990742515772581, + 0.17392516136169434, + 0.012909640558063984, + 0.19551080465316772, + 0.025309257209300995, + 0.009325964376330376, + 0.011800184845924377, + 0.009932042099535465, + 0.013128588907420635, + 0.009749689139425755, + 0.006500392686575651, + 0.007721839938312769, + 0.007932241074740887, + 0.006988962180912495, + 0.005451628472656012, + 0.006788208149373531, + 0.006407191976904869, + 0.008786150254309177, + 0.008333757519721985, + 0.00860561616718769, + 0.0076116579584777355, + 0.006913735065609217, + 0.00556530524045229, + 0.008209985680878162, + 0.00781440269201994, + 0.015863390639424324, + 0.012716834433376789, + 0.021992411464452744, + 0.008807359263300896, + 0.006513996049761772, + 0.0074728974141180515, + 0.0031344923190772533, + 0.004513682331889868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.012538649141788483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306488484144211 + ], + "opposite_logprob_tokens": [ + -3.168064832687378 + ], + "logprob_reward": -0.04306488484144211, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306488484144211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05877770483493805, + 0.05391603335738182, + 0.2144884616136551, + 0.13543666899204254, + 0.016076501458883286, + 0.07046613842248917, + 0.3362952470779419, + 0.015381215140223503, + 0.048714518547058105, + 0.0504474863409996 + ], + "credit_map": [ + 0.011297044344246387, + 0.014535119757056236, + 0.00465794000774622, + 0.003606007667258382, + 0.006109959911555052, + 0.006798832211643457, + 0.004025938920676708, + 0.007746861316263676, + 0.015095540322363377, + 0.012739420868456364, + 0.005048449616879225, + 0.003440448082983494, + 0.003799783531576395, + 0.005769322160631418, + 0.004595174454152584, + 0.003427897347137332, + 0.04191527143120766, + 0.03820761293172836, + 0.020555701106786728, + 0.016627606004476547, + 0.0301218181848526, + 0.03458631783723831, + 0.015269393101334572, + 0.017204752191901207, + 0.01513825822621584, + 0.018213169649243355, + 0.0261002779006958, + 0.014056027866899967, + 0.016947491094470024, + 0.016432024538517, + 0.01769806444644928, + 0.010851352475583553, + 0.0026050929445773363, + 0.0020167070906609297, + 0.0027176332660019398, + 0.001292646280489862, + 0.002257145708426833, + 0.0021405997686088085, + 0.0016747011104598641, + 0.001371974591165781, + 0.009020782075822353, + 0.011829246766865253, + 0.010449237190186977, + 0.007331093307584524, + 0.012363119050860405, + 0.007371782325208187, + 0.006141452584415674, + 0.0059594218619167805, + 0.12453877925872803, + 0.011988168582320213, + 0.13863204419612885, + 0.021268723532557487, + 0.00785946287214756, + 0.014728836715221405, + 0.007702968083322048, + 0.009576285257935524, + 0.0019296864047646523, + 0.0015825317241251469, + 0.0025135730393230915, + 0.00160084234084934, + 0.0015767799923196435, + 0.0020510584581643343, + 0.0022085432428866625, + 0.0019182012183591723, + 0.008321481756865978, + 0.005266837775707245, + 0.005998970940709114, + 0.006161476951092482, + 0.005398769862949848, + 0.0046254419721663, + 0.007703152019530535, + 0.005238386802375317, + 0.008579153567552567, + 0.009628592059016228, + 0.013062630780041218, + 0.004974133335053921, + 0.005448048003017902, + 0.003970674704760313, + 0.0018010108033195138, + 0.002983243204653263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.04306488484144211 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255098920315504 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255098920315504, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255098920315504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14980241656303406, + 0.10624108463525772, + 0.0706976130604744, + 0.027052778750658035, + 0.031002802774310112, + 0.05250024423003197, + 0.19923678040504456, + 0.024861566722393036, + 0.21049818396568298, + 0.12810657918453217 + ], + "credit_map": [ + 0.030805297195911407, + 0.026958201080560684, + 0.01289394311606884, + 0.011364820413291454, + 0.014756644144654274, + 0.016435902565717697, + 0.0148976044729352, + 0.02169000543653965, + 0.02913527563214302, + 0.02396601252257824, + 0.017288433387875557, + 0.005758909974247217, + 0.007853245362639427, + 0.007362570147961378, + 0.006727199535816908, + 0.008149437606334686, + 0.012374837882816792, + 0.010372198186814785, + 0.0066557638347148895, + 0.0064321099780499935, + 0.009301591664552689, + 0.01474366057664156, + 0.0042948611080646515, + 0.006522580981254578, + 0.0029527724254876375, + 0.0038029453717172146, + 0.0048521896824240685, + 0.002625588793307543, + 0.0033406042493879795, + 0.0037883007898926735, + 0.003302657511085272, + 0.0023877215571701527, + 0.003996518440544605, + 0.004467407241463661, + 0.0049784122966229916, + 0.0027827562298625708, + 0.00320912036113441, + 0.005011111963540316, + 0.003018743358552456, + 0.00353873404674232, + 0.008846070617437363, + 0.014494270086288452, + 0.00630574906244874, + 0.0038402273785322905, + 0.004378285259008408, + 0.006369258742779493, + 0.004617178812623024, + 0.0036492012441158295, + 0.07582534849643707, + 0.006834669038653374, + 0.08295813202857971, + 0.012644259259104729, + 0.0039715394377708435, + 0.006520829629153013, + 0.004757920280098915, + 0.005724084097892046, + 0.0034802209120243788, + 0.003756168531253934, + 0.004241977818310261, + 0.002574555343016982, + 0.0023892968893051147, + 0.004803923889994621, + 0.0017374433809891343, + 0.0018779805395752192, + 0.038015395402908325, + 0.025022605434060097, + 0.037569478154182434, + 0.02773243375122547, + 0.02426069788634777, + 0.021102383732795715, + 0.01809295266866684, + 0.01870224066078663, + 0.021341389045119286, + 0.0263688787817955, + 0.03225143998861313, + 0.012648039497435093, + 0.013025945983827114, + 0.01105888094753027, + 0.005313931033015251, + 0.006098069250583649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.005255098920315504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set against the chaotic backdrop of a fierce thunderstorm, this eye-level close-up focuses on a black ant desperately clinging to the smooth, trembling surface of a broad green leaf. The camera initially remains static, capturing the grimness of the scene. The ant's sleek black exoskeleton is speckled with tiny water droplets, its body twitching in jerky, unnatural motions, its legs twitching spasmodically against the leaf. A sinister, gray-white fungal stalk emerges eerily from the junction between its head and thorax, its alien form gleaming in the gloomy, stormy light. The camera then begins a slow, deliberate, and tight zoom, pushing past the ant's feeble flailing antennae. As the camera moves closer, the dark, rain-swept background of leaves fades into a complete blur, and the parasitic plant fills the frame. In this extreme close-up, stark, high-contrast shadows emphasize every gruesome detail: the fibrous texture of the stalk, the subtle pulsation at its base, and the eerie way it blends into the ant's body, creating a chilling atmosphere of biological terror.", + "seed": 15024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_9/index_290_seed_15024_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08470387738198042, + "vlm_logprob": -0.08470387738198042, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.08470387738198042, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012005925178527832 + ], + "opposite_logprob_tokens": [ + -5.262005805969238 + ], + "logprob_reward": -0.012005925178527832, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.012005925178527832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11030302941799164, + 0.04537971690297127, + 0.07623080909252167, + 0.051111094653606415, + 0.1301547884941101, + 0.15958771109580994, + 0.08060509711503983, + 0.210885688662529, + 0.05724628269672394, + 0.0784958004951477 + ], + "credit_map": [ + 0.016057763248682022, + 0.019266238436102867, + 0.010901331901550293, + 0.017466183751821518, + 0.014084998518228531, + 0.010271665640175343, + 0.009203989990055561, + 0.01305085327476263, + 0.006321997381746769, + 0.0060636186972260475, + 0.008459024131298065, + 0.004431452602148056, + 0.003800684353336692, + 0.005221371538937092, + 0.005546784494072199, + 0.005534786265343428, + 0.015212509781122208, + 0.011606802232563496, + 0.007927091792225838, + 0.006456532049924135, + 0.011452564969658852, + 0.006657944526523352, + 0.0062777637504041195, + 0.010639605112373829, + 0.007389899343252182, + 0.00797179527580738, + 0.005363962147384882, + 0.007212009280920029, + 0.00468925666064024, + 0.005440437234938145, + 0.006655836943536997, + 0.006387895438820124, + 0.01547609269618988, + 0.014815169386565685, + 0.025850174948573112, + 0.022454114630818367, + 0.012068495154380798, + 0.012782578356564045, + 0.014650656841695309, + 0.012057500891387463, + 0.02646041475236416, + 0.014016864821314812, + 0.022085370495915413, + 0.0273869801312685, + 0.010884911753237247, + 0.010346517898142338, + 0.03065682388842106, + 0.017749814316630363, + 0.012847010977566242, + 0.01159525103867054, + 0.012984038330614567, + 0.01139842625707388, + 0.006542020943015814, + 0.009266098029911518, + 0.008232834748923779, + 0.007739416789263487, + 0.039656560868024826, + 0.012847564183175564, + 0.06653335690498352, + 0.03168908506631851, + 0.010122830048203468, + 0.009957714937627316, + 0.0246884748339653, + 0.01539008691906929, + 0.01021934300661087, + 0.006159313954412937, + 0.008350910618901253, + 0.005958691239356995, + 0.006534338463097811, + 0.0038035816978663206, + 0.008385488763451576, + 0.007834612391889095, + 0.010209856554865837, + 0.009561345912516117, + 0.014266310259699821, + 0.011123058386147022, + 0.008154205977916718, + 0.008661857806146145, + 0.008594971150159836, + 0.00792420282959938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.012005925178527832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34870418906211853 + ], + "opposite_logprob_tokens": [ + -1.223704218864441 + ], + "logprob_reward": -0.34870418906211853, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34870418906211853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21742360293865204, + 0.09352266788482666, + 0.1302698254585266, + 0.13025200366973877, + 0.08644072711467743, + 0.04851768910884857, + 0.06333783268928528, + 0.15344087779521942, + 0.044034138321876526, + 0.032760653644800186 + ], + "credit_map": [ + 0.024091485887765884, + 0.03566637262701988, + 0.022296737879514694, + 0.024323951452970505, + 0.028891071677207947, + 0.024026619270443916, + 0.02027995139360428, + 0.03784741833806038, + 0.011205881834030151, + 0.010795638896524906, + 0.014110786840319633, + 0.010640967637300491, + 0.011584657244384289, + 0.014675169251859188, + 0.010315137915313244, + 0.010194431059062481, + 0.015361426398158073, + 0.023205483332276344, + 0.013958974741399288, + 0.010619526728987694, + 0.03156299889087677, + 0.012552366591989994, + 0.008428131230175495, + 0.01458091288805008, + 0.01871725544333458, + 0.012737274169921875, + 0.016052929684519768, + 0.01657564379274845, + 0.010872418992221355, + 0.02020358480513096, + 0.020780472084879875, + 0.014312425628304482, + 0.01162654161453247, + 0.008065273985266685, + 0.020385755226016045, + 0.011729012243449688, + 0.005820028483867645, + 0.007881955243647099, + 0.010835066437721252, + 0.010097095742821693, + 0.008907672949135303, + 0.004356912802904844, + 0.008414877578616142, + 0.005730575881898403, + 0.004202785901725292, + 0.003631445812061429, + 0.00784433912485838, + 0.0054290806874632835, + 0.009834479540586472, + 0.00577266002073884, + 0.007888856343925, + 0.007319086231291294, + 0.004553476348519325, + 0.00731278769671917, + 0.009504424408078194, + 0.01115206629037857, + 0.023812120780348778, + 0.011843863874673843, + 0.04859945923089981, + 0.024936387315392494, + 0.009189656004309654, + 0.0076133450493216515, + 0.01528860256075859, + 0.012157446704804897, + 0.013273057527840137, + 0.003911420702934265, + 0.003916269633919001, + 0.0030800544191151857, + 0.003788406029343605, + 0.00271563115529716, + 0.004586676601320505, + 0.008762619458138943, + 0.005096363835036755, + 0.003914727363735437, + 0.005143363960087299, + 0.0032265433110296726, + 0.0038850384298712015, + 0.003336333902552724, + 0.003412611549720168, + 0.004745671525597572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.34870418906211853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16125716269016266, + 0.1056450828909874, + 0.06746865808963776, + 0.10580938309431076, + 0.11469513177871704, + 0.11564517021179199, + 0.09478358924388885, + 0.12087657302618027, + 0.035556089133024216, + 0.07826317101716995 + ], + "credit_map": [ + 0.021180326119065285, + 0.020713994279503822, + 0.019959092140197754, + 0.017039045691490173, + 0.024228325113654137, + 0.01580861024558544, + 0.015707967802882195, + 0.026619812473654747, + 0.01418814156204462, + 0.015592722222208977, + 0.008420544676482677, + 0.011328032240271568, + 0.02344195544719696, + 0.01174531877040863, + 0.010258389636874199, + 0.010669983923435211, + 0.01087588258087635, + 0.009189723990857601, + 0.010147702880203724, + 0.005974072031676769, + 0.009119420312345028, + 0.006761796306818724, + 0.004447912331670523, + 0.010952148586511612, + 0.014243518933653831, + 0.01256316713988781, + 0.014957431703805923, + 0.02086971513926983, + 0.008286312222480774, + 0.01092176791280508, + 0.011990818195044994, + 0.011976656503975391, + 0.0159388929605484, + 0.014010404236614704, + 0.019862262532114983, + 0.02200324833393097, + 0.013797270134091377, + 0.006838652305305004, + 0.011488398537039757, + 0.01075600553303957, + 0.020039746537804604, + 0.012483736500144005, + 0.015785736963152885, + 0.018864508718252182, + 0.007696896325796843, + 0.00792813953012228, + 0.01999923400580883, + 0.01284718420356512, + 0.014053893275558949, + 0.012411341071128845, + 0.014739143662154675, + 0.010484093800187111, + 0.008197241462767124, + 0.010759339667856693, + 0.010843886062502861, + 0.013294653967022896, + 0.020952874794602394, + 0.010911771096289158, + 0.02491612918674946, + 0.0193024892359972, + 0.008733240887522697, + 0.009970102459192276, + 0.013108843006193638, + 0.01298112329095602, + 0.008985943160951138, + 0.0034231399185955524, + 0.0030471626669168472, + 0.003134120488539338, + 0.004129888489842415, + 0.0025452293921262026, + 0.004931952338665724, + 0.005358654074370861, + 0.01104053296148777, + 0.012790228240191936, + 0.008473711088299751, + 0.00902822520583868, + 0.010998927056789398, + 0.009258146397769451, + 0.009707857854664326, + 0.00696554547175765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810177743434906 + ], + "opposite_logprob_tokens": [ + -3.2881016731262207 + ], + "logprob_reward": -0.03810177743434906, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810177743434906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17780819535255432, + 0.05968007445335388, + 0.09938982129096985, + 0.08872582018375397, + 0.1798778623342514, + 0.10397402197122574, + 0.07526785880327225, + 0.09069109708070755, + 0.037740979343652725, + 0.08684425055980682 + ], + "credit_map": [ + 0.021279260516166687, + 0.02908838912844658, + 0.02020697109401226, + 0.021030930802226067, + 0.030221832916140556, + 0.017646288499236107, + 0.016732869669795036, + 0.021601635962724686, + 0.007591755595058203, + 0.006517975591123104, + 0.00929892435669899, + 0.006871548481285572, + 0.007544838823378086, + 0.008273623883724213, + 0.006638102233409882, + 0.006943301763385534, + 0.013172343373298645, + 0.01744106225669384, + 0.01010692585259676, + 0.008132143877446651, + 0.020585088059306145, + 0.00900882389396429, + 0.007113642059266567, + 0.013829795643687248, + 0.012209351174533367, + 0.014878041110932827, + 0.010054138489067554, + 0.012227477505803108, + 0.009888180531561375, + 0.010204636491835117, + 0.010557234287261963, + 0.008706762455403805, + 0.022488858550786972, + 0.02302650548517704, + 0.03251097723841667, + 0.02010636031627655, + 0.02516237460076809, + 0.015057877637445927, + 0.02359306439757347, + 0.017931843176484108, + 0.01727147214114666, + 0.010381323285400867, + 0.011793803423643112, + 0.019734691828489304, + 0.010374895296990871, + 0.0049905781634151936, + 0.018150467425584793, + 0.01127678994089365, + 0.011266565881669521, + 0.009886331856250763, + 0.011724594049155712, + 0.010402847081422806, + 0.00667858449742198, + 0.009074046276509762, + 0.007869615219533443, + 0.008365273475646973, + 0.019556501880288124, + 0.007505381479859352, + 0.016067450866103172, + 0.014790135435760021, + 0.006331232842057943, + 0.005901385564357042, + 0.010308163240551949, + 0.010230856016278267, + 0.011686026118695736, + 0.003067577723413706, + 0.0035359757021069527, + 0.002567639574408531, + 0.003331698477268219, + 0.0025421215686947107, + 0.0033004391007125378, + 0.007709501311182976, + 0.010818535462021828, + 0.014655979350209236, + 0.009616206400096416, + 0.009576503187417984, + 0.01242463756352663, + 0.012240584008395672, + 0.00811428390443325, + 0.009397519752383232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.03810177743434906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864217709749937 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.00864217709749937, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864217709749937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1818769872188568, + 0.0677923932671547, + 0.07154423743486404, + 0.043252721428871155, + 0.08691571652889252, + 0.06382102519273758, + 0.04365018010139465, + 0.15257160365581512, + 0.1591041088104248, + 0.1294710338115692 + ], + "credit_map": [ + 0.0196224357932806, + 0.032982829958200455, + 0.02144583873450756, + 0.027515796944499016, + 0.021406451240181923, + 0.01504439301788807, + 0.01713656634092331, + 0.026722678914666176, + 0.008362415246665478, + 0.0072495583444833755, + 0.009339941665530205, + 0.008611578494310379, + 0.008894518949091434, + 0.00976930558681488, + 0.007806616369634867, + 0.007758462335914373, + 0.008510733023285866, + 0.012453237548470497, + 0.008206482045352459, + 0.005961690563708544, + 0.01155814342200756, + 0.006857958156615496, + 0.007014182861894369, + 0.010981809347867966, + 0.006317295134067535, + 0.007199073676019907, + 0.004029543604701757, + 0.004655094817280769, + 0.0037296328227967024, + 0.004742885008454323, + 0.005751430056989193, + 0.0068277642130851746, + 0.011928192339837551, + 0.009607608430087566, + 0.019259637221693993, + 0.012735486030578613, + 0.008920848369598389, + 0.006790028885006905, + 0.0076825921423733234, + 0.009991319850087166, + 0.009923375211656094, + 0.006108307745307684, + 0.010825934819877148, + 0.0104086147621274, + 0.006251294631510973, + 0.003933332394808531, + 0.010533491149544716, + 0.005836677737534046, + 0.006808939855545759, + 0.004456163849681616, + 0.007882975041866302, + 0.004382897634059191, + 0.0035206156317144632, + 0.004884915426373482, + 0.005610314663499594, + 0.006103354971855879, + 0.02574053965508938, + 0.015845781192183495, + 0.03496553748846054, + 0.02472897246479988, + 0.01088356226682663, + 0.01152754295617342, + 0.014719457365572453, + 0.014160207472741604, + 0.051356054842472076, + 0.012321419082581997, + 0.012236800976097584, + 0.011190582998096943, + 0.017628174275159836, + 0.008659068495035172, + 0.015548641793429852, + 0.030163370072841644, + 0.019125228747725487, + 0.020923016592860222, + 0.015330640599131584, + 0.01418178528547287, + 0.017568549141287804, + 0.013144380412995815, + 0.015325162559747696, + 0.013872269541025162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.00864217709749937 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set against the chaotic backdrop of a fierce thunderstorm, this eye-level close-up focuses on a black ant desperately clinging to the smooth, trembling surface of a broad green leaf. The camera initially remains static, capturing the grimness of the scene. The ant's sleek black exoskeleton is speckled with tiny water droplets, its body twitching in jerky, unnatural motions, its legs twitching spasmodically against the leaf. A sinister, gray-white fungal stalk emerges eerily from the junction between its head and thorax, its alien form gleaming in the gloomy, stormy light. The camera then begins a slow, deliberate, and tight zoom, pushing past the ant's feeble flailing antennae. As the camera moves closer, the dark, rain-swept background of leaves fades into a complete blur, and the parasitic plant fills the frame. In this extreme close-up, stark, high-contrast shadows emphasize every gruesome detail: the fibrous texture of the stalk, the subtle pulsation at its base, and the eerie way it blends into the ant's body, creating a chilling atmosphere of biological terror.", + "seed": 15025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_9/index_290_seed_15025_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.25720888618379834, + "vlm_logprob": -0.25720888618379834, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.25720888618379834, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007842346094548702 + ], + "opposite_logprob_tokens": [ + -5.757842540740967 + ], + "logprob_reward": -0.007842346094548702, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007842346094548702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6078419089317322, + 0.035870250314474106, + 0.041396304965019226, + 0.041728224605321884, + 0.028146805241703987, + 0.11367523670196533, + 0.04415014013648033, + 0.031624309718608856, + 0.020406048744916916, + 0.03516073152422905 + ], + "credit_map": [ + 0.03127660974860191, + 0.22471436858177185, + 0.012322261929512024, + 0.011581460013985634, + 0.026728134602308273, + 0.0427931509912014, + 0.018102960661053658, + 0.24032291769981384, + 0.004414290655404329, + 0.004649478010833263, + 0.003912627696990967, + 0.004610264208167791, + 0.004292734432965517, + 0.004740861710160971, + 0.004637383855879307, + 0.004612612072378397, + 0.007848214358091354, + 0.0062692612409591675, + 0.005887439474463463, + 0.003420890774577856, + 0.003211161820217967, + 0.005526343826204538, + 0.0038762460462749004, + 0.005356750451028347, + 0.009132961742579937, + 0.0049321819096803665, + 0.0038546447176486254, + 0.004526112228631973, + 0.0039536659605801105, + 0.005085139535367489, + 0.002775224158540368, + 0.007468292489647865, + 0.0037966014351695776, + 0.005170020274817944, + 0.0026721986941993237, + 0.002193512162193656, + 0.003141490276902914, + 0.003631547326222062, + 0.0036124172620475292, + 0.003929017577320337, + 0.015372308902442455, + 0.014910460449755192, + 0.015719005838036537, + 0.012463885359466076, + 0.012101937085390091, + 0.01361326314508915, + 0.013068297877907753, + 0.01642609015107155, + 0.004983674269169569, + 0.006905768532305956, + 0.00503147067502141, + 0.0031873765401542187, + 0.0076985773630440235, + 0.006957704667001963, + 0.004195619374513626, + 0.005189942661672831, + 0.0052819750271737576, + 0.003909144084900618, + 0.003299906849861145, + 0.003142370842397213, + 0.0035651777870953083, + 0.0032653063535690308, + 0.003555543487891555, + 0.005604886449873447, + 0.002321152715012431, + 0.0025423907209187746, + 0.002753998851403594, + 0.003031735774129629, + 0.002776381326839328, + 0.0021134973503649235, + 0.002170366467908025, + 0.0026965271681547165, + 0.0030480101704597473, + 0.003479550126940012, + 0.004507256206125021, + 0.004330709110945463, + 0.005463660228997469, + 0.0035323614720255136, + 0.006302827037870884, + 0.004496357869356871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.007842346094548702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236380577087402 + ], + "opposite_logprob_tokens": [ + -0.3486380875110626 + ], + "logprob_reward": -1.2236380577087402, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236380577087402, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.32364511489868164, + 0.07359003275632858, + 0.18995210528373718, + 0.07031473517417908, + 0.10384119302034378, + 0.043512631207704544, + 0.03425244241952896, + 0.06058493256568909, + 0.04373463615775108, + 0.05657220259308815 + ], + "credit_map": [ + 0.026468997821211815, + 0.10677452385425568, + 0.012128234840929508, + 0.011411686427891254, + 0.016753366217017174, + 0.029671858996152878, + 0.011909144930541515, + 0.1085272878408432, + 0.008396952413022518, + 0.012626164592802525, + 0.006483853794634342, + 0.007109763100743294, + 0.010081413201987743, + 0.011357453651726246, + 0.008830307051539421, + 0.008704128675162792, + 0.019905613735318184, + 0.05017085373401642, + 0.009670278988778591, + 0.008492892608046532, + 0.016020989045500755, + 0.02546418271958828, + 0.01504478044807911, + 0.04518250375986099, + 0.014297464862465858, + 0.010391377843916416, + 0.005086020100861788, + 0.0064008887857198715, + 0.005564873572438955, + 0.006742485333234072, + 0.005120855290442705, + 0.016710767522454262, + 0.01648365706205368, + 0.013660432770848274, + 0.018845118582248688, + 0.00697378208860755, + 0.008112338371574879, + 0.015378713607788086, + 0.008207951672375202, + 0.01617920771241188, + 0.004003362730145454, + 0.005528181791305542, + 0.0035654669627547264, + 0.005500501021742821, + 0.007489367388188839, + 0.0056924414820969105, + 0.005355968605726957, + 0.006377344019711018, + 0.0038096203934401274, + 0.0037337683606892824, + 0.004613795783370733, + 0.0027067919727414846, + 0.004751782398670912, + 0.007303117774426937, + 0.0028553535230457783, + 0.004478210583329201, + 0.010636416263878345, + 0.007587202824652195, + 0.006173967849463224, + 0.005473925266414881, + 0.007082947064191103, + 0.009798584505915642, + 0.007012132555246353, + 0.006819756235927343, + 0.00661295372992754, + 0.003285389393568039, + 0.010884291492402554, + 0.0058673289604485035, + 0.004062206018716097, + 0.0024096330162137747, + 0.0031096083112061024, + 0.007503223139792681, + 0.006708556320518255, + 0.005597169045358896, + 0.01080621499568224, + 0.004796159919351339, + 0.007334727793931961, + 0.008309631608426571, + 0.005910962820053101, + 0.007108780089765787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -1.2236380577087402 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076274327933788 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076274327933788, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076274327933788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3998981714248657, + 0.03584297373890877, + 0.12055251002311707, + 0.033316720277071, + 0.07360629737377167, + 0.10994698852300644, + 0.11821352690458298, + 0.03623376786708832, + 0.03191516175866127, + 0.04047388583421707 + ], + "credit_map": [ + 0.018937785178422928, + 0.15100586414337158, + 0.007735728286206722, + 0.008024138398468494, + 0.017568526789546013, + 0.026518970727920532, + 0.01011610683053732, + 0.159991055727005, + 0.00368895404972136, + 0.0051743583753705025, + 0.004280949477106333, + 0.003019805997610092, + 0.005171856842935085, + 0.004618756007403135, + 0.005075013265013695, + 0.0048132832162082195, + 0.01408332772552967, + 0.0322624146938324, + 0.007891659624874592, + 0.007406293414533138, + 0.008893439546227455, + 0.011068550869822502, + 0.01142161712050438, + 0.02752520889043808, + 0.00803681556135416, + 0.004303800407797098, + 0.0023178434930741787, + 0.002932177623733878, + 0.002976920222863555, + 0.0042825667187571526, + 0.002806460252031684, + 0.005660140421241522, + 0.011928455904126167, + 0.009603436104953289, + 0.01183953694999218, + 0.004275180399417877, + 0.00550980307161808, + 0.009925921447575092, + 0.00671283807605505, + 0.013811135664582253, + 0.013465207070112228, + 0.01581118255853653, + 0.01412707008421421, + 0.012681153602898121, + 0.013640834018588066, + 0.013299593701958656, + 0.011717566289007664, + 0.015204383060336113, + 0.017443828284740448, + 0.013492354191839695, + 0.010224144905805588, + 0.00788799487054348, + 0.014876098372042179, + 0.02859550714492798, + 0.009269541129469872, + 0.016424065455794334, + 0.0062187002040445805, + 0.0043797786347568035, + 0.003570548491552472, + 0.0031137955375015736, + 0.0035146018490195274, + 0.005112834740430117, + 0.00381724932231009, + 0.0065062642097473145, + 0.0042486959137022495, + 0.0023031237069517374, + 0.0034027141518890858, + 0.004812232218682766, + 0.003333199070766568, + 0.00634700246155262, + 0.0034522993955761194, + 0.0040158978663384914, + 0.004325697664171457, + 0.004193785134702921, + 0.0049811555072665215, + 0.004995896015316248, + 0.00770312687382102, + 0.004543748218566179, + 0.005022899713367224, + 0.004707581363618374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.011076274327933788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033703748136758804 + ], + "opposite_logprob_tokens": [ + -3.4087038040161133 + ], + "logprob_reward": -0.033703748136758804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033703748136758804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34485405683517456, + 0.045015785843133926, + 0.19599837064743042, + 0.06772277504205704, + 0.10587483644485474, + 0.04925763979554176, + 0.03210867568850517, + 0.055435340851545334, + 0.04859089478850365, + 0.05514160916209221 + ], + "credit_map": [ + 0.01731623336672783, + 0.13088013231754303, + 0.006306301802396774, + 0.007295095361769199, + 0.01589340530335903, + 0.017561759799718857, + 0.010051838122308254, + 0.13954930007457733, + 0.004698545206338167, + 0.007602107711136341, + 0.005722530651837587, + 0.004068086855113506, + 0.00615608599036932, + 0.004826574120670557, + 0.005427564959973097, + 0.006514296866953373, + 0.022823913022875786, + 0.04663136601448059, + 0.0160333514213562, + 0.01320648193359375, + 0.015803247690200806, + 0.02361200377345085, + 0.017194950953125954, + 0.04069305211305618, + 0.016217704862356186, + 0.007501415908336639, + 0.0035192130599170923, + 0.005410115234553814, + 0.004863198380917311, + 0.00677330419421196, + 0.005346102640032768, + 0.018091728910803795, + 0.019233837723731995, + 0.01284851972013712, + 0.011377324350178242, + 0.0075834994204342365, + 0.012345402501523495, + 0.009346868842840195, + 0.009922458790242672, + 0.023216936737298965, + 0.005717611405998468, + 0.008745206519961357, + 0.004581514745950699, + 0.005299587734043598, + 0.006496835965663195, + 0.007693002466112375, + 0.0045622289180755615, + 0.006161658558994532, + 0.003519294550642371, + 0.005921040661633015, + 0.003034930443391204, + 0.0028515756130218506, + 0.004969550296664238, + 0.004744125995784998, + 0.0030635634902864695, + 0.00400459673255682, + 0.007748804520815611, + 0.0077556767500936985, + 0.005543197505176067, + 0.005884081590920687, + 0.0074198502115905285, + 0.008057170547544956, + 0.006005909759551287, + 0.0070206583477556705, + 0.004659100901335478, + 0.006473549176007509, + 0.009049070999026299, + 0.0061811781488358974, + 0.006180648226290941, + 0.0032424824312329292, + 0.005949944723397493, + 0.006854924373328686, + 0.004030220210552216, + 0.005839834455400705, + 0.004651468247175217, + 0.0076870061457157135, + 0.00985325500369072, + 0.006541791372001171, + 0.010063156485557556, + 0.006474880501627922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.033703748136758804 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784004651010036 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784004651010036, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784004651010036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13226318359375, + 0.10188835859298706, + 0.21736936271190643, + 0.1255505234003067, + 0.08239506185054779, + 0.03825562819838524, + 0.08262408524751663, + 0.14206624031066895, + 0.030712058767676353, + 0.046875495463609695 + ], + "credit_map": [ + 0.013019964098930359, + 0.03310086950659752, + 0.013557827100157738, + 0.00762442871928215, + 0.010841433890163898, + 0.011914031580090523, + 0.007440161891281605, + 0.034764472395181656, + 0.01291122566908598, + 0.01677967607975006, + 0.014045272022485733, + 0.008430033922195435, + 0.012292711064219475, + 0.013046151027083397, + 0.011867421679198742, + 0.012515859678387642, + 0.030853569507598877, + 0.0546027272939682, + 0.01574847288429737, + 0.01204992737621069, + 0.018014512956142426, + 0.023003555834293365, + 0.018284298479557037, + 0.04481228440999985, + 0.03481873497366905, + 0.014680221676826477, + 0.006806278135627508, + 0.00635976018384099, + 0.005815469194203615, + 0.010709308087825775, + 0.005723498295992613, + 0.04063726216554642, + 0.012685786932706833, + 0.01286850031465292, + 0.010653037577867508, + 0.006869093980640173, + 0.006671181879937649, + 0.01168868038803339, + 0.008373282849788666, + 0.01258549839258194, + 0.005537465214729309, + 0.005880014505237341, + 0.003237799508497119, + 0.003749686060473323, + 0.006612889003008604, + 0.003977806307375431, + 0.003974052146077156, + 0.0052859182469546795, + 0.010365893132984638, + 0.01160752959549427, + 0.007906967774033546, + 0.005835686344653368, + 0.014680151827633381, + 0.01591206155717373, + 0.006001900415867567, + 0.01031389832496643, + 0.022931797429919243, + 0.017738856375217438, + 0.014019458554685116, + 0.015148484148085117, + 0.018661607056856155, + 0.020954934880137444, + 0.01563810557126999, + 0.01697300188243389, + 0.0029536506626755, + 0.0023587029427289963, + 0.004630868788808584, + 0.004321690648794174, + 0.0028744973242282867, + 0.0049059768207371235, + 0.004032882861793041, + 0.004633789882063866, + 0.005719997454434633, + 0.004458575043827295, + 0.006686550565063953, + 0.0056064981035888195, + 0.005996439140290022, + 0.006978212855756283, + 0.005753861274570227, + 0.00567536149173975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.009784004651010036 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Set against the chaotic backdrop of a fierce thunderstorm, this eye-level close-up focuses on a black ant desperately clinging to the smooth, trembling surface of a broad green leaf. The camera initially remains static, capturing the grimness of the scene. The ant's sleek black exoskeleton is speckled with tiny water droplets, its body twitching in jerky, unnatural motions, its legs twitching spasmodically against the leaf. A sinister, gray-white fungal stalk emerges eerily from the junction between its head and thorax, its alien form gleaming in the gloomy, stormy light. The camera then begins a slow, deliberate, and tight zoom, pushing past the ant's feeble flailing antennae. As the camera moves closer, the dark, rain-swept background of leaves fades into a complete blur, and the parasitic plant fills the frame. In this extreme close-up, stark, high-contrast shadows emphasize every gruesome detail: the fibrous texture of the stalk, the subtle pulsation at its base, and the eerie way it blends into the ant's body, creating a chilling atmosphere of biological terror.", + "seed": 15026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000161/group_9/index_290_seed_15026_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11841584919020534, + "vlm_logprob": -0.11841584919020534, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.11841584919020534, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013594076037406921 + ], + "opposite_logprob_tokens": [ + -4.888594150543213 + ], + "logprob_reward": -0.013594076037406921, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.013594076037406921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44853419065475464, + 0.20594064891338348, + 0.03363555669784546, + 0.03456812724471092, + 0.047546979039907455, + 0.02264554426074028, + 0.06930023431777954, + 0.05025109276175499, + 0.03715670108795166, + 0.05042092502117157 + ], + "credit_map": [ + 0.04554221034049988, + 0.03804426267743111, + 0.0851208046078682, + 0.04864870011806488, + 0.06417575478553772, + 0.05849284306168556, + 0.04224621504545212, + 0.06626337021589279, + 0.012311429716646671, + 0.012634007260203362, + 0.01762302964925766, + 0.018034307286143303, + 0.031202971935272217, + 0.0375356487929821, + 0.02595536969602108, + 0.05064389109611511, + 0.004076034761965275, + 0.002065049484372139, + 0.0032452403102070093, + 0.0024624604266136885, + 0.003289905609562993, + 0.008167635649442673, + 0.0036640821490436792, + 0.006665148306638002, + 0.005052007734775543, + 0.00258314935490489, + 0.005354474764317274, + 0.002703990088775754, + 0.004662507213652134, + 0.005242657382041216, + 0.0034164891112595797, + 0.005552852526307106, + 0.007631790358573198, + 0.00823985505849123, + 0.00447623897343874, + 0.005432022735476494, + 0.00602827500551939, + 0.004377517383545637, + 0.005639518145471811, + 0.005721759051084518, + 0.004357117228209972, + 0.003045282093808055, + 0.0030490264762192965, + 0.0021631617564707994, + 0.0019364377949386835, + 0.002684714738279581, + 0.002854367718100548, + 0.0025554357562214136, + 0.005714213941246271, + 0.004736703354865313, + 0.005452865269035101, + 0.007399365771561861, + 0.006320240907371044, + 0.019525542855262756, + 0.005542794708162546, + 0.01460850890725851, + 0.009725764393806458, + 0.0034660384990274906, + 0.005191834177821875, + 0.005677184090018272, + 0.0042627230286598206, + 0.010514847002923489, + 0.0037847724743187428, + 0.007627931423485279, + 0.004221545998007059, + 0.002381374128162861, + 0.0035173201467841864, + 0.0038852295838296413, + 0.0027882950380444527, + 0.00793728418648243, + 0.0044483402743935585, + 0.007977309636771679, + 0.0056486790999770164, + 0.0020218193531036377, + 0.00339552597142756, + 0.0035405505914241076, + 0.0030080974102020264, + 0.016409974545240402, + 0.003232290269806981, + 0.013163989409804344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.013594076037406921 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233664512634277 + ], + "opposite_logprob_tokens": [ + -0.8983664512634277 + ], + "logprob_reward": -0.5233664512634277, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233664512634277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11971069872379303, + 0.35222306847572327, + 0.07693596184253693, + 0.04945738613605499, + 0.03011259064078331, + 0.042466241866350174, + 0.08964918553829193, + 0.12547919154167175, + 0.05469216778874397, + 0.05927347391843796 + ], + "credit_map": [ + 0.0124834468588233, + 0.01408612634986639, + 0.015166865661740303, + 0.01843303069472313, + 0.01642824150621891, + 0.014769438654184341, + 0.009789126925170422, + 0.01855442114174366, + 0.022800059989094734, + 0.024140868335962296, + 0.028436820954084396, + 0.041832342743873596, + 0.06750360876321793, + 0.06311792880296707, + 0.041318297386169434, + 0.06307315826416016, + 0.011833957396447659, + 0.004474026616662741, + 0.013912864029407501, + 0.009914924390614033, + 0.009534628130495548, + 0.01204877533018589, + 0.007138263434171677, + 0.00807852391153574, + 0.006063191685825586, + 0.0034320715349167585, + 0.0064513199031353, + 0.005944490898400545, + 0.00518326461315155, + 0.010678150691092014, + 0.00477988924831152, + 0.006925004534423351, + 0.00325581943616271, + 0.0035983009729534388, + 0.002669890644028783, + 0.004533458035439253, + 0.0033141295425593853, + 0.003888173960149288, + 0.004364499822258949, + 0.004488320555537939, + 0.006886189337819815, + 0.0034142907243222, + 0.004342494532465935, + 0.006617641542106867, + 0.004124497529119253, + 0.006438671611249447, + 0.005568995140492916, + 0.005073461215943098, + 0.016117874532938004, + 0.007971496321260929, + 0.013112320564687252, + 0.011229348368942738, + 0.007438166998326778, + 0.014863611198961735, + 0.008163287304341793, + 0.010753076523542404, + 0.029141269624233246, + 0.009649768471717834, + 0.016887713223695755, + 0.014636889100074768, + 0.009928208775818348, + 0.019930364564061165, + 0.010377487167716026, + 0.014927489683032036, + 0.004981884732842445, + 0.003916851710528135, + 0.004972480237483978, + 0.005086829885840416, + 0.003980055917054415, + 0.014671090990304947, + 0.004864701069891453, + 0.012218271382153034, + 0.004973744973540306, + 0.0015179679030552506, + 0.003006021026521921, + 0.003555438481271267, + 0.0026474585756659508, + 0.022005852311849594, + 0.0032503202091902494, + 0.01831667125225067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.5233664512634277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071322485804558 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071322485804558, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071322485804558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14377263188362122, + 0.2670252323150635, + 0.08543860912322998, + 0.09325488656759262, + 0.048558078706264496, + 0.06108970567584038, + 0.08456781506538391, + 0.09177181124687195, + 0.04428788274526596, + 0.08023339509963989 + ], + "credit_map": [ + 0.0131880147382617, + 0.017345434054732323, + 0.015729116275906563, + 0.01322690024971962, + 0.026112625375390053, + 0.018721120432019234, + 0.020828375592827797, + 0.01862102746963501, + 0.020574551075696945, + 0.020465707406401634, + 0.02293548919260502, + 0.037276044487953186, + 0.04785175248980522, + 0.042774006724357605, + 0.03533928468823433, + 0.03980838134884834, + 0.009262737818062305, + 0.006567984353750944, + 0.008759739808738232, + 0.006471976172178984, + 0.01010879036039114, + 0.01759963482618332, + 0.0090361088514328, + 0.01763163134455681, + 0.01429654285311699, + 0.009765451774001122, + 0.00913750845938921, + 0.006691501475870609, + 0.02077295258641243, + 0.011664560995995998, + 0.01081441156566143, + 0.010111945681273937, + 0.005094201769679785, + 0.004942851606756449, + 0.0030746343545615673, + 0.0046734935604035854, + 0.00580603675916791, + 0.009866331703960896, + 0.004342232830822468, + 0.010758290998637676, + 0.004932156298309565, + 0.002812835620716214, + 0.006044973619282246, + 0.005343206692487001, + 0.0034542400389909744, + 0.01831306703388691, + 0.004306616261601448, + 0.015882601961493492, + 0.011418187990784645, + 0.006779415532946587, + 0.009875082410871983, + 0.011155106127262115, + 0.011205234564840794, + 0.013658590614795685, + 0.008252150379121304, + 0.012224042788147926, + 0.018178127706050873, + 0.00520318141207099, + 0.011086715385317802, + 0.009305894374847412, + 0.005105428863316774, + 0.018697870895266533, + 0.008464667946100235, + 0.015729909762740135, + 0.005016701761633158, + 0.002026646863669157, + 0.005704223643988371, + 0.00520935608074069, + 0.003805209416896105, + 0.010051297955214977, + 0.003636130830273032, + 0.008838311769068241, + 0.009792223572731018, + 0.004346185829490423, + 0.007456935942173004, + 0.007184283807873726, + 0.007299987133592367, + 0.019314153119921684, + 0.010215118527412415, + 0.014624495059251785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.011071322485804558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097988814115524 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.038097988814115524, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097988814115524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14200527966022491, + 0.17860375344753265, + 0.1124555766582489, + 0.0746031180024147, + 0.040778446942567825, + 0.0968894436955452, + 0.08495716005563736, + 0.0737476497888565, + 0.06981190294027328, + 0.1261475831270218 + ], + "credit_map": [ + 0.014588097110390663, + 0.015489562414586544, + 0.020988862961530685, + 0.022034641355276108, + 0.01674271747469902, + 0.019609646871685982, + 0.01494751125574112, + 0.01760426163673401, + 0.010726386681199074, + 0.010779343545436859, + 0.01354246586561203, + 0.015979880467057228, + 0.02596486173570156, + 0.03722527623176575, + 0.018527349457144737, + 0.04585821181535721, + 0.011200537905097008, + 0.007589178625494242, + 0.009764423593878746, + 0.009595313109457493, + 0.015151397325098515, + 0.029765784740447998, + 0.009870132431387901, + 0.019518820568919182, + 0.010847390629351139, + 0.006554015912115574, + 0.00825333222746849, + 0.00653404975309968, + 0.010686988942325115, + 0.01427281741052866, + 0.0069566466845571995, + 0.01049788761883974, + 0.005002160090953112, + 0.006524336524307728, + 0.004171112086623907, + 0.004711294546723366, + 0.006344649475067854, + 0.005367425736039877, + 0.004738559015095234, + 0.003918914590030909, + 0.005027448292821646, + 0.002929392736405134, + 0.006197407841682434, + 0.007190898060798645, + 0.004348551854491234, + 0.03581893816590309, + 0.005685936659574509, + 0.029690882191061974, + 0.010111045092344284, + 0.0058797006495296955, + 0.007816740311682224, + 0.011640011332929134, + 0.009677568450570107, + 0.01575697585940361, + 0.011228803545236588, + 0.012846320867538452, + 0.006262354087084532, + 0.003936694003641605, + 0.0076580955646932125, + 0.007239402737468481, + 0.0068735689856112, + 0.019845716655254364, + 0.00648563914000988, + 0.015446186996996403, + 0.009999452158808708, + 0.006873625796288252, + 0.011411852203309536, + 0.009520470164716244, + 0.007226814515888691, + 0.00784546323120594, + 0.00861238781362772, + 0.008321838453412056, + 0.012088137678802013, + 0.005232036113739014, + 0.007934358902275562, + 0.009767715819180012, + 0.00946899689733982, + 0.03789149969816208, + 0.012313541024923325, + 0.031451307237148285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.038097988814115524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an eye-level close-up of a black ant on a broad green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic, rain-swept environment with a fierce thunderstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sinister gray-white fungal stalk emerge from the junction between the ant's head and thorax?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start static and then perform a slow, deliberate, tight zoom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature stark, high-contrast shadows emphasizing the fibrous texture of the stalk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594940735027194 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.00594940735027194, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594940735027194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.271893709897995, + 0.11272615939378738, + 0.05393756553530693, + 0.05117550119757652, + 0.07333562523126602, + 0.17892736196517944, + 0.05060029402375221, + 0.1091136783361435, + 0.0491202175617218, + 0.049169864505529404 + ], + "credit_map": [ + 0.024580838158726692, + 0.025169091299176216, + 0.042026668787002563, + 0.03936874121427536, + 0.03889249265193939, + 0.029695529490709305, + 0.031789813190698624, + 0.04037056118249893, + 0.007990062236785889, + 0.010193523950874805, + 0.012097418308258057, + 0.009694656357169151, + 0.011290941387414932, + 0.018425386399030685, + 0.012073749676346779, + 0.03096041828393936, + 0.006735357455909252, + 0.004117532633244991, + 0.008642768487334251, + 0.004669724032282829, + 0.006941977888345718, + 0.010249319486320019, + 0.005042448174208403, + 0.007538438308984041, + 0.007362636271864176, + 0.003083476098254323, + 0.007747582159936428, + 0.004036766476929188, + 0.003816382260993123, + 0.011521229520440102, + 0.004954757168889046, + 0.008652671240270138, + 0.0074263750575482845, + 0.005673302803188562, + 0.008856046944856644, + 0.0046346294693648815, + 0.005318997427821159, + 0.019986947998404503, + 0.005427454598248005, + 0.01601187326014042, + 0.0198461152613163, + 0.00982077606022358, + 0.01662062108516693, + 0.013281509280204773, + 0.007903916761279106, + 0.053920548409223557, + 0.009517031721770763, + 0.04801684245467186, + 0.010313505306839943, + 0.0029646370094269514, + 0.004746503196656704, + 0.008926043286919594, + 0.003559273201972246, + 0.008576140739023685, + 0.005363220348954201, + 0.006150968838483095, + 0.020904941484332085, + 0.005426516756415367, + 0.012409115210175514, + 0.011878504417836666, + 0.010664820671081543, + 0.019567007198929787, + 0.011248855851590633, + 0.017013918608427048, + 0.004837362095713615, + 0.004660297185182571, + 0.004959681537002325, + 0.006980007514357567, + 0.005586524493992329, + 0.009431094862520695, + 0.005700268782675266, + 0.006964977364987135, + 0.01058436930179596, + 0.0035125447902828455, + 0.005677790381014347, + 0.005931687541306019, + 0.005689218174666166, + 0.0057540396228432655, + 0.005066647659987211, + 0.006953564938157797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.00594940735027194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_0/index_tensor([9254, 9254])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_0/index_tensor([9254, 9254])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b7a04625b68d68dd2b608434fba3466c7fc17ffb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_0/index_tensor([9254, 9254])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly zooming camera, from an upward angle, pans to a towering, mist-shrouded mountaintop. The rising sun casts a golden glow across the craggy rock. At the summit's center, a massive, ancient immortal stone stands silently, its surface covered in moss and weathered, appearing incredibly heavy and spiritual. Suddenly, the stone begins to vibrate slightly, and a dazzling golden light permeates its interior, rapidly spreading a web of cracks across its dark gray surface. Then, in a silent yet earth-shattering burst of energy, the stone suddenly explodes outward, sending countless fragments of varying sizes scattering in a hail of dust. As the dazzling golden light and swirling smoke from the center of the explosion dissipate, a stone monkey, draped in dazzling golden hair, emerges from the stone. Its hair gleams like molten gold, its muscles defined, and parts of its body still bear the rough texture of freshly hewn rock. It landed steadily on the broken stone foundation in a half-squatting posture full of primitive power, looking around at this new world curiously and defiantly. The most eye-catching thing was its deep eyes, which radiated two almost tangible, all-piercing golden beams, injecting a sacred and wild vitality into the entire grand and tranquil scene, announcing the birth of a legend.", + "seed": 15485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_0/index_9254_seed_15485_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04824393914313987, + "vlm_logprob": -0.04824393914313987, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.04824393914313987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2210049331188202 + ], + "opposite_logprob_tokens": [ + -1.7210049629211426 + ], + "logprob_reward": -0.2210049331188202, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2210049331188202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20562966167926788, + 0.12284282594919205, + 0.07902086526155472, + 0.10298028588294983, + 0.08303102850914001, + 0.0963243842124939, + 0.08094484359025955, + 0.08455417305231094, + 0.06795244663953781, + 0.07671956717967987 + ], + "credit_map": [ + 0.019041528925299644, + 0.043798647820949554, + 0.023132801055908203, + 0.01634596288204193, + 0.028843538835644722, + 0.031799789518117905, + 0.023588279262185097, + 0.019079113379120827, + 0.01810884103178978, + 0.01438921969383955, + 0.01531876903027296, + 0.010085880756378174, + 0.017238426953554153, + 0.01662018895149231, + 0.01763145811855793, + 0.01345005352050066, + 0.008687831461429596, + 0.01449387613683939, + 0.008321963250637054, + 0.008526989258825779, + 0.011999713256955147, + 0.009372053667902946, + 0.010039009153842926, + 0.007579420693218708, + 0.01796743832528591, + 0.012331770732998848, + 0.017188575118780136, + 0.009979375638067722, + 0.011139431968331337, + 0.013284585438668728, + 0.013125116936862469, + 0.007963987998664379, + 0.010785798542201519, + 0.013866878114640713, + 0.012442522682249546, + 0.007716722320765257, + 0.014176327735185623, + 0.011184616014361382, + 0.007981681264936924, + 0.004876482300460339, + 0.01227941457182169, + 0.011050362139940262, + 0.015469088219106197, + 0.011686631478369236, + 0.012984002940356731, + 0.00897953286767006, + 0.014358560554683208, + 0.009516787715256214, + 0.011590241454541683, + 0.006861999165266752, + 0.010610769502818584, + 0.009967982769012451, + 0.008419809862971306, + 0.011744323186576366, + 0.01575651206076145, + 0.005993206519633532, + 0.016598964110016823, + 0.009574460797011852, + 0.012252178974449635, + 0.007125889416784048, + 0.009756444953382015, + 0.00989423505961895, + 0.009385008364915848, + 0.009966987185180187, + 0.007846803404390812, + 0.005449352785944939, + 0.01437543798238039, + 0.007722493261098862, + 0.006823337636888027, + 0.01029359083622694, + 0.007050598971545696, + 0.008390829898416996, + 0.00676349364221096, + 0.007989930920302868, + 0.016551537439227104, + 0.0059592691250145435, + 0.00775487395003438, + 0.012922964058816433, + 0.009778611361980438, + 0.008998888544738293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.2210049331188202 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00465927179902792 + ], + "opposite_logprob_tokens": [ + -5.379659175872803 + ], + "logprob_reward": -0.00465927179902792, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00465927179902792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20041829347610474, + 0.10454799979925156, + 0.07453285157680511, + 0.1518019586801529, + 0.03271670266985893, + 0.023628264665603638, + 0.11410441994667053, + 0.07128279656171799, + 0.17800037562847137, + 0.048966385424137115 + ], + "credit_map": [ + 0.013576998375356197, + 0.02999124862253666, + 0.015052503906190395, + 0.01083792932331562, + 0.03132336586713791, + 0.04125356674194336, + 0.032058458775281906, + 0.026324212551116943, + 0.008460409939289093, + 0.015234760008752346, + 0.011614909395575523, + 0.0072366585955023766, + 0.018247876316308975, + 0.014975066296756268, + 0.018550708889961243, + 0.010227610357105732, + 0.007825106382369995, + 0.013423318974673748, + 0.00764481071382761, + 0.007162750232964754, + 0.013143602758646011, + 0.008537297137081623, + 0.009068247862160206, + 0.007727714721113443, + 0.021177886053919792, + 0.03168746829032898, + 0.014178849756717682, + 0.018207358196377754, + 0.020972849801182747, + 0.017261328175663948, + 0.017776859924197197, + 0.010539354756474495, + 0.004267983138561249, + 0.00474898423999548, + 0.0031544954981654882, + 0.0029701374005526304, + 0.005262234713882208, + 0.005320700351148844, + 0.004559588618576527, + 0.0024325810372829437, + 0.0027561236638575792, + 0.003136603394523263, + 0.0025166531559079885, + 0.002899585058912635, + 0.0040094987489283085, + 0.0019566596020013094, + 0.0032383273355662823, + 0.0031148134730756283, + 0.012301945127546787, + 0.008382833562791348, + 0.012094366364181042, + 0.013252314180135727, + 0.011432508938014507, + 0.014551573432981968, + 0.025605373084545135, + 0.016483504325151443, + 0.012239675037562847, + 0.004692866466939449, + 0.011682250536978245, + 0.006970537360757589, + 0.007756974082440138, + 0.008680703118443489, + 0.011220659129321575, + 0.008039130829274654, + 0.015626607462763786, + 0.01387419831007719, + 0.04618591442704201, + 0.016223974525928497, + 0.01612132601439953, + 0.031187845394015312, + 0.02505328133702278, + 0.013727240264415741, + 0.0034695102367550135, + 0.0056488183327019215, + 0.00900321640074253, + 0.004422144033014774, + 0.007333533372730017, + 0.007218915969133377, + 0.007027267012745142, + 0.004842979833483696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.00465927179902792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002206512028351426 + ], + "opposite_logprob_tokens": [ + -6.127206325531006 + ], + "logprob_reward": -0.002206512028351426, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002206512028351426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19011488556861877, + 0.13816407322883606, + 0.07774228602647781, + 0.06882934272289276, + 0.12987585365772247, + 0.061771802604198456, + 0.07203423976898193, + 0.11497008055448532, + 0.09540294855833054, + 0.05109439045190811 + ], + "credit_map": [ + 0.012865527532994747, + 0.023006483912467957, + 0.01291618775576353, + 0.010102219879627228, + 0.030772658064961433, + 0.04144614562392235, + 0.0291043259203434, + 0.029901353642344475, + 0.013046876527369022, + 0.015236889943480492, + 0.018972989171743393, + 0.01556160394102335, + 0.018232595175504684, + 0.01855343207716942, + 0.02117467299103737, + 0.01738504320383072, + 0.008299235254526138, + 0.014058963395655155, + 0.006452504079788923, + 0.008786669000983238, + 0.013975933194160461, + 0.0100088557228446, + 0.009100495837628841, + 0.0070596425794065, + 0.007887518033385277, + 0.012318441644310951, + 0.008790230378508568, + 0.008139961399137974, + 0.008161215111613274, + 0.009387921541929245, + 0.007806533016264439, + 0.006337529513984919, + 0.012795811519026756, + 0.027751725167036057, + 0.010569472797214985, + 0.010909981094300747, + 0.01841561309993267, + 0.02437872812151909, + 0.014714464545249939, + 0.010340064764022827, + 0.009937208145856857, + 0.010678953491151333, + 0.006276728585362434, + 0.0069852713495492935, + 0.007134057115763426, + 0.006250514183193445, + 0.006946134380996227, + 0.007562947925180197, + 0.006635925732553005, + 0.007885311730206013, + 0.007442524656653404, + 0.009120899252593517, + 0.012666864320635796, + 0.009481552988290787, + 0.012911585159599781, + 0.005889586638659239, + 0.014686597511172295, + 0.006989112589508295, + 0.019735613837838173, + 0.007427423261106014, + 0.01006445474922657, + 0.01746683567762375, + 0.011066848412156105, + 0.027533210813999176, + 0.009088875725865364, + 0.006828344892710447, + 0.026529794558882713, + 0.006442073732614517, + 0.00849026720970869, + 0.01793682388961315, + 0.012438848614692688, + 0.007647937163710594, + 0.004440599586814642, + 0.006029301322996616, + 0.009818604215979576, + 0.004598811734467745, + 0.006668393034487963, + 0.006768734194338322, + 0.006877623498439789, + 0.005892329849302769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002206512028351426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008132726070471108 + ], + "opposite_logprob_tokens": [ + -7.1258134841918945 + ], + "logprob_reward": -0.0008132726070471108, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008132726070471108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12112110108137131, + 0.09861908853054047, + 0.0439419262111187, + 0.14924897253513336, + 0.0726543664932251, + 0.09478992968797684, + 0.06786259263753891, + 0.1299157291650772, + 0.1570693850517273, + 0.06477688252925873 + ], + "credit_map": [ + 0.00857471115887165, + 0.020721158012747765, + 0.00928756594657898, + 0.010260985232889652, + 0.01829143613576889, + 0.021677177399396896, + 0.018705859780311584, + 0.013602202758193016, + 0.008851459249854088, + 0.013905910775065422, + 0.009045884013175964, + 0.008108746260404587, + 0.016669264063239098, + 0.014318455941975117, + 0.015992680564522743, + 0.011726693250238895, + 0.004729968961328268, + 0.00886419229209423, + 0.004212349187582731, + 0.004301474429666996, + 0.00782843679189682, + 0.005432566627860069, + 0.004292609635740519, + 0.004280329681932926, + 0.020583996549248695, + 0.028514796867966652, + 0.019874464720487595, + 0.02005101554095745, + 0.01954270340502262, + 0.014040919952094555, + 0.017640618607401848, + 0.009000470861792564, + 0.006536576431244612, + 0.012026836164295673, + 0.0053462316282093525, + 0.007960026152431965, + 0.009480476379394531, + 0.011446177028119564, + 0.01316943857818842, + 0.006688605062663555, + 0.012858227826654911, + 0.01213468424975872, + 0.009414886124432087, + 0.012720867991447449, + 0.012227507308125496, + 0.00962925050407648, + 0.011574657633900642, + 0.014229845255613327, + 0.00952984020113945, + 0.008665960282087326, + 0.008442766964435577, + 0.007030135486274958, + 0.008860258385539055, + 0.0080734733492136, + 0.00960947759449482, + 0.007650678511708975, + 0.011990569531917572, + 0.012826421298086643, + 0.023623673245310783, + 0.014874933287501335, + 0.019815681502223015, + 0.014999732375144958, + 0.018390290439128876, + 0.013394431211054325, + 0.014279861934483051, + 0.01024628896266222, + 0.04176444187760353, + 0.013644540682435036, + 0.015187353827059269, + 0.027446644380688667, + 0.022566763684153557, + 0.011933487839996815, + 0.0050444346852600574, + 0.0073205167427659035, + 0.014088050462305546, + 0.005358374677598476, + 0.007832330651581287, + 0.010302936658263206, + 0.008245040662586689, + 0.006585196126252413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.0008132726070471108 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15287883579730988, + 0.06460227817296982, + 0.04618154093623161, + 0.08452313393354416, + 0.09915776550769806, + 0.1689871996641159, + 0.06218072399497032, + 0.1660591959953308, + 0.12808547914028168, + 0.027343817055225372 + ], + "credit_map": [ + 0.013071238994598389, + 0.02343437448143959, + 0.015313912183046341, + 0.00992488395422697, + 0.022368069738149643, + 0.026943422853946686, + 0.026009365916252136, + 0.015813570469617844, + 0.006245086435228586, + 0.006712484173476696, + 0.007836026139557362, + 0.00629675667732954, + 0.011124765500426292, + 0.008224086835980415, + 0.008557628840208054, + 0.009605443105101585, + 0.007532249670475721, + 0.005262686405330896, + 0.005600482225418091, + 0.004593913909047842, + 0.005387432407587767, + 0.00641721161082387, + 0.004599901847541332, + 0.006787664722651243, + 0.009197734296321869, + 0.013438655063509941, + 0.011805936694145203, + 0.011702697724103928, + 0.010907081887125969, + 0.009407862089574337, + 0.010599199682474136, + 0.007463972549885511, + 0.010813900269567966, + 0.01401599496603012, + 0.011300737969577312, + 0.011554297991096973, + 0.012257478199899197, + 0.015326142311096191, + 0.01576680690050125, + 0.008122404105961323, + 0.022223707288503647, + 0.01843186467885971, + 0.016079625114798546, + 0.021236423403024673, + 0.019713424146175385, + 0.017418401315808296, + 0.023717349395155907, + 0.03016640804708004, + 0.008251885883510113, + 0.004516627173870802, + 0.008029312826693058, + 0.00689029460772872, + 0.005825694650411606, + 0.007352628745138645, + 0.01237722672522068, + 0.008937051519751549, + 0.02557489648461342, + 0.016559336334466934, + 0.02307315543293953, + 0.015238669700920582, + 0.02008492313325405, + 0.019764995202422142, + 0.022110190242528915, + 0.02365303970873356, + 0.014137597754597664, + 0.009292139671742916, + 0.03686201944947243, + 0.010278451256453991, + 0.011167848482728004, + 0.02010936662554741, + 0.011519904248416424, + 0.014718158170580864, + 0.0022331068757921457, + 0.001999853877350688, + 0.004078006837517023, + 0.0034079430624842644, + 0.0040095713920891285, + 0.0036927941255271435, + 0.004224428907036781, + 0.003698111744597554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.012535706162452698 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming camera, from an upward angle, pans to a towering, mist-shrouded mountaintop. The rising sun casts a golden glow across the craggy rock. At the summit's center, a massive, ancient immortal stone stands silently, its surface covered in moss and weathered, appearing incredibly heavy and spiritual. Suddenly, the stone begins to vibrate slightly, and a dazzling golden light permeates its interior, rapidly spreading a web of cracks across its dark gray surface. Then, in a silent yet earth-shattering burst of energy, the stone suddenly explodes outward, sending countless fragments of varying sizes scattering in a hail of dust. As the dazzling golden light and swirling smoke from the center of the explosion dissipate, a stone monkey, draped in dazzling golden hair, emerges from the stone. Its hair gleams like molten gold, its muscles defined, and parts of its body still bear the rough texture of freshly hewn rock. It landed steadily on the broken stone foundation in a half-squatting posture full of primitive power, looking around at this new world curiously and defiantly. The most eye-catching thing was its deep eyes, which radiated two almost tangible, all-piercing golden beams, injecting a sacred and wild vitality into the entire grand and tranquil scene, announcing the birth of a legend.", + "seed": 15486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_0/index_9254_seed_15486_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.034640150726772845, + "vlm_logprob": -0.034640150726772845, + "vlm_yes_no_margin": 4.649999976158142, + "vlm_token_credit": -0.034640150726772845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14710482954978943 + ], + "opposite_logprob_tokens": [ + -2.1471047401428223 + ], + "logprob_reward": -0.14710482954978943, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.14710482954978943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1892845630645752, + 0.07210949808359146, + 0.05688716471195221, + 0.09486689418554306, + 0.02884923666715622, + 0.08437216281890869, + 0.3894769251346588, + 0.023477571085095406, + 0.03795836865901947, + 0.022717563435435295 + ], + "credit_map": [ + 0.02815874107182026, + 0.021355414763092995, + 0.02293340675532818, + 0.021271949633955956, + 0.026239493861794472, + 0.02109784260392189, + 0.025219466537237167, + 0.02300826646387577, + 0.007168986368924379, + 0.007921467535197735, + 0.010233788751065731, + 0.011711135506629944, + 0.011135770007967949, + 0.006845390889793634, + 0.007331553380936384, + 0.009761410765349865, + 0.007542256265878677, + 0.008668603375554085, + 0.010589556768536568, + 0.008627419359982014, + 0.00648836512118578, + 0.005774950608611107, + 0.0045099821873009205, + 0.004686032421886921, + 0.009660539217293262, + 0.00811777450144291, + 0.010420759208500385, + 0.005810926202684641, + 0.017004679888486862, + 0.013322803191840649, + 0.007810276933014393, + 0.022719135507941246, + 0.005316951312124729, + 0.004392771981656551, + 0.004829433746635914, + 0.00241225422360003, + 0.0035073449835181236, + 0.0032278888393193483, + 0.0027673221193253994, + 0.002395271323621273, + 0.01528793852776289, + 0.009190228767693043, + 0.009240678511559963, + 0.006885280832648277, + 0.008336352184414864, + 0.010451513342559338, + 0.005880218464881182, + 0.01909996196627617, + 0.04399251565337181, + 0.06692856550216675, + 0.035927917808294296, + 0.050792526453733444, + 0.06966787576675415, + 0.04078793153166771, + 0.05378131568431854, + 0.027598295360803604, + 0.0019504574593156576, + 0.003792101051658392, + 0.002351770643144846, + 0.00274668144993484, + 0.0048341103829443455, + 0.0030710087157785892, + 0.0026359413750469685, + 0.002095501869916916, + 0.0035361729096621275, + 0.003153890836983919, + 0.0051566665060818195, + 0.004185566678643227, + 0.0063108052127063274, + 0.004872372839599848, + 0.006162707228213549, + 0.0045801918022334576, + 0.0030471631325781345, + 0.0020031125750392675, + 0.0033828280866146088, + 0.0019426007056608796, + 0.0021389543544501066, + 0.004644503351300955, + 0.002936589065939188, + 0.0026218139100819826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.14710482954978943 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764752272516489 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.00764752272516489, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764752272516489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1486748456954956, + 0.09951959550380707, + 0.1299905925989151, + 0.07090111076831818, + 0.09190389513969421, + 0.13372868299484253, + 0.19168801605701447, + 0.022203974425792694, + 0.060069188475608826, + 0.05132005363702774 + ], + "credit_map": [ + 0.01798570714890957, + 0.016799863427877426, + 0.01713850535452366, + 0.01962887868285179, + 0.020008793100714684, + 0.017057470977306366, + 0.01612776145339012, + 0.023927873000502586, + 0.00743258697912097, + 0.013364966958761215, + 0.015360564924776554, + 0.01034361869096756, + 0.015949485823512077, + 0.01619631052017212, + 0.010259533300995827, + 0.010612529702484608, + 0.015854984521865845, + 0.015533814206719398, + 0.023816173896193504, + 0.01810654252767563, + 0.012664219364523888, + 0.017645571380853653, + 0.011939654126763344, + 0.014429633505642414, + 0.008556569926440716, + 0.01181300263851881, + 0.008099065162241459, + 0.008064060471951962, + 0.010152650997042656, + 0.012364601716399193, + 0.005782437045127153, + 0.006068718153983355, + 0.01326123159378767, + 0.01039278693497181, + 0.0161503404378891, + 0.008626807481050491, + 0.013856161385774612, + 0.012595679610967636, + 0.009239282459020615, + 0.007781602907925844, + 0.029311327263712883, + 0.010373186320066452, + 0.009978115558624268, + 0.006707255262881517, + 0.00872715562582016, + 0.017709938809275627, + 0.0076221888884902, + 0.043299511075019836, + 0.021939696744084358, + 0.0305646825581789, + 0.016555393114686012, + 0.025306059047579765, + 0.034551866352558136, + 0.022824933752417564, + 0.025975193828344345, + 0.013970186933875084, + 0.0018331357277929783, + 0.0032596916425973177, + 0.0023824991658329964, + 0.002636977471411228, + 0.004247983917593956, + 0.0032252129167318344, + 0.0027052105870097876, + 0.001913263346068561, + 0.007170318625867367, + 0.006771903019398451, + 0.011098183691501617, + 0.004295121878385544, + 0.006576193030923605, + 0.010179860517382622, + 0.005747102200984955, + 0.008230502717196941, + 0.005578577518463135, + 0.005765637848526239, + 0.006347381044179201, + 0.004729576874524355, + 0.005490528419613838, + 0.01173870638012886, + 0.005761885084211826, + 0.005907760001718998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.00764752272516489 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632262073457241 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632262073457241, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632262073457241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09292047470808029, + 0.09194906800985336, + 0.06549837440252304, + 0.08051729947328568, + 0.3005184233188629, + 0.07155971974134445, + 0.1400187760591507, + 0.04260130226612091, + 0.053713779896497726, + 0.06070282682776451 + ], + "credit_map": [ + 0.012791529297828674, + 0.014805846847593784, + 0.01166249718517065, + 0.010261029936373234, + 0.013139649294316769, + 0.011477581225335598, + 0.009720360860228539, + 0.009061978198587894, + 0.007659705355763435, + 0.00797189585864544, + 0.008951176889240742, + 0.012005356140434742, + 0.017373153939843178, + 0.013126367703080177, + 0.011218645609915257, + 0.013642764650285244, + 0.01180489081889391, + 0.007773511577397585, + 0.009310112334787846, + 0.008421235717833042, + 0.006819830741733313, + 0.00842170137912035, + 0.0063096098601818085, + 0.0066374861635267735, + 0.008440990932285786, + 0.01623520813882351, + 0.007432069629430771, + 0.007993240840733051, + 0.013155089691281319, + 0.011236065998673439, + 0.005535589065402746, + 0.01048904750496149, + 0.042885273694992065, + 0.07047482579946518, + 0.020953970029950142, + 0.013712470419704914, + 0.05074531212449074, + 0.02336975932121277, + 0.021456381306052208, + 0.05692045018076897, + 0.012963185086846352, + 0.006214095279574394, + 0.007569744251668453, + 0.004821610637009144, + 0.005544840823858976, + 0.01026330329477787, + 0.0049326661974191666, + 0.019250277429819107, + 0.01744862273335457, + 0.021973710507154465, + 0.013046610169112682, + 0.017728958278894424, + 0.021594787016510963, + 0.02105620503425598, + 0.017442015931010246, + 0.009727870114147663, + 0.004225979093462229, + 0.006976299919188023, + 0.006563351955264807, + 0.004329996649175882, + 0.004272625781595707, + 0.007149246986955404, + 0.005043758079409599, + 0.004040044732391834, + 0.006929181050509214, + 0.005608659237623215, + 0.0086860666051507, + 0.006184104830026627, + 0.006648642476648092, + 0.00754734268411994, + 0.005191342439502478, + 0.006918441504240036, + 0.004385880194604397, + 0.005637094844132662, + 0.0077436831779778, + 0.008547114208340645, + 0.008884979411959648, + 0.010635594837367535, + 0.007248864509165287, + 0.007619611453264952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.007632262073457241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010408704401925206 + ], + "opposite_logprob_tokens": [ + -6.876040935516357 + ], + "logprob_reward": -0.0010408704401925206, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010408704401925206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11117953807115555, + 0.08412017673254013, + 0.05458853393793106, + 0.07825274765491486, + 0.10326613485813141, + 0.04782668873667717, + 0.20812396705150604, + 0.07463369518518448, + 0.14679522812366486, + 0.09121324867010117 + ], + "credit_map": [ + 0.012351554818451405, + 0.014026185497641563, + 0.013275191187858582, + 0.014577656053006649, + 0.014835990965366364, + 0.011020206846296787, + 0.015691356733441353, + 0.01540139690041542, + 0.007458825130015612, + 0.007252884097397327, + 0.009387128986418247, + 0.012408357113599777, + 0.013319219462573528, + 0.0121259531006217, + 0.011160964146256447, + 0.011006851680576801, + 0.005915040150284767, + 0.0073852743953466415, + 0.008863582275807858, + 0.008901312947273254, + 0.006341500207781792, + 0.006211928557604551, + 0.004887789953500032, + 0.006082105450332165, + 0.008113667368888855, + 0.008177562616765499, + 0.00733485072851181, + 0.0056421649642288685, + 0.015534602105617523, + 0.008415463380515575, + 0.006122651044279337, + 0.01891178824007511, + 0.013495994731783867, + 0.026607144623994827, + 0.0099876020103693, + 0.005770707502961159, + 0.013634078204631805, + 0.008784826844930649, + 0.006182357203215361, + 0.018803419545292854, + 0.005333429668098688, + 0.005807452369481325, + 0.003990644123405218, + 0.0045707812532782555, + 0.006549286190420389, + 0.006289422977715731, + 0.006306302733719349, + 0.008979367092251778, + 0.02289520762860775, + 0.03431254252791405, + 0.0189005546271801, + 0.027971066534519196, + 0.03690772503614426, + 0.02111811935901642, + 0.030719079077243805, + 0.015299675054848194, + 0.005744350608438253, + 0.012592867016792297, + 0.0075854468159377575, + 0.007531919050961733, + 0.0168638676404953, + 0.010627097450196743, + 0.008026902563869953, + 0.005661248695105314, + 0.010620267130434513, + 0.016019372269511223, + 0.01823340356349945, + 0.01818062737584114, + 0.020658981055021286, + 0.02037176676094532, + 0.021547064185142517, + 0.02116374485194683, + 0.007282445207238197, + 0.006315510254353285, + 0.008175618015229702, + 0.01542375423014164, + 0.01595534197986126, + 0.012604553252458572, + 0.015315419994294643, + 0.01014060527086258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.0010408704401925206 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775268845260143 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775268845260143, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775268845260143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21353276073932648, + 0.0809064656496048, + 0.048720888793468475, + 0.07935741543769836, + 0.06371917575597763, + 0.0656922310590744, + 0.12880687415599823, + 0.11050151288509369, + 0.12183119356632233, + 0.08693143725395203 + ], + "credit_map": [ + 0.029330097138881683, + 0.022485820576548576, + 0.02563983015716076, + 0.02709866687655449, + 0.029028812423348427, + 0.019828688353300095, + 0.029851624742150307, + 0.030269235372543335, + 0.007179225794970989, + 0.007391320541501045, + 0.007087256293743849, + 0.0126343984156847, + 0.013281434774398804, + 0.01034194603562355, + 0.011794444173574448, + 0.011196441017091274, + 0.006129714660346508, + 0.005912735126912594, + 0.007683953270316124, + 0.007066249381750822, + 0.0055412775836884975, + 0.007180757820606232, + 0.004728148225694895, + 0.004478048998862505, + 0.009196126833558083, + 0.008361362852156162, + 0.008566193282604218, + 0.0056954738683998585, + 0.012763277627527714, + 0.009877972304821014, + 0.005993853788822889, + 0.018903151154518127, + 0.010167489759624004, + 0.006410946138203144, + 0.00648962426930666, + 0.006292321253567934, + 0.009455201216042042, + 0.007522457744926214, + 0.007672509178519249, + 0.009708622470498085, + 0.013285626657307148, + 0.005867767613381147, + 0.006476887967437506, + 0.005137674510478973, + 0.00540228420868516, + 0.009410240687429905, + 0.005253258161246777, + 0.014858491718769073, + 0.013181553222239017, + 0.015058599412441254, + 0.010867014527320862, + 0.01552274078130722, + 0.028780553489923477, + 0.018634017556905746, + 0.01701853610575199, + 0.00974385254085064, + 0.010383802466094494, + 0.014344587922096252, + 0.013104391284286976, + 0.011774377897381783, + 0.01900986209511757, + 0.016968417912721634, + 0.015242278575897217, + 0.009673800319433212, + 0.01081883069127798, + 0.015086968429386616, + 0.018271543085575104, + 0.01433450449258089, + 0.013625655323266983, + 0.017169322818517685, + 0.013284790329635143, + 0.019239576533436775, + 0.008328400552272797, + 0.008647242560982704, + 0.011092949658632278, + 0.00902643334120512, + 0.015781434252858162, + 0.01587465964257717, + 0.008791730739176273, + 0.009388583712279797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.009775268845260143 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming camera, from an upward angle, pans to a towering, mist-shrouded mountaintop. The rising sun casts a golden glow across the craggy rock. At the summit's center, a massive, ancient immortal stone stands silently, its surface covered in moss and weathered, appearing incredibly heavy and spiritual. Suddenly, the stone begins to vibrate slightly, and a dazzling golden light permeates its interior, rapidly spreading a web of cracks across its dark gray surface. Then, in a silent yet earth-shattering burst of energy, the stone suddenly explodes outward, sending countless fragments of varying sizes scattering in a hail of dust. As the dazzling golden light and swirling smoke from the center of the explosion dissipate, a stone monkey, draped in dazzling golden hair, emerges from the stone. Its hair gleams like molten gold, its muscles defined, and parts of its body still bear the rough texture of freshly hewn rock. It landed steadily on the broken stone foundation in a half-squatting posture full of primitive power, looking around at this new world curiously and defiantly. The most eye-catching thing was its deep eyes, which radiated two almost tangible, all-piercing golden beams, injecting a sacred and wild vitality into the entire grand and tranquil scene, announcing the birth of a legend.", + "seed": 15487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_0/index_9254_seed_15487_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025520227709785105, + "vlm_logprob": -0.025520227709785105, + "vlm_yes_no_margin": 5.325, + "vlm_token_credit": -0.025520227709785105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1143777146935463 + ], + "opposite_logprob_tokens": [ + -2.364377737045288 + ], + "logprob_reward": -0.1143777146935463, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1143777146935463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14870324730873108, + 0.05920254439115524, + 0.042265407741069794, + 0.06702612340450287, + 0.0805891677737236, + 0.06459388881921768, + 0.25919726490974426, + 0.12053787708282471, + 0.08741171658039093, + 0.07047271728515625 + ], + "credit_map": [ + 0.02952282875776291, + 0.01773637719452381, + 0.017109448090195656, + 0.014284609816968441, + 0.02287682145833969, + 0.018703317269682884, + 0.01604236476123333, + 0.012427467852830887, + 0.007409632671624422, + 0.0057034967467188835, + 0.005243416875600815, + 0.005416082683950663, + 0.013676977716386318, + 0.007809009402990341, + 0.008043032139539719, + 0.005900896154344082, + 0.004160598386079073, + 0.004544990602880716, + 0.003909120336174965, + 0.004087756387889385, + 0.007546364329755306, + 0.004874869715422392, + 0.007217059377580881, + 0.005924649070948362, + 0.00816921703517437, + 0.009978709742426872, + 0.0056157829239964485, + 0.008711550384759903, + 0.00877333339303732, + 0.007118516601622105, + 0.0101590845733881, + 0.008499931544065475, + 0.012155083939433098, + 0.0076185232028365135, + 0.011473594233393669, + 0.00855934526771307, + 0.007408908102661371, + 0.01021640282124281, + 0.009808878414332867, + 0.013348432257771492, + 0.01262036245316267, + 0.008258000016212463, + 0.006567113567143679, + 0.006305753719061613, + 0.008549603633582592, + 0.006354163400828838, + 0.010668640024960041, + 0.005270255729556084, + 0.04483554884791374, + 0.04018077999353409, + 0.024844292551279068, + 0.02758762054145336, + 0.04791226610541344, + 0.022718314081430435, + 0.02998943068087101, + 0.021129030734300613, + 0.016142919659614563, + 0.013935490511357784, + 0.014247522689402103, + 0.019097747281193733, + 0.011126314289867878, + 0.02074487693607807, + 0.009286224842071533, + 0.015956779941916466, + 0.013083918020129204, + 0.013128492049872875, + 0.0076725599355995655, + 0.0064404611475765705, + 0.009591429494321346, + 0.02013392187654972, + 0.007860710844397545, + 0.009500234387814999, + 0.007883173413574696, + 0.00918656401336193, + 0.008844847790896893, + 0.0070035336539149284, + 0.007662009447813034, + 0.00799567811191082, + 0.010424536652863026, + 0.01147237978875637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.1143777146935463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002837562235072255 + ], + "opposite_logprob_tokens": [ + -5.877837657928467 + ], + "logprob_reward": -0.002837562235072255, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002837562235072255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19062519073486328, + 0.1591493934392929, + 0.06531134992837906, + 0.02794620953500271, + 0.05775744095444679, + 0.07037975639104843, + 0.1957678198814392, + 0.0957021564245224, + 0.023158570751547813, + 0.1142021045088768 + ], + "credit_map": [ + 0.03486168384552002, + 0.020514274016022682, + 0.019059574231505394, + 0.02126125805079937, + 0.0274016372859478, + 0.01989126019179821, + 0.028729140758514404, + 0.018906369805336, + 0.012446070089936256, + 0.01708320342004299, + 0.009317999705672264, + 0.02078274078667164, + 0.03212031349539757, + 0.020459352061152458, + 0.030327243730425835, + 0.016612466424703598, + 0.004400522448122501, + 0.007998824119567871, + 0.0062184082344174385, + 0.008627147413790226, + 0.010387396439909935, + 0.011606662534177303, + 0.008197295479476452, + 0.00787508673965931, + 0.0021178298629820347, + 0.004304037429392338, + 0.0021570809185504913, + 0.0046725827269256115, + 0.003145986469462514, + 0.0028646369464695454, + 0.00393924955278635, + 0.004744805861264467, + 0.006720544770359993, + 0.00818302109837532, + 0.006842302158474922, + 0.008251002989709377, + 0.005411408841609955, + 0.008610662072896957, + 0.007267667446285486, + 0.006470832508057356, + 0.012240424752235413, + 0.010682931169867516, + 0.006187259219586849, + 0.008379893377423286, + 0.01025202963501215, + 0.006413733586668968, + 0.01039497833698988, + 0.005828502122312784, + 0.034393806010484695, + 0.029438132420182228, + 0.01916334219276905, + 0.019384104758501053, + 0.03201960399746895, + 0.023034777492284775, + 0.021566199138760567, + 0.016767870634794235, + 0.012179560959339142, + 0.010996156372129917, + 0.008194590918719769, + 0.021062208339571953, + 0.00921702105551958, + 0.014960614033043385, + 0.007888277061283588, + 0.011203726753592491, + 0.0028898403979837894, + 0.002112970920279622, + 0.00349108362570405, + 0.002546265721321106, + 0.0023707111831754446, + 0.004952212329953909, + 0.002487828256562352, + 0.002307660412043333, + 0.01449569407850504, + 0.00967512559145689, + 0.015016978606581688, + 0.01049837563186884, + 0.010313932783901691, + 0.019927693530917168, + 0.011369075626134872, + 0.022905224934220314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.002837562235072255 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002202229807153344 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.002202229807153344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002202229807153344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1348467618227005, + 0.07065196335315704, + 0.03541737422347069, + 0.04742666706442833, + 0.0727100819349289, + 0.0943499207496643, + 0.20384536683559418, + 0.15284134447574615, + 0.11558657884597778, + 0.07232391089200974 + ], + "credit_map": [ + 0.020387746393680573, + 0.016377894207835197, + 0.014495717361569405, + 0.013635573908686638, + 0.020938390865921974, + 0.016152050346136093, + 0.014696124941110611, + 0.01816326193511486, + 0.00613447604700923, + 0.0091109499335289, + 0.005148136056959629, + 0.008376005105674267, + 0.010622866451740265, + 0.010361974127590656, + 0.013192178681492805, + 0.007705372292548418, + 0.0026258996222168207, + 0.0044938744977116585, + 0.002888857387006283, + 0.0037484816275537014, + 0.006662404630333185, + 0.006018100306391716, + 0.004688922315835953, + 0.004290835466235876, + 0.004476496018469334, + 0.007020008750259876, + 0.0034750858321785927, + 0.005787595175206661, + 0.008746867999434471, + 0.004853368271142244, + 0.006073326803743839, + 0.006993914023041725, + 0.008295978419482708, + 0.008486500009894371, + 0.007247744593769312, + 0.0090776477009058, + 0.008442692458629608, + 0.011137774214148521, + 0.008563986048102379, + 0.01145776268094778, + 0.015118815936148167, + 0.01077545527368784, + 0.008235097862780094, + 0.008344179019331932, + 0.015537395142018795, + 0.011730143800377846, + 0.010156658478081226, + 0.01445217989385128, + 0.03830290585756302, + 0.0324007086455822, + 0.017761921510100365, + 0.017354436218738556, + 0.03141256794333458, + 0.027692480012774467, + 0.02236909419298172, + 0.016551237553358078, + 0.02093506045639515, + 0.015659328550100327, + 0.01609068177640438, + 0.030197743326425552, + 0.020249638706445694, + 0.01647757738828659, + 0.016704333946108818, + 0.016526972874999046, + 0.01461122464388609, + 0.010574365966022015, + 0.018877819180488586, + 0.008547233417630196, + 0.01253780908882618, + 0.02029293216764927, + 0.009635131806135178, + 0.020510056987404823, + 0.010216089896857738, + 0.006913809105753899, + 0.008599954657256603, + 0.005928076338022947, + 0.009430639445781708, + 0.011052707210183144, + 0.007817101664841175, + 0.012365525588393211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.002202229807153344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005597693379968405 + ], + "opposite_logprob_tokens": [ + -7.5005598068237305 + ], + "logprob_reward": -0.0005597693379968405, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005597693379968405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1522609442472458, + 0.06378446519374847, + 0.07117660343647003, + 0.03438446670770645, + 0.05343096703290939, + 0.05392226204276085, + 0.09730794280767441, + 0.09024067223072052, + 0.13624444603919983, + 0.24724721908569336 + ], + "credit_map": [ + 0.023544952273368835, + 0.017426300793886185, + 0.019114259630441666, + 0.013439234346151352, + 0.022371483966708183, + 0.01860063523054123, + 0.019557051360607147, + 0.01820702664554119, + 0.007191590499132872, + 0.011122842319309711, + 0.004268590360879898, + 0.006419030018150806, + 0.012142636813223362, + 0.00851336307823658, + 0.006763739045709372, + 0.007362674456089735, + 0.006222441792488098, + 0.007732638623565435, + 0.006726541090756655, + 0.011138548143208027, + 0.011503483168780804, + 0.011315455660223961, + 0.00781134283170104, + 0.008726148866117, + 0.00342478696256876, + 0.003925024997442961, + 0.0027704278472810984, + 0.0032398924231529236, + 0.0060412706807255745, + 0.0035364856012165546, + 0.006360971834510565, + 0.005085603799670935, + 0.006205033510923386, + 0.008659462444484234, + 0.006685164291411638, + 0.007533684838563204, + 0.0066461157985031605, + 0.00530253816395998, + 0.00539524806663394, + 0.007003716193139553, + 0.0077645485289394855, + 0.010297726839780807, + 0.004503295756876469, + 0.006336817983537912, + 0.008443924598395824, + 0.004729099106043577, + 0.006221055518835783, + 0.005625792313367128, + 0.015526888892054558, + 0.01309217605739832, + 0.011890796013176441, + 0.009446514770388603, + 0.013522271998226643, + 0.011847162619233131, + 0.010758254677057266, + 0.011223878711462021, + 0.011592318303883076, + 0.009343920275568962, + 0.008879861794412136, + 0.01936556212604046, + 0.009663988836109638, + 0.01096108928322792, + 0.01017087697982788, + 0.010263058356940746, + 0.016360629349946976, + 0.015522606670856476, + 0.01520371064543724, + 0.018319854512810707, + 0.015815339982509613, + 0.02238607034087181, + 0.016769811511039734, + 0.015866434201598167, + 0.03268149867653847, + 0.023923572152853012, + 0.0331701785326004, + 0.02147618681192398, + 0.02285279706120491, + 0.04527352377772331, + 0.024701731279492378, + 0.04316771775484085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0005597693379968405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623862475156784 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623862475156784, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623862475156784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07722417265176773, + 0.05076387897133827, + 0.05034380778670311, + 0.017126144841313362, + 0.05309422314167023, + 0.05503401532769203, + 0.09637880325317383, + 0.09299331158399582, + 0.10998933762311935, + 0.3970523178577423 + ], + "credit_map": [ + 0.013186226598918438, + 0.008041434921324253, + 0.010069805197417736, + 0.006187324412167072, + 0.011661541648209095, + 0.009017466567456722, + 0.009277059696614742, + 0.00978330709040165, + 0.006433437112718821, + 0.00551894586533308, + 0.0043175723403692245, + 0.004215052351355553, + 0.012089231051504612, + 0.007192592136561871, + 0.005435118917375803, + 0.005561926867812872, + 0.004137970041483641, + 0.005078683141618967, + 0.004436276853084564, + 0.007447581272572279, + 0.008790326304733753, + 0.009008647873997688, + 0.006133320741355419, + 0.00531100295484066, + 0.0022066181991249323, + 0.0018256078474223614, + 0.0021365690045058727, + 0.0029244699981063604, + 0.002065816428512335, + 0.001659800997003913, + 0.002249278360977769, + 0.0020579840056598186, + 0.006667048204690218, + 0.006016669794917107, + 0.007466780487447977, + 0.0058462899178266525, + 0.006094435229897499, + 0.006838080007582903, + 0.008064948953688145, + 0.006099972873926163, + 0.008794378489255905, + 0.007824080996215343, + 0.006004693452268839, + 0.006680374965071678, + 0.007611704058945179, + 0.004820272792130709, + 0.007231305819004774, + 0.006067205220460892, + 0.017651140689849854, + 0.013047069311141968, + 0.007371588610112667, + 0.009943350218236446, + 0.019248025491833687, + 0.007414789404720068, + 0.01482048537582159, + 0.006882356945425272, + 0.01260889507830143, + 0.010090624913573265, + 0.008144155144691467, + 0.017416972666978836, + 0.00971930380910635, + 0.013357658870518208, + 0.009637407958507538, + 0.012018299661576748, + 0.010789372958242893, + 0.0134165333583951, + 0.013165219686925411, + 0.013660706579685211, + 0.01573283039033413, + 0.015008166432380676, + 0.013496576808393002, + 0.014719923958182335, + 0.04731550067663193, + 0.0302362572401762, + 0.04746203497052193, + 0.041329484432935715, + 0.0358426459133625, + 0.07100927084684372, + 0.04543403908610344, + 0.07842306047677994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.007623862475156784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming camera, from an upward angle, pans to a towering, mist-shrouded mountaintop. The rising sun casts a golden glow across the craggy rock. At the summit's center, a massive, ancient immortal stone stands silently, its surface covered in moss and weathered, appearing incredibly heavy and spiritual. Suddenly, the stone begins to vibrate slightly, and a dazzling golden light permeates its interior, rapidly spreading a web of cracks across its dark gray surface. Then, in a silent yet earth-shattering burst of energy, the stone suddenly explodes outward, sending countless fragments of varying sizes scattering in a hail of dust. As the dazzling golden light and swirling smoke from the center of the explosion dissipate, a stone monkey, draped in dazzling golden hair, emerges from the stone. Its hair gleams like molten gold, its muscles defined, and parts of its body still bear the rough texture of freshly hewn rock. It landed steadily on the broken stone foundation in a half-squatting posture full of primitive power, looking around at this new world curiously and defiantly. The most eye-catching thing was its deep eyes, which radiated two almost tangible, all-piercing golden beams, injecting a sacred and wild vitality into the entire grand and tranquil scene, announcing the birth of a legend.", + "seed": 15488, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_0/index_9254_seed_15488_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0209600176429376, + "vlm_logprob": -0.0209600176429376, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.0209600176429376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0780303031206131 + ], + "opposite_logprob_tokens": [ + -2.8280303478240967 + ], + "logprob_reward": -0.0780303031206131, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0780303031206131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2559029161930084, + 0.1862955391407013, + 0.10988092422485352, + 0.15177994966506958, + 0.12577511370182037, + 0.054133497178554535, + 0.05505747348070145, + 0.028766028583049774, + 0.014599353075027466, + 0.017809201031923294 + ], + "credit_map": [ + 0.014315088279545307, + 0.010540551505982876, + 0.01379309594631195, + 0.015162059105932713, + 0.010841382667422295, + 0.09014803171157837, + 0.010624624788761139, + 0.09047810733318329, + 0.01859983243048191, + 0.014215563423931599, + 0.012343030422925949, + 0.018680475652217865, + 0.022274615243077278, + 0.0372302420437336, + 0.02720901183784008, + 0.03574277088046074, + 0.01785261183977127, + 0.008487693965435028, + 0.012294928543269634, + 0.011905690655112267, + 0.013996699824929237, + 0.01335111353546381, + 0.016556022688746452, + 0.015436170622706413, + 0.017770487815141678, + 0.015306268818676472, + 0.02146279253065586, + 0.0164906308054924, + 0.013926238752901554, + 0.028882237151265144, + 0.0168747641146183, + 0.021066533401608467, + 0.014491081237792969, + 0.012490559369325638, + 0.013814791105687618, + 0.014340550638735294, + 0.014431098476052284, + 0.016185546293854713, + 0.018921159207820892, + 0.021100327372550964, + 0.007366372272372246, + 0.005459599196910858, + 0.008269565179944038, + 0.005515214055776596, + 0.005069293547421694, + 0.010557662695646286, + 0.00670994445681572, + 0.005185847170650959, + 0.009948797523975372, + 0.005283751059323549, + 0.008250817656517029, + 0.004679005593061447, + 0.004405617248266935, + 0.011640528216958046, + 0.004906277637928724, + 0.005942681338638067, + 0.007117705885320902, + 0.0029786184895783663, + 0.003882673801854253, + 0.002720592310652137, + 0.0028316800016909838, + 0.003953623119741678, + 0.0021255919709801674, + 0.0031555432360619307, + 0.0018544963095337152, + 0.0027432062197476625, + 0.0008512368076480925, + 0.00149433349724859, + 0.0023234302643686533, + 0.0026178269181400537, + 0.0016597704961895943, + 0.0010550526203587651, + 0.002040694933384657, + 0.002774617401883006, + 0.001529628410935402, + 0.0033662973437458277, + 0.0024448710028082132, + 0.003079786663874984, + 0.0013404759811237454, + 0.0012328311568126082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0780303031206131 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006758449599146843 + ], + "opposite_logprob_tokens": [ + -5.006758213043213 + ], + "logprob_reward": -0.006758449599146843, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006758449599146843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24394363164901733, + 0.19143511354923248, + 0.029456553980708122, + 0.03189186006784439, + 0.13399577140808105, + 0.07673685252666473, + 0.10991257429122925, + 0.11516901105642319, + 0.03764958307147026, + 0.02980906516313553 + ], + "credit_map": [ + 0.010276767425239086, + 0.011087706312537193, + 0.011558475904166698, + 0.015395833179354668, + 0.015719132497906685, + 0.08048810809850693, + 0.014120539650321007, + 0.08529703319072723, + 0.020760467275977135, + 0.017239311710000038, + 0.019238270819187164, + 0.021942010149359703, + 0.021082913503050804, + 0.03432201221585274, + 0.02392806112766266, + 0.03292206674814224, + 0.004217283800244331, + 0.002972895046696067, + 0.004012269899249077, + 0.003731104312464595, + 0.0026799074839800596, + 0.0031818372663110495, + 0.004206288605928421, + 0.004454966634511948, + 0.0036276462487876415, + 0.0031460996251553297, + 0.003927718382328749, + 0.0032772645354270935, + 0.002866866998374462, + 0.0071516684256494045, + 0.0046548196114599705, + 0.003239777171984315, + 0.012507675215601921, + 0.015507228672504425, + 0.014258448034524918, + 0.021700015291571617, + 0.009730391204357147, + 0.015441546216607094, + 0.022435549646615982, + 0.02241491712629795, + 0.008050210773944855, + 0.010872244834899902, + 0.016824133694171906, + 0.007507950067520142, + 0.007017309777438641, + 0.011360904201865196, + 0.007137751672416925, + 0.007966344244778156, + 0.0205870121717453, + 0.009305316023528576, + 0.011037996038794518, + 0.009111303836107254, + 0.010116003453731537, + 0.027796268463134766, + 0.00887343194335699, + 0.013085230253636837, + 0.01304890401661396, + 0.012958072125911713, + 0.02371365576982498, + 0.023419572040438652, + 0.008261100389063358, + 0.015554477460682392, + 0.010194006375968456, + 0.0080192219465971, + 0.003728073090314865, + 0.004923657514154911, + 0.007486936170607805, + 0.004587316885590553, + 0.004258373286575079, + 0.0044297147542238235, + 0.0035798351746052504, + 0.004655675496906042, + 0.004701157100498676, + 0.005207540467381477, + 0.002183027332648635, + 0.0022374119143933058, + 0.00493868812918663, + 0.005695001222193241, + 0.0025562129449099302, + 0.0022900232579559088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.006758449599146843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644440021365881 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.004644440021365881, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644440021365881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18816863000392914, + 0.027427755296230316, + 0.03253147378563881, + 0.03457009419798851, + 0.22427701950073242, + 0.15375392138957977, + 0.22555139660835266, + 0.04254787787795067, + 0.03367914631962776, + 0.03749264404177666 + ], + "credit_map": [ + 0.00911862961947918, + 0.007389913313090801, + 0.006136487238109112, + 0.007570311892777681, + 0.008122696541249752, + 0.06946390867233276, + 0.00886299554258585, + 0.07150370627641678, + 0.003916963469237089, + 0.003039384726434946, + 0.003380522597581148, + 0.003461848245933652, + 0.0030816260259598494, + 0.004259904846549034, + 0.0029479756485670805, + 0.0033395346254110336, + 0.005122669506818056, + 0.002280471846461296, + 0.005310875829309225, + 0.002828380558639765, + 0.002819230780005455, + 0.004919950384646654, + 0.0030599392484873533, + 0.0061899591237306595, + 0.0037078692112118006, + 0.004622113890945911, + 0.004817703738808632, + 0.003936196677386761, + 0.003913864027708769, + 0.00540726026520133, + 0.0049080755561590195, + 0.003257013624534011, + 0.020264457911252975, + 0.026775414124131203, + 0.02187975123524666, + 0.039396122097969055, + 0.01649860292673111, + 0.02427465282380581, + 0.04424211382865906, + 0.030945923179388046, + 0.01777173951268196, + 0.016561195254325867, + 0.02782297134399414, + 0.015457005240023136, + 0.013638436794281006, + 0.02944268099963665, + 0.01627649925649166, + 0.016783418133854866, + 0.031094664707779884, + 0.017816836014389992, + 0.026233790442347527, + 0.023775767534971237, + 0.019778765738010406, + 0.03962719067931175, + 0.018558867275714874, + 0.04866555333137512, + 0.006122946739196777, + 0.004523162730038166, + 0.005603113677352667, + 0.005892036948353052, + 0.005670972168445587, + 0.005484235007315874, + 0.004591538105159998, + 0.004659874364733696, + 0.0027858088724315166, + 0.002999573014676571, + 0.005383782088756561, + 0.004478488117456436, + 0.006338296923786402, + 0.0037259049713611603, + 0.003812363836914301, + 0.004154928494244814, + 0.004327384755015373, + 0.005782269407063723, + 0.002733986359089613, + 0.005366594064980745, + 0.005874009802937508, + 0.007024524267762899, + 0.00308779114857316, + 0.0032960851676762104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.004644440021365881 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178999780677259 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.001178999780677259, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178999780677259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1883254200220108, + 0.09624836593866348, + 0.03414314240217209, + 0.020378490909934044, + 0.14552418887615204, + 0.047688078135252, + 0.20610228180885315, + 0.10681243985891342, + 0.0791303738951683, + 0.07564720511436462 + ], + "credit_map": [ + 0.010512575507164001, + 0.009624558500945568, + 0.009330837056040764, + 0.010522107593715191, + 0.014064894057810307, + 0.0625629797577858, + 0.009772068820893764, + 0.06193540617823601, + 0.011568010784685612, + 0.011616930365562439, + 0.007909981533885002, + 0.010744014754891396, + 0.011059707961976528, + 0.015586222521960735, + 0.013563062995672226, + 0.014200438745319843, + 0.004743719473481178, + 0.002530122408643365, + 0.006374834571033716, + 0.004004491493105888, + 0.0023726383224129677, + 0.00444720359519124, + 0.003961688373237848, + 0.005708445329219103, + 0.002172290813177824, + 0.0024480384308844805, + 0.0034557252656668425, + 0.0026688857469707727, + 0.002297322265803814, + 0.0028401417657732964, + 0.0024380767717957497, + 0.0020580075215548277, + 0.01588035561144352, + 0.017201589420437813, + 0.013851356692612171, + 0.019961604848504066, + 0.012970935553312302, + 0.015439769253134727, + 0.028793346136808395, + 0.021425241604447365, + 0.006190378684550524, + 0.009416453540325165, + 0.006573832128196955, + 0.005777650512754917, + 0.005453702062368393, + 0.0044187824241817, + 0.004998801741749048, + 0.004858477506786585, + 0.027446124702692032, + 0.020491328090429306, + 0.029432980343699455, + 0.021069707348942757, + 0.022183235734701157, + 0.03349529579281807, + 0.01889844425022602, + 0.033085163682699203, + 0.016508808359503746, + 0.01028214767575264, + 0.010077967308461666, + 0.013751051388680935, + 0.019034385681152344, + 0.017991436645388603, + 0.011267290450632572, + 0.007899357005953789, + 0.008930424228310585, + 0.006160026881843805, + 0.010183591395616531, + 0.008066610433161259, + 0.013623636215925217, + 0.01111014373600483, + 0.010857312008738518, + 0.010198629461228848, + 0.009426298551261425, + 0.008805377408862114, + 0.00575154647231102, + 0.009236677549779415, + 0.01225096546113491, + 0.014663969166576862, + 0.008446810767054558, + 0.007065561134368181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.001178999780677259 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with an upward angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a massive ancient stone covered in moss and weathered texture at the mountaintop summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the stone vibrate, crack with golden light, and then explode into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a stone monkey with molten gold hair and rough rock-textured body emerge from the explosion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stone monkey land in a half-squatting posture with deep eyes radiating golden beams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187895692884922 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.014187895692884922, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187895692884922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2166343629360199, + 0.06192289665341377, + 0.0355190746486187, + 0.04545037820935249, + 0.11495577543973923, + 0.06953497976064682, + 0.11402949690818787, + 0.10148520022630692, + 0.1397978663444519, + 0.10066995769739151 + ], + "credit_map": [ + 0.010500073432922363, + 0.01055978611111641, + 0.009179904125630856, + 0.0131525918841362, + 0.01658099703490734, + 0.06955371052026749, + 0.018440160900354385, + 0.06866713613271713, + 0.009561852551996708, + 0.00921616330742836, + 0.006475617177784443, + 0.006537254434078932, + 0.005480744875967503, + 0.009482204914093018, + 0.006736317183822393, + 0.008432739414274693, + 0.004972108639776707, + 0.0029897403437644243, + 0.005451613571494818, + 0.004506118129938841, + 0.0029311361722648144, + 0.004803557880222797, + 0.0044712163507938385, + 0.0053935847245156765, + 0.005734202917665243, + 0.005455383099615574, + 0.007182186935096979, + 0.004549264907836914, + 0.004934282042086124, + 0.006189254578202963, + 0.006195427849888802, + 0.005210375878959894, + 0.016058212146162987, + 0.011996792629361153, + 0.01266772486269474, + 0.013279722072184086, + 0.014236164279282093, + 0.013635195791721344, + 0.019545050337910652, + 0.013536917977035046, + 0.010405930690467358, + 0.00836870539933443, + 0.013439483940601349, + 0.006946768146008253, + 0.006415142212063074, + 0.01179331075400114, + 0.005346089601516724, + 0.006819551344960928, + 0.014304080046713352, + 0.020662853494286537, + 0.014316494576632977, + 0.011883150786161423, + 0.010283142328262329, + 0.02009234018623829, + 0.009941517375409603, + 0.012545917183160782, + 0.021129198372364044, + 0.011221693828701973, + 0.00795762613415718, + 0.007968086749315262, + 0.015497708693146706, + 0.016312871128320694, + 0.01144671905785799, + 0.009951293468475342, + 0.020677918568253517, + 0.031504709273576736, + 0.010004115290939808, + 0.012237769551575184, + 0.02021605335175991, + 0.024054404348134995, + 0.01156980823725462, + 0.009533093310892582, + 0.008496424183249474, + 0.009344445541501045, + 0.009592863731086254, + 0.015288404189050198, + 0.014070813544094563, + 0.023313693702220917, + 0.011735450476408005, + 0.008827862329781055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.014187895692884922 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_tensor([9873, 9873])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_tensor([9873, 9873])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..18df65559a79c47190686aab82653eef816e39a4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_tensor([9873, 9873])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, eye-level wide-angle tracking shot, filmed in visceral, high-frame slow motion in the style of motion photography, follows the raw, destructive force of a massive tornado as it carves a path through a dense forest of mature maples. The tornado itself is a dark, roiling funnel of dust, soil, and debris, its form silhouetted eerily against the ominous, bruised purple and green storm sky. As the vortex makes contact with the trees, the air erupts into a chaotic whirlwind, and vibrant autumn colors—fiery reds, vibrant oranges, and golden yellows—are violently torn from branches and pulled into its swirling maw. The camera meticulously tracks the wind's immense force as it first bends the sturdy, thick trunks to the breaking point. In hyper-detailed slow motion, the bark begins to peel and splinter, then the wood fibers themselves tear apart in a visceral display of natural strength, sending a torrent of wood splinters through the swirling chaos, creating a terrifying and awe-inspiring spectacle of nature's power.", + "seed": 637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_9873_seed_637_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1073933294042945, + "vlm_logprob": -0.1073933294042945, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.1073933294042945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08568764477968216 + ], + "opposite_logprob_tokens": [ + -2.5856876373291016 + ], + "logprob_reward": -0.08568764477968216, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08568764477968216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1431046426296234, + 0.19940704107284546, + 0.11492506414651871, + 0.10586825758218765, + 0.05405339226126671, + 0.041045576333999634, + 0.08073288202285767, + 0.09761784970760345, + 0.05564119666814804, + 0.10760409384965897 + ], + "credit_map": [ + 0.01138407550752163, + 0.029088875278830528, + 0.015189897269010544, + 0.01562894880771637, + 0.015432723797857761, + 0.013610925525426865, + 0.017839031293988228, + 0.024930184707045555, + 0.02182537317276001, + 0.04569701477885246, + 0.014412438496947289, + 0.013706372119486332, + 0.010765697807073593, + 0.018770592287182808, + 0.01739521138370037, + 0.05683435872197151, + 0.011570458300411701, + 0.010817979462444782, + 0.017827916890382767, + 0.022832011803984642, + 0.008649026975035667, + 0.016701379790902138, + 0.01224633026868105, + 0.014279969967901707, + 0.013591405935585499, + 0.028690533712506294, + 0.013248990289866924, + 0.008411345072090626, + 0.013891012407839298, + 0.004803780000656843, + 0.015353209339082241, + 0.007877991534769535, + 0.0069997431710362434, + 0.005152747966349125, + 0.010507794097065926, + 0.006659745704382658, + 0.0053017293103039265, + 0.003924727439880371, + 0.00872357189655304, + 0.0067833359353244305, + 0.0032575989607721567, + 0.003714567283168435, + 0.010059436783194542, + 0.005755208432674408, + 0.003321324707940221, + 0.00269978865981102, + 0.005116801708936691, + 0.007120853755623102, + 0.005671599414199591, + 0.006400028709322214, + 0.0240531824529171, + 0.010332814417779446, + 0.005637933034449816, + 0.005178735125809908, + 0.006382830906659365, + 0.017075764015316963, + 0.00485843513160944, + 0.003836813848465681, + 0.030792798846960068, + 0.006215044762939215, + 0.00539673725143075, + 0.0037717833183705807, + 0.03516647592186928, + 0.007579767145216465, + 0.005355472676455975, + 0.00834730640053749, + 0.005665346048772335, + 0.007088822778314352, + 0.005465466529130936, + 0.0074391914531588554, + 0.005969374906271696, + 0.0103102196007967, + 0.006540937349200249, + 0.013722830452024937, + 0.0282679945230484, + 0.01006898283958435, + 0.026593802496790886, + 0.008049857802689075, + 0.00682843616232276, + 0.007531254086643457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.08568764477968216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982516586780548 + ], + "opposite_logprob_tokens": [ + -3.529825210571289 + ], + "logprob_reward": -0.02982516586780548, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982516586780548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11229971796274185, + 0.1190372109413147, + 0.07529095560312271, + 0.05160815641283989, + 0.06445222347974777, + 0.06772365421056747, + 0.2536439895629883, + 0.1235281452536583, + 0.06589425355195999, + 0.06652170419692993 + ], + "credit_map": [ + 0.007721377070993185, + 0.014289404265582561, + 0.012863845564424992, + 0.016511032357811928, + 0.009785819798707962, + 0.008841637521982193, + 0.019306540489196777, + 0.02298005484044552, + 0.010817890986800194, + 0.016451051458716393, + 0.01578659377992153, + 0.018028391525149345, + 0.011417299509048462, + 0.017191199585795403, + 0.010236402973532677, + 0.019108369946479797, + 0.007965920493006706, + 0.015608659014105797, + 0.009592433460056782, + 0.008906152099370956, + 0.006118893623352051, + 0.006411654409021139, + 0.006548152770847082, + 0.014139088802039623, + 0.0069554513320326805, + 0.013148312456905842, + 0.006785175297409296, + 0.004077163990586996, + 0.005534471943974495, + 0.003679199144244194, + 0.006979071069508791, + 0.004449308849871159, + 0.004446194041520357, + 0.00835514161735773, + 0.010706042870879173, + 0.007515376899391413, + 0.010600992478430271, + 0.003858553245663643, + 0.013508849777281284, + 0.005461071152240038, + 0.0054941726848483086, + 0.0057463617995381355, + 0.00783565267920494, + 0.023731691762804985, + 0.004198800306767225, + 0.0031052706763148308, + 0.009943768382072449, + 0.0076679326593875885, + 0.01678358018398285, + 0.018795309588313103, + 0.0626474991440773, + 0.030037708580493927, + 0.032828278839588165, + 0.01965775527060032, + 0.030093511566519737, + 0.04280035197734833, + 0.008342531509697437, + 0.012360581196844578, + 0.03261430189013481, + 0.009459656663239002, + 0.008951679803431034, + 0.009156960994005203, + 0.030309220775961876, + 0.012333216145634651, + 0.006538549438118935, + 0.011838816106319427, + 0.006721331272274256, + 0.008527198806405067, + 0.006834371946752071, + 0.006007758434861898, + 0.011929763481020927, + 0.007496463600546122, + 0.0034886018838733435, + 0.006782554555684328, + 0.019231880083680153, + 0.006202006712555885, + 0.01639685221016407, + 0.00524967210367322, + 0.004534066189080477, + 0.004636066500097513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.02982516586780548 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023309852927923203 + ], + "opposite_logprob_tokens": [ + -3.7733099460601807 + ], + "logprob_reward": -0.023309852927923203, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023309852927923203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10402354598045349, + 0.1014447882771492, + 0.11903924494981766, + 0.0458608902990818, + 0.1554211527109146, + 0.08936677873134613, + 0.19787760078907013, + 0.049282271414995193, + 0.05743669345974922, + 0.08024702221155167 + ], + "credit_map": [ + 0.013203028589487076, + 0.019851775839924812, + 0.009559598751366138, + 0.012129351496696472, + 0.01044973824173212, + 0.007084117271006107, + 0.014768431894481182, + 0.01697751320898533, + 0.008316073566675186, + 0.012716957367956638, + 0.012963550165295601, + 0.017984718084335327, + 0.00856687594205141, + 0.012524562887847424, + 0.0088644502684474, + 0.019507603719830513, + 0.013175968080759048, + 0.03360193595290184, + 0.010369264520704746, + 0.010417967103421688, + 0.009830253198742867, + 0.012576810084283352, + 0.009808770380914211, + 0.019258277490735054, + 0.003709865268319845, + 0.006410484202206135, + 0.0067718736827373505, + 0.005026622675359249, + 0.006361719686537981, + 0.0034092606510967016, + 0.008642181754112244, + 0.005528884008526802, + 0.01614866405725479, + 0.016009248793125153, + 0.018261101096868515, + 0.020900562405586243, + 0.022003868594765663, + 0.006932184100151062, + 0.03982715681195259, + 0.015338369645178318, + 0.006148906424641609, + 0.007293894421309233, + 0.012687308713793755, + 0.028091706335544586, + 0.008094774559140205, + 0.0048097181133925915, + 0.011961056850850582, + 0.010279413312673569, + 0.020005108788609505, + 0.017181238159537315, + 0.02549244463443756, + 0.025044001638889313, + 0.03699710592627525, + 0.020014751702547073, + 0.031279128044843674, + 0.021863799542188644, + 0.004248144570738077, + 0.004467173013836145, + 0.007486181799322367, + 0.006738432217389345, + 0.0048286025412380695, + 0.0055776918306946754, + 0.0072978585958480835, + 0.008638186380267143, + 0.0052052149549126625, + 0.007401483133435249, + 0.006886331830173731, + 0.004605005029588938, + 0.006242967210710049, + 0.011278445832431316, + 0.007487162947654724, + 0.008330084383487701, + 0.004094836302101612, + 0.005993811879307032, + 0.02163369208574295, + 0.008366018533706665, + 0.02060960978269577, + 0.006289549171924591, + 0.006984321866184473, + 0.006275179795920849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.023309852927923203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075920425355434 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075920425355434, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075920425355434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11996471881866455, + 0.4059092402458191, + 0.05312390998005867, + 0.025470003485679626, + 0.0387411043047905, + 0.030152738094329834, + 0.1501505970954895, + 0.13060259819030762, + 0.020628217607736588, + 0.02525690384209156 + ], + "credit_map": [ + 0.00845912005752325, + 0.017094772309064865, + 0.014122504740953445, + 0.021766992285847664, + 0.007520030718296766, + 0.009353797882795334, + 0.018894042819738388, + 0.022753460332751274, + 0.03426571935415268, + 0.06181939318776131, + 0.04448607936501503, + 0.06398001313209534, + 0.028188038617372513, + 0.052717115730047226, + 0.03482501208782196, + 0.08562786877155304, + 0.006690457928925753, + 0.01510534342378378, + 0.004695225972682238, + 0.005931791849434376, + 0.004256348125636578, + 0.006210660561919212, + 0.003698419313877821, + 0.006535661872476339, + 0.002651261631399393, + 0.0038643330335617065, + 0.006381955463439226, + 0.0029312167316675186, + 0.003379927249625325, + 0.0012687008129432797, + 0.0029862539377063513, + 0.002006353810429573, + 0.004421499092131853, + 0.0035789122339338064, + 0.006838914472609758, + 0.0038113966584205627, + 0.0029686305206269026, + 0.0018345718272030354, + 0.011420018039643764, + 0.00386716122739017, + 0.0017146485624834895, + 0.0023686541244387627, + 0.005346683785319328, + 0.009489072486758232, + 0.0018181803170591593, + 0.0015402771532535553, + 0.0034628126304596663, + 0.004412410315126181, + 0.015074114315211773, + 0.017734259366989136, + 0.029407229274511337, + 0.0199190191924572, + 0.024617845192551613, + 0.01129505317658186, + 0.016815222799777985, + 0.015287832356989384, + 0.007378908805549145, + 0.005653548985719681, + 0.043030329048633575, + 0.007077978923916817, + 0.00535562913864851, + 0.0056848241947591305, + 0.04759417101740837, + 0.008827204816043377, + 0.0018944400362670422, + 0.002702597063034773, + 0.0029027287382632494, + 0.003191349096596241, + 0.0019612384494394064, + 0.002494602231308818, + 0.0025888625532388687, + 0.0028923999052494764, + 0.001962372800335288, + 0.004204019904136658, + 0.00549926795065403, + 0.0029132694471627474, + 0.004633876960724592, + 0.0020782523788511753, + 0.0019840775057673454, + 0.001981769222766161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.011075920425355434 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870680630207062 + ], + "opposite_logprob_tokens": [ + -1.1370680332183838 + ], + "logprob_reward": -0.3870680630207062, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870680630207062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06246339902281761, + 0.08222699910402298, + 0.03887476772069931, + 0.14663001894950867, + 0.14588747918605804, + 0.09538748860359192, + 0.19403265416622162, + 0.1281258463859558, + 0.07373137772083282, + 0.03263998404145241 + ], + "credit_map": [ + 0.0056684124283492565, + 0.01256616972386837, + 0.00788272637873888, + 0.010962175205349922, + 0.004064584616571665, + 0.005053886212408543, + 0.007833455689251423, + 0.008431989699602127, + 0.007147867698222399, + 0.015650926157832146, + 0.007213567849248648, + 0.009960532188415527, + 0.00619908981025219, + 0.008365214802324772, + 0.00863646063953638, + 0.019053343683481216, + 0.004291776102036238, + 0.007124095689505339, + 0.004754791967570782, + 0.0056874556466937065, + 0.0029498417861759663, + 0.0033159281592816114, + 0.004041030537337065, + 0.006709848530590534, + 0.02163759618997574, + 0.043758172541856766, + 0.014087328687310219, + 0.010361360386013985, + 0.013414920307695866, + 0.005864412523806095, + 0.024022208526730537, + 0.013484012335538864, + 0.011105628684163094, + 0.011091228574514389, + 0.02109847590327263, + 0.010841836221516132, + 0.009664024226367474, + 0.00670289434492588, + 0.05060673505067825, + 0.024776652455329895, + 0.010081860236823559, + 0.006409631576389074, + 0.011868984438478947, + 0.030763685703277588, + 0.005549301393330097, + 0.005596630275249481, + 0.01565713807940483, + 0.009460258297622204, + 0.016384761780500412, + 0.020312301814556122, + 0.04500933736562729, + 0.027914734557271004, + 0.016782430931925774, + 0.011663458310067654, + 0.01588420569896698, + 0.040081411600112915, + 0.007825568318367004, + 0.007784656714648008, + 0.03662944212555885, + 0.007822848856449127, + 0.006868292577564716, + 0.007167413365095854, + 0.041248854249715805, + 0.012778762727975845, + 0.007981871254742146, + 0.013328754343092442, + 0.007502014748752117, + 0.010777751915156841, + 0.009348887018859386, + 0.0071183075197041035, + 0.010365732945501804, + 0.0073080603033304214, + 0.0020524058490991592, + 0.004439794458448887, + 0.004536422900855541, + 0.005718658212572336, + 0.006249457597732544, + 0.002957782940939069, + 0.003642893163487315, + 0.003042570548132062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.3870680630207062 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, eye-level wide-angle tracking shot, filmed in visceral, high-frame slow motion in the style of motion photography, follows the raw, destructive force of a massive tornado as it carves a path through a dense forest of mature maples. The tornado itself is a dark, roiling funnel of dust, soil, and debris, its form silhouetted eerily against the ominous, bruised purple and green storm sky. As the vortex makes contact with the trees, the air erupts into a chaotic whirlwind, and vibrant autumn colors—fiery reds, vibrant oranges, and golden yellows—are violently torn from branches and pulled into its swirling maw. The camera meticulously tracks the wind's immense force as it first bends the sturdy, thick trunks to the breaking point. In hyper-detailed slow motion, the bark begins to peel and splinter, then the wood fibers themselves tear apart in a visceral display of natural strength, sending a torrent of wood splinters through the swirling chaos, creating a terrifying and awe-inspiring spectacle of nature's power.", + "seed": 638, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_9873_seed_638_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1685000468045473, + "vlm_logprob": -0.1685000468045473, + "vlm_yes_no_margin": 2.325000011920929, + "vlm_token_credit": -0.1685000468045473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3974866271018982 + ], + "opposite_logprob_tokens": [ + -1.147486686706543 + ], + "logprob_reward": -0.3974866271018982, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3974866271018982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30957332253456116, + 0.03354237973690033, + 0.07047515362501144, + 0.06009036675095558, + 0.04775683954358101, + 0.06626742333173752, + 0.14727628231048584, + 0.10348259657621384, + 0.031832631677389145, + 0.12970292568206787 + ], + "credit_map": [ + 0.017458293586969376, + 0.04205842316150665, + 0.03002360835671425, + 0.05127742514014244, + 0.03262629359960556, + 0.04532700404524803, + 0.034202124923467636, + 0.056600190699100494, + 0.00323456060141325, + 0.0026587704196572304, + 0.005211693234741688, + 0.007389119826257229, + 0.0042740898206830025, + 0.0025533968582749367, + 0.0051231905817985535, + 0.0030975581612437963, + 0.008850261569023132, + 0.005633845459669828, + 0.011127958074212074, + 0.00905569363385439, + 0.010254544205963612, + 0.010525960475206375, + 0.007392311468720436, + 0.007634581532329321, + 0.0029378945473581553, + 0.0024221164640039206, + 0.021359624341130257, + 0.014601597562432289, + 0.003703770227730274, + 0.004046586807817221, + 0.006354140117764473, + 0.004664642736315727, + 0.005985792260617018, + 0.004210362676531076, + 0.006756509654223919, + 0.005244473926723003, + 0.005605620797723532, + 0.005117541644722223, + 0.009423072449862957, + 0.005413469392806292, + 0.006302565801888704, + 0.014227479696273804, + 0.007213169243186712, + 0.004908501636236906, + 0.0033020959235727787, + 0.004689565394073725, + 0.009315014816820621, + 0.016309037804603577, + 0.012648518197238445, + 0.011653072200715542, + 0.015245791524648666, + 0.015557091683149338, + 0.012871169485151768, + 0.01263398677110672, + 0.037312425673007965, + 0.02935425005853176, + 0.011311894282698631, + 0.008054058067500591, + 0.008513509295880795, + 0.019044214859604836, + 0.008965922519564629, + 0.011240784078836441, + 0.023355688899755478, + 0.01299652736634016, + 0.0037625220138579607, + 0.003056133398786187, + 0.002121326746419072, + 0.0033830138854682446, + 0.004293309990316629, + 0.006971914321184158, + 0.004781551193445921, + 0.0034628650173544884, + 0.020459739491343498, + 0.01475415751338005, + 0.014619236811995506, + 0.011293124407529831, + 0.010603252798318863, + 0.014210837893188, + 0.0175328366458416, + 0.026229752227663994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.3974866271018982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381297767162323 + ], + "opposite_logprob_tokens": [ + -3.2881298065185547 + ], + "logprob_reward": -0.0381297767162323, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381297767162323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3956293761730194, + 0.045820482075214386, + 0.0858205035328865, + 0.09815976768732071, + 0.043879128992557526, + 0.04468246176838875, + 0.06579124182462692, + 0.12160056084394455, + 0.03624926507472992, + 0.062367163598537445 + ], + "credit_map": [ + 0.025685535743832588, + 0.05181393399834633, + 0.0385303869843483, + 0.05439542979001999, + 0.10146892815828323, + 0.04723283648490906, + 0.03286948427557945, + 0.04363284260034561, + 0.003014120040461421, + 0.0032947431318461895, + 0.0064913188107311726, + 0.009557390585541725, + 0.005578196607530117, + 0.009608704596757889, + 0.004722098354250193, + 0.0035539104137569666, + 0.006049038376659155, + 0.0059008109383285046, + 0.011292100884020329, + 0.011677082628011703, + 0.01842995546758175, + 0.010573780164122581, + 0.014688336290419102, + 0.007209399249404669, + 0.008236320689320564, + 0.0065949480049312115, + 0.013830882497131824, + 0.016763096675276756, + 0.010297782719135284, + 0.01151266135275364, + 0.01720980927348137, + 0.01371425949037075, + 0.006400919985026121, + 0.004143957514315844, + 0.00795831996947527, + 0.006142028607428074, + 0.002902309177443385, + 0.006061422172933817, + 0.006174656096845865, + 0.004095511045306921, + 0.004062402527779341, + 0.009682312607765198, + 0.00465375604107976, + 0.004899907391518354, + 0.003073203843086958, + 0.0073419492691755295, + 0.004985385108739138, + 0.005983545910567045, + 0.01167693454772234, + 0.014069054275751114, + 0.008158663287758827, + 0.006282830145210028, + 0.0033532234374433756, + 0.004232847131788731, + 0.010549311526119709, + 0.007468369789421558, + 0.012746216729283333, + 0.01392536610364914, + 0.012254633009433746, + 0.01817830093204975, + 0.014550743624567986, + 0.013848342001438141, + 0.02217358537018299, + 0.013923369348049164, + 0.003985857591032982, + 0.004057590384036303, + 0.002128727501258254, + 0.004264322109520435, + 0.006825460121035576, + 0.0035702914465218782, + 0.007227868773043156, + 0.004189147613942623, + 0.006696837022900581, + 0.007950103841722012, + 0.0036293042358011007, + 0.0055871764197945595, + 0.006561498623341322, + 0.01048838160932064, + 0.008675063960254192, + 0.01277879998087883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0381297767162323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298707127571106 + ], + "opposite_logprob_tokens": [ + -2.2379870414733887 + ], + "logprob_reward": -0.11298707127571106, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298707127571106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2369534969329834, + 0.07423581182956696, + 0.11641258746385574, + 0.03989056497812271, + 0.05953548476099968, + 0.10466573387384415, + 0.1272619068622589, + 0.08330655843019485, + 0.0446198545396328, + 0.1131180077791214 + ], + "credit_map": [ + 0.022862618789076805, + 0.013460681773722172, + 0.016741110011935234, + 0.06530001014471054, + 0.00963882077485323, + 0.012220029719173908, + 0.015812411904335022, + 0.0809178277850151, + 0.0058618769980967045, + 0.006786076817661524, + 0.011155259795486927, + 0.014656469225883484, + 0.009621388278901577, + 0.008595641702413559, + 0.011719769798219204, + 0.005839337129145861, + 0.008004380390048027, + 0.006621076725423336, + 0.01639297977089882, + 0.012787841260433197, + 0.029063500463962555, + 0.015589391812682152, + 0.018616903573274612, + 0.009336517192423344, + 0.0029429749120026827, + 0.002463876735419035, + 0.00967793446034193, + 0.0072951540350914, + 0.004154469817876816, + 0.002872671466320753, + 0.006052626296877861, + 0.004430860280990601, + 0.005318349227309227, + 0.005854686722159386, + 0.015290120616555214, + 0.010492806322872639, + 0.003409379394724965, + 0.005430290941148996, + 0.008263408206403255, + 0.005476446822285652, + 0.010379420593380928, + 0.026197856292128563, + 0.009963464923202991, + 0.005881732329726219, + 0.005263366270810366, + 0.013361134566366673, + 0.010753583163022995, + 0.0228651724755764, + 0.007599571254104376, + 0.00979409459978342, + 0.02497466839849949, + 0.007718746084719896, + 0.005366743542253971, + 0.004598409403115511, + 0.05019455403089523, + 0.01701512560248375, + 0.010876640677452087, + 0.009348011575639248, + 0.008450284600257874, + 0.010185997001826763, + 0.00882016122341156, + 0.010699832811951637, + 0.013190330006182194, + 0.01173530612140894, + 0.004816029686480761, + 0.005162587855011225, + 0.0027965379413217306, + 0.004296051803976297, + 0.011394781060516834, + 0.005214485805481672, + 0.006884433329105377, + 0.004054951947182417, + 0.013771727681159973, + 0.01001605112105608, + 0.0071203140541911125, + 0.008946292102336884, + 0.014278547838330269, + 0.023129146546125412, + 0.019306311383843422, + 0.016549626365303993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.11298707127571106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549126520752907 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012549126520752907, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549126520752907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3112482726573944, + 0.03905011713504791, + 0.02717095986008644, + 0.21666772663593292, + 0.10277179628610611, + 0.06708396226167679, + 0.05801010504364967, + 0.12085946649312973, + 0.028288640081882477, + 0.028848953545093536 + ], + "credit_map": [ + 0.02687116712331772, + 0.041564956307411194, + 0.030982093885540962, + 0.043720342218875885, + 0.03926656022667885, + 0.051633674651384354, + 0.0304635688662529, + 0.04674592241644859, + 0.0022344845347106457, + 0.002783732023090124, + 0.006377581972628832, + 0.009710738435387611, + 0.00552732078358531, + 0.0031757832039147615, + 0.004960726946592331, + 0.004279753193259239, + 0.001583293778821826, + 0.0018254402093589306, + 0.003743703942745924, + 0.005318441893905401, + 0.005297828931361437, + 0.003065595868974924, + 0.00395766319707036, + 0.0023789936676621437, + 0.01470401044934988, + 0.0072380066849291325, + 0.07986196130514145, + 0.0572064071893692, + 0.014160907827317715, + 0.011418548412621021, + 0.019744114950299263, + 0.0123337646946311, + 0.0075868843123316765, + 0.005726856645196676, + 0.03517208620905876, + 0.01783832535147667, + 0.005424792412668467, + 0.0077100833877921104, + 0.015362201258540154, + 0.00795057788491249, + 0.0055532269179821014, + 0.012488186359405518, + 0.00645667128264904, + 0.005616035312414169, + 0.0032911552116274834, + 0.009021449834108353, + 0.008138701319694519, + 0.01651853695511818, + 0.003732610959559679, + 0.004287211690098047, + 0.010800636373460293, + 0.004908295348286629, + 0.0033081178553402424, + 0.0026891916058957577, + 0.020956965163350105, + 0.007327082566916943, + 0.01261154655367136, + 0.01267824787646532, + 0.01045962143689394, + 0.024978501722216606, + 0.012122760526835918, + 0.01104644313454628, + 0.023167766630649567, + 0.013794578611850739, + 0.003468622686341405, + 0.004161834716796875, + 0.0030850123148411512, + 0.0026778155006468296, + 0.005051681771874428, + 0.0028992414008826017, + 0.004627308342605829, + 0.00231712544336915, + 0.004944473505020142, + 0.002761887852102518, + 0.0034286545123904943, + 0.0023817108012735844, + 0.0031347181648015976, + 0.004719209857285023, + 0.0038454229943454266, + 0.003632878651842475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.012549126520752907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813476324081421 + ], + "opposite_logprob_tokens": [ + -1.406347632408142 + ], + "logprob_reward": -0.2813476324081421, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813476324081421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3282181918621063, + 0.04192864149808884, + 0.0369148924946785, + 0.12166286259889603, + 0.11598759144544601, + 0.12068381905555725, + 0.12817057967185974, + 0.04645928367972374, + 0.026970570906996727, + 0.03300352394580841 + ], + "credit_map": [ + 0.03103255294263363, + 0.02131625823676586, + 0.02317775972187519, + 0.0815359354019165, + 0.02307821251451969, + 0.016177020967006683, + 0.0215816218405962, + 0.11031884700059891, + 0.0030587457586079836, + 0.002770376158878207, + 0.005768518429249525, + 0.00664679566398263, + 0.005378944333642721, + 0.007852353155612946, + 0.005753037985414267, + 0.0046998681500554085, + 0.0033444217406213284, + 0.00201793503947556, + 0.004893174394965172, + 0.005371611099690199, + 0.006194656249135733, + 0.005323733668774366, + 0.005400874186307192, + 0.004368487745523453, + 0.008823341690003872, + 0.006283079273998737, + 0.033981747925281525, + 0.02536345086991787, + 0.010889394208788872, + 0.010688398033380508, + 0.013873936608433723, + 0.011759515851736069, + 0.007058498915284872, + 0.00419126870110631, + 0.03941258788108826, + 0.022620348259806633, + 0.006134634371846914, + 0.006024103611707687, + 0.0197757501155138, + 0.010770403780043125, + 0.01043885201215744, + 0.013688975013792515, + 0.012989498674869537, + 0.013545402325689793, + 0.010335762985050678, + 0.035965945571660995, + 0.012305570766329765, + 0.011413820087909698, + 0.01999390870332718, + 0.02010657824575901, + 0.015061801299452782, + 0.012486540712416172, + 0.010840858332812786, + 0.009938723407685757, + 0.01921757310628891, + 0.020524591207504272, + 0.004995186813175678, + 0.005617598071694374, + 0.005417199339717627, + 0.007123797666281462, + 0.004863335285335779, + 0.004914473742246628, + 0.007899453863501549, + 0.005628242623060942, + 0.0028997217305004597, + 0.003330404171720147, + 0.0017144110752269626, + 0.002784512937068939, + 0.006087384186685085, + 0.002607188653200865, + 0.005048381630331278, + 0.0024985691998153925, + 0.004178300965577364, + 0.003212106879800558, + 0.003163361456245184, + 0.002751834224909544, + 0.0038984552957117558, + 0.005142447538673878, + 0.004945109598338604, + 0.005711907986551523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.2813476324081421 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, eye-level wide-angle tracking shot, filmed in visceral, high-frame slow motion in the style of motion photography, follows the raw, destructive force of a massive tornado as it carves a path through a dense forest of mature maples. The tornado itself is a dark, roiling funnel of dust, soil, and debris, its form silhouetted eerily against the ominous, bruised purple and green storm sky. As the vortex makes contact with the trees, the air erupts into a chaotic whirlwind, and vibrant autumn colors—fiery reds, vibrant oranges, and golden yellows—are violently torn from branches and pulled into its swirling maw. The camera meticulously tracks the wind's immense force as it first bends the sturdy, thick trunks to the breaking point. In hyper-detailed slow motion, the bark begins to peel and splinter, then the wood fibers themselves tear apart in a visceral display of natural strength, sending a torrent of wood splinters through the swirling chaos, creating a terrifying and awe-inspiring spectacle of nature's power.", + "seed": 639, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_9873_seed_639_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.27273509129881857, + "vlm_logprob": -0.27273509129881857, + "vlm_yes_no_margin": 1.6, + "vlm_token_credit": -0.27273509129881857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10340095311403275 + ], + "opposite_logprob_tokens": [ + -2.353400945663452 + ], + "logprob_reward": -0.10340095311403275, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10340095311403275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16886065900325775, + 0.05837959423661232, + 0.06631791591644287, + 0.04385184496641159, + 0.041584908962249756, + 0.03887510299682617, + 0.07921939343214035, + 0.11916057765483856, + 0.04026976227760315, + 0.3434801995754242 + ], + "credit_map": [ + 0.015992671251296997, + 0.01518949307501316, + 0.02528129145503044, + 0.02126368321478367, + 0.029541580006480217, + 0.01195839699357748, + 0.03675524890422821, + 0.01287829503417015, + 0.006811174564063549, + 0.006655626464635134, + 0.008077485486865044, + 0.005597240291535854, + 0.006879298482090235, + 0.009558970108628273, + 0.007762041408568621, + 0.007037760689854622, + 0.005731862969696522, + 0.008473900146782398, + 0.009087972342967987, + 0.006146501749753952, + 0.0066211470402777195, + 0.012398187071084976, + 0.008899454027414322, + 0.008958890102803707, + 0.0037527980748564005, + 0.004968324210494757, + 0.005553197115659714, + 0.0033134755212813616, + 0.005452840588986874, + 0.008576617576181889, + 0.005484027788043022, + 0.006750563159584999, + 0.003520313650369644, + 0.004001378547400236, + 0.0036861475091427565, + 0.004084710031747818, + 0.005279401782900095, + 0.009781202301383018, + 0.003307737410068512, + 0.007924016565084457, + 0.003733932040631771, + 0.0036143192555755377, + 0.004809508100152016, + 0.005092649254947901, + 0.004361617844551802, + 0.009407797828316689, + 0.004491901490837336, + 0.003363375784829259, + 0.003690046025440097, + 0.0037788376212120056, + 0.004741135518997908, + 0.02687801420688629, + 0.00456636818125844, + 0.016496112570166588, + 0.014685261994600296, + 0.004383620340377092, + 0.012446465902030468, + 0.007568887900561094, + 0.026435384526848793, + 0.011629030108451843, + 0.01028695609420538, + 0.010720496065914631, + 0.008950410410761833, + 0.03112294152379036, + 0.003508651163429022, + 0.005002719350159168, + 0.006330317351967096, + 0.008295894600450993, + 0.003666420467197895, + 0.0044938428327441216, + 0.004699737299233675, + 0.004272182937711477, + 0.023004552349448204, + 0.019782843068242073, + 0.041766513139009476, + 0.12445583194494247, + 0.08309364318847656, + 0.02502318285405636, + 0.015302256681025028, + 0.011051377281546593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.10340095311403275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632823646068573 + ], + "opposite_logprob_tokens": [ + -0.757823646068573 + ], + "logprob_reward": -0.632823646068573, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632823646068573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23570720851421356, + 0.061210472136735916, + 0.04457172751426697, + 0.03361429274082184, + 0.02291652001440525, + 0.0362568162381649, + 0.059378426522016525, + 0.05724548175930977, + 0.035548605024814606, + 0.4135504364967346 + ], + "credit_map": [ + 0.020121360197663307, + 0.013121013529598713, + 0.02281392738223076, + 0.014035193249583244, + 0.055752094835042953, + 0.020536670461297035, + 0.07273553311824799, + 0.016591407358646393, + 0.003182793501764536, + 0.005075227469205856, + 0.0029995592776685953, + 0.002753390232101083, + 0.007456976920366287, + 0.004606999922543764, + 0.02046143263578415, + 0.014674090780317783, + 0.003294950118288398, + 0.00902462936937809, + 0.005327505525201559, + 0.004997251555323601, + 0.005373066756874323, + 0.00591137632727623, + 0.0055313436314463615, + 0.0051116012036800385, + 0.0037062668707221746, + 0.0039847297593951225, + 0.003925564233213663, + 0.0017334946896880865, + 0.005494596436619759, + 0.006384124513715506, + 0.00364927900955081, + 0.004736237693578005, + 0.0013902845093980432, + 0.0015048335772007704, + 0.001981996465474367, + 0.0022994789760559797, + 0.0036568711511790752, + 0.005071595311164856, + 0.0016210126923397183, + 0.005390445701777935, + 0.006657572463154793, + 0.004436790011823177, + 0.005033187568187714, + 0.003769251750782132, + 0.003408663673326373, + 0.007064884062856436, + 0.0028844703920185566, + 0.0030019960831850767, + 0.003518729005008936, + 0.003163393819704652, + 0.006997169926762581, + 0.014846074394881725, + 0.003190054791048169, + 0.015962406992912292, + 0.008975712582468987, + 0.0027248836122453213, + 0.005191789008677006, + 0.00566380238160491, + 0.009578173980116844, + 0.005371278617531061, + 0.0046429140493273735, + 0.008414584212005138, + 0.0050505418330430984, + 0.013332394883036613, + 0.004365789704024792, + 0.004606164991855621, + 0.006438057404011488, + 0.004490754101425409, + 0.0048204269260168076, + 0.0029986512381583452, + 0.003720448352396488, + 0.004108311142772436, + 0.027675362303853035, + 0.02332347445189953, + 0.04910118505358696, + 0.14988359808921814, + 0.09799563884735107, + 0.033394210040569305, + 0.018164370208978653, + 0.01401259284466505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.632823646068573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604028344154358 + ], + "opposite_logprob_tokens": [ + -1.910402774810791 + ], + "logprob_reward": -0.1604028344154358, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604028344154358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35947442054748535, + 0.048373546451330185, + 0.08412247896194458, + 0.05932125076651573, + 0.04506721347570419, + 0.03943609446287155, + 0.06336625665426254, + 0.0672883689403534, + 0.033785369247198105, + 0.19976496696472168 + ], + "credit_map": [ + 0.03053726814687252, + 0.02315409481525421, + 0.02056247554719448, + 0.026908157393336296, + 0.08540579676628113, + 0.022539153695106506, + 0.12321434915065765, + 0.027153154835104942, + 0.003958623390644789, + 0.00568838557228446, + 0.004284895956516266, + 0.006168430671095848, + 0.006372583564370871, + 0.007127803284674883, + 0.008800028823316097, + 0.0059727998450398445, + 0.008240488357841969, + 0.015018939971923828, + 0.008869101293385029, + 0.008791702799499035, + 0.009404354728758335, + 0.013833360746502876, + 0.01084362156689167, + 0.009120911359786987, + 0.004284778144210577, + 0.005899698939174414, + 0.007988727651536465, + 0.0040133558213710785, + 0.007895365357398987, + 0.011817232705652714, + 0.008935416117310524, + 0.008486681617796421, + 0.0040683019906282425, + 0.005647909361869097, + 0.004690292291343212, + 0.00574356596916914, + 0.005709126591682434, + 0.005144841503351927, + 0.007974252104759216, + 0.006088927388191223, + 0.006072698626667261, + 0.005280764307826757, + 0.00707218749448657, + 0.004780490417033434, + 0.0033681835047900677, + 0.005544880870729685, + 0.0038173249922692776, + 0.0034995684400200844, + 0.004606529138982296, + 0.004178202711045742, + 0.006860565394163132, + 0.018727479502558708, + 0.004358818754553795, + 0.012969933450222015, + 0.007421509362757206, + 0.0042432257905602455, + 0.008490517735481262, + 0.006421742495149374, + 0.0072397710755467415, + 0.010112611576914787, + 0.009095323272049427, + 0.007628387305885553, + 0.008858365938067436, + 0.009441656991839409, + 0.004324671346694231, + 0.004045491572469473, + 0.0036836236249655485, + 0.0036397981457412243, + 0.00410934304818511, + 0.005460171960294247, + 0.003937798552215099, + 0.004584472626447678, + 0.014357205480337143, + 0.011946628801524639, + 0.025271430611610413, + 0.06553802639245987, + 0.04641343653202057, + 0.01575784757733345, + 0.011214802041649818, + 0.009265600703656673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.1604028344154358 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038114283233881 + ], + "opposite_logprob_tokens": [ + -3.288114309310913 + ], + "logprob_reward": -0.038114283233881, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038114283233881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2284150868654251, + 0.07663807272911072, + 0.09154218435287476, + 0.13731765747070312, + 0.05279489979147911, + 0.06635681539773941, + 0.043623004108667374, + 0.07913350313901901, + 0.023038843646645546, + 0.20113983750343323 + ], + "credit_map": [ + 0.02678067237138748, + 0.02191499061882496, + 0.046711936593055725, + 0.031721971929073334, + 0.0332491509616375, + 0.022521939128637314, + 0.02687891572713852, + 0.018635524436831474, + 0.007895167917013168, + 0.008126473054289818, + 0.0065489234402775764, + 0.006882015150040388, + 0.005778730381280184, + 0.0067533450201153755, + 0.019856423139572144, + 0.014797001145780087, + 0.008881139568984509, + 0.014951343648135662, + 0.009417135268449783, + 0.009862305596470833, + 0.011296972632408142, + 0.014267812483012676, + 0.0125079695135355, + 0.010357517749071121, + 0.015004276297986507, + 0.021126404404640198, + 0.015973955392837524, + 0.009094571694731712, + 0.02345879375934601, + 0.023637909442186356, + 0.01366607565432787, + 0.015355671755969524, + 0.004216648172587156, + 0.006612660828977823, + 0.005244675558060408, + 0.006281089968979359, + 0.008066736161708832, + 0.010326567105948925, + 0.005619695410132408, + 0.006426834035664797, + 0.0076584164053201675, + 0.006123439408838749, + 0.009731640107929707, + 0.00878289807587862, + 0.008964051492512226, + 0.007699196692556143, + 0.011266794987022877, + 0.006130384746938944, + 0.0029079450760036707, + 0.0017482705879956484, + 0.005445954389870167, + 0.008387453854084015, + 0.0024732856545597315, + 0.01093879621475935, + 0.009255294688045979, + 0.0024660094641149044, + 0.007718831766396761, + 0.008790692314505577, + 0.011467932723462582, + 0.010334552265703678, + 0.008802743628621101, + 0.008646789938211441, + 0.009206072427332401, + 0.01416589505970478, + 0.00342718418687582, + 0.0028715310618281364, + 0.004387340974062681, + 0.003470370313152671, + 0.0016580186784267426, + 0.002107839100062847, + 0.0024381219409406185, + 0.0026784390211105347, + 0.012875862419605255, + 0.011355843394994736, + 0.022087911143898964, + 0.06671302765607834, + 0.043286070227622986, + 0.022297071292996407, + 0.012627179734408855, + 0.00989691074937582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.038114283233881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289337396621704 + ], + "opposite_logprob_tokens": [ + -1.0539337396621704 + ], + "logprob_reward": -0.4289337396621704, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289337396621704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12154220789670944, + 0.041403889656066895, + 0.053405601531267166, + 0.03290602192282677, + 0.051741842180490494, + 0.02434614486992359, + 0.1463521420955658, + 0.03393027186393738, + 0.019419550895690918, + 0.4749522805213928 + ], + "credit_map": [ + 0.01259934064000845, + 0.00859079509973526, + 0.009469570592045784, + 0.006946763955056667, + 0.028110196813941002, + 0.009599214419722557, + 0.03840318322181702, + 0.007823164574801922, + 0.0027165834326297045, + 0.0029526937287300825, + 0.0034267196897417307, + 0.002965606516227126, + 0.0037229983136057854, + 0.003541660727933049, + 0.012168038636446, + 0.009909594431519508, + 0.00535653717815876, + 0.0074087283574044704, + 0.0062144105322659016, + 0.005511582363396883, + 0.005155669525265694, + 0.010234123095870018, + 0.0069665019400417805, + 0.006558052729815245, + 0.004601175431162119, + 0.005793152842670679, + 0.003673663130030036, + 0.002024636138230562, + 0.004689447116106749, + 0.004304289817810059, + 0.003861465025693178, + 0.003958197310566902, + 0.003906363621354103, + 0.004918342921882868, + 0.004794469103217125, + 0.005556467920541763, + 0.00673649413511157, + 0.012387904338538647, + 0.004126992076635361, + 0.00931481271982193, + 0.0028814966790378094, + 0.001696665189228952, + 0.0030005972366780043, + 0.0027672313153743744, + 0.0036928250920027494, + 0.005964189767837524, + 0.0024676057510077953, + 0.0018755372148007154, + 0.007688171695917845, + 0.008294911123812199, + 0.008691759780049324, + 0.05222230404615402, + 0.008856372907757759, + 0.0058873179368674755, + 0.0478227473795414, + 0.006888553500175476, + 0.003467060159891844, + 0.004087184090167284, + 0.004185056779533625, + 0.003297700546681881, + 0.0029241025913506746, + 0.004245458636432886, + 0.004013757221400738, + 0.007709953002631664, + 0.0020056087523698807, + 0.002008989453315735, + 0.003509379690513015, + 0.0020472006872296333, + 0.0019362352322787046, + 0.0029317610897123814, + 0.002073434414342046, + 0.0029069434385746717, + 0.03234165161848068, + 0.025900503620505333, + 0.056741438806056976, + 0.17260727286338806, + 0.115199014544487, + 0.03516654297709465, + 0.0221853107213974, + 0.01481057982891798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.4289337396621704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, eye-level wide-angle tracking shot, filmed in visceral, high-frame slow motion in the style of motion photography, follows the raw, destructive force of a massive tornado as it carves a path through a dense forest of mature maples. The tornado itself is a dark, roiling funnel of dust, soil, and debris, its form silhouetted eerily against the ominous, bruised purple and green storm sky. As the vortex makes contact with the trees, the air erupts into a chaotic whirlwind, and vibrant autumn colors—fiery reds, vibrant oranges, and golden yellows—are violently torn from branches and pulled into its swirling maw. The camera meticulously tracks the wind's immense force as it first bends the sturdy, thick trunks to the breaking point. In hyper-detailed slow motion, the bark begins to peel and splinter, then the wood fibers themselves tear apart in a visceral display of natural strength, sending a torrent of wood splinters through the swirling chaos, creating a terrifying and awe-inspiring spectacle of nature's power.", + "seed": 640, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_1/index_9873_seed_640_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15488542150706053, + "vlm_logprob": -0.15488542150706053, + "vlm_yes_no_margin": 2.424999988079071, + "vlm_token_credit": -0.15488542150706053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2575187385082245 + ], + "opposite_logprob_tokens": [ + -1.5075187683105469 + ], + "logprob_reward": -0.2575187385082245, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2575187385082245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22386735677719116, + 0.06409917771816254, + 0.045085009187459946, + 0.09842698276042938, + 0.037268929183483124, + 0.2063262164592743, + 0.03663815185427666, + 0.0922660380601883, + 0.07319911569356918, + 0.12282300740480423 + ], + "credit_map": [ + 0.03151988983154297, + 0.025632265955209732, + 0.020899835973978043, + 0.021304188296198845, + 0.022100506350398064, + 0.019458763301372528, + 0.03811175376176834, + 0.044840164482593536, + 0.004229544661939144, + 0.009347835555672646, + 0.005932332016527653, + 0.007984931580722332, + 0.007527596317231655, + 0.00481127854436636, + 0.011462565511465073, + 0.012803091667592525, + 0.0036607214715331793, + 0.004484111908823252, + 0.005312634631991386, + 0.005432030186057091, + 0.005068108439445496, + 0.0055420370772480965, + 0.007720810826867819, + 0.007864556275308132, + 0.009258212521672249, + 0.014256625436246395, + 0.012154738418757915, + 0.0111719174310565, + 0.013205248862504959, + 0.012951855547726154, + 0.01247928012162447, + 0.012949110008776188, + 0.004403047729283571, + 0.0036920751444995403, + 0.0029038030188530684, + 0.0051708221435546875, + 0.00785820372402668, + 0.004471376538276672, + 0.005405928008258343, + 0.0033636721782386303, + 0.01584557071328163, + 0.017618564888834953, + 0.021093785762786865, + 0.024259347468614578, + 0.03397541493177414, + 0.023195086047053337, + 0.043600622564554214, + 0.026737820357084274, + 0.0021056546829640865, + 0.003180619329214096, + 0.005090563558042049, + 0.003475729376077652, + 0.006056542508304119, + 0.00735249649733305, + 0.005349496845155954, + 0.004027049522846937, + 0.007897285744547844, + 0.014430356211960316, + 0.017027975991368294, + 0.008846358396112919, + 0.01466312725096941, + 0.013014719821512699, + 0.008702735416591167, + 0.007683479227125645, + 0.005770802963525057, + 0.00939320307224989, + 0.006042924709618092, + 0.014832702465355396, + 0.010744052939116955, + 0.00919028464704752, + 0.011070150882005692, + 0.006154995411634445, + 0.009799149818718433, + 0.01772087998688221, + 0.0151866115629673, + 0.02353418618440628, + 0.016115106642246246, + 0.014463461004197598, + 0.01347229816019535, + 0.012531310319900513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.2575187385082245 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869088530540466 + ], + "opposite_logprob_tokens": [ + -3.0486907958984375 + ], + "logprob_reward": -0.04869088530540466, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869088530540466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14003610610961914, + 0.14243724942207336, + 0.028650231659412384, + 0.07091433554887772, + 0.054708268493413925, + 0.03776164352893829, + 0.03550831601023674, + 0.06906750053167343, + 0.26155906915664673, + 0.15935729444026947 + ], + "credit_map": [ + 0.016239233314990997, + 0.020072882995009422, + 0.01720220223069191, + 0.01617354527115822, + 0.011027119122445583, + 0.011204667389392853, + 0.02491873688995838, + 0.023197710514068604, + 0.008724482730031013, + 0.01604405604302883, + 0.01708196848630905, + 0.017625795677304268, + 0.014896035194396973, + 0.012403206899762154, + 0.026426605880260468, + 0.029235102236270905, + 0.0022550122812390327, + 0.003036804497241974, + 0.003866627113893628, + 0.003662688657641411, + 0.0024337240029126406, + 0.0035357384476810694, + 0.003474058350548148, + 0.00638557830825448, + 0.006627252791076899, + 0.010234382934868336, + 0.010321218520402908, + 0.009299390017986298, + 0.009277711622416973, + 0.007994611747562885, + 0.00901295430958271, + 0.008146816864609718, + 0.003762683365494013, + 0.004412533715367317, + 0.004553086590021849, + 0.007751555182039738, + 0.01339736394584179, + 0.008103679865598679, + 0.007062182296067476, + 0.005665183067321777, + 0.0035597840324044228, + 0.0035528007429093122, + 0.003850259119644761, + 0.003937572240829468, + 0.005130124744027853, + 0.0076497080735862255, + 0.006446050480008125, + 0.0036353387404233217, + 0.0023074625059962273, + 0.003748617134988308, + 0.004617394879460335, + 0.0036228897515684366, + 0.007140111643821001, + 0.005694798193871975, + 0.0044214618392288685, + 0.003955577500164509, + 0.006011942867189646, + 0.008323261514306068, + 0.009598293341696262, + 0.0063613285310566425, + 0.014650469645857811, + 0.009981978684663773, + 0.007197486702352762, + 0.0069427466951310635, + 0.020681677386164665, + 0.02136744186282158, + 0.032853636890649796, + 0.02967275306582451, + 0.03656214475631714, + 0.04720866307616234, + 0.05133495107293129, + 0.021877804771065712, + 0.013840306550264359, + 0.01693481020629406, + 0.026444995775818825, + 0.016167763620615005, + 0.024596890434622765, + 0.025068743154406548, + 0.019925596192479134, + 0.016378184780478477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.04869088530540466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006660848855972 + ], + "opposite_logprob_tokens": [ + -2.6950666904449463 + ], + "logprob_reward": -0.07006660848855972, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006660848855972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1390247941017151, + 0.1192852184176445, + 0.0424787737429142, + 0.2008337378501892, + 0.025411764159798622, + 0.07801032811403275, + 0.08282914757728577, + 0.0932074561715126, + 0.15349064767360687, + 0.06542818248271942 + ], + "credit_map": [ + 0.0156732015311718, + 0.0188831128180027, + 0.01671145111322403, + 0.01707916334271431, + 0.010980661027133465, + 0.012074297294020653, + 0.0245807021856308, + 0.02304219827055931, + 0.008007439784705639, + 0.01331225037574768, + 0.012247792445123196, + 0.01449623890221119, + 0.012423324398696423, + 0.011882108636200428, + 0.020055828616023064, + 0.026860222220420837, + 0.0035509448498487473, + 0.004684991668909788, + 0.0062162973918020725, + 0.00803830474615097, + 0.004834866151213646, + 0.0034203154500573874, + 0.005417299456894398, + 0.006315754726529121, + 0.01627219282090664, + 0.03524879738688469, + 0.026791075244545937, + 0.02447742410004139, + 0.0289971474558115, + 0.022728944197297096, + 0.022831283509731293, + 0.02348685823380947, + 0.0025346344336867332, + 0.0023830474819988012, + 0.0018213001312687993, + 0.0022810972295701504, + 0.005601171404123306, + 0.0034740183036774397, + 0.003954531624913216, + 0.0033619613386690617, + 0.005583393387496471, + 0.005966056603938341, + 0.009718907065689564, + 0.009521632455289364, + 0.015965169295668602, + 0.009028260596096516, + 0.013586808927357197, + 0.008640101179480553, + 0.005601322278380394, + 0.0071647511795163155, + 0.010383988730609417, + 0.007306029554456472, + 0.014148346148431301, + 0.01841130293905735, + 0.00994760449975729, + 0.009865798987448215, + 0.007776537910103798, + 0.016131138429045677, + 0.018191898241639137, + 0.009659952484071255, + 0.0118899866938591, + 0.01296259369701147, + 0.007964839227497578, + 0.008630502037703991, + 0.011736069805920124, + 0.021972456946969032, + 0.01224395353347063, + 0.030908601358532906, + 0.022654104977846146, + 0.018541043624281883, + 0.025362741202116013, + 0.010071654804050922, + 0.0055183423683047295, + 0.008940189145505428, + 0.00539771281182766, + 0.010385174304246902, + 0.010328005068004131, + 0.009221253916621208, + 0.008905679918825626, + 0.006731818895787001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.07006660848855972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080636642873287 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080636642873287, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080636642873287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12246423959732056, + 0.06291943788528442, + 0.15742698311805725, + 0.09626949578523636, + 0.04227021709084511, + 0.11288470029830933, + 0.02376042865216732, + 0.07335659861564636, + 0.2622843086719513, + 0.04636356979608536 + ], + "credit_map": [ + 0.01650504767894745, + 0.01540988590568304, + 0.011304410174489021, + 0.011053601279854774, + 0.014586841687560081, + 0.009715728461742401, + 0.018705245107412338, + 0.02518349327147007, + 0.004215672612190247, + 0.00588681036606431, + 0.0051363748498260975, + 0.008246333338320255, + 0.005112775601446629, + 0.006061451975256205, + 0.01492687501013279, + 0.013333159498870373, + 0.017983706668019295, + 0.015328348614275455, + 0.014689786359667778, + 0.03889427334070206, + 0.012494713068008423, + 0.017880499362945557, + 0.023018095642328262, + 0.01713758148252964, + 0.00826490018516779, + 0.014270502142608166, + 0.010950127616524696, + 0.01038576290011406, + 0.01222695130854845, + 0.017593031749129295, + 0.010483327321708202, + 0.01209490280598402, + 0.005148956552147865, + 0.003941211849451065, + 0.00411658501252532, + 0.004650391172617674, + 0.006459060590714216, + 0.0035231178626418114, + 0.009156639687716961, + 0.005274258553981781, + 0.009949457831680775, + 0.008966479450464249, + 0.010199597105383873, + 0.010410415008664131, + 0.017920389771461487, + 0.01810253970324993, + 0.02483595348894596, + 0.012499884702265263, + 0.0013077636249363422, + 0.0025700803380459547, + 0.0027731468435376883, + 0.0027641316410154104, + 0.0027785361744463444, + 0.006057012360543013, + 0.0026466427370905876, + 0.0028631172608584166, + 0.006108677946031094, + 0.008973784744739532, + 0.012310471385717392, + 0.006551380269229412, + 0.015392418019473553, + 0.009770534932613373, + 0.007586571387946606, + 0.006662764586508274, + 0.020717062056064606, + 0.02347785048186779, + 0.035603661090135574, + 0.0384388193488121, + 0.04193132743239403, + 0.03640156239271164, + 0.03808698430657387, + 0.02762703411281109, + 0.0038300445303320885, + 0.006998637691140175, + 0.0041474755853414536, + 0.007869535125792027, + 0.006230405997484922, + 0.0061555844731628895, + 0.005405884236097336, + 0.005726006347686052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.011080636642873287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense forest of mature maple trees with vibrant autumn colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive tornado visible as a dark, roiling funnel of dust, soil, and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees shown bending to the breaking point and having bark peel and splinter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky depicted as ominous with bruised purple and green colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed as an eye-level wide-angle tracking shot following the tornado?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870702385902405 + ], + "opposite_logprob_tokens": [ + -1.1370701789855957 + ], + "logprob_reward": -0.3870702385902405, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3870702385902405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24705520272254944, + 0.22066636383533478, + 0.029374852776527405, + 0.029042880982160568, + 0.023689640685915947, + 0.07876060158014297, + 0.04497722536325455, + 0.08620857447385788, + 0.17496316134929657, + 0.06526152044534683 + ], + "credit_map": [ + 0.0354660227894783, + 0.03320273756980896, + 0.03578570485115051, + 0.02671075239777565, + 0.018735861405730247, + 0.02106323279440403, + 0.035066451877355576, + 0.04102443531155586, + 0.011949063278734684, + 0.017274584621191025, + 0.020332861691713333, + 0.031063534319400787, + 0.017651984468102455, + 0.021143749356269836, + 0.03915068134665489, + 0.062099892646074295, + 0.0038334205746650696, + 0.0036143441684544086, + 0.004236103501170874, + 0.005585989914834499, + 0.003600599244236946, + 0.0021359813399612904, + 0.0032290706876665354, + 0.003139343811199069, + 0.0019774220418184996, + 0.0030704878736287355, + 0.004342792090028524, + 0.0035604629665613174, + 0.0040710726752877235, + 0.004249777179211378, + 0.0041160378605127335, + 0.0036548275966197252, + 0.0016508660046383739, + 0.0019600745290517807, + 0.0026076578069478273, + 0.0029462354723364115, + 0.00590655580163002, + 0.003174396464601159, + 0.003207382746040821, + 0.0022364717442542315, + 0.005556220654398203, + 0.0072555942460894585, + 0.006083126179873943, + 0.012808081693947315, + 0.013015844859182835, + 0.012105615809559822, + 0.012395855970680714, + 0.009540267288684845, + 0.0034330948255956173, + 0.005279268138110638, + 0.00611736997961998, + 0.0029362868517637253, + 0.010067901574075222, + 0.006711462512612343, + 0.005644703283905983, + 0.0047871386632323265, + 0.0076072183437645435, + 0.013739360496401787, + 0.014148678630590439, + 0.00824093446135521, + 0.013416577130556107, + 0.012456593103706837, + 0.008938085287809372, + 0.007661125622689724, + 0.013673999346792698, + 0.02288692630827427, + 0.012701518833637238, + 0.04312657192349434, + 0.017847908660769463, + 0.021101858466863632, + 0.03276415914297104, + 0.010860213078558445, + 0.0062132240273058414, + 0.0076655675657093525, + 0.010245930403470993, + 0.009670674800872803, + 0.00936952792108059, + 0.008057286031544209, + 0.00792625080794096, + 0.006113053299486637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.3870702385902405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_tensor([3947, 3947])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_tensor([3947, 3947])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..deef9546f21b945dd06c08a1ed861131a4a42f1f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_tensor([3947, 3947])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Recorded in a simple yet expert style, a low-angle, wide-angle shot pans smoothly to the right, tracking the powerful takeoff roll of a massive C-17 Globemaster III cargo plane. Its massive fuselage, painted in a non-reflective matte gray with visible panel lines, fills the frame as it roars along the long concrete runway marked with faded white lines. Four powerful turbofan engines, suspended high beneath the wings, roar with incredible force, visibly distorting the air behind them and emitting a shimmering heat haze under the harsh light of the clear midday sun. The camera maintains a perfect frame of the fuselage as the landing gear tires spin blurringly, building immense speed until the nose wheel decisively lifts off the tarmac. The entire aircraft rotates steadily upward, its massive form outlined against the clear blue sky, before the main landing gear leaves the ground and begins a steep, powerful ascent, creating an atmosphere of immense mechanical power and operational precision.", + "seed": 30092, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_3947_seed_30092_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01890716850757599, + "vlm_logprob": -0.01890716850757599, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.01890716850757599, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046649783849716187 + ], + "opposite_logprob_tokens": [ + -3.296649694442749 + ], + "logprob_reward": -0.046649783849716187, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.046649783849716187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06609612703323364, + 0.023133425042033195, + 0.02598758041858673, + 0.04591855779290199, + 0.02486031875014305, + 0.024583620950579643, + 0.010849226266145706, + 0.013336838223040104, + 0.7209537625312805, + 0.04428047314286232 + ], + "credit_map": [ + 0.008659840561449528, + 0.009358742274343967, + 0.007181384600698948, + 0.007173541933298111, + 0.0082966648042202, + 0.010710629634559155, + 0.008567987941205502, + 0.006147340405732393, + 0.002586311660706997, + 0.002814393024891615, + 0.003536919364705682, + 0.0022997520864009857, + 0.003037883434444666, + 0.0033189631067216396, + 0.0022040933836251497, + 0.0033351087477058172, + 0.0057588424533605576, + 0.004557778127491474, + 0.0027708786074072123, + 0.0011703204363584518, + 0.0027477096300572157, + 0.004201437346637249, + 0.0015097787836566567, + 0.0032708351500332355, + 0.006609088741242886, + 0.005447853356599808, + 0.004434916656464338, + 0.0037832933012396097, + 0.005698718596249819, + 0.009014778770506382, + 0.006601681932806969, + 0.004328226204961538, + 0.0031653568148612976, + 0.003488871268928051, + 0.0028621514793485403, + 0.002713832538574934, + 0.0033189274836331606, + 0.002947380533441901, + 0.0037401472218334675, + 0.002623653272166848, + 0.002463444834575057, + 0.0025925482623279095, + 0.002194299828261137, + 0.002010887488722801, + 0.005824950989335775, + 0.005413554143160582, + 0.0021841218695044518, + 0.0018998143495991826, + 0.0013838295126333833, + 0.0015993635170161724, + 0.0010921286884695292, + 0.0012150410329923034, + 0.0018641846254467964, + 0.001968235010281205, + 0.0010150405578315258, + 0.0007114027976058424, + 0.0014956308295950294, + 0.0023237913846969604, + 0.0013753361999988556, + 0.0016107368282973766, + 0.0019024212379008532, + 0.0019181258976459503, + 0.0010030785342678428, + 0.0017077173106372356, + 0.035460397601127625, + 0.018872680142521858, + 0.17306281626224518, + 0.038237545639276505, + 0.11886844038963318, + 0.27672049403190613, + 0.017077889293432236, + 0.04265354201197624, + 0.0037215175107121468, + 0.002563543152064085, + 0.007945415563881397, + 0.004226371180266142, + 0.004162267316132784, + 0.0063052307814359665, + 0.005737964063882828, + 0.009618164040148258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.046649783849716187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610919088125229 + ], + "opposite_logprob_tokens": [ + -4.141108989715576 + ], + "logprob_reward": -0.01610919088125229, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610919088125229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06812448054552078, + 0.03912309184670448, + 0.0273857694119215, + 0.05882582068443298, + 0.022275593131780624, + 0.015690945088863373, + 0.02128804475069046, + 0.02193444035947323, + 0.6705253720283508, + 0.054826412349939346 + ], + "credit_map": [ + 0.008215715177357197, + 0.0086005087941885, + 0.009144730865955353, + 0.008084933273494244, + 0.008781414479017258, + 0.009434403851628304, + 0.008547545410692692, + 0.007315229624509811, + 0.005988783668726683, + 0.0049664624966681, + 0.004441540688276291, + 0.004600935615599155, + 0.00669053802266717, + 0.004678552504628897, + 0.0037336740642786026, + 0.004022602923214436, + 0.003472617594525218, + 0.005240974947810173, + 0.002365482971072197, + 0.0016393527621403337, + 0.0031881786417216063, + 0.006464310456067324, + 0.0026238143909722567, + 0.002391038928180933, + 0.007897666655480862, + 0.007479056250303984, + 0.005864865146577358, + 0.005303879734128714, + 0.006917486432939768, + 0.01125888992100954, + 0.007812525145709515, + 0.0062914518639445305, + 0.003351640421897173, + 0.0032100577373057604, + 0.0027126644272357225, + 0.001887042890302837, + 0.003590785665437579, + 0.002461837837472558, + 0.002806193195283413, + 0.002255370607599616, + 0.001977061154320836, + 0.0021846124436706305, + 0.0019660252146422863, + 0.0015347860753536224, + 0.0025855950079858303, + 0.002073579467833042, + 0.0018026875331997871, + 0.0015665990067645907, + 0.0017074267379939556, + 0.0023433405440300703, + 0.0035314634442329407, + 0.0015658586053177714, + 0.004462675657123327, + 0.004220485687255859, + 0.0013240247499197721, + 0.002132768975570798, + 0.003205046523362398, + 0.0041541182436048985, + 0.0025901098269969225, + 0.0027244864031672478, + 0.002459703478962183, + 0.0025983178056776524, + 0.0013851134572178125, + 0.0028175448533147573, + 0.032893043011426926, + 0.017099902033805847, + 0.16114841401576996, + 0.03519481420516968, + 0.11078459024429321, + 0.2573784589767456, + 0.016041377559304237, + 0.03998478129506111, + 0.004520117770880461, + 0.0035489352885633707, + 0.009198772720992565, + 0.005927780643105507, + 0.0046859909780323505, + 0.007049937266856432, + 0.006604288704693317, + 0.01329058688133955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.01610919088125229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259012337774038 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005259012337774038, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259012337774038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16372130811214447, + 0.052927128970623016, + 0.06923095881938934, + 0.07857530564069748, + 0.13341914117336273, + 0.025439836084842682, + 0.04247969016432762, + 0.16939789056777954, + 0.2275722622871399, + 0.037236426025629044 + ], + "credit_map": [ + 0.03405659645795822, + 0.029207563027739525, + 0.030022013932466507, + 0.01110018603503704, + 0.010392989963293076, + 0.023712679743766785, + 0.0072499229572713375, + 0.017979372292757034, + 0.006597655825316906, + 0.007013435009866953, + 0.006716791540384293, + 0.0052883205935359, + 0.007137551438063383, + 0.007436530198901892, + 0.004744251724332571, + 0.007992600090801716, + 0.01381056196987629, + 0.010309659875929356, + 0.00601264787837863, + 0.004370961803942919, + 0.008170644752681255, + 0.011559993028640747, + 0.005683799274265766, + 0.009312698617577553, + 0.010896551422774792, + 0.008729303255677223, + 0.011121056973934174, + 0.006106749642640352, + 0.008009049110114574, + 0.014796138741075993, + 0.008510089479386806, + 0.010406377725303173, + 0.017828190699219704, + 0.01653497852385044, + 0.02049924060702324, + 0.013738865964114666, + 0.019135411828756332, + 0.016158055514097214, + 0.013607325032353401, + 0.015917088836431503, + 0.0031089158728718758, + 0.002847611438483, + 0.003732136217877269, + 0.002104105893522501, + 0.0034343041479587555, + 0.0038638589903712273, + 0.0042725675739347935, + 0.002076340839266777, + 0.004527431912720203, + 0.006438538432121277, + 0.006538244895637035, + 0.003827191423624754, + 0.007154140155762434, + 0.006890980061143637, + 0.002744726138189435, + 0.004358440637588501, + 0.01755201816558838, + 0.019516464322805405, + 0.013027504086494446, + 0.02112545259296894, + 0.03726232796907425, + 0.03368983790278435, + 0.011090036481618881, + 0.01613426022231579, + 0.013667029328644276, + 0.008808336220681667, + 0.047954361885786057, + 0.014171767048537731, + 0.034201499074697495, + 0.07466838508844376, + 0.009890847839415073, + 0.024210063740611076, + 0.00393577478826046, + 0.0023880950175225735, + 0.006801900919526815, + 0.00394727336242795, + 0.0036868148017674685, + 0.006098656915128231, + 0.0041628265753388405, + 0.006215086206793785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.005259012337774038 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943956319242716 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943956319242716, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943956319242716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07555275410413742, + 0.0897921621799469, + 0.031413111835718155, + 0.03717545419931412, + 0.058144472539424896, + 0.041662294417619705, + 0.027158956974744797, + 0.014341081492602825, + 0.6022692322731018, + 0.022490546107292175 + ], + "credit_map": [ + 0.010257350280880928, + 0.012558907270431519, + 0.01183256320655346, + 0.006996233947575092, + 0.007575782015919685, + 0.010711274109780788, + 0.006838147062808275, + 0.008782492950558662, + 0.010559966787695885, + 0.01205801498144865, + 0.01911047101020813, + 0.009080789051949978, + 0.011303435079753399, + 0.011630098335444927, + 0.0062067243270576, + 0.009842662140727043, + 0.004820660222321749, + 0.004673745017498732, + 0.0031278315000236034, + 0.0017626873450353742, + 0.0053640794940292835, + 0.005383829586207867, + 0.0025013517588377, + 0.003778928890824318, + 0.007363773416727781, + 0.004491134081035852, + 0.005328511353582144, + 0.0026249573566019535, + 0.0036109574139118195, + 0.006654635537415743, + 0.0028249730821698904, + 0.00427651172503829, + 0.006655927747488022, + 0.00894848071038723, + 0.007145341020077467, + 0.006316769868135452, + 0.009434372186660767, + 0.007551135495305061, + 0.006474677007645369, + 0.005617773625999689, + 0.004875763785094023, + 0.006318673025816679, + 0.004029873292893171, + 0.0037321436684578657, + 0.008288153447210789, + 0.006954958662390709, + 0.0039664930664002895, + 0.0034962377976626158, + 0.002670065965503454, + 0.003535286523401737, + 0.0037564707454293966, + 0.004061566200107336, + 0.003719950094819069, + 0.003130950266495347, + 0.003475259291008115, + 0.0028094074223190546, + 0.0013100411742925644, + 0.0019428275991231203, + 0.002021706197410822, + 0.0016168730799108744, + 0.0019865334033966064, + 0.0018720190273597836, + 0.001359808025881648, + 0.0022312726359814405, + 0.02914694882929325, + 0.015218009240925312, + 0.14559830725193024, + 0.029485730454325676, + 0.10081548243761063, + 0.23623009026050568, + 0.012965782545506954, + 0.03280887007713318, + 0.0024983210023492575, + 0.00196169875562191, + 0.00315438536927104, + 0.0023660834413021803, + 0.001939420122653246, + 0.0028066029772162437, + 0.0026444057002663612, + 0.005119629204273224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.005943956319242716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573899149894714 + ], + "opposite_logprob_tokens": [ + -3.895573854446411 + ], + "logprob_reward": -0.020573899149894714, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573899149894714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05477754771709442, + 0.021070115268230438, + 0.02455058880150318, + 0.029469087719917297, + 0.021779341623187065, + 0.018311649560928345, + 0.03879633918404579, + 0.04459557682275772, + 0.606506884098053, + 0.14014291763305664 + ], + "credit_map": [ + 0.009891924448311329, + 0.009436484426259995, + 0.0070730154402554035, + 0.004811990074813366, + 0.004812227562069893, + 0.008171934634447098, + 0.004017457831650972, + 0.006562499329447746, + 0.003403640119358897, + 0.003021563868969679, + 0.0035402337089180946, + 0.0020917775109410286, + 0.002077233511954546, + 0.0032260152511298656, + 0.001828589360229671, + 0.0018810582114383578, + 0.0035634131636470556, + 0.004464421421289444, + 0.0017912054900079966, + 0.001650699763558805, + 0.0032412484288215637, + 0.006001108326017857, + 0.0013274986995384097, + 0.002510989550501108, + 0.0047350614331662655, + 0.0032214252278208733, + 0.004055164288729429, + 0.0016803106991574168, + 0.002562697743996978, + 0.005905972793698311, + 0.0024039645213633776, + 0.00490448297932744, + 0.003817141056060791, + 0.002365804510191083, + 0.003967869095504284, + 0.001612942898645997, + 0.002191154984757304, + 0.0032412640284746885, + 0.0023434257600456476, + 0.0022397348657250404, + 0.002029433846473694, + 0.0014254656853154302, + 0.00222032330930233, + 0.0011936486698687077, + 0.00437884358689189, + 0.0043878210708498955, + 0.001481338869780302, + 0.0011947720777243376, + 0.004203903954476118, + 0.005936244037002325, + 0.004050072282552719, + 0.004245953168720007, + 0.00755607383325696, + 0.0076439473778009415, + 0.0025221994146704674, + 0.0026379378978163004, + 0.004319948609918356, + 0.0037362126167863607, + 0.005333846900612116, + 0.004415818955749273, + 0.009886876679956913, + 0.009859687648713589, + 0.0027844097930938005, + 0.004258767701685429, + 0.03074353188276291, + 0.021756837144494057, + 0.1441696137189865, + 0.03333568572998047, + 0.09587913751602173, + 0.22546719014644623, + 0.015070773661136627, + 0.040084004402160645, + 0.011458612978458405, + 0.00891820713877678, + 0.023798039183020592, + 0.01424433384090662, + 0.011952722445130348, + 0.0200343057513237, + 0.015388155356049538, + 0.034348513931035995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.020573899149894714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Recorded in a simple yet expert style, a low-angle, wide-angle shot pans smoothly to the right, tracking the powerful takeoff roll of a massive C-17 Globemaster III cargo plane. Its massive fuselage, painted in a non-reflective matte gray with visible panel lines, fills the frame as it roars along the long concrete runway marked with faded white lines. Four powerful turbofan engines, suspended high beneath the wings, roar with incredible force, visibly distorting the air behind them and emitting a shimmering heat haze under the harsh light of the clear midday sun. The camera maintains a perfect frame of the fuselage as the landing gear tires spin blurringly, building immense speed until the nose wheel decisively lifts off the tarmac. The entire aircraft rotates steadily upward, its massive form outlined against the clear blue sky, before the main landing gear leaves the ground and begins a steep, powerful ascent, creating an atmosphere of immense mechanical power and operational precision.", + "seed": 30093, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_3947_seed_30093_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018502663634717465, + "vlm_logprob": -0.018502663634717465, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.018502663634717465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041692472994327545 + ], + "opposite_logprob_tokens": [ + -3.2916924953460693 + ], + "logprob_reward": -0.041692472994327545, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.041692472994327545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06477529555559158, + 0.03611680492758751, + 0.05983157083392143, + 0.25772207975387573, + 0.12345881760120392, + 0.08587747812271118, + 0.18025721609592438, + 0.097828708589077, + 0.04081200063228607, + 0.05332006886601448 + ], + "credit_map": [ + 0.008962352760136127, + 0.01001980621367693, + 0.0070686000399291515, + 0.007399879861623049, + 0.00998330395668745, + 0.006043591536581516, + 0.00861277710646391, + 0.0066849905997514725, + 0.005490626208484173, + 0.005769167095422745, + 0.005461219232529402, + 0.003909335471689701, + 0.0031052289996296167, + 0.002595803001895547, + 0.003378527704626322, + 0.006406897213310003, + 0.00928315706551075, + 0.010606423951685429, + 0.005599574185907841, + 0.005679982248693705, + 0.0057706329971551895, + 0.004984534811228514, + 0.00602411013096571, + 0.011883155442774296, + 0.06665842235088348, + 0.10987772047519684, + 0.00864295195788145, + 0.007904604077339172, + 0.02276509441435337, + 0.016758562996983528, + 0.008216431364417076, + 0.016898298636078835, + 0.020381027832627296, + 0.020152907818555832, + 0.013014704920351505, + 0.008746703155338764, + 0.0205447506159544, + 0.013731989078223705, + 0.010510441847145557, + 0.016376296058297157, + 0.008502267301082611, + 0.011437874287366867, + 0.009804245084524155, + 0.013261569663882256, + 0.008334717713296413, + 0.007403858937323093, + 0.01230169553309679, + 0.014831244945526123, + 0.009024403057992458, + 0.013306093402206898, + 0.038836460560560226, + 0.009457727894186974, + 0.07171431183815002, + 0.021430976688861847, + 0.0062402039766311646, + 0.01024702563881874, + 0.009496224112808704, + 0.009035369381308556, + 0.013882188126444817, + 0.013028295710682869, + 0.012589553371071815, + 0.008189728483557701, + 0.013955401256680489, + 0.01765194907784462, + 0.0027971689123660326, + 0.0018861460266634822, + 0.014598357491195202, + 0.0024867274332791567, + 0.0013982102973386645, + 0.002179044531658292, + 0.0017554302467033267, + 0.013710914179682732, + 0.0040399618446826935, + 0.003584114834666252, + 0.0023801010102033615, + 0.004675081465393305, + 0.004964465741068125, + 0.010479635559022427, + 0.01401772815734148, + 0.009178980253636837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.041692472994327545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009811745956540108 + ], + "opposite_logprob_tokens": [ + -4.634811878204346 + ], + "logprob_reward": -0.009811745956540108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009811745956540108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06934111565351486, + 0.04927298054099083, + 0.039729852229356766, + 0.03594205901026726, + 0.03583323955535889, + 0.03536340221762657, + 0.44492098689079285, + 0.05736524239182472, + 0.16636420786380768, + 0.06586690247058868 + ], + "credit_map": [ + 0.010654107667505741, + 0.00734707061201334, + 0.008319410495460033, + 0.007252503652125597, + 0.006864826660603285, + 0.007144800387322903, + 0.007608860731124878, + 0.014149529859423637, + 0.007634011562913656, + 0.0039084493182599545, + 0.00563718331977725, + 0.004471021704375744, + 0.004730481654405594, + 0.004568732809275389, + 0.004101418424397707, + 0.014221683144569397, + 0.00589728681370616, + 0.007014727685600519, + 0.006419437471777201, + 0.0035436658654361963, + 0.005067687481641769, + 0.0039707389660179615, + 0.003811886068433523, + 0.0040044235065579414, + 0.006276702508330345, + 0.008128929883241653, + 0.004673424642533064, + 0.0031502889469265938, + 0.002671689260751009, + 0.003836380084976554, + 0.0032680858857929707, + 0.003936556167900562, + 0.005416178610175848, + 0.00573087390512228, + 0.005602350924164057, + 0.003231954062357545, + 0.004421502351760864, + 0.0036629110109061003, + 0.003602422308176756, + 0.004165051504969597, + 0.003843875601887703, + 0.005708490964025259, + 0.004348402377218008, + 0.003952755592763424, + 0.003906626719981432, + 0.004815575201064348, + 0.004378496669232845, + 0.004409175366163254, + 0.01883210986852646, + 0.035908617079257965, + 0.09710599482059479, + 0.01662570610642433, + 0.19276566803455353, + 0.05544010177254677, + 0.010510713793337345, + 0.017732111737132072, + 0.01198889035731554, + 0.005725222174078226, + 0.00669505400583148, + 0.005489649251103401, + 0.005428857635706663, + 0.004852853715419769, + 0.0067969802767038345, + 0.010387735441327095, + 0.008427439257502556, + 0.0035133271012455225, + 0.06650137901306152, + 0.007603382226079702, + 0.0031185378320515156, + 0.00857460591942072, + 0.004670298658311367, + 0.06395523250102997, + 0.005897533148527145, + 0.005646516103297472, + 0.00411244947463274, + 0.00685855932533741, + 0.004472634755074978, + 0.005525059532374144, + 0.018347129225730896, + 0.015007019974291325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.009811745956540108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645367808640003 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645367808640003, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645367808640003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09499238431453705, + 0.055238813161849976, + 0.04958343878388405, + 0.103382907807827, + 0.048120684921741486, + 0.09193351864814758, + 0.24414867162704468, + 0.11804088205099106, + 0.1590505987405777, + 0.03550804406404495 + ], + "credit_map": [ + 0.01443851925432682, + 0.016524765640497208, + 0.012676391750574112, + 0.012262537144124508, + 0.011548057198524475, + 0.007547431625425816, + 0.012033614329993725, + 0.00796107854694128, + 0.009088799357414246, + 0.00762740196660161, + 0.006739712320268154, + 0.004920223727822304, + 0.004761004354804754, + 0.005687172524631023, + 0.004276372957974672, + 0.012138131074607372, + 0.006989296991378069, + 0.006891667377203703, + 0.005657986272126436, + 0.006309442222118378, + 0.006714590359479189, + 0.007371841464191675, + 0.005415607709437609, + 0.004233010578900576, + 0.028371376916766167, + 0.04095403477549553, + 0.004774264991283417, + 0.004894885700196028, + 0.008933939039707184, + 0.006297157611697912, + 0.004450462758541107, + 0.004706797190010548, + 0.006062801461666822, + 0.010497295297682285, + 0.0061270552687346935, + 0.004518060013651848, + 0.005181448068469763, + 0.0037683099508285522, + 0.004552406724542379, + 0.0074133118614554405, + 0.009571977891027927, + 0.011257275007665157, + 0.014156014658510685, + 0.012643133290112019, + 0.014284918084740639, + 0.010132218711078167, + 0.009428007528185844, + 0.010459979064762592, + 0.011534028686583042, + 0.021517829969525337, + 0.04606345668435097, + 0.01292842160910368, + 0.09997127950191498, + 0.030873334035277367, + 0.00856646616011858, + 0.01269389595836401, + 0.01533377543091774, + 0.009896144270896912, + 0.019611693918704987, + 0.013963213190436363, + 0.012335821986198425, + 0.010913207195699215, + 0.013655761256814003, + 0.022331269457936287, + 0.008181141689419746, + 0.003837005700916052, + 0.062423694878816605, + 0.0076697166077792645, + 0.003708010073751211, + 0.008026284165680408, + 0.004717670381069183, + 0.06048709899187088, + 0.0044600688852369785, + 0.0042161522433161736, + 0.003318964270874858, + 0.004830497317016125, + 0.002430833876132965, + 0.00456191273406148, + 0.00586857832968235, + 0.005821037106215954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.008645367808640003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817513071000576 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.01817513071000576, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817513071000576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1152368113398552, + 0.09151662886142731, + 0.06917044520378113, + 0.2229737937450409, + 0.031831491738557816, + 0.08954537659883499, + 0.14262962341308594, + 0.1204487681388855, + 0.059471383690834045, + 0.057175666093826294 + ], + "credit_map": [ + 0.018153555691242218, + 0.01071111485362053, + 0.01649576425552368, + 0.011756261810660362, + 0.016632482409477234, + 0.00779970595613122, + 0.011723391711711884, + 0.021964535117149353, + 0.013611972332000732, + 0.011167112737894058, + 0.01578344590961933, + 0.007271322887390852, + 0.011396845802664757, + 0.007718088570982218, + 0.008204968646168709, + 0.016362866386771202, + 0.009279179386794567, + 0.01196835283190012, + 0.006675104144960642, + 0.008158733136951923, + 0.007816093042492867, + 0.005776150617748499, + 0.009912557899951935, + 0.009584278799593449, + 0.06322977691888809, + 0.07597371935844421, + 0.01606593281030655, + 0.013628749176859856, + 0.014362696558237076, + 0.012157577089965343, + 0.012007439509034157, + 0.015547916293144226, + 0.004066700115799904, + 0.00620162533596158, + 0.004655675031244755, + 0.0027133547700941563, + 0.004103443119674921, + 0.0033096361439675093, + 0.0034512507263571024, + 0.003329806961119175, + 0.008991754613816738, + 0.010817589238286018, + 0.011260561645030975, + 0.014450978487730026, + 0.009549126960337162, + 0.007507001981139183, + 0.012776988558471203, + 0.014191370457410812, + 0.006827785167843103, + 0.011751705780625343, + 0.0329618863761425, + 0.006569758523255587, + 0.05660506337881088, + 0.015559135936200619, + 0.0057859416119754314, + 0.006568349897861481, + 0.01100198458880186, + 0.009486212395131588, + 0.02586904540657997, + 0.013228945434093475, + 0.011078116483986378, + 0.009389211423695087, + 0.012003542855381966, + 0.02839171327650547, + 0.0036815640050917864, + 0.0022082412615418434, + 0.021273361518979073, + 0.003765710396692157, + 0.003161095082759857, + 0.0032035920303314924, + 0.002363086212426424, + 0.019814735278487206, + 0.003157362574711442, + 0.0049771531485021114, + 0.0028422086033970118, + 0.006079995539039373, + 0.006883446592837572, + 0.01638663373887539, + 0.01136231329292059, + 0.00548655167222023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.01817513071000576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188600704073906 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188600704073906, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188600704073906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0881434828042984, + 0.055275652557611465, + 0.09134098142385483, + 0.0483013316988945, + 0.03534979000687599, + 0.06373526901006699, + 0.15923142433166504, + 0.07612735778093338, + 0.3458159565925598, + 0.03667878732085228 + ], + "credit_map": [ + 0.014416557736694813, + 0.008170965127646923, + 0.014637469314038754, + 0.007149658631533384, + 0.008165067993104458, + 0.007808439899235964, + 0.0072179511189460754, + 0.02057737112045288, + 0.007925068959593773, + 0.00667504221200943, + 0.008039877749979496, + 0.005163214169442654, + 0.005557643249630928, + 0.005192964803427458, + 0.0046629104763269424, + 0.012058932334184647, + 0.013172825798392296, + 0.013968056999146938, + 0.009594528004527092, + 0.010287902317941189, + 0.01234512496739626, + 0.011595360934734344, + 0.010906592942774296, + 0.009470588527619839, + 0.008228005841374397, + 0.009902571327984333, + 0.005903742276132107, + 0.0038392755668610334, + 0.005603541620075703, + 0.005948127247393131, + 0.003789866343140602, + 0.005086198914796114, + 0.004752928391098976, + 0.011523679830133915, + 0.003267758060246706, + 0.0024250915739685297, + 0.004537559114396572, + 0.0028953461442142725, + 0.0026315634604543447, + 0.0033158636651933193, + 0.00614364305511117, + 0.009811929427087307, + 0.006490583997219801, + 0.008408620953559875, + 0.008208098821341991, + 0.007964582182466984, + 0.008783779107034206, + 0.0079240333288908, + 0.007085578981786966, + 0.012999489903450012, + 0.036622337996959686, + 0.007973779924213886, + 0.062399860471487045, + 0.018148919567465782, + 0.0046287341974675655, + 0.009372727014124393, + 0.014580625109374523, + 0.010934794321656227, + 0.007893943227827549, + 0.007015347946435213, + 0.008201325312256813, + 0.00868610106408596, + 0.0070260800421237946, + 0.011789144016802311, + 0.019694142043590546, + 0.007332147564738989, + 0.13910989463329315, + 0.01723439246416092, + 0.007279064506292343, + 0.01777043752372265, + 0.009286263026297092, + 0.12810960412025452, + 0.0034674841444939375, + 0.005400117486715317, + 0.0030249443370848894, + 0.003122771391645074, + 0.003913114313036203, + 0.005943639203906059, + 0.0038154616486281157, + 0.007991251535713673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.014188600704073906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Recorded in a simple yet expert style, a low-angle, wide-angle shot pans smoothly to the right, tracking the powerful takeoff roll of a massive C-17 Globemaster III cargo plane. Its massive fuselage, painted in a non-reflective matte gray with visible panel lines, fills the frame as it roars along the long concrete runway marked with faded white lines. Four powerful turbofan engines, suspended high beneath the wings, roar with incredible force, visibly distorting the air behind them and emitting a shimmering heat haze under the harsh light of the clear midday sun. The camera maintains a perfect frame of the fuselage as the landing gear tires spin blurringly, building immense speed until the nose wheel decisively lifts off the tarmac. The entire aircraft rotates steadily upward, its massive form outlined against the clear blue sky, before the main landing gear leaves the ground and begins a steep, powerful ascent, creating an atmosphere of immense mechanical power and operational precision.", + "seed": 30094, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_3947_seed_30094_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0666441929526627, + "vlm_logprob": -0.0666441929526627, + "vlm_yes_no_margin": 3.974999952316284, + "vlm_token_credit": -0.0666441929526627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2911708354949951 + ], + "opposite_logprob_tokens": [ + -1.4161708354949951 + ], + "logprob_reward": -0.2911708354949951, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2911708354949951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047211337834596634, + 0.021377163007855415, + 0.0569911003112793, + 0.020763712003827095, + 0.5701695084571838, + 0.0086519168689847, + 0.027746887877583504, + 0.024249741807579994, + 0.02680956944823265, + 0.19602909684181213 + ], + "credit_map": [ + 0.0065584625117480755, + 0.005244365427643061, + 0.005716846324503422, + 0.004857753403484821, + 0.010274147614836693, + 0.0059461635537445545, + 0.0038387200329452753, + 0.004774877801537514, + 0.0026688671205192804, + 0.0031376220285892487, + 0.0020217914134263992, + 0.0021138405427336693, + 0.0031196007039397955, + 0.002655979711562395, + 0.0032824145164340734, + 0.002377047436311841, + 0.005238211248070002, + 0.007757098879665136, + 0.005255808122456074, + 0.004640596453100443, + 0.007949277758598328, + 0.010173536837100983, + 0.008150644600391388, + 0.007825925014913082, + 0.003365912940353155, + 0.0032609349582344294, + 0.0022199521772563457, + 0.0016051203710958362, + 0.004670646507292986, + 0.0026783915236592293, + 0.0017146270256489515, + 0.001248125801794231, + 0.0783466100692749, + 0.1615154892206192, + 0.025708919391036034, + 0.03580993413925171, + 0.17084577679634094, + 0.044708918780088425, + 0.031976647675037384, + 0.021257225424051285, + 0.0012132268166169524, + 0.0012850603088736534, + 0.0010569351725280285, + 0.0012088321382179856, + 0.0008851534803397954, + 0.0009053816320374608, + 0.0006421228172257543, + 0.001455204444937408, + 0.0030068191699683666, + 0.0028795315884053707, + 0.002355828881263733, + 0.003694827202707529, + 0.0016181421233341098, + 0.004132910165935755, + 0.00174171372782439, + 0.008317114785313606, + 0.0022879946045577526, + 0.003491155570372939, + 0.0034057607408612967, + 0.004978553857654333, + 0.0020586152095347643, + 0.002613759133964777, + 0.001975209219381213, + 0.0034386927727609873, + 0.0033840700052678585, + 0.0037932023406028748, + 0.004925202578306198, + 0.006202146876603365, + 0.0017486736178398132, + 0.0020058536902070045, + 0.002637802390381694, + 0.002112617949023843, + 0.01994730904698372, + 0.025250231847167015, + 0.054618146270513535, + 0.04281195253133774, + 0.01400931365787983, + 0.012812082655727863, + 0.014015712775290012, + 0.01256435178220272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.2911708354949951 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011122724041342735 + ], + "opposite_logprob_tokens": [ + -4.511122703552246 + ], + "logprob_reward": -0.011122724041342735, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011122724041342735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06835875660181046, + 0.032611995935440063, + 0.04637254402041435, + 0.04058312252163887, + 0.5837429165840149, + 0.04029582440853119, + 0.030156416818499565, + 0.052694376558065414, + 0.029964368790388107, + 0.07521963864564896 + ], + "credit_map": [ + 0.010332703590393066, + 0.00765950046479702, + 0.00786038488149643, + 0.0076331645250320435, + 0.014421594329178333, + 0.007956120185554028, + 0.006080484949052334, + 0.006414806470274925, + 0.004612230230122805, + 0.005107892211526632, + 0.0024089952930808067, + 0.003238213947042823, + 0.004657089244574308, + 0.0046904124319553375, + 0.004226373042911291, + 0.003670790931209922, + 0.004092883318662643, + 0.006033582147210836, + 0.0035742137115448713, + 0.003875168738886714, + 0.006105007603764534, + 0.008502860553562641, + 0.007300061639398336, + 0.006888768170028925, + 0.006083650514483452, + 0.005200705491006374, + 0.00480318209156394, + 0.004503748845309019, + 0.005871686153113842, + 0.007212574128061533, + 0.002917770529165864, + 0.003989804070442915, + 0.07885576039552689, + 0.16410323977470398, + 0.027538148686289787, + 0.03916025161743164, + 0.17089317739009857, + 0.04832508787512779, + 0.032657504081726074, + 0.02220982126891613, + 0.005998915992677212, + 0.009835446253418922, + 0.0038237483240664005, + 0.0041756778955459595, + 0.005803084932267666, + 0.0035479075741022825, + 0.002515390980988741, + 0.004595655482262373, + 0.004566482733935118, + 0.0033377991057932377, + 0.004445306491106749, + 0.004792343359440565, + 0.0036645731888711452, + 0.00335648562759161, + 0.002405193168669939, + 0.0035882366355508566, + 0.005112616810947657, + 0.007037706207484007, + 0.008071417920291424, + 0.011767475865781307, + 0.004268655087798834, + 0.0054203784093260765, + 0.0037988165859133005, + 0.0072173126973211765, + 0.003196677193045616, + 0.004074825439602137, + 0.005083873402327299, + 0.006328658666461706, + 0.002915245248004794, + 0.0035188777837902308, + 0.0018602663185447454, + 0.00298594799824059, + 0.007890080101788044, + 0.009175693616271019, + 0.019221842288970947, + 0.015716003254055977, + 0.007257071789354086, + 0.005458350758999586, + 0.005121506750583649, + 0.005379098001867533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.011122724041342735 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008654113858938217 + ], + "opposite_logprob_tokens": [ + -4.7586541175842285 + ], + "logprob_reward": -0.008654113858938217, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008654113858938217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05131782963871956, + 0.026999574154615402, + 0.03689270466566086, + 0.02426185831427574, + 0.5602071285247803, + 0.03526991233229637, + 0.017813406884670258, + 0.02896249294281006, + 0.023260025307536125, + 0.1950151026248932 + ], + "credit_map": [ + 0.006883717142045498, + 0.006579237524420023, + 0.00643892353400588, + 0.005544898100197315, + 0.010560642927885056, + 0.006733997259289026, + 0.0035282487515360117, + 0.005048157647252083, + 0.002960427664220333, + 0.003930502571165562, + 0.0020477480720728636, + 0.002954874886199832, + 0.004380752332508564, + 0.003737474326044321, + 0.003936301451176405, + 0.003051488194614649, + 0.003312129992991686, + 0.005024789832532406, + 0.0032633626833558083, + 0.0029408736154437065, + 0.004899561405181885, + 0.006756878923624754, + 0.005094380583614111, + 0.0056007253006100655, + 0.00355870695784688, + 0.0033981602173298597, + 0.0021905372850596905, + 0.0031873572152107954, + 0.002352508483454585, + 0.004169997293502092, + 0.0022816320415586233, + 0.0031229560263454914, + 0.07613303512334824, + 0.1564759761095047, + 0.02583962306380272, + 0.03635181114077568, + 0.1661432832479477, + 0.0459858737885952, + 0.031755223870277405, + 0.021522238850593567, + 0.004141739569604397, + 0.00572767062112689, + 0.003260459518060088, + 0.0034021951723843813, + 0.005194336641579866, + 0.0041713714599609375, + 0.0022744720336049795, + 0.007097661029547453, + 0.0028815651312470436, + 0.0020045768469572067, + 0.0017156220274046063, + 0.0014111693017184734, + 0.0020609283819794655, + 0.0025140189100056887, + 0.0013953327434137464, + 0.003830190282315016, + 0.004287842661142349, + 0.004897738341242075, + 0.0038270847871899605, + 0.0035239721182733774, + 0.001961473608389497, + 0.004289521370083094, + 0.0019502018112689257, + 0.004224653821438551, + 0.002883119275793433, + 0.003311338135972619, + 0.004459514282643795, + 0.004310018848627806, + 0.00163969153072685, + 0.0021983240731060505, + 0.002365442691370845, + 0.0020925726275891066, + 0.019862409681081772, + 0.024964269250631332, + 0.05397393926978111, + 0.042890697717666626, + 0.014010567218065262, + 0.0125714261084795, + 0.013803244568407536, + 0.012938525527715683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.008654113858938217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040932451374828815 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.0040932451374828815, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040932451374828815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05931694805622101, + 0.03648124635219574, + 0.03319891169667244, + 0.044313110411167145, + 0.39448651671409607, + 0.07384338974952698, + 0.0444226898252964, + 0.11619839817285538, + 0.1359405815601349, + 0.06179826706647873 + ], + "credit_map": [ + 0.007380265276879072, + 0.009645494632422924, + 0.003884090343490243, + 0.004699590150266886, + 0.01363638136535883, + 0.008955122902989388, + 0.004762351978570223, + 0.006353653036057949, + 0.00432888837531209, + 0.004494746681302786, + 0.002278959844261408, + 0.0033911315258592367, + 0.008537269197404385, + 0.00461928965523839, + 0.00491096219047904, + 0.0039199963212013245, + 0.0029275661800056696, + 0.004211208783090115, + 0.004794977139681578, + 0.0028620692901313305, + 0.004837386775761843, + 0.005555798765271902, + 0.004409180488437414, + 0.003600719850510359, + 0.0070710559375584126, + 0.006307541858404875, + 0.005610100459307432, + 0.0040345205925405025, + 0.004613477736711502, + 0.008667112328112125, + 0.003581835888326168, + 0.004427459090948105, + 0.05469290912151337, + 0.1105712279677391, + 0.017766103148460388, + 0.024758197367191315, + 0.11796220391988754, + 0.03206833824515343, + 0.022198732942342758, + 0.014468773268163204, + 0.008742006495594978, + 0.009925751015543938, + 0.0059562851674854755, + 0.007838110439479351, + 0.006845812778919935, + 0.011037987656891346, + 0.005449377465993166, + 0.018048053607344627, + 0.007026889361441135, + 0.008601379580795765, + 0.004473275970667601, + 0.006062170024961233, + 0.006766339763998985, + 0.004406300373375416, + 0.0037782960571348667, + 0.003308031940832734, + 0.012004222720861435, + 0.016828272491693497, + 0.015458629466593266, + 0.02288197912275791, + 0.009096818044781685, + 0.013284353539347649, + 0.009574939496815205, + 0.017069179564714432, + 0.01459091529250145, + 0.01797688752412796, + 0.025248045101761818, + 0.03394196555018425, + 0.008741979487240314, + 0.010068207047879696, + 0.012610088102519512, + 0.01276247389614582, + 0.00644754059612751, + 0.008224002085626125, + 0.014144694432616234, + 0.011483845300972462, + 0.006065571215003729, + 0.004061441868543625, + 0.00500305974856019, + 0.0063681043684482574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.0040932451374828815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818004623055458 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.01818004623055458, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818004623055458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05067756399512291, + 0.03482479229569435, + 0.020199602469801903, + 0.022384990006685257, + 0.42343565821647644, + 0.02983073517680168, + 0.13084076344966888, + 0.09110738337039948, + 0.11000470072031021, + 0.08669380843639374 + ], + "credit_map": [ + 0.005545018706470728, + 0.007546582259237766, + 0.005543740000575781, + 0.004076302517205477, + 0.011181549169123173, + 0.007492425385862589, + 0.004675684031099081, + 0.004616261459887028, + 0.003916142042726278, + 0.005781108513474464, + 0.004220650997012854, + 0.002428308129310608, + 0.004937299527227879, + 0.004432045388966799, + 0.0041023618541657925, + 0.0050068809650838375, + 0.0022121586371213198, + 0.002849027281627059, + 0.0018825300503522158, + 0.0019544176757335663, + 0.002618800848722458, + 0.0031793296802788973, + 0.0028780861757695675, + 0.0026252528186887503, + 0.0038268864154815674, + 0.004261407069861889, + 0.0018539221491664648, + 0.0021621042396873236, + 0.003410999896004796, + 0.003232513554394245, + 0.0018295188201591372, + 0.0018076386768370867, + 0.05753779411315918, + 0.11591774225234985, + 0.01907327212393284, + 0.028954945504665375, + 0.12449832260608673, + 0.03716753423213959, + 0.023782087489962578, + 0.016503935679793358, + 0.00400262838229537, + 0.005340866278856993, + 0.0036100095603615046, + 0.003657652297988534, + 0.004063891246914864, + 0.0040003773756325245, + 0.00252814800478518, + 0.0026271608658134937, + 0.015237981453537941, + 0.011579783633351326, + 0.010551631450653076, + 0.016200216487050056, + 0.00823314767330885, + 0.019977513700723648, + 0.008005719631910324, + 0.041054777801036835, + 0.012115894816815853, + 0.014466846361756325, + 0.016115838661789894, + 0.018213655799627304, + 0.007230773102492094, + 0.006896995473653078, + 0.006905423942953348, + 0.009161952883005142, + 0.012333721853792667, + 0.015050828456878662, + 0.020971717312932014, + 0.02475118637084961, + 0.008507892489433289, + 0.009400544688105583, + 0.00985176395624876, + 0.009137037210166454, + 0.009144940413534641, + 0.011155489832162857, + 0.02229589968919754, + 0.017616793513298035, + 0.00756325526162982, + 0.005746550392359495, + 0.006899900268763304, + 0.006270986050367355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.01818004623055458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Recorded in a simple yet expert style, a low-angle, wide-angle shot pans smoothly to the right, tracking the powerful takeoff roll of a massive C-17 Globemaster III cargo plane. Its massive fuselage, painted in a non-reflective matte gray with visible panel lines, fills the frame as it roars along the long concrete runway marked with faded white lines. Four powerful turbofan engines, suspended high beneath the wings, roar with incredible force, visibly distorting the air behind them and emitting a shimmering heat haze under the harsh light of the clear midday sun. The camera maintains a perfect frame of the fuselage as the landing gear tires spin blurringly, building immense speed until the nose wheel decisively lifts off the tarmac. The entire aircraft rotates steadily upward, its massive form outlined against the clear blue sky, before the main landing gear leaves the ground and begins a steep, powerful ascent, creating an atmosphere of immense mechanical power and operational precision.", + "seed": 30095, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_10/index_3947_seed_30095_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030198576115071773, + "vlm_logprob": -0.030198576115071773, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.030198576115071773, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09982733428478241 + ], + "opposite_logprob_tokens": [ + -2.474827289581299 + ], + "logprob_reward": -0.09982733428478241, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09982733428478241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07987818866968155, + 0.100540392100811, + 0.07454531639814377, + 0.1693599820137024, + 0.06345824152231216, + 0.029051806777715683, + 0.10845349729061127, + 0.035143762826919556, + 0.16149994730949402, + 0.17806889116764069 + ], + "credit_map": [ + 0.008660303428769112, + 0.008844036608934402, + 0.013005238026380539, + 0.009442592039704323, + 0.009380132891237736, + 0.010103898122906685, + 0.011060832999646664, + 0.00938115082681179, + 0.03928859159350395, + 0.039802733808755875, + 0.002873173914849758, + 0.002778455149382353, + 0.005780172534286976, + 0.003324041375890374, + 0.003325831377878785, + 0.0033673918806016445, + 0.012297717854380608, + 0.012693212367594242, + 0.01109628938138485, + 0.006894160062074661, + 0.009281545877456665, + 0.008150358684360981, + 0.008039298467338085, + 0.006092729978263378, + 0.02553773671388626, + 0.02227357029914856, + 0.026249492540955544, + 0.01891106739640236, + 0.019152076914906502, + 0.021819399669766426, + 0.01981012150645256, + 0.01560650672763586, + 0.00897302944213152, + 0.026143746450543404, + 0.0028402910102158785, + 0.00208662124350667, + 0.0167043786495924, + 0.0024174880236387253, + 0.002182450145483017, + 0.0021102363243699074, + 0.004664537496864796, + 0.0029997897800058126, + 0.0037327727768570185, + 0.003567029954865575, + 0.0050673773512244225, + 0.004395245108753443, + 0.0025580094661563635, + 0.002067042514681816, + 0.01700424775481224, + 0.018513891845941544, + 0.015169563703238964, + 0.016507692635059357, + 0.011678669601678848, + 0.006378759164363146, + 0.016482479870319366, + 0.006718181539326906, + 0.0053656380623579025, + 0.004533823579549789, + 0.0063332500867545605, + 0.006371335592120886, + 0.004271214362233877, + 0.0024927922058850527, + 0.0032395068556070328, + 0.0025362027809023857, + 0.020774876698851585, + 0.019620627164840698, + 0.022740857675671577, + 0.028175540268421173, + 0.01946672797203064, + 0.015099710784852505, + 0.02072042040526867, + 0.014901180751621723, + 0.0535045824944973, + 0.010391488671302795, + 0.06620276719331741, + 0.020502349361777306, + 0.006515793967992067, + 0.007045422215014696, + 0.0062489863485097885, + 0.007657492533326149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.09982733428478241 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009824494831264019 + ], + "opposite_logprob_tokens": [ + -4.634824275970459 + ], + "logprob_reward": -0.009824494831264019, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009824494831264019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2758491635322571, + 0.20455119013786316, + 0.03763921931385994, + 0.05241120606660843, + 0.14132185280323029, + 0.037461645901203156, + 0.07784835249185562, + 0.024468552321195602, + 0.031048893928527832, + 0.11739996820688248 + ], + "credit_map": [ + 0.03172855079174042, + 0.030972329899668694, + 0.05464326590299606, + 0.03206998482346535, + 0.0416080541908741, + 0.027318596839904785, + 0.02392399311065674, + 0.0335843600332737, + 0.0844866931438446, + 0.08533918857574463, + 0.007035144604742527, + 0.00590148102492094, + 0.009172504767775536, + 0.004186239559203386, + 0.004900519736111164, + 0.003529410110786557, + 0.005689922720193863, + 0.008117935620248318, + 0.004177699331194162, + 0.004737509414553642, + 0.004833076614886522, + 0.0026307208463549614, + 0.003793468466028571, + 0.0036588802468031645, + 0.006874015089124441, + 0.009945153258740902, + 0.006454086862504482, + 0.00419273367151618, + 0.00865904800593853, + 0.00585394399240613, + 0.004899138119071722, + 0.005533083342015743, + 0.021226370707154274, + 0.05706101655960083, + 0.006047593895345926, + 0.004310532007366419, + 0.03616581857204437, + 0.00704294303432107, + 0.0055419509299099445, + 0.003925608936697245, + 0.008226009085774422, + 0.004387538880109787, + 0.003682181006297469, + 0.003510415321215987, + 0.0079782260581851, + 0.003972234204411507, + 0.003040866693481803, + 0.002664172323420644, + 0.020040592178702354, + 0.022670120000839233, + 0.0111387949436903, + 0.004527018405497074, + 0.009394587017595768, + 0.0035544373095035553, + 0.0034371826332062483, + 0.003085614647716284, + 0.003839637618511915, + 0.003581300377845764, + 0.004138254094868898, + 0.005010894499719143, + 0.002626958070322871, + 0.0015338280936703086, + 0.0022602027747780085, + 0.001477475161664188, + 0.0038535320200026035, + 0.0030776788480579853, + 0.004301393870264292, + 0.006370606832206249, + 0.0031943300273269415, + 0.0035467161796987057, + 0.003111002966761589, + 0.003593627829104662, + 0.025649666786193848, + 0.01036771945655346, + 0.031020404770970345, + 0.014455371536314487, + 0.013927807100117207, + 0.007366266567260027, + 0.007740896660834551, + 0.006871820893138647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009824494831264019 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089242994785309 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.011089242994785309, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089242994785309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1340249478816986, + 0.1463688313961029, + 0.08146318048238754, + 0.12440364062786102, + 0.06134308502078056, + 0.035682231187820435, + 0.06697399169206619, + 0.026906194165349007, + 0.07482745498418808, + 0.2480064481496811 + ], + "credit_map": [ + 0.015143610537052155, + 0.015667319297790527, + 0.026139339432120323, + 0.01616431586444378, + 0.016843179240822792, + 0.01493289414793253, + 0.014606612734496593, + 0.014527671039104462, + 0.055152662098407745, + 0.05171060189604759, + 0.009777269326150417, + 0.005413087550550699, + 0.007934896275401115, + 0.005563641898334026, + 0.0061642020009458065, + 0.004652469418942928, + 0.010234840214252472, + 0.014649025164544582, + 0.009441905654966831, + 0.008633177727460861, + 0.010989799164235592, + 0.007134632673114538, + 0.011815190315246582, + 0.008564612828195095, + 0.017668701708316803, + 0.018592247739434242, + 0.01495389360934496, + 0.013464128598570824, + 0.016232116147875786, + 0.01625334471464157, + 0.014150969684123993, + 0.013088242150843143, + 0.009477327577769756, + 0.0185588076710701, + 0.004650026559829712, + 0.004050669725984335, + 0.011487201787531376, + 0.005457015708088875, + 0.003681495785713196, + 0.003980538807809353, + 0.006027895025908947, + 0.0038512861356139183, + 0.004320354666560888, + 0.0036491351202130318, + 0.0065817213617265224, + 0.0051127150654792786, + 0.003350134240463376, + 0.002788989804685116, + 0.015409384854137897, + 0.01613869145512581, + 0.00858392845839262, + 0.0064064376056194305, + 0.005774344317615032, + 0.0046195583418011665, + 0.00522296829149127, + 0.004818678367882967, + 0.0037103095091879368, + 0.0042395275086164474, + 0.005194506142288446, + 0.003664947347715497, + 0.002832457423210144, + 0.0025460971519351006, + 0.002654758980497718, + 0.002063588472083211, + 0.010139424353837967, + 0.008338027633726597, + 0.013651276007294655, + 0.011202136985957623, + 0.008367964997887611, + 0.00909309834241867, + 0.007214197888970375, + 0.006821331102401018, + 0.06678150594234467, + 0.016980096697807312, + 0.08218114823102951, + 0.028826134279370308, + 0.01984788104891777, + 0.011285213753581047, + 0.012447199784219265, + 0.009657277725636959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.011089242994785309 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062267124652863 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.016062267124652863, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062267124652863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10860222578048706, + 0.06575957685709, + 0.09656160324811935, + 0.03410285711288452, + 0.034248560667037964, + 0.10949723422527313, + 0.0665145292878151, + 0.14608833193778992, + 0.11759369820356369, + 0.2210313081741333 + ], + "credit_map": [ + 0.013763649389147758, + 0.013905465602874756, + 0.018878525123000145, + 0.01389528252184391, + 0.01565099135041237, + 0.013244652189314365, + 0.007960900664329529, + 0.011302754282951355, + 0.012423518113791943, + 0.011867864057421684, + 0.010776021517813206, + 0.005434058140963316, + 0.007530674338340759, + 0.005831603426486254, + 0.005804918240755796, + 0.006090917624533176, + 0.01267287228256464, + 0.015534854494035244, + 0.009864358231425285, + 0.012011758051812649, + 0.01253860630095005, + 0.007945715449750423, + 0.01691405102610588, + 0.009079386480152607, + 0.004965032450854778, + 0.004994950722903013, + 0.005123544950038195, + 0.005543840583413839, + 0.003983046393841505, + 0.0035198540426790714, + 0.003164195455610752, + 0.002808390883728862, + 0.0036388735752552748, + 0.008622074499726295, + 0.003977857995778322, + 0.002706687431782484, + 0.007204004563391209, + 0.0025993261951953173, + 0.0027638375759124756, + 0.0027358969673514366, + 0.013665185309946537, + 0.012036570347845554, + 0.010063625872135162, + 0.02058098278939724, + 0.01442723162472248, + 0.02039119228720665, + 0.010945603251457214, + 0.007386850658804178, + 0.009369823150336742, + 0.007862607017159462, + 0.010301969014108181, + 0.011769409291446209, + 0.012972372584044933, + 0.004801486153155565, + 0.005238350015133619, + 0.0041985162533819675, + 0.02102493681013584, + 0.02037065289914608, + 0.03168949484825134, + 0.02025933936238289, + 0.012890208512544632, + 0.014649765565991402, + 0.013741310685873032, + 0.011462616734206676, + 0.014523996040225029, + 0.01339700073003769, + 0.016389669850468636, + 0.01979248598217964, + 0.013041884638369083, + 0.014428550377488136, + 0.013360721059143543, + 0.012659391388297081, + 0.06314754486083984, + 0.011599446646869183, + 0.07659009844064713, + 0.020622681826353073, + 0.01709357090294361, + 0.010596317239105701, + 0.01135993655771017, + 0.010021721012890339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.016062267124652863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive C-17 Globemaster III cargo plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage painted in a non-reflective matte gray with visible panel lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four turbofan engines visible suspended beneath the wings, emitting a shimmering heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the plane accelerating on a concrete runway with faded white lines before lifting off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, wide-angle view that pans smoothly to the right while tracking the plane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189541339874268 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189541339874268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189541339874268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.178606778383255, + 0.14387457072734833, + 0.056967731565237045, + 0.03945056349039078, + 0.06460539996623993, + 0.06755879521369934, + 0.12265966087579727, + 0.03406592085957527, + 0.05276938155293465, + 0.2394411563873291 + ], + "credit_map": [ + 0.021140268072485924, + 0.02069001831114292, + 0.034461669623851776, + 0.02452229894697666, + 0.017229262739419937, + 0.02004433050751686, + 0.01681334339082241, + 0.023705599829554558, + 0.049385856837034225, + 0.05128030478954315, + 0.009582131169736385, + 0.005746311508119106, + 0.010329453274607658, + 0.00577117083594203, + 0.006503978278487921, + 0.005275358445942402, + 0.007979689165949821, + 0.010666092857718468, + 0.006003691349178553, + 0.008224350400269032, + 0.007938746362924576, + 0.005372024141252041, + 0.006632125470787287, + 0.004151011351495981, + 0.00443361047655344, + 0.0068650334142148495, + 0.0043295263312757015, + 0.004549085628241301, + 0.005033140070736408, + 0.005551987327635288, + 0.0046396516263484955, + 0.004048529546707869, + 0.012346239760518074, + 0.025010915473103523, + 0.0038592994678765535, + 0.002916655270382762, + 0.01315105427056551, + 0.002687253523617983, + 0.0024454323574900627, + 0.0021885475143790245, + 0.010027982294559479, + 0.007616469636559486, + 0.009310196153819561, + 0.008966749534010887, + 0.010541314259171486, + 0.009474601596593857, + 0.006693929899483919, + 0.004927547182887793, + 0.032584019005298615, + 0.039449043571949005, + 0.01022565457969904, + 0.010179613716900349, + 0.013386709615588188, + 0.004694671835750341, + 0.008317345753312111, + 0.0038226160686463118, + 0.005027644336223602, + 0.004837889689952135, + 0.0058520096354186535, + 0.004188977647572756, + 0.004097385331988335, + 0.003915979992598295, + 0.0032782405614852905, + 0.0028677911031991243, + 0.0066611869260668755, + 0.005384096875786781, + 0.007154614198952913, + 0.008316971361637115, + 0.005100946873426437, + 0.006791049148887396, + 0.006067274138331413, + 0.007293236907571554, + 0.0771268978714943, + 0.013982836157083511, + 0.08345983177423477, + 0.021316614001989365, + 0.014265133999288082, + 0.010768869891762733, + 0.008193029090762138, + 0.010327951982617378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.014189541339874268 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_11/index_tensor([9235, 9235])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_11/index_tensor([9235, 9235])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3351368cd8ad542b8a43a9e6c1a4ecdc83a43b1f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_11/index_tensor([9235, 9235])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot depicts two figures, dwarfed by the massive presence of a power plant's cooling towers. The scene is bathed in the soft, warm light of dusk, casting long, dramatic shadows across the gravelly ground. On the left stands an older man, his weathered face etched with the lines of experience. He wears a faded, dark blue work jumpsuit, a yellowed hard hat tucked under his arm. His posture is slightly relaxed, one hand gesturing deftly, his expression thoughtful. Beside him, a younger figure in a modern, dark green technical jacket and gray overalls listens intently, hands in pockets, head tilted slightly. They are engaged in a quiet conversation, their distance suggesting a shared, private moment. Massive concrete towers rise behind them, their richly textured surfaces capturing the golden hour, while white steam drifts lazily through vast openings, emitting pink and orange glows against the deepening twilight sky. The predominant color palette is a warm fusion of amber and rose, contrasted against the cool, industrial gray of the concrete, creating an atmosphere of deep contemplation and tranquility, highlighting moments of intimate human connection against a backdrop of immense power.", + "seed": 5695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_11/index_9235_seed_5695_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012775278743356466, + "vlm_logprob": -0.012775278743356466, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.012775278743356466, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02068273350596428 + ], + "opposite_logprob_tokens": [ + -6.0206828117370605 + ], + "logprob_reward": -0.02068273350596428, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.02068273350596428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3035834729671478, + 0.14173872768878937, + 0.06924014538526535, + 0.06360562890768051, + 0.04936590418219566, + 0.0804177075624466, + 0.11274628341197968, + 0.0551111064851284, + 0.04904179275035858, + 0.07514920085668564 + ], + "credit_map": [ + 0.05984712392091751, + 0.04937068372964859, + 0.056608300656080246, + 0.02815849334001541, + 0.031140713021159172, + 0.03168473392724991, + 0.017330607399344444, + 0.02944282442331314, + 0.027407728135585785, + 0.027855895459651947, + 0.021450143307447433, + 0.01610688865184784, + 0.01626557856798172, + 0.011821861378848553, + 0.008722923696041107, + 0.01210771594196558, + 0.013962152414023876, + 0.008676612749695778, + 0.01130338478833437, + 0.006943018641322851, + 0.010311865247786045, + 0.006457336712628603, + 0.005282666068524122, + 0.00630310270935297, + 0.011078665032982826, + 0.008019302040338516, + 0.009988605044782162, + 0.006588801741600037, + 0.012520328164100647, + 0.005948165897279978, + 0.0044074105098843575, + 0.005054353736341, + 0.010644598864018917, + 0.005572827998548746, + 0.00891700480133295, + 0.00580441951751709, + 0.0056775836274027824, + 0.005581921432167292, + 0.00339897139929235, + 0.0037685763090848923, + 0.018212731927633286, + 0.010006226599216461, + 0.013001842424273491, + 0.009754238650202751, + 0.006823794450610876, + 0.008433734998106956, + 0.005160776432603598, + 0.009024363942444324, + 0.02634827420115471, + 0.013464738614857197, + 0.01771015301346779, + 0.009239391423761845, + 0.009258254431188107, + 0.012145670130848885, + 0.007837343961000443, + 0.01674245484173298, + 0.011510219424962997, + 0.006693792529404163, + 0.008417442440986633, + 0.008017272688448429, + 0.004866195842623711, + 0.006470109801739454, + 0.003000079421326518, + 0.00613598944619298, + 0.008344043977558613, + 0.005106188356876373, + 0.006535024382174015, + 0.0056209382601082325, + 0.0054419320076704025, + 0.005018108990043402, + 0.0043166340328752995, + 0.008658922277390957, + 0.016645167022943497, + 0.012978928163647652, + 0.01124523114413023, + 0.008198929950594902, + 0.006787087768316269, + 0.007723284885287285, + 0.006740428972989321, + 0.004830147139728069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02068273350596428 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255748514086008 + ], + "opposite_logprob_tokens": [ + -4.387557506561279 + ], + "logprob_reward": -0.01255748514086008, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255748514086008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3178752362728119, + 0.18935857713222504, + 0.08755698055028915, + 0.06080576404929161, + 0.055089469999074936, + 0.048088591545820236, + 0.16605864465236664, + 0.018902812153100967, + 0.030315978452563286, + 0.025947947055101395 + ], + "credit_map": [ + 0.07356388866901398, + 0.03936770558357239, + 0.05726801976561546, + 0.03712982311844826, + 0.03336850926280022, + 0.027391545474529266, + 0.018805813044309616, + 0.030979935079813004, + 0.030341563746333122, + 0.02843434363603592, + 0.04219171032309532, + 0.027067959308624268, + 0.0151065057143569, + 0.01608436554670334, + 0.015113361179828644, + 0.015018770471215248, + 0.01660301722586155, + 0.011445554904639721, + 0.013959691859781742, + 0.010561635717749596, + 0.01170791033655405, + 0.007999733090400696, + 0.007919111289083958, + 0.00736032472923398, + 0.011617506854236126, + 0.0057400609366595745, + 0.006915333215147257, + 0.007872818037867546, + 0.011640847660601139, + 0.005982810631394386, + 0.004089246038347483, + 0.006947138346731663, + 0.010029755532741547, + 0.006483111996203661, + 0.007521717809140682, + 0.007514060474932194, + 0.006290940102189779, + 0.00549929728731513, + 0.004346202127635479, + 0.007404382806271315, + 0.006860032677650452, + 0.003676541382446885, + 0.01582709513604641, + 0.00464608334004879, + 0.00467645563185215, + 0.0057060676626861095, + 0.003305058693513274, + 0.0033912567887455225, + 0.04654689505696297, + 0.02475312538444996, + 0.03201313316822052, + 0.013744646683335304, + 0.008596040308475494, + 0.01929353177547455, + 0.00829600915312767, + 0.012815253809094429, + 0.003162223845720291, + 0.0017449123552069068, + 0.002391926711425185, + 0.0029629161581397057, + 0.0019695772789418697, + 0.0020245325285941362, + 0.001888340339064598, + 0.0027583828195929527, + 0.0069152289070189, + 0.003504929831251502, + 0.00496251042932272, + 0.0022119153290987015, + 0.0028198326472193003, + 0.0034629632718861103, + 0.0023668576031923294, + 0.0040717399679124355, + 0.005815149750560522, + 0.003139426466077566, + 0.0037726666778326035, + 0.0025238743983209133, + 0.0025716712698340416, + 0.0031509511172771454, + 0.001597795751877129, + 0.00337641267105937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01255748514086008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638040162622929 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008638040162622929, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638040162622929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31571903824806213, + 0.16397234797477722, + 0.0679595097899437, + 0.07542259246110916, + 0.07672516256570816, + 0.05546725541353226, + 0.11299631744623184, + 0.05814109370112419, + 0.042414817959070206, + 0.031181834638118744 + ], + "credit_map": [ + 0.04977548494935036, + 0.03703580051660538, + 0.03635481745004654, + 0.05108196288347244, + 0.059329744428396225, + 0.02340192347764969, + 0.03445839136838913, + 0.02428095042705536, + 0.027919858694076538, + 0.021610673516988754, + 0.035559672862291336, + 0.02434973046183586, + 0.013608118519186974, + 0.013845143839716911, + 0.011799057945609093, + 0.015280080027878284, + 0.013078708201646805, + 0.008082517422735691, + 0.010401396080851555, + 0.01205951813608408, + 0.007580510340631008, + 0.005801061168313026, + 0.005148707423359156, + 0.005807095672935247, + 0.014749246649444103, + 0.007582066580653191, + 0.00993078202009201, + 0.016252629458904266, + 0.010061548091471195, + 0.007347593549638987, + 0.005017830990254879, + 0.004480893258005381, + 0.009987864643335342, + 0.007882396690547466, + 0.0079875523224473, + 0.021727122366428375, + 0.009079709649085999, + 0.00511799193918705, + 0.00976053811609745, + 0.005181985441595316, + 0.008568689227104187, + 0.0055614979937672615, + 0.011766939423978329, + 0.009936089627444744, + 0.006152013782411814, + 0.004774151369929314, + 0.004620025400072336, + 0.004087850451469421, + 0.03203759342432022, + 0.017067240551114082, + 0.020503735169768333, + 0.009547952562570572, + 0.006050816271454096, + 0.013199878856539726, + 0.005800582002848387, + 0.0087885158136487, + 0.011116180568933487, + 0.007461715489625931, + 0.006636183243244886, + 0.013098515570163727, + 0.005008291453123093, + 0.004953729920089245, + 0.0062803709879517555, + 0.003586104605346918, + 0.007000254467129707, + 0.00495601212605834, + 0.005042232573032379, + 0.005939246620982885, + 0.007027851417660713, + 0.00387665000744164, + 0.005001700017601252, + 0.003570872824639082, + 0.006054328754544258, + 0.0036181474570184946, + 0.004430877976119518, + 0.0030151954852044582, + 0.00406775763258338, + 0.0036979992873966694, + 0.0023993959184736013, + 0.003898132126778364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.008638040162622929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016053704544901848 + ], + "opposite_logprob_tokens": [ + -4.141053676605225 + ], + "logprob_reward": -0.016053704544901848, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016053704544901848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2155883014202118, + 0.1548803746700287, + 0.040278419852256775, + 0.05735571309924126, + 0.06514899432659149, + 0.1447768211364746, + 0.20438039302825928, + 0.06734637916088104, + 0.029641011729836464, + 0.02060358040034771 + ], + "credit_map": [ + 0.035644855350255966, + 0.03247174993157387, + 0.04579803720116615, + 0.02148800529539585, + 0.0183329526335001, + 0.020352553576231003, + 0.015275050885975361, + 0.026225097477436066, + 0.03744751214981079, + 0.019668221473693848, + 0.03123370185494423, + 0.020488617941737175, + 0.00786744523793459, + 0.01677621342241764, + 0.0081697478890419, + 0.013228902593255043, + 0.005631611216813326, + 0.006154376547783613, + 0.008674446493387222, + 0.004097622353583574, + 0.0046756090596318245, + 0.003893887158483267, + 0.0033039553090929985, + 0.0038469117134809494, + 0.0075204093009233475, + 0.005346536636352539, + 0.013244127854704857, + 0.005608726758509874, + 0.007690164260566235, + 0.006144897546619177, + 0.005106234923005104, + 0.006694618612527847, + 0.015942959114909172, + 0.0090327812358737, + 0.011690659448504448, + 0.005878804251551628, + 0.005556395743042231, + 0.008383837528526783, + 0.0039046958554536104, + 0.004758861381560564, + 0.032167382538318634, + 0.018429892137646675, + 0.03939930722117424, + 0.012193290516734123, + 0.009794376790523529, + 0.016757093369960785, + 0.007430619094520807, + 0.008604856207966805, + 0.05784766748547554, + 0.03027527406811714, + 0.037801265716552734, + 0.01716199330985546, + 0.010638377629220486, + 0.024104973301291466, + 0.010230309329926968, + 0.01632053032517433, + 0.01787654310464859, + 0.01038572657853365, + 0.010106646455824375, + 0.0077490173280239105, + 0.00491938553750515, + 0.008083168417215347, + 0.003486786037683487, + 0.004739100579172373, + 0.0048843976110219955, + 0.0028974812012165785, + 0.006574604194611311, + 0.0029898916836827993, + 0.0033354659099131823, + 0.0032472664024680853, + 0.0027448327746242285, + 0.0029670705553144217, + 0.005500162485986948, + 0.003003227524459362, + 0.0028577058110386133, + 0.0022296151146292686, + 0.0016379745211452246, + 0.0026316142175346613, + 0.001132108038291335, + 0.0016111729200929403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.016053704544901848 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944430362433195 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944430362433195, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944430362433195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40582630038261414, + 0.1065770611166954, + 0.06540368497371674, + 0.04779001697897911, + 0.041203536093235016, + 0.05150894448161125, + 0.09090516716241837, + 0.09389207512140274, + 0.061886150389909744, + 0.03500712290406227 + ], + "credit_map": [ + 0.07032165676355362, + 0.07129494845867157, + 0.08320824801921844, + 0.0445253886282444, + 0.03628985583782196, + 0.032756030559539795, + 0.02543059177696705, + 0.04199958220124245, + 0.023247774690389633, + 0.0143752321600914, + 0.020367542281746864, + 0.013066214509308338, + 0.009865140542387962, + 0.009941523894667625, + 0.006287968717515469, + 0.00942566990852356, + 0.006963025312870741, + 0.007405959069728851, + 0.006894104648381472, + 0.014916826970875263, + 0.011620239354670048, + 0.005031676962971687, + 0.006602982059121132, + 0.005968867335468531, + 0.008427029475569725, + 0.0049308892339468, + 0.00820902083069086, + 0.00497135566547513, + 0.009309718385338783, + 0.00503069581463933, + 0.00313780689612031, + 0.00377350184135139, + 0.007050882559269667, + 0.004873193334788084, + 0.006637444719672203, + 0.004567903466522694, + 0.0044292570091784, + 0.0045421430841088295, + 0.003067033365368843, + 0.006035675760358572, + 0.008787555620074272, + 0.007006142754107714, + 0.01270409394055605, + 0.004813300911337137, + 0.004528159741312265, + 0.005487608257681131, + 0.0033834497444331646, + 0.004798637703061104, + 0.024841835722327232, + 0.012820670381188393, + 0.015809647738933563, + 0.007777572609484196, + 0.005080403294414282, + 0.01076497882604599, + 0.0048871878534555435, + 0.008922872133553028, + 0.02531050704419613, + 0.014590890146791935, + 0.015418239869177341, + 0.00974054541438818, + 0.006789287086576223, + 0.011229217983782291, + 0.00487422663718462, + 0.005939166992902756, + 0.010510972701013088, + 0.006794297136366367, + 0.013199521228671074, + 0.0054398332722485065, + 0.006774207577109337, + 0.006697340402752161, + 0.005319603718817234, + 0.007150379009544849, + 0.007235371042042971, + 0.003922491800040007, + 0.005373423919081688, + 0.0031170344445854425, + 0.0038461259100586176, + 0.0038791175466030836, + 0.002341833896934986, + 0.005291725043207407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.005944430362433195 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot depicts two figures, dwarfed by the massive presence of a power plant's cooling towers. The scene is bathed in the soft, warm light of dusk, casting long, dramatic shadows across the gravelly ground. On the left stands an older man, his weathered face etched with the lines of experience. He wears a faded, dark blue work jumpsuit, a yellowed hard hat tucked under his arm. His posture is slightly relaxed, one hand gesturing deftly, his expression thoughtful. Beside him, a younger figure in a modern, dark green technical jacket and gray overalls listens intently, hands in pockets, head tilted slightly. They are engaged in a quiet conversation, their distance suggesting a shared, private moment. Massive concrete towers rise behind them, their richly textured surfaces capturing the golden hour, while white steam drifts lazily through vast openings, emitting pink and orange glows against the deepening twilight sky. The predominant color palette is a warm fusion of amber and rose, contrasted against the cool, industrial gray of the concrete, creating an atmosphere of deep contemplation and tranquility, highlighting moments of intimate human connection against a backdrop of immense power.", + "seed": 5696, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_11/index_9235_seed_5696_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011743209697306157, + "vlm_logprob": -0.011743209697306157, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.011743209697306157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010840808041393757 + ], + "opposite_logprob_tokens": [ + -5.510840892791748 + ], + "logprob_reward": -0.010840808041393757, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010840808041393757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25601091980934143, + 0.1582041084766388, + 0.13856951892375946, + 0.06496091932058334, + 0.06644739955663681, + 0.08517471700906754, + 0.035199109464883804, + 0.07024092227220535, + 0.0632355809211731, + 0.06195680797100067 + ], + "credit_map": [ + 0.033938389271497726, + 0.039248693734407425, + 0.02987181581556797, + 0.03881483152508736, + 0.026474375277757645, + 0.047768108546733856, + 0.017071621492505074, + 0.022823091596364975, + 0.020669152960181236, + 0.024579310789704323, + 0.019057396799325943, + 0.02573150210082531, + 0.020905647426843643, + 0.023765157908201218, + 0.010516698472201824, + 0.0129792345687747, + 0.011321130208671093, + 0.017497854307293892, + 0.013647232204675674, + 0.009991418570280075, + 0.03276194632053375, + 0.027151675894856453, + 0.01021560188382864, + 0.015982672572135925, + 0.006308616604655981, + 0.006240495480597019, + 0.009525492787361145, + 0.009880347177386284, + 0.008027102798223495, + 0.012881901115179062, + 0.004681161139160395, + 0.007415807340294123, + 0.008426979184150696, + 0.007630256935954094, + 0.006605828180909157, + 0.010698010213673115, + 0.011305026710033417, + 0.009796952828764915, + 0.005486288573592901, + 0.006498057395219803, + 0.009840726852416992, + 0.01445605605840683, + 0.010833797045052052, + 0.015812234953045845, + 0.008824697695672512, + 0.009596656076610088, + 0.006787808146327734, + 0.009022748097777367, + 0.0038452313747256994, + 0.005222284700721502, + 0.004821616690605879, + 0.005643087439239025, + 0.0036535931285470724, + 0.005261566955596209, + 0.003172212978824973, + 0.00357951526530087, + 0.00789306964725256, + 0.009116598404943943, + 0.006916191894561052, + 0.007954263128340244, + 0.00892433151602745, + 0.014948978088796139, + 0.004773508757352829, + 0.009713985025882721, + 0.0074208565056324005, + 0.007468899246305227, + 0.009859642945230007, + 0.007828136906027794, + 0.007887516170740128, + 0.010404765605926514, + 0.004709556698799133, + 0.007656209636479616, + 0.007680620066821575, + 0.007038959767669439, + 0.008046681992709637, + 0.010877623222768307, + 0.008864290080964565, + 0.008043832145631313, + 0.0047980546951293945, + 0.006606749724596739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.010840808041393757 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609218120574951 + ], + "opposite_logprob_tokens": [ + -4.141092300415039 + ], + "logprob_reward": -0.01609218120574951, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609218120574951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3004406690597534, + 0.21723823249340057, + 0.06854532659053802, + 0.08031605929136276, + 0.03424340859055519, + 0.03556080162525177, + 0.05284048616886139, + 0.06109442561864853, + 0.10481036454439163, + 0.04491019994020462 + ], + "credit_map": [ + 0.0434330552816391, + 0.04402492567896843, + 0.03655540198087692, + 0.0629996508359909, + 0.033044345676898956, + 0.033761635422706604, + 0.019921084865927696, + 0.02670055627822876, + 0.029540833085775375, + 0.02954612672328949, + 0.02665896527469158, + 0.042886245995759964, + 0.02622266672551632, + 0.03167395293712616, + 0.014083992689847946, + 0.016625458374619484, + 0.007110089063644409, + 0.01090610958635807, + 0.011146617121994495, + 0.01187517587095499, + 0.008774114772677422, + 0.007171068340539932, + 0.0037330621853470802, + 0.007829086855053902, + 0.007970943115651608, + 0.00797885563224554, + 0.015262426808476448, + 0.008285622112452984, + 0.007275742944329977, + 0.01116764172911644, + 0.005160871893167496, + 0.017213962972164154, + 0.003783908672630787, + 0.0046063633635640144, + 0.003750643925741315, + 0.006151837762445211, + 0.004725157283246517, + 0.0038701826706528664, + 0.003503682091832161, + 0.0038516330532729626, + 0.0037140059284865856, + 0.00455156434327364, + 0.005619099363684654, + 0.0036847603041678667, + 0.0035610126797109842, + 0.005879445932805538, + 0.0026671963278204203, + 0.005883715581148863, + 0.0054523032158613205, + 0.006869782693684101, + 0.008597089909017086, + 0.008000781759619713, + 0.005609168671071529, + 0.007184974849224091, + 0.004643971100449562, + 0.0064824107103049755, + 0.0075522130355238914, + 0.005393998697400093, + 0.009186429902911186, + 0.007488266099244356, + 0.01026051864027977, + 0.011047612875699997, + 0.0035032969899475574, + 0.006662089843302965, + 0.013613336719572544, + 0.00995953194797039, + 0.0209907665848732, + 0.014885705895721912, + 0.008161594159901142, + 0.012377693317830563, + 0.007357284892350435, + 0.017464453354477882, + 0.004724616650491953, + 0.005143655464053154, + 0.007699972949922085, + 0.007058508228510618, + 0.004804292693734169, + 0.00485246442258358, + 0.0030558432918041945, + 0.007570847403258085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01609218120574951 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782115928828716 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782115928828716, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782115928828716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2494613081216812, + 0.144234761595726, + 0.10954473912715912, + 0.09872850775718689, + 0.07064324617385864, + 0.06370086222887039, + 0.05847710371017456, + 0.08192065358161926, + 0.059014275670051575, + 0.06427454203367233 + ], + "credit_map": [ + 0.034201689064502716, + 0.043189190328121185, + 0.02336370386183262, + 0.04498835653066635, + 0.04112556576728821, + 0.02342529408633709, + 0.019133996218442917, + 0.020033493638038635, + 0.0165215153247118, + 0.025586677715182304, + 0.015630612149834633, + 0.01437003631144762, + 0.02550382912158966, + 0.025307854637503624, + 0.010020569898188114, + 0.011293659918010235, + 0.010268674232065678, + 0.015955844894051552, + 0.009732889011502266, + 0.011517344042658806, + 0.022693432867527008, + 0.0195423886179924, + 0.007845649495720863, + 0.011988519690930843, + 0.009372367523610592, + 0.016448799520730972, + 0.01274613942950964, + 0.017623819410800934, + 0.006755612324923277, + 0.013460690155625343, + 0.006968694273382425, + 0.01535238977521658, + 0.006621356122195721, + 0.01332545280456543, + 0.01261446438729763, + 0.012259932234883308, + 0.006239783018827438, + 0.0063818576745688915, + 0.00585343549028039, + 0.0073469639755785465, + 0.005261039827018976, + 0.01164759136736393, + 0.011126955971121788, + 0.010257573798298836, + 0.004593248944729567, + 0.008483001962304115, + 0.005944141186773777, + 0.0063873110339045525, + 0.005049009807407856, + 0.012775070033967495, + 0.00830658059567213, + 0.011415264569222927, + 0.005964134819805622, + 0.005785102024674416, + 0.0041619534604251385, + 0.005019988864660263, + 0.007649474311619997, + 0.007117693778127432, + 0.0100823063403368, + 0.011407129466533661, + 0.011366662569344044, + 0.015728607773780823, + 0.0051574017852544785, + 0.013411376625299454, + 0.0066686668433249, + 0.004993891343474388, + 0.011621887795627117, + 0.00779287563636899, + 0.0035359200555831194, + 0.008652974851429462, + 0.00447859987616539, + 0.011269455775618553, + 0.005893407389521599, + 0.008015071973204613, + 0.01067026425153017, + 0.00791481975466013, + 0.006384236738085747, + 0.008915727026760578, + 0.005056316964328289, + 0.011424700729548931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009782115928828716 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016057105734944344 + ], + "opposite_logprob_tokens": [ + -4.141057014465332 + ], + "logprob_reward": -0.016057105734944344, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016057105734944344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4449097812175751, + 0.18701548874378204, + 0.03914865851402283, + 0.08997780829668045, + 0.060003023594617844, + 0.03647851571440697, + 0.025235753506422043, + 0.030631251633167267, + 0.05978268012404442, + 0.02681705541908741 + ], + "credit_map": [ + 0.06523450464010239, + 0.05056991055607796, + 0.05468083545565605, + 0.0646032765507698, + 0.07678790390491486, + 0.05333554744720459, + 0.04544830322265625, + 0.03424952179193497, + 0.024929633364081383, + 0.027628527954220772, + 0.019539063796401024, + 0.053020235151052475, + 0.022199885919690132, + 0.01271442323923111, + 0.01019347459077835, + 0.0167902372777462, + 0.004236581735312939, + 0.005599010270088911, + 0.006031187251210213, + 0.0042155128903687, + 0.00658847950398922, + 0.004607527982443571, + 0.003973357379436493, + 0.003896998940035701, + 0.008227699436247349, + 0.009304890409111977, + 0.014523988589644432, + 0.011024354957044125, + 0.007794461213052273, + 0.01525383535772562, + 0.006042963825166225, + 0.017805613577365875, + 0.010117958299815655, + 0.006819112692028284, + 0.011948755942285061, + 0.007436749991029501, + 0.005094115622341633, + 0.009049205109477043, + 0.004574109334498644, + 0.004963015206158161, + 0.004492597188800573, + 0.006073022726923227, + 0.0054555246606469154, + 0.0046358900144696236, + 0.0035155813675373793, + 0.005365283694118261, + 0.0022220578975975513, + 0.004718558397144079, + 0.0023991893976926804, + 0.0038172947242856026, + 0.004183144774287939, + 0.003967330325394869, + 0.0024308920837938786, + 0.003703779773786664, + 0.0019387014908716083, + 0.0027954212855547667, + 0.002986237872391939, + 0.0026218239217996597, + 0.005231302231550217, + 0.0035078264772892, + 0.004871961660683155, + 0.005034047644585371, + 0.0020796500612050295, + 0.004298401065170765, + 0.008160862140357494, + 0.00556698115542531, + 0.011564997024834156, + 0.0058492450043559074, + 0.0037136974278837442, + 0.009709488600492477, + 0.003723905421793461, + 0.011493504047393799, + 0.0027929467614740133, + 0.004084793385118246, + 0.0041623953729867935, + 0.004532034508883953, + 0.002508876146748662, + 0.0034808653872460127, + 0.0021891887299716473, + 0.0030659548938274384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.016057105734944344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943837575614452 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943837575614452, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943837575614452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20416367053985596, + 0.07182200253009796, + 0.09077586233615875, + 0.24096806347370148, + 0.1662231832742691, + 0.053504589945077896, + 0.023532535880804062, + 0.06446897238492966, + 0.05744796618819237, + 0.02709314227104187 + ], + "credit_map": [ + 0.024918438866734505, + 0.022348156198859215, + 0.030453821644186974, + 0.01697416976094246, + 0.03321203961968422, + 0.034567564725875854, + 0.024191508069634438, + 0.017497967928647995, + 0.007698944304138422, + 0.012290483340620995, + 0.006402584724128246, + 0.010178523138165474, + 0.015497563406825066, + 0.0068814679980278015, + 0.006924984976649284, + 0.005947454366832972, + 0.010063687339425087, + 0.01227855309844017, + 0.012705097906291485, + 0.010335031896829605, + 0.021304162219166756, + 0.008959677070379257, + 0.007058414630591869, + 0.008071237243711948, + 0.022135889157652855, + 0.025639252737164497, + 0.04547484591603279, + 0.029961636289954185, + 0.02007664181292057, + 0.04109938070178032, + 0.017448922619223595, + 0.039131492376327515, + 0.01715359278023243, + 0.014431171119213104, + 0.02771780639886856, + 0.019671406596899033, + 0.018877029418945312, + 0.03224319592118263, + 0.011791040189564228, + 0.02433793991804123, + 0.007413388695567846, + 0.00920353177934885, + 0.005800533574074507, + 0.009110532701015472, + 0.005857742857187986, + 0.006585114169865847, + 0.003918913658708334, + 0.0056148311123251915, + 0.0034414464607834816, + 0.002568967407569289, + 0.004763513337820768, + 0.0030712164007127285, + 0.002718627220019698, + 0.003248395398259163, + 0.0017327424138784409, + 0.0019876272417604923, + 0.007330260239541531, + 0.005614312365651131, + 0.006673878524452448, + 0.00492103211581707, + 0.009933996014297009, + 0.015943510457873344, + 0.0038986648432910442, + 0.010153314098715782, + 0.0067241559736430645, + 0.005765965674072504, + 0.011923082172870636, + 0.0053553800098598, + 0.0052022687159478664, + 0.008992145769298077, + 0.004055952187627554, + 0.009429019875824451, + 0.0038749556988477707, + 0.005247732158750296, + 0.00341459596529603, + 0.004003547132015228, + 0.002664274536073208, + 0.003230032278224826, + 0.002179564442485571, + 0.0024784409906715155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.005943837575614452 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot depicts two figures, dwarfed by the massive presence of a power plant's cooling towers. The scene is bathed in the soft, warm light of dusk, casting long, dramatic shadows across the gravelly ground. On the left stands an older man, his weathered face etched with the lines of experience. He wears a faded, dark blue work jumpsuit, a yellowed hard hat tucked under his arm. His posture is slightly relaxed, one hand gesturing deftly, his expression thoughtful. Beside him, a younger figure in a modern, dark green technical jacket and gray overalls listens intently, hands in pockets, head tilted slightly. They are engaged in a quiet conversation, their distance suggesting a shared, private moment. Massive concrete towers rise behind them, their richly textured surfaces capturing the golden hour, while white steam drifts lazily through vast openings, emitting pink and orange glows against the deepening twilight sky. The predominant color palette is a warm fusion of amber and rose, contrasted against the cool, industrial gray of the concrete, creating an atmosphere of deep contemplation and tranquility, highlighting moments of intimate human connection against a backdrop of immense power.", + "seed": 5697, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_11/index_9235_seed_5697_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012840440683066844, + "vlm_logprob": -0.012840440683066844, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.012840440683066844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01633368618786335 + ], + "opposite_logprob_tokens": [ + -5.26633358001709 + ], + "logprob_reward": -0.01633368618786335, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.01633368618786335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1843119114637375, + 0.03840601071715355, + 0.15263523161411285, + 0.05169602110981941, + 0.030981028452515602, + 0.047067057341337204, + 0.2317565381526947, + 0.15457962453365326, + 0.05125186964869499, + 0.057314708828926086 + ], + "credit_map": [ + 0.030205750837922096, + 0.01919987052679062, + 0.025226382538676262, + 0.019694825634360313, + 0.019187072291970253, + 0.02610684558749199, + 0.020020173862576485, + 0.024670979008078575, + 0.005550912581384182, + 0.004073688294738531, + 0.005959372967481613, + 0.0062058898620307446, + 0.0041602193377912045, + 0.0035560312680900097, + 0.002980619203299284, + 0.005919272545725107, + 0.023005584254860878, + 0.012415234930813313, + 0.028795151039958, + 0.020679937675595284, + 0.012603759765625, + 0.011947240680456161, + 0.017072996124625206, + 0.026115315034985542, + 0.005694101098924875, + 0.0052667162381112576, + 0.0077805533073842525, + 0.008175292983651161, + 0.004615139216184616, + 0.004875168204307556, + 0.005111307371407747, + 0.010177738033235073, + 0.003684302791953087, + 0.0037122820504009724, + 0.004031697288155556, + 0.004032815340906382, + 0.005239209625869989, + 0.003457730170339346, + 0.0031065992079675198, + 0.0037163887172937393, + 0.009539895690977573, + 0.00452025281265378, + 0.011163030751049519, + 0.004568830598145723, + 0.004972221329808235, + 0.0037832120433449745, + 0.003764370223507285, + 0.004755241330713034, + 0.05604740232229233, + 0.015601888298988342, + 0.09413954615592957, + 0.014214384369552135, + 0.011498847976326942, + 0.010200373828411102, + 0.014622535556554794, + 0.015431540086865425, + 0.02812625840306282, + 0.011323509737849236, + 0.04010598361492157, + 0.021667271852493286, + 0.012154349125921726, + 0.010277231223881245, + 0.012155345641076565, + 0.01876966282725334, + 0.009375440888106823, + 0.005149985663592815, + 0.011167556047439575, + 0.005942052695900202, + 0.0041032470762729645, + 0.004383785650134087, + 0.003107532160356641, + 0.008022267371416092, + 0.009695352055132389, + 0.005144904367625713, + 0.01389382965862751, + 0.007037545554339886, + 0.004723451565951109, + 0.005279813893139362, + 0.004082918632775545, + 0.007456890773028135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.01633368618786335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609441079199314 + ], + "opposite_logprob_tokens": [ + -4.141094207763672 + ], + "logprob_reward": -0.01609441079199314, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609441079199314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06682231277227402, + 0.02414523810148239, + 0.06998449563980103, + 0.019669359549880028, + 0.04930770397186279, + 0.0538548044860363, + 0.2294086217880249, + 0.4566632807254791, + 0.01589079201221466, + 0.014253373257815838 + ], + "credit_map": [ + 0.00946701131761074, + 0.006735024508088827, + 0.008423207327723503, + 0.00706502515822649, + 0.008782247081398964, + 0.008841263130307198, + 0.00874409731477499, + 0.00876444112509489, + 0.004110998939722776, + 0.0030596747528761625, + 0.003367068711668253, + 0.003499214770272374, + 0.0024071529041975737, + 0.0021045927423983812, + 0.0018694393802434206, + 0.003727095667272806, + 0.00915573351085186, + 0.007430885918438435, + 0.012971751391887665, + 0.010148733854293823, + 0.00670256931334734, + 0.005686260759830475, + 0.007033645641058683, + 0.010854916647076607, + 0.0039053072687238455, + 0.0021089662332087755, + 0.0037118704058229923, + 0.0021525253541767597, + 0.0015361574478447437, + 0.001661607064306736, + 0.0013936941977590322, + 0.0031992325093597174, + 0.004438032396137714, + 0.004584020469337702, + 0.004330216906964779, + 0.010787198320031166, + 0.005067664198577404, + 0.004017872270196676, + 0.010058693587779999, + 0.00602400628849864, + 0.012024824507534504, + 0.003831131150946021, + 0.01659999042749405, + 0.006542224902659655, + 0.0035972718615084887, + 0.00340447761118412, + 0.0033793668262660503, + 0.004475517198443413, + 0.05899091064929962, + 0.007996308617293835, + 0.10551294684410095, + 0.012324390932917595, + 0.008641556836664677, + 0.008709978312253952, + 0.013101001270115376, + 0.014131522737443447, + 0.07761712372303009, + 0.032175272703170776, + 0.1037662923336029, + 0.07568927854299545, + 0.03458390757441521, + 0.03110627830028534, + 0.039053283631801605, + 0.06267184764146805, + 0.002692019334062934, + 0.001119615277275443, + 0.0027731771115213633, + 0.0015951045788824558, + 0.0015111974207684398, + 0.0025568262208253145, + 0.00117972563020885, + 0.002463126555085182, + 0.002105200197547674, + 0.0011744784424081445, + 0.0022619564551860094, + 0.001573461457155645, + 0.001389360404573381, + 0.0022292404901236296, + 0.001074745669029653, + 0.0024449299089610577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.01609441079199314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082286946475506 + ], + "opposite_logprob_tokens": [ + -4.511082172393799 + ], + "logprob_reward": -0.011082286946475506, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082286946475506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09289859980344772, + 0.032597221434116364, + 0.08752718567848206, + 0.02708013728260994, + 0.042708802968263626, + 0.06138428673148155, + 0.24596713483333588, + 0.37501171231269836, + 0.0160477664321661, + 0.018777109682559967 + ], + "credit_map": [ + 0.013122840784490108, + 0.005873859394341707, + 0.014170299284160137, + 0.010105288587510586, + 0.008909342810511589, + 0.010971988551318645, + 0.018461985513567924, + 0.011282999068498611, + 0.004661571700125933, + 0.0030342494137585163, + 0.005464923568069935, + 0.003926368895918131, + 0.004096267744898796, + 0.004383082967251539, + 0.002963022096082568, + 0.004067736677825451, + 0.011219827458262444, + 0.009448735043406487, + 0.01479374524205923, + 0.013662289828062057, + 0.00758548267185688, + 0.006689609028398991, + 0.01137330662459135, + 0.012754199095070362, + 0.005474133882671595, + 0.002329290146008134, + 0.005267471540719271, + 0.0024506880436092615, + 0.0022120093926787376, + 0.0024646776728332043, + 0.0019188231090083718, + 0.004963046405464411, + 0.004202285781502724, + 0.004247148986905813, + 0.004381117876619101, + 0.009254204109311104, + 0.0046889884397387505, + 0.0043251654133200645, + 0.005984228570014238, + 0.005625666584819555, + 0.008855608291924, + 0.006283582653850317, + 0.016900597140192986, + 0.009180921129882336, + 0.004774763248860836, + 0.0046245623379945755, + 0.004916077014058828, + 0.005848177708685398, + 0.06276542693376541, + 0.008625106886029243, + 0.11222085356712341, + 0.014332707040011883, + 0.009604203514754772, + 0.009217234328389168, + 0.014304449781775475, + 0.01489719096571207, + 0.06619197875261307, + 0.02599332109093666, + 0.08933984488248825, + 0.05917990207672119, + 0.027151867747306824, + 0.0251605361700058, + 0.03170892596244812, + 0.05028536170721054, + 0.002688030945137143, + 0.0009569198591634631, + 0.004122892860323191, + 0.001282915472984314, + 0.001212062081322074, + 0.0020572447683662176, + 0.0010763363679870963, + 0.00265136593952775, + 0.002653033472597599, + 0.0015999010065570474, + 0.0032744777854532003, + 0.0034637246280908585, + 0.0012955504935234785, + 0.0019285714952275157, + 0.0018185923108831048, + 0.0027432595379650593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.011082286946475506 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605675369501114 + ], + "opposite_logprob_tokens": [ + -4.141056537628174 + ], + "logprob_reward": -0.01605675369501114, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605675369501114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05164142698049545, + 0.030337655916810036, + 0.04620169475674629, + 0.03134237974882126, + 0.05747508630156517, + 0.20320917665958405, + 0.3912517726421356, + 0.14218361675739288, + 0.025132615119218826, + 0.021224558353424072 + ], + "credit_map": [ + 0.007624389603734016, + 0.005877900402992964, + 0.009558788500726223, + 0.006251460872590542, + 0.005056001711636782, + 0.004965624772012234, + 0.004281218163669109, + 0.00802604015916586, + 0.005126993637531996, + 0.00532220583409071, + 0.005061986390501261, + 0.0031616082414984703, + 0.003214823780581355, + 0.0026397916954010725, + 0.0019341707229614258, + 0.003876074217259884, + 0.00667398190125823, + 0.005659534130245447, + 0.006887312978506088, + 0.0056726280599832535, + 0.004594953265041113, + 0.00435723876580596, + 0.004134019836783409, + 0.008222028613090515, + 0.004522234667092562, + 0.00452879723161459, + 0.003879099152982235, + 0.0033674670848995447, + 0.003514878684654832, + 0.0029161192942410707, + 0.002675803145393729, + 0.005937977693974972, + 0.008598082698881626, + 0.005932295229285955, + 0.007997619919478893, + 0.007785207126289606, + 0.004643025808036327, + 0.004093941766768694, + 0.010605055838823318, + 0.007819860242307186, + 0.05009955167770386, + 0.015049639157950878, + 0.08428405970335007, + 0.01239550020545721, + 0.0080331452190876, + 0.007158581167459488, + 0.0109885074198246, + 0.015200195834040642, + 0.09984391927719116, + 0.01760007068514824, + 0.17782831192016602, + 0.022824564948678017, + 0.013115860521793365, + 0.014148334972560406, + 0.021744269877672195, + 0.024146465584635735, + 0.023939955979585648, + 0.009750054217875004, + 0.03248574957251549, + 0.023638514801859856, + 0.01113195437937975, + 0.009504337795078754, + 0.011938685551285744, + 0.019794374704360962, + 0.004167732782661915, + 0.0022329494822770357, + 0.006143341306596994, + 0.00193401041906327, + 0.001838947762735188, + 0.0025624867994338274, + 0.0015784894349053502, + 0.0046746572479605675, + 0.0040067643858492374, + 0.002000667154788971, + 0.004928682465106249, + 0.0020690627861768007, + 0.001981962937861681, + 0.001988762989640236, + 0.001243166858330369, + 0.003005487844347954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.01605675369501114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635065793991089 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004635065793991089, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635065793991089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14443771541118622, + 0.029804455116391182, + 0.04249076545238495, + 0.059810247272253036, + 0.0512847863137722, + 0.08952540159225464, + 0.30774936079978943, + 0.22095943987369537, + 0.03285890072584152, + 0.021078867837786674 + ], + "credit_map": [ + 0.023612013086676598, + 0.020299741998314857, + 0.01915203593671322, + 0.016482533887028694, + 0.014924840070307255, + 0.018200861290097237, + 0.01394161581993103, + 0.017824072390794754, + 0.004947623237967491, + 0.0036880781408399343, + 0.004625922068953514, + 0.004964824765920639, + 0.003447841154411435, + 0.002757619135081768, + 0.001846029539592564, + 0.0035265162587165833, + 0.005369578953832388, + 0.006320205517113209, + 0.00764017179608345, + 0.004863325040787458, + 0.004775789100676775, + 0.0041645970195531845, + 0.0030010840855538845, + 0.006356015335768461, + 0.009997926652431488, + 0.00479398388415575, + 0.008324972353875637, + 0.007889879867434502, + 0.00498249800875783, + 0.004927631933242083, + 0.005916273221373558, + 0.012977084144949913, + 0.006887504830956459, + 0.005499731283634901, + 0.006279184017330408, + 0.004635712597519159, + 0.003918246831744909, + 0.004357466008514166, + 0.010882040485739708, + 0.008824902586638927, + 0.016568204388022423, + 0.008705933578312397, + 0.029082125052809715, + 0.008573045954108238, + 0.0068665798753499985, + 0.0054747965186834335, + 0.005832231603562832, + 0.00842248648405075, + 0.0773572027683258, + 0.014990226365625858, + 0.13833025097846985, + 0.01598464697599411, + 0.011402668431401253, + 0.011893629096448421, + 0.017190171405673027, + 0.02060057409107685, + 0.039456646889448166, + 0.015327798202633858, + 0.052830129861831665, + 0.03495073318481445, + 0.0159104373306036, + 0.014614146202802658, + 0.0185221079736948, + 0.029347434639930725, + 0.005409944336861372, + 0.0019859643653035164, + 0.008610093966126442, + 0.00345920049585402, + 0.0022174157202243805, + 0.0028142565861344337, + 0.002313141943886876, + 0.006048882845789194, + 0.003066203324124217, + 0.00218484690412879, + 0.004481097217649221, + 0.0026573925279080868, + 0.0016850346000865102, + 0.002151679480448365, + 0.001599217881448567, + 0.0032533977646380663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.004635065793991089 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot depicts two figures, dwarfed by the massive presence of a power plant's cooling towers. The scene is bathed in the soft, warm light of dusk, casting long, dramatic shadows across the gravelly ground. On the left stands an older man, his weathered face etched with the lines of experience. He wears a faded, dark blue work jumpsuit, a yellowed hard hat tucked under his arm. His posture is slightly relaxed, one hand gesturing deftly, his expression thoughtful. Beside him, a younger figure in a modern, dark green technical jacket and gray overalls listens intently, hands in pockets, head tilted slightly. They are engaged in a quiet conversation, their distance suggesting a shared, private moment. Massive concrete towers rise behind them, their richly textured surfaces capturing the golden hour, while white steam drifts lazily through vast openings, emitting pink and orange glows against the deepening twilight sky. The predominant color palette is a warm fusion of amber and rose, contrasted against the cool, industrial gray of the concrete, creating an atmosphere of deep contemplation and tranquility, highlighting moments of intimate human connection against a backdrop of immense power.", + "seed": 5698, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_11/index_9235_seed_5698_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012144836084917187, + "vlm_logprob": -0.012144836084917187, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.012144836084917187, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012608692981302738 + ], + "opposite_logprob_tokens": [ + -5.887608528137207 + ], + "logprob_reward": -0.012608692981302738, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.012608692981302738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3411228358745575, + 0.08037173002958298, + 0.035613641142845154, + 0.02964567206799984, + 0.09062469005584717, + 0.08119368553161621, + 0.25037989020347595, + 0.035773828625679016, + 0.022228503599762917, + 0.03304555267095566 + ], + "credit_map": [ + 0.040079936385154724, + 0.043012503534555435, + 0.05007389187812805, + 0.03515978902578354, + 0.04905164614319801, + 0.0477161668241024, + 0.0375199168920517, + 0.038508977741003036, + 0.011738400906324387, + 0.007388605270534754, + 0.01322397030889988, + 0.011393488384783268, + 0.0071472348645329475, + 0.006963103078305721, + 0.00839327648282051, + 0.014123646542429924, + 0.006973298266530037, + 0.003986923489719629, + 0.006552116014063358, + 0.002990574575960636, + 0.003995327744632959, + 0.002937549026682973, + 0.003310286905616522, + 0.004867562558501959, + 0.004895927384495735, + 0.0029851666186004877, + 0.005228940397500992, + 0.002158890012651682, + 0.002585293957963586, + 0.004504799377173185, + 0.0018703198293223977, + 0.00541633740067482, + 0.013361342251300812, + 0.006439405493438244, + 0.009496251121163368, + 0.007853428833186626, + 0.0061582159250974655, + 0.01897704228758812, + 0.007186840754002333, + 0.02115216851234436, + 0.009278268553316593, + 0.008264290168881416, + 0.011363699100911617, + 0.009962412528693676, + 0.007279671262949705, + 0.01134385634213686, + 0.008626824244856834, + 0.015074660070240498, + 0.030156662687659264, + 0.029091212898492813, + 0.043303072452545166, + 0.02096235752105713, + 0.024571092799305916, + 0.03641021251678467, + 0.022983143106102943, + 0.04290211945772171, + 0.0046161324717104435, + 0.0025626281276345253, + 0.003817385295405984, + 0.003851895919069648, + 0.004360191524028778, + 0.007088494021445513, + 0.0035233963280916214, + 0.005953703075647354, + 0.0025782205630093813, + 0.002986560808494687, + 0.002806613687425852, + 0.0034996839240193367, + 0.001999093685299158, + 0.0029452547896653414, + 0.00237935408949852, + 0.003033722983673215, + 0.005065735429525375, + 0.005165853537619114, + 0.004557709209620953, + 0.003440216649323702, + 0.003868731902912259, + 0.002879257081076503, + 0.004189997911453247, + 0.003878048388287425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.012608692981302738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215044677257538 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.014215044677257538, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215044677257538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20386214554309845, + 0.022246897220611572, + 0.042072758078575134, + 0.016980642452836037, + 0.3347853720188141, + 0.14444498717784882, + 0.12129808217287064, + 0.016244225203990936, + 0.04810573160648346, + 0.04995924234390259 + ], + "credit_map": [ + 0.01908181421458721, + 0.02230101265013218, + 0.025623897090554237, + 0.027395330369472504, + 0.02853330597281456, + 0.01921490766108036, + 0.032621853053569794, + 0.029090026393532753, + 0.0027565290220081806, + 0.0024075389374047518, + 0.00285004242323339, + 0.0029691276140511036, + 0.0032024793326854706, + 0.0018803478451445699, + 0.003311489475890994, + 0.002869340591132641, + 0.007248875685036182, + 0.004357277415692806, + 0.0068841297179460526, + 0.0032321803737431765, + 0.004438937641680241, + 0.004903679713606834, + 0.003960238303989172, + 0.007047436200082302, + 0.002720833057537675, + 0.002391089918091893, + 0.00282288552261889, + 0.0016721936408430338, + 0.001990711549296975, + 0.0017319796606898308, + 0.0016825087368488312, + 0.0019684401340782642, + 0.050998084247112274, + 0.025503825396299362, + 0.038832321763038635, + 0.025425653904676437, + 0.025281917303800583, + 0.0654730573296547, + 0.02566583827137947, + 0.07760464400053024, + 0.018661698326468468, + 0.013644332066178322, + 0.02695017121732235, + 0.01561234425753355, + 0.018711242824792862, + 0.01214855257421732, + 0.017475174739956856, + 0.021241480484604836, + 0.014894668012857437, + 0.01187958661466837, + 0.021153630688786507, + 0.010464504361152649, + 0.012231938540935516, + 0.01789761148393154, + 0.011453027836978436, + 0.02132311463356018, + 0.00196135719306767, + 0.002086889697238803, + 0.0019676978699862957, + 0.0013125480618327856, + 0.001990122487768531, + 0.0027840009424835443, + 0.0014447709545493126, + 0.0026968384627252817, + 0.007539556827396154, + 0.008083859458565712, + 0.009062791243195534, + 0.00501013221219182, + 0.004379082005470991, + 0.0045684813521802425, + 0.004661838989704847, + 0.004799986723810434, + 0.006395913660526276, + 0.012633646838366985, + 0.005629956256598234, + 0.006252654362469912, + 0.004923575092107058, + 0.0044359783641994, + 0.004241773393005133, + 0.005445742979645729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.014215044677257538 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537943199276924 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537943199276924, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537943199276924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4444178640842438, + 0.04627252742648125, + 0.0638241097331047, + 0.02256312407553196, + 0.19716668128967285, + 0.07484613358974457, + 0.07924092561006546, + 0.04678376764059067, + 0.01044954638928175, + 0.014435337856411934 + ], + "credit_map": [ + 0.04205507040023804, + 0.0489908903837204, + 0.05305120721459389, + 0.05301938205957413, + 0.06382699310779572, + 0.05734728276729584, + 0.06040354445576668, + 0.0657234862446785, + 0.00886678509414196, + 0.003127076430246234, + 0.007527416106313467, + 0.00484366063028574, + 0.0038118131924420595, + 0.004485315643250942, + 0.004228607751429081, + 0.009381855838000774, + 0.009792855940759182, + 0.006880329921841621, + 0.009903454221785069, + 0.005689695477485657, + 0.00815467070788145, + 0.007587432395666838, + 0.006721986923366785, + 0.009093689732253551, + 0.0034840372391045094, + 0.003226896747946739, + 0.0037030186504125595, + 0.002099626464769244, + 0.0036019047256559134, + 0.0015177989844232798, + 0.0028430833481252193, + 0.002086758380755782, + 0.03029542975127697, + 0.015403893776237965, + 0.023365171626210213, + 0.013823146000504494, + 0.015111973509192467, + 0.03858591988682747, + 0.014936555176973343, + 0.04564457759261131, + 0.014826027676463127, + 0.007630201987922192, + 0.013845865614712238, + 0.004911485128104687, + 0.00955352745950222, + 0.0065210615284740925, + 0.006831076927483082, + 0.010726883076131344, + 0.00960956048220396, + 0.00767466751858592, + 0.013487488962709904, + 0.006625300273299217, + 0.007926994003355503, + 0.01236893329769373, + 0.0071738422848284245, + 0.01437414437532425, + 0.006634620949625969, + 0.002540897112339735, + 0.007178331725299358, + 0.0031146067194640636, + 0.00454690121114254, + 0.009306361898779869, + 0.0037061513867229223, + 0.009755895473062992, + 0.0014577453257516026, + 0.001657735207118094, + 0.002003738656640053, + 0.0009263547835871577, + 0.0007772443350404501, + 0.0013487236574292183, + 0.0008804534445516765, + 0.001397551386617124, + 0.0022357702255249023, + 0.0027946478221565485, + 0.0023934070486575365, + 0.0016669537872076035, + 0.001199103775434196, + 0.0010548388818278909, + 0.0010935498867183924, + 0.00199706619605422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.012537943199276924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018173139542341232 + ], + "opposite_logprob_tokens": [ + -4.0181732177734375 + ], + "logprob_reward": -0.018173139542341232, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018173139542341232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3917182683944702, + 0.09378471225500107, + 0.08345701545476913, + 0.10916213691234589, + 0.11891680955886841, + 0.06360860168933868, + 0.05609128251671791, + 0.04649052023887634, + 0.01849118433892727, + 0.018279490992426872 + ], + "credit_map": [ + 0.03894875571131706, + 0.047816142439842224, + 0.04554315656423569, + 0.04750598222017288, + 0.0620204322040081, + 0.051321230828762054, + 0.05357583612203598, + 0.04498668387532234, + 0.01620415411889553, + 0.0071448227390646935, + 0.01781405508518219, + 0.011619939468801022, + 0.007592450361698866, + 0.007998054847121239, + 0.009899724274873734, + 0.015511496923863888, + 0.016767779365181923, + 0.007933668792247772, + 0.022027553990483284, + 0.006256698630750179, + 0.0072539448738098145, + 0.006169476080685854, + 0.007207509130239487, + 0.009840377606451511, + 0.024007761850953102, + 0.008975672535598278, + 0.029972145333886147, + 0.007044103927910328, + 0.00885895173996687, + 0.009025854989886284, + 0.007904141210019588, + 0.013373489491641521, + 0.017895227298140526, + 0.011115736328065395, + 0.014655555598437786, + 0.0074651953764259815, + 0.009220553562045097, + 0.023172006011009216, + 0.008531572297215462, + 0.02686094492673874, + 0.008957871235907078, + 0.01477219257503748, + 0.00848422758281231, + 0.005111120641231537, + 0.008566147647798061, + 0.004994822666049004, + 0.005846814718097448, + 0.006875399965792894, + 0.007898463867604733, + 0.004209334030747414, + 0.010481156408786774, + 0.004264902789145708, + 0.0058051301166415215, + 0.007551189977675676, + 0.005525038577616215, + 0.010356060229241848, + 0.005862956400960684, + 0.0036125604528933764, + 0.006554466672241688, + 0.0032372293062508106, + 0.0044007450342178345, + 0.009241510182619095, + 0.0037325983867049217, + 0.00984844658523798, + 0.0026169943157583475, + 0.0032038153149187565, + 0.0036085292231291533, + 0.0017290615942329168, + 0.0017837040359154344, + 0.0015094985719770193, + 0.0015325610293075442, + 0.0025070200208574533, + 0.0023980275727808475, + 0.003664448158815503, + 0.0028718477115035057, + 0.0018635985907167196, + 0.0015925528714433312, + 0.0015752753242850304, + 0.001989104552194476, + 0.0023246337659657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.018173139542341232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two figures visible in the scene, one older and one younger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the older man wearing a faded dark blue work jumpsuit with a yellowed hard hat under his arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the younger figure wearing a modern dark green technical jacket and gray overalls with hands in pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are massive concrete cooling towers visible in the background with white steam emitting pink and orange glows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the soft warm light of dusk with a color palette of amber, rose, and industrial gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189360024407506 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.003189360024407506, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189360024407506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5853924751281738, + 0.022058038040995598, + 0.03277704864740372, + 0.01556780282407999, + 0.16250379383563995, + 0.0561051182448864, + 0.05797938257455826, + 0.04327340051531792, + 0.012010627426207066, + 0.012332300655543804 + ], + "credit_map": [ + 0.06746906042098999, + 0.06810645014047623, + 0.08771469444036484, + 0.06716004759073257, + 0.08388084173202515, + 0.07047329097986221, + 0.07103919982910156, + 0.06954885274171829, + 0.003392553888261318, + 0.001909969374537468, + 0.003075894434005022, + 0.0023034687619656324, + 0.0035174081567674875, + 0.002040795050561428, + 0.0028749071061611176, + 0.002943041967228055, + 0.006853881757706404, + 0.0028815565165132284, + 0.00573594169691205, + 0.002474013715982437, + 0.0032765273936092854, + 0.0035118605010211468, + 0.002753376727923751, + 0.005289891269057989, + 0.0038201212882995605, + 0.0013715215027332306, + 0.0024084444157779217, + 0.001165473135188222, + 0.001477473764680326, + 0.0017897364450618625, + 0.0012289822334423661, + 0.002306049456819892, + 0.024663599207997322, + 0.01280425675213337, + 0.019905652850866318, + 0.011505438946187496, + 0.011885645799338818, + 0.0321248397231102, + 0.012218372896313667, + 0.03739597648382187, + 0.006118587218225002, + 0.008565996773540974, + 0.0067547280341386795, + 0.00431012362241745, + 0.0055398158729076385, + 0.009784668684005737, + 0.004230653401464224, + 0.01080054510384798, + 0.007353681605309248, + 0.007038315758109093, + 0.011456146836280823, + 0.004939776845276356, + 0.0059616826474666595, + 0.0067188250832259655, + 0.005386809818446636, + 0.009124143980443478, + 0.005171380005776882, + 0.004707422107458115, + 0.006394022144377232, + 0.0033179919701069593, + 0.004068860784173012, + 0.007829179987311363, + 0.0034397768322378397, + 0.00834476388990879, + 0.001323707401752472, + 0.0021457429975271225, + 0.0014604858588427305, + 0.0014791331486776471, + 0.001410432974807918, + 0.0009348620660603046, + 0.0011510697659105062, + 0.00210519228130579, + 0.0012415334349498153, + 0.0020718155428767204, + 0.0016282733995467424, + 0.001350408187136054, + 0.0013817257713526487, + 0.0007713879458606243, + 0.0014891527825966477, + 0.0023980033583939075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.003189360024407506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_12/index_tensor([541, 541])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_12/index_tensor([541, 541])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d53ce68f65c8562b2aff85419e17e390337ba45f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_12/index_tensor([541, 541])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful, dynamic tracking shot begins high above the clouds, then pitches downward at a steep swooping angle, closely following a massive flying dragon. It first folds its vast, sky-covering wings inward, clinging to its massive frame. The movement is swift and decisive, like folding steel. The scales covering its body are no ordinary substance, but a cold, metallic, obsidian. In the glaring sunlight filtering through the gaps in the clouds, they refract shifting shades of deep blue and purple, each subtle shift in angle creating a dazzling display of light. Its serpentine neck stretches forward, its spiky head lowered, its golden eyes, like molten lava, fixed fixedly on an unknown target below. As its wings fold, it instantly transforms into a living meteor, plummeting with astonishing acceleration, tearing through the thick, swirling cumulonimbus clouds and leaving a long, white trail of vapor and vortexes in its wake. The whole picture is full of a sense of speed and unparalleled power. The air makes a sharp whistling sound due to its high-speed movement, creating an epic atmosphere that is grand, magnificent and highly oppressive.", + "seed": 6565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_12/index_541_seed_6565_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01845066510140896, + "vlm_logprob": -0.01845066510140896, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.01845066510140896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006437635514885187 + ], + "opposite_logprob_tokens": [ + -5.6314377784729 + ], + "logprob_reward": -0.006437635514885187, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006437635514885187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09787257760763168, + 0.10560481995344162, + 0.2529281675815582, + 0.11495669931173325, + 0.05073970556259155, + 0.10653407126665115, + 0.053316548466682434, + 0.0832514539361, + 0.09876720607280731, + 0.03602874279022217 + ], + "credit_map": [ + 0.010508588515222073, + 0.008199254050850868, + 0.015736069530248642, + 0.013745114207267761, + 0.010878539644181728, + 0.012806825339794159, + 0.012391128577291965, + 0.013607059605419636, + 0.014508959837257862, + 0.014233832247555256, + 0.008665294386446476, + 0.016758538782596588, + 0.014024930074810982, + 0.009501838125288486, + 0.015546589158475399, + 0.012364831753075123, + 0.07584325224161148, + 0.022742431610822678, + 0.016577769070863724, + 0.013784762471914291, + 0.014256797730922699, + 0.017575975507497787, + 0.0679992288351059, + 0.024147942662239075, + 0.014876584522426128, + 0.012796384282410145, + 0.026146061718463898, + 0.013081748969852924, + 0.007255577016621828, + 0.013562750071287155, + 0.01037924736738205, + 0.01685834303498268, + 0.004753246437758207, + 0.0063689518719911575, + 0.007434382103383541, + 0.006880787201225758, + 0.005553630646318197, + 0.005408931523561478, + 0.006601126864552498, + 0.007738651242107153, + 0.006770676467567682, + 0.013154233805835247, + 0.011744031682610512, + 0.01365289743989706, + 0.010071380995213985, + 0.017221197485923767, + 0.008583134040236473, + 0.02533651888370514, + 0.0045360540971159935, + 0.007404661737382412, + 0.009841067716479301, + 0.009881717152893543, + 0.006287730298936367, + 0.003919963724911213, + 0.004848075564950705, + 0.00659727631136775, + 0.006827359087765217, + 0.009046602062880993, + 0.012008123099803925, + 0.011998014524579048, + 0.010996179655194283, + 0.01114679779857397, + 0.008640460669994354, + 0.012587917037308216, + 0.008872481994330883, + 0.009266743436455727, + 0.014984866604208946, + 0.029877562075853348, + 0.004856964107602835, + 0.014195692725479603, + 0.009116941131651402, + 0.00759595911949873, + 0.0044469949789345264, + 0.00470191752538085, + 0.007225835230201483, + 0.004821526352316141, + 0.003608450060710311, + 0.003010094165802002, + 0.003873273963108659, + 0.004340650513768196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.006437635514885187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865613952279091 + ], + "opposite_logprob_tokens": [ + -3.0486562252044678 + ], + "logprob_reward": -0.04865613952279091, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865613952279091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13373936712741852, + 0.047915540635585785, + 0.4832538962364197, + 0.07752948254346848, + 0.025998767465353012, + 0.02987557463347912, + 0.022095253691077232, + 0.07413160055875778, + 0.07303161174058914, + 0.032428935170173645 + ], + "credit_map": [ + 0.011561354622244835, + 0.010777300223708153, + 0.027263827621936798, + 0.0234365351498127, + 0.01244991272687912, + 0.01713821478188038, + 0.01339687965810299, + 0.017715346068143845, + 0.009009684436023235, + 0.007861585356295109, + 0.003697400214150548, + 0.006980761885643005, + 0.005271682050079107, + 0.00387545651756227, + 0.0065346877090632915, + 0.004684280138462782, + 0.14471203088760376, + 0.04366697371006012, + 0.03206060081720352, + 0.02511664852499962, + 0.02703148126602173, + 0.033314451575279236, + 0.1306028664112091, + 0.04674884304404259, + 0.004666128195822239, + 0.0036979233846068382, + 0.007139259949326515, + 0.02419157512485981, + 0.003768926952034235, + 0.005075555760413408, + 0.006657419726252556, + 0.022332698106765747, + 0.0019079791381955147, + 0.0018783134873956442, + 0.0030887338798493147, + 0.004190816078335047, + 0.0031862908508628607, + 0.0031319179106503725, + 0.003075665095821023, + 0.005539048928767443, + 0.0017921322723850608, + 0.0027298331260681152, + 0.003369480138644576, + 0.00536068482324481, + 0.002455637324601412, + 0.005192016717046499, + 0.0025350742507725954, + 0.00644071726128459, + 0.0018786382861435413, + 0.002493893960490823, + 0.0055454205721616745, + 0.0040894970297813416, + 0.0022038030438125134, + 0.0021133983973413706, + 0.0017818886553868651, + 0.0019887133967131376, + 0.009420971386134624, + 0.008174787275493145, + 0.011328351683914661, + 0.010906822048127651, + 0.008537882938981056, + 0.012583895586431026, + 0.006836943328380585, + 0.006341944914311171, + 0.006059322971850634, + 0.006411178968846798, + 0.011196873150765896, + 0.02359677478671074, + 0.0032782829366624355, + 0.011035734787583351, + 0.006847497541457415, + 0.004605945199728012, + 0.003998049069195986, + 0.00307682016864419, + 0.006890323478728533, + 0.00416959635913372, + 0.004203103948384523, + 0.0029905952978879213, + 0.003184688277542591, + 0.003915760666131973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.04865613952279091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280853405594826 + ], + "opposite_logprob_tokens": [ + -3.7732808589935303 + ], + "logprob_reward": -0.023280853405594826, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280853405594826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10420539975166321, + 0.04533546045422554, + 0.32818350195884705, + 0.03425963595509529, + 0.027666717767715454, + 0.02444128692150116, + 0.09791819006204605, + 0.1107717901468277, + 0.18656878173351288, + 0.040649209171533585 + ], + "credit_map": [ + 0.009638496674597263, + 0.008909827098250389, + 0.017021259292960167, + 0.016986852511763573, + 0.013580980710685253, + 0.011289185844361782, + 0.012579250149428844, + 0.014199549332261086, + 0.008827839978039265, + 0.008659848012030125, + 0.005522038787603378, + 0.004374906420707703, + 0.004294217564165592, + 0.00387703999876976, + 0.00529314111918211, + 0.00448642997071147, + 0.0987042635679245, + 0.03016718477010727, + 0.02126394584774971, + 0.018239259719848633, + 0.018002541735768318, + 0.0226103775203228, + 0.0873243510723114, + 0.03187157213687897, + 0.004857136867940426, + 0.0035345114301890135, + 0.0052079991437494755, + 0.004169608000665903, + 0.003542743157595396, + 0.005513323005288839, + 0.0027720534708350897, + 0.004662259016185999, + 0.0020435412880033255, + 0.0028885698411613703, + 0.0038288047071546316, + 0.004446996841579676, + 0.0030534821562469006, + 0.0030315460171550512, + 0.0030414974316954613, + 0.005332279484719038, + 0.0021464230958372355, + 0.002649371512234211, + 0.003414195729419589, + 0.002941424958407879, + 0.0028161611407995224, + 0.0032595363445580006, + 0.0037405486218631268, + 0.0034736276138573885, + 0.00749040674418211, + 0.009712706319987774, + 0.025593649595975876, + 0.01857704110443592, + 0.012730726972222328, + 0.008411872200667858, + 0.007103276904672384, + 0.00829851720482111, + 0.012070251628756523, + 0.012489641085267067, + 0.015321427956223488, + 0.013437805697321892, + 0.020197225734591484, + 0.015492388978600502, + 0.010564375668764114, + 0.011198683641850948, + 0.01745094358921051, + 0.01560168992727995, + 0.026855578646063805, + 0.05201881378889084, + 0.010350145399570465, + 0.02650711126625538, + 0.026361215859651566, + 0.011423282325267792, + 0.004147866275161505, + 0.005675370339304209, + 0.005939772352576256, + 0.00582065200433135, + 0.003945075906813145, + 0.0036681005731225014, + 0.006167316343635321, + 0.005285051651299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.023280853405594826 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008630831725895405 + ], + "opposite_logprob_tokens": [ + -4.758630752563477 + ], + "logprob_reward": -0.008630831725895405, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008630831725895405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3404427170753479, + 0.11419092863798141, + 0.08021609485149384, + 0.15805281698703766, + 0.0176673736423254, + 0.06221066787838936, + 0.06034010648727417, + 0.025440115481615067, + 0.1247384175658226, + 0.016700711101293564 + ], + "credit_map": [ + 0.030171474441885948, + 0.0297938734292984, + 0.06604541838169098, + 0.056722573935985565, + 0.032904114574193954, + 0.045037850737571716, + 0.03525690361857414, + 0.044510528445243835, + 0.027988160029053688, + 0.025011295452713966, + 0.016550567001104355, + 0.009778903797268867, + 0.00808695424348116, + 0.009516005404293537, + 0.008982216008007526, + 0.008276823908090591, + 0.022646993398666382, + 0.007069383282214403, + 0.006113559473305941, + 0.005895570386201143, + 0.004917082376778126, + 0.005772518925368786, + 0.019792761653661728, + 0.008008231408894062, + 0.007228270638734102, + 0.005365341901779175, + 0.008534196764230728, + 0.05725971609354019, + 0.0040617831982672215, + 0.007198677863925695, + 0.011461657471954823, + 0.05694316700100899, + 0.0016401403117924929, + 0.0018613868160173297, + 0.0024534750264137983, + 0.0024378506932407618, + 0.0023549089673906565, + 0.0023110422771424055, + 0.001992348115891218, + 0.0026162241119891405, + 0.0029434049502015114, + 0.0044045718386769295, + 0.005234997719526291, + 0.006974011659622192, + 0.004733057226985693, + 0.014122500084340572, + 0.003679742803797126, + 0.020118385553359985, + 0.0038075789343565702, + 0.007405790966004133, + 0.013736540451645851, + 0.012197596952319145, + 0.005508998408913612, + 0.005401770584285259, + 0.005669616162776947, + 0.0066122193820774555, + 0.002102868864312768, + 0.0025713511276990175, + 0.003661403665319085, + 0.004112744703888893, + 0.0031867362558841705, + 0.00468378234654665, + 0.0028408027719706297, + 0.0022804271429777145, + 0.00955207273364067, + 0.012110142968595028, + 0.018807359039783478, + 0.040950652211904526, + 0.006658380851149559, + 0.018700284883379936, + 0.010511194355785847, + 0.00744834216311574, + 0.002534510102123022, + 0.002281016670167446, + 0.002314318437129259, + 0.002403527032583952, + 0.0016230501933023334, + 0.001566035789437592, + 0.002173597225919366, + 0.0018046576296910644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.008630831725895405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247865337878466 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005247865337878466, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247865337878466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24908259510993958, + 0.031130945309996605, + 0.24570782482624054, + 0.06651099771261215, + 0.030137622728943825, + 0.09139220416545868, + 0.05175286531448364, + 0.0517755001783371, + 0.1698395311832428, + 0.012669961899518967 + ], + "credit_map": [ + 0.021223796531558037, + 0.02033103257417679, + 0.04936232045292854, + 0.035469818860292435, + 0.018059656023979187, + 0.04792747646570206, + 0.019256865605711937, + 0.03745163977146149, + 0.005388836842030287, + 0.005287650041282177, + 0.003273828886449337, + 0.004015506245195866, + 0.0030204798094928265, + 0.002618225524201989, + 0.004159505479037762, + 0.003366911318153143, + 0.07098239660263062, + 0.02203267067670822, + 0.018421288579702377, + 0.013918555341660976, + 0.013762221671640873, + 0.01805010251700878, + 0.06522130221128464, + 0.023319290950894356, + 0.0036255232989788055, + 0.0032520031090825796, + 0.003864888334646821, + 0.021963486447930336, + 0.0023054052144289017, + 0.003738023340702057, + 0.004888173192739487, + 0.022873492911458015, + 0.0028075685258954763, + 0.0024187983945012093, + 0.0034861317835748196, + 0.00515242014080286, + 0.002730481093749404, + 0.0035150416661053896, + 0.003035160480067134, + 0.006992020644247532, + 0.005825073458254337, + 0.009162535890936852, + 0.008752623572945595, + 0.011232824064791203, + 0.006889690179377794, + 0.017284980043768883, + 0.004672135692089796, + 0.02757234498858452, + 0.003568342886865139, + 0.00587891973555088, + 0.012900660745799541, + 0.01124434545636177, + 0.0063347285613417625, + 0.003531835740432143, + 0.00337006663903594, + 0.0049239653162658215, + 0.004943573381751776, + 0.004581969231367111, + 0.0078094773925840855, + 0.010730286128818989, + 0.0047750999219715595, + 0.007879392243921757, + 0.0063324966467916965, + 0.0047232056967914104, + 0.013650072738528252, + 0.01378925982862711, + 0.024440370500087738, + 0.04828784242272377, + 0.009553632698953152, + 0.02248556725680828, + 0.02691502869129181, + 0.010717754252254963, + 0.0017430440057069063, + 0.0015547795919701457, + 0.0019311447395011783, + 0.0019702448043972254, + 0.0010901789646595716, + 0.0013014745200052857, + 0.00157067715190351, + 0.0015084182377904654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.005247865337878466 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, dynamic tracking shot begins high above the clouds, then pitches downward at a steep swooping angle, closely following a massive flying dragon. It first folds its vast, sky-covering wings inward, clinging to its massive frame. The movement is swift and decisive, like folding steel. The scales covering its body are no ordinary substance, but a cold, metallic, obsidian. In the glaring sunlight filtering through the gaps in the clouds, they refract shifting shades of deep blue and purple, each subtle shift in angle creating a dazzling display of light. Its serpentine neck stretches forward, its spiky head lowered, its golden eyes, like molten lava, fixed fixedly on an unknown target below. As its wings fold, it instantly transforms into a living meteor, plummeting with astonishing acceleration, tearing through the thick, swirling cumulonimbus clouds and leaving a long, white trail of vapor and vortexes in its wake. The whole picture is full of a sense of speed and unparalleled power. The air makes a sharp whistling sound due to its high-speed movement, creating an epic atmosphere that is grand, magnificent and highly oppressive.", + "seed": 6566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_12/index_541_seed_6566_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01919291107915342, + "vlm_logprob": -0.01919291107915342, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.01919291107915342, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0063891904428601265 + ], + "opposite_logprob_tokens": [ + -5.756389141082764 + ], + "logprob_reward": -0.0063891904428601265, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0063891904428601265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.357786625623703, + 0.06539178639650345, + 0.10792846232652664, + 0.08700145781040192, + 0.061605289578437805, + 0.05551744997501373, + 0.07987578958272934, + 0.053429584950208664, + 0.05361214280128479, + 0.07785134017467499 + ], + "credit_map": [ + 0.033488862216472626, + 0.07234928011894226, + 0.03571167588233948, + 0.03772842511534691, + 0.04086029902100563, + 0.05889831483364105, + 0.03709080070257187, + 0.04165900871157646, + 0.005989941768348217, + 0.008428173139691353, + 0.006015102844685316, + 0.01093185693025589, + 0.010130424052476883, + 0.009098131209611893, + 0.008343481458723545, + 0.0064546819776296616, + 0.012915745377540588, + 0.016556059941649437, + 0.010596024803817272, + 0.013934571295976639, + 0.014657406136393547, + 0.01973838545382023, + 0.010584394447505474, + 0.008945885114371777, + 0.00829707644879818, + 0.009064657613635063, + 0.010037042200565338, + 0.00982903316617012, + 0.012251414358615875, + 0.014088130556046963, + 0.01109109167009592, + 0.012343019247055054, + 0.008126131258904934, + 0.007535073906183243, + 0.007560986094176769, + 0.009154018945991993, + 0.008380623534321785, + 0.00814664363861084, + 0.0064639742486178875, + 0.006237845402210951, + 0.007147852331399918, + 0.00649550836533308, + 0.00646285992115736, + 0.0042948042973876, + 0.0054844138212502, + 0.010152104310691357, + 0.005949858576059341, + 0.00953005626797676, + 0.00748899020254612, + 0.00817568227648735, + 0.012731785885989666, + 0.011030414141714573, + 0.010230612941086292, + 0.013274005614221096, + 0.008131035603582859, + 0.008813273161649704, + 0.00810234248638153, + 0.008993672206997871, + 0.004741317592561245, + 0.004238110035657883, + 0.010130194947123528, + 0.00550849037244916, + 0.005533056799322367, + 0.006182405166327953, + 0.006668605376034975, + 0.005571028683334589, + 0.0056222970597445965, + 0.009412379935383797, + 0.008318907581269741, + 0.0057736835442483425, + 0.005224952008575201, + 0.007020291406661272, + 0.00954537931829691, + 0.00797460786998272, + 0.01231665350496769, + 0.010905829258263111, + 0.011521709151566029, + 0.011670834384858608, + 0.007917826063930988, + 0.005998505745083094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0063891904428601265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866136610507965 + ], + "opposite_logprob_tokens": [ + -3.048661470413208 + ], + "logprob_reward": -0.04866136610507965, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866136610507965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1315133273601532, + 0.06171992048621178, + 0.10796406865119934, + 0.05613061413168907, + 0.08950039744377136, + 0.034876175224781036, + 0.13725709915161133, + 0.17934180796146393, + 0.10679173469543457, + 0.09490486234426498 + ], + "credit_map": [ + 0.015370327979326248, + 0.013103188015520573, + 0.01724029704928398, + 0.013969209976494312, + 0.017179271206259727, + 0.01961529813706875, + 0.01826620288193226, + 0.0167695302516222, + 0.005131293088197708, + 0.007629149593412876, + 0.007446644827723503, + 0.00891775730997324, + 0.008356552571058273, + 0.007695854175835848, + 0.007566915359348059, + 0.008975755423307419, + 0.013749876990914345, + 0.017766756936907768, + 0.01086574699729681, + 0.017005186527967453, + 0.012993774376809597, + 0.014913089573383331, + 0.009946499019861221, + 0.010723146609961987, + 0.005687730852514505, + 0.005788045935332775, + 0.009831403382122517, + 0.005177880637347698, + 0.008022035472095013, + 0.006388507783412933, + 0.007443233393132687, + 0.007791782729327679, + 0.00879861880093813, + 0.007277773227542639, + 0.011317411437630653, + 0.013182491064071655, + 0.018975261598825455, + 0.010340009815990925, + 0.009435160085558891, + 0.010173682123422623, + 0.00535886874422431, + 0.0034307301975786686, + 0.005146287381649017, + 0.004015752580016851, + 0.0027353710029274225, + 0.006210729479789734, + 0.003243420273065567, + 0.004735021386295557, + 0.009459939785301685, + 0.015102437697350979, + 0.016928505152463913, + 0.015350842848420143, + 0.02692580595612526, + 0.018720638006925583, + 0.014828769490122795, + 0.01994018442928791, + 0.014647913165390491, + 0.013141602277755737, + 0.0388442799448967, + 0.025593062862753868, + 0.030315348878502846, + 0.010669717565178871, + 0.023027662187814713, + 0.023102236911654472, + 0.012014630250632763, + 0.01321965828537941, + 0.015260267071425915, + 0.018266480416059494, + 0.011567448265850544, + 0.017496153712272644, + 0.007702076807618141, + 0.011265029199421406, + 0.013823135755956173, + 0.007167164236307144, + 0.019614892080426216, + 0.01150019932538271, + 0.008355136029422283, + 0.01452995277941227, + 0.007471427787095308, + 0.012442958541214466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.04866136610507965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979658544063568 + ], + "opposite_logprob_tokens": [ + -3.529796600341797 + ], + "logprob_reward": -0.02979658544063568, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979658544063568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12027869373559952, + 0.06641403585672379, + 0.0841001570224762, + 0.07846128940582275, + 0.0618387907743454, + 0.09195869415998459, + 0.11324622482061386, + 0.16253064572811127, + 0.10366132110357285, + 0.11751009523868561 + ], + "credit_map": [ + 0.012153764255344868, + 0.013345502316951752, + 0.014784027822315693, + 0.015074635855853558, + 0.01575247198343277, + 0.016544926911592484, + 0.017786415293812752, + 0.01483695488423109, + 0.005446956027299166, + 0.007876160554587841, + 0.0044914293102920055, + 0.009676223620772362, + 0.009271969087421894, + 0.00862693041563034, + 0.007485975977033377, + 0.013538396917283535, + 0.012317891232669353, + 0.014233043417334557, + 0.009209779091179371, + 0.00842155609279871, + 0.006918196566402912, + 0.013894097879529, + 0.006534244865179062, + 0.01257135346531868, + 0.006084587890654802, + 0.009576416574418545, + 0.010162864811718464, + 0.009724155068397522, + 0.009978828951716423, + 0.013078355230391026, + 0.008545037358999252, + 0.011311052367091179, + 0.00708949938416481, + 0.00699987169355154, + 0.00961625948548317, + 0.00776890991255641, + 0.011116704903542995, + 0.007409968879073858, + 0.005246818996965885, + 0.006590764503926039, + 0.013406551443040371, + 0.008005895651876926, + 0.01176653616130352, + 0.009383728727698326, + 0.008056389167904854, + 0.017537318170070648, + 0.008120170794427395, + 0.01568211428821087, + 0.009456939063966274, + 0.01204745564609766, + 0.01211595255881548, + 0.015348540619015694, + 0.019243601709604263, + 0.019150851294398308, + 0.01156043354421854, + 0.014322460629045963, + 0.013838673941791058, + 0.010001433081924915, + 0.054713599383831024, + 0.035133007913827896, + 0.0044542280957102776, + 0.012464880011975765, + 0.008165393956005573, + 0.02375945635139942, + 0.010669108480215073, + 0.01075899787247181, + 0.013972143642604351, + 0.01831183396279812, + 0.013658284209668636, + 0.013321644626557827, + 0.009776734746992588, + 0.01319259125739336, + 0.018671922385692596, + 0.009948871098458767, + 0.023399075493216515, + 0.012327607721090317, + 0.013196194544434547, + 0.01516382209956646, + 0.011772164143621922, + 0.013030444271862507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.02979658544063568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024850459303706884 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.0024850459303706884, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024850459303706884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1278989464044571, + 0.08292214572429657, + 0.19458173215389252, + 0.04282962158322334, + 0.0800209566950798, + 0.050370022654533386, + 0.07413294166326523, + 0.12501701712608337, + 0.13294586539268494, + 0.08928070217370987 + ], + "credit_map": [ + 0.011815005913376808, + 0.025934360921382904, + 0.01830202341079712, + 0.014461282640695572, + 0.015149792656302452, + 0.012725780718028545, + 0.01559216808527708, + 0.013918557204306126, + 0.00847737304866314, + 0.008590249344706535, + 0.009560367092490196, + 0.010448760353028774, + 0.011176330037415028, + 0.009722601622343063, + 0.012470336630940437, + 0.012476139701902866, + 0.029628116637468338, + 0.02650696039199829, + 0.013309021480381489, + 0.027343688532710075, + 0.025896010920405388, + 0.037465982139110565, + 0.018059082329273224, + 0.016372887417674065, + 0.003085190197452903, + 0.004950585309416056, + 0.00794108398258686, + 0.004288565367460251, + 0.005555570591241121, + 0.005886358674615622, + 0.00466763973236084, + 0.006454631220549345, + 0.007943246513605118, + 0.004838425666093826, + 0.012737495824694633, + 0.007817188277840614, + 0.012397431768476963, + 0.010588307864964008, + 0.006327027454972267, + 0.017371835187077522, + 0.00520691741257906, + 0.0041323499754071236, + 0.006244766991585493, + 0.005720941349864006, + 0.004620139487087727, + 0.00974308606237173, + 0.005052841268479824, + 0.009648981504142284, + 0.00584664661437273, + 0.005306861363351345, + 0.01013581920415163, + 0.00969018880277872, + 0.014104733243584633, + 0.012649237178266048, + 0.006517412606626749, + 0.009882044978439808, + 0.013302445411682129, + 0.008798815310001373, + 0.0363008938729763, + 0.024866502732038498, + 0.00730467215180397, + 0.011012091301381588, + 0.006674190983176231, + 0.0167574193328619, + 0.016435299068689346, + 0.018602639436721802, + 0.017906229943037033, + 0.022326543927192688, + 0.013923518359661102, + 0.01722024381160736, + 0.009953605011105537, + 0.016577791422605515, + 0.010245622135698795, + 0.008774511516094208, + 0.016334276646375656, + 0.01116959098726511, + 0.007501000538468361, + 0.018606793135404587, + 0.006686687469482422, + 0.009962229989469051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.0024850459303706884 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632367476820946 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.008632367476820946, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632367476820946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.163370743393898, + 0.05313558876514435, + 0.07850155234336853, + 0.08671025186777115, + 0.054770372807979584, + 0.13310988247394562, + 0.04965861141681671, + 0.10862696915864944, + 0.10003741830587387, + 0.17207863926887512 + ], + "credit_map": [ + 0.019200686365365982, + 0.016766760498285294, + 0.017021579667925835, + 0.011152679100632668, + 0.023474399000406265, + 0.027040092274546623, + 0.024868374690413475, + 0.023846182972192764, + 0.004305034410208464, + 0.005960592068731785, + 0.0031322245486080647, + 0.007655456196516752, + 0.006239774636924267, + 0.009742639027535915, + 0.006941131316125393, + 0.009158737026154995, + 0.009838176891207695, + 0.012425597757101059, + 0.00674138218164444, + 0.013011211529374123, + 0.007300574332475662, + 0.011723832227289677, + 0.008491422981023788, + 0.008969352580606937, + 0.0076288883574306965, + 0.012836998328566551, + 0.01713218167424202, + 0.009380999021232128, + 0.009583685547113419, + 0.007053222041577101, + 0.010308823548257351, + 0.012785454280674458, + 0.004890216048806906, + 0.004268168471753597, + 0.007572848815470934, + 0.00493940943852067, + 0.010479804128408432, + 0.007302275393158197, + 0.004748465958982706, + 0.010569185949862003, + 0.02448817901313305, + 0.014207224361598492, + 0.017665810883045197, + 0.010888847522437572, + 0.010738643817603588, + 0.024687493219971657, + 0.01218519825488329, + 0.018248485401272774, + 0.0039239306934177876, + 0.004670335911214352, + 0.0064838663674890995, + 0.005790296941995621, + 0.009708420373499393, + 0.005330855026841164, + 0.004971818532794714, + 0.008779089897871017, + 0.007705841679126024, + 0.008613849058747292, + 0.03411925584077835, + 0.02269039675593376, + 0.006577282678335905, + 0.007339502684772015, + 0.005171238444745541, + 0.016409598290920258, + 0.0068633356131613255, + 0.010501894168555737, + 0.012945486232638359, + 0.017182357609272003, + 0.016602683812379837, + 0.014196410775184631, + 0.010391461662948132, + 0.011353794485330582, + 0.034891240298748016, + 0.013689187355339527, + 0.031114494428038597, + 0.016659896820783615, + 0.024619784206151962, + 0.02231416292488575, + 0.013441009446978569, + 0.015348867513239384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.008632367476820946 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, dynamic tracking shot begins high above the clouds, then pitches downward at a steep swooping angle, closely following a massive flying dragon. It first folds its vast, sky-covering wings inward, clinging to its massive frame. The movement is swift and decisive, like folding steel. The scales covering its body are no ordinary substance, but a cold, metallic, obsidian. In the glaring sunlight filtering through the gaps in the clouds, they refract shifting shades of deep blue and purple, each subtle shift in angle creating a dazzling display of light. Its serpentine neck stretches forward, its spiky head lowered, its golden eyes, like molten lava, fixed fixedly on an unknown target below. As its wings fold, it instantly transforms into a living meteor, plummeting with astonishing acceleration, tearing through the thick, swirling cumulonimbus clouds and leaving a long, white trail of vapor and vortexes in its wake. The whole picture is full of a sense of speed and unparalleled power. The air makes a sharp whistling sound due to its high-speed movement, creating an epic atmosphere that is grand, magnificent and highly oppressive.", + "seed": 6567, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_12/index_541_seed_6567_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014991084951907397, + "vlm_logprob": -0.014991084951907397, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.014991084951907397, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004532302729785442 + ], + "opposite_logprob_tokens": [ + -6.629532337188721 + ], + "logprob_reward": -0.004532302729785442, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.004532302729785442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11591272056102753, + 0.09436694532632828, + 0.08381763100624084, + 0.12816892564296722, + 0.13634078204631805, + 0.11086790263652802, + 0.07545194774866104, + 0.10998363047838211, + 0.07756057381629944, + 0.06752895563840866 + ], + "credit_map": [ + 0.011621180921792984, + 0.015586151741445065, + 0.010661601088941097, + 0.017598632723093033, + 0.010996864177286625, + 0.013937331736087799, + 0.020562008023262024, + 0.014948958531022072, + 0.007928494364023209, + 0.010403390042483807, + 0.0130905257537961, + 0.017220966517925262, + 0.00745892059057951, + 0.012455622665584087, + 0.00891054980456829, + 0.01689848117530346, + 0.007619204465299845, + 0.009128355421125889, + 0.01454166043549776, + 0.011480209417641163, + 0.005484116263687611, + 0.011469763703644276, + 0.012393313460052013, + 0.011701008304953575, + 0.01591675728559494, + 0.030322514474391937, + 0.011374393478035927, + 0.012269995175302029, + 0.0107418829575181, + 0.019334308803081512, + 0.011539947241544724, + 0.01666911132633686, + 0.011596007272601128, + 0.011396556161344051, + 0.020382985472679138, + 0.010915623046457767, + 0.02771882154047489, + 0.018072620034217834, + 0.016982439905405045, + 0.01927572675049305, + 0.008434426039457321, + 0.017812591046094894, + 0.0156178530305624, + 0.015295598655939102, + 0.012944918125867844, + 0.012906973250210285, + 0.013908063061535358, + 0.013947474770247936, + 0.005912217777222395, + 0.006433138158172369, + 0.01266601961106062, + 0.018953721970319748, + 0.007923725061118603, + 0.007319414522498846, + 0.00619554053992033, + 0.010048167780041695, + 0.010767051950097084, + 0.01994745433330536, + 0.01244423259049654, + 0.018235618248581886, + 0.011969763785600662, + 0.015412673354148865, + 0.010679991915822029, + 0.010526837781071663, + 0.00812713336199522, + 0.00935232825577259, + 0.010083149187266827, + 0.010676036588847637, + 0.007286939304322004, + 0.01301241759210825, + 0.006924666929990053, + 0.012097900733351707, + 0.008234655484557152, + 0.006586137693375349, + 0.010777803137898445, + 0.005637523252516985, + 0.009969343431293964, + 0.00843792874366045, + 0.009913546033203602, + 0.007972023449838161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.004532302729785442 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810429945588112 + ], + "opposite_logprob_tokens": [ + -3.288104295730591 + ], + "logprob_reward": -0.03810429945588112, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810429945588112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06815115362405777, + 0.05683757737278938, + 0.12518982589244843, + 0.214097797870636, + 0.04106897488236427, + 0.13019777834415436, + 0.06194726377725601, + 0.03571420535445213, + 0.23149190843105316, + 0.03530348464846611 + ], + "credit_map": [ + 0.006585461553186178, + 0.0070366752333939075, + 0.004502618219703436, + 0.013025259599089622, + 0.008424715138971806, + 0.007414513733237982, + 0.011918414384126663, + 0.009243506006896496, + 0.0050015258602797985, + 0.006933016702532768, + 0.008288105949759483, + 0.008270321413874626, + 0.004246510565280914, + 0.008384926244616508, + 0.007000243756920099, + 0.008712932467460632, + 0.010043821297585964, + 0.01263892836868763, + 0.018042244017124176, + 0.018780121579766273, + 0.01757660135626793, + 0.014161442406475544, + 0.017135387286543846, + 0.016811290755867958, + 0.02570178546011448, + 0.06223544478416443, + 0.015284269116818905, + 0.02265777997672558, + 0.01659267023205757, + 0.028045015409588814, + 0.015041649341583252, + 0.02853918820619583, + 0.004863828420639038, + 0.004924527835100889, + 0.005817383527755737, + 0.003704481292515993, + 0.0051285503432154655, + 0.0059442720375955105, + 0.005312661174684763, + 0.0053732749074697495, + 0.014827095903456211, + 0.03298500180244446, + 0.013881426304578781, + 0.013590620830655098, + 0.018893137574195862, + 0.012159397825598717, + 0.015539695508778095, + 0.008321409113705158, + 0.00555255264043808, + 0.004916592966765165, + 0.010087973438203335, + 0.011413942091166973, + 0.008309151977300644, + 0.010519701056182384, + 0.004907050170004368, + 0.006240301765501499, + 0.003083054209128022, + 0.006304966751486063, + 0.00420188857242465, + 0.005531465169042349, + 0.0038836083840578794, + 0.004316606558859348, + 0.0038560465909540653, + 0.004536574240773916, + 0.01825203001499176, + 0.024876881390810013, + 0.021030867472290993, + 0.03337780386209488, + 0.018640704452991486, + 0.03164690360426903, + 0.04744343459606171, + 0.03622330352663994, + 0.003578459145501256, + 0.0026321234181523323, + 0.003755454672500491, + 0.003280648263171315, + 0.005343203898519278, + 0.004146188031882048, + 0.007435409817844629, + 0.005131999962031841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.03810429945588112 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066022217273712 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066022217273712, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066022217273712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13138261437416077, + 0.06432927399873734, + 0.0804542824625969, + 0.06912021338939667, + 0.08779970556497574, + 0.17060765624046326, + 0.11304095387458801, + 0.07785995304584503, + 0.11813333630561829, + 0.08727199584245682 + ], + "credit_map": [ + 0.014905523508787155, + 0.019404711201786995, + 0.011516534723341465, + 0.028150059282779694, + 0.013605833053588867, + 0.012172093614935875, + 0.01847049593925476, + 0.01315736398100853, + 0.004386625252664089, + 0.006432350259274244, + 0.010597044602036476, + 0.01033211499452591, + 0.008616069331765175, + 0.008953849785029888, + 0.006756739690899849, + 0.00825448241084814, + 0.0061689759604632854, + 0.0075819059275090694, + 0.014208909124135971, + 0.010022258386015892, + 0.008588477037847042, + 0.010564972646534443, + 0.012247792445123196, + 0.01107099186629057, + 0.0073995389975607395, + 0.01337348110973835, + 0.008044320158660412, + 0.008151255548000336, + 0.005926406476646662, + 0.011742757633328438, + 0.005585984326899052, + 0.00889646541327238, + 0.007945793680846691, + 0.00919369887560606, + 0.015225501731038094, + 0.006305404007434845, + 0.01798330433666706, + 0.011167770251631737, + 0.007795293349772692, + 0.012182945385575294, + 0.024108650162816048, + 0.036599401384592056, + 0.021966449916362762, + 0.01574091799557209, + 0.020928213372826576, + 0.015802618116140366, + 0.01899624429643154, + 0.016465162858366966, + 0.009356209076941013, + 0.0113489655777812, + 0.01959727145731449, + 0.02117893099784851, + 0.013570647686719894, + 0.01679409295320511, + 0.010104579851031303, + 0.011090252548456192, + 0.0063925208523869514, + 0.01514823455363512, + 0.013046251609921455, + 0.01226865779608488, + 0.008098753169178963, + 0.008483643643558025, + 0.008022604510188103, + 0.006399280391633511, + 0.009309777989983559, + 0.012998837046325207, + 0.011231409385800362, + 0.01696089841425419, + 0.01189323328435421, + 0.02089417353272438, + 0.015414852648973465, + 0.01943015120923519, + 0.007767234928905964, + 0.010468272492289543, + 0.01103190891444683, + 0.008405662141740322, + 0.00942792184650898, + 0.009867094457149506, + 0.017402509227395058, + 0.012901387177407742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.016066022217273712 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008630949072539806 + ], + "opposite_logprob_tokens": [ + -4.758630752563477 + ], + "logprob_reward": -0.008630949072539806, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008630949072539806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05699780583381653, + 0.05061240866780281, + 0.08040007948875427, + 0.05861881375312805, + 0.03945382684469223, + 0.022516299039125443, + 0.024341387674212456, + 0.03292662277817726, + 0.6120923757553101, + 0.02204042673110962 + ], + "credit_map": [ + 0.005844131112098694, + 0.006840767804533243, + 0.005355184432119131, + 0.00692738825455308, + 0.0070747146382927895, + 0.006731383968144655, + 0.009664294309914112, + 0.008559942245483398, + 0.0036550434306263924, + 0.0049638645723462105, + 0.005309330299496651, + 0.00584351597353816, + 0.004767960403114557, + 0.01110121700912714, + 0.008843036368489265, + 0.006128440611064434, + 0.0052985153160989285, + 0.009770520962774754, + 0.009801368229091167, + 0.013211547397077084, + 0.006890663877129555, + 0.013223182410001755, + 0.010587918572127819, + 0.011616360396146774, + 0.007053912151604891, + 0.01667299121618271, + 0.004084649961441755, + 0.006070148199796677, + 0.0046429927460849285, + 0.00839450303465128, + 0.003941275645047426, + 0.007758337538689375, + 0.0032959217205643654, + 0.004201982636004686, + 0.004635849501937628, + 0.002901202766224742, + 0.00986893754452467, + 0.004292752593755722, + 0.003948476631194353, + 0.006308702286332846, + 0.0018300124211236835, + 0.004619433544576168, + 0.0027210433036088943, + 0.0025301026180386543, + 0.0037683474365621805, + 0.0024503597524017096, + 0.002865455811843276, + 0.0017315420554950833, + 0.0024407056625932455, + 0.002265330869704485, + 0.0036537321284413338, + 0.003705864306539297, + 0.002510416554287076, + 0.004100024234503508, + 0.002587995259091258, + 0.0030773167964071035, + 0.0031256447546184063, + 0.004499695263803005, + 0.0039133718237280846, + 0.005228420253843069, + 0.004409727640450001, + 0.004109642002731562, + 0.0036020928528159857, + 0.004038031212985516, + 0.0441863052546978, + 0.07818812876939774, + 0.05774565041065216, + 0.07825209945440292, + 0.04918147251009941, + 0.07898811250925064, + 0.1408163458108902, + 0.08473428338766098, + 0.0021812347695231438, + 0.0019342508167028427, + 0.0033871964551508427, + 0.0018488761270418763, + 0.003444448346272111, + 0.002275806851685047, + 0.004288833122700453, + 0.0026797791942954063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.008630949072539806 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621851284056902 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.007621851284056902, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621851284056902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2936824560165405, + 0.06712564080953598, + 0.061135612428188324, + 0.08936351537704468, + 0.10262542963027954, + 0.09967179596424103, + 0.038101524114608765, + 0.07121919095516205, + 0.10260090231895447, + 0.07447392493486404 + ], + "credit_map": [ + 0.026335621252655983, + 0.02571372129023075, + 0.02287520281970501, + 0.06952520459890366, + 0.018943410366773605, + 0.025952108204364777, + 0.06537367403507233, + 0.03896350786089897, + 0.005733714904636145, + 0.009728217497467995, + 0.008077994920313358, + 0.009832820855081081, + 0.0057813203893601894, + 0.012226592749357224, + 0.007223202846944332, + 0.008521775715053082, + 0.004459370858967304, + 0.006345013156533241, + 0.011070759035646915, + 0.008082280866801739, + 0.004413201939314604, + 0.011019272729754448, + 0.007273060269653797, + 0.008472653105854988, + 0.008481143042445183, + 0.015521836467087269, + 0.011123545467853546, + 0.011404601857066154, + 0.008836419321596622, + 0.012733791023492813, + 0.008789779618382454, + 0.012472395785152912, + 0.010475627146661282, + 0.011722361668944359, + 0.011843943037092686, + 0.006093995179980993, + 0.02604103460907936, + 0.011102838441729546, + 0.009126023389399052, + 0.0162196047604084, + 0.007965541444718838, + 0.015758974477648735, + 0.013748600147664547, + 0.01302255131304264, + 0.010595147497951984, + 0.013430858962237835, + 0.011236125603318214, + 0.013913997448980808, + 0.0033289939165115356, + 0.004204987082630396, + 0.005283818580210209, + 0.006670390721410513, + 0.003955751191824675, + 0.004572483245283365, + 0.004016534890979528, + 0.006068564485758543, + 0.007546437904238701, + 0.01174529455602169, + 0.008434619754552841, + 0.012020791880786419, + 0.007567181717604399, + 0.009197790175676346, + 0.007061992306262255, + 0.007645075209438801, + 0.007035956252366304, + 0.01049306895583868, + 0.009277937933802605, + 0.024396147578954697, + 0.008595861494541168, + 0.015015553683042526, + 0.008260613307356834, + 0.019525760784745216, + 0.008794146589934826, + 0.006559495348483324, + 0.008518695831298828, + 0.006210027728229761, + 0.010493661276996136, + 0.009261264465749264, + 0.015336026437580585, + 0.009300610050559044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.007621851284056902 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, dynamic tracking shot begins high above the clouds, then pitches downward at a steep swooping angle, closely following a massive flying dragon. It first folds its vast, sky-covering wings inward, clinging to its massive frame. The movement is swift and decisive, like folding steel. The scales covering its body are no ordinary substance, but a cold, metallic, obsidian. In the glaring sunlight filtering through the gaps in the clouds, they refract shifting shades of deep blue and purple, each subtle shift in angle creating a dazzling display of light. Its serpentine neck stretches forward, its spiky head lowered, its golden eyes, like molten lava, fixed fixedly on an unknown target below. As its wings fold, it instantly transforms into a living meteor, plummeting with astonishing acceleration, tearing through the thick, swirling cumulonimbus clouds and leaving a long, white trail of vapor and vortexes in its wake. The whole picture is full of a sense of speed and unparalleled power. The air makes a sharp whistling sound due to its high-speed movement, creating an epic atmosphere that is grand, magnificent and highly oppressive.", + "seed": 6568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_12/index_541_seed_6568_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015550414612516762, + "vlm_logprob": -0.015550414612516762, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.015550414612516762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0035321249160915613 + ], + "opposite_logprob_tokens": [ + -6.8785319328308105 + ], + "logprob_reward": -0.0035321249160915613, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0035321249160915613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16878515481948853, + 0.08020125329494476, + 0.04392099007964134, + 0.06971879303455353, + 0.03705360367894173, + 0.042917486280202866, + 0.14223383367061615, + 0.10848294943571091, + 0.10198073834180832, + 0.2047051191329956 + ], + "credit_map": [ + 0.028903810307383537, + 0.021730024367570877, + 0.018414931371808052, + 0.013916507363319397, + 0.013977017253637314, + 0.019041942432522774, + 0.012981812469661236, + 0.03981912136077881, + 0.0075299604795873165, + 0.009329384192824364, + 0.009614176116883755, + 0.011073913425207138, + 0.00825508777052164, + 0.009798924438655376, + 0.01612800359725952, + 0.008471815846860409, + 0.004191071260720491, + 0.004117106553167105, + 0.0056808409281075, + 0.00585913797840476, + 0.005772420205175877, + 0.003935776185244322, + 0.006636377424001694, + 0.007728261407464743, + 0.00787613820284605, + 0.010417819954454899, + 0.010049469769001007, + 0.008064392022788525, + 0.011303006671369076, + 0.008523638360202312, + 0.005345787387341261, + 0.008138551376760006, + 0.004651500377804041, + 0.003087555756792426, + 0.00654437905177474, + 0.004082450643181801, + 0.003563407575711608, + 0.004121225792914629, + 0.00499865273013711, + 0.006004437804222107, + 0.0050824591889977455, + 0.005306963808834553, + 0.004907613154500723, + 0.0063197254203259945, + 0.005908893886953592, + 0.004946906119585037, + 0.005413443315774202, + 0.005031483247876167, + 0.015122243203222752, + 0.015687309205532074, + 0.01174220722168684, + 0.01608790084719658, + 0.016111286357045174, + 0.023182762786746025, + 0.015263819135725498, + 0.029036318883299828, + 0.010146849788725376, + 0.010657095350325108, + 0.00926926638931036, + 0.014373665675520897, + 0.010540184564888477, + 0.021764222532510757, + 0.007863348349928856, + 0.023868326097726822, + 0.01766185835003853, + 0.011152129620313644, + 0.015476061031222343, + 0.01986015774309635, + 0.010015962645411491, + 0.008827176876366138, + 0.00860810186713934, + 0.010379303246736526, + 0.019616184756159782, + 0.031022656708955765, + 0.025545375421643257, + 0.02500058524310589, + 0.023342233151197433, + 0.036250192672014236, + 0.023362930864095688, + 0.020564982667565346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0035321249160915613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305974766612053 + ], + "opposite_logprob_tokens": [ + -3.168059825897217 + ], + "logprob_reward": -0.04305974766612053, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305974766612053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22759613394737244, + 0.04838751256465912, + 0.038616277277469635, + 0.3026721775531769, + 0.03252197429537773, + 0.021457236260175705, + 0.10460249334573746, + 0.07316254824399948, + 0.06351467221975327, + 0.08746898174285889 + ], + "credit_map": [ + 0.07114949822425842, + 0.013525942340493202, + 0.020336518064141273, + 0.009254541248083115, + 0.009013492614030838, + 0.018139811232686043, + 0.0180001612752676, + 0.06817617267370224, + 0.0042152851819992065, + 0.007174645084887743, + 0.00707961106672883, + 0.005899661686271429, + 0.005177417770028114, + 0.0051221661269664764, + 0.006414741277694702, + 0.007303981110453606, + 0.0047433082945644855, + 0.0057637691497802734, + 0.003539647441357374, + 0.0050215283408761024, + 0.004643683787435293, + 0.0035559386014938354, + 0.004993844777345657, + 0.006354555021971464, + 0.03665538877248764, + 0.03636861592531204, + 0.07046739757061005, + 0.025255829095840454, + 0.0376991406083107, + 0.03508630394935608, + 0.023322781547904015, + 0.037816714495420456, + 0.0035095487255603075, + 0.002625065855681896, + 0.005965746007859707, + 0.00342402420938015, + 0.004413169343024492, + 0.003821222810074687, + 0.003224498126655817, + 0.005538698751479387, + 0.0026555003132671118, + 0.0021400107070803642, + 0.0015510666416957974, + 0.0021238389890640974, + 0.0036347433924674988, + 0.0030330722220242023, + 0.002413359936326742, + 0.0039056451059877872, + 0.007306529209017754, + 0.007334384601563215, + 0.009942985139787197, + 0.012084445916116238, + 0.0087057463824749, + 0.0217759869992733, + 0.010465086437761784, + 0.026987334713339806, + 0.007843655534088612, + 0.010415533557534218, + 0.010705387219786644, + 0.009321042336523533, + 0.009445920586585999, + 0.011162570677697659, + 0.0067558265291154385, + 0.007512605749070644, + 0.008087539114058018, + 0.004920731298625469, + 0.013255053199827671, + 0.017241552472114563, + 0.004932859446853399, + 0.005224702414125204, + 0.004968294873833656, + 0.004883941728621721, + 0.009026481769979, + 0.011763202957808971, + 0.010900823399424553, + 0.007694982457906008, + 0.010452693328261375, + 0.012861896306276321, + 0.01170819066464901, + 0.013060711324214935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.04305974766612053 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281434550881386 + ], + "opposite_logprob_tokens": [ + -3.7732813358306885 + ], + "logprob_reward": -0.023281434550881386, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281434550881386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16888804733753204, + 0.0777958482503891, + 0.05337768793106079, + 0.17386168241500854, + 0.06697432696819305, + 0.07166039198637009, + 0.030880387872457504, + 0.22413915395736694, + 0.0558643639087677, + 0.07655814290046692 + ], + "credit_map": [ + 0.05219504237174988, + 0.013168261386454105, + 0.018296033143997192, + 0.009190930053591728, + 0.013131881132721901, + 0.011115934699773788, + 0.015113212168216705, + 0.036676742136478424, + 0.008607311174273491, + 0.008656413294374943, + 0.010943496599793434, + 0.011988293379545212, + 0.008534553460776806, + 0.007248012814670801, + 0.009994623251259327, + 0.011823143810033798, + 0.0068087768740952015, + 0.007069949060678482, + 0.005556195043027401, + 0.0072758812457323074, + 0.007989338599145412, + 0.005513121839612722, + 0.006939517334103584, + 0.006224906537681818, + 0.018170740455389023, + 0.021453671157360077, + 0.030381010845303535, + 0.013798180036246777, + 0.018664749339222908, + 0.020425962284207344, + 0.014393734745681286, + 0.0365736223757267, + 0.007175894919782877, + 0.0074012246914207935, + 0.00978817418217659, + 0.0077285668812692165, + 0.009479627013206482, + 0.008294887840747833, + 0.005679570138454437, + 0.01142637524753809, + 0.007483118213713169, + 0.010732957161962986, + 0.006520554888993502, + 0.007719165179878473, + 0.01673470064997673, + 0.00837753340601921, + 0.006851390935480595, + 0.0072409650310873985, + 0.0033413926139473915, + 0.003369122976437211, + 0.0021701145451515913, + 0.002956670243293047, + 0.0032254324760288, + 0.005584927275776863, + 0.003362806746736169, + 0.00686991959810257, + 0.017189957201480865, + 0.015788620337843895, + 0.01585276424884796, + 0.03515728935599327, + 0.017304232344031334, + 0.05498035252094269, + 0.017742570489645004, + 0.05012336000800133, + 0.008852861821651459, + 0.004690401256084442, + 0.008172059431672096, + 0.014130487106740475, + 0.004092091228812933, + 0.005098703317344189, + 0.00643764715641737, + 0.004390110727399588, + 0.007778178434818983, + 0.005410651210695505, + 0.01098672952502966, + 0.005365435499697924, + 0.011239038780331612, + 0.009967364370822906, + 0.014649024233222008, + 0.011161711998283863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.023281434550881386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019369428046047688 + ], + "opposite_logprob_tokens": [ + -6.251936912536621 + ], + "logprob_reward": -0.0019369428046047688, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019369428046047688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11794141680002213, + 0.03770573064684868, + 0.05093779414892197, + 0.17734576761722565, + 0.22395020723342896, + 0.08416305482387543, + 0.13890963792800903, + 0.03696553036570549, + 0.04093663766980171, + 0.09114421904087067 + ], + "credit_map": [ + 0.0193156860768795, + 0.013646088540554047, + 0.014628907665610313, + 0.009538696147501469, + 0.013599138706922531, + 0.011800259351730347, + 0.014009316451847553, + 0.02140333317220211, + 0.0034190835431218147, + 0.004036827478557825, + 0.0053819091990590096, + 0.003735409351065755, + 0.003907314036041498, + 0.004783177748322487, + 0.006708483211696148, + 0.005733526777476072, + 0.005712892860174179, + 0.00614278856664896, + 0.005936238449066877, + 0.0049971635453403, + 0.006149837281554937, + 0.004207814112305641, + 0.0067766811698675156, + 0.011014381423592567, + 0.020349184051156044, + 0.022788062691688538, + 0.03045319765806198, + 0.018703967332839966, + 0.02072707563638687, + 0.01903168484568596, + 0.017674745991826057, + 0.02761784754693508, + 0.03181561827659607, + 0.01543017290532589, + 0.041324470192193985, + 0.018834326416254044, + 0.021349303424358368, + 0.025237862020730972, + 0.032504767179489136, + 0.03745367377996445, + 0.01243877224624157, + 0.009979726746678352, + 0.011157727800309658, + 0.01306209061294794, + 0.006846940610557795, + 0.01156865805387497, + 0.009149327874183655, + 0.009959809482097626, + 0.011177398264408112, + 0.009939797222614288, + 0.009150069206953049, + 0.014906699769198895, + 0.011324032209813595, + 0.03238152712583542, + 0.013728326186537743, + 0.036301787942647934, + 0.004450113046914339, + 0.0038705263286828995, + 0.004748527891933918, + 0.004724619444459677, + 0.005035798996686935, + 0.00555510725826025, + 0.0036657373420894146, + 0.004915096331387758, + 0.005223920103162527, + 0.0033304644748568535, + 0.005833948031067848, + 0.005228390917181969, + 0.00519802700728178, + 0.006144821178168058, + 0.0055702761746943, + 0.004406788386404514, + 0.00822894461452961, + 0.010151660069823265, + 0.014891047962009907, + 0.00942996609956026, + 0.010576329194009304, + 0.018754271790385246, + 0.008737693540751934, + 0.010374307632446289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.0019369428046047688 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive flying dragon with a serpentine neck and spiky head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales metallic and obsidian, refracting deep blue and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon fold its vast wings inward and transform into a living meteor while plummeting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sky with thick, swirling cumulonimbus clouds and glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin high above the clouds and pitch downward at a steep swooping angle to follow the dragon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005941823124885559 + ], + "opposite_logprob_tokens": [ + -5.130941867828369 + ], + "logprob_reward": -0.005941823124885559, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005941823124885559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20446538925170898, + 0.06704698503017426, + 0.07026653736829758, + 0.19761605560779572, + 0.06372792273759842, + 0.09160782396793365, + 0.07781320810317993, + 0.06394164264202118, + 0.052579086273908615, + 0.11093532294034958 + ], + "credit_map": [ + 0.06527724117040634, + 0.01613098382949829, + 0.01998373679816723, + 0.010645158588886261, + 0.010701730847358704, + 0.013943732716143131, + 0.013671081513166428, + 0.054111745208501816, + 0.006015261635184288, + 0.006703169085085392, + 0.009530025534331799, + 0.009609444066882133, + 0.006606453098356724, + 0.009578621946275234, + 0.009924732148647308, + 0.009079286828637123, + 0.008979047648608685, + 0.009056351147592068, + 0.00702953664585948, + 0.010066280141472816, + 0.008526545949280262, + 0.00712061021476984, + 0.010004659183323383, + 0.009483511559665203, + 0.023839710280299187, + 0.024084795266389847, + 0.053140778094530106, + 0.01839798502624035, + 0.020491980016231537, + 0.02219356782734394, + 0.015893075615167618, + 0.019574178382754326, + 0.008114905096590519, + 0.005320475436747074, + 0.010601956397294998, + 0.006409113761037588, + 0.008445247076451778, + 0.006940266117453575, + 0.00808938592672348, + 0.00980658270418644, + 0.00818447396159172, + 0.008587817661464214, + 0.005353947635740042, + 0.015359456650912762, + 0.008814724162220955, + 0.014485002495348454, + 0.016290443018078804, + 0.014531968161463737, + 0.010087631642818451, + 0.012212114408612251, + 0.0054065268486738205, + 0.01150799635797739, + 0.00875990092754364, + 0.010548828169703484, + 0.008001307025551796, + 0.011288920417428017, + 0.006061253137886524, + 0.005180079955607653, + 0.005964105017483234, + 0.008308472111821175, + 0.006866122595965862, + 0.013014689087867737, + 0.005713408812880516, + 0.012833517044782639, + 0.00905611552298069, + 0.005928537342697382, + 0.007998039945960045, + 0.00636536069214344, + 0.004918097518384457, + 0.006105659529566765, + 0.007501029409468174, + 0.004706250969320536, + 0.01172697450965643, + 0.011215035803616047, + 0.01111676823347807, + 0.012030607089400291, + 0.014925390481948853, + 0.01501645427197218, + 0.020551156252622604, + 0.014352953061461449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.005941823124885559 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_tensor([2916, 2916])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_tensor([2916, 2916])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..451d0899dfe207a9d1913c52dab5ccb6877ba790 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_tensor([2916, 2916])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up, synchronized with the circular motion of a carousel, focuses firmly on the back of a magnificent white painted wooden horse, where a young woman lies crouched. Her long, slightly tousled dark brown hair, a few strands clinging to her cheeks, a sleep-tinged flush. She wears a simple white cotton dress that seems incongruous with the scene. As the horse rises and falls steadily, as if awakened by the gentle jolt, her eyelids begin to tremble slightly, and then, with a touch of uncertainty, she raises her head. Her heavy eyelids slowly open, initially confused and empty, before adjusting for a moment, they gradually become occupied by the surrounding scene, reflecting a myriad of light. Her gaze is a world of flowing light, composed of thousands of warm incandescent bulbs and iridescent LED strips. These lights, as the carousel rises and spins, cast a shifting, dreamlike pattern of light on her pale skin and the hem of her pure white dress. Her lips are slightly parted, her expression shifting from initial confusion to an indescribable wonder. In the background, other wooden horses, lions, and ornate mirrored decorations merge into a brilliant halo due to the shallow depth of field and motion blur, creating a surreal atmosphere that is both disorienting and full of fantasy magic.", + "seed": 4399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_2916_seed_4399_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04247624948620796, + "vlm_logprob": -0.04247624948620796, + "vlm_yes_no_margin": 3.799999976158142, + "vlm_token_credit": -0.04247624948620796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015409108251333237 + ], + "opposite_logprob_tokens": [ + -5.015408992767334 + ], + "logprob_reward": -0.015409108251333237, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.015409108251333237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11943061649799347, + 0.04269589111208916, + 0.05321574956178665, + 0.13343608379364014, + 0.10224533081054688, + 0.04698532074689865, + 0.11593478918075562, + 0.047315701842308044, + 0.25961899757385254, + 0.07912153005599976 + ], + "credit_map": [ + 0.010213061235845089, + 0.010559841990470886, + 0.015312428586184978, + 0.018514426425099373, + 0.02053825743496418, + 0.01519515085965395, + 0.01233053207397461, + 0.01676691509783268, + 0.005831749178469181, + 0.003919720184057951, + 0.008105281740427017, + 0.006608275230973959, + 0.0030001820996403694, + 0.006668701767921448, + 0.004255854059010744, + 0.004306126851588488, + 0.00474336976185441, + 0.005604675970971584, + 0.00497099757194519, + 0.005536952055990696, + 0.011115290224552155, + 0.005513404030352831, + 0.007427540607750416, + 0.00830351933836937, + 0.012840282171964645, + 0.010060972534120083, + 0.009035306051373482, + 0.019947629421949387, + 0.011042064055800438, + 0.014277597889304161, + 0.02651260793209076, + 0.02971961721777916, + 0.011610960587859154, + 0.00807457696646452, + 0.00898901466280222, + 0.012524359859526157, + 0.010726510547101498, + 0.01498718187212944, + 0.015507280826568604, + 0.019825445488095284, + 0.0033263496588915586, + 0.0049621607176959515, + 0.004012200981378555, + 0.007214770652353764, + 0.005865630693733692, + 0.004125741310417652, + 0.00954565592110157, + 0.007932811975479126, + 0.009391752071678638, + 0.011773750185966492, + 0.01224864274263382, + 0.024873435497283936, + 0.01514542754739523, + 0.007758185733109713, + 0.017422249540686607, + 0.017321346327662468, + 0.0040938882157206535, + 0.0044762128964066505, + 0.004493880085647106, + 0.00779375247657299, + 0.0063233268447220325, + 0.006132082082331181, + 0.009306316263973713, + 0.004696245770901442, + 0.015118708834052086, + 0.011949971318244934, + 0.07186317443847656, + 0.02845228835940361, + 0.020819228142499924, + 0.008149654604494572, + 0.08855116367340088, + 0.014714816585183144, + 0.01006493903696537, + 0.011378088966012001, + 0.011130647733807564, + 0.013138441368937492, + 0.010324599221348763, + 0.007490973919630051, + 0.009917606599628925, + 0.005676234606653452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.015409108251333237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201468512415886 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201468512415886, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201468512415886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0832623615860939, + 0.02754201553761959, + 0.10252154618501663, + 0.023544058203697205, + 0.0487913154065609, + 0.08628669381141663, + 0.062199223786592484, + 0.037407949566841125, + 0.39061495661735535, + 0.13782989978790283 + ], + "credit_map": [ + 0.007778463419526815, + 0.009240654297173023, + 0.008171586319804192, + 0.014162402600049973, + 0.015226699411869049, + 0.007849260233342648, + 0.00954087357968092, + 0.011292424984276295, + 0.0026725446805357933, + 0.0024978602305054665, + 0.003025938756763935, + 0.003755462821573019, + 0.0031145529355853796, + 0.00488097220659256, + 0.003446707734838128, + 0.004147978033870459, + 0.00962651427835226, + 0.009626645594835281, + 0.009530017152428627, + 0.012867466546595097, + 0.016844989731907845, + 0.00980418547987938, + 0.016189461573958397, + 0.018032263964414597, + 0.0028586157131940126, + 0.0020357014145702124, + 0.0024702087976038456, + 0.003999676555395126, + 0.0024441636633127928, + 0.0028792847879230976, + 0.003572986228391528, + 0.0032834182493388653, + 0.0052202194929122925, + 0.004682841245085001, + 0.004732960369437933, + 0.005606873892247677, + 0.008980166167020798, + 0.008115549571812153, + 0.006705920677632093, + 0.0047467839904129505, + 0.007379783317446709, + 0.007672234904021025, + 0.011327498592436314, + 0.0137141989544034, + 0.013482419773936272, + 0.0060877311043441296, + 0.018054405227303505, + 0.008568421006202698, + 0.005220500286668539, + 0.006762639619410038, + 0.00750304339453578, + 0.011240703985095024, + 0.007521555293351412, + 0.005984737072139978, + 0.012285939417779446, + 0.005680107977241278, + 0.003096735803410411, + 0.0033721113577485085, + 0.005976377986371517, + 0.005175917875021696, + 0.004624934401363134, + 0.006303462665528059, + 0.005911210551857948, + 0.002947199856862426, + 0.022547932341694832, + 0.021172314882278442, + 0.1247079074382782, + 0.03797903656959534, + 0.028109125792980194, + 0.010698497295379639, + 0.12499314546585083, + 0.020407015457749367, + 0.010276921093463898, + 0.023283565416932106, + 0.014847829937934875, + 0.028615662828087807, + 0.020956773310899734, + 0.01101327408105135, + 0.020379668101668358, + 0.008456199429929256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.018201468512415886 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14289315044879913 + ], + "opposite_logprob_tokens": [ + -2.017893075942993 + ], + "logprob_reward": -0.14289315044879913, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14289315044879913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1353013515472412, + 0.019654909148812294, + 0.022410698235034943, + 0.026864295825362206, + 0.0643744096159935, + 0.07955421507358551, + 0.37883394956588745, + 0.06227976456284523, + 0.11053431034088135, + 0.10019205510616302 + ], + "credit_map": [ + 0.008838633075356483, + 0.012076732702553272, + 0.011050339788198471, + 0.025392424315214157, + 0.022763295099139214, + 0.017456799745559692, + 0.01758457161486149, + 0.0201385710388422, + 0.0020711752586066723, + 0.00216514035128057, + 0.0030036920215934515, + 0.00281758070923388, + 0.002395619172602892, + 0.0025730221532285213, + 0.0023888086434453726, + 0.002239871770143509, + 0.003046811558306217, + 0.002422997495159507, + 0.0031486989464610815, + 0.002990178996697068, + 0.002859621774405241, + 0.0024109170772135258, + 0.003102104179561138, + 0.0024293686728924513, + 0.0031213436741381884, + 0.0035395140293985605, + 0.002050460549071431, + 0.004161113873124123, + 0.003286075545474887, + 0.0021164235658943653, + 0.005177311133593321, + 0.003412053221836686, + 0.00353872193954885, + 0.004511740989983082, + 0.0038463438395410776, + 0.01708087883889675, + 0.009691769257187843, + 0.0049615115858614445, + 0.011604121886193752, + 0.009139320813119411, + 0.0059869857504963875, + 0.005385956261307001, + 0.010139419697225094, + 0.0083836754783988, + 0.009412853978574276, + 0.006454699672758579, + 0.017870474606752396, + 0.01592014729976654, + 0.023510290309786797, + 0.03268349915742874, + 0.025516968220472336, + 0.04773342236876488, + 0.08567404001951218, + 0.028795212507247925, + 0.039790138602256775, + 0.09513037651777267, + 0.0053937118500471115, + 0.005289989057928324, + 0.010529148392379284, + 0.011453384533524513, + 0.008714189752936363, + 0.007189203053712845, + 0.00928819365799427, + 0.004421949852257967, + 0.005218964535742998, + 0.009542027488350868, + 0.011904282495379448, + 0.0343913659453392, + 0.013095787726342678, + 0.005497168283909559, + 0.01974323019385338, + 0.011141483671963215, + 0.008732072077691555, + 0.0137266144156456, + 0.013753159902989864, + 0.018143953755497932, + 0.014643625356256962, + 0.008477685041725636, + 0.01532057672739029, + 0.007394373416900635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.14289315044879913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332615666091442 + ], + "opposite_logprob_tokens": [ + -3.7733261585235596 + ], + "logprob_reward": -0.02332615666091442, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332615666091442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07486570626497269, + 0.037066664546728134, + 0.05332407355308533, + 0.06949525326490402, + 0.12081387639045715, + 0.17851844429969788, + 0.07549287378787994, + 0.02974521741271019, + 0.3088608980178833, + 0.05181696638464928 + ], + "credit_map": [ + 0.008199168369174004, + 0.007267745677381754, + 0.008373880758881569, + 0.009416433982551098, + 0.013458577916026115, + 0.00808916985988617, + 0.011320041492581367, + 0.00874069333076477, + 0.003960075788199902, + 0.004527399782091379, + 0.0029729059897363186, + 0.005734351463615894, + 0.0037719549145549536, + 0.005999838933348656, + 0.004885013680905104, + 0.00521512608975172, + 0.0049163103103637695, + 0.006685126572847366, + 0.004816689994186163, + 0.007483005523681641, + 0.011721749790012836, + 0.004443757236003876, + 0.007873966358602047, + 0.005383466836065054, + 0.008219373412430286, + 0.005991687532514334, + 0.004513067658990622, + 0.01143677718937397, + 0.006100036203861237, + 0.0069272774271667, + 0.01296129822731018, + 0.013345742598176003, + 0.0069007668644189835, + 0.011246109381318092, + 0.010153649374842644, + 0.02213064767420292, + 0.024530023336410522, + 0.01016736775636673, + 0.02115398272871971, + 0.014531328342854977, + 0.012096327729523182, + 0.0035439408384263515, + 0.006543228402733803, + 0.009387177415192127, + 0.017696067690849304, + 0.007850098423659801, + 0.06311526894569397, + 0.05828633904457092, + 0.008077682927250862, + 0.011519481427967548, + 0.013395644724369049, + 0.007960989139974117, + 0.009763538837432861, + 0.00717127462849021, + 0.008062642998993397, + 0.009541616775095463, + 0.002333165379241109, + 0.004136783070862293, + 0.0026152024511247873, + 0.0041745929047465324, + 0.0030401493422687054, + 0.005473635159432888, + 0.004304802045226097, + 0.0036668875254690647, + 0.01773945987224579, + 0.021128736436367035, + 0.08825176954269409, + 0.03117239661514759, + 0.022853245958685875, + 0.013223960064351559, + 0.09584282338619232, + 0.018648510798811913, + 0.005507219582796097, + 0.0057764807716012, + 0.005142895504832268, + 0.006865374743938446, + 0.0100239347666502, + 0.004754127934575081, + 0.007736815605312586, + 0.006010120268911123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.02332615666091442 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551363557577133 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551363557577133, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551363557577133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22311349213123322, + 0.06402495503425598, + 0.04469161853194237, + 0.13656271994113922, + 0.1032215803861618, + 0.10759306699037552, + 0.08742839843034744, + 0.08182916045188904, + 0.0814645066857338, + 0.07007045298814774 + ], + "credit_map": [ + 0.014195248484611511, + 0.024612167850136757, + 0.020404523238539696, + 0.03331821784377098, + 0.0434807687997818, + 0.018667789176106453, + 0.04145854339003563, + 0.026976250112056732, + 0.00757015123963356, + 0.00541489664465189, + 0.00940506812185049, + 0.005927617195993662, + 0.005733281373977661, + 0.012797330506145954, + 0.008056532591581345, + 0.009120076894760132, + 0.0045349737629294395, + 0.005647709593176842, + 0.0055772121995687485, + 0.005765093490481377, + 0.006692873779684305, + 0.0036346542183309793, + 0.008209174498915672, + 0.004629929084330797, + 0.014926448464393616, + 0.00942874513566494, + 0.010158738121390343, + 0.02376522310078144, + 0.012488397769629955, + 0.011627286672592163, + 0.029834723100066185, + 0.024333158507943153, + 0.008210699073970318, + 0.010690981522202492, + 0.008712327107787132, + 0.014638425782322884, + 0.01899106614291668, + 0.013823006302118301, + 0.015467827208340168, + 0.012687241658568382, + 0.008966700173914433, + 0.007008449640125036, + 0.008296515792608261, + 0.009609473869204521, + 0.012098039500415325, + 0.00827584695070982, + 0.029051417484879494, + 0.024286622181534767, + 0.007555224001407623, + 0.011683735065162182, + 0.009906058199703693, + 0.010331862606108189, + 0.017137732356786728, + 0.00887003168463707, + 0.009595654904842377, + 0.012348095886409283, + 0.006324575748294592, + 0.009073982015252113, + 0.009579835459589958, + 0.011638264171779156, + 0.010332192294299603, + 0.012889770790934563, + 0.013750665821135044, + 0.00823987927287817, + 0.005146754905581474, + 0.007026636507362127, + 0.019960129633545876, + 0.01028089877218008, + 0.006467938423156738, + 0.0054781450890004635, + 0.018954280763864517, + 0.008149723522365093, + 0.007171886041760445, + 0.007333856076002121, + 0.016391579061746597, + 0.01344109047204256, + 0.007340490352362394, + 0.005636810790747404, + 0.007929803803563118, + 0.004824932664632797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.012551363557577133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, synchronized with the circular motion of a carousel, focuses firmly on the back of a magnificent white painted wooden horse, where a young woman lies crouched. Her long, slightly tousled dark brown hair, a few strands clinging to her cheeks, a sleep-tinged flush. She wears a simple white cotton dress that seems incongruous with the scene. As the horse rises and falls steadily, as if awakened by the gentle jolt, her eyelids begin to tremble slightly, and then, with a touch of uncertainty, she raises her head. Her heavy eyelids slowly open, initially confused and empty, before adjusting for a moment, they gradually become occupied by the surrounding scene, reflecting a myriad of light. Her gaze is a world of flowing light, composed of thousands of warm incandescent bulbs and iridescent LED strips. These lights, as the carousel rises and spins, cast a shifting, dreamlike pattern of light on her pale skin and the hem of her pure white dress. Her lips are slightly parted, her expression shifting from initial confusion to an indescribable wonder. In the background, other wooden horses, lions, and ornate mirrored decorations merge into a brilliant halo due to the shallow depth of field and motion blur, creating a surreal atmosphere that is both disorienting and full of fantasy magic.", + "seed": 4400, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_2916_seed_4400_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05825590677559376, + "vlm_logprob": -0.05825590677559376, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.05825590677559376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014425160363316536 + ], + "opposite_logprob_tokens": [ + -4.889425277709961 + ], + "logprob_reward": -0.014425160363316536, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.014425160363316536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17152439057826996, + 0.11636244505643845, + 0.05300923436880112, + 0.0861257016658783, + 0.036252476274967194, + 0.047608599066734314, + 0.2097092568874359, + 0.10000915825366974, + 0.12976542115211487, + 0.04963328316807747 + ], + "credit_map": [ + 0.017134591937065125, + 0.024620449170470238, + 0.01856883056461811, + 0.024346260353922844, + 0.017933327704668045, + 0.02996613085269928, + 0.013838701881468296, + 0.025116100907325745, + 0.012065479531884193, + 0.009226975962519646, + 0.018771085888147354, + 0.026539549231529236, + 0.009287758730351925, + 0.009079894050955772, + 0.01335355918854475, + 0.018038148060441017, + 0.004397351760417223, + 0.004754440393298864, + 0.0075388760305941105, + 0.010142550803720951, + 0.004929272923618555, + 0.006233502179384232, + 0.008725978434085846, + 0.0062872604466974735, + 0.01233217678964138, + 0.010008509270846844, + 0.01076761819422245, + 0.008968139998614788, + 0.014026531018316746, + 0.012772287242114544, + 0.008008675649762154, + 0.009241759777069092, + 0.0038574393838644028, + 0.003825483378022909, + 0.004426168277859688, + 0.003837789176031947, + 0.005159727763384581, + 0.0060469950549304485, + 0.0049814824014902115, + 0.004117389675229788, + 0.005383859388530254, + 0.004327593836933374, + 0.006157406140118837, + 0.007833966985344887, + 0.005474140401929617, + 0.004654277116060257, + 0.009119365364313126, + 0.004657989367842674, + 0.03885713219642639, + 0.024606037884950638, + 0.02575773186981678, + 0.025108372792601585, + 0.029891761019825935, + 0.019642148166894913, + 0.026947347447276115, + 0.018898719921708107, + 0.011220327578485012, + 0.008210347965359688, + 0.008503583259880543, + 0.021829253062605858, + 0.009616275317966938, + 0.009019928053021431, + 0.013319692574441433, + 0.018289754167199135, + 0.015230827033519745, + 0.00839382316917181, + 0.01613909937441349, + 0.013486404903233051, + 0.011159676127135754, + 0.01665439084172249, + 0.016114627942442894, + 0.03258657827973366, + 0.005855998955667019, + 0.0049947635270655155, + 0.00483278650790453, + 0.006961672566831112, + 0.007593270391225815, + 0.006528656464070082, + 0.005939139518886805, + 0.006926998030394316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.014425160363316536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371942415833473 + ], + "opposite_logprob_tokens": [ + -3.408719539642334 + ], + "logprob_reward": -0.03371942415833473, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371942415833473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06866199523210526, + 0.06108006834983826, + 0.024613860994577408, + 0.08866583555936813, + 0.02870328165590763, + 0.11378666758537292, + 0.07635693997144699, + 0.42776501178741455, + 0.07908210158348083, + 0.03128425031900406 + ], + "credit_map": [ + 0.006883414462208748, + 0.009357493370771408, + 0.00932312197983265, + 0.013207715936005116, + 0.006455598399043083, + 0.008685148321092129, + 0.007611058186739683, + 0.007138440385460854, + 0.006883463356643915, + 0.004037094302475452, + 0.011209170334041119, + 0.010648089461028576, + 0.0049686124548316, + 0.006628914270550013, + 0.0047934092581272125, + 0.011911315843462944, + 0.0024487944319844246, + 0.0029474813491106033, + 0.0035395228769630194, + 0.002549326280131936, + 0.002837181556969881, + 0.003275785129517317, + 0.002913339529186487, + 0.004102428909391165, + 0.012043367139995098, + 0.007834375835955143, + 0.011864474974572659, + 0.014501865021884441, + 0.009963924065232277, + 0.009630953893065453, + 0.008279421366751194, + 0.014547453261911869, + 0.0027849730104207993, + 0.0032020059879869223, + 0.003471862291917205, + 0.0030878742691129446, + 0.003390226047486067, + 0.005655658431351185, + 0.004164645913988352, + 0.0029460359364748, + 0.00752812996506691, + 0.004473507404327393, + 0.007370158564299345, + 0.013538530096411705, + 0.032102689146995544, + 0.0274307020008564, + 0.010838467627763748, + 0.010504486039280891, + 0.007634260226041079, + 0.007950074039399624, + 0.006655944976955652, + 0.012425978668034077, + 0.009713615290820599, + 0.006810035556554794, + 0.015413767658174038, + 0.00975326169282198, + 0.02883775718510151, + 0.02872740477323532, + 0.019733475521206856, + 0.12134145945310593, + 0.015990326181054115, + 0.04374736547470093, + 0.03904120624065399, + 0.13034602999687195, + 0.00613873265683651, + 0.006711696274578571, + 0.005442637484520674, + 0.009242176078259945, + 0.007742161862552166, + 0.010381896048784256, + 0.012675726786255836, + 0.020747072994709015, + 0.0022502122446894646, + 0.0023543096613138914, + 0.001335779088549316, + 0.004903677850961685, + 0.003661154070869088, + 0.006087902933359146, + 0.0066966768354177475, + 0.0039945365861058235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.03371942415833473 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160755515098572 + ], + "opposite_logprob_tokens": [ + -1.701607584953308 + ], + "logprob_reward": -0.20160755515098572, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160755515098572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051738061010837555, + 0.04460084065794945, + 0.0402233824133873, + 0.06690440326929092, + 0.024226555600762367, + 0.08675387501716614, + 0.06834188103675842, + 0.3654974699020386, + 0.18071426451206207, + 0.07099923491477966 + ], + "credit_map": [ + 0.0037497032899409533, + 0.004964739084243774, + 0.00617527961730957, + 0.010541797615587711, + 0.0053267995826900005, + 0.006778806913644075, + 0.007916362956166267, + 0.006284573581069708, + 0.005659905262291431, + 0.004723476245999336, + 0.006343530490994453, + 0.006940100342035294, + 0.006853251717984676, + 0.00393502926453948, + 0.005861115641891956, + 0.004284432623535395, + 0.003658610861748457, + 0.0033400871325284243, + 0.006465981714427471, + 0.006676627788692713, + 0.004362809471786022, + 0.0043908036313951015, + 0.005728327669203281, + 0.005600138567388058, + 0.008418040350079536, + 0.004762028343975544, + 0.008428586646914482, + 0.0124405762180686, + 0.006082207430154085, + 0.009465692564845085, + 0.007065421435981989, + 0.010241849347949028, + 0.002240490633994341, + 0.0019021658226847649, + 0.0035577290691435337, + 0.002075282856822014, + 0.0032244143076241016, + 0.0052932389080524445, + 0.0023181207943707705, + 0.0036151146050542593, + 0.005965488031506538, + 0.004677079617977142, + 0.0065790461376309395, + 0.014632470905780792, + 0.017598245292901993, + 0.015429784543812275, + 0.012651466764509678, + 0.009220294654369354, + 0.008111314848065376, + 0.009097401984035969, + 0.005953949876129627, + 0.009496555663645267, + 0.008780036121606827, + 0.006547755561769009, + 0.013126765377819538, + 0.007228102535009384, + 0.023297742009162903, + 0.029964018613100052, + 0.017864177003502846, + 0.09842746704816818, + 0.017203643918037415, + 0.035980045795440674, + 0.034415144473314285, + 0.10834524780511856, + 0.012877942994236946, + 0.015522351488471031, + 0.019879315048456192, + 0.02785777673125267, + 0.014906716533005238, + 0.018889794126152992, + 0.0321931317448616, + 0.03858723118901253, + 0.004783556796610355, + 0.0055601526983082294, + 0.0036898672115057707, + 0.010927565395832062, + 0.011856917291879654, + 0.01297057420015335, + 0.014630969613790512, + 0.006579631008207798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.20160755515098572 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018224529922008514 + ], + "opposite_logprob_tokens": [ + -4.018224716186523 + ], + "logprob_reward": -0.018224529922008514, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018224529922008514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10676208883523941, + 0.05826052278280258, + 0.05993398651480675, + 0.1325487494468689, + 0.06170300394296646, + 0.02825150266289711, + 0.2518566846847534, + 0.1370137631893158, + 0.09275379031896591, + 0.07091590762138367 + ], + "credit_map": [ + 0.009595232084393501, + 0.009505938738584518, + 0.018876200541853905, + 0.011736440472304821, + 0.009695605374872684, + 0.018130362033843994, + 0.010238506831228733, + 0.018983811140060425, + 0.007218095473945141, + 0.004577300976961851, + 0.011278988793492317, + 0.007920373231172562, + 0.005105967633426189, + 0.0061156898736953735, + 0.007272498682141304, + 0.008771606720983982, + 0.0067425742745399475, + 0.0058411345817148685, + 0.01249169372022152, + 0.004656415432691574, + 0.005363140255212784, + 0.009905551560223103, + 0.006138340104371309, + 0.008795137517154217, + 0.01820492371916771, + 0.013388462364673615, + 0.019094476476311684, + 0.022087031975388527, + 0.013799324631690979, + 0.013602644205093384, + 0.0110979825258255, + 0.021273907274007797, + 0.006628728937357664, + 0.006479429546743631, + 0.007106699049472809, + 0.005728151649236679, + 0.007205033674836159, + 0.015013959258794785, + 0.006199517287313938, + 0.007341485936194658, + 0.003040524199604988, + 0.00400887057185173, + 0.003110828110948205, + 0.004231219179928303, + 0.003147557144984603, + 0.0032318835146725178, + 0.0047529018484056, + 0.00272771785967052, + 0.03252272680401802, + 0.027241183444857597, + 0.03754725307226181, + 0.03305976465344429, + 0.03708808496594429, + 0.02195708639919758, + 0.03855552524328232, + 0.023885073140263557, + 0.009983444586396217, + 0.010088248178362846, + 0.0063080983236432076, + 0.037741437554359436, + 0.005957629531621933, + 0.01396652776747942, + 0.013683395460247993, + 0.03928496688604355, + 0.010297715663909912, + 0.007914538495242596, + 0.011340955272316933, + 0.012833399698138237, + 0.0076482114382088184, + 0.010637136176228523, + 0.013826803304255009, + 0.018255028873682022, + 0.007953260093927383, + 0.00649105804041028, + 0.004199400078505278, + 0.013661658391356468, + 0.009275185875594616, + 0.009386811405420303, + 0.014127238653600216, + 0.005821294616907835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.018224529922008514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302864283323288 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302864283323288, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302864283323288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10089673101902008, + 0.07845313102006912, + 0.06415325403213501, + 0.0901380404829979, + 0.07141326367855072, + 0.0701703354716301, + 0.05313682183623314, + 0.37123483419418335, + 0.059601277112960815, + 0.04080231860280037 + ], + "credit_map": [ + 0.009680849500000477, + 0.015354152768850327, + 0.012782631441950798, + 0.01592409983277321, + 0.011588692665100098, + 0.016394944861531258, + 0.008273939602077007, + 0.010897427797317505, + 0.009626081213355064, + 0.006390683818608522, + 0.01209805067628622, + 0.01731170527637005, + 0.007433239836245775, + 0.005699556320905685, + 0.009459974244236946, + 0.010433845221996307, + 0.006276880390942097, + 0.006793564185500145, + 0.00983849074691534, + 0.006999618839472532, + 0.007336973678320646, + 0.01088824775069952, + 0.00640989700332284, + 0.009609591215848923, + 0.013473792932927608, + 0.010807838290929794, + 0.01259899977594614, + 0.010739977471530437, + 0.011634813621640205, + 0.007315291557461023, + 0.007375197485089302, + 0.01619214192032814, + 0.006613894831389189, + 0.007437278050929308, + 0.009406893514096737, + 0.006894242018461227, + 0.008211647160351276, + 0.012983867898583412, + 0.006666403729468584, + 0.013199040666222572, + 0.005021438933908939, + 0.007052283268421888, + 0.005826527718454599, + 0.006391723174601793, + 0.013239946216344833, + 0.01416868343949318, + 0.006901788059622049, + 0.011567952111363411, + 0.007896030321717262, + 0.005654148291796446, + 0.007497577928006649, + 0.005434927996248007, + 0.00587352504953742, + 0.0074993036687374115, + 0.0077462452463805676, + 0.005535067990422249, + 0.03024965152144432, + 0.036679044365882874, + 0.03894984722137451, + 0.07850421965122223, + 0.030619865283370018, + 0.04062502458691597, + 0.04133686423301697, + 0.0742703229188919, + 0.00681429635733366, + 0.005356342066079378, + 0.010180302895605564, + 0.008344241417944431, + 0.005557287484407425, + 0.005167440511286259, + 0.008443736471235752, + 0.009737629443407059, + 0.003717019921168685, + 0.004409892950206995, + 0.0047429646365344524, + 0.008795920759439468, + 0.006028302013874054, + 0.0036679115146398544, + 0.00601711543276906, + 0.0034231971949338913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.023302864283323288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, synchronized with the circular motion of a carousel, focuses firmly on the back of a magnificent white painted wooden horse, where a young woman lies crouched. Her long, slightly tousled dark brown hair, a few strands clinging to her cheeks, a sleep-tinged flush. She wears a simple white cotton dress that seems incongruous with the scene. As the horse rises and falls steadily, as if awakened by the gentle jolt, her eyelids begin to tremble slightly, and then, with a touch of uncertainty, she raises her head. Her heavy eyelids slowly open, initially confused and empty, before adjusting for a moment, they gradually become occupied by the surrounding scene, reflecting a myriad of light. Her gaze is a world of flowing light, composed of thousands of warm incandescent bulbs and iridescent LED strips. These lights, as the carousel rises and spins, cast a shifting, dreamlike pattern of light on her pale skin and the hem of her pure white dress. Her lips are slightly parted, her expression shifting from initial confusion to an indescribable wonder. In the background, other wooden horses, lions, and ornate mirrored decorations merge into a brilliant halo due to the shallow depth of field and motion blur, creating a surreal atmosphere that is both disorienting and full of fantasy magic.", + "seed": 4401, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_2916_seed_4401_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04811453483998775, + "vlm_logprob": -0.04811453483998775, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04811453483998775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019469117745757103 + ], + "opposite_logprob_tokens": [ + -4.269469261169434 + ], + "logprob_reward": -0.019469117745757103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.019469117745757103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34508949518203735, + 0.18702906370162964, + 0.07994351536035538, + 0.036447394639253616, + 0.03304621949791908, + 0.1141803190112114, + 0.02555278316140175, + 0.036390386521816254, + 0.11158420145511627, + 0.030736573040485382 + ], + "credit_map": [ + 0.028451085090637207, + 0.05534767359495163, + 0.03130220249295235, + 0.03672366589307785, + 0.05779258906841278, + 0.04844866320490837, + 0.05375303328037262, + 0.033270612359046936, + 0.011760994791984558, + 0.011867445893585682, + 0.03133227303624153, + 0.016832925379276276, + 0.012600879184901714, + 0.009857209399342537, + 0.06639241427183151, + 0.026384945958852768, + 0.0045914314687252045, + 0.005375243257731199, + 0.008813739754259586, + 0.01446814090013504, + 0.009117419831454754, + 0.009071717038750648, + 0.01687980256974697, + 0.01162603311240673, + 0.0037702666595578194, + 0.004240294452756643, + 0.003987986594438553, + 0.003069523023441434, + 0.00880462396889925, + 0.0037264348939061165, + 0.0042126071639359, + 0.004635660909116268, + 0.003497383091598749, + 0.0025965096428990364, + 0.004308543633669615, + 0.003338016802445054, + 0.003727589501067996, + 0.004218647256493568, + 0.006161306519061327, + 0.0051982286386191845, + 0.011505021713674068, + 0.01155503187328577, + 0.005239028949290514, + 0.01631123758852482, + 0.01551109179854393, + 0.0204310342669487, + 0.017627649009227753, + 0.01600024476647377, + 0.0026672978419810534, + 0.0018779158126562834, + 0.00321821216493845, + 0.0037495503202080727, + 0.002907576272264123, + 0.002263119909912348, + 0.0057935225777328014, + 0.0030755915213376284, + 0.0034376257099211216, + 0.005351265426725149, + 0.0030921685975044966, + 0.0033854341600090265, + 0.006657286547124386, + 0.0031787811312824488, + 0.006632599979639053, + 0.0046552326530218124, + 0.013555662706494331, + 0.02313324250280857, + 0.012833490036427975, + 0.013210426084697247, + 0.013718383386731148, + 0.007636350579559803, + 0.011836661957204342, + 0.015660002827644348, + 0.0030001522973179817, + 0.0036252448335289955, + 0.002839120104908943, + 0.004627552814781666, + 0.005307257641106844, + 0.0031709554605185986, + 0.004185403231531382, + 0.003980891779065132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.019469117745757103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547831982374191 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547831982374191, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547831982374191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07809237390756607, + 0.5238105058670044, + 0.03184633329510689, + 0.05367433652281761, + 0.029495492577552795, + 0.12272729724645615, + 0.030351843684911728, + 0.029979081824421883, + 0.03489137440919876, + 0.06513144075870514 + ], + "credit_map": [ + 0.006943868473172188, + 0.013221967034041882, + 0.00620016036555171, + 0.008467013947665691, + 0.011273819953203201, + 0.008086038753390312, + 0.01597730629146099, + 0.007922184653580189, + 0.029832873493433, + 0.027046212926506996, + 0.09917522221803665, + 0.037991028279066086, + 0.0291957575827837, + 0.013549323193728924, + 0.20284709334373474, + 0.08417294919490814, + 0.0027014659717679024, + 0.0021999517921358347, + 0.003551436122506857, + 0.004717285744845867, + 0.004300225526094437, + 0.004034859593957663, + 0.006201913580298424, + 0.004139188211411238, + 0.005210506729781628, + 0.005564950872212648, + 0.0064317029900848866, + 0.0036084335297346115, + 0.014406410977244377, + 0.006900222972035408, + 0.004488624632358551, + 0.0070634763687849045, + 0.0033931927755475044, + 0.0023776667658239603, + 0.004121485631912947, + 0.0031768926419317722, + 0.0028429192025214434, + 0.003085350850597024, + 0.006564149633049965, + 0.003933829255402088, + 0.01159269455820322, + 0.012145541608333588, + 0.006171956192702055, + 0.01675030030310154, + 0.01860123686492443, + 0.021102961152791977, + 0.01891274005174637, + 0.017449846491217613, + 0.0015979268355295062, + 0.0022390091326087713, + 0.002840074012055993, + 0.0036412770859897137, + 0.005362709518522024, + 0.002529433462768793, + 0.0076120952144265175, + 0.004529315046966076, + 0.0024471732322126627, + 0.004440080374479294, + 0.003354061162099242, + 0.002578750718384981, + 0.004747910890728235, + 0.003414047881960869, + 0.005782917607575655, + 0.0032141373958438635, + 0.0036395785864442587, + 0.00930909812450409, + 0.0026585538871586323, + 0.004204866476356983, + 0.0053868661634624004, + 0.0020529895555227995, + 0.003969996701925993, + 0.0036694195587188005, + 0.006084993481636047, + 0.004792199470102787, + 0.004609367810189724, + 0.021993326023221016, + 0.006402676925063133, + 0.004970266949385405, + 0.011704778298735619, + 0.004573817364871502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.012547831982374191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991435527801514 + ], + "opposite_logprob_tokens": [ + -1.8049143552780151 + ], + "logprob_reward": -0.17991435527801514, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991435527801514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09343850612640381, + 0.26560959219932556, + 0.042213015258312225, + 0.051150862127542496, + 0.09672101587057114, + 0.024830417707562447, + 0.02241218462586403, + 0.06391863524913788, + 0.05504874140024185, + 0.2846570312976837 + ], + "credit_map": [ + 0.00712093710899353, + 0.016374977305531502, + 0.008343459106981754, + 0.009246224537491798, + 0.01684093475341797, + 0.01115361601114273, + 0.015218592248857021, + 0.00913977436721325, + 0.0154412891715765, + 0.013559086248278618, + 0.049884818494319916, + 0.01889452338218689, + 0.01595539227128029, + 0.008651702664792538, + 0.10124603658914566, + 0.041976746171712875, + 0.0027535255067050457, + 0.0031717566307634115, + 0.0033633073326200247, + 0.0038164788857102394, + 0.008084939792752266, + 0.008448639884591103, + 0.006651116069406271, + 0.005923251621425152, + 0.004682270810008049, + 0.0047441949136555195, + 0.007342988159507513, + 0.00350578804500401, + 0.013788522221148014, + 0.00663526076823473, + 0.003721793182194233, + 0.0067300498485565186, + 0.008707243017852306, + 0.007850444875657558, + 0.014045128598809242, + 0.010718708857893944, + 0.016748949885368347, + 0.01677812822163105, + 0.010748493485152721, + 0.01112392544746399, + 0.0026789591647684574, + 0.003123716451227665, + 0.0025385951157659292, + 0.003861404024064541, + 0.002524165902286768, + 0.0030463184230029583, + 0.003152555087581277, + 0.003904706100001931, + 0.0017613687086850405, + 0.0022562083322554827, + 0.002729212399572134, + 0.0022665208671242, + 0.0024897793773561716, + 0.0017189342761412263, + 0.005945992656052113, + 0.003244169056415558, + 0.0035762512125074863, + 0.00839301012456417, + 0.004033379256725311, + 0.015161918476223946, + 0.009468918666243553, + 0.004834782797843218, + 0.013484884053468704, + 0.004965493455529213, + 0.004271003417670727, + 0.011939497664570808, + 0.004571784753352404, + 0.009477314539253712, + 0.009584473446011543, + 0.00548776239156723, + 0.005682988092303276, + 0.004033924546092749, + 0.023325970396399498, + 0.029641827568411827, + 0.010233371518552303, + 0.07407615333795547, + 0.03464376553893089, + 0.020833782851696014, + 0.05660289153456688, + 0.03529930114746094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.17991435527801514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012566432356834412 + ], + "opposite_logprob_tokens": [ + -4.387566566467285 + ], + "logprob_reward": -0.012566432356834412, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012566432356834412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05513126030564308, + 0.3855448067188263, + 0.024810556322336197, + 0.09199252724647522, + 0.0541786290705204, + 0.07693216949701309, + 0.024092117324471474, + 0.050018228590488434, + 0.07542604953050613, + 0.1618736833333969 + ], + "credit_map": [ + 0.004938787315040827, + 0.010893292725086212, + 0.005870279390364885, + 0.007924539037048817, + 0.005783574655652046, + 0.004257082473486662, + 0.010984037071466446, + 0.004479671828448772, + 0.02225579507648945, + 0.019453154876828194, + 0.07198865711688995, + 0.02816878631711006, + 0.02178116701543331, + 0.010803861543536186, + 0.14911848306655884, + 0.061974912881851196, + 0.0016932677244767547, + 0.0018762053223326802, + 0.0024760779924690723, + 0.0036551873199641705, + 0.0027225043158978224, + 0.0026897720526903868, + 0.006022452376782894, + 0.003675088519230485, + 0.008559366688132286, + 0.00935045164078474, + 0.012816153466701508, + 0.006790051236748695, + 0.023485902696847916, + 0.01129141915589571, + 0.007829461246728897, + 0.01186972763389349, + 0.004425388760864735, + 0.00787353329360485, + 0.009753747843205929, + 0.005884783808141947, + 0.008051872253417969, + 0.00741786090657115, + 0.005633968394249678, + 0.00513747613877058, + 0.008865008130669594, + 0.008260924369096756, + 0.005245087668299675, + 0.010408403351902962, + 0.009960447438061237, + 0.013237343169748783, + 0.010883753187954426, + 0.010071204043924809, + 0.002158569870516658, + 0.0017311220290139318, + 0.00391450896859169, + 0.002554617589339614, + 0.002896166406571865, + 0.0030833694618195295, + 0.00446407962590456, + 0.0032896853517740965, + 0.004455301444977522, + 0.007310242857784033, + 0.004869602154940367, + 0.009399542585015297, + 0.006442148704081774, + 0.0057838670909404755, + 0.007373978849500418, + 0.004383542109280825, + 0.00891632679849863, + 0.011736051179468632, + 0.008932912722229958, + 0.00831297505646944, + 0.009465864859521389, + 0.011071458458900452, + 0.007857561111450195, + 0.009132899343967438, + 0.013196908868849277, + 0.017863202840089798, + 0.007965578697621822, + 0.049043845385313034, + 0.01571573317050934, + 0.016372090205550194, + 0.030878685414791107, + 0.01083762850612402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.012566432356834412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607493683695793 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.01607493683695793, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607493683695793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18446411192417145, + 0.4870642423629761, + 0.02607032284140587, + 0.1353307068347931, + 0.031701814383268356, + 0.024383949115872383, + 0.016284042969346046, + 0.02380099892616272, + 0.025050055235624313, + 0.045849766582250595 + ], + "credit_map": [ + 0.01495281234383583, + 0.02749488316476345, + 0.015213689766824245, + 0.020817028358578682, + 0.02921408787369728, + 0.02459889091551304, + 0.03438379615545273, + 0.01778894104063511, + 0.028165625408291817, + 0.024552252143621445, + 0.09139371663331985, + 0.037685781717300415, + 0.027116214856505394, + 0.012784495018422604, + 0.18837414681911469, + 0.07699207961559296, + 0.0031754549127072096, + 0.0031010834500193596, + 0.003964029718190432, + 0.004083884414285421, + 0.0028953866567462683, + 0.0021771311294287443, + 0.0037355839740484953, + 0.002937771612778306, + 0.013874245807528496, + 0.011910399422049522, + 0.021020065993070602, + 0.009479955770075321, + 0.034266553819179535, + 0.016100898385047913, + 0.010570007376372814, + 0.018108580261468887, + 0.004453486297279596, + 0.0026060768868774176, + 0.004093968775123358, + 0.0036094363313168287, + 0.002518863184377551, + 0.003136794548481703, + 0.0072886403650045395, + 0.003994552884250879, + 0.0027428404428064823, + 0.0023035327903926373, + 0.0032895400654524565, + 0.003134491154924035, + 0.003119444940239191, + 0.003503880463540554, + 0.002986196894198656, + 0.003304026322439313, + 0.0019279703265056014, + 0.0012244177050888538, + 0.0018351315520703793, + 0.001566269202157855, + 0.0018118146108463407, + 0.001895719557069242, + 0.0037906812503933907, + 0.002232041209936142, + 0.0025571107398718596, + 0.0033773991744965315, + 0.0021580003667622805, + 0.004039939027279615, + 0.002796491142362356, + 0.0034337707329541445, + 0.003736670594662428, + 0.0017016176134347916, + 0.0017716585425660014, + 0.004271094221621752, + 0.002858818741515279, + 0.0035307889338582754, + 0.004709223285317421, + 0.0020230975933372974, + 0.003707021940499544, + 0.0021783546544611454, + 0.003297618357464671, + 0.003809695364907384, + 0.0027754853945225477, + 0.012553677894175053, + 0.0050126854330301285, + 0.0037898279260843992, + 0.009413444437086582, + 0.005197334568947554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01607493683695793 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, synchronized with the circular motion of a carousel, focuses firmly on the back of a magnificent white painted wooden horse, where a young woman lies crouched. Her long, slightly tousled dark brown hair, a few strands clinging to her cheeks, a sleep-tinged flush. She wears a simple white cotton dress that seems incongruous with the scene. As the horse rises and falls steadily, as if awakened by the gentle jolt, her eyelids begin to tremble slightly, and then, with a touch of uncertainty, she raises her head. Her heavy eyelids slowly open, initially confused and empty, before adjusting for a moment, they gradually become occupied by the surrounding scene, reflecting a myriad of light. Her gaze is a world of flowing light, composed of thousands of warm incandescent bulbs and iridescent LED strips. These lights, as the carousel rises and spins, cast a shifting, dreamlike pattern of light on her pale skin and the hem of her pure white dress. Her lips are slightly parted, her expression shifting from initial confusion to an indescribable wonder. In the background, other wooden horses, lions, and ornate mirrored decorations merge into a brilliant halo due to the shallow depth of field and motion blur, creating a surreal atmosphere that is both disorienting and full of fantasy magic.", + "seed": 4402, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_13/index_2916_seed_4402_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05094751603901386, + "vlm_logprob": -0.05094751603901386, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.05094751603901386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039366696029901505 + ], + "opposite_logprob_tokens": [ + -3.7893667221069336 + ], + "logprob_reward": -0.039366696029901505, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.039366696029901505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03530346229672432, + 0.013266772963106632, + 0.00820353627204895, + 0.004116792231798172, + 0.009405450895428658, + 0.034877900034189224, + 0.026971446350216866, + 0.8366196751594543, + 0.026700731366872787, + 0.004534217528998852 + ], + "credit_map": [ + 0.0029133916832506657, + 0.004327932372689247, + 0.0036168661899864674, + 0.006452430505305529, + 0.004149460233747959, + 0.004853618796914816, + 0.00507865147665143, + 0.003911112900823355, + 0.0018434637458994985, + 0.0018359178211539984, + 0.0011140737915411592, + 0.0013761923182755709, + 0.0013080352218821645, + 0.0025555032771080732, + 0.0014699974562972784, + 0.0017635892145335674, + 0.0011997191468253732, + 0.0009629196720197797, + 0.0011933119967579842, + 0.0008889734162949026, + 0.0011084110010415316, + 0.0009614667505957186, + 0.0010863399365916848, + 0.0008023935952223837, + 0.0003156609018333256, + 0.0003784106520470232, + 0.0006607254617847502, + 0.0005269484245218337, + 0.0005028365412726998, + 0.0004634780634660274, + 0.0007906355313025415, + 0.00047809648094698787, + 0.0013833154225721955, + 0.0014479157980531454, + 0.000747346377465874, + 0.001164610031992197, + 0.000994331669062376, + 0.0007189457537606359, + 0.0021374248899519444, + 0.0008115602540783584, + 0.006366457790136337, + 0.0032851649448275566, + 0.004714861046522856, + 0.005688702687621117, + 0.0027582827024161816, + 0.0037581180222332478, + 0.004030218347907066, + 0.0042760963551700115, + 0.004344888962805271, + 0.002746068174019456, + 0.0020866761915385723, + 0.004196114372462034, + 0.0033125116024166346, + 0.002684888429939747, + 0.005589225795120001, + 0.0020110737532377243, + 0.030116558074951172, + 0.028289468958973885, + 0.047062188386917114, + 0.32021135091781616, + 0.036089058965444565, + 0.03963635861873627, + 0.05414588004350662, + 0.2810688018798828, + 0.0018468211637809873, + 0.0029394086450338364, + 0.005136323161423206, + 0.004374773241579533, + 0.004722189623862505, + 0.0020928948651999235, + 0.0038919681683182716, + 0.0016963527305051684, + 0.0006534408312290907, + 0.0009362942073494196, + 0.0003940642636734992, + 0.0004442778299562633, + 0.0005561434663832188, + 0.0005796466139145195, + 0.000541589455679059, + 0.00042876112274825573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.039366696029901505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086198389530182 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.016086198389530182, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086198389530182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027764499187469482, + 0.10182180255651474, + 0.024110114201903343, + 0.034998439252376556, + 0.010142556391656399, + 0.02613574080169201, + 0.07324080914258957, + 0.6687349081039429, + 0.023707354441285133, + 0.009343699552118778 + ], + "credit_map": [ + 0.002277630614116788, + 0.004443848039954901, + 0.0026067737489938736, + 0.003960121888667345, + 0.0051203761249780655, + 0.0025568909477442503, + 0.00373888504691422, + 0.0030599741730839014, + 0.007312757428735495, + 0.014503000304102898, + 0.010241590440273285, + 0.015151118859648705, + 0.017792843282222748, + 0.006946404464542866, + 0.020513080060482025, + 0.00936101283878088, + 0.0015552523545920849, + 0.0017797226319089532, + 0.002993124071508646, + 0.005637428257614374, + 0.0028609635774046183, + 0.002377768512815237, + 0.0037147975526750088, + 0.0031910559628158808, + 0.002474533161148429, + 0.002859630389139056, + 0.005263487808406353, + 0.004356000572443008, + 0.003081266302615404, + 0.003929115831851959, + 0.006032293196767569, + 0.00700211338698864, + 0.0012685670517385006, + 0.0014829167630523443, + 0.0010692788055166602, + 0.001158461906015873, + 0.001168500049971044, + 0.0010123332031071186, + 0.0020009533036500216, + 0.0009815447265282273, + 0.0028980886563658714, + 0.0020501662511378527, + 0.004128181841224432, + 0.0059944274835288525, + 0.002441065851598978, + 0.0021674928721040487, + 0.002756890607997775, + 0.003699427703395486, + 0.01134510152041912, + 0.007971104234457016, + 0.005719142500311136, + 0.013929641805589199, + 0.008392511866986752, + 0.007584062404930592, + 0.013876738958060741, + 0.0044225044548511505, + 0.023855650797486305, + 0.022405186668038368, + 0.03734150528907776, + 0.25133681297302246, + 0.028610119596123695, + 0.03399715945124626, + 0.043154411017894745, + 0.2280340939760208, + 0.0019513017032295465, + 0.0025378610007464886, + 0.003955575171858072, + 0.004054054617881775, + 0.0045472742058336735, + 0.002035323064774275, + 0.0033457272220402956, + 0.0012802379205822945, + 0.0009531588293612003, + 0.001051242696121335, + 0.0006166882230900228, + 0.0005598430288955569, + 0.0009440509020350873, + 0.0016806665807962418, + 0.0010796264978125691, + 0.0024584231432527304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.016086198389530182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035033762454987 + ], + "opposite_logprob_tokens": [ + -1.9103503227233887 + ], + "logprob_reward": -0.16035033762454987, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035033762454987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1456775963306427, + 0.19247756898403168, + 0.11923959851264954, + 0.02415267936885357, + 0.017056407406926155, + 0.20347432792186737, + 0.027954690158367157, + 0.21473869681358337, + 0.025936756283044815, + 0.02929171919822693 + ], + "credit_map": [ + 0.009284802712500095, + 0.022602416574954987, + 0.012087030336260796, + 0.025125931948423386, + 0.030063537880778313, + 0.01321173645555973, + 0.019224755465984344, + 0.014077373780310154, + 0.015345007181167603, + 0.030959704890847206, + 0.01811625435948372, + 0.03130752593278885, + 0.03131505474448204, + 0.014035345055162907, + 0.03350372985005379, + 0.017894955351948738, + 0.010417266748845577, + 0.009856807067990303, + 0.01596316136419773, + 0.01273429486900568, + 0.01809346117079258, + 0.010598787106573582, + 0.020191451534628868, + 0.021384362131357193, + 0.002125951461493969, + 0.002066160086542368, + 0.003870416199788451, + 0.002687381347641349, + 0.0025302101857960224, + 0.0034791426733136177, + 0.0029560287948697805, + 0.0044373865239322186, + 0.0018533001421019435, + 0.0018234489252790809, + 0.0014005012344568968, + 0.0037927136290818453, + 0.0018349969759583473, + 0.0011189122451469302, + 0.003432296682149172, + 0.001800237107090652, + 0.04084189236164093, + 0.020124046131968498, + 0.028906671330332756, + 0.031678859144449234, + 0.015523736365139484, + 0.020839298143982887, + 0.02270708605647087, + 0.02285272628068924, + 0.0034152560401707888, + 0.0030626009684056044, + 0.003180508268997073, + 0.004448144230991602, + 0.0038320363964885473, + 0.003154018661007285, + 0.004329708404839039, + 0.0025324132293462753, + 0.008006063289940357, + 0.0076978025026619434, + 0.011990142054855824, + 0.0797722116112709, + 0.009305664338171482, + 0.01127262506633997, + 0.01401346642524004, + 0.07268070429563522, + 0.0026176052633672953, + 0.004422192927449942, + 0.002619953593239188, + 0.003927403129637241, + 0.004353887867182493, + 0.0019296010723337531, + 0.004458138719201088, + 0.0016079717315733433, + 0.003630537074059248, + 0.0021830268669873476, + 0.002335588913410902, + 0.0016635345527902246, + 0.003039784263819456, + 0.005694033112376928, + 0.004833478480577469, + 0.005911733955144882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.16035033762454987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638416178524494 + ], + "opposite_logprob_tokens": [ + -3.651384115219116 + ], + "logprob_reward": -0.02638416178524494, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638416178524494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09736575931310654, + 0.08473137021064758, + 0.09783615916967392, + 0.04022178798913956, + 0.023562027141451836, + 0.17920364439487457, + 0.06102775037288666, + 0.26450324058532715, + 0.07939165085554123, + 0.07215657830238342 + ], + "credit_map": [ + 0.008893785066902637, + 0.014519345946609974, + 0.01170431636273861, + 0.012373706325888634, + 0.015835924074053764, + 0.010507223196327686, + 0.014167667366564274, + 0.009363789111375809, + 0.006628891918808222, + 0.01342613436281681, + 0.008217517286539078, + 0.01058640144765377, + 0.012032202444970608, + 0.008929948322474957, + 0.010999214835464954, + 0.013911064714193344, + 0.01566539891064167, + 0.009397492744028568, + 0.014247714541852474, + 0.015809105709195137, + 0.00886489637196064, + 0.012333021499216557, + 0.010135011747479439, + 0.01138351671397686, + 0.0033035350497812033, + 0.0034567266702651978, + 0.004348920192569494, + 0.004818757064640522, + 0.004884764552116394, + 0.006723802536725998, + 0.005567232146859169, + 0.00711805047467351, + 0.0026566251181066036, + 0.0031792763620615005, + 0.002194795524701476, + 0.004328029230237007, + 0.002348485868424177, + 0.0023185366299003363, + 0.004247282166033983, + 0.00228899740613997, + 0.036249104887247086, + 0.017375465482473373, + 0.026830419898033142, + 0.024698957800865173, + 0.014779576100409031, + 0.021118715405464172, + 0.021037260070443153, + 0.017114154994487762, + 0.008882537484169006, + 0.007443350274115801, + 0.004506053868681192, + 0.007459878455847502, + 0.0063137453980743885, + 0.01057606190443039, + 0.011493826285004616, + 0.004352300427854061, + 0.011328252963721752, + 0.010898925364017487, + 0.017166560515761375, + 0.1056976392865181, + 0.01111218798905611, + 0.03900083526968956, + 0.017909737303853035, + 0.051389120519161224, + 0.009904731065034866, + 0.009562504477798939, + 0.01609494909644127, + 0.012357478961348534, + 0.010920477099716663, + 0.006309675984084606, + 0.010240782052278519, + 0.004001051187515259, + 0.008427047170698643, + 0.006245707627385855, + 0.005310727283358574, + 0.00405040942132473, + 0.006685204803943634, + 0.015452230349183083, + 0.008635974489152431, + 0.017349276691675186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.02638416178524494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up focusing on the back of a white painted wooden horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman with long dark brown hair lying crouched on the horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman wear a simple white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's gaze reflect a pattern of warm incandescent bulbs and iridescent LED strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring other wooden horses and lions rendered with shallow depth of field and motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255018636584282 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.01255018636584282, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255018636584282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12048249691724777, + 0.10566552728414536, + 0.08849608153104782, + 0.06364212930202484, + 0.11207800358533859, + 0.3525528013706207, + 0.03047333098948002, + 0.033805232495069504, + 0.042533475905656815, + 0.05027085915207863 + ], + "credit_map": [ + 0.012861983850598335, + 0.014417889527976513, + 0.014566210098564625, + 0.01721007749438286, + 0.012185833416879177, + 0.022143973037600517, + 0.014859586954116821, + 0.012236951850354671, + 0.01193811185657978, + 0.007390622980892658, + 0.020281899720430374, + 0.01000916212797165, + 0.008924808353185654, + 0.016311924904584885, + 0.018361184746026993, + 0.01244781631976366, + 0.008947021327912807, + 0.008610204793512821, + 0.012295803055167198, + 0.011180461384356022, + 0.012888307683169842, + 0.017927469685673714, + 0.012003239244222641, + 0.004643584135919809, + 0.005891835317015648, + 0.007095483131706715, + 0.010064740665256977, + 0.007581626530736685, + 0.006982059683650732, + 0.009245770052075386, + 0.006570016033947468, + 0.010210600681602955, + 0.016082946211099625, + 0.0185799989849329, + 0.010748193599283695, + 0.014797345735132694, + 0.011796052567660809, + 0.00964314304292202, + 0.01813480816781521, + 0.012295523658394814, + 0.05998114496469498, + 0.029712481424212456, + 0.03952040150761604, + 0.05661897361278534, + 0.03914918750524521, + 0.03635183721780777, + 0.047861192375421524, + 0.04335763305425644, + 0.004056750796735287, + 0.00309929926879704, + 0.002186831319704652, + 0.005269870162010193, + 0.004166188649833202, + 0.004008140880614519, + 0.005298496689647436, + 0.002387754386290908, + 0.0026275331620126963, + 0.002024355111643672, + 0.0019566372502595186, + 0.0026273662224411964, + 0.00281386636197567, + 0.010283838957548141, + 0.003809315152466297, + 0.007662326097488403, + 0.008075731806457043, + 0.007032461930066347, + 0.00699482299387455, + 0.004264417104423046, + 0.0054136840626597404, + 0.003211754374206066, + 0.005091477185487747, + 0.002449127845466137, + 0.00741945905610919, + 0.0044530583545565605, + 0.0044262828305363655, + 0.002567398827522993, + 0.0037589736748486757, + 0.012735187076032162, + 0.004702149890363216, + 0.010208350606262684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.01255018636584282 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_14/index_tensor([4404, 4404])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_14/index_tensor([4404, 4404])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..59b1134893528304e257ce5f1f0a2b3806c0f41d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_14/index_tensor([4404, 4404])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a dramatic medium close-up, a lone warrior, clad in a weathered suit of ancient armor, stands like a rock atop a slippery black reef. The scene is bathed in a cold, moonlit blue hue, the cold light reflecting off the warrior's armor, woven from dark metal plates and deep blue cord, riddled with scratches and dents, creating a sharp metallic sheen. His black hair, tied in a high bun, is disheveled by the wind, a few wet strands plastering to his angular face. He purses his lips, his deep eyes revealing an unyielding resolve, as he lets the cold water splash against his resolute features. The camerawork is highly dynamic, switching rapidly and unsettlingly between the warrior's face, etched with determination, and the raging waves in the background, rolling and roaring, stirring up vast masses of white foam. Each change of focus is accompanied by a dazzling white flash that instantly cuts through the screen, like silent lightning, which not only illuminates the splashing water droplets, but also intensifies the tense atmosphere of the fierce confrontation between man and nature, forming an extreme visual contrast between the static tenacity of this lonely guardian and the dynamic and violent power of the sea.", + "seed": 28426, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_14/index_4404_seed_28426_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2865939604351297, + "vlm_logprob": -0.2865939604351297, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.2865939604351297, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5331645011901855 + ], + "opposite_logprob_tokens": [ + -0.9081645011901855 + ], + "logprob_reward": -0.5331645011901855, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5331645011901855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1410294622182846, + 0.09375974535942078, + 0.07257308065891266, + 0.041146162897348404, + 0.06486913561820984, + 0.21499885618686676, + 0.05506923422217369, + 0.05015576630830765, + 0.19213823974132538, + 0.07426033169031143 + ], + "credit_map": [ + 0.020681006833910942, + 0.016079001128673553, + 0.023999374359846115, + 0.02045949548482895, + 0.011709419079124928, + 0.018346961587667465, + 0.01870790310204029, + 0.011046304367482662, + 0.009037458337843418, + 0.009658725932240486, + 0.014723686501383781, + 0.022456234320998192, + 0.010775879956781864, + 0.00740062864497304, + 0.009458712302148342, + 0.01024842169135809, + 0.012957028113305569, + 0.006078292615711689, + 0.010568642057478428, + 0.009176484309136868, + 0.00735586928203702, + 0.008666740730404854, + 0.009822702035307884, + 0.007947325706481934, + 0.005073269363492727, + 0.003825374413281679, + 0.008783256635069847, + 0.004652968142181635, + 0.005302354693412781, + 0.004467874765396118, + 0.005343646742403507, + 0.0036974153481423855, + 0.009821658954024315, + 0.006352376192808151, + 0.017292914912104607, + 0.0058968085795640945, + 0.0064805722795426846, + 0.008689266629517078, + 0.005177084822207689, + 0.00515845837071538, + 0.04989401996135712, + 0.029786255210638046, + 0.038363758474588394, + 0.023415062576532364, + 0.016935789957642555, + 0.031462594866752625, + 0.015249463729560375, + 0.009891899302601814, + 0.008596973493695259, + 0.005421804264187813, + 0.008571149781346321, + 0.007272347342222929, + 0.006887386552989483, + 0.007608477026224136, + 0.0065051778219640255, + 0.004205918870866299, + 0.008927333168685436, + 0.006606845185160637, + 0.006188567262142897, + 0.006318789441138506, + 0.004714557435363531, + 0.010160996578633785, + 0.003961313981562853, + 0.003277360927313566, + 0.0351545624434948, + 0.018446771427989006, + 0.029487235471606255, + 0.02461356483399868, + 0.01673019863665104, + 0.029101837426424026, + 0.021985672414302826, + 0.01661839336156845, + 0.009656183421611786, + 0.007451054640114307, + 0.008644787594676018, + 0.010903243906795979, + 0.009854087606072426, + 0.009408933110535145, + 0.007395220920443535, + 0.010946819558739662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.5331645011901855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314744234085083 + ], + "opposite_logprob_tokens": [ + -3.7733147144317627 + ], + "logprob_reward": -0.023314744234085083, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314744234085083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36148881912231445, + 0.10716298222541809, + 0.09123603254556656, + 0.0809665396809578, + 0.04231778904795647, + 0.10172785818576813, + 0.05433377996087074, + 0.03898882865905762, + 0.06507007032632828, + 0.056707240641117096 + ], + "credit_map": [ + 0.034417394548654556, + 0.03921302780508995, + 0.07727678120136261, + 0.04443031921982765, + 0.0289614200592041, + 0.0434088297188282, + 0.05962028354406357, + 0.0341607891023159, + 0.009803736582398415, + 0.008986778557300568, + 0.025669123977422714, + 0.01659930869936943, + 0.009699790738523006, + 0.008948617614805698, + 0.014919540844857693, + 0.01253608986735344, + 0.014937440864741802, + 0.006065486930310726, + 0.009178461506962776, + 0.0153996292501688, + 0.013429088518023491, + 0.011944443918764591, + 0.011732249520719051, + 0.00854923203587532, + 0.009469416923820972, + 0.007788943126797676, + 0.012965057976543903, + 0.009580078534781933, + 0.01185050792992115, + 0.00855950266122818, + 0.011697965674102306, + 0.009055077098309994, + 0.005404478404670954, + 0.00471130758523941, + 0.007985241711139679, + 0.004251161590218544, + 0.005642509553581476, + 0.0051928069442510605, + 0.003995626699179411, + 0.005134657956659794, + 0.021034708246588707, + 0.01288608182221651, + 0.0161734726279974, + 0.010097741149365902, + 0.010432644747197628, + 0.013836205005645752, + 0.008363587781786919, + 0.008903424255549908, + 0.0073255072347819805, + 0.008907667361199856, + 0.006119978614151478, + 0.007576095871627331, + 0.007332175504416227, + 0.005805515218526125, + 0.006376801524311304, + 0.00489004235714674, + 0.007719090674072504, + 0.004354659002274275, + 0.00626040855422616, + 0.003546525724232197, + 0.003158816834911704, + 0.007975059561431408, + 0.003135284408926964, + 0.0028389859944581985, + 0.013853918761014938, + 0.006341135129332542, + 0.0074278065003454685, + 0.006443959195166826, + 0.005878467112779617, + 0.014636741019785404, + 0.005420433357357979, + 0.0050676120445132256, + 0.008029190823435783, + 0.0049662464298307896, + 0.004842966329306364, + 0.007575574330985546, + 0.005376811139285564, + 0.007716413587331772, + 0.011470640078186989, + 0.0067294007167220116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.023314744234085083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001716570113785565 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.001716570113785565, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001716570113785565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.179128497838974, + 0.08119149506092072, + 0.1233079731464386, + 0.030076440423727036, + 0.06561465561389923, + 0.06237310171127319, + 0.03571707010269165, + 0.09693099558353424, + 0.2564361095428467, + 0.06922364234924316 + ], + "credit_map": [ + 0.020570581778883934, + 0.030234599485993385, + 0.030926581472158432, + 0.027832794934511185, + 0.02132982574403286, + 0.017150072380900383, + 0.017836380749940872, + 0.013247658498585224, + 0.007973411120474339, + 0.007464317139238119, + 0.012652290984988213, + 0.015875348821282387, + 0.008511233143508434, + 0.008066347800195217, + 0.00940376240760088, + 0.011244780384004116, + 0.021890997886657715, + 0.012937825173139572, + 0.020989887416362762, + 0.02115236036479473, + 0.01188968401402235, + 0.015138490125536919, + 0.011768778786063194, + 0.007539952639490366, + 0.0034592454321682453, + 0.002905634231865406, + 0.0038203950971364975, + 0.00465080002322793, + 0.004970526788383722, + 0.0032040888909250498, + 0.004957417957484722, + 0.0021083340980112553, + 0.012991918250918388, + 0.007669287268072367, + 0.013506180606782436, + 0.00548333115875721, + 0.00640473747625947, + 0.009105431847274303, + 0.005622017662972212, + 0.004831753671169281, + 0.01326767448335886, + 0.008207292295992374, + 0.010488199070096016, + 0.006738276220858097, + 0.0063745854422450066, + 0.008531593717634678, + 0.005301213823258877, + 0.003464268520474434, + 0.005474228411912918, + 0.005993752274662256, + 0.0038562787231057882, + 0.004054752178490162, + 0.00424626050516963, + 0.004291965160518885, + 0.0036928756162524223, + 0.004106955602765083, + 0.01792810671031475, + 0.009561849758028984, + 0.016193710267543793, + 0.008725824765861034, + 0.009936047717928886, + 0.020671667531132698, + 0.008300965651869774, + 0.0056128245778381824, + 0.052549488842487335, + 0.02146996185183525, + 0.02841656282544136, + 0.024718884378671646, + 0.017231203615665436, + 0.05815073102712631, + 0.023796914145350456, + 0.030102375894784927, + 0.008801904506981373, + 0.007513165008276701, + 0.005438332911580801, + 0.008548609912395477, + 0.007076981943100691, + 0.00917717907577753, + 0.010462828911840916, + 0.01220464427024126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.001716570113785565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261196613311768 + ], + "opposite_logprob_tokens": [ + -0.5761196613311768 + ], + "logprob_reward": -0.8261196613311768, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261196613311768, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08199214190244675, + 0.0726563110947609, + 0.059827178716659546, + 0.03479502722620964, + 0.06650911271572113, + 0.05725346878170967, + 0.1092347577214241, + 0.13989174365997314, + 0.31045928597450256, + 0.06738098710775375 + ], + "credit_map": [ + 0.010359463281929493, + 0.008361207321286201, + 0.014394184574484825, + 0.011251700110733509, + 0.008255472406744957, + 0.008697190321981907, + 0.01024580653756857, + 0.010427122935652733, + 0.008879021741449833, + 0.005135267041623592, + 0.014962267130613327, + 0.014500216580927372, + 0.004790137521922588, + 0.00888807326555252, + 0.007720959838479757, + 0.007780374959111214, + 0.007812450639903545, + 0.004648763220757246, + 0.009202334098517895, + 0.010239538736641407, + 0.010263850912451744, + 0.005423474591225386, + 0.007672202307730913, + 0.00456456933170557, + 0.004015349317342043, + 0.003245361614972353, + 0.004855692386627197, + 0.004445207305252552, + 0.005886750761419535, + 0.0048689525574445724, + 0.004531023092567921, + 0.002946695080026984, + 0.0056693716906011105, + 0.005633447784930468, + 0.007370999548584223, + 0.007345723453909159, + 0.011643534526228905, + 0.0066493237391114235, + 0.008274619467556477, + 0.01392209529876709, + 0.007323989178985357, + 0.008004789240658283, + 0.00742843933403492, + 0.007970716804265976, + 0.009339622221887112, + 0.006937976460903883, + 0.005706014111638069, + 0.00454192329198122, + 0.014757291413843632, + 0.016532719135284424, + 0.017670009285211563, + 0.011704428121447563, + 0.014122179709374905, + 0.011798491701483727, + 0.01245336327701807, + 0.010196276940405369, + 0.02489902265369892, + 0.011939560994505882, + 0.03420846164226532, + 0.01359033863991499, + 0.010928671807050705, + 0.021961873397231102, + 0.01314543653279543, + 0.009218381717801094, + 0.06570138782262802, + 0.02585657685995102, + 0.0330340638756752, + 0.031273387372493744, + 0.02123168855905533, + 0.07306686788797379, + 0.028136001899838448, + 0.03215932473540306, + 0.006624958012253046, + 0.007048661820590496, + 0.00487779825925827, + 0.009345365688204765, + 0.006176655646413565, + 0.006460445001721382, + 0.013308040797710419, + 0.01353906374424696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.8261196613311768 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865432530641556 + ], + "opposite_logprob_tokens": [ + -3.048654317855835 + ], + "logprob_reward": -0.04865432530641556, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865432530641556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1843838095664978, + 0.09046171605587006, + 0.049209363758563995, + 0.04706074297428131, + 0.11396949738264084, + 0.03381569683551788, + 0.08604993671178818, + 0.04558412730693817, + 0.2874581515789032, + 0.062007032334804535 + ], + "credit_map": [ + 0.02539411000907421, + 0.016920393332839012, + 0.046920448541641235, + 0.02443939633667469, + 0.012967090122401714, + 0.020878992974758148, + 0.02389339543879032, + 0.012969964183866978, + 0.010151049122214317, + 0.0072149671614170074, + 0.020071258768439293, + 0.020190950483083725, + 0.006313714664429426, + 0.009101253002882004, + 0.008716054260730743, + 0.008702459745109081, + 0.008269023150205612, + 0.004258435219526291, + 0.006268779281526804, + 0.007821673527359962, + 0.006851118057966232, + 0.005303093697875738, + 0.006832432467490435, + 0.0036047983448952436, + 0.00578384380787611, + 0.004124418366700411, + 0.011801289394497871, + 0.005654127802699804, + 0.006534353364259005, + 0.004934921395033598, + 0.005355655215680599, + 0.002872124081477523, + 0.027940422296524048, + 0.014372468926012516, + 0.02330240048468113, + 0.00924411416053772, + 0.007604418322443962, + 0.018264953047037125, + 0.006317111197859049, + 0.00692360196262598, + 0.0045225066132843494, + 0.004675260279327631, + 0.00489999121055007, + 0.004818850662559271, + 0.0050513045862317085, + 0.0037876672577112913, + 0.0037666717544198036, + 0.002293440978974104, + 0.011586175300180912, + 0.01376635953783989, + 0.006438924465328455, + 0.01135717611759901, + 0.010715785436332226, + 0.010017014108598232, + 0.013116995804011822, + 0.009051489643752575, + 0.007485799491405487, + 0.005690675228834152, + 0.00576791912317276, + 0.004491160623729229, + 0.004734653979539871, + 0.010049305856227875, + 0.004536715801805258, + 0.002827889285981655, + 0.06349675357341766, + 0.02252374216914177, + 0.0293622687458992, + 0.026411699131131172, + 0.019023915752768517, + 0.07023272663354874, + 0.025586429983377457, + 0.030820555984973907, + 0.008239254355430603, + 0.006136964540928602, + 0.004137961193919182, + 0.007536493707448244, + 0.005348804406821728, + 0.009907609783113003, + 0.010845550335943699, + 0.009854382835328579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.04865432530641556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a dramatic medium close-up, a lone warrior, clad in a weathered suit of ancient armor, stands like a rock atop a slippery black reef. The scene is bathed in a cold, moonlit blue hue, the cold light reflecting off the warrior's armor, woven from dark metal plates and deep blue cord, riddled with scratches and dents, creating a sharp metallic sheen. His black hair, tied in a high bun, is disheveled by the wind, a few wet strands plastering to his angular face. He purses his lips, his deep eyes revealing an unyielding resolve, as he lets the cold water splash against his resolute features. The camerawork is highly dynamic, switching rapidly and unsettlingly between the warrior's face, etched with determination, and the raging waves in the background, rolling and roaring, stirring up vast masses of white foam. Each change of focus is accompanied by a dazzling white flash that instantly cuts through the screen, like silent lightning, which not only illuminates the splashing water droplets, but also intensifies the tense atmosphere of the fierce confrontation between man and nature, forming an extreme visual contrast between the static tenacity of this lonely guardian and the dynamic and violent power of the sea.", + "seed": 28427, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_14/index_4404_seed_28427_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12428158146794885, + "vlm_logprob": -0.12428158146794885, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.12428158146794885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.530978262424469 + ], + "opposite_logprob_tokens": [ + -0.905978262424469 + ], + "logprob_reward": -0.530978262424469, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.530978262424469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07876727730035782, + 0.3132230341434479, + 0.06535977125167847, + 0.04433463513851166, + 0.05606719106435776, + 0.04217172786593437, + 0.23081809282302856, + 0.0703655481338501, + 0.06377214938402176, + 0.03512055426836014 + ], + "credit_map": [ + 0.007048123516142368, + 0.013359284959733486, + 0.009340696968138218, + 0.007411560975015163, + 0.015199209563434124, + 0.008677114732563496, + 0.008934570476412773, + 0.008796721696853638, + 0.033881183713674545, + 0.018209382891654968, + 0.08032335340976715, + 0.0153480414301157, + 0.018613094463944435, + 0.11273728311061859, + 0.015095258131623268, + 0.0190154779702425, + 0.008790153078734875, + 0.009678065776824951, + 0.008155548013746738, + 0.0065858191810548306, + 0.009596702642738819, + 0.007920549251139164, + 0.00825856626033783, + 0.006374365650117397, + 0.0039850929751992226, + 0.006578193046152592, + 0.007563743740320206, + 0.0049991300329566, + 0.007209830917418003, + 0.0036065999884158373, + 0.005993884522467852, + 0.004398159217089415, + 0.005594079382717609, + 0.006077989470213652, + 0.00496141379699111, + 0.009563191793859005, + 0.008898932486772537, + 0.006862771697342396, + 0.00635336572304368, + 0.007755445782095194, + 0.003820738522335887, + 0.0049001253210008144, + 0.003957343753427267, + 0.005074291955679655, + 0.0073974705301225185, + 0.007277370430529118, + 0.0037957618478685617, + 0.005948621779680252, + 0.02594776451587677, + 0.02597994916141033, + 0.022209500893950462, + 0.025263715535402298, + 0.01927192322909832, + 0.04186631739139557, + 0.03808445483446121, + 0.03219447284936905, + 0.007430115714669228, + 0.00928787887096405, + 0.009691178798675537, + 0.007700884249061346, + 0.010189220309257507, + 0.008848512545228004, + 0.00838980358093977, + 0.008827954530715942, + 0.005621908698230982, + 0.008661353960633278, + 0.006054085213690996, + 0.005864286329597235, + 0.012668135575950146, + 0.007772095501422882, + 0.005992490332573652, + 0.011137797497212887, + 0.003599476534873247, + 0.004113261122256517, + 0.004054540302604437, + 0.0031889064703136683, + 0.0045821103267371655, + 0.00440734438598156, + 0.003482354572042823, + 0.00769256055355072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.530978262424469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810355976223946 + ], + "opposite_logprob_tokens": [ + -3.5298104286193848 + ], + "logprob_reward": -0.029810355976223946, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810355976223946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08613035082817078, + 0.22020617127418518, + 0.11587168276309967, + 0.09454500675201416, + 0.1468208283185959, + 0.062449343502521515, + 0.06665104627609253, + 0.03138723224401474, + 0.03905615955591202, + 0.13688220083713531 + ], + "credit_map": [ + 0.007545069791376591, + 0.014009788632392883, + 0.01023221667855978, + 0.006814072374254465, + 0.01500829216092825, + 0.013981059193611145, + 0.008381190709769726, + 0.010158657096326351, + 0.026198197156190872, + 0.01122243981808424, + 0.051073018461465836, + 0.013319912366569042, + 0.016038592904806137, + 0.07732031494379044, + 0.011686804704368114, + 0.013346864841878414, + 0.018172122538089752, + 0.0124525036662817, + 0.013475615531206131, + 0.012361481785774231, + 0.019869007170200348, + 0.013954845257103443, + 0.013755250722169876, + 0.011830847710371017, + 0.00808157678693533, + 0.007347384933382273, + 0.01832585223019123, + 0.008715800940990448, + 0.017700018361210823, + 0.007610031869262457, + 0.013119909912347794, + 0.013644430786371231, + 0.016325507313013077, + 0.021242236718535423, + 0.017689963802695274, + 0.016588954254984856, + 0.015929732471704483, + 0.01615247316658497, + 0.018238119781017303, + 0.0246538408100605, + 0.0068949502892792225, + 0.006430039647966623, + 0.006852049380540848, + 0.007077325601130724, + 0.00798586942255497, + 0.01234397292137146, + 0.004513582680374384, + 0.010351555421948433, + 0.0073758456856012344, + 0.010138742625713348, + 0.005844877567142248, + 0.007439549081027508, + 0.007203633431345224, + 0.013698872178792953, + 0.009668024256825447, + 0.005281500983983278, + 0.0028430521488189697, + 0.006630952004343271, + 0.00301011698320508, + 0.004874496720731258, + 0.004995779599994421, + 0.00331111834384501, + 0.0029985401779413223, + 0.002723174635320902, + 0.003807276487350464, + 0.005043128039687872, + 0.004199610557407141, + 0.003146073082461953, + 0.00824409443885088, + 0.004779106937348843, + 0.003804208477959037, + 0.006032662000507116, + 0.00704201590269804, + 0.010413090698421001, + 0.012666357681155205, + 0.008486361242830753, + 0.009482774883508682, + 0.04155747964978218, + 0.009569887071847916, + 0.03766423836350441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.029810355976223946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001042894902639091 + ], + "opposite_logprob_tokens": [ + -6.87604284286499 + ], + "logprob_reward": -0.001042894902639091, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001042894902639091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17166748642921448, + 0.055122919380664825, + 0.12740154564380646, + 0.07104621082544327, + 0.061657462269067764, + 0.09074677526950836, + 0.21023911237716675, + 0.055758893489837646, + 0.05465814471244812, + 0.10170148313045502 + ], + "credit_map": [ + 0.014810594730079174, + 0.02518421970307827, + 0.023386096581816673, + 0.0186512041836977, + 0.033298879861831665, + 0.021439053118228912, + 0.01669219136238098, + 0.01820524036884308, + 0.004676187876611948, + 0.006653275806456804, + 0.013740910217165947, + 0.006054173223674297, + 0.006143566220998764, + 0.008009035140275955, + 0.005156627390533686, + 0.004689145367592573, + 0.019552690908312798, + 0.01616836152970791, + 0.01583804003894329, + 0.011248618364334106, + 0.02379763498902321, + 0.015973778441548347, + 0.014628772623836994, + 0.010193643160164356, + 0.007939385250210762, + 0.00828658975660801, + 0.010437263175845146, + 0.008309470489621162, + 0.01237543672323227, + 0.006796679925173521, + 0.00921877846121788, + 0.007682609371840954, + 0.005353719927370548, + 0.008036037907004356, + 0.007668003905564547, + 0.007277108728885651, + 0.008689343929290771, + 0.006817431654781103, + 0.007686807308346033, + 0.010129008442163467, + 0.0073146820068359375, + 0.010745343752205372, + 0.01025699544698, + 0.026363138109445572, + 0.006592285353690386, + 0.009743616916239262, + 0.011228027753531933, + 0.008502687327563763, + 0.020438391715288162, + 0.024319659918546677, + 0.016667645424604416, + 0.02680620364844799, + 0.019836323335766792, + 0.04286424443125725, + 0.034696295857429504, + 0.024610351771116257, + 0.005053999368101358, + 0.006700389552861452, + 0.00725064892321825, + 0.005980133544653654, + 0.008872258476912975, + 0.008600746281445026, + 0.00538989994674921, + 0.007910815998911858, + 0.0039046716410666704, + 0.009290549904108047, + 0.005853554233908653, + 0.006111705210059881, + 0.008969479240477085, + 0.007428769022226334, + 0.005357361398637295, + 0.00774205569177866, + 0.007649128790944815, + 0.009855328127741814, + 0.006896111648529768, + 0.005895630456507206, + 0.01741577312350273, + 0.020221292972564697, + 0.00899122841656208, + 0.024776993319392204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.001042894902639091 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029786866158246994 + ], + "opposite_logprob_tokens": [ + -3.5297868251800537 + ], + "logprob_reward": -0.029786866158246994, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029786866158246994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08132616430521011, + 0.1241714283823967, + 0.1025720089673996, + 0.0895489752292633, + 0.032634805887937546, + 0.12085797637701035, + 0.10059075802564621, + 0.18138210475444794, + 0.12200793623924255, + 0.044907912611961365 + ], + "credit_map": [ + 0.00624872837215662, + 0.011014285497367382, + 0.012514249421656132, + 0.009385284967720509, + 0.018875837326049805, + 0.006501173134893179, + 0.008919437415897846, + 0.007867167703807354, + 0.007075021043419838, + 0.00649698032066226, + 0.040724147111177444, + 0.0067580631002783775, + 0.007554353214800358, + 0.04239870607852936, + 0.005560284946113825, + 0.007603862322866917, + 0.018132785335183144, + 0.01487694587558508, + 0.01457698829472065, + 0.008626689203083515, + 0.013573330827057362, + 0.012794280424714088, + 0.010117588564753532, + 0.009873396717011929, + 0.009900052100419998, + 0.006611275020986795, + 0.014362955465912819, + 0.007564852945506573, + 0.022483475506305695, + 0.007820043712854385, + 0.008007687516510487, + 0.012798630632460117, + 0.0036979923024773598, + 0.005129425320774317, + 0.003162198234349489, + 0.00414890144020319, + 0.00429979432374239, + 0.0032097664661705494, + 0.004907590802758932, + 0.004079133737832308, + 0.010493556037545204, + 0.014220903627574444, + 0.01429914403706789, + 0.03349665552377701, + 0.01006972324103117, + 0.012691419571638107, + 0.014725269749760628, + 0.010861298069357872, + 0.01341213844716549, + 0.017597565427422523, + 0.011411845684051514, + 0.010494405403733253, + 0.009306046180427074, + 0.01344178058207035, + 0.013767026364803314, + 0.011159938760101795, + 0.014002969488501549, + 0.02141571044921875, + 0.018172359094023705, + 0.01835920475423336, + 0.03285124525427818, + 0.03484829142689705, + 0.014487159438431263, + 0.02724515087902546, + 0.006872573867440224, + 0.023093663156032562, + 0.01266385056078434, + 0.019736258313059807, + 0.013612416572868824, + 0.015895584598183632, + 0.015054654330015182, + 0.01507892832159996, + 0.0027426481246948242, + 0.00289229117333889, + 0.003654329339042306, + 0.0023764465004205704, + 0.004406520165503025, + 0.011940662749111652, + 0.0024095484986901283, + 0.014485463500022888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.029786866158246994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789527878165245 + ], + "opposite_logprob_tokens": [ + -3.529789447784424 + ], + "logprob_reward": -0.029789527878165245, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789527878165245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10101432353258133, + 0.07655303180217743, + 0.1548669934272766, + 0.049042604863643646, + 0.04807085916399956, + 0.10916367918252945, + 0.12001898884773254, + 0.19116084277629852, + 0.13592863082885742, + 0.014180020429193974 + ], + "credit_map": [ + 0.009698000736534595, + 0.010987835936248302, + 0.015098722651600838, + 0.012984334491193295, + 0.01726803556084633, + 0.013534078374505043, + 0.011341447941958904, + 0.010101864114403725, + 0.013917684555053711, + 0.009236307814717293, + 0.015616164542734623, + 0.005547716747969389, + 0.00948314368724823, + 0.006736345123499632, + 0.007963643409311771, + 0.008052024990320206, + 0.03249066323041916, + 0.01812373287975788, + 0.022527793422341347, + 0.012651280499994755, + 0.019845053553581238, + 0.023895032703876495, + 0.014808096922934055, + 0.010525334626436234, + 0.0052580926567316055, + 0.004512780345976353, + 0.006623896770179272, + 0.00509561225771904, + 0.009657755494117737, + 0.004479516297578812, + 0.007192745804786682, + 0.0062222047708928585, + 0.005258543416857719, + 0.00573278171941638, + 0.004416473209857941, + 0.005749983713030815, + 0.005894594360142946, + 0.005665382836014032, + 0.00798221118748188, + 0.0073708887211978436, + 0.009759134612977505, + 0.012148087844252586, + 0.010437395423650742, + 0.0316532701253891, + 0.008668850176036358, + 0.012352567166090012, + 0.013307962566614151, + 0.010836414992809296, + 0.010594509541988373, + 0.01722843199968338, + 0.008617188781499863, + 0.018124055117368698, + 0.013942357152700424, + 0.02583232894539833, + 0.01537700742483139, + 0.010303106158971786, + 0.013368708081543446, + 0.022173848003149033, + 0.013063072226941586, + 0.020271996036171913, + 0.0371529720723629, + 0.03968648985028267, + 0.015129935927689075, + 0.03031381592154503, + 0.011636915616691113, + 0.02427602745592594, + 0.015840839594602585, + 0.015789669007062912, + 0.019040066748857498, + 0.017629660665988922, + 0.011651422828435898, + 0.020064042881131172, + 0.0016139039071276784, + 0.0013436087174341083, + 0.0012705117696896195, + 0.0012449686182662845, + 0.0023160725831985474, + 0.002622848143801093, + 0.0011415822664275765, + 0.0026265247724950314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.029789527878165245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dramatic medium close-up, a lone warrior, clad in a weathered suit of ancient armor, stands like a rock atop a slippery black reef. The scene is bathed in a cold, moonlit blue hue, the cold light reflecting off the warrior's armor, woven from dark metal plates and deep blue cord, riddled with scratches and dents, creating a sharp metallic sheen. His black hair, tied in a high bun, is disheveled by the wind, a few wet strands plastering to his angular face. He purses his lips, his deep eyes revealing an unyielding resolve, as he lets the cold water splash against his resolute features. The camerawork is highly dynamic, switching rapidly and unsettlingly between the warrior's face, etched with determination, and the raging waves in the background, rolling and roaring, stirring up vast masses of white foam. Each change of focus is accompanied by a dazzling white flash that instantly cuts through the screen, like silent lightning, which not only illuminates the splashing water droplets, but also intensifies the tense atmosphere of the fierce confrontation between man and nature, forming an extreme visual contrast between the static tenacity of this lonely guardian and the dynamic and violent power of the sea.", + "seed": 28428, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_14/index_4404_seed_28428_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21410879746545106, + "vlm_logprob": -0.21410879746545106, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.21410879746545106, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8326973915100098 + ], + "opposite_logprob_tokens": [ + -0.5826973915100098 + ], + "logprob_reward": -0.8326973915100098, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8326973915100098, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21939393877983093, + 0.09596537798643112, + 0.05028362199664116, + 0.0669812336564064, + 0.04401734471321106, + 0.07337073236703873, + 0.04347359016537666, + 0.05105864256620407, + 0.18651214241981506, + 0.16894342005252838 + ], + "credit_map": [ + 0.03420920670032501, + 0.04468585178256035, + 0.021185925230383873, + 0.025652164593338966, + 0.019054602831602097, + 0.027348292991518974, + 0.028141800314188004, + 0.019116081297397614, + 0.00973486341536045, + 0.010031801648437977, + 0.013735768385231495, + 0.015362690202891827, + 0.008195057511329651, + 0.014881143346428871, + 0.012459379620850086, + 0.011564667336642742, + 0.0065230922773480415, + 0.008280654437839985, + 0.0069543663412332535, + 0.008092239499092102, + 0.004395302385091782, + 0.005488160066306591, + 0.0050820242613554, + 0.005467779468744993, + 0.012848121114075184, + 0.01662382297217846, + 0.007054135203361511, + 0.0070824348367750645, + 0.005688241682946682, + 0.00586028303951025, + 0.005910937208682299, + 0.005913255736231804, + 0.003923204727470875, + 0.008715491741895676, + 0.01084335520863533, + 0.004852144047617912, + 0.0044487942941486835, + 0.00329957134090364, + 0.0038159959949553013, + 0.004118784796446562, + 0.010567102581262589, + 0.005593455862253904, + 0.01532040722668171, + 0.007872220128774643, + 0.008400402031838894, + 0.009026193991303444, + 0.00862179696559906, + 0.007969141937792301, + 0.007176538929343224, + 0.0040853023529052734, + 0.006499909795820713, + 0.005649592261761427, + 0.003781673964112997, + 0.006507991813123226, + 0.0045552439987659454, + 0.005217334255576134, + 0.006870010402053595, + 0.005614875815808773, + 0.005854577291756868, + 0.005613191984593868, + 0.008793226443231106, + 0.005399289075285196, + 0.006651907693594694, + 0.006261561997234821, + 0.025111237540841103, + 0.01914576254785061, + 0.02071681246161461, + 0.02887003868818283, + 0.019253741949796677, + 0.03027825988829136, + 0.024269042536616325, + 0.01886722818017006, + 0.03691544756293297, + 0.015022174455225468, + 0.032134007662534714, + 0.01338411495089531, + 0.014076294377446175, + 0.030473513528704643, + 0.012780317105352879, + 0.01415752898901701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.8326973915100098 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906083554029465 + ], + "opposite_logprob_tokens": [ + -2.4640607833862305 + ], + "logprob_reward": -0.08906083554029465, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906083554029465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08404073864221573, + 0.028613023459911346, + 0.02388831041753292, + 0.12296455353498459, + 0.03562114015221596, + 0.08393596857786179, + 0.05531642958521843, + 0.019767943769693375, + 0.49549081921577454, + 0.05036099627614021 + ], + "credit_map": [ + 0.01114390417933464, + 0.01254185289144516, + 0.009654400870203972, + 0.011932723224163055, + 0.013398166745901108, + 0.007445689290761948, + 0.009572839364409447, + 0.008351168595254421, + 0.003108101198449731, + 0.003371237078681588, + 0.002945025684311986, + 0.0023505722638219595, + 0.003826720640063286, + 0.003806705353781581, + 0.0057042925618588924, + 0.003500369843095541, + 0.002128343563526869, + 0.001935129170306027, + 0.003924467600882053, + 0.006686483044177294, + 0.0020302734337747097, + 0.0016311494400724769, + 0.0033372158650308847, + 0.002215250628069043, + 0.012526375241577625, + 0.022922959178686142, + 0.009741831570863724, + 0.019219567999243736, + 0.019901782274246216, + 0.011274225078523159, + 0.013637331314384937, + 0.01374049112200737, + 0.004630540031939745, + 0.009608585387468338, + 0.005372479557991028, + 0.0034411950036883354, + 0.0026092121843248606, + 0.003338576527312398, + 0.0026068915612995625, + 0.0040136645548045635, + 0.007122659124433994, + 0.004893770907074213, + 0.008436012081801891, + 0.01443573273718357, + 0.007039252202957869, + 0.019220784306526184, + 0.008889089338481426, + 0.013898676261305809, + 0.01125988271087408, + 0.0035727028734982014, + 0.006941355299204588, + 0.0051575531251728535, + 0.004409376531839371, + 0.014739492908120155, + 0.004179396200925112, + 0.005056676920503378, + 0.0030886689200997353, + 0.0019057454774156213, + 0.0025602439418435097, + 0.0018301305826753378, + 0.002579435473307967, + 0.002872135955840349, + 0.002091506030410528, + 0.0028400802984833717, + 0.07206285744905472, + 0.045338224619627, + 0.0569271445274353, + 0.07429107278585434, + 0.04379761591553688, + 0.09334508329629898, + 0.06012241169810295, + 0.04960642009973526, + 0.00628885580226779, + 0.004899467807263136, + 0.007393352687358856, + 0.007794036529958248, + 0.006178382784128189, + 0.0057238019071519375, + 0.005353705491870642, + 0.006729397457093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.08906083554029465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011819765204563737 + ], + "opposite_logprob_tokens": [ + -6.7511820793151855 + ], + "logprob_reward": -0.0011819765204563737, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011819765204563737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23488107323646545, + 0.13587389886379242, + 0.07136105000972748, + 0.05403633415699005, + 0.02983204834163189, + 0.1373865008354187, + 0.06463805586099625, + 0.1349792778491974, + 0.06204567849636078, + 0.07496605813503265 + ], + "credit_map": [ + 0.028287535533308983, + 0.03183883801102638, + 0.02862258441746235, + 0.04213758558034897, + 0.0243473369628191, + 0.02455832250416279, + 0.02960067242383957, + 0.025488201528787613, + 0.009202571585774422, + 0.014025256037712097, + 0.016837408766150475, + 0.016818145290017128, + 0.016848314553499222, + 0.02040363848209381, + 0.023797687143087387, + 0.017940878868103027, + 0.009388859383761883, + 0.014696326106786728, + 0.008492652326822281, + 0.009738768450915813, + 0.005806661210954189, + 0.007435896899551153, + 0.008491799235343933, + 0.007310092449188232, + 0.004785540048032999, + 0.011300682090222836, + 0.005421592853963375, + 0.015320057980716228, + 0.00490439822897315, + 0.0034275385551154613, + 0.003468535142019391, + 0.005407988093793392, + 0.0031657922081649303, + 0.0037311885971575975, + 0.004362830892205238, + 0.004311918281018734, + 0.004385271109640598, + 0.002607103670015931, + 0.0034067421220242977, + 0.003861201461404562, + 0.015551717951893806, + 0.010678917169570923, + 0.011634341441094875, + 0.020929435268044472, + 0.011386622674763203, + 0.02459469623863697, + 0.01960776187479496, + 0.02300301380455494, + 0.009797624312341213, + 0.005479144863784313, + 0.008392054587602615, + 0.00835664663463831, + 0.00791584700345993, + 0.009956258349120617, + 0.006893123500049114, + 0.007847362197935581, + 0.02036682516336441, + 0.010165153071284294, + 0.01719435304403305, + 0.014034915715456009, + 0.011972286738455296, + 0.015147064812481403, + 0.01576869934797287, + 0.03032999113202095, + 0.006300231441855431, + 0.00744542246684432, + 0.01057349145412445, + 0.007449276279658079, + 0.007260248530656099, + 0.009224820882081985, + 0.006251688580960035, + 0.007540497463196516, + 0.007560847792774439, + 0.006745726335793734, + 0.015185713768005371, + 0.013497999869287014, + 0.008008904755115509, + 0.007768592331558466, + 0.00854167714715004, + 0.007656599394977093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0011819765204563737 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702789902687073 + ], + "opposite_logprob_tokens": [ + -2.127027988433838 + ], + "logprob_reward": -0.12702789902687073, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702789902687073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0874854102730751, + 0.1680755615234375, + 0.08760228008031845, + 0.11630664765834808, + 0.07796014845371246, + 0.11020330339670181, + 0.07728267461061478, + 0.12757742404937744, + 0.12223204225301743, + 0.0252744909375906 + ], + "credit_map": [ + 0.013855023309588432, + 0.013913785107433796, + 0.008518627844750881, + 0.009037503041327, + 0.011816688813269138, + 0.007834520190954208, + 0.013470806181430817, + 0.009038456715643406, + 0.012435773387551308, + 0.019416386261582375, + 0.01564035564661026, + 0.01574883610010147, + 0.02652980573475361, + 0.0191009771078825, + 0.03565764054656029, + 0.023545796051621437, + 0.007527464535087347, + 0.014058041386306286, + 0.015158276073634624, + 0.020133227109909058, + 0.007021532393991947, + 0.005535366013646126, + 0.008645419962704182, + 0.00952294934540987, + 0.01404762826859951, + 0.03422524034976959, + 0.009183350019156933, + 0.028553027659654617, + 0.007872349582612514, + 0.006359754595905542, + 0.006628366187214851, + 0.009436924941837788, + 0.007676546927541494, + 0.0204736590385437, + 0.01152684073895216, + 0.009642207995057106, + 0.00846436433494091, + 0.00656132772564888, + 0.005782360211014748, + 0.007832842878997326, + 0.01121615245938301, + 0.010765805840492249, + 0.02271530218422413, + 0.014980544336140156, + 0.012352978810667992, + 0.011487026698887348, + 0.011792542412877083, + 0.014892954379320145, + 0.010262875817716122, + 0.006910094991326332, + 0.01175135187804699, + 0.010629423893988132, + 0.009923464618623257, + 0.009730895049870014, + 0.00835929811000824, + 0.009715269319713116, + 0.020381484180688858, + 0.011083784513175488, + 0.01831457018852234, + 0.012788176536560059, + 0.013010204769670963, + 0.015677068382501602, + 0.015389731153845787, + 0.020932404324412346, + 0.01681891269981861, + 0.010854213498532772, + 0.017005842179059982, + 0.016323281452059746, + 0.010808175429701805, + 0.02717694640159607, + 0.012675234116613865, + 0.010569444857537746, + 0.004079269710928202, + 0.0017823808593675494, + 0.004477451089769602, + 0.002559415763244033, + 0.002309439703822136, + 0.005493902135640383, + 0.0020118008833378553, + 0.002560832304880023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.12702789902687073 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575884729623795 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.020575884729623795, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575884729623795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15491491556167603, + 0.10127037763595581, + 0.0782092735171318, + 0.08479700982570648, + 0.04494252800941467, + 0.08825279027223587, + 0.20777557790279388, + 0.13544020056724548, + 0.021250765770673752, + 0.0831466093659401 + ], + "credit_map": [ + 0.027747519314289093, + 0.031772613525390625, + 0.016311248764395714, + 0.013579952530562878, + 0.016706153750419617, + 0.021726490929722786, + 0.012964787892997265, + 0.014106130227446556, + 0.01031902153044939, + 0.011908150278031826, + 0.012482927180826664, + 0.011305779218673706, + 0.012471307069063187, + 0.018189363181591034, + 0.013463316485285759, + 0.011130506172776222, + 0.0072169313207268715, + 0.004082901403307915, + 0.011305129155516624, + 0.025570454075932503, + 0.0053542847745120525, + 0.004942380823194981, + 0.010775983333587646, + 0.008961200714111328, + 0.007428551092743874, + 0.010314751416444778, + 0.007965505123138428, + 0.025712646543979645, + 0.011678863316774368, + 0.0046343267895281315, + 0.007684302516281605, + 0.009378058835864067, + 0.004278265405446291, + 0.009130935184657574, + 0.007169566582888365, + 0.00584739213809371, + 0.004415632225573063, + 0.0037802571896463633, + 0.005369505845010281, + 0.004950972273945808, + 0.012828621082007885, + 0.007814434356987476, + 0.01535661332309246, + 0.00948900356888771, + 0.010353455320000648, + 0.011126115918159485, + 0.01222056895494461, + 0.00906397495418787, + 0.04530176892876625, + 0.01215493492782116, + 0.024028385058045387, + 0.021447932347655296, + 0.014268976636230946, + 0.059857409447431564, + 0.01575288735330105, + 0.014963284134864807, + 0.018928272649645805, + 0.010506939142942429, + 0.01680036447942257, + 0.013942419551312923, + 0.011101782321929932, + 0.01618894562125206, + 0.01604093424975872, + 0.031930528581142426, + 0.002141385804861784, + 0.0025490778498351574, + 0.0032491071615368128, + 0.0029881445225328207, + 0.003356580389663577, + 0.002071609254926443, + 0.002769254148006439, + 0.0021256040781736374, + 0.01600857824087143, + 0.006763652432709932, + 0.01895633339881897, + 0.006764964200556278, + 0.006129927467554808, + 0.014435425400733948, + 0.006487563718110323, + 0.0076001533307135105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.020575884729623795 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dramatic medium close-up, a lone warrior, clad in a weathered suit of ancient armor, stands like a rock atop a slippery black reef. The scene is bathed in a cold, moonlit blue hue, the cold light reflecting off the warrior's armor, woven from dark metal plates and deep blue cord, riddled with scratches and dents, creating a sharp metallic sheen. His black hair, tied in a high bun, is disheveled by the wind, a few wet strands plastering to his angular face. He purses his lips, his deep eyes revealing an unyielding resolve, as he lets the cold water splash against his resolute features. The camerawork is highly dynamic, switching rapidly and unsettlingly between the warrior's face, etched with determination, and the raging waves in the background, rolling and roaring, stirring up vast masses of white foam. Each change of focus is accompanied by a dazzling white flash that instantly cuts through the screen, like silent lightning, which not only illuminates the splashing water droplets, but also intensifies the tense atmosphere of the fierce confrontation between man and nature, forming an extreme visual contrast between the static tenacity of this lonely guardian and the dynamic and violent power of the sea.", + "seed": 28429, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_14/index_4404_seed_28429_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4862679255194962, + "vlm_logprob": -0.4862679255194962, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.4862679255194962, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.1284632682800293 + ], + "opposite_logprob_tokens": [ + -0.12846322357654572 + ], + "logprob_reward": -2.1284632682800293, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.1284632682800293, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09587954729795456, + 0.19605472683906555, + 0.23709924519062042, + 0.17277459800243378, + 0.04802435636520386, + 0.052181147038936615, + 0.03913641348481178, + 0.017633739858865738, + 0.03976312652230263, + 0.10145309567451477 + ], + "credit_map": [ + 0.009446774609386921, + 0.012628273107111454, + 0.008294842205941677, + 0.004925424233078957, + 0.012442109175026417, + 0.019426370039582253, + 0.007040032651275396, + 0.0216757170855999, + 0.018684450536966324, + 0.01922871358692646, + 0.0235899705439806, + 0.02659900300204754, + 0.03721640259027481, + 0.02384781464934349, + 0.022488709539175034, + 0.024399643763899803, + 0.010449690744280815, + 0.00811801478266716, + 0.005838457029312849, + 0.008239700458943844, + 0.0984407439827919, + 0.030125990509986877, + 0.01213907077908516, + 0.0637475773692131, + 0.019551290199160576, + 0.021636754274368286, + 0.022298388183116913, + 0.020162953063845634, + 0.02842581272125244, + 0.023842638358473778, + 0.021923167631030083, + 0.0149335702881217, + 0.007934450171887875, + 0.003776004072278738, + 0.010672219097614288, + 0.004677406512200832, + 0.0031003861222416162, + 0.0037953192368149757, + 0.004782547242939472, + 0.009286021813750267, + 0.004202444572001696, + 0.00567967863753438, + 0.005214970093220472, + 0.008064460009336472, + 0.008891935460269451, + 0.00403122091665864, + 0.004627480171620846, + 0.011468956246972084, + 0.0033655192237347364, + 0.004840088542550802, + 0.006590149365365505, + 0.005589536391198635, + 0.0035069033037871122, + 0.006185332778841257, + 0.004876500926911831, + 0.004182379227131605, + 0.0018236110918223858, + 0.00295554893091321, + 0.0024859667755663395, + 0.003767801681533456, + 0.0011587463086470962, + 0.0016391961835324764, + 0.0023539657704532146, + 0.0014489012537524104, + 0.0036247691605240107, + 0.005235323682427406, + 0.005406726151704788, + 0.005100203212350607, + 0.0036854948848485947, + 0.005264166742563248, + 0.004802912939339876, + 0.0066435267217457294, + 0.01700499840080738, + 0.014836173504590988, + 0.011893601156771183, + 0.010702598839998245, + 0.009556952863931656, + 0.010731962509453297, + 0.013489973731338978, + 0.013236832804977894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -2.1284632682800293 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08910347521305084 + ], + "opposite_logprob_tokens": [ + -2.4641034603118896 + ], + "logprob_reward": -0.08910347521305084, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08910347521305084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0917552188038826, + 0.08675628155469894, + 0.05311096832156181, + 0.16512829065322876, + 0.0714210644364357, + 0.03747635334730148, + 0.06083821877837181, + 0.03658676519989967, + 0.054122745990753174, + 0.34280410408973694 + ], + "credit_map": [ + 0.011830667965114117, + 0.015055610798299313, + 0.011463862843811512, + 0.0066444952972233295, + 0.013287629932165146, + 0.013325694017112255, + 0.007303411141037941, + 0.012843848206102848, + 0.008165013045072556, + 0.008592033758759499, + 0.01085122674703598, + 0.012348657473921776, + 0.016289599239826202, + 0.010188979096710682, + 0.009292233735322952, + 0.011028537526726723, + 0.004874284379184246, + 0.006871645804494619, + 0.004850463010370731, + 0.006798010319471359, + 0.008601360954344273, + 0.008075954392552376, + 0.0064252582378685474, + 0.006613988429307938, + 0.016257338225841522, + 0.02558562159538269, + 0.02153944969177246, + 0.020856129005551338, + 0.023382244631648064, + 0.022269273176789284, + 0.020874885842204094, + 0.014363352209329605, + 0.018003983423113823, + 0.005742741283029318, + 0.008605854585766792, + 0.005625590216368437, + 0.0057698385789990425, + 0.00550826545804739, + 0.006816117092967033, + 0.01534866914153099, + 0.002937871031463146, + 0.0036696670576930046, + 0.004943770822137594, + 0.0071061719208955765, + 0.0035807944368571043, + 0.00533967325463891, + 0.0038916943594813347, + 0.006006711628288031, + 0.0042922357097268105, + 0.00934616569429636, + 0.009440711699426174, + 0.010849927552044392, + 0.004924362059682608, + 0.007931752130389214, + 0.007444585207849741, + 0.006608481053262949, + 0.003921917639672756, + 0.006146008148789406, + 0.005534220952540636, + 0.006569322198629379, + 0.0032801772467792034, + 0.002720779040828347, + 0.005472519434988499, + 0.002941820537671447, + 0.004806147422641516, + 0.00875063892453909, + 0.006460234522819519, + 0.0073789507150650024, + 0.00558664882555604, + 0.006027813069522381, + 0.006840216461569071, + 0.00827209372073412, + 0.07067607343196869, + 0.05227822810411453, + 0.03586812689900398, + 0.03556409105658531, + 0.029782840982079506, + 0.03504275530576706, + 0.04258786886930466, + 0.04100411757826805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.08910347521305084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099181387573481 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099181387573481, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099181387573481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09081925451755524, + 0.18100625276565552, + 0.2556081712245941, + 0.1375025063753128, + 0.08467182517051697, + 0.06881122291088104, + 0.02841952256858349, + 0.01821811869740486, + 0.024009644985198975, + 0.11093346029520035 + ], + "credit_map": [ + 0.010547077283263206, + 0.011517583392560482, + 0.007935171946883202, + 0.004123910795897245, + 0.013947758823633194, + 0.01578652299940586, + 0.007510217372328043, + 0.01945101097226143, + 0.0174548402428627, + 0.01714160293340683, + 0.021948879584670067, + 0.023975303396582603, + 0.034808363765478134, + 0.021609216928482056, + 0.021199380978941917, + 0.022868653759360313, + 0.011798838153481483, + 0.011171635240316391, + 0.007661323994398117, + 0.013921413570642471, + 0.10097166150808334, + 0.030907901003956795, + 0.013773933984339237, + 0.06540146470069885, + 0.015497596934437752, + 0.015930907800793648, + 0.017413657158613205, + 0.015962334349751472, + 0.023726144805550575, + 0.019228633493185043, + 0.017430022358894348, + 0.012313214130699635, + 0.022064972668886185, + 0.006068050395697355, + 0.012509332969784737, + 0.005751084070652723, + 0.004345110151916742, + 0.0060799820348620415, + 0.0056277792900800705, + 0.0222255177795887, + 0.006079318933188915, + 0.007298842538148165, + 0.008301799185574055, + 0.011841684579849243, + 0.009354657493531704, + 0.005642706528306007, + 0.00719478027895093, + 0.013097434304654598, + 0.002339123748242855, + 0.004268222954124212, + 0.003187731606885791, + 0.005388799589127302, + 0.002096880692988634, + 0.003554956754669547, + 0.004194748122245073, + 0.00338905886746943, + 0.0018150771502405405, + 0.003334138775244355, + 0.0023947940208017826, + 0.003954814746975899, + 0.0011099823750555515, + 0.0016070251585915685, + 0.002724135760217905, + 0.001278150244615972, + 0.0022953746374696493, + 0.002600855426862836, + 0.0033596805296838284, + 0.003282582387328148, + 0.0026340445037931204, + 0.0030662226490676403, + 0.002958795288577676, + 0.0038120895624160767, + 0.019091758877038956, + 0.015722336247563362, + 0.012969310395419598, + 0.011410394683480263, + 0.010143156163394451, + 0.013107222504913807, + 0.014967147260904312, + 0.013522142544388771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.004099181387573481 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987561225891113 + ], + "opposite_logprob_tokens": [ + -1.8048756122589111 + ], + "logprob_reward": -0.17987561225891113, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987561225891113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10379310697317123, + 0.06005709245800972, + 0.28239136934280396, + 0.05672586336731911, + 0.1039549931883812, + 0.08175741881132126, + 0.09922163933515549, + 0.018560515716671944, + 0.0414835549890995, + 0.15205448865890503 + ], + "credit_map": [ + 0.011811641044914722, + 0.013204571790993214, + 0.010732410475611687, + 0.007828280329704285, + 0.015837108716368675, + 0.01522979699075222, + 0.008862987160682678, + 0.020286303013563156, + 0.0059409732930362225, + 0.0053739058785140514, + 0.00705634243786335, + 0.00857400894165039, + 0.009913904592394829, + 0.008832696825265884, + 0.007010299246758223, + 0.00735495425760746, + 0.013256617821753025, + 0.01590697467327118, + 0.010118450038135052, + 0.020660053938627243, + 0.10420200973749161, + 0.03419453278183937, + 0.016291184350848198, + 0.06776153296232224, + 0.004647800698876381, + 0.009674401953816414, + 0.007687423378229141, + 0.009516650810837746, + 0.004583524074405432, + 0.006792631931602955, + 0.008067449554800987, + 0.005755973048508167, + 0.01589476689696312, + 0.00962676852941513, + 0.01551148109138012, + 0.007213465869426727, + 0.008463233709335327, + 0.01662490889430046, + 0.009713000617921352, + 0.02090735360980034, + 0.0046180738136172295, + 0.008776669390499592, + 0.007327843923121691, + 0.01456404197961092, + 0.009763169102370739, + 0.007575947791337967, + 0.005864992737770081, + 0.023266669362783432, + 0.006488005630671978, + 0.01615314744412899, + 0.017995379865169525, + 0.019625410437583923, + 0.006101661361753941, + 0.010053381323814392, + 0.011341794393956661, + 0.011462843976914883, + 0.0016456895973533392, + 0.0027300145011395216, + 0.002284678863361478, + 0.004164739511907101, + 0.0013623140985146165, + 0.0027827825397253036, + 0.0021278962958604097, + 0.001462397282011807, + 0.003747690701857209, + 0.011653623543679714, + 0.0038818444591015577, + 0.006185278762131929, + 0.004216517321765423, + 0.003368252655491233, + 0.0043820650316774845, + 0.004048279020935297, + 0.030075449496507645, + 0.02350982464849949, + 0.01653558947145939, + 0.017860135063529015, + 0.013941120356321335, + 0.014191115275025368, + 0.017722973600029945, + 0.01821826957166195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.17987561225891113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a slippery black reef with raging waves and white foam in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing weathered ancient armor made of dark metal plates and deep blue cord with visible scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a cold, moonlit blue hue with water splashing against the warrior's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera switch rapidly and unsettlingly between the warrior's face and the background waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are each of the camera focus changes accompanied by a dazzling white flash that illuminates water droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979809045791626 + ], + "opposite_logprob_tokens": [ + -3.5297980308532715 + ], + "logprob_reward": -0.02979809045791626, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979809045791626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12673698365688324, + 0.19555245339870453, + 0.14152392745018005, + 0.10059400647878647, + 0.11531540006399155, + 0.1528451293706894, + 0.03742602467536926, + 0.033351875841617584, + 0.05861543491482735, + 0.03803876414895058 + ], + "credit_map": [ + 0.018300477415323257, + 0.017256546765565872, + 0.012432205490767956, + 0.009788387455046177, + 0.020877955481410027, + 0.018719015643000603, + 0.01091083325445652, + 0.018451573327183723, + 0.02163783833384514, + 0.02020963653922081, + 0.023588355630636215, + 0.02058703824877739, + 0.025958634912967682, + 0.030133234336972237, + 0.026244904845952988, + 0.027192818000912666, + 0.00840076059103012, + 0.010873486287891865, + 0.006419404409825802, + 0.008336040191352367, + 0.048843637108802795, + 0.01900944858789444, + 0.009511792100965977, + 0.030129356309771538, + 0.010080481879413128, + 0.0175508763641119, + 0.012752423994243145, + 0.01224411278963089, + 0.01226053386926651, + 0.011453499086201191, + 0.014560184441506863, + 0.009691893123090267, + 0.02470366470515728, + 0.009012366645038128, + 0.020370502024888992, + 0.009790664538741112, + 0.007678395602852106, + 0.01006634347140789, + 0.009844370186328888, + 0.023849083110690117, + 0.01109074056148529, + 0.013878203928470612, + 0.015171205624938011, + 0.0208794716745615, + 0.02541988715529442, + 0.01303861290216446, + 0.013281146064400673, + 0.04008585959672928, + 0.0031121026258915663, + 0.003985354211181402, + 0.004344020504504442, + 0.004662316292524338, + 0.0038279721047729254, + 0.005257054697722197, + 0.005252597853541374, + 0.006984606850892305, + 0.0018136283615604043, + 0.002870971569791436, + 0.011504937894642353, + 0.009840615093708038, + 0.001138907391577959, + 0.0014715768629685044, + 0.002787229837849736, + 0.001924008596688509, + 0.0048373108729720116, + 0.009038446471095085, + 0.007340891286730766, + 0.009999210946261883, + 0.005282571539282799, + 0.005571373738348484, + 0.008472234010696411, + 0.008073396980762482, + 0.005837361793965101, + 0.003995702601969242, + 0.0051903799176216125, + 0.004160323180258274, + 0.004995673429220915, + 0.005268333945423365, + 0.004139720927923918, + 0.004451266024261713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.02979809045791626 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_15/index_tensor([5994, 5994])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_15/index_tensor([5994, 5994])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..97dfb2227737548057e4bf6832e3706a1a4ce88a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_15/index_tensor([5994, 5994])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, high-speed panning shot, held close to the ground, closely follows a dark green, modified 4x4, its body, fenders, and undercarriage coated in thick, dried mud. It's speeding down a sun-drenched, gravelly dirt road. As the vehicle violently cuts into a sharp curve, its rear wheels instantly lose grip, sending the rear end slinging outward at a precise and dramatic angle, creating a flawless power drift. In that instant, the massive, deeply treaded all-terrain tires, with astonishing force, fling a torrent of sand, dust, and small rocks backward and to the sides, forming a spectacular, opaque wall of dust. In the glaring afternoon sun, every particle of sand and mud is clearly visible, as if momentarily frozen in mid-air. The entire scene is imbued with a raw sense of power, speed, and passion, creating an atmosphere of intense excitement and adrenaline.", + "seed": 20345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_15/index_5994_seed_20345_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0384779060492292, + "vlm_logprob": -0.0384779060492292, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.0384779060492292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035402365028858185 + ], + "opposite_logprob_tokens": [ + -6.410402297973633 + ], + "logprob_reward": -0.035402365028858185, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.035402365028858185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055891234427690506, + 0.03311288729310036, + 0.03223707526922226, + 0.4908803105354309, + 0.026769420132040977, + 0.03711308538913727, + 0.03131896257400513, + 0.11254104226827621, + 0.08009304106235504, + 0.10004296898841858 + ], + "credit_map": [ + 0.006470398511737585, + 0.005399714689701796, + 0.007165619172155857, + 0.006426230538636446, + 0.006440319120883942, + 0.00794924981892109, + 0.005458065774291754, + 0.010581635870039463, + 0.003789549693465233, + 0.003647241508588195, + 0.002950110239908099, + 0.004241683054715395, + 0.006511681247502565, + 0.003948598634451628, + 0.004615765064954758, + 0.0034082597121596336, + 0.0025228136219084263, + 0.005699905566871166, + 0.002308690920472145, + 0.004420244600623846, + 0.004405088722705841, + 0.005935415159910917, + 0.0037354337982833385, + 0.0032094812486320734, + 0.01541488990187645, + 0.04085896909236908, + 0.008471420966088772, + 0.009291349910199642, + 0.2015749216079712, + 0.2073284238576889, + 0.0032891840673983097, + 0.004651164170354605, + 0.0033602239564061165, + 0.002774059073999524, + 0.0040859500877559185, + 0.0033805212005972862, + 0.003935279790312052, + 0.003536824369803071, + 0.003033876419067383, + 0.0026626847684383392, + 0.003914575558155775, + 0.0035968762822449207, + 0.0027942724991589785, + 0.004306270275264978, + 0.0056021446362137794, + 0.008790669031441212, + 0.005238937214016914, + 0.0028693408239632845, + 0.0020445173140615225, + 0.0032875952310860157, + 0.003384689334779978, + 0.0051786648109555244, + 0.0032842191867530346, + 0.00427712919190526, + 0.004301395267248154, + 0.005560749210417271, + 0.009771947748959064, + 0.011503827758133411, + 0.013760009780526161, + 0.019992683082818985, + 0.012512517161667347, + 0.018397893756628036, + 0.01399240642786026, + 0.012609758414328098, + 0.01072995737195015, + 0.009058126248419285, + 0.011537862941622734, + 0.010776814073324203, + 0.008291462436318398, + 0.006329800002276897, + 0.014236759394407272, + 0.009132258594036102, + 0.023225175216794014, + 0.027539610862731934, + 0.010799447074532509, + 0.01159636490046978, + 0.007026187144219875, + 0.005589735694229603, + 0.008914925158023834, + 0.005351529456675053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.035402365028858185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557720765471458 + ], + "opposite_logprob_tokens": [ + -4.387557506561279 + ], + "logprob_reward": -0.012557720765471458, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557720765471458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10493391007184982, + 0.06627962738275528, + 0.07871207594871521, + 0.1827702671289444, + 0.07221495360136032, + 0.031818851828575134, + 0.048984404653310776, + 0.12432985007762909, + 0.14134041965007782, + 0.14861562848091125 + ], + "credit_map": [ + 0.00953747145831585, + 0.013149905949831009, + 0.01115510519593954, + 0.016771838068962097, + 0.018472962081432343, + 0.015790574252605438, + 0.007824933156371117, + 0.012231114320456982, + 0.007179505191743374, + 0.009758327156305313, + 0.005462130531668663, + 0.006044880487024784, + 0.013450607657432556, + 0.0087401969358325, + 0.00560659309849143, + 0.010037385858595371, + 0.0059400866739451885, + 0.013289046473801136, + 0.004766057711094618, + 0.009704195894300938, + 0.012414015829563141, + 0.016521940007805824, + 0.00786503590643406, + 0.008211692795157433, + 0.01224997267127037, + 0.022819479927420616, + 0.008542290888726711, + 0.010084491223096848, + 0.06270623207092285, + 0.05147620290517807, + 0.010443185456097126, + 0.004448402673006058, + 0.0066500473767519, + 0.00732448510825634, + 0.00579787464812398, + 0.005798798520117998, + 0.011236917227506638, + 0.016323626041412354, + 0.011001245118677616, + 0.00808196421712637, + 0.0022560434881597757, + 0.0032134484499692917, + 0.001974383369088173, + 0.003329649567604065, + 0.007067524828016758, + 0.00686612818390131, + 0.0038091004826128483, + 0.0033025739248842, + 0.0032553577329963446, + 0.0053825038485229015, + 0.00370577210560441, + 0.0084494948387146, + 0.006647950503975153, + 0.006196713540703058, + 0.008465009741485119, + 0.006881603971123695, + 0.015858100727200508, + 0.009598677046597004, + 0.025325562804937363, + 0.02055414579808712, + 0.011546572670340538, + 0.01979830302298069, + 0.013455787673592567, + 0.008192699402570724, + 0.01839039847254753, + 0.01796291396021843, + 0.020509641617536545, + 0.030705135315656662, + 0.011040274053812027, + 0.009965748526155949, + 0.022207818925380707, + 0.010558493435382843, + 0.040991343557834625, + 0.045274823904037476, + 0.024879731237888336, + 0.013911300338804722, + 0.005968437064439058, + 0.0054560755379498005, + 0.006682132370769978, + 0.005451784003525972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.012557720765471458 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292169988155365 + ], + "opposite_logprob_tokens": [ + -2.237921714782715 + ], + "logprob_reward": -0.11292169988155365, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292169988155365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05279282107949257, + 0.045323457568883896, + 0.024024758487939835, + 0.0674477219581604, + 0.031689368188381195, + 0.10898352414369583, + 0.06335444003343582, + 0.0979660376906395, + 0.08922772854566574, + 0.4191901683807373 + ], + "credit_map": [ + 0.0066030132584273815, + 0.005256338510662317, + 0.0054353466257452965, + 0.00824002269655466, + 0.005577555857598782, + 0.006754235364496708, + 0.0053846994414925575, + 0.009541602805256844, + 0.004492307081818581, + 0.005919572431594133, + 0.004640295170247555, + 0.005372560583055019, + 0.010153395123779774, + 0.005332306493073702, + 0.003972386941313744, + 0.005440628156065941, + 0.0024990993551909924, + 0.005206032190471888, + 0.0014659828739240766, + 0.0023545592557638884, + 0.002735992893576622, + 0.0047083655372262, + 0.0025081669446080923, + 0.0025465572252869606, + 0.005912225693464279, + 0.007493256125599146, + 0.0032593179494142532, + 0.0037533422000706196, + 0.020575307309627533, + 0.021693889051675797, + 0.002962706843391061, + 0.0017976717790588737, + 0.00283900648355484, + 0.003569434629753232, + 0.004404849838465452, + 0.004169958643615246, + 0.005539657082408667, + 0.003476034849882126, + 0.003853756235912442, + 0.0038366697262972593, + 0.01579696498811245, + 0.006055521313101053, + 0.006993837654590607, + 0.006990435067564249, + 0.019979456439614296, + 0.037572089582681656, + 0.011039982549846172, + 0.004555235616862774, + 0.004714039620012045, + 0.007003781385719776, + 0.00480625731870532, + 0.010349994525313377, + 0.01092365849763155, + 0.00899670459330082, + 0.008003135211765766, + 0.008556871674954891, + 0.009465817362070084, + 0.009216586127877235, + 0.01963675767183304, + 0.018267443403601646, + 0.00944641325622797, + 0.015807058662176132, + 0.00789345521479845, + 0.008232499472796917, + 0.011699947528541088, + 0.01291215606033802, + 0.010400675237178802, + 0.011750273406505585, + 0.011619829572737217, + 0.011143479496240616, + 0.011293117888271809, + 0.008408238179981709, + 0.06991101056337357, + 0.07779591530561447, + 0.10307269543409348, + 0.06616614013910294, + 0.027926644310355186, + 0.01899769715964794, + 0.03359745442867279, + 0.02172258123755455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.11292169988155365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979438751935959 + ], + "opposite_logprob_tokens": [ + -3.529794454574585 + ], + "logprob_reward": -0.02979438751935959, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979438751935959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036573994904756546, + 0.02715197205543518, + 0.043523844331502914, + 0.2451237589120865, + 0.02966778725385666, + 0.1521759331226349, + 0.034939851611852646, + 0.09373991191387177, + 0.08719759434461594, + 0.2499053180217743 + ], + "credit_map": [ + 0.003444135654717684, + 0.0052796825766563416, + 0.0033572150859981775, + 0.004368906375020742, + 0.004789299797266722, + 0.004519201815128326, + 0.00346173788420856, + 0.007353818044066429, + 0.0032305384520441294, + 0.0034484195057302713, + 0.0029252469539642334, + 0.0033181114122271538, + 0.004511092323809862, + 0.002710896311327815, + 0.0037014964036643505, + 0.0033061767462641, + 0.0039547281339764595, + 0.007910382933914661, + 0.0036661389749497175, + 0.005506520159542561, + 0.005808101501315832, + 0.006303082220256329, + 0.004789137281477451, + 0.005585760809481144, + 0.008233287371695042, + 0.019866548478603363, + 0.004831178579479456, + 0.004236606415361166, + 0.09786910563707352, + 0.10558510571718216, + 0.002155771479010582, + 0.0023461715318262577, + 0.003054595785215497, + 0.003468683222308755, + 0.004140703473240137, + 0.002097842050716281, + 0.004488882143050432, + 0.00493667833507061, + 0.004462650511413813, + 0.0030177542939782143, + 0.019587794318795204, + 0.0077598607167601585, + 0.009650981053709984, + 0.01165417954325676, + 0.032640133053064346, + 0.04895004257559776, + 0.014059020206332207, + 0.00787393283098936, + 0.0019487958634272218, + 0.003603624179959297, + 0.002352149458602071, + 0.005502042826265097, + 0.00480794720351696, + 0.004284797236323357, + 0.005798002239316702, + 0.0066424948163330555, + 0.014052602462470531, + 0.009013430215418339, + 0.023829996585845947, + 0.014420527964830399, + 0.006159127224236727, + 0.011933036148548126, + 0.007937964051961899, + 0.006393235642462969, + 0.01456359587609768, + 0.01032609399408102, + 0.014316467568278313, + 0.014683535322546959, + 0.006882137153297663, + 0.00614727009087801, + 0.013617672957479954, + 0.006660826969891787, + 0.08038904517889023, + 0.08788272738456726, + 0.021563740447163582, + 0.018865175545215607, + 0.014503980986773968, + 0.008785631507635117, + 0.010352245531976223, + 0.0075628128834068775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.02979438751935959 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017133570509031415 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017133570509031415, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017133570509031415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018084995448589325, + 0.01598210446536541, + 0.013620126061141491, + 0.1872817426919937, + 0.0120614655315876, + 0.01762620359659195, + 0.034304216504096985, + 0.1255398541688919, + 0.1007140502333641, + 0.4747852385044098 + ], + "credit_map": [ + 0.002405463717877865, + 0.001640072907321155, + 0.0026021753437817097, + 0.0027164157945662737, + 0.002195334527641535, + 0.0019687237218022346, + 0.002112611895427108, + 0.0024441967252641916, + 0.0014722690684720874, + 0.001665903371758759, + 0.0015553742414340377, + 0.0018604538636282086, + 0.0033879794646054506, + 0.0022111828438937664, + 0.0016714809462428093, + 0.002157460432499647, + 0.001351508661173284, + 0.0018117731669917703, + 0.000961727520916611, + 0.0013981565134599805, + 0.0021679361816495657, + 0.0024555251002311707, + 0.0013916789321228862, + 0.002081820974126458, + 0.005840526428073645, + 0.014836429618299007, + 0.0031841613817960024, + 0.0033739858772605658, + 0.07683540880680084, + 0.08033590763807297, + 0.001124919974245131, + 0.0017503828275948763, + 0.0013121983502060175, + 0.0013522246154025197, + 0.002089680638164282, + 0.0012147193774580956, + 0.0014513226924464107, + 0.0021297456696629524, + 0.001679249107837677, + 0.0008323254296556115, + 0.0014714343706145883, + 0.0026703020557761192, + 0.0010185620049014688, + 0.0023595215752720833, + 0.00418523233383894, + 0.0031383587047457695, + 0.0017032480100169778, + 0.0010795448906719685, + 0.0020881551317870617, + 0.0034227599389851093, + 0.002742488868534565, + 0.005493778269737959, + 0.005008373409509659, + 0.004056528210639954, + 0.005168117117136717, + 0.006324016489088535, + 0.020156152546405792, + 0.011962376534938812, + 0.03882401064038277, + 0.018700366839766502, + 0.009550297632813454, + 0.008268503472208977, + 0.010196144692599773, + 0.007881999015808105, + 0.015252347104251385, + 0.013049015775322914, + 0.014763987623155117, + 0.01837538368999958, + 0.008364037610590458, + 0.006204205099493265, + 0.01568007841706276, + 0.009024993516504765, + 0.1504613161087036, + 0.1703798472881317, + 0.051932379603385925, + 0.04608761891722679, + 0.017691293731331825, + 0.011202924884855747, + 0.017054220661520958, + 0.009975608438253403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.0017133570509031415 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, high-speed panning shot, held close to the ground, closely follows a dark green, modified 4x4, its body, fenders, and undercarriage coated in thick, dried mud. It's speeding down a sun-drenched, gravelly dirt road. As the vehicle violently cuts into a sharp curve, its rear wheels instantly lose grip, sending the rear end slinging outward at a precise and dramatic angle, creating a flawless power drift. In that instant, the massive, deeply treaded all-terrain tires, with astonishing force, fling a torrent of sand, dust, and small rocks backward and to the sides, forming a spectacular, opaque wall of dust. In the glaring afternoon sun, every particle of sand and mud is clearly visible, as if momentarily frozen in mid-air. The entire scene is imbued with a raw sense of power, speed, and passion, creating an atmosphere of intense excitement and adrenaline.", + "seed": 20346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_15/index_5994_seed_20346_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022100334381684662, + "vlm_logprob": -0.022100334381684662, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.022100334381684662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02344028279185295 + ], + "opposite_logprob_tokens": [ + -5.273440361022949 + ], + "logprob_reward": -0.02344028279185295, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.02344028279185295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2228034883737564, + 0.10458632558584213, + 0.14416496455669403, + 0.05350511893630028, + 0.07291339337825775, + 0.11349572241306305, + 0.05127131566405296, + 0.06461063027381897, + 0.08105099201202393, + 0.0915980190038681 + ], + "credit_map": [ + 0.01781441643834114, + 0.04267279803752899, + 0.026660291478037834, + 0.020240163430571556, + 0.0354105606675148, + 0.03848065808415413, + 0.026613980531692505, + 0.014910629950463772, + 0.006818524096161127, + 0.009791046380996704, + 0.009004679508507252, + 0.016629591584205627, + 0.01591944880783558, + 0.01925324834883213, + 0.016593778505921364, + 0.010576009750366211, + 0.007127561140805483, + 0.008246936835348606, + 0.008600001223385334, + 0.010285409167408943, + 0.009394781664013863, + 0.04730027914047241, + 0.012499519623816013, + 0.040710486471652985, + 0.003702730406075716, + 0.00458828080445528, + 0.0026502730324864388, + 0.0037214974872767925, + 0.0067257629707455635, + 0.014698106795549393, + 0.010858233086764812, + 0.006560237612575293, + 0.004046231042593718, + 0.005640844814479351, + 0.006033352576196194, + 0.00574576947838068, + 0.010016064159572124, + 0.010093232616782188, + 0.017852189019322395, + 0.013485709205269814, + 0.010172396898269653, + 0.0077559067867696285, + 0.010372942313551903, + 0.015313182957470417, + 0.0158496405929327, + 0.014698468148708344, + 0.015092351473867893, + 0.024240830913186073, + 0.007109176833182573, + 0.00924581941217184, + 0.00417428370565176, + 0.0051748622208833694, + 0.009519558399915695, + 0.004362476523965597, + 0.005910086911171675, + 0.005775052588433027, + 0.004458438605070114, + 0.005814034026116133, + 0.008133716881275177, + 0.011701919138431549, + 0.007485829293727875, + 0.0072771538980305195, + 0.010797475464642048, + 0.008942068554461002, + 0.00828500185161829, + 0.011324435472488403, + 0.008586290292441845, + 0.008888048119843006, + 0.011905195191502571, + 0.012610438279807568, + 0.010170337744057178, + 0.00928124226629734, + 0.010964476503431797, + 0.011231611482799053, + 0.015965253114700317, + 0.016548139974474907, + 0.008420472033321857, + 0.008703372441232204, + 0.011099726893007755, + 0.00866497028619051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.02344028279185295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331451140344143 + ], + "opposite_logprob_tokens": [ + -3.7733144760131836 + ], + "logprob_reward": -0.02331451140344143, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331451140344143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20661291480064392, + 0.06968428194522858, + 0.2694402039051056, + 0.038711078464984894, + 0.1272154599428177, + 0.025730257853865623, + 0.09641043096780777, + 0.03768814727663994, + 0.021972307935357094, + 0.1065349131822586 + ], + "credit_map": [ + 0.015220163390040398, + 0.04650178924202919, + 0.022566799074411392, + 0.018852895125746727, + 0.03420679271221161, + 0.034154366701841354, + 0.024136975407600403, + 0.010973140597343445, + 0.003618858056142926, + 0.008858081884682178, + 0.004414378199726343, + 0.0049492665566504, + 0.01295080129057169, + 0.01689761132001877, + 0.011517258360981941, + 0.0064780330285429955, + 0.011891013942658901, + 0.024790648370981216, + 0.010748517699539661, + 0.01635914482176304, + 0.01857049949467182, + 0.08866890519857407, + 0.016137976199388504, + 0.08227351307868958, + 0.0031522936187684536, + 0.0036090246867388487, + 0.002604509936645627, + 0.002718941541388631, + 0.005388939753174782, + 0.009658455848693848, + 0.007693808060139418, + 0.003885104088112712, + 0.009039022028446198, + 0.012384123168885708, + 0.008989819325506687, + 0.010915892198681831, + 0.021868135780096054, + 0.016166236251592636, + 0.029687827453017235, + 0.01816440373659134, + 0.0020245425403118134, + 0.0032442077063024044, + 0.0020849942229688168, + 0.004144364036619663, + 0.0037770492490381002, + 0.004495103377848864, + 0.003215531585738063, + 0.0027444660663604736, + 0.008659996092319489, + 0.02221016399562359, + 0.00886599812656641, + 0.009569918736815453, + 0.012954751029610634, + 0.008097031153738499, + 0.013534336350858212, + 0.012518235482275486, + 0.004634414799511433, + 0.005063607357442379, + 0.004586329218000174, + 0.0034488097298890352, + 0.004946489818394184, + 0.0036972351372241974, + 0.005656974855810404, + 0.005654284730553627, + 0.003851865651085973, + 0.004275224171578884, + 0.0021501139272004366, + 0.002330298302695155, + 0.001978987827897072, + 0.0025137963239103556, + 0.002528693061321974, + 0.0023433277383446693, + 0.009660718962550163, + 0.0110819386318326, + 0.01260234136134386, + 0.015472102910280228, + 0.005796307232230902, + 0.007283649407327175, + 0.026397479698061943, + 0.018240366131067276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.02331451140344143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067626655101776 + ], + "opposite_logprob_tokens": [ + -3.168067693710327 + ], + "logprob_reward": -0.043067626655101776, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067626655101776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29655662178993225, + 0.05183129012584686, + 0.11357073485851288, + 0.11528557538986206, + 0.08666016161441803, + 0.09338407218456268, + 0.05387425050139427, + 0.04936810955405235, + 0.05336898937821388, + 0.0861002504825592 + ], + "credit_map": [ + 0.02192878909409046, + 0.07126850634813309, + 0.026691824197769165, + 0.023283857852220535, + 0.05538438633084297, + 0.05226365476846695, + 0.03258686512708664, + 0.01314869336783886, + 0.003794710850343108, + 0.009072274900972843, + 0.003671303391456604, + 0.004728635773062706, + 0.00837541464716196, + 0.007741931360214949, + 0.00852829497307539, + 0.005918717011809349, + 0.0070304106920957565, + 0.012917214073240757, + 0.006547595839947462, + 0.008919650688767433, + 0.01108110323548317, + 0.025241246446967125, + 0.018137920647859573, + 0.023695580661296844, + 0.007205659989267588, + 0.009353982284665108, + 0.005503639113157988, + 0.00580511474981904, + 0.014027578756213188, + 0.03454527258872986, + 0.02596721053123474, + 0.012877107597887516, + 0.003914420492947102, + 0.004991015885025263, + 0.003434196813032031, + 0.005203156266361475, + 0.012037898413836956, + 0.013720673508942127, + 0.02765577845275402, + 0.01570300944149494, + 0.006525433622300625, + 0.01168739516288042, + 0.007120131514966488, + 0.014450501650571823, + 0.012684565037488937, + 0.013027374632656574, + 0.012739893980324268, + 0.01514875702559948, + 0.0065866983495652676, + 0.011600843630731106, + 0.004737424198538065, + 0.004793023690581322, + 0.009652804583311081, + 0.005144915077835321, + 0.006403587758541107, + 0.004954943433403969, + 0.004206511657685041, + 0.0057651447132229805, + 0.0060364799574017525, + 0.007328014355152845, + 0.0061098141595721245, + 0.005660290829837322, + 0.007973751984536648, + 0.006288100033998489, + 0.003968470264226198, + 0.0045051672495901585, + 0.005140476394444704, + 0.006251666694879532, + 0.00884940568357706, + 0.010660532861948013, + 0.007853449322283268, + 0.00613981019705534, + 0.006740040611475706, + 0.008223328739404678, + 0.009363394230604172, + 0.011961973272264004, + 0.005630519241094589, + 0.0062834336422383785, + 0.022004442289471626, + 0.015893105417490005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.043067626655101776 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818707026541233 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818707026541233, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818707026541233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18589772284030914, + 0.05740534886717796, + 0.16948546469211578, + 0.09423942118883133, + 0.07364081591367722, + 0.08625193685293198, + 0.08399726450443268, + 0.06260375678539276, + 0.05233016610145569, + 0.1341480165719986 + ], + "credit_map": [ + 0.012483001686632633, + 0.020835086703300476, + 0.04029519855976105, + 0.027454117313027382, + 0.023282770067453384, + 0.023673469200730324, + 0.022447124123573303, + 0.015426963567733765, + 0.00682837376371026, + 0.009649690240621567, + 0.005488638300448656, + 0.0052119954489171505, + 0.008949106559157372, + 0.0061033000238239765, + 0.008310371078550816, + 0.006863870657980442, + 0.007323739118874073, + 0.010226943530142307, + 0.014539891853928566, + 0.012804885394871235, + 0.009860084392130375, + 0.044785428792238235, + 0.02723793499171734, + 0.04270656406879425, + 0.009120087139308453, + 0.009135962463915348, + 0.005557588301599026, + 0.006551696453243494, + 0.016962768509984016, + 0.01935754157602787, + 0.016639985144138336, + 0.010913786478340626, + 0.005673583131283522, + 0.006140054669231176, + 0.004508183337748051, + 0.006816614419221878, + 0.012292888946831226, + 0.010935146361589432, + 0.016870714724063873, + 0.010403627529740334, + 0.008063305169343948, + 0.005824565421789885, + 0.008095962926745415, + 0.012295686639845371, + 0.011619649827480316, + 0.00907540787011385, + 0.013194592669606209, + 0.018082763999700546, + 0.0067939069122076035, + 0.010020114481449127, + 0.007107091136276722, + 0.007047706283628941, + 0.02157249115407467, + 0.008218259550631046, + 0.012733826413750648, + 0.010503864847123623, + 0.007385624572634697, + 0.006910078693181276, + 0.01031225360929966, + 0.009656466543674469, + 0.00662168487906456, + 0.006795470602810383, + 0.005917150992900133, + 0.009005026891827583, + 0.0036834680940955877, + 0.005766082089394331, + 0.004697785247117281, + 0.0053584096021950245, + 0.008686979301273823, + 0.011789761483669281, + 0.007481983397156, + 0.004865697119385004, + 0.012776731513440609, + 0.01586451753973961, + 0.01640625111758709, + 0.01783820241689682, + 0.009936705231666565, + 0.011629956774413586, + 0.028987763449549675, + 0.020707882940769196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.01818707026541233 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024921807926148176 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.0024921807926148176, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024921807926148176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2693234384059906, + 0.046115677803754807, + 0.07765284180641174, + 0.07387559860944748, + 0.08097638934850693, + 0.1507469117641449, + 0.06184215843677521, + 0.06927080452442169, + 0.07794801145792007, + 0.09224820137023926 + ], + "credit_map": [ + 0.021915540099143982, + 0.0647352859377861, + 0.0183238722383976, + 0.0196640994399786, + 0.049655526876449585, + 0.04574539139866829, + 0.03566348925232887, + 0.013620225712656975, + 0.003446137299761176, + 0.005866677965968847, + 0.005081428214907646, + 0.0033133849501609802, + 0.007541390135884285, + 0.00945921428501606, + 0.006287646014243364, + 0.00511979591101408, + 0.005775423254817724, + 0.010982050560414791, + 0.00586383743211627, + 0.008708767592906952, + 0.00828458834439516, + 0.01332109235227108, + 0.012502341531217098, + 0.01221474725753069, + 0.005383441224694252, + 0.011468679644167423, + 0.0057207574136555195, + 0.00621769530698657, + 0.016020145267248154, + 0.009322654455900192, + 0.007925072684884071, + 0.01181715726852417, + 0.0038583644200116396, + 0.005634426139295101, + 0.004483015742152929, + 0.005992535501718521, + 0.012818410992622375, + 0.011693358421325684, + 0.02077404223382473, + 0.015722228214144707, + 0.010845726355910301, + 0.010812549851834774, + 0.011463052593171597, + 0.019234634935855865, + 0.014705706387758255, + 0.01994754560291767, + 0.027106665074825287, + 0.0366310253739357, + 0.007329321466386318, + 0.015219061635434628, + 0.004676404874771833, + 0.005474331323057413, + 0.008839610032737255, + 0.005944627337157726, + 0.008275010623037815, + 0.006083788350224495, + 0.004363915417343378, + 0.007003150414675474, + 0.006812816485762596, + 0.007790520321577787, + 0.005281299818307161, + 0.012409091927111149, + 0.009156141430139542, + 0.016453873366117477, + 0.007867753505706787, + 0.0115179643034935, + 0.008101429790258408, + 0.0095702288672328, + 0.009150621481239796, + 0.014002524316310883, + 0.01007378101348877, + 0.007663700729608536, + 0.0069837202318012714, + 0.011354897171258926, + 0.016305116936564445, + 0.014998866245150566, + 0.01107828039675951, + 0.010544958524405956, + 0.010886377654969692, + 0.010095981881022453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0024921807926148176 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, high-speed panning shot, held close to the ground, closely follows a dark green, modified 4x4, its body, fenders, and undercarriage coated in thick, dried mud. It's speeding down a sun-drenched, gravelly dirt road. As the vehicle violently cuts into a sharp curve, its rear wheels instantly lose grip, sending the rear end slinging outward at a precise and dramatic angle, creating a flawless power drift. In that instant, the massive, deeply treaded all-terrain tires, with astonishing force, fling a torrent of sand, dust, and small rocks backward and to the sides, forming a spectacular, opaque wall of dust. In the glaring afternoon sun, every particle of sand and mud is clearly visible, as if momentarily frozen in mid-air. The entire scene is imbued with a raw sense of power, speed, and passion, creating an atmosphere of intense excitement and adrenaline.", + "seed": 20347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_15/index_5994_seed_20347_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020459040813148022, + "vlm_logprob": -0.020459040813148022, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.020459040813148022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03687914088368416 + ], + "opposite_logprob_tokens": [ + -5.786879062652588 + ], + "logprob_reward": -0.03687914088368416, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.03687914088368416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07588467746973038, + 0.06132837384939194, + 0.04771694168448448, + 0.1004665270447731, + 0.10892584174871445, + 0.0791126936674118, + 0.10988741368055344, + 0.05024425685405731, + 0.19251780211925507, + 0.17391549050807953 + ], + "credit_map": [ + 0.00605737091973424, + 0.009188137948513031, + 0.006708627101033926, + 0.008711863309144974, + 0.009462359361350536, + 0.009306807070970535, + 0.015882549807429314, + 0.01056696381419897, + 0.010626344941556454, + 0.005798630882054567, + 0.010819267481565475, + 0.007372010033577681, + 0.004972479306161404, + 0.008824320510029793, + 0.004272456746548414, + 0.008642863482236862, + 0.005358981434255838, + 0.005805308930575848, + 0.00689998222514987, + 0.00558168301358819, + 0.005596923176199198, + 0.010055610910058022, + 0.004710034932941198, + 0.0037084179930388927, + 0.01481736171990633, + 0.017735840752720833, + 0.015664881095290184, + 0.012220188975334167, + 0.015181419439613819, + 0.010791871696710587, + 0.004338267259299755, + 0.009716697968542576, + 0.008830763399600983, + 0.0196013655513525, + 0.009356403723359108, + 0.014322451315820217, + 0.014570414088666439, + 0.023167293518781662, + 0.007439864799380302, + 0.011637283489108086, + 0.012287401594221592, + 0.00626484164968133, + 0.014095005579292774, + 0.007084476295858622, + 0.009735078550875187, + 0.006997061427682638, + 0.0037471242249011993, + 0.018901707604527473, + 0.00980860274285078, + 0.01588025502860546, + 0.015187948010861874, + 0.01301812194287777, + 0.014577255584299564, + 0.01328759454190731, + 0.0162370502948761, + 0.011890580877661705, + 0.003174437675625086, + 0.0055727483704686165, + 0.003935006447136402, + 0.008325645700097084, + 0.007399252150207758, + 0.007942155003547668, + 0.007344604004174471, + 0.006550407037138939, + 0.024063754826784134, + 0.01840265654027462, + 0.03393552079796791, + 0.02046375907957554, + 0.018164025619626045, + 0.03075757436454296, + 0.018216056749224663, + 0.02851446159183979, + 0.014416133053600788, + 0.02216224931180477, + 0.013101178221404552, + 0.014807896688580513, + 0.03521810099482536, + 0.02690916694700718, + 0.024744262918829918, + 0.0225564893335104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.03687914088368416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009804191067814827 + ], + "opposite_logprob_tokens": [ + -4.6348042488098145 + ], + "logprob_reward": -0.009804191067814827, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009804191067814827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1341651976108551, + 0.06286804378032684, + 0.03183157369494438, + 0.04862194135785103, + 0.03134462609887123, + 0.14347527921199799, + 0.11692845076322556, + 0.12862393260002136, + 0.22023984789848328, + 0.08190108835697174 + ], + "credit_map": [ + 0.010857506655156612, + 0.013040902093052864, + 0.010573525913059711, + 0.012191478163003922, + 0.02044019103050232, + 0.013017271645367146, + 0.03474929928779602, + 0.01929502747952938, + 0.0093154925853014, + 0.007659600581973791, + 0.010178469121456146, + 0.005096041597425938, + 0.006794479675590992, + 0.005541322287172079, + 0.007615943439304829, + 0.010666695423424244, + 0.00475283432751894, + 0.003478318452835083, + 0.005858269985765219, + 0.0031369004864245653, + 0.003650758182629943, + 0.005953194573521614, + 0.0025426517240703106, + 0.002458646660670638, + 0.009612186811864376, + 0.006269735749810934, + 0.009830194525420666, + 0.005688752047717571, + 0.005609029904007912, + 0.005090540274977684, + 0.0026176441460847855, + 0.003903862088918686, + 0.003479905426502228, + 0.003438348416239023, + 0.0028151555452495813, + 0.0039509921334683895, + 0.003171867225319147, + 0.00816283281892538, + 0.001756687299348414, + 0.004568838514387608, + 0.01929527148604393, + 0.014031732454895973, + 0.02183653973042965, + 0.01309578213840723, + 0.018353983759880066, + 0.015156819485127926, + 0.008543575182557106, + 0.033161576837301254, + 0.011828206479549408, + 0.01307107787579298, + 0.009375103749334812, + 0.014308285899460316, + 0.014454714953899384, + 0.008586854673922062, + 0.028496915474534035, + 0.016807302832603455, + 0.010305236093699932, + 0.008136754855513573, + 0.017200680449604988, + 0.015055285766720772, + 0.017050936818122864, + 0.034058503806591034, + 0.0070201498456299305, + 0.01979638636112213, + 0.02745058760046959, + 0.022253556177020073, + 0.038175951689481735, + 0.024224398657679558, + 0.024967411532998085, + 0.03464629501104355, + 0.01580977253615856, + 0.03271188959479332, + 0.007903959602117538, + 0.0120781771838665, + 0.011670038104057312, + 0.00767571059986949, + 0.012164685875177383, + 0.00953701976686716, + 0.011249451898038387, + 0.009622045792639256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.009804191067814827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798785224556923 + ], + "opposite_logprob_tokens": [ + -3.529798746109009 + ], + "logprob_reward": -0.029798785224556923, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798785224556923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28879523277282715, + 0.06746149063110352, + 0.04729330539703369, + 0.04813389107584953, + 0.0496382862329483, + 0.02739732153713703, + 0.18232361972332, + 0.03480685129761696, + 0.18335242569446564, + 0.07079759985208511 + ], + "credit_map": [ + 0.014849296770989895, + 0.019548587501049042, + 0.018091242760419846, + 0.024870531633496284, + 0.03287861868739128, + 0.03315670043230057, + 0.09258696436882019, + 0.05281328409910202, + 0.0206668172031641, + 0.003955462481826544, + 0.024838725104928017, + 0.003441046690568328, + 0.0033041839487850666, + 0.0021918255370110273, + 0.004014479927718639, + 0.005048946477472782, + 0.008039434440433979, + 0.005986466538161039, + 0.007487266790121794, + 0.006486316677182913, + 0.0048021613620221615, + 0.005174993071705103, + 0.0035324180498719215, + 0.005784243810921907, + 0.006680138409137726, + 0.009143699891865253, + 0.006071269046515226, + 0.006605398375540972, + 0.0064252158626914024, + 0.005138522479683161, + 0.002859722124412656, + 0.005209924653172493, + 0.0045204502530395985, + 0.006083112210035324, + 0.005835088435560465, + 0.004832711536437273, + 0.004837144631892443, + 0.011900319717824459, + 0.003335660556331277, + 0.008293798193335533, + 0.003192018950358033, + 0.0028104500379413366, + 0.004875366110354662, + 0.0026941311080008745, + 0.005151628982275724, + 0.00328140240162611, + 0.002168782753869891, + 0.0032235397957265377, + 0.011637356132268906, + 0.013834814541041851, + 0.018025847151875496, + 0.019045043736696243, + 0.016298439353704453, + 0.02248590812087059, + 0.03979584574699402, + 0.04120037332177162, + 0.002788671525195241, + 0.002838475862517953, + 0.002879398176446557, + 0.004245278891175985, + 0.004486194346100092, + 0.008615564554929733, + 0.003135360311716795, + 0.005817906931042671, + 0.028016211465001106, + 0.020823940634727478, + 0.03650706633925438, + 0.020196935161948204, + 0.02180572971701622, + 0.0206484142690897, + 0.014902069233357906, + 0.020452052354812622, + 0.0060912370681762695, + 0.010217471048235893, + 0.010157449170947075, + 0.005225758068263531, + 0.013884198851883411, + 0.008440776728093624, + 0.010711181908845901, + 0.0060695260763168335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.029798785224556923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184728920459747 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184728920459747, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184728920459747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10775724798440933, + 0.06231819465756416, + 0.1283782422542572, + 0.04274315759539604, + 0.09533195197582245, + 0.04992229864001274, + 0.26297876238822937, + 0.03337807208299637, + 0.1411399096250534, + 0.07605218142271042 + ], + "credit_map": [ + 0.009153291583061218, + 0.007912510074675083, + 0.009425129741430283, + 0.008081942796707153, + 0.011472157202661037, + 0.00922626443207264, + 0.03361599147319794, + 0.018869953230023384, + 0.01707194373011589, + 0.004223678261041641, + 0.02195635810494423, + 0.0033030996564775705, + 0.003422811394557357, + 0.00441042659804225, + 0.002044051419943571, + 0.005885825026780367, + 0.01716800220310688, + 0.013877660036087036, + 0.020119182765483856, + 0.021645953878760338, + 0.01512068510055542, + 0.017515994608402252, + 0.011792058125138283, + 0.011138697154819965, + 0.006904467940330505, + 0.0052256472408771515, + 0.007885275408625603, + 0.005000975914299488, + 0.00517637375742197, + 0.004710604902356863, + 0.002483995631337166, + 0.005355814937502146, + 0.00883909035474062, + 0.014854676090180874, + 0.007187069393694401, + 0.011984089389443398, + 0.012290572747588158, + 0.02314736135303974, + 0.006447098683565855, + 0.010581991635262966, + 0.0049585518427193165, + 0.006004648748785257, + 0.010186780244112015, + 0.005718799773603678, + 0.005349011160433292, + 0.005060352385044098, + 0.003229161724448204, + 0.009414993226528168, + 0.020794367417693138, + 0.02500828169286251, + 0.04262964800000191, + 0.02591930888593197, + 0.03721049055457115, + 0.0336703360080719, + 0.05642300099134445, + 0.021323340013623238, + 0.0024689012207090855, + 0.0023790423292666674, + 0.004307591821998358, + 0.003947757184505463, + 0.00432957150042057, + 0.00853851716965437, + 0.0026630908250808716, + 0.004743595607578754, + 0.014236231334507465, + 0.01241222582757473, + 0.021472174674272537, + 0.0166762787848711, + 0.014521409757435322, + 0.028041046112775803, + 0.012618906795978546, + 0.021161632612347603, + 0.0062282392755150795, + 0.008332767523825169, + 0.011158506385982037, + 0.007858320139348507, + 0.010523932985961437, + 0.009595142677426338, + 0.010016340762376785, + 0.012338930740952492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.018184728920459747 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628357969224453 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628357969224453, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628357969224453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08092152327299118, + 0.05248921364545822, + 0.04766715690493584, + 0.029088139533996582, + 0.008588043041527271, + 0.028755569830536842, + 0.039828211069107056, + 0.06156480684876442, + 0.4561989903450012, + 0.1948983371257782 + ], + "credit_map": [ + 0.006506482604891062, + 0.007479080930352211, + 0.007350353989750147, + 0.007209511008113623, + 0.009334697388112545, + 0.009334881789982319, + 0.022119218483567238, + 0.011587303131818771, + 0.010836376808583736, + 0.003395745065063238, + 0.013631198555231094, + 0.006281078793108463, + 0.00266797817312181, + 0.004895494785159826, + 0.003793434938415885, + 0.0069879102520644665, + 0.00548697542399168, + 0.005157096311450005, + 0.006019249092787504, + 0.008409328758716583, + 0.005552012473344803, + 0.007390983868390322, + 0.005690572317689657, + 0.00396093912422657, + 0.004401468206197023, + 0.0034399928990751505, + 0.0046341922134160995, + 0.004771222360432148, + 0.004900674801319838, + 0.0026152089703828096, + 0.0017380142817273736, + 0.002587367780506611, + 0.0007308151689358056, + 0.0013176655629649758, + 0.0008287019445560873, + 0.0008420268422923982, + 0.00116631924174726, + 0.0017543910071253777, + 0.00046377736725844443, + 0.0014843456447124481, + 0.004491142462939024, + 0.003063174430280924, + 0.0037234441842883825, + 0.0029257393907755613, + 0.0037004761397838593, + 0.002546807052567601, + 0.0017143839504569769, + 0.006590403150767088, + 0.0032104842830449343, + 0.0057657272554934025, + 0.003515403252094984, + 0.005600475240498781, + 0.004971007816493511, + 0.003863946069031954, + 0.009143464267253876, + 0.003757702885195613, + 0.004659091588109732, + 0.0034738383255898952, + 0.00812561996281147, + 0.008058246225118637, + 0.008748038671910763, + 0.016922542825341225, + 0.0029659192077815533, + 0.008611513301730156, + 0.07451646775007248, + 0.059609994292259216, + 0.09409830719232559, + 0.046452175825834274, + 0.042329102754592896, + 0.05001145973801613, + 0.0382864736020565, + 0.050895001739263535, + 0.015567447990179062, + 0.02333431877195835, + 0.02897140197455883, + 0.02366619184613228, + 0.02293894812464714, + 0.022473491728305817, + 0.025875188410282135, + 0.03207135573029518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.007628357969224453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, high-speed panning shot, held close to the ground, closely follows a dark green, modified 4x4, its body, fenders, and undercarriage coated in thick, dried mud. It's speeding down a sun-drenched, gravelly dirt road. As the vehicle violently cuts into a sharp curve, its rear wheels instantly lose grip, sending the rear end slinging outward at a precise and dramatic angle, creating a flawless power drift. In that instant, the massive, deeply treaded all-terrain tires, with astonishing force, fling a torrent of sand, dust, and small rocks backward and to the sides, forming a spectacular, opaque wall of dust. In the glaring afternoon sun, every particle of sand and mud is clearly visible, as if momentarily frozen in mid-air. The entire scene is imbued with a raw sense of power, speed, and passion, creating an atmosphere of intense excitement and adrenaline.", + "seed": 20348, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_15/index_5994_seed_20348_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04980811905115843, + "vlm_logprob": -0.04980811905115843, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.04980811905115843, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033388350158929825 + ], + "opposite_logprob_tokens": [ + -5.658388137817383 + ], + "logprob_reward": -0.033388350158929825, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.033388350158929825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16734382510185242, + 0.08910206705331802, + 0.0800199955701828, + 0.05489393323659897, + 0.0343623086810112, + 0.1738128811120987, + 0.07047507911920547, + 0.09711644053459167, + 0.19094176590442657, + 0.0419316366314888 + ], + "credit_map": [ + 0.014306080527603626, + 0.013006885536015034, + 0.0312536284327507, + 0.030006691813468933, + 0.013734656386077404, + 0.010419062338769436, + 0.015522070229053497, + 0.03909475728869438, + 0.009164088405668736, + 0.012058716267347336, + 0.022860873490571976, + 0.010849816724658012, + 0.004556361585855484, + 0.005635765381157398, + 0.005574772600084543, + 0.018401674926280975, + 0.006883981171995401, + 0.011661657132208347, + 0.012946098111569881, + 0.013509208336472511, + 0.004551953636109829, + 0.004183973651379347, + 0.014934327453374863, + 0.01134879793971777, + 0.007425062358379364, + 0.010055734775960445, + 0.006791840773075819, + 0.0065945847891271114, + 0.007663581054657698, + 0.004808672703802586, + 0.004610020667314529, + 0.0069444384425878525, + 0.004524542484432459, + 0.0053307791240513325, + 0.0042573316022753716, + 0.00422264076769352, + 0.0037666044663637877, + 0.004149711225181818, + 0.003892855951562524, + 0.004217840265482664, + 0.016399135813117027, + 0.014059513807296753, + 0.012107877060770988, + 0.04550508037209511, + 0.01837191917002201, + 0.024021901190280914, + 0.02938910387456417, + 0.013958354480564594, + 0.004123129416257143, + 0.01247604563832283, + 0.005240154918283224, + 0.003551554400473833, + 0.018253512680530548, + 0.005714258644729853, + 0.008584961295127869, + 0.012531455606222153, + 0.005219855345785618, + 0.005953064188361168, + 0.005180736072361469, + 0.006081705912947655, + 0.007105949800461531, + 0.007978403009474277, + 0.03288924694061279, + 0.026707470417022705, + 0.020768430083990097, + 0.0247208159416914, + 0.020815294235944748, + 0.0189963411539793, + 0.029870908707380295, + 0.012426836416125298, + 0.02472643367946148, + 0.038616713136434555, + 0.005011728033423424, + 0.005315059795975685, + 0.006098275538533926, + 0.005460233427584171, + 0.006395512726157904, + 0.004168170969933271, + 0.0036229793913662434, + 0.005859680473804474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.033388350158929825 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559250928461552 + ], + "opposite_logprob_tokens": [ + -4.387559413909912 + ], + "logprob_reward": -0.012559250928461552, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559250928461552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10704375058412552, + 0.10576004534959793, + 0.029681673273444176, + 0.04497120529413223, + 0.03268447145819664, + 0.08008512854576111, + 0.049915723502635956, + 0.2244776338338852, + 0.29332900047302246, + 0.032051365822553635 + ], + "credit_map": [ + 0.009454445913434029, + 0.01530629675835371, + 0.015425743535161018, + 0.023001739755272865, + 0.01153868529945612, + 0.007269255351275206, + 0.011902273632586002, + 0.013145317323505878, + 0.014729533344507217, + 0.0056062606163322926, + 0.038394127041101456, + 0.012867542915046215, + 0.003107151947915554, + 0.0034903339110314846, + 0.004683486185967922, + 0.022881608456373215, + 0.004070042632520199, + 0.005030189175158739, + 0.003183182096108794, + 0.004061768762767315, + 0.002609232673421502, + 0.0022954263258725405, + 0.003546342719346285, + 0.0048854900524020195, + 0.0037982596550136805, + 0.008052180521190166, + 0.005278066731989384, + 0.005218553822487593, + 0.009920633397996426, + 0.00421810382977128, + 0.003563945647329092, + 0.0049214609898626804, + 0.003042419906705618, + 0.003497289028018713, + 0.004301226697862148, + 0.004386222455650568, + 0.003451635828241706, + 0.0038561797700822353, + 0.005617927294224501, + 0.004531568847596645, + 0.005517303943634033, + 0.005650690756738186, + 0.005061769858002663, + 0.009373064152896404, + 0.015764987096190453, + 0.013558980077505112, + 0.014829805120825768, + 0.010328524746000767, + 0.003334092441946268, + 0.0074140592478215694, + 0.003799622878432274, + 0.002855609869584441, + 0.009450024925172329, + 0.004492480773478746, + 0.007543652784079313, + 0.011026178486645222, + 0.010827946476638317, + 0.011880025267601013, + 0.007934845983982086, + 0.012747677974402905, + 0.011540541425347328, + 0.013747069984674454, + 0.09039802849292755, + 0.06540150195360184, + 0.03135285526514053, + 0.039796382188797, + 0.031190648674964905, + 0.029219338670372963, + 0.04433578997850418, + 0.020679891109466553, + 0.03737742081284523, + 0.05937664583325386, + 0.004883972927927971, + 0.00305753736756742, + 0.003741408232599497, + 0.0038702005986124277, + 0.004589284770190716, + 0.00282680569216609, + 0.00386896263808012, + 0.005213193595409393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.012559250928461552 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035683453083038 + ], + "opposite_logprob_tokens": [ + -1.910356879234314 + ], + "logprob_reward": -0.16035683453083038, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035683453083038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14901573956012726, + 0.07993762195110321, + 0.1494673639535904, + 0.047285281121730804, + 0.03849097713828087, + 0.13431496918201447, + 0.10646841675043106, + 0.09130390733480453, + 0.16600969433784485, + 0.037706032395362854 + ], + "credit_map": [ + 0.01569795422255993, + 0.022883255034685135, + 0.023724675178527832, + 0.02719959244132042, + 0.01659189909696579, + 0.01002301275730133, + 0.015631411224603653, + 0.01726393587887287, + 0.008515169844031334, + 0.006501221563667059, + 0.02407492697238922, + 0.008762582205235958, + 0.0038455871399492025, + 0.00400649756193161, + 0.005171682219952345, + 0.019059954211115837, + 0.018951864913105965, + 0.012990700080990791, + 0.02136087976396084, + 0.02266284078359604, + 0.010762715712189674, + 0.009796381928026676, + 0.02312561310827732, + 0.029816368594765663, + 0.005488554481416941, + 0.0076620085164904594, + 0.00522109167650342, + 0.00665119057521224, + 0.006904973182827234, + 0.0044319950975477695, + 0.005301752593368292, + 0.005623712204396725, + 0.00805036909878254, + 0.004222256597131491, + 0.004879906307905912, + 0.004140770062804222, + 0.004575079772621393, + 0.004431819077581167, + 0.004106770269572735, + 0.0040840040892362595, + 0.006943955551832914, + 0.006805916782468557, + 0.006775818299502134, + 0.029066409915685654, + 0.012173461727797985, + 0.015882134437561035, + 0.021822189912199974, + 0.03484507277607918, + 0.006307393778115511, + 0.01896747201681137, + 0.005864291451871395, + 0.004421865101903677, + 0.025052279233932495, + 0.008695847354829311, + 0.014111029915511608, + 0.023048238828778267, + 0.006286787800490856, + 0.008226568810641766, + 0.005938332993537188, + 0.004814547020941973, + 0.008639843203127384, + 0.009844079613685608, + 0.027298076078295708, + 0.020255664363503456, + 0.018139895051717758, + 0.024916687980294228, + 0.016864027827978134, + 0.015372827649116516, + 0.023920530453324318, + 0.011373411864042282, + 0.022294700145721436, + 0.03312760591506958, + 0.003998471889644861, + 0.00353398104198277, + 0.005761345848441124, + 0.0031080807093530893, + 0.004452565219253302, + 0.0029513216577470303, + 0.00477313669398427, + 0.009127126075327396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.16035683453083038 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809741511940956 + ], + "opposite_logprob_tokens": [ + -3.288097381591797 + ], + "logprob_reward": -0.03809741511940956, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809741511940956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10187587887048721, + 0.08164525777101517, + 0.0576043464243412, + 0.061676327139139175, + 0.062429554760456085, + 0.1696529984474182, + 0.07765413075685501, + 0.17172540724277496, + 0.15050989389419556, + 0.06522613763809204 + ], + "credit_map": [ + 0.009399264119565487, + 0.014176551252603531, + 0.015302315354347229, + 0.019345475360751152, + 0.014305859804153442, + 0.007792554330080748, + 0.010796025395393372, + 0.010757839307188988, + 0.009246158413589, + 0.005683659575879574, + 0.02777869999408722, + 0.009743344970047474, + 0.0032333123963326216, + 0.0035621689166873693, + 0.004040950909256935, + 0.018356969580054283, + 0.006641009356826544, + 0.006535983644425869, + 0.008325733244419098, + 0.011720332317054272, + 0.004941619001328945, + 0.003994755446910858, + 0.007643349003046751, + 0.007801568601280451, + 0.00738149881362915, + 0.00854488555341959, + 0.009759338572621346, + 0.005898292642086744, + 0.010243027471005917, + 0.007238057907670736, + 0.006283184979110956, + 0.006328045390546322, + 0.008629069663584232, + 0.00775170186534524, + 0.007503175642341375, + 0.006772063672542572, + 0.006080951541662216, + 0.007039301097393036, + 0.009690545499324799, + 0.00896274670958519, + 0.01067768968641758, + 0.011386040598154068, + 0.012248503975570202, + 0.03532690554857254, + 0.01856851391494274, + 0.019306130707263947, + 0.0240121278911829, + 0.0381271131336689, + 0.0049308049492537975, + 0.011669077910482883, + 0.0046020736917853355, + 0.0035239080898463726, + 0.01671517640352249, + 0.005796484183520079, + 0.011960630305111408, + 0.018455982208251953, + 0.011273175477981567, + 0.016885090619325638, + 0.011423954740166664, + 0.009329680353403091, + 0.018798833712935448, + 0.02323964051902294, + 0.04619718715548515, + 0.034577857702970505, + 0.01688447967171669, + 0.02207706682384014, + 0.015657376497983932, + 0.01358817145228386, + 0.02197667397558689, + 0.01057629007846117, + 0.019369564950466156, + 0.030380282551050186, + 0.008422101847827435, + 0.00648821284994483, + 0.006499130744487047, + 0.006803951226174831, + 0.005602338816970587, + 0.005650720559060574, + 0.006424034014344215, + 0.019335653632879257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.03809741511940956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, held close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a dark green modified 4x4 vehicle covered in thick dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the vehicle speeding down a sun-drenched gravelly dirt road and cutting into a sharp curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the rear wheels lose grip and sling the rear end outward at a precise angle to create a power drift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a torrent of sand, dust, and small rocks flung backward forming an opaque wall of dust with visible particles in the afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463874451816082 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.00463874451816082, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463874451816082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10759540647268295, + 0.10078639537096024, + 0.07598410546779633, + 0.032233063131570816, + 0.046895500272512436, + 0.023026354610919952, + 0.050395213067531586, + 0.2631504237651825, + 0.19967010617256165, + 0.10026340186595917 + ], + "credit_map": [ + 0.011872565373778343, + 0.01343357004225254, + 0.018685808405280113, + 0.017458753660321236, + 0.008175498805940151, + 0.009617254137992859, + 0.01222524605691433, + 0.016126707196235657, + 0.015358302742242813, + 0.010200225748121738, + 0.022747192531824112, + 0.019731050357222557, + 0.004805991426110268, + 0.004431893117725849, + 0.008485057391226292, + 0.015026681125164032, + 0.005092809442430735, + 0.009153560735285282, + 0.01287267729640007, + 0.01897125504910946, + 0.00420003104954958, + 0.002892829943448305, + 0.011639668606221676, + 0.011161274276673794, + 0.00358764105476439, + 0.005748906172811985, + 0.0035485050175338984, + 0.004666588734835386, + 0.004230281803756952, + 0.0038019756320863962, + 0.0037142459768801928, + 0.0029349159449338913, + 0.004904808942228556, + 0.0052453624084591866, + 0.00652617122977972, + 0.005883840844035149, + 0.005563567858189344, + 0.004999993368983269, + 0.007449822034686804, + 0.006321932189166546, + 0.0017683753976598382, + 0.0019013418350368738, + 0.0017061491962522268, + 0.003349320264533162, + 0.003323119832202792, + 0.0037148892879486084, + 0.003953265957534313, + 0.0033098936546593904, + 0.004446160979568958, + 0.008742174133658409, + 0.004818675573915243, + 0.0032384907826781273, + 0.010702710598707199, + 0.004751939792186022, + 0.007727636024355888, + 0.005967425648123026, + 0.013136772438883781, + 0.016377970576286316, + 0.011365052312612534, + 0.012430675327777863, + 0.017025334760546684, + 0.020533094182610512, + 0.09903386980295181, + 0.07324764877557755, + 0.020609647035598755, + 0.026847312226891518, + 0.020795198157429695, + 0.021501725539565086, + 0.032116807997226715, + 0.015640676021575928, + 0.023336855694651604, + 0.038821883499622345, + 0.01121822465211153, + 0.008575938642024994, + 0.01371957827359438, + 0.01664745807647705, + 0.0065750619396567345, + 0.0074409968219697475, + 0.014628920704126358, + 0.02145722508430481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.00463874451816082 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_tensor([2108, 2108])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_tensor([2108, 2108])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b90f17d2119232a6cdee9db1688866047576bfc9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_tensor([2108, 2108])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The first-person point-of-view shot places the viewer directly in the heart of a tranquil, sunlit forest, the camera held steady to capture a moment of quiet connection. A slender, pale-skinned hand reaches forward from the lower portion of the frame with deliberate grace. The focus is on the hand as it approaches a tall birch tree, its iconic white bark rendered in sharp detail, showing the papery layers peeling and curling, revealing a darker, rougher texture beneath. As the hand approaches, its long, graceful fingers spread slightly in anticipation. The camera grips the fingertips, gently and tentatively touching each one, one by one, the unique surface of the tree trunk, a gesture that conveys a sense of curiosity and reverence. Dappled sunlight filters through the overhead canopy, casting ever-changing patterns of light and shadow on the hand and the tree, while a soft-focus background reveals the leaf-covered forest floor. The scene is imbued with a tranquil and contemplative atmosphere, emphasizing the tactile feel of skin against bark and the profound stillness of the natural world.", + "seed": 33223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_2108_seed_33223_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016552886553108694, + "vlm_logprob": -0.016552886553108694, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.016552886553108694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03632359579205513 + ], + "opposite_logprob_tokens": [ + -4.036323547363281 + ], + "logprob_reward": -0.03632359579205513, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03632359579205513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20782232284545898, + 0.031656187027692795, + 0.045669518411159515, + 0.15789850056171417, + 0.09002280235290527, + 0.04382861778140068, + 0.17805559933185577, + 0.044291671365499496, + 0.14838048815727234, + 0.05237427353858948 + ], + "credit_map": [ + 0.023134632036089897, + 0.07819873094558716, + 0.027507537975907326, + 0.012638435699045658, + 0.014297510497272015, + 0.013939558528363705, + 0.018976548686623573, + 0.019129374995827675, + 0.005191606003791094, + 0.007313404232263565, + 0.003972406964749098, + 0.00416382635012269, + 0.0028041712939739227, + 0.0026987469755113125, + 0.002980284858494997, + 0.0025317419786006212, + 0.006764276884496212, + 0.004480895586311817, + 0.006786054000258446, + 0.007048392202705145, + 0.006245250813663006, + 0.004749152343720198, + 0.006042034365236759, + 0.003553467569872737, + 0.01282567624002695, + 0.017689937725663185, + 0.012788167223334312, + 0.05020502954721451, + 0.013392499648034573, + 0.012392058037221432, + 0.027575267478823662, + 0.011029870249330997, + 0.004660825710743666, + 0.005763816647231579, + 0.011647433042526245, + 0.039988383650779724, + 0.004906894639134407, + 0.009311951696872711, + 0.010033213533461094, + 0.003710291115567088, + 0.003686465322971344, + 0.004273212514817715, + 0.0046963682398200035, + 0.008284234441816807, + 0.007184954360127449, + 0.004903074353933334, + 0.005030136555433273, + 0.005770175252109766, + 0.06097419559955597, + 0.011598510667681694, + 0.05008937790989876, + 0.007969992235302925, + 0.01005652453750372, + 0.012385779060423374, + 0.007439608685672283, + 0.01754162833094597, + 0.004936464596539736, + 0.003538592020049691, + 0.00837703887373209, + 0.005159188061952591, + 0.008631536737084389, + 0.00455035176128149, + 0.005374681670218706, + 0.0037238190416246653, + 0.011788171716034412, + 0.025088217109441757, + 0.015284695662558079, + 0.018911093473434448, + 0.014378346502780914, + 0.022175749763846397, + 0.019580256193876266, + 0.02117396704852581, + 0.0058318558149039745, + 0.008317512460052967, + 0.006099254824221134, + 0.00779120484367013, + 0.007096428889781237, + 0.005955989006906748, + 0.006178149953484535, + 0.005103882867842913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.03632359579205513 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005274902563542128 + ], + "opposite_logprob_tokens": [ + -5.255274772644043 + ], + "logprob_reward": -0.005274902563542128, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005274902563542128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07515905052423477, + 0.048596978187561035, + 0.02726445347070694, + 0.08361630141735077, + 0.07581229507923126, + 0.04953007027506828, + 0.10058268159627914, + 0.07440872490406036, + 0.3808256685733795, + 0.08420383930206299 + ], + "credit_map": [ + 0.008495870046317577, + 0.02811993844807148, + 0.011506689712405205, + 0.004402664490044117, + 0.005304341204464436, + 0.004623801913112402, + 0.006346649490296841, + 0.0063590994104743, + 0.005960199981927872, + 0.011329230852425098, + 0.005504344590008259, + 0.006333416793495417, + 0.004198487848043442, + 0.003644336713477969, + 0.005365180317312479, + 0.006261779926717281, + 0.0026758983731269836, + 0.0035579348914325237, + 0.0028070120606571436, + 0.004739279858767986, + 0.004720141179859638, + 0.002137290546670556, + 0.004262078553438187, + 0.002364818472415209, + 0.006781437899917364, + 0.007187377195805311, + 0.006778582464903593, + 0.026333939284086227, + 0.007308422587811947, + 0.005923695862293243, + 0.01567121222615242, + 0.007631637621670961, + 0.004746105056256056, + 0.004099579062312841, + 0.00970342755317688, + 0.021490221843123436, + 0.003822025377303362, + 0.019504522904753685, + 0.008357333950698376, + 0.0040890793316066265, + 0.0037165216635912657, + 0.0035588222090154886, + 0.008425015956163406, + 0.01538173109292984, + 0.0060124886222183704, + 0.003639677306637168, + 0.0049436879344284534, + 0.003852128516882658, + 0.036357082426548004, + 0.012043201364576817, + 0.023784196004271507, + 0.00433273846283555, + 0.007263953797519207, + 0.006301562767475843, + 0.005392162129282951, + 0.005107795353978872, + 0.006887425202876329, + 0.006368550006300211, + 0.008443252183496952, + 0.008511105552315712, + 0.02082221210002899, + 0.0073683918453752995, + 0.010212583467364311, + 0.005795203614979982, + 0.03215121850371361, + 0.061358626931905746, + 0.05059418827295303, + 0.03408883512020111, + 0.030189253389835358, + 0.05269277095794678, + 0.048184651881456375, + 0.07156611233949661, + 0.007737408392131329, + 0.018026389181613922, + 0.008683368563652039, + 0.008040215820074081, + 0.010461797006428242, + 0.008786913938820362, + 0.01296188309788704, + 0.0095058623701334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.005274902563542128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640880040824413 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640880040824413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640880040824413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06893368065357208, + 0.05427107214927673, + 0.06227148696780205, + 0.04175134748220444, + 0.029781745746731758, + 0.23966604471206665, + 0.027380172163248062, + 0.05187937989830971, + 0.3782786428928375, + 0.045786429196596146 + ], + "credit_map": [ + 0.009504390880465508, + 0.010161534883081913, + 0.011746139265596867, + 0.0051228697411715984, + 0.009575211443006992, + 0.007169094402343035, + 0.009303059428930283, + 0.006351381074637175, + 0.00552316103130579, + 0.014845302328467369, + 0.006366967223584652, + 0.007906565442681313, + 0.005261488724499941, + 0.005005350802093744, + 0.00541573716327548, + 0.0039464980363845825, + 0.007489550858736038, + 0.007459866814315319, + 0.007337561342865229, + 0.008111091330647469, + 0.00961985532194376, + 0.00579890888184309, + 0.010178932920098305, + 0.006275716237723827, + 0.004077179357409477, + 0.005128002259880304, + 0.004527651704847813, + 0.009498941712081432, + 0.004497468937188387, + 0.004761698190122843, + 0.005648280959576368, + 0.0036121264565736055, + 0.002474842593073845, + 0.0035312494728714228, + 0.004160371609032154, + 0.003244953230023384, + 0.0035865658428519964, + 0.006265257950872183, + 0.003655056469142437, + 0.0028634483460336924, + 0.015565297566354275, + 0.015682166442275047, + 0.038136087357997894, + 0.09892335534095764, + 0.02241966314613819, + 0.01442327257245779, + 0.024294450879096985, + 0.010221755132079124, + 0.0031249281018972397, + 0.0026313881389796734, + 0.004033421631902456, + 0.002626876812428236, + 0.005905934609472752, + 0.0029915564227849245, + 0.002727434504777193, + 0.0033386312425136566, + 0.004334456752985716, + 0.0040419818833470345, + 0.007287102751433849, + 0.005483497865498066, + 0.013541913591325283, + 0.006094782613217831, + 0.006900463718920946, + 0.0041951825842261314, + 0.030743498355150223, + 0.052861537784338, + 0.04922982677817345, + 0.03168579936027527, + 0.02782633900642395, + 0.04932839050889015, + 0.0472976453602314, + 0.08930561691522598, + 0.003978190943598747, + 0.012117698788642883, + 0.006661145016551018, + 0.003737581428140402, + 0.006526175420731306, + 0.00391518697142601, + 0.005390495993196964, + 0.003459957195445895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.004640880040824413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067427014000713825 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.0067427014000713825, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067427014000713825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050025295466184616, + 0.10416329652070999, + 0.022531213238835335, + 0.04344991222023964, + 0.13236261904239655, + 0.16052545607089996, + 0.06564386934041977, + 0.0742662250995636, + 0.24557992815971375, + 0.10145217925310135 + ], + "credit_map": [ + 0.008242919109761715, + 0.00622868025675416, + 0.006885378155857325, + 0.0044632856734097, + 0.006053514778614044, + 0.005922195501625538, + 0.006270955316722393, + 0.0059583671391010284, + 0.012976130470633507, + 0.019355658441781998, + 0.015218563377857208, + 0.013253653421998024, + 0.012400783598423004, + 0.009502789936959743, + 0.012672589160501957, + 0.008783121593296528, + 0.003534651594236493, + 0.002497368957847357, + 0.0027527364436537027, + 0.0029530893079936504, + 0.0036431103944778442, + 0.0022381653543561697, + 0.003203578293323517, + 0.001708511495962739, + 0.007054650690406561, + 0.0036641033366322517, + 0.005981104914098978, + 0.00569077068939805, + 0.0043274457566440105, + 0.005316652357578278, + 0.005832062102854252, + 0.005583126097917557, + 0.009060526266694069, + 0.004499032627791166, + 0.02828526683151722, + 0.021325932815670967, + 0.0056006647646427155, + 0.04633737727999687, + 0.006933433935046196, + 0.01032037939876318, + 0.009695269167423248, + 0.010287990793585777, + 0.027114691212773323, + 0.06338821351528168, + 0.015338638797402382, + 0.010852165520191193, + 0.016065116971731186, + 0.007783361244946718, + 0.024663232266902924, + 0.006315809674561024, + 0.01562752202153206, + 0.003477871185168624, + 0.0034397433046251535, + 0.004014025442302227, + 0.0033090603537857533, + 0.00479660602286458, + 0.007211846299469471, + 0.005462027620524168, + 0.009412996470928192, + 0.006856970023363829, + 0.022754808887839317, + 0.007705457042902708, + 0.009017810225486755, + 0.0058443075977265835, + 0.019565466791391373, + 0.022005118429660797, + 0.03681917488574982, + 0.02774631232023239, + 0.021845944225788116, + 0.033994756639003754, + 0.029516156762838364, + 0.05408700183033943, + 0.00885249674320221, + 0.022780030965805054, + 0.014315586537122726, + 0.008283767849206924, + 0.014101453125476837, + 0.011818574741482735, + 0.011711291037499905, + 0.009588988497853279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.0067427014000713825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029782352969050407 + ], + "opposite_logprob_tokens": [ + -3.529782295227051 + ], + "logprob_reward": -0.029782352969050407, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029782352969050407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.233415424823761, + 0.04018258675932884, + 0.020609404891729355, + 0.0912930816411972, + 0.09663356840610504, + 0.0950959324836731, + 0.037549395114183426, + 0.037977926433086395, + 0.2191917598247528, + 0.12805090844631195 + ], + "credit_map": [ + 0.02903156913816929, + 0.08281845599412918, + 0.028616303578019142, + 0.016072344034910202, + 0.01738111861050129, + 0.016539394855499268, + 0.02095732092857361, + 0.02199893817305565, + 0.006323855370283127, + 0.007310913875699043, + 0.004300981294363737, + 0.0035531469620764256, + 0.00520472414791584, + 0.004090399947017431, + 0.004724137019366026, + 0.004674434661865234, + 0.002521924441680312, + 0.0032346316147595644, + 0.0028242801781743765, + 0.0030857932288199663, + 0.0026360645424574614, + 0.0022754655219614506, + 0.0019431182881817222, + 0.0020881290547549725, + 0.008311242796480656, + 0.00795977283269167, + 0.009838652797043324, + 0.02581731602549553, + 0.0068169585429131985, + 0.011789382435381413, + 0.011090758256614208, + 0.009669005870819092, + 0.00722875539213419, + 0.0031778672710061073, + 0.022676050662994385, + 0.012927265837788582, + 0.003052850253880024, + 0.03520479053258896, + 0.004286143928766251, + 0.008079859428107738, + 0.007522196043282747, + 0.0065545993857085705, + 0.016907798126339912, + 0.027758225798606873, + 0.011208245530724525, + 0.008202820084989071, + 0.009575317613780499, + 0.00736673828214407, + 0.013053341768682003, + 0.002192853018641472, + 0.009837460704147816, + 0.0020173953380435705, + 0.003577277995646, + 0.0024588359519839287, + 0.0018912549130618572, + 0.0025209798477590084, + 0.0038122422993183136, + 0.0032016332261264324, + 0.004247875418514013, + 0.0053407857194542885, + 0.00881829485297203, + 0.004066737834364176, + 0.005022089928388596, + 0.0034682732075452805, + 0.017384056001901627, + 0.030539866536855698, + 0.021320955827832222, + 0.028697427362203598, + 0.018801651895046234, + 0.02519635483622551, + 0.028681641444563866, + 0.04856982082128525, + 0.012648448348045349, + 0.01799911819398403, + 0.015337199904024601, + 0.01147935539484024, + 0.018289364874362946, + 0.02075023576617241, + 0.015400253236293793, + 0.016146942973136902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.029782352969050407 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The first-person point-of-view shot places the viewer directly in the heart of a tranquil, sunlit forest, the camera held steady to capture a moment of quiet connection. A slender, pale-skinned hand reaches forward from the lower portion of the frame with deliberate grace. The focus is on the hand as it approaches a tall birch tree, its iconic white bark rendered in sharp detail, showing the papery layers peeling and curling, revealing a darker, rougher texture beneath. As the hand approaches, its long, graceful fingers spread slightly in anticipation. The camera grips the fingertips, gently and tentatively touching each one, one by one, the unique surface of the tree trunk, a gesture that conveys a sense of curiosity and reverence. Dappled sunlight filters through the overhead canopy, casting ever-changing patterns of light and shadow on the hand and the tree, while a soft-focus background reveals the leaf-covered forest floor. The scene is imbued with a tranquil and contemplative atmosphere, emphasizing the tactile feel of skin against bark and the profound stillness of the natural world.", + "seed": 33224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_2108_seed_33224_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.00894117052666843, + "vlm_logprob": -0.00894117052666843, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.00894117052666843, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023635663092136383 + ], + "opposite_logprob_tokens": [ + -4.5236358642578125 + ], + "logprob_reward": -0.023635663092136383, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.023635663092136383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1709786355495453, + 0.06130879372358322, + 0.08997223526239395, + 0.03518678992986679, + 0.18484991788864136, + 0.09875205159187317, + 0.06999421864748001, + 0.1724674552679062, + 0.07275582104921341, + 0.04373403266072273 + ], + "credit_map": [ + 0.018309246748685837, + 0.018691491335630417, + 0.015059596858918667, + 0.022731417790055275, + 0.02884957566857338, + 0.015858478844165802, + 0.02612314559519291, + 0.025355689227581024, + 0.0035305707715451717, + 0.005677259527146816, + 0.0036579135339707136, + 0.006153502035886049, + 0.014326211996376514, + 0.009384742006659508, + 0.007628353778272867, + 0.01095023937523365, + 0.00886168796569109, + 0.008391824550926685, + 0.008270558901131153, + 0.00902634672820568, + 0.011482572183012962, + 0.019426381215453148, + 0.010364077985286713, + 0.01414878573268652, + 0.004616136197000742, + 0.002550906501710415, + 0.0024734679609537125, + 0.005405085161328316, + 0.0030015716329216957, + 0.004422240890562534, + 0.004688540007919073, + 0.008028843440115452, + 0.015348033979535103, + 0.013796950690448284, + 0.018112676218152046, + 0.031585853546857834, + 0.01898634433746338, + 0.037777818739414215, + 0.024351703003048897, + 0.024890530854463577, + 0.011316729709506035, + 0.007360661402344704, + 0.009669842198491096, + 0.0204696636646986, + 0.009212056174874306, + 0.014982417225837708, + 0.01136051956564188, + 0.01438016165047884, + 0.00864448957145214, + 0.011813643388450146, + 0.006867501884698868, + 0.007804986089468002, + 0.006244912277907133, + 0.009767288342118263, + 0.00949303712695837, + 0.009358364157378674, + 0.01634065806865692, + 0.011088534258306026, + 0.018164727836847305, + 0.020544232800602913, + 0.03242405131459236, + 0.03000270389020443, + 0.027578741312026978, + 0.016323799267411232, + 0.005985386669635773, + 0.005577822215855122, + 0.012794619426131248, + 0.022101255133748055, + 0.006202510092407465, + 0.007432902697473764, + 0.006294327322393656, + 0.006367004942148924, + 0.005230471026152372, + 0.004622025415301323, + 0.003938743378967047, + 0.006697062402963638, + 0.00435930909588933, + 0.006391457747668028, + 0.004777406807988882, + 0.0077175572514534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.023635663092136383 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006753950379788876 + ], + "opposite_logprob_tokens": [ + -5.006753921508789 + ], + "logprob_reward": -0.006753950379788876, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006753950379788876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10839246958494186, + 0.04502175375819206, + 0.07476919889450073, + 0.037224527448415756, + 0.20537658035755157, + 0.1296781450510025, + 0.11637106537818909, + 0.11265590786933899, + 0.09912022203207016, + 0.07139011472463608 + ], + "credit_map": [ + 0.013851183466613293, + 0.015118262730538845, + 0.009587022475898266, + 0.01457173191010952, + 0.01594003103673458, + 0.010003921575844288, + 0.014904453419148922, + 0.014415865764021873, + 0.0034434155095368624, + 0.004731120076030493, + 0.003548471024259925, + 0.0030369828455150127, + 0.008394510485231876, + 0.0050361063331365585, + 0.00520522752776742, + 0.011625916697084904, + 0.004087340086698532, + 0.004345051012933254, + 0.005082078743726015, + 0.010213818401098251, + 0.009558402001857758, + 0.017550062388181686, + 0.00894821248948574, + 0.014984234236180782, + 0.002769084647297859, + 0.002157678594812751, + 0.0020265576895326376, + 0.00814933143556118, + 0.0027801564428955317, + 0.004876707214862108, + 0.006602505687624216, + 0.007862507365643978, + 0.015949511900544167, + 0.018814079463481903, + 0.01703627035021782, + 0.03679583594202995, + 0.01776224933564663, + 0.027087515220046043, + 0.03422219678759575, + 0.037708934396505356, + 0.01321057602763176, + 0.01026927586644888, + 0.012153789401054382, + 0.020434211939573288, + 0.015011315234005451, + 0.022360732778906822, + 0.014754132367670536, + 0.02148411236703396, + 0.009915074333548546, + 0.009668899700045586, + 0.013104848563671112, + 0.01468762755393982, + 0.011622663587331772, + 0.023697594180703163, + 0.010623190551996231, + 0.02305116504430771, + 0.006683514453470707, + 0.007989074103534222, + 0.012097415514290333, + 0.01792742684483528, + 0.018226785585284233, + 0.017984943464398384, + 0.0200610663741827, + 0.011685690842568874, + 0.007386548910290003, + 0.007548139430582523, + 0.01918470859527588, + 0.031528715044260025, + 0.007273735944181681, + 0.008244556374847889, + 0.006882591638714075, + 0.011071214452385902, + 0.008718891069293022, + 0.0077217151410877705, + 0.006929285358637571, + 0.010520086623728275, + 0.006230021361261606, + 0.008200417272746563, + 0.010343953035771847, + 0.012725751847028732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.006753950379788876 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010448002722114325 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010448002722114325, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010448002722114325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15634514391422272, + 0.13691818714141846, + 0.09879650920629501, + 0.14212758839130402, + 0.16453181207180023, + 0.0590067133307457, + 0.032657504081726074, + 0.10973528027534485, + 0.06050989776849747, + 0.03937142714858055 + ], + "credit_map": [ + 0.015124460682272911, + 0.016581213101744652, + 0.010655958205461502, + 0.017701860517263412, + 0.02169213257730007, + 0.019687281921505928, + 0.020731745287775993, + 0.0341704860329628, + 0.011200547218322754, + 0.011026528663933277, + 0.005749272182583809, + 0.01025165244936943, + 0.017919646576046944, + 0.027231579646468163, + 0.014289711602032185, + 0.0392492413520813, + 0.0073295398615300655, + 0.0071790143847465515, + 0.007475199643522501, + 0.01221439614892006, + 0.014396839775145054, + 0.022576406598091125, + 0.010989075526595116, + 0.016636034473776817, + 0.00903238169848919, + 0.007336318492889404, + 0.0066416761837899685, + 0.042342279106378555, + 0.008758831769227982, + 0.014258883893489838, + 0.033186640590429306, + 0.02057057060301304, + 0.014366867020726204, + 0.018733343109488487, + 0.015964478254318237, + 0.01825571432709694, + 0.018918612971901894, + 0.033882979303598404, + 0.025219716131687164, + 0.019190099090337753, + 0.005084121599793434, + 0.007222531829029322, + 0.0059282127767801285, + 0.014380821026861668, + 0.007606789469718933, + 0.006092868745326996, + 0.005424927920103073, + 0.007266434840857983, + 0.004640519618988037, + 0.0028470337856560946, + 0.005686413496732712, + 0.005290331784635782, + 0.0031366136390715837, + 0.003514087526127696, + 0.003589688567444682, + 0.0039528170600533485, + 0.012894650921225548, + 0.00864257663488388, + 0.011557871475815773, + 0.014360768720507622, + 0.015613047406077385, + 0.008726797066628933, + 0.027531083673238754, + 0.010408480651676655, + 0.004114412236958742, + 0.0038580656982958317, + 0.008463485166430473, + 0.010553402826189995, + 0.009431985206902027, + 0.0056266444735229015, + 0.009666664525866508, + 0.008795238099992275, + 0.005331582389771938, + 0.005048105958849192, + 0.004380050580948591, + 0.007080575451254845, + 0.0024196102749556303, + 0.0055095599964261055, + 0.004202785901725292, + 0.005399156361818314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0010448002722114325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022031813859939575 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.0022031813859939575, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022031813859939575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15391981601715088, + 0.15383312106132507, + 0.1416349709033966, + 0.08486679941415787, + 0.11041204631328583, + 0.059998925775289536, + 0.055193882435560226, + 0.11941416561603546, + 0.07137947529554367, + 0.04934675991535187 + ], + "credit_map": [ + 0.014962172135710716, + 0.017466027289628983, + 0.0124999163672328, + 0.01944240927696228, + 0.029913784936070442, + 0.012613586150109768, + 0.028119666501879692, + 0.0189022496342659, + 0.01335233449935913, + 0.016740547493100166, + 0.015783965587615967, + 0.025586899369955063, + 0.030455607920885086, + 0.015380367636680603, + 0.01934359408915043, + 0.01718979887664318, + 0.009547224268317223, + 0.008580398745834827, + 0.010493290610611439, + 0.025220422074198723, + 0.01635982282459736, + 0.029850291088223457, + 0.01768963411450386, + 0.023893890902400017, + 0.004958163946866989, + 0.005313357338309288, + 0.004268785007297993, + 0.021139249205589294, + 0.0054502640850842, + 0.00848484318703413, + 0.02130535989999771, + 0.013946766033768654, + 0.010923574678599834, + 0.007962677627801895, + 0.011279664933681488, + 0.015721719712018967, + 0.013025668449699879, + 0.0238026212900877, + 0.014498535543680191, + 0.013197586871683598, + 0.010130703449249268, + 0.004877099767327309, + 0.009077196009457111, + 0.005920258816331625, + 0.006534488871693611, + 0.008027691394090652, + 0.007246737368404865, + 0.008184750564396381, + 0.00519516272470355, + 0.009266272187232971, + 0.006472290493547916, + 0.007789460942149162, + 0.004378039389848709, + 0.00816781260073185, + 0.006731050554662943, + 0.007193795870989561, + 0.013826033100485802, + 0.008056595921516418, + 0.01309818122535944, + 0.015433820895850658, + 0.021230356767773628, + 0.01888493448495865, + 0.01741042546927929, + 0.011473813094198704, + 0.006079021841287613, + 0.005628813989460468, + 0.013103474862873554, + 0.023442301899194717, + 0.00449863076210022, + 0.004467271268367767, + 0.0054676649160683155, + 0.008692299947142601, + 0.006304766051471233, + 0.006943624932318926, + 0.007362473756074905, + 0.010779110714793205, + 0.002360805170610547, + 0.004738316405564547, + 0.004218719899654388, + 0.006638945080339909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0022031813859939575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068257503211498 + ], + "opposite_logprob_tokens": [ + -4.511068344116211 + ], + "logprob_reward": -0.011068257503211498, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068257503211498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1784113496541977, + 0.12244760245084763, + 0.07784653455018997, + 0.12614139914512634, + 0.09757780283689499, + 0.03747762739658356, + 0.09935975819826126, + 0.0431530736386776, + 0.07233521342277527, + 0.14524967968463898 + ], + "credit_map": [ + 0.01702866703271866, + 0.01973932981491089, + 0.016603482887148857, + 0.02342323027551174, + 0.030535142868757248, + 0.013733400031924248, + 0.033267535269260406, + 0.0240805521607399, + 0.011426769196987152, + 0.01506691612303257, + 0.012117278762161732, + 0.02230767346918583, + 0.01912517473101616, + 0.014124334789812565, + 0.014406387694180012, + 0.013873069547116756, + 0.008025293238461018, + 0.006305026821792126, + 0.006760749500244856, + 0.02179580181837082, + 0.008083131164312363, + 0.010399877093732357, + 0.0076029617339372635, + 0.008873691782355309, + 0.017067482694983482, + 0.0092051075771451, + 0.006561942398548126, + 0.04428087919950485, + 0.005784295964986086, + 0.007711450569331646, + 0.026203570887446404, + 0.009326671250164509, + 0.01276580709964037, + 0.008574292063713074, + 0.009561896324157715, + 0.010357131250202656, + 0.011122111231088638, + 0.019511329010128975, + 0.012804893776774406, + 0.012880350463092327, + 0.004369223956018686, + 0.006866119801998138, + 0.005519389640539885, + 0.0044055781327188015, + 0.00533278239890933, + 0.0036253798753023148, + 0.0041317506693303585, + 0.0032274022232741117, + 0.011561281979084015, + 0.021909493952989578, + 0.015368800610303879, + 0.011001902632415295, + 0.009967735037207603, + 0.009384412318468094, + 0.010318173095583916, + 0.009847951121628284, + 0.00438344432041049, + 0.003888819832354784, + 0.005411078687757254, + 0.006039308849722147, + 0.0051934197545051575, + 0.003811886068433523, + 0.01076467614620924, + 0.003660438349470496, + 0.006617505569010973, + 0.006335957907140255, + 0.011038566008210182, + 0.021966984495520592, + 0.0053842272609472275, + 0.006286346819251776, + 0.005185310263186693, + 0.009520314633846283, + 0.020927729085087776, + 0.021863799542188644, + 0.024135040119290352, + 0.037424277514219284, + 0.00527142034843564, + 0.010768505744636059, + 0.011943081393837929, + 0.012915827333927155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.011068257503211498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The first-person point-of-view shot places the viewer directly in the heart of a tranquil, sunlit forest, the camera held steady to capture a moment of quiet connection. A slender, pale-skinned hand reaches forward from the lower portion of the frame with deliberate grace. The focus is on the hand as it approaches a tall birch tree, its iconic white bark rendered in sharp detail, showing the papery layers peeling and curling, revealing a darker, rougher texture beneath. As the hand approaches, its long, graceful fingers spread slightly in anticipation. The camera grips the fingertips, gently and tentatively touching each one, one by one, the unique surface of the tree trunk, a gesture that conveys a sense of curiosity and reverence. Dappled sunlight filters through the overhead canopy, casting ever-changing patterns of light and shadow on the hand and the tree, while a soft-focus background reveals the leaf-covered forest floor. The scene is imbued with a tranquil and contemplative atmosphere, emphasizing the tactile feel of skin against bark and the profound stillness of the natural world.", + "seed": 33225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_2108_seed_33225_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012633725674822927, + "vlm_logprob": -0.012633725674822927, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.012633725674822927, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03933357447385788 + ], + "opposite_logprob_tokens": [ + -3.7893335819244385 + ], + "logprob_reward": -0.03933357447385788, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03933357447385788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07531819492578506, + 0.06074203550815582, + 0.016442300751805305, + 0.0743248239159584, + 0.03322528302669525, + 0.18220658600330353, + 0.07195179909467697, + 0.13583742082118988, + 0.31230270862579346, + 0.03764881193637848 + ], + "credit_map": [ + 0.009638027288019657, + 0.011970333755016327, + 0.008969027549028397, + 0.009241173043847084, + 0.006370238959789276, + 0.009213855490088463, + 0.008525027893483639, + 0.011390517465770245, + 0.0038336666766554117, + 0.006642807275056839, + 0.004658564459532499, + 0.007568544242531061, + 0.006811873987317085, + 0.011270381510257721, + 0.009516564197838306, + 0.010439636185765266, + 0.001430500647984445, + 0.0019010326359421015, + 0.0024878745898604393, + 0.0016008422244340181, + 0.0017834942555055022, + 0.002905401634052396, + 0.0022947099059820175, + 0.002038446022197604, + 0.006404594052582979, + 0.007266167085617781, + 0.0063330139964818954, + 0.007165184244513512, + 0.011741302907466888, + 0.015698306262493134, + 0.012410782277584076, + 0.0073054819367825985, + 0.0029984319116920233, + 0.004780094139277935, + 0.0032693292014300823, + 0.0043137394823133945, + 0.0049629067070782185, + 0.0057474421337246895, + 0.003259968012571335, + 0.0038933733012527227, + 0.01263639610260725, + 0.01259137224406004, + 0.01729760877788067, + 0.015089419670403004, + 0.02467467077076435, + 0.03813836723566055, + 0.02352355793118477, + 0.03825520724058151, + 0.0059027159586548805, + 0.00864118617027998, + 0.009599998593330383, + 0.009895767085254192, + 0.012125331908464432, + 0.007934750989079475, + 0.009236658923327923, + 0.008615396916866302, + 0.011950334534049034, + 0.013412135653197765, + 0.01520998403429985, + 0.01480973232537508, + 0.017179127782583237, + 0.030864292755723, + 0.016593841835856438, + 0.015817975625395775, + 0.03333260118961334, + 0.04246508330106735, + 0.029868820682168007, + 0.03913996368646622, + 0.05047796294093132, + 0.043337851762771606, + 0.0389871783554554, + 0.03469325602054596, + 0.002825135597959161, + 0.005328638944774866, + 0.0042552282102406025, + 0.006273613311350346, + 0.0060526542365550995, + 0.0045631397515535355, + 0.004257816821336746, + 0.004092590883374214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.03933357447385788 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00676188338547945 + ], + "opposite_logprob_tokens": [ + -5.0067620277404785 + ], + "logprob_reward": -0.00676188338547945, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00676188338547945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08596119284629822, + 0.09027153253555298, + 0.047311846166849136, + 0.05084889754652977, + 0.06751997023820877, + 0.18740545213222504, + 0.05137767270207405, + 0.059712111949920654, + 0.27942609786987305, + 0.08016519993543625 + ], + "credit_map": [ + 0.008602638728916645, + 0.009066720493137836, + 0.007298497948795557, + 0.009213542565703392, + 0.009746957570314407, + 0.011285601183772087, + 0.012823265045881271, + 0.017923973500728607, + 0.006567658390849829, + 0.012390135787427425, + 0.006902364082634449, + 0.014722364023327827, + 0.00866654608398676, + 0.011539178900420666, + 0.020079603418707848, + 0.009403679519891739, + 0.004704468417912722, + 0.005188259296119213, + 0.0039006685838103294, + 0.0037969781551510096, + 0.005890064872801304, + 0.00948248989880085, + 0.007463396061211824, + 0.006885519716888666, + 0.005176539067178965, + 0.005250121932476759, + 0.006143351085484028, + 0.005195036064833403, + 0.006166708189994097, + 0.00999688170850277, + 0.0067861974239349365, + 0.006134061608463526, + 0.005551809910684824, + 0.00847098883241415, + 0.0071389339864254, + 0.008752092719078064, + 0.012580908834934235, + 0.008771688677370548, + 0.008132164366543293, + 0.008121385239064693, + 0.018025385215878487, + 0.02135598286986351, + 0.018761983141303062, + 0.01566147617995739, + 0.028682881966233253, + 0.032292935997247696, + 0.02007937990128994, + 0.032545432448387146, + 0.006251016166061163, + 0.0050400239415466785, + 0.008561194874346256, + 0.005395694635808468, + 0.006298529449850321, + 0.006960596889257431, + 0.004321355372667313, + 0.008549261838197708, + 0.006487458478659391, + 0.005890079773962498, + 0.0057174284011125565, + 0.006784921046346426, + 0.008787715807557106, + 0.010200602933764458, + 0.008080503903329372, + 0.0077634030021727085, + 0.031627096235752106, + 0.03448928892612457, + 0.02507363259792328, + 0.03128447011113167, + 0.04025612026453018, + 0.04355614259839058, + 0.04370874539017677, + 0.029430601745843887, + 0.008715285919606686, + 0.0071431417018175125, + 0.00875622034072876, + 0.008270630612969398, + 0.006952815689146519, + 0.015997720882296562, + 0.016933631151914597, + 0.007395752239972353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.00676188338547945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001341396477073431 + ], + "opposite_logprob_tokens": [ + -6.626341342926025 + ], + "logprob_reward": -0.001341396477073431, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001341396477073431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2527906894683838, + 0.07749726623296738, + 0.08599372208118439, + 0.12792131304740906, + 0.06746302545070648, + 0.08263377845287323, + 0.05821672081947327, + 0.09177672117948532, + 0.11059771478176117, + 0.04510912671685219 + ], + "credit_map": [ + 0.020683130249381065, + 0.025880899280309677, + 0.028707975521683693, + 0.023036615923047066, + 0.027970079332590103, + 0.03968459367752075, + 0.04611561447381973, + 0.04071177542209625, + 0.005504751577973366, + 0.009669691324234009, + 0.006848054472357035, + 0.011911746114492416, + 0.00961302686482668, + 0.012451495975255966, + 0.012289833277463913, + 0.009208657778799534, + 0.012476888485252857, + 0.011847603134810925, + 0.010826225392520428, + 0.007521876599639654, + 0.008781829848885536, + 0.012406584806740284, + 0.009232979267835617, + 0.012899721972644329, + 0.014394240453839302, + 0.016538437455892563, + 0.011573259718716145, + 0.012262962758541107, + 0.016387442126870155, + 0.02035263553261757, + 0.022208604961633682, + 0.014203714206814766, + 0.005473432131111622, + 0.007524061948060989, + 0.0082077132537961, + 0.013000953942537308, + 0.007503355387598276, + 0.008485367521643639, + 0.010306119918823242, + 0.006962016690522432, + 0.008403923362493515, + 0.008442669175565243, + 0.00645406125113368, + 0.008008473552763462, + 0.014063809998333454, + 0.011436901986598969, + 0.010159963741898537, + 0.01566397026181221, + 0.006281617097556591, + 0.00478339521214366, + 0.006312888581305742, + 0.00783586222678423, + 0.008823267184197903, + 0.009154750034213066, + 0.007771546021103859, + 0.007253384217619896, + 0.006715434603393078, + 0.009474127553403378, + 0.006625413428992033, + 0.01226350013166666, + 0.013539929874241352, + 0.0147105036303401, + 0.017539935186505318, + 0.010907870717346668, + 0.007677226793020964, + 0.009857330471277237, + 0.009222314693033695, + 0.01207725889980793, + 0.014857360161840916, + 0.02316964417695999, + 0.021137511357665062, + 0.012599050998687744, + 0.006732384208589792, + 0.004411768633872271, + 0.005437750834971666, + 0.0053667728789150715, + 0.004664922598749399, + 0.006821881048381329, + 0.006286828313022852, + 0.005386810749769211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.001341396477073431 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003201718209311366 + ], + "opposite_logprob_tokens": [ + -5.753201484680176 + ], + "logprob_reward": -0.003201718209311366, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003201718209311366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12243720144033432, + 0.1774839609861374, + 0.056751467287540436, + 0.07195141166448593, + 0.08048654347658157, + 0.06398354470729828, + 0.06978685408830643, + 0.131682351231575, + 0.1955893486738205, + 0.029847335070371628 + ], + "credit_map": [ + 0.014226185157895088, + 0.01701373979449272, + 0.016184791922569275, + 0.01247230265289545, + 0.00864590983837843, + 0.013459542766213417, + 0.015458228066563606, + 0.02497650682926178, + 0.020431114360690117, + 0.02651628665626049, + 0.03056948259472847, + 0.02489387057721615, + 0.015131847932934761, + 0.018845338374376297, + 0.022408628836274147, + 0.01868738792836666, + 0.00727927265688777, + 0.00778728723526001, + 0.008161625824868679, + 0.0051088337786495686, + 0.005258501973003149, + 0.007780173793435097, + 0.0067473966628313065, + 0.008628373965620995, + 0.01008670125156641, + 0.009822092950344086, + 0.012447454035282135, + 0.005436710547655821, + 0.006119188386946917, + 0.01270255632698536, + 0.005886537954211235, + 0.00945016834884882, + 0.006480427924543619, + 0.009435772895812988, + 0.008032845333218575, + 0.013783380389213562, + 0.010747380554676056, + 0.012378991581499577, + 0.01112047303467989, + 0.00850727129727602, + 0.007492374628782272, + 0.005006211344152689, + 0.00643015606328845, + 0.005856935400515795, + 0.0074417307041585445, + 0.01318139024078846, + 0.008258175104856491, + 0.010316566564142704, + 0.006849025376141071, + 0.007171587087213993, + 0.007307384628802538, + 0.008315769024193287, + 0.013172237202525139, + 0.008282389491796494, + 0.00604889914393425, + 0.012639558874070644, + 0.00826167780905962, + 0.012114722281694412, + 0.009174712002277374, + 0.022814180701971054, + 0.015210909768939018, + 0.01790725812315941, + 0.03024362213909626, + 0.015955260023474693, + 0.01854725554585457, + 0.030204692855477333, + 0.017875859513878822, + 0.02419496700167656, + 0.025569019839167595, + 0.024022504687309265, + 0.03314093127846718, + 0.02203412912786007, + 0.0029311717953532934, + 0.003833244787529111, + 0.0029736405704170465, + 0.004687903448939323, + 0.004037937149405479, + 0.0041545359417796135, + 0.003290032735094428, + 0.003938867244869471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.003201718209311366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012530055828392506 + ], + "opposite_logprob_tokens": [ + -4.3875298500061035 + ], + "logprob_reward": -0.012530055828392506, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012530055828392506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09679330885410309, + 0.09206490218639374, + 0.1329926699399948, + 0.34785231947898865, + 0.03628088906407356, + 0.06725065410137177, + 0.056273747235536575, + 0.03676525875926018, + 0.09649581462144852, + 0.03723045811057091 + ], + "credit_map": [ + 0.012466466054320335, + 0.016436386853456497, + 0.01361255906522274, + 0.010207289829850197, + 0.0060077328234910965, + 0.013812967576086521, + 0.009111205115914345, + 0.015138691291213036, + 0.010154340416193008, + 0.01462059561163187, + 0.01419775653630495, + 0.015252895653247833, + 0.0059287650510668755, + 0.009825837798416615, + 0.011483905836939812, + 0.01060080248862505, + 0.022028017789125443, + 0.022665267810225487, + 0.028159772977232933, + 0.010957309044897556, + 0.0066985515877604485, + 0.01681555062532425, + 0.008326875977218151, + 0.017341317608952522, + 0.0473581925034523, + 0.037234775722026825, + 0.04498354345560074, + 0.02862207219004631, + 0.04442998766899109, + 0.06104462593793869, + 0.04628259316086769, + 0.037896528840065, + 0.003004408674314618, + 0.004302110988646746, + 0.004836998414248228, + 0.004916093312203884, + 0.0049247401766479015, + 0.004768662620335817, + 0.004816099070012569, + 0.004711776040494442, + 0.008358723483979702, + 0.008851775899529457, + 0.006973755080252886, + 0.006062932312488556, + 0.009471563622355461, + 0.009740089997649193, + 0.006747517269104719, + 0.01104429829865694, + 0.007410217542201281, + 0.005131733603775501, + 0.007339752744883299, + 0.005507724359631538, + 0.010150340385735035, + 0.008467406965792179, + 0.005961855407804251, + 0.006304718554019928, + 0.003998169209808111, + 0.0027094921097159386, + 0.003936495166271925, + 0.0048692054115235806, + 0.0032710558734834194, + 0.008548871614038944, + 0.004958755802363157, + 0.004473214037716389, + 0.007951715961098671, + 0.012032671831548214, + 0.00883598905056715, + 0.011751871556043625, + 0.015496492385864258, + 0.011491766199469566, + 0.01820828765630722, + 0.010727010667324066, + 0.005476674996316433, + 0.003752210410311818, + 0.003283571917563677, + 0.004215192515403032, + 0.0038933767937123775, + 0.007227487862110138, + 0.005973841995000839, + 0.0034081030171364546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.012530055828392506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The first-person point-of-view shot places the viewer directly in the heart of a tranquil, sunlit forest, the camera held steady to capture a moment of quiet connection. A slender, pale-skinned hand reaches forward from the lower portion of the frame with deliberate grace. The focus is on the hand as it approaches a tall birch tree, its iconic white bark rendered in sharp detail, showing the papery layers peeling and curling, revealing a darker, rougher texture beneath. As the hand approaches, its long, graceful fingers spread slightly in anticipation. The camera grips the fingertips, gently and tentatively touching each one, one by one, the unique surface of the tree trunk, a gesture that conveys a sense of curiosity and reverence. Dappled sunlight filters through the overhead canopy, casting ever-changing patterns of light and shadow on the hand and the tree, while a soft-focus background reveals the leaf-covered forest floor. The scene is imbued with a tranquil and contemplative atmosphere, emphasizing the tactile feel of skin against bark and the profound stillness of the natural world.", + "seed": 33226, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_2/index_2108_seed_33226_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016005830105859785, + "vlm_logprob": -0.016005830105859785, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.016005830105859785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04052802920341492 + ], + "opposite_logprob_tokens": [ + -3.6655280590057373 + ], + "logprob_reward": -0.04052802920341492, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04052802920341492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21922720968723297, + 0.11749745905399323, + 0.1392187923192978, + 0.03759613260626793, + 0.05353407934308052, + 0.07069519907236099, + 0.11256228387355804, + 0.0836375281214714, + 0.08793908357620239, + 0.07809226959943771 + ], + "credit_map": [ + 0.01824038289487362, + 0.02816007286310196, + 0.028256958350539207, + 0.08284083008766174, + 0.014275670051574707, + 0.018001873046159744, + 0.013675341382622719, + 0.01577608659863472, + 0.01847924292087555, + 0.04070249944925308, + 0.025260979309678078, + 0.006331776734441519, + 0.005119200795888901, + 0.009066295810043812, + 0.005621832329779863, + 0.006915627047419548, + 0.01638202741742134, + 0.03254568949341774, + 0.028020907193422318, + 0.008328269235789776, + 0.010115597397089005, + 0.013739993795752525, + 0.009982448071241379, + 0.020103855058550835, + 0.0038350557442754507, + 0.003996639046818018, + 0.004194870591163635, + 0.003536713309586048, + 0.007035811897367239, + 0.0053460486233234406, + 0.004023254849016666, + 0.005627738777548075, + 0.004493379965424538, + 0.004773333203047514, + 0.006696454249322414, + 0.0059223296120762825, + 0.010827391408383846, + 0.007751221768558025, + 0.006516093388199806, + 0.006553878076374531, + 0.0036677175667136908, + 0.0028885777574032545, + 0.022832632064819336, + 0.0030318149365484715, + 0.0053840759210288525, + 0.02543502487242222, + 0.0030733190942555666, + 0.004382038954645395, + 0.010687069967389107, + 0.007996927946805954, + 0.016590667888522148, + 0.009756263345479965, + 0.01441352628171444, + 0.021471213549375534, + 0.013842727057635784, + 0.01780387945473194, + 0.008029148913919926, + 0.0018854436930269003, + 0.030847422778606415, + 0.0028377233538776636, + 0.002254294231534004, + 0.025586692616343498, + 0.0030790134333074093, + 0.009117789566516876, + 0.006862822920084, + 0.011125719174742699, + 0.005501273553818464, + 0.005324896890670061, + 0.020523596554994583, + 0.011605612002313137, + 0.014676829800009727, + 0.012318328022956848, + 0.010388700291514397, + 0.00884838867932558, + 0.006238441448658705, + 0.006640813313424587, + 0.01287881564348936, + 0.006820344366133213, + 0.013921966776251793, + 0.012354799546301365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04052802920341492 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005279882811009884 + ], + "opposite_logprob_tokens": [ + -5.255280017852783 + ], + "logprob_reward": -0.005279882811009884, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005279882811009884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29651668667793274, + 0.07380890846252441, + 0.16054779291152954, + 0.04832925647497177, + 0.02574075385928154, + 0.05027289688587189, + 0.04487533122301102, + 0.027410803362727165, + 0.05776016414165497, + 0.21473738551139832 + ], + "credit_map": [ + 0.02327113039791584, + 0.038844574242830276, + 0.03862690180540085, + 0.12916837632656097, + 0.008814695291221142, + 0.021785056218504906, + 0.015932802110910416, + 0.02007312886416912, + 0.00900794193148613, + 0.025853652507066727, + 0.022105015814304352, + 0.004745542071759701, + 0.0026737607549875975, + 0.0036944623570889235, + 0.0026558588724583387, + 0.0030726769473403692, + 0.01440107449889183, + 0.05848808214068413, + 0.050449248403310776, + 0.008394292555749416, + 0.008770201355218887, + 0.008342817425727844, + 0.005213848315179348, + 0.006488223560154438, + 0.005172526463866234, + 0.0049420553259551525, + 0.006000360939651728, + 0.004133647307753563, + 0.008722629398107529, + 0.007754060905426741, + 0.0043606506660580635, + 0.007243323605507612, + 0.002694192109629512, + 0.002428109524771571, + 0.004113170318305492, + 0.0015571336261928082, + 0.0026711716782301664, + 0.00516114104539156, + 0.0023985616862773895, + 0.004717273637652397, + 0.005811895709484816, + 0.003319351701065898, + 0.007374478504061699, + 0.004175888374447823, + 0.007438022643327713, + 0.007454043719917536, + 0.005854514893144369, + 0.008844702504575253, + 0.0056977286003530025, + 0.0037715721409767866, + 0.006975721567869186, + 0.004108163993805647, + 0.005263239610940218, + 0.006275956518948078, + 0.006459247320890427, + 0.006323698442429304, + 0.003022255841642618, + 0.0012498304713517427, + 0.007994206622242928, + 0.0012863286538049579, + 0.0015402372227981687, + 0.006393008399754763, + 0.0019196512876078486, + 0.0040052845142781734, + 0.00427356269210577, + 0.008285103365778923, + 0.004301946610212326, + 0.006757226772606373, + 0.007307905703783035, + 0.007414767984300852, + 0.008160199038684368, + 0.011259449645876884, + 0.009316216222941875, + 0.015548795461654663, + 0.013800174929201603, + 0.011520163156092167, + 0.01797300949692726, + 0.013100549578666687, + 0.0682462602853775, + 0.06523220986127853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.005279882811009884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008169650682248175 + ], + "opposite_logprob_tokens": [ + -7.125816822052002 + ], + "logprob_reward": -0.0008169650682248175, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008169650682248175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.240604966878891, + 0.2572944760322571, + 0.03414024040102959, + 0.029920140281319618, + 0.02285604551434517, + 0.06523387879133224, + 0.0416656956076622, + 0.06968700885772705, + 0.08941178768873215, + 0.14918570220470428 + ], + "credit_map": [ + 0.01661193184554577, + 0.02672223374247551, + 0.024860715493559837, + 0.07818783074617386, + 0.015653546899557114, + 0.04385526478290558, + 0.015007966198027134, + 0.01970549114048481, + 0.04360300675034523, + 0.09087641537189484, + 0.05462384968996048, + 0.012888231314718723, + 0.010503694415092468, + 0.02012518234550953, + 0.010445850901305676, + 0.01422826200723648, + 0.00515852402895689, + 0.003795371390879154, + 0.005121905356645584, + 0.004682957194745541, + 0.0036701462231576443, + 0.0046470873057842255, + 0.0034560831263661385, + 0.003608166938647628, + 0.004285705741494894, + 0.00622784486040473, + 0.004340396262705326, + 0.002815396524965763, + 0.003550618654116988, + 0.0032385443337261677, + 0.0026950547471642494, + 0.002766578458249569, + 0.0025156689807772636, + 0.002285899594426155, + 0.0028518654871731997, + 0.0028196536004543304, + 0.004087781999260187, + 0.0032682521268725395, + 0.0017528702737763524, + 0.0032740519382059574, + 0.004094252362847328, + 0.0039826929569244385, + 0.011951185762882233, + 0.0063994587399065495, + 0.008854149840772152, + 0.012226272374391556, + 0.008912108838558197, + 0.008813761174678802, + 0.005523682571947575, + 0.003909470979124308, + 0.006108010187745094, + 0.0038406946696341038, + 0.0038405805826187134, + 0.008702595718204975, + 0.0034491780679672956, + 0.006291484460234642, + 0.006853026337921619, + 0.0021243735682219267, + 0.02344639226794243, + 0.0030193107668310404, + 0.00263326708227396, + 0.01889265514910221, + 0.005165243521332741, + 0.007552745286375284, + 0.007018812000751495, + 0.006781159900128841, + 0.006350452546030283, + 0.005056973081082106, + 0.018254149705171585, + 0.011197332292795181, + 0.016201240941882133, + 0.018551666289567947, + 0.0063788797706365585, + 0.0049893781542778015, + 0.006170470733195543, + 0.004396012052893639, + 0.008880765177309513, + 0.010186917148530483, + 0.05484798923134804, + 0.0533352866768837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0008169650682248175 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036216876469552517 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036216876469552517, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036216876469552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1997324526309967, + 0.27594831585884094, + 0.0247939582914114, + 0.014846067875623703, + 0.01882062666118145, + 0.05740203708410263, + 0.021364571526646614, + 0.037474267184734344, + 0.032166071236133575, + 0.31745171546936035 + ], + "credit_map": [ + 0.016936296597123146, + 0.027726823464035988, + 0.026995643973350525, + 0.0863480195403099, + 0.005811378359794617, + 0.009930165484547615, + 0.012425797991454601, + 0.01355828158557415, + 0.0302839707583189, + 0.0897316038608551, + 0.07866233587265015, + 0.02518227882683277, + 0.011185531504452229, + 0.015547090210020542, + 0.011633453890681267, + 0.013722024857997894, + 0.0031170453876256943, + 0.002487152349203825, + 0.0026657034177333117, + 0.0018572905100882053, + 0.0028133131563663483, + 0.004010877106338739, + 0.002942754654213786, + 0.004899818915873766, + 0.0023277313448488712, + 0.001897399197332561, + 0.0016430264804512262, + 0.0011709738755598664, + 0.0013948078267276287, + 0.002055334858596325, + 0.0016249286709353328, + 0.0027318638749420643, + 0.0023580454289913177, + 0.002110500121489167, + 0.0032542606350034475, + 0.0023016997147351503, + 0.0019098769407719374, + 0.002452538348734379, + 0.0016313064843416214, + 0.0028023964259773493, + 0.008327824994921684, + 0.005187362898141146, + 0.008915815502405167, + 0.00484324898570776, + 0.0082401717081666, + 0.007272345945239067, + 0.00585513049736619, + 0.008760127238929272, + 0.0022879568859934807, + 0.001914443215355277, + 0.0028984711971133947, + 0.002516179345548153, + 0.0020454665645956993, + 0.00384464836679399, + 0.0022741907741874456, + 0.0035832137800753117, + 0.003578860778361559, + 0.0013951624277979136, + 0.012583005242049694, + 0.002089578192681074, + 0.0011878611985594034, + 0.011021146550774574, + 0.0017301163170486689, + 0.0038885294925421476, + 0.0026035604532808065, + 0.0034399626310914755, + 0.004567804746329784, + 0.00237680459395051, + 0.0046930513344705105, + 0.003967984113842249, + 0.0046354844234883785, + 0.005881412886083126, + 0.016706310212612152, + 0.008396976627409458, + 0.00747703155502677, + 0.00704211974516511, + 0.016698721796274185, + 0.01820339262485504, + 0.12558573484420776, + 0.11734140664339066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0036216876469552517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale-skinned hand visible reaching forward from the lower portion of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall birch tree with white bark featuring peeling, papery layers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does dappled sunlight filter through the canopy, casting patterns of light and shadow on the hand and tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus view of a leaf-covered forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978258579969406 + ], + "opposite_logprob_tokens": [ + -3.52978253364563 + ], + "logprob_reward": -0.02978258579969406, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978258579969406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20292913913726807, + 0.11629106849431992, + 0.14155356585979462, + 0.06945235282182693, + 0.04589292034506798, + 0.04734320193529129, + 0.10356904566287994, + 0.041837893426418304, + 0.044042788445949554, + 0.1870880275964737 + ], + "credit_map": [ + 0.01778245158493519, + 0.02690826542675495, + 0.024994857609272003, + 0.07833153754472733, + 0.00838214810937643, + 0.0171089768409729, + 0.013447736389935017, + 0.015973180532455444, + 0.012347918935120106, + 0.02964232861995697, + 0.020389622077345848, + 0.017865817993879318, + 0.009107767604291439, + 0.010340997017920017, + 0.008334139361977577, + 0.008262477815151215, + 0.011088703759014606, + 0.04911546781659126, + 0.0444134846329689, + 0.007805223576724529, + 0.006364434026181698, + 0.007227399852126837, + 0.007572781294584274, + 0.007966061122715473, + 0.009531192481517792, + 0.0109860859811306, + 0.01348020974546671, + 0.004680258687585592, + 0.006865459959954023, + 0.008528763428330421, + 0.005119973793625832, + 0.010260408744215965, + 0.006004274822771549, + 0.005361075978726149, + 0.008299692533910275, + 0.006458160933107138, + 0.005354775115847588, + 0.005830156151205301, + 0.003450502175837755, + 0.005134289152920246, + 0.003962427843362093, + 0.0033922193106263876, + 0.00907099898904562, + 0.005711444653570652, + 0.005209251306951046, + 0.00915886927396059, + 0.0031911544501781464, + 0.007646837737411261, + 0.012466555461287498, + 0.008037467487156391, + 0.015261339955031872, + 0.008309071883559227, + 0.011638798750936985, + 0.02070511318743229, + 0.009065057151019573, + 0.01808563806116581, + 0.006254410371184349, + 0.0031155995093286037, + 0.008255669847130775, + 0.0023962068371474743, + 0.0045598093420267105, + 0.0059563834220170975, + 0.004986768122762442, + 0.006313048768788576, + 0.0032025722321122885, + 0.004561347886919975, + 0.0045358408242464066, + 0.0040749553591012955, + 0.007346468046307564, + 0.004746692255139351, + 0.008234817534685135, + 0.007340096402913332, + 0.018289534375071526, + 0.007287807762622833, + 0.011353564448654652, + 0.009350755251944065, + 0.01479057315737009, + 0.019066222012043, + 0.053614500910043716, + 0.05333505943417549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.02978258579969406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_tensor([3093, 3093])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_tensor([3093, 3093])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..48026cc4b2ffb0813429c67e6beeb0a7768df9c9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_tensor([3093, 3093])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a slightly elevated angle, this extreme close-up, static shot captures a moment of culinary joy, imbued with a cinematic quality and a hint of film grain. A polished silver fork, pressed deliberately, its tines gently cracking the taut, glistening white of a perfectly poached egg, resting on a thick slice of golden-brown artisan toast. The moment the membrane is pierced, a vibrant, deep-gold yolk breaks through the surface and begins to ooze out in a slow, viscous, and utterly satisfying flow. This creamy, liquid gold lazily overflows from the egg's edges, landing on the toast's crisp, richly textured crumb and pooling in its crevices. The scene is bathed in soft, warm morning light, filtering in from an off-camera window, creating delicate highlights on the glistening yolk and the metallic sheen of the fork. The shallow depth of field blurs the rustic wooden table in the background, focusing the viewer's attention squarely on the rich textures and movement, creating a tranquil, appetizing, and deeply sensory atmosphere.", + "seed": 30437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_3093_seed_30437_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0056429912336170675, + "vlm_logprob": -0.0056429912336170675, + "vlm_yes_no_margin": 5.875, + "vlm_token_credit": -0.0056429912336170675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009346893057227135 + ], + "opposite_logprob_tokens": [ + -6.384346961975098 + ], + "logprob_reward": -0.009346893057227135, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.009346893057227135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21179310977458954, + 0.045701660215854645, + 0.02017318271100521, + 0.029548535123467445, + 0.04963432252407074, + 0.043183356523513794, + 0.24574896693229675, + 0.17217214405536652, + 0.06616809219121933, + 0.11587665975093842 + ], + "credit_map": [ + 0.01298537664115429, + 0.027813443914055824, + 0.01773076318204403, + 0.03101770393550396, + 0.03314913436770439, + 0.01795213483273983, + 0.04464045539498329, + 0.026504075154662132, + 0.005249381065368652, + 0.004087178036570549, + 0.010797752067446709, + 0.005227168556302786, + 0.003272335510700941, + 0.0033854569774121046, + 0.005780626554042101, + 0.007901757024228573, + 0.002758399583399296, + 0.002668728120625019, + 0.003349103732034564, + 0.0021417313255369663, + 0.0024609414394944906, + 0.0015357170486822724, + 0.002318685408681631, + 0.002939875004813075, + 0.003244071966037154, + 0.0035533492919057608, + 0.005353992804884911, + 0.005031456705182791, + 0.0017720761243253946, + 0.0014646189520135522, + 0.0052770511247217655, + 0.0038519136141985655, + 0.004670339170843363, + 0.005398881621658802, + 0.008549882099032402, + 0.005794026423245668, + 0.004997202195227146, + 0.004424230195581913, + 0.007716261316090822, + 0.008083494380116463, + 0.004580875393003225, + 0.006826599128544331, + 0.009105089120566845, + 0.0035263989120721817, + 0.006533608306199312, + 0.0037873240653425455, + 0.0035712607204914093, + 0.0052521987818181515, + 0.010918964631855488, + 0.017693663015961647, + 0.019698545336723328, + 0.016460761427879333, + 0.06739442050457001, + 0.019100766628980637, + 0.0246004369109869, + 0.06988140195608139, + 0.01171077974140644, + 0.018631843850016594, + 0.02505439892411232, + 0.01756388507783413, + 0.024418998509645462, + 0.02831851877272129, + 0.026364367455244064, + 0.02010934054851532, + 0.00531024020165205, + 0.009567078202962875, + 0.013336250558495522, + 0.006285529118031263, + 0.007959102280437946, + 0.009774023666977882, + 0.006527816876769066, + 0.007408048026263714, + 0.008393639698624611, + 0.010543204843997955, + 0.009207683615386486, + 0.0058330632746219635, + 0.030376503244042397, + 0.02506857179105282, + 0.010903498157858849, + 0.015550470910966396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.009346893057227135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001061352901160717 + ], + "opposite_logprob_tokens": [ + -6.87606143951416 + ], + "logprob_reward": -0.001061352901160717, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001061352901160717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12921014428138733, + 0.04170897603034973, + 0.1262112557888031, + 0.030201712623238564, + 0.056874245405197144, + 0.03997669741511345, + 0.3654562830924988, + 0.1237948015332222, + 0.030753031373023987, + 0.05581290274858475 + ], + "credit_map": [ + 0.012238331139087677, + 0.017136599868535995, + 0.012262649834156036, + 0.014873947016894817, + 0.02064039371907711, + 0.011454390361905098, + 0.023185446858406067, + 0.01741837151348591, + 0.00497241597622633, + 0.003556613577529788, + 0.009825209155678749, + 0.005665839184075594, + 0.003306639613583684, + 0.0023905327543616295, + 0.0061820573173463345, + 0.005809670779854059, + 0.012285087257623672, + 0.021014265716075897, + 0.021067973226308823, + 0.02373330481350422, + 0.009441852569580078, + 0.005772787146270275, + 0.01801043003797531, + 0.014885549433529377, + 0.003422242822125554, + 0.004710181150585413, + 0.005405877251178026, + 0.0041493987664580345, + 0.00177581247407943, + 0.001719332067295909, + 0.00482345512136817, + 0.004195411689579487, + 0.006085485685616732, + 0.005045982543379068, + 0.007945191115140915, + 0.007771361153572798, + 0.004336442798376083, + 0.004783698823302984, + 0.008149644359946251, + 0.012756437994539738, + 0.004771330393850803, + 0.0038698872085660696, + 0.004185976926237345, + 0.003906942438334227, + 0.006359787192195654, + 0.006049969233572483, + 0.005585385486483574, + 0.0052474187687039375, + 0.016095366328954697, + 0.026387713849544525, + 0.03999302163720131, + 0.025998426601290703, + 0.09913980215787888, + 0.023868801072239876, + 0.03421573340892792, + 0.09975739568471909, + 0.007528419606387615, + 0.01380929909646511, + 0.01673400029540062, + 0.014958874322474003, + 0.01983281411230564, + 0.01793234795331955, + 0.017931582406163216, + 0.015067460015416145, + 0.0026288095396012068, + 0.0035440651699900627, + 0.005575763527303934, + 0.0029321066103875637, + 0.005062669515609741, + 0.004027185495942831, + 0.003609301056712866, + 0.003373129526153207, + 0.003939989488571882, + 0.006916336715221405, + 0.006425950676202774, + 0.004376353695988655, + 0.013529977761209011, + 0.005526800639927387, + 0.00651418836787343, + 0.008583304472267628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.001061352901160717 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015224544331431389 + ], + "opposite_logprob_tokens": [ + -6.501522541046143 + ], + "logprob_reward": -0.0015224544331431389, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015224544331431389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16575893759727478, + 0.06514997035264969, + 0.031738411635160446, + 0.018063537776470184, + 0.028272073715925217, + 0.03625524416565895, + 0.3743465542793274, + 0.13855557143688202, + 0.052552465349435806, + 0.08930730819702148 + ], + "credit_map": [ + 0.011076078750193119, + 0.021527012810111046, + 0.011588643305003643, + 0.017033686861395836, + 0.031152602285146713, + 0.01432835403829813, + 0.036238573491573334, + 0.022813962772488594, + 0.005391577258706093, + 0.007912020199000835, + 0.007295716553926468, + 0.017180122435092926, + 0.004382471553981304, + 0.004113360773772001, + 0.011647505685687065, + 0.0072271861135959625, + 0.0028060872573405504, + 0.0036048484034836292, + 0.007657105103135109, + 0.004600424785166979, + 0.0024523339234292507, + 0.0016322304727509618, + 0.004151977598667145, + 0.004833398852497339, + 0.0020257807336747646, + 0.0020066762808710337, + 0.002753431675955653, + 0.0026154210790991783, + 0.0018109827069565654, + 0.0011771671706810594, + 0.003088005119934678, + 0.0025860704481601715, + 0.0028362013399600983, + 0.0025873961858451366, + 0.003583865240216255, + 0.0030397071968764067, + 0.003145546419546008, + 0.003795346012338996, + 0.004977772478014231, + 0.004306234885007143, + 0.0033982950262725353, + 0.006310078781098127, + 0.004491717088967562, + 0.003589259460568428, + 0.005345385055989027, + 0.0029205642640590668, + 0.005424691829830408, + 0.004775249864906073, + 0.017373288050293922, + 0.028374623507261276, + 0.04665859416127205, + 0.030054420232772827, + 0.09392163902521133, + 0.02773570641875267, + 0.03456416726112366, + 0.09566406905651093, + 0.0084359310567379, + 0.02098034881055355, + 0.017838818952441216, + 0.015841839835047722, + 0.019161406904459, + 0.02368549257516861, + 0.016318429261446, + 0.016293294727802277, + 0.00580727681517601, + 0.006899083033204079, + 0.01144445315003395, + 0.005263098981231451, + 0.005389022175222635, + 0.004999947734177113, + 0.005938285030424595, + 0.006811291910707951, + 0.006402816157788038, + 0.010456271469593048, + 0.013495191000401974, + 0.009449224919080734, + 0.01631760410964489, + 0.006460854317992926, + 0.010095246136188507, + 0.016630087047815323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0015224544331431389 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639833260327578 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.007639833260327578, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639833260327578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2714296281337738, + 0.05768168717622757, + 0.04380017891526222, + 0.05524851009249687, + 0.029032234102487564, + 0.08923806250095367, + 0.1522454172372818, + 0.10741405934095383, + 0.08490045368671417, + 0.10900984704494476 + ], + "credit_map": [ + 0.021462172269821167, + 0.035749707370996475, + 0.024081947281956673, + 0.03138849139213562, + 0.05329683795571327, + 0.024603143334388733, + 0.04238153621554375, + 0.038465775549411774, + 0.00821329653263092, + 0.005524083506315947, + 0.014217293821275234, + 0.006341442000120878, + 0.0030010028276592493, + 0.002893848577514291, + 0.009089411236345768, + 0.008401310071349144, + 0.00544261559844017, + 0.005305876024067402, + 0.010298943147063255, + 0.005627722013741732, + 0.003674347186461091, + 0.00248633767478168, + 0.005500935483723879, + 0.005463402718305588, + 0.005648907274007797, + 0.007919739931821823, + 0.010570843704044819, + 0.007664981763809919, + 0.0026979984249919653, + 0.0023111437913030386, + 0.007822816260159016, + 0.010612080805003643, + 0.003015680005773902, + 0.002985118655487895, + 0.004235436674207449, + 0.0035643300507217646, + 0.003678214270621538, + 0.003085280070081353, + 0.00478525273501873, + 0.003682919777929783, + 0.009294741787016392, + 0.009576666168868542, + 0.01351966243237257, + 0.008338774554431438, + 0.013283848762512207, + 0.009606149047613144, + 0.013203231617808342, + 0.01241498626768589, + 0.011127983219921589, + 0.019664565101265907, + 0.022664545103907585, + 0.0163518525660038, + 0.029537376016378403, + 0.013487393036484718, + 0.0172073096036911, + 0.02220439538359642, + 0.00866308156400919, + 0.014854904264211655, + 0.021962283179163933, + 0.010243738070130348, + 0.015369855798780918, + 0.011522230692207813, + 0.011966473422944546, + 0.012831486761569977, + 0.0075993528589606285, + 0.011811913922429085, + 0.017819911241531372, + 0.008700191043317318, + 0.013235422782599926, + 0.007347288075834513, + 0.009596899151802063, + 0.008789479732513428, + 0.00804073829203844, + 0.013191082514822483, + 0.014260197058320045, + 0.008829737082123756, + 0.027339855208992958, + 0.011746704578399658, + 0.009562903083860874, + 0.016038624569773674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.007639833260327578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644422516226768 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644422516226768, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644422516226768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12141898274421692, + 0.047534916549921036, + 0.24529583752155304, + 0.17938750982284546, + 0.035477492958307266, + 0.03644537553191185, + 0.21417440474033356, + 0.04053812846541405, + 0.03736037388443947, + 0.042367029935121536 + ], + "credit_map": [ + 0.008975635282695293, + 0.015269467607140541, + 0.010420278646051884, + 0.014922299422323704, + 0.024319767951965332, + 0.009960656054317951, + 0.022065255790948868, + 0.01548562478274107, + 0.0033610383979976177, + 0.004751496482640505, + 0.008046790026128292, + 0.010935409925878048, + 0.0025454265996813774, + 0.0024892636574804783, + 0.00992851983755827, + 0.00547697301954031, + 0.024783432483673096, + 0.04234885051846504, + 0.043314460664987564, + 0.04642874002456665, + 0.0185993742197752, + 0.009425118565559387, + 0.03479376435279846, + 0.025602081790566444, + 0.015257546678185463, + 0.029113268479704857, + 0.03454769030213356, + 0.029517238959670067, + 0.008397537283599377, + 0.005820651073008776, + 0.030093930661678314, + 0.02663963846862316, + 0.003595708403736353, + 0.003334329230710864, + 0.005883808713406324, + 0.0036818685475736856, + 0.006075355224311352, + 0.004606676287949085, + 0.003876195987686515, + 0.004423549864441156, + 0.0025658910162746906, + 0.007159701082855463, + 0.004822876304388046, + 0.0038047689013183117, + 0.00455078762024641, + 0.0034231378231197596, + 0.004835170228034258, + 0.005283044185489416, + 0.009454043582081795, + 0.015001374296844006, + 0.023822816088795662, + 0.014866406098008156, + 0.055227652192115784, + 0.01966877281665802, + 0.018557175993919373, + 0.05757616087794304, + 0.0033007576130330563, + 0.004268133547157049, + 0.0068105245009064674, + 0.005095822270959616, + 0.005352727137506008, + 0.005024093668907881, + 0.004628987982869148, + 0.006057081278413534, + 0.0030480881687253714, + 0.005287186708301306, + 0.005643667187541723, + 0.0031472148839384317, + 0.007910537533462048, + 0.003671425161883235, + 0.004470124840736389, + 0.0041821300983428955, + 0.00397618068382144, + 0.0033007620368152857, + 0.007740284316241741, + 0.0036815290804952383, + 0.0067185149528086185, + 0.008124527521431446, + 0.003722682362422347, + 0.005102549679577351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.008644422516226768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, this extreme close-up, static shot captures a moment of culinary joy, imbued with a cinematic quality and a hint of film grain. A polished silver fork, pressed deliberately, its tines gently cracking the taut, glistening white of a perfectly poached egg, resting on a thick slice of golden-brown artisan toast. The moment the membrane is pierced, a vibrant, deep-gold yolk breaks through the surface and begins to ooze out in a slow, viscous, and utterly satisfying flow. This creamy, liquid gold lazily overflows from the egg's edges, landing on the toast's crisp, richly textured crumb and pooling in its crevices. The scene is bathed in soft, warm morning light, filtering in from an off-camera window, creating delicate highlights on the glistening yolk and the metallic sheen of the fork. The shallow depth of field blurs the rustic wooden table in the background, focusing the viewer's attention squarely on the rich textures and movement, creating a tranquil, appetizing, and deeply sensory atmosphere.", + "seed": 30438, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_3093_seed_30438_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.004765099857468158, + "vlm_logprob": -0.004765099857468158, + "vlm_yes_no_margin": 5.95, + "vlm_token_credit": -0.004765099857468158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010452880524098873 + ], + "opposite_logprob_tokens": [ + -5.885452747344971 + ], + "logprob_reward": -0.010452880524098873, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.010452880524098873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07980821281671524, + 0.0422988086938858, + 0.31825220584869385, + 0.016596730798482895, + 0.015617950819432735, + 0.01817331090569496, + 0.04643455147743225, + 0.12940464913845062, + 0.0935458391904831, + 0.23986776173114777 + ], + "credit_map": [ + 0.010308953002095222, + 0.00993841141462326, + 0.011759345419704914, + 0.01419609971344471, + 0.005999652203172445, + 0.006911857519298792, + 0.011343564838171005, + 0.009350329637527466, + 0.006206174846738577, + 0.004180260002613068, + 0.0069771273992955685, + 0.007713263854384422, + 0.003815033473074436, + 0.0031337207183241844, + 0.005401084665209055, + 0.004872143734246492, + 0.04050854220986366, + 0.03928358480334282, + 0.04960805922746658, + 0.059411339461803436, + 0.03180183842778206, + 0.021635478362441063, + 0.04379161074757576, + 0.03221176564693451, + 0.0024292534217238426, + 0.001732026576064527, + 0.0021313910838216543, + 0.0036689392291009426, + 0.001139412634074688, + 0.0015782804694026709, + 0.0023846386466175318, + 0.0015327887376770377, + 0.002245645970106125, + 0.0015631576534360647, + 0.0016318416455760598, + 0.002224200638011098, + 0.002457228023558855, + 0.0012475603725761175, + 0.0027815839275717735, + 0.0014667324721813202, + 0.002917859936133027, + 0.0015063928440213203, + 0.001670872326940298, + 0.003185415407642722, + 0.0017057096119970083, + 0.001131038530729711, + 0.004104317165911198, + 0.0019517046166583896, + 0.0066652665846049786, + 0.008511778898537159, + 0.004350024275481701, + 0.0054781679064035416, + 0.006258400157094002, + 0.0037302100099623203, + 0.006746646016836166, + 0.004694059025496244, + 0.0129052409902215, + 0.023615339770913124, + 0.01591748557984829, + 0.01439235545694828, + 0.020025385543704033, + 0.010669968090951443, + 0.01774921640753746, + 0.014129672199487686, + 0.014909484423696995, + 0.007318633142858744, + 0.017864078283309937, + 0.011955324560403824, + 0.013032305985689163, + 0.0075443340465426445, + 0.010175422765314579, + 0.010746251791715622, + 0.036112990230321884, + 0.02412692829966545, + 0.04900550842285156, + 0.03474630415439606, + 0.02647072821855545, + 0.020063092932105064, + 0.027747631072998047, + 0.021594591438770294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.010452880524098873 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019574069883674383 + ], + "opposite_logprob_tokens": [ + -6.251957416534424 + ], + "logprob_reward": -0.0019574069883674383, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019574069883674383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08895314484834671, + 0.062327850610017776, + 0.1985032856464386, + 0.11632425338029861, + 0.027251046150922775, + 0.03139443323016167, + 0.08219215273857117, + 0.09682504832744598, + 0.07850516587495804, + 0.21772368252277374 + ], + "credit_map": [ + 0.012601260095834732, + 0.01005956344306469, + 0.0090405885130167, + 0.01650436595082283, + 0.011537176556885242, + 0.007350655272603035, + 0.012274078093469143, + 0.009585450403392315, + 0.010010894387960434, + 0.006140846759080887, + 0.0062653315253555775, + 0.01204670500010252, + 0.005819105077534914, + 0.005202371161431074, + 0.009732388891279697, + 0.007110200822353363, + 0.027447499334812164, + 0.019518129527568817, + 0.03483493998646736, + 0.03530706837773323, + 0.01639116182923317, + 0.01746719889342785, + 0.025932619348168373, + 0.021604644134640694, + 0.016761429607868195, + 0.012266113422811031, + 0.010934612713754177, + 0.019989587366580963, + 0.01620045118033886, + 0.013393922708928585, + 0.014602137729525566, + 0.012175997719168663, + 0.0041757007129490376, + 0.0026702352333813906, + 0.003292459761723876, + 0.0036833398044109344, + 0.003448755480349064, + 0.0037695325445383787, + 0.0038157242815941572, + 0.002395295538008213, + 0.0050631859339773655, + 0.003379643661901355, + 0.002889081835746765, + 0.005039937328547239, + 0.004137633368372917, + 0.0026394599117338657, + 0.00555039756000042, + 0.002695087343454361, + 0.011755830608308315, + 0.00871414877474308, + 0.009196145460009575, + 0.014316576533019543, + 0.014945819973945618, + 0.005429555661976337, + 0.010998707264661789, + 0.006835360545665026, + 0.011432229541242123, + 0.017010455951094627, + 0.009602832607924938, + 0.011031288653612137, + 0.017534980550408363, + 0.010216932743787766, + 0.011802379041910172, + 0.008193937130272388, + 0.008920427411794662, + 0.006405979860574007, + 0.009072421118617058, + 0.012361472472548485, + 0.01373290829360485, + 0.006864573340862989, + 0.01136488001793623, + 0.00978249590843916, + 0.026505304500460625, + 0.01495115552097559, + 0.04432861506938934, + 0.032588329166173935, + 0.023065971210598946, + 0.02328292652964592, + 0.029568294063210487, + 0.023433057591319084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.0019574069883674383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008212531101889908 + ], + "opposite_logprob_tokens": [ + -7.125821113586426 + ], + "logprob_reward": -0.0008212531101889908, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008212531101889908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13276776671409607, + 0.028589289635419846, + 0.08183516561985016, + 0.02889147959649563, + 0.023253705352544785, + 0.047242820262908936, + 0.1520811915397644, + 0.08975837379693985, + 0.18149541318416595, + 0.23408479988574982 + ], + "credit_map": [ + 0.023327607661485672, + 0.014371908269822598, + 0.011979792267084122, + 0.026502633467316628, + 0.010615818202495575, + 0.009504814632236958, + 0.0227386262267828, + 0.013726554811000824, + 0.003946056589484215, + 0.0031446428038179874, + 0.0028907491359859705, + 0.005147484131157398, + 0.003119739471003413, + 0.0027294140309095383, + 0.004450387321412563, + 0.0031608175486326218, + 0.011897960677742958, + 0.008264969103038311, + 0.013468883000314236, + 0.01754835434257984, + 0.0061907037161290646, + 0.007242291700094938, + 0.009896439500153065, + 0.007325558457523584, + 0.003870203625410795, + 0.003609124105423689, + 0.003277202369645238, + 0.005253288894891739, + 0.0025266732554882765, + 0.002410471672192216, + 0.004792466759681702, + 0.0031520479824393988, + 0.0033679893240332603, + 0.003838376374915242, + 0.00243989541195333, + 0.0036775453481823206, + 0.002933490090072155, + 0.002619366394355893, + 0.0024252457078546286, + 0.0019517960026860237, + 0.006729130633175373, + 0.005565433297306299, + 0.0056497473269701, + 0.010730513371527195, + 0.004441752564162016, + 0.0026570255868136883, + 0.006981954909861088, + 0.004487264435738325, + 0.022662948817014694, + 0.027692683041095734, + 0.010579378344118595, + 0.021585306152701378, + 0.01852656900882721, + 0.01298647839576006, + 0.024759668856859207, + 0.013288167305290699, + 0.012675443664193153, + 0.011532577686011791, + 0.007786895614117384, + 0.01164743397384882, + 0.015990564599633217, + 0.011574462056159973, + 0.00965203158557415, + 0.008898966014385223, + 0.023899847641587257, + 0.015501291491091251, + 0.0201646126806736, + 0.021742023527622223, + 0.029191642999649048, + 0.012316422536969185, + 0.037087492644786835, + 0.02159208431839943, + 0.029137922450900078, + 0.016876595094799995, + 0.035328760743141174, + 0.03358234465122223, + 0.02788621559739113, + 0.026912109926342964, + 0.03706246614456177, + 0.027298400178551674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.0008212531101889908 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955214146524668 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955214146524668, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955214146524668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06724637746810913, + 0.05598381906747818, + 0.2405814528465271, + 0.027815338224172592, + 0.01965221017599106, + 0.019166866317391396, + 0.04498877376317978, + 0.23769837617874146, + 0.05975325033068657, + 0.2271135151386261 + ], + "credit_map": [ + 0.011524150148034096, + 0.005465512629598379, + 0.007548814173787832, + 0.012898290529847145, + 0.006673541851341724, + 0.00607130816206336, + 0.010502692312002182, + 0.006562072318047285, + 0.007674543187022209, + 0.005121571011841297, + 0.007298168260604143, + 0.009487285278737545, + 0.005184517242014408, + 0.004771786276251078, + 0.009111879393458366, + 0.0073340702801942825, + 0.03268125280737877, + 0.02379666082561016, + 0.0431564636528492, + 0.0442238412797451, + 0.025479871779680252, + 0.015547980554401875, + 0.03247590363025665, + 0.023219482973217964, + 0.003676431253552437, + 0.0033577033318579197, + 0.0035304201301187277, + 0.004277888219803572, + 0.0031750609632581472, + 0.002782513154670596, + 0.0035808300599455833, + 0.003434491576626897, + 0.002735715126618743, + 0.0031851695384830236, + 0.002412669127807021, + 0.0031441044993698597, + 0.0018433657241985202, + 0.001734048710204661, + 0.0028584953397512436, + 0.001738641643896699, + 0.0030281676445156336, + 0.0015492039965465665, + 0.0029752221889793873, + 0.0037898491136729717, + 0.00125376065261662, + 0.00235958443954587, + 0.002587215742096305, + 0.001623861608095467, + 0.0049514491111040115, + 0.008722850121557713, + 0.004080150276422501, + 0.006355005316436291, + 0.005185144953429699, + 0.002771160565316677, + 0.007942166179418564, + 0.004980843979865313, + 0.03306743502616882, + 0.061322472989559174, + 0.02825881354510784, + 0.03534906357526779, + 0.022960003465414047, + 0.016292504966259003, + 0.018663685768842697, + 0.021784404292702675, + 0.008338229730725288, + 0.004526167176663876, + 0.010637084022164345, + 0.0074339066632092, + 0.010881874710321426, + 0.004926455207169056, + 0.0066652740351855755, + 0.00634425925090909, + 0.024829810485243797, + 0.017466537654399872, + 0.03532613813877106, + 0.03826477378606796, + 0.027274109423160553, + 0.02446521818637848, + 0.03282913565635681, + 0.026657791808247566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.005955214146524668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463874451816082 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.00463874451816082, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463874451816082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058187276124954224, + 0.09724180400371552, + 0.38104262948036194, + 0.07384222000837326, + 0.031321220099925995, + 0.06141217425465584, + 0.0246257483959198, + 0.0818239077925682, + 0.04245660826563835, + 0.14804647862911224 + ], + "credit_map": [ + 0.008451438508927822, + 0.006177939940243959, + 0.009314783848822117, + 0.011647758074104786, + 0.005124852526932955, + 0.004358356352895498, + 0.006863436661660671, + 0.006248703226447105, + 0.013163378462195396, + 0.010203203186392784, + 0.01342411432415247, + 0.015368352644145489, + 0.009880957193672657, + 0.008887899108231068, + 0.013777527026832104, + 0.012536365538835526, + 0.05005161464214325, + 0.04165976122021675, + 0.0690995529294014, + 0.07021177560091019, + 0.03721747547388077, + 0.024781828746199608, + 0.0504649318754673, + 0.03755567595362663, + 0.01100460160523653, + 0.008464694954454899, + 0.012939208187162876, + 0.011298416182398796, + 0.0051846979185938835, + 0.005905217956751585, + 0.009907147847115993, + 0.009138235822319984, + 0.0041695572435855865, + 0.00421911058947444, + 0.004571886267513037, + 0.004361119586974382, + 0.002477966947481036, + 0.002379814861342311, + 0.005112102255225182, + 0.004029662348330021, + 0.009645945392549038, + 0.005813300143927336, + 0.008780866861343384, + 0.016091743484139442, + 0.003319047624245286, + 0.004594814497977495, + 0.008070400916039944, + 0.00509605510160327, + 0.0034009935334324837, + 0.0038383477367460728, + 0.002655522432178259, + 0.003378765657544136, + 0.0030077260453253984, + 0.002147551393136382, + 0.0036676116287708282, + 0.0025292285718023777, + 0.011260694824159145, + 0.006039475556463003, + 0.007274981588125229, + 0.011447914876043797, + 0.015618966892361641, + 0.008609066717326641, + 0.012249233201146126, + 0.009323570877313614, + 0.006218357011675835, + 0.004054886754602194, + 0.0051997629925608635, + 0.004865839146077633, + 0.009211808443069458, + 0.004595483653247356, + 0.004295683931559324, + 0.004014784935861826, + 0.021195344626903534, + 0.015328843146562576, + 0.027713650837540627, + 0.019742410629987717, + 0.015746189281344414, + 0.020246239379048347, + 0.015621974132955074, + 0.012451826594769955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.00463874451816082 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, this extreme close-up, static shot captures a moment of culinary joy, imbued with a cinematic quality and a hint of film grain. A polished silver fork, pressed deliberately, its tines gently cracking the taut, glistening white of a perfectly poached egg, resting on a thick slice of golden-brown artisan toast. The moment the membrane is pierced, a vibrant, deep-gold yolk breaks through the surface and begins to ooze out in a slow, viscous, and utterly satisfying flow. This creamy, liquid gold lazily overflows from the egg's edges, landing on the toast's crisp, richly textured crumb and pooling in its crevices. The scene is bathed in soft, warm morning light, filtering in from an off-camera window, creating delicate highlights on the glistening yolk and the metallic sheen of the fork. The shallow depth of field blurs the rustic wooden table in the background, focusing the viewer's attention squarely on the rich textures and movement, creating a tranquil, appetizing, and deeply sensory atmosphere.", + "seed": 30439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_3093_seed_30439_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.008927757595665753, + "vlm_logprob": -0.008927757595665753, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.008927757595665753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021831197664141655 + ], + "opposite_logprob_tokens": [ + -4.271831035614014 + ], + "logprob_reward": -0.021831197664141655, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.021831197664141655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5227928757667542, + 0.0373784601688385, + 0.03139081224799156, + 0.14940395951271057, + 0.04897066950798035, + 0.02915041707456112, + 0.05655204504728317, + 0.03023630939424038, + 0.023084284737706184, + 0.07104028761386871 + ], + "credit_map": [ + 0.05086826533079147, + 0.05800867825746536, + 0.054061539471149445, + 0.06369415670633316, + 0.08892670273780823, + 0.08039219677448273, + 0.05778515338897705, + 0.06905610114336014, + 0.004611829295754433, + 0.004739785101264715, + 0.005113665945827961, + 0.005768405273556709, + 0.0036423339042812586, + 0.004264927469193935, + 0.004585120361298323, + 0.00465239305049181, + 0.003165846224874258, + 0.0035135543439537287, + 0.0050251539796590805, + 0.005110385362058878, + 0.003173097036778927, + 0.0025338251143693924, + 0.00492133991792798, + 0.003947608172893524, + 0.014229459688067436, + 0.011486879549920559, + 0.024072883650660515, + 0.015654977411031723, + 0.023121491074562073, + 0.03305186331272125, + 0.011969645507633686, + 0.01581674814224243, + 0.006225633434951305, + 0.005297193303704262, + 0.008482675068080425, + 0.005804447457194328, + 0.004841444082558155, + 0.005004174541682005, + 0.006058443803340197, + 0.007256652694195509, + 0.0030833485070616007, + 0.0025209251325577497, + 0.0030335774645209312, + 0.0034288475289940834, + 0.0032962029799818993, + 0.0018510420341044664, + 0.005662946961820126, + 0.006273525767028332, + 0.005064176395535469, + 0.005346032325178385, + 0.007019398268312216, + 0.007167440373450518, + 0.007545065600425005, + 0.007798017468303442, + 0.008852017112076283, + 0.007759891916066408, + 0.0014044997515156865, + 0.001956767635419965, + 0.0021778715308755636, + 0.0020896673668175936, + 0.00257382751442492, + 0.006244342308491468, + 0.008966057561337948, + 0.004823272116482258, + 0.0015003333101049066, + 0.002087034285068512, + 0.0022373341489583254, + 0.002751800464466214, + 0.004252796061336994, + 0.005569250788539648, + 0.002279447391629219, + 0.002406287007033825, + 0.006083931773900986, + 0.00945449247956276, + 0.017122169956564903, + 0.008187645114958286, + 0.008644417859613895, + 0.008436313830316067, + 0.005602483171969652, + 0.007508825045078993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.021831197664141655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028398206923156977 + ], + "opposite_logprob_tokens": [ + -5.877840042114258 + ], + "logprob_reward": -0.0028398206923156977, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028398206923156977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22528333961963654, + 0.07763242721557617, + 0.04050886258482933, + 0.26366275548934937, + 0.08172484487295151, + 0.10552757978439331, + 0.041833844035863876, + 0.07364439964294434, + 0.0572725273668766, + 0.03290938958525658 + ], + "credit_map": [ + 0.023301107808947563, + 0.025070032104849815, + 0.021256113424897194, + 0.02699076011776924, + 0.041693512350320816, + 0.032584577798843384, + 0.026508882641792297, + 0.02787836454808712, + 0.0068160658702254295, + 0.008642462082207203, + 0.009537880308926105, + 0.009750640019774437, + 0.010267658159136772, + 0.009187090210616589, + 0.013210471719503403, + 0.01022016629576683, + 0.00511409156024456, + 0.003838670440018177, + 0.004150619730353355, + 0.004862544126808643, + 0.006881823297590017, + 0.0036690745037049055, + 0.006085895001888275, + 0.00590615114197135, + 0.023754822090268135, + 0.024702569469809532, + 0.04586496949195862, + 0.027013596147298813, + 0.038396064192056656, + 0.05264716595411301, + 0.021766606718301773, + 0.029517000541090965, + 0.008674079552292824, + 0.009213283658027649, + 0.011405796743929386, + 0.008886760100722313, + 0.009655754081904888, + 0.010806390084326267, + 0.011146855540573597, + 0.01193593442440033, + 0.010032535530626774, + 0.011064658872783184, + 0.010435974225401878, + 0.010946457274258137, + 0.011289973743259907, + 0.007095671724528074, + 0.02259690687060356, + 0.022065410390496254, + 0.0045979986898601055, + 0.003658176399767399, + 0.006815093569457531, + 0.00598170468583703, + 0.0052126203663647175, + 0.004435327369719744, + 0.005863861180841923, + 0.005269065033644438, + 0.0041276877745985985, + 0.003485403722152114, + 0.003822692669928074, + 0.004254513885825872, + 0.005625725258141756, + 0.010585098527371883, + 0.024700267240405083, + 0.017043016850948334, + 0.003393910126760602, + 0.002726042876020074, + 0.0028710055630654097, + 0.0030849098693579435, + 0.00990628357976675, + 0.0203421488404274, + 0.006762936245650053, + 0.008185294456779957, + 0.0026382943615317345, + 0.005116789136081934, + 0.004127084277570248, + 0.004933507181704044, + 0.003768523456528783, + 0.0048850709572434425, + 0.004070518538355827, + 0.0033696074970066547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0028398206923156977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001522811478935182 + ], + "opposite_logprob_tokens": [ + -6.501523017883301 + ], + "logprob_reward": -0.001522811478935182, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001522811478935182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10475622117519379, + 0.03353100270032883, + 0.07502011954784393, + 0.16132502257823944, + 0.06780951470136642, + 0.055976614356040955, + 0.09431526064872742, + 0.05775399133563042, + 0.26689809560775757, + 0.08261419832706451 + ], + "credit_map": [ + 0.010099036619067192, + 0.011575382202863693, + 0.008633334189653397, + 0.012942547909915447, + 0.019478855654597282, + 0.013575155287981033, + 0.01396825723350048, + 0.014483657665550709, + 0.0038781699258834124, + 0.00373347126878798, + 0.004701375961303711, + 0.005084540694952011, + 0.0032623654697090387, + 0.0029671380762010813, + 0.005730313248932362, + 0.0041736275888979435, + 0.006347971502691507, + 0.012168437242507935, + 0.009064901620149612, + 0.011842726729810238, + 0.006842137314379215, + 0.009217540733516216, + 0.009501367807388306, + 0.010035034269094467, + 0.01896665245294571, + 0.018763521686196327, + 0.029191549867391586, + 0.017982177436351776, + 0.022321468219161034, + 0.017986206337809563, + 0.015475396066904068, + 0.02063804119825363, + 0.0077342982403934, + 0.006586993578821421, + 0.006246075499802828, + 0.0057268161326646805, + 0.008626195602118969, + 0.015747448429465294, + 0.008365745656192303, + 0.008775943890213966, + 0.004845183342695236, + 0.003815164789557457, + 0.006160355638712645, + 0.0050396607257425785, + 0.004449116066098213, + 0.00326370308175683, + 0.012876011431217194, + 0.01552741788327694, + 0.008462131023406982, + 0.010012860409915447, + 0.012878765352070332, + 0.013477596454322338, + 0.012959963642060757, + 0.010616757906973362, + 0.015194294042885303, + 0.010712890885770321, + 0.0051798271015286446, + 0.004586444701999426, + 0.008015017956495285, + 0.0070034898817539215, + 0.0070436373353004456, + 0.01051421370357275, + 0.008277906104922295, + 0.0071334498934447765, + 0.015871170908212662, + 0.01923516020178795, + 0.012739818543195724, + 0.019435569643974304, + 0.02851329743862152, + 0.08798487484455109, + 0.04336962848901749, + 0.03974857181310654, + 0.0072124009020626545, + 0.007847419008612633, + 0.011287963017821312, + 0.012127234600484371, + 0.011882291175425053, + 0.012828292325139046, + 0.011035987176001072, + 0.008392608724534512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.001522811478935182 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796871803700924 + ], + "opposite_logprob_tokens": [ + -4.634797096252441 + ], + "logprob_reward": -0.009796871803700924, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796871803700924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13729871809482574, + 0.09936219453811646, + 0.0970199704170227, + 0.18299835920333862, + 0.08171825855970383, + 0.051987987011671066, + 0.06326794624328613, + 0.11734902113676071, + 0.09861049056053162, + 0.0703870877623558 + ], + "credit_map": [ + 0.013266908004879951, + 0.0118534741923213, + 0.019327249377965927, + 0.019219553098082542, + 0.015089140273630619, + 0.017498284578323364, + 0.023734312504529953, + 0.01730978861451149, + 0.008519673720002174, + 0.00962921790778637, + 0.016461895778775215, + 0.016095824539661407, + 0.011321538127958775, + 0.009555620141327381, + 0.018274080008268356, + 0.009504348039627075, + 0.007006749976426363, + 0.012045111507177353, + 0.014056355692446232, + 0.014995705336332321, + 0.010335749946534634, + 0.011171470396220684, + 0.017277253791689873, + 0.010131574235856533, + 0.018981434404850006, + 0.01576877012848854, + 0.023457149043679237, + 0.014275375753641129, + 0.027608921751379967, + 0.046894706785678864, + 0.016225742176175117, + 0.019786253571510315, + 0.008398477919399738, + 0.007556440308690071, + 0.014757107943296432, + 0.00813857652246952, + 0.006254413165152073, + 0.014514520764350891, + 0.010578591376543045, + 0.011520135216414928, + 0.005046098027378321, + 0.0037527133245021105, + 0.006968264002352953, + 0.005806050729006529, + 0.003869907232001424, + 0.005726889241486788, + 0.00926371943205595, + 0.011554348282516003, + 0.004824177362024784, + 0.0060352748259902, + 0.006994002964347601, + 0.009887795895338058, + 0.008375067263841629, + 0.00836531724780798, + 0.01107251551002264, + 0.007713791914284229, + 0.007500444073230028, + 0.007915325462818146, + 0.007104544900357723, + 0.01046033389866352, + 0.013622893951833248, + 0.021836355328559875, + 0.025866175070405006, + 0.023042937740683556, + 0.006279362365603447, + 0.007088675629347563, + 0.004976720083504915, + 0.008687247522175312, + 0.02407071925699711, + 0.025154011324048042, + 0.012128113768994808, + 0.01022564060986042, + 0.006735103204846382, + 0.006354426499456167, + 0.012249056249856949, + 0.01046403031796217, + 0.009291903115808964, + 0.009131724946200848, + 0.00768456282094121, + 0.00847628340125084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.009796871803700924 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648086339235306 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008648086339235306, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648086339235306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3000718057155609, + 0.08432332426309586, + 0.10597532987594604, + 0.07352620363235474, + 0.08479838073253632, + 0.04446043446660042, + 0.07548553496599197, + 0.08739273250102997, + 0.09019166976213455, + 0.05377461388707161 + ], + "credit_map": [ + 0.03240012377500534, + 0.03135499730706215, + 0.0414079949259758, + 0.03691747039556503, + 0.04268958792090416, + 0.04894482344388962, + 0.033647894859313965, + 0.032708898186683655, + 0.009861866943538189, + 0.009993314743041992, + 0.010108879767358303, + 0.015482242219150066, + 0.008204050362110138, + 0.009267930872738361, + 0.013367386534810066, + 0.008037650026381016, + 0.010307391174137592, + 0.010036748833954334, + 0.020812872797250748, + 0.015841253101825714, + 0.009494380094110966, + 0.011545144952833652, + 0.014846870675683022, + 0.013090666383504868, + 0.009775612503290176, + 0.00819734949618578, + 0.011182744055986404, + 0.008120167069137096, + 0.009063711389899254, + 0.006490013562142849, + 0.010533202439546585, + 0.01016340870410204, + 0.013449962250888348, + 0.007562704849988222, + 0.010083426721394062, + 0.007407130673527718, + 0.009069031104445457, + 0.013482734560966492, + 0.011733525432646275, + 0.012009862810373306, + 0.004296966828405857, + 0.005594372283667326, + 0.005822411738336086, + 0.005590108223259449, + 0.005341468378901482, + 0.006916687823832035, + 0.0054163746535778046, + 0.00548204593360424, + 0.00724437041208148, + 0.0061715589836239815, + 0.017980773001909256, + 0.008508819155395031, + 0.0074429563246667385, + 0.00908205471932888, + 0.006601236294955015, + 0.012453768402338028, + 0.005375184118747711, + 0.0041924589313566685, + 0.00438847579061985, + 0.005917105823755264, + 0.007234647870063782, + 0.01291747484356165, + 0.027850955724716187, + 0.01951642893254757, + 0.006123508792370558, + 0.006430027075111866, + 0.007492358330637217, + 0.006696636788547039, + 0.012578573077917099, + 0.022999843582510948, + 0.017758401110768318, + 0.010112310759723186, + 0.005710368510335684, + 0.006325936876237392, + 0.007172075565904379, + 0.007119109854102135, + 0.008543269708752632, + 0.0056572481989860535, + 0.0063975038938224316, + 0.006849096622318029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.008648086339235306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, this extreme close-up, static shot captures a moment of culinary joy, imbued with a cinematic quality and a hint of film grain. A polished silver fork, pressed deliberately, its tines gently cracking the taut, glistening white of a perfectly poached egg, resting on a thick slice of golden-brown artisan toast. The moment the membrane is pierced, a vibrant, deep-gold yolk breaks through the surface and begins to ooze out in a slow, viscous, and utterly satisfying flow. This creamy, liquid gold lazily overflows from the egg's edges, landing on the toast's crisp, richly textured crumb and pooling in its crevices. The scene is bathed in soft, warm morning light, filtering in from an off-camera window, creating delicate highlights on the glistening yolk and the metallic sheen of the fork. The shallow depth of field blurs the rustic wooden table in the background, focusing the viewer's attention squarely on the rich textures and movement, creating a tranquil, appetizing, and deeply sensory atmosphere.", + "seed": 30440, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_3/index_3093_seed_30440_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.005353982863016427, + "vlm_logprob": -0.005353982863016427, + "vlm_yes_no_margin": 5.725, + "vlm_token_credit": -0.005353982863016427, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01113557443022728 + ], + "opposite_logprob_tokens": [ + -6.011135578155518 + ], + "logprob_reward": -0.01113557443022728, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.01113557443022728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0627564936876297, + 0.040739934891462326, + 0.01675877906382084, + 0.03246336057782173, + 0.06781653314828873, + 0.07078192383050919, + 0.0922095775604248, + 0.06387361139059067, + 0.47428062558174133, + 0.07831919938325882 + ], + "credit_map": [ + 0.008938402868807316, + 0.007276732008904219, + 0.009456443600356579, + 0.006879865191876888, + 0.003848826279863715, + 0.004780778661370277, + 0.011074041947722435, + 0.01050140243023634, + 0.003992028068751097, + 0.0051327720284461975, + 0.006157051771879196, + 0.006514595355838537, + 0.0026937266811728477, + 0.004106021951884031, + 0.0070393807254731655, + 0.005104354582726955, + 0.001935289241373539, + 0.0024190868716686964, + 0.003970454912632704, + 0.001875683548860252, + 0.000745072087738663, + 0.0015538485022261739, + 0.0018656770698726177, + 0.0023936659563332796, + 0.004672987852245569, + 0.003252028953284025, + 0.006473638582974672, + 0.003367185592651367, + 0.0017397997435182333, + 0.0031521893106400967, + 0.004788122605532408, + 0.00501740537583828, + 0.011146622709929943, + 0.00833752378821373, + 0.014274432323873043, + 0.009650043211877346, + 0.0031575041357427835, + 0.006501265801489353, + 0.006099145393818617, + 0.008649990893900394, + 0.01485839206725359, + 0.009811444208025932, + 0.018071725964546204, + 0.005872991867363453, + 0.003058431204408407, + 0.003928157035261393, + 0.00868983194231987, + 0.006490950006991625, + 0.010274992324411869, + 0.012065478600561619, + 0.004593288525938988, + 0.004954155068844557, + 0.02353169582784176, + 0.022343438118696213, + 0.008976594544947147, + 0.005469922441989183, + 0.006925877183675766, + 0.009579033590853214, + 0.006549987010657787, + 0.004923895932734013, + 0.010326136834919453, + 0.01176904421299696, + 0.0076683503575623035, + 0.006131283938884735, + 0.047790247946977615, + 0.17446662485599518, + 0.009703452698886395, + 0.01003685686737299, + 0.024555305019021034, + 0.16841793060302734, + 0.020048916339874268, + 0.019261235371232033, + 0.010482831858098507, + 0.009018566459417343, + 0.0055455053225159645, + 0.004703325219452381, + 0.019652536138892174, + 0.015428546816110611, + 0.007344840560108423, + 0.006143040489405394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.01113557443022728 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002836611121892929 + ], + "opposite_logprob_tokens": [ + -5.87783670425415 + ], + "logprob_reward": -0.002836611121892929, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002836611121892929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07529374212026596, + 0.06330864131450653, + 0.03901508450508118, + 0.04396495223045349, + 0.07995880395174026, + 0.056087058037519455, + 0.04000779986381531, + 0.05668796971440315, + 0.45461440086364746, + 0.09106157720088959 + ], + "credit_map": [ + 0.008581145666539669, + 0.009461458772420883, + 0.010015033185482025, + 0.007656781934201717, + 0.00402269558981061, + 0.008178161457180977, + 0.013182410039007664, + 0.01419606152921915, + 0.007732503581792116, + 0.00898442417383194, + 0.013190621510148048, + 0.006493692751973867, + 0.0036346910055726767, + 0.006413080729544163, + 0.006811246275901794, + 0.010048375464975834, + 0.004812050610780716, + 0.007274276111274958, + 0.006272037047892809, + 0.0044766259379684925, + 0.0022194059565663338, + 0.003255642019212246, + 0.004614891484379768, + 0.00609015766531229, + 0.005747349001467228, + 0.004699856508523226, + 0.0065046935342252254, + 0.005197103600949049, + 0.003507377579808235, + 0.0046491604298353195, + 0.006942492909729481, + 0.006716920994222164, + 0.022732021287083626, + 0.022518452256917953, + 0.007786866743117571, + 0.007163377944380045, + 0.0033695970196276903, + 0.0060224030166864395, + 0.004781062249094248, + 0.005585019942373037, + 0.007960335351526737, + 0.004843744914978743, + 0.014570288360118866, + 0.007179365027695894, + 0.0030889841727912426, + 0.0038001425564289093, + 0.009267144836485386, + 0.005377051420509815, + 0.004869166295975447, + 0.0047280630096793175, + 0.007087322883307934, + 0.0023804272059351206, + 0.005563759710639715, + 0.007827186025679111, + 0.003387845354154706, + 0.004164028912782669, + 0.007780181244015694, + 0.00975546333938837, + 0.005212961696088314, + 0.004986745771020651, + 0.00730214873328805, + 0.01040074322372675, + 0.006172613240778446, + 0.005077114328742027, + 0.04575284942984581, + 0.1662246584892273, + 0.010752184316515923, + 0.009530849754810333, + 0.02514442428946495, + 0.158602774143219, + 0.01837983727455139, + 0.02022680826485157, + 0.009230822324752808, + 0.010075168684124947, + 0.010370563715696335, + 0.008233643136918545, + 0.015045708045363426, + 0.015398753806948662, + 0.011011454276740551, + 0.01169546414166689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.002836611121892929 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001949554425664246 + ], + "opposite_logprob_tokens": [ + -6.251949787139893 + ], + "logprob_reward": -0.001949554425664246, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001949554425664246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04325104132294655, + 0.009047633968293667, + 0.008821394294500351, + 0.01248104777187109, + 0.05885162949562073, + 0.01745455153286457, + 0.0603511780500412, + 0.08761271834373474, + 0.6551027297973633, + 0.047026075422763824 + ], + "credit_map": [ + 0.006767575163394213, + 0.00851089134812355, + 0.008116364479064941, + 0.0035452041774988174, + 0.0017999446718022227, + 0.0038792830891907215, + 0.004487457685172558, + 0.006144320126622915, + 0.0011216722195968032, + 0.001127864932641387, + 0.0014352718135342002, + 0.0012256046757102013, + 0.0008050434989854693, + 0.00090032845037058, + 0.0013488916447386146, + 0.001082957023754716, + 0.001220382866449654, + 0.000986343715339899, + 0.0015001894207671285, + 0.0009543490596115589, + 0.00046893145190551877, + 0.0009647521073929965, + 0.0010451271664351225, + 0.0016813183901831508, + 0.0016540603246539831, + 0.0014873422915115952, + 0.002186171943321824, + 0.0015771868638694286, + 0.0006434100214391947, + 0.0012260725488886237, + 0.0016294611850753427, + 0.0020773420110344887, + 0.018045324832201004, + 0.018881715834140778, + 0.004680366721004248, + 0.004381496924906969, + 0.0023055793717503548, + 0.0038962061516940594, + 0.003006894141435623, + 0.0036540478467941284, + 0.002469336148351431, + 0.001914061140269041, + 0.00408488605171442, + 0.0015224611852318048, + 0.0013892127899453044, + 0.0015158193418756127, + 0.002095475560054183, + 0.002463299548253417, + 0.006919190287590027, + 0.007577275857329369, + 0.004249495919793844, + 0.002621553372591734, + 0.012145440094172955, + 0.01611322909593582, + 0.005334269255399704, + 0.005390724167227745, + 0.012710127979516983, + 0.016701513901352882, + 0.008762997575104237, + 0.006404116749763489, + 0.01072826236486435, + 0.016954904422163963, + 0.008492154069244862, + 0.006858645938336849, + 0.06589950621128082, + 0.23863202333450317, + 0.01598702371120453, + 0.013382920064032078, + 0.03673779219388962, + 0.2299201637506485, + 0.02676474303007126, + 0.02777855284512043, + 0.005237170495092869, + 0.005368489306420088, + 0.0036239936016499996, + 0.004052051343023777, + 0.010730190202593803, + 0.007291603367775679, + 0.006417501252144575, + 0.004305075388401747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.001949554425664246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104998894035816 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.004104998894035816, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104998894035816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042072687298059464, + 0.019978776574134827, + 0.013932149857282639, + 0.038952574133872986, + 0.14731977880001068, + 0.04030187055468559, + 0.07285367697477341, + 0.03606724739074707, + 0.5601757168769836, + 0.028345482423901558 + ], + "credit_map": [ + 0.005267498083412647, + 0.005795879755169153, + 0.007006276398897171, + 0.003920001443475485, + 0.00195473269559443, + 0.005133946891874075, + 0.006472126115113497, + 0.006522227544337511, + 0.0026413272134959698, + 0.0033146291971206665, + 0.002738429233431816, + 0.0033916449174284935, + 0.0010870026890188456, + 0.001529889996163547, + 0.0023416667245328426, + 0.0029341867193579674, + 0.0017976437229663134, + 0.0014996189856901765, + 0.0035050157457590103, + 0.001485759043134749, + 0.0006679443758912385, + 0.0013929199194535613, + 0.0016835633432492614, + 0.0018996858270838857, + 0.004961477126926184, + 0.004360412713140249, + 0.01239085104316473, + 0.004195617511868477, + 0.0017392653971910477, + 0.0038127913139760494, + 0.0033815165515989065, + 0.004110642708837986, + 0.04240565747022629, + 0.04378661513328552, + 0.012878315523266792, + 0.013567632995545864, + 0.005414607003331184, + 0.009954588487744331, + 0.008710121735930443, + 0.010602246969938278, + 0.006947619374841452, + 0.0041952370665967464, + 0.010175238363444805, + 0.003543567843735218, + 0.002094583585858345, + 0.0021604534704238176, + 0.005714667495340109, + 0.005470503121614456, + 0.00850611086934805, + 0.010536102578043938, + 0.0069579267874360085, + 0.004749824292957783, + 0.01471125427633524, + 0.01564926654100418, + 0.006759727373719215, + 0.004983467049896717, + 0.004270621109753847, + 0.006113656330853701, + 0.004358106758445501, + 0.0026950242463499308, + 0.003965349402278662, + 0.007843764498829842, + 0.003651512088254094, + 0.0031692124903202057, + 0.05611734837293625, + 0.20290417969226837, + 0.015716440975666046, + 0.012573793530464172, + 0.03177754208445549, + 0.19374090433120728, + 0.023556753993034363, + 0.02378881350159645, + 0.0030688433907926083, + 0.003311769338324666, + 0.0022421851754188538, + 0.0015706145204603672, + 0.008764563128352165, + 0.005036105401813984, + 0.0025197335053235292, + 0.0018316684290766716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.004104998894035816 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a poached egg on artisan toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished silver fork visible pressing into the egg yolk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the yolk break and ooze out in a slow, viscous flow onto the toast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in soft, warm morning light creating highlights on the yolk and fork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static with a shallow depth of field that blurs the wooden table background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743175443261862 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006743175443261862, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743175443261862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022737525403499603, + 0.016989441588521004, + 0.010576792061328888, + 0.015522138215601444, + 0.05508467182517052, + 0.012468701228499413, + 0.021140841767191887, + 0.04582766070961952, + 0.7681297659873962, + 0.03152255341410637 + ], + "credit_map": [ + 0.0037309464532881975, + 0.0030148588120937347, + 0.004722542595118284, + 0.001726643298752606, + 0.0013135651824995875, + 0.002246143762022257, + 0.003398279892280698, + 0.002584542613476515, + 0.002124967984855175, + 0.0026616586837917566, + 0.004217714071273804, + 0.0016470253467559814, + 0.0006030622171238065, + 0.0013237078674137592, + 0.0017968558240681887, + 0.002614447148516774, + 0.0018586706137284636, + 0.0013159411028027534, + 0.0018977243453264236, + 0.0009319306700490415, + 0.0004932787269353867, + 0.0011104861041530967, + 0.0012288884026929736, + 0.0017398713389411569, + 0.001982836052775383, + 0.0020326024387031794, + 0.004951489623636007, + 0.0019850698299705982, + 0.0008181207813322544, + 0.0015003494918346405, + 0.0011214452097192407, + 0.0011302242055535316, + 0.009381385520100594, + 0.007858513854444027, + 0.011147940531373024, + 0.007579450495541096, + 0.0024602459743618965, + 0.005377287045121193, + 0.004760571289807558, + 0.006519267335534096, + 0.0017978663090616465, + 0.0011346619576215744, + 0.00297199422493577, + 0.0015138478484004736, + 0.0007782239699736238, + 0.0008878341759555042, + 0.0022142224479466677, + 0.001170049305073917, + 0.0024749303702265024, + 0.0026739640161395073, + 0.0033138133585453033, + 0.001343717915005982, + 0.0032153089996427298, + 0.00405864417552948, + 0.0018604343058541417, + 0.002200025599449873, + 0.005085154436528683, + 0.008465021848678589, + 0.007713963743299246, + 0.0038079170044511557, + 0.005718010477721691, + 0.006309710908681154, + 0.004328214563429356, + 0.004399662371724844, + 0.07881986349821091, + 0.2835158407688141, + 0.016018595546483994, + 0.015159078873693943, + 0.0408535934984684, + 0.2701423764228821, + 0.031345561146736145, + 0.0322747640311718, + 0.0028936509042978287, + 0.0036871680058538914, + 0.003601906355470419, + 0.002790234750136733, + 0.005735407583415508, + 0.0060818917118012905, + 0.002999895717948675, + 0.0037323946598917246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.006743175443261862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_4/index_tensor([5845, 5845])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_4/index_tensor([5845, 5845])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4f5010283790acccacdb95638f35c68dedd2909c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_4/index_tensor([5845, 5845])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, horizontal medium shot, the camera held still, captures a woman in an exquisite lake-blue silk cheongsam. She walks gracefully and leisurely along a winding bluestone path in an ancient courtyard. Her long, lustrous black hair is pulled into a delicate low bun, secured with a delicate jade hairpin. Her features are classically graceful, a faint smile playing on their lips. The cheongsam is tailored to her figure, with a high, buttoned collar and elaborately embroidered with gold and silver thread, featuring blooming peonies and fluttering butterflies. Soft afternoon sunlight filters through the gaps in the courtyard's bamboo grove, casting dappled, swaying shadows on the ground. The light outlines the delicate contours of her figure and imbues the silk and embroidery of the cheongsam with a delicate sheen. In her right hand, she gently grasps a small, round fan, made of translucent white silk, painted with a few elegant orchids in light ink. As she steps and gently shakes her wrist, the fan sways gracefully across her chest. In the background, a corner of a wooden corridor with raised eaves and meticulously arranged Taihu stones can be seen. The entire scene, dominated by lake blue, emerald green, and pewter, exudes a tranquil, elegant, and poetic atmosphere of classical oriental art.", + "seed": 6288, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_4/index_5845_seed_6288_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09034904018044472, + "vlm_logprob": -0.09034904018044472, + "vlm_yes_no_margin": 2.4749999761581423, + "vlm_token_credit": -0.09034904018044472, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07080654799938202 + ], + "opposite_logprob_tokens": [ + -2.9458065032958984 + ], + "logprob_reward": -0.07080654799938202, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07080654799938202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09798454493284225, + 0.6414710879325867, + 0.035742029547691345, + 0.036639533936977386, + 0.028271440416574478, + 0.028362831100821495, + 0.029979411512613297, + 0.034874290227890015, + 0.05112399905920029, + 0.015550905838608742 + ], + "credit_map": [ + 0.00872434675693512, + 0.013182350434362888, + 0.010817854665219784, + 0.013710982166230679, + 0.011385023593902588, + 0.012313276529312134, + 0.01325219962745905, + 0.014598502777516842, + 0.05112829804420471, + 0.084994837641716, + 0.052004002034664154, + 0.11307793855667114, + 0.06521329283714294, + 0.09185199439525604, + 0.11790639162063599, + 0.06529423594474792, + 0.003063485724851489, + 0.007885845378041267, + 0.0028611936140805483, + 0.0043549612164497375, + 0.004722196143120527, + 0.004513292107731104, + 0.004410644061863422, + 0.003930410370230675, + 0.00442546745762229, + 0.007654576562345028, + 0.0026738790329545736, + 0.00401837844401598, + 0.005692740902304649, + 0.004061283078044653, + 0.004377307370305061, + 0.0037358973640948534, + 0.0027924946043640375, + 0.004233478102833033, + 0.0033779351506382227, + 0.003464469453319907, + 0.0027161911129951477, + 0.0036786277778446674, + 0.0032812724821269512, + 0.004726970102638006, + 0.002231902675703168, + 0.0055566043592989445, + 0.0031296960078179836, + 0.002614258322864771, + 0.003538370132446289, + 0.003535438794642687, + 0.003409442724660039, + 0.004347116220742464, + 0.002042221836745739, + 0.004621115047484636, + 0.003052265616133809, + 0.0031492402777075768, + 0.004131374880671501, + 0.004629397764801979, + 0.004839808214455843, + 0.003513982752338052, + 0.0026560083497315645, + 0.0059906779788434505, + 0.0033300688955932856, + 0.005247175227850676, + 0.0038519923109561205, + 0.0044235652312636375, + 0.0059437220916152, + 0.0034310759510844946, + 0.0046361093409359455, + 0.012570274993777275, + 0.004106742329895496, + 0.0072265989147126675, + 0.004948250018060207, + 0.00690841069445014, + 0.00535074295476079, + 0.005376868415623903, + 0.001661184011027217, + 0.004240245092660189, + 0.0015007785987108946, + 0.001801139791496098, + 0.0014225253835320473, + 0.001898219808936119, + 0.0011913763592019677, + 0.0018354356288909912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.07080654799938202 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048684071749448776 + ], + "opposite_logprob_tokens": [ + -3.0486841201782227 + ], + "logprob_reward": -0.048684071749448776, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048684071749448776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09830867499113083, + 0.6961218118667603, + 0.03084176778793335, + 0.02601894736289978, + 0.02464870736002922, + 0.027228450402617455, + 0.02675679139792919, + 0.026542367413640022, + 0.029972920194268227, + 0.01355949230492115 + ], + "credit_map": [ + 0.007759655825793743, + 0.013190913945436478, + 0.010118023492395878, + 0.011060567572712898, + 0.019109709188342094, + 0.013365400023758411, + 0.009947561658918858, + 0.013756858184933662, + 0.05536952614784241, + 0.09083397686481476, + 0.05708407238125801, + 0.12362649291753769, + 0.07004139572381973, + 0.09998222440481186, + 0.12832292914390564, + 0.07086124271154404, + 0.0026404077652841806, + 0.004044898319989443, + 0.0024563802871853113, + 0.00474097765982151, + 0.0050519616343081, + 0.0036184419877827168, + 0.00553483422845602, + 0.002753868931904435, + 0.0022726182360202074, + 0.0028885272331535816, + 0.0025331685319542885, + 0.003040771931409836, + 0.0044547379948198795, + 0.0036512287333607674, + 0.004782600328326225, + 0.002395297633484006, + 0.003151020035147667, + 0.0029579417314380407, + 0.002778768539428711, + 0.003077325876802206, + 0.0028554650489240885, + 0.003002758603543043, + 0.003948301542550325, + 0.002877128543332219, + 0.0017179209971800447, + 0.0042451657354831696, + 0.0023073768243193626, + 0.004071695730090141, + 0.0038120874669402838, + 0.0033410668838769197, + 0.005321970675140619, + 0.002411168534308672, + 0.0018036781111732125, + 0.004697271157056093, + 0.0022602092940360308, + 0.002894802950322628, + 0.0032240322325378656, + 0.0037826532498002052, + 0.004198361653834581, + 0.0038957868237048388, + 0.0020712900441139936, + 0.004291610326617956, + 0.0027244025841355324, + 0.004072504583746195, + 0.003311960259452462, + 0.0032888755667954683, + 0.004094414878636599, + 0.00268731196410954, + 0.0025610881857573986, + 0.006790001876652241, + 0.00309009850025177, + 0.0032927768770605326, + 0.0031227802392095327, + 0.004063725005835295, + 0.00313940504565835, + 0.003913050051778555, + 0.0019550316501408815, + 0.0013589828740805387, + 0.0010573238832876086, + 0.0023392115253955126, + 0.002057454315945506, + 0.0015856858808547258, + 0.0016658356180414557, + 0.0015399691183120012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.048684071749448776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14289624989032745 + ], + "opposite_logprob_tokens": [ + -2.0178961753845215 + ], + "logprob_reward": -0.14289624989032745, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14289624989032745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13131628930568695, + 0.5368822813034058, + 0.05481646955013275, + 0.03700198978185654, + 0.06285210698843002, + 0.07647731900215149, + 0.031889691948890686, + 0.029741110280156136, + 0.025349242612719536, + 0.01367348711937666 + ], + "credit_map": [ + 0.010057829320430756, + 0.01631096936762333, + 0.013486381620168686, + 0.012485318817198277, + 0.027520444244146347, + 0.0185878686606884, + 0.01407668273895979, + 0.01879080943763256, + 0.042880259454250336, + 0.07041358947753906, + 0.044267553836107254, + 0.09551284462213516, + 0.05395948886871338, + 0.07706110179424286, + 0.09876299649477005, + 0.054024454206228256, + 0.0038862605579197407, + 0.006275474093854427, + 0.004236150067299604, + 0.008008060045540333, + 0.00919037964195013, + 0.0072297886945307255, + 0.009735764004290104, + 0.006254593376070261, + 0.003026599297299981, + 0.0032532874029129744, + 0.0029895470943301916, + 0.004036383703351021, + 0.008401838131248951, + 0.005562548525631428, + 0.006301312241703272, + 0.0034304701257497072, + 0.005770700052380562, + 0.0066991522908210754, + 0.0036056123208254576, + 0.007134584244340658, + 0.00932718813419342, + 0.008750025182962418, + 0.014839583076536655, + 0.006725261453539133, + 0.004691008478403091, + 0.007713696453720331, + 0.005937424022704363, + 0.012405428104102612, + 0.011223574168980122, + 0.009458823129534721, + 0.01762996055185795, + 0.007417406886816025, + 0.0017326698871329427, + 0.005202201660722494, + 0.0024567663203924894, + 0.004285438917577267, + 0.004109245724976063, + 0.004266579635441303, + 0.006800439208745956, + 0.0030363514088094234, + 0.002555601764470339, + 0.005689057055860758, + 0.003312024287879467, + 0.004455866292119026, + 0.0031570654828101397, + 0.0036101418081671, + 0.004000620450824499, + 0.002960731741040945, + 0.002080958103761077, + 0.005534677766263485, + 0.0026008104905486107, + 0.0027283683884888887, + 0.0027879984118044376, + 0.003532380796968937, + 0.0028660385869443417, + 0.0032180093694478273, + 0.0019830286037176847, + 0.002011646516621113, + 0.0011159466812387109, + 0.0023323155473917723, + 0.0016697918763384223, + 0.001673156046308577, + 0.001394444378092885, + 0.0014931573532521725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.14289624989032745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032574087381363 + ], + "opposite_logprob_tokens": [ + -2.3503258228302 + ], + "logprob_reward": -0.10032574087381363, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032574087381363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1063288152217865, + 0.5457807779312134, + 0.04444831237196922, + 0.03246120363473892, + 0.046609193086624146, + 0.027329662814736366, + 0.047195032238960266, + 0.06111174449324608, + 0.07994882762432098, + 0.008786525577306747 + ], + "credit_map": [ + 0.007978547364473343, + 0.011382713913917542, + 0.012396360747516155, + 0.022077858448028564, + 0.009655220434069633, + 0.01178695634007454, + 0.019027916714549065, + 0.012023240327835083, + 0.04393760487437248, + 0.07459837198257446, + 0.04528585076332092, + 0.09534011781215668, + 0.05508119985461235, + 0.07802088558673859, + 0.09888271242380142, + 0.0546339750289917, + 0.0033341010566800833, + 0.005608036648482084, + 0.0031243732664734125, + 0.00704155070707202, + 0.007022805046290159, + 0.005725640803575516, + 0.009058077819645405, + 0.0035337256267666817, + 0.003561778925359249, + 0.0039063417352736, + 0.003051956184208393, + 0.00442825723439455, + 0.005982503294944763, + 0.004457603674381971, + 0.003770300652831793, + 0.00330246239900589, + 0.00495385704562068, + 0.00583330262452364, + 0.004267209209501743, + 0.005213337484747171, + 0.0062464443035423756, + 0.006263767834752798, + 0.0075566405430436134, + 0.006274635437875986, + 0.0021334970369935036, + 0.004480123985558748, + 0.0032078891526907682, + 0.0031474267598241568, + 0.004404608625918627, + 0.0034521911293268204, + 0.003270744578912854, + 0.00323318294249475, + 0.003794012824073434, + 0.010589735582470894, + 0.004485752899199724, + 0.004768057726323605, + 0.006315693259239197, + 0.007027694024145603, + 0.00451571773737669, + 0.005698366556316614, + 0.004513774998486042, + 0.00846859160810709, + 0.005207704845815897, + 0.009971702471375465, + 0.008214104920625687, + 0.007449000608175993, + 0.01101651694625616, + 0.006270349025726318, + 0.007188776973634958, + 0.016565173864364624, + 0.005408282391726971, + 0.012285259552299976, + 0.008246534503996372, + 0.010534357279539108, + 0.01077677309513092, + 0.008943670429289341, + 0.0009238438797183335, + 0.0015462527517229319, + 0.0007631709449924529, + 0.0010053986916318536, + 0.0010732655646279454, + 0.0011111925123259425, + 0.001059482921846211, + 0.0013039185432717204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.10032574087381363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903259038925171 + ], + "opposite_logprob_tokens": [ + -2.4640326499938965 + ], + "logprob_reward": -0.08903259038925171, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903259038925171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11811074614524841, + 0.5836352705955505, + 0.04678525775671005, + 0.05575435981154442, + 0.031173640862107277, + 0.018632428720593452, + 0.029131658375263214, + 0.05893407389521599, + 0.03447212651371956, + 0.02337050810456276 + ], + "credit_map": [ + 0.011132236570119858, + 0.01821528747677803, + 0.014526063576340675, + 0.0193965882062912, + 0.013259293511509895, + 0.012799309566617012, + 0.015023653395473957, + 0.013758302666246891, + 0.04629302769899368, + 0.07663273811340332, + 0.04747358709573746, + 0.1035076379776001, + 0.058872077614068985, + 0.08375043421983719, + 0.10764247179031372, + 0.0594632551074028, + 0.004338446538895369, + 0.01022371742874384, + 0.004900463856756687, + 0.006658151280134916, + 0.004684111103415489, + 0.005434023682028055, + 0.006115426309406757, + 0.004430911969393492, + 0.005076280795037746, + 0.014783952385187149, + 0.004452220164239407, + 0.005068881902843714, + 0.005726470611989498, + 0.0070609720423817635, + 0.0041376929730176926, + 0.009447879157960415, + 0.0026347727980464697, + 0.0039037850219756365, + 0.002229083562269807, + 0.002928910544142127, + 0.004476600326597691, + 0.003940922673791647, + 0.006314972881227732, + 0.0047445897944271564, + 0.001661424059420824, + 0.0034212933387607336, + 0.0017236331477761269, + 0.001781438128091395, + 0.0025651874020695686, + 0.002292103599756956, + 0.002711503766477108, + 0.0024758426006883383, + 0.002064469736069441, + 0.005517173558473587, + 0.0030785200651735067, + 0.003129278076812625, + 0.003735072212293744, + 0.0041673206724226475, + 0.0033970524091273546, + 0.004042769782245159, + 0.005524920299649239, + 0.01736370287835598, + 0.00599127309396863, + 0.00663755415007472, + 0.006142073776572943, + 0.007396767381578684, + 0.004578403662890196, + 0.005299372132867575, + 0.003367635188624263, + 0.007866006344556808, + 0.00219741091132164, + 0.0042976378463208675, + 0.003891552099958062, + 0.00457589328289032, + 0.0042024040594697, + 0.00407358817756176, + 0.0020147839095443487, + 0.008071019314229488, + 0.0024734344333410263, + 0.002165297046303749, + 0.0019345354521647096, + 0.003225410357117653, + 0.0015355857321992517, + 0.0019504391821101308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.08903259038925171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, horizontal medium shot, the camera held still, captures a woman in an exquisite lake-blue silk cheongsam. She walks gracefully and leisurely along a winding bluestone path in an ancient courtyard. Her long, lustrous black hair is pulled into a delicate low bun, secured with a delicate jade hairpin. Her features are classically graceful, a faint smile playing on their lips. The cheongsam is tailored to her figure, with a high, buttoned collar and elaborately embroidered with gold and silver thread, featuring blooming peonies and fluttering butterflies. Soft afternoon sunlight filters through the gaps in the courtyard's bamboo grove, casting dappled, swaying shadows on the ground. The light outlines the delicate contours of her figure and imbues the silk and embroidery of the cheongsam with a delicate sheen. In her right hand, she gently grasps a small, round fan, made of translucent white silk, painted with a few elegant orchids in light ink. As she steps and gently shakes her wrist, the fan sways gracefully across her chest. In the background, a corner of a wooden corridor with raised eaves and meticulously arranged Taihu stones can be seen. The entire scene, dominated by lake blue, emerald green, and pewter, exudes a tranquil, elegant, and poetic atmosphere of classical oriental art.", + "seed": 6289, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_4/index_5845_seed_6289_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13375057503581048, + "vlm_logprob": -0.13375057503581048, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.13375057503581048, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05961478874087334 + ], + "opposite_logprob_tokens": [ + -3.059614896774292 + ], + "logprob_reward": -0.05961478874087334, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05961478874087334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14087221026420593, + 0.04213538393378258, + 0.07563634216785431, + 0.03846495971083641, + 0.07222971320152283, + 0.026377098634839058, + 0.1231837198138237, + 0.12142859399318695, + 0.09050297737121582, + 0.26916900277137756 + ], + "credit_map": [ + 0.016483832150697708, + 0.015852654352784157, + 0.02294117771089077, + 0.019655879586935043, + 0.018342269584536552, + 0.016262346878647804, + 0.017903083935379982, + 0.01343096699565649, + 0.004584915470331907, + 0.004057061858475208, + 0.005795864388346672, + 0.005021106917411089, + 0.005937164183706045, + 0.004936777986586094, + 0.007975734770298004, + 0.003826757660135627, + 0.005999288987368345, + 0.008827537298202515, + 0.008579811081290245, + 0.017007775604724884, + 0.007981747388839722, + 0.007310331799089909, + 0.011966961435973644, + 0.007962889038026333, + 0.004322440829128027, + 0.004853395279496908, + 0.004259146749973297, + 0.005578565411269665, + 0.005378312431275845, + 0.003697747364640236, + 0.005664099007844925, + 0.0047112563624978065, + 0.007497148122638464, + 0.0063886744901537895, + 0.01881340704858303, + 0.007295476738363504, + 0.00637631444260478, + 0.011778486892580986, + 0.007090081460773945, + 0.006990116089582443, + 0.004471062682569027, + 0.003554002847522497, + 0.0027763210237026215, + 0.0033317189663648605, + 0.002938545309007168, + 0.0034772332292050123, + 0.0027538342401385307, + 0.0030743791721761227, + 0.011459160596132278, + 0.014592193998396397, + 0.019173024222254753, + 0.011831392534077168, + 0.013524376787245274, + 0.01465341541916132, + 0.014249519445002079, + 0.023700634017586708, + 0.012056824751198292, + 0.014048896729946136, + 0.013879852369427681, + 0.01980280503630638, + 0.01655421406030655, + 0.008688658475875854, + 0.018865421414375305, + 0.017531927675008774, + 0.011793154291808605, + 0.008354133926331997, + 0.010407163761556149, + 0.013184349052608013, + 0.00776250334456563, + 0.007025966886430979, + 0.012230410240590572, + 0.01974528841674328, + 0.015066673047840595, + 0.046211887151002884, + 0.021667901426553726, + 0.026339342817664146, + 0.053479958325624466, + 0.03263797238469124, + 0.039667900651693344, + 0.034097373485565186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.05961478874087334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054995931684970856 + ], + "opposite_logprob_tokens": [ + -2.9299960136413574 + ], + "logprob_reward": -0.054995931684970856, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054995931684970856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09348170459270477, + 0.038849856704473495, + 0.5794439911842346, + 0.0772295668721199, + 0.03907477483153343, + 0.03499487042427063, + 0.06342552602291107, + 0.020642593502998352, + 0.020057998597621918, + 0.032799046486616135 + ], + "credit_map": [ + 0.015154323540627956, + 0.011365062557160854, + 0.015582199208438396, + 0.009777804836630821, + 0.014576085843145847, + 0.011496329680085182, + 0.007255230098962784, + 0.008274665102362633, + 0.005013055168092251, + 0.004521956667304039, + 0.004065022803843021, + 0.0050794584676623344, + 0.004585093818604946, + 0.004366603679955006, + 0.004368357826024294, + 0.006850308272987604, + 0.05626584589481354, + 0.06355255097150803, + 0.055002663284540176, + 0.10452315956354141, + 0.060714852064847946, + 0.055816736072301865, + 0.13174743950366974, + 0.05182080715894699, + 0.009418739937245846, + 0.0071472059935331345, + 0.014605306088924408, + 0.00987125001847744, + 0.0068674092181026936, + 0.011268792673945427, + 0.01086284127086401, + 0.007188018411397934, + 0.004349860362708569, + 0.0037035574205219746, + 0.0059659890830516815, + 0.0057664294727146626, + 0.003953009378165007, + 0.005284301470965147, + 0.006649434100836515, + 0.00340219191275537, + 0.002761123701930046, + 0.0037798157427459955, + 0.004292570985853672, + 0.005891794338822365, + 0.004543359391391277, + 0.003821480320766568, + 0.006087501998990774, + 0.0038172246422618628, + 0.006564727518707514, + 0.007021384779363871, + 0.010781039483845234, + 0.007293036673218012, + 0.008501502685248852, + 0.006136606447398663, + 0.007688273210078478, + 0.009438950568437576, + 0.0019789221696555614, + 0.0023933465126901865, + 0.0030495901592075825, + 0.003368778619915247, + 0.0022338973358273506, + 0.0024399503599852324, + 0.0029634390957653522, + 0.002214668085798621, + 0.0015991791151463985, + 0.002152792876586318, + 0.0031897500157356262, + 0.0023626317270100117, + 0.001918741618283093, + 0.0013624858111143112, + 0.0026629697531461716, + 0.004809447564184666, + 0.003127660835161805, + 0.0054323249496519566, + 0.003109517740085721, + 0.0033578253351151943, + 0.00501388730481267, + 0.003546597436070442, + 0.005537443794310093, + 0.0036737890914082527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.054995931684970856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1004244014620781 + ], + "opposite_logprob_tokens": [ + -2.350424289703369 + ], + "logprob_reward": -0.1004244014620781, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1004244014620781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05200904980301857, + 0.037603482604026794, + 0.7043722867965698, + 0.051596738398075104, + 0.03565898910164833, + 0.024321740493178368, + 0.03427451476454735, + 0.018927162513136864, + 0.015200533904135227, + 0.026035508140921593 + ], + "credit_map": [ + 0.008185554295778275, + 0.007220048923045397, + 0.008803557604551315, + 0.005542532075196505, + 0.006894104648381472, + 0.005079902242869139, + 0.004707162734121084, + 0.005576192867010832, + 0.005623668432235718, + 0.0038286091294139624, + 0.008283847942948341, + 0.0034782893490046263, + 0.0028726081363856792, + 0.004564289469271898, + 0.005465845111757517, + 0.003486327128484845, + 0.06746279448270798, + 0.07777981460094452, + 0.06689313799142838, + 0.12631843984127045, + 0.07473047077655792, + 0.06787312030792236, + 0.16025090217590332, + 0.0630636066198349, + 0.005419728346168995, + 0.004659717436879873, + 0.00847750436514616, + 0.00730228703469038, + 0.004497729707509279, + 0.0069151464849710464, + 0.00965806096792221, + 0.00466656731441617, + 0.005048037972301245, + 0.0045011937618255615, + 0.004432570189237595, + 0.004899122286587954, + 0.003924414981156588, + 0.00427993992343545, + 0.004995136521756649, + 0.0035785739310085773, + 0.002557384083047509, + 0.002273476216942072, + 0.003220475045964122, + 0.0036641608458012342, + 0.0028007016517221928, + 0.0026482283137738705, + 0.004847316071391106, + 0.00230999942868948, + 0.003938025329262018, + 0.004197843838483095, + 0.005534361582249403, + 0.004833798855543137, + 0.004625710193067789, + 0.0038157457020133734, + 0.0036881715059280396, + 0.0036408568266779184, + 0.002201422583311796, + 0.0026212474331259727, + 0.0021755427587777376, + 0.0026461973320692778, + 0.002614379860460758, + 0.0023522612173110247, + 0.0024376302026212215, + 0.0018784835701808333, + 0.001614117412827909, + 0.0013880584156140685, + 0.0020369598641991615, + 0.0017659771256148815, + 0.0012616435997188091, + 0.0012779702665284276, + 0.0022175905760377645, + 0.003638217691332102, + 0.0023487925063818693, + 0.004762263968586922, + 0.0017840127693489194, + 0.0021764005068689585, + 0.005664852913469076, + 0.002950986847281456, + 0.0036305240355432034, + 0.0027176751755177975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.1004244014620781 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521360218524933 + ], + "opposite_logprob_tokens": [ + -1.502135992050171 + ], + "logprob_reward": -0.2521360218524933, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521360218524933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.127986341714859, + 0.10573919117450714, + 0.1800813227891922, + 0.07019022107124329, + 0.04051220044493675, + 0.03406532481312752, + 0.15061624348163605, + 0.028741104528307915, + 0.09106022119522095, + 0.17100787162780762 + ], + "credit_map": [ + 0.019122086465358734, + 0.016956966370344162, + 0.019067902117967606, + 0.01393855269998312, + 0.01715252362191677, + 0.013373450376093388, + 0.010336501523852348, + 0.01803835667669773, + 0.010093017481267452, + 0.010290938429534435, + 0.021535729989409447, + 0.014091853983700275, + 0.00787576474249363, + 0.008730428293347359, + 0.01714733988046646, + 0.01597410999238491, + 0.017263514921069145, + 0.01785896345973015, + 0.01794823259115219, + 0.032624438405036926, + 0.017603779211640358, + 0.01765803061425686, + 0.04220680147409439, + 0.016917549073696136, + 0.013777450658380985, + 0.007661647628992796, + 0.007702426519244909, + 0.008879227563738823, + 0.007652570027858019, + 0.00991035345941782, + 0.008622311986982822, + 0.005984229501336813, + 0.005447015166282654, + 0.0038926454726606607, + 0.008433361537754536, + 0.005408003460615873, + 0.0038862822111696005, + 0.005081706680357456, + 0.004622613545507193, + 0.0037405702751129866, + 0.00398288294672966, + 0.003647204488515854, + 0.0035100062377750874, + 0.005060794297605753, + 0.003990231081843376, + 0.004590170923620462, + 0.00485540647059679, + 0.004428626038134098, + 0.016677014529705048, + 0.018677137792110443, + 0.028033696115016937, + 0.017390871420502663, + 0.018799448385834694, + 0.015550661832094193, + 0.014804773032665253, + 0.02068263478577137, + 0.002426612889394164, + 0.0037359637208282948, + 0.0037399353459477425, + 0.0033123379107564688, + 0.0036727760452777147, + 0.003179574152454734, + 0.005936522968113422, + 0.0027373814955353737, + 0.012198506854474545, + 0.00970611721277237, + 0.011526507325470448, + 0.010852784849703312, + 0.009683608077466488, + 0.009290425106883049, + 0.012742888182401657, + 0.015059386380016804, + 0.013007016852498055, + 0.034947995096445084, + 0.013133592903614044, + 0.01734275557100773, + 0.04876190051436424, + 0.019225159659981728, + 0.013197766616940498, + 0.011391674168407917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.2521360218524933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158173143863678 + ], + "opposite_logprob_tokens": [ + -1.7015817165374756 + ], + "logprob_reward": -0.20158173143863678, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158173143863678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1419908106327057, + 0.06579053401947021, + 0.3598802387714386, + 0.06717562675476074, + 0.026421966031193733, + 0.030393509194254875, + 0.10411673784255981, + 0.01571943238377571, + 0.026239609345793724, + 0.16227146983146667 + ], + "credit_map": [ + 0.023522863164544106, + 0.01789279468357563, + 0.02048630453646183, + 0.011273657903075218, + 0.019142748787999153, + 0.018835002556443214, + 0.009533978067338467, + 0.021303467452526093, + 0.00859922170639038, + 0.006341453175991774, + 0.011404952965676785, + 0.00617958465591073, + 0.005090177059173584, + 0.008322376757860184, + 0.007599092088639736, + 0.012253680266439915, + 0.034705083817243576, + 0.0376833938062191, + 0.03431103751063347, + 0.061130911111831665, + 0.036783039569854736, + 0.03663340210914612, + 0.08660256862640381, + 0.03203083574771881, + 0.008638016879558563, + 0.006396505516022444, + 0.009699194692075253, + 0.009340761229395866, + 0.0065400260500609875, + 0.009522506035864353, + 0.00930051226168871, + 0.0077381138689816, + 0.0031450956594198942, + 0.0023792439606040716, + 0.004037288948893547, + 0.0021866674069315195, + 0.0021635808516293764, + 0.004635707940906286, + 0.0022212485782802105, + 0.005653133615851402, + 0.0018466890323907137, + 0.0032584872096776962, + 0.0036848618183285, + 0.003328255144879222, + 0.003034815192222595, + 0.0057985526509583, + 0.003293053014203906, + 0.006148799788206816, + 0.01194771844893694, + 0.012127348221838474, + 0.019523177295923233, + 0.01335710659623146, + 0.012489762157201767, + 0.010653720237314701, + 0.00981875043362379, + 0.014199161902070045, + 0.0015001562424004078, + 0.0028313621878623962, + 0.0019379816949367523, + 0.002210841979831457, + 0.002002852503210306, + 0.0017573696095496416, + 0.001763968146406114, + 0.0017149008344858885, + 0.004284746944904327, + 0.002882034284994006, + 0.0036735341418534517, + 0.003088638884946704, + 0.0032374004367738962, + 0.002610453637316823, + 0.003152645193040371, + 0.0033101574517786503, + 0.013170385733246803, + 0.03382469341158867, + 0.011199493892490864, + 0.013013938441872597, + 0.04866933450102806, + 0.016507338732481003, + 0.01424308679997921, + 0.011643221601843834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.20158173143863678 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, horizontal medium shot, the camera held still, captures a woman in an exquisite lake-blue silk cheongsam. She walks gracefully and leisurely along a winding bluestone path in an ancient courtyard. Her long, lustrous black hair is pulled into a delicate low bun, secured with a delicate jade hairpin. Her features are classically graceful, a faint smile playing on their lips. The cheongsam is tailored to her figure, with a high, buttoned collar and elaborately embroidered with gold and silver thread, featuring blooming peonies and fluttering butterflies. Soft afternoon sunlight filters through the gaps in the courtyard's bamboo grove, casting dappled, swaying shadows on the ground. The light outlines the delicate contours of her figure and imbues the silk and embroidery of the cheongsam with a delicate sheen. In her right hand, she gently grasps a small, round fan, made of translucent white silk, painted with a few elegant orchids in light ink. As she steps and gently shakes her wrist, the fan sways gracefully across her chest. In the background, a corner of a wooden corridor with raised eaves and meticulously arranged Taihu stones can be seen. The entire scene, dominated by lake blue, emerald green, and pewter, exudes a tranquil, elegant, and poetic atmosphere of classical oriental art.", + "seed": 6290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_4/index_5845_seed_6290_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07108397483825683, + "vlm_logprob": -0.07108397483825683, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.07108397483825683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04914144054055214 + ], + "opposite_logprob_tokens": [ + -3.2991414070129395 + ], + "logprob_reward": -0.04914144054055214, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04914144054055214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1470687836408615, + 0.09291369467973709, + 0.06398726254701614, + 0.16255049407482147, + 0.07633841037750244, + 0.09096379578113556, + 0.058596983551979065, + 0.08597470819950104, + 0.11837346106767654, + 0.10323239862918854 + ], + "credit_map": [ + 0.014136713929474354, + 0.01751292683184147, + 0.01419051457196474, + 0.01475686114281416, + 0.015508079901337624, + 0.016994474455714226, + 0.024349462240934372, + 0.029619740322232246, + 0.01259083952754736, + 0.014141364023089409, + 0.008059408515691757, + 0.012980899773538113, + 0.011844756081700325, + 0.010488137602806091, + 0.00945678073912859, + 0.013351505622267723, + 0.00956053752452135, + 0.00918586179614067, + 0.007468874566257, + 0.006358305923640728, + 0.008852574974298477, + 0.007417189888656139, + 0.006342016160488129, + 0.008801903575658798, + 0.013726056553423405, + 0.01748400181531906, + 0.013887113891541958, + 0.032425135374069214, + 0.021854529157280922, + 0.020931312814354897, + 0.026969846338033676, + 0.01527250837534666, + 0.00956752896308899, + 0.008735778741538525, + 0.006358510814607143, + 0.007166311610490084, + 0.018459415063261986, + 0.010841423645615578, + 0.0072064087726175785, + 0.00800302904099226, + 0.015670398250222206, + 0.01009141281247139, + 0.009850206784904003, + 0.010642344132065773, + 0.010845194570720196, + 0.01112157478928566, + 0.012585609219968319, + 0.010157061740756035, + 0.007870038039982319, + 0.006742572411894798, + 0.0068653724156320095, + 0.008388497866690159, + 0.007001929916441441, + 0.00552435452118516, + 0.010850127786397934, + 0.005354090128093958, + 0.016560368239879608, + 0.01019018143415451, + 0.007664970587939024, + 0.010974768549203873, + 0.0145360566675663, + 0.00810379721224308, + 0.01057426817715168, + 0.00737029779702425, + 0.022244269028306007, + 0.017539381980895996, + 0.011880752630531788, + 0.012334597297012806, + 0.014797690324485302, + 0.01744716614484787, + 0.011118941009044647, + 0.011010662652552128, + 0.009373804554343224, + 0.004438122268766165, + 0.03418378531932831, + 0.007768967654556036, + 0.007771345321089029, + 0.028053561225533485, + 0.006176549941301346, + 0.0054662590846419334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.04914144054055214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207248568534851 + ], + "opposite_logprob_tokens": [ + -2.812072515487671 + ], + "logprob_reward": -0.06207248568534851, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207248568534851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1641387939453125, + 0.058182112872600555, + 0.04324224591255188, + 0.10397039353847504, + 0.0456579215824604, + 0.09919589757919312, + 0.045311763882637024, + 0.058161452412605286, + 0.13476093113422394, + 0.24737846851348877 + ], + "credit_map": [ + 0.03474055975675583, + 0.024011272937059402, + 0.01627279818058014, + 0.013180706650018692, + 0.023142391815781593, + 0.020500456914305687, + 0.01216822024434805, + 0.020122390240430832, + 0.008022341877222061, + 0.00714375264942646, + 0.005512258969247341, + 0.009077678434550762, + 0.0070539056323468685, + 0.006447445601224899, + 0.008625134825706482, + 0.006299596279859543, + 0.006185128819197416, + 0.0077211977913975716, + 0.0035126141738146544, + 0.0050132921896874905, + 0.005550006870180368, + 0.004658911842852831, + 0.005200350657105446, + 0.005400740075856447, + 0.009994534775614738, + 0.010477039031684399, + 0.009324893355369568, + 0.015062353573739529, + 0.017928682267665863, + 0.014215079136192799, + 0.017971331253647804, + 0.008996480144560337, + 0.006787335500121117, + 0.005568576045334339, + 0.004829478915780783, + 0.004876607563346624, + 0.006374369841068983, + 0.0043966746889054775, + 0.006421699654310942, + 0.006403177510946989, + 0.016665704548358917, + 0.013325056992471218, + 0.009892985224723816, + 0.00819916557520628, + 0.015393566340208054, + 0.0097091905772686, + 0.01518961787223816, + 0.010820611380040646, + 0.003878458170220256, + 0.004516708664596081, + 0.003964108414947987, + 0.007126845885068178, + 0.007533231284469366, + 0.004626674111932516, + 0.009536213241517544, + 0.004129523877054453, + 0.0067435926757752895, + 0.00531840231269598, + 0.004374281968921423, + 0.010610651224851608, + 0.009889406152069569, + 0.0069907791912555695, + 0.008751078508794308, + 0.005483262240886688, + 0.016790922731161118, + 0.018547620624303818, + 0.013460401445627213, + 0.017484404146671295, + 0.02145836316049099, + 0.01419580727815628, + 0.020729659125208855, + 0.012093756347894669, + 0.02822047844529152, + 0.024086542427539825, + 0.042647819966077805, + 0.02322726882994175, + 0.041596777737140656, + 0.03682872653007507, + 0.03543858602643013, + 0.015332271344959736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.06207248568534851 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12712301313877106 + ], + "opposite_logprob_tokens": [ + -2.1271231174468994 + ], + "logprob_reward": -0.12712301313877106, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12712301313877106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20544745028018951, + 0.05570182576775551, + 0.02093798853456974, + 0.11834544688463211, + 0.05567195639014244, + 0.09401019662618637, + 0.11328328400850296, + 0.10830731689929962, + 0.113743357360363, + 0.11455116420984268 + ], + "credit_map": [ + 0.04343600198626518, + 0.02810012549161911, + 0.024098630994558334, + 0.015146718360483646, + 0.028653105720877647, + 0.028259456157684326, + 0.015271320007741451, + 0.022482093423604965, + 0.00813411083072424, + 0.006383812054991722, + 0.004869138356298208, + 0.008568023331463337, + 0.006577649153769016, + 0.006729321088641882, + 0.008083966560661793, + 0.006355802528560162, + 0.0023782593198120594, + 0.004014964215457439, + 0.0019344741012901068, + 0.0026761204935610294, + 0.0020567539613693953, + 0.002692105248570442, + 0.0027478390838950872, + 0.0024374728091061115, + 0.011541694402694702, + 0.011179630644619465, + 0.008480899967253208, + 0.01590057648718357, + 0.02390000782907009, + 0.014663370326161385, + 0.021585557609796524, + 0.01109371054917574, + 0.006100860424339771, + 0.00603228434920311, + 0.004937431775033474, + 0.005053476896136999, + 0.013169961981475353, + 0.007336578331887722, + 0.006011457182466984, + 0.007029903586953878, + 0.013866226188838482, + 0.01490884181112051, + 0.010873659513890743, + 0.009704514406621456, + 0.012795720249414444, + 0.009276295080780983, + 0.012427248992025852, + 0.010157685726881027, + 0.024822920560836792, + 0.018790051341056824, + 0.015057132579386234, + 0.01035986840724945, + 0.008082280866801739, + 0.012709314003586769, + 0.010393364354968071, + 0.013068352825939655, + 0.008429056033492088, + 0.010779435746371746, + 0.00714966282248497, + 0.014378445222973824, + 0.031896937638521194, + 0.012777053751051426, + 0.01453203521668911, + 0.008364691399037838, + 0.011211445555090904, + 0.009103745222091675, + 0.011691474355757236, + 0.01698993146419525, + 0.015309120528399944, + 0.01674160175025463, + 0.02208457514643669, + 0.010611473582684994, + 0.01074758730828762, + 0.007225557696074247, + 0.024820171296596527, + 0.012169275432825089, + 0.012916217558085918, + 0.021216383203864098, + 0.018678363412618637, + 0.006777605973184109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.12712301313877106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810395672917366 + ], + "opposite_logprob_tokens": [ + -3.2881040573120117 + ], + "logprob_reward": -0.03810395672917366, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810395672917366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18315932154655457, + 0.11449456214904785, + 0.08440423756837845, + 0.11934827268123627, + 0.08524558693170547, + 0.10831429809331894, + 0.08723360300064087, + 0.07844020426273346, + 0.06263184547424316, + 0.07672807574272156 + ], + "credit_map": [ + 0.035366032272577286, + 0.028396034613251686, + 0.017543617635965347, + 0.015139589086174965, + 0.026647470891475677, + 0.022943750023841858, + 0.014092680998146534, + 0.023030148819088936, + 0.012641336768865585, + 0.015702582895755768, + 0.009643282741308212, + 0.01981808990240097, + 0.01166553795337677, + 0.012207799591124058, + 0.020338112488389015, + 0.01247782539576292, + 0.010422888211905956, + 0.01485824678093195, + 0.007483828812837601, + 0.010526727885007858, + 0.00982026755809784, + 0.009502296335995197, + 0.01110902614891529, + 0.01068095676600933, + 0.016604775562882423, + 0.014133439399302006, + 0.015038864687085152, + 0.02009081095457077, + 0.01078164391219616, + 0.013294151052832603, + 0.015978462994098663, + 0.013426119461655617, + 0.011233502998948097, + 0.010585784912109375, + 0.012299160473048687, + 0.01120492722839117, + 0.009628535248339176, + 0.009030992165207863, + 0.012163558974862099, + 0.009099130518734455, + 0.017313916236162186, + 0.016444852575659752, + 0.014301721937954426, + 0.010904444381594658, + 0.012912847101688385, + 0.011074546724557877, + 0.013108829036355019, + 0.012253140099346638, + 0.010544983670115471, + 0.009679319337010384, + 0.008511975407600403, + 0.010007592849433422, + 0.013859701342880726, + 0.010634694248437881, + 0.016154201701283455, + 0.007841133512556553, + 0.007168825250118971, + 0.009588152170181274, + 0.005889852065593004, + 0.013665879145264626, + 0.01200895942747593, + 0.007911192253232002, + 0.014953366480767727, + 0.007253982592374086, + 0.011899740435183048, + 0.010917666368186474, + 0.00636180629953742, + 0.0045633334666490555, + 0.00857075396925211, + 0.006409990135580301, + 0.006934954784810543, + 0.0069735972210764885, + 0.007042364217340946, + 0.0042587448842823505, + 0.023186488077044487, + 0.0061171529814600945, + 0.007771338801831007, + 0.019151294603943825, + 0.005596974398940802, + 0.003603717079386115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.03810395672917366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897897809743881 + ], + "opposite_logprob_tokens": [ + -2.578979015350342 + ], + "logprob_reward": -0.07897897809743881, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897897809743881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1504930853843689, + 0.0783109962940216, + 0.052423302084207535, + 0.10128969699144363, + 0.0738515630364418, + 0.1001487746834755, + 0.2154616266489029, + 0.06533664464950562, + 0.08471769839525223, + 0.0779666006565094 + ], + "credit_map": [ + 0.018019989132881165, + 0.023391535505652428, + 0.01522408239543438, + 0.015890996903181076, + 0.01746601238846779, + 0.01760181598365307, + 0.01786324940621853, + 0.02503540739417076, + 0.009690527804195881, + 0.009587281383574009, + 0.006116911303251982, + 0.011502147652208805, + 0.011653699912130833, + 0.011600427329540253, + 0.008541357703506947, + 0.009618643671274185, + 0.008047079667448997, + 0.005549882538616657, + 0.006381125655025244, + 0.005026461556553841, + 0.007840734906494617, + 0.006934915203601122, + 0.004509221762418747, + 0.008133881725370884, + 0.014067361131310463, + 0.01327401865273714, + 0.015620016492903233, + 0.01176761370152235, + 0.012137068435549736, + 0.012446385808289051, + 0.011732026934623718, + 0.010245206765830517, + 0.014455536380410194, + 0.009592140093445778, + 0.010760026052594185, + 0.007800529710948467, + 0.0069791944697499275, + 0.009258987382054329, + 0.006531187798827887, + 0.008473964408040047, + 0.019121451303362846, + 0.016797451302409172, + 0.01514480635523796, + 0.00902704056352377, + 0.008997575379908085, + 0.011056358925998211, + 0.008739812299609184, + 0.011264275759458542, + 0.05551746115088463, + 0.03360747918486595, + 0.034433089196681976, + 0.012724997475743294, + 0.015227129682898521, + 0.023018386214971542, + 0.011560243554413319, + 0.029372842982411385, + 0.012703755870461464, + 0.007395751774311066, + 0.008577276021242142, + 0.006252649240195751, + 0.008256187662482262, + 0.006691464222967625, + 0.007000486366450787, + 0.008459076285362244, + 0.00985666736960411, + 0.011585145257413387, + 0.0078390222042799, + 0.011612912639975548, + 0.017121100798249245, + 0.01019419264048338, + 0.009404883719980717, + 0.00710377236828208, + 0.0069407192058861256, + 0.003458375809714198, + 0.02645147778093815, + 0.005744345020502806, + 0.005688378121703863, + 0.021697944030165672, + 0.004399564117193222, + 0.003585801227018237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.07897897809743881 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, horizontal medium shot, the camera held still, captures a woman in an exquisite lake-blue silk cheongsam. She walks gracefully and leisurely along a winding bluestone path in an ancient courtyard. Her long, lustrous black hair is pulled into a delicate low bun, secured with a delicate jade hairpin. Her features are classically graceful, a faint smile playing on their lips. The cheongsam is tailored to her figure, with a high, buttoned collar and elaborately embroidered with gold and silver thread, featuring blooming peonies and fluttering butterflies. Soft afternoon sunlight filters through the gaps in the courtyard's bamboo grove, casting dappled, swaying shadows on the ground. The light outlines the delicate contours of her figure and imbues the silk and embroidery of the cheongsam with a delicate sheen. In her right hand, she gently grasps a small, round fan, made of translucent white silk, painted with a few elegant orchids in light ink. As she steps and gently shakes her wrist, the fan sways gracefully across her chest. In the background, a corner of a wooden corridor with raised eaves and meticulously arranged Taihu stones can be seen. The entire scene, dominated by lake blue, emerald green, and pewter, exudes a tranquil, elegant, and poetic atmosphere of classical oriental art.", + "seed": 6291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_4/index_5845_seed_6291_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07958537489175796, + "vlm_logprob": -0.07958537489175796, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.07958537489175796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07083175331354141 + ], + "opposite_logprob_tokens": [ + -2.945831775665283 + ], + "logprob_reward": -0.07083175331354141, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07083175331354141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32599928975105286, + 0.08298010379076004, + 0.06425586342811584, + 0.0748240277171135, + 0.07771693915128708, + 0.04492044821381569, + 0.11647381633520126, + 0.04750234633684158, + 0.07421647757291794, + 0.0911107137799263 + ], + "credit_map": [ + 0.05568854510784149, + 0.05044376105070114, + 0.03493625670671463, + 0.03396274894475937, + 0.04103322699666023, + 0.04820139333605766, + 0.030078522861003876, + 0.03165484964847565, + 0.011259181424975395, + 0.01636131852865219, + 0.008426419459283352, + 0.011599679477512836, + 0.008952150121331215, + 0.01174147892743349, + 0.007339031435549259, + 0.00730083929374814, + 0.0087317805737257, + 0.008982648141682148, + 0.008882264606654644, + 0.009170430712401867, + 0.006410462781786919, + 0.006312042474746704, + 0.008096679113805294, + 0.007669550832360983, + 0.010348757728934288, + 0.012129519134759903, + 0.009701029397547245, + 0.008147018030285835, + 0.011495311744511127, + 0.011363277211785316, + 0.007777562364935875, + 0.003861552570015192, + 0.014654278755187988, + 0.009884693659842014, + 0.012530360370874405, + 0.00931924395263195, + 0.00801896769553423, + 0.010264210402965546, + 0.006642422638833523, + 0.006402759347110987, + 0.00874381884932518, + 0.006568319629877806, + 0.005832272116094828, + 0.005023166537284851, + 0.004583744332194328, + 0.004747570957988501, + 0.004242240451276302, + 0.0051793185994029045, + 0.01535934116691351, + 0.01037987507879734, + 0.02312352880835533, + 0.011529266834259033, + 0.01438156422227621, + 0.014973952434957027, + 0.011770356446504593, + 0.014955934137105942, + 0.006404364015907049, + 0.0058415732346475124, + 0.003975173458456993, + 0.009045458398759365, + 0.006759251933544874, + 0.003705972107127309, + 0.008438446559011936, + 0.003332109423354268, + 0.017297113314270973, + 0.007170597556978464, + 0.011210085824131966, + 0.009651057422161102, + 0.010728755034506321, + 0.005003596190363169, + 0.0074475668370723724, + 0.005707703530788422, + 0.022592609748244286, + 0.008589575067162514, + 0.013855522498488426, + 0.008400985971093178, + 0.012874347157776356, + 0.0077647799625992775, + 0.007280825171619654, + 0.00975207332521677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.07083175331354141 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054998867213726044 + ], + "opposite_logprob_tokens": [ + -2.9299988746643066 + ], + "logprob_reward": -0.054998867213726044, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054998867213726044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30263423919677734, + 0.09483843296766281, + 0.05875508114695549, + 0.12572918832302094, + 0.08249787241220474, + 0.03472500666975975, + 0.10810884833335876, + 0.054161228239536285, + 0.06060291826725006, + 0.07794717699289322 + ], + "credit_map": [ + 0.03705073520541191, + 0.05682000890374184, + 0.029974769800901413, + 0.042074304074048996, + 0.04104265570640564, + 0.04270806536078453, + 0.024602601304650307, + 0.028361087664961815, + 0.0205093827098608, + 0.01403719186782837, + 0.011726212687790394, + 0.015420433133840561, + 0.008126339875161648, + 0.008381680585443974, + 0.010876824147999287, + 0.005760373082011938, + 0.007333435583859682, + 0.008198120631277561, + 0.007449916563928127, + 0.006430565845221281, + 0.008800488896667957, + 0.008299342356622219, + 0.007068905979394913, + 0.005174304358661175, + 0.017624566331505775, + 0.017547817900776863, + 0.01357151847332716, + 0.016160815954208374, + 0.02027156576514244, + 0.02040368691086769, + 0.01351996324956417, + 0.006629258394241333, + 0.020497208461165428, + 0.011493665166199207, + 0.011638777330517769, + 0.008366474881768227, + 0.00835665687918663, + 0.009152596816420555, + 0.005594700574874878, + 0.007397796493023634, + 0.005670167040079832, + 0.004063331056386232, + 0.004189685452729464, + 0.005323816556483507, + 0.0033427414018660784, + 0.003671040525659919, + 0.004377488512545824, + 0.004086733795702457, + 0.014579500071704388, + 0.009774385020136833, + 0.02037074789404869, + 0.010323680937290192, + 0.01530391164124012, + 0.01363225094974041, + 0.010333182290196419, + 0.013791190460324287, + 0.006290932651609182, + 0.006454129237681627, + 0.005512287840247154, + 0.008289435878396034, + 0.008455317467451096, + 0.003912941552698612, + 0.0095664719119668, + 0.005679714493453503, + 0.007821410894393921, + 0.005959001369774342, + 0.006155041046440601, + 0.011953725479543209, + 0.008942858316004276, + 0.005209046881645918, + 0.00858165044337511, + 0.005980184767395258, + 0.010458476841449738, + 0.009724589996039867, + 0.009327409788966179, + 0.010296422056853771, + 0.0135643957182765, + 0.0077329762279987335, + 0.00977791752666235, + 0.007064986042678356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.054998867213726044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08912026882171631 + ], + "opposite_logprob_tokens": [ + -2.464120388031006 + ], + "logprob_reward": -0.08912026882171631, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08912026882171631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29801511764526367, + 0.11703374236822128, + 0.10333223640918732, + 0.11350850760936737, + 0.08943590521812439, + 0.037742309272289276, + 0.06053836643695831, + 0.05022779479622841, + 0.058182742446660995, + 0.07198330760002136 + ], + "credit_map": [ + 0.034598708152770996, + 0.044181566685438156, + 0.031738560646772385, + 0.040379106998443604, + 0.03678825497627258, + 0.040423646569252014, + 0.030111605301499367, + 0.03979366645216942, + 0.022761890664696693, + 0.01599632203578949, + 0.012882702052593231, + 0.024179914966225624, + 0.01054622046649456, + 0.009178563952445984, + 0.012992440722882748, + 0.008495686575770378, + 0.012607293203473091, + 0.017069118097424507, + 0.009883507154881954, + 0.009267652407288551, + 0.01938384398818016, + 0.01589062064886093, + 0.012232745066285133, + 0.00699745723977685, + 0.016229981556534767, + 0.016345733776688576, + 0.01219128631055355, + 0.014310110360383987, + 0.018362967297434807, + 0.018196670338511467, + 0.012370887212455273, + 0.005500873550772667, + 0.01961282640695572, + 0.013115731999278069, + 0.012769673950970173, + 0.01081063225865364, + 0.008750512264668941, + 0.010325506329536438, + 0.007681646384298801, + 0.006369378883391619, + 0.005300867836922407, + 0.006446346640586853, + 0.0033190646208822727, + 0.005782573949545622, + 0.004770024213939905, + 0.004189580213278532, + 0.004946675151586533, + 0.0029871773440390825, + 0.0084653589874506, + 0.006059844046831131, + 0.010103446431457996, + 0.006040873471647501, + 0.009008835069835186, + 0.0072786868549883366, + 0.005911733955144882, + 0.007669585756957531, + 0.006272535305470228, + 0.006373531650751829, + 0.005124385934323072, + 0.006391249131411314, + 0.008173755370080471, + 0.0035617391113191843, + 0.007995471358299255, + 0.00633512856438756, + 0.007336771115660667, + 0.00603141263127327, + 0.006711015943437815, + 0.011619283817708492, + 0.008310549892485142, + 0.005275215022265911, + 0.008139067329466343, + 0.004759430885314941, + 0.009730536490678787, + 0.008402397856116295, + 0.009520580992102623, + 0.008021927438676357, + 0.011381951160728931, + 0.007861221209168434, + 0.009540639817714691, + 0.007524047512561083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.08912026882171631 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003682106733322 + ], + "opposite_logprob_tokens": [ + -2.6950368881225586 + ], + "logprob_reward": -0.07003682106733322, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003682106733322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1956968754529953, + 0.06383796036243439, + 0.061034414917230606, + 0.08679961413145065, + 0.10273666679859161, + 0.051734816282987595, + 0.1734871119260788, + 0.06289602071046829, + 0.08634024113416672, + 0.11543632298707962 + ], + "credit_map": [ + 0.026772424578666687, + 0.02634255401790142, + 0.02944670245051384, + 0.02333790622651577, + 0.021458512172102928, + 0.027210919186472893, + 0.025600939989089966, + 0.015526905655860901, + 0.013818635605275631, + 0.0109791848808527, + 0.008068372495472431, + 0.00926972646266222, + 0.006237601395696402, + 0.006330312695354223, + 0.005631098058074713, + 0.0035030250437557697, + 0.007875664159655571, + 0.008471189998090267, + 0.00907919555902481, + 0.006236533634364605, + 0.006149770226329565, + 0.005547259468585253, + 0.007221036124974489, + 0.010453764349222183, + 0.013664226047694683, + 0.011796822771430016, + 0.011286969296634197, + 0.01121819019317627, + 0.009793159551918507, + 0.010864676907658577, + 0.009193168953061104, + 0.008982396684587002, + 0.02453823946416378, + 0.013558571226894855, + 0.014294194988906384, + 0.011230649426579475, + 0.010883801616728306, + 0.010344798676669598, + 0.007098384667187929, + 0.010788020677864552, + 0.010486365295946598, + 0.005457915365695953, + 0.006713551003485918, + 0.006378537975251675, + 0.005056076217442751, + 0.005058342590928078, + 0.005383465904742479, + 0.007200564257800579, + 0.024517711251974106, + 0.01666570082306862, + 0.03279919549822807, + 0.014056258834898472, + 0.02604863792657852, + 0.02256139926612377, + 0.015322491526603699, + 0.021515728905797005, + 0.007094613276422024, + 0.008370982483029366, + 0.006935252342373133, + 0.009529917500913143, + 0.00907206255942583, + 0.004267355892807245, + 0.011090910993516445, + 0.006534927524626255, + 0.01307111606001854, + 0.008378923870623112, + 0.007546592503786087, + 0.0173936877399683, + 0.013232983648777008, + 0.006353006698191166, + 0.012673192657530308, + 0.0076907360926270485, + 0.01847120188176632, + 0.012768360786139965, + 0.013364529237151146, + 0.017570113763213158, + 0.019421793520450592, + 0.009877451695501804, + 0.014224881306290627, + 0.009737995453178883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.07003682106733322 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady, horizontal medium shot with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a lake-blue silk cheongsam with a high, buttoned collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cheongsam elaborately embroidered with gold and silver thread featuring blooming peonies and fluttering butterflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman walking gracefully along a winding bluestone path in an ancient courtyard while holding a translucent white silk fan with orchid designs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft afternoon sunlight filtering through a bamboo grove, casting dappled swaying shadows on the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293916404247284 + ], + "opposite_logprob_tokens": [ + -2.2379391193389893 + ], + "logprob_reward": -0.11293916404247284, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293916404247284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40304580330848694, + 0.10797920823097229, + 0.0367065966129303, + 0.032455120235681534, + 0.07447358220815659, + 0.03871855139732361, + 0.13197103142738342, + 0.05209839716553688, + 0.04661850258708, + 0.07593322545289993 + ], + "credit_map": [ + 0.048015810549259186, + 0.05655773729085922, + 0.039748936891555786, + 0.029809309169650078, + 0.059079576283693314, + 0.07591340690851212, + 0.04660481587052345, + 0.04731622338294983, + 0.015201087109744549, + 0.01577996276319027, + 0.013205118477344513, + 0.006881954614073038, + 0.013534975238144398, + 0.021516423672437668, + 0.00905031617730856, + 0.012809373438358307, + 0.00495579419657588, + 0.004010526463389397, + 0.0070458450354635715, + 0.0031080911867320538, + 0.0031768223270773888, + 0.0052489303052425385, + 0.0037272709887474775, + 0.005433319136500359, + 0.005732035730034113, + 0.0038169603794813156, + 0.005105217453092337, + 0.004488182254135609, + 0.002876752056181431, + 0.003508619964122772, + 0.004338017199188471, + 0.002589336596429348, + 0.013009951449930668, + 0.008832503110170364, + 0.011409414000809193, + 0.006475692614912987, + 0.008378186263144016, + 0.009618177078664303, + 0.004984142258763313, + 0.011765520088374615, + 0.00611292477697134, + 0.004236944019794464, + 0.004799457732588053, + 0.00402232026681304, + 0.003148743649944663, + 0.003794365096837282, + 0.0034477561712265015, + 0.009156039915978909, + 0.017977649345993996, + 0.011946790851652622, + 0.02635880745947361, + 0.01157198566943407, + 0.016620319336652756, + 0.01764972321689129, + 0.012761306948959827, + 0.017084449529647827, + 0.007410279009491205, + 0.005992709659039974, + 0.007842360995709896, + 0.005506732501089573, + 0.008207877166569233, + 0.004901278298348188, + 0.006023218855261803, + 0.00621394207701087, + 0.006458239164203405, + 0.00498453713953495, + 0.006247308105230331, + 0.007387637160718441, + 0.00559376971796155, + 0.003753958735615015, + 0.006020442117005587, + 0.006172609981149435, + 0.009890750050544739, + 0.00735442852601409, + 0.009420122019946575, + 0.0071280826814472675, + 0.009750764816999435, + 0.007826614193618298, + 0.00863461010158062, + 0.015927860513329506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.11293916404247284 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_5/index_tensor([6042, 6042])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_5/index_tensor([6042, 6042])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9a88f71640887faaf6d9dc710b8db13d62b4bfef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_5/index_tensor([6042, 6042])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "magnificent 360-degree slow-motion shot, shot from a slightly upward perspective, captures the epic battle between two mythical phoenixes amidst a turbulent sky swirling with dark storm clouds. One phoenix is composed entirely of flowing magma and incandescent fire, each feather a pure, pulsating crimson and gold flame. Its eyes are like blazing embers, and its flight leaves a dazzling trail of sparks and embers in its wake. The other phoenix, in stark contrast, is carved from translucent ice crystals, its feathers razor-sharp icicles gleaming with sapphire and cyan. Its wings, with their beats, scatter countless tiny, glittering frost particles, and its icy eyes are like distant stars. The camera smoothly rotates around them, capturing every detail as they swirl, dive, and claw at each other in a deadly aerial ballet: the fire phoenix folds its wings sharply, swooping down like a flaming meteor, while the ice phoenix gracefully sideways, blocking with the edge of its crystal wings. When the camera rotates to the optimal angle, the two divine birds have a fierce head-on collision. The burning feathers collide with the icy feathers, and a dazzling energy shock wave instantly erupts - the hot sparks hiss and turn into steam when they come into contact with the cold air, while the hard ice crystals explode into countless glittering fragments under the high temperature, intertwined with the falling and extinguished flame feathers, forming a destructive picture full of primitive power and magnificent beauty. The whole scene, with the extreme contrast of warm and cold colors, creates a tense, intense and sacred fantasy atmosphere.", + "seed": 32799, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_5/index_6042_seed_32799_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01182404845021665, + "vlm_logprob": -0.01182404845021665, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.01182404845021665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006600365042686462 + ], + "opposite_logprob_tokens": [ + -5.256600379943848 + ], + "logprob_reward": -0.006600365042686462, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006600365042686462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15502393245697021, + 0.05112180486321449, + 0.060035280883312225, + 0.08289790898561478, + 0.025475207716226578, + 0.44978395104408264, + 0.039890389889478683, + 0.04076211899518967, + 0.03731020167469978, + 0.057699207216501236 + ], + "credit_map": [ + 0.030609821900725365, + 0.022278349846601486, + 0.021915767341852188, + 0.0172478836029768, + 0.01608983986079693, + 0.015398206189274788, + 0.016796426847577095, + 0.014687634073197842, + 0.005712481215596199, + 0.006282282993197441, + 0.005365047603845596, + 0.006017313804477453, + 0.006308532319962978, + 0.006265741307288408, + 0.005743817891925573, + 0.00942658819258213, + 0.005069907754659653, + 0.006987821310758591, + 0.005771420896053314, + 0.007241468410938978, + 0.005032934248447418, + 0.0065001375041902065, + 0.012598305940628052, + 0.010833286680281162, + 0.011366995051503181, + 0.013976366259157658, + 0.010065787471830845, + 0.008161053992807865, + 0.007873764261603355, + 0.01171994861215353, + 0.008464987389743328, + 0.011269006878137589, + 0.0037963432259857655, + 0.001431053620763123, + 0.0030955267138779163, + 0.0021633575670421124, + 0.0037454457487910986, + 0.00478101521730423, + 0.002805622760206461, + 0.003656845074146986, + 0.0078887352719903, + 0.005256649572402239, + 0.0050254096277058125, + 0.006510668899863958, + 0.010068249888718128, + 0.18934734165668488, + 0.03219824284315109, + 0.19348862767219543, + 0.005289791151881218, + 0.004025180358439684, + 0.007328078616410494, + 0.004205248784273863, + 0.004460851661860943, + 0.0041690729558467865, + 0.00487512955442071, + 0.005537037272006273, + 0.0038478595670312643, + 0.004516614601016045, + 0.004450816195458174, + 0.005852132104337215, + 0.004310347139835358, + 0.00587909622117877, + 0.006425446830689907, + 0.005479807499796152, + 0.004452962893992662, + 0.007639159448444843, + 0.004532897844910622, + 0.0034030855167657137, + 0.005605876911431551, + 0.004282286390662193, + 0.003807616885751486, + 0.003586314618587494, + 0.006808083038777113, + 0.0062781027518212795, + 0.005739694926887751, + 0.010017379187047482, + 0.01114178728312254, + 0.005567474290728569, + 0.005675876047462225, + 0.006470805034041405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.006600365042686462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022054414730519056 + ], + "opposite_logprob_tokens": [ + -6.1272053718566895 + ], + "logprob_reward": -0.0022054414730519056, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022054414730519056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21907848119735718, + 0.07135839015245438, + 0.06898052990436554, + 0.05350155755877495, + 0.028580067679286003, + 0.2661280632019043, + 0.0330314040184021, + 0.14807729423046112, + 0.04661225527524948, + 0.0646519660949707 + ], + "credit_map": [ + 0.026339028030633926, + 0.02004411444067955, + 0.02847341075539589, + 0.02728579379618168, + 0.025268740952014923, + 0.029352379962801933, + 0.03866231068968773, + 0.02365269511938095, + 0.007111693732440472, + 0.006307872943580151, + 0.012432383373379707, + 0.012287287972867489, + 0.006218374241143465, + 0.008063153363764286, + 0.010027645155787468, + 0.008909976109862328, + 0.007455135229974985, + 0.008340646512806416, + 0.007967104203999043, + 0.008803858421742916, + 0.004952690098434687, + 0.009238096885383129, + 0.011177477426826954, + 0.011045520193874836, + 0.0066061848774552345, + 0.008134914562106133, + 0.006355020683258772, + 0.005971268285065889, + 0.004225056618452072, + 0.009722605347633362, + 0.004629105795174837, + 0.007857400923967361, + 0.0038785352371633053, + 0.002318101469427347, + 0.003913003019988537, + 0.0022861796896904707, + 0.00447915680706501, + 0.0050092460587620735, + 0.002510962774977088, + 0.004184883087873459, + 0.0060086939483881, + 0.0042525664903223515, + 0.004449318628758192, + 0.00576532119885087, + 0.005798667669296265, + 0.10908538848161697, + 0.019028961658477783, + 0.1117391362786293, + 0.0034566971007734537, + 0.002541020279750228, + 0.007191667333245277, + 0.004490380175411701, + 0.00328613817691803, + 0.00329533196054399, + 0.0042807720601558685, + 0.004489399027079344, + 0.010226507671177387, + 0.014864219352602959, + 0.01631884090602398, + 0.02053149603307247, + 0.01651613973081112, + 0.018709275871515274, + 0.029649771749973297, + 0.021261034533381462, + 0.007455515209585428, + 0.009816767647862434, + 0.004405525512993336, + 0.00381294684484601, + 0.007644715253263712, + 0.004911123309284449, + 0.004992933478206396, + 0.0035727277863770723, + 0.007378576789051294, + 0.004824340343475342, + 0.005835479125380516, + 0.011203490197658539, + 0.007820667698979378, + 0.007516501471400261, + 0.00763977225869894, + 0.012433137744665146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0022054414730519056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180398270487785 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180398270487785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180398270487785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09548791497945786, + 0.04700027406215668, + 0.20389877259731293, + 0.22469590604305267, + 0.01460949145257473, + 0.3141510784626007, + 0.021510664373636246, + 0.021586736664175987, + 0.02595711313188076, + 0.031102053821086884 + ], + "credit_map": [ + 0.011356402188539505, + 0.010539884679019451, + 0.01011265441775322, + 0.010519005358219147, + 0.012263968586921692, + 0.015946535393595695, + 0.01385102141648531, + 0.01089843176305294, + 0.0042104520834982395, + 0.005286269821226597, + 0.006264034658670425, + 0.00543449679389596, + 0.005523213651031256, + 0.00566642452031374, + 0.006861917208880186, + 0.007753460668027401, + 0.019865861162543297, + 0.022672774270176888, + 0.031055746600031853, + 0.034310244023799896, + 0.014086534269154072, + 0.022264692932367325, + 0.026483967900276184, + 0.033158935606479645, + 0.03633144125342369, + 0.022504715248942375, + 0.029796533286571503, + 0.049340054392814636, + 0.012477058917284012, + 0.018817313015460968, + 0.017311347648501396, + 0.0381174236536026, + 0.0013774486724287271, + 0.0010376943973824382, + 0.0019044902874156833, + 0.0012332516489550471, + 0.0021030567586421967, + 0.002446639584377408, + 0.0022386140190064907, + 0.0022682943381369114, + 0.005684196949005127, + 0.004143756348639727, + 0.004149516578763723, + 0.006653988268226385, + 0.012567867524921894, + 0.12198401987552643, + 0.028287028893828392, + 0.13068069517612457, + 0.0021706894040107727, + 0.0017161027062684298, + 0.004358956590294838, + 0.0024885646998882294, + 0.0023960373364388943, + 0.002346480032429099, + 0.0034335488453507423, + 0.002600284293293953, + 0.00202455697581172, + 0.0028480691835284233, + 0.002007570583373308, + 0.003994167782366276, + 0.0028237137012183666, + 0.0038252309896051884, + 0.0021327114664018154, + 0.0019307138863950968, + 0.0035500265657901764, + 0.0031378869898617268, + 0.002041048603132367, + 0.0022615408524870872, + 0.004666232969611883, + 0.0038955090567469597, + 0.0027868940960615873, + 0.0036179719027131796, + 0.003947192337363958, + 0.00283265789039433, + 0.003486646106466651, + 0.0053269583731889725, + 0.005979111418128014, + 0.003147994400933385, + 0.0029500797390937805, + 0.003431411460042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.018180398270487785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068367287516594 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068367287516594, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068367287516594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1035861074924469, + 0.05460808798670769, + 0.06388042122125626, + 0.09489542990922928, + 0.06530681252479553, + 0.34325119853019714, + 0.09206783771514893, + 0.07786065340042114, + 0.050619471818208694, + 0.05392393097281456 + ], + "credit_map": [ + 0.015315476804971695, + 0.013123832643032074, + 0.014719897881150246, + 0.011830303817987442, + 0.013072040863335133, + 0.010508760809898376, + 0.013672290369868279, + 0.011343501508235931, + 0.005552788730710745, + 0.005360329523682594, + 0.007255810312926769, + 0.007831140421330929, + 0.0053505259566009045, + 0.006669316440820694, + 0.008206499740481377, + 0.008381674066185951, + 0.005457194522023201, + 0.010169192217290401, + 0.005556619726121426, + 0.007108275778591633, + 0.007151288911700249, + 0.006510313134640455, + 0.012829286977648735, + 0.009098244830965996, + 0.01081868913024664, + 0.01600867509841919, + 0.012152443639934063, + 0.011381734162569046, + 0.010162353515625, + 0.011866508983075619, + 0.010084759443998337, + 0.012420265935361385, + 0.006240050308406353, + 0.006337516009807587, + 0.0081143444404006, + 0.0060368431732058525, + 0.012056732550263405, + 0.010557764209806919, + 0.008496744558215141, + 0.007466824259608984, + 0.007131726946681738, + 0.004217346198856831, + 0.004866433795541525, + 0.005989374127238989, + 0.007296213414520025, + 0.14262062311172485, + 0.024782449007034302, + 0.1463470458984375, + 0.010628351010382175, + 0.00756501592695713, + 0.012112598866224289, + 0.011815082281827927, + 0.01043833326548338, + 0.012353451922535896, + 0.012152248993515968, + 0.015002751722931862, + 0.00789436511695385, + 0.01040744874626398, + 0.0064934613183140755, + 0.014086761511862278, + 0.01387442834675312, + 0.009237876161932945, + 0.00921982154250145, + 0.0066465032286942005, + 0.007129326928406954, + 0.0076826950535178185, + 0.005983265582472086, + 0.006442250683903694, + 0.007983045652508736, + 0.004888381343334913, + 0.00523150572553277, + 0.0052789971232414246, + 0.006530087441205978, + 0.005154403392225504, + 0.005149223376065493, + 0.00912980455905199, + 0.008691910654306412, + 0.005751516669988632, + 0.007848293520510197, + 0.005668689962476492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.016068367287516594 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065670177340508 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065670177340508, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065670177340508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17313426733016968, + 0.05084359645843506, + 0.10909361392259598, + 0.18858212232589722, + 0.17438596487045288, + 0.08134596794843674, + 0.0356043204665184, + 0.03490009531378746, + 0.043461188673973083, + 0.10864883661270142 + ], + "credit_map": [ + 0.024692660197615623, + 0.01616237871348858, + 0.027098488062620163, + 0.024340016767382622, + 0.020753853023052216, + 0.020005282014608383, + 0.022702129557728767, + 0.017379462718963623, + 0.005611071828752756, + 0.004478067625313997, + 0.007521866820752621, + 0.007836142554879189, + 0.003350664395838976, + 0.006710936781018972, + 0.007878998294472694, + 0.007455843035131693, + 0.007187788840383291, + 0.013321266509592533, + 0.012501302175223827, + 0.01565755158662796, + 0.013158148154616356, + 0.013554469682276249, + 0.01683732494711876, + 0.016875755041837692, + 0.021136028692126274, + 0.030015988275408745, + 0.021166477352380753, + 0.0406283438205719, + 0.013244876638054848, + 0.01594253070652485, + 0.01451140083372593, + 0.03193647414445877, + 0.01119257416576147, + 0.017731785774230957, + 0.02408287301659584, + 0.018017634749412537, + 0.02069336175918579, + 0.03612050414085388, + 0.023985914885997772, + 0.02256130985915661, + 0.005367958918213844, + 0.0034529066178947687, + 0.004942455794662237, + 0.006570319179445505, + 0.012087044306099415, + 0.018384546041488647, + 0.01406478974968195, + 0.016475947573781013, + 0.004017437342554331, + 0.002485924633219838, + 0.00782324280589819, + 0.004019169602543116, + 0.0037889881059527397, + 0.0030228793621063232, + 0.005260339472442865, + 0.005186339840292931, + 0.003721281886100769, + 0.005134793929755688, + 0.004025978036224842, + 0.005042037460952997, + 0.005642126314342022, + 0.004002034198492765, + 0.0030440662521868944, + 0.004287777468562126, + 0.008135159499943256, + 0.009509171359241009, + 0.0041381376795470715, + 0.004665134474635124, + 0.005267301108688116, + 0.004387946799397469, + 0.004191284533590078, + 0.003167053684592247, + 0.011014284566044807, + 0.010038888081908226, + 0.012108506634831429, + 0.01794716715812683, + 0.016400832682847977, + 0.013549582101404667, + 0.01415946800261736, + 0.01343010924756527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.016065670177340508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "magnificent 360-degree slow-motion shot, shot from a slightly upward perspective, captures the epic battle between two mythical phoenixes amidst a turbulent sky swirling with dark storm clouds. One phoenix is composed entirely of flowing magma and incandescent fire, each feather a pure, pulsating crimson and gold flame. Its eyes are like blazing embers, and its flight leaves a dazzling trail of sparks and embers in its wake. The other phoenix, in stark contrast, is carved from translucent ice crystals, its feathers razor-sharp icicles gleaming with sapphire and cyan. Its wings, with their beats, scatter countless tiny, glittering frost particles, and its icy eyes are like distant stars. The camera smoothly rotates around them, capturing every detail as they swirl, dive, and claw at each other in a deadly aerial ballet: the fire phoenix folds its wings sharply, swooping down like a flaming meteor, while the ice phoenix gracefully sideways, blocking with the edge of its crystal wings. When the camera rotates to the optimal angle, the two divine birds have a fierce head-on collision. The burning feathers collide with the icy feathers, and a dazzling energy shock wave instantly erupts - the hot sparks hiss and turn into steam when they come into contact with the cold air, while the hard ice crystals explode into countless glittering fragments under the high temperature, intertwined with the falling and extinguished flame feathers, forming a destructive picture full of primitive power and magnificent beauty. The whole scene, with the extreme contrast of warm and cold colors, creates a tense, intense and sacred fantasy atmosphere.", + "seed": 32800, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_5/index_6042_seed_32800_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03972751856781542, + "vlm_logprob": -0.03972751856781542, + "vlm_yes_no_margin": 4.224999976158142, + "vlm_token_credit": -0.03972751856781542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00808644201606512 + ], + "opposite_logprob_tokens": [ + -5.008086204528809 + ], + "logprob_reward": -0.00808644201606512, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00808644201606512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19321946799755096, + 0.05838900804519653, + 0.04982610046863556, + 0.09300782531499863, + 0.19262686371803284, + 0.0628768801689148, + 0.0735272690653801, + 0.0701025128364563, + 0.1570548415184021, + 0.049369219690561295 + ], + "credit_map": [ + 0.01624860242009163, + 0.017525702714920044, + 0.037037525326013565, + 0.02845718152821064, + 0.015537378378212452, + 0.03344068303704262, + 0.020392587408423424, + 0.02457982674241066, + 0.008170753717422485, + 0.005414229817688465, + 0.006239026784896851, + 0.0059656985104084015, + 0.005776346195489168, + 0.008384057320654392, + 0.006689334288239479, + 0.011749565601348877, + 0.005984602961689234, + 0.004643741995096207, + 0.006821232847869396, + 0.004864863120019436, + 0.004466506652534008, + 0.00889322068542242, + 0.006688705179840326, + 0.007463226560503244, + 0.00927125196903944, + 0.016657060012221336, + 0.013246900402009487, + 0.0055323573760688305, + 0.015895668417215347, + 0.01007701177150011, + 0.011452657170593739, + 0.010874918662011623, + 0.02387397550046444, + 0.034313734620809555, + 0.02816724218428135, + 0.022677106782794, + 0.034592390060424805, + 0.02618725597858429, + 0.012529384344816208, + 0.010285773314535618, + 0.006479222327470779, + 0.008689617738127708, + 0.008432405069470406, + 0.010250627994537354, + 0.008879940025508404, + 0.006390455178916454, + 0.007599600125104189, + 0.006155009847134352, + 0.006657440215349197, + 0.013963703997433186, + 0.008918629959225655, + 0.00965110119432211, + 0.007329205051064491, + 0.006655330304056406, + 0.010001881048083305, + 0.010349986143410206, + 0.009175647981464863, + 0.011259508319199085, + 0.009881827048957348, + 0.012399120256304741, + 0.0069583444856107235, + 0.0061211371794342995, + 0.004386206157505512, + 0.009920720010995865, + 0.02393389493227005, + 0.014756259508430958, + 0.03485528752207756, + 0.024595381692051888, + 0.016830623149871826, + 0.010594542138278484, + 0.014838953502476215, + 0.016649892553687096, + 0.006788055412471294, + 0.00615348806604743, + 0.00851393211632967, + 0.006866346579045057, + 0.005340876989066601, + 0.006903471425175667, + 0.004740589298307896, + 0.004062460735440254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.00808644201606512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011893587652593851 + ], + "opposite_logprob_tokens": [ + -6.751189231872559 + ], + "logprob_reward": -0.0011893587652593851, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011893587652593851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17428207397460938, + 0.03709982708096504, + 0.041686397045850754, + 0.056371986865997314, + 0.06634580343961716, + 0.05497065186500549, + 0.06622155010700226, + 0.06182146444916725, + 0.34003621339797974, + 0.10116403549909592 + ], + "credit_map": [ + 0.017025383189320564, + 0.017494414001703262, + 0.029059123247861862, + 0.025924140587449074, + 0.014132644049823284, + 0.029112085700035095, + 0.01653478853404522, + 0.02499948814511299, + 0.00643605412915349, + 0.003488238900899887, + 0.004718288313597441, + 0.003199764294549823, + 0.002790526719763875, + 0.005429220851510763, + 0.004060028120875359, + 0.006977709010243416, + 0.0052719684317708015, + 0.0035169683396816254, + 0.006219947244971991, + 0.004097979050129652, + 0.004741381388157606, + 0.0062055992893874645, + 0.005322021432220936, + 0.0063105314038693905, + 0.006052081007510424, + 0.010181647725403309, + 0.0075105465948581696, + 0.004057401325553656, + 0.009856004267930984, + 0.005103115923702717, + 0.005121451336890459, + 0.00848973635584116, + 0.009137731976807117, + 0.010784105397760868, + 0.009220241568982601, + 0.010960184037685394, + 0.007713438477367163, + 0.00881741102784872, + 0.005753060337156057, + 0.003959629684686661, + 0.005434337072074413, + 0.006898725870996714, + 0.0065652597695589066, + 0.01053521316498518, + 0.0050603775307536125, + 0.0050065238028764725, + 0.008082150481641293, + 0.0073880646377801895, + 0.005934803280979395, + 0.011714909225702286, + 0.008401704020798206, + 0.009088374674320221, + 0.00659293495118618, + 0.00646234443411231, + 0.008830519393086433, + 0.009195961058139801, + 0.007358473259955645, + 0.010219467803835869, + 0.008658574894070625, + 0.012280869297683239, + 0.006866531912237406, + 0.004166504368185997, + 0.0044365376234054565, + 0.00783450435847044, + 0.060266561806201935, + 0.03849145770072937, + 0.027201494202017784, + 0.03044063039124012, + 0.0356299988925457, + 0.025273576378822327, + 0.04219067841768265, + 0.08054180443286896, + 0.011652807705104351, + 0.011809569783508778, + 0.016300296410918236, + 0.01913234405219555, + 0.009818846359848976, + 0.013375687412917614, + 0.00865461491048336, + 0.010419866070151329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0011893587652593851 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278587698936462 + ], + "opposite_logprob_tokens": [ + -2.0177857875823975 + ], + "logprob_reward": -0.14278587698936462, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14278587698936462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41178080439567566, + 0.033052653074264526, + 0.0290286336094141, + 0.1570894867181778, + 0.05783234164118767, + 0.07619427889585495, + 0.029027702286839485, + 0.04822070524096489, + 0.12175591289997101, + 0.036017466336488724 + ], + "credit_map": [ + 0.04525924474000931, + 0.03455163538455963, + 0.02706584520637989, + 0.029587415978312492, + 0.03366868942975998, + 0.09302999824285507, + 0.039258770644664764, + 0.10935920476913452, + 0.004731917288154364, + 0.004751074593514204, + 0.003247613552957773, + 0.0029959697276353836, + 0.004346476402133703, + 0.003508544759824872, + 0.004666516091674566, + 0.004804540425539017, + 0.004305172711610794, + 0.0030998357106000185, + 0.0031579190399497747, + 0.0024120397865772247, + 0.0035657642874866724, + 0.005222453270107508, + 0.0032089657615870237, + 0.004056483041495085, + 0.008894690312445164, + 0.013468906283378601, + 0.050188105553388596, + 0.0074975136667490005, + 0.013128984719514847, + 0.010071600787341595, + 0.013444476760923862, + 0.040395211428403854, + 0.0047348784282803535, + 0.01012150477617979, + 0.006837891880422831, + 0.010362470522522926, + 0.004895853344351053, + 0.009263011626899242, + 0.007550152949988842, + 0.0040665725246071815, + 0.008273603394627571, + 0.01142627838999033, + 0.009288232773542404, + 0.010199128650128841, + 0.010790287517011166, + 0.006537486799061298, + 0.009092890657484531, + 0.010586364194750786, + 0.0037020226009190083, + 0.003945774864405394, + 0.00383029249496758, + 0.0038623616565018892, + 0.002970272209495306, + 0.002235454274341464, + 0.003449206007644534, + 0.005032319109886885, + 0.005739268381148577, + 0.007035986986011267, + 0.004612292628735304, + 0.007277947850525379, + 0.0051138815470039845, + 0.006358540616929531, + 0.003395909210667014, + 0.00868687592446804, + 0.019360091537237167, + 0.010692580603063107, + 0.02715381607413292, + 0.019250286743044853, + 0.010390087030827999, + 0.008807065896689892, + 0.010892338119447231, + 0.015209649689495564, + 0.005104237701743841, + 0.005159600637853146, + 0.006577171850949526, + 0.00417060311883688, + 0.0024966418277472258, + 0.00415757903829217, + 0.003071904880926013, + 0.005279728211462498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.14278587698936462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288540542125702 + ], + "opposite_logprob_tokens": [ + -3.7732884883880615 + ], + "logprob_reward": -0.023288540542125702, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288540542125702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11657325178384781, + 0.08973339945077896, + 0.06387228518724442, + 0.051059916615486145, + 0.0874011293053627, + 0.10871004313230515, + 0.0704081580042839, + 0.06435318291187286, + 0.2723000645637512, + 0.07558845728635788 + ], + "credit_map": [ + 0.01726556569337845, + 0.013137913309037685, + 0.013335046358406544, + 0.013954096473753452, + 0.011510109528899193, + 0.018200693652033806, + 0.012898256070911884, + 0.016271591186523438, + 0.016550960019230843, + 0.009622528217732906, + 0.012584029696881771, + 0.01045859232544899, + 0.007466011215001345, + 0.010280435904860497, + 0.010107404552400112, + 0.012663448229432106, + 0.006804953794926405, + 0.005715453531593084, + 0.01030889805406332, + 0.007201073691248894, + 0.00685125170275569, + 0.011454150080680847, + 0.006418140605092049, + 0.009118371643126011, + 0.005310602951794863, + 0.006217741407454014, + 0.007221761625260115, + 0.004428774584084749, + 0.004833712242543697, + 0.007568950764834881, + 0.008608588948845863, + 0.006869788747280836, + 0.009712082333862782, + 0.013991866260766983, + 0.011139627546072006, + 0.014384075067937374, + 0.008723210543394089, + 0.015554743818938732, + 0.007513134274631739, + 0.006382404360920191, + 0.010801956988871098, + 0.01400159765034914, + 0.019618283957242966, + 0.011628656648099422, + 0.013894110918045044, + 0.00765865808352828, + 0.016505271196365356, + 0.01460152305662632, + 0.007835538126528263, + 0.011883067898452282, + 0.009067053906619549, + 0.009237999096512794, + 0.007786625996232033, + 0.00572745269164443, + 0.009452583268284798, + 0.009417844004929066, + 0.007272259332239628, + 0.010082203894853592, + 0.010102049447596073, + 0.01070600375533104, + 0.00849723257124424, + 0.005879694130271673, + 0.004789655562490225, + 0.00702408654615283, + 0.05216005817055702, + 0.02023906446993351, + 0.038507115095853806, + 0.02929084561765194, + 0.01742611452937126, + 0.02013474702835083, + 0.029255684465169907, + 0.06528648734092712, + 0.009768523275852203, + 0.008637391962110996, + 0.013686712831258774, + 0.00786014273762703, + 0.00785516481846571, + 0.007909582927823067, + 0.00784553587436676, + 0.012025408446788788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.023288540542125702 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287374526262283 + ], + "opposite_logprob_tokens": [ + -3.773287296295166 + ], + "logprob_reward": -0.023287374526262283, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287374526262283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09193599224090576, + 0.06749562919139862, + 0.07366100698709488, + 0.10844247043132782, + 0.10739430785179138, + 0.07541893422603607, + 0.05515804514288902, + 0.14143377542495728, + 0.16684658825397491, + 0.11221323162317276 + ], + "credit_map": [ + 0.013791979290544987, + 0.010686471126973629, + 0.014722569845616817, + 0.016493814066052437, + 0.0073843724094331264, + 0.008344481699168682, + 0.010160443373024464, + 0.01035186368972063, + 0.009714563377201557, + 0.010573671199381351, + 0.008459953591227531, + 0.006887116935104132, + 0.005613008048385382, + 0.008715912699699402, + 0.007486413232982159, + 0.010044997557997704, + 0.009170101955533028, + 0.005746860522776842, + 0.013717743568122387, + 0.007146203890442848, + 0.007228113245218992, + 0.011444834060966969, + 0.008095459081232548, + 0.011111699976027012, + 0.007477400824427605, + 0.012023315764963627, + 0.027905847877264023, + 0.00658007338643074, + 0.013501213863492012, + 0.008729114197194576, + 0.010237357579171658, + 0.021988149732351303, + 0.011462030932307243, + 0.01495717465877533, + 0.013878279365599155, + 0.015575852245092392, + 0.014509418047964573, + 0.021244237199425697, + 0.008488419465720654, + 0.007278904318809509, + 0.006945950444787741, + 0.01017734408378601, + 0.00967735331505537, + 0.010258124209940434, + 0.008412686176598072, + 0.006719349883496761, + 0.011512271128594875, + 0.011715862900018692, + 0.005989915691316128, + 0.011508640833199024, + 0.006563014350831509, + 0.007008485496044159, + 0.007225332781672478, + 0.00461291428655386, + 0.007349661085754633, + 0.004900089465081692, + 0.015813279896974564, + 0.02118421532213688, + 0.01573534868657589, + 0.018468691036105156, + 0.014047629199922085, + 0.019966308027505875, + 0.014968178234994411, + 0.02125013805925846, + 0.021796882152557373, + 0.018982306122779846, + 0.03120081126689911, + 0.023587839677929878, + 0.02953307516872883, + 0.014676788821816444, + 0.013147389516234398, + 0.013921498320996761, + 0.013550680130720139, + 0.013441672548651695, + 0.019559500738978386, + 0.017088524997234344, + 0.011299774050712585, + 0.013802848756313324, + 0.010990999639034271, + 0.01247923169285059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.023287374526262283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "magnificent 360-degree slow-motion shot, shot from a slightly upward perspective, captures the epic battle between two mythical phoenixes amidst a turbulent sky swirling with dark storm clouds. One phoenix is composed entirely of flowing magma and incandescent fire, each feather a pure, pulsating crimson and gold flame. Its eyes are like blazing embers, and its flight leaves a dazzling trail of sparks and embers in its wake. The other phoenix, in stark contrast, is carved from translucent ice crystals, its feathers razor-sharp icicles gleaming with sapphire and cyan. Its wings, with their beats, scatter countless tiny, glittering frost particles, and its icy eyes are like distant stars. The camera smoothly rotates around them, capturing every detail as they swirl, dive, and claw at each other in a deadly aerial ballet: the fire phoenix folds its wings sharply, swooping down like a flaming meteor, while the ice phoenix gracefully sideways, blocking with the edge of its crystal wings. When the camera rotates to the optimal angle, the two divine birds have a fierce head-on collision. The burning feathers collide with the icy feathers, and a dazzling energy shock wave instantly erupts - the hot sparks hiss and turn into steam when they come into contact with the cold air, while the hard ice crystals explode into countless glittering fragments under the high temperature, intertwined with the falling and extinguished flame feathers, forming a destructive picture full of primitive power and magnificent beauty. The whole scene, with the extreme contrast of warm and cold colors, creates a tense, intense and sacred fantasy atmosphere.", + "seed": 32801, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_5/index_6042_seed_32801_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08719009200576693, + "vlm_logprob": -0.08719009200576693, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.08719009200576693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008686376735568047 + ], + "opposite_logprob_tokens": [ + -4.883686542510986 + ], + "logprob_reward": -0.008686376735568047, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008686376735568047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12544703483581543, + 0.04746972769498825, + 0.08710117638111115, + 0.05906178057193756, + 0.04308793321251869, + 0.13766950368881226, + 0.03989870101213455, + 0.27747735381126404, + 0.06786687672138214, + 0.11491997539997101 + ], + "credit_map": [ + 0.011067486368119717, + 0.01414713729172945, + 0.02016826532781124, + 0.014633430168032646, + 0.013893961906433105, + 0.014202238991856575, + 0.0163411945104599, + 0.020993322134017944, + 0.005530188325792551, + 0.005153529345989227, + 0.005342432763427496, + 0.0060929241590201855, + 0.004820066969841719, + 0.006307124625891447, + 0.007752470672130585, + 0.006470990367233753, + 0.007244284264743328, + 0.008134325966238976, + 0.013361377641558647, + 0.01098456047475338, + 0.00681290403008461, + 0.008741996251046658, + 0.01759037934243679, + 0.014231354929506779, + 0.006524730008095503, + 0.005682556424289942, + 0.009711171500384808, + 0.00694917282089591, + 0.004463273100554943, + 0.004636761732399464, + 0.009587609209120274, + 0.011506509967148304, + 0.005606574006378651, + 0.004553297534584999, + 0.004614375066012144, + 0.007543232291936874, + 0.004793389234691858, + 0.003910134546458721, + 0.005893615540117025, + 0.006173311732709408, + 0.019073577597737312, + 0.01284823752939701, + 0.01296956930309534, + 0.02607754059135914, + 0.01784980297088623, + 0.013800861313939095, + 0.01732899621129036, + 0.0177209060639143, + 0.0067159682512283325, + 0.0037348587065935135, + 0.006374015938490629, + 0.004927522502839565, + 0.004624553490430117, + 0.0040881396271288395, + 0.005148810800164938, + 0.00428483122959733, + 0.024020111188292503, + 0.01844628155231476, + 0.07704783976078033, + 0.02855084463953972, + 0.026780618354678154, + 0.048837125301361084, + 0.03766675293445587, + 0.016127770766615868, + 0.00801515020430088, + 0.01608024723827839, + 0.009784187190234661, + 0.008281379006803036, + 0.005848532542586327, + 0.006772663444280624, + 0.009645291604101658, + 0.003439425490796566, + 0.013918586075305939, + 0.02014509029686451, + 0.021777253597974777, + 0.015146414749324322, + 0.01453320775181055, + 0.006466765888035297, + 0.01488524954766035, + 0.008047407492995262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.008686376735568047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001520549994893372 + ], + "opposite_logprob_tokens": [ + -6.50152063369751 + ], + "logprob_reward": -0.001520549994893372, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001520549994893372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1312883049249649, + 0.07168321311473846, + 0.0578714981675148, + 0.07292338460683823, + 0.06447509676218033, + 0.06305623799562454, + 0.07324974238872528, + 0.34606826305389404, + 0.059922512620687485, + 0.05946173146367073 + ], + "credit_map": [ + 0.010510648600757122, + 0.018411902710795403, + 0.01766936294734478, + 0.010763810947537422, + 0.021688979119062424, + 0.014442210085690022, + 0.01682944968342781, + 0.020971938967704773, + 0.007198932580649853, + 0.006754555739462376, + 0.007214956451207399, + 0.008317439816892147, + 0.010625677183270454, + 0.013277352787554264, + 0.008001046255230904, + 0.010293257422745228, + 0.004472322762012482, + 0.0053862398490309715, + 0.008236033841967583, + 0.008215808309614658, + 0.0051798392087221146, + 0.0058970064856112, + 0.010710706003010273, + 0.009773540310561657, + 0.006584065966308117, + 0.005165506154298782, + 0.013387046754360199, + 0.01093450840562582, + 0.0061912513338029385, + 0.007338516414165497, + 0.013786428608000278, + 0.009536059573292732, + 0.007236766163259745, + 0.007325927261263132, + 0.012812327593564987, + 0.009682433679699898, + 0.006005760747939348, + 0.004645733628422022, + 0.009159847162663937, + 0.007606303784996271, + 0.008655555546283722, + 0.00709711667150259, + 0.005127996206283569, + 0.010654961690306664, + 0.01033555343747139, + 0.00807999912649393, + 0.006760864518582821, + 0.006344187539070845, + 0.009497402235865593, + 0.004456114489585161, + 0.011535566300153732, + 0.011917629279196262, + 0.006299770902842283, + 0.008542128838598728, + 0.01109460461884737, + 0.009906519204378128, + 0.030918385833501816, + 0.019173163920640945, + 0.10512801259756088, + 0.03162180632352829, + 0.02299145981669426, + 0.08702565729618073, + 0.0323137491941452, + 0.01689605787396431, + 0.007270128466188908, + 0.014007829129695892, + 0.008744077757000923, + 0.00677605252712965, + 0.005028784275054932, + 0.00736719835549593, + 0.007487887982279062, + 0.003240553429350257, + 0.007424371317028999, + 0.008147727698087692, + 0.008974897675216198, + 0.006576779764145613, + 0.005953941959887743, + 0.006253066938370466, + 0.007613726891577244, + 0.00851721502840519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.001520549994893372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343165040016174 + ], + "opposite_logprob_tokens": [ + -1.3134316205978394 + ], + "logprob_reward": -0.31343165040016174, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343165040016174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0899265706539154, + 0.05237029120326042, + 0.13259470462799072, + 0.052085138857364655, + 0.0586465559899807, + 0.037814825773239136, + 0.02311149798333645, + 0.5059377551078796, + 0.02428039163351059, + 0.023232299834489822 + ], + "credit_map": [ + 0.008029327727854252, + 0.012452025897800922, + 0.00991495605558157, + 0.007407699711620808, + 0.011647351086139679, + 0.009021761827170849, + 0.010675733909010887, + 0.020777719095349312, + 0.006162890233099461, + 0.005855471361428499, + 0.006391329690814018, + 0.005460498388856649, + 0.006611963734030724, + 0.009060698561370373, + 0.006060871761292219, + 0.006766569800674915, + 0.01531465444713831, + 0.018700772896409035, + 0.018633168190717697, + 0.021675962954759598, + 0.012266621924936771, + 0.009401700459420681, + 0.020924165844917297, + 0.015677664428949356, + 0.006204437464475632, + 0.005423756316304207, + 0.01002789381891489, + 0.008071955293416977, + 0.003879352007061243, + 0.003504364751279354, + 0.007103224750608206, + 0.007870151661336422, + 0.007181559689342976, + 0.006516338791698217, + 0.009974981658160686, + 0.00943257100880146, + 0.005904250778257847, + 0.0037591566797345877, + 0.008413909934461117, + 0.007463783025741577, + 0.0046978299506008625, + 0.003149365773424506, + 0.0033915340900421143, + 0.009508507326245308, + 0.003917959053069353, + 0.0028139175847172737, + 0.0053779082372784615, + 0.0049578044563531876, + 0.0029647944029420614, + 0.002376949181780219, + 0.0037473184056580067, + 0.003733786754310131, + 0.0025062053464353085, + 0.00249123596586287, + 0.002581048058345914, + 0.002710159169510007, + 0.04482337087392807, + 0.022405317053198814, + 0.16767676174640656, + 0.027210397645831108, + 0.028618142008781433, + 0.09962687641382217, + 0.08357074111700058, + 0.032006099820137024, + 0.002627650508657098, + 0.005044423509389162, + 0.004360831342637539, + 0.0025993529707193375, + 0.0021376528311520815, + 0.003297277493402362, + 0.002422293880954385, + 0.0017909090965986252, + 0.0026850516442209482, + 0.0036395739298313856, + 0.0028439627494663, + 0.002601412357762456, + 0.0029781644698232412, + 0.002233468694612384, + 0.003208851907402277, + 0.0030418129172176123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.31343165040016174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902648836374283 + ], + "opposite_logprob_tokens": [ + -2.46402645111084 + ], + "logprob_reward": -0.08902648836374283, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902648836374283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14942337572574615, + 0.09415160119533539, + 0.07701817154884338, + 0.05031510815024376, + 0.10325643420219421, + 0.04783850908279419, + 0.08367261290550232, + 0.21573160588741302, + 0.12031668424606323, + 0.05827587470412254 + ], + "credit_map": [ + 0.01728065311908722, + 0.019317662343382835, + 0.020457496866583824, + 0.01391358207911253, + 0.01582239381968975, + 0.013320337049663067, + 0.013811926357448101, + 0.03549933806061745, + 0.00945349782705307, + 0.01039901189506054, + 0.013237779960036278, + 0.015496148727834225, + 0.009284866973757744, + 0.012011614628136158, + 0.012782363221049309, + 0.011486329138278961, + 0.011807975359261036, + 0.013713981956243515, + 0.013674699701368809, + 0.009827104397118092, + 0.00718877324834466, + 0.005749802105128765, + 0.008567458018660545, + 0.00648838235065341, + 0.004840982612222433, + 0.005310531239956617, + 0.008088118396699429, + 0.006725993473082781, + 0.004677062388509512, + 0.005706490948796272, + 0.007996134459972382, + 0.006969798821955919, + 0.01152002438902855, + 0.009803668595850468, + 0.015355809591710567, + 0.014878587797284126, + 0.010093128308653831, + 0.008991562761366367, + 0.012709836475551128, + 0.019903821870684624, + 0.006184694357216358, + 0.005292440764605999, + 0.005367710720747709, + 0.007278403267264366, + 0.005891548935323954, + 0.005842012818902731, + 0.004816533997654915, + 0.007165171205997467, + 0.009777961298823357, + 0.010077462531626225, + 0.014811770059168339, + 0.010961590334773064, + 0.009923765435814857, + 0.006535682827234268, + 0.012197165749967098, + 0.009387227706611156, + 0.014504154212772846, + 0.01653987541794777, + 0.045131370425224304, + 0.02426902763545513, + 0.025140542536973953, + 0.024374540895223618, + 0.051957447081804276, + 0.01381466444581747, + 0.013855444267392159, + 0.026794608682394028, + 0.017005661502480507, + 0.014104490168392658, + 0.012872463092207909, + 0.014241967350244522, + 0.013587859459221363, + 0.007854200899600983, + 0.007077977526932955, + 0.011974419467151165, + 0.01023411937057972, + 0.0059059299528598785, + 0.007905209437012672, + 0.003382194321602583, + 0.0065229800529778, + 0.005273048300296068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.08902648836374283 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328539453446865 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.02328539453446865, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328539453446865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1483398824930191, + 0.13336876034736633, + 0.09347540140151978, + 0.0789562538266182, + 0.04786211624741554, + 0.0333746112883091, + 0.10229883342981339, + 0.1247955784201622, + 0.10777487605810165, + 0.12975364923477173 + ], + "credit_map": [ + 0.013575525023043156, + 0.020817775279283524, + 0.01651310734450817, + 0.011924909427762032, + 0.02287396788597107, + 0.019118644297122955, + 0.01772201620042324, + 0.02579394169151783, + 0.010066603310406208, + 0.01882457546889782, + 0.016899267211556435, + 0.020365143194794655, + 0.016437377780675888, + 0.017777398228645325, + 0.016427500173449516, + 0.01657090149819851, + 0.008963818661868572, + 0.008870563469827175, + 0.010913338512182236, + 0.01197801809757948, + 0.00810819212347269, + 0.008823097683489323, + 0.01943320408463478, + 0.016385165974497795, + 0.009670769795775414, + 0.009576333686709404, + 0.015565531328320503, + 0.010292084887623787, + 0.005571537185460329, + 0.009411962702870369, + 0.011103206314146519, + 0.007764830254018307, + 0.006222998257726431, + 0.006596854422241449, + 0.008579985238611698, + 0.007389189675450325, + 0.004364829044789076, + 0.0034553182777017355, + 0.006007757969200611, + 0.005245186388492584, + 0.0038556167855858803, + 0.0034317283425480127, + 0.003309838240966201, + 0.004253562539815903, + 0.007815440185368061, + 0.002820600289851427, + 0.0034101910423487425, + 0.004477635491639376, + 0.011987690813839436, + 0.011716771870851517, + 0.012352848425507545, + 0.01541207917034626, + 0.011349870823323727, + 0.013110026717185974, + 0.012682527303695679, + 0.013687030412256718, + 0.016287991777062416, + 0.007031616289168596, + 0.020482463762164116, + 0.01182642113417387, + 0.013303877785801888, + 0.019908670336008072, + 0.02593609131872654, + 0.010018441826105118, + 0.014364921487867832, + 0.024152522906661034, + 0.014501167461276054, + 0.0162951797246933, + 0.009399259462952614, + 0.00791193637996912, + 0.015241391956806183, + 0.005908501334488392, + 0.01778278313577175, + 0.016109846532344818, + 0.0150138009339571, + 0.010940248146653175, + 0.012249636463820934, + 0.020001128315925598, + 0.01380705926567316, + 0.023849140852689743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02328539453446865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "magnificent 360-degree slow-motion shot, shot from a slightly upward perspective, captures the epic battle between two mythical phoenixes amidst a turbulent sky swirling with dark storm clouds. One phoenix is composed entirely of flowing magma and incandescent fire, each feather a pure, pulsating crimson and gold flame. Its eyes are like blazing embers, and its flight leaves a dazzling trail of sparks and embers in its wake. The other phoenix, in stark contrast, is carved from translucent ice crystals, its feathers razor-sharp icicles gleaming with sapphire and cyan. Its wings, with their beats, scatter countless tiny, glittering frost particles, and its icy eyes are like distant stars. The camera smoothly rotates around them, capturing every detail as they swirl, dive, and claw at each other in a deadly aerial ballet: the fire phoenix folds its wings sharply, swooping down like a flaming meteor, while the ice phoenix gracefully sideways, blocking with the edge of its crystal wings. When the camera rotates to the optimal angle, the two divine birds have a fierce head-on collision. The burning feathers collide with the icy feathers, and a dazzling energy shock wave instantly erupts - the hot sparks hiss and turn into steam when they come into contact with the cold air, while the hard ice crystals explode into countless glittering fragments under the high temperature, intertwined with the falling and extinguished flame feathers, forming a destructive picture full of primitive power and magnificent beauty. The whole scene, with the extreme contrast of warm and cold colors, creates a tense, intense and sacred fantasy atmosphere.", + "seed": 32802, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_5/index_6042_seed_32802_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04905865103937686, + "vlm_logprob": -0.04905865103937686, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04905865103937686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006983751431107521 + ], + "opposite_logprob_tokens": [ + -5.256983757019043 + ], + "logprob_reward": -0.006983751431107521, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006983751431107521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22293253242969513, + 0.10097941756248474, + 0.06287927180528641, + 0.14726825058460236, + 0.07970529049634933, + 0.06075718253850937, + 0.04639638215303421, + 0.04013695567846298, + 0.02949247881770134, + 0.20945225656032562 + ], + "credit_map": [ + 0.03297947347164154, + 0.01886064000427723, + 0.015321120619773865, + 0.014805867336690426, + 0.03909137845039368, + 0.05774448812007904, + 0.022369783371686935, + 0.021759793162345886, + 0.007837805896997452, + 0.0067609441466629505, + 0.0063131824135780334, + 0.007527455221861601, + 0.010297449305653572, + 0.02178788185119629, + 0.03172442689538002, + 0.008730268105864525, + 0.0061779748648405075, + 0.00689659034833312, + 0.007689340505748987, + 0.015288962982594967, + 0.004980999045073986, + 0.009665869176387787, + 0.006915593054145575, + 0.005263944622129202, + 0.011623612605035305, + 0.00951535627245903, + 0.0196934025734663, + 0.012993445619940758, + 0.019731586799025536, + 0.026570307090878487, + 0.019372297450900078, + 0.027768252417445183, + 0.010547989048063755, + 0.0066934856586158276, + 0.008148050867021084, + 0.010135098360478878, + 0.012277754954993725, + 0.007248683832585812, + 0.01322655938565731, + 0.01142766885459423, + 0.006544874515384436, + 0.003525595413520932, + 0.014626497402787209, + 0.008374263532459736, + 0.004704516846686602, + 0.011688455007970333, + 0.006004199851304293, + 0.0052887736819684505, + 0.0040890006348490715, + 0.0038646135944873095, + 0.007236090488731861, + 0.006758741568773985, + 0.00581359351053834, + 0.006346134934574366, + 0.007673748768866062, + 0.004614458419382572, + 0.0037147642578929663, + 0.003414965933188796, + 0.010536390356719494, + 0.005048003513365984, + 0.0030008580069988966, + 0.00782716739922762, + 0.0035800961777567863, + 0.0030147074721753597, + 0.004421534016728401, + 0.004249377176165581, + 0.004604072775691748, + 0.006022752728313208, + 0.0024133382830768824, + 0.0019340307917445898, + 0.0032864497043192387, + 0.0025609226431697607, + 0.05874387174844742, + 0.01397936325520277, + 0.07108239084482193, + 0.02186691015958786, + 0.006353038363158703, + 0.005630428437143564, + 0.024766957387328148, + 0.007029290311038494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.006983751431107521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002497175009921193 + ], + "opposite_logprob_tokens": [ + -6.00249719619751 + ], + "logprob_reward": -0.002497175009921193, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002497175009921193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27205032110214233, + 0.1403978019952774, + 0.05279533937573433, + 0.09894298017024994, + 0.029258551076054573, + 0.06045627221465111, + 0.06405161321163177, + 0.05634404718875885, + 0.01953011378645897, + 0.20617294311523438 + ], + "credit_map": [ + 0.05419924482703209, + 0.01599106378853321, + 0.014902926981449127, + 0.012800322845578194, + 0.053423184901475906, + 0.08813869953155518, + 0.016073593869805336, + 0.016521280631422997, + 0.010819658637046814, + 0.007806946989148855, + 0.005056183319538832, + 0.0058904471807181835, + 0.012025602161884308, + 0.034145258367061615, + 0.052824363112449646, + 0.011829350143671036, + 0.005001978017389774, + 0.005792389158159494, + 0.004889003001153469, + 0.014104337431490421, + 0.005157723557204008, + 0.006327597890049219, + 0.006043560337275267, + 0.0054787518456578255, + 0.007098953239619732, + 0.005427869968116283, + 0.01440766267478466, + 0.007953678257763386, + 0.01360283512622118, + 0.01766025833785534, + 0.013975274749100208, + 0.018816456198692322, + 0.00483681121841073, + 0.0021134696435183287, + 0.0028797569684684277, + 0.0030655006412416697, + 0.004337516613304615, + 0.002565349219366908, + 0.005149096250534058, + 0.00431105075404048, + 0.0068024322390556335, + 0.003815027419477701, + 0.015373281203210354, + 0.007866292260587215, + 0.005307406652718782, + 0.010655642487108707, + 0.0051943473517894745, + 0.00544184073805809, + 0.005971922073513269, + 0.008412166498601437, + 0.009405887685716152, + 0.007105764001607895, + 0.0051275454461574554, + 0.007878029718995094, + 0.01122500654309988, + 0.008925286121666431, + 0.0030922514852136374, + 0.002298135543242097, + 0.020613333210349083, + 0.004326464142650366, + 0.001876902999356389, + 0.019096314907073975, + 0.002436731941998005, + 0.002603909233585, + 0.0015997859882190824, + 0.003010029438883066, + 0.0024581286124885082, + 0.0035206058528274298, + 0.001961776055395603, + 0.0018370950128883123, + 0.002370063681155443, + 0.002772629726678133, + 0.058061327785253525, + 0.013864865526556969, + 0.06913058459758759, + 0.021243512630462646, + 0.0060601565055549145, + 0.005585333798080683, + 0.024456774815917015, + 0.0077703832648694515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.002497175009921193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703493237495422 + ], + "opposite_logprob_tokens": [ + -2.127034902572632 + ], + "logprob_reward": -0.12703493237495422, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703493237495422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49751824140548706, + 0.06330045312643051, + 0.14698395133018494, + 0.045570965856313705, + 0.021066896617412567, + 0.04016057774424553, + 0.018734904006123543, + 0.10942519456148148, + 0.007644048426300287, + 0.049594778567552567 + ], + "credit_map": [ + 0.09584496915340424, + 0.026058310642838478, + 0.02805335633456707, + 0.02085983008146286, + 0.1032819151878357, + 0.1699841171503067, + 0.029959144070744514, + 0.023476609960198402, + 0.006715371273458004, + 0.0048021962866187096, + 0.004465530626475811, + 0.005990311037749052, + 0.008436397649347782, + 0.010740871541202068, + 0.017059562727808952, + 0.005090209189802408, + 0.01640043407678604, + 0.021559787914156914, + 0.017953351140022278, + 0.029324468225240707, + 0.014320965856313705, + 0.01718616671860218, + 0.017928918823599815, + 0.012309854850172997, + 0.0043914769776165485, + 0.0030454599764198065, + 0.006205314304679632, + 0.0041257827542722225, + 0.005794628523290157, + 0.00766407698392868, + 0.006022167857736349, + 0.008322058245539665, + 0.0031226598657667637, + 0.0018993692938238382, + 0.0023559164255857468, + 0.002864579204469919, + 0.0029205975588411093, + 0.0017025290289893746, + 0.0028491949196904898, + 0.0033520504366606474, + 0.003408395918086171, + 0.0024020664859563112, + 0.011319126933813095, + 0.00413997657597065, + 0.0033179577440023422, + 0.0093910600990057, + 0.0029110198374837637, + 0.0032709725201129913, + 0.0022821766324341297, + 0.0032671652734279633, + 0.002831696765497327, + 0.0015777470543980598, + 0.0022953765001147985, + 0.0023498041555285454, + 0.0027841292321681976, + 0.001346808741800487, + 0.0046581365168094635, + 0.004051283001899719, + 0.04219287633895874, + 0.00644342415034771, + 0.002917720703408122, + 0.04150564596056938, + 0.003413704689592123, + 0.004242398776113987, + 0.0007560905651189387, + 0.0010663995053619146, + 0.0009653366287238896, + 0.0011847488349303603, + 0.0006426364416256547, + 0.0008004637202247977, + 0.0008833722095005214, + 0.0013449997641146183, + 0.013750025071203709, + 0.003353965003043413, + 0.016404926776885986, + 0.005065530072897673, + 0.0018110214732587337, + 0.0013216427760198712, + 0.005918204318732023, + 0.0019694624934345484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.12703493237495422 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897930592298508 + ], + "opposite_logprob_tokens": [ + -2.578979253768921 + ], + "logprob_reward": -0.07897930592298508, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897930592298508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2595694363117218, + 0.13228632509708405, + 0.0470123216509819, + 0.07673799991607666, + 0.060530491173267365, + 0.0604344941675663, + 0.05554380267858505, + 0.06948953121900558, + 0.03270906209945679, + 0.20568647980690002 + ], + "credit_map": [ + 0.051649682223796844, + 0.015745293349027634, + 0.01992214471101761, + 0.015144423581659794, + 0.05063251405954361, + 0.07592596858739853, + 0.015537727624177933, + 0.015011700801551342, + 0.009451152756810188, + 0.007646271027624607, + 0.006447385530918837, + 0.006720798090100288, + 0.015601418912410736, + 0.028561025857925415, + 0.04600127041339874, + 0.011857006698846817, + 0.00528997927904129, + 0.005671029910445213, + 0.006346038542687893, + 0.005634484812617302, + 0.004412394016981125, + 0.005359278991818428, + 0.006621695123612881, + 0.007677425630390644, + 0.0059908414259552956, + 0.005637068767100573, + 0.01071416400372982, + 0.007160631939768791, + 0.008784097619354725, + 0.012786445207893848, + 0.01113981194794178, + 0.01452493853867054, + 0.011137806810438633, + 0.006372635718435049, + 0.0071624889969825745, + 0.009271970018744469, + 0.005064208991825581, + 0.005255155730992556, + 0.0065750363282859325, + 0.009691187180578709, + 0.005368881858885288, + 0.007288205437362194, + 0.007128346711397171, + 0.006165720522403717, + 0.010281721130013466, + 0.0073441555723547935, + 0.009465786628425121, + 0.007391676772385836, + 0.005546044558286667, + 0.004592768382281065, + 0.008147720247507095, + 0.006132794078439474, + 0.010370888747274876, + 0.007729813456535339, + 0.006256622262299061, + 0.006767147220671177, + 0.003770790295675397, + 0.004327470902353525, + 0.023455094546079636, + 0.007029606495052576, + 0.0038442653603851795, + 0.019204651936888695, + 0.0041202520951628685, + 0.00373739842325449, + 0.0035861849319189787, + 0.003989888355135918, + 0.004655640572309494, + 0.004068737383931875, + 0.0025750072672963142, + 0.003705165581777692, + 0.004764245357364416, + 0.00536419078707695, + 0.057287655770778656, + 0.013463921844959259, + 0.06764949858188629, + 0.02072916366159916, + 0.006486633326858282, + 0.006927912589162588, + 0.02415521815419197, + 0.008986479602754116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07897930592298508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is one phoenix composed entirely of flowing magma, incandescent fire, and crimson/gold feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the other phoenix carved from translucent ice crystals with sapphire/cyan feathers and icy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a turbulent sky swirling with dark storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly rotate around the two phoenixes during the battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the phoenixes collide head-on causing sparks to hiss into steam and ice crystals to explode into fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979809045791626 + ], + "opposite_logprob_tokens": [ + -3.5297980308532715 + ], + "logprob_reward": -0.02979809045791626, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979809045791626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08333311975002289, + 0.26270318031311035, + 0.03218510001897812, + 0.05749962851405144, + 0.11175336688756943, + 0.0950334444642067, + 0.06923004239797592, + 0.08480511605739594, + 0.02834669128060341, + 0.1751103699207306 + ], + "credit_map": [ + 0.010927454568445683, + 0.012916205450892448, + 0.007440862711519003, + 0.005946225952357054, + 0.016581958159804344, + 0.010525640100240707, + 0.011710044927895069, + 0.007284731138497591, + 0.019136404618620872, + 0.014717409387230873, + 0.00836425181478262, + 0.010904790833592415, + 0.02195250615477562, + 0.06721965968608856, + 0.09858234226703644, + 0.021825825795531273, + 0.0029488662257790565, + 0.0032646378967911005, + 0.0034702783450484276, + 0.005696233361959457, + 0.002738810842856765, + 0.005635545589029789, + 0.004325421992689371, + 0.004105305299162865, + 0.005869904533028603, + 0.004478731192648411, + 0.009500750340521336, + 0.005718992091715336, + 0.0070752971805632114, + 0.008486011996865273, + 0.007502635009586811, + 0.00886730756610632, + 0.014055348932743073, + 0.010008021257817745, + 0.012292133644223213, + 0.01463952288031578, + 0.018719468265771866, + 0.009570732712745667, + 0.019017294049263, + 0.013450846076011658, + 0.007233554031699896, + 0.0048996624536812305, + 0.030382631346583366, + 0.009939239360392094, + 0.006049706134945154, + 0.024007031694054604, + 0.005175948143005371, + 0.007345674093812704, + 0.008059760555624962, + 0.011021319776773453, + 0.011863934807479382, + 0.007650204934179783, + 0.00832170620560646, + 0.008878463879227638, + 0.007808908354490995, + 0.0056257424876093864, + 0.003486422821879387, + 0.003350198734551668, + 0.03305640071630478, + 0.005531370639801025, + 0.002735320245847106, + 0.031211184337735176, + 0.0026151183992624283, + 0.002819100162014365, + 0.0026535477954894304, + 0.004023065324872732, + 0.0035653566010296345, + 0.004782342817634344, + 0.002705412218347192, + 0.002775569912046194, + 0.0036234718281775713, + 0.004217925481498241, + 0.04829913005232811, + 0.011762340553104877, + 0.058542292565107346, + 0.01808074489235878, + 0.0059079499915242195, + 0.005097258370369673, + 0.021079326048493385, + 0.00634132232517004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.02979809045791626 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_tensor([4999, 4999])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_tensor([4999, 4999])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2fc06da7281e9bf1f00a1d7bc175d8e7a4568278 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_tensor([4999, 4999])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic zoom out, the camera slowly panning back and upward, opens on a stunning woman clad in flowing ancient costume. Her cascading jet-black hair is meticulously tied into an intricate, classical bun, adorned with a few jade hairpins radiating a gentle glow. Strands of hair cascade down, floating gently in the weightless space. She wears a layered, moon-white silk Hanfu (Chinese traditional clothing), with wide sleeves and a long ribbon at her waist flowing silently and dancing behind her like ripples on water. In her arms, she clutches a snow-white jade rabbit, its eyes as shimmering as rubies. Curled in peace, it stands in stark contrast to its owner's anxiety. Involuntarily, slowly and helplessly, she drifts toward the depths of the frame, away from the shrinking mountains and rivers below, shrouded in the night. Against the backdrop of the distant universe, a majestic lunar palace, carved from ice crystals and white jade, quietly hovers, emitting a cold and lonely silver glow. This light is the primary source of the entire scene, outlining a sharp, icy outline on the woman's profile, the edge of her robe, and the fur of the Jade Rabbit. With difficulty, she turns her head behind her, casting a last, affectionate gaze back at the land that gave her birth. Her delicate face is tearless, but her slightly furrowed brow and her sparkling eyes, filled with endless reluctance and heartbreak, silently speak of a grand and sorrowful farewell. The entire scene is shrouded in a grand, serene, yet fatalistic atmosphere of sorrow.", + "seed": 12978, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_4999_seed_12978_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03895972166210413, + "vlm_logprob": -0.03895972166210413, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.03895972166210413, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08400828391313553 + ], + "opposite_logprob_tokens": [ + -2.70900821685791 + ], + "logprob_reward": -0.08400828391313553, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08400828391313553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12106902152299881, + 0.056233156472444534, + 0.045393142849206924, + 0.03971688821911812, + 0.14034247398376465, + 0.10829460620880127, + 0.12585392594337463, + 0.1608482450246811, + 0.16404804587364197, + 0.03820047900080681 + ], + "credit_map": [ + 0.014598693698644638, + 0.010972760617733002, + 0.016282806172966957, + 0.011825483292341232, + 0.013370903208851814, + 0.019834864884614944, + 0.014960018917918205, + 0.019223488867282867, + 0.008268844336271286, + 0.005237474571913481, + 0.007062908262014389, + 0.008426876738667488, + 0.005480257328599691, + 0.005861640442162752, + 0.008699469268321991, + 0.007195688784122467, + 0.004886094480752945, + 0.005249202251434326, + 0.003463295754045248, + 0.0047261668369174, + 0.009396049194037914, + 0.006079774349927902, + 0.006610339041799307, + 0.004982221405953169, + 0.005685984157025814, + 0.004498767666518688, + 0.005865997169166803, + 0.0037570844870060682, + 0.0056285071186721325, + 0.005242415238171816, + 0.004450345411896706, + 0.004587783943861723, + 0.024110952392220497, + 0.014688180759549141, + 0.024673888459801674, + 0.013753511011600494, + 0.013658403418958187, + 0.01991584151983261, + 0.01094035804271698, + 0.01860133744776249, + 0.017203373834490776, + 0.009729322977364063, + 0.011720474809408188, + 0.008506185375154018, + 0.011892327107489109, + 0.03205695003271103, + 0.007892017252743244, + 0.009293957613408566, + 0.016010094434022903, + 0.015275963582098484, + 0.016419032588601112, + 0.009296824224293232, + 0.018839791417121887, + 0.023199301213026047, + 0.014209725894033909, + 0.012603200040757656, + 0.053965531289577484, + 0.013091064058244228, + 0.05211584270000458, + 0.006905375514179468, + 0.006244991440325975, + 0.014212201349437237, + 0.005390518810600042, + 0.00892273522913456, + 0.0168455857783556, + 0.02864968776702881, + 0.036513037979602814, + 0.009740049950778484, + 0.00997835025191307, + 0.04324137791991234, + 0.007022189442068338, + 0.012057765386998653, + 0.0061731343157589436, + 0.0050222137942910194, + 0.005285524297505617, + 0.005533107556402683, + 0.0035984497517347336, + 0.004463629797101021, + 0.004169295076280832, + 0.003955123946070671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.08400828391313553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422950066626072 + ], + "opposite_logprob_tokens": [ + -4.2642292976379395 + ], + "logprob_reward": -0.01422950066626072, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422950066626072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15947376191616058, + 0.09654941409826279, + 0.04423751309514046, + 0.03222266212105751, + 0.1214321106672287, + 0.058593958616256714, + 0.11836200952529907, + 0.14353200793266296, + 0.16388070583343506, + 0.06171589717268944 + ], + "credit_map": [ + 0.019720766693353653, + 0.01039144117385149, + 0.023400789126753807, + 0.024091841652989388, + 0.016481559723615646, + 0.016907157376408577, + 0.031644128262996674, + 0.016836069524288177, + 0.01090006623417139, + 0.009548085741698742, + 0.013145600445568562, + 0.013022498227655888, + 0.015754058957099915, + 0.009290519170463085, + 0.012500371783971786, + 0.012388208881020546, + 0.004691167268902063, + 0.0065298499539494514, + 0.004467436578124762, + 0.004727618303149939, + 0.008097582496702671, + 0.005499699153006077, + 0.0056093959137797356, + 0.0046147676184773445, + 0.004804837517440319, + 0.003480693558230996, + 0.004575769882649183, + 0.0033762615639716387, + 0.003283449448645115, + 0.004677074961364269, + 0.0032693941611796618, + 0.004755182657390833, + 0.011467110365629196, + 0.009671867825090885, + 0.015222370624542236, + 0.014311307109892368, + 0.009142306633293629, + 0.025240419432520866, + 0.015750443562865257, + 0.020626282319426537, + 0.009078034199774265, + 0.005566950421780348, + 0.006286990828812122, + 0.003976034466177225, + 0.006381489802151918, + 0.014713277108967304, + 0.005159156396985054, + 0.007432021666318178, + 0.016096586361527443, + 0.012499185279011726, + 0.01162207406014204, + 0.008249395526945591, + 0.015841320157051086, + 0.031519223004579544, + 0.013912095688283443, + 0.008622130379080772, + 0.040646862238645554, + 0.011509686708450317, + 0.037876758724451065, + 0.00781678594648838, + 0.0074193538166582584, + 0.02054695598781109, + 0.008582810871303082, + 0.00913279503583908, + 0.015514956787228584, + 0.027720246464014053, + 0.03862522542476654, + 0.009441811591386795, + 0.010896464809775352, + 0.04057390242815018, + 0.009774641133844852, + 0.011333455331623554, + 0.008548990823328495, + 0.008690389804542065, + 0.008386391215026379, + 0.0053403726778924465, + 0.005181583110243082, + 0.010949366725981236, + 0.005444389767944813, + 0.009174413047730923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.01422950066626072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372495621442795 + ], + "opposite_logprob_tokens": [ + -3.4087250232696533 + ], + "logprob_reward": -0.03372495621442795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372495621442795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1021227166056633, + 0.15101365745067596, + 0.037318889051675797, + 0.023084085434675217, + 0.05605898052453995, + 0.05021122470498085, + 0.38885098695755005, + 0.05261530354619026, + 0.0929417535662651, + 0.04578239098191261 + ], + "credit_map": [ + 0.014606124721467495, + 0.006670704111456871, + 0.012015542015433311, + 0.01569291390478611, + 0.009592813439667225, + 0.01067377720028162, + 0.021320274099707603, + 0.011550567112863064, + 0.019695818424224854, + 0.014207880012691021, + 0.022644205018877983, + 0.017828967422246933, + 0.016016189008951187, + 0.018604377284646034, + 0.02430451288819313, + 0.017711706459522247, + 0.0059515866450965405, + 0.004546553827822208, + 0.003876936621963978, + 0.004412543959915638, + 0.006092926487326622, + 0.0049420674331486225, + 0.004552105907350779, + 0.0029441684018820524, + 0.0028785846661776304, + 0.0029325373470783234, + 0.002879312029108405, + 0.0022175689227879047, + 0.003444384317845106, + 0.0032827560789883137, + 0.002331083407625556, + 0.0031178579665720463, + 0.004669236950576305, + 0.006507770158350468, + 0.005429136101156473, + 0.007528832647949457, + 0.005814791657030582, + 0.010661441832780838, + 0.007420955691486597, + 0.008026812225580215, + 0.010811184532940388, + 0.005317963194102049, + 0.007186616770923138, + 0.0030108538921922445, + 0.004871990066021681, + 0.012325678020715714, + 0.002521415241062641, + 0.004165521822869778, + 0.04981241747736931, + 0.0500597208738327, + 0.04223279282450676, + 0.022671744227409363, + 0.05832531675696373, + 0.09635312855243683, + 0.03611107915639877, + 0.03328477591276169, + 0.0119909243658185, + 0.004958834964782, + 0.010956862010061741, + 0.0037573568988591433, + 0.0033676119055598974, + 0.01020145881921053, + 0.003620434319600463, + 0.003761818865314126, + 0.008742738515138626, + 0.01397024467587471, + 0.014617628417909145, + 0.005857751239091158, + 0.007186554372310638, + 0.02704816870391369, + 0.007915863767266273, + 0.007602805271744728, + 0.004994644317775965, + 0.005751462187618017, + 0.0061957319267094135, + 0.004906495567411184, + 0.004098064731806517, + 0.008874321356415749, + 0.0042235697619616985, + 0.00673809926956892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.03372495621442795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864444583654404 + ], + "opposite_logprob_tokens": [ + -3.048644542694092 + ], + "logprob_reward": -0.04864444583654404, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864444583654404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04625329002737999, + 0.027035336941480637, + 0.017470164224505424, + 0.01608462445437908, + 0.043583258986473083, + 0.10299486666917801, + 0.08219948410987854, + 0.38042017817497253, + 0.24554991722106934, + 0.03840888291597366 + ], + "credit_map": [ + 0.007042330224066973, + 0.002303928369656205, + 0.007894244976341724, + 0.005260760895907879, + 0.003446728689596057, + 0.008330143056809902, + 0.006886640563607216, + 0.005088512320071459, + 0.004915965721011162, + 0.0016300396528095007, + 0.0038279122672975063, + 0.0029869014397263527, + 0.0029034377075731754, + 0.003906619735062122, + 0.0030199619941413403, + 0.003844499122351408, + 0.003996819723397493, + 0.0016106002731248736, + 0.0015755537897348404, + 0.0016626531723886728, + 0.001626405050046742, + 0.002605991205200553, + 0.0019060283666476607, + 0.0024861120618879795, + 0.001897236448712647, + 0.0020532042253762484, + 0.0023747687228024006, + 0.0012271781452000141, + 0.0026038819923996925, + 0.0025357052218168974, + 0.0014855930348858237, + 0.001907057361677289, + 0.006386328488588333, + 0.003862327430397272, + 0.007233752869069576, + 0.003378656692802906, + 0.003374057123437524, + 0.007334137335419655, + 0.003907074220478535, + 0.008106924593448639, + 0.020223204046487808, + 0.009630986489355564, + 0.016118360683321953, + 0.005608770530670881, + 0.007904073223471642, + 0.029210861772298813, + 0.0049012391828000546, + 0.009397370740771294, + 0.010361223481595516, + 0.007639219984412193, + 0.010238499380648136, + 0.0052725328132510185, + 0.010725973173975945, + 0.020717760547995567, + 0.008146071806550026, + 0.009098201058804989, + 0.1259365975856781, + 0.02878396213054657, + 0.11525887995958328, + 0.01487897988408804, + 0.014749530702829361, + 0.04434638097882271, + 0.01484735868871212, + 0.021618492901325226, + 0.023428097367286682, + 0.04320559650659561, + 0.0615786612033844, + 0.015315723605453968, + 0.013633131049573421, + 0.060611870139837265, + 0.010617554187774658, + 0.017159273847937584, + 0.005594007205218077, + 0.006781204603612423, + 0.006304703652858734, + 0.0027358687948435545, + 0.002803461393341422, + 0.006223440635949373, + 0.0031047272495925426, + 0.004861467517912388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.04864444583654404 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191421680152416 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191421680152416, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191421680152416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1269240379333496, + 0.07002859562635422, + 0.05508328601717949, + 0.02043774165213108, + 0.08045082539319992, + 0.13295362889766693, + 0.08135145157575607, + 0.23237669467926025, + 0.1352805197238922, + 0.06511323153972626 + ], + "credit_map": [ + 0.016590677201747894, + 0.006339505314826965, + 0.02149413339793682, + 0.014188502915203571, + 0.007483739871531725, + 0.024030037224292755, + 0.020234329625964165, + 0.016563110053539276, + 0.012820987962186337, + 0.006484676618129015, + 0.009644138626754284, + 0.007033676374703646, + 0.0051172226667404175, + 0.00839998945593834, + 0.00721811642870307, + 0.013309785164892673, + 0.01223921962082386, + 0.0057357363402843475, + 0.004872255492955446, + 0.004814838990569115, + 0.0070343660190701485, + 0.008068811148405075, + 0.004780454561114311, + 0.0075376033782958984, + 0.0031141259241849184, + 0.002114933682605624, + 0.002847605152055621, + 0.002883191918954253, + 0.00241776742041111, + 0.0028055778238922358, + 0.002448608400300145, + 0.0018059294670820236, + 0.00827796384692192, + 0.006502319127321243, + 0.01197815965861082, + 0.008693345822393894, + 0.005361887626349926, + 0.01777227781713009, + 0.009699186310172081, + 0.012165682390332222, + 0.022665970027446747, + 0.01064847782254219, + 0.017194176092743874, + 0.00924758706241846, + 0.011501818895339966, + 0.04165642336010933, + 0.006272819358855486, + 0.013766353949904442, + 0.010068651288747787, + 0.00909423828125, + 0.009683907963335514, + 0.0057844617404043674, + 0.011005934327840805, + 0.019832663238048553, + 0.007384791970252991, + 0.00849680881947279, + 0.08902646601200104, + 0.018549393862485886, + 0.08283256739377975, + 0.007256637327373028, + 0.006053000222891569, + 0.009842637926340103, + 0.004961744416505098, + 0.013854247517883778, + 0.012697330676019192, + 0.023578930646181107, + 0.02688913233578205, + 0.008317939005792141, + 0.008404908701777458, + 0.03659915551543236, + 0.006991299334913492, + 0.01180182583630085, + 0.010075189173221588, + 0.009958893992006779, + 0.007977618835866451, + 0.004131775349378586, + 0.004673769697546959, + 0.012929386459290981, + 0.004397227894514799, + 0.01096937246620655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.014191421680152416 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic zoom out, the camera slowly panning back and upward, opens on a stunning woman clad in flowing ancient costume. Her cascading jet-black hair is meticulously tied into an intricate, classical bun, adorned with a few jade hairpins radiating a gentle glow. Strands of hair cascade down, floating gently in the weightless space. She wears a layered, moon-white silk Hanfu (Chinese traditional clothing), with wide sleeves and a long ribbon at her waist flowing silently and dancing behind her like ripples on water. In her arms, she clutches a snow-white jade rabbit, its eyes as shimmering as rubies. Curled in peace, it stands in stark contrast to its owner's anxiety. Involuntarily, slowly and helplessly, she drifts toward the depths of the frame, away from the shrinking mountains and rivers below, shrouded in the night. Against the backdrop of the distant universe, a majestic lunar palace, carved from ice crystals and white jade, quietly hovers, emitting a cold and lonely silver glow. This light is the primary source of the entire scene, outlining a sharp, icy outline on the woman's profile, the edge of her robe, and the fur of the Jade Rabbit. With difficulty, she turns her head behind her, casting a last, affectionate gaze back at the land that gave her birth. Her delicate face is tearless, but her slightly furrowed brow and her sparkling eyes, filled with endless reluctance and heartbreak, silently speak of a grand and sorrowful farewell. The entire scene is shrouded in a grand, serene, yet fatalistic atmosphere of sorrow.", + "seed": 12979, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_4999_seed_12979_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03577783480286598, + "vlm_logprob": -0.03577783480286598, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.03577783480286598, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09512866288423538 + ], + "opposite_logprob_tokens": [ + -2.8451287746429443 + ], + "logprob_reward": -0.09512866288423538, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09512866288423538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11447785049676895, + 0.06404466181993484, + 0.08406417816877365, + 0.11566488444805145, + 0.11104078590869904, + 0.2389165312051773, + 0.05957050621509552, + 0.10266180336475372, + 0.08132702112197876, + 0.028231775388121605 + ], + "credit_map": [ + 0.017896508798003197, + 0.008899413049221039, + 0.013939733617007732, + 0.008103491738438606, + 0.011806024238467216, + 0.01374061219394207, + 0.016855547204613686, + 0.02323651686310768, + 0.007306235376745462, + 0.007589559070765972, + 0.008309371769428253, + 0.009061317890882492, + 0.005705576390028, + 0.007619153242558241, + 0.00795377604663372, + 0.010499673895537853, + 0.010487573221325874, + 0.009111292660236359, + 0.013363482430577278, + 0.012280664406716824, + 0.010639391839504242, + 0.009154727682471275, + 0.009358299896121025, + 0.0096687451004982, + 0.01878039911389351, + 0.011293692514300346, + 0.01812363602221012, + 0.01577615924179554, + 0.009784393012523651, + 0.01354959886521101, + 0.01157963927835226, + 0.016777370125055313, + 0.010747832246124744, + 0.009063070639967918, + 0.01752668432891369, + 0.013950994238257408, + 0.012414931319653988, + 0.018695134669542313, + 0.015682416036725044, + 0.01295972615480423, + 0.029687922447919846, + 0.019587010145187378, + 0.05092591419816017, + 0.031172972172498703, + 0.02187477983534336, + 0.028199022635817528, + 0.02961971051990986, + 0.02784919925034046, + 0.00791512243449688, + 0.007975559681653976, + 0.005254332907497883, + 0.006196617614477873, + 0.005573201458901167, + 0.007462381385266781, + 0.010351964272558689, + 0.008841324597597122, + 0.011670568026602268, + 0.007530603092163801, + 0.01343671977519989, + 0.01048265676945448, + 0.007489507552236319, + 0.02241443283855915, + 0.0156876090914011, + 0.013949709944427013, + 0.009722749702632427, + 0.008165495470166206, + 0.01190154068171978, + 0.009154013358056545, + 0.00791757833212614, + 0.014829195104539394, + 0.01091549452394247, + 0.008720951154828072, + 0.002601680113002658, + 0.0033322195522487164, + 0.0031843571923673153, + 0.003973385319113731, + 0.0031376900151371956, + 0.00460862135514617, + 0.004384677857160568, + 0.0030091432854533195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.09512866288423538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023325342684984207 + ], + "opposite_logprob_tokens": [ + -3.7733254432678223 + ], + "logprob_reward": -0.023325342684984207, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023325342684984207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08011764287948608, + 0.07572007924318314, + 0.050850287079811096, + 0.05477677285671234, + 0.08036380261182785, + 0.1189403310418129, + 0.21545957028865814, + 0.21609105169773102, + 0.07597878575325012, + 0.031701602041721344 + ], + "credit_map": [ + 0.00755148334428668, + 0.008842244744300842, + 0.009527178481221199, + 0.01471018884330988, + 0.0076593863777816296, + 0.009280439466238022, + 0.01163258496671915, + 0.010914133861660957, + 0.010464498773217201, + 0.009874548763036728, + 0.005414563696831465, + 0.0074507975950837135, + 0.006775420159101486, + 0.0097538772970438, + 0.008887412957847118, + 0.017098959535360336, + 0.0054457662627100945, + 0.00424523139372468, + 0.008058004081249237, + 0.005949581973254681, + 0.00893096812069416, + 0.0063383495435118675, + 0.004980492405593395, + 0.00690189516171813, + 0.008752922527492046, + 0.004422270692884922, + 0.007829644717276096, + 0.00715668685734272, + 0.00513167679309845, + 0.007733007427304983, + 0.005990003701299429, + 0.007760563399642706, + 0.008022196590900421, + 0.008276318199932575, + 0.011068115942180157, + 0.009117146022617817, + 0.008264408446848392, + 0.015176177024841309, + 0.009753350168466568, + 0.01068609394133091, + 0.016542958095669746, + 0.008049186319112778, + 0.024077191948890686, + 0.013639585115015507, + 0.009692658670246601, + 0.016632165759801865, + 0.015166543424129486, + 0.015140051953494549, + 0.04619057849049568, + 0.021366283297538757, + 0.0216447114944458, + 0.019061045721173286, + 0.013756089843809605, + 0.015372229740023613, + 0.026940198615193367, + 0.051128435879945755, + 0.03240908682346344, + 0.017400039359927177, + 0.02205638214945793, + 0.01736081764101982, + 0.015160671435296535, + 0.06224687024950981, + 0.021182671189308167, + 0.02827451378107071, + 0.011424808762967587, + 0.007670385297387838, + 0.010416303761303425, + 0.009593176655471325, + 0.005770835094153881, + 0.013269404880702496, + 0.008533953689038754, + 0.009299919940531254, + 0.003044315380975604, + 0.00325377588160336, + 0.0035875183530151844, + 0.004049451090395451, + 0.0034785845782607794, + 0.005009753629565239, + 0.0047277482226490974, + 0.004550455138087273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.023325342684984207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033712275326251984 + ], + "opposite_logprob_tokens": [ + -3.408712387084961 + ], + "logprob_reward": -0.033712275326251984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033712275326251984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09224382787942886, + 0.1491035372018814, + 0.08532475680112839, + 0.04945337772369385, + 0.14422646164894104, + 0.10417668521404266, + 0.25638139247894287, + 0.04101192206144333, + 0.0542939119040966, + 0.023784134536981583 + ], + "credit_map": [ + 0.010100162588059902, + 0.00924085546284914, + 0.010660035535693169, + 0.013402743265032768, + 0.008594965562224388, + 0.010909444652497768, + 0.011408458463847637, + 0.01792716234922409, + 0.016860488802194595, + 0.015550335869193077, + 0.013491985388100147, + 0.019646700471639633, + 0.014036247506737709, + 0.017400968819856644, + 0.019677691161632538, + 0.03243912756443024, + 0.009596336632966995, + 0.008440054021775723, + 0.012685404159128666, + 0.016317352652549744, + 0.00956051703542471, + 0.008227412588894367, + 0.010809427127242088, + 0.009688254445791245, + 0.006171651184558868, + 0.0039731590077281, + 0.009390775114297867, + 0.00688230199739337, + 0.004688747692853212, + 0.007153387181460857, + 0.004609872121363878, + 0.006583480630069971, + 0.012489117681980133, + 0.019202755764126778, + 0.018189160153269768, + 0.022640075534582138, + 0.012094711884856224, + 0.016787881031632423, + 0.02542939782142639, + 0.017393359914422035, + 0.015443031676113605, + 0.006682055536657572, + 0.020134516060352325, + 0.010771703906357288, + 0.008501790463924408, + 0.015881646424531937, + 0.01239976566284895, + 0.014362169429659843, + 0.04191744327545166, + 0.015898318961262703, + 0.048248302191495895, + 0.024034878239035606, + 0.030977092683315277, + 0.03716370835900307, + 0.021610287949442863, + 0.03653135523200035, + 0.005064710043370724, + 0.003502525622025132, + 0.006497806869447231, + 0.0035352364648133516, + 0.0026019206270575523, + 0.009581619873642921, + 0.00421589519828558, + 0.0060122087597846985, + 0.007770748343318701, + 0.005101669114083052, + 0.009096715599298477, + 0.006245503667742014, + 0.0053704045712947845, + 0.00690390495583415, + 0.006107967346906662, + 0.007696998305618763, + 0.0024503995664417744, + 0.0027827585581690073, + 0.0025866935029625893, + 0.0032402705401182175, + 0.0024412996135652065, + 0.005013203714042902, + 0.0025109013076871634, + 0.002758605871349573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.033712275326251984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533470056951046 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533470056951046, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533470056951046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06680405139923096, + 0.031406790018081665, + 0.024833522737026215, + 0.03501381725072861, + 0.1392439305782318, + 0.062069982290267944, + 0.05705530196428299, + 0.41413792967796326, + 0.12739679217338562, + 0.04203785955905914 + ], + "credit_map": [ + 0.007935326546430588, + 0.0054878657683730125, + 0.006455492693930864, + 0.006708817090839148, + 0.004649359267205, + 0.008928005583584309, + 0.010621107183396816, + 0.016018083319067955, + 0.0034185447730123997, + 0.0022123006638139486, + 0.0036689743865281343, + 0.0028322485741227865, + 0.003597539383918047, + 0.004266417119652033, + 0.004215588793158531, + 0.0071951779536902905, + 0.004015390761196613, + 0.0022892779670655727, + 0.0026470550801604986, + 0.002974497387185693, + 0.003668016055598855, + 0.004159847274422646, + 0.0020817427430301905, + 0.002997696865350008, + 0.005578706506639719, + 0.003016358008608222, + 0.003855114569887519, + 0.003930892329663038, + 0.00323949521407485, + 0.005998475011438131, + 0.003417469095438719, + 0.005977308377623558, + 0.014734713360667229, + 0.00893299002200365, + 0.020758355036377907, + 0.012786640785634518, + 0.018057288601994514, + 0.033990196883678436, + 0.011575139127671719, + 0.018408626317977905, + 0.006851377431303263, + 0.0042469990439713, + 0.012099109590053558, + 0.007225691340863705, + 0.0057280114851891994, + 0.010100474581122398, + 0.007138851098716259, + 0.008679470047354698, + 0.008013498038053513, + 0.004891602322459221, + 0.009631152264773846, + 0.005831384100019932, + 0.008351684547960758, + 0.005878291092813015, + 0.006764687597751617, + 0.007693009451031685, + 0.05070572346448898, + 0.028785616159439087, + 0.057421427220106125, + 0.03979097679257393, + 0.024400847032666206, + 0.10189525783061981, + 0.05437539145350456, + 0.05676266551017761, + 0.01997723989188671, + 0.010615191422402859, + 0.02358141541481018, + 0.011221420019865036, + 0.005893445573747158, + 0.024218706414103508, + 0.011423178017139435, + 0.020466193556785583, + 0.0037052594125270844, + 0.004754828289151192, + 0.004427371080964804, + 0.004488908685743809, + 0.002401728415861726, + 0.012258500792086124, + 0.003387743839994073, + 0.006613525096327066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.012533470056951046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189423061907291 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189423061907291, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189423061907291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07910405844449997, + 0.06507545709609985, + 0.04243608936667442, + 0.14319074153900146, + 0.0783543735742569, + 0.12317927926778793, + 0.07839836925268173, + 0.17309489846229553, + 0.15716533362865448, + 0.06000136956572533 + ], + "credit_map": [ + 0.009603493846952915, + 0.008511751890182495, + 0.0068422527983784676, + 0.015787223353981972, + 0.009096157737076283, + 0.00915309228003025, + 0.010374577715992928, + 0.009735512547194958, + 0.005715594626963139, + 0.0049380348064005375, + 0.006048919167369604, + 0.006397734396159649, + 0.008294308558106422, + 0.008917947299778461, + 0.009048307314515114, + 0.0157146193087101, + 0.005189849529415369, + 0.003952106460928917, + 0.004342340398579836, + 0.004955030046403408, + 0.006775336340069771, + 0.007263295818120241, + 0.0040147872641682625, + 0.00594334676861763, + 0.018413301557302475, + 0.007505440153181553, + 0.023536378517746925, + 0.015653038397431374, + 0.020358169451355934, + 0.027164721861481667, + 0.010799187235534191, + 0.019760511815547943, + 0.00637131417170167, + 0.006747256964445114, + 0.011917293071746826, + 0.009609224274754524, + 0.009531383402645588, + 0.016981376335024834, + 0.007432358339428902, + 0.00976417027413845, + 0.01760811358690262, + 0.009247051551938057, + 0.019754426553845406, + 0.012039453722536564, + 0.013176665641367435, + 0.018009280785918236, + 0.013982484117150307, + 0.01936180330812931, + 0.009046305902302265, + 0.011021974496543407, + 0.009364538826048374, + 0.00740879587829113, + 0.01060996949672699, + 0.013064511120319366, + 0.009955598041415215, + 0.007926691323518753, + 0.020148511976003647, + 0.016603590920567513, + 0.030172618106007576, + 0.016061916947364807, + 0.014067135751247406, + 0.03118840791285038, + 0.022734999656677246, + 0.022117726504802704, + 0.020237190648913383, + 0.0148629080504179, + 0.033383458852767944, + 0.012048512697219849, + 0.009916022419929504, + 0.03324601426720619, + 0.012270483188331127, + 0.02120075561106205, + 0.005486650392413139, + 0.0060997819527983665, + 0.007261547725647688, + 0.007525676861405373, + 0.008330224081873894, + 0.010379846207797527, + 0.007065728306770325, + 0.0078519182279706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.014189423061907291 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic zoom out, the camera slowly panning back and upward, opens on a stunning woman clad in flowing ancient costume. Her cascading jet-black hair is meticulously tied into an intricate, classical bun, adorned with a few jade hairpins radiating a gentle glow. Strands of hair cascade down, floating gently in the weightless space. She wears a layered, moon-white silk Hanfu (Chinese traditional clothing), with wide sleeves and a long ribbon at her waist flowing silently and dancing behind her like ripples on water. In her arms, she clutches a snow-white jade rabbit, its eyes as shimmering as rubies. Curled in peace, it stands in stark contrast to its owner's anxiety. Involuntarily, slowly and helplessly, she drifts toward the depths of the frame, away from the shrinking mountains and rivers below, shrouded in the night. Against the backdrop of the distant universe, a majestic lunar palace, carved from ice crystals and white jade, quietly hovers, emitting a cold and lonely silver glow. This light is the primary source of the entire scene, outlining a sharp, icy outline on the woman's profile, the edge of her robe, and the fur of the Jade Rabbit. With difficulty, she turns her head behind her, casting a last, affectionate gaze back at the land that gave her birth. Her delicate face is tearless, but her slightly furrowed brow and her sparkling eyes, filled with endless reluctance and heartbreak, silently speak of a grand and sorrowful farewell. The entire scene is shrouded in a grand, serene, yet fatalistic atmosphere of sorrow.", + "seed": 12980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_4999_seed_12980_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05206274446099997, + "vlm_logprob": -0.05206274446099997, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.05206274446099997, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08805345743894577 + ], + "opposite_logprob_tokens": [ + -2.8380534648895264 + ], + "logprob_reward": -0.08805345743894577, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08805345743894577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14688892662525177, + 0.18262122571468353, + 0.20206762850284576, + 0.08888224512338638, + 0.040744680911302567, + 0.0670674592256546, + 0.05730068311095238, + 0.04658959433436394, + 0.08231784403324127, + 0.08551976829767227 + ], + "credit_map": [ + 0.018675610423088074, + 0.010406487621366978, + 0.00592576339840889, + 0.014081800356507301, + 0.0172142181545496, + 0.031154148280620575, + 0.030832432210445404, + 0.018598467111587524, + 0.011177600361406803, + 0.013163959607481956, + 0.00800608191639185, + 0.01744138076901436, + 0.0408349335193634, + 0.040019962936639786, + 0.029754692688584328, + 0.022222604602575302, + 0.016675375401973724, + 0.019285596907138824, + 0.016767892986536026, + 0.026042480021715164, + 0.02696278877556324, + 0.037587713450193405, + 0.03199045732617378, + 0.02675533853471279, + 0.008035866543650627, + 0.008133897557854652, + 0.01060265302658081, + 0.011922813951969147, + 0.012147205881774426, + 0.012769349850714207, + 0.011579158715903759, + 0.013691302388906479, + 0.004969436675310135, + 0.005960886366665363, + 0.004350391216576099, + 0.005453480873256922, + 0.005824851803481579, + 0.005124859977513552, + 0.004698859062045813, + 0.0043619172647595406, + 0.008788613602519035, + 0.009284267202019691, + 0.009230665862560272, + 0.006919605657458305, + 0.009475449100136757, + 0.005706689320504665, + 0.010582608170807362, + 0.0070795645006000996, + 0.0057928552851080894, + 0.008781103417277336, + 0.0053254677914083, + 0.005012524779886007, + 0.011022697202861309, + 0.00565686821937561, + 0.010065256617963314, + 0.005643912591040134, + 0.00839083082973957, + 0.005927732214331627, + 0.00469845999032259, + 0.0036282239016145468, + 0.00647192494943738, + 0.010728124529123306, + 0.003242853097617626, + 0.0035014422610402107, + 0.008710553869605064, + 0.009001655504107475, + 0.01386185735464096, + 0.008767825551331043, + 0.010613217949867249, + 0.012737585231661797, + 0.008370538242161274, + 0.010254607535898685, + 0.007669001817703247, + 0.015881096944212914, + 0.008064046502113342, + 0.0060462336987257, + 0.017333703115582466, + 0.012977232225239277, + 0.008359955623745918, + 0.009188495576381683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.08805345743894577 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062072791159153 + ], + "opposite_logprob_tokens": [ + -3.895620822906494 + ], + "logprob_reward": -0.02062072791159153, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062072791159153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10108473896980286, + 0.05776684358716011, + 0.13107888400554657, + 0.247430682182312, + 0.07746751606464386, + 0.05159992724657059, + 0.10999736189842224, + 0.0866677463054657, + 0.07333458214998245, + 0.06357166916131973 + ], + "credit_map": [ + 0.016386546194553375, + 0.009178238920867443, + 0.007622020784765482, + 0.013677554205060005, + 0.01583433523774147, + 0.015076098032295704, + 0.011352818459272385, + 0.011957136914134026, + 0.003391014412045479, + 0.006369195878505707, + 0.004577494226396084, + 0.006799220107495785, + 0.01254234928637743, + 0.007064862176775932, + 0.008385413326323032, + 0.008637295104563236, + 0.007627598475664854, + 0.012565028853714466, + 0.009805374778807163, + 0.01768241822719574, + 0.020357495173811913, + 0.027863455936312675, + 0.02038531005382538, + 0.014792211353778839, + 0.017417721450328827, + 0.0178128220140934, + 0.022261707112193108, + 0.019867531955242157, + 0.06211460754275322, + 0.03557803854346275, + 0.03542265668511391, + 0.03695560619235039, + 0.006421098951250315, + 0.0043978882022202015, + 0.006347595248371363, + 0.011253047734498978, + 0.015436442568898201, + 0.01248205080628395, + 0.009524203836917877, + 0.011605189181864262, + 0.0038552654441446066, + 0.0036742144729942083, + 0.0033982801251113415, + 0.005564005579799414, + 0.009453651495277882, + 0.004611687734723091, + 0.012910292483866215, + 0.008132528513669968, + 0.009967243298888206, + 0.014176884666085243, + 0.009680768474936485, + 0.01494921837002039, + 0.0229664109647274, + 0.013287495821714401, + 0.011548199690878391, + 0.013421153649687767, + 0.013290010392665863, + 0.009041284210979939, + 0.020106438547372818, + 0.006670191418379545, + 0.010509868152439594, + 0.01867583580315113, + 0.004123348277062178, + 0.004250776022672653, + 0.011932534165680408, + 0.010310494340956211, + 0.0116098802536726, + 0.0056620691902935505, + 0.007793115917593241, + 0.011606311425566673, + 0.007254210766404867, + 0.007165968883782625, + 0.007145989686250687, + 0.006144244689494371, + 0.010367345996201038, + 0.00634485948830843, + 0.00655552139505744, + 0.008506777696311474, + 0.00569399306550622, + 0.012812944129109383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.02062072791159153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498148873448372 + ], + "opposite_logprob_tokens": [ + -2.9299814701080322 + ], + "logprob_reward": -0.05498148873448372, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498148873448372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1856289803981781, + 0.09131541103124619, + 0.04509464278817177, + 0.47934436798095703, + 0.061319027096033096, + 0.042310930788517, + 0.015771375969052315, + 0.023463943973183632, + 0.03680622577667236, + 0.018945124000310898 + ], + "credit_map": [ + 0.013199137523770332, + 0.010603342205286026, + 0.008419772610068321, + 0.014800095930695534, + 0.022567376494407654, + 0.04961137846112251, + 0.04856467619538307, + 0.017863193526864052, + 0.0060582286678254604, + 0.006183323450386524, + 0.0040995534509420395, + 0.011360906064510345, + 0.024125533178448677, + 0.012521038763225079, + 0.015410895459353924, + 0.011555940844118595, + 0.003337391884997487, + 0.0053748381324112415, + 0.003702606773003936, + 0.006016157101839781, + 0.006441311910748482, + 0.009167338721454144, + 0.006129635497927666, + 0.004925366025418043, + 0.03160955756902695, + 0.03413219377398491, + 0.04082075506448746, + 0.050370581448078156, + 0.11528348922729492, + 0.06867051869630814, + 0.06952117383480072, + 0.06893609464168549, + 0.0033971441444009542, + 0.005066294688731432, + 0.003972847014665604, + 0.007122739218175411, + 0.011487486772239208, + 0.008470720611512661, + 0.013476848602294922, + 0.008324948139488697, + 0.0029758336022496223, + 0.00364857935346663, + 0.0025704368017613888, + 0.005015724338591099, + 0.005173006560653448, + 0.0036367648281157017, + 0.012559832073748112, + 0.006730755791068077, + 0.0014268660452216864, + 0.0018492820672690868, + 0.0017129791667684913, + 0.002268237294629216, + 0.0027843178249895573, + 0.0017445811536163092, + 0.001685464521870017, + 0.0022996480111032724, + 0.004970615729689598, + 0.0030159137677401304, + 0.0035051123704761267, + 0.001632066909223795, + 0.00252044340595603, + 0.0028117853216826916, + 0.0018000053241848946, + 0.0032080018427222967, + 0.004385548643767834, + 0.004618239589035511, + 0.005932713858783245, + 0.0034349579364061356, + 0.005147258285433054, + 0.005340395960956812, + 0.004068171139806509, + 0.0038789408281445503, + 0.0017875665798783302, + 0.003986954689025879, + 0.002482540439814329, + 0.0012778291711583734, + 0.002569338772445917, + 0.002846845192834735, + 0.0017085728468373418, + 0.0022854767739772797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.05498148873448372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903335779905319 + ], + "opposite_logprob_tokens": [ + -2.464033365249634 + ], + "logprob_reward": -0.08903335779905319, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903335779905319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26893889904022217, + 0.047808874398469925, + 0.1612965166568756, + 0.11514849215745926, + 0.05306198075413704, + 0.031073780730366707, + 0.026977267116308212, + 0.08997046947479248, + 0.07083134353160858, + 0.13489237427711487 + ], + "credit_map": [ + 0.02694186381995678, + 0.0160612091422081, + 0.012396893464028835, + 0.022221524268388748, + 0.030090928077697754, + 0.06771152466535568, + 0.06347935646772385, + 0.030035609379410744, + 0.0031640350352972746, + 0.004689882975071669, + 0.0039292904548347, + 0.005918087437748909, + 0.007938429713249207, + 0.006920799612998962, + 0.008341235108673573, + 0.00690711522474885, + 0.011731326580047607, + 0.01905687339603901, + 0.013339955359697342, + 0.0178852379322052, + 0.023574544116854668, + 0.03181679546833038, + 0.025177381932735443, + 0.01871439814567566, + 0.009754033759236336, + 0.010908498428761959, + 0.009824594482779503, + 0.02015581727027893, + 0.01893066056072712, + 0.01573915034532547, + 0.015032509341835976, + 0.014803222380578518, + 0.005724692717194557, + 0.0033772869501262903, + 0.0049303509294986725, + 0.007261461578309536, + 0.009383604861795902, + 0.007828949019312859, + 0.007650216110050678, + 0.006905419286340475, + 0.003230794332921505, + 0.0019370028749108315, + 0.0034621593076735735, + 0.0031284592114388943, + 0.004210345912724733, + 0.002809707075357437, + 0.007770594209432602, + 0.004524718038737774, + 0.002706520725041628, + 0.0048606423661112785, + 0.0030124534387141466, + 0.0037782106082886457, + 0.003701933892443776, + 0.0027494817040860653, + 0.003354980843141675, + 0.002813042374327779, + 0.01455699186772108, + 0.00977575033903122, + 0.02598685398697853, + 0.00830979272723198, + 0.009814423508942127, + 0.009941446594893932, + 0.0057204654440283775, + 0.005864741280674934, + 0.008709512650966644, + 0.009827183559536934, + 0.01097814179956913, + 0.004880159627646208, + 0.008953739888966084, + 0.010636094026267529, + 0.008129667490720749, + 0.008716844953596592, + 0.01503741554915905, + 0.02571979910135269, + 0.020541643723845482, + 0.01165758166462183, + 0.011788150295615196, + 0.0212032962590456, + 0.007815149612724781, + 0.021129345521330833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.08903335779905319 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624690420925617 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624690420925617, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624690420925617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1622532457113266, + 0.09214236587285995, + 0.07627744972705841, + 0.24022085964679718, + 0.06875530630350113, + 0.03949462249875069, + 0.03840434551239014, + 0.08499780297279358, + 0.10601311177015305, + 0.09144087135791779 + ], + "credit_map": [ + 0.01933165267109871, + 0.014365460723638535, + 0.010830530896782875, + 0.023269830271601677, + 0.024603407829999924, + 0.03152032569050789, + 0.022679917514324188, + 0.015652112662792206, + 0.006370083428919315, + 0.010838326998054981, + 0.008346468210220337, + 0.01151480246335268, + 0.015512293204665184, + 0.014313154853880405, + 0.014379303902387619, + 0.010867932811379433, + 0.007053336128592491, + 0.009273440577089787, + 0.006031691562384367, + 0.01127541996538639, + 0.010476132854819298, + 0.014117100276052952, + 0.009498967789113522, + 0.008551357313990593, + 0.01671683043241501, + 0.017461229115724564, + 0.024134691804647446, + 0.02578476257622242, + 0.053240276873111725, + 0.032585509121418, + 0.03470579907298088, + 0.03559177741408348, + 0.005760928150266409, + 0.003961195703595877, + 0.006165012717247009, + 0.011409496888518333, + 0.011590675450861454, + 0.009237077087163925, + 0.011325143277645111, + 0.009305774234235287, + 0.005137878470122814, + 0.003514196490868926, + 0.0037839622236788273, + 0.004171004984527826, + 0.006048779934644699, + 0.004103614948689938, + 0.007203821558505297, + 0.005531363654881716, + 0.0033399933017790318, + 0.004594974685460329, + 0.004181958269327879, + 0.00586872361600399, + 0.005404965486377478, + 0.003374220570549369, + 0.005019645672291517, + 0.006619867868721485, + 0.013675498776137829, + 0.008705800399184227, + 0.020356465131044388, + 0.005736981984227896, + 0.00910875666886568, + 0.01864868402481079, + 0.003716332372277975, + 0.005049287807196379, + 0.014789381995797157, + 0.011236215941607952, + 0.020418863743543625, + 0.008156194351613522, + 0.012159468606114388, + 0.017598025500774384, + 0.008424630388617516, + 0.013230325654149055, + 0.00930425338447094, + 0.014964963309466839, + 0.014161705039441586, + 0.005766463465988636, + 0.014071717858314514, + 0.013576067052781582, + 0.009080996736884117, + 0.010514707304537296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.007624690420925617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic zoom out, the camera slowly panning back and upward, opens on a stunning woman clad in flowing ancient costume. Her cascading jet-black hair is meticulously tied into an intricate, classical bun, adorned with a few jade hairpins radiating a gentle glow. Strands of hair cascade down, floating gently in the weightless space. She wears a layered, moon-white silk Hanfu (Chinese traditional clothing), with wide sleeves and a long ribbon at her waist flowing silently and dancing behind her like ripples on water. In her arms, she clutches a snow-white jade rabbit, its eyes as shimmering as rubies. Curled in peace, it stands in stark contrast to its owner's anxiety. Involuntarily, slowly and helplessly, she drifts toward the depths of the frame, away from the shrinking mountains and rivers below, shrouded in the night. Against the backdrop of the distant universe, a majestic lunar palace, carved from ice crystals and white jade, quietly hovers, emitting a cold and lonely silver glow. This light is the primary source of the entire scene, outlining a sharp, icy outline on the woman's profile, the edge of her robe, and the fur of the Jade Rabbit. With difficulty, she turns her head behind her, casting a last, affectionate gaze back at the land that gave her birth. Her delicate face is tearless, but her slightly furrowed brow and her sparkling eyes, filled with endless reluctance and heartbreak, silently speak of a grand and sorrowful farewell. The entire scene is shrouded in a grand, serene, yet fatalistic atmosphere of sorrow.", + "seed": 12981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_6/index_4999_seed_12981_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029500660486519337, + "vlm_logprob": -0.029500660486519337, + "vlm_yes_no_margin": 3.824999952316284, + "vlm_token_credit": -0.029500660486519337, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05943642184138298 + ], + "opposite_logprob_tokens": [ + -3.559436321258545 + ], + "logprob_reward": -0.05943642184138298, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.05943642184138298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354423463344574, + 0.11999309808015823, + 0.16991567611694336, + 0.14547471702098846, + 0.051061924546957016, + 0.08571242541074753, + 0.10246818512678146, + 0.04012645408511162, + 0.05878208205103874, + 0.09102310985326767 + ], + "credit_map": [ + 0.01655774749815464, + 0.019812684506177902, + 0.014906603842973709, + 0.013988543301820755, + 0.015060308389365673, + 0.022839246317744255, + 0.01730075478553772, + 0.014976456761360168, + 0.021655671298503876, + 0.013306290842592716, + 0.013768435455858707, + 0.00816124863922596, + 0.022986574098467827, + 0.014982708729803562, + 0.013295143842697144, + 0.011837027035653591, + 0.019902300089597702, + 0.012950864620506763, + 0.017234910279512405, + 0.015842491760849953, + 0.02388407289981842, + 0.025298425927758217, + 0.029917193576693535, + 0.024885421618819237, + 0.01811116933822632, + 0.0276231300085783, + 0.014452903531491756, + 0.016280965879559517, + 0.018066510558128357, + 0.014524376019835472, + 0.020692607387900352, + 0.015723051503300667, + 0.008494722656905651, + 0.005517404060810804, + 0.0046457466669380665, + 0.00431823218241334, + 0.006872222293168306, + 0.005856390576809645, + 0.007104730233550072, + 0.008252475410699844, + 0.007348941173404455, + 0.005625904072076082, + 0.00714862858876586, + 0.006428953260183334, + 0.015375308692455292, + 0.007139348424971104, + 0.021106187254190445, + 0.01553915161639452, + 0.01645297184586525, + 0.012981968000531197, + 0.013993141241371632, + 0.007882732897996902, + 0.01015342678874731, + 0.014603203162550926, + 0.013295465148985386, + 0.013105278834700584, + 0.006849713157862425, + 0.004692931193858385, + 0.006419494282454252, + 0.0036932246293872595, + 0.004969405941665173, + 0.0058431681245565414, + 0.0036006940063089132, + 0.004057823680341244, + 0.00674453703686595, + 0.00898442417383194, + 0.005345441866666079, + 0.010843428783118725, + 0.007518026977777481, + 0.007557732984423637, + 0.0066072698682546616, + 0.0051812282763421535, + 0.011136396788060665, + 0.01047139335423708, + 0.012510878965258598, + 0.020469270646572113, + 0.008055884391069412, + 0.010661708191037178, + 0.008817711845040321, + 0.008899874985218048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.05943642184138298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018225817009806633 + ], + "opposite_logprob_tokens": [ + -4.01822566986084 + ], + "logprob_reward": -0.018225817009806633, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018225817009806633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24592024087905884, + 0.19819745421409607, + 0.1292884349822998, + 0.06967226415872574, + 0.055670786648988724, + 0.04268564283847809, + 0.03527397662401199, + 0.11165636032819748, + 0.05975644290447235, + 0.05187835544347763 + ], + "credit_map": [ + 0.013818829320371151, + 0.015346652828156948, + 0.024025967344641685, + 0.03401661664247513, + 0.030036216601729393, + 0.04318242892622948, + 0.04292188584804535, + 0.04257163777947426, + 0.022137466818094254, + 0.014675443060696125, + 0.014107874594628811, + 0.021009448915719986, + 0.03593665361404419, + 0.0360274575650692, + 0.0249861441552639, + 0.029316987842321396, + 0.01776888407766819, + 0.009237760677933693, + 0.012880285270512104, + 0.008940858766436577, + 0.018303873017430305, + 0.02214386686682701, + 0.014154958538711071, + 0.025857960805296898, + 0.010877490974962711, + 0.013907210901379585, + 0.008262738585472107, + 0.008199270814657211, + 0.007433708757162094, + 0.007026162464171648, + 0.007162883412092924, + 0.006802796386182308, + 0.010101943276822567, + 0.007863059639930725, + 0.007025658152997494, + 0.0035155205987393856, + 0.00641791895031929, + 0.005265542771667242, + 0.00665128743276, + 0.008829851634800434, + 0.006781463976949453, + 0.003568169428035617, + 0.00990019366145134, + 0.0030863373540341854, + 0.00443377997726202, + 0.005161228124052286, + 0.004883310291916132, + 0.004871157929301262, + 0.003344923723489046, + 0.004318580497056246, + 0.00427184347063303, + 0.0026895625051110983, + 0.004659725818783045, + 0.004976674448698759, + 0.0042417943477630615, + 0.006770872976630926, + 0.014302420429885387, + 0.012131533585488796, + 0.017186174169182777, + 0.016290826722979546, + 0.011460323818027973, + 0.01588430069386959, + 0.011442289687693119, + 0.012958493083715439, + 0.005452972836792469, + 0.0076347533613443375, + 0.005826519802212715, + 0.011754277162253857, + 0.006136422976851463, + 0.006898188032209873, + 0.00858396291732788, + 0.007469348609447479, + 0.005316433031111956, + 0.004818700719624758, + 0.00793513935059309, + 0.01074537169188261, + 0.005362167488783598, + 0.006382765248417854, + 0.005388552322983742, + 0.0059292251244187355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018225817009806633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371873125433922 + ], + "opposite_logprob_tokens": [ + -3.4087188243865967 + ], + "logprob_reward": -0.03371873125433922, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371873125433922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08750931173563004, + 0.5586512088775635, + 0.10482499748468399, + 0.04372941330075264, + 0.03313962742686272, + 0.027412600815296173, + 0.025275588035583496, + 0.03518405556678772, + 0.03856256231665611, + 0.045710641890764236 + ], + "credit_map": [ + 0.005442483816295862, + 0.005804528947919607, + 0.009083379991352558, + 0.01064744871109724, + 0.008717670105397701, + 0.018355613574385643, + 0.014385741204023361, + 0.015072444453835487, + 0.05755697190761566, + 0.037212617695331573, + 0.05445177108049393, + 0.052551910281181335, + 0.1276128590106964, + 0.09044157713651657, + 0.06704002618789673, + 0.07178348302841187, + 0.014980247244238853, + 0.007573187351226807, + 0.011240409687161446, + 0.012275814078748226, + 0.015149715356528759, + 0.01247255876660347, + 0.015999477356672287, + 0.015133597888052464, + 0.006258344743400812, + 0.006631402298808098, + 0.005185652524232864, + 0.004237407818436623, + 0.005878590978682041, + 0.004918142221868038, + 0.0056028664112091064, + 0.0050170025788247585, + 0.005869686603546143, + 0.0043380106799304485, + 0.00333594367839396, + 0.0019223561976104975, + 0.004873399622738361, + 0.002585831331089139, + 0.004247266333550215, + 0.005967131815850735, + 0.002473028376698494, + 0.0021030546631664038, + 0.006543641909956932, + 0.0022457153536379337, + 0.0036809800658375025, + 0.0034107407554984093, + 0.003346554469317198, + 0.003608885919675231, + 0.003271046094596386, + 0.003342886222526431, + 0.004017021041363478, + 0.002280465792864561, + 0.003134222235530615, + 0.0040872711688280106, + 0.0026503680273890495, + 0.0024923072196543217, + 0.003467096481472254, + 0.0048224302008748055, + 0.0033811568282544613, + 0.006916276179254055, + 0.004078713245689869, + 0.004488580394536257, + 0.004902603570371866, + 0.0031272012274712324, + 0.003090172540396452, + 0.005100936163216829, + 0.0038880067877471447, + 0.006659213919192553, + 0.0039977519772946835, + 0.005037674214690924, + 0.0039011419285088778, + 0.006887662224471569, + 0.005147886928170919, + 0.003714048769325018, + 0.007635948713868856, + 0.008877531625330448, + 0.003095347201451659, + 0.005673113279044628, + 0.005520595703274012, + 0.006046167109161615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.03371873125433922 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343639940023422 + ], + "opposite_logprob_tokens": [ + -3.651343584060669 + ], + "logprob_reward": -0.026343639940023422, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343639940023422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09379744529724121, + 0.08898122608661652, + 0.24524441361427307, + 0.09231264889240265, + 0.08319761604070663, + 0.23114506900310516, + 0.0438297800719738, + 0.04307665675878525, + 0.03150564804673195, + 0.046909552067518234 + ], + "credit_map": [ + 0.010316923260688782, + 0.011396766640245914, + 0.008456908166408539, + 0.009393922053277493, + 0.013013326562941074, + 0.015097119845449924, + 0.013489965349435806, + 0.012632504105567932, + 0.013425315730273724, + 0.009301283396780491, + 0.011547956615686417, + 0.0075250547379255295, + 0.014709245413541794, + 0.011981832794845104, + 0.011018902994692326, + 0.009471619501709938, + 0.029821202158927917, + 0.020062806084752083, + 0.02229723148047924, + 0.020949263125658035, + 0.029713913798332214, + 0.034679677337408066, + 0.046264562755823135, + 0.041455745697021484, + 0.014357179403305054, + 0.018528031185269356, + 0.01209648884832859, + 0.011499791406095028, + 0.009545035660266876, + 0.0064101554453372955, + 0.011654295027256012, + 0.008221661671996117, + 0.013638801872730255, + 0.012601740658283234, + 0.013089601881802082, + 0.006511223502457142, + 0.008246996439993382, + 0.007702403236180544, + 0.008636722341179848, + 0.01277011539787054, + 0.048731185495853424, + 0.02158081904053688, + 0.04171567037701607, + 0.016137897968292236, + 0.021644996479153633, + 0.027043957263231277, + 0.03403793275356293, + 0.020252592861652374, + 0.0058546243235468864, + 0.005270316265523434, + 0.008756773546338081, + 0.0030473601073026657, + 0.004952569957822561, + 0.006637255195528269, + 0.0040702736005187035, + 0.005240599159151316, + 0.006849450990557671, + 0.0046765487641096115, + 0.008055584505200386, + 0.003949947655200958, + 0.00370641122572124, + 0.0058981445617973804, + 0.004600543063133955, + 0.005340022500604391, + 0.0032791439443826675, + 0.0034136215690523386, + 0.0028028858359903097, + 0.005690224003046751, + 0.0036998712457716465, + 0.0040235803462564945, + 0.0038854884915053844, + 0.004710828885436058, + 0.005119608715176582, + 0.00424862140789628, + 0.007374870590865612, + 0.008571544662117958, + 0.0039313677698373795, + 0.0057891011238098145, + 0.005656743422150612, + 0.006217686925083399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.026343639940023422 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a slow zoom out while panning back and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a layered, moon-white silk Hanfu with wide sleeves and a long ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman clutching a snow-white jade rabbit with ruby-colored eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a majestic lunar palace carved from ice crystals and white jade hovering in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a primary silver light source outlining the woman's profile, robe, and the rabbit's fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977869238704443 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.00977869238704443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977869238704443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12105042487382889, + 0.23295587301254272, + 0.11194323748350143, + 0.10599566996097565, + 0.04811205714941025, + 0.04936641827225685, + 0.051352448761463165, + 0.08466305583715439, + 0.11137232184410095, + 0.08318847417831421 + ], + "credit_map": [ + 0.015268862247467041, + 0.009039909578859806, + 0.018134579062461853, + 0.015193031169474125, + 0.010828830301761627, + 0.018091242760419846, + 0.015151608735322952, + 0.01934237778186798, + 0.02451542764902115, + 0.016430072486400604, + 0.027136512100696564, + 0.02939358539879322, + 0.0522599034011364, + 0.03264829143881798, + 0.027011146768927574, + 0.02356095239520073, + 0.01836046203970909, + 0.006986869964748621, + 0.012221151031553745, + 0.010514836758375168, + 0.018536416813731194, + 0.013293322175741196, + 0.015208537690341473, + 0.01682165078818798, + 0.01677371934056282, + 0.02249475196003914, + 0.011546801775693893, + 0.010755159892141819, + 0.01191611960530281, + 0.010415184311568737, + 0.011557336896657944, + 0.010536596179008484, + 0.007301182020455599, + 0.006237328518182039, + 0.004427030216902494, + 0.004768447019159794, + 0.0068764956668019295, + 0.004315944388508797, + 0.007079108618199825, + 0.007106527220457792, + 0.0051483106799423695, + 0.004589334595948458, + 0.01060539297759533, + 0.004391574766486883, + 0.006732653360813856, + 0.0056609464809298515, + 0.00528852641582489, + 0.006949680857360363, + 0.006860732566565275, + 0.006643084809184074, + 0.010837046429514885, + 0.0035998777020722628, + 0.005558962933719158, + 0.007767595816403627, + 0.004984679631888866, + 0.005100471433252096, + 0.013302349485456944, + 0.009303728118538857, + 0.01866982877254486, + 0.009874248877167702, + 0.00634479196742177, + 0.010412181727588177, + 0.006052186246961355, + 0.01070375181734562, + 0.012590290047228336, + 0.01190291065722704, + 0.01262442022562027, + 0.014658307656645775, + 0.012211164459586143, + 0.014440183527767658, + 0.01384492963552475, + 0.01910012774169445, + 0.010560459457337856, + 0.006744913756847382, + 0.008139877580106258, + 0.010625510476529598, + 0.009929671883583069, + 0.013610986061394215, + 0.007669865153729916, + 0.01590719260275364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.00977869238704443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_tensor([3807, 3807])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_tensor([3807, 3807])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bf9721cee565184a2fcb4a265414eb37110f4727 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_tensor([3807, 3807])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle drone tracking shot, the camera practically touching the murky green river surface, moves alongside a massive traditional dragon boat, rapidly advancing. The boat's hull is painted a vibrant red and gold, its sides adorned with dragon scales that gleam in the sunlight. A majestic dragon head, eyes glaring, is intricately carved at the prow. The boat is filled with muscular rowers, uniformly dressed in white vests. They sway back and forth in unison to the powerful beats of the bow drummer, their oars thrusting into the water with astonishing unison. They thrust their paddles back and forth, pushing the boat through the narrow waterway, leaving a trail of surging whitewater in their wake. Lush subtropical vegetation lines the riverbanks, the dense canopy nearly obscuring the sky and adding a sense of oppression to the scene. With every sprint of the dragon boat, the huge splashes of water stirred up by the oars splashed violently onto the drone's lens, instantly blurring the field of vision. The water droplets then quickly slid down, leaving dynamic water marks on the screen. The entire picture, illuminated by the bright afternoon sunlight, was full of high-contrast light and shadow and saturated colors, conveying a raw, explosive energy and a tense and intense competitive atmosphere, full of unparalleled sense of power.", + "seed": 14229, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_3807_seed_14229_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02679753052070737, + "vlm_logprob": -0.02679753052070737, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.02679753052070737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035683080554008484 + ], + "opposite_logprob_tokens": [ + -5.1606831550598145 + ], + "logprob_reward": -0.035683080554008484, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.035683080554008484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19585393369197845, + 0.0497932992875576, + 0.08645890653133392, + 0.12500163912773132, + 0.12825632095336914, + 0.087826669216156, + 0.08724630624055862, + 0.059874195605516434, + 0.07484925538301468, + 0.1048395037651062 + ], + "credit_map": [ + 0.0111629543825984, + 0.016291914507746696, + 0.016555463895201683, + 0.05281182378530502, + 0.010480242781341076, + 0.009612126275897026, + 0.02645483799278736, + 0.052484579384326935, + 0.007790558505803347, + 0.007288339082151651, + 0.008880476467311382, + 0.005690868012607098, + 0.004264308605343103, + 0.004047519061714411, + 0.006689119152724743, + 0.005142112262547016, + 0.010701875202357769, + 0.01009152177721262, + 0.01085920911282301, + 0.009833177551627159, + 0.008889117278158665, + 0.010184206068515778, + 0.016507180407643318, + 0.009392611682415009, + 0.012722994200885296, + 0.02082192711532116, + 0.020176716148853302, + 0.015203452669084072, + 0.011335344985127449, + 0.0147042665630579, + 0.014439146965742111, + 0.015597803518176079, + 0.008514342829585075, + 0.014837518334388733, + 0.010293436236679554, + 0.01805349439382553, + 0.008616629056632519, + 0.025098420679569244, + 0.028526782989501953, + 0.014315692707896233, + 0.009606702253222466, + 0.007929324172437191, + 0.01190165150910616, + 0.010589324869215488, + 0.011904713697731495, + 0.0157270897179842, + 0.009442710317671299, + 0.010725149884819984, + 0.007684974931180477, + 0.012020216323435307, + 0.008617197163403034, + 0.008296804502606392, + 0.007468726951628923, + 0.018500475212931633, + 0.01068840641528368, + 0.01396950427442789, + 0.006907815113663673, + 0.006394666153937578, + 0.009131953120231628, + 0.00995604321360588, + 0.0059104133397340775, + 0.007370224688202143, + 0.007759942207485437, + 0.006443136837333441, + 0.007613170426338911, + 0.012658999301493168, + 0.007725080940872431, + 0.009054017253220081, + 0.007006787229329348, + 0.00788192916661501, + 0.010758655145764351, + 0.012150619179010391, + 0.009251649491488934, + 0.010330552235245705, + 0.017482491210103035, + 0.013025869615375996, + 0.010061588138341904, + 0.018987830728292465, + 0.013110833242535591, + 0.012588690035045147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.035683080554008484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003627982921898365 + ], + "opposite_logprob_tokens": [ + -5.628627777099609 + ], + "logprob_reward": -0.003627982921898365, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003627982921898365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2366485744714737, + 0.0616876557469368, + 0.03181413188576698, + 0.0597083754837513, + 0.05245908349752426, + 0.10971605777740479, + 0.08572757989168167, + 0.268525630235672, + 0.036899421364068985, + 0.05681348964571953 + ], + "credit_map": [ + 0.023698516190052032, + 0.03083086758852005, + 0.03218887746334076, + 0.04246075451374054, + 0.02152755670249462, + 0.018109912052750587, + 0.03762524202466011, + 0.03020685352385044, + 0.008363211527466774, + 0.0064842780120670795, + 0.009608830325305462, + 0.00930397305637598, + 0.005150472745299339, + 0.004726938903331757, + 0.01012273970991373, + 0.007927211001515388, + 0.0035108875017613173, + 0.0034934887662529945, + 0.0042083957232534885, + 0.0037718615494668484, + 0.00311982911080122, + 0.0035480165388435125, + 0.0065941521897912025, + 0.00356749864295125, + 0.0043367487378418446, + 0.006431238260120153, + 0.012142262421548367, + 0.008767428807914257, + 0.005101945251226425, + 0.00719512440264225, + 0.00733815785497427, + 0.008395467884838581, + 0.004309941548854113, + 0.007043096702545881, + 0.0041215731762349606, + 0.006734143011271954, + 0.004874332342296839, + 0.01250888966023922, + 0.008219338022172451, + 0.004647769033908844, + 0.012776827439665794, + 0.010053124278783798, + 0.012706413865089417, + 0.014662697911262512, + 0.011751369573175907, + 0.014246727339923382, + 0.02006673812866211, + 0.013452163897454739, + 0.011838707141578197, + 0.010491961613297462, + 0.00909098144620657, + 0.006823619361966848, + 0.007578514516353607, + 0.013757609762251377, + 0.010170122608542442, + 0.015976062044501305, + 0.025819219648838043, + 0.032286640256643295, + 0.028392475098371506, + 0.029945889487862587, + 0.039795130491256714, + 0.036870960146188736, + 0.042070530354976654, + 0.033344779163599014, + 0.003941465634852648, + 0.005540814716368914, + 0.004263998940587044, + 0.004696886520832777, + 0.0037794860545545816, + 0.003727610455825925, + 0.005225248169153929, + 0.005723914597183466, + 0.006569247227162123, + 0.004335097502917051, + 0.00964260008186102, + 0.004345152992755175, + 0.005730907898396254, + 0.01319152768701315, + 0.004399390425533056, + 0.008599559776484966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.003627982921898365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620541125535965 + ], + "opposite_logprob_tokens": [ + -2.81205415725708 + ], + "logprob_reward": -0.0620541125535965, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620541125535965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32788655161857605, + 0.05460908263921738, + 0.040401965379714966, + 0.17883212864398956, + 0.046249743551015854, + 0.0420953743159771, + 0.05226501449942589, + 0.10245028138160706, + 0.0752437561750412, + 0.07996610552072525 + ], + "credit_map": [ + 0.03584400936961174, + 0.04349970817565918, + 0.043486129492521286, + 0.05426809564232826, + 0.03048948384821415, + 0.01839405857026577, + 0.046884410083293915, + 0.05502067133784294, + 0.00611927779391408, + 0.006060609128326178, + 0.010908635333180428, + 0.006610135547816753, + 0.004481419920921326, + 0.004073368851095438, + 0.009751353412866592, + 0.0066042798571288586, + 0.005752186756581068, + 0.0044847815297544, + 0.006881203968077898, + 0.004179741255939007, + 0.0042596738785505295, + 0.004781876690685749, + 0.005501110106706619, + 0.004561390727758408, + 0.013274681754410267, + 0.014118362218141556, + 0.0310239065438509, + 0.020171500742435455, + 0.024894561618566513, + 0.016064438968896866, + 0.02867310121655464, + 0.030611570924520493, + 0.003616976784542203, + 0.004796783905476332, + 0.004336994141340256, + 0.00671447953209281, + 0.004196367226541042, + 0.008447801694273949, + 0.009252214804291725, + 0.004888125229626894, + 0.003460070351138711, + 0.0028880585450679064, + 0.006406588014215231, + 0.003470389638096094, + 0.004235453903675079, + 0.005958895664662123, + 0.0074075618758797646, + 0.008268355391919613, + 0.004953214433044195, + 0.005523460917174816, + 0.004901379346847534, + 0.00546026136726141, + 0.0042305560782551765, + 0.009939247742295265, + 0.005612652283161879, + 0.011644240468740463, + 0.01162702776491642, + 0.011882769875228405, + 0.012651626020669937, + 0.012660949490964413, + 0.015239777974784374, + 0.01078892219811678, + 0.016851624473929405, + 0.010747578926384449, + 0.007280096877366304, + 0.01566566526889801, + 0.008393598720431328, + 0.009170594625175, + 0.00962397363036871, + 0.006710236892104149, + 0.008620215579867363, + 0.009779375977814198, + 0.008535469882190228, + 0.009042536839842796, + 0.01037609577178955, + 0.012178209610283375, + 0.0077095902524888515, + 0.010107983835041523, + 0.013224177993834019, + 0.008792036212980747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0620541125535965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806073755025864 + ], + "opposite_logprob_tokens": [ + -3.529806137084961 + ], + "logprob_reward": -0.029806073755025864, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806073755025864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2757651209831238, + 0.05602454021573067, + 0.11389312893152237, + 0.05465998128056526, + 0.2422303855419159, + 0.10536403208971024, + 0.04463603347539902, + 0.0403679758310318, + 0.019871722906827927, + 0.04718710482120514 + ], + "credit_map": [ + 0.01321641355752945, + 0.016123328357934952, + 0.018796531483530998, + 0.06058485433459282, + 0.03281392157077789, + 0.020831065252423286, + 0.03516734018921852, + 0.07823164761066437, + 0.007391379214823246, + 0.004657514858990908, + 0.009121385402977467, + 0.005592905916273594, + 0.0067108552902936935, + 0.00499681243672967, + 0.008322212845087051, + 0.00923147238790989, + 0.014986344613134861, + 0.015185734257102013, + 0.017537672072649002, + 0.016055429354310036, + 0.010327125899493694, + 0.011177407577633858, + 0.01289072260260582, + 0.01573268696665764, + 0.0039841909892857075, + 0.006707090884447098, + 0.008573037572205067, + 0.010373735800385475, + 0.004778602160513401, + 0.005408639553934336, + 0.008237200789153576, + 0.006597479805350304, + 0.01985972747206688, + 0.02269628830254078, + 0.016786083579063416, + 0.03506417199969292, + 0.02474595233798027, + 0.038465630263090134, + 0.053316496312618256, + 0.031296029686927795, + 0.008037103340029716, + 0.007702474948018789, + 0.015452831983566284, + 0.011279817670583725, + 0.01108177937567234, + 0.015548179857432842, + 0.016316751018166542, + 0.019945094361901283, + 0.003706764429807663, + 0.004111740738153458, + 0.00521930493414402, + 0.006047962233424187, + 0.0038494206964969635, + 0.008520510978996754, + 0.005747563671320677, + 0.007432766258716583, + 0.007187962997704744, + 0.006225643679499626, + 0.0037840118166059256, + 0.004071020986884832, + 0.005610097665339708, + 0.005137424450367689, + 0.004584101960062981, + 0.003767710877582431, + 0.0024735068436712027, + 0.002882214728742838, + 0.002577965147793293, + 0.002580675296485424, + 0.0019110784633085132, + 0.0018650825368240476, + 0.0019067834364250302, + 0.0036744147073477507, + 0.004598075523972511, + 0.005169875454157591, + 0.007271016947925091, + 0.005558183416724205, + 0.00548999710008502, + 0.008453701622784138, + 0.0061286333948373795, + 0.004517624620348215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.029806073755025864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816402819007635 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816402819007635, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816402819007635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11502309143543243, + 0.061321552842855453, + 0.11818739771842957, + 0.33679965138435364, + 0.07726132124662399, + 0.054494623094797134, + 0.09196396172046661, + 0.04636768251657486, + 0.06015908718109131, + 0.038421619683504105 + ], + "credit_map": [ + 0.007570710964500904, + 0.01287218276411295, + 0.010916076600551605, + 0.025813955813646317, + 0.011812027543783188, + 0.005684845615178347, + 0.018578415736556053, + 0.021774878725409508, + 0.005261666141450405, + 0.005552668124437332, + 0.00880546122789383, + 0.009302197955548763, + 0.00472656125202775, + 0.0036605389323085546, + 0.01263411808758974, + 0.011378346011042595, + 0.022585976868867874, + 0.014665922150015831, + 0.018496990203857422, + 0.011142722330987453, + 0.011190163902938366, + 0.01719478890299797, + 0.013309961184859276, + 0.009600864723324776, + 0.032018743455410004, + 0.03027573972940445, + 0.06739296019077301, + 0.04022975265979767, + 0.02624249830842018, + 0.033906903117895126, + 0.046961888670921326, + 0.05977117642760277, + 0.006284442264586687, + 0.006754410918802023, + 0.007531016133725643, + 0.01474535372108221, + 0.007281789556145668, + 0.011196237057447433, + 0.014838023111224174, + 0.008630051277577877, + 0.004157783463597298, + 0.004205742850899696, + 0.0053169261664152145, + 0.008613809011876583, + 0.0050113932229578495, + 0.007825109176337719, + 0.012624162249267101, + 0.006739697884768248, + 0.007487154565751553, + 0.009944339282810688, + 0.014339324086904526, + 0.014118662104010582, + 0.00569361075758934, + 0.011755816638469696, + 0.016061056405305862, + 0.012563999742269516, + 0.006155559327453375, + 0.006901118438690901, + 0.00637377705425024, + 0.006887566763907671, + 0.00547745730727911, + 0.005909468047320843, + 0.0040926383808255196, + 0.004570099990814924, + 0.005456086248159409, + 0.013475734740495682, + 0.006886710878461599, + 0.007963230833411217, + 0.006089254282414913, + 0.00578264007344842, + 0.005960613023489714, + 0.008544817566871643, + 0.003951591439545155, + 0.00354898814111948, + 0.003627185709774494, + 0.006451592780649662, + 0.0038125452119857073, + 0.00400400860235095, + 0.006624679546803236, + 0.00640102569013834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.002816402819007635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle drone tracking shot, the camera practically touching the murky green river surface, moves alongside a massive traditional dragon boat, rapidly advancing. The boat's hull is painted a vibrant red and gold, its sides adorned with dragon scales that gleam in the sunlight. A majestic dragon head, eyes glaring, is intricately carved at the prow. The boat is filled with muscular rowers, uniformly dressed in white vests. They sway back and forth in unison to the powerful beats of the bow drummer, their oars thrusting into the water with astonishing unison. They thrust their paddles back and forth, pushing the boat through the narrow waterway, leaving a trail of surging whitewater in their wake. Lush subtropical vegetation lines the riverbanks, the dense canopy nearly obscuring the sky and adding a sense of oppression to the scene. With every sprint of the dragon boat, the huge splashes of water stirred up by the oars splashed violently onto the drone's lens, instantly blurring the field of vision. The water droplets then quickly slid down, leaving dynamic water marks on the screen. The entire picture, illuminated by the bright afternoon sunlight, was full of high-contrast light and shadow and saturated colors, conveying a raw, explosive energy and a tense and intense competitive atmosphere, full of unparalleled sense of power.", + "seed": 14230, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_3807_seed_14230_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021870619151741266, + "vlm_logprob": -0.021870619151741266, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.021870619151741266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06754126399755478 + ], + "opposite_logprob_tokens": [ + -3.3175413608551025 + ], + "logprob_reward": -0.06754126399755478, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06754126399755478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14096246659755707, + 0.052689630538225174, + 0.0706048309803009, + 0.048658858984708786, + 0.1457449197769165, + 0.04481128230690956, + 0.036956537514925, + 0.09462642669677734, + 0.319562166929245, + 0.045382898300886154 + ], + "credit_map": [ + 0.026097266003489494, + 0.01655544713139534, + 0.03943685069680214, + 0.012771951034665108, + 0.01584332250058651, + 0.009058192372322083, + 0.009919466450810432, + 0.011279965750873089, + 0.008443409577012062, + 0.01110901404172182, + 0.005088323261588812, + 0.0036287251859903336, + 0.008266677148640156, + 0.007182634435594082, + 0.005765183363109827, + 0.0032056651543825865, + 0.008757558651268482, + 0.008517774753272533, + 0.010180631652474403, + 0.007559525780379772, + 0.011687370017170906, + 0.00792977586388588, + 0.009254667907953262, + 0.0067175282165408134, + 0.003978133667260408, + 0.005033506546169519, + 0.00486811064183712, + 0.003796423552557826, + 0.006686782464385033, + 0.014555606991052628, + 0.004737484734505415, + 0.005002810154110193, + 0.019466184079647064, + 0.011272363364696503, + 0.018309075385332108, + 0.01891440898180008, + 0.017915690317749977, + 0.011052901856601238, + 0.021262502297759056, + 0.02755177952349186, + 0.004482177086174488, + 0.0072983186691999435, + 0.006524206604808569, + 0.004703307989984751, + 0.004925592802464962, + 0.0067275953479111195, + 0.006117038894444704, + 0.004033044446259737, + 0.0031988692935556173, + 0.004317667800933123, + 0.0051795500330626965, + 0.006208639591932297, + 0.002743747318163514, + 0.004316863603889942, + 0.005838354118168354, + 0.005152846686542034, + 0.005878947209566832, + 0.006678070407360792, + 0.012841560877859592, + 0.01988772675395012, + 0.0061263833194971085, + 0.005834035109728575, + 0.02472659945487976, + 0.012653102166950703, + 0.014627174474298954, + 0.028080914169549942, + 0.08518676459789276, + 0.10040154308080673, + 0.01671012118458748, + 0.015286039561033249, + 0.030189523473381996, + 0.02908010594546795, + 0.004510505590587854, + 0.0035426421090960503, + 0.007233289070427418, + 0.005810892675071955, + 0.0030368545558303595, + 0.0035846340470016003, + 0.008194070309400558, + 0.009470013901591301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.06754126399755478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747792940586805 + ], + "opposite_logprob_tokens": [ + -5.006747722625732 + ], + "logprob_reward": -0.006747792940586805, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747792940586805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10284271836280823, + 0.07782119512557983, + 0.09552987664937973, + 0.09272558242082596, + 0.10285613685846329, + 0.02068580873310566, + 0.02136804908514023, + 0.07182364165782928, + 0.38469505310058594, + 0.029651960358023643 + ], + "credit_map": [ + 0.017247892916202545, + 0.010187608189880848, + 0.013350130058825016, + 0.00974786002188921, + 0.014684611931443214, + 0.008168571628630161, + 0.019190233200788498, + 0.010265814140439034, + 0.012345452792942524, + 0.01103927381336689, + 0.006302356254309416, + 0.005835618358105421, + 0.017264559864997864, + 0.011226366274058819, + 0.00836450606584549, + 0.005443062167614698, + 0.012839251197874546, + 0.011495350860059261, + 0.013910462148487568, + 0.01143979374319315, + 0.01304384134709835, + 0.009911755099892616, + 0.013692930340766907, + 0.009196487255394459, + 0.00686484994366765, + 0.009415815584361553, + 0.01304477546364069, + 0.009754778817296028, + 0.013168912380933762, + 0.01476172637194395, + 0.013520115055143833, + 0.012194612063467503, + 0.016610248014330864, + 0.009730798192322254, + 0.014449560083448887, + 0.01461528055369854, + 0.011603213846683502, + 0.007278474047780037, + 0.01399071142077446, + 0.014577852562069893, + 0.0026576672680675983, + 0.0035202850122004747, + 0.0023044724948704243, + 0.002227859105914831, + 0.0018429794581606984, + 0.002475976012647152, + 0.0023919101804494858, + 0.0032646588515490294, + 0.001405459363013506, + 0.0022518420591950417, + 0.0035385461524128914, + 0.0028152044396847486, + 0.001837567426264286, + 0.002466522855684161, + 0.003561361925676465, + 0.0034915453288704157, + 0.004401260521262884, + 0.009091106243431568, + 0.010283357463777065, + 0.014499853365123272, + 0.006088285706937313, + 0.005968267098069191, + 0.012034354731440544, + 0.009457156993448734, + 0.019702209159731865, + 0.03174061328172684, + 0.10703196376562119, + 0.11865634471178055, + 0.017423437908291817, + 0.019604971632361412, + 0.03551871329545975, + 0.03501678630709648, + 0.0024268601555377245, + 0.0031605411786586046, + 0.003090634709224105, + 0.0028130514547228813, + 0.004139769822359085, + 0.004290272016078234, + 0.004680593963712454, + 0.005050236359238625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.006747792940586805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202821999788284 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202821999788284, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202821999788284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09538646042346954, + 0.07097383588552475, + 0.048436492681503296, + 0.21230506896972656, + 0.13816216588020325, + 0.02366776205599308, + 0.01765063963830471, + 0.1294698864221573, + 0.19433464109897614, + 0.0696130320429802 + ], + "credit_map": [ + 0.01588325761258602, + 0.009269095957279205, + 0.008374875411391258, + 0.006777749862521887, + 0.015580661594867706, + 0.020525790750980377, + 0.008582218550145626, + 0.0103928092867136, + 0.012116181664168835, + 0.016867585480213165, + 0.005798483733087778, + 0.005471596494317055, + 0.01206823531538248, + 0.007034111302345991, + 0.007526185363531113, + 0.004091460257768631, + 0.006347599904984236, + 0.00493317237123847, + 0.006781234871596098, + 0.004845964256674051, + 0.005145614501088858, + 0.006048144772648811, + 0.008982435800135136, + 0.005352322477847338, + 0.014803312718868256, + 0.019518110901117325, + 0.025958837941288948, + 0.022683676332235336, + 0.0412018820643425, + 0.027104580774903297, + 0.027180150151252747, + 0.03385450690984726, + 0.017595583572983742, + 0.011210381053388119, + 0.016499115154147148, + 0.016844283789396286, + 0.019879721105098724, + 0.013331408612430096, + 0.01899772137403488, + 0.023803962394595146, + 0.0024301887024194, + 0.0030019839759916067, + 0.002483358373865485, + 0.002205626107752323, + 0.0034244232811033726, + 0.005029191728681326, + 0.002330612624064088, + 0.002762377727776766, + 0.0012329522287473083, + 0.0011295790318399668, + 0.00297274230979383, + 0.0028718796093016863, + 0.0012801425764337182, + 0.00164306303486228, + 0.0036319722421467304, + 0.002888308372348547, + 0.010854463092982769, + 0.019228743389248848, + 0.019167602062225342, + 0.022881686687469482, + 0.01129187922924757, + 0.01400319579988718, + 0.01602698490023613, + 0.01601533219218254, + 0.01167175080627203, + 0.01188761554658413, + 0.05989157781004906, + 0.06338620185852051, + 0.007816904224455357, + 0.010712084360420704, + 0.012236708775162697, + 0.016731802374124527, + 0.006403823383152485, + 0.009641030803322792, + 0.009493459947407246, + 0.008342538960278034, + 0.00926386471837759, + 0.009180470369756222, + 0.007849149405956268, + 0.009438691660761833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.014202821999788284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081107892096043 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081107892096043, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081107892096043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14840100705623627, + 0.07951618731021881, + 0.12982212007045746, + 0.05495747923851013, + 0.15071488916873932, + 0.07129960507154465, + 0.0557505302131176, + 0.061286672949790955, + 0.2029566913843155, + 0.04529479146003723 + ], + "credit_map": [ + 0.02543175034224987, + 0.019693784415721893, + 0.030719004571437836, + 0.014658045955002308, + 0.017929114401340485, + 0.010809928178787231, + 0.011995829641819, + 0.017163556069135666, + 0.012995067052543163, + 0.014102240093052387, + 0.008832874707877636, + 0.010164772160351276, + 0.00858326070010662, + 0.009791601449251175, + 0.008112152107059956, + 0.006934225559234619, + 0.019089987501502037, + 0.017140235751867294, + 0.025657903403043747, + 0.01828102581202984, + 0.010928824543952942, + 0.012359553016722202, + 0.012489231303334236, + 0.013875358738005161, + 0.0048383306711912155, + 0.005503370426595211, + 0.005498138722032309, + 0.010056106373667717, + 0.009304233826696873, + 0.007885275408625603, + 0.007016843184828758, + 0.004855175968259573, + 0.02009575255215168, + 0.011583322659134865, + 0.015245312824845314, + 0.014375816099345684, + 0.026113413274288177, + 0.016579613089561462, + 0.023429179564118385, + 0.023292481899261475, + 0.00703995767980814, + 0.00896302331238985, + 0.0071534039452672005, + 0.007243738509714603, + 0.009423499926924706, + 0.015530627220869064, + 0.008139737881720066, + 0.007805619388818741, + 0.0058757951483130455, + 0.005643180571496487, + 0.008937695994973183, + 0.008777155540883541, + 0.004446277394890785, + 0.0078012216836214066, + 0.006656292360275984, + 0.0076129077933728695, + 0.003370011458173394, + 0.0041379379108548164, + 0.008649432100355625, + 0.011783810332417488, + 0.0046294080093503, + 0.0032421669457107782, + 0.017452344298362732, + 0.008021561428904533, + 0.010071635246276855, + 0.014328979887068272, + 0.06407850235700607, + 0.06729234009981155, + 0.008995642885565758, + 0.009561539627611637, + 0.013228239491581917, + 0.015399829484522343, + 0.004599663428962231, + 0.0072158267721533775, + 0.0050079296343028545, + 0.004230463411659002, + 0.008158305659890175, + 0.007185614202171564, + 0.004530373960733414, + 0.004366609733551741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.011081107892096043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030904926359653473, + 0.027045831084251404, + 0.05198851227760315, + 0.0550423301756382, + 0.30037781596183777, + 0.018120767548680305, + 0.02802235446870327, + 0.0618472695350647, + 0.3903403580188751, + 0.0363098680973053 + ], + "credit_map": [ + 0.0053006624802947044, + 0.004519971087574959, + 0.005077209789305925, + 0.0027022655121982098, + 0.0035351819824427366, + 0.002634999807924032, + 0.0038847248069941998, + 0.00324990670196712, + 0.005315511487424374, + 0.0032024963293224573, + 0.0032302194740623236, + 0.0024766679853200912, + 0.004732741974294186, + 0.002745560137555003, + 0.002714145928621292, + 0.0026284849736839533, + 0.007934832945466042, + 0.00582208251580596, + 0.0070737143978476524, + 0.004684965591877699, + 0.007948044687509537, + 0.006460846401751041, + 0.006582805421203375, + 0.005481213331222534, + 0.005209988448768854, + 0.006395021919161081, + 0.007166965864598751, + 0.007838383316993713, + 0.005645371042191982, + 0.00963581819087267, + 0.00751976715400815, + 0.005631012376397848, + 0.0534345768392086, + 0.02853817492723465, + 0.0531897209584713, + 0.04864345118403435, + 0.020436860620975494, + 0.013294806703925133, + 0.043293241411447525, + 0.03954695165157318, + 0.0016467204550281167, + 0.0019156006164848804, + 0.001923225587233901, + 0.002457416383549571, + 0.0019170378800481558, + 0.0027838111855089664, + 0.00321285636164248, + 0.002264097798615694, + 0.002599346684291959, + 0.002709904219955206, + 0.0038529622834175825, + 0.004407077096402645, + 0.0019697705283761024, + 0.002756480360403657, + 0.0045520346611738205, + 0.005174776539206505, + 0.007996470667421818, + 0.008550419472157955, + 0.009256797842681408, + 0.009771434590220451, + 0.0036197274457663298, + 0.004262819420546293, + 0.007135218940675259, + 0.011254372075200081, + 0.020731333643198013, + 0.0335068553686142, + 0.10871069878339767, + 0.1197676807641983, + 0.01673794724047184, + 0.0202212855219841, + 0.03607180342078209, + 0.03459273651242256, + 0.0031662732362747192, + 0.003667023964226246, + 0.004399674944579601, + 0.00482593197375536, + 0.0034487275406718254, + 0.0025319375563412905, + 0.0074746995232999325, + 0.006795598194003105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.00978010892868042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle drone tracking shot, the camera practically touching the murky green river surface, moves alongside a massive traditional dragon boat, rapidly advancing. The boat's hull is painted a vibrant red and gold, its sides adorned with dragon scales that gleam in the sunlight. A majestic dragon head, eyes glaring, is intricately carved at the prow. The boat is filled with muscular rowers, uniformly dressed in white vests. They sway back and forth in unison to the powerful beats of the bow drummer, their oars thrusting into the water with astonishing unison. They thrust their paddles back and forth, pushing the boat through the narrow waterway, leaving a trail of surging whitewater in their wake. Lush subtropical vegetation lines the riverbanks, the dense canopy nearly obscuring the sky and adding a sense of oppression to the scene. With every sprint of the dragon boat, the huge splashes of water stirred up by the oars splashed violently onto the drone's lens, instantly blurring the field of vision. The water droplets then quickly slid down, leaving dynamic water marks on the screen. The entire picture, illuminated by the bright afternoon sunlight, was full of high-contrast light and shadow and saturated colors, conveying a raw, explosive energy and a tense and intense competitive atmosphere, full of unparalleled sense of power.", + "seed": 14231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_3807_seed_14231_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.036699846666306254, + "vlm_logprob": -0.036699846666306254, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.036699846666306254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09529058635234833 + ], + "opposite_logprob_tokens": [ + -2.8452906608581543 + ], + "logprob_reward": -0.09529058635234833, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09529058635234833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09692375361919403, + 0.1930476874113083, + 0.027725376188755035, + 0.11380454897880554, + 0.05959276854991913, + 0.17115753889083862, + 0.07034523040056229, + 0.05130275711417198, + 0.05422188341617584, + 0.16187846660614014 + ], + "credit_map": [ + 0.012710199691355228, + 0.013946865685284138, + 0.012869857251644135, + 0.009419640526175499, + 0.014119841158390045, + 0.011888033710420132, + 0.01244206540286541, + 0.009527243673801422, + 0.008714861236512661, + 0.009308134205639362, + 0.023350266739726067, + 0.012597527354955673, + 0.01259545236825943, + 0.05432078614830971, + 0.04682615399360657, + 0.025334523990750313, + 0.002843905007466674, + 0.002886376576498151, + 0.0035463201347738504, + 0.0038209869526326656, + 0.003907199949026108, + 0.00441677076742053, + 0.003254261566326022, + 0.0030495538376271725, + 0.010143237188458443, + 0.011323758400976658, + 0.017955521121621132, + 0.015392868779599667, + 0.014357475563883781, + 0.02388126589357853, + 0.011248542927205563, + 0.009501873515546322, + 0.009073570370674133, + 0.005701512563973665, + 0.007035147864371538, + 0.006311375182121992, + 0.0048065995797514915, + 0.012111500836908817, + 0.008085636422038078, + 0.006467427592724562, + 0.015657547861337662, + 0.023602139204740524, + 0.01467811781913042, + 0.014174344949424267, + 0.032285694032907486, + 0.022279206663370132, + 0.021400999277830124, + 0.027079487219452858, + 0.007600152399390936, + 0.006950805429369211, + 0.012948197312653065, + 0.0077454145066440105, + 0.006806168705224991, + 0.007073821034282446, + 0.012311351485550404, + 0.00890931487083435, + 0.005056176800280809, + 0.005422554444521666, + 0.008026919327676296, + 0.007649831473827362, + 0.0083908187225461, + 0.004543766379356384, + 0.007265258114784956, + 0.004947434179484844, + 0.0067144096828997135, + 0.007059402763843536, + 0.004991384223103523, + 0.006932456977665424, + 0.006782644893974066, + 0.00804857723414898, + 0.006628883071243763, + 0.007064126431941986, + 0.013087310828268528, + 0.036304570734500885, + 0.0060353949666023254, + 0.007847968488931656, + 0.011344230733811855, + 0.005348930601030588, + 0.06223857030272484, + 0.0196714848279953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.09529058635234833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214222319424152 + ], + "opposite_logprob_tokens": [ + -4.264214038848877 + ], + "logprob_reward": -0.014214222319424152, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214222319424152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06850361078977585, + 0.21655169129371643, + 0.06760959327220917, + 0.22487924993038177, + 0.0599137544631958, + 0.03727507218718529, + 0.044266823679208755, + 0.03002738580107689, + 0.02849690243601799, + 0.22247588634490967 + ], + "credit_map": [ + 0.006779167801141739, + 0.00696736853569746, + 0.00763506256043911, + 0.009087366051971912, + 0.0077512930147349834, + 0.009665335528552532, + 0.010330411605536938, + 0.010287608951330185, + 0.016214480623602867, + 0.014631161466240883, + 0.03407807648181915, + 0.014197529293596745, + 0.01835714280605316, + 0.06703168898820877, + 0.023084748536348343, + 0.028956882655620575, + 0.005910477600991726, + 0.005515610799193382, + 0.009757407009601593, + 0.01152012962847948, + 0.007241200190037489, + 0.007175448350608349, + 0.009026255458593369, + 0.011463061906397343, + 0.023650387302041054, + 0.015251015312969685, + 0.03825380280613899, + 0.04543577507138252, + 0.022334560751914978, + 0.03527113422751427, + 0.022208692505955696, + 0.022473875433206558, + 0.009254205040633678, + 0.006687360815703869, + 0.008403164334595203, + 0.009774290025234222, + 0.004088444635272026, + 0.009488009847700596, + 0.006206896156072617, + 0.006011385470628738, + 0.003732060780748725, + 0.004783304873853922, + 0.003005067817866802, + 0.0025984765961766243, + 0.009913172572851181, + 0.004480708856135607, + 0.004232191015034914, + 0.004530090373009443, + 0.004040301311761141, + 0.005441311746835709, + 0.006360620725899935, + 0.00449470616877079, + 0.005509277805685997, + 0.006485950667411089, + 0.006622320972383022, + 0.005312334280461073, + 0.0021732982713729143, + 0.002839796943590045, + 0.005250788293778896, + 0.004612770862877369, + 0.004299539607018232, + 0.002703328849747777, + 0.004460398107767105, + 0.0036874664947390556, + 0.005622164811939001, + 0.002876145066693425, + 0.0034647402353584766, + 0.0032584539148956537, + 0.0034334235824644566, + 0.002486530691385269, + 0.004137543495744467, + 0.0032178987748920918, + 0.018970467150211334, + 0.05139337107539177, + 0.008420242927968502, + 0.010537656955420971, + 0.01819404773414135, + 0.011235374957323074, + 0.0825483575463295, + 0.021176369860768318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.014214222319424152 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329576015472412 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.02329576015472412, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329576015472412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322917491197586, + 0.262030690908432, + 0.02122151106595993, + 0.045973651111125946, + 0.011758322827517986, + 0.1191837340593338, + 0.018641043454408646, + 0.06635581701993942, + 0.022779198363423347, + 0.2997643053531647 + ], + "credit_map": [ + 0.012063165195286274, + 0.007345149293541908, + 0.008967743255198002, + 0.011605893261730671, + 0.008230600506067276, + 0.022953525185585022, + 0.02523636631667614, + 0.03588930889964104, + 0.022139599546790123, + 0.019283119589090347, + 0.033405303955078125, + 0.01824784465134144, + 0.02150030806660652, + 0.04631400108337402, + 0.04919471591711044, + 0.05194578692317009, + 0.0014211866073310375, + 0.0016595993656665087, + 0.002881534630432725, + 0.003232348943129182, + 0.0015120359603315592, + 0.00248113926500082, + 0.0043300436809659, + 0.0037036240100860596, + 0.00466510932892561, + 0.004601125605404377, + 0.006266801152378321, + 0.005121659487485886, + 0.00607677549123764, + 0.010188599117100239, + 0.004998844116926193, + 0.0040547410026192665, + 0.0017306423978880048, + 0.0015431197825819254, + 0.0020880529191344976, + 0.001535264658741653, + 0.0010986934648826718, + 0.0012449499918147922, + 0.0011130187194794416, + 0.0014045810094103217, + 0.009587671607732773, + 0.015118747018277645, + 0.011237173341214657, + 0.00838079396635294, + 0.02796984650194645, + 0.014198703691363335, + 0.01480797864496708, + 0.01788281835615635, + 0.001530767185613513, + 0.001946563832461834, + 0.0028918155003339052, + 0.0027526149060577154, + 0.002558404114097357, + 0.002057852456346154, + 0.0027438851539045572, + 0.002159140771254897, + 0.004768658429384232, + 0.005974181462079287, + 0.013301558792591095, + 0.009935216046869755, + 0.008862772025167942, + 0.006149883382022381, + 0.009984067641198635, + 0.007379480171948671, + 0.0028123341035097837, + 0.0021561773028224707, + 0.0024852785281836987, + 0.002535030711442232, + 0.002462048316374421, + 0.0029053420294076204, + 0.0043672532774508, + 0.0030557329300791025, + 0.024092694744467735, + 0.05109878256917, + 0.011011817492544651, + 0.018350141122937202, + 0.027536068111658096, + 0.022928299382328987, + 0.10537311434745789, + 0.039373382925987244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.02329576015472412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043071962893009186 + ], + "opposite_logprob_tokens": [ + -3.168071985244751 + ], + "logprob_reward": -0.043071962893009186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043071962893009186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2675281763076782, + 0.13119608163833618, + 0.07415372878313065, + 0.10067442804574966, + 0.08160736411809921, + 0.11161261051893234, + 0.02518164925277233, + 0.054332759231328964, + 0.02423710562288761, + 0.12947607040405273 + ], + "credit_map": [ + 0.04073180630803108, + 0.026459690183401108, + 0.03250934183597565, + 0.0218062661588192, + 0.027636108919978142, + 0.05355287715792656, + 0.023664765059947968, + 0.04116734117269516, + 0.01068862434476614, + 0.010848985984921455, + 0.014026283286511898, + 0.014470125548541546, + 0.008490191772580147, + 0.009825781919062138, + 0.03271512687206268, + 0.03013097681105137, + 0.00708493497222662, + 0.0049417451955378056, + 0.006921011954545975, + 0.00674734590575099, + 0.005016806535422802, + 0.0074635716155171394, + 0.017110634595155716, + 0.01886768266558647, + 0.008157859556376934, + 0.009807280264794827, + 0.013980844058096409, + 0.013976098969578743, + 0.011753364466130733, + 0.015479903668165207, + 0.014607880264520645, + 0.012911207042634487, + 0.009564687497913837, + 0.009962807409465313, + 0.010377698577940464, + 0.013993697240948677, + 0.008375831879675388, + 0.012825663201510906, + 0.009555555880069733, + 0.006951430346816778, + 0.013263409957289696, + 0.014395426027476788, + 0.012846414931118488, + 0.009402326308190823, + 0.016312431544065475, + 0.014499732293188572, + 0.013535802252590656, + 0.01735706813633442, + 0.003562071593478322, + 0.0025595214683562517, + 0.0022677264641970396, + 0.0023652345407754183, + 0.0025811681989580393, + 0.003147278679534793, + 0.004375787917524576, + 0.0043228622525930405, + 0.0052731637842953205, + 0.007106120232492685, + 0.006776368711143732, + 0.006968734320253134, + 0.0053797499276697636, + 0.0035912073217332363, + 0.010705922730267048, + 0.008531498722732067, + 0.004032589495182037, + 0.0032457001507282257, + 0.0019105856772512197, + 0.002547117182984948, + 0.0037153870798647404, + 0.0036950111389160156, + 0.003055403009057045, + 0.002035314915701747, + 0.010258762165904045, + 0.026257995516061783, + 0.004710665438324213, + 0.006100937258452177, + 0.01037019956856966, + 0.00742634292691946, + 0.04833729565143585, + 0.01601388305425644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.043071962893009186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626701612025499 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626701612025499, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626701612025499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13788719475269318, + 0.25940021872520447, + 0.09223955869674683, + 0.15948256850242615, + 0.11252249032258987, + 0.031003760173916817, + 0.047314319759607315, + 0.08877982199192047, + 0.03621339425444603, + 0.03515665605664253 + ], + "credit_map": [ + 0.018735947087407112, + 0.01411457173526287, + 0.01730271801352501, + 0.02051001414656639, + 0.012522931210696697, + 0.013646794483065605, + 0.019469337537884712, + 0.021584881469607353, + 0.01606806181371212, + 0.013173560611903667, + 0.02486041933298111, + 0.015717756003141403, + 0.01663847081363201, + 0.02672390826046467, + 0.07548678666353226, + 0.0707312822341919, + 0.007243600208312273, + 0.005818639881908894, + 0.006197994574904442, + 0.008616317063570023, + 0.010985157452523708, + 0.00859461072832346, + 0.02138844132423401, + 0.023394795134663582, + 0.015806695446372032, + 0.01611889712512493, + 0.023660417646169662, + 0.02031071111559868, + 0.023204125463962555, + 0.02962409146130085, + 0.016640661284327507, + 0.0141169847920537, + 0.017921289429068565, + 0.0117460573092103, + 0.016070663928985596, + 0.018414126709103584, + 0.00760718435049057, + 0.012595664709806442, + 0.015541896224021912, + 0.012625613249838352, + 0.0022024614736437798, + 0.003683214308694005, + 0.00344404811039567, + 0.0020266228821128607, + 0.008390042930841446, + 0.003896595910191536, + 0.0035651549696922302, + 0.0037956228479743004, + 0.006549304351210594, + 0.005180789157748222, + 0.005494000390172005, + 0.0049806744791567326, + 0.0043724700808525085, + 0.008362441323697567, + 0.006884161848574877, + 0.005490483716130257, + 0.006745216436684132, + 0.007280097343027592, + 0.022139232605695724, + 0.01081234309822321, + 0.010600948706269264, + 0.0075684902258217335, + 0.012680821120738983, + 0.010952670127153397, + 0.006837093736976385, + 0.003977272659540176, + 0.003847041167318821, + 0.0039236294105648994, + 0.004992909263819456, + 0.0033176455181092024, + 0.004541255533695221, + 0.004776551853865385, + 0.0031950599513947964, + 0.00889891479164362, + 0.0032750538084656, + 0.0029131202027201653, + 0.003795685712248087, + 0.002304601948708296, + 0.004146547522395849, + 0.006627674680203199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.007626701612025499 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle drone tracking shot, the camera practically touching the murky green river surface, moves alongside a massive traditional dragon boat, rapidly advancing. The boat's hull is painted a vibrant red and gold, its sides adorned with dragon scales that gleam in the sunlight. A majestic dragon head, eyes glaring, is intricately carved at the prow. The boat is filled with muscular rowers, uniformly dressed in white vests. They sway back and forth in unison to the powerful beats of the bow drummer, their oars thrusting into the water with astonishing unison. They thrust their paddles back and forth, pushing the boat through the narrow waterway, leaving a trail of surging whitewater in their wake. Lush subtropical vegetation lines the riverbanks, the dense canopy nearly obscuring the sky and adding a sense of oppression to the scene. With every sprint of the dragon boat, the huge splashes of water stirred up by the oars splashed violently onto the drone's lens, instantly blurring the field of vision. The water droplets then quickly slid down, leaving dynamic water marks on the screen. The entire picture, illuminated by the bright afternoon sunlight, was full of high-contrast light and shadow and saturated colors, conveying a raw, explosive energy and a tense and intense competitive atmosphere, full of unparalleled sense of power.", + "seed": 14232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_7/index_3807_seed_14232_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021294608246535063, + "vlm_logprob": -0.021294608246535063, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.021294608246535063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05398107320070267 + ], + "opposite_logprob_tokens": [ + -4.178981304168701 + ], + "logprob_reward": -0.05398107320070267, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.05398107320070267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19351574778556824, + 0.03935696557164192, + 0.05707086622714996, + 0.03712243214249611, + 0.06035009026527405, + 0.0902562141418457, + 0.08910544216632843, + 0.30332863330841064, + 0.08824478089809418, + 0.04164876043796539 + ], + "credit_map": [ + 0.00822747964411974, + 0.007874570786952972, + 0.017781192436814308, + 0.0662059485912323, + 0.005791012663394213, + 0.00952004175633192, + 0.020470893010497093, + 0.05764460563659668, + 0.0027468015905469656, + 0.002155780326575041, + 0.002313112374395132, + 0.009617680683732033, + 0.003687458811327815, + 0.003808779176324606, + 0.0030143330805003643, + 0.012013023719191551, + 0.0024360783863812685, + 0.0028147841803729534, + 0.005642836447805166, + 0.0028002727776765823, + 0.002066887216642499, + 0.002781256102025509, + 0.021948087960481644, + 0.01658066175878048, + 0.0055926586501300335, + 0.0038777044974267483, + 0.004802891053259373, + 0.00433606281876564, + 0.004103145096451044, + 0.0043355198577046394, + 0.005378332454711199, + 0.004696123767644167, + 0.0047560022212564945, + 0.003966732881963253, + 0.004149055574089289, + 0.0072385151870548725, + 0.0036483777221292257, + 0.004282309208065271, + 0.017306897789239883, + 0.01500220037996769, + 0.0032838976476341486, + 0.0023947793524712324, + 0.004467966966331005, + 0.010740529745817184, + 0.0037517547607421875, + 0.005355465225875378, + 0.03295173868536949, + 0.02731008268892765, + 0.00948988925665617, + 0.009341944009065628, + 0.01258771400898695, + 0.01471230573952198, + 0.010579980909824371, + 0.005961991846561432, + 0.015760265290737152, + 0.010671349242329597, + 0.012824560515582561, + 0.00457972614094615, + 0.005314100533723831, + 0.024685878306627274, + 0.010156598873436451, + 0.010743076913058758, + 0.12110275775194168, + 0.11392196267843246, + 0.005236287601292133, + 0.0032325806096196175, + 0.00548715703189373, + 0.005663398653268814, + 0.0038978976663202047, + 0.004868634045124054, + 0.029850192368030548, + 0.03000863827764988, + 0.0034977570176124573, + 0.003551255678758025, + 0.008484510704874992, + 0.007919996045529842, + 0.0042714388109743595, + 0.003464006818830967, + 0.004717317409813404, + 0.005742479581385851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.05398107320070267 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005267668981105089 + ], + "opposite_logprob_tokens": [ + -5.25526762008667 + ], + "logprob_reward": -0.005267668981105089, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005267668981105089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4703880548477173, + 0.059349097311496735, + 0.05288287252187729, + 0.033431313931941986, + 0.015195555984973907, + 0.04067651927471161, + 0.022948604077100754, + 0.18203052878379822, + 0.10007686167955399, + 0.023020632565021515 + ], + "credit_map": [ + 0.012417029589414597, + 0.01749993860721588, + 0.04456225410103798, + 0.16635432839393616, + 0.01513032615184784, + 0.020849481225013733, + 0.06317499279975891, + 0.130399689078331, + 0.00550842983648181, + 0.00404183566570282, + 0.00439323578029871, + 0.009938955307006836, + 0.007833883166313171, + 0.008365385234355927, + 0.005491224583238363, + 0.013776145875453949, + 0.0028351186774671078, + 0.0032428265549242496, + 0.0035563455894589424, + 0.0027243541553616524, + 0.003126209368929267, + 0.00451276870444417, + 0.017577067017555237, + 0.015308180823922157, + 0.0032579959370195866, + 0.0026067185681313276, + 0.005099148023873568, + 0.004583992063999176, + 0.004042200278490782, + 0.005593256093561649, + 0.0035083009861409664, + 0.004739703610539436, + 0.002719509182497859, + 0.0015627160901203752, + 0.002039653016254306, + 0.0016256425296887755, + 0.0015312839532271028, + 0.0016123306704685092, + 0.0025293913204222918, + 0.0015750279417261481, + 0.005147313233464956, + 0.0042968979105353355, + 0.006227725185453892, + 0.004483857657760382, + 0.005229319911450148, + 0.004746528808027506, + 0.00489908317103982, + 0.00564579339697957, + 0.00161256012506783, + 0.0018744419794529676, + 0.0031884508207440376, + 0.004474708344787359, + 0.0020730469841510057, + 0.003252808004617691, + 0.0032719832379370928, + 0.0032006038818508387, + 0.008008401840925217, + 0.0031193201430141926, + 0.0035706274211406708, + 0.015043719671666622, + 0.006166469305753708, + 0.007092619314789772, + 0.07187386602163315, + 0.06715550273656845, + 0.004020963329821825, + 0.0037613895256072283, + 0.003184692468494177, + 0.00649329973384738, + 0.004487302619963884, + 0.005934289190918207, + 0.03620295226573944, + 0.03599197417497635, + 0.0019118533236905932, + 0.0017535685328766704, + 0.004986044019460678, + 0.0041845981031656265, + 0.002338804304599762, + 0.0019861699547618628, + 0.0028609323780983686, + 0.0029986605513840914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.005267668981105089 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787782095372677 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787782095372677, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787782095372677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4174059331417084, + 0.024782776832580566, + 0.04253848269581795, + 0.04744873568415642, + 0.051415275782346725, + 0.14869588613510132, + 0.03469327464699745, + 0.07289667427539825, + 0.1432158350944519, + 0.016907110810279846 + ], + "credit_map": [ + 0.013955380767583847, + 0.018067991361021996, + 0.03980562463402748, + 0.14262567460536957, + 0.013173980638384819, + 0.02196001634001732, + 0.038551829755306244, + 0.12926539778709412, + 0.0033663802314549685, + 0.0018074411200359464, + 0.004212153144180775, + 0.002674280432984233, + 0.0023682783357799053, + 0.0029567002784460783, + 0.0037957916501909494, + 0.0036017487291246653, + 0.002526030410081148, + 0.003824929241091013, + 0.005008351989090443, + 0.002222891664132476, + 0.002560933819040656, + 0.0038039113860577345, + 0.013842186890542507, + 0.008749246597290039, + 0.004109763540327549, + 0.006451422814279795, + 0.007363371085375547, + 0.0070183780044317245, + 0.0050736404955387115, + 0.006785152480006218, + 0.005578356329351664, + 0.0050686500035226345, + 0.0069581735879182816, + 0.007964941672980785, + 0.0061346800066530704, + 0.0067708417773246765, + 0.005539396777749062, + 0.004291011020541191, + 0.008014252409338951, + 0.005741974804550409, + 0.008231306448578835, + 0.005302610341459513, + 0.00957543309777975, + 0.01499385479837656, + 0.008084200322628021, + 0.011413981206715107, + 0.04895410314202309, + 0.04214039817452431, + 0.002399799646809697, + 0.0024573265109211206, + 0.007924288511276245, + 0.007904737256467342, + 0.0026189046911895275, + 0.002338751684874296, + 0.005036975722759962, + 0.0040124934166669846, + 0.004152878187596798, + 0.0032137653324753046, + 0.004643247928470373, + 0.008149683475494385, + 0.002878946717828512, + 0.0036214301362633705, + 0.02400214783847332, + 0.02223457768559456, + 0.006859006360173225, + 0.005895831622183323, + 0.004885424394160509, + 0.007308158557862043, + 0.006607698276638985, + 0.008156424388289452, + 0.05175085365772247, + 0.051752444356679916, + 0.00191914988681674, + 0.0014329096302390099, + 0.00275451154448092, + 0.0024358786176890135, + 0.001927850884385407, + 0.0019587392453104258, + 0.0022740426938980818, + 0.0022040277253836393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.009787782095372677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980804070830345 + ], + "opposite_logprob_tokens": [ + -3.5298080444335938 + ], + "logprob_reward": -0.02980804070830345, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980804070830345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34002968668937683, + 0.11791742593050003, + 0.05158351734280586, + 0.042785245925188065, + 0.0465303398668766, + 0.17914073169231415, + 0.05696393921971321, + 0.08452486246824265, + 0.05106866732239723, + 0.029455585405230522 + ], + "credit_map": [ + 0.015581968240439892, + 0.015493543818593025, + 0.03203998878598213, + 0.11341332644224167, + 0.011835016310214996, + 0.017959341406822205, + 0.03315737470984459, + 0.1005491390824318, + 0.013432526029646397, + 0.009774253703653812, + 0.013791093602776527, + 0.014478392899036407, + 0.008279787376523018, + 0.011806695722043514, + 0.017649579793214798, + 0.028705090284347534, + 0.005446759983897209, + 0.0037802981678396463, + 0.010224498808383942, + 0.0049391151405870914, + 0.0032943799160420895, + 0.005009785760194063, + 0.011175299063324928, + 0.0077133784070611, + 0.0061790174804627895, + 0.0038759491872042418, + 0.005178515333682299, + 0.004671116825193167, + 0.004737246781587601, + 0.004781914409250021, + 0.007177450228482485, + 0.006184036377817392, + 0.004232475534081459, + 0.003775480668991804, + 0.005935526452958584, + 0.004606584087014198, + 0.004864583723247051, + 0.008658355101943016, + 0.006481063552200794, + 0.007976274006068707, + 0.00898802187293768, + 0.006499968469142914, + 0.00951236393302679, + 0.02064388245344162, + 0.00833309069275856, + 0.011884956620633602, + 0.05893569812178612, + 0.05434275418519974, + 0.006025916896760464, + 0.005521579645574093, + 0.006147191394120455, + 0.008448269218206406, + 0.008846922777593136, + 0.006929130759090185, + 0.007218008860945702, + 0.007826914079487324, + 0.0033244939986616373, + 0.0015212907455861568, + 0.002087385393679142, + 0.007602323312312365, + 0.0029244571924209595, + 0.0036115827970206738, + 0.03253670409321785, + 0.030916621908545494, + 0.0023917846847325563, + 0.0016246538143604994, + 0.002496416447684169, + 0.00432236585766077, + 0.002237966051325202, + 0.002698096912354231, + 0.017889464274048805, + 0.017407914623618126, + 0.002845645882189274, + 0.0029328549280762672, + 0.005547827109694481, + 0.005357418209314346, + 0.002636625664308667, + 0.003133227117359638, + 0.003407395211979747, + 0.0035945922136306763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02980804070830345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, practically touching the murky green river surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive traditional dragon boat with a hull painted vibrant red and gold visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are muscular rowers uniformly dressed in white vests rowing in unison?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright afternoon sunlight creating high-contrast light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do water droplets from the oars splash onto the lens, causing blurring and dynamic water marks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628476247191429 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628476247191429, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628476247191429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12139404565095901, + 0.21830891072750092, + 0.03695450350642204, + 0.02538658306002617, + 0.016075221821665764, + 0.05570672079920769, + 0.023825306445360184, + 0.25994765758514404, + 0.21964207291603088, + 0.02275896444916725 + ], + "credit_map": [ + 0.004157960880547762, + 0.00525163346901536, + 0.011118865571916103, + 0.04052354022860527, + 0.005507207475602627, + 0.005857142154127359, + 0.021151825785636902, + 0.02782587520778179, + 0.03652282804250717, + 0.021740077063441277, + 0.02875809371471405, + 0.02843579277396202, + 0.018622733652591705, + 0.02351321280002594, + 0.023611998185515404, + 0.03710417449474335, + 0.002680395497009158, + 0.0025238425005227327, + 0.0031207550782710314, + 0.0019657141529023647, + 0.0015987305669113994, + 0.002050950890406966, + 0.012417006306350231, + 0.010597110725939274, + 0.0020321959163993597, + 0.002432087203487754, + 0.004708338528871536, + 0.005405776668339968, + 0.0017590404022485018, + 0.002173064276576042, + 0.004169244319200516, + 0.0027068378403782845, + 0.0019032190321013331, + 0.0018027541227638721, + 0.0014238350559026003, + 0.0017564318841323256, + 0.0011963126016780734, + 0.0012855675304308534, + 0.0036010537296533585, + 0.0031060471665114164, + 0.005132324993610382, + 0.003691279562190175, + 0.005689598619937897, + 0.005197717808187008, + 0.004792363848537207, + 0.004962888080626726, + 0.01429988257586956, + 0.011940669268369675, + 0.0023350913543254137, + 0.002632709452882409, + 0.002404367784038186, + 0.0031367954798042774, + 0.0044790818355977535, + 0.002470665145665407, + 0.003223562380298972, + 0.0031430323142558336, + 0.011273743584752083, + 0.00403169309720397, + 0.004420491401106119, + 0.02116207405924797, + 0.008755617775022984, + 0.009221687912940979, + 0.1036011129617691, + 0.09748122841119766, + 0.007190563250333071, + 0.006357759237289429, + 0.008755377493798733, + 0.02088376134634018, + 0.008669114671647549, + 0.009726278483867645, + 0.0788218230009079, + 0.07923739403486252, + 0.0016535279573872685, + 0.0019469715189188719, + 0.004575649742037058, + 0.004532965831458569, + 0.0019466480007395148, + 0.002124951221048832, + 0.0027385037392377853, + 0.0032397466711699963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.007628476247191429 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_8/index_tensor([3645, 3645])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_8/index_tensor([3645, 3645])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..79936971ac81136417087bd909f38b34573588c9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_8/index_tensor([3645, 3645])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a static, high-angle, top-down bird's-eye view, the camera gazes upon a vast, unbroken ocean of lush green foliage stretching to a soft horizon. The scene is a vibrant, dense tapestry of foliage, with an infinite variety of sizes and shapes, creating a random, natural pattern with no discernible organization. Rich, complex shades of green dominate the frame, ranging from deep emerald and jade in the shadowy depths to brilliant lime and chartreuse in the bright, clear sunlight, making some leaves appear almost translucent. The light creates a speckled, shimmering effect across the landscape, highlighting the waxy texture of some leaves and the delicate, intricate veining of others. The entire scene unfolds in perfect, serene silence, capturing the profound tranquility and overwhelming richness of nature in a single, breathtaking shot.", + "seed": 9396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_8/index_3645_seed_9396_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09700159849599004, + "vlm_logprob": -0.09700159849599004, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.09700159849599004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025862248614430428 + ], + "opposite_logprob_tokens": [ + -3.900862216949463 + ], + "logprob_reward": -0.025862248614430428, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025862248614430428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15641099214553833, + 0.10512614250183105, + 0.06322643160820007, + 0.06442902982234955, + 0.1961861252784729, + 0.06712574511766434, + 0.06954726576805115, + 0.16792649030685425, + 0.08488325029611588, + 0.02513856254518032 + ], + "credit_map": [ + 0.009092763066291809, + 0.01754549704492092, + 0.01864859275519848, + 0.014895531348884106, + 0.023463174700737, + 0.032403238117694855, + 0.024582477286458015, + 0.01577969640493393, + 0.010093031451106071, + 0.010744307190179825, + 0.016399450600147247, + 0.018938371911644936, + 0.015244513750076294, + 0.010620061308145523, + 0.012540381401777267, + 0.01054601650685072, + 0.004782477859407663, + 0.005753468256443739, + 0.015226299874484539, + 0.009983178228139877, + 0.004512855317443609, + 0.0045007565058767796, + 0.01092537771910429, + 0.007542008068412542, + 0.0051189386285841465, + 0.009011585265398026, + 0.00796380266547203, + 0.011722022667527199, + 0.00567656010389328, + 0.007782295346260071, + 0.0094929039478302, + 0.007660917937755585, + 0.015688052400946617, + 0.027513545006513596, + 0.02694627083837986, + 0.019947180524468422, + 0.015990613028407097, + 0.051579590886831284, + 0.016486866399645805, + 0.02203400619328022, + 0.004833759739995003, + 0.004924739710986614, + 0.005920253228396177, + 0.011743007227778435, + 0.004292478784918785, + 0.0035980138927698135, + 0.018952133134007454, + 0.012861350551247597, + 0.006712721660733223, + 0.014060488902032375, + 0.0078064086847007275, + 0.009906817227602005, + 0.009388192556798458, + 0.007484264671802521, + 0.008009908720850945, + 0.0061784591525793076, + 0.006115226540714502, + 0.006748192012310028, + 0.007121628150343895, + 0.011437898501753807, + 0.013882779516279697, + 0.008300702087581158, + 0.05939318984746933, + 0.05492687225341797, + 0.00460566021502018, + 0.00514525780454278, + 0.005854707211256027, + 0.008554711006581783, + 0.017574457451701164, + 0.005074064712971449, + 0.027882959693670273, + 0.010191427543759346, + 0.002498104702681303, + 0.004168231971561909, + 0.0022293173242360353, + 0.004520413000136614, + 0.002483452670276165, + 0.003748501418158412, + 0.0024032730143517256, + 0.003087266581133008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.025862248614430428 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104880150407553 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.004104880150407553, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104880150407553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23378470540046692, + 0.09111688286066055, + 0.0772336795926094, + 0.07019294053316116, + 0.13056150078773499, + 0.06686341017484665, + 0.11650045961141586, + 0.1509331464767456, + 0.03534708917140961, + 0.027466244995594025 + ], + "credit_map": [ + 0.015325119718909264, + 0.030005844309926033, + 0.043277956545352936, + 0.03246915712952614, + 0.016939587891101837, + 0.03547954559326172, + 0.03205196559429169, + 0.028235487639904022, + 0.007440079469233751, + 0.00943790003657341, + 0.012935636565089226, + 0.016020137816667557, + 0.01252990122884512, + 0.01049079466611147, + 0.010841766372323036, + 0.011420655064284801, + 0.004920999985188246, + 0.00799314770847559, + 0.01707414537668228, + 0.009477703832089901, + 0.009779300540685654, + 0.00639238441362977, + 0.013190720230340958, + 0.008405264467000961, + 0.006280206609517336, + 0.008781291544437408, + 0.006163702346384525, + 0.012395531870424747, + 0.005243939347565174, + 0.010198721662163734, + 0.01151934452354908, + 0.00961019191890955, + 0.01005095336586237, + 0.012446044944226742, + 0.02137075737118721, + 0.017927242442965508, + 0.013732286170125008, + 0.02427872270345688, + 0.01450255885720253, + 0.016252920031547546, + 0.0055463602766394615, + 0.006492663640528917, + 0.010259019210934639, + 0.013568730093538761, + 0.004464755300432444, + 0.005286315456032753, + 0.01441525761038065, + 0.006830298341810703, + 0.009464418515563011, + 0.027231091633439064, + 0.010825702920556068, + 0.02723993919789791, + 0.01192702166736126, + 0.010960147716104984, + 0.01039640698581934, + 0.008455716073513031, + 0.0056096650660037994, + 0.004997959826141596, + 0.006001045927405357, + 0.010427840985357761, + 0.008770826272666454, + 0.006720885168761015, + 0.05672681704163551, + 0.051678091287612915, + 0.0032205989118665457, + 0.004303151275962591, + 0.004559616092592478, + 0.004830806981772184, + 0.0060344175435602665, + 0.003412606194615364, + 0.00441849185153842, + 0.004567395895719528, + 0.002129144500941038, + 0.0023418772034347057, + 0.0030637159943580627, + 0.006289748474955559, + 0.0025603233370929956, + 0.002318798564374447, + 0.004613146185874939, + 0.004149488639086485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.004104880150407553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893946170806885 + ], + "opposite_logprob_tokens": [ + -1.0539394617080688 + ], + "logprob_reward": -0.42893946170806885, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42893946170806885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14342717826366425, + 0.0551174134016037, + 0.07841441035270691, + 0.10828377306461334, + 0.05647023022174835, + 0.08000052720308304, + 0.029723994433879852, + 0.4095546305179596, + 0.015896625816822052, + 0.023111218586564064 + ], + "credit_map": [ + 0.010693063028156757, + 0.018801214173436165, + 0.017892146483063698, + 0.018516026437282562, + 0.023847205564379692, + 0.015762006863951683, + 0.024559631943702698, + 0.013355881907045841, + 0.004491689149290323, + 0.005674058571457863, + 0.009083843789994717, + 0.009701929986476898, + 0.006888143718242645, + 0.00592547282576561, + 0.006565675139427185, + 0.006786597426980734, + 0.006072384305298328, + 0.008300267159938812, + 0.012032845057547092, + 0.008490107022225857, + 0.015957549214363098, + 0.009447050280869007, + 0.011143719777464867, + 0.006970485206693411, + 0.008694144897162914, + 0.015273701399564743, + 0.014077379368245602, + 0.01623472012579441, + 0.013510075397789478, + 0.01559621375054121, + 0.013151275925338268, + 0.01174626499414444, + 0.0038697337731719017, + 0.005905576050281525, + 0.010036666877567768, + 0.007026516832411289, + 0.006645600777119398, + 0.011100061237812042, + 0.005502458196133375, + 0.0063836160115897655, + 0.005824211519211531, + 0.005431225057691336, + 0.012126288376748562, + 0.016993725672364235, + 0.006941125262528658, + 0.005117453169077635, + 0.019405553117394447, + 0.008160950616002083, + 0.002646960783749819, + 0.003726927563548088, + 0.004239024594426155, + 0.0036477050743997097, + 0.0030516658443957567, + 0.004285086877644062, + 0.004369435366243124, + 0.003757188329473138, + 0.014697317034006119, + 0.012934924103319645, + 0.016773933544754982, + 0.02979147993028164, + 0.020960530266165733, + 0.01925896666944027, + 0.15803879499435425, + 0.13709868490695953, + 0.001274602604098618, + 0.002075832337141037, + 0.0016223298152908683, + 0.002023858716711402, + 0.002807418117299676, + 0.0012947841314598918, + 0.0035468190908432007, + 0.001250979956239462, + 0.0016190481837838888, + 0.0025004916824400425, + 0.002710950095206499, + 0.005053134635090828, + 0.0027117806021124125, + 0.002722619567066431, + 0.003725104732438922, + 0.0020680874586105347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.42893946170806885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028198501095175743 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.0028198501095175743, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028198501095175743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16964659094810486, + 0.12413117289543152, + 0.08923327922821045, + 0.10822305083274841, + 0.054868508130311966, + 0.06999611854553223, + 0.0677768662571907, + 0.23154118657112122, + 0.01997755840420723, + 0.06460563093423843 + ], + "credit_map": [ + 0.013078014366328716, + 0.019913900643587112, + 0.035233575850725174, + 0.01924838498234749, + 0.01527449581772089, + 0.021988287568092346, + 0.02471453696489334, + 0.02019539475440979, + 0.008818456903100014, + 0.011490143835544586, + 0.018032701686024666, + 0.02223452553153038, + 0.018442805856466293, + 0.011227592825889587, + 0.016745178028941154, + 0.01713975891470909, + 0.005577401723712683, + 0.007937723770737648, + 0.017352206632494926, + 0.011329781264066696, + 0.016303399577736855, + 0.0095896665006876, + 0.012420530430972576, + 0.008722568862140179, + 0.007489557843655348, + 0.01418384350836277, + 0.016990048810839653, + 0.015365594066679478, + 0.007682344410568476, + 0.015357044525444508, + 0.01800445280969143, + 0.013150166720151901, + 0.005893557798117399, + 0.0061517562717199326, + 0.01123492605984211, + 0.008902515284717083, + 0.004017951898276806, + 0.006974991410970688, + 0.005317429546266794, + 0.006375380791723728, + 0.005817264784127474, + 0.006537133362144232, + 0.010998839512467384, + 0.01303110457956791, + 0.004473655950278044, + 0.005958513356745243, + 0.014482894912362099, + 0.008696713484823704, + 0.005330629646778107, + 0.01780940778553486, + 0.004715802613645792, + 0.0099415248259902, + 0.0067094857804477215, + 0.008094376884400845, + 0.007498348131775856, + 0.007677297107875347, + 0.008979007601737976, + 0.0063636694103479385, + 0.010468472726643085, + 0.016101593151688576, + 0.01249539665877819, + 0.010666396468877792, + 0.08836760371923447, + 0.07809903472661972, + 0.0019864090718328953, + 0.0026847373228520155, + 0.0028001710306853056, + 0.0029664733447134495, + 0.0029579652473330498, + 0.0018841188866645098, + 0.0023972177878022194, + 0.0023004666436463594, + 0.004143645521253347, + 0.006051020231097937, + 0.007617332972586155, + 0.018416067585349083, + 0.006779802963137627, + 0.006267371121793985, + 0.00906622689217329, + 0.006264162249863148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0028198501095175743 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281551897525787 + ], + "opposite_logprob_tokens": [ + -3.7732815742492676 + ], + "logprob_reward": -0.023281551897525787, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281551897525787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09888160973787308, + 0.1689232438802719, + 0.048299551010131836, + 0.0916077196598053, + 0.03925637900829315, + 0.04938792809844017, + 0.023281140252947807, + 0.41472339630126953, + 0.024145841598510742, + 0.04149321839213371 + ], + "credit_map": [ + 0.007297446019947529, + 0.009297557175159454, + 0.014466499909758568, + 0.01204004418104887, + 0.013063815422356129, + 0.014382833614945412, + 0.016542332246899605, + 0.011791071854531765, + 0.009588886052370071, + 0.015481870621442795, + 0.025501525029540062, + 0.02468075044453144, + 0.03541497141122818, + 0.017686206847429276, + 0.02047073282301426, + 0.02009829878807068, + 0.0030619697645306587, + 0.004464657045900822, + 0.009914268739521503, + 0.006365128327161074, + 0.005127218551933765, + 0.004932223353534937, + 0.009068050421774387, + 0.005366031546145678, + 0.006721470970660448, + 0.009808116592466831, + 0.010240888223052025, + 0.012269101105630398, + 0.011340205557644367, + 0.016400234773755074, + 0.01272385660558939, + 0.012103837914764881, + 0.0042773522436618805, + 0.005314284469932318, + 0.007756651844829321, + 0.005025640595704317, + 0.0036881330888718367, + 0.0054401857778429985, + 0.0043005989864468575, + 0.0034535301383584738, + 0.0038798414170742035, + 0.0030082433950155973, + 0.01045189518481493, + 0.011693904176354408, + 0.002953428542241454, + 0.004633719567209482, + 0.006260856986045837, + 0.006506034173071384, + 0.001727550057694316, + 0.004134561866521835, + 0.0025003342889249325, + 0.0031867497600615025, + 0.0026393563020974398, + 0.0024662206415086985, + 0.003613966517150402, + 0.0030123991891741753, + 0.015149563550949097, + 0.01377243734896183, + 0.01498574297875166, + 0.027652239426970482, + 0.04245122894644737, + 0.021660231053829193, + 0.14762084186077118, + 0.13143108785152435, + 0.0013371536042541265, + 0.0021108000073581934, + 0.0015517785213887691, + 0.002176395384594798, + 0.005557413212954998, + 0.001764090615324676, + 0.007635187357664108, + 0.0020130216144025326, + 0.00273072998970747, + 0.0053738076239824295, + 0.00440441258251667, + 0.009249717928469181, + 0.003903106087818742, + 0.005576763302087784, + 0.006083881948143244, + 0.004170794039964676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.023281551897525787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, high-angle, top-down bird's-eye view, the camera gazes upon a vast, unbroken ocean of lush green foliage stretching to a soft horizon. The scene is a vibrant, dense tapestry of foliage, with an infinite variety of sizes and shapes, creating a random, natural pattern with no discernible organization. Rich, complex shades of green dominate the frame, ranging from deep emerald and jade in the shadowy depths to brilliant lime and chartreuse in the bright, clear sunlight, making some leaves appear almost translucent. The light creates a speckled, shimmering effect across the landscape, highlighting the waxy texture of some leaves and the delicate, intricate veining of others. The entire scene unfolds in perfect, serene silence, capturing the profound tranquility and overwhelming richness of nature in a single, breathtaking shot.", + "seed": 9397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_8/index_3645_seed_9397_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05345335749443621, + "vlm_logprob": -0.05345335749443621, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05345335749443621, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01783008873462677 + ], + "opposite_logprob_tokens": [ + -4.267829895019531 + ], + "logprob_reward": -0.01783008873462677, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01783008873462677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27145108580589294, + 0.07238009572029114, + 0.12639372050762177, + 0.14756418764591217, + 0.04418652504682541, + 0.13502344489097595, + 0.051546577364206314, + 0.047994114458560944, + 0.07248491793870926, + 0.030975323170423508 + ], + "credit_map": [ + 0.021707870066165924, + 0.021876709535717964, + 0.051655206829309464, + 0.024198412895202637, + 0.029965439811348915, + 0.07857419550418854, + 0.026912154629826546, + 0.016561102122068405, + 0.007997899316251278, + 0.010860847309231758, + 0.00720532750710845, + 0.008532395586371422, + 0.010947148315608501, + 0.011562242172658443, + 0.009478027001023293, + 0.005796209443360567, + 0.017846066504716873, + 0.016821861267089844, + 0.020504172891378403, + 0.017330067232251167, + 0.011248565278947353, + 0.013614661060273647, + 0.016725867986679077, + 0.012302454560995102, + 0.026616517454385757, + 0.013150324113667011, + 0.02417704649269581, + 0.01703125797212124, + 0.018093734979629517, + 0.017272695899009705, + 0.01867680437862873, + 0.012545800767838955, + 0.006032042670994997, + 0.004415549803525209, + 0.006862820591777563, + 0.003016827628016472, + 0.007220868021249771, + 0.004846886266022921, + 0.006758283823728561, + 0.005033244378864765, + 0.01167344395071268, + 0.004450521431863308, + 0.04391852393746376, + 0.006877637468278408, + 0.010531515814363956, + 0.047326166182756424, + 0.00577592896297574, + 0.004469715058803558, + 0.00793931633234024, + 0.0056081353686749935, + 0.007893039844930172, + 0.0034174341708421707, + 0.009249323979020119, + 0.0069439527578651905, + 0.006354411598294973, + 0.00414096238091588, + 0.004220815375447273, + 0.0058264886029064655, + 0.008022150956094265, + 0.0033492436632514, + 0.010148550383746624, + 0.007040370721369982, + 0.005438297055661678, + 0.003948197700083256, + 0.007684518117457628, + 0.006560963578522205, + 0.011257153004407883, + 0.006532716099172831, + 0.006197819020599127, + 0.012366578914225101, + 0.007523978129029274, + 0.014361189678311348, + 0.002968752756714821, + 0.00409078411757946, + 0.00463144201785326, + 0.003098625224083662, + 0.005325209815055132, + 0.003978986758738756, + 0.003285548184067011, + 0.0035959735978394747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.01783008873462677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082797199487686 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082797199487686, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082797199487686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5314926505088806, + 0.07335460186004639, + 0.05873909965157509, + 0.06578727066516876, + 0.05772501230239868, + 0.04556792601943016, + 0.03960517793893814, + 0.06106990575790405, + 0.029794078320264816, + 0.036864276975393295 + ], + "credit_map": [ + 0.02909156121313572, + 0.018387405201792717, + 0.1573585718870163, + 0.028240734711289406, + 0.04081067815423012, + 0.21573467552661896, + 0.01732511632144451, + 0.02454390935599804, + 0.008958284743130207, + 0.009657800197601318, + 0.008985579945147038, + 0.008598851040005684, + 0.010008375160396099, + 0.01143412385135889, + 0.009395827539265156, + 0.006315763108432293, + 0.008107108063995838, + 0.007584607228636742, + 0.008334453217685223, + 0.006316183600574732, + 0.00867156870663166, + 0.007880953140556812, + 0.006872774567455053, + 0.004971454385668039, + 0.010835163295269012, + 0.006109538953751326, + 0.009807400405406952, + 0.005908644758164883, + 0.009811783209443092, + 0.010431621223688126, + 0.007055905647575855, + 0.0058272103779017925, + 0.008753739297389984, + 0.00629929406568408, + 0.007336004171520472, + 0.004489906132221222, + 0.007706828415393829, + 0.006892035249620676, + 0.007870110683143139, + 0.008377090096473694, + 0.003552883630618453, + 0.0039009032770991325, + 0.011017462238669395, + 0.003037652000784874, + 0.005567405838519335, + 0.012225989252328873, + 0.0036671352572739124, + 0.0025984924286603928, + 0.004340812098234892, + 0.0038100923411548138, + 0.007142013870179653, + 0.003331004874780774, + 0.004978520330041647, + 0.009414952248334885, + 0.003801218466833234, + 0.002786563243716955, + 0.005048253573477268, + 0.0072312443517148495, + 0.011726420372724533, + 0.0034023975022137165, + 0.0052649155259132385, + 0.020886987447738647, + 0.003609569976106286, + 0.003900114679709077, + 0.0024404956493526697, + 0.0017730247927829623, + 0.004133881069719791, + 0.00296209380030632, + 0.003541278187185526, + 0.0032755068968981504, + 0.006955548655241728, + 0.004712248686701059, + 0.003525703214108944, + 0.004237169865518808, + 0.00997683685272932, + 0.002271786332130432, + 0.004030697979032993, + 0.004639125894755125, + 0.0028817818965762854, + 0.005301172845065594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.016082797199487686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161476731300354 + ], + "opposite_logprob_tokens": [ + -1.7016147375106812 + ], + "logprob_reward": -0.20161476731300354, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161476731300354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6829254031181335, + 0.04275005683302879, + 0.017083333805203438, + 0.03328605741262436, + 0.013152736239135265, + 0.10264647006988525, + 0.010396028868854046, + 0.039904192090034485, + 0.032466258853673935, + 0.025389457121491432 + ], + "credit_map": [ + 0.036991655826568604, + 0.01911611296236515, + 0.20240971446037292, + 0.03416775166988373, + 0.05265824496746063, + 0.2847815155982971, + 0.018874166533350945, + 0.03392629697918892, + 0.005011443980038166, + 0.0043784696608781815, + 0.0051949648186564445, + 0.003094784915447235, + 0.007724699564278126, + 0.007741338573396206, + 0.0052353860810399055, + 0.004368969704955816, + 0.00260869855992496, + 0.0021131604444235563, + 0.002050480805337429, + 0.0016112487064674497, + 0.002835831604897976, + 0.0023887779098004103, + 0.001949008321389556, + 0.0015261269873008132, + 0.0029404088854789734, + 0.0040405625477433205, + 0.004039830062538385, + 0.0023992930073291063, + 0.0064224801026284695, + 0.007193222641944885, + 0.0036556150298565626, + 0.0025946449022740126, + 0.0019439479801803827, + 0.0013341581216081977, + 0.0017013121396303177, + 0.0010659464169293642, + 0.0021469276398420334, + 0.0014728876994922757, + 0.002042528474703431, + 0.00144502695184201, + 0.008746003732085228, + 0.0031713347416371107, + 0.03332787752151489, + 0.004826833494007587, + 0.008969930931925774, + 0.03757484629750252, + 0.0031632683239877224, + 0.002866369439288974, + 0.001438598264940083, + 0.0010229295585304499, + 0.0019241117406636477, + 0.0007769408402964473, + 0.001721830340102315, + 0.0015072626993060112, + 0.001179849379695952, + 0.00082450604531914, + 0.002614562166854739, + 0.002479839138686657, + 0.011689203791320324, + 0.0020203152671456337, + 0.0034995852038264275, + 0.013361558318138123, + 0.0023015986662358046, + 0.0019375282572582364, + 0.00346361449919641, + 0.0038144129794090986, + 0.0046562133356928825, + 0.0024019121192395687, + 0.005162008106708527, + 0.004584000911563635, + 0.004201760981231928, + 0.004182336386293173, + 0.0027930892538279295, + 0.002048131078481674, + 0.0066843898966908455, + 0.001472180592827499, + 0.002268440555781126, + 0.0030916703399270773, + 0.0018113243859261274, + 0.0052202302031219006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.20161476731300354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1962200254201889, + 0.06738069653511047, + 0.05179540440440178, + 0.04224533587694168, + 0.12481314688920975, + 0.06877173483371735, + 0.045051027089357376, + 0.05021345987915993, + 0.2871294319629669, + 0.06637980043888092 + ], + "credit_map": [ + 0.015882691368460655, + 0.013650836423039436, + 0.04517161846160889, + 0.016478825360536575, + 0.01707448810338974, + 0.062206994742155075, + 0.012432768940925598, + 0.01332179270684719, + 0.007515530101954937, + 0.0076015666127204895, + 0.00909651443362236, + 0.009778458625078201, + 0.007898088544607162, + 0.008943859487771988, + 0.00874976348131895, + 0.0077969105914235115, + 0.005618119612336159, + 0.0067639718763530254, + 0.007000403478741646, + 0.004721475299447775, + 0.009258726611733437, + 0.007748022675514221, + 0.0060541448183357716, + 0.0046305349096655846, + 0.00449738372117281, + 0.005283619277179241, + 0.007035969290882349, + 0.0031111238058656454, + 0.007869570516049862, + 0.00535549083724618, + 0.004956638906151056, + 0.004135532304644585, + 0.012823035009205341, + 0.014876098372042179, + 0.017702454701066017, + 0.012094453908503056, + 0.01971663162112236, + 0.016077004373073578, + 0.016657255589962006, + 0.014866212382912636, + 0.00567497918382287, + 0.003390699392184615, + 0.018862271681427956, + 0.0036939631681889296, + 0.007284542080014944, + 0.02251793071627617, + 0.0044478196650743484, + 0.0028995180036872625, + 0.00564974918961525, + 0.005189085379242897, + 0.0063706557266414165, + 0.00500124366953969, + 0.006329625379294157, + 0.006478461902588606, + 0.006307527888566256, + 0.0037246737629175186, + 0.004194550681859255, + 0.00417441688477993, + 0.014747748151421547, + 0.002635653130710125, + 0.00576649559661746, + 0.012336624786257744, + 0.003208138979971409, + 0.003149826545268297, + 0.028321051970124245, + 0.02072710543870926, + 0.035693373531103134, + 0.023335788398981094, + 0.024605507031083107, + 0.03931434825062752, + 0.04855603724718094, + 0.06657617539167404, + 0.006685479078441858, + 0.007335460279136896, + 0.02108892984688282, + 0.005030408967286348, + 0.006051558535546064, + 0.007294982206076384, + 0.005209836643189192, + 0.0076831369660794735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.001943010720424354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796123504638672 + ], + "opposite_logprob_tokens": [ + -3.5297961235046387 + ], + "logprob_reward": -0.029796123504638672, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796123504638672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5487077236175537, + 0.0683511421084404, + 0.0366487130522728, + 0.03250174969434738, + 0.05976354330778122, + 0.027476487681269646, + 0.03186941146850586, + 0.06449129432439804, + 0.08104343712329865, + 0.04914650693535805 + ], + "credit_map": [ + 0.03042985312640667, + 0.019973019137978554, + 0.15593817830085754, + 0.030610952526330948, + 0.04316851124167442, + 0.22266621887683868, + 0.018230557441711426, + 0.027690468356013298, + 0.00694121140986681, + 0.008651956915855408, + 0.00898278784006834, + 0.010341349989175797, + 0.007794903591275215, + 0.009954959154129028, + 0.008684922941029072, + 0.006999053526669741, + 0.0038606738671660423, + 0.004355390090495348, + 0.005043513607233763, + 0.0037653809413313866, + 0.00565217062830925, + 0.005821630358695984, + 0.004588488023728132, + 0.0035614653024822474, + 0.003629476297646761, + 0.003456816775724292, + 0.004608625546097755, + 0.0022382845636457205, + 0.006120414473116398, + 0.003608099650591612, + 0.00430524256080389, + 0.004534788895398378, + 0.008567404933273792, + 0.00766467722132802, + 0.00855150818824768, + 0.005545068997889757, + 0.007369498256593943, + 0.007554055657237768, + 0.006757894530892372, + 0.007753435056656599, + 0.003471576841548085, + 0.002383099403232336, + 0.0050412737764418125, + 0.002766366582363844, + 0.003489692695438862, + 0.0055733947083354, + 0.002240956760942936, + 0.0025101250503212214, + 0.003845126600936055, + 0.003955572843551636, + 0.004890329670161009, + 0.003778299316763878, + 0.003584617283195257, + 0.004370702896267176, + 0.0044172112829983234, + 0.0030275536701083183, + 0.005208026152104139, + 0.006645962130278349, + 0.01135089248418808, + 0.003028025384992361, + 0.00599568709731102, + 0.022463809698820114, + 0.004538624081760645, + 0.005260269623249769, + 0.007561549544334412, + 0.007394124288111925, + 0.01182436104863882, + 0.0069161634892225266, + 0.011393270455300808, + 0.008886178024113178, + 0.015894366428256035, + 0.011173425242304802, + 0.005801551975309849, + 0.004397060722112656, + 0.011667325161397457, + 0.004579521249979734, + 0.005353305488824844, + 0.005541064776480198, + 0.0037835733965039253, + 0.008023104630410671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.029796123504638672 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, high-angle, top-down bird's-eye view, the camera gazes upon a vast, unbroken ocean of lush green foliage stretching to a soft horizon. The scene is a vibrant, dense tapestry of foliage, with an infinite variety of sizes and shapes, creating a random, natural pattern with no discernible organization. Rich, complex shades of green dominate the frame, ranging from deep emerald and jade in the shadowy depths to brilliant lime and chartreuse in the bright, clear sunlight, making some leaves appear almost translucent. The light creates a speckled, shimmering effect across the landscape, highlighting the waxy texture of some leaves and the delicate, intricate veining of others. The entire scene unfolds in perfect, serene silence, capturing the profound tranquility and overwhelming richness of nature in a single, breathtaking shot.", + "seed": 9398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_8/index_3645_seed_9398_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10477101439610123, + "vlm_logprob": -0.10477101439610123, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.10477101439610123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16970711946487427 + ], + "opposite_logprob_tokens": [ + -1.9197070598602295 + ], + "logprob_reward": -0.16970711946487427, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16970711946487427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21508391201496124, + 0.0865120068192482, + 0.16231095790863037, + 0.038275402039289474, + 0.06135837733745575, + 0.07525159418582916, + 0.05900023132562637, + 0.05138818547129631, + 0.18132081627845764, + 0.06949843466281891 + ], + "credit_map": [ + 0.020150700584053993, + 0.02827008254826069, + 0.045266229659318924, + 0.018613437190651894, + 0.01797652617096901, + 0.033668555319309235, + 0.024446051567792892, + 0.02669234201312065, + 0.0079009048640728, + 0.007469924166798592, + 0.008944526314735413, + 0.009096171706914902, + 0.00848058145493269, + 0.01952364481985569, + 0.00858229584991932, + 0.016513975337147713, + 0.012623539194464684, + 0.005777521058917046, + 0.0710597112774849, + 0.024058088660240173, + 0.006919022649526596, + 0.02487475797533989, + 0.007846616208553314, + 0.009151711128652096, + 0.004211637657135725, + 0.0053316084668040276, + 0.0076682609505951405, + 0.0038895036559551954, + 0.0044752461835742, + 0.004641158971935511, + 0.003969505429267883, + 0.004088482819497585, + 0.0068455287255346775, + 0.007820401340723038, + 0.01000606082379818, + 0.008503250777721405, + 0.007518682163208723, + 0.007355688605457544, + 0.0062233577482402325, + 0.007085410878062248, + 0.010173073969781399, + 0.007362185511738062, + 0.012382980436086655, + 0.00929379090666771, + 0.008985817432403564, + 0.008247549645602703, + 0.008531730622053146, + 0.010274467058479786, + 0.0040567247197031975, + 0.0059617431834340096, + 0.009698446840047836, + 0.01108036283403635, + 0.00741998665034771, + 0.007300993427634239, + 0.007135194726288319, + 0.0063467868603765965, + 0.006140219047665596, + 0.0052095758728682995, + 0.00895540602505207, + 0.005347725469619036, + 0.006258619017899036, + 0.008360969834029675, + 0.004831773228943348, + 0.006283903028815985, + 0.01958451420068741, + 0.016894496977329254, + 0.01904824562370777, + 0.01891578920185566, + 0.03646814450621605, + 0.0399780198931694, + 0.014667262323200703, + 0.01576436124742031, + 0.008466563187539577, + 0.008569179102778435, + 0.02127198502421379, + 0.006440350320190191, + 0.007900165393948555, + 0.00617821142077446, + 0.005682239308953285, + 0.004989743698388338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.16970711946487427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869554191827774 + ], + "opposite_logprob_tokens": [ + -3.0486955642700195 + ], + "logprob_reward": -0.04869554191827774, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869554191827774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07337528467178345, + 0.06056104972958565, + 0.14630837738513947, + 0.02326946333050728, + 0.03727138042449951, + 0.027526622638106346, + 0.05053791031241417, + 0.063630111515522, + 0.458307683467865, + 0.05921216309070587 + ], + "credit_map": [ + 0.007571694906800985, + 0.013015247881412506, + 0.008620204403996468, + 0.005864484701305628, + 0.006827837787568569, + 0.010075784288346767, + 0.012241702526807785, + 0.00915832445025444, + 0.006881857290863991, + 0.006841386202722788, + 0.010078388266265392, + 0.005491734482347965, + 0.005339710041880608, + 0.00816456601023674, + 0.006427278742194176, + 0.011336124502122402, + 0.012295563705265522, + 0.0054077403619885445, + 0.06347235292196274, + 0.022862646728754044, + 0.006526004057377577, + 0.01941116899251938, + 0.00773618882521987, + 0.008596700616180897, + 0.0022471053525805473, + 0.0025345361791551113, + 0.0044151064939796925, + 0.002542564645409584, + 0.0016818922013044357, + 0.002452194457873702, + 0.002440875396132469, + 0.004955186974257231, + 0.00420753238722682, + 0.003982491325587034, + 0.004735657479614019, + 0.005264312494546175, + 0.005042949691414833, + 0.004844936542212963, + 0.003815161995589733, + 0.0053783380426466465, + 0.0034528085961937904, + 0.0026345658116042614, + 0.004236229695379734, + 0.0032617952674627304, + 0.003524281783029437, + 0.0035867912229150534, + 0.0035607984755188227, + 0.0032693520188331604, + 0.005959922447800636, + 0.005805969703942537, + 0.007175586652010679, + 0.007859861478209496, + 0.007678261026740074, + 0.0051992908120155334, + 0.00505107082426548, + 0.005807943642139435, + 0.007030074950307608, + 0.0074009052477777, + 0.010534998029470444, + 0.005841698031872511, + 0.00907494779676199, + 0.0072910478338599205, + 0.00793166272342205, + 0.008524773642420769, + 0.048090774565935135, + 0.04602436348795891, + 0.04834694787859917, + 0.05203690379858017, + 0.08956743776798248, + 0.088447704911232, + 0.042126890271902084, + 0.04366663470864296, + 0.006041656248271465, + 0.0084971459582448, + 0.013352084904909134, + 0.006994269322603941, + 0.0063788979314267635, + 0.006489583756774664, + 0.006074592471122742, + 0.00538392411544919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.04869554191827774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252151221036911 + ], + "opposite_logprob_tokens": [ + -1.5021512508392334 + ], + "logprob_reward": -0.252151221036911, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252151221036911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11067081242799759, + 0.11932256817817688, + 0.02399381436407566, + 0.06720133870840073, + 0.04622002691030502, + 0.05560552328824997, + 0.2913021743297577, + 0.06395304203033447, + 0.1133919432759285, + 0.10833876579999924 + ], + "credit_map": [ + 0.012462960556149483, + 0.015533060766756535, + 0.017209233716130257, + 0.008480148389935493, + 0.01094167958945036, + 0.016233744099736214, + 0.014087225310504436, + 0.01572275161743164, + 0.00900435820221901, + 0.012457268312573433, + 0.013125035911798477, + 0.00871890876442194, + 0.01077559869736433, + 0.033931780606508255, + 0.011925294995307922, + 0.019384313374757767, + 0.0031353302765637636, + 0.002371155424043536, + 0.003211020492017269, + 0.0023301474284380674, + 0.0022437504958361387, + 0.00403881911188364, + 0.003390437923371792, + 0.0032731518149375916, + 0.007359933573752642, + 0.006961555685847998, + 0.008871329948306084, + 0.007844151929020882, + 0.007166384253650904, + 0.010311109013855457, + 0.006999094970524311, + 0.011687769554555416, + 0.004910983610898256, + 0.004603161476552486, + 0.008925181813538074, + 0.005484468303620815, + 0.004309593699872494, + 0.006008570082485676, + 0.005964673589915037, + 0.006013388745486736, + 0.005023092031478882, + 0.004392058588564396, + 0.00833780225366354, + 0.008411255665123463, + 0.006836815737187862, + 0.009220201522111893, + 0.006941420957446098, + 0.0064428700134158134, + 0.022856395691633224, + 0.027216430753469467, + 0.055170536041259766, + 0.054781682789325714, + 0.033214639872312546, + 0.031513117253780365, + 0.03752486780285835, + 0.02902449294924736, + 0.007053903304040432, + 0.0061541590839624405, + 0.010393361561000347, + 0.005927505437284708, + 0.011320430785417557, + 0.009276245720684528, + 0.0058225784450769424, + 0.008004853501915932, + 0.011322597041726112, + 0.01281811110675335, + 0.011978686787188053, + 0.015481395646929741, + 0.021390950307250023, + 0.016448095440864563, + 0.012113988399505615, + 0.011838100850582123, + 0.010215158574283123, + 0.012104745954275131, + 0.02510388195514679, + 0.009729300625622272, + 0.015402662567794323, + 0.016081642359495163, + 0.012051375582814217, + 0.007649990264326334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.252151221036911 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645389039069414 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645389039069414, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645389039069414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1588510423898697, + 0.0809485986828804, + 0.05904148891568184, + 0.05826675146818161, + 0.035119082778692245, + 0.08076409995555878, + 0.14316387474536896, + 0.09111449867486954, + 0.07431711256504059, + 0.21841344237327576 + ], + "credit_map": [ + 0.011875642463564873, + 0.015582784079015255, + 0.042338382452726364, + 0.01097045373171568, + 0.015463386662304401, + 0.018492359668016434, + 0.012817198410630226, + 0.03131083399057388, + 0.008621051907539368, + 0.010176463983952999, + 0.010349973104894161, + 0.010619663633406162, + 0.007926003076136112, + 0.010669730603694916, + 0.01262829639017582, + 0.009957416914403439, + 0.004867045674473047, + 0.0038449487183243036, + 0.02272617444396019, + 0.007462989538908005, + 0.0034747093450278044, + 0.006948288530111313, + 0.0048128534108400345, + 0.004904479719698429, + 0.004566969349980354, + 0.005204356275498867, + 0.008360722102224827, + 0.004952859133481979, + 0.0046135177835822105, + 0.010680019855499268, + 0.008536342531442642, + 0.011351962573826313, + 0.004411201924085617, + 0.0038382606580853462, + 0.005923137068748474, + 0.0038877171464264393, + 0.0037765391170978546, + 0.005015788599848747, + 0.004149650223553181, + 0.004116787109524012, + 0.007576332427561283, + 0.009806430898606777, + 0.009437422268092632, + 0.012747545726597309, + 0.011051611043512821, + 0.010811196640133858, + 0.008924838155508041, + 0.010408720932900906, + 0.01102355308830738, + 0.015224887058138847, + 0.029139503836631775, + 0.02352072112262249, + 0.016188200563192368, + 0.014813427813351154, + 0.01816294901072979, + 0.015090635977685452, + 0.009352943859994411, + 0.012102119624614716, + 0.012813142500817776, + 0.009272482246160507, + 0.013352611102163792, + 0.014437529258430004, + 0.009820458479225636, + 0.009963215328752995, + 0.007659090217202902, + 0.010074571706354618, + 0.011433413252234459, + 0.008837894536554813, + 0.012207716703414917, + 0.010972881689667702, + 0.0056281075812876225, + 0.007503439672291279, + 0.026048079133033752, + 0.026310736313462257, + 0.06728962808847427, + 0.022165615111589432, + 0.023996684700250626, + 0.02141156978905201, + 0.017849169671535492, + 0.013341963291168213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.004645389039069414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865580052137375 + ], + "opposite_logprob_tokens": [ + -3.0486557483673096 + ], + "logprob_reward": -0.04865580052137375, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865580052137375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10659393668174744, + 0.035843972116708755, + 0.06077589467167854, + 0.06810615211725235, + 0.12524712085723877, + 0.07938602566719055, + 0.062373463064432144, + 0.06864795088768005, + 0.06836465746164322, + 0.32466086745262146 + ], + "credit_map": [ + 0.009094027802348137, + 0.01556246168911457, + 0.014106250368058681, + 0.007566407788544893, + 0.010618789121508598, + 0.021724557504057884, + 0.013017346151173115, + 0.014904101379215717, + 0.0036610832903534174, + 0.004249775782227516, + 0.005185374990105629, + 0.0037195617333054543, + 0.0037593974266201258, + 0.005732591263949871, + 0.003289867890998721, + 0.00624631904065609, + 0.004616564605385065, + 0.0035887856502085924, + 0.019140852615237236, + 0.008329015225172043, + 0.0035389491822570562, + 0.012817593291401863, + 0.003589828033000231, + 0.005154307000339031, + 0.0067225173115730286, + 0.008092312142252922, + 0.011069700121879578, + 0.007226895075291395, + 0.006618645042181015, + 0.010985996574163437, + 0.010134251788258553, + 0.007255840115249157, + 0.014201798476278782, + 0.0174444317817688, + 0.016023408621549606, + 0.014783474616706371, + 0.014102901332080364, + 0.01483707595616579, + 0.012384593486785889, + 0.02146943099796772, + 0.006771979853510857, + 0.01061839610338211, + 0.01069228257983923, + 0.011893446557223797, + 0.012370898388326168, + 0.009696501307189465, + 0.009264549240469933, + 0.008077973499894142, + 0.007487304508686066, + 0.0073016975075006485, + 0.010155185125768185, + 0.0070909615606069565, + 0.008534575812518597, + 0.007305938750505447, + 0.007595154922455549, + 0.006902645342051983, + 0.006799950730055571, + 0.008226582780480385, + 0.014200487174093723, + 0.006448716856539249, + 0.009505469352006912, + 0.009947434067726135, + 0.007012936286628246, + 0.006506375502794981, + 0.0068483673967421055, + 0.008697761222720146, + 0.009657559916377068, + 0.007067129481583834, + 0.009724564850330353, + 0.014394810423254967, + 0.0060988240875303745, + 0.005875642877072096, + 0.04000519961118698, + 0.038916561752557755, + 0.10601889342069626, + 0.029865596443414688, + 0.037167031317949295, + 0.029360488057136536, + 0.024697378277778625, + 0.018629711121320724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.04865580052137375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, high-angle, top-down bird's-eye view, the camera gazes upon a vast, unbroken ocean of lush green foliage stretching to a soft horizon. The scene is a vibrant, dense tapestry of foliage, with an infinite variety of sizes and shapes, creating a random, natural pattern with no discernible organization. Rich, complex shades of green dominate the frame, ranging from deep emerald and jade in the shadowy depths to brilliant lime and chartreuse in the bright, clear sunlight, making some leaves appear almost translucent. The light creates a speckled, shimmering effect across the landscape, highlighting the waxy texture of some leaves and the delicate, intricate veining of others. The entire scene unfolds in perfect, serene silence, capturing the profound tranquility and overwhelming richness of nature in a single, breathtaking shot.", + "seed": 9399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_8/index_3645_seed_9399_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09200885156169533, + "vlm_logprob": -0.09200885156169533, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.09200885156169533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1181163415312767 + ], + "opposite_logprob_tokens": [ + -2.2431163787841797 + ], + "logprob_reward": -0.1181163415312767, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1181163415312767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26364684104919434, + 0.06788979470729828, + 0.03831734508275986, + 0.11523199081420898, + 0.2833995521068573, + 0.053748246282339096, + 0.02983808144927025, + 0.04720790684223175, + 0.05812317132949829, + 0.04259698465466499 + ], + "credit_map": [ + 0.021194428205490112, + 0.03727603703737259, + 0.02062048763036728, + 0.02476809173822403, + 0.06459017097949982, + 0.049181144684553146, + 0.02241380140185356, + 0.023602686822414398, + 0.00579358683899045, + 0.011339845135807991, + 0.009376930072903633, + 0.008320383727550507, + 0.013429372571408749, + 0.0070189787074923515, + 0.006786103826016188, + 0.0058246031403541565, + 0.004211588762700558, + 0.0056467908434569836, + 0.00640604505315423, + 0.005455668084323406, + 0.004870664793998003, + 0.005160540342330933, + 0.0032542431727051735, + 0.0033118075225502253, + 0.0089733200147748, + 0.013259726576507092, + 0.026712339371442795, + 0.017808299511671066, + 0.015172909013926983, + 0.015128279104828835, + 0.01015789620578289, + 0.008019222877919674, + 0.020137326791882515, + 0.042447350919246674, + 0.05468877777457237, + 0.02444782853126526, + 0.03472736477851868, + 0.045078061521053314, + 0.03864166513085365, + 0.023231210187077522, + 0.0042409468442201614, + 0.004791272804141045, + 0.013648507185280323, + 0.004499007947742939, + 0.013746477663516998, + 0.004872579593211412, + 0.00438166968524456, + 0.003567787818610668, + 0.004412333946675062, + 0.004647152964025736, + 0.00490151671692729, + 0.003752312157303095, + 0.0024511185474693775, + 0.0038082741666585207, + 0.0032604518346488476, + 0.0026049241423606873, + 0.004282232839614153, + 0.008852958679199219, + 0.004864846356213093, + 0.006474664900451899, + 0.011054672300815582, + 0.004075965844094753, + 0.0035820913035422564, + 0.00402047298848629, + 0.005512644536793232, + 0.005552025977522135, + 0.01415122952312231, + 0.006245334632694721, + 0.014245604164898396, + 0.0041939918883144855, + 0.005170186050236225, + 0.003052158048376441, + 0.00465520890429616, + 0.004526181146502495, + 0.005434408318251371, + 0.0049733384512364864, + 0.007107362151145935, + 0.005621687509119511, + 0.006314738653600216, + 0.003964063711464405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.1181163415312767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330542728304863 + ], + "opposite_logprob_tokens": [ + -3.7733054161071777 + ], + "logprob_reward": -0.02330542728304863, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330542728304863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10746697336435318, + 0.1072249785065651, + 0.09843853861093521, + 0.046095311641693115, + 0.2476699948310852, + 0.19199761748313904, + 0.022203344851732254, + 0.057182032614946365, + 0.06798139214515686, + 0.05373978242278099 + ], + "credit_map": [ + 0.008683104999363422, + 0.011282999999821186, + 0.023594887927174568, + 0.010962084867060184, + 0.020445195958018303, + 0.014153040945529938, + 0.010016459971666336, + 0.008329210802912712, + 0.007950318977236748, + 0.02779575251042843, + 0.008625032380223274, + 0.008543629199266434, + 0.02537783980369568, + 0.013131237588822842, + 0.008644149638712406, + 0.00715703284367919, + 0.008490930311381817, + 0.013470984995365143, + 0.019548650830984116, + 0.018860897049307823, + 0.011845269240438938, + 0.009935429319739342, + 0.008283786475658417, + 0.008002599701285362, + 0.003919033333659172, + 0.005829180125147104, + 0.0064741577953100204, + 0.005263093858957291, + 0.004902317188680172, + 0.008027022704482079, + 0.005862501449882984, + 0.005818010773509741, + 0.013085858896374702, + 0.017752552404999733, + 0.0631391778588295, + 0.020970163866877556, + 0.050607096403837204, + 0.034647535532712936, + 0.03390653803944588, + 0.013561086729168892, + 0.016955995932221413, + 0.014870129525661469, + 0.05250125750899315, + 0.016488390043377876, + 0.04886040464043617, + 0.012680768966674805, + 0.015276489779353142, + 0.014364189468324184, + 0.0027628783136606216, + 0.0028096730820834637, + 0.002866250928491354, + 0.003961828537285328, + 0.002404206432402134, + 0.0016556564951315522, + 0.0033609913662075996, + 0.00238186027854681, + 0.0066049592569470406, + 0.015195728279650211, + 0.004906805697828531, + 0.006031856406480074, + 0.011448536068201065, + 0.004283504094928503, + 0.004832146689295769, + 0.003878505900502205, + 0.0060646808706223965, + 0.01268088724464178, + 0.005765577312558889, + 0.006071378011256456, + 0.015637751668691635, + 0.008076169528067112, + 0.008042178116738796, + 0.005642773117870092, + 0.006287488620728254, + 0.008359171450138092, + 0.007144616451114416, + 0.00645119184628129, + 0.0063302358612418175, + 0.00627143494784832, + 0.007058699615299702, + 0.005836948286741972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.02330542728304863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160122215747833 + ], + "opposite_logprob_tokens": [ + -1.701601266860962 + ], + "logprob_reward": -0.20160122215747833, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160122215747833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1904224008321762, + 0.11420363187789917, + 0.12335062772035599, + 0.04066922515630722, + 0.06502532213926315, + 0.08522865176200867, + 0.01706121861934662, + 0.0560702309012413, + 0.24545051157474518, + 0.06251818686723709 + ], + "credit_map": [ + 0.01597440615296364, + 0.020653817802667618, + 0.03576859086751938, + 0.015506191179156303, + 0.05979488044977188, + 0.01683666557073593, + 0.012341968715190887, + 0.013545883819460869, + 0.00979468785226345, + 0.025697994977235794, + 0.010061787441372871, + 0.010599454864859581, + 0.027028290554881096, + 0.011226312257349491, + 0.010652365162968636, + 0.009142736904323101, + 0.00962154008448124, + 0.01952611282467842, + 0.015087489038705826, + 0.01802169159054756, + 0.013440561480820179, + 0.022858090698719025, + 0.010691942647099495, + 0.014103204943239689, + 0.0031308287288993597, + 0.005369415041059256, + 0.007219052407890558, + 0.005658578127622604, + 0.005765331443399191, + 0.006872315891087055, + 0.0032795381266623735, + 0.0033741644583642483, + 0.006186248268932104, + 0.011685584671795368, + 0.00743747828528285, + 0.00651510851457715, + 0.00815972127020359, + 0.009246955625712872, + 0.010195227339863777, + 0.005598996300250292, + 0.008391180075705051, + 0.0101582957431674, + 0.022527264431118965, + 0.007394273765385151, + 0.017899587750434875, + 0.005817938596010208, + 0.0056242565624415874, + 0.007415854837745428, + 0.0027380366809666157, + 0.001822170801460743, + 0.0022747369948774576, + 0.0017515873769298196, + 0.001815451541915536, + 0.0026534327771514654, + 0.001687017036601901, + 0.0023187845945358276, + 0.004945402964949608, + 0.011104912497103214, + 0.005290466360747814, + 0.004872013349086046, + 0.016190679743885994, + 0.004987179767340422, + 0.005138862878084183, + 0.0035407128743827343, + 0.03426370397210121, + 0.013137195259332657, + 0.10121022164821625, + 0.036660704761743546, + 0.030274594202637672, + 0.01009050291031599, + 0.009487171657383442, + 0.010326401330530643, + 0.009338270872831345, + 0.012117034755647182, + 0.009982207790017128, + 0.007984514348208904, + 0.00766192888841033, + 0.004512323532253504, + 0.005815783049911261, + 0.005106122698634863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.20160122215747833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097756464034319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097756464034319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38896459341049194, + 0.08483777195215225, + 0.1660567820072174, + 0.08697942644357681, + 0.05752085521817207, + 0.08729530870914459, + 0.024380873888731003, + 0.030735844746232033, + 0.041304778307676315, + 0.03192371129989624 + ], + "credit_map": [ + 0.021587438881397247, + 0.03538952395319939, + 0.08847316354513168, + 0.025855105370283127, + 0.13794520497322083, + 0.04200400784611702, + 0.01675122231245041, + 0.0209589134901762, + 0.006454544607549906, + 0.01923607476055622, + 0.00611876230686903, + 0.006914202589541674, + 0.017610909417271614, + 0.012292752973735332, + 0.008778016082942486, + 0.007432514801621437, + 0.012974074110388756, + 0.02519758604466915, + 0.02046933025121689, + 0.02440955489873886, + 0.018841877579689026, + 0.03028147853910923, + 0.01499705296009779, + 0.018885837867856026, + 0.0069034770131111145, + 0.012298998422920704, + 0.010493708774447441, + 0.011018438264727592, + 0.010766791179776192, + 0.017928743734955788, + 0.009291468188166618, + 0.008277803659439087, + 0.004396941512823105, + 0.009385429322719574, + 0.013847201131284237, + 0.006062776781618595, + 0.00564920948818326, + 0.008397708646953106, + 0.006353392265737057, + 0.00342819862999022, + 0.008183613419532776, + 0.008216057904064655, + 0.022411448881030083, + 0.007195961661636829, + 0.02103515900671482, + 0.006685326807200909, + 0.006252986844629049, + 0.007314750924706459, + 0.003252363530918956, + 0.004415333271026611, + 0.0033773137256503105, + 0.003925802186131477, + 0.002369566587731242, + 0.0019276615930721164, + 0.0029518376104533672, + 0.002160994801670313, + 0.002847915282472968, + 0.006636008620262146, + 0.0029290798120200634, + 0.0036242911592125893, + 0.003758077509701252, + 0.003845405299216509, + 0.004102824255824089, + 0.002992244204506278, + 0.0036186040379107, + 0.005221047438681126, + 0.005012910813093185, + 0.003240636084228754, + 0.014283054508268833, + 0.0032483800314366817, + 0.00410641822963953, + 0.0025737229734659195, + 0.003942023031413555, + 0.00411445926874876, + 0.005436264909803867, + 0.004548949189484119, + 0.0041200825944542885, + 0.0031607155688107014, + 0.0035738316364586353, + 0.003027387196198106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.004097756464034319 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, high-angle, top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene consist of a vast, unbroken ocean of lush green foliage stretching to a soft horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves depicted with a random, natural pattern showing no discernible organization?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves display rich, complex shades of green ranging from deep emerald to brilliant lime and chartreuse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a speckled, shimmering light effect highlighting the waxy texture and intricate veining of the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129235103726387 + ], + "opposite_logprob_tokens": [ + -2.2379236221313477 + ], + "logprob_reward": -0.1129235103726387, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129235103726387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4140537977218628, + 0.06431122869253159, + 0.12462449818849564, + 0.07503341138362885, + 0.07287698984146118, + 0.08221970498561859, + 0.018269602209329605, + 0.026499805971980095, + 0.0885334461927414, + 0.033577483147382736 + ], + "credit_map": [ + 0.024959353730082512, + 0.03644257038831711, + 0.09962984174489975, + 0.02699890360236168, + 0.14716944098472595, + 0.039452143013477325, + 0.01758592389523983, + 0.021815624088048935, + 0.005207380745559931, + 0.008072306402027607, + 0.011231301352381706, + 0.007206792011857033, + 0.009951181709766388, + 0.0085378373041749, + 0.007571608759462833, + 0.006532822735607624, + 0.009009062312543392, + 0.015408683568239212, + 0.021318119019269943, + 0.019616873934864998, + 0.014064772985875607, + 0.018625598400831223, + 0.01212261151522398, + 0.01445878017693758, + 0.0059296260587871075, + 0.010847803205251694, + 0.00838983990252018, + 0.007204113062471151, + 0.008648889139294624, + 0.016541043296456337, + 0.009231633506715298, + 0.008240454830229282, + 0.0038328594528138638, + 0.022016925737261772, + 0.017263898625969887, + 0.0070785898715257645, + 0.010833331383764744, + 0.004672112874686718, + 0.0035323738120496273, + 0.0036468964535743, + 0.005690560210496187, + 0.0053880587220191956, + 0.02672509104013443, + 0.0068952776491642, + 0.022198308259248734, + 0.005426489748060703, + 0.004998688120394945, + 0.004897232633084059, + 0.0019731794018298388, + 0.002928640926256776, + 0.0025948502589017153, + 0.0032969193998724222, + 0.0022834253031760454, + 0.0015286622801795602, + 0.0016968115232884884, + 0.0019671113695949316, + 0.0027628708630800247, + 0.005577320232987404, + 0.0025723695289343596, + 0.003729978110641241, + 0.0042602429166436195, + 0.002152498811483383, + 0.002664014231413603, + 0.00278051127679646, + 0.014401893131434917, + 0.01046641357243061, + 0.022139623761177063, + 0.011760899797081947, + 0.015238604508340359, + 0.004800343886017799, + 0.005289739463478327, + 0.004435930401086807, + 0.003951109480112791, + 0.0036879933904856443, + 0.005106854252517223, + 0.004589905962347984, + 0.00459335558116436, + 0.004096247721463442, + 0.003896851558238268, + 0.003655166132375598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.1129235103726387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_9/index_tensor([9445, 9445])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_9/index_tensor([9445, 9445])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7a2cc8af35bccfa577fc8bd8aa61dcb364ed37 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_9/index_tensor([9445, 9445])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth, low-angle, 360-degree panning shot closely follows a classic Porsche 911 Carrera RS 3.8, painted a vibrant signal yellow that gleams brilliantly in the harsh afternoon sun. The iconic, massive \"whale tail\" spoiler and the \"Carrera RS\" lettering on the side are outlined in harsh, yet unmistakable black lines, lending the image a rugged, retro, anime-inspired quality. As it drifts through a corner at high speed on a rough, gravel mountain road, the car's entire body experiences a violent jolting, each compression and rebound stroke of the suspension powerful. The wide rear wheels furiously paw at the loose gravel, sending clouds of thick, yellow-brown dust cascading into the air, forming a billowing cloud that nearly obscures the rear half of the car. The backdrop, a jagged gray rock face and sparse, dark green pine trees, contrasts starkly against the almost unrealistic blue of the sky, rendering every element vibrant and vibrant. As the camera moves steadily and gracefully around the car body, light flows on the paint, creating sharp highlights and blocky deep shadows. The entire picture is full of strong visual impact, perfectly capturing this adrenaline-pumping and blood-boiling racing moment.", + "seed": 14215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_9/index_9445_seed_14215_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033716813381761315, + "vlm_logprob": -0.033716813381761315, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.033716813381761315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012276679277420044 + ], + "opposite_logprob_tokens": [ + -5.387276649475098 + ], + "logprob_reward": -0.012276679277420044, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012276679277420044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7130553722381592, + 0.05570603907108307, + 0.026386573910713196, + 0.025663604959845543, + 0.020776106044650078, + 0.02176707610487938, + 0.04138060286641121, + 0.04706519469618797, + 0.020776420831680298, + 0.027422979474067688 + ], + "credit_map": [ + 0.05360002815723419, + 0.06964089721441269, + 0.07682891190052032, + 0.13370059430599213, + 0.0879155620932579, + 0.08259083330631256, + 0.11565250158309937, + 0.093126080930233, + 0.005731330253183842, + 0.005219305865466595, + 0.013528089970350266, + 0.005206756293773651, + 0.005248138681054115, + 0.00596368545666337, + 0.00470696808770299, + 0.010101762600243092, + 0.00252680410631001, + 0.0031728879548609257, + 0.0027378457598388195, + 0.005348665174096823, + 0.003367119003087282, + 0.0034449293743819, + 0.0031683563720434904, + 0.0026199661660939455, + 0.0034831140656024218, + 0.003727282863110304, + 0.0019261478446424007, + 0.0022798588033765554, + 0.0032098963856697083, + 0.0024140437599271536, + 0.004827365279197693, + 0.0037958952598273754, + 0.002291907323524356, + 0.0032398987095803022, + 0.002283815061673522, + 0.0020553586073219776, + 0.002608211012557149, + 0.003116252599284053, + 0.00278817443177104, + 0.0023924880661070347, + 0.003158930689096451, + 0.005168666131794453, + 0.002367323264479637, + 0.0018207503017038107, + 0.0025386640336364508, + 0.002535618143156171, + 0.0019417655421420932, + 0.0022353585809469223, + 0.01068840641528368, + 0.011415524408221245, + 0.002009208081290126, + 0.0030861077830195427, + 0.0037154906895011663, + 0.005082827527076006, + 0.0025555770844221115, + 0.002827458316460252, + 0.006027981638908386, + 0.004990539979189634, + 0.003122338792309165, + 0.004753173794597387, + 0.008271426893770695, + 0.010911116376519203, + 0.0040860227309167385, + 0.0049025979824364185, + 0.0027278983034193516, + 0.0027488796040415764, + 0.0025359138380736113, + 0.0016858333256095648, + 0.002331323688849807, + 0.0026893175672739744, + 0.0019450323889032006, + 0.004112221300601959, + 0.004497762303799391, + 0.005470708478242159, + 0.0032675780821591616, + 0.00407096603885293, + 0.0024426933377981186, + 0.002748832106590271, + 0.0026809449773281813, + 0.0022434929851442575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012276679277420044 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979793444275856 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.00979793444275856, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979793444275856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7847520112991333, + 0.034610848873853683, + 0.016588017344474792, + 0.03502276912331581, + 0.03663504868745804, + 0.01783757098019123, + 0.025605985894799232, + 0.019985835999250412, + 0.017620248720049858, + 0.011341697536408901 + ], + "credit_map": [ + 0.05557926371693611, + 0.07536324113607407, + 0.08129696547985077, + 0.14965210855007172, + 0.09987171739339828, + 0.08738712221384048, + 0.12962770462036133, + 0.10597386211156845, + 0.0047822436317801476, + 0.004135147202759981, + 0.006752254907041788, + 0.0028667286969721317, + 0.0034630559384822845, + 0.004951131530106068, + 0.00251224753446877, + 0.005148040130734444, + 0.001804161467589438, + 0.0019779037684202194, + 0.0019726629834622145, + 0.003232691204175353, + 0.0017601585714146495, + 0.0018718801438808441, + 0.002063255524262786, + 0.0019053034484386444, + 0.004123931284993887, + 0.0057141948491334915, + 0.0030822171829640865, + 0.00316886929795146, + 0.005629589315503836, + 0.004953516647219658, + 0.0036825640127062798, + 0.004667884670197964, + 0.003289945889264345, + 0.003694552695378661, + 0.0030362263787537813, + 0.0025274467188864946, + 0.004808973055332899, + 0.005055882968008518, + 0.007241921033710241, + 0.006980098783969879, + 0.0021703310776501894, + 0.0020832347217947245, + 0.0012296813074499369, + 0.001740913954563439, + 0.003438791260123253, + 0.0032292488031089306, + 0.0020754344295710325, + 0.001869935542345047, + 0.005487649701535702, + 0.005815419368445873, + 0.0014682990731671453, + 0.00280865840613842, + 0.002285195514559746, + 0.0039087338373064995, + 0.002050498966127634, + 0.0017815319588407874, + 0.002075209515169263, + 0.0027215497102588415, + 0.0014651681995019317, + 0.0015069462824612856, + 0.0025419469457119703, + 0.004709981847554445, + 0.002528943819925189, + 0.002436089562252164, + 0.0017865161644294858, + 0.0025473046116530895, + 0.0024026744067668915, + 0.001417868072167039, + 0.002121919533237815, + 0.0029771518893539906, + 0.0016285551246255636, + 0.002738259732723236, + 0.0021415571682155132, + 0.001973152859136462, + 0.0012814539950340986, + 0.0015392806380987167, + 0.0010597080690786242, + 0.0011248694499954581, + 0.0012437229743227363, + 0.0009779521496966481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.00979793444275856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430426187813282 + ], + "opposite_logprob_tokens": [ + -3.1680426597595215 + ], + "logprob_reward": -0.0430426187813282, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430426187813282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3170095682144165, + 0.2902132272720337, + 0.041836511343717575, + 0.056350838392972946, + 0.026881618425250053, + 0.025143878534436226, + 0.05096292868256569, + 0.16287130117416382, + 0.016822578385472298, + 0.011907466687262058 + ], + "credit_map": [ + 0.030064264312386513, + 0.03402763605117798, + 0.02413860335946083, + 0.03961913287639618, + 0.03700878471136093, + 0.039535410702228546, + 0.040353816002607346, + 0.07226195186376572, + 0.030039185658097267, + 0.029260512441396713, + 0.06968435645103455, + 0.026467176154255867, + 0.03409162908792496, + 0.031680215150117874, + 0.016069510951638222, + 0.052920691668987274, + 0.004250653088092804, + 0.004512694664299488, + 0.004404277075082064, + 0.005047381855547428, + 0.0073769111186265945, + 0.0063005294650793076, + 0.004797025583684444, + 0.005147043615579605, + 0.007127344608306885, + 0.008340728469192982, + 0.005683078896254301, + 0.004507460631430149, + 0.008078690618276596, + 0.0068325623869895935, + 0.007657099515199661, + 0.008123879320919514, + 0.0028381417505443096, + 0.002863343106582761, + 0.003141129855066538, + 0.0022547917906194925, + 0.003107571741566062, + 0.003965850453823805, + 0.0050825923681259155, + 0.0036282013170421124, + 0.0039124502800405025, + 0.0035187597386538982, + 0.0016960870707407594, + 0.002910114359110594, + 0.003650861093774438, + 0.0032032672315835953, + 0.0035333463456481695, + 0.002718995325267315, + 0.012552989646792412, + 0.010777708142995834, + 0.0041075460612773895, + 0.0032743909396231174, + 0.005185656249523163, + 0.006933034397661686, + 0.004043301567435265, + 0.004088310524821281, + 0.02071569673717022, + 0.01586228422820568, + 0.010113205760717392, + 0.013030369766056538, + 0.02401573210954666, + 0.041416388005018234, + 0.01396953221410513, + 0.02374810166656971, + 0.002668125322088599, + 0.0025065524969249964, + 0.0018740257946774364, + 0.001457089208997786, + 0.001983265858143568, + 0.003119861474260688, + 0.001344388583675027, + 0.0018692732555791736, + 0.001480559236370027, + 0.0017833574675023556, + 0.0014417162165045738, + 0.0012298371875658631, + 0.0011439456138759851, + 0.0015000118874013424, + 0.0013363080797716975, + 0.0019917325116693974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0430426187813282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031898352317512035 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.0031898352317512035, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031898352317512035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.807518720626831, + 0.016040407121181488, + 0.011182799935340881, + 0.013807464390993118, + 0.06780611723661423, + 0.015006558038294315, + 0.018752627074718475, + 0.021135693415999413, + 0.009757639840245247, + 0.018991919234395027 + ], + "credit_map": [ + 0.058280058205127716, + 0.07708772271871567, + 0.0871533527970314, + 0.15086723864078522, + 0.09864963591098785, + 0.09266357868909836, + 0.1294422596693039, + 0.11337493360042572, + 0.0017684734193608165, + 0.0013565340777859092, + 0.003999247215688229, + 0.001354197389446199, + 0.0017396040493622422, + 0.0020798672921955585, + 0.001206533401273191, + 0.0025359492283314466, + 0.0013965752441436052, + 0.0012658396735787392, + 0.0012787928571924567, + 0.0019051783019676805, + 0.0014195592375472188, + 0.0016594987828284502, + 0.0012387020979076624, + 0.0010186537401750684, + 0.0016147460555657744, + 0.0015281777596101165, + 0.000809294346254319, + 0.0008354288875125349, + 0.0015542027540504932, + 0.0018085066694766283, + 0.0026426631957292557, + 0.003014445072039962, + 0.0068262186832726, + 0.0072176214307546616, + 0.007104974705725908, + 0.0054949005134403706, + 0.007757172454148531, + 0.009991864673793316, + 0.012429940514266491, + 0.010983427055180073, + 0.0021938737481832504, + 0.003151951590552926, + 0.0008586298208683729, + 0.0007250630878843367, + 0.003358711488544941, + 0.002776231849566102, + 0.000987434177659452, + 0.000954663148149848, + 0.0037138189654797316, + 0.004184963647276163, + 0.0012821628479287028, + 0.0018150504911318421, + 0.0015114207053557038, + 0.003165019443258643, + 0.0014310134574770927, + 0.0016491785645484924, + 0.0030496222898364067, + 0.002586034359410405, + 0.001361141912639141, + 0.00142476346809417, + 0.002961084945127368, + 0.0051501113921403885, + 0.0016131328884512186, + 0.0029898013453930616, + 0.001167617621831596, + 0.0008716400479897857, + 0.0010355048580095172, + 0.0009058937430381775, + 0.0012866768520325422, + 0.0016617834335193038, + 0.0011100205592811108, + 0.0017185027245432138, + 0.003855789778754115, + 0.004919794853776693, + 0.0015615782467648387, + 0.0014651514356955886, + 0.0016698059625923634, + 0.002744875615462661, + 0.0013497048057615757, + 0.001425218302756548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0031898352317512035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027699917554855 + ], + "opposite_logprob_tokens": [ + -2.3502769470214844 + ], + "logprob_reward": -0.10027699917554855, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027699917554855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6449364423751831, + 0.028241468593478203, + 0.019714875146746635, + 0.04076666757464409, + 0.04827495664358139, + 0.051970768719911575, + 0.041632842272520065, + 0.04858105629682541, + 0.06753258407115936, + 0.008348251692950726 + ], + "credit_map": [ + 0.04600805789232254, + 0.06139358878135681, + 0.06956782937049866, + 0.11823616176843643, + 0.08009945601224899, + 0.07391443848609924, + 0.10059763491153717, + 0.09511936455965042, + 0.0032916306518018246, + 0.0028185099363327026, + 0.002591741504147649, + 0.002375421579927206, + 0.004859172739088535, + 0.003972593694925308, + 0.0033275026362389326, + 0.005004898179322481, + 0.0023693980183452368, + 0.0017507525626569986, + 0.0032046143896877766, + 0.003056825138628483, + 0.002592559438198805, + 0.0028405196499079466, + 0.002104919869452715, + 0.0017952893394976854, + 0.005497056990861893, + 0.004995222669094801, + 0.003648568643257022, + 0.002878715982660651, + 0.005233857315033674, + 0.004158776719123125, + 0.006688715424388647, + 0.007665760815143585, + 0.004291321616619825, + 0.0074637215584516525, + 0.004537953529506922, + 0.003999791108071804, + 0.00608575064688921, + 0.007560485042631626, + 0.0084355054423213, + 0.005900435149669647, + 0.009302952326834202, + 0.01414488349109888, + 0.002986261621117592, + 0.0034121605567634106, + 0.007682807743549347, + 0.006408081855624914, + 0.003092999802902341, + 0.004940629471093416, + 0.008840729482471943, + 0.008366396650671959, + 0.003313243854790926, + 0.003014463698491454, + 0.005041345953941345, + 0.005802475847303867, + 0.003242896869778633, + 0.0040112994611263275, + 0.006448449566960335, + 0.004665134474635124, + 0.003329680534079671, + 0.003390629542991519, + 0.006959403399378061, + 0.011925514787435532, + 0.004161652643233538, + 0.0077005974017083645, + 0.007793772034347057, + 0.007477226201444864, + 0.006656663026660681, + 0.006212408188730478, + 0.010272915475070477, + 0.01159921195358038, + 0.0065016113221645355, + 0.011018789373338223, + 0.0014937480445951223, + 0.0014153604861348867, + 0.0013843560591340065, + 0.000889000715687871, + 0.0006424682796932757, + 0.0010091576259583235, + 0.0006399144767783582, + 0.0008742476929910481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.10027699917554855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, low-angle, 360-degree panning shot closely follows a classic Porsche 911 Carrera RS 3.8, painted a vibrant signal yellow that gleams brilliantly in the harsh afternoon sun. The iconic, massive \"whale tail\" spoiler and the \"Carrera RS\" lettering on the side are outlined in harsh, yet unmistakable black lines, lending the image a rugged, retro, anime-inspired quality. As it drifts through a corner at high speed on a rough, gravel mountain road, the car's entire body experiences a violent jolting, each compression and rebound stroke of the suspension powerful. The wide rear wheels furiously paw at the loose gravel, sending clouds of thick, yellow-brown dust cascading into the air, forming a billowing cloud that nearly obscures the rear half of the car. The backdrop, a jagged gray rock face and sparse, dark green pine trees, contrasts starkly against the almost unrealistic blue of the sky, rendering every element vibrant and vibrant. As the camera moves steadily and gracefully around the car body, light flows on the paint, creating sharp highlights and blocky deep shadows. The entire picture is full of strong visual impact, perfectly capturing this adrenaline-pumping and blood-boiling racing moment.", + "seed": 14216, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_9/index_9445_seed_14216_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029774953611195087, + "vlm_logprob": -0.029774953611195087, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.029774953611195087, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017418738454580307 + ], + "opposite_logprob_tokens": [ + -4.89241886138916 + ], + "logprob_reward": -0.017418738454580307, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017418738454580307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07312842458486557, + 0.049216561019420624, + 0.06619980186223984, + 0.07020479440689087, + 0.337352454662323, + 0.061898402869701385, + 0.10568201541900635, + 0.08847138285636902, + 0.08909524232149124, + 0.058750975877046585 + ], + "credit_map": [ + 0.007819881662726402, + 0.0080407178029418, + 0.01040306780487299, + 0.013526088558137417, + 0.008975494652986526, + 0.007257073651999235, + 0.00547613063827157, + 0.011629960499703884, + 0.004386020824313164, + 0.00384759739972651, + 0.004903565160930157, + 0.006563351023942232, + 0.011198753491044044, + 0.008300410583615303, + 0.005141910165548325, + 0.00487494794651866, + 0.012805674225091934, + 0.007187551353126764, + 0.007247757166624069, + 0.0049820938147604465, + 0.0066423723474144936, + 0.013912391848862171, + 0.005656791850924492, + 0.007765165064483881, + 0.009551163762807846, + 0.01049833744764328, + 0.010877684690058231, + 0.00834411196410656, + 0.010643905028700829, + 0.007020141463726759, + 0.0049189371056854725, + 0.008350510150194168, + 0.047982603311538696, + 0.013586100190877914, + 0.14968238770961761, + 0.02226211689412594, + 0.025759313255548477, + 0.010819933377206326, + 0.054922133684158325, + 0.01233779825270176, + 0.01232015062123537, + 0.003926612436771393, + 0.008176460862159729, + 0.005046194419264793, + 0.0049612135626375675, + 0.009003572165966034, + 0.006773951463401318, + 0.011690240353345871, + 0.008808954618871212, + 0.005939166061580181, + 0.01729300431907177, + 0.013725103810429573, + 0.01054829079657793, + 0.011089988984167576, + 0.022025568410754204, + 0.016251929104328156, + 0.00787645485252142, + 0.005532249808311462, + 0.010747566819190979, + 0.008524201810359955, + 0.00768838869407773, + 0.014720051549375057, + 0.015354202128946781, + 0.0180282611399889, + 0.010554186068475246, + 0.012174886651337147, + 0.010909968987107277, + 0.017298057675361633, + 0.006639507133513689, + 0.009595492854714394, + 0.009503427892923355, + 0.012419701553881168, + 0.006310120690613985, + 0.008174249902367592, + 0.007450869306921959, + 0.012711115181446075, + 0.00624184450134635, + 0.0046811895444989204, + 0.006039992906153202, + 0.0071415891870856285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.017418738454580307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801357984542847 + ], + "opposite_logprob_tokens": [ + -4.634801387786865 + ], + "logprob_reward": -0.009801357984542847, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801357984542847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15357576310634613, + 0.15959107875823975, + 0.0545920692384243, + 0.06438951194286346, + 0.35154232382774353, + 0.04040367528796196, + 0.05428048223257065, + 0.036710213869810104, + 0.05189870670437813, + 0.03301623463630676 + ], + "credit_map": [ + 0.010444381274282932, + 0.015604779124259949, + 0.024570263922214508, + 0.028161881491541862, + 0.02055291272699833, + 0.016944849863648415, + 0.010605677030980587, + 0.026690995320677757, + 0.014985158108174801, + 0.011209842748939991, + 0.01567922718822956, + 0.01446162536740303, + 0.03781941160559654, + 0.03569083288311958, + 0.016516728326678276, + 0.013228225521743298, + 0.0068927938118577, + 0.005892919842153788, + 0.006007912568747997, + 0.0068777576088905334, + 0.005387281067669392, + 0.010976619087159634, + 0.0059898775070905685, + 0.006566901691257954, + 0.009038292802870274, + 0.007106203120201826, + 0.012104755267500877, + 0.006850449368357658, + 0.007285220082849264, + 0.006448578089475632, + 0.004557632375508547, + 0.010998377576470375, + 0.046819351613521576, + 0.01416650228202343, + 0.15410394966602325, + 0.021160541102290154, + 0.02113077975809574, + 0.011309238150715828, + 0.07089754939079285, + 0.011954376474022865, + 0.007669971790164709, + 0.002394885988906026, + 0.007433466613292694, + 0.004121728707104921, + 0.003328037215396762, + 0.0055376687087118626, + 0.0038691379595547915, + 0.00604877807199955, + 0.006573214195668697, + 0.003928018268197775, + 0.00845146831125021, + 0.005850447807461023, + 0.004631607793271542, + 0.0069961873814463615, + 0.010598302818834782, + 0.007251233793795109, + 0.0032281591556966305, + 0.0021711618173867464, + 0.00567464716732502, + 0.0036177090369164944, + 0.002704012207686901, + 0.005788275972008705, + 0.004960501566529274, + 0.008565742522478104, + 0.005913604982197285, + 0.007361854892224073, + 0.005908039398491383, + 0.007896478287875652, + 0.004472623113542795, + 0.007522334810346365, + 0.005398781970143318, + 0.007424983661621809, + 0.0033010109327733517, + 0.00485219107940793, + 0.004059509839862585, + 0.006208507809787989, + 0.0028288394678384066, + 0.003131680656224489, + 0.003824786748737097, + 0.004809705540537834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.009801357984542847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179461359977722 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.018179461359977722, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018179461359977722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06050291657447815, + 0.03502534702420235, + 0.030890749767422676, + 0.037566978484392166, + 0.48273739218711853, + 0.03513467311859131, + 0.05227280035614967, + 0.143550306558609, + 0.07628586143255234, + 0.04603296518325806 + ], + "credit_map": [ + 0.004348054528236389, + 0.004868000280112028, + 0.008935229852795601, + 0.010317000560462475, + 0.006376993376761675, + 0.008716786280274391, + 0.006303591653704643, + 0.010637270286679268, + 0.0033115348778665066, + 0.0023330645635724068, + 0.0033073201775550842, + 0.00224706600420177, + 0.00620277738198638, + 0.008995638228952885, + 0.0034928740933537483, + 0.00513507379218936, + 0.00522128539159894, + 0.0033459519036114216, + 0.0029324982315301895, + 0.004089667461812496, + 0.0026777395978569984, + 0.005296390503644943, + 0.0029458252247422934, + 0.004381396342068911, + 0.005348604172468185, + 0.005321341101080179, + 0.005393818020820618, + 0.0051689306274056435, + 0.005608535371720791, + 0.003686296520754695, + 0.003027232363820076, + 0.004012223333120346, + 0.0786687433719635, + 0.021081646904349327, + 0.21753618121147156, + 0.034427765756845474, + 0.027495715767145157, + 0.018258821219205856, + 0.06995555758476257, + 0.015313005074858665, + 0.006239739712327719, + 0.002319025108590722, + 0.0041013434529304504, + 0.002546396804973483, + 0.0036820177920162678, + 0.005002141464501619, + 0.004645831882953644, + 0.006598183419555426, + 0.00583876296877861, + 0.003631005296483636, + 0.006438895128667355, + 0.006095668766647577, + 0.004396718926727772, + 0.00785269495099783, + 0.010896324180066586, + 0.007122740149497986, + 0.012199189513921738, + 0.009420878253877163, + 0.017528317868709564, + 0.015799017623066902, + 0.01058010570704937, + 0.020274002104997635, + 0.02735884115099907, + 0.03038998320698738, + 0.00576621200889349, + 0.007811584044247866, + 0.009265946224331856, + 0.010394702665507793, + 0.007240439299494028, + 0.006368993315845728, + 0.01518817339092493, + 0.014249817468225956, + 0.003720696782693267, + 0.009421929717063904, + 0.005444840993732214, + 0.008355632424354553, + 0.003037570044398308, + 0.003942318260669708, + 0.005637436639517546, + 0.00647254753857851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.018179461359977722 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191736526787281 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003191736526787281, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191736526787281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032273370772600174, + 0.03998540714383125, + 0.0368257500231266, + 0.09397601336240768, + 0.6291351914405823, + 0.034777622669935226, + 0.03437589481472969, + 0.029223939403891563, + 0.029532108455896378, + 0.03989468514919281 + ], + "credit_map": [ + 0.0026546975132077932, + 0.0036584155168384314, + 0.004725226201117039, + 0.0050154863856732845, + 0.004529979545623064, + 0.003964495379477739, + 0.00231302622705698, + 0.00541204446926713, + 0.0032680060248821974, + 0.0023424061946570873, + 0.004052088130265474, + 0.004112646449357271, + 0.009396529756486416, + 0.00878642126917839, + 0.004135710187256336, + 0.0038915975019335747, + 0.004790663719177246, + 0.00500076450407505, + 0.004194880835711956, + 0.003189916256815195, + 0.004676083568483591, + 0.005908645689487457, + 0.0032834794837981462, + 0.005781315732747316, + 0.012638094834983349, + 0.011747879907488823, + 0.017061026766896248, + 0.009352070279419422, + 0.017107637599110603, + 0.008639714680612087, + 0.00974471215158701, + 0.0076848710887134075, + 0.0945654883980751, + 0.02530970796942711, + 0.2850693166255951, + 0.04235487058758736, + 0.03342939168214798, + 0.02279483713209629, + 0.1067081168293953, + 0.018903512507677078, + 0.006681033875793219, + 0.002135451417416334, + 0.004370453301817179, + 0.0034873047843575478, + 0.0030721439979970455, + 0.00488647073507309, + 0.00399612495675683, + 0.006148639600723982, + 0.0026833578012883663, + 0.002216540277004242, + 0.004839745350182056, + 0.0032410253770649433, + 0.0034864984918385744, + 0.005660206079483032, + 0.007324677892029285, + 0.004923842381685972, + 0.001947971060872078, + 0.001915531698614359, + 0.006157897878438234, + 0.0024241781793534756, + 0.002261048648506403, + 0.0033981844317167997, + 0.004266277886927128, + 0.006852850783616304, + 0.0025189619045704603, + 0.004195036832243204, + 0.0038412194699048996, + 0.004479646682739258, + 0.002819209359586239, + 0.00382693437859416, + 0.003827017033472657, + 0.004024086054414511, + 0.003633641405031085, + 0.011627135798335075, + 0.004982480313628912, + 0.006526063662022352, + 0.0021683538798242807, + 0.0025330570060759783, + 0.0032695895060896873, + 0.005154366139322519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.003191736526787281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028347373008728 + ], + "opposite_logprob_tokens": [ + -2.35028338432312 + ], + "logprob_reward": -0.10028347373008728, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028347373008728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021357610821723938, + 0.01930946484208107, + 0.04376314952969551, + 0.029430318623781204, + 0.758711040019989, + 0.042028799653053284, + 0.009977440349757671, + 0.022196752950549126, + 0.04792021960020065, + 0.005305166356265545 + ], + "credit_map": [ + 0.0016774807590991259, + 0.0019634293857961893, + 0.002884031506255269, + 0.0024489746429026127, + 0.0036204324569553137, + 0.0026663111057132483, + 0.002279342617839575, + 0.003817607183009386, + 0.0015929907094687223, + 0.0010827904334291816, + 0.0015910333022475243, + 0.0010529612191021442, + 0.003318033879622817, + 0.005895023699849844, + 0.0015415732050314546, + 0.003235058393329382, + 0.006306116934865713, + 0.0035317600704729557, + 0.0044867428950965405, + 0.0035746414214372635, + 0.003933098167181015, + 0.01073055062443018, + 0.0036367815919220448, + 0.007563457358628511, + 0.0040773916989564896, + 0.004198460374027491, + 0.005006639752537012, + 0.003255303716287017, + 0.004284332972019911, + 0.0028313121292740107, + 0.0022955769672989845, + 0.003481301013380289, + 0.11785902827978134, + 0.03132607042789459, + 0.3440137505531311, + 0.05255557596683502, + 0.04356048256158829, + 0.02775820530951023, + 0.11834639310836792, + 0.02329152449965477, + 0.007755090482532978, + 0.002563742222264409, + 0.005631215870380402, + 0.004260255489498377, + 0.0035372627899050713, + 0.005371933802962303, + 0.005184783134609461, + 0.007724514696747065, + 0.0009113922715187073, + 0.0006381398416124284, + 0.0013161778915673494, + 0.0010225522564724088, + 0.0011618018615990877, + 0.001478317310102284, + 0.0021879824344068766, + 0.0012610765406861901, + 0.00175150774884969, + 0.0013154143234714866, + 0.0028102961368858814, + 0.0030342552345246077, + 0.0018409049371257424, + 0.0024083589669317007, + 0.003989420365542173, + 0.005046594422310591, + 0.0039025694131851196, + 0.004659346770495176, + 0.0071135954931378365, + 0.008186792023479939, + 0.004127617459744215, + 0.004743760917335749, + 0.007364095654338598, + 0.007822440937161446, + 0.0005258371238596737, + 0.0007120741065591574, + 0.0005222510662861168, + 0.0006710770539939404, + 0.0006659043137915432, + 0.0006522180628962815, + 0.0007608296000398695, + 0.0007949749706313014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.10028347373008728 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, low-angle, 360-degree panning shot closely follows a classic Porsche 911 Carrera RS 3.8, painted a vibrant signal yellow that gleams brilliantly in the harsh afternoon sun. The iconic, massive \"whale tail\" spoiler and the \"Carrera RS\" lettering on the side are outlined in harsh, yet unmistakable black lines, lending the image a rugged, retro, anime-inspired quality. As it drifts through a corner at high speed on a rough, gravel mountain road, the car's entire body experiences a violent jolting, each compression and rebound stroke of the suspension powerful. The wide rear wheels furiously paw at the loose gravel, sending clouds of thick, yellow-brown dust cascading into the air, forming a billowing cloud that nearly obscures the rear half of the car. The backdrop, a jagged gray rock face and sparse, dark green pine trees, contrasts starkly against the almost unrealistic blue of the sky, rendering every element vibrant and vibrant. As the camera moves steadily and gracefully around the car body, light flows on the paint, creating sharp highlights and blocky deep shadows. The entire picture is full of strong visual impact, perfectly capturing this adrenaline-pumping and blood-boiling racing moment.", + "seed": 14217, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_9/index_9445_seed_14217_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05216166814789176, + "vlm_logprob": -0.05216166814789176, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.05216166814789176, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014373810961842537 + ], + "opposite_logprob_tokens": [ + -4.889373779296875 + ], + "logprob_reward": -0.014373810961842537, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.014373810961842537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1793554723262787, + 0.06498479843139648, + 0.04795132577419281, + 0.056433215737342834, + 0.07939410209655762, + 0.06890106946229935, + 0.11782953143119812, + 0.09530824422836304, + 0.22408297657966614, + 0.06575924903154373 + ], + "credit_map": [ + 0.02475270815193653, + 0.01871929131448269, + 0.01375353243201971, + 0.021323271095752716, + 0.03172583132982254, + 0.029652854427695274, + 0.015503283590078354, + 0.023924706503748894, + 0.00901342649012804, + 0.010827096179127693, + 0.00632266141474247, + 0.005674448329955339, + 0.007385925855487585, + 0.01400832086801529, + 0.004684900864958763, + 0.007068028673529625, + 0.006396619137376547, + 0.007227825932204723, + 0.006840516347438097, + 0.009572199545800686, + 0.0039981696754693985, + 0.0048570712096989155, + 0.003578478004783392, + 0.005480451043695211, + 0.007944034412503242, + 0.00907171331346035, + 0.009359443560242653, + 0.004308598581701517, + 0.008511933498084545, + 0.007086276542395353, + 0.005094296298921108, + 0.005056925117969513, + 0.006852003280073404, + 0.004086137283593416, + 0.012757143937051296, + 0.026503127068281174, + 0.005961144343018532, + 0.006514008156955242, + 0.010497227311134338, + 0.006223316304385662, + 0.017272667959332466, + 0.005908881779760122, + 0.010409827344119549, + 0.011257007718086243, + 0.005113346502184868, + 0.006838877219706774, + 0.006480301730334759, + 0.005620163399726152, + 0.006549036130309105, + 0.011323727667331696, + 0.024118958041071892, + 0.03285627439618111, + 0.008334135636687279, + 0.007300386670976877, + 0.01920097880065441, + 0.00814603641629219, + 0.006288515869528055, + 0.008289915509521961, + 0.016558300703763962, + 0.018070900812745094, + 0.010298429057002068, + 0.006458987016230822, + 0.019601408392190933, + 0.00974180456250906, + 0.01282633375376463, + 0.03471244126558304, + 0.04402940720319748, + 0.06256427615880966, + 0.02036431059241295, + 0.01713654398918152, + 0.020364193245768547, + 0.01208548154681921, + 0.00911716092377901, + 0.011757710948586464, + 0.008154098875820637, + 0.013321807608008385, + 0.007076397072523832, + 0.00488530145958066, + 0.005753920413553715, + 0.00569285498932004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.014373810961842537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659903891384602 + ], + "opposite_logprob_tokens": [ + -4.758659839630127 + ], + "logprob_reward": -0.008659903891384602, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008659903891384602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13130857050418854, + 0.11343918740749359, + 0.04596935212612152, + 0.03905007243156433, + 0.07137554883956909, + 0.04638029634952545, + 0.09236382693052292, + 0.10695837438106537, + 0.2931056022644043, + 0.06004917249083519 + ], + "credit_map": [ + 0.01563163287937641, + 0.012538300827145576, + 0.009342222474515438, + 0.012186570093035698, + 0.026786180213093758, + 0.02997405081987381, + 0.011806467548012733, + 0.013043131679296494, + 0.013909554108977318, + 0.01574569195508957, + 0.012482364661991596, + 0.009705749340355396, + 0.009991856291890144, + 0.028266659006476402, + 0.0094802500680089, + 0.01385706290602684, + 0.006511169485747814, + 0.010173878632485867, + 0.004817588720470667, + 0.007580670993775129, + 0.0045605977065861225, + 0.005101905204355717, + 0.002448074985295534, + 0.0047754645347595215, + 0.006907192058861256, + 0.0036564418114721775, + 0.011162061244249344, + 0.0026017294730991125, + 0.0033739465288817883, + 0.005246158689260483, + 0.0034473692066967487, + 0.002655175281688571, + 0.005034194327890873, + 0.004084940068423748, + 0.010833107866346836, + 0.027689049020409584, + 0.003920563496649265, + 0.004868647083640099, + 0.009496296755969524, + 0.005448759067803621, + 0.00775403156876564, + 0.003979278262704611, + 0.006408997345715761, + 0.010786720551550388, + 0.00427046837285161, + 0.00427555525675416, + 0.0050852494314312935, + 0.003819995792582631, + 0.006594076287001371, + 0.012119708582758904, + 0.015385989099740982, + 0.022982710972428322, + 0.007896320894360542, + 0.005987755488604307, + 0.014296007342636585, + 0.007101261522620916, + 0.006831869948655367, + 0.009393628686666489, + 0.019706763327121735, + 0.02319846861064434, + 0.012716087512671947, + 0.007819851860404015, + 0.01930372230708599, + 0.007987987250089645, + 0.018158311024308205, + 0.054286230355501175, + 0.05308232456445694, + 0.07537415623664856, + 0.027421465143561363, + 0.02398708090186119, + 0.025408567860722542, + 0.015387461520731449, + 0.0070843021385371685, + 0.010850494727492332, + 0.006325425114482641, + 0.010140595957636833, + 0.007986445911228657, + 0.005576817784458399, + 0.006198725197464228, + 0.005886364728212357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.008659903891384602 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635086007416248 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008635086007416248, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635086007416248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11366084218025208, + 0.02517283894121647, + 0.023805292323231697, + 0.02385942079126835, + 0.0651392787694931, + 0.04832630231976509, + 0.11777392029762268, + 0.06066043674945831, + 0.43604978919029236, + 0.08555181324481964 + ], + "credit_map": [ + 0.014518852345645428, + 0.00794292613863945, + 0.009894637390971184, + 0.013585347682237625, + 0.016707757487893105, + 0.02474757470190525, + 0.008571218699216843, + 0.01769254542887211, + 0.002735731191933155, + 0.0030247848480939865, + 0.0021750808227807283, + 0.0022340202704072, + 0.00449070381000638, + 0.0037757691461592913, + 0.002945150714367628, + 0.003791599767282605, + 0.00262360624037683, + 0.004600230138748884, + 0.002469909144565463, + 0.004031071905046701, + 0.0034718860406428576, + 0.0027636149898171425, + 0.0016463208012282848, + 0.0021986556239426136, + 0.0055481502786278725, + 0.003929224796593189, + 0.0037241915706545115, + 0.0016148233553394675, + 0.0024912834633141756, + 0.003749086754396558, + 0.0016038228059187531, + 0.0011988405603915453, + 0.005065422039479017, + 0.0035231225192546844, + 0.010660793632268906, + 0.022806596010923386, + 0.004595000762492418, + 0.005368685349822044, + 0.008112267591059208, + 0.005007393658161163, + 0.007535944227129221, + 0.0051779309287667274, + 0.006526012439280748, + 0.010547344572842121, + 0.004348915070295334, + 0.005866341758519411, + 0.0048886812292039394, + 0.0034351383801549673, + 0.008099035359919071, + 0.012271611019968987, + 0.021152224391698837, + 0.03360563889145851, + 0.008797330781817436, + 0.007819772697985172, + 0.017586449161171913, + 0.008441870100796223, + 0.00497895572334528, + 0.005819510668516159, + 0.010849207639694214, + 0.014788839966058731, + 0.008056876249611378, + 0.0035084893461316824, + 0.009748820215463638, + 0.002909741597250104, + 0.031346485018730164, + 0.08652784675359726, + 0.07125458121299744, + 0.10464955866336823, + 0.04186023399233818, + 0.037259142845869064, + 0.04033629596233368, + 0.02281567081809044, + 0.010141165927052498, + 0.010250789113342762, + 0.014192886650562286, + 0.022106841206550598, + 0.006727449130266905, + 0.00515849981456995, + 0.010203411802649498, + 0.006770779844373465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.008635086007416248 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036139669828116894 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036139669828116894, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036139669828116894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10786920040845871, + 0.03224215656518936, + 0.047654397785663605, + 0.07309027761220932, + 0.06389133632183075, + 0.05961427092552185, + 0.2443641722202301, + 0.15927737951278687, + 0.17471757531166077, + 0.03727922961115837 + ], + "credit_map": [ + 0.01229813601821661, + 0.009315340779721737, + 0.009337291121482849, + 0.012570166029036045, + 0.016581684350967407, + 0.01795298233628273, + 0.008701147511601448, + 0.021112460643053055, + 0.0036193111445754766, + 0.004931851755827665, + 0.0020266715437173843, + 0.0031209366861730814, + 0.0035750707611441612, + 0.006174631416797638, + 0.004040705971419811, + 0.004752976819872856, + 0.004917844198644161, + 0.010206300765275955, + 0.0055429707281291485, + 0.008909818716347218, + 0.0050391643308103085, + 0.004846989642828703, + 0.0032037291675806046, + 0.00498758303001523, + 0.015190556645393372, + 0.00966587383300066, + 0.014729087240993977, + 0.004891184624284506, + 0.007086115423589945, + 0.012039409019052982, + 0.00505472719669342, + 0.0044333250261843204, + 0.005348681937903166, + 0.003709654323756695, + 0.009460506029427052, + 0.022153768688440323, + 0.004599005449563265, + 0.005681370850652456, + 0.008310365490615368, + 0.004627990536391735, + 0.005457765888422728, + 0.005119325127452612, + 0.0092215146869421, + 0.017535585910081863, + 0.007042920682579279, + 0.00400547543540597, + 0.007173458114266396, + 0.004058223217725754, + 0.014000029303133488, + 0.023778919130563736, + 0.04257936030626297, + 0.07228110730648041, + 0.026592714712023735, + 0.017236342653632164, + 0.032398853451013565, + 0.015496846288442612, + 0.014191126450896263, + 0.015570680610835552, + 0.02882910892367363, + 0.042807210236787796, + 0.02081650123000145, + 0.011267522349953651, + 0.017681961879134178, + 0.008113261312246323, + 0.00986025482416153, + 0.025451676920056343, + 0.03305643051862717, + 0.04722912237048149, + 0.01726391166448593, + 0.013420584611594677, + 0.018150154501199722, + 0.010285435244441032, + 0.0049461377784609795, + 0.0047636013478040695, + 0.005090468097478151, + 0.008111817762255669, + 0.0035709154326468706, + 0.004509427584707737, + 0.0035872419830411673, + 0.0026996235828846693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.0036139669828116894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22552557289600372 + ], + "opposite_logprob_tokens": [ + -1.6005256175994873 + ], + "logprob_reward": -0.22552557289600372, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22552557289600372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06421735137701035, + 0.06893233209848404, + 0.08540906757116318, + 0.04296359419822693, + 0.11738330870866776, + 0.08115027844905853, + 0.2367623895406723, + 0.07242877781391144, + 0.1864459067583084, + 0.04430701211094856 + ], + "credit_map": [ + 0.007409435696899891, + 0.005611903965473175, + 0.008849434554576874, + 0.008262881077826023, + 0.007111833430826664, + 0.012644704431295395, + 0.006883019581437111, + 0.00744414608925581, + 0.008915645070374012, + 0.007730251178145409, + 0.006136439274996519, + 0.0069793518632650375, + 0.00801452063024044, + 0.01453996542841196, + 0.007851450704038143, + 0.00876470748335123, + 0.008634601719677448, + 0.010370281524956226, + 0.012400097213685513, + 0.01843472756445408, + 0.010324631817638874, + 0.010356210172176361, + 0.005648714490234852, + 0.009239804930984974, + 0.010045676492154598, + 0.006499792914837599, + 0.005283030215650797, + 0.0037604968529194593, + 0.003606726648285985, + 0.006582287140190601, + 0.00357457366771996, + 0.0036110123619437218, + 0.011437919922173023, + 0.007275470998138189, + 0.013391600921750069, + 0.03256223350763321, + 0.011055098846554756, + 0.018280193209648132, + 0.01437927782535553, + 0.009001517668366432, + 0.013842722401022911, + 0.009102036245167255, + 0.010354731231927872, + 0.01377926766872406, + 0.011898324824869633, + 0.007525057066231966, + 0.008649961091578007, + 0.005998177919536829, + 0.0143346032127738, + 0.021026592701673508, + 0.04232554882764816, + 0.06943931430578232, + 0.021699856966733932, + 0.011107956990599632, + 0.0398944690823555, + 0.016934029757976532, + 0.010240613482892513, + 0.00918100867420435, + 0.012203126214444637, + 0.012751465663313866, + 0.007382974959909916, + 0.0061553469859063625, + 0.010508803650736809, + 0.004005438648164272, + 0.008546940982341766, + 0.017494451254606247, + 0.04195264354348183, + 0.05403626337647438, + 0.01605982519686222, + 0.013983038254082203, + 0.021006185561418533, + 0.013366566970944405, + 0.006405504886060953, + 0.007598112802952528, + 0.004084365908056498, + 0.008308804593980312, + 0.004678062628954649, + 0.0035582587588578463, + 0.003583379555493593, + 0.006090521812438965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.22552557289600372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, low-angle, 360-degree panning shot closely follows a classic Porsche 911 Carrera RS 3.8, painted a vibrant signal yellow that gleams brilliantly in the harsh afternoon sun. The iconic, massive \"whale tail\" spoiler and the \"Carrera RS\" lettering on the side are outlined in harsh, yet unmistakable black lines, lending the image a rugged, retro, anime-inspired quality. As it drifts through a corner at high speed on a rough, gravel mountain road, the car's entire body experiences a violent jolting, each compression and rebound stroke of the suspension powerful. The wide rear wheels furiously paw at the loose gravel, sending clouds of thick, yellow-brown dust cascading into the air, forming a billowing cloud that nearly obscures the rear half of the car. The backdrop, a jagged gray rock face and sparse, dark green pine trees, contrasts starkly against the almost unrealistic blue of the sky, rendering every element vibrant and vibrant. As the camera moves steadily and gracefully around the car body, light flows on the paint, creating sharp highlights and blocky deep shadows. The entire picture is full of strong visual impact, perfectly capturing this adrenaline-pumping and blood-boiling racing moment.", + "seed": 14218, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000166/group_9/index_9445_seed_14218_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08219227204099297, + "vlm_logprob": -0.08219227204099297, + "vlm_yes_no_margin": 3.624999976158142, + "vlm_token_credit": -0.08219227204099297, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019284849986433983 + ], + "opposite_logprob_tokens": [ + -4.394284725189209 + ], + "logprob_reward": -0.019284849986433983, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.019284849986433983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11367145925760269, + 0.040949784219264984, + 0.06362023204565048, + 0.18771393597126007, + 0.052049607038497925, + 0.059762123972177505, + 0.1821897327899933, + 0.07769810408353806, + 0.09443188458681107, + 0.12791310250759125 + ], + "credit_map": [ + 0.019848419353365898, + 0.013920385390520096, + 0.012353014200925827, + 0.012151473201811314, + 0.01341584138572216, + 0.017613934352993965, + 0.01033006701618433, + 0.014038334600627422, + 0.0071250093169510365, + 0.005989112891256809, + 0.004899826366454363, + 0.004404197447001934, + 0.005775973666459322, + 0.00475865975022316, + 0.0034399766009300947, + 0.00455703167244792, + 0.01213454082608223, + 0.004572381731122732, + 0.006090050563216209, + 0.008434218354523182, + 0.005810583010315895, + 0.006654834840446711, + 0.005664735101163387, + 0.01425889227539301, + 0.016009598970413208, + 0.014010762795805931, + 0.00961619894951582, + 0.05931977182626724, + 0.00934469886124134, + 0.018207047134637833, + 0.019183486700057983, + 0.04202238470315933, + 0.0061499797739088535, + 0.009597421623766422, + 0.004498685710132122, + 0.00582406111061573, + 0.01013695914298296, + 0.005733660887926817, + 0.004981501027941704, + 0.005127338692545891, + 0.008414005860686302, + 0.008939280174672604, + 0.0060269697569310665, + 0.007909287698566914, + 0.0078286649659276, + 0.005615246016532183, + 0.00839913822710514, + 0.006629538722336292, + 0.017707742750644684, + 0.016577675938606262, + 0.01137113943696022, + 0.010572243481874466, + 0.010475152172148228, + 0.032730832695961, + 0.02071303129196167, + 0.06204194203019142, + 0.015234200283885002, + 0.004843906033784151, + 0.018972091376781464, + 0.005659981165081263, + 0.005295579321682453, + 0.0060374122112989426, + 0.01182603556662798, + 0.009828910231590271, + 0.01076667569577694, + 0.010811874642968178, + 0.013917602598667145, + 0.01100120134651661, + 0.009835752658545971, + 0.01198724564164877, + 0.010260919108986855, + 0.015850622206926346, + 0.009683915413916111, + 0.024265818297863007, + 0.022111160680651665, + 0.01643608883023262, + 0.0119019141420722, + 0.014473910443484783, + 0.011227450333535671, + 0.017812853679060936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.019284849986433983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009799704886972904 + ], + "opposite_logprob_tokens": [ + -4.634799480438232 + ], + "logprob_reward": -0.009799704886972904, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009799704886972904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12834280729293823, + 0.06379909068346024, + 0.0718119665980339, + 0.24547743797302246, + 0.0353102944791317, + 0.054788243025541306, + 0.06766922026872635, + 0.09402338415384293, + 0.17998038232326508, + 0.05879717692732811 + ], + "credit_map": [ + 0.02579534985125065, + 0.012872147373855114, + 0.014377825893461704, + 0.013889968395233154, + 0.009916733019053936, + 0.0200481079518795, + 0.011352636851370335, + 0.020090030506253242, + 0.011097380891442299, + 0.007407845463603735, + 0.007029680069535971, + 0.008076785132288933, + 0.007198987063020468, + 0.006335541605949402, + 0.005533722694963217, + 0.011119143106043339, + 0.008615363389253616, + 0.0058304728008806705, + 0.006478609051555395, + 0.025837751105427742, + 0.004698717966675758, + 0.0063694193959236145, + 0.005932480096817017, + 0.008049143478274345, + 0.024544907733798027, + 0.040403928607702255, + 0.013746513053774834, + 0.048361796885728836, + 0.019733041524887085, + 0.03209864720702171, + 0.020438645035028458, + 0.04614994674921036, + 0.005424547009170055, + 0.008188100531697273, + 0.0023011311423033476, + 0.004178396891802549, + 0.004740472417324781, + 0.0034008913207799196, + 0.0037652032915502787, + 0.0033115497790277004, + 0.006642740685492754, + 0.012014026753604412, + 0.003989566117525101, + 0.012143869884312153, + 0.005317470990121365, + 0.004804625175893307, + 0.005470350384712219, + 0.004405587911605835, + 0.008550125174224377, + 0.005567551124840975, + 0.00423724390566349, + 0.003413184080272913, + 0.005365687422454357, + 0.012798754498362541, + 0.006725320592522621, + 0.0210113488137722, + 0.02051379345357418, + 0.006583245936781168, + 0.022572694346308708, + 0.006050734780728817, + 0.005010799504816532, + 0.005682863760739565, + 0.015753554180264473, + 0.01185569167137146, + 0.017214348539710045, + 0.014506257139146328, + 0.027910219505429268, + 0.02345866896212101, + 0.020245879888534546, + 0.018672771751880646, + 0.030341822654008865, + 0.027630388736724854, + 0.006058342754840851, + 0.00894953217357397, + 0.0069126239977777, + 0.010388567112386227, + 0.007543759886175394, + 0.005841549020260572, + 0.006498380564153194, + 0.006604418158531189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.009799704886972904 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25205329060554504 + ], + "opposite_logprob_tokens": [ + -1.5020532608032227 + ], + "logprob_reward": -0.25205329060554504, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25205329060554504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09461568295955658, + 0.043650101870298386, + 0.09880202263593674, + 0.08075006306171417, + 0.029086334630846977, + 0.047035254538059235, + 0.04228712618350983, + 0.11260659247636795, + 0.314156174659729, + 0.13701066374778748 + ], + "credit_map": [ + 0.014640138484537601, + 0.009773639030754566, + 0.010222864337265491, + 0.009382735937833786, + 0.00754487793892622, + 0.009654587134718895, + 0.010105765424668789, + 0.02329106628894806, + 0.007341891061514616, + 0.005371628794819117, + 0.0047638448886573315, + 0.005838586017489433, + 0.004903140477836132, + 0.00456055486574769, + 0.0038370948750525713, + 0.007033357862383127, + 0.0120103619992733, + 0.006643710192292929, + 0.007767674047499895, + 0.04380317032337189, + 0.004135361406952143, + 0.006687598302960396, + 0.009535498917102814, + 0.008218644186854362, + 0.006052584853023291, + 0.007516462821513414, + 0.004406173713505268, + 0.031930893659591675, + 0.004693551920354366, + 0.005527863744646311, + 0.010592170059680939, + 0.010030360892415047, + 0.005084847565740347, + 0.005901321768760681, + 0.0023279248271137476, + 0.00315688899718225, + 0.004020575433969498, + 0.0024368194863200188, + 0.002553465310484171, + 0.0036044896114617586, + 0.007188688963651657, + 0.012854460626840591, + 0.003193566808477044, + 0.006424435414373875, + 0.005792005453258753, + 0.0034168330021202564, + 0.0038397791795432568, + 0.004325484856963158, + 0.005624113138765097, + 0.0033071471843868494, + 0.004300951026380062, + 0.004300690721720457, + 0.0023235559929162264, + 0.00347115658223629, + 0.0048305983655154705, + 0.014128913171589375, + 0.02104360982775688, + 0.008945178240537643, + 0.02478082850575447, + 0.008492004126310349, + 0.007101176306605339, + 0.0077508967369794846, + 0.01456074882298708, + 0.019932150840759277, + 0.030085189267992973, + 0.03480834141373634, + 0.037969641387462616, + 0.03881761431694031, + 0.029633233323693275, + 0.05273619294166565, + 0.0404944121837616, + 0.04961152747273445, + 0.013477342203259468, + 0.018822895362973213, + 0.021332086995244026, + 0.021623646840453148, + 0.009253437630832195, + 0.012618969194591045, + 0.011154690757393837, + 0.028727587312459946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.25205329060554504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816402819007635 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816402819007635, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816402819007635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1497175693511963, + 0.07743395119905472, + 0.03150078281760216, + 0.27219530940055847, + 0.06878848373889923, + 0.03733547031879425, + 0.019341284409165382, + 0.03159203752875328, + 0.18050327897071838, + 0.13159184157848358 + ], + "credit_map": [ + 0.028307823464274406, + 0.013366472907364368, + 0.01703992486000061, + 0.01207585446536541, + 0.009843984618782997, + 0.032474350184202194, + 0.011319134384393692, + 0.02529001608490944, + 0.011357401497662067, + 0.00957441609352827, + 0.009690389968454838, + 0.009020758792757988, + 0.008388474583625793, + 0.008670850656926632, + 0.005500108934938908, + 0.015231546945869923, + 0.004579400643706322, + 0.0032697219867259264, + 0.003705733921378851, + 0.00394404586404562, + 0.004158752504736185, + 0.003654157742857933, + 0.002464050194248557, + 0.005724916700273752, + 0.024385781958699226, + 0.018646402284502983, + 0.014053001999855042, + 0.08623337000608444, + 0.012300748378038406, + 0.024113696068525314, + 0.027287323027849197, + 0.06517497450113297, + 0.00954851321876049, + 0.017304377630352974, + 0.005022452678531408, + 0.007216398138552904, + 0.012637009844183922, + 0.005702487193048, + 0.004821049980819225, + 0.006536186672747135, + 0.004220154136419296, + 0.006329095922410488, + 0.003186470828950405, + 0.008743911981582642, + 0.0037480078171938658, + 0.0033673702273517847, + 0.004430017899721861, + 0.003310436848551035, + 0.004196326248347759, + 0.0023776069283485413, + 0.0022963036317378283, + 0.0013870998518541455, + 0.001886969548650086, + 0.0018016444519162178, + 0.002943126019090414, + 0.002452206565067172, + 0.004051695112138987, + 0.003162149339914322, + 0.005436195991933346, + 0.003642435185611248, + 0.0029832879081368446, + 0.002318275859579444, + 0.004641229752451181, + 0.005356765352189541, + 0.016659948974847794, + 0.019341181963682175, + 0.024422839283943176, + 0.02976101264357567, + 0.019201789051294327, + 0.013914166949689388, + 0.03331706300377846, + 0.02388525754213333, + 0.018532125279307365, + 0.02181948535144329, + 0.015623755753040314, + 0.017515376210212708, + 0.015888173133134842, + 0.016846567392349243, + 0.009818083606660366, + 0.015548263676464558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.002816402819007635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic Porsche 911 Carrera RS 3.8 painted in vibrant signal yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car feature a massive 'whale tail' spoiler and 'Carrera RS' lettering outlined in black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car drifting at high speed on a rough gravel mountain road, kicking up thick yellow-brown dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background consist of a jagged gray rock face, sparse dark green pine trees, and an unrealistic blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree panning shot that moves steadily around the car body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270071119070053 + ], + "opposite_logprob_tokens": [ + -2.127007007598877 + ], + "logprob_reward": -0.1270071119070053, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270071119070053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07592427730560303, + 0.03293420374393463, + 0.11094938218593597, + 0.20513083040714264, + 0.07232443243265152, + 0.07078652828931808, + 0.10565881431102753, + 0.05659421533346176, + 0.20299141108989716, + 0.06670593470335007 + ], + "credit_map": [ + 0.010321537964046001, + 0.007220374885946512, + 0.009840131737291813, + 0.0065438030287623405, + 0.008952109143137932, + 0.013574210926890373, + 0.006857272703200579, + 0.012614840641617775, + 0.004924597684293985, + 0.004408808425068855, + 0.004607052076607943, + 0.00441387016326189, + 0.0035639952402561903, + 0.0034346957691013813, + 0.002867564093321562, + 0.0047136214561760426, + 0.014195745810866356, + 0.01009285543113947, + 0.010049926117062569, + 0.03810908645391464, + 0.009348313324153423, + 0.00811847485601902, + 0.008800037205219269, + 0.012234947644174099, + 0.017227087169885635, + 0.0141953956335783, + 0.01032180618494749, + 0.06628555804491043, + 0.008955768309533596, + 0.017741840332746506, + 0.02113630808889866, + 0.049267061054706573, + 0.013040319085121155, + 0.015632856637239456, + 0.006100103259086609, + 0.00938277319073677, + 0.008599468506872654, + 0.006486484315246344, + 0.005604302044957876, + 0.007478124927729368, + 0.008344344794750214, + 0.009785916656255722, + 0.006838609464466572, + 0.012751088477671146, + 0.009236159734427929, + 0.006469491869211197, + 0.009390891529619694, + 0.007970019243657589, + 0.01510972436517477, + 0.009087816812098026, + 0.008782311342656612, + 0.005837826989591122, + 0.0068006874062120914, + 0.016315286979079247, + 0.010818001814186573, + 0.032907161861658096, + 0.010055589489638805, + 0.005325570702552795, + 0.010694747790694237, + 0.004842808004468679, + 0.0035995824728161097, + 0.0045353868044912815, + 0.009474854916334152, + 0.008065671660006046, + 0.021491238847374916, + 0.024469058960676193, + 0.026128560304641724, + 0.02688986249268055, + 0.02247351035475731, + 0.019175242632627487, + 0.030737657099962234, + 0.03162628784775734, + 0.004919051192700863, + 0.006478427443653345, + 0.011080394499003887, + 0.007753902580589056, + 0.006273340433835983, + 0.008610253222286701, + 0.006615182384848595, + 0.01497538760304451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.1270071119070053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file